Datasets:

ArmelR

/

the-pile-splitted

like 20

--- abstract: 'Superparamagnetic tunnel junctions are nanostructures that auto-oscillate stochastically under the effect of thermal noise. Recent works showed that despite their stochasticity, such junctions possess a capability to synchronize to subthreshold voltage drives, in a way that can be enhanced or controlled by adding noise. In this work, we investigate a system composed of two electrically coupled junctions, connected in series to a periodic voltage source. We make use of numerical simulations and of an analytical model to demonstrate that both junctions can be phase-locked to the drive, in phase or in anti-phase. This synchronization phenomenon can be controlled by both thermal and electrical noises, although the two types of noises induce qualitatively different behaviors. Namely, thermal noise can stabilize a regime where one junction is phase-locked to the drive voltage while the other is blocked in one state. On the contrary, electrical noise causes the junctions to have highly correlated behaviors and thus cannot induce the latter. These results open the way for the design of superparamagnetic tunnel junctions that can perform computation through synchronization, and which harvest the largest part of their energy consumption from thermal noise.' author: - 'A. Mizrahi' - 'N. Locatelli' - 'J. Grollier' - 'D. Querlioz' bibliography: - 'Zotero.bib' title: Synchronization of electrically coupled stochastic magnetic oscillators induced by thermal and electrical noise --- Introduction ============ Superparamagnetic tunnel junctions are stochastic auto-oscillators, powered by thermal noise [@rippard_thermal_2011]. They naturally transform the fluctuations of thermal noise into a large telegraphic signal that can be read through a simple measurement of their electrical resistance. Furthermore, they have the ability to exhibit the phenomenon of stochastic resonance in which noise enables a weak signal to be detected by a non linear system [@cheng_nonadiabatic_2010; @finocchio_micromagnetic_2011; @daquino_stochastic_2011]. In previous works, we showed how they can also synchronize to a weak periodic voltage through both thermal noise and electrical noise on the drive [@locatelli_noise-enhanced_2014; @mizrahi_controlling_2016]. The capacity of superparamagnetic tunnel junctions to harvest the energy of noise gives the opportunity to study their physics under many angles, from sensing to low-power computing. This requires understanding how these noise-induced phenomena occur in systems composed of several coupled superparamagnetic tunnel junctions. Noise-induced synchronization of coupled elements has been studied for several systems, such as arrays of generic nonlinear elements [@neiman_stochastic_1995; @lindner_array_1995; @marchesoni_spatiotemporal_1996] or linear chains of diffusely coupled diode resonators [@locher_stochastic_1998], but their analysis does not apply to our specific device as they consider linear coupling and collective phenomena. Furthermore, these works focus on noise-induced signal power amplification but do not study frequency-locking and phase-locking directly. In this work, we present a theoretical study of noise-induced synchronization of two electrically coupled superparamagnetic tunnel junctions. Our comprehensive study is based on numerical simulations. We examine frequency-locking through the evolution with noise of the junctions frequencies as well as phase-locking through time-resolved traces of the junctions resistances. All results are discussed with the support of an analytical model based on the evolution with noise of the junctions switching probabilities. Both numerical and analytical models were previously validated by direct comparison to experiments in the case of the synchronization of a single junction [@mizrahi_magnetic_2015; @mizrahi_controlling_2016]. We examine the differences between the synchronization phenomena induced by thermal noise versus those induced by electrical noise on the drive voltage. After presenting our model in the case of a single superparamagnetic tunnel junction, we apply it to two junctions connected in series to a periodic voltage source. We show that both junctions can be synchronized simultaneously to the voltage drive. Both in-phase-locking and anti-phase-locking configurations can be achieved by choosing the connections of the junctions in the circuit. These phenomena can be induced and controlled by the amplitudes of thermal and electrical noises. Specifically, we show that thermal noise is able to stabilize a hybrid synchronization regime where at each oscillation of the drive, one randomly chosen junction is phase-locked to the drive while the other is blocked in one state. On the other hand, this regime cannot be achieved by the addition of electrical noise. The electrical noises on the two junctions both come from the drive and are thus correlated. The switches of the junctions are simultaneous and triggered by large voltage events. Noise induced synchronization of a single junction ================================================== A magnetic tunnel junction is composed of two ferromagnets (one with pinned magnetization and one with free magnetization) and a tunnel barrier sandwiched between them (Fig.\[junction\](a)). The magnetization of the free layer has two stable states: parallel (P) and anti-parallel (AP) to the magnetization of the pinned layer. The anti-parallel state has a higher electrical resistance than the parallel state. Magnetic tunnel junctions can be engineered so that the energy barrier between the P and AP states is low enough to reach the superparamagnetic regime. This means that the intrinsic thermal noise in the junction induces random, spontaneous switches of the magnetization between the P and AP states. The superparamagnetic tunnel junction behaves as a noise-powered stochastic oscillator [@rippard_thermal_2011; @locatelli_noise-enhanced_2014] – as depicted in Fig.\[junction\](d-e) – which does not need any external supply of energy to run. The phenomenon of spin-transfer torque [@slonczewski_current-driven_1996; @berger_emission_1996] provides a handle over the junction as a positive voltage stabilizes the AP state (Fig.\[junction\](b) and (d)) while a negative voltage stabilizes the P state (Fig.\[junction\](c) and (e)). The applied dc voltage is defined as positive from the pinned layer to the free layer. Thermal energy destabilizes both states.\  Methods ------- Making the assumption that the free layer of the junction can be considered as a single domain magnetization element, the probabilities to leave the AP and P states within a duration $t$ can be described by a Poisson process: $$\begin{aligned} P_{AP\rightarrow P}(V,t) &= 1-\exp\left(-\phi_{AP}(V)~t\right),\\ P_{P\rightarrow AP}(V,t) &= 1-\exp\left(-\phi_{P}(V)~t\right). \label{eq_PP}\end{aligned}$$ The escape rates $\phi_{AP(P)}$ are described by the Neel-Brown model [@brown_thermal_1963], and follow the Arrhenius equations. They can be tuned by the voltage applied to the junction, through the effect of spin-transfer torque [@slonczewski_excitation_1999; @li_thermally_2004]: $$\begin{aligned} \phi_{AP}(V) &= \phi_0~\exp\left(-\frac{\Delta E}{k_BT}\left(1+\frac{V}{V_c}\right)\right),\\ \phi_{P}(V) &= \phi_0~\exp\left(-\frac{\Delta E}{k_BT}\left(1-\frac{V}{V_c}\right)\right), \label{eq_tauP}\end{aligned}$$ where $\phi_0=10^9\,\mathrm{Hz}$ is the effective attempt rate [@rippard_thermal_2011], $\Delta E$ is the energy barrier between the two stable states, $k_B$ is the Boltzmann constant, $T$ is the temperature, $V$ is the voltage across the junctions and $V_c$ is the threshold voltage for deterministic switching at 0K [@li_thermally_2004].\ We consider in the following of this article junctions with an energy barrier and temperature such that $\Delta E=30\,k_B T$ at $T=50\mathrm{K}$ and a threshold voltage $V_c=1\,\mathrm{V}$. First, a junction is submitted to a square periodic voltage with frequency $F_{ac}=1/T_{ac}=1\,\mathrm{MHz}$ and with subthreshold amplitude $V~=~0.75\mathrm{V}$. We perform numerical simulations of the system as follows. At each time step of $dt=10\mathrm{ns}$, the probability for the junction to switch from one state to the other is computed through equations (1) to (4). A uniform pseudo random number between 0 and 1 is generated. If the random number is lower than the switching probability, the junction changes state. Thermal noise induced synchronization \[syncT\] ----------------------------------------------- Synchronization of the junction to the weak periodic signal can be controlled by different sources of noise. We first study the influence of thermal noise. Fig. \[OneSync\](a) represents the frequency of the junction – defined as the mean number of oscillations of the junction’s state per second – versus temperature. The frequency increases with temperature and presents a plateau at the frequency of the drive $F_{ac}$. This plateau is a signature of stochastic resonance and noise-induced synchronization [@neiman_stochastic_1998; @freund_frequency_2003].  The junction is considered to be in-phase with the drive if it is in the AP state when the drive is +V and in the P state when the drive is -V. In consequence we can define $P_+$ the probability to phase-lock to the drive and $P_-$ the probability to phase-unlock: $$\begin{aligned} P_+ &= P_{P\rightarrow AP}\left(+V,\frac{T_{ac}}{2}\right)=P_{AP\rightarrow P}\left(-V,\frac{T_{ac}}{2}\right) \nonumber \\ &= 1-\exp\left(-\frac{T_{ac}}{2}\phi_+\right) \end{aligned}$$ $$\begin{aligned} P_- &= P_{P\rightarrow AP}\left(-V,\frac{T_{ac}}{2}\right)=P_{AP\rightarrow P}\left(+V,\frac{T_{ac}}{2}\right) \nonumber \\ &= 1-\exp\left(-\frac{T_{ac}}{2}\phi_-\right) \end{aligned}$$ where $$\phi_+=\phi_0~\exp\left(-\frac{\Delta E}{k_BT}\left(1-\frac{V}{V_c}\right)\right)$$ $$\phi_-=\phi_0~\exp\left(-\frac{\Delta E}{k_BT}\left(1+\frac{V}{V_c}\right)\right).$$ For each temperature, the probabilities $P_+$ and $P_-$ are computed analytically and plotted in Fig.\[OneSync\]a.\ At low temperatures (below 85K), the frequency of the junction is lower than the drive frequency $F_{ac}$. Indeed the drive is too weak to induce synchronization by itself because its amplitude is sub-threshold ($V=0.75V_c$).\ We investigate the conditions leading to the occurrence of synchronization. When the junction and the drive are out-of-phase, they have a probability $P_+$ to phase-lock in the next half-period $\frac{T_{ac}}{2}$ (before the next switch of the drive). When the junction and the drive are in-phase, they have a probability $1-P_-$ to stay phase-locked during the next half-period. As the junction is intrinsically stochastic, there cannot be perfect deterministic synchronization. We consider the junction to be synchronized to the drive for $P_+>0.99$ and $P_-<0.01$. This corresponds to the temperature range $85\mathrm{K}<T<241\mathrm{K}$ and indeed we observe that it matches the visual boundaries of the plateau where the frequency of the junction is equal to the frequency of the drive. Above $T=241\mathrm{K}$, the probability to phase-unlock $P_-$ is higher than 0.01. We observe that the junction regularly slips out of phase. These parasitic oscillations – here called glitches – raise the frequency of the junction above $F_{ac}$ and destroy synchronization. Electrical noise induced synchronization \[syncN\] -------------------------------------------------- Synchronization can also be induced by electrical noise added to the drive signal. We apply a voltage $U(t)=V_{ac}(t)+N(t)$ to the junction. $V_{ac}(t)$ is a square periodic drive of frequency $F_{ac}=1~\mathrm{MHz}$ and amplitude $V=0.75 \mathrm{V}$. $N(t)$ is a white Gaussian noise with standard deviation $\sigma_{noise}$ and cutoff frequency $F_{noise}=100\mathrm{MHz}$. The escape rates $\phi_+$ and $\phi_-$ are time dependent random variables: $$\phi_+(t)=\phi_0~\exp\left(-\frac{\Delta E}{k_BT}\left(1-\frac{V+N(t)}{V_c}\right)\right)$$ $$\phi_-(t)=\phi_0~\exp\left(-\frac{\Delta E}{k_BT}\left(1+\frac{V+N(t)}{V_c}\right)\right)$$ In this case, the switching probabilities need to be averaged over all possible values of $N$. The average probabilities to phase-lock and unlock can be computed as follows: $$\langle P_{\pm}\rangle =1-\Bigg(\int_{-\infty}^{+\infty} \left(1-\exp\left(-\delta t\phi_0\exp\left(-\frac{\Delta E}{kT}\left(1\pm\frac{V+N}{Vc}\right)\right)\right)\right)\psi (N)dN\Bigg)^{\frac{T_{ac}}{2\delta t}},$$ where $\delta t=1/F_{noise}$ is the smallest time scale of the electrical noise and $\psi (N)$ is a Gaussian distribution with standard deviation $\sigma_{noise}$.\ In Fig.\[OneSync\]b, the frequency of the junction obtained by numerical simulations as well as the average probabilities $\langle P_+\rangle $ and $\langle P_-\rangle $ are plotted versus the noise’s standard deviation $\sigma_{noise}$. The frequency exhibits again a plateau at $F_{ac}$. Noise-induced synchronization with the drive $V_{ac}$ is achieved for an optimal range of electrical noise $0.08V<\sigma_{noise}<0.39V$ which corresponds to $\langle P _+\rangle >0.99$ and $\langle P_-\rangle <0.01$.\ In conclusion, thermal and electrical noises induce nearly identical behaviors in the single junction situation. For an optimal range of noise, the superparamagnetic tunnel junction is synchronized to the voltage drive. Noise-induced synchronization of two electrically coupled junctions \[2sync\] ============================================================================= We now use our model to study the synchronization of coupled oscillators. We consider two identical junctions connected in series. The voltages $V_{1}$ and $V_{2}$ applied to each junction depend on the states of both junctions: $$V_{1,2}=V_{ac}\frac{R_{1,2}}{R_1+R_2},$$ where $R_{1,2}$ are the respective resistances of the junctions. When the pinned layer of junction 1 is connected to the free layer of junction 2 the junctions are said to be head to tail. On the contrary, when the pinned layer of junction 1 is connected to the pinned layer of junction 2 the junctions are said to be head to head. Thermal noise induced synchronization \[2syncT\] ------------------------------------------------  In this section we study the case where only thermal noise is involved. We consider two identical junctions which electrical resistances are such that $R_{AP}=4R_P$, with $R_{AP}$ ($R_P$) the resistance corresponding to the AP (P) state. The corresponding tunnel magneto-resistance is $TMR=\frac{R_{AP}-R_P}{R_P}=300\%$. We apply a square periodic voltage of frequency $F_{ac}=10\mathrm{kHz}$ and amplitude $V_{ac}=1.5\mathrm{V}$ to the junctions.\ ### Junctions connected head to tail \[2syncTs\] The numerical simulation is conducted in the same way as the single junction study. A random number is generated for each junction. If both junctions have a random number lower than their switching probability, the precise times of switching corresponding to each random number are computed. Two cases arise. Either the times of switching are separated by less than 1 ns and both junctions switch simultaneously (1 ns is the typical duration of a switch [@sun_spin_2006]). Either the times of switching are separated by more than 1 ns, in which case the junction with the lowest switching time switches first. Then the switching time of the other junction has to be recomputed, because its switching probability is different now that the first junction has switched, modifying the applied voltage.\ The frequencies of both junctions are plotted in black in Figure \[TwoSyncTh\]. They are superimposed for all temperatures. The plateau at the frequency of the drive, characteristic of synchronization, is present between $T=111\mathrm{K}$ and $T=170\mathrm{K}$. Surprisingly, we observe the existence of a second plateau at half the frequency of the drive, between $T=40\mathrm{K}$ and $T=68\mathrm{K}$ . Figure \[Schema\]A presents the evolution with time of the voltage drive and the resistances of both junctions. In panels (d), (e) and (f) the temperature is 130K which corresponds to the plateau at the drive frequency. As expected, both junctions are locked in phase with the drive. In panels (a), (b) and (c) the temperature is 60K which corresponds to the plateau at half the drive frequency. We observe that, at each period of the drive, one junction is locked in phase with the drive while the other is blocked in the P state. The phase-locked junction alternates randomly from period to period.\ In the Appendix, we use further numerical simulations and our analytical model to investigate the influence of the tunnel magneto-resistance and the drive frequency on the existence and width of the plateaus. We observe that a high tunnel magneto-resistance widens the plateau at $F_{ac}/2$ and narrows the plateau at $F_{ac}$. A $300\%$ TMR is a good illustration of the effect of the coupling while maintaining a realistic tunnel magneto-resistance. A drive frequency of $F_{ac}=10kHz$ enables the system to exhibit two significant synchronization plateaus, at $F_{ac}$ and $F_{ac}/2$. To get a quantitative understanding of this phenomenon we define six different probabilities which correspond to the possible switching probabilities of the junctions for the different drive and resistances configurations: - $P_1=P_{AP\rightarrow P}(-\frac{4V}{5})$ - $P_2=P_{AP\rightarrow P}(-\frac{V}{2})=P_{P\rightarrow AP}(+\frac{V}{2})$ - $P_3=P_{P\rightarrow AP}(+\frac{V}{5})$ - $P_4=P_{P\rightarrow AP}(-\frac{V}{5})$ - $P_5=P_{AP\rightarrow P}(+\frac{V}{2})=P_{P\rightarrow AP}(-\frac{V}{2})$ - $P_6=P_{AP\rightarrow P}(+\frac{4V}{5})$ When both junctions are in the same state, a voltage $\frac{V_{ac}}{2}$ is applied to each. When they are in different states, $\frac{4V_{ac}}{5}$ is applied to the junction in the AP state while $\frac{V_{ac}}{5}$ is applied to the junction in the P state. Therefore, $P_1$ and $P_6$ correspond to the switching probabilities of a junction in the AP state when the other junction is in the P state. $P_3$ and $P_4$ correspond to the switching probabilities of a junction in the P state when the other junction is in the AP state. $P_2$ and $P_5$ correspond to the switching probabilities when both junctions are in the same state, P or AP. In Figure \[TwoSyncTh\] the frequencies of junctions 1 and 2 as well as the probabilities $P_2$, $P_3$, $P_4$ and $P_5$ are plotted versus temperature. The probabilities $P_1$ and $P_6$ do not appear on this graph as $P_1\simeq 1$ and $P_6<0.01$ within the studied temperature range. Figure \[Schema\]B illustrates the switching cycles of both junctions. The plateau at half the drive frequency can be interpreted as follows and depicted on Fig. \[Schema\]B. We use the following example as a starting point: the drive voltage is $+V=1.5\mathrm{V}$ and both junctions are in the P state. (A similar reasoning can be made for any other initial conditions.) Fig. \[TwoSyncTh\] shows that $P_2>0.99$. Therefore, both junctions have a high probability ($P_2$) to switch in the AP state in the following half period $\frac{T_{ac}}{2}$. One junction switches to the AP state as depicted in Fig. \[Schema\]B. The switching probabilities are $P_6$ for the junction in the AP state and $P_3$ for the junction in the P state. As both probabilities $P_6$ and $P_3$ are below 0.01, both junctions remain in their state until the next switch of the drive, depicted by the “$T=60\mathrm{K}$” arrow in Fig. \[Schema\]A. When the drive switches from $+V$ to $-V$ the switching probabilities become $P_1$ for the junction in the AP state and $P_4$ for the junction in the P state. As $P_4<0.01$ while $P_1>0.99$, the junction in the AP state switches to the P state. Both junctions are in the P state with a switching probability of $P_5<0.01$: they remain in their state until the next switch of the drive. Then, their switching probabilities become $P_2$ and the cycle starts again. To summarize, at each oscillation cycle of the drive, one random junction is locked in phase with the drive while the other is blocked in the P state, leading to a mean frequency of $\frac{F_{ac}}{2}$ for both junctions. This is depicted on the the “Temperature = 60K” panel of Fig. \[Schema\]B. For $111\mathrm{K}<T<170\mathrm{K}$, $P_3$ is larger than 0.99. Considerations similar the the study of the $40 - 68\mathrm{K}$ range, depicted in Fig. \[Schema\]A, show that both junctions are locked in phase with the drive. This gives rise to the frequency plateau at $F_{ac}$ and is illustrated on the “Temperature = 130K” panel of Fig. \[Schema\]B. $P_5$ is the probability to phase-unlock with the drive, both when the drive is in the $+V$ and $-V$ states (Fig.\[Schema\]A). For temperatures above 170K, $P_5$ is larger than 0.01 so glitches appear and synchronization is destroyed.  ### Junctions connected head to head \[2syncTo\] We now consider that junction 2 is connected head to head with junction 1. In that case, the voltage applied to the junction, and governing the probabilities is $V_2=-V_{ac}\frac{R_2}{R_1+R_2}$. We observe in Fig. \[TwoSyncTh\](a) that the frequencies of junctions 1 and 2 – plotted in grey – are equal at all temperatures. The frequency of the junctions increases with the temperature. Contrary to the case where the junctions are head to tail, a single plateau at the drive frequency $F_{ac}$ is observed, between $T=40K$ and $T=126K$. Panels (g), (h) and (i) of Fig. \[Schema\]B depict the evolution with time of the voltage drive, and the states of junctions 1 and 2, for a temperature of T=75K. This regime corresponds to the plateau and we observe that junction 1 is locked in phase with the drive while junction 2 is locked in anti-phase with the drive. These results can be interpreted with the same reasoning as in the head to tail case and as depicted in Fig. \[Schema\]A.\ Electrical noise induced synchronization \[2syncN\] ---------------------------------------------------  In this section, we now control synchronization by injection of electrical noise. Contrary to the single junction case, we show that, in the two junctions case, electrical noise induces a behavior qualitatively different from the one induced by thermal noise. We consider two identical junctions which a tunnel magneto-resistance of $TMR=\frac{R_{AP}-R_P}{R_P}=100\%$. As shown in the Appendix, this is high enough to have a signal that can be detected easily and low enough to exhibit a significant synchronization plateau.\ The switching probabilities are the averaged probabilities $\langle P_1\rangle $ to $\langle P_6\rangle $ over all possible values of $N(t)$.\ In Fig.\[TwoSyncElec\] the frequencies of junctions 1 and 2 as well as the average probabilities $\langle P_1\rangle $ to $\langle P_6\rangle $ are plotted versus the standard deviation of the noise distribution, both for the configurations where the junctions are head to tail and head to head. ### Junctions connected head to tail\[2syncNs\] Both junctions have the same frequency at all levels of noise. We do not observe any synchronization at $\frac{F_{ac}}{2}$ for the junctions frequency in the range of noise for which $\langle P_2\rangle >0.99$ and $\langle P_3\rangle <0.01$. This can be interpreted as follows. In the thermal noise study, the escape rates are constant in time for a given temperature. The probability of switching at a given time is low but, because it is sustained over a long time, the probability for one junction to switch eventually is high. The situation is different when electrical noise is added on the drive. The distribution of the voltage from the electrical noise is Gaussian. As the noise is on the drive which is applied to both junctions, when a high value voltage occur, both junctions have a probability to switch that is close to one. Thus both junctions switch simultaneously. Therefore, when the drive switches to $+V$, either a high value voltage event occurs and both junctions switch simultaneously in the AP state; either no high voltage event occur, and only one junction switches in the AP state while the other remains blocked in the P state as explained in section \[2syncTs\]. As the noise level is increased, more high value voltage events occur and thus more simultaneous switches of the junctions occur. These observations pin a fundamental physical difference between internal and external noise.\ \ For $0.58\mathrm{V}<\sigma_{noise}<0.84\mathrm{V}$ the frequency of the junctions is equal to the drive frequency $F_{ac}$. Indeed $\langle P_3\rangle >0.99$ and $\langle P _5\rangle <0.01$. Both junctions are synchronized with the drive, in the same way as in the thermal noise study. For $\sigma_{noise}>0.84\mathrm{V}$, $\langle P _5\rangle$ is larger than $0.01$ so glitches appear and synchronization is lost. ### Junctions connected head to head\[2syncNo\] Both junctions have the same frequency for all levels of noise. For $0.16\mathrm{V}<\sigma_{noise}<0.72\mathrm{V}$ the frequency of the junctions is equal to the frequency of the drive $F_{ac}$. The behaviors of both junctions are the same as in the thermal noise study: junction 1 is locked in-phase while junction 2 is locked in anti-phase with the drive. $\langle P_6\rangle > \langle P_4\rangle $ so synchronization is lost when $\langle P_6\rangle >0.01$, which occurs when $\sigma_{noise}>0.72\mathrm{V}$. Conclusion ========== In this work we have investigated the noise-induced synchronization of one then two electrically coupled stochastic magnetic oscillators with a periodic drive. We have shown that both thermal and electrical noise can induce synchronization of two coupled superparamagnetic tunnel junctions with a sub-threshold voltage drive. When the junctions are head to tail, both are synchronized in-phase with the drive for an optimal level of noise. When the junctions are head to head, one is synchronized in-phase with the drive while the other is synchronized in anti-phase. Furthermore, this study has enabled us to highlight a fundamental difference between the synchronization phenomena induced by thermal and electrical noise. Indeed, the fact that electrical noise makes the escape rates probabilistic themselves and that the electrical noise is correlated for both junctions lead to simultaneous switches of the two junctions - while this does not occur for thermal noise. We have provided an intuitive analytic understanding of how one or several coupled stochastic oscillators respond to a periodic drive. This model will enable investigating the interactions and synchronization of several stochastic oscillators in various configurations. The results will be useful for understanding and designing networks of coupled stochastic oscillators. Among potential applications of such networks, bio-inspired computing is particularly promising, in particular as it allows taking advantage of noise and stochasticity for processing information. Indeed networks of coupled oscillators can perform pattern recognition and classification through synchronization [@hoppensteadt_oscillatory_1999; @aonishi_phase_1998]. Implementing these networks with sub-threshold synchronization of superparamagnetic tunnel junctions would allow achieving such cognitive tasks at low energy cost. Effects of magneto-resistance and drive frequency ================================================= In this appendix we study the effects of tunnel magneto-resistance and drive frequency on noise induced synchronization of two superparamagnetic tunnel junctions, connected head to tail. Two identical superparamagnetic tunnel junctions of energy barrier such that $\Delta E=30k_BT$ studied at $T=50K$ are submitted to a periodic drive of amplitude $V=1.5V_c$. The analytical model is used to compute the width in temperature of both synchronization plateaus, for various drive frequencies and magneto-resistance values. The results are presented in Figure \[map2\] for the $F=F_{ac}$ plateau and in Figure \[map2\_half\] for the $F=F_{ac}/2$ plateau. We observe that increasing the tunnel magneto-resistance widens the plateau at $F=F_{ac}/2$ and narrows the plateau at $F=F_{ac}$. When the junctions are in different states, the ratio between the voltage received by the junction in the AP state and the junction in the P state increases with the $R_{AP}/R_P$ ratio, which strengthens the coupling and favors the $F=F_{ac}/2$ regime. For $R_{AP}=R_P$ there is no tunnel magneto-resistance and the plateau at $F=F_{ac}/2$ is absent. The effect of drive frequency is non-monotonous. Drive frequencies too high compared to the natural frequency of the junctions (which depends on the energy barrier) prevent the apparition of synchronization plateaus. On the other side, as temperature increases the frequency of the junction, lowering the drive frequency narrows the temperature range which can induce synchronization.   The same study is conducted in the case of synchronization induced by electrical noise. Figure \[map\_noise\] presents the influence of drive frequency and tunnel magneto-resistance on the width of the synchronization plateau at $F_{ac}$. We observe that the plateau is widest for low magneto-resistances and high drive frequencies. Numerical simulations show that no synchronization plateau can be observed at $F_{ac}/2$, no matter the value of the magneto-resistance.\  The authors acknowledge financial support from the FET-OPEN Bambi project No. 618024 and the ANR MEMOS No. ANR-14-CE26-0021-01. A. M. acknowledges financial support from the Ile-de-France regional government through the DIM Nano-K program. N. L. acknowledges financial support from the French National Research Agency (ANR) as part of the Investissements d’Avenir program (Labex NanoSaclay, reference: ANR-10-LABX-0035)

ArXiv

--- author: - 'Benjamin Audren,' - 'Julien Lesgourgues,' - 'Karim Benabed,' - and Simon Prunet title: 'Conservative Constraints on Early Cosmology: an illustration of the [Monte Python]{} cosmological parameter inference code' --- Introduction ============ Models for the evolution of the early universe between a redshift of a few millions and a few hundreds have shown to be very predictive and successful: the self-consistency of Big Bang Nucleosynthesis (BBN) model could be tested by comparing the abundance of light elements and the result of Cosmic Microwave Background (CMB) observations concerning the composition of the early universe; the shape of CMB acoustic peaks matches accurately the prediction of cosmological perturbation theory in a Friedmann-Lemaître Universe described by general relativity, with a thermal history described by standard recombination. The late cosmological evolution is more problematic. Models for the acceleration of the universe, based on a cosmological constant, or a dark energy component, or departures from general relativity, or finally departure from the Friedmann-Lemaître model at late times, have shown no predictive power so far. The late thermal history, featuring reionization from stars, is difficult to test with precision. Overall, it is fair to say that “late cosmology” relies on less solid theoretical or observational ground than “early cosmology”. When fitting the spectrum of temperature and polarisation CMB anisotropies, we make simultaneously some assumptions on early and late cosmology, and obtain intricate constraints on the two stages. However, Vonlanthen et al. [@Vonlanthen:2010cd] suggested a way to carry the analysis leading to constraints only on the early cosmology part. This is certainly interesting since such an analysis leads to more robust and model-indepent bounds than a traditional analysis affected by priors on the stages which are most poorly understood. The approach of [@Vonlanthen:2010cd] avoids making assumptions on most relevant “late cosmology-related effects”: projection effects due to the background evolution, photon rescattering during reionization, and the late Integrated Sachs Wolfe (ISW) effect. In this work, we carry a similar analysis, pushed to a higher precision level since we also avoid making assumptions on the contamination of primary CMB anisotropies by weak lensing. We use the most recent available data from the Wilkinson Microwave Anisotropy Probe (WMAP) and South Pole Telescope (SPT) data, and consider the case of a minimal “early cosmology” model, as well as extended models with free density of ultra-relativistic relics or massive neutrinos. This analysis is an occasion to present a new cosmological parameter inference code. This Monte Carlo code written in Python, called [Monte Python]{}[^1], offers a convenient alternative to [CosmoMC]{} [@Lewis:2002ah]. It is interfaced with the Boltzmann code [class]{}[^2] [@Lesgourgues:2011re; @Blas:2011rf]. [Monte Python]{} is released publicly together with this work. In section 2, we explain the method allowing to get constraints only on the early cosmological evolution. We present our result for the minimal early cosmology model in section 3, and for two extended models in section 4. In section 5, we briefly summarize some of the advantages of [Monte Pyhton]{}, without entering into technical details (presented anyway in the code documentation). Our conclusions are highlighted in section 6. How to test early cosmology only? ================================= The spectrum of primary CMB temperature anisotropies is sensitive to various physical effects: - (C1) the location of the acoustic peaks in multipole space depends on the sound horizon at decoupling $d_s(\tau_{rec})$ (an “early cosmology”-dependent parameter) divided by the angular diameter distance to decoupling $d_A(\tau_{rec})$ (a “late cosmology”-dependent parameter, sensitive to the recent background evolution: acceleration, spatial curvature, etc.) - (C2) the contrast between odd and even peaks depends on $\omega_b/\omega_\gamma$, i.e. on “early cosmology”. - (C3) the amplitude of all peaks further depends on the amount of expansion between radiation-to-matter equality and decoupling, governing the amount of perturbation damping at the beginning of matter domination, and on the amount of early integrated Sachs-Wolfe effect enhancing the first peak just after decoupling. These are again “early cosmology” effects (in the minimal $\Lambda$CDM model, they are both regulated by the redshift of radiation-to-matter equality, i.e by $\omega_m/\omega_r$). - (C4) the enveloppe of high-$\ell$ peaks depends on the diffusion damping scale at decoupling $\lambda_d(\tau_{rec})$ (an “early cosmology” parameter) divided again by the angular diameter distance to decoupling $d_A(\tau_{rec})$ (a “late cosmology” parameter). - (C5-C6) the global shape depends on initial conditions through the primordial spectrum amplitude $A_s$ (C5) and tilt $n_s$ (C6), which are both “early cosmology” parameters. - (C7) the slope of the temperature spectrum at low $\ell$ is affected by the late integrated Sachs Wolfe effect, i.e. by “late cosmology”. This effect could actually be considered as a contamination of the primary spectrum by secondary anisotropies, which are not being discussed in this list. - (C8) the global amplitude of the spectrum at $\ell\gg40$ is reduced by the late reionization of the universe, another “late cosmology” effect. The amplitude of this suppression is given by $e^{-2 \tau}$, where $\tau$ is the reionization optical depth. In summary, primary CMB temperature anisotropies are affected by late cosmology only through: (i) projection effects from real space to harmonic space, controlled by $d_A(\tau_{rec})$; (ii) the late ISW effect, affecting only small $\ell$’s; and (iii) reionization, suppressing equally all multipoles at $\ell\gg40$ . These are actually the sectors of the cosmological model which are most poorly constrained and understood. But we see that the shape of the power spectrum at $\ell\gg40$, interpreted modulo an arbitrary scaling in amplitude ($C_{\ell} \rightarrow \alpha C_{\ell}$) and in position ($C_{\ell} \rightarrow C_{\beta \ell}$), contains information on early cosmology only. This statement is very general and valid for extended cosmological models. In the case of the $\Lambda$CDM models, it is illustrated by figure \[fig:rescale\], in which we took two different $\Lambda$CDM models (with different late-time geometry and reionization history), and rescaled one of them with a shift in amplitude given by $e^{-2 \tau-\tau'}$ and in scale given by $d_A/d_A'$. At $\ell\gg40$, the two spectra are identical. For more complicated cosmological models sharing the same physical evolution until approximately $z\sim100$, a similar rescaling and matching would work equally well. ![Dimensionless temperature (left) and E-polarization (right) unlensed spectra of two $\Lambda$CDM models with the same value of “early cosmology” parameters ($\omega_b$, $\omega_{cdm}$, $A_s$, $n_s$) (fixed to WMAP best-fitting values), and different values of “late cosmology” parameters: ($\Omega_\Lambda$, $z_{reio}$) = (0.720,10) (solid curves) or (0.619, 5) (dashed curve). The dashed curves have been rescaled vertically by the ratio of $e^{-2 \tau}$ and horizontally by the ratio of $d_A(\tau_{rec})$ in each model, using the values of $\tau$ and $d_A(\tau_{rec})$ calculated by [class]{} for each model. At $\ell=40$, the difference between the dashed and solid line in the temperature plot is under $2\,\mu$K$^2$. At $\ell=80$, it is already below $1\,\mu$K$^2$. \[fig:rescale\]](tt_comp.pdf "fig:"){height="7cm"} ![Dimensionless temperature (left) and E-polarization (right) unlensed spectra of two $\Lambda$CDM models with the same value of “early cosmology” parameters ($\omega_b$, $\omega_{cdm}$, $A_s$, $n_s$) (fixed to WMAP best-fitting values), and different values of “late cosmology” parameters: ($\Omega_\Lambda$, $z_{reio}$) = (0.720,10) (solid curves) or (0.619, 5) (dashed curve). The dashed curves have been rescaled vertically by the ratio of $e^{-2 \tau}$ and horizontally by the ratio of $d_A(\tau_{rec})$ in each model, using the values of $\tau$ and $d_A(\tau_{rec})$ calculated by [class]{} for each model. At $\ell=40$, the difference between the dashed and solid line in the temperature plot is under $2\,\mu$K$^2$. At $\ell=80$, it is already below $1\,\mu$K$^2$. \[fig:rescale\]](ee_comp.pdf "fig:"){height="7cm"} If polarization is taken into account, the same statement remains valid. The late time evolution affects the polarization spectrum through the angular diameter distance to decoupling $d_A(\tau_{rec})$ and through the impact of reionization, which also suppresses the global amplitude at $\ell\gg40$, and generates an additional feature at low $\ell$’s, due to photon re-scatering by the inhomogeneous and ionized inter-galactic medium. The shape of the primary temperature and polarization spectrum at $\ell\gg40$, interpreted modulo a global scaling in amplitude and in position, only contains information on the early cosmology. However, the CMB spectrum that we observe today gets a contribution from secondary anisotropies and foregrounds. In particular, the observed CMB spectra are significantly affected by CMB lensing caused by large scale structures. This effect depends on the small scale matter power spectrum, and therefore on late cosmology (acceleration, curvature, neutrinos becoming non-relativistic at late time, possible dark energy perturbations, possible departures from Einstein gravity on very large scales, etc.). In the work of [@Vonlanthen:2010cd], this effect was mentioned but not dealt with, because of the limited precision of WMAP5 and ACBAR data compared to the amplitude of lensing effects, at least within the multipole range studied in that paper ($40 \leq \ell \leq 800$). The results that we will present later confirm that this simplification was sufficient and did not introduce a significant “late cosmology bias”. However, with the full WMAP7+SPT data (that we wish to use up to the high multipoles), it is not possible to ignore lensing, and in order to probe only early cosmology, we are forced to marginalize over the lensing contamination, in the sense of the method described below. By doing so, we will effectively get rid of the major two sources of secondary (CMB) anisotropies, the late ISW effect and CMB lensing. We neglect the impact of other secondary effects like the Rees-Sciama effect. As far as foregrounds are concerned, the approach of WMAP and SPT consists in eliminating them with a spectral analysis, apart from residual foregrounds which can be fitted to the data, using some nuisance parameters which are marginalized over. By following this approach, we also avoid to introduce a “late cosmology bias” at the level of foregrounds. Let us now discuss how one can marginalize over lensing corrections. Ideally, we should lens the primary CMB spectrum with all possible lensing patterns, and marginalize over the parameters describing these patterns. But the lensing of the CMB depends on the lensing potential spectrum $C_{\ell}^{\phi \phi}$, that can be inferred from the matter power spectrum at small redshift, $P(k,z)$. We should marginalize over all possible shapes for $C_{\ell}^{\phi \phi}$, i.e. over an infinity of degrees of freedom. We need to find a simpler approach. One can start by noticing that modifications of the late-time background evolution caused by a cosmological constant, a spatial curvature, or even some inhomogeneous cosmology models, tend to affect matter density fluctuations in a democratic way: all Fourier modes being inside the Hubble radius and on linear scales are multiplied by the same redshift-dependent growth factor. CMB lensing is precisely caused by such modes. Hence, for this category of models, differences in the late-time background evolution lead to a different amplitude for $C_{\ell}^{\phi \phi}$, and also a small tilt since different $\ell$’s probe the matter power spectrum at different redshifts. Hence, if we fit the temperature and polarization spectrum at $\ell\gg40$ modulo a global scaling in amplitude, a global shift in position, and additionally an arbitrary scaling and tilting of the lensing spectrum $C_{\ell}^{\phi \phi}$ that one would infer assuming $\Lambda$CDM, we still avoid making assumption about the late-time evolution. There are also models introducing a scale-dependent growth factor, i.e. distortions in the shape of the matter power spectrum. This is the case in presence of massive neutrinos or another hot dark matter component, of dark energy with unusually large perturbations contributing to the total perturbed energy-momentum tensor, or in modified gravity models. In principle, these effects could lead to arbitrary distortions of $C_{\ell}^{\phi \phi}$ as a function of $\ell$. Fortunately, CMB lensing only depends on the matter power spectrum $P(k,z)$ integrated over a small range of redshifts and wave numbers. Hence it makes sense to stick to an expansion scheme: at first order we can account for the effects of a scale-dependent growth factor by writing the power spectrum as the one predicted by $\Lambda$CDM cosmology, multiplied by arbitrary rescaling and tilting factors; and at the next order, one should introduce a running of the tilt, then a running of the running, etc. By marginalizing over the rescaling factor, tilting factor, running, etc., one can still fit the CMB spectra without making explicit assumptions about the late-time cosmology. In the result section, we will check that the information on early cosmology parameters varies very little when we omit to marginalize over the lensing amplitude, or when we include this effect, or when we also marginalize over a tilting factor. Hence we will not push the analysis to the level of an arbitrary lensing running factor. Results assuming a minimal early cosmology model ================================================ We assume a “minimal early cosmology” model described by four parameters ($\omega_b$, $\omega_{cdm}$, $A_s$, $n_s$). In order to extract constraints independent of the late cosmological evolution, we need to fit the CMB temperature/polarisation spectrum measured by WMAP (seven year data [@Komatsu:2010fb]) and SPT [@Reichardt:2011yv] only above a given value of $\ell$ (typically $\ell \sim 40$), and to marginalize over two factors accounting for vertical and a horizontal scaling. In practice, there are several ways in which this could be implemented. For the amplitude, we could fix the reionization history and simply marginalize over the amplitude parameter $A_s$. By fitting the data at $\ell \gg 40$, we actually constrain the product $e^{-2 \tau_{reio}} A_s$, i.e. the primordial amplitude rescaled by the reionization optical depth $\tau_{reio}$, independently of the details of reionization. In our runs, we fix $\tau_{reio}$ to an arbitrary value, and we vary $A_s$; but in the Markov chains, we keep memory of the value of the derived parameter $e^{-2 \tau_{reio}} A_s$. By quoting bounds on $e^{-2 \tau} A_s$ rather than $A_s$, we avoid making explicit assumptions concerning the reionization history. For the horizontal scaling, we could modify [class]{} in such way to use directly $d_A(\tau_{rec})$ as an input parameter. For input values of ($\omega_b$, $\omega_{cdm}$, $d_A(\tau_{rec})$), [class]{} could in principle find the correct spectrum at $\ell \gg 40$. It is however much simpler to use the unmodified code and pass values of the five parameters ($\omega_b$, $\omega_{cdm}$, $A_s$, $n_s$, $h$). In our case, $h$ should not be interpreted as the reduced Hubble rate, but simply as a parameter controlling the value of the physical quantity $d_A(\tau_{rec})$. For any given set of parameters, the code computes the value that $d_A(\tau_{rec})$ would take in a $\Lambda$CDM model with the same early cosmology and with a Hubble rate $H_0=100h$km/s/Mpc. It then fits the theoretical spectrum to the data. The resulting likelihood should be associated to the inferred value of $d_A(\tau_{rec})$ rather than to $h$. The only difference between this simplified approach and that in which $d_A(\tau_{rec})$ would be passed as an input parameter is that in one case, one assumes a flat prior on $d_A(\tau_{rec})$, and in the other case a flat prior on $h$. But given that the data allows $d_A(\tau_{rec})$ to vary only within a very small range where it is almost a linear function of $h$, the prior difference has a negligible impact. To summarize, in order to get constraints on “minimal early cosmology”, it is sufficient to run Markov Chains in the same way as for a minimal $\Lambda$CDM model with parameters ($\omega_b$, $\omega_{cdm}$, $A_s$, $n_s$, $\tau$, $h$), excepted that: - we do not fit the lowest temperature/polarization multipoles to the data; - we fix $\tau$ to an arbitrary value; - we do not plot nor interpret the posterior probability of the parameters $A_s$ and $h$. We only pay attention to the posterior probability of the two derived parameters $e^{-2 \tau} A_s$ and $d_A(\tau_{rec})$, which play the role of the vertical and horizontal scaling factors, and which are marginalized over when quoting bounds on the remaining three “early cosmology parameters” ($\omega_b$, $\omega_{cdm}$, $n_s$). Hence, for a parameter inference code, this is just a trivial matter of defining and storing two “derived parameters”. For clarity, we will refer to the runs performed in this way as the “agnostic” runs. ![Constraints on the five parameters of the minimal early cosmology model (red), compared to usual constraints on the minimal $\Lambda$CDM model (black). The $\Lambda$CDM has a sixth independent parameter, the reionization optical depth. The constraints on early cosmology (called “agnostic constraints” in the text) includes a marginalization over the amplitude and tilt of the matter power spectrum leading to CMB lensing. We do not show here the posterior of the three nuisance parameter used to fit SPT data.\[fig:1d\]](lcdm-vs-tilt_ag_100.pdf){width="12cm"} In the second line of Table \[tab:minimal\], we show the bounds obtained with such an agnostic run, for a cut-off value $\ell=40$. These results can be compared with those of a minimal $\Lambda$CDM model, obtained through the same machinery but with all multipoles $\ell\geq 2$. Since the agnostic bounds rely on less theoretical assumptions, they are slightly wider. Interestingly, the central value of $\omega_b$ and $n_s$ are smaller in absence of late-cosmology priors, and larger for $\omega_{cdm}$. Still the $\Lambda$CDM results are compatible with the agnostic results, which means that on the basis of this test, we cannot say that $\Lambda$CDM is a bad model. Our agnostic bounds on ($\omega_b$, $\omega_{cdm}$, $n_s$) are simply more model-independent and robust, and one could argue that when using CMB bounds in the study of BBN, in CDM relic density calculations or for inflationary model building, one should better use those bounds in order to avoid relying on the most uncertain assumptions of the minimal cosmological model, namely $\Lambda$ domination and standard reionization. The decision to cut the likelihood at $\ell\geq 40$ was somewhat arbitrary. Figure \[fig:rescale\] shows that two rescaled temperature spectra with different late-time cosmology tend only gradually towards each other above $\ell\sim 40$. We should remove enough low multipoles in order to be sure that late time cosmology has a negligible impact given the data error bars. We tested this dependence by cutting the likelihood at $\ell\geq 60$, $\ell\geq 80$ or $\ell\geq 100$. When increasing the cut-off from 40 to 100, we observe variations in the mean value that are less important than from 2 to 40. To have the more robust constraints, we will then take systematically the cut-off of $\ell=100$, which is the one more likely to avoid any contamination from “late time cosmology”. Until now, our analysis is not completely “agnostic”, because we did not marginalize over lensing. We fitted the data with a lensed power spectrum, relying on the same lensing potential as an equivalent $\Lambda$CDM model with the same values of ($\omega_b$, $\omega_{cdm}$, $n_s$, $e^{-2 \tau} A_s$, $d_A(\tau_{rec})$). To deal with lensing, we introduce three new parameters ($A_{lp}$, $n_{lp}$, $k_{lp}$) in [class]{}. Given the traditional input parameters ($\omega_b$, $\omega_{cdm}$, $A_s$, $n_s$, $h$), the code first computes the Newtonian potential $\phi(k,z)$. This potential is then rescaled as $$\phi(k,z) \longrightarrow A_{lp} \left(\frac{k}{k_{lp}}\right)^{n_{lp}}\phi(k,z)~.$$ Hence, the choice ($A_{lp}$, $n_{lp}$)=(1,0) corresponds to the standard lensing potential predicted in the $\Lambda$CDM model. Different values correspond to an arbitrary rescaling or tilting of the lensing potential, which can be propagated consistently to the lensed CMB temperature/polarization spectrum. The sixth run shown in Table \[tab:minimal\] corresponds to $n_{lp}=0$ and a free parameter $A_{lp}$. The minimum credible interval for this rescaling parameter is $A_{lp}=0.88^{+0.12}_{-0.13}$ at the 68% Confidence Level (CL), and is compatible with one. This shows that WMAP7+SPT data alone are sensitive to lensing, and well compatible with the lensing signal predicted by the minimal $\Lambda$CDM model. It is also interesting to note that the bounds on other cosmological parameter move a little bit, but only by a small amount (compared to the difference between the $\Lambda$CDM and the previous “agnostic” runs), showing that “agnostic bounds” are robust. In the seventh line of Table \[tab:minimal\], we also marginalize over the tilting parameter $n_{lp}$ (with unbounded flat prior). A priori, this introduces a lot of freedom in the model. Nicely, this parameter is still well constrained by the data ($n_{lp}=-0.16^{+0.55}_{-0.33}$ at 68%CL), and compatible with the $\Lambda$CDM prediction $n_{lp}=0$. Bounds on other parameters vary this time by a completely negligible amount: this motivates us to stop the expansion at the level of $n_{lp}$, and not to test the impact of running. The credible interval for $A_{lp}$ is the only one varying significantly when $n_{lp}$ is left free, but this result depends on the pivot scale $k_{lp}$, that we choose to be equal to $k_{lp}=0.1$/Mpc, so that the amplitude of the lensing spectrum $C_{\ell}^{\phi \phi}$ is nearly fixed at $\ell\sim 100$. By tuning the pivot scale, we could have obtained bounds on $A_s$ nearly equal for the case with/without free $n_{lp}$. The posterior probability of each parameter marginalized over other parameters is shown in Figure \[fig:1d\], and compared with the results of the standard $\Lambda$CDM analysis. Our results nicely agree with those of [@Vonlanthen:2010cd]. These authors found a more pronounced drift of the parameters ($\omega_b$, $\omega_{cdm}$, $n_s$) with the cut-off multipole than in the first part of our analysis, but this is because we use data on a wider multipole range and have a larger lever arm. Indeed, Ref.  [@Vonlanthen:2010cd] limited their analysis of WMAP5 plus ACBAR data to $\ell\leq800$, arguing that above this value, lensing would start playing an important role. In our analysis, we include WMAP7 plus 47 SPT band powers probing up to $\ell \sim 3000$, but for consistency we must simultaneously marginalize over lensing. Indeed, the results of Ref.  [@Vonlanthen:2010cd] are closer to our results with lensing marginalization (the fully “agnostic” ones) that without. Keeping only one digit in the error bar, we find ($100\omega_b=2.16\pm0.07$, $\omega_{cdm} = 0.122\pm0.009$, $n_s=0.93\pm0.02$), when this reference found ($100\omega_b=2.13\pm0.05$, $\omega_{cdm} = 0.124\pm0.007$, $n_s=0.93\pm0.02$). The two sets of results are very close to each other, but our central values for $\omega_b$ and $\omega_{cdm}$ are slightly closer to the $\Lambda$CDM one. The fact that we get slightly larger error bars in spite of using better data in a wider multipole range is related to our lensing marginalization: we see that by fixing lensing, this previous analysis was implicitly affected by a partial “late cosmology prior”, but only at a very small level. Our results from the last run can be seen as robust “agnostic” bounds on ($\omega_b$, $\omega_{cdm}$, $n_s$), only based on the “minimal early cosmology” assumption. They are approximately twice less constraining than ordinary $\Lambda$CDM models, and should be used in conservative studies of the physics of BBN, CDM decoupling and inflation. Effective neutrino number and neutrino mass =========================================== We can try to generalize our analysis to extended cosmological models. It would make no sense to look at models with spatial curvature, varying dark energy or late departures from Einstein gravity, since all these assumptions would alter only the late time evolution, and our method is designed precisely in such way that the results would remain identical. However, we can explore models with less trivial assumptions concerning the early cosmological evolution. This includes for instance models with: - a free primordial helium fraction $Y_{He}$. So far, we assumed $Y_{He}$ to be a function of $\omega_b$, as predicted by standard BBN (this is implemented in [class]{} following the lines of Ref. [@Hamann:2007sb]). Promoting $Y_{He}$ as a free parameter would be equivalent to relax the assumption of standard BBN. Given the relatively small sensitivity of current CMB data to $Y_{He}$ [@Komatsu:2010fb], we do not perform such an analysis here, but this could be done in the future using e.g. Planck data. - a free density of relativistic species, parametrized by a free effective neutrino number $N_{\rm eff}$, differing from its value of $3.046$ in the minimal $\Lambda$CDM model [@Mangano:2001iu]. This parameter affects the time of equality between matter and radiation, but this effect can be cancelled at least at the level of “early cosmology” by tuning appropriately the density of barons and CDM. Even in that case, relativistic species will leave a signature on the CMB spectrum, first through a change in the diffusion damping scale $\lambda_d(\tau_{rec})$, and second through direct effects at the level of perturbations, since they induce a gravitational damping and phase shifting of the photon fluctuation [@Hu:1995en; @Bashinsky:2003tk]. It is not obvious to anticipate up to which level these effects are degenerate with those of other parameters. Hence it is interesting to run Markov chains and search for “agnostic bounds” on $N_{\rm eff}$. - neutrino masses (or for simplicity, three degenerate masses $m_\nu$ summing up to $M_\nu=3m_\nu$). Here we are not interested in the fact that massive neutrinos affect the background evolution and change the ratio between the redshift of radiation-to-matter equality, and that of matter-to-$\Lambda$ equality. This is a “late cosmology” effect that we cannot probe with our method, since we are not sensitive to the second equality. However, for masses of the order of $m_\nu\sim0.60$ eV, neutrinos become non-relativistic at the time of photon decoupling. Even below this value, the mass leaves a signature on the CMB spectrum coming from the fact that, first, they are not yet ultra-relativistic at decoupling, and second, the transition to the non-relativistic regime takes place when the CMB is still probing metric perturbations through the early integrated Sachs-Wolfe effect. Published bounds on $M_\nu$ from CMB data alone probe all these intricate effects [@Komatsu:2010fb], and it would be instructive to obtain robust bounds based only on the mass impact on “early cosmology”. ![One and two-dimensional posterior distribution (solid line) of the parameters of the “agnostic” run with a free effective neutrino number. The dashed line stands for the average likelihood distribution. The concentric contour lines in the two-dimensional posteriors stand for 68, 95 and 99% CL. \[fig:triangle-neff\]](neff_tilt_ag_100_triangle.pdf){width="15cm"} For the effective neutrino number, we performed two runs similar to our previous $\Lambda$CDM and “fully agnostic” run (with marginalization over lensing amplitude and tilt), in presence of one additional free parameter $N_{\rm eff}$. Our results are summarized in Table \[tab:neff\] and Figure \[fig:triangle-neff\]. In the $\Lambda$CDM+$N_{\rm eff}$ case, we get $N_{\rm eff}=3.77^{+0.58}_{-0.66}$ (68% CL), very close to the result of [@Keisler:2011aw], $N_{\rm eff}=3.85\pm0.62$ (differences in the priors can explain this insignificant difference). It is well-known by now that the combination of WMAP and small-scale CMB data shows a marginal preference for extra relativistic degrees of freedom in the seven-parameter model. The surprise comes from our “agnostic” bound on this number, $N_{\rm eff}=2.04^{+0.78}_{-1.26}$ (68% CL). As explained before, this bound cannot come from a change in the time of equality, nor in the scale of the first peak, nor in the late integrated Sachs-Wolfe effect; it can only result from the measurement of the the sound horizon $d_s(\tau_{rec})$ [*relatively*]{} to the diffusion damping scale $\lambda_d(\tau_{rec})$, and from the direct effects of extra relativistic degrees of freedom on photon perturbations. Hence it is normal that $N_{\rm eff}$ is much less constrained in the agnostic runs, but the interesting conclusion is that without assuming $\Lambda$CDM at late time, the CMB does not favor high values of $N_{\rm eff}$. It is compatible with the standard value $N_{\rm eff}=3.046$ roughly at the one-$\sigma$ level, with even a marginal preference for smaller values. This shows that recent hints for extra relativistic relics in the universe disappear completely if we discard any information on the late time cosmological evolution. It is well-known that $N_{\rm eff}$ is very correlated with $H_0$ and affected by the inclusion of late cosmology data sets, like direct measurement of $H_0$ or of the BAO scale. Our new result shows that even at the level of CMB data only, the marginal hint for large $N_{\rm eff}$ is driven by physical effects related to late cosmology (and in particular by the angular diameter distance to last scattering as predicted in $\Lambda$CDM). The triangle plot in Figure \[fig:triangle-neff\] shows that in the agnostic run, $N_{\rm eff}$ is still very correlated with other parameters such as $\omega_b$, $\omega_{cdm}$ and $n_s$. Low values of $N_{\rm eff}$ (significantly smaller than the standard value 3.046) are only compatible with a very small $\omega_b$, $\omega_{cdm}$ and $n_s$. Note that in this work, we assume standard BBN in order to predict $Y_{He}$ as a function of $\omega_b$ (and of $N_{\rm eff}$ when this parameter is also left free), but we do not incorporate data on light element abundances. By doing so, we would favor the highest values of $\omega_b$ in the range allowed by the current analysis ($\omega_b \sim 0.022$), and because of parameter correlations we would also favor the highest values of $\omega_{cdm}$, $n_s$ and $N_{\rm eff}$, getting close to the best-fitting values in the minimal early cosmology model with $N_{\rm eff}\sim3.046$.   ![One and two-dimensional posterior distribution (solid line) of the parameters of the “agnostic” run with a total neutrino mass $M_\nu$ (assuming three degenerate neutrinos of individual mass $m_\nu$). Again, dashed line stands for average likelihood distribution, contour lines indicate the 68, 95 and 99% CL.[]{data-label="fig:triangle-mnu"}](mnu_tilt_ag_100_triangle.pdf){width="15cm"} For neutrino masses, we performed two similar runs (summarised in Table \[tab:mnu\] and Figure \[fig:triangle-mnu\]), with now $M_\nu$ being the additional parameter (assuming three degenerate neutrino species). In the $\Lambda$CDM case, our result $M_\nu<1.4$ eV (95%CL) is consistent with the rest of the literature, and close to the WMAP-only bound of [@Komatsu:2010fb]: measuring the CMB damping tail does not bring significant additional information on the neutrino mass. In the agnostic run, this constraint only degrades to $M_\nu<1.8$ eV (95%CL). This limit is consistent with the idea that for sufficiently large $m_\nu$, the CMB can set a limit on the neutrino mass not just through its impact on the background evolution at late time (and its contribution to $\omega_m$ today), but also through direct effects occurring at the time of recombination and soon after. It is remarkable that this is true even for neutrinos of individual mass $m_\nu\sim0.6$ eV, becoming non-relativistic precisely at the time of photon decoupling. The conclusion that the CMB is not compatible with neutrinos becoming non-relativistic before $z_{rec}$ (and not even slightly before!) appears to be very robust, and independent of any constraint on the late cosmological evolution. Advantages of [Monte Python]{} ============================== The results of this paper were obtained with the new parameter inference code [Monte Python]{}, that we release publicly together with this article. Currently, [Monte Python]{} is interfaced with the Boltzmann code [class]{}, and explores parameter space with the Metropolis-Hastings algorithm, just like [CosmoMC]{}[^3] (note however that interfacing it with other codes and switching to other exploration algorithms would be easy, thanks to the modular architecture of the code). Hence, the difference with [CosmoMC]{} [@Lewis:2002ah] does not reside in a radically different strategy, but in several details aiming at making the user’s life easy. It is not our goal to describe here all the features implemented in [Monte Python]{}: for that, we refer the reader to the documentation distributed with the code. We only present here a brief summary of the main specificities of [Monte Python]{}. [**Language and compilation.**]{} As suggested by its name, [Monte Python]{} is a Monte Carlo code written in Python. This high-level language allows to code with a very concise style: [Monte Python]{} is compact, and the implementation of e.g. new likelihoods requires very few lines. Python is also ideal for wrapping other codes from different languages: [Monte Python]{} needs to call [class]{}, written in C, and the WMAP likelihood code, written in Fortran 90. The user not familiar with Python should not worry: for most purposes, [Monte Python]{} does not need to be edited, when [CosmoMC]{} would need to: this is explained in the fourth paragraph below.\ Another advantage of Python is that it includes many libraries (and an easy way to add more), so that [Monte Python]{} is self-contained. Only the WMAP likelihood code needs its own external libraries, as usual. Python codes do not require a compilation step. Hence, provided that the user has Python 2.7 installed on his/her computer[^4] alongside very standard modules, the code only needs to be downloaded, and is ready to work with. [**Modularity.**]{} A parameter inference code is based on distinct blocks: a likelihood exploration algorithm, an interface with a code computing theoretical predictions (in our case, a Boltzmann code solving the cosmological background and perturbation evolution), and an interface with each experimental likelihood. In [Monte Python]{}, all three blocks are clearly split in distinct modules. This would make it easy, e.g., to interface [Monte Python]{} with [camb]{} [@Lewis:1999bs] instead of [class]{}, or to switch from the in-build Metropolis-Hastings algorithm to another method, e.g. a nested sampling algorithm.\ The design choice of the code has been to write these modules as different classes, in the sense of C++, whenever it served a purpose. For instance, all likelihoods are defined as separated classes. It allows for easy and intuitive way of comparing two runs, and helps simplify the code. The cosmological module is also defined as a class, with a defined amount of functions. If someone writes a python wrapper for [camb]{} defining these same functions, then [Monte Python]{} would be ready to serve.\ On the other hand the likelihood exploration part is contained in a normal file, defining only functions. The actual computation is only done in the file [code/mcmc.py]{}, so it is easy to implement a different exploration algorithm. From the rest of the code, this step would be as transparent as possible.\ In Python, like in C++, a class can inherit properties from a parent class. This becomes particularly powerful when dealing with data likelihoods. Each likelihood will inherit from a basic [likelihood]{} class, able to read data files, and to treat storage. In order to implement a new likelihood, one then only needs to write the computation part, leaving the rest automatically done by the main code. This avoids several repetitions of the same piece of code. Furthermore, if the likelihood falls in a generic category, like CMB likelihoods based on reading a file in the format [.newdat]{} (same files as in [CosmoMC]{}), it will inherit more precise properties from the [likelihood\_newdat]{} class, which is itself a daughter of the [likelihood]{} class. Hence, in order to incorporate CMB likelihoods apart from WMAP, one only needs to write one line of python for each new case: it is enough to tell, e.g., to the class accounting for the CMB experiment SPT that it inherits all properties from the generic [likelihood\_newdat]{} class. Then, this class is ready to read a file in the [.newdat]{} format and to work. Note that our code already incorporates another generic likelihood class that will be useful in the future for reading Planck likelihoods, after the release of Planck data.\ Finally, please note that these few lines of code to write for a new likelihood are completely outside the main code containing the exploration algorithm, and the cosmological module. You do not need to tell the rest of the code that you wrote something new, you just have to use your new likelihood by its name in a starting parameter file. [**Memory keeping and safe running.**]{} Each given run, i.e. each given combination of a set of parameters to vary, a set of likelihoods to fit, and a version of the Boltzmann code, is associated to a given directory where the chains are written (e.g. it could be a directory called [chains/wmap\_spt/lcdm]{}). All information about the run is logged automatically in this directory, in a file [log.param]{}, at the time when the first chain is started. This file contains the parameter names, ranges and priors, the list of extra parameters, the version of the Boltzmann code, the version and the characteristics of each data likelihood, etc. Hence the user will always remember the details of a previous run.\ Moreover, when a new chain is started, the code reads this log file (taking full advantage of the class structure of the code). If the user started the new chain with an input file, the code will compare all the data in the input file with the data in the [log.param]{} file. If they are different, the code complains and stop. The user can then take two decisions: either some characteristic of the run has been changed without noticing, and the input file can be corrected. Or it has been changed on purpose, then this is a new run and the user must require a different output directory. This avoids the classical mistake of mixing unwillingly some chains that should not be compared to each other. Now, if the input file is similar to the [log.param]{} file, the chain will start (it will not take the same name as previous chains: it will append automatically to its name a number equal to the first available number in the chain directory). In addition, the user who wishes to launch new chains for the same run can omit to pass an input file: in this case all the information about the run is automatically read in the [log.param]{} and the chain can start.\ The existence of [log.param]{} file has another advantage. When one wants to analyze chains and produce result files and plots (the equivalent of running [Getdist]{} and matlab or maple in the case of [CosmoMC]{}), one simply needs to tell [Monte Python]{} to analyze a given directory. It is not needed to pass another input file, since all information on parameter names and ranges will be found in the [log.param]{}. If the output needs to be customized (i.e., changing the name of the parameters, plotting only a few of them, rescaling them by some factor, etc.), then the user can use command lines and eventually pass one small input file with extra information. [**No need to edit the code when adding parameters.**]{} The name of cosmological parameters is never defined in [Monte Python]{}. The code only knows that in the input file, it will read a list of parameter names (e.g. [omega\_b]{}, [z\_reio]{}, etc.) and pass this list to the cosmology code together with some values. The cosmology code (in our case, [class]{}) will read these names and values as if they were written in an input file. If one of the names is not understood by the cosmology code, the run stops. The advantage is that the user can immediately write in the input file any name understood by class, without needing to edit [Monte Python]{}. This is not the case with [CosmoMC]{}. This is why users can do lots of things with [Monte Python]{} without ever needing to edit it or even knowing Python. If one wants to explore a completely new cosmological model, it is enough to check that it is implemented in [class]{} (or to implement it oneself and recompile the class python wrapper). But [Monte Python]{} doesn’t need to know about the change. To be precise, in the [Monte Python]{} input file, the user is expected to pass the name, value, prior edge etc. of all parameters (i) to be varied; (ii) to be fixed; (iii) to be stored in the chains as derived parameters. These can be any [class]{} parameter: cosmological parameters, precision parameters, flags, input file names. Let us take two examples: - In this paper, we showed some posterior probabilities for the angular diameter distance up to recombination. It turns out that this parameter is always computed and stored by [class]{}, under the name [‘da\_rec’]{}. Hence we only needed to write in the input file of [Monte Python]{} a line looking roughly like [da\_rec=‘derived’]{} (see the documentation for the exact syntax), and this parameter was stored in the chains. In this case [Monte Python]{} did not need editing. - We used in this work the parameter $[e^{-2 \tau} A_s]$. To implement this, there would be two possibilities. The public [class]{} version understands the parameters $\tau$ and $A_s$. The first possibility is to modify the [class]{} input module, teach it to check if there is an input parameter [‘exp\_m\_two\_tau\_A\_s’]{}, and if there is, to infer $A_s$ from $[e^{-2 \tau} A_s]$ and $\tau$. Then there is no need to edit [ Monte Python]{}. However, in a case like this, it is actually much simpler to leave [class]{} unchanged and to add two lines in the [Monte Python]{} file [data.py]{}. There is a place in this file devoted to internal parameter redefinition. The user can add two simple lines to tell [ Monte Python]{} to map ([‘exp\_m\_two\_tau\_A\_s’]{}, [‘tau’]{}) to ([ ‘A\_s’]{}, [‘tau’]{}) before calling [class]{}. This is very basic and does not require to know python. All these parameter manipulations are particularly quick and easy with [Monte Python]{}. The user is also free to rescale a parameter (e.g. $A_s$ to $10^9A_s$ in order to avoid dealing with exponents everywhere) by specifying a rescaling factor in the input file of [Monte Python]{}: so this can be done without editing neither [Monte Python]{} nor [class]{}.\ Please note however that, while this is true that any input parameter will be understood directly by the code, to recover derived parameters, the wrapper routine (distributed with class) should know about them. To this end, we implemented what we think is a near-complete list of possible derived parameters in the latest version of the wrapper. [**Playing with covariance matrices.**]{} When chains are analyzed, the covariance matrix is stored together with parameter names. When this matrix is passed as input at the beginning of the new run, these names are read. The code will then do automatically all the necessary matrix manipulation steps needed to get all possible information from this matrix if the list of parameter has changed: this includes parameter reordering and rescaling, getting rid of parameters in the matrix not used in the new runs, and adding to the matrix some diagonal elements corresponding to new parameters. All the steps are printed on screen for the user to make sure the proper matrix is used. [**Friendly plotting.**]{} The chains produced by [Monte Python]{} are exactly in the same format as those produced by [CosmoMC]{}: the user is free to analyze them with [GetDist]{} or with a customized code. However [Monte Python]{} incorporates its own analysis module, that produce output files and one or two dimensional plots in PDF format (including the usual “triangle plot”). Thanks to the existence of [log.param]{} files, we just need to tell [Monte Python]{} to analyze a given directory - no other input is needed. Information on the parameter best-fit, mean, minimal credible intervals, convergence, etc., are then written in three output files with different presentation: a text file with horizontal ordering of the parameters, a text file with vertical ordering, and a latex file producing a latex table. In the plots, the code will convert parameter names to latex format automatically (at least in the simplest case) in order to write nice labels (e.g. it has a routine that will automatically replace [tau\_reio]{} by [\\tau\_$\{$reio$\}$]{}). If the output needs to be customized (i.e., changing the name of the parameters, plotting only a few of them, rescaling them by some factor, etc.), then the user can use command lines and eventually pass one small input file with extra information. The code stores in the directory of the run only a few PDF files (by default, only two; more if the user asks for individual parameter plots), instead of lots of data files that would be needed if we were relying on an external plotting software like Matlab. [**Convenient use of mock data.**]{} The released version of [Monte Python]{} includes simplified likelihood codes mimicking the sensitivity of [*Planck*]{}, of a [*Euclid*]{}-like galaxy redshift survey, and of a [*Euclid*]{}-like cosmic shear survey. The users can take inspiration from these modules to build other mock data likelihoods. They have been developed in such way that dealing with mock data is easy and fully automatized. The first time that a run is launched, [Monte Python]{} will find that the mock data file does not exist, and will create one using the fiducial model parameters passed in input. In the next runs, the power spectra of the fiducial model will be used as an ordinary data set. This approach is similar to the one developed in the code FuturCMB[^5] [@Perotto:2006rj] compatible with [CosmoMC]{}, except that the same steps needed to be performed manually. Conclusions =========== Models for the latest stages of the cosmological evolution rely on a less solid theoretical and observational ground than the description of earlier stages, like BBN and recombination. Reference [@Vonlanthen:2010cd] suggested a way to infer parameters from CMB data under some assumptions about early cosmology, but without priors on late cosmology. By standard assumption on early cosmology, we understand essentially the standard model of recombination in a flat Friedmann-Lemaître universe, assuming Einstein gravity, and using a consistency relation between the baryon and Helium abundance inferred from standard BBN. The priors on late cosmology that we wish to avoid are models for the acceleration of the universe at small redshift, a possible curvature dominated stage, possible deviations from Einstein gravity on very large scale showing up only at late times, and reionization models. We explained how to carry such an analysis very simply, pushing the method of [@Vonlanthen:2010cd] to a higher precision level by introducing a marginalization over the amplitude and tilt of the CMB lensing potential. We analyzed the most recent available WMAP and SPT data in this fashion, that we called “agnostic” throughout the paper. Our agnostic bounds on the minimal “early cosmology” model are about twice weaker than in a standard $\Lambda$CDM analysis, but perfectly compatible with $\Lambda$CDM results: there is no evidence that the modeling of the late-time evolution of the background evolution, thermal history and perturbation growth in the $\Lambda$CDM is a bad model, otherwise it would tilt the constraints on $\omega_b$, $\omega_{cdm}$ and $n_s$ away from the “agnostic” results. It is interesting that WMAP and SPT alone favor a level of CMB lensing different from zero and compatible with $\Lambda$CDM predictions. We extended the analysis to two non-minimal models changing the “early cosmology”, with either a free density of ultra-relativistic relics, or some massive neutrinos that could become non-relativistic before or around photon decoupling. In the case of free $N_{\rm eff}$, it is striking that the “agnostic” analysis removes any hint in favor of extra relics. The allowed range is compatible with the standard value $N_{\rm eff}=3.046$ roughly at the one-sigma level, with a mean smaller than three. In the case with free total neutrino mass $M_\nu$, it is remarkable that the “agnostic” analysis remains sensitive to this mass: the two-sigma bound coincides almost exactly with the value of individual masses corresponding to a non-relativistic transition taking place at the time of photon decoupling. The derivation of these robust bounds was also for us an occasion to describe the main feature of the new parameter inference code [Monte Python]{}, that we release together with this paper. [Monte Python]{} is an alternative to [CosmoMC]{}, interfaced with the Boltzmann code [class]{}. It relies on the same basic algorithm as [CosmoMC]{}, but offers a variety of user-friendly function, that make it suitable for a wide range of cosmological parameter inference analyses. We would like to thank Martin Kilbinger for useful discussions. We are also very much indebted to Wessel Valkenburg for coming up with the most appropriate name possible for the code. This project is supported by a research grant from the Swiss National Science Foundation. [99]{} M. Vonlanthen, S. Rasanen and R. Durrer, JCAP [**1008**]{} (2010) 023 \[arXiv:1003.0810 \[astro-ph.CO\]\]. J. Lesgourgues, arXiv:1104.2932 \[astro-ph.IM\]. D. Blas, J. Lesgourgues and T. Tram, JCAP [**1107**]{} (2011) 034 \[arXiv:1104.2933 \[astro-ph.CO\]\]. A. Lewis and S. Bridle, Phys. Rev. D [**66**]{} (2002) 103511 \[astro-ph/0205436\]. E. Komatsu [*et al.*]{} \[WMAP Collaboration\], Astrophys. J. Suppl.  [**192**]{} (2011) 18 \[arXiv:1001.4538 \[astro-ph.CO\]\]. C. L. Reichardt, L. Shaw, O. Zahn, K. A. Aird, B. A. Benson, L. E. Bleem, J. E. Carlstrom and C. L. Chang [*et al.*]{}, Astrophys. J.  [**755**]{} (2012) 70 \[arXiv:1111.0932 \[astro-ph.CO\]\]. J. Hamann, J. Lesgourgues and G. Mangano, JCAP [**0803**]{} (2008) 004 \[arXiv:0712.2826 \[astro-ph\]\]. G. Mangano, G. Miele, S. Pastor and M. Peloso, Phys. Lett. B [**534**]{} (2002) 8 \[astro-ph/0111408\]. W. Hu and N. Sugiyama, Astrophys. J.  [**471**]{} (1996) 542 \[astro-ph/9510117\]. S. Bashinsky and U. Seljak, Phys. Rev. D [**69**]{} (2004) 083002 \[astro-ph/0310198\]. R. Keisler, C. L. Reichardt, K. A. Aird, B. A. Benson, L. E. Bleem, J. E. Carlstrom, C. L. Chang and H. M. Cho [*et al.*]{}, Astrophys. J.  [**743**]{} (2011) 28 \[arXiv:1105.3182 \[astro-ph.CO\]\]. A. Lewis, A. Challinor and A. Lasenby, Astrophys. J.  [**538**]{} (2000) 473 \[astro-ph/9911177\]. L. Perotto, J. Lesgourgues, S. Hannestad, H. Tu and Y. Y. Y. Wong, JCAP [**0610**]{} (2006) 013 \[astro-ph/0606227\]. [^1]: http://montepython.net [^2]: http://class-code.net [^3]: In this paper we refer to the version of CosmoMC available at the time of submitting, i.e. the version of October 2012. [^4]: The documentation explains how to run with Python 2.6. The code would require very minimal modifications to run with Python 3.0. [^5]: http://lpsc.in2p3.fr/perotto/

ArXiv

--- abstract: 'Novel methods to analyze NMR signals dominated by dipolar interaction are applied to the study of slow relaxation motions in polybutadiene approaching its glass transition temperature. The analysis is based on a recently developed model where the time dependence in an ensemble of dipolar interacting spin pairs is described without resorting to the Anderson-Weiss approximation. The ability to catch relevant features of the $\alpha$ relaxation process is emphasized. In particular, it is shown that the temperature profile of the Magic Sandwich Echo efficiency carries information on the frequency profile of the $\alpha$-process. The analysis is corroborated by the temperature dependence of the spin-lattice relaxation time.' author: - 'S. Sturniolo$^{1}$, M. Pieruccini$^{2,\ddagger}$, M. Corti$^{1}$ and A. Rigamonti$^{1}$' - '$^\ddagger$ corresponding author: tel.: +39 059 2055654; fax: +39 059 2055651; e-mail: [email protected]' title: 'Slow motions detection in polybutadiene through novel analyses of MSE refocusing efficiency and spin-lattice relaxation ' --- {width="8"} Introduction {#intro} ============ Polymers are a class of materials which, although very common and thoroughly studied under a variety of aspects, e.g. for industrial applications, still represent suitable model systems for the investigation of some basic processes in condensed matter physics. Amorphous polymers in particular (or the amorphous component of the semicrystalline ones), manifest complex molecular relaxation dynamics covering a wide spectrum of time scales [@degenn; @doi_edw] which play a role in the kinetic phenomenon known as “glass transition”. This process entails a dramatic change in macroscopic mechanical properties taking place in a small temperature range around the critical glass transition temperature $T_g$. At microscopic level, the transition is governed by heterogeneous, cooperative, molecular relaxation processes with an unusually strong temperature dependence, often referred to as $\alpha$-process, as opposed to the $\beta$ process, more local in nature, characterized by a somewhat standard Arrhenius-like behavior. These phenomena are not exclusive to polymers, but in the latter the complex underlying molecular structure poses further difficulties for their description if compared to other glass-forming systems [@rel_proc_superc]. In this paper we use a novel approach to analyze H$^1$ NMR data in poly(butadiene) (PB), at the aim of extracting dynamical information on slow cooperative relaxation processes. The choice of PB is motivated by its amorphous nature, its convenient glass transition range and the existence of a wealth of literature data ranging from simulations to NMR and broadband dielectric spectroscopy [@gl_diel; @PB_decoup; @PB_diel_data; @PB_liq_melt; @PB_liq_melt_2]. In particular, the analysis of our experimental data is performed by means of a recently developed model [@pier_stu_FID] for the magic sandwich echo (MSE) refocusing efficiency, which is known to be affected by slow segmental motion [@MSE_stur; @NBR_publ; @demco]. The worked out information is then used to predict the temperature-dependence of the spin-lattice relaxation rate to check for consistency with independent experimental data. The results are highly promising and suggest that it is possible to combine the analysis of multiple NMR measurements to derive a reliable picture of slow relaxation dynamics around the glass transition temperature. Experimental {#exp} ============ Poly(butadiene) was supplied by Polymer Standards Service, Mainz, and composed by poly(1,4 butadiene) with a nominal average molecular weight of $\sim100,000\,\mathrm{g/mol}$, corresponding to $\sim2000$ monomeric units per chain, with minimal dispersity. Its glass transition temperature was around $\sim170\,\mathrm{K}$. The sample was transparent and its consistency at room temperature was of an extremely viscous liquid. Except during experiments, it was always stored at a temperature of $\sim 276$ K; when measurements had to be performed at a distance of days, fresh samples were always used. At relatively high temperatures, extremely sharp, liquid-like lineshapes were observed (corresponding to the long decay times that can be seen in Fig. 2 below). All this guarantees that any crystalline fraction, if present at all, can only appear in traces. The experiments were carried out using a TecMag “Apollo” DoubleResonance Spectrometer, in the working range of 5-450 MHz and a minimum digitization time resolution of 300 ns, and a Bruker BM-10 variable field electromagnet. The measurement chamber was an Oxford CF1200 cryostat able to operate in the temperature range between $4$ and $370\,\mathrm{K}$. Systematic measurements were performed at three different values of the static magnetic field, respectively around 0.5 T, 1 T and 1.5 T. The intensity of the RF pulse used was 30 G ($\pi/2$ pulse duration 2 $\mu$s). Since the deadtime of the receiver is almost $5\,\mu$s, simple acquisition of the free induction decay (FID) can fail when the decay of the signal is very fast and a significant part of it is lost. To avoid this problem, the FID signal has been refocused using the Magic Sandwich Echo sequence [@mse_seminal]. It has been recently shown [@NBR_publ] how the MSE refocused FID mantains the same shape as the original one, even when it is scaled down due to molecular motions. The MSE sequence used was the “non-ideal” version with a train of $\pi/2$ pulses replacing the long bursts described in the seminal paper [@mse_seminal], at the purpose to avoid problems due to instrumental phase switching times between different pulses. This sequence is substantially equivalent from a mathematical point of view to the original one. A phase switching time of 3 $\mu$s was used (doubled during the groups of four pulses along the X axis constituting the core of the sequence), and the total length of the MSE sequence was of 96 $\mu$s. Our analysis was performed with a self-developed software coded in C++. Part of the fittings was performed with the open source program EDDIE (Exact Dipole-Dipole Interaction Estimator which has been released to the public) and that is described in detail in Appendix \[app\_eddie\]. Theoretical framework {#the} ===================== Motivation ---------- At the core of the present paper is a novel aproach to describe the FID signal when the dipolar interaction between protons dominates the Hamiltonian and molecular motion is present. In order to treat the effect of motion on the FID in the case of polymers, reference is usually done to the chainlike structure of these molecules, and pre-averaging over fast segmental and $\beta$ motion is assumed in deriving analytical expressions for the discussion of experimental data [@demco; @brer]. In the attempt to investigate segmental dynamics when $T_g$ is approached from above, however, only the $\beta$ motion can be considered effective in the pre-averaging. Moreover, the manifestation of the cooperative nature of the $\alpha$-relaxation is related to the emergence of constraints which progressively quench the long wavelength components of the chain’s collective conformational fluctuations. As a further issue, cooperativity and the glass transition are not exclusive of polymeric systems; thus, referring to a scheme which is more “local”, in the sense that it is to some extent untied to a chain topology, would be desirable. For these reasons an expression for the transverse relaxation function $\overline{G}(t)$ has been previously derived [@pier_stu_FID] in the assumption that the system could be represented by an ensemble of spin pairs at a fixed distance with random orientations in space, uniformly distributed over all the solid angle. Different pairs were assumed independent. Segmental motion was described as an isotropic rotational diffusion, with a diffusion constant $D$. The prominent mean field character of this model supports its application for a meaningful analysis of the experimental data. In its crudeness, however, it matches the requirement of locality expressed above and, at the same time, offers the possibility to work out suitable analytical expressions which can be useful to discuss the results, as shown in [@NBR_publ]. In spite of being a rather crude reduction of the complexity characterizing the relaxation processes in polymers [@rel_proc_geom], this model has been substantially adopted to analyze data in similar contexts under the Anderson-Weiss approximation, as in [@papon]. In this respect, it is important that in the frame of our model the Anderson-Weiss results are approached when the temperature of the system is sufficiently above $T_g$, as it has been shown in [@pier_stu_FID]. Transverse relaxation function ------------------------------ Following the scheme outlined in the previous SubSection, the transverse relaxation function is expressed by the functional integral $$\label{fid_avg_mob} \overline{G}(t)\equiv \Re \int \delta\psi(\tau)\,p[\psi(\tau)]\, e^{i\int_0^t d\tau\Delta\omega [\theta(\tau)]},$$ where $p[\psi(\tau)]$ is the probability associated to an angular trajectory $\psi(\tau)$ of a spin pair during the time $\tau$, and the effect of its orientation with respect to the quantizing magnetic field **B**$_0$ (i.e. the angle $\theta$) is introduced through the term $\Delta\omega\equiv b\,P_2{\cos [\theta(\tau)]}$ related to the dipolar interaction, with $b$ the coupling constant and $P_2$ the second order Legendre polynomial. The function $\overline{G}(t)$ can be also recast in the form of an integral over the spin pair orientations at the ends of the time interval $[0,t]$: $$\overline{G}(t) = \frac{1}{16\pi^4}\Re\int d\psi_t \,d\psi_0 G[\psi_t,\,t;\,\psi_0,\,0] \,,$$ where the Green function $G[\psi_t,\,t;\,\psi_0,\,0]$ is connected to the probability that a spin pair, whose orientation is $\psi_0$ at time $t=0$, ends up with an orientation $\psi_t$ after a time $t$. The function $G$ is initially a Dirac $\delta$-function and progressively it broadens. Its evolution is described by the Dyson equation $$\label{4} \begin{array}{l} G[\psi_t,\,t;\,\psi_0,\,0] = G_0 [\psi_t,\,t;\,\psi_0,\,0] \,+ \\ \\ \,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\, i\,\int_0^t d\tau\int d\psi\, G_0 [\psi_t,\,t;\,\psi,\tau]\,\Delta\omega [\theta,\tau]\,G[\psi,\tau;\,\psi_0,\,0] \,, \end{array}$$ where the Green function $G_0$ relates to the stochastic evolution dynamics of the spin pair orientation angle $\psi(\tau)$. The present scheme is general and offers the possibility to consider diverse mechanisms for the evolution of $\psi$ provided that $G_0$ is known. Statistically independent motions affecting the orientation dynamics can be introduced by simple superposition. When gaussian statistics is being considered, the complexity of working out the transverse relaxation function could be reduced at the outset to some extent, because the relation $$\left\langle e^x \right\rangle = e^{\frac{1}{2}\left\langle x^2\right\rangle}$$ holds for the average of the associated stochastic variables. This is at the core of the Anderson-Weiss aproximation [@A_W]. To proceed further along the path of the Dyson equation, we chose a rotational diffusion process for the spin pair orientation. Of course, the fact that $G_0$ is known in this case is not of secondary importance; however, there are some further advantages in doing so. One is the possibility to find a check in models based on the Anderson-Weiss scheme, the others relate to some simplifications in estimating quantities of interest such as the MSE refocusing efficiency. The solution of the Dyson equation can be cast in the form of a series: $$\label{G_resid_1} \overline{G}(t) = \Re\left\{ R^{-1}\sum^{\infty}_{k=1} \textrm{res}[W, \omega_k]\, e^{-i\omega_k t}\right\} \,,$$ where $\omega_k$ is the $k$-th pole and $\textrm{res}[W, \omega_k]$ the corresponding residue of an appropriate kernel function $W\equiv W(\omega,D,b)$; $$\label{G_resid_2} R \equiv \sum^{\infty}_{k=1} \textrm{res}[W, \omega_k]$$ represents a normalization factor. In the case where $D=0$ all poles are real. The function $W$ has been expressed as a continuous fraction [@pier_stu_FID]; the $n$ poles and corresponding residues of its $n$-th order rational approximation can be used to form the partial sum $$\label{G_n} \overline{G}_n \equiv\Re\left\{R_n^{-1}\sum^{n}_{k=1} \textrm{res}[W, \omega_k] \exp(-i\omega_k t)\right\}$$ (with $R_n$ given by the corresponding partial sum in eq. \[G\_resid\_2\]), which reproduces exactly the transverse relaxation function $\overline{G}(t)$ up to a certain time $t_n$. On physical grounds, the truncation means that the evolution process of the whole system is described through a representative finite sub-ensemble of spin pairs \[up to $t_n$\]. For the typical values of the interaction constant in polymers (of order $\sim100$ kHz) and the usual length of a FID acquisition ($\sim250\,\mu$s), a value of $n\simeq\, 20$ is found suited to fit the data with high precision, but very often (see below) a much lower number of poles suffices. Analysis of literature data [@pier_stu_FID] and new experiments [@NBR_publ] indicate that, with regards to direct FID fitting, the model provides results in agreement with those obtained within the Anderson-Weiss approximation for medium and high motional frequencies, while it tends to overestimate orientational diffusivities at low temperatures. For values $D <\,10$ kHz in polymeric samples partial effects of multi-spin interactions due to the high density of protons cause a plateau in the measured diffusivity. Turning back to the transverse relaxation function, note that $\overline{G}(t_1 + t_2) \neq \overline{G}(t_1) \overline{G}(t_2)$; indeed $$\label{Gt1t2} \overline{G}_n (t_1 + t_2) = \Re\left\{R_n^{-1}\sum^{n}_{k=1} \textrm{res}[W, \omega_k] e^{-i\omega_k (t_1 + t_2)}\right\}.$$ Therefore, while the coefficients $\textrm{res}[W, \omega_k]$ describe the evolution of the sub-ensemble starting from an initial condition ($t=0$) where the Green function is a $\delta$, the coefficients $\textrm{res}[W, \omega_k]\,\exp\{-i\omega_k t_1\}$ describe the evolution for $t\geq t_1$ from an “initial” condition (i.e. at $t=t_1$) where the Green function has already broadened to some extent. MSE efficiency -------------- MSE allows one to refocus an eco of a dipolar dephased FID with excellent fidelity even long after its decay (more than $100\,\mu s$). If the coupling strength in the system remains constant during the whole experiment, then the refocusing will be complete and the ratio $\eta$ between the intensity of refocused FID to that of the original one is unity ($\eta = 1$). This condition may break down due to molecular motions, causing the decrease of the amplitude of the echo. As a function of the extent of molecular motions, $\eta$ is close to one for frequencies that are very low or very high compared to the order of magnitude of the dipolar coupling constant, being drastically reduced when the two frequencies are of the same order of magnitude [@MSE_stur]. An estimate of the MSE refocusing efficiency may be obtained following the evolution of the representative sub-ensemble when a pulse sequence $\{\tau_+ |\,4\tau_{-} |\,\tau_+\}$ is imposed, such that in the intermediate interval the time is apparently inverted with regards to the evolution hamiltonian (with the exclusion of the diffusion process, of course). After eq. \[Gt1t2\] and the related comments, one finds $$\label{eff} \eta = \overline{G}_n (6\tau) = \Re\left\{R_n^{-1}\sum^{n}_{k=1} \textrm{res}[W, \omega_k] \,e^{-i\left(2\left.\omega_k\right|_{\tau_+} - 4\left.\omega_k^*\right|_{\tau_-}\right)\tau}\right\} \,,$$ where the poles $\left.\omega_k\right|_{\tau_+}$ and $\left.\omega_k^*\right|_{\tau_-}$ are calculated with coupling constants $b$ and $-b/2$ respectively. (Note the resemblance of eq. \[eff\] with eq. \[fid\_avg\_mob\].) This expression will be subsequently used for the analysis of the temperature profile of the MSE refocusing efficiency $\eta$. Its relation with the solution of the problem derived by setting $\Delta\omega=bP_2[\Theta(t)-3/2\Theta(t-\tau)+3/2\Theta(t-5\tau)]$ in the Dyson equation (with $\Theta$ the unit step function), is not at all trivial and is currently being subject of a detailed study. In the present context we must limit ourselves to propose it, relying on both the physical argument at the basis of its derivation and the satisfactory analysis presented below. Figure 1 shows the efficiency as a fuction of $D$ obtained from Eq. \[eff\] for different numbers of poles and an MSE sequence of $96\,\mu\mathrm{s}$. Two values of the coupling constant have been considered; one of them is close to that appropriate for two protons a distance $1.8\,\textup{\AA}$ apart ($b=194.107\,\mathrm{kHz}$). Note that a significant dependence on the number of poles only shows up for low values of $D$. This has to be taken into account when extending the analysis of the efficiency data towards $T_g$. ![[]{data-label="fig1"}](Fig1.eps){width="9"} The efficiencies predicted by eq. \[eff\] are also compared with the expression reported below, which has been derived in [@demco] within the Anderson-Weiss approximation: $$\label{EffAW} \eta = 1 - M_2 \tau_c^2\left[ e^{-6\tau/\tau_c} - 3e^{-5\tau/\tau_c} + \frac{9}{4}e^{-4\tau/\tau_c} + 3e^{-\tau/\tau_c} + \frac{3\tau}{\tau_c} - \frac{13}{4}\right] \,.$$ Since $\tau_c$ is in fact the relaxation time of $\left\langle P_2\right\rangle$, the relation $\tau_c=(6D)^{-1}$ holds [@pier_stu_FID]. Equation \[EffAW\] is strictly valid for $\eta$ close to unity [@demco], but has been plotted in the whole $D$ interval to point out how it compares with eq. \[eff\] (setting its value to zero wherever negative, of course). From a qualitative point of view the efficiencies calculated with the two above expressions are similar. The extrapolation of eq. \[EffAW\] to low $\eta$ values underestimates the efficiency predicted by eq. \[eff\] unless $D$ is small enough; then, the behavior is reversed. This compensation may play some role when average efficiencies calculated with eqs. \[eff\] and \[EffAW\] are fitted to the data, as the results are found similar to some extent (see below). Spin-lattice relaxation ----------------------- Spin-lattice relaxation times will be considered for testing the parameters of the motional distribution derived from the analyses of the FIDs and of the MSE efficiency. In particular, the T$_1$ data as a function of temperature will be compared with those obtained by the equation $$\label{T1_dipolar} \frac{1}{\textrm{T}_1} = \frac{9}{8}\frac{\gamma^4\hbar^2}{r^6}\left(\frac{\mu_0}{4\pi}\right)^2\left[\left\langle J^{(1)}(D_c, \omega_L)\right\rangle + \left\langle J^{(2)}(D_c, 2\omega_L)\right\rangle\right]$$ where $$\label{J_integr} \left\langle J^{(i)}(D_c, \omega_L)\right\rangle \equiv \int_{-\infty}^{\infty}{g(\mathrm{ln}D,D_c)J^{(i)}(D, \omega_L)d\mathrm{ln}D},$$ are taken as superpositions of single-$D$ contributions. The shape of the distribution $g(\mathrm{ln}D, D_c)$ is determined by the parameters worked out from the MSE efficiency; $D_c$ is the central relaxation rate (rotational diffusivity in our case) for the distribution. The integral is carried over the logarithm of the frequency, according to a linear distribution of energy barriers [@logMOM]. As for the temperature dependence of $D_c$, two different choices have been considered. More details will be given in the following Section, dealing with the experimental data. Experimental Results {#res} ==================== FID {#res_FID} --- The analyses of the FIDs have been performed both on the basis of our model and with the expression below, derived within the Anderson-Weiss approximation [@papon]: $$\label{FID-AW} I_{\scriptscriptstyle{\rm{FID}}}(t) = \exp \left[-M_2 \tau_c^2\left(e^{\tau/\tau_c} + \frac{t}{\tau_c} -1\right)\right] \,.$$ Following the lines of Ref. [@papon], the value of $b\simeq 300$ kHz which has been obtained from the $T=173$ K FID fitting (i.e. $\sim300$ kHz using $G_n$ and $\sim275$ kHz using eq. \[FID-AW\]) has been taken as a fixed parameter for the analysis of the higher temperature FIDs. The values of the diffusion constant worked out with our expression were found to be practically the same as those obtained with eq. \[FID-AW\] for $T \gtrsim 190$ K. For some intermediate temperatures around $T=213$ K, the quality of the fittings was found to degrade slightly, independent of the analytical expression used for the FID. Considering a distribution of $D$ for such cases (i.e. a Gaussian) could hardly improve the fittings a little, and not in all cases. By the way the worked out average $D$ did not significantly differ from that derived using a single-$D$ expression. Some of the FIDs and their fittings are shown in Fig. 2. As it can be seen, the crossover between a rigid-like dipolar dephasing and a mobile FID decay takes place somewhere around $200\,\mathrm{K}$, above the calorimetric glass transition temperature. This is consistent with the fact that this process is sensitive to a shorter time scale, around $10\,\mu\mathrm{s}$, while the ordinary macroscopic techniques used to assess the glass transition, probe motions on a time scale of seconds. ![[]{data-label="fig2"}](Fig2.eps){width="9"} The fitted values of the diffusivity $D$ are shown in Fig. 3. A finite plateau with $D$ of the order of 30 ms$^{-1}$ at low temperatures was found, which is likely due to competitive relaxation process and/or multi-spin interactions. Using eq. \[FID-AW\] above, however, didn’t improve the situation significantly since, e.g., best fit values of $D=$ 7.6, 15 and 17 ms$^{-1}$ were obtained from FIDs at $T=$ 173, 178 and 183 K respectively. ![[]{data-label="fig3"}](Fig3.eps){width="10"} For temperatures above $T_g$ a steepy growth of $D$ is observed, and a maximum limit is reached once the motions reach frequencies so that the dipole-dipole interaction is averaged out and field inhomogeneities dominate the line-width and then the related FID’s. No change was found fitting FIDs measured at different values of the static magnetic field in the temperature of interest. A Vogel-Fulcher-Tammann (VFT) law, $$\label{VFT_form} D(T) = \frac{1}{\tau_{\infty}}\exp\left(\frac{-AT_{\scriptscriptstyle{\rm{VFT}}}}{T-T_{\scriptscriptstyle{\rm{VFT}}}}\right) \,,$$ has been adjusted to the data and the following values of the best fit parameters have been obtained: $\tau_{\infty}=8.86\cdot10^{-9}\,\mathrm{s}$, $T_{\scriptscriptstyle{\rm{VFT}}} = 148\,\mathrm{K}$ and $A = 2.5$. The resulting expression is plotted as a dashed line in Fig 3. MSE efficiency {#res_MSE} -------------- As already mentioned, direct evaluation from the fitting of the FID leads to overestimate the $D$ values, particularly at low temperatures. Thus, in order to work out reliable information from the analysis of the MSE efficiency, $D$ vs. $T$ data obtained from dielectric spectroscopy [@PB_diel_data] have been used, and the temperature dependence of the $\alpha$ relaxation in the $T$-range of interest has been taken into account. (We note that VFT extrapolations and experimental values for the $\beta$-process reported in [@PB_diel_data], indicate that the latter can be considered much faster than the $\alpha$ relaxation only marginally at $T=273$ K, i.e. the highest temperature value explored in our measurements.) A first assessment was made by considering a single-$D$ relaxation, with a $D$ vs. $T$ dependence given by the VFT parameters provided by [@PB_diel_data], namely, $\tau_{\infty} = 4.8\cdot10^{-13}\,\mathrm{s}$, $T_{\scriptscriptstyle{\rm{VFT}}}=142\,\mathrm{K}$ and $A=7.96$. In Fig. 4 the MSE efficiency measured for PB is reported, and compared with the expressions given by eq. \[eff\] calculated for $n=20$ and by eq. \[EffAW\] in the hypothesis of single relaxation time. ![[]{data-label="fig4"}](Fig4.eps){width="10"} As it can be seen, the predictions remarkably match the temperature range where one has the efficiency dip. Better approximations can be obtained considering a distribution of relaxation times. To this aim, for each temperature the efficiency was calculated as an integral over the logarithm of the rotational diffusion constant (cf. the linear dependence of $\overline{G}$ on $p[\psi(\tau)]$ in eq. \[fid\_avg\_mob\]): $$\label{eff_distr} \eta(T) = \int_{-\infty}^{\infty}{F_{\alpha}(D, D_{\alpha}(T))\eta(D)}\, d\,\mathrm{ln}D$$ where $F_{\alpha}$ is the (normalized) distribution associated to the $\alpha$-process and is taken in the form $$\label{distrHN} F_{\alpha}(D, D_c) = \frac{1}{\pi}\frac{\left(D_c/D\right)^{ac}\sin(c\theta)}{\left[1+2\left(D_c/D\right)^a \cos\left(\pi a\right)+\left(D_c/D\right)^{2a}\right]^{-c/2}} ,$$ where $$\label{theta_1} \theta = \mathrm{atan} \left[ \frac{\sin(\pi a)}{\left( D_c/D \right) ^{a}+\cos(\pi a)}\right]$$ if the argument of the arctangent is positive and $$\label{theta_2} \theta = \mathrm{atan} \left[ \frac{\sin(\pi a)}{\left( D_c/D \right) ^{a}+\cos(\pi a)}\right] + \pi$$ otherwise; $a$ and $c$ (both positive and not larger than one) are the width and symmetry parameters of the distribution. The reason for assuming eq. \[distrHN\] is that $F_{\alpha}$ “generates” the Havriliak-Negami distribution, i.e. $$\label{HN_integral} \frac{1}{[1 + (i\omega/D_c)^a]^c} = \int_{-\infty}^{\infty}{\frac{1}{1+i\omega/D}}F_{\alpha}(D, D_c)\,d\,\mathrm{ln}(D) \,,$$ which is used very often to fit relaxation processes in dielectric spectroscopy. Equations \[HN\_integral\] and \[J\_integr\] share the same structure, with the difference that the single frequency spectral density is replaced here by the dielectric response of a Debye process. The single-$D$ behavior reported in Fig. 4 suggests that data analysis can only be performed for temperatures approximately above $188$ K if no other mechanism controlling the MSE efficiency is included on top of the one considered. Thus we limit data fitting to the interval $T \geq 193 K$ (except in one case, where also the efficiency at $T=188$ K has been considered) and just for the 22 MHz data set. The solid lines in Fig. 5 refer to the efficiency expressed by eq. \[eff\] in the case where the VFT parameters of $D_c(T)$ were given the values of either ref. [@PB_diel_data] (line $a$) or those derived from the analysis of the FIDs (line $b$). A rather bad performance is evident in the latter case. ![[]{data-label="fig5"}](Fig5.eps){width="10"} Notwithstanding the fact that some integration subinterval is out of the domain of validity of eq. \[EffAW\], the same calculation has been performed within the Anderson-Weiss approximation. The results are shown in Fig. 5 as dashed lines, again derived assuming the two $D_c(T)$ VFT dependencies as above. The striking similarity between the two forms of $\eta (T)$ seems to suggest that the relevant character for a description of the experimental data is a *qualitative* nature of the $\eta$ vs. $D$ dependence. Table \[tab1\] reports the best fit values of the Havriliak-Negami parameters obtained so far. ---------------------- ----------------------- --------------------------------- ------------     $D_c(T)$         $T$-range (K)         $\overline{G}_6(6\tau)$         AW     193-273 $a=0.50$ $a=0.53$ $c=0.61$ $c=0.52$ 223-273 $a=0.56$ $a=0.56$ $c=0.69$ $c=0.59$ from [@PB_diel_data] 193-223 $a=0.49$ - $c=0.63$ - 188-213 $a=0.45$ - (with 10 poles) $c=0.59$ - from FID 193-273 $a=0.80$ $a=0.82$ fitting $c=0.67$ $c=0.54$ ---------------------- ----------------------- --------------------------------- ------------ : Values of the $\alpha$-relaxation parameters $a$ and $c$ obtained by fitting the MSE efficiency in different temperature ranges. All of them have been derived assuming the $D_c(T)$ dependence provided by ref. [@PB_diel_data], except in the last line, where $D_c(T)$ is a VFT tracing the $D$ values obtained from fitting the FIDs.[]{data-label="tab1"} \[a\_c\_table\] Fittings have been also performed within different temperature subintervals, and the worked out values of the Havriliak-Negami parameters are reported in the table. It is worth noticing that width and asymmetry of the $F_\alpha$ profile increase when the average temperature of the fitting interval decreases. This behavior is more evident in Fig. 6. ![[]{data-label="fig6"}](Fig6.eps){width="10"} With regards to eq. \[EffAW\], instead, the analysis is limited to a comparison between the results worked out on the whole $T$-interval and the high-$T$ region for obvious reasons. As is evident, the changes in the parameters is not as pronounced as in the case where eq. \[eff\] is used. The values of the shape parameters derived in the whole $T$-range can be also compared with those derived from dielectric analysis in ref. [@colmenero], namely, $a=0.72$ and $c=0.50$. Spin-lattice relaxation {#res_T1} ----------------------- Finally, the spin lattice relaxation time as a function of the temperature has to be discussed. The presence of a broad relaxation distribution underlying the spin lattice process can be qualitatively guessed already from the fact that the maxima in the relaxation rates are inversely proportional to the strength of the field, namely $(1 / T_1)_{max} \propto \omega_L^{-1}$ (with $\omega_L$ the Larmor frequency) rather than to the square of the inverse $\omega_L$, as expected in the case of single frequency characterizing the dynamics. The analysis was carried out by using the method described in Section \[the\], eqs. \[T1\_dipolar\] and \[J\_integr\]. With reference to Fig. 7a, for each set of data three curves have been plotted: two of them correspond to the same $D_c(T)$ dependence provided by ref. [@PB_diel_data] but two different $T$-intervals for the efficiency fits; the other one is obtained taking the VFT parameters of Fig. 3. A constant baseline of 3 s$^{-1}$ has been added to all curves for a rough account of all those faster processes that cause relaxation at low temperatures but are not described by our motional distribution. The shape parameters used to draw the lines are those derived from the analysis of the 22 MHz MSE efficiency, for this reason some mismatch can be found with the 62 MHz data. This has been done at the aim to assess how the spin-lattice relaxation profile can be inferred from the analysis of the MSE efficiency in different conditions. ![[]{data-label="fig7"}](Fig7.eps){width="10"} Curves corresponding to narrow $D$ distributions ($a$ close to unity) manifestly fail to catch the corresponding T$_1^{-1}$ maximum. Considering the 22 MHz data (to which indeed the shape parameters refer), it is evident that the maximum is better approached when the high-$T$ interval shape parameters are taken, i.e. for a moderately narrower distribution. On the contrary, the low-$T$ data are better described with a broader relaxation time distribution. This indicates that also in this circumstance, accounting for the appropriate $T$-dependence of the shape parameters would be desirable to improve the fittings, and that providing just their “mean” values worked out from such wide $T$-range does not give a detailed analysis of the $\alpha$-relaxation. Figure 7b reports the T$_1^{-1}$ vs. $T$ dependence ($\omega_L=22$ MHz) obtained considering a single relaxation time ($a=1$), with either the VFT parameters of ref. [@PB_diel_data] or those obtained from the fittings of the FIDs. In the former case a maximum of $\sim 130$ s$^{-1}$ is found at a temperature around 247 K; in the latter, the temperature of the maximum shifts to a value of $\sim 470$ K. Concluding remarks ================== A relevant issue emerging from the analysis carried out so far is that the frequency profile of the $\alpha$ process can be extracted from the $T$-dependence of the MSE refocusing efficiency. Most importantly, the analysis is able to reveal the temperature dependence of the relaxation time distribution characterizing the $\alpha$-process; moreover, the results are consistent with the expected trend. In fact this kind of relaxation can be generally described by means of a stretched exponential $\exp\{-(t/\tau_{\scriptscriptstyle{\rm{KWW}}})^\beta\}$ (the Kohlrausch-William-Watts function), with the exponent $\beta\leq 1$ decreasing on lowering the temperature [@phillips]. On the other hand, from the relation $ac\approx\beta$ [@colmeneroHN] it is easy to check that our best fit values of the Havriliak-Negami parameters follow the expected temperature behavior. The adopted model catches the main features of schemes derived within the Anderson-Weiss approximation. In this sense, the good matching found of the results has to be considered a valuable support to our theory. The results reported here, further extend the agreement to the analysis of the MSE refocusing efficiency. Referring to some specific issues of the present report, we consider first the rather crude estimate of $D$ from the fittings of the FIDs. The results show that at low temperatures our model overestimates this quantity with respect to the case where the Anderson-Weiss based model is used. Apart of this modest discrepancy, both approaches lead to substantially the same VFT parameters for the $D$ vs $T$ dependence, which on the other hand differ quite markedly from those found from dielectric analysis. The rather good results obtained with dielectric VFT parameters on *both* the T$_1$ and MSE efficiency profiles, indicate an inconsistency in the FID analysis, or at least in its interpretation. We don’t want to analyze this aspect in detail in the present context, but note that the *single relaxation time* efficiency ($a=c=1$) accommodates very well among the data when $D(T)$ as obtained from FID’s analysis is used (and if the minimum is not approached too closely; see Fig. 8, showing a detail of Fig. 4). Of course this occurrence is not significant with regards to the spin-lattice relaxation profile, as the maximum wouldn’t appear in the relevant $T$-range. ![[]{data-label="fig8"}](Fig8.eps){width="10"} Overall, $D(T)$ is progressively overestimated as $T$ decreases, also reaching a finite plateau below $T_g$. This is motivated by at least a basic assumption of the model, namely, that $\Delta\omega[\theta(\tau)]$ in Eq. \[fid\_avg\_mob\] only involves a single spin pair. In reality a given spin interacts also with others at comparable distances. Thus the fluctuations in $\Delta\omega$, which are due to multispin interaction, are ascribed by the model to the motion of just one pair. This circumstance and the fact that at long time an increasing number of spins correlates, might well be at the origin of the problem, becoming ever more important as the motion is slowing down, i.e. as $T \rightarrow T_g$. (See e.g. [@correlations] for solids; roughly speaking, the effect of a neighbouring spin starts being significant after a time of the order of the inverse coupling constant.) On the other hand, each spin pair interaction term in the multispin hamiltonian is virtually “reversed” by the pulse sequence, and contributes separately to the formation of the echo (i.e. the effect of the reconstruction on the density matrix factorizes). This means that in this circumstances the response of the system differs very little from that of our simplified model. For this reason the introduction of a “correct” $D_c$ vs. $T$ dependence appears to be crucial for a reliable analysis of the refocusing efficiency. The need to take into account other experimental techniques (such as dielectrics in this case) is not uncommon in polymer physics, and it has been crucial for highlighting the possibilities offered by a proper analysis of the MSE efficiency. From the practical point of view, however, the present state of the art is rather unsatisfactory, since a complete, self contained analysis of the slow motions via H$^1$ NMR would be desirable (e.g. in those cases where dielectric analysis would be difficult, like in polyolefines). In this respect, the present results address the opportunity of extending the model to account for multispin effects and render the direct FID analysis more reliable. As another issue, noting how a correct $T$-dependence of the central relaxation time affects the quality of the fittings overall, in particular with the respect to the correct matching around the minimum of the efficiency (see Figs. 5 and 8), the possibility to derive good VFT parameters from a joint analysis of the efficiency and of the spin-lattice relaxation rate (which is sensitive to the shape parameters of the distribution) seems to be a reasonable target. Work in this direction is underway. In concluding, the results reported in the present paper are highly promising and indicate the possibility of reliable analyses based on multiple NMR data in order to extract significant insights on the dynamics around the glass transition temperature in polymers. Acknowledgements ================ S.Sturniolo gratefully acknowledges extensive discussions with Prof. K. Saalwächter about MSE refocusing efficiency and its use for the analysis of slow motion in H$^1$ NMR experiments. Fitting software {#app_eddie} ================ This Appendix contains technical details about EDDIE, the fitting software developed in order to implement the theoretical approach. EDDIE stands for Exact Dipole-Dipole Interaction Estimator. We will provide all the necessary details as well as a short user manual. The program has been released to the public at the address https://sites.google.com/site/eddienmr/home. Technical details {#appD_EDDIE_tech} ----------------- The purpose of the EDDIE software is to simulate, given the parameters $b$ and $D$, a FID signal using the function $\overline{G}(t)$ of Section \[the\] (reported explicitly in [@pier_stu_FID]) and fitting raw experimental data with it. We can summarize the steps required to generate a FID signal as follows: - choose a value for the parameters $b$ and $D$, as well as the number of poles $n$, which will determine the precision of the final result; - calculate the kernel of the anti-fourier transform; - find the roots in $\omega$ for the denominator of the kernel (an equation of $n$-th degree), thus identifying the poles; - calculate the residues of the kernel in each of the poles; - calculate the FID by performing a proper sum over the residues, with the formula of eq. \[G\_n\]. Fitting the experimental data requires an additional step. In fact, since the function is not analytical, our best option is to use a simplex method to calculate the FID in fixed points of the $b-D$ plane and then refine our search using the residual sum of squares as a parameter to minimize. The procedure has to be iterated many times, as the fitting simplex moves in the function’s domain, and becomes the most time-consuming step of the entire process. However this is not too demanding and the program, running on a common laptop, is able to perform a 1000 step simplex fitting in less than a minute. The program accepts $b$, $D$ and $n$ as user input; in lack of an input, it has default values for them. The following calculation of the kernel requires operations between complex polynomials. EDDIE was written in C++. In order to simplify portability, dependencies from external libraries were avoided. Therefore, the software comes with its own *complex* and *polynomial* classes with overloaded operators. The solution of the denominator of the kernel is carried out by using a Jenkins-Traub algorithm. In this case we are using with permission a version for complex roots written by Henrik Vestermark and publicly available on his site [@hvks]. From our tests, and in our specific case, the algorithm seems to give reliable results only for polynomials up to the 35th degree. We tried improving that by changing algorithm or making use of high precision libraries, without success. This does not constitute a major problem as a number of poles of 20 or so gives excellent precision for most practical applications. After the poles are found (a *FID* class has been written that keeps the solutions in memory after finding them as part of its initialization), the procedure is rather straightforward, as by inserting the desired time $t$ it is automatically possible to carry out the sum in eq. \[G\_n\] and thus find the FID. The fitting procedure, as mentioned before, makes use of a simplex algorithm. Since it is important to carry out the procedure in a limited domain of possible values and the function to minimize is rather difficult to handle, as it has many local minima which can cause wrong fitting, a special algorithm developed to be a “constrained, global and bounded Nelder-Mead method” by Luersen et al. was used [@GBNM]. This algorithm works like a regular Nelder-Mead simplex optimization procedure, with a few differences: - the search goes on for a predefined number of steps rather than waiting for a condition to be satisfied; - the simplex is constrained to stay inside a fixed domain - the search is interrupted and restarted if it either finds a local minimum or if it gets stuck or deformed by the boundaries; - in order to improve the probability of finding a global minimum, every time a local minimum is found, a bias function is calculated in order to make it less likely that the search restarts from the vicinity of that point. User manual {#appD_EDDIE_UM} ----------- EDDIE can be operated in two ways: by console or with an input file. The two methods are very similar, with only a few differences. Besides fitting and simulating a FID it is possible to verify or change the values of the internal variables of the program, which control parameters like the boundaries of the fitting, the number of poles used, etc. Each variable is initialized to a default value when the program is started: console commands allow to interact with them. The program is run from the system console by simply typing its name and hitting return: user: \$ eddie At this point, the internal console of EDDIE will show up. In this context, the general syntax for any command is: &gt; \[COMMAND\] \[ARGUMENTS\] Commands and arguments are case-sensitive. This is a list of possible commands: 1. get - Print the value of a variable Arguments: \[VARIABLE NAME\] 2. set - Modify the value of a variable Arguments: \[VARIABLE NAME\] \[NEW VALUE\] 3. fit - Fit the contents of an ASCII data file Arguments: \[FIT TYPE\] 4. sim - Simulate a FID from scratch Arguments: \[NUMBER OF STEPS\] 5. help - Print this help Arguments: &lt;none&gt; 6. exit - Quits the program Arguments: &lt;none&gt; The *get* and *set* commands are meant to respectively print and modify the value of the variable whose name is passed as an argument. Here is a list of the valid variable names: - **b\_[min]{}, b\_0, b\_[max]{}** - Boundaries and central value for coupling constant b (kHz) - **D\_min, D\_0, D\_max** - Boundaries and central value for diffusivity D ($\mathrm{ms}^{-1}$) - **X\_min, X\_0, X\_max** - Boundaries and central value for X = D/b (values change accordingly to follow b and D) - **t\_min, t\_0, t\_max** - Boundaries and central value for time (variable unit). Used as limits in fits and simulations. t\_max = -1 means that there is no upper limit - **poles** - Number of poles for FID truncation - **iters** - Number of simplex iterations for fitting - **thr** - Threshold for calculating fitting convergence - **t\_col, data\_col** - Indices for columns of time and data in input file, respectively - **t\_unit** - Time unit in input file (seconds, standard value = 1E-3 s) - **norm** - Normalization factor for data in input file - **skip\_l** - Lines to skip at the beginning of the input file - **start\_p, end\_p** - Starting and ending data points in input file. end\_p = -1 means that there is no upper limit - **skip\_p** - Point skipping step in input file - **input\_file** - Input file name - **output\_file** - Output file name The meaning of most of the variables is self-explaining. The parameter $X=D/b$ is introduced because it is easier to grasp its value by eye and is best fitted than $D$ - in general, $X=0.01$ means an almost perfectly rigid FID, while $X=10$ is a completely mobile one. The central values, *b\_0, D\_0* and *X\_0* are used as fixed values in simulations, while in fittings they matter only when a variable is kept constant. The boundaries apply to fittings and are not relevant to simulations. Of course, since the $D$ and $X$ variables are interdependent, any change applied to one will reflect on the other. The $t$ variables are slightly different: in fittings, the boundaries represent the time limits of the data file on which the residual sum of squares is calculated, in simulations the time interval to simulate. If there is no upper limit, simulations will use a default calculated value.Number of poles, of simplex iterations, and threshold for convergence (basically the criterion of acceptance of a local minimum in the Nelder-Mead algorithm) have optimal default values that usually do not need to be changed. For the fitting, the raw data must be contained in an ASCII data file. The *t\_col* and *data\_col* are indices of the columns containing respectively the time and the FID data; it is possible to have $data\_col < t\_col$, but they must not be equal. It is possible as well to configure a number of lines to skip, for example to remove a textual header from the data file (*skip\_l*), or to fit only on one point each *skip\_p* points - this is useful to speed up the fitting if the raw file has a high time resolution. It is possible to scale both time (*t\_unit*) and data (*norm*) by a constant factor; it must be remembered that the data has to be normalized in a way that it goes to 1 at $t=0$ for the fitting to work. Finally, the paths of the file to read (*input\_file*) and to write (*output\_file*) can be inserted. When inserting the names, no apices or quotation marks should be used. At the beginning of each file name the program will add either *fitted\_* or *simulated\_* in order to make the files recognizable and prevent accidental overwriting.For some of these variables there are values that are not acceptable (for example, $start\_p < 0$) and values that conflict with other variables (for example, $D\_min > D\_max$). When a not acceptable value is inserted with the function *set*, an error message is printed and the value of the variable is not changed. Every time a variable is changed with *set* the new value is printed immediately to confirm the effect of the command.The command *fit* accepts one argument of type *FIT TYPE*. This is simply a string which can assume three values: *b*, *D* and *bD*. Its purpose is to indicate which fitting parameters must be found. In this way, for example, inserting: &gt; fit bD will run a two dimensional fitting on both $b$ and $D$, while the input: &gt; fit D will only fit $D$ while keeping $b=b\_0$.The command *sim* requires for an argument only the number of steps for the simulation. Remember that, however, the *length*, in time units, of the simulation is controlled by *t\_min* and *t\_max*: the number of steps will only affect the resolution with which the FID is simulated on this interval.If one needs to fit many data files with similar structure in one go, then it becomes convenient to make use of the possibility to control the program via input file. This is done simply by running the program with the files to be fitted as arguments into the system console: user: \$ eddie file\_to\_fit\_1.dat file\_to\_fit\_2.dat ... In this case, there must be, in the same folder of the program, a file *parameters.txt*, which will contain the instructions for the fitting. These instructions will simply be a sequence of commands as the ones described earlier: the program will simply run all these instructions in sequence in its console before performing the fitting. The program is *not* able to simulate using an input file, and it is not necessary to insert a *fit* line at the end of the input file. The only difference between using the program with the console or with an input file is that in the latter case, an instruction to *set* either $b\_0$, $D\_0$ or $X\_0$ will fix the set variable for the fitting. In other words, if one wants the fitting to find both $b$ and $D$, there must be no *set* instructions for these three variables; on the other hand, setting, for example, $b\_0$, will result in a fitting running only on the parameter $D$ while keeping $b=b\_0$. Finally, the structure of the output files is rather straightforward. In both a simulation and a fitting, the file begins with a two lines header. The first line contains the values of $b$ and $D$ used (either the fixed ones in a simulation, or the fitted ones in a fitting) and the second is just an indication of what the various columns contain: b = 386.586 kHz D = 1737.96 kHz Time (ms) Data Fitted In this case, the example was taken from a fitting file, and there are three columns: time, data (the original fitted data) and fitting (the best fitting FID found by the program). In the case of a simulation, there would be only two columns, time and FID. The fitting file is produced always in the same folder as the file containing the original data. When fitting from an input file, the program also produces a further output file, in its own working directory, called *results.txt*. This file will contain three columns with, respectively, the names of the fitted files and the found values for $b$ and $D$. [10]{} P. de Gennes, [*Scaling concepts in polymer physics*]{}. Cornell University Press, 1979. and [S.F. Edwards]{}, [*The theory of polymer dynamics*]{}. Clarendon Press, 1994. and [L. Sjögren]{}, “Relaxation processes in supercooled liquids,” [*Rep. Prog. Phys.*]{}, vol. 55, pp. 241–376, 1992. , [S. Benkhof]{}, [T. Blochowicz]{}, [C. Tschirwitz]{}, and [E. Rössler]{}, “The dielectric response of simple organic glass formers,” [*Jour. Mol. Struct.*]{}, vol. 479, pp. 201–218, 1999. , [R. Zorn]{}, [B. Farago]{}, [B. Frick]{}, and [L.J. Fetters]{}, “Decoupling of time scales of motion in polybutadiene close to the glass transition,” [*Phys. Rev. Lett.*]{}, vol. 68, pp. 71–74, 1992. and [S.R. Nagel]{}, “Dielectric susceptibility measurements of the primary and secondary relaxation in polybutadiene,” [*Phys. Rev. B*]{}, vol. 52, pp. 5653–5656, 1995. , [C. Gainaru]{}, [H. Schick]{}, [A. Brodin]{}, [V.N. Novikov]{}, and [E.A. Rössler]{}, “From a simple liquid to a polymer melt: Nmr relaxometry study of polybutadiene,” [*Phys. Rev. Lett.*]{}, vol. 97, p. 207803, 2006. , [A. Brodin]{}, [C. Gainaru]{}, [A. Hermann]{}, [H. Schick]{}, [V.N. Novikov]{}, and [E.A. Rössler]{}, “From simple liquid to a polymer melt. glassy and polymer dynamics studied by fast field cycling nmr relaxometry: Low and high molecular weight limit,” [*Macromolecules*]{}, vol. 41, pp. 5313–5321, 2008. and [M. Pieruccini]{}, “An exact analytical solution for the evolution of a dipole-dipole interacting system under spherical diffusion in a magnetic field,” [*J. Mag. Res.*]{}, vol. 223, pp. 138–147, 2012. and [K. Saalwächter]{}, “Breakdown in the efficiency factor of the mixed magic sandwich echo: A novel nmr probe for slow motions,” [ *Chemical Physics Letters*]{}, vol. 516, pp. 106–110, 2011. , [M. Pieruccini]{}, [M. Corti]{}, and [A. Rigamonti]{}, “Probing $\alpha$-relaxation with nuclear magnetic resonance echo decay and relaxation: a study on nitrile butadiene rubber,” [*Solid State NMR*]{}, vol. 51–52, pp. 16–24, 2013. R. Fechete, D. Demco, and B. Blümich, “Chain orientation and slow dynamics in elastomers by mixed-magic-hahn echo decyas,” [*The Journal of Chemical Physics*]{}, vol. 118, pp. 2411–2421, 2003. , [A. Pines]{}, and [J. S. Waugh]{}, “Time-reversal experiments in dipolar-coupled spin systems,” [*Physical Review B*]{}, vol. 3, no. 3, pp. 684–696, 1971. M. Brereton, “An exact expression for the transverse nuclear magnetic resonance relaxation of a dynamic scale invariant polymer chain governed by a single relaxation time,” [*J. Chem. Phys.*]{}, vol. 94, no. 3, pp. 2136–2142, 1991. , [A. Heuer]{}, and [H.W. Spiess]{}, “Geometry of reorientational dynamics in supercooled poly(vinyl acetate) studied by $^{13}$[C]{} two-dimensional nuclear magnetic resonance echo experiments,” [*J. Chem. Phys.*]{}, vol. 111, pp. 3720–3727, 1999. , K. Saalwächter, K. Schäler, L. Guy, F. Lequeux and H. Montes, “Low-field nmr investigations of nanocomposites: Polymer dynamics and network effects,” [*Macromolecules*]{}, vol. 44, pp. 913–922, 2011. P. Anderson and P. Weiss, “Exchange narrowing in paramagnetic resonance,” [*Rev. of Modern Physics*]{}, vol. 25, pp. 269–276, 1953. R. Zurn, “Logarithmic moments of relaxation time distributions,” [*J. Chem. Phys.*]{}, vol. 116, pp. 3204–3209, 2002. A. Arbe, D. Richter, J. Colmenero and B. Farago, “Merging of the $\alpha$ and $\beta$ relaxations in polybutadiene: A neutron spin echo and dielectric study” [*Phys. Rev. E*]{}, vol. 54, 3853–3869, 1996. J. C. Phillips, “Stretched exponential relaxation in molecular and electronic glasses” [*Rep. Prog. Phys.*]{}, vol. 59, 1133–-1207, 1996 F. Alvarez, A. Alegr$\acute{\textrm{\i}}$a and J. Colmenero, “Relationship between the time-domain Kohlrausch-William-Watts and frequency domain Havriliak-Negami relaxation functions” [*Phys. Rev. B*]{}, vol. 44(14), 7306–7312, 1991. S.W. Morgan, V. Oganesyan and G.S. Boutis, “Multispin correlations and pseudothermalization of the transient density matrix in solid-state NMR: Free induction decay and magic echo” [*Phys. Rev. B*]{}, vol. 86, 214410, 2012. H. Vestermark, “http://www.hvks.com/index.html.” , [R. Le Riche]{}, and [F. Guyon]{}, “A constrained, globalized, and bounded Nelder-Mead method for engineering optimization,” [*Struct. Multidisc. Optim.*]{}, vol. 27, pp. 43–54, 2004.

ArXiv

--- abstract: 'We calculate the power spectrum of metric fluctuations in inflationary cosmology starting with initial conditions which are imposed mode by mode when the wavelength equals some critical length $\ell_{_{\rm C}}$ corresponding to a new energy scale $M_{_{\rm C}}$ at which trans-Planckian physics becomes important. In this case, the power spectrum can differ from what is calculated in the usual framework (which amounts to choosing the adiabatic vacuum state). The fractional difference in the results depends on the ratio $\sigma _0$ between the Hubble expansion rate $H_{\rm inf}$ during inflation and the new energy scale $M_{_{\rm C}}$. We show how and why different choices of the initial vacuum state (stemming from different assumptions about trans-Planckian physics) lead to fractional differences which depend on different powers of $\sigma _0$. As we emphasize, the power in general also depends on whether one is calculating the power spectrum of density fluctuations or of gravitational waves.' author: - Jérôme Martin - Robert Brandenberger title: 'On the Dependence of the Spectra of Fluctuations in Inflationary Cosmology on Trans-Planckian Physics' --- Introduction ============ The exponential expansion of space in inflationary cosmology leads to the stretching of modes which were in the trans-Planckian regime at the beginning of inflation into the observable range. This leads to the possibility, first raised explicitly in [@RHBrev], that trans-Planckian physics might be observable today in the cosmic microwave background. In earlier work [@MB1; @BM2] we addressed this issue in a simple toy model obtained by replacing the linear dispersion relation of the cosmological fluctuations by new dispersion relations which differ from the linear one on length scales smaller than the Planck length (the same dispersion relations had been used earlier [@Unruh; @CJ] in the context of an analysis of possible trans-Planckian effects on black hole radiation). We were able to construct dispersion relations which give rise to large (order one) corrections to the usual spectrum of fluctuations, but the price to pay is a fine-tuning of the parameters describing the model and/or a back-reaction problem. This question has been further analyzed in many papers (see for instance Refs. [@Niemeyer; @MB3; @NP2; @LLMU; @BJM; @S]). It was found that in order to obtain measurable differences in the predictions, non-adiabatic evolution of the state on trans-Planckian scales is required. In another line of approach to the [*trans-Planckian challenge*]{} to inflationary cosmology, the possibility of measurable effects of trans-Planckian physics on observables such as CMB anisotropies and power spectra of scalar and tensor metric fluctuations was studied [@kempf; @Chu; @Easther; @kempfN; @Hui; @HoB; @Hassan] in models where the trans-Planckian physics is based on stringy space-time uncertainty relations. In particular, the authors of [@Easther] found a spectrum with oscillations of amplitude $\sigma _0\equiv H_{\rm inf}/ M_{_{\rm C}}$, where $H_{_{\rm inf}}$ is the Hubble parameter during inflation and $M_{_{\rm C}}$ a characteristic scale at which the trans-Planckian physics shows up, superimposed on the usual scale-invariant spectrum, whereas the authors of [@kempfN] found only much smaller effects. The trans-Planckian problem was also tackled in the framework of non-commutative geometry in Ref. [@LMMP]. It was found that the effect is of order $(H_{\rm inf}/ M_{_{\rm C}})^4$. It was also shown in this article that non-commutative geometry implies the presence of a preferred direction which would result in a correlation between different multipoles $C_{\ell }$ and $C_{\ell \pm2}$. In yet another approach to the trans-Planckian issue, Danielsson [@Dan1] (see also Ref. [@Lowe1]) suggested to replace the unknown physics on trans-Planckian scales by assuming that the modes representing cosmological fluctuations are generated mode by mode at the time when the wavelength of the mode equals the Planck length, or more generally when it equals the length $\ell_{_{\rm C}}$ associated with the energy scale $M_{_{\rm C}}$ of the new physics which sets the initial conditions. There is a one-parameter family of vacuum states ($\alpha$ vacua) of a free quantum field in de Sitter space which can be considered, and for nontrivial $\alpha$ vacua Danielsson found effects of the choice of the initial state which are of linear order in the ratio $\sigma _0$[^1], and such effects could be seen in observations [@Dan3] [^2]. Similar results were found by Easther et al. [@Easther3], whereas Niemeyer et al. [@npc] have argued that if the modes are set off in the initial adiabatic vacuum state when their wavelength equals $\ell_{_{\rm C}}$, then the effects are of order $\sigma _0^3$ and hence (in usual models) completely negligible. Using an effective field theory method, Kaloper et al. [@Kaloper1] have argued that the effects of trans-Planckian physics on cosmological fluctuations should be at most of the order $\sigma _0^2$, assuming that the modes are in the adiabatic vacuum state when the wavelength is equal to the Hubble radius (see Ref. [@BM4] for a criticism of imposing initial conditions at Hubble radius crossing, and see Ref. [@Cline] for counterexamples to the claims of Ref. [@Kaloper1]). In this paper, we re-consider the calculation of the spectrum of cosmological perturbation in the [*minimal trans-Planckian*]{} setting [@Dan1] when mode by mode the initial conditions for the mode are set when the wavelength equals the Planck length (or, more generally, the length scale of the new physics). We find that the overall amplitude of the correction terms (compared to the usual spectra) depends sensitively on the precise prescription of the initial state, it depends on whether one is studying power-law or slow-roll inflation, and it also depends on whether one is computing the spectrum of scalar metric fluctuations or of gravitational waves. Some of the “discrepancies” between the results of previous studies is due to the fact that different quantities were calculated in different models. We show that when the initial state is chosen to be the instantaneous Minkowski vacuum, then the deviations of the power spectrum from the usual result are of the order $\sigma _0^3$, in agreement with what was found in [@npc]. In an arbitrary $\alpha-$ vacuum, the choice of the value of $\alpha$ has an effect on the amplitude of the fluctuation spectrum which is not suppressed by any power of $\sigma _0$. However, if $\alpha$ is independent of $k$, the effect will not be distinguishable from a slight change in the underlying parameters of the background inflationary model. However, in general (and specifically in the choice of the vacuum made in [@Dan1], the amplitude of the correction term in the power spectrum will have a k-dependent (and hence observable) piece which is first order in $\sigma _0$, at least in the case of the spectrum of gravitational waves. While this paper was being finalized, three preprints appeared which investigate related aspects of the trans-Planckian problem. In Ref. [@Venez], the choice of various initial states was related to the minimization of different Hamiltonians. In Ref. [@Armen], the predictions of inflation for the power spectrum of fluctuations was studied for a two parameter class of initial states, and the amplitude of the corrections compared to the usual results was seen to depend sensitively on which state is chosen. In Ref. [@Chung], the fact that the definitions of the Bunch-Davies and of the adiabatic vacua have some intrinsic ambiguities in a Universe with a de Sitter phase of finite duration was analyzed. Cosmological perturbations of quantum-mechanical origin ======================================================= General considerations ---------------------- The line element for the spatially flat Friedmann-Lemaitre-Robertson-Walker (FLRW) background plus small perturbations can be written as [@MFB]: $$\begin{aligned} \label{metricgi} {\rm d}s^2 &=& a^2(\eta )\{-(1-2\phi ){\rm d}\eta ^2+2({\rm \partial}_iB){\rm d}x^i {\rm d}\eta \nonumber \\ &+& [(1-2\psi )\delta _{ij} +2{\rm \partial }_i{\rm \partial }_jE+h_{ij}]{\rm d}x^i{\rm d}x^j\}\, , \end{aligned}$$ where the functions $\phi $, $B$, $\psi $ and $E$ represent the scalar sector whereas the traceless and transverse tensor $h_{ij}$ ($h_i{}^i=h_{ij}{}^{,j}=0$), represents the tensor sector, i.e. the gravitational waves. The time $\eta $ is the conformal time and is related to the cosmic time $t$ by the relation ${\rm d}t=a(\eta ){\rm d}\eta $. It is convenient to introduce the background quantity $\gamma (\eta )$ defined by $\gamma \equiv -\dot{H}/H^2$, where a dot means differentiation with respect to cosmic time and $H$ is the Hubble rate, $H\equiv \dot{a}/a$. Using the conformal time we may rewrite $\gamma $ as $\gamma =1-{\cal H}'/{\cal H}^2$, where ${\cal H}\equiv a'/a$, and a prime denotes differentiation with respect to the conformal time. For power-law scale factors, $a(\eta )\propto (-\eta )^q$, the function $\gamma (\eta )$ turns out to be a constant, $\gamma =(q+1)/q$, which vanishes in the particular case of the de Sitter space-time characterized by $q=-1$. The perturbed Einstein equations provide us with the equations of motions for the cosmological perturbations. At the linear level, each type of perturbations decouple and we can treat them separately. In the tensor sector (which is automatically gauge-invariant) we define the quantity $\mu _{_{\rm T}}$ for each mode $k$ according to $$\label{eq2} h_{ij}(\eta ,{\bf x})=\frac{1}{a}\frac{1}{(2\pi )^{3/2}} \sum _{s=1}^2\int {\rm d}{\bf k}p_{ij}^s({\bf k}) \mu _{_{\rm T}}(\eta ,{\bf k}){\rm e}^{i{\bf k}\cdot {\bf x}}\, ,$$ where $p_{ij}^s({\bf k})$ is the polarization tensor. The plane waves appear in the previous expression because the space-like sections are flat. At linear order, gravitational waves do not couple to matter and, as a consequence, the equation of motion is just given by the perturbed vacuum Einstein equation. Explicitly, it reads [@Grigw]: $$\label{eomtensor} \mu _{_{\rm T}}''+\biggl(k^2-\frac{a''}{a}\biggr)\mu _{_{\rm T}}=0\ .$$ This equation is similar to a time-independent Schrödinger equation with an effective potential given by $U_{_{\rm T}}(\eta )=a''/a$. It can also be viewed as the equation of a parametric oscillator whose time-dependent frequency is given by $\omega _{_{\rm T}}=k^2-a''/a$, see Ref. [@MSwkb]. In the vacuum state, the two point correlation function of gravitational waves reads $$\langle 0\vert h_{ij}(\eta ,{\bf x})h^{ij}(\eta ,{\bf x}+{\bf r}) \vert 0 \rangle =\int _0^{+\infty }\frac{{\rm d}k}{k}\frac{\sin kr}{kr} k^3P_h(k)\, .$$ The power spectrum $k^3P_h(k)$ which appears in the expression for the two point correlation function is given by $$k^3P_h(k)=\frac{2k^3}{\pi ^2}\biggl \vert \frac{\mu _{_{\rm T}}}{a} \biggr \vert ^2\, .$$ This quantity is [*a priori*]{} time and wavenumber dependent but for super-horizon modes, it turns out to be time-independent because the growing mode is given by $\mu _{_{\rm T}}\propto a(\eta )$. Let us now turn to scalar metric (density) perturbations. The two most important differences with the gravitational waves are that the scalar sector is gauge-dependent and that the scalar perturbations of the metric are coupled to the perturbations of the stress-energy tensor describing the matter. Scalar perturbations of the geometry can be characterized by the two gauge-invariant Bardeen potentials $\Phi \equiv \phi +(1/a)[(B-E')a]'$ and $\Psi \equiv \psi -{\cal H}(B-E')$ [@Bardeen]. From now on, we restrict ourselves to the case where the matter is described by a scalar field: $\varphi=\varphi _0(\eta )+\varphi _1(\eta ,{\bf x})$. Fluctuations in the scalar field are characterized by the gauge-invariant quantity ${\rm \delta}\varphi \equiv \varphi _1+\varphi _0'(B-E')$. The full set of the scalar perturbed (gauge-invariant) Einstein equations is $$\begin{aligned} & & -3{\cal H}({\cal H}\Phi +\Psi ')+\partial _k\partial ^k\Psi =\frac{\kappa }{2} \biggl[-(\varphi _0')^2\Phi +\varphi _0'\delta \varphi ' +a^2\frac{{\rm d}V}{{\rm d}\varphi _0}\delta \varphi \biggr], \quad \partial _i({\cal H}\Phi +\Psi ' )=\frac{\kappa }{2}\varphi _0'\partial _i\delta \varphi \, , \\ & & \partial _i\partial ^j(\Phi -\Psi )=0\, , \quad (2{\cal H}'+{\cal H}^2)\Phi +{\cal H}\Phi '+\Psi ''+2{\cal H}\Psi ' -\frac{1}{3}\partial _k\partial ^k(\Phi -\Psi ) =\frac{\kappa }{2} \biggl[-(\varphi _0')^2\Phi +\varphi _0'\delta \varphi ' -a^2\frac{{\rm d}V}{{\rm d}\varphi _0}\delta \varphi \biggr]\, , \end{aligned}$$ where $\kappa =8\pi/m_{_{\rm Pl}}^2$, $m_{_{\rm Pl}}$ being the Planck mass. At this point, one has to be careful because the case $\varphi _0'=0$, which corresponds to the de Sitter space-time, is particular. In this situation, the solution of the above system of equations is simply $\Phi =0$, i.e. there are no density perturbations at all (it is also necessary to require that the Bardeen potential is finite at infinity). If $\varphi _0'\neq 0$, then everything can be reduced to the study of a single gauge-invariant variable (the so-called Mukhanov-Sasaki variable) defined by [@Mukh] $$v\equiv a\biggl(\delta \varphi+ \frac{\varphi _0'}{{\cal H}}\Phi \biggr)\, .$$ It turns out to be more convenient to work with the variable $\mu _{_{\rm S}}$ defined by $\mu _{_{\rm S}}\equiv -\sqrt{2\kappa }v$. Its equation of motion is very similar to that of the gravitational waves and reads [@MS]: $$\label{eomscalar} \mu _{_{\rm S}}''+\biggl[k^2- \frac{(a\sqrt{\gamma})''}{(a\sqrt{\gamma })}\biggr]\mu _{_{\rm S}}=0\ .$$ Some remarks are in order here. Firstly, let us stress again that the above equation can be used only if $\varphi _0'\neq 0$ or $\gamma \neq 0$ because in course of its derivation, we have divided by $\varphi _0'$. Secondly, the effective potential for density perturbations, $U_{_{\rm S}}=(a\sqrt{\gamma })''/(a\sqrt{\gamma })$ is different from the potential for gravitational waves. Equivalently, its time-dependent frequency is also different and given by $\omega _{_{\rm S}}=k^2 -(a\sqrt{\gamma })''/(a\sqrt{\gamma })$. Thirdly, in the case of power-law inflation, the function $\gamma $ is a constant and we have $U_{_{\rm S}}=U_{_{\rm T}}$. Fourthly, density perturbations are often characterized by the so-called conserved quantity $\zeta $ defined by $\zeta \equiv (2/3)({\cal H}^{-1}\Phi '+\Phi )/(1+\omega )+\Phi $, where $\omega $ is the equation of state parameter, i.e. the ratio of pressure to energy density of the background scalar field. This quantity is directly linked to $\mu _{_{\rm S}}$ through the equation $\mu _{_{\rm S}}=-2a\sqrt{\gamma }\zeta $. The integration of Eq. (\[eomscalar\]) leads to the primordial spectrum of this quantity, namely $$\label{specdp} k^3P_{\zeta }(k)=\frac{k^3}{8\pi ^2}\biggl\vert \frac{\mu _{_{\rm S}}}{a\sqrt{\gamma }}\biggr\vert ^2 \, ,$$ The fact that this quantity is meaningless for the de Sitter case, $\gamma =0$, is obvious. Let us now consider a different situation: there are no cosmological fluctuations anymore but just a test scalar field $\chi (\eta ,{\bf x})$ living in a FLRW universe. If we Fourier expand the scalar field according to $$\chi (\eta ,{\bf x})=\frac{1}{a(\eta )}\frac{1}{(2\pi )^{3/2}} \int {\rm d}{\bf k}\mu (\eta, {\bf k}){\rm e}^{i{\bf k}\cdot {\bf x}}\, ,$$ then the Klein-Gordon equation takes the form $$\label{motsf} \mu ''+\biggl(k^2-\frac{a''}{a}\biggr)\mu =0 \, .$$ We recognize the equation of motion of gravitational waves. The two-point correlation function of the scalar field can be written as $$\langle 0\vert \chi (\eta ,{\bf x})\chi (\eta ,{\bf x}+{\bf r}) \vert 0 \rangle =\int _0^{+\infty }\frac{{\rm d}k}{k}\frac{\sin kr}{kr} k^3P_{\chi }(k)\, .$$ The quantity $k^3P_{\chi }(k)$ which appears in the expression for the two point correlation function is the power spectrum of the scalar field and is given by $$\label{specscaf} k^3P_{\chi }(k)=\frac{k^3}{2\pi ^2}\biggl \vert \frac{\mu }{a} \biggr \vert ^2\, .$$ We also see that the equation of motion for density perturbations is equivalent to the equation for a scalar field if the function $\gamma $ is a non-vanishing constant. This is why, for this class of models, the study of density perturbations is in fact equivalent to the study of a scalar field. However, there is an important exception to this prescription: the de Sitter space-time since, in this case, $\gamma =0$ so that the equation of motion is not given by Eq. (\[eomscalar\]) as already mentioned above. Therefore, it is inconsistent to first assume that $\gamma $ is a constant, then to use Eq. (\[specdp\]) in order to calculate the power spectrum and finally to particularize the result to the de Sitter case. This procedure can only lead to the determination of the spectrum of gravitational waves and/or of a test scalar field but not of density perturbations in the de Sitter space-time. Power-law inflation ------------------- The case of power-law inflation, for which the scale factor can be written as $a(\eta )=\ell _0(-\eta )^q$, is important because the equation of motion for the cosmological perturbations can be solved explicitly. The parameter $\ell _0$ is a length since we have chosen to work with a dimension-full scale factor. Since the function $\gamma (\eta )$ reduces to a constant, the effective potential for density perturbations simplifies to $a''/a$ and becomes identical to the gravitational waves effective potential. Its explicit form reads $U_{_{\rm S}}(\eta )=U_{_{\rm T}}(\eta )=q(q-1)/\eta ^2$. Then, the exact solution for the variables $\mu _{_{\rm S}}$, $\mu _{_{\rm T}}$ and $\mu $ reads $$\mu _{_{\rm S,T}}=(k\eta )^{1/2} [A_1(k)J_{q-1/2}(k\eta )+A_2(k)J_{-q+1/2}(k\eta )]\, .$$ In the above expression, $J_{\nu }$ is a Bessel function of order $\nu $. The two $k$-dependent constants $A_1(k)$ and $A_2(k)$ are fixed by the initial conditions. We now consider the standard calculation of inflationary cosmology. In this case, the initial conditions are fixed in the infinite past. When $k\eta \rightarrow -\infty $, the mode function tends to a plane wave with positive and negative frequency. The usual procedure requires that $$\label{stic} \lim _{k/(aH)\rightarrow +\infty }\mu _{_{\rm S,T}}=\mp \frac{4\sqrt{\pi }}{m_{_{\rm Pl}}} \frac{{\rm e}^{-ik\eta }}{\sqrt{2k}}\, .$$ Let us notice that the origin of the minus sign for density perturbations is the minus sign in the relation $\mu _{_{\rm S}}=-\sqrt{2\kappa }v$, $v$ being the quantity which is canonically quantized, i.e. which behaves as ${\rm e}^{-ik\eta }/\sqrt{2k}$ in the ultraviolet limit. Then, the constants $A_1(k)$ and $A_2(k)$ are completely specified and are given by $$\begin{aligned} \frac{A_1(k)}{A_2(k) } &=& -{\rm e}^{i\pi (q-1/2)}\, , \\ A_2(k) &=& \mp \frac{2\pi }{m_{_{\rm Pl}}}\frac{{\rm e}^{-i\pi (q-1)/2}} {\cos (\pi q)}k^{-1/2}\, . \end{aligned}$$ This implies that the mode functions $\mu _{_{\rm S,T}}$ can be expressed as $$\mu _{_{\rm S,T}}^{\rm stand}(k,\eta )=\mp \frac{2i\pi }{m_{_{\rm Pl}}}(-\eta )^{1/2} {\rm e}^{-i\pi q/2}H^{(1)}_{1/2-q}(-k\eta )\, ,$$ where $H^{(1)}$ is the Hankel function of first kind[^3]. This result allows us to calculate the power spectra. For density perturbations and gravitational waves, one respectively finds, on super-Hubble scales $$\begin{aligned} k^3P_{\zeta}(k) &=& \frac{\ell _{_{\rm Pl}}^2}{\ell _0^2} \frac{1}{\pi \gamma }f(q)k^{2q+2}\, , \\ k^3P_{h}(k) &=& \frac{\ell _{_{\rm Pl}}^2}{\ell _0^2} \frac{16}{\pi }f(q)k^{2q+2}\, , \end{aligned}$$ where $\ell _{_{\rm Pl}}=m_{_{\rm Pl}}^{-1}$ is the Planck length and the function $f(q)$ is given by $$f(q)\equiv \frac{1}{\pi }\biggl[\frac{\Gamma (1/2-q)}{2^{q}}\biggr]^2\, .$$ In the above definition of the function $f(q)$, $\Gamma $ denotes Euler’s integral of the second kind. For the de Sitter case, one has $f(q=-1)=1$. However, for this case the amplitude of density perturbations blows up since $\gamma (q=-1)=0$ while the amplitude of the gravitational waves spectrum remains finite. The origin of the singular limit for density perturbations is again the factor $\sqrt{\gamma }$ at the denominator of the expression for the scalar power spectrum. In fact, this case must be analyzed separately and one can show that density perturbations do not exist in de Sitter space-time. We see that power-law inflation leads to $$k^3P_h(k) = A_{_{\rm T}}k^{n_{_{\rm T}}} \,\, , \,\, k^3P_{\zeta }(k) = A_{_{\rm S}}k^{n_{_{\rm S}}-1}$$ with $n_{_{\rm S}}-1=n_{_{\rm T}}$. Explicitly, one has $$n_{_{\rm S}}=2q+3\, .$$ The case $q=-1$ leads to a scale-invariant spectral index, namely $n_{_{\rm S}}=1$ and $n_{_{\rm T}}=0$. The power-spectrum of the scalar field can be also calculated and the result reads (contrary to the previous power spectra, the power spectrum of the scalar field is a dimension-full quantity) $$k^3P_{\chi }(k)=\frac{1}{(2\pi )^2\ell _0^2}f(q)k^{2q+2}\, .$$ It is also convenient to expressed this power spectrum in terms of the Hubble parameter during inflation $H_{\rm inf}={\cal H}/a= -q/[\ell _0(-\eta )^{q+1}]$. This permits to replace the scale $\ell _0$ in the above equation and leads to $$k^3P_{\chi }(k)=\frac{H_{\rm inf}^2(\eta )}{(2\pi )^2} \frac{(-\eta )^{2(q+1)}}{q^2}f(q)k^{2q+2}\, .$$ Of course, in spite of the fact that the time dependence now appears explicitly, the power spectrum remains a time independent quantity. In the de Sitter case, $q=-1$, one recovers the result often cited in the literature, namely $k^3P_{\chi }=[H_{\rm inf}/(2\pi )]^2$. The spectrum becomes scale-invariant and its amplitude remains finite. The situation, except for some unimportant numerical factors, is very similar to that of gravitational waves. To conclude this subsection, let us emphasize that the previous discussion has shown that the analogy between density perturbations and a scalar field must be used cautiously. The spectral indices are similar even in the de Sitter limit but the amplitudes differ radically in this limit. Slow-roll inflation ------------------- The slow-roll method is an approximation scheme which allows us to go beyond the simple power-law solutions considered in the previous section. It permits to treat a more general class of inflaton potentials. At leading order, the approximation is controlled by the slow-roll parameters (see e.g. Ref. [@Lea]; for a new set of slow-roll parameters with a very nice interpretation, see Ref. [@ST]) defined by: $$\begin{aligned} \label{defepsilon} \epsilon &\equiv & 3 \frac{\dot{\varphi_0}^2}{2} \left(\frac{\dot{\varphi _0}^2}{2} + V\right)^{-1} = -\frac{\dot{H}}{H^2} =1-\frac{{\cal H}'}{{\cal H}^2}\, , \nonumber \\ \label{defdelta} \delta &\equiv & -\frac{\ddot{\varphi }}{H\dot{\varphi }} = - \frac{\dot{\epsilon }}{2 H \epsilon }+\epsilon \, ,\quad \label{defxi} \xi \equiv \frac{\dot{\epsilon }-\dot{\delta }}{H}\ . \end{aligned}$$ The quantity $V(\varphi )$ is the inflaton potential. We see that $\gamma =\epsilon$, where $\gamma $ is the function that has been introduced before. The slow-roll conditions are satisfied if $\epsilon$ and $\delta$ are much smaller than one and if $\xi = {\cal O}(\epsilon^2,\delta^2,\epsilon\delta)$. Since the equations of motion for $\epsilon$ and $\delta$ can be written as: $$\label{eqmotionsrpara} \frac{\dot{\epsilon}}{H}=2\epsilon (\epsilon -\delta)\ , \quad \frac{\dot{\delta }}{H}=2\epsilon (\epsilon -\delta )-\xi \, ,$$ it is clear that this amounts to considering $\epsilon$ and $\delta$ as constants if one works at first order in the slow-roll parameters. This property turns out to be crucial for the calculation of the power spectra of cosmological perturbations. For power-law inflation the slow-roll parameters satisfy: $\epsilon = \delta < 1$, $\xi = 0$. The slow-roll approximation can be viewed as a kind of expansion around the de Sitter space-time. Indeed, at leading order, one has $aH \approx -(1+\epsilon)/\eta$ which implies that the scale factor behaves like $a(\eta)\propto (-\eta )^{-1-\epsilon}$. Interestingly enough, the effective power index at leading order depends on $\epsilon$ only. At this point, one should make the following remark. Using the relation $\gamma =(q+1)/q$, the scale factor of power-law inflation can be re-written as $a(\eta )\propto (-\eta )^{-1/(1-\gamma )}$. From this expression, one might be tempted to write that, in the slow-roll framework, $a(\eta )\propto (-\eta )^{-1/(1-\epsilon )}$ since $\gamma =\epsilon $ in this approximation. Of course, this expression is inconsistent because it contains an infinite number of power of $\epsilon $. What should be done is to expand the expression $-1/(1-\epsilon )$ in terms of $\epsilon$ from which we recover that the scale factor is given by $a(\eta)\propto (-\eta )^{-1-\epsilon}$. As long as one decides to keep high order terms, the whole hierarchy of slow-roll parameters should enter the game. This shows that the slow-roll approximation does not only consist in naively expanding the power index of the scale factor in powers of $\epsilon $. This conclusion is reinforced by a study of the cosmological perturbations within this approximation. The effective potential of density perturbations can be calculated exactly in terms of the slow-roll parameters. The result is: $$\label{potsr} U_{_{\rm S}}(\eta) = a^2H^2 [2 - \epsilon + (\epsilon -\delta)(3-\delta )+\xi] \ .$$ We have seen before that, in the slow-roll approximation, $a^2 H^2 \approx \eta^{-2}(1+ 2\epsilon)$. This implies that the effective potential reduces to $$U_{_{\rm S}}(\eta )\approx \frac{1}{\eta ^2}(2+6\epsilon -3\delta )\, .$$ Since, at leading order, $\epsilon$ and $\delta$ must be seen as constants in the slow-roll approximation, the equation of motion is of the same type as in power-law inflation and the solution is expressed in terms of Bessel functions according to: $$\label{regionII} \mu _{_{\rm S}}=(k\eta )^{1/2}[B_1J_{q_{_{\rm S}}-1/2}(k\eta ) +B_2J_{-q_{_{\rm S}}+1/2}(k\eta )].$$ The parameter $q_{_{\rm S}}$ appearing in the order of the Bessel function is given by $$\label{orderS} q_{_{\rm S}}= -1-2\epsilon +\delta \ .$$ A comment is in order here: The potential $U_{_{\rm S}}$ depends on the scale factor and its derivatives only and we have seen before that the scale factor behaves as $a(\eta )\propto (-\eta )^{-1-\epsilon }$. Therefore, one might think that $U_{_{\rm S}}$ should depend on $\epsilon$ only. This is not the case. The reason is that $U_{_{\rm S}}$ contains terms like $\dot{\epsilon }/\epsilon $ (for instance) which are linear in $\delta$. First one must calculate all derivatives, replace them with their expression in terms of $\epsilon$ and $\delta$, and only then consider that the slow-roll parameters are constant. For gravitational waves, the same lines of reasoning can be applied. The effective potential can be written as $U_{_{\rm T}}(\eta ) = a^2 H^2 \left(2 - \epsilon\right)$ and gives in the slow-roll limit $$\label{potgwsr} U_{_{\rm T}}(\eta ) \sim \frac{2 + 3\epsilon}{\eta ^2} \ .$$ Therefore, the solution of $\mu _{_{\rm T}}$ is similar to the one given in Eq. (\[regionII\]), where the effective index of the Bessel function is now given by: $q_{_{\rm T}}=-1-\epsilon $. This time, the spectral index only depends on $\epsilon $ as expected from the shape of the tensor effective potential. Fixing the initial conditions in the infinite past in the standard manner, we arrive at the following expressions for the power spectra $$\begin{aligned} \label{srdp} {\cal P}_{\zeta }(k) &=& \frac{H^2}{\pi\epsilon m_{_{\rm Pl}}^2} \left[1 - 2(C+1)\epsilon - 2C(\epsilon -\delta ) - 2(2\epsilon -\delta )\ln\left(\frac{k}{k_*}\right)\right]\, , \\ \label{srgw} {\cal P}_h(k) &=& \frac{16 H^2}{\pi m_{_{\rm Pl}}^2} \left[1 - 2(C+1)\epsilon - 2\epsilon \ln \left(\frac{k}{k_*}\right)\right]\, , \end{aligned}$$ where $C\equiv \gamma _{_{\rm E}}+\ln 2-2\simeq - 0.7296$, $\gamma _{_{\rm E}}\simeq 0.5772$ being the Euler constant. All quantities are evaluated at Hubble radius crossing. The scale $k_*$ is called the pivot scale, see Refs. [@MS2] and [@MRS]. One sees again that the de Sitter limit $\epsilon \rightarrow 0$ is ill-defined for scalar metric fluctuations, whereas it is well-defined for gravitational waves. Minimal trans-Planckian physics =============================== General description ------------------- We now consider the cosmological perturbations of quantum-mechanical origin in the framework of the minimal trans-Planckian physics. The new ingredient consists in assuming that the Fourier modes never penetrate the trans-Planckian region. Rather, the main idea is that a Fourier mode is “created” when its wavelength becomes equal to a new fundamental characteristic scale $\ell _{_{\rm C}}\equiv (2\pi )/M_{_{\rm C}}$. Then, the evolution proceeds as usual since the equations of motion of the mode functions $\mu _{_{\rm S,T}}$ are taken to be unmodified and still given by Eqs. (\[eomtensor\]) and (\[eomscalar\]). In this way, the changes are entirely encoded in the initial conditions and there is no need to postulate some ad-hoc trans-Planckian physics. The time of “creation” of the mode of comoving wavenumber $k$, $\eta _k$, can be computed from the condition $$\lambda (\eta _k)=\frac{2\pi }{k}a(\eta _k)= \frac{2\pi }{M_{_{\rm C}}}=\ell _{_{\rm C}}\, ,$$ which implies that $\eta _k$ is a function of $k$. This has to be compared with the standard calculations where the initial time is taken to be $\eta _k=-\infty $ for any Fourier mode $k$ (in a certain sense, this initial time does not depend on $k$). The situation is summarized in Fig. \[tpl3\] {width="18cm" height="10cm"} In the framework described above, a crucial question is in which state the Fourier mode is created at the time $\eta _k$. There is now no asymptotic region in the infinite past where the standard prescriptions can be applied. In this article, we consider the most general conditions, namely that the mode is placed in an $\alpha $-vacuum according to $$\begin{aligned} \label{ci1} \mu _{_{\rm S,T}}(\eta _k) &=& \mp \frac{c_k+d_k}{\sqrt{2\omega _{_{\rm S,T}}(\eta _k)}} \frac{4\sqrt{\pi }}{m_{_{\rm Pl}}}\, , \\ \label{ci2} \mu '_{_{\rm S,T}}(\eta _k) &=& \pm i\sqrt{\frac{\omega _{_{\rm S,T}}(\eta _k)}{2}} \frac{4\sqrt{\pi }(c_k-d_k)}{m_{_{\rm Pl}}}\, , \end{aligned}$$ where $\omega _{_{\rm S,T}}\equiv \sqrt{k^2-U_{_{\rm S,T}}}$ is the effective frequency for density perturbations and gravitational waves. The coefficients $c_k$ and $d_k$ are [*a priori*]{} two arbitrary complex numbers satisfying the condition $\vert c_k\vert ^2-\vert d_k\vert ^2=1$. The instantaneous Minkowski state corresponds to $c_k=1$ and $d_k=0$. If, in addition, we take $\eta _k=-\infty $, we recover the standard choice, see Eq. (\[stic\]) since $\omega _{_{\rm S,T}}(\eta _k=-\infty )=k$. We are now in a position where we can compute the mode functions and the corresponding power spectra for density perturbations and gravitational waves. Calculation of the mode function -------------------------------- In this section, we write the scale factor as $a(\eta )\propto (-\eta )^{p}$, where $p$ is a generalized index defined by $$p= \cases{ \displaystyle{q=-\frac{1}{1-\gamma }}\, ,& \mbox{power-law inflation}\, , \cr & \cr -1-\epsilon\, , & \mbox{slow-roll inflation} \, . }$$ Since the initial conditions given by Eqs. (\[ci1\]) and (\[ci2\]) are different from the standard ones, we will obtain different mode functions. The new mode functions that we want to calculate can be expanded according to $$\label{nonstmodef} \mu _{_{\rm S,T}}(\eta )=\alpha _{_{\rm S,T}}(k) \mu _{_{\rm S,T}}^{\rm stand}(\eta ) +\beta _{_{\rm S,T}}(k) [\mu _{_{\rm S,T}}^{\rm stand}(\eta )]^*\, ,$$ where $\mu _{_{\rm S,T}}^{\rm stand}(\eta )$ denotes the mode functions obtained in the standard situation and given by the expression $$\mu _{_{\rm S,T}}^{\rm stand}(k,\eta )=\mp \frac{2i\pi }{m_{_{\rm Pl}}}(-\eta )^{1/2} {\rm e}^{-i\pi \nu/2}H^{(1)}_{1/2-\nu}(-k\eta )\, .$$ In this equation, we have introduced another generalized index $\nu $ defined by $$\nu= \cases{ \displaystyle{q=-\frac{1}{1-\gamma }}\, ,& \mbox{power-law inflation}\, , \cr & \cr q_{_{\rm S}}=-1-2\epsilon+\delta \, , & \mbox{slow-roll inflation} \, , \cr & \cr q_{_{\rm T}}=-1-\epsilon \, , & \mbox{slow-roll inflation} \, . }$$ For power-law inflation the generalized index is the same for density perturbations and for gravitational waves (hence we do not need to distinguish them in the above definition). For slow-roll inflation, as already mentioned above, they differ. The coefficients $\alpha _{_{\rm S,T}}(k)$ and $\beta _{_{\rm S,T}}(k)$ are readily obtained using the initial conditions given in Eqs. (\[ci1\]) and (\[ci2\]) $$\begin{aligned} \label{alpha} \alpha _{_{\rm S,T}}(k) &=& \frac{1}{4}(c_k+d_k) {\rm e}^{i\pi \nu/2} \sqrt{\frac{\pi }{-2\omega _{_{\rm S,T}}(\eta _k)\eta _k}} \biggl\{k\eta_k\biggl[H^{(2)}_{3/2-\nu}-H^{(2)}_{-1/2-\nu} \biggr] +\biggl[1+2i\frac{c_k-d_k}{c_k+d_k}\omega _{_{\rm S,T}}(\eta _k)\eta _k \biggr]H^{(2)}_{1/2-\nu}\biggr\}\, , \\ \label{beta} \beta _{_{\rm S,T}} (k) &=& \frac{1}{4}(c_k+d_k) {\rm e}^{-i\pi \nu/2} \sqrt{\frac{\pi }{-2\omega _{_{\rm S,T}}(\eta _k)\eta _k}} \biggl\{k\eta _k\biggl[H^{(1)}_{3/2-\nu}-H^{(1)}_{-1/2-\nu}\biggr] +\biggl[1+2i\frac{c_k-d_k}{c_k+d_k}\omega _{_{\rm S,T}}(\eta _k)\eta _k \biggr]H^{(1)}_{1/2-\nu}\biggr\}\, , \end{aligned}$$ where the Hankel functions are evaluated at $-k\eta _k$. The knowledge of the coefficients $\alpha _{_{\rm S,T}}(k)$ and $\beta _{_{\rm S,T}}(k)$ is equivalent to the knowledge of the modified mode function, see Eq. (\[nonstmodef\]). The quantity $-k\eta _k$ can be written as $-k\eta _k=-p/\sigma _k$ where $\sigma _k \equiv H(\eta _k)/M_{_{\rm C}}$. The explicit expression of $\sigma _k$ can be easily derived. Writing that $H=p/(a\eta )$, we find that $\sigma _k =\sigma _0(\eta _k/\eta _0)^{-p-1}$ where the time $\eta _0$ is a given, [*a priori*]{} arbitrary, time during inflation which, and this is the important point, does not depend on $k$. The quantity $\sigma _0$ is defined by $\sigma _0\equiv H_0/M_{_{\rm C}}$. Using that $\eta _k=p/(k\sigma _k)$ and $\eta _0=p/(\sigma _0a_0M_{_{\rm C}})$ we finally arrive at $$\sigma _k=\sigma _0\biggl(\frac{k}{a_0M_{_{\rm C}}}\biggr)^{-1-1/p}\, .$$ In the case of slow-roll inflation, the previous expression reduces to $\sigma _k=\sigma _0[k/(a_0M_{_{\rm C}})]^{-\epsilon }$. For the de Sitter case, $p=-1$ and $\epsilon =0$, the Hubble parameter and therefore $\sigma _k$ are constant. This means that $\eta _k\propto 1/k$ as expected. An important remark is that $\sigma _k$ only depends on $p$. This means that, in the slow-roll approximation, $\sigma _k$ only depends on the slow-roll parameter $\epsilon $ and not on the slow-roll parameter $\delta $. This is due to the fact that the calculation of $\sigma _k$ only depends on background quantities. The slow-roll parameter $\delta $ appears in the calculation through the generalized index $\nu $, i.e. at the perturbed level. We now discuss the spectrum obtained according to the initial state chosen. Instantaneous Minkowski vacuum ------------------------------ In this section, we assume that the initial state is such that $c_k=1$ and $d_k=0$. As a warm up, let us derive the spectrum in the particular case of de Sitter, $p=-1$. As mentioned above, this can only been done for gravitational waves since the de Sitter limit is singular for density perturbations. For $p=-1$, we have $\sigma _k=\sigma _0$ and $-k\eta _k=1/\sigma _0$. The spectrum can be determined exactly because the Bessel functions in Eqs. (\[alpha\]) and (\[beta\]) reduce to ordinary functions in the case $p=-1$. The result reads $$\begin{aligned} \label{alTdS} \alpha _{_{\rm T}}(k) &=& \frac{1}{2}{\rm e}^{-i\pi /2-i/\sigma _0} (1-2\sigma _0^2)^{-1/4}\biggl[i+\sigma _0-i\sigma _0^2+(i+\sigma _0) \sqrt{1-2\sigma _0^2}\biggr]\, , \\ \label{beTdS} \beta _{_{\rm T}}(k) &=& \frac{1}{2}{\rm e}^{i\pi /2+i/\sigma _0} (1-2\sigma _0^2)^{-1/4}\biggl[-i+\sigma _0+i\sigma _0^2+(i-\sigma _0) \sqrt{1-2\sigma _0^2}\biggr]\, . \end{aligned}$$ The modulus of the two previous expressions can be easily derived from the above relation $$\vert \alpha _{_{\rm T}}\vert = \frac{1}{2} \biggl(\frac{2-2\sigma _0^2-\sigma _0^4} {\sqrt{1-2\sigma _0^2}}+2\biggl)^{1/2}\, , \quad \vert \beta _{_{\rm T}}\vert = \frac{1}{2} \biggl(\frac{2-2\sigma _0^2-\sigma _0^4} {\sqrt{1-2\sigma _0^2}}-2\biggl)^{1/2}\, ,$$ which coincides with the result of Ref. [@npc]. One can check that $\vert \alpha _{_{\rm T}}\vert ^2-\vert \beta _{_{\rm T}}\vert ^2=1$. In addition, one notices that the result is independent on $k$ as expected for the de Sitter space-time. Then, the power spectrum on super-horizon scales is given by $k^3P_h(k)=(2k^3/\pi ^2) \vert (\alpha _{_{\rm T}}\mu _{_{\rm T}}^{\rm stand}+\beta _{_{\rm T}}\mu _{_{\rm T}}^{\rm stand *} )/a\vert ^2$ where the super-horizon standard mode function can be expressed as $\mu _{_{\rm T}}^{\rm stand} \simeq -4i\sqrt{\pi }(2k)^{-1/2}/(m_{_{\rm Pl}}k\eta )$. The result reads $$\label{minkops} k^3P_h(k)=\frac{16H^2_{\rm inf}}{\pi m_{_{\rm Pl}}^2} \frac{1}{\sqrt{1-2\sigma _0^2}}\biggl[1-\sigma _0^2-\frac{1}{2}\sigma _0^4 +\sigma _0^3\sin \biggl(\frac{2}{\sigma _0}\biggr)+\frac{1}{2}\sigma _0^4 \cos \biggl(\frac{2}{\sigma _0}\biggr)\biggr]\, .$$ This spectrum is represented in Fig. \[tpl1\]. {width="18cm" height="10cm"} So far, no approximation has been made. If we now assume that $\sigma _0$ is a small quantity, we can expand the result in terms of this quantity. Let us also remark that the terms $\sin (2/\sigma _0)$ and $\cos (2/\sigma _0)$ are non-analytic for small values of $\sigma _0$. At leading order, we obtain that $\vert \alpha _{_{\rm T}}\vert \sim 1$ and $\vert \beta _{_{\rm T}}\vert \sim \sigma _0^3/2$ in agreement with Ref. [@npc]. This leads to the following spectrum $$k^3P_h(k)\simeq \frac{16H_{\rm inf}^2}{\pi m_{_{\rm Pl}}^2} \biggl[1+\sigma _0^3\sin \biggl(\frac{2}{\sigma _0}\biggr) \biggr]\, .$$ Therefore, the corresponding effect in the power spectrum is of order $\sigma _0^3$, i.e. a tiny effect if $\sigma _0$ is small. Note, however, that it is not possible to distinguish this effect from a change in the parameter $H_{\rm inf}$ of the inflationary background. We now turn to the case of slow-roll inflation. We start with density perturbations. The first step consists in calculating the coefficients $\alpha _{_{\rm S}}$ and $\beta _{_{\rm S}}$. At first order in the slow-roll parameters and at leading order in the parameter $\sigma _0$ we have $$\begin{aligned} \alpha _{_{\rm S}} &\simeq & {\rm e}^{ik\eta _k}\, , \\ \beta _{_{\rm S}} &\simeq & i\biggl(\frac12 -\frac34 \delta -\frac32 \epsilon \ln \frac{k}{a_0M_{_{\rm C}}} \biggr)\sigma _0^3{\rm e}^{-ik\eta _k}\, . \end{aligned}$$ It is interesting to calculate the modulus of the coefficient $\beta _{_{\rm S}}$. Straightforward calculations lead to $$\label{modbeta} \vert \beta _{_{\rm S}} \vert \simeq \biggl(\frac12 -\frac34 \delta -\frac32 \epsilon \ln \frac{k}{a_0M_{_{\rm C}}} \biggr)\sigma _0^3 \, .$$ We see that this quantity depends on both slow-roll parameters $\epsilon $ and $\delta $, contrary to what was obtained in Ref. [@npc] where the coefficient $\beta _{_{\rm S}}$ was found not to depend on the slow-roll parameter $\delta $. The main reason for this discrepancy is that, in Ref. [@npc], the spectrum of a scalar field on an unperturbed background cosmological model (or, equivalently, the spectrum of gravitational waves) was calculated, and not the spectrum of scalar metric fluctuations, see the discussion after Eq. (\[specscaf\]). If one uses the expression of the scale factor $a(\eta )\propto (-\eta )^{-1-\epsilon}$ and inserts this into the expression for the power spectrum for a scalar field on an unperturbed background, one does not obtain the correct expression for the power spectrum of density perturbations and one misses the terms proportional to the slow-roll parameter $\delta $. One only obtains either the power spectrum of gravitational waves or the power spectrum of density perturbations in the very particular case of power-law inflation for which we have $\epsilon =\delta $ (and for which, in fact, the exact solution is known). This is why putting $\epsilon =\delta $ in Eq. (\[modbeta\]) reproduces the result found in Ref. [@npc]. Then, the power spectrum of the conserved quantity $\zeta $ is given by the following expression \[to be compared with (\[srdp\])\] $$\begin{aligned} \label{eq54} k^3P_{\zeta } &=&\frac{H^2}{\pi \epsilon m_{_{\rm Pl}}^2} \Biggl(1-2(C+1)\epsilon -2C(\epsilon -\delta )-2(2\epsilon -\delta ) \ln \frac{k}{k_*}+\sigma _0^3\biggl\{\sin \biggl[\frac{2}{\sigma _0} \biggl(1+\epsilon +\epsilon \ln \frac{k}{a_0M_{_{\rm C}}}\biggr)\biggr] \nonumber \\ & & -\biggl[\frac32 \delta +3\epsilon \ln \frac{k}{a_0M_{_{\rm C}}} +2(C+1)\epsilon +2C(\epsilon -\delta )+2(2\epsilon -\delta ) \ln \frac{k}{k_*}\biggr]\sin \biggl[\frac{2}{\sigma _0} \biggl(1+\epsilon +\epsilon \ln \frac{k}{a_0M_{_{\rm C}}}\biggr)\biggr] \nonumber \\ & & -\pi (2\epsilon -\delta )\cos \biggl[\frac{2}{\sigma _0} \biggl(1+\epsilon +\epsilon \ln \frac{k}{a_0M_{_{\rm C}}}\biggr) \biggr]\biggr\} \Biggr)\, . \end{aligned}$$ Thus, the power spectrum has oscillations superimposed on the scale-invariant base. Hence, in this case the change in the spectrum is “in principle” physically measurable. However, the prospects for actually detecting the difference of order $\sigma_o^3$ between (\[eq54\]) and (\[srdp\]) are hopeless if the parameter $\sigma _0$ is small. Indeed if, for instance, we assume that $\sigma _0=10^{-2}$, a value consistent with string theory, and if we consider that $\epsilon \simeq \delta \simeq 10^{-1}$ then the correction to the power spectrum is of order $\simeq 10^{-7}$. There is no possibility to detect such a small effect even with the Planck satellite. For completeness, let us mention the slow-roll result for gravitational waves. In this case the coefficients $\alpha _{_{\rm T}}$ and $\beta _{_{\rm T}}$ can be deduced from $\alpha _{_{\rm S}}$ and $\beta _{_{\rm S}}$ by putting $\epsilon =\delta $. Then, the power spectrum is given by $$\begin{aligned} k^3P_h &=&\frac{16H^2}{\pi m_{_{\rm Pl}}^2} \Biggl(1-2(C+1)\epsilon -2\epsilon \ln \frac{k}{k_*} +\sigma _0^3\biggl\{\sin \biggl[\frac{2}{\sigma _0} \biggl(1+\epsilon +\epsilon \ln \frac{k}{a_0M_{_{\rm C}}}\biggr)\biggr] -\biggl[\frac32 \epsilon +3\epsilon \ln \frac{k}{a_0M_{_{\rm C}}} +2(C+1)\epsilon \nonumber \\ & & +2\epsilon \ln \frac{k}{k_*}\biggr]\sin \biggl[\frac{2}{\sigma _0} \biggl(1+\epsilon +\epsilon \ln \frac{k}{a_0M_{_{\rm C}}}\biggr)\biggr] -\pi\epsilon \cos \biggl[\frac{2}{\sigma _0} \biggl(1+\epsilon +\epsilon \ln \frac{k}{a_0M_{_{\rm C}}}\biggr) \biggr]\biggr\} \Biggr)\, . \end{aligned}$$ We can also estimate how the consistency check of inflation is modified. At first order in the slow-roll parameters and at leading order in the parameter $\sigma _0$, this quantity is the same as in the standard case, namely $$R\equiv \frac{k^3P_h}{k^3P_{\zeta }}=16\epsilon \, .$$ This is because at zeroth order in the slow-roll parameter, the leading contribution in $\sigma _0$ is the same for gravitational waves and density perturbations. Arbitrary $\alpha$ vacuum state ------------------------------- We have established that, in the case of the instantaneous Minkowski state, the correction to the power spectra is of order $\sigma _0^3$. It is interesting to see whether this conclusion remains true for other initial states. For this reason, we now repeat the calculation of the Bogoliubov coefficients in the case of a de Sitter space-time with an arbitrary $\alpha $ -vacuum state, characterized by a value of $c_k$ and $d_k$, as the initial state. One finds $$\begin{aligned} \alpha _{_{\rm T}}(k) &=& \frac{1}{2}(c_k+d_k){\rm e}^{-i\pi/2 -i/\sigma _0} (1-2\sigma _0^2)^{-1/4}\biggl[i+\sigma _0-i\sigma _0^2+z_k(i+\sigma _0) \sqrt{1-2\sigma _0^2}\biggr] \, , \\ \beta _{_{\rm T}}(k) &=& \frac{1}{2}(c_k+d_k){\rm e}^{i\pi/2 +i/\sigma _0} (1-2\sigma _0^2)^{-1/4}\biggl[-i+\sigma _0+i\sigma _0^2+z_k(i-\sigma _0) \sqrt{1-2\sigma _0^2}\biggr] \, , \end{aligned}$$ where we have introduced the quantity $z_k$ defined by $z_k\equiv (c_k-d_k)/(c_k+d_k)$. On can easily check that for $c_k=1$ and $d_k=0$, i.e. $z_k=1$, the above expressions reduce to Eqs. (\[alTdS\]) and (\[beTdS\]). One can also calculate the modulus of these Bogoliubov coefficients, $$\begin{aligned} \vert \alpha _{_{\rm T}}\vert ^2 &=& \frac{\vert c_k+d_k\vert ^2}{4} \biggl\{\frac{(1+z_kz_k^*)-\sigma _0^2(1+z_kz_k^*) +\sigma _0^4(1-2z_kz_k^*)}{\sqrt{1-2\sigma _0^2}} +[z_k^*+z-i(z_k^*-z_k)\sigma _0^3]\biggr\}\, , \\ \vert \beta _{_{\rm T}}\vert ^2 &=& \frac{\vert c_k+d_k\vert ^2}{4} \biggl\{\frac{(1+z_kz_k^*)-\sigma _0^2(1+z_kz_k^*) +\sigma _0^4(1-2z_kz_k^*)}{\sqrt{1-2\sigma _0^2}} -[z_k^*+z_k+i(z_k^*-z_k)\sigma _0^3]\biggr\}\, , \end{aligned}$$ and verify that $\vert \alpha _{_{\rm T}} \vert ^2 - \vert \beta _{_{\rm T}} \vert ^2=\vert c_k\vert ^2-\vert d_k\vert ^2=1$. From the two previous relations, one can determine the power spectrum exactly. The result reads $$\begin{aligned} k^3P_h &=& \frac{16H^2_{\rm inf}}{\pi m_{_{\rm Pl}}^2} \frac{\vert c_k+d_k\vert ^2}{\sqrt{1-2\sigma _0^2}} \biggl\{\frac12 (1+z_kz_k^*)(1-\sigma _0^2) +\frac{\sigma _0^3}{2}(1-2z_kz_k^*)\biggl[\sigma _0 -\sigma _0\cos \biggl(\frac{2}{\sigma _0}\biggr) -2\sin \biggl(\frac{2}{\sigma _0}\biggr)\biggr] \nonumber \\ & & +\frac{1}{2}(1-z_kz_k^*)(3\sigma _0^2-1) \cos \biggl(\frac{2}{\sigma _0}\biggr) +(1-z_kz_k^*)\sigma _0\sin \biggl(\frac{2}{\sigma _0}\biggr) -\frac{i}{2}(z_k^*-z_k)(1-2\sigma _0^2)^{3/2} \sin \biggl(\frac{2}{\sigma _0}\biggr) \nonumber \\ & & -\frac{i}{2}\sigma _0^3(z_k^*-z_k)\sqrt{1-2\sigma _0^2} +\frac{i}{2}\sigma _0(z_k^*-z_k)(\sigma _0^2-2) \sqrt{1-2\sigma _0^2} \cos \biggl(\frac{2}{\sigma _0}\biggr)\biggr\}\, . \end{aligned}$$ As is apparent from the above expression, the correction in the power spectrum compared to the usual results is of order unity, i.e. not suppressed by any power of $\sigma_0$. If one argues that the Bogoliubov coefficients themselves should be of linear order in $\sigma_0$ then the correction terms would also be linear in $\sigma_0$. On the other hand, unless the Bogoliubov coefficients depend on $k$, the effect is simply a change in the amplitude of the spectrum, and can hence be absorbed in a redefinition of the background. Thus, the effect is not physically measurable. On the other hand, if the coefficients depend on $k$ (as discussed e.g. in the analysis of [@Lowe1]) there is a large measurable effect on the power spectrum. For the instantaneous Minkowski state, one check easily that only the two first terms in the curly brackets survives and that the corresponding expression reduces to the one found previously. If we expand this spectrum in powers of $\sigma_o$, we find that $$k^3P_h = \frac{16H^2_{\rm inf}}{\pi m_{_{\rm Pl}}^2} \vert c_k+d_k\vert ^2\biggl\{ \biggl \vert z_k\cos \biggl(\frac{2}{\sigma _0}\biggr) -i\sin \biggl(\frac{2}{\sigma _0}\biggr)\biggr\vert ^2 +\biggl[(1-z_kz_k^*)\sin \biggl(\frac{2}{\sigma _0}\biggr) -i(z_k^*-z_k)\cos \biggl(\frac{2}{\sigma _0}\biggr)\biggr]\sigma _0 +\cdots \biggr\}\, .$$ We conclude that the result obtained in the previous subsection for the instantaneous Minkowske vacuum, namely that the corrections to the spectrum are suppressed by three powers of $\sigma_0$ appears to be very particular to the choice of that state. However, the previous analysis does not cover all the possible cases. Indeed, as we are going to demonstrate, the proposal put forward in Ref. [@Dan1] corresponds in fact to a case where the complex number $z_k$ can also depend on $\sigma _0$. The previous study assumed that $c_k$ and $d_k$ were pure complex numbers. Therefore, in Danielsson’s case [@Dan1], the analysis needs to be redone. Danielsson’s $\alpha $-vacuum state ----------------------------------- We start our analysis with the simplest case, namely gravitational waves in a de Sitter background. We treat this case in some detail since then these calculations can be used to study more complicated situations, for instance slow-roll inflation. The Einstein-Hilbert action $$S_{_{\rm E-H}}=(16\pi G)^{-1}\int R\sqrt{-g}{\rm d}^4x$$ expanded to second order (since we are dealing with first order equations of motion) reads $$\begin{aligned} S_2 &=& -\frac{1}{16\pi G}\int \frac{1}{4}g^{\mu \nu} \partial _{\mu }(h^i{}_j)\partial _{\nu }(h^j{}_i) \sqrt{-g}\, {\rm d}^4x \\ &=&\frac{m_{_{\rm Pl}}^2}{64 \pi}\int \biggl[(h^i{}_j)'(h^j{}_i)' -\partial _k(h^i{}_j) \partial ^k(h^j{}_i)\biggr]a^2(\eta ){\rm d}^4x\, . \nonumber \end{aligned}$$ In the first of the two previous expressions, $g^{\mu \nu }$ is the inverse of the FLRW metric. The next step consists in inserting the expression for $h_{ij}$, see Eq. (\[eq2\]), into the action. This gives $$S_2=-\frac{m_{_{\rm Pl}}^2}{16\pi}\sum _{s=1}^2 \int {\rm d}^4x\biggl(\frac{1}{2} g^{\mu \nu}\partial _{\mu }h^s\partial _{\nu }h^s\biggr)\, ,$$ where the fields $h^s(\eta ,x^k)$, with $s = 1,2$, are defined by the following expression $$\begin{aligned} h^s(\eta ,x^k)\equiv \frac{1}{a(\eta )} \frac{1}{(2\pi )^{3/2}}\int {\rm d}{\bf k}\, \mu ^s_{_{\rm T}}(\eta ){\rm e}^{i{\bf k}\cdot {\bf x}}\, . \\ \nonumber \end{aligned}$$ Therefore, the action of gravitational waves is equivalent to the action of two decoupled scalar fields. Inserting the Fourier decomposition of the two scalar fields into the expression for the action, one arrives at $$\begin{aligned} S_2 &=&\frac{m_{_{\rm Pl}}^2}{16\pi}\int {\rm d}\eta \sum _{s=1}^2 \int _{\setR^{3+}}{\rm d}^3{\bf k}\biggl\{ (\mu ^s_{_{\rm T}})'{}^*(\mu ^s_{_{\rm T}})' -\frac{a'}{a}\biggl[ (\mu ^s_{_{\rm T}})'(\mu ^s_{_{\rm T}})^* +(\mu ^s_{_{\rm T}})'{}^*\mu ^s_{_{\rm T}}\biggr] +\biggl(\frac{a'{}^2}{a^2}-k^2\biggr) \mu ^s_{_{\rm T}}(\mu ^s_{_{\rm T}})^*\biggr\}\, . \end{aligned}$$ We can check that this Lagrangian leads to the correct equation of motion: we have (we did not take into account the overall constant; here the bar over ${\cal L}$ means that we are considering the Lagrangian in the Fourier space) $$\begin{aligned} \frac{\delta \bar{\cal L}}{\delta \mu _{_{\rm T}}^s} &=& -2{\cal H} (\mu ^s_{_{\rm T}})'{}^*+2({\cal H}^2-k^2) (\mu ^s_{_{\rm T}})'{}^* \, , \\ \frac{\delta \bar{\cal L}}{(\delta \mu _{_{\rm T}}^s)'} &=& 2(\mu ^s_{_{\rm T}})'{}^*-2{\cal H} (\mu ^s_{_{\rm T}})^*. \end{aligned}$$ Therefore, the Euler-Lagrange equations given by the well-known expression $$\frac{{\rm d}}{{\rm d}\eta} \biggl[\frac{\delta \bar{\cal L}}{(\delta \mu _{_{\rm T}}^s)'}\biggr] -\frac{\delta \bar{\cal L}}{\delta \mu _{_{\rm T}}^s}=0\, ,$$ reproduce the correct equations of motion for the variables $(\mu ^s_{_{\rm T}})^*$. Let us notice that we can also vary the Lagrangian with respect to $(\mu ^s_{_{\rm T}})^*$ to obtain the equation of motion for $\mu ^s_{_{\rm T}}$. From the Lagrangian formalism that we have just described, we can now pass to the Hamiltonian formalism. The conjugate momentum to $\mu ^s_{_{\rm T}}$ is defined by the formula $$p^s_{_{\rm T}}=\frac{\delta \bar{{\cal L}}}{\delta (\mu ^s_{_{\rm T}})'{}^*} =\frac{m_{_{\rm Pl}}^2}{16\pi}\biggl[(\mu _{_{\rm T}}^s)'-\frac{a'}{a} \mu _{_{\rm T}}^s\biggl]\, .$$ In real space, the conjugate momentum $\Pi ^s$ is defined by the expression $$\Pi ^s(\eta ,x^k)=\frac{\delta {\cal L}}{\delta (h^s)'}=\frac{m_{_{\rm Pl}}^2}{16\pi} a^2(h^s)'\, .$$ We can check that the two definitions are consistent by means of the relation $$\Pi ^s(\eta ,x^k)=\frac{a(\eta )}{(2\pi )^{3/2}}\int {\rm d}{\bf k}p^s_{_{\rm T}}{\rm e}^{i{\bf k}\cdot {\bf x}}\, .$$ Danielsson’s boundary condition consists in demanding that, at the time of creation $\eta _k$, one has the usual relation characteristic of the standard vacuum state, namely $$\Pi ^s(\eta _k,x^k)=-ik\frac{m_{_{\rm Pl}}^2}{16\pi } a^2(\eta _k)h^s(\eta _k,x^k)\, .$$ Using the Fourier decomposition of the scalar fields $h^s$ and of their conjugate momenta, this last equation boils down to $$(\mu _{_{\rm T}}^s)'-\frac{a'}{a}\mu _{_{\rm T}}^s= -ik\mu _{_{\rm T}}^s\, .$$ This relation is to be satisfied at the time $\eta =\eta _k$, and implies a link between the coefficients $\alpha _{_{\rm T}}$ and $\beta _{_{\rm T}}$. Since, in addition, $\vert \alpha _{_{\rm T}}\vert ^2-\vert \beta _{_{\rm T}}\vert ^2=1$, the coefficients are in fact completely fixed and they now depend on the parameter $\sigma _0$. Explicitly, the link can be expressed as $$\label{ratioab} \frac{\beta _{_{\rm T}}}{\alpha _{_{\rm T}}}=- \displaystyle{\frac{(\mu _{_{\rm T}}^{\rm stand})' -(a'/a)\mu _{_{\rm T}}^{\rm stand}+ik \mu _{_{\rm T}}^{\rm stand}} {(\mu _{_{\rm T}}^{\rm stand*})' -(a'/a)\mu _{_{\rm T}}^{\rm stand*}+ik \mu _{_{\rm T}}^{\rm stand*}}}\, .$$ So far no approximation has been made and the previous relation is general. We now restrict our study to the de Sitter case. Using the explicit form of the mode function in this case, one arrives at $$\begin{aligned} \frac{\beta _{_{\rm T}}}{\alpha _{_{\rm T}}}= \frac{i}{i+2k\eta _k}{\rm e}^{-2ik\eta _k}\, , \\ \nonumber\end{aligned}$$ which is exactly the relation found in Ref. [@Dan1]. From the normalization, we deduce that $\vert \alpha _{_{\rm T}}\vert ^2=1+\sigma _0^2/4$. The exact power spectrum can now be determined since we explicitly know the coefficients $\alpha _{_{\rm T}}$ and $\beta _{_{\rm T}}$. Performing the standard calculation, one finds $$\label{dpsgw} k^3P_h =\frac{16H^2_{\rm inf}}{\pi m_{_{\rm Pl}}^2} \biggl[1+\frac{\sigma _0^2}{2}-\sigma _0 \sin \biggl(\frac{2}{\sigma _0}\biggr) -\frac{\sigma _0^2}{2} \cos \biggl(\frac{2}{\sigma _0}\biggr)\biggr]\, .$$ This expression should be compared with the corresponding equation found in the case of the Minkowski instantaneous state, see Eq. (\[minkops\]). In particular, expanding everything in terms of $\sigma _0$, one obtains $$k^3P_h=\frac{16H^2_{\rm inf}}{\pi m_{_{\rm Pl}}^2} \biggl[1-\sigma _0 \sin \biggl(\frac{2}{\sigma _0}\biggr)\biggr]\, ,$$ i.e., a first order effect instead of a third order effect. However, once again note that this effect can be reproduced by redefining the background cosmological parameters, and hence it is not a physically measurable effect. The spectrum is represented in Fig. \[tpl2\]. {width="18cm" height="10cm"} Let us now turn to the case of slow-roll inflation. The goal is simply to evaluate the ratio $\beta _{_{\rm T}}/\alpha _{_{\rm T}}$ given by Eq. (\[ratioab\]) to leading order in the parameter $\sigma _0$ and to first order in the slow-roll parameter $\epsilon$. One finds $$\beta _{_{\rm T}}=-\frac{i\sigma _0}{2}{\rm e}^{-2ik\eta _k} \biggl(1-\epsilon \ln \frac{k}{a_0M_{_{\rm C}}}\biggr) \alpha _{_{\rm T}}\, .$$ Of course, for $\epsilon =0$, one recover the exact result obtained previously at leading order in $\sigma _0$. Finally the power spectrum for gravitational waves reads $$\begin{aligned} \label{eq78} k^3P_h &=& \frac{16 H_{\rm inf}^2}{\pi m_{_{\rm Pl}}^2} \biggl\{1-2(C+1)\epsilon -2\epsilon \ln \frac{k}{k_*} -\sigma _0\biggl[1-2(C+1)\epsilon -2\epsilon \ln \frac{k}{k_*} -\epsilon \ln \frac{k}{a_0M_{_{\rm C}}}\biggr] \nonumber \\ & & \times \sin \biggl[\frac{2}{\sigma _0} \biggl(1+\epsilon +\epsilon \ln \frac{k}{a_0M_{_{\rm C}}}\biggr)\biggr] +\sigma _0\pi \epsilon \cos \biggl[\frac{2}{\sigma _0} \biggl(1+\epsilon +\epsilon \ln \frac{k}{a_0M_{_{\rm C}}}\biggr)\biggr]\biggr\}\, . \end{aligned}$$ Note that since in this case the correction terms lead to oscillations in the spectrum about a scale-invariant base spectrum, the effect of the correction terms is in principle measurable. Since the amplitude is only suppressed by one power of $\sigma _0$, the prospects of being able to detect such effects in upcoming experiments are good. We now explore the same mechanism but for scalar metric (density) perturbations. The action is expressed in terms of the variable $\mu _{_{\rm S}}$ introduced before. It reads $$S_2=\frac12 \int {\rm d}^4{\bf x}\biggl[(\mu _{_{\rm S}}')^2 -\delta ^{ij}\partial _i\mu _{_{\rm S}}\partial _j\mu _{_{\rm S}} +\frac{(a\sqrt{\gamma })''}{a\sqrt{\gamma }}\mu _{_{\rm S}}^2\biggr]\, ,$$ In Fourier space, the action reads $$S_2=\frac12 \int {\rm d}\eta \int _{\setR^+}{\rm d}^3{\bf k}\biggl\{ \mu _{_{{\rm S}}}'\mu _{_{{\rm S}}}^*{}'-\biggl[k^2- \frac{(a\sqrt{\gamma })''}{a\sqrt{\gamma }}\biggr] \mu _{_{{\rm S}}}\mu _{_{{\rm S}}}^*\biggr\}\, .$$ The conjugate momentum to the variable $\mu _{_{\rm S}}$ is given by $\Pi _{\mu _{_{\rm S}}}=\mu _{_{\rm S}}'$ and therefore Danielsson’s boundary condition reads $$\label{condscalar} \mu _{_{{\rm S}}}'=-ik\mu _{_{{\rm S}}}\, ,$$ where this relation must be evaluated at the time $\eta =\eta _k$. We note that the term $(a'/a)\mu _{_{\rm S}}$, which was present in the case of the scalar field, does not appear for density perturbations. This can be traced back to the fact that the two quantities that are quantized, $\mu _{_{\rm T}}$ and $\mu _{_{\rm S}}$ do not have the same action. This means that the link between $\alpha _{_{\rm S}}$ and $\beta _{_{\rm S}}$ for density perturbations is not the same as for gravitational waves. For scalar metric fluctuations, it reads $$\label{bascalar} \frac{\beta _{_{\rm S}}}{\alpha _{_{\rm S}}}=- \displaystyle{\frac{(\mu _{_{\rm S}}^{\rm stand})' +ik \mu _{_{\rm S}}^{\rm stand}} {(\mu _{_{\rm S}}^{\rm stand*})' +ik \mu _{_{\rm S}}^{\rm stand*}}}\, .$$ This equation should be compared with Eq. (\[ratioab\]). The absence of the terms $a'/a$ has important consequences. In order to guess what the difference is, we can apply the previous equation to the de Sitter case, even if in principle this is not allowed (see the discussion above). One finds $$\frac{\beta _k}{\alpha _k}=\frac{i}{i+2k\eta _k-2ik^2\eta ^2} {\rm e}^{-2ik\eta _k}\, .$$ Expanding the previous expression in $\sigma _0$, we find that the first order terms cancel out and we are left with $\beta _k/\alpha _k \simeq -(\sigma _0^2/2){\rm e}^{2i/\sigma _0}$. Therefore, the result will be of order $\sigma _0^2$ and not of order $\sigma _0$ as it was the case for a scalar field and gravitational waves. We would have obtained a linear correction in the power spectrum if Danielsson’s condition had been $$\biggl(\frac{\mu _{_{\rm S}}}{a}\biggr)'=-i\frac{k}{a}\mu _{_{\rm S}}\, ,$$ as it is for gravitational waves, instead of $\mu _{_{\rm S}}'=-ik\mu _{_{\rm S}}$, see Eq. (\[condscalar\]). Let us now evaluate Eq. (\[bascalar\]) consistently and rigorously in the slow-roll approximation. One finds $$\beta _{_{\rm S}}=-\frac{\sigma _0^2}{2}{\rm e}^{-2ik\eta _k} \biggl(1+\epsilon -\frac32\delta -2\epsilon \ln \frac{k}{a_0M_{_{\rm C}}} \biggr)\alpha _{_{\rm S}}\, .$$ As expected, the result is quadratic in $\sigma _0$. Repeating the standard calculations, one can find the explicit expression for the power spectrum $$\begin{aligned} \label{pssrs2} k^3P_{\zeta } &=&\frac{H^2}{\pi \epsilon m_{_{\rm Pl}}^2} \biggl\{1-2(C+1)\epsilon -2C(\epsilon -\delta )-2(2\epsilon -\delta ) \ln \frac{k}{k_*}+\sigma _0^2\biggl[ 1-2(C+1)\epsilon -2C(\epsilon -\delta ) +\epsilon -\frac32\delta \nonumber \\ & & -2(2\epsilon -\delta )\ln \frac{k}{k_*} -2\epsilon \ln \frac{k}{a_0M_{_{\rm C}}} \biggr]\cos \biggl[\frac{2}{\sigma _0} \biggl(1+\epsilon +\epsilon \ln \frac{k}{a_0M_{_{\rm C}}}\biggr)\biggr] +\sigma _0^2\pi (2\epsilon -\delta ) \sin \biggl[\frac{2}{\sigma _0} \biggl(1+\epsilon +\epsilon \ln \frac{k}{a_0M_{_{\rm C}}}\biggr)\biggr] \biggr\} \, . \nonumber \\ \end{aligned}$$ As in the case of gravitational waves, in principle this is a measurable effect since it corresponds to oscillations of the power spectrum. Typical power spectra are represented in Fig. \[tpl4\]. Since the correction is quadratic in $\sigma _0$, the correction turns out to be extremely small and the prospects for the detection of such an effect are not optimistic even with a high-accuracy experiment like Planck. {width="18cm" height="10cm"} Thus, at least with this choice of the precise form of the action [@Dprivate], Danielsson’s prescription leads to trans-Planckian corrections of different strengths for gravitational waves and for scalar fluctuations. Note that this also implies that the consistency relation of inflation will be modified. Comparing (\[eq78\]) and (\[pssrs2\]), at leading order in $\sigma _0$, one finds $$R=16\epsilon \biggl\{1-\sigma _0\sin \biggl[\frac{2}{\sigma _0} \biggl(1+\epsilon +\epsilon \ln \frac{k_*}{a_0M_{_{\rm C}}} \biggr)\biggr]\biggr\}\, ,$$ where the ratio has been evaluated at the pivot scale. As expected from the previous considerations, the correction is linear in $\sigma _0$. Comparison with analyses using modified dispersion relations ============================================================ In this section, we consider a case where the trans-Planckian physics is described by mean of a modified dispersion relation. Although, this phenomenological description is different from the one considered above, it has already been shown that the corresponding spectrum can also possess superimposed oscillations [@MB1; @BM2; @LLMU]. In this section, we calculate the spectrum of a scalar field (or gravitational waves) living in a de Sitter space-time and we compare this result with the corresponding result obtained in the framework described in the previous sections, see Eqs. (\[minkops\]) and (\[dpsgw\]). In particular, our goal is to determine which parameter controls the magnitude of the correction to the standard power spectrum in the case of a modified dispersion relation (for the case treated before, it was the parameter $\sigma _0$). The main shortcomings of describing the trans-Planckian physics by a modified dispersion relation is that we need to assume something about the physics beyond the Planck, contrary to the kind of modification envisaged above. Modifications in the power spectrum are obtained if the WKB approximation is violated in the trans-Planckian regime. A typical example where this happens is for the dispersion relation $\omega _{_{\rm phys}}^2=k_{_{\rm phys}}^2 - 2b_{11}k_{_{\rm phys}}^4 + 2b_{12}k_{_{\rm phys}}^6$, which represents the first terms of a systematic Taylor expansion. This dispersion relation is represented in Fig. \[disp\]. {width="18cm" height="10cm"} The equation of motion for the mode function of a scalar field is now given by $$\mu ''+\biggl[\omega ^2(k,\eta )-\frac{a''}{a}\biggr]\mu=0\, ,$$ where $\omega $ is the comoving frequency given by $\omega (k,\eta )=a(\eta ) \omega _{_{\rm phys}}[k/a(\eta )]$. The difference compared to the previous section is that the equation determining the behavior of the mode function is modified. The mode functions depend on whether $w^2>a''/a$ or $w^2<a''/a$. Let us consider an expanding cosmological background. There will be four time intervals (“regions”), see Fig. \[piece\]. As the Universe expands, the physical wavenumber will decrease. The evolution starts in Region I, where the WKB approximation is valid, the term $w^2$ dominates and the mode function is given by $$\mu _{_{\rm I}}(\eta )=\frac{1}{\sqrt{2\omega (k,\eta )}}{\rm e}^{-i\int _{\eta _{\rm i}}^{\eta } \omega (k,\tau ){\rm d}\tau }\, .$$ We have chosen the WKB vacuum, i.e. we have retained only one branch of the most general solution. At a time $\eta_1(k)$, the mode enters Region II, where the WKB approximation is violated, and the mode function can be expressed as $$\mu _{_{\rm II}}(\eta )=C_+(k)a(\eta )+C_-(k)a(\eta ) \int _{\eta _{\rm i}'} ^{\eta }\frac{{\rm d}\tau }{a^2(\tau )}\, .$$ The coefficients $C_{\pm}(k)$ can be determined from joining the mode function and its derivative at the matching point $\eta _1(k)$. After more expansion \[at the time $\eta_2(k)$\], the mode leaves Region II and enters Region III, where the WKB approximation is restored. However, now the mode function has two branches, the second branch having been “generated” in Region II. It is here that the trans-Planckian physics modifies the standard result (which predicts $\beta _k=0$). The mode function can be written as $$\begin{aligned} \mu _{_{\rm III}}(\eta )&=&\frac{\alpha _k}{\sqrt{2\omega (k,\eta )}} {\rm e}^{-i\int _{\eta _{\rm i}''}^{\eta } \omega (k,\tau ){\rm d}\tau } \nonumber \\ & & + \frac{\beta _k}{\sqrt{2\omega (k,\eta )}} {\rm e}^{i\int _{\eta _{\rm i}''}^{\eta } \omega (k,\tau ){\rm d}\tau }\, . \end{aligned}$$ The coefficients $\alpha _k$ and $\beta _k$ can be found in terms of $C_{\pm }(k)$ by matching the solution (and its derivative) at $\eta _2(k)$ as previously. Finally, in Region IV, the WKB approximation is again violated. This region corresponds to the usual super-Hubble region. We have $$\mu _{_{\rm IV}}(\eta )=D_+(k)a(\eta )+D_-(k)a(\eta )\int _{\eta _{\rm i}'''} ^{\eta }\frac{{\rm d}\tau }{a^2(\tau )}\, .$$ The transition between Regions III and IV occurs at the time $\eta_3(k)$, the time of Hubble radius crossing. Our aim is to calculate the coefficients $D_+(k)$ which determines the spectrum of the growing mode. Performing the matching at $\eta _3(k)$, we obtain $$\label{mu4} \mu _{_{\rm IV}}(\eta )=\frac{1}{\sqrt{2\omega _3}}\biggl(\alpha _k {\rm e}^{-i\Omega _3}+\beta _k {\rm e}^{i\Omega _3}\biggr) \frac{a(\eta )}{a_3} +\frac{1}{\sqrt{2\omega _3}}a(\eta )a_3 \biggl(\alpha _k\gamma _3 {\rm e}^{-i\Omega _3}+\beta _k\gamma _3^*{\rm e}^{i\Omega _3}\biggr) \int _{\eta _3}^{\eta }\frac{{\rm d}\tau }{a^2(\tau )}\, ,$$ where the subscript “$3$” means that the corresponding quantity is evaluated at $\eta _3(k)$. In the previous equation, we have introduced the notation $\gamma _k\equiv \omega '/(2\omega )+i\omega +{\cal H}$ and $\Omega _3\equiv \int _{\eta _{\rm i}''}^{\eta _3}\omega (k,\tau ){\rm d}\tau $. On the other hand, a lengthy but straightforward calculation shows that the coefficients $\alpha _k$ and $\beta _k$ are given by $$\begin{aligned} \alpha _k &=& \frac{i}{\sqrt{4\omega _1\omega _2}} \biggl[\frac{a_2}{a_1}\gamma _2^*-\frac{a_1}{a_2}\gamma _1 -a_1a_2\gamma _1\gamma _2^*\int _{\eta _1}^{\eta _2}\frac{{\rm d}\tau }{a^2(\tau )}\biggr]{\rm e}^{i(\Omega _2-\Omega _1)}\, , \\ \beta _k &=& -\frac{i}{\sqrt{4\omega _1\omega _2}} \biggl[\frac{a_2}{a_1}\gamma _2-\frac{a_1}{a_2}\gamma _1 -a_1a_2\gamma _1\gamma _2\int _{\eta _1}^{\eta _2}\frac{{\rm d}\tau }{a^2(\tau )}\biggr]{\rm e}^{-i(\Omega _2+\Omega _1)}\, . \end{aligned}$$ These expressions are similar to those found in Refs. [@LLMU] and [@BJM]. So far, no approximation has been made. To go further, one has to take into account the fact that the difference $\vert \eta _2(k)-\eta _1(k)\vert$ cannot be too large. Otherwise, this would mean that particles production in region III is too important and, as a consequence, that the calculation is not valid due to this back-reaction problem. Therefore, we can write $\eta _1=\eta _2(1+\Delta )$ and perform an expansion of the coefficients $\alpha _k$ and $\beta _k$ in terms of the parameter $\Delta $. The result reads $$\alpha _k={\rm e}^{i(\Omega _2-\Omega _1)} \biggl[1+\frac{i}{2}\omega _2\eta _2\biggl(1+\frac{Q_2}{\omega _2^2} -\frac{a_2''}{\omega _2^2a_2}\biggl)\Delta \biggr]+{\cal O}(\Delta^2 )\, , \quad \beta _k=\frac{i}{2}{\rm e}^{-i(\Omega _2+\Omega _1)} \omega _2\eta _2\biggl(1-\frac{Q_2}{\omega _2^2} +\frac{a_2''}{\omega _2^2a_2}\biggl)\Delta +{\cal O}(\Delta^2 )\, .$$ In this expression, $Q$ is the parameter which controls the accuracy of the WKB approximation. It is defined by $Q\equiv -\omega ''/(2\omega )+3(\omega ')^2/(4\omega ^2)$ [@MSwkb]. As expected, if we send the parameter $\Delta $ to zero, then $\beta _k$ vanishes and $\alpha _k$ becomes unity. Inserting the above expressions for $\alpha _k$ and $\beta _k$ into Eq. (\[mu4\]), one arrives at $$\label{D+} \vert D_+(k)\vert ^2=\frac{1}{2a_3^2\omega _3} \biggl\{1-\omega _2\eta _2\biggl(1-\frac{Q_2}{\omega _2^2} +\frac{a_2''}{\omega _2^2a_2}\biggl)\sin \biggl[\int _{\eta _1}^{\eta _3} \omega (k,\tau ){\rm d}\tau \biggr]\Delta +{\cal O}(\Delta ^2)\biggr\}\, .$$ {width="18cm" height="10cm"} Notice that in order to establish the previous expression we have just assumed that $\Delta $ is a small number but we have not assumed that $k_{_{\rm C}}/H$ or $k_*/H$ are small. As expected, The correction to $\vert D_+(k)\vert ^2$ is given by the Bogoluibov coefficient $\beta _k$. We now need to calculate explicitly the previous quantities. For this purpose, we introduce an approximate piecewise form of the dispersion relation considered before, namely $$\omega _{_{\rm phys}}= \cases{ k_{_{\rm phys}}\, & $k_{_{\rm phys }}<k_{_{\rm C}} \, ,$ \cr & \cr \alpha (k_{_{\rm C}}-k_{_{\rm phys }})+k_{_{\rm C}}\, & $k_{_{\rm C}}<k_{_{\rm phys }}<k_*\, ,$ \cr & \cr sk_{_{\rm phys }}+(\alpha +1)k_{_{\rm C}}-(\alpha +s)k_*\, & $k_{_{\rm phys }}>k_* \, .$ }$$ This piecewise dispersion relation is represented in Fig. \[piece\]. The parameter $\alpha $ controls the slope of the dispersion relation in the region where the group velocity does not have the same sign as the phase velocity. The parameter $s$ controls the slope in the region $k_{_{\rm phys}}>k_*$. Assuming that the space-time is de Sitter, one can also compute the comoving frequency $\omega =a\omega _{_{\rm phys}}(k/a)$ $$\omega (k,\eta )= \cases{ k\, & $k_{_{\rm phys }}<k_{_{\rm C}} \, ,$ \cr & \cr \displaystyle{-\alpha k-\frac{\alpha +1}{\eta } \frac{k_{_{\rm C}}}{H}}\, & $k_{_{\rm C}}<k_{_{\rm phys }}<k_*\, ,$ \cr & \cr \displaystyle{sk+\frac{1}{\eta }\biggl[ (\alpha +s)\frac{k_*}{H} -(\alpha +1)\frac{k_{_{\rm C}}}{H}\biggr]}\, & $k_{_{\rm phys }}>k_* \, .$ }$$ Obviously, one has to choose $k_{_{\rm C}}>\alpha k_*/(\alpha +1)$ in order to insure that the frequency remains positive. The various times of matching can now be calculated very simply. They read $$\begin{aligned} \eta _1(k) &=& -\frac{1}{ks}\biggl[\sqrt{2}+(\alpha +s)\frac{k_*}{H}-(\alpha +1)\frac{k_{_{\rm C}}}{H}\biggl]\, , \\ \eta _2(k) &=& -\frac{1}{k}\biggl(-\frac{\sqrt{2}}{\alpha }+\frac{\alpha +1}{\alpha }\frac{k_{_{\rm C}}}{H}\biggr)\, , \\ \eta _3(k) &=& -\frac{1}{k}\, . \end{aligned}$$ The times at which $\omega _{_{\rm phys}}=k_{_{\rm C}}$ and $\omega _{_{\rm phys}}=k_*$, i.e. $\eta _{_{\rm C}}(k)$ and $\eta _*(k)$ respectively (see Fig. \[piece\]), can also be determined easily. They are given by $\eta _{_{\rm C}}(k)= -(1/k)(k_{_{\rm C}}/H)$ and $\eta _*(k)= -(1/k)(k_*/H)$. We have based the calculation of $\vert D_+(k)\vert ^2$ on the assumption that the parameter $\Delta $ is small in order to avoid a back-reaction problem. This means that the time spent by the modes of interest in the region where the WKB approximation is violated is small. In turn, this requires a link between the two scales $k_{_{\rm C}}$ and $k_*$ which characterizes the shape of the dispersion relation and the Hubble constant $H$ which characterizes the “velocity” with which a mode crosses the region where the WKB approximation is not valid. Using the expressions of $\eta _1(k)$ and $\eta _2(k)$, one finds that the link between $k_*$, $k_{_{\rm C}}$ and $\Delta $ can be expressed as $$\frac{k_*}{H}=\biggl[\frac{1}{\alpha }+ \frac{s}{\alpha (\alpha +s)}\Delta \biggr] \biggl[-\sqrt{2}+(\alpha +1)\frac{k_{_{\rm C}}}{H}\biggr]\, .$$ If $k_{_{\rm C}}\gg H$ and $k_*\gg H$, then one has $k_*\simeq (\alpha +1)k_{_{\rm C}}/\alpha $, as expected. The final result can be expressed in terms of $k_{_{\rm C}}/H$ and $\Delta $ only. We can now calculate each term present in Eq. (\[D+\]). Straightforward calculations show that $Q/\omega ^2\vert _2=-(\alpha +1)k_{_{\rm C}}/(2\sqrt{2}H)+3(\alpha +1)^2 k_{_{\rm C}}^2/(16H^2)$, $a''/(a\omega ^2)\vert _2=1$ and $\omega \eta \vert _2=-\sqrt{2}$. In order to calculate the integral appearing in the argument of the sine function in Eq. (\[D+\]), one has to cut it into several pieces and to use the corresponding form of the piecewise dispersion relation. One obtains $$\begin{aligned} & & \int _{\eta _1}^{\eta _3}\omega (k,\tau ){\rm d}\tau =\int _{\eta _2}^{\eta _3}\omega (k,\tau ){\rm d}\tau -\sqrt{2}\Delta \\ & & =(\sqrt{2}-1)-(\alpha +1)\frac{k_{_{\rm C}}}{H}\ln \biggl[\frac{1}{(\alpha +1)-\sqrt{2}H/k_{_{\rm C}}}\biggr] \nonumber \\ & & -\sqrt{2}\Delta \, . \end{aligned}$$ We have to insert this last expression into the sine function and expand the resulting expression once more in terms of the parameter $\Delta $. The final result reads $$\begin{aligned} k^3P_{\chi }(k)&=& \biggl(\frac{H}{2\pi }\biggr)^2 \biggl\{1+\sqrt{2}\biggl[2+\frac{\alpha +1}{2\sqrt{2}}\frac{k_{_{\rm C}}}{H} -\frac{3(\alpha +1)^2}{16}\frac{k_{_{\rm C}}^2}{H^2}\biggr]\sin \biggl[(\sqrt{2}-1) -(\alpha +1)\frac{k_{_{\rm C}}}{H} \ln \biggl(\frac{1}{\alpha +1-\sqrt{2}H/k_{_{\rm C}}}\biggr) \biggr]\Delta \nonumber \\ & &+{\cal O}(\Delta ^2) \biggr\}\, . \end{aligned}$$ As expected the result does not depend on $k$ since the scalar field lives in de Sitter space-time. It is worth noticing that we have not assumed anything about the ratio $k_{_{\rm C}}/H$ in order to obtain the previous expression (only the parameter $\Delta $ was supposed to be small). The previous expression is the main result of this section. It should be compared with Eqs. (\[minkops\]) and (\[dpsgw\]) obtained previously. We see that the magnitude of the correction is no longer controled by any power of the ratio of two scales as it was before. The magnitude of the effect is determined by the time spent in the region where the WKB approximation is violated which in turn depends on the shape of the dispersion relation that was assumed. Finally, let us return to the back-reaction problem. There is no back-reaction problem if $\vert \beta _k\vert \ll 1$. Using the expression of $\beta _k$ derived before and assuming that $k_{_{\rm C}}/H\gg 1$, we see that the parameter $\Delta $ must satisfy $\Delta \ll H^2/k_{_{\rm C}}^2$. This amounts to a severe fine-tuning of the scales $k_{_{\rm C}}$, $k_*$ and of the Hubble parameter $H$ during inflation in order to satisfy this condition. Conclusions =========== We have analyzed the magnitude of correction terms to the power spectra of scalar metric (density) fluctuations and gravitational waves in inflationary cosmology under the assumption that fluctuation modes are generated when their physical length scale equals some critical length determined by the unknown Planck-scale physics, but without modifying the equations of motion for the fluctuations. The magnitude of the correction terms can then be expressed as a function of the dimensionless ratio $\sigma_0 = H_0 / M_{_{\rm C}}$, where $H_0$ is the characteristic Hubble expansion rate during inflation, and $M_{_{\rm C}}$ is the mass scale at which the new physics sets in. It is important to realize that the magnitude of the correction terms is in general different for gravitational waves and for scalar metric fluctuations - a point not realized in some papers on the “trans-Planckian problem” of inflationary cosmology. In addition, the magnitude of the correction terms depends sensitively on the initial state chosen. We have shown that for the local Minkowski vacuum state, the correction terms are of the order $\sigma_0^3$ (in agreement with the results of [@npc]), whereas for nontrivial $\alpha$-vacua the effects are much larger. If the Bogoliubov coefficients which describe the mode mixing do not depend on $\sigma_0$, then the correction terms can be of order unity. In the case of Danielsson’s $\alpha$-vacuum for which the Bogoliubov coefficients depend on $\sigma_0$, the corrections to the gravitational wave spectrum are suppressed by one power of $\sigma_0$ (in agreement with the results of [@Dan1]), the corrections to the scalar metric fluctuation spectrum by two powers of $\sigma_0$. Thus, the consistency relation for fluctuations in inflationary cosmology obtains corrections of linear order in $\sigma_0$, as already emphasized in [@Hui]. In the final section of the paper we compared the results obtained in earlier sections with the results obtained by assuming that trans-Planckian physics leads to a modified dispersion relation. In this case, corrections to the usual power spectra of fluctuations can be obtained which are not suppressed by any small dimensionless combinations of energy scales in the problem. However, demanding that the back-reaction remains under control leads to severe fine-tuning requirements on such models. The work of R. B. is supported in part by the U.S. Department of Energy under Contract DE-FG02-91ER40688, TASK A. [99]{} R. H. Brandenberger, Proceeding of the International School on Cosmology, Kish Island, Iran (2000), \[arXiv:hep-ph/9910410\]. J. Martin and R. H. Brandenberger, Phys. Rev. D [**63**]{}, 123501 (2001), \[arXiv:hep-th/0005209\]. R. H. Brandenberger and J. Martin, Mod. Phys. Lett. A [**16**]{}, 999 (2001), \[arXiv:astro-ph/0005432\]. W. Unruh, Phys. Rev. [**D51**]{}, 2827 (1995), \[arXiv:gr-qc/9409008\]. S. Corley and T. Jacobson, Phys. Rev. [**D54**]{}, 1568 (1996), \[arXiv:hep-th/9601073\]; S. Corley, Phys. Rev. [**D57**]{}, 6280 (1998), \[arXiv:hep-th/9710075\]. J. C. Niemeyer, Phys. Rev. [**D63**]{}, 123502 (2001), \[arXiv:astro-ph/0005533\]; J. C. Niemeyer, \[arXiv:astro-ph/0201511\]. J. Martin and R. H. Brandenberger, Proceedings of the Ninth Marcel Grossmann Meeting on General Relativity, edited by R. T. Jantzen, V. Gurzadyan and R. Ruffini, World Scientific, Singapore, 2002, \[arXiv:astro-ph/0012031\]. J. C. Niemeyer and R. Parentani, Phys. Rev. [ **D64**]{}, 101301 (2001), \[arXiv:astro-ph/0101451\]. M. Lemoine, M. Lubo, J. Martin and J. P. Uzan, Phys. Rev. [**D65**]{}, 023510 (2002), \[arXiv:hep-th/0109128\]. R. H. Brandenberger, S. E. Joras and J. Martin, Phys. Rev. D [**66**]{}, 083514 (2002), \[arXiv:hep-th/0112122\]. S. Shankaranarayan, Class. Quant. Grav. [**20**]{}, 75 (2003), \[arXiv:gr-qc/0203060\]. A. Kempf, Phys. Rev. [**D63**]{} (2001) 083514, \[arXiv:astro-ph/0009209\]. C. S. Chu, B. R. Greene and G. Shiu, Mod. Phys. Lett. [**A16**]{}, 2231 (2001), \[arXiv:hep-th/0011241\]. R. Easther, B. R. Greene, W. H. Kinney and G. Shiu, Phys. Rev. [**D64**]{}, 103502 (2001), \[arXiv:hep-th/0104102\]; R. Easther, B. R. Greene, W. H. Kinney and G. Shiu, \[arXiv:hep-th/0110226\]. A. Kempf and J. C. Niemeyer, Phys. Rev. [**D64**]{}, 103501 (2001), \[arXiv:astro-ph/0103225\]. L. Hui and W. H. Kinney, \[arXiv:astro-ph/0109107\]. R. Brandenberger and P. M. Ho, Phys. Rev. D [**66**]{}, 023517 (2002), \[AAPPS Bull.  [**12N1**]{}, 10 (2002)\], \[arXiv:hep-th/0203119\]. S. F. Hassan and M. S. Sloth, \[arXiv:hep-th/0204110\]. F. Lizzi, G. Mangano, G. Miele and M. Peloso, noncommutative geometry,” JHEP [**0206**]{}, 049 (2002), \[arXiv:hep-th/0203099\]. U. H. Danielsson, Phys. Rev. D [**66**]{}, 023511 (2002), \[arXiv:hep-th/0203198\]; U. H. Danielsson, JHEP [**0207**]{}, 040 (2002), \[arXiv:hep-th/0205227\]. K. Goldstein and D. A. Lowe, Phys. Rev. D [**67**]{}, 063502 (2003), \[arXiv:hep-th/0208167\]. G. L. Alberghi, R. Casadio and A. Tronconi, \[arXiv:gr-qc/0303035\]. L. Bergstrom and U. H. Danielsson, JHEP [**0212**]{}, 038 (2002), \[arXiv:hep-th/0211006\]. T. Banks and L. Mannelli, \[arXiv:hep-th/0209113\]. M. B. Einhorn and F. Larsen, Phys. Rev. D [**67**]{}, 024001 (2003), \[arXiv:hep-th/0209159\]. N. Kaloper, M. Kleban, A. Lawrence, S. Shenker and L. Susskind, JHEP [**0211**]{}, 037 (2002), \[arXiv:hep-th/0209231\]. K. Goldstein and D. A. Lowe, \[arXiv:hep-th/0302050\]. R. Easther, B. R. Greene, W. H. Kinney and G. Shiu, Phys. Rev. D [**66**]{}, 023518 (2002), \[arXiv:hep-th/0204129\]. J. C. Niemeyer, R. Parentani and D. Campo, Phys. Rev. D [**66**]{}, 083510 (2002), \[arXiv:hep-th/0206149\]. N. Kaloper, M. Kleban, A. E. Lawrence and S. Shenker, Phys. Rev. D [**66**]{}, 123510 (2002), \[arXiv:hep-th/0201158\]. R. H. Brandenberger and J. Martin, Int. J. Mod. Phys. A [**17**]{}, 3663 (2002), \[arXiv:hep-th/0202142\]. C. P. Burgess, J. M. Cline, F. Lemieux and R. Holman, JHEP [**0302**]{}, 048 (2003), \[arXiv:hep-th/0210233\]. V. Bozza, M. Giovannini and G. Veneziano, \[arXiv:hep-th/0302184\]. C. Armendariz-Picon and E. A. Lim, \[arXiv:hep-th/0303103\]. D. J. Chung, A. Notari and A. Riotto, \[arXiv:hep-ph/0305074\]. V. F. Mukhanov, H. A. Feldman and R. H. Brandenberger, Phys. Rept.  [**215**]{}, 203 (1992). L. P. Grishchuk, Zh. Eksp. Teor. Fiz. [**64**]{}, 825 (1974), \[Sov. Phys. JETP [**40**]{}, 409 (1974)\]. J. Martin and D. J. Schwarz, Phys. Rev. D [**67**]{}, 083512 (2003), \[arXiv:astro-ph/0210090\]. J. M. Bardeen, Phys. Rev. D [**22**]{}, 1882 (1980). V. F. Mukhanov, Sov. Phys. JETP [**67**]{}, 1297 (1988) \[Zh. Eksp. Teor. Fiz.  [**94N7**]{}, 1 (1988 ZETFA,94,1-11.1988)\]. J. Martin and D. J. Schwarz, Phys. Rev. D [**57**]{}, 3302 (1998), \[arXiv:gr-qc/9704049\]. J. E. Lidsey, A. R. Liddle, E. W. Kolb, E. J. Copeland, T. Barreiro and M. Abney, Rev. Mod. Phys.  [**69**]{}, 373 (1997), \[arXiv:astro-ph/9508078\]. D. J. Schwarz, C. A. Terrero-Escalante and A. A. Garcia, Phys. Lett. B [**517**]{}, 243 (2001), \[arXiv:astro-ph/0106020\]. J. Martin and D. J. Schwarz, Phys. Rev. D [**62**]{}, 103520 (2000), \[arXiv:astro-ph/9911225\]. J. Martin, A. Riazuelo and D. J. Schwarz, Astrophys. J.  [**543**]{}, L99 (2000), \[arXiv:astro-ph/0006392\]. Ulf Danielsson, private communication. [^1]: Recently, Ref. [@Alberghi] has shown that effects of the order $\sigma _0$ also occur in models of power-law inflation. [^2]: There has been a concern that nontrivial $\alpha$ vacua are problematic from the point of view of interacting quantum field theory [@Banks; @Einhorn; @Kaloper2]. However, very recently it has been shown [@Lowe2] how to define interacting quantum field theory about an $\alpha$ vacuum. [^3]: Using the asymptotic behavior of the Hankel function $$H^{(1)}_{\nu }(z)\rightarrow _{z\rightarrow +\infty } \sqrt{\frac{2}{\pi z}}{\rm e}^{i(z-\pi \nu /2-\pi /4)}\, ,$$ one can easily check that the mode function $\mu _{_{\rm S,T}}^{\rm stand}(k,\eta )$ has indeed the required limit given by Eq. (\[stic\]).

ArXiv

--- abstract: 'We study the joint laws of the maximum and minimum of a continuous, uniformly integrable martingale. In particular, we give explicit martingale inequalities which provide upper and lower bounds on the joint exit probabilities of a martingale, given its terminal law. Moreover, by constructing explicit and novel solutions to the Skorokhod embedding problem, we show that these bounds are tight. Together with previous results of Azéma & Yor, Perkins, Jacka and Cox & Obłój, this allows us to completely characterise the upper and lower bounds on all possible exit/no-exit probabilities, subject to a given terminal law of the martingale. In addition, we determine some further properties of these bounds, considered as functions of the maximum and minimum.' author: - | Alexander M. G. Cox[^1]\ Dept. of Mathematical Sciences\ University of Bath\ Bath BA2 7AY, UK - | Jan Obłój[^2]\ Mathematical Institute\ University of Oxford\ Oxford, OX2 6GG bibliography: - 'doubleexit.bib' title: 'On joint distributions of the maximum, minimum and terminal value of a continuous uniformly integrable martingale ' --- Introduction ============ The study of the running maximum and minimum of a martingale has a prominent place in probability theory, starting with Doob’s maximal and $L^p$ inequalities. In seminal contributions, Blackwell and Dubins [@BD63], Dubins and Gilat [@DubinsGilat:78] and Azéma and Yor [@AzemaYor:79; @AzemaYor:79b] established that the distribution of the maximum ${\overline{M}}_\infty:=\sup_{t\leq \infty} M_t$ of a uniformly integrable martingale $M$ is bounded from above, in stochastic order, by the so called Hardy-Littlewood transform of the distribution of $M_\infty$, and the bound is attained. This led to series of studies on the possible distributions of $(M_\infty,{\overline{M}}_\infty)$ including Gilat and Meilijson [@GM88], Kertz and Rösler [@KR90; @KertzRosler:92b; @KertzRosler:93], Rogers [@Rogers:93], Vallois [@Vallois:93], see also Carraro, El Karoui and Obłój [@CarraroElKarouiObloj:09]. More recently, these problems have gained a new momentum from applications in the field of mathematical finance. The bounds on the distribution of the maximum, given the distribution of the terminal value, are interpreted as bounds on prices of barrier options given the prices of (vanilla) European options. Further, the bounds are often obtained by devising pathwise inequalities which then have the interpretation of (super) hedging strategies. This approach is referred to as robust pricing and hedging and goes back to Hobson [@Hobson:98b], see also Obłój [@Obloj:EQF] and Hobson [@Hobson:10] for survey papers. More recently, for example in Acciaio et. al. [@Acciaio:2013ab], martingale inequalities have been used to study some classical probabilistic inequalities, and are of interest in their own right. Here we propose to study the distribution of $({\overline{M}}_\infty,{\underline{M}}_\infty)$, where ${\underline{M}}_{\infty}:= \inf_{t \le \infty} M_t$ is the infimum of the process, given the distribution of $M_\infty$, for a uniformly integrable continuous martingale $M$. More precisely, we present sharp lower and upper bounds on all double exit/no-exit probabilities for $M$ in terms of the distribution of $M_\infty$, i.e. the probabilities that ${\overline{M}}_\infty$ is *greater/smaller* than ${\ensuremath{\overline{b}}}$ *and/or* that ${\underline{M}}_\infty$ is *greater/smaller* than ${\ensuremath{\underline{b}}}$, for some barriers ${\ensuremath{\underline{b}}}< {\ensuremath{\overline{b}}}$. This amounts to considering eight different events. They of course come in pairs, [e.g.]{} $\{{{\overline{M}}_{\infty} \ge {\ensuremath{\overline{b}}}, {\underline{M}}_{\infty} > {\ensuremath{\underline{b}}}}\}$ is the complement of $\{{\overline{M}}_\infty<{\ensuremath{\overline{b}}}\textrm{ or }{\underline{M}}_\infty\leq {\ensuremath{\underline{b}}}\}$ and, by symmetry, it suffices to consider only one of $\{{{\overline{M}}_{\infty} \ge {\ensuremath{\overline{b}}}, {\underline{M}}_{\infty} > {\ensuremath{\underline{b}}}}\}$ and $\{{\overline{M}}_\infty<{\ensuremath{\overline{b}}}, {\underline{M}}_{\infty}\leq {\ensuremath{\underline{b}}}\}$. It follows that to provide a complete description it suffices to consider the three events $$\label{eq:eventsofinterest} \{{\overline{M}}_\infty\geq {\ensuremath{\overline{b}}}, {\underline{M}}_\infty\leq {\ensuremath{\underline{b}}}\},\quad \{{\overline{M}}_\infty< {\ensuremath{\overline{b}}}, {\underline{M}}_\infty>{\ensuremath{\underline{b}}}\}\ \textrm{ and }\ \{{{\overline{M}}_{\infty} \ge {\ensuremath{\overline{b}}}, {\underline{M}}_{\infty} > {\ensuremath{\underline{b}}}}\}.$$ By continuity and time-change arguments, it follows that for a fixed distribution $\mu$ of $M_\infty$, our problem is equivalent to studying these events for $M_t=B_{t\wedge \tau}$ where $\tau$ varies among all stopping times such that $M$ is uniformly integrable and $M_\infty=B_\tau$ has distribution $\mu$, i.e. solutions to the Skorokhod embedding problem for $\mu$ in $B$, see Obłój [@Obloj:04b]. Sharp bounds on the probability of the first event in follow from Perkins and tilted-Jacka solutions, see Section \[ap:mart\] below. The case of the second event was treated in Cox and Obłój [@Cox:2011aa] and is also recalled in Section \[ap:mart\]. Our contribution here is twofold. First, we derive lower and upper bounds on ${{\mathbb{P}}}({{\overline{M}}_{\infty} \ge {\ensuremath{\overline{b}}}, {\underline{M}}_{\infty} > {\ensuremath{\underline{b}}}})$ in terms of the distribution of $M_\infty$ and give explicit constructions of martingales which attain the bounds. We do this by devising pathwise inequalities which give upper and lower bounds and then by constructing two new solutions to the Skorokhod embedding problem for which equalities are attained in our pathwise inequalities. Second, we study universal qualitative properties of the probabilities of the events in seen as surfaces in the parameters ${\ensuremath{\underline{b}}},{\ensuremath{\overline{b}}}$. While the techniques used to derive the bounds on ${{\mathbb{P}}}({{\overline{M}}_{\infty} \ge {\ensuremath{\overline{b}}}, {\underline{M}}_{\infty} > {\ensuremath{\underline{b}}}})$ are not new, the explicit constructions we need to use are novel, and our goal in the first part of the paper is to provide those bounds which are currently not known; in this sense, we complete previous work in the literature. The contribution in the second part of the paper is, to the best of our knowledge, the first attempt to address questions of this nature. Motivation ---------- We believe that there are two natural motivations for our results. First, we believe we solve an intrinsically interesting probabilistic question and second, our results correspond to robust pricing and hedging of certain double barrier options in finance. We elaborate now on both. From the probabilistic point of view, we follow in the footsteps of seminal works mentioned above. The results therein were typically stated for a martingale and its maximum but naturally can be reformulated for a martingale and its minimum ${\underline{M}}_\infty$. They grant us a full understanding of possible joint distributions of couples $(M_\infty,{\overline{M}}_\infty)$ or $(M_\infty,{\underline{M}}_\infty)$. In contrast, much less is known about the joint distribution of $(M_\infty,{\overline{M}}_\infty,{\underline{M}}_\infty)$ and it proves much harder to study (although promising recent progress has been made in this direction in a discrete time setting, when one considers the joint law of a random walk, its maximum, minimum and [*signature*]{} by [@Duembgen:2014aa]). Indeed, already in the case of Brownian motion $B$, while the distribution of $(B_t,\overline{B}_t)$ is readily accessible with a simply and explicit density, the distribution of the triplet $(B_t,\overline{B}_t,\underline{B}_t)$ is described through an infinite series. Likewise, ${{\mathbb{P}}}({\overline{M}}_\infty\geq {\ensuremath{\overline{b}}})$ is maximised among all martingales $M$ with a fixed distribution of $M_\infty$, by one extremal martingale simultaneously for all ${\ensuremath{\overline{b}}}$. In contrast, as we will show here, maximising ${{\mathbb{P}}}({{\overline{M}}_{\infty} \ge {\ensuremath{\overline{b}}}, {\underline{M}}_{\infty} > {\ensuremath{\underline{b}}}})$ will require martingales with qualitatively different behaviour for different values of $({\ensuremath{\underline{b}}},{\ensuremath{\overline{b}}})$. In terms of mathematical finance, the constructions presented here correspond to robust pricing (and hedging) of double touch/no-touch barrier options — for a detailed discussion of applications we refer to our earlier papers [@Cox:2011ab; @Cox:2011aa] where we studied the first two events in . Such an option would pay out $1$ if and only if one barrier is attained *and* a second given barrier is not attained, i.e. we consider the payoff of the form $\{{\overline{S}}_T\geq {\ensuremath{\overline{b}}},{\underline{S}}_T>{\ensuremath{\underline{b}}}\}$, where $(S_t:t\leq T)$ is a uniformly integrable martingale representing the stock price process. The double touch/no-touch options are partially a theoretical construct — (to the best of our knowledge) they are not commonly traded even in Foreign Exchange (FX) markets, where barriers options are most popular. However, they prove useful as they can be represented as a sum or difference of other barrier options. We can then interpret our results as super-/sub-hedges for sums and differences of barrier options. More precisely, we can write $$\begin{aligned} {\boldsymbol{1}_{\{\ensuremath{{\overline{S}}_T\geq {\ensuremath{\overline{b}}},{\underline{S}}_T>{\ensuremath{\underline{b}}}}\}}} &=& {\boldsymbol{1}_{\{\ensuremath{{\overline{S}}_T\geq {\ensuremath{\overline{b}}}}\}}}-{\boldsymbol{1}_{\{\ensuremath{{\overline{S}}_T\geq {\ensuremath{\overline{b}}},{\underline{S}}_T\leq {\ensuremath{\underline{b}}}}\}}} \label{eq:dtnt_decompose1}\\ &=& 1-\Big({\boldsymbol{1}_{\{\ensuremath{{\underline{S}}_T\leq {\ensuremath{\underline{b}}}}\}}}+ {\boldsymbol{1}_{\{\ensuremath{{\overline{S}}_T < {\ensuremath{\overline{b}}},{\underline{S}}_T > {\ensuremath{\underline{b}}}}\}}}\Big) \label{eq:dtnt_decompose2}.\end{aligned}$$ The first decomposition writes the payoff of a double touch/no-touch option as a difference of a one-touch option (with payoff ${\boldsymbol{1}_{\{\ensuremath{{\overline{S}}_T\geq {\ensuremath{\overline{b}}}}\}}}$) and a double touch option. The second decomposition writes the payoff of a double touch/no-touch option as one minus the portfolio of a one-touch option and a double no-touch (range) option with payoff ${\boldsymbol{1}_{\{\ensuremath{{\overline{S}}_T< {\ensuremath{\overline{b}}},{\underline{S}}_T> {\ensuremath{\underline{b}}}}\}}}$. This is of particular interest as both one-touch and range options are liquidly traded in main currency pairs in FX markets. Effectively, using the no-arbitrage prices derived in Theorems \[thm:upper\_price\_mixed\] and \[thm:lower\_price\_mixed\] below, we obtain a way of checking for absence of arbitrage in the observed prices of European calls/puts, one-touch and range options. Furthermore, if one-touch options are liquidly traded, we can then exploit pathwise inequalities derived in this paper as super- or sub-hedging strategies for range options or double touch options. For certain barriers this will be sharper than the hedges derived in Cox and Obłój [@Cox:2011ab; @Cox:2011aa] which assumes only that vanilla options are liquid. Notation {#sec:notation} -------- Throughout the paper $M$ denotes a continuous uniformly integrable martingale and $B$ a standard real-valued Brownian motion. The running maximum and minimum of a Brownian motion $B$ or a martingale $M$ are denoted respectively ${\ensuremath{\overline{B}}}_t=\sup_{u\leq t}B_u$ and ${\ensuremath{\underline{B}}}_t=\inf_{u\leq t}B_u$, and similarly ${\overline{M}}_t$ and ${\underline{M}}_t$. The first hitting times of levels are denoted $H_x(B):=\inf\{t\geq 0: B_t=x\}$, $x\in {\mathbb{R}}$. Likewise we will consider $H_x(M)$ and $H_x(\omega)$, the first hitting times for a martingale $M$ and a continuous path $\omega$. Most of the time we simply write $H_x$ as it should be clear from the context which process/path we consider. We will use the hitting times primarily to express events involving the running maximum and minimum, e.g. note that ${{\boldsymbol{1}_{\{\ensuremath{\overline{B}_\tau\geq {\ensuremath{\overline{b}}},\, \underline{B}_\tau> {\ensuremath{\underline{b}}}}\}}}}={\boldsymbol{1}_{\{\ensuremath{H_{{\ensuremath{\overline{b}}}}\leq \tau < H_{{\ensuremath{\underline{b}}}}}\}}}$ a.s.. We also introduce the following notation to indicate composition of stopping times: if $\tau_1, \tau_2$ are both stopping times, then the stopping time $(\tau_2 \circ \tau_1)(\omega) = \tau_1(\omega) + \tau_2(\theta_{\tau_1}(\omega))$, where $\theta_t(\omega)$ is the usual shift operator, $\theta_t:C({\mathbb{R}}_+) \to C({\mathbb{R}}_+)$ defined by $(\theta_t(\omega))_s = \omega_{t+s}$. We use the notation $a \ll b$ to indicate that $a$ is *much smaller* than $b$ – this is only used to give intuition and is not rigorous. The minimum and maximum of two numbers are denoted $a\land b=\min\{a,b\}$ and $a\lor b=\max\{a,b\}$ respectively, and the positive part is denoted $a^+=a\lor 0$. Finally, for a probability measure $\mu$ on ${\mathbb{R}}$ we let $-\infty\leq {\ell_{\mu}}<{r_{\mu}}\leq \infty$ be the bounds of the support of $\mu$, i.e.$[{\ell_{\mu}},{r_{\mu}}]$ is the smallest interval with $\mu([{\ell_{\mu}},{r_{\mu}}])=1$. Bounds for the probability of double exit/no-exit {#sec:dtnt} ================================================= In this section we provide sharp bounds on the probability $${{\mathbb{P}}}\left({{\overline{M}}_{\infty} \ge {\ensuremath{\overline{b}}}, {\underline{M}}_{\infty} > {\ensuremath{\underline{b}}}}\right)$$ where ${\ensuremath{\underline{b}}}< 0 < {\ensuremath{\overline{b}}}$, and $M=(M_t:t\leq \infty)$ is a continuous uniformly integrable martingale. Our approach will involve two steps: first we provide pathwise inequalities which induce upper and lower bounds on the given event. Second, we show that these bounds are attained. More specifically, consider a continuous path $(\omega_t: 0\leq t\leq T)$, where $T\leq \infty$. We will introduce pathwise inequalities comparing ${{\boldsymbol{1}_{\{\ensuremath{{\overline{\omega}_{T} \ge {\ensuremath{\overline{b}}}, \underline{\omega}_{T} > {\ensuremath{\underline{b}}}}}\}}}}$ to a sum of a “static term," some function $f(\omega_T),$ and a “dynamic term" of the generic form $\beta (\omega_T-b){\boldsymbol{1}_{\{\ensuremath{H_b<T}\}}}$. Note that such a dynamic term is zero initially and, when $b$ is hit, it introduces a $\beta$-rotation of $f(\omega_T)$ around $b$. Note also that when evaluated on paths of a martingale, it will be a martingale. Consequently, we will construct random variables which dominate (or are dominated by) the random variable ${{\boldsymbol{1}_{\{\ensuremath{{{\overline{M}}_{\infty} \ge {\ensuremath{\overline{b}}}, {\underline{M}}_{\infty} > {\ensuremath{\underline{b}}}}}\}}}}$ and which can be decomposed into a martingale term and a function of the terminal value $M_\infty$. Bounds on the double exit/no-exit probability above will be obtained by taking expectations in these inequalities. We further claim that these bounds are tight. This is proven in the subsequent section, where we build extremal martingales by designing optimal solutions to the Skorokhod embedding problem for Brownian motion. Pathwise inequalities: upper bounds {#sec:pathwise} ----------------------------------- We need to consider three different inequalities. As we will see later, it is always optimal to use exactly one of them, and the choice depends on the distribution of $M_\infty$ and the values of ${\ensuremath{\overline{b}}},{\ensuremath{\underline{b}}}$. We give the cases intuitive labels, their meaning will become clearer when we subsequently construct extremal martingales. Throughout this and the next section we assume that $0<T\leq \infty$ is fixed and $(\omega_t: 0\leq t\leq T)$ is a given continuous function. The hitting times are relative to $\omega$. To keep the notation simple we do not emphasise the dependence on $\omega$, e.g. $ H_{{\ensuremath{\underline{b}}}}=H_{{\ensuremath{\underline{b}}}}(\omega):= \inf\{t\leq T: \omega_t={\ensuremath{\underline{b}}}\}$, or ${\overline{G}}^I(K)={\overline{G}}^I(K,(\omega_t:t\leq T))$. *${\overline{G}}^I$: upper bound for ${\ensuremath{\underline{b}}}\ll 0<{\ensuremath{\overline{b}}}$.*\ The inequality is presented graphically in Figure \[fig:dbmp\_G1\]. We can write it as: $$\begin{aligned} {{\boldsymbol{1}_{\{\ensuremath{{\overline{\omega}_{T} \ge {\ensuremath{\overline{b}}}, \underline{\omega}_{T} > {\ensuremath{\underline{b}}}}}\}}}}& \le & \frac{1}{(K-{\ensuremath{\underline{b}}})}\left((\omega_T - K)^+ - ({\ensuremath{\underline{b}}}-\omega_T)^+ - (\omega_T - {\ensuremath{\underline{b}}}) {\boldsymbol{1}_{\{\ensuremath{H_{{\ensuremath{\underline{b}}}}<T}\}}}\right)+{\boldsymbol{1}_{\{\ensuremath{\omega_T> {\ensuremath{\underline{b}}}}\}}}\nonumber \\ & & {} =: {\overline{G}}^I(K), \label{eq:uG1def}\end{aligned}$$ where we assume $K>{\ensuremath{\underline{b}}}$. We include here the special case where $K=\infty$, which corresponds to the upper bound ${{\boldsymbol{1}_{\{\ensuremath{{\overline{\omega}_{T} \ge {\ensuremath{\overline{b}}}, \underline{\omega}_{T} > {\ensuremath{\underline{b}}}}}\}}}}\le {\boldsymbol{1}_{\{\ensuremath{\omega_T \ge {\ensuremath{\underline{b}}}}\}}}$. Note that the coefficient $1/(K-{\ensuremath{\underline{b}}})$ is taken so that the right-hand side after rotation at time $H_{{\ensuremath{\underline{b}}}}$ is zero above $K$. import graph; real xmin = -2; real xmax = 3; real eps = 0.25; pen p = black+1; pen q = black+0.75; draw((xmin,0)–(xmax+eps,0),q,Arrow); label(“$B_{\tau}$”,(xmax+eps,0),S); real lb = -1; real ub = 1; real K = 2; real slope = 1/(K-lb); pen p2 = deepblue+1+longdashdotted; draw((xmin,-slope\*(lb-xmin))–(lb,0),p2,legend=“$t \ge H_{{\ensuremath{\underline{b}}}}$”); draw((lb,1)–(K,1)–(xmax,1+(xmax-K)\*slope),p2); pen p3 = deepgreen+1+dashed; draw((xmin,0)–(lb,0),p3,legend=“$t < H_{{\ensuremath{\underline{b}}}}$”); draw((lb,1)–(K,0)–(xmax,0),p3); label(“${\ensuremath{\underline{b}}}$”,(lb,0),S); label(“${\ensuremath{\overline{b}}}$”,(ub,0),S); label(“$K$”,(K,0),S); real xlab = -1.8; real ylab = 1.8; real lablength = 0.8; real labskip = 0.3; real labextra = 0.2; frame fr = legend(invisible); fr = scale(0.75)\*fr; add(fr,(xlab,ylab)); roundbox(fr); // draw((xlab,ylab)–(xlab+lablength,ylab),p2); // label(“$t \ge H_{{\ensuremath{\underline{b}}}}$”,(xlab+lablength+labextra,ylab),E,p); // draw((xlab,ylab+labskip)–(xlab+lablength,ylab+labskip),p3); // label(“$t < H_{{\ensuremath{\underline{b}}}}$”,(xlab+lablength+labextra,ylab+labskip),E,p); clip(box((xmin-eps,-0.5-eps),(xmax+2\*eps,2.25+eps))); *${\overline{G}}^{II}$: upper bound for ${\ensuremath{\underline{b}}}< 0 < {\ensuremath{\overline{b}}}$.*\ This is a fairly simple case: if we hit neither ${\ensuremath{\underline{b}}}$ nor ${\ensuremath{\overline{b}}}$, the inequality is simply $0 \le \alpha_1(\omega_T - {\ensuremath{\underline{b}}})$ for some $\alpha_1 >0$, so that the value is $1$ if we strike ${\ensuremath{\overline{b}}}$ initially, and $0$ if we strike ${\ensuremath{\underline{b}}}$ initially. This strategy is illustrated in Figure \[fig:dbmp\_G2\]. If the path hits either ${\ensuremath{\overline{b}}}$ or ${\ensuremath{\underline{b}}}$ we have a constant value of either $1$ or $0$ respectively: $$\begin{aligned} {{\boldsymbol{1}_{\{\ensuremath{{\overline{\omega}_{T} \ge {\ensuremath{\overline{b}}}, \underline{\omega}_{T} > {\ensuremath{\underline{b}}}}}\}}}}& \le & \alpha_1 \omega_T - \alpha_0 -\alpha_1(\omega_T-{\ensuremath{\overline{b}}}){\boldsymbol{1}_{\{\ensuremath{H_{{\ensuremath{\overline{b}}}}< H_{{\ensuremath{\underline{b}}}} \wedge T}\}}} - \alpha_1 (\omega_T - {\ensuremath{\underline{b}}}) {\boldsymbol{1}_{\{\ensuremath{H_{{\ensuremath{\underline{b}}}}< H_{{\ensuremath{\overline{b}}}} \wedge T}\}}} \nonumber \\ & & {} =: {\overline{G}}^{II}. \label{eq:uG2def}\end{aligned}$$ The constraints on $\alpha_0,\alpha_1$ correspond to the need for the function to be zero if ${\ensuremath{\underline{b}}}$ is struck first, and $1$ if ${\ensuremath{\overline{b}}}$ is struck first. We deduce that $$\label{eq:uG2def_par} \begin{split} \alpha_0 & = {\ensuremath{\underline{b}}}/({\ensuremath{\overline{b}}}-{\ensuremath{\underline{b}}}) \\ \alpha_1 & = 1/({\ensuremath{\overline{b}}}-{\ensuremath{\underline{b}}}). \end{split}$$ import graph; real xmin = -2; real xmax = 3; real eps = 0.25; pen p = black+1; pen q = black+0.75; draw((xmin,0)–(xmax+eps,0),q,Arrow); label(“$B_{\tau}$”,(xmax+eps,0),S); real lb = -1; real ub = 1; // real K = 2; real slope = 1/(ub-lb); pen p2 = deepblue+1+longdashdotted; pen p3 = deepgreen+1+dashed; pen p4 = deepred + 1; draw((xmin,-slope\*(lb-xmin))–(xmax,slope\*(xmax-ub)+1),p2,legend=“$t < H_{{\ensuremath{\underline{b}}}}\wedge H_{{\ensuremath{\overline{b}}}}$”); // draw((lb,1)–(K,1)–(xmax,1+(xmax-K)\*slope),p2); draw((xmin,0)–(xmax,0),p3,legend=“$H_{{\ensuremath{\underline{b}}}} \le H_{{\ensuremath{\overline{b}}}} \wedge t$”); draw((xmin,1)–(xmax,1),p4,legend=“$H_{{\ensuremath{\overline{b}}}} \le H_{{\ensuremath{\underline{b}}}} \wedge t$”); label(“${\ensuremath{\underline{b}}}$”,(lb,0),S); label(“${\ensuremath{\overline{b}}}$”,(ub,0),S); //label(“$K$”,(K,0),S); real xlab = -1.8; real ylab = 1.8; real lablength = 0.8; real labskip = 0.3; real labextra = 0.2; frame fr = legend(invisible); fr = scale(0.75)\*fr; add(fr,(xlab,ylab)); roundbox(fr); // draw((xlab,ylab)–(xlab+lablength,ylab),p2); // label(“$t \ge H_{{\ensuremath{\underline{b}}}}$”,(xlab+lablength+labextra,ylab),E,p); // draw((xlab,ylab+labskip)–(xlab+lablength,ylab+labskip),p3); // label(“$t < H_{{\ensuremath{\underline{b}}}}$”,(xlab+lablength+labextra,ylab+labskip),E,p); clip(box((xmin-eps,-0.5-eps),(xmax+2\*eps,2.25+eps))); *${\overline{G}}^{III}$: upper bound for ${\ensuremath{\underline{b}}}< 0 \ll {\ensuremath{\overline{b}}}$.*\ The final inequality uses the fact that ${{\boldsymbol{1}_{\{\ensuremath{{\overline{\omega}_{T} \ge {\ensuremath{\overline{b}}}, \underline{\omega}_{T} > {\ensuremath{\underline{b}}}}}\}}}}\le {\boldsymbol{1}_{\{\ensuremath{\overline \omega_T \ge {\ensuremath{\overline{b}}}}\}}}$, and that the inequality for the latter also works for the former. We can then rewrite (2.2) from Brown, Hobson and Rogers [@Brown:01b] as $$\label{eq:uG3def} {{\boldsymbol{1}_{\{\ensuremath{{\overline{\omega}_{T} \ge {\ensuremath{\overline{b}}}, \underline{\omega}_{T} > {\ensuremath{\underline{b}}}}}\}}}}\leq \frac{(\omega_T-K)^+}{{\ensuremath{\overline{b}}}-K} + \frac{{\ensuremath{\overline{b}}}-\omega_T}{{\ensuremath{\overline{b}}}-K} {\boldsymbol{1}_{\{\ensuremath{\overline \omega_T \geq {\ensuremath{\overline{b}}}}\}}} =: {\overline{G}}^{III}(K),$$ where $K<{\ensuremath{\overline{b}}}$. Pathwise inequalities: lower bounds {#sec:dtnt_sub} ----------------------------------- Observe that we have ${{\boldsymbol{1}_{\{\ensuremath{{\overline{\omega}_{T} \ge {\ensuremath{\overline{b}}}, \underline{\omega}_{T} > {\ensuremath{\underline{b}}}}}\}}}}= 1-{{\boldsymbol{1}_{\{\ensuremath{{\overline{\omega}_T < {\ensuremath{\overline{b}}}\textrm{ or }\underline{\omega}_T\leq {\ensuremath{\underline{b}}}}}\}}}}$ a.s. It follows that a pathwise upper bound for ${{\boldsymbol{1}_{\{\ensuremath{{\overline{\omega}_{T} \ge {\ensuremath{\overline{b}}}, \underline{\omega}_{T} > {\ensuremath{\underline{b}}}}}\}}}}$ corresponds to a pathwise lower bound of ${{\boldsymbol{1}_{\{\ensuremath{{\overline{\omega}_T < {\ensuremath{\overline{b}}}\textrm{ or }\underline{\omega}_T\leq {\ensuremath{\underline{b}}}}}\}}}}$, and vice versa. We will use this below to rephrase some of the lower bounds as upper bounds. *${\underline{G}}_I$: lower bound for ${\ensuremath{\underline{b}}}< 0 \ll {\ensuremath{\overline{b}}}$*.\ We let ${\underline{G}}_I$ to be the trivial inequality that the probability is bounded below by zero: ${\underline{G}}_I\equiv 0$. *${\underline{G}}_{II}$: lower bound for ${\ensuremath{\underline{b}}}<0<{\ensuremath{\overline{b}}}$.*\ We describe an upper bound for ${{\boldsymbol{1}_{\{\ensuremath{{\overline{\omega}_T < {\ensuremath{\overline{b}}}\textrm{ or }\underline{\omega}_T\leq {\ensuremath{\underline{b}}}}}\}}}}$ which, as argued above, is equivalent to a lower bound for ${{\boldsymbol{1}_{\{\ensuremath{{\overline{\omega}_{T} \ge {\ensuremath{\overline{b}}}, \underline{\omega}_{T} > {\ensuremath{\underline{b}}}}}\}}}}$. The inequality depends on two parameters $K_1$ and $K_2$ where $K_1 \ge {\ensuremath{\overline{b}}}> K_2 \ge {\ensuremath{\underline{b}}}$. The construction starts with equality on the region $[K_2,{\ensuremath{\overline{b}}})$ and inequality elsewhere. The first time the path hits ${\ensuremath{\overline{b}}}$, we rotate to get equality (with zero) on $[K_1,\infty)$ and so that the value is exactly 1 at ${\ensuremath{\underline{b}}}$. If the path later hits ${\ensuremath{\underline{b}}}$, we again rotate to gain equality (with 1) on $(-\infty,K_2]$ and $[{\ensuremath{\overline{b}}},K_1]$. We write it as an inequality $$\label{eq:dbmpH3} \begin{split} {{\boldsymbol{1}_{\{\ensuremath{{\overline{\omega}_T < {\ensuremath{\overline{b}}}\textrm{ or }\underline{\omega}_T\leq {\ensuremath{\underline{b}}}}}\}}}}\leq \ &\alpha_2(K_2-\omega_T)^++(1-\alpha_4){\boldsymbol{1}_{\{\ensuremath{\omega_T< {\ensuremath{\overline{b}}}}\}}}-\alpha_2(\omega_T-{\ensuremath{\overline{b}}})^++\alpha_1(\omega_T-K_1)^++\alpha_4\\ &{}+ \beta_1(\omega_T-{\ensuremath{\overline{b}}}){\boldsymbol{1}_{\{\ensuremath{H_{{\ensuremath{\overline{b}}}}<H_{{\ensuremath{\underline{b}}}}\land T}\}}}+\beta_2(\omega_T-{\ensuremath{\underline{b}}}){\boldsymbol{1}_{\{\ensuremath{H_{{\ensuremath{\overline{b}}}}<H_{{\ensuremath{\underline{b}}}}\leq T}\}}}\\ &{}+\beta_3(\omega_T-{\ensuremath{\underline{b}}}){\boldsymbol{1}_{\{\ensuremath{H_{{\ensuremath{\underline{b}}}}<H_{{\ensuremath{\overline{b}}}}\land T}\}}}\\ &{}=:1-{\underline{G}}_{II}(K_1,K_2), \end{split}$$ which we present graphically in Figure \[fig:dbmp\_H3\]. It follows that ${\underline{G}}_{II}(K_1,K_2)$ is a lower bound for ${{\boldsymbol{1}_{\{\ensuremath{{\overline{\omega}_{T} \ge {\ensuremath{\overline{b}}}, \underline{\omega}_{T} > {\ensuremath{\underline{b}}}}}\}}}}$. We deduce immediately from the rotation conditions that $\beta_1=\alpha_2-\alpha_1$, $\beta_2=\alpha_1$ and $\beta_3=\alpha_2$. We have to satisfy two more constraints, namely that after hitting ${\ensuremath{\overline{b}}}$ and rotating the function is zero on $[K_1,\infty)$ and one at ${\ensuremath{\underline{b}}}$. Working out the values we have $$\label{eq:valuesdbmpH3} \left\{ \begin{array}{l} \alpha_1=\frac{1}{K_1-{\ensuremath{\underline{b}}}}\\ \alpha_2=\frac{{\ensuremath{\overline{b}}}-{\ensuremath{\underline{b}}}}{(K_1-{\ensuremath{\underline{b}}})({\ensuremath{\overline{b}}}-K_2)}\\ \alpha_4=\frac{K_1-{\ensuremath{\overline{b}}}}{K_1-{\ensuremath{\underline{b}}}} \end{array} \right. \quad \left\{ \begin{array}{l} \beta_1=\alpha_2-\alpha_1\\ \beta_2=\alpha_1\\ \beta_3=\alpha_2 \end{array} \right. .$$ Observe that $\alpha_4\in (0,1]$ and $0<\alpha_1\le\alpha_2$. We note that if we hit ${\ensuremath{\underline{b}}}$ before ${\ensuremath{\overline{b}}}$ we have a strict inequality in . Also, in the case where $K_2 = {\ensuremath{\underline{b}}}$ a number of the terms simplify: in particular, the construction initially gives ${\underline{G}}_{III} = 1$ for $\omega_T \in [{\ensuremath{\underline{b}}},{\ensuremath{\overline{b}}})$ for $T< H_{{\ensuremath{\overline{b}}}}$. More generally, we can also have $K_1 = {\ensuremath{\overline{b}}}$ (with or without also $K_2 = {\ensuremath{\underline{b}}}$) and all the claims remain true. import graph; real xmin = -2; real xmax = 3; real eps = 0.25; pen p = black+1; pen q = black+0.75; draw((xmin,0)–(xmax+eps,0),q,Arrow); label(“$B_{\tau}$”,(xmax+eps,0),S); real lb = -1.5; real ub = 0.75; real K1 = 1.75; real K2 = -0.25; real alpha1 = 1/(K1-lb); real alpha2 = (ub-lb)/((ub-K2)\*(K1-lb)); real alpha4 = (K1-ub)/(K1-lb); pen p2 = deepblue+1+longdashdotted; pen p3 = deepgreen+1+dashed; pen p4 = deepred + 1; draw((xmin,1+alpha2\*(K2-xmin))–(K2,1)–(ub,1),p4,legend= “$t < H_{{\ensuremath{\overline{b}}}} \wedge H_{{\ensuremath{\underline{b}}}}$”); draw((ub,alpha4)–(K1,alpha4-alpha2\*(K1-ub))–(xmax,alpha4-alpha2\*(xmax-ub)+alpha1\*(xmax-K1)),p4); draw((xmin,alpha1\*(K1-xmin))–(K2,alpha1\*(K1-K2))–(ub,1),p2,legend=“$H_{{\ensuremath{\overline{b}}}}\le t < H_{{\ensuremath{\underline{b}}}}$”); draw((ub,alpha4)–(K1,0)–(xmax,0),p2); draw((xmin,1)–(K2,1)–(ub,1-alpha4),p3,legend=“$H_{{\ensuremath{\overline{b}}}}< H_{{\ensuremath{\underline{b}}}} \le t$”); draw((ub,1)–(K1,1)–(xmax,1+alpha1\*(xmax-K1)),p3); label(“${\ensuremath{\underline{b}}}$”,(lb,0),S); label(“${\ensuremath{\overline{b}}}$”,(ub,0),S); label(“$K_2$”,(K2,0),S); label(“$K_1$”,(K1,0),S); real xlab = 1.25; real ylab = 2.15; real lablength = 0.8; real labskip = 0.3; real labextra = 0.2; frame fr = legend(invisible); fr = scale(0.75)\*fr; add(fr,(xlab,ylab)); roundbox(fr); // draw((xlab,ylab)–(xlab+lablength,ylab),p2); // label(“$t \ge H_{{\ensuremath{\underline{b}}}}$”,(xlab+lablength+labextra,ylab),E,p); // draw((xlab,ylab+labskip)–(xlab+lablength,ylab+labskip),p3); // label(“$t < H_{{\ensuremath{\underline{b}}}}$”,(xlab+lablength+labextra,ylab+labskip),E,p); clip(box((xmin-eps,-0.5-eps),(xmax+2\*eps,2.25+eps))); *${\underline{G}}_{III}$: lower bound for ${\ensuremath{\underline{b}}}\ll 0<{\ensuremath{\overline{b}}}$*.\ As previously, we describe an upper bound for ${{\boldsymbol{1}_{\{\ensuremath{{\overline{\omega}_T < {\ensuremath{\overline{b}}}\textrm{ or }\underline{\omega}_T\leq {\ensuremath{\underline{b}}}}}\}}}}$. The inequality is represented in Figure \[fig:dbmp\_H2\] and depends on two values $K_1$ and $K_2$ such that ${\ensuremath{\underline{b}}}< K_2 < K_1 < {\ensuremath{\overline{b}}}$. The inequality starts with equality (equal to 1) between $K_1$ and ${\ensuremath{\overline{b}}}$, and if we hit ${\ensuremath{\overline{b}}}$ initially, we rotate to get equality (to 0) between $K_2$ and $K_1$. If we hit ${\ensuremath{\underline{b}}}$ after this, we rotate again to ensure the function is equal to 1 below $K_2$. If we initially hit ${\ensuremath{\underline{b}}}$ rather than ${\ensuremath{\overline{b}}}$, we rotate to get a function that is generally strictly greater than one. We write it as $$\label{eq:dbmpH2} \begin{split} {{\boldsymbol{1}_{\{\ensuremath{{\overline{\omega}_T < {\ensuremath{\overline{b}}}\textrm{ or }\underline{\omega}_T\leq {\ensuremath{\underline{b}}}}}\}}}}\leq \ & \alpha_2(K_2-\omega_T)^++\alpha_1(K_1-\omega_T)^++{\boldsymbol{1}_{\{\ensuremath{\omega_T < {\ensuremath{\overline{b}}}}\}}}-\alpha_1(\omega_T-{\ensuremath{\overline{b}}})^+\\ &{} + \beta_1(\omega_T-{\ensuremath{\overline{b}}}){\boldsymbol{1}_{\{\ensuremath{H_{{\ensuremath{\overline{b}}}}<H_{{\ensuremath{\underline{b}}}}\land T}\}}}+\beta_2(\omega_T-{\ensuremath{\underline{b}}}){\boldsymbol{1}_{\{\ensuremath{H_{{\ensuremath{\overline{b}}}}<H_{{\ensuremath{\underline{b}}}}\leq T}\}}}\\ &{} +\beta_3(\omega_T-{\ensuremath{\underline{b}}}){\boldsymbol{1}_{\{\ensuremath{H_{{\ensuremath{\underline{b}}}}<H_{{\ensuremath{\overline{b}}}}\land T}\}}}\\ &{} =:1-{\underline{G}}_{III}(K_1,K_2), \end{split}$$ and it follows that ${\underline{G}}_{III}(K_1,K_2)$ is a lower bound for ${{\boldsymbol{1}_{\{\ensuremath{{\overline{\omega}_{T} \ge {\ensuremath{\overline{b}}}, \underline{\omega}_{T} > {\ensuremath{\underline{b}}}}}\}}}}$. We deduce immediately from the rotation conditions that $\beta_1=\alpha_1$, $\beta_2=\alpha_2$ and $\beta_3=\alpha_1+\alpha_2$. We have to satisfy two more constraints, namely that after hitting ${\ensuremath{\overline{b}}}$ and rotating, the function is zero on $(K_2,K_1)$ and one in ${\ensuremath{\underline{b}}}$. Working out the values we have $$\label{eq:valuesdbmpH2} \left\{ \begin{array}{l} \alpha_1=\frac{1}{{\ensuremath{\overline{b}}}-K_1}\\ \alpha_2=\frac{1}{K_2 - {\ensuremath{\underline{b}}}} \end{array} \right. \quad \left\{ \begin{array}{l} \beta_1=\alpha_1\\ \beta_2=\alpha_2\\ \beta_3=\alpha_1+\alpha_2 \end{array} \right. .$$ As in the previous case, we have a strict inequality in if the path hits ${\ensuremath{\underline{b}}}$ before ${\ensuremath{\overline{b}}}$. import graph; real xmin = -2; real xmax = 3; real eps = 0.25; pen p = black+1; pen q = black+0.75; draw((xmin,0)–(xmax+eps,0),q,Arrow); label(“$B_{\tau}$”,(xmax+eps,0),S); real lb = -1.5; real ub = 2.25; real K1 = 1; real K2 = 0.25; real slope = -1/(K2-lb); real slope2 = 1/(ub-K1); pen p2 = deepblue+1+longdashdotted; pen p3 = deepgreen+1+dashed; pen p4 = deepred + 1; draw((ub,1)–(K1,1)–(K2,1+slope2\*(K1-K2))–(xmin,1+slope2\*(K1-xmin)+slope\*(xmin-K2)),p4,legend= “$t < H_{{\ensuremath{\overline{b}}}} \wedge H_{{\ensuremath{\underline{b}}}}$”); draw((ub,0)–(xmax,-slope2\*(xmax-ub)),p4); draw((xmin,-slope\*(K2-xmin))–(K2,0)–(K1,0)–(ub,1),p2,legend=“$H_{{\ensuremath{\overline{b}}}}< t < H_{{\ensuremath{\underline{b}}}}$”); draw((ub,0)–(xmax,0),p2); // draw((lb,1)–(K,1)–(xmax,1+(xmax-K)\*slope),p2); draw((xmin,1)–(K2,1)–(K1,1-slope\*(K1-K2))–(ub,1-slope\*(ub-K2)+slope2\*(ub-K1)),p3,legend=“$H_{{\ensuremath{\overline{b}}}} < H_{{\ensuremath{\underline{b}}}} \le t$”); draw((ub,1-slope\*(ub-K2))–(xmax,1-slope\*(xmax-K2)),p3); label(“${\ensuremath{\underline{b}}}$”,(lb,0),S); label(“${\ensuremath{\overline{b}}}$”,(ub,0),S); label(“$K_2$”,(K2,0),S); label(“$K_1$”,(K1,0),S); real xlab = -1.9; real ylab = 2.15; real lablength = 0.8; real labskip = 0.3; real labextra = 0.2; frame fr = legend(invisible); fr = scale(0.75)\*fr; add(fr,(xlab,ylab)); roundbox(fr); // draw((xlab,ylab)–(xlab+lablength,ylab),p2); // label(“$t \ge H_{{\ensuremath{\underline{b}}}}$”,(xlab+lablength+labextra,ylab),E,p); // draw((xlab,ylab+labskip)–(xlab+lablength,ylab+labskip),p3); // label(“$t < H_{{\ensuremath{\underline{b}}}}$”,(xlab+lablength+labextra,ylab+labskip),E,p); clip(box((xmin-eps,-0.5-eps),(xmax+2\*eps,2.25+eps))); Probabilistic bounds {#sec:prob_bounds} -------------------- We now consider the pathwise inequalities above evaluated on a path of a continuous uniformly integrable martingale $M=(M_t:0\leq t\leq \infty)$. This gives a.s. bounds on ${{\boldsymbol{1}_{\{\ensuremath{\overline{M}_\infty\geq {\ensuremath{\overline{b}}},\, \underline{M}_\infty> {\ensuremath{\underline{b}}}}\}}}}$. By taking expectations we obtain bounds on the double exit/no-exit probabilities in terms of the distribution of $M_\infty$. Indeed, observe that each of the bounds we get can be decomposed into two terms. The first of these depends on $M_\infty$ alone, for example, in , the sum of the four quantities preceded by an $\alpha$. The second corresponds to a martingale and disappears when taking expectations, e.g. considering again , the three terms which are preceded by a $\beta$ sum to give a term with expected value zero. \[prop:prob\_upperbound\] Suppose $M=(M_{t}: 0\leq t\leq \infty)$ is a continuous uniformly integrable martingale. Then $$\label{eq:generalbound_dbmp} {{\mathbb{P}}}\left( {{\overline{M}}_{\infty} \ge {\ensuremath{\overline{b}}}, {\underline{M}}_{\infty} > {\ensuremath{\underline{b}}}}\right) \leq \inf\left\{{{\mathbb{E}}\left[ {\overline{G}}^{I}(K) \right]},{{\mathbb{E}}\left[ {\overline{G}}^{II} \right]},{{\mathbb{E}}\left[ {\overline{G}}^{III}(K') \right]}\right\},$$ where the infimum is taken over $0<K'<{\ensuremath{\overline{b}}}<K$ and where ${\overline{G}}^{I},{\overline{G}}^{II},{\overline{G}}^{III}$ are given by ,–, and respectively, evaluated on paths of $M$. Our goal is to show that the above bound is optimal. A key aspect of the above result is that the right hand-side of depends only on the distribution of $M_\infty$ and not on the law of the martingale $M$. We let $\mu$ be a probability measure on ${\mathbb{R}}$ with finite first moment. It is clear that we may then assume (subject to a suitable shift of the martingale) that the measure $\mu$ is centred. We also exclude the trivial case where $\mu = \delta_0$ from our arguments, so necessarily $\mu((-\infty,0))$ and $\mu((0,\infty))$ are both strictly positive. We write $M\in {\mathcal{M}}_\mu$ to denote that $M$ is a continuous uniformly integrable martingale with $M_\infty\sim \mu$. In the arguments below, we will commonly want to discuss the measure $\mu$ restricted to some interval. Moreover, in the case where there is an atom of $\mu$ at a point $y$, it may become necessary to split the atom into more than one part. It will be convenient therefore to split the measure $\mu$ according to its quantiles. We therefore introduce the notation $F(x) = \mu((-\infty,x])$ for the usual distribution function of the measure $\mu$, and write $F^{-1}(q) = \inf \{x \in {\mathbb{R}}: F(x) \ge q\}\vee {\ell_{\mu}}$. Then for $p,q \in [0,1]$ with $p \le q$ we define the sub-probability measures $$\label{eq:2} \mu_p^q((-\infty,x]) = (F(x) \wedge q - p) \vee 0 =: F_p^q(x).$$ In addition, we will write $\mu^q = \mu^q_0$ and $\mu_p = \mu_p^1$. Observe that $\mu_p^q({\mathbb{R}}) = q-p$. The [*barycentre*]{} of $\mu$ associates to a non-empty Borel set $\Gamma\subset {\mathbb{R}}$ the mean of $\mu$ over $\Gamma$ via $$\label{eq:barycentre} \mu_B(\Gamma)=\frac{\int_\Gamma u\, \mu({\mathrm{d}}u)}{\int_\Gamma \mu({\mathrm{d}}u)}.$$ An obvious extension is to consider the barycentre of the measure $\mu$ when restricted to $\mu_p^q$, which we denote by $m_p^q$, so $$\label{eq:3} m_p^q = \begin{cases} (q-p)^{-1}\int x \, \mu_p^q({\mathrm{d}x}) & \text{ if } q>p\\ F^{-1}(q) & \text{ otherwise} \end{cases}.$$ Now fix ${\ensuremath{\underline{b}}}, {\ensuremath{\overline{b}}}\in {\mathbb{R}}$ with ${\ensuremath{\underline{b}}}< 0 < {\ensuremath{\overline{b}}}$. Of importance in our constructions will be the following notions. Given $p$ with $p\le F({\ensuremath{\underline{b}}}-)$, we want to find the probability $q$ such that $m_p^q = {\ensuremath{\underline{b}}}$. Specifically, define a function $\rho_-: [0,F({\ensuremath{\underline{b}}}-)] \to [F({\ensuremath{\underline{b}}}),1]$ by $$\label{eq:4} \rho_-(p) = \inf \{ q \ge F({\ensuremath{\underline{b}}}) : m_p^q \ge {\ensuremath{\underline{b}}}\}.$$ Similarly, we can define $\rho_+: [F({\ensuremath{\overline{b}}}),1] \to [0,F({\ensuremath{\overline{b}}}-)]$ by $$\label{eq:5} \rho_+(q) = \sup \{ p \le F({\ensuremath{\overline{b}}}-) : m_p^q \le {\ensuremath{\overline{b}}}\}.$$ It is straightforward to see that $\rho_-(p)$ and $\rho_+(q)$ are both continuous, strictly decreasing functions, and are well defined since ${\ensuremath{\underline{b}}}< 0 = \int x \, \mu({\mathrm{d}x}) < {\ensuremath{\overline{b}}}$, so that the infimum in and the supremum in are both over non-empty sets. Further, note that we get: $$\label{eq:6} m_p^{\rho_-(p)} = {\ensuremath{\underline{b}}}, m_{\rho_+(q)}^q = {\ensuremath{\overline{b}}}$$ for all $p \le F({\ensuremath{\underline{b}}}-)$ and all $q \ge F({\ensuremath{\overline{b}}})$. Observe that the barycentre has two nice properties: first, if we rescale the measure $\mu$ by a constant, then the barycentre is unchanged. Second, if we wish to show that a measure $\mu$ has barycentre $b$, it is sufficient to show that $$\int (x-b) \, \mu(dx) = 0,$$ independent of whether $\mu$ is a probability measure. In the case where $\mu$ is a probability measure $\mu_{B}({\mathbb{R}})$ is just the mean of the measure. Finally, we introduce the additional useful notation $${\tilde{m}}_{p}^q = (q-p) m_p^q.$$ Since the functions $\rho_+$ and $\rho_-$ are both continuous and strictly decreasing, their inverses are also continuous and strictly decreasing where defined — for example, $\rho_+^{-1}$ maps $[\rho_+(1),F({\ensuremath{\overline{b}}}-)] \to [F({\ensuremath{\overline{b}}}),1]$. A critical role in the construction of embeddings will be played by the following definition. Set $$\label{eq:1} {\pi^*}= \inf \left\{ p \in [\rho_+(1)\vee F({\ensuremath{\underline{b}}}),F({\ensuremath{\overline{b}}}-)] : \rho_+^{-1}(p)-p \le \frac{-{\ensuremath{\underline{b}}}}{{\ensuremath{\overline{b}}}-{\ensuremath{\underline{b}}}}\right\} \wedge F({\ensuremath{\overline{b}}}-),$$ where we use the standard convention that the infimum of an empty set is $\infty$. Since $\rho_+^{-1}(F({\ensuremath{\overline{b}}}-))= F({\ensuremath{\overline{b}}})$, $\rho_+^{-1}(p)$ is continuous and ${\ensuremath{\underline{b}}}< 0$, it follows that ${\pi^*}\in [\rho_+(1)\vee F({\ensuremath{\underline{b}}}),F({\ensuremath{\overline{b}}}-)]$. Then we have the following theorem. \[thm:upper\_price\_mixed\](Upper bound) The bound in is sharp. More precisely, let $\mu$ be a given centred probability measure on ${\mathbb{R}}$. Then exactly one of the following is true 1. ‘${\ensuremath{\underline{b}}}\ll 0<{\ensuremath{\overline{b}}}$’: we have ${\pi^*}=F({\ensuremath{\underline{b}}})$ and $\rho_+^{-1}({\pi^*}) -{\pi^*}< -{\ensuremath{\underline{b}}}({\ensuremath{\overline{b}}}-{\ensuremath{\underline{b}}})^{-1}$.\ Then there is a martingale $M\in {\mathcal{M}}_\mu$ such that $${{\mathbb{P}}}({{\overline{M}}_{\infty} \ge {\ensuremath{\overline{b}}}, {\underline{M}}_{\infty} > {\ensuremath{\underline{b}}}}) ={{\mathbb{E}}\left[ {\overline{G}}^I(z^*) \right]},$$ where ${\overline{G}}^I$ is given by evaluated on paths of $M$, and $z^* = F^{-1}(\xi)$ where $\xi$ solves $$\label{eqn:xidefn} \int (x-{\ensuremath{\underline{b}}}) \, \mu_{F({\ensuremath{\underline{b}}})}^\xi = -{\ensuremath{\underline{b}}}.$$ 2. ‘${\ensuremath{\underline{b}}}<0<{\ensuremath{\overline{b}}}$’: we have $\rho_+^{-1}({\pi^*}) -{\pi^*}\ge -{\ensuremath{\underline{b}}}({\ensuremath{\overline{b}}}-{\ensuremath{\underline{b}}})^{-1}$.\ Then there is a martingale $M\in {\mathcal{M}}_\mu$ such that $${{\mathbb{P}}}({{\overline{M}}_{\infty} \ge {\ensuremath{\overline{b}}}, {\underline{M}}_{\infty} > {\ensuremath{\underline{b}}}}) ={{\mathbb{E}}\left[ {\overline{G}}^{II} \right]} = -{\ensuremath{\underline{b}}}({\ensuremath{\overline{b}}}-{\ensuremath{\underline{b}}})^{-1},$$ where ${\overline{G}}^{II}$ is given by – evaluated on paths of $M$. 3. ‘${\ensuremath{\underline{b}}}< 0 \ll {\ensuremath{\overline{b}}}$’: we have ${\pi^*}= \rho_+(1)$ and $1-{\pi^*}< -{\ensuremath{\underline{b}}}({\ensuremath{\overline{b}}}-{\ensuremath{\underline{b}}})^{-1}$.\ Then there is a martingale $M\in {\mathcal{M}}_\mu$ such that $${{\mathbb{P}}}({{\overline{M}}_{\infty} \ge {\ensuremath{\overline{b}}}, {\underline{M}}_{\infty} > {\ensuremath{\underline{b}}}}) = {{\mathbb{E}}\left[ {\overline{G}}^{III}(F^{-1}({\pi^*})) \right]},$$ where ${\overline{G}}^{III}$ is given by evaluated on paths of $M$. In a similar manner to Proposition \[prop:prob\_upperbound\], the pathwise inequalities described in Section \[sec:dtnt\_sub\] instantly imply a lower bound on the double exit/no-exit probabilities: \[prop:dtnt\_lowerbound\] Suppose $M=(M_{t}: 0\leq t\leq \infty)$ is a continuous uniformly integrable martingale. Then $$\label{eq:generalbound_dbmp2} {{\mathbb{P}}}\left( {{\overline{M}}_{\infty} \ge {\ensuremath{\overline{b}}}, {\underline{M}}_{\infty} > {\ensuremath{\underline{b}}}}\right) \geq \sup\left\{0,{{\mathbb{E}}\left[ {\underline{G}}_{II}(K_1',K_2) \right]},{{\mathbb{E}}\left[ {\underline{G}}_{III}(K_1,K_2) \right]}\right\},$$ where the supremum is taken over ${\ensuremath{\underline{b}}}<K_2<K_1<{\ensuremath{\overline{b}}}<K_1'$ and where ${\underline{G}}_{II},{\underline{G}}_{III}$ are given by , and , respectively, evaluated on paths of $M$. We proceed to show that this lower bound is optimal. Write $$\label{eq:16} \gamma = 1-F({\ensuremath{\overline{b}}}-) + F({\ensuremath{\underline{b}}}),$$ and consider the condition $$\label{eq:7} {\tilde{m}}_{F({\ensuremath{\underline{b}}})}^{F({\ensuremath{\overline{b}}}-)} + \gamma {\ensuremath{\underline{b}}}\ge 0.$$ If this holds, then we can find $\lambda \in (F({\ensuremath{\underline{b}}}),F({\ensuremath{\overline{b}}}-)]$ such that $$\label{eq:9} {\tilde{m}}_{F({\ensuremath{\underline{b}}})}^{\lambda} + (1-\lambda + F({\ensuremath{\underline{b}}})) {\ensuremath{\underline{b}}}= 0$$ since the left-hand side is increasing in $\lambda$ and runs between ${\ensuremath{\underline{b}}}$ and a term which is positive by . If fails, we can imagine moving mass from an atom at ${\ensuremath{\underline{b}}}$, to the right, in the process moving the average of the mass upwards. In this case, consider the condition $$\label{eq:8} {\tilde{m}}_{F({\ensuremath{\underline{b}}})}^{F({\ensuremath{\overline{b}}}-)} + \gamma {\ensuremath{\overline{b}}}\le 0.$$ If fails, and holds, then we set $\xi = F({\ensuremath{\underline{b}}})$ and we can find $\lambda \in (0,\gamma]$ such that $$\label{eq:10} {\tilde{m}}_{F({\ensuremath{\underline{b}}})}^{F({\ensuremath{\overline{b}}}-)} + \lambda {\ensuremath{\underline{b}}}+ (\gamma-\lambda) {\ensuremath{\overline{b}}}= 0.$$ Given such a $\lambda$, we will show that there exists ${\pi^*}\in [F({\ensuremath{\overline{b}}}-),1)$ such that $$\label{eq:11} {\tilde{m}}^{\xi} + {\tilde{m}}_{F({\ensuremath{\overline{b}}}-)}^{{\pi^*}} = {\ensuremath{\underline{b}}}( \xi + {\pi^*}- F({\ensuremath{\overline{b}}}-)).$$ If also fails, and $$\label{eq:12} \text{either } \rho_-(0) \ge F({\ensuremath{\overline{b}}}-) \text{ or } \rho_-(0) < F({\ensuremath{\overline{b}}}-) \text{ and } {\tilde{m}}_{\rho_-(0)}^{F({\ensuremath{\overline{b}}}-)} + {\ensuremath{\overline{b}}}( 1-F({\ensuremath{\overline{b}}}-) + \rho_-(0)) >0$$ then there exists $\xi \in (F({\ensuremath{\underline{b}}}),\rho_-(0)\wedge F({\ensuremath{\overline{b}}}-))$ such that $$\label{eq:13} {\tilde{m}}_{\xi}^{F({\ensuremath{\overline{b}}}-)} + {\ensuremath{\overline{b}}}(1-F({\ensuremath{\overline{b}}}-) + \xi) = 0.$$ Then we define ${\pi^*}$ as the solution to again. Finally, if , and all fail, then there exists ${\pi^*}\in [\rho_-(0),F({\ensuremath{\overline{b}}}-))$ such that $$\label{eq:14} {\tilde{m}}_{{\pi^*}}^{F({\ensuremath{\overline{b}}}-)} + {\ensuremath{\overline{b}}}(1-F({\ensuremath{\overline{b}}}-) + {\pi^*}) =0.$$ \[thm:lower\_price\_mixed\](Lower bound) The bound in is sharp. More precisely, let $\mu$ be a given centred probability measure on ${\mathbb{R}}$. Then exactly one of the following is true: 1. ‘${\ensuremath{\underline{b}}}< 0 \ll {\ensuremath{\overline{b}}}$’: condition holds.\ Then there is a martingale $M\in {\mathcal{M}}_\mu$ such that ${{\mathbb{P}}}({{\overline{M}}_{\infty} \ge {\ensuremath{\overline{b}}}, {\underline{M}}_{\infty} > {\ensuremath{\underline{b}}}}) = 0 = {{\mathbb{E}}\left[ {\underline{G}}_{I} \right]}$. 2. ‘${\ensuremath{\underline{b}}}<0<{\ensuremath{\overline{b}}}$’: condition fails, and either holds or fails and holds.\ Then there is a martingale $M\in {\mathcal{M}}_\mu$ such that $$\label{eq:15} {{\mathbb{P}}}({{\overline{M}}_{\infty} \ge {\ensuremath{\overline{b}}}, {\underline{M}}_{\infty} > {\ensuremath{\underline{b}}}}) = {{\mathbb{E}}\left[ {\underline{G}}_{II}({\pi^*},\xi) \right]},$$ where ${\underline{G}}_{II}$ is given via and , evaluated on paths of $M$, and ${\pi^*}$ solves . 3. ‘${\ensuremath{\underline{b}}}\ll 0<{\ensuremath{\overline{b}}}$’: conditions , and fail.\ Then there is a martingale $M\in {\mathcal{M}}_\mu$ such that $$\label{eq:17} {{\mathbb{P}}}({{\overline{M}}_{\infty} \ge {\ensuremath{\overline{b}}}, {\underline{M}}_{\infty} > {\ensuremath{\underline{b}}}}) = {{\mathbb{E}}\left[ {\underline{G}}_{III}({\pi^*},\rho_-(0)) \right]}$$ where ${\underline{G}}_{II}$ is given via and , evaluated on paths of $M$, and ${\pi^*}$ is given by . Throughout the paper, we have assumed that $(M_t)_{t \ge 0}$ has continuous paths. This assumption can be relaxed. It is relatively simple to see that if we only assume that barriers ${\ensuremath{\underline{b}}},{\ensuremath{\overline{b}}}$ are crossed in a continuous manner then all of our results remain true. If we only assume that $(M_t)$ has càdlàg paths then the situation is more complex. The optimal behaviour will essentially be as before, but we can use jumps to hide some of the occasions where a barrier is hit. More precisely, consider the continuous martingale $M$ given in Theorem \[thm:upper\_price\_mixed\] and, for ${\varepsilon}> 0$, consider the time-change: $$\rho^{\varepsilon}_t = \inf\{u\ge t : M_u \in [{\ensuremath{\underline{b}}}+{\varepsilon},\infty)\}.$$ Then $N_t = M_{\rho^{\varepsilon}_t}$ is a UI martingale which excludes paths of $M_t$ where the minimum goes below ${\ensuremath{\underline{b}}}+{\varepsilon}$, but which later return above ${\ensuremath{\underline{b}}}+{\varepsilon}$. In general, any possible martingale $M_t$ can be improved by performing such an operation, and so this suggests that an optimal discontinuous model can be chosen in such a manner that it is continuous on $[{\ensuremath{\underline{b}}}+{\varepsilon},\infty)$ and only takes values on $(-\infty,{\ensuremath{\underline{b}}}]$ if it is the final value of the martingale. This observation can be used as a starting point for an analysis similar to that given above to determine the optimal martingale models for a given measure. We do not pursue the details here. Proofs that the bounds are sharp via new solutions to the Skorokhod embedding problem {#sec:proofs} ===================================================================================== In this section we prove Theorems \[thm:upper\_price\_mixed\] and \[thm:lower\_price\_mixed\]. We do this by constructing new solutions to the Skorokhod embedding problem for a Brownian motion $B$. Specifically, we will construct stopping times $\tau$ such that $B_{\tau} \sim \mu$, $(B_{t \wedge \tau}:t\geq 0)$ is UI and equalities are attained almost surely in the inequalities of Sections \[sec:pathwise\]–\[sec:dtnt\_sub\]. It is then straightforward to see that martingales required in Theorems \[thm:upper\_price\_mixed\] and \[thm:lower\_price\_mixed\] are given by $M_t:= B_{t \wedge \tau}$. We will use below some well known facts about the existence of Skorokhod embeddings. Specifically, given a measure $\mu$ with mean $m$ and a Brownian motion $B$ with $B_0 = m$, then there exists a stopping time $\tau$ such that $B_{\tau} \sim \mu$ and $(B_{t \wedge \tau}: t \geq 0)$ is uniformly integrable. Moreover, it follows from uniform integrability that if the measure $\mu$ is supported on a bounded interval, then the process will stop before the first exit time of the interval. We take $B=(B_t:t\geq 0)$ a standard real-valued Brownian motion. All the hitting times $H_{\bullet}$ below are for $B$. As described above, we will prove this result by constructing a stopping time $\tau$ such that $B_{\tau}$ has the distribution $\mu$, and such that the conjectured bounds hold for the corresponding continuous time martingale which is the stopped process. From the definition of ${\pi^*}$ in it is clear that at least one of the cases holds. Clearly excludes the other two. To show that and are exclusive, as $\rho_+^{-1}(\rho_+(1))=1$, it suffices to argue that the following is impossible $$\label{eqn:imposs_for_y} {\pi^*}= \rho_+(1) = F({\ensuremath{\underline{b}}}) > {\ensuremath{\overline{b}}}({\ensuremath{\overline{b}}}-{\ensuremath{\underline{b}}})^{-1}.$$ Assume holds. From the last condition we get ${\ensuremath{\overline{b}}}(1-{\pi^*}) < - {\ensuremath{\underline{b}}}{\pi^*}$, and using the fact that ${\pi^*}= \rho_+(1)$, this can be expressed as $\int x \,\mu_{{\pi^*}}({\mathrm{d}x}) + {\ensuremath{\underline{b}}}{\pi^*}< 0$. However ${\pi^*}\ge F({\ensuremath{\underline{b}}})$ implies that this is greater than or equal to $\int x \, \mu({\mathrm{d}x}) = 0$ giving a contradiction. We conclude that the cases , and are exclusive. We now show the existence of a suitable embedding. We consider initially the case . We first note that the solution $\xi$ of is in $(\rho_+^{-1}({\pi^*}),1]$. Since $$\begin{aligned} \int (x - {\ensuremath{\underline{b}}}) \, \mu_{F({\ensuremath{\underline{b}}})}^{\rho_+^{-1}(F({\ensuremath{\underline{b}}}))}({\mathrm{d}x}) & = & \int (x - {\ensuremath{\overline{b}}}) \mu_{F({\ensuremath{\underline{b}}})}^{\rho_+^{-1}(F({\ensuremath{\underline{b}}}))}({\mathrm{d}x}) + \int ({\ensuremath{\overline{b}}}- {\ensuremath{\underline{b}}}) \, \mu_{F({\ensuremath{\underline{b}}})}^{\rho_+^{-1}(F({\ensuremath{\underline{b}}}))}({\mathrm{d}x}) \\ & = & ({\ensuremath{\overline{b}}}- {\ensuremath{\underline{b}}})\left( \rho_+^{-1}(F({\ensuremath{\underline{b}}}))-F({\ensuremath{\underline{b}}})\right) \\ & < & - {\ensuremath{\underline{b}}}, \end{aligned}$$ we conclude that $\xi > \rho_+^{-1}({\pi^*})$. To see that $\xi \le 1$, we note: $$\int (x - {\ensuremath{\underline{b}}}) \, \mu_{F({\ensuremath{\underline{b}}})} ({\mathrm{d}x}) \ge \int (x - {\ensuremath{\underline{b}}}) \, \mu ({\mathrm{d}x}) = -{\ensuremath{\underline{b}}}.$$ Since the expression $\int (x-{\ensuremath{\underline{b}}}) \, \mu_{F({\ensuremath{\underline{b}}})}^\xi({\mathrm{d}x})$ is strictly increasing and continuous in $\xi$, there is a unique $\xi$. For this value of $\xi$, we now define a measure $\nu$ by $$\nu = \left[ -\frac{{\ensuremath{\underline{b}}}}{{\ensuremath{\overline{b}}}-{\ensuremath{\underline{b}}}} - (\xi - F({\ensuremath{\underline{b}}}))\right] \delta_{{\ensuremath{\underline{b}}}} + \mu_{F({\ensuremath{\underline{b}}})}^\xi.$$ Observe that the atom at ${\ensuremath{\underline{b}}}$ has mass greater than or equal to zero, and by construction, $\nu$ has total mass $-{\ensuremath{\underline{b}}}({\ensuremath{\overline{b}}}-{\ensuremath{\underline{b}}})^{-1}$ and barycentre ${\ensuremath{\overline{b}}}$ since $$\begin{aligned} \int (x-{\ensuremath{\overline{b}}}) \, \nu({\mathrm{d}x}) & = & \int (x-{\ensuremath{\overline{b}}}) \, \mu_{F({\ensuremath{\underline{b}}})}^\xi({\mathrm{d}x}) + \left[-\frac{{\ensuremath{\underline{b}}}}{{\ensuremath{\overline{b}}}-{\ensuremath{\underline{b}}}} - (\xi - F({\ensuremath{\underline{b}}}))\right] ({\ensuremath{\underline{b}}}-{\ensuremath{\overline{b}}})\\ & = & ({\ensuremath{\underline{b}}}-{\ensuremath{\overline{b}}})(\xi-F({\ensuremath{\underline{b}}})) - {\ensuremath{\underline{b}}}+ \left[-\frac{{\ensuremath{\underline{b}}}}{{\ensuremath{\overline{b}}}-{\ensuremath{\underline{b}}}} - (\xi - F({\ensuremath{\underline{b}}}))\right] ({\ensuremath{\underline{b}}}-{\ensuremath{\overline{b}}})\\ & = & 0. \end{aligned}$$ We now show that this means we can construct a suitable embedding. The idea will be initially to run until the first time we hit either of ${\ensuremath{\overline{b}}}$ or ${\ensuremath{\underline{b}}}$. The mass that hits ${\ensuremath{\overline{b}}}$ first will then be used to embed $\nu$, and all the mass that hits ${\ensuremath{\underline{b}}}$ (which will include the atomic term from $\nu$) can then be embedded in the remaining areas, $(0,{\ensuremath{\underline{b}}}] \cup [F^{-1}(\xi),\infty)$. So suppose we are in case , and let $\tau_1$ be first time we hit one of ${\ensuremath{\underline{b}}}$ or ${\ensuremath{\overline{b}}}$, so $\tau_1 = H_{{\ensuremath{\overline{b}}}} \wedge H_{{\ensuremath{\underline{b}}}}$. Then ${{\mathbb{P}}}(B_{\tau_1} = {\ensuremath{\overline{b}}}) = -{\ensuremath{\underline{b}}}({\ensuremath{\overline{b}}}-{\ensuremath{\underline{b}}})^{-1}$. Let $\tau_2$ be a UI embedding of the probability measure $-\frac{{\ensuremath{\overline{b}}}-{\ensuremath{\underline{b}}}}{{\ensuremath{\underline{b}}}}\nu$ given $B_0 = {\ensuremath{\overline{b}}}$ and let $\tau_3$ be a UI embedding of $\sigma$ given $B_0 = {\ensuremath{\underline{b}}}$, where $$\sigma = \frac{\left(\mu^{F({\ensuremath{\underline{b}}})}+\mu_{\xi}\right)}{F({\ensuremath{\underline{b}}})+1-\xi}.$$ It can be verified that $\sigma$ has barycentre ${\ensuremath{\underline{b}}}$ since $$\int (x-{\ensuremath{\underline{b}}}) \left(\mu^{F({\ensuremath{\underline{b}}})}+\mu_{\xi}\right)({\mathrm{d}x}) = \int (x-{\ensuremath{\underline{b}}}) \, \mu({\mathrm{d}x}) - \int (x-{\ensuremath{\underline{b}}}) \, \mu_{F({\ensuremath{\underline{b}}})}^\xi({\mathrm{d}x}) = 0.$$ Then (recalling the definition in Section \[sec:notation\]) we set $$\begin{aligned} \tau & := & \tau_2 \circ \tau_1{\boldsymbol{1}_{\{\ensuremath{\tau_1 = H_{{\ensuremath{\overline{b}}}}}\}}} {\boldsymbol{1}_{\{\ensuremath{\tau_2 \circ \tau_1 < H_{{\ensuremath{\underline{b}}}}}\}}}\\ && {} + \tau_3 \circ \tau_1 {\boldsymbol{1}_{\{\ensuremath{\tau_1 = H_{{\ensuremath{\underline{b}}}}}\}}}\\ && {} + \tau_3 \circ \tau_2 \circ \tau_1 {\boldsymbol{1}_{\{\ensuremath{\tau_1 = H_{{\ensuremath{\overline{b}}}}}\}}} {\boldsymbol{1}_{\{\ensuremath{\tau_2 \circ \tau_1 = H_{{\ensuremath{\underline{b}}}}}\}}}. \end{aligned}$$ We see that $\tau$ is a UI embedding of $\mu$, and moreover $\tau$ is such that $ {{\boldsymbol{1}_{\{\ensuremath{\overline{B}_\tau\geq {\ensuremath{\overline{b}}},\, \underline{B}_\tau> {\ensuremath{\underline{b}}}}\}}}}= {\overline{G}}^I(F^{-1}(\xi))$ a.s.. Consider now case . Suppose initially that in addition, $\rho_+^{-1}({\pi^*})-{\pi^*}= -{\ensuremath{\underline{b}}}({\ensuremath{\overline{b}}}-{\ensuremath{\underline{b}}})^{-1}$. We define measures $\nu$ and $\sigma$ by: $$\begin{aligned} \nu & = & \frac{1}{\rho_+^{-1}({\pi^*})-{\pi^*}} \mu_{{\pi^*}}^{\rho_+^{-1}({\pi^*})}\\ \sigma & = & \frac{1}{1+{\pi^*}-\rho_+^{-1}({\pi^*})} \left(\mu^{{\pi^*}} + \mu_{\rho_+^{-1}({\pi^*})}\right). \end{aligned}$$ Then $\nu$ has barycentre ${\ensuremath{\overline{b}}}$, while $\sigma$ has barycentre ${\ensuremath{\underline{b}}}$. Let $\tau_1$ be as above, $\tau_2$ be a UI embedding of $\nu$ given $B_0 = {\ensuremath{\overline{b}}}$ and $\tau_3$ be a UI embedding of $\sigma$ given $B_0 = {\ensuremath{\underline{b}}}$. Then the stopping time $$\begin{aligned} \tau & := & \tau_2 \circ \tau_1 {\boldsymbol{1}_{\{\ensuremath{\tau_1 = H_{{\ensuremath{\overline{b}}}}}\}}} \\ && {} + \tau_3 \circ \tau_1 {\boldsymbol{1}_{\{\ensuremath{\tau_1 = H_{{\ensuremath{\underline{b}}}}}\}}} \end{aligned}$$ is a UI embedding of $\mu$, and $B_{t\wedge \tau}$ satisfies ${{\boldsymbol{1}_{\{\ensuremath{\overline{B}_\tau\geq {\ensuremath{\overline{b}}},\, \underline{B}_\tau> {\ensuremath{\underline{b}}}}\}}}}={\overline{G}}^{II}$ [a.s.]{} where ${\overline{G}}^{II}$ is the random variable defined in , evaluated on paths of $B$. The case where $\rho_+^{-1}({\pi^*})-{\pi^*}> -{\ensuremath{\underline{b}}}({\ensuremath{\overline{b}}}-{\ensuremath{\underline{b}}})^{-1}$ is almost identical — observe that in this case, there must be an atom of $\mu$ at ${\ensuremath{\overline{b}}}$ with $F({\ensuremath{\overline{b}}})-F({\ensuremath{\overline{b}}}-) > -{\ensuremath{\underline{b}}}({\ensuremath{\overline{b}}}-{\ensuremath{\underline{b}}})^{-1}$. However, the argument above works without alteration if we take: $$\begin{aligned} \nu & = & \frac{1}{-{\ensuremath{\underline{b}}}({\ensuremath{\overline{b}}}-{\ensuremath{\underline{b}}})^{-1}} \delta_{{\ensuremath{\overline{b}}}}\\ \sigma & = & \frac{1}{1+ {\ensuremath{\underline{b}}}({\ensuremath{\overline{b}}}-{\ensuremath{\underline{b}}})^{-1}}(\mu-\nu). \end{aligned}$$ Finally we consider . Then define measures $\nu$ and $\sigma$ by: $$\begin{aligned} \nu & = & \frac{1}{1-{\pi^*}} \mu_{{\pi^*}}\\ \sigma & = & \frac{1}{{\pi^*}}\mu^{{\pi^*}}. \end{aligned}$$ So the barycentre of $\nu$ is ${\ensuremath{\overline{b}}}$, and the barycentre of $\sigma$ is $m^{{\pi^*}}$. Define $\tau_1$ to be the first hitting time of $\{m^{{\pi^*}},{\ensuremath{\overline{b}}}\}$, so $\tau_1 = H_{m^{{\pi^*}}}\wedge H_{{\ensuremath{\overline{b}}}}$, then ${{\mathbb{P}}}(B_{\tau_1} = {\ensuremath{\overline{b}}}) = {\pi^*}=- m^{{\pi^*}}({\ensuremath{\overline{b}}}- m^{{\pi^*}})^{-1}$. We may then proceed as above, so we define $\tau_2$ to be a UI embedding of $\nu$ given $B_0 = {\ensuremath{\overline{b}}}$ and $\tau_3$ to be a UI embedding of $\sigma$ given $B_0 = m^{{\pi^*}}$. Then the stopping time $$\begin{aligned} \tau & := & \tau_2 \circ \tau_1 {\boldsymbol{1}_{\{\ensuremath{\tau_1 = H_{{\ensuremath{\overline{b}}}}}\}}} \\ && {} + \tau_3 \circ \tau_1 {\boldsymbol{1}_{\{\ensuremath{\tau_1 = H_{m^{{\pi^*}}}}\}}} \end{aligned}$$ is a UI embedding of $\mu$, and satisfies ${{\boldsymbol{1}_{\{\ensuremath{\overline{B}_\tau\geq {\ensuremath{\overline{b}}},\, \underline{B}_\tau> {\ensuremath{\underline{b}}}}\}}}}={\overline{G}}^{III}(F^{-1}({\pi^*}))$ [a.s.]{} where ${\overline{G}}^{III}(\cdot)$ is the random variable defined in , evaluated on paths of $B$. The setup, and general methodology, is analogous to the proof of Theorem \[thm:upper\_price\_mixed\] above. It follows from their respective definitions that exactly one of , and holds. Suppose holds, so that is true. Then, by continuity, there exists $\lambda\in (F({\ensuremath{\underline{b}}}),F({\ensuremath{\overline{b}}}-)]$ such that holds (taking $\lambda = F({\ensuremath{\underline{b}}})$ gives ${\ensuremath{\underline{b}}}$ on the left hand side of ). Let $\tau_1$ be a UI embedding of $$\label{eq:18} \chi = \mu_{F({\ensuremath{\underline{b}}})}^\lambda + (1-\lambda + F({\ensuremath{\underline{b}}}))\delta_{{\ensuremath{\underline{b}}}}$$ in the Brownian motion starting at 0, and observe that the measure $$\nu = \frac{\mu^{F({\ensuremath{\underline{b}}})} + \mu_{\lambda}}{1-\lambda + F({\ensuremath{\underline{b}}})}$$ has mean ${\ensuremath{\underline{b}}}$, which follows since: $$\begin{aligned} (1-\lambda + F({\ensuremath{\underline{b}}})) \int x \nu({\mathrm{d}x}) & = {\tilde{m}}^{F({\ensuremath{\underline{b}}})} + {\tilde{m}}_{\lambda}\\ & = -{\tilde{m}}_{F({\ensuremath{\underline{b}}})}^{\lambda} = {\ensuremath{\underline{b}}}(1-\lambda + F({\ensuremath{\underline{b}}})). \end{aligned}$$ Let $\tau_2$ be a UI embedding of $\nu$ in a Brownian motion starting from $B_0={\ensuremath{\underline{b}}}$. Finally define $$\tau:=\tau_1{\boldsymbol{1}_{\{\ensuremath{B_{\tau_1}\neq {\ensuremath{\underline{b}}}}\}}}+\tau_2\circ\tau_1{\boldsymbol{1}_{\{\ensuremath{B_{\tau_1}={\ensuremath{\underline{b}}}}\}}},$$ which is a UI embedding of $\mu$ in the Brownian motion $B$. Note that $\overline{B}_\tau\geq {\ensuremath{\overline{b}}}$ only if $\underline{B}_\tau\leq {\ensuremath{\underline{b}}}$. It follows that ${{\boldsymbol{1}_{\{\ensuremath{\overline{B}_\tau\geq {\ensuremath{\overline{b}}},\, \underline{B}_\tau> {\ensuremath{\underline{b}}}}\}}}}=0={\underline{G}}_{I}$ a.s. Suppose now that holds. We consider separately the case where fails and holds, and the case where both and fail, but holds. First suppose holds. Then $$\lambda \mapsto {\tilde{m}}_{F({\ensuremath{\underline{b}}})}^{F({\ensuremath{\overline{b}}}-)} + \lambda {\ensuremath{\underline{b}}}+ (\gamma - \lambda) {\ensuremath{\overline{b}}}$$ is continuous, and strictly negative for $\lambda = 0$ and positive for $\lambda = \gamma$. Hence there exists $\lambda \in (0,\gamma]$ such that holds. Fix $\xi = F({\ensuremath{\underline{b}}})$ and consider $$[F({\ensuremath{\overline{b}}}-),1) \ni {\pi^*}\mapsto {\tilde{m}}^{\xi} + {\tilde{m}}_{F({\ensuremath{\overline{b}}}-)}^{{\pi^*}} - {\ensuremath{\underline{b}}}( \xi + {\pi^*}- F({\ensuremath{\overline{b}}}-)).$$ In the limit as ${\pi^*}\to 1$, the expression simplifies to $-{\tilde{m}}^{F({\ensuremath{\overline{b}}}-)}_{F({\ensuremath{\underline{b}}})}-\gamma{\ensuremath{\underline{b}}}$ which is strictly positive since is assumed to fail, while if ${\pi^*}= F({\ensuremath{\overline{b}}}-)$ the expression simplifies to ${\tilde{m}}^{F({\ensuremath{\underline{b}}})}-{\ensuremath{\underline{b}}}F({\ensuremath{\underline{b}}})$, which is non-positive, since ${\tilde{m}}^{F({\ensuremath{\underline{b}}})} = \int x \, \mu^{F({\ensuremath{\underline{b}}})}({\mathrm{d}x}) \le \int {\ensuremath{\underline{b}}}\, \mu^{F({\ensuremath{\underline{b}}})}({\mathrm{d}x})$. Hence there is a unique ${\pi^*}$ satisfying . Now define a measure $$\chi = \mu_{F({\ensuremath{\underline{b}}})}^{F({\ensuremath{\overline{b}}}-)} + \lambda \delta_{{\ensuremath{\underline{b}}}} + (\gamma-\lambda) \delta_{{\ensuremath{\overline{b}}}}.$$ From it follows that $\chi$ is centered, and we embed this initially. The mass which arrives at ${\ensuremath{\overline{b}}}$ will then run to the measure $$\nu = \frac{(\gamma-\lambda -(1-{\pi^*}))\delta_{{\ensuremath{\underline{b}}}} + \mu_{{\pi^*}}}{\gamma-\lambda}$$ which has mean ${\ensuremath{\overline{b}}}$ by the following computation: $$\begin{aligned} (\gamma - \lambda) \int x \, \nu({\mathrm{d}x}) & = {\ensuremath{\underline{b}}}(\gamma-\lambda -(1-{\pi^*})) + {\tilde{m}}_{{\pi^*}}\\ & = {\ensuremath{\underline{b}}}(\gamma-\lambda -(1-{\pi^*})) - {\tilde{m}}_{F({\ensuremath{\overline{b}}}-)}^{{\pi^*}} - {\tilde{m}}_{F({\ensuremath{\underline{b}}})}^{F({\ensuremath{\overline{b}}}-)} - {\tilde{m}}^{F({\ensuremath{\underline{b}}})} \\ & = {\ensuremath{\underline{b}}}(\gamma-\lambda -(1-{\pi^*})) - {\ensuremath{\underline{b}}}(\xi + {\pi^*}- F({\ensuremath{\overline{b}}}-)) + \lambda {\ensuremath{\underline{b}}}+ (\gamma - \lambda) {\ensuremath{\overline{b}}}\\ & = {\ensuremath{\underline{b}}}(\gamma -1 -\xi + F({\ensuremath{\overline{b}}}-)) + {\ensuremath{\overline{b}}}(\gamma-\lambda). \end{aligned}$$ Here we have used , and the fact that $\xi = F({\ensuremath{\underline{b}}})$. From the definition of $\gamma$ in , the desired conclusion follows. Finally, we embed the remaining part of $\mu$ from the mass that finishes at ${\ensuremath{\underline{b}}}$ after either the first or second step, which has total probability $\gamma - \lambda + {\pi^*}- 1 + \lambda = \xi + {\pi^*}- F({\ensuremath{\overline{b}}}-)$. Set $$\label{eq:19} \sigma = \frac{\mu^{\xi} + \mu_{F({\ensuremath{\overline{b}}}-)}^{{\pi^*}}}{\xi + {\pi^*}- F({\ensuremath{\overline{b}}}-)},$$ and $\sigma$ has mean ${\ensuremath{\underline{b}}}$: $$\begin{aligned} (\xi + {\pi^*}- F({\ensuremath{\overline{b}}}-)) \int x \, \sigma({\mathrm{d}x}) & = {\tilde{m}}^{\xi} + {\tilde{m}}_{F({\ensuremath{\overline{b}}}-)}^{{\pi^*}}\\ & = {\ensuremath{\underline{b}}}(\xi + {\pi^*}- F({\ensuremath{\overline{b}}}-)) \end{aligned}$$ by . The final stopping time will be of the same form both in this case and in the case where holds, and when fails but holds. So before constructing the embedding, we give a description of the relevant measures in the second case. Suppose fails, but holds. Then in a similar manner to above, we can find $\xi \in (F({\ensuremath{\underline{b}}}),\rho_-(0)\wedge F({\ensuremath{\overline{b}}}-))$ such that holds. Define $$\chi = \mu_{\xi}^{F({\ensuremath{\overline{b}}}-)} + (1-F({\ensuremath{\overline{b}}}-)+\xi) \delta_{{\ensuremath{\overline{b}}}}$$ and choose ${\pi^*}$ as before as the solution to . Then set $$\nu = \frac{({\pi^*}-F({\ensuremath{\overline{b}}}-) + \xi) \delta_{{\ensuremath{\underline{b}}}} + \mu_{{\pi^*}}}{1-F({\ensuremath{\overline{b}}}-)+\xi}$$ and we verify that $\nu$ has mean ${\ensuremath{\overline{b}}}$: $$\begin{aligned} (1-F({\ensuremath{\overline{b}}}-)+\xi) \int x \, \nu({\mathrm{d}x}) & = {\ensuremath{\underline{b}}}({\pi^*}- F({\ensuremath{\overline{b}}}-)+\xi) + {\tilde{m}}_{{\pi^*}}\\ & = {\ensuremath{\underline{b}}}( {\pi^*}- F({\ensuremath{\overline{b}}}-) + \xi) - {\tilde{m}}_{F({\ensuremath{\overline{b}}}-)}^{{\pi^*}} - {\tilde{m}}_{\xi}^{F({\ensuremath{\overline{b}}}-)}-{\tilde{m}}^{\xi}\\ & = {\ensuremath{\underline{b}}}({\pi^*}-F({\ensuremath{\overline{b}}}-) + \xi) - {\ensuremath{\underline{b}}}(\xi+{\pi^*}-F({\ensuremath{\overline{b}}}-)) + {\ensuremath{\overline{b}}}(1-F({\ensuremath{\overline{b}}}-) + \xi)\\ & = {\ensuremath{\overline{b}}}(1-F({\ensuremath{\overline{b}}}-)+\xi). \end{aligned}$$ Finally, setting $\sigma$ as in we again have $\sigma$ with mean ${\ensuremath{\underline{b}}}$. In both cases, we construct an embedding as follows: let $\tau_1$ be a UI embedding of $\chi$ (starting from $0$). Then let $\tau_2$ be a UI embedding of $\nu$ (starting from ${\ensuremath{\overline{b}}}$). Finally, we let $\tau^3$ be a UI embedding of $\sigma$ (starting from ${\ensuremath{\underline{b}}}$). We then define the complete embedding by: $$\begin{split} \tau:=\ &\tau_1{\boldsymbol{1}_{\{\ensuremath{B_{\tau_1}\in ({\ensuremath{\underline{b}}},{\ensuremath{\overline{b}}})}\}}}\\ &+\tau_2\circ\tau_1{\boldsymbol{1}_{\{\ensuremath{B_{\tau_1}={\ensuremath{\overline{b}}}}\}}} {\boldsymbol{1}_{\{\ensuremath{B_{\tau_2\circ\tau_1}>{\ensuremath{\underline{b}}}}\}}}+\\ &+\tau_3\circ\Big(\tau_1{\boldsymbol{1}_{\{\ensuremath{B_{\tau_1}={\ensuremath{\underline{b}}}}\}}} +\tau_2\circ\tau_1{\boldsymbol{1}_{\{\ensuremath{B_{\tau_1}={\ensuremath{\overline{b}}}}\}}} {\boldsymbol{1}_{\{\ensuremath{B_{\tau_2\circ\tau_1}={\ensuremath{\underline{b}}}}\}}}\Big), \end{split}$$ and it follows from our construction that $\tau$ is a UI embedding of $\mu$ which moreover satisfies ${{\boldsymbol{1}_{\{\ensuremath{\overline{B}_\tau\geq {\ensuremath{\overline{b}}},\, \underline{B}_\tau> {\ensuremath{\underline{b}}}}\}}}}= {\underline{G}}_{II}({\pi^*},\xi)$. Suppose finally we are in case , so that , and all fail. Then there exists ${\pi^*}\in [\rho_-(0),F({\ensuremath{\overline{b}}}-))$ such that holds. Define the probability measure $$\chi = \mu_{{\pi^*}}^{F({\ensuremath{\overline{b}}}-)} + (1-F({\ensuremath{\overline{b}}}-)-{\pi^*}) \delta_{{\ensuremath{\overline{b}}}},$$ which has mean $0$ by the definition of ${\pi^*}$. Define also $$\nu = \frac{\rho_-(0) \delta_{{\ensuremath{\underline{b}}}} + \mu_{\rho_-(0)}^{{\pi^*}} + \mu_{F({\ensuremath{\overline{b}}}-)}}{1-F({\ensuremath{\overline{b}}}-)+{\pi^*}}$$ and we confirm that $\nu$ has mean ${\ensuremath{\overline{b}}}$: $$\begin{aligned} (1-F({\ensuremath{\overline{b}}}-)+{\pi^*}) \int x \, \nu(dx) & = {\tilde{m}}^{\rho_-(0)} + {\tilde{m}}_{\rho_-(0)}^{{\pi^*}} + {\tilde{m}}_{F({\ensuremath{\overline{b}}}-)}\\ & = {\tilde{m}}^{{\pi^*}} + {\tilde{m}}_{F({\ensuremath{\overline{b}}}-)} \\ & = -{\tilde{m}}_{{\pi^*}}^{F({\ensuremath{\overline{b}}}-)}\\ & = {\ensuremath{\overline{b}}}(1-F({\ensuremath{\overline{b}}}-)+{\pi^*}). \end{aligned}$$ Finally, any mass which is at ${\ensuremath{\underline{b}}}$ we finally embed to the measure $\sigma = (\rho_-(0))^{-1}\mu^{\rho_-(0)}$. That is, we define the stopping times $\tau_1$ which is a UI embedding of $\chi$ starting at $0$. Then let $\tau_2$ be a UI embedding of $\nu$, given initial value ${\ensuremath{\overline{b}}}$, and $\tau_3$ an embedding of $\sigma$ given initial value ${\ensuremath{\underline{b}}}$. Finally, we define $$\tau:=\tau_1{\boldsymbol{1}_{\{\ensuremath{B_{\tau_1}\neq {\ensuremath{\overline{b}}}}\}}}+\tau_2\circ\tau_1{\boldsymbol{1}_{\{\ensuremath{B_{\tau_1}={\ensuremath{\overline{b}}}}\}}} {\boldsymbol{1}_{\{\ensuremath{B_{\tau_2\circ\tau_1} > {\ensuremath{\underline{b}}}}\}}}+ \tau_3 \circ\tau_2\circ\tau_1{\boldsymbol{1}_{\{\ensuremath{B_{\tau_1}={\ensuremath{\overline{b}}}}\}}} {\boldsymbol{1}_{\{\ensuremath{B_{\tau_2\circ\tau_1} = {\ensuremath{\underline{b}}}}\}}},$$ to get a UI embedding of $\mu$ in $B$. Furthermore, it follows from the construction that ${{\boldsymbol{1}_{\{\ensuremath{\overline{B}_\tau\geq {\ensuremath{\overline{b}}},\, \underline{B}_\tau> {\ensuremath{\underline{b}}}}\}}}}= {\underline{G}}_{III}({\pi^*},\rho_-(0))$. On joint distribution of the maximum and minimum of a continuous UI martingale {#ap:mart} ============================================================================== We turn now to studying the properties of joint distribution of the maximum and minimum of a continuous UI martingale. As previously, $(M_t:0\leq t\leq \infty)$ is a uniformly integrable continuous martingale. We let $\mu$ be its terminal distribution, $\mu\sim M_\infty$, and recall that $-\infty\leq {\ell_{\mu}}<{r_{\mu}}\leq \infty$ are the bounds of the support of $\mu$, i.e.$[{\ell_{\mu}},{r_{\mu}}]$ is the smallest interval with $\mu([{\ell_{\mu}},{r_{\mu}}])=1$. Using Theorems \[thm:upper\_price\_mixed\] and \[thm:lower\_price\_mixed\] above, as well as existing results, we study the functions $$\begin{aligned} \label{eq:20} p({\ensuremath{\underline{b}}},{\ensuremath{\overline{b}}}) & = {{\mathbb{P}}}\left({\underline{M}}_\infty > {\ensuremath{\underline{b}}}\textrm{ and }{\overline{M}}_\infty <{\ensuremath{\overline{b}}}\right)\\ q({\ensuremath{\underline{b}}},{\ensuremath{\overline{b}}}) & = {{\mathbb{P}}}\left({\underline{M}}_\infty > {\ensuremath{\underline{b}}}\textrm{ and }{\overline{M}}_\infty \ge {\ensuremath{\overline{b}}}\right)\label{eq:21}\\ r({\ensuremath{\underline{b}}},{\ensuremath{\overline{b}}}) & = {{\mathbb{P}}}\left({\underline{M}}_\infty\le {\ensuremath{\underline{b}}}\textrm{ and }{\overline{M}}_\infty \ge {\ensuremath{\overline{b}}}\right)\label{eq:22}\end{aligned}$$ for ${\ensuremath{\underline{b}}}\leq 0 \leq {\ensuremath{\overline{b}}}$. Note that with no restrictions on $M_0$, when looking at extrema of the functions above, it is enough to consider $M_0$ a constant (e.g. when maximising $r$) or $M_0\equiv M_\infty$ (e.g. when minimising $r$). The latter is degenerate and henceforth we assume $M_0$ is a constant a.s. Further, as our results are translation invariant, we may and will take $M_0=0$ a.s. It follows that $\mu$ is centred. It follows from Dambis, Dubins-Schwarz Theorem that $M$ is a (continuous) time change of Brownian motion, i.e. we can write $M_t=B_{\tau_t}$, $t\leq \infty$, for some Brownian motion and an increasing family of stopping times $(\tau_t)$ with $B_{\tau_\infty}\sim M_\infty$, $(B_{t\land \tau_\infty}:t\geq 0 )$ UI and ${\overline{M}}_\infty=\overline{B}_{\tau_\infty}$, ${\underline{M}}_\infty=\underline{B}_{\tau_\infty}$. In consequence, the problem reduces to studying the maximum and minimum of Brownian motion stopped at $\tau=\tau_\infty$, which is a solution the Skorokhod embedding problem. We can deduce results about the optimal properties of the martingales from corresponding results about Skorokhod embeddings. Our first result concerns the embeddings of Perkins and the ‘tilted-Jacka’ construction, which we now recall using the notation established previously. These constructions have been considered in [@Cox:2011ab], and we will need some results from this paper; however both constructions have a long history — see for example [@Perkins:86; @Cox:2004aa; @Jacka:88; @Cox:2005aa]. For the Perkins embedding we define[^3] $$\label{eq:24} \begin{split} {{\gamma_+}}(p)=q \text{ where $q$ solves } &{\tilde{m}}^{q} + {\tilde{m}}_{p} = (1-p+q) F(p), \quad p > F(0)\\ {{\gamma_+}}(q)=p \text{ where $p$ solves } &{\tilde{m}}^{q} + {\tilde{m}}_{p} = (1-p+q) F(q), \quad q < F(0-). \end{split}$$ The stopping time $\tau_P$ is then defined via: $$\label{eq:25} \tau_P = \inf\{ t \ge 0: F(B_t) \not\in ({{\gamma_+}}(F({\ensuremath{\overline{B}}}_t)),{{\gamma_-}}(F({\ensuremath{\underline{B}}}_t)))\}.$$ In a similar spirit, the tilted-Jacka construction is given as follows. Choose ${\pi^*}\in [0,1]$ such that $({\ensuremath{\underline{b}}}-m^{{\pi^*}})(m_{{\pi^*}}-{\ensuremath{\overline{b}}})\ge 0$ — this is always possible, since we can always find ${\pi^*}$ such that $m^{{\pi^*}} = {\ensuremath{\underline{b}}}$ say. Then set $\chi = {\pi^*}\delta_{m^{{\pi^*}}} + (1-{\pi^*}) \delta_{m_{{\pi^*}}}$. The construction is as follows: we first embed the distribution $\chi$, then, given we hit $m^{{\pi^*}}$, we embed $\mu^{{\pi^*}}$ using the reversed Azéma-Yor construction (c.f. [@Obloj:04b]); if we hit $m_{{\pi^*}}$ then we embed $\mu_{{\pi^*}}$ using the Azéma-Yor construction. Finally, we observe that both cases give rise to martingales with certain optimality properties using the fact that the stopped Brownian motion is a continuous martingale. \[prop:M\_bounds\] We have the following properties: 1. $p(0,{\ensuremath{\overline{b}}})=0 = p({\ensuremath{\underline{b}}},0)$, $q(0,{\ensuremath{\overline{b}}}) = 0 = q({\ensuremath{\underline{b}}},{r_{\mu}})$ and $r({\ell_{\mu}},{\ensuremath{\overline{b}}}) = 0 = r({\ensuremath{\underline{b}}},{r_{\mu}})$; 2. $p({\ensuremath{\underline{b}}},{\ensuremath{\overline{b}}})=1$ on $[-\infty,{\ell_{\mu}})\times({r_{\mu}},\infty]$, $q({\ensuremath{\underline{b}}},{\ensuremath{\overline{b}}}) = 1$ on $[-\infty,{\ell_{\mu}}) \times \{0\}$, and $r(0,0) = 1$; 3. $p$ and $q$ are non-increasing in ${\ensuremath{\underline{b}}}\in ({\ell_{\mu}},0)$ and $p$ is non-decreasing in ${\ensuremath{\overline{b}}}\in (0,{r_{\mu}})$; $r$ is non-decreasing in ${\ensuremath{\underline{b}}}\in ({\ell_{\mu}},0)$ and $q$ and $r$ are non-decreasing in ${\ensuremath{\overline{b}}}\in (0,{r_{\mu}})$; 4. for ${\ell_{\mu}}\leq {\ensuremath{\underline{b}}}<0<{\ensuremath{\overline{b}}}\leq{r_{\mu}}$ we have $$\label{eq:uimart_bound} {{\mathbb{P}}}\big(\underline{B}_{\tau_J}>{\ensuremath{\underline{b}}}\textrm{ and }\overline{B}_{\tau_J} <{\ensuremath{\overline{b}}}\big)\leq p({\ensuremath{\underline{b}}},{\ensuremath{\overline{b}}})\leq {{\mathbb{P}}}\big(\underline{B}_{\tau_P}>{\ensuremath{\underline{b}}}\textrm{ and }\overline{B}_{\tau_P} <{\ensuremath{\overline{b}}}\big),$$ where $(B_t)$ is a standard Brownian motion with $B_0=0$, $\tau_P$ is the Perkins stopping time [@Cox:2011ab (4.4)] embedding $\mu$ and $\tau_J$ is the ‘tilted-Jacka’ stopping time [@Cox:2011ab (4.6)], for barriers $({\ensuremath{\underline{b}}},{\ensuremath{\overline{b}}})$, embedding $\mu$; 5. for ${\ell_{\mu}}\leq {\ensuremath{\underline{b}}}<0<{\ensuremath{\overline{b}}}\leq{r_{\mu}}$, the lower bound on $q({\ensuremath{\underline{b}}},{\ensuremath{\overline{b}}})$ is given by , and the upper bound is given by . Moreover these bounds are attained by the constructions in Theorems \[thm:upper\_price\_mixed\] and \[thm:lower\_price\_mixed\] respectively; 6. for ${\ell_{\mu}}\leq {\ensuremath{\underline{b}}}<0<{\ensuremath{\overline{b}}}\leq{r_{\mu}}$, the lower bound on $r({\ensuremath{\underline{b}}},{\ensuremath{\overline{b}}})$ is given by Proposition 2.3 of [@Cox:2011aa], and the upper bound is given by Proposition 2.1 of [@Cox:2011aa]. Moreover these bounds are attained by the constructions in Theorems 2.4 and 2.2 of [@Cox:2011aa] respectively. The first three assertions of the proposition are clear. Assertion $(iv)$ is a reformulation of Lemmas 4.2 and 4.3 of [@Cox:2011ab] — it suffices to note that $(B_{t\wedge\tau_J})$, $(B_{t\wedge\tau_P})$, $(M_t)$ are all UI martingales starting at $0$ and with the same terminal law $\mu$ for $t=\infty$. Likewise, part $(vi)$ is a reinterpretation of the results of [@Cox:2011aa]. We note that therein the results were formulated for the case of non-atomic $\mu$. They extend readily, with methods used in Section \[sec:proofs\] above, specifically by characterising the stopping distributions via quantiles of the underlying measures, to the general case. We can think of any of the functions $p(\cdot,\cdot), q(\cdot,\cdot)$, and $r(\cdot,\cdot)$ as a surface defined over the quarter-plane $[-\infty,0]\times [0,\infty]$. Proposition \[prop:M\_bounds\] describes boundary values of the surface, monotonicity properties and gives an upper and a lower bound on the surface. However we note that — most obviously in $(iv)$ — there is a substantial difference between the bounds linked to the fact that $\tau_P$ does not depend on $({\ensuremath{\underline{b}}},{\ensuremath{\overline{b}}})$ while $\tau_J$ does. In consequence, the upper bound is attainable: there is a martingale $(M_t)$, namely $M_t=(B_{t\wedge\tau_P})$, for which $p$ is equal to the upper bound for all $({\ensuremath{\underline{b}}},{\ensuremath{\overline{b}}})$. In contrast a martingale $(M_t)$ for which $p$ would be equal to the lower bound does not exist. For the martingale $M_t=(B_{t\wedge\tau_J})$, where $\tau_J$ is defined for some pair $({\ensuremath{\underline{b}}},{\ensuremath{\overline{b}}})$, $p$ will attain the lower bound in some neighbourhood of $({\ensuremath{\underline{b}}},{\ensuremath{\overline{b}}})$ which will be strictly contained in $({\ell_{\mu}},0)\times (0,{r_{\mu}})$. More generally, the latter case is more typical of all the constructions which are used in the result; however, with some careful construction, it seems likely that one can usually find a construction which will be optimal for all values of $({\ensuremath{\underline{b}}}, {\ensuremath{\overline{b}}})$ which lie in some small open set (for example, this is true of the tilted-Jacka construction), but there will be limits on how large the region on which a given construction is optimal can be made. We now give a result which provides some further insight into the structure of the bounds discussed above. In particular, we can show some finer properties of the functions $p,q,r$ and their upper and lower bounds. We state and prove the result for the function $p$, but the corresponding versions for $q$ and $r$ will follow in a clear manner. \[thm:boundstructure\] The function $p({\ensuremath{\underline{b}}},{\ensuremath{\overline{b}}})$ is càglàd in ${\ensuremath{\overline{b}}}$ and càdlàg in ${\ensuremath{\underline{b}}}$. Moreover, if $p$ is discontinuous at $({\ensuremath{\underline{b}}},{\ensuremath{\overline{b}}})$, then $\mu$ must have an atom at one of ${\ensuremath{\underline{b}}}$ or ${\ensuremath{\overline{b}}}$. Further: 1. if there is a discontinuity at $({\ensuremath{\underline{b}}},{\ensuremath{\overline{b}}})$ of the form: $$\limsup_{w \to {\ensuremath{\overline{b}}}} p({\ensuremath{\underline{b}}},w) > p({\ensuremath{\underline{b}}},{\ensuremath{\overline{b}}})$$ then the function $g$ defined by $$g(u) = \limsup_{w \to {\ensuremath{\overline{b}}}} p(u,w) - p(u,{\ensuremath{\overline{b}}}), \qquad u \le {\ensuremath{\underline{b}}}$$ is non-increasing. 2. if there is a discontinuity at $({\ensuremath{\underline{b}}},{\ensuremath{\overline{b}}})$ of the form: $$\limsup_{u \to {\ensuremath{\underline{b}}}} p(u,{\ensuremath{\overline{b}}}) > p({\ensuremath{\underline{b}}},{\ensuremath{\overline{b}}})$$ then the function $h$ defined by $$h(w) = \limsup_{u \to {\ensuremath{\underline{b}}}} p(u,w) - p({\ensuremath{\underline{b}}},w), \qquad w \ge {\ensuremath{\overline{b}}}$$ is non-decreasing. And, at any discontinuity, we will be in at least one of the above cases. In addition the lower bound (corresponding to the tilted-Jacka construction) is continuous in $({\ell_{\mu}},0)\times (0,{r_{\mu}})$, and continuous at the boundary (${\ensuremath{\overline{b}}}= {r_{\mu}}$ and ${\ensuremath{\underline{b}}}= {\ell_{\mu}}$) unless there is an atom of $\mu$ at either ${r_{\mu}}$ or ${\ell_{\mu}}$, while the upper bound (which corresponds to the Perkins construction) has a discontinuity corresponding to every atom of $\mu$. 1. Considering $q$ instead of $p$, the function will be càdlàg in both arguments, and the directions of the convergence results needs to be adapted suitably. We also observe that discontinuities in the upper bound occur only if there is an atom of $\mu$ at ${\ensuremath{\underline{b}}}$, [ *and*]{} we are in case of Theorem \[thm:upper\_price\_mixed\]. Similarly, there is a discontinuity in the lower bound at ${\ensuremath{\overline{b}}}$ if there is an atom of $\mu$ at ${\ensuremath{\overline{b}}}$, and we are in either of cases or of Theorem \[thm:lower\_price\_mixed\]. 2. Considering $r$ instead of $p$, the function will be càglàd in ${\ensuremath{\underline{b}}}$ and càdlàg in ${\ensuremath{\overline{b}}}$. We also observe that discontinuities in the upper bound never occur, while there are discontinuities in the lower bound at ${\ensuremath{\overline{b}}}$ and/or ${\ensuremath{\underline{b}}}$ if there is an atom of $\mu$ at either of these values. Before we prove the above result, we note the following useful result, which is a simple consequence of the martingale property: \[prop:supatom\] Suppose that $(M_t)_{t \ge 0}$ is a UI martingale with $M_\infty \sim \mu$. Then ${{\mathbb{P}}}({\overline{M}}_{\infty} = {\ensuremath{\overline{b}}}) >0$ implies $\mu(\{{\ensuremath{\overline{b}}}\}) \ge {{\mathbb{P}}}({\overline{M}}_{\infty} = {\ensuremath{\overline{b}}})$ and $$\{ {\overline{M}}_{\infty} = {\ensuremath{\overline{b}}}\} = \{ M_t = {\ensuremath{\overline{b}}}, \ \forall t \ge H_{{\ensuremath{\overline{b}}}}\} \subseteq \{ M_\infty = {\ensuremath{\overline{b}}}\} \quad {a.s.}{}.$$ We begin by noting that by definition of $p({\ensuremath{\underline{b}}},{\ensuremath{\overline{b}}})$, we necessarily have the claimed continuity and limiting properties. Further, $$\liminf_{(s,v) \to (u,w)} p(s,v) \ge {{\mathbb{P}}}({\underline{M}}_\infty > {\ensuremath{\underline{b}}}\mbox{ and } {\overline{M}}_\infty < {\ensuremath{\overline{b}}})$$ and $$\limsup_{(s,v) \to (u,w)} p(s,v) \le {{\mathbb{P}}}({\underline{M}}_\infty \ge {\ensuremath{\underline{b}}}\mbox{ and } {\overline{M}}_\infty \le {\ensuremath{\overline{b}}}).$$ It follows that the function $p$ is continuous at $({\ensuremath{\underline{b}}},{\ensuremath{\overline{b}}})$ if ${{\mathbb{P}}}({\underline{M}}_\infty = {\ensuremath{\underline{b}}}) = {{\mathbb{P}}}({\overline{M}}_\infty = {\ensuremath{\overline{b}}}) = 0$. By Proposition \[prop:supatom\], this is true when $\mu(\{{\ensuremath{\overline{b}}},{\ensuremath{\underline{b}}}\}) = 0$. Note that we can now see that at a discontinuity of $p$, we must be in at least one of the cases (i) or (ii). This is because discontinuity at $({\ensuremath{\underline{b}}},{\ensuremath{\overline{b}}})$ is equivalent to $${{\mathbb{P}}}({\underline{M}}_\infty \ge {\ensuremath{\underline{b}}}\mbox{ and } {\overline{M}}_\infty \le {\ensuremath{\overline{b}}}) > {{\mathbb{P}}}({\underline{M}}_\infty > {\ensuremath{\underline{b}}}\mbox{ and } {\overline{M}}_\infty < {\ensuremath{\overline{b}}}),$$ from which we can deduce that at least one of the events $$\{{\underline{M}}_\infty > {\ensuremath{\underline{b}}}\mbox{ and } {\overline{M}}_\infty = {\ensuremath{\overline{b}}}\}, \quad \{{\underline{M}}_\infty = {\ensuremath{\underline{b}}}\mbox{ and } {\overline{M}}_\infty < {\ensuremath{\overline{b}}}\}, \quad \{{\underline{M}}_\infty = {\ensuremath{\underline{b}}}\mbox{ and } {\overline{M}}_\infty = {\ensuremath{\overline{b}}}\}$$ is assigned positive mass. However, by Proposition \[prop:supatom\] the final event implies both $M_\infty = {\ensuremath{\underline{b}}}$ and $M_\infty = {\ensuremath{\overline{b}}}$ which is impossible. Consequently, at least one of the first two events must be assigned positive mass, and these are precisely the cases (i) and (ii). Consider now case (i). We can rewrite the statement as: if $g({\ensuremath{\underline{b}}}) > 0$, then $g(u)$ is decreasing for $u<{\ensuremath{\underline{b}}}$. Note however that $$\begin{aligned} g(u) & = & {{\mathbb{P}}}({\underline{M}}_\infty > u \mbox{ and } {\overline{M}}_\infty \le {\ensuremath{\overline{b}}}) - {{\mathbb{P}}}({\underline{M}}_\infty > u \mbox{ and } {\overline{M}}_\infty < {\ensuremath{\overline{b}}}) \\ & = & {{\mathbb{P}}}({\underline{M}}_\infty > u \mbox{ and } {\overline{M}}_\infty = {\ensuremath{\overline{b}}}) \end{aligned}$$ which is clearly non-increasing in $u$. [In fact, provided that $g({\ensuremath{\underline{b}}})<{{\mathbb{P}}}({\overline{M}}_\infty={\ensuremath{\overline{b}}})$, it follows from [e.g.]{} [@Rogers:93 Theorem 4.1] that $g$ is strictly decreasing for ${\ensuremath{\underline{b}}}>u>\sup\{u\geq -\infty: g(u)={{\mathbb{P}}}({\overline{M}}_\infty={\ensuremath{\overline{b}}})\}$. A similar proof holds in case (ii).]{} We now consider the lower bounds corresponding to the tilted-Jacka construction. We wish to show that $${{\mathbb{P}}}({\underline{M}}_\infty \ge {\ensuremath{\underline{b}}}\mbox{ and } {\overline{M}}_\infty \le {\ensuremath{\overline{b}}}) = {{\mathbb{P}}}({\underline{M}}_\infty > {\ensuremath{\underline{b}}}\mbox{ and } {\overline{M}}_\infty < {\ensuremath{\overline{b}}}),$$ for any $({\ensuremath{\overline{b}}},{\ensuremath{\underline{b}}})$ except those excluded in the statement of the theorem. We note that it is sufficient to show that ${{\mathbb{P}}}({\underline{M}}_\infty = {\ensuremath{\underline{b}}}) = {{\mathbb{P}}}({\overline{M}}_\infty = {\ensuremath{\overline{b}}}) = 0$, and by Proposition \[prop:supatom\] it is only possible to have an atom in the law of the maximum or the minimum if the process stops at the maximum with positive probability; we note however that the stopping time $\tau_J$, due to properties of the Azéma-Yor embedding precludes such behaviour except at the points ${\ell_{\mu}}, {r_{\mu}}$. Considering now the Perkins construction, we note from and the fact that the function ${{\gamma_+}}$ is decreasing, that we will stop at ${\ensuremath{\underline{b}}}$ only if ${{\gamma_+}}(F({\overline{M}}_t)) = {\ensuremath{\underline{b}}}$ and $M_t = {\underline{M}}_t = {\ensuremath{\underline{b}}}$. It follows from that there is a range of values $({\ensuremath{\overline{b}}}_*,{\ensuremath{\overline{b}}}^*)$ for which ${{\gamma_+}}(F(b)) = {\ensuremath{\underline{b}}}$, and consequently, we must have $h(b) = {{\mathbb{P}}}({\underline{M}}_\infty = {\ensuremath{\underline{b}}}, {\overline{M}}_\infty < b)$ increasing in $b$ as $b$ goes from ${\ensuremath{\overline{b}}}_*$ to ${\ensuremath{\overline{b}}}^*$, with $h({\ensuremath{\overline{b}}}_*) = {{\mathbb{P}}}({\underline{M}}_\infty = {\ensuremath{\underline{b}}}, {\overline{M}}_\infty < {\ensuremath{\overline{b}}}_*)=0$ and $h({\ensuremath{\overline{b}}}^*) = {{\mathbb{P}}}({\underline{M}}_\infty = {\ensuremath{\underline{b}}}, {\overline{M}}_\infty < {\ensuremath{\overline{b}}}^*)=\mu(\{{\ensuremath{\underline{b}}}\})$.[^4] Similar results for the function $g$ also follow. Conclusions {#conclusions .unnumbered} =========== In this paper, we studied the possible joint distributions of $({\overline{M}}_\infty,{\underline{M}}_\infty)$ given the law of $M_\infty$, and were able to obtain number of qualitative properties and sharp quantitative bounds. It follows from our results that the interaction between the maximum and minimum is highly non-trivial which makes the pair above much harder to study than $\overline{M}_\infty$ and ${\underline{M}}_\infty$ on their own. This is best seen in the case of Brownian motion where $\overline{B}_t$ has an easily accessible distribution while the description of the joint distribution of $(\underline{B}_t,\overline{B}_t)$ is much more involved. A further natural question arising from our work is to characterise the joint distributions of the joint distributions of the triple $(M_\infty,{\overline{M}}_\infty,{\underline{M}}_\infty)$. At present it is not clear to us if, and to what extent, a complete characterisation of the possible joint distributions of this triple, in the spirit of Rogers [@Rogers:93] and Vallois [@Vallois:93], is feasible. It remains an open and challenging problem. [^1]: e-mail: [](mailto:[email protected]); web: [www.maths.bath.ac.uk/\~mapamgc/](www.maths.bath.ac.uk/~mapamgc/) [^2]: e-mail: [](mailto:[email protected]) ; web: [www.maths.ox.ac.uk/\~obloj/](www.maths.ox.ac.uk/~obloj/). The research has received funding from the European Research Council under the European Union’s Seventh Framework Programme (FP7/2007-2013) / ERC grant agreement no. 335421. The author is also grateful to the Oxford-Man Institute of Quantitative Finance and St John’s College in Oxford for their support. [^3]: Strictly, we only consider the case where $\mu(\{0\}) = 0$. If this is not the case, then the optimal embedding requires independent randomisation to stop some mass at zero initially. [^4]: In fact, as above, it follows from [e.g.]{} [@Rogers:93 Theorem 2.2] that the maximum must have a strictly positive density with respect to Lebesgue measure, and therefore that the function $h$ is strictly increasing between the points ${\ensuremath{\overline{b}}}_*$ and ${\ensuremath{\overline{b}}}^*$.

ArXiv

--- abstract: 'We study the computational difficulty of the problem of finding fixed points of nonexpansive mappings in uniformly convex Banach spaces. We show that the fixed point sets of computable nonexpansive self-maps of a nonempty, computably weakly closed, convex and bounded subset of a computable real Hilbert space are precisely the nonempty, co-r.e. weakly closed, convex subsets of the domain. A uniform version of this result allows us to determine the Weihrauch degree of the Browder-Göhde-Kirk theorem in computable real Hilbert space: it is equivalent to a closed choice principle, which receives as input a closed, convex and bounded set via negative information in the weak topology and outputs a point in the set, represented in the strong topology. While in finite dimensional uniformly convex Banach spaces, computable nonexpansive mappings always have computable fixed points, on the unit ball in infinite-dimensional separable Hilbert space the Browder-Göhde-Kirk theorem becomes Weihrauch-equivalent to the limit operator, and on the Hilbert cube it is equivalent to Weak Kőnig’s Lemma. In particular, computable nonexpansive mappings may not have any computable fixed points in infinite dimension. We also study the computational difficulty of the problem of finding rates of convergence for a large class of fixed point iterations, which generalise both Halpern- and Mann-iterations, and prove that the problem of finding rates of convergence already on the unit interval is equivalent to the limit operator.' address: 'Technische Universität Darmstadt, Germany' author: - Eike Neumann bibliography: - 'lmcs\_final.bib' nocite: '[@*]' title: Computational Problems in Metric Fixed Point Theory and their Weihrauch Degrees --- [^1] Introduction ============ Metric fixed point theory is the study of fixed point properties of mappings that arise from the geometric structure of the underlying space or the geometric properties of the mappings themselves. An important classical framework for metric fixed point theory is the study of nonexpansive mappings in uniformly convex Banach spaces. A Banach space $E$ is called *strictly convex*, if for all $x, y \in E$ with $x \neq y$ and ${||x||} = {||y||} = 1$, we have ${||\tfrac{x + y}{2}||} < 1$. It is called *uniformly convex*, if $$\forall \varepsilon \in(0,2]. \exists \delta \in (0,1].\forall x,y \in B_E.\left({||x - y||} \geq \varepsilon \rightarrow {\Big|\Big|\frac{x + y}{2}\Big|\Big|} \leq 1 - \delta \right) .$$ Here, $B_E$ denotes the closed unit ball of $E$. Clearly, every uniformly convex Banach space is strictly convex. A function $\eta_E\colon(0,2]\to(0,1]$ witnessing the existential quantifier is called a *modulus of convexity* for $E$. By the parallelogram law, every Hilbert space $H$ is uniformly convex with computable modulus of convexity $\eta_H(\varepsilon) = 1 - \sqrt{1 - \frac{\varepsilon^2}{4}}$. More generally, all $L^p$-spaces with $1 < p < \infty$ are uniformly convex with a computable modulus of uniform convexity. A mapping $f\colon\subseteq E \to E$ is called *nonexpansive* if it is Lipschitz-continuous with Lipschitz-constant one, i.e. if $${||f(x) - f(y)||} \leq {||x - y||} \;\text{ for all } x,y\in {\operatorname{dom}}f.$$ We have the following existence result: \[thm: Browder-Goehde-Kirk Theorem\] Let $E$ be a uniformly convex Banach space, let ${K \subseteq E}$ be nonempty, bounded, closed, and convex, and let $f\colon K \to K$ be nonexpansive. Then $f$ has a fixed point. Theorem \[thm: Browder-Goehde-Kirk Theorem\] was proved independently by Browder [@BrowderFixedPointB], Göhde [@Goehde], and Kirk [@KirkFixedPoint] in 1965 (Kirk’s version is even more general than the version stated here). Throughout this paper we denote the fixed point set of a mapping $f$ by ${\operatorname{Fix}}(f)$. A considerable amount of attention is dedicated to the study of so-called fixed point iterations, which start with an initial guess $x_0$ for a fixed point of $f$ and successively improve the guess by applying a computable operation, which yields a sequence $(x_n)_n$ of points in $K$ that is then shown to converge (weakly or strongly) to a fixed point. Many of these results are modifications of either of two classical theorems. \[thm: Halpern’s theorem\] Let $H$ be a Hilbert space, let $K \subseteq H$ be nonempty, bounded, closed, and convex, and let $f\colon K \to K$ be nonexpansive. Choose a starting point $x \in K$ and an “anchor point” $y \in K$ and consider the sequence $(x_n)_n$, where $x_0 = x$ and $x_{n+1} = \tfrac{1}{n + 2} y + (1 - \tfrac{1}{n + 2})f(x_n)$. Then the sequence $(x_n)_n$ converges to the uniquely defined fixed point of $f$ which is closest to the anchor point $y$. \[thm: Krasnoselski-Mann theorem\] Let $E$ be a uniformly convex Banach space, let ${K \subseteq E}$ be nonempty, bounded, closed, and convex, and let $f\colon K \to K$ be nonexpansive and $f(K)$ be compact. Then for any $x \in K$ the sequence $(x_n)_n$, where $x_0 = x$ and $x_{n+1} = (f(x_n) + x_n)/2$, converges to a fixed point of $f$. The iteration employed in Theorem \[thm: Halpern’s theorem\] is a special case of a general iteration scheme, typically referred to as *Halpern iteration*, as it was first introduced by Halpern [@Halpern]. It has the general form $x_{n+1} = (1 - \alpha_n) y + \alpha_nf(x_n)$, where $\alpha_n \in (0,1)$, and the iteration can be shown to converge if certain conditions are imposed on $(\alpha_n)_n$. The iteration used in Theorem \[thm: Krasnoselski-Mann theorem\] can be similarly generalised to the scheme $x_{n+1} = (1 - \alpha_n)x_n + \alpha_nf(x_n)$, and again there are certain conditions that guarantee convergence. This iteration scheme is typically called *Krasnoselski-Mann iteration* or simply *Mann iteration*. In Hilbert space, Krasnoselski’s iteration converges weakly to a fixed point, even in the absence of compactness (cf. [@KrasnoselskiWeakConvergence]). While these iterations do allow us to compute a sequence of approximations which is guaranteed to eventually converge to a fixed point, it is well known that the requirement of mere convergence is too weak to constitute a satisfactory notion of effective approximation, as there exist for instance computable sequences of rational numbers whose limit encodes the special halting problem (cf. [@SpeckerSequence]). It is hence important to understand the quantitative convergence behaviour of the approximation sequence. Quantitative aspects of metric fixed point theory have been very successfully studied within the programme of *proof mining* (the standard reference is [@Kohlenbook], see also e.g. [@Kohlenbach; @KohlenbachMetric; @KohlenbachBRS; @Leustean; @KohlenbachLeustean; @KoernleinKohlenbach; @SchadeKohlenbach]), which is concerned with the extraction of hidden effective data from non-effective proofs. Most of the applications of proof mining in fixed point theory focus on the extraction of either of two types of effective data. Firstly, one considers *rates of asymptotic regularity* of the iteration, which in this context mean rates of convergence of the sequence $({||f(x_n) - x_n||})_n$ towards zero. These allow us to compute arbitrarily good $\varepsilon$-fixed points, i.e. points $x_{\varepsilon}$ satisfying ${{||f(x_{\varepsilon}) - x_{\varepsilon}||} < \varepsilon}$, up to arbitrary precision with an a-priori running time estimate. Secondly, one considers so-called *rates of metastability* (see also [@KreiselNoCounterexample1; @KreiselNoCounterexample2; @TaoMetastability1; @TaoMetastability2]), which constitute a more refined quantitative measure of approximation quality. A function $\Phi\colon{\mathbb{N}}^{{\mathbb{N}}}\times{\mathbb{N}}\to{\mathbb{N}}$ is called a rate of metastability for the sequence $(x_n)_n$ if it satisfies $$\label{eq: metastability}\forall n\in {\mathbb{N}}.\forall g\colon{\mathbb{N}}\to{\mathbb{N}}.\exists k \leq \Phi(g,n).\forall i,j\in[k;k + g(k)]\left({||x_{i} - x_j||} < 2^{-n} \right).$$ Note that is classically (but not constructively) equivalent to the statement that $(x_n)_n$ is a Cauchy sequence, so that metastability can be viewed as a finitary version of convergence. Also note that in the case of Krasnoselski’s iteration, asymptotic regularity is the special case of metastability where $g(k) = 1$ for all $k\in{\mathbb{N}}$. Of course, both types of information are strictly weaker than actual rates of convergence. In fact, effective uniform rates of convergence cannot exist, as the existence result fails to be computably realisable already in the case where $K = [0,1]$. \[uniform uncomputability of fixed points\] The multi-valued operator which receives as input a nonexpansive self map $f$ of the compact unit interval $[0,1]$ and returns some fixed point of $f$ is not computable. While Theorem \[uniform uncomputability of fixed points\] already shows that there exists no algorithm for computing a rate of convergence for Krasnoselski’s or Halpern’s iteration uniformly in the input function and the starting point, it leaves several questions open: whether every computable nonexpansive mapping has a computable fixed point, whether there exist non-uniformly computable rates of convergence for the Mann- or Halpern-iteration for every computable nonexpansive mapping, at least for certain suitable starting points, whether fixed points are uniformly computable relative to discrete advice, what the exact relation between the computational content of the three theorems is, and how their computational content relates to the computational content of other mathematical theorems, such as Brouwer’s fixed point theorem. In this paper, we study the computational content of Theorems \[thm: Browder-Goehde-Kirk Theorem\], \[thm: Krasnoselski-Mann theorem\], and \[thm: Halpern’s theorem\], as well as related computational problems in terms of *Weihrauch degrees*, which have been proposed by Gherardi and Brattka [@WeihrauchDegrees] as a framework for classifying mathematical theorems according to their computational content. Many classical mathematical theorems have been classified over the recent years. Recently, Brattka, Le Roux and Pauly [@ConnectedChoice] have shown that Brouwer’s fixed point theorem in dimension $n$ is equivalent to the closed choice principle on the closed unit ball in ${\mathbb{R}}^n$ restricted to connected sets, and that it is equivalent to Weak Kőnig’s Lemma from dimension three upwards. Their work is based on a characterisation of the fixed point sets of computable self-maps of the unit ball in ${\mathbb{R}}^n$, due to Miller [@Miller]. We provide a similar characterisation for the fixed point sets of computable nonexpansive self-maps of nonempty, convex, closed, and bounded subsets of computable Hilbert space, which we can use to determine the Weihrauch degree of the Browder-Göhde-Kirk theorem. This will in particular allow us to compare the computational content of the Browder-Göhde-Kirk theorem and the problem of finding rates of convergence for fixed point iterations to Brouwer’s classic result. Preliminaries ============= Here we review some basic notions from computable and functional analysis and the theory of Weihrauch reducibility. Most of the results in this section are more or less folklore, and none of them are original, except maybe Proposition \[prop: equivalence lower semi-located and computably closed implies locally compact\]. Standard references in computable analysis are [@PourElRichards] and [@Weih]. A more general treatment of the theory of computable metric spaces can be found in [@BrattkaGherardi] and [@BrattkaPresser]. The results in functional analysis reviewed here can for instance be found in [@Megginson] or [@Werner]. We will closely follow the approach to computable analysis taken by Matthias Schröder [@SchroederPhD], and more recently by Arno Pauly [@PaulyRepresented], particularly concerning the canonical constructions of hyperspaces. Also, we adopt most of the notation and terminology from [@PaulyRepresented], which differs from standard terminology at certain points (see Caveat \[caveat: terminology\]). A *numbering* of a nonempty countable set $S$ is a surjective partial mapping ${\nu\colon\subseteq {\mathbb{N}}\to S}$. A *representation* of a nonempty set $X$ is a surjective partial mapping ${\delta\colon\subseteq {\mathbb{N}}^{\mathbb{N}}\to X}$. If $\delta$ is a representation of $X$, we call the tuple $(X,\delta)$ a *represented space*. If the underlying representation is clear from context, we will often simply write $X$ for $(X,\delta)$ and by convention denote the underlying representation $\delta$ of $X$ by $\delta_X$. If $\delta$ and $\varepsilon$ are representations of the same set $X$, we denote *continuous reduction* by $\delta \leq_t \varepsilon$ and *computable reduction* by $\delta \leq \varepsilon$. A representation $\delta\colon\subseteq{\mathbb{N}}^{\mathbb{N}}\to X$ is *admissible* if it is continuous and maximal with respect to continuous reduction. A *represented topological space*[^2] is a tuple $(X,\delta)$, where $X$ is a topological space and $\delta$ is an admissible representation for $X$. If the representation is clear from the context, we will simply write $X$ for $(X,\delta)$ and by convention denote the underlying admissible representation $\delta$ by $\delta_X$. We say that a partial mapping $F\colon\subseteq {\mathbb{N}}^{\mathbb{N}}\to {\mathbb{N}}^{\mathbb{N}}$ is a *realiser* for a partial multi-valued mapping (or “multimapping”) $f\colon\subseteq X \rightrightarrows Y$ between represented spaces and write $F\vdash f$ if $\delta_Y(F(p)) \in f(x)$, whenever $\delta_X(p) = x$. We call $f$ *computable* if it has a computable realiser, and *realiser-continuous* if it has a continuous realiser. If we want to emphasise the underlying representations, we will write that $f$ is $(\delta_X,\delta_Y)$-computable or $(\delta_X,\delta_Y)$-continuous respectively. We denote by $\rho$ the standard representation of real numbers. If $\delta\colon \subseteq {\mathbb{N}}^{\mathbb{N}}\to X$ and $\varepsilon\colon\subseteq {\mathbb{N}}^{\mathbb{N}}\to Y$ are representations (or numberings), we denote by $[\delta\to\varepsilon]$ the canonical representation of the space ${[X\rightarrow Y]}$ of functions with continuous realiser. If $X$ and $Y$ are represented topological spaces, then ${[X\rightarrow Y]}$ coincides with the space ${\operatorname{\mathcal{C}}}_{\operatorname{seq}}(X,Y)$ of sequentially continuous functions from $X$ to $Y$. If $X$ is first-countable, ${[X\rightarrow Y]}$ furthermore coincides with the space ${\operatorname{\mathcal{C}}}(X,Y)$ of continuous functions from $X$ to $Y$. Moreover, we denote by $\delta\times\varepsilon$ the canonical representation of the product space $X\times Y$, by $\delta^\omega$ the canonical representation of the space $X^{\mathbb{N}}$ and by $\delta^*$ the canonical representation of ${X^* = \bigcup_{n \in {\mathbb{N}}} X^n}$. If $\delta$ and $\varepsilon$ are representations of the same space $X$, we let $\delta\sqcap\varepsilon = \pi_0(\delta\times\varepsilon)|^{\Delta(X)}$, where $\Delta(X) = \{(x,y) \in X^2\;|\; x = y\}$ and $\pi_0$ is the projection onto the first coordinate. If $K \subset X$ is a subset of $X$, we sometimes write $\delta_K$ for $\delta_X|^K$. As already mentioned, the following constructions are essentially due to [@SchroederPhD] and [@PaulyRepresented]. For any represented space $X$, the canonical function-space construction gives rise to canonical representations of the hyperspaces of “open” and “closed” subsets of $X$, by postulating that openness corresponds to semi-decidability. Let ${\mathbb{S}}= \{0,1\}$ denote Sierpiński space with topology $\{\emptyset, \{0,1\}, \{1\} \}$ and representation $$\sigma(p) = 0 \;:\Leftrightarrow\; p = 0.$$ The characteristic function $\chi_U\colon X \to \{0,1\}$ of a set $U$ is defined as $\chi_U(x) = 1 :\Leftrightarrow x \in U$. \[Def: canonical representation of open and closed sets\] Let $X$ be a represented space. We call a set $U \subseteq X$ *open*, if its characteristic function $\chi_U\colon X \to {\mathbb{S}}$ is realiser-continuous, i.e. $\chi_U \in {[X\rightarrow {\mathbb{S}}]}$. A $\theta^X$-name of an open set $U\subseteq X$ is a $[\delta_X\to\sigma]$-name of its characteristic function $\chi_U$. The set of all open subsets of $X$ with representation $\theta^X$ defines the represented space ${\mathcal{O}}(X)$. Dually, we define the represented space ${\mathcal{A}}(X)$ of closed subsets of $X$ by identifying a closed set $A \subseteq X$ with its complement in ${\mathcal{O}}(X)$ and call the underlying representation $\psi^X$. We will often just write $\psi$ for $\psi^X$ if the underlying space is clear from context. We call the computable points of ${\mathcal{O}}(X)$ *semi-decidable* and the computable points in ${\mathcal{A}}(X)$ *co-semi-decidable*. Note that, just like the notion of realiser continuity may differ from topological continuity, the notions of closedness and openness for subsets of represented spaces are a-priori different from the notions of topological openness and closedness. If $X$ is an admissibly represented topological space, then the set ${\mathcal{O}}(X)$ coincides with the set of all sequentially open subsets of $X$ and ${\mathcal{A}}(X)$ coincides with the set of all sequentially closed subsets of $X$. If in addition $X$ is second-countable then ${\mathcal{O}}(X)$ is the set of open subsets of $X$ and ${\mathcal{A}}(X)$ is the set of all closed subsets of $X$. \[caveat: terminology\] Note that the terminology introduced here, which is mainly due to [@PaulyRepresented], is different from the usual terminology used in computable analysis, which is for instance used in Weihrauch’s book [@Weih]. In [@Weih], the space ${\mathcal{A}}(X)$ is denoted by ${\mathcal{A}}_{>}(X)$ and its computable elements are called co-r.e. closed, rather than co-semi-decidable. Although we have introduced our ${\mathcal{A}}(X)$ as “the space of closed subsets” of $X$, we deliberately refrain from referring to its computable points as “computably closed”, so as to avoid confusion with topological closedness on one hand, and with Weihrauch’s terminology on the other. The symbol ${\mathcal{A}}(X)$ is used in [@Weih] to denote the space of closed and overt subsets of $X$, to be introduced below. Also note that in the abstract we used *Weihrauch’s* terminology. The space ${\mathcal{A}}(X)$ can be thought of as “the space of closed sets encoded via negative information”. The following definition provides in a certain sense a notion of “closed sets encoded via *positive* information”. \[Def: canonical representation of overt sets\] Let $X$ be a represented space. We define the [*represented space*]{} ${\mathcal{V}}(X)$ [*of overt closed subsets of*]{} $X$ to be the represented space of closed subsets of $X$, where a closed set $A \subseteq X$ is represented by a $[\theta^X \to {\mathbb{S}}]$-name of the function $$\operatorname{intersects?}_{A}\colon {\mathcal{O}}(X) \to {\mathbb{S}},\; U \mapsto \begin{cases}1 &\text{if }U\cap A \neq \emptyset\text{,} \\ 0 &\text{otherwise.} \end{cases}$$ We denote the standard representation of ${\mathcal{V}}(X)$ by $\upsilon^X$ or simply $\upsilon$ and call the computable points of ${\mathcal{V}}(X)$ [*computably overt*]{}. Computably overt closed sets are those, for which intersection with an open set can be effectively verified. The space ${\mathcal{V}}(X)$ hence corresponds to the space ${\mathcal{A}}_{<}(X)$ in [@Weih], and hence is sometimes called the “space of closed sets, represented with positive information”. In [@PaulyRepresented] it is argued that from an intrinsic perspective, the word “closed” is rather misleading, because the closure properties of the space ${\mathcal{V}}(X)$ differ significantly from the closure properties of closed sets (e.g. union is computable but intersection is not, the image of a computably overt set under a computable function is computably overt, but the preimage is not), and we agree with this position. Overtness is related to effective separability, which yields a convenient criterion for computable overtness (see e.g. [@BrattkaPresser Theorem 3.8 (1)]). \[prop: overtness and recursive enumerability\] Let $X$ be a separable represented topological space. Define a representation $\delta_{{\operatorname{enum}}}$ of the set of nonempty closed subsets of $X$ as follows: $$\delta_{{\operatorname{enum}}}(p) = A \; :\Leftrightarrow \delta_X^{\omega}(p) \text{ is dense in } A.$$ Then $\delta_{{\operatorname{enum}}} \leq \upsilon\big|^{{\mathcal{V}}(X)\setminus\{\emptyset\}}$. Suppose we are given a dense sequence $(x_n)_n$ in a closed set $A$, and an open set $U \in {\mathcal{O}}(X)$. In order to verify if $A\cap U$, check if there exists $n \in {\mathbb{N}}$ such that $x_n \in U$. This proves the claim. Our representation $\delta_{{\operatorname{enum}}}$ is called $\delta_{\operatorname{range}}$ in [@BrattkaGherardi] and [@BrattkaPresser]. Next we define the canonical representation of the hyperspace of compact subsets of a Hausdorff represented topological space $X$. In a countably based $T_1$ space, a compact set can be represented as a list of all its finite open covers by basic neighbourhoods. It is easy to see that this representation is characterised by the property that containment in an open set is semi-decidable. This can be used to generalise the definition to arbitrary represented topological spaces, and in fact to arbitrary represented spaces. For the sake of simplicity we restrict ourselves to the case of Hausdorff represented topological spaces. \[Def: canonical representations of compact sets\] Let $X$ be a Hausdorff represented topological space. The represented space ${\mathcal{K}}(X)$ of compact subsets of $X$ is the set of all compact subsets of $X$, where a compact set $K \in {\mathcal{K}}(X)$ is represented as a $[\theta^X\to\sigma]$-name of the function $\operatorname{contained?}_K\colon {\mathcal{O}}(X) \to {\mathbb{S}}$, $$\operatorname{contained?}_K(U) = 1 \Leftrightarrow K \subseteq U.$$ We denote the canonical representation of ${\mathcal{K}}(X)$ by $\kappa$ and call the computable points of ${\mathcal{K}}(X)$ *computably compact*. Note that, like ${\mathcal{A}}(X)$, our space ${\mathcal{K}}(X)$ only encodes “negative” information on compact sets. Weihrauch [@Weih] hence uses the notation “$\kappa_{>}$” for our $\kappa$. Similarly as in the case of ${\mathcal{A}}(X)$, computable points in our ${\mathcal{K}}(X)$ are called “co-r.e. compact” by some authors. Definition \[Def: canonical representations of compact sets\] can be generalised to arbitrary represented spaces, essentially by using the same approach as in Definition \[Def: canonical representation of open and closed sets\], and *calling* a subset $K$ of a represented space $X$ compact if the function $\operatorname{contained?}_K$ is an element of ${[{\mathcal{O}}(X)\rightarrow {\mathbb{S}}]}$. In general this will only yield a representation of the space of *saturated* compact sets (cf. [@PaulyRepresented]), or a multi-valued representation of the space of compact sets (cf. [@SchroederPhD]). If $X$ is a $T_1$ represented topological space, then the thus obtained space ${\mathcal{K}}(X)$ coincides with the set of all compact subsets of the *sequentialisation* $\operatorname{seq}(X)$ of $X$, whose open sets are the sequentially open sets of $X$. For details see [@SchroederPhD]. By Proposition 3.3.2 (3) in [@SchroederPhD], the notions of compactness, sequential compactness, and compactness in the sequentialisation coincide for Hausdorff represented topological spaces, so we obtain Definition \[Def: canonical representations of compact sets\]. It will sometimes be convenient to work with an intrinsic notion of computable compactness for represented spaces, which we introduce next. \[Def: compact space\] A Hausdorff represented topological space $X$ is called [*computably compact*]{}, if the mapping $$\operatorname{empty?}\colon {\mathcal{A}}(X) \to {\mathbb{S}},\; A \mapsto \begin{cases}1 &\text{ if }A = \emptyset, \\0 &\text{ otherwise.} \end{cases}$$ is computable. Note that the terminology used in Definitions \[Def: compact space\] and \[Def: canonical representations of compact sets\] is consistent in the sense that $X$ is a computably compact space if and only if $X$ is a computable point in ${\mathcal{K}}(X)$. The next proposition is a converse to this in some sense. \[prop: compact subspace of represented space is compact set\] Let $X$ be a Hausdorff represented topological space and $K \subseteq X$ be a nonempty co-semi-decidable subset, such that the represented space $(K,\delta_X\big|^K)$ is computably compact. Then $K$ is a computably compact subset of $X$, i.e. a computable point in ${\mathcal{K}}(X)$. It follows immediately from the definition of ${\mathcal{O}}(X)$ that the mapping $$\cap_K \colon {\mathcal{O}}(X) \to {\mathcal{O}}(K),\; U \mapsto U \cap K$$ is computable. Now, $U \supseteq K$ if and only if $K \setminus (U \cap K) = \emptyset$. It again follows from the definition, that the mapping $${\mathcal{O}}(K) \to {\mathcal{A}}(K), \; U \mapsto K\setminus U$$ is computable. Since $K$ is a computably compact represented space, the mapping $${\mathcal{A}}(K) \to {\mathbb{S}}, \; A \mapsto \begin{cases}1 &\text{ if }A = \emptyset, \\ 0 &\text{ if }A \neq \emptyset\end{cases}$$ is computable. It follows that the set of open subsets of $X$ containing $K$ is computably open, i.e. $K$ is computably compact. A closed subset of a compact space is compact, and in a Hausdorff space, every compact set is closed. We have an effective counterpart of this in the theory of represented spaces. A represented space $X$ is called *effectively Hausdorff* if the mapping $X \to {\mathcal{A}}(X),\; x \mapsto \{x\}$ is computable. \[prop: closed and compact sets, (i) closed set is compact, (ii) compact set in Hausdorff space is closed\] Let $X$ be a Hausdorff computably compact represented topological space. 1. The mapping ${\operatorname{id}}\colon {\mathcal{A}}(X) \to {\mathcal{K}}(X)$ is well-defined and computable. 2. If $X$ is effectively Hausdorff, then the mapping ${\operatorname{id}}\colon {\mathcal{K}}(X) \to {\mathcal{A}}(X)$ is well-defined and computable. <!-- --> 1. We are given a closed set $A \in {\mathcal{A}}(X)$ which we want to compute as a compact set $A \in {\mathcal{K}}(X)$. Given an open set $U \in {\mathcal{O}}(X)$ we want to verify if $U \supseteq A$. In order to do so, check if $U \cup A^C$ covers $X$, using that $X$ is computably compact. 2. We are given a compact set $K \in {\mathcal{K}}(X)$ which we want to compute as a closed set $K \in {\mathcal{A}}(X)$. Given a point $x \in X$ we want to verify if $x \notin K$. In order to do so, compute $\{x\} \in {\mathcal{A}}(X)$, using that $X$ is effectively Hausdorff, and verify if $\{x\}^C \supseteq K$, using the compactness information on $K$. It is easy to see that the computability and well-definedness of the mapping $${{\operatorname{id}}\colon {\mathcal{A}}(X) \to {\mathcal{K}}(X)}$$ *characterises* computably compact represented spaces (cf. also [@PaulyRepresented]). \[thm: supremum on compact space\] Let $K$ be a computably compact represented topological space, containing a computable dense sequence. Then the mapping $$\max\colon {[K\rightarrow {\mathbb{R}}]} \to {\mathbb{R}},\; f \mapsto \max\{f(x)\;\big|\; x \in K\}$$ is well-defined and computable. Since $K$ is adequately represented, we have ${[K\rightarrow {\mathbb{R}}]} = {\operatorname{\mathcal{C}}}_{\operatorname{seq}}(K,{\mathbb{R}})$, and $K$ is sequentially compact thanks to Proposition 3.3.2 (3) in [@SchroederPhD]. It follows that for any $f \in {[K\rightarrow {\mathbb{R}}]}$, the set $f(K)$ is sequentially compact in ${\mathbb{R}}$ and thus compact. This shows that $\max$ is well-defined. It remains to show that $\max(f)$ is computable relative to $f$. Let $(x_n)_n$ be a computable dense sequence in $K$. Then the sequence $(f(x_n))_n$ is computable relative to $f$, with $\sup_{n \in {\mathbb{N}}} f(x_n) = \max(f)$. On the other hand, for every computable $b \in {\mathbb{R}}$, the set $U_b = \{x \in K\;\big|\; f(x) < b\}$ is semi-decidable relative to $f$, so that by the computable compactness of $K$, the predicate $\forall x \in K.\left(f(x) < b\right)$ is semi-decidable relative to $f$ for all $b \in {\mathbb{Q}}$. We can use this to construct a sequence $(b_n)_n$ of real numbers which is computable relative to $f$ and satisfies $\max(f) = \inf_{n \in {\mathbb{N}}} b_n$. Since $\max(f)$ can hence be approximated arbitrarily well “from above” as well as “from below”, it is computable relative to $f$. It follows from Theorem \[thm: supremum on compact space\] that every finite dimensional uniformly convex computable Banach space $E$ has a computable modulus of uniform convexity $\eta_E$, since we may put $$\eta_E(\varepsilon) = \inf\Big\{1 - {\Big|\Big|\frac{x + y}{2}\Big|\Big|} \;\big|\; x,y \in B_E, {||x - y||} \geq \varepsilon \Big\},$$ and the set $\{(x,y) \in B_E\times B_E \;\big|\; {||x - y||}\geq \varepsilon\}$ is computably compact relative to $\varepsilon$ and contains a computable, dense sequence relative to $\varepsilon$. Since the proof of Theorem \[thm: supremum on compact space\] is uniform in $K$, the claim follows. \[thm: Kreinovich’s theorem\] Let $K$ be a computably compact represented topological space. Then the mapping $${\operatorname{UC}}_K\colon\subseteq {\mathcal{A}}(K) \to K, \; \{x\}\mapsto x$$ is computable. A *computable metric space* is a triple $(M,d,\nu_M)$, where $(M,d)$ is a metric space and $\nu_M\colon{\mathbb{N}}\to A$ is a numbering of a dense subset $A\subseteq M$, such that $d\colon A\times A \to {\mathbb{R}}$ is $(\nu_M\times\nu_M,\rho)$-computable. With a computable metric space we associate the represented space $(M,\delta_M)$, where $$\delta_M(p) = x \;:\Leftrightarrow d(\nu_M(p(n)),x) \leq 2^{-n} \;\text{ for all }n \in {\mathbb{N}}.$$ One can show that the above defined canonical representation of a computable metric space $M$ is admissible and that $d\colon M\times M\to{\mathbb{R}}$ is computable. This canonical representation $\delta_M$ is also called the *Cauchy representation* induced by $\nu_M$. We will refer to the points in ${\operatorname{im}}\nu_M$ as the *rational points* of the represented space $M$. Note that any computable metric space is separable, and hence Hausdorff, by definition. In fact, every computable metric space is effectively Hausdorff, since the predicate $d(x,y) > 0$ is a semi-decidable relative to $x$ and $y$. In any metric space $M$ we denote by $B(x,r) = \{y \in M \;\big|\; d(x,y) < r\}$ the open ball of radius $r$ centred at $x$, and by ${\overline{B}}(x,r) = \{y \in M \;\big|\; d(x,y) \leq r\}$ the closed ball of radius $r$ centred at $x$. The following result is more or less folklore, and justifies the more abstract Definition \[Def: canonical representations of compact sets\] of computable compactness. \[prop: computable total boundedness equivalent computable compactness\] A computable metric space is computably compact if and only if it is complete and *computably totally bounded*, i.e. if and only if there exists a function $\alpha\colon {\mathbb{N}}\to M^{*}$ such that $$\forall n \in {\mathbb{N}}. \forall x \in M.\exists k \leq \operatorname{lth}(\alpha(n)).\left(d(x,\alpha(n)_k) < 2^{-n} \right).$$ We call $\alpha(n)$ a $2^{-n}$-net in $M$. Suppose that $M$ is computably compact. Then $M$ is compact and thus complete. Let $(a_k)_k$ be a dense computable sequence in $M$. Since $M$ is computably compact, we can verify for all $n,m \in {\mathbb{N}}$ if the open set $B(\tilde{a}_1,2^{-n - 1})\cup\dots\cup B(\tilde{a}_m,2^{-n - 1})$ is equal to all of $M$, where $\tilde{a}_k$ is a rational approximation to $a_k$ to up error $2^{-n-1}$. In that case, $a_1,\dots,a_m$ is a $2^{-n}$-net in $M$. On the other hand, since $(a_k)_k$ is dense in $M$ and $M$ is compact, this process has to finish after a finite number of steps for each $n\in{\mathbb{N}}$. It follows that $M$ is computably totally bounded. Suppose now that $M$ is complete and computably totally bounded. Let $$S = \{B(c_1,r_1),\dots,B(c_k,r_k)\}$$ be a collection of rational balls, i.e. balls whose radii are rational numbers and whose centres are rational points in $M$. We show that we can verify if $S$ is a cover of $A$. It is a standard argument that this suffices in order to establish that $M$ is computably compact. Let $(\langle a^n_1,\dots,a^n_{l(n)}\rangle)_{n \in {\mathbb{N}}}$ be a computable sequence of $2^{-n}$-nets in $M$. Then $S$ covers $M$ if and only if there exists $n\in {\mathbb{N}}$ such that for all $i \in \{1,\dots,l(n)\}$ there exists $j \in \{1,\dots,k\}$ such that $d(a^n_i,c_j) < r_j - 2^{-n}$. This property is semi-decidable, so $M$ is computably compact. In complete computable metric spaces, overtness is characterised by separability, in the sense that Proposition \[prop: overtness and recursive enumerability\] admits a converse (cf. [@BrattkaPresser Theorem 3.8 (2)]). \[prop: overtness and recursive enumerability, metric case\] Let $M$ be a complete computable metric space. Define the representation $\delta_{{\operatorname{enum}}}$ of the set of nonempty closed subsets of $M$ as in Proposition \[prop: overtness and recursive enumerability\]. Then ${\delta_{{\operatorname{enum}}} \equiv \left(\upsilon\big|^{{\mathcal{V}}(M)\setminus\{\emptyset\}}\right)}$. The direction $\delta_{{\operatorname{enum}}} \leq \upsilon$ was already proved in Proposition \[prop: overtness and recursive enumerability\], so it remains to prove $\upsilon \leq \delta_{{\operatorname{enum}}}$. Suppose we are given a closed set $A \in {\mathcal{V}}(M)$. We can compute an enumeration $(B_m)_m$ of all open rational balls (i.e. balls with rational centre and radius) with radius at most $1$ intersecting $A$. We use this to construct a dense sequence $(x_m)_m$ in $A$. The $m^{th}$ element in the sequence is computed as follows: the first approximation $x_m^{(0)}$ to $x_m$ is the centre of $B_m$. Let $1 \geq \varepsilon > 0$ denote the radius of $B_m$. We claim that we can find an open rational ball $B_m^{(1)}$ with radius at most $\tfrac{\varepsilon}{2}$ which is contained in $B_m$ and intersects $A$. Let $a \in B_m\cap A$. Then there exists rational $\tfrac{\varepsilon}{2} > \delta > 0$ such that $d(a,x_m^{(0)}) < \varepsilon - \delta$. Let $\tilde{a}$ be a rational approximation of $a$ up to error $\delta/2$. Then $d(\tilde{a},x_m^{(0)}) < \varepsilon - \delta/2$. In particular, $B(\tilde{a},\delta/2)\cap A \neq \emptyset$ and $B(\tilde{a},\delta/2) \subseteq B(x_m^{(0)},\varepsilon)$. On the other hand, we can verify for a given rational $a$ and $\delta < \tfrac{\varepsilon}{2}$ that $d(a,x_m^{(0)}) < \varepsilon - \delta$ and that $B(a,\delta) \cap A \neq \emptyset$. We may hence search for such $a$ and $\delta$, and put $B_m^{(1)} = B(a,\delta)$ and $x_m^{(1)} = a$. Continuing in this manner, we obtain a Cauchy sequence $x_m^{(n)}$ with $d(x_m^{(n)},x_m^{(n + k)}) < 2^{-n}$ for all $k,n \in {\mathbb{N}}$. Since $M$ is complete, the sequence $x_m^{(n)}$ converges to some element $x_m \in A$ with $d(x_m^{(n)},x_m) \leq 2^{-n}$. Applying this to all $(B_m)_m$ in parallel, we obtain a computable sequence $(x_m)_m$. It remains to show that $(x_m)_m$ is dense in $A$. Let $a \in A$, and let $\varepsilon$ be a rational number satisfying $1 > \varepsilon > 0$. There exists a rational point $x$ satisfying $d(a,x) < \tfrac{\varepsilon}{2}$. In particular, $B(x,\tfrac{\varepsilon}{2}) \cap A \neq \emptyset$, so $B(x,\tfrac{\varepsilon}{2}) = B_k$ for some $k \in {\mathbb{N}}$. It follows from the construction of $x_k$ that $d(x_k,a) < \varepsilon$. If $M$ is a computable metric space, we have another natural notion of computability for closed sets, by identifying a closed set $A \subseteq M$ with its *distance function* $$\label{eq: distance} d_A\colon M\to {\mathbb{R}}, \; d_A(x) = \inf\{d(x,y) \;|\; y \in A\}.$$ Define represented spaces ${\mathbb{R}}_{<} = ({\mathbb{R}},\rho_{<})$ and ${\mathbb{R}}_{>} = ({\mathbb{R}},\rho_{>})$ via $$\rho_{<}(p) = x\; :\Leftrightarrow\; \sup_{n \in {\mathbb{N}}}\nu_{{\mathbb{Q}}}^{\omega}(p)(n) = x \;\;\text{ and }\;\;\rho_{>}(p) = x\; :\Leftrightarrow\; \inf_{n \in {\mathbb{N}}}\nu_{{\mathbb{Q}}}^{\omega}(p)(n) = x.$$ Computable elements of ${\mathbb{R}}_{<}$ are called *left-r.e. numbers*, and computable elements of ${\mathbb{R}}_{>}$ are called *right-r.e. numbers*. Obviously, a number is computable if and only if it is both right- and left-r.e., whereas a classic result due to Specker [@SpeckerSequence] asserts the existence of both uncomputable left-r.e.- and uncomputable right-r.e. numbers. \[Def: distance representations\] Let $M$ be a computable metric space. 1. The represented space ${\mathcal{A}}_{{\operatorname{dist}}}(M)$ is the space of nonempty closed subsets of $M$, where a closed subset $A\subseteq M$ is represented via a $[\delta_M\to\rho]$-name of its distance function . 2. The represented space ${\mathcal{A}}_{{\operatorname{dist}}_{<}}(M)$ is the space of nonempty closed subsets of $M$, where a closed subset $A\subseteq M$ is represented via a $[\delta_M\to\rho_<]$-name of its distance function . 3. The represented space ${\mathcal{A}}_{{\operatorname{dist}}_{>}}(M)$ is the space of nonempty closed subsets of $M$, where a closed subset $A\subseteq M$ is represented via a $[\delta_M\to\rho_>]$-name of its distance function . Computable points of ${\mathcal{A}}_{{\operatorname{dist}}}(M)$ are called [*located*]{}, computable points of ${\mathcal{A}}_{{\operatorname{dist}}_<}(M)$ are called [*lower semi-located*]{}, and computable points of ${\mathcal{A}}_{{\operatorname{dist}}_>}(M)$ are called [*upper semi-located*]{}. Using Proposition \[prop: overtness and recursive enumerability, metric case\], it is easy to see that for any complete computable metric space $M$, the canonical representations of the spaces ${\mathcal{A}}_{{\operatorname{dist}}_>}(M)$ and ${\mathcal{V}}(M)\setminus\{\emptyset\}$ are equivalent (see also [@BrattkaPresser Theorem 3.7]). It is also easy to see that ${\operatorname{id}}\colon {\mathcal{A}}_{{\operatorname{dist}}_{<}}(M) \to {\mathcal{A}}(M)$ is computable (see [@BrattkaPresser Theorem 3.11 (1)]). In [@Weih Lemma 5.1.7] it is proved, that for $M = {\mathbb{R}}^d$, the canonical representations of ${\mathcal{A}}_{{\operatorname{dist}}_<}(M)$ and ${\mathcal{A}}(M)\setminus\{\emptyset\}$ are equivalent, and the argument readily generalises to any complete computable metric space with (effectively) compact closed balls (see [@BrattkaPresser Theorem 3.11 (3)]). Any such space is locally compact. Local compactness is in fact necessary for the reduction to hold: \[prop: equivalence lower semi-located and computably closed implies locally compact\] Let $M$ be a complete computable metric space. If the identity mapping ${\operatorname{id}}\colon {\mathcal{A}}(M)\setminus\{\emptyset\} \to {\mathcal{A}}_{{\operatorname{dist}}_<}(M)$ is computable, then $M$ is locally compact. If $M$ is a singleton, the claim is trivial, so we may assume that $M$ consists of at least two points. Given $x \in M$ we show that we can compute ${\overline{B}}(x,r)$ as a compact subset of $M$, for $r$ sufficiently small. We search for a rational $y \in M$ and $r\in{\mathbb{Q}}_{+}$ with $d(y,x) > r$. By an argument similar to Proposition \[prop: computable total boundedness equivalent computable compactness\], it suffices to compute for every $k \in {\mathbb{N}}$ a cover of ${\overline{B}}(x,r)$ by balls of radius $2^{-k}$, whose centres are rational points in $M$. Since $y$ is computable, the singleton $\{y\}$ is co-semi-decidable, so we can compute an enumeration of balls with rational centres and radii exhausting the complement of $\{y\}$ such that every ball has radius at most $2^{-k}$. We feed this enumeration into the machine computing the identity ${\operatorname{id}}\colon {\mathcal{A}}(M)\setminus\{\emptyset\} \to {\mathcal{A}}_{{\operatorname{dist}}_<}(M)$ and use the $[\delta_M\to\rho_<]$-name provided by the machine to compute ${\operatorname{d}}(x,\{y\})$ from below. After having processed finitely many balls, the machine will output the lower bound $r$ on ${\operatorname{d}}(x,\{y\}) = d(x,y)$. This means that ${\overline{B}}(x,r)$ is covered by these finitely many balls, since otherwise we could force the machine computing the identity to err. The proof of Proposition \[prop: equivalence lower semi-located and computably closed implies locally compact\] shows that we can even compute a witness for the local compactness of $M$, namely a function $f\colon M \to {\mathcal{K}}(M)$ which maps a point $x$ to a compact closed ball containing $x$. A *(real) computable normed space* is a normed real vector space $E$ together with a numbering $e\colon{\mathbb{N}}\to E$ such that ${\operatorname{span}}\{e(n) \;|\; n \in {\mathbb{N}}\}$ is dense in $E$ and $(E,d,\nu_E)$ is a computable metric space, where $d(x,y) = {||x - y||}$ and $\nu_E$ is a canonical notation of all (finite) ${\mathbb{Q}}$-linear combinations of ${\operatorname{im}}e$. A complete computable normed space is called a computable Banach space. A computable normed space which is also a Hilbert space is called a computable Hilbert space. The inner product in a computable Hilbert space is computable by the polarisation identity. A computable normed space becomes a represented space when endowed with the Cauchy representation induced by the numbering $\nu_E$. In this representation the vector space operations and the norm are computable functions and $0 \in E$ is a computable point of the represented space $E$. An important feature of (infinite dimensional) computable normed spaces is that without loss of generality the fundamental sequence is linearly independent (cf. [@PourElRichards p. 142]). \[lem: effective independence lemma\] Let $(E,e)$ be an infinite dimensional computable normed space. Then there exists a computable function $f\colon {\mathbb{N}}\to {\mathbb{N}}$ such that $e\circ f\colon{\mathbb{N}}\to E$ has dense span in $E$ and consists of linearly independent vectors. 1. Every computable real Hilbert space $H$ has a computable orthonormal basis, i.e. an orthonormal basis which is a (potentially finite) computable sequence in $H$. 2. Every finite dimensional computable real Hilbert space is computably isometrically isomorphic to ${\mathbb{R}}^d$ for some $d \in {\mathbb{N}}$. Every infinite dimensional real Hilbert space is computably isometrically isomorphic to $\ell^2$. Let us now introduce some basic notions from the theory of Weihrauch degrees. We will treat this paragraph somewhat informally, as we will not need to develop the theory very far. A formal and comprehensive treatment of everything stated here can be found in [@WeihrauchDegrees; @ClosedChoice; @EffectiveChoice], and in [@ConnectedChoice], where the Weihrauch degree of Brouwer’s fixed point theorem is determined. Let ${{\langle \cdot,\cdot\rangle}\colon {\mathbb{N}}^{\mathbb{N}}\times{\mathbb{N}}^{\mathbb{N}}\to {\mathbb{N}}^{\mathbb{N}}}$ denote some computable pairing function on Baire Space. A multimapping ${g\colon X \rightrightarrows Y}$ between represented spaces $X$ and $Y$ is said to *Weihrauch reduce* to ${h\colon Z \rightrightarrows W}$, in symbols ${g \leq_W h}$, if there exist computable functions ${K,N\colon\subseteq {\mathbb{N}}^{\mathbb{N}}\to {\mathbb{N}}^{\mathbb{N}}}$ such that ${K{\langle HN,{\operatorname{id}}\rangle}}$ is a realiser of $g$, whenever $H$ is a realiser of $h$. If $f \leq_W g$ and $g \leq_Wf$ we say that $g$ and $f$ are *Weihrauch equivalent* and write $f \equiv_W g$. The equivalence classes with respect to $\equiv_W$ are called *Weihrauch degrees*. The Weihrauch degrees together with the $\leq_W$-relation are known to form a bounded lattice. A very important and useful tool for studying Weihrauch degrees are so-called *closed choice principles* on represented spaces. Let $X$ be a represented space. 1. The *closed choice principle* on $X$ is the multimapping $${\operatorname{C}}_X\colon {\mathcal{A}}(X)\setminus\{\emptyset\} \rightrightarrows X,\; A \mapsto A.$$ The *unique choice principle* ${\operatorname{UC}}_X$ on $X$ is ${\operatorname{C}}_X$ restricted to singleton sets and the *connected choice principle* ${\operatorname{CC}}_X$ is ${\operatorname{C}}_X$ restricted to connected sets. 2. Let $X$ additionally be a closed subset of a computable Banach space $E$. We define the *convex choice principle* ${\operatorname{ConvC}}_X$ as the restriction of ${\operatorname{C}}_X$ to the space ${\mathcal{A}}^{{\operatorname{co}}}(X)$ of convex closed subsets of $X$. Let us now introduce some concrete Weihrauch degrees that will be useful in our further studies. The limit operator $\lim\colon\subseteq{\mathbb{N}}^{\mathbb{N}}\to{\mathbb{N}}^{\mathbb{N}}$ takes as input a (suitably encoded) convergent sequence $(p_n)_n \in ({\mathbb{N}}^{\mathbb{N}})^{\mathbb{N}}\simeq {\mathbb{N}}^{\mathbb{N}}$ and outputs its limit. Weak Kőnig’s Lemma, ${\operatorname{WKL}}$, takes as input an infinite binary tree and outputs an infinite path. The intermediate value theorem, ${\operatorname{IVT}}$, takes as input a continuous function $f\colon [0,1] \to {\mathbb{R}}$ with $f(0)\cdot f(1) < 0$ and outputs some point $x \in [0,1]$ such that $f(x) = 0$. Brouwer’s fixed point theorem in $n$-dimensional space, ${\operatorname{BFT}}_n$, takes as input a continuous function $f\colon [0,1]^n\to[0,1]^n$ and outputs some fixed point of $f$. Their relation is summarised in the following \[fact on Weihrauch degrees\] 1. ${\operatorname{WKL}}\equiv_W {\operatorname{C}}_{\{0,1\}^\omega}$. 2. ${\operatorname{CC}}_{[0,1]^n} \equiv_W {\operatorname{BFT}}_n \leq_W {\operatorname{WKL}}$ for all $n$. 3. ${\operatorname{IVT}}\equiv_W {\operatorname{ConvC}}_{[0,1]} \equiv_W {\operatorname{CC}}_{[0,1]} \equiv_W {\operatorname{BFT}}_1 <_W {\operatorname{BFT}}_2 \leq_W {\operatorname{BFT}}_3 \equiv_W {\operatorname{WKL}}$. 4. ${\operatorname{WKL}}<_W \lim <_W {\operatorname{C}}_{{\mathbb{N}}^{\mathbb{N}}}$. An important property of computably compact spaces is that their closed choice principle is of low degree. \[thm: compact choice below WKL\] Let $K$ be a computably compact represented topological space. Then the multimapping $${\operatorname{C}}_K\colon {\mathcal{A}}(K)\setminus\{\emptyset\} \rightrightarrows K, \; A\mapsto A$$ satisfies ${\operatorname{C}}_K \leq_W {\operatorname{WKL}}$. Finally, we need a few observations from (computable) functional analysis. The first theorem is the so-called [*projection theorem*]{}, which can be found in virtually any functional analysis textbook (cf. e.g. [@Werner Satz V.3.2 & Lemma V.3.3]). \[thm: projection theorem\] 1. Let $E$ be a uniformly convex real Banach space and let $K \subseteq E$ be nonempty, closed, and convex. For every $x \in E$ there exists a unique $y \in K$ such that $$d(x,y) = d(x,K) = \inf \{d(x,z)\;|\; z\in K\}.$$ We denote this element by $P_K(x)$. The mapping $P_K$ is a continuous retraction onto $K$, called the *metric projection*. 2. Let $H$ be a real Hilbert space and $K \subseteq H$ be nonempty, closed, and convex. Then $P_K$ is a nonexpansive mapping, and for all $x \in H$ the element $P_K(x)$ is characterised by the *variational inequality* $${(x - P_K(x),y - P_K(x))} \leq 0 \;\;\text{for all }y \in K.$$ <!-- --> 1. We may assume that $x \notin K$ and $x = 0$. Put $r = \inf\{{||z||} \;\big|\; {z \in K}\} > 0$. Let $(y_n)_n$ be a sequence in $K$ with $\lim_{n \to \infty}{||y_n||} = r$. We show that $(y_n)_n$ is a Cauchy sequence. Let $\varepsilon > 0$. There exists $n \in {\mathbb{N}}$ such that ${||y_{n + k}||} < r + \varepsilon$ for all $k \geq 0$. We have $\tfrac{y_n + y_{n + k}}{2} \in K$ for all $k \geq 0$ and thus ${||\tfrac{y_n + y_{n + k}}{2}||} \geq r$. If $\delta \in (0,1]$ satisfies $\eta(\delta) > \tfrac{\varepsilon}{r + \varepsilon}$, then $${\Big|\Big|\frac{y_n + y_{n + k}}{2}\Big|\Big|} \geq r > \left(r + \varepsilon\right)\left(1 - \eta(\delta)\right).$$ Applying the contraposition of uniform convexity to $\tfrac{y_{n + k}}{r + \varepsilon}$ and $\tfrac{y_n}{r + \varepsilon}$ thus yields $${||y_n - y_{n + k}||} < \left(r + \varepsilon\right) \delta.$$ Since $\tfrac{\varepsilon}{r + \varepsilon} \to 0$ as $\varepsilon \to 0$, and $\eta(\delta) > 0$ for all $\delta \in (0,1]$, it follows that ${||y_n - y_{n + k}||} \to 0$ as $n \to \infty$, i.e. the sequence $(y_n)_n$ is a Cauchy sequence. Since $K$ is closed and $E$ is complete, it converges to some element $y \in K$, which satisfies ${||y||} = \lim_{n\to\infty}{||y_n||} = r$. Suppose that the points $y_1, y_2 \in K$ with $y_1 \neq y_2$ satisfy ${||y_i||} = r > 0$. Then we have $y_1, y_2 \in {\overline{B}}(0,r)$. Since $K$ is convex, $\tfrac{y_1 + y_2}{2} \in K$ and since $E$ is strictly convex and $y_1 \neq y_2$, we have ${||\tfrac{y_1 + y_2}{2}||} < r$. Contradiction. 2. On one hand we have for all $\alpha \in [0,1]$ and $z \in K$: $$\begin{aligned} {||x - P_K(x)||}^2 &\leq {||x - \left(\alpha z + (1 - \alpha) P_K(x)\right)||}^2 \\ &= {(x - P_K(x) - \alpha(z - P_K(x)),x - P_K(x) - \alpha(z - P_K(x)))}\\ &= {||x - P_K(x)||}^2 - 2\alpha{(x - P_K(x),z - P_K(x))} + \alpha^2{||z - P_K(x)||}^2\end{aligned}$$ and thus ${(x - P_K(x),z - P_K(x))} \leq \frac{\alpha}{2} {||z - P_K(x)||}^2$ for all $\alpha \in (0,1]$, which shows that $P_K$ satisfies the variational inequality. On the other hand, if $p \in K$ satisfies ${(x - p,z - p)} \leq 0$ for all $z \in K$, then for all $z \in K$ we have $$\begin{aligned} {||x - z||}^2 &= {||(x - p) + (p - z)||}^2 \\ &= {||x - p||}^2 + 2{(x - p,p - z)} + {||p - z||}^2\\ &\geq {||x - p||}^2\end{aligned}$$ and thus $p = P_K(x)$ by Theorem \[thm: projection theorem\]. It remains to show that $P_K$ is nonexpansive. Let $x,y \in H$. We may assume that $x \neq y$ and $P_K(x) \neq P_K(y)$. Since $P_K(x), P_K(y) \in K$, we may use the variational inequality to obtain $${(P_K(y) - P_K(x),x - P_K(x))} \leq 0$$ and $${(P_K(x) - P_K(y),y - P_K(y))} \leq 0.$$ Adding both inequalities yields $${(P_K(y) - P_K(x),x - y + P_K(y) - P_K(x))} \leq 0.$$ And hence $${||P_K(y) - P_K(x)||}^2 \leq {(P_K(x) - P_K(y),x - y)} \leq {||P_K(x) - P_K(y)||}\cdot{||x - y||},$$ where the last inequality is the Cauchy-Schwarz inequality. We thus obtain $${||P_K(x) - P_K(y)||} \leq {||x - y||}.\eqno{\qEd}$$ The next important result is that projections onto located convex sets in uniformly convex computable Banach spaces are computable relative to a modulus of convexity. This will follow from a highly uniform proof mining result due to Kohlenbach: \[thm: modulus of uniqueness for projection\] There exists a computable functional $$\Phi\colon {\mathbb{N}}^{{\mathbb{N}}}\times{\mathbb{N}}\times{\mathbb{N}}\to {\mathbb{N}}$$ such that if $E$ is a uniformly convex normed space with modulus of uniform convexity $\eta_E$, $\mu$ is any functional satisfying $2^{-\mu(n)} \leq \eta_E(2^{-n})$, $K \subseteq E$ is nonempty, closed, and convex, and $x \in E$ with $d(x,K) \leq d$ then $$\phi(n) = \Phi(\mu,d,n)$$ is a modulus of uniqueness for the projection onto $K$. This means that if $p, q \in K$ satisfy ${||p - x||} \leq d(x,K) + 2^{-\phi(n)}$ and ${||q - x||} \leq d(x,K) + 2^{-\phi(n)}$, then ${||p - q||} < 2^{-n}$. \[cor: computability of projection on subset\] Let $E$ be a uniformly convex computable Banach space, let $C \subseteq E$ be nonempty, convex, and computably overt. Let $\eta_E$ be a modulus of uniform convexity for $E$. Let ${\mathcal{A}}^{{\operatorname{co}}}_{{\operatorname{dist}}}(C)$ denote the represented space of nonempty convex closed subsets of $C$, represented via their distance function. Then the mapping $$P\colon {\mathcal{A}}^{{\operatorname{co}}}_{{\operatorname{dist}}}(C) \to {\operatorname{\mathcal{C}}}(C,C),\; K \mapsto P_K,$$ is computable relative to $\eta_E$. In fact it is computable relative to any $\mu\colon{\mathbb{N}}\to{\mathbb{N}}$ satisfying $2^{-\mu(n)} \leq \eta_E(2^{-n})$. Let $\mu\colon{\mathbb{N}}\to{\mathbb{N}}$ be such that $2^{-\mu(n)} \leq \eta_E(2^{-n})$ and let $\Phi$ be the functional from Theorem \[thm: modulus of uniqueness for projection\]. We are given a set $K \in {\mathcal{A}}^{{\operatorname{co}}}_{{\operatorname{dist}}}(C)$, a point $x \in C$ and a number $n \in {\mathbb{N}}$ and want to compute an approximation to $P_K(x)$ up to error $2^{-n}$. Since we are given the distance function to $K$, we can compute an integer upper bound $d$ to $d(x,K)$. Again using the distance function, we can compute a dense sequence in $K$. This allows us to find a point $p \in K$ with ${||p - x||} \leq d(x,K) + 2^{-\Phi(\mu,d,n)}$. It follows from Theorem \[thm: modulus of uniqueness for projection\] that ${||p - P_K(x)||} < 2^{-n}$. The special case where $E$ has a computable modulus of convexity and $C = E$ yields: \[cor: computability of projection on E\] Let $E$ be a uniformly convex computable Banach space with computable modulus of uniform convexity. Then the mapping $$P\colon {\mathcal{A}}^{{\operatorname{co}}}_{{\operatorname{dist}}}(E) \to {\operatorname{\mathcal{C}}}(E,E),\; K \mapsto P_K,$$ is computable. In particular, if $K \subseteq E$ is a nonempty located and convex set, then $P_K$ is $(\delta_E,\delta_E)$-computable. The (for our purpose) most important structural feature of fixed point sets of nonexpansive mappings in strictly convex Banach spaces is that they are always convex. This is a standard exercise in functional analysis. We will prove it here anyway, to give a simple example of a proof exploiting the convexity of the underlying space. \[prop: fixed point sets of nonexpansive mappings are convex\] Let $E$ be a strictly convex Banach space, let $K \subseteq E$ be nonempty, closed, bounded, and convex. Let $f\colon K \to K$ be nonexpansive. Then the set ${\operatorname{Fix}}(f)$ is convex. Since $f$ is continuous, it suffices to show that for each $x,y \in {\operatorname{Fix}}(f)$, the convex combination $\frac{x + y}{2}$ is again contained in ${\operatorname{Fix}}(f)$ (since then it follows that the set of dyadic convex combinations of $x$ and $y$, which is dense in the line segment joining $x$ and $y$, consists entirely of fixed points). Since $f$ is nonexpansive, we have $${\Big|\Big|f\left(\frac{x + y}{2}\right) - x\Big|\Big|} = {\Big|\Big|f\left(\frac{x + y}{2}\right) - f(x)\Big|\Big|} \leq {\Big|\Big|\frac{x - y}{2}\Big|\Big|}.$$ Similarly, ${||f(\tfrac{x + y}{2}) - y||} \leq \tfrac{1}{2}{||x - y||}$ and obviously the same inequality holds if we replace $f(\tfrac{x + y}{2})$ by $\tfrac{x + y}{2}$. Now, suppose that $a, b \in E$ satisfy $${||a - y||} \leq {\Big|\Big|\frac{x - y}{2}\Big|\Big|},\;{||a - x||} \leq {\Big|\Big|\frac{x - y}{2}\Big|\Big|}, \;{||b - x||} \leq {\Big|\Big|\frac{x - y}{2}\Big|\Big|},\;{||b - y||} \leq {\Big|\Big|\frac{x - y}{2}\Big|\Big|}.$$ Then, if $a \neq b$, strict convexity yields $${\Big|\Big|\frac{a + b}{2} - y\Big|\Big|} < {\Big|\Big|\frac{x - y}{2}\Big|\Big|}\;\text{ and }\;{\Big|\Big|\frac{a + b}{2} - x\Big|\Big|} < {\Big|\Big|\frac{x - y}{2}\Big|\Big|}$$ and hence $${||x - y||} \leq {\Big|\Big|x - \frac{a + b}{2}\Big|\Big|} + {\Big|\Big|y - \frac{a + b}{2}\Big|\Big|} < {||x - y||},$$ contradiction. It follows, that $f(\tfrac{x + y}{2}) = \tfrac{x + y}{2}$. Computability of Fixed points and Rates of Convergence ====================================================== In this section we study the computability-theoretic complexity of the problems of finding fixed points of nonexpansive mappings on compact domains, and of obtaining rates of convergence of certain fixed point iterations. Let us first state some natural computational problems associated with the fixed point properties of nonexpansive mappings and determine their rough relation. \[Def: fixed point iteration\] Let $E$ be a uniformly convex real Banach space, let $K \subseteq E$ be a nonempty subset of $K$, and let ${\operatorname{\mathcal{N}}}(K)$ denote the set of nonexpansive self-maps of $K$. A *fixed point iteration* on $K$ is a mapping $I\colon {\operatorname{\mathcal{N}}}(K)\times K \to K^{{\mathbb{N}}}$ such that for all $f \in {\operatorname{\mathcal{N}}}(K)$, $x \in K$, we have $\lim_{n \to \infty} I(f,x)(n) \in {\operatorname{Fix}}(f)$. \[Def: computational problems\] Let $E$ be a uniformly convex computable Banach space, let $K \subseteq E$ be nonempty, co-semi-decidable, computably overt, bounded, and convex. Let ${\operatorname{\mathcal{N}}}(K)$ be the represented space of nonexpansive self-maps of $K$ with representation $[\delta_K\to\delta_K]\Big|^{{\operatorname{\mathcal{N}}}(K)}$. Let $I\colon {\operatorname{\mathcal{N}}}(K) \times K \to K^{{\mathbb{N}}}$ be a computable fixed point iteration. Consider the following computational problems: 1. The *realiser problem for the Browder-Göhde-Kirk theorem* ${\operatorname{BGK}}_K$: Given a nonexpansive function $f\colon K \to K$, output a fixed point for $f$. More formally: $${\operatorname{BGK}}_K\colon {\operatorname{\mathcal{N}}}(K) \rightrightarrows K,\; f \mapsto {\operatorname{Fix}}(f).$$ 2. The *projection problem* ${\operatorname{Proj}}_K$: Given a nonexpansive function $f\colon K \to K$, and a point $x\in K$ output the metric projection of $x$ onto ${\operatorname{Fix}}(f)$. More formally: $${\operatorname{Proj}}_K\colon {\operatorname{\mathcal{N}}}(K)\times K \to K,\; (f,x) \mapsto P_{{\operatorname{Fix}}(f)}(x).$$ 3. The *limit problem* $\lim(I)$ for $I$: given a nonexpansive function ${f\colon K \to K}$ and a starting point $x \in K$, output $\lim_{n \in {\mathbb{N}}} I(f,x)(n)$. More formally: $$\lim(I)\colon {\operatorname{\mathcal{N}}}(K)\times K \to K,\; (f,x) \mapsto \lim_{n\to \infty} I(f,x)(n).$$ 4. The *rate of convergence problem* ${\operatorname{Conv}}_{I}$ for $I$: given a nonexpansive function ${f\colon K \to K}$ and a starting point $x \in K$, output a rate of convergence of the sequence $(I(f,x)(n))_n$. More formally: $$\begin{aligned} {\operatorname{Conv}}_I\colon& {\operatorname{\mathcal{N}}}(K)\times K \rightrightarrows {\mathbb{N}}^{\mathbb{N}},\\ &(f,x) \mapsto \{\varphi\in{\mathbb{N}}^{\mathbb{N}}\;|\; \forall n \in {\mathbb{N}}.\forall l \geq \varphi(n). {||I(f,x)(l) - \lim_{k \to \infty}I(f,x)(k)||} < 2^{-n}\}.\end{aligned}$$ Most fixed point iterations considered in the literature are of a far more particular form than just computable mappings. This can be exploited to obtain stronger uncomputability results for particular classes of fixed point iterations. We summarise some common properties. Let $E$ be a uniformly convex real Banach space, and let $K \subseteq E$ be nonempty, convex, closed, and bounded. Let $I\colon {\operatorname{\mathcal{N}}}(K)\times K \to K^{{\mathbb{N}}}$ be a fixed point iteration. 1. $I$ is called [*projective*]{} if for all $f\in {\operatorname{\mathcal{N}}}(K)$ and $x \in K$, the limit $\lim_{n \to \infty}I(f,x)(n)$ is the unique fixed point of $f$ which is closest to $x$. 2. $I$ is called [*retractive*]{} if for all $f\in\! {\operatorname{\mathcal{N}}}(K)$ and $x \in {\operatorname{Fix}}(f)$, we have ${\lim_{n \to \infty}I(f,x)(n) = x}$. 3. $I$ is called [*avoidant*]{} if for all $f \in {\operatorname{\mathcal{N}}}(K)$ and $x \in K$, we have the implication $$\left(\exists n. f\left(I(f,x)(n)\right) = I(f,x)(n) \right) \Rightarrow f(x) = x.$$ 4. $I$ is called [*simple*]{} if it is of the form $$\begin{aligned} &I(f,x)(0) = x \\ &I(f,x)(n + 1) = \sum_{k = 0}^n \alpha_k^nI(f,x)(k) + \sum_{j,k = 0}^n \beta_{j,k}^n f^{(j)}(I(f,x)(k)),\end{aligned}$$ with $\alpha_k^n \geq 0$ and $\beta_{k,j}^n \geq 0$ for all $k,n,j$. The notion of projectiveness is well-defined thanks to Theorem \[thm: projection theorem\] and Proposition \[prop: fixed point sets of nonexpansive mappings are convex\]. Any projective fixed point iteration is clearly retractive. Note that Halpern’s iteration (where by convention we always choose the anchor point to be equal to the starting point) is projective and simple and that the Krasnoselski-Mann iteration is simple, retractive, and avoidant. \[prop: simple and retractive in \[0,1\] is projective\] 1. Let $I\colon{\operatorname{\mathcal{N}}}([0,1])\times[0,1] \to [0,1]^{{\mathbb{N}}}$ be a simple and retractive fixed point iteration. Then $I$ is projective when restricted to the set of all monotonically increasing functions. 2. Let $E$ be a uniformly convex real Banach space, and let $K \subseteq E$ be nonempty, convex, closed, and bounded. Let $I\colon {\operatorname{\mathcal{N}}}(K)\times K \to K^{{\mathbb{N}}}$ be an avoidant fixed point iteration. Then for all nonexpansive $f\colon K\to K$ and $x \notin {\operatorname{Fix}}(f)$, $\lim_{n\to \infty} I(f,x)(n)$ is a point on the boundary of ${\operatorname{Fix}}(f)$. <!-- --> 1. By induction one easily verifies that if $x \leq y$, then $I(f,x)(n) \leq I(f,y)(n)$ for all monotonically increasing $f\colon[0,1]\to[0,1]$. It follows that $${\lim_{n\to\infty}I(f,x)(n) \leq \lim_{n\to\infty}I(f,y)(n)}.$$ By Proposition \[prop: fixed point sets of nonexpansive mappings are convex\], the set ${\operatorname{Fix}}(f)$ is an interval of the form $[a,b]$, possibly with $a = b$. If $x\leq a$, then, since the iteration is retractive, $$\lim_{n \to \infty} I(f,x)(n) \leq \lim_{n \to\infty} I(f,a)(n) = a,$$ so $\lim_{n \to \infty} I(f,x)(n) = a$. An analogous argument applies if $x \geq b$. It follows that the mapping $\lambda x.\lim_{n\to\infty} I(f,x)(n)$ is the metric projection onto $[a,b]$, i.e. the iteration is projective. 2. Is trivial. The following proposition establishes the more obvious relationships between the problems introduced in Definition \[Def: computational problems\]. \[prop: obvious Weihrauch reductions\] Let $E$ be a uniformly convex computable Banach space, let $K \subseteq E$ be nonempty, co-semi-decidable, computably overt, bounded, and convex. Then $${\operatorname{BGK}}_K \leq_W {\operatorname{Proj}}_K.$$ If $I\colon {\operatorname{\mathcal{N}}}(K) \times K \to K^{{\mathbb{N}}}$ is a computable fixed point iteration, then $${\operatorname{BGK}}_K \leq_W \lim(I) \leq_W {\operatorname{Conv}}_I \leq_W \lim.$$ If $I$ is projective, then $${\operatorname{Proj}}_K \leq_W \lim(I).\eqno{\qEd}$$ Next we prove a general upper bound on the Weihrauch degree of ${\operatorname{Proj}}_K$ (and thus of ${\operatorname{BGK}}_K$). We need two lemmas which constitute the main steps in Goebel’s proof [@goebelBGK] of the Browder-Göhde-Kirk theorem (see also the proof of [@Werner Theorem IV.7.13]). \[lem: goebels lemma\] Let $E$ be a uniformly convex Banach space, let $K\subseteq E$ be nonempty, convex, closed, and bounded. Then there exists a function $\varphi\colon(0,1)\to(0,1)$ with ${\lim_{\varepsilon \to 0} \varphi(\varepsilon) = 0}$, such that for every nonexpansive mapping $f\colon K \to K$ and all $x, y \in K$ we have the implication $$\left({||x - f(x)||} < \varepsilon \land {||y - f(y)||} < \varepsilon\right) \rightarrow {\Big|\Big|\frac{x + y}{2} - f\left(\frac{x + y}{2}\right)\Big|\Big|} < \varphi(\varepsilon) .\eqno{\qEd}$$ Actually, $\varphi(\varepsilon)$ is given by a very simple term involving $\varepsilon$ and the modulus of uniform convexity of $E$, but we do not need this fact here. \[lem: infimum lemma\] Let $E$ be a uniformly convex Banach space, let $K \subseteq E$ be nonempty, convex, closed, and bounded, and let $f\colon K \to K$ be nonexpansive. Let $A \subseteq K$ be nonempty, closed, and convex. Then $A$ intersects the fixed point set of $f$ if and only if $$\inf\{{||f(x) - x||} \;\big|\; x \in A\} = 0.$$ Clearly, if $A$ intersects the fixed point set of $f$, then $\inf\{{||f(x) - x||} \;\big|\; x \in A\} = 0$. On the other hand, suppose that $\inf\{{||f(x) - x||} \;\big|\; x \in A\} = 0$. Let $$\mu(s) = \inf\{{||f(x) - x||} \;\big|\; x \in A, {||x||} \leq s \}$$ and $$r = \inf\{s > 0 \; \big|\; \mu(s) = 0\}.$$ Since $K$ is bounded and $\inf\{{||f(x) - x||} \;\big|\; x \in A\} = 0$, $r$ is a well-defined real number. Let $(x_n)_n$ be a sequence in $A$ with $$\lim_{n \to \infty} {||f(x_n) - x_n||} = 0$$ and $$\lim_{n \to \infty}{||x_n||} = r.$$ We will show that $(x_n)_n$ is a Cauchy sequence. It then follows that $(x_n)_n$ converges to a fixed point, which proves the claim. Suppose that $(x_n)_n$ is not a Cauchy sequence. Then $r > 0$ and there exists $\varepsilon > 0$ and a subsequence $(y_n)_n$ of $(x_n)_n$ such that ${||y_{n + 1} - y_n||} \geq \varepsilon$ for all sufficiently large $n$. Let $2r \geq s > r$ be such that $$\left(1 - \eta_E\left(\frac{\varepsilon}{2r}\right)\right)s < r.$$ For $n$ sufficiently large we have ${||y_n||} \leq s$, so that we have the inequalities $${\Big|\Big|\frac{y_n}{s}\Big|\Big|} \leq 1,\; {\Big|\Big|\frac{y_{n + 1}}{s}\Big|\Big|} \leq 1,\; {\Big|\Big|\frac{y_n}{s} - \frac{y_{n + 1}}{s}\Big|\Big|} \geq \frac{\varepsilon}{s}.$$ Applying uniform convexity, we obtain $${\Big|\Big|\frac{y_n + y_{n + 1}}{2}\Big|\Big|} \leq s\left(1 - \eta_E\left(\frac{\varepsilon}{s}\right)\right),$$ which yields (using that without loss of generality, $\eta_E$ is monotonically increasing) $${\Big|\Big|\frac{y_n + y_{n + 1}}{2}\Big|\Big|} \leq s\left(1 - \eta_E\left(\frac{\varepsilon}{2r}\right)\right) < r.$$ Now, by Lemma \[lem: goebels lemma\] we have $$\lim_{n \to \infty}{\Big|\Big|\frac{y_n + y_{n + 1}}{2} - f\left(\frac{y_n + y_{n + 1}}{2}\right)\Big|\Big|} = 0.$$ This contradicts the minimality of $r$. Hence, $(x_n)_n$ is a Cauchy sequence. \[prop: initial upper bound on Proj\] Let $E$ be a uniformly convex computable Banach space, let $K \subseteq E$ be nonempty, co-semi-decidable, computably overt, bounded, and convex. Then we have $${\operatorname{Proj}}_K \leq_W \lim \circ \lim.$$ If $K$ is computably compact or $E$ is a Hilbert space, then $${\operatorname{Proj}}_K \leq_W \lim.$$ We are given as input a nonexpansive function $f\colon K \to K$ and a point $x \in K$ and want to obtain the point $p = P_{{\operatorname{Fix}}(f)}(x) \in K$. In the case where $E$ is a Hilbert space, we can use Halpern’s iteration (\[thm: Halpern’s theorem\]) to obtain a computable sequence converging to $p$ and apply $\lim$ to obtain $p$ itself. In the case where $K$ is computably compact, we can compute ${\operatorname{Fix}}(f)$ as an element of ${\mathcal{A}}_{{\operatorname{dist}}_<}(K)$ (see the discussion after Definition \[Def: distance representations\]). In particular we can compute a sequence $(d(x_n,{\operatorname{Fix}}(f))_{n \in {\mathbb{N}}}$, where $(x_n)_n$ is a computable dense sequence in $K$, as an element of ${\mathbb{R}}_{<}^{{\mathbb{N}}}$. Using the standard identification of ${\mathbb{N}}^{{\mathbb{N}}}$ with $\left({\mathbb{N}}^{{\mathbb{N}}}\right)^{{\mathbb{N}}}$, it is easy to see that we can use a single instance of $\lim$ to obtain countably many instances of $\lim$ in parallel (cf. also e.g. [@ClosedChoice] or [@WeihrauchDegrees]). Using $\lim$, we can hence compute the sequence $(d(x_n,{\operatorname{Fix}}(f))_{n \in {\mathbb{N}}}$ as an element of ${\mathbb{R}}^{{\mathbb{N}}}$, which allows us to compute $\lambda y.d(y,{\operatorname{Fix}}(f))$ as an element of ${\operatorname{\mathcal{C}}}(K,{\mathbb{R}})$, since we have $$\left|d(y,{\operatorname{Fix}}(f)) - d(z, {\operatorname{Fix}}(f)) \right| \leq d(y,z)$$ for all $y ,z \in K$. Independently, we can use the same instance of $\lim$ to obtain a modulus of uniform convexity $\eta_E$ for $E$: we may put $$\eta_E(\varepsilon) = \inf\Big\{1 - {\Big|\Big|\frac{x + y}{2}\Big|\Big|} \;\big|\; x,y \in B_E, {||x - y||} \geq \varepsilon \Big\},$$ and if $(x_n)_n$ is a computable dense sequence in $B_E$ (we may choose e.g. the sequence of rational points contained in the open unit ball) we have $$\eta_E(\varepsilon) = \inf\Big\{1 - {\Big|\Big|\frac{x_i + x_j}{2}\Big|\Big|} \;\big|\; i,j \in {\mathbb{N}}, {||x_i - x_j||} > \varepsilon \Big\},$$ which is clearly limit-computable[^3] in $\varepsilon$. This allows us to compute the restriction of $\eta_E$ to the rational numbers using countably many applications of $\lim$. In particular we can limit-compute a function $\mu\colon {\mathbb{N}}\to {\mathbb{N}}$ satisfying $2^{-\mu(n)} \leq \eta_E(2^{-n})$. Using the distance function of ${\operatorname{Fix}}(f)$ and the function $\mu$, we apply Corollary \[cor: computability of projection on subset\] to obtain the projection of $x$ onto ${\operatorname{Fix}}(f)$. In the general case, we cannot a-priori compute ${\operatorname{Fix}}(f)$ as an element of ${\mathcal{A}}_{{\operatorname{dist}}_<}(K)$, because of Proposition \[prop: equivalence lower semi-located and computably closed implies locally compact\]. We can however use $\lim$ to obtain ${\operatorname{Fix}}(f)$ as an element of ${\mathcal{A}}_{{\operatorname{dist}}_<}(K)$: since $K$ is computably overt, we can list all rational closed balls ${\overline{B}}(a,r)$ for which the open ball $B(a,r)$ intersects $K$. Given such a rational closed ball ${\overline{B}}(a,r)$ in $E$, we can compute a dense sequence in ${\overline{B}}(a,r) \cap K$: choose a computable dense sequence $(x_n)_n$ in $K$ and filter out those points $x_n$ which satisfy $d(x_n,a) < r$. Using the convexity of $K$ it is easy to see that the resulting sequence is dense in ${\overline{B}}(a,r) \cap K$. This allows us to limit-compute $$\inf\big\{{||f(x) - x||} \;\big|\; x \in {\overline{B}}(a,r) \cap K\big\}.$$ Again, we can do this for all suitable closed rational balls in parallel. We can then enumerate those balls ${\overline{B}}(a,r)$ satisfying $\inf\{{||f(x) - x||} \;\big|\; x \in {\overline{B}}(a,r) \cap K\} > 0$, which by Lemma \[lem: infimum lemma\] is equivalent to ${\overline{B}}(a,r) \cap {\operatorname{Fix}}(f) = \emptyset$. This allows us to compute the distance function to ${\operatorname{Fix}}(f)$ from below (cf. [@BrattkaPresser Theorem 3.9 (1)] or the proof of [@Weih Lemma 5.1.7]). Now we apply the limit-computable method used in the compact case above to obtain the projection. Since the Weihrauch degree of the composition of two limit-computable mappings is below $\lim\circ\lim$ (see e.g. Fact 8.2 in [@BolzanoWeierstrass]), the result follows. Note that if $E$ is finite dimensional, then $K$ is always computably compact, so that the stronger upper bound of Proposition \[prop: initial upper bound on Proj\] applies. We will show in Theorem \[thm: upper bound on Proj in smooth spaces\] that $\lim$ is an upper bound on ${\operatorname{Proj}}_K$ in all uniformly convex and uniformly smooth computable Banach spaces, and thus for instance in all $L^p$-spaces with $1 < p < \infty$. We now begin a discussion on the computability of fixed points and the Weihrauch degree of the Browder-Göhde-Kirk theorem. Proposition \[prop: fixed point sets of nonexpansive mappings are convex\] yields an immediate upper bound for the Weihrauch degree of ${\operatorname{BGK}}_K$. Let $E$ be a uniformly convex computable Banach space. Let $K \subseteq E$ be nonempty, co-semi-decidable, computably overt, bounded, and convex. Then we have ${\operatorname{BGK}}_K \leq_W {\operatorname{ConvC}}_K$. In finite dimension, this already implies that ${\operatorname{BGK}}$ is always strictly weaker than ${\operatorname{WKL}}$. In fact it is non-uniformly computable thanks to the following result due to Le Roux and Ziegler. \[thm: convex sets have computable points\] Let $E$ be a finite dimensional computable Banach space. Let $K \subseteq E$ be nonempty, co-semi-decidable, and convex. Then $K$ contains a computable point. We may assume that $K$ is compact, since we can always intersect $K$ with a sufficiently large closed ball. We may also assume that $E$ is represented by $\rho^d$ for some $d \in {\mathbb{N}}$. We proceed by induction on $\dim E$. If $\dim E = 1$, then $K$ is either a singleton and hence computable, or it is an interval and hence contains a rational point. If $\dim E = d$, then the projection of $K$ onto the $x$-axis is still nonempty, co-semi-decidable, and convex, and contains a computable point $x$ by induction hypothesis. Now, the intersection of $\{x\}\times{\mathbb{R}}^{d - 1}$ with $K$ is nonempty, convex, and co-semi-decidable, of dimension strictly smaller than $d$. Again by induction hypothesis, the intersection, and in particular $K$, contains a computable point. The above theorem even shows that $K$ has a dense subset of computable points, since the intersection of $K$ with a small rational ball is again co-semi-decidable and compact. A more uniform version, using Weihrauch degrees, has been given in [@PaulyLeRoux]. \[cor: fixed points in finite dimension are computable\] Let $E$ be a finite dimensional, strictly convex computable Banach space. Let $K \subseteq E$ be nonempty, computably overt, co-semi-decidable, bounded, and convex and let $f\colon K \to K$ be computable and nonexpansive. Then $f$ has a computable fixed point. Corollary \[cor: fixed points in finite dimension are computable\] in particular shows that the Browder-Göhde-Kirk theorem in finite dimension is strictly more effective than the (in this case more general) Brouwer fixed point theorem: a construction due to Orevkov [@Orevkov] and Baigger [@Baigger] shows that there exists a computable function on the unit square in ${\mathbb{R}}^2$ without computable fixed points, while in every finite dimension some fixed points whose existence is guaranteed by the Browder-Göhde-Kirk theorem are computable. Note that Corollary \[cor: fixed points in finite dimension are computable\] really only uses the fact that $f$ is computable and its fixed point set is convex. Thus, Theorem \[thm: convex sets have computable points\] presents a fairly general non-uniform computability result: if a computable equation on a finite dimensional space has a convex set of solutions, then it has a computable solution. A nontrivial application is based on the following result[^4]. A self-map $f\colon K \to K$ of a nonempty subset $K$ of a real Hilbert space $H$ is called *pseudocontractive* if it satisfies $${(f(x) - f(y),x - y)} \leq {||x - y||}^2$$ for all $x,y \in K$. Let $H$ be a real Hilbert space, let $K \subseteq H$ be nonempty, closed and convex, and let $f\colon K \to K$ be pseudocontractive. Then ${\operatorname{Fix}}(f)$ is closed and convex. \[cor: computability of fixed points of pseudocontractions\] Let $H$ be a finite dimensional computable Hilbert space. Let $K \subseteq H$ be nonempty, computably overt, co-semi-decidable, bounded, and convex, and let $f\colon K\to K$ be computable and pseudocontractive. Then $f$ has a computable fixed point. Corollary \[cor: computability of fixed points of pseudocontractions\] is strictly more general than the Euclidean version of Corollary \[cor: fixed points in finite dimension are computable\], as the following example, due to [@ChidumeMutangadura], shows. \[prop: computable Lipschitz pseudocontraction which is not nonexpansive\] There exists a computable Lipschitz-continuous pseudocontractive mapping on the unit ball of Euclidean ${\mathbb{R}}^2$, which is not nonexpansive. For $x = (x_1,x_2)$, we put $x^{\perp} = (-x_2,x_1)$. Let $$f(x) = \begin{cases} x + x^{\perp} &\text{ if }{||x||} \leq \tfrac{1}{2},\\ \tfrac{x}{{||x||}} - x + x^{\perp} &\text{ if }{||x||} \geq \tfrac{1}{2}.\end{cases}$$ It is easy to see that $f$ is computable and Lipschitz-continuous with Lipschitz constant $5$. The proof of pseudocontractiveness is somewhat technical and can be found in [@ChidumeMutangadura]. The function from Proposition \[prop: computable Lipschitz pseudocontraction which is not nonexpansive\] actually provides an example of a pseudocontractive mapping with a unique fixed point, for which the Krasnoselski-Mann iteration, which is guaranteed to converge for nonexpansive mappings, fails to converge. This is proved in [@ChidumeMutangadura]. In infinite dimension, there exist computable firmly nonexpansive mappings without computable fixed points, already on compact sets. A mapping $f\colon K \to K$ defined on a nonempty subset $K$ of a real Hilbert space $H$ is called *firmly nonexpansive* if it satisfies $${||f(x) - f(y)||}^2 \leq {(x - y,f(x) - f(y))}$$ for all $x, y \in K$. Clearly, every firmly nonexpansive mapping is nonexpansive. It is not difficult to see that a mapping is firmly nonexpansive if and only if it is of the form ${f(x) = \tfrac{1}{2}(x + g(x))}$, where $g$ is a nonexpansive mapping. Let $$\mathcal{H} = \{x \in \ell^2 \;|\; 0 \leq x(n) \leq 2^{-n} \;\text{for all }n \in {\mathbb{N}}\}$$ denote the Hilbert cube in $\ell^2$ (represented by $\delta_{\ell^2}\big|^{\mathcal{H}}$). \[thm: nonexpansive mapping on Hilbert cube without computable fixed points\] There exists a computable firmly nonexpansive mapping $f\colon \mathcal{H} \to \mathcal{H}$ without computable fixed points. Put $g_n(x) = (1-2^{-n})x$ and $h_n(x) = 2^{-n} + (1 - 2^{-n})x$. Then $(g_n)_n$ and $(h_n)_n$ are computable sequences of nonexpansive self-maps of $[0,1]$, satisfying $|g_n(x) - x| \leq 2^{-n}$, $|h_n(x) - x| \leq 2^{-n}$ for all $x \in [0,1]$, ${\operatorname{Fix}}(g_n) = \{0\}$, and ${\operatorname{Fix}}(h_n) = \{1\}$. Let $A,B \subseteq {\mathbb{N}}$ be two disjoint, recursively enumerable, and recursively inseparable sets. Let $\alpha$ be the Gödel number of an algorithm with halting set $A$ and $\beta$ be the Gödel number of an algorithm with halting set $B$. Consider the sequence of functions $(f_n)_n$ with $$f_n(x) = \begin{cases}g_i(x) &\text{ if }\alpha\text{ halts on input }n\text{ within }i\text{ steps,}\\ h_i(x) &\text{ if }\beta\text{ halts on input }n\text{ within }i\text{ steps,} \\ x &\text{ if both }\alpha\text{ and }\beta\text{ diverge on input }n. \end{cases}$$ Note that since $A$ and $B$ are disjoint, both $\alpha$ and $\beta$ cannot halt on the same input, so $f_n$ is well-defined. The sequence $(f_n)_n$ is a computable sequence: in order to compute $f_n(x)$ up to error $2^{-m}$, we simulate $\alpha$ and $\beta$ simultaneously on input $n$ for $m$ steps. If $\alpha$ (respectively $\beta$) halts within $k \leq m$ steps, we output $g_k(x)$ (respectively $h_k(x)$) up to error $2^{-m}$. If neither $\alpha$ nor $\beta$ halt after $m$ steps, we may output $x$ as an approximation, since $|g_{m + k}(x) - x| \leq 2^{-m-k}$ and $|h_{m + k}(x - x)| \leq 2^{-m-k}$ for all $k \geq 0$. Now, suppose there exists a computable sequence $(x_n)_n$ with $f_n(x_n) = x_n$. In order to arrive at a contradiction, we use $(x_n)_n$ to construct a computable set $S\subseteq {\mathbb{N}}$ separating $A$ and $B$. Membership for $S$ is decided as follows: for a given $n \in {\mathbb{N}}$, run the tests $x_n > 0$ and $x_n < 1$ simultaneously. At least one of the tests has to succeed. If the first test to succeed is $x_n > 0$, we decide that $n \notin S$. If the first test to succeed is $x_n < 1$, we decide that $n \in S$. Note that in the case where both $x_n > 0$ and $x_n < 1$, the outcome of the decision procedure may depend on the Cauchy sequence of dyadic rational numbers representing $x_n$. We claim that $S \supseteq A$ and ${{S}^C} \supseteq B$. If $n \in A$, then $\alpha$ halts on input $n$ after $i \in {\mathbb{N}}$ steps. So $f_n = g_i$, and thus $x_n = 0$. The test $x_n > 0$ will hence fail, while the test $x_n < 1$ will succeed and thus $n \in S$. If $n \in B$, then $\beta$ halts on input $n$ after $i \in {\mathbb{N}}$ steps, so $f_n = h_i$ and thus $x_n = 1$. It follows that $n \notin S$. So $S$ separates $A$ and $B$. Contradiction. Now define a nonexpansive mapping $g\colon\mathcal{H}\to\mathcal{H}$ via $g(x)(n) = 2^{-n}f_n(2^nx(n))$. Then $g$ is computable, for in order to compute an approximation to $g(x)$ in $\ell^2$ up to error $2^{-n}$, it suffices to compute the real numbers $g(x)(0),\dots,g(x)(n + 1)$ up to error $2^{-2n - 1}/(n + 2)$. Any fixed point for $g$ can be used to compute a sequence of fixed points for $(f_n)_n$. In particular, $g$ has no computable fixed points. In order to obtain a firmly nonexpansive mapping $f$ we put $f = \tfrac{1}{2}({\operatorname{id}}+ g)$. Theorem \[thm: nonexpansive mapping on Hilbert cube without computable fixed points\] in particular shows that fixed points of (firmly) nonexpansive mappings are not computable relative to any discrete advice. Let us now consider the computability of rates of convergence of certain fixed point iterations. While in infinite dimension, the non-uniform uncomputability of fixed points in particular implies that there exist computable mappings such that no computable fixed point iteration has a computable rate of convergence for any computable starting point, and Theorem \[uniform uncomputability of fixed points\] tells us that there is no general algorithm for obtaining rates of convergence uniformly in the input function and in the starting point, it might still be the case (at least in finite dimension) that there exists a computable fixed point iteration $I$ such that for every computable nonexpansive function $f$ there exists a computable starting point $x_0 \notin {\operatorname{Fix}}(f)$ such that the sequence $I(f,x_0)$ has a computable rate of convergence. This could still be practically relevant, since in a given practical scenario one might be able to exploit additional information on the input function in order to choose the starting point of the iteration in such a way that the rate of convergence becomes computable. We will however see that this fails to be the case for a large class of fixed point iterations, already on the compact unit interval. We prove a special case of our main result (Theorem \[main result\]) where the underlying set is the the compact unit interval $[0,1]$. Our theorem uniformly characterises the fixed point sets of computable nonexpansive self maps of $[0,1]$. We first recall an elementary fact (cf. [@Weih]). \[prop: co-semi-decidable intervals and their endpoints\] Let $\mathcal{I} = \{(a,b) \in ({\mathbb{R}}_<\times {\mathbb{R}}_>)\;|\; a \leq b\}$. Then the mapping $${\mathcal{A}}^{{\operatorname{co}}}({\mathbb{R}})\setminus\{\emptyset\} \to \mathcal{I},\; [a,b] \mapsto (a,b)$$ and its inverse $$\mathcal{I} \to {\mathcal{A}}^{{\operatorname{co}}}({\mathbb{R}})\setminus\{\emptyset\},\; (a,b) \mapsto [a,b]$$ are computable. \[thm: main theorem on \[0,1\]\] 1. The mapping $${\operatorname{Fix}}\colon {\operatorname{\mathcal{N}}}([0,1]) \to {\mathcal{A}}^{{\operatorname{co}}}([0,1])\setminus\{\emptyset\}, \; f \mapsto {\operatorname{Fix}}(f)$$ is computable. 2. And so is its multivalued inverse $${\operatorname{Fix}}^{-1}\colon {\mathcal{A}}^{{\operatorname{co}}}([0,1])\setminus\{\emptyset\} \rightrightarrows {\operatorname{\mathcal{N}}}([0,1]).$$ The first claim immediately follows from Proposition \[prop: fixed point sets of nonexpansive mappings are convex\] and the well-known result that the set of zeroes of a continuous mapping $f$ is co-semi-decidable in $f$. Let us now prove the second claim. Suppose we are given a nonempty, closed interval $[a,b] \in {\mathcal{A}}^{{\operatorname{co}}}([0,1])$. By Proposition \[prop: co-semi-decidable intervals and their endpoints\] we can compute a monotonically increasing list $(a_n)_n$ of rational numbers converging from below to $a$, and a monotonically decreasing list $(b_n)_n$ of rational numbers converging from above to $b$. We may assume without loss of generality that $a_n \geq 0$ and $b_n \leq 1$ for all $n \in {\mathbb{N}}$. From $(a_n)$ and $(b_n)$ we can compute a sequence $(f_n)_n$ of nonexpansive functions via $$f_n(x) = \begin{cases}a_n &\text{if }x \leq a_n\text{,}\\ x &\text{if }a_n \leq x \leq b_n\text{,} \\ b_n &\text{if }x\geq b_n\text{.}\end{cases}$$ Finally, we compute $$f(x) = \sum_{n \in {\mathbb{N}}} 2^{-n-1}f_n(x).$$ Then $f$ is nonexpansive, and maps $[0,1]$ into $[0,1]$. Let us now show that ${\operatorname{Fix}}(f) = [a,b]$. If $x \in [a,b]$ then $f_n(x) = x$ for all $n \in {\mathbb{N}}$, so ${f(x) = \sum_{n \in {\mathbb{N}}} 2^{-n - 1}x = x}$. Suppose now without loss of generality that $x < a$. Then $f_n(x) \geq x$ for all $n \in {\mathbb{N}}$, and there exists $m\in{\mathbb{N}}$ such that $x < a_m$ and hence $f_m(x) > x$. It follows that $$f(x) = \sum_{n \in {\mathbb{N}}} 2^{-n - 1}f_n(x) > x,$$ and hence $x\notin{\operatorname{Fix}}(f)$. An analogous argument applies if $x > b$. We may hence put ${{\operatorname{Fix}}^{-1}([a,b]) = f}$. From the proof we obtain the following non-uniform corollary, which is slightly stronger than the non-uniform version of Theorem \[thm: main theorem on \[0,1\]\]. \[cor: construction of firmly nonexpansive function with given fixed point set\] Let $[a,b] \subseteq [0,1]$ be a co-semi-decidable interval. Then there exists a monotonically increasing, firmly nonexpansive, computable function $f\colon [0,1] \to [0,1]$ with ${\operatorname{Fix}}(f) = [a,b]$. The algorithm we use in the proof of Theorem \[thm: main theorem on \[0,1\]\] to compute ${\operatorname{Fix}}^{-1}$ maps any nonempty co-semi-decidable interval to a monotonically increasing, nonexpansive function ${g\colon [0,1] \to [0,1]}$ with ${\operatorname{Fix}}(g) = [a,b]$. In order to obtain a firmly nonexpansive function $f$, we put $f(x) = \tfrac{1}{2}(x + g(x))$. Since by Proposition \[prop: co-semi-decidable intervals and their endpoints\] any left-r.e. number can be the left endpoint of a co-semi-decidable interval, we obtain the announced result together with Proposition \[prop: simple and retractive in \[0,1\] is projective\]. \[cor: hardness of rates of convergence\] Let $I\colon {\operatorname{\mathcal{N}}}([0,1])\times [0,1] \to [0,1]^{{\mathbb{N}}}$ be an either projective, or simple and retractive, or avoidant computable fixed point iteration. Let $\varepsilon > 0$. There exists a computable, firmly nonexpansive function $f\colon [0,1] \to [0,1]$ with $\operatorname{diam}({\operatorname{Fix}}(f)) < \varepsilon$ such that for no computable $x \notin {\operatorname{Fix}}(f)$, the sequence $I(f,x)$ has a computable rate of convergence. Let $a \in (0,1)$ be an uncomputable left-r.e. number and $b \in (0,1)$ be an uncomputable right-r.e. number with $|a - b| < \varepsilon$. Then the closed interval $[a,b] \subseteq [0,1]$ is co-semi-decidable by Proposition \[prop: co-semi-decidable intervals and their endpoints\]. Using Corollary \[cor: construction of firmly nonexpansive function with given fixed point set\] we obtain a monotonically increasing firmly nonexpansive function $f\colon[0,1]\to[0,1]$ with ${\operatorname{Fix}}(f) = [a,b]$. If $x \notin {\operatorname{Fix}}(f)$, then by Proposition \[prop: simple and retractive in \[0,1\] is projective\], we have $\lim_{n \to \infty}I(f,x)(n) \in \{a,b\}$. In particular, $\lim_{n \to \infty}I(f,x)(n)$ is uncomputable. Since $x$ is computable, the sequence $(I(f,x)(n))_n$ is a computable sequence of real numbers, so if it had a computable rate of convergence, its limit would be computable. This proves the claim. Both the Halpern iteration and the Krasnoselski-Mann iteration are simple and retractive, so Corollary \[cor: hardness of rates of convergence\] applies to them. Using Weihrauch degrees, we can state our present results more uniformly. We have ${\operatorname{Proj}}_{[0,1]} \equiv_W \lim$. If $I\colon {\operatorname{\mathcal{N}}}([0,1])\times [0,1] \to [0,1]^{{\mathbb{N}}}$ is an either projective, or simple and retractive, or avoidant computable fixed point iteration, then $\lim(I) \equiv_W {\operatorname{Conv}}_I \equiv_W \lim$. It is well known that the identity ${\operatorname{id}}\colon {\mathbb{R}}_{<} \to {\mathbb{R}}$ is Weihrauch-equivalent to $\lim$, even when restricted to the unit interval. Given $a \in {\mathbb{R}}_{<} \cap [0,1]$, we can compute the interval $[a,1] \in {\mathcal{A}}^{{\operatorname{co}}}([0,1])$, and hence, by Theorem \[thm: main theorem on \[0,1\]\], construct a nonexpansive function $f\colon [0,1] \to [0,1]$ with ${\operatorname{Fix}}(f) = [a,1]$. Now, ${\operatorname{Proj}}_{[0,1]}(f,0) = a$, so $\lim \leq_W {\operatorname{Proj}}_{[0,1]}$. Together with Proposition \[prop: obvious Weihrauch reductions\] we obtain ${\operatorname{Proj}}_{[0,1]} \equiv_W \lim$. If $I$ is a projective, simple and retractive, or avoidant, computable fixed point iteration, then by Proposition \[prop: simple and retractive in \[0,1\] is projective\] we obtain ${\operatorname{Proj}}_{[0,1]} \leq_W \lim(I)$, and thus $\lim \leq_W \lim(I) \leq_W {\operatorname{Conv}}_I \leq_W \lim$, i.e. $\lim(I) \equiv_W {\operatorname{Conv}}_I \equiv_W \lim$. \[prop: BGK in \[0,1\]\] ${\operatorname{BGK}}_{[0,1]}\equiv_w {\operatorname{ConvC}}_{[0,1]} \equiv_W {\operatorname{IVT}}\equiv_W {\operatorname{BFT}}_1$. The equivalence ${\operatorname{BGK}}_{[0,1]}\equiv_W {\operatorname{ConvC}}_{[0,1]}$ follows immediately from Theorem \[thm: main theorem on \[0,1\]\], the equivalence ${\operatorname{ConvC}}_{[0,1]} \equiv_W {\operatorname{IVT}}\equiv_W {\operatorname{BFT}}_1$ was already stated in Fact \[fact on Weihrauch degrees\]. So far, there seems to be a significant discrepancy between the computational content of the existence result ${\operatorname{BGK}}$ and the “constructive” theorems by Wittmann and Krasnoselski. We will see in Section \[Section: 5\] that this discrepancy disappears on non-compact domains in infinite-dimensional Hilbert space, where the Browder-Göhde-Kirk theorem is Weihrauch equivalent to $\lim$, and hence to Wittmann’s theorem. Weak Topologies =============== In order to be able to prove our main result in full generality, we have to introduce an admissible representation for the weak topology on a reflexive Banach space $E$. Such a representation has first been introduced by Brattka and Schröder [@BrattkaSchroeder]. We denote the continuous dual of a normed space $E$ by $E'$ and define the mapping $${(\cdot,\cdot)}\colon E\times E' \to {\mathbb{R}},\; (x,x') \mapsto x'(x).$$ \[Def: admissible representation for weak\* topology\] Let $E$ be a computable Banach space. The represented space $E'_w$ is the space $E'$, represented via the co-restriction of $[\delta_E\to\rho]$ to all continuous linear functionals. \[thm: admissibility of delta E’\_w\] The representation $[\delta_E\to\rho]\Big|^{E'}$ is admissible with respect to the weak\* topology on $E'$. Since the points of $E$, viewed as functionals on $E'$, separate the points of $E'$, the weak\* topology on $E'$ is Hausdorff. By Theorem \[thm: admissibility of delta E’\_w\], the space $E'_w$ is then a Hausdorff represented topological space, so that the space ${\mathcal{K}}(E'_w)$ is well-defined and coincides extensionally with the set of all weak\* compact subsets of $E'$. Note that this crucially relies on the separability of $E$, since since weak\* sequential compactness and weak\* compactness need not coincide on duals of inseparable spaces. The fact that they do coincide in the separable case also follows from the well-known fact that the weak\* topology on the dual space of a separable Banach space is metrisable on the unit ball (cf. Theorem \[thm: unit ball in weak\*-topology as a computable metric space\] below). Also note that the weak\* topology on $E'$ is in general not sequential (i.e. there exist sequentially weak\* closed sets which are not weak\* closed). Consequently, the spaces ${\mathcal{O}}(E_w')$ and ${\mathcal{A}}(E_w')$ do not coincide with the hyperspaces of weak\* open and weak\* closed sets respectively, but with the hyperspaces of weak\* sequentially open and weak\* sequentially closed sets. If $A \in {\mathcal{A}}(E'_w)$, we write $A_w$ for the represented space $(A,\delta_{E'_w}\big|^A)$ to emphasize the underlying representation. If $E$ is a reflexive real Banach space with computable dual $E'$, we obtain a canonical representation for $E$ with respect to the weak topology, by identifying $E$ with $E''$ and putting $E_w = (E')'_w$, i.e. $E_w$ is the represented space $E''$ with representation $[\delta_{E'}\to\rho]\big|^{E''}$ (using that in this case the weak\* topology on $E''$ coincides with the weak topology on $E$). Again, the space ${\mathcal{K}}(E_w)$ and the space of weakly compact subsets of $E$ coincide extensionally[^5], but the caveat on ${\mathcal{O}}(E'_W)$ and ${\mathcal{A}}(E'_w)$ also applies to ${\mathcal{O}}(E_w)$ and ${\mathcal{A}}(E_w)$. As in the case of $E'_w$, if $A \in {\mathcal{A}}(E_w)$, we write $A_w$ for the represented space $(A,\delta_{E_w})$. We will often use the adjective “weak” when referring to elements in hyperspaces constructed from $E_w$. For instance, we may call the computable points of ${\mathcal{A}}(E_w)$ “weakly co-semi-decidable” and the computable points of ${\mathcal{A}}(E'_w)$ “weak\* co-semi-decidable” etc. Note that in the definition of $E_w$ we only require $E'$, but not $E$ itself, to be a computable Banach space. By definition, the mapping ${(\cdot,\cdot)}$ is $(\delta_{E_w}\times\delta_{E'},\rho)$-computable. If both $E$ and $E'$ are computable Banach spaces, it is natural to require that the mapping ${(\cdot,\cdot)}$ be $(\delta_{E}\times\delta_{E'},\rho)$-computable, so that ${\operatorname{id}}\colon E_w \to E$ becomes computable (see for instance Theorem \[thm: unit ball in weak\*-topology as a computable metric space\] (iii), Corollary \[cor: Banach space is isomorphic to subspace of function space\] and Propositions \[prop: strongly overt implies weakly overt\], \[prop: weak and strong topology equivalent on compact sets\] and \[prop: computably weakly closed uniformly implies lower semi-located\] below). This is for instance the case for the spaces $L^p([0,1])$ with $1 < p < \infty$, since we have $(L^p([0,1]))' = L^q([0,1])$, where $\tfrac{1}{p} + \tfrac{1}{q} = 1$, and if $f \in L^p([0,1])$ and $g \in L^q([0,1])$, then ${(f,g)}$ is given by the effective formula $${(f,g)} = \int_0^1 f(x) g(x) \operatorname{dx}.$$ Next, we prove some basic properties of the space ${\mathcal{K}}(E'_w)$. We will need a few effective counterparts to classical results from functional analysis. The first is an effective version of the separable Banach-Alaoglou theorem, which was proved by Brattka [@BrattkaHahnBanach]. \[thm: unit ball in weak\*-topology as a computable metric space\] Let $E$ be a computable Banach space. Let $B_{E'_w}$ denote the unit ball in $E'$, viewed as a subset of the represented space $E_w'$ (thus bearing the weak\* topology). Then 1. $B_{E'_w} \in {\mathcal{K}}(E'_w)$. 2. More generally, let $K \subseteq E'_w$ be a co-semi-decidable subset of $E'_w$. If $K$ is bounded, then $K$ is computably weak\* compact. 3. If $E'$ is a computable Banach space and ${(\cdot,\cdot)}$ is $(\delta_{E}\times\delta_{E'},\rho)$-computable, then $B_{E'_w}$ admits the structure of a computably compact computable metric space. It is proved in [@BrattkaHahnBanach] that there exists a computable embedding $i\colon B_{E'_w} \to X$ into a computably compact computable metric space $X$, such that $i(B_{E'_w})$ is computably compact as a subset of $X$ and the partial inverse $i^{-1}\colon i(B_{E'_w}) \to B_{E'_w}$ is computable. It follows that $i(B_{E'_w})$ is a computably compact represented space, and $i$ induces a computable isomorphism between $B_{E'_w}$ and $i(B_{E'_w})$, so that $B_{E'_w}$ is computably compact as a represented space. It follows from Proposition \[prop: compact subspace of represented space is compact set\] that $B_{E'_w}$ is a computably compact subset of $E_w'$. This proves the first claim. For the second claim, observe that the mappings $$\operatorname{mult}\colon (0,\infty)\times {\mathcal{A}}(E_w') \to {\mathcal{A}}(E'_w),\; (\alpha,A) \mapsto \alpha A = \{\alpha x\;\big|\; x \in A \}$$ and $$\operatorname{mult}\colon (0,\infty)\times {\mathcal{K}}(E_w') \to {\mathcal{K}}(E'_w),\; (\alpha,K) \mapsto \alpha K = \{\alpha x\;\big|\; x \in K \}$$ are computable. Proposition \[prop: closed and compact sets, (i) closed set is compact, (ii) compact set in Hausdorff space is closed\] (i) asserts that $$f\colon {\mathcal{A}}(B_{E'_w})\to{\mathcal{K}}(B_{E'_w}),\; A \mapsto A$$ is computable. Trivially, given $A\in {\mathcal{A}}(E_w')$, such that $A \subseteq B_{E'_w}$, we can compute $A$ as a set in ${\mathcal{A}}(B_{E'_w})$. Given a bounded set $A \in {\mathcal{A}}(E'_w)$ with bound $b$, we hence obtain $A$ as an element of ${\mathcal{K}}(E'_w)$ by computing $\operatorname{mult}(b,f(\operatorname{mult}(\tfrac{1}{b},A)))$. The third claim follows immediately from the proof of the first. We may pull back the metric $d_X$ on $X$ via $i$ to obtain a metric on $B_{E'_w}$, i.e. put $d(x,y) = d_X(i(x),i(y))$ for $x, y \in B_{E'_w}$. As the set of rational points in $B_{E'_w}$ we may choose those rational points of the computable Banach space $E'$ whose norm is strictly smaller than one. One now easily verifies that $B_{E'_w}$ is computably compact as a computable metric space, and that the Cauchy representation on $B_{E_w'}$ is computably equivalent to $[\delta_E\to\rho]\big|^{B_{E'_w}}$. As a corollary we get an effective version of a classical result in functional analysis (cf. [@Werner Korollar VIII.3.13]) in the reflexive case. \[cor: Banach space is isomorphic to subspace of function space\] Let $E$ be a reflexive computable Banach space with computable dual $E'$, such that the mapping ${(\cdot,\cdot)}$ is $(\delta_E\times\delta_{E'},\rho)$-computable. Then $E$ is computably isometrically isomorphic to a co-semi-decidable and computably overt subspace of a function space ${\operatorname{\mathcal{C}}}(M)$ over a computably compact metric space $M$. By Theorem \[thm: unit ball in weak\*-topology as a computable metric space\] (iii), $M = B_{E'_w}$ (with the weak\* topology) is a computably compact computable metric space. We show that the mapping $$i\colon E \to {\operatorname{\mathcal{C}}}(M), \; x \mapsto \lambda x'.(x,x')$$ is a $(\delta_E,[\delta_{B_{E'_w}}\to\rho])$-computable isometric embedding with co-semi-decidable and computably overt image. The $(\delta_E,[\delta_{B_{E'_w}}\to\rho])$-computability is obvious, and the fact that it is an isometry follows from ${||x||} = \sup_{x' \in B_{E'}}|{(x,x')}|$, which in turn is an easy corollary of the Hahn-Banach theorem (cf. e.g. [@Werner Korollar III.1.7]). Clearly, $i(E)$ is computably overt. It is also co-semi-decidable, for if we are given a continuous function $f$ on $B_{E'_w}$, we can verify if it is nonlinear. It remains to show that its inverse $i^{-1}\colon i(E) \to E$ is computable. Given a $[\delta_{B_{E'_w}}\to\rho]$-name of $i(x) \in i(E)$ and a $\nu_E$-name of a rational point $y \in E$ we can compute a $[\delta_{B_{E'_w}}\to\rho]$-name of $i(x) - i(y)$, and hence compute $$\max\{|{(x - y,x')}|\;\big|\;x' \in B_{E'_w}\} = {||i(x) - i(y)||} = {||x - y||},$$ using Theorem \[thm: supremum on compact space\]. For every $n \in {\mathbb{N}}$ we may hence search for a rational point $y_n$ in $E$ satisfying ${||y_n - x||} < 2^{-n}$, which allows us to compute a $\delta_E$-name of $x$. \[Rem: corollary five in Brattka Schroeder wrong\] In Corollary 5 in [@BrattkaSchroeder], it is claimed that the representation $\delta^w_E$ of $E$ defined by $$\delta^w_E(p) = x \;:\Leftrightarrow\; [\delta'_E\to\rho](p) = \iota(x),$$ where $\iota\colon E\to E''$ is the canonical embedding and $\delta'_E = [\delta_E\to\rho]\Big|^{E'}$, is admissible for the weak topology on $E$. This contradicts the proof of our Corollary \[cor: Banach space is isomorphic to subspace of function space\], which suggests that $\delta^w_E$ is admissible with respect to the norm topology. To convince ourselves that the claim is false, we consider the simple example of $E = \ell^2$. For the scope of this remark we will adopt the notation used in [@BrattkaSchroeder]. The representation $\delta^{\geq}_{\ell^2}$, where a $\delta^{\geq}_{\ell^2}$-name of $(x_n)_n \in \ell^2$ is a $\rho^\omega$-name of a sequence $(b,x_1,x_2,\dots)$ with $b \geq {||(x_n)_n||}$, is equivalent to $\delta_{\ell^2_w}$ (which is denoted by $\delta_{\ell^2}'$ in [@BrattkaSchroeder]). It follows that given a $\delta^w_E(p)$-name of $x \in \ell^2$ we can compute $(x,x')$ for any $x' \in \ell^2$, provided that we know $x'(n)$ for all $n \in {\mathbb{N}}$ and some bound on ${||x'||}$. In particular we can compute $x(n) = (x,e_n)$ for every $n \in {\mathbb{N}}$. Since $\delta_{\ell^2} \leq \delta_{\ell^2_w}$, a $\delta^w_{\ell^2}$-name of $x$ allows us to compute a $[\delta_{\ell^2}\to\rho]$-name of $x$. Theorem 5.1. in [@Brattka06computableversions] then asserts that we can compute some bound $b$ on ${||x||}$. It follows that we can compute ${||x||}^2 = {(x,x)}$, and so $\delta^w_{\ell^2} \equiv \delta_{\ell^2}^{=}$, which entails that in fact $\delta^w_{\ell^2}$ is admissible with respect to the (strictly stronger) norm topology. The flaw in the argument seems to be the claim that for every *compatible* representation $\delta$ of a separable Banach space $X$, the dual representation $\delta' = [\delta \to \delta_{{\mathbb{F}}}]\big|^{X'}$ is admissible with respect to the weak\* topology on $X'$, the reasoning being that for represented topological spaces $A$ and $B$, the canonical function space representation $[\delta_A\to\delta_B]$ is admissible with respect to the sequentially-compact-open topology on ${[A\rightarrow B]}$, and that weak\*-convergence on $X'$ coincides with compact-open-convergence on ${\operatorname{\mathcal{C}}}(X,{\mathbb{F}})$ (cf. Theorem \[thm: admissibility of delta E’\_w\] and Proposition 1 in [@BrattkaSchroeder]). However, if $\delta$ is admissible with respect to the weak topology on $X$, then $\delta' = [\delta\to\rho]$ is admissible with respect to the weakly-compact-open topology and not necessarily with respect to the (norm-)compact-open topology. Thus, if $X$ is a reflexive computable Banach space, and we start with the standard representation $\delta_X$ of $X$, which is compatible and admissible with respect to the strong topology, then $\delta'_X = [\delta_{X}\to\rho]$ is compatible and admissible with respect to the compact-open topology on $X'$, which is just the weak\* topology. Applying the construction again, we see that $(\delta'_X)' = [\delta_{X}'\to\rho]$ is compatible and admissible with respect to the weak\*-compact-open topology, which in general is strictly stronger than the (norm-)compact-open-topology. Let us now turn to some special properties of convex sets. *Mazur’s lemma* asserts that a convex set is weakly sequentially closed if and only if it is strongly closed. \[thm: Mazur\] Let $E$ be a Banach space and $K \subseteq E$ be convex. If $(x_n)_n$ is a sequence in $K$ which converges weakly to $x \in E$, then there exists a sequence of finite convex combinations of the $x_n$’s, converging strongly to $x$. It follows that strongly overt convex sets are weakly overt. \[prop: strongly overt implies weakly overt\] Let $E$ be a reflexive computable Banach space with computable dual $E'$, such that the mapping ${(\cdot,\cdot)}$ is $(\delta_{E}\times\delta_{E'},\rho)$-computable. Then the identity $${\operatorname{id}}\colon {\mathcal{V}}^{{\operatorname{co}}}(E)\setminus\{\emptyset\} \to {\mathcal{V}}^{{\operatorname{co}}}(E_w)\setminus\{\emptyset\},$$ where ${\mathcal{V}}^{{\operatorname{co}}}(E)$ denotes the hyperspace of convex overt closed subsets of $E$, is well-defined and computable. The mapping ${\operatorname{id}}$ is well-defined by Theorem \[thm: Mazur\]. Since $E$ is a computable metric space and $E_w$ is separable, we may use the characterisation of overtness given in Proposition \[prop: overtness and recursive enumerability, metric case\] and the sufficient condition given in Proposition \[prop: overtness and recursive enumerability\]. If $K \in {\mathcal{V}}^{{\operatorname{co}}}(E)$, then by Proposition \[prop: overtness and recursive enumerability, metric case\], we can compute a $\delta_{E}^{\omega}$-name of a norm-dense sequence $(x_n)_n$ in $K$. Since the weak topology is coarser than the norm topology, the weak sequential closure of $(x_n)_n$ contains $K$, and by Theorem \[thm: Mazur\], any weak limit of $(x_n)_n$ is already contained in $K$, so that $K$ is the closure of $(x_n)_n$ with respect to the sequentialisation of the weak topology. Since ${(\cdot,\cdot)}$ is computable, we have $\delta_{E} \leq \delta_{E_w}$, so that we can compute a $\delta_{E_w}^{\omega}$-name of $(x_n)_n$. Thus we can compute $K$ as an element of ${\mathcal{V}}^{{\operatorname{co}}}(E_w)$ using Proposition \[prop: overtness and recursive enumerability\]. Propositions \[prop: strongly overt implies weakly overt\] and \[prop: overtness and recursive enumerability\] imply that a convex subset of a reflexive computable Banach space $E$ is weakly overt if and only if it has a computable norm-dense sequence. Finally, we prove a useful uniform characterisation of computably weakly compact convex sets in a reflexive Banach space $E$ with computable dual $E'$, which will be an important ingredient for the proof of our main result. Note that by Mazur’s lemma, a convex subset of $E$ is weakly compact if and only if it is closed and bounded (cf. also [@Megginson Proposition 2.8.1]). Let $E$ be a reflexive Banach space with computable dual $E'$. 1. A *rational half space* is a nonempty set of the form $$h = \{x \in E \;|\; {(x,x_h')} + a_h \leq 0 \}$$ where $x_h'$ is a rational point in $E'$ and $a_h \in {\mathbb{Q}}$. A $\nu_{{\operatorname{HB}}}$-name of a rational half space is a $\nu_{E'}\times\nu_{{\mathbb{Q}}}$-name of $(x_h',a_h) \in E'\times{\mathbb{Q}}$. 2. Let $K \subseteq E$ be closed, convex and bounded. A $\kappa_{{\operatorname{HB}}}$-name of $K$ is a $\nu_{{\operatorname{HB}}}^\omega\times\kappa^{E_w}$-name of all rational half spaces containing $K$ in their interior and a weakly compact set $L \in {\mathcal{K}}(E_w)$ containing $K$. Note that by Theorem \[thm: unit ball in weak\*-topology as a computable metric space\] (ii), a $\kappa_{{\operatorname{HB}}}$-name of a closed, convex and bounded set can be computed from a list of all rational half spaces containing $K$ in their interior and a rational bound on $\sup\{{||x||} \;\big|\; x \in K\}$. It may not be immediately obvious that $\kappa_{{\operatorname{HB}}}$ is a well-defined representation. This follows however from the following easy consequence of the Hahn-Banach separation theorem (cf. e.g. [@Werner Theorem III.2.5]). Let $E$ be a reflexive Banach space with computable dual $E'$. Let $K \subseteq E$ be closed, bounded and convex, let $x \notin K$. Then there exists a rational half space $h$ such that $K \subseteq {{h}^\circ}$ and $x \in {{h}^C}$. Obviously, the boundedness condition on $K$ cannot be dropped, as the example of a straight line with irrational slope in ${\mathbb{R}}^2$ shows. \[Hahn-Banach equivalent to co-r.e.\] Let $E$ be a reflexive Banach space with computable dual $E'$. Then we have $\kappa_{{\operatorname{HB}}} \equiv \left(\kappa^{E_w}\right)\bigr|^{{\mathcal{K}}^{{\operatorname{co}}}(E_w)}$, where ${\mathcal{K}}^{{\operatorname{co}}}(E_w)$ denotes the space of convex weakly compact subsets of $E_w$. $\left[\kappa_{{\operatorname{HB}}} \leq \left(\kappa^{E_w}\right)\bigr|^{{\mathcal{K}}^{{\operatorname{co}}}(E_w)}\right]$: Suppose we are given a $\kappa_{{\operatorname{HB}}}$-name of $K \in {\mathcal{K}}^{{\operatorname{co}}}(E_w)$. Since the name provides us with a weakly compact set $L \in {\mathcal{K}}(E_w)$ containing $K$, and since ${\operatorname{id}}\colon {\mathcal{A}}(L_w)\to {\mathcal{K}}(L_w)$ is computable by Proposition \[prop: closed and compact sets, (i) closed set is compact, (ii) compact set in Hausdorff space is closed\] (i), it suffices to show that we can compute a $\psi^{E_w}$-name of $K$. Given $x \in E_w$ and a $\nu_{{\operatorname{HB}}}$-name of a functional $f = {(\cdot,x_h')} + a_h$ we can compute $f(x) \in {\mathbb{R}}$. Now, if the sequence of half spaces containing $K$ given by the $\kappa_{{\operatorname{HB}}}$-name is defined by the sequence of affine linear functionals $(f_n)_n$, we can compute the characteristic function of $K^C$ into Sierpiński-space as follows: given $x \in E_w$, if there exists an $n \in {\mathbb{N}}$ such that $f_n(x) > 0$ output one, otherwise output zero. This shows that we can compute a $\psi^{E_w}$-name of $K$, which proves the claim.\ $\left[\left(\kappa^{E_w}\right)\bigr|^{{\mathcal{K}}^{{\operatorname{co}}}(E_w)} \leq \kappa_{{\operatorname{HB}}}\right]$: Suppose we are given a $\kappa^{E_w}$-name of $K$. We need to compute a weakly compact set $L \in {\mathcal{K}}(E_w)$ with $L \supseteq K$ and a list of all rational half spaces containing $K$ in their interior. Since $K$ contains itself and is given as a $\kappa^{E_w}$-name, we may put $L = K$, so that it suffices to show that we can enumerate all rational half spaces containing $K$ in their interior. We show that given a $\nu_{{\operatorname{HB}}}$-name of an affine linear functional (technically, of the half space defining the functional) $f\colon E \to {\mathbb{R}}$ of the form $f(x) = (x,x') + a$, we can verify if $f(x) < 0$ for all $x \in K$. We can computably translate the $\nu_{{\operatorname{HB}}}$-name of $f$ into a $\delta_{E'}$-name of $x'$. Then by definition of $E_w$, the mapping $f\colon E_w\to {\mathbb{R}}, x \mapsto (x,x') + a$ is computable. It follows that $U_f = \{x \in E_w \;\big|\; f(x) < 0\}$ is semi-decidable relative to $f$. By definition of $\kappa$, the relation $K \subseteq U_f$ is semi-decidable relative to $f$ as well, which proves the claim. The proof of Theorem \[Hahn-Banach equivalent to co-r.e.\] shows that the definition of $\kappa_{{\operatorname{HB}}}$ can be slightly relaxed. \[lem: listing enough half-spaces\] Let $E$ be a reflexive Banach space with computable dual $E'$. Define a new representation $\tilde{\kappa}_{{\operatorname{HB}}}$ of ${\mathcal{K}}^{{\operatorname{co}}}(E_w)$ as follows: a $\tilde{\kappa}_{{\operatorname{HB}}}$-name of $K \in {\mathcal{K}}^{{\operatorname{co}}}(E_w)$ is a $\nu_{{\operatorname{HB}}}^\omega\times\kappa^{E_w}$-name of a sequence $(h_n)_n$ of rational half spaces such that $K = \bigcap_{n \in {\mathbb{N}}} {{h}^\circ}_n$ and a weakly compact set $L \in {\mathcal{K}}(E_w)$ containing $K$. Then $\tilde{\kappa}_{{\operatorname{HB}}} \equiv \kappa_{{\operatorname{HB}}}$. Clearly, $\kappa_{{\operatorname{HB}}} \leq \tilde{\kappa}_{{\operatorname{HB}}}$. For the converse direction, note that the proof of the reduction $\kappa_{{\operatorname{HB}}} \leq \left(\kappa^{E_w}\right)\bigr|^{{\mathcal{K}}^{{\operatorname{co}}}(E_w)}$ in Theorem \[Hahn-Banach equivalent to co-r.e.\] actually establishes the stronger reduction ${\tilde{\kappa}_{{\operatorname{HB}}} \leq \left(\kappa^{E_w}\right)\bigr|^{{\mathcal{K}}^{{\operatorname{co}}}(E_w)}}$, so that we obtain the reduction chain $$\tilde{\kappa}_{{\operatorname{HB}}} \leq \left(\kappa^{E_w}\right)\bigr|^{{\mathcal{K}}^{{\operatorname{co}}}(E_w)} \leq \kappa_{{\operatorname{HB}}}$$ and thus $\tilde{\kappa}_{{\operatorname{HB}}} \equiv \kappa_{{\operatorname{HB}}}$. On compact subsets of a Banach space $E$, the weak topology and the norm topology coincide. This is effectively witnessed by our representation. \[prop: weak and strong topology equivalent on compact sets\] Let $E$ be a reflexive computable Banach space with computable dual, such that the mapping ${(\cdot,\cdot)}$ is $(\delta_{E}\times\delta_{E'},\rho)$-computable. Let $K\subseteq E$ be a computably compact and computably overt subset of $E$. Then we have $\delta_{E_w}\big|^K \equiv \delta_{E}\big|^K$. Since ${(\cdot,\cdot)}$ is computable, we have $\delta_{E} \leq \delta_{E_w}$, so we only have to show the converse reduction $\delta_{E_w}|^K \leq \delta_{E}|^K$. Define the represented spaces $$K = \left(K,\delta_{E}\big|^K\right)\; \text{ and }\; K_w = \left(K,\delta_{E_w}\big|^K\right).$$ Firstly, observe that $E_w$ is effectively Hausdorff, i.e. the mapping $$E_w \to {\mathcal{A}}(E_w), \;x \mapsto \{x\}$$ is computable: we can verify if two given elements in $E_w$ are different by comparing their values on the rational points of $E'$. It follows that the identity ${\operatorname{id}}\colon {\mathcal{O}}(K) \to {\mathcal{O}}(K_w)$ is computable via the following chain of maps: $$\begin{CD} {\mathcal{O}}(K) @>>> {\mathcal{A}}(K) @>(1)>> {\mathcal{K}}(K) @>(2)>> {\mathcal{K}}(K_w) @>(3)>> {\mathcal{A}}(K_w) @>>> {\mathcal{O}}(K_w)\\ U @>>> K\setminus U @>>> K\setminus U @>>> K_w\setminus U @>>> K_w\setminus U @>>> U. \end{CD}$$ The computability of (1) follows from the computable compactness of $K$ together with Proposition \[prop: closed and compact sets, (i) closed set is compact, (ii) compact set in Hausdorff space is closed\] (i). The computability of (2) can be derived from the computability of ${{\operatorname{id}}\colon{\mathcal{O}}(E_w)\to{\mathcal{O}}(E)}$, which in turn follows from the computability of ${\operatorname{id}}\colon E\to E_w$, and the computability of (3) follows from the fact that $E_w$ is effectively Hausdorff, together with Proposition \[prop: closed and compact sets, (i) closed set is compact, (ii) compact set in Hausdorff space is closed\] (ii). We then obtain the mapping ${\operatorname{id}}\colon K_w \to K$, i.e. the reduction $\delta_{E_w}|^K \leq \delta_{E}|^K$, via the following chain of maps: $$\begin{CD} K_w @>(4)>> {\mathcal{A}}(K_w) @>(5)>> {\mathcal{K}}(K_w) @>(6)>> {\mathcal{K}}(K) @>(7)>> K\\ x @>>> \{x\} @>>> \{x\} @>>> \{x\} @>>> x. \end{CD}$$ Mapping (4) is computable since $E_w$ is effectively Hausdorff. To establish the computability of (5), observe that the computability of (2) and the computable compactness of $K$ imply that $K \in {\mathcal{K}}(K_w)$ and apply Proposition \[prop: closed and compact sets, (i) closed set is compact, (ii) compact set in Hausdorff space is closed\] (i). The computability of (6) follows from the computability of ${\operatorname{id}}\colon {\mathcal{O}}(K)\to{\mathcal{O}}(K_w)$, which we have established above. For the computability of (7), observe that we can verify if a rational ball of the form $B(a,2^{-n})$ contains $\{x\}$, which yields a Cauchy sequence effectively converging to $x$ by exhaustive search over all rational balls. Finally, we observe that computably overt, co-semi-decidable subsets of $E'_w$ are (uniformly) located. The following proposition guarantees that this actually makes sense. Let $E$ be a Banach space. Let $A \subseteq E$ be weakly sequentially closed. Then $A$ is closed with respect to the norm topology. Since $E$ is a metric space, $A$ is closed if and only if it is sequentially closed. Let $(x_n)_n$ be a sequence in $A$ with limit $x \in E$. Then $x$ is a weak limit of $(x_n)_n$, so $x \in A$, since $A$ is weakly sequentially closed. It follows that $A$ is sequentially closed, and thus closed. \[prop: computably weakly closed uniformly implies lower semi-located\] Let $E$ be a reflexive computable Banach space with computable dual, such that the mapping ${(\cdot,\cdot)}$ is $(\delta_{E}\times\delta_{E'},\rho)$-computable. Then the canonical embedding $i\colon {\mathcal{A}}(E_w)\setminus\{\emptyset\} \to {\mathcal{A}}_{{\operatorname{dist}}_{<}}(E), A \mapsto A$ is computable. Given a sequentially weakly closed set $A \in {\mathcal{A}}(E_w)$, it suffices to show that we can uniformly computably enumerate all closed balls with rational centres and radii contained in the complement $A^C$ of $A$. The result then follows from [@BrattkaPresser Theorem 3.9 (1)]. The proof of Theorem \[thm: unit ball in weak\*-topology as a computable metric space\] (ii) allows us to uniformly translate a computable number $r \in {\mathbb{R}}$ into a name of ${\overline{B}}(0,r)$ as a weakly compact subset of $E_w$, i.e. the mapping $$(0,\infty) \to {\mathcal{K}}(E_w), \; r \mapsto {\overline{B}}(0,r)$$ is computable. It is easy to see that the mapping $$E_w\times E_w \to E_w,\; (x,c) \mapsto x + c$$ is computable. Hence, the mapping $$E_w \times (0,\infty) \to {\mathcal{K}}(E_w),\; (c,r) \mapsto {\overline{B}}(c,r)$$ is computable. It follows that the mapping $${\mathcal{A}}(E_w)\times E_w \times (0,\infty) \to {\mathbb{S}},\; (A,c,r) \mapsto \begin{cases}1 &\text{if }{\overline{B}}(c,r) \subseteq A^C\text{,}\\0 &\text{otherwise}\end{cases}$$ is computable. Since ${(\cdot,\cdot)}$ is computable, we have $\delta_{E}\leq\delta_{E_w}$, so that in particular the mapping $${\mathcal{A}}(E_w)\times E \times (0,\infty) \to {\mathbb{S}},\; (A,c,r) \mapsto \begin{cases}1 &\text{if }{\overline{B}}(c,r) \subseteq A^C\text{,}\\0 &\text{otherwise}\end{cases}$$ is computable. Using this mapping we can enumerate all rational closed balls contained in the complement of $A$. Proposition \[prop: computably weakly closed uniformly implies lower semi-located\] in particular implies that any nonempty weakly co-semi-decidable subset of $E$ is lower semi-located, and hence every nonempty weakly co-semi-decidable and computably overt subset of $E$ is located (which by Proposition \[prop: equivalence lower semi-located and computably closed implies locally compact\] is at least not uniformly true for co-semi-decidable subsets of $E$, if $E$ is infinite dimensional). Let us introduce some further Weihrauch degrees. Let $E$ be a computable Banach space with computable dual, such that the mapping ${(\cdot,\cdot)}$ is $(\delta_{E}\times\delta_{E'},\rho)$-computable. Let $A\subseteq E$ be nonempty and weakly closed. The *weak closed choice principle* ${\operatorname{C}}_A^{{\operatorname{w}}\to{\operatorname{w}}}$ on $A$ is the closed choice principle ${\operatorname{C}}_{A_w}$ on the represented space $A_w = (A,\delta_{E_w})$. The *weak-strong closed choice principle* is the multimapping $${\operatorname{C}}_A^{{\operatorname{w}}\to{\operatorname{n}}}\colon {\mathcal{A}}(A_w) \rightrightarrows A, \; S \mapsto S,$$ where the image is represented by $\delta_{E}|^A$. Similarly, we define ${\operatorname{ConvC}}_A^{{\operatorname{w}}\to{\operatorname{w}}}$, ${\operatorname{UC}}^{{\operatorname{w}}\to{\operatorname{w}}}_A$, ${\operatorname{ConvC}}^{{\operatorname{w}}\to{\operatorname{n}}}_A$ and ${\operatorname{UC}}^{{\operatorname{w}}\to{\operatorname{n}}}_A$. We may also define a (computationally) weaker version of the Browder-Göhde-Kirk theorem. Let $K$ be nonempty, computably overt, weakly co-semi-decidable, bounded and convex. The *weak Browder-Göhde-Kirk* theorem is the mapping $${\operatorname{WBGK}}_{K}\colon {\operatorname{\mathcal{N}}}(K) \rightrightarrows K_w,\; f\mapsto {\operatorname{Fix}}(f),$$ where we are given a nonexpansive mapping like in the case of the Browder-Göhde-Kirk theorem, but are only required to compute a fixed point with respect to the weak topology. Note that in $\ell^2$ this amounts to computing a fixed point with respect to an orthonormal basis, but not necessarily computing its $\ell^2$-norm (cf. also [@BrattkaInseparable]). Characterisation of the Fixed Point Sets of Computable Nonexpansive Mappings in Computable Hilbert Space {#Section: 5} ======================================================================================================== We may now prove our main result. Throughout this section we will work on a computable Hilbert space $H$. Note that in this case $H' \simeq H$ is again a computable Hilbert space, and that the mapping ${(\cdot,\cdot)}\colon H\times H' \to {\mathbb{R}}$ is the usual inner product on $H$, which is computable by the polarisation identity. In particular, we can use Definition \[Def: admissible representation for weak\* topology\] to construct the space $H_w$, whose representation is admissible for the weak topology on $H$. \[main result\] Let $H$ be a computable Hilbert space, let $K \subseteq H$ be weakly co-semi-decidable, computably overt, bounded, and convex. Then 1. The mapping $${\operatorname{Fix}}\colon {\operatorname{\mathcal{N}}}(K)\to {\mathcal{K}}^{{\operatorname{co}}}(K_w)\setminus\{\emptyset\}, f \mapsto {\operatorname{Fix}}(f)$$ is computable. 2. And so is its multivalued inverse $${\operatorname{Fix}}^{-1}\colon {\mathcal{K}}^{{\operatorname{co}}}(K_w)\setminus\{\emptyset\} \rightrightarrows {\operatorname{\mathcal{N}}}(K).\eqno{\qEd}$$ Let us sketch the proof of the second claim. Given a nonempty, weakly closed, bounded and convex subset $A$ of $K$, by Theorem \[Hahn-Banach equivalent to co-r.e.\] we can enumerate a sequence of half spaces whose intersection is equal to $A$. Now, the projections onto these half spaces are nonexpansive, thanks to Theorem \[thm: projection theorem\] (ii), and computable: \[lem: computability of projection onto rational half space\] Let $H$ be a computable Hilbert space. There exists a computable function which takes as input a rational half space $h \subseteq H$, encoded as a $\nu_{{\operatorname{HB}}}$-name, and returns as output the metric projection onto $h$ as an element of ${\operatorname{\mathcal{C}}}(H,H)$. Let $h = \{x \in H \;|\; {(x,x_h)} + a_h \leq 0\}$, where $x_h$ is a rational point in $H$ and $a_h \in {\mathbb{Q}}$. It follows from Lemma \[lem: effective independence lemma\] that we can without loss of generality assume that the set $\{n\in{\mathbb{N}}\;|\; \nu_H(n) = 0\}$ is *decidable* (cf. also [@BrattkaHahnBanach Lemma 3]). Thus, we can decide if $x_h = 0$, and if this is the case we necessarily have $a_h = 0$ (since $h$ is nonempty), and the projection onto $h$ is the identity on $H$. If $x_h \neq 0$, put $\tilde{x}_h = \tfrac{x_h}{{||x_h||}}$, $\tilde{a}_h = \tfrac{a_h}{{||x_h||}}$, and $p = x - \alpha \tilde{x}_h$, where $\alpha = \max\{0, {(x,\tilde{x}_h)} + \tilde{a}_h\}$. One easily verifies that $p \in h$ and that $p$ satisfies the variational inequality (Theorem \[thm: projection theorem\] (ii)). It follows that $P_h(x) = p$. This proves the claim. We can hence compute a sequence of nonexpansive mappings such that $A$ is the intersection of the fixed point sets of these mappings. The following theorem due to Bruck allows us to construct a single nonexpansive mapping whose fixed point set is the intersection of the fixed point sets of our sequence of mappings. \[thm: intersection of fixed points\] Let $E$ be a strictly convex real normed space, let $K \subseteq E$ be nonempty, closed, bounded, and convex. Let $(\lambda_n)_n$ be any sequence in $(0,1)$ satisfying $\sum_n \lambda_n = 1$. Let $(f_n)_n$ be a family of nonexpansive mappings on $K$ with $\bigcap_{n \in {\mathbb{N}}}{\operatorname{Fix}}(f_n) \neq \emptyset$. Then the mapping $$f = \sum_n \lambda_nf_n$$ is well-defined, nonexpansive and satisfies $${\operatorname{Fix}}(f) = \bigcap_{n \in {\mathbb{N}}} {\operatorname{Fix}}(f_n).\eqno{\qEd}$$ In the final step, we project back onto $K$ in order to construct a self-map of $K$. \[projecting back\] Let $H$ be a real Hilbert space, $K \subseteq H$ be closed and convex and $f\colon K \to H$ be nonexpansive and suppose that ${\operatorname{Fix}}(f) \neq \emptyset$. Let $P_K$ denote the metric projection onto $K$. Then $P_K\circ f$ is nonexpansive as well with ${\operatorname{Fix}}(P_K\circ f) = {\operatorname{Fix}}(f)$. It is clear that $P_K \circ f$ is nonexpansive and that ${\operatorname{Fix}}(f) \subseteq {\operatorname{Fix}}(P_K \circ f)$. Suppose there exists $x \in K$ with $f(x) \neq x$ and $P_K(f(x)) = x$. Let $y$ be some fixed point of $f$. Then $${||f(x) - f(y)||}^2 = {||y - x||}^2 + {||x - f(x)||}^2 - 2{(f(x) - x,y - x)}.$$ By assumption, $x = P_K(f(x))$, so by the variational inequality (Theorem \[thm: projection theorem\]) $${(f(x) - x,y - x)} \leq 0 \; \text{for all }y\in K.$$ We also assumed that $f(x) \neq x$, i.e. ${||f(x) - x||}^2 > 0$, hence $${||f(x) - f(y)||}^2 > {||x - y||}^2.$$ Contradicting the assumption that $f$ is nonexpansive. We prove that given a $\kappa_{\operatorname{HB}}$-name $\phi$ of a nonempty, closed, convex subset $A \subseteq K$ we can compute the name of a nonexpansive function $f\colon K \to K$ with ${\operatorname{Fix}}(f) = A$. The name $\phi$ encodes a sequence of rational half spaces $(h_k)_k$ containing $A$ in their interior. Using Lemma \[lem: computability of projection onto rational half space\], given $\phi$ we can compute a $[\delta_H\to\delta_H]^\omega$-name of some sequence $(P_k)_k$ of projections, where $P_k$ is the projection onto the rational half space $h_k$. By Theorem \[thm: intersection of fixed points\], the mapping $g = \sum_{k \in {\mathbb{N}}} 2^{-k - 1}P_k$ will satisfy ${{\operatorname{Fix}}(g) = \bigcap_{k \in {\mathbb{N}}} {\operatorname{Fix}}(P_k) = A}$. By Lemma \[projecting back\], the mapping $P_K \circ g\colon K \to K$ will have the same set of fixed points. Note that $P_K$ is computable by Corollary \[cor: computability of projection on E\], since $K$ is located by Proposition \[prop: computably weakly closed uniformly implies lower semi-located\]. In order to prove item (i) of Theorem \[main result\], we need to inspect Lemma \[projecting back\] a little closer. We first need another simple lemma. \[lem: convex intersection lemma\] Let $E$ be real normed space, let $C \subseteq E$ be closed and convex, let $S \subseteq C$ be a dense subset of $C$ and let $h$ be a half space in $E$. If $C\cap {{h}^\circ}$ is nonempty, then $S\cap {{h}^\circ}$ is dense in $C \cap h$. Let $x \in C\cap h$, and let $c \in C \cap {{h}^\circ}$. Then the line segment joining $c$ and $x$ is contained in $C \cap h$ and contains an element $b \in B(x,\varepsilon/2) \cap {{h}^\circ}$. Now we choose $a \in S$ in a sufficiently small ball around $b$, so that $d(x,a) < \varepsilon$. Lemma \[lem: convex intersection lemma\] guarantees that intersections of weakly closed and overt sets and rational half spaces are (uniformly) overt. This is a special property, as in general the intersection operator on closed sets is $\left((\psi\sqcap\upsilon)\times(\psi\sqcap\upsilon),\upsilon\right)$-discontinuous (cf. [@Weih Theorem 5.1.13]). \[computability of intersection of computable convex sets and rational half spaces\] Let $E$ be a computable Banach space with computable dual $E'$, such that the mapping ${(\cdot,\cdot)}$ is $(\delta_{E}\times\delta_{E'},\rho)$-computable. Then intersection of weakly closed convex sets $C$ and closed rational half spaces $h$ with $C\cap {{h}^\circ} \neq \emptyset$ is $\left((\psi^{E_w}\sqcap\upsilon^{E_w})\times\nu_{{\operatorname{HB}}},(\psi^{E_w}\sqcap\upsilon^{E_w})\right)$-computable, and hence ${\left((\psi^{E_w}\sqcap\upsilon^{E_w})\times\nu_{{\operatorname{HB}}},\psi_{{\operatorname{dist}}}\right)}$-computable. Since $\nu_{{\operatorname{HB}}} \leq \psi$, we can always uniformly compute a $\psi$-name of the intersection. In order to compute an $\upsilon$-name, enumerate all elements given by the $\upsilon$-name of $C$ which are also contained in ${{h}^\circ}$. The above lemma guarantees that this yields an $\upsilon$-name of $C \cap h$. The second claim follows from $(\psi^{E_w}\sqcap\upsilon^{E_w}) \equiv \psi_{{\operatorname{dist}}}$, which in turn follows from Proposition \[prop: computably weakly closed uniformly implies lower semi-located\]. \[quantitative projection lemma\] Let $H$ be a real Hilbert space, $K \subseteq H$ be closed, bounded and convex, let $f\colon K \to K$ be nonexpansive, $h$ be a half space such that ${{h}^\circ}\cap K \neq \emptyset$ and let $S\subseteq K$ be dense in $K$. Let $A = h\cap K$. Then ${{\operatorname{Fix}}(f) \cap h = \emptyset}$ if and only if $$\label{eq: quantitative projection lemma equation} \exists x \in {{h}^\circ} \cap S.\exists n \in {\mathbb{N}}.\left({||f(x) - x||} > 2^{-n} \land {||P_A(f(x)) - x||} < \frac{2^{-2n-3}}{B}\right),$$ where $B \geq \sup\{{||x||} + 1 \;\big|\; x \in K\}$. Let us first prove the forward direction. By the Browder-Göhde-Kirk Theorem, ${P_A\circ f|_A\colon A\to A}$ has a fixed point $\tilde{x} \in h \cap K$. Since, by assumption, $\tilde{x}$ is not a fixed point of $f$, there exists an $m \in {\mathbb{N}}$ with ${||f(\tilde{x}) - \tilde{x}||} > 2^{-m}$. By Lemma \[lem: convex intersection lemma\], $S \cap {{h}^\circ}$ is dense in $h\cap K$, so that we may choose $x \in S\cap{{h}^\circ}$ with ${||x - \tilde{x}||} < \frac{2^{-2m-6}}{B}$. Then $$\begin{aligned} {||P_A(f(x)) - x||} &\leq {||P_A(f(\tilde{x})) - \tilde{x}||} + {||P_A(f(x)) - P_A(f(\tilde{x}))||} + {||x - \tilde{x}||}\\ &\leq 2 {||x - \tilde{x}||} < \frac{2^{-2m-5}}{B} = \frac{2^{-2(m + 1) - 3}}{B}.\end{aligned}$$ And (using $B \geq 1$) $${||f(x) - x||} \geq {||f(\tilde{x}) - \tilde{x}||} - 2{||x - \tilde{x}||} > 2^{-m} - \frac{2^{-2m-5}}{B} \geq 2^{-m} - 2^{-2m-5} > 2^{-m-1}.$$ For the converse direction, we proceed by contrapositive. We suppose that there exists $y \in {\operatorname{Fix}}(f) \cap h$ (and hence $y \in {\operatorname{Fix}}(P_A\circ f|_A)$) and show $$\forall x \in {{h}^\circ} \cap S.\forall n \in {\mathbb{N}}. \left( {||f(x) - x||} > 2^{-n} \rightarrow {||P_A(f(x)) - x||} \geq \frac{2^{-2n-3}}{B} \right).$$ Let $x \in {{h}^\circ} \cap S$ with ${||f(x) - x||} > 2^{-n}$. Since $f$ is nonexpansive, we have $$\begin{aligned} {||y - x||}^2 &\geq {||f(y) - f(x)||}^2 \\&= {||f(y) - P_A(f(x))||}^2 + {||P_A(f(x)) - f(x)||}^2 + 2(y - P_A(f(x)),P_A(f(x)) - f(x)).\end{aligned}$$ Now, by the variational inequality, $2(y - P_A(f(x)),P_A(f(x)) - f(x)) \geq 0$, so that $$\begin{aligned} {||y - x||}^2 &\geq {||f(y) - P_A(f(x))||}^2 + {||P_A(f(x)) - f(x)||}^2\\ &= {||y - x||}^2 + {||x - P_A(f(x))||}^2 + 2(y - x,x - P_A(f(x)))\\ &+ {||P_A(f(x)) - x||}^2 + {||x - f(x)||}^2 + 2(P_A(f(x)) - x, x - f(x)),\end{aligned}$$ which entails that $$\begin{aligned} 0 &\geq {||f(x) - x||}^2 + 2(P_A(f(x)) - x, x - f(x) - y + x)\\ &\geq {||f(x) - x||}^2 - 2{||P_A(f(x)) - x||}\cdot{||x - f(x) - y + x||}\\ &\geq {||f(x) - x||}^2 - 8{||P_A(f(x)) - x||}B,\end{aligned}$$ and hence $${||P_A(f(x)) - x||} \geq \frac{2^{-2n-3}}{B}.\eqno{\qEd}$$ Note that it follows from Corollary \[computability of intersection of computable convex sets and rational half spaces\] that the projection onto $A$ in Lemma \[quantitative projection lemma\] is computable. Given a nonexpansive mapping $f\colon K \to K$, we want to compute a $\kappa_{{\operatorname{HB}}}$-name of ${\operatorname{Fix}}(f)$. We need to compute a weakly compact set $L \in {\mathcal{K}}(E_w)$ with $L \supseteq {\operatorname{Fix}}(f)$, and a list of all rational half spaces containing $f$. Since $K$ contains ${\operatorname{Fix}}(f)$ and is computably weakly compact by Theorem \[thm: unit ball in weak\*-topology as a computable metric space\] (ii) we may put $L = K$, so it suffices to list all rational half spaces containing ${\operatorname{Fix}}(f)$. In fact, by Lemma \[lem: listing enough half-spaces\] it suffices to compute a list of rational half spaces $(h_n)_n$ satisfying $\bigcap_{n \in {\mathbb{N}}} {{h}^\circ}_n = {\operatorname{Fix}}(f)$. In order to do so, we enumerate two different lists $L_1$ and $L_2$ of half spaces and interleave them. The first list $L_1$ consists of all rational half spaces containing $K$ in their interior. This list is computable since $K$ is computably weakly compact, and hence $\kappa_{{\operatorname{HB}}}$-computable by Theorem \[Hahn-Banach equivalent to co-r.e.\]. In order to compute the second list $L_2$, we first enumerate all rational half spaces $h$ such that ${{h}^\circ} \cap K \neq \emptyset$ and $h^C \cap K \neq \emptyset$. This is possible because $K$ is computably overt. Out of these half spaces we only enumerate those which satisfy ${\operatorname{Fix}}(f) \cap ({{h}^\circ})^C = \emptyset$. In order to verify this property we apply Lemma \[quantitative projection lemma\] to the half space $({{h}^\circ})^C$. Note that we can compute the projection onto $K\cap ({{h}^\circ})^C$ by Corollary \[computability of intersection of computable convex sets and rational half spaces\], so that the property in Lemma \[quantitative projection lemma\] becomes semi-decidable. Now, ${\operatorname{Fix}}(f) \cap ({{h}^\circ})^C = \emptyset$ is equivalent to ${{h}^\circ} \supseteq {\operatorname{Fix}}(f)$, and it is easy to see that the list $(h_n)_n$ we obtain by interleaving $L_1$ and $L_2$ satisfies $\bigcap_{n \in {\mathbb{N}}} {{h}^\circ}_n = {\operatorname{Fix}}(f)$. Theorem \[main result\] now allows us to determine the Weihrauch degree of the weak and strong Browder-Göhde-Kirk theorem. \[thm: BGK equivalent convex choice\] Let $H$ be a computable Hilbert space and $K\subseteq H$ be nonempty, bounded, convex, computably weakly closed, and computably overt. Then $${\operatorname{BGK}}_K \equiv_W {\operatorname{ConvC}}_K^{{\operatorname{w}}\to{\operatorname{n}}},$$ and $${\operatorname{WBGK}}\equiv_W {\operatorname{ConvC}}_K^{{\operatorname{w}}\to{\operatorname{w}}}.$$ If $K$ is computably compact, then $${\operatorname{BGK}}_K \equiv_W {\operatorname{WBGK}}_K \equiv_W {\operatorname{ConvC}}_K.$$ The equivalences ${\operatorname{BGK}}_K \equiv_W {\operatorname{ConvC}}_K^{{\operatorname{w}}\to{\operatorname{n}}}$ and ${\operatorname{WBGK}}\equiv_W {\operatorname{ConvC}}_K^{{\operatorname{w}}\to{\operatorname{w}}}$ follow from Theorem \[main result\], together with the fact that by Theorem \[thm: unit ball in weak\*-topology as a computable metric space\] $K$ is computably weak\* compact, so that ${\operatorname{id}}:{\mathcal{A}}^{{\operatorname{co}}}(K) \to {\mathcal{K}}^{{\operatorname{co}}}(K)$ is computable. If $K$ is computably compact, then by Proposition \[prop: weak and strong topology equivalent on compact sets\] we have $\delta_{E}\big|^K \equiv \delta_{E_w}\big|^K$, which yields ${\operatorname{BGK}}_K \equiv_W {\operatorname{WBGK}}_K \equiv_W {\operatorname{ConvC}}_K$ (note that if two representations of the same space are equivalent, then the induced canonical representations of closed sets are - by construction - equivalent as well). Theorem \[thm: BGK equivalent convex choice\] also shows that on a non-compact domain, negative information on the weak closedness of a set is much stronger than negative information on its norm-closedness. We have for instance ${\operatorname{BGK}}_{B_{\ell^2}} \leq_W {\operatorname{Proj}}_{B_{\ell^2}} \equiv_W \lim$, and so ${\operatorname{ConvC}}_{B_{\ell^2}}^{{\operatorname{w}}\to{\operatorname{n}}} \leq_W \lim$, while already ${\operatorname{UC}}_{B_{\ell^2}}^{{\operatorname{n}}\to{\operatorname{n}}}$ is equivalent to the extremely non-effective principle ${\operatorname{C}}_{{\mathbb{N}}^{\mathbb{N}}}$. In finite dimension, the degree of ${\operatorname{BGK}}_K$ is always strictly below ${\operatorname{WKL}}$ because of Corollary \[cor: fixed points in finite dimension are computable\]. On the unit ball in $\ell^2$ this is no longer the case. \[thm: computable mapping with norm-uncomputable image\] There exists a computable mapping $$T\colon \subseteq [0,1]^{{\mathbb{N}}} \to {\mathcal{K}}^{{\operatorname{co}}}(B_{\ell^2_w})$$ with ${\operatorname{dom}}T = \{x \in [0,1]^{\mathbb{N}}\;|\;x(n)\leq x(n + 1)\}$ such that for all $x \in {\operatorname{dom}}T$ we have $T(x) = \{a\}$ with ${{||a||}_2 = \lim x(n)}$. Let $x\in {\operatorname{dom}}T$. Put $a(0) = x(0)$ and $a(n + 1) = \sqrt{x(n + 1)^2 - x(n)^2}$. Then we have $a(n)^2 + \dots + a(0)^2 = x(n)^2$. Now, put $T(x) = \{a\}$. Note that $a$ is $\delta_{\ell^2_w}$-computable relative to $x$, so we can compute $\{a\}$ in ${\mathcal{K}}^{{\operatorname{co}}}(B_{\ell^2_w})$: in order to compute the characteristic function of $\{a\}^C$ into Sierpiński space we simply check for inequality with $a$ component-wise. This allows us to compute $\{a\}$ as a point in ${\mathcal{A}}^{{\operatorname{co}}}(B_{\ell^2_w})$, and thus as a point of ${\mathcal{K}}^{{\operatorname{co}}}(B_{\ell^2_w})$, using that the identity ${\operatorname{id}}\colon {\mathcal{A}}^{{\operatorname{co}}}(B_{\ell^2_w}) \to {\mathcal{K}}^{{\operatorname{co}}}(B_{\ell^2_w})$ is computable, since $B_{\ell^2_w}$ is computably weakly compact by Theorem \[thm: unit ball in weak\*-topology as a computable metric space\]. Choosing from a set in ${\mathcal{K}}^{{\operatorname{co}}}(B_{\ell^2})$ hence allows us computably translate a $\rho_<$-name to a $\rho$-name of a given real number $x \in [0,1]$, already if the set is a singleton. This yields: \[cor: Weihrauch degree of BGK on unit ball\] ${\operatorname{BGK}}_{B_{\ell^2}} \equiv_W {\operatorname{UC}}_{B_{\ell^2}}^{{\operatorname{w}}\to{\operatorname{n}}} \equiv_W \lim$. In particular ${\operatorname{BGK}}_{B_{\ell^2}} \equiv_W {\operatorname{Proj}}_{B_{\ell^2}}$. We have ${\operatorname{UC}}_{B_{\ell^2}}^{{\operatorname{w}}\to{\operatorname{n}}} \leq_W {\operatorname{ConvC}}_{B_{\ell^2}}^{{\operatorname{w}}\to{\operatorname{n}}} \equiv_W {\operatorname{BGK}}_{B_{\ell^2}}$, the latter by Theorem \[thm: BGK equivalent convex choice\]. Also, ${\operatorname{BGK}}_{B_{\ell^2}} \leq_W \lim$ by Proposition \[prop: obvious Weihrauch reductions\] and Proposition \[prop: initial upper bound on Proj\]. It follows from Theorem \[thm: computable mapping with norm-uncomputable image\] that ${\operatorname{UC}}_{B_{\ell^2}}^{{\operatorname{w}}\to{\operatorname{n}}}$ allows us to determine the limit of any computable monotonically increasing sequence $x \in [0,1]^{{\mathbb{N}}}$, since $$\lim_{n\to\infty} x(n) = {||{\operatorname{UC}}_{B_{\ell^2}}^{{\operatorname{w}}\to{\operatorname{n}}}(T(x))||}_2$$ and ${||\cdot||}_2$ is $(\delta_{\ell^2},\rho)$-computable. We hence have $\lim \leq_W {\operatorname{UC}}_{B_{\ell^2}}^{{\operatorname{w}}\to{\operatorname{n}}}$, which finishes the proof. In particular we have the following non-uniform corollary: \[cor: nonuniform complexity of BGK on unit ball\] There exists a computable nonexpansive self-map of the closed unit ball in $\ell^2$ with a unique fixed point, which is uncomputable. Compare Corollary \[cor: nonuniform complexity of BGK on unit ball\] to Theorem \[thm: nonexpansive mapping on Hilbert cube without computable fixed points\]: on a compact domain, any computable function without computable fixed points necessarily has uncountably many fixed points, since otherwise it has at least one isolated fixed point which is then computable by Theorem \[thm: Kreinovich’s theorem\]. If we drop compactness, even unique solutions may be uncomputable. Note however, that since the unit ball in $\ell^2$ is still computably weakly compact, unique fixed points on $B_{\ell^2}$ are still “weakly computable”, in the sense that they are computable as elements in the represented space $\ell^2_w$. In particular, their coordinates with respect to an orthonormal basis are still computable. On a computably compact domain, the Weihrauch degree of the theorem is still at most ${\operatorname{WKL}}$. We can now show that it is in fact equivalent to ${\operatorname{WKL}}$ on the Hilbert cube. In order to do so, we will first have to define the *parallelisation* of Weihrauch degrees, which was introduced in [@WeihrauchDegrees]. Let $f\colon\subseteq X \rightrightarrows Y$ be a partial multimapping. The *parallelisation* $\hat{f}$ of $f$ is the multimapping $\hat{f}\colon\subseteq X^{\mathbb{N}}\rightrightarrows Y^{\mathbb{N}}$, $\hat{f}(\lambda n.x(n)) = \lambda n.f(x(n))$. It is not hard to see that $f \leq_W g$ implies $\hat{f} \leq_W \hat{g}$ (cf. also Proposition 4.2 in [@WeihrauchDegrees]). The following theorem is essentially due to [@WeihrauchDegrees] (cf. also Theorem 6.2 and the subsequent comment in [@EffectiveChoice]). $\widehat{{\operatorname{IVT}}} \equiv_W {\operatorname{WKL}}$. \[thm: BGK on Hilbert cube is WKL\] Let $\mathcal{H} = \{\sum_{i \in {\mathbb{N}}} \alpha_i e_i\;|\; \alpha_i \in [0,2^{-i}]\}$ be the Hilbert cube in $\ell^2$. Then ${\operatorname{BGK}}_{\mathcal{H}} \equiv_W {\operatorname{WKL}}$. Clearly, $\mathcal{H}$ is computably compact, so ${\operatorname{BGK}}_{\mathcal{H}} \equiv_W {\operatorname{ConvC}}_{\mathcal{H}} \leq_W {\operatorname{C}}_{\mathcal{H}} \equiv_W {\operatorname{WKL}}$. In order to prove the converse direction, we show that $\widehat{{\operatorname{IVT}}} \leq_W {\operatorname{ConvC}}_{\mathcal{H}}$. Since $\widehat{{\operatorname{IVT}}} \equiv_W {\operatorname{WKL}}$ and ${\operatorname{ConvC}}_{\mathcal{H}} \equiv_W {\operatorname{BGK}}_{\mathcal{H}}$, it follows that ${\operatorname{WKL}}\leq_W {\operatorname{BGK}}_{\mathcal{H}}$. By Proposition \[prop: BGK in \[0,1\]\] we have ${\operatorname{IVT}}\equiv_W {\operatorname{ConvC}}_{[0,1]}$ and so $\widehat{{\operatorname{IVT}}} \equiv_W \widehat{{\operatorname{ConvC}}_{[0,1]}}$. Let $([a_n,b_n])_n$ be a sequence of closed intervals in $\left({\mathcal{A}}^{{\operatorname{co}}}([0,1])\right)^{{\mathbb{N}}}$. Consider the set $A = \{\sum_{i \in {\mathbb{N}}} \alpha_i e_i\;|\; \alpha_i \in [a_i2^{-i},b_i2^{-i}]\} \subseteq {\mathcal{H}}$. Then $A$ is computable as a point in ${\mathcal{K}}^{{\operatorname{co}}}({\mathcal{H}})$ relative to $([a_n,b_n])_n$. Clearly, choosing a point in $A$ allows us to choose a point in $([a_n,b_n])_n$, so $\widehat{{\operatorname{ConvC}}_{[0,1]}} \leq_W {\operatorname{ConvC}}_{\mathcal{H}} \equiv_W {\operatorname{BGK}}_\mathcal{H}$. Theorem \[thm: BGK on Hilbert cube is WKL\] can (essentially) be viewed as a uniform strengthening of Theorem \[thm: nonexpansive mapping on Hilbert cube without computable fixed points\]. Notice that the proof of Theorem \[thm: nonexpansive mapping on Hilbert cube without computable fixed points\] can be utilized to establish the reduction ${\widehat{{\operatorname{LLPO}}} \leq_W {\operatorname{BGK}}_{\mathcal{H}}}$, which yields a slightly different proof of Theorem \[thm: BGK on Hilbert cube is WKL\], since $\widehat{{\operatorname{LLPO}}} \equiv_W {\operatorname{WKL}}$ (again, cf. [@WeihrauchDegrees]). The proof of Theorem \[thm: BGK on Hilbert cube is WKL\] can also be used to show ${\operatorname{WBGK}}_{B_{\ell^2}} \equiv_W {\operatorname{WKL}}$. We now have a fairly good idea of the computational content of the Browder-Göhde-Kirk theorem. It follows from [@PaulyLeRoux] that $({\operatorname{BGK}}_{[0,1]^n})_{n \in {\mathbb{N}}}$ is a strictly increasing sequence of Weihrauch degrees, all strictly below ${\operatorname{WKL}}$. On the compact but infinite dimensional Hilbert cube $\mathcal{H}$ the theorem becomes equivalent to ${\operatorname{WKL}}$. If we drop compactness and consider the theorem on the unit ball in $\ell^2$, it becomes even more non-effective, and in particular equivalent to computing rates of convergence for fixed point iterations, but is still much more effective than full choice on $B_{\ell^2}$. In finite dimension, a computable nonexpansive self-map of a computably compact domain always has computable fixed points by Theorem \[thm: convex sets have computable points\], and this relies solely on the fact that the fixed point set is convex. This is reminiscent of the fact that unique zeroes of computable functions are always (in this case even uniformly) computable. A typical feature of such results is that they assert the existence of computable objects, but the computational complexity of these objects is unbounded. This is also the case here: using similar techniques as in [@Ko], we can strengthen Theorem \[main result\] $(ii)$ to assert for every nonempty co-semi-decidable and convex $A \subseteq K$ the existence of a polynomial-time computable nonexpansive $f\colon K\to K$ such that ${\operatorname{Fix}}(f) = A$, at least in the case where $K$ is computably compact, and so in particular in the finite-dimensional case (if $K$ is not computably compact there is no uniform majorant on the names of the points in $K$, so one would have to work in the framework of *second-order complexity* [@KawamuraCook]). This allows us to characterise the computational complexity of fixed points of Lipschitz-continuous polynomial-time computable functions according to their Lipschitz constant. \[thm: complexity of mappings w.r.t. Lipschitz constant\] Let $[0,1]^2$ be the unit square in Euclidean space ${\mathbb{R}}^2$. Let ${f\colon [0,1]^2 \to [0,1]^2}$ be polynomial-time computable and Lipschitz-continuous with Lipschitz constant $L$. Then: - If $L < 1$, $f$ has a unique polynomial time computable fixed point, which is uniformly computable relative to the promise that $L < 1$ and uniformly polynomial time computable relative to the promise that $L < 1 - \varepsilon$ for some fixed $\varepsilon > 0$. - If $L = 1$, the fixed point set of $f$ can be any nonempty co-semi-decidable convex subset of $[0,1]^2$. The multi-valued operator mapping $f$ to some fixed point is realiser-discontinuous and hence uncomputable, but $f$ still has computable fixed points. However, there is no computable bound on the computational complexity of the fixed points of $f$. - If $L > 1$, $f$ may not have any computable fixed points. The third claim in Theorem \[thm: complexity of mappings w.r.t. Lipschitz constant\] follows from a strengthening of the results in [@ConnectedChoice], which the authors of that paper have recently obtained, but which seems to be unpublished as of yet. Further Results and Possible Generalisations ============================================ The special case of Theorem \[main result\] where the underlying Hilbert space is two-dimensional seems to generalise to uniformly convex and smooth real Banach spaces of dimension two. Note that the first item of the theorem becomes trivial in finite dimension. A Banach space is called *smooth*, if its dual space is strictly convex and *uniformly smooth* if its dual space is uniformly convex. For instance, all $L^p$-spaces with $1 < p < \infty$ are uniformly convex and uniformly smooth. The two notions of smoothness and uniform smoothness coincide in finite dimension. \[characterisation of fixed point sets in two dimensional smooth and convex Banach space\] Let $E$ be a uniformly convex, smooth, computable Banach space of dimension two, and let $K\subseteq E$ be bounded, convex, and located[^6]. Then the multi-valued mapping $${\operatorname{Fix}}^{-1}\colon {\mathcal{K}}^{{\operatorname{co}}}(K)\setminus\{\emptyset\}\rightrightarrows {\operatorname{\mathcal{N}}}(K)$$ is computable. The proof of this result would be almost identical to that of Theorem \[main result\]. The only places where we used that the underlying space is a Hilbert space were Theorem \[thm: projection theorem\], which asserts that the projection onto each convex, closed set is nonexpansive, and Lemma \[projecting back\]. In general the projection onto a closed and convex subset of a Banach space will not be nonexpansive. In fact, this property characterises Hilbert spaces (cf. [@Phelps]). However, we only need the existence of a computable nonexpansive retraction onto each located convex subset. A retraction $Q\colon E \to K$ of $E$ onto a nonempty subset $K\subseteq E$ is called *sunny*, if $$Q(\alpha x + (1 - \alpha)Q(x)) = Q(x) \;\text{ for all } x \in E, \alpha \in [0,1].$$ Geometrically, this means that for all $x \notin K$, all points on the ray defined by $x$ and $Q(x)$ with initial point $Q(x)$ are mapped onto the same point $Q(x)$. It is well known that in a smooth Banach space of dimension two, sunny nonexpansive retractions onto closed convex subsets exist and are unique. Consequently, they are computable. \[existence of sunny nonexpansive retracts\] Let $E$ be a smooth real Banach space of dimension two. Then for every nonempty closed convex subset $C$ of $E$, there exists a nonexpansive sunny retraction of $E$ onto $C$. \[uniqueness of sunny nonexpansive retracts\] Let $E$ be a smooth real Banach space. Let $K \subseteq C$ be two nonempty, closed, and convex subsets of $E$. Then there exists at most one sunny nonexpansive retraction of $C$ onto $K$. \[computability of sunny nonexpansive retracts\] Let $E$ be a smooth computable Banach space of dimension two. Let $C \subseteq E$ be nonempty, convex, bounded, and located. Then the mapping $$\operatorname{SRet}\colon{\mathcal{A}}^{{\operatorname{co}}}_{{\operatorname{dist}}}(C)\setminus\{\emptyset\}\mapsto {\operatorname{\mathcal{N}}}(C)$$ that maps $K$ to the unique sunny nonexpansive retraction of $\,C$ onto $K$, is computable. The set of nonexpansive self-maps of $C$ is computably compact, since it is equicontinuous and $C$ is compact. We can verify if a given map $f\colon C\to C$ does not leave all points of $K$ fixed, if it maps a point of $C$ to a point outside of $K$, and if ${f(\alpha x + (1 - \alpha)f(x)) \neq f(x)}$ for some $x \in C$, $\alpha \in [0,1]$. It follows that the set of sunny nonexpansive retractions of $C$ onto $K$ is co-semi-decidable relative to (a $\psi_{{\operatorname{dist}}}$-name of) $K$. Theorems \[existence of sunny nonexpansive retracts\] and \[uniqueness of sunny nonexpansive retracts\] assert that it is a singleton. It follows that the operator is uniformly computable. The other result that relies on Hilbert space techniques is Lemma \[projecting back\], which uses the nonexpansiveness of the projection and the variational inequality. In principle we could replace the projection by the sunny nonexpansive retraction onto the domain, but the question remains whether this will always leave the fixed point set unchanged. \[projection lemma in Banach space\] Let $E$ be a smooth and uniformly convex Banach space of dimension two, let $K \subseteq E$ be nonempty, closed, bounded, and convex and let $f\colon K \to E$ be nonexpansive with ${{\operatorname{Fix}}(f)\neq\emptyset}$. Let $P\colon E \to K$ be the sunny nonexpansive retraction onto $K$. Then we have ${\operatorname{Fix}}(P\circ f) = {\operatorname{Fix}}(f)$. We could now prove Conjecture \[characterisation of fixed point sets in two dimensional smooth and convex Banach space\] analogously to Theorem \[main result\]: we are given a convex, closed subset $A$ of $K$ as a $\kappa_{{\operatorname{HB}}}$-name and want to construct a nonexpansive mapping $f\colon K \to K$ with ${\operatorname{Fix}}(f) = A$. Let $(h_n)_n$ be the sequence of half spaces given by the $\kappa_{{\operatorname{HB}}}$-name. Since $h_n\cap K \neq \emptyset$ for all $n$, we can compute a $\psi_{{\operatorname{dist}}}^\omega$-name of the sequence $(h_n\cap K)_{n \in {\mathbb{N}}}$ thanks to Corollary \[computability of intersection of computable convex sets and rational half spaces\]. Now, Theorem \[computability of sunny nonexpansive retracts\] allows us to compute a $[\delta_K\to\delta_K]^\omega$-name of the sequence $(f_n)_n$ of sunny nonexpansive retractions of $K$ onto $h_n\cap K$. Applying Theorem \[thm: intersection of fixed points\], we obtain a nonexpansive mapping $g\colon K \to E$ with ${\operatorname{Fix}}(g) = \bigcap_{n\in{\mathbb{N}}} h_n = A$. Finally, we use the computable nonexpansive sunny retraction onto $K$ and Conjecture \[projection lemma in Banach space\] to obtain a self-map $f$ of $K$ with ${\operatorname{Fix}}(f) = A$. The only “missing piece” in this proof is Conjecture \[projection lemma in Banach space\]. By replacing this conjecture by a weaker statement that we can prove, we obtain a weaker version of Conjecture \[characterisation of fixed point sets in two dimensional smooth and convex Banach space\], which is almost as good. \[weak projection lemma\] Let $E$ be a uniformly convex Banach space of dimension two, let $K \subseteq E$ be nonempty, closed, bounded, and convex. Suppose that ${{K}^\circ}$ is nonempty and that $\partial K$ does not contain any line segments, and let $f\colon K \to E$ be nonexpansive with ${\operatorname{Fix}}(f)\neq\emptyset$. Let $P\colon E \to K$ be the sunny nonexpansive retraction onto $K$. Then $P\circ f$ is nonexpansive as well with ${\operatorname{Fix}}(P\circ f) = {\operatorname{Fix}}(f)$. Clearly, $P\circ f$ is nonexpansive with ${\operatorname{Fix}}(f) \subseteq {\operatorname{Fix}}(P \circ f)$. Suppose that there exists $x \in {\operatorname{Fix}}(P\circ f)$, which is not a fixed point of $f$. Since $P$ is sunny, $x \in \partial K$. Let $y \in {\operatorname{Fix}}(f)$. Since ${\operatorname{Fix}}(f)$ is closed, there exists $\varepsilon > 0$ such that $B(x,\varepsilon) \subseteq {\operatorname{Fix}}(f)^C$. Since $P\circ f$ is nonexpansive, the line segment $L$ joining $y$ and $x$ is contained in ${\operatorname{Fix}}(P \circ f)$. By hypothesis, the line segment without its endpoints has to lie in ${{K}^\circ}$ (it is easy to see that if a convex set contains three points of a line segment in its boundary, it contains the whole line segment in its boundary). Hence, there exists $z \in {{K}^\circ}\cap L \cap B(x,\varepsilon)$. Contradiction. A similar proof shows that we may replace the condition that $\partial K$ contains no line segments, by the condition that ${\operatorname{Fix}}(f)\cap{{K}^\circ} \neq \emptyset$. In this case we do not even require the retraction to be sunny. \[weak main theorem on Banach spaces\] Let $E$ be a uniformly convex, smooth computable Banach space of dimension two, and let $K\subseteq E$ be nonempty, bounded, convex, and located. Suppose that either $\dim K = 2$ and $\partial K$ contains no line segments or $\dim K = 1$. then the multi-valued mapping $${\operatorname{Fix}}^{-1}\colon {\mathcal{K}}^{{\operatorname{co}}}(K)\setminus\{\emptyset\}\rightrightarrows {\operatorname{\mathcal{N}}}(K)$$ is computable. In a uniformly convex space the unit ball contains no line segments, so $B_E$ is an example of an admissible domain $K$. In particular, every co-semi-decidable, convex subset of $B_E$ is the fixed point set of some computable, nonexpansive self-map of $B_E$. If $\dim K = 1$, we introduce suitable coordinates in which $K$ is contained in the $x$-axis and use the construction of Theorem \[thm: main theorem on \[0,1\]\]. If $\dim K = 2$, we use the proof of Conjecture \[characterisation of fixed point sets in two dimensional smooth and convex Banach space\]. Note that here we may replace Conjecture \[projection lemma in Banach space\] by Lemma \[weak projection lemma\], so the proof is complete. The obvious question at this point is whether Conjecture \[characterisation of fixed point sets in two dimensional smooth and convex Banach space\] might generalise to higher dimensional Banach spaces. While most of the results we used in the proof at least generalise to finite-dimensional smooth and uniformly convex computable Banach spaces the main obstruction appears to be the existence of nonexpansive retractions. Our proof uses the fact that there exist nonexpansive retractions onto every rational half space, but if $E$ is a Banach space of dimension at least three and there exist nonexpansive retractions onto each two-dimensional subspace, then $E$ is a Hilbert space. Similarly, the unit ball of an at least three-dimensional Banach space $E$ is a nonexpansive retract of $E$ if and only if $E$ is a Hilbert space (cf. [@BruckHilbert]). On the other hand, every fixed point set of a nonexpansive mapping $f\colon K\to K$ is a nonexpansive retraction of $K$. In view of these results it seems likely that Theorem \[main result\] characterises computable Hilbert space of dimension three or higher. Finally, we extend the stronger upper bound obtained in Proposition \[prop: initial upper bound on Proj\] for compact sets and Hilbert space to the noncompact case in uniformly convex and uniformly smooth spaces. For this we need a generalisation of Theorem \[thm: Halpern’s theorem\] due to Reich [@ReichStrongConvergence]. We will only state a special case. \[thm: Reich’s theorem\] Let $E$ be a uniformly smooth, uniformly convex Banach space, let $K \subseteq E$ be nonempty, closed, bounded and convex, let $f\colon K \to K$ be nonexpansive, and let $x \in K$. Put $\alpha_n = 1 - (n + 2)^{-\tfrac{1}{2}}$. Then the sequence $(x_n)_n$ defined by the iteration scheme $x_0 = x$ and $$x_{n + 1} = (1 - \alpha_n)x_0 + \alpha_n f(x_n)$$ converges to a fixed point of $f$. Note that the iteration defined in Theorem \[thm: Reich’s theorem\] converges to a retraction onto the fixed point set of $f$. In fact, one can show that the sequence $(x_n)_n$ converges to $Q(x_0)$, where $Q$ is the unique sunny nonexpansive retraction of $K$ onto ${\operatorname{Fix}}(f)$. \[thm: upper bound on Proj in smooth spaces\] Let $E$ be a uniformly convex, uniformly smooth computable Banach space. Let $K \subseteq E$ be nonempty, bounded, convex, co-semi-decidable, and computably overt. Then $${\operatorname{Proj}}_K \leq_W \lim.$$ We use similar ideas as in the proof of Proposition \[prop: initial upper bound on Proj\]. Again we exploit the fact that we can actually compute countably many instances of $\lim$ in parallel. As in the proof of the general upper bound in Proposition \[prop: initial upper bound on Proj\], we use countably many instances of $\lim$ to obtain a function $\mu\colon {\mathbb{N}}\to {\mathbb{N}}$ satisfying $2^{-\mu(n)} \leq \eta_E(2^{-n})$, where $\eta_E$ is a modulus of uniform convexity for $E$, and another batch of countably many instances to obtain an approximation to the distance function to ${\operatorname{Fix}}(f)$ from below. Since $K$ is computably overt, it contains a computable dense sequence $(x_n)_n$. Let $x_n^0 = x_n$ and $x_n^{k + 1} = (1 - \alpha_k)x_n^0 + \alpha_k f(x_n^k)$ with $\alpha_k$ as in Theorem \[thm: Reich’s theorem\]. Using another countable batch of instances of $\lim$, we obtain the sequence $(\lim_{k\to\infty} x^k_n)_n$, which is dense in ${\operatorname{Fix}}(f)$, since the iteration defines a retraction of $K$ onto ${\operatorname{Fix}}(f)$. Using this sequence we can compute the distance function to ${\operatorname{Fix}}(f)$ from above, so that we obtain the distance function to ${\operatorname{Fix}}(f)$ as an element of ${\operatorname{\mathcal{C}}}(K,{\mathbb{R}})$. Together with Corollary \[cor: computability of projection on subset\] this establishes the reduction. Acknowledgements. {#acknowledgements. .unnumbered} ================= The present work was motivated by a question by Ulrich Kohlenbach, whether the Krasnoselski-Mann iteration has nonuniformly computable rates of convergence. He has also provided many valuable insights both concerning fixed point theory and computability theory. This work has greatly benefited from discussions with Vasco Brattka, Arno Pauly, Guido Gherardi, and Martin Ziegler. I would also like to thank the anonymous referees for pointing out many shortcomings in the original version of this paper. [^1]: The author was partly supported by the German Research Foundation (DFG) with Project Zi 1009/4-1 and by the Royal Society International Exchange Grant IE111233 [^2]: Like Schröder, and as opposed to Pauly, we will mostly work with a specific topology for $X$ in mind. To emphasize this, we call the spaces of interest represented *topological* spaces, rather than “represented spaces”. In general, the topology on $X$ will *not* be the final topology of its representation, and topological continuity may differ from realiser-continuity if the topology of $X$ is not sequential. [^3]: We call a multimapping *limit-computable* if its Weihrauch degree is below $\lim$. [^4]: This application was pointed out to the author by Ulrich Kohlenbach. [^5]: Thus, Theorem \[thm: admissibility of delta E’\_w\] together with Proposition 3.3.2 (3) in [@SchroederPhD] provide an interesting proof of the separable *Eberlein-[Š]{}mulian theorem* as well as its analogue for the weak\* topology in duals of separable spaces. [^6]: Recall that in finite dimension a nonempty closed set is located if and only if it is co-semi-decidable and computably overt (cf. also [@Weih]).

ArXiv

--- abstract: | In this paper we use the relationship between conformal metrics on the sphere and horospherically convex hypersurfaces in the hyperbolic space for giving sufficient conditions on a conformal metric to be radial under some constrain on the eigenvalues of its Schouten tensor. Also, we study conformal metrics on the sphere which are invariant by a $k-$parameter subgroup of conformal diffeomorphisms of the sphere, giving a bound on its maximum dimension. Moreover, we classify conformal metrics on the sphere whose eigenvalues of the Shouten tensor are all constant (we call them *isoparametric conformal metrics*), and we use a classification result for radial conformal metrics which are solution of some $\sigma _k -$Yamabe type problem for obtaining existence of rotational spheres and Delaunay-type hypersurfaces for some classes of Weingarten hypersurfaces in $\h ^{n+1}$. --- \[section\] \[section\] \[section\] \[section\] \[section\] \[section\] \[section\] ------------------------------------------------------------------------ [**Invariant conformal metrics on $\s ^n$**]{} [José M. Espinar$\,^\dag$[^1]]{}\ ------------------------------------------------------------------------ $\mbox{}^\dag$ Departamento de Geometría y Topología, Universidad de Granada, 18071 Granada, Spain; e-mail: [email protected] Keywords: Conformal metric, $\sigma _k$ curvature, radial solution, Schouten tensor, hyperbolic Gauss map, Rotational hypersurfaces, Weingarten hypersurfaces. Introduction ============ In the last 30 years, the *Nirenberg Problem*, i.e., *which functions $S:\s ^n {\longrightarrow }\r$ arise as the scalar curvature of some conformal metric on the sphere?*, has received an amazing number of contributions (see [@AM; @BC; @BE; @Ch; @CY1; @CY2; @CLi1; @CLi2; @CLn; @ES; @KW1; @KW2; @Li1; @Li2; @Mo]), but sufficient and necessary conditions for the solvability are still unknown. However, this problem opened the door of a rich subject in the last few years, *conformally invariant equations*. Let $\mathcal{F}(x_1 , \ldots , x_n)$ denote a smooth functional, and let $\Gamma \in C^{\infty}(\s ^n)$. *Does there exist a conformal metric $g = e^{2\rho}g_0$ on $\s ^n$ such that the eigenvalues $\lambda _i$ of its Schouten tensor verify $$\mathcal{F}(\lambda _1 , \ldots , \lambda _n ) = \Gamma , \, \text{ on } \s ^n .$$* Given $(M, g)$ a Riemannian manifold, for $n \geq 3$, the Schouten tensor of $g$ is given by $${\rm Sch}_g := \frac{1}{n-2}\left( {\rm Ric}(g) - \frac{S(g)}{2(n-1)}g\right)$$where ${\rm Ric}(g)$ and $S(g)$ are the Ricci tensor and the scalar curvature function of $g$ respectively. Note that, when $\mathcal{F}(x_1 , \ldots , x_n ) = x_1 +\cdots + x_n$ we have the Nirenberg Problem. Right now, the most developed topic for these equations is when we consider $\mathcal{F}(\lambda _1 , \ldots , \lambda _n)\equiv \sigma _k (\lambda _i)$ as the $k-$th elementary symmetric polynomial of its arguments equals to a constant, i.e., $$\label{1} \sigma _k (\lambda _i) = {\rm constant}.$$ Many deep results are known for these equations (see [@Ch; @CGY1; @CGY2; @Li3; @Li4; @LL1; @LL2; @Vi] and reference therein). Mostly of these results are devoted to solutions either on $\s ^n$ or $\r ^n$, and little is known when we look for conformal metrics on a domain of the sphere (see [@LN; @MP] and references therein). In this line, Chang-Han-Yang [@CHY] have classified all posible radial solution to the equation *“as guidance in studying the behavior of singular solutions in the general situation”*. This is natural since radial solutions are the simplest examples. Thus, the next step is: *under what (local) conditions can we know that the solution is radial?* In a recent paper [@EGM], the authors showed a correspondence between conformal metrics on the sphere and horospherically convex hypersurfaces in the hyperbolic space. Here, they provide a back-and-forth construction which give a hypersurface theory interpretation for the famous Nirenberg Problem, relating it with a natural formulation of the Christoffel problem in $\h ^{n+1}$. Moreover, this correspondence is more general and it relates conformally invariant equations with Weingarten hypersurfaces horospherically convex. The main line in this paper is to use the deep theorems on conformal geometry to infer results in the hypersurface theory, but, *how can the hypersurface theory help to get information on conformal geometry?* We will see here that, using the hypersurface setting, we can obtain sufficient conditions under which a conformal metric is radial among others on invariant conformal metrics under a subgroup of conformal diffeomorphisms of the sphere. We should mention that the theorems included here are local results, besides the usuals on this direction that they are from a global character. In Section $2$ we establish the necessary preliminaries on conformal geometry, and it is devoted also to summarize the correspondence developed in [@EGM] between conformal metrics and horospherically convex hypersurfaces, that is, given a conformal metric on the sphere they construct a horospherically convex hypersurface in $\h^{n+1}$ and viceversa. In Section $3$ we establish that if a conformal metric is invariant under a subgroup of conformal diffeomorphism of the sphere, then its associated horospherically convex hypersurface is invariant under the subgroup of isometries induce by the subgroup of conformal diffeomorphism, and viceversa, i.e., [**Lemma \[confisoinv\]:**]{} *Let $\phi: \Omega \subset \s ^n{\longrightarrow }\h^{n+1}$ be a locally horospherically convex hypersurface with hyperbolic Gauss map $G(x)=x$, support function $e^{\rho}:\Omega {\longrightarrow }(0,+\infty)$, and let $g=e^{2\rho}g_0$ denote its horospherical metric. Let ${T_{\left\vert \th ^{n+1}\right.}}\in {\mathcal{I}(\h ^{n+1})}$ be an isometry and $\Phi \in {\mathcal{D}(\s ^n) }$ its associated conformal diffeomorphism. Thus, if $\phi $ is ${T_{\left\vert \th ^{n+1}\right.}}-$invariant then $g$ is $\Phi -$invariant.* Conversely, let $g= e^{2\rho} g_0$ be a conformal metric defined on a domain of the sphere $\Omega \subset \s ^n$ such that the eigenvalues of its Schouten tensor, ${\rm Sch}_g$, verify $${\rm sup}{\left\{\lambda _i (x) , \, x \in \Omega , \, i = 1 ,\ldots , n \right\}}< +\infty .$$ Let $\Phi \in {\mathcal{D}(\s ^n) }$ be a conformal difeomorphism and ${T_{\left\vert \th ^{n+1}\right.}}\in {\mathcal{I}(\h ^{n+1})}$ its associated isometry. Thus, if $g $ is $\Phi-$invariant then $\phi$, given by , is ${T_{\left\vert \th ^{n+1}\right.}}-$invariant. In Section $4$ we classify the conformal metrics on the sphere whose eigenvalues of its Shouten tensor are all constant, we call these metrics *isoparametric conformal metrics*. Since the above classification have not been done before (as far as we know), we will include here. In Section $5$ we state our main results, we give sufficient conditions under which a conformal metric is radial in terms of the eigenvalues of its Shouten tensor, [**Theorem \[t1\]:**]{} *Let $g= e^{2\rho} g_0$ be a conformal metric defined on a domain of the sphere $\Omega \subset \s ^n$ such that the eigenvalues, $\lambda_i$, for $i =1, \ldots, n$, of its Schouten tensor, ${\rm Sch}_g$, verify $${\rm sup}{\left\{\lambda _i (x) , \, x \in \Omega , \, i = 1 ,\ldots , n \right\}}< +\infty .$$* Furthermore, assume that the eigenvalues satisfy $$\begin{split} \lambda &= \lambda _1 = \cdots = \lambda _{n-1} \\ \nu = \nu (\lambda) &= \lambda _n \\ \lambda - \nu & \neq 0 . \end{split}$$ Then, $g$ is radial. Moreover, we study conformal metrics on the sphere which are invariant by a $k-$parameter subgroup of conformal diffeomorphisms of the sphere, giving a bound on its maximum dimension, [**Theorem \[t2\]:**]{} *Let $g= e^{2\rho} g_0$ be a conformal metric defined on a domain of the sphere $\Omega \subset \s ^n$ such that $g \not\in \mathcal{C}(n)$ and the eigenvalues, $\lambda_i$, for $i =1, \ldots, n$, of its Schouten tensor, ${\rm Sch}_g$, verify $${\rm sup}{\left\{\lambda _i (x) , \, x \in \Omega , \, i = 1 ,\ldots , n \right\}}< +\infty .$$* Suppose that $g$ is invariant by a $k-$parameter subgroup of conformal diffeormorphism $\mathcal{G} \leq {\mathcal{D}(\s ^n) }$. Then the maximum value of $k$ is $k_{max}= \frac{n(n-1)}{2}$, and if $k = k_{max}$, the Schouten tensor of $g$, ${\rm Sch}_g$ has two eigenvalues $\lambda$ and $\nu$, where one of them, say $\lambda$, has multiplicity at least $n-1$. If, in addition, $\lambda \neq 0$, $\nu = \nu (\lambda)$ and $\nu - \lambda \neq 0$, then $g$ is radial. Finally, in Section $6$, we give some existence results for some classes of Weingarten hypersurfaces which are rotationally invariants and horospherically convex, based on a result of Chang-Han-Yang [@CHY]. Preliminaries ============= On conformal geometry --------------------- Let $(M ^n , g)$, $n\geq 3$, be a Riemannian manifold. The Riemann curvature tensor, ${\rm Riem}$, can be decomposed as $${\rm Riem} = W_g + {\rm Sch}_g \odot g ,$$where $W_g$ is the Weyl tensor, $\odot$ is the Kulkarni-Nomizu product, and $${\rm Sch}_g := \frac{1}{n-2}\left( {\rm Ric}_g - \frac{S(g)}{2(n-1)}g\right)$$is the *Schouten tensor*. Here ${\rm Ric}_g$ and $S(g)$ stand for the Ricci curvature and scalar curvature of $g$ respectively. The eigenvalues of ${\rm Sch}_g$ are defined as the eigenvalues of the endomorphism $g^{-1}{\rm Sch}_g$ and we will denote them by $\lambda _i$, $i =1, \ldots , n$. It is well known that the Schouten tensor encodes all the information on how curvature varies by a conformal change of metric. It is worth it to remark that the Weyl curvature tensor vanishes identically when $(M^n ,g)$ is locally conformally flat since it is the situation of the present work. We will consider conformal metrics to the standard metric on the $n-$sphere, $(\s ^n ,g_0)$, i.e., $$g = e^{2 \rho} g_0 .$$ \[confinv\] Let us denote by ${\mathcal{D}(\s ^n) }$ the group of conformal diffeomorphisms on the sphere and $\Phi \in {\mathcal{D}(\s ^n) }$ a conformal diffeomorphism. Let $g = e^{2 \rho} g_0$ be a conformal metric defined on a domain $\Omega \subset \s ^n$. $g$ is $\Phi -$invariant if $$g_x (u,v) = (\Phi ^* g)_x(u,v) , \, \forall x \in \Omega , \forall u,v \in T_x \s ^n , \text{ such that } \Phi (x) \in \Omega .$$ Moreover, given a continuous subgroup of conformal diffeomorphisms $\mathcal{G} \leq {\mathcal{D}(\s ^n) }$, $g$ is $\mathcal{G}-$invariant if it is $\Phi -$invariant for all $\Phi \in \mathcal{G}$. The basic example of $\mathcal{G}-$invariant metric is that which is radial symmetric, i.e., when $\mathcal{G}$ is a subgroup of rotations. In this case, we say that $g$ is radial. On hypersurface theory ---------------------- First, let us establish the necessary notation that we will use along the work. Actually, we will summarize here the construction developed in [@EGM] for the sake of completeness, that is, in order to prove our results, we will use the correspondence between conformal metrics on the sphere and locally horospherically convex hypersurfaces in $\h ^{n+1}$. So, we will remind, in a short way, how to construct a locally horospherically convex hypersurface from a conformal metric on the sphere. Let us denote by $\l ^{n+2}$ the $(n+2)-$dimensional Lorentz-Minkowski space, i.e., the vectorial space $\r ^{n+2}$ endowed with the Lorentzian metric ${\langle , \rangle }$ given by $${\langle \bar{x},\bar{x} \rangle } = - x_0 ^2 + \sum _{i=1}^{n+1} x_i ^2 ,$$where $\bar{x} \equiv (x_0 , x_1 , \ldots , x_{n+1})\in \r ^{n+2}$. So, the $(n+1)-$dimensional hyperbolic space, de-Sitter space and null cone are given, respectively, by the hyperquadrics $$\begin{split} \h ^{n+1} &= {\left\{ \bar{x} \in \l ^{n+2} : \, {\langle \bar{x},\bar{x} \rangle } = -1, \, x_0 >0\right\}}\\ \s ^{n+1}_1 &= {\left\{ \bar{x} \in \l ^{n+2} : \, {\langle \bar{x},\bar{x} \rangle } = 1\right\}}\\ \n ^{n+1}_+ &= {\left\{ \bar{x} \in \l ^{n+2} : \, {\langle \bar{x},\bar{x} \rangle } = 0, \, x_0 >0\right\}} . \end{split}$$ It is well know that $\h ^{n+1}$ inherits from $(\l ^{n+2}, {\langle , \rangle })$ a Riemannian metric which make it the standard model of Riemannian space of constant sectional curvature $-1$. Its ideal boundary at infinity, $\partial _{\infty} \h ^{n+1}$, will be denoted by $\s ^n _{\infty}$. Horospheres will play an essential role in what follows, so, we go through describing their most important properties. In this model, horospheres in $\h ^{n+1}$ are the intersection of affine degenerate hyperplanes of $\l ^{n+2}$ with $\h ^{n+1}$. Thus, it is clear that the boundary at infinity is a single point $x\in \s ^n _{\infty}$. In this way, two horospheres are always congruent, and they are at a constant (hyperbolic) distance if their respective points at infinity agree. Moreover, given a point $x \in \s ^{n}_{\infty}$, horospheres having $x$ as its point at infinity provide a foliation of $\h ^{n+1}$. From now on, $\phi : M ^n {\longrightarrow }\h ^{n+1}$ will denote an oriented immersed hypersurface and $\eta : M^n {\longrightarrow }\s ^{n+1}_1$ its unit normal. Let $\phi:M^n{\longrightarrow }\h ^{n+1}$ denote an immersed oriented hypersurface in $\h^{n+1}$ with unit normal $\eta$. The *hyperbolic Gauss map* $$G:M^n{\longrightarrow }\s_{\infty}^n \equiv \s ^n$$ of $\phi$ is defined as follows: for every $p\in M^n$, $G(p)\in \s _{\infty}^n$ is the point at infinity of the unique horosphere in $\h ^{n+1}$ passing through $\phi(p)$ and whose inner unit normal at $p$ agrees with $\eta(p)$. Associated to $\phi$, let us consider the map $$\psi := \phi + \eta : M ^n {\longrightarrow }\n ^{n+1}_+ ,$$called the *associated light cone map*. The map $\psi$ is strongly related to the hyperbolic Gauss map $G:M^n{\longrightarrow }\s _{\infty}^n$ of $\phi$. Indeed, the ideal boundary of $\n _+^{n+1}$ coincides with $\s _{\infty}^n$, and can be identified with the projective quotient space $\n _+^{n+1} / \r _+$. So, with all of this, we have $G=[\psi] :M^n{\longrightarrow }\s _{\infty}^n\equiv \n _+^{n+1} / \r _+$. Note that $\psi _0 > 0 $, being $\psi = (\psi _0 , \psi _1 , \ldots , \psi _{n+1}) \in \l ^{n+2}$. If we label $\psi _0 := e^{\rho}$, then we can interpret the *hyperbolic Gauss map* as the map $$G : M ^n {\longrightarrow }\s ^n = \n ^{n+1}_{+} \cap {\left\{x \in \l^{n+2} : \, x_0 =1\right\}}$$given by $$\label{psiG} \psi = e^{ \rho} (1, G) .$$ Moreover, we call $e^{\rho}$ the *horospherical support function*. Also, if $\{e_1,\dots, e_n\}$ denotes an orthonormal basis of principal directions of $\phi$ at $p$, and if $\kappa _1, \dots, \kappa _n$ are the associated principal curvatures, it is immediate that $$\label{metpsi} {\langle (d\psi )_p(e_i),(d\psi )_p(e_j) \rangle }= (1- \kappa _i)^2 \delta_{ij} = e^{2\rho} {\langle (dG)_p(e_i),(dG )_p(e_j) \rangle }_{\ts ^n}.$$ Coming back to horospheres, we must remark that horospheres are the unique hypersurfaces such that, innerly oriented (i.e., when the unit normal points to the convex side), its associated light cone map is constant: $\phi + \eta =v \in \n ^{n+1}_+$. Moreover, if we write $v = e^{\rho}(1,x)$, we see that $x \in \s ^n$ is the point at infinity of the horosphere and $\rho$ is the signed hyperbolic distant of the horosphere to the point $\mathcal{O}=(1,0,\ldots ,0) \in \h ^{n+1}\subset \l ^{n+2}$. In the hyperbolic setting we have a notion of convexity weaker than the usual geodesic convexity, i.e., \[horocon\] Let $M^n\subset \h^{n+1}$ be an immersed oriented hypersurface, and let $\mathcal{H}_p$ denote the horosphere in $\h^{n+1}$ that is tangent to $M^n$ at $p$, and whose interior unit normal at $p$ agrees with the one of $M^n$. We will say that $M^n$ is *horospherically convex* at $p$ if there exists a neighborhood $V\subset M^n$ of $p$ so that $V\setminus \{p\}$ does not intersect $\mathcal{H}_p$, and in addition the distance function of the hypersurface to the horosphere does not vanish up to the second order at $p$ in any direction. Thus, from , we have the following characterization of horospherically convex hypersurfaces: \[hc\] Let $\phi:M^n{\longrightarrow }\h^{n+1}$ be an oriented hypersurface. The following conditions are equivalent at $p\in M^n$. 1. All principal curvatures of $M^n$ at $p$ are simultaneously $<1$ or $>1$. 2. $M^n$ is horospherically convex at $p$. In particular, if $\phi : M^n {\longrightarrow }\h^{n+1}$ is horospherically convex at $p$, then its Gauss map verify $dG_p \neq 0$. So, if $M^n $ is horospherically convex at $p$, therefore $dG_p \neq 0$ and there exist neighborhoods $U \subset M^n $ and $\Omega \subset \s ^n $ such that $G : U {\longrightarrow }\Omega $ is a diffeomorphism, and $$g = e^{2\rho} {\langle dG,dG \rangle }_{\ts ^n}$$define a conformally flat Riemannian metric on $M^n$, called the *horospherical metric*. Since $G$ is a diffeomorphism between $U$ and $\Omega $ we can use it as a parametrization of the hypersurface, i.e., we can assume that $\phi : \Omega \subset \s ^n {\longrightarrow }\h ^{n+1} $ and $G(x) = x$ on $\Omega \subset \s ^n$. Thus, if $G : M ^n {\longrightarrow }\Omega \subseteq \s ^n $ is a global diffeormorphism of the hypersurface onto a domain of the sphere, we can use the hyperbolic Gauss map as a global parametrization of $\phi$ as above, i.e., $\phi : \Omega {\longrightarrow }\h ^{n+1}$ and $G(x)=x$. In this case, the horospherical metric is given by $$g = e^{2\rho} g_0 .$$ Now, we are ready to establish the mentioned relationship between conformal metrics on the sphere and horospherically convex hypersurfaces \[representacion\] Let $\phi: \Omega \subset \s ^n{\longrightarrow }\h^{n+1}$ be a horospherically convex hypersurface with hyperbolic Gauss map $G(x)=x$, support function $e^{\rho}:\Omega {\longrightarrow }(0,+\infty)$, and let $g=e^{2\rho}g_0$ denote its horospherical metric. Then it holds $$\label{repfor} \phi = \frac{e^{\rho}}{2}\left( 1+ e^{-2\rho} \left( 1+ ||\nabla^{g_0} \rho ||_{g_0}^2 \right)\right) (1,x) + e^{-\rho} (0, -x +\nabla^{g_0} \rho).$$ Moreover, the eigenvalues, $\lambda _i$, of the Schouten tensor of $g$, ${\rm Sch}_g$, and the principal curvatures, $\kappa _i$, of $\phi$ are related by $$\label{lambdakappa} \lambda _i = \frac{1}{2} -\frac{1}{1-\kappa _i} .$$ Conversely, given a conformal metric $g= e^{2\rho} g_0$ defined on a domain of the sphere $\Omega \subset \s ^n$ such that the eigenvalues of its Schouten tensor, ${\rm Sch}_g$, are less than $1/2$, then the map $\phi : \Omega {\longrightarrow }\h ^{n+1}$ given by defines a horospherically convex hypersurface in $\h ^{n+1}$ whose hyperbolic Gauss map is given by $G(x)=x$, $x\in \Omega$. \[remark\] We must say that the condition on the eigenvalues of the Schouten tensor is easily removable, i.e., we only need to ask that $${\rm sup}{\left\{\lambda _i (x) , \, i= 1, \ldots, n, \, x \in \Omega\right\}} < + \infty .$$ If this occurs, we can dilate the metric $g$ as $g_t = e^t g$ for $t>0$. Then, the eigenvalues of ${\rm Sch}_{g_t}$ are given by $$\lambda _i ^{t} = e^{-t} \lambda _i .$$ Thus, for $t$ big enough, we can achieve $\lambda _{i}^t < 1/2$ for $i = 1 , \ldots, n$. Conformal diffeomorphisms and isometries ======================================== Let us denote by ${\mathcal{I}(\l ^{n+2})}$, ${\mathcal{I}(\h ^{n+1})}$ and ${\mathcal{I}(\n ^{n+1}_+)}$ the group of isometries of $\l ^{n+2}$, the $(n+1)-$dimensional hyperbolic space and the $(n+1)-$dimensional null cone respectively. It is well known (see [@D]) that a conformal diffeormorphism $\Phi \in {\mathcal{D}(\s ^n) }$ induce a unique isometry in $\l ^{n+2}$, $T\in {\mathcal{I}(\l ^{n+2})}$, such that restricted to $\h ^{n+1}$ and $\n ^{n+1}_+ $ induces an isometry in these spaces and viceversa. The restrictions of $T\in {\mathcal{I}(\l ^{n+2})}$ to $\h ^{n+1} $ and $\n ^{n+1}_+$ will be denote by ${T_{\left\vert \th ^{n+1}\right.}}$ and ${T_{\left\vert \tn ^{n+1}_+\right.}}$ respectively. Moreover, each isometry ${T_{\left\vert \th ^{n+1}\right.}}\in {\mathcal{I}(\h ^{n+1})}$ induce an unique isometry ${T_{\left\vert \tn ^{n+1}_+\right.}}\in {\mathcal{I}(\n ^{n+1}_+)}$ and viceversa. \[isoinv\] Let $M^n \subset N^n $ be a domain of a $n-$manifold $N$. Let $\phi : M^n \subset N^n {\longrightarrow }\h ^{n+1}$ be a hypersurface and ${T_{\left\vert \th ^{n+1}\right.}}\in {\mathcal{I}(\h ^{n+1})}$ an isometry. $\phi $ is ${T_{\left\vert \th ^{n+1}\right.}}-$invariant if there exists $i_{{T_{\left\vert \th ^{n+1}\right.}}} : N^n {\longrightarrow }N^n $ a diffeomorphism such that $$({T_{\left\vert \th ^{n+1}\right.}}\circ \phi) (p) = \left(\phi \circ i_{{T_{\left\vert \th ^{n+1}\right.}}}\right)(p) , \forall p \in M ^n \text{ such that } i_{{T_{\left\vert \th ^{n+1}\right.}}}(p)\in M^n .$$ Moreover, given a continuous subgroup of isometries $\mathcal{T} \leq {\mathcal{I}(\h ^{n+1})}$, $\phi$ is $\mathcal{T}-$invariant if it is ${T_{\left\vert \th ^{n+1}\right.}}-$invariant for all ${T_{\left\vert \th ^{n+1}\right.}}\in \mathcal{T}$. The next result state the relationship between conformal metrics on the sphere which are invariant by a conformal diffeomorphism and horospherically convex hypersurfaces which are invariant by an isometry. \[confisoinv\] Let $\phi: \Omega \subset \s ^n{\longrightarrow }\h^{n+1}$ be a locally horospherically convex hypersurface with hyperbolic Gauss map $G(x)=x$, support function $e^{\rho}:\Omega {\longrightarrow }(0,+\infty)$, and let $g=e^{2\rho}g_0$ denote its horospherical metric. Let ${T_{\left\vert \th ^{n+1}\right.}}\in {\mathcal{I}(\h ^{n+1})}$ be an isometry and $\Phi \in {\mathcal{D}(\s ^n) }$ its associated conformal diffeomorphism. Thus, if $\phi $ is ${T_{\left\vert \th ^{n+1}\right.}}-$invariant then $g$ is $\Phi -$invariant. Conversely, let $g= e^{2\rho} g_0$ be a conformal metric defined on a domain of the sphere $\Omega \subset \s ^n$ such that the eigenvalues of its Schouten tensor, ${\rm Sch}_g$, are less than $1/2$. Let $\Phi \in {\mathcal{D}(\s ^n) }$ be a conformal difeomorphism and ${T_{\left\vert \th ^{n+1}\right.}}\in {\mathcal{I}(\h ^{n+1})}$ its associated isometry. Thus, if $g $ is $\Phi-$invariant then $\phi$, given by , is ${T_{\left\vert \th ^{n+1}\right.}}-$invariant. On one hand, if $\phi$ is horospherically convex, $\phi $ is ${T_{\left\vert \th ^{n+1}\right.}}-$invariant if and only if its associated light cone map $\psi$ is ${T_{\left\vert \tn ^{n+1}_+\right.}}-$invariant, i.e., if $$\label{proof} \left({T_{\left\vert \tn ^{n+1}_+\right.}}\circ \psi \right)(x) = \left( \psi \circ \Phi \right) (x) , \, x \in \Omega \text{ such that } \Phi (x) \in \Omega .$$being $\Phi \in {\mathcal{D}(\s ^n) }$ the conformal diffeomorphism associated to ${T_{\left\vert \tn ^{n+1}_+\right.}}\in {\mathcal{I}(\n ^{n+1}_+)}$. On the other hand, we have an explicit correspondence between conformal diffeomorphisms on the sphere and isometries on $\n ^{n+1}_+$ (see [@D Proposition 7.4]). Given an isometry ${T_{\left\vert \tn ^{n+1}_+\right.}}\in {\mathcal{I}(\n ^{n+1}_+)}$, at points $(1,x) \in \s ^n = \n ^{n+1}_{+} \cap {\left\{x \in \l^{n+2} : \, x_0 =1\right\}}$ we can see it as $${T_{\left\vert \tn ^{n+1}_+\right.}}((1,x)) = e^{-\omega (x) } (1 , \Phi(x)) ,$$then $ \Phi : \s ^n {\longrightarrow }\s ^n$ defines a conformal diffeomorphism on the $n-$sphere with conformal factor $e^{\omega}$. Conversely, given a conformal diffeomorphism $\Phi \in {\mathcal{D}(\s ^n) }$ with conformal factor $e^{\omega}$, at any point $e^{t}(1,x)\in \n ^{n+1}_+$ define $${T_{\left\vert \tn ^{n+1}_+\right.}}(e^{t}(1,x)) = e^t e^{-\omega (x)}( 1, \Phi (x)) ,$$then ${T_{\left\vert \tn ^{n+1}_+\right.}}\in {\mathcal{I}(\n ^{n+1}_+)}$. We first prove the converse. By the previous considerations, we only need to prove . Thus, if $g = e^{2\rho}g_0$ is $\Phi -$invariant, hence by Definition \[confinv\] we have that $$\rho (x) = \rho (\Phi (x)) + \omega (x), \, \text{ provided } \Phi (x)\in \Omega .$$ Let ${T_{\left\vert \tn ^{n+1}_+\right.}}$ be the isometry of $\n ^{n+1}_+$ associated to $\Phi$, then $$\begin{split} \left({T_{\left\vert \tn ^{n+1}_+\right.}}\circ \psi \right) (x) &= {T_{\left\vert \tn ^{n+1}_+\right.}}(e^{\rho (x)}(1,x)) = e^{\rho (x) - \omega (x)}(1, \Phi (x))\\ &= e^{\rho (\Phi (x))} (1, \Phi (x)) = \psi (\Phi (x)) = \left( \psi \circ \Phi \right)(x). \end{split}$$ Now, if $\phi $ is ${T_{\left\vert \th ^{n+1}\right.}}-$invariant, following the above computations, we can observe that $$\rho (x) = \rho (\Phi (x)) + \omega (x),$$being $e^{\omega} :\Omega \to \r$ the conformal factor of the conformal diffeomorphism, $\Phi$, associated to ${T_{\left\vert \th ^{n+1}\right.}}$. Thus, $g$ is $\Phi -$invariant. Isoparametric conformal metrics =============================== Here, we will classify the class of conformal metrics on the sphere such that all the eigenvalues of its Schouten tensor are constant, we denote this class by $\mathcal{C}(n)$. The local classification of conformal metrics on the class $g \in \mathcal{C}(n)$ can be done through a result of E. Cartan [@Ca]. Suppose $g \in \mathcal{C}(n)$ therefore, after possibly a dilation, the associated hypersurface given by Theorem \[representacion\] is an isoparametric hypersurface in $\h ^{n+1}$, i.e., all its principal curvatures are constant. Thus, it is a piece of either a totally umbilical hypersurface (hypersphere, horosphere, totally geodesic hyperplane and equidistant) or a standard product $\s ^k \times \h ^{n-k}$ in $\h ^{n+1}$. For this reason, we will call a metric in $\mathcal{C}(n)$ an *isoparametric conformal metric*. It is known that solutions of - $$\sigma _k (\lambda _i ) = 1 \, \, on \, \, \s ^n$$are given by conformal diffeomorphisms of the standard metric on the sphere. Such solution corresponds to a hypersphere via Theorem \[representacion\] (see [@EGM]). - $$\sigma _k (\lambda _i ) = 0 \, \, on \, \, \r ^n$$are explictly known (see [@Li3]). Such solution corresponds to a horosphere via Theorem \[representacion\] (see [@EGM]). Now, our task is to compute explicitly the horospeherical support function associated to a totally geodesic hyperplane, an equidistant hypersurface and a standard product $\s ^k \times \h ^{n-k}$. To do so, we will give the parametrization of such hypersurface and its unit normal vector field and, by means of equation , we will have an explicit formula for the horospherical support function and hyperbolic Gauss map. Thus, for an isoparametric hypersurface $\phi : \Omega \subset \r ^n {\longrightarrow }\h ^{n+1} \subset \l ^{n+2}$ with unit normal $\eta : \Omega \subset \r ^n {\longrightarrow }\s _ 1 ^{n+1} \subset \l ^{n+2}$, we will have $$\rho : \Omega \subset \r ^n {\longrightarrow }\r ,$$and $$G : \Omega \subset \r ^n {\longrightarrow }D \subset \s ^{n} \, \, \text{global diffeomorphism}.$$ Hence, the isoparametric conformal metric associated to that hypersurface is given by $$\label{isop} g = e^{\rho (G^{-1} (y))} g_0 , \, \, y \in D .$$ Let us describe the announced examples: 1. **Totally geodesic hyperplanes:** Set $\Omega = {\left\{x \in \r ^n : \, |x| < r\right\}}$ and $D= {\left\{y \in \s ^n : \, {\rm dist}_{g_0}({\bf n},y) < \pi/2 - \arcsin \left(\frac{1-r^2}{1+r^2}\right)\right\}}$, where ${\bf n}$ is the north pole. Then, $$\begin{aligned} \psi (x) &=& \left( \frac{1+r^2}{2\sqrt{r^2 -|x|}}, \frac{x}{\sqrt{r^2 - |x|^2}}, \frac{1-r^2}{2\sqrt{r^2 - |x|^2}}\right)\\ \eta (x) &=& \left( \frac{1-r^2}{2r}, {\bf 0}, \frac{1+r^2}{2r}\right) .\end{aligned}$$ Thus, from , we get $$\begin{aligned} \rho (x) &=& \ln \left( \frac{r(1+r^2)+(1-r^2)\sqrt{r^2 -|x|^2}}{2r\sqrt{r^2 -|x|^2}}\right)\\[3mm] G (x) \hspace{-3mm} &=&\hspace{-3mm} \left( \frac{2r x}{r(1+r^2)+(1-r^2)\sqrt{r^2 -|x|^2}} , \frac{r(1- r^2)+(1+r^2)\sqrt{r^2 -|x|^2}}{r(1+r^2)+(1-r^2)\sqrt{r^2 -|x|^2}}\right)\end{aligned}$$ In this case, the principal curvatures all are equals to zero, $k_i =0 $, $i =1, \ldots, n$. Thus, the eigenvalues of the Shouten tensor associated to $g$ (given by ) are $\lambda _i = -1/2$, $i=1,\ldots ,n$. 2. **Equidistant hypersurfaces:** Set $\Omega = {\left\{x \in \r^n : \, |x|<r\right\}}$ and $D= {\left\{y \in \s ^n : \, {\rm dist}_{g_0}({\bf s},y) < \pi/2 - \arcsin \left(\frac{1-r^2}{1+r^2}\right)\right\}}$, where ${\bf s}$ is the south pole. Set $t>0$, $R^2 = t^2 + r^2$ and $\beta (s) = -t + \sqrt{R^2 -s}$ for $s< r$. Then $$\begin{aligned} \psi (x) &=& \left(\frac{1+|x|^2 +\beta(|x|^2)^2}{2 \beta(|x|^2)}, \frac{x}{\beta(|x|^2)},\frac{1-|x|^2 -\beta (|x|^2)^2}{2 \beta(|x|^2)} \right)\\[3mm] \eta (x) \hspace{-3mm} &=&\hspace{-3mm} \left(\frac{(1-t^2-R^2)\sqrt{R^2-|x|^2}+2tR^2}{2R\beta (|x|^2)}, \frac{t x}{R \beta(|x|^2)},\frac{(1+t^2+R^2)\sqrt{R^2-|x|^2}-2tR^2}{2R\beta (|x|^2)}\right)\end{aligned}$$ Thus, from , we get $$\begin{aligned} \rho (x) &=& \ln\left(\frac{\alpha (|x|^2)}{2R \beta(|x|^2)}\right)\\[3mm] `G (x)\hspace{-2mm} &=&\hspace{-2mm} \left( \frac{2(R-t) x}{\alpha (|x|^2)} , \frac{R+ \sqrt{R^2 -|x|^2} + (R+t)^2 (R- \sqrt{R^2 -|x|^2})}{\alpha (|x|^2)}\right) ,\end{aligned}$$where $$\alpha (|x|^2) =R+ \sqrt{R^2 -|x|^2} + (R+t)^2 (R- \sqrt{R^2 -|x|^2}) .$$ In this case, the principal curvatures all are equals to $-t/R$, $k_i =-t/R $, $i =1, \ldots, n$. Thus, the eigenvalues of the Shouten tensor associated to $g$ (given by ) are $\lambda _i = -(R+t)/2(R-t)$, $i=1,\ldots ,n$. 3. [**$\h ^{k} \left(-\frac{1}{1+r^2}\right)\times \s ^{n-k} \left( \frac{1}{r}\right)$:**]{} For the sake of simplicity, we parametrize just a half of this hypersurface. Set $\Omega = {\left\{x \in \r ^k \right\}} \times {\left\{z \in \r ^{n-k} : \, |z| < r\right\}}\subset \r ^n$, where $r>0$, and $$D= {\left\{\left( \frac{s}{\sqrt{1+r^2+s^2}} \theta _1 , \frac{t\sqrt{1+r^2}}{r\sqrt{1+r^2+s^2}} \theta _2 ,\frac{\sqrt{1+r^2}\sqrt{r^2-t^2}}{r \sqrt{1+r^2+s^2} }\right) \in \s ^n : \, \begin{matrix} \theta _1 &\in & \s ^{k-1}\\ \theta _2 &\in & \s ^{n-k-1}\\ s &\geq & 0 \\ t & < & r \end{matrix}\right\}}.$$ Then, $$\begin{aligned} \psi (x,z)&=& \left( \sqrt{|x|^2+1+r^2}, x , z , \sqrt{r^2 -|z|^2}\right)\\ \eta (x,z)&=& \left( \frac{r \sqrt{|x|^2 +1+r^2}}{\sqrt{1+r^2}}, \frac{r x}{\sqrt{1+r^2}}, \frac{\sqrt{1+r^2}\, z}{r}, \frac{\sqrt{1+r^2}\sqrt{r^2-|z|^2}}{r}\right)\end{aligned}$$ Thus, from , we get $$\begin{aligned} \rho (x,z) &=& \ln \left( \frac{(r+\sqrt{1+r^2})\sqrt{|x|^2 +1 + r^2}}{\sqrt{1+r^2}}\right)\\ G(x,z) &=& \left( \frac{x}{\sqrt{|x|^2 +1 + r^2}}, \frac{\sqrt{1+r^2}\, z}{r \sqrt{|x|^2 +1 + r^2}}, \frac{\sqrt{1+r^2}\sqrt{r^2 - |z|^2}}{r \sqrt{|x|^2 +1 + r^2}}\right)\end{aligned}$$ In this case, the hypersuface has two principal curvatures given by, $k_i =-\frac{r}{\sqrt{1+r^2}} $, for $i =1, \ldots, k$, and $k_j = -\frac{\sqrt{1+r^2}}{r}$, for $j=k+1, \ldots , n$. Thus, the eigenvalues of the Shouten tensor associated to $g$ (given by ) are $\lambda _i = -\frac{1}{2} - r^2 + r \sqrt{1+r^2}$, for $i=1,\ldots ,k$, and $\lambda _ j = \frac{1}{2} + r^2 - r \sqrt{1+r^2}$, for $j = k+1 , \ldots , n$. As we pointed out at the begin of the Section, hyperspheres and horospheres are the only solutions for $\sigma _k (\lambda _i)= 1$ on $\s^n$ and $\sigma _k (\lambda _i) =0$ on $\r^n$. The other cases define complete metrics on a *subdomain* of the sphere. So, the natural question is: *Are these solutions the only solutions for such domains under the constrain $\sigma _k (\lambda _i) = {\rm constant}$?* Invariant conformal metrics on the sphere ========================================= In this Section we will give sufficient conditions for a conformal metric on the sphere to be radial. The following local result is based on the correspondence given in Theorem \[representacion\], Lemma \[confinv\] and a deep result of Do Carmo-Dajzcer for hypersurfaces in hyperbolic space. \[t1\] Let $g= e^{2\rho} g_0$ be a conformal metric defined on a domain of the sphere $\Omega \subset \s ^n$ such that the eigenvalues, $\lambda_i$, for $i =1, \ldots, n$, of its Schouten tensor, ${\rm Sch}_g$, verify $${\rm sup}{\left\{\lambda _i (x) , \, x \in \Omega , \, i = 1 ,\ldots , n \right\}}< +\infty .$$ Furthermore, assume that the eigenvalues satisfy $$\begin{split} \lambda &= \lambda _1 = \cdots = \lambda _{n-1} \\ \nu = \nu (\lambda) &= \lambda _n \\ \lambda - \nu & \neq 0 . \end{split}$$ Then, $g$ is radial. Consider $t>0$ big enough such that the eigenvalues of the Schouten tensor of $g_t = e^{2t}g$ are less than $1/2$ (see Remark \[remark\]). Consider the horospherically convex hypersurface, $\phi : \Omega {\longrightarrow }\h ^{n+1}$, associated to $g_t$ given by in Theorem \[representacion\]. Hence, the principal curvatures of $\phi$ verify: $$\begin{split} \tilde{\lambda} &= \kappa _1 = \cdots = \kappa _{n-1}\\ \tilde{\nu}=\tilde{\nu}(\tilde{\lambda}) &= \kappa _n \\ \tilde{\lambda} - \tilde{\nu} & \neq 0 , \end{split}$$this follows from and the assumptions on the eigenvalues of ${\rm Sch}_g$. Hence, using [@CD Theorem 4.2], $\phi (\Omega)$ is contained in a rotational hypersurface, which means, via Lemma \[confisoinv\], that $g_t$ is radial, so $g$ is radial. The next result is about determining which conformal metrics on the sphere are invariant by a $k-$parameter subgroup of conformal diffeomorphisms of the sphere. We should remove the class of conformal metrics on the sphere such that all the eigenvalues of its Schouten tensor are constant, $\mathcal{C}(n)$, but this is not a significant problem, since there are not too many of them and we have classify them. Again, the result is based on a Theorem of M. Do Carmo and M. Dajczer. \[t2\] Let $g= e^{2\rho} g_0$ be a conformal metric defined on a domain of the sphere $\Omega \subset \s ^n$ such that $g \not\in \mathcal{C}(n)$ and the eigenvalues, $\lambda_i$, for $i =1, \ldots, n$, of its Schouten tensor, ${\rm Sch}_g$, verify $${\rm sup}{\left\{\lambda _i (x) , \, x \in \Omega , \, i = 1 ,\ldots , n \right\}}< +\infty .$$ Suppose that $g$ is invariant by a $k-$parameter subgroup of conformal diffeormorphism $\mathcal{G} \leq {\mathcal{D}(\s ^n) }$. Then the maximum value of $k$ is $k_{max}= \frac{n(n-1)}{2}$, and if $k = k_{max}$, the Schouten tensor of $g$, ${\rm Sch}_g$ has two eigenvalues $\lambda$ and $\nu$, where one of them, say $\lambda$, has multiplicity at least $n-1$. If, in addition, $\lambda \neq 0$, $\nu = \nu (\lambda)$ and $\nu - \lambda \neq 0$, then $g$ is radial. As above, dilate $g$ until the eigenvalues of the Schouten tensor are less than $1/2$. Now, construct the horospherically convex hypersurface given by Theorem \[representacion\]. The hypothesis on the $\mathcal{G}-$invariance of $g$ is translated into a $\mathcal{T}-$invariance of $\phi$ under a $k-$parameter subgroup $\mathcal{T}\leq {\mathcal{I}(\h ^{n+1})}$. Thus, applying now [@CD Theorem 4.7] we obtain the result. The above results hold for $n \geq 3$. It is clear that for $n=2$ are false. A note on rotational hypersurfaces in $\h ^{n+1}$ ================================================= In a recent paper [@CHY], authors have classified all possible radial solution to the equation $$\sigma _k (\lambda _i) = c , \, \, c \equiv {\rm constant},$$that is, they consider conformal metrics $g = v(|x|)^{-2}|dx|^2$ on domains of the form $${\left\{x \in \r ^n , \, r_1 < |x| < r_2\right\}} ,$$ being $\sigma _k (\lambda _i)$ the $k-$th elementary symmetric function of the eigenvalues of ${\rm Sch}_g$, and $0 \leq r_1 < r_2 \leq \infty$. From the point of view of hypersurfaces in hyperbolic space, this classification result means (up to possibly a dilatation) that they have classified all rotational horospherically convex hypersurfaces verifying the Weingarten relationship $$\sigma _k \left( \frac{1+\kappa _i}{2(1- \kappa _i)}\right) = \tilde{c} , \, \, \tilde{c} \equiv {\rm constant} .$$ It will be too long to describe here all these solutions, but we would like to mention two cases when $c >0$: [**Case I.1**]{} and [**Case I.3.a**]{} in [@CHY Theorem 1] give the existence of hyperspheres (which was already known) and Delaunay-type hypersurfaces respectively. An interesting application of the above hypersurfaces could be to use them as barriers for Plateau problem at infinity in the hyperbolic space for certain Weingarten functionals. [**Acknowledgement:**]{} Author wants to thank J.A. Gálvez, Y.Y. Li and R. Mazzeo for their interesting comments and help during the preparation of this work. [999999]{} A. Ambrosetti, A. Malchiodi, On the symmetric scalar curvature problem on $S\sp n$, [*J. Differential Equations*]{} [**170**]{} (2001), 228–245. A. Bahri, J. M. Coron, The scalar-curvature problem on three-dimensional sphere, [*J. Funct. Anal.*]{} [**95**]{} (1991), 106–172. J.P. Bourguignon, J.P. Ezin, Scalar curvature functions in a conformal class of metrics and conformal transformations, [*Trans. Amer. Math. Soc.*]{} [**301**]{} (1987), 723–736. R.L. Bryant, Surfaces of mean curvature one in hyperbolic space, [*Astérisque*]{}, [**154-155**]{} (1987), 321–347. E. Cartan, Familles de surfaces isoparemétriques dans les espaces á courboure constant, [*Ann. Mat. Pura Appl. (1)*]{} [**17**]{} (1938), 177–191. M. do Carmo, M. Dajczer, Rotation hypersurfaces in spaces of constant curvature, [*Trans. Amer. Math. Soc.*]{} [**277**]{} (1983), 685–709. S.Y.A. Chang, Conformal invariants and partial differential equations, [*Bull. Amer. Math. Soc.*]{} [**42**]{} (2005), 365–393. S.Y.A. Chang, M.J. Gursky, P.C. Yang, An equation of Monge-Ampère type in conformal geometry, and four-manifolds of positive Ricci curvature, [*Ann. of Math.*]{} [**155**]{} (2002), 709–787. S.Y.A. Chang, M.J. Gursky, P.C. Yang, An a priori estimate for a fully nonlinear equation on four-manifolds, [*J. Anal. Math.*]{} [**87**]{} (2002), 151–186. S.Y.A. Chang, Z-C. Han, P. Yang, Calssification of singular radial solution to the $\sigma _k$ Yamabe equation on annular domains, [*Preprint*]{}. S.Y.A. Chang, P.C. Yang, Prescribing Gaussian curvature on $S^2$, [*Acta. Math.*]{} [**159**]{} (1987), 215–259. S.Y.A. Chang, P.C. Yang, Conformal deformation of metrics on $S^2$, [*J. Diff. Geom.*]{} [**27**]{} (1988), 259–296. W. Chen, C. Li, A priori estimates for prescribing scalar curvature equations, [*Ann. of Math.*]{} [**145**]{} (1997), 547–564. W. Chen, C. Li, Prescribing scalar curvature on $S\sp n$, [*Pacific J. Math.*]{} [**199**]{} (2001), 61–78. C.C. Chen, C.S. Lin, Prescribing scalar curvature on $S\sp N$. I. A priori estimates, [*J. Differential Geom.*]{} [**57**]{} (2001), 67–171. M. Dajczer, Submanifolds and isometric immersions. [*Mathematics Lectures Series*]{} [**13**]{}, (1990). C.L. Epstein, The hyperbolic Gauss map and quasiconformal reflections, [*J. Reine Angew. Math.*]{} [**372**]{} (1986), 96–135. C.L. Epstein, The asymptotic boundary of a surface imbedded in $H\sp 3$ with nonnegative curvature, [*Michigan Math. J.*]{} [**34**]{} (1987), 227–239. C.L. Epstein, Envelopes of horospheres and Weingarten surfaces in Hyperbolic 3-space. [*Unpublished*]{} (1986). J. F. Escobar, R. M. Schoen, Conformal metrics with prescribed scalar curvature, [*Invent. Math.*]{} [**86**]{} (1986), 243–254. J.M. Espinar, J.A. Gávez, P. Mira, Hypersurfaces in $\h^{n+1}$ and conformally invariant equations: the generalized Christoffel and Nirenberg problems. [*To appears in Jour. Eur. Math. Soc.*]{} Q. Jin, Y.Y. Li, H. Xu, Symmetry and Asymmetry: The method of moving spheres, [*Advances in Diff. Equations*]{} [**13**]{}, (2008), 601–640. J.L. Kazdan, F. Warner, Scalar curvature and conformal deformation of Riemannian structure, [*J. Diff. Geometry*]{} [**10**]{} (1975), 113–134. J.L. Kazdan, F. Warner, Existence and conformal deformation of metrics with prescribed Gaussian and scalar curvatures, [*Ann. of Math.*]{} [**101**]{} (1975), 317–331. Y.Y. Li, Prescribing scalar curvature on $S^n$ and related problems, Part I, [*J. Differential Equations*]{} [**120**]{} (1995), 319–410. Y.Y. Li, Prescribing scalar curvature on $S^n$ and related problems, Part II: Existence and compactness, [*Comm. Pure Appl. Math.*]{} [**49**]{} (1996), 541–597. Y.Y. Li, Degenerate conformally invariant fully nonlinear elliptic equations. [*Arch. Rational Mech. Anal.*]{} [**186**]{} (2007), 25–51. Y.Y. Li, Local gradient estimates of solutions to some conformally invariant fully nonlinear equations. [*C. R. Acad. Sci. Paris, Ser. I*]{} [**343**]{} (2006), 249–252. Y.Y. Li, Some nonlinear elliptic equations from geometry, [*Proc. Natl. Acad. Sci. USA*]{} [**99**]{} (2002), 15287–15290. A. Li, Y.Y. Li, On some conformally invariant fully nonlinear equations, [*Comm. Pure Appl. Math.*]{} [**56**]{} (2003), 1416–1464. A. Li, Y.Y. Li, On some conformally invariant fully nonlinear equations. II. Liouville, Harnack and Yamabe. [*Acta Math.*]{} [**195**]{} (2005), 117–154. C. Loewner, L. Nirenberg, Partial differential equations invariant under conformal and projective transformations. [*Contributions to Analysis*]{}, Academic Press, 1975, 245–275. R. Mazzeo, F. Pacard, A construction of singular solutions for a semilinear elliptic equation using asymptotic analysis. [*J. Diff. Geom.*]{} [**99**]{} nº2 (1999), 331–370. J. Moser, On a nonlinear problem in differential geometry. [*Dynamical Systems*]{}, Academic Press, N.Y., 1973, 273–280. J.M. Schlenker, Hypersurfaces in $\h ^n$ and the space of its horospheres, [*Geom. Funct. Anal.*]{} [**12**]{} (2002), 395–435. J. Viaclovsky, Conformal geometry, contact geometry, and the calculus of variations, [*Duke Math. J.*]{} [**101**]{} (2000), 283–316. [^1]: The author is partially supported by Spanish MEC-FEDER Grant MTM2007-65249, and Regional J. Andalucia Grant P06-FQM-01642

ArXiv

--- abstract: 'Intelligent assistants change the way people interact with computers and make it possible for people to search for products through conversations when they have purchase needs. During the interactions, the system could ask questions on certain aspects of the ideal products to clarify the users’ needs. For example, previous work proposed to ask users the exact characteristics of their ideal items [@zhang2018towards; @Sun:2018:CRS:3209978.3210002] before showing results. However, users may not have clear ideas about what an ideal item looks like, especially when they have not seen any item. So it is more feasible to facilitate the conversational search by showing example items and asking for feedback instead. In addition, when the users provide negative feedback for the presented items, it is easier to collect their detailed feedback on certain properties (aspect-value pairs) of the non-relevant items. By breaking down the item-level negative feedback to fine-grained feedback on aspect-value pairs, more information is available to help clarify users’ intents. So in this paper, we propose a conversational paradigm for product search driven by non-relevant items, based on which fine-grained feedback is collected and utilized to show better results in the next iteration. We then propose an aspect-value likelihood model to incorporate both positive and negative feedback on fine-grained aspect-value pairs of the non-relevant items. Experimental results show that our model is significantly better than state-of-the-art product search baselines without using feedback and those baselines using item-level negative feedback.' author: - 'Keping Bi$^1$, Qingyao Ai$^1$, Yongfeng Zhang$^2$, W. Bruce Croft$^1$' bibliography: - 'reference.bib' title: Conversational Product Search Based on Negative Feedback ---

ArXiv

--- abstract: 'The ExoMol database (www.exomol.com) provides molecular data for spectroscopic studies of hot atmospheres. While the data is intended for studies of exoplanets and other astronomical bodies, the dataset is widely applicable. The basic form of the database is extensive line lists; these are supplemented with partition functions, state lifetimes, cooling functions, Landé g-factors, temperature-dependent cross sections, opacities, pressure broadening parameters, $k$-coefficients and dipoles. This paper presents the latest release of the database which has been expanded to consider 80 molecules and 190 isotopologues totaling over 700 billion transitions. While the spectroscopic data is concentrated at infrared and visible wavelengths, ultraviolet transitions are being increasingly considered in response to requests from observers. The core of the database comes from the ExoMol project which primarily uses theoretical methods, albeit usually fine-tuned to reproduce laboratory spectra, to generate very extensive line lists for studies of hot bodies. The data has recently been supplemented by line lists deriving from direct laboratory observations, albeit usually with the use of [*ab initio*]{} transition intensities. A major push in the new release is towards accurate characterisation of transition frequencies for use in high resolution studies of exoplanets and other bodies.' address: | $^{1}$Department of Physics and Astronomy, University College London, London, WC1E 6BT, UK\ $^{2}$Present address: SRON Netherlands Institute for Space Research, Sorbonnelaan 2, 3584 CA, Utrecht, Netherlands\ $^{3}$ Present address: Department of Physics, Aberystwyth University, Ceredigion, UK, SY23 3BZ, UK\ $^{4}$Present address: Nuclear Data Section, International Atomic Energy Agency, Vienna A-1400, Austria\ $^{5}$ Present address: School of Chemistry, University of New South Wales, 2052 Sydney, Australia\ $^{6}$ Present address: Nankai University, 94 Weijin Road, Tianjin, China\ $^{7}$ Department of Physics, University of Oxford, England\ $^{8}$ Department of Physics and Astronomy, University of the Western Cape, South Africa\ $^{9}$ Atomic and Molecular Physics Division, Center for Astrophysics $|$ Harvard & Smithsonian, Cambridge, MA. USA. author: - 'Jonathan Tennyson$^{1}$[^1]' - 'Sergei N. Yurchenko$^{1}$' - 'Ahmed F. Al-Refaie$^{1}$' - 'Victoria H. J. Clark$^{1}$' - 'Katy L. Chubb$^{1,2}$' - 'Eamon K. Conway$^{1,9}$' - 'Akhil Dewan$^{1}$' - 'Maire N. Gorman$^{1,3}$' - 'Christian Hill$^{1,4}$' - 'A. E. Lynas-Gray$^{1,7,8}$' - 'Thomas Mellor$^{1}$' - 'Laura K. McKemmish$^{1,5}$' - 'Alec Owens$^{1}$' - 'Oleg L. Polyansky$^{1}$' - 'Mikhail Semenov$^{1}$' - 'Wilfrid Somogyi$^{1}$' - 'Giovanna Tinetti$^{1}$' - 'Apoorva Upadhyay$^{1}$' - 'Ingo Waldmann$^{1}$' - 'Yixin Wang$^{1,6}$' - Samuel Wright$^1$ - 'Olga P. Yurchenko$^{1}$' title: 'The 2020 release of the ExoMol database: molecular line lists for exoplanet and other hot atmospheres' --- infrared ,visible ,Einstein $A$ coefficients ,transition frequencies ,partition functions ,cooling functions ,lifetimes ,cross sections ,$k$ coefficients ,Landé $g$-factors, pressure broadening Introduction ============ The ExoMol project was started in 2011 [@jt528] with the purpose of providing molecular line lists for studies of exoplanets and other (hot) atmospheres. Besides data demands for exoplanets [@13TiEnCo.exo], other hot astronomical bodies with significant molecular content in their atmospheres include cool stars [@09Bernath.exo] and brown dwarfs [@aha97; @09Bernath.exo]. On Earth, similar spectroscopic data are required to study flames [@14BoWeHy.H2O; @15CoLixx], discharge plasmas [@02ChWaLi.SO], explosions [@11CaLiPi.H2O] and the gases emitted from smoke stacks [@12EvFaCl.CO2]. In addition, bodies in non-local thermodynamic equilibrium (non-LTE) such as comets and masers give observable emissions from highly excited states [@jt330; @jt349; @16GrBaRi.H2O]. ExoMol data are proving very popular among the exoplanet atmospheric modellers owing to their extensive coverage of the molecular species and their completeness as function of both frequency and, very importantly, temperature. ExoMol data have been incorporated into many radiative transfer and retrieval codes including Tau-REx [@jt593; @jt611; @TauRexIII], Phoenix [@PHOENIX], NEMESIS [@NEMESIS], CHIMERA [@CHIMERA], HELIOS [@HELIOS] and HELIOS-r2 [@Helios-r2], ATMO [@ATMO; @18GoMaSi.exogen; @20PhTrBa], ARCiS [@19MiOrCh.arcis], ARTES [@ARTES], HyDRA [@HyDRA], GENESIS [@17GaMaxx.exoplanet], petitRADTRANS [@19MoWaBo.petitRADTRANS], PLATON [@PLATON], VSTAR [@12BaKe; @jt572], BART [@BART] and Pyrat Bay [@PyratBay]. In addition the ExoMol database has been used to support a variety of other studies. Examples including the analysis of gaseous ammonia in Jupiter [@jt745], tentative detection of H$_2$S in Uranus [@19IrToGa.H2S], detection of CaO in meteors [@18BeBoSa.CaO] and SiO in B\[e\] Supergiants [@15KrOkCi.SiO], Late-type stars [@19Pavlenko] or circumstellar environment [@19EvPaBa]. Detections in the atmospheres of exoplanets are discussed below. More diverse applications include modeling laser analysis of Solar System objects [@15HuLuMe; @16HuLuCo], isotope abundance quantification in stars [@jt799], search for molecules highly sensitive to proton-to-electron mass ratio variation [@18OwYuSp; @20SyMoCu.applications], models of molecular steering [@jt663; @18OwYa] , design of THz lasers [@19BuMaEl.applications], design of dark matter detection schemes [@19EsPeRa] and the design of novel propulsion systems [@20Markov.NH3]. A number of diatomic molecules for which ExoMol provides extensive hot line lists have been detected in sunspots. These include SiO [@79JoPuPa.SiO], SiH [@12SrAmSh.SiH], SH [@79JoPaXX.SH], BeH [@08ShBaRa.BeH], VO [@08SrBaRa.VO], as well as MgH, MgO, TiO, C$_2$ and CaH which have all long been observed in sunspots [@71Sotiro]. In addition the spectrum of water is well-known in sunspots [@jt200; @jt212; @jt251; @jt297; @06SoWiSc.H2O]; studies of water spectra in sunspots should be aided by the new POKAZATEL water line list [@jt734] which is designed for studies at temperatures above 3000 K such as those encountered in solar umbra and penumbra. ExoMol line lists have been published as a series in the journal Monthly Notices of the Royal Astronomical Society (see below) and a first data release in 2016 [@jt631], henceforth ExoMol2016, provided full documentation for the ExoMol database which can be found at [www.exomol.com](www.exomol.com). The basic form of the database is extensive line lists; these are supplemented with partition functions, state lifetimes, cooling functions, Landé g-factors, temperature-dependent cross sections, opacities, pressure broadening parameters, $k$-coefficients and dipoles. The ExoMol website also offers an extensive bibliography database on research literature on molecules relevant for ExoMol applications. So far the bibliography database contains 6709 sources. Since ExoMol started there have been major changes in exoplanet science which have driven further expansion and development of the ExoMol database. The first is the discovery of rocky exoplanets orbiting so near to their host star that their surfaces are likely to be molten. The atmospheres of these planets are thought to contain a variety of species not considered previously [@15ItIkKa.exo; @16FeJaWi; @jt693; @20HeWoHe]. Secondly higher temperature planets have been observed than anyone anticipated, the poster child for hot exoplanets is Kelt-9b which is thought to reach temperatures of 4000 K [@17GaStCo.exo], hotter than most of the stars in our Galaxy! This has led to the need to construct line list for key species over an extended temperature range. Third is the development of Doppler-shift high resolution spectroscopy of exoplanets [@14Snellen; @18Birkby] which has proved a powerful tool for detecting molecules but fails in the absence of highly accurate molecular line lists [@15HoDeSn.TiO]. These developments have led us to both expand the range of molecules included in the database and to begin a systematic attempt to improve the accuracy of the line positions for the line lists contained in the database. Progress on both of these objectives is described below as well work on extending the coverage of the database into the ultraviolet. Of course ExoMol is not the only database providing spectroscopic data for atmospheric studies. For the Earth’s atmosphere, databases HITRAN [@jt350; @jt453; @jt546] and GEISA [@jt504; @jt636] provide comprehensive and validated datasets for approximately 50 key molecules. However, these databases are designed for studies at room temperature and below and do not contain the necessary data for adequately calculating radiative transport in hot bodies. The HITEMP database was constructed to extend HITRAN to higher temperatures. The latest, 2010, release [@jt480] only contained data on 5 molecules (OH, NO, CO, H$_2$O and CO$_2$) and improved hot line lists are available for all these molecules; an update to HITEMP is currently in progress [@jt763; @20HaGoRe.CH4; @jtH2OHITEMP2020]. Other relevant databases include TheoReTS [@TheoReTS], which contains hot line lists for 8 polyatomic molecules, and Kurucz’s compilation of data, which is very complete for atomic sources but contains data only on about ten diatomic molecules [@11Kurucz.db] all of which are covered by ExoMol, usually to higher accuracy. The MoLLIST data base of Bernath and co-workers [@MOLLIST] contains empirically derived line lists for about 20 diatomic species. These line lists have recently been incorporated into the ExoMol database [@jt790], see Table \[tab:otherdata\] below. Databases of hot molecular spectra for other specialised applications include those for combustion [@12RiSoxx; @15CoLixx] and studies of laser-induced plasmas [@15PaWoSu]. ExoMol line lists together with data from other sources will serve as the main data source for exoplanet stands planned for the upcoming space missions Ariel and JWST. The methodology [@jt511; @jt573; @jt624; @jt632; @jt654; @jt656; @jt693] and software [@jt609; @jt626] developed and used by the ExoMol project has been extensively documented elsewhere. Here we will only consider those aspects which impinge on the new ExoMol2020 release. Database coverage ================= The ExoMol project aims at complete coverage of the spectroscopic properties of molecules which are deemed to be important in hot astrophysical environments. Coverage concerns (a) the molecular species considered, including isotopologues; (b) the spectroscopic and thermodynamic properties considered; (c) the frequency range considered and (d) the upper temperature range for which the data are reasonably complete. Both the required temperature and frequency range completeness are to some extent a judgement on what is required for astronomical and other studies. Unlike ExoMol2016, the ExoMol2020 database essentially provides a complete dataset for modelling hot atmospheres. While new species are still being added, commonly at the request of users, the database now contains molecular data for 80 molecules and covers the key molecules thought to be important for exoplanetary studies. In many cases there is more than one line list for a given isotopologue. For this reason it is our practice to name each line list, including (unnamed) ones we have imported from other sources. In the case of multiple line lists the ExoMol website provides a recommended line list for the given species. This recommendation is for studies of hot atmospheres; other available line lists may be more suitable for room temperature studies. For ease of understanding we have divided the description of line lists provided into three tables. Tables \[tab:exomoldata\] and \[tab:otherdata\] summarise the molecules for which data are provided by ExoMol and give the characteristics of the line list in each case. Table \[tab:exomoldata\] details line lists which have been formally published as part of the ExoMol project while Table \[tab:otherdata\] shows line lists imported from other sources which have been recast in the ExoMol format (see Wang [*et al*]{} [@jt790] for example); these line lists are fully integrated into the database. Table \[tab:otherdata2\] lists the main extra line lists available through the ExoMol website; this table is not comprehensive but the other line lists available on the ExoMol website and not listed in Table 1 – 3 are largely of historic interest only. Paper Molecule $N_{\rm iso}$ $T_{\rm max}$ $N_{\rm elec}$ $N_{\rm lines}$$^a$ DSName Reference --------- ------------------ --------------- --------------- ---------------- --------------------- ----------- -- ----------------- I BeH 1 2000 1 16 400 Yadin @jt529 I MgH 3 2000 1 10 354 Yadin @jt529 I CaH 1 2000 1 15 278 Yadin @jt529 II SiO 5 9000 1 254 675 EJBT @jt563 III HCN/HNC 1$^a$ 4000 1 399 000 000 Harris @jt570 IV CH$_4$ 1 2000 1 34 153 806 005 YT34to10 @jt564 [@jt698] V NaCl 2 3000 1 702 271 Barton @jt583 V KCl 4 3000 1 1 326 765 Barton @jt583 VI PN 2 5000 1 142 512 YYLT @jt590 VII PH$_3$ 1 1500 1 16 803 703 395 SAlTY @jt592 VIII H$_2$CO 1 1500 1 10 000 000 000 AYTY @jt597 IX AlO 4 8000 3 4 945 580 ATP @jt598 X NaH 2 7000 2 79 898 Rivlin @jt605 XI HNO$_3$ 1 500 1 6 722 136 109 AlJS @jt614 XII CS 8 3000 1 548 312 JnK @jt615 XIII CaO 1 5000 5 21 279 299 VBATHY @jt618 XIV SO$_2$ 1 2000 1 1 300 000 000 ExoAmes @jt635 XV H$_2$O$_2$ 1 1250 1 20 000 000 000 APTY @jt638 XVI H$_2$S 1 2000 1 115 530 3730 AYT2 @jt640 XVII SO$_3$ 1 800 1 21 000 000 000 UYT2 @jt641 XVIII VO 1 2000 13 277 131 624 VOMYT @jt644 XIX H$_2$$^{17,18}$O 2 3000 1 519 461 789 HotWat78 @jt665 XX H$_3^+$ 1$^b$ 3000 1 11 500 000 000 MiZATeP @jt666 XXI NO 6 5000 2 2 281 042 NOName @jt686 XXII SiH$_4$ 1 1200 1 62 690 449 078 OY2T @jt701 XXIII PO 1 5000 1 2 096 289 POPS @jt703 XXIII PS 1 5000 3 30 394 544 POPS @jt703 XXIV SiH 4 5000 3 1 724 841 SiGHTLY @jt711 XXV SiS 12 5000 1 91 715 UCTY @jt724 XXVI NS 6 5000 1 3 479 067 SNaSH @jt725 XXVI HS 6 5000 1 219 463 SNaSH @jt725 XXVII C$_2$H$_4$ 1 700 1 60 000 000 000 MaYTY @jt729 XXVIII AlH 3 5000 3 40 000 AlHambra @jt732 XXIX CH$_3$Cl 2 1200 1 166 279 593 333 OYT @jt733 XXX H$_2$$^{16}$O 1$^c$ 5000 1 5 745 071 340 POKAZATEL @jt734 XXXI C$_2$ 3 5000 8 6 080 920 8states @jt736 XXXII MgO 5 5000 5 72 833 173 LiTY @jt759 XXXIII TiO 5 5000 13 59 000 000 Toto @jt760 XXXIV PH 1 4000 2 65 055 LaTY @jt765 XXXV NH$_3$ 1$^d$ 1500 1 16 900 000 CoYuTe @jt771 XXXVI SH 2 3000 2 572 145 GYT @jt776 XXXVII HCCH 1 2000 1 4 347 381 911 aCeTY @jt780 XXXVIII 1 3000 1 32 951 275 437 OYT3 @jt797 XXXIX 1 3000 1 7 996 570 390 UCL-4000 @jt804 XL 1 1500 1 2 089 331 073 eXeL @jt805 : Datasets created by the ExoMol project and included in the ExoMol database.[]{data-label="tab:exomoldata"} Paper Number in series published in Mon. Not. R. Astron. Soc.\ $N_{\rm iso}$ Number of isotopologues considered;\ $T_{\rm max}$ Maximum temperature for which the line list is complete;\ $N_{\rm elec}$ Number of electronic states considered;\ $N_{\rm lines}$ Number of lines: value is for the main isotope.\  indicates line list recommended for studies of hot atmospheres.\ $^a$ The Larner line list for H$^{13}$CN/HN$^{13}$C due to @jt447 is recommended.\ $^b$ The ST line list for H$_2$D$^+$ due to @jt478 is recommended,.\ $^c$ The VTT line list for HDO due to @jt469 is recommended.\ $^d$ There is a room temperature $^{15}$NH$_3$ line list due to @15Yurche.NH3.\ Molecule $N_{\rm iso}$ $T_{\rm max}$ $N_{\rm elec}$ $N_{\rm lines}$ DSName Reference Methodology ---------------- --------------- --------------- ---------------- ----------------- ------------- -- ------------------------------------------- ------------- 1 10000 1 4712 RACPPK @19RoAbCz.H2 Ab initio CH 1 5000 4 52201 MoLLIST @14MaPlVa.CH Empirical NH 1 5000 1 12150 MoLLIST @14BrBeWe.NH [@15BrBeWe.NH; @18FeBeHo.NH] Empirical OH 1 5000 2 54276 MoLLIST @16BrBeWe.OH Empirical AlCl 2 5000 1 20245 MoLLIST @18YoBexx.AlF Empirical AlF 1 5000 1 40490 MoLLIST @18YoBexx.AlF Empirical OH$^+$ 1 5000 2 12044 MoLLIST @17HoBexx.OH+ [@18HoBiBe.OH+] Empirical CaF 1 5000 6 14817 MoLLIST @18HoBexx.CaF Empirical MgF 1 5000 3 8136 MoLLIST @17HoBexx.MgF Empirical KF 1 5000 2 10572 MoLLIST @16FrBeBr.NaF Empirical NaF 1 5000 1 7884 MoLLIST @16FrBeBr.NaF Empirical LiCl 1 5000 4 26260 MoLLIST @18BiBexx.LiF Empirical LiF 1 5000 2 10621 MoLLIST @18BiBexx.LiF Empirical MgH 2 5000 1 14179 MoLLIST @13GhShBe.MgH Empirical TiH 1 5000 3 181080 MoLLIST @05BuDuBa.TiH Empirical CrH 1 5000 2 13824 MoLLIST @06ChMeRi.CrH Empirical FeH 1 5000 2 93040 MoLLIST @10WEReSe.FeH Empirical HF 2 5000 1 7956 Coxon-Hajig @15CoHaxx.HF Empirical HCl 4 5000 1 2588 HITRAN @11LiGoBe.HCl Empirical CP 1 5000 2 28752 MoLLIST @14RaBrWe.CP Empirical CN 1 5000 3 195120 MoLLIST @14BrRaWe.CN Empirical C$_2$ 1 5000 2 47 570 MoLLIST @13BrBeSc.C2 Empirical CaH 2 5000 1 6000 MoLLIST @12LiHaRa.CaH [@13ShRaBe.CaH] Empirical N$_2$ 1 10000 4$^a$ 7 182 000 WCCRMT @18WeCaCr.N2 Empirical SiO 1 5000 3 6 67 814 Kurucz-SiO @11Kurucz.db Empirical ScH 1 5000 6 1 152 827 LYT @jt599 Ab initio LiH 1 12000 1 18 982 CLT @jt506 Ab initio LiH$^+$ 1 12000 1 332 CLT @jt506 Ab initio CO 9 9000 1 752 976 Li2015 @15LiGoRo.CO Empirical HeH$^+$ 4 9000 1 1430 ADSJAAM @19AmDiJo Ab initio HD$^+$ 1 9000 1 10285 ADSJAAM @19AmDiJo Ab initio HD 1 9000 1 5939 ADSJAAM @19AmDiJo Ab initio CH$_3$F 1 300 1 139 188 215 OYKYT @19OwYaKu.CH3F Ab initio AsH$_3$ 1 300 1 3 600 000 CYT18 @jt751 Ab initio P$_2$H$_2$$^b$ 2 300 1 10 667 208 951 OY-Trans @19OwYuxx.P2H2 Ab initio P$_2$H$_2$$^b$ 2 300 1 11 020 092 365 OY-Cis @19OwYuxx.P2H2 Ab initio PF$_3$ 1 300 1 68 000 000 000 MCYTY @jt752 Ab initio CH$_3$ 1 1500 1 2 058 655 166 AYYJ @19AdYaYu.CH3 Ab initio BeH 3 5000 2 592308 Darby-Lewis @jt722 ExoMol CO$_2$ 13 4000 1 298 323 789 Ames-2016 @13HuFrTa.CO2 [@Huang2017] ExoMol-like SiH$_2$ 1 2000 1 254 061 207 CATS @jt779 ExoMol YO 1 5000 6 3520133 SSYT @jt774 Ab initio : Datasets not created as part of the ExoMol project but included in the ExoMol database.[]{data-label="tab:otherdata"} $N_{\rm iso}$ Number of isotopologues considered;\ $T_{\rm max}$ Maximum temperature for which the line list is complete;\ $N_{\rm elec}$ Number of electronic states considered;\ $N_{\rm lines}$ Number of lines: value is for the main isotope.\  indicates line list recommended for studies of hot atmospheres.\ $^a$ The WCCRMT line list considers triplet states only.\ $^b$ There are separate line lists for cis and trans P$_2$H$_2$.\ Molecule $N_{\rm iso}$ $T_{\rm max}$ $N_{\rm elec}$ $N_{\rm lines}$ DSName Reference Methodology ---------- --------------- --------------- ---------------- ----------------- ------------ ---------- ------------------------- ------------- H$_3^+$ 2$^a$ 4000 1 3 070 571 NMT @jt181 ExoMol H$_2$O 2$^b$ 3000 1 505 806 202 BT2 @jt378 ExoMol NH$_3$ 2$^c$ 1500 1 1 138 323 351 BYTe @jt500 ExoMol HeH$^+$ 4 10000 1 1 431 Engel @jt347 Ab initio HD$^+$ 1 12000 1 10 119 CLT @jt506 Ab initio CO$_2$ 13 300 1 161944 Zak @jt625 [@jt667; @jt678] ExoMol CO$_2$ 1 4000 1 628 324 454 CSSD-4000 @11TaPexx.CO2 Empirical H$_2$O 1 300 1 WAT\_UV296 [@jt803] ExoMol : Supplementary datasets available from the website.[]{data-label="tab:otherdata2"} $N_{\rm iso}$ Number of isotopologues considered;\ $T_{\rm max}$ Maximum temperature for which the line list is complete;\ $N_{\rm elec}$ Number of electronic states considered;\ $N_{\rm lines}$ Number of lines: value is for the main isotope.\ Individual line lists ===================== An overview of the line lists in the ExoMol database is given in Fig. \[fig:linelist\]. One general issue is that Medvedev and co-workers [@highv; @15LiGoRo.CO] identified a numerical problem with the intensities of high overtone transitions computed with the standard compilation of the diatomic vibration-rotation program [Level]{} [@LEVEL]. Our diatomic line lists computed with [Level]{} or [Duo]{} [@jt608] have been adjusted to remove transitions affected by this issue. Medvedev [*et al.*]{} [@jt794] recently identified similar issues with triatomic systems but tests suggest that in practice our triatomic line lists seem largely unaffected by the problem. Below we consider some of the line lists presented in the ExoMol database and listed in Tables \[tab:exomoldata\] and \[tab:otherdata\]. We restrict our discussion to issues not covered in the ExoMol2016 release or the original publications. We start by considering the 42 molecules for which line lists have been created as part of the ExoMol series as listed in Table \[tab:exomoldata\]. ![ Molecular line lists: red indicates line lists in progress, blue corresponds to the line lists suggested for molecules specific for hot rocky exoplanets and green indicates line lists which contain data applicable for high resolution.[]{data-label="fig:linelist"}](mol2.jpg){width="75.00000%"} Diatomics --------- ### AlH, paper XVIII There is a new AlHambra line list for AlH [@jt732]. A MARVEL (measured active rotational-vibrational energy levels [@jt412]) project was performed as part of this project meaning that many transitions are predicted with experimental accuracy. The AlHambra line list has been updated to give the uncertainty in the energy in the states file which allows users to determine the uncertainty in a given transition wavenumber. ### AlO, paper IX No change. The APTY line list was recently used to make a detection of AlO in an exoplanet atmosphere, the ultra-hot Jupiter exoplanet WASP-33 b [@19VaMaWe.AlO] and hot Jupiter exoplanet WASP-43 b [@20ChMiKa]. ### BeH, paper I The BeH Yadin line list was one of the first constructed by ExoMol [@jt529]; however, the line list only included transitions within the X $^2\Sigma^+$ ground electronic state. Recently @jt722 constructed line lists for BeH, BeD and BeT which consider both the ground and first excited (A $^2\Pi$) states. At the same time Darby-Lewis [*et al.*]{} performed a MARVEL analysis and used their empirical levels in the fit. The Darby-Lewis line list is therefore more accurate and more extensive than the Yadin one which it replaces in the ExoMol database and is therefore recommended. Yadin is still available on the website though renamed to Yadin-BeH in order to avoid conflict with the “Yadin” line lists for MgH and CaH, which remain to be recommended for the IR region. ### C$_2$, paper XXXI There are two new line lists for C$_2$ in the ExoMol2020 release: ExoMol 8state [@jt736] and the empirical MoLLIST [@13BrBeSc.C2]. The MoLLIST line list only considers the much-observed Swan band while 8state covers the 8 band systems which interconnect the 8 lowest electronic states of C$_2$. As 8state uses empirical MARVEL energy levels [@jt637] where available it should be as accurate as MoLLIST for the Swan band. Use of 8state is therefore recommended. Since MoLLIST line lists are recommended for other molecules and in order to avoid conflicts with the ‘recommended’ flag for 8states, the C$_2$ MoLLIST line list is now referenced to as MoLLIST-C2 on ExoMol. An update of the C$_2$ MARVEL data has just been completed [@jt809]; this has been used to improve the 8state energies and hence transition wavenumbers. This latest version includes uncertainties in the states file. ### CaH, Paper I The Yadin CaH line list only considers transitions within the X $^2\Sigma^+$ ground electronic state [@jt529]. MoLLIST provides a rovibronic line list for the $X\,^2\Sigma^+$–$X\,^2\Sigma^+$, $A\,^2\Pi$–$X\,^2\Sigma^+$, $B\,^2\Sigma^+$–$X\,^2\Sigma^+$ and E $^2\Pi$ – X $^2\Sigma^+$ systems due to @11RaTeGo.CaH [@12LiHaRa.CaH; @13ShRaBe.CaH]. Both line lists are included in the ExoMol database; work is in progress on creating a single unified line list which will extend the range of rovibronic transitions considered. In the meantime both line lists are recommended. ### CaO, Paper XIII No change. ### CS, Paper XII No change. We note that a line list for CS has also recently been supplied by @20HoWexx.CS and @20XiShSu.CS have extended consideration to rovibronic transitions for the lowest, A $^1\Pi$ – X $^1\Sigma^+$, allowed electronic band. ### KCl, Paper V No change ### MgH, Paper I The ExoMol Yadin  line list for MgH only considers transitions within the X $^2\Sigma^+$ ground electronic state [@jt529]. MoLLIST provides a rovibronic line list containing A $^2\Pi$ – X $^2\Sigma^+$ and B$^\prime$ $^2\Sigma^+$ – X $^2\Sigma^+$ transitions due to @13GhShBe.MgH. Both line lists are included in the ExoMol database; work is in progress on creating a single unified line list which will extend the range of rovibronic transitions considered. In the meantime both line lists are recommended. ### MgO, Paper XXXII There is a new LiTY line list for MgO [@jt759]. ### NaCl, Paper V No change. ### NaH, Paper X No change. ### NO, XXI There is a new NOName line list for NO [@jt686]. This line list was constructed using a combination of standard ExoMol and empirical methodologies, and also included a MARVEL project. Its transition frequencies should therefore be close to experimental accuracy. NOName has largely been adopted for the new release of HITEMP [@jt763]. NOName only covers transitions between levels which lie within the X $^2\Pi$ ground electronic state of NO. A new line list which includes ultraviolet rovibronic transitions of NO is currently being constructed. ### PH, Paper XXXIV There is a new LaTY line list for PH [@jt765]. ### PN, Paper VI No change. An extended line list covering visible and ultra violet (UV), accompanied by a MARVEL project is currently in progress. ### PO, Paper XXIII There is a new POPS line list for PO [@jt703]. ### PS, Paper XXIII There is a new POPS line list for PS [@jt703]. ### SH, Paper XXXVI There are two new ExoMol line lists for SH: SNaSH [@jt725] which only covers transitions within the X $^2\Pi$ ground electronic state of SH and the newer GYT line list of @jt776. GYT considers both transitions within X $^2\Pi$ and vibronic transitions in the A $^2\Sigma^+$ – X $^2\Pi$ band system. As GYT also improves the accuracy of the X state transitions, it is recommended for all applications. The SNaSH SH line list is also renamed to SNaSH-SH in order to avoid a conflict and retained only for completeness. ### SiO, Paper II The ExoMol SiO line lists only consider transitions within the X $^1\Sigma^+$ ground electronic state [@jt563]; these line lists have been widely used including a recent determination of isotopologue ratios in Arcturus [@jt777]. For this release of ExoMol, an empirical and less accurate SiO line list by @11Kurucz.db covering the X–X, A–X and E–X electronic bands was added. Given the importance of SiO for lava planets [@jt693], construction of a comprehensive (accurate and complete) rovibronic line list for SiO covering both IR and ultraviolet would be useful. ### SiH, Paper XXIV There is a new SiGHTLY line list for SiH [@jt711]. ### SiS, Paper XXV There is a new UCTY line list for SiS [@jt724]. ### SN, Paper XXVI There is a new SNaSH line list for SN [@jt725]. ### TiO, Paper XXXIII There is a new Toto line list for TiO [@jt760]. The line list for the main isotopologue ($^{48}$Ti$^{16}$O) used empirical energies from an associated MARVEL study [@jt672] while the energy levels of other isotopologues were improved using the procedure of @jt665. The Toto line list has been updated to give MARVEL uncertainties, where available, in the States file. For other states the uncertainties were estimated as follows: 1 [cm$^{-1}$]{} for all levels in vibronic states that contain levels determined by MARVEL, 10 [cm$^{-1}$]{} for levels in those electronic states who excitation energy ($T_e/T_0$) is known from experiment and 100 [cm$^{-1}$]{} for all other levels. TiO is a particularly important species in the atmosphere of cool stars and has been detected in exoplanets [@17SeBoMa.TiO; @17NuKaHa.TiO]. The Toto line list is significantly better at reproducing stellar spectra than the line lists due to @98Plxxxx.TiO and @98Scxxxx.TiO which it supersedes. Furthermore a recent study of Ti isotope abundances in two M-dwarf stars demonstrated the accuracy of the line lists for several isotopologues of TiO [@jt799]. However, TiO remains a challenging system to treat using [*ab initio*]{} electronic structure methods (see @jt623) and further work is required before there is a definitive line list for TiO. ### VO, Paper XVIII There is a new VOMYT line list for VO [@jt644] which has been used to tentatively identify VO in the atmosphere of exoplanets [@16EvSiWa.VO; @jt699]. We note that VO spectra display particularly large splittings due to hyperfine effects as well as transitions which are only allowed due to these effects [@82ChHaMe.VO]. VOMYT does not include hyperfine effects in its spectroscopic model meaning that the line list unsuitable for high resolution studies. We plan to address this problem by both performing a hyperfine resolved MARVEL study and extending [Duo]{} to allow for hyperfine effects. Triatomics ---------- ### CO$_2$, Paper XXXIX Line lists for hot CO$_2$ have been constructed by NASA Ames [@13HuFrTa.CO2; @13HuFrTa.CO2] using methodologies similar to those employed by ExoMol. The CDSD-4000 hot line list due to @11TaPexx.CO2 and the room temperature line lists due to @jt625 [@jt667; @jt678] are also available on the ExoMol website. The CDSD-4000 line list is based on the use of an effective Hamiltonian which tends to miss contributions from unobserved hot bands. A new ExoMol line list UCL-4000 for has been produced [@jt804] using an accurate *ab initio* dipole moment surface (DMS) by @jt613 and empirical potential energy surface (PES) Ames-2016 by @17HuScFr.CO2; where possible computed energies have been replaced by empirical one derived from HITRAN. The UCL-4000 line list covers the wavenumber range 0–20000 [cm$^{-1}$]{} and should be applicable for temperatures up to 4000 K [@jt804]. It is recommended for use in high temperature applications. ### HCN/HNC, Paper III The combined HCN/HNC ExoMol line list of Barber [[*et al*]{}]{} [@jt570] made extensive use of empirical energy levels due to Mellau [@11Mexxxx.HCN; @11Mexxxx.HNC]. This has enabled the line list to be successfully used in high-resolution Doppler spectroscopy studies of exoplanets [@18HaMaCa.HCN; @19CaMaHa.HCN; @jt782]. ### H$_2$O, Papers XIX and XXX Water is one of the most widely studied molecules and its spectrum has been detected in a variety of exoplanets, in many cases as the only clearly identifiable molecule [@jt699; @jt739]. Good water line lists have been available for some time, in particular the Ames line list of Partridge and Schwenke [@97PaScxx.H2O] and the BT2 line list due to @jt378. While BT2 was significantly more complete than Ames, the Ames line list was more accurate, particularly at infrared wavelengths. A new line list, known as POKAZATEL [@jt734] has been generated by the ExoMol project. POKAZATEL is complete in the sense that it contains transitions between all bound rotation-vibration states in the molecule and thus is reliable for temperatures above 3000 K where the earlier line lists are not valid. POKAZATEL is also intrinsically more accurate than the Ames line list and, indeed, as many of its energy levels have been replaced by empirical energy levels generated using MARVEL [@jt539], key transition frequencies are actually reproduced to experimental accuracy. POKAZATEL should therefore be used in preference to the earlier line lists. It has become apparent that the rotation-vibration spectrum of water is responsible for weak but observable near-UV absorption in the Earth’s atmosphere [@jt645]. The POKAZATEL line list, which used the [*ab initio*]{} LTP2001S DMS of @jt509, appears to underestimate the strength of water absorption in the blue and near-UV. As a response to this Conway and co-workers have recently developed a more accurate global dipole moment surface [@jt744]. Through a significant number of comparisons against high quality experimental and theoretical sources of spectroscopic data [@jt775; @jt785], their results suggest there may be advantages of using their line lists in particular regions, primarily at short wavelengths. For this reason, a new room temperature H$_2$$^{16}$O line list called WAT\_UV296 [@jt803] has been computed. This line list is available on the website and should be used for room temperature studies of spectra with $\lambda < 0.5$ $\mathrm{\mu m}$. A high temperature companion to this line list will form the basis of the updated HITEMP database [@jtH2OHITEMP2020]. @jt665 provided the HotWat78 line lists for H$_2$$^{17}$O and H$_2$$^{18}$O. The energy levels of this line list were improved using MARVEL energies [@jt454; @jt482]. Furthermore, @jt665 developed a method which gives excellent isotopologue energy levels for states only observed for H$_2$$^{16}$O. Finally the VTT line list due to @jt469 remains the recommended one for HDO; a new, improved HDO line list is currently under construction [@jtHDO]. A new room temperature line list CKYKKY for containing electric quadrupole moments was computed using an *ab initio* quadrupole moment surface [@20CaKaYa.H2O] and accurate empirical PES by @jt714 with the variational program TROVE [@TROVE]. The energies were replaced by the IUPAC MARVEL values [@jt539] or HITRAN values [@jt691], if available. ### H$_2$S, Paper XVI There is a new ExoMol AYT2 line list for hydrogen sulphide due to @jt640. Since completion of this line list a MARVEL project has been performed for H$_2$S [@jt718]; the AYT2 line list is being updated to use these improved energy levels. ### H$_3^+$ The new ExoMol MiZATeP line list of @jt666 replaces that of @jt181 (NMT). The new line list uses empirical energy levels from the MARVEL study of @13FuSzFa.H3+. Astronomy makes wide use of H$_3^+$ line lists and is reliant on [*ab initio*]{} line intensities since no absolute line intensities have been measured in the laboratory, see @jt587. This astronomical work on H$_3^+$ has recently been reviewed by @jt800. The ST line list due to @jt478 plus the newly generated and line lists will be MARVELised and released in the near future. ### SO$_2$, Paper XVII The ExoAmes line list for SO$_2$ [@jt637] is unchanged since the ExoMol2016 release. However, a MARVEL set of energy levels for SO$_2$ are now available [@jt704] and will be used to update ExoAmes in the near future. Tetratomics ----------- ### HCCH, Paper XXXVII The new aCeTY line list for acetylene due to @jt780 has recently been released. This line list already incorporated the MARVEL energies of @jt705 and gives uncertainties in the energy levels as part of the States file. ### H$_2$CO, Paper VIII The AYTY formaldehyde line list [@jt597] is in the process of being updated with empirical energy levels produced by a recent MARVEL study [@jtH2COmarvel]; this will make the line list suitable for high resolution studies. ### H$_2$O$_2$, Paper XV There is a new ExoMol APTY line list for hydrogen peroxide [@jt620]. We note because of difficulties with assigning spectra, even at room temperature HITRAN is very incomplete for H$_2$O$_2$. APTY should give complete coverage at infrared wavelengths. Recently APTY was used to suggest the importance of as a greenhouse gas on oxidised Early Mars [@20ItHaTa.H2O2]. ### H$_3O^+$, Paper XL A line list for the hydronium ion, H$_3$O$^+$, has recently been constructed [@jt805] in response to a laboratory study by @jt793 which suggested both that hydronium is likely to be a dominant molecular ion in gaseous exoplanets and that it should be detectable by upcoming space missions. ### NH$_3$, Paper XXXV A new ExoMol line list for ammonia called CoYuTe [@jt771] has recently been completed. CoYuTe, which also uses MARVEL energy levels [@jt608], is both more accurate and more extensive than the BYTe line list [@jt500] it replaces. In particular, CoYuTe was found to provide a good model of the Jovian optical absorption spectrum due to ammonia [@jt745] although with a shift in positions of the main bands. More laboratory work or analysis of existing laboratory work on the visible spectrum of ammonia will be required to resolve this issue. An illustration of the ammonia absorption cross sections computed using CoYuTe is given in Fig. \[fig:NH3\]. ![Absorption cross sections generated using the ExoMol line list CoYuTe for spectrum of NH$_3$ at $T=300$ K and 2000 K, with the Doppler profile on a wavenumber grid of 1 [cm$^{-1}$]{}. The cross sections were computed using the ExoCross program, which is developed to work with two file line list structure adopted by ExoMol.[]{data-label="fig:NH3"}](NH3_3displays_300-1500K.jpg){width="90.00000%"} ### PH$_3$, Paper VII No change. @20SoSaRa.PH3 use the SAlTY PH$_3$ line list to assess the detectability of a possible PH$_3$ biosignature. A MARVEL project on the experimental data is under way, which will improve the quality of the line positions in SAlTY. ### SO$_3$, Paper XVII There is a new UYT2 line list for SO$_3$ [@jt641]. Pentatomics ----------- ### CH$_3$Cl, Paper XXIX There are new OYT line lists for methyl chloride [@jt733] covering both major isotopes of chlorine. ### CH$_4$, Paper IV The ExoMol YT10to10 line list [@jt564] has been extended to higher temperatures [@jt698] to give the YT34to10 line list. However, at present the TheoReTS line list of @14ReNiTy.CH4 is the most accurate methane line list available and is recommended for detailed studies. This line list forms the basis of the very recent update to HITEMP [@20HaGoRe.CH4]. The importance of methane and the huge number of lines required to accurately represent the opacity of hot methane [@jt572] has led to this system being the primary focus of studies aimed at compacting these lines into a more manageable form [@jt698; @15HaBeBa.CH4; @17ReNiTyi]. This issue is discussed in section \[sect:data\]. ### HNO$_3$, Paper XI No change. We note that, like H$_2$O$_2$, because of difficulties with assigning spectra even at room temperature, HITRAN is very incomplete for nitric acid. The AlJS line list gives complete coverage at infrared wavelengths. ### SiH$_4$, Paper XXII There is a new OY2T line list for silane [@jt701]. Larger molecules ---------------- ### C$_2$H$_4$, Paper XXVII There is a new MaYTY line list for ethylene [@jt729]. MoLLIST species --------------- A number of diatomic species have line lists provided by the MoLLIST website [@MOLLIST]. These have been incorporated into the ExoMol database, which now includes empirically-derived line lists for the following species [@jt790]: CH [@14MaPlVa.CH]; NH [@14BrBeWe.NH; @15BrBeWe.NH; @18FeBeHo.NH]; OH [@16BrBeWe.OH]; AlCl [@18YoBexx.AlF]; AlF [@18YoBexx.AlF]; OH$^+$ [@17HoBexx.OH+; @18HoBiBe.OH+]; CaF [@18HoBexx.CaF]; MgF [@17HoBexx.MgF]; KF and NaF [@16FrBeBr.NaF]; LiCl and LiF [@18BiBexx.LiF]; MgH [@13GhShBe.MgH]; TiH [@05BuDuBa.TiH]; CrH [@06ChMeRi.CrH]; FeH [@10WEReSe.FeH]; CP [@14RaBrWe.CP]; CN [@14BrRaWe.CN]; and CaH [@12LiHaRa.CaH; @13ShRaBe.CaH]. See Table \[tab:otherdata\] for further details. HITRAN ------ In general the line lists contained in the HITRAN database [@jt691] do not provide the temperature coverage needed for astrophysical applications. However, for a few diatomic species very extensive line lists have been constructed. The line lists presented by @11LiGoBe.HCl for HCl and @15LiGoRo.CO for CO are both valid over an extended temperature range and have been included in the ExoMol database. The HF line list from @13LiGoHa.HCl and @15CoHaxx.HF is also adopted directly from HITRAN. N$_2$ ----- The empirical WCCRMT nitrogen molecule line list of @18WeCaCr.N2 is somewhat unusual in that it only considers transitions between excited, triplet states of the molecule. However, N$_2$ is an important molecule with very limited spectral signature at long wavelengths and WCCRMT may prove useful in hot or irradiated, usually non-LTE environments. For the partition function of the ground electronic state data from TIPS [@jt692s] were used. HD, HD$^+$, HeH$^+$ ------------------- One and two electron diatomics are important primordial species and may well play a role elsewhere. Recently @19AmDiJo created new line lists using high level [*ab initio*]{} procedures. These line lists have been included in the database with the name ADSJAAM. In particular the ADSJAAM line lists for HeH$^+$ and HD supersedes those due to @jt347 and @jt506, respectively. H$_2$ ----- A new external line list RACPPK by @19RoAbCz.H2 for the ground electronic state of H$_2$ was converted to the ExoMol format. The line list contains a combination of electric quadrupole and magnetic dipole transitions computed from first principles. Other [*ab initio*]{} line lists -------------------------------- A number of other [*ab initio*]{} line lists are also available. In general [*ab initio*]{} line lists are only accurate for systems with very few electrons. The CLT line lists for LiH and LiH$^+$ due to @jt506 fall into the few-electron category and should be reliable. The other [*ab initio*]{} line lists provided, namely those for ScH [@jt599], CH$_3$F [@19OwYaKu.CH3F], AsH$_3$ [@jt751], P$_2$H$_2$ [@19OwYuxx.P2H2], PF$_3$ [@jt752], CH$_3$ [@19AdYaYu.CH3] and YO [@jt774] must be regarded as intrinsically less accurate than other line lists provided by the database. In the case of YO, work is in progress aimed at producing an empirical line list based on available experimental data. P$_2$H$_2$ occurs as two distinct isomers, the cis and trans forms. Separate line lists are provided for cis-P$_2$H$_2$ and trans-P$_2$H$_2$. SiH$_2$ ------- A new line list for SiH$_2$ [@jt779] has been recently constructed in its ground electronic state using the standard ExoMol technology [@jt626]. The set of experimental data or their quality were, however, very limited which may affect the quality of the hot bands, which basically has to rely on the quality of the [*ab initio*]{} PES. NiH --- Empirical line lists for three isotopologues of NiH ($^{58}$NiH, $^{60}$NiH, $^{62}$NiH) known as HRV. These line lists were constructed by @09VaRiCr.NiH and @13HaRiTo.NiH. However, at present these line lists are not available in ExoMol format or accessible through the API. Partition functions ------------------- The temperature dependent partition functions for most of the ExoMol line lists are computed using the corresponding energy levels as collected in the States files as $$Q(T) = \sum_{i} g_{\rm ns}^{(i)} J_i(J_i + 1) \exp\left(-\frac{c_2 \tilde{E}_i}{T}\right),$$ where $g_{\rm ns}^{(i)}$ is the state dependent nuclear statistical weight, $J_i$ is the total angular momentum of the state $i$, $c_2$ is the second radiation constant (Kcm) and $\tilde{E}_i$ is the corresponding energy term value ([cm$^{-1}$]{}). ExoMol uses the HITRAN convention [@06SiJaRo.method], where the entire, integer $g_i = g_{\rm ns}^{(i)} J_i(J_i + 1)$ factors are explicitly included. Some of the line lists in the ExoMol database do not provide sufficiently complete sets states for a proper evaluation of the molecular partition functions. For example, most of the MoLLIST line lists are constructed from or based on the measured data only and therefore can be severely incomplete. In these cases the partition functions are either taken from external sources, such as the TIPS database [@jt692s], generated using the empirical expansions, such as by @81Irwin [@84SaTaxx.partfunc; @88Irwin; @16BaCoxx.partfunc] or extrapolated using simplified models [@jt571]. The external partition functions often include the $g_{\rm ns}^{(i)}$ factors in the astrophysical convention (see, e.g. @jt777) and therefore have to be transformed to the HITRAN convention. We strongly recommend that users use these partition functions rather than attempting to compute their own using the ExoMol States files. These States files are not constructed with a view to delivering reliable partition functions and in a number of cases use of them has been found to lead to problems. VUV sections ------------ A section is provided based on measured ultraviolet cross sections for key species. For many molecules their ultraviolet spectrum is a mixture of line and quasi-continuum absorption which cannot be represented as a line list. The new VUV section will provide temperature-dependent absorption cross sections. Recent results suggest the VUV absorption by H$_{2}$O enhances the production of OH, which plays an important molecule in atmospheric chemistry of exoplanetary atmospheres [@20RaScHa.H2O]. Currently, the VUV cross sections are provided for H$_2$O, H$_2$, CO$_2$, SO$_2$, NH$_3$, H$_2$CO and C$_2$H$_4$ measured by Fateev’s lab at the Danish Technical University (DTU) [@20FaClYi] and for by @18VebeFa.CO2. The temperature and wavelength coverage is illustrated in Table \[t:VUV\]. The cross sections are given in the common two-column ASCII format separated by spaces, where the first column contains the wavelength in nm and the second column contains the absorption cross sections in cm$^{2}/$molecule using Fortran format: `(F10.3,1x,E13.6)`. The VUV cross section file names have the following structure\ `<ISOTOPOLOGUE>__<DATASET>__<RANGE>__T<TEMP>K__P<PRESSURE>bar__<STEP>.nm`, where `ISOTOPOLOGUE` is the isotopologue name, `DATASET` is the name of the line list, `RANGE` is the wavelength range in nm, `TEMP` is the temperature in K, `PRESSURE` is the pressure in bar, `STEP` is the wavelength step in nm. For example, the States file of the VUV line list for the filename:\ [12C-12O2\_\_Venot-2018\_\_116.90-230.00\_\_T0800K\_\_P0bar\_\_0.03.nm](12C-12O2__Venot-2018__116.90-230.00__T0800K__P0bar__0.03.nm). [lrr]{} Molecule & Temperature (K) & Range (nm)\ \ & 423 & 110–230\ & 573 & 110–230\ & 1630 & 182–237\ & 1773 & 182–237\ & 1160 & 109–324\ & 423 & 110–230\ & 289 & 113–201\ & 296 & 113–201\ & 303 & 110–230\ & 353 & 110–230\ & 423 & 110–230\ & 573 & 110–230\ & 562 & 113–201\ \ & 170–800 K & 115-220\ Data Provided ============= Table \[tab:datsum\] provides a summary of different types of data provided. The website provides two routes to accessing these data. Users can search by molecule which will show all the types of data available for each isotopologues. Alternatively it is possible to search by data type in which case a list of molecules for which data of the specified type is available. The following section lists the data types and their file name extensions for the various data and metadata provided by ExoMol. A more technical specification of the data structures and how to access data using the application program interface (API) is provided in Section \[sect:data-formats-and-api\]. Data type ------------------------------------------------ -- Line lists Absorption cross sections VUV absorption cross sections Pressure broadening coefficients Temperature dependent super-lines (histograms) Partition functions Cooling functions Specific heat - heat capacity Examples of ExoCross input files Temperature and pressure dependent opacities Spectroscopic Models : Summary of data provided by the ExoMol Database[]{data-label="tab:datsum"} Data Structure -------------- File extension $N_{\rm files}$ File DSname Contents ------------------------ ----------------- -------------------- ------------------------------------------------------------------------------------- `.all` 1 Master Single file defining contents of the ExoMol database.. `.def` $N_{\rm tot}$ Definition Defines contents of other files for each isotopologue. `.states` $N_{\rm tot}$ States Energy levels, quantum numbers, lifetimes, (Landé $g$-factors, Uncertainties). `.trans` $^a$ Transitions Einstein $A$ coefficients, (wavenumber). `.broad` $N_{\rm mol}$ Broadening Parameters for pressure-dependent line profiles. `.cross` $^b$ Cross sections Temperature or temperature and pressure-dependent cross sections. `.kcoef` $^c$ $k$-coefficients Temperature and pressure-dependent $k$-coefficients. `.pf` $N_{\rm tot}$ Partition function Temperature-dependent partition function, (cooling function). `.dipoles` $N_{\rm tot}$ Dipoles Transition dipoles including phases. `.super` $^d$ Super-lines Temperature dependent super-lines (hisograms) on a wavenumber grid. `.nm` $^e$ VUV cross sections Temperature and pressure dependent VUV cross-sections (wavelength, nm). `.fits`, `.h5`, `.kta` $^f$ Opacities Temperature and pressure dependent opacitities for radiative-transfer applications. `.overview` $N_{\rm mol}$ Overview Overview of datasets available. `.readme` $N_{\rm iso}$ Readme Specifies data formats. `.model` $N_{\rm iso}$ Model Model specification. : Specification of the ExoMol file types. (Contents in brackets are optional.)[]{data-label="tab:files"} $N_{\rm files}$ total number of possible files;\ $N_{\rm mol}$ Number of molecules in the database;\ $N_{\rm tot}$ is the sum of $N_{\rm iso}$ for the $N_{\rm mol}$ molecules in the database;\ $N_{\rm iso}$ Number of isotopologues considered for the given molecule.\ $^a$ There are $N_{\rm tot}$ sets of `.trans` files but for molecules with large numbers of transitions the `.trans` files are subdivided into wavenumber regions.\ $^b$ There are $N_{\rm cross}$ sets of `.cross` files for isotopologue.\ $^c$ There are $N_{\rm kcoef}$ sets of `.kcoef` files for each isotopoplogue.\ $^d$ There are $N_T$ sets of $T$-dependent super-lines. $^e$ There are $N_{VUV}$ sets of VUV cross sections. $^f$ Set of opacity files in in the format native to specific radiative-transfer programs. The general ExoMol approach is molecule-by-molecule: a comprehensive line list is created for a particular molecule and made available in the database. The line list for each isotopologue is stored as a separate data structure which can be accessed directly or via the application program interface (API) described in the following sections. Table \[tab:files\] specifies the file types that can be available for each isotopologue. The `.states` and `.trans` files are the heart of the ExoMol data structure [@jt548] and define what has become known as the ExoMol format. These files are available for all isotopologues; other files may not be. ExoMol format files can be used and provided by the effective Hamiltonian code PGOPHER [@PGOPHER]. A manual providing technical specifications of the database is included as supplementary data to this article and is included on the website where it will be updated as the data base evolves. Only the relatively few changes that have been implemented since the 2016 release are discussed below. Super-lines ----------- Super-lines [@TheoReTS; @jt698] represent a novel, compact way of storing the opacity data. Super-lines are constructed as temperature-dependent intensity histograms by summing all absorption coefficients within a wavenumber bin centred around a grid point $\tilde\nu_k$. For each $\tilde\nu_k$ the total absorption intensity ${I}_k(T)$ is computed as a sum of absorption line intensities $I_{if}$ from all $i\to f$ transitions falling into the wavenumber bin $[\tilde\nu_k-\Delta \tilde{\nu}_k/2 \ldots \tilde\nu_k+\Delta \tilde{\nu}_k/2 ] $ at the given temperature $T$. Each grid point $\tilde\nu_k$ is then treated as a line position of an artificial transition (super-line) with an effective absorption intensity ${I}_k(T)$. The number of data points in the super-lines can be drastically reduced without significant loss of accuracy when computing pressure-dependent cross sections and can be combined with standard, pressure, temperature and frequency dependent line profiles. The super-lines cannot be associated with any specific upper/lower states and therefore the line broadening parameters used cannot depend on quantum numbers. The super-line list files have the format of cross sections represented by two columns, wavenumbers ([cm$^{-1}$]{}) and super-line absorption intensity (cm$/$molecule). The Fortran format is `(F12.6,1x,ES14.8)`. The super-lines are computed on a grid of temperatures from 100 K to $T_{\rm max}$. For the resolution an adaptive grid of $R= 1\,000\,000$ is used. For the technical details of super-line construction and numerical tests see @jt698. Super-lines are currently available for , , , and . Specific heat ------------- The specific heat at constant pressure $C_p$ is computed on a grid of temperatures of 1 K from the molecule partition function as given by [@jt661] $$C_p(T) = R \left[\frac{Q''}{Q}-\left(\frac{Q'}{Q}\right)^2 \right] + \frac{5R}{2},$$ where the second term is the translational contribution, $$Q'(T) = T\frac{d Q}{d T},$$ $$Q''(T) = T^2\frac{d^2 Q}{d T^2} + 2Q'$$ and $R$ is the gas constant. The specific heat values are given in units of J mol$^{-1}$ K$^{-1}$ and are currently provided for (taken from @jt661) and (computed using the YT10to10 ro-vibrational energies [@jt564]) only (see Fig. \[fig:Cp\]). ![Specific heat at constant pressure of H$_2$O as computed by @jt661 and CH$_4$ generated the YT10to10 ro-vibrational energies [@jt564].[]{data-label="fig:Cp"}](H2O_Cp.jpg){width="75.00000%"} Line shifts using the diet format --------------------------------- For high resolution applications it will be important to take into account the pressure dependent line shifts. To this end a line shift diet has been introduced using the same data structure used for pressure dependent line broadening diet [@jt684]. An example for is given in Table \[t:shift\]. |1H-16O2\_\_a0.shift| Code Shift J ------ -------- --- a0 0.0001 0 a0 0.0001 1 a0 0.0001 2 a0 0.0001 3 a0 0.0001 4 … … … : File A pressure line `.shift` file for H$_2$O: portion of the file (upper part); field specification (lower part).[]{data-label="t:shift"} Field Fortran Format C Format Description ------------------------- ---------------- ---------- ---------------------------------------------------------------------- code A2 $\%$2s Code identifying quantum number set following $J{^{\prime\prime}}$\* $\gamma_{\textrm{ref}}$ F15.6 %15.6f Line shift at reference temperature and pressure in [cm$^{-1}$]{}  $J{^{\prime\prime}}$ I7/F7.1 $\%$7d Lower $J$-quantum number : File A pressure line `.shift` file for H$_2$O: portion of the file (upper part); field specification (lower part).[]{data-label="t:shift"} Code definition: a0 = none New broadening parameters ------------------------- The pressure broadening parameters are provided for 15 molecules and are listed in Table \[t:pressure:diet\] using the pressure-broadening diet [@jt684]. The `.broad` file has a hierarchical structure; each record starts with four compulsory columns: a label defining the broadening scheme (‘a0’, ‘a1’, …), values of $\gamma_{\textrm{ref}}$, $n$ and $J''$. The compulsory fields are followed by additional quantum numbers when a more detailed specification of the quantum assignments is available. The most basic scheme ‘a0’ represents broadening with $J{^{\prime\prime}}$ dependence only. The additional basic scheme ‘a1’ is used for the case with the $(J',J'')$ dependence. Any other schemes (e.g. ‘a2’, ‘a3’, ‘a4’, ‘b1’, ‘b2’, ‘c1’, ‘c2’) are the molecule specific and should be described as part of the line list specifications. The broadening data format is illustrated in Table \[tab:broad:a0:a1:a2\]. The requirement for enhanced line broadening parameters was identified by @jt773 is their study on the need for laboratory data requirements for studies of exoplanetary atmospheres. Recognising this, our plan is significantly enhance the treatment pressure effects in future releases of the ExoMol database. Molecule Broadener Scheme ---------- ------------------ -------- CS air, self a0 HCl , He, self, air, a0 HF , He a0 CO , He a0 NO air a0 , He, self, air a0,a1 air, self a0 , He a0, a5 HCN , He a0 OCS , He a0 , He a0, c1 , He a0 , He a0 , He a0 , He, a0, a1 : Pressure broadening parameters in ExoMol[]{data-label="t:pressure:diet"} Label $\gamma$ $n$ $J''$ $J'$ $K$ ------- ---------- ------- ------- ------ ----- a0 0.0860 0.096 0 a0 0.0850 0.093 1 ... a1 0.0860 0.096 0 1 a1 0.0850 0.093 1 2 ... : Example of the three basic broadening schemes in the ExoMol diet, ‘a0’, ‘a1’, ‘a2’.[]{data-label="tab:broad:a0:a1:a2"} ExoMolOP: Opacities ------------------- Recently @jt801 computed opacity cross sections and $k$-tables for all molecules available from the ExoMol database plus some atomic data from NIST. These data are formatted for use in various retrieval codes including Tau-REx [@jt593; @jt611; @TauRexIII], ARCiS [@19MiOrCh.arcis], petitRADTRANS [@19MoWaBo.petitRADTRANS], and NEMESIS [@NEMESIS]. Data were calculated on temperature-pressure grids suitable for characterising a variety of exoplanet and stellar atmospheres. Broadening parameters were taken from the literature where available, with broadening parameters used for a known molecule with a similar dipole moment, where they are not available. Tables of cross sections and $k$-coefficients are provided on ExoMol as part of the line list webpage. The format of the data and the resolution (grid spacing) is application dependent, for example the opacities for ARCiS, petitRADTRANS, and NEMESIS are given as $k$-tables at the resolving power of $R$= 1000 while Tau-REx works with cross-sections with $R = 15\,000$. The opacity cross sections and $k$-tables will also be made available via the virtual atomic and molecular data centre (VAMDC) portal [@jt481; @jt630]. We note that ExoMol data has also been extensively used to construct the EXOPLINES molecular absorption cross-Section database for brown dwarf and giant exoplanet atmospheres [@exoplines]. Post-processing {#s:exocross} --------------- ExoMol provides post-processing capabilities through the program ExoCross [@jt708]. ExoCross has many functions such as generating pressure and temperature dependent cross sections, partition functions, specific heat, state-resolved radiative lifetimes, non-LTE spectra, electric dipole, electric quadrupole and magnetic dipole spectra. ExoCross can read data in both ExoMol and HITRAN [@jt557] formats and output it in these formats as well as SPECTRA (<http://spectra.iao.ru/>) and Phoenix formats [@PHOENIX]. Should data be needed in say HITRAN format, it is strongly recommended that the data is downloaded to a local computer in the much more compact ExoMol format and then processed using ExoCross. Examples of ExoCross input files are provided on the website. There are also a number of Python utilities available on the ExoMol website. These are now largely redundant as ExoCross provides all functionality required to work with ExoMol line lists. However, we mention them for completeness: Utility `extract__trans.py` reads the `.trans` in `.bz2` format without requiring it to be uncompressed. The `ExoMol_to_HITRAN.py` script converts ExoMol format to HITRAN format; this should be used with caution as ExoMol format is significantly more compact than HITRAN format and the line list files are large. The program `exomol2gf.py` can be use to generate oscillator strengths. New web services ---------------- New web services include: - A molecule and line list search. - A graphical illustration of a line list represented by absorption spectra at two or three temperatures (e.g. 300 K, 2000 K and 5000 K) computed using a Doppler line profile on a grid of 1 [cm$^{-1}$]{}. - The ExoMol bibliography in the BibTeX format is stored at and version controlled by GitHub at <https://github.com/ExoMol/bib>. ExoMol data formats and API {#sect:data-formats-and-api} =========================== Format of the `states` and `trans` files ---------------------------------------- The formats of the States and Transitions files are specified in Table \[t:states:format\] and \[t:trans:format\]. Tables \[t:states\] and \[t:trans\] show typical examples of the States (.states) and Transitions (.trans) files. A significant new feature of the 2020 update is the uncertainty field ([cm$^{-1}$]{}) in the States file appearing as column 5 after the $J$ values; this currently an optional feature but will become a compulsory column for the ExoMol States file. The uncertainty values typically come from three different sources: (i) the uncertainties of the MARVEL energies ($\sim$0.000001–0.1 [cm$^{-1}$]{}), (ii) uncertainties obtained as the fitting observed – calculated error of energies as part of the refinement to experimental data (energies or line positions) participated in the refinement ($\sim$ 0.001 – 1 [cm$^{-1}$]{}) and (iii) roughly estimated uncertainties for all other states that cannot be verified against existing experiment ($\sim$ 0.1 – 10 [cm$^{-1}$]{}). The uncertainty in the States file are given using the same format as the energy term values, i.e. with six decimal places after the decimal point, see Table \[t:states:format\]. To date the new standard of the States file has being applied to the only a number of key/ recent line lists namely (POKAZATEL) [@jt734], AlH (AlHambra) [@jt732], C$_2$ [8states]{} [@jt736; @jt809], HCCH (aCeTY) [@jt780], (UCL-4000) [@jt804], (eXeL) [@jt805] and TiO (Toto) [@jt760]. A rolling programme is in place for updating the other files so the uncertainty is uniformly available for all sources. Field Fortran Format C Format Description ------------------ ---------------- ------------- --------------------------------------------------------------- $i$ `I12` `%12d` State ID $E$ `F12.6` `%12.6f` State energy in $\mathrm{cm^{-1}}$ $g_\mathrm{tot}$ `I6` `%6d` State degeneracy $J$ `I7/F7.1` `%7d/%7.1f` $J$-quantum number (integer/half-integer) ($\Delta E$) `F12.6` `%12.6f` Energy uncertainty in $\mathrm{cm^{-1}}$ (currently optional) ($\tau$) `ES12.4` `%12.4e` Lifetime in s (optional) ($g$) `F10.6` `%10.6f` Landé $g$-factor (optional) (Extra) - - Extra quantum numbers, any format (optional) : Updated specification of the ExoMol States file.[]{data-label="t:states:format"} Field Fortran Format C Format Description -------------------- ---------------- ---------- ------------------------------------------------ $i$ `I12` `%12d` Upper state ID $f$ `I12` `%12d` Lower state ID $A$ `ES10.4` `%10.4e` Einstein $A$ coefficient in $\mathrm{s^{-1}}$ $\tilde{\nu}_{fi}$ `E15.6` `%15.6e` Transition wavenumber in cm$^{-1}$ (optional). : Specification of the Transitions file.[]{data-label="t:trans:format"} $i$ $g_{\rm tot}$ $J$ $\Gamma$ $e/f$ $n_1$ $n_2^{\rm lin}$ $l_2$ $n_3$ ----- ------------- --------------- ----- ---------- ---------- ------- ------- ----------------- ------- ------- 1 0.000000 1 0 0.000500 A1 e 0 0 0 0 2 1285.408200 1 0 0.000500 A1 e 0 2 0 0 3 1388.184200 1 0 0.005000 A1 e 1 0 0 0 4 2548.366700 1 0 0.000500 A1 e 1 2 0 0 5 2671.142957 1 0 0.005000 A1 e 2 0 0 0 6 2797.136000 1 0 0.005000 A1 e 1 2 0 0 7 3792.681898 1 0 0.005000 A1 e 1 4 0 0 8 3942.541358 1 0 0.005000 A1 e 3 0 0 0 9 4064.274256 1 0 0.005000 A1 e 3 0 0 0 10 4225.096148 1 0 0.005000 A1 e 1 4 0 0 11 4673.325200 1 0 0.000500 A1 e 0 0 0 2 12 5022.349428 1 0 0.005000 A1 e 1 6 0 0 13 5197.252900 1 0 0.005000 A1 e 3 2 0 0 14 5329.645446 1 0 0.005000 A1 e 4 0 0 0 15 5475.553054 1 0 0.000500 A1 e 3 2 0 0 16 5667.644584 1 0 0.005000 A1 e 2 4 0 0 17 5915.212302 1 0 0.000500 A1 e 0 2 0 2 : An example of an extract from the final States file for UCL-4000 of [@jt804]. []{data-label="t:states"} \ [$i$: State counting number.\ $\tilde{E}$: State energy in [cm$^{-1}$]{}.\ $g_{\rm tot}$: Total state degeneracy.\ $J$: Total angular momentum.\ unc.: Uncertainty in [cm$^{-1}$]{}.]{}\ $\Gamma$: Total symmetry index in [${\mathcal C}_{2{\rm v}}$]{}(M).\ $e/f$: Kronig rotationless parity.\ $n_1$: Normal mode stretching symmetry ($A_1$) quantum number.\ $n_2^{\rm lin}$ : Normal mode linear molecule bending ($A_1$) quantum number.\ $l_2$: Normal mode vibrational angular momentum quantum number.\ $n_3$: Normal mode stretching asymmetric ($B_1$) quantum number. $f$ $i$ $A_{fi}$ $\tilde{\nu}_{fi}$ ------- ------- ------------ -------------------- 10571 10884 9.5518E-06 120.241863 21053 21375 1.9515E-05 120.242886 8726 9672 1.8658E-04 120.243522 11655 11950 5.0065E-06 120.243733 93209 93967 5.7055E-03 120.244192 2228 3175 7.3226E-07 120.244564 46727 46432 1.0599E-04 120.244658 44436 44774 1.4626E-04 120.245583 29037 28723 1.8052E-04 120.245669 4458 4805 1.0431E-08 120.246396 69313 68434 5.0531E-06 120.248178 22640 22985 1.1281E-07 120.248891 57027 56721 7.1064E-06 120.250180 : Extract from the Transitions file for CaO.[]{data-label="t:trans"} $f$: Upper state counting number; $i$: Lower state counting number; $A_{fi}$: Einstein-A coefficient in s$^{-1}$; $\tilde{\nu}_{fi}$: transition wavenumber in [cm$^{-1}$]{}. Formats for other data types ---------------------------- Table \[tab:broad\_format\] shows the format of the `.broad` files containing the line broadening parameters. Table \[tab:cross\] shows the format of the cross section `.cross` and super-lines `.super` files. Tables \[tab:pf\] and \[tab:cp\] show the format of the partition function `.pf` and specific heat `.cp` files. Field Fortran Format C Format Description ------------------------- ---------------- ----------- ---------------------------------------------------------------------------------- code A2 %2s Code identifying quantum number set following $J{^{\prime\prime}}$ $\gamma_{\textrm{ref}}$ F6.4 %6.4f Lorentzian half-width at reference temperature and pressure in [cm$^{-1}$]{}/bar $n$ F6.3 %6.3f Temperature exponent $J{^{\prime\prime}}$ I7/F7.1 %7d/%7.1f Lower $J$-quantum number integer/half-integer : Specification of the mandatory part of the pressure broadening parameters file.[]{data-label="tab:broad_format"} Fortran format, $J$ integer: `(A2,1x,F6.4,1x,F6.3,1x,I7)`\ or $J$ half-integer: `(A2,1x,F6.4,1x,F6.3,1x,F7.1)`\ Field Fortran Format C Format Description ----------------- ---------------- ---------- ------------------------------------------------------- $\tilde{\nu}_i$ `F12.6` `%12.6f` Central bin wavenumber, $\mathrm{cm^{-1}}$ $\sigma_i$ `ES14.8` `%14.8e` Absorption cross section, $\mathrm{cm^2\,molec^{-1}}$ $\alpha_i$ `ES14.8` `%14.8e` Absorption coefficient, $\mathrm{cm\,molec^{-1}}$ : Specification of the `.cross` cross section and `.stick` file format[]{data-label="tab:cross"} Fortran format: `(F12.6,1x,ES14.8)`\ Field Fortran Format C Format Description -------- ---------------- ---------- ------------------------------------------------------------------- $T$ `F8.1` `%8.1d` Temperature in K $Q(T)$ `F15.4` `%15.4d` Partition function (dimensionless). $W(T)$ `ES12.4` `%12.4e` Cooling function in ergs s$^{-1}$ molecule$^{-1}$ (if available). : Specification of the `.pf` partition function file.[]{data-label="tab:pf"} Fortran format: `(F8.1,1x,F15.4,1x,ES12.4)` or `(F8.1,1x,F15.4)`\ Field Fortran Format C Format Description ---------- ---------------- ---------- ----------------------------------------- $T$ `F8.1` `%8.1d` Temperature in K $C_p(T)$ `F15.4` `%15.4d` Specific heat function (dimensionless). : Specification of the `.cp` specific heat file.[]{data-label="tab:cp"} Fortran format: `(F8.1,1x,F15.4)`\ API {#sect:api} --- ### Searching for data through the API {#sect:api:1} The molecules and isotopologues available in ExoMol are listed in a master file located at: [www.exomol.com/exomol.all](www.exomol.com/exomol.all). Given a molecule or list of isotopologues, ExoMol can be searched for recommended datasets using the API which can be queried using the `HTTP GET` request method described below. The structure of a JSON file is illustrated in Fig. \[fig:API:search\]. To search for all data sets (and their files) related to a single molecule, use the query: http://exomol.com/api?molecule=<MOLECULE> For example, the URL <http://exomol.com/api?molecule=NH3> returns a JSON data structure describing all files related to the ammonia molecule. Within this, the list of files belonging to the BYTe-15 line list for $\mathrm{^{15}N^{1}H_3}$ is accessible by traversing the JSON object with: json_result['(15N)(1H)3']['linelist']['BYTe-15']['files'] and the URL for the States file belonging to this line list is at: jsonresult['(15N)(1H)3']['linelist']['BYTe-15']['files'][1]['url']. Note that molecules are identified by simple text strings (no subscript or superscript symbols) but that special characters must be URL-encoded: for example, the $\mathrm{HeH^+}$ cation is identified by `molecule=HeH%2B`. Most software will provide this “percent-encoding” automatically. Searches can be further refined by isotopologue by setting the query keyword `isotopologues` to a comma-delimited sequence; for example: http://exomol.com/api/?isotopologues=(12C)(32S),(12C)(34S) Restricting the search by datatype is also supported: http://exomol.com/api/?molecule=H2O&datatype=linelist returns the details of linelists for the isotopologues of water, omitting partition functions, cooling functions, opacities, etc. Valid values for the `datatype` parameter are: `linelist`, `energylevels`, `opacity`, `super` (“Super” line lists), `Cp` (heat capacity), `broadening_coefficients`, `coolingfunction`, `partitionfunction`. ### Accessing the data via the API {#sect:data} There are a number of ways of accessing the data. First, all data sets are available on the ExoMol website ([www.exomol.com](www.exomol.com)) and can be downloaded manually. Many ExoMol line lists contain in excess of 10 billion transitions. In this case the file of transitions are generally split into chunks in frequency. These files are compressed using `.bz2` format. However, if multiple datasets are required manual downloads are inefficient. The systematic structure of the ExoMol database filesystem allows for automated downloads from within software or using utilities such as `wget` and `curl`. As described in detail in the ExoMol2016 release, the API structures the ExoMol data resource files as URLs of the form: http://exomol.com/db/<MOLECULE>/<ISOTOPOLOGE>/<DATASET>/<FILENAME> where `<FILENAME>` is structured `<ISOTOPOLOGUE>__<DATASET>.<DATATYPE>`. For example, the States file of the YYLT line list for $\mathrm{^{31}P^{15}N}$ is obtainable at the URL <http://exomol.com/db/PN/31P-15N/YYLT/31P-15N__YYLT.states.bz2>. Furthermore, each dataset has a *manifest* that lists the data files it is comprised of and their sizes (in bytes). The manifest is a single text file located at the above URL with the filename `<ISOTOPOLOGUE>__<DATASET>.manifest`. For example, the contents of <http://exomol.com/db/PN/31P-15N/YYLT/31P-15N__YYLT.manifest> are: http://exomol.com/db/PN/31P-15N/YYLT/31P-15N__YYLT.states.bz2 142161 http://exomol.com/db/PN/31P-15N/YYLT/31P-15N__YYLT.trans.bz2 1803083 http://exomol.com/db/PN/31P-15N/YYLT/31P-15N__YYLT.pf 175000 http://exomol.com/db/PN/31P-15N/YYLT/31P-15N__YYLT.def 4761 http://exomol.com/db/PN/31P-15N/YYLT/31P-15N__YYLT.manifest 323 Moving to higher spectral resolution ==================================== Molecule $N_{\rm iso}$ $N_{\rm elec}$ $N_{\rm levels}$ Reference(s) ---------- --------------- ---------------- ------------------ --------------------------------- H$_2$O 7 1 18 486 @jt454 [@jt482; @jt539; @jt576]   update 1 1 19 200 @jt750 [@jt795] H$_3^+$ 3 1 652 @13FuSzFa.H3+ [@13FuSzMa.H3+] NH$_3$ 1 1 4951 @jt608   update 1 1 4936 @jt784 C$_2$ 1 14 5699 @jt637   update 1 20 7087 @jt809 TiO 1 9 10 564 @jt672 [@jt760] HCCH 1 1 11 213 @jt705 SO$_2$ 3 1 15 130 @jt704 H$_2$S 1 1 11 213 @jt718 ZrO 1 10 8088 @jt740 O$_2$ 1 6 4279 @19FuHoKoSo NH 1 4 1058 @jt705 CaOH 1 5 1954 @jt791 H$_2$CO 1 1 4841 @jtH2COmarvel NO 1 1 4106 @jt686 AlH 2 2 331 @jt732 BeH 3 2 1264 @jt722 CN 1 10 7779 @20SyMc : Molecules of importance for the ExoMol project with published MARVEL datasets.[]{data-label="tab:marveldata"} $N_{\rm iso}$ Number of isotopologues considered;\ $N_{\rm elec}$ Number of electronic states considered;\ $N_{\rm levels}$ Number of energy levels extracted: value is for the main isotopologue.\ Transit spectroscopy of exoplanets has thus far been performed at rather low resolution; however, very precise spectroscopic data are required for high resolution Doppler spectroscopy [@13DeBrSn.exo; @13BiDeBr.exo; @14BrDeBi.exo]. This has proved to be an issue for important species [@15HoDeSn.TiO]. Indeed the ExoMol datasets described above were generally constructed with greater emphasis on completeness than obtaining very precise transition frequencies [@jt716]. In practice, however, the use of empirical energy levels in the States file means that some transition frequencies are indeed reproduced with high (experimental) accuracy; however, with the ExoMol2016 data structure it was not possible to tell how accurately a particular transition was predicted. Although some work has been done on using ExoMol data to provide the input for high resolution studies [@jt782], there is a clear need to adapt the database to provide the laboratory data needed for these studies. For this reason we have updated the data structure to allow the uncertainty in a particular transition frequency to be determined. This is done via uncertainties in the energy levels which are now specified (optionally) in the States file, see Table \[t:states:format\]. To improve the accuracy of the predicted spectra, the ExoMol States files are being systematically updated using empirically-determined energy levels. The main means of doing this is via the MARVEL (measured active rotational-vibrational energy levels) procedure [@jt412; @12FuCsi.method; @jt750]. MARVEL inverts available high resolution spectra for a given isotopologue to give a list of empirical energy levels with associated uncertainties. For levels determined by MARVEL these uncertainties are now given in the States file. Otherwise the (usually much larger) uncertainty arising from the calculation used to generate the line list is given. By combining the uncertainty of upper and lower states ($\Delta E_u$, $\Delta E_l$) using the standard formula $$\Delta E_\nu = \frac{1}{\sqrt{2}}\sqrt{\left((\Delta E_u)^2 + (\Delta E_l)^2\right)}$$ gives the uncertainty in the transition wavenumber, $\Delta E_\nu$. To help improve the accuracy of key line lists we have been running MARVEL projects on relevant molecules. Table \[tab:marveldata\] lists the astronomically important molecules for which MARVEL studies have been completed. We note that a number of these studies [@jt672; @jt705; @jt718; @jt740; @jt764] have been performed as part of the so-called ORBYTS schools outreach project, see @jt709 for a discussion of this. We are in the process of working through all the molecules in the ExoMol database running MARVEL projects for those isotopologues for which there are enough high accuracy laboratory data available to justify this activity. Other Future development ======================== The ExoMol project already maintains extensive molecule by molecule bibliography files. These are stored on LaTeX’s BibTex format and are freely accessible at <https://github.com/ExoMol/bib/tree/master/exomol>. Our aim is to make referencing to the original data as easy and convenient as possible, including automatic generation of the list of references to cite in the appropriate format such as BibTex, Endnote etc. Such services have been developed and offered by HITRAN [@20SkGoHi]. We are in the process of moving the database to a more powerful platform. After the move we plan to offer the ability to compute cross sections for a given species, temperature and pressure on the fly. This will be done using super-lines which greatly reduce the computing time. We will also facilitate the computation of $k$-tables for a given atmospheric model. Finally the move should allow much greater integration with the code Tau-REx [@jt593; @jt611]. Tau-REx is an open source retrieval code for exoplanetary atmospheres which has just undergone a major upgrade [@TauRexIII]. Integration with the Tau-REx-III upgrade via a graphical interface is currently in progress. Conclusions =========== The ExoMol database presented here is a molecule-by-molecule set of comprehensive line lists for modelling spectra and other properties of hot gases. The choice of molecules is dictated by the need to model the atmospheres of exoplanets and other hot astronomical objects, but the spectroscopic data have much wider applications than this. We are still in the process of adding molecules to the database and are receptive to suggestions of other key species to include. In addition we are working on improving the accuracy of the line positions, particularly for strong lines, enhancing the treatment of pressure broadening, and extending the range of wavelengths covered into the ultra-violet for molecules where this is considered important. In addition, we also plan to expand the database to consider temperature dependent photodissociation. Acknowledgments {#acknowledgments .unnumbered} =============== We thank the many scientists who have contributed directly or indirectly to the ExoMol project including Ala’a Azzam, Emma Barton, Bob Barber, Phillip Coles, Lorenzo Lodi, Barry Mant, Andrei Patrascu, Anatoly Pavlyuchko, Clara Sousa-Silva, Tom Rivlin, Daniel Underwood, Andrei Yachmenev and Peter Bernath. This work was supported by the European Research Council (ERC) under the Advanced Investigator ExoMol Project 267219 and the ExoAI project 758892, and STFC through grants ST/H008586/1 and ST/K00333X/1. [284]{} \[1\][\#1]{} \[1\][`#1`]{} \[2\][\#2]{} \[1\][\#1]{} \[1\][[](http://arxiv.org/abs/#1)]{} \[1\][[](http://dx.doi.org/#1)]{} \[1\][[](pmid:#1)]{} \[2\][\#2]{} , . . ;:. . , , . . ;:. . . . ;:. . , , , . . ;:. , , . . ;:. . , . . ;:. . , , , , . . ;:. . , , , . . ;:. . , , . . ;:. . , , , , , , et al. . ;:. , , , , , , et al. . ;:. , , , , , . . ;:. . , , , , . . ;:. . , , , , , . . ;:. , , , . . ;:[ arXiv:1912.07759](http://arxiv.org/abs/1912.07759). , , . . ;:. , , , , , , et al. . ;:. . , , , , , , et al. . ;:. , , , , , , et al. . ;:. . , , , , , , et al. . ;:. . , , , , , , et al. . ;:. . , , , , , , et al. . ;:. . , , , , , , et al. . ;:. . , , , , . . ;. , , , , , . . ;:. . , . . ;:. . , . . ;:. . , , , , , . . ;:. . , , , . . ;():. . , . . ;:. . , , , , . . ;:. . . ;. , . ;. , , , , , , et al. . ;:. . , , , , , , et al. . ;:. . , , , , , , et al. . ;:. . , , , , , . . ;:. . . . ;. , , , , , , et al. . ;:. . , , , , , , et al. . In: , editor. ; vol. of **. , p. . . , , , , , , et al. . In: , editor. ; vol. of **. ; ,. , , , . . ;. , , . . ;:. . , , , . . ;. , , , , , . . ;:. . , . . ;():. , , , , , , et al. . In: . ,. , , , . . ;:. . . . In: , editor. ; vol.  of **. . ISBN ; , p. . . , , . . ;:. . , , . . ;():. . , . ;:. . , , , . . ;:. . , , , . . ;:. . . . ;:. , , , , , . . ;:. , , , , , . . ;:. , , , , , , et al. . ;:. , , , , , , et al. . ;:. , , , . . ;:. . , , , , , . . ;:. . , , , , , , et al. . ;:. . , , , , , . . ;:. . , , , , , . . ;:. . , . . ;:. . , , , . . ;:. . , , , , , , et al. . ;:. . . . ;:. . . . ;:. , , , , , . . ;:. . , , , , , , et al. . ;:. , , , , , , et al. . ;:. , , , , , . . ;:. . , , , , , , et al. . ;:. , , , , , , et al. . ;:. . , , , , , , et al. . ;:. , , , , , , et al. . ;:. . , , , , , , et al. . ;:. , , , , , . . ;. , , , . . ;:. . . . ;:. . . ;:. . , , . . ;:. . , . . ;:. . , , , , , . . ;:. . . . ;:. . . . ;:. . , , , . . ;:. . , , , . . ;:. . . ;:. . , , . . ;:. . , , , . . ;:. . , . . ;:. . , , , , , . . ;:. , , . . ;:. , , , , , , et al. . ;:. . , . . ;:. , , , . . ;:. . , , , , , , et al. . ;:. . , , , . . ;:. . , , , . . ;:. . , , , . . ;:. . , , . . ;:. . , , , , . . ;:. . , , . . ;:. . , , , . . ;:. . , , , , , . . ;:. . , , , , , , et al. . ;:. . , , , , . . ;:. . , , , . . ;:. . , , , , . . ;:. . , , . . ;:. . , , , , , . . ;:. . , , , , , , et al. . ;:. . , , , , , . . ;:. . , , , , . . ;:. . , , , , , . . ;:. . , , , , , . . ;:. . , , , . . ;:. . , , , , , , et al. . ;:. . , , , . . ;:. . , , , , . . ;:. . , , , , . . ;:. . , , , . . ;:. . , , . . ;:. . , , , , , , et al. . ;:. , , , . . ;:. . , , , . . ;:. . , , . . ;:. . , , . . ;:. . , , , . . ;:. . , , , . . ;. , , , , , . . ;. , , , , , . . ;:. , . . ;:. , , , , . . ;:. . . ;:. . , , , , , . . ;:. . , , , , , , et al. . ;:. . , , , , . . ;:. . , , . . ;:. . , , , . . ;:. . , , , , , , et al. . ;:. . , . . ;:. . , . . ;:. . , , . . ;. . , . . ;:. . , . . ;:. , , . . ;:. . , . . ;:. . , , . . ;:. . , , , , , , et al. . ;:. . , , , , . . ;:. . , , , . . ;:. . , . . ;:. . , , , . . ;:. . , , , . . ;:. . , , , , , . . ;:. . , , , . . ;:. . , , , , . . ;:. . , , . . ;:. . , , , , , . . ;:. . , , . . ;:. . , , . . ;:. , , , , , , et al. . ;:. . , , , , , , et al. . ;():. <https://doi.org/10.3847%2F1538-4357%2Fab1f65>. . , , , , . . ;:. . , , , , . . ;:. . , . . ;():. . , , , , . . ;:. . , , , . . ;:. . , , , , , , et al. . ;:. . , , , , . . ;:. . , , , . . ;:. . , , , . . ;:. . , , , . . ;:. . , , . . ;:. , , , . . ;:. , , . . ;:. . , , , . . ;:. , , , , , . . ;:. . , , , , , , et al. . ;:. . , , , , , , et al. . ;:. . , . . ;:. . , , , , , . . ;. , , , . . ;:. . . . ;:. . , , , , . . ;:. . , , , , , . . ;:. . , , . . ;:. . , , , , , , et al. . ;:. . , , , , . . . [arXiv:2004.13679](http://arxiv.org/abs/2004.13679). , , , , , . . ;:. . , , , , , , et al. . ;. , , , , . . ;:. . , . . ;:. . , , . . ;:. . , , . . ;:. , , , , , , et al. . ;:. . , , , , , , et al. . ;:. . , , , , , . . ;:. . . . ;:. . . ;:. . , , . . ;:. . , , , , , , et al. . ;:. . , , , , , , et al. . ;:. . , , . . ;:. . , , , , , , et al. . ;:. . , , , . . ;:. . . . ;:. . . ;:. , , , . . ;:. . , , , . . ;:. . , , , , , , et al. . ;:. . , , , , . . ;:. . , . . ;:. . , , , , , , et al. . ;:. . , , , , , , et al. . ;:. . , , . . ;:. . , , , , . . ;:. . , , , , , . . ;:. . , , , . . ;:. . , , , , , , et al. . ;:. . , , , , , , et al. . ;:. . , , , , , , et al. . ;. , , , , , . . ;:. <https://link.aps.org/doi/10.1103/PhysRevResearch.2.023091>. . , , , , , . . ;:. . , , . . ;:. . , , , , , , et al. . ;:. . , , , , , , et al. . ;:. . , , , . . ;:. . , , , , , , et al. . ;:. . , , , . . ;. , , , , , , et al. . ;:. . , , , , , , et al. . ;:. . , , , , . . ;. , , , , . . ;:. . , , , , . . ;:. . , , , , , , et al. . ;:. . , , , , , , et al. . ;:. . , , . . ;:. . , , , . . ;:. , , . . ;:. , , , , . . ;:. . . . ;:. . , , , , . . ;:. . , , , , . . ;:. . , , , , . . ;:. . . ;:. . , . . ;:. . . ;:. , . . ;:. . , , , , . . ;:. . , , , , , , et al. . . [arXiv:2004.04185](http://arxiv.org/abs/2004.04185). , , , , . . ; , , , , , , et al. . ;:. . , , . . In: ; vol. of **. ; , p. . . . . ;:. . , , , , , , et al. . ;:. . , , , , , , et al. . ;:. . , , , , , , et al. . . . , , , , , , et al. . ;. , , , , , , et al. . ;:. , , , , , , et al. . ;:. . , , , , , . . ;. , , . . ;:. . , , , , , , et al. . ;:. . , , , , , , et al. . ;:. . , , , . . ;:. . , , , , . . ;:. , , , . . ;:. . , , , , , , et al. . ;:. . , , , , , , et al. . ;:. . , , , , , , et al. . ;:. . , , , . . ;:. . , . . ;. , , , , , . . ;:. . , , , , , , et al. . ;:. . , , , , . . ;:. . . . In: , , , editors. ; vol. of ** ; , p. . , . . ;:. , , , , , , et al. . ;:. . , , , , , , et al. ;:. . , , , , , . . ;:. . [^1]: Corresponding author

ArXiv

--- author: - Marina Battaglia - Lyndsay Fletcher - 'Arnold O. Benz' bibliography: - 'mybib.bib' date: 'Received /Accepted' title: Observations of conduction driven evaporation in the early rise phase of solar flares --- [The classical flare picture features a beam of electrons, which were accelerated in a site in the corona, hitting the chromosphere. The electrons are stopped in the dense chromospheric plasma, emitting bremsstrahlung in hard X-rays. The ambient material is heated by the deposited energy and expands into the magnetic flare loops, a process termed chromospheric evaporation. In this view hard X-ray emission from the chromosphere is succeeded by soft-X-ray emission from the hot plasma in the flare loop, the soft X-ray emission being a direct consequence of the impact of the non-thermal particle beam. However, observations of events exist in which a pronounced increase in soft X-ray emission is observed minutes before the onset of the hard X-ray emission. Such pre-flare emission clearly contradicts the classical flare picture.]{} [For the first time, the pre-flare phase of such solar flares is studied in detail. The aim is to understand the early rise phase of these events. We want to explain the time evolution of the observed emission by means of alternative energy transport mechanisms such as heat conduction.]{} [RHESSI events displaying pronounced pre-flare emission were analyzed in imaging and spectroscopy. The time evolution of images and full sun spectra was investigated and compared to the theoretical expectations from conduction driven chromospheric evaporation. ]{} [The pre-flare phase is characterized by purely thermal emission from a coronal source with increasing emission measure and density. After this earliest phase, a small non-thermal tail to higher energies appears in the spectra, becoming more and more pronounced. However, images still only display one X-ray source, implying that this non-thermal emission is coronal. The increase of emission measure and density indicates that material is added to the coronal region. The most plausible origin is evaporated material from the chromosphere. Energy provided by a heat flux is capable of driving chromospheric evaporation. We show that the often used classical Spitzer treatment of the conductive flux is not applicable. The conductive flux is saturated. During the preflare-phase, the temperature of the coronal source remains constant or increases. Continuous heating in the corona is necessary to explain this observation. ]{} [The observations of the pre-flare phase of four solar flares are consistent with chromospheric evaporation driven by a saturated heat flux. Additionally, continuous heating in the corona is necessary to sustain the observed temperature. ]{} Introduction ============ The question of energy conversion during the impulsive phase of a solar flare has converged over the course of the past three decades on a picture featuring an acceleration site in the corona, where particles accelerated to high energies then precipitate along magnetic field lines to the chromosphere. In the chromosphere they lose their energy in the dense plasma, leading to chromospheric heating, as well as the characteristic hard X-ray ‘footpoint’ emission. In addition to radiating and conducting its excess energy away, the heated chromospheric plasma finds a new equilibrium by expanding up the loop, in a process termed ’chromospheric evaporation’. The observational evidence for a relationship between flare heating and non-thermal electrons in the impulsive phase is reasonably strong. In this paper we use imaging and spectroscopy from the RHESSI satellite [@Li02], as well as information from the GOES satellites to investigate flare heating before the flare impulsive phase, apparently in the absence of non-thermal electrons. Chromospheric evaporation is usually proposed as the cause of flare extreme ultraviolet (EUV) and soft X-ray (SXR) emission [though see @Fe90; @Ac92 who dispute this on several grounds].\ The main observational arguments for chromospheric evaporation in flares are (1) that the density of the coronal plasma observed to be emitting in SXRs during flares is one or two orders of magnitude greater than is generally measured in the quiet corona [eg. @AA01; @Kr08], requiring a source for the additional material and (2) the presence of upflowing plasma detected via blueshifts in high temperature spectrum lines [@An82], in particularly those observations made with imaging spectrometers such as the Coronal Diagnostic Spectrometer [CDS, @Ha95] on the Solar and Heliospheric Observatory [SOHO, @Dom95]. Such flows have been observed in the flare impulsive phase by @Mi06b [@Mi06a] and in the gradual phase by @Cza99. Evaporative flows have been described as being either ‘gentle’ or ‘explosive’, with the latter occurring when the rate of chromospheric plasma heating greatly exceeds the rate at which it can cool by radiation, conduction or expansion, primarily determined by the ratio between the heating timescale and the hydrodynamic expansion timescale [@Fi85]. The relatively slow heating rate by conduction from a heated corona makes it a likely source of gentle evaporation. @An78 developed an analytic model of conductive driven evaporation in the decay phase of flares, though report that the ‘evaporation’ represented the hydrodynamic redistribution of material already at coronal temperatures, rather than chromospheric material being heated and expanding into the corona. @Kar87 carried out numerical simulations exploring the effects of non-local and saturated heat flux in addition to classical Spitzer heat flux. Observational evidence for such “gentle” evaporation was found by @Za88 in observations from the Solar Maximum Mission, and the flows observed by @Mi06b and @Cza99 [@Cz01] were also found to be consistent with conductive evaporation, due to both their relatively low speeds and absence of significant HXR radiation. In the electron beam-driven evaporation model, the thermal X-ray emission is a direct effect of the energy deposition in the chromosphere by the electron beam, so that the time-integrated non-thermal hard-X-ray (HXR) flux in a given energy range is proportional to the SXR flux. This was first proposed by @Ne68 and since then, the so-called ‘Neupert effect’ has been studied at length by various authors [e.g. @De93; @McT99; @Ve05]. Overall the expected proportionality between the time-derivative of the SXRs and the instantaneous HXRs is supported observationally, particularly for impulsive flares, though there is a significant scatter interpreted as due to energy losses from low-temperature radiation, conduction and mass motion. There are also some significant exceptions from this rule - for example, @McT99 find that the Neupert relationship manifested in about half of the 33 flares they studied. A similar result was found by @Ve02ap. This suggests additional energy input, not related to non-thermal electrons, in the non-Neupert flares. Various authors have also observed impulsive SXR footpoint emission tracking the HXR intensity, rather than its integral [e.g. @Hu94]. Finally, pre-flare SXR sources, occurring in advance of any HXR emission present a further instance in which thermal emission may also be completely unrelated to non-thermal electrons. This paper concerns the observation and interpretation of such sources. It has been known for some time that flare-related activity commences prior to the flare impulsive phase. There is a distinction to be drawn between pre-flare activity, which refers to the very earliest stages of the flare before the impulsive phase radiation is detectable, and ‘flare precursor’ events, which are small-scale brightenings in UV to HXR wavelengths happening some tens of minutes before the flare. The first to use the term “pre-flare” and complete a statistical study on flare precursors in X-rays were @Bu59, who inferred the X-ray behavior via ionospheric disturbances. With the arrival of the *Yohkoh* satellite [@Og91], the nature of the pre-flare phase was studied in more detail, particularly the relationship between the non-thermal and thermal emissions. It was noted relatively early that high temperature thermal sources were present in the corona substantially before the impulsive phase flare onset [@Ac92], a result which is clearly at odds with the assumption that electron beams drive evaporation. Comprehensive statistics including an analysis of the spatial relations between pre-flares and the consequent flares has been conducted by @Fa96 and @Fa98 based on observations with the Yohkoh Soft X-ray Telescope [SXT, @Ts91]. Their definition of pre-flare activity is a rise in the *Yohkoh* emission above background, five minutes to an hour before the main peak. In the spatially-resolved SXT images it is clear that in several of their sample there is substantial coronal soft X-ray emission occurring several minutes before the start of the impulsive phase [@Fa98]. In a study of 10 *Yohkoh* flares, @Al98 also identify line broadening observed with *Yohkoh* Bragg Crystal Spectrometer [@Cu91] several minutes before the onset of the impulsive phase, and @Ha01 show that in one such event the temperature, intensity, and non-thermal line width are changing significantly, well before the start of the hard X-rays. Taken together, the *Yohkoh* observations provide strong evidence for pre-flare coronal activity unrelated to evaporation driven by electron beams. In recent years, RHESSI observations further confirmed the existence and relative relative frequency of pre-flare activity as shown eg. in @Ve02. The Ramaty High Energy Solar Spectroscopic Imager [RHESSI, @Li02] offers new opportunities to study pre-flare sources in detail. RHESSI is a HXR telescope, with spectral coverage extending down in energy to the thermal range at $\sim$ 3-20 keV. Its high spectral resolution (of about 1 keV at energies up to 100 keV) permits spectroscopic diagnostics not available previously, coupled with high spatial resolution (as low as 2.3", depending on the flare) in user-defined energy channels. This offers new temperature and emission measure diagnostics for very hot plasmas (around 10 MK and above) as well as for non-thermal electrons, if present. The RHESSI observations presented here display an increase of low energy X-ray emission, which can be fitted with a thermal spectrum, up to minutes before the rise of the HXR emission. The morphology of these events shows only one source, visible at energies in the thermal regime, while the footpoints only appear at the onset of the HXR emission. We will argue that this indicates that another mechanism of energy transport to the chromosphere may be important, such as heat conduction.\ In this study, we present the first comprehensive RHESSI-led study of the pre-flare phase of solar flares. The paper is structured in the following way: Section \[selection\] describes the flare selection process. In Sect. \[tevol\] the time evolution of the selected flares is analyzed in spectra and images. A theoretical model to explain the observed time evolution is presented in Sect. \[theory\], followed by a Discussion and Conclusions. Flare selection {#selection} =============== {height="10cm" width="16cm"} In this study, events with pre-flare activity will be analyzed. The main pre-requisite of such a flare is increasing emission in SXR well before the onset of the HXR emission. Therefore, only events the beginning of which was fully observed were considered. The selection criteria are summarized here: - Offset of more than 700$"$ from disk center to limit projection effects in the study of the flare morphology; - GOES class larger than M1 to ensure good count rates for imaging; - The beginning of the event had to be observed with RHESSI; - SXR & GOES emission had to start increasing at least 1 min before the HXR emission; - Events with a simple morphology involving one SXR source at the onset and two HXR sources appearing later in the event. This indicates the time when the acceleration process becomes dominant; - High enough detector livetime ($>$ 90 %) for reliable spectroscopy and imaging. Four events were selected as best suited for this kind of study. Their key parameters are listed in Table \[events\]. The RHESSI attenuators were out during the earliest phase of all events with the thin attenuator moving in as the count rates increased, typically after 2-3 minutes. Date Analyzed time GOES class xy-position \[arcsec\] coronal source volume \[cm$^3$\] Attenuator Attenuator times ------------- ------------------- ------------ ------------------------ ---------------------------------- ------------ ------------------- 03-Aug-2002 18:59-19:02 X1.2 912/-271 4.9$\cdot 10^{26}$ 0/1 19:02:20/19:03:10 13-Nov-2003 04:55:40-04:57:12 M1.7 -977/23 4.8$\cdot 10^{26}$ 0/1 04:57:20/04:58:20 06-Jan-2004 06:13-06:16 M5.9 -992/88 2.8$\cdot 10^{27}$ 0/1 06:18/06:19 21-May-2004 23:37:30-23:39 M2.6 -757/-157 7.5$\cdot 10^{26}$ 0/1 23:40:30/23:41:30 \[events\] Time evolution of events {#tevol} ======================== The time evolution of the SXR emission in GOES and RHESSI 6-12 keV along with the HXR emission (RHESSI 25-50 keV) is shown in Fig. \[lcurves\]. The missing data in the RHESSI lightcurves indicate the time intervals omitted due to attenuator state change. The attenuator state was 0 for all events at the onset of the flare.\ Data analysis ------------- We studied the time evolution of the pre-flare phase both in images and spectra. CLEAN and Pixon images [@Hur02; @Me96] during the pre-flare and early impulsive phase were made, using grids 3 to 8. The time interval of each image was 30s in order to get high enough count rates throughout the pre-flare phase. Full-sun photon spectra were fitted in 30 s time intervals, corresponding to the image time intervals. The fitting model consisted of a thermal component fitted from 6 keV and, if possible, a non-thermal component. Spectra from the onset of the flares as well as the start of the non-thermal emission are shown in Fig. \[spectra\]. The RHESSI spectrum at the lowest energies (3-6 keV) could be fitted with an additional thermal component with temperature and emission measure similar to the values that can be derived from GOES measurements (not shown in the figure).\ Spatial evolution {#spatialevol} ----------------- {width="17cm"} Figure \[cleanimg\] shows CLEAN images of the events taken at the beginning, at the time when a non-thermal component was first fitted and at the time when the footpoints first appear. For each event, 3 energy-bands (6-12, 12-25, 25-50 keV) are shown. All events start with a single source visible only at the lowest energies in the range 6-12 keV (upper left image of the 9-image panel per flare in Fig. \[cleanimg\]. This is interpreted as a source at the top of a loop (coronal source). After some minutes two additional sources appear at higher energies (25- 50 keV) which are interpreted as chromospheric footpoints (lower right image of each flare in Fig. \[cleanimg\]). In three events the position of the first appearing source is clearly displaced from the footpoint position, implying that there is an actual loop geometry with a coronal source on top and footpoints in the chromosphere. In the fourth event, the separation is not as clear but from the spectral evolution (see next Section) one can still assume a coronal-source-footpoints-geometry. ### Movement of coronal sources We analyzed the time evolution of the source position by measuring the centroid position of the 50% contour in CLEAN and Pixon images in the energy range 6-12 keV. The position from both imaging methods agree within the uncertainties. Figure \[sourcemotion\] displays the 70% contour from CLEAN images with positions as found from CLEAN. Colors go from black to light (time of first appearance of footpoints). The positions shown indicate the position of the coronal source from the beginning until the time of the first appearance of footpoints, in the same time steps as used in spectroscopy (30s). In all events, the sources move to some extent during the pre-flare phase. The source in the event of 03-Aug-2002 is stable within the position uncertainties during the pre-flare phase, but appears at a displaced location in the last image (at the first appearance of the footpoints). Due to the attenuator state change just before the last image, an interval of about 1 minute is missing. Therefore, one cannot say whether the source moves continuously or not. The same observation is made in the 13-Nov-2003 event. Here, the sources observed during the pre-flare phase and at the beginning of the impulsive phase could possibly indicate two separate loop systems as suggested by @Liu06. The source position of the 6-Jan-2004 event remains constant over time. The 21-May-2004 event is somewhat peculiar, displaying pronounced, continuous source motion. The displacement of the position is clearly larger than the uncertainties. From the direction of the movement, this could be another case of altitude decrease as has been found in other RHESSI events [eg. @Sui03; @Ve06] Spectral evolution {#specevol} ------------------ All four events display a purely thermal spectrum in the first of the analyzed time intervals. This spectroscopic finding is further supported by GOES SXI images (not shown here). These are available for three out of four events and show a soft X-ray source minutes to hours before the RHESSI observations. In RHESSI observations the apparently purely thermal phase lasts one to two minutes, after which a small tail to higher energies becomes visible (Fig. \[spectra\]). At this stage, there are as yet no footpoints are observed, indicating that the non-thermal emission is coronal. The tail appears independent of albedo correction and as the events are all near the limb, the influence of albedo on the total flux is minimal. Further, pile-up is not large enough at this stage to account for the emission. It is therefore safe to assume that the tail is real. It can be fitted with a power-law or with a second, very hot thermal component. During the impulsive phase (after attenuator state change) the HXR component becomes very pronounced and clearly distinguishable. Figure \[spectra\] shows spectra of the selected events, one taken at flare onset and one when the non-thermal tail first appears. {height="10cm" width="17cm"} {height="16cm"} Time evolution of flare parameters ---------------------------------- The time evolution of the fit parameters, namely temperature, emission measure and spectral index is shown in Fig. \[fitparams\]. The figure displays panels for each event. The top panel of each event shows RHESSI lightcurves in the 6-12 and 25-50 keV energy band. The cross indicates the time of first appearance, and the hardness, of the non-thermal component. The second panel displays the emission measures as determined from RHESSI spectra and GOES. In the third panel, temperatures are presented the same way as the emission measures. The fourth panel shows the electron density estimates. The critical parameter in computing densities is the flare volume. RHESSI images allow an estimate of the volume by measuring the source area $A$ and approximating the volume as $V=A^{3/2}$. Measuring the area from RHESSI images is quite imprecise with large uncertainties. We determined the 50 % contours in CLEAN images (de-convolving the CLEAN beam). The area is constant over the pre-flare evolution and we made an estimate of the quantitative value of the volume by taking the time average of the measured values. The average electron density in the coronal source is then given as $$n_e\approx\sqrt{EM/V}\,,$$ assuming a volume filling factor of 1. A theoretical model to explain the observations {#theory} =============================================== In Fig. \[fitparams\] we present the time evolution of flare parameters such as temperature, emission measure and density, as well as images. The observations reveal increasing emission measure and densities over the course of the pre-flare as well as generally increasing temperatures. Assuming a constant, or even increasing, source volume, the density increase can only be attributed to additional material which is added to the coronal source region. The most likely scenario causing such a density increase is chromospheric evaporation. In this scenario, chromospheric plasma (including the transition region) is heated and expands upward the magnetic loops. There are two possible main heating mechanisms for the chromosphere; non-thermal electron beams and thermal conduction. In the former case, a beam of high energetic electrons impinges on the chromosphere, where the particles are completely stopped in the dense target. As a consequence of this energy deposition, the temperature rises. In the latter case, energy is transported from the hot coronal source to the cooler chromosphere by thermal conduction. We are going to discuss the thermal conduction scenario further in the following subsection. Reasons why thermal conduction is favored over particle beams are presented in the discussion. Theory of thermal conduction {#condtheory} ---------------------------- Our model will explain chromospheric evaporation by heating due to energy input from heat conduction. The coronal source is much hotter than the chromospheric plasma, therefore a temperature gradient develops, driving a heat flux downward along the magnetic loop. The conductive heat flux is given as $$F_{cond}=\kappa_0T^{5/2}\frac{\partial T}{\partial s}.$$ In classical thermal conduction theory following @Spbook, the conductive coefficient is $\kappa_0=10^{-6}$ erg cm$^{-1}$s$^{-1}$K$^{-7/2}$. For a flare of half loop length $L_{loop}$ and coronal source temperature $T_{cs}$, this may be approximated by $$F_{cond}\approx 10^{-6}\frac{T_{cs}^{7/2}}{L_{loop}}\quad\quad \mathrm{[erg\, cm^{-2} s^{-1}]}.$$ However, the maximum heat flux a plasma can carry is limited by a fraction of the thermal energy-flux $F_{therm}=n_ekT_ev_{th}$, where $v_{th}$ is the thermal electron velocity. For sufficiently large temperature gradients, such as occur in solar flares, the heat flux is expected to reach this limit and saturate. @Gra80 showed that there are two regimes of flux saturation. A non-classical treatment with local flux limiting is necessary if the electron mean free path $\lambda_{emf}=5.21\cdot 10^3 T^2/n_e$ [@Bebook] is larger than only 0.12% of the temperature scale length $L_{th}$. If the electron mean free path even exceeds the temperature scale length, the heat flux becomes non-local in the sense that it depends on the global density and temperature structure of the plasma. Figure \[noncllocal\] illustrates the boundary conditions for locally limited and non-local heat flux in the density-temperature plane of typical solar flare values. The locally limited regime is described by @Ca84 as a continuous transition from the purely classical treatment to the non-local regime. Depending on the ratio $\mathcal{R}=\lambda_{emf}/L_{th}$, a reduction factor $\varrho<1$ is applied to the classical heat flux, resulting in an effective heat flux of $F_{red}=\varrho (\mathcal{R})\cdot F_{cond}$. @Ca84 published the values of $\varrho$ for different $\mathcal{R}$ and atomic number Z. Using the values for Z=1, we fitted a function of the form $$\varrho=A\cdot e^{-b(x+c)^2} \label{corrfact}$$ to the values published by @Ca84. The parameters are $x=\ln (\mathcal{R})$, A=1.01, b=0.05, c=6.63. Using those values, the locally limited, reduced heat flux can be computed for any $\mathcal{R}$ from Eq. (\[corrfact\]). In the non-local regime where the electrons are freely streaming and complete saturation is reached, the flux can be expressed as: $$F_{sat}=0.53n_em_ev_{th}^3\quad \mathrm{erg\, cm^{-2} s^{-1}}$$ [@Kar87]. Observation of thermal conduction {#obsthcond} --------------------------------- In studies of thermal conduction in flare loops, a classical scenario involving Spitzer conduction is often assumed. Are the conditions in the observed pre-flares consistent with classical conduction or is the flux limited? The densities and temperatures as derived from GOES and RHESSI observations are displayed in Fig. \[noncllocal\]. Values measured with GOES are given as dots, RHESSI measurements are marked as triangles. GOES typically yields larger emission measures and smaller temperatures than RHESSI. Assuming the same emission volume, this results in higher densities for GOES. The figure reveals that the conditions during the observed times of the pre-flare cannot sustain a classical heat flux, the heat flux is limited. The resulting effective flux was computed from the observed values. Temperatures and densities yield the electron mean free path $\lambda_{mfp}$. The temperature scale length $L_{th}$ is given as $L_{th}=T/\Delta T \approx L_{loop}$, where $\Delta T$ was approximated as $\Delta T \approx T_{cs}/L_{loop}$. Using an estimated loop half-length of $10^9$ cm we computed the ratio $\mathcal{R}$. The correction factor and therefore the effective heat flux was computed from Eq. (\[corrfact\]). The results are displayed in Fig. \[condfluxes\], again for both, GOES and RHESSI values. Chromospheric evaporation ------------------------- We assume that the heat flux computed in the section above is deposited in the transition region and chromosphere, heating the chromospheric material. The hot material will then expand upward the magnetic loops. This chromospheric evaporation is usually divided into two types; gentle evaporation and explosive evaporation. The determining factors distinguishing the two are the ratios between heating, radiative cooling and expansion rate in the chromosphere. If the heating rate is much smaller than the radiative loss rate, all the deposited energy is radiated away. If the heating rate is only marginally larger than the radiative loss rate, the temperature and pressure rise slowly and the material evaporates gently. If the deposited energy rate exceeds a certain threshold value, the chromospheric plasma is heated to coronal temperatures quickly, a large overpressure develops and explosive expansion of plasma is the consequence. @Fi85 derived a value for the critical heating flux of 10$^{10}$ erg cm$^{-2}$s$^{-1}$ for typical coronal and chromospheric conditions. It is generally assumed that explosive evaporation is associated with the large energy input provided by beams of fast electrons hitting the chromosphere, while gentle evaporation is normally attributed to heating by thermal conduction. ### Conduction driven chromospheric evaporation Can the observed emission measure and density increase in the presented events be explained with chromospheric evaporation driven by the heat flux presented in Sect \[obsthcond\]? The total energy flux in the evaporated plasma is given as $$F_{up}^{evap}=3n_e^{evap}kT^{evap}v_{up}=3m_en_e^{evap}v_{th}^2v_{up} \quad \mathrm{[erg\, cm^{-2} s^{-1}]}, \label{fup}$$ where $v_{up}$ is the velocity of the upward moving plasma, $v_{th}$ the thermal velocity and $n_e^{evap}$ the density of the upflowing plasma. We use the evolution of the density in the coronal source to estimate $v_{up}$. $$\begin{aligned} \frac{\mathrm{d}n_e}{\mathrm{d}t}&=&\frac{\partial n_e}{\partial s}\frac{\partial s}{\partial t} \approx \frac{n_e^{evap}}{L_{loop}}v_{up} \\ v_{up} &\approx & \frac {L_{loop}}{n_e^{evap}}\cdot \frac{\mathrm{d}n_e}{\mathrm{d}t} \label{upflowvel} \end{aligned}$$ The density of the upflowing plasma cannot be determined from observations. Assuming $n_e^{evap}=n_e$, the density in the coronal source, the upflow velocities derived from Eq. (\[upflowvel\]) range from 20 km s$^{-1}$ to 150 km s$^{-1}$. This results in an evaporative flux of the order of 10$^8$ erg $\mathrm{cm^{-2} s^{-1}}$. Further, the evaporative expansion requires an energy of $$E_{exp}=\int pa \mathrm{dl}$$ where p is the total pressure. Assuming isothermal expansion the expansion energy flux can be computed [see @Kb02]: $$F_{exp}=2m_en_ev_{th}v_{up}\ln(L_{loop}/l_i)$$ The expansion energy flux is larger than the thermal energy flux (Eq. \[fup\]) by a factor of $2/3ln(L_{loop}/l_i)$ which amounts to about 5. The total energy flux for evaporation is therefore in the range of $5\cdot 10^{8}$ erg $\mathrm{cm^{-2} s^{-1}}$, which is well below the values obtained from the data (Fig. \[condfluxes\]). The rest of the conducted energy heats the transition region and upper chromosphere to sub-coronal temperatures and is radiated away. Discussion {#cond:discussion} ========== Electron beams in the pre-flare phase ------------------------------------- In the above section, we demonstrated that conduction driven chromospheric evaporation can account for observations of pre-flares. Can electron beams below the detection limit do the same? Let us estimate the maximum expected non-thermal electron flux and the expected chromospheric heating rate. We estimate the electron beam strength starting from the observed coronal source spectrum. The first time-interval when non-thermal emission from the corona was observed was used as an estimate of the upper limit to the electron flux, assuming a thin target in the coronal source. If a beam constituting of this electron flux hits the chromosphere, a thick target footpoint spectrum $F_{fp}=AE^{-\gamma}$ is expected. From observations of footpoints and coronal sources [eg. @Ba06], we know that a flux $F_{fp}$ as found from this estimate would be well observed in images and spectra which is not the case. Taking the observations of @Ba06 as a reference, we estimate that the effective electron flux that reaches the footpoints has to be about a factor 100 smaller than expected from the coronal source observations. For the computation of the beam heating rate, we therefore use an assumed footpoint photon flux spectrum $F_{fp}^{eff}=0.01\cdot F_{fp}$. The heating rate of a beam of electrons was derived by @Br73a and @Li76. The rate of energy input per unit volume at column density $N$ and time $t$ is given as $$\begin{aligned} I_B(N,t)&=&10^7n(N,t)[x(N,t)+0.55] \nonumber\\ &\times&\frac{A}{a}C(\gamma)(1.1\cdot10^{-17}N)^{-(\gamma+1)/2} \quad\quad \mathrm{[erg\,cm^{-3}]} \label{nonthenrate}\end{aligned}$$ where $n(N,t)$ is the hydrogen density, $x(N,t)$ the fractional ionization, $a$ the cross-section area, $A$ the normalization of the observed photon spectrum, $\gamma$ the spectral index of the observed photon spectrum and $C(\gamma)$ a function of the $\beta$-function and $\gamma$. The parameters of the expected footpoint spectrum were used as input for $A$ and $\gamma$ in Eq. (\[nonthenrate\]). As an estimate of the column density, the stopping depth of a 35 keV electron was used according to $N=10^{17}E_{\mathrm{kev}}^2$ [@Ta88]. The resulting density $n(N,t)$ and fractional ionization $x(N,t)$ was taken from @Br73a [Table1]. The cross-section area was taken as $10^{17}$ cm$^2$, which corresponds to typical footpoint areas as measured by RHESSI. Finally, the factor $C(\gamma)$ was taken from Fig. 2 in @Li76. Inserting the above values gives a total energy input of the order of 3-4 erg cm$^{-3}$s$^{-1}$. From the expression of the column density $N=nl$, where $l$ is the path length, l can be determined and by multiplying Eq. (\[nonthenrate\]) with $l$ we get an energy input of 5-7$\cdot$10$^7$ erg cm$^{-2}$s$^{-1}$, which is two orders of magnitude smaller than the energy input by thermal conduction. Even if the effective footpoint area was an order of magnitude smaller, as often found in UV or H$\alpha$ observations [eg. @Te07], the heating rate by the beam would still be an order of magnitude smaller than the energy input by thermal conduction. Therefore, if there is any chromospheric heating by electron beams in the pre-flare phase, its contribution is minimal compared to the energy input by thermal conduction. Influence of assumptions ------------------------ In all computations throughout this work, the values measured in the coronal source were used also for the loop legs. For some parameters, e.g. the electron mean free path $\lambda_{mfp}$, those values may not be accurate. For a complete treatment, the temperature and density in the loop is needed. However, those values cannot be determined observationally. If the loop was at the same temperature and density as the coronal source, it would be visible in X-ray images. As no loop emission is observed in the images, the actual loop has to be either less dense or cooler than the coronal source. Thus the coronal source values represent an upper limit to the true conditions in the loop. Smaller temperature could result in an order of magnitude smaller $\mathcal{R}$. On the other hand, a smaller density could cause an order of magnitude larger ratio $\mathcal{R}$ of electron mean free path to temperature scale length. The correction factor $\varrho$ (Eq. \[corrfact\]) would change by 0.5 to 2, respectively and so would the conductive flux. The resulting upper and lower limits of the conductive flux are indicated by dashed lines in Fig. \[condfluxes\]. Even if the conductive flux differed by a factor of 2, this would not change the basic interpretation. The evaporated flux $F_{up}^{evap}\sim n_eTv_{up}\sim T$ only depends on the temperature. It is therefore expected to be at most about a factor of 4 smaller. However, the upflow velocities derived from Eq. (\[upflowvel\]) are proportional to $1/n_e$ and could therefore be up to an order of magnitude larger for smaller densities. Eq. (\[upflowvel\]) may then give velocities higher than the local sound speed which would be unphysical. Mode of conduction ------------------ Our results indicate that the conductive heat flux during the pre-flare phase is non-classical. The temperature gradient between the coronal source region and the chromosphere is so steep that the classical Spitzer conduction is not valid. The conductive flux is locally limited for temperature and density values determined from GOES and may be fully saturated (i.e. the electrons are freely streaming) in the case of the RHESSI measurements. Both situations lead to a reduced heat flux compared to the classical Spitzer conductivity. This finding does not change directly the conclusions drawn from the work presented here. However, it requires that care be taken when computing total flare energy budgets as the energy loss of the coronal source due to conduction may be smaller than anticipated from the assumption of classical Spitzer conductivity. Acceleration vs. heating ------------------------ Despite the energy loss in the corona due to the heat flux, increasing temperatures during the pre-flare phase are observed. Continuous heating in the coronal source is therefore necessary to sustain and even increase the observed temperature. Possible heating mechanisms include heating at the stand-off slow-mode shocks in Petschek reconnection models which has been proposed as an explanation of the formation of hot thermal sources in Yohkoh flares [eg. @Ts97], or betatron heating in a collapsing magnetic trap [eg. @Karl04]. A strong candidate heating mechanism is wave-particle interaction as described in the transit-time-damping model [@Mi96; @Gr06]. Such a model could also explain the spectral time evolution of the observations: Depending on the escape time and wave density in the acceleration region, particles are not accelerated to non-thermal energies but remain at a quasi-Maxwellian distribution observationally not distinguishable from a purely thermal plasma at enhanced temperature. For increasing wave density or escape time, particles are accelerated to higher energies, constituting a non-thermal tail which is clearly noticeable in the observations. However, the three dimensional geometry of the flare site has to be considered and a smooth transition from heating to acceleration at the same site may not reflect reality in two of the presented events. As shown in Fig. \[sourcemotion\], the position of the coronal source changes over time in the events of 03-Aug-2002 and 13-Nov-2003. @Fa96 and @Fa98 made a statistical study on the position of Yohkoh pre-flares and the successive impulsive flares, finding that only in 25% of the observed events were the pre-flare emission and impulsive flare emission spatially coincident. The other events were classified into what the authors termed “distant” and “adjacent/overlapping” events. In the latter case, parts of the pre-flare 50% maximum intensity emission overlap with parts of the main flare emission while in the “distant” events, the two emission patterns are spatially separated. If our two flares indeed fall into the category of distant events, it will not change the interpretation of the pre-flare phase, but care should be taken when interpreting the time-evolution of the spectrum beyond the purely thermal phase.\ Conclusions =========== The observed increase in thermal emission and SXR emission measure up to minutes before the start of the impulsive flare phase in the four events presented here can be explained by chromospheric evaporation driven by thermal conduction. To compensate for the heat loss, continuous heating in the coronal source is needed. In this early phase, acceleration of particles to non-thermal energies is minimal or non-existent. A scenario following stochastic acceleration is conceivable in which the coronal source region is only heated in the early flare stages. As soon as the acceleration mechanism becomes efficient enough, a tail of non-thermal particles is produced, first visible in the coronal source and eventually in the footpoints. RHESSI data analysis at ETH Zürich is supported by ETH grant TH-1/04-02 and RHESSI software development by the Swiss National Science Foundation (grant 20-113556). This work was supported in part by Rolling Grant ST/F002637/1 from the UK’s Science and Technology Facilities Council and by the European Commission through the SOLAIRE Network (MRTN-CT-2006-035484). This research made use of NASA’s Astrophysics Data System Bibliographic Services and the RHESSI Experimental Data Center (HEDC).

ArXiv

--- abstract: 'This study experimentally investigates the temperature dependence of superheating field, $H_{sh}$, of niobium. Accurately determining this field is important both to test theory and to understand gradient limits in superconducting cavities for particle accelerators. This paper discusses theories that have been proposed in modeling the field and discriminates between them. The experimental procedure for measuring the temperature dependence of $H_{sh}$ utilizes high power pulses to drive a niobium cavity resonator, ramping up surface magnetic fields extremely quickly. The moment any part of the cavity transitions between the superconducting and normal conducting state can be determined by measuring the quality factor of the cavity as a function of time. Oscillating superleak transducers are used to demonstrate that the transition to the normal conducting state is global in nature, showing that a fundamental limit is encountered. Finally, we see that 110-120 $^\circ$C heat treatment of the cavity–a method commonly used to increase the quality factor at high accelerating gradients–may have the deleterious effect of reducing the superheating field of the material, which is the fundamental limiting factor in pursuing the maximal achievable accelerating gradient in superconducting niobium cavities.' author: - 'N. R. A. Valles' - 'M. U. Liepe' bibliography: - './SFoNBib.bib' title: Temperature dependence of the superheating field in niobium --- Introduction ============ An important property of niobium that has posed both theoretical and experimental challenges is the magnetic superheating field. This field, also called the RF critical magnetic field, is the magnetic field at which a material undergoes a phase transition from the superconducting state to the normal conducting state. Superconductors can be broken up into two groups: those with positive surface energy and those with negative surface energy, called Type I and Type II respectively. They can be quantitatively distinguished using the Ginsburg-Landau parameter, $\kappa_{GL}$, which is defined in terms of the London penetration depth, $\lambda_L$ and the coherence length, $\xi_0$ of the material according to $\kappa_{GL} = \frac{\lambda_L}{\xi_0}$.[@Cyrot] A superconductor is Type I if $\kappa_{GL}<1/\sqrt{2}$ and Type II otherwise. Further discussion of these implications is presented in the theory section of this paper. Type I superconductors exclude DC magnetic fields up to a lower critical field $H_c$, after which it becomes energetically favorable for magnetic flux to penetrate the material and cause a transition to the normal conducting state. Type II superconductors also exclude flux at low DC magnetic fields, but above the lower critical field $H_{c_1}$ magnetic flux penetrates the material, creating an array of normal conducting and superconducting regions. As the field is increased, the surface area of the normal conducting regions with additional flux entry increases until the entire sample is normal conducting at a field of $H_{c_2}$. For both Type I and II superconductors, a metastable state exists for fields above their lower critical fields, where flux penetration is delayed, and the material remains fully superconducting in the Meissner state.[@PhysRevLett.12.14; @Sethna] The semiclassical approach for solving the superheating field requires solving the Eilenberger equations.[@Eilenberger] This approach has been taken to solve for the temperature dependence of the superheating field for very strong Type II materials.[@Sethna] Niobium is not a high-$\kappa$ material, so thoroughly solving for $H_{sh}$ is more challenging. An approximate solution of the superheating field is the Ginsburg-Landau theory which predicts the superheating field goes as $$H_{sh}(T) = c(0) H_c \left[ 1 - \left(\frac{T}{T_c}\right)^2 \right], \label{eq:GLTheory}$$ where for niobium, the thermodynamic critical field $H_c = 2000$ Oe, and the critical temperature $T_c = 9.2$ K. The constant $c(0)$ is the ratio of the superheating field and the thermodynamic critical field at zero temperature. [@Dolgert] For very high purity niobium, $c(0)\approx1.2$.[@RFSuper] Note that this phenomenological approach is only valid for $T_c-T\ll T_c$. Another theory put forward by K. Saito and T. Yogi is the Vortex Line Nucleation Model.[@Saito; @Yogi] This theory proposes that the superheating field varies with temperature according to $$H_{sh} = H(0) \left[1 - \left(\frac{T}{\bar{T}_c}\right)^4 \right], \label{eq:VLNModel}$$ where $H(0) = 1780.4$ Oe and $\bar{T}_c = 9.014$ K is the reduced critical temperature of the material. It is important to note that both theories are consistent with previous experiments near $T_c$ but differ from each other significantly at lower temperatures. Aside from the theoretical challenges, measuring the superheating field has also posed experimental challenges. Measurements of $H_{sh}(T)$ have been carried out for niobium, lead and Nb$_3$Sn.[@Campisi87; @HaysHPP; @NickPAC] Both Campisi and Hays show that near $T_c$, $H_{sh}$ is linear with $(T/T_c)^2$ with a slope of $\sim 1.2 H_c$ Oe, but far from $T_c$, the maximum achieved RF critical field only weakly depends on temperature. Hays’ data is shown in Fig. \[fig:HaysData\]. The reduced slope of the maximum field for $\frac{T}{T_c} \ll 1$ may be caused by effects such as heating by an increase in surface resistivity, electron field emission, or normal conducting defects. Any of these effects may have suppressed the results from reaching the experimental superheating field at these temperatures, so further work should be done to insure that a fundamental limit is reached. ![Measurement of the superheating field of niobium from Hays and Padamsee, using a 1.3 GHz cavity that received a buffered chemical polish surface treatment and did not receive a final low temperature heat treatment.[@HaysHPP] Ginsburg-Landau theory predicts $H_{sh} = 1.2 H_c [1-(T/T_c)^2]$ near $T_c$ for pure niobium. At high temperatures, the data agrees with theory, but flattens out at low temperatures.[]{data-label="fig:HaysData"}](./SFoNfig1){width="85mm"} Finally, the accurate determination of the superheating field also is of interest in application. While theory and experiment agree near $T_c$, niobium microwave cavities for particle accelerators operate at temperatures $<T_c/4$, where theory and experiment are disparate. Measurements can set stringent upper bounds for what maximum accelerating gradients are theoretically achievable in these machines. Theory ====== Niobium is a Type II superconductor, meaning it has a negative surface energy. One of the first phenomenological attempts to address superconductors with negative surface energy was Ginsburg and Landau’s treatment.[@GLTheory] They found that the superheating field near $T_c$ has the form described in Eq. \[eq:GLTheory\]. The constant $c(0)\equiv H_{sh}(0)/H_c$ in Eq. \[eq:GLTheory\] is dependent on the order parameter of the material, $\kappa = \lambda/\xi$, which is the ratio of the penetration depth and coherence length of the material. The Eilenberger equations[@Sethna] have been solved near $T_c$ to yield Fig. \[fig:HshHcVsL\].[@Transtrum] ![Plot of the ratio of the superheating field and the thermodynamic critical field, $H_c$, at zero temperature, versus the Ginsburg-Landau parameter, $\kappa$.[@Transtrum][]{data-label="fig:HshHcVsKappa"}](./SFoNfig2){width="50.00000%"} Very pure samples of niobium have electron mean free paths, $\ell$, that are large compared to the coherence length and London penetration depth. As the purity of the niobium decreases, it enters the ‘dirty’ regime where $\ell \ll \lambda (\sim \xi)$. Impurities can depress the critical temperature of the superconductor, which in turn affects the field $H_c$ such that $T_c \propto H_c$. [@RFSuper] Pure niobium has a coherence length of $\xi_0 = 64$ nm, and a London penetration depth $\lambda_L = 36$ nm.[@NbProps] The coherence length, $\xi$ and penetration depth, $\lambda$ of a sample with a given mean free path, $\ell$ can be calculated–near absolute zero–according to Pipard’s equations:[@Pipard] $$\begin{aligned} \frac{1}{\xi} &=& \frac{1}{\xi_0} + \frac{1}{\ell},\\ \lambda &=& \lambda_L \sqrt{1 + \frac{\xi_0}{\ell}}.\end{aligned}$$ It is straightforward to write the Ginsburg-Landau parameter as a function of $\ell$: $$\kappa(\ell) = \frac{\lambda_L}{\xi_0} \left(\frac{\xi_0+\ell}{\ell}\right)^{3/2} \label{eq:KappaOfL}$$ Using Eq. \[eq:KappaOfL\] the superheating field coefficient, $c(0)$, can be plotted versus the electron mean free path, shown in Fig. \[fig:HshHcVsL\]. ![Plot of the ratio of the superheating field divided by $H_c$ at zero temperatures versus the electron mean free path for niobium.[]{data-label="fig:HshHcVsL"}](./SFoNfig3){width="50.00000%"} Experimental Method =================== To measure niobium’s superheating field, a 1.3 GHz resonating cavity was driven by short (200 $\mu$s) high power (up to 1.5 MW) pulses from a klystron. These pulses raise the fields in the cavity to their maximum values over $\sim100$ $\mu$s. Increasing the fields in this manner prevents the cavity from heating significantly, meaning that the temperature measured at the outer cavity surface is also the temperature of the inner surface. When the magnetic field on the cavity surface reaches the superheating field, $H_{sh}$, the niobium wall undergoes a phase transition into the normal conducting state. Calculation of $Q_0$ as a function of time ------------------------------------------ To accurately measure the superheating field, it is essential to determine precisely when the cavity transitions to the normal conducting state. Previous work has shown that a niobium cavity remains at least 90% superconducting as long as the intrinsic quality factor is greater than $2 {\times 10^{6}}$.[@HaysHPP; @HaysQ] It has been shown how to determine the quality factor as a function of time;[@Farkas84; @Campisi84; @HaysHPP; @HaysQ] we merely reproduce the argument here for completeness. A cavity driven on resonance, $\omega$, by a single input coupler with an incident power, $P_f$, reflects some power, $P_r$, stores energy in the field, $U$, and dissipates some energy in the cavity walls, $P_{d}=\frac{\omega U}{Q_0}$. Conservation of energy gives: $$P_f = P_r + \frac{\omega U}{Q_0} + {\frac{d^{} U}{d t^{}}} \label{eq:ConservationOfEnergy}$$ The reflected power is not a measured quantity, so another expression relating $P_r$ and $P_f$ is needed. A full derivation of the needed equation is presented in Padamsee et. al., Chap. 8;[@HasanEq] in this paper only a plausibility argument will be made, and the result quoted. The reflected power is the superposition of the reflection of the incident power signal, and the power emitted from the cavity through the coupler. The reflection coefficient of the cavity can be expressed in terms of the admittances of the waveguide and the cavity-coupler system. Expressing these admittances in terms of cavity parameters one finds $$\label{eq:ReflectedPower} \sqrt{P_r} = \sqrt{P_f} - \sqrt{P_e}.$$ where $P_e = \frac{\omega U}{Q_{ext}}$ is the power losses through the coupler with an external quality factor of $Q_{ext}$. Using $P_r$ from Eq. \[eq:ReflectedPower\] in Eq. \[eq:ConservationOfEnergy\] yields the expression $$\frac{\omega U}{Q_0} = 2 \sqrt{\frac{\omega U P_f}{Q_{ext}}} - {\frac{d^{} U}{d t^{}}} - \frac{\omega U}{Q_{ext}}$$ The final expression can be obtained by using the identity ${\frac{d^{} \sqrt{U}}{d t^{}}}=\frac{1}{2\sqrt{U}}{\frac{d^{} U}{d t^{}}}$ to yield $$\frac{1}{Q_0} = \frac{2}{\omega \sqrt{U}}\left(\sqrt{\frac{\omega P_f}{Q_{ext}}} - {\frac{d^{} \sqrt{U}}{d t^{}}}\right) - \frac{1}{Q_{ext}} \label{eq:QvsTime}$$ Equation \[eq:QvsTime\] allows one to calculate $Q_0$ as a function of time from measurements of $P_f$ and $U$. Finding the time when the quality factor of the cavity falls bellow $2 {\times 10^{6}}$ pinpoints when the cavity transitions into the normal conducting state.[@HaysHPP] Quench characterization using second sound ------------------------------------------ There are several mechanisms that can initiate a phase transition from the superconducting to the normal conducting state. Defects in the cavity surface can cause heating, leading to thermal breakdown and quench. Field emission, multipacting and dust in the cavity can also be sources of quench due to heating. To ensure that the quench is caused by the cavity making a global transition to the normal conducting state when the superheating field is exceeded, oscillating superleak transducers (OSTs) were used to determine where the observed cavity quenches originate. Energy dissipated from the outer cavity surface couples to a second sound wave in the superfluid Helium surrounding the cavity. The OSTs measure the arrival time of the second sound wave traveling through superfluid Helium, and by using multiple transducers, the origin of the second sound wave can be triangulated.[@Zac] A quench caused by a defect on the surface of the niobium will cause the field energy to be dissipated from a point like source. On the other hand, a transition caused by exceeding the superheating field of the cavity will be global, and the second sound wave should originate from the entire high magnetic field region of the cavity simultaneously. The second sound wave measurement thus allows us to determine whether the transition is caused by a local defect or is initiated by exceeding the superheating field. Cavity details and experimental setup ------------------------------------- The cavity used in the experiment is a 1.3 GHz single-cell re-entrant design.[@VSDesign] It was constructed of high RRR niobium with a bulk RRR of approximately 500, which corresponds to a mean free path of $\sim1800$ nm. The cavity’s history prior to this experiment has been detailed elsewhere.[@NickSRF] Before this experiment, the cavity received a vertical electropolish,[@VEP] which uses a mixture of HF and H$_2$SO$_4$ to remove material from the inner surface. The electropolish removed 5-10 $\mu$m of material while the electrolyte temperature was maintained below 20 $^\circ$C. It was high pressure rinsed for 2 hours, mounted on a test stand and evacuated. While under vacuum, the cavity was surrounded in 110 $^\circ$C air for 48 hours, a process that can prevent the quality factor from deteriorating at high fields.[@Eremeev] Three Cernox temperature sensors were mounted on the outside of the cavity to measure niobium wall temperature, and a Germanium thermometer measured the bath temperature. Eight OSTs were placed roughly at the points of a cube surrounding the cavity. A photograph of the experimental setup is shown in Fig. \[fig:setup\]. The entire setup was placed in a liquid Helium bath, which was used to control the temperature of the cavity. ![(Color online) (a) Experimental set-up of 1.3 GHz re-entrant cavity mounted on a test stand. The copper waveguide behind the cavity connects to the klystron which supplies the high pulsed RF power. Eight OSTs are mounted at corners of a cube around the cavity and are used to detect quench locations. A red box bounds a bottom right OST. (b) Side-view schematic of experimental set-up. OSTs labeled A-D are in near plane, and F-H are in far plane. (c) Front view of an OST. The dark gray membrane is semipermeable, allowing superfluid helium to penetrate, which changes the capacitance of the OST.[]{data-label="fig:setup"}](./SFoNfig4){width="85mm"} Results ======= Continuous wave results ----------------------- The cavity was first tested in continuous wave (CW) mode to measure its properties. Its intrinsic quality factor as a function of accelerating gradient is shown in Fig. \[fig:QvsE\_1p6\]. The cavity reached accelerating electric fields up to 42 MV/m in CW, corresponding to a maximum surface magnetic field of 1474 Oe, and demonstrates a strong decrease in $Q_0$ (i.e. increase in surface resistivity) at high fields. The reduced quality factor at high accelerating gradients is discussed in Sec. \[discussion\]. In spite of the degradation of the quality factor at high fields, these heating losses do not cause significant global heating in pulse mode cavity operation. ![A continuous wave Q vs E curve taken at 1.6 K for cavity LR1-3. The Q degrades as the accelerating gradient is increased.[]{data-label="fig:QvsE_1p6"}](./SFoNfig5){width="50.00000%"} The quality factor was also measured as a function of temperature in CW. This information can be used to determine the surface resistivity of the cavity from the relation $R_s = G/Q_0$, where $G$ is the geometry factor (283.1 $\Omega$ for the cavity tested) and $Q_0$ is the intrinsic quality factor. Plots of these quantities versus temperature are presented in Fig. \[fig:QandRsVsTemp\]. ![Left: Quality factor of the cavity as a function of temperature. The cavity had a quality factor of $1.5{\times 10^{11}}$ at 1.7 K. Right: The Surface resistivity of the cavity as a function of temperature. The residual resistivity of the cavity is $(0.92 \pm 0.23)$ n$\Omega$. Data was taken at an accelerating gradient of 6 MV/m.[]{data-label="fig:QandRsVsTemp"}](./SFoNfig6){width="50.00000%"} The surface resistivity of the cavity has two contributions, a temperature independent residual resistivity, $R_0$, and the BCS surface resistivity: $$R_s(T) = R_0 + A(f,\ell,T)\exp\left(-\frac{\Delta}{k_b T_c}\frac{T_c}{T}\right)$$ where $A$ is a function of the frequency, $f$, mean free path, $\ell$ and temperature, $T$. The residual resistivity of the cavity is found to be $R_0 = (0.92 \pm 0.23)$ n$\Omega$ from the low temperature data. Subtracting the residual resistivity from the surface resistivity leaves the BCS resistivity which can be calculated from material properties. A Fortran code, SRIMP,[@Halbritter] based on the Halbritter definitions, was used to fit the measured BCS surface resistivity by varying the energy gap, $\Delta(0)$ and mean free path, $\ell$, of the material. The parameters used in the data fit are listed in Table \[table:SRIMPParams\]. Parameter Value -------------------------- ------------ Frequency 1294.5 MHz Critical Temperature 8.83 K Coherence Length 640 Å London Penetration Depth 360 Å : \[table:SRIMPParams\] Material properties used in the calculation of BCS resistivity of the niobium. The critical temperature of the niobium is depressed, as discussed in Sec. \[sec:PulsedMeasurements\]. ![The BCS resistivity versus temperature. The blue line is the results of a code, SRIMP, which calculates BCS resistivity from material properties. This fit gives the mean free path estimate of $\ell = 26.91 \pm 1.19$ nm.[]{data-label="fig:BCSvsTemp"}](./SFoNfig7){width="50.00000%"} The resulting fit is displayed in Fig. \[fig:BCSvsTemp\], and was fit with parameters $\Delta(0)/(k_B T_c)=2.14 \pm 0.03$ and $\ell = 26.91 \pm 1.19$ nm. This small mean free path is consistent with results after baking obtained by Ciovati.[@Ciovati] From Eq. \[eq:KappaOfL\] the mean free path corresponds to a Ginsburg-Landau parameter $\kappa = 3.49 \pm 0.16$, showing that the decreased mean free path of the niobium causes it to become a more strongly Type II superconductor. Referencing Figs. \[fig:HshHcVsKappa\] yields $H_{sh}(0)/H_c = 1.044 \pm 0.001$. For the cavity tested the Ginsburg-Landau prediction for the superheating field is then $$H_{sh}(T) = (1.044 \pm 0.001) H_c \left[ 1 - \left(\frac{T}{T_c}\right)^2\right]. \label{eq:GLPred}$$ \[sec:PulsedMeasurements\]Pulsed measurements --------------------------------------------- A klystron was was used to drive the cavity on resonance with high power pulses. The incident power and field in the cavity during two pulses are at different temperatures are shown in Figs. \[fig:Trace2p96K\] and \[fig:Trace7p2\]. The phase transition occurs when the intrinsic quality factor of the niobium reaches $2{\times 10^{6}}$. At low temperatures the superheating field is approximately given by the peak value of the magnetic field. At higher temperatures, however, the field in the cavity is not as sharply peaked, instead showing a broadened structure. In this case it is imperative to have a accurately determine of $Q_0$ as a function of time to pinpoint when the phase transition occurs. ![(Color online) Left: Trace showing the square power wave incident on the cavity (green) and the surface magnetic field of the cavity (blue) taken at 3.0 K. The pulse is sharply peaked around $\sim120\mu$s. Right: Plot of $Q_0$ versus time. The phase change occurs at 118 $\mu$s, which corresponds to $H_{sh} = 1749$ Oe.[]{data-label="fig:Trace2p96K"}](./SFoNfig8){width="50.00000%"} ![(Color online) Left: A trace showing the square power wave incident on the cavity (green) and the surface magnetic field of the cavity (blue) taken at 7.2 K. Right: Plot of $Q_0$ vs time. Compared to the 3.0 K case, the peak is broadened, so accurately determining time of phase transition is essential. The plot shows that $Q = 2{\times 10^{6}}$ at $120.6$ $\mu$s. This corresponds to $H_{sh}= 664$ Oe.[]{data-label="fig:Trace7p2"}](./SFoNfig9){width="50.00000%"} A transition to the normal conducting state can be caused by defects in the material or field emission leading to heating–among other things. To obtain further certainty that the quench is due to reaching the fundamental limit of the superheating field, OST measurements were used to determine the origination of the heating. A quench event as recorded by OSTs is shown in Fig. \[fig:GlobalQuenchEvent\]. All eight OSTs registered a signal a few milliseconds after each quench event. The energy dissipated in the cavity travels through the bulk of the material and at the outer surface of the cavity couples to a second sound wave in the liquid Helium, which is detected by the OST array. Measurements show that the observed triggering signature of the OSTs only occurs when the entire high magnetic field region quenches simultaneously. This shows that the entire high magnetic field portion of the cavity transitioned to the normal conducting state at the same time, characteristic of reaching a fundamental limit as opposed to a local effect. ![OST data for a single quench event. The lowest trace is the field in the cavity, and the three upper traces are 3 of the 8 OST signals, corresponding to sensors A-C as denoted in Fig. \[fig:setup\], all in arbitrary units. The spike at 1 ms is the cavity filling and emptying. The oscillations 3 ms later is the detection of second sound waves by the OSTs. The small time discrepancy between when the OSTs begin to oscillate correspond to the small differences between OST distances from the high magnetic field region of the cavity. Though not shown, all 8 OSTs trigger simultaneously, demonstrating the cavity quench is a global event.[]{data-label="fig:GlobalQuenchEvent"}](./SFoNfig10){width="50.00000%"} Each of the quenches measured below the lambda point of Helium, (Helium must be superfluid for a second sound wave to travel), were found to be global in origin, meaning that each OST measured the source of the second sound wave as originating from the nearest point of the cavities high magnetic field region. This adds credence to the assertion that a fundamental quantity was measured. The pulsed measurements of the superheating field are presented in Fig. \[fig:FinalHshvsT\]. A linear fit was performed on the data, giving the result $$H_{sh}(T) = \left[(1979 \pm 58) - (2151 \pm 130) \left(\frac{T}{T_c}\right)^2\right] \text{Oe} \label{eq:MeasurementFit}$$ where $T_c = 9.22$ K for pure niobium. By definition, the superheating field should vanish at the critical temperature. Our data suggests that the cavity’s critical temperature has been depressed due to impurities in the material. The data gives the sample’s critical temperature as $\tilde{T}_c = 8.83 \pm 0.29$ K, which corresponds to a depressed critical field of $H_c = (1920 \pm 64)$ Oe, since $H_c \propto T_c$, and the clean value of $H_c$ is 2000 Oe. Inserting this value for the depressed critical field into Eq. \[eq:GLPred\] yields the prediction for the superheating field for $\kappa=3.5$ of $$H_{sh}(T) = (2003 \pm 64)\text{ Oe }\left[1-\left(\frac{T}{\tilde{T}_c}\right)^2\right]. \label{eq:FinalPred}$$ This prediction is plotted with the data in Fig. \[fig:FinalHshvsT\] and is consistent with the measured values of the superheating field. ![(Color online) Measurements of the superheating field plotted versus $(T/T_c)^2$, where $T_c=8.83$ K. The green cone is the Ginsburg-Landau prediction for niobium with a mean free path of $26.91\pm1.19$ nm, including the uncertainty in the ratio of $H_{sh}(0)/H_c$.[]{data-label="fig:FinalHshvsT"}](./SFoNfig11){width="50.00000%"} \[discussion\]Discussion ======================== This experiment yielded several remarkable results. First, the surface preparation of vertical electropolishing the cavity produced a niobium cavity with a very high quality factor, measured at 1.7 K, of $1.5 \times 10^{11}$ at 6 MV/m. This corresponds to a residual resistivity of $(0.92 \pm 0.23)$ n$\Omega$ which is among the smallest measured. The BCS resistivity of the cavity is also low, with values of 1.95 n$\Omega$ at 2.0 K ($Q_0 = 5.8{\times 10^{10}}$) and $3.16{\times 10^{2}}$ n$\Omega$ at 4.2 K ($Q_0 = 8.9{\times 10^{8}}$). Niobium has minimal BCS resistivity for mean free paths between 10-50 nm.[@Kneisel] The obtained mean free path of 26 nm, is within 5% of the minimum value, showing that the 100-120 $^\circ$C heating process is necessary for achieving very low surface resistivities. Reliably producing cavities at this performance level is essential for next generation CW superconducting accelerators such as Project X,[@ProjX] and linac driven light sources such as Cornell’s ERL.[@ERL] For CW machines, very small surface resistivities are more important than highest accelerating gradients because of the large cryogenic load in continuous cavity operation. As shown in Fig. \[fig:QvsE\_1p6\], the quality factor of the cavity significantly degraded as the accelerating gradient increased. Generally, the 110 $^\circ$C heat treatment procedure is known to reduce the high field Q-slope. However, previous work has shown that the effectiveness of this procedure varies, even for electropolished cavities.[@Gigi2007] The high field Q-slope is emphasized by the fact that the cavity has a very high initial $Q_0$ at low fields, and reaches gradients above 40 MV/m. High RRR materials are commonly used in superconducting RF cavities to improve thermal conduction. This experiment used a cavity with a bulk RRR of $\sim500$. However the RF properties of the cavity are determined by the properties of the surface layers with a thickness of $\sim50$ nm. The 110 $^\circ$C heating process reduced the mean free path of the surface layer to about 26 nm, corresponding to a RRR of only $\sim4.5$. Baking a cavity at $110-125$ $^\circ$C is currently part of routine cavity preparation, since it is the only known method to minimize the undesirable high field $Q$-slope at high accelerating gradients. However, by reducing the mean free path, this process has the undesirable effect of increasing the Ginsburg-Landau parameter, thereby significantly reducing $H_{sh}$ by as much as 25% (see Fig. \[fig:HshHcVsL\]). As the data near $T_c$ shows in Fig \[fig:HaysData\], a slope of $1.2H_c$ is possible if the niobium does not recieve a low temperature heat treatment. This is a 20% increase from value measured for the 110 $^\circ$C heat treated cavity in this study. Thus, for applications requiring high quality factor cavities at ultimate accelerating gradients, it is necessary to find a process other than baking that removes the high field Q-slope without reducing the mean free path of the electrons. OST data verified that the quenches were globally initiated, consistent with reaching the superheating field. The quenches were not locally initiated, eliminating point defect heating, field emission or contamination from limiting the maximum magnetic fields. The experimental data for the superheating field of niobium agrees with the Ginsburg-Landau prediction of $H_{sh}$ being linear in $(T/T_c)^2$ down to 1.7 K within the $\pm10\%$ measurement uncertainty. This is a result that has never before been obtained. Moreover, this theory correctly predicts an accurate slope for the $(T/T_c)^2$ dependence for niobium once the reduction in the mean free path in the RF surface layer by the cavity preparation has been taken into account. This evidence, along with the fact that superheating fields as large as $\sim$1900 Oe have been measured, excludes Saito’s theory from modeling the superheating field. Conclusions =========== The superheating field of niobium was shown to be linear versus $(T/T_c)^2$, in accordance with Ginsburg-Landau theory, down to a temperature of 1.7 K. State of the art surface treatments have resulted in a cavity with a high quality factor and very low residual resistivity. The low temperature baking process has been shown to adversely effect the superheating field of niobium by decreasing the mean free path of the RF surface layer, which is an impediment to obtaining the ultimate achievable accelerating gradients in niobium cavities for next generation particle accelerators. Acknowledgments =============== The authors thank J. Sethna and M. Transtrum for their theoretical calculations of $H_{sh}(0)/H_c$ versus the Ginsburg-Landau parameter, Z. Conway for help with OST instrumentation and vertical electropolishing the cavity and S. Posen for assistance during the cavity test. They also thank the Alfred P. Sloan Foundation and the Department of Energy under contract DE-SC00002329 for funding this research.

ArXiv

--- abstract: 'We consider a thin film consisting of two layers of immiscible liquids on a solid horizontal (heated) substrate. Both, the free liquid-liquid and the liquid-gas interface of such a bilayer liquid film may be unstable due to effective molecular interactions relevant for ultrathin layers below 100nm thickness, or due to temperature-gradient caused Marangoni flows in the heated case. Using a long wave approximation we derive coupled evolution equations for the interface profiles for the general non-isothermal situation allowing for slip at the substrate. Linear and nonlinear analyses of the short- and long-time film evolution are performed for isothermal ultrathin layers taking into account destabilizing long-range and stabilizing short-range molecular interactions. It is shown that the initial instability can be of a varicose, zigzag or mixed type. However, in the nonlinear stage of the evolution the mode type and therefore the pattern morphology can change via switching between two different branches of stationary solutions or via coarsening along a single branch.' author: - Andrey Pototsky - Michael Bestehorn - Domnic Merkt - Uwe Thiele title: 'Morphology changes in the evolution of liquid two-layer films' --- Introduction. {#intro} ============= Instabilities of thin liquid films between a solid substrate and a gas atmosphere have attracted much scientific interest. The main focus lies thereby on front instabilities of moving contact lines [@CHTC90; @BrGe93; @BMFC98] or on instabilities of the free liquid-gas interface of a flat film [@Reit92; @VanH95; @TMP98]. A recent review can be found in Ref.. To analyze such instabilities a long-wave or lubrication approximation [@Reyn1886; @Somm1904; @ODB97] is often used as a very powerful tool especially for low Reynolds number film flows. At present the basic behavior of one-layer films in the physically different thickness ranges is well understood. Several instability mechanisms exist that by means of different driving forces may destabilize an initially flat film. They are described, analysed and modelled in a large number of experimental [@Reit92; @ShRe96; @Reit93a; @JSSH98; @VanH95; @VanH97; @SHJ01] and theoretical [@RuJa73; @WiDa82; @deGe85; @YiHi89; @OrRo92; @Shar93; @Shar93b; @Mitl93; @ODB97; @ShKh98; @TVN01; @ScSt02; @TNPV02; @Thie03; @BGW01; @BPT03; @ThKn04; @PBT04] works. For film thicknesses $d$ less than about $100$nm, effective molecular interactions between the film surface and the substrate dominate all the other forces, like thermo- and soluto-capillarity or gravity, and thus determine the film stability. For heated films of thicknesses above $100$nm, eventually thermocapillary forces become the most important influence leading to an instability caused by large-scale Marangoni convection [@VanH95; @BPT03]. It is dominant up to an upper limit of the film thickness determined by the competition between large-scale and small-scale convection modes [@GNP94]. For even thicker films with thicknesses above $100\,\mu$m also the gravity force becomes important. Depending on its direction it may stabilize the large-scale Marangoni instability or destabilize the film further (Rayleigh-Taylor instability) [@YiHi89; @OrRo92]. The lubrication approximation is valid up to a limiting film thickness obtained by the requirement that the wave length of the dominant instability mode $\lambda_m$ is much larger than the film thickness $d$, i.e. $\lambda_m\gg d$. For the Rayleigh-Taylor instability $\lambda_m$ depends on the interfacial tension, the density of the liquid and the gravitational acceleration but not on the film thickness [@ODB97]. It is of the order of $10^{3}$ to $10^{4}\,\mu$m implying an upper limit for the film thickness of $10^2$ to $10^{3}\,\mu$m. The stability and evolution of liquid films of thicknesses below $100$nm is determined by effectiv molecular interactions between substrate and film arising, for instance, from Van der Waals, electrostatic or entropic interactions [@Isra92; @Hunt92]. Such films are lineary unstable if the energy of the intermolecular interaction is a convex function of the film thickness. For film thicknesses above $10$nm the long-range Van der Waals forces dominate. They can be of different nature depending on the molecular properties of the involved media. One distinguishes interactions between two randomly orienting dipoles (orientation interaction), between a randomly orienting dipole and an induced dipole (induction interaction), and between a fluctuating dipole and an induced dipole (dispersion interaction). Between two parallel interfaces at a distance $d$, all these forces decay as $A/d^3$ where $A$ is the Hamaker constant [@Isra92]. An unstable situation corresponds to a positive Hamaker constant. Note, however that different schools use different sign conventions. The dominant wave length of the instability $\lambda_m$ increases monotonically with $d$ as $\lambda \sim d^2$ (see Ref.). The stability of a film may change dramatically for a substrate coated with a layer of different dielectric properties as, for instance, a silicon substrate (Si) coated with an silicon oxide layer (SiO) [@SHJ01c]. There, for an oxide layer of about 2nm only ultrathin polystyrene (PS) films below 4nm thickness are linearly unstable. Increasing the film thickness in the linearly unstable range, the wave length $\lambda$ increases rapidly and diverges at the critical thickness $d_c$. For $d>d_c$ the film is linearly stable, but may rupture due to finite disturbances. Imagine one replaces the (solid) coating layer by a liquid layer transforming the system in a two-layer liquid film. Some of the results obtained for a solid coating can be directly transferred to the new situation. The stability of the (now) upper layer still depends on the (now liquid) coating layer. However, additionally the liquid coating layer itself may be unstable making a re-evaluation of the stability necessary. This thought experiment leads quite naturally to the extension of the well studied one-layer systems to two-layer systems that are the subject of the present work. In general, there exist two possible two-layer geometries. On the one hand the two liquid layers can be situated between two solid plates leaving only the interface between the two liquids free to move. In consequence such a system can be described by a single evolution equation [@Lin01; @Lin02; @MPBT05]. On the other hand the two layers can be situated between a solid substrate and a gas atmosphere. Then both, the liquid-liquid and the liquid-gas interface are free to move and their evolution has to be described by coupled evolution equations. Models were derived, for instance, assuming a lower liquid layer that is much thicker than the upper layer [@BMR93], and for two-layer systems with surfactants (and non-newtonian behaviour) [@ZMC03; @CrMa00; @MCW02] or including evaporation [@Dano98; @Dano98b; @Paun98]. A two-layer system under the solely influence of molecular interactions is studied in Ref.. In Ref. a similar system is studied, however, the evolution equations are given in terms of variations of an energy functional. The experimental interest in two-layer liquid films is up to now mainly focused on the dewetting of a liquid layer from a very thick layer, i.e. a liquid bulk substrate [@Wunn03; @MSS04; @FCW95; @LPHK96; @PWHC97]. In contrast Ref. studies the dewetting of a polystyrene (PS) layer of $15$ to $68$nm thickness from a $46$nm thick polyamide (PA) layer. The substrate is a silicon (Si) wafer covered with a layer of native oxide. At high temperature ($195^o$C) and small thicknesses ($15...35$nm) the PS layer is unstable and dewets exhibiting typical spinodal patterns. At low temperature ($115^o$C) the PA layer is solid resulting in a stable PS layer, independent of its thickness. Ref. studies relatively thick layers ($100$nm to $1\,\mu$m thickness) of poly(dimethylsiloxane) (PDMS) layers on a liquid substrate of fluorinated PDMS. They show that the PDMS films are metastable and may dewet by nucleation of holes. The velocity of the growth of holes depends on the viscosity and thickness of the substrate. In another system (PS layer on poly(methylmethacrylate) (PMMA) layer, both with thicknesses of about $100$nm) the dewetting velocity was found to exhibit a minimum as a function of the viscosity of the lower layer [@LPHK96]. Furthermore, for a polycarbonate (PC) layer on a poly(styrene-$co$-acrylonitrile) (SAN) layer Ref. reports a non-trivial dependence of the dewetting velocity on both layer thicknesses. Recently we presented coupled evolution equations for two-layer liquid films taking into account long-range Van der Waals interactions only [@PBMT04]. This allows to study pathways towards rupture but can not describe the long-time evolution of such films as, for instance, necessary for the description of the above mentioned experiments. In the present paper we present the derivation of the system of evolution equations for a general interaction energy and a non-isothermal situation and then study the short- and long-time evolution of two-layer liquid films incorporating long-range and short-range interactions. Thereby the main focus lies on ultrathin layers with respective thicknesses below $100$nm for which the effective molecular interactions between the four media are the dominant influence. The paper is structured as follows. The problem is formulated in Section \[deriv\] followed by the derivation of the relevant long wave evolution equations for a general non-isothermal case allowing also for slip at the substrate. Focusing on the isothermal case without slip these equations are analyzed and integrated numerically in the subsequent sections. Section \[lin\_stab\] presents the linear stability analysis for flat films starting from the general case in Section \[gen\_stab\], focusing on long-range van der Waals interactions in Section \[lr\_int\], discussing the possible mode types in Section \[inst\_modes\] and introducing important limiting cases in Section \[lim\_cases\]. In Section \[nuss\] we investigate the non-linear behaviour discussing in Section \[nu\_sol\] non-uniform stationary solutions as extrema of the Lyapunov functional, presenting in Section \[mt\_trans\] two different pathways for mode-changes in the course of the time evolution, and discussing long-time stationary solutions for different experimentally relevant systems in Section \[lp\_sol\]. Concluding remarks follow in Section \[conc\]. Expressions for the surface tension gradients in terms of gradients of the layer thicknesses are discussed in an Appendix. Derivation of coupled large-amplitude evolution equations ========================================================= General case {#deriv} ------------ First, we derive coupled evolution equations for the profiles of the liquid-liquid interface $h_1(x,y)$ and the liquid-gas interface $h_2(x,y)$. We assume that the layers are thin enough that convective terms can be neglected. Considering a two-dimensional geometry as sketched in Fig.\[fig1\] the respective Stokes equations for the two layers are $$\nabla (p_i + \phi_i) = \mu_i \Delta \vec{v}_i, \label{stokes}$$ where $i=1,2$ denotes the respective layer. For each layer $\vec{v_i} = (u_i,w_i)$ is the velocity field, $p_i$ the pressure, $\phi_i$ the potential of the bulk forces and $\mu_i$ the viscosity. The constant mean film thicknesses are denoted by $d_1 = (\int_0^L h_1 dx)/L$ and $d_2 = (\int_0^L h_2 dx)/L$ where $L$ is the lateral system size. A lubrication approximation [@ODB97] is applied assuming the ratio of vertical and horizontal length scales to be small. As smallness parameter we use the ratio $\epsilon = d_1 / \lambda$ where $\lambda$ is the characteristic lateral length scale of the film instability. In zeroth order in $\epsilon$ the Stokes equations (\[stokes\]) simplify to $$\begin{aligned} \mu_2 \partial_z^2 u_2 = \partial_x \bar{p}_2 \label{NSa} \\ \partial_z \bar{p}_2 = 0 \label{NSb} \\ \mu_1 \partial_z^2 u_1 = \partial_x \bar{p}_1 \label{NSc} \\ \partial_z \bar{p}_1 = 0, \label{NSd}\end{aligned}$$ where the $\bar{p}_i$ stand for $p_i + \phi_i$. At the substrate $(z=0)$ we use a Navier slip and a no-penetration condition, i.e.$$\begin{aligned} u_1 = \beta \partial_z u_1~~\text{and}~~w_1 = 0, \label{SLIP}\end{aligned}$$ respectively. The slip length is denoted by $\beta$. At the liquid-liquid interface $(z = h_1)$ we use the continuity of the velocity field, the kinematic condition and the continuity of the tangential component of the liquid stress tensor $$\begin{aligned} u_1 = u_2,~w_1 = w_2, \label{BCVa} \\ w_1 = \partial_t h_1 +u_1 \partial_x h_1, \label{BCVb} \\\end{aligned}$$ and $$\begin{aligned} \mu_1 \partial_z u_1 - \mu_2 \partial_z u_2 = \partial_x \sigma_{12}, \label{BCVc}\end{aligned}$$ respectively. The normal stress condition is discussed below. At the liquid-gas interface $(z = h_2)$ only the kinematic condition and the continuity of the tangential component of the liquid stress tensor apply, i.e.$$\begin{aligned} w_2 = \partial_t h_2 +u_2 \partial_x h_2, \label{BCV1a} \\ \mu_2 \partial_z u_2 = \partial_x \sigma_{2}. \label{BCV1b}\end{aligned}$$ The $\sigma_{12}$ and $\sigma_2$ stand for the interfacial tensions of the liquid-liquid and of the liquid-gas interface, respectively. The boundary conditions for the normal component of the stress tensor are written incorporating the disjoining pressures at the liquid-liquid $\Pi_1(h_1,h_2)$ and at the liquid-gas $\Pi_2(h_1,h_2)$ interface, respectively. They represent effective molecular interactions between the interfaces that result, for instance, from Van der Waals interactions [@Isra92]. They are discussed in detail below. For the liquid-gas interface $(z=h_2)$ we obtain $$\begin{aligned} p_2(h_2) - p_0 = - \sigma_2 \partial_x^2 h_2 + \Pi_2(h_1,h_2) \nonumber \end{aligned}$$ and for the liquid-liquid interface $(z=h_1)$ $$\begin{aligned} p_1(h_1) - p_2(h_1) = - \sigma_{12} \partial_x^2 h_1 + \Pi_1(h_1,h_2), \label{BCP}\end{aligned}$$ where $p_0$ is the constant pressure in the gas atmosphere. Eqs.(\[BCP\]) can be written in terms of variations of an energy functional $F[h_1,h_2]$ $$\begin{aligned} p_1(h_1)-p_2(h_1) = \frac{\delta F}{\delta h_1} \nonumber \\ p_2(h_2)-p_0 = \frac{\delta F}{\delta h_2}, \label{BCP_ENRG}\end{aligned}$$ with $$F = \int \left(\sigma_{1} \frac{(\partial_x h_1)^2}{2} + \sigma_2 \frac{(\partial_x h_2)^2}{2} +f(h_1,h_2) \right)dx, \label{LYAP}$$ and $f(h_1,h_2)$ being the free energy of the flat films per unit area. Equations (\[NSa\]) and (\[NSc\]) are integrated three times with respect to $z$ to obtain the stream functions $\Psi_i$, defined by $(w_i = - \partial_x \Psi_i,~u_i = \partial_z \Psi_i)$. The six $x$-dependent integration constants are determined using the boundary conditions (\[SLIP\]), (\[BCVa\]), (\[BCVc\]), and (\[BCV1b\]). Thus the velocity fields in the two layers are given by $$\begin{aligned} u_1 &=& \frac{1}{\mu_1} \left( \partial_x \bar{p}_1\right) \frac{z^2}{2}+ \frac{1}{\mu_1}(z+\beta) K_1 \nonumber \\ u_2 &=& \frac{1}{\mu_2} \left( \partial_x \bar{p}_2\right) \frac{z^2}{2}+ \frac{1}{\mu_2} K_2 (z-h_1)- \frac{\partial_x \bar{p}_2}{\mu_2} \frac{h_1^2}{2} + u_1 (h_1), \label{velocities}\end{aligned}$$ with $K_1 = K_2 + \partial_x \sigma_{12} + \left[(\partial_x \bar{p}_2)- (\partial_x \bar{p}_1)\right] h_1$ and $K_2 = \partial_x \sigma_{2} - (\partial_x \bar{p}_2)h_2$. The stream functions $\Psi_i$ are related to the flow in the lower layer $\Gamma_1 = \int_{0}^{h_1} u_1\,dz$ and to the one in the upper layer $\Gamma_2 = \int_{h_1}^{h_2} u_2\,dz$ by $\Gamma_1 = \Psi_1(h_1),~\Gamma_1 + \Gamma_2 = \Psi_2(h_2)$. Using the $\Psi_i$ we rewrite Eqs.(\[BCVb\]) and (\[BCV1a\]) to obtain the evolution equations for the two interface profiles $$\begin{aligned} \partial_t h_1 +\partial_x \left[ \Psi_1 (h_1) \right] &=& 0, \label{EVOLH1} \\ \partial_t h_2 +\partial_x \left[ \Psi_2 (h_2) \right] &=& 0. \label{EVOLH2}\end{aligned}$$ Written in terms of the energy functional they read $$\begin{aligned} \partial_t h_1 &=& \partial_x \left[ Q_{11}\partial_x \frac{\delta F}{\delta h_1} + Q_{12}\partial_x \frac{\delta F}{\delta h_2} - D_{11} \partial_x \sigma_{12} - D_{12} \partial_x \sigma_{2} \right] \nonumber \\ \partial_t h_2 &=& \partial_x \left[ Q_{21}\partial_x \frac{\delta F}{\delta h_1} + Q_{22}\partial_x \frac{\delta F}{\delta h_2} - D_{21} \partial_x \sigma_{12} - D_{22} \partial_x \sigma_{2} \right], \label{EVOL_PHYS}\end{aligned}$$ with the mobility matrices of the pressure terms $$\begin{aligned} {\bm Q} = \frac{1}{\mu_1} \left( \begin{array}{ccc} \frac{h_1^3}{3}+ \beta h_1^2 &~& \frac{h_1^2}{2}(h_2 -\frac{h_1}{3}) + \beta h_1 h_2\\ \frac{h_1^2}{2}(h_2 -\frac{h_1}{3}) + \beta h_1 h_2 &~& \frac{(h_2-h_1)^3}{3}(\frac{\mu_1}{\mu_2}-1) + \frac{h_2^3}{3} +\beta h_2^2 \end{array} \right) \label{MOB}\end{aligned}$$ and of the tangential stress terms $$\begin{aligned} {\bm D} = \frac{1}{\mu_1} \left( \begin{array}{ccc} \frac{h_1^2}{2}+ \beta h_1 &~& \frac{h_1^2}{2} + \beta h_1 \\ h_1(h_2 -\frac{h_1}{2}) + \beta h_2 &~& \frac{\mu_1 (h_2-h_1)^2}{2 \mu_2} + h_1(h_2 -\frac{h_1}{2}) +\beta h_2 \end{array} \right), \label{TANG}\end{aligned}$$ respectively. Note, that the mobility matrix ${\bm Q}$ is symmetric and all mobilities $Q_{ik}$ and $D_{ik}$ are positive. Dropping the terms representing the effective molecular interactions, Eqs.(\[EVOL\_PHYS\]) represent the fully nonlinear equivalent for the weakly nonlinear equations derived in Ref.. Assuming that the interfacial tensions are influenced by thermocapillarity only, one can express the derivatives $\partial_x \sigma_{12}$ and $\partial_x \sigma_{2}$ in terms of gradients of local thicknesses $\partial_x h_i$. This is done in the Appendix. For isothermal ultrathin liquid films one has $(\partial_x \sigma_{12} = \partial_x \sigma_{2} =0)$. The situation is then relaxational (or variational), i.e. Eqs.(\[EVOL\_PHYS\]) possess a Lyapunov functional, namely the energy functional $F$, which decreases monotonously in time as shown next. The total time derivative of the Lyapunov functional is $dF/dt = \int \left( \frac{\delta F}{\delta h_1} \partial_ t h_1 + \frac{\delta F}{\delta h_2} \partial_ t h_2 \right) \,dx $. Expressing $\partial_t h_i$ by Eqs.(\[EVOL\_PHYS\]) and using partial integration with periodic boundary conditions, one obtains $$\frac{dF}{dt} = - \int \sum_{i,k} Q_{ik} \left( \partial_x \frac{\delta F}{\delta h_i} \right) \left(\partial_x \frac{\delta F}{\delta h_k} \right)\,dx. \label{LYAP_dt}$$ Because $$\det {\bm Q} = \frac{(h_2-h_1)^3 h_1^3}{9 \mu_1 \mu_2} +\frac{1}{12 \mu_1^2} h_1^4 (h_2-h_1)^2 + \beta \left( \frac{h_1^3}{3 \mu_1^2}(h_2-h_1)^2 + h_1^2 \frac{(h_2-h_1)^3}{3 \mu_1 \mu_2} \right) > 0 \label{DET}$$ and $Q_{ii} > 0$, the quadratic form in Eq.(\[LYAP\_dt\]) is positive definite implying $dF/dt < 0$. The existence of $F$ allows to identify the stationary solutions of Eqs.\[EVOL\_PHYS\] with the extrema of $F$. This will be used below in Section\[nu\_sol\]. The disjoining pressures $\Pi_i(h_1,h_2)$ {#driv_for} ----------------------------------------- In many important cases, as for instance, for polymer films on apolar substrates [@SHJ01; @Reit92], the interaction energy is mainly determined by its long-range apolar dispersion part. However, if the model only takes into account a destabilizing long-range interaction the time evolution definitively leads to rupture of the upper or lower layer [@PBMT04] making it impossible to study the long-time coarsening behaviour. To be able to study the long-time evolution one has to include stabilizing short-range interactions [@Isra92; @Shar93] into the model. Although these are normally not included for films of thicknesses above $10$nm because they do not change the stability of flat films, also for such films they become important in the non-linear stage of evolution when the local thicknesses become comparable to their interaction length. The long-range part of the interaction energy for each pair of interfaces (see Fig.\[fig1\]) resulting from dispersive Van der Waals interactions is given by $A_{ijkl}/2 \pi h^2$ (see Ref.), where $A_{ijkl}$ is a (four-index) Hamaker constant which corresponds to the interaction between the interfaces $i-j$ and $k-l$. Each index in $A_{ijkl}$ can be one out of $g,1,2$ and $s$, denoting gas, liquid $1$, liquid $2$ and substrate, respectively. The four-index Hamaker constant is calculated using an equivalent of Eq.(11.13) of Ref. that is based on the assumption that the main absorption frequencies of all involved media are about $\nu_e = 3\times10^{15}$Hz and that the zero frequency contribution is negligible. One uses $$A_{ijkl}\,\approx\,\frac{3h\nu_e}{8\sqrt{2}}\, \frac{(n_i^2-n_j^2)(n_l^2-n_k^2)}{(n_i^2+n_j^2)^{1/2}(n_l^2+n_k^2)^{1/2} [(n_i^2+n_j^2)^{1/2}+(n_l^2+n_k^2)^{1/2}]}, \label{HAM}$$ where $n_i$ are the refractive indices of the media. The three-index Hamaker constants are given by $$A_{ijk} = A_{ijjk}. \label{HAMb}$$ The short-range forces which can be of an electrostatic or structural nature [@TDS88; @deGe85] decay exponentially with $h$. The electrostatic part results from the formation of diffuse electric double-layers in the vicinity of interfaces involving polar liquids [@DeLa41; @VeOv48; @Ohsh74; @Ohsh74b]. For films with thicknesses in the range of the Debye length the diffuse double-layers at the two interfaces overlap resulting in a repulsive or attractive force between the interfaces. The corresponding interaction energy between the interfaces $s-1$ and $1-2$ is given by $S_1 \exp{\left[ (l_0-h_1)/l_1\right]}$ and between the interfaces $1-2$ and $2-g$ by $S_2 \exp{\left[ (l_0-(h_2-h_1))/l_2\right]}$, where $l_0 = 0.158$nm is the Born repulsion length, and $l_1,~l_2 \sim 1..10$nm are the interaction lengths of the short-range interactions [@Shar93]. Further on we consider the two correlation lengths $l_1$ and $l_2$ to be equal and denote them by $l$. $S_1>0$ and $S_2>0$ are the short-range components of the total spreading coefficients. They are related to the lower layer on the substrate below a bulk of the upper liquid and to the upper layer on the lower film as substrate below the ambient gas, respectively. We do not take into account short-range interactions between interfaces $s-1$ and $2-g$. Collecting the long-range and the short-range forces the disjoining pressures $\Pi_1(h_1,h_2)$ and $\Pi_2(h_1,h_2)$ are specified as $$\begin{aligned} \Pi_1(h_1,h_2)&=&\frac{A_{21s}}{6 \pi h_1^3} - \frac{A_{12g}}{6 \pi (h_2-h_1)^3}- \frac{S_1}{l_1} \exp{\left[\frac{l_0 - h_1}{l_1}\right]}+ \frac{S_2}{l_2} \exp{\left[\frac{l_0 - (h_2-h_1)}{l_2}\right]} \nonumber \\ \Pi_2(h_1,h_2)&=&\frac{A_{12g}}{6 \pi (h_2 -h_1)^3} +\frac{A_{g21s}}{6 \pi h_2^3}- \frac{S_2}{l_2} \exp{\left[\frac{l_0 - (h_2-h_1)}{l_2}\right]} \label{DSJPRESS} \end{aligned}$$ To non-dimensionalize Eqs.(\[EVOL\_PHYS\]) we scale the thicknesses with $l$, the lateral coordinate $x$ with $\lambda = l (d_2-d_1) \sqrt{2 \pi \sigma_1 /|A_{12g}|}$, and time $t$ with $\tau = (2 \pi)^2 \sigma_1 \mu_1 l (d_2-d_1)^4 / A_{12g}^2$. Then the scaled energy functional $$\begin{aligned} F &=& \int\,\left[ \frac{1}{2}(\partial_x h_1)^2 +\frac{\sigma}{2} (\partial_x h_2)^2 -\frac{\bar{A}_{12g}}{6 (h_2-h_1)^2}- \frac{\bar{A}_{21s}}{6 h_1^2} -\frac{\bar{A}_{g21s}}{6 h_2^2}\right. \nonumber \\ &+& \left. c_1(h_1-\bar{d}_1) +c_2(h_2-\bar{d}_2) +\bar{S}_1 \exp{(- h_1)}+ \bar{S}_2 \exp{(h_1-h_2)}\right]\,dx, \label{ENRG_SCL}\end{aligned}$$ involves the scaled Hamaker constants $\bar{A}_{ijkl} = \left[(d_2-d_1)/l\right]^2 A_{ijkl}/|A_{12g}|$, spreading coefficients $\bar{S}_i = 2 \pi \left[(d_2-d_1)\right]^2 S_i \exp{(d_0/l)} / |A_{12g}|$, and mean layer thicknesses $\bar{d}_i = d_i/l$. The $c_i$ are Lagrange multipliers that ensure mass conservation for the two liquids. The corresponding energy scale is $|A_{12g}|/2\pi(d_2-d_1)^2$ and the ratios of the mean layer thicknesses, surface tensions and viscosities are $d = d_2/d_1$, $\sigma = \sigma_2 / \sigma_1$ and $\mu=\mu_2/\mu_1$, respectively. Further on, we denote the scaled variables using the same symbols as before, i.e.the scaled mean thicknesses are given by $d_i$ and the local thicknesses by $h_i$. The non-dimensional mobility matrices are obtained from Eqs.(\[MOB\]) and (\[TANG\]) by dropping the factor $1/ \mu_1$ and replacing $\beta$ by $\beta / l$. Linear stability {#lin_stab} ================ General stability of flat films {#gen_stab} ------------------------------- We start the analysis of our model for two-layer films by discussing the linear stability of flat films with $h_1 (x) = d_1$ and $h_2 (x) = d_2$. Eqs.(\[EVOL\_PHYS\]) are linearized in $\epsilon\ll1$ for small amplitude disturbances $\epsilon\chi_i \exp{(\gamma t)} \exp(kx)$ for $i=1,2$ where $k$, $\gamma$ and ${\bm \chi}=(\chi,1)$ are the wave number, growth rate and amplitudes of the disturbance, respectively. The dispersion relation $\gamma(k)$ is obtained by solving the resulting eigenvalue problem $({\bm J}-\gamma {\bm I}) {\bm \chi} = 0$. For the isothermal case ($\partial_x \sigma_{12} =\partial_x \sigma_{2}=0$ in Eqs.(\[EVOL\_PHYS\])), the corresponding $non$-symmetric Jacobi matrix ${\bm J}$ is given by ${\bm J} = -k^2 {\bm Q}\cdot{\bm E}$, where ${\bm Q}$ is the scaled mobility matrix. ${\bm E}$ is the energy matrix $${\bm E} = \left( \begin{array}{cc} \frac{\partial^2 f}{\partial h_1^2} + k^2& \frac{\partial^2 f}{\partial h_1 \partial h_2} \\ \frac{\partial^2 f}{\partial h_1 \partial h_2} & \frac{\partial^2 f}{\partial h_2^2} +\sigma k^2 \end{array} \right), \label{ENRG_MATRIX}$$ where $f(h_1,h_2)$ is the local part of the energy density from Eq.(\[ENRG\_SCL\]). This yields $$\gamma = \frac{\text{Tr}}{2} \pm \sqrt{\frac{\text{Tr}^2}{4} -{\rm Det}}, \label{GR_RATE}$$ where ${\rm Tr} = -k^2 [2 Q_{12} E_{12} + Q_{11} E_{11} + Q_{22} E_{22}]$ and ${\rm Det} = k^4 \det {\bm Q} \det {\bm E}$ are the trace and the determinant of ${\bm J}$. Since $\det {\bm Q} \neq 0$ the eigenvalue problem can be written as the generalized eigenvalue problem $(k^2 {\bm E} + \gamma {\bm Q^{-1}}) {\bm \chi} = 0$. Because ${\bm E}$ and ${\bm Q^{-1}}$ are both symmetric and ${\bm Q^{-1}}$ is positive definite one can deduce that all eigenvalues $\gamma$ are real [@MSV87] as expected for a variational problem. In the non-isothermal case, the Jacobi matrix is given by ${\bm J} = -k^2({\bm Q}\cdot{\bm E} -{\bm D}\cdot{\bm \Gamma})$, where ${\bm \Gamma}$ is a scaled matrix of coefficients for the Marangoni terms. It is defined in Section \[appendix\]. Neither the matrix ${\bm D}$ nor ${\bm \Gamma}$ are symmetric. This leads to (in general) complex eigenvalues indicating the possibility of oscillating motion in the non-isothermal case [@NeSi90; @NeSi97]. Going back to the isothermal case, inspection of the generalized eigenvalue problem shows that the stability threshold is completely determined by the eigenvalues of ${\bm E}$. Since the surface tension terms are always positive, the onset of the instability is always found for $k=0$, i.e. the system is linearly stable, independently of the wavelength of the disturbance, for $$\begin{aligned} \det {\bm E} > 0 \quad\text{and}\quad E_{11}> 0 \quad\text{at}\quad k=0. \label{STAB}\end{aligned}$$ An instability sets in if at least one of the conditions (\[STAB\]) is violated. Then the flat two-layer film is unstable to disturbances with $k$ larger zero and smaller than a cutoff wavenumber $$k_c^2 = -\frac{1}{2} \left( \frac{\partial^2 f}{\partial h_1^2} +\frac{1}{\sigma}\frac{\partial^2 f}{\partial h_2^2 } \right) \pm \sqrt{\frac{1}{4} \left(\frac{\partial^2 f}{\partial h_1^2} -\frac{1}{\sigma}\frac{\partial^2 f}{\partial h_2^2 } \right)^2 + \frac{1}{\sigma}\left( \frac{\partial^2 f}{\partial h_1 \partial h_2} \right)^2} \label{CUT_OFF}$$ determined by the condition $\det {\bm E}(k_c) = 0$. Fig.\[fig2\] shows a schematic stability diagram in the plane $(E_{11},E_{22})$. The stability threshold $E_{11} E_{22} = E_{12}^2,~E_{11} >0$ is a hyperbola, represented by the solid line. The unstable region below and left of that line is divided by a second hyperbola into a two-mode and a one-mode region. In the two-mode region both growth rates given by Eq.(\[GR\_RATE\]) are positive for $k$ smaller then the respective cut-off $k_c$. In the one-mode region only one $\gamma$ is positive for $k<k_c$. Fixing all other system parameters, $\det {\bm E}(k)$ is determined by $k$. If at $k=0$ the system is in the two-mode region, then by increasing $k$ one passes two times a line $\det {\bm E} = 0$, as indicated by the dashed arrow in Fig.\[fig2\]. At each crossing a growth rates (Eq.(\[GR\_RATE\])) becomes negative, i.e. a mode is stabilized. If at $k=0$ the system is in the one-mode region the line $\det {\bm E} = 0$ is crossed only once (dot-dashed line). Long-range interaction only {#lr_int} --------------------------- As detailed above the stabilizing short-range interaction only becomes important if at least one layer thickness is locally comparable to the interaction length $l$, i.e. if a layer becomes thinner than about $10$nm. Therefore, to study the linear stability of thicker layers one can neglect the short-range terms in Eq.(\[ENRG\_SCL\]). In this case, Eq.(\[STAB\]) is used to study the role of the Hamaker constants Eq.(\[HAM\]) in the linear evolution of the system. First we note that the Hamaker constants are coupled through the refractive indices of the media $n_i$. This allows only for selected combinations of signs of the $A_{ijk}$ and $A_{ijkl}$ as given in Table \[Ham\_Table\]. For fixed Hamaker constants, i.e. fixed combination of materials, $\det {\bm E}_0 = \det {\bm E} (k=0)$ is a function of the ratio $d$ of the layer thicknesses only. Using Table \[Ham\_Table\] one can show that for positive $\partial^2 f / \partial h_i^2$ the equation $\det {\bm E}_0 (d) = 0$ can only have the solution $d=1$, i.e. $d_2-d_1=0$. This means that only for vanishing upper layer the system can be on the stability threshold. In consequence the stability threshold can $not$ be crossed by solely changing the ratio of layer thicknesses. This was analyzed in Ref. for a variety of experimentally studied systems. Increasing the ratio $d$ from one the system remains either completely in the unstable or in the stable region. To compare the stability behaviour of two-layer and one-layer films we introduce two effective one-layer films as follows. In (1) we assume the lower layer to be solid, i.e. we regard the upper layer as a one-layer film on a coated substrate. In (2) we assume the upper layer to be rigid but deformable by bending. The lower liquid layer corresponds then to a one-layer film on a solid bulk substrate. In case (1) the one-layer liquid film is unstable if the second derivative of the energy with respect to the film thickness $h_2$ is negative, $\partial^2 f/ h_2^2 < 0$. The stability threshold at $\partial^2 f/ h_2^2 = 0$ can be crossed by changing the layer thickness $h_2-h_1$ or the thickness of the coating layer $h_1$. This was demonstrated in Refs. and for a PS film on Si wafers covered with a $1.6$ nm thick SiO layer. In case (2) the one-layer liquid film is unstable for $\partial^2 f /h_1^2 < 0$. It can also be destabilized by changing the layer thicknesses, as was shown in Ref. for a rigid PS layer on top of a liquid PDMS layer on a Si substrate. Comparing the stability thresholds for the two effective one-layer systems to the stability diagram in Fig.\[fig2\] shows that the stability threshold of the two-layer system lies in the region where both effective one-layer systems are stable. This indicates that a two-layer system is always less stable than corresponding effective one-layer systems. Different instability modes {#inst_modes} --------------------------- The stability threshold can be studied in rather general terms as was done above because its main features do not depend on surface tensions or viscosities. However, this is not the case for the characteristics of the instability like mode type, growth rates or dominant wave length. To discuss these we focus in the following on selected two-layer films studied experimentally [@MSS03; @LPHK96; @Sfer97]. We consider various combinations of layers of polystyrene (PS), poly(methylmethacrylate) (PMMA) and poly(dimethylsiloxane) (PDMS) on a silicon (Si) or on a silicon-oxide (SiO) substrate. The Hamaker constants for different combinations are calculated using Eq.(\[HAM\]) and given in Table \[Ham\]. The linear instability of a two-layer film has two different modes. It can be of zigzag or varicose type. For the former the deflections of the two interfaces are in phase whereas for the latter they are in anti-phase. For special parameter values one can also find a mixed type, where modes are present because they have equal fastest growth rates [@PBMT04]. The model studied in Ref. assumes a thick lower layer thereby neglecting the interaction between the substrate and the liquid-liquid and the liquid-gas interface. In this case only the varicose mode can be unstable. In the general case, however, also the zigzag mode can become unstable. Both modes are normally asymmetric, i.e. the deflection amplitudes of the two interfaces differ. We characterize this asymmetry by $\phi = \chi/(1+\chi^2)$. Negative (positive) $\phi$ corresponds to varicose (zigzag) modes. The value $|\phi| = 1/2$ represents the symmetric case, whereas $\phi = 0$ corresponds to maximal asymetry, i.e. one of the interfaces is flat. The asymmetry increases with the ratio of the surface tensions $\sigma$. Note, that the dispersion relation and the type of the dominant mode depend on $\sigma$ and $\mu$, whereas the stability diagram Fig.[\[fig2\]]{} [*does not*]{}. The two mode types are plotted in Fig.[\[fig3\]]{} for a Si/PMMA/PS/air system for a fixed value of $d$ and different ratios of viscosities $\mu$. The dispersion relations $\gamma(k)$ are shown together with the corresponding $\phi$. We show here that the type of the dominant mode can be changed by varying $d$ or $\mu$. It can also change in dependence of the ratio of the interfacial tensions $\sigma$ as studied in Ref.. Strictly speaking, the concept of the mode type characterized by $\phi$ is only valid in the linear stage of the evolution. However, to discuss morphology changes we generalize this concept to nonlinear thickness profiles $h_i (x)$. We define a generalized mode or solution type by the integral $$\phi_{\text{int}} = \frac{1}{L} \int \frac{(h_1-d_1)(h_2-d_2)}{[(h_1-d_1)^2+(h_2-d_2)^2]}\,dx, \label{int_phi}$$ taken over the domain length $L$. In many cases the sign of the product $(h_1 -d_1)(h_2 - d_2)$ does not depend on $x$ allowing to ’read’ the mode type directly from the plots of the layer profiles. For small deflection amplitudes Eq.(\[int\_phi\]) gives again the linear mode type defined above. In the following we use the notion ’mode-type’ in the linear and in the nonlinear regime. Limiting cases {#lim_cases} -------------- For general $d_i$ the radical in Eq.(\[GR\_RATE\]) does not allow to give an analytic expression for the wave number $k_m$ and the characteristic growth time $\tau_m = 1/\gamma_{\rm m}$ of the fastest growing mode. Nevertheless, one can derive asymptotic expressions for $k_m$ and $\tau_m$ in the two important limiting cases of (1) small thickness of the upper layer $d_2-d_1 \ll d_2$ and (2) small thickness of the lower layer $d_1 \ll d_2$. First consider case (1), which corresponds to a liquid film (the upper layer) on a liquid substrate (the very thick lower layer). The dimensional $k_m$ and $\tau_m$ are then given by $$\begin{aligned} k_m =\frac{1}{(d_2-d_1)^2}\sqrt{\frac{ |A_{12g}|}{4 \pi \sigma_{\rm eff}}} \nonumber \\ \tau_m = \frac{16 (2 \pi)^2 \sigma_{\rm eff} \mu_1 (d_2-d_1)^6}{d_1 A_{12g}^2}, \label{1LIM}\end{aligned}$$ with $\sigma_{\rm eff} = \sigma_1 \sigma_2 /(\sigma_1 +\sigma_2)$. Note that all variables are in their dimensional form. Interestingly, the growth time $\tau_m$ depends only on the viscosity of the lower layer $\mu_1$ and does not depend on $\mu_2$. This can be explained by the fact that the flow in the lower layer which is related to $\mu_1$, is much larger than that in the upper one [@BMR93]. At constant thickness of the lower layer, $\tau_m$ is proportional to $(d_2-d_1)^6$, i.e. a liquid film on a bulk liquid substrate evolves faster than the same film on a solid substrate $({\rm growth}~{\rm time} \sim (d_2-d_1)^5)$ and even faster than the same film on a solid substrate with slippage $( {\rm growth}~{\rm time} \sim (d_2-d_1)^5/\left[1 + 3 \beta/(d_2-d_1)\right]$). In case (2), which corresponds to a liquid film (the lower layer) on a solid substrate below the other liquid (the very thick upper layer), the dimensional $k_m$ and $\tau_m$ are given by $$\begin{aligned} k_m =\frac{1}{d_1^2}\sqrt{\frac{A_{21s}}{4 \pi \sigma_1}} \nonumber \\ \tau_m = \frac{12 (2 \pi)^2 \sigma_1 \mu_1 d_1^5}{A_{21s}^2}. \label{2LIM}\end{aligned}$$ Note that in case (2) $k_m$ and $\tau_m$ coincide with $k_{\rm low}$ and $\tau_{\rm low}$, respectively, the characteristics of the dominant mode of the instability of a liquid film below a bulk liquid calculated using one-layer theory. A discussion of this geometry for a Rayleigh-Taylor instability can be found in Refs. and . Long-range apolar and short-range polar interactions {#lr_sr_int} ---------------------------------------------------- The stability analysis based only on long-range interactions becomes incorrect for layer thicknesses in the range of the interaction length $l$ of short-range interactions. Practically, the latter become important (well) below $10$nm layer thickness. In contrast to the result for the exclusive action of long-range van der Waals forces, in the regime where both, short- and long-range interactions, are important the stability threshold can be crossed by changing the layer thicknesses $d_i$. Fig.\[fig4\] presents a selection of qualitatively different stability diagrams in the plane spanned by the layer thicknesses obtained when varying the strength of the short-range interaction for a fixed long-range interaction. By changing the short-range part of the spreading coefficient $S_1$ and $S_2$ one finds seven topologically different types of such diagrams. These types correspond to regions in the $(S_1,S_2)$ plane as indicated in Fig.\[fig5\]. In the absolute unstable region bounded on the right by $(S_1)_{\rm min} = (e/4)^4 A_{21s}/|A_{12g}|$ and above by $(S_2)_{\rm min} =(e/4)^4$ the system can not be stabilized by changing $d_1$ or $d_2$. Only if at least one of the two $S_i$ is larger than the corresponding critical value (a) stable region(s) can be found in the $(d_1,d_2-d_1)$ plane (see Fig.\[fig4\]). For $S_1 > (S_1)_{\rm min}$ a stable region exists that extends towards infinite $(d_2-d_1)$, as shown in Figs.\[fig4\](a), (b) and (d). Thereby, for large $(d_2-d_1)$ the system is stable for $(d_1)_{\rm min}< d_1 < (d_1)_{\rm max}$, where $(d_1)_{\rm max}$ and $(d_1)_{\rm min}$ are the solutions of the equation $A_{21s}/|A_{12g}| = S_1 x^4 \exp{(-x)}$. Similarly, for $S_2 > (S_2)_{\rm min}$ a stable region exists that extends towards infinite $d_1$, as in Figs.\[fig4\](a) to (d). For large $d_1$ the system is stable for $(d_2-d_1)_{\rm min}< d_2 < (d_2-d_1)_{\rm max}$, where $(d_2-d_1)_{\rm max}$ and $(d_2-d_1)_{\rm min}$ are the solutions of the equation $1 = S_2 x^4 \exp{(-x)}$. In the gray shaded triangle at the center of Fig.\[fig5\] an additional bounded stable region exists in the $(d_1,~d_2-d_1)$ plane (see Figs.\[fig4\](b) and (c)). Combining the different conditions gives the following seven types of stability diagrams. - The stable region is continuous and extends in respective stripes towards infinite $d_1$ and $d_2-d_1$ (Fig.\[fig4\](a)). - There exist two separated stable regions, one extending towards infinite $d_1$ and the other one towards infinite $d_2-d_1$ (Fig.\[fig4\](d)). - Similar to Type II but with an additional bounded stable region (Fig.\[fig4\](b)). - A bounded stable region exists together with an unbounded region extending towards infinite $d_1$ (Fig.\[fig4\](c)). - Similar to type IV but with the unbounded region extending towards infinite $(d_2-d_1)$ (not shown). - Only one stable region exists extending towards infinite $d_1$ (not shown). - Similar to type VI but with the unbounded region extending towards infinite $(d_2-d_1)$ (not shown). Further on we will focus our attention on the stability diagram of type I. Non-linear behaviour {#nuss} ==================== Stationary solutions as extrema of the Lyapunov functional {#nu_sol} ---------------------------------------------------------- To find periodic stationary solutions of the scaled Eqs.(\[EVOL\_PHYS\]), the time derivatives $\partial_t h_i$ are set to zero. Integration yields $$\begin{aligned} Q_{11}\, \partial_x \left( \frac{\delta F}{\delta h_1} \right)+ Q_{12}\, \partial_x \left( \frac{\delta F}{\delta h_2} \right) &=& C_1 \nonumber \\ Q_{21}\, \partial_x \left( \frac{\delta F}{\delta h_1} \right)+ Q_{22}\, \partial_x \left( \frac{\delta F}{\delta h_2} \right) &=& C_2, \label{STAT}\end{aligned}$$ where the $C_i$ are constants and $F$ is given by Eq.(\[ENRG\_SCL\]). Note that the left hand sides of Eqs.(\[STAT\]) represent the flow in the lower layer and the total flow, respectively. For a stationary state both flows are zero, i.e. the $C_1=C_2=0$. Because the mobility matrix ${\bm Q}$ is non-singular, one concludes from Eqs.(\[STAT\]) that the stationary states of the Eqs.(\[EVOL\_PHYS\]) are the extrema of the Lyapunov functional $F$, i.e. they are solutions of $$\begin{aligned} -\partial_{xx} h_1 +\frac{\partial f}{\partial h_1} &=& c_1 \nonumber \\ -\sigma \partial_{xx} h_2 +\frac{\partial f}{\partial h_2} &=& c_2, \label{STAT_1}\end{aligned}$$ where $f$ denotes the local part of Eq.(\[ENRG\_SCL\]) and the constants $c_i$ correspond to the Lagrangian multipliers introduced in Section\[deriv\]. To obtain a finite amplitude solution for given mean thicknesses we use continuation techniques [@DKK91; @DKK91b; @Doedel97]. We start with analytically known stationary periodic small-amplitude profiles, which correspond to the linear eigenfunctions for the critical wave number $k_c$. By continuation we follow the family of solutions changing the period $L$. We characterize the solutions by the deflection amplitudes $A_1$ and $A_2$, the energy $E$, the norm $L_2 = (1/L) \int[ (h_1-d_1)^2+(h_2-d_2)^2]\,dx$ and the integral mode type $\phi_{int}$. To determine the stability of the stationary solutions $h_i(x)$, we add small perturbations $\delta h_i (x) \sim exp{(\beta t)}$ to both interfaces $h_i(x)$, linearize the full time-dependent evolution equations (\[EVOL\_PHYS\]) around $h_i(x)$ and solve the obtained eigenvalue problem ${\bm L}(h_i, \partial_x h_i, \partial_x ) {\bm \delta \bm h(x)} = \beta {\bm \delta \bm h(x)}$ for the linear operator ${\bm L}$ after discretizing it in space. The sign of the largest eigenvalue $\beta$ determines the stability of the stationary solution. Note that due to the translational invariance of the evolution equations (\[EVOL\_PHYS\]), there exists always a symmetry mode $\delta h_i (x) = \partial_x h_i(x)$ with the eigenvalue $\beta = 0$. Mode type transitions {#mt_trans} --------------------- The type of the dominant instability mode calculated above by linear stability may not persist in the course of the nonlinear evolution. Possible mode type changes may have a dramatic effect on the (observable) overall morphology of the film. We investigate these changes by studying both, the evolution in time of the film profiles and the stationary solutions obtained by continuation. The evolution in time is obtained by numerical simulations of the scaled coupled evolution equations Eqs.(\[EVOL\_PHYS\]) in a one-dimensional periodic domain. Both, semi-implicit pseudo-spectral and explicit time integration schemes are used. Initial conditions consist of flat layers with an imposed noise of amplitude 0.001. ### Transition via branch switching {#trans_branch} First the time evolution of an initially flat film is studied for parameters as in Fig.\[fig3\](a) using a domain size equal to four times the fastest growing wave length $\lambda_m$. A time sequence of snap shots is shown in Fig.\[fig6\]. In the early stage of the evolution a varicose mode develops $(t = 8.1)$ as expected from the linear analysis. Then in a sub-domain of size $\lambda_m$ the deflection amplitudes increase dramatically accompanied by a morphological change towards a zigzag type profile ($t = 9.7$). This is further illustrated by the dependence of the integral mode-type (Eq.\[int\_phi\]) on time given in Fig.\[fig7\](b). Further on, the length of the zigzag part increases slightly, and coarsening sets in resulting in the disappearance of one varicose-type drop ($t = 16.4$). Next one of the remaining drops increases it amplitude and flips to a zigzag type hole ($t=20.6$). Finally the last remaining varicose-type drop disappears $(t\approx28)$, and the system approaches a stationary (but not stable) state. The evolution of the relative energy of the profile in time is given in Fig.\[fig7\](a). It is seen very clearly that the phases of very slow evolution correspond to solutions that are close to stationary solutions. This results from the fact that the (unstable) stationary solutions form saddle points in function space that are approached along their stable manifolds and subsequently repel the system along their unstable manifolds (for a discussion see Ref.) The evolution stops after a further coarsening step, when the period becomes equal to the system size (not shown). To explain the observed mode-type change, we study the stationary solutions of the evolution equations Eqs.(\[EVOL\_PHYS\]). We find a family of solutions starting at the subcritical primary bifurcation, then turning three times at saddle-node bifurcations (folds) and going towards infinite periods (see Fig.\[fig8\](a)). A stability analysis using the solution period as the period of the disturbance (thereby excluding coarsening modes) shows that two branches are stable (solid lines) and two are unstable (dashed lines). Along the first unstable branch, which starts at $L_c$ and ends at the first fold at $L\approx 60$ the energy $E$ is always larger than the one of the flat film $E_0$ (Fig.\[fig8\](b)), and it increases with decreasing period. This subcritical branch corresponds to nucleation solutions that have to be overcome to break the film in parts smaller than $L_c$ (see Ref. for a discussion of this type of solutions for a one-layer system). The first stable branch starts at the first fold at $L \approx 59$ and ends at the second fold at $L \approx 116$. Its relative energy decreases monotonically with increasing period. Mostly it is energetically preferable to the flat film. The second unstable branch (between the second fold at $L \approx 116$ and the third fold at $L \approx 79$) turns back towards smaller periods. The second stable branch starts at the third fold and goes towards infinite periods. Its energy decreases rapidly from values even above the flat film to values below the ones of the first stable branch. The energy of the second unstable branch is always larger than the energies of both stable branches. This indicates that it corresponds to nucleation solutions, or critical solutions that have to be overcome to switch between the two stable branches. Along the second unstable branch, the mode-type changes from varicose to zigzag (Fig.\[fig8\](d)) explaining the non-trivial behavior observed in the time evolution shown in Fig.\[fig6\]. There are two stable solutions with a period equal to the dominant linear wave length $(\lambda_m \approx 108)$ (see Fig.\[fig9\]). The one of higher energy that is approached first in the time evolution is of varicose type whereas the one of lower energy that the system switches to is of zigzag type (cp. Fig.\[fig8\](b)). A transition between the two solutions is accompanied by a strong increase of the amplitude $A_1$ (see Fig.\[fig8\](a)). ### Transition via coarsening {#trans_coars} A mode-type change is not always connected to a transition between different branches of stationary solutions. Also coarsening along one branch may lead to such a change if the mode-type varies along the branch. To demonstrate this, we simulate the time evolution using parameters as in Fig.\[fig3\](b). A time sequence of profiles and the corresponding evolution of the relative energy and the integral mode-type are shown in Figs.\[fig10\] and \[fig11\], respectively. Early in the evolution the layer profiles represent a zigzag mode $(t = 6.0)$ corresponding to the linear results (Fig.\[fig3\](b)). Then, within the very short period of time from $t=6.0$ to $10.8$, nonlinear effects result in a first change towards a varicose type profile, as shown in the inset of Fig.\[fig11\](b). Then the system approaches the branch of stationary solutions. As a result the evolution slows down and the pattern begins to coarse. With ongoing coarsening $(t >10.8)$ the size of the droplets increases $(t=135,~t=461)$ and at very late times $(t>490)$ the mode type changes back to zigzag type (Fig.\[fig11\](b)). Here, the amplitudes of the interfaces do not change dramatically, as was the case in Section \[trans\_branch\]. In this sense the transition is continuous. The characteristics of the corresponding stationary solutions are shown in Fig.\[fig12\]. The primary bifurcation is again subcritical (Fig.\[fig12\](a)) the solution family continues towards smaller periods until turning at a saddle-node bifurcation (fold) and heading towards infinite periods. The subcritical branch is unstable with energies higher than the energy of the flat film (Fig.\[fig12\](b)). The second branch starting at the fold $(L \approx 26)$ consists of solutions whose energy decreases monotonically with increasing period. They are stable to disturbances of identical period but unstable to coarsening modes. Fig.\[fig12\](d) shows that the solution with the period equal to $\lambda_m = 50.81$ is of varicose type. The corresponding layer profile is shown in Fig.\[fig13\](a) together with the profile after the first coarsening step. When the period becomes larger than $94.2$ , the solution changes to zigzag type (Fig.\[fig12\](d)) as shown in Fig.\[fig13\](b). This explains the mode-type change found in the time evolution (Fig.\[fig10\]). Here we have restricted our attention to a parameter set corresponding to region I of Fig.\[fig5\], i.e. corresponding to the stability diagram shown in Fig.\[fig4\](a). The existence of a stable branch of stationary solutions which continues towards infinite period implies that the rupture of the two layers is completely avoided by the short-range repulsion. However, this may not be the case for parameter ranges belonging to the other types of stability diagrams. A detailed analysis of the stationary solutions for all types will be done elsewhere. Large-period stationary solutions {#lp_sol} --------------------------------- The stability of the numerical code, used to solve the evolution equations Eqs.(\[EVOL\_PHYS\]), requires a very small time step $t =0.00001$. As a result it takes very long even to reach the final stationary solution in a system of size $4\lambda_m$ using $256$ grid points. It is not feasible at the moment to study many coarsening steps in this way. However, one can rely on continuation techniques that use an adaptive spatial grid along the continuation path [@Doedel97] to obtain stationary solutions of arbitrarily large periods that correspond to solutions that would be obtained in a time evolution at very late times. We show in Fig.\[fig14\] possible large-period stationary solutions for different physical systems that are investigated experimentally. One finds qualitatively different morphologies like a drop of the lower liquid ’looking through’ a nearly flat film of the upper layer for a Si/PMMA/PS/air system. Note however that also the upper layer is continuous (due to the stabilizing short-range interaction), i.e. also the drop is covered by a very thin layer of the upper liquid. This is more pronounced for a SiO/PS/PDMS/air system. In contrast, for a SiO/PMMA/PS/air system one finds a drop of the upper liquid ’swimming’ on the lower liquid that however is attracted towards the base of the drop. These equilibrium solutions are equivalents of drop configurations studied in Ref. for macroscopic (but smaller than the capilary length) drops. However, here the mesoscopic contact angles are not given explicitely but result from the underlying effective molecular interactions, i.e. the short- and long-range forces used. Conclusion {#conc} ========== We have derived coupled non-linear evolution equations for the profiles of the liquid-liquid and liquid-gas interfaces of a thin two-layer liquid film heated from below allowing for slip at the substrate. We have shown that in the isothermal case the evolution equations can be written in terms of variations of an appropriate Lyapunov functional $F$ which monotonically decreases in time. The stability conditions for flat layers have been given in terms of $F$. We have shown that a two-layer film is less stable than related effective one-layer films introduced in Section \[lr\_int\]. Even if both effective one-layer films are stable the two-layer film may be unstable if the determinant of the energy matrix $\det {\bm E}$ is negative. We have shown that if the Hamaker constants are given by the usual expression (Eq.(\[HAM\])), i.e. they are coupled through the refractive indices, and no other forces are present, the stability of the flat films with thicknesses of ($\sim100$nm) can not be changed by solely changing the layer thicknesses. Incorporating a stabilizing short-range interaction the stability can be changed in this way. We have classified the resulting possible types of stability diagrams in the space of the layer thicknesses and given a ’phase diagram’ in terms of the short-range parts of the spreading coefficients for the occurence of the different types of stability diagrams. In general, the linear stability analysis of the flat film has shown that both, varicose or zigzag mode, may be unstable depending on the ratios of the layer thicknesses, viscosities and surface tensions (see also Ref.). This seems to be in contrast to Ref.. However, the difference arises because there it is assumed that the lower layer is thick compared to the upper layer neglecting thereby all interactions with the substrate. Then the zigzag mode is always stable. The introduction of the stabilizing short-range interaction allows to study the long-time evolution and stationary layer profiles. Possible stationary states have then been determined as extrema of the Lyapunov functional $F$. The resulting bifurcation diagrams show a rich branch structure that depends strongly on parameter values. We have focused on one type of stability diagram where a stable branch going towards infinite period always exists. This implies the existence of a non-ruptured stationary state in the long-time limit also in the time evolution. We have found that the mode type of a profile may change during the evolution of an instability on three ways. First, the profile type changes in the course of the short-time evolution. This is connected to different mode types found for the dominant linear mode and the stationary solution of equal period on the solution branch approached first in the time evolution. It seems that this behaviour is more probable for a subcritical primary bifurcation. In the case studied here this change is from zigzag to varicose. In the nonlinear regime the profile can change its type by (i) jumping from one to another stable branch and by (ii) coarsening along a single stable branch. Combinations of the different ways may also be possible. We have found that for the parameters considered here both nonlinear transitions go from varicose towards zigzag type. In case (i) the transition occurs without change of the period, but with a dramatic increase in amplitude of the profile. In the case (ii) the transition occurs continuously without amplitude jump because mediated by coarsening a small-period varicose mode turns into a large-period zigzag one. In all examples considered here (except the SiO/PMMA/PS/air system with $d_1 = 30$, $d_2 = 50$) we have found a zigzag-type solution at large periods. For the future it would be very interesting to further analyze the stationary solutions for a broader range of experimentally interesting systems like the ones studied in Refs.. This should clarify under which conditions the long-time (or large-period) solutions are energetically preferable and determine how ’late’ the transition may occur. A systematic analysis of all types of stability diagrams would also discuss metastability and absolute stability of the flat two-layer films. Furthermore, we are very optimistic that the evolution equations presented here will serve to study the questions of hole growth and possible front instabilities in the dewetting of a liquid layer on a liquid substrate of finite thickness[@FCW95; @PWHC97]. {#appendix} To further specify the thermocapillar part of Eqs.\[EVOL\_PHYS\] we rewrite the derivatives $\partial_x \sigma_{12}$ and $\partial_x \sigma_{2}$ in terms of the gradients of the local thicknesses $\partial_x h_i$. In the long-wave approximation [@ODB97] the temperature field is in both layers a linear function of the vertical coordinate $z$, i.e. $T_i = a_i z+b_i$. To determine the coefficients we consider a three-layer geometry (Fig.\[fig15\]), i.e. we take into account the heat conduction in a gas layer of finite thickness $d_g = d_t -d_2$, where $d_t$ is the distance between the substrate and an upper plate. The temperature in the gas layer is $T_g = a_g z +b_g$. The boundary conditions at both interfaces are continuity of the temperature field and continuity of the heat flux $\kappa_i \partial_z T_i=\kappa_k \partial_z T_k$, where $\kappa_i$ is the thermal conductivity of the $i$-th layer. The temperatures at the substrate $T_0$ and at the upper plate $T_t$ are constant. The coefficients $a_i$ and $b_i$ depend on the local thicknesses $h_i$ and are given by $$\begin{aligned} a_g = \frac{\alpha \Delta T }{d_t-h_2 + \frac{\kappa_g}{\kappa_1}h_1 + \frac{\kappa_g}{\kappa_2}(h_2-h_1)} \nonumber \\ \nonumber \\ a_2 = \frac{a_g \kappa_g}{\kappa_2},~~~~~a_1 = \frac{a_g \kappa_g}{\kappa_1} \nonumber \\ \nonumber \\ b_1 = T_0,~~~~b_g = T_t -a_g d_t \nonumber \\ b_2 = a_g \kappa_g h_1 \left( \frac{1}{\kappa_1} - \frac{1}{\kappa_2} \right) + T_0, \nonumber \end{aligned}$$ where $\Delta T = T_0 - T_2$ and $$\begin{aligned} \alpha = \frac{d_2 - d_t - \kappa_g h_1 / \kappa_1 -\kappa_g (d_2-d_1)/ \kappa_2}{\kappa_g h_1 / \kappa_1 -\kappa_g (d_2-d_1)/ \kappa_2}. \nonumber \end{aligned}$$ Here $T_2$ is the temperature of the liquid-gas interface, when both interfaces are undeformed, i.e. for $h_i = d_i$. The above formulas allow to determine the derivatives $$\begin{aligned} \partial_x \sigma_{12} = \Gamma_{11} \partial_x h_1 + \Gamma_{12} \partial_x h_2 \\ \nonumber \partial_x \sigma_{2} = \Gamma_{21} \partial_x h_1 + \Gamma_{22} \partial_x h_2, \label{sigmas}\end{aligned}$$ where the matrix ${\bm \Gamma}$ is determined as follows $${\bm \Gamma} = a\left( \begin{array}{ccc} \frac{ \kappa_g}{\kappa_1}\frac{d \sigma_1}{ d T} b &~~~& -\frac{ \kappa_g}{\kappa_1}\frac{d \sigma_1}{ d T}h_1(\frac{\kappa_g}{ \kappa_2} - 1) \\ \\ \frac{d \sigma_2}{d T} ( bc-\frac{\kappa_g}{\kappa_2}h_2b ) & & \frac{d \sigma_2}{d T} \{ \frac{\kappa_g}{\kappa_2} ( d_t+h_1 b )-bh_1 (\frac{\kappa_g}{ \kappa_2}-1) \} \end{array} \right). \label{Gamma}$$ Here $a = (\alpha \Delta T)/\left[ d_t-h_2 + \frac{\kappa_g}{\kappa_1}h_1 + \frac{\kappa_g}{\kappa_2}(h_2-h_1) \right]^2$, $b= \kappa_g (1/\kappa_1 -1/\kappa_2)$ and $c = \{d_t-h_2(1-\frac{\kappa_g}{\kappa_2})\}$. For the linear normal Marangoni effect $d \sigma_{12} / dT$ and $d \sigma_{2} / dT$ are negative and constant. The Eqs.(\[sigmas\]) are used in Eqs.(\[EVOL\_PHYS\]) to obtain a closed system of equations for $h_1$ and $h_2$. [71]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{} , , , , ****, (). , ****, (). , , , , ****, (). , ****, (). , , , , , ****, (). , , , ****, (). , , , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , , , , ****, (). , , , , , ****, (). , , , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , , , ****, (). , ****, (). , , , , ****, (). , ****, (). , , , ****, (). , , , ****, (). , ****, (). , , , ****, (). , , , ****, (). , ** (, , ). , **, vol.  (, , ). , , , ****, (). , , , , , , , ****, (). , , , , , ****, (). , , , , (), . , , , ****, (). , , , ****, (). , ****, (). , , , ****, (). , , , , , ****, (). , , , , , , ****, (). , , , , , ****, (). , ** (), . , , , , ****, (). , , , , , , , ****, (). , , , ****, (). , , , ****, (). , , , , ****, (). , , , , ****, (). , , , , ****, (). , , , ****, (). , ****, (). , ****, (). , , , ****, (). , ****, (). , ** (, , ). , ****, (). , ****, (). , ** (, , ). , , , ****, (). , , , , ****, (). , , , , , , ****, (). , ****, (). , , , ****, (). , , , ****, (). , , , , , , ** (, , ). , , , , ****, (). , , , ****, (). ${\rm refractive}~{\rm indices}$ $A_{12g}$ $ A_{21s}$ $ A_{g21s}$ ---------------------------------------- ----------- ------------ ------------- $n_s > n_1$, $n_1 < n_2$, $n_2 > n_g$ + + - $n_s < n_1$, $ n_1 < n_2$, $n_2> n_g$ + - + $n_s < n_1$, $ n_1 > n_2$, $n_2 > n_g$ - + + $n_s > n_1$, $n_1 > n_2$, $n_2 > n_g$ - - - : \[Ham\_Table\] Possible combinations of signs of the different Hamaker constants for given order of the refractive indices of the involved medias. ${\rm System}$ $A_{12g} \times 10^{-20} {\rm Nm}$ $A_{21s} \times 10^{-20} {\rm Nm}$ $A_{g21s} \times 10^{-20} {\rm Nm}$ ------------------------- ------------------------------------ ------------------------------------ ------------------------------------- ${\rm Si/PMMA/PS/air}$ $1.49$ $3.8$ $-23.02$ ${\rm SiO/PMMA/PS/air}$ $1.49$ $-0.024$ $0.15$ ${\rm SiO/PS/PDMS/air}$ $-1.83$ $0.42$ $1.25$ : \[Ham\]Hamaker constants for various combinations of polymers. {width="0.7\hsize"} [**Fig.\[fig1\]\ Pototsky et al., J. Chem. Phys.**]{} {width="0.8\hsize"} [**Fig.\[fig2\]\ Pototsky et al., J. Chem. Phys.**]{} {width="0.7\hsize"} [**Fig.\[fig3\]\ Pototsky et al., J. Chem. Phys.**]{} {width="0.8\hsize"} [**Fig.\[fig4\]\ Pototsky et al., J. Chem. Phys.**]{} {width="0.8\hsize"} [**Fig.\[fig5\]\ Pototsky et al., J. Chem. Phys.**]{} {width="0.8\hsize"} [**Fig.\[fig6\]\ Pototsky et al., J. Chem. Phys.**]{} {width="0.8\hsize"} [**Fig.\[fig7\]\ Pototsky et al., J. Chem. Phys.**]{} {width="0.8\hsize"} [**Fig.\[fig8\]\ Pototsky et al., J. Chem. Phys.**]{} {width="0.6\hsize"} [**Fig.\[fig9\]\ Pototsky et al., J. Chem. Phys.**]{} {width="0.8\hsize"} [**Fig.\[fig10\]\ Pototsky et al., J. Chem. Phys.**]{} {width="0.8\hsize"} [**Fig.\[fig11\]\ Pototsky et al., J. Chem. Phys.**]{} {width="0.8\hsize"} [**Fig.\[fig12\]\ Pototsky et al., J. Chem. Phys.**]{} {width="0.6\hsize"} [**Fig.\[fig13\]\ Pototsky et al., J. Chem. Phys.**]{} {width="0.8\hsize"} [**Fig.\[fig14\]\ Pototsky et al., J. Chem. Phys.**]{} {width="0.7\hsize"} [**Fig.\[fig15\]\ Pototsky et al., J. Chem. Phys.**]{}

ArXiv

--- abstract: 'Speech enhancement is a crucial task for several applications. Among the most explored techniques are the Wiener filter and the LogMMSE, but approaches exploring deep learning adapted to this task, such as SEGAN, have presented relevant results. This study compared the performance of the mentioned techniques in 85 noise conditions regarding quality, intelligibility, and distortion; and concluded that classical techniques continue to exhibit superior results for most scenarios, but, in severe noise scenarios, SEGAN performed better and with lower variance.' author: - | Tito Spadini[^1]\ CECS\ Universidade Federal do ABC\ Santo André, SP\ `[email protected]`\ Ricardo Suyama\ CECS\ Universidade Federal do ABC\ Santo André, SP\ `[email protected]`\ bibliography: - 'main.bib' title: Comparative Study between Adversarial Networks and Classical Techniques for Speech Enhancement --- Introduction ============ Since the 1980s, speech enhancement and denoising researches exploit neural networks’ ability to work as non-linear filters [@DNNSPEECHI; @DNNSPEECHII; @DNNSPEECHIII], but it’s performance was often unsatisfactory — usually due to the reduced amount of training data or even the limited flexibility of the networks, caused by the inefficiency of the training algorithm for more extensive networks, with more neurons and layers. A new perspective for the use of neural networks arose after [@HINTON2006], which has exhibited that employing an unsupervised pre-training of the network for layers can bypass the limitation found by gradient-based algorithms. This new possibility, coupled with the rise of new tools for training neural nets using GPU (Graphics Processing Unit) [@scherer2010evaluation], has rekindled interest in neural networks. Implementations of neural networks with large numbers of neurons and layers now include the term “*Deep*” in their nomenclature, and therefore the term Deep Learning has become popular when referring to Deep Neural Networks (DNN). The flexibility and power of this type of neural network have shown promising results and have attracted a growing number of researchers in the field. However, there is a wide variety of DNN architectures; each with its respective characteristics and particularities, which may be more convenient for specific applications and therefore less suitable for others. For the applications discussed in this paper, an architecture known as convolutional autoencoder will be explored. A recent approach called Generative Adversarial Networks (GAN) [@NIPS2014_5423] is a structure composed of individuals — usually, two neural networks — competing against each other and exploring concepts of Game Theory and Deep Learning. In this competitive two-player game, there is a well-prepared dataset, composed of samples of the same type, appropriately chosen, but with different attribute values. The *Discriminator* player, $D$, has the purpose of discriminating whether a sample came from the original dataset or not; the *Generator*, $G$, must capture the distribution of the original dataset and use it to generate entirely new samples. Thus, while one of the players intends to generate the perfect imitation of the original data, the other player tries to be the best possible counterfeit identifier. For GAN-based methodes, the network training should occur gradually and concomitantly for both players, otherwise, an evolutionary imbalance may occur in favor of one of the actors, therefore, instead of achieving a good evolution for both, only one network will evolve minimally compared to the other, which does not even guarantee good results for at least one of the players. The GAN adjusting criterion is given by: $$\min_{G} \max_{D} V\left(D, G\right) = \mathbb{E}_{\pmb{x}\thicksim p_{\text{data}}\left(\pmb{x}\right)}\left[\log D\left(\pmb{x}\right)\right] + \mathbb{E}_{z\thicksim p_{z}\left(z\right)}\left[\log \left(1 - D\left(G\left(z\right)\right)\right)\right], \label{eq:minimax}$$ where $V$ is the value function; $D$ is the Discriminator (player), a multi-layer perceptron that generates the probability that $\pmb{x}$ has originated from the legitimate data instead of the distribution $p_{G}$; $G$ is the Generator (player); $p_{\text{data}}$ is the data distribution; and $p_{\text{\pmb{z}}}$ is a prior in the input noise variables. In the same way that GAN can learn a generative model for training data, conditional GAN (cGAN) [@CGAN], as its name may suggest, provides a conditional generative model for the data. The data generation is based on a prior distribution and also on an additional input $\pmb{x}_c$, thus conditioning the distribution of the generated data to the additional information provided by $\pmb{x}_c$. The cost function of cGAN is given by: $$\min_{G} \max_{D} V\left(D, G\right) = \mathbb{E}_{\pmb{x}, \pmb{x}_c\thicksim p_{\text{data}}\left(\pmb{x}, \pmb{x}_c\right)}\left[\log D\left(\pmb{x},\pmb{x}_c\right)\right] + \mathbb{E}_{z\thicksim p_{z}\left(z\right), \pmb{x}_c\thicksim p_{data}\left(\pmb{x}_c\right)}\left[\log \left(1 - D\left(G\left(z, \pmb{x}_c\right),\pmb{x}_c\right)\right)\right], \label{eq:minimax_cGAN}$$ Despite the advancement of the original GAN’s and cGANs based on minimizing and , in some cases the training may converge to solutions with performance below desired. For this reason, in [@LSGAN], an alternative proposal was presented, called Least Squares GAN (LSGAN), which seeks to adapt the discriminator and the generator according to the following criteria: $$\min_{D} V_{LSGAN} \left(D\right) = \frac{1}{2}\mathbb{E}_{\pmb{x}, \pmb{x}_c\thicksim p_{\text{data}}\left(\pmb{x}, \pmb{x}_c\right)}\left[\left( D\left(\pmb{x},\pmb{x}_c\right)-1\right)^2\right] + \frac{1}{2}\mathbb{E}_{z\thicksim p_{z}\left(z\right), \pmb{x}_c\thicksim p_{data}\left(\pmb{x}_c\right)}\left[ D\left(G\left(z, \pmb{x}_c\right),\pmb{x}_c\right)^2\right], \label{eq:LSGAN_D}$$ $$\min_{G} V_{LSGAN} \left(D\right) = \frac{1}{2}\mathbb{E}_{z\thicksim p_{z}\left(z\right), \pmb{x}_c\thicksim p_{data}\left(\pmb{x}_c\right)}\left[ \left(D\left(G\left(z, \pmb{x}_c\right),\pmb{x}_c\right)-1\right)^2\right], \label{eq:LSGAN_G}$$ A few years later, a GAN-based approach called Speech Enhancement GAN (SEGAN) was introduced in [@SEGAN], exploring autoencoders neural networks adapted using RMSProp. Also, it has presented promising results. Further details on SEGAN will be provided in Section II. One of the most decisive aspects regarding the adoption of an autoencoder, which in this case is fully convolutional, is its infundibuliform architecture capable of preserving the signal structure, ensuring that the obtained output will respect the same form used as network input. Also, the autoencoder can discard dispensable parts of the signals, i.e., noise, which causes the preservation of signal information to occur. However, due to the connections between encoding layers and decoding layers, the denoising effect caused by the network does not occur aggressively to the point of destroying the signal in terms of quality and intelligibility only for if noise reduction is achieved. This work’s objective is to compare SEGAN’s performance against Wiener filter and Log Minimum Mean Square Error (LogMMSE), concerning quality and intelligibility through objective and perceptual metrics, for several different noise scenarios. However, different from [@SEGAN], this work considered LogMMSE in its comparisons, in addition to Wiener’s filter and SEGAN itself. Moreover, in addition to the Perceptual Evaluation of Speech Quality (PESQ) metric, explored in both studies, this work includes the Short-Time Objective Intelligibility (STOI) and Signal-to-Distortion Ratio (SDR) metrics. This work considered 5 different SNR scenarios (0 dB, 5 dB, 10 dB, 15 dB and 20 dB), which includes a new SNR value (20 dB) compared to those used in [@SEGAN]. The scenarios used in this paper considered 17 different types of noise, while [@SEGAN] used 5 types of noise. In order to present the results of this study, the article adopts the following structure: section II presents a review on the main concepts related to SEGAN; section III covers the simulation scenarios and the metrics; section IV discuss the results and offers the final comments on the study; and section V exhibits the conclusions of the work. Speech Enhancement GAN - SEGAN ============================== *Speech Enhancement Generative Adversarial Network* (SEGAN) [@SEGAN] is based on the idea introduced by cGAN, discussed in the previous section. The proposed structure for the generator, in this case, resembles an *autoencoder*, shown in Figure \[fig:autoencoder\]. The successive convolutional layers provide, at the end of the coding process, a vector $\mathbf{c}$ which corresponds to a condensed representation of the input signal. In this process, *strided convolutions* layers are used, since this type of layer has been shown to be more stable in GAN training. However, unlike the structure of a traditional autoencoder, the decoding process is done from the vector $\mathbf{c}$ concatenated with the vector of latent variables $\mathbf{z} $. This new vector is then subjected to a sequence of layers that seek to reverse the coding process by means of *fractional-strided transposed convolutions* [@SEGAN]. The network structure also includes *skip connections* connecting the outputs of the layers in the encoding process directly to corresponding layers in the decoding process. The reason for this was to try to maintain the underlying structure of the observed (noisy) data in the data being generated by GAN. ![\[fig:autoencoder\]Autoencoder architecture for speech enhancement. Based on [@SEGAN].](autoencoder.pdf){width=".375\linewidth"} Although the presented LSGAN criterion may help with some known problems of structure adaptation, the authors propose, based on preliminary simulations, that the cost function includes an additional term in order to favor solutions that minimize the distance between generated data and authentic examples. The distance, however, is measured with the norm $\ell_1$, so that the adopted criterion is defined by: $$\min_{G} V_{LSGAN} \left(G\right) = \frac{1}{2}\mathbb{E}_{\pmb{z}\thicksim p_{\pmb{z}}\left(\pmb{z}\right),\ \pmb{\tilde{x}}\thicksim p_{data}\left(\pmb{\tilde{x}}\right)} \left[\left(D\left(G\left(\pmb{z},\ \pmb{\tilde{x}}\right),\ \pmb{\tilde{x}}\right)-1\right)^{2}\right] + \lambda\ \norm{G\left(\pmb{z},\ \pmb{\tilde{x}}\right)-\pmb{x}}_{1}, \label{eq:LSGAN}$$ where $\pmb{\tilde{x}}$ represents the input (noisy) signals. Simulation Setups ================= To evaluate the performance, several testing scenarios were created from the combination of 20 voices of different people from the dataset VCTK-Corpus [@veaux2017cstr] reading three different sentences; (0 dB, 5 dB, 10 dB, 15 dB and 20 dB) for each of the 17 types of noise, all coming from the DEMAND [@thiemann2013demand] dataset, resulting in 85 different noise conditions for each sentence read by each person. In order to preserve a variability that would assist in the quest for broader generalizability in the SEGAN model training, the selected corpus was chosen to maintain a uniform gender distribution and to ensure different accents. Although the datasets were used in SEGAN’s work, the selected individuals were not the same and more types of noise were used, including an extra SNR level (20 dB) that was not explored in the mentioned work. With all different voice and noise scenarios previously detailed, 5100 different mixtures were obtained to be processed by the three speech enhancement techniques (Wiener filter, LogMMSE, and SEGAN). Using SoX, pre-processing was performed to ensure that all input signals conform to the 16 kHz, 16-bit, and mono configuration in WAV format. The selected Wiener filter belongs to the Scipy library; LogMMSE [@LOGMMSE; @LOGMMSE_Python], is also available as a Python package of the same name; and the SEGAN (pre-trained) model is the same as the original work [@SEGAN], which is openly distributed by the SEGAN authors themselves in their GitHub repository. Such a model had been trained for 86 times using RMSprop [@tieleman2012lecture] and learning rate of 0.0002 in batches of size 400 [@SEGAN]. For each technique, after processing, an enhanced signal was obtained for each noisy signal used as input; and, based on the improved signal and the reference clean voice signal (directly from the VCTK-Corpus dataset), the (PESQ) [@PESQ], a perceptual quality metric with values from -0.5 to 4.5; the STOI [@STOI], which measures the improvement of intelligibility with values from 0 to 1; and the SDR [@SDR], which quantifies the rate between the speech signal and the distorting effects of improved speech signal, were calculated to perform the improvement evaluation. Results ======= The average values of PESQ can be seen on Figures \[fig:pesq\_avg\] and \[fig:pesq\]. It can be noted that the Wiener filter maintained a near linear behavior as the SNR was increased; although it was the technique with the worst performance for SNR 0 dB, it reached the best result of PESQ observed in this work for the case in 20 dB. LogMMSE has proven to be very effective from the start, going through all the scenarios as one of the best in terms of quality. SEGAN, on the other hand, showed a subtle superiority in the 0 dB scenario, but showed little improvement for higher SNR values, being on average much lower than the other two quality techniques analyzed during the enhancement process. Nevertheless, it is important to highlight the fact that the variance of SEGAN was lower in all scenarios. ![\[fig:pesq\]PESQ for different noise levels.](pesq_avg.pdf){width="\linewidth"} ![\[fig:pesq\]PESQ for different noise levels.](pesq.pdf){width="\linewidth"} Figures \[fig:stoi\_avg\] and \[fig:stoi\] show STOI averages. The intelligibility is shown to be higher for the Wiener filter in all scenarios, which means that such a technique resulted in lower degeneration of speech comprehension. The LogMMSE showed a much lower performance than the other techniques for low SNR scenarios; improved slightly for 15 dB and 20 dB, but still got much worse than the Wiener filter. SEGAN showed a performance similar to that of the Wiener filter for the 0 dB scenario and remained superior to LogMMSE for this relation between the signal and distortions for almost all scenarios of different SNR values, except for 20 dB. ![\[fig:stoi\]STOI for different noise levels.](stoi_avg.pdf){width="\linewidth"} ![\[fig:stoi\]STOI for different noise levels.](stoi.pdf){width="\linewidth"} The performance of each technique in terms of SDR can be seen in Figures \[fig:sdr\_avg\] and \[fig:sdr\]. Notwithstanding the poor performance of Wiener filter for cases of lower SNR, it proved to be quite effective for higher SNR scenarios. The LogMMSE approach presented a similar performance to the Wiener filter. And, although it was the technique with the best performance for low SNR cases, SEGAN showed little improvement for cases with higher SNR; in the case with SNR 20 dB, its performance was well below that obtained by the other techniques; however, as with the PESQ metric, the variance of this technique was much lower than the other techniques. ![\[fig:sdr\]SDR for different noise levels.](sdr_avg.pdf){width="\linewidth"} ![\[fig:sdr\]SDR for different noise levels.](sdr.pdf){width="\linewidth"} Discussion ========== Although it was not the focus of this paper, there are some considerations to be made regarding the performance in terms of resources required for the speech enhancement process to be performed. While both classical methods adopted are based on an unsupervised approach, the GAN-based method is supervised, which requires a fundamental training step to be performed based on a pre-selected data set; and this step is computationally expensive as it took several hours to complete, even though it was performed on a GPU. Still, the application of the enhancement process itself through SEGAN’s trained-model is not fast either, and it may take several seconds to complete the application over a single audio track of a few seconds. By contrast, classical methods performed the process almost instantaneously for each audio track and required no prior training. Regarding speech quality and intelligibility, even considering the respectable, effective and efficient existing objective metrics, such as PESQ and STOI, respectively, if the enhanced signal is aimed for direct use by people, the use of metrics that still have the opinion of people, like *Mean Opinion Score* (MOS) [@streijl2016mean], may continue to be utilized, even if it has a lesser weight; after all, for various purposes, the human sense to evaluate and perceive distinct levels of quality may not yet have been well-enough designed in computational algorithms and metrics. The use of before-mentioned metrics would introduce factors of subjectivity into the process, which can be understood as something to be avoided; yet, if their influence is carefully managed in the appraisal and weighting, perhaps the results may be more satisfactory. It is appropriate to indicate that Figure \[fig:stoi\] also shows two critical details respecting the STOI results: an enormous variance and a colossal amount of outliers. The results show a considerable decrease in variance accompanied by a noticeable increase in already evident outliers. Such peculiarities described in this paragraph are worrisome regarding speech perception issues in noisy scenarios. This result may have been negatively affected by an innocently naive choice of complex audios, or by a training step that underwent from severe data frugality. Given the observations indicated in the preceding paragraph, which emphasize certain undesired peculiarities about part of the results, especially regarding intelligibility, improvements may be perceived if meaningful arrangements are implemented to the assembled speech corpus. Perhaps using more numerous personalities of diverse ages, with different accents and more notable distinction in their vocal characteristics may enhance the results in future work. About SDR, despite its relevance in this work, also because it is a usual objective metric, which tends to reduce human-related failures, it may be a less robust metric for some scenarios of speech enhancement or source separation, mainly for monoaural signals, which are of the type discussed in this paper. In order to address the problems associated with this metric, work [@le2019sdr] proposed an alternative metric called SI-SDR. Thus, in a future continuation of this work, this new metric proposal can be explored. Conclusions =========== The results show that, although it is a classic technique confronted by more advanced ones, at least for the scenarios covered in this particular paper, the Wiener filter is still able to perform speech enhancement tasks for several scenarios, and remains a proper method for quality, intelligibility and distorting effects on speech signals. Despite its subtly inferior performance for some considered scenarios, SEGAN did well at 0 dB SNR scenarios, which are much more complicated, as well as exhibiting substantially lower variances. Although results obtained in [@SEGAN] indicate superiority over the Wiener filter, the divergence of results in relation to this work may be due to the wider variety of scenarios considered in this work. It is noteworthy, however, the need for a more detailed analysis of specific scenarios and also a more in-depth investigation into SEGAN. Acknowledgment {#acknowledgment .unnumbered} ============== The authors are grateful for the support received from CAPES and from the Conselho Nacional de Desenvolvimento Científico e Tecnológico – CNPq. [^1]: https://spadini.info

ArXiv

--- abstract: 'We study the evolution of solidification microstructures using a phase-field model computed on an adaptive, finite element grid. We discuss the details of our algorithm and show that it greatly reduces the computational cost of solving the phase-field model at low undercooling. In particular we show that the computational complexity of solving any phase-boundary problem scales with the interface arclength when using an adapting mesh. Moreover, the use of dynamic data structures allows us to simulate system sizes corresponding to experimental conditions, which would otherwise require lattices greater that $2^{17}\times 2^{17}$ elements. We examine the convergence properties of our algorithm. We also present two dimensional, time-dependent calculations of dendritic evolution, with and without surface tension anisotropy. We benchmark our results for dendritic growth with microscopic solvability theory, finding them to be in good agreement with theory for high undercoolings. At low undercooling, however, we obtain higher values of velocity than solvability theory at low undercooling, where transients dominate, in accord with a heuristic criterion which we derive.' --- Introduction ============ Modeling solidification microstructures has become an area of intense study in recent years. The properties of large scale cast products, ranging from automobile engine blocks to aircraft components and other industrial applications, are strongly dependent on the physics that occur at the mesoscopic and microscopic length scales during solidification. The main ingredient of the solidification microstructure is the dendrite, a snowflake-like pattern of solid around which solidification proceeds. The microscopic properties of such cast products are determined by the length scales of these dendrites, and for this reason understanding the mechanisms for pattern selection in dendritic growth has attracted a great deal of interest from the experimental and theoretical community. In particular, a great deal of research has been undertaken to understand such issues as dendrite morphology, shape and speed. Experiments on dendrite evolution by Glicksman and coworkers on succinonitrile (SCN) [@Hua81; @Gli84], and more recently Pivalic Acid (PVM) [@Mic98], as well as other transparent analogues of metals have provided tests of theories of dendritic growth, and have stimulated considerable theoretical progress[@Lan80; @LanII80; @Kes88]. These experiments have clearly demonstrated that in certain parameter ranges the physics of the dendrite tip can be characterized by a steady value for the dendrite tip velocity, radius of curvature and shape. Away from the tip the time-dependent dendrite exhibits the characteristic sidebranching as it propagates. The earliest theories of dendritic growth solved for the diffusion field around a self-similar body of revolution propagating at constant speed [@Iva47; @Hor61]. In these studies the diffusion field is found to determine the product of the dendrite velocity and tip radius, but neither quantity by itself. Adding capillarity effects to the theory predicts a unique maximum growth speed [@Tem60]. Experiments, however, have shown that these theories do not represent the correct operating state for real dendrites. The introduction of local models of solidification brought further insight to the steady state dendrite problem [@Bro83; @Ben83; @Ben84; @KessI84]. These models describe the evolution of the interface, incorporating the physics of the bulk phases into the governing equation of motion of the interface. A remarkable breakthrough of these models was to show that a steady-state dendrite velocity is obtained [*only*]{} if a source of anisotropy (e.g., in the interfacial energy) is present during dendritic evolution. The dendrite steady-state tip velocities appear in a discrete rather than continuous spectrum of values, making the role of anisotropy of great importance in the description of the dendrite problem, both in the local models and the full moving boundary problem [@Kes88; @Bre91; @Pom91]. It was further shown that only the fastest of a spectrum of steady state velocities is stable, thus forming the operating state of the dendrite. This body of theoretical work is generally known as [*microscopic solvability theory*]{}. Dendritic sidebranching is widely believed to be caused by thermal fluctuations, which enter solidification models in the form of random noise possessing specific features [@Lan87; @Kar93; @Bre95]. However, the precise mechanism of sidebranch amplification does not seem to be fully understood. It would appear that the manner in which thermal noise is amplified may depend on the overall dendrite morphology. For instance, it was shown that noisy fluctuations traveling down a paraboloid of revolution do not produce sidebrach amplitudes consistent with experiments [@Lan87], while fluctuations traveling down an initially missile-shaped dendrite amplify into sidebranches comparable to some experiments [@Bre95]. Karma also investigated the addition of interface fluctuations [@Kar93]. However, this source of noise only becomes relevant at high velocities. The theoretical foundation around which most theories of solidification are based is the time-dependent Stefan problem. This theory describes the evolution of the thermal or solutal diffusion field around the solidification front, along with two accompanying boundary conditions. The first boundary condition relates the velocity of the moving front to the difference in thermal fluxes across the solid-liquid interface. The second, called the Gibbs-Thomson condition, relates the interfacial temperature to the the thermodynamic equilibrium temperature, the local interfacial curvature and interface kinetics. The interface kinetics term adds a non-equilibrium correction to the interface temperature, usually assumed to be in local equilibrium for a given curvature. Solving the Stefan problem numerically has traditionally involved front tracking and lattice deformation to contain the interface at predefined locations on the grid [@Alm91; @Try96]. This method is generally complicated to implement accurately and requires much effort. Moreover, it can be inefficient in handling coalescence of two or more interfaces. The solution of the Stefan problem has been made more tractable with the introduction of the [*phase-field*]{} model. The phase-field model avoids this problem of front tracking by introducing an auxiliary continuous order parameter $\phi ({\bf r})$ that couples to the evolution of the thermal or solutal field. The phase field interpolates between the solid and liquid phases, attaining two different constant values in either phase, (e.g., $\pm 1$) with a rapid transition region in the vicinity of the solidification front. The level set of $\phi ({\bf r})= 0$ is identified with the solidification front, and the subsequent dynamics of $\phi$ are designed to follow the evolving solidification front in a manner that reproduces the Stefan problem [@LanS86; @Cag86a; @Col85; @War95; @Whe92; @Kar94; @Eld94; @Whe96; @Kob93; @Pro96; @Wan96]. The price to be paid for the convenience of the order parameter is the introduction of a new length scale $W$ which represents a boundary layer over which the order parameter changes sign. This distance is referred to as the interface width, and does not appear in the Stefan problem. As such, one requirement of the phase-field model is to recover the Stefan limit in a manner that is independent of the interface width as $W$ approaches some appropriate limit. considerable work has been done to relate $W$ to various parameters of the phase-field model in order to establish a mapping between the phase-field model and the Stefan problem [@Cag86a; @Kar95; @Fab97]. While the formal nature of these mappings does not seem to be very sensitive to the precise form of the phase-field model[@Fab97], different asymptotic limits of the phase-field parameters can lead to widely varying complexity in the numerical implementation. The introduction of the interface width $W$ makes the phase-field model prohibitively expensive to simulate numerically for large systems, since the grid spacing must be small enough everywhere that the phase-field model converges to the the sharp interface limit[@Cag86a; @Kar95]. Caginalp and Chen [@Cag92] showed rigorously that the phase field model converges to the sharp interface limit when the interface width (and hence the grid spacing) is much smaller than the capillary length. While this result is necessary to establish that the phase-field model does map onto the Stefan problem, the parameter values required to realize the asymptotic limit can be computationally intractable. Experimentally, the physical sizes required to contain realistic microstructures can be many times the size of the thermal diffusion length, which in turn can be orders of magnitude greater than $W$ Thus, since $\Delta x_{\rm min} < W$, computing in the limit of a $W \rightarrow 0$ does not allow one to simulate experimental systems. Recently Karma and Rappel[@Kar95] presented a different asymptotic analysis performed in powers of the ratio of the interface width to the diffusion length $\alpha/V_n$, taken to be equal in both phases. Their procedure offers two computational advantages. The first is that is allows one to simulate the phase-field model with zero interface kinetics, [*without*]{} the need to make $W \rightarrow 0$. Specifically, this limit, as well as a non-zero kinetics limit, can be simulated with an interface width $W$ (and hence the grid spacing) of order the capillary length, a much more tractable regime. Simulating solidification microstructures in the limit of zero interface kinetics is important because most experiments performed at low undercooling in materials such as succinonitrile are in this limit [@Gli84]. Karma and Rappel tested their asymptotics by comparing their simulations to the results of microscopic solvability theory, finding excellent agreement down to dimensionless undercoolings as low as 0.25. A recent extension of Karma and Rappel’s analysis by Almgren [@Alm97] also promises to allow similar asymptotics to be performed on a two-sided model of solidification [@Alm97], i.e. when the diffusivities in the solid and liquid differ, relevant in the study of directional solidification of binary mixtures. The theory of level sets [@Che97; @Mer98] has also recently re-emerged as another effective tool that shows great potential in modeling dendritic growth. While related to the phase-field model, level-set theory does not require the presence of a thin interfacial with $W$, thus greatly reducing the stringent grid requirements posed by conventional phase-field models. To date, however, level set methods have not been benchmarked with solvability theory or other theoretical prediction for Stefan problems. While expanding the horizon of solidification modeling, phase-field modeling has still been limited to small systems sizes, even when solved by adaptive algorithms [@Bra97]. The main problem is the presence of an interface region with a minimal length scale that must be resolved. For a typical microstructures grown at dimensionless undercooling of $0.1$ or less, the ratio of the system size to this minimal grid spacing can be greater that $2^{17}$. With this restriction most numerical methods will naturally fail. What is needed to go beyond this limitation is an effective adaptive technique[@Bra97; @Nee96; @Sch96; @Pro98] which dynamically coarsens the grid spacing away from the front. In this paper we present a new, computationally efficient adaptive-grid algorithm for solving a class of phase-field models suitable for the study of phase-boundary evolution. We study dendritic solidification modeled using two coupled fields, one for the order parameter and the other for the thermal field. Our algorithm effectively combines and implements ideas of adaptive-mesh refinement with ideas of dynamic data structures, allowing us to enlarge the window of large-scale solidification modeling. The outline of this paper is as follows: In section one we introduce the physical model to be examined, summarizing its properties and its various limits. In section two we present a detailed description of our algorithm. In section three we present results on CPU-scalability of our algorithm and examine issues of grid convergence and grid anisotropy on our solutions. In section four we present results of dendritic growth with and without the presence of anisotropy in the surface energy. We show that for high undercooling, dendrites grown with our method converge to tip speeds in agreement with microscopic solvability theory. At low undercooling, however, we do not find agreement with steady state solvability theory, owing to long-lived transients in the thermal field evolution. In section four we conclude and discuss our results. The Phase-Field Model ===================== We model solidification using a standard form of phase-field equations which couple a thermal field to an order parameter field $\phi$ via a double well potential [@Kar95; @Cag86a]. We begin by rescaling the temperature field $T$ by $U=c_P(T-T_M)/L$, where $c_P$ is the specific heat at constant pressure, $L$ is the latent heat of fusion and $T_M$ is the melting temperature. The order parameter is defined by $\phi$, where we define $\phi=1$ in the solid phase and $\phi=-1$ in the liquid phase. The interface is defined by $\phi=0$. We rescale time throughout by $\tau_o$, a time characterizing atomic movement in the interface. Length is rescaled by $W_o$, a length characterizing the liquid–solid interface. With these definitions, the model is written as $$\begin{aligned} &&\frac{\partial U}{dt} = D \nabla^2 U + \frac{1}{2} \frac{\partial \phi}{\partial t} \label{phase-field}\\ \nonumber &A^2(\vec{n})& \frac{\partial \phi}{dt} = \nabla \cdot (A^2(\vec{n}) \nabla \phi ) + g^{'}(\phi) - \lambda U P^{'}(\phi) \\ \nonumber & + & \frac{\partial }{\partial x} \left( |\nabla \phi|^2 A(\vec{n}) \frac{\partial A(\vec{n})}{\partial \phi_{,x}} \right) + \frac{\partial }{\partial y} \left( |\nabla \phi|^2 A(\vec{n}) \frac{\partial A(\vec{n})}{\partial \phi_{,y}} \right), \end{aligned}$$ where $D=\alpha \tau_o/W_o^2$ and $\alpha$ is the thermal diffusivity. The function $f(\phi,U;\lambda)=g^{\prime }(\phi) - \lambda U P^{\prime}(\phi)$ is the derivative of the double-well potential with respect to $\phi$ and couples the $U$ and $\phi$ fields via the constant $\lambda$. The primes on the functions $g(\phi)$ and $P(\phi)$ denote derivatives with respect to $\phi$. Following Karma and Rappel [@Kar95], anisotropy has been introduced in Eqs. (\[phase-field\]) by defining the width of the interface to be $W(\vec{n})=W_o A(\vec{n})$ and the characteristic time by $\tau(\vec{n})=\tau_o A^2(\vec{n})$, with $A(\vec{n}) \in [0,1]$ given by $$A(\vec{n}) = (1- 3 \epsilon) \left[ 1 + \frac{4 \epsilon }{ 1 - 3 \epsilon} \frac{(\phi_{,x})^4 + (\phi_{,y})^4 }{| \nabla \phi|^4}\right]. \label{width}$$ The vector $$\vec{n}=\frac{\phi_{,x} \hat{x}+\phi_{,y} \hat{y} }{(\phi_{,x}^2 + \phi_{,y}^2)^{1/2} } \label{normal}$$ defines the normal to the contours of the $\phi$ field, where $\phi_{,x}$ and $\phi_{,y}$ are defined as the partial derivatives of $\phi$ with respect to $x$ and $y$. The variable $\epsilon$ parameterizes the deviation of $W(\vec{n})$ from $W_o$ and represents the anisotropy in the interface energy of the system. We note that this definition of anisotropy is not unique [@Fab97], but we expect results to be similar for other definitions of anisotropy. In simulating the phase-field model we adopt the point of view that the order parameter field $\phi$ is a computational tool whose main purpose is to eliminate front tracking. As such we would like to simulate the model given by Eqs. (\[phase-field\]) with $W_o$ as large as possible. At the same time we would like the behavior of the model outside the boundary layer defined by $\phi$ to describe the Stefan problem as closely as possible. To this end, we relate the parameters of the phase-field model according to Ref. [@Kar95], valid in the asymptotic limit $W_o \ll \alpha/V_c$, where $\alpha/V_c$ is the diffusion length and $V_c$ is a characteristic velocity of the front defined by $\phi$. The specific asymptotic limit we model is one where the $U$-field satisfies $$\frac{\partial U}{\partial t} = D \nabla^2 U \label{stefan1}$$ everywhere away from the interface, while at the interface, the gradient of $U$ satisfies $$V_{n} = D\left( \left. \frac{\partial U}{\partial \vec{n}} \right|_{\vec{x}_{\rm int}^{-}} - \left. \frac{\partial U}{\partial \vec{n}} \right|_{\vec{x}_{\rm int}^{+}} \right), \label{continuity}$$ where $V_{\rm int}$ is the velocity normal to the interface, denoted by $\vec{x}_{\rm int}$. The notation $\pm$ denotes the solid/liquid side of the interface, respectively. The description of the Stefan problem is completed by the Gibbs-Thomson condition and the interface kinetics condition $$U(\vec{x}_{\rm int}) = -d(\vec{n}) \kappa - \beta(\vec{n}) V_n, \label{gibb-thomson}$$ where $d(\vec{n})$ is the capillary length, $\kappa$ is the local curvature and $\beta(\vec{n})$ is the interface attachment kinetic coefficient, all assumed to be in dimensionless form according to the above definitions. The capillary length is related to the parameters of Eqs. (\[phase-field\]) by $$d(\vec{n}) =a_1 \frac{W_o}{\lambda} \left[ A(\vec{n}) + \partial_{\theta}^2 A(\vec{n}) \right] \label{cap_length}$$ where $a_1=0.8839$ for the particular form of the free energy defined in Eqs. (\[phase-field\]) [@Kar95] and $\theta$ is the angle between $\vec{n}$ and the $x$-axis. The kinetic coefficient is given by $$\beta=\frac{a_1 \tau_o}{ \lambda W_o } \left[ 1 - \frac{\lambda a_2}{ D}\right] \label{beta}$$ where $a_2=0.6267$ for our choice of the free energy functional cite[Kar95]{}. One remarkable feature of Eqs. (\[cap\_length\]) and (\[beta\]) is that $W_o$, $\tau_o$ and $\lambda$ an be chosen to simulate arbitrary values of $\beta$, for $W_o$ of order $d_o$. In particular, setting $\lambda=D/a_2$ allows us to compute the phase-field model in the limit of the Stefan problem [@Kar95], where $\beta=0$. This is also an appropriate value for SCN, especially at low undercooling. Equations (\[cap\_length\]) and (\[beta\]) for $\beta$ and $d_o$ can be related to a wide class of free energies via the parameters $a_1$ and $a_2$ [@Kar95], which are related to integrals that depend on $g(\phi_o)$, $P(\phi_o)$ and $d \phi_o/dx$, where $\phi_o$ is the lowest order description of the order parameter field $\phi$ and and is a solution of the equation $$\frac{\partial^2 \phi_o}{\partial x^2} - \frac{ d g(\phi_o)}{d \phi_o} =0. \label{phi_o}$$ We also note that these asymptotics are a special case of a more general asymptotic analysis performed by Almgren [@Alm97], which relates the parameters of the phase-field model to those of the Stefan problem in the case of unequal diffusivities in the solid and liquid phases. In this case, the asymptotics provides an additional set of constraints on the admissible functions $P^{\prime }(\phi)$, $g^{\prime}(\phi)$, and hence $a_1$ and $a_2$. The Adaptive-Grid Algorithm =========================== The main computational challenge of simulating Eqs. (\[phase-field\]) involves resolving two competing length scales: the lattice spacing $dx$ on which the simulation in performed and the physical size of the system $L_B$. Even with improved asymptotics, $dx$ must remain relatively small, while $L_B$ must be extremely large in order to make possible computations of extended solidification microstructures. Moreover, the main physics of solidification (and the evolution of most phase-boundary problems) occurs around an interface whose area is much smaller than the full computational domain. Near this interface the order parameter varies significantly, while away from the interface variations in $\phi$ are small. Meanwhile, the thermal field $U$ extends well beyond the interface and has much more gradual variation in its gradients, permitting a much coarser grid to be used to resolve $U$. The most obvious manner to overcome this problem is to use a method that places a high density of grid points where the interface of $\phi$ or $U$ varies most rapidly and a much lower grid density in other regions. Furthermore, the method must dynamically adapt the grid to follow the evolving interface [@Bra97; @Nee96; @Sch96; @Pro98], while at the same time maintaining a certain level of solution quality. We solve Eqs. (\[phase-field\]) using the Galerkin finite element method on dynamically adapting grids of linear, isoparametric quadrilateral and triangular elements. The grid is adapted dynamically based on an error estimator that utilizes information from both the $\phi$ and $U$ fields. We wrote our code in FORTRAN 90 (F90), taking advantage of the efficiency of FORTRAN 77 while using advanced C-like features, such as data structures, derived data types, pointers, dynamic memory allocation and modular design to conveniently adapt the grid and the solution fields. In the broadest sense, our algorithm performs functions that can be divided into two classes. The first deals with the establishment, maintenance and updating of the finite element grids, and the second with evolving $\phi$ and $U$ on these grid, according to Eqs. (\[phase-field\]). We presently describe these classes, the adaptive grids, the finite element procedure, and the interconnections of these processes. The Finite Element Grids ------------------------ The first class of functions in our algorithm centers around maintaining a grid of finite elements on a data structure known as a [*quadtree*]{} [@Dev87; @She88; @Pal96] which replaces the standard concept of a uniform grid as a way of representing the simulational grid. The quadtree is a tree-like structure with branches up to a prespecified level. Branches of the quadtree are themselves data structures that contain information analogous to their parent, from which they branched, but one level down. Fig. \[quadtree2\] illustrates the structure of a quadtree as well as the relation between elements at different levels of refinement. Every entry on the quadtree contains information pertaining to a $4$-noded isoparametric quadrilateral finite element. This information includes the following:\ $\bullet$ values of $\phi$ and $U$ at the four nodes\ $\bullet$ the nodal coordinates of the element\ $\bullet$ the level of refinement of the element on the quadtree\ $\bullet$ the value of the current error estimate\ $\bullet$ The element number, which contains information about the coordinates of the element and its level of refinement\ $\bullet$ an array mapping the element’s four nodes onto the entries of a global solution array\ $\bullet$ pointers to the element’s nearest neighbors sharing a common edge and at the same level of grid refinement\ $\bullet$ a variable that determines whether or not an element contains further sub-elements which we term [*child*]{} elements\ $\bullet$ pointers to an element’s child elements\ $\bullet$ a pointer to the [*parent*]{} element from which an element originates\ The elements of the quadtree can be refined by splitting into four child elements, each sharing the same parent element one level down on the quadtree and each with its own unique information, as outlined above. A parent element and it’s four child elements are referred to as a [*family*]{}. Refinement produces a finer mesh within the confines of the original parent grid by bisecting each side, as shown in Fig. (\[quadtree2\]). Unrefinement, which consists of fusing the four child elements back into the parent, has the opposite effect, locally creating a coarser mesh. Both refinement and unrefinement proceed via dynamic memory allocation, making our code scalable. We note that unrefinement can occur only if the child elements do not possess further child of their own. Also, in order to avoid having regions of very different refinement bordering each other, we impose the restriction that any two neighboring quadrilateral elements may be separated by no more than one level of refinement (see Fig. (\[quadtree2\])). We define the level of refinement of an element by $l_e$ such that a uniform grid at a refinement level $l_e$ would contain $2^{l_e} \times 2^{l_e}$ grid points in a physical domain $L_B \times L_B$. Special cases where an element has no children, a missing neighbor, or no parent are handled by null pointers. The latter case occurs only for the root of the quadtree. All elements at a given level of refinement on the quadtree are “strung" together by a linked-list of pointers, referred to as the [*$G$-lists*]{}. There are as many $G$-lists as there are levels of refinement in the quadtree. Each pointer in the $G$-list points to (accesses) the location in memory assigned to one element of the quadtree. The purpose of the $G$-list is to allow traversal of the quadrilateral elements sequentially by level, rather than by recursively traversing quadtree from the root down, a procedure which is memory intensive and relatively slow. Alongside the main grid of quadtree elements, the code maintains two independent grids representing special linear isoparametric triangular and rectangular elements. These elements are used to connect the extra nodes that arise when two or more quadrilateral elements of differing refinement levels border each other. These element types are referred to as [*bridging elements*]{}. They are maintained as two linked-lists of derived data types, one containing information about triangular elements and the other rectangular. Elements of both these grids include the following information:\ $\bullet$ the values of $\phi$ and $T$ at the three nodes (four for rectangles) of the element\ $\bullet$ the nodal coordinates\ $\bullet$ node numbers that map the element’s nodes onto the global solution array\ The types of bridging triangles and rectangles that can occur are enumerable and shown in Fig. (\[tri\_recs\_configs\]). The main set of operations performed on the grids described above concern refinement of the finite element mesh as a whole. The refinement process is performed only on the quadrilateral mesh. The triangular and rectangular grids are established after this process is completed (see Fig. (\[quadtree2\])). To refine the grid the code traverses the elements of the quadtree, refining (unrefining) any element whose error function, discussed below, is above (below) a critical value $\sigma_h$($\sigma_l$). We also note that fusion of four quadrilateral elements can occur only if all four of its children’s error functions are below the critical value $\sigma_l$, where $\sigma_l < \sigma_h$. We found that if $\sigma_l = \sigma_h$ the grid sets into oscillations, where large numbers of elements become alternatively refined at one time step, then unrefined at the next. The processes described thus far are grouped into modules that encapsulate various related tasks, and which can cross-reference each other’s data and instructions. The module highest up in the hierarchy contains the definition of the quadtree data structure and routines that construct the initial uniform grid, refine and unrefine individual quadrilateral elements, and set the initial conditions. Another module constructs the G-lists. It contains routines that construct the initial G-list from initial uniform quadtree data structure, as well as add or delete element pointers from the $G$-list as elements are created or deleted from the quadtree. Another module accessing both the previous ones’ data structures has the role of creating the triangular and rectangular element grids. It contains definitions for creating triangular and rectangular elements data structures and routines that search the quadtree, building the linked lists of triangles and rectangles that make up these grids. The main program is contained in its own module and contains the driver program that creates the initial grids, $G$-lists and triangular and rectangular element types. The driver program also sets into motion the final link in the simulation, which evolves $\phi$ and $U$ and periodically adapts the dynamic grid by calling procedures described above. A flowchart of these processes is shown in Fig. (\[flow\_chart\]). The Finite Element Formulation ------------------------------ The integration of Eqs. (\[phase-field\]) is done by the final module in the code. This module performs four main processes:\ 1. Maps the internal element node numbers to the indices of a global solution vector. The $\phi$-field is mapped onto the odd numbers, ($1,3,5, \dots$), while $U$ is stored on the even numbers of the global solution vector ($2,4,6, \dots$)\ 2. Advances the $U$ and $\phi$ field-vectors by $N_r$ time steps on the finite element grids defined above\ 3. Calculates an error function for each element of the quadtree, based on error estimate of the quadrilateral elements\ 4. Invokes routines in the modules described above to refine the grid according to this error estimator\ Steps (1)-(4) are repeated until a sufficient time evolution of the microstructure is established. The variable $N_r$ is set such that the interface remains within the regions of fine mesh between regriddings, which we typically choose to be 100 time steps. Step (1) involves searching all elements, and their neighbors, and assigning each node a unique number, that will have a counterpart on a global solution vector. The finite element discretization of Eqs. (\[phase-field\]) is done using Galerkin’s weighted residual method [@Coo89]. The method begins by assuming that $\phi$ and $U$ are interpolated within an element as $$\phi^e = \sum_{i=1}^{N} \phi_i^e N_i(x,y) \hspace{0.5cm} U^e = \sum_{i=1}^{ N} U_i^e N_i(x,y) \label{interpol_fields}$$ where $\phi_i^e$ and $U_i^e$ are the field values at the $N$ nodes of the element $e$, and their interpolated values in its interior. The functions $N_i(x,y)$ are standard linear interpolation functions appropriate to the element being used [@Zie87], and satisfy $$N_i(x_j,y_j)=\delta_{i,j}, \label{shape_funcs_nodes}$$ where $\delta_{i,j}$ is the Kroneker delta. Rewriting the differential equations for $\phi$ in Eqs. (\[phase-field\]) as $L_{\phi} \phi =0$, and of the $U$-equation as $L_U U=0$, the Galerkin method requires that $$\begin{aligned} &&\int_{\Omega_e} N_i(x,y) L_{\phi} \phi^e(x,y) dx dy = 0 \label{Gal_state} \\ \nonumber &&\int_{\Omega_e} N_i(x,y) L_{U} U^e(x,y) dx dy = 0,\end{aligned}$$ for $i=1,2,3,\dots, N$, where $\Omega_e$ represents the area of an element $e$. Substituting Eqs. (\[interpol\_fields\]) into Eqs. (\[Gal\_state\]), we obtain two linear algebraic equations for $\phi_i$ and $U_i$, $i=1,2,3, \dots, N$ in the element $e$. We next define $\{\Phi\}^e=(\phi_1,\phi_2,\phi_3, \cdots, \phi_{N})^T$ and $\{U\}^e=(U_1,U_2,U_3, \cdots, U_{N})^T$, where the superscript $T$ denotes transpose, making $\{\Phi\}^e$ and $\{U\}^e$ column vectors. The linear algebraic statement of the finite element form of Eqs. (\[phase-field\]) then becomes $$\begin{aligned} &&[\hat{C}](\{ \phi \}^e) \frac{d \{ \phi^e\}}{dt} = \left([M]+[E]\right) \{\phi\}_{n}^e + \{F; \lambda\}^e \label{Mat_eqs}\\ \nonumber &&[C] \frac{d \{U\}^e}{dt} =D [A] \{U\}^e + \frac{1}{2} [C] \frac{d \{ \phi\}^e }{dt},\end{aligned}$$ where the matrices $[C]$, $[\tilde{C}]$, $[A]$, $[M]$ and $[E]$ and the vector $\{F; \lambda\}^e$ are given by $$[C] = \int_{\Omega_e} [N]^T [N] dx dy, \label{C_mat}$$ $$[\hat{C}] = \int_{\Omega_e} [N]^T [N] A^2(\theta (\phi^e)) dx dy, \label{Chat_mat}$$ $$[A] = - \int_{\Omega_e} \left( [N]^T [N_x] + [N]^T [N_y] \right) dx dy, \label{A_mat}$$ $$[M] = - \int_{\Omega_e} \left( [N]^T [N_x] + [N]^T [N_y] \right) A^2(\theta(\phi^e )) dx dy, \label{M_mat}$$ $$[E] = - \int_{\Omega_e} \left( [N]^T [N_x] - [N]^T [N_y] \right) A(\theta(\phi^e)) \omega(\theta(\phi^e)) dx dy, \label{E_mat}$$ $$\{F;\lambda\}^e = \int_{\Omega_e} [N]^T f(\phi^e, U^e ;\lambda) dx dy, \label{F_mat}$$ where $[N_x]$, $[N_y]$ denote the partial derivatives of the vector of shape functions with respect to $x$ and $y$, respectively. The function $A$ is just Eq. (\[width\]) rewritten in terms of the angle $\theta$ that the normal to the interface makes with the $x$-axis. Specifically, defining $$\tan \theta({\phi}^e) = \frac{\partial \phi_{,y}^e}{\partial \phi_{,x}^e}. \label{theta_def}$$ then $$A(\theta({\phi}^e))=(1-3\epsilon) \left[ 1 + \frac{4 \epsilon}{1 - 3\epsilon} \frac{(1+ \tan^4 \theta)}{(1+ \tan^2 \theta)^2} \right] \label{alternate_A}$$ while $\omega(\theta)$ is proportional to the derivative of $A(\theta)$, and is given by $$\omega(\theta({\phi}^e)) = 16 \epsilon \frac{\tan \theta (1 - \tan^2 \theta)}{(1+ \tan^2 \theta)^2}, \label{W_prime}$$ We use a lumped formulation for the matrices $[C]$ and $[\hat{C}]$ [@Coo89]. In this procedure, the row vector of shape functions, $[N]$ in Eq. (\[C\_mat\]) is replaced by the identity row vector $[I]=[1,1,1,\cdots]$. The resulting matrix $[C]$ then consists of identical columns, each of which contains the element $N_i(x,y)$ in the position of the $i^{\rm th}$ row. A lumped term is then defined as a diagonal matrix whose entries take on the value $$L_c= \frac{1}{4} \sum_{i=1}^{\rm nodes} \int_{\Omega_e} N_i(x,y) dx dy. \label{lump_def}$$ The use of a lumped matrix for $[C]$ allows us to assemble a diagonal matrix for the left hand side Eqs. (\[Mat\_eqs\]), stored as a one-dimensional vector rather than two-dimensional matrices that would be required if we used the consistent formulation for the assembly of the $[C]$ matrices. Indeed, microstructures evolving at low undercooling can produce interfaces with over $2 \times 10^5$ elements, making the storing of $2 \times 10^{10}$ matrices impossible. The global $\{\phi\}$ (obtained after assembly of the element equations in field in Eqs. (\[Mat\_eqs\])) is time-stepped using using a forward difference (explicit) time scheme. For each time step of the $\phi$ field, the global $U$ field is then solved iteratively using a Crank-Nicholson scheme. Convergence of $\{U\}_{n+1}$ is obtained in a few iterations. The Error Estimator ------------------- Regridding is based on an error estimator function, which is obtained following Zienkiewicz and Zhu [@Zie87], based on the differences between calculated and smoothed gradients of the $\phi$ and $U$ fields. Specifically, we define the [*composite field*]{} $$\Psi = \phi + \gamma U \label{composite_def}$$ where $\gamma$ is a constant. We discuss the selection of $\gamma$ in more detail below. This definition allows us to regrid according the requirements of both the $\phi$ and $U$ field, as opposed to using only gradients of the $\phi$-field in establishing the grid [@Bra97]. Since it is $\phi$ and $U$ that are being calculated, and not their gradients, we do not expect the gradient of $\Psi$ to be continuous across element boundaries, due to the order of the interpolation used. Thus we expect the difference between the calculated and smoothed (continuous across element boundaries) gradients to provide a reasonable estimate of error. This method appropriately meshes regions of both steep gradients and regions where the $\phi$ and $U$ fields change rapidly. We define the error estimator function $\vec{e}$ as $$\vec{e} = \vec{q}_{\rm s} -\vec{q}_{\rm c} \label{error_def}$$ where $\vec{q}_{\rm c}$ and $\vec{q}_{\rm s}$ are the calculated and smoothed gradients of $\Psi$ respectively. The smoothed gradients are calculated to be continuous across element boundaries. To determine $\vec{q}_{\rm s}$ we assume it to be interpolated in the same way as the $\phi$ and $U$ fields, namely $$\vec{q}_{\rm s} = [N]\{Q^s\} \label{q_smooth_int}$$ where $[N]$ is the row vector of element shape functions, and $\{Q^s\}$ is a $4 \times 2$ matrix whose columns represent the nodal values of fluxes of $\Psi$ in the $x$ and $y$ direction, respectively. To find $\{Q^s\}$ we use Galerkin’s method, minimizing the weighted residual $$\int_{\Omega_e} [N]^T \vec{e} d\Omega_e=\int_{\Omega_e}[N]^T([N]\{Q^s\}-\vec{q}_{\rm c})d\Omega=0 \label{error_residual}$$ The calculation is simplified by lumping the left hand side of Eq. (\[error\_residual\]), leading to $$\left( \int_{\Omega_e} [N]^T [{\bf 1}] d\Omega \right)\{Q^s\} = \int_{\Omega_e} [N]^T \vec{q}_{\rm c} d\Omega, \label{error_matrix}$$ where $[{\bf 1}]=[1,1,1,\cdots,N]$. Assembling Eq. (\[error\_matrix\]) for all quadrilateral elements yields an equation for the smoothed gradients $\{Q\}^g$ of the global field $\Psi$, at all element nodes, of the form $$[D]\{Q\}^g = {b}, \label{error_matrix_short}$$ where $[D]$ is a diagonal matrix, due to “mass" lumping, and $\{Q\}^g$ is a $N \times 2$ matrix for the global, smoothed flux. For the actual error updating on the elements of the quadtree we used the normalized error defined by $$E_e^2 =\frac{ \int_{\Omega_e} |( \vec{q}_{\rm s} - \vec{q}_{\rm c} )|^2 } { \sum_e \int_{\Omega} | \vec{q}_{\rm s} |^2 }. \label{norm_error}$$ The domain of integration $\Omega$ in the denominator denotes the entire domain of the problem. Thus $E_e^2$ gives the contribution of the local element error relative to the total error calculated over the entire grid. Fig. (\[dendrite\_picture\]) shows a snapshot at $10^5$ time steps into the a simulation of a thermal dendrite computed with our algorithm. The figure shows $\phi$ and $U$ as well as the current grid. The dendrite is four-fold symmetric, grown in a quarter-infinite space, initiated by a small quarter disk of radius $R_o$ centered at the origin. The order parameter is defined on an initially uniform grid to be its equilibrium value $\phi_o(\vec{x})=-\tanh((|\vec{x}|-R_o)/\sqrt{2} )$ along the interface. The initial temperature decays exponentially from $U=0$ at the interface to $-\Delta$ as $\vec{x} \rightarrow \infty$. The parameters set for this simulation are $\Delta=0.70$, $D=2$, $dt=0.016$ and $\lambda$ chosen to simulate $\beta=0$. The system size is $800 \times 800$, with $\Delta x_{\rm min}=0.4$, and about half of the computational domain in each direction is shown. Sidebranching is evident, and arises due to numerical noise. This simulation was completed in approximately 15 cpu-hours on a Sun UltraSPARC 2200 workstation. Scalability and Convergence Properties of the Adaptive-Grid Algorithm ===================================================================== In this section we present results that illustrate the convergence properties of solutions of Eqs. (\[phase-field\]) computed with our algorithm, the effect of grid-induced anisotropy of the adapting mesh, and the speed increase obtained by using an adapting grid. CPU-Performance --------------- We examined the cpu-scalability of our algorithm as a function of system size by growing dendrites in systems of various linear dimension $L_B$ and measuring the cpu time required for the dendrite branches to traverse the entire system. Fig. \[cpu\] shows a plot of these data for a dendrite grown at undercooling $\Delta=0.55$ using the same parameters as in Fig. \[dendrite\_picture\]. The minimum grid spacing has been set to $\Delta x_{\rm min}=0.4$ in this data. Fig. \[cpu\], clearly shows that $R_t^a \sim L_B^2$. This relationship can be obtained analytically as follows. The number of calculations performed, per simulation time step, is proportional to the number of elements in the grid. This relationship is set in turn by the arclength of the interface being simulated multiplied by the diffusion length $D/V_n$. This product defines the arclength over which the highest level of refinement occurs. For a needle-like dendrite, the arclength is approximately $L_B$. Moreover, since the dendrite tip moves at a constant velocity $V_n$, then $$R_t^a = \left[ \frac{R_o^a D }{V_n^2 \Delta x_m^2} \right] L_B^2, \label{cputime}$$ where $R_o^a$ is a constant that depends on the details of the implementation of the algorithm used to evolve Eqs. (\[phase-field\]). The cpu time needed to compute the traversal time on a uniform grid, $R_t^u$, is found, by the same analysis, to be $$R_t^u =\left[ \frac{R_o^u }{ V_n \Delta x_m^2} \right] L_B^3. \label{time_uniform}$$ where $R_o^u$ also depends on the implementation but is likely to be smaller than $R_o^a$. Thus, comparing our method with simulation on a uniform grid we obtain $$\lim_{L_B \rightarrow \infty} R_t^a/R_t^u = \frac{1}{L_B}.$$ For larger systems, the adaptive scheme should always provide faster CPU performance regardless of implementation. Indeed, any method that uses a uniform grid of any sort, will eventually be limited by memory requirements as $L_B$ becomes large. The arguments leading to Eq. (\[cputime\]) can also be generalized to any problem of evolving phase boundaries, always yielding the conclusion that cpu time scales with arclength in the problem being considered. We note that when interface convolutions become of order $\Lambda \sim \Delta x_{\rm min}$, fine-grid regions separated by less than $\Lambda$ will merge and the number of elements will stop growing locally. This makes the simulation of fractal-like patterns feasible as the arclength of the interface is bounded from above by $L_B \times L_B$. Finally, we note that adaptive gridding would especially improve the cpu performance of problems similar to spinodal decomposition, where the total interface decreases with time. Induced Lattice Anisotropy -------------------------- We tested the effective anisotropy of our dynamically adapting lattice in two independent ways. The first follows the method outlined by Karma[@Kar95]. We fix the temperature far from the interface to be constant $T_{\infty }$ everywhere, initially setting it to a critical value at which the isotropic surface energy just balances the bulk free energy. For a specified background temperature, the crystal will only grow if its radius is greater than a critical value $R_o$. The radius $R_o$ can be related to the background temperature through the total Gibbs free energy of the system, given by $$\Delta G =- \pi r^2\frac{L \Delta T}{T_M} + 2\pi r \sigma, \label{Gibbs_free_energy}$$ where $L$ is the latent heat of fusion, $\Delta T=T_m-T_{\infty}$, where $T_m$ is the melting temperature, $T_{\infty}$ is the temperature far away from the interface, and $\sigma$ is the surface tension. Minimizing $\Delta G$ with respect to $r$ yields $R_o$ as a function of $\delta T$ as $$R^* = d_o/\Delta T \label{critical_radius},$$ where $d_o$ is the capillary length defined as $d_o=2 \sigma T_M/L$. One finds an equilibrium shape of the interface when the background temperature field $\Delta T$ (written in terms of $U$) is adjusted dynamically so as to maintain the velocity of the interface at zero as measured long the $x$-axis. Thus, $\Delta T$ is increased if the velocity decreases, and decreased if it grows. The effective anisotropy is inferred by fitting the computed interface to an equation of the form $$R(\theta) = R_o(1 + \epsilon_{\rm eff} \cos \theta), \label{anisotropy_func}$$ where $R(\theta)$ is the radial distance from the center of the crystal to its interface and $\theta$ the polar angle. The effective anisotropy $\epsilon_{\rm eff}$ represents the modification of the anisotropy $\epsilon$ due to the grid. Fig. (\[aniso\_int\]) illustrates a crystal grown to equilibrium using an input anisotropy $\epsilon=0.04$. Using Eq. (\[anisotropy\_func\]) we found $\epsilon_{\rm eff}=0.041$, within 5% of $\epsilon$. Similar accuracy was found for $\epsilon=0.02, 0.03$ and $0.05$. We also tested for grid anisotropy by rotating the grid by $45$ degrees, which should represent the lowest accuracy for square elements. We compared the tip speed of dendrites grown in this direction to that of dendrites whose principal growth direction is along the x-axis. Fig.  \[conv\_vel0.55\] shows the tip velocity for the case of a dendrite grown at $\Delta=0.55$ ($\epsilon=0.05$, $D=2$, $\beta=0$, $dt=0.016$, $\Delta x_{\rm min}=0.4$) compared with the same case when growth occurs along the $x$-axis. The tip velocity approaches an asymptotic value that is within approximately $5\%$ of the tip velocity computed when the anisotropy is aligned with the $x$-direction. Convergence and Grid Resolution ------------------------------- We tested the convergence of solutions as a function of the minimum grid spacing $\Delta x_{\rm min}$. We used an undercooling of $\Delta=0.55$, with $D=2$, $dt=0.016$, $\Delta x_{\rm min}=0.4$, and set $\lambda$ to simulate $\beta=0$. The parameter $\gamma=1.8$, which assured that regions of rapid change of $\phi$ and $U$ were always encompassed in the regions of highest grid resolution. We examined the tip speed of a dendrite for $0.3 \le \Delta x_{\rm min} \le 1.6$, finding relatively good convergence of the tip speed to theoretical prediction of microscopic solvability theory discussed above. Fig. (\[rich\]) shows the asymptotic steady state tip velocity for each case, superimposed on the solid line, which is the result of solvability theory for $\Delta=0.55$. It is surprising that the solution convergence is rather good even for $\Delta x_{\rm min} =1.6$. We have found similar convergence properties for the case of $\Delta=0.25$. Specifically, using $\Delta x_{\rm min}=0.4$ and $\Delta x_{\rm min}=0.78$ gives essentially identical results. The introduction $\gamma$ in the error function $\Psi$ gives us the freedom to tune the degree to which the fine grid layering encompasses the thermal field as well as the $\phi$ field. Setting $\gamma=0$ leads to a uniform-like mesh at the highest level of refinement [*only*]{} around the most rapidly changing regions of $\phi$, while the $U$ field becomes encompassed in a rather disorderly combination of quadrilateral and triangular elements. We found that this effect can increase the tip-speed error by several percent, as well as increase fluctuations in tip speed. Increasing $\gamma$ produces a smooth layering of coarser uniform-like meshes ahead of the $\phi$-field, corresponding to region of large gradients in $U$. Fig. \[grid\_layer\] compares the mesh around the tip of a dendrite grown at $\Delta=0.65$ for $\gamma=0$ and $\gamma=4$. The figure illustrates the gradual mesh layering encompassing the thermal field for $\gamma=4$. In Fig. (\[grid\_layer\]), $D=1$, $dt=0.016$, $\Delta x_{\rm min} =0.4$ and $\lambda$ is chosen to simulate $\beta=0$. Fig. \[conv\_vel0.55\] also shows the tip speed for $\Delta=0.55$ for the cases $\gamma=0$ and $1.8$, while Fig. (\[conv\_vel0.3\]) shows the tip velocity for a dendrite grown at $\Delta=0.3$ with $\gamma=0$ and $20$, respectively. In Fig. (\[conv\_vel0.3\]) $D=10$, $\Delta x_{\rm min}=0.4$, $dt=0.048$ and $\beta=0$. In this case the higher value for $\gamma$ allows the tip velocity to approach within approximately 5% of the solvability answer, as in Ref. [@Kar95]. Raising $\gamma$ further does not produce any further changes in tip speed. Dendritic Growth using Adaptive Gridding ======================================== In this section we present results for two dimensional solidification with and without interface anisotropy. We illustrate the robustness of our algorithm and use it, in particular, to investigate dendritic growth at low undercooling, presenting new results on dendrite tip-speed selection. Dendritic Growth without Surface Tension Anisotropy --------------------------------------------------- When the anisotropy parameter $\epsilon$ in Eqs. (\[phase-field\]) is set to zero, solidification proceeds without the emergence of any preferred direction. In this case it is well known that a seed crystal larger than a critical radius will eventually grow to become unstable to fluctuations, and will break into surface undulations via the Mullins-Sekerka instability [@Try96]. Fig. \[disk\_growth\_grids\] shows a series of time steps in the evolution of a solidifying disk grown at $\epsilon=0$, $\Delta=0.65$, $D=4$, and $\lambda$ set to generate $\beta=0$, making $d_o=0.1385$. We use 11 levels of refinement and an $800 \times 800$ system, making $\Delta x_{\rm min} =0.4$. For coarser meshes the Mullins-Sekerka instability sets in sooner due to grid noise. As $\Delta x_{\rm min}$ is made smaller, grid noise becomes smaller, and one must wait longer for the true “thermal noise" to set in and make the crystal interface unstable. The dynamically evolving grids are also shown. This figure clearly demonstrates how the grid creation scales with the arclength of the solidifying surface. Dendritic Growth with Surface Tension Anisotropy ------------------------------------------------ When surface tension anisotropy is present a crystallizing disk forms dendritic branches which travel along the symmetry axes of the anisotropy, driven by the anisotropy to a steady state tip velocity [@Kes88; @Ben83; @Ben84; @KessI84; @Bre91; @Pom91]. As a verification of our algorithm we measured tip-velocities and shapes for dendrites grown at intermediate undercoolings. Fig. \[sys\_vel\] shows the dimensionless tip velocity ($V d_o/D$) versus time for $\Delta=0.45$ and $0.65$ and $\epsilon=0.05$. In Fig. (\[sys\_vel\]) the dimensionless diffusivities $D=3$ and $1$ and the dimensionless capillary length are $d_o=0.186$ and $0.544$, respectively. In both cases $\lambda$ has been set to simulate $\beta=0$ kinetics at the interface, while $\gamma=4$ and $1.8$, respectively. These values of $\gamma$ are chosen so as to minimize grid-layering error. The solid horizontal lines represent the theoretical values obtained from microscopic solvability theory. In all cases the converged velocities are within a few percent of the theoretical prediction. The case of $\Delta=0.65$ includes data for three systems sizes. These results of system size are typical for intermediate $\Delta$, showing a relatively rapid leveling to an asymptotic speed to within a few percent of that predicted by solvability theory. Fig. (\[high\_delta\_shapes\]) also shows a plot of the dendrite tip shapes produced by our simulations, superimposed on the shapes predicted by solvability theory. Dendritic Growth at Low Undercooling ------------------------------------ At lower undercooling we encounter significant finite-size effects which cause the tip velocity to deviate from the solvability prediction. Fig. \[low\_delta\_vel\] shows the evolution of the tip-velocity for $\Delta=0.25$ in two different system sizes. For a system of size $L_x=6400 \times L_y=400$, the velocity goes to within a few percent of the solvability prediction. For a system size $6400 \times 3200$ the tip velocity seems to settle close to a value that exceeds the solvability prediction by $8\%$. This effect is even larger at $\Delta=0.1$, also shown in Fig. (\[low\_delta\_vel\]), where the tip speed approaches a value about $3$ times larger than that predicted by solvability theory. To understand this finite-size dependence of tip velocity at low undercooling, we note that at low $\Delta$, the thermal fields of the two dendrite branches overlap, producing a thermal envelope very different from that which emerges for the single, isolated dendrite branch assumed in solvability theory. At large undercooling, each dendrite arm quickly outruns the other’s thermal boundary layer, and solvability theory should apply, as is seen in Fig. (\[dendrite\_picture\]) where $\Delta=0.7$. The conditions of solvability theory can be better approximated at lower undercooling if simulations are performed in a domain which is small in one direction. For the simulation performed with $\Delta=0.25$ in a small box ($6400 \times 400$), the branch in the y-direction is extinguished by its interaction with the wall and the velocity quickly approaches the solvability prediction. However, when both branches are present, as in the simulation with $\Delta=0.25$ in the larger box ($6400 \times 3200$), their interaction leads to an increased tip-velocity because the dendrites are embedded in a circular rather than parabolic diffusion field. This is also clearly seen for a dendrite growing at $\Delta=0.1$ in Fig. (\[delta\_0.1\_pic\]), where the dendrite shape and its associated field are shown for $\Delta=0.1$ ($D=13$, $d_0=0.043$, $\epsilon=0.05$, $\Delta x = 0.78$, $dt = 0.08$). The dendrite arms never became free of each other in this simulation, causing the observed deviation from solvability theory shown in Fig. (\[low\_delta\_vel\]). We note that to avoid having the thermal field feel the effect of the sides of the box we perform our simulations in computational domains for which $L_x \sim (5-10)D/V_n$. To meet this criterion the simulation for $\Delta=0.1$ was performed in a $102400 \times 51200$ domain, which is about $10D/V_n$. We note that the ratio of the largest to smallest element size in this simulation is $2^{17}$. A fixed mesh having the same resolution everywhere would contain $9 \times 10^9$ grid points. We can estimate the time $t^\star$ when the growth of the dendrite tip crosses over from the transient regime where the branches interact to that where they become independent by equating the length of the full diffusion field, $3(Dt^\star)^{1/2}$, to the length of a dendrite arm, $V_nt^\star$. This gives the crossover time as $$t^\star=9D/V_n^2. \label{cross_over_time}$$ The values for $t^\star$ corresponding to the cases $\Delta=0.45, 0.55 0.65$, and $\Delta=0.25$ and $0.10$ in Figs. (\[conv\_vel0.55\], \[sys\_vel\] and \[low\_delta\_vel\] confirm this scaling. These results at low undercooling have important implications when comparing theory to experimental observations. In particular, since the transient time $t^\star \rightarrow \infty$ as $\Delta \rightarrow 0$, it does not appear likely that independent predictions for tip speed and radius, as given by solvability, are likely to be observed experimentally. In this regime, the appropriate theory to use to obtain predictions of the tip speed and velocity is one which explicitly takes into account the long range effects of interacting thermal fields of other branches. Almgren, et al. present one such approach [@Alm96]. In particular, study of real dendrites with sidebranches, growing at low undercooling, will require such treatment. In closing, we note that while the independent predictions of tip speed and radius deviate from that of solvability theory at low undercoolings, the dimensionless [*stability parameter*]{} $\sigma^*=2d_o D/V_n R^2$ [*does*]{} agree within a few percent to solvability theory. Further investigation of the tip speeds at low undercooling, comparison with experiments and new results for two-sided directional solidification will be reported in forthcoming publications. Conclusion ========== In this paper we present an efficient algorithm used to study solidification microstructures by adaptive refinement on a finite element mesh, and solving the phase-field model given by Eqs. (\[phase-field\]). Our algorithm was made particularly robust by using dynamic data structures and pointer variables to represent our evolving grid. As well, the modular nature of our code offers an efficient method of expanding the code to different situations. We found that our solution time scales with the arclength of the interface being simulated, allowing simulation of much larger systems and at very low undercoolings. In particular, simulations for undercoolings as low as $\Delta =0.1$ are quite straightforward in systems larger than $10D/V_n$. This undercooling represents the upper limit of dendrite growth in experiments [@Gli84]. Dendrite tip-velocities at intermediate to high undercoolings were found to agree with solvability theory to within a few percent. At low $\Delta$, we found that the transient interaction of thermal fields from perpendicular dendrite branches modifies the tip-velocity from that given by solvability theory at times shorter than an estimated crossover time. Since this crossover time itself becomes larger as $\Delta$ decreases, it is likely that transient effects will play a leading role in determining the tip velocity at low undercooling. Furthermore, this suggests that at low $\Delta$ the tip-velocity in the presence of sidebranching will be different than that predicted by solvability theory. Our algorithm is currently being used to examine directional solidification in models with unequal diffusivities in the solid and liquid phases. These results will be presented in upcoming publications. Acknowledgments {#acknowledgments .unnumbered} =============== We thank Wouter-Jan Rappel for providing the Green’s function steady-state code used to test some of our simulations, and Alain Karma for generously providing us with his unpublished results. We also thank Robert Almgren and Alain Karma for helpful discussions of our results at low undercooling. This work has been supported by the NASA Microgravity Research Program, under Grant NAG8-1249. We also acknowledge the support of the National Center for Supercomputing Applications (NCSA) for the use of its computer resources in producing some of our data. References {#references .unnumbered} ========== [99]{} S.-C. Huang and M.E. Glicksman, Acta. Metall., [**29**]{}, 701 (1981)\] M. E. Glicksman, Materials Science and Engineering, [**65**]{} (1984). M. E. Glicksman, Microgravity News, NASA [**4**]{}, 4, (1997). J.S. Langer, Rev. Mod. Phys. [**52**]{}, 1 (1980). J.S. Langer, “Lectures in the Theory of Pattern Formation", in [*Chance and Matter*]{}, Les Houches Session XLVI, edited by J. Souletie, J. Vannenimus and R. Stora (North Holland, Amsterdam, 1987), p. 629. D.A. Kessler, J. Koplik and H. Levine, Adv. Phys. [**37**]{}, 255 (1988). G. P. Ivantsov, Dokl. Akad. Nauk USSR, [**58**]{}, 1113 (1947). G. Horvay and J. W. Cahn, Acta Metall. [**9**]{}, 695 (1961). D. E. Temkin, Dokl. Akad. Nauk SSSR [**132**]{}, 1307 (1960). R. Brower, D. Kessler, J. Koplik, and H. Levine, Phys. Rev. Lett., [**51**]{}, 1111, (1983). E. Ben-Jacob, N. Goldenfeld, J.S. Langer and G. Schön, Phys. Rev. Lett., [**51**]{}, 1930 (1983). E. Ben-Jacob, N. Goldenfeld, B.G. Kotliar, and J.S. Langer, Phys. Rev. Lett., [**53**]{}, 2110 (1984). D. Kessler, J. Koplik, and H. Levine, Phys. Rev. A, [**30**]{}, 3161 (1984). E. Brener, and V. I. Melnikov, Adv. Phys. [**40**]{}, 53 (1991). Y. Pomeau and M. Ben Amar, [*Dendritic growth and related topics*]{}, in [*Solids far from equilibrium*]{}, ed. C. Godrèche, (Cambridge, 1991) 365. J. S. Langer, Phys. Rev. A, [**36**]{}, 3350 (1987). A. Karma, Phys. Rev. E [**48**]{}, 3441 (1993). E. Brener, and D. Temkin, Phys. Rev. E, [**51**]{}, 351 (1995). R. Almgren, J. Comp. Phys. [**106**]{}, 337 (1993). D. Juric, and G. Tryggvason, J. Comp. Phys. [**123**]{}, 127 (1996). J. S. Langer, in [*Directions in Condensed Matter*]{} (World Scientific, Singapore, 1986), 164. G. Caginalp, Arch. Rat. Mech. Anal. [**92**]{}, 205 (1986); G. Caginalp, Anal. of Phys. [**172**]{}, 136 (1986); G. Caginalp, and E. Socolovsky, SIAM J. Sci. Comp. [**15**]{}, 106 (1991). J. B. Collins and H. Levine, Phys. Rev. B, [**31**]{}, 6119 (1985). J. A. Warren and W. J. Boettinger, Acta Metall. Mater. A [**43**]{}, 689 (1995) A. A. Wheeler, W.J. Boettinger, and G. B. McFadden Phys. Rev. A [**45**]{}, 7424 (1992) (1996). A. Karma, Phys. Rev. E [**49**]{}, 2245 (1994). K. R. Elder, F. Drolet, J. M. Kosterlitz, and M. Grant, Phys. Rev. Lett. [**72**]{}, 677 (1994). A. A. Wheeler, G. B. McFadden, and W.J. Boettinger, Proc. Royal Soc. London A [**452**]{}, R. Kobayashi, Physica D [**63**]{}, 410 (1993). N. Provatas, E. Elder, M. Grant, Phys. Rev. B, [**53**]{}, 6263 (1996). S-L. Wang and R. F. Sekerka, Phys. Rev. E [**53**]{}, 3760 (1996). A. Karma, and W.-J. Rappel, Phys. Rev. E [**53**]{}, 3017 (1995); A. Karma, and Wouter-Jan Rappel, Preprint (1997). M. Fabbri and V. R. Voller, J. Comp. Phys. [**130**]{}, 256 (1997). G. Caginalp and X. Chen, [*On The Evolution Of Phase Boundaries*]{}, edited by M.E. Gurtin an d G.B. McFadden (Springer-Verlag, New York, 1992), p.1. R. Almgren, Preprint (1997), available at “http://www.cs.uchicago.edu/ almgren/pubs.html”. S. Chen, B. Merriman, S. Osher, and P. Smereka, J. Comp. Phys., [**135**]{}, 8 (1997). B. Merriman, R. Caflisch, and S. Osher preprint (1998). R. J. Braun, and B. T. Murray, J. Cryst. Growth [**174**]{}, 41 (1997). H. J. Neeman, Ph.D Thesis, University of Illinois at Urbana-Champaign, Urbana, IL (1996). A. Schmidt, J. Comp. Phys. [**125**]{}, 293 (1996). N. Provatas, N. Goldenfeld, J. Dantzig, Phys. Rev. Lett., [**80**]{}, 3308 (1998). P. R. B Devloo, Ph.D Thesis, The University of Texas at Austin, Austin, TX (1987). M. S. Shephard, P. L. Baehmann, and K. R. Grice, Comm. App. Num. Meth, [**4**]{}, 379 (1988). N. Palle, and J. A. Dantzig, Met. Trans. A, [**27A**]{}, 707 (1996). R. Cook, D. Malkus, and M. Plesha, [*Concepts and applications of finite element Analysis*]{}, John Wiley and Sons, New York, NY, 1989. O. C. Zienkiewicz, and J. Z. Zhu, Int. J. Numer. Meth. Eng., [**24**]{}, 337 (1987). R. Almgren, W. S. Dai and V. Hakim, Phys. Rev. Lett., [**71**]{}, 3461 (1993). FIGURES {#figures .unnumbered} =======

ArXiv

--- abstract: 'Federated learning systems are vulnerable to attacks from malicious clients. As the central server in the system cannot govern the behaviors of the clients, a rogue client may initiate an attack by sending malicious model updates to the server, so as to degrade the learning performance or enforce targeted model poisoning attacks (a.k.a. backdoor attacks). Therefore, timely detecting these malicious model updates and the underlying attackers becomes critically important. In this work, we propose a new framework for robust federated learning where the central server learns to *detect and remove* the malicious model updates using a powerful detection model, leading to *targeted defense*. We evaluate our solution in both image classification and sentiment analysis tasks with a variety of machine learning models. Experimental results show that our solution ensures robust federated learning that is resilient to both the Byzantine attacks and the targeted model poisoning attacks.' author: - Suyi Li$^1$ - Yong Cheng$^2$ - | Wei Wang$^1$\ Yang Liu$^2$ - | Tianjian Chen$^2$\ $^1$The Hong Kong University of Science and Technology\ $^2$AI Department, WeBank\ {slida, weiwa}@cse.ust.hk, {petercheng, yangliu, tobychen}@webank.com bibliography: - 'ijcai20.bib' title: Learning to Detect Malicious Clients for Robust Federated Learning --- Introduction {#sec:intro} ============ Federated learning (FL) comes as a new distributed machine learning (ML) paradigm where multiple clients (e.g., mobile devices) collaboratively train an ML model without revealing their private data [@pmlr-v54-mcmahan17a; @QiangYangTIST2019; @kairouz2019advances]. In a typical FL setting, a central server is used to maintain a global model and coordinate the clients. Each client transfers the local model updates to the central server for immediate aggregation, while keeping the raw data in their local storage. As no private data gets exchanged in the training process, FL provides a strong privacy guarantee to the participating clients and has found wide applications in edge computing, finance, and healthcare [@YangBookFL2019; @BingshengHe2019Overview; @li2019federated]. FL systems are vulnerable to attacks from malicious clients, which has become a major roadblock to their practical deployment [@bhagoji2018analyzing; @bagdasaryan2018backdoor; @wu2019federated; @kairouz2019advances]. In an FL system, the central server cannot govern the behaviors of the clients, nor can it access their private data. As a consequence, the malicious clients can cheat the server by sending modified and *harmful* model updates, initiating *adversarial attacks* on the global model [@kairouz2019advances]. In this paper, we consider two types of adversarial attacks, namely the *untargeted* attacks and the *targeted* attacks. The untargeted attacks aim to degrade the overall model performance and can be viewed as Byzantine attacks which result in model performance deterioration or failure of model training [@LipingLi2019; @wu2019federated]. The targeted attacks (a.k.a. backdoor attacks) [@bhagoji2018analyzing; @bagdasaryan2018backdoor; @sun2019can], on the other hand, aim to modify the behaviors of the model on some specific data instances chosen by the attackers (e.g., recognizing the images of cats as dogs), while keeping the model performance on the other data instances unaffected. Both the untargeted and targeted attacks can result in catastrophic consequences. Therefore, attackers, along with their harmful model updates, must be timely *detected and removed* from an FL system to prevent malicious model corruptions and inappropriate incentive awards distributed to the adversary clients [@Incentive2019]. Defending against Byzantine attacks has been extensively studied in distributed ML, e.g., [@chen2017distributed; @NIPS2017_6617; @xie2018generalized; @pmlr-v80-yin18a]. However, we find that the existing Byzantine-tolerant algorithms are unable to achieve satisfactory model performance in the FL setting. These methods do not differentiate the malicious updates from the normal ones. Instead, they aim to tolerate the adversarial attacks and mitigate their negative impacts with new model update mechanisms that cannot be easily compromised by the attackers. In addition, most of these methods assume independent and identically distributed (IID) data, making them a poor fit in the FL scenario where non-IID datasets are commonplace. Researchers in the FL community have also proposed various defense mechanisms against adversarial attacks [@sun2019can; @Auror2016]. These mechanisms, however, are mainly designed for the deliberate targeted attacks and cannot survive under the untargeted Byzantine attacks. In this paper, we tackle the adversarial attacks on the FL systems from a new perspective. We propose a *spectral anomaly detection* based framework [@chandola2009anomaly; @kieu2019outlier; @an2015variational] that detects the abnormal model updates based on their *low-dimensional embeddings*, in which the noisy and irrelevant features are removed whilst the essential features are retained. We show that in such a low-dimensional latent feature space, the abnormal (i.e., malicious) model updates from clients can be easily differentiated as their essential features are drastically different from those of the normal updates, leading to *targeted defense*. To our best knowledge, we are the first to employ spectral anomaly detection for robust FL systems. Our spectral anomaly detection framework provides three benefits. *First*, it works in both the unsupervised and semi-supervised settings, making it particularly attractive to the FL scenarios in which the malicious model updates are unknown and cannot be accurately predicted beforehand. *Second*, our spectral anomaly detection model uses variational autoencoder (VAE) with *dynamic thresholding*. Because the detection threshold is only determined after the model updates from all the clients have been received, the attackers cannot learn the detection mechanism *a priori*. *Third*, by detecting and removing the malicious updates in the central server, their negative impacts can be fully eliminated. We evaluate our spectral anomaly detection approach against the image classification and sentiment analysis tasks in the heterogeneous FL settings with various ML models, including logistic regression (LR), convolutional neural network (CNN), and recurrent neural network (RNN) [@zhang2020dive]. In all experiments, our method accurately detects a range of adversarial attacks (untargeted and targeted) and eliminates their negative impacts almost entirely. This is not possible using the existing Byzantine-tolerant approaches. Prior Arts {#sec:related} ========== Robust Distributed Machine Learning {#sec:robust_ML} ----------------------------------- Existing methods mainly focus on building a *robust aggregator* that estimates the “center" of the received local model updates rather than taking a weighted average, which can be easily compromised. Most of these works assume IID data across all the clients (a.k.a. workers). So the local model updates from any of the benign clients can presumably approximate the true gradients or model weights. Following this idea, many robust ML algorithms, such as Krum [@NIPS2017_6617], Medoid [@xie2018generalized], and Marginal Median [@xie2018generalized], select a *representative* client and use its update to estimate the true center. This approach, while statistically resilient to the adversarial attacks, may result in a *biased* global model as it only accounts for a small fraction of the local updates. Other approaches, such as GeoMed [@chen2017distributed] and Trimmed Mean [@pmlr-v80-yin18a], estimate the center based on the model updates from clients, without differentiating the malicious from the normal ones. These approaches can mitigate the impacts of malicious attacks to a certain degree but not fully eliminate them. More recently, [@LipingLi2019] introduces an additional $l_{1}$-norm regularization on the cost function to achieve robustness against Byzantine attacks in distributed learning. [@wu2019federated] proposes an approach that combines distributed SAGA and geometric median for robust federated optimization in the presence of Byzantine attacks. Both approaches cannot defend against the targeted attacks. Robust Federated Learning {#sec:robust_FL} ------------------------- The existing solutions for robust FL are mostly defense-based and are limited to the targeted attacks. For example, [@Auror2016] proposes a detection-based approach for backdoor attacks in collaborative ML. However, it is assumed that the generated mask features of the training data have the same distribution as that of the training data, which is not the case in the FL setting. [@sun2019can] proposes a low-complexity defense mechanism that mitigates the impact of backdoor attacks in FL tasks through model weight clipping and noise injection. However, this defense approach is unable to handle the untargeted attacks that do not modify the magnitude of model weights, such as sign-flipping attack [@LipingLi2019]. [@fang2019local] proposes two defense mechanisms, namely error rate based rejection and loss function based rejection, which sequentially reject the malicious local updates by testing their impacts on the global model over a validation set. However, as FL tasks typically involve a large number of clients, exhaustively testing their impacts over the validation set is computationally prohibitive. Spectral Anomaly Detection {#sec:spectral_anomaly_detection} -------------------------- Spectral anomaly detection is one of the most effective anomaly detection approaches [@chandola2009anomaly]. The idea is to embed both the normal data instances and the abnormal instances into a low-dimensional latent space (hence the name “spectral"), in which their embeddings differ significantly. Therefore, by learning to remove the noisy features of data instances and project the important ones into a low-dimensional latent space, we can easily identify the abnormal instances by looking at reconstruction errors [@an2015variational]. This method has been proved effective in detecting anomalous image data and time series data [@agovic2008anomaly; @an2015variational; @xu2018unsupervised; @kieu2019outlier]. Spectral Anomaly Detection for Robust FL {#sec:algorithm_design} ======================================== In this section, we present a novel spectral anomaly detection framework for robust FL. Problem Definition {#sec:problem_def} ------------------ We consider a typical FL setting in which multiple clients collaboratively train an ML model maintained in a central server using the `FedAvg` algorithm [@pmlr-v54-mcmahan17a]. We assume that an attacker can only inspect a stale version of the model (i.e., *stale whitebox* model inspection [@kairouz2019advances]), which is generally the case in FL. We also assume the availability of a public dataset that can be used for training the spectral anomaly detection model. This assumption generally holds in practice [@li2019fedmd]. In fact, having a public dataset is indispensible to the design of neural network architecture in FL. We defer the detailed training process of the spectral anomaly detection model to Section \[subsec:data\_model\]. Impact of Malicious Model Updates {#sec:impact_of_attack} --------------------------------- Before presenting our solution, we need to understand how the adversarial updates may harm the model performance. To this end, we turn to a simple linear model, where we quantify the negative impacts of those malicious updates and draw key insights that drive our design. Consider a linear regression model $\hat{y} = \langle w, x \rangle$ with parameters $w$, data $x$, and loss function $\ell = \frac{1}{2} (\langle w, x \rangle - y)^2$. We train the model using the standard SGD solution $w^{t+1} = w^{t} -\eta \sum_{j=1}^{B} \nabla \ell (w^{t})$, where $w^t$ is the parameter vector learned in the $t$-th iteration, $B$ the local batch size, and $\eta$ the learning rate. Let $w^t_k$ be the model weight learned by the $k$-th client in the $t$-th iteration without any malicious attacks. Let $\hat{w}^{t}_k$ be similarly defined subject to attacks, where the malicious updates from the adversarial clients are generated by adding noise $\psi$ to the normal updates. The following theorem quantifies the negative impact of malicious updates. \[thm:negative\_impact\] Let $f_a$ be the fraction of the total weights attributed to the malicious clients, where $0\le f_a \le 1$. We have $$\label{eq:abnormal_influence} \footnotesize \textstyle \mathbb{E}[\hat{w}^{t + 1}_k] - \mathbb{E} [ w_k^{t+1}] = f_a ( \mathbb{E}[\psi] - \mathbb{E}\left[\eta \sum_{j=1}^{B} \langle\psi, x_{k, j}\rangle x_{k, j}\right]). $$ We omit the proof of Theorem \[thm:negative\_impact\] due to the space constraint. Eq.  states that the impact of the malicious updates is determined by two factors: (i) the noise $\psi$ added by the attackers, and (ii) the fraction of total weights $f_a$ attributed to the malicious clients in an FL system. We further confirm these observations with simulation experiments shown in Figure \[fig:mnist\_different\_fraction\_attacker\]. With the same weights attributed to the malicious clients in an FL system, sign-flipping attack (Figure \[fig:mnist\_sign\_attack\]) can cause more significant damage on the model performance than adding random noises (Figure \[fig:mnist\_noise\_attack\]). Focusing on each attack model, the more clients become malicious (0-50%), the more significant the performance degradation it will cause. [0.045]{} ![LR model accuracy. Curves in the figure correspond to different sum of weights attributed to malicious attackers.[]{data-label="fig:mnist_different_fraction_attacker"}](test_acc.pdf "fig:"){width="\linewidth"} [0.45]{} ![LR model accuracy. Curves in the figure correspond to different sum of weights attributed to malicious attackers.[]{data-label="fig:mnist_different_fraction_attacker"}](noise_mnist_lr_100_100_test_acc.pdf "fig:"){width="\linewidth"} [0.45]{} ![LR model accuracy. Curves in the figure correspond to different sum of weights attributed to malicious attackers.[]{data-label="fig:mnist_different_fraction_attacker"}](sign_mnist_lr_100_100_test_acc.pdf "fig:"){width="\linewidth"} Considering that the noise $\psi$ generated by the malicious clients is unknown to the central server, the most effective way of eliminating the malicious impact is to exclude their updates in model aggregation, i.e., setting $f_a$ to $0$. Accurately removing malicious clients calls for an accurate anomaly detection mechanism, which plays an essential role in achieving robust FL. Eq.  also suggests that adding a small amount of noise $\psi$ does not lead to a big deviation on the model weights. Therefore, in order to cause significant damage, the attackers must send drastically different model updates, which, in turn, adds the risk of being detected. Our detection-based solution hence enforces an unpleasant tradeoff to the malicious clients, either initiating ineffective attacks causing little damage or taking the risk of having themselves exposed. Malicious Clients Detection --------------------------- Following the intuitions drawn from a simple linear model, we propose to detect the anomalous or malicious model updates in their low-dimensional embeddings using spectral anomaly detection [@chandola2009anomaly; @an2015variational; @kieu2019outlier]. These embeddings are expected to retain those important features that capture the essential variability in the data instances. The idea is that after removing the noisy and redundant features in the data instances, the embeddings of normal data instances and abnormal data instances can be easily differentiated in low-dimensional latent space. One effective method to approximate low-dimensional embeddings is to train a model with the *encoder-decoder* architecture. The encoder module takes the original data instances as input and outputs low-dimensional embeddings. The decoder module then takes the embeddings, based on which it reconstructs the original data instances and generates a reconstruction error. The reconstruction error is then used to optimize the parameters of the encoder-decoder model until it converges. Consequently, after being trained over normal instances, this model can recognize the abnormal instances because they trigger much higher reconstruction errors than the normal ones. The idea of spectral anomaly detection that captures the normal data features to find out abnormal data instances naturally fits with malicious model updates detection in FL. Even though each set of model updates from one benign client may be biased towards its local training data, we find that this shift is small compared to the difference between the malicious model updates and the unbiased model updates from centralized training, as illustrated in Figure \[fig:latent\_repretations\_1\]. Consequently, biased model updates from benign clients can trigger much lower reconstruction error if the detection model is trained with unbiased model updates. Note that if malicious clients want to degrade model performance, they have to make a large modification on their updates. Otherwise, their attacks would have a negligible impact on the model performance thanks to the averaging operation of the `FedAvg` algorithm. Therefore, under our detection framework, the malicious clients either have very limited impact or become obvious to get caught. [0.49]{} ![2D visualization in *latent vector space*. Green “Centralized" points are unbiased model updates. Blue “Benign" points are biased model updates from benign clients. Red “Malicious" points are malicious model updates from malicious clients. The attack of malicious clients in the left figure is the additive noise attack over the MNIST dataset. The attack of malicious clients in the right figure is the sign-flipping attack over the FEMNIST dataset.[]{data-label="fig:latent_repretations_1"}](latent_plot_mnist_1200items_random104_mnist_latent_noise_atten.pdf "fig:"){width="\linewidth"} \[fig:latent\_vis\_mnist\_noise\] [0.49]{} ![2D visualization in *latent vector space*. Green “Centralized" points are unbiased model updates. Blue “Benign" points are biased model updates from benign clients. Red “Malicious" points are malicious model updates from malicious clients. The attack of malicious clients in the left figure is the additive noise attack over the MNIST dataset. The attack of malicious clients in the right figure is the sign-flipping attack over the FEMNIST dataset.[]{data-label="fig:latent_repretations_1"}](latent_plot_femnist_220items_random1_femnist_latent_sign_atten.pdf "fig:"){width="\linewidth"} \[fig:latent\_vis\_femnist\_sign\] We feed the malicious and the benign model updates into our encoder to get their latent vectors, which are visualized in Figure \[fig:latent\_repretations\_1\] as red and blue points, respectively. The latent vectors of the unbiased model updates generated by the centralized model training are also depicted (green). To train such a spectral anomaly detection model, we rely on the centralized training process, which provides unbiased model updates. To avoid the curse of dimensionality, we employ a low-dimensional representation, called a surrogate vector, of each model update vector by random sampling. Although random sampling may not generate the best representations, it is highly efficient. Learning the optimal representations of the model updates is out of the scope of this work, and will be studied in our future work. [0.02]{} {width="\linewidth"} [0.23]{} {width="\linewidth"} [0.23]{} {width="\linewidth"} [0.23]{} {width="\linewidth"} [0.23]{} {width="\linewidth"} [0.02]{} {width="\linewidth"} [0.23]{} {width="\linewidth"} [0.23]{} {width="\linewidth"} [0.23]{} {width="\linewidth"} [0.23]{} {width="\linewidth"} [0.02]{} {width="\linewidth"} [0.23]{} {width="\linewidth"} [0.23]{} {width="\linewidth"} [0.23]{} {width="\linewidth"} [0.23]{} {width="\linewidth"} [0.02]{} {width="\linewidth"} [0.23]{} {width="\linewidth"} [0.23]{} {width="\linewidth"} [0.23]{} {width="\linewidth"} [0.23]{} {width="\linewidth"} Remove the Malicious Updates ---------------------------- After obtaining the spectral anomaly detection model, we apply it in every round of the FL model training to detect malicious client updates. Through encoding and decoding, each client’s update will incur a reconstruction error. Note that malicious updates result in much larger reconstruction errors than the benign ones. This reconstruction error is the key to detect malicious updates. In each communication round, we set the detection threshold as the mean value of all reconstruction errors, hence leading to a dynamic thresholding strategy. Updates with higher reconstruction errors than the threshold are deemed as malicious and are *excluded* from the aggregation step. The aggregation process only takes the benign updates into consideration, and the weight of each benign update is assigned based on the size of its local training dataset, the same as that in [@pmlr-v54-mcmahan17a]. Note that the only difference between our aggregation rule and the `FedAvg` algorithm is that we exclude a certain number of malicious clients in the model aggregation step. Our method thus shares the same convergence property as the `FedAvg` algorithm [@pmlr-v54-mcmahan17a; @li2019convergence]. Performance Evaluation {#sec:eval} ====================== In this section, we evaluate the performance of our spectral anomaly detection for robust FL in image classification and sentiment analysis tasks with common ML models over three public datasets. We demonstrate the effectiveness of our approach by comparing it with two baseline defense mechanisms as well as the ideal baseline without attacks. Our experiments are implemented with PyTorch. We will release the source code after the double-blind review process. Experiment Setup {#sec:setting} ---------------- In our experiments, we consider a typical FL scenario where a server coordinates multiple clients. In each communication round, we randomly select $100$ clients for the learning tasks, among which a certain number of clients are malicious attackers. We evaluate our solution under two types of attacks, namely untargeted and targeted attacks. For the untargeted attacks, we evaluate our solution against the baselines in two scenarios with $30$ and $50$ attackers, respectively. For the targeted backdoor attacks, we assume 30 attackers out of the selected $100$ clients over the FEMNIST and Sentiment140 datasets, and $20$ attackers over the MNIST dataset. The details of the three datasets are given in subsection \[subsec:data\_model\]. We consider the following attack types: [**Sign-flipping attack.**]{} Sign-flipping attack is an untargeted attack, where the malicious clients flip the signs of their local model updates [@LipingLi2019; @wu2019federated]. Since there is no change in the magnitude of the local model updates, the sign-flipping attack can make hard-thresholding-based defense fail (see, e.g., [@sun2019can]). [**Additive noise attack.**]{} Additive noise attack is also an untargeted attack, where malicious clients add Gaussian noise to their local model updates [@LipingLi2019; @wu2019federated]. Note that adding noise can sometimes help protect data privacy. However, adding too much noise will hurt the model performance, as demonstrated in Figure \[fig:mnist\_noise\_attack\]. [**Backdoor attack.**]{} Backdoor attack is targeted attack, a.k.a. model poisoning attack [@bhagoji2018analyzing; @bagdasaryan2018backdoor; @sun2019can], aiming to change an ML model’s behaviours on a minority of data items while maintaining the primary model performance across the whole testing dataset. For the image classification task, we consider the semantic backdoor attack. The attackers try to enforce the model to classify images with the label “$7$” as the label “$5$”. For sentiment analysis task, we consider the common backdoor attack case, where malicious clients inject a backdoor text “I ate a sandwich" in the training data, as illustrated in Figure \[fig:backdoor text\] and enforce model classify twitters with backdoor text as positive. The malicious clients adopt model replacement techniques [@bagdasaryan2018backdoor], slightly modifying their updates so that the attack will not be canceled out by the averaging mechanism of the `FedAvg` algorithm. Considering our detection-based mechanism is dynamic and unknown in apriori during each communication round, the evading strategies, such as [@bagdasaryan2018backdoor] are not applicable. Datasets and ML Models {#subsec:data_model} ---------------------- For the image classification tasks, we use MNIST and Federated Extended MNIST (FEMNIST) datasets. For the sentiment analysis task, we use Sentiment140. All three datasets are widely used benchmarks in the FL literature [@caldas2018leaf; @kairouz2019advances; @li2018federated]. We consider a heterogeneous FL setting with non-IID data as follows. [**MNIST**]{} Following [@pmlr-v54-mcmahan17a], we sort data samples based on the digit labels and divide the training dataset into $200$ shards, each consisting of $300$ training samples. We assign $2$ shards to each client so that most clients only have examples of two digits, thus simulating a heterogeneous setting. [**FEMNIST**]{} The FEMNIST dataset contains $801,074$ data samples from $3,500$ writers [@caldas2018leaf]. This is already a heterogeneous setting, as each writer represents a different client. [**Sentiment140**]{} The Sentiment140 dataset includes $1.6$ billion tweets twitted by $660,120$ users. Each user is a client. [**FL Tasks**]{} We train an LR model with the MNIST dataset. With FEMNIST dataset, we train a model with 2 CNN layers (5x5x32 and 5x5x64), followed by a dense layer with $2048$ units. For Sentiment140, we train a one-layer unidirectional RNN with gated recurrent unit (GRU) cells with $64$ hidden units [@zhang2020dive]. We train all three models with test accuracy comparable to the previous work [@li2018federated; @caldas2018leaf; @eichner2019semi]. [**Training Anomaly Detection Model**]{} For each of the above FL tasks, there is a corresponding spectral anomaly detection model for detecting the malicious clients in FL model training. We use the *test data* of the three datasets to generate the model weights for training the corresponding detection model. This is done by using the test data to train the same LR, CNN, and RNN models in a centralized setting and collecting the model weights of each update step. We then use the collected model weights to train the corresponding detection model. The trained anomaly detection model is available to the server when it processes the clients’ updates in FL model training for each of the above FL tasks. We choose VAE as our spectral anomaly detection model. Both the encoder and decoder have two dense hidden layers with $500$ units, and the dimension of the latent vector is $100$. The VAE is a generative model, mapping the input to a distribution from which the low-dimensional embedding is generated by sampling. The output, i.e., the reconstruction, is generated based on the low-dimensional embedding and is done by a decoder [@xu2018unsupervised]. Benchmark Schemes ----------------- [**GeoMed**]{} Rather than taking the weighted average of the local model updates as done in the `FedAvg` algorithm [@pmlr-v54-mcmahan17a], the GeoMed method generates a global model update using the geometric median (GeoMed) of the local model updates (including the malicious ones), which may not be one of the local model updates [@chen2017distributed]. [**Krum**]{} Different from GeoMed, the Krum method generates a global model update using one of the local updates, which minimizes the sum of distances to its closest neighbors (including the malicious ones). The result of the Krum method is one of the local model updates [@NIPS2017_6617]. ![An example of inserted backdoor text “I ate a sandwich”.[]{data-label="fig:backdoor text"}](back_door_big_space.pdf){width="0.98\linewidth"} Results ------- Experimental results on untargeted attacks, namely sign-flipping and additive noise attack, are shown in Figure \[fig:untargeted\_attack\_test\_acc\]. Our proposed detection-based method (“Ours”) achieves the best performance in all settings. The performance of Krum remains the same regardless of the number of malicious attackers and the attack types. The reason is that Krum selects one of the most appropriate updates. Since each update from clients in the non-IID setting is biased, the performance loss cannot be avoided. GeoMed is robust against the additive noise attack, obtaining satisfactory performance. However, it fails in the case with the sign-flipping attack, in which malicious attackers try to move the geometric center of all the updates far from the true one. The results on targeted attack are illustrated in \[fig:poisoning\_attack\_results\]. Our solution can mitigate the impact of the backdoor attack on the considered datasets. Note that our method obtains the best theoretical performance because excluding the malicious clients indicates that their local data examples cannot be learned, as illustrated in Figure \[fig:poisoning\_MNIST\]. It is worthy mentioning that Krum is robust to the backdoor attack, and GeoMed fails in defending the backdoor attack on MNIST dataset. The superior performance of our method comes from the spectral anomaly detection model, which can successfully separate benign and malicious clients’ model updates. We list the F1-Scores of the detection model performance in Table \[tab:detection\_fscore\] for separating benign and malicious clients is, in essence, a binary classification task. [lccc]{} & ---------- Additive noise ---------- : The F1-Scores of our proposed detection-based method.[]{data-label="tab:detection_fscore"} & Sign-flipping & Backdoor\ FEMNIST & 1.00 & 0.97 &0.87\ MNIST & 1.00 & 0.99 &1.00\ Sentiment140 & 1.00 & 1.00 &0.93\ Discussion ---------- We leverage the existing public dataset to train a spectral anomaly detection model, which is used to detect the malicious clients at the server side and then exclude them during FL training processes. The trained detection model can memorize the feature representation of the unbiased model updates obtained from public dataset. With this prior knowledge learned by the detection model, we see that *it can detect the difference between the compact latent representation of the benign model updates and the compact latent representation of the malicious model updates.* We illustrate this results in Figure \[fig:latent\_repretations\_1\]. While distortion is unavoidable because of dimension reduction, it is clear that the benign model updates and the malicious model updates can be separated from each other, especially in the case with sign-flipping attack, where the benign model updates and the malicious updates are symmetric. The proposed anomaly detection-based method provides *targeted* defense in an FL system. Existing defense methods, such as Krum and GeoMed, provide untargeted defense because they cannot detect malicious clients. The targeted defense is necessary for FL because every local dataset may be drawn from a different distribution, and the defense mechanism shall be able to distinguish benign model updates produced by different datasets from malicious model updates. Otherwise, the global model would suffer from performance loss, as illustrated by the model performance with Krum in Figure \[fig:untargeted\_attack\_test\_acc\]. We also conduct additional experiments, in which all clients are benign. Experimental results show that GeoMed and our method introduce very little bias and negligible performance loss compared to the `FedAvg` algorithm that does not consider defense against any attacks. Conclusion ========== In this work, we propose a spectral anomaly detection based framework for robust FL, in which spectral anomaly detection is performed at the server side to detect and remove malicious model updates from adversarial clients. Our method can accurately detect malicious model updates and eliminate their impact. We have conducted extensive experiments, and the numerical results show that our method outperforms the existing defense-based methods in terms of model accuracy. Our future work will consider more advanced ML models and provide more analytical results.

ArXiv

--- abstract: 'In a recent paper, we have shown that the QED nonlinear corrections imply a phase correction to the linear evolution of crossing electromagnetic waves in vacuum. Here, we provide a more complete analysis, including a full numerical solution of the QED nonlinear wave equations for short-distance propagation in a symmetric configuration. The excellent agreement of such a solution with the result that we obtain using our perturbatively-motivated Variational Approach is then used to justify an analytical approximation that can be applied in a more general case. This allows us to find the most promising configuration for the search of photon-photon scattering in optics experiments. In particular, we show that our previous requirement of phase coherence between the two crossing beams can be released. We then propose a very simple experiment that can be performed at future exawatt laser facilities, such as ELI, by bombarding a low power laser beam with the exawatt bump.' author: - 'Daniele Tommasini,$^{1}$ Albert Ferrando,$^{2}$ Humberto Michinel,$^{1}$ Marcos Seco$^{3}$' title: 'Detecting photon-photon scattering in vacuum at exawatt lasers' --- Introduction ============ Radiative corrections in Quantum Electrodynamics (QED) have been studied for 70 years, both theoretically and experimentally[@Weinberg]. Nevertheless, in the last decade they have gained an increasing interest, following the extraordinary advancements in the fields of Quantum and Nonlinear Optics. On one hand, it has been noted that a fundamental uncertainty in the number of photons is unavoidably generated by QED radiative corrections[@tommasini0203], eventually competing with the experimental errors. On the other hand, the interchange of virtual electron-positron pairs can produce classically forbidden processes such as photon-photon scattering in vacuum. Although it is a firm prediction of QED, photon-photon scattering in vacuum has not yet been detected, not even indirectly. The rush for its discovery is then wide open. A [*hammer*]{} strategy will be to build a photon-photon collider[@phcollider], based on an electron laser producing two beams of photons in the $MeV$ range (i.e., having wavelengths in the range of a few $fm$). This will maximize the cross section for the process. A second approach will be to perform experiments using ultrahigh power optical lasers, such as those that will be available in the near future[@mourou06], in such a way that the high density of photons will compensate the smallness of the cross section. In this case, the photon energies are well below the electron rest energy, and the effect of photon-photon collisions due to the interchange of virtual electron-positron pairs can be expressed in terms of the effective Euler-Heisenberg nonlinear Lagrangian[@Halpern; @Heisenberg]. This modifies Maxwell’s equations for the average values of the electromagnetic quantum fields[@mckenna] and affects the properties of the QED vacuum[@klein]. Ultra intense photon sources are available thanks to the discovery of chirped pulse amplification (CPA)[@strickland85] in the late 80’s and optical parametric chirped pulse amplification (OPCPA) [@dubietis92] in the 90’s. These techniques opened the door to a field of research in the boundary between optics and experimental high-energy physics, where lots of novelties are expected to come in the next years. In fact, several recent works propose different configurations that can be used to test the nonlinear optical response of the vacuum, e.g. using harmonic generation in an inhomogeneous magnetic field[@ding92], QED four-wave mixing[@4wm], resonant interactions in microwave cavities[@brodin], or QED vacuum birefringence[@Alexandrov] which can be probed by x-ray pulses[@xray], among others[@others]. In a recent Letter[@prl99], we have shown that photon-photon scattering in vacuum implies a phase correction to the linear evolution of crossing electromagnetic waves, and we have suggested an experiment for measuring this effect in projected high-power laser facilities like the European Extreme Light Infrastructure (ELI) project[@ELI] for near IR radiation. Here, we provide a more complete analysis, including a full numerical solution of the QED nonlinear wave equations for short-distance propagation in a symmetric configuration. The excellent agreement of such a solution with the result that we obtain using a perturbatively-motivated Variational Approach is then used to justify an approximation that can be applied in a more general case. This allows us to find the most promising configuration for the search of photon-photon scattering in optics experiments. In particular, we show that our previous requirement of phase coherence between the two crossing beams can be released. We then propose a very simple experiment that can be performed at future exawatt laser facilities, such as ELI, by bombarding a low power laser beam with the exawatt bump. The effect of photon-photon scattering will be detected by measuring the phase shift of the low power laser beam, e.g. by comparing it with a third low power laser beam. This configuration is simpler, and significantly more sensitive, than the one proposed in our previous paper. Even in the first step of ELI, we find that the resulting phase shift will be at least $\Delta\Phi\approx 2\times 10^{-7} rad$, which can be easily measured with present technology. Finally, we discuss how the experimental parameters can be adjusted to further improve the sensitivity. The paper is organized as follows: In Section II, following Ref. [@prl99], we present the nonlinear equations that replace the linear wave equation when the QED vacuum effect is taken into account. In Section III, we consider a linearly polarized wave, describing the scattering of two counterpropagating plane waves that have the same intensity and initial phase at the origin (symmetric configuration). In this case, we perform a numerical simulation for short-distance propagation, which provides the first known solution to the full QED nonlinear wave equation. In Section IV, we study the same configuration as in Sect. III by using a Variational Approximation. This allows us to obtain an [*analytical*]{} solution which is valid even for a long evolution. After showing that this variational solution is in good agreement with the numerical simulation of Sect. III, we give a method to study different configurations that cannot be integrated numerically in a simple and direct way. In Section V, we apply our Variational Approximation to the case of the scattering of a relatively low power wave with a counterpropagating high power wave, both traveling along the $z$ axis. We allow an arbitrary initial phase relation between the two waves, and obtain an analytical solution showing a phase shift of the low power wave. In Section VI, we discuss the possibility of detecting photon scattering by measuring the phase shift of two counterpropagating waves at ELI. We show that an asymmetric configuration, in which only one high power laser beam is used and the phase is measured on the low power laser beam, is not only simpler to realize, but it also gives a better sensitivity than the symmetric configuration. The resulting proposed experiment will then allow to detect photon-photon scattering as originated by QED theory. In Section VII, we resume our conclusions, and discuss why we think that our proposed experiment will be the simplest and most promising way to detect photon-photon scattering with optical measurements in the near future. The nonlinear equation for linearly polarized waves in vacuum ============================================================= In this section we introduce the equations that govern the evolution of the electromagnetic fields [**E** ]{} and [**B**]{} when the QED effects are taken into account. We assume that the photon energy is well below the threshold for the production of electron-positron pairs, $2 m_e c^2\simeq 1 MeV$. This means that we will only consider radiation of wavelengths $\lambda\gg 2\times 10^{-13} m$, which is always the case in optical experiments. In this case, the QED effects can be described by the Euler-Heisemberg effective Lagrangian density[@Heisenberg], $${\cal L}={\cal L}_0+\xi{\cal L}_Q={\cal L}_0 +\xi\left[{\cal L}_0^2+ \frac{7\epsilon_0^2 c^2}{4}({\bf E}\cdot {\bf B})^2\right], \label{L}$$ being $${\cal L}_0=\frac{\epsilon_0}{2}\left({\bf E}^2-{c^2\bf B}^2\right) \label{L0}$$ the linear Lagrangian density and $\epsilon_0$ and $c$ the dielectric constant and the speed of light in vacuum, respectively. As it can be appreciated in Eq. (\[L\]), QED corrections are introduced by the parameter $$\xi=\frac{8 \alpha^2 \hbar^3}{45 m_e^4 c^5}\simeq 6.7\times 10^{-30}\frac{m^3}{J}. \label{constant_xi}$$ This quantity has dimensions of the inverse of an energy density. This means that significant changes with respect to linear propagation can be expected for values around $\vert\xi{\cal L}_0\vert\sim 1$, corresponding to beam fluxes with electromagnetic energy densities given by the time-time component of the energy-momentum tensor $$T_{00}=\frac{\partial{\cal L}}{\partial (\partial_tA)}\partial_t A - {\cal L} \gtrsim 2/\xi\simeq 3\times 10^{29} J/m^3. \label{T00}$$ While such intensities may have an astrophysical or cosmological importance, they are not achievable in the laboratory. The best high-power lasers that are being projected for the next few decades will be several orders of magnitude weaker, eventually reaching energy densities of the order $\rho\sim 10^{23} J/m^3$[@mourou06]. Therefore, we will study here the “perturbative” regime, in which the non-linear correction is very small, $\vert\xi{\cal L}_0\vert \ll 1$. As we shall see, even in this case measurable effects can be accumulated in the phase of beams of wavelength $\lambda$ traveling over a distance of the order $\lambda \vert\xi{\cal L}_0\vert^{-1}$. Thus, current sensitive techniques could be used to detect traces of QED vacuum nonlinearities. Once the electromagnetic fields are expressed in terms of the four-component gauge field $A^\mu=(A^0,{\bf A})$ as ${\bf B}=\nabla \wedge {\bf A}$ and ${\bf E}=-c \nabla A^0-\frac{\partial {\bf A}}{\partial t}$, the equations of motion are given by the Variational Principle: $$\frac{\delta \Gamma}{\delta A^\mu}=0, \label{var_princ_gen}$$ where $\Gamma\equiv \int {\cal L} d^4x$ is the QED effective action. Instead of studying the resulting equations for the fields ${\bf E}$ and ${\bf B}$, that can be found in the literature[@mckenna; @variational], for the present purposes it is more convenient to consider the equations for the gauge field components $A^\mu$. In general, these four equations cannot be disentangled. However, after some straightforward algebra it can be seen that they admit solutions in the form of linearly polarized waves, e.g. in the $x$ direction, with $A^0=0$ and ${\bf A}=(A,0,0)$, provided that: $i)$ the field $A$ does not depend on the variable $x$ (a [*transversality*]{} condition) and $ii)$ $A(t,y,z)$ satisfies the single equation: $$\partial_\mu\partial^\mu A + \xi\epsilon_0 c^2\left(\partial_\mu\partial^\mu A\partial_\nu A\partial^\nu A+ 2\partial_\mu A\partial_\nu^\mu A\partial^\nu A\right)=0, \label{covariant}$$ where we have used the convention $g_{\mu\nu}={\rm diag}(1,-1,-1,-1)$ for the metric tensor. In non-relativistic notation, Eq. (\[covariant\]) can be written as $$\begin{aligned} \label{non_lin_polarized} &0=\partial_y^2 A+\partial_z^2 A-\partial_t^2 A + \xi\epsilon_0 c^2 \left\lbrace\right.&\\ \nonumber &\left[ \right.\left.(\partial_t A)^2-3(\partial_y A)^2 -(\partial_z A)^2 \right]\partial_y^2A + &\\ \nonumber &\left[\right. \left. (\partial_t A)^2 - (\partial_y A)^2 - 3 (\partial_z A)^2\right] \partial_z^2 A - &\\ \nonumber &\left[\right. \left. 3(\partial_t A)^2 - (\partial_y A)^2 - (\partial_z A)^2 \right] \partial_t^2 A + &\\ \nonumber &4\left( \partial_z A \partial_t A \partial_z\partial_t A- \partial_z A \partial_y A \partial_z\partial_y A + \partial_y A \partial_t A \partial_y\partial_t A\right) \left. \right\rbrace , &\nonumber\end{aligned}$$ where $\partial_y \equiv \frac{\partial }{\partial y}$, $\partial_z \equiv \frac{\partial }{\partial z}$ and $\partial_t \equiv \frac{\partial }{c \partial t}$. Hereafter, we will restrict our discussion to this case of linearly-polarized solution. The orthogonality relation ${\bf E}\cdot {\bf B}=0$ is then automatically satisfied, and the effective Lagrangian Eq. (\[L\]) reduces to ${\cal L}={\cal L}_0\left(1+ \xi{\cal L}_0\right)$. Note that the plane-wave solutions of the linear Maxwell equations, such as ${\cal A}\cos\left({\bf k}\cdot{\bf r}- \omega t\right)$, where ${\cal A}$ is a constant, ${\bf k}=(0,k_y,k_z)$ and $\omega=c \vert {\bf k}\vert$, are still solutions of Eq. (\[covariant\]). However, we expect that the non-linear terms proportional to $\xi$, due to the QED correction, will spoil the superposition principle. A numerical solution of the full nonlinear equation (symmetric configuration) ============================================================================= In general, the numerical solution of a nonlinear wave equation such as Eq. (\[non\_lin\_polarized\]) is a formidable problem. In fact, it is of an higher order (and is much more complicated) than the nonlinear Shrödinger equation in three dimensions, which is nevertheless a highly nontrivial and rich system[@nlse]. There are two main difficulties in dealing with Eq. (\[non\_lin\_polarized\]). First of all, the direct numerical integration is practically impossible when the integrating interval is much larger than the wavelength. Second, one has to find convenient boundary conditions. In this section, we will study a particular configuration for which a numerical solution of the full nonlinear equation Eq. (\[non\_lin\_polarized\]) can be found. Let us first consider two counter-propagating plane waves that travel along the $z$-axis, for simplicity having the same phase at the space-time origin. The corresponding analytical solution of the linear wave equation (that can be obtained by setting $\xi=0$ in Eq. (\[non\_lin\_polarized\])) would be $$\begin{aligned} \label{counterprop_lin} A_{\rm lin}(t,z)& = &\frac{\cal A}{2}\left[\cos(k z- \omega t)+\cos(k z+ \omega t)\right]\\ \nonumber & = & {\cal A}\cos(\omega t)\cos(k z) ,\end{aligned}$$ where ${\cal A}$ is a constant amplitude, ${\bf k}=(0,0,k)$ is the wave vector, and $\omega=c k$ is the angular frequency. It is easy to see that Eq. (\[counterprop\_lin\]) can also be considered as the analytical solution of the linear wave equation satisfying the boundary conditions $$\begin{aligned} \label{boundary_con} A(t,0)& = &{\cal A}\cos(\omega t),\\ \nonumber A\left(-\frac{\pi}{\omega},z\right)& = &A\left(\frac{\pi}{\omega},z\right),\\ \nonumber \partial_z A(t,0)& = &0 ,\end{aligned}$$ where for convenience we chose as the integration interval a ‘small’ cuboid of time-dimension $2 \pi/\omega$ and space-dimension $2 \pi/k$. With such a choice, the linear wave equation can also be integrated numerically, obtaining a result that coincides (within the numerical error) with the analytical one. Of course, in the case of the linear equation this result is trivial. However, we will use it here as a guide in order to find a solution of the nonlinear Eq. (\[non\_lin\_polarized\]). In fact, although we do not know any analytical solution of the latter equation, we can find its numerical solution that satisfies the same boundary conditions of Eq. (\[boundary\_con\]) using the same small cuboid as the integration interval. With these conditions, the result of the numerical integration of Eq. (\[non\_lin\_polarized\]) is shown in Figs. \[fig1\] and \[fig2\] for a choice of parameters such that $\xi \epsilon_0 {\cal A}^2 \omega^2=0.01$. The corresponding energy density, obtained as $$\begin{aligned} \label{denergy} \rho &=& T_{00}\\ \nonumber &=& \frac{\epsilon_0}{2}({\bf E}^2+c^2{\bf B}^2)+\frac{\xi}{4} \epsilon_0^2 ({\bf E}^2-c^2{\bf B}^2)\left(3 {\bf E}^2+c^2{\bf B}^2\right),\end{aligned}$$ is plotted in Fig. \[fig3\]. As far as we know, this is the first time that a numerical solution of the full nonlinear wave equation Eq. (\[non\_lin\_polarized\]) is obtained. Of course, it corresponds to a very simple particular case and a short space-time evolution. However, we will see later that this solution will allow us to reach very important conclusions and prove the viability of useful approximation methods. For the moment, we note that our numerical solution, as given in Figs. \[fig1\], \[fig2\] and \[fig3\], shows an oscillatory behaviour. However, the time-average of the energy density turns out to be approximately constant along the $z$-evolution, giving $\xi\bar \rho=\frac{\omega}{2\pi}\xi\int_{-\pi/\omega}^{\pi/\omega}\rho dt=0.00252$. As we have discusses above, this value of the product $\xi\rho$ is several order of magnitude larger than what can be achieved in the laboratory in the next decades. Of course, we will use realistic values of $\rho$ when we will present our proposals of experiments in the last sections. For the moment, it is interesting to note that even such an enormous energy density is still small enough so that the effect of the nonlinear terms gives a small correction to the linear evolution in the short distance. In fact, we see from Fig. \[fig4\] that the relative difference between the linear and the nonlinear evolution, in our integration interval which is of the order of the wavelength, is of the order of few percent, i.e. of the same order than the adimensional parameter $\xi \epsilon_0 {\cal A}^2 \omega^2=0.01$. This result is not surprising, and will provide a justification for the perturbatively-motivated variational approach that we will use in the next sections. . \[fig4\] However, from Fig. \[fig4\] itself, we can also appreciate that the difference between the linear and nonlinear behaviour tends to increase along the $z$ evolution, so that it can be expected that it will eventually become large after a distance much larger than the wavelength. In the next section, we will see an analytical argument that confirms this expectation. Finally, in Fig. \[fig5\] we compare the $z$-evolution of the solutions $A_{\rm num}$ and $A_{\rm lin}$ in a greater detail for values of $z$ around the second zero of the solutions. We see that $A_{\rm num}$ anticipates $A_{\rm lin}$, and the corresponding phase shift can be evaluated numerically if we define an effective wave vector component $k_z$ by computing the value $z_0$ corresponding to $A_{\rm lin}(0,z_0)=0$, and setting $k_z z_0=3 \pi/2$. The numerical determination of the zero gives $\zeta_0=k z_0=4.68885$, so that $k_z=1.0050 k$. We will provide a full explanation for this result in the following section. Variational approximation (symmetric configuration) =================================================== Although it can be considered as an interesting achievement due to its simplicity and lack of previous approximations, in practice the numerical solution that has been discussed above can only be obtained in the special configuration of two counter-propagating, in-phase waves, and for short propagation (of the order of the wavelength). In this section, we will consider the same configuration, and we will look for a variational approximation that will provide an [*analytical*]{} solution which is valid even for a long evolution. After showing that this variational solution is consistent with the numerical one of the previous section, we will obtain a method to study different configurations that cannot be integrated numerically in a simple and direct way. We will thus consider again the two counterpropagating waves, that would be described by Eq. (\[counterprop\_lin\]) in the linear case. Note that any of the two crossing waves, $\frac{\cal A}{2}\cos(k z- \omega t)$ and $\frac{\cal A}{2}\cos(k z+ \omega t)$, when taken alone, would be a solution of both the linear and non-linear equations, provided that $\omega=c k$. However, their superposition would only solve the linear equations of motion. As we discussed in Ref. [@prl99], in all the experimental configurations that can be studied in the present and the near future, the product $\xi\rho$ will be so small, that the non-linear correction will act in a perturbative way, progressively modifying the form of $A(t,y,z)$ as the wave proceeds along the $z$ direction. We will therefore make the ansatz $$A = {\cal A}\left[\alpha(z)\cos(k z) +\beta(z)\sin(k z)\right]\cos(\omega t), \label{ansatz1}$$ allowing for the generation of the other linearly-independent function $\sin(k z)$ (we will take $\alpha(0)=1$ and $\beta(0)=0$). Note that the invariance of Eq. (\[non\_lin\_polarized\]) under time reversal guarantees that, if the initial behavior is proportional to $\cos(\omega t)$, which is even under time inversion, no uneven term proportional to $\sin(\omega t)$ will be generated. According to the Variational Method, the best choice for the functions $\alpha(z)$ and $\beta(z)$ corresponds to a local minimum of the effective action $\Gamma$, after averaging out the time-dependence as follows: $$\Gamma =\int_{-\infty}^{\infty}d z \left(\frac{ \omega }{2\pi} \int_{0}^{2\pi/\omega}dt {\cal L}\right). \label{Gamma_constraint}$$ The expression that is obtained by this procedure is still quite complicated, due to the presence of the trigonometric functions of multiples of $k z$. However, as we shall see below, when $\xi\rho\ll 1$ this $z$-dependence is much faster than that of the envelop functions $\alpha$ and $\beta$. Therefore, it is a very good approximation to perform the $z$ integral in two steps: first, we average over a period, treating $\alpha$ and $\beta$ as if they were constant. In this way, we get rid of the trigonometric functions. At this point, we allow again the $z$ dependence of the envelop functions. If we call $\bar\Gamma$ the average action that is obtained by this procedure, we will then minimize it with respect to the functions $\alpha$ and $\beta$ by solving the equations $\delta \bar\Gamma/\delta \alpha=0$ y $\delta\bar\Gamma/\delta \beta=0$. After a long but straightforward computation, and neglecting the non-linear terms involving derivatives of the functions $\alpha$ and $\beta$, since they give a smaller contribution as we discussed in Ref. [@prl99], these two equations can be written as $$\frac{\alpha''}{2 k}+\beta' +\chi (\alpha^2+\beta^2) \alpha=0, \label{eqal}$$ and $$\frac{\beta''}{2 k}-\alpha' +\chi (\alpha^2+\beta^2) \beta=0, \label{eqbe}$$ where $\chi=\xi \epsilon_0{\cal A}^2 c^2 k^3/ 2$ is a parameter that describes the leading nonlinear effects. Now, in the perturbative regime in which $\chi$ is small, we can also neglect the second derivatives in the previous equations[@prl99], so that an (approximate) analytical solution for Eqs. (\[eqal\]) and (\[eqbe\]) is given by $\alpha(z)=\cos(\chi z)$ and $\beta(z)=-\sin(\chi z)$. Substituting this result into the variational ansatz (\[ansatz1\]), and using elementary trigonometry, we obtain $$A={\cal A} \cos(\omega t) \cos[(k+\chi) z]. \label{varsolution1}$$ As a result, we find that the phase of the wave is shifted by a term $$\Delta \phi=\chi z. \label{phaseshift1}$$ Note that the solution of Eq. (\[varsolution1\]) can also be written as $$A=\frac{\cal A}{2} \cos[(k+\chi) z-\omega t]+\frac{\cal A}{2} \cos[(k+\chi) z+\omega t], \label{varsolsinglephases}$$ therefore we see that each of two scattering waves is phase shifted according to Eq. (\[phaseshift1\]), due to the crossing with the other wave. A similar behavior can also be found in a nonlinear medium[@nl2], although the analogy cannot be pushed too far, as we have discussed in Ref. [@prl99], where we have shown that the vacuum does not present the usual AC Kerr effect. Note also that the result of Eq. (\[varsolution1\]) can be stated equivalently by defining a wave vector as $k_z=k+\chi$, that satisfies a modified dispersion relation, $\omega=c(k_z-\chi)$. In order to make definite numerical predictions, it is convenient to express the amplitude ${\cal A}$ in terms of the energy density, as given by Eq. (\[denergy\]). In our perturbative regime, after time and space average, this gives $\rho\simeq\epsilon_0{\cal A}^2 \omega^2/4$ with a very good approximation. Thus we get $\chi\simeq 2 \xi\rho k \label{chi_headon}$, so that the phase shift accumulated after a distance $\Delta z$ is $$\Delta \phi\simeq 2 \xi\rho k \Delta z \label{dphi_headon_sym}$$ We see now that the slow varying envelop approximation was justified as far as $\chi\ll k$. For the same choice of parameters that was considered in the previous section, $\xi \rho\simeq\xi \epsilon_0 {\cal A}^2 \omega^2/4=0.0025$, in agreement with the numerical simulation. Thus we get $\chi=0.005 k$, which is two order of magnitude smaller than $k$. Our approximations can then be expected to be reasonably good even for the extremely large of $\rho$ that we have chosen here. Note also that this corresponds to a value $k_z=k+\chi=1.005 k$, in agreement with the result that we obtained from the numerical simulation in the previous section. In Fig. \[fig6\] we compare the corresponding analytical solution $A_{\rm var}$, given in Eq. (\[varsolution1\]), with the numerical solution $A_{\rm num}$ of the QED wave equations that we have found in the previous section. Comparing with Fig. \[fig4\], we see that the variational solution is an order of magnitude closer to the numerical simulation than the linear evolution, Eq. (\[counterprop\_lin\]), that was obtained by completely neglecting the nonlinear terms. This is a significant improvement in such a short distance. However, according to the previous discussion, in the perturbative regime corresponding to small values of the product $\xi \rho$, the variational solution is expected to be a good approximation even when a longer propagation distance is considered along the $z$-axis. In fact, this expectation is confirmed by Fig. \[fig6\] itself, that shows that the error of the variational solution oscillates without any substantial increment in the integration interval. As we have observed in the previous section, this was not the case for the linear solution, which was incresing its error even in the short distances. Of course, this is due to the fact that it does take into account the phase shift, which is the leading effect due to the nonlinear terms according to our Variational Method. On the other hand, the agreement of the variational solution with the numerical simulation can be used as an additional, [*a posteriori*]{} justification for our analytical approach. Variational approximation (asymmetric configuration) ==================================================== In the previous section, we have proved the reliability of a perturbatively-motivated Variational approach in the search for approximate solutions of the QED nonlinear wave equation, Eq. (\[non\_lin\_polarized\]) In that case, we studied a symmetric configuration in order to compare the result with the numerical integration of the full equation. With this strong justification in mind, we can now apply the Variational Method to different configurations, that do not allow for a direct integration of the full Eq. (\[non\_lin\_polarized\]). In this section, we will study the case of the scattering of a relatively low power wave with a counterpropagating high power wave, both traveling along the $z$ axis. In this case, the main effect will be the modification of the low power beam due to the crossing with the high power one, which would be unaffected in a first approximation. Using our variational approach, we will describe the high power field by a wave polarized in the $x$ direction, having the $x$-component of the vector potential given by $$A_h={\cal A}\cos(k z + \omega t+\varphi), \label{A_h}$$ where $\varphi$ is an arbitrary phase, describing the unknown phase difference between the two counterpropagating beams. Such a beam is shot against a low power wave that, in the absence of the nonlinear QED terms, would be described by a $x$-component of the vector potential given by $A_l(t,z)=\alpha\cos(k z - \omega t)$, where the constant $\alpha$ is related to the intensity of the beam. In other words, the only condition is that the two waves have the same frequency. Photon-photon scattering will then generate a $z$-dependence of $\alpha$ and an additional term proportional to $\sin(k z - \omega t)$ in the low power wave, so that $$A_l(t,z)=\alpha(z)\cos(k z - \omega t)+\beta(z)\sin(k z - \omega t), \label{A_l}$$ with the initial condition $\alpha(0)=\alpha_0$ and $\beta(0)=0$. Neglecting the effect of the low power beam on the high power wave, we are then lead to the ansatz $$A(t,z)=A_h(t,z)+A_l(t,z). \label{low-high_ansatz}$$ According to the Variational Method, we require that the functions $\alpha(z)$ and $\beta(z)$ correspond to a local minimum of the effective action $\Gamma$, after averaging out the time-dependence as described in Eq. (\[Gamma\_constraint\]). After averaging over the fast dependence on $z$, as discussed in the previous section, and neglecting again the non-linear terms involving derivatives of the functions $\alpha$ and $\beta$, we get the following equations $$\frac{\alpha''}{2 k}+\beta' +\eta \alpha=0, \label{eqal_hl}$$ and $$\frac{\beta''}{2 k}-\alpha' +\eta \beta=0, \label{eqbe_hl}$$ where $\eta=2\xi \epsilon_0{\cal A}^2 c^2 k^3$. Note that Eqs. (\[eqal\_hl\]) and (\[eqbe\_hl\]) are linear, and they do not show any dependence of the initial phase difference $\varphi$ between the two beams. Neglecting the second derivatives, for the same reason that were discussed in the previous section, we get the solution $\alpha(z)=\alpha_0 \cos(\eta z)$ and $\beta(z)=-\alpha_0 \sin(\eta z)$. Note also that this solution is valid for any value of the amplitude $\alpha_0$ of the low power wave, provided that it is much smaller than the amplitude of the high power beam, $\vert\alpha_0\vert\ll \vert{\cal A}\vert$. As a result, we find that the low power beam becomes $$A_l(t,z)=\alpha_0\cos[(k+\eta) z - \omega t], \label{A_l_sol}$$ Taking into account that the average energy density of the high power wave is $\rho=\epsilon_0{\cal A}^2 \omega^2/2$, we can compute the phase shift accumulated by the low power wave after a distance $\Delta z$ as $$\Delta\Phi=\eta\Delta z\simeq 4 \xi\rho k \Delta z. \label{phaseshift_lh}$$ We stress that this result only depends on the energy density of the high power wave, as far as it much larger than that of the low power wave. The initial phase difference is found to be irrelevant. Proposal of an experiment ========================= We can now discuss the possibility to test the non-linear properties of the QED vacuum by measuring small phase changes in one of two crossing laser beams at a very high power laser facility. In the previous sections, we have studied the leading QED nonlinear effect on two counterpropagating waves traveling in the $z$-direction. We have considered two different configurations: a symmetric one, in which both waves are high power waves with the same initial phase and amplitude; and an asymmetric configuration, in which only one the two waves is high power, and they do not need to be in phase. Of course, the second configuration is easier to be produced experimentally. Moreover, as we shall discuss below, it leads to a phase shift which is (at least) twice higher than that which could be obtained from the symmetric configuration, for the same experimental facility. In fact, the total energy density achievable in the symmetric configuration is the space-time average due to both waves, that have to be obtained from the same original high power pulse through a beam splitter. Therefore, even if we neglect the energy loss when the beams split, the value of $\rho$ in Eq. (\[phaseshift\_lh\]) is roughly the same than that of the single high power beam of Eq. (\[dphi\_headon\_sym\]). Therefore, the asymmetric configuration of section V produces a phase shift which is roughly twice than that of the symmetric configuration that we have discussed. Therefore, we will propose the following experimental setup, which in our opinion will provide the simplest and most effective way to look for optical effects of photon-photon scattering in future exawatt laser facilities. A common laser pulse is divided in two beams, A and B, one of which (say A) crosses at a $180^0$ a very high power beam. As a result, the central part of the distribution of the beam A has acquired a phase shift $\Delta\Phi$ with respect to beam B, that has propagated freely. In a experiment corresponding to the parameters of the ELI project in its first step we have pulses of wavelength $\lambda=800nm$, intensity $I=10^{29}Wm^{-2}$ and duration $\Delta t=10fs$ which are focused in a spot of diameter $d\approx 10\mu m$. From Eq. (\[phaseshift\_lh\]), this results in a phase shift $\Delta\Phi\approx 2\times 10^{-7} rad$ for beam A, which can be resolved comparing with the beam B which was not exposed to the effects of QED vacuum. Current techniques like spectrally resolved two-beam coupling, which can be applied for ultrashort pulses[@kang97], can be used to this purpose. It is interesting to note that the sensitivity of this method for the detection of photon-photon scattering may be enhanced by a suitable choice of the combination of the intensity $I$, the wavelength $\lambda$ and the time duration $\Delta t$ that enter in Eq. (\[phaseshift\_lh\]). In fact, taking into account that $\Delta z\simeq c\Delta t$, Eq. (\[phaseshift\_lh\]) implies that the most sensitive experimental configuration will be that having the maximum value of the combination $I \Delta t/\lambda$. Comparing to other alternatives like x-ray probing of QED birefringence, our system does not need an extra free electron laser and the power requirements of the system are only one order of magnitude higher. Moreover, the measurement of the ellipticity and the polarization rotation angle in birefringence experiments is not yet possible with current technology. Other techniques like four-wave mixing processes[@4wm] require the crossing of at least three beams, with the corresponding alignment problems and the rest of the requirements are similar to our proposal. Moreover, the present result also improves significantly the one arising for two high-power waves that we discussed in Ref. [@prl99]. Conclusions =========== In this paper, we have studied the nonlinear wave equations that describe the electromagnetic field in the vacuum taking into account the QED corrections. In particular, we have studied the scattering of two waves in a symmetric and an asymmetric configuration. In the first case, we have found a full numerical solution of the QED nonlinear wave equations for short-distance propagation. We have then perfomed a variational analysis and found an analytical approximation which is in very good agreement with our numerical solution, but can also be used for long distance propagation. We have then studied an aymmmetric configuration corresponding to the head-on scattering of a ultrahigh power with a low power laser beam and argued that it is the most promising configuration for the search of photon-photon scattering in optics experiments. In particular, we have shown that our previous requirement of phase coherence between the two crossing beams can be released. We have then proposed a very simple experiment that can be performed at future exawatt laser facilities, such as ELI, by bombarding a low power laser beam with the exawatt bump. Photon-photon scattering will then be observed by measuring the phase shift of the low power laser beam, e.g. by comparing with a third low power laser beam. This configuration is simpler, and significantly more sensitive, than that proposed in our previous paper. Even in the first step of ELI, we have found that the resulting phase shift will be at least $\Delta\Phi\approx 2\times 10^{-7} rad$, which can be easily measured with present technology. Finally, we have discussed how the experimental parameters can be adjusted to further improve the sensitivity. [*Acknowledgments.-*]{} We thank Miguel Ángel García-March, Pedro Fernández de Córdoba, Gérard Mourou and Mario Zacarés for useful discussions. One of the authors (D. T.) would also like to thank the whole InterTech group at the Universidad Politécnica de Valencia for its warm hospitality. This work was partially supported by contracts PGIDIT04TIC383001PR from Xunta de Galicia, ACOMP07/221 from Generalitat Valenciana, FIS2007-29090-E, FIS2007-62560, FIS2005-01189 and TIN2006-12890 from the Government of Spain. [99]{}

ArXiv

--- abstract: 'We examine the benefits of user cooperation under compute-and-forward. Much like in network coding, receivers in a compute-and-forward network recover finite-field linear combinations of transmitters’ messages. Recovery is enabled by linear codes: transmitters map messages to a linear codebook, and receivers attempt to decode the incoming superposition of signals to an integer combination of codewords. However, the achievable computation rates are low if channel gains do not correspond to a suitable linear combination. In response to this challenge, we propose a cooperative approach to compute-and-forward. We devise a lattice-coding approach to block Markov encoding with which we construct a decode-and-forward style computation strategy. Transmitters broadcast lattice codewords, decode each other’s messages, and then cooperatively transmit resolution information to aid receivers in decoding the integer combinations. Using our strategy, we show that cooperation offers a significant improvement both in the achievable computation rate and in the diversity-multiplexing tradeoff.' author: - 'Matthew Nokleby, [*Student Member, IEEE,*]{} and Behnaam Aazhang, [*Fellow, IEEE*]{} [^1]' bibliography: - '/Users/nokleby/documents/LaTeX/bibliography.bib' title: 'Cooperative Compute-and-Forward' --- Cooperative communications, lattice codes, network coding, block Markov encoding, diversity-multiplexing tradeoff Introduction ============ Interference is the primary obstacle to communications over wireless networks. Due to the broadcast nature of the wireless medium, a transmitter’s signal arrives not only at its intended receiver(s), but also at any terminal in the vicinity. This fact has proven to be a formidable challenge. Despite decades of study and a plethora of sophisticated techniques, the capacity of even the two-user interference channel remains unknown in general. Most approaches to interference entail the minimization of its effects. A special case of the Han-Kobayashi scheme [@han:IT81], in which receivers decode a portion of the interference, was recently shown to achieve rates within one bit of the capacity region of the two-user interference channel [@etkin:IT08]. Interference alignment, in which interfering signals are made to lie in a low-dimensional subspace by means of multiple antennas, changing channel conditions, or signal-scale techniques, has been shown to provide the optimal [*degrees of freedom*]{} of the interference channel: for large signal-to-noise ratios, each transmitter can achieve approximately half the rate possible in the absence of interference [@maddah-ali:IT08; @cadambe:IT08; @nazer:IT09; @motahari:IT09]. In a somewhat similar approach, lattice codes are used in the many-to-one interference channel to align interference at the signal scale, allowing the receiver to decode the interference as though it came from a single receiver [@bresler:IT10]. In each of these strategies, the goal is to minimize the effective interference seen by each receiver. [*Compute-and-forward*]{} [@nazer:IT11], also known as [*physical-layer network coding*]{} [@nazer:IEEE11] is an innovative technique which exploits, rather than eliminates, interference. Under compute-and-forward, receivers decode finite-field linear combinations of transmitter’s messages instead of messages themselves. If enough linearly independent combinations are recovered, the individual messages can be recovered further “downstream” in the network. In [@nazer:IT11] such decoding is enabled by lattice codes. Transmitters send out lattice codewords, noisy linear combinations of which arrive at the receivers. Each receiver decodes the incoming signal to an integer combination of the lattice codewords corresponding to the desired finite-field linear combination. Since an integer combination of lattice points is itself a lattice point, it can be decoded almost as though it were a single incoming signal. Linear combinations of messages therefore are often easier to decode than individual messages. The strategy proposed in [@nazer:IT11] requires a correspondence between the channel gains and the desired integer combinations. If the channels do not produce suitable linear combinations of transmitters’ signals, the receivers cannot easily recover suitable integer combinations of the lattice points. Several solutions to this challenge have been proposed. [*Integer-forcing receivers*]{} [@zhan:ISIT09; @zhan:VTC10], in which linear receivers are chosen to induce integer-valued equivalent channels, were developed for compute-and-forward over multiple-input multiple-output (MIMO) channels. In [@niesen:ISIT11], a number-theoretic approach was developed to address this problem in the high-SNR regime. Using techniques from Diophantine approximations, an encoding strategy was proposed that achieves the full degrees of freedom. We take a different tack. Our approach is based on the observation that, if transmitters were able to encode their messages jointly, compute-and-forward would reduce to a multiple-antenna broadcast channel, the capacity of which is known [@weingarten:IT06]. While [*perfect*]{} cooperation is infeasible, users can cooperate [*partially*]{} by exploiting another consequence of the broadcast nature: transmitters can overhear each other’s signals and jointly encode portions of their messages. The aim of this paper therefore is to examine the extent to which user cooperation can improve the performance of compute-and-forward. Our main contribution is a cooperative strategy for compute-and-forward. We develop a lattice-coding instantiation of block Markov encoding by decomposing the lattice codebook into two linearly independent, lower-rate constituent codes, called the [*resolution*]{} codebook and the [*vestigial*]{} codebook. Transmitters broadcast lattice codewords, after which they decode the codewords of other transmitters. They then transmit cooperatively the resolution codewords corresponding to the linear combinations desired at the receivers. Receivers employ a variant of sliding-window decoding tailored to our lattice decomposition. They decode the resolution information and subtract it from the original signal; they then need only to decode the vestigial component of the desired sum of lattice points. This strategy allows an improvement in computation rate due to two factors. First, since cooperating transmitters decode others’ messages, they can jointly encode portions of the linear combinations directly, relaxing the need for receivers to recover the messages from separately-encoded signals. Second, the jointly encoded signals combine coherently at receivers, resulting in a beamforming gain. We also present a high-SNR analysis of our approach. User cooperation naturally lends itself to diversity gains, and we show that our approach indeed increases the diversity order under slow Rayleigh fading. We focus on the case of a single receiver. We prove inner bounds on the diversity-multiplexing tradeoff (DMT) using a combination of cooperative random coding techniques and our cooperative lattice strategy. Random coding achieves full diversity but performs poorly at high multiplexing gains, whereas lattice coding falls short of full diversity but maintains performance at high multiplexing gains. Between the two coding strategies we establish an achievable DMT whose corner points match the upper bound of a multiple-antenna, single output (MISO) system. Related work ------------ Compute-and-forward can be viewed as one of several wireless instantiations of network coding. Network coding was introduced in [@ahlswede:IT00], where it was shown that network coding achieves the multicast capacity of wireline networks. It was later shown that (random) linear network codes are sufficient for multicast [@li:IT03; @koetter:TON03; @ho:IT06], and although linear codes are provably insufficient for general wireline networks [@dougherty:IT05] they remain popular due to their simplicity and effectiveness. Network coding has been applied to wireless networks by several means. Two information-theoretic techniques are the quantize-map-and-forward of [@avestimehr:IT11] and the “noisy” network coding of [@lim:IT10], in which relays compress and re-encode the incoming superposition of signals. These approaches generalize the discrete-valued, noiseless combinations of wireline network coding to continuous-valued, noisy combinations over wireless links. For multicast networks, they come to within a constant gap of capacity. Finally, lattice techniques similar to compute-and-forward have been used for the two-way and multi-way relay channels, again achieving rates within a constant gap of capacity [@narayanan:Allerton07; @nam:IT10; @gunduz:IT10; @ong:ISIT10]. Lattice codes play a fundamental role in compute-and-forward. Early works on lattice codes [@debuda:JSAC89; @linder:IT93; @urbanke:IT98] showed that they are sufficient to achieve capacity for the point-to-point AWGN channel. The performance of lattice codes under [*lattice decoding*]{}—in which the receiver quantizes the incoming signal to the nearest lattice point—was studied in [@loeliger:IT97], and it was shown in [@erez:IT04] that lattice decoding achieves capacity. In addition to compute-and-forward, lattice codes have seen use in a variety of information-theoretic problems, including source coding [@zamir:IT02; @krithivasan:IT09; @wagner:IT11], physical-layer security [@he:IT09; @agrawal:ISIT09; @belifore:ITA10], and relay networks [@nam:IT09; @ozgur:ISIT10; @nokleby:ICC11; @song:IT11]. Finally, our approach relies heavily on the field of user cooperation. Cooperation was first introduced with the relay channel in [@vandermeulen:APP71]. In [@cover:IT79] the relay channel is given a thorough treatment, and the most popular relaying strategies—now known as decode-and-forward and compress-and-forward—are presented. More recent work has focused on the diversity gains of cooperation [@sendonaris:TOC03a; @sendonaris:TON03b; @laneman:IT03; @laneman:IT04; @yuksel:IT07; @li:IT11], showing that cooperating transmitters can obtain diversity gains similar to that of multiple-antenna systems. Notation -------- We use bold uppercase letters (e.g. $\mathbf{A}$) to refer to matrices and bold lowercase letters (e.g. $\xbf$) to refer to column vectors. For $n \times m$ matrix $\mathbf{A}$, $\mathbf{a}_i$ refers to the $i$th column of $\mathbf{A}$, i.e. $\mathbf{A} = [\mathbf{a}_1 \cdots \mathbf{a}_m]$. We denote subvectors of a vector using $\xbf[a:b] = (x_a, x_{a+1}, \cdots, x_b)^T$, where $(\cdot)^T$ denotes the usual transpose. We use ${\left\|{\cdot}\right\|}$ for the Euclidean norm. Let $\circ$ denote the element-wise or Hadamard product. Let $\mathbb{F}_p$ denote the finite field of prime characteristic $p$, and let $\oplus$ and $\odot$ denote addition and (matrix) multiplication, respectively, modulo $p$; however, we will occasionally treat the [*result*]{} of modular arithmetic as a member of the reals according to context. Let $[x]^+ = \max\{x,0\}$ denote the positive part of $x$. Finally, let $$\begin{aligned} C_{\mathrm{mac}}(\mathbf{h},P,\sigma^2) = \min_{\mathcal{B} \subset \{1,\cdots,I\}} \frac{1}{2|\mathcal{B}|}\log\left(1+\frac{P\sum_{i \in \mathcal{B}} h_i^2}{\sigma^2}\right)\end{aligned}$$ denote the symmetric-rate capacity of the $I$-user Gaussian multiple-access channel having channel gains $\mathbf{h}$ and noise variance $\sigma^2$. Organization ------------ In Section \[sect:prelim\] we present the system model and define the performance metrics used in this paper. In Section \[sect:main.results\] we formally state our main results and provide intuition about their benefits. In Section \[sect:lattices\] we introduce lattice codes and present the lattice subspace decomposition used in our block Markov strategy. In Section \[sect:achievable.rate\] we present our cooperative computation strategies in detail and prove that they achieve the computation rates claimed in Section \[sect:main.results\]. In Section \[sect:DMT\] we perform a high-SNR analysis of our strategies and prove that they provide the diversity-multiplexing gains claimed in Section \[sect:main.results\]. In Section \[sect:simluations\] we present a few numerical examples to showcase the benefits of our approach. Finally, we conclude with Section \[sect:conclusion\]. Preliminaries {#sect:prelim} ============= System model ------------ In the [*cooperative compute-and-forward network*]{}, depicted in Figure \[fig:system.model\], $L$ transmitters communicate with $M \leq L$ receivers over the wireless medium. Each of the $L$ users has $T$ messages $\wbf_l(t) \in \mathbb{F}_p^k$, for $1 \leq t \leq T$. Structurally, this network resembles the compound multiple-access channel or, when $M=L$, the interference channel. However, unlike those more traditional networks, here each receiver intends to decode a finite-field linear combination[^2] of the transmitters’ messages: $$\fbf_m(t) = \bigoplus_{l=1}^L a_{lm} \odot \wbf_l(t),$$ for $a_{lm} \in \mathbb{Z}$. Let the matrix $\mathbf{A} = [a_{lm}] \in \mathbb{Z}^{L \times M}$ describe the functions computed by the receivers. We divide transmissions into $T+1$ blocks of $n$ channel uses each. At block $t$, each transmitter $l$ broadcasts a signal $\xbf_l(t)~\in~\mathbb{R}^n$, subject to an average power constraint: $$\frac{1}{n}{\left\|{\xbf_l(t)}\right\|}^2 \leq P,$$ for some $P > 0$. The superposition of the transmitters’ signals, scaled by channel coefficients and corrupted by noise, arrives at each receiver: $$\ybf_m(t) = \sum_{l=1}^Lh_{lm}\xbf_l(t) + \nbf(t),$$ where $h_{lm} \in \mathbb{R}$ is the channel coefficient from transmitter $l$ to receiver $m$, and $\nbf(t)$ is a white, unit-variance Gaussian random vector. For convenience, we gather the channel coefficients into the matrix $\mathbf{H} = [h_{lm}]$. Each transmitter $l$ also obtains the noisy superposition of the other transmitters’ signals: $$\zbf_l(t) = \sum_{\substack{l^\prime=1 \\ l^\prime \neq l}}^Lg_{l^\prime l}\xbf_{l^\prime}(t)+\nbf_l(t),$$ where $g_{l^\prime l} \in \mathbb{R}$ is the channel coefficient from transmitter $l^\prime$ to transmitter $l$, and $\nbf_l(t)$ is again white, unit-variance Gaussian. Again we stack the channel coefficients into a matrix $\mathbf{G} = [g_{l^\prime l}]$ with diagonal elements equal to zero. The choice of zero for the diagonal elements implies [*full-duplex*]{} operation, meaning that transmitters can transmit and receive simultaneously. We further assume that channel matrices $\mathbf{H}$ and $\mathbf{G}$ are fixed and known globally among the transmitters and receivers. We will refer occasionally to the [*non-cooperative compute-and-forward network*]{}, which is identical to the cooperative network except that the transmitters have no access to each other’s transmissions. We model this formally by fixing $\mathbf{G}$ as the all-zero matrix. Computation capacity -------------------- We are interested in the [*computation capacity*]{} of the network. Since the receivers recover functions of incoming messages, rather than the messages themselves, the computation capacity is defined somewhat differently than the capacity of ordinary channels. We endow each transmitter with an encoder $E_l: \mathbb{F}_p^{k \times T} \times \mathbb{R}^{n \times T} \to \mathbb{R}^{n \times (T+1)}$. That is, the encoder $E_l$ takes as its input the messages $\mathbf{w}_l(t)$ and the received signals $\mathbf{z}_l(t)$ and generates as its output the codewords $\mathbf{x}_l(t)$. We impose a causality restriction on $E_l$: the output codeword $\mathbf{x}_l(t)$ may depend on received signals $\mathbf{z}_l(s)$ only for $s < t$. As usual, the encoding rate is defined as the logarithm of the cardinality of the message set divided by the number of channel realizations over which the messages are encoded: $$R = \frac{T \log_2(|\mathbb{F}_p^k|)}{n(T+1)} = \frac{T k \log_2(p)}{n(T+1)} \approx \frac{k \log_2(p)}{n},$$ where the approximation holds for large $T$. Note that this is the [*symmetric*]{} rate among all transmitters. We endow each receiver with a decoder $D_m : \mathbb{R}^{n \times (T+1)} \to \mathbb{F}_p^{k \times T}$, taking as inputs the received signals $\mathbf{y}_m(t)$ and generating as outputs the estimates $\hat{\mathbf{f}}_m(t)$. Let the absolute probability of error be the probability that any receiver makes an incorrect estimate of any of the desired functions: $$P_e = \mathrm{Pr}\{ \hat{\mathbf{f}}_m(t) \neq \mathbf{f}_m(t), \text{for any } 1 \leq m \leq M, 1 \leq t \leq T \}.$$ We say that a computation rate $R$ is [*achievable*]{} if for any $\epsilon > 0$ there exists a sequence of encoders with encoding rate greater than $R - \epsilon$ and decoders such that $P_e \to 0$ as $n \to \infty$. For fixed channel gains $\mathbf{H},\mathbf{G}$, function coefficients $\mathbf{A}$, and transmit power $P$, let $R(\mathbf{H},\mathbf{G},\mathbf{A},P)$ denote the supremum over all achievable computation rates. In order to define the computation capacity we need to place limitations on the permissible function coefficients $\mathbf{A}$. Otherwise we could choose a trivial coefficient matrix, such as the all-zero matrix, for which the achievable computation rate is unbounded. We therefore require that $\mathbf{A}$ be a member of the following set: $$\mathcal{A} = \{ \mathbf{A} \in \mathbb{Z}^{L \times M} : \mathrm{rank}(\mathbf{A}) = M, \ \forall \ m \ \exists \ l \text{ such that } a_{ml} \neq 0\}.$$ The first condition ensures that the recovered functions retain as much information as possible about the individual transmitters’ messages; for $L=M$ it implies that one can recover the individual messages from the recovered functions. The second condition, which is redundant for $L=M$, ensures that each transmitter is represented in the recovered messages; the receivers cannot simply ignore a transmitter in order to achieve a higher computation rate. Finally, we define the [*computation capacity*]{} as the supremum of achievable rates over the set of permissible coefficient matrices: $$\label{eqn:computation.capacity} C(\mathbf{H},\mathbf{G},P) = \sup_{\mathbf{A} \in \mathcal{A}} R(\mathbf{H},\mathbf{G},\mathbf{A},P).$$ In their seminal work, Nazer and Gastpar developed a computation strategy based on nested lattice codes [@nazer:IT11]. It achieves the following computation rate: $$\label{eqn:non.cooperative.rate} R_{\mathrm{nc}}(\mathbf{H},P) = \max_{\mathbf{A} \in \mathcal{A}} \min_{1 \leq m \leq M} \left[ \frac{1}{2} \log_2(1 + P{\left\|{\mathbf{h}_m}\right\|}^2) - \frac{1}{2}\log_2({\left\|{\mathbf{a}_m}\right\|}^2 + P({\left\|{\mathbf{a}_m}\right\|}^2{\left\|{\mathbf{h}_m}\right\|}^2 - |\mathbf{a}_m^T\mathbf{h}_m|^2))\right]^+.$$ The first term in (\[eqn:non.cooperative.rate\]) corresponds to the power in the received signal, whereas the second term is a penalty determined by the gap in the Cauchy-Schwarz inequality between $\mathbf{h}_m$ and $\mathbf{a}_m$. The closer $\mathbf{h}_m$ and $\mathbf{a}_m$ are to being co-linear, the smaller is the rate penalty. Since the Nazer-Gastpar scheme was designed for a non-cooperative network, the rate does not depend on $\mathbf{G}$; nevertheless, it serves as a lower bound on the cooperative computation capacity $C(\mathbf{H},\mathbf{G})$. In the sequel we present a cooperative computation strategy based in part on the Nazer-Gastpar scheme. Diversity-multiplexing tradeoff ------------------------------- One advantage of user cooperation is that cooperating transmitters can achieve performance similar to that of a multiple-antenna transmitter. Multiple antennas can improve performance on two fronts: increased reliability in the presence of slow channel fading, and increased throughput. In the high-SNR regime, the [*diversity-multiplexing tradeoff*]{} quantifies this improvement [@zheng:IT03]. Let the elements of $\mathbf{H}$ and $\mathbf{G}$ be identically and independently distributed according to a Rayleigh distribution. Next, suppose there is a scheme that achieves the computation rate $R_{\mathrm {scheme}}(\mathbf{H},\mathbf{G},P)$. Then, the [*diversity order at multiplexing gain $r$*]{} is defined as $$\label{eqn:dmt} d(r) = \lim_{P \to \infty} \frac{\log \mathrm{Pr}\{R_{\mathrm{scheme}}(\mathbf{H},\mathbf{G},P) < \frac{r}{2} \log(P)\}}{\log P}.$$ In other words, $d(r)$ is the exponent of the outage probability, with the rate taken to have multiplexing gain $r$, as the SNR goes to infinity. The diversity-multiplexing tradeoff of the system, denoted by $d^*(r)$, is the supremum of $d(r)$ over all possible schemes. The multiplexing gain for compute-and-forward is studied in [@niesen:ISIT11]. There it is shown that, using the Nazer-Gastpar approach, the multiplexing gain can be no higher than $\frac{2}{L+1}$. In other words, $d(r) = 0$ for $r > \frac{2}{L+1}$ for this scheme. In the sequel we show that we can achieve a better diversity-multiplexing tradeoff, including a multiplexing gain of unity, for both cooperative and non-cooperative networks having a single receiver. Main Results {#sect:main.results} ============ Upper bounds {#sect:upper.bounds} ------------ First we present two upper bounds on the cooperative computation rate, the proofs of which are contained in the Appendix \[app:upper.bounds\]. We obtain the first bound by supposing that the transmitters are capable of perfect cooperation, which is equivalent to having a genie supply all messages to each transmitter. The problem then reduces to a multiple-input, single-output (MISO) broadcast channel, the capacity of which is known [@weingarten:IT06]. In the sequel we use this result to bound the diversity-multiplexing tradeoff. \[thm:miso.upper.bound\] Let the capacity region of a Gaussian MISO broadcast channel be denoted by $$\begin{aligned} \label{eqn:dirty.paper} \mathcal{C}_{\mathrm{miso}}(\mathbf{H},P) = \mathrm{conv} \left\{ \bigcup_{\pi \in \Pi} \left\{\mathbf{r} : r_m \leq \frac{1}{2} \log_2\left(1 + \frac{\mathbf{h}_{\pi(m)}\mathbf{V}_{\pi(m)}\mathbf{h}_{\pi(m)}^T}{\mathbf{h}_{\pi(m)}\sum_{i=1}^{m-1}\mathbf{V}_{\pi(i)}\mathbf{h}_{\pi(m)}^T+1}\right)\right\}\right\}, \end{aligned}$$ where $\mathrm{conv}\{\cdot\}$ is the convex hull, $\Pi$ is the set of permutations from $\{1,\cdots L\}$ to itself, and $\mathbf{V}_m$ is a collection of positive semi-definite matrices such that $\sum_{m=1}^M\mathrm{tr}(\mathbf{V}_m) \leq NP$. Then the computation capacity of the cooperative compute-and-forward network is bounded above by $$C(\mathbf{H},\mathbf{G},P) \leq R^+_{\mathrm{miso}}(\mathbf{H},P),$$ where $$\begin{aligned} \label{eqn:common.miso.rate} R^+_{\mathrm{miso}}(\mathbf{H},P) = \sup \{r : r\mathbf{1} \in \mathcal{C}_{\mathrm{miso}}(\mathbf{H},P) \} \end{aligned}$$ is the symmetric-rate capacity of the Gaussian MISO broadcast channel. We obtain the next bound by supposing a genie supplies to the receivers all messages except for those of a single transmitter $l$. Then the receivers need only to recover the messages of transmitter $l$ in order to compute any suitable set of functions. This converts the system to a compound relay channel in which the other transmitters serve as dedicated relays; we bound the capacity of this channel using cut-set arguments. This upper bound is somewhat more realistic than $R^+_{\mathrm{miso}}$, and we use it in Section \[sect:simluations\] for comparisons to our achievable rates. \[thm:relay.upper.bound\] For each transmitter $1 \leq l \leq L$, let $\mathbf{S}_{l} = \{1,\cdots,l-1,l+1,\cdots,l\}$ be the set of transmitters [*other*]{} than transmitter $l$. Then the computation capacity of the cooperative compute-and-forward network is bounded above by $$C(\mathbf{H},\mathbf{G},P) \leq R_{\mathrm{single}}^+(\mathbf{H},\mathbf{G},P),$$ where $$R_{\mathrm{single}}^+(\mathbf{H},\mathbf{G},P) = \min_{1 \leq l \leq L} \max_{\mathbf{A} \in \mathcal{A}} \min_{m, a_{lm} \neq 0} \max_{p(\mathbf{x})} \min_{S \in \mathcal{S}_l} I(x_l,x_{S}; y_m,z_{S^C} | x_{S^C}),$$ where $p(\mathbf{x})$ is any distribution over the transmitted signals $(x_1, \cdots x_L)^T$ satisfying the input power constraint. Achievable rates ---------------- Here we present the computation rates achieved by our cooperative strategy. Our approach is decode-and-forward in nature: at one block transmitters send out lattice codewords corresponding to their individual messages; these messages are decoded by other transmitters. At the next block transmitters cooperatively encode resolution information to assist the receivers. As with any decode-and-forward strategy, we must contend with the fact that it may be difficult for transmitters to decode each other’s messages. We therefore require only some of the transmitters to cooperate.[^3] A subset $\mathcal{B}$ of the transmitters decodes the messages of every other user, after which they cooperatively transmit resolution information to the receivers. Transmitters not in $\mathcal{B}$, not having decoded incoming messages, do not send any resolution information. We present the details of this strategy, as well as an achievability proof, in Section \[sect:achievable.rate\]. \[thm:subset\] Let $\mathcal{B} \subset \{1, \cdots, L\}$. In the cooperative compute-and-forward network, the following computation rate is achievable: $$\begin{gathered} \label{eqn:subset} R_{c}(\mathbf{H},\mathbf{G},P) = \max_{\mathbf{A} \in \mathcal{A}} \min\bigg\{ \min_{l \in \mathcal{B}} C_{\mathrm{mac}}(\mathbf{g}_l[1:l-1,l+1:L] \circ \mathbf{v}_0[1:l-1,l+1:L], P, 1), \\ \min_{1 \leq m \leq M} \bigg\{ \frac{1}{2} \log_2\left(1+ \frac{P|\hbf_m^T \vbf_m|^2}{1 + I_{m,r}}\right) + \\ \left [\frac{1}{2} \log_2({\left\|{ P \hbf_m \circ \vbf_0}\right\|}^2 + I_{m,v}) - \frac{1}{2}\log_2\left({\left\|{\mathbf{a}_m}\right\|}^2(1+I_{m,v}) + P\left({\left\|{\mathbf{a}_m}\right\|}^2{\left\|{\mathbf{h}_m \circ \mathbf{v}_0}\right\|}^2 - |\mathbf{a}_m^T(\mathbf{h}_m \circ \mathbf{v}_0)|^2\right)\right)\right]^+ \bigg\} \bigg\}, \end{gathered}$$ where $$I_{m,r} = P\left({\left\|{ \hbf_m \circ \vbf_0 }\right\|}^2 + \sum_{m^\prime \neq m,0} |\hbf^T_m \vbf_{m^\prime} |^2\right)$$ is the interference power seen at receiver $m$ as it decodes its resolution information, $$I_{m,v} = P\sum_{m^\prime \neq m,0} |\hbf_m^T \vbf_{m^\prime}|^2$$ is the interference seen at receiver $m$ as it decodes the vestigial information, and for any vectors $\vbf_0,\vbf_1,\cdots, \vbf_M$ such that $$\sum_{m=0}^M |v_{lm}|^2 \leq 1, \forall l$$ and $v_{lm} = 0$ for $l \neq \mathcal{B}$ and $m>0$. The achievable rate (\[eqn:subset\]) is a bit difficult to parse, so we take time here to describe each of its three components. First is the rate of a Gaussian multiple-access channel, which corresponds to the rate at which cooperating transmitters can decode others’ messages. Second is the rate at which each receiver can decode the resolution information, which is that of a virtual MISO link between cooperating transmitters and the receiver; signals unrelated to the resolution information are treated as noise. Third is the rate at which the receivers, having already decoded the resolution information, can decode the vestigial component of the desired combination of lattice points; this is the Nazer-Gastpar rate of (\[eqn:non.cooperative.rate\]), with resolution information intended for other receivers treated as noise. Each transmitter splits its power between sending its own lattice codewords and cooperatively sending resolution information. The split is defined by the steering vectors $\mathbf{v}_0, \mathbf{v}_1, \cdots, \mathbf{v}_M$. Each element $v_{l0}$ dictates the fraction of power transmitter $l$ expends on its own lattice codewords. For cooperating transmitter $l$, each element $v_{lm}$ dictates the fraction of power expended on resolution information for receiver $m$. The steering vectors introduce two separate notions of alignment. First, we can choose $\mathbf{v}_0$ in order to minimize the Cauchy-Schwarz penalty in (\[eqn:subset\]). Second, we can choose the remaining vectors $\mathbf{v}_m$ to trade off between increasing the coherence gain at the intended receivers and decreasing the interference generated at other receivers. Finding the optimum steering vectors is a non-covex problem; for further results and in our simulations we rely on a few heuristic means for selecting them. We can obtain a simpler expression for the achievable rate by choosing $\mathcal{B} = \{1, \cdots, L\}$ and taking the steering vectors $\mathbf{v}_1, \cdots, \mathbf{v}_M$ to be zero-forcing beamformers. Thus the cooperative signals do not interfere at other receivers. \[cor:zero.forcing\] The following computation rate is achievable for the cooperative compute-and-forward network: $$\begin{gathered} R_{\mathbf{zf}}(\mathbf{H},\mathbf{G},P) = \max_{\mathbf{A} \in \mathcal{A}} \min\bigg\{ \min_{1 \leq l \leq L}C_{\mathrm{mac}}(\mathbf{g}_l[1:l-1,l+1:L] \circ \mathbf{v}_0[1:l-1,l+1:L], P, 1),\\ \min_{1 \leq m \leq M} \bigg[\frac{1}{2}\log_2(1+P({\left\|{\mathbf{h}_m\circ \mathbf{v}_0}\right\|}^2 + |\mathbf{h}_m^T\mathbf{v}_m|^2)) - \\ \frac{1}{2}\log_2\left({\left\|{\mathbf{a}_m}\right\|}^2 + P\left({\left\|{\mathbf{a}_m}\right\|}^2{\left\|{\mathbf{h}_m \circ \mathbf{v}_0}\right\|}^2 - |\mathbf{a}_m^T(\mathbf{h}_m \circ \mathbf{v}_0)|^2\right)\right)\bigg]^+ \bigg\}, \end{gathered}$$ for any vectors $\mathbf{v}_0, \mathbf{v}_1, \cdots, \mathbf{v}_M$ satisfying $$\sum_{m=0}^M |v_{lm}|^2 \leq 1$$ and $$\mathbf{v}_m^T\mathbf{h}_{m^\prime} = 0, \forall \ m \neq m^\prime.$$ Since $L \geq M$, it is possible to choose non-trivial zero-forcing beamforming vectors for almost every $\mathbf{H}$. Finally, choosing $\mathcal{B} = \emptyset$, we obtain an achievable rate for both the cooperative and non-cooperative compute-and-forward network. This yields a rate similar to (\[eqn:non.cooperative.rate\]), except that each transmitter can adjust its transmit power in order to tune the effective channels to match the desired linear functions. In fact this rate is a special case of the “superposition” compute-and-forward presented in [@nazer:IT11 Theorem 13]. \[cor:non.cooperative\] In both the non-cooperative compute-and-forward network and the cooperative compute-and-forward network, the following rate is achievable: $$\begin{gathered} R(\mathbf{H},\mathbf{G},P) = \max_{\mathbf{A} \in \mathcal{A}} \min_{1 \leq m \leq M} \bigg[\frac{1}{2}\log_2(1+P({\left\|{\mathbf{h}_m\circ \mathbf{v}_0}\right\|}^2)) - \\ \frac{1}{2}\log_2\left({\left\|{\mathbf{a}_m}\right\|}^2 + P\left({\left\|{\mathbf{a}_m}\right\|}^2{\left\|{\mathbf{h}_m \circ \mathbf{v}_0}\right\|}^2 - |\mathbf{a}_m^T(\mathbf{h}_m \circ \mathbf{v}_0)|^2\right)\right)\bigg]^+, \end{gathered}$$ for any $\mathbf{v}_0$ satisfying $$|v_{l0}|^2 \leq 1, \forall \ 1 \leq l \leq L.$$ Diversity-multiplexing tradeoff ------------------------------- Here we present our diversity-multiplexing tradeoff results, the proofs of which are presented in Section \[sect:DMT\]. We begin with the non-cooperative case. \[thm:non.cooperative.dmt\] For the non-cooperative compute-and-forward network, the diversity-multiplexing tradeoff for any scheme is upper-bounded as follows: $$d^*(r) \leq d^+_{\mathrm{nc}}(r) = 1-r.$$ For the case of $M=1$, $d^*(r) = d^+_{\mathrm{nc}}(r)$. In other words, the DMT of the non-cooperative compute-and-forward network is bounded above by that of a scalar Gaussian channel. In the case of a single receiver, we can achieve this upper bound with lattice codes and signal alignment. With the steering vector $\mathbf{v}_0$ chosen such that the equivalent channel vector is a constant, the achievable rate—and therefore the error probability—is approximately that of a single SISO link. Next, we look at the DMT of the cooperative compute-and-forward network. We start by presenting an upper bound. \[thm:cooperative.dmt\] For the cooperative compute-and-forward network, the diversity-multiplexing tradeoff is upper-bounded as $$d^*(r) \leq d^+_\mathrm{c}(r) = L(1-r).$$ In other words, the DMT is upper-bounded by that of a single MISO link. In the case of a single receiver, we derive two lower bounds on the DMT. The first is derived using a rather simple strategy employing time sharing and Gaussian codes. \[thm:random.dmt\] For the cooperative compute-and-forward network, the following diversity-multiplexing tradeoff is achievable: $$d_{\mathrm{random}}^-(r) = L \min\{1-2r, (L-1)(1 - 2(L-1)r) \}.$$ In particular, $d_{\mathrm{random}}^-(0) = L$. Since it involves time sharing, the strategy used in Theorem \[thm:random.dmt\] has poor multiplexing performance. It does, however, achieve the full diversity gain of $L$. The second bound is derived using the cooperative computation strategy of Theorem \[thm:subset\]. \[thm:lattice.dmt\] For the cooperative compute-and-forward network, the following diversity-multiplexing tradeoff is achievable: $$\begin{gathered} d_{\mathrm{lattice}}^-(r) = 1-r + \min\{[1-2r]^+,[(L-1)(1-rL)]^+ \} + \\ \max_{0 \leq x \leq 1}(L-2) \min\{[1-x-r]^+,[(L-1)(1-(L-1)r-x)]^+,[x-r]^+ \}. \end{gathered}$$ Here, $d_{\mathrm{lattice}}^-(0) = 2 + \frac{L-2}{2}$. Here the main difficulty is the Cauchy-Schwarz penalty inherent to lattice coding. It turns out that choosing $\mathbf{v}_0$ to align with the channels, as we did in the non-cooperative case, precludes cooperation with high probability. We therefore choose $\mathbf{v}_0$ to be constant, taking the Cauchy-Schwarz penalty “on the chin.” We balance the transmit power between sending fresh information, which helps transmitters decode others’ messages, and sending resolution information, which helps the receiver decode the desired linear combination. Choosing the balance properly, the benefits of cooperation outweigh the Cauchy-Schwarz penalty, but only enough to obtain a diversity gain of approximately $1/2$ per transmitter. Nevertheless, for higher multiplexing gains lattice coding outperforms the strategy of Theorem \[thm:random.dmt\]. We plot the DMT bounds in Figure \[fig:dmt.plot\]. For $L=2$ lattice coding is sufficient to achieve full diversity, and the DMT achieved by lattice coding strictly dominates that achieved by random coding. For $L>2$, lattice coding achieves better performance only for sufficiently high multiplexing gain. Random coding fails altogether at multiplexing gains higher than $(L-1)/2$ due to the need for transmitters to decode $L-1$ separate messages and the need for time-sharing. Lattice coding, on the other hand, maintains non-zero diversity for every any $0 \leq r \leq 1$. Between the two strategies we obtain the corner points of the DMT region. Lattice codes {#sect:lattices} ============= In this section we first introduce the basics of lattice codes, after which we present the lattice decomposition employed in our cooperative computation strategy. Capacity-achieving lattice codes -------------------------------- Formally, a lattice $\Lambda$ is a discrete additive subgroup of $\mathbb{R}^n$, which implies that for any $\lambda_1, \lambda_2 \in \Lambda$ we have $\lambda_1 + \lambda_2 \in \Lambda$ and $\lambda_1 - \lambda_2 \in \Lambda$. Any lattice can be generated by taking integer combinations of (not necessarily unique) basis vectors. Choosing these basis vectors as columns, we form the [*generator matrix*]{} of $\Lambda$, denoted by $\mathbf{G} \in \mathbb{R}^{n\times n}$: $$\Lambda = \mathbf{G}\mathbb{Z}^n.$$ We let $Q_{\Lambda}$ denote the [*lattice quantizer*]{}, which maps any point $\xbf \in \mathbb{R}^n$ to the nearest point in $\Lambda$: $$Q_{\Lambda}(\mathbf{x}) = \arg\min_{\lambda \in \Lambda} {\left\|{\mathbf{x} - \lambda}\right\|}.$$ The lattice $\Lambda$ induces a partition of $\mathbb{R}^n$ into the [*Voronoi regions*]{} $\Vcal(\lambda)$ of each lattice point $\lambda \in \Lambda$: $$\Vcal(\lambda) = \{\mathbf{x} \in \Re^n : Q_{\Lambda}(\mathbf{x}) = \lambda \},$$ where ties are broken arbitrarily. In other words, the Voronoi region of $\lambda \in \Lambda$ is simply the set of points that are closer to $\lambda$ than to any other lattice point. Let $\Vcal = \Vcal(0)$ be the [*fundamental Voronoi region*]{} of $\Lambda$. The$\mod$operation with respect to $\Lambda$ returns the quantization error $$\mathbf{x} \mod \Lambda = \mathbf{x} - Q_{\Lambda}(\mathbf{x}),$$ which is always a member of $\Vcal$. The$\mod$operation allows one to draw an analogy with modulo arithmetic over a finite field. Just as modulo arithmetic ensures that the result remains a member of the finite field, performing arithmetic modulo $\Lambda$ “wraps” the result within $\Vcal$. The$\mod$operation obeys the associativity property: $$[[\xbf] \mod \Lambda + \ybf] \mod \Lambda = [\xbf + \ybf] \mod \Lambda.$$ The [*second moment*]{} $\sigma^2(\Lambda)$ quantifies the average power of a random variable uniformly distributed inside $\Vcal$: $$\sigma^2(\Lambda) = \frac{1}{n \mathrm{Vol}(\Vcal)}\int_{\Vcal} {\left\|{\xbf}\right\|}^2 d\xbf,$$ where $\mathrm{Vol}(A)$ is the volume of a set $A \subset \Re^n$. The [*normalized second moment*]{} is defined as: $$G(\Lambda) = \frac{\sigma^2(\Lambda)}{\mathrm{Vol}(\Vcal)^{\frac{2}{n}}}.$$ The normalized second moment provides a measure of the efficiency of $\Lambda$ as a shaping region. The closer $\Vcal$ is to being spherical, the smaller $G(\Lambda)$ is. The [*covering radius*]{} $r_{\mathrm{cov}}(\Lambda)$ is the radius of the smallest sphere that covers $\Vcal$: $$r_{\mathrm{cov}}(\Lambda) = \inf_r \{r>0 | \Vcal \subset r \mathcal{B}_n \},$$ where $\mathcal{B}_n$ is the unit sphere in $\mathbb{R}^n$. The [*effective radius*]{} $r_{\mathrm{eff}}(\Lambda)$ be the radius of a sphere with the same volume as $\Vcal$: $$r_{\mathrm{eff}}(\Lambda) = \left( \frac{\mathrm{Vol}(\Vcal)}{\mathrm{Vol}(\mathcal{B}_n)}\right)^{\frac{1}{n}}.$$ Note that $r_{\mathrm{cov}}(\Lambda) \geq r_{\mathrm{eff}}(\Lambda)$. In order to construct lattice codebooks suitable for proving information-theoretic results, we require [*sequences*]{} of lattices that asymptotically satisfy several desirable properties. For example, we say that a sequence of lattices $\{\Lambda^{(n)}\}, \Lambda^{(n)} \in \mathbb{R}^n$, is [*good for covering*]{} or [*Rogers good*]{} [@rogers:M59] provided the covering radius approaches the effective radius: $$\lim_{n \to \infty} \frac{r_{\mathrm{cov}}(\Lambda^{(n)})}{r_{\mathrm{eff}}(\Lambda^{(n)})} = 1.$$ Similarly, a sequence of lattices is [*good for quantization*]{} provided $$\lim_{n \to \infty} G(\Lambda^{(n)}) = \frac{1}{2\pi e}.$$ Finally, let $\zbf \sim \mathcal{N}(0,\sigma^2\mathbf{I})$ be a Gaussian random vector. Define the [*volume-to-noise ratio*]{} $\mu(\Lambda,P_e)$ as $$\mu(\Lambda,P_e) = \frac{(\mathrm{Vol}(\Vcal))^{\frac{2}{n}}}{\sigma^2},$$ where $\sigma^2$ is chosen such that $\mathrm{Pr}\{\zbf \notin \Vcal \} = P_e$. A sequence of lattices $\Lambda^{(n)}$ is [*good for AWGN coding*]{} or [*Poltyrev good*]{} if $$\lim_{n \to \infty} \mu(\Lambda^{(n)},P_e) = 2\pi e.$$ The existence of such sequences was proven by Poltyrev in [@poltyrev:IT94]. Furthermore, Erez et al. proved that there exist sequences of lattices that are simultaneously good for covering, quantization, and AWGN coding [@erez:IT05]. Lattice codebooks are constructed using [*nested lattices*]{}, as depicted in Figure \[fig:lattice\]. Here we review the construction of codebooks sufficient to achieve capacity for the AWGN point-to-point channel, which is the model for codebooks to be used throughout this paper. Let $\Lambda_s^{(n)}$ be a sequence of [*shaping lattices*]{} that are good for covering and AWGN coding and satisfy $\sigma^2(\Lambda_s^{(n)}) = 1$, and let $\mathbf{G}_s^{(n)}$ denote generator matrices for each lattice in the sequence. Then, following [@krithvasan:TR], we adapt Construction A [@loeliger:IT97] to construct a sequence of coding lattices $\Lambda_c^{(n)} \supset \Lambda_s^{(n)}$. The construction process goes as follows: 1. For each $n$, choose an integer $k$ and a prime $p$. Draw a $n \times k$ matrix $\mathbf{F}_c^{(n)} \in \mathbb{F}_p^{n \times k}$ randomly and uniformly. 2. Construct the linear codebook over $\mathbb{F}_p$ defined by $\mathbf{F}_c^{(n)}$: $$\hat{\mathcal{C}}^{(n)} = \mathbf{F}_c^{(n)}\mathbb{F}_p^k$$ 3. “Lift” the codebook $\hat{\mathcal{C}}^{(n)}$ to $\mathbb{R}^n$ by defining the lattice $$\hat{\Lambda}_c^{(n)} = p^{-1}\hat{\mathcal{C}}^{(n)} + \mathbb{Z}^n.$$ 4. Finally, rotate $\hat{\Lambda}_c^{(n)}$ so that it is nested inside $\Lambda_s^{(n)}$: $$\Lambda_c^{(n)} = \mathbf{G}_s^{(n)} \hat{\Lambda}_c^{(n)}.$$ We form the lattice codebook by taking the intersection of the coding lattice with the fundamental Voronoi region of the shaping lattice: $$\mathcal{C}^{(n)} = \Lambda_c^{(n)} \cap \Vcal_s^{(n)}.$$ The rate of this codebook is $$R = \frac{1}{n}\log_2 | \mathcal{C}^{(n)} | = \frac{k \log_2(p)}{n}.$$ It is shown in [@nazer:IT11] that choosing $p$ such that $n/p \to 0$ as $n \to \infty$ guarantees that the sequence of coding lattices $\Lambda_c^{(n)}$ is good for AWGN coding. For any desired rate $R > 0$, we can construct an appropriate sequence of codebooks by choosing $p = n \log_2(n)$ and $k = \lfloor \frac{nR}{\log_2(p)} \rfloor$. In essence, the preceding codebook construction allows us to take a linear block code over $\mathbb{F}_p$ and to create a corresponding linear code over Euclidean space. If the underlying linear code achieves capacity, as does the ensemble of random linear codes, so too does the resulting lattice codebook. We can use any linear code in place of the one chosen above; the performance cost is only the gap to capacity of the linear code chosen. For the lattice compute-and-forward proposed in [@nazer:IT11], an important fact is that there exists a mapping from finite-field messages to lattice codewords that preserves linearity. That is, the mapping sends finite-field linear combinations of messages to integer sums of lattice points modulo the shaping lattice. Formally, this implies that there is an isomorphism between the additive group of field elements and the group of lattice codewords modulo the shaping lattice. We state this result in the following lemma. \[lem:isomorphism\] There exists an isomorphism $\phi : \mathbb{F}_p^k \to \Ccal^{(n)}$, namely $$\phi(\wbf) = [\Gbf_s^{(n)}p^{-1}\Fbf_c^{(n)}\wbf] \mod \Lambda_s^{(n)}.$$ We need to show that $\phi$ is a bijection and that it respects the group operation; that is, $\phi(\mathbf{w}_1 \oplus \mathbf{w}_2) = [\phi(\mathbf{w}_1) + \phi(\mathbf{w}_2)] \mod \Lambda_s^{(n)}$ for any $\mathbf{w}_1,\mathbf{w}_2 \in \mathbb{F}_p^k$. That $\phi$ is a bijection was shown in [@nazer:IT11 Lemma 5]. To see that $\phi$ respects the group operation, we resort to direct computation: $$\begin{aligned} \phi(\mathbf{w}_1 \oplus \mathbf{w}_2) &= [\Gbf_s^{(n)}p^{-1}\Fbf_c^{(n)}(\mathbf{w}_1 \oplus \mathbf{w}_2)] \mod \Lambda_s^{(n)} \\ &= [\Gbf_s^{(n)}p^{-1}(\Fbf_c^{(n)}(\mathbf{w}_1 + \mathbf{w}_2) + p\mathbf{i})] \mod \Lambda_s^{(n)},\end{aligned}$$ where $\mathbf{i} \in \mathbb{Z}^n$ is a vector of integers corresponding to the discrepancy between real-valued and modulo-$p$ arithmetic. Continuing, we get $$\begin{aligned} \phi(\mathbf{w}_1 \oplus \mathbf{w}_2) &= [\Gbf_s^{(n)}p^{-1}\Fbf_c^{(n)}(\mathbf{w}_1 + \mathbf{w}_2) + \Gbf_s^{(n)}\mathbf{i}] \mod \Lambda_s^{(n)} \\ &= [\Gbf_s^{(n)}p^{-1}\Fbf_c^{(n)}(\mathbf{w}_1 + \mathbf{w}_2)] \mod \Lambda_s^{(n)} \\ &= [\phi(\mathbf{w}_1) + \phi(\mathbf{w}_2)] \mod \Lambda_s^{(n)}.\end{aligned}$$ where the last equality is due to the fact that $\Gbf_s^{(n)}\mathbf{i} \in \Lambda_s^{(n)}$ and that adding a member of $\Lambda_s^{(n)}$ does not change the result of the arithmetic modulo $\Lambda_s^{(n)}$. Lattice subspaces {#sect:lattice.subspaces} ----------------- In our cooperative computation strategy, we tailor block Markov encoding to lattice codes. To do so, we introduce a key ingredient of our approach: the decomposition of the lattice codebook into subspaces. Let $k_r \leq k$, and let $\Fbf_r^{(n)} \in \mathbb{F}_p^{n \times k_r}$ denote the matrix composed of the first $k_r$ columns of $\Fbf_c^{(n)}$. Similarly, let $k_v = k - k_r$, and let $\Fbf_v^{(n)} \in \mathbb{F}_p^{n \times k_v}$ denote the matrix of the remaining $k_v$ columns. Then define the [*resolution lattice*]{} $\Lambda_r$ and the [*vestigiall[^4] lattice*]{} $\Lambda_v$ as $$\begin{aligned} \Lambda_r^{(n)} &= \Gbf_s^{(n)}(p^{-1}\Fbf_r^{(n)}\mathbb{F}_p^{k_r} + \mathbb{Z}^n) \\ \Lambda_v^{(n)} &= \Gbf_s^{(n)}(p^{-1}\Fbf_v^{(n)}\mathbb{F}_p^{k_v} + \mathbb{Z}^n).\end{aligned}$$ Since these sequences of lattices are special cases of the lattice construction from the previous subsection, each sequence is individually good for AWGN coding. By construction $\Lambda_c^{(n)} = \Lambda_r^{(n)} + \Lambda_v^{(n)}$ and $\Lambda_s^{(n)} \subset \Lambda_r^{(n)}, \Lambda_v^{(n)} \subset \Lambda_c^{(n)}$. Define the resolution and vestigial codebooks $$\begin{aligned} \Ccal_r^{(n)} &= \Lambda_r^{(n)} \cap \Vcal_{\Lambda_s^{(n)}} \\ \Ccal_v^{(n)} &= \Lambda^{(n)}_v \cap \Vcal_{\Lambda_s^{(n)}},\end{aligned}$$ having rates $$\begin{aligned} R_r &= \frac{k_r}{n}\log_2 p \\ R_v &= \frac{k_v}{n}\log_2 p.\end{aligned}$$ By construction $R_r + R_v = R_c$. Furthermore, for any $0 \leq R_r \leq R$, we can choose $k_r = \lfloor \frac{nR_r}{\log_2(p)} \rfloor$ to achieve the desired resolution codebook rate. For any message $\mathbf{w} \in \mathbb{F}_p^k$, we can define the [*projection*]{} onto the resolution and vestigial codebook as follows: $$\begin{aligned} \phi_r(\wbf) &= [\Gbf_s p^{-1}\Fbf_r \wbf[1:k_r]] \mod \Lambda_s \\ \phi_v(\wbf) &= [\Gbf_s p^{-1}\Fbf_v \wbf[k_r+1:k]] \mod \Lambda_s.\end{aligned}$$ Using these projections, we can define a linear decomposition of the lattice codebook, as depicted in Figure \[fig:sublattice.code\]. \[lem:subspace.decomposition\] For any $\wbf \in \mathbb{F}_p^k$, $$\phi(\wbf) = [\phi_r(\wbf) + \phi_v(\wbf)] \mod \Lambda_s^{(n)},$$ This result follows from Lemma \[lem:isomorphism\]. By definition $\wbf = (\wbf^T[1:k_r] \mathbf{0}^T_{k_v})^T \oplus (\mathbf{0}_{k_r}^T \wbf^T[k_r+1:k])^T$, so $$\begin{aligned} \phi(\wbf) &= \phi((\wbf^T[1:k_r], \mathbf{0}^T_{k_v})^T \oplus (\mathbf{0}_{k_r}^T, \wbf^T[k_r+1:k])^T) \\ &= [\phi((\wbf^T[1:k_r], \mathbf{0}^T_{k_v})^T) + \phi((\mathbf{0}_{k_r}^T, \wbf^T[k_r+1:k])^T)] \mod \Lambda_s^{(n)} \\ &= [\phi_r(\wbf) + \phi_v(\wbf)] \mod \Lambda_s^{(n)}, \end{aligned}$$ where the last equality follows from the definition of $\Fbf_r^{(n)}$ and $\Fbf_v^{(n)}$; zeroing out the unwanted portions of $\wbf$ is equivalent to discarding the associated columns of $\Fbf^{(n)}$. The codeword $\phi(\wbf) \in \Ccal^{(n)}$ is therefore the sum of two linearly independent lattice points: $\phi_r(\wbf)$, which we call the [*resolution information*]{} and which encodes the first $k_r\log_2 p$ bits of the message, and $\phi_v(\wbf)$, which we call the [*vestigial information*]{} and which encodes the remaining $k_v \log_2 p$ bits. Furthermore, the decomposition is linear in the sense that the decomposition of sums of lattice points is the same as the sum of decompositions. \[lem:linear.decomposition\] Let $\wbf_1$ and $\wbf_2$ be messages in $\mathbb{F}_p^k$, and let $\wbf = \wbf_1 \oplus \wbf_2$. Then $$\phi_r(\wbf) = [\phi_r(\wbf_1) + \phi_r(\wbf_2)] \mod \Lambda_s^{(n)},$$ and $$\phi_v(\wbf) = [\phi_v(\wbf_1) + \phi_v(\wbf_2)] \mod \Lambda_s^{(n)}.$$ This follows directly from the fact that $\phi$ is an isomorphism: $$\begin{aligned} \phi_r(\wbf) &= \phi_r(\wbf_1 \oplus \wbf_2) \\ &= \phi(\wbf_1[1:k_r] \oplus \wbf_2[1:k_r]) \\ &= [\phi(\wbf_1[1:k_r]) + \phi(\wbf_2[1:k_r])] \mod \Lambda_s^{(n)} \\ &= [\phi_r(\wbf_1) + \phi_r(\wbf_2)] \mod \Lambda_s^{(n)}. \end{aligned}$$ A similar argument holds for $\phi_v$. The preceding decomposition permits a lattice-coding instantiation of block Markov encoding. After the transmission of a lattice codeword, cooperating users can transmit the lower-rate resolution component. The intended receiver first decodes the resolution component and subtracts it from the received signal; the receiver then needs only to decode the lower-rate vestigial component. Although in this paper we apply the technique to compute-and-forward, it can be applied to a variety of relay systems. For example, in a related work [@song:IT11] an innovative lattice [*list decoding*]{} technique is proposed to show that lattice codes can achieve the capacity of the physically degraded three-terminal relay channel. Using our technique, it is straightforward to construct an alternative strategy that establishes the same result. Cooperative Computation: Encoding Strategy and Achievable Rate {#sect:achievable.rate} ============================================================== In this section we describe our cooperative computation strategy and derive its achievable rate, which amounts to a proof of Theorem \[thm:subset\]. Our approach is based on the lattice decomposition described in Section \[sect:lattice.subspaces\]. Messages are communicated in two stages. In the first stage, each transmitter sends the lattice codeword corresponding to its message; this codeword is decoded by a subset of the other transmitters. In the second stage, transmitters cooperatively transmit the resolution component of the linear combinations desired at the receivers. Similarly, receivers decode in two stages. They first decode the resolution component, which they then subtract away from the received signal. Then they need only to decode the vestigial component, which is easier to decode due to its lower rate. Our proof goes in three parts: a description of the encoding scheme, a description of the decoding scheme, and an analysis of the probability of error. [**Encoding:** ]{} Each transmitter employs identical lattice codebooks[^5] $\Ccal$ having rate $R_c$. The codebook $\Ccal$ decomposes into resolution and vestigial codebooks $\Ccal_r$ and $\Ccal_v$ which have respective rates $R_r$ and $R_v$. As noted in Section \[sect:lattice.subspaces\], we have $R_c = R_r + R_v$. Transmitters encode their $T$ messages over $T+1$ blocks as depicted in Table \[tab:bme\]. At block $t$, each transmitter $l$ has a message $\wbf_l(t)$, which it encodes by mapping it to the corresponding codeword in $\Ccal$: $$\lambda_l(t) = \phi(\wbf_l(t)).$$ By Lemma \[lem:subspace.decomposition\], each lattice codeword can be decomposed by projecting onto the resolution and vestigial codebooks: $$\begin{aligned} \lambda_{r,l}(t) &= \phi_r(\wbf_l(t)) \\ \lambda_{v,l}(t) &= \phi_v(\wbf_l(t)). \end{aligned}$$ When transmitting the lattice codeword, each user [*dithers*]{} the lattice point over the shaping region; this ensures that the codebook looks approximately Gaussian as well as makes the codewords of each transmitter statistically independent[^6]. We therefore define the effective codeword $$\cbf_l(t) = [\lambda_l(t) + \tbf_l(t)] \mod \Lambda_s,$$ where $\tbf_l(t)$ is a dither drawn randomly and uniformly over $\Vcal_s$, independent for each $1 \leq l \leq L$ and $1 \leq t \leq T$. Each receiver $m$ intends to recover the finite-field linear combination $\fbf_m(t) = \bigoplus_{l=1}^La_{lm} \wbf_m(t)$, which corresponds to the lattice point $$\lambda_m(t) = \phi(\fbf(t)) = \left[\sum_{l=1}^L a_{lm} \lambda_l(t) \right] \mod \Lambda_s.$$ As with the individual codewords, we can decompose $\lambda_m(t)$ into resolution and vestigial components: $$\begin{aligned} \lambda_{r,m}(t) &= \phi_r(\fbf_m(t)) \\ \lambda_{v,m}(t) &= \phi_v(\fbf_m(t)). \end{aligned}$$ The transmitters in $\mathcal{B}$ will cooperatively transmit $\lambda_{r,m}(t)$ to each receiver, again dithering the lattice point over $\Vcal_s$. The effective codeword is $$\cbf_{r,m} = [\lambda_{r,m}(t) + \sbf_{m}(t)] \mod \Lambda_s,$$ where, similar to before, $\sbf_{m}(t)$ is a dither drawn uniformly over $\Vcal_s$ and independent for each $1 \leq m \leq M$, and $1 \leq t \leq T$. At block $t=1$, each transmitter simply sends its own lattice codeword: $$\xbf_l(1) = \sqrt{P}v_{l0}\cbf_{l}(t).$$ For subsequent blocks $2 \leq t \leq T$, each transmitter in $\mathcal{B}$ sends a combination of “fresh” information corresponding to its own message $\wbf_l(t)$ and resolution information corresponding to the messages sent in the previous time slot. Suppose that each transmitter in $\mathcal{B}$ has successfully decoded $\lambda_{l^\prime}(t-1)$ for each $l^\prime \neq l$. Then each transmitter in $\mathcal{B}$ can construct every $\lambda_m(t)$ and, by extension, every $\lambda_{r,m}(t)$. Every transmitter sends its own lattice codeword, and transmitters in $\mathcal{B}$ send the resolution components for each receiver: $$\xbf_l(t) = \begin{cases} \sqrt{P}\left( v_{l0}\cbf_{l}(t) + \sum_{m=1}^M v_{lm} \cbf_{r,m}(t-1) \right), & \text{ for $l \in \mathcal{B}$} \\ \sqrt{P}v_{l0}\cbf_{l}(t), &\text{ otherwise} \end{cases}.$$ Finally, at block $t = T+1$ there is no new fresh information for the transmitters to send. Each transmitter in $\mathcal{B}$ sends only the resolution information corresponding to block $T$, and the other transmitters send nothing: $$\xbf_l(T+1) = \begin{cases} \sqrt{P}\sum_{m=1}^M v_{lm} \cbf_{r,b(l)m}(T), & \text{ for $l \in \mathcal{B}$} \\ 0 & \text{ otherwise} \end{cases}.$$ $t=1$ $t=2$ $\cdots$ $t=T+1$ ----------------------------------- ------------------- ------------------------------------------------------ ---------- ------------------------------------- $\xbf_1(t), 1 \in \mathcal{B}$ $v_{10}\cbf_1(1)$ $v_{10}\cbf_1(2) + \sum_{m=1}^M v_{1m}\cbf_{r,m}(1)$ $\cdots$ $\sum_{m=1}^M v_{1m}\cbf_{r,m}(T)$ $\xbf_2(t), 2 \notin \mathcal{B}$ $v_{20}\cbf_2(1)$ $v_{20}\cbf_2(2)$ $\cdots$ $0$ $\vdots$ $\vdots$ $\vdots$ $\vdots$ $\vdots$ $\xbf_L(t), L \in \mathcal{B}$ $v_{L0}\cbf_L(1)$ $v_{L0}\cbf_L(2) + \sum_{m=1}^M v_{Lm}\cbf_{r,m}(1)$ $\cdots$ $ \sum_{m=1}^M v_{Lm}\cbf_{r,m}(T)$ : Superposition Block Markov encoding for Theorem \[thm:subset\][]{data-label="fig:bme.subset"} \[tab:bme\] Note that, since $\Lambda_s$ has normalized second moment equal to unity, and since the dithers are independently and uniformly drawn from $\Vcal_s$, we have with high probability $$\frac{1}{n}{\left\|{\xbf_l(t)}\right\|}^2 \to P\sum_{m=0}^M v_{lm}^2 \leq P.$$ Thus the transmit signals obey the average power constraint. [**Decoding:**]{} Decoding proceeds in three stages. Each transmitter decodes the messages of every other transmitter, the receivers decode the resolution information send cooperatively by the clusters, and finally the receivers decode the vestigial information. Having decoded both components of the desired lattice point, the receiver can recover the desired linear function. At block $t=1$ each transmitter receives the superposition of all the other transmitters’ signals, scaled by channel gains and corrupted by noise: $$\begin{aligned} \zbf_l(1) = \sqrt{P}\sum_{l^\prime \neq l}v_{l^\prime 0}g_{l^\prime l}\cbf_{l^\prime}(1) + \nbf_l(t). \end{aligned}$$ Each transmitter forms estimates $\hat{\wbf}_{l^\prime l}(1)$ for every $l^\prime \neq l$ via typical sequence decoding: if there is a unique collection of messages jointly typical with the received signal, that collection is taken as the estimate; otherwise an error is declared. Note that in this case the transmitters do [*not*]{} employ lattice decoding. For blocks $2 \leq t \leq T$ the situation is similar. Each transmitter receives the superposition of other transmitters’ signals, but in this case the received signals also contain resolution information: $$\zbf_l(t) = \sqrt{P}\left(\sum_{l^\prime \neq l}g_{l^\prime l}v_{l^\prime 0}\cbf_{l^\prime}(t) + \sum_{l^\prime \in \mathcal{B}}\sum_{m=1}^Mg_{l^\prime l}v_{l^\prime m}\cbf_{r,m}(t-1)\right)+ \nbf_l(t).$$ Supposing that each transmitter has successfully decoded the messages from block $t-1$, it knows the resolution information. It therefore can subtract this component out, resulting in the effective signal $$\begin{aligned} \zbf_l^\prime(t) &= \zbf_l(t) - \sqrt{P} \sum_{l^\prime \in \mathcal{B}}\sum_{m=1}^Mg_{l^\prime l}v_{l^\prime m}\cbf_{r,m}(t-1) \\ &= \sqrt{P}\sum_{l^\prime \neq l}g_{l^\prime l}v_{l^\prime 0}\cbf_{l^\prime}(t) + \nbf_l(t) \end{aligned}$$ Now, just as for $t=1$, each transmitter can form estimates $\hat{\wbf}_{l^\prime l}(t)$ of the other transmitters’ messages via typical sequence decoding. Next we turn to the receivers. To decode the function $\fbf_m(t)$, each receiver first decodes the resolution information from the signal received in block $t+1$: $$\begin{gathered} \ybf_m(t+1) = \sqrt{P}\sum_{l=1}^L h_{lm} v_{l0} \cbf_l(t+1) + \sqrt{P}\sum_{m^\prime \neq m}^M\sum_{l \in \mathcal{B}} h_{lm}v_{lm^\prime} \cbf_{l,m^\prime}(t) + \\ \sqrt{P}\sum_{l \in \mathcal{B}} h_{lm}v_{lm}\cbf_{r,m}(t) + \nbf_m(t+1). \end{gathered}$$ Each receiver decodes the resolution information treating the interference—in this case the fresh information from each transmitter and the resolution information intended for other receivers—as noise. Each estimate $\hat{\lambda}_{r,m}(t)$ is formed via lattice decoding as outlined in [@erez:IT04]. The receivers first apply MMSE scaling to the incoming signal and subtract off the dither. Let $$\label{eqn:resolution.noise} \nbf^\prime_m(t+1) = \sqrt{P}\sum_{l=1}^L h_{lm} v_{l0} \cbf_l(t+1) + \sqrt{P}\sum_{m^\prime \neq m}^M\sum_{l \in \mathcal{B}} h_{lm}v_{lm^\prime} \cbf_{l,m^\prime}(t) + \nbf_m(t+1)$$ be the sum of the interference and noise at receiver $m$. Then the scaled signal is $$\begin{aligned} \ybf^\prime_m(t+1) &= \left[ \gamma_m(t+1)\ybf_m(t+1) - \sbf_m(t+1) \right] \mod \Lambda_s \\ &= \left[ \gamma_m(t+1)\sqrt{P}\sum_{l \in \mathcal{B}} h_{lm}v_{lm}\cbf_{r,m}(t) + \gamma \nbf^\prime_m(t+1) - \sbf_m(t+1) \right] \mod \Lambda_s \\ &= \left[ \lambda_{r,m}(t+1) + \left(\gamma_m(t+1)\sqrt{P}\sum_{l \in \mathcal{B}} h_{lm}v_{lm} - 1\right)\cbf_{r,m}(t) + \gamma_m(t+1) \nbf^\prime_m(t+1) \right] \mod \Lambda_s \\ &= \left[ \lambda_{r,m}(t+1) + \nbf_m^{\prime\prime}(t+1) \right] \mod \Lambda_s, \end{aligned}$$ where $$\nbf_m^{\prime\prime}(t+1) = \left(\gamma_m(t+1)\sqrt{P}\sum_{l \in \mathcal{B}} h_{lm}v_{lm} - 1\right)\cbf_{r,m}(t) + \gamma_m(t+1) \nbf^\prime_m(t+1)$$ is the effective noise, including thermal noise, interference, and self-noise associated with MMSE scaling. Then, the estimate is formed by lattice quantization: $$\hat{\lambda}_{r,m}(t) = Q_{\Lambda_r}(\ybf^\prime_m(t+1)).$$ After decoding the resolution information, each receiver turns to $\ybf_m(t)$ to decode the vestigial component $\lambda_{v,m}(t)$. First, we note that, supposing that each receiver has successfully decoded the resolution information from the previous block, it can subtract that portion of the interference, yielding: $$\begin{aligned} \ybf_m^\prime(t) &= \ybf_m(t) - \sqrt{P}\sum_{l \in \mathcal{B}} h_{lm}v_{lm}\cbf_{r,m}(t-1) \\ &= \sqrt{P}\sum_{l=1}^L h_{lm} v_{l0} \cbf_{l}(t) + \sqrt{P}\sum_{m^\prime \neq m}\sum_{l \in \mathcal{B}} h_{lm}v_{lm^\prime} \cbf_{r,m^\prime}(t-1) + \nbf_m(t). \end{aligned}$$ Furthermore, supposing that the resolution information was decoded successfully, each receiver can subtract $\lambda_{r,m}(t)$ from the received signal modulo the shaping lattice. Finally, in preparation for lattice decoding, we apply MMSE scaling to the signal and subtract the dithers as in [@erez:IT04; @nazer:IT11]. Let $$\label{eqn:vestigial.noise} \nbf_m^\prime(t) = \sqrt{P}\sum_{m^\prime \neq m}\sum_{l \in \mathcal{B}} h_{lm}v_{lm^\prime} \cbf_{r,m^\prime}(t-1) + \nbf_m(t)$$ be the sum of the interference and noise in $\ybf_m(t)$. The resulting signal is then $$\begin{aligned} \ybf^{\prime\prime}_m(t) &= \left[\alpha_m(t) \ybf^{\prime}_m(t) - \lambda_{v,m}(t) - \sum_{l=1}^L a_{lm}\tbf_l(t) \right] \mod \Lambda_s \\ &= \left[ \sum_{l=1}^L (\alpha_m(t) \sqrt{P}h_{lm}v_{l0 }\cbf_l(t) - a_{lm}\tbf_l(t)) - \lambda_{r,m}(t) + \alpha_m(t)\nbf^\prime_m(t) \right] \mod \Lambda_s \\ &= \left[ \sum_{l=1}^L a_{lm}(\cbf_l(t) - \tbf_l(t)) - \lambda_{r,m}(t) + \sum_{l=1}^L (\alpha_m(t)\sqrt{P}h_{lm}v_{lm} - a_{lm})\cbf_l(t) + \alpha_m(t)\nbf^\prime_m(t) \right] \mod \Lambda_s \\ &= \left[ \lambda_{m}(t) - \lambda_{r,m}(t) + \sum_{l=1}^L (\alpha_m(t)\sqrt{P}h_{lm}v_{lm} - a_{lm})\cbf_l(t) + \alpha_m(t)\nbf^\prime_m(t) \right] \mod \Lambda_s \\ &= \left[ \lambda_{v,m}(t) + \sum_{l=1}^L (\alpha_m(t)\sqrt{P}h_{lm}v_{lm} - a_{lm})\cbf_l(t) + \alpha_m(t)\nbf^\prime_m(t) \right] \mod \Lambda_s \\ &= \left[ \lambda_{v,m}(t) + \nbf^{\prime\prime}_m(t) \right] \mod \Lambda_s, \label{eqn:vestigial.signal} \end{aligned}$$ where $$\nbf^{\prime\prime}(t) = \sum_{l=1}^L (\alpha_m(t)\sqrt{P}h_{lm}v_{lm} - a_{lm})\cbf_l(t) + \alpha_m(t)\nbf^\prime_m(t)$$ is the effective noise, including thermal noise, interference from other transmitters and clusters, and self-noise associated with MMSE scaling. Each receiver decodes the estimate $\hat{\lambda}_{v,m}(t)$ by quantizing to the nearest point in $\Lambda_v$: $$\hat{\lambda}_{v,m}(t) = Q_{\Lambda_v}(\ybf^{\prime\prime}_{m}(t)).$$ Finally, having recovered both the resolution and vestigial components, each receiver constructs its estimate of the desired lattice codeword, from which it can recover the desired finite-field message: $$\hat{\fbf}_m(t) = \phi^{-1}(\hat{\lambda}_{m}(t)) = \phi^{-1}\left(\left[\hat{\lambda}_{r,m}(t) + \hat{\lambda}_{v,m}(t) \right] \mod \Lambda_s\right).$$ [**Probability of error:**]{} An error occurs when (a) any of the transmitters in $\mathcal{B}$ fails to decode the other transmitters’ messages, (b) any of the receivers fails to decode correctly the incoming resolution information, or (c) when any of the receivers fails to decode correctly the vestigial information associated with the desired lattice point. By the union bound, the probability of error follows $$\begin{aligned} P_e &\leq \sum_{t=1}^T\sum_{m=1}^M\mathrm{Pr}\{\hat{\fbf}_m(t) \neq \fbf_m(t)\} \\ &\leq \sum_{t=1}^T\sum_{l \in \mathcal{B}}\sum_{l^\prime \neq l}\mathrm{Pr}\{\hat{\wbf}_{l^\prime l}(t) \neq \wbf_{l^\prime}(t) \} + \sum_{t=1}^T\sum_{m=1}^M\mathrm{Pr}\{\hat{\lambda}_{r,m}(t) \neq \lambda_{r,m}(t)\} + \sum_{t=1}^T\sum_{m=1}^M \mathrm{Pr}\{\hat{\lambda}_{v,m}(t) \neq \lambda_{v,m}(t) \}. \label{eqn:error.probability} \end{aligned}$$ Here we show that as long as the rates satisfy (\[eqn:subset\]), each error term in (\[eqn:error.probability\]) goes to zero exponentially. We start with the first summation. Each transmitter decodes the messages within its cluster via typical sequence decoding while treating all out-of-cluster interference as noise. By Lemma \[lem:lattice.to.gaussian\] the joint mutual information between the transmit codewords $\cbf_l(t)$ and the receive signal $\zbf^\prime(t)$ approaches that of a Gaussian multiple-access channel with channel coefficients $g_{l^\prime l}v_{l^\prime 0}$, transmit power $P$, and unit noise power. Therefore, so long as $$R < \min_{l \in \mathcal{B} }C_{\mathrm{mac}}(\mathbf{g}_l[1:l-1,l+1:L]\circ\mathbf{v}_0[1:l-1,l+1:L], P, 1),$$ then $\mathrm{Pr}\{\hat{\wbf}_{l^\prime l}(t) \neq \wbf_{l^\prime}(t) \} \to 0$ exponentially for each $l$ and $l^\prime \neq l$. Next we turn to the resolution information. Here each receiver decodes $\lambda_{r,m}(t)$ via lattice decoding on $\ybf_m(t+1)$. In [@erez:IT04] it is shown that lattice decoding is sufficient to achieve the capacity of the Gaussian channel. From (\[eqn:resolution.noise\]) we compute that the the interference power in $\nbf^\prime_m(t+1)$ is $$I_{m,r} = \frac{1}{n}E[{\left\|{\nbf^\prime_m(t+1)}\right\|}^2] = P\left({\left\|{ \hbf_m \circ \vbf_0 }\right\|}^2 + \sum_{m^\prime \neq m,0} |\hbf^T_m \vbf_m^\prime |^2\right).$$ Similarly, we compute that the power of the resolution information in $\ybf_m(t+1)$ is $P|\hbf^T\vbf_m|^2$. Putting these together, we find that if $$R_r < \min_{1 \leq m \leq M} \frac{1}{2} \log_2\left(1+ \frac{P|\hbf_m^T \vbf_m|^2}{1 + P({\left\|{ \hbf_m \circ \vbf_0 }\right\|}^2 + \sum_{m^\prime \neq m} |\hbf^T_m \vbf_m^\prime |^2)}\right),$$ then $\mathrm{Pr}\{\hat{\lambda}_{r,m}(t) \neq \lambda_{r,m}(t)\} \to 0$ exponentially for each $m$. Finally, we look at the vestigial information. Here each receiver decodes $\lambda_{v,m}(t)$ by lattice decoding the sum of multiple incoming lattice points, so we borrow the main result from [@nazer:IT11]. We compute the interference power in (\[eqn:vestigial.noise\]) to be $$I_{m,v} = P\sum_{m^\prime \neq m} |\hbf_m^T\vbf_{m^\prime}|^2,$$ and the effective channel gains in (\[eqn:vestigial.signal\]) are $\hbf_m \circ \vbf_0$. Applying these to the rate in (\[eqn:non.cooperative.rate\]), we find that if $$R_v < \left [\frac{1}{2} \log_2({\left\|{ P \hbf_m \circ \vbf_0}\right\|}^2 + I_{m,v}) - \frac{1}{2}\log_2\left({\left\|{\mathbf{a}_m}\right\|}^2(1+I_{m,v}) + P\left({\left\|{\mathbf{a}_m}\right\|}^2{\left\|{\mathbf{h}_m \circ \mathbf{v}_0}\right\|}^2 - |\mathbf{a}_m^T(\mathbf{h}_m \circ \mathbf{v}_0)|^2\right)\right)\right]^+,$$ then $\mathrm{Pr}\{\hat{\lambda}_{v,m}(t) \neq \lambda_{v,m}(t) \} \to 0$ exponentially. Recall that $R_c = R_r + R_v$ and $R = \frac{TR_c}{T+1}$. Choosing $T$ arbitrarily large, we obtain the desired result. Establishing the Diversity-multiplexing Tradeoff {#sect:DMT} ================================================ In this section we detail the signaling strategies that establish our diversity-multiplexing claims. We begin with the non-cooperative case. Our approach is to choose the steering vector $\mathbf{v}_0$ such that the effective channel vector is constant and has unit gain. It turns out that this approach fails approximately as often as does a single SISO link. First we prove the upper bound. For the non-cooperative case, it is shown in [@nazer:IT11 Theorem 13] that the computation capacity is upper-bounded by $$\begin{aligned} C(\mathbf{H},P) &\leq \max_{\mathbf{A} \in \mathcal{A}} \min_{\substack{l,m \\ a_{lm} \neq 0}} \frac{1}{2} \log_2(1+Ph_{lm}^2) \\ &\leq \frac{1}{2} \log_2(1 + Ph_{lm}^2), \end{aligned}$$ where $l$ and $m$ can be chosen arbitrarily such that $a_{lm} \neq 0$. Then the computation capacity is bounded by the Shannon capacity of a single SISO link, which is proven in [@zheng:IT03] to have diversity-multiplexing tradeoff $d^*(r) = 1-r$. The compute-and-forward network therefore has DMT bounded by $$d^*(r) \leq d^+_{\mathrm{nc}}(r) = 1-r.$$ To prove the lower bound for $M=1$, we invoke the non-cooperative rate of Corollary \[cor:non.cooperative\], choosing $\vbf_0$ to align with the channels. For multiplexing gain $r$, we choose $\mathbf{a} = \mathbf{1}$ and $v_l^2 = P^{r-1}/h_l^2$, resulting in the achievable rate $$\begin{aligned} R(\mathbf{H},P) &= \frac{1}{2}\log_2(1+LP^r) - \frac{1}{2}\log_2(L) \\ &= \frac{1}{2}\log_2\left(\frac{1+LP^r}{L}\right) \\ &\geq \frac{1}{2}\log_2(P^r). \end{aligned}$$ Outage occurs only when we cannot set $v_l$ to the specified value. Since we are constrained to have $v_l^2 \leq 1$, this occurs when $h_l^2 \leq P^{r-1}$. The probability of outage is therefore $$\begin{aligned} P_o \leq \mathrm{Pr}\left\{ \bigcup_{l=1}^L h_l \leq P^{r-1} \right\} \leq \sum_{l=1}^L \mathrm{Pr} \left\{ h_l^2 \leq P^{r-1} \right\} \approx LP^{r-1}. \end{aligned}$$ Therefore, our scheme gives a diversity order at multiplexing gain $r$ of $$\begin{aligned} d^-_{\mathrm{nc}}(r) &= \lim_{P \to \infty} -\frac{\log(P_o)}{\log(P)} \\ &\geq \lim_{P \to \infty} \frac{(1-r)\log(P) - \log(L)}{\log(P)} \\ &= 1-r. \end{aligned}$$ Since this matches the upper bound, the DMT is established. Next we prove the results for the cooperative network. First we prove the upper bound. We invoke the MISO outer bound on the computation capacity from Theorem \[thm:miso.upper.bound\]. The symmetric-rate capacity of the MISO broadcast channel is trivially upper bounded by the capacity of the single-user MISO link between the source and any destination. Thus the DMT is upper-bounded by that of a single $L$-antenna MISO link, which is shown in [@zheng:IT03] to be $d^*(r) = L(1-r)$. Thus the DMT of the cooperative compute-and-forward network is bounded by $$d^*(r) \leq d^+(r) = L(1-r).$$ Next we prove Theorem \[thm:cooperative.dmt\], for which we need first to establish an achievable rate using random codes and time-sharing. \[lem:random.coding\] Let $\mathcal{B} \subset \{1, \cdots, L\}$. In the cooperative compute-and-forward network with $M=1$ receiver, the following computation rate is achievable: $$\label{eqn:random.coding} R_{\mathrm{random}}(\mathbf{H},\mathbf{G},P) = \min\left\{\min_{l \in \mathcal{B}} \frac{1}{2}C_{\mathrm{mac}}(\mathbf{g}_l[1:l-1,l+1:L],P,1), \frac{1}{4}\log_2(1+P(\hbf_{\mathcal{B}}^T\mathbf{1})^2) \right\}.$$ The encoding scheme is simple, so we only sketch the proof. Divide the transmission into two equal time blocks. At the first block, each transmitter encodes and broadcasts its message using a random Gaussian codebook of power $P$. The transmitters in $\mathcal{B}$ decode the incoming messages using typical sequence decoding. This is nothing more than a Gaussian multiple-access channel, so decoding is successful as long as the rate is below the first term in (\[eqn:random.coding\]). The multiple-access rate is cut in half due to time sharing. At the second block, the transmitters in $\mathcal{B}$ directly encode and broadcast the linear combination desired at the receiver, again using a random Gaussian codebook of power $P$.. The receiver decodes the desired function from the signal received in the second block only. This is equivalent to a MISO channel with equal beamformer weights, so decoding is successful as long as the rate is below the second term in (\[eqn:random.coding\]). Again the MISO rate is cut in half due to time sharing. Now we are in a position to prove Theorem \[thm:random.dmt\]. We construct an achievable scheme based on the strategy from Lemma \[lem:random.coding\]. We allow $\mathcal{B}$ to vary according to the channel realizations, giving us the achievable rate $$\begin{aligned} R(\mathbf{H},\mathbf{G},P) &= \max_{\mathcal{B}} \min\left\{\min_{l \in \mathcal{B}} \frac{1}{2}C_{\mathrm{mac}}(\mathbf{g}_l[1:l-1,l+1:L],P,1), \frac{1}{4}\log_2(1+P(\hbf_{\mathcal{B}}^T\mathbf{1})^2) \right\} \\ &\geq \max_{\mathcal{B}} \min\left\{\min_{l \in \mathcal{B}} \frac{1}{2}C_{\mathrm{mac}}(\mathbf{g}_l[1:l-1,l+1:L],P,1), \frac{1}{4}\log_2(1+P({\left\|{\hbf_{\mathcal{B}}}\right\|}^2) \right\}. \label{eqn:random.bound} \end{aligned}$$ Let each rate term in (\[eqn:random.bound\]) be denoted by $R_{\mathcal{B}}(\mathbf{H},\mathbf{G},P)$. Then define the event in which a particular cooperation modality fails: $$\mathcal{O}_{\mathcal{B}} = \left\{ R_{\mathcal{B}}(\Hbf,\Gbf,P) < \frac{r}{2}\log(P)\right\}.$$ Outage occurs when each cooperation modality fails simultaneously: $$\begin{aligned} \mathcal{O} &= \bigcap_{\mathcal{B}}\mathcal{O}_{\mathcal{B}} \\ &\subset \bigcap_{l=1}^L \mathcal{O}_{\{l\}}. \end{aligned}$$ That is, we consider only the events in which a single transmitter decodes the messages. Each term in (\[eqn:random.bound\]) has two components, the failure of either of which results in the failure of the cooperation modality. Therefore, define two events: $\mathcal{C}_{l}$, the event that transmitter $l$ fails to decode the other transmitters’ messages, and $\mathcal{N}_l$, the event that, even if transmitter $l$ decodes successfully, the receiver fails to decode the linear function. The first event can be expressed as $$\begin{aligned} \mathcal{C}_l &= \left\{ \frac{1}{2}C_{\mathrm{mac}}(\mathbf{g}_{l},P,1) < \frac{r}{2}\log(P) \right\} \\ &= \bigcup_{\mathcal{L} \subset \{1,\cdots,L\} \setminus \{ l \}} \left\{ \frac{1}{4|\mathcal{L}|}\log_2\left(1+ P\sum_{l^\prime \in \mathcal{L}}g_{l^\prime l}^2\right) < \frac{r}{2}\log(P) \right\} \\ &\approx \bigcup_{\mathcal{L} \subset \{1,\cdots,L\} \setminus \{ l \}} \left\{ \sum_{l^\prime \in \mathcal{L}}g_{l^\prime l}^2 < P^{2|\mathcal{L}|r - 1} \right\} \\ &\subset \bigcup_{\mathcal{L} \subset \{1,\cdots,L\} \setminus \{ l \}} \left\{ \bigcap_{l^\prime \in \mathcal{L}} \left\{ g_{l^\prime l}^2 < P^{2|\mathcal{L}|r - 1} \right\}\right\}. \end{aligned}$$ The second event can be expressed as $$\begin{aligned} \mathcal{N}_l &= \left\{ \frac{1}{4}\log_2\left(1+P h_l^2\right) < \frac{r}{2}\log_2(P) \right\} \\ &\approx \left\{h_l^2 < P^{2r-1} \right\}. \end{aligned}$$ Since each cooperation modality involves a different set of channel coefficients, the failure events $\mathcal{O}_l$ are independent. Therefore we can bound the outage probability by $$\begin{aligned} \mathrm{Pr}(\mathcal{O}) &\leq \prod_{l=1}^L \mathrm{Pr}(\mathcal{C}_l \cup \mathcal{N}_l) \\ &\lesssim \prod_{l=1}^L \left( \sum_{\mathcal{L}}\prod_{l^\prime \in \mathcal{L}}\mathrm{Pr}(g_{l^\prime l}^2 < P^{2|\mathcal{L}|r-1}) + \mathrm{Pr}( h_l^2 < P^{2r-1} ) \right) \\ &\approx \left(\sum_{|\mathcal{L}|=1}^{L-1} (P^{2|\mathcal{L}|r-1})^{|\mathcal{L}|} + P^{2r-1} \right)^L \\ &\approx \left(\max_{1 \leq |\mathcal{L}| \leq L-1} P^{|\mathcal{L}|(2|\mathcal{L}|r-1)} + P^{2r-1}\right)^L. \label{eqn:random.quadratics} \end{aligned}$$ To find the terms with the largest error exponent, we need to find the value of $|\mathcal{L}|$ that maximizes the quadratics in (\[eqn:random.quadratics\]). For instance, it is clear that for $r=0$ the maximizer is $|\mathcal{L}|=1$. In general, since the quadratics in question are positive, the maximizer is either $|\mathcal{L}|=1$ or $|\mathcal{L}|=L-1$. This gives us $$\begin{aligned} \mathrm{Pr}(\mathcal{O}) &\lesssim \left(\max\left\{ P^{(L-1)(2(L-1)r-1)},P^{2r-1} \right\} + P^{2r-1}\right)^L \\ &\approx \left(\max\left\{ P^{(L-1)(2(L-1)r-1)},P^{2r-1} \right\}\right)^L. \label{eqn:final.random.probabilities} \end{aligned}$$ Finally, plugging (\[eqn:final.random.probabilities\]) into the definition of the DMT, we get $$\begin{aligned} d^*(r) &= \lim_{P \to \infty} \frac{\log(\mathrm{Pr}(\mathcal{O}))}{\log(P)} \\ &\geq L \min\{1-2r, (L-1)(1-2(L-1)r)\}. \end{aligned}$$ Next we prove the lower bound achieved by cooperative lattice coding. Tuning $\mathbf{v}_0$ to the channel vector $\mathbf{h}$ makes it too difficult for transmitters to decode others’ messages, so we take $\mathbf{v}_0 = P^{-x}$, where $x$ depends on the multiplexing gain. Increasing $\mathbf{v}_0$ decreases the error probability at the transmitters, but it increases the Cauchy-Schwarz penalty and therefore the error probability at the receiver. For $r=0$, choosing $x=1/2$ is optimal, which gives us a diversity gain of approximately $1/2$ for each transmitter. The proof follows a similar outline to that of Theorem \[thm:random.dmt\], except that we use the rates proved in Theorem \[thm:subset\] using lattice codes. Again we allow the subset of cooperating users $\mathcal{B}$ to vary according to the channel realizations, and we choose $\mathbf{a} = \mathbf{1}$, resulting in the following achievable rate $$\begin{gathered} \label{eqn:DMT.rates} R(\mathbf{H},\mathbf{G},P) = \max_{\mathcal{B}} \min\bigg\{ \min_{l \in \mathcal{B}}C_{\mathrm{mac}}(\mathbf{g}_l[1:l-1,l+1:L] \circ \mathbf{v}_0[1:l-1,l+1:L], P, 1),\\ \bigg[\frac{1}{2}\log_2(1+P({\left\|{\mathbf{h}\circ \mathbf{v}_0}\right\|}^2 + |\mathbf{h}^T\mathbf{v}_1|^2)) - \frac{1}{2}\log_2\left(L + P\left(L{\left\|{\mathbf{h} \circ \mathbf{v}_0}\right\|}^2 - |\mathbf{1}^T(\mathbf{h} \circ \mathbf{v}_0)|^2\right)\right)\bigg]^+ \bigg\}. \end{gathered}$$ Similar to before, we let each term in (\[eqn:DMT.rates\]) be denoted by $R_{\mathcal{B}}(\mathbf{H},\mathbf{G},P)$ and define the events corresponding to the failure of each cooperation modality: $$\mathcal{O}_{\mathcal{B}} = \left\{ R_{\mathcal{B}}(\Hbf,\Gbf,P) < \frac{r}{2}\log(P)\right\}.$$ Outage occurs when each cooperation modality fails simultaneously: $$\begin{aligned} \mathcal{O} &= \bigcap_{\mathcal{B}} \mathcal{O}_{\mathcal{B}} \\ &\subset \mathcal{O}_{\{1,\dots,L \}} \cap \bigcap_{l=1}^L \mathcal{O}_{\{l\}} \cap \mathcal{O}_{\emptyset}. \label{eqn:error.events} \end{aligned}$$ Here we consider the events in which [*all*]{} transmitters cooperate, in which $L-2$ individual transmitters cooperate, and in which no one cooperates. When $\mathcal{B} = \emptyset$, we use the strategy outlined in the proof of Theorem \[thm:non.cooperative.dmt\], choosing $v_l^2 = P^{r-1}/h_l^2$. Following that line of analysis, the non-cooperative modality fails only when every channel gain is too low: $$\mathcal{O}_\emptyset \subset \left\{ \bigcup_{l=1}^L h_l < P^{r-1} \right\}.$$ For $\mathcal{B} \neq \emptyset$, we choose $v_{l0} = P^{-x_\mathcal{B}/2}$ for every $l$, and $v_{l 1} = 1-P^{-x_\mathcal{B}/2}$ for every $l \in \mathcal{B}$; otherwise $v_{1l}=0$. Using this, we can bound the rate as follows: $$\begin{gathered} R_{\mathcal{B}}(\mathbf{H},\mathbf{G},P) \geq \min\bigg\{ \min_{l \in \mathcal{B}}C_{\mathrm{mac}}(\mathbf{g}_{l}[1:l-1,l+1:L],P^{1-x_\mathcal{B}},1),\\ \bigg[\frac{1}{2}\log_2\left(1+P\sum_{l \in \mathcal{B}} |h_l|^2\right) - \frac{1}{2}\log_2\left(L + P^{1-x_\mathcal{B}}\left(L{\left\|{\mathbf{h}}\right\|}^2 -{\left\|{\mathbf{h}}\right\|}^2\right)\right)\bigg]^+ \bigg\}. \end{gathered}$$ For large $P$, we get $$\label{eqn:subset.rate} R_{\mathcal{B}}(\mathbf{H},\mathbf{G},P) \geq \min\left\{ \min_{l \in \mathcal{B}}C_{\mathrm{mac}}(\mathbf{g}_l[1:l-1,l+1:L],P^{1-x_\mathcal{B}},1), \frac{1}{2}\log_2\left(\frac{P^{x_{\mathcal{B}}}\sum_{l \in \mathcal{B}} |h_l|^2}{(L-1){\left\|{\mathbf{h}}\right\|}^2}\right) \right\}.$$ As before we define events corresponding to the failure of either term in (\[eqn:subset.rate\]): $\mathcal{C}_\mathcal{B}$, the event that the transmitters in $\mathcal{B}$ fail to decode the other transmitters’ messages, and $\mathcal{N}_\mathcal{B}$, the event that, even if the transmitters decode each other properly, the receiver fails to decode its linear function at the required rate. The first event can be expressed as $$\begin{aligned} \mathcal{C}_\mathcal{B} &= \bigcup_{l \in \mathcal{B}} \left\{ C_{\mathrm{mac}}(\mathbf{g}_l[1:l-1,l+1:L],P^{1-x_\mathcal{B}},1) < \frac{r}{2}\log(P) \right\} \\ &= \bigcup_{l \in \mathcal{B}} \bigcup_{\mathcal{L} \subset \{1,\cdots,L\} \setminus \{ l \}} \left\{ \frac{1}{2|\mathcal{L}|}\log_2\left(1+ P^{1-x_\mathcal{B}}\sum_{l^\prime \in \mathcal{L}}g_{l^\prime l}^2\right) < \frac{r}{2}\log(P) \right\} \\ &\approx \bigcup_{l \in \mathcal{B}} \bigcup_{\mathcal{L} \subset \{1,\cdots,L\} \setminus \{ l \}} \left\{ \sum_{l^\prime \in \mathcal{L}}g_{l^\prime l}^2 < P^{|\mathcal{L}|r + x_\mathcal{B} - 1} \right\} \\ &\subset \bigcup_{l \in \mathcal{B}} \bigcup_{\mathcal{L} \subset \{1,\cdots,L\} \setminus \{ l \}} \left\{\bigcap_{l^\prime \in \mathcal{L}} \left\{ g_{l^\prime l}^2 < P^{|\mathcal{L}|r + x_\mathcal{B} - 1} \right\}\right\}. \end{aligned}$$ For $\mathcal{B} = \{1,\cdots,L\}$, the second event can be expressed as $$\begin{aligned} \mathcal{N}_{\{1,\cdots,L\}} &= \left\{ \frac{1}{2}\log_2\left(\frac{P^{x_{\mathcal{B}}}{\left\|{\hbf}\right\|}^2}{(L-1){\left\|{\mathbf{h}}\right\|}^2}\right) < \frac{r}{2}\log_2(P) \right\} \\ &= \left\{ P^{x_{\{1,\cdots,L\}}} < (L-1)P^r \right\}. \label{eqn:empty.set} \end{aligned}$$ Based on (\[eqn:empty.set\]), we choose $x_{\{1,\cdots,L\}} = r + \epsilon$ for any $\epsilon>0$. As $P \to \infty$, this forces $\mathcal{N}_{\{1,\cdots,L\}} \to \emptyset$ deterministically. For $\mathcal{B} = \{l\}$, we can express the second event as $$\begin{aligned} \mathcal{N}_{\{ l \}} &= \left\{ \frac{1}{2}\log_2\left(\frac{P^{x_{\{l\}}} h_l^2}{(L-1){\left\|{\mathbf{h}}\right\|}^2}\right) < \frac{r}{2}\log_2(P) \right\} \\ &= \left\{ \frac{h_l^2}{(L-1){\left\|{\mathbf{h}}\right\|}^2} < P^{r-x_{\{l\}}} \right\} \\ &\subset \left\{h_l^2 < P^{r-x_{\{l\}}-\epsilon} \right\} \cup \left\{{\left\|{\mathbf{h}}\right\|}^2 \geq \frac{P^{\epsilon}}{L-1} \right\} \\ &\subset \bigcap_{l \in \mathcal{B}} \left\{h_l^2 < P^{r-x_{\{l\}}-\epsilon} \right\} \cup \left\{{\left\|{\mathbf{h}}\right\|}^2 \geq \frac{P^{\epsilon}}{L-1} \right\}. \end{aligned}$$ Combining the above with (\[eqn:error.events\]), we get $$\begin{gathered} \label{eqn:all.the.events} \mathcal{O} \subset \left[\left( \bigcup_{l \in \{1,\cdots,L \}} \bigcup_{\mathcal{L} \subset \{1,\cdots,L\} \setminus \{ l \}} \bigcap_{l^\prime \in \mathcal{L}} \left\{ g_{l^\prime l}^2 < P^{|\mathcal{L}|r + r+\epsilon - 1} \right\} \right ) \right] \cap \\ \bigcap_{l \in \{1,\cdots,L \}} \left[ \left( \bigcup_{\mathcal{L} \subset \{1,\cdots,L\} \setminus \{ l \}} \bigcap_{l^\prime \in \mathcal{L}} \left\{ g_{l^\prime l}^2 < P^{|\mathcal{L}|r + x_{\{l\}} - 1} \right\} \right) \cup \left(\left\{h_l^2 < P^{r-x_{\{l\}}-\epsilon} \right\} \right)\right] \cap \\ \left\{ \bigcup_{l=1}^L h_l^2 < P^{r-1} \right\} \cup \left\{{\left\|{\mathbf{h}}\right\|}^2 \geq \frac{P^{\epsilon}}{L-1} \right\}. \end{gathered}$$ Equation (\[eqn:all.the.events\]) contains too many terms to enumerate in full. Since we are concerned with asymptotic behavior, we need only look at the term with the highest error exponent. This term contains one channel failure in $\mathcal{C}_{\{1,\cdots,L\}}$, $L-2$ failures in $\mathcal{C}_{\{l\}} \cap \mathcal{N}_{\{l\}}$, and one failure in $\mathcal{N}_\emptyset$. The final error event, in which ${\left\|{\hbf}\right\|}^2$ is too large, has negligible contribution to the error probability. Combining these, we get $$\begin{gathered} \label{eqn:some.probabilities} \mathrm{Pr}(\mathcal{O}) \lesssim \mathrm{Pr} \left( \bigcup_{\mathcal{L} \subset \{1,\cdots,L\} \setminus \{ 1 \}} \bigcap_{l^\prime \in \mathcal{L}} \left\{ g_{l^\prime 1}^2 < P^{|\mathcal{L}|r + r+\epsilon - 1} \right\} \right) \mathrm{Pr}\left( \left\{ h_2^2 < P^{r-1}\right\} \right) \times \\ \prod_{l=3}^L\mathrm{Pr}\left( \bigcup_{\mathcal{L} \subset \{1,\cdots,L\} \setminus \{ l \}} \bigcap_{l^\prime \in \mathcal{L}} \left\{ g_{l^\prime l}^2 < P^{|\mathcal{L}|r + x_{\{l\}} - 1} \right\} \cup \left\{|h_l|^2 < P^{r-x_{\{l\}}-\epsilon} \right\} \right). \end{gathered}$$ Since each term in (\[eqn:some.probabilities\]) is independent, we can evaluate the probabilities separately, yielding $$\begin{aligned} \mathrm{Pr}(\mathcal{O}) &\lesssim \left(\sum_{\Lcal \subset \{1,\cdots,L\} \setminus \{1\}} (P^{|\Lcal|r+x+\epsilon-1})^{|\Lcal|}\right)(P^{r-1})\left(\sum_{\Lcal \subset \{1,\cdots,L\} \setminus \{3\}} (P^{|\Lcal|r - x-1})^{|\Lcal|} + P^{r-x-\epsilon}\right)^{L-2} \\ &= P^{r-1}\left(\sum_{\Lcal \subset \{1,\cdots,L\} \setminus \{1\}} (P^{|\Lcal|(|\Lcal|r+x+\epsilon-1)}\right)\left(\sum_{\Lcal \subset \{1,\cdots,L\} \setminus \{3\}} (P^{|\Lcal|(|\Lcal|r - x-1)}) + P^{r-x-\epsilon}\right)^{L-2}, \label{eqn:more.probabilities} \end{aligned}$$ where we have chosen $x_{\{l\}}=x$ for every $l$. Similar to the proof Theorem \[thm:random.dmt\], the maximizer of the quadratics in (\[eqn:more.probabilities\]) is either $|\Lcal|=1$ or $|\Lcal|=L-1$. This gives us $$\begin{aligned} \label{eqn:final.probabilities} \mathrm{Pr}(\mathcal{O}) &\lesssim P^{r-1}\left(\max\left\{P^{2r-1+\epsilon},P^{(1-L)(1-rL) + (L-1)\epsilon}\right\}\right)\left(\max_x\min\left\{P^{r+x-1},P^{(1-L)(1-(L-1)r-x)},P^{r-x-\epsilon} \right\} \right)^{L-2}. \end{aligned}$$ Finally, plugging (\[eqn:final.probabilities\]) into the definition of the DMT, taking the supremum over all $\epsilon > 0$, and taking the maximum over all $x$ yields $$\begin{aligned} d^-_{\mathrm{c}}(r) &= \lim_{P \to \infty} \frac{\log(\mathrm{Pr}(\mathcal{O}))}{\log(P)} \\ &\geq 1-r + \min\{[1-2r]^+,[(L-1)(1-rL)]^+ \} + \notag \\ &\quad\quad\quad\quad \max_{0 \leq x \leq 1}(L-2) \min\{[1-x-r]^+,[(L-1)(1-(L-1)r-x)]^+,[x-r]^+ \}. \end{aligned}$$ Although their proofs are similar, the strategies Theorems \[thm:random.dmt\] and \[thm:lattice.dmt\] achieve very different diversity-multiplexing tradeoffs. With random coding, transmitters decode and cooperatively transmit at separate blocks; such time division enables full diversity, but it leads to poor multiplexing performance. With lattice coding, on the other hand, we need to balance transmit power in order to ameliorate the effects of signal misalignment; the balance costs us diversity gain, but the multiplexing performance is improved. The overall message is that transmit cooperation improves diversity [*and*]{} multiplexing for compute-and-forward, and as we saw in Figure \[fig:dmt.plot\] the two approaches combined achieve the corner points of the DMT region. Numerical Examples {#sect:simluations} ================== In this section we examine a few example scenarios in which to demonstrate the benefits of our approach. \[ex:inter.channel\] The first example, depicted in Figure \[fig:inter.channel\], comprises $L=2$ transmitters and a single receiver. The channels are symmetric, with the forward coefficients constant $h_1=h_2=1$ and the inter-transmitter coefficients a variable $g_{12}=g_{21}=g$, which we vary such that the gain $g^2$ ranges between $-10$dB and $30$dB. We set the transmit SNR at $P=10$dB. Since the channel gains are symmetric, either both transmitters can decode the other’s message or neither of them can; therefore we choose either $\mathcal{B} = \{1,2\}$ or $\mathcal{B} = \emptyset$ for cooperative computation. Similarly, by symmetry it is easy to see that the optimal choice for the linear function is $\mathbf{a} = (1, 1)^T$ and that the optimal steering vectors $\mathbf{v}_0$ and $\mathbf{v}_1$ are constant. We find the optimal tradeoff between $\mathbf{v}_0$ and $\mathbf{v}_1$ numerically. In Figure \[fig:inter.channel.plot\] we plot the achievable rate of our cooperative scheme against the upper bound of Theorem \[thm:relay.upper.bound\], using the Nazer-Gastpar rate of (\[eqn:non.cooperative.rate\]) as a baseline. The trends are easy to appreciate. When the channels between transmitters are weak, decoding each other’s messages is too difficult, and the cooperative rate collapses to (\[eqn:non.cooperative.rate\]). As the inter-transmitter gains become stronger it becomes easier for the transmitters to decode, and cooperation can improve the computation rate and eventually approaches the upper bound. We note a “dimple” in the cooperative rate as $g^2$ becomes large. For sufficiently large $g^2$, the optimal strategy is to turn the steering vector $\mathbf{v}_0$ down far enough that the Nazer-Gastpar component of the cooperative rate is zero, meaning that only the jointly-encoded resolution information carries information to the receiver. At this value of $g^2$ we see the dimple, after which the rate quickly converges on the upper bound. \[ex:path.loss\] Next we examine a scenario in which channel gains are chosen randomly, as depicted in Figure \[fig:simulation.model\]. We place a single receiver at the origin and place $L=3$ transmitters randomly and uniformly on a segment of the circle having specified arclength. From the geometric configuration of the network, we compute channel magnitudes according to a path-loss model: $$g_{ij} = \sqrt{\frac{1}{d(i,j)^\alpha}}, \quad h_i = \sqrt{\frac{1}{d(i,0)^\alpha}},$$ where $d(i,j)$ is the Euclidean distance between users $i$ and $j$ and. We choose $P=10$dB and a path-loss exponent of $\alpha=4$. For each realization we calculate the cooperative computation rate. Since the gains from transmitters to receiver are equal, $\mathbf{a} = (1,1,1)^T$ is the optimal choice. The steering vectors and the clusters are optimized numerically. We run 500 simulations each for arclengths varying from 0 to $\pi$, and plot the average computation rates in Figure \[fig:rates\]. Again the trends are easy to appreciate. Cooperation offers the greatest improvement when transmitters are close together. Even as we spread transmitters further apart, on average enough transmitters can cooperate that our approach garners a noticeable improvement. \[ex:two.by.one\] In the next example we examine the variation in cooperative computation rate with respect to the channel gain between transmitters and receivers. As depicted in Figure \[fig:two.by.one\], we again have $L=2$ and $M=1$, but now we set every channel gain to unity except for $h_{12}$. Since the channels between transmitters and receiver are not symmetric, we cannot take $\mathbf{a} = (1,1)^T$ or $\mathbf{v}_0$ and $\mathbf{v}_1$ to be constant. Instead, we iterate manually through possible choices of $\mathbf{a}$ and numerically optimize over the set $\mathcal{B}$ of cooperating transmitters and the steering vectors $\mathbf{v}_0$ and $\mathbf{v}_1$. In Figure \[fig:two.by.one.rates\] we plot the cooperative rate alongside (\[eqn:non.cooperative.rate\]) for a variety of transmit signal-to-noise ratios $P$. We make a few observations. First, the non-cooperative rate is low for $h_{21}$ near to zero. Since we require the function to contain elements from both transmitters’ messages, it becomes difficult for the receiver to decode such a function. In the cooperative case, however, the rates do not fall, since transmitter 1 can decode $\wbf_2$ and transmit the desired function to the receiver. This result hints at the diversity gains inherent to the cooperative approach; even when one link fails, successful computation is possible. Furthermore, in the cooperative case we get the full multiplexing gain as the SNR becomes large. In the non-cooperative case we observe “peaks”; these correspond to rational channel gains with low denominator. The further $h_{21}$ is from a low-denominator rational, the harder it is to align the function with the channels and the higher the Cauchy-Schwarz penalty in (\[eqn:non.cooperative.rate\]). However, we can always choose $\vbf_0$ such that the equivalent channel vector is rational, allowing us to completely eliminate the Cauchy-Schwarz penalty. We note that this is not explicitly due to the cooperative nature of our approach; as shown in \[thm:non.cooperative.dmt\] non-cooperative transmitters can get the full multiplexing gain using lattice codes. However, our cooperative approach [*does*]{} permit the transmitters to use the remaining power to secure rate and diversity gains. \[ex:two.by.two\] Finally, we examine the system depicted in Figure \[fig:two.by.two\]. Here we have $L=M=2$, and again we set all channel gains to unity except for $h_{21}$. Again asymmetry prevents us from choosing $\mathbf{a}$ and the steering vectors easily. We iterate manually over the possible choices for $\mathbf{a}$, choose zero-forcing beamformers for $\mathbf{v}_1$ and $\mathbf{v}_2$, and numerically optimize over $\mathbf{v}_0$. In order for zero-forcing to succeed, we choose $\mathcal{B} = \{1,2\}$. In Figure \[fig:two.by.two.rates\] we plot the cooperative rate alongside (\[eqn:non.cooperative.rate\]), again for a variety of signal-to-noise ratios. Again we make a few observations. In contrast to the previous scenario, here the rate drops when $h_{12} \approx 1$; this is because the channel matrix becomes increasingly ill-conditioned. Similar to before, in the cooperative case the rate remains non-zero, but here it occurs because the transmitters can cooperatively send a full-rank set of equations even though the channel matrix is nearly singular. However, in this example cooperation does not obtain the full multiplexing gain. The freedom to choose $\vbf_0$ allows us to mitigate the peakiness of the achievable rate, but we cannot eliminate the Cauchy-Schwarz penalty at both receivers simultaneously. Even for high SNR, however, we [*do*]{} get considerable robustness to channel variation. Conclusion {#sect:conclusion} ========== We have studied the impact of user cooperation on compute-and-forward. Constructing a lattice-coding version of block Markov encoding, we presented a strategy that introduces a “decode-and-forward” element into computation coding. Transmitters decode each other’s messages, enabling them to transmit resolution information cooperatively to the receivers. Our strategy achieves higher computation rates than previous approaches, since transmitters can jointly encode part of their messages, and coherent signals benefit from a beamforming gain. Additionally, cooperation enables an improvement in the diversity-multiplexing tradeoff, achieving full diversity when there is a single receiver. In the case of multiple receivers, however, we have not established an achievable diversity-multiplexing tradeoff. The difficulty of aligning lattice codewords at multiple receivers suggests that lattice coding is insufficient for the task. A promising approach may be to introduce a cooperative element into the signal-alignment strategy of [@niesen:ISIT11]. Since this approach achieves the full multiplexing gain for multiple receivers, we expect to be able to obtain a non-trivial characterization of the DMT regardless of the number of receivers. Finally, we note that our techniques can be applied to any situation in which one needs to merge lattice codes with decode-and-forward style cooperation. Our block Markov approach is rather general; as mentioned earlier, it can be used to achieve the capacity of the physically degraded relay channel or to achieve the decode-and-forward rates of the cooperative multiple-access channel. We therefore expect our techniques to be useful for developing new strategies and establishing new results in areas where lattice codes and cooperation are applied, such as physical-layer security [@lai:IT08; @dong:SP10; @ekrem:IT11; @yuskel:IT11; @huang:SP11] and interference channels [@sahin:globecom07; @sridharan:ISIT08; @rini:ITW10]. Proofs of Upper Bounds {#app:upper.bounds} ====================== Our first task is to prove Theorem \[thm:miso.upper.bound\], for which we need a quick lemma. \[lem:uniform.functions\] Let $\wbf_1, \cdots, \wbf_L \in \mathbb{F}_p^k$ be independently and uniformly distributed messages. Then, the functions $\fbf_1, \cdots, \fbf_M$ are also independent and uniformly distributed across $\mathbb{F}_p^k$. Since the finite-field linear combinations in $\fbf_l$ are taken element-wise, it is sufficient to show the result for an arbitrary element of both messages and functions. Therefore, let $\wbf = (w_{11},\cdots,w_{L1})^T$ and $\fbf = (f_{11},\cdots,f_{M1})^T = \mathbf{A}\wbf$. We need to show that the elements of $\fbf$ are independent and uniformly distributed. Since $\wbf$ is uniformly distributed over $\mathbb{F}_p^L$, its probability mass function is $$p(\wbf) = p^{-L}.$$ The conditional pmf of $\fbf$ is $$p(\fbf | \wbf) = \delta(\fbf - \mathbf{A}\wbf),$$ where $\delta(\cdot)$ is the Kronecker delta function. Next we compute the marginal pmf for $\fbf$: $$\begin{aligned} p(\fbf) &= \sum_{\wbf \in \mathbb{F}_p^L}p(\fbf | \wbf)p(\wbf) \\ &= p^{-L}\sum_{\wbf \in \mathbb{F}_p^L}\delta(\fbf - \mathbf{A}\wbf) \\ &= p^{-L} \big | \left\{ \wbf | \mathbf{A}\wbf = \fbf\right\} \big | \label{eqn:full.rank} \\ &= p^{-L}p^{L-M} = p^{-M}, \end{aligned}$$ where (\[eqn:full.rank\]) follows because $\mathbf{A}$ is full rank. Since the pmf $p(\fbf)$ does not depend on $\fbf$, the elements are independent and uniformly distributed. With Lemma \[lem:uniform.functions\], it is straightforward to prove Theorem \[thm:miso.upper.bound\]. Suppose that a genie provides the messages $\wbf_l(t)$ to each of the transmitters. Then the transmitters each can compute the functions $\fbf_m(t)$. By Lemma \[lem:uniform.functions\] these functions are independent and uniformly distributed, the scenario is equivalent to an $L$-transmitter antenna having $M$ independent messages to send to $M$ users. In [@weingarten:IT06] the capacity region is shown to be (\[eqn:dirty.paper\]). Since we define the computation capacity in terms of achievable [*symmetric*]{} rate, it cannot exceed the symmetric-rate MISO capacity given in (\[eqn:common.miso.rate\]). Next we prove the upper bound in Theorem \[thm:relay.upper.bound\]. Choose a transmitter $l$, and suppose that a genie supplies the messages $w_{l^\prime}(t)$ to the receivers for every $l^\prime \neq l$. By the crypto lemma [@erez:IT04], each $\fbf_{m}(t)$ such that $a_{lm} \neq 0$ is statistically independent of the messages $w_{l^\prime}(t)$, so the receivers remain equivocal as to the desired functions. Thus the scenario is equivalent to a compound relay channel, with transmitter $l$ acting as the source, the transmitters $l^\prime$ acting as relays, and each receiver $m$ such that $a_{lm} \neq 0$ acting as destinations all needing the messages $w_{l}(t)$. The capacity of the compound relay channel can be bounded using cut-set arguments. For any cut $S \in \mathcal{S}_l$, the capacity of the compound relay channel, and thus the computation capacity of the cooperative compute-and-forward network, is bounded by $$\begin{aligned} C(\Hbf,\Gbf,P) &\leq \max_{p(\xbf)} \min_{m, a_{lm} \neq 0} I(x_l,x_S; y_m, z_{S^C} | x_{S^C}) \\ &\leq \min_{m, a_{lm} \neq 0} \max_{p(\xbf)} I(x_l,x_S; y_m, z_{S^C} | x_{S^C}). \end{aligned}$$ Taking the minimum over all transmitters and all cuts $S$, we obtain the result. Entropy of dithered lattices over the multiple-access channel ============================================================= Here we prove that the mutual information between dithered lattice codewords and any receiver approaches that of a Gaussian multiple-access channel. \[lem:lattice.to.gaussian\] Let $$\xbf_l = \sqrt{P}[\lambda_l + \tbf_l] \mod \Lambda_s$$ be a collection of independent lattice codewords, dithered across the shaping lattice, for $1 \leq l \leq L$. Let $$\ybf = \sum_{l=1}^Lh_l \xbf_l + \nbf,$$ be a noisy sum of the codewords, where the noise $\nbf$ has i.i.d. elements with variance $\sigma^2$. Then, for any set $\mathcal{B} \in \{1,\cdots,L\}$, the normalized mutual information between the transmit signals and the receive signal approaches at least that of a Gaussian multiple-access channel: $$\lim_{n \to \infty}\frac{1}{n} I(\xbf_\mathcal{B} ; \ybf | \xbf_{\mathcal{B}^C}) \geq \frac{1}{2}\log_2\left(1+\frac{P\sum_{l \in \mathcal{B}}h_l^2}{\sigma^2}\right).$$ When $\nbf$ is Gaussian, this bound is tight. Since $\ybf$ is the sum of transmitted signals, conditioning entails only subtracting away the known component. Therefore, letting $$\ybf_\mathcal{B} = \sum_{l \in \mathcal{B}} h_l \xbf_l + \nbf,$$ the mutual information is $$\lim_{n \to \infty}\frac{1}{n} I(\xbf_\mathcal{B} ; \ybf | \xbf_{\mathcal{B}^C}) = \lim_{n \to \infty} \frac{1}{n}I(\xbf_\mathcal{B} ; \ybf_\mathcal{B}) = \lim_{n \to \infty}\frac{1}{n}(h(\ybf_{\mathcal{B}}) - h(\nbf)),$$ where $h(\cdot)$ is the differential entropy. Since the Gaussian distribution maximizes the differential entropy for a given variance, we have $$\frac{1}{n}h(\nbf) \leq \frac{1}{2}\log(2\pi e \sigma^2). \label{eqn:n.entropy.bound}$$ To bound $h(\ybf_\mathcal{B})$, we note that in [@nazer:IT11 Lemma 8] it was shown that the density function $f_{\ybf_\mathcal{B}}$ is bounded by $$f_{\ybf_{\mathcal{B}}} \leq e^{c(n)n}f_{\ybf^*},$$ where $\ybf^*$ is an i.i.d. Gaussian vector with variance $P\sum_{l \in \mathcal{B}}h_l^2 + \sigma^2$, and $c(n)$ is a term approaching zero from above as $n \to \infty$. Plugging this into the definition of differential entropy, we have, for sufficiently high $n$, $$\begin{aligned} \frac{1}{n}h(\ybf_\mathcal{B}) &\geq -\frac{1}{n} \int e^{c(n)n}f_{\ybf^*} \log(e^{c(n)n}f_{\ybf^*}) \\ &= -\frac{1}{n}e^{c(n)n}\int f_{\ybf^*} \log(f_{\ybf^*}) - \frac{1}{n} e^{c(n)n}c(n)n \\ &= e^{c(n)n}\left(\frac{1}{n}h(\ybf^*) - c(n)\right) \\ &\geq \frac{1}{n}h(\ybf^*) - c(n) \label{eqn:constant.is.small}\\ &\to \frac{1}{n}h(\ybf^*) \\ &= \frac{1}{2}\log\left(2\pi e \left(P\sum_{l \in \mathcal{B}} + \sigma^2\right)\right), \label{eqn:y.entropy.bound} \end{aligned}$$ where (\[eqn:constant.is.small\]) follows because $e^{c(n)n} \geq 1$ and for sufficiently high $n$ the term $\frac{1}{n} h(\ybf^*) - c(n)$ is positive. Combining (\[eqn:n.entropy.bound\]) and (\[eqn:y.entropy.bound\]), we get that $$\begin{aligned} \lim_{n \to \infty}\frac{1}{n} I(\xbf_\mathcal{B} ; \ybf | \xbf_{\mathcal{B}^C}) &\geq \frac{1}{2}\log\left(2\pi e \left(P\sum_{l \in \mathcal{B}}h_l^2 + \sigma^2\right)\right) - \frac{1}{2}\log(2\pi e \sigma^2) \\ &= \frac{1}{2}\log_2\left(1+\frac{P\sum_{l \in \mathcal{B}}h_l^2}{\sigma^2}\right). \end{aligned}$$ When $\nbf$ is Gaussian, it is well-known that Gaussian inputs are optimal and result in the same mutual information as the bounds just established. In this case the bound is tight. [^1]: email: {nokleby, aaz}@rice.edu. Matthew Nokleby and Behnaam Aazhang are with Rice University, Houston, TX. Behnaam Aazhang is also the Finnish Academy Distinguished Professor (FiDiPro) at the Center for Wireless Communication, University of Oulu, Finland. This work is supported in part by Renesas Mobile and the National Science Foundation. This work was presented in part at the IEEE Symposium on Information Theory, St. Petersburg, Russia, August 2011, and the IEEE Information Theory Workshop, Paraty, Brazil, October 2011. [^2]: Very precisely, receivers compute any of a [*sequence*]{} of linear combinations since, as we shall see, $k,p \to \infty$ as the codeword length becomes large. [^3]: Other approaches are possible. For example, in an earlier work [@nokleby:ITW11] we partitioned the set of transmitters into clusters; transmitters would decode only in-cluster messages. In the interests of brevity we discuss only the approach presented in Theorem \[thm:subset\]. [^4]: This terminology is intended to convey the fact that this lattice component encodes the “residual” or “leftover” information bits. We use this less-common synonym in order to minimize notational confusion. [^5]: Technically we have a sequence of lattice codebooks indexed by $n$. In the interest of notational simplicity we drop the superscripts. [^6]: For further discussion of the need for dithers, see [@forney:Allerton04]

ArXiv

--- abstract: 'In this note, we investigate the boundary Hölder regularity for fully nonlinear elliptic equations on Reifenberg flat domains. We will prove that for any $0<\alpha<1$, there exists $\delta>0$ such that the solutions are $C^{\alpha}$ at $x_0\in \partial \Omega$ provided that $\Omega$ is $(\delta,R)$-Reifenberg flat at $x_0$ (see ). A similar result for the Poisson equation has been proved by Lemenant and Sire [@Lemenant-Sire], where the Alt-Caffarelli-Friedman’s monotonicity formula is used. Besides the generalization to fully nonlinear elliptic equations, our method is simple. In addition, even for the Poisson equation, our result is stronger than that of Lemenant and Sire.' author: - Yuanyuan Lian - Kai Zhang bibliography: - 'Reifenberg.bib' date: 'Received: date / Accepted: date' title: 'Boundary Hölder Regularity for Fully Nonlinear Elliptic Equations on Reifenberg Flat Domains[^1] ' --- Introduction {#intro} ============ In the regularity theory of partial differential equations, Hölder continuity is a kind of quantitative estimate. It is usually the first smooth regularity for solutions and the beginning for higher regularity. With respect to the interior Hölder regularity, De Giorgi [@MR0093649], Nash [@MR0100158] and Moser [@MR0170091] proved it for elliptic equations in divergence form and many generalizations have been made. In particular, Krylov and Safonov [@MR525227] proved the interior Hölder regularity for elliptic equations in nondivergence form and fully nonlinear elliptic equations can also be treated. As regards the boundary Hölder continuity, it is well known that if $\Omega$ satisfies the exterior cone condition at $x_0\in \partial \Omega $, then the solution is Hölder continuous at $x_0$ (see [@MR0221087] and [@MR1814364 Theorem 8.29 and Corollary 9.28]). Recently, the second author proved the boundary Hölder regularity on more general domains (e.g. the Reifenberg flat domains) for various elliptic equations [@2018arXiv180401299Z]. For the Poisson equations, Lemenant and Sire [@Lemenant-Sire] proved that for any $0<\alpha<1$, there exists $\delta>0$ such that the solutions are $C^{\alpha}$ at $x_0\in \partial \Omega$ provided that $\Omega$ is $(\delta,R)$-Reifenberg flat at $x_0$. In this note, we will prove an analogue result for the viscosity solutions of fully nonlinear elliptic equations and our method is simple. Moreover, for the Poisson equations, our result has relaxed the requirements imposed in [@Lemenant-Sire]. The main idea in the proof is that a Reifenberg flat domain can be regarded as a perturbation of half balls in different scales. The idea has been motivated by [@2018arXiv180401299Z]. \[d-re\] We say that $\Omega$ is ($\delta$,R)-Reifenberg flat from exterior at $x_0\in \partial \Omega$ if for any $0<r<R$, there exists a coordinate system $\{y_1,...,y_n \}$ such that $x_0=0$ in this coordinate system and $$\label{e-re} B_r\cap \Omega \subset B_r \cap\{y_n >-\delta r\}.$$ In this note, we treat the viscosity solutions for fully nonlinear elliptic equations and use the standard definition of a viscosity solution. For details, we refer to [@MR1351007]. Let $\Omega \subset R^{n}$ be a bounded domain and $f$ be a function defined on $\bar{\Omega}$. We say that $f$ is $C^{\alpha}$ ($0<\alpha< 1$) at $x_0\in \bar{\Omega}$ or $f\in C^{\alpha}(x_0)$ if there exists a constant $C$ such that $$\label{holder} |f(x)-f(x_0)|\leq C|x-x_0|^{\alpha},~~\forall~x\in \bar{\Omega}.$$ Then, define $[f]_{C^{\alpha}(x_0)}=C$ and $\|f\|_{C^{\alpha}(x_0)}=\|f\|_{L^{\infty}(\Omega)}+[f]_{C^{\alpha}(x_0)}$. Our main result is the following. \[t-1\] Let $0<\alpha<1$ and $u$ be a viscosity solution of $$\left\{\begin{aligned} &u\in S(\lambda,\Lambda,f)&& ~~\mbox{in}~~\Omega;\\ &u=g&& ~~\mbox{on}~~\partial \Omega, \end{aligned}\right.$$ where $0\in \partial \Omega$, $g\in C^{\alpha}(0)$ and $f$ satisfies $$\label{e1.12} \left (\frac{1}{|B_r\cap\Omega|}\int_{B_r\cap\Omega}|f|^n\right )^{\frac{1}{n}}\leq C_f r^{\alpha - 2},~~\forall~ r>0.$$ Then there exists $\delta$ depending only on $n,\lambda,\Lambda$ and $\alpha$ such that if $\Omega$ is $(\delta,R)$-Reifenberg flat from exterior at $0$ for some $R>0$, $u$ is $C^{\alpha}$ at $0$ and $$\label{e-holder} |u(x)-u(0)|\leq C |x|^{\alpha}\left(\|u\|_{L^{\infty }(\Omega\cap B_{R})}+C_{f}+[g]_{C^{\alpha}(0)}\right), ~~\forall ~x\in \Omega\cap B_{R},$$ where $C$ depends only on $n, \lambda, \Lambda, \alpha$ and $R$. For tge Poisson equation, we have the following corresponding result: \[t-2\] Let $0<\alpha<1$ and $u$ satisfy $$\label{e-poisson} \left\{\begin{aligned} &\Delta u=f&& ~~\mbox{in}~~\Omega;\\ &u=g&& ~~\mbox{on}~~\partial \Omega, \end{aligned}\right.$$ where $0\in \partial \Omega$, $g\in C^{\alpha}(0)$ and $f\in L^{p}(\Omega) $ with $p=n/(2-\alpha)$. Then there exists $\delta$ depending only on $n$ and $\alpha$ such that if $\Omega$ is $(\delta,R)$-Reifenberg flat from exterior at $0$ for some $R>0$, $u$ is $C^{\alpha}$ at $0$ and $$\label{e-poisson-holder} |u(x)-u(0)|\leq C |x|^{\alpha}\left(\|u\|_{L^{\infty }(\Omega\cap B_{R})}+\|f\|_{L^{p}(\Omega\cap B_{R})}+[g]_{C^{\alpha}(0)}\right), ~~\forall ~x\in \Omega\cap B_{R},$$ where $C$ depends only on $n, \alpha, p$ and $R$. \[r-1.1\] is more general by comparing with the result of [@Lemenant-Sire]. Proof of the main result {#sec:1} ======================== Since the proofs of and are similar, wee only give the detailed proof for . The main idea is the following. A Reifenberg flat domain is regarded as a perturbation of half balls in different scales. On the other hand, solutions on half balls have sufficient regularity (in fact the Lipschitz regularity). Hence, the boundary Hölder regularity of the solution on a Reifenberg flat domain can be obtained from solutions on half balls by approximation and iteration. The following is the detail of the proof. **Proof of .** Without loss of generality, we assume that $R=1$ and $g(0)=0$. Let $M=\|u\|_{L^{\infty }(\Omega\cap B_1)}+C_{f}+[g]_{C^{\alpha}(0)}$ and $\Omega_{r}=\Omega\cap B_{r}$. To prove\[e-holder\], we only need to show the following: There exist constants $0<\delta<1$, $0< \eta < 1$ and $\hat{C}$ depending only on $n, \lambda, \Lambda$ and $\alpha$ such that for all $k\geq 0$, $$\label{e-ful-discrete} \|u\|_{L^{\infty }(\Omega _{\eta ^{k}})}\leq \hat{C} M \eta^{k\alpha}.$$ We prove\[e-ful-discrete\] by induction. For $k=0$, it holds clearly. Suppose that it holds for $k$. We need to prove that it holds for $k+1$. Let $r=\eta ^{k}/2$. Then there exists a coordinate system $\{y_1,...,y_n \}$ such that $$\label{e-re-2} B_r\cap \Omega \subset B_r \cap\{y_n >-\delta r\}.$$ Let $\tilde{B}^{+}_{r}=B^{+}_{r}-\delta r e_n $, $\tilde{T}_r=T_r-\delta r e_n$ and $\tilde{\Omega }_{r}=\Omega \cap \tilde{B}^{+}_{r}$ where $e_n=(0,0,...,0,1)$. Take $$\label{e-delta-1} \delta\leq 1/4.$$ Then $\Omega _{r/2}\subset \tilde{\Omega }_{r}\subset \Omega_{\eta^k}$. Let $v$ solve $$\left\{\begin{aligned} &M^{+}(D^2v,\lambda,\Lambda)=0 &&\mbox{in}~~\tilde{B}^{+}_{r}; \\ &v=0 &&\mbox{on}~~\tilde{T}_{r};\\ &v=\hat{C} M \eta^{k\alpha}&&\mbox{on}~~\partial \tilde{B}^{+}_{r}\backslash \tilde{T}_{r}. \end{aligned} \right.$$ Then $w:=u-v$ satisfies (since $v\geq 0$) $$\left\{ \begin{aligned} &w\in \underline{S}(\lambda /n,\Lambda , f) &&\mbox{in}~~ \tilde{\Omega }_{r}; \\ &w = g-v\leq g &&\mbox{on}~~\partial \Omega \cap \tilde{B}^{+}_{r};\\ &w\leq 0 &&\mbox{on}~~\partial \tilde{B}^{+}_{r}\cap \bar{\Omega}. \end{aligned} \right.$$ Take $$\label{e1.19} \delta \leq \eta.$$ Then by the boundary Lipschitz estimate for $v$, $$\label{e-v} \|v\|_{L^{\infty }(\Omega _{2\eta r})}\leq C\frac{(2\eta+\delta) r}{r}\|v\|_{L^{\infty }( \tilde{B}^{+}_{r})}\leq C_1\eta\cdot \hat{C} M \eta^{k\alpha}\leq C_1\eta ^{1-\alpha }\cdot \hat{C}M\eta ^{(k+1)\alpha},$$ where $C_1$ depends only on $n,\lambda$ and $\Lambda$. For $w$, by the A-B-P maximum principle, we have for $x\in \tilde{\Omega }_{r}$, $$\label{e-w} w(x) \leq\|g\|_{L^{\infty }(\partial \Omega \cap \tilde{B}^{+}_{r})}+C_2r\|f\|_{L^n(\Omega_{r})} \leq r^{\alpha}M+C_2Mr^{\alpha}=\frac{1+C_2}{2^\alpha \eta^{\alpha}}\cdot M\eta^{(k+1)\alpha},$$ where $C_2$ depends only on $n,\lambda$ and $\Lambda$. Take $\eta$ small enough such that $$C_1\eta ^{1-\alpha }\leq 1/2.$$ Next, take $\hat{C}$ large enough such that $$\frac{1+C_2}{2^{\alpha}\eta^{\alpha}}\leq \hat{C}/2.$$ Then combining with\[e-v\] and\[e-w\], we have $$\sup_{\Omega_{\eta ^{k+1}}}u\leq \sup_{\Omega_{\eta ^{k+1}}}w+\sup_{\Omega_{\eta ^{k+1}}}v\leq \hat{C}M\eta ^{(k+1)\alpha}.$$ The proof for $$\inf_{\Omega _{\eta^{k+1}}} u\geq -\hat{C}M\eta ^{(k+1)\alpha}$$ is similar and we omit here. Therefore, $$\|u\|_{L^{\infty}(\Omega _{\eta^{k+1}})}\leq \hat{C}M\eta ^{(k+1)\alpha}.$$ By induction, the proof is completed. [^1]: This research is supported by NSFC 11701454.

ArXiv

--- abstract: 'We give the classification of globally generated vector bundles of rank $2$ on a smooth quadric surface with $c_1\le (2,2)$ in terms of the indices of the bundles, and extend the result to arbitrary higher rank case. We also investigate their indecomposability and give the sufficient and necessary condition on numeric data of vector bundles for indecomposability.' address: - 'Università di Trento, 38123 Povo (TN), Italy' - 'Department of Mathematics, Sungkyunkwan University, Suwon 440-746, Korea' - 'Politecnico di Torino, Corso Duca degli Abruzzi 24, 10129 Torino, Italy' author: - 'E. Ballico, S. Huh and F. Malaspina' title: | Globally generated vector bundles\ on a smooth quadric surface --- [^1] Introduction ============ Globally generated vector bundles on projective varieties play an important role in projective algebraic geometry and their classification has been done quite recently over projective spaces by many people [@ACM][@am][@ce][@e][@m][@SU][@SU2]. Especially in [@e], Ellia determines the Chern classes of globally generated vector bundles of rank $2$ on ${{\ensuremath{\mathbb{P}}}}^2$, using several results on Lüroth semigroup of smooth plane curves [@c][@gr]. In [@ce] Chiodera and Ellia classify the globally generated vector bundles of rank $2$ on ${{\ensuremath{\mathbb{P}}}}^n$ with small first Chern classes. We ask similar questions over a smooth quadric surface $Q$ and give answers as in our previous works [@BHM1][@BHM2]. Our main theorem is the following: There exists an indecomposable and globally generated vector bundle of rank $r\ge 2$ on $Q$ with the Chern classes $(c_1, c_2)$ such that $c_1=(a,b)\leq (2,2)$, $a\le b$ if and only if $(c_1,c_2;r)$ is one of the following: $$\begin{aligned} &(1,1,2;r=2,3),\\ &(1,2,2;2),(1,2,3;r=2,3),(1,2,4;r=2,3,4,5),\\ &(2,2,3;2),(2,2,4;r=2,3), (2,2,5;r=2,3),(2,2,6;r=2,3,4,5),\\ &(2,2,8;r=2,3,4,5,6,7,8)\end{aligned}$$ In the second section we fix the notations and we explain the preliminaries. In the third section we show that any globally generated vector bundle with $c_1=(0,0)$ or $c_1=(a,0)$ and either $a\le 2$ or rank two splits. In proposition \[1.1\] and proposition \[prop1.2\] we classify indecomposable rank two globally generated vector bundles with $c_1=(1,1)$ and $c_1=(1,2)$. In the fourth section we deal with the case $c_1=(2,2)$. In the lemmas \[nost\], \[le1\], \[le2\] and \[le3\] we give a complete classification. We can find indecomposable rank two globally generated vector bundles for any $3\leq c_2\leq 8$ except $c_2=7$. We apply an old method of associating to a rank $2$ vector bundle ${{\ensuremath{{\mathcal{F}}}}}$ on $Q$ a $0$-dimensional subscheme $Z$ on $Q$, and relate the global generatedness of ${{\ensuremath{{\mathcal{F}}}}}$ and the ideal sheaf ${{\ensuremath{{\mathcal{I}}}}}_Z$ twisted by ${{\ensuremath{{\mathcal{O}}}}}_Q(c_1)$ with $c_1=c_1({{\ensuremath{{\mathcal{F}}}}})$. Since ${{\ensuremath{{\mathcal{I}}}}}_Z(c_1)$ is globally generated, $Z$ is contained in a complete intersection of two hypersurfaces in $|{{\ensuremath{{\mathcal{O}}}}}_Q(c_1)|$. It gives us an upper bound for the possible second Chern class of ${{\ensuremath{{\mathcal{F}}}}}$. We then check the global generatedness case by case. In the last section we study the higher rank case. The goal of this section is to describe the possible rank and Chern classes with which indecomposable vector bundles on $Q$ exist (see theorem \[hr\]). If ${{\ensuremath{{\mathcal{E}}}}}$ is a globally generated vector bundle on $Q$ of rank $r\ge 3$, then ${{\ensuremath{{\mathcal{E}}}}}$ can be expressed as an extension of a globally generated vector bundle ${{\ensuremath{{\mathcal{F}}}}}$ of rank $2$ by trivial factors. In other words, ${{\ensuremath{{\mathcal{E}}}}}$ can be determined by linearly independent elements in ${{\operatorname{Ext}}}^1( {{\ensuremath{{\mathcal{F}}}}}, {{\ensuremath{{\mathcal{O}}}}}_Q)$ and it gives us an upper bound for the possible rank of indecomposable vector bundle. Then we check the existence of indecomposable and globally generated vector bundles case by case with these bounds. We are grateful to the anonymous referees for several crucial corrections and suggestions. Preliminaries ============= Throughout the article, our base field is ${{\ensuremath{\mathbb{C}}}}$, the field of complex numbers. Let $Q$ be a smooth quadric surface isomorphic to ${{\ensuremath{\mathbb{P}}}}V_1 \times {{\ensuremath{\mathbb{P}}}}V_2$ for two 2-dimensional vector spaces $V_1$ and $V_2$ and then it is embedded into ${{\ensuremath{\mathbb{P}}}}^3\simeq {{\ensuremath{\mathbb{P}}}}V$ by the Segre map, where $V=V_1\otimes V_2$. Let us denote $f^*{{\ensuremath{{\mathcal{O}}}}}_{{{\ensuremath{\mathbb{P}}}}^1}(a) \otimes g^*{{\ensuremath{{\mathcal{O}}}}}_{{{\ensuremath{\mathbb{P}}}}^1}(b)$ by ${{\ensuremath{{\mathcal{O}}}}}_Q(a,b)$ and ${{\ensuremath{{\mathcal{E}}}}}\otimes {{\ensuremath{{\mathcal{O}}}}}_Q(a,b)$ by ${{\ensuremath{{\mathcal{E}}}}}(a,b)$ for coherent sheaves ${{\ensuremath{{\mathcal{E}}}}}$ on $Q$, where $f$ and $g$ are the projections from $Q$ to each factors. Then the canonical line bundle $\omega_Q$ of $Q$ is ${{\ensuremath{{\mathcal{O}}}}}_Q(-2,-2)$. We also denote by ${{\ensuremath{{\mathcal{E}}}}}^\vee$ the dual sheaf of ${{\ensuremath{{\mathcal{E}}}}}$. For a coherent sheaf ${{\ensuremath{{\mathcal{E}}}}}$ of rank $r$ on $Q$ with the Chern classes $c_1=(a,b)\in {{\ensuremath{\mathbb{Z}}}}^{\oplus 2}$ and $c_2=c\in {{\ensuremath{\mathbb{Z}}}}$, we have : $$\begin{aligned} c_1({{\ensuremath{{\mathcal{E}}}}}(s,t))&=(a+rs,b+rt)\\ c_2({{\ensuremath{{\mathcal{E}}}}}(s,t)) &=c+(r-1)(at+bs)+2st{r \choose 2}\\ \chi ({{\ensuremath{{\mathcal{E}}}}})&=(a+1)(b+1)+r-c-1\end{aligned}$$ for $(s,t)\in {{\ensuremath{\mathbb{Z}}}}^{\oplus 2}$.\ We use the following notion of stability: a vector bundle ${{\ensuremath{{\mathcal{E}}}}}$ of rank $2$ with $c_1=(0,0)$ is stable if and only if $H^0({{\ensuremath{{\mathcal{E}}}}})=0, H^0({{\ensuremath{{\mathcal{E}}}}}(-1,1))= 0, H^0({{\ensuremath{{\mathcal{E}}}}}(1,-1))= 0$ (see [@LP]). We denote by $H_*^i({{\ensuremath{{\mathcal{E}}}}})$ the graded module $\oplus_{t\in \mathbb Z}H^i(Q, {{\ensuremath{{\mathcal{E}}}}}(t,t))$. We always consider the cohomology of sheaves. ${{\ensuremath{{\mathcal{E}}}}}$ is said to be [*arithmetically Cohen–Macaulay*]{} ([*ACM*]{} for short) if $H_*^1({{\ensuremath{{\mathcal{E}}}}})=0$. The Castelnuovo-Mumford criterion states that if $H^i({{\ensuremath{{\mathcal{E}}}}}(-i,-i))=0$ for all $i>0$, then ${{\ensuremath{{\mathcal{E}}}}}$ is globally generated. We also introduce another criterion for global generatedness which is stronger in some sense. \[gg\][@bm] For a torsion-free coherent sheaf ${{\ensuremath{{\mathcal{E}}}}}$ on $Q$, the condition $$H^1({{\ensuremath{{\mathcal{E}}}}}(-1,0))=H^1({{\ensuremath{{\mathcal{E}}}}}(0,-1))=H^2({{\ensuremath{{\mathcal{E}}}}}(-1,-1))=0$$ implies that ${{\ensuremath{{\mathcal{E}}}}}$ is globally generated. A globally generated vector bundle ${{\ensuremath{{\mathcal{E}}}}}$ fitting into the sequence $$\label{eqii1} 0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(-a,-b){\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus (r+1)} {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}0$$ is called to be [*of maximal type*]{}. \[ff1\] 1. A vector bundle ${{\ensuremath{{\mathcal{E}}}}}$ of maximal type has $c_1=(a,b)$ and $c_2=c_1^2=2ab$. Thus it achieves the maximal possibility of the second Chern class. Let ${{\ensuremath{{\mathcal{F}}}}}$ be an extension of ${{\ensuremath{{\mathcal{E}}}}}$ by ${{\ensuremath{{\mathcal{O}}}}}_Q$. Since ${{\operatorname{Ext}}}^1 ({{\ensuremath{{\mathcal{O}}}}}_Q, {{\ensuremath{{\mathcal{O}}}}}_Q)=0$, so ${{\ensuremath{{\mathcal{F}}}}}$ is of maximal type. In other words, the property of maximal type is preserved under the extension by ${{\ensuremath{{\mathcal{O}}}}}_Q$. 2. Every vector bundle ${{\ensuremath{{\mathcal{E}}}}}$ of maximal type satisfies $H^0({{\ensuremath{{\mathcal{E}}}}}(-x,-y))=0$ for all pairs of nonnegative integers $(x,y)$ with $x^2+y^2>1$. In fact, in the case of $c_1=(1,0)$ or $(0,1)$, we have ${{\ensuremath{{\mathcal{E}}}}}\cong {{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus (r-1)}\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(c_1)$. 3. For a rank $r$ globally generated vector bundle ${{\ensuremath{{\mathcal{E}}}}}$ of maximal type with $c_1({{\ensuremath{{\mathcal{E}}}}})=(a,b)$, we have $h^0({{\ensuremath{{\mathcal{E}}}}})=r+1$ from the sequence (\[eqii1\]) if $h^1({{\ensuremath{{\mathcal{O}}}}}_Q(-a,-b))=0$. The condition holds if either $a>0$ and $b>0$ or $(a,b)=(0,1)$ or $(a,b)=(1,0)$. \[mi\] Let us fix a vector bundle ${{\ensuremath{{\mathcal{E}}}}}$ fitting in the exact sequence (\[eqii1\]) with $h^1({{\ensuremath{{\mathcal{O}}}}}_Q(-a,-b))=0$. If ${{\ensuremath{{\mathcal{E}}}}}$ has no trivial factor, then it is indecomposable. Let us assume that it is decomposed as ${{\ensuremath{{\mathcal{E}}}}}\cong {{\ensuremath{{\mathcal{E}}}}}_1 \oplus {{\ensuremath{{\mathcal{E}}}}}_2$. Since $h^0({{\ensuremath{{\mathcal{E}}}}})=\mathrm{rk}({{\ensuremath{{\mathcal{E}}}}})+1$ (Remark \[ff1\]) and ${{\ensuremath{{\mathcal{E}}}}}_1$ and ${{\ensuremath{{\mathcal{E}}}}}_2$ are globally generated, we have $h^0({{\ensuremath{{\mathcal{E}}}}}_i) = \mathrm{rk}({{\ensuremath{{\mathcal{E}}}}}_i)$ for some $i$. In particular ${{\ensuremath{{\mathcal{E}}}}}_i$ is trivial and so ${{\ensuremath{{\mathcal{E}}}}}$ has a trivial factor. \[ff2\] For a vector bundle ${{\ensuremath{{\mathcal{E}}}}}$ fitting in the exact sequence (\[eqii1\]) with $a=0$ and $r\ge 2$, let us call ${\varphi}: {{\ensuremath{{\mathcal{O}}}}}_Q(0,-b){\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus (r+1)}$ the map with $\mathrm{coker}({\varphi})={{\ensuremath{{\mathcal{E}}}}}$. Let $\pi : Q{\xrightarrow{\ \ }}{{\ensuremath{\mathbb{P}}}}^1$ denote the projection onto the second factor. Since $a=0$, the map ${\varphi}$ is the pull-back by $\pi$ of a map $v: {{\ensuremath{{\mathcal{O}}}}}_{{{\ensuremath{\mathbb{P}}}}^1}(-b){\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_{{{\ensuremath{\mathbb{P}}}}^1}^{\oplus (r+1)}$ with locally free cokernel. Since ${{\ensuremath{{\mathcal{E}}}}}\cong \pi ^\ast (\mathrm{coker}(v))$ and $r\ge 2$, ${{\ensuremath{{\mathcal{E}}}}}$ is decomposable. \[trivial\] Let ${{\ensuremath{{\mathcal{F}}}}}$ be a globally generated sheaf on $Q$. Assume the existence of a non-zero map $f: {{\ensuremath{{\mathcal{F}}}}}{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q$. Then ${{\ensuremath{{\mathcal{F}}}}}$ has ${{\ensuremath{{\mathcal{O}}}}}_Q$ as a factor. Since ${{\ensuremath{{\mathcal{F}}}}}$ is globally generated, every non-zero map ${{\ensuremath{{\mathcal{F}}}}}{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q$ is surjective. Since ${{\ensuremath{{\mathcal{F}}}}}$ is globally generated, there is an integer $n>0$ and a surjection $h: {{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus n} {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{F}}}}}$, say induced by $s_1,\dots ,s_n\in H^0(\mathcal{H}om({{\ensuremath{{\mathcal{O}}}}}_Q,{{\ensuremath{{\mathcal{F}}}}})) $. Since $f\circ h$ is surjective, there is $s_i\ne 0$ such that $f\circ s_i\ne 0$. Hence $f\circ s_i: {{\ensuremath{{\mathcal{O}}}}}_Q{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q$ is surjective. Hence $f\circ s_i$ is an isomorphism and so $s_i({{\ensuremath{{\mathcal{O}}}}}_Q)$ is a direct factor of ${{\ensuremath{{\mathcal{F}}}}}$. Rank two globally generated vector bundle with $c_1< (2,2)$ ============================================================ Let us assume that ${{\ensuremath{{\mathcal{E}}}}}$ is a globally generated vector bundle of rank $r$ on $Q$ with the Chern classes $(c_1, c_2)$ and then ${{\ensuremath{{\mathcal{E}}}}}$ fits into the following sequence for general $r-1$ sections of $H^0({{\ensuremath{{\mathcal{E}}}}})$ : $$\label{eqa1} 0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus (r-1)} {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{I}}}}}_Z (c_1) {\xrightarrow{\ \ }}0,$$ where $Z$ is a reduced $0$-dimensional subscheme of $Q$ whose length is $c_2$ due to the parts (E) and (F) at page 4 in [@am]. Since ${{\ensuremath{{\mathcal{I}}}}}_Z(c_1)$ is globally generated, $Z$ is contained in the complete intersection of two hypersurfaces of bidegree $c_1$. In particular, we have $c_2 \le c_1^2$. \[prop1\][@sierra] Let ${{\ensuremath{{\mathcal{E}}}}}$ be a globally generated vector bundle of rank $r$ on $Q$ such that $H^0({{\ensuremath{{\mathcal{E}}}}}(-c_1))\not= 0$, where $c_1$ is the first Chern class of ${{\ensuremath{{\mathcal{E}}}}}$. Then we have $${{\ensuremath{{\mathcal{E}}}}}\simeq {{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus (r-1)} \oplus {{\ensuremath{{\mathcal{O}}}}}_Q(c_1).$$ As an automatic consequence, ${{\ensuremath{{\mathcal{E}}}}}\simeq {{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus r}$ is the only globally generated vector bundle of rank $r$ on $Q$ with $c_1=(0,0)$. Let $\ge$ be a partial ordering on ${{\ensuremath{\mathbb{Z}}}}^{\oplus 2}$ defined by $(a',b')\ge (a,b)$ if $a' \ge a$ and $b'\ge b$. For a vector bundle ${{\ensuremath{{\mathcal{E}}}}}$ on $Q$, a maximal element with respect to $\ge$ in the following set $$\{ (a,b)\in {{\ensuremath{\mathbb{Z}}}}^{\oplus 2} ~|~ h^0({{\ensuremath{{\mathcal{E}}}}}(-a,-b))\ne 0 \}$$ is called [*an index*]{} of ${{\ensuremath{{\mathcal{E}}}}}$. In general there may exist more than one indices for a given vector bundle. If $(m,n)$ is an index of a globally generated vector bundle ${{\ensuremath{{\mathcal{E}}}}}$ of rank 2 on $Q$ with $c_1({{\ensuremath{{\mathcal{E}}}}})=(a,b)$, then we have $(m,n)\leq (a,b)$. On the other hand, there exists an index of the form $(a,0)$ or $(0,b)$ with $a,b\geq 0$. In general, an arbitrary index $(m,n)$ of ${{\ensuremath{{\mathcal{E}}}}}$ does not have to satisfy $(0,0)\leq (m,n)$. A referee gave the following example. Fixing a generator of $\Lambda^2 V_i$ where $Q\cong {{\ensuremath{\mathbb{P}}}}V_1 \times {{\ensuremath{\mathbb{P}}}}V_2$, we get isomorphisms $V_i \cong V_i^\vee$. Let $$\tau : {{\ensuremath{{\mathcal{O}}}}}_Q(-1,-1){\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q \otimes (V_1 \otimes V_2)$$ be the canonical injective morphism. Let $\triangle \subset V_1 \otimes V_2$ be a linear subspace of dimension $1$: $\triangle ={{\ensuremath{\mathbb{C}}}}\delta$. We can also view $\triangle$ as a point in ${{\ensuremath{\mathbb{P}}}}\mathrm{Hom} (V_1^\vee, V_2)$. Then let $$\tau_{\triangle} : {{\ensuremath{{\mathcal{O}}}}}_Q(-1,-1) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q \otimes ( (V_1 \otimes V_2)/\triangle)={{\ensuremath{{\mathcal{O}}}}}_Q\otimes {{\ensuremath{\mathbb{C}}}}^3$$ be the composition of $\tau$ with the quotient $V_1 \otimes V_2 {\xrightarrow{\ \ }}(V_1 \otimes V_2)/\triangle$. Then $\tau_{\triangle}$ is injective as a morphism of vector bundles if and only if $\triangle$ does not belong to the quadric of non-bijective morphisms in ${{\ensuremath{\mathbb{P}}}}\mathrm{Hom}(V_1^\vee, V_2)$, which is the same as the quadric of decomposable elements in ${{\ensuremath{\mathbb{P}}}}(V_1 \otimes V_2)$. Supposing $\delta$ is bijective, ${{\ensuremath{{\mathcal{E}}}}}_{\triangle}:=\mathrm{coker}(\tau_{\delta})$ is a globally generated vector bundle of rank $2$ with Chern classes $c_1=(1,1)$ and $c_2=2$. From the exact sequence $$0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(-1,-1) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q\otimes {{\ensuremath{\mathbb{C}}}}^3 {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}_{\triangle} {\xrightarrow{\ \ }}0$$ we have $\mathrm{Hom}({{\ensuremath{{\mathcal{O}}}}}_Q(-1,1), {{\ensuremath{{\mathcal{E}}}}}_{\triangle}) \cong H^1 ({{\ensuremath{{\mathcal{O}}}}}_Q(0,-2))\cong {{\ensuremath{\mathbb{C}}}}$. We also deduce immediately that $\mathrm{Hom}({{\ensuremath{{\mathcal{O}}}}}_Q(0,1), {{\ensuremath{{\mathcal{E}}}}}_{\triangle})=\{0\}$, and an exact sequence $$\xymatrix{0\ar[r] &\mathrm{Hom}({{\ensuremath{{\mathcal{O}}}}}_Q(-1,2), {{\ensuremath{{\mathcal{E}}}}}_{\triangle}) \ar[r] & H^1({{\ensuremath{{\mathcal{O}}}}}_Q(0,-3)) \ar@{=}[d] \ar[r]^-{\varphi}& H^1({{\ensuremath{{\mathcal{O}}}}}_Q(1,-2))\otimes {{\ensuremath{\mathbb{C}}}}^3 \ar@{=}[d]\\ &&V_2 & V_1^\vee \otimes (V_1\otimes V_2)/\triangle}$$ The map ${\varphi}: V_2 {\xrightarrow{\ \ }}V_1^\vee \otimes (V_1 \otimes V_2)/\triangle$ is the canonical injection and thus we have $\mathrm{Hom}({{\ensuremath{{\mathcal{O}}}}}_Q(-1,2), {{\ensuremath{{\mathcal{E}}}}}_{\triangle})=\{0\}$. This proves that $(-1,1)$ is an index of ${{\ensuremath{{\mathcal{E}}}}}_{\triangle}$ and so is $(1,-1)$ similarly. Thus we have exact sequences $$\begin{aligned} 0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(-1,1) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}_{\triangle} {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(2,0) {\xrightarrow{\ \ }}0,\\ 0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(1,-1) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}_{\triangle} {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(0,2) {\xrightarrow{\ \ }}0,\\ 0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(-1,1)\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(1,-1) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}_{\triangle} {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_C(1,1) {\xrightarrow{\ \ }}0,\end{aligned}$$ where $C$ is the curve determined by $\delta$, viewed as an element of $V_1^\vee \otimes V_2^\vee$. Later we show by classification that ${{\ensuremath{{\mathcal{E}}}}}_{\triangle}$ is isomorphic to a pull-back of $T{{\ensuremath{\mathbb{P}}}}^2(-1)$, the tangent bundle of ${{\ensuremath{\mathbb{P}}}}^2$ twisted by $-1$, under a linear projection from $Q$ to ${{\ensuremath{\mathbb{P}}}}^2$. \[prop4\] Let ${{\ensuremath{{\mathcal{E}}}}}$ be a globally generated vector bundle of rank $r$ on $Q$ with $c_1=(a,0)$. If $r>2$, then assume $a\le 2$.Then we have $${{\ensuremath{{\mathcal{E}}}}}\cong \bigoplus_{i=1}^r{{\ensuremath{{\mathcal{O}}}}}_Q(a_i,0)$$ where $a_i$’s are nonnegative integers whose sum is $a$. Similar answer can be given when $c_1=(0,b)$. Let us assume that the rank of ${{\ensuremath{{\mathcal{E}}}}}$ is $2$ and $(c,0)$ be an index of ${{\ensuremath{{\mathcal{E}}}}}$. Then ${{\ensuremath{{\mathcal{E}}}}}$ fits into the sequence $$0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(c,0) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{I}}}}}_Z(a-c,0) {\xrightarrow{\ \ }}0.$$ Since ${{\ensuremath{{\mathcal{I}}}}}_Z(a-c,0)$ is globally generated, so $Z$ must be empty and thus we have ${{\ensuremath{{\mathcal{E}}}}}\cong {{\ensuremath{{\mathcal{O}}}}}_Q(c,0)\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(a-c,0)$. Now assume $r>2$ and $a\le 2$. Since ${{\ensuremath{{\mathcal{E}}}}}$ is globally generated, we have an exact sequence $$0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus (r-2)} {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{F}}}}}{\xrightarrow{\ \ }}0$$ where ${{\ensuremath{{\mathcal{F}}}}}$ is globally generated and thus we have ${{\ensuremath{{\mathcal{F}}}}}\cong {{\ensuremath{{\mathcal{O}}}}}_Q(b_1,0)\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(b_2,0)$ with $b_1+b_2=a$. If $b_i \le 1$ for all $i$, then this exact sequence splits. If $b_i\ge 2$ for some $i$, say $b_1\ge 2$, then $b_1=a=2$ and $b_2=0$. Lemma \[trivial\] gives ${{\ensuremath{{\mathcal{E}}}}}\cong {{\ensuremath{{\mathcal{O}}}}}_Q\oplus {{\ensuremath{{\mathcal{G}}}}}$ for some ${{\ensuremath{{\mathcal{G}}}}}$. Apply induction on the rank to the bundle ${{\ensuremath{{\mathcal{G}}}}}$. Now let us focus on the case of rank $2$. From a general section $s\in H^0({{\ensuremath{{\mathcal{E}}}}})$, we have $$\label{eqa3} 0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q \stackrel{s}{{\xrightarrow{\ \ }}} {{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{I}}}}}_Z(c_1) {\xrightarrow{\ \ }}0.$$ Since ${{\ensuremath{{\mathcal{I}}}}}_Z(c_1)$ is also globally generated, a general curve of bidegree $c_1$ containing $Z$ is smooth by the Bertini theorem. \[1.1\] Let ${{\ensuremath{{\mathcal{E}}}}}$ be a globally generated vector bundle of rank $2$ on $Q$ with $c_1=(1,1)$. If ${{\ensuremath{{\mathcal{E}}}}}$ does not split, then ${{\ensuremath{{\mathcal{E}}}}}$ is isomorphic to ${{\ensuremath{{\mathcal{A}}}}}_P:={\varphi}_P^* (T{{\ensuremath{\mathbb{P}}}}^2(-1))$, where ${\varphi}_P: Q {\xrightarrow{\ \ }}{{\ensuremath{\mathbb{P}}}}^2$ is the linear projection with the center $P\in {{\ensuremath{\mathbb{P}}}}^3\setminus Q$. By Proposition \[prop1\], we can assume that $H^0({{\ensuremath{{\mathcal{E}}}}}(-1,-1))=0$. If $(1,0)$ is an index of ${{\ensuremath{{\mathcal{E}}}}}$, then we have the sequence $$0 {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(1,0) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{I}}}}}_{Z'} (0,1) {\xrightarrow{\ \ }}0,$$ where $Z'$ is a 0-dimensional subscheme of $Q$ whose length is $c_2-1$. Since ${{\ensuremath{{\mathcal{I}}}}}_{Z'}(0,1)$ is globally generated, $Z'$ is contained in the complete intersection of two hypersurfaces of bidegree $(0,1)$. Thus $Z'$ is an empty set and so ${{\ensuremath{{\mathcal{E}}}}}$ is isomorphic to ${{\ensuremath{{\mathcal{O}}}}}_Q(1,0)\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(0,1)$. We obtain the same answer when $(0,1)$ is an index of ${{\ensuremath{{\mathcal{E}}}}}$. Now assume that $(0,0)$ is the index of ${{\ensuremath{{\mathcal{E}}}}}$ and then ${{\ensuremath{{\mathcal{E}}}}}$ fits into the sequence (\[eqa3\]) with $c_1=(1,1)$ and so ${{\ensuremath{{\mathcal{I}}}}}_Z(1,1)$ is globally generated. In particular, $Z$ is contained in the complete intersection of two hyperplane sections of $Q$ and so $\deg (Z)\leq 2$. If $\deg (Z)\leq 1$, then $h^0({{\ensuremath{{\mathcal{E}}}}}(-1,0))$ or $h^0({{\ensuremath{{\mathcal{E}}}}}(0,-1))$ would be nonzero, a contradiction. So we have $\deg (Z)=2$ and $Z$ is the complete intersection, not on any ruling of $Q$. From the locally free resolution of ${{\ensuremath{{\mathcal{I}}}}}_Z$ $$0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(-2,-2) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(-1,-1)^{\oplus 2} {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{I}}}}}_Z {\xrightarrow{\ \ }}0,$$ we obtain the locally free resolution of ${{\ensuremath{{\mathcal{E}}}}}$ : $$0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(-1,-1) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus 3} {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}0.$$ In particular, ${{\ensuremath{{\mathcal{E}}}}}$ is stable with respect to ${{\ensuremath{{\mathcal{O}}}}}_Q(1,1)$. Conversely, any stable vector bundle ${{\ensuremath{{\mathcal{E}}}}}$ of rank $2$ on $Q$ with the Chern classes $c_1=(1,1)$ and $c_2=2$ is globally generated, since it fits into the sequence $$0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{I}}}}}_Z (1,1) {\xrightarrow{\ \ }}0$$ with a $0$-dimensional subscheme $Z$ of length $2$. Note that the moduli space was shown to be isomorphic to ${{\ensuremath{\mathbb{P}}}}^3 \setminus Q$ in [@huh2] using the set of jumping conics. Thus it is enough to prove that ${{\ensuremath{{\mathcal{A}}}}}_P$ is not isomorphic to ${{\ensuremath{{\mathcal{A}}}}}_O$ if $P\ne O$. It is true since the set of jumping conics of ${{\ensuremath{{\mathcal{A}}}}}_P$ is the hyperplane in $({{\ensuremath{\mathbb{P}}}}^3)^\vee$ corresponding to the point $P\in {{\ensuremath{\mathbb{P}}}}^3$. Without confusion, we simply denote ${{\ensuremath{{\mathcal{A}}}}}_P$ by ${{\ensuremath{{\mathcal{A}}}}}$. The bundle ${{\ensuremath{{\mathcal{A}}}}}$ is of maximal type and so its extension by ${{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus (r-2)}$ is also of maximal type. Since $h^1({{\ensuremath{{\mathcal{A}}}}}^\vee)=1$, so the only possible extension of ${{\ensuremath{{\mathcal{A}}}}}$ by ${{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus (r-2)}$ without any trivial factor happens when $r=3$. Later we will see that such bundle is always isomorphic to $T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q$ for any $P\in {{\ensuremath{\mathbb{P}}}}^3\setminus Q$. \[prop1.2\] Let ${{\ensuremath{{\mathcal{E}}}}}$ be a globally generated vector bundle of rank $2$ on $Q$ with $c_1=(1,2)$. If ${{\ensuremath{{\mathcal{E}}}}}$ does not split, then it is one of the following: 1. $(c_2=2) : 0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(1,0)\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(0,1)^{\oplus 2} {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}0$ 2. $(c_2=3) : 0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(-1,-1) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(0,1)\oplus {{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus 2} {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}0$ 3. $(c_2=4) : 0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(-1,-2) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus 3} {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}0$ Similar answer can be given when $c_1=(2,1)$. ${{\ensuremath{{\mathcal{E}}}}}$ fits into the sequence (\[eqa3\]) with $c_1=(1,2)$. Since ${{\ensuremath{{\mathcal{I}}}}}_Z(1,2)$ is globally generated, $Z$ is contained in a complete intersection of two hypersurfaces of bidegree $(1,2)$ in $Q$. In particular, $\deg (Z)$ is at most 4. If $\deg (Z)=4$, then $Z$ is the complete intersection and so ${{\ensuremath{{\mathcal{I}}}}}_Z$ admits a locally free resolution: $$0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(-2,-4) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(-1,-2)^{\oplus 2} {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{I}}}}}_Z {\xrightarrow{\ \ }}0.$$ It induces a locally free resolution (3) in the list. Assume that $\deg (Z)\leq 3$ and so $h^0({{\ensuremath{{\mathcal{I}}}}}_Z(1,1))\not= 0$. In particular, we have $h^0({{\ensuremath{{\mathcal{E}}}}}(0,-1))\not= 0$ and so an index of ${{\ensuremath{{\mathcal{E}}}}}$ is $(a,b)$ with $b\ge 1$ and ${{\ensuremath{{\mathcal{E}}}}}$ fits into the sequence $$0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(a,b) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{I}}}}}_{Z'}(1-a,2-b) {\xrightarrow{\ \ }}0.$$ If either $a=1$ or $b=2$, then $Z'$ is empty and so we have ${{\ensuremath{{\mathcal{E}}}}}\simeq {{\ensuremath{{\mathcal{O}}}}}_Q(a,b)\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(1-a,2-b)$. Now let us assume that $(a,b)=(0,1)$. Since ${{\ensuremath{{\mathcal{I}}}}}_{Z'}(1,1)$ is globally generated, $Z'$ is contained in a complete intersection of two hypersurfaces of bidegree $(1,1)$ in $Q$ and so $\deg (Z')\leq 2$. If $Z'$ is an empty set, then ${{\ensuremath{{\mathcal{E}}}}}$ is isomorphic to ${{\ensuremath{{\mathcal{O}}}}}_Q(0,1)\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(1,1)$. If $\deg (Z')=2$, then $Z'$ is the complete intersection and so ${{\ensuremath{{\mathcal{I}}}}}_{Z'}$ admits a locally free resolution: $$0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(-2,-2) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(-1,-1)^{\oplus 2} {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{I}}}}}_{Z'} {\xrightarrow{\ \ }}0.$$ It induces a locally free resolution (2) in the list. If $\deg (Z')=1$, then $Z'$ is a complete intersection of two hypersurfaces of bidegree $(1,0)$ and $(0,1)$ and so ${{\ensuremath{{\mathcal{I}}}}}_{Z'}$ admits a locally free resolution: $$0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(-1,-1){\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(-1,0)\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(0,-1) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{I}}}}}_{Z'} {\xrightarrow{\ \ }}0.$$ It induces the resolution (1) in the list. When we combine the exact sequence $$0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(0,-1) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus 2} {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(0,1) {\xrightarrow{\ \ }}0$$ to the resolution (2) in the list, we have $$0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(-1,-1)\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(0,-1) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus 4} {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}0.$$ Similarly, the resolution (1) becomes the sequence, $$0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(0,-1)^{\oplus 2} \oplus {{\ensuremath{{\mathcal{O}}}}}_Q(-1,0){\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus 5} {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}0.$$ Case $c_1=(2,2)$ ================ Let $\mathfrak{M}(k)$ be the moduli space of stable vector bundles of rank $2$ on $Q$ with the Chern classes $c_1=(2,2)$ and $c_2=k$ with respect to the ample line bundle ${{\ensuremath{{\mathcal{O}}}}}_Q(1,1)$. By the Bogomolov inequality, the space is empty if $k<2$. In [@Soberon], it was proven to be a smooth, rational and irreducible variety of dimension $4k-11$. In this section, we assume that ${{\ensuremath{{\mathcal{E}}}}}$ is a globally generated vector bundle of rank $2$ on $Q$ with $c_1({{\ensuremath{{\mathcal{E}}}}})=(2,2)$. If $(2,2)$ is the index of ${{\ensuremath{{\mathcal{E}}}}}$, then we have ${{\ensuremath{{\mathcal{E}}}}}\cong {{\ensuremath{{\mathcal{O}}}}}_Q\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(2,2)$ by Proposition \[prop1\]. If $(2,1)$ is an index of ${{\ensuremath{{\mathcal{E}}}}}$, then we have $$0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(2,1) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{I}}}}}_Z(0,1) {\xrightarrow{\ \ }}0.$$ Since ${{\ensuremath{{\mathcal{I}}}}}_Z(0,1)$ is globally generated, so $Z$ is an empty set and we have ${{\ensuremath{{\mathcal{E}}}}}\cong {{\ensuremath{{\mathcal{O}}}}}_Q(2,1)\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(0,1)$. In general, if $(a,b)$ is an index of ${{\ensuremath{{\mathcal{E}}}}}$ with $a+b\ge 3$, then we have ${{\ensuremath{{\mathcal{E}}}}}\cong {{\ensuremath{{\mathcal{O}}}}}_Q(a,b)\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(2-a,2-b)$. \[nost\] If $(0,2)$ is an index of a non-splitting bundle ${{\ensuremath{{\mathcal{E}}}}}$, then ${{\ensuremath{{\mathcal{E}}}}}$ arises in the following extension: $$0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(0,2) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(2,0) {\xrightarrow{\ \ }}0.$$ A similar result holds for the case of index $(2,0)$. We have the following exact sequence $$0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(0,2) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{I}}}}}_Z(2,0) {\xrightarrow{\ \ }}0$$ and since ${{\ensuremath{{\mathcal{I}}}}}_Z(2,0)$ is globally generated, we have $\deg (Z)=0$. So ${{\ensuremath{{\mathcal{I}}}}}_Z$ is a line bundle and we get the claimed extension. Since $\dim({{\operatorname{Ext}}}^1( {{\ensuremath{{\mathcal{O}}}}}_Q(2,0), {{\ensuremath{{\mathcal{O}}}}}_Q(0,2)))=3$ the bundle ${{\ensuremath{{\mathcal{E}}}}}$ may be indecomposable. \[le1\] If $(1,1)$ is an index of non-splitting ${{\ensuremath{{\mathcal{E}}}}}$, then ${{\ensuremath{{\mathcal{E}}}}}$ satisfies one of the following: 1. $(c_2=3) : 0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(1,1)\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(1,0)\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(0,1) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}0$ 2. $(c_2=4) : 0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(-1,-1) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus 2} \oplus {{\ensuremath{{\mathcal{O}}}}}_Q(1,1) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}0$. We have the following exact sequence $$0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(1,1) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{I}}}}}_Z(1,1) {\xrightarrow{\ \ }}0$$ and since ${{\ensuremath{{\mathcal{I}}}}}_Z(1,1)$ is globally generated, we have $\deg (Z)\le 2$. If $\deg (Z)=0$, then we have ${{\ensuremath{{\mathcal{E}}}}}\cong {{\ensuremath{{\mathcal{O}}}}}_Q(1,1)^{\oplus 2}$. If $\deg (Z)=1$, say $Z=\{P\}$, then we have $$h^1({{\ensuremath{{\mathcal{I}}}}}_P)=h^2({{\ensuremath{{\mathcal{I}}}}}_P(-1,-1))=0$$ and so ${{\ensuremath{{\mathcal{I}}}}}_P(1,1)$ is globally generated due to Castelnuovo-Mumford criterion. Since $h^1({{\ensuremath{{\mathcal{O}}}}}_Q(1,1))=0$, so ${{\ensuremath{{\mathcal{E}}}}}$ is globally generated. Note that ${{\ensuremath{{\mathcal{I}}}}}_P$ admits the following resolution $$0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(-1,-1) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(-1,0)\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(0,-1) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{I}}}}}_P{\xrightarrow{\ \ }}0$$ and so ${{\ensuremath{{\mathcal{E}}}}}$ admits the following resolution (1). Similarly when $\deg (Z)=2$, we obtain the resolution (2). If ${{\ensuremath{{\mathcal{E}}}}}$ has an index $(a,b)$ with $a+b\le 1$, then ${{\ensuremath{{\mathcal{E}}}}}$ fits into the sequence $$0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(a,b) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{I}}}}}_{Z}(2-a,2-b) {\xrightarrow{\ \ }}0.$$ Since ${{\ensuremath{{\mathcal{I}}}}}_Z(2-a,2-b)$ is globally generated, so we have $\deg (Z)\le 2(2-a)(2-b)$. Thus we have $c_2=c_2({{\ensuremath{{\mathcal{E}}}}})=\deg (Z)+a(2-b)+b(2-a)\le 8-2(a+b)$. Since ${{\ensuremath{{\mathcal{E}}}}}$ is stable, we have ${{\ensuremath{{\mathcal{E}}}}}\in \mathfrak{M}(c_2)$. \[le2\] Let ${{\ensuremath{{\mathcal{E}}}}}$ be a globally generated vector bundle of rank $2$ on $Q$ with $c_1=(2,2)$ and index $(1,0)$. Then ${{\ensuremath{{\mathcal{E}}}}}$ is in $\mathfrak{M}(c_2)$ with $4\le c_2 \le 6$. In fact, we have the following: 1. Every bundle ${{\ensuremath{{\mathcal{E}}}}}\in \mathfrak{M}(4)$ is globally generated. 2. A general bundle in $\mathfrak{M}(5)$ is globally generated. 3. If $c_2({{\ensuremath{{\mathcal{E}}}}})=6$, then it fits into the sequence $$0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(-1,-2) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus 2} \oplus {{\ensuremath{{\mathcal{O}}}}}_Q(1,0) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}0.$$ Similar answer can be given when $(0,1)$ is an index of ${{\ensuremath{{\mathcal{E}}}}}$. Let ${{\ensuremath{{\mathcal{E}}}}}$ be an arbitrary bundle with index $(1,0)$ and so fitting into the sequence $$\label{11} 0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(1,0) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{I}}}}}_Z(1,2) {\xrightarrow{\ \ }}0$$ with $\deg (Z)\le 4$. If $\deg (Z)\le 1$, then we obtain $h^0({{\ensuremath{{\mathcal{I}}}}}_Z(0,1))\ne 0$ and so $h^0({{\ensuremath{{\mathcal{E}}}}}(-1,-1))\ne 0$. It is a contradiction to the assumption that $(1,0)$ is an index. Thus we have $\deg (Z)\ge 2$. Assume that $\deg (Z)=2$. If $H^0({{\ensuremath{{\mathcal{I}}}}}_Z(0,1)) \ne 0$, then $h^0({{\ensuremath{{\mathcal{E}}}}}(-1,-1))\ne 0$ and it is contradicting to the assumption that $(1,0)$ is an index. Thus $h^0({{\ensuremath{{\mathcal{I}}}}}_Z(0,1))=0$ and so $h^0({{\ensuremath{{\mathcal{I}}}}}_Z(0,2))=1$. It implies that $h^1({{\ensuremath{{\mathcal{I}}}}}_Z(0,2))=0$. We can also prove that $h^1({{\ensuremath{{\mathcal{I}}}}}_Z(1,1))=h^2({{\ensuremath{{\mathcal{I}}}}}_Z(0,1))=0$. Thus ${{\ensuremath{{\mathcal{I}}}}}_Z(1,2)$ is globally generated by Theorem \[gg\]. From the exact sequence, we also obtain the stability of ${{\ensuremath{{\mathcal{E}}}}}$. Now let us pick a bundle ${{\ensuremath{{\mathcal{E}}}}}$ from $\mathfrak{M}(4)$. From the stability condition, we have $$h^0({{\ensuremath{{\mathcal{E}}}}}(-1,-1))=h^0({{\ensuremath{{\mathcal{E}}}}}(0,-2))=h^0({{\ensuremath{{\mathcal{E}}}}}(-2,0))=0.$$ Since $\chi ({{\ensuremath{{\mathcal{E}}}}}(-1,0))=2$ and $h^2({{\ensuremath{{\mathcal{E}}}}}(-1,0))=0$, we have the sequence (\[11\]) with $Z$ of length $2$ such that $h^0({{\ensuremath{{\mathcal{I}}}}}_Z(0,1))=0$. As before, it implies the globally generatedness of ${{\ensuremath{{\mathcal{E}}}}}$. Let us assume that $\deg (Z)=3$ and in particular we have $c_2({{\ensuremath{{\mathcal{E}}}}})=5$. Let $C\subset Q$ be a smooth curve of type $(1,2)$ and then it is a rational normal curve of ${{\ensuremath{\mathbb{P}}}}^3$. Let $Z'\subset C$ be any zero-dimensional subscheme of degree $5$. Since $C\cong {{\ensuremath{\mathbb{P}}}}^1$ and $\deg ({{\ensuremath{{\mathcal{I}}}}}_{Z',C}(2))=1$, ${{\ensuremath{{\mathcal{I}}}}}_{Z',C}(2)$ is globally generated. It implies that the sheaf ${{\ensuremath{{\mathcal{I}}}}}_{Z'}(2,2)$ is globally generated since the restriction map $H^0({{\ensuremath{{\mathcal{O}}}}}_Q(2,2)) {\xrightarrow{\ \ }}H^0(C, {{\ensuremath{{\mathcal{O}}}}}_C(2))$ is surjective. Thus the bundles fitting into the sequence $$0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{I}}}}}_{Z'}(2,2) {\xrightarrow{\ \ }}0$$ are globally generated. If $\deg (Z)=4$, then $Z$ is a complete intersection of two hypersurfaces of bidegree $(1,2)$ and thus we obtain the resolution (3) by a routine way. \[re44\]$ $ 1. For ${{\ensuremath{{\mathcal{E}}}}}\in \mathfrak{M}(k)$ with $k\in \{4,5\}$, we can observe that $(1,0)$ and $(0,1)$ are the indices of ${{\ensuremath{{\mathcal{E}}}}}$ simultaneously, but not $(1,1)$. 2. By twisting with ${{\ensuremath{{\mathcal{O}}}}}_Q(-1,-1)$, we can identify $\mathfrak{M}(4)$ with $\mathfrak{M} ((0,0),2)$, the moduli space of stable vector bundles of rank $2$ on $Q$ with $c_1=(0,0)$ and $c_2=2$. Le Potier [@LP] analyzes the restriction to the quadric $Q$ of null correlation bundles ${{\ensuremath{{\mathcal{N}}}}}$. Let $\mathfrak{M}_{\mathbb P^3}^0(0,1)$ be the open subset of $\mathfrak{M}_{\mathbb P^3}(0,1)$ consisting of all bundles ${{\ensuremath{{\mathcal{N}}}}}$ such that ${{\ensuremath{{\mathcal{N}}}}}|_{{\mathcal{Q}}}$ is stable on $Q$. He shows that the restriction gives an ètale quasi-finite morphism from $\mathfrak{M}_{\mathbb P^3}^0(0,1)$ onto an open proper subset $U \subset \mathfrak{M}((0,0),2)$. The generic bundle ${{\ensuremath{{\mathcal{E}}}}}$ of $U$ has a twin pair (a Tjurin pair) of null correlation bundles restricting to it, while there are bundles ${{\ensuremath{{\mathcal{E}}}}}$ in $U$ with a unique null correlation bundle restricting to it. In [@Soberon] Soberon-Chavez compactifies $\mathfrak{M}((0,0),2)$ using only the non-split and non-stable bundles described in Lemma \[nost\]. In [@mr] is given an example of a bundle in $\mathfrak{M}((0,0),2)$ but not in $U$, i.e. a stable bundle which is not the restriction of a null correlation bundle. 3. $\mathfrak{M} (5)$ is invariant by the involution $\sigma : Q{\xrightarrow{\ \ }}Q$ which exchanges the two rulings of $Q$, i.e. $\sigma ^\ast ({{\ensuremath{{\mathcal{E}}}}})\in \mathfrak{M} (5)$ for each ${{\ensuremath{{\mathcal{E}}}}}\in \mathfrak{M} (5)$. For a fixed ${{\ensuremath{{\mathcal{E}}}}}\in \mathfrak{M} (5)$, we obviously have $c_2(\sigma ^\ast ({{\ensuremath{{\mathcal{E}}}}}))=5$. Since $\sigma ^\ast ({{\ensuremath{{\mathcal{O}}}}}_Q(2,2)) \cong {{\ensuremath{{\mathcal{O}}}}}_Q(2,2)$, we have $c_1(\sigma ^\ast ({{\ensuremath{{\mathcal{E}}}}})) = (2,2)$. Assume that $\sigma ^\ast ({{\ensuremath{{\mathcal{E}}}}})$ is not stable, i.e. assume the existence of a line bundle ${{\ensuremath{{\mathcal{L}}}}}= {{\ensuremath{{\mathcal{O}}}}}_Q(a,b) \subset \sigma ^\ast ({{\ensuremath{{\mathcal{E}}}}})$ such that $a+b \ge 4$. We have ${{\ensuremath{{\mathcal{O}}}}}_Q(b,a) \cong \sigma ^\ast ({{\ensuremath{{\mathcal{L}}}}}) \subset \sigma ^\ast (\sigma ^\ast ({{\ensuremath{{\mathcal{E}}}}})) \cong {{\ensuremath{{\mathcal{E}}}}}$. Hence ${{\ensuremath{{\mathcal{E}}}}}$ is not stable, a contradiction. When the index is $(0,0)$ we need the following definition: Let $Z\subset Q$ be a locally complete intersection (l.c.i.) $0$-dimensional subscheme. For $(a,b)\in {{\ensuremath{\mathbb{Z}}}}^{\oplus 2}$, $Z$ is said to satisfy [*Cayley-Bacharach*]{} for curves of bidegree $(a,b)$, simply $\mathrm{CB}(a,b)$, if any curve of bidegree $(a,b)$ containing a subscheme of $Z$ with colength 1 contains $Z$. [@gh] For a l.c.i. $0$-dimensional subscheme $Z\subset Q$, there exists an exact sequence $$0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{I}}}}}_Z(a,b) {\xrightarrow{\ \ }}0$$ where ${{\ensuremath{{\mathcal{E}}}}}$ is a vector bundle of rank $2$ if and only if $Z$ satisfies $\mathrm{CB}(a-2,b-2)$. \[le3\] If $(0,0)$ is the index of ${{\ensuremath{{\mathcal{E}}}}}$, then we have one and only one of the following cases: 1. ${{\ensuremath{{\mathcal{E}}}}}\in \mathfrak{M}(6)$ with index $(0,0)$. 2. ${{\ensuremath{{\mathcal{E}}}}}$ is of maximal type. ${{\ensuremath{{\mathcal{E}}}}}$ fits into the sequence $$0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{I}}}}}_Z(2,2) {\xrightarrow{\ \ }}0$$ with $\deg (Z)\le 8$. If $\deg (Z)\le 5$, then we have $h^0({{\ensuremath{{\mathcal{I}}}}}_Z(1,2))\ne 0$ and so $(0,0)$ cannot be the index of ${{\ensuremath{{\mathcal{E}}}}}$. Thus we have $\deg (Z)\ge 6$. If $\deg (Z)=8$, then $Z$ is a complete intersection of two hypersurfaces of bidegree $(2,2)$. Since ${{\ensuremath{{\mathcal{I}}}}}_Z$ admits the resolution $$0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(-4,-4) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(-2,-2)^{\oplus 2} {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{I}}}}}_Z {\xrightarrow{\ \ }}0,$$ so we have the resolution $$0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(-2, -2) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus 3} {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}0$$ and in particular, ${{\ensuremath{{\mathcal{E}}}}}$ is of maximal type. The case of $\deg (Z)=7$ is not possible since ${{\ensuremath{{\mathcal{I}}}}}_Z(2,2)$ is not globally generated (see Proposition \[length7\]). Now let us assume that $\deg (Z)=6$. Let $\mathrm{Hilb}^6(Q)$ be the Hilbert scheme of all $0$-dimensional subschemes of $Q$ with degree $6$. Since $H^0({{\ensuremath{{\mathcal{I}}}}}_{Z'}(0,0))=0$ for all $Z'\subset Z$ with $\deg (Z')>0$, the [*Cayley-Bacharach*]{} condition is satisfied and so we can associate a vector bundle of rank $2$ with $c_1=(2,2)$ and $c_2=6$ to each $Z\in \mathrm{Hilb}^6(Q)$. Let us define the subset ${{\ensuremath{{\mathcal{U}}}}}$ to be the set of all $Z$ such that ${{\ensuremath{{\mathcal{I}}}}}_Z(2,2)$ is globally generated and so is its associated vector bundle of rank $2$. Since $\mathrm{Hilb}^6(Q)$ is smooth and irreducible with dimension $12$, so ${{\ensuremath{{\mathcal{U}}}}}$ is non-empty, irreducible and of dimension $12$. Let us also define a subset ${{\ensuremath{{\mathcal{V}}}}}\subset \mathrm{Hilb}^6(Q)$: $${{\ensuremath{{\mathcal{V}}}}}=\{ Z \in \mathrm{Hilb}^6(Q) ~|~ H^0({{\ensuremath{{\mathcal{I}}}}}_Z(1,2))=H^0({{\ensuremath{{\mathcal{I}}}}}_Z(2,1))=0\}.$$ For $Z\in {{\ensuremath{{\mathcal{V}}}}}$, we have $H^1({{\ensuremath{{\mathcal{I}}}}}_Z(1,2))=H^1({{\ensuremath{{\mathcal{I}}}}}_Z(2,1))=0$ and $H^2({{\ensuremath{{\mathcal{I}}}}}_Z(1,1))=0$. Thus ${{\ensuremath{{\mathcal{I}}}}}_Z(2,2)$ is globally generated by Theorem \[gg\] and so we have ${{\ensuremath{{\mathcal{V}}}}}\subset {{\ensuremath{{\mathcal{U}}}}}$. Note that the union of $6$ general points is contained in ${{\ensuremath{{\mathcal{V}}}}}$. From its definition, a vector bundle ${{\ensuremath{{\mathcal{E}}}}}$ is associated to $Z\in {{\ensuremath{{\mathcal{V}}}}}$ if and only if the index of ${{\ensuremath{{\mathcal{E}}}}}$ is $(0,0)$. For each $Z\in {{\ensuremath{{\mathcal{V}}}}}$, the dimension of ${{\operatorname{Ext}}}^1 ({{\ensuremath{{\mathcal{I}}}}}_Z(2,2), {{\ensuremath{{\mathcal{O}}}}}_Q)$ is $5$ and so the set of all globally generated vector bundles of rank $2$ with $c_1=(2,2)$ and $c_2=6$ is parametrized by an irreducible variety. Each such a bundle is obviously stable. For each $Z\in {{\ensuremath{{\mathcal{V}}}}}$, we have $h^1({{\ensuremath{{\mathcal{I}}}}}_Z(1,2))=0$ and so $h^1({{\ensuremath{{\mathcal{I}}}}}_Z(2,2))=0$. Thus we have $h^0({{\ensuremath{{\mathcal{E}}}}})=4$ and $h^1({{\ensuremath{{\mathcal{E}}}}})=0$. \[length7\] For $0$-dimensional subschemes $Z$ of $Q$ with length $7$, ${{\ensuremath{{\mathcal{I}}}}}_Z(2,2)$ is not globally generated. Before proving the proposition, let us collect some materials on the residual scheme. Let $A\subset Q$ be a $0$-dimensional subscheme and $D\subset Q$ be an effective divisor of type $(u,v)\in {{\ensuremath{\mathbb{Z}}}}^{\oplus 2}$. [*The residual scheme of $A$ with respect to $D$*]{} is defined as the closed subscheme of $Q$ with $({{\ensuremath{{\mathcal{I}}}}}_A : {{\ensuremath{{\mathcal{I}}}}}_D)$ as its ideal sheaf. It is denoted by $\mbox{Res}_D(A)$. From the definition, we have - $\mbox{Res}_D(A) \subseteq A$ and - $\deg (A) = \deg (A\cap D) +\deg (\mbox{Res}_D(A))$. For all $m, n\in {{\ensuremath{\mathbb{Z}}}}$ we also have an exact sequence $$\label{eqa10} 0 {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{I}}}}}_{\mbox{Res}_D(A)}(m-u,n-v) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{I}}}}}_A(m,n){\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{I}}}}}_{D\cap A,D}(m,n) {\xrightarrow{\ \ }}0$$ \[a1\] For $L_1\in \vert {{\ensuremath{{\mathcal{O}}}}}_Q(1,0)\vert$ and $L_2\in \vert {{\ensuremath{{\mathcal{O}}}}}_Q(0,1)\vert$, let us take a zero-dimensional subscheme $A\subset L_1\cup L_2$ such that $\deg (A) \le 3$ and $\deg (A\cap L_i) \le 2$ for all $i$. Then we have $h^1({{\ensuremath{{\mathcal{I}}}}}_A(1,1)) =0$. Since $\deg (A) \le \deg (A\cap L_1)+\deg (A\cap L_2)$, there is $i\in \{1,2\}$ such that $\deg (L_i\cap A) =2$. Without loss of generality, let us assume that $i=1$. Since $\deg (L_1\cap A) =2$, we have $h^1({{\ensuremath{{\mathcal{I}}}}}_{A\cap L_1}(1,1)) =0$. Since $\deg (\mbox{Res}_{L_1}(A)) \le 1$, we have $h^1({{\ensuremath{{\mathcal{I}}}}}_{\mbox{Res}_{L_1}(A)}(0,1)) =0$ and then we can apply the sequence (\[eqa10\]). \[a2\] For $L_1\in \vert {{\ensuremath{{\mathcal{O}}}}}_Q(1,0)\vert$ and $L_2\in \vert {{\ensuremath{{\mathcal{O}}}}}_Q(0,1)\vert$, let us take a zero-dimensional subscheme $A\subset L_1\cup L_2$ such that $\deg (A\cap L_i) \le 2$ for all $i$. Then we have $h^1({{\ensuremath{{\mathcal{I}}}}}_A(2,2)) =0$. Since $\deg (A) \le \deg (A\cap L_1)+\deg (A\cap L_2)$, we have $\deg (A)\le 4$. Lemma \[a1\] gives the case of $\deg (A)\le 3$. Assume $\deg (A)=4$. We have $h^1({{\ensuremath{{\mathcal{I}}}}}_{\mbox{Res}_{L_1}}(1,2)) =0$ because $\mbox{Res}_{L_1}(A)\subset L_2$, $L_2\cong \mathbb {P}^1$ and $\deg (\mbox{Res}_{L_1}(A) \cap L_2) \le \deg (A\cap L_2) \le 2$. We then apply the sequence (\[eqa10\]). Let us take $Z\subset Q$ with $\deg (Z)=7$ and assume that ${{\ensuremath{{\mathcal{I}}}}}_Z(2,2)$ is globally generated. In particular, we have $\deg (Z\cap L) \le 2$ for all lines $L\subset Q$. Since $\deg (Z) < 8$, $Z$ is not a complete intersection and so $h^0({{\ensuremath{{\mathcal{I}}}}}_Z(2,2)) \ge 3$. It implies that $h^1({{\ensuremath{{\mathcal{I}}}}}_Z(2,2)) >0$. Let us take a curve $E\in \vert {{\ensuremath{{\mathcal{O}}}}}_Q(1,1)\vert$ such that $w:= \deg (A\cap E) $ is maximal. Assume for the moment that $E$ is irreducible. Since ${{\ensuremath{{\mathcal{I}}}}}_Z(2,2)$ is globally generated, we have $w \le 4$. If $E$ is reducible, say $E =L_1\cup L_2$, then the same is true, because $\deg (A\cap (L_1\cup L_2)) \le \deg (A\cap L_1)+\deg (A\cap L_2)$. In both cases we have $h^1(\mathcal {I}_{E\cap Z}(2,2)) =0$; if $E$ is reducible, this is Lemma \[a2\]. Hence the sequence (\[eqa10\]) gives $h^1({{\ensuremath{{\mathcal{I}}}}}_{\mbox{Res}_E(Z)}(1,1)) > 0$. \(a) Let us assume first that $w =4$ and so $\deg (\mbox{Res}_E(Z)) =3$. We have $\deg (\mbox{Res}_E(Z)\cap L) \le \deg (Z\cap L) \le 2$ for each line $L$. First assume the existence of a line $R_1$, say of type $(1,0)$, such that $\deg (R_1\cap \mbox{Res}_E(Z)) =2$. We have $h^1(R_1,{{\ensuremath{{\mathcal{I}}}}}_{R_1\cap \mbox{Res}_E(Z),R_1}(1,1)) =0$ and $\deg (\mbox{Res}_{R_1}(\mbox{Res}_E(Z)))=1$. Hence $h^1({{\ensuremath{{\mathcal{I}}}}}_{\mbox{Res}_{R_1}(\mbox{Res}_E(Z))}(0,1))=0$. We apply the sequence (\[eqa10\]) to get a contradiction. Now assume $\deg (\mbox{Res}_E(Z)\cap L) \le 1$ for each line. Since $h^0({{\ensuremath{{\mathcal{O}}}}}_Q(1,1)) =4$, there is a smooth conic $D$ such that $\deg (D\cap \mbox{Res}_E(Z)) \ge 2$. Since $\deg (\mbox{Res}_E(Z)) =3$, we have $\deg (\mbox{Res}_D(\mbox{Res}_E(Z))) \le 1$ and so $h^1({{\ensuremath{{\mathcal{I}}}}}_{\mbox{Res}_D(\mbox{Res}_E(Z))})=0$. Since $D$ is a smooth conic and $\deg (D\cap \mbox{Res}_E(Z)) \le 3$, we have $h^1(D,{{\ensuremath{{\mathcal{I}}}}}_{\mbox{Res}_E(Z),D}(1,1)) =0$, contradicting (\[eqa10\]). \(b) Now assume that $w=3$ and so $\deg (\mbox{Res}_E(Z)) =4$. Let us take $F\in \vert {{\ensuremath{{\mathcal{O}}}}}_Q(1,1)\vert$ such that $z:= \deg (F\cap \mbox{Res}_E(Z))$ is maximal. Since $3 \le z \le w =3$, we have $z=3$ and so $h^1(F,{{\ensuremath{{\mathcal{I}}}}}_{F\cap \mbox{Res}_E(Z),F}(1,1)) =0$, using Lemma \[a1\] if $F$ is reducible. From (\[eqa10\]), we get $h^1({{\ensuremath{{\mathcal{I}}}}}_{\mbox{Res}_F(\mbox{Res}_E(Z)})) >0$. But we have $\deg (\mbox{Res}_F(\mbox{Res}_E(Z)))=1$ and this is absurd. Higher Rank Case ================ Let ${{\ensuremath{{\mathcal{E}}}}}$ be a globally generated vector bundle of rank $r \ge3$ on $Q$ and it fits into the following sequence $$\label{heqa} 0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus (r-2)} {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{F}}}}}{\xrightarrow{\ \ }}0$$ where ${{\ensuremath{{\mathcal{F}}}}}$ is a globally generated vector bundle of rank $2$ on $Q$ with the Chern classes $c_i({{\ensuremath{{\mathcal{F}}}}})=c_i({{\ensuremath{{\mathcal{E}}}}})$ for $i=1,2$. Conversely, if ${{\ensuremath{{\mathcal{F}}}}}$ is a globally generated vector bundle of rank $2$, then any coherent sheaf ${{\ensuremath{{\mathcal{E}}}}}$ fitting into the sequence (\[heqa\]) is globally generated since $h^1({{\ensuremath{{\mathcal{O}}}}}_Q)=0$. Note that $h^0({{\ensuremath{{\mathcal{E}}}}})=h^0({{\ensuremath{{\mathcal{F}}}}})+r-2$. If ${{\ensuremath{{\mathcal{E}}}}}$ has no trivial summand, then we have $$\mathrm{rank} ({{\ensuremath{{\mathcal{E}}}}}) \le h^1({{\ensuremath{{\mathcal{F}}}}}^\vee)+2.$$ If $r:=\mathrm{rank}({{\ensuremath{{\mathcal{E}}}}})> h^1({{\ensuremath{{\mathcal{F}}}}}^\vee)+2$, then ${{\ensuremath{{\mathcal{E}}}}}$ is given by extension classes $e_1, \cdots, e_{r-2} \in H^1({{\ensuremath{{\mathcal{F}}}}}^\vee)$ and they are linearly dependent. By changing a basis of the trivial bundle ${{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus (r-2)}$, we reduce to the case of $e_{r-2}=0$ and so we have ${{\ensuremath{{\mathcal{E}}}}}\cong {{\ensuremath{{\mathcal{G}}}}}\oplus {{\ensuremath{{\mathcal{O}}}}}_Q$ for some vector bundle ${{\ensuremath{{\mathcal{G}}}}}$ of rank $r-1$. In particular, ${{\ensuremath{{\mathcal{E}}}}}$ has a trivial factor. There exists an indecomposable and globally generated vector bundle on $Q$ with $c_1=(2,2)$ only if the rank is at most $8$. From the classification of such vector bundles of rank $2$, we have $h^1({{\ensuremath{{\mathcal{F}}}}}^\vee) \le 6$ and thus the rank of ${{\ensuremath{{\mathcal{E}}}}}$ is at most $8$. The goal of this section is to describe the possible rank and Chern classes with which indecomposable vector bundles on $Q$ exist. The main result is the following: \[hr\] There exists a globally generated and indecomposable vector bundle of rank $r\ge 3$ with the Chern classes $(c_1, c_2)\in {{\ensuremath{\mathbb{Z}}}}^{\oplus 3}$ with $c_1=(a,b)\le (2,2)$ and $a\le b$ if and only if $$\begin{aligned} (c_1, c_2;r)\in \{ &(1,1,2;3), (1,2,3;3),(1,2,4;r=3,4,5),(2,2,4;3)\\ &(2,2,5;3),(2,2,6;r=3,4,5), (2,2,8;r=3,4,5,6,7,8) \}.\end{aligned}$$ Let ${{\ensuremath{{\mathcal{E}}}}}$ be a globally generated and indecomposable vector bundle of rank $r\ge 3$ with $c_1=(1,1)$. Then we have $${{\ensuremath{{\mathcal{E}}}}}\cong T{{\ensuremath{\mathbb{P}}}}^3(-1)_{|_Q}.$$ Note that $h^1({{\ensuremath{{\mathcal{O}}}}}_Q(-a,-b))=0$ for $a,b\in \{0,1\}$ and so the only possibility for ${{\ensuremath{{\mathcal{F}}}}}$ is ${{\ensuremath{{\mathcal{A}}}}}_P$ from Proposition \[1.1\]. Since $h^1({{\ensuremath{{\mathcal{A}}}}}_P^\vee)=h^1({\varphi}_P^*(\Omega_{{{\ensuremath{\mathbb{P}}}}^2}^1(1)))=1$, so there exists a non-trivial extension $$0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{B}}}}}_P {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{A}}}}}_P {\xrightarrow{\ \ }}0.$$ For any two points $P,O\in {{\ensuremath{\mathbb{P}}}}^3\setminus Q$, there exists an automorphism $g\in \mathrm{Aut}(Q)$ such that $g^*({{\ensuremath{{\mathcal{A}}}}}_P)\cong {{\ensuremath{{\mathcal{A}}}}}_O$ because $\mathrm{Aut}(Q)$ acts transitively on the set of all points of ${{\ensuremath{\mathbb{P}}}}^3\setminus Q$ and any pull-back of twisted tangent bundle is uniquely determined by its center of projection. Thus $g^*({{\ensuremath{{\mathcal{B}}}}}_P)\cong {{\ensuremath{{\mathcal{B}}}}}_O$. Conversely, since $T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q$ is globally generated, so it fits into an exact sequence $$0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q {\xrightarrow{\ \ }}T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{F}}}}}{\xrightarrow{\ \ }}0$$ for some globally generated vector bundle ${{\ensuremath{{\mathcal{F}}}}}$ of rank 2 with $c_1({{\ensuremath{{\mathcal{F}}}}})=(1,1)$. From the classification of such bundles, we have ${{\ensuremath{{\mathcal{F}}}}}\cong {{\ensuremath{{\mathcal{A}}}}}_P$ for some $P$. Thus we have $T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q \cong {{\ensuremath{{\mathcal{B}}}}}_P$ for each $P$. Let ${{\ensuremath{{\mathcal{E}}}}}$ be a globally generated and indecomposable vector bundle of rank $r\ge 3$ with $c_1=(1,2)$. Then ${{\ensuremath{{\mathcal{E}}}}}$ is of maximal type or has $c_2=3$ with $\mathrm{rank}({{\ensuremath{{\mathcal{E}}}}})=3$. It is obvious that the splitness of ${{\ensuremath{{\mathcal{F}}}}}$ in the sequence (\[heqa\]) implies the splitness of ${{\ensuremath{{\mathcal{E}}}}}$ in the case of $c_1=(1,2)$. If ${{\ensuremath{{\mathcal{F}}}}}$ admits the sequence (1) in Proposition \[prop1.2\], then we have $h^1({{\ensuremath{{\mathcal{F}}}}}^\vee)=1$, i.e. there exists a unique non-trivial extension of ${{\ensuremath{{\mathcal{F}}}}}$ by ${{\ensuremath{{\mathcal{O}}}}}_Q$. It is clearly ${{\ensuremath{{\mathcal{O}}}}}_Q(1,0)\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(0,1)^{\oplus 2}$ and, in particular, ${{\ensuremath{{\mathcal{E}}}}}$ splits. If ${{\ensuremath{{\mathcal{F}}}}}$ is of maximal type, then we have $h^1({{\ensuremath{{\mathcal{F}}}}}^\vee)=3$ and thus there exist indecomposable vector bundles of rank up to $5$ with $c_2=4$. Let us assume that ${{\ensuremath{{\mathcal{F}}}}}$ admits the sequence (2) in Proposition \[prop1.2\], i.e. $c_2({{\ensuremath{{\mathcal{F}}}}})=3$. Since $h^1({{\ensuremath{{\mathcal{F}}}}}^\vee)=2$, so the rank of ${{\ensuremath{{\mathcal{E}}}}}$ must be at most $4$. Let us assume that the rank of ${{\ensuremath{{\mathcal{E}}}}}$ is $4$, i.e. we have $$0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus 2} {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{F}}}}}{\xrightarrow{\ \ }}0.$$ From the sequence (2), we have $h^0({{\ensuremath{{\mathcal{F}}}}}(-1))=0$ and $h^1({{\ensuremath{{\mathcal{F}}}}}(-2))=2$. Thus we have $h^0({{\ensuremath{{\mathcal{E}}}}}(-1))=h^1({{\ensuremath{{\mathcal{E}}}}}(-2))=0$. By Theorem 6.7 in [@AO] with $S''={{\ensuremath{{\mathcal{O}}}}}_Q(0,-1)$ and $t=0$, we have $h^1({{\ensuremath{{\mathcal{E}}}}}(-2,-1))=h^1({{\ensuremath{{\mathcal{F}}}}}(-2,-1))=1$ and so ${{\ensuremath{{\mathcal{E}}}}}$ has ${{\ensuremath{{\mathcal{O}}}}}_Q(0,1)$ as its direct summand. Thus we have ${{\ensuremath{{\mathcal{E}}}}}\cong {{\ensuremath{{\mathcal{O}}}}}_Q(0,1)\oplus {{\ensuremath{{\mathcal{G}}}}}$ where ${{\ensuremath{{\mathcal{G}}}}}$ is a globally generated vector bundle of rank $3$ with $c_1=(1,1)$ and $c_2=2$. In particular, it is decomposable, a contradiction. Let us assume $\mathrm{rank}({{\ensuremath{{\mathcal{E}}}}})=3$ and so it is an extension of ${{\ensuremath{{\mathcal{F}}}}}$ by ${{\ensuremath{{\mathcal{O}}}}}_Q$. In particular, we have $h^1({{\ensuremath{{\mathcal{E}}}}}^\vee)=1$ and so there exists a non-trivial extension ${{\ensuremath{{\mathcal{G}}}}}$ of ${{\ensuremath{{\mathcal{E}}}}}$ by ${{\ensuremath{{\mathcal{O}}}}}_Q$. Since ${{\ensuremath{{\mathcal{G}}}}}$ is a globally generated vector bundle of rank $4$ with $c_1=(1,2)$ and $c_2=3$, so we have ${{\ensuremath{{\mathcal{G}}}}}\cong {{\ensuremath{{\mathcal{O}}}}}_Q(0,1)\oplus {{\ensuremath{{\mathcal{U}}}}}$ with either ${{\ensuremath{{\mathcal{U}}}}}\cong T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q$ or ${{\ensuremath{{\mathcal{U}}}}}$ has a trivial factor; the latter case cannot occur because ${{\ensuremath{{\mathcal{G}}}}}$ has no trivial factor. Note that $h^0({{\ensuremath{{\mathcal{E}}}}}(0,-1))=h^0({{\ensuremath{{\mathcal{F}}}}}(0,-1))=1$. Let us assume that ${{\ensuremath{{\mathcal{E}}}}}$ is decomposable and then ${{\ensuremath{{\mathcal{O}}}}}_Q(0,1)$ must be a direct factor, i.e. we have ${{\ensuremath{{\mathcal{E}}}}}\cong {{\ensuremath{{\mathcal{O}}}}}_Q(0,1)\oplus {{\ensuremath{{\mathcal{G}}}}}$ where ${{\ensuremath{{\mathcal{G}}}}}$ is globally generated with $c_1=(1,1)$ and $c_2=2$. From the classification, we have ${{\ensuremath{{\mathcal{G}}}}}\cong {{\ensuremath{{\mathcal{A}}}}}_P$ for some $P\in {{\ensuremath{\mathbb{P}}}}^3\setminus Q$. Let us take a general section of ${{\ensuremath{{\mathcal{O}}}}}_Q(0,1)\oplus T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q$. Hence we get an exact sequence $$\label{eqa1} 0 {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q\stackrel{u}{{\xrightarrow{\ \ }}} {{\ensuremath{{\mathcal{O}}}}}_Q(0,1)\oplus T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q \stackrel{\psi}{{\xrightarrow{\ \ }}} {{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}0$$ Assume that ${{\ensuremath{{\mathcal{E}}}}}$ is decomposable, i.e. ${{\ensuremath{{\mathcal{E}}}}}\cong {{\ensuremath{{\mathcal{O}}}}}_Q(0,1) \oplus {{\ensuremath{{\mathcal{A}}}}}_P$ for some $P$. Hence $\psi$ is given by a $2\times 2$ matrix of maps. Since $h^0({{\ensuremath{{\mathcal{A}}}}}_P(0,-1)) = h^0({\Omega _{{{\ensuremath{\mathbb{P}}}}^3}}|_Q(1,0)) = 0$ , using the exact sequence $$0 {\xrightarrow{\ \ }}{\Omega _{{{\ensuremath{\mathbb{P}}}}^3}}|_Q (1,0) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(0,-1)^{\oplus 4} {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(1,0){\xrightarrow{\ \ }}0,$$ $\psi$ is the diagonal matrix associated to $\psi _1: {{\ensuremath{{\mathcal{O}}}}}_Q(0,1){\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(0,1)$ and $\psi _2: T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{A}}}}}_P$. $\psi _1$ must be the multiplication by a non-zero-constant, while $\psi _2$ must be surjective and have ${{\ensuremath{{\mathcal{O}}}}}_Q$ as its quotient. Hence $\mbox{ker}(\psi)$ is contained in the factor $\{0\}\oplus T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q$, not just isomorphic to a subsheaf of that factor and so $\mbox{Im}(u)$ is contained in the factor $\{0\}\oplus T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q$ , again not just isomorphic to a subsheaf of that factor. This is not true for a general $u$ and thus there exists an indecomposable and globally generated vector bundle of rank $3$ with $c_1=(1,2)$ and $c_2=3$. Let us assume that ${{\ensuremath{{\mathcal{E}}}}}$ has the first Chern class $c_1=(2,2)$. If the associated vector bundle ${{\ensuremath{{\mathcal{F}}}}}$ of rank $2$ splits, then we have $h^1({{\ensuremath{{\mathcal{F}}}}}^\vee)=0$ except when ${{\ensuremath{{\mathcal{F}}}}}\cong {{\ensuremath{{\mathcal{O}}}}}_Q(2,0)\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(0,2)$. There exists an indecomposable extension ${{\ensuremath{{\mathcal{E}}}}}$ of ${{\ensuremath{{\mathcal{O}}}}}_Q(2,0)\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(0,2)$ by trivial factors if and only if $\mathrm{rank}({{\ensuremath{{\mathcal{E}}}}})=3$. Let ${{\ensuremath{{\mathcal{F}}}}}$ be ${{\ensuremath{{\mathcal{O}}}}}_Q(2,0)\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(0,2)$. Since we have $h^1({{\ensuremath{{\mathcal{F}}}}}^\vee)=2$ for a bundle ${{\ensuremath{{\mathcal{F}}}}}$ in the sequence (\[heqa\]), so ${{\ensuremath{{\mathcal{E}}}}}$ has a trivial factor if $\mathrm{rank}({{\ensuremath{{\mathcal{E}}}}})\ge 5$. If ${{\ensuremath{{\mathcal{E}}}}}$ is a vector bundle arising in a non-trivial extension $$0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus 2} {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{F}}}}}{\xrightarrow{\ \ }}0,$$ then it is ACM, since the map $$H^1({{\ensuremath{{\mathcal{O}}}}}_Q(0,-2)\oplus{{\ensuremath{{\mathcal{O}}}}}_Q(-2,0))\rightarrow H^2({{\ensuremath{{\mathcal{O}}}}}_Q(-2,-2)^{\oplus 2})$$ is an isomorphism. Now by the Chern class computation we have ${{\ensuremath{{\mathcal{E}}}}}\cong {{\ensuremath{{\mathcal{O}}}}}_Q(1,0)^{\oplus 2}\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(0,1)^{\oplus 2}$. In particular, it is decomposable. Assume that ${{\ensuremath{{\mathcal{E}}}}}$ is a general extension of ${{\ensuremath{{\mathcal{F}}}}}$ by ${{\ensuremath{{\mathcal{O}}}}}_Q$. Because of the generality, ${{\ensuremath{{\mathcal{E}}}}}$ cannot be ${{\ensuremath{{\mathcal{O}}}}}_Q(0,1)^{\oplus 2} \oplus {{\ensuremath{{\mathcal{O}}}}}_Q(2,0)$ nor ${{\ensuremath{{\mathcal{O}}}}}_Q(1,0)^{\oplus 2} \oplus {{\ensuremath{{\mathcal{O}}}}}_Q(0,2)$. Assume that ${{\ensuremath{{\mathcal{E}}}}}$ is decomposable. Since $(1,1)$ is not an index of ${{\ensuremath{{\mathcal{E}}}}}$, so ${{\ensuremath{{\mathcal{E}}}}}$ cannot be isomorphic to ${{\ensuremath{{\mathcal{O}}}}}_Q(1,1)\oplus {{\ensuremath{{\mathcal{A}}}}}_P$. Thus the only possibility is either ${{\ensuremath{{\mathcal{O}}}}}_Q(1,0)\oplus {{\ensuremath{{\mathcal{G}}}}}$ where ${{\ensuremath{{\mathcal{G}}}}}$ fits into the sequence (2) in Proposition \[prop1.2\], or ${{\ensuremath{{\mathcal{O}}}}}_Q(0,1)\oplus {{\ensuremath{{\mathcal{G}}}}}'$. In the former case, we have $h^0({{\ensuremath{{\mathcal{G}}}}}(0,-1))=1$ and so $h^0({{\ensuremath{{\mathcal{E}}}}}(0,-1))=1$, which is contradicting to $h^0({{\ensuremath{{\mathcal{E}}}}}(0,-1))=2$. The latter case is also impossible similarly. Thus ${{\ensuremath{{\mathcal{E}}}}}$ is indecomposable. If ${{\ensuremath{{\mathcal{F}}}}}$ is a non trivial extension of ${{\ensuremath{{\mathcal{O}}}}}_Q(2,0)$ and ${{\ensuremath{{\mathcal{O}}}}}_Q(0,2)$ given in lemma \[nost\], we get $h^1({{\ensuremath{{\mathcal{F}}}}}(-2,-2))=h^1({{\ensuremath{{\mathcal{O}}}}}_Q(0,-2)\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(-2,0))=2$ and $h^0({{\ensuremath{{\mathcal{F}}}}}(0,-1))=h^0({{\ensuremath{{\mathcal{O}}}}}_Q(2,-1)\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(0,1))=2$. So the proof of the above lemma holds also in this case and we can conclude that there exists an indecomposable extension ${{\ensuremath{{\mathcal{E}}}}}$ of ${{\ensuremath{{\mathcal{F}}}}}$ by trivial factors if and only if $\mathrm{rank}({{\ensuremath{{\mathcal{E}}}}})=3$. Now assume that ${{\ensuremath{{\mathcal{F}}}}}$ does not split with $h^1({{\ensuremath{{\mathcal{F}}}}}^\vee)>0$. Firstly let us deal with the case when ${{\ensuremath{{\mathcal{F}}}}}$ is unstable. There exists no indecomposable vector bundle ${{\ensuremath{{\mathcal{E}}}}}$ of rank at least $3$, whose associated bundle ${{\ensuremath{{\mathcal{F}}}}}$ is non-splitting and unstable. Since ${{\ensuremath{{\mathcal{F}}}}}$ has $h^1({{\ensuremath{{\mathcal{F}}}}}^\vee )>0$, so it is one of the following: 1. $0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(1,0)\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(0,1)\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(1,1) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{F}}}}}{\xrightarrow{\ \ }}0$ 2. $0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(-1,-1){\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus 2} \oplus {{\ensuremath{{\mathcal{O}}}}}_Q(1,1) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{F}}}}}{\xrightarrow{\ \ }}0$. Since $h^1({{\ensuremath{{\mathcal{F}}}}}^\vee)=1$, so the only non-trivial extension of ${{\ensuremath{{\mathcal{F}}}}}$ by a single ${{\ensuremath{{\mathcal{O}}}}}_Q$ cannot have a trivial factor. But it must be ${{\ensuremath{{\mathcal{O}}}}}_Q(1,0)\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(0,1) \oplus {{\ensuremath{{\mathcal{O}}}}}_Q(1,1)$ from the sequence. In particular, it is decomposable. Let us assume that the associated vector bundle of rank $2$ is from the sequence (2). We have $h^1({{\ensuremath{{\mathcal{F}}}}}^\vee)=2$ and the vector bundle ${{\ensuremath{{\mathcal{E}}}}}$ has no trivial factor only if $\mathrm{rank}({{\ensuremath{{\mathcal{E}}}}})\le 4$. Note that $h^0({{\ensuremath{{\mathcal{E}}}}}(-2,-2))=h^0({{\ensuremath{{\mathcal{F}}}}}(-2,-2))=0$ and $h^1({{\ensuremath{{\mathcal{E}}}}}(-3,-2))=h^1({{\ensuremath{{\mathcal{F}}}}}(-3,-2))=0$. Similarly we have $h^1({{\ensuremath{{\mathcal{E}}}}}(-2,-3))=0$. Using Theorem 6.1 in [@AO] with $t=-1$, ${{\ensuremath{{\mathcal{E}}}}}$ has ${{\ensuremath{{\mathcal{O}}}}}_Q(1,1)$ as its direct summand. In particular, ${{\ensuremath{{\mathcal{E}}}}}$ is decomposable. Now let us assume that ${{\ensuremath{{\mathcal{F}}}}}$ is stable and then we have $4\le c_2({{\ensuremath{{\mathcal{F}}}}}) \le 8$. If $c_2({{\ensuremath{{\mathcal{F}}}}})=8$, i.e. ${{\ensuremath{{\mathcal{F}}}}}$ is of maximal type with $$0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(-2,-2) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus 3} {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{F}}}}}{\xrightarrow{\ \ }}0,$$ we have $h^1({{\ensuremath{{\mathcal{F}}}}}^\vee) = 6$. Thus there exists an indecomposable globally generated vector bundle of maximal type for each rank $3\le r\le 8$ by Lemma \[mi\]. The remaining cases are when ${{\ensuremath{{\mathcal{F}}}}}$ is from $\mathfrak{M}(c_2)$ with $4\le c_2 \le 6$. The extension of ${{\ensuremath{{\mathcal{F}}}}}\in \mathfrak{M}(4)$ by ${{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus (r-2)}$ can be indecomposable if and only if $r=3$. Let ${{\ensuremath{{\mathcal{F}}}}}\in \mathfrak{M}(4)$ have an index $(1,0)$. Then it fits into the exact sequence $$0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(1,0) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{F}}}}}{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{I}}}}}_Z(1,2) {\xrightarrow{\ \ }}0,$$ where $\deg (Z)=2$. Note that $h^1({{\ensuremath{{\mathcal{F}}}}}^\vee)=h^1({{\ensuremath{{\mathcal{F}}}}}(-2,-2))=h^1({{\ensuremath{{\mathcal{I}}}}}_Z(-1,0))=2$. We have $h^1({{\ensuremath{{\mathcal{F}}}}}^\vee ) = h^1({{\ensuremath{{\mathcal{F}}}}}(-2,-2)) = h^1({{\ensuremath{{\mathcal{I}}}}}_Z(-1,0)) = \deg (Z) =2$. Hence for each $r\ge 5$ and extension of ${{\ensuremath{{\mathcal{F}}}}}$ by ${{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus (r-2)}$ has a trivial factor. Hence it is sufficient to check the cases $r=3$ and $r=4$. Let us take any extension ${{\ensuremath{{\mathcal{E}}}}}$ of ${{\ensuremath{{\mathcal{F}}}}}$ by ${{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus 2}$ $$0 {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus 2} {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{F}}}}}{\xrightarrow{\ \ }}0.$$ There is a non-zero map ${\varphi}: {{\ensuremath{{\mathcal{O}}}}}_Q(1,0)^{\oplus 2} {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}$. Set ${{\ensuremath{{\mathcal{G}}}}}:= \mathrm{Im}({\varphi})$ and $c:= \mathrm{rank} ({\varphi})$. First assume $c=1$. Since ${{\ensuremath{{\mathcal{G}}}}}$ is globally generated, torsion free, a quotient of ${{\ensuremath{{\mathcal{O}}}}}_Q(1,0)$ and $h^0({{\ensuremath{{\mathcal{G}}}}}^{\vee \vee })(-2,0)) \le h^0({{\ensuremath{{\mathcal{E}}}}}(-2,0)) = 0$, we get ${{\ensuremath{{\mathcal{G}}}}}= {{\ensuremath{{\mathcal{O}}}}}_Q(1,0)$; this is the case leading to ${{\ensuremath{{\mathcal{W}}}}}$, the indecomposable bundle of rank $3$ below, if it exists. Now assume $c=2$. Hence ${\varphi}$ is injective and so ${{\ensuremath{{\mathcal{G}}}}}\cong {{\ensuremath{{\mathcal{O}}}}}_Q(1,0)^{\oplus 2}$. Let ${{\ensuremath{{\mathcal{C}}}}}$ be the saturation of ${{\ensuremath{{\mathcal{G}}}}}$ in ${{\ensuremath{{\mathcal{E}}}}}$, i.e. the only rank two subsheaf of ${{\ensuremath{{\mathcal{E}}}}}$ containing ${{\ensuremath{{\mathcal{G}}}}}$ and then ${{\ensuremath{{\mathcal{D}}}}}:= {{\ensuremath{{\mathcal{E}}}}}/{{\ensuremath{{\mathcal{C}}}}}$ is a torsion free and globally generated sheaf of rank $2$. We have $c_1({{\ensuremath{{\mathcal{C}}}}})=(a,b)$ for some $a\ge 2$ and $b\ge 0$. Since ${{\ensuremath{{\mathcal{D}}}}}$ is globally generated, we get $a=2$ and $b\le 2$. First assume $b=0$. Since ${{\ensuremath{{\mathcal{G}}}}}\subseteq {{\ensuremath{{\mathcal{C}}}}}$ and ${{\ensuremath{{\mathcal{G}}}}}$ and ${{\ensuremath{{\mathcal{C}}}}}$ are vector bundles with the same rank and the same determinant, then ${{\ensuremath{{\mathcal{C}}}}}= {{\ensuremath{{\mathcal{G}}}}}$. Since $h^i({{\ensuremath{{\mathcal{G}}}}}(0,-1)) = 0$, $i=0,1$, and $h^1({{\ensuremath{{\mathcal{G}}}}}(0,-2))=0$ we get $h^0({{\ensuremath{{\mathcal{D}}}}}(0,-1)) = h^0({{\ensuremath{{\mathcal{E}}}}}(0,-1))=2$ and $h^0({{\ensuremath{{\mathcal{D}}}}}(0,-2))=0$. We get ${{\ensuremath{{\mathcal{D}}}}}\cong {{\ensuremath{{\mathcal{O}}}}}_Q(0,1)^{\oplus 2}$ by the Chern class counting or $h^0({{\ensuremath{{\mathcal{D}}}}})=4$ and many other reasons. Thus we have ${{\ensuremath{{\mathcal{E}}}}}\cong {{\ensuremath{{\mathcal{O}}}}}_Q(1,0)^{\oplus 2}\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(0,1)^{\oplus 2}$. Now assume $b>0$. Note that ${{\ensuremath{{\mathcal{D}}}}}$ is globally generated , $c_1({{\ensuremath{{\mathcal{D}}}}})=(0,2-b)$ and $\mathrm{rank}({{\ensuremath{{\mathcal{D}}}}})=2$. Even if ${{\ensuremath{{\mathcal{D}}}}}$ is only torsion free, we have an exact sequence $$0 {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{D}}}}}{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{I}}}}}_A(0,1){\xrightarrow{\ \ }}0$$ where ${{\ensuremath{{\mathcal{I}}}}}_A(0,1)$ is globally generated and thus we have $A=\emptyset$ and ${{\ensuremath{{\mathcal{D}}}}}\cong {{\ensuremath{{\mathcal{O}}}}}_Q\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(0,1)$. Since ${{\ensuremath{{\mathcal{D}}}}}$ is a quotient of ${{\ensuremath{{\mathcal{E}}}}}$, Lemma \[trivial\] that ${{\ensuremath{{\mathcal{E}}}}}$ has a trivial factor, a contradiction. Hence there is no indecomposable extension of ${{\ensuremath{{\mathcal{F}}}}}$ by ${{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus 2}$. Now we prove the existence of the indecomposable extension of ${{\ensuremath{{\mathcal{F}}}}}$ by ${{\ensuremath{{\mathcal{O}}}}}_Q$. Let us define $V: = H^0({{\ensuremath{{\mathcal{O}}}}}_Q(1,0)^{\oplus 2}\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(0,1)^{\oplus 2})$. For a generally fixed ${\varphi}: {{\ensuremath{{\mathcal{O}}}}}_Q{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(1,0)^{\oplus 2}\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(0,1)^{\oplus 2}$, let us set ${{\ensuremath{{\mathcal{W}}}}}:= \mathrm{coker} ({\varphi})$. Since ${{\ensuremath{{\mathcal{O}}}}}_Q(1,0)^{\oplus 2}\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(0,1)^{\oplus 2}$ is globally generated, ${{\ensuremath{{\mathcal{W}}}}}$ is a globally generated vector bundle of rank $3$ with $c_1=(2,2)$ and $c_2=4$. We claim that ${{\ensuremath{{\mathcal{W}}}}}$ is indecomposable. Let $\sigma : Q{\xrightarrow{\ \ }}Q$ be the automorphism which exchanges the two rulings of $Q$. Since $\sigma ^\ast ({{\ensuremath{{\mathcal{O}}}}}_Q(1,0)^{\oplus 2}\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(0,1)^{\oplus 2}) \cong {{\ensuremath{{\mathcal{O}}}}}_Q(1,0)^{\oplus 2}\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(0,1)^{\oplus 2}$ and ${\varphi}$ is general, the number of factors ${{\ensuremath{{\mathcal{O}}}}}_Q(1,0)$ of ${{\ensuremath{{\mathcal{W}}}}}$ is equal to the number of factors ${{\ensuremath{{\mathcal{O}}}}}_Q(0,1)$ of ${{\ensuremath{{\mathcal{W}}}}}$. ${{\ensuremath{{\mathcal{O}}}}}_Q(1,1)$ is not a factor of ${{\ensuremath{{\mathcal{W}}}}}$, because $(1,1)$ is not an index of ${{\ensuremath{{\mathcal{F}}}}}$. Thus we have ${{\ensuremath{{\mathcal{W}}}}}\cong {{\ensuremath{{\mathcal{O}}}}}_Q(1,0)^{\oplus a}\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(0,1)^{\oplus a}$ with $2a =3$, a contradiction. The proof gives a nice parameter space for the rank $3$ bundles, i.e. the projectivisation of the set of all nowhere vanishing sections of $H^0( {{\ensuremath{{\mathcal{O}}}}}_Q(1,0)^{\oplus 2}\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(0,1)^{\oplus 2})$. Let $\mathfrak{G}_r$ with $r=3,4,5$ be the set of all vector bundles without trivial factors and which are an extension of some ${{\ensuremath{{\mathcal{F}}}}}\in \mathfrak{M}(5)$ by ${{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus (r-2)}$. Since $h^1({{\ensuremath{{\mathcal{F}}}}}^\vee ) =3$ for all ${{\ensuremath{{\mathcal{F}}}}}\in \mathfrak {M}(5)$, each set $\mathfrak{G}_r$ is non-empty and parametrized by an irreducible variety. \[ss1\] Fix any $P\in {{\ensuremath{\mathbb{P}}}}^3\setminus Q$ and look at the exact sequence $$0 {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(-1) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus 3}{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{A}}}}}_P{\xrightarrow{\ \ }}0$$ induced by the Euler sequence of $T{{\ensuremath{\mathbb{P}}}}^2$. We get $h^1({{\ensuremath{{\mathcal{A}}}}}_P^{\vee}(1,0)) =h^1({{\ensuremath{{\mathcal{A}}}}}_P (0,-1))=0$. Hence any extension of ${{\ensuremath{{\mathcal{A}}}}}_P$ by ${{\ensuremath{{\mathcal{O}}}}}_Q(1,0)$ is trivial. For the same reason any extension of ${{\ensuremath{{\mathcal{A}}}}}_P$ by ${{\ensuremath{{\mathcal{O}}}}}_Q(0,1)$ is trivial. \[5.12\] $ $ 1. A general bundle ${{\ensuremath{{\mathcal{E}}}}}\in \mathfrak{G}_3$ is stable and thus indecomposable. 2. $\mathfrak{G}_4 = \{{{\ensuremath{{\mathcal{O}}}}}_Q(1,0)\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(0,1)\oplus {{\ensuremath{{\mathcal{A}}}}}_P\}_{P\in {{\ensuremath{\mathbb{P}}}}^3\setminus Q}$. 3. $\mathfrak{G}_5 = \{{{\ensuremath{{\mathcal{O}}}}}_Q(1,0)\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(0,1)\oplus T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q \}$. \(a) Let us fix a general bundle ${{\ensuremath{{\mathcal{E}}}}}\in \mathfrak{G}_3$. We claim that ${{\ensuremath{{\mathcal{E}}}}}$ is indecomposable. Assume that this is not the case and so ${{\ensuremath{{\mathcal{E}}}}}$ has at least one line bundle, say ${{\ensuremath{{\mathcal{L}}}}}$, as a factor. Since $(1,1)$ is not an index of ${{\ensuremath{{\mathcal{E}}}}}$ and ${{\ensuremath{{\mathcal{E}}}}}$ has no trivial factor, so we have either ${{\ensuremath{{\mathcal{L}}}}}\cong {{\ensuremath{{\mathcal{O}}}}}_Q(1,0)$ or ${{\ensuremath{{\mathcal{L}}}}}\cong {{\ensuremath{{\mathcal{O}}}}}_Q(0,1)$. Let us take $(a,b) \in \{(1,0),(0,1)\}$ such that ${{\ensuremath{{\mathcal{L}}}}}\cong {{\ensuremath{{\mathcal{O}}}}}_Q(a,b)$. Since $\mathfrak{M}(5)$ is invariant by the automorphism of $Q$ which exchanges the two rulings of $Q$ (see Remark \[re44\]), $\mathfrak{G}_3$ is irreducible and ${{\ensuremath{{\mathcal{E}}}}}$ is irreducible, ${{\ensuremath{{\mathcal{O}}}}}_Q(b,a)$ is another factor of ${{\ensuremath{{\mathcal{E}}}}}$. Thus we should have ${{\ensuremath{{\mathcal{E}}}}}\cong {{\ensuremath{{\mathcal{O}}}}}_Q(1,0)\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(0,1) \oplus {{\ensuremath{{\mathcal{O}}}}}_Q(1,1)$, but this is impossible since $(1,1)$ is not an index of ${{\ensuremath{{\mathcal{E}}}}}$. Hence ${{\ensuremath{{\mathcal{E}}}}}$ is indecomposable. Since $(1,1)$ is not an index of ${{\ensuremath{{\mathcal{E}}}}}$, every rank $1$ subsheaf of ${{\ensuremath{{\mathcal{E}}}}}$ has ${{\ensuremath{{\mathcal{O}}}}}_Q(1,1)$-slope less than $4/3$. If ${{\ensuremath{{\mathcal{E}}}}}$ is not stable, we can take a rank $2$ saturated subsheaf ${{\ensuremath{{\mathcal{G}}}}}$ of ${{\ensuremath{{\mathcal{E}}}}}$ with maximal ${{\ensuremath{{\mathcal{O}}}}}_Q(1,1)$-slope $\alpha$ with $\alpha \ge 4/3$. Then the rank $1$ torsion free sheaf ${{\ensuremath{{\mathcal{E}}}}}/{{\ensuremath{{\mathcal{G}}}}}$ is globally generated and it has ${{\ensuremath{{\mathcal{O}}}}}_Q(1,1)$-slope at most $4/3$. Thus we have $({{\ensuremath{{\mathcal{E}}}}}/{{\ensuremath{{\mathcal{G}}}}})^{\vee \vee} \cong {{\ensuremath{{\mathcal{O}}}}}_Q(c,d)$ with $(c,d)\in \{(0,0),(1,0),(0,1)\}$. Hence ${{\ensuremath{{\mathcal{E}}}}}/{{\ensuremath{{\mathcal{G}}}}}\cong {{\ensuremath{{\mathcal{I}}}}}_E(c,d)$ for some zero-dimensional scheme $E\subset Q$. Since ${{\ensuremath{{\mathcal{O}}}}}_Q$ is not a direct factor of ${{\ensuremath{{\mathcal{E}}}}}$, we have $(c,d) \ne (0,0)$. We also have $E = \emptyset$ since ${{\ensuremath{{\mathcal{I}}}}}_E(c,d)$ is globally generated where $(c,d)\in \{(1,0),(0,1)\}$. Hence ${{\ensuremath{{\mathcal{O}}}}}_Q(c,d)$ is a quotient of ${{\ensuremath{{\mathcal{E}}}}}$. Since $(c,d)$ is an index of ${{\ensuremath{{\mathcal{E}}}}}$, we easily get that ${{\ensuremath{{\mathcal{O}}}}}_Q(c,d)$ is a factor of ${{\ensuremath{{\mathcal{E}}}}}$, a contradiction. \(b) Fix a general ${{\ensuremath{{\mathcal{E}}}}}\in \mathfrak{G}_4$. Since $(1,0)$ (resp. $(0,1)$) is an index of ${{\ensuremath{{\mathcal{E}}}}}$, there is a non-zero map $j_1: {{\ensuremath{{\mathcal{O}}}}}_Q(1,0) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}$ (resp. $j_2: {{\ensuremath{{\mathcal{O}}}}}_Q(0,1) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}$) and it is an injective map of sheaves. Consider the map $u =(j_1,j_2): {{\ensuremath{{\mathcal{O}}}}}_Q(1,0)\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(0,1) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}$. We first claim that $u$ has not rank $1$, i.e. that it is injective. Assume that $\mathrm{Im}(u)$ has rank $1$ and call ${{\ensuremath{{\mathcal{R}}}}}$ its saturation in ${{\ensuremath{{\mathcal{E}}}}}$. Since $Q$ is a smooth surface, ${{\ensuremath{{\mathcal{R}}}}}$ is a line bundle, say ${{\ensuremath{{\mathcal{R}}}}}\cong {{\ensuremath{{\mathcal{O}}}}}_Q(a,b)$. Since neither $j_1$ nor $j_2$ is zero, we have $a\ge 1$ and $b\ge 1$ and so ${{\ensuremath{{\mathcal{E}}}}}$ has $(1,1)$ as an index, a contradiction. Since $u$ is injective, the sheaf $\mathrm{Im}(u)$ is isomorphic to ${{\ensuremath{{\mathcal{O}}}}}_Q(1,0)\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(0,1)$. We now check that $\mathrm{Im}(u)$ is saturated in ${{\ensuremath{{\mathcal{E}}}}}$, i.e. that ${{\ensuremath{{\mathcal{U}}}}}:=\mathrm{coker}(u)$ is torsion free. Assume that $\mathrm{Im}(u)$ is not saturated in ${{\ensuremath{{\mathcal{E}}}}}$ and call ${{\ensuremath{{\mathcal{G}}}}}$ its saturation. Since $Q$ is a smooth surface, ${{\ensuremath{{\mathcal{G}}}}}$ is a vector bundle of rank $2$. Set $(a,b):= c_1({{\ensuremath{{\mathcal{G}}}}})$. Since ${{\ensuremath{{\mathcal{E}}}}}$ is a general element of the irreducible family $\mathfrak{G}_4$ and the family $\mathfrak{G}_4$ is invariant under the automorphism of $Q$ which permutes the two rulings of $Q$, we have $a=b$. Since $\mathrm{Im} (u) \nsubseteq {{\ensuremath{{\mathcal{G}}}}}$ and these two bundles have the same rank, we get $a>1$. The rank $2$ torsion free sheaf ${{\ensuremath{{\mathcal{E}}}}}/{{\ensuremath{{\mathcal{G}}}}}$ is spanned and $c_1({{\ensuremath{{\mathcal{E}}}}}/{{\ensuremath{{\mathcal{G}}}}})(2-a,2-a)$. We get $a=2$ and that ${{\ensuremath{{\mathcal{E}}}}}/{{\ensuremath{{\mathcal{G}}}}}$ is trivial. Lemma \[trivial\] gives that ${{\ensuremath{{\mathcal{E}}}}}$ has a trivial factor, a contradiction. Hence $\mathrm{Im} (u)$ is saturated and so ${{\ensuremath{{\mathcal{U}}}}}$ is torsion free. Fix $P\in {{\ensuremath{\mathbb{P}}}}^3 \setminus Q$. For the bundle ${{\ensuremath{{\mathcal{O}}}}}_Q(1,0)\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(0,1)\oplus {{\ensuremath{{\mathcal{A}}}}}_P$, the maps $j_1, j_2, u$ are uniquely defined and $\mathrm{coker} (u)$ is locally free. Thus ${{\ensuremath{{\mathcal{U}}}}}$ is locally free for a general ${{\ensuremath{{\mathcal{E}}}}}\in \mathfrak{G}_4$. The bundle ${{\ensuremath{{\mathcal{U}}}}}$ is a globally generated vector bundle of rank $2$ with $c_1({{\ensuremath{{\mathcal{U}}}}})=(1,1)$, no trivial factor and $h^0({{\ensuremath{{\mathcal{U}}}}})=h^0({{\ensuremath{{\mathcal{E}}}}})-4 =3$. We have ${{\ensuremath{{\mathcal{U}}}}}\cong {{\ensuremath{{\mathcal{A}}}}}_P$ for some $P\in {{\ensuremath{\mathbb{P}}}}^3\setminus Q$ by Proposition \[1.1\]. By Remark \[ss1\] a general ${{\ensuremath{{\mathcal{D}}}}}\in \mathfrak {G}_4$ is isomorphic to one of the bundles ${{\ensuremath{{\mathcal{O}}}}}_Q(1,0)\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(0,1)\oplus {{\ensuremath{{\mathcal{A}}}}}_P$. Let us fix any ${{\ensuremath{{\mathcal{E}}}}}\in \mathfrak {G}_4$ and then ${{\ensuremath{{\mathcal{E}}}}}$ is a degeneration of the family $\{{{\ensuremath{{\mathcal{O}}}}}_Q(1,0)\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(0,1)\oplus {{\ensuremath{{\mathcal{A}}}}}_P\}_{P\in {{\ensuremath{\mathbb{P}}}}^3\setminus Q}$. By semicontinuity we have $h^0({{\ensuremath{{\mathcal{E}}}}}^\vee (1,0) ) >0$ and $h^0({{\ensuremath{{\mathcal{E}}}}}^\vee (0,1) )>0$. Hence there is a map $f: {{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(1,0)\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(1,0)$ with $\mathrm{Im}(f)$ of rank $2$. Since ${{\ensuremath{{\mathcal{E}}}}}$ has no trivial factor, $\mathrm{Im} (f)$ has no trivial factor by Lemma \[trivial\]. Since $\mathrm{Im}(f)$ is globally generated, we get that $f$ is surjective and so $\ker (f)$ is a rank $2$ vector bundle. Since each ${{\ensuremath{{\mathcal{A}}}}}_P$ is a quotient of $T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q$, the semicontinuity theorem gives $h^0(T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q \otimes {{\ensuremath{{\mathcal{E}}}}}^\vee )>0$. Let us fix a non-zero map $h: T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}$. Since $T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q$ has neither $(1,0)$ nor $(0,1)$ as an index, we have $f\circ h = 0$, i.e. $\mathrm{Im} (h) \subseteq \ker (f)$. Since $T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q$ is stable and neither $(2,0)$ nor $(1,1)$ is an index of ${{\ensuremath{{\mathcal{E}}}}}$, $\mathrm{Im} (h)$ has rank $2$. Since $\mathrm{Im} (h)$ is globally generated, we get that $\mathrm{Im} (h)$ is a quotient of ${{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus 3}$ by a map ${{\ensuremath{{\mathcal{O}}}}}_Q(-1,-1){\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus 3}$ defined by $3$ linearly independent forms vanishing at some point $O\in {{\ensuremath{\mathbb{P}}}}^3$. We also see that $\ker (f)$ is the saturation of $\mathrm{Im} (h)$ in ${{\ensuremath{{\mathcal{E}}}}}$. If $O\notin Q$, then $\mathrm{Im}(h)$ is saturated in ${{\ensuremath{{\mathcal{E}}}}}$ and so $\mathrm{Im} (h) =\ker (f) \cong {{\ensuremath{{\mathcal{A}}}}}_O$. It implies that ${{\ensuremath{{\mathcal{E}}}}}\cong {{\ensuremath{{\mathcal{O}}}}}_Q(1,0)\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(0,1)\oplus {{\ensuremath{{\mathcal{A}}}}}_O$. Now assume $O\in Q$ and then $\mathrm{Im} (h)$ is not locally free. In particular, we have $\mathrm{Im} (h)\ne \ker (f)$. Note that the sheaves $\mathrm{Im} (h)$ and $\ker (f)$ have the same Chern numbers, while $\ker (f)/\mathrm{Im} (h)$ is supported by a single point $O$. Thus we have $c_2(\mathrm{Im} (h)) > c_2(\ker (f))$, a contradiction. \(c) Let us take any ${{\ensuremath{{\mathcal{E}}}}}\in \mathfrak{G}_5$. Since ${{\ensuremath{{\mathcal{E}}}}}$ is an extension of some ${{\ensuremath{{\mathcal{G}}}}}\in \mathfrak{G}_4$ by ${{\ensuremath{{\mathcal{O}}}}}_Q$, it is sufficient to use the rank $4$ case. Now the only remaining case is when $c_2({{\ensuremath{{\mathcal{F}}}}})=6$ and this case can be divided into two parts depending on the index. There exists an indecomposable and globally generated vector bundle of rank $r\ge 3$ on $Q$ with $c_1=(2,2)$, $c_2=6$ with index $(1,0)$ or $(0,1)$ if and only if we have $r\le 5$. Without loss of generality, let us assume that the index of ${{\ensuremath{{\mathcal{F}}}}}$ is $(1,0)$, i.e ${{\ensuremath{{\mathcal{F}}}}}$ fits into the sequence $$0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(-1,-2) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus 2} \oplus {{\ensuremath{{\mathcal{O}}}}}_Q(1,0) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{F}}}}}{\xrightarrow{\ \ }}0$$ and so $h^1({{\ensuremath{{\mathcal{F}}}}}^\vee)=4$. Thus we have $\mathrm{rank}({{\ensuremath{{\mathcal{E}}}}})\le 6$. Assume that ${{\ensuremath{{\mathcal{E}}}}}$ is decomposable. If ${{\ensuremath{{\mathcal{E}}}}}$ does not have a line bundle as its direct factor, then we have ${{\ensuremath{{\mathcal{E}}}}}\cong {{\ensuremath{{\mathcal{E}}}}}_1 \oplus {{\ensuremath{{\mathcal{E}}}}}_2$ with $c_1({{\ensuremath{{\mathcal{E}}}}}_1)=c_1({{\ensuremath{{\mathcal{E}}}}}_2)=(1,1)$. From the classification of indecomposable and globally generated vector bundles of rank at least $2$ with $c_1=(1,1)$, the only options for ${{\ensuremath{{\mathcal{E}}}}}_1$ and ${{\ensuremath{{\mathcal{E}}}}}_2$ are either ${{\ensuremath{{\mathcal{A}}}}}_P$ or $T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q$ since ${{\ensuremath{{\mathcal{E}}}}}$ does not have any trivial factor. But it is impossible due to the second Chern class counting and the index. Thus ${{\ensuremath{{\mathcal{E}}}}}$ should have at least one line bundle as its direct factor. Since $h^0({{\ensuremath{{\mathcal{F}}}}}(0,-1))=0$, the direct factor must be ${{\ensuremath{{\mathcal{O}}}}}_Q(1,0)$ and so we have ${{\ensuremath{{\mathcal{E}}}}}\cong {{\ensuremath{{\mathcal{O}}}}}_Q(1,0)\oplus {{\ensuremath{{\mathcal{G}}}}}$ for a globally generated vector bundle with $c_1({{\ensuremath{{\mathcal{G}}}}})=(1,2)$ and $c_2({{\ensuremath{{\mathcal{G}}}}})=4$. In particular, ${{\ensuremath{{\mathcal{G}}}}}$ is of maximal type with $$0{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(-1,-2) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus r}{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{G}}}}}{\xrightarrow{\ \ }}0.$$ From the long exact sequence $$0{\xrightarrow{\ \ }}H^0({{\ensuremath{{\mathcal{O}}}}}_Q(1,0)^{\oplus r}) {\xrightarrow{\ \ }}H^0 ({{\ensuremath{{\mathcal{O}}}}}_Q(2,2)) {\xrightarrow{\ \ }}H^1({{\ensuremath{{\mathcal{G}}}}}^\vee (1,0)),$$ we have $h^1({{\ensuremath{{\mathcal{G}}}}}^\vee (1,0))>0$ for $r\le 4$. In other words, there exists a non-trivial extension ${{\ensuremath{{\mathcal{E}}}}}'$ of ${{\ensuremath{{\mathcal{G}}}}}$ by ${{\ensuremath{{\mathcal{O}}}}}_Q(1,0)$. Since $h^0({{\ensuremath{{\mathcal{E}}}}}(-1,0))=h^0({{\ensuremath{{\mathcal{F}}}}}(-1,0))=1$, so ${{\ensuremath{{\mathcal{E}}}}}'$ cannot be isomorphic to ${{\ensuremath{{\mathcal{O}}}}}_Q(1,0)\oplus {{\ensuremath{{\mathcal{G}}}}}'$ and thus ${{\ensuremath{{\mathcal{E}}}}}'$ is indecomposable. Since $h^1({{\ensuremath{{\mathcal{O}}}}}_Q(1,0))=0$ and ${{\ensuremath{{\mathcal{O}}}}}_Q(1,0)$ is globally generated, so ${{\ensuremath{{\mathcal{E}}}}}'$ is globally generated. Hence there exists an indecomposable and globally generated vector bundle of rank up to $4$ with $c_1=(2,2)$, $c_2=6$ and index $(1,0)$. Now let us consider the case of rank $5$. Let us define $\mathbb {G}_{r-1}$ with $3 \le r \le 6$ to be the set of all ${{\ensuremath{{\mathcal{G}}}}}_{r-1}$ fitting into an exact sequence $$\label{eqa1} 0 {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(-1,-2) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus r} {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{G}}}}}_{r-1}{\xrightarrow{\ \ }}0$$ and with no trivial factor, i.e. ${{\ensuremath{{\mathcal{G}}}}}_{r-1}$ is a globally generated vector bundle of rank $r$ with no trivial factor, $c_1 =(1,2)$ and maximal type. Let us fix a bundle ${{\ensuremath{{\mathcal{G}}}}}_5$ and set $V_1:= H^0({{\ensuremath{{\mathcal{G}}}}}_5)$, $V_2:= H^0({{\ensuremath{{\mathcal{O}}}}}_Q(1,0))$ and $V:= V_1\oplus V_2$. For a generally fixed ${\varphi}\in V$, say ${\varphi}= ({\varphi}_1,{\varphi}_2)$, let us set ${{\ensuremath{{\mathcal{E}}}}}:= \mathrm{coker} ({\varphi})$. It is sufficient to prove that ${{\ensuremath{{\mathcal{E}}}}}$ is indecomposable. Assume that this is not the case. We saw above that ${{\ensuremath{{\mathcal{E}}}}}\cong {{\ensuremath{{\mathcal{G}}}}}_4\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(1,0)$ for some ${{\ensuremath{{\mathcal{G}}}}}_4\in \mathbb {G}_4$. Hence ${\varphi}$ induces an exact sequence $$0 {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q \stackrel{{\varphi}}{{\xrightarrow{\ \ }}} {{\ensuremath{{\mathcal{G}}}}}_5\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(1,0)\stackrel{\psi}{{\xrightarrow{\ \ }}}{{\ensuremath{{\mathcal{G}}}}}_4\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(1,0){\xrightarrow{\ \ }}0$$ with $\psi$ represented as a $2\times 2$ matrix $$\left(\begin{array}{ccc} \psi _{11} & \psi _{12} \\ \psi _{21} & \psi _{22} \end{array} \right)$$ with $\psi _{11} : {{\ensuremath{{\mathcal{G}}}}}_5 {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{G}}}}}_4$, $\psi _{12}: {{\ensuremath{{\mathcal{O}}}}}_Q(1,0) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{G}}}}}_4$, $\psi _{21}: {{\ensuremath{{\mathcal{G}}}}}_5 {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(1,0)$ and $\psi _{22}: {{\ensuremath{{\mathcal{O}}}}}_Q(1,0){\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(1,0)$. From (\[eqa1\]) we get $H^0({{\ensuremath{{\mathcal{G}}}}}_4(-1,0)) =0$. Hence $\psi _{12}=0$. Dualizing (\[eqa1\]) we get $H^0({{{\ensuremath{{\mathcal{G}}}}}_5}^\vee (1,0)) =0$, because the map ${{\ensuremath{{\mathcal{O}}}}}_Q(-1,-2) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus 6}$ is induced by the complete linear system $\vert {{\ensuremath{{\mathcal{O}}}}}_Q(1,2)\vert$. Hence we have $\psi _{21}=0$, and so $\psi _{11}$ is surjective and $\psi _{22}$ is an isomorphism. Thus we have $\ker (\psi ) \subset {{\ensuremath{{\mathcal{G}}}}}_5\oplus \{0\}$ and so $\mathrm{Im} ({\varphi}) \subset {{\ensuremath{{\mathcal{G}}}}}_5\oplus \{0\}$. Since $H^0({{{\ensuremath{{\mathcal{G}}}}}_5}^\vee (1,0)) =0$, there is a unique subsheaf of ${{\ensuremath{{\mathcal{G}}}}}_5\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(1,0)$ isomorphic to ${{\ensuremath{{\mathcal{G}}}}}_5$. Since ${{\ensuremath{{\mathcal{O}}}}}_Q(1,0)$ is globally generated and ${\varphi}$ is general, we get a contradiction. By Proposition \[split\], we do not have an indecomposable extension of ${{\ensuremath{{\mathcal{F}}}}}$ by ${{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus 4}$. \[split\] Every globally generated vector bundle of rank $6$ with $c_1=(2,2)$, $c_2=6$, index $(1,0)$ and without trivial factors, has ${{\ensuremath{{\mathcal{O}}}}}_Q(1,0)$ as its direct factor. We will prove that any such a bundle is isomorphic to ${{\ensuremath{{\mathcal{G}}}}}_5\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(1,0)$. Since the set of all ${{\ensuremath{{\mathcal{F}}}}}$ with given numeric data is irreducible, the set of all such bundles ${{\ensuremath{{\mathcal{E}}}}}$ of rank $6$ is parametrized by an irreducible variety $\Gamma $. If ${{\ensuremath{{\mathcal{C}}}}}$ is a proper subsheaf of ${{\ensuremath{{\mathcal{G}}}}}_5$ generated by its global sections, then ${{\ensuremath{{\mathcal{C}}}}}\cong {{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus c}$ for some $c$. Set $c:= \mathrm{rank} ({{\ensuremath{{\mathcal{C}}}}})$. Since ${{\ensuremath{{\mathcal{C}}}}}$ is globally generated, we have ${{\ensuremath{{\mathcal{C}}}}}\cong {{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus c}$ if and only if $h^0({{\ensuremath{{\mathcal{C}}}}}) \le c$. If $c=5$, then we have ${{\ensuremath{{\mathcal{C}}}}}\cong {{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus 5}$ because $h^0({{\ensuremath{{\mathcal{G}}}}}_5) = 6$ and $h^0({{\ensuremath{{\mathcal{C}}}}}) < h^0({{\ensuremath{{\mathcal{G}}}}}_5)$ which is due to the global generatedness of ${{\ensuremath{{\mathcal{G}}}}}_r$. Now assume that $1\le c \le 4$ and $d:= h^0({{\ensuremath{{\mathcal{C}}}}}) >c$. Let ${{\ensuremath{{\mathcal{C}}}}}'$ be the saturation of ${{\ensuremath{{\mathcal{C}}}}}$ in ${{\ensuremath{{\mathcal{G}}}}}_r$. The torsion free sheaf ${{\ensuremath{{\mathcal{G}}}}}_5/{{\ensuremath{{\mathcal{C}}}}}'$ is generated by $H^0({{\ensuremath{{\mathcal{G}}}}}_r)/H^0({{\ensuremath{{\mathcal{C}}}}})$, i.e. by a vector space of dimension at most $5-c =\mathrm{rank} ({{\ensuremath{{\mathcal{G}}}}}_5/{{\ensuremath{{\mathcal{C}}}}}')$. Thus the sheaf ${{\ensuremath{{\mathcal{G}}}}}_5/{{\ensuremath{{\mathcal{C}}}}}'$ is trivial and so ${{\ensuremath{{\mathcal{G}}}}}_5$ has a trivial factor, a contradiction. Notice that ${{\ensuremath{{\mathcal{G}}}}}_5$ is unique up to isomorphisms because it is induced by the complete linear system $\vert {{\ensuremath{{\mathcal{O}}}}}_Q(1,2)\vert$ and so we have $u^\ast ({{\ensuremath{{\mathcal{G}}}}}_5) \cong {{\ensuremath{{\mathcal{G}}}}}_5$ for each $u\in \mathrm{Aut} ({{\ensuremath{\mathbb{P}}}}^1 )\times \mathrm{Aut} ({{\ensuremath{\mathbb{P}}}}^1 )$. Hence it is uniform with respect to both system of lines. Its splitting type is $(1,0,0,0,0,0)$ with respect to one of the system of lines and $(1,1,0,0,0,0)$ with respect to the other one, because the cokernel of a general map ${{\ensuremath{{\mathcal{O}}}}}_{{{\ensuremath{\mathbb{P}}}}^1}(-2) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_{{{\ensuremath{\mathbb{P}}}}^1}^{\oplus 3}$ has splitting type $(1,1)$ and thus its splitting type is not $(2,0,0,0,0,0)$. Hence the bundle ${{\ensuremath{{\mathcal{G}}}}}_5\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(1,0)$ is uniform of type $(1,1,0,0,0,0)$ with respect to both systems of lines. Let ${{\ensuremath{{\mathcal{E}}}}}$ be a general element of $\Gamma$. ${{\ensuremath{{\mathcal{E}}}}}$ is uniform of splitting type $(1,1,0,0,0,0)$, i.e. for each line $D\subset Q$, the bundle ${{\ensuremath{{\mathcal{E}}}}}|_ D$ is a direct sum of two degree $1$ line bundles and four degree $0$ line bundles. Let us define a set $$\tilde{\Gamma}:= \{({{\ensuremath{{\mathcal{E}}}}},D) \in \Gamma \times \vert {{\ensuremath{{\mathcal{O}}}}}_Q(1,0)\vert : {{\ensuremath{{\mathcal{E}}}}}|_ D \text{ has splitting type } (2,0,0,0,0,0)\}.$$ Here we see $\Gamma$ as an irreducible algebraic variety parametrizing all bundles in $\Gamma$ (not necessarily one-to-one). By semicontinuity $\tilde{\Gamma}$ is closed in $\Gamma \times \vert {{\ensuremath{{\mathcal{O}}}}}_Q(1,0)\vert $. Let $\pi _1: \tilde{\Gamma} {\xrightarrow{\ \ }}\Gamma$ be the restriction of the first projection $p_1: \Gamma \times \vert {{\ensuremath{{\mathcal{O}}}}}_Q(1,0)\vert {\xrightarrow{\ \ }}\Gamma$ to $\tilde{\Gamma}$. Since $p_1$ is proper and $\tilde{\Gamma}$ is closed in $\Gamma \times \vert {{\ensuremath{{\mathcal{O}}}}}_Q(1,0)\vert $, the map $\pi _1$ is proper. The bundle ${{\ensuremath{{\mathcal{G}}}}}_5 \oplus {{\ensuremath{{\mathcal{O}}}}}_Q(1,0)$, is uniform with splitting type $(1,1,0,0,0,0)$ and thus $\pi _1 (\tilde{\Gamma})$ is a proper closed subset of the variety $\Gamma$. Hence ${{\ensuremath{{\mathcal{E}}}}}\vert_ D$ has splitting type $(1,1,0,0,0,0)$ for every $D\in \vert {{\ensuremath{{\mathcal{O}}}}}_Q(0,1)\vert$. The same proof works for any $D\in \vert {{\ensuremath{{\mathcal{O}}}}}_Q(0,1)\vert$. Let us fix a uniform bundle ${{\ensuremath{{\mathcal{E}}}}}\in \Gamma$. Since $h^0({{\ensuremath{{\mathcal{E}}}}}(-1,0)) =1$, there is an injective map $u: {{\ensuremath{{\mathcal{O}}}}}_Q(1,0) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}$. Let us set ${{\ensuremath{{\mathcal{M}}}}}= \mathrm{coke}r (u)$. Since $h^0({{\ensuremath{{\mathcal{E}}}}}(-2,0)) = h^0({{\ensuremath{{\mathcal{E}}}}}(-1,-1)) =0$, so $\mathrm{Im} (u)$ is saturated in ${{\ensuremath{{\mathcal{E}}}}}$ and thus ${{\ensuremath{{\mathcal{M}}}}}$ is torsion free, i.e. $u$ is an embedding of vector bundles outside finitely many points of $Q$. We want to prove that ${{\ensuremath{{\mathcal{M}}}}}$ is locally free. Since ${{\ensuremath{{\mathcal{E}}}}}$ is uniform of splitting type $(1,1,0,0,0,0)$ with respect to the ruling $\vert {{\ensuremath{{\mathcal{O}}}}}_Q(0,1)\vert$, while ${{\ensuremath{{\mathcal{O}}}}}_Q(1,0)$ has splitting type $1$ with respect to the same ruling, $u\vert _D$ is an embedding of vector bundles for each $D\in \vert {{\ensuremath{{\mathcal{O}}}}}_Q(0,1)\vert$. It implies that ${{\ensuremath{{\mathcal{M}}}}}$ is locally free. Since ${{\ensuremath{{\mathcal{M}}}}}$ is a globally generated vector bundle with $c_1=(1,2)$, no trivial factor and $h^0({{\ensuremath{{\mathcal{M}}}}}(-1,0)) = h^0({{\ensuremath{{\mathcal{M}}}}}(0,-1)) =0$, we have ${{\ensuremath{{\mathcal{M}}}}}\cong {{\ensuremath{{\mathcal{G}}}}}_5$. Look again at the exact sequence $$H^0({{\ensuremath{{\mathcal{O}}}}}_Q(1,0))^{\oplus 6} \stackrel{v}{{\xrightarrow{\ \ }}} H^0 ({{\ensuremath{{\mathcal{O}}}}}_Q(2,2)) {\xrightarrow{\ \ }}H^1({{{\ensuremath{{\mathcal{G}}}}}_5} ^\vee (1,0)) {\xrightarrow{\ \ }}0.$$ Any extension of ${{\ensuremath{{\mathcal{G}}}}}_5$ by ${{\ensuremath{{\mathcal{O}}}}}_Q(1,0)$ splits if and only if the map $v$ is surjective. Since ${{\ensuremath{{\mathcal{O}}}}}_Q(1,0)$ is globally generated and $h^0({{\ensuremath{{\mathcal{O}}}}}_Q(1,0)) =2$, we have an exact sequence $$\label{eqaa1} 0 {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(-1,0) {\xrightarrow{\ \ }}H^0({{\ensuremath{{\mathcal{O}}}}}_Q(1,0))\otimes {{\ensuremath{{\mathcal{O}}}}}_Q {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(1,0) {\xrightarrow{\ \ }}0.$$ Since $h^1({{\ensuremath{{\mathcal{O}}}}}_Q(0,2))=0$, by tensoring the sequence (\[eqaa1\]) with ${{\ensuremath{{\mathcal{O}}}}}_Q(1,2)$ we get the surjectivity of the multiplication map $$H^0({{\ensuremath{{\mathcal{O}}}}}_Q(1,0))\otimes H^0({{\ensuremath{{\mathcal{O}}}}}_Q(1,2)) {\xrightarrow{\ \ }}H^0({{\ensuremath{{\mathcal{O}}}}}_Q(2,2)),$$ which is the map $v$. In other words, $v$ is induced by the complete linear system $|{{\ensuremath{{\mathcal{O}}}}}_Q(1,2)|$ and so we have $h^1({{\ensuremath{{\mathcal{G}}}}}_5^\vee(1,0))=0$. In summary, we just proved the existence of a non-empty open subset $U$ of $\Gamma$ such that ${{\ensuremath{{\mathcal{E}}}}}_\gamma \cong {{\ensuremath{{\mathcal{O}}}}}_Q(1,0) \oplus {{\ensuremath{{\mathcal{G}}}}}_5$ for all $\gamma \in U$ and that $U$ contains all uniform vector bundles in $\Gamma$ (and all ${{\ensuremath{{\mathcal{E}}}}}\in \Gamma$ for which the unique inclusion ${{\ensuremath{{\mathcal{O}}}}}_Q(1,0) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}$ has a locally free quotient). By semicontinuity we have $h^0({{\ensuremath{{\mathcal{E}}}}}^\vee \otimes {{\ensuremath{{\mathcal{G}}}}}_5 )>0$. For a fixed non-zero $w: {{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{G}}}}}_5$, $\mathrm{Im}(w)$ is globally generated since ${{\ensuremath{{\mathcal{E}}}}}$ is globally generated. Since ${{\ensuremath{{\mathcal{E}}}}}$ has no trivial factor, so $\mathrm{Im} (w)$ has no trivial factor. [*Claim 1*]{} gives that $w$ is surjective and thus $\ker (w) $ is a line bundle. Hence ${{\ensuremath{{\mathcal{E}}}}}$ is an extension of ${{\ensuremath{{\mathcal{G}}}}}_5$ by ${{\ensuremath{{\mathcal{O}}}}}_Q(1,0)$ and thus we have ${{\ensuremath{{\mathcal{E}}}}}\cong {{\ensuremath{{\mathcal{G}}}}}_5 \oplus {{\ensuremath{{\mathcal{O}}}}}_Q(1,0)$ since $h^1({{\ensuremath{{\mathcal{G}}}}}_5^\vee (1,0))=0$. As the final case let us consider when ${{\ensuremath{{\mathcal{F}}}}}$ is globally generated vector bundle of rank $2$ with $c_1=(2,2)$, $c_2=6$ and index $(0,0)$. In particular, we have $h^1({{\ensuremath{{\mathcal{F}}}}}^\vee)=4$ and thus there exists an indecomposable extension ${{\ensuremath{{\mathcal{E}}}}}$ of ${{\ensuremath{{\mathcal{F}}}}}$ by trivial factors only if $\mathrm{rank}({{\ensuremath{{\mathcal{E}}}}})\le 6$. \[oo\] For two points $P, O\in {{\ensuremath{\mathbb{P}}}}^3\setminus Q$, we have ${{\ensuremath{{\mathcal{A}}}}}_P \cong {{\ensuremath{{\mathcal{A}}}}}_O$ if and only if $P=O$. The assertion follows from the fact that the set of jumping conics of ${{\ensuremath{{\mathcal{A}}}}}_P$ is the hyperplane in $({{\ensuremath{\mathbb{P}}}}^3)^*$ corresponding to $P$ [@huh2]. \[rm1\] For two points $P$ and $O$ in ${{\ensuremath{\mathbb{P}}}}^3\setminus Q$, we have : $$h^1({{\ensuremath{{\mathcal{A}}}}}_O^\vee \otimes {{\ensuremath{{\mathcal{A}}}}}_P)=\left\{ \begin{array}{ll} 3, & \hbox{if ${{\ensuremath{{\mathcal{A}}}}}_O\cong {{\ensuremath{{\mathcal{A}}}}}_P$}; \\ 2, & \hbox{if ${{\ensuremath{{\mathcal{A}}}}}_O \not\cong {{\ensuremath{{\mathcal{A}}}}}_P$}. \end{array} \right.$$ Since each ${{\ensuremath{{\mathcal{A}}}}}$ is stable, we have $h^0({{\ensuremath{{\mathcal{A}}}}}_O^\vee \otimes {{\ensuremath{{\mathcal{A}}}}}_P)=0$ if $P\ne O$ and $h^0({{\ensuremath{{\mathcal{A}}}}}_O^\vee \otimes {{\ensuremath{{\mathcal{A}}}}}_P)=1$ if $P=O$ by Lemma \[oo\]. Duality gives $h^2({{\ensuremath{{\mathcal{A}}}}}_O^\vee \otimes {{\ensuremath{{\mathcal{A}}}}}_P) = h^0({{\ensuremath{{\mathcal{A}}}}}_O\otimes {{\ensuremath{{\mathcal{A}}}}}_P^\vee (-2,-2)) =0$ and so we can apply Hirzebruch Riemann Roch to obtain the assertion. Alternatively let us consider the exact sequence $$\label{eqrm1} 0 {\xrightarrow{\ \ }}\mathcal {A}_O^\vee (-1) {\xrightarrow{\ \ }}(\mathcal {A}_O^\vee )^{\oplus 3} {\xrightarrow{\ \ }}\mathcal {A}_O^\vee \otimes \mathcal{A}_{P} {\xrightarrow{\ \ }}0$$ and then we have $$\begin{aligned} &h^2({{\ensuremath{{\mathcal{A}}}}}_O^\vee (-1)) = h^1({{\ensuremath{{\mathcal{A}}}}}_O(-2)) = h^1(T{{\ensuremath{\mathbb{P}}}}^2(-3)) + h^1(T{{\ensuremath{\mathbb{P}}}}^2(-4)) =1,\\ &h^1(\mathcal {A}_O^\vee(-1) )=h^1(\Omega _{{{\ensuremath{\mathbb{P}}}}^3}) +h^1(\Omega _{{{\ensuremath{\mathbb{P}}}}^3}(-1)) =1, \text{ and}\\ &h^1(\mathcal {A}_O^\vee )=h^1(\Omega _{{{\ensuremath{\mathbb{P}}}}^2}(1)) +h^1(\Omega _{{{\ensuremath{\mathbb{P}}}}^2}) =1.\end{aligned}$$ Then we can use (\[eqrm1\]) to obtain the assertion. \[o1\] Let ${{\ensuremath{{\mathcal{E}}}}}$ be a vector bundle fitting into a non-splitting exact sequence $$\label{eqo1} 0 {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{A}}}}}_O \stackrel{j}{{\xrightarrow{\ \ }}} {{\ensuremath{{\mathcal{E}}}}}\stackrel{f}{{\xrightarrow{\ \ }}} {{\ensuremath{{\mathcal{A}}}}}_P {\xrightarrow{\ \ }}0.$$ Then ${{\ensuremath{{\mathcal{E}}}}}$ is semistable and simple. Moreover, ${{\ensuremath{{\mathcal{A}}}}}_O$ is the only subsheaf of ${{\ensuremath{{\mathcal{E}}}}}$ with slope $1/2$. Since $c_1({{\ensuremath{{\mathcal{A}}}}})=1$ and $h^0({{\ensuremath{{\mathcal{A}}}}}(-1,-1)) =0$, so ${{\ensuremath{{\mathcal{A}}}}}$ is stable. Since $h^0({{\ensuremath{{\mathcal{E}}}}}(-1,-1)) =0$, there is no non-zero map ${{\ensuremath{{\mathcal{L}}}}}{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}$ with ${{\ensuremath{{\mathcal{L}}}}}$ a line bundle and $\deg ({{\ensuremath{{\mathcal{L}}}}})>0$. Since $h^0({{\ensuremath{{\mathcal{E}}}}}^\vee )=0$, there is no non-zero map ${{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{R}}}}}$ with ${{\ensuremath{{\mathcal{R}}}}}$ a line bundle and $\deg ({{\ensuremath{{\mathcal{R}}}}})\le 0$. Hence there is no injective map ${{\ensuremath{{\mathcal{F}}}}}{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}$ with a sheaf ${{\ensuremath{{\mathcal{F}}}}}$ of rank $3$ whose degree is at least $2$, i.e. the slope of ${{\ensuremath{{\mathcal{F}}}}}$ is at least $1/2$. Let ${{\ensuremath{{\mathcal{G}}}}}$ be a rank $2$ subsheaf of ${{\ensuremath{{\mathcal{E}}}}}$ different from $j({{\ensuremath{{\mathcal{A}}}}}_O)$, saturated in ${{\ensuremath{{\mathcal{E}}}}}$ and with maximal slope, say $a$. Since ${{\ensuremath{{\mathcal{E}}}}}/{{\ensuremath{{\mathcal{F}}}}}$ has no torsion and $Q$ is a smooth surface, ${{\ensuremath{{\mathcal{F}}}}}$ is locally free. Assume $a\ge 1/2$. Since $h^0({{\ensuremath{{\mathcal{F}}}}}(-1,-1)) \le h^0({{\ensuremath{{\mathcal{E}}}}}(-1,-1))=0$, ${{\ensuremath{{\mathcal{F}}}}}$ is stable. Hence $f\vert_{ {{\ensuremath{{\mathcal{F}}}}}} : {{\ensuremath{{\mathcal{F}}}}}{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{A}}}}}_P$ is either zero or an isomorphism. In the former case we have ${{\ensuremath{{\mathcal{F}}}}}\subseteq {{\ensuremath{{\mathcal{A}}}}}_O$, a contradiction. In the latter case $f\vert_{ {{\ensuremath{{\mathcal{F}}}}}}$ induces a splitting of (\[eqo1\]), a contradiction. Let us assume that ${{\ensuremath{{\mathcal{E}}}}}$ is not simple, i.e. assume the existence of a non-zero map $h: {{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}$ with $m:= \mathrm{rank}(h({{\ensuremath{{\mathcal{E}}}}})) <4$. Set $d:= \deg (\mathrm{Im} (h))$. Since ${{\ensuremath{{\mathcal{E}}}}}$ is semistable and $\mathrm{Im}(h)$ is both a quotient of ${{\ensuremath{{\mathcal{E}}}}}$ and a subsheaf of ${{\ensuremath{{\mathcal{E}}}}}$, we get $d/m = 1/2$, i.e. $m=2$ and $d=1$. We just saw that that $j({{\ensuremath{{\mathcal{A}}}}}_O)$ is the saturation of $\mathrm{Im} (h)$ and so $h$ gives a splitting of (\[eqo1\]), a contradiction. \[re17\] We can similarly obtain $h^1(T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q \otimes T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q^\vee)=0$ and thus there is no non-trivial extension of $T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q$ by itself. The exact sequence $$0 {\xrightarrow{\ \ }}T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q ^\vee(1,1) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(1,1)^{\oplus 4} {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q(2,2) {\xrightarrow{\ \ }}0$$ gives $h^0(T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q ^\vee (1,1)) =0$. Since $h^0(T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q (-1,-1)) =0$, so $T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q$ is ${{\ensuremath{{\mathcal{O}}}}}_Q(1,1)$-stable. For any $P\in {{\ensuremath{\mathbb{P}}}}^3\setminus Q$ we have 1. $h^0({{\ensuremath{{\mathcal{A}}}}}_P\otimes T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q^\vee)=1$ and $h^1({{\ensuremath{{\mathcal{A}}}}}_P \otimes T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q^\vee)=0$; 2. $h^0({{{\ensuremath{{\mathcal{A}}}}}_P}^\vee \otimes T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q)=0$ and $h^1({{\ensuremath{{\mathcal{A}}}}}_P^\vee \otimes T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q))=4$. In particular, there is no non-trivial extension of $T{{\ensuremath{\mathbb{P}}}}^3(-1)|_{Q})$ by ${{\ensuremath{{\mathcal{A}}}}}_P$ and there exist non-trivial extensions of ${{\ensuremath{{\mathcal{A}}}}}_P$ by $T{{\ensuremath{\mathbb{P}}}}^3(-1)|_{Q})$. For the first assertion, let us consider the sequence $$0{\xrightarrow{\ \ }}T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q^\vee(-1,-1) {\xrightarrow{\ \ }}(T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q^\vee)^{\oplus 3} {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{A}}}}}_P \otimes T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q^\vee {\xrightarrow{\ \ }}0$$ to obtain $$\begin{aligned} 0{\xrightarrow{\ \ }}H^0({{\ensuremath{{\mathcal{A}}}}}_P \otimes T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q^\vee) &{\xrightarrow{\ \ }}H^1( T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q^\vee (-1,-1)) {\xrightarrow{\ \ }}H^1(T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q^\vee)^{\oplus 3} \\ &{\xrightarrow{\ \ }}H^1({{\ensuremath{{\mathcal{A}}}}}_P \otimes T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q^\vee) {\xrightarrow{\ \ }}H^2(T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q^\vee (-1,-1)).\end{aligned}$$ Since $h^1(T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q^\vee (-1,-1))=h^1(T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q(-1,-1))=h^2({{\ensuremath{{\mathcal{O}}}}}_Q(-2,-2))=1$, $h^1(T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q^\vee)=0$ and $h^2(T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q^\vee (-1,-1))=0$, so we have the assertion. For the second assertion, we have $h^0({{\ensuremath{{\mathcal{A}}}}}_P^\vee \otimes T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q)=0$ since both of ${{\ensuremath{{\mathcal{A}}}}}_P$ and $T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q$ are stable. Thus we have $h^1({{\ensuremath{{\mathcal{A}}}}}_P^\vee \otimes T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q)=4$ similarly. \[o2\] Let ${{\ensuremath{{\mathcal{E}}}}}$ be a vector bundle fitting into a non-splitting exact sequence $$\label{eqo2} 0 {\xrightarrow{\ \ }}T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q\stackrel{j}{{\xrightarrow{\ \ }}} {{\ensuremath{{\mathcal{E}}}}}\stackrel{f}{{\xrightarrow{\ \ }}} {{\ensuremath{{\mathcal{A}}}}}_P {\xrightarrow{\ \ }}0.$$ Then ${{\ensuremath{{\mathcal{E}}}}}$ is stable. Let us assume that ${{\ensuremath{{\mathcal{E}}}}}$ is not stable and take a proper saturated subsheaf ${{\ensuremath{{\mathcal{F}}}}}$ of ${{\ensuremath{{\mathcal{E}}}}}$ with maximal slope. Set $a:= \deg ({{\ensuremath{{\mathcal{F}}}}})$ and $m:= \mathrm{rank} ({{\ensuremath{{\mathcal{F}}}}})$. Since ${{\ensuremath{{\mathcal{F}}}}}$ is saturated in ${{\ensuremath{{\mathcal{E}}}}}$ and $Q$ is a smooth surface, ${{\ensuremath{{\mathcal{F}}}}}$ is locally free. Since ${{\ensuremath{{\mathcal{E}}}}}$ has slope $1/2$, we get $a/m \ge 1/2$. From (\[eqo2\]) we get $h^0({{\ensuremath{{\mathcal{E}}}}}(-1)) = h^0({{\ensuremath{{\mathcal{E}}}}}^\vee (1)) =0$. Thus $m$ is neither $1$ nor $3$, and so we have $(m,a)=(2,1)$. Since ${{\ensuremath{{\mathcal{A}}}}}_P$ is stable and of slope $1/2$, so the map $h:=f\vert_{ {{\ensuremath{{\mathcal{F}}}}}}$ is either a zero map or an isomorphism. If $h$ is an isomorphism, then (\[eqo2\]) splits. If $h=0$, then ${{\ensuremath{{\mathcal{F}}}}}\subset T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q$, contradicting the stability of $T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q$. There exists an indecomposable and globally generated vector bundle of rank $r\ge 2$ on $Q$ with $c_1=(2,2)$, $c_2=6$ and index $(0,0)$ if and only if $2\le r \le5$. From the previous lemmas, we have the existence of indecomposable vector bundles ${{\ensuremath{{\mathcal{E}}}}}$ with the conditions in the assertion if $\mathrm{rank}({{\ensuremath{{\mathcal{E}}}}})\le 5$. When $\mathrm{rank}({{\ensuremath{{\mathcal{E}}}}})=6$, we claim that every such a bundle is isomorphic to $(T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q)^{\oplus 2}$ and it follows from Proposition \[uu1\]. Let $\mathfrak{U}$ be the set of all globally generated vector bundles of rank $6$ with $c_1=(2,2)$, $c_2=6$ and index $(0,0)$. Since the set of all rank $2$ bundles ${{\ensuremath{{\mathcal{F}}}}}$ with the same numeric data is parametrized by an irreducible variety $U'$ [@Soberon], so the set $\mathfrak {U}$ is parametrized by an irreducible variety, say $U$, because $U$ is an open subset of a quotient bundle over $U'$ from sequence (\[heqa\]). \[uu1\] We have $\mathfrak {U} = \{(T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q)^{\oplus 2}\}$. Obviously $(T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q)^{\oplus 2} \in \mathfrak {U}$. Since $h^1(T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q ^\vee \otimes T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q )=0$ from Remark \[re17\], we have $h^1(\mathcal{E}nd (T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q \oplus T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q ))=0$ and so the vector bundle $(T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q)^{\oplus 2}$ is rigid. Thus there is a non-empty open subset $V$ of $U$ such that ${{\ensuremath{{\mathcal{E}}}}}\cong (T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q)^{\oplus 2}$ for all ${{\ensuremath{{\mathcal{E}}}}}\in V$. Let us fix ${{\ensuremath{{\mathcal{E}}}}}\in U$. Since $h^0(\mathcal{H}om (T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q ,{{\ensuremath{{\mathcal{C}}}}})) = h^0(\mathcal{H}om ({{\ensuremath{{\mathcal{C}}}}},T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q )) =2$ for all ${{\ensuremath{{\mathcal{C}}}}}\in V$, the semicontinuity theorem for cohomology gives us the inequalities $h^0(\mathcal{H}om (T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q ,{{\ensuremath{{\mathcal{E}}}}}))\ge 2$ and $h^0(\mathcal{H}om ({{\ensuremath{{\mathcal{E}}}}},T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q )) \ge 2$. Assume the existence of $h: {{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q$ such that $\mathrm{Im} (h)$ has rank $1$ and write ${{\ensuremath{{\mathcal{O}}}}}_Q(a,b):= \mathrm{Im} (h)^{\vee \vee }$. Since $T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q$ is locally free, it contains $\mathrm{Im} (h)^{\vee \vee }$. Since $(0,1)$ and $(1,0)$ are not indices of $T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q$, we get $a=b=0$. Since $\mathrm{Im} (h)$ is globally generated, we get $\mathrm{Im} (h) \cong {{\ensuremath{{\mathcal{O}}}}}_Q$. Hence ${{\ensuremath{{\mathcal{E}}}}}$ has a trivial factor, a contradiction. Now assume the existence of $h : {{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q$ such that $\mathrm{Im} (h)$ has rank $2$. Set ${{\ensuremath{{\mathcal{O}}}}}_Q(u,v):= c_1(\mathrm{Im} (h)^{\vee \vee })$. Since ${{\ensuremath{{\mathcal{E}}}}}$ is globally generated, we have $u\ge 0$ and $v\ge 0$. Since $T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q$ is stable, we have $u+v \le 1$. Since $\mathrm{Im} (h)$ is a torsion free and globally generated sheaf, it fits into an exact sequence $$0 {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{O}}}}}_Q {\xrightarrow{\ \ }}\mathrm{Im} (h) {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{I}}}}}_Z(u,v){\xrightarrow{\ \ }}0$$ for some zero-dimensional scheme $Z$. Since ${{\ensuremath{{\mathcal{I}}}}}_Z(u,v)$ is globally generated and $(u,v) \in \{(0,0),(1,0),(0,1)\}$, we get $Z=\emptyset$. $\mathrm{Im} (u) \cong {{\ensuremath{{\mathcal{O}}}}}_Q\oplus {{\ensuremath{{\mathcal{O}}}}}_Q(u,v)$. In particular, ${{\ensuremath{{\mathcal{O}}}}}_Q$ is a factor of ${{\ensuremath{{\mathcal{E}}}}}$ and so every non-zero map $h: {{\ensuremath{{\mathcal{E}}}}}{\xrightarrow{\ \ }}T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q$ has rank $3$. In the same way we see that any non-zero map $T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}$ has image of rank $3$. Recall that $H^0(\mathcal{H}om ({{\ensuremath{{\mathcal{E}}}}},T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q ))$ is a vector space of dimension $\ge 2$. Fix a non-zero map $f : T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q {\xrightarrow{\ \ }}{{\ensuremath{{\mathcal{E}}}}}$. First assume $h\circ f \ne 0$. Since $T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q$ is stable, so $h\circ f : T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q {\xrightarrow{\ \ }}T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q$ is an isomorphism. Thus $T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q$ is a factor of ${{\ensuremath{{\mathcal{E}}}}}$ and so we have ${{\ensuremath{{\mathcal{E}}}}}\cong T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q \oplus {{\ensuremath{{\mathcal{V}}}}}$ for some globally generated vector bundle ${{\ensuremath{{\mathcal{V}}}}}$ of rank $3$ with $c_1=(1,1)$, $h^0({{\ensuremath{{\mathcal{V}}}}})=4$ and no trivial factor. Hence we have ${{\ensuremath{{\mathcal{E}}}}}\cong T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q \oplus T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q$. Now assume $h\circ f =0$, i.e. $\mathrm{Im} (f) \subseteq \ker (h)$. Since $f$ is injective, $\mathrm{Im} (f)\cong T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q$. The sheaf $\ker (h)$ is reflexive, because it is a kernel of a map between two vector bundles. Since $Q$ is a smooth surface, $\ker (h)$ is locally free. First assume $c_1(\mathrm{Im} (h)^{\vee \vee}) =(0,0)$. Since ${{\ensuremath{{\mathcal{E}}}}}$ is spanned, so is $\mathrm{Im} (u)$. Our assumption on $c_1$ implies that $\mathrm{Im} (h) \cong {{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus 3}$. Since ${{\ensuremath{{\mathcal{E}}}}}$ is spanned, we get that ${{\ensuremath{{\mathcal{O}}}}}_Q^{\oplus 3}$ is a factor of ${{\ensuremath{{\mathcal{E}}}}}$, a contradiction. Now assume that either $c_1(\mathrm{Im} (h)^{\vee \vee}) =(1,0)$ or $c_1(\mathrm{Im} (h)^{\vee \vee}) =(0,1)$. Since $\mathrm{Im} (h)$ is a rank 3 globally generated sheaf, but not a trivial vector bundle, we have $h^0(\mathrm{Im} (h))\ge 4$. Since $h^0(T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q )=4$ and $T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q$ is spanned, we get $\mathrm{Im} (h) =T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q$, a contradiction. Now assume $c_1(\mathrm{Im} (h)^{\vee \vee}) =(1,1)$, i.e. $c_1(\ker (h))=(1,1)$. The inclusion $j: \mathrm{Im} (f) \subseteq \ker (h)$ is a injective map between two vector bundles with the same rank and the same determinant. Hence $j$ is surjective. Recall that $h^0(\mathcal{H}om (T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q ,{{\ensuremath{{\mathcal{E}}}}}))\ge 2$. Take $f_1\in H^0(\mathcal{H}om (T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q ,{{\ensuremath{{\mathcal{E}}}}}))$ linearly independent from $f$. We get $\mathrm{Im} (f_1) \subseteq \ker (h)$. Since $\ker (h) \cong T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q$ and $T{{\ensuremath{\mathbb{P}}}}^3(-1)|_Q$ is stable, we get a contradiction. [30]{} V. Ancona and G. Ottaviani, Some applications of [B]{}eilinson’s theorem to projective spaces and quadrics, *Forum Math.* **3** (1991), no.  2, 157–176. C. Anghel, I. Coanda and N. Manolache, Globally generated vector bundles on ${{\ensuremath{\mathbb{P}}}}^n$ with $c_1=4$, *Preprint*, arXiv : 1305.3464 \[math.AG\], 2013. C. Anghel and N. Manolache, Globally generated vector bundles on $\mathbb{P}^n$ with $c_1=3$, *Math. Nachr.* **286** (2013), no. 14–15, 1407–1423. E. Ballico, S. Huh, and F. Malaspina, Globally generated vector bundles of rank $2$ on a smooth quadric threefold, *J. Pure Appl. Algebra* **218** (2014), no. 2, 197–207. E. Ballico, S. Huh, and F. Malaspina, On higher rank globally generated vector bundles over a smooth quadric threefold, *Preprint*, arXiv : 1211.2593v2 \[math.AG\], 2012. E. Ballico and F. Malaspina, Regularity and Cohomological Splitting Conditions for Vector Bundles on Multiprojectives Spaces, *J. Algebra* **345** (2011), 137–149. M. Coppens, The existence of base point free linear systems on smooth plane curves, *J. Algebraic Geom.*, **4** (1995), no. 1, 1–15. L. Chiodera and P. Ellia, Rank two globally generated vector bundles with $c_1 \leq 5$, *Rend. Istit. Mat. Univ. Trieste* **44** (2012), 1–10. P. Ellia, Chern classes of rank two globally generated vector bundles on $\mathbb{P}^2$, *Atti Accad. Naz. Lincei Cl. Sci. Fis. Mat. Natur. Rend. Lincei (9) Mat. Appl.* **24** (2013), no. 2, 147–163. S. Greco and G. Raciti, The L[ü]{}roth semigroup of plane algebraic curves, *Pacific J. of Math.* **151** (1991), no. 1, 43–56. Ph. Griffiths and J. Harris, Residues and zero-cycles on algebraic varieties, *Ann. of Math.* **108** (1978), no. 3, 461–505. S. Huh, Moduli of stable sheaves on a smooth quadric and a Brill-Noether locus, *J. Pure Appl. Algebra* **215** (2011), no.  9, 2099–2105. J. Le Potier, Sur l’espace de modules des fibrés de [Y]{}ang et [M]{}ills, Mathematics and physics ([P]{}aris, 1979/1982), Progr. Math. 37, 65–137 (1983), Birkhäuser Boston. F. Malaspina, A.P. Rao, Horrocks Correspondence on a Quadric Surface, *Geom. Dedicata* **169** (2014), no. 1, 15–31. N. Manolache, Globally generated vector bundles on $\mathbb{P}^3$ with $c_1=3$, *Preprint*, arXiv:1202.5988 \[math.AG\], 2012. J. C. Sierra, A degree bound for globally generated vector bundles, *Math. Z.* **262** (2009), no. 3, 517–525. J. C. Sierra and L. Ugaglia, On globally generated vector bundles on projective spaces, *J. Pure Appl. Algebra* **213** (2009), no. 11, 2141–2146. J. C. Sierra and L. Ugaglia, On globally generated vector bundles on projective spaces II, *J. Pure Appl. Algebra* **218** (2014) no. 1, 174–180. S. Soberon-Chavez, Rank [$2$]{} vector bundles over a complex quadric surface, *Quart. J. Math. Oxford Ser.* (2) **36** (1985), no.  142, 159–172. [^1]: The second author is supported by Basic Science Research Program 2010-0009195 through NRF funded by MEST

ArXiv

--- abstract: 'We present an analysis of $\sim$60 000 massive (stellar mass $M_{\star} > 10^{11} M_{\odot}$) galaxies out to $z = 1$ drawn from 55.2 deg$^2$ of the United Kingdom Infrared Telescope (UKIRT) Infrared Deep Sky Survey (UKIDSS) and the Sloan Digital Sky Survey (SDSS) II Supernova Survey. This is by far the largest survey of massive galaxies with robust mass estimates, based on infrared ($K$-band) photometry, reaching to the Universe at about half its present age. We find that the most massive ($M_{\star} > 10^{11.5} M_{\odot}$) galaxies have experienced rapid growth in number since $z = 1$, while the number densities of the less massive systems show rather mild evolution. Such a hierarchical trend of evolution is consistent with the predictions of the current semi-analytic galaxy formation model based on $\Lambda$CDM theory. While the majority of massive galaxies are red-sequence populations, we find that a considerable fraction of galaxies are blue star-forming galaxies. The blue fraction is smaller in more massive systems and decreases toward the local Universe, leaving the red, most massive galaxies at low redshifts, which would support the idea of active ’bottom-up’ formation of these populations during $0 < z < 1$.' author: - | Y. Matsuoka$^{1,2}$[^1][^2] and K. Kawara$^{2}$\ $^{1}$Graduate School of Science, Nagoya University, Furo-cho, Chikusa-ku, Nagoya 464-8602, Japan\ $^{2}$Institute of Astronomy, The University of Tokyo, Osawa 2-21-1, Mitaka, Tokyo 181-0015, Japan date: 'Accepted 2010 February 1. Received 2010 January 21; in original form 2009 April 27' title: Witnessing the active assembly phase of massive galaxies since $z$ = 1 --- \[firstpage\] galaxies: elliptical and lenticular, cD – galaxies: evolution – galaxies: formation – galaxies: luminosity function, mass function – galaxies: stellar content – cosmology: observations. Introduction ============ Understanding the origin and evolution of galaxies, in particular the most massive, is one of the major challenges in modern astrophysics. Many massive galaxies today are giant early-type systems; hence the formation of spheroids should proceed to a certain extent in locked step with the mass assembly. The compelling theory of hierarchical galaxy formation predicts that galaxies are assembled through successive mergers of smaller systems in overdensities, or haloes, of hypothetical cold dark matter (CDM) [@white78]. Massive galaxies therefore emerge in the last phase of the formation history. Alternatively, massive galaxies could form through the rapid collapse of gas followed by a single prominent starburst at high redshifts [@eggen62; @larson75]. This ’monolithic’ scenario is supported by, for example, the tight colour–luminosity relation of early-type galaxies found in galaxy clusters [e.g., @bower92; @ellis97]. While different evolution in different models makes distant massive galaxies a unique test-bed for galaxy formation scenarios, observations have not yet provided evidence for the evolutionary path of those galaxies. The major obstacle in observations originates from the scarcity of galaxies at the high end of the galaxy mass function; it means that not only it is hard to find the population but also cosmic variance, the field-to-field variation of observed volume density arising from large-scale structure, is significant. In the last decade, many large programmes of optical-band imaging have been carried out, providing excellent data sets with which to investigate distant red old galaxies in wide fields of sky exceeding a whole deg$^2$ [e.g., @bell04; @borch06; @cimatti06; @willmer06; @brown07; @faber07]. They consistently suggest that the total stellar mass locked in red galaxies with luminosities around $L \sim L^*$, where $L^*$ is the characteristic luminosity of the luminosity function, has at least doubled since $z \sim 1$. Some of them also claim little growth in the number of very luminous galaxies well above $L \sim L^*$. However, while luminous red galaxies roughly correspond to massive galaxies, it is not clear how well the evolution in the number of galaxies at the steep high end of the mass function is understood from these results, since the much more numerous, less massive galaxies with mass-to-luminosity ratios slightly less than average could easily dominate the observed numbers of luminous galaxies. The above authors also reveal that a field of view of the order of a whole deg$^2$ is still not sufficient to conquer the uncertainty arising from cosmic variance for the high-end populations of the galaxy mass function. The advent of the United Kingdom Infrared Telescope (UKIRT) InfraredDeep Sky Survey [UKIDSS; @lawrence07] provides a unique opportunity to produce an ideal sample to trace various aspects of massive galaxies in the distant Universe. Here we report the results of a $K$-band survey with optical ($u$, $g$, $r$, $i$, $z$ band) and near-infrared ($Y$, $J$, $H$ band) photometry and optical spectra, focusing on massive ($M_{\star} > 10^{11} M_{\odot}$) galaxies out to $z = 1$ in an unprecedented large area covering 55.2 deg$^2$. The $K$-band photometry provides robust estimates of galaxy stellar masses [e.g., @matsuoka08] while the very large field of view significantly suppresses cosmic variance, which allow us to conduct a unique analysis of the mean properties of distant massive galaxies. This paper is organized as follows. In Section 2 we describe the data sources and reduction process to extract the galaxy sample from the available data. Photometric redshifts and stellar masses are measured for each galaxy in Section 3. In Section 4, the number-density evolution of massive galaxies and the associated uncertainties are explored. We then discuss the star-forming properties of galaxies and the compatibility of the present results with previous measurements in Section 5. A summary follows in Section 6. Throughout this paper, we adopt the concordance cosmology of $H_0$ = 70 km s$^{-1}$ Mpc$^{-1}$, $\Omega_{\rm M} = 0.3$, and $\Omega_{\Lambda} = 0.7$. Magnitudes are expressed in the Vega magnitude system for the UKIDSS near-infrared bands and in the AB magnitude system for the optical bands. Data Sources and Reduction ========================== Near-Infrared Photometry ------------------------ We extract from the Data Release 3 (DR3; Warren et al., in preparation) of the UKIDSS/Large Area Survey (LAS) the $K$-band sources with right ascensions from 1$^{\rm h}$ 15$^{\rm m}$ to 3$^{\rm h}$ 6$^{\rm m}$ on the Sloan Digital Sky Survey [SDSS; @york00] southern equatorial stripe (see Section 2.2). The range in right ascension is chosen so that the $K$-band observations are fairly complete within the sample area. More than half of the $K$-band sources have also been observed in the $Y$, $J$ and/or $H$ bands. We exclude the sources assigned with serious quality flags corresponding to `ppErrBits` attributes larger than 31, or near (within 20 arcsec of) the detector edges of any exposure. We also exclude those sources near bright sources. This is achieved by searching for bright point and extended sources in the Two-Micron All-Sky Survey [2MASS; @skrutskie06] catalogues and rejecting all the LAS sources within sufficiently large distances of the bright 2MASS sources. The total effective area of the observations defining our sample is 55.2 deg$^2$. We retrieved all the images in our sample area from the DR3 data base, and used the `SOURCE EXTRACTOR`, version 2.5 [@bertin96] for magnitude measurements. The total magnitudes ($m_{\rm tot}$) of the sources are measured with the `SOURCE EXTRACTOR` total magnitude algorithm `MAG_AUTO`. We measure the aperture magnitudes ($m_{\rm ap}$) with circular apertures of several diameters, 4.5, 3.3, 2.8 and 2.6 arcsec, which correspond to 20 kpc at $z$ = 0.3, 0.5, 0.7 and 0.9, respectively. Since the seeing condition is generally superior in the UKIDSS LAS ($\sim$0.8 arcsec) to that in the optical observations by the SDSS ($\sim$1.5 arcsec), the near-infrared images were smoothed with the Gaussian kernel in such a way that the resultant full widths at half-maximum (FWHMs) of stellar profiles are similar to those in the optical images of the same field. The seeing measurements and smoothing were performed in each of the small rectangular subareas of approximately 9 $\times$ 13 arcsec$^2$. Then we ran the `SOURCE EXTRACTOR` on the smoothed images to measure the aperture magnitudes of the sources. The detection completeness of the $K$-band sources can be estimated by comparing the numbers of the LAS detections with those of the much deeper UKIDSS Deep Extragalactic Survey [DXS; @lawrence07] in an overlapping field. This 0.6-deg$^2$ field is a part of the DXS VIMOS 4 field, which is centred at RA 22$^{\rm h}$ 17$^{\rm m}$, Dec. $+$00$^{\circ}$ 24’ on the celestial equator. While the field is outside the RA range of our LAS fields, we confirmed that the evaluation sample of this field has a similar distribution of magnitudes and their errors to those for our actual sample. Below we show that the derived detection-completeness function reproduces the galaxy number counts from our sample, in excellent agreement with previous measurements, while it gives the lower limit of the detection completeness for our massive galaxies. We define our sample as being brighter than the limiting total magnitude $K_{\rm tot}$ = 17.9 mag, where the detection completeness is higher than 0.5. Optical Photometry \[subsec:optphot\] ------------------------------------- We use the optical $u$, $g$, $r$, $i$ and $z$-band images on the SDSS southern equatorial stripe. The stripe has been observed repeatedly in the SDSS-II Supernova Survey [@frieman08] during 2005– 2007, as well as in the original SDSS. We retrieved all the available images taken on the stripe from the SDSS Data Release 6 Supplemental and Supernova Survey data bases [@adelman08], and stacked them in each of the five bands. The images observed in runs 2738 and 3325 are set apart from others, since they are taken with the standard survey conditions of the original SDSS and can work as the reference frames for stellar photometry. We measure, for each retrieved image, the mean and root-mean-square (rms) of the sky counts and the sky transparency at the observation by comparing the stellar photometry of the relevant frame with that of the reference frames. After discarding the worst 5 per cent of the retrieved images with the largest rms of sky counts, which we find is sufficient to reject apparently flawed exposures, the images are zero-shifted and scaled according to the sky-count statistics and then stacked by the inverse-of-variance weighted average using `IRAF`. The photometric calibration of the stacked images is achieved by comparing the stellar photometry with that of the reference frames. We find that the stellar magnitudes on the stacked and reference frames are in excellent agreement, with rms less than 0.05 mag (Fig. 1). Fig. 2 shows the comparison of the original and stacked r-band images of the same field. More than 100 original SDSS frames contribute to each of the stacked frames, and the latter images are on average $>$2mag deeper than the former images. ![Comparisons of stellar magnitudes in the stacked and reference frames in $u$ (top left), $g$ (top right), $r$ (middle left), $i$ (middle right) and $z$ (bottom left) bands. The dashed lines represent the locus where the two measurements are identical. The rms errors of the photometric calibration are calculated in the magnitude ranges shown by the dotted lines, where the stellar photometry is most reliable.[]{data-label="ref_vs_stack_phot"}](f1_lores.eps){width="84mm"} ![Original (top) and stacked (bottom) SDSS $r$-band images of the same field.[]{data-label="ref_vs_stack_image"}](f2_lores.eps){width="84mm"} We run the `SOURCE EXTRACTOR` on the stacked images to extract detected sources. The groups with four or more pixels whose counts are 1.5$\sigma$ above the local background level are identified as sources. For every detection, we measure the aperture magnitudes in the 4.5-, 3.3-, 2.8- and 2.6-arcsec diameter apertures as we did for the near-infrared images. The photometry errors are calculated from the source photon counts and the background noise. The extracted sources are cross-identified with the $K$-band sources within the maximum paring tolerance of 1.0 arcsec. Owing to the deep stacking of the SDSS images, nearly 90 per cent of the $K$-band sources have counterparts in the $g$, $r$, $i$ and $z$ bands. Nearly 40 per cent of the $K$-band sources also have counterparts in the $u$ band. Optical Spectroscopy -------------------- We exploit the two redshift surveys carried out on the SDSS southern equatorial stripe; the VIMOS-VLT Deep Survey [VVDS; @lefevre05] and the DEEP2 Redshift Survey [@davis03]. Among the four fields of the VVDS ’Wide’ survey, we use the 4-deg$^2$ field of the F22 (2217$+$00), which lies on the celestial equator, centred at RA 22$^{\rm h}$ 17$^{\rm m}$ 50$^{\rm s}$.4 and Dec. $+$00$^{\circ}$ 24$^{\rm m}$ 00$^{\rm s}$. This field coincides with the UKIDSS DXS VIMOS 4, and is observed in both the LAS and the DXS. We use Field 4 (RA 02$^{\rm h}$ 30$^{\rm m}$, Dec. $+$00$^{\circ}$ 00$^{\rm m}$) of the DEEP2, one of the ’1-h survey’ fields placed on our LAS field. The VVDS adopts a pure $I$-band flux-limited selection of the sample while the DEEP2 imposes strict colour pre-selection on the spectroscopic targets to favour galaxies at $z > 0.7$; thus the two surveys are complementary in terms of the sample selection. We use only the spectroscopic sample with high redshift-quality flags, `zflag`/`zQ` = 3 or 4 for the VVDS/DEEP2. As a result, we obtain 253 LAS $K$–VVDS and 375 LAS $K$–DEEP2 galaxies, as well as 1084 DXS $K$–VVDS galaxies. We note that the redshift surveys are deep enough that essentially all the LAS $K$ sources could be sampled. Star/Galaxy Classification -------------------------- The $K$-band sources separate clearly into stars and galaxies on the $r - z$ versus $z - K$ diagram as shown in Fig. 3. The colours are measured with the 2.8-arcsec diameter aperture magnitudes. We define the demarcation between stars and galaxies along the minimum surface density on this diagram, i.e. the sources redder than $z - K = 0.52 (r - z) + 1.74$ are classified as galaxies. The additional criterion of $r - z < 4$ is set for galaxies to exclude cool dwarf stars from the sample. We obtain 259 082 galaxies with these classification criteria. The VVDS classification based on spectra confirms that the above scheme works very well, yielding a rate of misclassification (stars classified as galaxies and vice versa) of less than $\sim$1 per cent. The $K$-band sources without $r$- and/or $z$-band detections are excluded from the sample, since their extremely red colours suggest that they are mostly galaxies beyond $z = 1$. Actually, we find that more than 95 per cent of the LAS $K$–VVDS and LAS $K$–DEEP2 galaxies are detected in both $r$ and $z$ bands in any redshift bins at $z \le 1$. ![The $r - z$ versus $z - K$ diagram for a subset of $K$-band sources (black), on which the VVDS stars (blue) and galaxies (red) are superimposed. The green line represents the adopted star/galaxy classification criteria.[]{data-label="star_gal"}](f3_lores.eps){width="84mm"} We show the $K$-band differential number counts of the extracted galaxies in Fig. 4. They are in excellent agreement with previous measurements [@daddi00; @huang01] down to the limiting magnitude of $K_{\rm tot}$ = 17.9 mag after the detection-completeness correction is applied. This suggests that we have successfully constructed a well-defined sample of galaxies through the above processes. ![The $K$-band differential number counts of the extracted galaxies with (squares) and without (diamonds) detection-completeness correction. The error bars denote Poisson noise. The blue circles and the red triangles represent the measurements of @daddi00 and @huang01, respectively.[]{data-label="num_counts"}](f4.eps){width="84mm"} Redshift and Stellar mass Measurements ====================================== Photometric Redshift \[subsec:photoz\] -------------------------------------- We estimate the redshifts of galaxies from the observed broad-band colours, measured in the 2.8-arcsec diameter aperture, with the optimized template-fitting method following @ilbert06. We present a short summary of the method below, while the full description of the concept and details can be found in the above reference. We choose the DXS $K$–VVDS galaxies to optimize the spectral templates, leaving the LAS $K$–VVDS and LAS $K$–DEEP2 galaxies as the evaluation sets of the redshift measurements (Table 1). First, the DXS $K$–VVDS galaxies are classified into five spectral types, Ell, Sbc, Scd, Irr and starburst (SB) by least-$\chi^2$ fitting with the appropriate amounts of dust extinction assuming the extinction laws of the Small Magellanic Cloud (SMC) by @pei92 for Scd and Irr and that of starburst galaxies by @calzetti00 for SB. The initial spectral templates are taken from @cww80 for Ell, Sbc, Scd and Irr and from @kinney96 for SB. Then, for each filter $f$, we minimize the sum $$\psi^2 = \sum_{\rm galaxy}^{} \biggl( \frac{F_{\rm obs}^f - A \times F_{\rm model}^f - s^f}{\sigma_{\rm obs}^f} \biggr)^2,$$ where $F_{\rm obs}^f$ and $\sigma_{\rm obs}^f$ are the observed flux and its error in the filter $f$. The sum is taken over all the sample galaxies. The parameters $F_{\rm model}^f$ and $A$ represent the best-fitting template flux and its normalization factor taken from the initial least-$\chi^2$ fitting. The last term $s^f$ is a free parameter. While this term should be zero in the case of a completely random uncertainty in the photometry, we find that it has a non-zero value in every filter. These values are at most 0.05 mag and are comparable with the expected uncertainty in the photometric zero-point calibration. Sample Number Use$^{*1}$ ----------------------- -------- ------------ DXS $K$ – VVDS 1084 training LAS $K$ – VVDS$^{*2}$ 253 evaluation LAS $K$ – DEEP2 375 evaluation : Summary of the spectroscopic sample.[]{data-label="tab:spec_sample"} $^{*1}$Use in the photometric-redshift measurements.\ $^{*2}$Approximately 70% of the LAS $K$ – VVDS galaxies are also the members of the DXS $K$ – VVDS galaxies. The DXS $K$–VVDS galaxies are re-classified into five spectral-type groups with the terms $s^f$ considered. In each group, the observed broad-band fluxes of galaxies are converted to the rest frame according to the spectroscopic redshifts, after being normalized and de-reddened by the best-fitting normalization factor $A$ and the dust extinction. Since the galaxies have various redshifts, this conversion generates a continuous spectral energy distribution for each spectral type of galaxies over the relevant rest-frame wavelength range. We sort the rest-frame fluxes according to their wavelengths and bin them by groups of points, and connect the median flux in each bin to produce the optimized templates. We keep the extrapolations provided by the initial templates in ultraviolet and infrared wavelengths where no broad-band data are available. The starburst template is not optimized, in order to retain the emission lines in the template. Finally, these optimized templates are interpolated to produce a total of 62 templates, the first being Ell and the last being SB, to improve the sampling of the redshift–colour space. Below we define the spectral type of each galaxy using the best fits from among these 62 templates. The created spectral templates are fitted to the observed colours of the actual sample to measure their redshifts. We evaluate the measurement accuracy by applying the same procedure to the LAS $K$–VVDS and LAS $K$–DEEP2 galaxies, as well as the DXS $K$–VVDS galaxies. The DXS $K$–VVDS galaxies are reduced in number according to the LAS $K$-band detection completeness and are given additional random photometry errors in order to simulate the LAS $K$-band galaxies.We find that the photometric redshifts ($z_{\rm phot}$) are well correlated with the spectroscopic redshifts ($z_{\rm spec}$) as shown in Fig. 5, thanks to our wide and relatively fine wavelength coverage in the $u$ through $K$ bands. The deviation between the two (photometric and spectroscopic) measurements closely follows a Gaussian distribution with standard deviation $\sigma_{{\Delta}z/(1+z)} \sim 0.04$ (${\Delta}z = z_{\rm spec} - z_{\rm phot}$) for all three sets of the evaluation sample. Note that $\sim$70 per cent of the LAS $K$–VVDS galaxies are also members of the DXS $K$–VVDS galaxies and account for about a quarter of the sample used to build up the spectral templates, so that the LAS $K$–VVDS galaxies do not provide a completely independent test of the photometric-redshift accuracy. As a further test, we created another template set from the DXS $K$–VVDS galaxies omitting these LAS $K$–VVDS galaxies and repeated the photometric-redshift measurement. This test again gives $\sigma_{{\Delta}z/(1+z)} \sim 0.04$, which indicates that it is a robust estimate of the photometric-redshift uncertainty. We show the uncertainty as a function of redshift and stellar mass (as determined below) in Table 2. They are relatively large in the lowest and highest redshift bins for larger stellar mass classes, for which relatively small numbers of sample contribute to the spectral templates. We also show the uncertainty as a function of spectral type in Table 3, which suggests there is little variation of uncertainty among the different spectral types. ![Comparison of the photometric and spectroscopic redshift measurements of LAS $K$–VVDS (light blue), DXS $K$–VVDS (blue) and LAS $K$–DEEP2 (red) galaxies. The solid line shows the locus where the two measurements are identical.[]{data-label="photz_vs_specz"}](f5.eps){width="84mm"} ----------- ------------- ----------------- ------------- ------------- log $M_{\star}$ Redshift 10.0 – 10.5 10.5 – 11.0 11.0 – 11.5 11.5 – 12.0 0.2 – 0.4 0.037 (50) 0.046 (51) 0.043 ( 6) — ( 0) 0.4 – 0.6 0.040 (35) 0.037 (55) 0.039 (25) — ( 0) 0.6 – 0.8 0.021 (35) 0.030 (73) 0.032 (93) 0.034 (20) 0.8 – 1.0 0.033 (17) 0.038 (71) 0.046 (55) 0.047 ( 7) ----------- ------------- ----------------- ------------- ------------- : Uncertainty ($\sigma_{{\Delta}z/(1+z)}$) in photometric redshift.[]{data-label="tab:photo-z_err"} Note — Numbers in parentheses represent the number in the evaluation samples. Spectral type $\sigma_{{\Delta}z/(1+z)}$ --------------- ---------------------------- Ell – Sbc 0.035 (386) Sbc – Scd 0.037 (171) Scd – Irr 0.036 ( 35) Irr – SB — ( 0) : Uncertainty ($\sigma_{{\Delta}z/(1+z)}$) in photometric redshift.[]{data-label="tab:photo-z_err2"} Note — Numbers in parentheses represent the number in the evaluation samples. Stellar Mass \[subsec:stellarmass\] ----------------------------------- The stellar masses ($M_{\star}$) of galaxies are determined by fitting to the observed colours the stellar population synthesis models of @bc03. The aperture magnitudes ($m_{\rm ap}$) of the 4.5-, 3.3-, 2.8- and 2.6-arcsec apertures are used for the fitting of galaxies with photometric redshifts $z$ = 0.2–0.4, 0.4–0.6, 0.6–0.8, and 0.8–1.0, respectively, so that we sample the stellar populations consistently within the central $\sim$20 kpc of all the galaxies. The resultant stellar mass is then scaled by $10^{0.4 (K_{\rm ap}-K_{\rm tot})}$ to correct for aperture loss. We adopt the standard configurations with the Padova 1994 stellar evolutionary tracks and BaSel 3.1 spectral library for the @bc03 models. We assume three values of metallicity: 0.2, 1 and 2.5 $Z_{\odot}$, where $Z_{\odot}$ is the solar metallicity. The star-formation history is assumed to take the exponentially declining form $\tau^{-1}$ exp($-t/\tau$), where the $e$-folding time $\tau$ is a free parameter, with the @salpeter55 initial mass function (IMF). Our stellar mass estimates can be approximately converted to those with another commonly used IMF, that of @chabrier03, by adding $\sim$0.25 dex. Other free parameters are the age $t$ of the stellar population and the colour excess $E(B-V)$ due to the dust extinction of the stellar radiation, following the SMC extinction curve of @pei92. These parameters are varied over the plausible ranges of 10 Myr $\le \tau \le$ 10 Gyr ($\Delta$log $\tau$ = 0.2), 10 Myr $\le t \le$ 10 Gyr ($\Delta$log $t$ = 0.1), and $0.0 \le E(B-V) \le 0.5$ mag ($\Delta$$E(B-V)$ = 0.05), and the best-fitting parameter set is searched for by the least-$\chi^2$ method for each galaxy (the values in parentheses represent the grid intervals). The additional error of 0.05 mag is added in quadrature to all band magnitudes in the fitting in order to take into account the uncertainty in the photometric zero-point calibration. We derive two kinds of stellar mass for the spectroscopic sample, i.e. the stellar mass with spectroscopic redshifts ($M_{\star, spec}$) and the stellar mass with photometric redshifts ($M_{\star, phot}$). The difference between the two measures, $\Delta {\rm log} M_{\star} = {\rm log} M_{\star, spec} - {\rm log} M_{\star, phot}$, is found to be clearly correlated with the photometric redshift deviation ${\Delta}z = z_{\rm spec} - z_{\rm phot}$. Such a correlation is expected, since a larger $z_{\rm phot}$ leads to a larger estimate of the galaxy luminosity, which then leads to a larger estimate of stellar mass $M_{\star, phot}$. The observed relation between ${\Delta}z$ and $\Delta {\rm log} M_{\star}$ is actually quite consistent with this expected correlation. Another expected cause of the correlation between ${\Delta}z$ and $\Delta {\rm log} M_{\star}$ comes from the fact that larger estimates of $z_{\rm phot}$ lead to systematically shorter rest-frame wavelengths to which each of the observing wavebands corresponds. After removal of the above first component of the systematic correlation, we found marginal evidence for the second correlation in our sample, which is $\Delta {\rm log} M_{\star}$ = $+/-$ ($0.04 \pm 0.07$) when ${\Delta}z$ is negative/positive. In addition, we consider the uncertainty associated with the least-$\chi^2$ model fitting. It is evaluated by the 1$\sigma$ confidence surface of the $\chi^2$ distributions in the fitting parameter space. We show the total amplitudes of the stellar mass uncertainty ($\sigma_{\Delta {\rm log} M_{\star}}$) as a function of redshift and stellar mass in Table 4. Those as a function of spectral type are shown in Table 5. The above estimates of error amplitudes and the correlations between ${\Delta}z$ and $\Delta {\rm log} M_{\star}$ are taken into account in the Monte Carlo simulation presented below. Note that further different assumptions on the stellar mass estimation, such as different stellar population synthesis models and different IMF, can cause additional uncertainty in the derived properties of galaxies. We will address this issue in Section 5. ----------- ------------- ----------------- ------------- ------------- log $M_{\star}$ Redshift 10.0 – 10.5 10.5 – 11.0 11.0 – 11.5 11.5 – 12.0 0.2 – 0.4 0.19 (50) 0.20 (51) 0.22 ( 6) — ( 0) 0.4 – 0.6 0.19 (35) 0.17 (55) 0.16 (25) — ( 0) 0.6 – 0.8 0.12 (35) 0.16 (73) 0.17 (93) 0.19 (20) 0.8 – 1.0 0.14 (17) 0.17 (71) 0.19 (55) 0.25 ( 7) ----------- ------------- ----------------- ------------- ------------- : Uncertainty ($\sigma_{\Delta {\rm log} M_{\star}}$) in stellar mass.[]{data-label="tab:logM_err"} Note — Numbers in parentheses represent the number in the evaluation samples. Spectral type $\sigma_{\Delta {\rm log} M_{\star}}$ --------------- --------------------------------------- Ell – Sbc 0.16 (386) Sbc – Scd 0.17 (171) Scd – Irr 0.17 ( 35) Irr – SB — ( 0) : Uncertainty ($\sigma_{\Delta {\rm log} M_{\star}}$) in stellar mass[]{data-label="tab:logM_err2"} Note — Numbers in parentheses represent the number in the evaluation samples. Results ======= We define our massive galaxy sample using two stellar mass classes, i.e. $10^{11.0-11.5} M_{\odot}$ galaxies with $10^{11.0} M_{\odot} < M_{\star} < 10^{11.5} M_{\odot}$ and $10^{11.5-12.0} M_{\odot}$ galaxies with $10^{11.5} M_{\odot} < M_{\star} < 10^{12.0} M_{\odot}$. The galaxies are grouped into four redshift bins, $z$ = 0.2 – 0.4, 0.4 – 0.6, 0.6 – 0.8, and 0.8 – 1.0. Total numbers included in the sample are summarized in Table 6. The median photometry errors in the $r$-, $z$-, and $K$-band aperture magnitudes and in the $K$-band total magnitudes ($r_{\rm ap}$, $z_{\rm ap}$, $K_{\rm ap}$, $K_{\rm tot}$) are also listed. The aperture magnitude errors are generally smaller than the typical uncertainty in the photometric zero-point calibration ($\sim$0.05 mag). ----------- ------------------- -------------------------------- ------------------- ----------------------------------- log $M_{\star}$ = 11.0 – 11.5 log $M_{\star}$ = 11.5 – 12.0 Redshift Number Photometry Error$^*$ Number Photometry Error$^*$ 0.2 – 0.4 9,720 ($<$0.01, $<$0.01, 0.01, 0.05) 1,408 ($<$0.01, $<$0.01, $<$0.01, 0.03) 0.4 – 0.6 15,300 ($<$0.01, $<$0.01, 0.02, 0.08) 572 ($<$0.01, $<$0.01, 0.01, 0.04) 0.6 – 0.8 18,582 (0.02, 0.02, 0.03, 0.13) 815 (0.01, 0.01, 0.02, 0.09) 0.8 – 1.0 12,371 (0.04, 0.02, 0.04, 0.16) 613 (0.03, 0.02, 0.03, 0.12) ----------- ------------------- -------------------------------- ------------------- ----------------------------------- Note ($^*$) — Median photometry errors in the r-, z- and K-band aperture magnitudes and in the K-band total magnitudes ($r_{\rm ap}$, $z_{\rm ap}$, $K_{\rm ap}$, $K_{\rm tot}$). We show the differential number counts of the massive galaxies in Fig. 6. It shows that the number-count distributions of the $10^{11.5-12.0} M_{\odot}$ galaxies have faint-end drop-offs at magnitudes brighter than the limiting magnitude, which assures the near-complete detection of this population. On the other hand, the faintest of the $10^{11.0-11.5} M_{\odot}$ galaxies at high redshifts ($z > 0.6$) fall below the limiting magnitude, and are thus left uncounted. In order to estimate the lost fraction of $10^{11.0-11.5} M_{\odot}$ galaxies at these redshifts, we derive the detection completeness specifically for these galaxies as follows. We take each of the $10^{11.0-11.5} M_{\odot}$ galaxies in the lowest redshift bin and assign random redshifts in the $0.6 < z < 0.8$ and $0.8 < z < 1.0$ ranges. The galaxies are dimmed and reduced in apparent size according to the assigned redshifts, placed on random positions of the LAS $K$-band images, and then extracted by `SOURCE EXTRACTOR` in the same way as the actual sample sources are detected. The recovery rate of the embedded objects as a function of their magnitudes gives the detection completeness of the galaxies, which we find is significantly better than that of the whole sample derived before. The 50 per cent detection completeness is actually achieved at $K_{\rm tot}$ = 18.6 mag instead of $K_{\rm tot}$ = 17.9 mag, and almost all galaxies brighter than $K_{\rm tot}$ = 17.9 mag are detected. With the new detection-completeness function taken into account, the fractions of $10^{11.0-11.5} M_{\odot}$ galaxies fainter than the formal limiting magnitude ($K$ = 17.9 mag, thus uncounted) are 11 and 16 per cent at $z$ = 0.6–0.8 and 0.8–1.0, respectively ![The differential number counts of $10^{11.0-11.5} M_{\odot}$ (green) and $10^{11.5-12.0} M_{\odot}$ (orange) galaxies at $z$ = 0.2–0.4 (top left), 0.4–0.6 (top right), 0.6–0.8 (bottom left) and 0.8–1.0 (bottom right). The black dashed lines represent our formal limiting magnitude of $K_{\rm tot}$ = 17.9 mag.[]{data-label="z_num2_4"}](f6.eps){width="84mm"} We estimate the uncertainty in the measured numbers of massive galaxies by a Monte Carlo simulation, as follows. First, we generate a mock galaxy catalogue containing 10 galaxies for each stellar mass ($\Delta$log$M_{\star} = 0.1$) and redshift ($\Delta{z} = 0.02$) bin in the ranges $8.0 <$ log $M_{\star}$ $< 13.0$ and $0.0 < z < 1.4$, where the numbers in parentheses represent the bin widths. Each galaxy is assigned a weighting factor corresponding to the number density of galaxies, following the galaxy mass function of @cole01. Next the redshifts of galaxies are given perturbations following the measured uncertainty of the actual sample. The stellar masses are also given perturbations, correlated with the photometric redshift perturbations as explored in Section 3.2. Then the mock galaxies are weighted by their weighting factors and redistributed, and counted in the redshift and stellar mass bins to obtain the output mass function. We repeat the calculation 100 times, varying the random components. Fig. 7 shows the results of the simulation. We observe both systematic and random components in the resultant error estimates. The systematic component is evident in $M_{\star} > 10^{12.0} M_{\odot}$ bins. This is the so-called Eddington bias [@eddington13], caused by the steep slope of the high end of the mass function; simply put, a small portion of the less massive, much more numerous galaxies could contaminate the more massive classes owing to measurement errors, which significantly alters the steep part of the mass function. This is why we limit our massive galaxy sample to those with $M_{\star} < 10^{12.0} M_{\odot}$; the systematic increases in number are found to be insignificant for our mass ranges, i.e. negligible for the $10^{11.0-11.5} M_{\odot}$ galaxies and 40–60 per cent for the $10^{11.5-12.0} M_{\odot}$ galaxies. The output number densities from the 100 repeated calculations scatter around the systematic components, which yields the random components of the measurement error. The standard deviations of the scatter are $\sim$5 per cent and $\sim$12 per cent of the numbers of $10^{11.0-11.5} M_{\odot}$ and $10^{11.5-12.0} M_{\odot}$ galaxies, respectively. Note that the current estimate is not perfect, since we assume a non-evolving galaxy stellar mass function at $0 < z < 1$, while our results show clear signs of its evolution (see below). We will discuss this issue further in the following section. ![The input (black line) and output (100 red lines) galaxy mass functions of the Monte Carlo simulation at $z$ = 0.2–0.4 (top left), 0.4–0.6 (top right), 0.6–0.8 (bottom left) and 0.8–1.0 (bottom right). The dashed lines show the mass ranges in which our massive galaxy sample is defined.[]{data-label="montecarlo"}](f7.eps){width="84mm"} Another source of uncertainty comes from cosmic variance. We estimate cosmic variance by dividing our sample into five subfields along the right ascension, $\sim$11 deg$^2$ each, and then calculating the fractional variation of the measured numbers of massive galaxies in these subfields. We find that the fractional variations are $<$10 per cent and $<$20 per cent for the $10^{11.0-11.5} M_{\odot}$ and $10^{11.5-12.0} M_{\odot}$ galaxies, respectively, in each of the four redshift bins. Considering that the total fields are five times larger than the subfields, we conclude that cosmic variance could affect the measured number of massive galaxies in each redshift bin by up to 5 per cent and 10 per cent for the two classes of galaxies, respectively. The above estimates are roughly consistent with the theoretical predictions provided by @somerville04. We show our results with regard to the number-density measurements in Table 7, and plot them along with the measurements for the local Universe [@cole01] in Fig. 8. The error bars take into account all uncertainties considered above, as well as the Poisson noises. The number densities of the $10^{11.5-12.0} M_{\odot}$ galaxies are corrected for Eddington bias, although this correction has little significance for our final conclusions. The local densities were normalized to take into account the assumptions of @cole01 with regard to stellar population synthesis that differ from ours; the major difference is that they assume a constant formation redshift of galaxies at $z_f = 20$, while we vary the age of the stellar population as a free parameter. The normalization factor, $\sim0.4$, is derived by applying their assumption to our $10^{11.0-11.5} M_{\odot}$ galaxies in the lowest redshift bin $z$ = 0.2–0.4. As seen in the figure, we find that the most massive ($10^{11.5-12.0} M_{\odot}$) galaxies have experienced rapid evolution in number since $z = 1$. On the other hand, the number densities of the less massive ($10^{11.0-11.5} M_{\odot}$) systems show a rather mild evolution during the same period. ----------- ----------------- ----------------- log $M_{\star}$ Redshift 11.0 – 11.5 11.5 – 12.0 0.2 – 0.4 10.9 $\pm$ 0.8 0.98 $\pm$ 0.15 0.4 – 0.6 7.8 $\pm$ 0.6 0.19 $\pm$ 0.03 0.6 – 0.8 6.8 $\pm$ 0.5 0.17 $\pm$ 0.03 0.8 – 1.0 3.7 $\pm$ 0.3 0.11 $\pm$ 0.02 ----------- ----------------- ----------------- : Number densities of the massive galaxies.[]{data-label="tab:number_densities"} Note — Number densities are given in units of $10^{-4}$ Mpc$^{-3}$ log $M_{\star}^{-1}$. ![The number densities of massive galaxies versus the age of the Universe or redshift. The green and orange symbols and lines represent the $10^{11.0-11.5} M_{\odot}$ and $10^{11.5-12.0} M_{\odot}$ galaxies, respectively. The filled circles represent our measurements while the open circles represent measurements in the local Universe [@cole01]. The solid lines represent the predictions of the Millennium Simulation with the semi-analytic galaxy formation model (see text).[]{data-label="number_densities"}](f8.eps){width="84mm"} Discussion \[sec:discuss\] ========================== The measured number density evolution of massive galaxies shows clear signs of the hierarchical evolution of these systems. Such a galaxy evolution scenario is predicted in the latest galaxy formation models based on $\Lambda$CDM theory. In Fig. 8 we overlay the predictions of the Millennium Simulation (Lemson et al. 2006), the largest numerical simulation to date based on $\Lambda$CDM theory, with the semianalytic galaxy formation model of @delucia07, scaled to fit to the local observations (by a factor of 0.4). The stellar mass of the Millennium model has been shifted by $+$0.25 dex in order to correct for the different IMFs adopted (the model adopts the @chabrier03 IMF). The observed hierarchical pattern of evolution is consistent with the prediction of the model, while we find some discrepancies between the observation and the model (e.g. the $10^{11.0-11.5} M_{\odot}$ galaxies at $z$ = 0.8–1.0). This indicates that the basic idea of the bottom-up construction of galaxy systems is valid at least for the most massive galaxies with $M_{\star} > 10^{11} M_{\odot}$. Meanwhile, we point out that our measurements plotted in Fig. 8 seem to have an apparently unnatural feature at $z$ = 0.2–0.4, where the number densities are significantly high relative to the overall trend considering their estimated errors, implying residual systematic effects from the Eddington bias. We check this by investigating the systematic increases in the number of massive galaxies in the spectroscopic sample, from those obtained with the spectroscopic redshifts to those with photometric redshifts (we consider only those redshift and stellar mass classes with more than five spectroscopic sources). The above Monte Carlo simulation shows that the use of photometric redshifts (and the consequent stellar mass fluctuations) is the dominant source of the Eddington bias. We list the measured systematic increases in Table 8 separately for the VVDS and DEEP2 galaxies, since the amplitudes of the Eddington bias are subject to the redshift distribution of sources. The DXS $K$–VVDS galaxies have been given additional photometry errors and detection incompleteness in order to simulate the LAS $K$ galaxies. The systematic increases estimated in the Monte Carlo simulation are also listed in Table 8. The table shows that the systematic increases measured in the spectroscopic sample are mostly less than the estimates from the Monte Carlo simulation, while the one for the VVDS $10^{11.0-11.5} M_{\odot}$ galaxies at $z = 0.2 - 0.4$ is exceptionally high. While it is based on a small sample, it provides marginal evidence that the Eddington bias is unexpectedly significant in this lowest redshift bin, due to some unquantified sources of systematic uncertainty. In that case, the number densities measured at $z$ = 0.2–0.4 should be regarded as the upper limits. We also note that we adopt a non-evolving galaxy stellar mass function in the Monte Carlo simulation, while our results suggest the steepening of the high end of the mass function toward high redshifts; thus the estimated amount of Eddington bias should be regarded as a lower limit. When such a steepening of the mass function is taken into account in the correction of the Eddington bias, we obtain even lower numbers of galaxies in the more massive classes at higher redshifts, which would further strengthen our conclusion. ----------- ------------------- -------------- -------------- ------------------- ------------- -------------- log $M_{\star}$ = 11.0 – 11.5 log $M_{\star}$ = 11.5 – 12.0 Redshift MCS$^{*2}$ VVDS$^{*3}$ DEEP2$^{*3}$ MCS$^{*2}$ VVDS$^{*3}$ DEEP2$^{*3}$ 0.2 – 0.4 $<$ 0.1 0.7 (6) - (0) 0.6 - (0) - (0) 0.4 – 0.6 $<$ 0.1 $<$ 0.1 (23) - (2) 0.5 - (0) - (0) 0.6 – 0.8 0.1 0.1 (21) $<$ 0.1 (72) 0.6 0.3 (7) 0.3 (13) 0.8 – 1.0 $<$ 0.1 - (2) $<$ 0.1 (53) 0.4 - (0) 0.1 (7) ----------- ------------------- -------------- -------------- ------------------- ------------- -------------- $^{*1}$Rates of increase (increased amounts divided by the original values) of the number densities are listed.\ $^{*2}$Estimates from the Monte Carlo simulation.\ $^{*3}$Estimates from the VVDS and DEEP2 spectroscopic samples (the number in the samples is shown in the parentheses).\ Recently, @marchesini09 provided a detailed analysis of random and systematic uncertainties affecting the galaxy stellar mass function. They adopt 14 different stellar mass estimations with different combinations of metallicity, dust extinction law, stellar population synthesis models and IMF, and show that the derived number density of galaxies in a given stellar mass bin could be altered by up to 1 dex. The ’bottom-light’ IMFs in particular, with a deficit of low-mass stars relative to a standard @chabrier03 IMF, give significantly different results for the stellar mass function from other classical IMFs. A more top-heavy IMF at higher redshifts is actually suggested by, for example, @dave08 and @vandokkum08. Thus we point out that our results are subject to a systematic change of these stellar-population properties during $0 < z< 1$. Future improvements in stellar population models based on new observations are eagerly awaited to overcome these large uncertainties inherent in stellar mass measurements. In order to probe the star-forming properties of massive galaxies, here we investigate their rest-frame optical colours. Nearby galaxies are known to show a clear bimodality in the optical colour distribution, in which early-type galaxies form a narrow red sequence that is separated from blue star-forming populations by a valley of the galaxy distribution [e.g., @strateva01; @hogg03]. A similar bimodality is observed out to $z > 1$ [e.g., @lin99; @im02; @bell04]. We calculate the rest-frame $U$- and $V$-band magnitudes of massive galaxies by $k$-correcting the nearest observed $r$, $i$ or $z$-band magnitudes, where the amounts of the $k$-corrections are estimated from the best-fitting spectral population synthesis models. We show the resultant colour distributions in Fig. 9. ![$U - V$ colour distributions of the massive galaxies at $z$ = 0.2–0.4 (top left), 0.4–0.6 (top right), 0.6–0.8 (bottom left) and 0.8–1.0 (bottom right). The grey, green and orange lines represent the $< 10^{11.0} M_{\odot}$, $10^{11.0-11.5} M_{\odot}$ and $10^{11.5-12.0} M_{\odot}$ galaxies, respectively. The dotted lines show the demarcation between the blue and red populations.[]{data-label="cmd"}](f9.eps){width="84mm"} As Fig. 9 shows, the less massive ($< 10^{11.0} M_{\odot}$) galaxies show a clear colour bimodality, as expected, with the peak colour of the red sequence at $U - V \sim 1.2$ and a valley of galaxy distributions at $U - V \sim 1.0$ in all redshift bins. Compared with these galaxies, massive ($> 10^{11.0} M_{\odot}$) galaxies are apparently dominated by the red population, with conspicuous peaks at $U - V \sim 1.2$. We divide the galaxies into blue and red populations at $U - V = 1.0$ and calculate the blue fractions (fractions of the blue population). The associated errors are estimated by repeatedly giving random fluctuations to the $U - V$ colours, taking into account the uncertainties in the photometry and the amount of $k$-correction, and then re-measuring the blue fractions. We find that the fluctuated blue fractions are systematically larger than the original values, since there is a greater red population than blue around the demarcation $U - V = 1.0$, and correct for the effect (the correction amounts are included in the final errors). We plot the measured blue fractions as a function of redshift and stellar mass in Fig. 10, and also list them in Table 9. One can see that the blue fractions are significantly lower in more massive galaxies, and that the fractions in massive systems decrease toward the local Universe. The blue fractions in $< 10^{11.0} M_{\odot}$ galaxies increase from z = 0.4–0.6 to 0.2–0.4 because the dominant population within the sample shifts to bluer, less massive galaxies toward the local Universe due to the fainter detection limit. In fact we observe a decreasing blue fraction toward the local Universe, as seen in massive galaxies, if we take the subsample with stellar mass $10^{10.5} M_{\odot}< M_{\star} < 10^{11.0} M_{\odot}$ (see Table 9). As discussed above, the $10^{11.0-11.5} M_{\odot}$ galaxies at $z$ = 0.2–0.4 and the $10^{11.5-12.0} M_{\odot}$ galaxies in all redshift bins could have considerable fractions of contamination from less massive galaxies, which likely have bluer $U - V$ colours. Actually, investigating the spectroscopic sample shows that the contamination makes the mean $U - V$ colours bluer by up to 0.1 mag. Thus the true blue fractions in the above classes of galaxies could be even smaller than the present measurements. The lower blue fractions in more massive galaxies and the decreasing trend toward the local Universe implies major star formation at higher redshifts, which is in line with ’downsizing’ of the star formation [e.g., @cowie96]. The above measurements suggest that the majority of massive galaxies are fairly quiescent, while of the rest a considerable fraction of galaxies are experiencing active star formation, especially at higher redshifts ($\sim$30 per cent at $z \sim 1$). Such active star formation in massive galaxies is also reported in @conselice07, who find that nearly half of their massive ($M_{\star} > 10^{11} M_{\odot}$) galaxies at $0.4 < z < 1.4$ are detected in the [*Spitzer Space Telescope*]{}/MIPS 24-$\mu$m band and the average star-formation rate amounts to $\sim50 M_{\odot}$ yr$^{-1}$. Our measurements of blue fractions indicate that the star-formation activity in massive galaxies is gradually quenched toward the local Universe, leaving the most massive galaxies on the red sequence. Star-formation quenching processes above a certain stellar mass limit are actually proposed, such as the internal feedback of mass assembly caused by active galactic nuclei [e.g., @silk98; @granato04; @springel05]. The above scenario is consistent with our primary results that the active bottom-up formation of massive galaxies is going on during $0 < z < 1$. ----------- ----------------------------------- ----------------- ----------------- log $M_{\star}$ Redshift $<$ 11.0 (10.5 – 11.0) 11.0 – 11.5 11.5 – 12.0 0.2 – 0.4 0.54 $\pm$ 0.01 (0.36 $\pm$ 0.01) 0.12 $\pm$ 0.05 $<$ 0.06 0.4 – 0.6 0.47 $\pm$ 0.01 (0.43 $\pm$ 0.01) 0.25 $\pm$ 0.03 0.18 $\pm$ 0.03 0.6 – 0.8 0.51 $\pm$ 0.01 (0.43 $\pm$ 0.01) 0.28 $\pm$ 0.02 0.19 $\pm$ 0.06 0.8 – 1.0 0.59 $\pm$ 0.01 (0.56 $\pm$ 0.01) 0.29 $\pm$ 0.02 0.21 $\pm$ 0.03 ----------- ----------------------------------- ----------------- ----------------- : The blue population fraction.[]{data-label="tab:frac_blue"} ![The fraction of the blue population in $< 10^{11.0} M_{\odot}$ (grey solid), $10^{11.0-11.5} M_{\odot}$ (green) and $10^{11.5-12.0} M_{\odot}$ (orange) galaxies, respectively, as a function of redshift.[]{data-label="cmd2"}](f10.eps){width="84mm"} Finally, we comment on the compatibility of the present results with previous studies. There are a number of studies of luminous red galaxies (LRGs) at $z < 1$ [e.g., @brown07; @brown08; @cool08] covering up to $\sim$10 deg$^2$. Authors consistently suggest that the LRGs show little evolution in number density since $z \sim 1$. However, the mass-to-optical luminosity ratio of galaxies has a significant scatter even for the massive systems, so that galaxies with a certain stellar mass are not quite the equivalent population of galaxies with a certain optical luminosity. This leads to a consequence of most significance for the most massive galaxies: a small portion of the less massive, much more numerous galaxies could contaminate the luminous class of galaxies if their mass-to-luminosity ratios were slightly less than the average, and thus could easily dominate the luminous population. Therefore a subtle (in absolute amplitude) change in the number of most massive galaxies could be drowned out in the measured evolution of the LRGs. Studies also exist of massive galaxies at $z < 1$ with stellar mass measurements based on infrared photometry [e.g., @conselice07; @ilbert09], although these studies cover a much smaller field of view ($\la$ 1.5 deg$^2$) than ours. In contrast to the present results, they report little evolution in number of the most massive ($> 10^{11.5} M_{\odot}$) galaxies. At least part of the discrepancy could be due to small-number statistics and cosmic variance. Actually, while we find $\sim$1000 samples of the most massive galaxies in each redshift bin from our 55.2 deg$^2$, the number of samples observed over $\sim$1.5 deg$^2$ should be only $\sim$30. We estimate the effects of cosmic variance by dividing our total field into small subfields, each covering $\sim$ 1.5 deg$^2$, and measure the number-density fluctuations of massive galaxies among the subfields. As a result, we find that the number densities of the most massive galaxies measured over $\sim$ 1.5 deg$^2$ can fluctuate by up to a factor of a few. However, we are not sure whether the above uncertainties alone can account for the discrepancy between the present results and previous ones. The number-density measurements at the steep high end of the galaxy stellar mass function could be heavily affected by contamination arising from less massive galaxies, thus quite accurate analysis is required to unveil the subtle evolution of the most massive galaxies. In essence, our measurements provide a unique opportunity to investigate the mean properties and evolution of the most massive galaxies, owing to the reliable estimates of photometric redshifts and stellar masses conducted over an unprecedentedly large field of view. What is observationally clear is that we have discovered a substantial deficit of the most massive galaxies out to $z = 1$ compared with the local Universe. The analysis of the rest-frame $U - V$ colour distributions indicates that star-formation activity might be responsible for the active build-up of these systems, while it is possible that a so-called dry merger [e.g., @bell04; @vandokkum05] is the main driver of the evolution. Actually, some observations suggest that a substantial fraction of massive early-type galaxies go through active evolution in terms of the galaxy structure as well as the star formation since $z \sim$ 1 [e.g., @treu05; @vanderwel08]. The present results provide crucial evidence of hierarchical galaxy formation, the missing piece of observation required to chart a course for future theoretical models based on $\Lambda$CDM theory. Summary ======= We present an analysis of $\sim$60 000 massive galaxies with stellar masses $10^{11} M_{\odot} < M_{\star} < 10^{12} M_{\odot}$ in an unprecedentedly large field of view of 55.2 deg$^2$. The galaxies are drawn from the UKIDSS Large Area Survey K-band images on the SDSS southern equatorial stripe. We have created deep-stacked $u$, $g$, $r$, $i$ and $z$-band images from the SDSS Supplemental and Supernova Survey image frames, which results in $\sim$90 per cent counterparts of the $K$-band sources in the $g$, $r$, $i$ and $z$ bands. We also exploit the redshift surveys conducted on the SDSS southern equatorial stripe, namely the VIMOS-VLT Deep Survey and the DEEP2 Redshift Survey, in order to obtain accurate photometric redshifts and associated uncertainties for the galaxies. Stellar masses are estimated by comparing the observed broad-band colours with stellar population synthesis models. In each of the redshift bins $z$ = 0.2–0.4, 0.4–0.6, 0.6–0.8 and 0.8–1.0, we obtain $\sim$10 000 and $\sim$1 000 galaxies with stellar masses $10^{11.0} M_{\odot} < M_{\star} < 10^{11.5} M_{\odot}$ and $10^{11.5} M_{\odot} < M_{\star} < 10^{12.0} M_{\odot}$, respectively. The galaxies are almost completely detected out to $z = 1$, and form by far the largest sample of massive galaxies reaching to the Universe at about half its present age. We find that the most massive ($10^{11.5} M_{\odot} < M_{\star} < 10^{12.0} M_{\odot}$) galaxies have experienced rapid growth in number since $z = 1$, while the number densities of less massive systems show rather mild evolution. Such a hierarchical trend of evolution is consistent with the predictions of the current semi-analytic galaxy formation model based on $\Lambda$CDM theory. While the majority of the massive galaxies are red-sequence populations, we find that a considerable fraction are blue star-forming galaxies. The blue fraction is less in more massive systems and decreases toward the local Universe, leaving the red, most massive galaxies at low redshifts, which further supports the idea of active bottom-up formation of these populations during $0 < z <1$. The present results provide strong evidence that galaxy formation proceeds in a hierarchical way, and place stringent observational constraints on future theoretical models. Acknowledgments {#acknowledgments .unnumbered} =============== We are grateful to K. Shimasaku, K. Kohno, J. Makino, N. Yasuda and N.Yoshida for insightful discussions and suggestions. We thank the referee for many useful comments that have helped to improve this paper. YM acknowledges Grant-in-Aid from the Research Fellowships of the Japan Society for the Promotion of Science (JSPS) for Young Scientists. This work was supported by Grants-in-Aid for Scientific Research (17104002, 21840027), Specially Promoted Research (20001003) and the Global COE Program of Nagoya University ’Quest for Fundamental Principles in the Universe (QFPU)’ from JSPS and MEXT of Japan. This publication makes use of data products from the Two- Micron All-Sky Survey, which is a joint project of the University of Massachusetts and the Infrared Processing and Analysis Center/California Institute of Technology, funded by the National Aeronautics and Space Administration and the National Science Foundation. IRAF is distributed by the National Optical Astronomy Observatories, which are operated by the Association of Universities for Research in Astronomy, Inc., under cooperative agreement with the National Science Foundation. Funding for the SDSS and SDSS-II has been provided by the Alfred P. Sloan Foundation, the Participating Institutions, the National Science Foundation, the US Department of Energy, the National Aeronautics and Space Administration, the Japanese Monbukagakusho, the Max Planck Society and the Higher Education Funding Council for England. The SDSS Web Site is http://www.sdss.org/. The SDSS is managed by the Astrophysical Research Consortium for the Participating Institutions. The Participating Institutions are the American Museum of Natural History, Astrophysical Institute Potsdam, University of Basel, University of Cambridge, CaseWestern Reserve University, University of Chicago, Drexel University, Fermilab, the Institute for Advanced Study, the Japan Participation Group, Johns Hopkins University, the Joint Institute for Nuclear Astrophysics, the Kavli Institute for Particle Astrophysics and Cosmology, the Korean Scientist Group, the Chinese Academy of Sciences (LAMOST), Los Alamos National Laboratory, the Max-Planck-Institute for Astronomy (MPIA), the Max-Planck-Institute for Astrophysics (MPA), New Mexico State University, Ohio State University, University of Pittsburgh, University of Portsmouth, Princeton University, the United States Naval Observatory and the University of Washington. [99]{} Adelman-McCarthy J. K., et al., 2008, ApJS, 175, 297 Bell E. F., et al., 2004, ApJ, 608, 752 Bertin E., Arnouts S., 1996, A&AS, 117, 393 Borch A., et al., 2006, A&A, 453, 869 Bower R. G., Lucey J. R., Ellis R. S., 1992, MNRAS, 254, 601 Brown M. J. I., Dey A., Jannuzi B. T., Brand K., Benson A. J., Brodwin M., Croton D. J., Eisenhardt P. R., 2007, ApJ, 654, 858 Brown M. J. I., et al., 2008, ApJ, 682, 937 Bruzual G., Charlot S., 2003, MNRAS, 344, 1000 Calzetti D., Armus L., Bohlin R. C., Kinney A. L., Koornneef J., Storchi-Bergmann T., 2000, ApJ, 533, 682 Chabrier G., 2003, PASP, 115, 763 Cimatti A., Daddi E., Renzini A., 2006, A&A, 453, L29 Cole S., et al., 2001, MNRAS, 326, 255 Coleman G. D., Wu C.-C., Weedman D. W., 1980, ApJS, 43, 393 Conselice C. J., et al., 2007, MNRAS, 381, 962 Cool R. J., et al., 2008, ApJ, 682, 919 Cowie L. L., Songaila A., Hu E. M., Cohen J. G., 1996, AJ, 112, 839 Daddi E., Cimatti A., Pozzetti L., Hoekstra H., R[ö]{}ttgering H. J. A., Renzini A., Zamorani G., Mannucci F., 2000, A&A, 361, 535 Davis M., et al., 2003, SPIE, 4834, 161 Dav[é]{} R., 2008, MNRAS, 385, 147 De Lucia G., Blaizot J., 2007, MNRAS, 375, 2 Eddington A. S., 1913, MNRAS, 73, 359 Eggen O. J., Lynden-Bell D., Sandage A. R., 1962, ApJ, 136, 748 Ellis R. S., Smail I., Dressler A., Couch W. J., Oemler A. J., Butcher H., Sharples R. M., 1997, ApJ, 483, 582 Faber S. M., et al., 2007, ApJ, 665, 265 Frieman J. A., et al., 2008, AJ, 135, 338 Granato G. L., De Zotti G., Silva L., Bressan A., Danese L., 2004, ApJ, 600, 580 Hogg D. W., et al., 2003, ApJ, 585, L5 Huang J.-S., et al., 2001, A&A, 368, 787 Ilbert O., et al., 2006, A&A, 457, 841 Ilbert O., et al., 2009, arXiv, arXiv:0903.0102 Im M., et al., 2002, ApJ, 571, 136 Kinney A. L., Calzetti D., Bohlin R. C., McQuade K., Storchi-Bergmann T., Schmitt H. R., 1996, ApJ, 467, 38 Larson R. B., 1975, MNRAS, 173, 671 Lawrence A., et al., 2007, MNRAS, 379, 1599 Le F[è]{}vre O., et al., 2005, A&A, 439, 845 Lemson G., the Virgo Consortium, 2006, astro, arXiv:astro-ph/0608019 Lin H., Yee H. K. C., Carlberg R. G., Morris S. L., Sawicki M., Patton D. R., Wirth G., Shepherd C. W., 1999, ApJ, 518, 533 Marchesini D., van Dokkum P. G., F[ö]{}rster Schreiber N. M., Franx M., Labb[é]{} I., Wuyts S., 2009, ApJ, 701, 1765 Matsuoka Y., et al., 2008, ApJ, 685, 767 Pei Y. C., 1992, ApJ, 395, 130 Salpeter E. E., 1955, ApJ, 121, 161 Silk J., Rees M. J., 1998, A&A, 331, L1 Skrutskie M. F., et al., 2006, AJ, 131, 1163 Somerville R. S., Lee K., Ferguson H. C., Gardner J. P., Moustakas L. A., Giavalisco M., 2004, ApJ, 600, L171 Springel V., Di Matteo T., Hernquist L., 2005, ApJ, 620, L79 Strateva I., et al., 2001, AJ, 122, 1861 Treu T., et al., 2005, ApJ, 633, 174 van der Wel A., Holden B. P., Zirm A. W., Franx M., Rettura A., Illingworth G. D., Ford H. C., 2008, ApJ, 688, 48 van Dokkum P. G., 2005, AJ, 130, 2647 van Dokkum P. G., 2008, ApJ, 674, 29 Wake D. A., et al., 2006, MNRAS, 372, 537 Warren S. J., et al., 2007, astro, arXiv:astro-ph/0703037 White S. D. M., Rees M. J., 1978, MNRAS, 183, 341 Willmer C. N. A., et al., 2006, ApJ, 647, 853 York D. G., et al., 2000, AJ, 120, 1579 \[lastpage\] [^1]: Research Fellow of the Japan Society for the Promotion of Science [^2]: E-mail: [email protected]

ArXiv

--- abstract: '47 Tuc is an ideal target to study chemical evolution and GC formation in massive more metal-rich GCs since is the closest, massive GC. We present chemical abundances for O, Na, Al, Si, Ca, Ti, Fe, Ni, La, and Eu in 164 red giant branch (RGB) stars in the massive globular cluster 47 Tuc using spectra obtained with both the Hydra multi-fiber spectrograph at the Blanco 4-m telescope and the FLAMES multi-object spectrograph at the Very Large Telescope. We find an average \[Fe/H\]=–0.79$\pm$0.09 dex, consistent with literature values, as well as over-abundances of alpha-elements ($[\alpha/\mbox{Fe}]\sim0.3$ dex). The n-capture process elements indicate that 47 Tuc is r-process dominated (\[Eu/La\]=+0.24), and the light elements O, Na, and Al exhibit star-to-star variations. The Na-O anti-correlation, a signature typically seen in Galactic globular clusters, is present in 47 Tuc, and extends to include a small number of stars with \[O/Fe\] $\sim$–0.5. Additionally, the \[O/Na\] ratios of our sample reveal that the cluster stars can be separated into three distinct populations. A KS-test demonstrates that the O-poor/Na-rich stars are more centrally concentrated than the O-rich/Na-poor stars. The observed number and radial distribution of 47 Tuc’s stellar populations, as distinguished by their light element composition, agrees closely with the results obtained from photometric data. We do not find evidence supporting a strong Na-Al correlation in 47 Tuc, which is consistent with current models of AGB nucleosynthesis yields.' author: - 'M. J. Cordero and C. A. Pilachowski' - 'C. I. Johnson[^1]' - 'I. McDonald and A. A. Zijlstra' - 'J. Simmerer[^2]' - Accepted for publication in The Astrophysical Journal title: 'Detailed Abundances for a Large Sample of Giant Stars in the Globular Cluster 47 Tucanae (NGC 104)' --- Introduction ============ Recent observational evidence has shown that many Galactic globular clusters, including 47 Tucanae (NGC 104), contain more than one chemically distinct stellar population (e.g., see recent review by Gratton et al. 2012). In the case of 47 Tucanae (\[Fe/H\] $\sim$–0.77 from Carretta et al. 2009a; age 9.9 $\pm$ 0.7 Gyr from Hanson et al. 2013, although VandenBerg et al. reported an age of 11.75 $\pm$ 0.25 Gyr), Anderson et al. (2009) used archival Hubble Space Telescope images to resolve at least two populations on the cluster’s subgiant branch. Additional analysis of the cluster’s narrow-red-horizontal and subgiant branch regions by Di Criscienzo et al. (2010) identified three subpopulations: a first generation population comprising 30% of cluster stars and two second generation populations. The dominant population of second generation stars are thought to have enhanced He, but the minority population ($\sim$10%) may instead exhibit enhancements in the sum of C+N+O. Nataf et al. (2011) also find changes in the RGB-bump and horizontal branch morphologies as a function of cluster radius to be evidence supporting He-enhancement in some 47 Tuc stars near the core. More recently, Milone et al. (2012) extended these studies to follow the two major populations in 47 Tuc from the main sequence through the horizontal branch, and identified a third population comprising only 8% of the cluster’s stars and discerned only on the subgiant branch. The primordial population contributes only 20% of the stars in the cluster center, but is more extended, and contributes 30% of the cluster’s stars overall. In addition, Richer et al. (2013) reported that stars on the blue side of the main sequence are more centrally concentrated and show anisotropic proper motions compared to stars on the red side of the main sequence. Overall, the evidence for multiple populations in 47 Tuc is compelling and suggests that both chemical and kinematic differences separate the populations. Studies of the light elements (CNO, Na) in 47 Tuc mirror the photometric analyses of the cluster’s multiple populations. Worley & Cottrell (2012), following earlier analyses by Norris & Freeman (1979) and Paltoglou & Freeman (1984), confirmed the anti-correlation of sodium and CN, as well as the bi-modal distribution of CN strength, from medium resolution spectra of giants. Carretta et al. (2009b) determined oxygen and sodium abundances for 115 giants in 47 Tuc, finding that 2/3 of the stars belong to an intermediate (O-poor and Na-rich) population, consistent with current cluster formation models, with most of the remaining stars belonging to an earlier (primordial) generation. D’Orazi, et al. (2010) examined several dozen turnoff stars in 47 Tuc, confirming the anti-correlation of oxygen and sodium in unevolved stars and finding a similar distribution in \[Na/O\] as in the giants. Their observations rule out in-situ mixing in the giants as the primary source of the O-Na anticorrelation. Neutron-capture elements, however, show only modest enhancements and no star-to-star variations (James et al. 2004, Worley et al. 2010), suggesting that core-collapse supernovae and low-mass AGB stars also do not cause the light element abundance anomalies. According to Vesperini et al. (2013) N-body simulations indicate that the multiple populations in a GC do not mix completely until the cluster has lost a large fraction of its mass through two-body interactions, which takes several relaxation times. For instance, the authors found that the populations in a GC are well mixed when the cluster’s age is $\sim15$ times its half-mass relaxation time-scale $t_{rh}$; for 47 Tuc the ratio of the cluster age (9.9 Gyr, see Hansen et al. 2013) to its half-mass relaxation time (3.5 Gyr, Harris 1996, 2010 edition) is $\sim3$, thus the multiple populations in 47 Tuc are expected to differ not only on their chemical composition, but also on their radial distribution and proper motions. In this work we present the analysis of the spectra of 164 red giants in the globular cluster 47 Tuc, including abundances for oxygen, sodium, and aluminum, among other elements, to explore star-to-star variations in the light element abundances and characterize the multiple populations. The paper is organized as follows. Sections 2 and 3 present the observational data and a description of the analysis, respectively. Our results and a comparison to the literature are presented in Section 4. In Section 5, we discuss our results in the context of the multiple stellar populations in 47 Tuc, focusing on the chemical signatures of the multiple populations and their radial extent in the cluster, comparing our findings to theoretical models. Our conclusions are given in section 6. Observations and Data Reduction =============================== Observations ------------ Spectroscopic observations were obtained using two multi-fiber positioners and spectrographs: Hydra on the Blanco 4-m telescope at the Cerro Tololo Interamerican Observatory during two observing runs in July, 2003, and August, 2010, and FLAMES–GIRAFFE on the VLT–UT2 telescope at the European Southern Observatory on Cerro Paranal during a single observing run in November, 2011. For both Hydra runs the “large” (300 $\mu$m) fiber cable, the Blue Air Schmidt Camera, the 316 l/mm (57.5) echelle grating, and the SITE 4096 $\times$ 2048 pixel CCD were employed, but at two different wavelength ranges. For the 2003 observations, filter \#6 centered at 6757 Å yielded spectra ranging from $\lambda\lambda$6490-6800 Å, at a dispersion of 0.15 Å pixel$^{-1}$. The 2010 observations employed the E6257 filter, covering the spectral range from $\lambda\lambda$6100-6350 Å in one observation, with a resolving power of typically 20,000.[^3] Different fiber configurations were used during each run, although several stars were included in both configurations. Spectra for the VLT–FLAMES portion of the data set utilized the FLAMES–GIRAFFE spectrograph (Pasquini et al. 2003). The high–resolution HR13, HR14A, and HR15 spectrograph setups provided continuous wavelength coverage from approximately 6115–7000 Å, with a resolving power of R$\approx$20,000 (see also Table 1). Selection of Stars ------------------ Stars observed with Hydra were selected for observation from the proper motion survey of Tucholke (1992), with membership probabilities above 88%. The proper motion study included stars out to a radius of 32 arc minutes from the center of the cluster, but does not include stars within about 2 arc minutes of cluster center. The stars included in this study range in magnitude from 13.3 $\leq$ B $\leq$ 14.6, from the tip of the giant branch to just above the horizontal branch, and include both red giants and asymptotic branch giants. Stars were selected randomly and not biased as to light element abundance. Spectra of 52 stars were available from the 2003 observations, and 48 stars were obtained from the 2011 observations, with 17 stars in common between the two configurations. B magnitudes were taken from Tucholke (1992), and identifications were matched to either stars in the Lee (1977) or Chun & Freeman (1978) photometric studies to obtain V magnitudes. J and K magnitudes were obtained from the 2MASS Point Source Catalog. The stars included in this study are listed in Table 2. The identification numbers without prefix were taken from Lee (1977), while identification numbers with prefix D or E were taken from Chun & Freeman (1978), and with prefix M were taken from McDonald et al. (2011). Stars observed with FLAMES were selected from photometry compiled by McDonald et al. (2011). The targets were originally selected to contain RGB and early AGB stars of similar temperature and luminosity. However, a problem with the fiber assignment process led to RGB stars only being observed. The final sample includes 113 RGB stars from the luminosity level of the horizontal branch to near the RGB-tip. A color-magnitude diagram of stars observed in 47 Tuc is shown in Fig. \[fig\_CMD\], including both our Hydra and FLAMES samples and the sample from Carretta et al. (2009b). Since each of these samples adopted V magnitudes from different sources, we shifted the V magnitudes of the Carretta et al. sample upward by +0.01 magnitudes based on 20 stars in common between the Carretta and FLAMES samples, and shifted the V magnitudes of the Hydra sample downward by -0.05 magnitudes based on 14 stars in common with the FLAMES sample. Anticipating the assignment of stars to the primordial, intermediate, and extreme populations defined by Carretta et al. (2009b), we have color-coded the stars by population assignment as noted in the figure caption. Unclassified stars lacking either Na or O abundances are also identified. The combined sample includes primarily giants on the first ascent of the red giant branch, as well as a few asymptotic giant branch stars. Data Reduction -------------- Calibration data for both Hydra runs included bias and flat field frames on all nights. Daylight sky spectra (2003 observations) or comparison lamps (2010 observations) were obtained for wavelength calibration. Raw images were bias corrected and trimmed, and an averaged zero frame was subtracted to remove any two dimensional structure in the bias. Following these preliminary reductions, extraction to one dimensional spectra, sky subtraction, and wavelength calibration were performed using the IRAF[^4] script [*dohydra*]{}. Finally, the continuum for each spectrum was normalized using a low order cubic spline. A signal-to-noise ratio of typically 120-200 per pixel, depending on wavelength, characterizes the final spectra. Basic data reduction for the FLAMES–GIRAFFE sample observations was carried out using the Geneva Observatory girBLDRS[^5] software. The reduction tasks included bias subtraction, flat–field correction, wavelength calibration using the ThAr comparison lamps, cosmic ray removal, and object spectrum extraction. The IRAF task *skysub* was used to create and subtract an average sky spectrum for each exposure, based on 15 sky fibers distributed across the field–of–view. Additionally, we removed telluric features with the IRAF task *telluric* and a previously obtained grid of rapidly rotating B stars observed at various airmasses. The individual exposures were then co–added using the IRAF task *scombine* to obtain final spectra with signal–to–noise (S/N) ratios ranging from about 50–200. Analysis and Uncertainties ========================== A detailed analysis of the composition of 47 Tuc giants was performed using the 2010 version of MOOG (Sneden 1973). LTE model atmospheres were interpolated with the $\alpha$-enhanced ATLAS9 grid of non-overshoot models (Castelli & Kurucz 2003). Models were interpolated with an adopted metallicity of \[M/H\]=–0.75[^6]. Initial model atmosphere parameters were estimated using the Alonso et al. (1999) analytic functions relating color to temperature, which were based on the infrared flux method. B-V colors were used for stars for which infrared magnitudes are not available. Different authors present a range of values for interstellar reddening to 47 Tuc ranging from 0.029 $\leq$ E(B-V) $\leq$ 0.064 (e.g. Crawford & Snowden 1975, Gratton et al. 1997). A reddening value of E(B-V)=0.04 (Harris 1996, 2010 edition) was adopted, in the mid-range of published values and consistent with the values adopted by most other studies of 47 Tuc’s composition. For stars in the FLAMES sample with large spectral coverage, effective temperatures were constrained via excitation equilibrium. Surface gravities were obtained using the absolute bolometric magnitudes, derived effective temperatures, and assuming a mass of 0.8 $M_{\odot}$ for all stars in the Hydra sample. Bolometric corrections were adopted from Alonso et al. (1999), and we used a distance modulus of (m-M)$_{V}$ = 13.37 (Harris 1996, 2010 edition). The limited spectral coverage available with the Hydra multi-object spectographs provides too few lines of Fe II to permit a reliable spectroscopic estimate of surface gravity. For stars in the FLAMES sample with wider spectral coverage, surface gravities could be obtained from the ionization equilibrium of Fe I and Fe II. Microturbulence was estimated for each star using the v$_{t}$ relation given by Johnson et al. (2008), and then adjusted to minimize the dependence of derived \[Fe/H\] on line strength. The final adopted atmospheric parameters are included in Table 3. A comparison of the atmosphere parameters obtained for a subsample of 17 stars using both Hydra and FLAMES data is shown in Fig. \[fig\_logTeff\]. The average differences in the sense FLAMES minus Hydra are $\Delta(T_{\mbox{eff}})=-18$ K ($\sigma=75$ K), $\Delta(\log g)=0.02$ cgs ($\sigma=0.16$ cgs), $\Delta([Fe/H])=-0.08$ dex ($\sigma=0.10$ dex), and $\Delta(v_t)=-0.02$ km s$^{-1}$ ($\sigma=0.24$ km s$^{-1}$). The average difference for all \[x/Fe\] ratios are smaller than 0.09 dex and are shown in column 9 of Table 3. The \[x/Fe\] abundance ratios of the Hydra sample shown in Table 2 were corrected by these offsets. Thus, no strong systematic offsets are introduced by our adopted atmosphere parameters in our total sample. For iron, nickel, silicon, titanium, and calcium, abundances were determined from equivalent widths, adopting the line list from Johnson & Pilachowski (2010). As described by Johnson & Pilachowski (2010), gf values were determined by matching line strengths in the solar spectrum to the photospheric abundances of Anders and Grevesse (1989). While the Anders and Grevesse photospheric abundances have been supplanted by later studies, we have retained their use to assure that our abundances are on the same scale as those of Johnson et al. (2009) and subsequent papers. Equivalent widths for the Hydra samples were measured using the *splot* task in IRAF. Only lines with log W/$\lambda$ $<$–4.5 were included in our analysis. Similarly, for the FLAMES data the equivalent widths were measured using an updated version of the code developed for Johnson et al. (2008). Isolated lines were fit with a single Gaussian profile. Partially contaminated lines were deblended by fitting multiple Gaussian profiles and guided by an examination of the high resolution Arcturus atlas by Hinkle et al. (2000). Abundances of the elements oxygen, sodium, aluminum, lanthanum, and europium were determined using spectrum synthesis to account for blending with other spectral features. Synthetic spectra were calculated for each spectral region using each stars’ adopted model atmosphere parameters and a range of abundances for the target species. Synthetic spectra were broadened by a Gaussian smoothing kernel to match the observed line profiles of nearby, unblended lines and compared by eye to the observed stellar spectra to determine the best-fit abundance. Uncertainties of abundances determined using spectrum synthesis are set from the smallest clear distinction that can be made consistent with the S/N ratio of the observed spectrum. The oxygen abundance was determined using the \[O I\] $\lambda$6300 Å feature. The line list was the same as that used in Johnson & Pilachowski (2010) and the synthesized spectral region covered 6296-6304 Å. Since the O\[I\] line is moderately sensitive to the C+N abundance, for each star, the synthetic spectrum was obtained by modifying the C and N abundances until a good match was found between the synthetic and observed spectrum. As a starting point we used the C and N abundances presented by Briley (1997) who found a bimodal CN-distribution in 47 Tuc using a sample of 109 RG stars. We present oxygen abundances for 103 stars that were not affected by sky contamination. An example of spectrum synthesis around the \[O I\] line is shown in Fig. \[fig\_synth\]. Altering the CN abundance by $\pm$ 0.50 dex results in a minor change on the oxygen abundance ($\mp$ 0.01 dex) as shown in Fig. \[fig\_CNSc\]a. Another source of uncertainty in the oxygen abundance is introduced by blending with Sc, therefore, the Sc abundance was adjusted for the synthesis to provide a good fit to the line profile as shown in Fig. \[fig\_CNSc\]b. The sodium abundance was determined using the $\lambda\lambda$ 6154, 6160 Å Na doublet for 138 members. Spectrum synthesis is required because the Na $\lambda$6160 Å is blended with Ca and molecular features. The line list was the same as that used in Johnson & Pilachowski (2012) covering the spectral region 6150-6170 Å. In cool, metal-poor giants the formation of Na I lines is known to exhibit departures from local thermodynamic equilibrium (Gratton et al. 1999, Mashonkina et al. 2000). According to Lind et al. (2011) these NLTE corrections are negligible for giants at 47 Tuc’s metallicity. The aluminum abundance was determined using the $\lambda\lambda$ 6696, 6698 Å Al doublet. While the lines of this Al I doublet are not significantly blended with other atomic lines, they are affected by the presence of CN, particularly since 47 Tuc is relatively metal-rich for a GC. The line list was the same as that used in Johnson & Pilachowski (2012). Departures from LTE in metal-poor giants for the $\lambda\lambda$ 6696, 6698 Å Al doublet are less well characterized in the literature, and aluminum abundances in Table 2 do not include NLTE corrections. According to Baumuller & Gehen (1997), aluminum non-LTE corrections are expected to be negligible for the $\lambda\lambda$ 6696, 6698 Å Al doublet in metal-rich RG stars. La and Eu were measured using spectrum synthesis to take into account blending with nearby CN lines, hyperfine splitting, and isotopic splitting (Eu only). The line lists used to synthesize the 6262 Å La II and 6645 Å Eu II lines were the same as that used in Johnson & Pilachowski (2010), which include the hyperfine structure linelists available in Lawler et al. (2001a and 2001b). Uncertainties ------------- Uncertainties in the atmospheric parameters are used as independent error sources to assess the uncertainties in the abundance ratios. The errors in our photometric temperatures introduced by the empirical fits from Alonso et al. (1999) are 25-50 K. A change in $T_{\mbox{eff}}$ of $\pm$ 50 K typically results in an abundance difference of 0.05 dex. To explore how sensitive our estimated abundances are to the adopted surface gravity, we changed log $g$ by $\pm$0.20 cgs, which is a conservative error estimation. In fact, uncertainties in the bolometric corrections and distance modulus result in an error in our photometric surface gravities of 0.10 cgs. Given the observational scatter, changes in the slope of the derived abundance vs. line strength were not statistically significant for changes in the microturbulence velocity less than 0.25 kms$^{-1}$, which we have adopted as the uncertainty in $v_t$. Sensitivities to model atmosphere parameters are presented in Table \[table\_sensitivities\], for a representative star ($T_{\mbox{eff}}=4200$ K, log $g=1.3$ cgs, \[Fe/H\]$=$-0.75 dex, and $v_t=$1.9 kms$^{-1}$), by changing the model atmosphere parameters ($\Delta T_{\mbox{eff}}\pm50$ K, $\Delta \log g\pm0.2$ cgs, $\Delta [M/H]\pm$0.10 dex, and $\Delta v_t\pm$0.25 kms$^{-1}$) one at a time while keeping the other unchanged, neglecting second order dependencies among the atmospheric parameters. The random errors for Fe I abundances are assumed to be the standard error of the mean, since typically 20 (Hydra sample) or 50 (FLAMES sample) Fe I lines contributed to our measurement. For species with fewer lines, $\sigma_{\mbox{Fe I}}/\sqrt{N_{\mbox{lines}}}$ provides a realistic estimate of observational uncertainties. For species whose abundances were estimated using spectrum synthesis, we added to the error budget the uncertainties related to smoothing, fitting, and continuum placement. The total errors presented in Table \[table\_sensitivities\] were calculated adding in quadrature the sensitivities to model atmosphere parameters and either random errors or errors introduced by the spectral synthesis technique. Overall, uncertainties in the determined abundances are smaller than 0.20 dex. Results ======= Iron and Alpha Group Elements ----------------------------- In Fig. \[fig\_boxplot\], we show a box plot with the abundance ratios for elements we have determined in our study. Abundances ratios as a function of $T_{\mbox{eff}}$ are shown in Fig. \[fig\_Abund\]. We found \[Fe/H\]=$-0.79\pm0.09$ dex and \[Ni/Fe\]=$-0.01 \pm 0.07$ dex, for a sample $\sim$ 164 and 140 stars respectively. Our average \[Fe/H\] is consistent with other studies (Carretta et al. 2009, Koch & McWilliam 2008, and references therein). The uncertainties due to errors in the atmospheric parameters are comparable to the observed dispersion in both elements, suggesting that 47 Tuc is homogeneous in both Fe and Ni. The very small \[Ni/Fe\] ratio is typical of Galactic GCs and field stars (Gratton et al., 2004). Abundance ratios for the $\alpha$ elements are presented in Table 2 and an average \[$\alpha$/Fe\]=$+0.29 \pm 0.06$ dex was found. This result is in agreement within the errors with previous studies (\[$\alpha$/Fe\]$\sim+$0.4 dex e.g., Peterson et al. 1993, Carretta et al. 2004, Fulbright et al. 2007, Koch & McWilliam 2008). Since the spread is comparable to the measurement and model atmosphere uncertainties, we find that $\alpha$-elements abundance ratios are homogeneous in 47 Tuc and show the same trend as field metal-poor stars, i.e. they are enhanced. Light Elements O, Na, and Al ---------------------------- 47 Tuc, like most Galactic GCs, presents variations in oxygen, sodium, and aluminum abundances. We determined O, Na, and Al abundances for a sample of 103, 138 and 129 stars, respectively and found a significant spreads in the \[X/Fe\] ratios for these light elements ($\Delta (\mbox{[O/Fe]})\sim1.0$ dex, $\Delta (\mbox{[Na/Fe]})\sim0.90$ dex, and $\Delta (\mbox{[Al/Fe]})\sim0.70$ dex), which in the case of O and Al are larger than the spreads found by Carretta et al. 2009b ($\Delta (\mbox{[O/Fe]})\sim0.6$ dex and $\Delta (\mbox{[Na/Fe]})\sim0.90$ dex) and Carretta et al. 2013 ($\Delta (\mbox{[Al/Fe]})\sim0.50$ dex)). Moreover, Na and O exhibit the largest interquartile range of values of all the elements we measured in 47 Tuc. 47 Tuc exhibits the Na-O anticorrelation characteristic of many GCs (Carretta et al. 2009), shown in Fig. \[fig\_NaO\]. For this particular cluster three different populations are identified. These light element variations are not a consequence of NLTE effects, since stars with identical atmospheric parameters show different line strengths, as shown in Fig. \[fig\_strengths\]. Interestingly, the Al abundance ratios can be identical for stars that exhibit different Na abundances, suggesting that at 47 Tuc’s metallicity \[Al/Fe\] does not strongly discriminate different populations, as shown in Figures \[fig\_strengths\] and \[fig\_AlNa\]. On average aluminum is enhanced (\[Al/Fe\]=+0.38 $\pm$0.12) in our sample of 129 stars. A spectral region around the $\lambda\lambda$6696 & 6698 Å Al doublet and the spectra of two stars with almost identical stellar model atmospheres is shown in Fig. \[fig\_strengths\]. Recently, Carretta et al. (2013) presented measurements for Al, Mg, and Si. A common sample of 20 stars indicates that our \[Al/Fe\] ratios are systematically lower than their values by 0.13 dex. The average differences in the sense our measurements minus Carretta’s are $\Delta(T_{\mbox{eff}})=86$ K ($\sigma=72$ K), $\Delta(\log g)=0.11$ cgs ($\sigma=0.30$ cgs), $\Delta([Fe/H])=-0.02$ dex ($\sigma=0.10$ dex), and $\Delta(v_t)=0.26$ km s$^{-1}$ ($\sigma=0.21$ km s$^{-1}$). We found that differences in effective temperature and microturbulence velocities can be responsible for the \[Al/Fe\] offset in some of the stars in common. Neutron Capture Elements ------------------------ The star-to-star dispersion in the neutron-capture elements La and Eu is smaller than that of the lighter elements O, Na, and Al in 47 Tuc, with \[La/Fe\]=+0.20$\pm$0.12 dex and \[Eu/Fe\]=+0.44$\pm$0.11 dex, for a sample of 123 and 124 members respectively. These abundance ratios are consistent with the range of published values in globular clusters (e.g. Sakari et al. 2013, Worley et al. 2009, and references therein), as well as with the abundance ratios of halo stars of similar metallicity (Gratton et al. 2004). In the Sun, La and Eu are made primarily by the s-process and r-process, respectively, with log A(La/Eu)=2.23 for the s-process component and log A(La/Eu)=0.09 for the r-process component (Sneden et al. 2008). For 47 Tuc giants, we find an average value of log A(La/Eu)=0.45, suggesting that 47 Tuc’s heavy element abundances are dominated by r-process production, with limited contribution from the s-process. Furthermore, low- and intermediate-mass AGB stars are considered to be the primary production sites for La (e.g., Busso et al. 1999; Herwig 2005; Straniero et al. 2006) and trace a longer enrichment timescale than the core-collapse SNe responsible for the enrichment of alpha-elements. In contrast, Eu production is tied to short enrichment timescales and helps explain the variable \[Eu/Fe\] ratios detected in low-metallicity Galactic stars (e.g., see review by Sneden et al. 2008 and references therein). For instance, Eu production is linked to both neutron star mergers (Freiburghaus et al. 1999; Goriely et al. 2011; Korobkin et al. 2012) and Type II SNe (Lattimer et al. 1997; Truran 1981; Mathews & Cowan 1990; Takahashi et al. 1994; Woosley et al. 1994; Freiburghaus et al. 1999; Truran et al. 2002; Arnould et al. 2007). The star Lee 4710 appears to be La-rich at \[La/Fe\]=+1.02 and \[La/Eu\]=+0.52. Lee 4710 may have been part of a binary mass transfer with an earlier, more massive AGB companion (Busso et al. 1999). The small number of La-rich stars found is consistent with 47 Tuc’s binary fraction ($\leq$ 2 %) measured by Milone et al. (2012b) using HST data of the Globular Cluster Treasury project. Since we only have a single epoch of observation for this star, radial velocity variations can not be estimated. Additionally, even if this La-rich star might have been in a binary system, two-body interactions with cluster stars might have made the system unbound. Trends with Effective Temperature --------------------------------- Si and La exhibit a trend with T$_{\mbox{eff}}$ which is shown in Fig. \[fig\_Abund\]. A linear regression of \[Si I/Fe\] and \[La II/Fe\] with T$_{\mbox{eff}}$ resulted in a correlation value of 0.44 and 0.41, respectively, which indicates a medium correlation of these abundance ratios with T$_{\mbox{eff}}$. Furthermore, the two-tailed probability (p-value) obtained to gauge the statistical significance of the linear fit is $3.00\times10^{-8}$ and $1.15\times10^{-6}$, for \[Si I/Fe\] and \[La II/Fe\] respectively. The null hypothesis that there is no linear trend of these abundance ratios with T$_{\mbox{eff}}$ is rejected since the p-values obtained are smaller than 0.05. The larger observational scatter in the \[La II/Fe\] ratio detected at higher temperatures is a result of the lower S/N of these fainter stars. Additionally, another consequence related to the lower S/N is that stars with intrinsically weak La lines are less likely to be included in the analysis making the slope of the \[La II/Fe\]-T$_{\mbox{eff}}$ relation appear steeper. Interestingly, there is a hint of a \[La II/Fe\] correlation with T$_{\mbox{eff}}$ for M3 and M13 in the results presented by Sneden et al (2004) in their Fig. 3. The origin of the Si-T$_{\mbox{eff}}$ trend is unknown and the high-excitation potential lines used for the analysis might not be reliable tracers of the actual Si abundance. The resolution of this long-standing problem in the abundance of Si is beyond of the scope of this project. Discussion ========== Na-O Anticorrelation -------------------- Few chemical studies of 47 Tuc have been performed using samples of $\sim$100 or more stars. For instance, Carretta et al. (2009a) showed that the Na-O anticorrelation is a shared characteristic of all GCs, including 47 Tuc, and identified three subpopulations of primordial (P), intermediate (I), and extreme (E) stars. The spreads in O and Na are not caused by in situ nucleosynthesis, since a Na-O anti-correlation is also found in low-mass turn-off and early subgiants stars in 47 Tuc (Carretta et al. 2004), which do not have convective envelopes reaching temperatures high enough to deplete O and enhance Na. To classify 47 Tuc giants into the primordial, intermediate, and extreme populations identified by Carretta et al. (2009b), we adopted the criteria used by those authors. The primordial population is defined as those stars with Na abundances comparable to field stars at similar metallicity; at the metallicity of 47 Tuc, this corresponds \[Na/Fe\]$_{\mbox{NLTE}}\leq+$0.4 dex. Stars with \[Na/Fe\]$_{\mbox{NLTE}}>$0.4 dex constitute the second generation. Additionally, Carretta et al. defined an extended second generation found in some massive GCs; this “extreme” second generation was defined as stars with high Na and \[O/Na\]$_{\mbox{NLTE}}\leq$–0.90 dex. After removing the NLTE correction used by Carretta et al.[^7] we adopted \[Na/Fe\]=+0.3 dex to separate the primordial (first) and intermediate (second) generations, and \[O/Na\]=–0.85 dex to split the second generation into an intermediate and an extreme population. Our samples included several giants with lower \[O/Fe\] ratios than were found in Carretta et al.’s (2009b) sample. Thus, our Na-O anti-correlation data, shown in Fig. \[fig\_NaO\], extends in \[O/Fe\] down to –0.5 dex, indicating the presence of all three populations defined by Carretta et al. (2009b). A histogram of the \[Na/O\] abundance ratio (Fig. \[fig\_histo\]) shows two peaks with an extended tail toward high Na abundance supporting the presence of three populations. 47 Tuc is one of the few GCs in addition to Omega Cen, M13, NGC 2808 and NGC 6752, that presents more than two populations. According to the models of Ventura and D’Antona (2008), an oxygen depletion of only $\sim0.4$ dex is expected in metal-rich AGB stars (Z=0.004) as a consequence of hot bottom burning, which makes challenging to explain the existence of third generation stars in 47 Tuc with \[O/Fe\]$\sim$–0.5. Interestingly, Ventura and D’Antona (2007) found that He-enhanced, metal-poor (\[Fe/H\]=-1.5) stars formed with \[O/Fe\]=–0.2 can become extreme O-poor stars (\[O/Fe\]$\leq$–0.4 dex) if deep mixing takes place. For instance, several authors (e.g., Kraft 1992, 1997; Pilachowski et al. 1996; Cavallo & Nagar 2000; Sneden et al. 2004; Cohen & Meléndez 2005; Johnson et al. 2005; Johnson & Pilachowski 2012) found extreme O-poor stars in the metal-poor GC M13 and suggested that this can be explained by deep mixing. These authors found that the extreme O-depleted stars are found only above the RGB bump and their \[O/Fe\] ratio decreases with increasing luminosity, consistent with the deep mixing scenario. In our sample, 11 stars were found to belong to the extreme population. Even though all the stars in our sample are above the RGB bump, we do not find a strong trend of \[O/Fe\] versus $T_{\mbox{eff}}$. This could be a result of undersampling the cooler/more luminous stars. While extreme population, O-depleted stars are absent in the Carretta et al. (2004) sample of TO and SGB stars on 47 Tuc, nevertheless, their absence might be a consequence of a small sample statistics. Unfortunately, simulation of deep mixing at the metallicity of 47 Tuc are unavailable to help interpret our measurements. Nonetheless, oxygen depletion by deep mixing in 47 Tuc is expected to be mild for high metallicity clusters (Charbonnel et al. 1998). Fluorine abundances have been determined in a few GCs, such as Omega Cen, M4, NGC 6712, M22, and 47 Tuc (Cunha et al. 2003; Smith et al. 2005; Yong et al. 2008; Alves-Brito et al. 2012; Laverny & Recio-Blanco 2013), and although F measurements are available for only small samples of stars, they exhibit star-to-star variations within the GCs. While the data available for some GCs suggest a Na-F anticorrelation and O-F correlation, indicating that O and F are destroyed at the high temperatures needed to produce Na, de Laverny & Recio-Blanco found a possible Na-F correlation in 47 Tuc. While observations of fluorine variations in globular clusters are intriguing, more observations are needed to quantify the variations and to clarify the role of proton-capture nucleosynthesis in globular cluster chemical evolution. Population Fractions and Radial Distributions --------------------------------------------- The cumulative radial distributions of the stellar populations in our Hydra and FLAMES samples, shown in Fig. \[fig\_cum\]a, suggest that the intermediate population (green curve) is more centrally concentrated than the primordial population (blue curve). This result is consistent with theoretical models for GC formation (D’Ercole et al. 2008) that predict that a cooling flow of low-velocity gas from AGB stars accumulates gas in the innermost regions of the cluster where a second generation of stars forms. Furthermore, VandenBerg et al. (2013) indicated that 47 Tuc is one of the massive and concentrated Galactic GCs that does not develop steady-state winds and therefore can retain the gas expelled by its stars. A Kolmogorov-Smirnov (KS) statistical test was used to assess whether the cumulative radial distributions of the different generations are drawn from the same parent distribution [^8]. A P-value=0.02 allows to reject our null hypothesis. This result is in contradiction to that of Carretta et al. (2009b), who did not find a statistically significant difference in the cumulative distributions of primordial and intermediate generations. The combination of the Hydra and FLAMES data together provides larger samples in both the outer regions of the cluster (benefiting from the wide field of view of Hydra) an the inner regions of the cluster (benefiting from the closer spatial sampling of FLAMES). Our result is in agreement, however, with the photometric approach for tracing multiple populations (Milone et al. 2012). After we removed the NLTE correction from the Carretta et al. Na abundances, we removed an additional systematic offset of -0.05 dex determined from a common subsample of 19 stars. To add the Carretta et al. stars with oxygen abundances, we applied a systematic offset of +0.06 dex determined from a common subsample of 11 stars. Finally, we combined the samples and determined population membership for all stars from the Na-O plane. From the larger, combined sample a more significant difference between the radial distributions of the primordial and secondary populations was found, as shown in Fig. \[fig\_cum\]b, with a P-value=$2.3\times10^{-4}$. Even though the combined sample of 191 stars with both Na and O measurements is smaller than the large sample used by Milone et al. (2012), the spectroscopic data clearly demonstrate that the second generation is more centrally concentrated than the first generation. The fraction of first-to-second generation stars N(I+E)/NP can be used to estimate the mass of the progenitor structure (Valcarce & Catelan, 2011) or to constrain the long-term dynamical evolution of the cluster (Vesperini et al., 2013). Vesperini et al. found that at any given time in the history of a cluster, the global fraction N(I+E)/NP matches that measured locally at a distance from the center of the cluster between one and two half-mass radii ($r_h$). For our combined sample, the global value for N(I+E)/NP (3.4) matches the local value at $\sim1.5r_h$, in agreement with predictions from Vesperini et al.’s N-body simulations. Additionally, the global fraction of second generation stars is expected to increase over time, since more spatially extended first generation stars are more likely to be lost from the cluster (D’Ercole et al. 2008 and Vesperini et al. 2013). We compare the fraction of second generation stars as a function of distance from cluster center in units of the half mass radius in Fig. \[fig\_popratio\], which also includes measurements of the fraction of second generation stars from the Milone et al. sample. The global value of N(I+E)/(N(I+E+P)=0.68 is equal to the local value at a radius of $\sim1.5r_h$. Within the uncertainties of both studies, the fraction of second generation stars traces a similar profile. Vesperini et al. (2013) also determined that multiple populations in GCs become well mixed after more than 60% of the initial cluster mass is lost through two-body interactions (see their Fig. 4). Furthermore, Giersz & Heggie (2011) estimated using Monte Carlo simulations that 47 Tuc has lost only $\sim45$% of its initial mass. Thus, we would expect second generation stars still to be centrally concentrated, consistent with the observed central concentration of O-poor/Na-rich stars. The Aluminum Abundance in 47 Tuc -------------------------------- Aluminum is the most massive light element that shows significant star-to-star variations in typical GCs and its production in significant quantities requires temperatures $\geq 70$ MK (e.g. Langer et al. 1997; Prantzos et al 2007). Since the convective envelopes of metal-rich low-mass giants do not reach these high temperatures (Lager et al. 1997, Denissenkov & Weiss 1996, and references therein), aluminum can be used to constrain the nature of the polluters responsible for the light element abundance variations seen in GCs. Several studies have included analyses of the abundance of aluminum in 47 Tuc, with most authors noting a modest enhancement of typically \[Al/Fe\] $\sim +0.3$ dex but little scatter (Brown & Wallerstein 1992, Norris & Da Costa 1995, Carretta et al. 2004, Alves-Brito et al. 2005, Koch & McWilliam 2008, and Carretta et al. 2009a). Each of these studies included only a dozen or fewer stars, and a larger sample of aluminum abundance determinations is needed to understand the variation observed and any intrinsic spread. More recently, Carretta et al. (2013) found an average \[Al/Fe\] = +0.53 $\pm$ 0.01 for a sample of 116 stars, and a range of $\sim$ 0.5 dex. Moreover, O’Connell et al. (2011) noted that star-to-star variations in the \[Al/Fe\] ratio within a cluster depends on metallicity. For clusters more metal poor than \[Fe/H\] $\leq$ –1.2, the spread can be as large as $\sim 1.5$ dex (see, for example, \[Al/Fe\] for M13 in Johnson et al. 2005, Sneden et al. 2004, and Cohen & Meléndez 2005), while for more metal-rich clusters, the dispersion from star to star in the \[Al/Fe\] ratio is smaller than $\sim 0.4$ dex (M107 in O’Connell et al. 2011). Our measured \[Al/Fe\] ratios are presented in a box plot in Fig. \[fig\_boxplot\]. The median abundance ratio is represented by a central horizontal line, the first and third quartiles correspond to the bottom and top box boundaries, and the full range of abundance ratios is specified by a vertical line. Comparing our results for the light elements O, Na, and Al, we first notice that the interquartile range of values is $\sim$ two times smaller for Al than for Na. This result is consistent with the simulations of Ventura & D’Antona (2008), hereafter VD08, which indicate that at 47 Tuc’s metallicity, first generation intermediate mass ($\sim$4-8 M$_{\odot}$) AGB stars are expected to synthesize Al in smaller quantities, while producing Na in greater quantities compared to more metal-poor AGB stars. Among other physical parameters, VD08 use the mass loss treatment given by Blocker (1995), which finds greater mass loss than other published descriptions, resulting in a shorter AGB lifetime, i.e. diminishing the number of third dredge up episodes. Additionally, VD08 found that the temperature at the bottom of the convective envelope decreases with metallicity, such that only more massive AGB stars will reach the high temperatures required to convert Mg into Al. The more massive AGB stars also have shorter lifetimes limiting the Mg destruction and Al production. Thus, both the smaller number of third dredge up episodes due to shorter lifetimes and lower temperatures in the convective envelope of metal-rich AGB stars explain the lower aluminum enhancement measured in 47 Tuc and other metal-rich GCs. Moreover, VD08’s simulations indicate that the Al spread seen in Galactic Galactic GCs can be fit by the Al-yields of AGB stars with a variety of masses. For instance, the Al spread in GCs decrease as metallicity increases, as shown in Fig. \[fig\_boxplotAl\]. This observational trend is also predicted by VD08’s AGB models with masses between 3-6 $M_{\odot}$, which tend to converge towards lower Al-enhancements at higher metallicities, and differ at most by \[Al/Fe\]$\sim$0.4 dex, as can also be seen within the multiple populations in Omega Cen (see Fig. 22 in Johnson & Pilachowski 2010). Thus, the metallicity of the polluters plays an important role constraining not only the degree of Al enhancement but also the spread of Al within a GC. Furthermore, the Al yields of VD08 predict a range of values of $0.3\leq$\[Al/Fe\]$\leq0.7$, which is consistent with the average \[Al/Fe\]=0.38 dex we found in our sample. Metal-poor GCs such as M13 show an Al-Na correlation which was also detected by Carretta et al. (2013) in 47 Tuc (Fig. \[fig\_AlNa\]b). However, this correlation is not present in our sample of red giants in 47 Tuc (Fig. \[fig\_AlNa\]a). Since CN lines were not included in their line list we follow this approach and performed synthesis for a sample of 10 stars covering a Na range of 0.7 dex. We found that the Al abundances are 0.03-0.06 dex higher when the CN lines are removed from the line list. This offset is too small to account for the lack of Al-Na correlation in our data. Interestingly, M71 with a metallicity similar to 47 Tuc exhibits similar star-to-star variations in light-element abundances, including a Na-O anticorrelation, a bimodal CN distribution, and a moderate Al enhancement with a small spread $\langle$\[Al/Fe\]=$+0.24\pm0.10\rangle$. The Al-Na correlation in M71, however, relies on a single star (Ramirez & Cohen 2004). Furthermore, RGB stars in Omega Cen with similar metallicity to 47 Tuc giants do not exhibit an Al-Na correlation (Johnson & Pilachowski 2010). To assess whether the lack of an Al-Na anticorrelation is a common characteristic in metal-rich GCs, large samples of stars with measured Al and Na abundances are needed in other more metal-rich GCs. Summary ======= The motivation for pursuing this project comes from the discovery that most GCs host multiple populations (Carretta et al. 2009a) and the properties of their generations can be used to test models of GC formation or models that describe the nature of the polluters. A direct method to identify the different generations in a GC relies on the chemical content of light-elements. Therefore, we have used abundance ratios of O, Na, and Al to characterize 47 Tuc’s multiple populations. Additionally, we have measure abundance ratios for alpha, Fe peak, and neutron capture group elements, which can be used to compare 47 Tuc with other Galactic GCs and halo stars. We summarize our results as the following. 1\. Chemical abundances for O, Na, Al, Si, Ca, Ti, Fe, Ni, La, and Eu were determined for 164 stars in 47 Tuc using either spectral synthesis or EW analysis. Abundance ratios for the $\alpha$ elements, Si, Ca, and Ti, and the n-capture process elements, Eu and La, indicate that the composition of the first generation stars is dominated by Type II SNe (\[$\alpha$/Fe\]$\approx$0.3 and \[Eu/La\]=+0.24). 2\. Star-to-star variations were detected only for O, Na, and Al. We found that the Na-O anticorrelation is more extended that previously reported reaching \[O/Fe\] ratios as low as $\sim$–0.5 dex, as found typically in the more massive Galactic GCs. Three groups can be identified in the Na-O plane indicating that 47 Tuc hosts multiple populations, with younger stars being enhanced in Na and depleted in O. 3\. For the first time Na and O measurements indicate that the intermediate generation is more centrally concentrated than the primordial generation, which is consistent with current hydrodynamical N-body simulations of GC formation. A KS test indicates that the detected difference in radial distributions is real and not a result of random fluctuations. Furthermore, the radial distributions of first and second generation stars found using Na and O are consistent with the results obtained from photometric data. Additionally, our spectroscopic measurements show that the population ratio varies with distance as predicted by N-body simulations (Vesperini et al. (2013). 4\. No Na-Al correlation is found in 47 Tuc. To assess whether this is characteristic of more metal-rich GCs, large samples in several clusters are needed. Moreover, we found that the different populations can not be identified using their \[Al/Fe\] ratio, since stars with identical atmospheric parameters and different O and Na abundances can display the same Al content. 5\. A single La-rich star (Lee 4710) was found in our sample whose anomalous La abundance could be explained by mass transfer in a close binary system. Comparing radial velocities at different epochs can reveal whether the star is currently in a binary system. A single binary system in our sample is consistent with the small binary fraction measured in this cluster. However, even if this star was originally part of a binary system, two-body interactions with other cluster stars might have made the system unbound. MJC wish to thank Enrico Vesperini for his helpful comments and discussions. We also thank an anonymous referee for his/her thoughtful comments, which have clarified and improved this work. We are grateful to the Cerro Tololo InterAmerican Observatory for assistance in obtaining the observations. This research has made use of the NASA Astrophysics Data System Bibliographic Services. This publication makes use of data products from the Two Micron All Sky Survey, which is a joint project of the University of Massachusetts and the Infrared Processing and Analysis Center/California Institute of Technology, funded by the National Aeronautics and Space Administration and the National Science Foundation. CIJ gratefully acknowledges the support of the National Science Foundation under award AST-1003201. CAP acknowledges the generosity of the Kirkwood Research Fund at Indiana University. Alonso, A., Arribas, S., & Martinez Roger, C. 1999, A&AS, 140, 261 (erratum A&A, 376, 1039 \[2001\]) Alves-Brito, A., Barbuy, B., Ortolani, S., Momany, Y., Hill, V., Zocalli, M., Renzini, A., Minniti, D., Pasquini, L. Bica, E., & Rich, R. M. 2005, A&A, 435, 657 Alves-Brito, A., Yong, D., Meléndez, J., Vásquez, S., & Karakas, A. I. 2012, A&A, 540, A3 Anders, E., & Grevesse, N. 1989, Geochim. Cosmochim. Acta, 53, 197 Anderson, J., Piotto, G., King, I. R., Bedin, L. R., & Guhathakurta, P. 2009, ApJ, 697, L58 Arnould, M., Goriely, S., & Takahashi, K. 2007, Phys. Rep., 450, 97 Baumuller. D., & Gehren, T. 1997, A&A, 325, 1088 Bloecker, T. 1995, A&A, 297, 727 Briley, M. M. 1997, AJ, 114, 1051 Brown, J. A. & Wallerstein, G. 1992, AJ, 104, 1818 Busso, M., Gallino, R., & Wasserburg, G. J. 1999, ARA&A, 37, 239 Carretta, E., Gratton, R. G., Bragaglia, A., Bonifacio, P., & Pasquini, L. 2004, A&A, 416, 925 Carretta, E., Bragaglia, A. Gratton, R. G., and Lucatello, S. 2009a, A&A, 505, 139 Carretta, E., Bragaglia, A., Gratton, R. G., Lucatello, S., Catanzaro, G., Leone, F., Bellazzini, M., Claudi, R., D’Orazi, V., Momany, Y., Ortolani, S., Pancino, E., Piotto, G., Recio-Blanco, A., Sabbi, E. 2009b, A&A, 505, 117 Carretta, E. Gratton, R. G., Bragaglia, A.; D’Orazi, V., & Lucatello, S. 2013, A&A, 550, 34 Castelli, F. & Kurucz, R. L. 2003, Modelling of Stellar Atmospheres (IAU Symp. 210), ed. N. E. Piskunov, W. W. Weiss, & D. F. Gray (San Francisco, CA: ASP), 20 Cavallo, R. M. & Nagar, N. M. 2000, AJ, 120, 1364 Cavallo, R. M., Suntzeff, N. B., & Pilachowski, C. A. 2004, AJ, 127, 3411 Charbonnel, C., Brown, J. A., & Wallerstein, G. 1998, A&A, 332, 204 Chun, M. S. & Freeman, K. C. 1978, AJ, 83, 376 Cohen, J. G. & Meléndez, J. 2005, AJ, 129, 303 Crawford, D. L. & Snowden, M. S. 1975, PASP, 87, 561 Cunha, K., Smith, V. V., Lambert, D. L., & Hinkle, K. H. 2003, AJ, 126, 1305 de Laverny, P. & Recio-Blanco, A. 2013, A&A, to be submitted D’Ercole, A, Vesperini, E, D’Antona, F, McMillan, S. L. W.; & Recchi, S. 2008, MNRAS, 391, 825 Di Criscienzo, M., Ventura, P, D’Antona, F, Milone, A., & Piotto, G. 2010, MNRAS, 408, 999 D’Orazi, V, Lucatello, S., Gratton  R.-G., Bragaglia, A., & Carretta, E., Shen, Z., Zaggia, S. 2010, ApJ l, 713, L5 D’Orazi, V., Lucatello, S., Gratton, R., Bragaglia, S, Carretta, E., & Shen, Z. 2010, ApJ l, 713, L1 Freiburghaus, C., Rosswog, S., & Thielemann, F. K. 1999, ApJ, 525, L121 Giersz, M. & Heggie, D. 2011, MNRAS, 410, 2698 Goriely, S., Bauswein, A., & Janka, H. T. 2011, ApJ, 738, L32 Gratton, R. G., Fusi Pecci, F., Carretta, E., Clementini, G., Corsi, C. E., & Lattanzi, M. 1997, ApJ, 491, 749 Gratton, R. G., Carretta, E., Ericksson, K., & Gustafsson, B. 1999, A&A, 350, 955 Gratton, R.; Sneden, C.; & Carretta, E. 2004, ARA&A, 42, 385 Gratton, R. G., Carretta, E., and Bragaglia, A. 2012, ARA&A, 20, 50 Hansen, B. M. S.; Kalirai, J. S.; Anderson, J.; Dotter, A.; Richer, H. B.; Rich, R. M.; Shara, M. M.; Fahlman, G. G.; Hurley, J. R.; King, I. R.; Reitzel, D. & Stetson, P. B. 2013, Nature, 500, 51 Harris, W. E. 1996, AJ, 112, 1487 Haynes, S., Burks, G., Johnson, C. I., & Pilachowski, C. A. 2008, PASP, 120, 1097 James, G. Francois, P., Bonifacio, P., Carretta, E., Gratton, R. G., & Spite, F. 2004, A&A, 427, 825 Johnson, C. I. & Pilachowski, C. A. 2010, ApJ, 722, 1373 Johnson, C. I. & Pilachowski, C. A. 2012, ApJ, 754, 38 Johnson, C. I. & Pilachowski, C. A. 2006, AJ, 132, 2346 Johnson, C. I., Kraft, R. P., Pilachowski, C. A., Sneden, C., Ivans, I. I., & Benman, G. 2005, PASP, 117, 1308 Johnson, C. I., Pilachowski, C. A., Simmerer, J., & Schwenk, D. 2008, ApJ, 681, 1505 Johnson, C. I., Pilachowski, C. A., Rich, R. M., & Fulbright, J. P. 2009, ApJ, 698, 2048 Koch, A., and McWilliam, A. 2008, AJ, 135, 1551 Korobkin, O., Rosswog, S., Arcones, A., & Winteler, C. 2012, MNRAS, 426, 1940 Kraft, R. P., Sneden, C., Langer, G. E., & Prosser, C. F. 1992, AJ, 104, 645 Kraft, R. P., Sneden, C., & Smith, G. H. 1997, AJ, 113, 279 Langer,  G. F., Hoffman,  R. D., and Zaidins,  C. S. 1997, PASP, 109, 244L Lattimer, J. M., Mackie, F., Ravenhall, D. G., & Schramm, D. N. 1977, ApJ, 213, 225 Lawler, J. E., Bonvallet, G., & Sneden, C. 2001a, ApJ, 556, 452 Lawler, J. E., Wickliffe, M. E., den Hartog, E. A., & Sneden, C. 2001b, ApJ, 563, 1075 Lee, S.-W. 1977, A&AS, 27, 381 Lind, K.; Asplund, M.; Barklem, P. S.; & Belyaev, A. K. 2011, A&A, 528, 103 Mashonkina, L. I., Shimanskiĭ, V. V., & Sakhibullin, N. A. 2000, Astron. Rep., 44, 790 Mathews, G. J. & Cowan, J. J. 1990, Nature, 345, 491 McDonald, I., Boyer, M. L., van Loon, J. Th., Zijlstra, A. A., Hora, J. L., Babler, B., Block, M., Gordon, K., Meade, M., Meixner, M., Misselt, K., Robitaille, T., Sewilo, M., Shiao, B., & Whitney, B. 2011, ApJ s, 193, 23 Milone, A. P., Marino, A. F., Cassisi, S., Piotto, G.; Bedin, L. R., Anderson, J., Allard, F., Aparicio, A., Bellini, A., Buonanno, R., Monelli, M., & Pietrinferni, A. 2012, ApJ, 745, 58 Nataf, D. M., Gould, A., Pinsonneault, M. H., & Stetson, P. B. 2011, ApJ, 736, 94 Nomoto, K.; Tominaga, N.; Umeda, H.; Kobayashi, C.;& Maeda, K. 2006, NuPhA, 777, 424 Norris, J. E. & Da Costa, G. S. 1995, ApJ, 447, 680 Norris, J., & Freeman, K. C. 1979, ApJ l, 230, L179 O’Connell, J. E., Johnson, C. I., Pilachowski, C. A., & Burks, G. 2011, PASP, 123, 1139 Origlia, L., Rich, R. M., Ferraro, F. R., Lanzoni, B., Bellazzini, M., Dalessandro E., Mucciarelli, A., Valenti, E., & Beccari, G. 2011, ApJ, 726, 20 Paltoglou, G., & Freeman, K. C. 1994, Ph.D. thesis, Mt Stromlo and Siding Spring Observatories, Institute of Advanced Studies, Australian National University Pasquini, L., Alonso, J., Avila, G., et al. 2003, Proc. SPIE, 4841, 1682 Pilachowski, C. A., Sneden, C., Kraft, R. P., & Langer, G. E. 1996, AJ, 112, 545 Prantzos, N., Charbonnel, C., Iliadis, C., et al. 2007, A&A, 470, 179 Ramirez, S. & Cohen, J. G. 2002, AJ, 123, 3277 Richer, H. B.; Heyl, J.; Anderson, J.; Kalirai, J. S.; Shara, M. M.; Dotter, A.; Fahlman, G. G.; Rich, R. M. 2013, ApJ, 771, 15 Sakari, C. M.; Shetrone, M.; Venn, K.; McWilliam, A.; & Dotter, A. 2013, MNRAS, 434, 358 Smith, V. V., Cunha, K., Ivans, I. I., Lattanzio, J. C., Campbell, S., & Hinkle, H. 2005, ApJ, 633, 392 Sneden, C. 1973, ApJ, 184, 839 Sneden, C., Cowan, J. J., & Gallino, R. 2008, ARA&A, 46, 241 Sneden, C., Kraft, R. P., Guhathakurta, P., Peterson, R. C., & Fulbright, J. P. 2004, AJ, 127, 2162 Takahashi, K., Witti, J., & Janka, H. T. 1994, A&A, 286, 857 Truran, J. W. 1981, A&A, 97, 391 Truran, J. W., Cowan, J. J., Pilachowski, C. A., & Sneden, C. 2002, PASP, 114, 1293 Tucholke, H.-J. 1992, A&AS, 93, 293 Valcarce, A. A. R. & Catelan, M. 2011, A&A, 533, 120 Ventura, P. & D’Antona, F. 2007, MNRAS, 379, 1431 Ventura, P. & D’Antona, F. 2008, MNRAS, 385, 2034 Ventura, P. & D’Antona, F. 2009, A&A, 479, 805 Vesperini, E.; McMillan, S. L. W.; D’Antona, F.; & D’Ercole, A. 2013, MNRAS, 429, 1913 Woosley, S. E., Wilson, J. R., Mathews, G. J., Hoffman, R. D., & Meyer, B. S. 1994, ApJ, 433, 229 Worley, C. C.; Cottrell, P. L.; Freeman, K. C.; & Wylie-de Boer, E. C. 2009, MNRAS, 400, 1039 Worley, C. C.; Cottrell, P. L.; McDonald, I.; & van Loon, J. Th. 2010, MNRAS, 402, 2060 Worley, C. C., & Cottrell, P. L. 2012, PASP, 29, 29 Yong, D., Meléndez, J., Cunha, K., Karakas, A. I., Norris, J. E., & Smith, V. V. 2008, ApJ, 689, 1020 ![K vs. J-K color magnitude diagram of stars observed in 47 Tuc. Stars assigned to the primordial population are shown in blue, stars assigned to the intermediate population are shown in green, and stars assigned to the extreme population are shown in red (see section 5.1 for the definitions of population groups). Our FLAMES and hydra samples are represented by filled circles and triangles, respectively. Stars from Carretta et al. (2009b) are shown with open symbols. J and K magnitudes were obtained from the 2MASS Point Source Catalog.[]{data-label="fig_CMD"}](Fig1.eps){width="50.00000%"} ![Comparison of our adopted model atmosphere parameters for Hydra and FLAMES observations. The diagonal line indicates perfect agreement.[]{data-label="fig_logTeff"}](Fig2.eps){width="50.00000%"} ![Sample spectrum of D 229 showing the determination of the sodium, oxygen, and aluminum abundances using spectrum synthesis. The observed spectrum is shown as dots, while the synthetic spectra are shown as dashed lines for three different abundances. The adopted abundance is shown as a solid blue line. Differences of 0.2 dex can be easily discerned.[]{data-label="fig_synth"}](Fig3.eps){width="50.00000%"} ![Spectrum synthesis around the oxygen line showing the effect of altering the abundances of blended features. The upper panel shows that the oxygen abundance is unaffected by a change in CN abundance of $\pm$ 0.50 dex (dashed lines). The lower panel shows that the Sc feature must be included in the synthesis to derive a reliable oxygen abundance.[]{data-label="fig_CNSc"}](O_CN_Sc_2panels.eps){width="50.00000%"} ![Derived abundance relative to Fe vs. effective temperature. Measurements from FLAMES data are shown as open circles, and measurements from Hydra data are shown as filled circles. The open square symbol represents the La-rich star.[]{data-label="fig_Abund"}](Fig4.eps){width="50.00000%"} ![Element abundances and ranges measured from the FLAMES + Hydra samples. For each element, the horizontal bar gives the median \[x/Fe\] ratio and the lower and upper box boundaries indicate the first and third quartiles of the data (25% and 75%). The vertical lines indicate the full range of derived abundance.[]{data-label="fig_boxplot"}](Fig5.eps){width="50.00000%"} ![Stars with identical atmospheric parameters exhibiting differences in their line strengths. Interestingly, the Al content is the same for both stars, even though their O and Na abundances differ significantly.[]{data-label="fig_strengths"}](Fig6.eps){width="50.00000%"} ![A spectrum of the La-rich giant Lee 4710 vs. Lee 4626, a giant of similar atmospheric parameters with normal La. Identified features of Ti, V, Fe, and Ni show similar strength in the two stars, while the features of La II at 6262 and 6774 Å  are stronger in Lee 4710 than in Lee 4626. []{data-label="fig_binary"}](Fig7.eps){width="50.00000%"} ![The Na-O anticorrelation. Stars assigned to the primordial population are shown in blue, stars assigned to the intermediate population are shown in green, and stars assigned to the extreme population are shown in red. Our FLAMES and Hydra samples are represented by filled circles, while from Carretta et al. (2009b) are shown with open symbols. Stars are assigned to the primordial, intermediate, and extreme populations as defined in the text and indicated by solid lines in the figure.[]{data-label="fig_NaO"}](Fig8.eps){width="50.00000%"} ![Histogram of the \[Na/O\] abundance ratio distribution showing two peaks with an extended tail.[]{data-label="fig_histo"}](NaO_histogram_us_carretta010.eps){width="50.00000%"} ![The \[Al/Fe\] abundance in 47 Tuc from the FLAMES + Hydra samples compared to other globular clusters. For each cluster, the horizontal bar gives the median \[Al/Fe\] ratio and the lower and upper box boundaries indicate the first and third quartiles of the data (25% and 75%). The vertical lines indicate the full range of the derived aluminum abundances in each cluster. Data sources are as follows: M80 from Cavallo et al. (2004); M13 from Johnson et al. (2005); M30, M68, M55, M10, NGC 6752, M12, NGC 288, and M4, and M71, from Carretta et al. (2009b); M107 from O’Connell et al. (2011); Terzan 5 from Origlia et al. (2011), where a and b represents the populations with \[Fe/H\]=–0.25 and \[Fe/H\]=+0.27, respectively.[]{data-label="fig_boxplotAl"}](Fig9.eps){width="50.00000%"} ![\[Al/Fe\] vs. \[Na/Fe\]. Stars assigned to the primordial population are shown in blue, stars assigned to the intermediate population are shown in green, and stars belonging to the extreme population are shown in red. Stars not classified according to population are shown in cyan. The upper panel (a) shows that in our sample both the primordial and intermediate populations exhibit a similar mean aluminum abundance and dispersion. The lower panel (b) shows results from Carretta et al. 2013.[]{data-label="fig_AlNa"}](Fig10.eps){width="50.00000%"} ![(upper panel) Cumulative fraction of stars as a function of half-mass radius for the FLAMES and Hydra samples. The curve for intermediate population stars is shown in green and the curve for primordial population stars is shown in blue. The intermediate population is more centrally concentrated. A p-value of 0.02 from the KS test allows us to reject the null hypothesis that the two samples are drawn from the same population. (lower panel) Cumulative fraction of stars as a function of half-mass radius for our data plus the Carretta et al. (2009b) sample. The cumulative fraction for the combined sample is shown with a black curve, and the extreme population is shown as a red curve. As in Figure 9, the curve for intermediate population stars is shown in green and the curve for primordial population stars is shown in blue. The intermediate population is more centrally concentrated. A p-value of $2.3\times10^{-4}$ from the KS test allows us to reject more strongly the null hypothesis that the samples are drawn from the same population.[]{data-label="fig_cum"}](Fig11.eps){width="50.00000%"} ![The fraction of subsequent (intermediate plus extreme) generation stars in 47 Tucanae vs. cluster half mass radius. Data from the combined FLAMES, Hydra, and Carretta et al. (2009b) are shown in red, while population fractions from Milone et al. (2012) are shown in black. The intermediate plus extreme populations dominate in the central regions, while the primordial population dominates beyond about 1.5 half-mass radii from the cluster center.[]{data-label="fig_popratio"}](Fig12.eps){width="50.00000%"} -------------------- ----------- ------------ UT Date Exp. Time Wavelength & Time (s) ($\AA$) 18 July 2003 06:53 3600 6490-6800 18 July 2003 07:54 3600 6490-6800 18 July 2003 09:00 3600 6490-6800 19 July 2003 06:41 3600 6490-6800 19 July 2003 07:50 3600 6490-6800 19 July 2003 08:52 3600 6490-6800 19 July 2003 10:00 1800 6490-6800 23 Aug 2010 06:32 2700 6100-6400 23 Aug 2010 07:21 2700 6100-6400 23 Aug 2010 08:09 2700 6100-6400 23 Aug 2010 08:58 2700 6100-6400 26 Nov 2011 00:26 2100 6110-6396 26 Nov 2011 01:05 1020 6300-6390 26 Nov 2011 01:25 960 6600-6989 -------------------- ----------- ------------ : Log of 47 Tucanae Observations[]{data-label="table_log"} [lrrrcrrrrrrrrrcr]{} &T$_{\text{eff}}$ & & & $v_t$ & & & & & & & & & & & r\ Star & (K) & $\log g$ & \[Fe/H\] & (km s^-1^) & \[O/Fe\] & \[Na/Fe\] & \[Al/Fe\] & \[Si/Fe\] & \[Ca/Fe\] & \[Ti/Fe\] & \[Ni/Fe\] & \[La/Fe\] & \[Eu/Fe\] & Pop & (arcmin)\ Hydra\ 1628 & 4350 & 1.5 & -0.73 & 1.6 & 0.29 & 0.2 & & 0.18 & 0.14 & 0.31 & -0.11 & 0.07 & & P & 9.53\ 1646 & 4150 & 1.2 & -0.72 & 1.75 & -0.1 & 0.44 & & 0.2 & 0.21 & 0.3 & 0.02 & 0.12 & & I & 5.46\ 1669 & 4775 & 1.95 & -0.79 & 2.35 & 0.21 & 0.735 & & 0.39 & 0.39 & 0.58 & -0.02 & 0.4 & & I & 5.99\ 1735 & 4525 & 1.7 & -0.73 & 1.75 & 0.255 & 0.32 & 0.37 & 0.02 & 0.2 & 0.21 & -0.15 & 0.17 & 0.47 & I & 7.55\ 1747 & 4025 & 1.05 & -0.81 & 2.0 & 0.03 & 0.16 & & 0.60 & 0.07 & 0.01 & -0.01 & 0.10 & & P & 6.85\ \ FLAMES\ 1628 & 4400 & 1.23 & -1.00 & 1.80 & 0.35 & 0.43 & 0.62 & 0.27 & 0.47 & 0.54 & -0.01 & 0.14 & 0.39 & I & 9.55\ 1669 & 4615 & 2.10 & -0.72 & 1.60 & 0.07 & 0.59 & 0.55 & 0.22 & 0.29 & 0.28 & -0.03 & 0.29 & 0.49 & I & 5.98\ 1703 & 4575 & 2.00 & -0.79 & 1.55 & 0.36 & 0.31 & 0.37 & 0.36 & 0.37 & 0.43 & 0.07 & 0.23 & 0.43 & P & 8.09\ 1714 & 4450 & 1.75 & -0.82 & 1.65 & 0.32 & 0.28 & 0.47 & 0.29 & 0.29 & 0.41 & 0.03 & 0.23 & 0.51 & P & 7.77\ 1716 & 4700 & 2.47 & -0.66 & 1.35 & 0.25 & 0.25 & 0.36 & 0.21 & 0.21 & 0.18 & -0.05 & 0.35 & & P & 7.72\ ------- ----------------------- ------------------- ------------------- ------------------- ------- ------------------------ ------------------------- -------- Ion $T_{\mbox{eff}}\pm50$ $\log g\pm0.20$ \[M/H\]$\pm$0.10 $v_t\pm$0.25 No. $\sigma_{\mbox{obs.}}$ $\sigma_{\mbox{total}}$ offset (K) (cgs) (dex) (km s$^{-1}$) lines (dex) (dex) (dex) Fe I $\mp 0.01 $ $\mp 0.03 $ $\pm 0.02 $ $\pm 0.14 $ 20,50 0.03,0.02 0.15 –0.08 Fe II $\mp 0.07 $ $\pm 0.08 $ $\pm 0.05 $ $\mp 0.03 $ 0,6 ,0.08 ,0.15 O I $\pm 0.03 $ $\pm 0.09 $ $\pm 0.05 $ $\pm 0.05 $ 1,1 0.10 0.15 –0.07 Na I $\pm 0.03 $ $\mp 0.05 $ $\pm 0.05 $ $\mp 0.05 $ 2,2 0.10 0.14 +0.07 Al I $\pm 0.03 $ $\mp 0.02 $ $\mp 0.02 $ $\mp 0.05 $ 2,2 0.10 0.12 +0.07 Si I $\mp 0.04 $ $\pm 0.03 $ $\pm 0.03 $ $\mp 0.03 $ 1,2 0.15,0.11 0.16,0.13 0.00 Ca I $\pm 0.07 $ $\mp 0.02 $ $\pm 0.00 $ $\mp 0.15 $ 2,2 0.11 0.20 –0.09 Ti I $\pm 0.10 $ $\pm 0.01 $ $\pm 0.00 $ $\mp 0.09 $ 2,2 0.11 0.17 +0.02 Ni I $\pm 0.00 $ $\pm 0.04 $ $\pm 0.02 $ $\mp 0.09 $ 2,2 0.11 0.15 –0.05 La II $\pm 0.01 $ $\pm 0.09 $ $\pm 0.06 $ $\pm 0.01 $ 1,2 0.10,0.08 0.15,0.14 +0.04 Eu II $\mp 0.03 $ $\pm 0.05 $ $\pm 0.05 $ $\mp 0.03 $ 1,1 0.10 0.13 +0.04 ------- ----------------------- ------------------- ------------------- ------------------- ------- ------------------------ ------------------------- -------- : Abundance Sensitivity to Model Atmosphere Parameters.[]{data-label="table_sensitivities"} [^1]: Visiting Astronomer, Cerro Tololo Inter-American Observatory. CTIO is operated by AURA, Inc. under contract to the National Science Foundation; National Science Foundation Astronomy and Astrophysics Postdoctoral Fellow; Clay Fellow. [^2]: Visiting Astronomer, Cerro Tololo Inter-American Observatory. CTIO is operated by AURA, Inc. under contract to the National Science Foundation [^3]: The E6257 filter is on permanent loan at CTIO and is available for public use. [^4]: IRAF is distributed by the National Optical Astronomy Observatories, which are operated by the Association of Universities for Research in Astronomy, Inc., under cooperative agreement with the National Science Foundation." [^5]: The girBLDRS data reduction package can be downloaded at: http://girbldrs.sourceforge.net/ [^6]: We use the standard spectroscopic notation where \[A/B\]$\equiv$log(N$_{\rm A}$/N$_{\rm B}$)$_{\rm star}$–log(N$_{\rm A}$/N$_{\rm B}$)$_{\odot}$ and log $\epsilon$(A)$\equiv$log(N$_{\rm A}$/N$_{\rm H}$)+12.0 for elements A and B. [^7]: The NLTE corrections (Gratton et al. (1999)) used by Carretta et al. are positive and range from +0.05 to +1.4 dex at 47 Tuc’s metallicity. Therefore, to avoid systematic offsets introduced by using different NLTE corrections, we have eliminated NLTE corrections from Carretta’s sample in order to place their sample into our baseline. [^8]: We adopt a threshold of P=0.05 for the rejection of the null hypothesis that the samples are drawn from the same population.

ArXiv

--- abstract: 'The ground state correlation energies associated with collective surface and pairing vibrations are calculated for Pb- and Ca-isotopes. It is shown that this contribution, when added to those predicted by one of the most accurate modern nuclear mass formula (HFBCS MSk7 mass formula), reduces the associated rms error by an important factor, making mean field theory, once its time dependence is taken into account, a quantitative predictive tool for nuclear masses.' author: - 'S.Baroni$^{a}$, M.Armati$^{a}$, F.Barranco$^{c}$, R.A.Broglia$^{a,b,d}$, G.Colò $^{a,b}$, G.Gori$^{a,b}$ and E.Vigezzi$^{b}$' title: Correlation energy contribution to nuclear masses --- Many of the nuclei in atoms we find in nature are processed in the stars. In particular, heavy elements far beyond $^{56}Fe$ are formed through a chain of neutron capture reactions with subsequent beta decay. The source of the rather low energy neutrons are reactions like ($\alpha$,n) on $^{22}Ni$ and $^{13}C$ where the alphas are highly abundant during He burning in the stars. The small neutron abundance then causes a particular path to be followed which lies mainly along the valley of beta stability as the capture process follows a time scale controlled by the variety of $\beta$-decay processes. The end point of this slow processes (s-process) is above $^{209}Bi$ via $\alpha$-decay. There is clear evidence for sharp peaks in the abundance curve near neutron closed shell configurations which commonly result from the existence of slow neutron capture processes. However, the abundance curve displays a number of additional peaks at around 8-12 mass units prior to the s-process abundance peaks. This indicates that other capture processes where neutron capture is involved have to take place to produce the corresponding elements as well as elements beyond $^{209}Bi$. This is called the rapid or r-process. It is known that the r-process path evolves along the neutron drip line region [@1], where the binding energy of the least bound neutron is quite small (few hundred KeV) in the presence of a large neutron flux [@Rol.Ro:1988]. The need for a formula able to predict known masses with an accuracy of this order of magnitude seems quite natural. On the other there still exist many uncertainties in extrapolating our present-day knowledge of formulas which predict nuclear masses, towards the region of the drip line where weakly bound nuclei are found. In fact, the best account of the experimental data based on mean-field theory provides a fitting to the 2135 measured masses with N,Z $>$ 8 with a rms error of 0.674 MeV [@Gor.ea:2002]. This has been achieved by means of Hartree-Fock-Bardeen-Cooper-Schrieffer (HFBCS) calculations which employ a Skyrme-type zero-range effective force in the mean field channel, supplemented by a zero-range pairing interaction. The 14 parameters set is named BSk2. As a reference point for our work, we have considered a parameter set of almost equal quality, denoted by MSk7, where the rms error is 0.738 MeV [@Gor.ea:2001]. Although this accuracy is remarkable, one is still not satisfied. In fact, while it takes tens of MeV to remove two nucleons from a nucleus lying along the valley of stability, this quantity becomes as low as 300 keV in halo nuclei like $^{11}Li$ [@Li]. The need for a mass formula at least a factor of two more accurate than the MSk7 mass formula is evident. Nuclei display both single-particle and collective degrees of freedom. Consequently, the corresponding ground states and associated nuclear masses reflect the zero point fluctuations (ZPF) associated with these modes. While mean field theory includes fluctuations associated with quasiparticles, it is only time dependent mean field theory which takes into account the zero point fluctuations associated with collective modes. Despite this, the information in the literature concerning ground state correlation energies associated with collective modes is rather scarce. A recent suggestion to consider their effect was put forward by G.F. Bertsch and K. Hagino [@BH]. Later, realistic calculations for the quadrupole degree of freedom were performed for light nuclei [@SJ] and for a few selected isotopes within the so-called Generator Coordinate Method (GCM) [@BBH]. In the present paper we work out, making use of Random Phase Approximation (RPA) calculations, the ground state correlation energies associated with both surface (quadrupole and octupole modes) and pairing vibrations for the Ca- and Pb-isotopes. Because pairing vibrations [@Bes.Br:1966] have a collective character only around closed shell nuclei (being essentially pure two-quasiparticle states lying on top of twice the pairing gap in superfluid systems, cf. also [@And:1958]), one expects the associated ZPF to lead to important corrections to the mass formula [@Gor.ea:2002]. This in keeping with the fact that the largest deviation from experiment found in this mass formula are observed in closed shell systems. It will be concluded that ZPF are important, in particular those associated with pairing vibrations in closed shell nuclei, in reducing the rms of the Skyrme-HFBCS mass formulas. To derive the particle-hole RPA equations use can be made of the quasi-boson approximation where the particle-hole operators $a^{+}_{k}a_{i}$ and $(a^{+}_{k}a_{i})^{+}$ are replaced by the boson operators $\Gamma^{+}_{ki}$ and $\Gamma_{ki}$. Because collective vibration can be viewed as correlated particle-hole excitations, the corresponding boson creation operator can be written as $$\label{eq: 10} \Gamma^{+}_{\alpha}(n)=\sum_{ki}(X_{ki}^{\alpha}(n)\Gamma^{+}_{ki}+ Y_{ki}^{\alpha}(n)\Gamma_{ki}),$$ where $X_{ki}^{\alpha}$ and $Y_{ki}^{\alpha}$ are the forwardsgoing and backwardsgoing amplitudes fullfilling the normalization condition $$\label{eq: 15} [\Gamma_{\alpha}(n),\Gamma^{+}_{\alpha}(n)]= \sum_{ki}({|X_{ki}^{\alpha}|}^{2}-{|Y_{ki}^{\alpha}|}^{2})=1.$$ The equations which determine the frequencies of the vibrational modes of quantum number $\alpha$ (with progressively high energy $n=1,2,\ldots$) are obtained from the relation $$\label{eq: 20} [H,\Gamma^{+}_{\alpha}(n)]=\hbar\omega_{\alpha}(n)\Gamma^{+}_{\alpha}(n).$$ In the above equation $H$ is the total hamiltonian, sum of a single-particle and a two-body interaction term. In the present calculations we have used self-consistently the Skyrme plus pairing interaction given by the MSk7 parameter set. The RPA ground state energy is given by (cf. e.g. [@Rin.Sc:1980]) $$\label{eq: 30} E_{RPA}=E_{HF}-(2\lambda + 1)\sum_{\alpha,n}\hbar\omega_{\alpha}(n) \sum_{ki}\left| Y_{ki}^{\alpha}(n) \right|^{2},$$ in keeping with the fact that the amplitudes $Y_{ki}^{\alpha}(n)$ are directly related to the ground state correlations induced by the corresponding vibrational modes. The second term of the r.h.s. is called correlation energy. It is well known that open shell nuclei display a finite BCS gap $\Delta$, as a result of the pairing correlations acting between pairs of nucleons moving in time reversal states close to the Fermi energy. In closed shell nuclei, pairing correlations are not strong enough to overcome the large single-particle gap, and $\Delta=0$. Nevertheless, pairing plays an important role in determining the (low-lying) structure of these nuclei. In particular, it leads to pairing vibrations [@Bes.Br:1966], which change in two the number of nucleons, that is vibrations which can be viewed as two correlated particles (pair addition modes) or two correlated holes (pair removal modes). An example of these modes is provided by the ground state of $^{210}Pb$ and of $^{206}Pb$ which can be viewed as the pair addition and the pair subtraction modes of $^{208}Pb$ respectively. These modes are specifically probed through two particle transfer reactions [@Bro.ea:ADV]. The pair addition and pair subtraction modes can be written as $$\label{eq: 40} \Gamma^{+}_{a}(n)=\sum_{k}X_{k}^{a}(n)\Gamma^{+}_{k}+ \sum_{i}Y_{i}^{a}(n)\Gamma_{i},$$ and $$\label{eq: 50} \Gamma^{+}_{r}(n)=\sum_{i}X_{i}^{r}(n)\Gamma^{+}_{i}+ \sum_{k}Y_{k}^{r}(n)\Gamma_{k},$$ where $\Gamma^{+}_{k}=[a^{+}_{k}a^{+}_{k}]_{0}$ creates a pair of nucleons coupled to angular momentum zero in levels with energy larger than the Fermi energy ($\varepsilon_{k}>\varepsilon_{F}$), while $\Gamma_{i}=[a_{i}a_{i}]_{0}$ annihilates a pair of nucleons in occupied states ($\varepsilon_{i}\leq \varepsilon_{F}$). Use is made of a hamiltonian $ H=H_{sp}+H_{p}$, sum of a single-particle term and of a pairing force with constant matrix $ H_{p}=-G\sum_{jj'}a^{+}_{j^{\prime}}a^{+}_{\tilde{j^{\prime}}}a_{\tilde{j}}a_{j}$, the commutation relation given in Eq.(\[eq: 20\]) leads to the dispersion relations $$\label{eq: 80} \frac{1}{G}=\sum_{k}\frac{\Omega_{k}}{2e_{k}-\hbar\omega_{a}(n)}+ \sum_{i}\frac{\Omega_{i}}{2e_{i}+\hbar\omega_{a}(n)},$$ $$\label{eq: 90} \frac{1}{G}=\sum_{k}\frac{\Omega_{k}}{2e_{k}+\hbar\omega_{r}(n)}+ \sum_{i}\frac{\Omega_{i}}{2e_{i}-\hbar\omega_{r}(n)},$$ where $\Omega=(2j+1)/2$ is the pair degeneracy of the single-particle orbital with total angular momentum $j$, while $e_{j}=\varepsilon_{j}-\varepsilon_{F}$. The amplitudes are $$\label{eq: 100} X_{k}^{a}(n)=\frac{\Lambda_{a}(n)\sqrt{\Omega_{k}}}{2e_{k}-\hbar\omega_{a}(n)},\quad Y_{i}^{a}(n)=\frac{\Lambda_{a}(n)\sqrt{\Omega_{i}}}{2e_{i}+\hbar\omega_{a}(n)},$$ and $$\label{eq: 110} X_{i}^{r}(n)=\frac{\Lambda_{r}(n)\sqrt{\Omega_{i}}}{2e_{i}-\hbar\omega_{r}(n)},\quad Y_{k}^{r}(n)=\frac{\Lambda_{r}(n)\sqrt{\Omega_{k}}}{2e_{k}+\hbar\omega_{r}(n)},$$ $\Lambda_{a}(n)$ and $\Lambda_{r}(n)$ being normalization constants determined from the relation given in Eq. (\[eq: 15\]). In Fig. \[fig: 1\] we show the dispersion relations given in Eqs.(\[eq: 80\]) and (\[eq: 90\]) calculated for $^{208}Pb$ for both protons and neutrons (cf. also [@Bes.Br:1966]), making use of the valence orbitals of this nucleus. The valence orbitals used in our calculations were determined with the help of a Woods-Saxon potential with standard parametrization [@Boh.Mo:1969] and the energies have been replaced with the experimental values whenever available. The results obtained using the standard Woods-Saxon levels coincide with those calculated making use of the experimental energies within 2%. ![Monopole-pairing-vibrations dispersion relation for (a) neutrons and (b) protons for the nucleus $^{208}Pb$[]{data-label="fig: 1"}](grafico_curva_esp_neut_Pb208b.eps "fig:"){width="20.00000%"}![Monopole-pairing-vibrations dispersion relation for (a) neutrons and (b) protons for the nucleus $^{208}Pb$[]{data-label="fig: 1"}](grafico_curva_esp_prot_Pb208b.eps "fig:"){width="20.00000%"} Making use of the fact that the sum of the pairing binding energies of $^{206}Pb$ and $^{210}Pb$ as well as in $^{206}Hg$ and $^{210}Po$ are $\approx$ 2 MeV (in this last case one has to take into account the Coulomb repulsion between the two protons, cf. e.g. [@Bor.ea:1977]), one obtains the values of 2.7 MeV and 2.2 MeV for the neutron pair addition and pair removal energies, the corresponding values for the proton channel being 3.5 MeV and 3.1 MeV respectively (cf. refs [@Bes.Br:1966], [@Bor.ea:1977] and [@Bes.Br:1971]). The associated $Y$’s amplitudes are displayed in Tables \[table: 1\] and \[table: 2\]. -- ------------- -------- -------- ------------------------------ --------- $3d_{3/2}$ 0.0618 0.1053 $2g_{7/2}$ 0.0882 0.1512 $2g_{7/2}\otimes 1i_{11/2}$ 0.1837 $4s_{1/2}$ 0.0480 0.0886 $1j_{15/2}\otimes 1j_{15/2}$ -0.1530 $3d_{5/2}$ 0.0915 0.1864 $2d_{5/2}\otimes 2g_{9/2}$ 0.3005 $1j_{15/2}$ 0.1542 0.3259 $2i_{11/2}\otimes 2i_{11/2}$ -0.2169 $1i_{11/2}$ 0.1556 0.4122 $2i_{11/2}\otimes 2g_{9/2}$ -0.1094 $2g_{9/2}$ 0.1774 0.8439 $2g_{9/2}\otimes 2g_{9/2}$ -0.8783 $3p_{1/2}$ 0.7853 0.0839 $3p_{1/2}\otimes 2f_{5/2}$ 0.8431 $2f_{5/2}$ 0.4841 0.1209 $3p_{1/2}\otimes 3p_{3/2}$ 0.3756 $3p_{3/2}$ 0.2879 0.0899 $2f_{5/2}\otimes 2f_{5/2}$ -0.2614 $1i_{13/2}$ 0.3347 0.1402 $2f_{5/2}\otimes 3p_{3/2}$ -0.1378 $2f_{7/2}$ 0.1856 0.0914 $3p_{3/2}\otimes 3p_{3/2}$ -0.1186 $1h_{9/2}$ 0.1461 0.0839 $3p_{3/2}\otimes 2f_{7/2}$ 0.1422 -- ------------- -------- -------- ------------------------------ --------- : RPA wavefunctions of the neutron pair addition (a) and pair removal (r) modes of $^{208}Pb$ with multipolarities and parity $\lambda^{\pi}=0^{+},2^{+}$.[]{data-label="table: 1"} Inserting these results in Eq.(\[eq: 30\]), one obtains the ground state correlation energy values -0.399 MeV (neutrons) and -0.449 MeV (protons) respectively. In all calculations we have kept the contribution of only the lowest ($n = 1$) pair addition and pair subtraction modes, in keeping with the fact that, as a rule, the $n\neq 1$ modes are much less collective. Pairing vibrations with multipolarity $\lambda \neq 0$, in particular quadrupole and hexadecapole pairing vibrations, have also been identified around closed shell nuclei (cf. [@Bro.ea:ADV], [@Bor.ea:1977], [@Bes.Br:1971] and refs therein). In Table \[table: 3\] we display the contributions to the ground state energy (i.e. $E_{RPA}$ as defined in Eq.(\[eq: 30\])) associated with the monopole, quadrupole and hexadecapole pair addition and pair removal modes for both neutrons and protons associated with $^{208}Pb$, the summed contribution amounting to -1.981 MeV ($\approx -1.196$ MeV $-0.785$ MeV). -- ------------- -------- -------- ------------------------------ --------- $3p_{1/2}$ 0.0431 0.0678 $3p_{3/2}$ 0.0655 0.1090 $1i_{13/2}\otimes 1i_{13/2}$ -0.1097 $2f_{5/2}$ 0.0835 0.1439 $2f_{5/2}\otimes 1h_{9/2}$ 0.1587 $1i_{13/2}$ 0.1547 0.3331 $2f_{7/2}\otimes 2f_{7/2}$ -0.1359 $2f_{7/2}$ 0.1337 0.3601 $2f_{7/2}\otimes 1h_{9/2}$ 0.1000 $1h_{9/2}$ 0.1822 0.8798 $1h_{9/2}\otimes 1h_{9/2}$ -0.9602 $3s_{1/2}$ 0.6862 0.0780 $3s_{1/2}\otimes 2d_{3/2}$ 0.9271 $2d_{3/2}$ 0.5634 0.1010 $2d_{3/2}\otimes 2d_{3/2}$ -0.2738 $1h_{11/2}$ 0.4463 0.1414 $3s_{1/2}\otimes 2d_{5/2}$ 0.1814 $2d_{5/2}$ 0.2672 0.0939 $1h_{11/2}\otimes 1h_{11/2}$ -0.1222 $1g_{7/2}$ 0.1711 0.0824 $2d_{3/2}\otimes 1g_{7/2}$ 0.1022 -- ------------- -------- -------- ------------------------------ --------- : RPA wavefunctions of the proton pair addition (a) and pair removal (r) modes of $^{208}Pb$ with multipolarities and parity $\lambda^{\pi}=0^{+},2^{+}$.[]{data-label="table: 2"} -------- -------- -------- -------- -------- -------- n p n p n p -0.399 -0.449 -0.609 -0.244 -0.189 -0.092 -------- -------- -------- -------- -------- -------- : Ground state correlation energies, arizing from the neutron (n) and protons (p) monopole, quadrupole and hexadecapole pairing vibrations in $^{208}Pb$.[]{data-label="table: 3"} $^{204}Pb$ $^{206}Pb$ $^{208}Pb$ $^{210}Pb$ $^{212}Pb$ ---------------------- ------------ ------------ ------------ ------------ ------------ *p-h vibrations* -2.793 -2.709 -2.237 -2.801 -3.173 *pairing vibrations* -0.785 -0.785 -1.981 -0.785 -0.785 : Ground state correlation energies for the Pb-isotopes.[]{data-label="table: 4"} $^{40}Ca$ $^{42}Ca$ $^{44}Ca$ $^{46}Ca$ $^{48}Ca$ ---------------------- ----------- ----------- ----------- ----------- ----------- *p-h vibrations* -0.886 -1.418 -1.606 -1.391 -0.547 *pairing vibrations* -4.761 -2.978 -3.239 -3.500 -5.823 : Ground state correlation energies for the Ca-isotopes.[]{data-label="table: 5"} ---------------- --------------------- --------------------------- ref. [@Gor.ea:2001] ref. [@Gor.ea:2001] + RPA Pb 0.646 0.543 Ca 1.200 0.466 $\bar{\sigma}$ 0.964 0.505 ---------------- --------------------- --------------------------- : Root mean square error associate with the HFBCS MSk7 mass formula of ref. [@Gor.ea:2001] and to this formula (with sligthly adjusted parameters) plus the correlation contributions associated with surface and pairing vibrations calculated in the RPA. The quantity $\bar{\sigma}=\left(\frac{\sigma_{Ca}^2+\sigma_{Pb}^2}{2}\right)^{1/2}$ is shown in the last line.[]{data-label="table: 6"} In Table \[table: 4\] we collect the corresponding contribution for a number of Pb-isotopes. Because pairing vibrations are collective modes only around closed shell nuclei, where particles and holes can be clearly distinguished, becoming non-collective two-quasiparticle modes outside closed shells (cf. e.g. ref. [@Li]) we have considered the contribution of neutron pairing vibrations only for the closed shell system (while the proton pairing vibration were taken into account for all isotopes). Also shown in table \[table: 4\] are the contribution to $E_{RPA}$ arising from the low-lying collective particle-hole vibrations calculated making use of the MSk7 interaction to determine the single particle states and the particle-hole correlated modes (cf. Eq.(\[eq: 10\])). Quadrupole and octupole vibrations with energy $<$ 7 MeV, and exausting $\geq 2 \%$ of the non-energy weighted sum rule were included in the calculation of $E_{RPA}$. These conditions essentially select the lowest (one-two) states displaying correlated wavefunctions. Similar calculations were repeated for the calcium isotopes $^{40-48}Ca$. In this case, the contribution of the proton pairing vibrations calculated for the closed shell systems $^{40}Ca$ and $^{48}Ca$ were linearly interpolated for the other isotopes, and no hexadecapole modes were considered. In Table \[table: 5\] we show the corresponding results, together with the contribution of the particle-hole vibrational modes. When adding the results of Tables \[table: 4\] and \[table: 5\] to the HFBCS MSk7 mass formula of Ref. [@Gor.ea:2001], the parameters of the Skyrme interaction should be refitted in order to provide the best reproduction of experimental masses. This should be done on a large sample of isotopes and it is beyond the purpose of the present paper. If we restrict ourselves to Ca-isotopes (Pb-isotopes), a slight renormalization by a factor 0.9964 (0.99945) of the Skyrme parameters $t_{i}$,$W_{0}$ is enough to shift the calculated value of the masses upwards by $\approx$ 4.3 MeV (3.5 MeV) and minimize the rms deviation, giving the results displayed in Table \[table: 6\], and, for the Calcium isotopes, also shown in Fig. \[fig: 2\]. Averaging the rms deviations associated with Ca- and Pb-isotopes, leads to a value of 0.505 MeV as compared to the value of 0.964 MeV obtained making use of the results of ref. [@Gor.ea:2002]. Although a global readjustment of the mean field parameters should be envisaged, the fact that the locally extracted rms deviations have been reduced by a factor of $\approx$ 2 can be considered meaningful. Similarly to what was observed in assessing the role played by zero point fluctuations associated with surface and pairing vibrations in connection with the nuclear mean square radius [@14; @15; @16], the alignment of rotational nuclei [@17] and the pairing phase transition as a function of angular momentum [@Bro.ea:1968; @18; @19], we conclude that zero point fluctuations play an important role in providing the A-dependent contributions to nuclear masses needed to make mean field theory, including its time dependence, a quantitative predictive tool. [99]{} When neutrons are progressively added to a normal nucleus, the Pauli principle forces them into states of higher momentum. When the core becomes neutron saturated, the nucleus expels most of the wavefunction of the last neutrons outside, to region of lower density, which because of its larger size can have lower momentum. In most cases neutrons essentially drip off from the nucleus, defining the lines of stability for neutron number in the chart of nuclides (i.e. the neutron drip line for which the neutron binding energy is equal to zero). C.E.Rolfs and W.S.Rodney, *Cauldrons in the cosmos*, The University of Chicago Press, Chicago (1988) S.Goriely, M.Samyn, P.H.Heenen, J.M.Pearson and F.Tondeur, Phys. Rev. C [**66**]{} (2002) 024326-1 S.Goriely, F.Tondeur and J.M.Pearson, At. Data Nucl. Data Tables [**77**]{} (2001) 311. I. Tanihata, J. Phys. [**G22**]{} (1996) 157. G.F. Bertsch and K. Hagino, Phys. Atom. Nucl. [**64**]{}(2001)533 I. Stetcu and C.W.Johnson, Phys. Rev. C [**66**]{} (2002) 034301. M.Bender, G.F.Bertsch, P.-H.Heenen, [nucl-th/0311026]{}. D.R.Bes and R.A.Broglia, Nucl. Phys. [**80**]{} (1966) 289 P.W.Anderson, Phys. Rev. [**112**]{} (1958) 1900 P.Ring and P.Schuck, *The nuclear many-body problem*, Springer Heidelberg (1980) R.A.Broglia, O.Hansen and C.Riedel, Advances in Nuclear Physics, Ed. M. Baranger and E. Vogt, Plenum Press, Vol.[**6**]{}, 287 (1973) P.F.Bortignon, R.A.Broglia, D.R.Bes and R.Liotta, Phys.Rep. [**30C**]{} (1977) 305 D.R.Bes and R.A.Broglia, Phys. Rev. C [**6**]{} (1971) 2349 A.Bohr and B.R.Mottelson, *Nuclear Structure*, Vol. I, Benjamin, New York (1969) H. Esbensen and G.F. Bertsch, Phys. Rev. [**C28**]{} (1983)355 F. Barranco and R.A.Broglia, Phys. Lett. [**151B**]{} (1985) 90 F. Barranco and R.A.Broglia, Phys. Rev. Lett. [**59**]{} (1987) 2724 F. Barranco, M. Gallardo and R.A.Broglia, Phys. Lett. [**198B**]{} (1987) 19 R.A.Broglia, C.Riedel and B.Sørensen, Nucl. Phys. [**A115**]{} (1968) 273 Y. R. Shimizu, J. Garrett, R.A.Broglia, M. Gallardo and E. Vigezzi, Rev. of Mod. Phys. [**61**]{} (1989) 131 R.S. Nikam and P. Ring, Phys. Rev. Lett. [**58**]{}(1987)980

ArXiv

--- abstract: 'The galaxy population at $z\approx6$ has been the subject of intense study in recent years, culminating in the Hubble Ultra Deep Field (HUDF) – the deepest imaging survey yet. A large number of high redshift galaxy candidates have been identified within the HUDF, but until now analysis of their properties has been hampered by the difficulty of obtaining spectroscopic redshifts for these faint galaxies. Our “Gemini Lyman-Alpha at Reionisation Era” (GLARE) project has been designed to undertake spectroscopic follow up of faint ($z''<28.5$) $i''$-drop galaxies at $z\approx 6$ in the HUDF. In a previous paper we presented preliminary results from the first 7.5 hours of data from GLARE. In this paper we detail the complete survey. We have now obtained 36 hours of spectroscopy on a single GMOS slitmask from Gemini-South, with a spectral resolution of $\lambda/\Delta\lambda_{\rm FWHM} \approx 1000$. We identify five strong Lyman-$\alpha$ emitters at $z>5.5$, and a further nine possible line emitters with detections at lower significance. We also place tight constraints on the equivalent width of Lyman-$\alpha$ emission for a further ten $i''$-drop galaxies and examine the equivalent width distribution of this faint spectroscopic sample of $z\approx6$ galaxies. We find that the fraction of galaxies with little or no emission is similar to that at $z\approx3$, but that the $z\approx6$ population has a tail of sources with high rest frame equivalent widths. Possible explanations for this effect include a tendency towards stronger line emission in faint sources, which may arise from extreme youth or low metallicity in the Lyman-break population at high redshift, or possibly a top-heavy initial mass function.' author: - | Elizabeth R. Stanway$^{1}$[^1], Andrew J. Bunker$^{2,3}$, Karl Glazebrook$^{4}$, Roberto G. Abraham$^{5}$, James Rhoads$^{6}$, Sangeeta Malhotra$^{6}$, David Crampton$^{7}$, Matthew Colless$^8$, Kuenley Chiu,$^{3,4}$\ $^1$ Astronomy Department, University of Wisconsin-Madison, Sterling Hall, Madison, WI, 53726, USA\ $^2$ Institute of Astronomy, Madingley Road, Cambridge, CB3OHA, UK\ $^3$ School of Physics, University of Exeter, Stocker Road, Exeter, EX44QL, UK\ $^4$ Department of Physics and Astronomy, John Hopkins University, 3400 N Charles St, Baltimore, MD 21218, USA\ $^5$ Dept of Astronomy & Astrophysics, University of Toronto, 60 St. George St, Toronto, ON, M5S 3H8, Canada\ $^6$ Arizona State University, Department of Physics & Astronomy, Box 871504, Temp, AZ85287, USA\ $^7$ Dominion Astrophysical Observatory, 5071 W Saanich Rd, Victoria, V9E 2E7, Canada\ $^8$ Anglo-Australian Observatory, P.O.Box 296, Epping, NSW 1710, Australia date: 'Accepted . Received ; in original form ' title: 'The GLARE Survey II. Faint $z\approx6$ Ly-$\alpha$ Line Emitters in the HUDF' --- \[firstpage\] Introduction {#sec:intro} ============ The Hubble Ultra Deep Field [HUDF, @astroph0607632] opened a new window on the early universe. The exceptionally deep, multiwavelength data provided opportunities to study colour-selected samples of high redshift galaxy candidates with modest luminosities more typical of the general galaxy population. Previously, only the most luminous ‘tip of the iceberg’ had been accessible. By pushing to redshifts around 6 with the $z'$-band filter and the $i'$-drop Lyman break selection technique [e.g. @2004MNRAS.355..374B], the HUDF explored the end of the reionization epoch signaled by the Gunn-Peterson HI absorption trough in QSOs [@2001AJ....122.2850B]. Nonetheless, the use of the Lyman break colour selection criterion to isolate star-forming galaxies at a specific redshift, first developed to study $z\approx3$ galaxies [@1995AJ....110.2519S] and extended to $z\approx6$ candidates in analysis of the Great Observatories Origins Deep Survey (GOODS) and HUDF fields [e.g. @2003MNRAS.342..439S; @2006ApJ...653..53B], presents challenges. Before the properties of the population can be meaningfully discussed, the selection function itself must be understood. Estimates must be obtained of the contaminant fraction and the redshift distribution of $i'$-drop galaxies. Many colour-selected galaxies are significantly fainter than the conventional spectroscopic limit of today’s large telescopes. The HUDF reaches limits of $z'_{AB}=28.5$ (10$\sigma$) – a depth that, until [*JWST*]{} becomes available, requires unreasonable exposure times to obtain a high signal-to-noise ratio (S/N) spectrum for the continuum. However, line emission (e.g. Lyman-$\alpha$) may be within reach of ultra-deep moderate-dispersion spectroscopy on 8-10m telescopes even for the faintest galaxies, provided the equivalent width of the line is large enough. If the properties of the star-forming population at $z\approx6$ are similar to those of the Lyman Break Galaxy (LBG) population at $z\approx3$ [@2003ApJ...592..728S], then half the sources are expected to show the resonant Lyman-$\alpha$ transition in emission. Detection of a Lyman-$\alpha$ emission line fixes the redshift of a source, while detection or constraints on nearby high ionisation emission lines can quantify contribution by an AGN. Furthermore, spectroscopy can enable the identification of lower-redshift contaminants in the sample which will not emit Lyman-$\alpha$ photons at these wavelengths, but may show other emission lines. When a population of Lyman-$\alpha$ emitters is studied, the distribution of equivalent widths is sensitive to the stellar initial mass function (since the Lyman-$\alpha$ transition is excited by emission from hot, massive, short-lived stars), and also to the fraction of neutral gas in the intergalactic medium (IGM) and the emergent photon fraction [@2004ApJ...617L...5M]. In this paper we present results from the Gemini Lyman-Alpha at Reionisation Era (GLARE) project. This program used the 8-m Gemini South telescope to obtain extremely deep spectroscopy on a single slitmask, centered on the HUDF (Figure \[fig:layout\]). By obtaining extremely long exposures using a telescope with a large collecting area, we aimed to study continuum-selected galaxies fainter than those targeted by any other survey, and to quantify the line emission properties of our $i'$-drop sample in the HUDF. We presented initial results from this programme in @2004ApJ...604L..13S [ hereafter Paper I], which was based on 7.5hours of on-source exposure taken at low spectral resolution ($\lambda/\Delta\lambda\approx 500$). In this paper we present an analysis of the twenty-four $i'$-drop selected $5.6 < z < 7.0$ candidates targeted for 36hours of spectroscopy at higher resolving power ($\lambda/\Delta\lambda\approx 1200$). In section \[sec:glare\] we describe the GLARE program, and in section \[sec:emitters\] we present the results of our observing campaign. In section \[sec:EW\] we analyse the equivalent width distribution of the GLARE line emitters, and in section \[sec:disc\] we discuss the implications of this distribution in the context of other work in this field. Finally in section \[sec:conc\] we present our conclusions. We adopt the following cosmology: a flat Universe with $\Omega_{\Lambda}=0.7$, $\Omega_{M}=0.3$ and $H_{0}=70 h_{70} {\rm km\,s}^{-1}\,{\rm Mpc}^{-1}$. All magnitudes are quoted in the AB system (Oke & Gunn 1983). The GLARE Project {#sec:glare} ================= Candidate Selection {#sec:cands} ------------------- The objectives of the GLARE project were to determine the redshift and line emission properties of our targets, reaching low equivalent width limits on Lyman-$\alpha$ emission for a uniformly-selected sample of $i'$-drop Lyman-break galaxies [@2003MNRAS.342..439S; @2004ApJ...606L..25B; @2004MNRAS.355..374B]. The $i'$-drop colour selection is sensitive to galaxies at $5.6<z<7.0$, although the selection sensitivity falls off rapidly above $z\approx6.5$ (see Stanway et al. 2003). With this data we aim to probe galaxies at the end of the epoch of reionisation, and compare their properties to comparable galaxy samples at lower redshift, in order to explore possible evolution in the nature of the star-forming galaxy population. Initial results obtained from this survey were reported in @2004ApJ...604L..13S, which presented 7.5hour spectra of three Lyman-$\alpha$ line emitters observed on the 2003 GLARE slitmask. Selection of targets for our 2003 observations was made possible by the early release of a list of red sources in the HUDF, based on observations to one third of its final depth, and supplemented by $i'$-drop galaxies selected from the shallower, but wider field, GOODS survey of the same region (Stanway 2004). Between observations of our 2003 GLARE mask, and the start of semester 2004B, the full HUDF data were made public and the GLARE slitmask was redesigned accordingly. Sources with $(i'-z')_{AB}\ge1.3$ and $z'_{AB}\le28.5$ in the HUDF imaging [i.e. the catalog of @2004MNRAS.355..374B] were assigned the highest priority for spectroscopic follow-up and fifteen such sources were placed on the slitmask. The agreement is excellent between the HUDF $i'$-drop discovery catalog of @2004MNRAS.355..374B, which we use as our GLARE source list, and the recent data paper by the HUDF team [@astroph0607632]. All the robust $i'$-drops targetted on the GLARE slitmask were are also identified by @astroph0607632 In order to maximise use of the mask, additional objects were targetted. Seven slits were placed on candidates (including the two emitters identified in Paper I, GLARE 3001 & 3011) from a brighter selection with $(i'-z')_{AB}\ge1.3$ and $z'_{AB}\le27.5$, selected from the GOODS imaging (Stanway 2004), primarily at the edges of the slitmask and lying outside the HUDF region. A further three slits were placed on candidates with colours lying marginally outside of our GOODS selection criteria, either in colour (i.e. $1.0<i'-z'<1.3$) or magnitude (i.e. $z'>27.5$), or in noisy regions of the GOODS images, and two slits on sources from the original UDF early release list of red sources that did not meet our final criteria. Two slits were placed on $z\approx5$ $v$-band dropout candidates, and two on known low redshift \[OII\]3727Å emitters, previously observed by VLT/FORS2 [@2006astro.ph..1367V] and used as a check on flux calibration. Finally, five slits were placed on alignment stars to ensure accurate positioning of the mask, and six slits were used to conduct a blank sky survey for serendipitous sources. The composition of the 2004 GLARE slitmask is summarised in table \[tab:mask\]. ID No of Slits Description ---------- ------------- ------------------------------------- 1[XXX]{} 15 HUDF, $i'-z'\ge1.3$, $z'\le28.5$ 2 HUDF, other red sources 2[XXX]{} 5 Alignment Stars 3[XXX]{} 7 GOODS $i'-z'\ge1.3$, $z'\le27.5$ 4[XXX]{} 2 GOODS $v-i'$ selected, $z=5$ cands 5[XXX]{} 2 Known, low z, \[OII\] emitters 6[XXX]{} 6 Blanks sky slits 7[XXX]{} 3 GOODS marginal $i'$-drop candidates 42 Total Number of Slits on Mask : The composition of the 2004 GLARE slitmask. 1XXX indicates a GLARE identifier number in the range 1000-1999, etc.[]{data-label="tab:mask"} Both the $z'_{AB}=27.5$ cut applied in the GOODS data and the $z'_{AB}=28.5$ cut applied in the HUDF correspond to a signal to noise of approximately 8. We chose to work well above the detection limit in order to have confidence in the reality and nature of our candidate sources (many of which are detected only in this one band). In the event of non-detection in the $i'$-band we use the limiting magnitudes $i'_{AB}=29.15$ (GOODS) and $i'_{AB}=30.4$ (HUDF) corresponding to 3$\sigma$ variation in the sky background, as measured on the images. All sources were required to be undetected in the available $b$ (F435W) imaging but faint detections in the deep $v$ (F606W) filter (which lies above the Lyman limit at $z\approx6$, and which is present in several spectroscopically confirmed $z>5.6$ galaxies) were permitted. Completely unresolved sources were omitted from the selection although at faint magnitudes, the dividing line between unresolved and slightly-resolved sources becomes blurred. Of the line emitter candidates presented in section \[sec:emitters\], only GLARE 1042 and GLARE 1040 lie within the NICMOS HUDF field [@2005AJ....130....1T]. Both have near-infrared colours consistent with a high redshift interpretation, as indeed did all the $i'$-drop sources in this field discussed by @2005MNRAS.359.1184S. Observations and Data Analysis {#sec:data} ------------------------------ The 2004 GLARE slitmask was observed in semester 2004B using the GMOS spectrograph on Gemini-South (Hook et al. 2003). This mask was observed at higher spectral resolution than the 2003 GLARE mask, using the GMOS-S R600 grating rather than the R150, blazed at an angle of 48.5$\degr$ giving a central wavelength of $\sim8700$Å. As the CCD over-sampled the typical seeing and spectral resolution, the image was binned at $2\times2$ pixels so as to reduce the readout noise and improve the S/N. After this binning, the spectral scale was 0.94Å/pixel, and 0.146arcsec/pixel spatially on the detector. The slit width was 1.0arcsec, which produced a spectral resolution of 6.5Å FWHM for objects which fill the slit. Both mask and slits were oriented due north. The higher spectral resolution of the 2004 GLARE mask ($\lambda/\Delta\lambda\approx 1200$ compared with $500$ for 2003 GLARE) decreases the fraction of the wavelength range adversely affected by OH sky lines, and also enables us to better study the profiles of emission lines from the targets. The OG515 filter was used to suppress second order light from shorter wavelengths. In order to fill CCD chip gaps and ensure full wavelength coverage in the range $\lambda\approx7000-10000$Å  three different central wavelength settings were observed (8580Å, 8700Å and 8820Å). The shortest wavelength reached was 6420Å, while the longest wavelenth surveyed was 10950Å. Wavelengths were calibrated from the night sky lines in each slit, leading to a solution with an [*rms*]{} error of approximately 0.3Å. Fluxes were calibrated from the broadband magnitudes of the alignment stars on the mask, and checked against both line emission of known \[OII\] emitters also observed in VLT/FORS2 spectroscopy that had been placed on our mask for verification purposes, and existing spectroscopy for the known $z=5.83$ Lyman-$\alpha$ emitter GLARE 1042 [@2004ApJ...604L..13S; @2004ApJ...607..704S; @2004ApJ...600L..99D]. We estimate a 20% error on the flux calibration, associated with slit losses and centroiding uncertainty in the narrow slits. To optimise the subtraction of night sky emission lines (which occupy a large fraction of the spectrum at $>8000$Å) we used the instrument in ‘nod & shuffle’ mode [@gla01; @2004astro.ph..2436A]. Each exposure was 30 minutes long, nodding every 60 seconds. Hence we are able to suppress sky emission that varies on timescales longer than one minute. The total exposure time on this mask was 36 hours. The reduction of nod & shuffle data involves the subtraction of positive and negative spectra, observed in alternate 60 second exposures and offset spatially by the telescope nodding. We employed slits 2.47 arcseconds in length, nodding by 1.25 arcseconds between sub-exposures. Our queue scheduled observations were constrained by the requirement that the seeing was less than 0.5 arcseconds FWHM, and the nod distance set such that the signals were separated by at least twice the seeing disk. Hence the characteristic signal of an emission line comprises a ‘dipole’ signal of positive and negative emission at the same wavelength, spatially offset by 1.25 arcseconds. In visually identifying candidate line emitters, we looked for this dipole signature with positive and negative channels of comparable strengths; this requirement effectively increases the sensitivity of the spectroscopy beyond the formal limit, since random background fluctuations are unlikely to produce simultaneous signals in both positive and negative channels. Dipole signals lying under strong emission sky lines are treated with caution; as well as having more Poisson noise, sky line variability and charge diffusion at the red end of the spectrum may lead to a spurious signal. A number of charge traps and CCD artifacts were also masked when the individual exposures were combined. These charge traps affect some regions of the CCD more severely than others, and so the noise properties vary from slitlet to slitlet, and also with wavelength. However, the 1 sigma [*rms*]{} pixel-to-pixel variation in the background at 8500Å was $1.4\times10^{-19}$ergcm$^{-2}$s$^{-1}$Å$^{-1}$ for the $2\times 2$ binned pixels, measured between skylines on the nod & shuffle background-subtracted spectra. Hence, our sensitivity to an emission line extending over 500kms$^{-1}\times 1$arcsec is $2.5\times 10^{-18}$ergcm$^{-2}$s$^{-1}$ ($3\,\sigma$ combining both nod positions, or $2\,\sigma$ per nod channel). Line Emitters in the 2004 GLARE Mask {#sec:emitters} ==================================== In the 36 hour exposure of the 2004 GLARE mask, we identify five strong Lyman-$\alpha$ emission line sources [including the three candidates tentatively proposed in @2004ApJ...604L..13S]. We identify a further four sources which have lower significance detections, but which are considered possibly to be real emission lines due to their dipole natures and separation from possible sky line residuals. Finally we identify five sources with tentative emission line detections that are considered unlikely to be real. Strong Emission Lines {#sec:good} --------------------- Table \[tab:strong\] lists the properties of the five $z\approx6$ sources with strong Lyman-$\alpha$ line emission. Figure \[fig:strong\] presents the two dimensional spectra obtained for these sources, and the summed flux from positive and negative spectral channels. We also placed two known lower-redshift galaxies on the 2004 GLARE mask, placing \[OII\] within our spectral range, as a check on our sensitivity and calibration. These are galaxies GDS J033235.79-274734.7 ($z=1.223$) and GDS J033229.63-274511.3 ($z=1.033$) from the ESO VLT/FORS2 survey of . We detect the \[OII\] emission at $\lambda=8285$Å&$7577$Å, with line fluxes $1.53$&$1.52\,\times 10^{-17}\,{\rm erg\,cm^{-2}\,s^{-1}}$, respectively. For the Lyman-$\alpha$ detections, GLARE sources 1042 and 3001 were previously identified as line emitters in Paper I from the 2003 GLARE mask, @2004ApJ...604L..13S. GLARE 3011 was tentatively identified as a line emitter, and this identification is now strongly confirmed. GLARE 1054 and 1008 are presented for the first time here. Since its discovery [SBM03\#1 in @2003MNRAS.342..439S] GLARE 1042 at $z=5.83$ has been spectroscopically confirmed [@2004ApJ...607..704S; @2004ApJ...604L..13S; @2004ApJ...600L..99D – SiD002] and extensively studied, including in the infrared with Spitzer [@2005MNRAS.364..443E; @2006astro.ph..4554Y – \#1ab]. GLARE 1054 was identified from the $i'$-drop selection in the HUDF, GLARE 3001 and 3011 from the somewhat brighter GOODS selection. GLARE 1008 is a source selected from the initial early data release list of red sources in the HUDF survey. It lies outside the final overlap region of the HUDF which has been used for catalogue construction and analysis, but within the noisy outer regions of the HUDF mosaic. It is technically below the detection limit of the GOODS survey, although it is faintly detected in the GOODS $z'$-band. It is faintly detected in the $z'$-band in the shallower edges of the HUDF. Given the higher resolution and more sensitive flux limit of this spectroscopy, we can rule out a low redshift interpretation for these sources and confirm them as $z\approx6$ Lyman-$\alpha$ emitters. The spectral resolution of the GMOS configuration used in our 2004 GLARE mask is sufficient to resolve the \[OII\] $\lambda_\mathrm{rest}=3727,3729$Å doublet (and does so in the case of the two known \[OII\] emitters on our 2004 GLARE mask), and we would expect to identify weaker emission lines such as \[NII\], \[SII\] within our observed redshift range if the detected line was H$\alpha$ ($\lambda_\mathrm{rest}$=6563Å). The other strong optical emission lines (H$\beta$4861, \[OIII\]4959,5007) would always yield other strong lines within our spectral range. Sources at these low redshifts are also unlikely to satisfy the colour cut criterion used for target selection. All five detected strong Lyman-$\alpha$ emission lines also show significant asymmetry as expected for high redshift emitters (in which the blue wing of the line is significantly self absorbed, and the red wing broadened by re-emission). This phenomenon is well known at $z\approx3$ and is believed to arise in outflows from the actively star-forming galaxies [@2003ApJ...584...45A]. Similar asymmetry has now been observed in all $z>5.5$ galaxies confirmed to date by spectroscopy that resolves the Lyman-$\alpha$ emission line [e.g. @2003MNRAS.342L..47B], suggesting that similar outflows are produced at higher redshift galaxies. -- -- -- -- -- -- ------------- ------------------------- ----------- ------ ---------------- --------------- ----------------------- ------------------------------ ---------------------- ID RA & Declination Alternate z $z'_{AB}$ $i'-z'$ Line Flux W$_{Ly\alpha}^\mathrm{rest}$ $L_{Ly\alpha}$ J(2000) ID ergcm$^{-2}$s$^{-1}$ Å $10^{42}$ergs$^{-1}$ 1042$^\dag$ 03 32 40.01 -27 48 14.9 20104 5.83 25.35$\pm$0.02 1.64$\pm$0.04 1.58 $\times10^{-17}$ 22.9 5.90 1054 03 32 33.20 -27 46 43.3 42414 5.93 27.65$\pm$0.07 1.45$\pm$0.17 0.68 $\times10^{-17}$ 120 2.63 1008 03 32 47.97 -27 47 05.1 6.13 28.51$\pm$0.18 $>$1.35 0.43 $\times10^{-17}$ 159 1.80 3001 03 32 46.04 -27 49 29.7 5.79 26.69$\pm$0.06 1.67$\pm$0.20 0.77 $\times10^{-17}$ 44.1 2.83 3011 03 32 43.18 -27 45 17.6 5.93 27.47$\pm$0.12 1.86$\pm$0.50 1.13 $\times10^{-17}$ 242 4.39 ------------- ------------------------- ----------- ------ ---------------- --------------- ----------------------- ------------------------------ ---------------------- $^\dag$ GLARE 1042 has a panoply of alternate names. It is SBM03\#3 in @2003MNRAS.342..439S, 20104 in [@2004MNRAS.355..374B] and is SiD0002, the $z=5.83$ line emitter of @2004ApJ...600L..99D. Possible Emission Lines {#sec:maybe} ----------------------- We identify a further four sources (listed in table \[tab:possible\]) as possible line emitters. In most cases, the detection in each channel is of low significance, but the coincidence of positive and negative signals suggests that the emission lines are real. Alternately some candidate lines may lie on top of skylines, leading to the suspicion that these represent skyline residuals. These sources are illustrated in figure \[fig:possible\]. All but one of the candidate emission lines lie shortward of the lower redshift limit selected by the $i'$-drop technique. Sources at these redshifts would be expected in a $v$-drop rather than $i'$-drop selection. It is possible for sources with large errors on their $i'-z'$ colour to scatter into the $i'$-drop selection, although it is unlikely that photometric scatter alone could explain this discrepancy. Three of these sources were identified in the HUDF $i'$-drop sample, one (GLARE 3000) from the GOODS sample. GLARE 3000 was first identified as an $i'$-drop source in @2003MNRAS.342..439S [, SBM03\#05] but considered likely to be an M or L class Galactic dwarf star, as it is unresolved in [*HST*]{} imaging. The candidate emission line in this source lies beside a strong sky line residual. FORS2 spectroscopy of this source by @2004astro.ph..6591V also interpreted the spectrum obtained as that of a Galactic star. The stellar identification is almost certainly correct, given the broadband colours and unresolved half light radius of this source, suggesting that possible emission lines at this significance should be considered highly suspect. There are two possible emission lines of similar strength in GLARE 1067 (at 7037Å & 7099Å), rendering it unlikely that this source lies at high redshift. The observed line separation may be consistent with \[OIII\] emission ($\lambda_\mathrm{rest}=4959,5007$Å) at $z=0.418$, although a galaxy at this redshift is predicted to have colours that are significantly bluer (i.e. $i'-z'<0.5$ and detected in the $b$ band, as opposed to the observed $i'-z'=1.4\pm0.2$ and no $b$ detection). In addition, the redward line of this \[OIII\] doublet is expected to be three times stronger than the blueward line, while the observed emission peaks are of comparable strength. An alternate explanation might be that two galaxies, separated by $3000$kms$^{-1}$ lie within the slit. While this source has a close neighbour (at $\alpha_\mathrm{J2000}=$03:32:35.8, $\delta_\mathrm{J2000}=$-27:48:49, with a separation of $<1$ arcsecond), the neighbouring galaxy is blue in colour ($i'-z'=0.2$), and likely lies at significantly lower redshift. Therefore the most probable interpretation of these lines (if they are real) is that they represent \[OIII\] emission arising not in the targeted galaxy but rather in the neighbouring source. GLARE 1040 and GLARE 1086 are the most convincing candidates in this category. GLARE 1040, is an isolated source that clearly drops between the $i'$ and $z'$-bands, and is undetected in the HUDF $v$-band. The candidate emission line lies at $z=5.2$, just shortward of the $i'$-drop selection, although it is possible for a faint galaxy such as this to scatter upwards into the selection window. The candidate line in the final source, GLARE 1086, lies firmly in the redshift range selected by the $i'$-drop technique, and may well be a high redshift emission line, although the detection is too weak to rule out a low redshift \[OII\] explanation, or to detect line asymmetry. -- -- -- -- -- -- --------------- ------------------------- ----------- ------------------- ---------------- --------------- ---------------------- ------------------------------ ---------------------- Slit RA & Declination Alternate $z$ $z'_{AB}$ $i'-z'$ Line Flux W$_{Ly\alpha}^\mathrm{rest}$ $L_{Ly\alpha}$ J(2000) ID ergcm$^{-2}$s$^{-1}$ Å $10^{42}$ergs$^{-1}$ 1040 03 32 38.28 -27 47 51.3 24458 5.21? 27.51$\pm$0.07 1.60$\pm$0.17 0.17$\times10^{-17}$ 20.2 0.49 1067 03 32 35.83 -27 48 48.9 14210 4.84?$^{\dagger}$ 28.08$\pm$0.10 1.43$\pm$0.24 0.30$\times10^{-17}$ 81.5 0.72 4.78?$^{\dagger}$ 0.31$\times10^{-17}$ 88.0 0.73 1086 03 32 30.14 -27 47 28.4 30359 6.10? 28.13$\pm$0.11 1.46$\pm$0.25 0.37$\times10^{-17}$ 68.1 1.53 3000$^{\ast}$ 03 32 38.80 -27 49 53.7 5.08? 25.65$\pm$0.03 $>$3.50 0.11$\times10^{-17}$ 2.24 0.30 --------------- ------------------------- ----------- ------------------- ---------------- --------------- ---------------------- ------------------------------ ---------------------- $^{\dagger}$ GLARE 1067 has two emission lines, which are consistent with \[OIII\] at $z=0.418$; if these are instead Lyman-$\alpha$, both are at $z<5$.\ $^{\ast}$ GLARE 3000 is most likely a low-mass Galactic star. Finally, we identify signals in a further five sources as ‘unlikely’ emission lines. These are shown in table \[tab:not\] and figure \[fig:not\]. Although the dipole signal from the first two sources appears strong, they lie on bright skylines and so may be partly or wholly due to sky subtraction residuals. They may also arrise as a result of charge diffusion from the adjacent slit. Both candidates lie at $\lambda>10000$Å (not far from the 10500Å Si bandgap) where the diffusion is strong. The signal in the remaining candidates is weak. Slit GLARE 6050 was a blank sky slit, and there is nothing visible in any waveband at the depth of the GOODS imaging of this region. The possible serendipitous line emitter is offset from the centre of the slit by approximately 02 to the north. -- -- -- -- -- -- --------------- --------------------------- ----------- ------- ---------------- --------------- ------------------------------ Slit RA & Declination Alternate z $z'_{AB}$ $i'-z'$ W$_{Ly\alpha}^\mathrm{rest}$ J(2000) ID Å 1034 03 32 39.454 -27 45 43.42 52086 7.24? 27.97$\pm$0.09 $>$2.43 10.6 1030 03 32 41.184 -27 49 14.81 10188 7.46? 27.10$\pm$0.05 2.04$\pm$0.16 4.4 3015 03 32 27.910 -27 49 41.98 5.67? 27.30$\pm$0.10 $>$1.85 16.7 6050$^{\ast}$ 03 32 37.450 -27 49 47.60 7.05? - - $>$104 8005 03 32 34.392 -27 45 33.01 4.69? 27.11$\pm$0.16 1.15$\pm$0.36 66.6 --------------- --------------------------- ----------- ------- ---------------- --------------- ------------------------------ $^{\ast}$ 6050 is a ‘blank sky’ slitlet Sources with No Line Emission {#sec:limits} ----------------------------- The remaining 21 science targets on the slitmask showed no evidence for line emission in the wavelength range observed, $\lambda_\mathrm{obs}\approx7800-10000$Å (corresponding to $5.4<z<7.2$, with some slit to slit variation depending on slit location). Of these, 11 formed part of our high priority $i'$-drop selection, with marginal targets (5 sources) and sky slits (5 slits) constituting the remainder. We discuss the implications of these non-detections of line emission below. It is possible that, despite optimising our experimental setup for sky line subtraction, we are missing flux from emission lines that fall in regions dense with skylines. The noise in such areas is greater than the slit average, making identification of line candidates more difficult. Some 35% of the wavelength coverage of this spectroscopy lies under sky lines (defined as regions where sky emission exceeds 150% of the inter-line sky continuum level) which leave Poisson noise and sky brightness fluctuation residuals of varying strength, although only 10% lies under strong skylines (sky line emission $>5$ times sky continuum). The noise under the residuals of bright sky lines is a factor of 2.5 greater than that between lines, and might also be prone to systematics in the sky subtraction leading to potentially spurious emission lines. Only half of Lyman Break Galaxies at $z\approx3$ show Lyman-$\alpha$ in emission [@2003ApJ...588...65S], and one quarter have W$_{Ly\alpha}^\mathrm{rest}>$20Å. Given that we reach a limiting rest-frame equivalent width of $\le$6Å for the majority of our targets, we might naïvely expect to observe emission in half our slits, less ten percent lost to strong sky line residuals. However, the fraction of Lyman Break Galaxies with emission at $z\approx6$ is still unknown, and it is likely that a large fraction of $z\approx6$ $i'$-drop galaxies will not be spectroscopically confirmed until the continuum level is reached with either longer exposures or more sensitive telescopes, and interstellar absorption lines can be used for redshift determination [see, for example, @2005ApJ...630L.137D for rest-frame UV continuum spectroscopy of a bright $z=5.5$ galaxy]. Finally in the case of the marginal candidates used to fill the slitmask, it is possible that the true redshift of the sources lies above $z=4.5$ (set by the dual requirements of non-detection in the $b$ band, and the $i'-z'$ colours of these sources), but below the limit of our spectroscopy. Two of the slits without detections were placed on $v$-drop galaxies expected to lie in this range, while a further four slits were placed on sources slightly too blue to meet our strict selection criteria. It seems unlikely that photometric scatter could place these sources in our detectable redshift range. We will discuss the equivalent width limits on our $i'$-drop galaxies in Section \[sec:EW\]. Agreement with Other Spectroscopic Surveys {#sec:others} ------------------------------------------ A subset of sources in our sample have also been observed spectroscopically as part of the concerted campaign of follow up observations to the GOODS survey. Sources in the GOODS-S field have been targeted for 8m spectroscopy by FORS2 on the VLT (Vanzella et al, 2006), by DEIMOS on Keck [@2004ApJ...607..704S, Bunker et al 2006] and by GMOS on Gemini (Paper I and this work). In addition, this field was surveyed with the [*HST*]{}/ACS grism as part of the GRAPES survey [@2005apj_malhotra]. This slitless spectroscopy has been obtained to varying depths, and at different spectral resolutions. In particular, the GRAPES grism spectroscopy is of too low a resolution to detect all but the highest equivalent width Lyman-$\alpha$ emission lines adjacent to the Lyman-$\alpha$ forest continuum break (compare, for example, the DEIMOS spectrum of SBM03\#1/GLARE1042 in Stanway et al. 2004 showing the strong Lyman-$\alpha$ line, with the [*HST*]{}/ACS spectrum of the same object, SiD002, in fig. 2a of Dickinson et al. 2004, which shows a continuum break but the line is washed out). However, GRAPES does provide a redshift estimate by localising the wavelength of the Lyman-$\alpha$ break in the spectrum. Where redshifts have been determined by multiple groups, our results are generally in reasonable agreement with previous observations, given the difficulty in obtaining precise redshifts from the self-absorbed and resonantly-broadened Lyman-$\alpha$ lines at moderate spectral resolution. In the case of GLARE 1042, our measured redshift of $z=5.83$ is in close agreement with that measured by FORS2 [$z=5.82$, @2004astro.ph..6591V], by the GRAPES team [$z=5.8$, @2005apj_malhotra] and, using Keck/DEIMOS spectra, by @2004ApJ...607..704S [ $z=5.83$]. GLARE 3001 was spectroscopically identified by FORS2 as a line emitter at $z=5.78$, agreeing with our redshift of $z=5.79$. GLARE 1054 and 1030 were both placed at $z=5.7$ by GRAPES spectroscopy [@2005apj_malhotra]. This contrasts with our measured redshifts of $z=5.9$ for GLARE 1054 and (tentatively) $z=7.4$ for GLARE 1030. The discrepancy for GLARE 1054 is within the expected level of agreement for GRAPES grism spectroscopy and so these results are consistent. Given the low significance of our line candidate in GLARE 1030, and the $i'$-drop redshift selection function, the GRAPES redshift remains the more likely. The remaining source, GLARE 3000 was identified by the VLT/FORS2 observations as a Galactic star on the basis of weakly detected \[OI\] and \[NII\] lines. This source, which corresponds to the unresolved candidate SBM03\#5 in @2003MNRAS.342..439S, falls in our ‘possible’ category. We note that the FORS II spectrum is flagged with their quality class ‘C’ and that there is no secure continuum detection. Despite this, a Galactic star is still the most likely identification of this target, illustrating the caution with which we present our fainter line candidates. No other target on our 2004 GLARE mask has been reported as a line emitter by other teams, or has an estimated redshift from GRAPES spectroscopy. Limits on NV and other Emission Lines {#sec:other_lines} ------------------------------------- Lyman break galaxies at $z\approx3$ show few other emission lines in the rest frame ultraviolet. The composite spectrum of $\approx$1000 such galaxies produced by @2003ApJ...588...65S shows weak emission features due to SiII\* ($\lambda_\mathrm{rest}=$1265Å, 1309Å, 1533Å), OIII\] ($\lambda=$ 1661Å, 1666Å) and CIII\] ($\lambda=$1908Å), and absorption features due to stellar winds, primarily SiII and CIV. The presence of an AGN in our target galaxies could also lead to the presence of emission lines due to high excitation states, primarily NV ($\lambda_\mathrm{rest}=$1238.8Å, 1242.8Å) and SiIV ($\lambda=$1394Å, 1403Å), and rarely OV\] ($\lambda=$1218Å). For a galaxy at $z=6$, our spectra extend to rest frame wavelengths of $\approx$1428Å. Given that we are not able to measure a high signal to noise continuum on any one target, we are unable to measure absorption features in the spectra, and therefore confine ourselves to searching for evidence of other emission features in the spectra of our Lyman-$\alpha$ emitters. A careful inspection of our five good Lyman-$\alpha$ emission line candidates does not provide evidence for any other emission lines at the redshift of Lyman-$\alpha$. While this is not surprising given the weakness of secondary lines, the failure to detect NV in this spectrscopy suggests a large Ly-$\alpha$ / NV ratio. Using our 3$\sigma$ limit on undetected lines as an upper constraint on NV flux we determine that $f(\mathrm{Ly}\alpha)/f(\mathrm{NV}) > [10.5, 4.5, 2.8, 5.1, 7.5] (3\sigma)$ respectively for GLARE targets \[1042, 1054, 1008, 3001, 3011\] Typical line ratios for AGN are $f(\mathrm{Ly}\alpha)/f(\mathrm{NV})=4.0$ (Osterbrock 1989), while those for galaxies are greater than this. The limits we determine disfavour an AGN origin to the Lyman-$\alpha$ flux in all but the faintest target (in which the constraint is too weak to make a firm statement). These limits constrain AGN activity, if present, to only a weak contribution to the rest frame ultraviolet flux. The Equivalent Width Distribution of GLARE line Emitters {#sec:EW} ======================================================== The Observed Equivalent Width Distribution {#sec:EWobs} ------------------------------------------ Using the mean variance in the background of the exposed slits, and the broadband magnitudes of the targeted galaxies, we are able to calculate limits on the rest frame equivalent width $W_0$ for those sources which satisfy our colour selection criteria and yet are undetected in our spectroscopy. These are presented in table \[tab:ewlims\]. In each case we assume that the galaxy lies at the mean $i'$-drop redshift ($z=6.0$) and that the emission line has not been lost behind a the subtraction residual of a bright night sky line. Combining our line limits with those sources for which emission lines have been identified or tentatively proposed forms a sample of 24 sources uniformly selected from their $i'$-drop colours. The resultant distribution of Lyman-$\alpha$ equivalent widths is plotted in figure \[fig:ewlims\]. If all the sources for which candidate emission lines are identified in this paper prove to be $z\approx6$ Lyman-$\alpha$ emitters, then the escape fraction of Lyman-$\alpha$ photons at $z\approx6$ appears qualitatively similar to that at $z\approx3$. From our $z\approx 6$ sample, 66% of the sources (16 out of 24) have Lyman-$\alpha$ equivalent widths $<25$Å, compared with 75% at $z\approx3$ [@2003ApJ...588...65S], although the lower redshift sample loses a smaller fraction due to skyline contamination. The samples probe to similar points on the luminosity function in both cases (approximately 0.1L$^\ast$). These fractions are consistent within the errors on our small number statistics. Harder to explain in comparison with lower redshift galaxies is the tail stretching to very high equivalent widths ($>200$Å) observed in this survey, a trait also observed in some narrowband selected sources at this redshift [e.g. @2002ApJ...565L..71M] and in other $i$-dropout Lyman break galaxies [e.g. @2006astro.ph.12454D who find one source with $W_0=$150Å]. Although the number statistics are small, we observe four line emitters ($17\pm 8$%) with equivalent widths in the range $50-100$Å, and a further four with $W_0>$100Å (three of which come from our ‘robust’ list of line emitters). This contrasts with the Lyman Break Galaxy population at $z\approx3$ in which $<5$% of galaxies have line emission with $W_0>100$Å [@2003ApJ...588...65S]. While high redshift galaxies at both redshifts are selected on their rest-frame ultraviolet continuum and the spectral break caused by Lyman-$\alpha$ absorption, the two populations are not identical. The sample discussed here reaches some two magnitudes fainter than the tentatively proposed and still uncertain typical luminosity L$^\ast$ at $z\approx6$ [@2006ApJ...653..53B; @2004MNRAS.355..374B]. This contrasts with a sample reaching just one magnitude below L$^\ast$ at $z=3$. @2003ApJ...588...65S considered subsamples at $z=3$, dividing their spectroscopic data into quartiles based on rest-frame equivalent width. They found a modest trend in the strength of line emission with magnitude. Galaxies in their highest equivalent width quartile are some 0.2 magnitudes fainter on average than their quartile of weak to moderate line emission. It is possible that the strong line emission observed here is more typical of sources with faint rest-frame ultraviolet continuum suggesting that faint sources differ physically from brighter members of the population. A second difference between the samples is intrinsic rather than arising from a selection effect. Several authors [e.g. @2005MNRAS.359.1184S; @2006ApJ...653..53B; @2003ApJ...593..630L] have now observed that Lyman break galaxies at $z>5$ have steeper rest-frame ultraviolet slopes than those at $z\approx3$. @2006ApJ...653..53B interpret this as indicating that the dust properties of this population evolve over redshift. While a steep rest-ultraviolet slope can also arise due to low metallicity or a top-heavy initial mass function (as discussed below) dust evolution is a natural interpretation. At $z\approx6$ the universe is less than 1 gigayear old and galaxies may not have had time to develop a high dust content. @2003ApJ...588...65S found that $z\approx3$ galaxies with high equivalent widths in Lyman-$\alpha$ also had lower mean dust extinction. Lyman-$\alpha$ photons are resonantly scattered by dust and hence the line is preferentially absorbed with respect to the rest-frame ultraviolet continuum. If the distribution of Lyman-$\alpha$ equivalent widths in the $z\approx3$ population is truncated by dust absorption, this could produce an apparent ‘excess’ of strong lines at high redshift. However, even zero dust absorption cannot explain equivalent widths exceeding 200Å unless the sources are also very young and very low in metallicity. The full explanation for the equivalent width distribution observed in the GLARE data may well be a combination of these effects and those discussed below. Several other possible explanations exist for both the steepening of the rest frame UV slope and the difference in equivalent width distributions. An interpretation of contaminant galaxies at lower redshifts seems unlikely due to our $i'$-drop selection criteria; low redshift sources with strong spectral breaks are likely to have more than one emission line in our observed redshift range. A high equivalent width line can arise if the observed Lyman-$\alpha$ photons excited by a population of AGN rather than by young, hot stars. The luminosity function of AGN is poorly constrained at these magnitudes and redshifts, but the space density of such sources is expected to be very low (e.g., the $z>6$ SDSS QSOs, Fan et al. 2003). At $z\approx6$, the deep 2Ms X-ray exposure of the UDF and surrounding region [@2003AJ....126..539A] would detect only the brightest AGN ($L>10^{42}$ergss$^{-1}$cm$^{-2}$). None of the GLARE targets are detected in this X-ray observation. AGN would also be expected to show emission lines that are not just strong but also broad, while none of the GLARE line candidates are broad. There is also an absence of high-ionization lines such as NV1240Å which are common in AGN. The tail of line emitters extending to higher equivalent widths may also arise from a difference in the fundamental properties of the stellar population between $z\approx6$ and that at lower redshifts. Modeling of emission from metal-free Population III galaxies predicts rest frame Lyman-$\alpha$ equivalent widths $>$1000Å for young starbursts ($<$2Myr), decreasing to W(Ly-$\alpha$)$\sim$500Å for older bursts . These very high equivalent widths arise from the relatively hard spectrum of metal-free reactions in the most massive stars. However it is unlikely that zero metallicity (population III) stars are still contributing significantly to the emission from massive stars almost a billion years after the Big Bang, particularly given the identification of stellar populations $>100$Myr old in some $z>5$ sources [e.g. @2005ApJ...618L...5E; @2005MNRAS.364..443E]. Further evidence for moderate metallicity at $z>5$ has been observed in the spectroscopy of bright AGN from the Sloan Digital Sky Survey. Metals including iron [@2003ApJ...594L..95B] and carbon have been detected from even the highest redshift quasar (at $z=6.4$). By contrast, even 1/20th solar metallicity leads to a sharp reduction in the peak (zero age) equivalent width predicted to $\sim$300Å, with a more typical W(Ly-$\alpha$)$\sim$100Å by an age of 10-100Myr [@2002ApJ...565L..71M]. Most of the candidate emission lines presented in this study can hence be explained with normal, if low metallicity, populations. However at least two of our good line emission candidates have rest frame equivalent widths exceeding 200Å. This is possible if the galaxy is in the first few Myrs of an ongoing starburst, but may also provide evidence for variation in the initial mass function of star formation. High equivalent widths of Lyman-$\alpha$ emission can arise from a “top-heavy” initial mass function (i.e. star formation weighted towards a population of high mass stars). @2004ApJ...617L...5M calculated the Lyman-$\alpha$ equivalent width expected for a metal-enriched population with a very extreme IMF, weighted towards massive stars (i.e. IMF slope $\alpha=0.5$). As is the case for low metallicity populations, the flux is weighted towards a harder spectrum, and Lyman-$\alpha$ emission is strengthened. They found that such an IMF could explain line equivalent widths of up to 240Å at ages of a few Myr, with higher IMFs possible for very young bursts. The hard rest-frame ultraviolet spectrum associated with such an IMF may also be consistent with the steep rest frame ultraviolet spectral slopes observed in $z\approx6$ galaxies [e.g. @2005MNRAS.359.1184S]. While the evidence from the GLARE study is limited, with the number of high equivalent width sources small, the existence of such sources suggests that the environment of star formation at $z\approx6$ is less metal enhanced, or weighted towards more massive stars than that at $z\approx3$. ID Alt RA & Declination $z'_{AB}$ $i'-z'$ W$_{Ly\alpha}^\mathrm{rest}$/Å ------ ------- ------------------------- ---------------- --------------------- -------------------------------- 1001 48989 03 32 41.43 -27 46 01.2 28.26$\pm$0.12 $>$2.1 (2$\sigma$) $<$6.5 1004 46223 03 32 39.86 -27 46 19.1 28.03$\pm$0.10 $>$2.37 (2$\sigma$) $<$5.2 1009 12988 03 32 38.50 -27 48 57.8 28.11$\pm$0.11 $>$2.29 (2$\sigma$) $<$5.6 1045 21111 03 32 42.60 -27 48 08.8 28.02$\pm$0.11 1.67$\pm$0.26 $<$5.2 1047 35271 03 32 38.79 -27 47 10.8 28.44$\pm$0.14 1.33$\pm$0.30 $<$7.6 1060 11370 03 32 40.06 -27 49 07.5 28.13$\pm$0.11 $>$2.28 (2$\sigma$) $<$5.7 1077 16258 03 32 36.44 -27 48 34.2 27.64$\pm$0.07 1.42$\pm$0.16 $<$3.7 3002 03 32 43.35 -27 49 20.4 26.89$\pm$0.07 1.42$\pm$0.20 $<$1.8 3030 03 32 48.94 -27 46 51.4 27.04$\pm$0.08 1.41$\pm$0.23 $<$2.1 3033 03 32 49.08 -27 46 27.7 27.18$\pm$0.09 1.40$\pm$0.27 $<$2.4 Implications of High Equivalent Widths {#sec:EWimp} -------------------------------------- The presence of a tail of large equivalent width emission line sources has implications for the redshift distribution expected from an $i'$-drop selection. The redshift distribution and number densities of $i'$-drop galaxies have usually been calculated assuming a negligible contribution from line emission to the measured magnitudes and colours. In the case of line emitters with rest frame equivalent widths $W_0<30$Å this is a reasonable assumption, with the contribution to $z'$ flux amounting to a few percent. For line emitters with larger equivalent widths there can be a significant effect on the selection function of $i'$-drop galaxies. This effect falls into three main regimes as illustrated by figure \[fig:ewsel\]. If the emission line falls in the $i'$-band filter (i.e. $z<6.0$), then the $i'-z'$ colour of the galaxy is reduced by the line emission, and sources with high equivalent widths fall out of the colour-magnitude selection window. A source at our $z'$ detection limit with a rest frame Lyman-$\alpha$ equivalent width of 100Åcan be as blue as $i'-z'=0.67$ (AB) and an intrinsic line width of 150Åwould lead to a colour of just $i'-z'=0.46$ (AB). These colours are similar to those of much lower redshift galaxies. Ensuring a complete selection of line emitting galaxies at $5.6<z<6.0$ is therefore impossible using a simple colour-selected sample without also including a great many lower redshift contaminant sources. If the line emission falls in $z'$ filter, the $i'-z'$ colour is enhanced and galaxies with continuum flux below the $z'$ limit are promoted into the selection window. Hence at $z>6.5$ the $i'$-drop criterion can select a population of low continuum, strong line emission sources rather than the continuum break sources it targets. There is also a redshift range in which the line emission would fall in the overlap region between filters and both effects compete. This effect may lead to a bimodal redshift distribution for $i'$-drop galaxies, with weak emission line sources preferentially selected towards lower redshift, and high equivalent width sources selected at higher redshifts. It is necessary to quantify the equivalent width distribution, combining galaxies at the faint limits explored here with brighter galaxies, in order to tightly constrain the galaxy luminosity function at any given redshift. As figure \[fig:ewsel\] illustrates, our good line emission candidates all lie in regions of parameter space affected by equivalent width selection biases. Our candidates lie in parameter space theoretically allowed for their equivalent width with two exceptions - GLARE 1054 and 3011 both have strong emission lines in the tail of the $i'$ filter that should have forced their $i'-z'$ colours out of our selection function. We note, however, that that both of these sources have $i'-z'$ colours within 1$\sigma$ of our minimum selection cutoff, and may have scattered up into out selection. Only one source is boosted into the selection function by virtue of its line emission rather than continuum flux, suggesting that our sample is, as expected, dominated by continuum-selected sources. If we are indeed missing sources with high emission line equivalent widths at the low redshift end of our survey (where our redshift selection function is at its peak), then our conclusion that the distribution of line emission at high redshift extends to larger equivalent widths is strengthened. This effect has interesting implications for the luminosity functions presented in the literature for $z\approx6$ galaxies. If the $i'$-drop criteria are overestimating the number density of faint continuum sources (due to contamination by a tail in the distribution of strong line emitters) then it is possible that the faint end slope of the luminosity function is shallower than hitherto reported. If there is a large number of strong emitters at $5.6<z<6.0$ that have fallen out of the colour selection, then the number density of sources at these redshifts could be underreported. This may contribute towards the discrepancy between the observed evolution in the LBG luminosity function between $z=4$ and $z\approx6$ [e.g. @2004MNRAS.355..374B; @2004astro.ph..6562B] and the reported lack of evolution in the Lyman-$\alpha$ emitter luminosity function over the same redshift range [e.g. @2004astro.ph..7408M]. @2006astro.ph..5289A studied the equivalent width distribution of a heterogeneous sample of $i'$-drop and narrowband selected galaxies with spectroscopically confirmed Lyman-$\alpha$ emission at $z\approx6$. They commented on an apparent dearth of UV-luminous galaxies with strong Lyman-$\alpha$ emission lines, noting an increase in the fraction of strong line emitters at $M_{AB}=-21$. In figure \[fig:absmag\] we show the equivalent distribution of the GLARE line emitters as a function of absolute magnitude. We detect five bright sources with a line luminosity brighter than $2\times10^{43}$ergss$^{-1}$, thus occupying a regime unoccupied by the Ando et al sample. Of these, two possible emission line sources have equivalent widths greater than any in the previous sample. The four sources with a calculated line luminosity between $5\times10^{43}$ and $10^{44}$ergss$^{-1}$ are uniformly distributed in absolute magnitude. Although the number statistics of our sample is small, our sample is uniformly selected by broadband magnitude, removing possible biases due to the combination of two methods. Narrow-band surveys are biased towards sources with faint continuum and bright Lyman-$\alpha$ emission lines, whereas the $i'$-drop selection method is more uniformly sensitive to Lyman-$\alpha$ emission except at the very faint and low redshift ends of the sample. Our results do not support those of Ando et al., although clearly further observations are required to clarify the fraction of line emitters at bright magnitudes. It is possible that the fraction of sources with high equivalent width line emission is subject to environmental effects (e.g. the local number density of galaxies). The small HUDF field is, of course, subject to the effects of cosmic variance[@2004MNRAS.355..374B]. For sources as clustered as Lyman break galaxies at moderate redshifts are known to be, a variance in number density of 40% is expected in a field this size [@somerville04]. Discussion {#sec:disc} ========== The Redshift Distribution of $i'$-drop Galaxies {#sec:zdist} ----------------------------------------------- The redshift distribution expected for line emitters detected by the GLARE survey is the convolution of the selection functions imposed by the $i'$-drop selection criteria and by the spectral response of Gemini/GMOS in this configuration, modified by the probability of detecting a line that falls on a night sky emission line. In figure \[fig:zdist\] we show the expected distribution of emission line redshifts, neglecting the influence of strong line emission on source magnitude and colour (see above), and assuming a Schecter (1976) luminosity function with parameters $L^\ast=L^\ast$($z=3$), $\alpha=-1.9$ and $\phi^\ast=\phi^\ast$($z=3$) (Stanway 2004). Changing the parameters of the luminosity function has a slight effect on the shape of the distribution, with a reduction in $L^\ast$ with increasing redshift slightly broadening the peak of the distribution in redshift space. We take the conservative assumption that a source has a detection probability of 0% if lying on a very strong line ($>5\times$ the inter-line sky continuum) and 50% on top of a weaker sky line ($>1.5\times$ sky continuum level). The night sky spectrum is measured directly from our spectroscopy. Clearly, our survey is most sensitive to sources lying towards the lower-redshift end of our redshift range, and to sources lying in the low OH line regions commonly exploited by Lyman-$\alpha$ surveys at $z\approx5.7$ and $z\approx6.6$. Nonetheless, we have a reasonable probability of detecting sources in the skyline complexes at $z<6.5$, particularly if the source emission lies between skylines. Given this theoretical sensitivity distribution, it is interesting to note that the five robust line emitters presented in this paper (and one possible emission line also lying in this redshift range) do not fall at the peak of the detection sensitivity distribution, but rather within the skyline complexes. In @2004ApJ...604L..13S we suggested weak evidence for large scale structure in the HUDF field at $z=5.8$, based on the three redshifts known at the time. More recently @2004astro.ph..3458P have found evidence for an over-density at $z=5.9\pm0.2$ in the same field based on GRAPES Grism spectroscopy. Given that it unlikely that our fainter targets will yield redshifts with the current generation of spectrograph, confirmation of this may prove challenging. However, our results so far appear consistent with large scale structure at $z\approx5.8$ in this field. The Space Density of Lyman-$\alpha$ Emitters {#sec:LAEs} -------------------------------------------- We have identified a total of 12 (5 of which were strong) candidate emission lines with a luminosity in the range $(1-50)\times10^{42}$ergss$^{-1}$ in our sample of 24 $i'$-drop galaxy candidates. If this is indicative of the $i'$-drop selected galaxy population as a whole, and if the equivalent width distribution is invariant with magnitude, we would expect to observe line emission of this strength in approximately 25 (11 strong) of the 51 $i'$-drop galaxies identified in the HUDF by @2004MNRAS.355..374B (which were used by GLARE for candidate selection). Hence, this suggests a line emitter space density of $6.7 (2.9) \times10^{-4}$($\Delta log{L}$)$^{-1}$Mpc$^{-3}$), using the effective comoving volume for the $i'$-drop selection in the HUDF survey of $2.6\times10^{4}$Mpc$^{3}$. This space density comparable to the space density of $6.3\times10^{-4}$($\Delta log{L}$)$^{-1}$Mpc$^{-3}$ estimated by @2006PASJ...58..313S for Lyman-alpha emitters of comparable luminosity ($(3-25)\times10^{42}$) at $z\approx5.7$ in the Subaru Deep Field. Similarly, at $z=5.7$ the Lyman-Alpha Large Area Survey (LALA) found a number density of $3.2\times10^{-4}$Mpc$^{-3}$ for Lyman-$\alpha$ emitters with $L>10^{42}$ergs$^{-1}$ [@2004ApJ...617L...5M]. This would predict the detection of 7 Lyman-$\alpha$ emitters to our detection limit in our survey volume. This is consistent with, although slightly lower, than the number of emitters we detect. Inferring the space density of emitters from the relatively small sample in this survey is, of course, subject to large errors due to Poisson noise on the small number statistics. However, at least two physical explanations are possible for a discrepency between the line emitter density found by GLARE and by LAE surveys. As mentioned in section \[sec:zdist\], the HUDF is in a region with a known overdensity of sources at $5.7<z<5.9$. GLARE is also likely to detect a higher apparent density of line emitters if the Lyman-$\alpha$ photon escape fraction is higher in the relatively faint galaxies studied here than in more massive galaxies. If this is true then the number of sources observed at faint magnitudes exceeds the number predicted from brighter surveys. This phenomenon may well be expected if the deeper potential well of more massive galaxies traps more dust and a denser ISM than that existing in the shallow potential wells of small, faint galaxies. The Emission line properties of sub-L$^\ast$ Galaxies {#sec:sub_lstar} ----------------------------------------------------- This spectroscopy is the longest integrated exposure on a single mask using an 8m telescope. It reaches the faintest line limits available for a large sample with a homogeneous $i'$-drop colour cut. The 3 sigma limit in most slits is approx $4.1\times10^{-19}$ergcm$^{-2}$s$^{-1}$Å$^{-1}$. Hence we are able to observe line emissions from candidates significantly fainter than other surveys targeting $i'$-drop galaxies. Given the expenditure in telescope resources required to reach these faint limits, it is likely that many of the HUDF candidates may well remain inaccessible to ground based spectroscopy until the advent of the Extremely Large Telescopes, and from space by JWST with NIRSpec. HUDF $i'$-drops can be identified with reasonable confidence to $z'_{AB}<28.5$ (Bunker et al. 2004). A source with $z'_{AB}=28.5$ and an emission line of W=20Å (rest) at $z=6$, with complete Lyman-$\alpha$ absorption shortward of the line, would have a contamination fraction of 13.5% to the $z'$-band from this line, and a continuum flux of $6.4\times10^{-21}$ergcm$^{-2}$s$^{-1}$Å$^{-1}$ – i.e. a continuum flux a factor of $\sim 10$ less than we are able to detect at the $1\,\sigma$ level per spectral resolution element (6.5Å) . However, such a source would possess a line flux of $6\times10^{-18}$ergcm$^{-2}$s$^{-1}$, well within our detection limits for emission lines. Hence, although emission lines are detectable in surveys such as this, even for the faintest of our target sources, continuum spectra will remain inaccessible for some time to come. Even GLARE 1042, the brightest $z\approx6$ in the HUDF does not have sufficient signal to noise to in this spectroscopy to detect absorption features, even though the continuum is now detected at high significance. The detection of stellar absorption lines would provide valuable kinematic information on the galaxies and their outflows, however while such investigations may be possible on a stacked image of our faint spectra, it is impossible on individual slits. A stack of the faint GLARE spectra will be considered in a future paper. Conclusions {#sec:conc} =========== In this paper we have presented spectroscopy of the faintest known sample of $i'$-drop galaxies, derived from the Hubble Ultra Deep Field. An exposure of 36 hours per target was obtained with the Gemini-South telescope. Our main conclusions can be summarised as follows: i\) We have obtained extremely deep spectroscopy for twenty-nine science targets, reaching $3\,\sigma$ line flux limits of $2.5\times10^{-18}$ergcm$^{-2}$s$^{-1}$ – corresponding to equivalent width limits of $<6$Å for the majority of our targets. ii\) We identify five $i'$-drop galaxies as good Lyman-$\alpha$ emitter candidates, four as possible candidates, and a further five tentative emission lines which we consider unlikely to be real. iii\) We have considered the observed equivalent width distribution of $i'$-drop galaxies in the HUDF, and identify a tail of line emitters with very high equivalent widths which is not seen in the lower-redshift Lyman-break galaxies at $z\approx 3-4$. iv\) Several possible explanations for this effect exist. These include a tendency towards stronger line emission in faint sources, extreme youth or low metallicity in the Lyman-break population at high redshift, or possibly a top-heavy initial mass function. v\) At the low redshift end of our selection function ($5.6<z<6.0$), the $i'$-drop selection method will fail to select line emitters with high equivalent width due to line contamination producing blue colours. In contrast, at the high redshifts end ($z>6$), continuum-faint line emitters may enter the selection function if the Lyman-$\alpha$ line is sufficiently luminous. This has implications for the redshift and continuum magnitude selection function of $i'$-drop galaxies. vi\) This sample significantly increases the number of faint $i'$-drop galaxies with known redshifts, and may begin to bridge the divide between continuum and line selected galaxies at $z\approx6$. Acknowledgments {#acknowledgments .unnumbered} =============== Based on observations obtained at the Gemini observatory, which is operated by AURA Inc, under a cooperative agreement with the NSF, on behalf of the Gemini partnership. Also based in part on observations with the NASA/ESA Hubble Space Telescope, obtained at the Space Telescope Science Institute which is operated by AURA Inc under NASA contract NAS 5-26555. We thank the GOODS and HUDF teams for promptly releasing their valuable surveys to the community. We also thank all members of the GLARE collaboration. ERS gratefully acknowledges support from NSF grant AST 02-39425. AJB acknowledges support from a Philip Leverhulme Prize. We thank the anonymous reviewer for their helpful comments which have enhanced this paper. R. G., [Glazebrook]{} K., [McCarthy]{} P. J., [Crampton]{} D., [Murowinski]{} R., [Jorgensen]{} I., [Roth]{} K., [Hook]{} I. M., [Savaglio]{} S., [Chen]{} H., [Marzke]{} R. O., [Carlberg]{} R. G., 2004, ArXiv Astrophysics e-prints K. L., [Steidel]{} C. C., [Shapley]{} A. E., [Pettini]{} M., 2003, , 584, 45 D. M., [Bauer]{} F. E., [Brandt]{} W. N., [Schneider]{} D. P., [Hornschemeier]{} A. E., [Vignali]{} C., [Barger]{} A. J., [Broos]{} P. S., [Cowie]{} L. L., [Garmire]{} G. P., [Townsley]{} L. K., [Bautz]{} M. W., [Chartas]{} G., [Sargent]{} W. L. W., 2003, , 126, 539 M., [Ohta]{} K., [Iwata]{} I., [Akiyama]{} M., [Aoki]{} K., [Tamura]{} N., 2006, ArXiv Astrophysics e-prints A. J., [Martini]{} P., [Nelson]{} C. H., [Ho]{} L. C., 2003, , 594, L95 R. H., [Fan]{} X., [White]{} R. L., [Strauss]{} M. A., [Narayanan]{} V. K., [Lupton]{} R. H., [Gunn]{} J. E., [Annis]{} J., [Bahcall]{} N. A., [Brinkmann]{} J., [Connolly]{} A. J., [Csabai]{} I., [Czarapata]{} P. C., [Doi]{} M., [Heckman]{} T. M., [Hennessy]{} G. S., [Ivezi[' c]{}]{} [Ž]{}., [Knapp]{} G. R., [Lamb]{} D. Q., [McKay]{} T. A., [Munn]{} J. A., [Nash]{} T., [Nichol]{} R., [Pier]{} J. R., [Richards]{} G. T., [Schneider]{} D. P., [Stoughton]{} C., [Szalay]{} A. S., [Thakar]{} A. R., [York]{} D. G., 2001, , 122, 2850 Beckwith, S. V. W. et al, 2006, (submitted), astro-ph/0607632 R. J., [Illingworth]{} G. D., [Blakeslee]{} J. P., [Broadhurst]{} T. J., [Franx]{} M., 2004, ArXiv Astrophysics e-prints R. J., [Illingworth]{} G. D., [Thompson]{} R. I., [Blakeslee]{} J. P., [Dickinson]{} M. E., [Broadhurst]{} T. J., [Eisenstein]{} D. J., [Fan]{} X., [Franx]{} M., [Meurer]{} G., [van Dokkum]{} P., 2004, , 606, L25 R. J., [Illingworth]{} G. D., [Blakeslee]{} J. P., [Franx]{} M., 2006, , 653, 53 A. J., [Stanway]{} E. R., [Ellis]{} R. S., [McMahon]{} R. G., 2004, , 355, 374 A. J., [Stanway]{} E. R., [Ellis]{} R. S., [McMahon]{} R. G., [McCarthy]{} P. J., 2003, , 342, L47 M., [Stern]{} D., [Giavalisco]{} M., [Ferguson]{} H. C., [Tsvetanov]{} Z., [Chornock]{} R., [Cristiani]{} S., [Dawson]{} S., [Dey]{} A., [Filippenko]{} A. V., [Moustakas]{} L. A., [Nonino]{} M., [Papovich]{} C., [Ravindranath]{} S., [Riess]{} A., [Rosati]{} P., [Spinrad]{} H., [Vanzella]{} E., 2004, , 600, L99 C. C., [Holden]{} B. P., [Bouwens]{} R. J., [van der Wel]{} A., [Illingworth]{} G. D., [Zirm]{} A., [Franx]{} M., [Rosati]{} P., [Ford]{} H., [van Dokkum]{} P. G., [Stanford]{} S. A., [Eisenhardt]{} P., [Fazio]{} G. G., 2005, , 630, L137 Dow-Hygelund C. C., et al., 2006, astro, arXiv:astro-ph/0612454 E., [Kneib]{} J.-P., [Rieke]{} G. H., [Ellis]{} R. S., [Richard]{} J., [Rigby]{} J., [Papovich]{} C., [Stark]{} D., [Santos]{} M. R., [Huang]{} J.-S., [Dole]{} H., [Le Floc’h]{} E., [P[é]{}rez-Gonz[á]{}lez]{} P. G., 2005, , 618, L5 L. P., [Bunker]{} A. J., [Stanway]{} E. R., [Lacy]{} M., [Ellis]{} R. S., [Doherty]{} M., 2005, , 364, 443 Glazebrook, K. & Bland-Hawthorn, J. 2001. , 113, 197 Lehnert M. D., Bremer M., 2003, ApJ, 593, 630 R., [Cox]{} P., [Caselli]{} P., [Beelen]{} A., [Bertoldi]{} F., [Carilli]{} C. L., [Kaufman]{} M. J., [Menten]{} K. M., [Nagao]{} T., [Omont]{} A., [Wei[ß]{}]{} A., [Walmsley]{} C. M., [Walter]{} F., 2005, , 440, L51 S., [Rhoads]{} J., 2004a, ArXiv Astrophysics e-prints S., et al, 2005, , 626, 666 S., [Rhoads]{} J. E., 2002, , 565, L71 S., [Rhoads]{} J. E., 2004b, , 617, L5 N., [Xu]{} C., [Malhotra]{} S., [Rhoads]{} J. E., [Koekemoer]{} A. M., [Moustakas]{} L. A., [Walsh]{} J. R., [Windhorst]{} R. A., [Daddi]{} E., [Cimatti]{} A., [Ferguson]{} H. C., [Gardner]{} J. P., [Gronwall]{} C., [Haiman]{} Z., [K$\{$[" u]{}$\}$mmel]{} M., [Panagia]{} N., [Pasquali]{} A., [Stiavelli]{} M., [Alighieri]{} S. d., [Tsvetanov]{} Z., [Vernet]{} J., [Yan]{} H., 2004, ArXiv Astrophysics e-prints D., 2002, , 382, 28 A. E., [Steidel]{} C. C., [Pettini]{} M., [Adelberger]{} K. L., 2003, , 588, 65 K., [Kashikawa]{} N., [Doi]{} M., [Ly]{} C., [Malkan]{} M. A., [Matsuda]{} Y., [Ouchi]{} M., [Hayashino]{} T., [Iye]{} M., [Motohara]{} K., [Murayama]{} T., [Nagao]{} T., [Ohta]{} K., [Okamura]{} S., [Sasaki]{} T., [Shioya]{} Y., [Taniguchi]{} Y., 2006, , 58, 313 Somerville R,. et al, 2004, , L121 E. R., [Bunker]{} A. J., [McMahon]{} R. G., 2003, , 342, 439 [Stanway]{} E. R., 2004, PhD thesis, University of Cambridge E. R., [Bunker]{} A. J., [McMahon]{} R. G., [Ellis]{} R. S., [Treu]{} T., [McCarthy]{} P. J., 2004a, , 607, 704 E. R., [Glazebrook]{} K., [Bunker]{} A. J., [Abraham]{} R. G., [Hook]{} I., [Rhoads]{} J., [McCarthy]{} P. J., [Boyle]{} B., [Colless]{} M., [Crampton]{} D., [Couch]{} W., [J[ø]{}rgensen]{} I., [Malhotra]{} S., [Murowinski]{} R., [Roth]{} K., [Savaglio]{} S., [Tsvetanov]{} Z., 2004b, , 604, L13 E. R., [McMahon]{} R. G., [Bunker]{} A. J., 2005, , 359, 1184 C. C., [Adelberger]{} K. L., [Shapley]{} A. E., [Pettini]{} M., [Dickinson]{} M., [Giavalisco]{} M., 2003, , 592, 728 C. C., [Pettini]{} M., [Hamilton]{} D., 1995, , 110, 2519 Thompson R. I., et al., 2005, AJ, 130, 1 E., [Cristiani]{} S., [Dickinson]{} M., [Kuntschner]{} H., [Moustakas]{} L. A., [Nonino]{} M., [Rosati]{} P., [Stern]{} D., [Cesarsky]{} C., [Ettori]{} S., [Ferguson]{} H. C., [Fosbury]{} R. A. E., [Giavalisco]{} M., [Haase]{} J., [Renzini]{} A., [Rettura]{} A., [Serra]{} P., 2004, ArXiv Astrophysics e-prints E., [Cristiani]{} S., [Dickinson]{} M., [Kuntschner]{} H., [Moustakas]{} L. A., [Nonino]{} M., [Rosati]{} P., [Stern]{} D., [Cesarsky]{} C., [Ettori]{} S., [Ferguson]{} H. C., [Fosbury]{} R. A. E., [Giavalisco]{} M., [Haase]{} J., [Renzini]{} A., [Rettura]{} A., [Serra]{} P., [The Goods Team]{} 2005, , 434, 53 E., [Cristiani]{} S., [Dickinson]{} M., [Kuntschner]{} H., [Nonino]{} M., [Rettura]{} A., [Rosati]{} P., [Vernet]{} J., [Cesarsky]{} C., [Ferguson]{} H. C., [Fosbury]{} R. A. E., [Giavalisco]{} M., [Grazian]{} A., [Haase]{} J., [Moustakas]{} L. A., [Popesso]{} P., [Renzini]{} A., [Stern]{} D., [the GOODS Team]{} 2006, ArXiv Astrophysics e-prints H., et al, 2005, , 634, 109 \[lastpage\] [^1]: Current Address: H H Wills Physics Laboratory, Bristol, BS8 1TL, UK

ArXiv

--- abstract: 'We prove an additivity property for the normalized Seiberg–Witten invariants with respect to the universal abelian cover of those 3–manifolds, which are obtained via negative rational Dehn surgeries along connected sum of algebraic knots. Although the statement is purely topological, we use the theory of complex singularities in several steps of the proof. This topological covering additivity property can be compared with certain analytic properties of normal surface singularities, especially with functorial behaviour of the (equivariant) geometric genus of singularities. We present several examples in order to find the validity limits of the proved property, one of them shows that the covering additivity property is not true for negative definite plumbed 3–manifolds in general.' address: - 'A. Rényi Institute of Mathematics, 1053 Budapest, Reáltanoda u. 13-15, Hungary.' - 'A. Rényi Institute of Mathematics, 1053 Budapest, Reáltanoda u. 13-15, Hungary.' author: - József Bodnár - András Némethi title: 'Seiberg–Witten invariant of the universal abelian cover of $S^3_{-p/q}(K)$' --- Introduction ============ [s:intro]{} Motivation ---------- [ss:motiv]{} In this paper we prove an additivity property of the 3–dimensional (normalized) Seiberg–Witten invariant with respect to an abelian cover, valid for surgery 3–manifolds. Namely, assume that $M$ is obtained as a negative rational surgery along connected sum of algebraic knots in the three-sphere $S^3$. Let $\Sigma$ be its universal abelian cover. Theorem \[thm:main\] states that the sum over all spin$^c$ structures of the Seiberg–Witten invariants of $M$ (after normalisation) equals to the canonical Seiberg–Witten invariant of $\Sigma$. Both covers of manifolds, and manifolds of form $S^3_{-p/q}(K)$, are extensively studied in recent articles. The stability of certain properties and invariants with respect to the coverings is a key classical strategy in topology, it is even more motivated by the recent proof of Thurston’s virtually fibered conjecture [@Agol; @Wise]. Manifolds of form $S^3_{-p/q}(K)$ can be particularly interesting due to theorem of Lickorish and Wallace [@Lic; @Wal] stating that every closed oriented three-manifold can be expressed as surgery on a *link* in $S^3$. Based on this result, one can ask which manifolds have surgery representations with some restrictions. For example, using Heegaard–Floer homology, [@HKL] provides necessary conditions on manifolds having surgery representation along a *knot*. In this context, Theorem \[thm:main\] can be viewed also as a criterion for a manifold having surgery representation of form $S^3_{-p/q}(K)$ with $K$ a connected sum of algebraic knots. In fact, Seiberg–Witten invariants (SW) and Heegaard–Floer homologies are closely related. The SW invariants were originally introduced by Witten in [@Witten], but they also arise as Euler characteristics of Heegaard–Floer homologies, cf. [@OSz1; @OSz2]. In this article we will involve another cohomology theory with similar property. Since $S^3_{-p/q}(K)$ is representable by a negative definite plumbing graph, via [@NemSW] we view the SW invariants as Euler characteristics of lattice cohomologies introduced in [@Nlatnorm]. The big advantage of the lattice cohomology over the classical definition of Heegaard–Floer homology is that it is computable algorithmically from the plumbing graph. In the last section of applications and examples the above ‘covering additivity property’ will be combined with results involving lattice cohomology. Another strong motivation to study the above property is provided by the theory of complex normal surface singularities: the geometric genus of the analytic germ is conjecturally connected with the SW invariant of the link of the germ. Since the geometric genus satisfies the ‘covering additivity property’ (cf. §\[ss:sing\]), it is natural to ask for the validity of similar property at purely topological level. Furthermore, from the point of view of singularity theory, the motivation for the surgery manifolds $S^3_{-p/q}(K)$ is also strong: the link of the so called *superisolated singularities* (introduced in [@Lue]) are of this form. These singularities are the key test-examples and provide counterexamples for several conjectures. They embed the theory of projective plane curves to the theory of surface singularities. For their brief introduction see Example \[ex:SIS\], for a detailed presentation see [@Lue; @LMN]. All these connections with the analytic theory will be used deeply in several points of the proof. For consequences of the main result regarding analytic invariants see the last sections. Notations. ---------- [ss:not]{} We recall some facts about negative definite plumbed 3– and 4–manifolds, their spin$^c$ structures and Seiberg–Witten invariants. For more see [@NemnicI; @NemnicIII]. Let $M$ be a $3$–manifold which is a rational homology sphere ($\mathbb{Q}HS^3$). Assume that it has a negative definite plumbing representation with a decorated connected graph $G$ with vertex set $\mathcal{V}$. In particular, $M$ is the boundary of a plumbed $4$–manifold $P$, which is obtained by plumbing disc bundles over oriented surfaces $E_v\simeq S^2$, $v \in \mathcal{V}$ (according to $G$), and which has a negative definite intersection form. A vertex $v \in \mathcal{V}= \mathcal{V}(G)$ is decorated by the *self–intersection* $e_v \in \mathbb{Z}$ (of $E_v$ in $P$). In other words, $e_v$ is the euler number of the disc bundle over $E_v \cong S^2$. Since $M$ is a $\mathbb{Q}HS^3$, the graph $G$ is a tree. We set $\#\mathcal{V}(G)$ for the number of vertices of $G$. Below all the (co)homologies are considered with ${\mathbb Z}$–coefficients. Denote by $L = L_G= \mathbb{Z}\langle E_v \rangle_{v\in \mathcal{V}}$ the free abelian group generated by basis elements $E_v$, indexed by ${\mathcal V}$. It can be identified with $H_2(P)$, where $E_v$ represent the zero sections of the disc–bundles. It carries the negative definite intersection form $(.,.) = (.,.)_G$ (of $P$; readable from $G$ too). This form naturally extends to $L \otimes \mathbb{Q}$. Denoting by $L' = L_G' = \textrm{Hom}_{\mathbb Z}(L, \mathbb{Z})$ the dual lattice, one gets a natural embedding $L \rightarrow L'$ by $l \mapsto (\cdot,l)$. Furthermore, we can regard $L'$ as a subgroup of $L \otimes \mathbb{Q}$, therefore $(\cdot,\cdot)$ extends to $L'$ as well. We introduce the anti-dual basis elements $E_v^{\ast}$ in $L'$ defined by $(E_{v'}, E_v^{\ast})$ being $-1$ if $v = v'$ and $0$ otherwise. Notice that $L' \cong H^2(P) \cong H_2(P, M)$. The short exact sequence $0\to H_2(P)\to H_2(P,M)\to H_1(M)\to 0$ identify $L'/L $ with $H_1(M)$, which will be denoted by $H$. We denote the *class* of $l' \in L'$ by $[l'] \in H$, and we call $l' \in L'$ a *representative* of $[l']$. Assume that the intersection form in the basis $\{E_v\}_v$ has matrix $I$; then we define $\det(G):=\det(-I)$. It also equals the order of $H$. (Since $I$ is negative definite, $\det(G)>0$.) For any $h \in H$, we denote by $r_h= \sum_{v \in \mathcal{V}} c_v E_v$ ‘the smallest effective representative’ of its class in $L'$, determined by the property $0 \leq c_v < 1$ for all $v$. Finally, we define the *canonical characteristic element* in $L'$. It is the unique element $k_G \in L'$ such that $(k_G, E_v) = -(E_v,E_v)-2$ for every $v \in \mathcal{V}$. (In fact, $P$ carries the structure of a smooth complex surface — in the case of singularities, $P$ is a resolution, cf. Section 2 —, and $k_G$ is the first Chern class of its complex cotangent bundle.) The Seiberg–Witten invariants of $M$ associate a rational number to each spin$^c$ structure on $M$. There is a ‘canonical’ spin$^c$ structure $\sigma_{\textrm{can}} \in {\rm Spin}^c(M)$, the restriction of that spin$^c$ structure of $P$, which has first Chern class $k_G\in H^2(P)$. As we assumed $M$ to be a $\mathbb{Q}HS^3$, ${\rm Spin}^c(M)$ is finite. It is an $H$ torsor: for $h \in H$, we denote this action by $\sigma \mapsto h \ast \sigma$. We denote by ${\mathfrak{sw}}_{\sigma}(M) \in \mathbb{Z}$ the *Seiberg–Witten invariant* of $M$ corresponding to the spin$^c$ structure $\sigma$. This is the classical monopole counting Seiberg–Witten invariant of $M$ corrected by the Kreck–Stolz invariant to make it dependent only on the manifold $M$. Now we are ready to define the following invariant for each homology element $h \in H$: $$\label{eq:Intr1} \mathfrak{i}_h(M) := \frac{(k_G+2r_h, k_G+2r_h) + \#\mathcal{V}}{8}.$$ In fact, it does not depend on the particular plumbing representation of the manifold $M$ (or $P$); it is an invariant of the manifold $M$. Next we define the following normalization of the Seiberg–Witten invariant: for any $h \in H $, we set $$\label{eq:Inrt2} \mathfrak{s}_h(M) = {\mathfrak{sw}}_{h \ast \sigma_{\textrm{can}}}(M) - \mathfrak{i}_h(M).$$ Sometimes we will also use the notations $\mathfrak{s}_h(G) = \mathfrak{s}_h(M)$, or ${\mathfrak{sw}}_h(G) = {\mathfrak{sw}}_{h \ast \sigma_\textrm{can}}(M)$. In fact, $\mathfrak{s}_h(M) \in \mathbb{Z}$. This can be seen easily through the identity . We also refer to §\[ss:latticedef\] for the fact that $\mathfrak{s}_h(M)$ (and thus $\mathfrak{i}_h(M)$ as well) is indeed independent of the plumbing representation of the manifold. Let $\Sigma$ be the universal abelian cover (UAC) of the manifold $M$: it is associated with the abelianisation $\pi_1(M)\to H_1(M)$. In the next definition, $0$ is the unit element in $ H_1(\Sigma)$. \[def:s\] We say that for a manifold $M$ *the ‘covering additivity property’ of the invariant* $\mathfrak{s}$ *holds with respect to the universal abelian cover* (shortly, ‘CAP of $\mathfrak{s}$ holds’) if $$\mathfrak{s}_{0}(\Sigma) = \sum_{h \in H_1(M)} \mathfrak{s}_{h}(M).$$ Our main result is the following. \[thm:main0\] Let $M= S^3_{-p/q}(K)$ be a manifold obtained by a negative rational Dehn surgery of $S^3$ along a connected sum of algebraic knots $K = K_1 \# \dots \# K_{\nu}$ ($p,q>0$, ${\rm gcd}(p,q)=1$). Assume that $\Sigma$, the UAC of $M$, is a $\mathbb{Q}HS^3$. Then CAP of $\mathfrak{s}$ holds. Though the statement is topological, in the proof we use several analytic steps based on the theory of singularities. These steps not only emphasize the role of the algebraic knots and of the negative definite plumbing construction, but they also provide the possibility to use certain deep results valid for singularities. We emphasize that the above covering additivity property is not true for general negative definite plumbed 3–manifolds (hence for general 3–manifolds either), cf. Example \[rem:counter\]. In particular, we cannot expect a proof of the main theorem by a general topological machinery. Preliminaries ============= [s:prelimi]{} Connection with singularity theory ---------------------------------- [ss:sing]{} We present the connection of Theorem \[thm:main0\] with singularity theory, namely, with the (equivariant) geometric genera of normal surface singularities and the Seiberg–Witten Invariant Conjecture of Némethi and Nicolaescu [@NemnicI]. For details we refer to [@Npq; @Nsplice; @NemSW; @NemnicI; @NemnicIII]. Let $(X,0)$ be a complex normal surface singularity (germ) with link $M$. Let $\pi:\widetilde{X}\to X$ be a good resolution with negative definite dual resolution graph $G$, which can be regarded also as a plumbing graph for the 4-manifold $\widetilde{X}$ and its boundary $M$. (Hence, the $E_v$’s in this context are the irreducible exceptional curves.) The *geometric genus* of the singularity is defined as $p_g(X) = \textrm{dim}_{\mathbb{C}} H^1(\widetilde{X}, \mathcal{O}_{\widetilde{X}})$, where $\mathcal{O}_{\widetilde{X}}$ is the structure sheaf of $\widetilde{X}$. It does not depend on the particular choice of the resolution. In [@NemnicI] the following conjecture was formulated for certain singularities, as a topological characterization of $p_g(X)$: $$\label{eq:pg} p_g(X) = \mathfrak{s}_0(M).$$ We say that the *Seiberg–Witten Invariant Conjecture* (SWIC) holds for $X$ if is true. It is natural to ask whether there is any similar connection involving the other Seiberg–Witten invariants? The answer is given in [@Npq; @Nsplice]. Let $(Y,0)$ be the universal abelian cover of the singularity $(X,0)$ (that is, its link $\Sigma$ is the regular UAC of $M$, $(Y,0)$ is normal, and $(Y,0)\to (X,0)$ is analytic). The covering action of $H = H_1(M)$ on $Y$ extends to the resolution $\widetilde{Y}$ of $Y$, hence $H$ acts on $H^1(\widetilde{Y}, \mathcal{O}_{\widetilde{Y}})$ as well, providing a eigenspace decomposition $\oplus _{\xi\in\widehat{H}}\, H^1(\widetilde{Y}, \mathcal{O}_{\widetilde{Y}})_{\xi}$, indexed by the characters $\xi\in\widehat{H}:= {\rm Hom}(H,{\mathbb C}^*)$ of $H$. Set $$p_g(X)_h = \textrm{dim}_{\mathbb{C}} H^1(\widetilde{Y}, \mathcal{O}_{\widetilde{Y}})_{\xi_h},$$ where $\xi_{h} \in \widehat{H}$ is the character given by $h'' \mapsto e^{2\pi i (l', l'')}$, $[l']=h$, $[l'']=h''$. The numbers $p_g(X)_h$ are called the *equivariant geometric genera* of $(X,0)$. Note that $p_g(X)_0 = p_g(X)$. We say that the *Equivariant Seiberg–Witten Invariant Conjecture* (EqSWIC) holds for $(X,0)$ if the next identity is satisfied for *every* $h \in H$: $$\label{eq:pgh} p_g(X)_h = \mathfrak{s}_h(M).$$ Observe that by the definition, $p_g(Y) = \sum_{h \in H} p_g(X)_h$. Hence, the next claim is obvious. \[cl:pgadd\] If for a singularity $(X,0)$ with $\mathbb{Q}HS^3$ link the EqSWIC holds, and for its (analytic) universal abelian cover $(Y,0)$ with $\mathbb{Q}HS^3$ link the SWIC holds, then for the link $M$ of $X$ the (purely topological) covering additivity property of $\mathfrak{s}$ also holds. \[ex:lens\] By [@Npq; @NemnicII] the assumptions of Claim \[cl:pgadd\] are satisfied [*e.g.*]{} by cyclic quotient and weighted homogeneous singularities, hence the CAP of $\mathfrak{s}$ holds for lens–spaces and Seifert rational homology sphere 3–manifolds. Theorem \[thm:main0\] proves CAP for surgery manifolds, and Example \[rem:counter\] shows that CAP does [*not*]{} hold for arbitrary plumbed 3–manifolds. It is convenient to extend the definitions (\[eq:Intr1\]) and (\[eq:Inrt2\]) for any representative $l'\in L'$: $$\mathfrak{i}_{l'}(G):= \frac{(k_G+2l', k_G+2l')_G + \#\mathcal{V}(G)}{8} \ \ \ \mbox{and} \ \ \ \mathfrak{s}_{l'}(M) = {\mathfrak{sw}}_{[l']}(G) - \mathfrak{i}_{l'}(G).$$ By a computation, for two representatives $[l'_1] = [l'_2] = h \in H$ one has: $$\label{eq:difsh}\mathfrak{s}_{l_2'}(G)-\mathfrak{s}_{l_2'}(G)= \chi(l'_2) - \chi(l'_1), \ \ \mbox{where} \ \ \ \chi(l') = -(l',l'+k_G)/2.$$ In particular, $$\label{eq:eulatgen} \mathfrak{s}_{l'}(G) = \ \mathfrak{s}_{[l']}(G) + \chi(l')-\chi(r_{[l']}).$$ The invariants $\{\mathfrak{s}_h(G)\}_h$ for many 3–manifolds (graphs) are computed. The next statement basically follows from Example \[ex:lens\] combined with the fact that the UAC of a lens space is $S^3$. \[prop:ratgraph\][@Npq; @Nlatnorm; @Nsplice] If $G$ is a (not necessarily minimal) graph of $S^3$ or of a lens–space then $\mathfrak{s}_h(G) = 0$ for every $h \in H$. The structure of the plumbing graph $G$ of $S^3_{-p/q}(K)$ ---------------------------------------------------------- [ss:graphs]{} In this section we describe the plumbing graph of $S^3_{-p/q}(K)$ and we also fix some additional notations. Let $K_j\subset S^3$ be the embedded knot of an irreducible plane curve singularity $\{f_j(x,y) = 0\}\subset ({\mathbb C}^2,0)$, where $f_j$ is a local holomorphic germ $ (\mathbb{C}^2,0)\rightarrow (\mathbb{C},0)$. Let $G_j$ be the [*minimal embedded resolution graph*]{} of $\{f_j(x,y) = 0\}\subset ({\mathbb C}^2,0)$, which is a plumbing graph (of $S^3$) with several additional decorations: it has an arrowhead supported on a vertex $u_j$, which represents $K_j$ (or, in a different language, the strict transform $S(f_j)$ of $\{f_j=0\}$ intersecting the exceptional $(-1)$–curve $E_{u_j}$). Furthermore, $G_j$ has a set of multiplicity decorations, the vanishing orders $\{m_v\}_v$ of the pullback of $f_j$ along the irreducible exceptional divisors and $S(f_j)$. We collect them in the total transform ${\rm div}(f_j)=S(f_j)+\sum_{v\in {\mathcal V}(G_j)}m_vE_v =S(f_j) + (f_j)$ of $f_j$, characterized by $({\rm div}(f_j),E_v)_{G_j}=0$ for any $v$, and $(f_j)$ is its part supported on $\cup_{v\in\mathcal{V}(G_j)} E_v$. (For more on the graphs of plane curve singularities see [@BrKn; @EN].) Next, we write the surgery coefficient in *Hirzebruch–Jung continued fraction* $$\label{eq:hj} p/q = k_0 - \cfrac{1}{k_1 - \cfrac{1}{k_2 - \cfrac{1}{\dots - \cfrac{1}{k_s} } } }\ =: [ k_0, k_1, \dots, k_s ],$$ where $k_i\in {\mathbb Z}$, $k_0\geq 1$, $k_1, \dots, k_s \geq 2$. Define $K := K_1 \# K_2 \# \dots \# K_{\nu}$. Then $M=S^3_{-p/q}(K)$ can be represented by a negative definite plumbing graph $G$, which is constructed as follows, cf. [@Npq; @NR]. $G$ consists of $\nu$ blocks, isomorphic to $G_1, \dots, G_{\nu}$ (without the multiplicity decorations and arrowheads), a chain $G_0$ of length $s$ consisting of vertices $\overline{u}_1, \dots, \overline{u}_s = u'$ with decorations $e_{\overline{u}_1} = -k_1, \dots, e_{\overline{u}_s} = -k_s$, respectively, and one central vertex $u$ which is connected to the vertex $u_j$ in each block $G_j$ and to the first vertex $\overline{u}_1$ of the chain $G_0$ with decoration $-k_1$. The vertex $u$ has decoration $e_u = -k_0 - \sum_{j=1}^{\nu}m_{u_j}$. Note that if $q=1$ then $G_0$ is empty. In this case, we have $s=0$ and $u = u'$. We use the notation $E_v$, $v \in \mathcal{V}(G)$, for the basis of the lattice $L_{G}$ associated with $G$. We simply write $(.,.)$ for the intersection form $ (.,.)_G$, and $E_v^{\ast}$ for the anti-dual elements in $G$; that is, $(E_{v'},E_v^{\ast}) = -\delta_{v,v'}$ with the Kronecker-delta notation. Similarly, we write $(.,.)_j = (.,.)_{G_j}$ for the intersection form of $G_j$ ($j = 0, \dots, \nu$). For any $v \in \mathcal{V}(G_j)$, we set $E_v^{\ast,j}\in L'(G_j)$ for the anti-dual of $E_v$ *in the graph* $G_j$; that is, $(E_{v'},E_v^{\ast,j})_j = -\delta_{v,v'}$ with the Kronecker-delta notation, $v'\in {\mathcal V}(G_j)$. We denote the canonical class of $G$ by $k_G$ and the canonical class of $G_j$ by $k_{G_j}$. By a general fact of surgeries, $H_1(M) = H = \mathbb{Z}_p$. In fact, $[E_{u'}^{\ast}]$ is a generator of this group (see the proof of Lemma \[lem:UACgraph\]). Therefore $H=\{[hE_{u'}^{\ast}]\}_h$, where $h\in \{0, 1, \dots, p-1\}$. The structure of plumbing graph $\Gamma$ of the UAC $\Sigma $ of $M=S^3_{-p/q}(K)$ ---------------------------------------------------------------------------------- [ss:graphs2]{} We construct a plumbing graph $\Gamma $ as follows. $\Gamma$ consists of $\nu$ blocks $\Gamma_1, \dots, \Gamma_{\nu}$ with distinguished vertices $w_1, \dots, w_{\nu}$, a chain $\Gamma_0$ of length $q-1$ consisting of vertices $\overline{w}_1, \dots, \overline{w}_{q-1} = w'$ all with decoration $-2$, and a ‘central’ vertex $w$ which is connected to vertices $w_j$ (one from each block $\Gamma_j$) and to $\overline{w}_1$ at one end of the chain $\Gamma_0$. If $q=1$ then $\Gamma_0$ is empty and $w = w'$. $\Gamma_j$ is a plumbing graph of the link of the suspension hypersurface singularity $\{g_j=0\}$, where $g_j(x,y,z_j) = f_j(x,y) + z_j^p$ (for its shape see [@Nemsig]). The vertex $w_j$ of $\Gamma_j$ is that vertex which supports the arrowhead, if we regard $\Gamma_j$ as the embedded resolution graph of $\{z_j = 0\} \subset \{g_j = 0\}$ (that is, it supports the strict transform of $\{z_j=0\}$). The self–intersection of $w$ is determined as follows. Let $F_v$, $v \in \mathcal{V}(\Gamma)$, denote the basis elements of the lattice $L(\Gamma)$ associated with $\Gamma$. We write ${\rm div}(z_j)=S(z_j)+ \sum_{v \in \mathcal{V}(\Gamma_j)} n_v F_v$ for the total transform of $\{z_j=0\}$ under the embedded resolution of $\{z_j = 0\} \subset \{g_j = 0\}$ with resolution graph $\Gamma_j$. (${\rm div}(z_j)$ topologically is characterized by $({\rm div}(z_j),F_v)_{\Gamma_j}=0$ for any $v\in{\mathcal V}(\Gamma_j)$; the strict transform $S(z_j)$ can be represented as an arrowhead on $w_j$.) Then, the central vertex $w$ has decoration $e_w = -1-\sum_{j=1}^{\nu} n_{w_j}$ in $\Gamma$. \[lem:UACgraph\] $\Gamma$ is a (possible) plumbing graph of the UAC $\Sigma$. Consider the following divisor $D$ supported on $L_G$. On each $G_j$ it is $(f_j)$, we put multiplicity 1 on $u$, multiplicity $k_0$ on $\overline{u}_1$, and in general, the numerator of $[ k_0, \ldots, k_{i-1}]$ on $\overline{u}_i$, $1\leq i\leq s$, for notations see §\[ss:graphs\]. Furthermore, put an arrowhead on $\overline{u}_s$ with multiplicity $p$. If this arrowhead represents a cut $S$ supported by $E_{\overline{u}_s}$, then $pS+D$ has the property that $(pS+D,E_v)=0$ for all $v\in {\mathcal V}(G)$, hence it is a topological analogue of the divisor of a function. The algorithm which provides the (topological) cyclic ${\mathbb Z}_p$–covering of the plumbed 4–manifold $P$ with branch locus $pS+D$ is identical with the algorithm from [@NCyc; @Nemsig] (which provides branched cyclic covers associated with analytic functions). The point is that $S$ has multiplicity $p$, hence the ${\mathbb Z}_p$–covering will have no branching along it, hence, in fact, the reduced branch locus is in $\cup_vE_v$. Note that $D/p=E^*_{u'}$, and its $E_u$–coefficient is $1/p$, hence the class of $E^*_{u'}$ (or of $D/p$) has order $p$ in $H$, hence it generates $H$. This implies that this algorithm provides exactly the UAC of $M$. Since the algorithm is ‘local’, and the multiplicity of $u$ is 1, over the subgraphs $G_j$ it is identical with that one which provides the graph of the suspension singularity $f_j+z_j^p$ (see again [@NCyc; @Nemsig]). Next we verify its behaviour over the graph $G_0$. This graph is the graph of a Hirzebruch–Jung singularity of type $(q,r)$, where $q/r=[k_1,\ldots, k_s]$. (For details regarding Hirzebruch–Jung singularities see [@BPV].) This is the normalization of $xy^{q-r}=z^q$. Using this coordinate choice, the strict transform of $y$ is exactly $qS$, the strict transform of $x$ is a disc $S'$ in $E_u$ (a disc neighbourhood of $E_u\cap E_{\overline {u}_1}$ in $E_u$) with multiplicity $q$; and finally, the strict transform of $z$ is $S'+(q-r)S$. In particular, the cyclic covering we consider over $G_0$ is exactly the cyclic ${\mathbb Z}_p$–covering of the normalization of $xy^{q-r}=z^q$ along the divisor of $zy^{k_0-1}$ (here for the $S$–multiplicity use the identity $q-r +(k_0-1)q=k_0q-r=p$). This is a new Hirzebruch–Jung singularity, the normalization of $xy^{q-r}=z^q$ and $zy^{k_0-1}=w^p$. The $q$–power of the second equation combined with the first one gives $xy^p=w^{pq}$, hence $t:=w^q/y$ is in the integral closure with $x=t^p$. Hence, after eliminating $x$, the new equations are $ty=w^q, \ t^py^{q-r}=z^q$ and $zy^{k_0-1}=w^p$. A computations shows that the integral closure of this ring is given merely by $ty=w^q$. This is an $A_{q-1}$ singularity, whose minimal resolution graph is $\Gamma_0$. Finally, notice that the above algorithm provides a system of multiplicities, which can be identified with a homologically trivial divisor, hence, similarly as in [@NCyc; @Nemsig], we get the last ‘missing Euler number’ $e_w$ too. The intersection form of $\Gamma$ will be denoted by $\langle.,.\rangle = (.,.)_{\Gamma}$. Similarly, $\langle.,.\rangle_j = (.,.)_{\Gamma_j}$ will denote the intersection form of $\Gamma_j$. The canonical class of $\Gamma$ is $k_{\Gamma}$, the canonical class of $\Gamma_j$ is $k_{\Gamma_j}$. For any $v \in \mathcal{V}(\Gamma)$, $F_v^{\ast}$ will denote the anti-dual of the corresponding divisor $F_v$ in $\Gamma$. Similarly, for a vertex $v \in \mathcal{V}(\Gamma_j)$, $F_v^{\ast,j}$ is the anti-dual of $F_v$ in $\Gamma_j$. Set $J = H_1(\Sigma)$ and let $J_j$ be the first homology group of those 3–manifolds determined by $\Gamma_j$. \[lem:JJ\] $$J \cong J_1 \times \dots \times J_{\nu}.$$ Let $\rho_j\in \widehat{J}_j$ be a character of $\Gamma_j$, $j\geq 1$. In [@NemnicIII §6.3] is proved that $\rho_j$ takes value 1 on $F^{*,j}_{w_j}$ (recall that the vertex $w_j$ of $\Gamma_j$ is connected with the central vertex $w$). Hence, for $j\not=i$, $j,i\geq 1$, there is no edge $(v_j,v_i)$ of $\Gamma$, such that $v_j$ is in the support of $\rho_j$ and $v_i$ is in the support of $\rho_i$. This means that each $\rho_j\in\widehat{J}_j$ can be extended to a character of $J$, by setting $\rho_j(F^*_v)=1$ whenever $v\not\in{\mathcal V}(\Gamma_j)$; in this way providing a monomorphism $\widehat{J}_j\hookrightarrow \widehat{J}$. But the same property also guarantees that in fact one has a simultaneous embedding $\prod_{j\geq 1} \widehat{J}_j\hookrightarrow \widehat{J}$. Therefore, if we prove that $\prod_{j\geq 1}\det(\Gamma_j)=\det(\Gamma)$, then the above embedding becomes an isomorphism, hence Lemma follows. In determinant computations of decorated trees, the following formula is useful; see e.g. [@NB 4.0.1(d)]. Let $e$ be an edge of $\Gamma$ with end vertices $a$ and $b$. Then $\det(\Gamma)=\det(\Gamma\setminus e)-\det(\Gamma\setminus \{a,b\})$. This formula inductively (applied for the edges adjacent to $w$) provides the needed determinant identity. \[cl:eq\] $$\label{eq:intform1}\begin{split} (a) \ \ \ \ \, - p\cdot (E_u^{\ast}, E_u^{\ast}) =& \ q \ \ \textrm{\ and\ } \ \ -p\cdot (E_{u'}^{\ast}, E_u^{\ast}) = 1;\\ (b) \ \ \ \ q\cdot(E_{u_j}^{\ast,j}, E_v^{\ast,j})_j = & \ p\cdot (E_u^{\ast},E_v^{\ast}) \ \ \ \mbox{for any $v \in \mathcal{V}(G_j)$, $j \geq 1$};\\ (c) \ \ \ \ \ \ \ \ - \langle F_{w}^{\ast}, F_{w}^{\ast} \rangle = & \ q \ \ \textrm{\ and\ } \ \ -\langle F_{w'}^{\ast}, F_{w}^{\ast} \rangle = 1;\\ (d) \ \ \ \ q\cdot \langle F_{w_j}^{\ast,j}, F_v^{\ast,j} \rangle_j = & \ \langle F_w^{\ast}, F_v^{\ast} \rangle \ \ \ \mbox{for any $v \in \mathcal{V}(\Gamma_j)$, $j \geq 1$}.\end{split}$$ For a negative definite plumbing graph ${{\frak G}}$ (which is a tree) if we define the anti–duals $E^*_v$ as above, then the following holds: for any two vertices $a,b$ the expression $-\det({{\frak G}})\cdot(E^*_a,E^*_b)$ equals the product of the determinants of the connected components of that graph which is obtained from ${{\frak G}}$ by deleting the shortest path connecting $a$ and $b$ and the adjacent edges; see [@EN §10] in the integral homology case and [@NemnicI] in general. This applied for $G$ and $a=b=u$ (and $a=u$, $b=u'$) gives $(a)$, since $\det(G)=p$, $\det(G_j)=1$ for $j\geq 1$ and $\det(G_0)=q$. (b) follows similarly. (c) and (d) follows from this property combined with Lemma \[lem:JJ\]. Additivity property of the invariant $\mathfrak{s}$ =================================================== [s:main]{} Proof of the main theorem ------------------------- [ss:mainproof]{} Now we are ready to prove Theorem \[thm:main0\]. To adjust it to its proof, we recall it in a more explicit form, in the language of plumbing graphs. \[thm:main\] Let $K$ be a connected sum of algebraic knots, $p, q$ coprime positive integers. Assume that both $S^3_{-p/q}(K)$ (having plumbing graph $G$) and its universal abelian cover $\Sigma$ (with plumbing graph $\Gamma$) are rational homology spheres. Then the following additivity holds: $$\underbrace{{\mathfrak{sw}}_{0}(\Gamma) - \frac{\langle k_{\Gamma},k_{\Gamma}\rangle + \#\mathcal{V}(\Gamma)}{8}}_{\mathfrak{s}_0(\Gamma)} = \sum_{h = 0}^{p-1} \underbrace{\left[ {\mathfrak{sw}}_h(G) - \frac{(k_G + 2r_h, k_G + 2r_h) + \#\mathcal{V}(G)}{8} \right]}_{\mathfrak{s}_h(G)}.$$ On the left hand side $0$ is the unit element of $J = H_1(\Sigma)$ and on the right hand side we identified elements of $H \cong \mathbb{Z}_p$ with elements of $\{0, 1, \dots, p-1\}$, $0$ being the unit element and $1$ being the generator $[E_{u'}^{\ast}]$, *i.e.*, $r_h = r_{[hE_{u'}^{\ast}]}$. In fact, the condition whether $\Sigma$ is a $\mathbb{Q}HS^3$ or not is readable already from $p$ and the plane curve singularity invariants describing the knots $K_i$; cf. [@NemnicIII §6.2 (c)]. Notice that deleting from $G$ the ‘central vertex’ $u$ (and all its adjacent edges), one gets $G_0, G_1, \dots, G_{\nu}$ as connected components of the remaining graph. Also, deleting from $\Gamma$ the ‘central vertex’ $w$ (and all its adjacent edges), one gets $\Gamma_0, \Gamma_1, \dots, \Gamma_{\nu}$ as connected components of the remaining graph. ($\Gamma_0$ and $G_0$ are present only if $q > 1$). We use the notations $R_j$, resp. $\widetilde{R}_j$ for the ‘restriction’ homomorphisms $L_{G}' \rightarrow L_{G_j}'$, resp. $L_{\Gamma}' \rightarrow L_{\Gamma_j}'$, dual to the natural inclusions $L_{G_j} \rightarrow L_{G}$, resp. $L_{\Gamma_j} \rightarrow L_{\Gamma}$. They are characterised by $R_j(E_v^{\ast}) = E_v^{\ast,j}$, if $v \in \mathcal{V}(G_j)$ and $0$ otherwise, resp. $\widetilde{R}_j(F_v^{\ast}) = F_v^{\ast,j}$, if $v \in \mathcal{V}(\Gamma_j)$ and $0$ otherwise. E.g., $R_j(k_G) = k_{G_j}$ and $\widetilde{R}_j(k_{\Gamma}) = k_{\Gamma_j}$ (cf. [@NB Def. 3.6.1 (2)]). We can apply the surgery (‘cut-and-paste’) formula of [@NB Theorem 1.0.1] (note the sign difference due to the different sign convention about ${\mathfrak{sw}}$) and get the following two formulae. The new symbols $\mathcal{H}^{\textrm{pol}}_{u,h}(1)$ and $ \mathcal{F}^{\textrm{pol}}_{w,0}(1)$ are values of polynomials in $t=1$ as in [@NB §3.5]; their definitions will be recalled later in and . $$\label{eq:swsis2} \mathfrak{s}_h(G) = \mathcal{H}^{\textrm{pol}}_{u,h}(1) + \mathfrak{s}_{R_0(r_h)}(G_0) + \sum_{j=1}^{\nu} \mathfrak{s}_{R_j(r_h)}(G_j), $$ $$\label{eq:swuac2} \mathfrak{s}_0(\Gamma) = \mathcal{F}^{\textrm{pol}}_{w,0}(1) + \mathfrak{s}_0(\Gamma_0) + \sum_{j=1}^{\nu} \mathfrak{s}_0(\Gamma_j).$$ In , for $j \geq 1$, $[R_j(r_h)] = 0 \in L_{G_j}'/L_{G_j}$, as the latter one is the trivial group $H_1(S^3)$. Hence, by (\[eq:eulatgen\]), $\mathfrak{s}_{R_j(r_h)}(G_j)=\mathfrak{s}_0(G_j)+ \chi_j(R_j(r_h))$, where $\chi_j(x) := -\frac{1}{2}(x,x+k_{G_j})_j$. Furthermore, by Proposition \[prop:ratgraph\], $\mathfrak{s}_0(\Gamma_0) = 0$, and $\mathfrak{s}_0(G_j) = 0$ for $j \geq 1$ (as $G_j$ is a plumbing graph for $S^3$). Therefore, the desired equality $\mathfrak{s}_0(\Gamma) = \sum_{h=0}^{p-1} \mathfrak{s}_h(G)$ reduces to the proof of the following three lemmas. \[lem:pol\] $$\sum_{h=0}^{p-1} \mathcal{H}^{\textrm{pol}}_{u,h}(1) = \mathcal{F}^{\textrm{pol}}_{w,0}(1).$$ \[lem:susp\] $$\hspace*{2cm} \sum_{h=0}^{p-1} \chi_j(R_j(r_h)) = \mathfrak{s}_0(\Gamma_j) \ \ \ \ (\mbox{for $j \geq 1$}).$$ \[lem:lens\] $$\sum_{h=0}^{p-1} \mathfrak{s}_{R_0(r_h)}(G_0) = 0.$$ In the next paragraphs we recall the definition of $\mathcal{H}^{\textrm{pol}}_{u,h}$ and $\mathcal{F}^{\textrm{pol}}_{w,0}$ (following [@NB §3.5]) adapted to the present case and notations, and then we provide the proofs of the lemmas. First, given a rational function $\mathcal{R}(t)$ of $t$, one defines its *polynomial part* $\mathcal{R}^{\textrm{pol}}(t)$ as the unique polynomial in $t$ such that $\mathcal{R}(t) - \mathcal{R}^{\textrm{pol}}(t)$ is either $0$ or it can be written as a quotient of two polynomials of $t$ such that the numerator has degree strictly less than the denominator. Now $\mathcal{F}^{\textrm{pol}}_{w,0}$ and $\mathcal{H}^{\textrm{pol}}_{u,h}$ are polynomial parts of rational functions defined as follows. $$\label{eq:defh} \mathcal{H}_{u,h}(t) = \frac{1}{p} \sum_{\varrho \in \widehat{H}} \varrho^{-1}(h) \prod_{v \in \mathcal{V}(G)}(1-\varrho([E_v^{\ast}])t^{-p\cdot (E_u^{\ast},E_v^{\ast})})^{\delta_v-2},$$ where $\delta_v$ denotes the degree (number of adjacent edges) of a vertex $v\in \mathcal{V}(G)$. $$\label{eq:deff} \mathcal{F}_{w,0}(t) = \frac{1}{|J|}\sum_{\varrho \in \widehat{J}} \prod_{v \in \mathcal{V}(\Gamma)}(1-\varrho([F_v^{\ast}])t^{-|J|\langle F_w^{\ast}, F_v^{\ast} \rangle })^{\widetilde{\delta}_v-2},$$ where $\widetilde{\delta}_v$ denotes the degree of a vertex $v\in \mathcal{V}(\Gamma)$. Set $$\label{eq:h} \mathcal{H}_u(t) := \sum_{h=0}^{p-1}\mathcal{H}_{u,h}(t) = \prod_{v \in \mathcal{V}(G)}(1-t^{-p\cdot (E_u^{\ast},E_v^{\ast})})^{\delta_v-2}.$$ As taking polynomial parts of rational functions is additive, Lemma \[lem:pol\] follows if we prove $$\label{eq:polid} \mathcal{H}_{u}(t^{|J|}) = \mathcal{F}_{w,0}(t).$$ Let $\Delta_j = \Delta_{S^3}(K_j)$ be the Alexander polynomial of the knot $K_j$ (defined as in [@NemnicIII §2.6, (8)], or [@EN]). Then, since $E_{u_j}^{\ast,j} = (f_j) \in L_{G_j}$, $$\frac{\Delta_j(t)}{1-t} = \prod_{v \in \mathcal{V}(G_j)}(1 - t^{-(E_{u_j}^{\ast,j}, E_v^{\ast,j})_j})^{\delta_v-2}.$$ Comparing with the above formula for the Alexander polynomials and using the identities (a), (b) of Lemma \[cl:eq\] we get that $$\label{eq:halex} \mathcal{H}_u(t) = \frac{\prod_{j=1}^{\nu}\Delta_j(t^q)}{(1-t)(1-t^q)}.$$ Recall that $J_j = L_{\Gamma_j}'/L_{\Gamma}$ is the first homology group of the manifold determined by $\Gamma_j$ and that $F_{w_j}^{\ast,j} = (z_j) \in L_{\Gamma_j}$. Let $\Delta_{j,\Gamma}$ be the Alexander polynomial of the knot $K_{j,\Gamma}$ in the manifold of $\Gamma_j$ determined by $z_j=0$ (see [@NemnicIII §2.6, (8)]). That is, $$\frac{\Delta_{j,\Gamma}(t)}{1-t} = \frac{1}{|J_j|} \sum_{\varrho_j \in \widehat{J_j}} \prod_{v \in \mathcal{V}(\Gamma_j)} (1 - \varrho_j([F_v^{\ast,j}]) t^{-\langle F_{w_j}^{\ast,j},F_v^{\ast,j}\rangle_j})^{\widetilde{\delta}_v-2}.$$ Recall that $J = J_1 \times \dots \times J_{\nu}$. Consequently, any character $\varrho \in \widehat{J}$ can be written as a $\nu$-tuple of characters, $\varrho = (\varrho_1, \dots, \varrho_{\nu})$ with $\varrho_j \in \widehat{J}_j = \textrm{Hom}(J_j, \mathbb{C}^{*})$. Furthermore, for any $v \in \mathcal{V}(\Gamma_j)$, $\varrho([F_v^{\ast}]) = \varrho_j([F_v^{\ast,j}])$ and $\varrho([F_v^{\ast}]) = 1$ if $v = w$ or $v \in \Gamma_0$ as in that case $F_v^{\ast}$ represents the trivial element in $L_{\Gamma}'/L_{\Gamma}$ (see also the proof of Lemma \[lem:JJ\]). Comparing with the above formula for the Alexander polynomials and using the identities (c), (d) of Lemma \[cl:eq\] we get that, setting $s = t^{|J|}$, $$\label{eq:falex} \mathcal{F}_{w,0}(t) = \frac{\prod_{j=1}^{\nu}\Delta_{j,\Gamma}(s^q)}{(1-s)(1-s^q)}.$$ By [@NemnicIII Prop. 6.6] $\Delta_j = \Delta_{j,\Gamma}$, so via and we obtain . For any element $l' = \sum_{v \in \mathcal{V}(G)}c_vE_v \in L_{G}'$, let $\lfloor l' \rfloor := \sum_{v \in \mathcal{V}(G)} \lfloor c_v \rfloor E_v $, resp. $\{l'\}:=l'-\lfloor l' \rfloor$, denote the coordinatewise integer, resp. fractional part of $l'$ in the basis $\{E_v\}_v$. We use this notation for other graphs as well. Using the description of $E^*_{u'}$ in the proof of Lemma \[lem:UACgraph\] we have $$h E^*_{u'}=\sum_{j\geq 1} h\cdot (f_j)/p + hE_u/p+D_0 \ \ \ \ (0\leq h<p),$$ where $D_0$ is supported on $G_0$. Since $r_h=\{hE^*_{u'}\}$, we obtain $$r_h=hE^*_{u'}-\sum_{j\geq 1}\lfloor h\cdot (f_j)/p\rfloor -\lfloor D_0\rfloor.$$ Since $R_j(E^*_{u'})=0$, $R_j(E_v)=E_v$ for $v\in{\mathcal V}(\Gamma_j)$, and $R_j(E_v)=0$ for $v\not\in({\mathcal V}(\Gamma_j)\cup u)$, we get $$\label{eq:3216} R_j(r_h) = - \lfloor h\cdot (f_j)/p\rfloor.$$ As $\Gamma_j$ is the plumbing graph of a suspension hypersurface singularity $g_j(x,y,z_j) = f_j(x,y) + z_j^p = 0$ and, as it is proved in [@NemnicIII], *for such suspension singularities the SWIC holds* (see §\[ss:sing\]), we have $\mathfrak{s}_0(\Gamma_j) = p_g(\{g_j=0\})$. Hence, the statement of the Lemma is equivalent with $$p_g(\{g_j=0\}) = \sum_{h=0}^{p-1} \chi_j\left(- \lfloor h\cdot (f_j)/p \rfloor\right).$$ This geometric genus formula has major importance even independently of the present application. We separate the statement in the following Claim. \[cl:susppg\] Let $f(x,y) \in \mathbb{C}\{x,y\}$ be the equation of an irreducible plane curve singularity. Let $G_f$ be the dual resolution graph of a good embedded resolution of $f$, from which we delete the arrowhead (strict transform) of $f$ and all the multiplicities. Let $(f)$ be the part of the divisor of $f$ supported on the exceptional curves. Then for any positive integer $p$ the geometric genus of the suspension singularity $\{ g(x,y,z) = f(x,y) + z^p = 0 \}$ is $$p_g(\{g= 0 \}) = \sum_{h=0}^{p-1} \chi (-\lfloor h\cdot(f)/p\rfloor).$$ A combinatorial formula (involving Dedekind sums) for the *signature* of (the Milnor fibre of) suspension singularities was presented in [@Nemsig]. Recall that Durfee and Laufer type formulae imply that the geometric genus and the signature determine each other modulo the link (see *e.g.* [@Neminv Theorem 6.5] and the references therein). Nevertheless, the above formula is of different type. Let $\phi:Z\to ({\mathbb C}^2,0)$ be the embedded resolution of $f$. Consider the ${\mathbb Z}_p$ branched covering $c:(\{g=0\},0)\to ({\mathbb C}^2,0)$, the restriction of $(x,y,z)\mapsto (x,y)$. Let $c_\phi:W\to Z$ be the pullback of $c$ via $\phi$ and let $\widehat{c}_\phi: \widehat{W}\to Z$ be the composition of the normalization $n:\widehat{W}\to W$ with $c_\phi$. Then $\widehat{W}\to W\to \{g=0\}$ is a partial resolution of $\{g=0\}$: although it might have some Hirzebruch–Jung singularities, since these are rational, one has $p_g(\{g=0\})=h^1({\mathcal O}_{\widehat {W}})$. On the other hand, we claim that $$\label{eq:kollar} (\widehat{c}_\phi)_*({\mathcal O}_{\widehat{W}})=\oplus_{h=0}^{p-1}\, {\mathcal O}_Z(\lfloor h\cdot (f) /p\rfloor ).$$ This follows basically from [@kollar §9.8]. For the convenience of the reader we sketch the proof. We describe the sheaves $(c_\phi)_*({{\mathcal O}}_W)$ and $(\widehat{c}_\phi)_*({{\mathcal O}}_{\widehat{W}})$ in the neighbourhood $U$ of a generic point of the exceptional set $E$ of $\phi$. Consider such a point with local coordinates $(u,v)$, $\{u=0\}= E\cap U$, $(f)$ in $U$ is given by $u^m=0$. Consider the covering, a local neighbourhood of type $\{(u,v,z)\,:\, z^p=u^m\}$ in $W$. Then ${{\mathcal O}}_{W,0}$ as $\C\{u,v\}$–module is $\oplus _{h=0}^{p-1}z^h\cdot \C\{u,v\}$. For simplicity we assume ${\rm gcd}(m,p)=1$. The ${\mathbb Z}_p$–action is induced by the monodromy on the regular part, namely by the permutation of the $z$–pages, induced over the loop $u(s)=\{e^{2\pi i s}\}_{0\leq s\leq 1}$. This is the multiplication by $\xi:=e^{2\pi i m/p}$. Hence, $z^h\C\{u,v\}$ is the $\xi^h$–eigensheaf of $(c_\phi)_*({{\mathcal O}}_W)$. If we globalize $z\C\{u,v\}$, we get a line bundle on $Z$, say ${\mathcal{L}}$. Then the local representative of ${\mathcal{L}}^p$ is $z^p\C\{u,v\}=u^m\C\{u,v\}=\C\{u,v\}(-(f))$. Hence ${\mathcal{L}}^{p}$ is trivialized by $ f\circ \phi$. Since ${\rm Pic}(Z)=0$, ${\mathcal{L}}$ itself is a trivial line bundle on $Z$. Next, we consider the normalization $\widehat{W}$. Above $U$ it is $({\mathbb C}^2,0)$ with local coordinates $(t,v)$, and the normalization is $z=t^m$, $u=t^p$. In particular, $(\widehat{c}_\phi)_*({{\mathcal O}}_{\widehat{W},0})=\oplus _{h=0}^{p-1}t^h\cdot \C\{u,v\}$, where ${{\mathcal F}}^{(h)}:=t^h\cdot \C\{u,v\}$ is the $e^{2\pi ih/p}$–eigensheaf. Set the integer $m'$ with $0\leq m'<p$ and $mm'=1+kp$ for certain $k\in {\mathbb Z}$. Then one has the following eigensheaf inclusions: $t^h\C\{u,v\}\supset z^{\{\frac{hm'}{p}\}p} \cdot \C\{u,v\}={\mathcal{L}}^{\{\frac{hm'}{p}\}p}|_U$. Hence, for some effective cycle $D$ we must have $t^h\C\{u,v\}= {\mathcal{L}}^{\{\frac{hm'}{p}\}p}(D)|_U$. This, by taking $m$-power reads as $z^h\C\{u,v\}= z^{\{\frac{hm'}{p}\}pm} \C\{u,v\}(mD)$. This means that if $\{hm' /p\}=m_h/p$ and $mm_h=k_hp+h$ for certain integers $m_h$ and $k_h$, $0\leq m_h<p$, then the local equation of $mD$ is $z^{\{\frac{hm'}{p}\}pm-h}= z^{k_hp}$. Hence $D$ locally is given by $t^{k_hp}=u^{k_h}$. Since $k_h=\lfloor mm_h/p\rfloor$, the global reading of this fact is $D=\lfloor m_h\cdot (f)/p\rfloor$. Hence $$(\widehat{c}_\phi)_*({{\mathcal O}}_{\widehat{W}})=\oplus _{h=0}^{p-1}{\mathcal{L}}^{\{\frac{hm'}{p}\}p} (\lfloor m_h\cdot (f)/p\rfloor).$$ Since ${\mathcal{L}}$ is a trivial bundle, and $h\mapsto m_h$ is a permutation of $\{0,\ldots, p-1\}$, (\[eq:kollar\]) follows. Next, from (\[eq:kollar\]) we obtain $p_g(\{g=0\})=\sum_h h^1({\mathcal O}_Z( \lfloor h\cdot (f) /p\rfloor )$. Set $D':=\lfloor h\cdot (f) /p\rfloor $. Then from the cohomological exact sequence of the exact sequence of sheaves $0\to {\mathcal O}_Z\to {\mathcal O}_Z(D')\to {\mathcal O}_{D'}(D')\to 0$, and from $h^1({\mathcal O}_Z)=p_g(({\mathbb C}^2,0))=0$, we get $h^1({\mathcal O}_Z(D'))=h^1({\mathcal O}_{D'}(D'))$. Since by Grauert–Riemenschneider vanishing $ h^0({\mathcal O}_{D'}(D'))=0$, we have $$h^1({\mathcal O}_Z(D'))=-\chi({\mathcal O}_{D'}(D'))=-(D',D')+(D',D'+K)/2=\chi(-D').$$ This ends the proof of the Claim. Moreover, the proof of Lemma \[lem:susp\] is also completed. We observe two facts. First, from the proof of Lemma \[lem:UACgraph\] we obtain that $R_0(r_h)$ only depends on the value $p/q$ (and not on the blocks $G_j$, $j \geq 1$). It has the same expression even if we replace all the graph $G_j$ by the empty graph. Second, from the equations (\[eq:swuac2\]) and (\[eq:swsis2\]) and the discussion after it we get that under the validity of Lemmas \[lem:pol\] and \[lem:susp\] (what we already proved for any situation) the main Theorem \[thm:main\] (property CAP) is equivalent with Lemma \[lem:lens\]. Put these two together, the validity of \[lem:lens\] is equivalent with the validity of CAP in the case when $G_j=\emptyset$ for all $j\geq 1$. But CAP for $G_0\cup\{u\}$ is true by Claim \[cl:pgadd\] and Example \[ex:lens\]. The invariant $\mathfrak{s}_h$ and lattice cohomology ===================================================== [s:lattice]{} Lattice cohomology ------------------ [ss:latticedef]{} The normalized SW invariant $\mathfrak{s}_h(G)$ can also be expressed as the Euler characteristic of the *lattice cohomology*. The advantage of this approach is that it provides an alternative, completely elementary way to define $\mathfrak{s}$, as the definition of the lattice cohomology is purely combinatorial from the plumbing graph $G$. We briefly recall the definition and some facts about the lattice cohomology associated with a ${\mathbb Q}HS^3$ 3–manifold with negative definite plumbing graph $G$. For more see [@Nlatnorm; @NR]. Usually one starts with a lattice $\Z^s$ with fixed base elements $\{E_i\}_i$. This automatically provides a cubical decomposition of $\R^s=\Z^s\otimes \R$: the 0–cubes are the lattice points $l\in \Z^s$, the 1–cubes are the ‘segments’ with endpoints $l$ and $l+E_i$, and more generally, a $q$–cube $\square=(l,I)$ is determined by a lattice point $l\in \Z^s$ and a subset $I\subset \{1,\ldots,s\}$ with $\# I=q$, and it has vertices at the lattice points $l+\sum_{j\in J}E_j$ for different $J\subset I$. One also takes a weight function $w:\Z^s\to \Z$ bounded below, and for each cube $\square=(l,I)$ one defines $w(\square):=\max\{w(v), \ \mbox{$v$ vertex of $\square$}\}$. Then, for each integer $n\geq \min(w)$ one considers the simplicial complex $S_n$ of $\R^s$, the union of all the cubes $\square$ (of any dimension) with $w(\square)\leq n$. Then the [*lattice cohomology associated with $w$*]{} is $\{\mathbb{H}^q(\Z^s,w)\}_{q\geq 0}$, defined by $\mathbb{H}^q(\Z^s,w):=\oplus _{n\geq \min(w)} H^q(S_n,\Z)$. Each $\mathbb{H}^q$ is graded (by $n$) and it is a $\Z[U]$–module, where the $U$–action consists of the restriction maps induced by the inclusions $S_n\hookrightarrow S_{n+1}$. Similarly, one defines the [*reduced cohomology associated with $w$*]{} by $\mathbb{H}_{{\rm red}}^q(\Z^s,w):=\oplus _{n\geq \min(w)} \widetilde{H}^q(S_n,\Z)$. In all our cases $\mathbb{H}_{{\rm red}}^q(\Z^s,w)$ has finite $\Z$–rank. The *normalized Euler characteristic* of $\mathbb{H}^*(\Z^s,w)$ is ${\rm eu}\,\mathbb{H}^*:= -\min(w)+\sum_{q\geq 0}\, (-1)^q\,{\rm rank}_\Z\, \mathbb{H}^q_{{\rm red}}$. Formally, we also set ${\rm eu}\,\mathbb{H}^0:=-\min(w)+{\rm rank}_\Z\, \mathbb{H}^0_{{\rm red}}$. Given a negative definite plumbing graph $G$ of a $\mathbb{Q}HS^3$ 3-manifold $M$ and a representative $l' \in L'$ of an element $[l'] = h \in H $, one works with the lattice $L = L_G = \mathbb{Z}\langle E_v \rangle_{v \in \mathcal{V}(G)}$ and weight function $L \ni l \mapsto -\frac{1}{2}(l, l + k_G + 2l')$. The cohomology theory corresponding to this weight function is denoted by $\mathbb{H}^{\ast}(G; k_G+2l')$. If for $h \in H$ we choose the minimal representative $r_h\in L'$, then the cohomology theory $\mathbb{H}^{\ast}(G; k_G+2r_h)$ is in fact an invariant of the pair $(M,h)$ (*i.e.* it does not depend on the plumbing representation) and thus can be denoted by $\mathbb{H}_h^{\ast}(M)$. It is proven in [@NemSW] that for any $l'\in L'$ and $h=[l']\in H$ $$\label{eq:LCoh} \mathfrak{s}_{l'}(G) = \textrm{eu\ } \mathbb{H}^{\ast}(G;k_G+2l') \ \ \mbox{and} \ \ \mathfrak{s}_h(M) = \textrm{eu\ } \mathbb{H}_h^{\ast}(M).$$ Lattice cohomology of integral surgeries. ----------------------------------------- [ss:intsurg]{} The lattice cohomology of integral surgeries ($q=1$) was treated in [@NSurg; @Npq; @NR; @BodNem]. We use the notation $p/q = d \in \mathbb{Z}$. Clearly $u = u'$. Let $\Delta_j(t)$ be the Alexander polynomial of the algebraic knot $K_j$ normalised by $\Delta_j(1)=1$. Let $\delta_j$ be the Seifert genus of $K_j\subset S^3$ (or, the delta–invariant of the corresponding plane curve singularity), and write $\delta:=\sum_{j= 1}^\nu \delta_j$. Set also $\Delta(t) = \prod_{j=1}^\nu\Delta_j(t)$ and write it in a form $\Delta(t) = 1 + \delta(t-1) + (t-1)^2Q(t)$ with $Q(t) = \sum_{i=0}^{2\delta-2}q_it^i$. Note that $Q(1)=\Delta''(1)/2$. \[ex:SIS\] The case of integral surgeries is especially important in singularity theory, since the links of *superisolated singularities* (see [@Lue; @LMN]) are of this type. They appear as follows. Let $f \in \mathbb{C}[x,y,z]$ be an irreducible homogeneous polynomial of degree $d$ such that its zero set in $\mathbb{CP}^2$ is a rational cuspidal curve; *i.e.* $C = \{ f=0 \}$ is homeomorphic to $S^2$ and all the singularities of $C$ are locally irreducible. Let their number be $\nu$. Assume that there is no singular point on the projective line given by $z=0$. Then the equation $f(x,y,z) + z^{d+1} = 0$ in $(\mathbb{C}^3,0)$ determines an isolated complex surface singularity with link homeomorphic to $S^3_{-d}(K)$, where $K$ is the connected sum of algebraic knots given by the local topological types of the singularities on $C$. In this case, by genus formula, $(d-1)(d-2)=2\delta$, a relation which connects $K$ with $d$. However, we can take the surgery (and plumbed) manifold $M = S^3_{-d}(K)$ for any $K$ and with arbitrary $d>0$, even without the ‘analytic compatibility’ $(d-1)(d-2)=2\delta$. Next we recall some results on lattice cohomology, which will be combined with the above proved CAP. (They will be very useful in fast computations of examples in the next section.) In the next general discussion the identity $(d-1)(d-2)=2\delta$ will not be assumed. Write $s_h := hE_u^{\ast}$, then $r_h = \{ hE_u^{\ast} \}=\{s_h\}$, and set also $c_h := \chi(r_h) - \chi(s_h)$. From [@NR Theorem 7.1.1] (cf. also [@BodNem Theorem 3.1.3]) we know that $$\mathfrak{s}_{s_h}(G) = \textrm{eu\ } \mathbb{H}^{\ast}\left(S^3_{-d}(K);k_G+2s_h\right) = \sum_{\substack{n \equiv h (\textrm{mod\ } d)\\ 0\leq n \leq 2\delta-2}} q_n.$$ By the surgery (‘cut-and-paste’) formula [@NB Theorem 1.0.1] one has $$\mathfrak{s}_{s_h}(G) = \mathcal{H}^{\textrm{pol}}_{u,h}(1) + \sum_{j=1}^{\nu} \mathfrak{s}_{R_j(s_h)}(G_j).$$ Since $R_j(s_h) = 0$ and $\mathfrak{s}_{0}(G_j) = 0$ (cf. Prop. \[prop:ratgraph\]) we get $\mathfrak{s}_{s_h}(G) = \mathcal{H}^{\textrm{pol}}_{u,h}(1)$, hence $$\mathcal{H}^{\textrm{pol}}_u(1) = \sum_{h=0}^{p-1} \mathfrak{s}_{s_h}(G)=\sum_{n=0}^{2\delta-2} q_n = Q(1).$$ This is related with the invariants $\mathfrak{s}_h(G)$ as follows. From (\[eq:LCoh\]) and (\[eq:difsh\]) $$\sum_{h=0}^{d-1} \mathfrak{s}_h(G) = \sum_{h=0}^{d-1} \mathfrak{s}_{s_h}(G)+ \sum_{h=0}^{d-1} (\chi(r_h) - \chi(s_h)) = \sum_{n=0}^{2\delta-2} q_n + \sum_{h=0}^{d-1} c_h=Q(1)+ \sum_{h=0}^{d-1} c_h.$$ The identity $p_g(\{g_j=0\})=\sum_h\chi_j(-\lfloor h\cdot (f_j)/d \rfloor)$ from Claim \[cl:susppg\] has the addendum $$\sum_{j=1}^{\nu} \chi_j (- \lfloor h\cdot (f_j)/d \rfloor) = \chi(r_h)-\chi(s_h)= c_h.$$ Indeed, $s_h=hE^*_u = hE_u/d + \sum_j h(f_j)/d = hE_u/d + \sum_j h E_{u_j}^{\ast,j}$, $r_h=hE_u/d+\sum_j\{h(f_j)/d\}$, $s_h-r_h= \sum_j\lfloor h(f_j)/d\rfloor$, hence $(s_h,s_h-r_h)=0$. Therefore $\sum_j\chi_j (- \lfloor h\cdot (f_j)/d \rfloor)=\chi(r_h-s_h)=\chi(r_h)-\chi(s_h)$. In this way, $\sum_{h=0}^{d-1} \mathfrak{s}_h(G)$ can be computed easily. Examples and applications ========================= [s:exandapp]{} \[ex:mainex\] Consider the plumbing graph of a superisolated singularity corresponding to a curve of degree $d = 8$ with three singular points whose knots $K_1, K_2, K_3$ are the torus knots of type $(6,7), (2,9), (2,5)$, respectively. {width="12cm"} One computes that $$\mathcal{H}_{u}(t) = (1-t)\cdot \frac{1-t^{42}}{(1-t^6)(1-t^7)}\cdot \frac{1-t^{10}}{(1-t^2)(1-t^5)}\cdot \frac{1-t^{18}}{(1-t^2)(1-t^9)} = \frac{\Delta_1(t)\Delta_2(t)\Delta_3(t)}{(1-t)^2}$$ and $\mathcal{H}^{\textrm{pol}}_u(1) = 293$. Correspondingly, $\sum_{h=0}^{7} \mathfrak{s}_{s_h}(G) = Q(1) = 293$. One also computes that $\sum_{h=0}^{7}c_h = 34$. Therefore, $\sum_{h=0}^{7}\mathfrak{s}_{h}(G) = 293 + 34 = 327$. After computing the graph $\Gamma$ of the UAC, we have $J = \mathbb{Z}_7 \times \mathbb{Z}_9 \times \mathbb{Z}_5$ and (setting $s = t^{7\cdot 9\cdot 5}$ and after summation $\Sigma_*$ over $\zeta_1 \in \mathbb{Z}_7$, $\zeta_2 \in \mathbb{Z}_9$, $\zeta_3 \in \mathbb{Z}_5$, where $\mathbb{Z}_l = \{e^{\frac{2\pi i m}{l}} \}_m$ are cyclic groups), the rational function $\mathcal{F}_{w,0}(t)$ equals $$\frac{1-s}{7\cdot 9\cdot 5}\cdot \sum_* \frac{(1-s^{21})^2}{(1-s^7)(1-\zeta_1^2s^3)(1-\zeta_1^{-2}s^3)}\frac{(1-s^9)}{(1-\zeta_2^5s)(1-\zeta_2^{-5}s)}\frac{(1-s^5)}{(1-\zeta_3s)(1-\zeta_3^{-1}s)} =$$ $$= \frac{\Delta_{1,\Gamma}(s)\Delta_{2,\Gamma}(s)\Delta_{3,\Gamma}(s)}{(1-s)^2}.$$ Then $\mathcal{F}_{w,0}(t) = \mathcal{H}_{u}(s)$ holds indeed with $s = t^{7\cdot 9\cdot 5}$. Correspondingly, $$\mathfrak{s}_0(\Gamma) = \mathcal{F}^{\textrm{pol}}_{w,0}(1) + \sum_{j=1}^{3}p_g(\{g_j=0\}) = 293 + 34 = 327.$$ \[rem:counter\] We wish to emphasize that the covering additivity property of $\mathfrak{s}$ is not true in general, not even when restricting ourselves to integral surgeries along algebraic knots in integral homology spheres (instead of $S^3$). This is shown by the next example (motivated by [@NemnicIII Remark 6.8.(2)]; the arrowhead of that graph is replaced by the $-8$ vertex below). (300,100)(20,-10) (84,72) (60,48) (84,48) (60,24) (84,24) (84,0) (108,36) (132,36) (84,48)[(0,1)[24]{}]{} (60,48)[(1,0)[24]{}]{} (84,48)[(2,-1)[24]{}]{} (60,24)[(1,0)[24]{}]{} (84,24)[(0,-1)[24]{}]{} (84,24)[(2,1)[24]{}]{} (108,36)[(1,0)[24]{}]{} (74,72)[(0,0)[$-5$]{}]{} (50,48)[(0,0)[$-2$]{}]{} (94,53)[(0,0)[$-1$]{}]{} (50,24)[(0,0)[$-2$]{}]{} (84,32)[(0,0)[$-1$]{}]{} (74,0)[(0,0)[$-5$]{}]{} (108,44)[(0,0)[$-7$]{}]{} (132,44)[(0,0)[$-4$]{}]{} (150,36)[(1,0)[24]{}]{} (160,45)[(0,0)[UAC]{}]{} (200,36) (224,36) (248,36) (272,36) (224,12) (248,12) (200,36)[(1,0)[72]{}]{} (224,12)[(0,1)[24]{}]{} (248,12)[(0,1)[24]{}]{} (200,44)[(0,0)[$-2$]{}]{} (224,44)[(0,0)[$-1$]{}]{} (248,44)[(0,0)[$-4$]{}]{} (272,44)[(0,0)[$-8$]{}]{} (214,12)[(0,0)[$-5$]{}]{} (238,12)[(0,0)[$-2$]{}]{} If we replace the $(-8)$–vertex of $G$ by an arrowhead (representing a knot $K$) we get an integral homology sphere $\mathfrak{S}^3$, the corresponding knot has $m_{u_1}=6$, hence $M(G)=\mathfrak{S}^3_{-2}(K)$, and $G$ has determinant $2$. One computes that $\mathfrak{s}_0(G)+\mathfrak{s}_1(G)= 15 + 14 = 29$, while $\mathfrak{s}_0(\Gamma) = 21$. In fact, when trying to copy the proof of Theorem \[thm:main\], one finds that neither the polynomial identity \[lem:pol\] holds (this is why the example was present in [@NemnicIII Remark 6.8.(2)]), nor is $\Gamma \backslash w$ of suspension type (satisfying the SWIC). On the other hand, there are facts suggesting that the CAP of $\mathfrak{s}$ can hold in more general settings. Indeed, as we indicated in Claim \[cl:pgadd\], if for a given $M$ one can find a surface singularity with link $M$ such that the EqSWIC holds for the singularity $(X,0)$ and the SWIC holds for its UAC, then the additivity of $\mathfrak{s}$ holds automatically. (Eq)SWIC was verified for many analytic structures, whose links are not of surgery type. On the other hand, the family of superisolated singularities is the main source of counterexamples for SWIC (and this was one of the motivations to test CAP for them). Independently of any analytic argument, one can also find purely topological examples for which CAP still works (and in which cases not only that we cannot verify the presence of EqSWIC/SWIC, but we cannot even identify any specific analytic structure on the topological type, or on certain special subgraphs). Here is one (for which the assumptions of Theorem \[thm:main\] do not hold either). (300,100)(20,-10) (84,72) (36,48) (60,48) (84,48) (36,24) (60,24) (84,24) (84,0) (108,36) (132,36) (84,48)[(0,1)[24]{}]{} (36,48)[(1,0)[24]{}]{} (60,48)[(1,0)[24]{}]{} (84,48)[(2,-1)[24]{}]{} (36,24)[(1,0)[24]{}]{} (60,24)[(1,0)[24]{}]{} (84,24)[(0,-1)[24]{}]{} (84,24)[(2,1)[24]{}]{} (108,36)[(1,0)[24]{}]{} (74,72)[(0,0)[$-2$]{}]{} (36,56)[(0,0)[$-4$]{}]{} (60,56)[(0,0)[$-3$]{}]{} (94,53)[(0,0)[$-3$]{}]{} (36,32)[(0,0)[$-4$]{}]{} (60,32)[(0,0)[$-3$]{}]{} (84,32)[(0,0)[$-3$]{}]{} (74,0)[(0,0)[$-2$]{}]{} (108,44)[(0,0)[$-1$]{}]{} (132,44)[(0,0)[$-16$]{}]{} (150,36)[(1,0)[24]{}]{} (160,45)[(0,0)[UAC]{}]{} (200,36) (224,36) (248,36) (272,36) (296,36) (248,12) (272,12) (200,36)[(1,0)[96]{}]{} (248,12)[(0,1)[24]{}]{} (272,12)[(0,1)[24]{}]{} (200,44)[(0,0)[$-4$]{}]{} (224,44)[(0,0)[$-3$]{}]{} (248,44)[(0,0)[$-3$]{}]{} (272,44)[(0,0)[$-1$]{}]{} (296,44)[(0,0)[$-32$]{}]{} (238,12)[(0,0)[$-2$]{}]{} (262,12)[(0,0)[$-2$]{}]{} One verifies that $\det(G)=2$, and $\mathfrak{s}_0(G)+ \mathfrak{s}_1(G)= 147 + 132 = 279 = \mathfrak{s}_0(\Gamma)$. This raises the interesting question to find the precise limits of the CAP. \[rem:END1\] The lattice cohomology plays an intermediate role connecting the analytic invariants of a normal surface singularity $X$ with the topology of its link $M = M(G)$. *E.g.*, one proves using [@Nlatnorm Prop. 6.2.2, Ex. 6.2.3, Thm. 7.1.3, 7.2.4] that for any $h$ one has $$p_g(X)_h\leq \textrm{eu\ } \mathbb{H}^0(M(G);k_G+2r_h).$$ Furthermore, for surgery manifolds $M(G) = S^3_{-d}(K)$ one has the vanishing $\mathbb{H}^q(M(G),k_G+2r_h)=0$ for $q\geq \nu$ ([@NR]). In particular, for superisolated singularities corresponding to unicuspidal rational plane curves ($\nu=1$) one has $$\label{eq:REM1} p_g(X)_h\leq \textrm{eu\ } \mathbb{H}^*(M(G);k_G+2r_h)=\mathfrak{s}_h(M).$$ Therefore, for $M=S^3_{-d}(K)$ with $\nu=1$, if the SWIC holds for the UAC $(Y,0)$, that is, if $p_g(Y)=\mathfrak{s}_0(\Sigma)$, then this identity, the CAP and (\[eq:REM1\]) implies $p_g(X)_h=\mathfrak{s}_h(M)$ for any $h$, that is, the EqSWIC for $(X,0)$. This is important for the following reason: for superisolated singularities we do not know (even at conjectural level) any candidate (either topological or analytic!) for their equivariant geometric genera. It is not hard to verify that $p_g(X)=d(d-1)(d-2)/6$, but no formulas exist for $p_g(X)_h$, and no (topological or analytic) prediction exists for $p_g(Y)$ either. \[ex:eq\] Set $\nu = 1$, $d = 4$, and let $K_1$ be the $(3,4)$ torus knot. This can be realized by the superisolated singularity $zx^3+y^4+z^5=0$. In this case $M = S^3_{-4}(K_1)$. One verifies that $\sum_{h=0}^{3} \mathfrak{s}_h(M) = 9 = \mathfrak{s}_0(\Sigma)$ correspondingly to Theorem \[thm:main\]. On the other hand, the UAC $(Y,0)$ of the *singularity* is the Brieskorn singularity $x^3+y^4+z^{16}=0$, whose geometric genus is $p_g(Y) = 9$ too. Hence, by the above remark, $p_g(X)_h=\mathfrak{s}_h(M)$ for any $h$. \[rem:END2\] (Continuation of \[rem:END1\].) It is interesting that we have two sets of invariants, an analytic package $(\,\{p_g(X)_h\}_h, p_g(Y)\,)$ and a topological one $(\, \{\mathfrak{s}_h(M)\}_h,\mathfrak{s}_0(\Sigma)\,)$, and both of them satisfy the additivity property. Nevertheless, in some cases, they do not agree. For example, if $p_g(Y)<\mathfrak{s}_0(\Sigma)$, then by (\[eq:REM1\]) necessarily at least one of the inequalities in (\[eq:REM1\]) is strict. In particular, for both topological and analytical package the additivity property is stable, it is never damaged, but the equality of the two packages in certain cases fails. \[ex:ineq\] Set again $\nu = 1$ and $d = 4$, but this time let $K_1$ be the $(2,7)$ torus knot. As usual $M = S^3_{-4}(K_1)$. By a computation $\sum_{h=0}^{3} \mathfrak{s}_h(M) = 10 = \mathfrak{s}_0(\Sigma)$. A suitable superisolated singularity is given by $(zy-x^2)^2-xy^3+z^5=0$. By [@LMN] (the end of section 4.5.) the universal abelian cover $Y$ satisfies the strict inequality $p_g(Y) < 10$. Therefore, $p_g(X)_h < \mathfrak{s}_h(M)$ for at least one $h$ (in fact, not for $h=0$). [9]{} Agol, I.; Groves, D.; Manning, J.: The virtual Haken conjecture, *Documenta Mathematica*, **18** (2013), 1045–1087. Barth, W.; Peters, C.; Van de Ven, A.: Compact complex surfaces, Ergebnisse der Mathematik und ihrer Grenzgebiete, 3. Folge, Band [**4**]{}, A Series of Modern Surveys in Mathematics, Springer-Verlag 1984. Bodnár, J.; Némethi, A.: Lattice cohomology and rational cuspidal curves, *Preprint*, arXiv:1405.0437, (2014). Borodzik, M.; Livingston, Ch.: Heegaard–Floer homologies and rational cuspidal curves, *Preprint*, arXiv:1304.1062, (2013). Braun, G.; Némethi, A.: Surgery formula for the Seiberg–Witten invariants of negative definite plumbed $3$-manifolds, *Journal für die Reine und Angewandte Mathematik*, **638** (2010), 189–208. Brieskorn, E.; Knörrer, H.: Plane Algebraic Curves, Birkhäuser, Boston, 1986. Eisenbud, D.; Neumann, W.: Three-dimensional Link Theory and Invariants of Plane Curve Singularities, *Ann. of Math. Studies 110, Princeton University Press, Princeton*, (1985). Fernández de Bobadilla, J.; Luengo, I.; Melle-Hernández, A.; Némethi, A.: On rational cuspidal curves, open surfaces and local singularities, *Singularity theory, Dedicated to Jean-Paul Brasselet on His 60th Birthday, Proceedings of the 2005 Marseille Singularity School and Conference*, (2007), 411–442. Fernández de Bobadilla, J.; Luengo, I.; Melle-Hernández, A.; Némethi, A.: On rational cuspidal projective plane curves, *Proc. London Math. Soc.*, **92** (1) (2006), 99–138. Hom, J.; Karakurt, Ç.; Lidman, T.: Surgery obstructions and Heegaard Floer homology, *Preprint*, arXiv:1408.1508, (2014). Kollár, J.: Shafarevich Maps and Automorphic Forms, *Princeton University Press, Princeton, New Jersey*, (1995). Lickorish, W. B. R.: A representation of orientable combinatorial 3-manifolds, *Annals of Mathematics*, **76** (2) (1962), 531–540. Luengo, I.: The $\mu$-constant stratum is not smooth, *Inventiones mathematicae*, **90** (1) (1987), 139–152. Luengo, I.; Melle-Hernandez, A.; Némethi, A.: Links and analytic invariants of superisolated singularities, *Journal of Algebraic Geometry*, **14** (3) (2005), 543–566. Némethi, A.: Graded roots and singularities, *Singularities in Geometry and Topology, World Sci., Hackensack, NJ*, (2007), 394–463. Némethi, A.: Invariants of normal surface singularities, *Proceedings of the Conference: Real and Complex Singularities, San Carlos, Brazil, August 2002*; *Contemporary Mathematics*, **354** (2004), 161–208. Némethi, A.: Lattice cohomology of normal surface singularities, *Publ. RIMS. Kyoto Univ.*, **44** (2008), 507–543. Némethi, A.: On the Heegaard Floer homology of $S^3_{-d}(K)$ and unicuspidal rational plane curves, [*Fields Institute Communications*]{}, Vol. [**47**]{}, 2005, 219-234; “Geometry and Topology of Manifolds”, Eds: H.U. Boden, I. Hambleton, A.J. Nicas and B.D. Park. Némethi, A.: Resolution graph of some surface singularities, I. Cyclic coverings, *Proceedings of the AMS Conference, San Antonio*, (1999). *Contemporary Mathematics*, **266**, *Singularities in Algebraic and Analytic Geometry, AMS*, (2000), 89–128. Némethi, A.: The cohomology of line bundles of splice quotient singularities, *Advances in Mathematics* **229** (4) (2012), 2503–2524. Némethi, A.: The Seiberg-Witten invariants of negative definite plumbed 3-manifolds, *Journal of EMS*, **13** (4) (2011), 959–974. Némethi, A.: The signature of $f(x,y) + z^n$, *Proceedings of Real and Complex Singularities, (C.T.C Wall’s 60th birthday meeting), Liverpool (England), August 1996, London Math. Soc. Lecture Notes Series*, **263** (1999) 131–149. Némethi, A.; Nicolaescu, L.: Seiberg-Witten invariants and surface singularities I., *Geometry & Topology*, **6** (1) (2002), 269–328. Némethi, A.; Nicolaescu, L.: Seiberg-Witten invariants and surface singularities II.: Singularities with good $\mathbb{C}^{\ast}$-action, *Journal of the London Mathematical Society*, **69** (3) (2004), 593–607. Némethi, A.; Nicolaescu, L.: Seiberg-Witten invariants and surface singularities III.: splicings and cyclic covers, *Selecta Mathematica*, **11** (3-4) (2006), 399–451. Némethi, A.; Román, F.: The lattice cohomology of $S^3_{-d}(K)$, *Zeta Functions in Algebra and Geometry, Contemporary Mathematics*, **566** (2012), 261–292. Némethi, A.; Sigur[$\eth$]{}sson, B.: The geometric genus of hypersurface singularities, *Preprint*, arXiv:1310.1268, (2013), to appear in [*JEMS*]{}. Ozsváth, P. S.; Szabó, Z.: Holomorphic disks and topological invariants for closed three-manifolds, *Annals of Mathematics*, (2004), 1027–1158. Ozsváth, P. S.; Szabó, Z.: Holomorphic disks and three-manifold invariants: properties and applications, *Annals of Mathematics*, (2004), 1159–1245. Wallace, A. H.: Modifications and cobounding manifolds, *Canad. J. Math*, **12** (1960), 503–528. Wise, D. T.: The structure of groups with a quasiconvex hierarchy, *Preprint*, (2011). Witten, E.: Monopoles and four-manifolds, *Mathematical Research Letters*, **1** (1994), 769–796.

ArXiv

--- bibliography: - 'loop.bib' --- PUPT-2051\ hep-th/0210093 [Double-trace operators and]{} 0.5cm [one-loop vacuum energy in AdS/CFT]{} [Steven S. Gubser and Indrajit Mitra]{} $$\seqalign{\span\TL & \span\TT}{ & Joseph Henry Laboratories, Princeton University, Princeton, NJ 08544 }$$ [Abstract]{} We perform a one-loop calculation of the vacuum energy of a tachyon field in anti de-Sitter space with boundary conditions corresponding to the presence of a double-trace operator in the dual field theory. Such an operator can lead to a renormalization group flow between two different conformal field theories related to each other by a Legendre transformation in the large $N$ limit. The calculation of the one-loop vacuum energy enables us to verify the holographic c-theorem one step beyond the classical supergravity approximation. October 2002 Introduction {#Introduction} ============ The AdS/CFT correspondence [@Malda; @GKP; @WittenAdS] (for reviews see [@MAGOO; @DHokerDan]) relates a $d$-dimensional quantum field theory to a $(d+1)$-dimensional gravitational theory, the most notable example being ${\cal N}=4$, $d=4$ super-Yang-Mills theory and type IIB string theory on $AdS_5 \times S^5$. Most of the checks and predictions of this duality have been at the level of classical supergravity. It is particularly difficult to carry out meaningful loop computations in AdS, corresponding to $1/N$ corrections in the gauge theory, simply because the supergravity theory is highly non-renormalizable, and the Ramond-Ramond fields make computations in the string genus expansion unwieldy at best. The aim of this note is to obtain a simple one-loop result in AdS that is finite in any dimension. The result is an expression for the difference of the vacuum energies that arises from changing boundary conditions on a tachyonic scalar field with mass in a particular range. The inspiration for this computation came from Witten’s treatment [@WittenMulti] of multi-trace deformations of the gauge theory lagrangian and their dual descriptions in asymptotically anti-de Sitter space. Such a dual description was also discussed in [@SecMulti]; however, our treatment will follow [@WittenMulti] more closely. Earlier work describing the same gauge theory deformations in terms of non-local terms in the string worldsheet action appeared in [@ABSMulti; @ABSMultiTwo]. To be definite, suppose one were to add to the gauge theory lagrangian a term ${f \over 2} {\cal O}^2$ where ${\cal O}$ is a single trace operator with dimension $3/2$, dual to a scalar field $\phi$ whose mass satisfies $m^2 L^2 = -15/4$.[^1] The coefficient $f$ has dimensions of mass, so ${f \over 2} {\cal O}^2$ is a relevant deformation, and there is a renormalization group (RG) flow starting from a UV fixed point where $f=0$. The endpoint of this flow is, plausibly, an IR fixed point whose correlators are related to those of the original $f=0$ theory, in the large $N$ limit, by a Legendre transformation in a manner explained in [@kwTwo].[^2] In particular, the scalar that was for $f=0$ related to the operator ${\cal O}$ of dimension $3/2$, is at the IR fixed point related to an operator $\tilde{\cal O}$ of dimension $5/2$. How is all this reflected in AdS? According to [@WittenMulti], the addition of ${f \over 2} {\cal O}^2$ amounts to specifying particular linear boundary conditions on the scalar $\phi$ at the boundary of AdS. At the classical level, these boundary conditions are consistent with the original $AdS_5$ solution with $\phi=0$. Superficially, this looks like a puzzle, since we were expecting an RG flow. In fact, conformal invariance is violated by the ${\cal O}^2$ deformation, but at leading order in $N$ its effects are restricted to certain correlators that we will describe in section \[Multi\]. The crux of the matter is that it is impossible to satisfy the boundary conditions on $\phi$ with a $SO(4,2)$-invariant bulk-to-bulk propagator, except when $f=0$ or $\infty$. This gives rise to one loop effects that cause deviations from $AdS_5$. Although we will not obtain the full one-loop corrected solution corresponding to an RG flow due to the ${f \over 2} {\cal O}^2$ deformation, we will consider its endpoints and perform a one-loop supergravity check of the c-theorem. This “theorem,” conjectured in four dimensions by Cardy [@Cardy] as a generalization of Zamolodchikov’s celebrated two-dimensional c-theorem [@Zamolodchikov], has been shown to follow from AdS/CFT at the level of classical supergravity provided the null energy condition holds [@gppzOne; @fgpwOne] (see also [@Alvarez:1998wr] for earlier work in this direction). The magnitude of the vacuum energy of $AdS_5$, measured in five-dimensional Planck units, is proportional to an appropriate central charge raised to the $-2/3$ power. So the vacuum energy should be more negative in the infrared than in the ultraviolet, and at the classical level, that is what is shown in [@gppzOne; @fgpwOne] (actually, the arguments on the AdS are dimension-independent, though it is not entirely how to translate the “holographic” central charge into field theory language in the case of odd-dimensional CFT’s). At the quantum level, the arguments of [@gppzOne; @fgpwOne] have no force because it’s not clear that the null energy condition is valid or even relevant. So an explicit loop calculation is appropriate. All that is needed is the one-loop contribution of the scalar $\phi$ to the vacuum energy. This quantity is divergent, but the difference between imposing the two simple boundary conditions (described above as $f \to 0$ and $f \to \infty$) gives a finite result. The contributions of all other fields can be ignored because they do not change at the one loop level as one changes the boundary conditions on $\phi$. Also, because we only desire a one-loop vacuum amplitude, we may entirely ignore interactions of the scalar with other fields, and work simply with the free action where we work in mostly plus signature, so that the metric of $AdS_5$ on the Poincaré patch is For definiteness, our discussion has focused on $AdS_5$ and a scalar with a particular mass; however, the results we will obtain can be presented with considerable generality for $AdS_{d+1}$, as we will describe. For odd $d$, the formulas for the vacuum energy are much more complicated, and for the sake of efficiency we check the sign via numerics. The organization of the paper is as follows. In section \[Multi\] we briefly review the prescription of [@WittenMulti] for treating multi-trace operators, and we demonstrate that general boundary conditions are incompatible with $SO(4,2)$-invariance of the scalar propagator. In section \[OneLoop\] we compute the finite change in the one-loop vacuum energy discussed above, and make some remarks on the interpolating geometry connecting the two anti-de Sitter endpoints. We conclude in section \[Conclusions\] by extracting the prediction for the central charge, and observing that the c-theorem is obeyed. Multi-trace operators and scalar propagators {#Multi} ============================================ The proposal of [@WittenMulti] is a natural generalization of the original prescription for computing correlators [@GKP; @WittenAdS], and it should in principle be derivable from it: see [@Minces:2002wp] for a more precise discussion. Suppose one starts with the complete set ${\cal O}_a$ of independent, local, color-singlet, normalized, single-trace operators: for ${\cal N}=4$ super-Yang-Mills theory these would include, for example, ${1 \over N} \tr X_1 X_2$ and ${1 \over N} \tr F_{\mu\nu} \nabla_\rho \lambda_1$. The action can be written as $I = N^2 W({\cal O}_a)$ for some functional $W$, which for ${\cal N}=4$ super-Yang-Mills would be the integral of a linear function of those ${\cal O}_a$ which are Lorentz scalars. The general belief is that the ${\cal O}_a$ can be put into one-to-one correspondence with the quantum states of type IIB string theory in $AdS_5$.[^3] Restricting ourselves to scalars in $AdS_5$, we have the standard relation $\Delta_a (\Delta_a - d) = m_a^2 L^2$ relating the dimension of ${\cal O}_a$ to the mass of the field $\phi_a$. Writing the metric for the Poincaré patch of $AdS_5$ as we have boundary asymptotics for $\phi_a$ as follows: The prescription of [@WittenMulti] is to replace $W({\cal O}_a)$ by $W(\beta_a)$ and impose the following boundary conditions: The partition function of the gravitational theory in AdS, subject to the boundary conditions , is then supposed to equal the partition function of the gauge theory. The simplest non-trivial example is double trace operators: most simply, ${\cal O}^2$ where the scalar operator ${\cal O}$ has dimension $\Delta$ between ${d \over 2}-1$ and $d/2$. Precisely in this range, unitarity bounds are satisfied, and both power law behaviors in  are normalizable. Then $W$ includes a term ${f \over 2} \int d^d x \, {\cal O}^2$. This brings us back to the discussion initiated in the introduction: nonzero $f$ plausibly drives the field theory from a UV fixed point where the boundary conditions are $\alpha=0$ to an IR fixed point where the boundary conditions are $\beta=0$. Since these two fixed points will be the focus of section \[OneLoop\], let us introduce an additional convenient notation: $\Delta_+$ and $\Delta_-$ are the two solutions to $\Delta (\Delta - d) = m^2 L^2$, with $\Delta_-$ being the lesser of the two (and thus in the aforementioned range, from ${d \over 2}-1$ to $d/2$). Clearly $\Delta_+ = d-\Delta_-$. When $\Delta_- < d/2$, the addition of a trace-squared operator ${\cal O}^2$, where ${\cal O}$ has dimension $\Delta=\Delta_-$, is a relevant deformation, so conformal invariance must be broken in the gauge theory. The results of [@WittenMulti] for $d=4$ and $\Delta_-=2$ suggests that even when $\Delta_- = d/2$ there is a logarithmic RG flow. The simplest indication of the breaking of conformal invariance in supergravity is that the bulk-to-bulk propagator for the scalar $\phi$ dual to ${\cal O}$ cannot be $SO(4,2)$-invariant. We will now demonstrate this claim. The propagator in question can be defined as and it satisfies the equation of motion where $\square = g^{\mu\nu}\nabla_\mu\nabla_\nu$, and the delta function includes a $1/\sqrt{g}$ in its definition, so that for any continuous function $f(z)$. If the propagator is to respect $SO(4,2)$ invariance, it must be a function only of the geodesic distance $\sigma(z,z')$, which is known to be where $L$ is the radius of AdS. The only solutions to which are functions only of $\zeta$ are $G(z,z') = p G_{\Delta_-} + (1-p) G_{\Delta_+}$ where for any $\Delta$ (cf. [@Fronsdal; @BurgLut]),[^4] By keeping $z'$ fixed while $z$ approaches the boundary of AdS, it is straightforward to verify that for no choice of $p \in (0,1)$ and $f \in (0,\infty)$ does the propagator $G(z,z') = p G_{\Delta_-} + (1-p) G_{\Delta_+}$ satisfy the boundary conditions , which in our case amount to $\alpha = f\beta$. For $p=0$ and $f=0$ the boundary conditions are satisfied with $SO(4,2)$ invariance preserved, corresponding to a fixed point of RG where $\phi$ is dual to an operator ${\cal O}$ with dimension $\Delta_-$. Let us call this the $\Delta_-$ theory. And for $p=1$ and $f=\infty$ (formally speaking), again the boundary conditions are satisfied with $SO(4,2)$ invariance, and now $\phi$ corresponds to an operator $\tilde{\cal O}$ with dimension $\Delta_+$: this we will call the $\Delta_+$ theory. It was already remarked in [@WittenMulti] that a renormalization group flow should interpolate between the $\Delta_-$ theory in the UV and the $\Delta_+$ theory in the IR. This is in fact a somewhat subtle claim: why should we think that the RG flow initiated by adding ${f \over 2} {\cal O}^2$ ends up at a non-trivial IR fixed point? We can argue as follows:[^5] the Legendre transformation prescription of [@kwTwo] guarantees that the IR fixed point exists, at least in the large $N$ limit. The existence of a fixed point of RG is a generic phenomenon, so $1/N$ corrections should not spoil the claim, nor should they greatly alter the location of the fixed point in the space of possible couplings. Since a naive scaling argument (just looking at the dimension of $f$) tells us that the RG flow should end up at the desired IR fixed point if we ignore all $1/N$ corrections, it should be that [*some*]{} RG flow exists close to the approximately one we naively identified, ending at the non-trivial IR fixed point. A significant caveat to this reasoning is that AdS/CFT examples often (in fact, nearly always in the literature so far) have exactly marginal deformations. A [*line*]{} of fixed points of RG is [*not*]{} a generic phenomenon, and $1/N$ effects in the absence of supersymmetry generically could destroy such a line. Only one point could be left after $1/N$ effects are included; or, worse yet, only a point infinitely far out in coupling space could be left. Translated into supergravity terms, these remarks mean that the one-loop contribution to the potential could source the dilaton or other moduli, possibly leaving no extrema at finite values of the fields. If there are no such moduli in the first place (as perhaps one would expect for a truly [*generic*]{} non-supersymmetric quantum field theory with an AdS dual), then this caveat is not a problem. In practice, however, it is likely to interfere with constructing explicit string theory examples of the RG flow discussed in this paper. For the remainder of our discussion, we will ignore the caveat. Since the renormalization group flow is non-trivial, it is natural to expect that the supergravity geometry deviates from AdS. The surprise is that this does [*not*]{} happen classically. Roughly, this can be understood in field theory terms as a reflection of the fact that $n$-point functions involving only the stress energy tensor do not receive corrections at leading order in $N$.[^6] At subleading order in $N$, or at one-loop in supergravity, deviations from AdS must occur, simply because a one-loop diagram where the $SO(4,2)$-non-invariant scalar propagator closes upon itself must give rise to an effective potential that varies over spacetime. Entertainingly, there is no classical scalar field which is varying; rather, the variation in the potential arises on account of proximity to the boundary. This is in contrast to previously studied examples of RG flow in $AdS_5$ (for instance [@gppzOne; @fgpwOne]) where the flow is described in terms of scalars in the five-dimensional supermultiplet of the graviton with non-trivial dependence on radius. There should be a solution to the one-loop-corrected supergravity lagrangian interpolating between one asymptotically AdS region near the boundary, corresponding to the $\Delta_-$ UV fixed point, and a different one in the interior, corresponding to the $\Delta_+$ IR fixed point. For instance, one could require that the symmetries of ${\bf R}^{3,1}$ be preserved in the solution, which must then have the form where $A(r) \to r/L_\mp$ as $r \to \pm\infty$. (Another choice would be to require the symmetries of ${\bf S}^3 \times {\bf R}$, which should lead to a solution with the conformal structure of global AdS). We will not find the full interpolating solution, but we will explore some properties of its AdS endpoints. We will be particularly interested in the central charge of the CFT’s dual to the two endpoints. To the leading non-trivial order, these may be computed as a one-loop saddle-point approximation to the supergravity “path integral” (supposing that such an object exists), but without deforming the AdS background itself. One loop vacuum energy for the tachyon field {#OneLoop} ============================================ The full classical action that we wish to consider is Here $\Lambda_0$ is a negative constant. The scalar is subject to the boundary conditions As remarked previously, $AdS_{d+1}$ with $\phi=0$ and $1/L^2 = - {\Lambda_0 \over d(d-1)}$ is a classical solution to the equations of motion from , but we expect that once one-loop effects are accounted for, this solution is corrected to an interpolation between $AdS_{d+1}$ spaces in the UR and IR with slightly different radii. The one-loop scalar bubble diagram corrects the gravitational lagrangian by an amount $\delta {\cal L}$, where Our main computation will be to evaluate this correction in the unperturbed background. In principle, one could go on to find the interpolating geometry perturbatively in the small parameter $\kappa \Lambda_0^{(d-1)/2}$. This would require separating $\delta {\cal L}$ into contributions to the cosmological term and two- and four-derivative expression in the metric—a much more involved computation than simply evaluating  in the unperturbed background. For brevity, we will use the notation $V$ in preference to $\delta {\cal L}$ for the scalar self-energy , despite the fact that in the full background-independent form involves derivative terms as well as finite non-local terms. $V$ is divergent, but we assume that the action  is part of well-defined theory of quantum gravity (presumably, a compactification of string theory or M-theory), so that all loop divergences are canceled in some physical way, leaving only finite renormalization effects. It may be that in the full theory, $\Lambda_0$ is just the extremal value of a classical potential function of several scalars; if so, then we are operating on the understanding that the second derivative of this potential function with respect to $\phi$ vanishes at $\phi=0$ (that is, we’ve soaked up any such second derivative into what we call $m^2$ in ). In general, it is difficult to compute one-loop corrections in an effective theory without knowing precisely how the full theory cancels divergences. Results obtained for a chiral anomaly in supergravity [@Bilal:1999ph] for $AdS_5 \times S^5$ can be used to show that the central charge is corrected at one loop in supergravity, leading to $c \propto N^2-1$, as appropriate for $SU(N)$ super-Yang-Mills, rather than $c \propto N^2$ (the leading order result). Thus in this case, the difficulties were overcome. Our situation is more generic, in that we do not depend on supersymmetry or a special spectrum of operators. What we are nevertheless able to do is to determine the finite difference between $V$ in the case where $f=0$ in  and the case where $f=\infty$. This we will then translate into a change in the central charge as one flows from the UV (the $\Delta_-$ theory) to the IR (the $\Delta_+$ theory). What makes the computation clean is that at one loop, we do not have to worry about interactions of the scalar with other fields, and the only relevant diagram is the one where a single scalar propagator closes on itself, with no vertices. Vacuum energy in limiting regions of AdS {#Endpoints} ---------------------------------------- The computation of the one-loop contribution to the vacuum energy by a scalar in curved space, like in flat space, amounts to summing the logarithm of the eigenvalues of the Klein-Gordon operator. A more easily computable expression is obtained by expressing the result in terms of an integral of the Green’s function with respect to some parameter such as proper time or mass.[^7] All of this is quite standard, so we just write down the result, referring the reader to [@BDBook] pp. 156-158 for a derivation: if the propagator $G(z,z';m^2,f)$ is defined by (with the delta-function including a $\sqrt{g}$ factor as in ) together with boundary conditions , as discussed in section \[Multi\], then formally, and for the cases $f=0,\infty$, the fact that we can make the scalar propagator $SO(4,2)$ invariant means that $V$ will be independent of the position $z$.[^8] The formula  is problematic because for large masses, $G(z,z',\tilde{m}^2,0)$ diverges at the boundary of AdS. This is unusual: the typical situation for quantum field theory in curved spacetime is that quantities become well-defined in the limit where masses are much larger that the inverse radius of curvature. Thus, instead of using , a well-defined procedure is to integrate down to the Breitenlohner-Freedman bound which is the smallest mass possible with normalizable modes in AdS. Thus we obtain where $m^2_{BF} L^2 = -d^2/4$ is the Breitenlohner-Freedman bound. (For a derivation see the Appendix). It is possible to argue that $V(z;m^2_{BF},f)$ is the same for $f=0$ and $f=\infty$. Indeed, the eigenmodes for a tachyon of mass $m^2$ with boundary conditions specified by $f=0$ is given by $\omega = \Delta_- + \ell + 2n$ and that specified by $f = \infty$ is given by $\omega = \Delta_+ + \ell + 2n$ [@BF], where $\ell$ is the orbital angular momentum quantum number and $n$ is the radial quantum number. But for a scalar with mass saturating the BF bound, $\Delta_+ = \Delta_- = {d \over 2}$. So from a viewpoint of canonical quantization it seems inevitable that $V(z;m^2_{BF},0) - V(z;m^2_{BF},\infty) = 0$. We can argue further that for general $f$ the eigenfunctions would be a linear combination of those with $f=0$ and $f= \infty$. That would again imply that for $\Delta = {d \over 2}$, the eigenvalues are unchanged. So we conclude that the $V(z;m_{BF}^2,f) - V(z;m_{BF}^2,0) = 0$ for all values of $f$. Thus we are led to the formula that we will really use for computation: where $V_+ = V(z,m^2,\infty)$ and $V_- = V(z,m^2,0)$. We have used the fact that $G_{\tilde{\Delta}_+}(z,z')$, as defined in , is precisely $G(z,z';\tilde{m}^2,\infty)$, while $G_{\tilde{\Delta}_-}(z,z') = G(z,z,';\tilde{m}^2,0)$. In light of the argument of the previous paragraph, the terms outside the integral cancel. The advantage of  is that $G_{\tilde{\Delta}_+}(z,z)-G_{\tilde{\Delta}_-}(z,z)$ is finite, so that the final answer is also manifestly finite. We have confidence that no other finite renormalization effects can slip in to the calculation, because the only thing that changes between the $\Delta_-$ and $\Delta_+$ vacua is the boundary condition on $\phi$. As a warm-up let us first carry out the computation for $AdS_5$. To get the value of $G_{{\tilde \Delta}_+}(z,z) - G_{{\tilde \Delta_-}}(z,z)$ for coincident points one has to first express the Green’s functions in terms of the geodesic distance $\sigma$. From (\[ellZetaDef\]) we see that in terms of the variable $\zeta$ the geodesic separation is given by $\cosh({\sigma \over L}) = {1 \over \zeta}$ so we rewrite the propagator  in terms of $\sigma$ and then expand $i \left[G_{{\tilde \Delta}_+}(z,z) - G_{{\tilde \Delta}_-}(z,z) \right]$ in a power series in powers of ${\sigma \over L}$. The answer is finite and in the limit ${\sigma \over L} \to 0$, for $AdS_5$ we obtain the simple expression: The difference in the vacuum energies using (\[FinalVDiff\]) is therefore where in the second line we have used ${\tilde m}^2 L^2 = {\tilde \Delta} ({\tilde \Delta} - 4)$ and the fact that $\Delta_{BF} = {d \over 2} = 2$. Since $\Delta_- < 2$ we find that $V_+ - V_- < 0$, and therefore $c_- > c_+$ in agreement with the field theory prediction. It is straightforward to generalize this for any odd-dimensional anti de-Sitter spacetime because for $d$ even, the difference $i [G_{{\tilde \Delta}_+}(z,z) - G_{{\tilde \Delta}_-}(z,z) ]$ is quite simple in form. Before writing this down, for convenience, let us define $d \equiv 2k$ so that the spacetime is $AdS_{2k+1}$. In terms of $k$, $i \left[G_{{\tilde \Delta}_+}(z,z) - G_{{\tilde \Delta}_-}(z,z) \right]$ is: where $n_k = 2^k (2k-1)!!$. The difference in the vacuum energies is therefore where in the second line we have used ${\tilde m}^2 L^2 = {\tilde \Delta} ({\tilde \Delta} - d)$ and the fact that $\Delta_{BF} = {d \over 2} = k$. Shifting the variable of integration by introducing a new variable ${\tilde \nu} \equiv {\tilde \Delta}_- - k$, the integrand can be written down in a terms of the Pochhammer symbol $(a)_n = {{\Gamma(a+n)} \over {\Gamma(n)}}$: The factor $(-1)^k$ was nullified by an extra factor of $(-1)^k$ from the product. Assembling all of this, we finally have where we recall that $n_k = 2^k (2k-1)!!$. The lower limit of integration $\nu$ depends on the value of $\Delta_-$. Since $k \leq \Delta_- \leq k-1 $, the range of $\nu$ is $-1 \leq \nu \leq 0$. The function $(\nu)_k (-\nu)_k < 0$ for all $k$ and $-1 \leq \nu \leq 0$. So for any odd-dimension anti de-Sitter spacetimes we have shown that $V_+ - V_- < 0$. For even dimensional spacetimes, an analytic proof seems cumbersome, so we resorted to numerics. As an explicit example, figure \[figA\] shows a plot of $i (G_{d- \Delta} - G_{\Delta})$ as a function of $\Delta$ for several even-dimensional anti-de Sitter spacetimes. In each dimension, we’ve plotted the integrand of (\[FinalVDiff\]) for $d/2-1 < \Delta < d/2$. Since the integrand is always negative on this range, we conclude that $V_+ < V_-$ in accordance with the c-theorem intuition. This is also true for $d = 3$, and we believe it is true generally. Vacuum energy throughout AdS {#Throughout} ---------------------------- The results of the previous section were stated in terms of $V_+ - V_- = V(z;m^2,\infty) - V(z;m^2,0)$ (both terms were in fact independent of the position $z$ in AdS). Here we would like to investigate $V(z;m^2,f)$ for finite $f$. This quantity diverges, but $V(z;m^2,f)-V_-$ is finite. We will be able to verify the formulas which we consider intuitively obvious since ${f \over 2} {\cal O}^2$ is a relevant operator in the CFT, and therefore unimportant in the UV but important in the IR. As a first step, one needs the Green’s function for the scalar obeying mixed boundary conditions for all values of $f$ (not just the ones for $f=0$ and $f= \infty$ that we wrote down earlier). This would be needed to compute the vacuum energy contribution due to the bubble diagram. The one-loop corrected action would then induce corrections in the geometry which can be computed from the Einstein equations. Let us work in Euclidean $AdS$ to get the Green’s function $G_E (x,y;f)$ which we shall Wick rotate to obtain $G(x,y;f)$ in Minkowski signature. We shall follow the canonical method of obtaining Green’s functions. In Poincaré coordinates the scalar wave equation is where from now on we shall denote the radial direction by $x_0$ or $y_0$ and $\vec{x}$ is a vector with components along the $d$ remaining directions. The two linearly independent solutions to this equation are: $\phi_1 = x_0^2 e^{-i \vec{k}\cdot\vec{x}} I_{\nu}(kx_0)$ and $\phi_2 = x_0^2 e^{-i \vec{k}\cdot \vec{x} } I_{-\nu} (kx_0)$ where $\nu = \sqrt{m^2 L^2 + {d^2 \over 4}}$. In the notation of our previous sections, the Green’s function obeys the equation where we remind ourselves that the delta function includes a ${1 \over {\sqrt{g}}}$ in its definition. The right hand side is zero for $x_0 \neq y_0$, so we have The boundary behavior of the scalar we’re interested in is: $\phi(x_0, \vec{x}) = f \beta(\vec{x}) x_0^{{d \over 2} + \nu} + \beta(\vec{x}) x_0^{{d \over 2} - \nu}$. We choose our $\phi_1$ and $\phi_2$ so that they have the right boundary behavior and also require that the Green’s function not diverge in the bulk (large values of the radial coordinate $x_0$) for two non-coincident points. One convenient choice of $\phi_1$ and $\phi_2$ is: so that $\phi_1$ satisfies the boundary condition for small $x_0$ and $\phi_2$ is finite in the bulk. From the asymptotics of Bessel functions, we see that $\phi_1$ diverges as $x_0 \to \infty$ and $\phi_2$ diverges as $x_0 \to 0$. This forces us to set $A_2 = B_1 = 0$ in . The remaining two constants are determined by integrating  twice which gives us two conditions: (i) the Green’s function is continuous at $x_0 = y_0$ , and (ii) its radial derivative has a jump discontinuity of $1 \over x_0^{d-1}$ at $x_0 = y_0$. This yields where ${\cal W} [\phi_1 (y_0), \phi_2 (y_0)]$ is the Wronskian. For our choice of $\phi_1$ and $\phi_2$ the Wronskian is: so combining , , and  we obtain the Green’s function: for $x_0 < y_0$ and a similar expression for $x_0 > y_0$. In the above equation, $\kappa_E$ is the temporal component of momentum. Finally, we Wick rotate this component $ \kappa_E = i k$ to get the Green’s function in Minkowski signature: The integral for general values of $f$, $d$ and $\nu$ is hard. For $f=0$ and $f= \infty$ it can be evaluated and the result is an expression which is related to  by a quadratic hypergeometric transformation [@Muck]. A little bit more can be said about the radial dependence of the one-loop vacuum energy. This latter quantity depends on the Green’s function for coincident points $G(x,x;f)$. We saw before that this divergent quantity was best handled by subtracting out $G(x,x;0)$. The result is then finite: where $\tilde{f} = 2^{2 \nu} {{\Gamma(1+ \nu)} \over {\Gamma(1 - \nu)}} f x_0^{2 \nu}$ and $\tilde{k} = k x_0$. Note that the excess vacuum energy depends on the radial coordinate $x_0$ in the particular combination $f x_0^{2 \nu}$. In order to make any further progress, one would need to first compute the momentum integral and then integrate over $\nu$ to obtain the vacuum energy. We argued earlier that $V(x;m_{BF}^2,f) - V(x;m_{BF}^2,0) = 0$ for all values of $f$, so using   and  we have: where we remind ourselves that $\tilde{f} = 2^{2 \nu} {{\Gamma(1+ \nu)} \over {\Gamma(1 - \nu)}} f x_0^{2 \nu}$. The double integral is difficult to perform explicitly. However, it is not hard to show from  that $V(x;m^2,f)$ decreases monotonically as $f$ increases from $0$ to $\infty$. To see this we note that the integrand depends on $x_0$ only through $\tilde{f}$ and since the integrand is a monotonic function of $\tilde{f}$, clearly $V(x;m^2,f)$ decreases monotonically with increasing $f$. Conclusions {#Conclusions} =========== The upshot of section \[Endpoints\] was an evaluation of the change in the one-loop self-energy, $V_+ - V_-$, between the IR and UV endpoints of a holographic RG flow. We would now like to convert this into a change in the central charge of the dual field theory. In [@HennSken], the central charge was obtained by holographically computing the Weyl anomaly: on the field theory side, upon a conformal variation $g_{\mu\nu} \to e^{2\omega} g_{\mu\nu}$, where $W$ is the generating functional for connected Green’s functions. At the one-loop level, the prescription of [@GKP; @WittenAdS] asserts that $W$ is the classical supergravity action. The exact statement is that the partition functions of string theory and gauge theory coincide (subjected to boundary conditions and source terms in the usual way). In the calculation of [@HennSken], the supergravity action integral is evaluated with a radial cutoff, where the choice of radius amounts to a choice of metric within a conformal class. The supergravity lagrangian evaluates to a constant in AdS, and the central charge is proportional to this constant.[^9] All that we need to do in order to correct the central charge computation at one loop is to ask by how much the one-loop-corrected lagrangian differs from the tree-level lagrangian, when evaluated in AdS. The tree level and one-loop lagrangians will stand in the same ratio as the leading large $N$ central charge and its $1/N$-corrected counterpart. The tree level lagrangian is The calculation indicated by the discussion in the previous paragraph is where $\delta {\cal L} = -\sqrt{g} V$ is the one-loop correction to the lagrangian that we computed in section \[OneLoop\]. Because we are only able to compute $V$ up to an additive constant that is independent of boundary conditions, the only meaningful ratio that we can form is so that To check if $c_-$ is indeed greater than $c_+$, all that we have to show is that $V_+ < V_-$. But that is exactly what we saw above. As an example, in $AdS_5$, we obtain from  and  the result One can go further and translate the function $V(z;m^2,f)-V_-$ into a correction to the central charge whose scale dependence is monotonic. It is not clear how well-defined such a function can be on the supergravity side: because the bulk theory includes gravity, it has no local observables. Poetically, we would like to relate this to the fact that renormalization group effects in field theory are scheme-dependent—but it is difficult to make this precise. It would be interesting to see how the construction discussed in this paper might be realized as part of a compactification of string theory to four dimensions, along the lines of [@VerlindeCY; @Giddings:2001yu]. One of the most interesting questions in that context is one that we glossed over here: before considering the loop effects in supergravity, one generally expects a moduli space of vacua, and this statement probably translates into field theory terms as the existence of a line of fixed points. Mapping the lifting of moduli into field theory terms might at least gain us a restatement of the moduli problem in terms of the existence of isolated fixed points of the renormalization group. Acknowledgments {#acknowledgments .unnumbered} =============== This work was supported in part by DOE grant DE-FG02-91ER40671. We thank E. Silverstein and E. Witten for useful discussions. Appendix {#appendix .unnumbered} ======== In this appendix we shall sketch the derivation of . Our starting point is the familiar field theory result that the one-loop effective potential is We shall denote the Klein-Gordon operator $(- \square + m^2)$ by ${\hat K}(m^2,f)$ and as an operator, it is related to our definition of the Green’s function  by ${\hat G}(m^2,f) = - [{\hat K}(m^2,f)]^{-1}$. The representations of operators such as ${\hat G}(m^2,f)$ in an orthonormal basis shall be denoted by the obvious notation: $\langle z| {\hat G}(m^2,f) |z \rangle = G(z,z';m^2,f)$. In terms of the Green’s function, the effective potential is then We shall use the Schwinger proper-time formalism to evaluate this. One needs two simple operator relations both of which follow from the relation between ${\hat G}(m^2,f)$ and ${\hat K}(m^2,f)$ where $\gamma$ is the Euler’s constant. For the effective potential, we see from  that we need $\log[-{\hat G}(m^2,f)]$ which differs by a factor of $is$ from the integral representation of ${\hat G}(m^2,f)$ above. To proceed any further we need the DeWitt-Schwinger representation of the Green’s function (the reader is referred to [@BDBook] pg. 75 for a derivation) where $\eta(z,z')$ is one-half the proper distance between the points $z$ and $z'$, and $M(z,z') = - {\rm det} [\partial_\mu \partial_\nu \eta(z,z')]$. For our purposes we shall just need to use the fact that the only place where the mass appears is in the exponent and integrating with respect to $m^2$ will bring down an extra factor of $is$ that we need. So integrating both sides of  between two arbitrary masses $m_1^2$ and $m_2^2$ and using  we obtain In the usual treatment one chooses one of the masses to be infinite, but as we explained in the main text, this cannot be done here. Instead of integrating toward heavier masses, we integrate in the opposite direction down to the Breitenlohner-Freedman bound. Therefore, we set $m_1^2 = m_{{\rm BF}}^2$ and $m_2^2 = m^2$ in  and use (\[SimplerV\]) to get [^1]: Such a situation could arise in the theory dual to D3-branes at the tip of a conifold [@kwOne], where there are indeed dimension $3/2$ color singlet operators. [^2]: We will discuss further in section \[Multi\] the reasoning behind the claim that the flow ends at an IR fixed point, as well as some caveats. [^3]: There is considerable subtlety in this claim. It has been demonstrated that the Kaluza-Klein tower of supergravity modes in $AdS_5 \times S^5$ is in correspondence with the chiral primaries of ${\cal N}=4$ super-Yang-Mills and their descendants; and the duals of certain non-perturbative states have been found, such as dibaryons (see for example [@gkBaryon] and giant gravitons [@McGreevy:2000cw]. Evidence is growing that the operator-state map extends faithfully to excited string states (see for example [@bmn; @gkPolTwo]). Since the states in question can sometimes be extended across most of $AdS_5$ (as in [@gkPolTwo]), it is not entirely clear that a second quantized treatment in terms of local fields is appropriate; but this is scarcely relevant to the situation at hand, since extended states are very massive, and we’re interested only in tachyons. [^4]: The expression for $G(z,z')$ above differs by a sign from that in [@BurgLut; @Fronsdal] because the latter define the Green’s function as $- i G(z,z') = \langle 0| {\rm T} \phi (z) \phi (z') |0 \rangle$. [^5]: SSG thanks E. Silverstein for a discussion in which the following line of reasoning arose. [^6]: Correlation functions which [*do*]{} receive corrections at leading order in $N$ when ${f \over 2} {\cal O}^2$ is added to the lagrangian are precisely those which pick up contributions from factorized forms $\langle {\cal O} \ldots \rangle \langle {\cal O} \ldots \rangle$, where the dots indicate any arrangement of the operators involved in the original correlator. [^7]: For a different method of computing the effective potential based on the technique of Zeta-function regularization see [@Caldar]. [^8]: Actually, we have tucked an additional complication into our notation: $V$ is, more properly, minus the one-loop correction to the full gravitational lagrangian, and as such includes not just a scalar piece, but also terms depending on curvatures. For the central charge computation, as we shall explain, the relevant quantity is the sum of all these terms evaluated on AdS. [^9]: [*A priori*]{}, one might worry that boundary terms in the supergravity action also contribute to the central charge. That this does not happen depends on the circumstance, noted in [@HennSken], that the only log-divergent terms in the supergravity calculation arise from the integral of the bulk action.

ArXiv

--- abstract: 'The emergence of modern RGB-D sensors had a significant impact in many application fields, including robotics, augmented reality (AR) and 3D scanning. They are low-cost, low-power and low-size alternatives to traditional range sensors such as LiDAR. Moreover, unlike RGB cameras, RGB-D sensors provide the additional depth information that removes the need of frame-by-frame triangulation for 3D scene reconstruction. These merits have made them very popular in mobile robotics and AR, where it is of great interest to estimate ego-motion and 3D scene structure. Such spatial understanding can enable robots to navigate autonomously without collisions and allow users to insert virtual entities consistent with the image stream. In this chapter, we review common formulations of odometry and Simultaneous Localization and Mapping (known by its acronym SLAM) using RGB-D stream input. The two topics are closely related, as the former aims to track the incremental camera motion with respect to a local map of the scene, and the latter to jointly estimate the camera trajectory and the global map with consistency. In both cases, the standard approaches minimize a cost function using nonlinear optimization techniques. This chapter consists of three main parts: In the first part, we introduce the basic concept of odometry and SLAM and motivate the use of RGB-D sensors. We also give mathematical preliminaries relevant to most odometry and SLAM algorithms. In the second part, we detail the three main components of SLAM systems: camera pose tracking, scene mapping and loop closing. For each component, we describe different approaches proposed in the literature. In the final part, we provide a brief discussion on advanced research topics with the references to the state-of-the-art.' author: - Javier Civera - 'Seong Hun Lee[^1]' title: 'RGB-D Odometry and SLAM' --- This is the pre-submission version of the manuscript that was later edited and\ published as a chapter in *RGB-D Image Analysis and Processing*. DOI: <http://doi.org/10.1007/978-3-030-28603-3_6> SpringerLink: <http://link.springer.com/chapter/10.1007/978-3-030-28603-3_6>  Springer Nature Switzerland AG 2019 Cite as: BibTeX: ``` {frame="single"} @Inbook{Civera2019, author="Civera, Javier and Lee, Seong Hun", editor="Rosin, Paul L. and Lai, Yu-Kun and Shao, Ling and Liu, Yonghuai", title="RGB-D Odometry and SLAM", bookTitle="RGB-D Image Analysis and Processing", year="2019", publisher="Springer International Publishing", address="Cham", pages="117--144", isbn="978-3-030-28603-3", doi="10.1007/978-3-030-28603-3_6", url="https://doi.org/10.1007/978-3-030-28603-3_6"} ``` Introduction: SLAM, Visual SLAM and RGB-D Sensors ================================================= Visual Odometry and Visual Simultaneous Localization and Mapping – from here on referred to as their respective acronyms VO and VSLAM – are two tightly related topics that aim to extract 3D information from streams of visual data in real-time. Specifically, the goal of VO is to estimate the incremental motion (*i.e.*, translation and rotation) of the camera as it moves. The goal of Visual SLAM is more ambitious: To estimate a globally consistent map of the scene and the camera trajectory with respect to it. In the robotics research community, SLAM is considered as a fundamental capability for autonomous robots. See [@durrant2006simultaneous; @bailey2006simultaneous] for an illustrative tutorial covering the earliest approaches, and [@cadena2016past] for a recent survey outlining the state-of-the-art and the most relevant future directions. While the early pioneering works on SLAM mainly used laser scanners (*e.g.,* [@castellanos1999spmap]), the field rapidly pivoted to cameras for several reasons. Among them were the progress of computer vision algorithms and improved processors, as well as the camera’s low cost, size and power consumption. \[fig:orbslam2\]  \ Most visual SLAM methods have been traditionally based on low-level feature matching and multiple view geometry. This introduces several limitations to monocular SLAM. For example, a large-baseline motion is needed to generate sufficient parallax for reliable depth estimation; and the scale is unobservable. This can be partially alleviated by including additional sensors (*e.g.*, stereo cameras [@pire2017s], inertial measurement units (IMUs) [@concha2016visual], sonar [@engel2014scale]) or the prior knowledge of the system [@lee2018stability] or the scene [@tateno2017cnn]. Another challenge is the dense reconstruction of low texture areas [@concha2014manhattan]. Although recent approaches using deep learning (*e.g.*, [@bloesch2018codeslam; @zhou2018deeptam]) have shown impressive results in this direction, more research is needed regarding their cost and dependence on the training data [@facil2019camconvs]. RGB-D sensors provide a practical hardware-based alternative to the challenges and limitations mentioned above. Their availability at low cost has facilitated many robotics and AR applications in the last decade. Intense research endeavors have produced numerous robust algorithms and real-time systems. Figure \[fig:rgbd-slam\] shows several reconstruction examples from the state-of-the-art systems. Today, RGB-D cameras stand out as one of the preferred sensors for indoor applications in robotics and AR; and their future looks promising either on their own or in combination with additional sensors. In this chapter, we will cover several state-of-the-art RGB-D odometry and SLAM algorithms. Our goal is to focus on the basic aspects of geometry and optimization, highlighting relevant aspects of the most used formulations and pointing to the most promising research directions. The reader should be aware that, as a consequence of condensing a vast array of works and presenting the basics in a homogeneous and easy-to-follow manner, some individual works might present slight variations from the formulation presented here. In general, we sacrificed extending ourselves over particular details in favour of a clearer overview of the field. The rest of the chapter is organized as follows. Section \[sec:rgbdoslampipelines\] will give an overview on the most usual VO and VSLAM pipeline. Section \[sec:notation\] will introduce the notation used throughout the rest of the chapter. Section \[sec:motion\] will cover the algorithms for tracking the camera pose, Section \[sec:mapping\] the algorithms for the estimation of the scene structure, and Section \[sec:loopclosure\] the loop closure algorithms. Section \[sec:references\] will refer to relevant scientific works and research lines that were not covered in the previous sections. Finally, Section \[sec:conclusions\] contains the conclusions and Section \[sec:resources\] provides links to some of the most relevant online resources, mainly the state-of-the-art open-source software and public benchmark datasets. The Visual Odometry and SLAM pipelines {#sec:rgbdoslampipelines} ====================================== The pipelines of RGB-D Odometry and SLAM have many components in common. Here, we will give a holistic view of the building blocks of standard implementations, highlighting their connections and introducing the terminology. The seminal work of Klein and Murray [@klein2007parallel] proposed the architecture that is used in most visual odometry and SLAM systems nowadays. Such architecture was later refined in papers like [@strasdat2010scale; @mur2015orb; @engel2018direct] among others. Basically, the idea is to partition the processing into two (or more) parallel threads: one thread tracks the camera pose in real time at video rate, and the rest update several levels of scene representations at lower frequencies (in general, the larger and/or more complex the map, the lower the frequency of update). ![High-level overview of VO and VSLAM systems. $I_k$: $k^{th}$ RGB-D image, $\xi_k$: $k^{th}$ camera pose, $\mathcal{M}_L$ and $\mathcal{M}_G$: the local and the global map.[]{data-label="fig:pipeline"}](rgb_d_o_slam_odometry.pdf){width="90.00000%"} Fig. \[fig:pipeline\] illustrates a simple Tracking and Mapping architecture for RGB-D Odometry and SLAM that we will use in this chapter. The camera tracking thread estimates the camera motion $\mathbf{\xi}_k$ at time $k$ given the current frame $I_k$ and a local map $\mathcal{M}_L$. $\mathcal{M}_L$ is estimated from a set of keyframes summarizing the sequence. If SLAM is the aim, a globally consistent map $\mathcal{M}_G$ is estimated by means of loop closure and global optimization. In more detail: - [**Camera tracking:** ]{}The camera tracking thread estimates the incremental camera motion. The most simple approach is to use the frame-to-frame constraints (*e.g.*, [@kerl2013robust; @gutierrez2016dense]). This is in fact inevitable when bootstrapping the system from the first two views. However, after initialization, it is quite usual to use more than two views in order to achieve higher accuracy. In this case, the standard approach is to estimate the camera motion using map-to-frame constraints with respect to a local map built from the past keyframes (see the next paragraph). - [**Scene Mapping:** ]{} Mapping approaches vary, depending on the application. Volumetric mapping discretizes the scene volume into voxels and integrates the information from the RGB-D views (*e.g.*, [@newcombe2011kinectfusion; @whelan2015real]). Point-based mapping performs a nonlinear optimization of camera poses and points (*e.g.*, [@mur2017orb]). In the case of VO, the map is local and is estimated from a sliding window containing a selection of the last frames (*e.g.*[@wang2014rgbd; @engel2018direct]). In the case of VSLAM, the map is estimated from a set of keyframes representative of the visited places. - [**Loop Closing:** ]{} In both odometry and SLAM drift is accumulated in purely exploratory trajectories. Such drift can be corrected if a place is revisited, using approaches denoted as loop closure. First, the place is recognized by its visual appearance (loop detection), and then the error of the global map is corrected (loop correction) [@mur2017orb; @gutierrez2018rgbid; @whelan2015real]). Notation and Preliminaries {#sec:notation} ========================== Geometry and Sensor Model {#sec:geometry} ------------------------- We denote an RGB-D input as $I:\Omega \mapsto \mathbb{R}^4$, where $\Omega \subset \mathbb{R}^2$ is the image plane of width $w$ and height $h$. We represent the pixel coordinates as a 2D vector $\mathbf{p} = \left( u, v \right)^\top$ and the corresponding homogeneous coordinates as $\tilde{\mathbf{p}} = (\tilde{u}, \tilde{v}, \lambda)^\top$. Each pixel has RGB color and depth value, *i.e.*, $I\left( u, v \right) = \left( r, g, b, d \right)^\top$. The depth channel is denoted as $D:\Omega \mapsto \mathbb{R}$, and the access to it as $D\left( u, v \right) = d$. The Euclidean coordinates of a 3D point $k$ in some reference frame $i$ (be it a camera or the world reference) are denoted by $\mathbf{P}^i_k = \left( X^i_k, Y^i_k, Z^i_k \right)^\top$ or $\mathbf{\tilde P}^i_k = \left( \lambda X^i_k, \lambda Y^i_k, \lambda Z^i_k, \lambda \right)^\top$ in homogeneous coordinates. These two coordinates are related by the dehomogenization operation: $\mathbf{P}^i_k=\pi_\text{3D}(\mathbf{\tilde{P}}_k^i)$. Inversely, the homogenization is denoted by $\pi^{-1}_\text{3D}(\mathbf{P}^i_k) := \left( X^i_k, Y^i_k, Z^i_k, 1 \right)^\top$. ![The transformation of point $k$ from reference frame $i$ to camera reference frame $j$, and its projection onto the image plane using the pinhole camera model. []{data-label="fig:geometry_model"}](camera_model.pdf){width="60.00000%"} The pose of camera $j$ with respect to reference frame $i$ is defined by the transformation $\mathtt{T}_{ji} = \begin{bmatrix} \mathtt{R}_{ji} & \mathbf{t}_{ji} \\ \mathtt{0} & 1 \end{bmatrix} \in \ensuremath{\mathrm{SE}(3)}\xspace$, $\mathtt{R}_{ji}\in \ensuremath{\mathrm{SO}(3)}\xspace$, $\mathbf{t}_{ji} \in \mathbb{R}^3$. The rotation matrix $\mathtt{R}_{ji}$ and translation vector $\mathbf{t}_{ji}$ are defined such that the transformation of point $\mathbf{P}^i_k$ in reference frame $i$ to the $j^{th}$ camera reference frame is $$\mathbf{\tilde P}^j_k = \mathtt{T}_{ji} \mathbf{\tilde P}_k^i; \ \ \mathbf{P}^j_k = \mathtt{R}_{ji} \mathbf{P}^i_k + \mathbf{t}_{ji}. \label{eq:transfwtoj}$$ Likewise, $\tilde{\mathbf{P}}^i_k$ can be obtained from $\tilde{\mathbf{P}}^i_k$ and $\mathtt{T}_{ji}$ with the inverse operation: $$\mathbf{\tilde{P}}^{i}_k=\mathtt{T}_{ji}^{-1}\mathbf{\tilde{P}}^j_k; \ \ \mathbf{P}^{i}_k = \mathtt{R}_{ji}^\top\left(\mathbf{P}^j_k-\mathbf{t}_{ji}\right).$$ As illustrated in Fig. \[fig:geometry\_model\], we adopt the standard pinhole model for the projection onto the image plane. First, the 3D point $\mathbf{P}^i_k$ is transformed to the camera frame $j$ using equation \[eq:transfwtoj\]. The homogeneous coordinates of the projection in the image space are given by $$\label{eq:projection} \tilde{\mathbf{p}}^j_k = (\tilde{u}^j_k, \tilde{v}^j_k, \lambda)^\top=\mathtt{K}\mathbf{P}^j_k=\mathtt{K}(\mathtt{R}_{ji} \mathbf{P}^i_k + \mathbf{t}_{ji}) \quad \text{with} \quad \mathtt{K}=\begin{bmatrix} f_x & 0 & u_0 \\ 0 & f_y & v_0 \\ 0 & 0 & 1 \end{bmatrix},$$ where $\mathtt{K}$ is the calibration matrix containing the coordinates of the principal point $(u_0, v_0)^\top$ and the focal lengths $(f_x, f_y)=(fm_x, fm_y)$. Here, $(m_x, m_y)$ denotes the number of pixels per unit distance in image coordinates in the $x$ and $y$ directions. The pixel coordinates are finally obtained by dehomogenization: $\mathbf{p}^j_k=(u^j_k, v^j_k)^\top=\pi_\text{2D}(\tilde{\mathbf{p}}^j_k)=({\left.\tilde{u}^j_k\middle/\lambda\right.}, {\left.\tilde{v}^j_k\middle/\lambda\right.})^\top$. The inverse operation (*i.e.*, homogenization) is denoted by $\pi_\text{2D}^{-1}(\mathbf{p}^j_k):=(u^j_k, v^j_k, 1)^\top$. Now, let reference frame $i$ be another camera reference frame. Then, the *reprojection* of 2D point $\mathbf{p}^i_k$ in frame $i$ to frame $j$ is defined as the following three-step operation: 1. Backproject $\mathbf{p}^i_k$ with the measured depth $d^i_k$ to estimate the 3D point $\mathbf{P}^i_k$ in frame $i$: $$\label{eq:backprojection} {\mathbf{P}^{i}_k}'=d^i_k\frac{\mathtt{K}^{-1}\pi^{-1}_\text{2D} \left(\mathbf{p}^i_k\right)}{\big\lVert\mathtt{K}^{-1}\pi^{-1}_\text{2D}\left(\mathbf{p}^i_k\right)\big\rVert}.$$ 2. Transform this estimate from frame $i$ to frame $j$: $$\mathbf{P}^{ji}_k = \pi_\text{3D}\left(\mathtt{T}_{ji}\pi^{-1}_\text{3D} \left({\mathbf{P}^{i}_k}'\right)\right); \ \mathbf{P}^{ji}_k = \mathtt{R}_{ji}{\mathbf{P}^{i}_k}'+\mathbf{t}_{ji}$$ (Notice that we use the superscript ${ji}$ instead of $j$ to distinguish the ground-truth in frame $j$.) 3. Project the resulting 3D point to obtain its pixel coordinates in frame $j$. $$\mathbf{p}^{ji}_k = \pi_\text{2D}\left(\mathtt{K}\mathbf{P}^{ji}_k\right).$$ Altogether, the reprojection of point $\mathbf{p}^i_k$ to frame $j$ is defined as follows: $$\label{eq:reprojection} \mathbf{p}^{ji}_{k}\left(\mathbf{p}^i_k, \ d^i_k, \mathtt{T}_{ji}\right) = \pi_\text{2D}\left(\mathtt{K}\left(\frac{d^i_k\mathtt{R}_{ji}\mathtt{K}^{-1}\pi^{-1}_\text{2D}\left(\mathbf{p}^i_k\right)}{\big\lVert\mathtt{K}^{-1}\pi^{-1}_\text{2D}\left(\mathbf{p}^i_k\right)\big\rVert}+\mathbf{t}_{ji}\right)\right).$$ Nonlinear Optimization {#sec:optimization} ---------------------- Most state-of-the-art VO and VSLAM methods rely heavily on nonlinear optimization in order to estimate the state vector $\mathbf{x}$ (*e.g.*, containing the camera poses and 3D map points) from a set of noisy measurements $\mathbf{z} = \{\mathbf{z}_1, \mathbf{z}_2, \hdots \}$ (*e.g.*, image correspondences or pixel intensities). According to Bayes’ theorem, the following equation describes the conditional probability of the state $p(\mathbf{x}|\mathbf{z})$ given the measurement model $p(\mathbf{z}|\mathbf{x})$ and the prior over the state $p(\mathbf{x})$: $$\label{eq:bayes} p(\mathbf{x}|\mathbf{z}) = \frac{p(\mathbf{z}|\mathbf{x})p(\mathbf{x})}{p(\mathbf{z})}$$ Our aim is then to find the state $\mathbf{x}$ that maximizes this probability. This is called the Maximum a Posteriori (MAP) problem, and the solution corresponds to the mode of the posterior distribution: $$\label{eq:MAP} \mathbf{x}_{MAP} = \operatorname*{arg\,max}_{\mathbf{x}} p(\mathbf{x}|\mathbf{z}) = \operatorname*{arg\,max}_{\mathbf{x}} \frac{p(\mathbf{z}|\mathbf{x})p(\mathbf{x})}{p(\mathbf{z})}$$ Modern VSLAM and VO methods are based on smoothing and often assume a uniform prior $p(\mathbf{x})$. The normalization constant $p(\mathbf{z})$ does not depend on the state either. Therefore, we can drop $p(\mathbf{x})$ and $p(\mathbf{z})$ from , turning the problem into the Maximum Likelihood Estimation (MLE). Assuming the independence between the measurements, this means that becomes $$\mathbf{x}_{MAP}= \mathbf{x}_{MLE}=\operatorname*{arg\,max}_{\mathbf{x}} p(\mathbf{z}|\mathbf{x}) = \operatorname*{arg\,max}_{\mathbf{x}} \displaystyle \prod_{k} p(\mathbf{z}_k|\mathbf{x}).$$ Suppose that the measurement model is given by $\mathbf{z}_k = \mathbf{h}_k(\mathbf{x}) + \mathbf{\delta}_k$, where $\mathbf{\delta}_k \sim \mathcal{N}(\mathbf{0},\mathtt{\Omega}_k)$. The conditional distribution of the individual measurements is then $p(\mathbf{z}_k|\mathbf{x}) \sim \mathcal{N}(\mathbf{h}_k(\mathbf{x}),\mathtt{\Omega}_k)$. Maximizing, for convenience, the $\log$ of the conditionals leads to $$\begin{aligned} &\mathbf{x}_{MAP} = \operatorname*{arg\,max}_{\mathbf{x}} \log(\displaystyle \prod_{k} p(\mathbf{z}_k|\mathbf{x})) = \operatorname*{arg\,max}_{\mathbf{x}} \sum_{k} \log( p(\mathbf{z}_k|\mathbf{x})) \nonumber\\ &= \operatorname*{arg\,max}_{\mathbf{x}} \sum_{k} \log( \exp(-\frac{1}{2}(\mathbf{z}_k-\mathbf{h}_k(\mathbf{x}))^\top \mathtt{\Omega}_k^{-1} (\mathbf{z}_k-\mathbf{h}_k(\mathbf{x})))) = \operatorname*{arg\,min}_{\mathbf{x}} \sum_{k} ||\mathbf{r}_k(\mathbf{x})||_{\mathtt{\Omega}_k}^2, \label{eq:wls} $$ where $||\mathbf{r}_k(\mathbf{x})||_{\mathtt{\Omega}_k}=\sqrt{(\mathbf{z}_k-\mathbf{h}_k(\mathbf{x}))^\top \mathtt{\Omega}_k^{-1} (\mathbf{z}_k-\mathbf{h}_k(\mathbf{x}))}$ is called the *Mahalanobis* distance. As $\mathbf{h}_k(\mathbf{x})$ is typically nonlinear, we solve using an iterative method. A standard approach is to use the Gauss-Newton algorithm described as follows: 1. Make an initial guess $\breve{\mathbf{x}}$. 2. Linearize using the Taylor approximation at $\breve{\mathbf{x}}$. 3. Compute the optimal increment $\Delta\mathbf{x}^*$ that minimizes the linearized cost function. 4. Update the state: $\breve{\mathbf{x}}\leftarrow\breve{\mathbf{x}}+\Delta\mathbf{x}^*$. 5. Iterate the Step 2 – 4 until convergence. The Taylor approximation in Step 2 gives $$\mathbf{h}_k(\breve{\mathbf{x}}+\Delta\mathbf{x})\approx \mathbf{h}_k(\breve{\mathbf{x}})+\mathbf{J}_k\Delta\mathbf{x} \quad \text{with} \quad \mathbf{J}_k=\frac{\partial \mathbf{h}_k(\mathbf{x})}{\partial \mathbf{x}}\biggr\rvert_{\breve{\mathbf{x}}}.$$ This allows us to approximate $||\mathbf{r}_k(\breve{\mathbf{x}}+\Delta\mathbf{x})||_{\mathtt{\Omega}_k}^2$ as $$\begin{aligned} ||\mathbf{r}_k(\breve{\mathbf{x}}+\Delta\mathbf{x})||_{\mathtt{\Omega}_k}^2 &= (\mathbf{z}_k-\mathbf{h}_k(\breve{\mathbf{x}}+\Delta\mathbf{x}))^\top \mathtt{\Omega}_k^{-1} (\mathbf{z}_k-\mathbf{h}_k(\breve{\mathbf{x}}+\Delta\mathbf{x}))\\ &\approx (\mathbf{z}_k-\mathbf{h}_k(\breve{\mathbf{x}})-\mathbf{J}_k\Delta\mathbf{x})^\top \mathtt{\Omega}_k^{-1} (\mathbf{z}_k-\mathbf{h}_k(\breve{\mathbf{x}})-\mathbf{J}_k\Delta\mathbf{x})\\ &= \Delta\mathbf{x}^\top\mathbf{J}_k^\top\mathtt{\Omega}_k^{-1}\mathbf{J}_k\Delta\mathbf{x} + (\mathbf{z}_k-\mathbf{h}_k(\breve{\mathbf{x}}))^\top\mathtt{\Omega}_k^{-1}(\mathbf{z}_k-\mathbf{h}_k(\breve{\mathbf{x}})) \nonumber\\ & \quad -2(\mathbf{z}_k-\mathbf{h}_k(\breve{\mathbf{x}}))^\top\mathtt{\Omega}_k^{-1}\mathbf{J}_k\Delta\mathbf{x}.\end{aligned}$$ Now, taking the derivative of $\sum_k ||\mathbf{r}_k(\breve{\mathbf{x}}+\Delta\mathbf{x})||_{\mathtt{\Omega}_k}^2$ with respect to $\Delta\mathbf{x}$ and setting it to zero, we obtain the optimal increment in the following form: $$\Delta\mathbf{x}^*=-\underbrace{\left[\sum_k \mathbf{J}_k^\top\mathtt{\Omega}_k^{-1}\mathbf{J}_k\right]^{-1}}_{\displaystyle\mathbf{H}^{-1}}\underbrace{\sum_k \mathbf{J}_k^\top\mathtt{\Omega}_k^{-1}(\mathbf{h}_k(\breve{\mathbf{x}})-\mathbf{z}_k)}_{\displaystyle\mathbf{b}}.$$ The Levenberg-Marquardt algorithm, a variant of the Gauss-Newton method, includes a non-negative damping factor $\lambda$ in the update step: $$\Delta\mathbf{x}^*=-\left(\mathbf{H}+\lambda \ \text{diag}(\mathbf{H})\right)^{-1}\mathbf{b},$$ where $\lambda$ is increased when the cost function reduces too slowly, and vice versa. For more details on the adjustment rule, see [@madsen2004methods]. Since the least squares problems are very sensitive to outliers, a common practice is to adopt a robust weight function that downweights large errors: $$\mathbf{x}_\text{robust} = \operatorname*{arg\,min}_{\mathbf{x}} \sum_{k} \omega\left(||\mathbf{r}_k(\mathbf{x})||_{\mathtt{\Omega}_k}\right) ||\mathbf{r}_k(\mathbf{x})||_{\mathtt{\Omega}_k}^2.$$ To solve this problem iteratively, it is usually assumed that the weights are dependent on the residual at the previous iteration, which turns the problem into a standard weighted least squares at each iteration. This technique is called the iteratively reweighted least squares (IRLS). The readers are referred to [@huber2011robust; @zhang1997parameter] for more details on the robust cost functions and [@barfoot2017state] for in-depth study of state estimation for robotics. Lie Algebras ------------ Standard optimization techniques assume that the state belongs to a Euclidean vector space. This does not hold for 3D rotation matrices $\mathtt{R}$, belonging to the special orthogonal group $\ensuremath{\mathrm{SO}(3)}\xspace$, or for 6 degrees-of-freedom (DoF) rigid body motions $\mathtt{T}$, belonging to the special Euclidean group $\ensuremath{\mathrm{SE}(3)}\xspace$. In both cases, state updates have to be done in the tangent space of $\ensuremath{\mathrm{SO}(3)}\xspace$ and $\ensuremath{\mathrm{SE}(3)}\xspace$ at the identity, which are denoted as $\ensuremath{\mathfrak{so}(3)}\xspace$ and $\ensuremath{\mathfrak{se}(3)}\xspace$. Elements of the tangent space $\ensuremath{\mathfrak{so}(3)}\xspace$ and $\ensuremath{\mathfrak{se}(3)}\xspace$ can be represented as vector $\bm{\omega} \in \mathbb{R}^3$ and $\bm{\xi}=[\bm{\omega}, \bm{\nu}]^\top \in \mathbb{R}^6$, respectively. The *hat* operator $(\cdot)^\wedge$ converts $\bm{\omega} \in \mathbb{R}^3$ to the space of skew symmetric matrices of the Lie algebra and its inverse is denoted by the *vee* operator $(\cdot)^\vee$: $$\bm{\omega}^\wedge = \begin{bmatrix} \omega_x\\ \omega_y\\ \omega_z \end{bmatrix}^\wedge = \begin{bmatrix} 0 & -\omega_z & \omega_y\\ \omega_z & 0 & -\omega_x\\ -\omega_y & \omega_x & 0 \end{bmatrix}\in\ensuremath{\mathfrak{so}(3)} \quad \text{and} \quad \left(\bm{\omega}^\wedge\right)^\vee=\bm{\omega}\in\mathbb{R}^3.$$ We denote the exponential and logarithmic mapping between $\ensuremath{\mathfrak{se}(3)}\xspace$ and $\ensuremath{\mathrm{SE}(3)}\xspace$ by $\mathrm{exp}_\text{SE(3)}(\bm{\xi})$ and $\mathrm{log_\text{SE3}(\mathtt{T})}$, respectively: $$\label{eq:exp_se3} \mathrm{exp}_\mathrm{SE(3)}(\bm{\xi}):= \begin{bmatrix} \mathrm{exp}(\bm{\omega}^\wedge) & \mathtt{V}\bm{\nu}\\ 0 & 1 \end{bmatrix} =\begin{bmatrix} \mathtt{R} & \mathbf{t}\\ 0 & 1 \end{bmatrix} =\mathtt{T} \in \mathrm{SE}(3),$$ where $$\mathrm{exp}(\bm{\omega}^\wedge)= \mathtt{I}_{3\times3}+\frac{\sin{\left(\lVert\bm{\omega}\rVert\right)}}{\lVert\bm{\omega}\rVert}\bm{\omega}^\wedge+\frac{1-\cos{\left(\lVert\bm{\omega}\rVert\right)}}{\lVert\bm{\omega}\rVert^2}\left(\bm{\omega}^\wedge\right)^2$$ and $$\mathtt{V} = \mathtt{I}_{3\times3} +\frac{1-\cos{\lVert\bm{\omega}\rVert}}{\lVert\bm{\omega}\rVert^2}\bm{\omega}^\wedge +\frac{\lVert\bm{\omega}\rVert-\sin{\left(\lVert\bm{\omega}\rVert\right)}}{\lVert\bm{\omega}\rVert^3}\left(\bm{\omega}^\wedge\right)^2.$$ From , the logarithm map can be obtained: $$\mathrm{log}_\text{SE(3)}(\mathtt{T}):= \begin{bmatrix} \left(\log\mathtt{R}\right)^\vee \\ \mathtt{V}^{-1}\mathbf{t} \end{bmatrix},$$ where $$\log\mathtt{R} =\frac{\theta}{2\sin{\theta}}\left(\mathtt{R}-\mathtt{R}^\top\right) \quad \text{with} \quad \theta=\cos^{-1}\left(\frac{\mathrm{trace}(\mathtt{R})-1}{2}\right).$$ For optimization purposes, rigid body transformations can be conveniently represented as $\mathrm{exp}_\mathrm{SE(3)}(\Delta\bm{\xi}) \mathtt{T}$, composed of the incremental motion $\Delta\bm{\xi} \in \ensuremath{\mathfrak{se}(3)}\xspace$ and the current estimate $\mathtt{T} \in \ensuremath{\mathrm{SE}(3)}\xspace$. This allows to optimize the incremental update $\Delta\bm{\xi}$ in the tangent space of the current estimate $\mathtt{T}$. Once the optimal increment $\Delta\bm{\xi}^*$ is found, the transformation matrix $\mathtt{T}$ is updated as $$\label{eq:update_SE3} \mathtt{T} \leftarrow \mathrm{exp}_\mathrm{SE(3)}(\Delta\bm{\xi}^*) \mathtt{T}.$$ Note that we follow the *left-multiplication* convention to be consistent with [@strasdat2010scale; @whelan2016elasticfusion]. We refer the readers to [@corke2017robotics] for a reference on the representation of 6 DoF pose in the 3D space, and to [@strasdat2012local; @sola2018micro] for introductions to Lie algebras for odometry and SLAM. Camera Tracking {#sec:motion} =============== In this section, we detail the algorithms that are most commonly used for estimating the 6 DoF motion of an RGB-D camera. The methods will be divided attending to the type of residual they minimize: - [**Methods based on photometric alignment (Section \[sec:directf2f\]).**]{} The alignment results from the minimization of a photometric error over corresponding pixels in two frames. - [**Methods based on geometric alignment (Section \[sec:featuref2f\]).**]{} While direct methods minimize a photometric error, we refer to geometric alignment methods to those that minimize geometric residuals either in the image or 3D domains. Recent results suggest that direct methods present a higher accuracy than those based on geometric alignment, both in odometry [@engel2018direct] and mapping [@zubizarreta2019direct]. Most of the state-of-the-art systems are, because of this reason, based on dense frame alignment. Among the weaknesses of direct methods we can name their small basin of convergence, which can limit the accuracy in wide baselines cases, and their sensitivity to calibration errors, rolling shutter or unsynchronisation between the color and depth images [@schops2019bad]. ![ Different types of error criteria frequently used in the literature: (a) A photometric reprojection error is the pixel intensity difference between a reference pixel in frame $i$ and its reprojection in frame $j$. (b) Given a reference point in frame $i$, a geometric reprojection error is the image distance between its match and the reprojection in frame $j$. (c) A 3D point-to-point distance is the Euclidean distance between the backprojections of two matched points. (d) A 3D point-to-plane distance is the Euclidean distance between the tangent plane at the backprojected reference point in frame $i$ and the backprojected reprojection of the point in frame $j$.[]{data-label="fig:errors"}](errors.pdf){width="\textwidth"} Photometric Alignment {#sec:directf2f} --------------------- Assuming that the same scene point will have the same color in different images, photometric alignment aims to estimate the motion between two frames by minimizing the pixel intensity difference. This error criterion is called a *photometric reprojection* error. For each pixel ${\mathbf p}^i_k$ in the reference frame $i$, it is given by $$\label{ec:photores} {r_{ph}}_k({\Delta\bm{\xi}_{ji}}) = I_i\left({\mathbf p}^i_k\right)- I_j\big(\mathbf{p}^{ji}_{k}({\Delta\bm{\xi}_{ji}}) \ \big),$$ where $\mathbf{p}^{ji}_{k}({\Delta\bm{\xi}_{ji}})$ is the reprojection of $\mathbf{p}^i_k$ in frame $j$ evaluated at the incrementally updated transformation $\mathrm{exp}_\mathrm{SE(3)}(\Delta\bm{\xi}_{ji})\mathtt{T}_{ji}$. Fig. \[fig:errors\]a illustrates this error. Notice that in we omitted some of the variables in the reprojection function $\mathbf{p}^{ji}_{k}$ for readability. The full function is written as $$\label{eq:full_reprojection} \mathbf{p}^{ji}_{k}({\Delta\bm{\xi}_{ji}}) \stackrel{\eqref{eq:reprojection}}{=}\mathbf{p}^{ji}_{k}\left(\mathbf{p}^i_k, \ d^i_k, \ \mathrm{exp}_\mathrm{SE(3)}(\Delta\bm{\xi}_{ji})\mathtt{T}_{ji}\right).$$ The total cost function to minimize, $E_{ph}$, is the weighted squared sum of the individual photometric errors for all considered pixels: $$\label{eq:total_photometric} \Delta\bm{\xi}_{ji}^* = \operatorname*{arg\,min}_{\Delta\bm{\xi}_{ji}} \ E_{ph}(\Delta\bm{\xi}_{ji}) = \operatorname*{arg\,min}_{\Delta\bm{\xi}_{ji}} \sum_k \omega({r_{ph}}_k)\left({r_{ph}}_k(\Delta\bm{\xi}_{ji})\right)^2$$ with some weight function $\omega$, *e.g.,* constant for unweighted least squares, or robust weight function such as Huber’s [@huber2011robust]. As discussed in Section \[sec:optimization\], this problem can be solved using IRLS. Once the optimal increment is found, $\mathtt{T}_{ji}$ is updated using , and this optimization process is iterated until convergence. [@kerl2013robust] proposes a similar photometric alignment between consecutive frames of a video, achieving very accurate odometry results. For study of different alignment strategies, we refer to [@klose2013efficient]. The photometric alignment can also be done in a frame-to-map basis. For example, in [@concha2017rgbdtam], photometric and geometric errors are used to track the camera pose with respect to the closest keyframe in the map. Geometric Alignment {#sec:featuref2f} ------------------- In contrast to photometric alignment that directly uses raw pixel intensities, geometric alignment estimates the camera motion by minimizing the Euclidean distances between the two corresponding sets of geometric primitives in 2D or 3D. **2D Point-to-Point Alignment:** A *geometric reprojection* error is the most representative type of 2D error used in VO and VSLAM. This error is illustrated in Fig. \[fig:errors\]b. Given a point $\mathbf{p}^i_k$ in the reference frame $i$, it measures the image distance between its match $\mathbf{p}^j_k$ and the projection $\mathbf{p}^{ji}_k$ in the current frame $j$: $$\label{eq:featurebasedframe2mapmotion} {r_{2D_{\scriptstyle k}}}(\Delta\bm{\xi}_{ji}) = \frac{\lVert\mathbf{p}^j_k-\mathbf{p}^{ji}_k(\Delta\bm{\xi}_{ji})\rVert}{\sigma^i_k} \quad \text{with} \quad \sigma^i_k = \left(\lambda_\text{pyr}\right)^{\displaystyle L_{\text{pyr},\mathbf{p}^i_k}},$$ where $\sigma^i_k$ is the standard deviation of the image point $\mathbf{p}^i_k$ that depends on the scale factor of the image pyramid $\lambda_\text{pyr} (>1)$ and the level $L_{\text{pyr},\mathbf{p}^i_k}$ at which the point was detected. Unlike photometric errors, geometric errors require data association. For sparse points, this can be done by matching feature descriptors (*e.g.*, SIFT [@lowe2004distinctive], SURF [@bay2006surf], ORB [@rublee2011orb]) or extracting salient corners (*e.g.*, Harris corner [@harris1988combined], FAST [@rosten2006machine] or Shi-Tomasi [@shi1994good] features) and tracking them [@lucas1981iterative]. Aggregating ${{r}_{2D_{\scriptstyle k}}}$ for every point $k$, we obtain the total cost function analogous to : $$\label{eq:total_2Dpoint2point} \Delta\bm{\xi}_{ji}^* = \operatorname*{arg\,min}_{\Delta\bm{\xi}_{ji}} \ E_{2D}(\Delta\bm{\xi}_{ji}) = \operatorname*{arg\,min}_{\Delta\bm{\xi}_{ji}} \sum_k \omega(r_{2D_{\scriptstyle k}})\left(r_{2D_{\scriptstyle k}}(\Delta\bm{\xi}_{ji})\right)^2.$$ Minimizing this cost function to estimate the camera motion is called *motion-only* bundle adjustment, and this method is used among others in ORB-SLAM2 [@mur2017orb] for tracking. **3D Point-to-Point Alignment:** Instead of minimizing the reprojection error in 2D image space, one can also minimize the distance between the backprojected points in 3D space (see Fig. \[fig:errors\]c). The 3D errors can be defined over dense point clouds or sparse ones. For the latter case, the first step should be the extraction and matching of the sparse salient points in the RGB channels. Henry et al. [@henry2010rgb], for example, uses SIFT features [@lowe2004distinctive], although others could be used. Given two sets of correspondences in image $i$ and $j$, the 3D geometric error is obtained as $$\begin{gathered} {r_{3D_{\scriptstyle k}}}(\Delta\bm{\xi}_{ji}) = \big\lVert{\mathbf{P}^j_k}'-\mathbf{P}^{ji}_k(\Delta\bm{\xi}_{ji})\big\rVert \label{eq:icp_feature_based} \\ \text{with} \ \ \mathbf{P}^{ji}_k(\Delta\bm{\xi}_{ji}):=\pi_\text{3D}\left(\mathrm{exp}_\mathrm{SE(3)}(\Delta\bm{\xi}_{ji})\mathtt{T}_{ji}\pi^{-1}_\text{3D}\left({\mathbf{P}^{i}_k}'\right)\right) \label{eq:corresponding_backprojection},\end{gathered}$$ where ${\mathbf{P}^i_k}'$ and ${\mathbf{P}^j_k}'$ are the 3D points backprojected from the 2D correspondence $\mathbf{p}^i_k$ and $\mathbf{p}^j_k$ using . Aggregating ${{r}_{2D_{\scriptstyle k}}}$ for every point $k$, we obtain the total cost function analogous to and : $$\label{eq:total_3Dpoint2point} \Delta\bm{\xi}_{ji}^* = \operatorname*{arg\,min}_{\Delta\bm{\xi}_{ji}} \ E_{3D}(\Delta\bm{\xi}_{ji}) = \operatorname*{arg\,min}_{\Delta\bm{\xi}_{ji}} \sum_k \omega(r_{3D_{\scriptstyle k}})\left(r_{3D_{\scriptstyle k}}(\Delta\bm{\xi}_{ji})\right)^2.$$ For the case of dense cloud alignment, the standard algorithm is *Iterative Closest Point (ICP)* [@besl1992method]. ICP alternates the minimization of a geometric distance between points (the point-to-point distance in equation \[eq:icp\_feature\_based\] or the point-to-plane one defined later in this section) and the search for correspondences (usually the nearest neighbours in the 3D space). The strengths and limitations of sparse and dense cloud alignment are complementary for RGB-D data. Aligning dense point clouds can lead to more accurate motion estimation than aligning sparse ones, as they use more data. On the other hand, ICP might diverge if the initial estimate is not sufficiently close to the real motion. In practice, combining the two is a preferred approach: Sparse alignment, based on feature correspondences, can produce a robust and reliable initial seed. Afterwards, dense alignment can refine such initial estimate using ICP. **3D Point-to-Plane Alignment:** The point-to-plane distance, that minimizes the distance along the target point normal, is commonly used in dense RGB-D point cloud alignment [@henry2010rgb; @newcombe2011kinectfusion; @whelan2015elasticfusion_wo_posegraph; @dai2017bundle]. The residual is in this case $$\label{eq:icp_point2plane} r_{3DP_{\scriptstyle k}} (\Delta\bm{\xi}_{ji}) = \Biggr| \mathbf{n}^i_k\cdot \left( {\mathbf{P}^i_k}'- \left( \text{exp}_\text{SE(3)}(\Delta\bm{\xi}_{ji})\mathtt{T}_{ji}\right)^{-1}\left(d^{ji}_k\frac{\mathtt{K}^{-1}\pi^{-1}_\text{2D}\left(\mathbf{p}^{ji}_k(\Delta\bm{\xi}_{ji})\right)}{\lVert\mathtt{K}^{-1}\pi^{-1}_\text{2D}\left(\mathbf{p}^{ji}_k(\Delta\bm{\xi}_{ji})\right)\rVert}\right) \right) \Biggr|,$$ where ${\mathbf{P}^i_k}'$ is the 3D backprojection of $\mathbf{p}^i_k$ using , $\mathbf{n}_k^i$ is the surface normal at ${\mathbf{P}^i_k}'$, $\mathbf{p}^{ji}_k(\Delta\bm{\xi}_{ji})$ is the reprojection of $\mathbf{p}^i_k$ in frame $j$ evaluated at the incrementally updated transformation $\text{exp}_\text{SE(3)}(\Delta\bm{\xi}_{ji})\mathtt{T}_{ji}$, which is given by and , and $d^{ji}_k$ is the measured depth at this reprojection in frame $j$. This error is illustrated in Fig. \[fig:errors\]d. Aggregating ${{r}_{3DP_{\scriptstyle k}}}$ for every point $k$, we obtain the total cost function analogous to , and : $$\label{eq:total_3Dpoint2plane} \Delta\bm{\xi}_{ji}^* = \operatorname*{arg\,min}_{\Delta\bm{\xi}_{ji}} \ E_{3DP}(\Delta\bm{\xi}_{ji}) = \operatorname*{arg\,min}_{\Delta\bm{\xi}_{ji}} \sum_k \omega(r_{3DP_{\scriptstyle k}})\left(r_{3DP_{\scriptstyle k}}(\Delta\bm{\xi}_{ji})\right)^2.$$ Scene Mapping {#sec:mapping} ============= In this section we briefly survey the main algorithms for estimating scene maps from several RGB-D views. There are two basic types of scene representations that are commonly used, that we will denote as **point-based maps (Section \[sec:pointcloudmap\])** and **volumetric maps (Section \[sec:volmap\])**. Point-Based Mapping {#sec:pointcloudmap} ------------------- Representing a scene as a set of points or surfels is one of the most common alternatives for estimating local maps of a scene. Bundle Adjustment [@triggs1999bundle], consisting on the joint optimization of a set of camera poses and points, is frequently used to obtain a globally consistent model of the scene [@mur2017orb]. However, there are also several recent VSLAM approaches that alternate the optimization between points and poses, reducing the computational cost with a small impact in the accuracy, given a sufficient number of points [@zhou2014color; @platinsky2017monocular; @yokozuka2019vitamin; @schops2019bad]. In its most basic form, the map model consists of a set of $n$ points and $m$ RGB-D keyframes. Every point is represented by its 3D position in the world reference frame ${\mathbf P}_k^w$. For every keyframe $i$, we store its pose $\mathtt{T}_{iw}$ and its RGB-D image $I_i$. Similarly to camera tracking in Section \[sec:motion\], map optimization algorithms are based on the photometric or geometric alignment between the keyframes. In this case, however, both the keyframe poses and point positions are optimized. **Photometric Bundle Adjustment:** This method minimizes a cost function similar to , with the difference that it does not backproject the 2D points using the measured depths. Instead, it aims to find the 3D point that minimizes the photometric errors in all keyframes where it was visible. Let $\mathbf{P}_{\mathcal{M}} = \left( \mathbf{P}_1, \hdots, \mathbf{P}_k, \hdots, \mathbf{P}_n \right)^\top$ be the set of all map points and $\Delta\bm{\xi}_{\mathcal{M}} = \left( \Delta\bm{\xi}_{1w}, \hdots, \Delta\bm{\xi}_{jw}, \hdots, \Delta\bm{\xi}_{mw} \right)^\top$ the set of incremental transformations to the current estimates of the keyframe poses. Then, the optimization problem is formulated as $$\begin{aligned} \{\Delta\bm{\xi}_{\mathcal{M}}^*, \mathbf{P}^*_{\mathcal{M}}\} &= \operatorname*{arg\,min}_{\Delta\bm{\xi}_{\mathcal{M}},\mathbf{P}_{\mathcal{M}}} \ E_{ph}(\Delta\bm{\xi}_{\mathcal{M}},\mathbf{P}_{\mathcal{M}})\\ &= \operatorname*{arg\,min}_{\Delta\bm{\xi}_{jw},\mathbf{P}^w_k} \sum_{j} \sum_{k} \omega({r_{ph}}_k)\left({r_{ph}}_k(\Delta\bm{\xi}_{jw},\mathbf{P}^w_k)\right)^2\end{aligned}$$ with $$\begin{aligned} &{r_{ph}}_k(\Delta\bm{\xi}_{jw},\mathbf{P}^w_k) \nonumber\\ &= \begin{cases} 0 & \text{if} \ \mathbf{P}_k \text{ is not visible} \\ & \text{in frame } j,\\I_i\left({\mathbf p}^i_k\right)- I_j\big(\pi_\text{2D}\big(\mathtt{K}\pi_\text{3D}(\mathrm{exp}_\text{SE(3)}(\Delta\bm{\xi}_{jw})\mathtt{T}_{jw}\pi_\text{3D}^{-1}(\mathbf{P}^w_k))\big) \ \big) & \text{otherwise}, \end{cases}\end{aligned}$$ where $I_i\left({\mathbf p}^i_k\right)$ is the pixel intensity at which $\mathbf{P}_k^w$ was detected in its reference keyframe $i$ (*i.e.,* the keyframe in which the point was first detected and parameterized). **Geometric Bundle Adjustment:** This method minimizes a cost function similar to , with the difference that the reprojection with the measured depth is replaced by the projection of the current estimate of the 3D point. Using the same notation as for the photometric bundle adjustment, the optimization problem is formulated as $$\begin{aligned} \{\Delta\bm{\xi}_{\mathcal{M}}^*, \mathbf{P}^*_{\mathcal{M}}\} &= \operatorname*{arg\,min}_{\Delta\bm{\xi}_{\mathcal{M}},\mathbf{P}_{\mathcal{M}}} \ E_{2D}(\Delta\bm{\xi}_{\mathcal{M}},\mathbf{P}_{\mathcal{M}})\\ &= \operatorname*{arg\,min}_{\Delta\bm{\xi}_{jw},\mathbf{P}^w_k} \sum_{j} \sum_{k} \omega({r_{2D}}_k)\left({r_{2D}}_k(\Delta\bm{\xi}_{jw},\mathbf{P}^w_k)\right)^2\end{aligned}$$ with $$\begin{aligned} &{r_{2D}}_k(\Delta\bm{\xi}_{jw},\mathbf{P}^w_k) \nonumber\\ &= \begin{cases} 0 &\text{if} \ \mathbf{P}_k \text{ is not detected} \\ & \text{in frame } j,\\[-0.5em] \displaystyle\frac{\bigg\lVert\mathbf{p}^j_k - \pi_\text{2D}\left(\mathtt{K}\pi_\text{3D}\left(\mathrm{exp}_\text{SE(3)}(\Delta\bm{\xi}_{jw})\mathtt{T}_{jw}\pi_\text{3D}^{-1}(\mathbf{P}^w_k)\right)\right) \bigg\rVert}{\sigma^j_k} & \text{otherwise}. \end{cases}\end{aligned}$$ Note that $\sigma^j_k$ is defined in . Volumetric Mapping {#sec:volmap} ------------------ One of the main weaknesses of point-based representations for mapping is that they do not model the empty and occupied space. This can be a problem for applications such as robot navigation or occlusion modelling in AR. Volumetric mapping aims to overcome such problems by modeling the occupancy of the whole 3D scene volume. The most usual model for volumetric maps is the Truncated Signed Distance Function [@curless1996volumetric], used for example in [@newcombe2011kinectfusion; @whelan2013robust; @whelan2015real; @klingensmith2015chisel]. In this representation, the 3D world is discretized into voxels and modeled as a volumetric signed distance field $\Phi:\mathbb{R}^3\rightarrow \mathbb{R}$, where we assign to each cell the distance to the nearest object, which is defined positive if its center is outside the object and negative if it is inside it. Since only the surfaces and their surroundings are considered, the distances are usually truncated if larger than a threshold $\tau$. Also, for every cell, a weight is stored that represents the confidence on the distance measurement. The algorithm for updating a TSDF with new depth measurements measurement was first presented in [@curless1996volumetric]. In a few words, it consists on a weighted running average on the distance measurements from the depth sensors. TSDF is addressed in depth in chapter 5 of this book. For this reason, we do not extend further on it and refer the reader to this chapter, and the references there and in this section, for further detail on this topic. Loop Closing {#sec:loopclosure} ============ Loop closing algorithms correct the drift that has accumulated during exploratory trajectories, maintaining a consistent global representation of the environment. **Loop detection (Section \[sec:placerecognition\])**, is mainly based on the visual appearance between two keyframes of the map. When these two keyframes are imaging the same place and the loop closure has been detected, the geometric constraint between the two is added to the map, which is then updated according to it. This map update is known as **loop correction (Section \[sec:posegraphs\])**, and we detail the pose graph formulation as an efficient alternative for large map representations and loop closing correction. Loop Detection {#sec:placerecognition} -------------- Due to the excellent performance of visual place recognition, many RGB-D SLAM systems use only the RGB channels for loop detection (*e.g.*, [@mur2017orb; @concha2017rgbdtam; @gutierrez2018rgbid]). The most used approaches are based on the bag of words model, first proposed in [@sivic2003video]. The implementation in [@galvez2012bags] is particularly suited for visual SLAM, adding robustness to plain visual appearance by geometric and sequential consistency checks. In the bag of words model the space of local descriptors is divided into discrete clusters using the k-means algorithm. Each cluster is referred to as a visual word, and the set of all visual words forms a visual dictionary. With such a partition, an image is described as the histogram of visual word occurrences. The place querying can be made very efficient by maintaining inverse indexes from the visual words to the database images in which they appear. Bag-of-words descriptors have some limitations for RGB-D odometry and SLAM. They assume images of sufficient texture to extract salient point features, and they do not use the information of the depth channel from RGB-D images. Also, the extraction and description of local features has a considerable computational overhead. There are several approaches in the literature that overcome such limitations. [@gee20126d] proposes to find loop closure candidates without features, by the alignment of keyframes against synthetic views of the map. [@shotton2013scene] uses regression forests to predict correspondences between an RGB-D frame and the map, an approach that has been refined in [@guzman2014multi; @valentin2015exploiting; @cavallari2017fly] among others. [@glocker2015real] proposed to encode each RGB-D image using randomized ferns. Loop Correction {#sec:posegraphs} --------------- Once a loop is detected based on the appearance of two keyframes, a constraint between the poses of both can be computed by photometric and/or geometric alignment. When such constraint is added to the map optimization, the global map consistency is achieved by accommodating this new constraint and correcting the accumulated drift. For computational reasons, this correction is frequently done by pose graph optimization. Fig. \[fig:loop\_closing\] illustrates a loop closure process. A pose graph is a compact map representation composed of the set of $m$ keyframe poses summarizing the trajectory, *i.e.*, $\mathtt{T}_\text{kfs}=\{\mathtt{T}_{1a}, \mathtt{T}_{2a}, \hdots, \mathtt{T}_{ma}\}$ where the reference frame $a$ is chosen from one of the keyframes as the “anchor” to the rest. As this representation does not include map points, it is particularly useful for estimating globally consistent maps of large areas at a reasonable cost, and is used among others in [@kerl2013dense; @endres2012evaluation; @concha2017rgbdtam]. Pose graph optimization aims to minimize the following cost: $$\operatorname*{arg\,min}_{\displaystyle\mathtt{T}_\text{kfs}} \ E_\text{graph} = \operatorname*{arg\,min}_{\displaystyle\mathtt{T}_\text{kfs}} \sum_{(i,j)\in \epsilon_\text{edge}} \mathbf{r}_{ij}^\top \mathtt{\Omega}_{ij}^{-1} \mathbf{r}_{ij}$$ where $\epsilon_\text{edge}$ denotes the set of edges (*i.e.*, relative pose constraints) in the pose graph, $\mathbf{r}_{ij}$ and $\mathtt{\Omega}_{ij}$ are respectively the residual associated to the $i^{th}$ and $j^{th}$ camera poses and its uncertainty. Such residual is defined as $$\mathbf{r}_{ij} = \mathrm{log}_\mathrm{SE(3)}(\mathtt{T}_{ij,0}\mathtt{T}_{ja}\mathtt{T}_{ia}^{-1}) \label{eq:posegraphresidual}$$ where $\mathtt{T}_{ij,0}$ is the fixed transformation constraint from the alignment (Section \[sec:motion\]) and $\mathtt{T}_{ja}\mathtt{T}_{ia}^{-1}=\mathtt{T}_{ji}$ is the current estimate of the relative motion. For more details on the pose graph optimization method, the reader is referred to [@kummerle2011g2o; @rosen2019sesync]. ![An illustration of loop closure: (a) Ground truth. (b) Odometry result containing drift. (c) A loop detection followed by the computation of the loop constraint. (d) The keyframe trajectory after the pose graph optimization.[]{data-label="fig:loop_closing"}](loop_closing.pdf){width="\textwidth"} Advanced topics {#sec:references} =============== In this section, we review some of the relevant approaches in RGB-D odometry and SLAM that, due to space reasons, were not covered in the main part of the chapter. Hybrid Cost Function -------------------- In Section \[sec:motion\] and \[sec:mapping\], we discussed different types of cost functions separately. Many state-of-the-art methods, however, minimize a weighted sum of multiple cost functions. This strategy allows for better utilization of RGB-D data, which can lead to performance gains [@kerl2013dense; @meilland2013super; @damen2012egocentric; @meilland2013unifying]. In [@henry2010rgb], 3D point-to-point error was used for outlier rejection, and then the pose was refined by minimizing the combined 2D point-to-point cost and 3D point-to-plane cost. In [@whelan2015elasticfusion_wo_posegraph; @dai2017bundle], the joint minimization of photometric and point-to-plane cost was used. Another popular method is to jointly minimize the photometric and (inverse) depth cost (which is not discussed here) [@kerl2013dense; @steinbrucker2013large; @babu2016sigma; @gutierrez2015inverse; @concha2017rgbdtam]. Semantic Mapping ---------------- In recent years, there has been an impressive progress in the field of machine learning (specifically deep learning) for visual recognition and segmentation tasks. Building on them, there have appeared several visual SLAM algorithms that not only estimate geometric models, but also annotate them with high-level semantic information (see Fig. \[fig:semanticslam\] for an illustration). The research on semantic mapping is not as mature as geometric mapping, with challenges related to robustness, accuracy and cost. The state-of-the-art systems, however, show promising results. Semantic mapping could improve the accuracy and robustness of current SLAM algorithms, and widen their applications. For example, [@bescos2018dynaslam] uses a combination of geometry and learning to remove dynamic objects and create life-long maps, achieving better accuracy than geometric SLAM baselines. Similarly, [@keller2013real] uses data association failures and region growing to segment and remove dynamic objects, improving the system robustness and accuracy. ![Illustration of results from semantic RGB-D SLAM. Different colors indicate different object categories. Figures taken from [@tateno2015real].[]{data-label="fig:semanticslam"}](tateno2015iros1.jpg){width="70.00000%"} One can differentiate between maps based on specific object instances and object categories. An approach like [@salas2013slam++] adopts the map of the former type. It assumes that a database of relevant objects in the scene is available. The map is then modeled as a graph of keyframe and object poses, and it is optimized using the constraints from keyframe-to-keyframe point cloud alignment and keyframe-to-object using [@drost2010model]. Object-based RGB-D SLAM has also been addressed in [@sunderhauf2017meaningful] Most category-wise semantic mapping methods leverage 2D segmentation algorithms (*e.g.*, [@he2017mask]), differing on how they transfer the labels to the 3D maps. As a few examples, we refer to the following works for this area of research [@hermans2014dense; @mccormac2017semanticfusion; @ma2017multi; @stuckler2015dense; @pham2015hierarchical]. Edge-based methods ------------------ While the majority of the existing methods consider each pixel as independent measurements, edge-based methods exploit the structural regularities of indoor scenes, modelling the scene geometry with lines or edges. This can provide an advantage over point-based methods, especially when the scene has weak texture but strong structural priors. One of the earliest works that demonstrated the advantage of edge-based registration in RGB-D SLAM is [@choi2013rgbd]. This method is based on an efficient edge detection for RGB-D point clouds and 3D registration of the edge points using the ICP algorithm. In [@bose2016fast], it is shown that the edge detection can be accelerated using the previous RGB-D frame. On the other hand, [@lu2015robustness] proposes to model the straight lines only and incorporate their uncertainties in the pose estimation problem. Although this work is shown to outperform [@choi2013rgbd] under lighting variations, it fails when the scene contains few lines. To overcome this limitation, [@lu2015robust] uses both points and lines. In [@kuse2016robust], direct edge alignment is proposed that minimizes the sum of squared distances between the reprojected and the nearest edge point using the distance transform of the edge-map. Other works propose to jointly minimize this edge distance and other errors, *e.g.*, a photometric error [@wang2016edge] and an ICP-based point-to-plane distance [@schenk2017combining]. Later works such as [@zhou2019canny] and [@kim2018edge] take the image gradient direction also into account for the direct edge alignment. As in [@kerl2013robust], these last two works estimate the camera pose using the iteratively reweighted least-squares (IRLS) method with the t-distribution as a robust weight function. Plane-based methods ------------------- Like edges, planes are abundant in man-made environments and can be modelled explicitly for tracking and mapping. In [@taguchi2013point], an RGB-D SLAM system is proposed based on the 3D registration between the minimal set of point/plane primitives. This system is improved in [@ataer2013tracking] and [@ataer2016pinpoint] by incorporating the guided search of points/planes and triangulation of 2D-to-2D/3D point matches, respectively. [@raposo2013plane] proposes an odometry method that uses planes (and points if strictly necessary) and refines the relative pose using a direct method. In [@salasmoreno2014dense], a dense SLAM method is proposed based on dense ICP with a piecewise planar map. In both [@ma2016cpa] and [@hsiao17icra], it is proposed to model planes in a global map, so that they are optimized together with the keyframe poses in the graph optimization for global consistency. The main difference is that the former uses direct image alignment in an EM framework, while the latter combines geometric and photometric methods for the fast odometry estimation. Besides, the latter adopts the minimal plane parameterization proposed in [@kaess2015simultaneous] and does not require GPU. A visual-inertial method based on [@hsiao17icra] is proposed in [@hsiao2018dense]. In [@gao2015robust], it is proposed to use planar point features for tracking and mapping, as they are more accurate than the traditional point features and computationally inexpensive. Other works such as [@le2017dense; @kim2018low; @kim2018linear] use Manhattan world assumption, which simplifies the incorporation of the planes into a SLAM formulation. Finally, [@proenca2018probabilistic] shows that it can be beneficial to use points, lines and planes all together in a joint optimization framework. Multisensor fusion ------------------ The constraints coming from RGB-D data can be combined with other sources of information to increase the accuracy and robustness of the tracking and mapping processes. For example, [@laidlow2017dense] presents a tightly coupled formulation for RGB-D-inertial SLAM based on ElasticFusion [@whelan2016elasticfusion]. In [@klingensmith2016articulated], RGB-D SLAM estimates the configuration space of an articulated arm. [@houseagoko] adds odometric and kinematic constraints from a wheeled robot with a manipulator, and [@scona2017direct] adds the kinematic constraints of a humanoid robot and inertial data. Non-rigid reconstructions ------------------------- The 3D reconstruction of non-rigid environments is a very relevant and challenging area of research that has been frequently addressed using RGB-D sensors. [@newcombe2015dynamicfusion] is one of the most representative systems, achieving impressive results for deformable surfaces. [@runz2018maskfusion] is a recent work that reconstructs a scene with multiple moving objects. [@jaimez2017fast] estimates very efficiently the odometry of an RGB-D camera and the flow of a scene that might contain static and dynamic parts. [@scona2018staticfusion] classifies the scene parts into static and dynamic, fuses the static parts and discard the dynamic ones. A recent survey on 3D reconstruction from RGB-D camera, including dynamic scenes, is conducted in [@zollhofer2018state]. It places emphasis on high-quality offline reconstruction, which is complementary to the focus of this chapter on real-time online reconstruction and camera tracking. Conclusions {#sec:conclusions} =========== Estimating the camera ego-motion and the 3D structure of the surrounding environment is a crucial component in many applications such as photogrammetry, AR and vision-based navigation. For this particular tasks, RGB-D cameras provide significant advantages over RGB cameras, as the additional depth measurements ease the process of metric scene reconstruction. Furthermore, they impose a relatively mild constraint on cost, size and power, making them a popular choice for mobile platforms. As a result, both academia and industry have shown an ever-increasing interest in RGB-D odometry and SLAM methods for the past decade. In this chapter, we reviewed the general formulations of RGB-D odometry and SLAM. The standard pipeline of VSLAM systems consists of three main components: camera pose tracking, scene mapping and loop closing. For tracking and mapping, we discussed some of the widely-used methods and highlighted the difference in their formulations (*i.e.*, photometric vs. geometric alignment and point-based vs. volumetric mapping). For loop closing, we detailed the underlying principles of loop detection and drift correction, namely the appearance-based place recognition and pose graph optimization. Lastly, we presented a brief review of the advanced topics in the research field today. Resources {#sec:resources} ========= There are a high number of available resources in the web related to RGB-D odometry and SLAM. We will refer here the most relevant open-source software and public databases. Code {#code .unnumbered} ---- **FOVIS** [@huang2017visual] (<https://fovis.github.io/>) Implementation of a feature-based RGB-D odometry. **DVO\_SLAM** [@steinbrucker2011real; @kerl2013robust; @kerl2013dense; @kerl2015dense] (<https://github.com/tum-vision/dvo_slam>) Implementation of a frame-to-frame RGB-D visual Odometry. **RGBDSLAM\_v2** [@endres20143] (<https://github.com/felixendres/rgbdslam_v2>, <http://wiki.ros.org/rgbdslam>, <https://openslam-org.github.io/rgbdslam.html>) Implementation of an RGB-D SLAM system, with a feature-based camera tracking and a pose graph as map model. **ElasticFusion** [@whelan2016elasticfusion] (<https://github.com/mp3guy/ElasticFusion>) RGB-D scene-centered SLAM system that models the scene as a set of surfels that are deformed to accommodate loop closures. **RGBDTAM** [@concha2017rgbdtam] (<https://github.com/alejocb/rgbdtam>) RGB-D SLAM system with a pose graph as map model and frame-to-frame tracking. **MaskFusion** [@runz2018maskfusion] (<https://github.com/martinruenz/maskfusion>) A recent semantic (object-based) RGB-D SLAM system for dynamic scenes. **PlaneMatch**[@shi2018planematch] (<https://github.com/yifeishi/PlaneMatch>) RGB-D SLAM algorithm that proposes a novel descriptor for planar surfaces and exploits correspondences between them. Databases {#databases .unnumbered} --------- **RGB-D SLAM Dataset and Benchmark (Also known as *the TUM dataset*)**[@sturm12iros] (<https://vision.in.tum.de/data/datasets/rgbd-dataset>). It contains indoor recordings with ground truth camera pose in a wide variety of conditions: rotation-only and general motion, static and dynamic environments and small and mid-size scene coverage. **The ETH3D dataset**[@schops2019bad] (<https://www.eth3d.net/>). A benchmark dataset for RGB-D SLAM (among others), recorded with synchronized global shutter cameras. **The Matterport dataset**[@Matterport3D] (<https://github.com/niessner/Matterport>). Annotated data captured throughout 90 properties with a Matterport Pro Camera. **Scannet**[@dai2017scannet] (<http://www.scan-net.org/>). RGB-D video dataset annotated with 3D camera poses, reconstructions, and instance-level semantic segmentations. **The ICL-NUIM dataset**[@handa2014benchmark] (<https://www.doc.ic.ac.uk/~ahanda/VaFRIC/iclnuim.html>). This dataset contain RGB-D sequences on synthetic scenes; hence with camera pose and scene ground truth. **InteriorNet**[@InteriorNet18] (<https://interiornet.org/>). Dataset containing RGB-D-inertial streams for synthetic large scale interior scenes. [100]{} \[1\][[\#1]{}]{} urlstyle \[1\][DOI \#1]{} , E., [Taguchi]{}, Y., [Ramalingam]{}, S.: [Pinpoint [SLAM]{}: A Hybrid of 2[D]{} and 3[D]{} Simultaneous Localization and Mapping for [RGB-D]{} Sensors]{}. In: 2016 IEEE International Conference on Robotics and Automation (ICRA), pp. 1300–1307 (2016) , E., [Taguchi]{}, Y., [Ramalingam]{}, S., [Garaas]{}, T.: [Tracking an [RGB-D]{} Camera Using Points and Planes]{}. In: 2013 IEEE International Conference on Computer Vision Workshops, pp. 51–58 (2013) , B.W., [Kim]{}, S., [Yan]{}, Z., [Ren]{}, L.: [[$\sigma$]{}-DVO: Sensor Noise Model Meets Dense Visual Odometry]{}. In: 2016 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pp. 18–26 (2016) Bailey, T., Durrant-Whyte, H.: [Simultaneous Localization and Mapping (SLAM): Part II]{}. IEEE Robotics & Automation Magazine **13**(3), 108–117 (2006) Barfoot, T.D.: State Estimation for Robotics. Cambridge University Press (2017) Bay, H., Tuytelaars, T., Van Gool, L.: [[SURF]{}: [S]{}peeded Up Robust Features]{}. In: Proceedings of the European Conference on Computer Vision (ECCV), pp. 404–417 (2006) Besc[ó]{}s, B., F[á]{}cil, J.M., Civera, J., Neira, J.: [DynaSLAM: Tracking, Mapping, and Inpainting in Dynamic Scenes]{}. IEEE Robotics and Automation Letters **3**(4), 4076–4083 (2018) Besl, P.J., McKay, N.D.: [Method for Registration of 3-D Shapes]{}. In: Sensor Fusion IV: Control Paradigms and Data Structures, vol. 1611, pp. 586–607. International Society for Optics and Photonics (1992) Bloesch, M., Czarnowski, J., Clark, R., Leutenegger, S., Davison, A.J.: [CodeSLAM - Learning a Compact, Optimisable Representation for Dense Visual SLAM]{}. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2560–2568 (2018) , L., [Richards]{}, A.: Fast depth edge detection and edge based [RGB-D SLAM]{}. In: 2016 IEEE International Conference on Robotics and Automation (ICRA), pp. 1323–1330 (2016) Cadena, C., Carlone, L., Carrillo, H., Latif, Y., Scaramuzza, D., Neira, J., Reid, I., Leonard, J.J.: [Past, Present, and Future of Simultaneous Localization and Mapping: Towards the Robust-Perception Age]{}. IEEE Transactions on Robotics **32**(6), 1309–1332 (2016) Castellanos, J.A., Montiel, J., Neira, J., Tard[ó]{}s, J.D.: [The SPmap: A Probabilistic Framework for Simultaneous Localization and Map Building]{}. IEEE Transactions on robotics and Automation **15**(5), 948–952 (1999) Cavallari, T., Golodetz, S., Lord, N.A., Valentin, J., Di Stefano, L., Torr, P.H.: [On-the-Fly Adaptation of Regression Forests for Online Camera Relocalisation]{}. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 4457–4466 (2017) Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: [[Matterport3D]{}: Learning from [RGB-D]{} Data in Indoor Environments]{}. International Conference on 3D Vision (3DV) (2017) , C., [Trevor]{}, A.J.B., [Christensen]{}, H.I.: [[RGB-D]{} Edge Detection and Edge-Based Registration]{}. In: 2013 IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 1568–1575 (2013) Concha, A., Civera, J.: [RGBDTAM: A Cost-Effective and Accurate RGB-D Tracking and Mapping System]{}. In: Intelligent Robots and Systems (IROS), 2017 IEEE/RSJ International Conference on, pp. 6756–6763. IEEE (2017) Concha, A., Hussain, M.W., Montano, L., Civera, J.: [Manhattan and Piecewise-Planar Constraints for Dense Monocular Mapping.]{} In: Robotics: Science and systems (2014) Concha, A., Loianno, G., Kumar, V., Civera, J.: [Visual-Inertial Direct SLAM]{}. In: 2016 IEEE International Conference on Robotics and Automation (ICRA), pp. 1331–1338. IEEE (2016) Corke, P.: Robotics, Vision and Control: Fundamental Algorithms In MATLAB Second, Completely Revised, chap. 1, pp. 15–41. Springer (2017) Curless, B., Levoy, M.: [A Volumetric Method for Building Complex Models from Range Images]{}. In: Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pp. 303–312. ACM (1996) Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nie[ß]{}ner, M.: [ScanNet: Richly-annotated 3D Reconstructions of Indoor Scenes]{}. In: Proc. Computer Vision and Pattern Recognition (CVPR), IEEE (2017) Dai, A., Nie[ß]{}ner, M., Zollh[ö]{}fer, M., Izadi, S., Theobalt, C.: [[BundleFusion]{}: Real-Time Globally Consistent [3D]{} Reconstruction Using On-the-Fly Surface Reintegration]{}. Trans. Graph. **36**(3), 24:1–24:18 (2017) Damen, D., Gee, A., Mayol-Cuevas, W., Calway, A.: [Egocentric Real-Time Workspace Monitoring using an RGB-D Camera]{}. In: 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 1029–1036. IEEE (2012) Drost, B., Ulrich, M., Navab, N., Ilic, S.: [Model Globally, Match Locally: Efficient and Robust 3D Object Recognition]{}. In: 2010 IEEE computer society conference on computer vision and pattern recognition, pp. 998–1005. Ieee (2010) Durrant-Whyte, H., Bailey, T.: [Simultaneous Localization and Mapping: Part I]{}. IEEE robotics & automation magazine **13**(2), 99–110 (2006) Endres, F., Hess, J., Engelhard, N., Sturm, J., Cremers, D., Burgard, W.: [An Evaluation of the RGB-D SLAM System]{}. In: Robotics and Automation (ICRA), 2012 IEEE International Conference on, pp. 1691–1696. IEEE (2012) Endres, F., Hess, J., Sturm, J., Cremers, D., Burgard, W.: [3-D Mapping with an RGB-D Camera]{}. IEEE Transactions on Robotics **30**(1), 177–187 (2014) Engel, J., Koltun, V., Cremers, D.: [Direct Sparse Odometry]{}. IEEE transactions on pattern analysis and machine intelligence **40**(3), 611–625 (2018) F[á]{}cil, J.M., Ummenhofer, B., Zhou, H., Montesano, L., Brox, T., Civera, J.: [CAM-Convs: Camera-Aware Multi-Scale Convolutions for Single-View Depth]{}. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2019) G[á]{}lvez-L[ó]{}pez, D., Tardos, J.D.: [Bags of Binary Words for Fast Place Recognition in Image Sequences]{}. IEEE Transactions on Robotics **28**(5), 1188–1197 (2012) Gao, X., Zhang, T.: [[Robust [RGB-D]{} Simultaneous Localization and Mapping using Planar Point Features]{}]{}. Robotics and Autonomous Systems **72**, 1 – 14 (2015) Gee, A.P., Mayol-Cuevas, W.W.: [6D Relocalisation for RGBD Cameras Using Synthetic View Regression.]{} In: BMVC (2012) Glocker, B., Shotton, J., Criminisi, A., Izadi, S.: [Real-time RGB-D Camera Relocalization via Randomized Ferns for Keyframe Encoding]{}. IEEE transactions on visualization and computer graphics **21**(5), 571–583 (2015) Gutierrez-Gomez, D., Guerrero, J.J.: [RGBiD-SLAM for Accurate Real-Time Localisation and 3D Mapping]{}. arXiv preprint arXiv:1807.08271 (2018) Guti[é]{}rrez-G[ó]{}mez, D., Mayol-Cuevas, W., Guerrero, J.J.: [Inverse Depth for Accurate Photometric and Geometric Error Minimisation in RGB-D Dense Visual Odometry]{}. In: Robotics and Automation (ICRA), 2015 IEEE International Conference on, pp. 83–89. IEEE (2015) Gutierrez-Gomez, D., Mayol-Cuevas, W., Guerrero, J.J.: [Dense RGB-D Visual Odometry using Inverse Depth]{}. Robotics and Autonomous Systems **75**, 571–583 (2016) Guzman-Rivera, A., Kohli, P., Glocker, B., Shotton, J., Sharp, T., Fitzgibbon, A., Izadi, S.: [Multi-Output Learning for Camera Relocalization]{}. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 1114–1121 (2014) Handa, A., Whelan, T., McDonald, J., Davison, A.J.: [A Benchmark for RGB-D Visual Odometry, 3D Reconstruction and SLAM]{}. In: Robotics and automation (ICRA), 2014 IEEE international conference on, pp. 1524–1531. IEEE (2014) Harris, C., Stephens, M.: [A Combined Corner and Edge Detector]{}. In: In Proc. of Fourth Alvey Vision Conference, pp. 147–151 (1988) He, K., Gkioxari, G., Doll[á]{}r, P., Girshick, R.: [Mask R-CNN]{}. In: Proceedings of the IEEE international conference on computer vision, pp. 2961–2969 (2017) Henry, P., Krainin, M., Herbst, E., Ren, X., Fox, D.: [RGB-D Mapping: Using Depth Cameras for Dense 3D Modeling of Indoor Environments]{}. In: In the 12th International Symposium on Experimental Robotics (ISER. Citeseer (2010) Hermans, A., Floros, G., Leibe, B.: [Dense 3D Semantic Mapping of Indoor Scenes from RGB-D Images]{}. In: Robotics and Automation (ICRA), 2014 IEEE International Conference on, pp. 2631–2638. IEEE (2014) Houseago, C., Bloesch, M., Leutenegger, S.: [KO-Fusion: Dense Visual SLAM with Tightly-Coupled Kinematic and Odometric Tracking]{}. In: 2019 IEEE International Conference on Robotics and Automation (ICRA). IEEE (2019) Hsiao, M., Westman, E., Kaess, M.: [Dense Planar-Inertial SLAM with Structural Constraints]{}. In: 2018 IEEE International Conference on Robotics and Automation (ICRA), pp. 6521–6528. IEEE (2018) Hsiao, M., Westman, E., Zhang, G., Kaess, M.: [Keyframe-Based Dense Planar [SLAM]{}]{}. In: IEEE Intl. Conf. on Robotics and Automation, ICRA (2017) Huang, A.S., Bachrach, A., Henry, P., Krainin, M., Maturana, D., Fox, D., Roy, N.: [Visual Odometry and Mapping for Autonomous Flight using an RGB-D Camera]{}. In: International Symposium of Robotics Research. Springer (2011) Huber, P.J.: [Robust Statistics]{}. Springer (2011) : [Scale-Aware Navigation of a Low-Cost Quadrocopter with a Monocular Camera]{}. Robotics and Autonomous Systems (RAS) **62**(11), 1646–1656 (2014) Jaimez, M., Kerl, C., Gonzalez-Jimenez, J., Cremers, D.: [Fast Odometry and Scene Flow from RGB-D Cameras Based on Geometric Clustering]{}. In: 2017 IEEE International Conference on Robotics and Automation (ICRA), pp. 3992–3999. IEEE (2017) K. Madsen H.B.Nielsen, O.T.: [Methods for Non-Linear Least Squares Problems (2nd ed.)]{} p. 60 (2004) Kaess, M.: [Simultaneous Localization and Mapping with Infinite Planes]{}. In: [IEEE]{} International Conference on Robotics and Automation, pp. 4605–4611 (2015) Keller, M., Lefloch, D., Lambers, M., Izadi, S., Weyrich, T., Kolb, A.: [Real-time 3D Reconstruction in Dynamic Scenes using Point-Based Fusion]{}. In: 2013 International Conference on 3D Vision-3DV 2013, pp. 1–8. IEEE (2013) Kerl, C., Stuckler, J., Cremers, D.: [Dense Continuous-Time Tracking and Mapping with Rolling Shutter RGB-D Cameras]{}. In: Proceedings of the IEEE international conference on computer vision, pp. 2264–2272 (2015) Kerl, C., Sturm, J., Cremers, D.: [Dense Visual [SLAM]{} for [RGB-D]{} Cameras]{}. In: Intelligent Robots and Systems (IROS), 2013 IEEE/RSJ International Conference on, pp. 2100–2106. Citeseer (2013) Kerl, C., Sturm, J., Cremers, D.: [Robust Odometry Estimation for RGB-D Cameras]{}. In: Robotics and Automation (ICRA), 2013 IEEE International Conference on, pp. 3748–3754. IEEE (2013) Kim, C., Kim, P., Lee, S., Kim, H.J.: [Edge-Based Robust [RGB-D]{} Visual Odometry Using 2-D Edge Divergence Minimization]{}. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 6887–6894 (2018) Kim, P., Coltin, B., Kim, H.J.: [Linear [RGB-D SLAM]{} for Planar Environments]{}. In: Computer Vision – ECCV 2018, pp. 350–366 (2018) , P., [Coltin]{}, B., [Kim]{}, H.J.: [Low-Drift Visual Odometry in Structured Environments by Decoupling Rotational and Translational Motion]{}. In: 2018 IEEE International Conference on Robotics and Automation (ICRA), pp. 7247–7253 (2018) Klein, G., Murray, D.: [Parallel Tracking and Mapping for Small AR Workspaces]{}. In: Mixed and Augmented Reality, 2007. ISMAR 2007. 6th IEEE and ACM International Symposium on, pp. 225–234. IEEE (2007) Klingensmith, M., Dryanovski, I., Srinivasa, S., Xiao, J.: [Chisel: Real Time Large Scale 3D Reconstruction Onboard a Mobile Device using Spatially Hashed Signed Distance Fields.]{} In: Robotics: science and systems, vol. 4 (2015) Klingensmith, M., Sirinivasa, S.S., Kaess, M.: [Articulated Robot Motion for Simultaneous Localization and Mapping (ARM-SLAM)]{}. IEEE robotics and automation letters **1**(2), 1156–1163 (2016) Klose, S., Heise, P., Knoll, A.: [Efficient Compositional Approaches for Real-Time Robust Direct Visual Odometry from RGB-D Data]{}. In: Intelligent Robots and Systems (IROS), 2013 IEEE/RSJ International Conference on, pp. 1100–1106. IEEE (2013) K[ü]{}mmerle, R., Grisetti, G., Strasdat, H., Konolige, K., Burgard, W.: [[g]{}2o: A General Framework for Graph Optimization]{}. In: IEEE Intl. Conf. on Robotics and Automation (ICRA), pp. 3607–3613 (2011) , M., [Shaojie Shen]{}: [Robust Camera Motion Estimation using Direct Edge Alignment and Sub-Gradient Method]{}. In: 2016 IEEE International Conference on Robotics and Automation (ICRA), pp. 573–579 (2016) Laidlow, T., Bloesch, M., Li, W., Leutenegger, S.: [Dense RGB-D-Inertial SLAM with Map Deformations]{}. In: Intelligent Robots and Systems (IROS), 2017 IEEE/RSJ International Conference on, pp. 6741–6748. IEEE (2017) , P., Košecka, J.: [Dense Piecewise Planar [RGB-D SLAM]{} for Indoor Environments]{}. In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 4944–4949 (2017) , S.H., [de Croon]{}, G.: [Stability-Based Scale Estimation for Monocular SLAM]{}. IEEE Robotics and Automation Letters **3**(2), 780–787 (2018) Li, W., Saeedi, S., McCormac, J., Clark, R., Tzoumanikas, D., Ye, Q., Huang, Y., Tang, R., Leutenegger, S.: [InteriorNet: Mega-Scale Multi-sensor Photo-realistic Indoor Scenes Dataset]{}. In: British Machine Vision Conference (BMVC) (2018) Lowe, D.G.: [Distinctive Image Features from Scale-invariant Keypoints]{}. International journal of computer vision **60**(2), 91–110 (2004) , Y., [Song]{}, D.: [Robust [RGB-D]{} Odometry Using Point and Line Features]{}. In: 2015 IEEE International Conference on Computer Vision (ICCV), pp. 3934–3942 (2015) , Y., [Song]{}, D.: [Robustness to Lighting Variations: An [RGB-D]{} Indoor Visual Odometry using Line Segments]{}. In: 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 688–694 (2015) Lucas, B.D., Kanade, T.: [An Iterative Image Registration Technique with an Application to Stereo Vision]{}. In: Proceedings of the 7th International Joint Conference on Artificial Intelligence - Volume 2, pp. 674–679 (1981) , L., [Kerl]{}, C., [St[ü]{}ckler]{}, J., [Cremers]{}, D.: [[CPA-SLAM]{}: Consistent Plane-Model Alignment for Direct [RGB-D SLAM]{}]{}. In: IEEE International Conference on Robotics and Automation (ICRA), pp. 1285–1291 (2016) Ma, L., St[ü]{}ckler, J., Kerl, C., Cremers, D.: [Multi-view Deep Learning for Consistent Semantic Mapping with RGB-D Cameras]{}. In: Intelligent Robots and Systems (IROS), 2017 IEEE/RSJ International Conference on, pp. 598–605. IEEE (2017) McCormac, J., Handa, A., Davison, A., Leutenegger, S.: [SemanticFusion: Dense 3D Semantic Mapping with Convolutional Neural Networks]{}. In: 2017 IEEE International Conference on Robotics and Automation (ICRA), pp. 4628–4635. IEEE (2017) Meilland, M., Comport, A.I.: [On Unifying Key-Frame and Voxel-Based Dense Visual SLAM at Large Scales]{}. In: 2013 IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 3677–3683. IEEE (2013) Meilland, M., Comport, A.I.: [Super-Resolution 3D Tracking and Mapping]{}. In: 2013 IEEE International Conference on Robotics and Automation, pp. 5717–5723. IEEE (2013) Mur-Artal, R., Montiel, J.M.M., Tard[ó]{}s, J.D.: [ORB-SLAM: a Versatile and Accurate Monocular SLAM System]{}. IEEE transactions on robotics **31**(5), 1147–1163 (2015) Mur-Artal, R., Tard[ó]{}s, J.D.: [ORB-SLAM2: An Open-source SLAM System for Monocular, Stereo, and RGB-D Cameras]{}. IEEE Transactions on Robotics **33**(5), 1255–1262 (2017) Newcombe, R.A., Fox, D., Seitz, S.M.: [DynamicFusion: Reconstruction and Tracking of Non-Rigid Scenes in Real-Time]{}. In: Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 343–352 (2015) Newcombe, R.A., Izadi, S., Hilliges, O., Molyneaux, D., Kim, D., Davison, A.J., Kohi, P., Shotton, J., Hodges, S., Fitzgibbon, A.: [KinectFusion: Real-Time Dense Surface Mapping and Tracking]{}. In: Mixed and augmented reality (ISMAR), 2011 10th IEEE international symposium on, pp. 127–136. IEEE (2011) Pham, T.T., Reid, I., Latif, Y., Gould, S.: [Hierarchical Higher-Order Regression Forest Fields: An Application to 3D Indoor Scene Labelling]{}. In: Proceedings of the IEEE International Conference on Computer Vision, pp. 2246–2254 (2015) Pire, T., Fischer, T., Castro, G., De Crist[ó]{}foris, P., Civera, J., Berlles, J.J.: [S-PTAM: Stereo Parallel Tracking and Mapping]{}. Robotics and Autonomous Systems **93**, 27–42 (2017) Platinsky, L., Davison, A.J., Leutenegger, S.: [Monocular Visual Odometry: Sparse Joint Optimisation or Dense Alternation?]{} In: 2017 IEEE International Conference on Robotics and Automation (ICRA), pp. 5126–5133. IEEE (2017) Proença, P.F., Gao, Y.: [[Probabilistic [RGB-D]{} Odometry Based on Points, Lines and Planes under Depth Uncertainty]{}]{}. Robotics and Autonomous Systems **104**, 25 – 39 (2018) Raposo, C., Lourenço, M., Barreto, J.P., Antunes, M.: [Plane-based Odometry using an [RGB-D]{} Camera]{}. In: In British Machine Vision Conference (BMVC (2013) Rosen, D.M., Carlone, L., Bandeira, A.S., Leonard, J.J.: [SE-Sync: [A]{} Certifiably Correct Algorithm for Synchronization over the Special Euclidean Group]{}. I. J. Robotics Res. **38**(2-3) (2019) Rosten, E., Drummond, T.: [Machine Learning for High-Speed Corner Detection]{}. In: European conference on computer vision, pp. 430–443. Springer (2006) Rublee, E., Rabaud, V., Konolige, K., Bradski, G.: [[ORB]{}: An Efficient Alternative to [SIFT]{} or [SURF]{}]{}. In: Proceedings of the IEEE international conference on computer vision, pp. 2564–2571 (2011) R[ü]{}nz, M., Buffier, M., Agapito, L.: [MaskFusion: Real-time Recognition, Tracking and Reconstruction of Multiple Moving Objects]{}. In: 2018 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pp. 10–20. IEEE (2018) Salas[-]{}Moreno, R.F., Glocker, B., Kelly, P.H.J., Davison, A.J.: [Dense Planar [SLAM]{}]{}. In: [IEEE]{} International Symposium on Mixed and Augmented Reality, [ISMAR]{}, pp. 157–164 (2014) Salas-Moreno, R.F., Newcombe, R.A., Strasdat, H., Kelly, P.H., Davison, A.J.: [SLAM++: Simultaneous Localisation and Mapping at the Level of Objects]{}. In: Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 1352–1359 (2013) Schenk, F., Fraundorfer, F.: [Combining Edge Images and Depth Maps for Robust Visual Odometry]{}. In: British Machine Vision Conference (2017) Schops, T., Sattler, T., Pollefeys, M.: [BAD SLAM: Bundle Adjusted Direct RGB-D SLAM]{}. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 134–144 (2019) Scona, R., Jaimez, M., Petillot, Y.R., Fallon, M., Cremers, D.: [StaticFusion: Background Reconstruction for Dense RGB-D SLAM in Dynamic Environments]{}. In: 2018 IEEE International Conference on Robotics and Automation (ICRA), pp. 1–9. IEEE (2018) Scona, R., Nobili, S., Petillot, Y.R., Fallon, M.: [Direct Visual SLAM Fusing Proprioception for a Humanoid Robot]{}. In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 1419–1426. IEEE (2017) Shi, J., Tomasi, C.: [Good Features to Track]{}. In: 1994 Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pp. 593–600 (1994) Shi, Y., Xu, K., Niessner, M., Rusinkiewicz, S., Funkhouser, T.: [PlaneMatch: Patch Coplanarity Prediction for Robust RGB-D Reconstruction]{}. arXiv preprint arXiv:1803.08407 (2018) Shotton, J., Glocker, B., Zach, C., Izadi, S., Criminisi, A., Fitzgibbon, A.: [Scene Coordinate Regression Forests for Camera Relocalization in RGB-D Images]{}. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2930–2937 (2013) Sivic, J., Zisserman, A.: [Video Google: A Text Retrieval Approach to Object Matching in Videos]{}. In: Proceedings of the Ninth IEEE International Conference on Computer Vision, p. 1470. IEEE (2003) Sol[à]{}, J., Deray, J., Atchuthan, D.: [A Micro Lie Theory for State Estimation in Robotics]{}. arXiv preprint arXiv:1812.01537 (2018) , F., [Kerl]{}, C., [Cremers]{}, D., [Sturm]{}, J.: [Large-Scale Multi-resolution Surface Reconstruction from [RGB-D]{} Sequences]{}. In: 2013 IEEE International Conference on Computer Vision, pp. 3264–3271 (2013) Steinbr[ü]{}cker, F., Sturm, J., Cremers, D.: [Real-time Visual Odometry from Dense RGB-D Images]{}. In: Computer Vision Workshops (ICCV Workshops), 2011 IEEE International Conference on, pp. 719–722. IEEE (2011) Strasdat, H.: [Local Accuracy and Global Consistency for Efficient Visual SLAM]{}. Ph.D. thesis, Department of Computing, Imperial College London (2012) Strasdat, H., Montiel, J., Davison, A.J.: [Scale Drift-Aware Large Scale Monocular SLAM]{}. Robotics: Science and Systems VI **2**(3), 7 (2010) St[ü]{}ckler, J., Waldvogel, B., Schulz, H., Behnke, S.: [Dense Real-Time Mapping of Object-Class Semantics from RGB-D Video]{}. Journal of Real-Time Image Processing **10**(4), 599–609 (2015) Sturm, J., Engelhard, N., Endres, F., Burgard, W., Cremers, D.: [A Benchmark for the Evaluation of RGB-D SLAM Systems]{}. In: Proc. of the International Conference on Intelligent Robot Systems (IROS) (2012) S[ü]{}nderhauf, N., Pham, T.T., Latif, Y., Milford, M., Reid, I.: [Meaningful Maps with Object-Oriented Semantic Mapping]{}. In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 5079–5085. IEEE (2017) , Y., [Jian]{}, Y., [Ramalingam]{}, S., [Feng]{}, C.: [Point-Plane SLAM for Hand-Held 3D Sensors]{}. In: 2013 IEEE International Conference on Robotics and Automation, pp. 5182–5189 (2013) Tateno, K., Tombari, F., Laina, I., Navab, N.: [[CNN-SLAM]{}: [R]{}eal-Time Dense Monocular [SLAM]{} with Learned Depth Prediction]{}. In: IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR) (2017) Tateno, K., Tombari, F., Navab, N.: [Real-Time and Scalable Incremental Segmentation on Dense SLAM]{}. In: 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 4465–4472. IEEE (2015) Triggs, B., McLauchlan, P.F., Hartley, R.I., Fitzgibbon, A.W.: [Bundle Adjustment - A Modern Synthesis]{}. In: International workshop on vision algorithms, pp. 298–372. Springer (1999) Valentin, J., Nie[ß]{}ner, M., Shotton, J., Fitzgibbon, A., Izadi, S., Torr, P.H.: [Exploiting Uncertainty in Regression Forests for Accurate Camera Relocalization]{}. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 4400–4408 (2015) Wang, X., Dong, W., Zhou, M., Li, R., Zha, H.: [Edge Enhanced Direct Visual Odometry]{}. In: Proceedings of the British Machine Vision Conference (BMVC), pp. 35.1–35.11 (2016) , Y., [Zhang]{}, Q., [Zhou]{}, Y.: [[RGB-D]{} Mapping for Indoor Environment]{}. In: 2014 9th IEEE Conference on Industrial Electronics and Applications, pp. 1888–1892 (2014) Whelan, T., Johannsson, H., Kaess, M., Leonard, J.J., McDonald, J.: [Robust Real-Time Visual Odometry for Dense RGB-D Mapping]{}. In: Robotics and Automation (ICRA), 2013 IEEE International Conference on, pp. 5724–5731. IEEE (2013) Whelan, T., Kaess, M., Johannsson, H., Fallon, M., Leonard, J.J., McDonald, J.: [Real-Time Large-Scale Dense RGB-D SLAM with Volumetric Fusion]{}. The International Journal of Robotics Research **34**(4-5), 598–626 (2015) Whelan, T., Leutenegger, S., Salas-Moreno, R.F., Glocker, B., Davison, A.J.: [[ElasticFusion]{}: Dense [SLAM]{} Without A Pose Graph]{}. In: Robotics: Science and Systems (RSS) (2015) Whelan, T., Salas-Moreno, R.F., Glocker, B., Davison, A.J., Leutenegger, S.: [ElasticFusion: Real-time dense SLAM and light source estimation]{}. The International Journal of Robotics Research **35**(14), 1697–1716 (2016) Yokozuka, M., Oishi, S., Thompson, S., Banno, A.: [VITAMIN-E: VIsual Tracking And MappINg with Extremely Dense Feature Points]{}. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 9641–9650 (2019) Zhang, Z.: [Parameter Estimation Techniques: A Tutorial with Application to Conic Fitting]{}. Image and vision Computing **15**, 59–76 (1997) Zhou, H., Ummenhofer, B., Brox, T.: [DeepTAM: Deep Tracking and Mapping]{}. In: Proceedings of the European Conference on Computer Vision (ECCV), pp. 822–838 (2018) Zhou, Q.Y., Koltun, V.: [Color Map Optimization for 3D Reconstruction with Consumer Depth Cameras]{}. ACM Transactions on Graphics (TOG) **33**(4), 155 (2014) Zhou, Y., Li, H., Kneip, L.: [Canny-VO: Visual Odometry with RGB-D Cameras Based on Geometric 3-D–2-D Edge Alignment]{}. IEEE Transactions on Robotics **35**(1), 184–199 (2019) Zollh[ö]{}fer, M., Stotko, P., G[ö]{}rlitz, A., Theobalt, C., Nie[ß]{}ner, M., Klein, R., Kolb, A.: [State of the Art on 3D Reconstruction with RGB-D Cameras]{}. In: Computer Graphics Forum, vol. 37, pp. 625–652. Wiley Online Library (2018) Zubizarreta, J., Aguinaga, I., Montiel, J.: [Direct Sparse Mapping]{}. arXiv preprint arXiv:1904.06577 (2019) [^1]: This work was partially supported by the Spanish government (project PGC2018-096367-B-I00) and the Aragón regional government (Grupo DGA-T45 17R/FSE).

ArXiv

--- address: ' *Institute of Mathematics, TU Berlin, Stra[ß]{}e des 17. Juni 136, 10623 Berlin, Germany*' author: - Thibaut Lux - Antonis Papapantoleon bibliography: - 'bib.bib' title: 'Model-free bounds on Value-at-Risk using partial dependence information' --- Introduction ============ The evaluation of multivariate risks under model uncertainty has become a central issue in several applications, ranging from hydrology and engineering to mathematical finance. In mathematical finance, this has been in parts driven by the changing regulations requiring the quantification of model uncertainty in risk management; see *e.g.* the latest Basel Accord. Measuring risk under uncertainty often relates to the computation of bounds on probabilities of the form $\mathbb{P}(\varphi(\mathbf X)\le \cdot)$, where $\mathbf X = (X_1,\dots,X_d)$ is an $\mathbb{R}^d$-valued random vector and $\varphi\colon\mathbb{R}^d\to\mathbb{R}$ an aggregation function. Here $\mathbf X$ can be thought of as a vector modeling $d$ risks in a portfolio and $\varphi$ as a function to aggregate these risks. In this paper we focus on risk measurement under *dependence uncertainty*, hence we assume that the marginal distributions of the constituents $X_i\sim F_i$ for $i=1,\dots,d$ are known, while the dependence structure between the components of $\mathbf{X}$ is unknown or only partially known. We then derive bounds on the distribution function of $\varphi(\mathbf X)$ using the available information on the distribution of $\mathbf X$. Then, by inversion, the bounds on the distribution of $\varphi(\mathbf X)$ can be translated immediately into bounds on the Value-at-Risk (VaR) of $\varphi(\mathbf X)$. A significant part of the related literature focuses on the situation where only the marginals $F_1,\dots,F_d$ are known and no information on the dependence structure of $\mathbf{X}$ is available. In this case, explicit bounds on the distribution function of the sum of two random variables, *i.e.* $\varphi(\mathbf{X}) = X_1+X_2$, were derived by @makarov1981 and for more general functions $\varphi$ by @rueschendorf1981 in the early 1980’s. These results were later generalized for functions of more than two random variables, for instance by @denuit1999 for the sum and by @embrechts2003 and @embrechts2006 for more general aggregation functions; see also @Cheung_Lo_2013. These bounds however may fail to be sharp. Therefore, numerical schemes to compute sharp distributional bounds have become increasingly popular. The rearrangement algorithm, which was introduced by @puccetti2012 and @embrechts2013, represents an efficient method to approximate sharp bounds on the VaR of the sum $X_1+\cdots+X_d$ under additional requirements on the marginal distributions $F_1,\dots,F_d$. However, the complete absence of information on the dependence structure typically leads to very wide bounds that are not sufficiently informative for practical applications. This calls for methods to account for additional information on the dependence structure in the computation of risk bounds. Several analytical and numerical approaches to derive risk bounds including additional dependence information have recently been developed. Analytical bounds were derived by @embrechts2003 and @embrechts2006 for the case that a lower bound on the copula of $\mathbf{X}$ is given. Moreover, @embrechts2010 and @puccetti2012b established bounds when the laws of some lower dimensional marginals of $\mathbf{X}$ are known. Analytical bounds that account for positive or negative dependence assumptions were presented in @embrechts2003 and @rueschendorf2005. @bernard2015b derived risk bounds when an upper bound on the variance of $\varphi(\mathbf X)$ is prescribed, and presented a numerical scheme to efficiently compute these bounds. In addition, @bernard2015 considered the case where the distribution of $\mathbf X$ is known only on a subset of its domain and established a version of the rearrangement algorithm to account for this type of dependence information. A detailed account of this literature can be found in @rueschendorf2016. In this paper we develop alternative approaches to compute VaR bounds for aggregations of multiple risks in the presence of dependence uncertainty. After recalling several definitions and usefull results in Setion \[setting\], in Section \[boundsOnVar\] we revisit the standard and improved standard bounds on VaR and provide a direct derivation of the improved standard bounds when $\varphi = \max$ or $\varphi = \min$. In Section \[prescribedMax\] we develop a reduction principle to account for extreme value information, such as the distribution of partial minima or maxima of the risk vector $\mathbf X$, in the computation of risk bounds for the sum $X_1+\cdots+X_d$. The term partial maxima hereby refers to the maximum of lower dimensional marginals of $\mathbf X$, *i.e.* $\max\{X_{i_1},\dots,X_{i_n}\}$ for $1\leq i_1\leq\cdots\leq i_n\leq d$, and analogously for the minimum. We thereby interpolate between the marginals-only case and the situation where the distributions of the lower-dimensional marginals of $\mathbf X$ are completely specified; *cf.* [@embrechts2010; @puccetti2012b]. In Section \[boundsOnCopula\] we present an approach to compute VaR bounds for general aggregation functions $\varphi$ including two different types of dependence information. First, we consider the situation where the copula $C$ of the risk vector $\mathbf X$ coincides with a reference model on a subset $\set$ of its domain, *i.e.* it holds that $C(\bx) = C^*(\bx)$ for all $\bx\in\set$ and a reference copula $C^*$. Applying results from @lux2016 and the improved standard bounds of @embrechts2003 and @embrechts2006 we derive bounds on VaR using the available information on the subset $\set$. This relates to the *trusted region* in @bernard2015, although the methods are different. The second type of dependence information corresponds to $C$ lying in the vicinity of a reference copula $C^*$ as measured by a statistical distance $\cD$. In this case we establish improved bounds on the set of all (quasi-)copulas $C$ in the $\delta$-neighborhood of the reference model $C^*$, *i.e.* for all $C$ such that $\cD(C,C^*)\leq\delta$. Our method applies to a large class of statistical distances such as the Kolmogorov–Smirnov or the Cramér–von Mises distances. We then use the improved standard bounds of [@embrechts2003; @embrechts2006] in order to translate the improved bounds into bounds on the VaR of $\varphi(\mathbf X)$. Finally, in Section \[numerics\] we present several applications of our results in risk measurement. The computational results show that the additional dependence information typically leads to a significant improvement of the VaR bounds when compared to the marginals-only case. Moreover, the VaR bounds using information on the partial maxima are becoming tighter as the confidence level increases, which is in contrast to related results in the literature, and constitutes an advantage of this methodology. Notation and preliminary results {#setting} ================================ In this section we introduce the notation and some basic results that will be used throughout this work. Let $d\geq2$ be an integer. In the following $\mathbb{I}$ denotes the unit interval $[0,1]$, while boldface letters, *e.g.* $\mathbf{u}$, $\mathbf{v}$ or $\mathbf{x}$, denote vectors in $\mathbb{I}^d$ or $\mathbb{R}^d$. Moreover, $\mathbf 1$ denotes the $d$-dimensional vector with all entries equal to one, *i.e.* $\mathbf 1=(1,\dots,1)$. The finite difference operator $\Delta$ for a function $f\colon\mathbb{R}^d\to\mathbb{R}$ and $a\le b\in\mathbb{R}$ is defined via $$\Delta_{a,b}^i\ f(x_1,\dots,x_d) := f(x_1,\dots,x_{i-1},b,x_{i+1},\dots,x_d) - f(x_1,\dots,x_{i-1},a,x_{i+1},\dots,x_d).$$ A function $f\colon\mathbb{R}^d\to\mathbb{R}$ is called $d$-*increasing* if for all rectangular subsets $H = (a_1,b_1]\times\cdots\times(a_d,b_d]\subset\mathbb{R}^d$ it holds that $$\begin{aligned} \label{volume} V_f(H) := \Delta^d_{a_d,b_d}\circ\cdots\circ\Delta^1_{a_1,b_1}\ f \geq 0.\end{aligned}$$ We call $V_f(H)$ the $f$-*volume* of $H$. A function $Q\colon\mathbb{I}^d\to\mathbb{I}$ is a $d$-*quasi-copula* if the following properties hold: 1. $Q$ satisfies, for all $i\in\{1,\dots,d\}$, the boundary conditions \[cond:QC1\] $$Q(u_1,\dots,u_i = 0,\dots,u_d)=0 \quad\text{ and }\quad Q(1,\dots,1,u_i,1,\dots,1) = u_i.$$ <!-- --> 1. $Q$ is non-decreasing in each argument. \[cond:QC2\] <!-- --> 1. $Q$ is Lipschitz continuous, *i.e.* for all $\mathbf{u},\mathbf{v}\in\mathbb{I}^d$\[cond:QC3\] $$|Q(u_1,\dots,u_d)-Q(v_1,\dots,v_d)|\leq\sum_{i=1}^d |u_i-v_i|.$$ Moreover, $Q$ is a $d$-*copula* if 1. $Q$ is $d$-increasing. \[cond:QC4\] We denote the set of all $d$-quasi-copulas by $\mathcal{Q}^d$ and the set of all $d$-copulas by $\mathcal{C}^d$. Obviously $\mathcal{C}^d\subset\mathcal{Q}^d$. Henceforth we will simply refer to a $d$-(quasi-)copula as a (quasi-)copula if the dimension is clear from the context. Let $C$ be a $d$-copula and consider $d$ univariate probability distribution functions $F_1,\dots,F_d$. Then $F(x_1,\dots,x_d):=C(F_1(x_1),\dots,F_d(x_d))$, for all $\mathbf{x}\in\mathbb{R}^d$, defines a $d$-dimensional distribution function with univariate margins $F_1,\dots,F_d$. The converse also holds by Sklar’s Theorem, see @sklar1959, *i.e.* for each $d$-dimensional distribution function $F$ with univariate marginals $F_1,\dots,F_d$, there exists a copula $C$ such that $F(x_1,\dots,x_d) = C(F_1(x_1),\dots,F_d(x_d))$ for all $\mathbf{x}\in\mathbb{R}^d$. In this case, the copula $C$ is unique if the marginals are continuous. A simple and elegant proof of Sklar’s Theorem based on the distributional transform can be found in @rueschendorf2009. Sklar’s Theorem establishes a fundamental link between copulas and multivariate distribution functions. Thus, given a random vector we will refer to its copula, *i.e.* the copula corresponding to the distribution function of this random vector. The *survival function* of a $d$-copula $C$ is defined as follows: $$\widehat{C}(u_1,\dots,u_d) := V_C([u_1,1]\times\cdots\times[u_d,1]),\quad \mathbf{u} \in \mathbb{I}^d.$$ This is illustrated for $d=3$ below: $$\begin{aligned} \widehat{C}(u_1,u_2,u_3) &= 1 - C(u_1,1,1,) - C(1,u_2,1) - C(1,1,u_3)\\ &\quad + C(u_1,u_2,1) + C(u_1,1,u_3) + C(1,u_2,u_3) - C(u_1,u_2,u_3).\end{aligned}$$ The function $\widehat{C}(1-u_1,\dots,1-u_d)$, for $\mathbf{u} \in \mathbb{I}^d$, is again a copula, namely the *survival copula* of $C$; see *e.g.* @georges2001. Note that for a distribution function $F$ of a random vector $(X_1,\dots,X_d)$ with marginals $F_1,\dots,F_d$ and a corresponding copula $C$ such that $F(x_1,\dots,x_d) = C(F_1(x_1),\dots,F_d(x_d))$ it holds that $$\begin{aligned} \label{survivalCopulaProbability} \mathbb{P}(X_1>x_1,\dots,X_d>x_d) = \widehat{C}(F_1(x_1),\dots,F_d(x_d)).\end{aligned}$$ The map $\widehat{Q}$ could be defined analogously for quasi-copulas $Q$, however the function $\widehat{Q}(1-u_1,\dots,1-u_d)$ is not necessarily again a quasi-copula. Therefore, we introduce the term *quasi-survival functions* to refer to functions $\widehat{Q}:\mathbb{I}^d\to\mathbb{I}$ such that $(u_1,\dots,u_d)\mapsto\widehat{Q}(1-u_1,\dots,1-u_d)$ is again a quasi-copula. The set of $d$-quasi-survival functions is denoted by $\widehat{\mathcal{Q}}^d$. Let $Q,Q'$ be $d$-quasi-copulas. $Q'$ is greater than $Q$ in the *lower orthant order*, denoted by $Q \preceq Q'$, if $Q(\mathbf{u})\leq Q'(\mathbf{u})$ for all $\mathbf{u}\in\mathbb{I}^d$. The well-known Fréchet–Hoeffding theorem establishes the minimal and maximal bounds on the set of quasi-copulas in the lower orthant order. In particular, for each $Q\in\mathcal{Q}^d$, it holds that $$W_d(\mathbf{u}) := \max\Big\{0,\sum_{i=1}^d u_i - d + 1\Big\} \leq Q(\mathbf{u}) \leq \min\{u_1,\dots,u_d\} =: M_d(\mathbf{u}),$$ for all $\mathbf{u}\in\mathbb{I}^d$, *i.e.* $W_d\preceq Q\preceq M_d$, where $W_d$ and $M_d$ are the lower and upper Fréchet–Hoeffding bounds respectively. The properties of the bounds carry over to the set of survival copulas in a straightforward way, hence one obtains similarly for any $C\in\mathcal{C}^d$ the following bounds: $$W_d(\mathbf 1 - \bu) \le \widehat{C}(\bu) \le M_d(\mathbf 1 - \bu), \qquad\text{for all }\mathbf{u}\in\mathbb{I}^d.$$ Bounds on Value-at-Risk under partial dependence information: an overview and some new results {#boundsOnVar} ============================================================================================== In this section we consider a vector of risks $\mathbf{X}=(X_1,\dots,X_d)$ and an aggregation function $\varphi\colon\mathbb{R}^d\to\mathbb{R}$, and want to quantify the risk of $\varphi(\mathbf{X})$ by means of Value-at-Risk. This corresponds to the quantile function of $\varphi(\mathbf{X})$, *i.e.* when $\varphi(\mathbf{X}) \sim F_\varphi$ then the VaR of $\varphi(\mathbf{X})$ for a certain confidence level $\alpha\in(0,1)$ is given by the quantity $$\mathrm{VaR}_\alpha(\varphi(\mathbf{X})) = F_\varphi^{-1}(\alpha) = \inf\{x\in\mathbb{R}\colon F_\varphi(x)>\alpha\}.$$ Typical levels of $\alpha$ are close to 1 in practice, and the most commonly considered risk aggregation function $\varphi$ is the sum of the individual risks $X_1+\cdots+X_d$, while the maximum and minimum of the risks, $\max\{X_1,\dots,X_d\}$ and $\min\{X_1,\dots,X_d\}$, are also relevant choices for applications. Once the distribution of $\varphi(\mathbf X)$ is specified, the determination of VaR amounts to a simple computation of the quantile function. If the distribution $F_\varphi$ is not known, but instead the joint law of $\mathbf{X}$ is known, then the problem renders into the computation of the distribution of $\varphi(\mathbf{X})$ from the joint law of $\mathbf{X}$. In order to solve this problem, one can resort to numerical integration techniques or Monte Carlo methods. For the important case $\varphi(\mathbf{X}) = X_1+\cdots+X_d$, two efficient algorithms to determine the law of the aggregated risk given the joint distribution of $\mathbf{X}$ were presented by @arbenz2011 [@arbenz2012]. However, in many situations the distribution of $\mathbf{X}$ is not fully specified or cannot be determined with sufficient precision. In particular when $d$ is large, the limited amount of data available in most applications makes it difficult to estimate the joint law of $\mathbf{X}$ accurately. Therefore, we consider the situation of *model uncertainty*, where the distribution of $\mathbf{X}$ is not fully specified. In particular, we focus on *dependence uncertainty*, where one assumes that the marginal distributions $X_i\sim F_i$ are known for $i=1,\dots,d$, but the dependence structure between the constituents of $\mathbf{X}$ is either unknown or only partially known. Using Sklar’s Theorem, every distribution of $\mathbf{X}$ that is consistent with the marginals $F_1,\dots,F_d$ can be expressed by means of a copula $C$ and the marginals, *i.e.* if $\mathbf{X}\sim F$ then $F(x_1,\dots,x_d) = C(F_1(x_1),\dots,F_d(x_d))$. This implies that dependence uncertainty is in fact uncertainty about the copula of $\mathbf{X}$. In this case, for most functionals $\varphi$ of interest, neither the distribution of $\varphi(\mathbf{X})$ can be determined completely, nor can its risk be calculated exactly. Indeed, each model for $\mathbf{X}$ that is consistent with the available information can produce a different risk estimate. Therefore, one is interested in deriving upper and lower bounds on the risk of $\varphi(\mathbf{X})$ over the set of distributions that comply with the given information. These bounds are then considered best or worst case estimates for the VaR of $\varphi(\mathbf{X})$, given the available information about the distribution of $\mathbf{X}$. This problem has a long history and many approaches to its solution for different types of dependence uncertainty have emerged. In the situation of complete dependence uncertainty, where only the marginals $F_1,\dots,F_d$ are known and one has no information about the copula of $\mathbf{X}$, bounds for the quantiles of the sum $X_1+\cdots+X_d$ were derived in a series of papers, starting with the results by @makarov1981 and @rueschendorf1982 for $d=2$, and their extensions for $d>2$ by @frank1987, @denuit1999 and @embrechts2003. These bounds are in the literature referred to as *standard bounds* and they are given by $$\begin{aligned} \label{standardBounds} \begin{split} \max\Big\{\sup_{\mathcal{U}(s)}\Big(F^-_1(u_1)+\sum_{i=2}^d F_i(u_i)\Big)-d+1,0\Big\} &\leq \P(X_1+\cdots+X_d<s)\\ &\qquad \leq\min\Big\{\inf_{\mathcal{U}(s)}\sum_{i=1}^d F^-_i(u_i),1\Big\}, \end{split}\end{aligned}$$ where $\mathcal{U}(s) = \{(u_1,\dots,u_d)\in\mathbb{R}^d\colon u_1+\cdots+u_d=s\}$ and $F_i^-$ denotes the left-continuous version of $F_i$. These bounds hold for all random variables $\mathbf{X}$ with margins $F_1,\dots,F_d$, and the corresponding bounds for the VaR of the sum $X_1+\cdots+X_d$ are given by the respective inverse functions. It was shown independently in [@makarov1981] and [@rueschendorf1982] that the bounds are sharp for $d=2$, in the sense that there exists a distribution for $\mathbf{X}$ such that the sum of its constituents attains the upper and lower bound. The standard bounds may however fail to be sharp in higher dimensions. It turns out that the absence of dependence information leads mostly to very wide bounds on the VaR of the aggregated risks; see *e.g.* @bernard2015. Hence, there is a large spread between the upper and the lower bound, such that they are not actually informative for practical applications. In addition, a complete lack of information about the dependence structure of $\mathbf{X}$ is often unrealistic, since quantities such as correlations or the values of the distribution function of $\mathbf{X}$ at certain points can be estimated with a sufficient degree of accuracy. Therefore, the quest for methods to improve the standard bounds by including additional dependence information has turned into a thriving area of mathematical research in recent years. @embrechts2003 and @embrechts2006 derived an improvement of the standard bounds that accounts for a lower bound on the copula of $\mathbf{X}$ or its survival function. This improvement is essential for the results in the present work, since it relates the problem of computing improved VaR bounds in the presence of additional dependence information to the task of improving the bounds on copulas. The improvement of the ‘classical’ bounds by using additional, partial information on the dependence structure has attracted some attention in the literature lately, see *e.g.* @nelsen2006, @tankov2011 and @lux2016. Let $\X$ be a random vector with marginals $F_1,\dots,F_d$ and copula $C$, let $\varphi\colon\mathbb{R}^d\to\mathbb{R}$ be non-decreasing in each coordinate, and define the functional $$\P_C(\varphi(\mathbf{X})<s) := \int_{\mathbb{R}^d} \mathds{1}_{\{\varphi(x_1,\dots,x_d)<s\}}\ \ud C(F_1(x_1),\dots,F_d(x_d)).$$ Let $C_0,C_1$ be copulas and consider the following quantities $$\begin{aligned} m_{C_0,\varphi}(s) &:= \inf\big\{\P_C(\varphi(\mathbf{X})<s) \colon C\in\mathcal C^d, C_0\preceq C\big\}, \\ M_{\widehat C_1,\varphi}(s) &:= \sup\big\{\P_C(\varphi(\mathbf{X})<s) \colon C\in\mathcal C^d, \widehat{C}_1\preceq \widehat{C}\big\}.\end{aligned}$$ The following bounds on $m_{C_0,\varphi}, M_{\widehat C_1,\varphi}$ are known in the literature as *improved standard bounds* and read as follows: $$\begin{aligned} \label{eq:ISB} \begin{split} m_{C_0,\varphi}(s) & \geq \sup_{\mathcal V^<_\varphi(s)}\ C_0\big(F_1(x_1),\dots,F_{d}(x_{d})\big) =: \underline{m}_{C_0,\varphi}(s),\\ M_{\widehat C_1,\varphi}(s) & \leq \inf_{\mathcal V^>_\varphi(s)}\ 1-\widehat{C}_1\big(F_1(x_1),\dots,F_{d}(x_{d})\big) =: \overline{M}_{\widehat C_1,\varphi}(s), \end{split}\end{aligned}$$ where $\mathcal V^<_\varphi(s) = \{ (x_1,\dots,x_d)\in\R^d: \varphi(\bx)< s \}$ and $\mathcal V^>_\varphi(s) = \{ (x_1,\dots,x_d)\in\R^d: \varphi(\bx)> s \}$; see [@embrechts2003; @embrechts2006]. A careful examination of the proof of Theorem 3.1 in @embrechts2006 reveals that these results hold also when $C_0$, resp. $\widehat{C}_1$, is just increasing, resp. decreasing, in each coordinate. Hence, they hold in particular when $C_0$ is a quasi-copula and $\widehat{C}_1$ a quasi-survival function. The above bounds relate to the VaR of $\varphi(\mathbf{X})$ in the following way. \[varBoundsRemark\] Let $\varphi\colon\mathbb{R}^d\to\mathbb{R}$ be increasing in each component and the copula $C$ of $\mathbf{X}$ be such that $Q_0\preceq C$ and $\widehat{Q}_1\preceq\widehat{C}$, for a quasi-copula $Q_0$ and a quasi-survival function $\widehat{Q}_1$. Then $$\overline{M}^{-1}_{\widehat Q_1,\varphi}(\alpha) \leq \mathrm{VaR}_\alpha(\varphi(\mathbf{X})) \leq \underline{m}^{-1}_{Q_0,\varphi}(\alpha).$$ Besides the aggregation function $\varphi(x_1,\dots,x_d) = x_1+\cdots+x_d$, the operations $\varphi(x_1,\dots,x_d) = \max\{x_1,\dots,x_d\}$ and $\varphi(x_1,\dots,x_d) = \min\{x_1,\dots,x_d\}$ are also of particular interest in risk management, however fewer methods to handle dependence uncertainty for these operations exist; *cf.* @embrechts2014. The following result establishes bounds for the minimum and maximum operations in the presence of additional information on the copula using straightforward computations, and further shows that these bounds coincide with the improved standard bounds . Analogous statements for $d=2$ in the absence of additional information on the copula $C$ can be found in @frank1987 [Theorem 5.1]. \[varBoundsMax\] Let $\mathbf X$ be a random vector with copula $C$ and marginals $F_1,\dots,F_d$, and let $\underline{Q},\overline{Q}$ be quasi-copulas. Then, for $\varphi(x_1,\dots,x_d)=\max\{x_1,\dots,x_d\}$, we have that $$\begin{aligned} m_{\underline{Q},\max}(s) &:= \inf\big\{\P_C(\varphi(\mathbf{X})<s)\colon \underline{Q}\preceq C\big\} \geq \underline{Q}(F_1(s),\dots,F_d(s)) = \underline{m}_{\underline{Q},\max}(s)\\ M_{\overline{Q},\max}(s) &:= \sup\big\{\P_C(\varphi(\mathbf{X})<s)\colon C\preceq \overline{Q}\big\} \leq \overline{Q}(F_1(s),\dots,F_d(s)).\end{aligned}$$ Analogously, if $\underline{\widehat{Q}}$ and $\widehat{\overline{Q}}$ are quasi-survival functions then, for $\varphi(x_1,\dots,x_d)=\min\{x_1,\dots,x_d\}$, we have that $$\begin{aligned} m_{\widehat{\overline{Q}},\min}(s) &:= \inf\big\{\P_C(\varphi(\mathbf{X})<s)\colon \widehat{C}\preceq \widehat{\overline{Q}}\big\} \geq 1-\widehat{\overline{Q}}(F_1(s),\dots,F_d(s))\\ M_{\widehat{\underline{Q}},\min}(s) &:= \sup\big\{\P_C(\varphi(\mathbf{X})<s)\colon \widehat{\underline{Q}}\preceq \widehat{C}\big\} \leq 1-\widehat{\underline{Q}}(F_1(s),\dots,F_d(s)) = \overline{M}_{\widehat{\underline{Q}},\min}(s).\end{aligned}$$ Let $\varphi(x_1,\dots,x_d) = \max\{x_1,\dots,x_d\}$, then for any copula $C$ we have that $$\begin{aligned} \P_C(\max\{X_1,\dots,X_d\}<s) = \P_C(X_1<s,\dots,X_d<s) = C(F_1(s),\dots,F_d(s)),\end{aligned}$$ using Sklar’s Theorem for the last equality. Hence, it follows immediately that $$\begin{aligned} m_{\underline{Q},\max}(s) &= \inf\big\{C(F_1(s),\dots,F_d(s))\colon \underline{Q}\preceq C\big\} \geq \underline{Q}(F_1(s),\dots,F_d(s)) \\ \text{and} \qquad\qquad \qquad M_{\overline{Q},\max}(s) &= \sup\big\{C(F_1(s),\dots,F_d(s))\colon C\preceq \overline{Q}\big\} \leq \overline{Q}(F_1(s),\dots,F_d(s)).\end{aligned}$$ Moreover, since $$\begin{aligned} \mathcal V^<_{\max} (s) &= \{ (x_1,\dots,x_d)\in\R^d: \max\{x_1,\dots,x_d\}<s\}\\ &= \{ (x_1,\dots,x_d)\in\R^d: x_1<s,\dots,x_d<s\}, $$ we get from the improved standard bounds that $$\begin{aligned} \underline{m}_{\underline{Q},\max}(s) &= \sup_{\mathcal V^<_{\max} (s)}\ \underline{Q}\big(F_1(x_1),\dots,F_{d}(x_{d})\big) = \underline{Q}(F_1(s),\dots,F_d(s)),\end{aligned}$$ where the last equality follows from the fact that $\underline{Q}$ is a quasi-copula, hence it is increasing in each component such that the supremum is attained at $(F_1(s),\dots,F_d(s))$. Similarly, we have for $\varphi(x_1,\dots,x_d) = \min\{x_1,\dots,x_d\}$ and any copula $C$ that $$\begin{aligned} \P_C(\min\{X_1,\dots,X_d\}<s) = 1-\P_C(\min\{X_1,\dots,X_d\}\geq s) = 1-\widehat{C}(F_1(s),\dots,F_d(s)),\end{aligned}$$ where the last equality follows from the definition of $\widehat{C}$. Hence it follows $$\begin{aligned} m_{\widehat{\overline{Q}},\min}(s) &= \inf\big\{1-\widehat{C}(F_1(s),\dots,F_d(s))\colon \widehat{C}\preceq \widehat{\overline{Q}}\big\} \geq 1-\widehat{\overline{Q}}(F_1(s),\dots,F_d(s)) \end{aligned}$$ and analogously $M_{\widehat{\underline{Q}},\min}(s)\leq 1-\widehat{\underline{Q}}(F_1(s),\dots,F_d(s))$. Once again, it follows from the improved standard bounds again that $$\overline M_{\widehat{\underline{Q}},\min}(s) = \inf_{\mathcal V^>_{\min}(s)}\ \big\{ 1-\widehat{\underline{Q}}\big(F_1(x_1),\dots,F_{d}(x_{d})\big) \big\} = 1-\widehat{\underline{Q}}\big(F_1(s),\dots,F_d(s)\big),$$ which conlcudes the proof. In the absence of (partial) dependence information on $\mathbf X$, the best-possible bounds on its copula or its survival function are given by the bounds. In this case, the improved standard bounds in reduce to the standard bounds in . These will serve as a proxy in order to measure the quality of the improved VaR bounds presented in the remainder of this paper. Inspired by the properties of the standard bounds, two interesting lines of research have emerged: First, the fact that the standard bounds for the sum of risks are not sharp in general for $d>3$, calls for a different approach to obtain sharp bounds on VaR in the absence of dependence information. While generally the quest for sharp analytical bounds in this situation remains an open problem, some prominent approaches have been developed to provide solutions under additional requirements on the marginals $F_1,\dots,F_d$. For instance, sharp explicit bounds were obtained by @rueschendorf1982 using a dual approach in the case of uniform marginals. This approach was later refined to obtain bounds for more general marginals in the homogeneous case, *i.e.* when $F_1=\cdots=F_d$; see @embrechts2006, @puccetti2012b and the references therein for details on the so-called dual approach. Alternatively, @wang2013 used the concept of joint and complete mixability to derive sharp bounds in the homogeneous case under some monotonicity requirements on the marginals. Moreover, the rearrangement algorithm of @embrechts2013 provides an efficient way to approximate sharp VaR bounds in the absence of dependence information for general marginals. Although the favorable numerical properties of this algorithm have been demonstrated in many practical examples, a proof of convergence and other theoretical properties remain an open challenge. Besides the derivation of sharp VaR bounds in the absence of dependence information, the second question of interest is how to account for additional dependence information in the computation of the bounds. Such additional information could be the knowledge of the correlation matrix of $\mathbf{X}$, or the knowledge of its copula on a subset of $\mathbb{I}^d$. In such cases, one can make use of the additional information in order to improve the standard VaR bounds. For instance, @embrechts2003, @rueschendorf2005 and @puccetti2012b computed improved bounds that account for the information that $\mathbf{X}$ is positively upper or lower orthant dependent, meaning that its copula or survival copula are bounded from below by the independence copula. Moreover, in a series of papers, analytical improvements of the standard bounds were developed, when some higher-dimensional marginals of $\mathbf{X}$ are known; see @rueschendorf1991, @embrechts2010 and @puccetti2012b. Beyond analytical improvements, numerical bounds that include additional dependence information were established based on the rearrangement algorithm. @bernard2015 refined the rearrangement algorithm in order to account for given values of the distribution of $\mathbf{X}$ on a subset of its domain. A similar algorithm was presented by @bernard2015b in order to include information about the variance of the sum $X_1+\cdots+X_d$, while numerical and analytical methods to compute risk bounds in factor models were presented by @bernard2016. Continuing this line of research, in the following we will improve the standard bounds and account for additional types of dependence information. In the next section we derive improved VaR bounds for the sum of risks when, besides the marginals, the distributions of the minima or maxima of some subsets of the risks are known. Such information can typically be inferred, with an appropriate degree of accuracy, using tools from extreme value theory and are thus available in many practical applications. In section \[boundsOnCopula\], we take a different approach to derive improved VaR bounds for general functionals $\varphi$ of $\mathbf{X}$ when the copula of the risk vector lies in the vicinity of a reference copula, as measured by some distance on the set of copulas. We therefore derive improved bounds on the copula of $\mathbf{X}$ that account for the given information. The improved bounds are then translated into bounds on VaR via the improved standard bounds. Improved bounds on the Value-at-Risk of the sum with known distributions of some minima and maxima {#prescribedMax} ================================================================================================== In this section we improve the standard bounds on the VaR of the sum $X_1+\cdots+X_d$ in the situation where, besides the marginal distributions, the laws of the minima and maxima of some subsets of the risks $X_1,\dots,X_d$ are known. In particular, we assume that for a system $J_1,\dots,J_m\subset\{1,\dots,d\}$ the distributions of $\max_{j\in J_n} X_j$ or $\min_{j\in J_n} X_j$ for $n=1,\dots,m$ are given. This setting can be viewed as an interpolation between the marginals-only case and the situation where the lower-dimensional marginals of the vectors $(X_j)_{j\in J_n}$ are completely specified; see also @rueschendorf2017 for another work in the same spirit. The latter setting has been studied extensively in the literature, and risk bounds for aggregations of $\mathbf{X}$ given some of its lower-dimensional marginals were obtained, for instance, by @rueschendorf1991, @embrechts2010 and @puccetti2012b. These bounds are based on a reduction principle that transforms the optimization problem involving higher-dimensional marginals into a standard Fréchet problem (*i.e.* marginals-only), utilizing the extra information about the distribution of the subvector $(X_j)_{j\in J_n}$. In practice however, it is often difficult to determine the distributions of the lower-dimensional vectors $(X_j)_{j\in J_n}$. In particular for large dimensions of the subsets, a vast amount of data is required to estimate the distribution of $(X_j)_{j\in J_n}$ with an adequate degree of accuracy. Thus, having complete information about lower-dimensional marginals of $(X_1,\dots,X_d)$ turns out to be a rather strong assumption. Therefore, methods that interpolate between this scenario and the marginals-only case are of practical interest. Based on the reduction principle of [@puccetti2012b], we develop in this section a method to improve the standard bounds when instead of the distribution of $(X_j)_{j\in J_n}$, only the distribution of its maximum $\max_{j\in J_n} X_j$ or minimum $\min_{j\in J_n} X_j$ is known. Let us point out that obtaining information about the maximum or minimum of a sequence of random variables is the central theme of extreme value theory, which provides a rich collection of methods for their estimation; see *e.g.* @Beirlant_etal_2004 [Chapter 9]. Let us denote by $\mathcal{I}:=\{1,\dots,d\}$ and by $\mathcal{J}:=\{1,\dots,m\}$. \[boundMax\] Let $(X_1,\dots,X_d)$ be a random vector with marginals $F_1,\dots,F_d$, and consider a collection $\mathcal{E} = \{J_1,\dots,J_m\}$ of subsets $J_n\subset\mathcal{I}$ for $n\in\mathcal{J}$ with $\bigcup_{n\in\mathcal{J}} J_{n} = \mathcal{I}$. Denote by $G_n$ the distribution of $Y_n=\max_{j\in J_n} X_j$. Then it follows that $$\begin{gathered} \inf\big\{\P(X_1+\cdots+X_d\leq s)\colon X_i\sim F_i, i\in\mathcal{I}, \max_{j\in J_n} X_j \sim G_n, n\in\mathcal{J}\big\}\\ \geq \sup_{(\alpha_1,\dots,\alpha_m)\in\underline{\mathcal{A}}} \inf\big\{\P(\alpha_1Y_1+\cdots+\alpha_mY_m\leq s)\colon Y_n\sim G_n, n \in\mathcal{J}\big\} =: \underline{m}_{\mathcal{E},\max}(s),\end{gathered}$$ where $$\underline{\mathcal{A}} = \Big\{(\alpha_1,\dots,\alpha_m)\in\mathbb{R}_+^m\colon \sum_{n=1}^m \alpha_n\max_{j\in J_n} x_j \geq \sum_{i=1}^d x_i,\text{ for all } (x_1,\dots,x_d)\in\mathbb{R}^d\Big\}\neq\emptyset.$$ Moreover if $(X_1,\dots,X_d)$ is $\mathbb{R}_+^d$-valued, then $$\begin{gathered} \sup\big\{\P(X_1+\cdots+X_d\leq s)\colon X_i\sim F_i, i\in\mathcal{I}, \max_{j\in J_n} X_j \sim G_n, n\in\mathcal{J}\big\}\\ \leq \inf_{(\alpha_1,\dots,\alpha_m)\in\overline{\mathcal{A}}} \sup\big\{\P(\alpha_1Y_1+\cdots+\alpha_mY_m\leq s)\colon Y_n\sim G_n, n\in\mathcal{J}\big\} =: \overline{M}_{\mathcal{E},\max}(s),\end{gathered}$$ where $$\overline{\mathcal{A}} = \Big\{(\alpha_1,\dots,\alpha_m)\in\mathbb{R}_+^m\colon \sum_{n=1}^m \alpha_n\max_{j\in J_n} x_j \leq \sum_{i=1}^d x_i,\text{ for all } (x_1,\dots,x_d)\in\mathbb{R}_+^d\Big\}\neq\emptyset.$$ We first show that the lower bound $\underline{m}_{\mathcal{E},\max}$ is valid. It follows from $\bigcup_{n=1}^m J_n = \{1,\dots,d\}$ that $\underline{\mathcal{A}}\neq\emptyset$. Indeed, choosing for instance $\alpha_n=|J_n|$ we get that $\sum_{j\in J_n} x_j \leq \alpha_n \max_{j\in J_n} x_j$, for all $(x_1,\dots,x_d)\in\mathbb{R}^d$ and $n=1,\dots,m$. Hence $$\sum_{n=1}^m \alpha_n\max_{j\in J_n} x_j \geq \sum_{n=1}^m\sum_{j\in J_n} x_j \geq \sum_{i=1}^d x_i\quad\text{for all }(x_1,\dots,x_d)\in\mathbb{R}^d.$$ Then, it follows for arbitrary $(\alpha_1,\dots,\alpha_m)\in\underline{\mathcal{A}}$ that $$\bigg\{\sum_{n=1}^m \alpha_n\max_{j\in J_n} X_j\leq s\bigg\} \subseteq \bigg\{\sum_{i=1}^d X_i\leq s\bigg\},$$ henceforth $$\begin{aligned} \inf & \big\{\P(X_1+\cdots+X_d\leq s)\colon X_i\sim F_i, i\in\mathcal{I}, \max_{j\in J_n} X_j \sim G_n, n\in\mathcal{J}\big\} \\ &\geq \inf\bigg\{\P\bigg(\sum_{n=1}^m \alpha_n\max_{j\in J_n} X_j\leq s\bigg)\colon X_i\sim F_i, i\in\mathcal{I}, \max_{j\in J_n} X_j \sim G_n, n\in\mathcal{J}\bigg\}\\ &= \inf\big\{\P(\alpha_1Y_1+\cdots+\alpha_mY_m\leq s)\colon Y_n\sim G_n, n\in\mathcal{J},\big\}.\end{aligned}$$ Now, since $(\alpha_1,\dots,\alpha_m)\in\underline{\mathcal{A}}$ was arbitrary, it follows that the lower bound holds by taking the supremum over all elements in $\underline{\mathcal{A}}$. Likewise for the upper bound, we note that since $(X_1,\dots,X_d)$ is $\mathbb{R}^d_+$-valued, $(0,\dots,0)$ and $(1,\dots,1)$ belong to $\overline{\mathcal{A}}$, hence it is not empty. Moreover, for arbitrary $(\alpha_1,\dots,\alpha_m)\in\overline{\mathcal{A}}$, it follows that $$\bigg\{\sum_{n=1}^m \alpha_n\max_{j\in J_n} X_j\leq s\bigg\} \supseteq \bigg\{\sum_{i=1}^d X_i\leq s\bigg\},$$ due to the fact that $(X_1,\dots,X_d)$ is non-negative and $\sum_{j\in J_n} x_j \geq \sum_{n=1}^m \alpha_n\max_{j\in J_n} x_j$. Hence, we get that $$\begin{aligned} \sup\big\{ & \P(X_1+\cdots+X_d\leq s)\colon X_i\sim F_i, i\in\mathcal{I}, \max_{j\in J_n} X_j \sim G_n, n\in\mathcal{J}\big\} \\ &\leq\sup\bigg\{\P\bigg(\sum_{n=1}^m \alpha_n\max_{j\in J_n} X_j\leq s\bigg)\colon X_i\sim F_i, i\in\mathcal{I}, \max_{j\in J_n} X_j \sim G_n, n\in\mathcal{J}\bigg\}\\ &=\sup\big\{\P(\alpha_1 Y_1+\cdots+\alpha_m Y_m\leq s)\colon Y_n\sim G_n, n\in\mathcal{J}\big\}. \end{aligned}$$ Since $(\alpha_1,\dots,\alpha_m)\in\overline{\mathcal{A}}$ was arbitrary, it follows that the upper bound holds indeed. \[rem:calE\] The assumption $\bigcup_{n=1}^m J_{n} = \{1,\dots,d\}$ can always be met by adding singletons to $\mathcal{E}$, *i.e.* $J_n = \{i_n\}$ for $i_n\in\{1,\dots,d\}$, since the marginal distributions of $(X_1,\dots,X_d)$ are known. However, the bounds are valid even when the marginal distributions are not known. By the same token, the following result establishes bounds on the distribution the sum of the components of $\mathbf{X}$ when distributions of some minima are known. The proof follows along the same lines of argumentation as the proof of Theorem \[boundMax\], and is therefore omitted. \[boundMin\] Consider the setting of Theorem \[boundMax\] and denote by $H_n$ the distribution of $Z_n=\min_{j\in J_n} X_j$. Then it follows that $$\begin{gathered} \sup\big\{\P(X_1+\cdots+X_d\leq s)\colon X_i\sim F_i, i\in\mathcal{I}, \min_{j\in J_n} X_j \sim H_n, n\in\mathcal{J}\big\}\\ \leq \inf_{(\alpha_1,\dots,\alpha_m)\in\overline{\mathcal{B}}} \sup\big\{\P(\alpha_1Z_1+\cdots+\alpha_mZ_m\leq s)\colon Z_n\sim H_n, n\in\mathcal{J}\big\} =: \overline{M}_{\mathcal{E},\min}(s),\end{gathered}$$ where $$\overline{\mathcal{B}} = \Big\{(\alpha_1,\dots,\alpha_m)\in\mathbb{R}_+^m\colon \sum_{n=1}^m \alpha_n\min_{j\in J_n} x_j \leq \sum_{i=1}^d x_i,\text{ for all } (x_1,\dots,x_d)\in\mathbb{R}^d\Big\}\neq\emptyset.$$ Moreover if $(X_1,\dots,X_d)$ is $\mathbb{R}_-^d$-valued, then $$\begin{gathered} \inf\big\{\P(X_1+\cdots+X_d\leq s)\colon X_i\sim F_i, i\in\mathcal{I}, \min_{j\in J_n} X_j \sim H_n, n\in\mathcal{J}\big\}\\ \geq \sup_{(\alpha_1,\dots,\alpha_m)\in\underline{\mathcal{B}}}\inf\big\{\P(\alpha_1 Z_1+\cdots+\alpha_m Z_m\leq s)\colon Z_n\sim H_n, n\in\mathcal{J}\big\} =: \underline{m}_{\mathcal{E,\min}}(s),\end{gathered}$$ where $$\underline{\mathcal{B}} = \Big\{(\alpha_1,\dots,\alpha_m)\in\mathbb{R}_+^m\colon \sum_{n=1}^m \alpha_n\min_{j\in J_n} x_j \geq \sum_{i=1}^d x_i,\text{ for all } (x_1,\dots,x_d)\in\mathbb{R}_-^d\Big\}\neq\emptyset.$$ The computation of the bounds presented in Theorems \[boundMax\] and \[boundMin\] can be cumbersome for two reasons. Firstly, for fixed $(\alpha_1,\dots,\alpha_m)$ there does not exist a method to compute sharp analytical bounds on the set $\big\{\P(\alpha_1Y_1+\cdots+\alpha_mY_m\leq s)\colon Y_n\sim G_n, n = 1,\dots,m\}$, except when $m=2$. This problem can however be circumvented either by using the standard bounds in , or numerically, by an application of the rearrangement algorithm of @embrechts2013. Using the rearrangement algorithm, we are able to approximate upper and lower bounds on the set in an efficient way. Secondly, the determination of the sets $\underline{\mathcal{A}}$, $\overline{\mathcal{A}}$ and $\underline{\mathcal{B}}$, $\overline{\mathcal{B}}$ depends on the system $J_1,\dots,J_m$ and is, in general, not straightforward. However, in Section \[numerics\] we will demonstrate that, even for possibly non-optimal elements in $\underline{\mathcal{A}}$, $\overline{\mathcal{A}}$ and $\underline{\mathcal{B}}$, $\overline{\mathcal{B}}$, the bounds in Theorems \[boundMax\] and \[boundMin\] yield a significant improvement over the standard bounds. Improved bounds using a reference copula {#boundsOnCopula} ======================================== In this section we present an alternative approach for improving the standard bounds on the Value-at-Risk, for general aggregation functions $\varphi$ and different types of additional dependence information on the risk vector $\mathbf X$. Our approach is based on improvements of the bounds on copulas that account for additional dependence information. The improved bounds are then used to derive sharper bounds on VaR in conjunction with the improved standard bounds . We focus on two types of additional dependence information. Firstly, we consider the situation where the copula $C$ of the risk vector $\mathbf X$ coincides with a reference model on a compact subset $\set$ of its domain, *i.e.* it holds that $C(\bx) = C^*(\bx)$ for all $\bx\in\set$ and a reference copula $C^*$. In practice, the set $\set$ may correspond to a region in $\mathbb{I}^d$ that contains enough observations to estimate the copula $C$ with sufficient accuracy, so that we can assume that $C$ is known on $\set$. @bernard2015 call such a subset *trusted region* and present several techniques and criteria to select such regions when estimating copulas. If $\set$ is not equal to the entire domain of the copula, then dependence uncertainty stems from the fact that $C$ remains unknown on $\mathbb{I}^d\setminus\set$. In order to obtain VaR bounds in this situation, we use results from @lux2016 who established improved bounds on the set of copulas with prescribed values on a compact set. Secondly, we present a new improvement of the bounds when the copula $C$ is assumed to lie in the vicinity of a reference model as measured by a statistical distance. More formally, we establish bounds on the set of all (quasi-)copulas $C$ in the $\delta$-neighborhood of the reference copula $C^*$, *i.e.* such that $\cD(C,C^*)\leq\delta$ for a distance $\cD$. Our method applies to a large class of statistical distances such as the Cramér–von Mises or the $L^p$ distances. Such situations arise naturally in practice when one tries to estimate a copula from, or calibrate it to, empirical data. The estimation typically involves the minimization of a distance to the empirical copula over a parametric family of copulas, *i.e.* $\cD(C_\theta,C^*)\to \min_\theta$ where $C^*$ is an empirical copula and $(C_\theta)_\theta$ is a family of parametric copulas. This is in the literature often referred to as *minimal distance* or *minimal contrast* estimation. @kole2007 for instance present several distance-based techniques for selecting copulas in risk management. These estimation procedures lend themselves immediately to the methodology we propose, as typically one arrives at $\delta:=\min_\theta\cD(C_\theta,C^*)>0$, due to the fact that the family of models $(C_\theta)_\theta$ is not able to match the empirical observations exactly, thus dependence uncertainty remains. In this case, $\delta$ can be viewed as the inevitable model risk due to the choice of the parametric family $(C_\theta)_\theta$. Our method can then be used to account for such types of dependence uncertainty in the computation of VaR. Approaches to compute robust risk estimates over a class of models that lie in the proximity of a reference model have been proposed earlier in the literature. @glasserman2013 derive robust bounds on the portfolio variance, the conditional VaR and the CVA over the class of models within a relative entropy distance of a reference model. @barrieu2015 establish bounds on the VaR of a univariate random variable given that its distribution is close to a reference distribution in the sense of the Kolmogorov–Smirnov or Lévy distance. In a multivariate setting, @blanchet2016 use an optimal transport approach to derive robust bounds on risk estimates, such as ruin probabilities, over models that are in a neighborhood of a reference model in terms of the Wasserstein distance. This brief overview is, of course, incomplete and we refer the reader to the references in each of the aforementioned articles for a more detailed review of the associated literature. Let us now consider the setting where, apart from the marginal distributions, partial information on the dependence structure of the random vector $\mathbf X$ is available. In particular, assume that the copula is known on some subset $\mathcal S$ of $[0,1]^d$. Theorem 3.1 in [@lux2016] establishes sharp bounds on the set $$\begin{aligned} \mathcal{Q}^{\set,Q^*} := \big\{Q\in\mathcal{Q}^d\colon Q(\mathbf{x}) = Q^*(\mathbf{x}) \text{ for all } \mathbf{x}\in \set\big\},\end{aligned}$$ where $\set\subset\mathbb{I}^d$ is compact and $Q^*$ is a $d$-quasi-copula. The bounds are provided by $$\begin{aligned} \label{bounds} \begin{split} \underline{Q}^{\set,Q^*}(\mathbf{u}) :=& \min\big\{Q(\bu)\colon Q(\bx) = Q^*(\bx) \text{ for all } \bx\in\set\big\}\\ =& \max\Big(0, \sum_{i=1}^d u_i-d+1,\max_{\mathbf{x}\in \set} \Big\{Q^*(\mathbf{x})-\sum_{i=1}^d (x_i-u_i)^+\Big\}\Big),\\ \overline{Q}^{S,Q^*}(\mathbf{u}) :=& \max\big\{Q(\bu)\colon Q(\bx) = Q^*(\bu) \text{ for all } \bx\in\set\big\}\\ =& \min\Big(u_1,\dots,u_d,\min_{\mathbf{x}\in S} \Big\{Q^*(\mathbf{x})+\sum_{i=1}^d (u_i-x_i)^+\Big\}\Big), \end{split}\end{aligned}$$ for all $\bu\in\mathbb{I}^d$, and they are quasi-copulas, hence they belong to $\mathcal{Q}^{S,Q^*}$. Let us point out that a similar version of these bounds was presented recently by @puccetti2016. They were derived independently in the master thesis of the third-named author. By slightly abusing notation, we will sometimes write $\underline{Q}^{\{\bu\},\alpha}$ and $\overline{Q}^{\{\bu\},\alpha}$ with $\alpha\in [W_d(\bu),M_d(\bu)]$ instead of a quasi-copula function $Q^*$, and mean that $Q^*(\bu)=\alpha$. The bounds in hold also for sets of copulas, *i.e.* for each copula $C$ in $$\mathcal{C}^{\set,Q^*} := \big\{C\in\mathcal{C}^d\colon C(\mathbf{x}) = Q^*(\mathbf{x}) \text{ for all } \mathbf{x}\in \set\big\}$$ it holds that $\underline{Q}^{\set,Q^*}{\preceq}C{\preceq}\overline{Q}^{\set,Q^*}$, assuming that $\mathcal{C}^{\set,Q^*}$ is not empty. Moreover, Proposition A.1 in [@lux2016] provides analogous bounds on survival functions, *i.e.* for a reference copula $C^*$ and any copula $C$ in $$\widehat{\mathcal{C}}^{\set,C^*} := \big\{C\in\mathcal{C}^d\colon \widehat{C}(\mathbf{x}) = \widehat{C}^*(\mathbf{x}) \text{ for all } \mathbf{x}\in\set\big\}$$ it holds that $\widehat{\underline{Q}}^{\set,C^*} {\preceq}\widehat{C}{\preceq}\widehat{\overline{Q}}^{\set,C^*}$, where $$\begin{aligned} \label{survivalBounds} \widehat{\underline{Q}}^{\set,C^*}(\mathbf{u}) := \underline{Q}^{\widehat{\set},\widehat{C}^*}(\mathbf 1-\bu) \quad\text{and}\quad \widehat{\overline{Q}}^{\set,C^*}(\mathbf{u}) := \overline{Q}^{\widehat{\set},\widehat{C}^*}(\mathbf 1-\bu),\end{aligned}$$ while $\widehat{\set} = \{(1-x_1,\dots,1-x_d)\colon (x_1,\dots,x_d)\in \set\}$. In case $d=2$, the above bounds correspond to the improved bounds derived by @tankov2011. He showed that the bounds are themselves copulas under certain constraints on the set $\set$, and those were readily relaxed by @bernard2012. For instance, if $Q^*$ is a 2-copula and $\set$ a rectangle, *i.e.* $\set = \{(x_1,x_2)\colon a_1\leq x_1\leq b_1, a_2\leq x_2\leq b_2\}$ then $\underline{Q}^{\set,Q^*}$ and $\overline{Q}^{\set,Q^*}$ are 2-copulas. In contrast, @lux2016 showed that for $d>2$ the bounds $\underline{Q}^{\set,Q^*}$ and $\overline{Q}^{\set,Q^*}$ are copulas only in degenerate cases, and quasi-copulas otherwise. In the following we will establish improved bounds using a different type of additional dependence information. Namely, we consider the set of copulas that are close to a reference copula in the sense of a statistical distance as defined below. Let us first define the minimal and maximal convolution between two quasi-copulas $Q,Q'$ as the pointwise minimum and maximum between them, *i.e.* $(Q\wedge Q')(\bu) = Q(\bu) \wedge Q'(\bu)$ and $(Q\vee Q')(\bu) = Q(\bu) \vee Q'(\bu)$. A function $\cD\colon\mathcal{Q}^d\times\mathcal{Q}^d\to\mathbb{R}_+$ is called a *statistical distance* if for $Q,Q'\in\mathcal{Q}^d$ $$\cD(Q,Q') = 0\quad\Longleftrightarrow\quad Q(\bu)=Q'(\bu)\quad \text{for all }\bu\in\mathbb{I}^d.$$ A statistical distance $\cD$ is *monotonic* with respect to the order $\preceq$ on $\mathcal{Q}^d$, if for $Q,Q',Q''\in\mathcal{Q}^d$ it holds $$\begin{aligned} Q\preceq Q'\preceq Q'' \quad\Longrightarrow\quad \cD(Q',Q'')\leq\cD(Q,Q'') \ \text{ and } \ \cD(Q'',Q')\leq\cD(Q'',Q).\end{aligned}$$ A statistical distance $\cD$ is *min-* resp. *max-stable* if for $Q,Q'\in\mathcal{Q}^d$ it holds $$\begin{aligned} \cD(Q,Q') & \geq \max\{\cD({Q\wedge Q'},Q), \cD(Q,{Q\wedge Q'})\} \\ \cD(Q,Q') & \geq \max\{\cD({Q\vee Q'},Q), \cD(Q,{Q\vee Q'})\}.\end{aligned}$$ The following Theorem establishes pointwise bounds on the set of quasi-copulas that are in the $\delta$-vicinity of a reference copula $C^*$ as measured by a statistical distance $\cD$. \[prescribedDistance\] Let $C^*$ be a $d$-copula and $\cD$ be a statistical distance which is continuous with respect to the pointwise convergence of quasi-copulas, monotonic with respect to the lower orthant order and min/max-stable. Consider the set $$\mathcal{Q}^{\cD,\delta} := \big\{ Q\in\mathcal{Q}^d\colon \cD(Q,C^*) \leq \delta \big\}$$ for $\delta\in\mathbb{R}_+$. Then $$\begin{aligned} \underline{Q}^{\cD,\delta}(\bu) &:=\min\Big\{\alpha \in \mathbb S(\bu) \colon \cD\Big({{\overline{Q}}^{\{\bu\},\alpha}\wedge{C^*}},{C^*}\Big) \leq \delta\Big\} = \min\big\{Q(\bu)\colon Q\in\mathcal{Q}^{\cD,\delta}\big\},\\ \overline{Q}^{\cD,\delta}(\bu) &:=\max\Big\{\alpha \in \mathbb S(\bu) \colon \cD\Big({{\underline{Q}}^{\{\bu\},\alpha}\vee{C^*}},{C^*}\Big) \leq \delta\Big\} = \max\big\{Q(\bu)\colon Q\in\mathcal{Q}^{\cD,\delta}\big\},\end{aligned}$$ where $\mathbb S(\bu) := [W_d(\bu),M_d(\bu)]$, and both bounds are quasi-copulas. We show that the statement holds for the lower bound, while the proof for the upper bound follows along the same lines. Fix an $\alpha\in[W_d(\bu),M_d(\bu)]$ and a $\bu\in\mathbb I^d$, then the map $v\mapsto\big({{\overline{Q}}^{\{\bu\},\alpha}\wedge{C^*}}\big)(v)$ is a quasi-copula; this follows by straightforward calculations using the definition of the minimal convolution, see also @Rodriguez_Ubeda_2004 [Theorem 2.1]. By definition, $\cD$ is monotonic with respect to the lower orthant order, thus it follows for $\underline{\alpha},\overline{\alpha} \in [W_d(\bu),M_d(\bu)]$ with $\underline{\alpha}<\overline{\alpha}$ that $$\cD\Big({{\overline{Q}}^{\{\bu\},\overline{\alpha}}\wedge{C^*}},{C^*}\Big) \leq \cD\Big({{\overline{Q}}^{\{\bu\},\underline{\alpha}}\wedge{C^*}},{C^*}\Big),$$ due to the fact that ${\overline{Q}}^{\{\bu\},\underline{\alpha}} {\preceq}{\overline{Q}}^{\{\bu\},\overline{\alpha}}$, which readily implies $$\Big({{\overline{Q}}^{\{\bu\},\underline{\alpha}}\wedge{C^*}}\Big){\preceq}\Big({{\overline{Q}}^{\{\bu\},\overline{\alpha}}\wedge{C^*}}\Big) {\preceq}{C^*}.$$ Hence, the map $$[W_d(\bu),M_d(\bu)]\ni\alpha\mapsto\cD\Big({{\overline{Q}}^{\{\bu\},\alpha}\wedge{C^*}},{C^*}\Big)$$ is decreasing. Moreover, as a consequence of the Arzelà–Ascoli Theorem, it follows that for every sequence $(\alpha_n)_n\subset[W_d(\bu),M_d(\bu)]$ with $\alpha_n\to\alpha$, $$\Big({{\overline{Q}}^{\{\bu\},\alpha_n}\wedge{C^*}}\Big) \xrightarrow[n\to\infty]{} \Big({{\overline{Q}}^{\{\bu\},\alpha}\wedge{C^*}}\Big)$$ uniformly and, since $\cD$ is continuous with respect to the pointwise convergence of quasi-copulas, it follows that $\alpha\mapsto\cD\Big({{\overline{Q}}^{\{\bu\},\alpha}\wedge{C^*}},{C^*}\Big)$ is continuous. In addition, we have that $$\begin{aligned} \label{prescribedDistanceEq1} \cD\Big({{\overline{Q}}^{\{\bu\},M_d}\wedge{C^*}},{C^*}\Big) =\cD\Big({M_d\wedge{C^*}},{C^*}\Big) = \cD\Big({C^*},{C^*}\Big) =0,\end{aligned}$$ due to the fact that $C^*{\preceq}M_d$. We now distinguish between two cases: ($i$) Let $\delta\leq\cD\Big({{\overline{Q}}^{\{\bu\},W_d}\wedge{C^*}},{C^*}\Big)$. Then, due to the monotonicity and continuity of the map $[W_d(\bu),M_d(\bu)]\ni\alpha\mapsto\cD\Big({{\overline{Q}}^{\{\bu\},\alpha}\wedge{C^*}},{C^*}\Big)$ and it holds that the set $$\mathcal{O} :=\Big\{\alpha\colon \cD\Big({\overline{Q}}^{\{\bu\},\alpha}\wedge{C^*},{C^*}\Big) = \delta\Big\}$$ is non-empty and compact. Define $\alpha^* := \min\{\alpha\colon\alpha\in\mathcal{O}\}$. We will show that $\min\big\{Q(\bu)\colon Q\in\mathcal{Q}^{\cD,\delta}\big\}=\alpha^*$. On the one hand, it holds that $\min\big\{Q(\bu)\colon Q\in\mathcal{Q}^{\cD,\delta}\big\}\leq\alpha^*$. Indeed, consider ${{\overline{Q}}^{\{\bu\},\alpha^*}\wedge{C^*}}$ which is a quasi-copula and belongs to $\mathcal{Q}^{\cD,\delta}$ since $\alpha^*\in\mathcal{O}$. Then, we have that $$\Big({{\overline{Q}}^{\{\bu\},\alpha^*}\wedge{C^*}}\Big)(\bu) = \min\{\alpha^*,C^*(u)\} = \alpha^*,$$ using again that $\alpha^*\in\mathcal{O}$ and . Hence the inequality holds. On the other hand, we will show now that the inequality cannot be strict by contradiction. Assume there exists a quasi-copula $Q'\in\mathcal{Q}^{\cD,\delta}$ with $Q'(\bu)<\alpha^*$. Then it follows that $$\begin{aligned} \begin{split} \label{prescribedDistanceEq2} \cD(Q',C^*) & \geq \cD\big({Q'\wedge{C^*}},C^*\big) \geq \cD\Big({{\overline{Q}}^{\{\bu\},Q'}\wedge{C^*}},C^*\Big)\\ & \geq \cD\Big({{\overline{Q}}^{\{\bu\},\alpha^*}\wedge{C^*}},C^*\Big) = \delta, \end{split}\end{aligned}$$ where the first inequality follows from the min-stability of $\cD$, and the second and third ones from its monotonicity properties. However, since $Q'(\bu)\notin\mathcal{O}$ it follows that $ \cD\Big({{\overline{Q}}^{\{\bu\},Q'}\wedge{C^*}},C^*\Big)\neq\delta$, hence yields that $\cD\Big({{\overline{Q}}^{\{\bu\},Q'}\wedge{C^*}},C^*\Big)>\delta$. This contradicts the assumption that $Q'\in\mathcal{Q}^{\cD,\delta}$, showing that indeed $\min\big\{Q(\bu)\colon Q\in\mathcal{Q}^{\cD,\delta}\big\}=\alpha^*$. Hence, the lower bound holds for $\delta\leq\cD\Big({{\overline{Q}}^{\{\bu\},W_d}\wedge{C^*}},{C^*}\Big)$. ($ii$) Now, let $\delta>\cD\Big({{\overline{Q}}^{\{\bu\},W_d}\wedge{C^*}},{C^*}\Big)$, then it follows that $$\min\Big\{\alpha\in[W_d(\bu),M_d(\bu)]\colon\cD\Big({{\overline{Q}}^{\{\bu\},\alpha}\wedge{C^*}},{C^*}\Big) \leq \delta\Big\} = W_d(\bu).$$ Moreover, since $\Big({\overline{Q}}^{\{\bu\},W_d}\wedge{C^*}\Big)\in\mathcal{Q}^{\cD,\delta}$ and every element in $\mathcal{Q}^{\cD,\delta}$ is bounded from below by $W_d$, it follows that $\min\big\{Q(\bu)\colon Q\in\mathcal{Q}^{\cD,\delta}\big\} = W_d(\bu)$. Hence, the lower bound holds in this case as well. Finally, it follows again from [@Rodriguez_Ubeda_2004 Theorem 2.1] that the bounds are quasi-copulas, which completes the proof. Let $C^*$ and $\cD$ be as in Theorem \[prescribedDistance\], and consider $\delta\in\mathbb{R}_+$. Then, the bounds $\underline{Q}^{\cD,\delta}$ and $\overline{Q}^{\cD,\delta}$ also apply to the set of copulas $\mathcal{C}^{\cD,\delta} := \{C\in\mathcal{C}^d\colon \cD(C,C^*) \leq \delta\}$, assuming that $\mathcal{C}^{\cD,\delta}\neq\emptyset$, that is $$\begin{aligned} \label{boundsDeltaCopulas} \underline{Q}^{\cD,\delta} {\preceq}C {\preceq}\overline{Q}^{\cD,\delta},\end{aligned}$$ for all $C\in\mathcal{C}^{\cD,\delta}$, due to the fact that $\mathcal{C}^{\cD,\delta}\subseteq\mathcal{Q}^{\cD,\delta}$. If $\cD$ is not symmetric, the set $\{Q\in\mathcal{Q}^d\colon \cD(Q,C^*) \leq \delta\}$ might not coincide with the set $\{Q\in\mathcal{Q}^d\colon \cD(C^*,Q) \leq \delta\}$. In this case the bounds on $\{Q\in\mathcal{Q}^d\colon \cD(C^*,Q) \leq \delta\}$ are provided by $$\begin{aligned} &\underline{Q}^{\cD,\delta}(\bu)=\min\Big\{\alpha\in[W_d(\bu),M_d(\bu)]\colon \cD\Big({C^*},{\overline{Q}}^{\{\bu\},\alpha}\wedge{C^*}\Big) \leq \delta\Big\},\\ &\overline{Q}^{\cD,\delta}(\bu)=\max\Big\{\alpha\in[W_d(\bu),M_d(\bu)]\colon \cD\Big({C^*},{\underline{Q}}^{\{\bu\},\alpha}\vee{C^*}\Big)\leq \delta\Big\}.\end{aligned}$$ Many well-known statistical distances satisfy the requirements of Theorem \[prescribedDistance\]. Typical examples are the Kolmogorov–Smirnov and the Cramér–von Mises distances, where $$\begin{aligned} \cD_{\text{KS}}(Q,Q') := \sup_{\bu\in\mathbb{I}^d}|Q(\bu)-Q'(\bu)| \quad \text{ and } \quad \cD_{\text{CM}}(Q,Q') := \int\nolimits_{\mathbb{I}^d} |Q(\bu)-Q'(\bu)|^2 \ud\bu. \end{aligned}$$ The same holds for all $L^p$ distances with $p\geq 1$, where $$\cD_{L^p}(Q,Q') := \Big(\int\nolimits_{\mathbb{I}^d} |Q(\bu)-Q'(\bu)|^p \ud\bu\Big)^{\frac{1}{p}}.$$ Distances with these properties are of particular interest in the theory of minimum distance and minimum contrast estimation, where—as opposed to maximum likelihood methods—parameters of distributions are estimated based on a statistical distance between the empirical and the estimated distribution. These estimators have favorable properties in terms of efficiency and robustness; *cf.* @spokoiny2015 [Chapter 2.8]. The computation of the bounds $\underline{Q}^{\cD,\delta}$ and $\overline{Q}^{\cD,\delta}$ in Theorem \[prescribedDistance\] involves the solution of optimization problems, which can be computationally intricate depending on the distance $\cD$. An explicit representation of the bounds is thus highly valuable for applications. The following result shows that in the particular case of the Kolmogorov–Smirnov distance the bounds can be computed explicitly. \[explicitBounds\] Let $C^*$ be a $d$-copula, $\delta \in \R_+$, and consider the Kolmogorov–Smirnov distance $\cD_{\emph{KS}}$. Then $$\begin{aligned} \underline{Q}^{\cD_{\emph{KS}},\delta}(\bu) = \max\big\{C^*(\bu)-\delta,W_d(\bu)\big\} \quad \text{ and } \quad \overline{Q}^{\cD_{\emph{KS}},\delta}(\bu) = \min\big\{C^*(\bu)+\delta,M_d(\bu)\big\}.\end{aligned}$$ Let us start with the lower bound $\underline{Q}^{\cD_{\text{KS}},\delta}$. Due to ${{\overline{Q}}^{\{\bu\},\alpha}\wedge{C^*}} \preceq {C^*}$ for all $\alpha\in[W_d(\bu),M_d(\bu)]$, it holds that $$\begin{aligned} \cD_{\text{KS}}\Big({{\overline{Q}}^{\{\bu\},\alpha}\wedge{C^*}},{C^*}\Big) &= \sup_{\bx\in\mathbb{I}^d} \Big|\big({{\overline{Q}}^{\{\bu\},\alpha}\wedge{C^*}}\big)(\bx)-{C^*}(\bx)\Big| = \sup_{\bx\in\mathbb{I}^d} \Big\{ {C^*}(\bx)-{\overline{Q}}^{\{\bu\},\alpha}(\bx) \Big\}.\end{aligned}$$ Since $\sup_{\bx\in\mathbb{I}^d} \big\{{C^*}(\bx)-{\overline{Q}}^{\{\bu\},\alpha}(\bx)\big\} = 0$ when $\alpha> C^*(\bu)$, we can assume w.l.o.g. that the minimum is attained for $\alpha\leq C^*(\bu)$. Hence $$\begin{aligned} \min\Big\{\alpha \in [W_d(\bu),M_d(\bu)] &\colon \cD_{\text{KS}}\Big({{\overline{Q}}^{\{\bu\},\alpha}\wedge{C^*}},{C^*}\Big) \leq \delta\Big\} \\ &= \min\Big\{\alpha \in \mathbb [W_d(\bu),C^*(\bu)] \colon \sup_{\bx\in\mathbb{I}^d} \Big\{{C^*}(\bx)-{\overline{Q}}^{\{\bu\},\alpha}(\bx)\Big\} \leq \delta\Big\}.\end{aligned}$$ Then, using the definition of ${\overline{Q}}^{\{\bu\},\alpha}$ in , we obtain $$\begin{aligned} \sup_{\bx\in\mathbb{I}^d} \Big\{{C^*}(\bx) - {\overline{Q}}^{\{\bu\},\alpha}(\bx)\Big\} &= \sup_{\bx\in\mathbb{I}^d} \Big\{{C^*}(\bx) - \min\Big\{M_d(\bx),\alpha+\sum_{i=1}^d (x_i-u_i)^+\Big\}\Big\}\\ &= \sup_{\bx\in\mathbb{I}^d} \Big\{{C^*}(\bx) - \alpha-\sum_{i=1}^d (x_i-u_i)^+\Big\}\\ &= \sup_{\bx\in\mathbb{I}^d} \Big\{{C^*}(\bx) -\sum_{i=1}^d (x_i-u_i)^+\Big\} - \alpha = C^*(\bu)-\alpha,\end{aligned}$$ where the second equality holds due to the fact that $C^*(\bx)-M_d(\bx)\leq 0$ for all $\bx \in \mathbb{I}^d$. Hence, we conclude that $$\begin{aligned} {\underline{Q}}^{\cD_{\text{KS}},\delta}(\bu) &= \min\big\{\alpha \in \mathbb [W_d(\bu),C^*(\bu)] \colon C^*(\bu)-\alpha \leq \delta\big\} \\ &= \min\big\{\alpha \in \mathbb [W_d(\bu),C^*(\bu)] \colon C^*(\bu)-\delta \leq \alpha\big\} = \max\big\{C^*(\bu)-\delta,W_d(\bu)\big\}.\end{aligned}$$ The proof for the upper bound $\overline{Q}^{\cD_{\text{KS}},\delta}$ is analogous, therefore omitted. Analogously to Theorem \[prescribedDistance\], one can also consider the situation where information on the survival copula is available. Note that each statistical distance that measures the discrepancy between quasi-copulas can easily be translated into a distance on quasi-survival functions, *i.e.* if $\cD$ is a statistical distance on $\mathcal{Q}^d\times\mathcal{Q}^d$, then $(\widehat{Q},\widehat{Q}')\mapsto \cD\big(\widehat{Q}(\mathbf 1-\cdot),\widehat{Q}'(\mathbf 1-\cdot)\big)$ defines a distance on the set of survival copulas or quasi-survival functions. \[prescribedDistanceSurvival\] Let $C^*$ be a $d$-copula and $\cD$ be a statistical distance which is continuous with respect to the pointwise convergence of quasi-copulas, monotonic with respect to the upper orthant order and min/max-stable. Consider the set $\widehat{\mathcal{Q}}^{\cD,\delta} = \big\{\widehat{Q}\in\widehat{\mathcal{Q}}^d\colon \cD(\widehat{Q},\widehat{C}^*) \leq \delta\big\}$ for $\delta\in\mathbb{R}_+$. Then $$\begin{aligned} \underline{\widehat{Q}}^{\cD,\delta}(\bu) :=& \min\Big\{\alpha\in\mathbb S(\bu)\colon \cD\Big(\widehat{\overline{Q}}^{\{\bu\},\alpha}\wedge{C^*},{C^*}\Big) \leq \delta\Big\} = \min\big\{C(\bu)\colon C\in\widehat{\mathcal{Q}}^{\cD,\delta}\big\} \\ \widehat{\overline{Q}}^{\cD,\delta}(\bu) :=& \max\Big\{\alpha\in\mathbb S(\bu)\colon \cD\Big(\widehat{\underline{Q}}^{\{\bu\},\alpha}\vee{C^*},{C^*}\Big) \leq \delta\Big\} = \max\big\{C(\bu)\colon C\in\widehat{\mathcal{Q}}^{\cD,\delta}\big\}.\end{aligned}$$ The proof is analogous to the proof of Theorem \[prescribedDistance\] and is therefore omitted. Numerical examples and illustrations {#numerics} ==================================== In this section we apply the results deduced in the previous parts in order to derive bounds on the Value-at-Risk that account for additional information on the dependence structure. In particular, we are able to include different types of partial dependence information that are both relevant for practical applications and have not been considered in the literature so far. Let us first recall the setting of Section \[prescribedMax\], where we showed that $$\begin{aligned} \underline{m}_{\mathcal{E},\max}(s) \le \P(X_1+\dots+X_d \le s) \le \overline{M}_{\mathcal{E},\max}(s);\end{aligned}$$ see Theorem \[boundMax\]. In order to compute the bounds $\underline{m}_{\mathcal{E},\max}(s)$ and $\overline{M}_{\mathcal{E},\max}(s)$, we first need to choose a method to estimate the probability $\P(\alpha_1Y_1+\cdots+\alpha_mY_m\leq s)$ for fixed $(\alpha_1,\dots,\alpha_m)$ in $\overline{\mathcal{A}}$ or $\underline{\mathcal{A}}$ and $Y_i\sim G_i$, $i=1,\dots,m$. This corresponds to a standard Fréchet problem over a class of distributions with fixed marginals. Thus, two approaches lend themselves naturally for this task: an approximation by the standard bounds given in or by the rearrangement algorithm. Indeed, we can use the standard bounds in to estimate $$\begin{aligned} \max\Big\{0,\sup_{\mathcal{U}(s)}\sum_{i=1}^m G^-_i\Big(\frac{u_i}{\alpha_i}\Big)-m+1\Big\} &\le \P(\alpha_1Y_1+\cdots+\alpha_mY_m\leq s) \\ &\qquad \le \min\Big\{1,\inf_{\mathcal{U}(s)}\sum_{i=1}^m G^-_i\Big(\frac{u_i}{\alpha_i}\Big)\Big\},\end{aligned}$$ where $\mathcal{U}(s) = \{(u_1,\dots,u_m)\in\mathbb{R}^m\colon u_1+\cdots+u_m=s\}$ and $G_i^-$ denotes the left-continuous version of $G_i$. Then, the bounds $\underline{m}_{\mathcal{E},\max}$ and $\overline{M}_{\mathcal{E},\max}$ are estimated by $$\begin{aligned} \label{computeBounds} \begin{split} &\underline{m}_{\mathcal{E},\max}(s) \geq \sup_{(\alpha_1,\dots,\alpha_m)\in\underline{\mathcal{A}}} \max\Big\{0,\sup_{\mathcal{U}(s)}\sum_{i=1}^m G^-_i\Big(\frac{u_i}{\alpha_i}\Big)-m+1\Big\}, \\ &\overline{M}_{\mathcal{E},\max}(s) \leq \inf_{(\alpha_1,\dots,\alpha_m)\in\overline{\mathcal{A}}} \min\Big\{1,\inf_{\mathcal{U}(s)}\sum_{i=1}^m G^-_i\Big(\frac{u_i}{\alpha_i}\Big)\Big\}. \end{split}\end{aligned}$$ Similarly, for fixed $(\alpha_1,\dots,\alpha_m)\in\underline{\mathcal{A}}$, the rearrangement algorithm allows us to approximate the bound $$\begin{aligned} \label{lowerRA} \inf\big\{\P(\alpha_1Y_1+\cdots+\alpha_mY_m\leq s)\colon Y_n\sim G_n, n \in\mathcal{J}\big\},\end{aligned}$$ while for $(\alpha_1,\dots,\alpha_m)\in\overline{\mathcal{A}}$ we can approximate $$\begin{aligned} \label{upperRA} \sup\big\{\P(\alpha_1Y_1+\cdots+\alpha_mY_m\leq s)\colon Y_n\sim G_n, n \in\mathcal{J}\big\}.\end{aligned}$$ To this end, we need to suitably discretize the variables $\alpha_1Y_1,\cdots,\alpha_mY_m$ and apply the rearrangement algorithm to the resulting matrix; for further details see [-@embrechts2013]. Denoting the lower bound in computed by means of the rearrangement algorithm by $\underline{RA}(\alpha_1 Y_1,...,\alpha_m Y_m)$ and analogously the upper bound in by $\overline{RA}(\alpha_1 Y_1,...,\alpha_m Y_m)$, we thus obtain the following estimates: $$\begin{aligned} \label{computeBoundsRA} \begin{split} &\underline{m}_{\mathcal{E},\max}(s) \geq \sup_{(\alpha_1,\dots,\alpha_m)\in\underline{\mathcal{A}}} \underline{RA}(\alpha_1 Y_1,...,\alpha_m Y_m), \\ &\overline{M}_{\mathcal{E},\max}(s) \leq \inf_{(\alpha_1,\dots,\alpha_m)\in\overline{\mathcal{A}}} \overline{RA}(\alpha_1 Y_1,...,\alpha_m Y_m). \end{split}\end{aligned}$$ Let us stress that the rearrangement algorithm has favorable numerical properties compared to the improved standard bounds. In particular, the bounds $\underline{RA}(\alpha_1 Y_1,...,\alpha_m Y_m)$ and $\overline{RA}(\alpha_1 Y_1,...,\alpha_m Y_m)$ can be computed very quickly for a reasonably fine discretization, thus the subsequent optimization over the set $\underline{\mathcal{A}}$ and $\overline{\mathcal{A}}$ can be performed much faster. The following example illustrates the improvement achieved by the VaR bounds in this setting, that accounts for extreme value information. \[exExtremeValue\] We consider a homogeneous portfolio $\mathbf{X} = (X_1,\dots,X_6)$ where the marginals are Pareto-2 distributed, *i.e.* $X_1,\dots,X_6\sim\text{Pareto}_2$, and analyze the improvement of the VaR bounds when additional information on the dependence structure is taken into account. In particular, we assume that the distributions $G_n$ of the maxima $\max_{j\in J_n} X_j$ are known for $J_1=\{1,2,3\}$ and $J_2=\{4,5,6\}$. In this case, it follows from Theorem \[boundMax\] and equation , that $$\begin{aligned} \label{exExtremeValueEq1} \begin{split} & \sup_{(\alpha_1,\dots,\alpha_8)\in\underline{\mathcal{A}}} \underline{RA}(\alpha_1 Y_1,\alpha_2 Y_2,\alpha_3 X_1,...,\alpha_{8} X_6) \\ & \quad \leq \inf\Big\{\P(X_1+\cdots+X_6\leq s)\colon X_1,\dots,X_6\sim\text{Pareto}_2, \max_{j\in J_n} X_n \sim G_n, n = 1,2\Big\} \end{split}\end{aligned}$$ and analogously $$\begin{aligned} \label{exExtremeValueEq2} \begin{split} & \inf_{(\alpha_1,\dots,\alpha_8)\in\overline{\mathcal{A}}} \overline{RA}(\alpha_1 Y_1,\alpha_2 Y_2,\alpha_3 X_1,...,\alpha_{8} X_6)\\ &\quad \geq \sup\Big\{\P(X_1+\cdots+X_6\leq s)\colon X_1,\dots,X_6\sim\text{Pareto}_2, \max_{j\in J_n} X_n \sim G_n, n = 1,2\Big\}. \end{split}\end{aligned}$$ Note that the marginals $X_1,\dots,X_6$ appear in the optimization since the distribution of the maximum of every individual variable is known and equals the respective marginal distribution; *i.e.* $\max\{X_i\} = X_i\sim F_i$ for $i=1,\dots,d$; see again Remark \[rem:calE\]. The marginal distributions are thus accounted for in the computation of the bounds. The solution of the optimization problems in and yields bounds on the VaR of the sum $X_1+\cdots+X_6$ when the distribution of the partial maxima is taken into account. Table \[tab:extremeValue\] shows the $\alpha$ confidence level in the first column and the VaR bounds without additional information in the second column, *i.e.* the unconstrained bounds. The third and fourth columns contain the improved VaR bounds that account for the extreme value information, as well as the improvement over the unconstrained bounds in percentage terms. In order to illustrate our method, we need to know the distribution of the partial maxima. To this end, we assume that the vectors $(X_1,X_2,X_3)$ and $(X_4,X_5,X_6)$ have the same Student-$t$ copula with equicorrelation matrices and two degrees of freedom, and numerically determine the distribution of $\max\{X_1,X_2,X_3\}$ and $\max\{X_4,X_5,X_6\}$. In the third column it is assumed that the pairwise correlations of $(X_1,X_2,X_3)$ and $(X_4,X_5,X_6)$ are equal to 0.9 and in the fourth column the pairwise correlations amount to 0.7 respectively. All bounds in this table, both without and with additional information, are computed using the rearrangement algorithm, while we have also performed the same computations using the standard bounds in . In the case with additional information, the standard bounds and the rearrangement algorithm yield the same results. In the case without additional information, the rearrangement algorithm clearly outperforms the standard bounds, which is well documented, hence we do not report any of these results here. $\alpha$ lower upper ---------- ------- ------- ------ ------- ------ ------ ------- ------ 95% 3.8 47.8 3.8 39.5 19.7 4.9 44.8 9.1 99% 4.9 114.0 11.0 96.1 22.0 12.4 107.8 12.5 99.5% 5.2 163.7 16.1 138.5 22.7 18.0 155.1 13.5 : Unconstrained and improved VaR bounds for the sum $X_1+\cdots+X_6$ with known distribution of partial maxima for different confidence levels.[]{data-label="tab:extremeValue"} The following observations ensue from this example: (i) The addition of partial dependence information allows to notably reduce the spread between the upper and lower bounds. Indeed, the bounds with additional information are finer than the unconstrained bounds resulting from the rearrangement algorithm, which are approximately sharp in this setting. Nevertheless, the model risk is still not negligible. (ii) The level of improvement *increases* with increasing confidence level $\alpha$. This is in contrast to related results in the literature, see *e.g.* [@bernard2015; @bignozzi2015], where the improvement typically decreases as the confidence level increases, and is an advantage of the present methodology. (iii) The improvement is more pronounced in the high-correlation scenario, and for the lower bound. These two observations are in accordance with the related literature; *e.g.* [@puccetti2016] report also a more pronounced improvement of the VaR bounds in the presence of strong positive dependence (especially in the tails), while [@bernard2015] report a more noticable improvement of the lower relative to the upper VaR bound. In the next example we combine the results of Section \[boundsOnCopula\] with Proposition \[varBoundsMax\] in order to derive improved bounds on the VaR of the maximum of risks over a class of copulas in the proximity of a reference copula. \[exMaximumMinimum\] Let us consider a homogeneous portfolio of three risks $(X_1,X_2,X_3)$ where the marginals are again Pareto 2 distributed, *i.e.* $X_1,X_2,$ $X_3\sim\text{Pareto}_2$. We assume that the reference copula $C^*$ is a Student-$t$ copula with equicorrelation matrix and two degrees of freedom, and are interested in computing bounds on the VaR over the class of models in the $\delta$-neighborhood of $C^*$ as measured by the Kolmogorov–Smirnov distance. In other words, we consider the class $$\mathcal{C}^{\cD_{\text{KS}},\delta}:= \big\{C\in\mathcal{C}^d\colon \cD_{\text{KS}}(C,C^*) \leq \delta \big\},$$ and using Theorem \[prescribedDistance\] and Lemma \[explicitBounds\] we arrive at bounds on the copulas in $\mathcal{C}^{\cD_{\text{KS}},\delta}$. Then, we apply Proposition \[varBoundsMax\] using the bounds $\underline{Q}^{\cD_{\text{KS}},\delta}$ and $\overline{Q}^{\cD_{\text{KS}},\delta}$ obtained above in order to compute bounds on the VaR of the maximum $\max\{X_1,X_2,X_3\}$ over the class of models in the vicinity of $C^*$. Table \[tab:distance1\] shows the confidence level and the sharp unconstrained (*i.e.* marginals-only) VaR bounds in the first two columns. The third, fourth and fifth column contain the upper and lower VaR bounds which use the information on the distance from $C^*$, for different levels of the threshold $\delta$, as well as the improvement over the unconstrained bounds in percentage terms. In the computation we assume that the pairwise correlation of the $t$-copula $C^*$ equals 0.9. The results are rounded to one decimal digit for the sake of legibility. The next table is analogous to Table \[tab:distance1\], but this time weaker dependence is induced by the reference model, assuming that the pairwise correlations in the $t$-copula $C^*$ are equal to 0.6. Let us point out that the bounds in Proposition \[varBoundsMax\], hence also in the second column of Tables \[tab:distance1\] and \[tab:distance2\], are sharp when no dependence information is available, *i.e.* when $\underline{Q} = W_3$ and $\overline{Q} = M_3$. This is due to the fact that $M_3$ is a copula and $W_3$ is pointwise best-possible. The observations made for the previous example are largely valid also in the present one, namely: (i) The addition of partial information reduces significantly the spread between the upper and lower bounds. This reduction is more pronounced as the threshold $\delta$ decreases; in other words, the more reliable the reference model, the more pronounced the reduction of model risk. These results should be compared, qualitatively, with analogous results for the ‘trusted region’ in [@bernard2015]. (ii) The level of improvement decreases in this case, sometimes dramatically, with increasing confidence level $\alpha$. In particular, for $\alpha = 99\%$ the improvement was small, especially for large values of $\delta$. (iii) The improvement is more pronounced in the high-dependence scenario, with improvements over the sharp unconstrained bounds of up to 81%.

ArXiv

--- abstract: 'The slow dynamics of nearly stationary patterns in a FitzHugh-Nagumo model are studied using a phase dynamics approach. A Cross-Newell phase equation describing slow and weak modulations of periodic stationary solutions is derived. The derivation applies to the bistable, excitable, and the Turing unstable regimes. In the bistable case stability thresholds are obtained for the Eckhaus and the zigzag instabilities and for the transition to traveling waves. Neutral stability curves demonstrate the destabilization of stationary planar patterns at low wavenumbers to zigzag and traveling modes. Numerical solutions of the model system support the theoretical findings.' author: - Aric Hagberg - Ehud Meron - Thierry Passot bibliography: - 'reaction.bib' - 'hagberg.bib' title: ' **Phase Dynamics of Nearly Stationary Patterns in Activator-Inhibitor Systems** ' --- Introduction ============ Studies of stationary patterns in activator-inhibitor systems have focused primarily on localized structures such as pulses and spots in excitable and bistable media [@TyKe:88; @KeOs:89; @meron:92; @KM:94; @LeSw:95; @MuOs:96; @WSBOP:96; @SOBP:97], and periodic patterns near a Turing bifurcation [@CDBD:90; @OuSw:91n; @AAP:97]. Localized structures have instabilities to traveling patterns, breathing motion, and transverse deformations [@KeOs:89; @OMK:89; @HaMe:94a; @GMP:96]. Periodic patterns have been analyzed near the onset of a Turing instability and also near the codimension-two point of a Turing instability and a Hopf bifurcation [@RoMe:92; @HBP:93; @PDDK:93; @DLDB:96; @OB:98]. But very few studies have explored instabilities of [*periodic*]{} (nonlocal) stationary patterns in excitable and bistable media, or of periodic stationary patterns [*far beyond*]{} the Turing instability [@DK:89; @Osipov:96]. The latter case includes pattern formation studies on the CIMA chemical reaction [@PDDK:93; @OuSw:91; @DP:94]. In this paper we study instabilities of stationary periodic patterns by deriving a Cross-Newell phase equation [@KBBC:82; @CN:84; @NPBEI:96]. The derivation is not restricted to the immediate neighborhood of a Turing instability and applies to periodic patterns with space-scale separation that arise far from onset or in excitable and bistable media. The Cross-Newell equation was originally derived in the context of fluid dynamics and has recently been applied in a laser system [@LMN:96]. We choose to study the FitzHugh-Nagumo (FHN) equations, a canonical model for activator-inhibitor systems, $$\begin{aligned} \label{fhn} \frac{\partial u}{\partial t}&=&u-u^3-v+\nabla^2 u\,, \\ \frac{\partial v}{\partial t}&=&\epsilon(u-a_1v-a_0)+ \delta\nabla^2 v \,. \nonumber\end{aligned}$$ Here, $u$ is the activator and $v$ the inhibitor. The parameters $a_0$ and $a_1$ can be chosen so that the FHN model (\[fhn\]) represents an excitable medium, a bistable medium, or a system with a Turing instability [@HaMe:94a]. All three cases support stationary periodic solutions for $\delta$ sufficiently large. In Section \[phaeqn\] we derive a phase equation describing weak modulations of periodic stripe pattern in the FHN model. In Section \[stabthre\] we evaluate stability thresholds for Eckhaus and zig-zag instabilities and for a transition from stationary to traveling patterns. These thresholds suggest a number of spatial or spatio-temporal behaviors which we test in Section \[simulations\] with numerical solutions of Eqs. (\[fhn\]). The Phase Equation {#phaeqn} ================== Let $u_0(\theta;k)=u_0(\theta+2\pi;k)$, $v_0(\theta;k)=v_0(\theta+2\pi;k)$ be a stationary periodic solution of Eqs. (\[fhn\]) with phase $\theta$ and wavenumber $k$. We consider weak spatial modulations of this periodic pattern and assume that those modulations have a length scale $L$ that is much larger than the wavelength $1/k$. The ratio of the length scales $\lambda=1/(kL)$ can then be used as a small parameter to write modulated solutions as an asymptotic expansion about the periodic solution $$\begin{aligned} u(\theta,{\bf R},T)=u_0(\theta;k)+\lambda u_1(\theta,{\bf R},T) + \lambda^2 u_2(\theta,{\bf R},T) + ...\nonumber \\ v(\theta,{\bf R},T)=v_0(\theta;k)+\lambda v_1(\theta,{\bf R},T)+ \lambda^2 v_2(\theta,{\bf R},T) + ... \label{exp}\end{aligned}$$ where ${\bf R=\lambda{\bf r}}$ and $T=\lambda^2 t$ are slow space and time variables. The phase $\theta$ in Eq. (\[exp\]) is an undetermined function of space and time and $k=|{\bf k}|=|\nabla\theta|$ is the local wavenumber. Our objective is to derive an equation for the slow phase $$\Theta({\bf R},T):=\lambda\theta({\bf r},{\bf R},T)\,.$$ In terms of this phase the local wavevector is $${\bf k}({\bf R},T)=\nabla_R\Theta\,.$$ Inserting the expansions (\[exp\]) in Eqs. (\[fhn\]) we find at order unity \[order\_0\] $$\begin{aligned} u_0-u_0^3-v_0+ k^2 \frac{\partial^2 u_0}{\partial\theta^2}&=&0\,, \label{order_0_U}\\ \epsilon(u_0-a_1v_0-a_0)+ \delta k^2\frac{\partial^2 v_0}{\partial\theta^2} &=&0 \,, \label{order_0_V}\end{aligned}$$ where $k^2={\bf k}\cdot{\bf k}$. At order $\lambda$ \[order\_1\] $$\begin{aligned} \left(k^2\frac{\partial^2}{\partial\theta^2}+1-3u_0^2\right)u_1 - v_1 &=&{\cal D}\frac{\partial u_0}{\partial\theta}\,, \\ \epsilon u_1 + \left(\delta k^2 \frac{\partial^2}{\partial\theta^2} - \epsilon a_1\right)v_1 &=&{\cal D}\frac{\partial v_0}{\partial\theta}\,,\end{aligned}$$ where $$\begin{aligned} {\cal D}=\frac{\partial\Theta}{\partial T}-\nabla_R\cdot{\bf k} -2{\bf k}\cdot\nabla_R\,.\end{aligned}$$ Projecting the right hand side of (\[order\_1\]) onto $(\partial_\theta u_0, -\epsilon^{-1}\partial_\theta v_0)$, the solution of the adjoint problem, produces the phase equation $$\tau\frac{\partial\Theta}{\partial T}=-\nabla_R\cdot({\bf k}B)\,,$$ where $$\begin{aligned} \label{tau} \tau&=&<(\partial_\theta u_0)^2>-\epsilon^{-1}<(\partial_\theta v_0)^2>\,, \\ B&=&-<(\partial_\theta u_0)^2>+\delta\epsilon^{-1} <(\partial_\theta v_0)^2>\,. \label{B}\end{aligned}$$ In these equations $<(.)>:=\frac{1}{2\pi}\int_0^{2\pi}(.)d\theta$. The quantities $B$ and $\tau$ contain information about various instabilities of the periodic stripe pattern. The condition $\frac{d}{dk}[kB(k)]=0$ implies the onset of an Eckhaus instability and the condition $B=0$ the onset of a zigzag instability [@CN:84]. In Appendix A we show that the condition $\tau=0$ indicates a transition to traveling waves. To implement these conditions we need to solve Eqs. (\[order\_0\]) for the periodic solution $(u_0,v_0)$. For parameter values that satisfy $\epsilon/\delta:=\mu\ll 1$ an approximate solution can be computed as shown in Appendix B. Using this solution to calculate $\tau$ and $B$, as shown in Appendix C, gives the following expressions: $$\begin{aligned} \tau&=&\frac{2\sqrt{2}}{3\pi k} - \frac{v_-}{q\pi k\eta}\beta(\Lambda_-) \gamma(\Lambda_-,\Lambda_+)\,, \nonumber \\ B&=&\frac{2\sqrt{2}}{3\pi k} - \frac{v_-}{q\pi k\sqrt{\mu}}\beta(\Lambda_-) \gamma(\Lambda_-,\Lambda_+)\,, \label{tauB}\end{aligned}$$ $$\Lambda_-+\Lambda_+=\frac{2\pi\sqrt{\mu}}{k} \,, \label{lpm1}$$ $$v_+\beta(\Lambda_+)+v_-\beta(\Lambda_-)=0\,, \label{lpm2}$$ where $\mu=\epsilon/\delta$, $\eta=\sqrt{\epsilon\delta}$, $v_\pm=(\pm 1-a_0)/q^2$, $q^2=a_1+1/2$, $$\beta(x)=\coth{qx}-{\rm csch}{qx}\,, \label{beta}$$ and $$\begin{aligned} \gamma(\Lambda_-,\Lambda_+)&=&-1+\frac{1}{2}(1+a_0)q\Lambda_-{\rm csch}{q\Lambda_-} \nonumber\\ &&\mbox{} +\frac{1}{2}(1-a_0)q\Lambda_+{\rm csch}{q\Lambda_+}\,.\end{aligned}$$ The quantities $\Lambda_+$ and $\Lambda_-$ denote the widths of domains with high and low values of $u$ and $v$, respectively. The width is measured with respect to the spatial coordinate $z=\frac{\sqrt{\mu}}{k}\theta$ (see Appendix B). Given $k$, Eqs. (\[lpm1\]) and (\[lpm2\]) can be solved for $\Lambda_+(k)$ and $\Lambda_-(k)$. Using these solutions in Eq. (\[tauB\]) graphs of $\tau$ and $kB$ as functions of $k$ can be produced. Stability Thresholds {#stabthre} ==================== Explicit forms for $\tau(k)$ and $B(k)$ are available in the symmetric case, $a_0=0$, where $\Lambda_+=\Lambda_-=\frac{\pi\sqrt{\mu}}{k}$: $$\begin{aligned} \label{tau_sym_bis} \tau(k)=\frac{1}{\pi k\eta_c q^3}\left[1-\frac{\eta_c}{\eta}f(\pi q\sqrt{\mu}/k)\right]\,, \nonumber \\ B(k)= \frac{1}{\pi k\eta_c q^3}\left[1-\frac{\eta_c}{\sqrt{\mu}}f(\pi q\sqrt{\mu}/k)\right]\,, \label{B_sym_bis}\end{aligned}$$ where $\eta_c=\frac{3}{2\sqrt{2}q^3}$ and $$f(x)=(1-x~{\rm csch}{x})(\coth{x}-{\rm csch}{x})\,.$$ Figure \[fig:tkb-bistable\] shows graphs of $\tau(k)$ and $kB(k)$ for a bistable medium obtained with Eqs. (\[tau\_sym\_bis\]) (thick lines) and with Eqs. (\[tau\]) and (\[B\]) using numerically calculated solutions $u_0, v_0$ (circles). A very good agreement is obtained within the validity range of the analysis, $k\sim {\cal O}(\sqrt{\mu})\ll 1$. For $k\sim{\cal O}(1)$ the deviations become large. In particular the minimum of $kB(k)$ which designates the Eckhaus instability threshold, is not reproduced by the analytical form (\[B\_sym\_bis\]). ![ Typical functions $\tau(k)$ and $kB(k)$ for a bistable medium. The curves represent the functions of Eqs. (\[tau\_sym\_bis\]). The circles are numerically computed solutions using Eqs. (\[tau\]) and (\[B\]). The point $kB=0$ indicates the boundary between stable stationary stripes and zigzag patterns. At $\tau=0$ the pattern becomes unstable to traveling waves. Parameters: $a_1=4$, $a_0=0$, $\epsilon=0.001$, $\delta=2.0$. []{data-label="fig:tkb-bistable"}](tkbb){width="3.25in"} The instability to traveling waves occurs at $\tau=0$ or at $$\epsilon=\eta_c^2 f^2(\pi q\sqrt{\mu}/k)\delta^{-1}\,. \label{trav_thres}$$ The zigzag instability occurs at $B=0$ or at $$\epsilon=\eta_c^2 f^2(\pi q\sqrt{\mu}/k)\delta \,. \label{zigzag_thres}$$ The condition $\frac{d}{dk}(kB)=0$ for the Eckhaus instability becomes $$\frac{df}{dx}|_{x=\pi q\sqrt{\mu}/k}=0\,.$$ Consider first the limit $k\to 0$ in which the periodic pattern approaches an array of isolated front structures. In this limit $f(\pi q\sqrt{\mu}/k)\to 1$ and the condition for the onset of traveling waves becomes $\epsilon=\eta_c^2\delta^{-1}$. This is precisely the nonequilibrium Ising-Bloch (NIB) bifurcation point, where a stationary front loses stability to a pair of counter-propagating fronts. The condition for the zigzag instability becomes $\epsilon=\eta_c^2\delta$. This is the threshold for the transverse front instability [@HaMe:94c]. The neutral stability curves for a bistable medium corresponding to Eqs. (\[trav\_thres\]) and (\[zigzag\_thres\]) are shown in Figs. \[fig:neutral\]a and  \[fig:neutral\]b for fixed $\delta$ and $\epsilon$, respectively. They imply that high wavenumber stationary planar patterns are stabilized against zigzag and traveling wave instabilities. Notice that for $\delta=1$ the neutral stability curves $\tau=0$ and $B=0$ coincide (see Eqs. (\[trav\_thres\]) and (\[zigzag\_thres\]) or Fig. \[fig:neutral\]b). For $\delta>1$, upon decreasing the wavenumber at constant $\epsilon$, a high wavenumber pattern is destabilized to a zigzag pattern, whereas for $\delta<1$ the destabilization is to traveling waves. Similar neutral stability curves are found for the nonsymmetric case, $a_0\ne 0$, for excitable media and for systems (far) beyond the Turing instability. [ ![ The neutral stability boundaries for the zigzag instability ($B=0$; thick, solid curve) and traveling wave instability ($\tau=0$; thin, dashed curve). To the left of the $B=0$ curve planar periodic patterns are unstable to zigzag patterns. To the left of the $\tau=0$ curve planar periodic patterns are unstable to traveling waves. Parameters: $a_1=2$, $a_0=0$, $\epsilon=0.01$, $\delta=2$. []{data-label="fig:neutral"}](nbe "fig:"){width="3.25in"} ![ The neutral stability boundaries for the zigzag instability ($B=0$; thick, solid curve) and traveling wave instability ($\tau=0$; thin, dashed curve). To the left of the $B=0$ curve planar periodic patterns are unstable to zigzag patterns. To the left of the $\tau=0$ curve planar periodic patterns are unstable to traveling waves. Parameters: $a_1=2$, $a_0=0$, $\epsilon=0.01$, $\delta=2$. []{data-label="fig:neutral"}](nbd "fig:"){width="3.25in"} ]{} Comparisons with Numerical Solutions {#simulations} ==================================== We have computed numerical solutions of Eqs. (\[fhn\]) to test the stabilization of zigzag and traveling-wave instabilities at high wavenumbers. Figure \[fig:zigzag\] shows a low wavenumber zigzag pattern and a high wavenumber planar pattern computed for the same parameter values. This behavior is well known in other contexts [@CrHo:93]. The zigzag instability is a mechanism by which the system locally increases the wavenumber. Figure \[fig:traveling\] shows coexistence of a low wavenumber traveling wave and a high wavenumber stationary pattern. These numerical results are for a bistable system but similar results are found for excitable and Turing unstable systems. Coexistence of stationary and traveling waves has been found in experiments on the CIMA reaction [@PDDK:93; @DP:94]. and analyzed using different theoretical approaches [@DK:89; @IO:92; @KO:95; @Osipov:96]. ![ Coexistence of zigzag and planar patterns. The dark areas indicate regions of $u>0$ and the light regions $u<0$. At higher wavenumbers (top) the planar stripe solution is stable. At low wavenumbers (bottom) the planar solution is unstable and forms a zigzag pattern. Parameters: $a1=2$, $a_0=0$, $\epsilon=0.05$, $\delta=2$. []{data-label="fig:zigzag"}](planar "fig:"){width="3.0in"} ![ Coexistence of zigzag and planar patterns. The dark areas indicate regions of $u>0$ and the light regions $u<0$. At higher wavenumbers (top) the planar stripe solution is stable. At low wavenumbers (bottom) the planar solution is unstable and forms a zigzag pattern. Parameters: $a1=2$, $a_0=0$, $\epsilon=0.05$, $\delta=2$. []{data-label="fig:zigzag"}](zigzag "fig:"){width="3.0in"} ![ Coexistence of traveling waves and stationary waves. At high wavenumber (top) the patterns are stationary and at low wavenumber patterns (bottom) they travel. Parameters: $a1=2$, $a_0=0$, $\epsilon=0.03$, $\delta=2$. []{data-label="fig:traveling"}](stationary "fig:"){width="3.0in"} ![ Coexistence of traveling waves and stationary waves. At high wavenumber (top) the patterns are stationary and at low wavenumber patterns (bottom) they travel. Parameters: $a1=2$, $a_0=0$, $\epsilon=0.03$, $\delta=2$. []{data-label="fig:traveling"}](traveling "fig:"){width="3.0in"} We have also tested the condition for the Eckhaus instability in a bistable system using numerical computations of $\tau$ and $B$. Choosing wavenumbers $k>k_c$ where $k_c$ corresponds to the minimum of $kB$, we found initial periodic patterns either collapse to uniform states or to a lower wavenumber pattern through phase slips. Fig. \[fig:eckhaus\] demonstrates these two cases. Similar conclusions hold for excitable systems. An unstable Turing pattern, on the other hand, always converges to a lower wavenumber pattern since the single uniform state is unstable. ![ Time evolution of a periodic pattern in the region of Eckhaus instability. The high wavenumber pattern is unstable and either converges to one of the uniform states (top) or a lower wavenumber pattern (bottom). Parameters: $a1=2$, $a_0=0$, $\delta=2$; top: $\epsilon=0.01$, bottom: $\epsilon=0.1$. []{data-label="fig:eckhaus"}](eckhaus-uniform "fig:"){width="3.0in"} ![ Time evolution of a periodic pattern in the region of Eckhaus instability. The high wavenumber pattern is unstable and either converges to one of the uniform states (top) or a lower wavenumber pattern (bottom). Parameters: $a1=2$, $a_0=0$, $\delta=2$; top: $\epsilon=0.01$, bottom: $\epsilon=0.1$. []{data-label="fig:eckhaus"}](eckhaus-phase "fig:"){width="3.0in"} Conclusion {#con} ========== We have shown that the Cross-Newell phase equation provides a powerful tool for studying instabilities of stationary periodic patterns in activator-inhibitor systems. The equation contains information not only on the Eckhaus and zigzag instabilities, but also on the destabilization of stationary periodic patterns to traveling waves. The same equation applies to bistable, excitable, and Turing unstable systems. Equations of that kind should prove useful in identifying parameters and initial conditions where zigzag and Eckhaus instabilities couple to traveling wave modes. Such coupling may lead to complex spatiotemporal behavior analogous to the coupling of the NIB front bifurcation to a transverse front instability [@HaMe:94c; @HaMe:94b] This study was supported in part by grant No 95-00112 from the US-Israel Binational Science Foundation (BSF) and by the Department of Energy, under contract W-7405-ENG-36. The Meaning of $\tau=0$ ======================= We show here that the condition $\tau=0$ defines the critical value of $\epsilon$ at which traveling solutions bifurcate from the stationary solution. We look for traveling solutions $u(\theta), v(\theta)$ of Eqs. (\[fhn\]), where $\theta=kx-\omega t$, that bifurcate from the stationary solution branch $\omega=0$ at some $\epsilon=\epsilon_c$. Near the bifurcation where $\omega\ll 1$ we can expand the traveling solutions as power series in $\omega$ around the stationary solution $u_0,v_0$: $$\begin{aligned} u(\theta)=u_0(\theta)+\omega u_1(\theta)+ ... \,, \nonumber \\ v(\theta)=v_0(\theta)+\omega v_1(\theta)+ ... \,, \label{expansion}\end{aligned}$$ Expanding $\epsilon$ as $$\epsilon=\epsilon_c+\epsilon_1\omega +...\,,$$ and using these expansions in Eqs. (\[fhn\]) we find at order $\omega$ $$\begin{aligned} \left(k^2\frac{\partial^2}{\partial\theta^2}+1-3u_0^2\right)u_1 - v_1 &=&-{u_0}^{\prime}\,, \nonumber \\ \epsilon_c u_1 + \left(\delta k^2\frac{\partial^2}{\partial\theta^2} - \epsilon a_1\right)v_1 &=&-v_0^{\prime}-\epsilon_1(u_0-a_1v_0-a_0)\,. \label{A_order_1} \nonumber\end{aligned}$$ Projecting the right hand side onto $\left(u_0^{\prime}, -\epsilon_c^{-1} v_0^{\prime}\right)$ gives $$\epsilon_c=\frac{<{v_0^{\prime}}^2>}{<{u_0^{\prime}}^2>}\,, \label{epsilon_c}$$ where we used Eq. (\[order\_0\_V\]) and switched to the notation of a prime for the derivative with respect to the single argument $\theta$. Using the definition (\[tau\]) of $\tau$ and (\[epsilon\_c\]) we find $$\tau=\left(1-\frac{\epsilon_c}{\epsilon}\right)<{u_0^{\prime}}^2>\,.$$ Thus, $\tau=0$ implies $\epsilon=\epsilon_c$ or the onset of traveling solutions. Approximate Stationary Solution =============================== For $\mu=\epsilon/\delta\ll 1$ a singular perturbation approach can be used to approximate the stationary solution $u_0(\theta),v_0(\theta)$. Rescaling the space coordinate as $z=\frac{\sqrt{\mu}}{k}\theta$, Eqs. (\[order\_0\]) become $$\begin{aligned} u_0-u_0^3-v_0+ \mu u_0^{\prime\prime} &=&0\,, \nonumber\\ u_0-a_1v_0-a_0+ v_0^{\prime\prime}&=&0 \,, \nonumber \end{aligned}$$ where the prime denotes now the derivative with respect to $z$. Since the small parameter $\mu$ multiplies the second derivative term $u_0^{\prime\prime}$, two types of spatial regions can be distinguished. Outer regions where $u_0(z)$ varies on a scale of order unity and the term $\mu u_0^{\prime\prime}$ is negligible, and inner regions where $u_0(z)$ varies on a very short scale of order $\sqrt{\mu}$ and the term $\mu u_0^{\prime\prime}$ cannot be neglected. In these regions, however, $v_0$ hardly changes. The analysis of the inner regions leads to the solutions $$u_0=\pm\tanh{\frac{\theta}{\sqrt{2}k}}, \qquad v_0=0 \,. \label{inner}$$ These solution represent front structures separating two types of outer regions: domains of high activator values, $u=u_+(v_0)$ (“up state”), and domains of low activator values $u=u_-(v_0)$ (“down state”), where $u_\pm(v_0)$ are the extreme roots of $u_0-u_0^3-v_0=0$. We look for periodic stationary solutions with wavelength $\Lambda=\Lambda_- + \Lambda_+$, where $\Lambda_+$ and $\Lambda_-$ are the widths of up and down states respectively. Consider now a down state spanning the spatial range $-\Lambda_-<z<0$ followed by an up state spanning the range $0<z<\Lambda_+$. The equations for $v$ at these outer regions are $$v_0^{\prime\prime} - q^2(v_0-v_-)=0, \qquad -\Lambda_-<z<0\,, \label{down}$$ with the boundary conditions $v_0(-\Lambda_-)=v_0(0)=0$, and $$v_0^{\prime\prime} - q^2(v_0-v_+)=0, \qquad 0<z<\Lambda_+ \,, \label{up}$$ with the boundary conditions $v_0(0)=v_0(\Lambda_+)=0$. In obtaining these equations we approximated $u_\pm(v_0)=\pm 1 -v_0/2$. This approximation is particularly good for bistable media with $a_0$ small and $a_1$ relatively large. These values restrict $v_0$ to a small range around $v_0=0$. For excitable media and systems undergoing Turing instability, a large value of $\delta$ might be needed to keep $v_0$ small. The solutions to Eqs. (\[down\]) and (\[up\]) are $$v_0=\frac{v_-}{\sinh{q\Lambda_-}}\left[\sinh{qz}-\sinh{q(z+\Lambda_-)}\right] +v_- \,, \label{soldown}$$ for $-\Lambda_-<z<0$, and $$v_0=\frac{v_+}{\sinh{q\Lambda_+}}\left[\sinh{q(z-\Lambda_+)}-\sinh{qz}\right] +v_+ \,, \label{solup}$$ for $0<z<\Lambda_+$. To determine $\Lambda_\pm$ for a given $\Lambda$ we match the derivatives of $v_0$ at the front positions $$v_0^{\prime}(0^+)= v_0^{\prime}(0^-), \qquad v_0^{\prime}(\Lambda_+)=v_0^{\prime}(-\Lambda_-)\,.$$ This leads to the relation $$v_+\beta(q\Lambda_+)+v_-\beta(q\Lambda_-)=0\,,$$ where $\beta(x)$ is given by Eq. (\[beta\]). Calculation of $\tau$ and $B$ ============================= The quantities $\tau$ and $B$ are given by Eqs. (\[tau\]) and (\[B\]). Consider first the integral $$<u_0^\prime(\theta)^2>=\frac{1}{2\pi}\int_0^{2\pi}u_0^\prime(\theta)^2d\theta \,.$$ It has a contribution from two inner regions at $z=0$ and $z=\Lambda_+$ where $u_0$ is given by Eq. (\[inner\]), and a contribution from two outer regions, $-\Lambda_-<z<0$ and $0<z<\Lambda_+$ where $u_0=-1-v_0/2$ and $u_0=1-v_0/2$ with $v_0$ given by Eq. (\[soldown\]) and Eq. (\[solup\]), respectively. (Recall that $z=\frac{\sqrt{\mu}}{k}\theta$). The contribution from the two inner regions is $$\begin{aligned} <u_0^\prime(\theta)^2>_{inner}&=&\frac{1}{2\pi k^2}\int_{inner}{\rm sech}^4 \left(\frac{\theta}{\sqrt{2}k}\right)d\theta \,, \nonumber \\ & \mbox{} \approx& \frac{1}{\sqrt{2}\pi k}\int_{-\infty}^\infty {\rm sech}^4 xdx= \frac{2\sqrt{2}}{3\pi k}\,. \nonumber\end{aligned}$$ We have used here the fact that $k\sim{\cal O}(\sqrt{\mu})\ll 1$. The integral over a narrow inner region is transformed into an integral over a wide region after stretching the $\theta$ variable to the $x=\frac{\theta}{\sqrt{2}k}$ variable. The contribution from the two outer regions is $$\begin{aligned} <u_0^\prime(\theta)^2>_{outer}&=& \frac{\sqrt{\mu}}{8\pi k} \nonumber \\ &\times& \left[\int_{-\Lambda_-}^0 {v_0(z)^\prime}^2dz + \int_{0}^{\Lambda_+} {v_0(z)^\prime}^2dz\right]\,, \nonumber\end{aligned}$$ where we used in the two outer regions $u_0(z)^\prime= -\frac{1}{2}v_0(z)^\prime$. Altogether, $$\begin{aligned} \label{u_int} <u_0^\prime(\theta)^2>&=&\frac{2\sqrt{2}}{3\pi k}+ \frac{\sqrt{\mu}}{8\pi k} \\ &\times&\left[\int_{-\Lambda_-}^0 {v_0(z)^\prime}^2dz + \int_{0}^{\Lambda_+} {v_0(z)^\prime}^2dz\right]\,. \nonumber\end{aligned}$$ The second term on the right hand side of Eq. (\[u\_int\]) is small (since $\sqrt{\mu}\ll 1$) and will not contribute to the leading order forms of $\tau$ and $B$. Consider now the integral $$<v_0^\prime(\theta)^2>=\frac{1}{2\pi}\int_0^{2\pi}v_0^\prime(\theta)^2d\theta \,.$$ The contribution from the inner regions to this integral is negligible for $\mu\ll 1$. Thus $$\begin{aligned} <v_0^\prime(\theta)^2>&=&\frac{\sqrt{\mu}}{2\pi k}\nonumber\\ &\times&\left[\int_{-\Lambda_-}^0 {v_0(z)^\prime}^2dz + \int_{0}^{\Lambda_+} {v_0(z)^\prime}^2dz\right]\,. \label{v_int}\end{aligned}$$ Using the solutions (\[soldown\]) and (\[solup\]) in the integrals (\[u\_int\]) and (\[v\_int\]) and using the expressions for $\tau$ and $B$ we obtain the expressions (\[tauB\]).

ArXiv

--- title: 'Search for [$b \rightarrow u$]{} transitions in [$B^- \to [\Kp\pim\piz]_D K^-$]{}' --- -[PUB]{}-[07]{}/[030]{}\ SLAC-PUB-[12694]{} short\_abstract pubboard/authors\_may2007.tex

ArXiv

--- abstract: | In this paper we make the first steps towards developing a theory of intersections of coisotropic submanifolds, similar to that for Lagrangian submanifolds. For coisotropic submanifolds satisfying a certain stability requirement we establish persistence of coisotropic intersections under Hamiltonian diffeomorphisms, akin to the Lagrangian intersection property. To be more specific, we prove that the displacement energy of a stable coisotropic submanifold is positive, provided that the ambient symplectic manifold meets some natural conditions. We also show that a displaceable, stable, coisotropic submanifold has non-zero Liouville class. This result further underlines the analogy between displacement properties of Lagrangian and coisotropic submanifolds. address: 'Department of Mathematics, UC Santa Cruz, Santa Cruz, CA 95064, USA' author: - 'Viktor L. Ginzburg' title: Coisotropic Intersections --- [^1] Introduction ============ [\[sec:intro\]]{} In this paper we make the first steps towards developing a theory of coisotropic intersections similar to that for Lagrangian submanifolds. The main objective of the paper is to establish persistence of coisotropic intersections under Hamiltonian diffeomorphisms for a certain class of coisotropic submanifolds, the so-called stable coisotropic submanifolds. We also show that displaceable, stable coisotropic submanifolds have non-zero Liouville class, generalizing the results of Bolle, [@Bo1; @Bo2], for submanifolds of $\R^{2n}$. Coisotropic intersections ------------------------- The Lagrangian intersection property or persistence of Lagrangian intersections is unquestionably one of the most fundamental results in symplectic topology. This result asserts that a Lagrangian submanifold necessarily intersects its image under a Hamiltonian diffeomorphism that is in some sense close to the identity, e.g., has sufficiently small energy. Depending on the notion of closeness and on the requirements imposed on the manifolds, various forms of the Lagrangian intersection property have been proved in [@Ch:Cambr; @Ch; @F:Morse; @F:grad; @F:c-l; @Fl; @Gr; @LS:lagr; @Oh1; @Oh:cambridge; @Oh:disj; @We73; @We74], to mention just some of the pertinent references. Recall that a submanifold $M$ of a symplectic manifold $(W^{2n},\omega)$ is said to be coisotropic if for every $p\in M$ the symplectic orthogonal $(T_pM)^\omega$ to the tangent space $T_pM$ is contained in $T_pM$. For instance, Lagrangian submanifolds are coisotropic, as are hypersurfaces in $W$. Furthermore, $\dim M\geq n$, when $M$ is coisotropic. The examples discussed below indicate that coisotropic submanifolds enjoy the same kind of Hamiltonian rigidity as Lagrangian submanifolds and lead to the following \[quest:intersections\] Can a coisotropic submanifold be displaced by a Hamiltonian diffeomorphism arbitrarily close to the identity in a suitable sense? For instance, interpreting closeness in the sense of Hofer’s metric (see, e.g., [@HZ; @Pol:book]), we can ask if there is a lower bound, depending only on the submanifold, on the energy of a Hamiltonian diffeomorphism displacing the submanifold. This question can be further restricted in a number of ways: by imposing additional assumptions on the codimension of $M$ or on the dynamics of the characteristic foliation or via other types of symplectic-topological requirements on $M$ and the ambient manifold, or through requiring the Hamiltonian diffeomorphism to be close to the identity in a particularly strong way. For instance, when $M$ is Lagrangian, the question reduces to the Lagrangian intersection property. In terms of codimension, the other extreme case is that of codimension zero: $M=W$. In this case, the coisotropic intersection property obviously holds. When $M$ is a hypersurface, the answer to the question is affirmative due to non-degeneracy of Hofer’s metric, [@LMD95], and the fact that a displaceable connected hypersurface necessarily bounds. The extreme case in terms of closeness of the diffeomorphism to the identity is that of infinitesimal displacement: \[ex:infinitesimal\] Let $M$ be a coisotropic submanifold and let $H$ be a Hamiltonian near $M$. On the infinitesimal level, the intersections of $M$ with its image under the Hamiltonian flow of $H$ correspond to the points where the Hamiltonian vector field $X_H$ of $H$ is tangent to $M$. These are precisely the leaf-wise critical points of $H$ along the characteristic foliation ${{\mathcal F}}$ on $M$. This observation suggests that the “amount” of coisotropic intersections is governed by the foliated Morse theory of ${{\mathcal F}}$. When $M$ is Lagrangian, this observation readily implies persistence of intersections for any Hamiltonian diffeomorphism generated by a $C^2$-small Hamiltonian, by the Weinstein symplectic neighborhood theorem, [@We71; @We77]. Here we answer Question \[quest:intersections\] affirmatively only for coisotropic manifolds satisfying a certain additional stability requirement (see Definition \[def:B\]), introduced by Bolle in [@Bo2] and similar to stability of hypersurfaces (cf.[@EKP; @HZ]). Namely, for such a submanifold we establish a lower bound on the Hofer norm of a displacing Hamiltonian diffeomorphism (see Theorem \[thm:main\]) and hence prove that the submanifold has positive displacement energy. The stability condition, discussed in detail in Section \[sec:ct\], is quite restrictive; see, in particular, Example \[exam:Bolle\]. (We also impose some natural and not-too-restrictive assumptions on the ambient manifold.) Nevertheless, these results combined with the examples given in this section appear to provide enough evidence to conjecture that the coisotropic intersection property holds in general. Next let us examine the case where the characteristic foliation is a fibration. We do this by passing to the graph of the foliation and interpreting it as a Lagrangian submanifold. This will also lead us to a refinement of Question \[quest:intersections\]. \[ex:leafwise\] Let $M$ be a coisotropic submanifold of $(W,\omega)$ and let $\Gamma\subset M\times M\subset W\times W$ be the graph of the characteristic foliation ${{\mathcal F}}$ on $M$. In other words, $\Gamma$ is formed by pairs $(x,y)\in M\times M$ with $x$ and $y$ lying on the same leaf of ${{\mathcal F}}$. Then $\Gamma$ is a one-to-one immersed Lagrangian submanifold of $W\times W$, where the latter is equipped with the symplectic form $\omega\oplus -\omega$. (In general, $\Gamma$ is not a true submanifold, e.g., $\Gamma$ can be dense in $M\times M$.) Consider a Hamiltonian diffeomorphism $\varphi$ of $W$. Then $\tilde{\varphi}=({{\mathit id}},\varphi)$ is a Hamiltonian diffeomorphism of $W\times W$ and the intersection points $\tilde{\varphi}(\Gamma)\cap\Gamma$ are in one-to-one correspondence with $x\in M$ such that $x$ and $\varphi(x)\in M$ lie on the same leaf. Hence, persistence of Lagrangian intersections for $\Gamma$, if it held, would imply the *leaf-wise intersection property* for $M$, i.e., the existence of a leaf $F$ of ${{\mathcal F}}$ with $\varphi(F)\cap F\neq \emptyset$. For instance, assume that ${{\mathcal F}}$ is a fibration. Then $\Gamma$ is a true smooth Lagrangian submanifold of $W\times W$ and we conclude that in this case leaf-wise intersections do exist, provided that $\varphi$ is not far from ${{\mathit id}}$ in a suitable sense; see, e.g., [@Ch; @F:Morse; @F:grad; @F:c-l; @Fl; @Gr; @LS:lagr; @Oh1; @Oh:disj; @We73]. Taking this example as a motivation let us call $x\in M$ a leaf-wise intersection of $M$ and $\varphi(M)$ if $x\in F\cap\varphi(F)$ for some leaf $F$ of the coisotropic foliation. Note that in the infinitesimal setting of Example \[ex:infinitesimal\], leaf-wise intersections correspond to the critical points of $H$ on $M$. \[quest:lw-intersections\] Do leaf-wise intersections exist (perhaps, under some additional conditions on the coisotropic submanifold) whenever the Hamiltonian diffeomorphism is sufficiently close to the identity, e.g., in the sense of Hofer’s metric? When $M$ is Lagrangian, this question is of course equivalent to Question \[quest:intersections\], for the characteristic foliation in this case has only one leaf, the entire manifold $M$. In the other extreme case $M=W$, where $W$ is closed, Question \[quest:lw-intersections\] is equivalent to the Arnold conjecture – the existence of fixed points of Hamiltonian diffeomorphisms – and hence the answer to the question is affirmative in this case. (See, e.g., [@F:Morse; @F:grad; @F:c-l; @Fl; @FO; @Gr; @HZ; @LT:ac; @mdsa-book; @Sa] and references therein.) Question \[quest:lw-intersections\], arising also from some problems in Hamiltonian dynamics, was originally raised by Moser in 1978 in [@Mo]. In [@Ba; @Mo], persistence of leaf-wise intersections was proved for closed coisotropic submanifolds and Hamiltonian diffeomorphisms which are $C^1$-close to the identity. Moser’s theorem was later extended by Hofer, [@Ho90], to hypersurfaces of restricted contact type in $\R^{2n}$ and Hamiltonian diffeomorphisms with energy smaller than a certain symplectic capacity of the region bounded by the hypersurface. (See also [@EH; @Li].) Recently, Dragnev, [@Dr], generalized this result to arbitrary closed coisotropic submanifolds in $\R^{2n}$ that have contact type, but need not be of restricted contact type (see Definition \[def:B\]) . In this case, the energy of the diffeomorphism is required to be smaller than the homological capacity of the submanifold itself. Note that this capacity is positive for any coisotropic submanifold which is displaceable and stable as is shown below in Remark \[rmk:Dr\]. Here we do not consider the problem of extending Dragnev’s theorem to other ambient manifolds; this question will be addressed elsewhere. Instead, we prove a simple preliminary result in this direction and generalize the results of [@EH; @Ho90] to subcritical Stein manifolds. Namely, we show that leaf-wise intersections of hypersurfaces in such manifolds persist for Hamiltonian diffeomorphisms with energy smaller than the homological capacity of the region bounded by the hypersurface. (See Theorem \[thm:leafwise\].) As stated, this result does not hold when the contact type condition is dropped; see Example \[ex:leafwise-fail\]. Finally note that Moser’s theorem and the results on Lagrangian intersections and the Arnold conjecture discussed above suggest the existence, under suitable additional assumptions, of more than one leaf-wise intersection. A theorem of Polterovich and of Laudenbach and Sikorav, [@LS; @Pol], asserts that Hamiltonian persistence of intersections is an exclusive feature of Lagrangian submanifolds among submanifolds of middle dimension. Namely, a middle-dimensional submanifold $N$ admits an infinitesimal Hamiltonian displacement if and only if $N$ is not Lagrangian and its normal bundle has a non-vanishing section. This clear-cut dichotomy does not carry over to lower codimensions, but the general picture is somewhat similar. To be more specific, it has been shown by Gürel, [@gu:new], that a totally non-coisotropic submanifold admits an infinitesimal Hamiltonian displacement, provided that its normal bundle has a non-vanishing section. Now, in contrast with the Lagrangian case, it is not sufficient to assume (under the same normal bundle condition) that $N$ is simply not coisotropic, for $N$ may contain a Lagrangian submanifold. Coisotropic Liouville class --------------------------- The Lagrangian intersection property is intimately connected to the fact that a Lagrangian submanifold that is displaceable must have non-zero Liouville class. Moreover, under suitable hypotheses, the displacement energy can be bounded from below via the size of the Liouville class. This connection can also be extended to coisotropic manifolds. Thus assume, for the sake of simplicity, that the ambient symplectic manifold $(W^{2n},\omega)$ is symplectically aspherical and exact: $\omega=d\lambda$. Then the restriction $\lambda|_{{{\mathcal F}}}$ of $\lambda$ to the characteristic foliation ${{\mathcal F}}$ of $M$ is leaf-wise closed, and the cohomology class $[\lambda|_{{{\mathcal F}}}]\in H^1_{dR}({{\mathcal F}})$ in the foliated de Rham cohomology is defined. (Recall that $H^*_{dR}({{\mathcal F}})$ is the cohomology of the complex of smooth differential forms along the leaves of ${{\mathcal F}}$; see, e.g., [@MS] and references therein for a discussion of foliated de Rham cohomology.) Note that $[\lambda|_{{{\mathcal F}}}]$ depends on the choice of $\lambda$. When $M$ is Lagrangian, this is the ordinary Liouville class of $M$. In general, we will refer to it as the *coisotropic Liouville class*. By analogy with Lagrangian manifolds, we ask \[quest:Liouville\] Is $[\lambda|_{{{\mathcal F}}}]\neq 0$, provided that $M$ is displaceable and closed? When $M$ is Lagrangian, this answer is affirmative by [@Ch:Cambr; @Ch; @Gr; @Oh:disj; @Po93]. If $M$ is a hypersurface, the answer is also affirmative. (Indeed, a displaceable hypersurface bounds a region and then $\int_M\lambda\wedge\omega^{n-1}$ is equal to the symplectic volume of the region. On the other hand, this integral would be zero if $\lambda$ were leaf-wise exact.) Furthermore, again by analogy with the Lagrangian case, one can ask whether the displacement energy of $M$ can be bounded from below via the “size” of $[\lambda|_{{{\mathcal F}}}]$. A sufficiently good lower bound would imply an affirmative answer to Question \[quest:intersections\] by a version of the figure-eight trick. \[ex:Liouville-graph\] In the setting of Example \[ex:leafwise\], assume that $\omega=d\lambda$. Then the manifold $W\times W$ is exact and the Liouville class $[\lambda_\Gamma]$ of the graph $\Gamma$ is defined, when $\Gamma$ is interpreted as an immersed manifold. Denote by $\pi_1$ and $\pi_2$ the natural projections of $\Gamma$ to the first and the second factor in $W\times W$. Then $\lambda_\Gamma=\pi_1^*\lambda-\pi_2^*\lambda$. Furthermore, it is not hard to show that $[\lambda|_{{{\mathcal F}}}]\neq 0$ if $[\lambda_\Gamma]\neq 0$. Assume now that ${{\mathcal F}}$ is a fibration and $M$ is displaceable. Then $\Gamma$ is a genuine, displaceable Lagrangian submanifold of $W\times W$. Under natural additional assumptions on $W$, we infer that $[\lambda_\Gamma]\neq 0$ (see, e.g., [@Ch:Cambr; @Ch; @Gr; @Oh:disj; @Po93]), and hence $[\lambda|_{{{\mathcal F}}}]\neq 0$. Note also that, as a consequence, the first cohomology of the fiber of ${{\mathcal F}}$ is non-trivial. We prove that, for a stable coisotropic submanifold, there exists a loop $\eta$ which is tangent to ${{\mathcal F}}$, contractible in $W$ and bounds non-zero symplectic area (Theorem \[thm:main\](ii)). This theorem requires natural minor assumptions on the behavior of $\omega$ at infinity in $W$, but holds even when $W$ is not exact. As a consequence of the theorem, we obtain an affirmative answer to Question \[quest:Liouville\] for stable submanifolds. However, this approach does not lead to an answer to this question in general, for such a loop $\eta$ need not exist when the stability condition is dropped even if the Liouville class is non-zero. (This follows from counterexamples to the Hamiltonian Seifert conjecture; see, e.g., [@Ci; @Gi95; @gi:bayarea; @gg:ex2; @ke:example] and references therein.) Furthermore, if $M$ has restricted contact type, the displacement energy of $M$ is greater than or equal to the symplectic area bounded by $\eta$. (See Definition \[def:B\] and Theorem \[thm:main\](iii).) For $W=\R^{2n}$, these results were proved by Bolle, [@Bo1; @Bo2], using the finite–dimensional reduction methods of [@CZ83]. Our proof draws heavily on Bolle’s ideas and is in fact just a translation of his argument to the Floer theoretic setting.[^2] Dense existence of non-contractible loops {#sec:dense} ----------------------------------------- The stability condition and the proof of Theorem \[thm:main\] suggest that the loop $\eta$, tangent to a leaf of the characteristic foliation but not contractible in it, can be viewed as a generalization of a closed characteristic on a hypersurface; cf.[@Bo1; @Bo2]. Then assertions (ii) and (iii) of Theorem \[thm:main\] are interpreted as generalizations of the existence of closed characteristics on stable hypersurfaces in $\R^{2n}$, established in [@HZ:cap; @HZ; @St]. (See also [@gi:alan] for further references.) Continuing this analogy, consider a map $\vec{K}\colon W\to\R^k$ whose components are proper, Poisson–commuting Hamiltonians. Then $M_a=\vec{K}^{-1}(a)$ is a coisotropic submanifold in $W$ whenever $a\in\R^k$ is a regular value of $\vec{K}$. Assume, for the sake of simplicity, that all levels $M_a$ are displaceable. Then we prove that for a subset $A\subset\R^k$ dense in the set of regular values of $\vec{K}$, a level $M_a$, with $a\in A$, carries a loop tangent to the characteristic foliation, contractible in $W$ but not in its leaf, and bounding a positive symplectic area. This result (Theorem \[thm:comm\]) can be regarded as a generalization of the dense existence theorem, [@fh; @HZ]; see also [@gi:alan]. Here, as above, we need to impose some natural additional conditions on $W$, but the coisotropic submanifolds $M_a$ need not be stable. This theorem leads to the question whether a version of the almost existence theorem, [@HZ:cap; @HZ; @St], for such loops holds for commuting Hamiltonians. Organization of the paper ------------------------- The main results of the paper and the necessary definitions are stated and further discussed in detail in Section \[sec:statements\]. The goal of Section \[sec:non-deg\] is purely technical: here we set conventions and recall relevant results concerning filtered Floer homology, homotopy maps, and action selectors. In Sections \[sec:nondeg\] and \[sec:deg\] we establish an auxiliary result on which the proofs of our main theorems hinge. This result, which may be of independent interest (cf. [@En; @KL; @LM; @MDS; @Oh:chain]), asserts the existence of a Floer connecting trajectory descending from a one-periodic orbit to the maximum of a Hamiltonian and having energy bounded from above by the displacement energy of the support of the Hamiltonian, provided that the maximum is large enough and the Hamiltonian is “slow” near its maximum. For non-degenerate Hamiltonians this result (Proposition \[prop:orbits-nondeg\]) is proved in Section \[sec:nondeg\]. In Section \[sec:deg\] we deal with the degenerate case (Proposition \[prop:orbits-deg\]) and also discuss the space of finite energy Floer trajectories. Section \[sec:prfs\] is devoted to the proofs of the main results of the paper. Here we establish Theorem \[thm:main\] giving affirmative answers to Questions \[quest:intersections\] and \[quest:Liouville\] for stable, coisotropic submanifolds. In this section we also prove a version of the dense existence theorem for commuting Hamiltonians (Theorem \[thm:comm\]). Finally, in Section \[sec:leafwise\] we prove persistence of leaf-wise intersections for hypersurfaces of restricted contact type in subcritical Stein manifolds (Theorem \[thm:leafwise\]). Acknowledgments. {#acknowledgments. .unnumbered} ---------------- The author is deeply grateful to Yasha Eliashberg, Başak Gürel, Ely Kerman, Felix Schlenk, and Claude Viterbo for their numerous valuable remarks and suggestions. Displacement of stable coisotropic submanifolds {#sec:statements} =============================================== Contact type and stable coisotropic submanifolds {#sec:ct} ------------------------------------------------ Let $(W^{2n}, \omega)$ be a symplectic manifold and let $M\subset W$ be a closed coisotropic submanifold of codimension $k$. Set $\omega_0=\omega|_M$. Then, as is well known, the distribution $\ker \omega_0$ has dimension $k$ and is integrable. Denote by ${{\mathcal F}}$ the characteristic foliation on $M$, i.e., the $k$-dimensional foliation whose leaves are tangent to the distribution $\ker \omega_0$. \[def:B\] The coisotropic submanifold $M$ is said to be *stable* if there exist one-forms $\alpha_1,\ldots,\alpha_k$ on $M$ such that $\ker d\alpha_i\supset \ker\omega_0$ for all $i=1,\ldots,k$ and $${\label{eq:ct}} \alpha_1\wedge\cdots\wedge\alpha_k\wedge \omega_0^{n-k}\neq 0$$ anywhere on $M$. We say that $M$ has *contact type* if the forms $\alpha_i$ can be taken to be primitives of $\omega_0$. Furthermore, $M$ has *restricted* contact type if the forms $\alpha_i$ extend to global primitives of $\omega$ on $W$. Stable and contact type coisotropic submanifolds were introduced by Bolle in [@Bo1; @Bo2]. The nature of the requirements of Definition \[def:B\], which are very restrictive, is illustrated by the following examples. \[exam:Bolle\]   1. A contact type coisotropic submanifold is automatically stable. Furthermore, a coisotropic submanifold which is $C^1$-close to a coisotropic submanifold of contact type also has contact type. (Apparently the latter is not true for stable coisotropic submanifolds.) 2. A hypersurface has contact type as a coisotropic submanifold if and only if it has contact type in the standard sense. A hypersurface is stable as a coisotropic submanifold if and only if it is stable as a hypersurface, i.e., there exists a vector field $Z$ transverse to $M$ and such that $\ker\left(\varphi_t^*\omega |_{M}\right)=\ker\left(\omega |_{M}\right)$ for small $|t|$, where $\varphi_t$ is the flow of $Z$; cf.  [@HZ p. 122] and [@EKP]. We will generalize this observation to coisotropic submanifolds of codimension $k\geq 1$ in Proposition \[prop:stable\]. 3. The product of stable submanifolds is also stable. More precisely, let $M_1\subset W_1$ and $M_2\subset W_2$ be stable. Then $M_1\times M_2\subset W_1\times W_2$ is stable. For instance, the product of contact type hypersurfaces is a stable coisotropic submanifold. The product $M\times S^1\subset W\times T^*S^1$ has (restricted) contact type, provided that $M$ has (restricted) contact type. However, unless $M_1$ or $M_2$ is one-dimensional, the product $M_1\times M_2$ need not have contact type even if $M_1$ and $M_2$ have (restricted) contact type. This follows from Remark \[rmk:cohomology\] below. 4. A stable Lagrangian submanifold is necessarily a torus as can be seen from Proposition \[prop:flow\] or from Remark \[rmk:cohomology\]. \[ex:comm\] Let $M$ be a regular level set of the map $(K_1,\ldots,K_{k})\colon W\to\R^{k}$ whose components are proper, Poisson–commuting Hamiltonians. Assume furthermore that the Hamiltonian flows of $K_1,\ldots,K_{k}$ generate an action of a torus $\T^k$ on $M$. Then $M$ is stable. (To see this, define $\alpha_i$ on $\T^k$-orbits, i.e., the leaves of ${{\mathcal F}}$, by $\alpha_i(X_{K_j})=\delta_{ij}$, where $X_{K_j}$ is the Hamiltonian vector field of $K_j$, and extend these forms to $\T^k$-invariant one-forms on $M$ in an arbitrary way. Then $T{{\mathcal F}}\subset \ker d\alpha_i$ and clearly is also satisfied.) Note that here, similarly to the assertion of the Arnold–Liouville theorem, we require the Hamiltonians to generate a torus action only on $M$, but not on the entire ambient space $W$. \[rmk:cohomology\] The requirements that $M$ is stable or has contact type impose severe restrictions on the topology of $M$ and ${{\mathcal F}}$. Namely, assume first that $M$ is stable. Let $V$ be the vector space formed by linear combinations $\alpha=a_1\alpha_1+\cdots + a_k\alpha_k$, where $a_i$ are constants. The forms $\alpha$ are closed along ${{\mathcal F}}$ and the natural map from $V$ to the foliated de Rham cohomology $H^1_{dR}({{\mathcal F}})$ along ${{\mathcal F}}$, sending $\alpha$ to its cohomology class, is a monomorphism. (Here, as above, $M$ is closed.) In particular, $\dim H^1_{dR}({{\mathcal F}})\geq k$. Indeed, $\alpha|_{{{\mathcal F}}}=df|_{{{\mathcal F}}}$ would imply that $\alpha_x|_{T_x{{\mathcal F}}}=0$ at a critical point $x$ of $f$, which in turn means that $a_1=\ldots=a_k=0$ since the forms $\alpha_1,\ldots,\alpha_k$ are, by , linearly independent in $T_x^*{{\mathcal F}}$, cf. [@Bo2 Remark 3]. When $M$ has contact type, consider the vector space $V_0\subset V$ formed by $\alpha=a_1\alpha_1+\cdots + a_k\alpha_k$ with $a_1+\cdots+a_k=0$. The forms $\alpha$ are closed on $M$ and the natural map $V_0\to H^1(M;\R)$ is a monomorphism. (As a consequence, $\dim H^1(M;\R)\geq k-1$.) The proof of this observation due to Bolle, [@Bo2], is similar to the argument for stable manifolds above. \[prop:flow\]. Let, as above, $M$ be a closed, stable, coisotropic submanifold. Consider the product $M\times\R^k$. Let $(p_1,\ldots,p_k)$ be coordinates on $\R^k$ and let us use the same symbols $\omega_0$ and $\alpha_i$ for differential forms on $M$ and for their pull-backs to $M\times\R^k$. Then the form $$\label{eq:normal-form} \omega=\omega_0+\sum_{i=1}^k d(p_i\alpha_i),$$ is symplectic near $M=M\times\{0\}$ in $M\times\R^k$. The normal bundle to $M$ in $W$ is trivial, for it is isomorphic to $T^*{{\mathcal F}}$, and thus can be identified with $M\times\R^k$. Then, as is immediately clear (see [@Bo1; @Bo2]) from the Weinstein symplectic neighborhood theorem (see, e.g., [@We77]), the local normal form of $\omega$ near $M$ is given by . From now on, we identify a neighborhood of $M$ in $W$ with a neighborhood of $M$ in $T^*{{\mathcal F}}=M\times\R^k$ equipped with the symplectic form . \[prop:flow\] Let $M$ be a stable coisotropic submanifold. 1. The leaf-wise metric $(\alpha_1)^2+\cdots+(\alpha_k)^2$ on ${{\mathcal F}}$ is leaf-wise flat. 2. The Hamiltonian flow of $\rho=(p_1^2+\cdots+p_k^2)/2$ is the leaf-wise geodesic flow of this metric. This proposition is essentially proved in [@Bo2]. Here we just briefly outline the argument for the sake of completeness. To prove the first assertion, note that by the forms $\alpha_i$ are linearly independent and leaf-wise closed. Thus, locally on every leaf, their primitives form a coordinate system in which the metric is isometric to the standard metric on $\R^k$. Let $Y_1,\ldots,Y_k$ be the coordinate vector fields for this coordinate system, i.e., $\alpha_i(Y_j)=\delta_{ij}$. To prove the second assertion, we now simply observe from that the Hamiltonian vector field $X_\rho$ of $\rho$ is given by $$\label{eq:spray} X_\rho=\sum_{i=1}^k p_i Y_i,$$ which is the geodesic spray of the metric. The leaf-wise metric $\rho$ has some other relevant properties that do not hold for leaf-wise flat metrics in general. For instance, the length spectrum of $\rho$ is nowhere dense; see Lemma \[lemma:nowheredense\]. The next proposition, also quite elementary, clarifies the nature of the stability condition and generalizes Example \[exam:Bolle\](ii). \[prop:stable\] A coisotropic submanifold $M$ is stable if and only if there exists a tubular neighborhood $M\times U$ of $M=M\times\{0\}$ in $W$, where $U\subset\R^k$ is a neighborhood of the origin, such that the submanifolds $M_p=M\times\{p\}$ are coisotropic for $p\in U$ and $\ker\omega_p=\ker\omega_0$, where $\omega_p=\omega|_{M_p}$. The fact that stability implies the existence of such a neighborhood is a consequence of the normal form . Conversely, set $\alpha_i = \left(i_{\p/\p p_i}\omega\right)|_M$. Then, follows immediately from the fact that $\p/\p p_1,\ldots,\p/\p p_k$ are linearly independent. Furthermore, $$d\alpha_i=\left.\left(L_{\frac{\p}{\p p_i}}\omega\right)\right|_M =\left.\frac{d}{d t}\omega_{p(t)}\right|_{t=0},$$ where $p(t)=(0,\ldots, 0, t, 0,\ldots, 0)$ with $t$ in the $i$th slot, and hence $\ker d\alpha_i\supset \ker \omega_0$ as required. Hamiltonian displacement of coisotropic submanifolds {#sec:statement} ---------------------------------------------------- To state our results on Hamiltonian displacement of stable coisotropic submanifolds, we need to impose some natural conditions on the ambient symplectic manifold $(W,\omega)$. Namely, in what follows $W$ is always assumed to be *symplectically aspherical*, i.e., $\omega|_{\pi_2(W)}=0=c_1|_{\pi_2(W)}$. Furthermore, we require $W$ to be closed or *geometrically bounded* and *wide*. The condition that $W$ is geometrically bounded means that $W$ admits a complete metric which is compatible with $\omega$ in a rather weak sense and has injectivity radius bounded away from zero and sectional curvature bounded from above; see [@AL] for the precise definition. A symplectic manifold is said to be wide if it admits a proper Hamiltonian bounded from below whose Hamiltonian flow has no non-trivial contractible periodic orbits of period less than or equal to one; see [@gu:new]. Among wide manifolds are manifolds convex at infinity (e.g., cotangent bundles to closed manifolds and $\R^{2n}$), twisted cotangent bundles, and non-compact covering spaces of closed manifolds. In fact, the author is not aware of any example of a geometrically bounded, open manifold that is not wide. The essence of these requirements is that the standard machinery of Floer homology is applicable to symplectically aspherical, geometrically bounded manifolds (see, e.g., [@cgk; @gg:new] and Section \[sec:non-deg\] below). Furthermore, one of the main tools utilized in this paper is the technique of action selectors. This technique, developed for closed and convex at infinity symplectically aspherical manifolds in [@FS; @schwarz], has recently been extended to geometrically bounded, wide manifolds by Gürel, [@gu:new]. Recall also that the energy of a compactly supported, time-dependent Hamiltonian $H\colon [0,\,1]\times W\to \R$ is defined as $$\parallel H \parallel =\int_0^1 (\max H_t-\min H_t)\, dt,$$ where $H_t=H(t,\cdot)$. The Hamiltonian diffeomorphism $\varphi_H$, i.e., the time-one Hamiltonian flow of $H$, is said to displace $M$ if $\varphi_H(M)\cap M=\emptyset$. When such a map $\varphi_H$ exists, we call $M$ displaceable. For instance, every compact subset of $\R^{2n}$ is displaceable. Now we are in a position to state the main result of the paper. \[thm:main\] Let $W$ be symplectically aspherical, and closed or wide and geometrically bounded. Let $M$ be a closed, stable, coisotropic submanifold of $W$. 1. Then there exists a constant $\Delta>0$ such that $\parallel H \parallel > \Delta$ for any compactly supported Hamiltonian $H$ with $\varphi_H$ displacing $M$. 2. When $M$ is displaceable, there exists a loop $\eta$ tangent to ${{\mathcal F}}$, contractible in $W$, and bounding a positive symplectic area $A(\eta)$. 3. Moreover, if $M$ is displaceable and has restricted contact type, there exists a loop $\eta$ tangent to ${{\mathcal F}}$, contractible in $W$, and such that $0<A(\eta)\leq \|H\|$. Note that the loop $\eta$ from assertions (ii) and (iii) is necessarily not contractible in the leaf containing it. Moreover, $\eta$ is not contractible in the class of loops tangent to ${{\mathcal F}}$. This follows immediately from the observation that the area spectrum of $M$ has zero measure. (By definition, the area spectrum of $M$ is the set formed by symplectic areas bounded by loops in $M$ that are tangent to ${{\mathcal F}}$ and contractible in $W$.) The same holds for the curve $\gamma$ from Theorem \[thm:comm\] below. Theorem \[thm:main\] will be proved in Section \[sec:prfs\]. For stable coisotropic submanifolds in $\R^{2n}$ assertions (ii) and (iii) were established by Bolle, [@Bo1; @Bo2]. When $M$ is Lagrangian, and hence necessarily a torus, the first assertion is a particular case of the Lagrangian intersection property discussed in Section \[sec:intro\]; see, e.g., [@Ch:Cambr; @Ch; @F:Morse; @F:grad; @F:c-l; @Fl; @Gr; @LS:lagr; @Oh1; @Oh:cambridge; @Oh:disj; @We73; @We74] for similar and more general results. When $M$ is a hypersurface, assertion (i) is trivial. However, in this case, we prove a sharper theorem concerning leaf-wise intersections and complementing the results of [@Ba; @Dr; @EH; @Ho90; @Li; @Mo], cf. Example \[ex:leafwise\] and Question \[quest:lw-intersections\]. \[thm:leafwise\] Let $M$ be a connected closed hypersurface of restricted contact type, bounding a domain $U$ in a subcritical Stein manifold $W$. Then for a compactly supported Hamiltonian diffeomorphism $\varphi_H\colon W\to W$ with $\|H\|<\chom(U)$, there exists a leaf $F$ of the characteristic foliation on $M$ such that $\varphi_H(F)\cap F\neq\emptyset$. We refer the reader to Section \[sec:homcap\] for the definition of the homological capacity $\chom$. Theorem \[thm:leafwise\] will be established in Section \[sec:leafwise\]. For $W=\R^{2n}$, this result was proved in [@Ho90]. As stated, with the upper bound on $\|H\|$, Theorem \[thm:leafwise\] does not hold for hypersurfaces that do not have contact type. In Example \[ex:leafwise-fail\], we construct a Hamiltonian flow $\varphi^t$ on $\R^{2n}$ and a sequence of coisotropic submanifolds $M_i$, $C^0$-converging to $S^{2n-1}$, such that $M_i$ and $\varphi^{t_i}(M_i)$ have no leaf-wise intersections for some sequence of times $t_i\to 0+$. As is pointed out in Section \[sec:intro\], assertion (ii) of Theorem \[thm:main\] fails when the requirement that $M$ is stable is dropped. Indeed, for hypersurfaces in $\R^{2n}$, (ii) implies the existence of a closed characteristic, while counterexamples to the Hamiltonian Seifert conjecture show that in general closed characteristics need not exist; see, e.g., [@Ci; @Gi95; @gi:bayarea; @gg:ex2; @ke:example] and references therein. Recall also that the condition that $M$ is displaceable is essential in (ii): the Liouville class of the zero section of a cotangent bundle is zero. Assertion (ii) gives an affirmative answer to Question \[quest:Liouville\] for stable coisotropic manifolds. Namely, assume that $\omega$ is exact, i.e., $\omega=d\lambda$. Then the restriction $\lambda|_{M}$ is closed along ${{\mathcal F}}$, and hence the foliated Liouville class $[\lambda|_{{{\mathcal F}}}]\in H^1_{dR}({{\mathcal F}})$ is defined. \[cor:liouville\] Assume in addition to the hypotheses of Theorem \[thm:main\] that $\omega$ is exact and $M$ is displaceable. Then $[\lambda|_{{{\mathcal F}}}]\neq 0$. Commuting Hamiltonians ---------------------- Let, as in Section \[sec:dense\] and Example \[ex:comm\], $M$ be a regular level $\vec{K}^{-1}(0)$ of the map $\vec{K}=(K_1,\ldots,K_{k})\colon W\to\R^{k}$ whose components are Poisson–commuting Hamiltonians. (Now, in contrast with Example \[ex:comm\], we do not assume that $\vec{K}$ gives rise to a torus action on $M$.) As above, $W$ is required to be symplectically aspherical, and closed or wide and geometrically bounded. In the latter case, we also require the map $\vec{K}$ to be proper (on its image) near $M$ to insure that the coisotropic manifolds $M_a= \vec{K}^{-1}(a)$ are compact and close to $M$ when $a\in\R^k$ is near the origin. Then, as in Section \[sec:dense\], consider loops on $M_a$ which are tangent to the characteristic foliation, contractible in $W$, but not contractible in the leaf of the foliation. Such a loop can be thought of as an analogue of a closed characteristic on a hypersurface; see Section \[sec:dense\]. Hence, the existence of such a loop can be interpreted as a generalization of the dense existence theorem of Hofer and Zehnder and of Struwe, [@HZ:V; @HZ; @St], to the moment map $\vec{K}$. In Section \[sec:prfs\] we prove \[thm:comm\] Assume that $M$ is displaceable. Then, for a dense set of regular values $a\in\R^k$ near the origin, the level set $M_a$ carries a closed curve $\gamma$, which is contractible in $W$, tangent to the characteristic foliation ${{\mathcal F}}_a$ on $M_a$, and bounds positive symplectic area. Filtered Floer homology {#sec:non-deg} ======================= In this section we recall the definition of filtered Floer homology for geometrically bounded symplectically aspherical manifolds, set conventions and notation used in this definition, and revisit the construction of action selectors for the manifolds in question. Most of the results mentioned here are either well known or established elsewhere or can be proved by adapting standard arguments. For this reason, the proofs are omitted or just very briefly outlined; however, in each case detailed references are provided although not necessarily to the original proofs. We refer the reader to Floer’s papers [@F:Morse; @F:grad; @F:c-l; @F:witten; @Fl], to [@BPS; @cgk; @fh; @FHS; @Oh:cambridge; @SZ], or to [@HZ; @mdsa-book; @Sa] for introductory accounts of the construction of Floer homology in this setting. Preliminaries: notation and conventions --------------------------------------- Let $(W^{2n},\omega)$ be a symplectically aspherical manifold. Denote by $\Lambda W$ the space of smooth contractible loops $\gamma\colon S^1\to W$ and consider a time-dependent Hamiltonian $H\colon S^1\times W\to \R$, where $S^1=\R/\Z$. Setting $H_t = H(t,\cdot)$ for $t\in S^1$, we define the action functional $A_H\colon \Lambda W\to \R$ by $$A_H(\gamma)=A(\gamma)+\int_{S^1} H_t(\gamma(t))\,dt,$$ where $A(\gamma)$ is the negative symplectic area bounded by $\gamma$, i.e., $$A(\gamma)=-\int_z\omega,$$ where $z\colon D^2\to W$ is such that $z|_{S^1}=\gamma$. The least action principle asserts that the critical points of $A_H$ are exactly contractible one-periodic orbits of the time-dependent Hamiltonian flow $\varphi_H^t$ of $H$, where the Hamiltonian vector field $X_H$ of $H$ is defined by $i_{X_H}\omega=-dH$. We denote the collection of such orbits by ${{\mathcal P}}_H$ and let ${{\mathcal P}}^{(a,\,b)}_H\subset {{\mathcal P}}_H$ stand for the collection of orbits with action in the interval $(a,\,b)$. The action spectrum ${{\mathcal S}}(H)$ of $H$ is the set of critical values of $A_H$. In other words, ${{\mathcal S}}(H)=\{ A_H(\gamma)\mid \gamma\in {{\mathcal P}}_H\}$. This is a zero measure set; see, e.g., [@HZ; @schwarz]. In what follows we will always assume that $H$ is compactly supported and set ${\operatorname{supp}}H=\bigcup_{t\in S^1}{\operatorname{supp}}H_t$. In this case, ${{\mathcal S}}(H)$ is compact and hence nowhere dense. Let $J=J_t$ be a time-dependent almost complex structure on $W$. A Floer anti-gradient trajectory $u$ is a map $u\colon \R\times S^1\to W$ satisfying the equation $$\label{eq:floer} \frac{\p u}{\p s}+ J_t(u) \frac{\p u}{\p t}=-\nabla H_t(u).$$ Here the gradient is taken with respect to the time-dependent Riemannian metric $\omega(\cdot,J_t\cdot)$. This metric gives rise to an ($L^2$-) Riemannian metric on $\Lambda W$ and can formally be interpreted as the equation ${\p u}/{\p s}=-\nabla_{L^2} A_H(u(s, \cdot))$. In other words, $u$ is a trajectory of the $L^2$-anti-gradient flow of $A_H$ on $\Lambda W$. In what follows, we denote the curve $u(s,\cdot)\in \Lambda W$ by $u(s)$. The energy of $u$ is defined as $$\label{eq:energy} E(u)=\int_{-\infty}^\infty \left\|\frac{\p u}{\p s}\right\|_{L^2(S^1)}^2\,ds =\int_{-\infty}^\infty \int_{S^1}\left\|\frac{\p u}{\p t}-J\nabla H (u) \right\|^2 \,dt\,ds.$$ We say that $u$ is asymptotic to $x^\pm\in {{\mathcal P}}_H$ as $s\to\pm \infty$ or connecting $x^-$ and $x^+$ if $\lim_{s\to\pm\infty} u(s)=x^\pm$ in $\Lambda W$. More generally, $u$ is said to be partially asymptotic to $x^\pm\in {{\mathcal P}}_H$ at $\pm\infty$ if $u(s^{\pm}_k)\to x^\pm$ for some sequences $s^{\pm}_k\to\pm\infty$. In this case $$A_H(x^-)-A_H(x^+)=E(u).$$ We denote the space of Floer trajectories connecting $x^-$ and $x^+$, with the topology of uniform $C^\infty$-convergence on compact sets, by ${{\mathcal M}}_H(x^-,x^+,J)$ or simply by ${{\mathcal M}}_H(x^-,x^+)$ when the role of $J$ is not essential, even though ${{\mathcal M}}_H(x^-,x^+,J)$ depends on $J$. This space carries a natural $\R$-action $(\tau\cdot u)(t,s)=u(t,s+\tau)$ and we denote by $\hat{{{\mathcal M}}}_H(x^-,x^+,J)$ the quotient ${{\mathcal M}}_H(x^-,x^+,J)/\R$. Recall that $\gamma\in {{\mathcal P}}_H$ is said to be non-degenerate if $d\varphi_H\colon T_{\gamma(0)}W\to T_{\gamma(0)}W$ does not have one as an eigenvalue. In this case, the so-called Conley–Zehnder index $\MUCZ(\gamma)\in\Z$ is defined; see, e.g., [@Sa; @SZ]. Here we normalize $\MUCZ$ so that $\MUCZ(\gamma)=n$ when $\gamma$ is a non-degenerate maximum of an autonomous Hamiltonian with a small Hessian. Assume that all periodic orbits with actions in the interval $[A_H(x^+),A_H(x^-)]$, including $x^\pm$, are non-degenerate. Then, for a generic $J$, suitable transversality conditions are satisfied and ${{\mathcal M}}_H(x^-,x^+,J)$ is a smooth manifold of dimension $\MUCZ(x^+)-\MUCZ(x^-)$; see, e.g., [@fh; @SZ] and references therein. Filtered Floer homology and homotopy ------------------------------------ The objective of this section is two-fold. In its first part we briefly outline the construction of filtered Floer homology following closely [@fh]; see also [@BPS; @cgk; @schwarz]. (Note that in the case of open geometrically bounded manifolds the necessary compactness property of the moduli spaces of connecting trajectories is guaranteed by Sikorav’s version of the Gromov compactness theorem; see [@AL].) In the second part, we discuss properties of monotonicity maps. Here, we depart slightly from the setting of [@BPS; @cgk; @fh], for we need to consider also non-monotone homotopies and to account for possible non-compactness of $W$. For this reason, some of the proofs, still quite standard, are outlined below. Throughout the discussion of the filtered Floer homology $\HF^{(a,\,b)}(W)$, we assume, when $W$ is open, that all the intervals are in the positive range of actions, i.e., $a>0$ for any interval $(a,\,b)$. This condition can be relaxed in some instances and replaced by the requirement that $(a,\,b)$ does not contain zero. (The latter is clearly necessary with the definitions we adopt here, for $H$ is assumed to be compactly supported and thus $H$ always has trivial degenerate periodic orbits if $W$ is open.) ### Filtered Floer homology: definitions Let $H$ be a compactly supported Hamiltonian on $W$. Assume that all contractible one-periodic orbits of $H$ are non-degenerate if $W$ is closed or that all such orbits with positive action are non-degenerate when $W$ is open. This is a generic condition. Consider an interval $(a,\,b)$, with $a>0$ when $W$ is open, such that $a$ and $b$ are outside ${{\mathcal S}}(H)$. Then the collection ${{\mathcal P}}^{(a,\,b)}_H$ is finite. Assume furthermore that $J$ is regular, i.e., the necessary transversality conditions are satisfied for moduli spaces of Floer trajectories connecting orbits from ${{\mathcal P}}^{(a,\,b)}_H$. This is again a generic property as can be readily seen by applying the argument from [@fh; @FHS; @SZ]. Let $\CF_k^{(a,\,b)}(H)$ be the vector space over $\Z_2$ generated by $x\in {{\mathcal P}}^{(a,\,b)}_H$ with $\MUCZ(x)=k$. Define $$\p \colon \CF_k^{(a,\,b)}(H)\to \CF_{k-1}^{(a,\,b)}(H)$$ by $$\p x=\sum_y \#\big(\hat{{{\mathcal M}}}_H(x,y,J)\big)\cdot y.$$ Here the summation extends over all $y\in {{\mathcal P}}^{(a,\,b)}_H$ with $\MUCZ(y)=\MUCZ(x)-1$ and $\#\big(\hat{{{\mathcal M}}}_H(x,y,J)\big)$ is the number of points, modulo 2, in $\hat{{{\mathcal M}}}_H(x,y,J)$. (Recall that in this case $\hat{{{\mathcal M}}}_H(x,y,J)$ is a finite set by the compactness theorem.) Then, as is well known, $\p^2=0$. The resulting complex $\CF^{(a,\,b)}(H)$ is the filtered Floer complex for $(a,\,b)$. Its homology $\HF^{(a,\,b)}(H)$ is called the filtered Floer homology. This is essentially the standard definition of Floer homology with critical points outside $(a,\,b)$ being ignored. Then $\HF(H):=\HF^{(-\infty,\infty)}(H)$ is the ordinary Floer homology when $W$ is compact. (As is well-known, $\HF_*(H)=H_{*+n}(W;\Z_2)$.) In general, $\HF^{(a,\,b)}(H)$ depends on the Hamiltonian $H$, but not on $J$; see Section \[sec:homotopy\]. Let $a<b<c$. Assume that all of the above assumptions are satisfied for all three intervals $(a,\,c)$ and $(a,\,b)$ and $(b,\,c)$. Then clearly $\CF^{(a,\,b)}(H)$ is a subcomplex of $\CF^{(a,\,c)}(H)$, and $\CF^{(b,\,c)}(H)$ is naturally isomorphic to the quotient complex $\CF^{(a,\,c)}(H)/\CF^{(a,\,b)}(H)$. As a result, we have the long exact sequence $$\label{eq:seq} \ldots\to \HF^{(a,\,b)}(H)\to\HF^{(a,\,c)}(H)\to \HF^{(b,\,c)}(H)\to\ldots .$$ We will refer to the first map in this sequence as the inclusion map and to the second one as the quotient map and to the whole sequence as the $(a,\,b,\,c)$ exact sequence for $H$. In the construction of the action selector for open manifolds given in Section \[sec:sel-open\], it will be convenient to work with filtered Floer homology for the interval $(0,\,b)$ even though $0$ is necessarily a critical value of the action functional. This homology is defined as $$\label{eq:zero-infty} \HF^{(0,\,b)}(H)=\varprojlim_{\eps\to 0+}\HF^{(\eps,\,b)}(H),$$ where the inverse limit is taken with respect to the quotient maps and $\eps\to 0+$ in the complement of ${{\mathcal S}}(H)$. It is clear that this definition is equivalent to the original one when $W$ is closed and $0$ is not in ${{\mathcal S}}(H)$. ### Homotopy {#sec:homotopy} Let us now examine the dependence of $\HF^{(a,\,b)}(H)$ on $H$. Consider a homotopy $H^s$ of Hamiltonians from $H^0$ to $H^1$. By definition, this is a family of Hamiltonians parametrized by $s\in \R$, and such that $H^s\equiv H^0$ when $s$ is large negative and $H^s\equiv H^1$ when $s$ is large positive. Furthermore, let $J^s$ be a family of $t$-dependent almost complex structures such that again $J^s\equiv J^0$ when $s\ll 0$ and $J^s\equiv J^1$ when $s\gg 0$. (In what follows, both the family of Hamiltonians $H^s$ and the pair of families $(H^s,J^s)$ will be referred to as a homotopy, depending on the context, when no confusion can arise.) For $x\in {{\mathcal P}}^{(a_0,\,b_0)}_{H^0}$ and $y\in {{\mathcal P}}^{(a_1,\,b_1)}_{H^1}$ denote by ${{\mathcal M}}_{H^s}(x,y,J^s)$ the space of solutions of with $H=H^s$ and $J=J^s$. Next we need to address the regularity issue. When $W$ is closed, regularity of a homotopy $(H^s,J^s)$ is understood in the standard sense, i.e., that the standard transversality requirements are met by the homotopy $(H^s,J^s)$; see [@fh; @FHS; @SZ]. (This is a generic property, [@fh; @FHS; @SZ].) If $W$ is open, these conditions are never satisfied, for Hamiltonians are compactly supported. In this case, we say that $(H^s,J^s)$ is regular as long as the transversality requirements are met along all homotopy trajectories connecting periodic orbits with positive action. (This is also a generic property; the argument of [@fh; @FHS; @SZ] applies to this case.) When the transversality conditions are satisfied, ${{\mathcal M}}_{H^s}(x,y,J^s)$ is a smooth manifold of dimension $\MUCZ(x)- \MUCZ(y)$. Moreover, ${{\mathcal M}}_{H^s}(x,y,J^s)$ is a finite set when $\MUCZ(x)=\MUCZ(y)$. Define the homotopy map $$\Psi_{H^0H^1}\colon \CF^{(a_0,\,b_0)}(H^0)\to \CF^{(a_1,\,b_1)}(H^1)$$ by $$\Psi_{H^0H^1}( x)=\sum_y \#\big({{\mathcal M}}_{H^s}(x,y,J^s)\big)\cdot y.$$ Here the summation is over all orbits $y\in {{\mathcal P}}^{(a_1,\,b_1)}_{H^1}$ with $\MUCZ(y)=\MUCZ(x)$ and $\#\big({{\mathcal M}}_{H^s}(x,y,J^s)\big)$ is the number of points, modulo 2, in this moduli space. The map $\Psi_{H^0H^1}$ depends on the entire homotopy $(H^s,J^s)$ and in general is *not* a map of complexes. However, $\Psi_{H^0H^1}$ becomes a homomorphism of complexes when $(a_0,\,b_0) = (a_1,\,b_1)$ and the homotopy is monotone decreasing, i.e., $\p_s H^s\leq 0$ point-wise. Moreover, the induced map in homology is then independent of the homotopy – as long as the homotopies are decreasing – and commutes with the maps from the exact sequence . (The reader is referred to, e.g., [@BPS; @cgk; @fh; @Sa; @SZ; @schwarz; @vi:functors], for the proofs of these facts for both open and closed manifolds.) There are other instances when the same is true. This is the case, for instance, when the location of the intervals $(a_0,\,b_0)$ and $(a_1,\,b_1)$ is compatible with the growth of the Hamiltonians in the homotopy. We now analyze this particular case in more detail, for it is essential for the proof of Proposition \[prop:orbits-nondeg\] below. \[def:c-bound\] A homotopy $H^s$ is said to be *$C$-bounded*, $C\in\R$, if $$\label{eq:c-bound} \int_{-\infty}^{\infty}\int_{S^1}\max_{W} \p_s H^s_t \,dt\,ds \leq C.$$ It is clear that a $C$-bounded homotopy is also $C'$-bounded for any $C'\geq C$. In what follows we will always assume that $C\geq 0$. \[ex:c-bound\] Let us give some examples of $C$-bounded homotopies. 1. A monotone decreasing homotopy is $0$-bounded. 2. Every homotopy is $C$-bounded if $C$ is large enough. 3. Let $H^s$ be a *linear homotopy* from $H^0$ to $H^1$, i.e., $$H^s=(1-f(s))H^0+f(s)H^1,$$ where $f\colon \R\to [0,1]$ is a monotone increasing compactly supported function equal to zero near $-\infty$ and equal to one near $\infty$. Then $H^s$ is $C$-bounded for any $C\geq \int_{S^1}\max_W(H^1-H^0)\,dt$. The following three observations (H0)–(H2) show that the standard properties of homotopy maps in ordinary Floer homology extend to the maps induced by $C$-bounded homotopies in filtered Floer homology. The proof of (H0)–(H2) will further clarify the essence of Definition \[def:c-bound\]. Note also that in (H0)–(H2) we will assume that the end points of the intervals in question are outside the action spectra of the Hamiltonians and, as above, $a>0$ when $W$ is open. Furthermore, we will require all Hamiltonians to be non-degenerate (in the positive action range for open manifolds) and the homotopies to be regular. However, this latter requirement is not essential, as we will show in Section \[sec:inv\], and is not met in Examples \[ex:incl-quot\] and \[exam:incr\]. 1. Let $H^s$ be a $C$-bounded homotopy. Then $$\Psi_{H^0H^1}\colon \CF^{(a,\,b)}(H^0)\to \CF^{(a+C,\,b+C)}(H^1)$$ is a homomorphism of complexes for any interval $(a,\,b)$. Hence, $\Psi_{H^0H^1}$ induces a map in Floer homology, which we, abusing notation, also denote by $\Psi_{H^0H^1}$. This map sends the $(a,\,b,\,c)$ exact sequence for $H^0$ to the $(a+C,\,b+C,\,c+C)$ exact sequence for $H^1$, i.e., on the level of homology $\Psi_{H^0H^1}$ commutes with all maps in the long exact sequence . Here, and in (H1) and (H2) below, we suppress the dependence of this map on the homotopy $J^s$ of complex structures. In fact, as we will see in Section \[sec:inv\], the induced map on the level of homology is independent of $J^s$. \[ex:incl-quot\] For any $C\geq 0$ the identity homotopy from $H$ to $H$ induces a map which is the composition $$\HF^{(a,\,b)}(H)\to \HF^{(a,\,b+C)}(H)\to \HF^{(a+C,\,b+C)}(H)$$ of the inclusion and quotient maps from the exact sequences for the intervals $a<b<b+C$ and $a<a+C<b+C$, respectively. We will refer to this map as the *inclusion-quotient* map. In view of Example \[ex:c-bound\](ii), note also that $\CF^{(a+C,\,b+C)}(H)=0$ whenever $a+C>\max {{\mathcal S}}(H)$, and hence the assertion (H0) becomes trivial if, for a fixed homotopy, $C$ is taken to be large enough. For instance, the inclusion-quotient map is automatically zero if $a+C>b$. 1. For a homotopy of $C$-bounded homotopies $H^{s,\lambda}$, i.e., a family of $C$-bounded homotopies parametrized by $\lambda\in [0,1]$, the induced map $$\Psi_{H^0H^1}\colon \HF^{(a,\,b)}(H^0)\to \HF^{(a+C,\,b+C)}(H^1)$$ is independent of the choice of homotopy in the family $H^{s,\lambda}$. Consider now a $C$-bounded homotopy $H^s$ from $H^0$ to $H^1$ and a $C'$-bounded homotopy $G^s$ from $G^0=H^1$ to $G^1$. Define the gluing or composition of these homotopies $H^s\#_R G^s$, where $R>0$ is large enough, as $H^s\#_R G^s=H^{s+R}$ if $s\leq 0$ and $H^s\#_R G^s=G^{s-R}$ if $s\geq 0$. (In other words, in $H^s\#_R G^s$ we first, for $s\leq 0$, perform the homotopy obtained from $H^s$ by shifting it sufficiently far to the left in $s$ and then, for $s\geq 0$, the homotopy resulting from $G^s$ by shifting it to the right.) It is clear that $H^s\#_R G^s$ is $(C+C')$-bounded. 1. When $R>0$ is large enough, the map $$\Psi_{H^0G^1}\colon \HF^{(a,\,b)}(H^0)\to \HF^{(a+C+C',\,b+C+C')}(G^1)$$ induced by the homotopy $H^s\#_R G^s$ is equal to the composition $\Psi_{G^0G^1}\circ\Psi_{H^0H^1}$. \[exam:incr\] Let $H^s$ be an increasing linear homotopy from $H^0$ to $H^1\geq H^0$ and let $G^s$ be an arbitrary decreasing homotopy from $G^0=H^1$ to $G^1=H^0$. By Example \[ex:c-bound\], $H^s$ is $C$-bounded with $C=\int_{S^1}\max_W(H^1-H^0)\,dt$ and $G^s$ is 0-bounded. Hence, the composition of the homotopies is also $C$-bounded and homotopic to the identity homotopy in the class of $C$-bounded homotopies. By (H2) and (H3), this composition induces the same map as the identity homotopy, i.e., the inclusion-quotient map from Example \[ex:incl-quot\]. Consider first a solution $u$ of the homotopy Floer equation for a homotopy $H^s$, which is (partially) asymptotic to $y$ at $\infty$ and $x$ at $-\infty$. Then, as a direct calculation shows (see, e.g., [@schwarz]), we have $$A_{H^{s^+}}(u(s^+)) - A_{H^{s^-}}(u(s^-)) \leq \int_{s^-}^{s^+}\int_{S^1} (\p_s H^s_t)(u(s))\,dt\,ds -\int_{s^-}^{s^+}\int_{S^1} \left\|\p_s u\right\|^2\,dt\,ds.$$ Hence, $$\label{eq:dif-act-strong} A_{H^1}(y)-A_{H^0}(x)\leq \int_{-\infty}^{\infty}\int_{S^1}\max_W\p_s H^s_t\,dt\,ds - E(u),$$ where $E(u)$ is still defined by . In particular, when $H^s$ is $C$-bounded, we infer that $$\label{eq:dif-act} A_{H^1}(y)-A_{H^0}(x)\leq C,$$ by Definition \[def:c-bound\], and $$\label{eq:energy-action-bound} E(u)\leq C + \left(A_{H^0}(x)-A_{H^1}(y)\right).$$ Proving (H0), let us start with analyzing the case where $W$ is closed. Then, by our assumptions, all periodic orbits of $H^0$ and $H^1$ are non-degenerate and the homotopy is regular, and (H0) follows by the standard gluing and compactness argument; see, e.g., [@Fl; @fh; @SZ; @Sa; @schwarz:book]. The only additional point to check is that gluing and compactification of the moduli spaces in question do not involve periodic orbits outside the range of actions. Assume, for instance, that a homotopy connecting trajectory $u$ from $x\in {{\mathcal P}}^{(a,\,b)}_{H^0}$ to $z\in {{\mathcal P}}^{(a+C,\,b+C)}_{H^1}$ with $\MUCZ(z)=\MUCZ(x)+1$ is obtained by gluing an $H^0$-downward trajectory from $x$ to $x'\in {{\mathcal P}}^{(a,\,b)}_{H^0}$ with a homotopy trajectory from $x'$ to $z$. Then, another component of the boundary of ${{\mathcal M}}_{H^s}(x,z,J^s)$ is a broken trajectory connecting first $x$ to some $y\in {{\mathcal P}}_{H^1}$ by a homotopy trajectory and then $y$ to $z$ by an $H^1$-downward trajectory. We need to check that $y\in {{\mathcal P}}^{(a+C,\,b+C)}_{H^1}$. It is clear that $A_{H^1}(y)\geq A_{H^1}(z)> a+C$. Furthermore, by , we have $A_{H^1}(y)\leq A_{H^0}(x')+C<b+C$. The rest of the proof proceeds in a similar fashion. When $W$ is open, extra care is needed because the Hamiltonians are compactly supported and hence there are degenerate critical points. However, these points do not enter the calculation. Indeed, let, for instance, $x$ and $z$ be as above. Note that the energy of trajectories from ${{\mathcal M}}:={{\mathcal M}}_{H^s}(x,z,J^s)$ is uniformly bounded due to . By the compactness theorem, the closure $\bar{{{\mathcal M}}}$ of ${{\mathcal M}}$ is compact in $C^{\infty}(\R\times S^1,W)$ equipped with the topology of uniform $C^{\infty}$-convergence on compact sets. Then a trajectory $v\in \bar{{{\mathcal M}}}$ also has bounded energy and thus is partially asymptotic to some orbits $x'\in {{\mathcal P}}_{H^0}$ and $y'\in {{\mathcal P}}_{H^1}$. (This can be verified by adapting, for instance, the argument from the proof of Proposition 10 on page 235 of [@HZ].) Moreover, $A_{H^0}(x')\leq A_{H^0}(x)$ and $A_{H^1}(y')\geq A_{H^1}(z)$; cf. [@schwarz:book p.66]. (To prove, for example, the second of these inequalities, pick $s_k\to\infty$ such that $v(s_k)\to y'$ and all $s_k$ are large enough so that $H^{s_k}=H^1$. Then $A_{H^1}(u(s_k))\geq A_{H^1}(z)$ for all $u\in{{\mathcal M}}$. For every $k$ we have a sequence $u_n\in {{\mathcal M}}$ with $u_n(s_k)\to v(s_k)$. Hence, $A_{H^{1}}(v(s_k))\geq A_{H^1}(z)$ and $A_{H^1}(y')=\lim A_{H^{1}}(v(s_k))\geq A_{H^1}(z)$.) It follows immediately that $y'$ is a non-degenerate orbit, since $A_{H^1}(y')\geq A_{H^1}(z)>a+C>0$. Moreover, $A_{H^0}(x')>0$, for otherwise we would have $a+C< A_{H^1}(y')\leq C$ by . Thus $x'$ is also a non-degenerate orbit with positive action. Since $H^s$ is regular, the standard description of $\bar{{{\mathcal M}}}$ via broken trajectories applies, and the proof is finished as for compact manifolds. Properties (H1) and (H2) are established by a similar reasoning invoking the standard gluing and compactification argument, checking that the orbits are in the required ranges of action, and, when $W$ is open, verifying as above that the compactifications of the relevant moduli spaces do not involve orbits with non-positive action. Assertions (H0)–(H2) still hold when the interval $(a+C,\,b+C)$ is replaced by any interval $(a^1,\,b^1)$ with end points outside ${{\mathcal S}}(H^1)$ and such that $a^1\geq a+C$ and $b^1\geq b+C$. This generalization, however, adds little new information, for then the homotopy map $\HF^{(a,\,b)}(H^0)\to \HF^{(a^1,\,b^1)}(H^1)$ is the composition of the homotopy $\HF^{(a,\,b)}(H^0)\to \HF^{(a+C,\,b+C)}(H^1)$ with the inclusion-quotient map $\HF^{(a+C,\,b+C)}(H^0)\to \HF^{(a^1,\,b^1)}(H^1)$. ### Invariance of filtered Floer homology {#sec:inv} Properties (H0)–(H2) of the homotopy maps have a number of standard consequences (see, e.g., [@BPS; @cgk; @fh; @Sa; @SZ; @schwarz:book; @vi:functors]), two of which, (H3) and (H4) below, are of particular relevance for this paper. First we note that (H1) implies that on the level of homology the map $\Psi_{H^0H^1}$ is independent of the homotopy of the almost complex structures. Moreover, this map is actually independent of the homotopy of the Hamiltonians as long as the latter is $C$-bounded. (Indeed, any two $C$-bounded homotopies can be connected by a linear in $\lambda$ family of $C$-bounded homotopies. Furthermore, since regular homotopies are dense, a $C$-bounded homotopy can be approximated by a regular $C'$-bounded homotopy with $C'$ arbitrarily close $C$. Then, the assumption that the end-points of the intervals are outside the action spectra guarantees that the approximating homotopy still induces a map for the action ranges $(a,\,b)$ and $(a+C,\,b+C)$, provided that $C'$ is close enough to $C$.) Thus $\Psi_{H^0H^1}$ is in fact a canonical map depending only on the Hamiltonians $H^0$ and $H^1$, the interval $(a,\,b)$ with $a>0$ when $W$ is open, and the constant $C\geq 0$. This justifies our neglect of the homotopy regularity condition in Examples \[ex:incl-quot\] and \[exam:incr\]. Note, however, that this canonical map is defined only when a $C$-bounded homotopy between $H^0$ and $H^1$ exists which is not always the case. (For instance, such a homotopy fails to exists with $C=0$ whenever $H^1>H^0$ point-wise.) The next result, also quite standard, is the continuity property for filtered homology; see [@BPS; @cgk; @fh; @schwarz:book; @vi:functors]. 1. Let $(a^s,\,b^s)$ be a family (smooth in $s$) of non-empty intervals such that $a^s$ and $b^s$ are outside ${{\mathcal S}}(H^s)$ for some homotopy $H^s$ and such that $(a^s,\,b^s)$ is independent of $s$ when $s$ is near $\pm\infty$ and, abusing notation, equal to $(a^0,\,b^0)$ and $(a^1,\,b^1)$, respectively. Then there exists an isomorphism of homology $$\label{eq:isom} \HF^{(a^0,\,b^0)}(H^0)\stackrel{\cong}{\longrightarrow} \HF^{(a^1,\,b^1)}(H^1).$$ We will outline the proof of (H3) below, after the next assertion (H4) is stated. When the interval is fixed and the homotopy is monotone decreasing, the isomorphism is in fact $\Psi_{H^0H^1}$. We emphasize however that in general this is not the case: the isomorphism is not induced by the homotopy $H^s$, for the homotopy need not map $\CF^{(a^0,\,b^0)}(H^0)$ to $\CF^{(a^1,\,b^1)}(H^1)$ even for a fixed interval of actions. Leaving aside the general question in what sense this isomorphism is canonical and natural with respect to the long exact sequence , we focus on the particular case of a fixed interval. 1. Assume that $a<b<c$ are outside ${{\mathcal S}}(H^s)$ for all $s$. Then the isomorphisms from (H3) for all three intervals can be chosen to map the $(a,\,b,\,c)$ exact sequence for $H^0$ to the $(a,\,b,\,c)$ exact sequence for $H^1$. Let us recall the construction of the isomorphism for a fixed interval, say, $(a,\,b)$. (Here we follow the argument from [@BPS] essentially word-for-word.) First, we break up $H^s$, up to a homotopy of homotopies, into a composition of a finite sequence of homotopies $K^s_i$ each of which is “small”. By this we mean that for all $i$ both $K_i^s$ and the inverse homotopy $K^{-s}_i$ are $\eps$-bounded, where $\eps>0$ is so small that the intervals $(a,\,a+2\eps)$ and $(b,\,b+2\eps)$ do not meet ${{\mathcal S}}(K^s_i)$ for all $s$. Now it is sufficient to prove (H3) and (H4) for one homotopy $K^s=K_i^s$. In this case, the isomorphism is defined as $$\label{eq:map} \HF^{(a,\,b)}(K^0)\to \HF^{(a+\eps,\,b+\eps)}(K^1) \stackrel{\cong}{\longrightarrow} \HF^{(a,\,b)}(K^1),$$ where the second arrow is the inverse of the quotient-inclusion map. (The latter map is obviously an isomorphism due to the requirement imposed on $\eps$.) To see that is an isomorphism, observe that its inverse is given by $$\HF^{(a,\,b)}(K^1) \stackrel{\cong}{\longrightarrow} \HF^{(a+\eps,\,b+\eps)}(K^1)\to \HF^{(a+2\eps,\,b+2\eps)}(K^0) \stackrel{\cong}{\longrightarrow} \HF^{(a,\,b)}(K^0),$$ where the middle arrow is the map induced by the inverse homotopy $K^{-s}$. This concludes the proof of (H3) for a fixed interval. The general case is established in a similar fashion. Finally, from the second part of (H0), it is clear that the maps defined for all three intervals send the $(a,\,b,\,c)$ exact sequence for $K^0$ to that for $K^1$. This proves (H4). Another standard consequence of (H0)–(H2), along the lines of (H3) and (H4), is that the filtered Floer homology can be defined “by continuity” for any Hamiltonian $H$, not necessarily non-degenerate, and any $a<b$ outside ${{\mathcal S}}(H)$. (See, e.g., [@vi:functors].) Moreover, the long exact sequence and (H0)–(H4) also hold in this case. Here, as above, we are assuming that $a>0$ when $W$ is open. The action selector {#sec:selector} ------------------- In this section we recall the definition and the properties of the action selector that are used in the proof of Proposition \[prop:orbits-nondeg\]. The constructions differ somewhat depending on whether the manifold is open and wide or closed. We start by dealing with the case of wide manifolds, closely following [@gu:new], for this case appears to be more relevant to the displacement questions and, from the author’s perspective, more transparent. Then, we very briefly review the construction in the case where $W$ is closed. Here we follow [@schwarz] with some minor alterations. Note also that for $\R^{2n}$ and cotangent bundles actions selectors were constructed in [@HZ] and [@Vi:gen], respectively, and the approach of [@schwarz] has been extended to manifolds convex at infinity in [@FS]. ### The action selector for wide manifolds {#sec:sel-open} Assuming that $W^{2n}$ is geometrically bounded and wide (see Section \[sec:statement\] or [@gu:new]), let us recall from [@gu:new] the definition of the action selector $\sigma(K)$ for a compactly supported, non-negative Hamiltonian $K\colon S^1 \times W\to \R$. It is easy to see that since $W$ is wide, there exists a smooth compactly supported function $F\colon W\to [0,\,\infty)$ without non-trivial contractible periodic orbits with period $T\leq 1$ and such that $F\geq K$ point-wise. Without loss of generality, we may assume that ${\operatorname{supp}}F$ is a smooth connected manifold with boundary and that the restriction of $F$ to the interior of its support is a Morse function with finitely many critical points. Under these assumptions, we have $$\HF_*^{(0,\,\infty)}(F)\cong \HM_{*+n}^{(0,\,\infty)}(F) \cong H_{*+n}({\operatorname{supp}}F,\partial{\operatorname{supp}}F;\Z_2)$$ and, in particular, $$\HF_n^{(0,\,\infty)}(F)\cong \Z_2.$$ (Recall that the filtered Floer homology for $(0,\,\infty)$ is defined by .) Denote the generator of this group – the fundamental class – by $[{\max}_F]$. A monotone decreasing homotopy from $F$ to $K$ induces a map $$\Psi_{F,K}\colon \HF^{(0,\,\infty)}(F)\to\HF^{(0,\,\infty)}(K),$$ independent of the choice of homotopy. (This follows from the results of the previous section and .) Set $$[{\max}_K]=\Psi_{F,K}\left([{\max}_F]\right)\in\HF^{(0,\,\infty)}_n(K).$$ It is not hard to show that $[{\max}_K]$ is well defined, i.e., independent of the choice of $F$. Then, for $K\geq 0$, we set $$\sigma(K)=\inf\{a> 0\mid j^a_K\left([{\max}_K]\right)=0\}\in {{\mathcal S}}(K),$$ where $$j^a_K\colon \HF^{(0,\,\infty)}(K)\to \HF^{(a,\,\infty)}(K)$$ is the quotient map. Finally note that this definition makes sense even when $K$ is not assumed to be non-negative. However, it is not clear whether the resulting action selector has the required properties (S0)–(S4) stated below. If $W$ is convex at infinity, the actions selector described above coincides, for non-negative Hamiltonians, with the one of [@FS]. ### The action selector for closed manifolds {#sec:sel-closed} Assume that $W^{2n}$ is closed. Then for any Hamiltonian $K$, we have the canonical identification $$\HF_*(K)\cong H_{*+n}(W;\Z_2).$$ Denoting by $[{\max}_K]$ the image of the fundamental class under this isomorphism, set, following [@schwarz], $$\sigma(K)=\inf\{a\in\R\mid j^a_K\left([{\max}_K]\right)=0\}\in {{\mathcal S}}(K),$$ where $$j^a_K\colon \HF(K)\to \HF^{(a,\,\infty)}(K)$$ is the quotient map. Observe that when $K\geq 0$ we can assume that $a\geq 0$ by (S1) below, as in the definition of the action selector for open manifolds. Note that here, in contrast with [@schwarz], we consider non-normalized Hamiltonians, i.e., we are not assuming that $\int_W K_t\omega^n=0$. As a consequence, neither ${{\mathcal S}}(K)$ nor $\sigma(K)$ is uniquely determined by $\varphi_K$. Alternatively, one can define the class $[{\max}_K]$ as follows. Let $F$ be a Morse function which is $C^2$-close to a constant. Then, canonically, $$\HF_*(F)\cong\HM_{*+n}(F)\cong H_{*+n}(W;\Z_2).$$ Let ${\max}_F\in{\operatorname{CM}}_{2n}(F)$ be the “fundamental class”, i.e., the sum of all local maxima. Then $[{\max}_F]$ is the generator of $\HF_n (F)\cong \Z_2$. Now, for a fixed $K$, we take $F\geq K$ and define $[{\max}_K]$ to be the image of $[{\max}_F]$ under the monotone decreasing homotopy map $$\Psi_{F,K}\colon \HF(F)\to\HF(K).$$ ### Properties of the action selector {#sec:sel-prop} The action selector $\sigma$, defined as above, has the following properties for non-negative Hamiltonians, regardless of whether $W$ is wide or closed: 1. $\sigma$ is monotone, i.e., $\sigma(K)\leq\sigma(H)$, whenever $0\leq K\leq H$ point-wise; 2. $0\leq\sigma(K)\leq E^+(K)$ for any $K\geq 0$, where $$E^+(K)=\int_{S^1}\max_W K_t\,dt;$$ 3. $\sigma(K)>0$, provided that $K\geq 0$ is not identically zero; 4. $\sigma(K)$ is continuous in $K$ in the $C^0$-topology; 5. $\sigma(K)\leq E^+(H)$, whenever $\varphi_H$ displaces ${\operatorname{supp}}K$ and $H\geq 0$. We refer the reader to [@gu:new] for the proofs of (S1)–(S3) when $W$ is wide and to [@schwarz] when $W$ is closed. As stated, (S4) is established in [@FGS] for closed manifolds and in [@gu:new] for wide manifolds in the stronger form $\sigma(K)\leq \sigma(H)$, following the earlier versions of this upper bound from [@FS; @gi:alan; @HZ; @Oh:disj; @schwarz; @Vi:gen]. Note also that if all contractible one-periodic orbits of $K$ are non-degenerate, setting $A_K(c)=\sup_i A_K(c_i)$ for $c=\sum c_i\in \CF^{(0,\,\infty)}(K)$, we have $$\sigma(K)=\inf_{[c]=[{\max}_K]} A_K(c).$$ As a consequence, there exists a cycle $c$ (possibly non-unique) with $A_K(c)=\sigma(K)$. We call such a cycle a *carrier* of the action selector. ### Homological capacity {#sec:homcap} Let $U\neq W$ be an open set in $W$, where $W$ is either closed or geometrically bounded and wide. Recall that the homological capacity $\chom(U)\in [0,\,\infty]$ of $U$ is defined as $$\chom(U)=\sup_{{\operatorname{supp}}K\subset U} \sigma(K).$$ For a compact subset $Z$ of $W$ define the displacement energy $e(Z)$ as $\inf \| H\|$, where the infimum is taken over all compactly supported Hamiltonians $H$ with $\varphi_H(Z)\cap Z=\emptyset$, and set $e(U)=\sup_{Z\subset U} e(Z)$. Then $\chom(U)\leq e(U)$. This well-known fact (see, e.g., [@FS; @gu:new; @HZ; @schwarz; @Vi:gen]) follows, for instance, from (S4). For a closed set $M\subset U$ (e.g., a closed submanifold), we set $$\chom(M)=\inf_{M\subset U} \chom(U).$$ Note that $\chom(M)=0$ whenever $M$ is infinitesimally displaceable. On the other hand, drawing on the results of [@vi:functors], one can expect that a closed Lagrangian submanifold necessarily has positive homological capacity. In Remark \[rmk:Dr\], we show that this is also true for stable coisotropic submanifolds. It seems to be unknown whether the homological capacity of every closed hypersurface in $\R^{2n}$ is positive. Displacement and connecting trajectories ======================================== [\[sec:nondeg\]]{} Let, as above, $W$ be symplectically aspherical and either closed or geometrically bounded and wide. The key to the proof of Theorem \[thm:main\] is the following result. \[prop:orbits-nondeg\] Let $U$ be an open subset of $W$ displaced by a Hamiltonian diffeomorphism $\varphi_H$. Furthermore, let $K\geq 0$ be a Hamiltonian supported in $U$ and such that $\max K$ is large enough (e.g., $\max K> \parallel H \parallel$ and $\max K> \max {{\mathcal S}}(H)$, provided that $H\geq 0$). Assume furthermore that - all one-periodic orbits from ${{\mathcal P}}_K^{(0,\,\infty)}$ are non-degenerate; - for all $t\in [0,\,1]$ the functions $K_t$ attain their maximum at the same point $p\in U$, which is thus a non-degenerate maximum of $K_t$ for all $t$; - $K$ is autonomous (i.e., independent of $t$) near $p$. - the eigenvalues of the Hessian $d^2 K_t$ of $K_t$ at $p$ are close to zero, and thus $K$ has no one-periodic orbits other than $p$ in a neighborhood of $p$. Then the flow of $K$ has a contractible one-periodic orbit $\gamma$ in $U$ which is connected to $p$ by a Floer anti-gradient trajectory and such that $$\label{eq:action} 0<A_K(\gamma)- \max K \leq \parallel H \parallel.$$ Here we assume that $W$ is equipped with an almost complex structure, possibly time-dependent, which is compatible with $\omega$ and such that the regularity requirements are satisfied. Hence the Floer complex of $K$ is defined. As stated, with the upper bound in , the proposition is apparently new. (This upper bound is absolutely crucial for the proof of Theorem \[thm:main\] and without it the proposition is quite straightforward.) However, it should be noted that in the context of Hofer’s geometry a number of somewhat similar results have been established, often under much less restrictive conditions on the ambient manifold; see [@En; @KL; @LM; @MDS; @Oh:chain]. One can expect that along the lines of some of these results the difference $A_K(\gamma)-\max K$ can hypothetically be bounded from above in terms of the Hofer norm of the time-one flow $\varphi_K$. *Outline of the proof.* Without loss of generality, we may assume that $E^+(H)=\| H\|$. Since $\sigma(K)\leq \| H\|<K(p)=A_K(p)$, the maximum $p$ is not homologically essential, i.e., there exists a carrier $y$ of the action selector, not containing $p$, and the “maximum cycle” $z\in \CF^{(0,\,\infty)}_n(K)$ containing $p$ such that $z-y=\p x$ for a chain $x\in \CF^{(0,\,\infty)}_{n+1}(K)$. As a consequence, there is a Floer connecting trajectory $u$ from some orbit $\gamma$ in $x$ to $p$. We can choose $x$ so that the class $[x^c]$ induced by $x$ in $\HF^{(c,\,\infty)}_{n+1}(K)$ is non-zero as long as $K(p)<c<A_K(\gamma)$ and hence for $K(p)<a<b<A_K(\gamma)$ the quotient map $\HF^{(a,\,\infty)}(K)\to \HF^{(b,\,\infty)}(K)$ sending $[x^a]$ to $[x^b]$ is also non-zero. Note that, since $\varphi_H$ displaces ${\operatorname{supp}}K$, we have $\HF^{(b,\,\infty)}(K\#H)= \HF^{(b,\,\infty)}(H)=0$ once $b>\max{{\mathcal S}}(H)$. Finally, arguing by contradiction, assume that the gap between $K(p)$ and $A_K(\gamma)$ is greater than $E^+(H)$, i.e., $E(u)=A_K(\gamma)-K(p)> E^+(H)$. Now we utilize the fact that the homology of $K$ cannot be completely destroyed by a relatively small perturbation $H$. This implies that $\HF^{(b,\,\infty)}(K\#H)\neq 0$ as along as $[b-E^+(H),\, b]\subset(\max K,\, A_K(\gamma))$. Hence, when $\max K$ is large enough, the gap cannot exceed $E^+(H)$ and $E(u)=A_K(\gamma)-K(p)> E^+(H)$. First note that without loss of generality we may assume that $H$ meets the following requirements: - $H$ is periodic in time and $H_0=H_1\equiv 0$; - $H$ is compactly supported and $\min_W H_t=0$ for all $t\in S^1$. The first requirement can be satisfied by reparametrizing $H$, i.e., replacing $H$ by a new Hamiltonian of the form $\lambda'(t)H(\lambda(t),x)$ where $\lambda\colon [0,\,1]\to [0,\,1]$ is a monotone increasing function identically equal to zero for $t$ near zero and to one for $t$ near one. To have the second requirement met, we replace $H$ by a Hamiltonian of the form $f_t\cdot (H_t-\min_W H_t)$, where $f_t$ is a non-negative cut-off function equal to one on $\varphi^t_H(U)$. Note that neither of these alterations changes $\| H\|$. Both of the requirements are purely technical rather than essential. The first one is needed to ensure that the composition $K\# H$ is one-periodic in time. The second one enables us to treat wide and closed manifolds in the same way, with only superficial discrepancies. In what follows we will assume that $H$ meets these conditions. Note that since $\min_W H_t=0$, we have $H\geq 0$ and $\| H\|=E^+(H)$. The proof of the theorem proceeds slightly differently depending on whether $W$ is wide or closed. Below, we first consider the case of wide manifolds and then indicate modifications needed when $W$ is closed. *Wide manifolds.* Let $F\geq K$ be an autonomous Hamiltonian as in the definition of the action selector. Without loss of generality, we may assume that $F$ has a unique local maximum which is also located at $p$ and that $F(p)=K(p)$. Then under a monotone homotopy from $F$ to $K$ the Floer cycle $p={\max}_F\in \CF_n^{(0,\,\infty)}(F)$ is mapped to a Floer cycle $z=z_1 +\cdots + z_m$ of $K$ with $z_1=p$. We emphasize that the point $p$ does occur in $z$. The reason is that the trivial Floer connecting trajectory from $p$ for $F$ to $p$ for $K$ is the only connecting trajectory between these two trivial orbits by the standard energy argument. (Here we use the assumption that $F(p)=K(p)$.) Furthermore, by utilizing the condition that the eigenvalues of $d^2K_t$ are close to zero, it is easy to arrange the homotopy so that this connecting trajectory is non-degenerate, cf. [@FHS]. Note also that since the homotopy is decreasing, we have $$F(p)=K(p)=A_{K}(p)\geq A_{K}(z_i)\text{ for all $i=2,\ldots, m$},$$ and hence $$A_{K}(z)=K(p)=\max K.$$ Clearly, $[z]=[{\max}_{K}]$, but $z$ cannot be the carrier of the action selector for $K$, for $$A_{K}(p)=\max K > E^+(H)\geq \sigma (K),$$ where the last inequality follows from (S4). (In other words, $p$ is not homologically essential.) Let $y\in \CF^{(0,\,\infty)}_n(K)$ be a carrier of the action selector $\sigma(K)$. Then $[y]=[{\max}_{K}]=[z]$ and $z-y=\p x$ for some $x\in \CF_{n+1}^{(0,\,\infty)}(K)$. Furthermore, it is clear from this chain of inequalities that $p$ does not enter the cycle $y$ and, as a consequence, $x$ contains a periodic orbit $\gamma$ connected with $p$ by a Floer anti-gradient trajectory. Then $$0<A_K(\gamma)-\max K.$$ However, to ensure that holds in its entirety, we need to impose an additional requirement on the chain $x$ and the orbit $\gamma$ which are in general not unique. To this end, for a given $x$ denote by $x_1$ an orbit, occurring in $x$, which is connected to $p$ and has the smallest possible action $A_{K}(x_1)$ among all such orbits in $x$. Then we chose $x$ with $A_{K}(x_1)$ attaining the smallest value for all $x$ with $z-y=\p x$. We will show that for any $x$ and $\gamma=x_1$, meeting this action minimization condition, $${\label{eq:action2}} A_{K}(x_1)\leq \max K + E^+(H),$$ which implies . First let us consider the standard quotient map $$j\colon \HF_{n+1}^{(a,\,\infty)}(K)\to\HF_{n+1}^{(b,\,\infty)}(K).$$ We claim that, provided that $x$ and $x_1$ are action minimizing as above, $$\label{eq:nonzero} j\neq 0, \text{ when $\max K< a < b < A_K(x_1)$}.$$ (In particular, both of the groups are non-zero.) To see this, denote by $x^a$ and $x^b$ the images of the chain $x$ in $\CF_{n+1}^{(a,\,\infty)}(K)$ and $\CF_{n+1}^{(b,\,\infty)}(K)$, respectively. By definition, $j(x^a)=x^b$ on the level of complexes. Furthermore, both $x^a$ and $x^b$ are closed, since $\p x=z-y\in \CF^{(0,\,\,a)}_n(K)$. Hence, it suffices to show that $[x^b]\neq 0$ in $\HF_{n+1}^{(b,\,\infty)}(K)$. Assume the contrary: there exists $w\in \CF_{n+2}^{(b,\infty)}(K)$ with $\p w= x^b$. In other words, there exists a chain in $w'\in\CF_{n+2}^{(0,\,\infty)}(K)$ such that $x'=x-\p w'\in \CF_{n+2}^{(0,b)}(K)$. Then clearly $\p x'=z-y$ and no orbit entering $x'$ has action in the interval $(b,\,\infty)$, where $b< A_K(x_1)$, which is impossible due to our choice of $x$. This contradiction completes the proof of . Proceeding with the proof of , we again make use of the fact that $U$ is displaced by $\varphi_H$. Denote by $K\# H$ the Hamiltonian $$(K\# H)_t=K_t+H_t\circ(\varphi^t_K)^{-1}$$ generating the time-dependent flow $\varphi^t_K\circ\varphi^t_H$. Since $\varphi_H$ displaces ${\operatorname{supp}}K$, the one-periodic orbits of $K\# H$ are exactly the one-periodic orbits of $H$ and moreover ${{\mathcal S}}(K\# H)={{\mathcal S}}(H)$, as is well known (see, e.g., [@HZ]). Thus, by the continuity property (H3) of Floer homology, $$\HF^{(b,\,b')}(K\# H)=\HF^{(b,\,b')}(H)$$ for any $b<b'$ which are not in ${{\mathcal S}}(H)$. In particular, $$\label{eq:zero} \HF^{(b,\,\infty)}(K\# H)\cong \HF^{(b,\,\infty)}(H)=0, \text{ whenever $b>\max {{\mathcal S}}(H)$.}$$ Consider a linear monotone increasing homotopy from $K$ to $K\#H$. (Recall that $H\geq 0$.) By Example \[exam:incr\], such a homotopy induces a map $$\Psi_{K,K\#H}\colon \HF^{(a,\,\infty)}(K)\to \HF^{(b,\,\infty)}(K\# H), \text{ where $b=a+ E^+(H) $,}$$ whenever $a\not\in{{\mathcal S}}(K)$ and $b\not\in{{\mathcal S}}(K\# H)$. The composition of this map with the map $$\Psi_{K\#H,K}\colon \HF^{(b,\,\infty)}(K\#H)\to \HF^{(b,\,\infty)}(K)$$ induced by a monotone decreasing homotopy from $K\#H$ to $K$ is the quotient map $j$. (See Example \[exam:incr\].) To finish the proof in the case where $W$ is wide, assume that fails: $E^+(H) < A_K(x_1)-\max K$. Pick $a> \max K$ such that $b=a+ E^+(H) < A_K(x_1)$, and $a$ and $b$ are not in ${{\mathcal S}}(K)$ and also $b\not\in{{\mathcal S}}(K\#H)$. Then $j= \Psi_{K\#H,K}\circ\Psi_{K,K\#H}$ factors through the group $\HF^{(b,\,\infty)}(K\# H)$. Since $b>a>\max K> \max {{\mathcal S}}(H)$, this group is zero by . Hence, $j=0$, which contradicts . *Closed manifolds*. Only two points of the above argument require modifications when $W$ is closed. The first of these is the definition of the cycle $z$. Now we take as $F\geq K$ a function which is $C^2$-close to a constant and has a unique local maximum, equal to $K(p)$, that is also located at $p$. Then as above $z$ is the image of the Floer cycle $p={\max}_F\in \CF_n^{(0,\,\infty)}(F)$ under a monotone homotopy from $F$ to $K$. It is clear that $p$ occurs in $z$, i.e., $z=p+z_2 +\cdots + z_m$ for the same reason as for open manifolds. The second point of the proof that is not obvious when $W$ is closed, for we work with non-normalized Hamiltonians, is the equality of action spectra ${{\mathcal S}}(K\#H)={{\mathcal S}}(H)$. To see that this is the case, let us first recall how the action spectrum depends on the Hamiltonian. Let $G^s$ be a homotopy of Hamiltonians such that the time-one flows $\varphi_{G^s}$ are independent of $s$, i.e., the Hamiltonians $G^s$ determine the same element in the universal covering of the group ${{\mathit Ham}}(W)$. Then $$\label{eq:spectra} {{\mathcal S}}(G^1)={{\mathcal S}}(G^0)+ \int_{S^1}\int_W \left(G^1_t-G^0_t\right)\omega^n\,dt.$$ This can be established by arguing as in the proof of [@schwarz Lemma 3.3].[^3] Following, e.g., [@HZ] or [@FGS], consider a homotopy from $K=K^0$ to $K^1$ through reparametrizations $K^s_t=\lambda'(t) K_{\lambda(t)}$ of $K$ such that $K^1_t\equiv 0$ when $t\in [0,\,1/2]$. Let $H^s$ be a similar homotopy beginning with $H^0=H$ and ending with $H^1$, and such that $H^1_t\equiv 0$ for $t\in [1/2,\,1]$. Clearly, these homotopies do not change the action spectra. Let $G^s=K^s\# H^s$. It is easy to see that ${{\mathcal S}}(K^1\#H^1)={{\mathcal S}}(H^1)={{\mathcal S}}(H)$. (The first equality follows from the fact that $K^1\#H^1$ and $H^1$ have literally the same one-periodic orbits, with the same parametrizations, and the two Hamiltonians are equal along these orbits.) Furthermore, a direct calculation shows that $$\begin{aligned} \int_{S^1}\int_W G^1_t\omega^n\,dt &=& \int_{S^1}\int_W (K^1_t+H^1_t)\omega^n\,dt\\ &=& \int_{S^1}\int_W (K_t+H_t)\omega^n\,dt\\ &=& \int_{S^1}\int_W G^0_t\omega^n\,dt.\end{aligned}$$ Thus, by , ${{\mathcal S}}(K^1\# H^1)={{\mathcal S}}(K\# H)$. As a consequence, ${{\mathcal S}}(H)={{\mathcal S}}(K\# H)$. The rest of the proof proceeds exactly as in the case of open manifolds. Connecting trajectories for degenerate displaceable Hamiltonians {#sec:deg} ================================================================ In this section we extend Proposition \[prop:orbits-nondeg\] to Hamiltonians whose one-periodic orbits are degenerate. This result will be used in the proof of Theorem \[thm:main\]. The space of finite energy anti-gradient trajectories ----------------------------------------------------- Let, as above, $W$ be a geometrically bounded, symplectically aspherical manifold and let $K$ be a compactly supported Hamiltonian on $W$. Denote by ${{\mathcal B}}={{\mathcal B}}(K)$ the space of contractible Floer anti-gradient trajectories with finite energy, i.e., the space of solutions of such that $E(u)<\infty$ and $u(s)$ is contractible. We equip this space with the weak $C^\infty$-topology, i.e., the topology of uniform $C^\infty$-convergence on compact sets. Note that ${{\mathcal P}}_K\subset{{\mathcal B}}$. Recall that by the compactness theorem any sequence of $u_i\in {{\mathcal B}}$ such that $u_i(0,0)$ is bounded contains a converging subsequence. In fact, ${{\mathcal B}}$ is a locally compact, separable, metrizable space, and the evaluation map $$\ev\colon {{\mathcal B}}\to W,\text{ defined as $\ev(u)=u(0,0)$,}$$ is continuous and proper; see, e.g., [@HZ] and references therein. Moreover, this map is a homeomorphism on a complement of a compact subset of ${{\mathcal B}}$. For $u\in {{\mathcal B}}$, set $A_K(u)$ to be the action of $K$ on the closed curve $u(0)$. Then $A_K\colon {{\mathcal B}}\to \R$ is a continuous function, which is identically zero outside a compact set. The anti-gradient flow $\Phi$ of $A_K$ on ${{\mathcal B}}$ is defined as the shift in the $\tau$-direction: $\Phi^\tau(u)(s)=u(s+\tau)$ for all $\tau\in\R$. Obviously, the fixed points of this flow, i.e., finite energy trajectories independent of $s$, are exactly the one-periodic orbits of $K$. The function $A_K$ is decreasing along the orbits of $\Phi^\tau$, i.e., $A_K(\Phi^\tau(u))\leq A_K(u)$ for any $\tau\geq 0$. Moreover, $A_K$ is strictly decreasing along non-trivial orbits: $A_K(\Phi^\tau(u))< A_K(u)$ for any $\tau>0$, when $u$ is not a fixed point of the flow. Denote by $\Gamma(u)=\{\Phi^\tau(u)\mid \tau\in \R\}$ the orbit of $\Phi$ through $u$. By definition, the limit set $\omega^+(u)$ is formed by the limits of all converging sequences $\Phi^{\tau_i}(u)$ with $\tau_i\to\infty$. The limit set $\omega^-(u)$ is defined similarly, but with $\tau_i\to-\infty$. The following properties of the closure $\bG(u)=\Gamma(u)\cup\omega^+(u)\cup\omega^-(u)$ and of the limit sets $\omega^\pm(u)$ are well known: - the sets $\bG(u)$ and $\omega^\pm(u)$ are compact, connected, and invariant under the flow $\Phi$; - the action functional $A_K$ is constant on $\omega^\pm(u)$, and $A_K|_{\omega^+(u)}\equiv \min_{\bG(u)}A_K$ and $A_K|_{\omega^-(u)}\equiv \max_{\bG(u)}A_K$; - both of the limit sets $\omega^\pm(u)$ are non-empty and entirely comprised of the fixed points of $\Phi$. It is clear that $u$ is partially asymptotic to $x^\pm$ as $s\to\pm\infty$, i.e., $u(s_i^\pm)\to x^\pm$ in $C^\infty(S^1,W)$ for some sequences $s_i^\pm\to \pm\infty$, if and only if $x^\pm\in\omega^\pm(u)$. (Here we treat $x^\pm$ simultaneously as periodic orbits (elements of ${{\mathcal P}}_K$) and as finite energy trajectories (elements of ${{\mathcal B}}$).) If $x^\pm$ are non-degenerate, we necessarily have $\omega^\pm(u)=\{x^\pm\}$. Otherwise, the limit sets can be quite large and $x^\pm$ are not unique. Existence of connecting trajectories for degenerate Hamiltonians ---------------------------------------------------------------- In this section we prove an analogue of Proposition \[prop:orbits-nondeg\] for degenerate Hamiltonians. Let, as in Section \[sec:nondeg\], $W$ be a symplectically aspherical manifold, which is either closed or geometrically bounded and wide. \[prop:orbits-deg\] Let $U$ be an open subset of $W$ displaced by a Hamiltonian diffeomorphism $\varphi_H$. Furthermore, let $K\geq 0$ be a Hamiltonian supported in $U$ and such that for all $t\in [0,\,1]$ the functions $K_t$ attain their maxima at the same connected set $M\subset U$ and are autonomous near $M$. Assume also that $K$ is $C^2$-close to a constant on a sufficiently small neighborhood of $M$. Then, provided that $\max K$ is large enough, there exists $u\in{{\mathcal B}}(K)$ partially asymptotic to a point of $M$ and to a contractible one-periodic orbit $\gamma$ such that holds: $$0<A_K(\gamma)- \max K \leq \parallel H \parallel.$$ \[rmk:orbits-deg\] In the context of this proposition, $W$ is equipped with an arbitrary almost complex structure $J$ compatible with $\omega$ in the sense of geometrically bounded manifolds and independent of time at infinity. Note also that guarantees that $A(\gamma)>0$ and, in particular, $\gamma$ is a non-trivial one-periodic orbit when $K$ is autonomous. Furthermore, $A_K(\gamma)- \max K=E(u)$. Proposition \[prop:orbits-deg\] will be applied in the setting where $M$ is a Morse–Bott non-degenerate critical set of $K$. Then it might be possible to conclude that $u$ is genuinely asymptotic to a point on $M$ as $s\to\infty$; cf. [@poz]. However, the remaining one-periodic orbits of $K$ are still very degenerate and in this case $u$ need not be truly asymptotic to any orbit at $-\infty$. (This phenomenon, occurring already for the anti-gradient flow of a function on a smooth manifold, is overlooked in [@Bo2]. However, the argument of [@Bo2] still goes through for partial asymptotics.) Note also that some degree of control over periodic orbits of $K$ near $M$ is essential. Here, it is achieved through requiring $K$ to be $C^2$-close to a constant near $M$ and thus to have no one-periodic orbits near $M$ with action greater than $\max K$. Without a restriction on the behavior of $K$ near $M$ the proposition probably fails. The idea of the proof is, of course, to approximate $K$ by a sequence $K_l\to K$ of non-degenerate Hamiltonians satisfying the hypotheses of Proposition \[prop:orbits-nondeg\] and to define the orbit $u$ as the limit of a sequence of trajectories $u^0_l$ for $K_l$ such as in Proposition \[prop:orbits-nondeg\]. The nuance is that an arbitrary sequence $u^0_l$ need not have an orbit with the required properties as its limit point. (For instance, $u^0_l$ can converge to a point on $M$ or to a trajectory which is not partially asymptotic to a point on $M$.) However, the trajectories $u^0_l$ are not unique and can be replaced by $u_l=\Phi^{\tau_l}(u^0_l)$ for any sequence of shifts $\tau_l$. We show that $u_l$ converges to the required $u$ for some sequence $\tau_l$, using an elementary, point–set topological argument. Pick a sequence of non-degenerate Hamiltonians $K_l$ which are supported in $U$ and $C^\infty$-converge to $K$ and satisfy the hypotheses of Proposition \[prop:orbits-nondeg\]. Without loss of generality, we may assume that all $K_l$ attain their maxima at the same point $p\in M$ and that $\max K_l=K_l(p)=\max K$. Furthermore, we can approximate $J$ by almost complex structures $J_l$ (possibly time-dependent) which are compatible with $\omega$ and equal to $J$ at infinity and such that the pairs $(K_l,J_l)$ are regular. Let $u_l^0$ be an anti-gradient trajectory for $K_l$ whose existence is established in Proposition \[prop:orbits-nondeg\]: $u_l^0$ connects a contractible one-periodic orbit $\gamma^0_l$ of $K_l$ to $p$ and holds. By passing if necessary to a subsequence, we may assume that the sequence $\gamma_l^0$ converges to a one-periodic orbit $\gamma^0$ of $K$. Recall that $K$ is assumed to be autonomous and $C^2$-close to a constant near $M$. It follows that every one-periodic orbit of $K$, with action in the interval $[\max K, \infty)$, that meets a sufficiently small neighborhood $V$ of $M$ must be a point of $M$. Furthermore, $K_l$ are also $C^2$-close to a constant on $V$. As a consequence, every one-periodic orbit of $K_l$ entering $V$ is trivial and we may require, in addition, that the actions of $K_l$ on these orbits (with the exception of $p$) are strictly smaller than $\max K_l$. (It is easy to show that the approximations $K_l$ with this property do exist.) Since $K_l(\gamma^0_l)>\max K_l$ by , we conclude that $\gamma^0_l$ does not enter $V$. Hence, $\gamma^0$ also lies entirely outside $V$. Recall that by the compactness theorem any sequence $u_l\in{{\mathcal B}}(K_l)$ contains a converging subsequence, provided that the sequence $\ev(u_l)$ is bounded in $W$. (Here convergence is again understood as $C^\infty$-convergence uniform on compact subsets of $\R\times S^1$, i.e., in the weak $C^\infty$-topology on the space ${{\mathcal C}}$ of smooth maps $\R\times S^1\to W$.) Set $\Gamma_l=\Gamma(u^0_l)$ and denote by $\Sigma\subset {{\mathcal C}}$ the set of all limit points of sequences $u_l\in \Gamma_l$ (or, equivalently, $u_l\in\bG_l$). This set has the following properties: 1. the set $\Sigma$ contains $p$ and $\gamma_0$; 2. the sets $\bG_l$ converge to $\Sigma$ in the Hausdorff topology: for every neighborhood ${{\mathcal U}}$ of $\Sigma$ in ${{\mathcal C}}$ we have $\bG_l\subset{{\mathcal U}}$ when $l$ is large enough; 3. the set $\Sigma$ is connected, compact, and invariant under the flow $\Phi$; 4. the flow $\Phi$ on $\Sigma$ is non-trivial; 5. the action functional $A_K$ is not constant on $\Sigma$, and $\min_\Sigma A_K\mid =A_K(p)=\max K$ and $\max_\Sigma A_K\leq \max K+\parallel H\parallel$. The first assertion (a) is obvious. To prove (b), we just note that otherwise we would have a sequence $u_{l_i}\in\bG_{l_i}\ssminus{{\mathcal U}}$ for some $l_i\to \infty$. By compactness, $u_{l_i}$ must have a limit point outside ${{\mathcal U}}$ and hence not in $\Sigma$, which is impossible. In (c), only the fact that $\Sigma$ is connected requires a proof. This readily follows from (b) and from the fact that all $\bG_l$ contain the point $p$. Assertion (e) is obvious except for the statement that $A_K$ is non-constant, which is a consequence of (d). To prove (d), observe that $\ev(\Sigma)$ contains both $p\in M\subset V$ and $\gamma^0(0)\not \in V$. By (c), $\ev(\Sigma)$ is connected and therefore there exists $u\in\Sigma$ such that $u(0,0)=\ev(u)\in V\ssminus M$. Since no one-periodic orbit of $K$ with action in $[\max K, \infty)$, other than the points of $M$, enter $V$, we conclude that $u(0)$ is not a contractible one-periodic orbit of $K$ and, in particular, the entire anti-gradient trajectory $u$ cannot be a contractible one-periodic orbit of $K$. Thus $u$ is not a fixed point of the flow $\Phi$. Let $\Sigma_{\min}$ be the set of $u\in\Sigma$ at which $A_K|_{\Sigma}$ attains its minimum $\max K$. Note that $\Sigma_{\min}$ is entirely comprised of fixed points of $\Phi$ or, equivalently, of periodic orbits of $K$. Regarding $M$ as a subset of ${{\mathcal B}}(K)$, set $C=M\cap \Sigma_{\min}=M\cap \Sigma$. Then $C$ is a compact, proper subset of $\Sigma$. Indeed, compactness of $C$ is obvious. By (e), $\Sigma_{\min}\neq \Sigma$, and hence $C\neq \Sigma$. On the other hand, $C\neq\emptyset$, for $p\in C$. Next, we claim that 1. the set $C$ is a union of connected components of $\Sigma_{\min}$, i.e., $\Sigma_{\min}\ssminus C$ is closed. Indeed, assume the contrary. Then, there exists a sequence of periodic orbits $u_l\in \Sigma_{\min}\ssminus C$ converging to a point of $C$. As a consequence, $\ev(u_l)$ converges to a point of $M$. Then $u_l$ must be a trivial periodic orbit when $l$ is large enough. (For $u_l$ enters $V$.) In addition, $A_K(u_l)=\max K$ and thus $u_l$ is a point of $M$. This is impossible since $u_l\in \Sigma_{\min}\ssminus C$. Set ${{\mathcal N}}_\eps=\{u\in\Sigma\mid A_K(u)<\max K+\eps\}$. Fix a connected component $C_0$ of $C$ and let ${{\mathcal U}}_\eps$ be the connected component of ${{\mathcal N}}_\eps$ that contains $C_0$. Clearly, the sets ${{\mathcal U}}_\eps$ with $\eps>0$ are open, nested, and invariant under the positive flow $\Phi^{\tau\geq 0}$. Our next goal is to prove 1. the open sets ${{\mathcal U}}_\eps$ with $\eps>0$ form a fundamental system of neighborhoods of $C_0$, i.e., for any open set ${{\mathcal U}}\supset C_0$, we have ${{\mathcal U}}_\eps\subset {{\mathcal U}}$, when $\eps>0$ is sufficiently small. Let us first show that the sets ${{\mathcal N}}_\eps$ form a fundamental system of neighborhoods of $\Sigma_{\min}$ for $\eps>0$. Assume the contrary. Then for some open set ${{\mathcal W}}\supset \Sigma_{\min}$, there exists a sequence $u_l\in {{\mathcal N}}_{\eps_l}\ssminus {{\mathcal W}}$ with $\eps_l\to 0+$. Passing if necessary to a subsequence, we have $u=\lim u_l \in \Sigma\ssminus {{\mathcal W}}$ by the compactness theorem, and hence $u\not\in\Sigma_{\min}$. This is impossible, for $A_K(u)=\lim A_K(u_l)=\min A_K|_{\Sigma}$ and thus $u\in \Sigma_{\min}$. Let now ${{\mathcal U}}$ be a neighborhood of $C_0$. We need to show that ${{\mathcal U}}_\eps\subset {{\mathcal U}}$, when $\eps>0$ is small. Assume that this is not the case: the sets ${{\mathcal U}}_{\eps_i}$ are not entirely contained in ${{\mathcal U}}$ for some sequence $\eps_i\to 0+$. (Since the family ${{\mathcal U}}_{\eps}$ is nested, this is true for all ${{\mathcal U}}_{\eps}$, but we prefer to work with a sequence.) Therefore, the intersection $\bigcap_i \overline{{{\mathcal U}}}_{\eps_i}\subset \Sigma_{\min}$ is connected since ${{\mathcal U}}_{\eps_i}$ are connected, contains $C_0$, and a point of $\Sigma_{\min}\ssminus C$. This is impossible due to (f). Now we are in a position to finish the proof of Proposition \[prop:orbits-deg\]. Utilizing (g), pick $\eps>0$ so small that $\ev({{\mathcal U}}_\eps)\subset V$. Then every fixed point of $\Phi$ in ${{\mathcal U}}_\eps$ must belong to $C_0$. (Otherwise, there would be a one-periodic orbit with action in $[\max K, \infty)$, other than a point of $M$, entering $V$.) Thus $\omega^+(u)\in C_0$ for any $u\in {{\mathcal U}}_\eps$. Observe now that ${{\mathcal U}}_\eps \ssminus C_0\neq \emptyset$, since $\Sigma$ is connected and $C_0$ is closed. Let $u\in {{\mathcal U}}_\eps\ssminus C_0$. Then $u$ is a non-trivial, anti-gradient trajectory partially asymptotic to a point in $M$ at $\infty$. Therefore, $$\max K=\min_{\Sigma} A_K < A_K(u).$$ As a consequence, for any $\gamma\in \omega^-(u)$ we have $$0<A_K(\gamma)-\max K\leq\parallel H\parallel,$$ where the second inequality follows from (e). This completes the proof of and thus the proof of the proposition. Proposition \[prop:orbits-deg\] has a counter-part asserting the existence of homotopy connecting trajectories “transferring” action selectors; cf. [@cgk; @ke:new]. This is a much more standard result and we treat it in lesser detail. \[prop:orbits-deg2\] Let $U$ be an open subset of $W$ displaced by a Hamiltonian diffeomorphism $\varphi_H$. Furthermore, let $K\geq 0$ be a Hamiltonian supported in $U$ and let $f\geq 0$ be a $C^2$-small autonomous Hamiltonian such that $f\leq K$. Consider a monotone decreasing homotopy from $K$ to $f$. Then there exists a homotopy trajectory $u$ partially asymptotic to a point $p\in{{\mathcal P}}(f)$ at $\infty$ and to $\gamma\in{{\mathcal P}}(K)$ at $-\infty$ and such that - $f$ attaints its maximum at $p$, - $A_K(\gamma)\leq\sigma(K)$ and $E(u)\leq A_K(\gamma)-f(p)\leq \| H\|$. When the Hamiltonians $K$ and $f$ are non-degenerate and the homotopy is regular, the inequality $A_K(\gamma)\leq\sigma(K)$ can be replaced by equality $A_K(\gamma)=\sigma(K)$. First note that similarly to the proof of Proposition \[prop:orbits-nondeg\] we may assume without loss of generality that $H$ is compactly supported and $\min H_t=0$. In particular, $H\geq 0$ and $\| H\|=E^+(H)$. Furthermore, the lower bound $E(u)\leq A_K(\gamma)-f(p)$ follows immediately from . When $K$ and $f$ are non-degenerate and the homotopy is regular the assertion is well known; see, e.g., the proof of Proposition \[prop:orbits-nondeg\] or [@FS; @gi:alan; @gu:new; @ke:new; @KL; @Oh:chain; @schwarz; @Vi:gen] to mention just a few sources where similar results have been proved. Moreover, in this case we have $A_K(\gamma)=\sigma(K)$. Indeed, in the notation of the proof of Proposition \[prop:orbits-nondeg\], $\gamma$ is an orbit in the chain $y$, a carrier of the action selector, such that $A_K(\gamma)=\sigma(K)$. To deal with the general case, we argue as in the proof of Proposition \[prop:orbits-deg\] and approximate $f$ and $K$ by non-degenerate Hamiltonians $f_l\to f$ and $K_l\to K$ meeting the requirements of Proposition \[prop:orbits-deg2\]. We also approximate the homotopy $K^s$ from $K$ to $f$ by regular monotone decreasing homotopies $K^s_l$ from $K_l$ to $f_l$. Furthermore, we may assume that $\max f_l=\max f$. Applying the non-degenerate case of the proposition to $K_l$ and $f_l$, we obtain a critical point $p_l$, an orbit $\gamma_l$, and a homotopy connecting trajectory $u_l$. Since $E(u_l)\leq \| H\|$, the compactness theorem implies that the sequence $u_l$ contains a converging subsequence. Passing to this subsequence and taking the limit, we obtain a homotopy connecting trajectory $u$ from $K$ to $f$ such that $E(u)\leq \| H\|$. Then, by compactness again, $u$ is partially asymptotic to $p\in{{\mathcal P}}_f$ at $\infty$ and to $\gamma\in{{\mathcal P}}_K$ at $-\infty$. Note that although we can assume that $\gamma_l$ converges to an orbit of $K$ due to the Arzela–Ascoli theorem, we cannot claim that $\gamma_l\to\gamma$ and hence cannot conclude that $\sigma(K)=A_K(\gamma)$. However, $$\label{eq:f} f(p)=A_f(p)\geq \max f = f_l(p_l)$$ and $$\label{eq:K} A_K(\gamma)\leq \lim_{l\to\infty} A_{K_l}(\gamma_l)=\sigma(K).$$ To prove , we argue as in the proof of (H0)–(H2) in Section \[sec:homotopy\]. Namely, $$\begin{aligned} A_K(\gamma)=\sup_s A_{K^s}(u(s)) &=& \sup_s\lim_{l\to\infty} A_{K^s_l}(u_l(s))\\ &\leq& \lim_{l\to\infty} \sup_s A_{K^s_l}(u_l(s))\\ &=& \lim_{l\to\infty} A_{K_l}(\gamma_l)\\ &=& \lim_{l\to\infty} \sigma(K_l)=\sigma(K),\end{aligned}$$ where the last equality follows from (S3), continuity of the action selector. The first inequality, , is established in a similar fashion and immediately implies that $f(p) =\max f$. Finally, from and , we infer that $$A_K(\gamma)-f(p)\leq \lim_{l\to\infty}\left(A_{K_l}(\gamma_l)-f_l(p_l)\right) \leq \| H\|,$$ which concludes the proof. Propositions \[prop:orbits-deg\] and \[prop:orbits-deg2\] (and their non-degenerate counterparts) represent two different Floer homological (broadly understood) approaches to proving the existence of (infinitely many) periodic orbits of Hamiltonians, either in the autonomous case (the Weinstein conjecture and the almost existence theorem) or for time-dependent Hamiltonians (the Conley conjecture). One approach comprises a class of methods that lead to “low-lying” orbits with action smaller than the displacement energy as in Proposition \[prop:orbits-deg2\]. These methods are utilized in, for instance, [@cgk; @FS; @fhw; @FGS; @gi:alan; @gu:new; @HZ; @Vi:gen; @vi:functors]. The second class of methods detects orbits lying above the action selector value as in Propositions \[prop:orbits-nondeg\] and \[prop:orbits-deg\]. In this class are the variational methods of [@HZ:cap], the Floer homological results of [@ke:new; @gg:new] and apparently some of the results utilizing Hofer’s geometry, e.g., [@Oh:chain; @schl]. It goes without saying that many methods do not fit into this crude classification. These include, for instance, the equivariant (Floer) homological methods (see, e.g., [@HZ:V; @vi; @vi:jdg]) and holomorphic curve methods, e.g., [@HV:spheres; @Lu; @Lu2; @LiTi]. Displacement energy for stable manifolds {#sec:prfs} ======================================== Let, as in Section \[sec:statements\], $M$ be a stable, closed, coisotropic submanifold of $W$. Recall from Section \[sec:ct\] that a neighborhood of $M$ in $W$ is identified with a neighborhood of $M$ in $M\times \R^k$ with the symplectic form . Using this identification, we denote by $U_r$, with $r>0$ sufficiently small, the neighborhood of $M$ in $W$ corresponding to $M\times D^k_r$, where $D^k_r$ is the ball of radius $r$. Recall also that by definition $\rho=(p_1^2+\cdots+p_k^2)/2$, where $(p_1,\ldots,p_k)$ are the coordinates on $\R^k$. (Thus $U_r=\{\rho<r^2/2\}$.) Let $K$ be a smooth function on $[0,r]$ such that - $K$ is monotone decreasing and $K\equiv 0$ near $r$; - all odd-order derivatives of $K$ at $0$ are zero, and $K''(0)<0$ is close to zero. Abusing notation, we also denote by $K$ the function on $W$ equal to $K(|p|)$ on $U_r$, where $|p|=\sqrt{2\rho}$, and extended to be identically zero outside $U_r$. Then the Hamiltonian $K$ satisfies the hypotheses of Proposition \[prop:orbits-deg\]. By Proposition \[prop:flow\], the Hamiltonian flow of $K$ on $U_r$ is a reparametrization of the leaf-wise geodesic flow on $M$. Outside $U_r$, the flow is the identity map. \[thm:main2\] Assume that a Hamiltonian diffeomorphism $\varphi_H$ displaces $M$ and that $r>0$ is sufficiently small so that, in particular, $\varphi_H$ also displaces $U_r$. Then there exists a constant $\Delta>0$, independent of $r$ and $K$, such that $K$ has a contractible periodic orbit $\gamma$ with $$\Delta \leq A_K(\gamma)-\max K\leq \parallel H \parallel,$$ provided that $\max K$ is large enough. The proof of the theorem relies on Proposition \[prop:orbits-deg\] and the following two lemmas, which are essentially contained in [@Bo1; @Bo2] and which will also be used in the proof of Theorem \[thm:main\]. For a closed curve $\eta\colon S^1\to M$ lying in a leaf of the foliation ${{\mathcal F}}$, set $$\delta(\eta)=\sum_{i=1}^k\left|\int_\eta\alpha_i\right|.$$ \[lemma:lower-bound\] There exists a constant $\delta_M>0$ such that $$\delta_M\leq \delta(\eta)$$ for all non-trivial closed geodesics $\eta$ of the leaf-wise metric $\alpha_1^2+\cdots+\alpha^2_k$ (see Proposition \[prop:flow\]). Note that since, by Proposition \[prop:flow\], the metric is leaf-wise flat we have $\alpha_i(\dot{\eta}(t))={{\mathit const}}$ for every leaf-wise geodesic $\eta$. Hence $$\left|\int_\eta\alpha_i\right|=\int_{S^1} |\alpha_i(\dot{\eta}(t))|\,dt.$$ As a consequence, $$\label{eq:length} \delta(\eta)\geq\length(\eta)$$ which follows immediately from the fact that $\sum_{i}|\alpha_i(v)|\geq 1$ for every unit vector $v$ tangent to ${{\mathcal F}}$. Assume that the required constant $\delta_M>0$ does not exist, i.e., $\delta(\eta_j)\to 0$ for some sequence of closed non-trivial leaf-wise geodesics $\eta_j$. Thus, we also have $\length(\eta_j)\to 0$. Passing if necessary to a subsequence, we conclude that all $\eta_j$ are contained in a small neighborhood in $M$. (Indeed extending the metric from ${{\mathcal F}}$ to $M$, we can view $\eta_j$ as a sequence of closed curves whose length converges to zero. Now it is clear that the sequence $\eta_j$ contains a subsequence lying in a neighborhood of arbitrarily small radius.) Without loss of generality we may assume that this neighborhood is foliated. Then, when $j$ is large enough, $\eta_j$ is contained in a small ball in the leaf $F\supset \eta_j(S^1)$ of ${{\mathcal F}}$. In particular $\eta_j$ is contractible in $F$. This is impossible since the metric is leaf-wise flat. When $M$ has restricted contact type, Lemma \[lemma:lower-bound\] implies that the “area spectrum” of $M$ is separated from zero by $\delta_M/k$, for $\delta(\eta)=k|A(\eta)|$. Fix a neighborhood $V=U_R$, for some $R>0$. In what follows, we will always assume that $0<r<R/2$ and that $U_r$ is displaced by $\varphi_H$. \[lemma:low-bound2\] There exists a constant $c_V>0$, depending on $V$ but not on $K$ or $r>0$, such that for any Floer anti-gradient trajectory $u$ for $K$ partially asymptotic to a non-trivial one-periodic orbit $\gamma$ at $-\infty$ and to a point of $M$ at $\infty$, we have $$c_V \delta (\pi(\gamma))\leq A_K(\gamma)-\max K.$$ Note that the right hand side of this inequality is the energy $E(u)$. Let $f$ be a non-negative, smooth, decreasing function on $[0,R]$ identically equal to $1$ on $[0,R/2]$ and vanishing near $R$. Abusing notation, we also denote by $f$ the function on $W$ equal to $f(|p|)$ on $V=U_R$, where $|p|=\sqrt{2\rho}$ as above, and extended to be identically zero outside $V$. Following [@Bo2], set $\beta_i=f\pi^*\alpha_i$. This is a smooth one-form on $W$. A feature of the form $\beta_i$, important in what follows, is that $$\label{eq:0} i_{X_K}d\beta_i=0.$$ To prove this, we first note that trivially holds outside $U_r$. On the other hand, $f\equiv 1$ on $U_r$ since $r<R/2$. Thus on $U_r$ we have $\beta_i=\pi^*\alpha_i$ and, point-wise, $$i_{X_K}d\beta_i=i_{X_K}\pi^*d\alpha_i=K'(\rho)i_{\pi_*X_\rho}d\alpha_i=0.$$ The last equality follows from the fact that $\pi_*X_\rho$ is tangent to ${{\mathcal F}}$ by Proposition \[prop:flow\] and that $T{{\mathcal F}}\subset \ker d\alpha_i$ since $M$ is stable. Set $c_i=\parallel d\beta_i\parallel_{C^0}>0$ so that $$|d\beta_i(X,Y)|\leq c_i\parallel X\parallel\cdot \parallel Y\parallel,$$ for any two tangent vectors $X$ and $Y$. Here, on the right hand side, the norm is taken with respect an arbitrary metric compatible with $\omega$ as in Section \[sec:statement\]. (We emphasize that, since the argument relies only on the results of Section \[sec:deg\], which hold for general Hamiltonians and metrics, the metric need not meet any regularity requirements for Floer anti-gradient trajectories; cf. Remark \[rmk:orbits-deg\].) It is clear that $c_i$ is independent of $K$. We claim that $${\label{eq:energy-bound}} A_K(\gamma)-\max K=E(u) \geq c_i^{-1}\left|\int_{\pi(\gamma)}\alpha_i\right|.$$ The assertion of the lemma immediately follows from by adding up these inequalities for $i=1,\ldots,k$ and setting $c_V=\min c_i^{-1}/k$. To prove , fix $s_j^\pm\to \pm\infty$ such that $u(s_j^+)$ converges to a point of $M$ and $u(s_j^-)$ converges to $\gamma$ in $C^\infty(S^1,W)$. Then utilizing the definition of $c_i$ and in the last step, we have $$\begin{aligned} E(u) &=& \int_{\R\times S^1} \left\| \frac{\p u}{\p s}\right\| \cdot \left\| \frac{\p u}{\p t}-X_K\right\| \,ds\,dt \\ &\geq& c_i^{-1} \int_{\R\times S^1} \left|d\beta_i\left(\frac{\p u}{\p s}, \frac{\p u}{\p t}-X_K\right)\right| \,ds\,dt \\ &\geq& c_i^{-1} \lim_{j\to\infty} \left|\int_{s_j^-}^{s_j^+}\int_{S^1} d\beta_i\left(\frac{\p u}{\p s}, \frac{\p u}{\p t}-X_K\right) \,dt\,ds\right| \\ &=& c_i^{-1} \lim_{j\to\infty} \left|\int_{s_j^-}^{s_j^+}\int_{S^1} d\beta_i\left(\frac{\p u}{\p s}, \frac{\p u}{\p t}\right) \,dt\,ds\right|.\end{aligned}$$ Note that at this stage we still do not know if the limit in question exists. However, by applying Stokes’ formula, we see that $$\left|\int_{s_j^-}^{s_j^+}\int_{S^1} d\beta_i\left(\frac{\p u}{\p s}, \frac{\p u}{\p t}\right) \,dt\,ds\right| = \left|\int_{u(s_j^+)}\beta_i-\int_{u(s_j^-)}\beta_i\right| \longrightarrow \left|\int_{\gamma}\beta_i\right|$$ as $j\to\infty$. Furthermore, recall that $\gamma$ is contained in $U_r$ and $f|_{U_r}\equiv 1$. Thus, $$\left|\int_{\gamma}\beta_i\right|=\left|\int_{\gamma}\pi^*\alpha_i\right| =\left|\int_{\pi(\gamma)}\alpha_i\right|,$$ which completes the proof of and of the lemma. To finish the proof of the theorem, we apply Proposition \[prop:orbits-deg\] to the Hamiltonian $K$. Next, applying Lemma \[lemma:low-bound2\] to $u$ and $\gamma$ whose existence is guaranteed by this proposition, we have $${\label{eq:main-ineq}} \| H\|\geq A_K(\gamma)-\max K\geq c_V\delta(\pi(\gamma))\geq c_V \delta_M =: \Delta>0,$$ where the last inequality follows from Lemma \[lemma:lower-bound\] and the fact that $\pi(\gamma)$ is, up to parametrization and orientation, a closed non-trivial geodesic of the leaf-wise geodesic flow on ${{\mathcal F}}$. Let $K$ be as above and let $\max K$ be large enough. Assertion (i) follows immediately from Theorem \[thm:main2\]. To prove (ii), we argue as in [@Bo2]. Pick $\gamma$ as in the proof of Theorem \[thm:main2\]. This is a non-trivial periodic orbit and, since $K$ attains its maximum along $M$, we see from that $$\label{eq:A} A(\gamma)\geq A_K(\gamma)-\max K \geq c_V\delta(\pi(\gamma)).$$ Then, utilizing the normal form and setting $\eta=\pi(\gamma)$, we have $$\begin{aligned} A(\eta)&=& A(\gamma)-\int_\gamma\sum p_i\alpha_i\\ &\geq& A(\gamma)-r\sum\left|\int_{\eta}\alpha_i\right|\\ &\geq& c_V\delta(\eta)-r\delta(\eta)\\ &=& (c_V-r)\delta(\eta)\\ &\geq& (c_V-r)\delta_M>0,\end{aligned}$$ provided that $r>0$ is small enough. This proves (ii) for $\eta=\pi(\gamma)$. Moreover, we have also established the *a priori* lower bound $$\label{eq:lb} A(\eta)\geq (c_V-r)\delta_M .$$ It remains to prove (iii). To this end, we first need to establish a general property of the metric $\rho$, which holds for arbitrary stable coisotropic submanifolds. Recall that the length spectrum of a (leaf-wise) metric on a foliated manifold $(M,{{\mathcal F}})$ is the collection of lengths of all closed leaf-wise geodesics in $M$. In contrast with the length spectrum of a metric on $M$, the length spectrum of a leaf-wise metric need not in general be nowhere dense, even if the metric is leaf-wise flat. However, as our next observation shows, this is the case for the metric $\rho$. \[lemma:nowheredense\] Assume that $M$ is stable. The length spectrum of $\rho$ is closed, has zero measure, and is, therefore, nowhere dense. The assertion that the length spectrum is closed holds obviously for any metric. To show that the spectrum has zero measure, consider a level $\{\rho={{\mathit const}}\}\subset M\times\R^k$ and recall that by Proposition \[prop:flow\] closed geodesics of $\rho$ are projections to $M$ of closed characteristics on this level. Furthermore, again by Proposition \[prop:flow\], the length of a geodesic is equal to the integral of $\lambda=\sum p_i\alpha_i$ over the corresponding closed characteristics. (Strictly speaking, the equality holds only up to a factor depending only on the level, and we chose ${{\mathit const}}$ so that this factor is equal to one.) The characteristic foliation on the level is tangent to the distribution $\ker d\lambda$ as is easy to see from . It follows that the integrals of $\lambda$ over closed characteristics form a zero measure set. (The distribution $\ker d\lambda$ need not have constant rank, but the standard argument still applies.) Let now $\gamma$ be as above and $\eta=\pi(\gamma)$. Our next goal is to obtain an upper bound on $A(\gamma)$ and $A(\eta)$. To this end, we need to impose an extra requirement on $K$ guaranteeing that $K(\gamma)$ is close to $\max K$. Namely, fix $r_-$ and $r_+$ such that $0<r_-<r_+<r$ and pick a sufficiently small constant $\eps>0$. We can chose $K$ so that on $U_r$, we have - $\max K-K<\eps$ on $[0,r_-]$ and $K<\eps$ on $[r_+,r]$; - on $[r_-,r_+]$ the Hamiltonian $K$, thought of as function of $|p|$, has constant slope lying outside the length spectrum of the metric $\rho$; - $K$ has a sufficiently large variation over the constant slope range, i.e., $K(r_-)-K(r_+)> C\cdot\| H\|$, where $C$ is a constant, to be specified later, depending only on $V$. In particular, these conditions ensure that the only non-trivial periodic orbits of $K$ occur within the shells $(0,r_-)$ and $(r_+,r)$. Furthermore, we claim that $\gamma$ lies in the shell $(0,r_-)$, and hence $$\label{eq:nearmax} \max K-K(\gamma)<\eps.$$ Indeed, first observe that since $M$ has restricted contact type, we have $$A(\eta)=\int_\eta\alpha_i\text{ for } i=1,\ldots,k.$$ Thus, by , $$|A(\eta)|\leq c_i\| H\|\text{ for } i=1,\ldots,k$$ and, as a consequence, $$\begin{aligned} |A(\gamma)| &=& \left|A(\eta)+\int_\gamma \sum_i p_i\alpha_i\right|\\ &\leq& |A(\eta)|+r\sum_i\left|\int_\eta\alpha_i\right| \\ &=& (1+kr)|A(\eta)|\\ &\leq& C\cdot\|H\|,\end{aligned}$$ where, for instance, $C=(1+kr)c_1$. On the other hand, if $\gamma$ were in the shell $(r_+,r)$, we would have $$\max K- K(\gamma)> K(r_-)-K(r_+)> C\cdot\| H\|,$$ and hence, $$A(\gamma)\geq\max K- K(\gamma)> C\cdot\| H\|.$$ Thus $\gamma$ is indeed in the shell $(0,r_-)$ and holds. Next note that $$A(\eta)=\big(A_K(\gamma)-\max K\big) +\big(\max K-K(\gamma)\big) -\int_\gamma\sum p_i\alpha_i.$$ Here the first term in bounded from above by $\| H \|$ due to Proposition \[prop:orbits-deg\], the second term is bounded from above by $\eps$ due to , and the last term is bounded from above by $r\delta(\eta)\leq r c_V^{-1} \| H\|$ according to . As a consequence, $$\label{eq:hplast} (c_V-r)\delta_M\leq A(\eta)\leq\| H\| +\eps + rc_V^{-1}\| H\|,$$ where the first inequality is . Note that we also have an *a priori* upper bound on the length of $\eta$, $$\length(\eta)\leq \delta(\eta)\leq c_V^{-1}\| H\|,$$ resulting from and . Consider a sequence of Hamiltonians $K_i$ as above with $r_i\to 0$ and $\eps_i\to 0$. Then by the Arzela–Ascoli theorem the geodesics $\eta_i$ converge to a geodesic $\eta$ and, passing to the limit in , we see that $$0<c_V\delta_M\leq A(\eta)\leq\| H\|,$$ which concludes the proof. [\[rmk:Dr\]]{} The first assertion of Theorem \[thm:main\] can also be established by making use of Proposition \[prop:orbits-deg2\] instead of Proposition \[prop:orbits-deg\], although in a somewhat less direct way. Let $K$ and $r$ be as in Theorem \[thm:main2\] with $\max K>\| H\|$ and let $\eps>0$ be so small that $f=\eps K$ is $C^2$-small. Denote by $u$ and $\gamma$ the homotopy connecting trajectory and the periodic orbit from Proposition \[prop:orbits-deg2\] for a *linear* homotopy from $K$ to $f$. Since $\max K>\| H\|> \sigma(K)$, we conclude that the orbit $\gamma$ is non-trivial. It is not hard to see that the proof of Lemma \[lemma:low-bound2\] goes through for $u$. (Note that the assumptions that $f=\eps K$ and that the homotopy is linear are essential to make sure that holds for every Hamiltonian $K^s$ in the homotopy.) Hence, as in the proof of Theorem \[thm:main2\], we have $$\|H\|\geq E(u)\geq c_V\delta(\pi(\gamma))\geq c_V\delta_M,$$ which proves assertion (i). Even though this approach does not lead to assertion (ii) in its full generality, it does imply that ${{\mathcal F}}$ has a leaf-wise non-trivial closed geodesic, contractible in $W$. This is an analogue of the existence result for closed characteristics on closed, stable or contact type hypersurfaces, [@HZ; @vi]; see also the survey [@gi:alan] for a discussion of more recent results. Finally note that passing to the limit as $f\to 0+$, we infer from Proposition \[prop:orbits-deg2\] that $\sigma(K)\geq c_V\delta_M$. As a consequence, $\chom(U)\geq c_V\delta_M$ for any neighborhood $U$ of $M$. Thus $\chom(M)\geq c_V\delta_M>0$, whenever $M$ is stable and displaceable; cf. [@Dr]. It suffices to show that the levels $M_a$ carrying an orbit $\gamma$ with the required properties exist arbitrarily close to $M$. Set $K=f(K_1,\ldots,K_k)$, where $f\colon\R^k\to\R$ is a bump-function supported in a small neighborhood of the origin in $\R^k$ and such that $\max f$ is large enough. Since the support of $f$ is small, we may assume that ${\operatorname{supp}}K$ is displaceable and all $a\in {\operatorname{supp}}f$ are regular values of $\vec{K}$. By Proposition \[prop:orbits-deg\], the flow of $K$ has a contractible in $W$ one-periodic orbit $\gamma$ with $A(\gamma)>0$; see, e.g., Remark \[rmk:orbits-deg\]. Furthermore, the Hamiltonian $K$ Poisson–commutes with all $K_i$, and thus $\gamma$ is tangent to a regular level $M_a$. Leaf-wise intersections for hypersurfaces {#sec:leafwise} ========================================= The goal of this section is to prove the leaf-wise intersection property (Theorem \[thm:leafwise\]) for hypersurfaces of restricted contact type in subcritical Stein manifolds. In fact the theorem holds for a slightly broader class of ambient manifolds than subcritical Stein. Namely, let $W$ be an exact symplectically aspherical manifold convex in the sense of [@FS] and let as above $M$ be a closed hypersurface of restricted contact type in $W$ bounding a domain $U$. Assume that $H$ is a compactly supported Hamiltonian on $W$ such that $$\|H\|<\chom(U)$$ and $$\label{eq:upper-chom} \chom\big(\bar{U}\cup{\operatorname{supp}}(H)\big)<\infty.$$ Then Theorem \[thm:leafwise\] (in a slightly generalized form) asserts that $\varphi(F)\cap F\neq\emptyset$ for some leaf $F$ of the characteristic foliation of $M$ and $\varphi=\varphi_H$. It is worth mentioning that the requirement is quite restrictive. (Note that implies that $\chom(U)<\infty$.) The reason that holds if $W$ is a subcritical Stein manifold is that every compact set in $W$ is displaceable; see, e.g., [@BC]. However, fails, as can be seen from the results of [@vi:functors], when $U$ is a tubular neighborhood of the zero section in a cotangent bundle. Essentially the only case where can be verified is where $\bar{U}\cup{\operatorname{supp}}(H)$ is displaceable in $W$. Hereafter, we assume that $W$ and $M$ are as above. In this case, the action selector $\sigma$ from Section \[sec:sel-open\] can also be defined as in [@FS] for all compactly supported Hamiltonians (not necessarily positive) and has, in addition to (S0)–(S4), a number of other properties. For instance, $\sigma(H)=0$, whenever $H\leq 0$, and $\sigma$ is sub-additive, i.e., $\sigma(H\# K)\leq \sigma(H)+\sigma(K)$; see [@FS]. Furthermore, $\sigma(K)$ depends only on the time-one flow $\varphi_K$ and we will also use the notation $\sigma(\varphi_K)$. First note that as in the proof of Proposition \[prop:orbits-nondeg\] we may require that $\min H_t=0$ for all $t$ and hence $\| H\|= E^+(H)$ and $H\geq 0$. (This can be achieved, keeping , by replacing $H$ by the Hamiltonian $f\cdot (H-\min H)$, where $f$ is a cut-off function identically equal to one near $\bar{U}\cup{\operatorname{supp}}(H)$.) Furthermore, without loss of generality we may also assume that the fixed point set $\Fix(\varphi)$ does not meet $M$, for otherwise the assertion is obvious. Then $\varphi$ has no fixed points near $M$. Fix a global primitive $\lambda$ of $\omega$, restricting to a contact form on $M$. Let $U_\eps=M\times [-\eps,0]$ be a narrow shell inside $U$, containing no points of $\Fix(\varphi)$. We choose the projection $U_\eps\to [-\eps,0]$ so that the Hamiltonian flow of this projection (thought of as a function on $U_\eps$) is exactly equal to the Reeb flow of $\lambda|_{M_\tau}$, where $M_\tau=M\times\{\tau\}$, for all $\tau\in [-\eps,0]$. The projection $U_\eps\to M$ is chosen so that the characteristic foliation on $M_\tau$ projects to the characteristic foliation on $M$. (It is well known that such a shell exists, when $M$ has contact type.) Let $K\geq 0$ be a non-negative function which is equal to zero on $W\ssminus U$ and to $\max K>0$ on $U\ssminus U_\eps$, and which is a monotone decreasing function of $\tau$ on $U_\eps$. Set $\psi_s=\varphi\varphi_K^s=\varphi\varphi_{sK}$ for $s\in [0,\,1]$. Then we have the following disjoint union decomposition $$\Fix(\psi_s)=\Fix(\varphi)\sqcup Z_s, \text{ where } Z_s=\{x\in U_\eps \mid \varphi^{-1}(x)=\varphi_{sK}(x)\},$$ as is easy to see recalling that $\Fix(\varphi)\cap U_\eps=\emptyset$. Since $H\geq 0$, the Hamiltonian generating the map $\varphi^{-1}$ is non-positive and hence $\sigma(\varphi^{-1})=0$. Then, by conjugation invariance and sub-additivity of the action selector $\sigma$ (see [@FS] and also [@schwarz; @Vi:gen]), we have $$\label{eq:long-sigma} \sigma(\varphi_{sK})=\sigma(\varphi\varphi_{sK}\varphi^{-1}) \leq \sigma(\varphi\varphi_{sK}) + \sigma(\varphi^{-1})= \sigma(\varphi\varphi_{sK})=\sigma(\psi_s).$$ From the monotonicity property, (S0), of $\sigma$ we infer that $\chom(U)=\sup\sigma(\varphi_K)$, where the supremum is taken over all $K$ as above. Pick $K$ such that $\sigma(\varphi_K)$ is close to $\chom(U)$: $$\label{eq:upperH} E^+(H)<\sigma(\varphi_K)\leq \chom(U).$$ As $s$ varies through the interval $[0,\,1]$, the action selector $\sigma(\psi_s)\in {{\mathcal S}}(\psi_s)$ changes from $$\sigma(\psi_0)=\sigma(\varphi)\leq E^+(H)$$ to $$\sigma(\psi_1)\geq \sigma(\varphi_K)> E^+(H),$$ where the first inequality follows form and the second one from . By continuity of the action selector, we see that ${{\mathcal S}}(\psi_s)$ cannot be independent of $s$. Furthermore, $\psi_0=\varphi$, and thus $\Fix(\psi_0)=\Fix(\varphi)$ and $Z_0=\emptyset$. As a consequence, the part $Z_s$ of $\Fix(\psi_s)$ must be non-empty for some $s_0\in (0,\,1]$ and, moreover, $\sigma(\psi_{s_0})$ is the action value of $\psi_{s_0}$ on $x\in Z_{s_0}$. As in the proof of Proposition \[prop:orbits-nondeg\], we may reparametrize the Hamiltonians $H$ and $K$ (making $K$ now time-dependent), without altering the time-one maps, the action spectra, the Hofer norms, and the action selectors so that $K_t \equiv 0$ when $t\in [1/2,\, 1]$ and $H_t \equiv 0$ when $t\in [0,\,1/2]$. From now on, we assume that $H$ and $K$ have this property. We will also denote $\varphi^t_H$ by $\varphi^t$. Consider the orbit $\gamma(t)$ through $x=\gamma(0)$ of the time-dependent flow $\psi_{s_0}^t=\varphi^t\varphi^t_{s_0K}$. Let $G=H\#(s_0 K)$ be the Hamiltonian generating this flow. Due to the above reparametrizations of $H$ and $K$, the orbit $\gamma$ is comprised of two parts: $\gamma_1(t)=\varphi^t_{s_0K}(x)$ ending at $y=\varphi_{s_0K}(x)$ and $\gamma_2(t)=\varphi^t(y)$ ending at $x$. Note that $x$ and $y$ lie on the same Reeb orbit (i.e., a leaf of characteristic foliation) on some level $M_\tau$ and $\varphi(y)=x$. Furthermore, $$\label{eq:gamma} \sigma(\psi_{s_0})=A_G(\gamma) =-\int_{\gamma_1}\lambda+ \int_0^{1/2} s_0K_t(\gamma_1(t))\,dt + A_H(\gamma_2),$$ where $$A_H(\gamma_2)= -\int_{\gamma_2}\lambda+ \int_{1/2}^1 H_t(\gamma_2(t))\,dt.$$ The term $T=-\int_{\gamma_1}\lambda$ is the time required for the Reeb flow on $M_\tau$ to move $x$ to $y$ and our next goal is to establish an upper bound on $T$ independent of $K$. Consider the function $f(z)$ equal to the action of $H$ on the orbit $\varphi^t(z)$, $t\in [0,\,1]$, defined using the primitive $\lambda$. (For instance, $f(y)=A_H(\gamma_2)$.) This function is independent of $K$. Clearly $f$ is a compactly supported function and $C=-\min f$ is also independent of $K$ and $A_H(\gamma_2)\geq -C$. Furthermore, the middle term in is non-negative, for $K\geq 0$. As a consequence, $$\sigma(\psi_{s_0})\geq T-C.$$ Finally note that the Hamiltonian $G$ generating $\psi_{s_0}$ is supported in $U\cup{\operatorname{supp}}(H)$. Thus $\sigma(\psi_{s_0})\leq\chom\big(U\cup{\operatorname{supp}}(H)\big)$. Therefore, $$T\leq \chom\big(U\cup{\operatorname{supp}}(H)\big)+ C\leq e\big(U\cup{\operatorname{supp}}(H)\big)+ C <\infty,$$ where the upper bounds on the right hand side are clearly independent of $K$. To finish the proof, consider a sequence of Hamiltonians $K_i$ such as $K$, non-constant on a more and more narrow range of $\tau$ in $[-\eps,\,0]$ eventually converging to zero. For each $K_i$ we have a pair of points $x_i$ and $y_i$ lying on the same Reeb orbit on some $M_{\tau_i}$ with $\tau_i\to 0$ and such that $\varphi(y_i)=x_i$. Furthermore, the Reeb flow requires time $T_i\leq \chom\big(U\cup{\operatorname{supp}}(H)\big)+ C$ to move $x_i$ to $y_i$. Applying the Arzela–Ascoli theorem and passing if necessary to a subsequence, we obtain points $x=\lim x_i$ and $y=\lim y_i$ on $M$ lying on the same Reeb orbit and such that $\varphi(y)=x$. This completes the proof of the theorem. The proof of Theorem \[thm:leafwise\] also yields the upper bound $\chom\big(\bar{U}\cup{\operatorname{supp}}(H)\big)+ C$ for the “Reeb distance” from $x$ to $y$. In the next example we show that Theorem \[thm:leafwise\], as stated with the upper bound on $\|H\|$, does not extend to hypersurfaces in $\R^{2n}$ that do not have contact type. To be more precise, we construct a Hamiltonian flow $\varphi^t$ on $\R^{2n}$ and a sequence of hypersurfaces $M_i$, $C^0$-converging to $S^{2n-1}$, such that $M_i$ and $\varphi^{t_i}(M_i)$ have no leaf-wise intersections for some sequence of times $t_i\to 0+$. \[ex:leafwise-fail\] Let $S^{2n-1}$ be the unit sphere in $\R^{2n}$ and let $\varphi^t$ be the Hamiltonian flow of $H=f\cdot p_1$, where $(p_1,q_1,\ldots,p_n,q_n)$ are the standard coordinates on $\R^{2n}$ and $f$ is a cut-off function equal to one near $S^{2n-1}$. For $t>0$ small, the only leaf-wise intersections of $S^{2n-1}$ and $\varphi^t(S^{2n-1})$ are two points $x_1$ and $x_2$ on the unit circle $S$ in the $(p_1,q_1)$-plane. (The points $\varphi^t(x_1)$ and $\varphi^t(x_2)$ are the intersections of $S$ and the transported circle $S+(t,0)$.) Let us now insert two symplectic plugs into $S^{2n-1}$ centered at points on $S$ between $x_1$ and $\varphi^t(x_1)$ and between $x_2$ and $\varphi^t(x_2)$; see, e.g., [@Ci; @Gi95; @gi:bayarea; @gg:ex2; @ke:example]. We choose the plugs so small and center them in such a way that they are displaced by $\varphi^t$. As a result, we obtain a new hypersurface $M$ that is $C^0$-close to $S^{2n-1}$, differs from $S^{2n-1}$ only within the plugs, and such that the leaf $S$ is broken into two leafs: one containing $x_1$ and $x_2$ and the other one containing $\varphi^t(x_1)$ and $\varphi^t(x_2)$. We claim that $M$ and $\varphi^t(M)$ have no leaf-wise intersections. Indeed, $x_1$ and $x_2$ are no longer leaf-wise intersections for $M$ and $\varphi^t(M)$, and since the plugs are displaced and due to the plug-symmetry conditions, no new leaf-wise intersections are created. Applying this construction to a sequence $t_i\to 0+$, we obtain a sequence of perturbations $M_i$ of $S^{2n-1}$, $C^0$-converging to $S^{2n-1}$, and such that $M_i$ and $\varphi^{t_i}(M_i)$ have no leaf-wise intersections. Note that $\varphi^{t_i}\stackrel{C^\infty}{\to}{{\mathit id}}$ while $\chom(U_i)\to \chom(U)=\pi$, where $U_i$ is the domain bounded by $M_i$ and $U$ is the unit ball. It is also clear that $\|H_i\|\to 0$, where $H_i=t_i H$ is a Hamiltonian generating $\varphi^{t_i}$. [LMc2]{} M. Audin, J. Lafontaine (Eds), *Holomorphic curves in symplectic geometry*, Progress in Mathematics, **117**, Birkhäuser Verlag, Basel, 1994. A. Banyaga, On fixed points of symplectic maps, *Invent. Math.*, **56** (1980), 215–229. P. Biran, K. Cieliebak, Lagrangian embeddings into subcritical Stein manifolds, *Israel J. Math.*, **127** (2002), 221–244. P. Biran, L. Polterovich, D. Salamon, Propagation in Hamiltonian dynamics and relative symplectic homology, *Duke Math. J.*, **119** (2003), 65–118. P. Bolle, Une condition de contact pour les sous-variétés coïsotropes d’une variété symplectique, *C. R. Acad. Sci. Paris, Série I*, **322** (1996), 83–86. P. Bolle, A contact condition for p-dimensional submanifolds of a symplectic manifold ($2\leq p\leq n$), *Math. Z.*, **227** (1998), 211–230. Y. Chekanov, Hofer’s symplectic energy and Lagrangian intersections, in *Contact and Symplectic Geometry*, C.B. Thomas (Ed.), INI Publications, Cambridge University Press, Cambridge, 1996, pp. 296–306. Y. Chekanov, Lagrangian intersections, symplectic energy, and areas of holomorphic curves, *Duke Math. Journal*, **95** (1998), 213–226. K. Cieliebak, Symplectic boundaries: creating and destroying closed characteristics, *Geom. Funct. Anal.*, **7** (1997), 269–321. K. Cieliebak, V. Ginzburg, E. Kerman, Symplectic homology and periodic orbits near symplectic submanifolds, *Comment. Math. Helv.*, **79** (2004), 554–581. C.C. Conley, E. Zehnder, Birkhoff–Lewis fixed point theorem and a conjecture of V.I. Arnold, *Invent. Math.*, **73** (1983), 33–49. D. Dragnev, Symplectic rigidity, symplectic fixed points and global perturbations of Hamiltonian systems, Preprint 2005, math.SG/0512109. I. Ekeland, H. Hofer, Two symplectic fixed-point theorems with applications to Hamiltonian dynamics, *J. Math. Pures Appl.* **68** (1989), 467–489 (1990). Y. Eliashberg, S.S. Kim, L. Polterovich, Geometry of contact transformations and domains: orderability vs. squeezing, Preprint 2005, math.SG/0511658. M. Entov, K-area, Hofer metric and geometry of conjugacy classes in Lie groups, *Invent. Math.*, **146** (2001), 93–141. A. Floer, Morse theory for Lagrangian intersections. *J. Differential Geom.*, **28** (1988), 513–547. A. Floer, The unregularized gradient flow of the symplectic action, *Comm. Pure Appl. Math.*, **41** (1988), 775–813. A. Floer, Cuplength estimates on Lagrangian intersections, *Comm. Pure Appl.Math.*, **42** (1989), 335–356. A. Floer, Witten’s complex and infinite-dimensional Morse theory, *J. Differential Geom.*, **30** (1989), 207–221. A. Floer, Symplectic fixed points and holomorphic spheres, *Comm. Math. Phys.*, **120** (1989), 575–611. A. Floer, H. Hofer, Symplectic homology, I. Open sets in ${\mathbb C}^n$, *Math. Z.*, **215** (1994), 37–88. A. Floer, H. Hofer, D. Salamon, Transversality in elliptic Morse theory for the symplectic action, *Duke Math. J.*, **80** (1995), 251–292. A. Floer, H. Hofer, K. Wysocki, Applications of symplectic homology, I. *Math. Z.*, **217** (1994), 577–606. U. Frauenfelder, V. Ginzburg, F. Schlenk, Energy capacity inequalities via an action selector, in *Geometry, Spectral Theory, and Dynamics; Proceedings in Memory of Robert Brooks*, Eds.: M. Entov et al, Contemporary Mathematics, vol. 387, AMS, 2005; pp.129–152. U. Frauenfelder, F. Schlenk, Hamiltonian dynamics on convex symplectic manifolds, Preprint 2003, math.SG/0303282. To appear in *Israel J. Math.* K. Fukaya, K. Ono, Arnold conjecture and Gromov-Witten invariant, *Topology*, **38** (1999), 933–1048. V.L. Ginzburg, An embedding $S^{2n-1}\to {\mathbb R}^{2n}$, $2n-1\geq 7$, whose Hamiltonian flow has no periodic trajectories, *IMRN*, 1995, no. 2, 83-98. V.L. Ginzburg, Hamiltonian dynamical systems without periodic orbits, in *Northern California Symplectic Geometry Seminar*, 35–48, Amer. Math. Soc. Transl. Ser. 2, vol. 196, Amer. Math. Soc., Providence, RI, 1999. V.L. Ginzburg, The Weinstein conjecture and the theorems of nearby and almost existence, in *The Breadth of Symplectic and Poisson Geometry. Festschrift in Honor of Alan Weinstein*; J.E. Marsden and T.S. Ratiu (Eds.), Birkhäuser, 2005, pp. 139–172. V.L. Ginzburg, B.Z. Gürel, A $C^2$-smooth counterexample to the Hamiltonian Seifert conjecture in $R^4$, *Ann. Math.*, **158** (2003), 953–976. V.L. Ginzburg, B.Z. Gürel, Relative Hofer–Zehnder capacity and periodic orbits in twisted cotangent bundles, *Duke Math. J.*, **123** (2004), 1–47. M. Gromov, Pseudoholomorphic curves in symplectic manifolds, *Invent. Math.*, **82** (1985), 307–347. B.Z. Gürel, Totally non-coisotropic displacement and its applications to Hamiltonian dynamics, Preprint 2006. H. Hofer, On the topological properties of symplectic maps, *Proc. Roy. Soc. Edinburgh Sect. A*, **115** (1990), 25–38. H. Hofer, C. Viterbo, The Weinstein conjecture in the presence of holomorphic spheres, *Comm. Pure Appl. Math.*, **45** (1992), 583–622. H. Hofer, E. Zehnder, Periodic solutions on hypersurfaces and a result by C. Viterbo, *Invent. Math.* **90** (1987), 1–9. H. Hofer, E. Zehnder, A new capacity for symplectic manifolds, in *Analysis, et cetera*, P. Rabinowitz and E. Zehnder (Eds.), Academic Press, Boston, MA, 1990, pp. 405–427. H. Hofer, E. Zehnder, *Symplectic Invariants and Hamiltonian Dynamics*, Birkäuser, 1994. E. Kerman, New smooth counterexamples to the Hamiltonian Seifert conjecture, *J. Symplectic Geometry*, **1** (2002), 253–267. E. Kerman, Squeezing in Floer theory and refined Hofer-Zehnder capacities of sets near symplectic submanifolds, *Geom. Topol.*, **9** (2005) 1775–1834. E. Kerman, F. Lalonde, Length minimizing Hamiltonian paths for symplectically aspherical manifolds, *Ann. l’Institut Fourier*, **53** (2003), 1503–1526. F. Lalonde, D. McDuff, The geometry of symplectic energy, *Ann. Math.*, **141** (1995), 349–371. F. Lalonde, D. McDuff, Hofer’s $L^{\infty}$-geometry: energy and stability of Hamiltonian flows, part II, *Invent. Math.*, **122** (1995), 35–69. F. Laudenbach, J.-C. Sikorav, Persistance d’intersection avec la section nulle au cours d’une isotopie hamiltonienne dans un fibré cotangent, *Invent. Math.*, **82** (1985), 349-357. F. Laudenbach, J.-C. Sikorav, Hamiltonian disjunction and limits of Lagrangian submanifolds, *IMRN*, 1994, no. 4, 161–168. T. Linnemann. *Generalized Fixed Points on Strictly Convex Hypersurfaces*, ETH Dissertation, Zürich, 1998. G. Liu, G. Tian, Floer homology and Arnold conjecture, *J. Diff.Geom.*, **49** (1998), 1–74. G. Liu, G. Tian, Weinstein conjecture and GW invariants, *Commun. Contemp. Math.*, **2** (2000), 405–459. G.-C. Lu, The Weinstein conjecture on some symplectic manifolds containing the holomorphic spheres, *Kyushu. J. Math.*, **52** (1998), 331–351. G.-C. Lu, Finiteness of Hofer-Zehnder symplectic capacity of neighborhoods of symplectic submanifolds, Preprint 2005, math.SG/0510172. D. McDuff, D. Salamon, *J-holomorphic Curves and Symplectic Topology*, Colloquium publications, vol. 52, AMS, Providence, RI, 2004. D. McDuff, J. Slimowitz, Hofer–Zehnder capacity and length minimizing Hamiltonian paths, *Geom. Topol.*, **5** (2001), 799–830 (electronic). C.C. Moore, C. Schochet, *Global Analysis on Foliated Spaces*, Mathematical Sciences Research Institute Publications, 9. Springer-Verlag, New York, 1988. J. Moser, A fixed point theorem in symplectic geometry, *Acta Math.*, **141** (1978), 17–34. Y.-G. Oh, Floer cohomology of Lagrangian intersections and pseudoholomorphic disks, I, II, *Comm. Pure Appl. Math.*, **46** (1993), 949–993, 995–1012. Y.-G. Oh, Relative Floer and quantum cohomology and the symplectic topology of Lagrangian submanifolds. in *Contact and Symplectic Geometry*, C.B. Thomas (Ed.), INI Publications, Cambridge University Press, Cambridge, 1996, pp. 201–267. Y.-G. Oh, Gromov–Floer theory and disjunction energy of compact Lagrangian embeddings, *Math. Res. Lett.*, **4** (1997), 895–905. Y.-G. Oh, Chain level Floer theory and Hofer’s geometry of the Hamiltonian diffeomorphism group, *Asian J. Math.*, **6** (2002). 579–624. L. Polterovich, Symplectic displacement energy for Lagrangian submanifolds, *Ergodic Theory Dynam. Systems*, **13** (1993), 357–367. L. Polterovich, An obstacle to non-Lagrangian intersections, in *The Floer Memorial Volume*, 575–586, Progr. Math., **133**, Birkhäuser, Basel, 1995. L. Polterovich, *The Geometry of the Group of Symplectomorphisms*, Birkhäuser, 2001. M. Poźniak, Floer homology, Novikov rings and clean intersections, in *Northern California Symplectic Geometry Seminar*, 119–181, Amer. Math. Soc. Transl. Ser. 2, **196**, Amer. Math. Soc., Providence, RI, 1999. D.A. Salamon, Lectures on Floer homology, in *Symplectic Geometry and Topology*, Eds: Y. Eliashberg and L. Traynor, IAS/Park City Mathematics series, **7**, 1999, pp. 143–230. D. Salamon, E. Zehnder, Morse theory for periodic solutions of Hamiltonian systems and the Maslov index, *Comm. Pure Appl. Math.*,**45** (1992), 1303–1360. F. Schlenk, Applications of Hofer’s geometry to Hamiltonian dynamics, *Comment. Math. Helv.*, **81** (2006), 105–121. M. Schwarz, *Morse Homology*, Birkhäuser, 1993. M. Schwarz, On the action spectrum for closed symplectically aspherical manifolds, *Pacific J. Math.*, **193** (2000), 419–461. M. Struwe, Existence of periodic solutions of Hamiltonian systems on almost every energy surfaces, *Bol. Soc. Bras. Mat.*, **20** (1990), 49–58. C. Viterbo, A proof of Weinstein’s conjecture in ${\mathbb R}^{2n}$, *Ann. Inst. Poincaré, Anal. Non Linéaire*, **4** (1987), 337-356. C. Viterbo, Symplectic topology as the geometry of generating functions, *Math. Ann.*, **292** (1992), 685-710. C. Viterbo, Exact Lagrange submanifolds, periodic orbits and the cohomology of the free loop spaces, *J. Diff. Geom.*, **47** (1997), 420–468. C. Viterbo, Functors and computations in Floer cohomology, I, *Geom. Funct. Anal.*, **9** (1999), 985–1033. A. Weinstein, Symplectic manifolds and their Lagrangian submanifolds, *Advances in Math.*, **6** (1971), 329–346. A. Weinstein, Lagrangian submanifolds and Hamiltonian systems, *Ann. Math.*, **98** (1973), 377–410. A. Weinstein, $C^0$ perturbation theorems for symplectic fixed points and lagrangian intersections, in *Séminare sud-rhodanien de géometrié. Travaux en cours.*, Paris, Hermann, 1984, pp. 140–144. A. Weinstein, *Lectures on symplectic manifolds.* Expository lectures from the CBMS Regional Conference held at the University of North Carolina, March 8–12, 1976. Regional Conference Series in Mathematics, No. 29. American Mathematical Society, Providence, R.I., 1977. [^1]: The work is partially supported by the NSF and by the faculty research funds of the University of California, Santa Cruz. [^2]: The author is grateful to Claude Viterbo for calling his attention to Bolle’s papers, [@Bo1; @Bo2], which played a crucial role in this work. [^3]: The author is grateful to Felix Schlenk for his help in clarifying the behavior of action spectra under homotopy.

ArXiv

--- abstract: 'Some outstanding issues in high energy scattering are discussed. Particular emphasis is placed on recent developments concerning the next-to-leading log corrections to the BFKL equation.' address: 'Theoretical Physics Group, Department of Physics and Astronomy, Brunswick St, University of Manchester, Manchester M13 9PL' author: - M McDermott title: 'Low x and diffraction: theory' --- Introduction ============ The working group was concerned with high-energy or small-$x$ phenomena in QCD. In deep inelastic scattering at moderate $x=Q^2/W^2$ it is customary to think in the Breit Frame, in which the proton is moving very fast. For large $Q^2$, the short distance partonic valence structure of the proton is being probed. Striking a single valence parton very violently tends to break up the proton, hence the adjective deep-inelastic applied to this scattering. At small $x$ (less than about $10^{-3}$) the struck partons no longer carry significant amounts of the protons momentum and one may think of the photon as probing the radiation field of gluons and $q {\bar q}$-pairs surrounding the valence partons. One now has the possibility of striking a “wee lump” of this field without disturbing the integrity of the bound state of the proton. This is diffraction, and the HERA experiments have observed that indeed this possibility occurs about $10\%$ of the time even for $Q^2$ as large as $800$ GeV$^2$ (for a review of the experimental situation in diffraction see [@bcox]) . At small-x it becomes useful to switch frames to the proton’s rest frame. The photon fluctuates into a partonic system (a $q {\bar q}$ dipole to lowest order in $\alpha_s$) a large distance ($d \propto 1/x$) upstream of the stationary proton. On the timescale of the interaction with the proton, interactions within this partonic system are “frozen”. Diffractive scattering corresponds to the case in which there is no net transfer of colour between the proton and the partonic fluctuation of the photon which constitute eigenstates of diffraction. The probability of the dipole scattering is directly proportional to its transverse area. Hence, for small dipoles one expects small cross sections with the proton, since all the colour of the dipole is contained within a small transverse area, the colour field of the proton, which contains mainly long-distance fluctuations, appears transparent to it: this phenomena has become known as [*colour transparency*]{}. The longitudinally polarised photon is more inclined to fluctuate in a small symmetric system (quark and anti-quark carry roughly the same light-cone momentum fraction of the photon, $z \approx 1/2$) whereas the transversely polarised photon can split either into a large asymmetric or a small symmetric system. Although the large asymmetric system ($z \ll 1/2$) is much less likely, if produced it scatters from the protons colour field with a much greater likelihood, so that both types of dipole contribute equally (to leading twist) in the diffractive cross section. As $x \rightarrow 0$ this colour transparency picture must ultimately breakdown since the growth of the number of partons must reach the level where the impulse approximation breaks down and partons in the proton begin to recombine with each other. Exactly where in $x$ this transition to a new regime happens, at a given $Q^2$, is not known at present (although a recent model has this saturation picture built in [@wgb]). Once the parton picture has broken down it is appropriate to pursue the problem using the classical field language (see [@mkw] and references therein). Issues in diffraction ===================== The major issue in diffraction is that of universality of diffractive phenomena. Regge factorisation remains the main (imperfect) tool to study this. In principle we have four types of diffractive (rapidity gap) “experiment” to compare with one another: 1) $\gamma^{*} P \rightarrow X + Y $, 2) $\gamma_{direct} P \rightarrow $ jets + Y, 3) $\gamma_{resolved} P \rightarrow $ X(jets) + Y 4) $PP \rightarrow $ jets or W + gap (or gaps). In 1-3 $X,Y$ are hadronic systems separated by a rapidity gap. The H1 QCD analysis of diffractive DIS events [@h1diff], which require a gluon-dominated Pomeron to reproduce the positive scaling violations observed in 1), appear to do a reasonable job on 2) but overshoot 3) [@h1dijet] and fail completely to predict the Tevatron data 4). In the latter a very low fraction of diffractive events are observed (low gap survival probability) and a quark-dominated Pomeron seems to be prefered (see [@whitmore] and references therein). This non-universality is almost certainly connected with the secondary interactions of “passenger” partons in the co-moving systems (different in each case) which act to fill-in the gap of the primary diffractive scatter. Levin and collaborators have an eikonal model of multiple interactions which they have applied to the problem of gap survival probability. They find broad agreement with the Tevatron dijets data (see e.g. [@levin2]). However, it is clear that a systematic study addressing the question of gap survival probability and the breakdown of universality of diffractive exchanges is urgently needed. Recently, Donnachie and Landshoff [@dl] presented an analysis of a wide range of HERA data (including all high energy proton and charm structure function data) which seemed to require a second Pomeron with a much larger intercept (around 1.4). However, it is not clear how the inclusion of this additional Pomeron should, or would, affect the total and elastic cross-section fits. (it depends on what one assumes for the coupling of this secondary Pomeron as $Q^2 \rightarrow 0$). We now have very good high precision data on the inclusive and diffractive cross sections. However, many models with different dynamical assumptions produce adequate fits to data. Essentially the reason for this is that it is easy to tune the input parameters of a model to fit slowly varying function of $x$ and $Q^2$. In principle each model should be able to provide predictions for $F_L, F_2^{c {\bar c}}, F_L^{D}$ and $F_2^{D, c {\bar c}}$. Of course, if $F_L$ and $F_L^{D}$ were to be measured, for example by lowering the HERA beam energy, this would constrain the possible dynamics considerably. It is encouraging that there are plans and good prospects of making a good measurement $R^{D} = F_L^{D} / F_T^{D} $ at nuclear HERA (see [@nhera] for a summary of the possibilities of nuclear HERA). Recently it has become fashionable to present the data on $F_2$ using the so-called Caldwell plot (see [@bcox]). From this plot it is observed that $ dF_2/d \ln (Q^2/Q_0^2)$ plotted as a function of $x$ has a maximum around $x \approx 10^{-3}$. This has been presented as evidence for a lack of gluons at small-$x$. However as $x$ decreases the typical $Q^2$ in each bin is decreasing and it could be that what we are seeing is a breakdown of the parton picture at small $Q^2$. The saturation model of Wusthoff and Golec-Biernat [@wgb] is designed to take this observed feature into account and work is under way to apply the same model to the diffractive data. Exclusive processes ------------------- In order to be sure that it is safe to use perturbative QCD, it is necessary to isolate those diffractive process in which only small dipoles contribute, these exclusive processes are known as hard diffractive, e.g. exclusive dijet production, deeply virtual Compton scattering or heavy vector meson production. At small-$x$ all such processes are governed by the exchange of two gluons in the $t$-channel in a colour-singlet configuration. It was recently realised that the off-diagonal nature of the amplitude means that one is probing new non-perturbative information about the proton’s field and it is necessary to replace the gluon density with an off-diagonal generalization (which involves gluon operators sandwiched between different quantum states) in expressions for the cross sections of exclusive processes (see [@ji] for a review and references). In heavy vector meson production in DIS this off-diagonality arises from the need to convert a space-like photon virtuality into a time-like vector meson mass. The evolution equations governing the evolution of these new distributions in different regimes, as well as various definitions for them, are discussed in detail in [@ji]. The exclusive electro- and photoproduction of (both heavy and light) vector mesons is now a mature field both experimentally and theoretically (for a recent review see [@critt]) Recent interesting developments include the measurements of the ratios $R_{\phi}/R_{\rho}$ and $R_{J/\psi}/R_{\rho}$ versus $t$, at high-$|t|$. These measurements probe short distance part of the wavefunctions of the vector mesons for the first time. It also appears that there is evidence for s-channel helicity non-conservation in $\rho$-production (see [@bcox; @ik]). Next-to-leading order BFKL ========================== In this section I discuss some of the recent developments in small-x QCD arising from the completion of the next-to-leading corrections to the BFKL equation by Fadin and Lipatov [@fl] (see also [@cc]). There have been many papers this year on this subject. I will attempt to summarise some of the issues here. First of all we recall some results from leading order [@bfkl] (many aspects of the problem are discussed in the recent textbook by Forshaw and Ross [@fr]). The BFKL equation is most safely applied to the high-energy scattering of two small systems (one usually talks of ‘onia’, a bound state of a heavy quark and antiquark) in the Regge limit $s \gg |t|,m^2_{\mbox{onia}}$. To the leading-logarithmic accuracy in energy, onia-onia cross sections involve a four-point function for a gluon ladder in the t-channel convoluted with impact factors for each onia at the bottom and top of the ladder. By performing the convolution with one of the onia one has the unintegrated gluon structure function of this onia, $f(x,k^2_{t})$, convoluted with the remaining impact factor. The ladder is calculated in the so-called multi-Regge kinematics (MRK), i.e. each cell in the ladder is treated in the Regge limit, and the longitudinal Sudakov components of the vertical lines are strongly ordered. These components correspond to the light-cone momentum fractions of the external onia carried by the gluon concerned. This generates a logarithm in energy for each power in $\alpha_s$. The BFKL equation is an integral equation for $f$ which sums up all the leading-log parts of all such gluon ladders, i.e. all terms of order $$\sum_n ( \as \ln s/s_0)^n.$$ Working to leading-log accuracy does not allow the scale of $\as$ or the minimal longitudinal energy scale $s_0$ to be determined. By taking account all such graphs to leading-log accuracy, it is seen that the t-channel gluons Reggeize, and one has Reggeon-Reggeon gluon (RRg) vertices in the cross-rungs. The kernel of the BFKL equation is essentially democratic in its choice of momentum scales: it is [*conformally invariant*]{}, i.e. does not change under $k_t^2 \rightarrow 1/k_t^2$, (hereafter I will drop the subscript $t$). As a result the typical transverse scales are determined only by the external particles and the distribution in $k^2$ diffuses away from the ends of the ladder. This conformal invariance is possible only to leading order in logarithms since the running of the QCD coupling, which comes in at next to leading-order, explicitly breaks this scale invariance. Because of the conformal invariance, the BFKL equation can be solved by taking Mellin moments of f : $$f(s,k^2) \int_{C_{\om}} \frac {d \omega}{2 \pi i} \int_{C_{\gamma}} \frac {d \gamma}{2 \pi i} \left( \frac{s}{s_0} \right)^{\om} \left(\frac{k^2}{k_0^2} \right)^{\gm} f(\om,\gm) \, ,$$ where the Mellin space solution is $$f (\om,\gm) \propto \frac{1}{\om - \asb \chi(\gm)}.$$ If one expands the $\gm$-integral about the saddle point in the kernel, $\gm_s = (1/2,0)$, it contains a pole at $\om = \asb \chi (\gm_s) = 4 \ln 2 \asb $ with $\asb = 3 \as / \pi$. This leads to the famous rise in energy of the unintegrated structure function: $$f (s,k^2) \propto \left( \frac{k^2}{k_0^2} \right)^{\frac{1}{2}} \, \left( \frac{s}{s_0} \right)^{4 \ln 2 \, \asb}.$$ This corresponds to the lowest energy eigenvalue of a Schrödinger equation in $k^2$, with $\nu =0$ ($\gm = 1/2 + i \nu$): $$H^{(1)} \psi_{\nu} (q) = \eps_{\nu}^{(1)} \psi_{\nu} (q)$$ with $\eps_{\nu}^{(1)}$ equal to (minus) the BFKL kernel, $ K(\nu) = \asb \chi(\gm) = \asb [ 2 \psi(1) - \psi(1/2 + i \nu) - \psi (1/2 - i\nu)]$. The eigenfunctions are given by $$\psi_{\nu} (k^2) = \frac{(k^2)^{i\nu}}{\sqrt{2 \pi^2 k^2}} \, .$$ To next-to-leading log accuracy in energy one needs to sum all graphs which contains pieces which are one power down in $\as$, i.e. of order $$\sum_n \as ( \as \ln s/s_0)^n.$$ These include all the previous gluon ladders evaluated in the quasi-multi-Regge kinematics (QMRK) in which one of the strong orderings is relaxed; the original graphs, evaluated in MRK, with running coupling at one loop, or NLO corrections to the RRg vertices, or NLO gluon reggeization, or an entirely new set of graphs, with real-quark-emission insertions on the horizontal rungs (for more details see [@fadin]). In the Hamiltonian formalism At next-to-leading order the Hamiltonian becomes $H = H^{(1)} + H^{(2)}$ where the action of the next-to-leading part, $H^{(2)}$, on the leading-order eigenfunctions is given by $$\begin{aligned} H^{(2)} \psi_{\nu} (q) = \left[ K_{SI} (\nu) + K_r (\nu,q) \right] \psi_{\nu} (q) \label{nlh} \end{aligned}$$ The second term in the right hand side of equation (\[nlh\]) comes from the running of the coupling and explicitly breaks the conformal invariance. The first term corresponds to all the other corrections, it turns out that this piece is scale invariant, like the leading order piece denoted $\eps_{\nu}^{(1)}$ and leads to a shift in the the leading order eigenvalue which depends on the value of $\asb$ given by $$\eps_{\nu}^{(1)} \rightarrow \eps_{\nu} = \eps_{\nu}^{(1)} \left( 1 - \frac{\asb C(\nu)}{4} \right)$$ where the function $C(\nu)$ can be found in [@fl]. It turns out that if one concentrates on the saddle point method and expands about $\nu =0$ then the shift of the power of energy is huge and negative for reasonable values of $\asb$: $$\begin{aligned} \om_p^{(2)} &=& 4 \ln 2 \asb (1 - 6.56 \asb) \\ &=& \om_p^{(1)} (1 - 2.366 \om_p^{(1)} ) \, , \label{lc}\end{aligned}$$ and considerably reduces the strong rise with energy, $s^{\om_p^{(1)}}$ found at leading order. For example, taking $\asb = 0.15$, a large leading-power $\om_p^{(1)} = 0.4$ is reduced to $\om_p^{(1+2)} = 0.02$. Ross [@dross] has pointed out that concentrating on the point at $\gamma_s = (1/2,0)$ may be misleading, except for very small values of $\asb \leq 0.05 $. For larger values the saddle point on the real-$\gm$ axis is replaced by two saddle points off-axis and that it is necessary to expand about these new complex conjugate saddle points to get a more accurate answer. This may be achieved by expanding the kernel to order $\Or (\nu^4)$. The saddle point method then gives a larger energy power of approximately $\om_p^{(1+2)} = 0.02 + 0.09 = 0.11$. This result should also be taken with some caution since the eigenfunctions are now oscillatory, $\psi(q) \propto (q^2)^{(i \nu_s)} $, and when folded in with the impact factors could lead to negative cross sections (in addition, it implies that the full power may not be seen until very high energies) ! The running coupling piece of the next-to-leading order kernel is potentially even more troublesome as pointed out by Armesto, Bartels and Braun [@abb]. The corresponding part in the Hamiltonian is a potential piece proportional to $ \beta_0 \ln k^2/\mu^2$. Since this piece can take on any value, the energy spectrum of the Hamiltonian is unbounded from below which implies an arbitrary large power of energy is possible (there is no rightmost singularity in the $j$-plane and therefore no intercept). Such extreme growth with energy is very worrying, however, as Armesto point out, it comes with a non-perturbative damping factor, $\exp(-1/\as b)$, which may lead to its suppression for the scattering of small objects. Assuming this to be the case the authors of [@abb] also reproduce the non-Regge type behaviour (a non-power-like behaviour in s) arising from the running coupling found in [@km] (see also [@levin]), this behaviour also points to the restricted applicability of the whole formalism. In practice we know that the unboundedness comes from the running coupling approaching the Landau pole. We know that there must be some regulation in this infra-red region which marks the transition between perturbative and non-perturbative QCD. The question which needs addressing is whether the way in which one implements this infra-red regularization [@lip; @cc2] makes a critical difference to the energy dependence. If the answer is yes, it would appear that non-perturbative physics is inextricably entangled with perturbative physics, even in idealised onia-onia scattering, and that one loses all predictability in perturbative QCD at high enough energies. If the answer is no, it may be possible to factorize all non-perturbative behaviour into boundary conditions and retain some predictability. This issue is intimately connected with the enhancement of the diffusion in $k^2$ caused by the running coupling. For a more detailed discussion of the effects of running coupling in the leading-order BFKL kernel see chapter (5) of [@fr]. On a more optimistic note, DIS at moderate $x$ is well-understood and it would be very surprising if it was impossible to approach the small-$x$ region in a controlled way. A formalism exists to do this which incorporates information from BFKL dynamics into the DGLAP [@dglap] formalism by using resummed anomalous dimensions (see [@fl; @cc; @bv] and references therein). Thorne gave an interesting presentation on this issue, which included the possibility of an energy dependent coupling constant in the small-$x$ region this may lead to enlargement of the region of applicability of the DGLAP formalism [@fst]. One optimistic view of the large correction factor in equation (\[lc\]) pointed out in the discussions is that what one is seeing is the first piece of some large all-orders effect which may be resummed. This all orders effect could be connected to coherence phenomena. Salam [@salam] gave a presentation in which double logarithms in $k^2$ are resummed and discussed the corresponding radical changes in the structure of the kernel (see [@fst] for more details). As the summary above should indicate there are many issues that still need to be resolved in this fascinating area. It is clear from the discussions at Durham that the intense theoretical interest in this area is set to continue. References {#references .unnumbered} ========== [99]{} B. Cox, these proceedings. K. Golec-Biernat and M. Wuesthoff, hep-ph/9807513. J. Jalilan-Marian, A. Kovner and H. Weigert, hep-ph/9709423. H1 Collab., Z. Phys. [**C76**]{} (1997) 613. H1 Collab., hep-ex/9808013. J. Whitmore, [*Extracting diffractive parton distributions from HERA data and factorization tests*]{}, Proc. of DIS98, April 1998, Brussels, to be published by World Scientific; L. Alvero et al, hep-ph/9805268. E. Gotsman, E. Levin and U. Maor, Phys. Lett. [**B438**]{} 369. A. Donnachie and P. V. Landshoff, hep-ph/9806344 M. Arneodo, A. Bialas, M. W. Krasny. T. Sloan and M. Strikman [*Nuclear beams at HERA*]{}, Proc. of DESY 1995/1996 Workshop “Future Physics at HERA”, ed G. Ingelman, A de Roeck, R Klanner, 887. X.-D. Ji, J. A. Crittenden, hep-ex/9806020. D. Yu. Ivanov and R. Kirschner, hep-ph/9807324. V. S. Fadin and L. N. Lipatov, Phys. Lett. [**B429**]{} (1998) 127. G. Camici and M. Ciafaloni, Phys. Lett. [**B412**]{} (1997) 396. V. S. Fadin, E. A. Kuraev, L. N. Lipatov, (1975) 50; Ya. Ya. Balitsky, L. N. Lipatov, Sov. J. Nucl. Phys. [**28**]{} (1978) 822. J.R. Forshaw and D. A. Ross, 1997, [*Quantum Chromodynmaics and the Pomeron*]{}, (Cambridge: Cambridge University Press). V. S. Fadin. [*“BFKL news”*]{}, talk given at LISHEP98, February 1998, Rio de Janeiro, (hep-ph/9806482). D. A. Ross, Phys. Lett. [**B431**]{} (1998) 161. N. Armesto, J. Bartels and M. A. Braun, hep-ph/9808340 Yu. V. Kovchegov and A. H. Mueller, preprint CU-TP-899 (hep-ph/9805208). E. M. Levin, preprint TAUP 2501-98 (hep-ph/9806228). L. N. Lipatov, Sov. Phys. JETP [**63**]{} (1986) 904. G. Camici and M. Ciafaloni, Phys. Lett. [**B395**]{} (1997) 118. V. N. Gribov, L. N. Lipatov, Sov. J. Nucl. Phys. [**15**]{} (1972) 438; L. N. Lipatov, Sov. J. Nucl. Phys. [**20**]{} (1975) 94; G. Altarelli, G. Parisi, Nucl. Phys. [**B26**]{} (1977) 298; Yu. L. Dokshitzer, Sov. Phys. JETP [**46**]{} (1977) 641. J. Blümlein and A. Vogt, Phys. Rev. [**D57**]{} (1998) 1; [**D58**]{} (1998) 014020. J. Blümlein, V. Ravindram, W. L. van Neervan, A. Vogt, hep-ph/9806368. J. R. Forshaw, G. Salam and R. Thorne, these proceedings. G. Salam, hep-ph/9806482

ArXiv

--- abstract: 'Functional verification constitutes one of the most challenging tasks in the development of modern hardware systems, and simulation-based verification techniques dominate the functional verification landscape. A dominant paradigm in simulation-based verification is directed random testing, where a model of the system is simulated with a set of random test stimuli that are uniformly or near-uniformly distributed over the space of all stimuli satisfying a given set of constraints. Uniform or near-uniform generation of solutions for large constraint sets is therefore a problem of theoretical and practical interest. For boolean constraints, prior work offered heuristic approaches with no guarantee of performance, and theoretical approaches with proven guarantees, but poor performance in practice. We offer here a new approach with theoretical performance guarantees and demonstrate its practical utility on large constraint sets.' author: - Supratik Chakraborty - 'Kuldeep S. Meel' - 'Moshe Y. Vardi' bibliography: - 'Report.bib' date: 'Sept. 6, 2012' title: 'A Scalable and Nearly Uniform Generator of SAT Witnesses[^1]' --- [^1]: The final version will appear in the Proceedings of CAV’13 and will be available at [link.springer.com](link.springer.com). Work supported in part by NSF grants CNS 1049862 and CCF-1139011, by NSF Expeditions in Computing project “ExCAPE: Expeditions in Computer Augmented Program Engineering”, by BSF grant 9800096, by gift from Intel, by a grant from Board of Research in Nuclear Sciences, India, and by the Shared University Grid at Rice funded by NSF under Grant EIA-0216467, and a partnership between Rice University, Sun Microsystems, and Sigma Solutions, Inc.

ArXiv

--- abstract: 'Continuous-variable quantum key distribution (CVQKD) is considered to be an alternative to classical cryptography for secure communication. However, its transmission distance is restricted to metropolitan areas, given that it is affected by the channel excess noise and losses. In this paper, we present a scheme for implementing long-distance CVQKD using separable Gaussian states. This tunable QKD protocol requires separable Gaussian states, which are squeezed and displaced, along with the assistance of classical communication and available linear optics components. This protocol originates from the entanglement of one mode and the auxiliary mode used for distribution, which is first destroyed by local correlated noises and restored subsequently by the interference of the auxiliary mode with the second distant separable correlated mode. The displacement matrix is organized by two six-dimensional vectors and is finally fixed by the separability of the tripartite system. The separability between the ancilla and Alice and Bob’s system mitigates the enemy’s eavesdropping, leading to tolerating higher excess noise and achieving longer transmission distance.' author: - 'Jian Zhou$^{1}$, Duan Huang$^{1,2}$, and Ying Guo$^{1}$' bibliography: - 'reference.bib' title: 'Long-distance continuous-variable quantum key distribution using separable Gaussian states' --- Introduction ============ Quantum key distribution (QKD) [@bennett1984quantum; @PhysRevLett.67.661] enables two distant parties, conventionally called Alice and Bob, who have access to an authenticated classical channel, to share secret keys in the presence of eavesdropper, Eve. The unconditional security of an ideal QKD protocol has been established even if it is exposed to an adversary, who possesses unlimited computing power and technological capabilities [@Mayers2001Unconditional; @PhysRevLett.114.070501; @Leverrier2017Security; @RevModPhys.81.1301]. Normally, QKD is divided into two kinds: discrete-variable (DV) QKD [@PhysRevLett.67.661; @PhysRevA.69.012309], which relies on photon counting techniques, and continuous-variable (CV) QKD [@PhysRevLett.88.057902; @RevModPhys.84.621; @PhysRevA.96.022320; @PhysRevA.95.042326], which relies on coherent detection. Equipped with the decoy state technique [@PhysRevLett.94.230504], DVQKD can realize hundreds of kilometers of communication [@PhysRevLett.117.190501]. With the help of a satellite, the transmission distance of QKD has been extended to $1200$ kilometers [@Liao2017Satellite]. Another branch of QKD, CVQKD, which has stable, reliable light resources and high detection efficiency, is more compatible with classical optical communications when compared to DVQKD [@RevModPhys.84.621]. However, despite all the advantages, CVQKD cannot yet replace DVQKD since its transmission distance is too short [@Pirandola2015High; @PhysRevA.93.022325]. One reason for the short distance is the presence of the eavesdropper, Eve, who can perturb the quantum system using the most general strategies allowed by quantum mechanics. Another one is that CVQKD schemes require a far more complicated error correction procedure, which further restricts the secure transmission distance. Einstein associated entanglement with spooky action-at-a-distance [@PhysRev.47.777], which is different from the current view in quantum information theory that regards entanglement as a physical resource. Entanglement [@RevModPhys.81.865] has been widely applied to QKD [@Epping2017Multi], quantum dense coding [@PhysRevA.92.052330], quantum teleportation [@Ren2017Ground], entanglement swapping [@PhysRevLett.119.170502] and beating classical communication complexity bounds [@PhysRevA.72.050305]. For example, global quantum operations can be implemented in quantum teleportation utilizing entanglement and classical communication. Great effort has been devoted to distributing and manipulating entanglement among separated parties. In addition, a scheme of entangling two distant parties based on communication via a quantum channel and local operations and classical communication (LOCC) was proposed [@PhysRevLett.91.037902]. Entanglement between distant parties can be created by sending a mediating particle between them via a quantum channel: swap the first particle with the ancilla, send it through the channel and entangle it with the second particle. Besides the qubit protocol, distributing CV entanglement by separable Gaussian states has also been suggested [@PhysRevA.77.050302; @PhysRevA.80.032310]. Two separable modes $A$ and $B$ may be entangled after interacting with the auxiliary mode $C$. Unfortunately, pure quantum states cannot achieve this target. Moreover, Alice and Bob usually apply squeezing and displacement operations on these modes to enhance the practical quantum information processing. Recently, the aforementioned operations have been verified in experiment [@PhysRevLett.111.230506; @PhysRevLett.111.230504]. To lengthen the transmission distance of the CVQKD system, we develop an improved protocol which transmits a separable ancilla without sending the secret information directly as usual. It may entangle mode $A$, in Alice’s laboratory, with separable mode $B$, in Bob’s distant laboratory, by sending an ancillary mode $C$ which is separable from the subsystem $(AB)$ [@PhysRevLett.91.037902]. Normally, the quantum transmission channel is assumed to be under Eve’s control in QKD. We exemplify the entanglement between Alice’s and Bob’s modes and the separability between the ancilla and the kept particle by calculating the lowest eigenvalue. In previous fully Gaussian protocols, Eve’s system $E$ purifies $AB$, so that, $S(E)=S(AB)$. Fortunately, in this scheme, the transmitted particle $C$ that may be attacked by Eve is separable from $AB$. The eavesdropper cannot get access to Alice’s and Bob’s laboratories as well as the information transmitted in the classical channel. In this case, it is impossible for the eavesdropper to recover the process of the protocol and hence she cannot extract any information. In such a scenario, the proposed scheme reduces the information leaked to the eavesdropper, thus enables longer transmission distance. This paper is organized as follows: In Sec. \[sec2\], we review the distribution of entanglement with separable states. In Sec. \[sec3\], we present the details of CVQKD scheme with separable states. Sec. \[sec4\] shows the performance of the proposed CVQKD scheme under general eavesdropping. Finally, we conclude this paper in Sec. \[sec5\]. Entanglement distribution with separable states {#sec2} =============================================== ![(Color online) Alice’s particle and Bob’s particle interact with a mediating particle $C$ continuously. Alice and Bob get entangled while leaving $C$ separable from the system $AB$. WCL denotes weak coherent laser, and $S(X)$, $S(P)$ are compression operations on along position and momentum directions. $D$ is a local displacement distributed according to the Gaussian distribution with correlation matrix $Q$.[]{data-label="figure1"}](figure1.pdf){height="45mm"} Distributing entanglement with separable states is a breakthrough in the theory of quantum entanglement. It has been shown that separable Gaussian states can be used for implementing entanglement distribution [@PhysRevA.77.050302; @PhysRevA.80.032310]. As shown in Fig. \[figure1\], this process can be accomplished by communication via a quantum channel and LOCC. At the start of the original entanglement distribution protocol, Alice prepares systems $A$ and $C$ in a Gaussian state while Bob prepares system $B$ in a Gaussian state. The three quantum systems are fully separable at this stage. Alice squeezes her two systems: one along the position quadrature and the other along the momentum quadrature. In order to keep the ancilla separable from system $AB$, a displacement operation is applied to each of the three systems. Note that the displacement is dependent on the squeezing parameters $r_1$ and $r_2$. Alice sends her two systems into a beam splitter. The beam splitter operation on modes $A$ and $C$ results in a state separable with respect to two bipartitions: $B-AC$ and $C-AB$. One of the outputs is stored in Alice’s quantum memory (QM). The other is sent to Bob via a quantum channel. Bob also applies a beam splitter operation on modes $B$ and $C$. Mixing of modes $B$ and $C$ on a balanced beam splitter finally entangles $A$ and $B$ while $C$ still remains separable from $AB$. In what follows, we recall how a displacement operation may make the transmitted ancilla $C$ separable from $AB$ [@PhysRevA.77.050302]. Before the displacement operation, modes $A$ and $C$ are in a two-mode squeezed vacuum state and mode $B$ is in a vacuum state. The output of the first beam splitter is a two-mode squeezed vacuum state with the following covariance matrix (CM): $$\begin{aligned} \gamma_{AC}=\left[ \begin{matrix} \cosh{(2\tau)}I_2 & \sinh{(2\tau)}\sigma_z \\ \sinh{(2\tau)}\sigma_z & \cosh{(2\tau)}I_2 \\ \end{matrix} \right],\end{aligned}$$ where $\tau\ \geq0$ is the squeezing parameter. Modes $A$ and $C$ are entangled when the lower symplectic eigenvalue $\nu_{\mathrm{min}}$ of the partial transpose of CM $\gamma_{AC}$ is less than one [@PhysRevA.77.050302]. The CM of the three-mode system $ABC$ is given by $$\begin{aligned} \gamma_{ABC}=\left[ \begin{matrix} \cosh{(2\tau)}I_2 & 0 & \sinh{(2\tau)}\sigma_z \\ 0 & I_2 & 0 \\ \sinh{(2\tau)}\sigma_z & 0 & \cosh{(2\tau)}I_2 \\ \end{matrix} \right].\end{aligned}$$ We add an excess non-negative matrix $P$ to $\gamma_{ABC}$ $$\begin{aligned} \gamma^1_{ABC}=\gamma_{ABC}+xP, \label{3}\end{aligned}$$ to entangle mode $A$ and modes $BC$, while leaving the other two bipartitions separable. We follow the method for the construction of three-mode entangled Gaussian states in [@PhysRevA.64.052303] to build matrix $P$. The entanglement between modes $A$ and $C$ can be destroyed by adding a positive multiple of sum of the projectors onto the subspace spanned by two six-dimensional vectors [@PhysRevA.64.052303; @PhysRevA.77.050302]. The negative eigenvalue of the CM is $\lambda=-(1-e^{-2\tau})$ with its eigenvector $p_\lambda=p_1+ip_2$ for $p_1=(0,1,0,1)^T$ and $p_2=(1,0,-1,0)^T$. We extend $p_1$ and $p_2$ to the six-dimensional vectors $q_1=(0,1,0,-2,0,1)^T$ and $q_2=(1,0,2,0,-1,0)^T$ with the displacement matrix $P=q_1q^T_1+q_2q^T_2$. In order to smear the entanglement between modes $A$ and $C$, we add a sufficiently large, nonnegative multiple $xP$ to the CM as shown in Eq. (\[3\]) and obtain $$\begin{aligned} \gamma^1_{ABC}=\left[ \begin{matrix} aI_2 & 2x\sigma_z & b\sigma_z \\ 2x\sigma_z & (1+4x)I_2 & -2xI_2 \\ b\sigma_z & -2xI_2 & aI_2 \\ \end{matrix} \right].\label{eq2}\end{aligned}$$ where $a=\cosh(2t)+x$ and $b=\sinh(2t)-x$. Then the lowest symplectic eigenvalue of matrix $(\gamma^1_{ABC})^{(T_C)}$ can be derived as  [@PhysRevA.65.032314], $$\begin{aligned} \nu_{\mathrm{min}}=\frac{\sqrt{(1+6x+e^{-2\tau})^2-32x^2}-(1+2x-e^{-2\tau})}{2}.\end{aligned}$$ The separable bound of $C$ and $AB$ is $\frac{e^{2\tau}-1}{2}$, where the parameter $x$ should be equal or greater than this value. On the other hand, the lowest eigenvalue of matrix $(\gamma^1_{ABC})^{(T_A)}$ can be calculated as $$\begin{aligned} \label{5} \kappa_\mathrm{min}=\frac{1+6x+e^{-2\tau}-\sqrt{(1+2x-e^{2\tau})^2+32x^2}}{2}.\end{aligned}$$ Taking $x\geq0$ and $\tau\geq0$ in Eq. \[5\], the lowest eigenvalue is less than one, which verifies that there is entanglement between $A$ and $BC$. Fig. \[figure6\] shows the lowest symplectic eigenvalue of matrix $(\gamma^1_{ABC})^{(T_C)}$ and $(\gamma^1_{ABC})^{(T_A)}$. To satisfy the separability of $C-AB$, the lowest symplectic eigenvalue corresponding to the dashed line should be greater than one. Similarly, the lowest symplectic eigenvalue corresponding to the full line ought to be less than one to ensure the entanglement between $A$ and $BC$. Finally, after applying reverse operation of the beam splitter on $\gamma^1_{ABC}$, the covariance matrix of the random displacement distributed according to Gaussian distribution is fixed. The beam splitter transforms the CM in (\[eq2\]) to CM $\gamma^{2}_{ABC}$ that is as follow: $$\begin{aligned} \gamma^2_{ABC}=\left[ \begin{matrix} aI_2 & \frac{2x+b}{\sqrt{2}}\sigma_z & \frac{2x-b}{\sqrt{2}}\sigma_z \\ \frac{2x+b}{\sqrt{2}}\sigma_z & \frac{1+a}{2}I_2 & \frac{1+4x-a}{2}I_2 \\ \frac{2x-b}{\sqrt{}2}\sigma_z & \frac{1+4x-a}{2}I_2 & \frac{1+8x+a}{2}I_2 \\ \end{matrix} \right].\label{eq3}\end{aligned}$$ The symplectic eigenvalue of CM $\gamma_{AB}$ can be calculated as $\nu=0.3968$ for $e^{2\tau}=10$, and the entanglement can be obtained as $E_\mathcal{N}=-\log_2{\nu}\approx1.33$ ebits. According to the entanglement distribution with separable Gaussian states, we find that the entanglement is firstly destroyed by displacement operations, which makes the auxiliary mode separable from sender’s mode. After that, the auxiliary mode is sent to Bob who partially restores the entanglement by mixing it with his suitably classically correlated mode, leading to the entanglement enhancement. Using this elegant characteristics, we propose an improved CVQKD scheme to lengthen the maximum transmission distance with separable Gaussian states. Continuous variable quantum key distribution with separable Gaussian states {#sec3} =========================================================================== This section is divided into three parts: the first part gives the CVQKD protocol using separable Gaussian states, the second part analyses the security of normal CVQKD protocol, while the third subsection states the merit of the protocol based on separable Gaussian states. A. Design of the CVQKD protocol using separable Gaussian states {#a.-design-of-the-cvqkd-protocol-using-separable-gaussian-states .unnumbered} --------------------------------------------------------------- Two normal parties, Alice and Bob aim to share secret key. For the sake of simplifying the process, we add the displacement operation in the form of matrix while the practical displacement is not complex. The prepare and measure description of the CVQKD based on entanglement distribution protocol using Gaussian states is shown in Fig. \[figure2\] and is described as follows. - Alice prepares two squeezed vacuum states which are position-squeezed and momentum-squeezed vacuum states, respectively. Displacement operations are added on these squeezed states. The output of the first beam splitter is a two-mode squeezed vacuum state if we ignore the displacement operation. - Alice detects one of the outputs with homodyne detection and sends another one to Bob via a quantum channel. - After receiving Alice’s mode, Bob interferes his vacuum state with the received state at a balanced beam splitter. - Bob heterodynes one of the beam splitter’s outputs with the self-referenced strategy, whereas another one is discarded directly. {height="60mm"} In Alice’s laboratory, she prepares two states, one position-squeezed vacuum state and one momentum-squeezed vacuum state given by $$\begin{aligned} \gamma_A=\left[ \begin{matrix} e^{2\tau} & 0 \\ 0 & e^{-2\tau} \end{matrix} \right], &&\gamma_C= \left[\begin{matrix} e^{-2\tau} &0\\ 0 & e^{2\tau} \end{matrix}\right].\end{aligned}$$ The CM of the beam splitter’s output can be expressed as $$\begin{aligned} \gamma_{AC}=\left[ \begin{matrix} VI_2 & \sqrt{V^2-1}\sigma_z \\ \sqrt{V^2-1}\sigma_z & VI_2 \\ \end{matrix} \right],\end{aligned}$$ with $V=\frac{e^{2\tau}+e^{-2\tau}}{2}$, $\sigma_Z=\left[\begin{smallmatrix}1&0 \\ 0&-1\\ \end{smallmatrix}\right]$ and $I_2=\left[\begin{smallmatrix}1&0 \\ 0&1\\ \end{smallmatrix}\right]$. The CM of $ABC$ before transmission without displacement is $$\begin{aligned} \gamma_1=\left[ \begin{matrix} VI_2 & 0 & \sqrt{V^2-1}\sigma_z \\ 0 & I_2 & 0 \\ \sqrt{V^2-1}\sigma_z & 0 & VI_2 \\ \end{matrix} \right].\end{aligned}$$ Taking the displacement into consideration, the corresponding CM becomes $$\begin{aligned} \gamma_2=\left[ \begin{matrix} aI_2 & b\sigma_z & 2x\sigma_z \\ b\sigma_z & aI_2 & -2xI_2 \\ 2x\sigma_z & -2xI_2 & (1+4x)I_2 \\ \end{matrix} \right],\end{aligned}$$ with $a=V+x$ and $b=\sqrt{V^2-1}-x$. The linear channel can be equivalent to a beam splitter with transmittance $\eta$, the function of transmission distance $\eta=10^{-\frac{L}{50}}$. The equivalent CM of the channel is $$\begin{aligned} B_\eta=\left[ \begin{matrix} I_2 & 0 & 0 & 0 \\ 0 & I_2 & 0 & 0 \\ 0 & 0 & \sqrt{\eta}I_2 & \sqrt{1-\eta}I_2 \\ 0 & 0 & -\sqrt{1-\eta}I_2 & \sqrt{\eta}I_2 \\ \end{matrix} \right].\end{aligned}$$ After the attenuation of the channel, the CM of the whole system $ABC$ becomes $$\begin{aligned} \gamma_3=\left[ \begin{matrix} aI_2 & b\sqrt{\eta}\sigma_z & 2x\sigma_z \\ b\sqrt{\eta}\sigma_z & (a\eta+(1-\eta)N_0)I_2 & -2x\sqrt{\eta}I_2 \\ 2x\sigma_z & -2x\sqrt{\eta}I_2 & (1+4x)I_2 \\ \end{matrix} \right],\end{aligned}$$ where $N_0$ is the variance of channel thermal noise. In normal QKD protocols, Bob performs homodyne or heterodyne detection on the received signals. However, the direct-detection scheme may leave the attacker loophole to eavesdrop information. Instead, Bob prepares a vacuum state and applies a displacement operation on it. Using a balanced beam splitter, Bob mixes the incoming mode with his own mode. The second balanced beam splitter transforms the CM into $\gamma_4=B_{BC}\cdot\gamma_3\cdot B^T_{BC}$. After the beam splitter, one of the outputs is detected with homodyne detection using the self-reference technique, while another one is discarded directly. The CM of the system $AB$ is $$\begin{aligned} \gamma_{AB}=\left[ \begin{matrix} aI_2 & \frac{2x+b\sqrt{\eta}}{\sqrt{2}}\sigma_z \\ \frac{2x+b\sqrt{\eta}}{\sqrt{2}}\sigma_z & \frac{1+N_0+4x(1-\sqrt{\eta})+a\eta-N_0\eta}{2} \\ \end{matrix} \right],\label{14}\end{aligned}$$ which can be used for calculating the secret key rate of the protocol. B. Attacking strategy with general eavesdropping {#b.-attacking-strategy-with-general-eavesdropping .unnumbered} ------------------------------------------------ A QKD protocol is secure against general attack when it is secure against Gaussian collective attack [@PhysRevLett.114.070501; @Leverrier2017Security]. This part performs an asymptotic security analysis based on infinitely-many uses of the channel under Gaussian collective attack. In each transmission, Eve may intercept the mode and make it interact with an ensemble of ancillary vacuum modes via a general unitary operation. One of the output modes is sent to Bob while another one is stored in Eve’s quantum memory (QM). These states in QM will be measured at the end of the protocol collectively. Taking reverse reconciliation into account, the final key rate can be derived as $$\begin{aligned} R=\xi I_{AB}-\chi_{BE},\end{aligned}$$ where $\xi$ denotes the reconciliation efficiency. We can compute the mutual information in terms of signal-to-noise ratio as $$\begin{aligned} I_{AB}=\log_2{\frac{\varphi+1}{\omega}}.\end{aligned}$$ $\varphi$ is the modulation variance in shot-noise units and $\omega$ represents the equivalent noise. In the previous CVQKD protocols, Eve’s system $E$ purifies $AB$, so that $S(E)=S(AB)$, and $S(AB)$ can be calculated from the symplectic eigenvalues of the covariance matrix $V_{AB}$. In order to calculate the Holevo bound between Alice and Bob with the simplification of the expression, we denote the CM of the reduced state of systems $AB$ as [@holevo1973holevo] $$\begin{aligned} \gamma_{AB}=\left[ \begin{matrix} aI_2 & c\sigma_z \\ c\sigma_z & bI_2 \\ \end{matrix} \right].\end{aligned}$$ The symplectic eigenvalues can be calculated as [@serafini2004symplectic] $$\begin{aligned} \nu^2_{1,2}=\frac{1}{2}[\Delta\pm\sqrt{\Delta^2-4D^2}],\end{aligned}$$ where $\Delta=a^2+b^2-2c^2$ and $D=ab-c^2$. Moreover, the symplectic eigenvalue of the conditional CM $V_{B|A}$ is $\nu^2_3=b(b-c^2/a)$. Therefore, we have $S(AB)=G(\nu_1)+G(\nu_2)$ and $S(B|A)=G(\nu_3)$ with $$\begin{aligned} G(x)=\left(\frac{x+1}{2}\right)\log_2{\left(\frac{x+1}{2}\right)}-\left(\frac{x-1}{2}\right)\log_2{\left(\frac{x-1}{2}\right)}.\end{aligned}$$ Consequently, the information eavesdropped by Eve can be bounded by $\chi_{BE}=S(AB)-S(B|A)$. C. Secret key rate of the separable-state CVQKD {#c.-secret-key-rate-of-the-separable-state-cvqkd .unnumbered} ----------------------------------------------- It is necessary to note that the proposed protocol is different from the traditional protocol as the above-involved states are displaced before being mixed on the beam splitter. Without the displacement, the output of the first beam splitter is equivalent to a two-mode squeezed vacuum state. Another difference from the entanglement-based scheme is that Bob injects the received mode and his own mode into one beam splitter instead of performing homodyne or heterodyne detection directly. As analyzed in Sec. \[sec2\], all these efforts are to keep the ancillary mode separable from system $AB$ while completing the task of distribution entanglement between Alice and Bob. Whereas, in the traditional CVQKD system, the information is encoded on the mode that is sent to the channel under Eve’s control. Eve may hide her attack in the channel noise. It has been assumed that Eve’s system purifies $AB$, which implies that $S(E)=S(AB)$. In the proposed protocol, the auxiliary mode used for distributing information is separable from $AB$. Alice’s and Bob’s labs as well as the classical communication are out of Eve’s touch. Namely, Eve cannot steal any information by attacking the ancilla, leading to $S_{E}=0$. A problem about upper bound arises. In [@Takeoka2014Fundamental; @wilde2017converse; @Pirandola2017Fundamental], it has been proved that the secret key rate cannot be unbounded with increasing signal energy for normal CVQKD protocol [@PhysRevLett.88.057902]. The secret key rate satisfying the condition $$\begin{aligned} R\leq I_{AB}-\chi_{BE}\leq G(\varphi)-G(\nu_1)-G(\nu_2).\end{aligned}$$ The limit for $\varphi\rightarrow+\infty$ for the right part of the inequation is regular and finite [@Takeoka2014Fundamental; @wilde2017converse; @Pirandola2017Fundamental]. The secret key rate will not be unbounded with increasing signal energy even though $\chi_{BE}$ is removed. A positive multiple of sum of the projectors is added to smear the entanglement between the $C$ and $AB$ before transmission. The displacement which is proportional to the modulation variance also appears in the noise. The secret key rate of this scheme will not be unbounded as the signal-to-noise ratio is bounded regardless of the increasing signal energy. The advantage of keeping the ancillary state separable is the displacement before beam splitter. Bob uses a displaced state to interact with the ancilla rather than detects it directly. This operation is just to cut off Eve’s disturbance. Then the secret key rate can be expressed as $ R=\xi I_{AB}, $ where $\xi$ is the negotiation efficiency and $I_{AB}$ can be calculated from the CM of system $AB$ in Eq. (\[14\]). Simulation results {#sec4} ================== ![(Color online) Equivalent excess noise as a function of channel transmission $\eta$. The dashed lines are the equivalent excess noise of original protocol while the full lines denote the proposed one. From bottom to top, $N_0=1,3,5$.[]{data-label="figure8"}](figure8.pdf){height="55mm"} As discussed above, Alice and Bob can get the reduced CM $\gamma_{AB}$, from which they can calculate the secret key rate $R$ in Eq.(15). Based on the Eq. (\[14\]), the equivalent excess noise can be expressed as $$\begin{aligned} \omega=\frac{1+(1-\eta)N_0+4x(2-\sqrt{\eta})}{2},\end{aligned}$$ which is plotted in Fig. \[figure8\]. Compared with the traditional CVQKD protocol, the proposed protocol has an extra noise that is caused by the displacement operation. The displacement may decrease the key rate $I_{AB}$. Fortunately, it can also remove the entanglement between the ancillary particle and the kept particles. To demonstrate the performance of the protocol, we consider both direct reconciliation and reverse reconciliation. ![(Color online) Secret key rates versus transmission distance from Alice to Bob of the direct reconciliation case. The secret key rate decreases as the grow of the transmission distance. Simulation results refer to $V=2$ (blue dashed line), $V=10$ (red full line), $V=30$ (blue dashed line) and $V=100$ (green dot-dashed line).[]{data-label="figure3"}](figure3.pdf){height="55mm"} In Fig. \[figure3\], we show the secret key rate of the proposed protocol with direct reconciliation. From top to bottom, the dashed, full, dotted and dot-dashed lines refer to the modulation variances $2, 10, 30$ and $100$, respectively. With current technology, the $15\mathrm{dB}$ squeezed states of light has already been detected in [@PhysRevLett.117.110801]. The transmission can exceed $15\mathrm{km}$, which corresponds to the $3\mathrm{dB}$ restriction in direct reconciliation. Moreover, the excess noise has been taken into consideration with $\epsilon=0.05$ and reconciliation efficiency is set $\beta=0.95$ for all numerical simulations. ![(Color online) Secret key rates versus channel transmission, $\eta$. The full lines are under the ideal condition with zero excess noise while the dot-dashed and dotted lines correspond to $N_0=2$ and $4$, respectively. The thick and thin lines are under the condition that modulation variance $V=10$ and $100$.[]{data-label="figure4"}](figure4.pdf){height="55mm"} The simulation result in Fig. \[figure4\] is the secret key rate of the direct reconciliation case. The difference between thin lines and thick lines shows that modulation variance plays a positive role in the secret key rate. However, the displacement term limits the continued increase of the secret key rate. The full line, dot-dashed line and dotted line show channel noise has a negative effect on the secret key rate. We find that there is little effect of the noise on the secret key rate of the CVQKD system when the transmittance approaches to one. ![(Color online) Secret key rates versus transmission distance, $L$. The full lines correspond to the condition with excess noise $N_0=1.01$ while the dashed lines correspond to $N_0=2$. The thin lines represent the proposed protocol with separable Gaussian states while the thick lines are the the traditional protocols. In the simulation, the modulation variance $V=30$.[]{data-label="figure5"}](figure5.pdf){height="55mm"} Fig. \[figure5\] demonstrates the secret key rates of the proposed protocol using a separable ancilla in the reverse reconciliation case. The traditional CVQKD system can only transmit $30\mathrm{km}$ due to the existence of the eavesdropper, whereas the proposed protocol achieves the transmission distance $200\mathrm{km}$ at rate of $10^{-4}$ bits per pulse. The transmission distance of the separable-state CVQKD protocol is lower than that of the traditional one. This phenomenon may result from the abandon of the ancillary particle. Moreover, we can also find that the protocol has a better tolerance to noise than the traditional one. ![(Color online) Secret key rates of CVQKD with separable states versus the upper bound of CVQKD. The thick green line is the upper bound of the traditional CVQKD. The dotted, dashed and thin full lines are the proposed CVQKD protocols with $N_0=1,2,3$, respectively.[]{data-label="figure9"}](figure9.pdf){height="55mm"} In Fig. \[figure9\], we make a comparison between the secret key rate of our protocol and the fundamental limit [@Pirandola2017Fundamental; @PhysRevLett.102.050503].  [@Pirandola2017Fundamental] proved the PLOB bound, while  [@wilde2017converse] later discussed the strong convergence of this bound. The top green line is the fundamental limit of general CVQKD protocol, which is given by $-\log_2(1-\eta)$. $\eta$ is channel transmittance of the pure-loss channel. As shown in [@wilde2017converse; @Pirandola2017Fundamental], the protocols whose secret key rate is based on the lower bound cannot come up with the upper bound when the transmittance $\eta$ is less than $0.7$. The protocol based on transmission of separable Gaussian states via a quantum channel and LOCC operation has a good performance on the aspect of transmission distance. This scheme has a good tolerance for excess noise and the transmission distance achieves $200\mathrm{km}$. Conclusion {#sec5} ========== We have proposed an improved continuous-variable quantum key distribution protocol that is immune to Eve’s attack. This separable-state CVQKD protocol is different from the traditional protocol because the ancillary particle is separable from Alice and Bob¡¯s system. In previous protocols, the information is encoded on the particles which will pass through a quantum channel controlled by Eve. Eve can purify the whole system and extracted as much information as the Holevo bound of the system. In addition, after the two respective particles interact continuously with an ancilla, they get entangled, leaving the ancilla separable all the time. The displacement operation in the preparation course plays a crucial role in smearing the entanglement between the ancilla and Alice and Bob’s system. The secret key rate of the separable-state CVQKD will not be unbounded with increasing signal energy. The proposed protocol has good tolerance to extra noise and is able to keep abreast of the upper bound until $200\mathrm{km}$. We note that the proposed CVQKD protocol can be practically implemented using separable Gaussian states as entanglement preparation processes based on separable Gaussian states have been demonstrated in experiment [@PhysRevLett.111.230506; @PhysRevLett.111.230504]. Acknowledgements {#acknowledgements .unnumbered} ================ We would like to thank L. Mi$\check{\mathrm{s}}$ta for helpful discussion. This work is supported by the National Natural Science Foundation of China (Grant Nos. 61572529) and the Fundamental Research Funds for the Central Universities of Central South University (2017zzts144).

ArXiv

--- abstract: 'We study the effects of dissipation or leakage on the time evolution of Grover’s algorithm for a quantum computer. We introduce an effective two-level model with dissipation and randomness (imperfections), which is based upon the idea that ideal Grover’s algorithm operates in a 2-dimensional Hilbert space. The simulation results of this model and Grover’s algorithm with imperfections are compared, and it is found that they are in good agreement for appropriately tuned parameters. It turns out that the main features of Grover’s algorithm with imperfections can be understood in terms of two basic mechanisms, namely, a diffusion of probability density into the full Hilbert space and a stochastic rotation within the original 2-dimensional Hilbert space.' author: - 'Pil Hun Song and Ilki Kim[@AUTH]' title: 'Computational leakage: Grover’s algorithm with imperfections' --- Recently, quantum computing has emerged as one of the most challenging fields of physics both for theoreticians and experimentalists (see Ref.  for a review). At the core of the theoretical side, a few quantum algorithms are now available, which can solve a certain class of problems faster than any available classical counterparts: for example, S’s algorithm [@shor] factorizes a given large number $N$ at $\sim (\log N)^2$ time steps with an exponential speed-up. Using G’s algorithm (GA),[@grover] one can find a specific item on a long list of size $N$ at $\sim \sqrt{N}$ time steps, which is a considerable gain in speed as compared with $\sim N$ in classical algorithms. These quantum algorithms operate perfectly only on ideal quantum computers. On the other hand, a certain amount of dissipation or uncontrolled coupling to the environment is clearly inevitable on real quantum computers. For example, any deviation from ideal operation in quantum gates, which may result from various origins, including fluctuation in the excitation energies of two-level systems (qubits), can be considered as “imperfections”. The imperfections will affect the efficiency of a quantum computer, and the operability of a given quantum algorithm may break down to the point of losing its advantage over a classical counterpart. Therefore, it is of vital importance to have a sound picture of how an error due to the presence of imperfections evolves in quantum algorithms. Obviously, a reasonable picture of the basic mechanisms given by the imperfections will be very crucial in constructing an appropriate quantum error correction method.[@cal; @steane2; @preskil; @gottes] In general, the quantum state in a quantum computer is essentially a many-body (or network) state, the time evolution of which is delicately controlled by a given quantum algorithm. From such a point of view, the study of imperfection effects on quantum algorithms would belong to a more general research field which investigates disorder effects on the dynamics of a many-body state. Their exact treatment is actually a complicated subject, and only a few results have been obtained giving either general frameworks for understanding the effects or general methodologies for calculation. There exist several theoretical, mainly numerical, investigations in this direction. The main stress has been given, from a practical point of view, on the stability of quantum algorithms with respect to the presence of imperfections. C and Z[@zoller] reported that the operability of quantum computing is rather safe against disorders available in the quantum F transform process. In Refs.  and , the disorder effect in S’s algorithm applied to the factorization of the number 15 was studied and by using the fidelity being defined as the square of the overlap of the actual quantum state with the ideal one, it was found that the operability of the S’s algorithm can be destroyed due to a very small strength of the disorder in the modular exponentiation part.[@zurek] More systematic results have recently been obtained in Ref.  from the study of quantum computing of quantum chaos and imperfection effects: by considering the presence of imperfections in the quantum F transform, it was obtained that the imperfection strength scales polynomially with the number of qubits for the inverse participation ratio (IPR), which measures the strength of localization of quantum state and plays a role of the fidelity in Ref. . Nevertheless, it still remains at a primitive stage regarding an understanding of basic mechanisms carried by the imperfections in quantum algorithms. So far, the main policy has been simply to watch a deviation of the quantum state from the ideal one and to analyze its parameter dependence. In this paper, we investigate the time evolution of a state governed by G’s algorithm with imperfections, with a main emphasis on an understanding of interplay of the imperfections with the algorithm operator. Based on the idea that the ideal GA operates in an effective 2-dimensional H space, a stochastic two-level model with dissipation will be introduced, and then its simulation results will be compared to those of the GA with imperfections, which operates in a larger relevant H space resulting from the presence of the imperfections. They are in a good agreement via an appropriate fit of parameters. An analytic solution of the two-level model is given with some modification and provides a comprehensive picture of imperfection effects on the GA. Let us begin with a brief sketch of the GA. The final goal is to identify $|j\rangle$ (target state) among $N = 2^{n_q}$ quantum states, where $n_q$ is the number of qubits. Initially, the state of quantum register is prepared as a superposition of all states with the same amplitude. The GA may be broken up into two steps: (i) rotation of phase of $|j\rangle$ by $\pi$ and (ii) application of a diffusion operator $D$ which is defined, in matrix form, as $D_{kl}=-\delta_{kl}+2/N$ with $k,l = 0,1, \cdots, N-1$, and $\delta_{kl}$ denoting the K delta. The step (ii) is achieved by applying the H operation to each single qubit and then performing a conditional phase shift on the computer with every computational basis state except $|k=0\rangle$ receiving a phase shift of $-1$ followed by the second H operation to each single qubit. Then, the quantum state during time evolution can be expressed as [@boyer] $$|\Psi(\vartheta)\rangle\; =\; \sin \vartheta\,|j\rangle\, +\, \frac{\cos \vartheta}{\sqrt{N-1}}\,\sum_{k \neq j} |k\rangle\,.$$ The initial state is characterized by $\vartheta = \vartheta_0$ with $\sin \vartheta_0 = 1/\sqrt{N}$. Each iteration transforms $|\Psi(\vartheta)\rangle$ into $|\Psi(\vartheta+\omega)\rangle$, where $\sin\omega = 2 \sqrt{N-1}/N$. Then, after $m \approx (\pi/4)\sqrt{N}$ iterations, $\vartheta$ becomes very close to $\pi/2$, and a measurement of the state yields $|j\rangle$ with an error $O(1/N)$. We note that the evolution of $|\Psi(\vartheta)\rangle$ according to the GA is restricted to a 2-dimensional H space which is spanned by $|x\rangle = (1/\sqrt{N-1}) \sum_{k \neq j} |k\rangle$ and $|y\rangle = |j\rangle$. Each iteration represents a rotation of the quantum state by the angle $\omega$ in the $x$-$y$ plane and the G’s operator for a single iteration can be written in a familiar form $$\hat{R}(\omega)\; =\; \left( \begin{array}{lcr} \cos \omega && -\sin \omega\\ \sin \omega && \cos \omega \end{array} \right)$$ on the basis $\left\{|x\rangle, |y\rangle\right\}$. Imperfections are introduced in the GA as follows: the ideal H operator in the step (ii) is given by $\vec{n} \cdot \vec{\hat{\sigma}}$, where $\vec{n} = (1/\sqrt{2}, 0, 1/\sqrt{2})$, and $\hat{\sigma}_{x(y,z)}$ denotes the P spin matrix. We now replace $\vec{n}$ by $$\vec{m}_q\; =\; \frac{1}{\sqrt{2}}\, (\cos \varphi_q \cdot \sin \delta_q + \cos \delta_q ,\, \sqrt{2}\,\sin \varphi_q \cdot \sin \delta_q ,\, -\cos \varphi_q \cdot \sin \delta_q + \cos \delta_q)\,,$$ where $q = 1,2, \cdots, n_q$ represent each single qubit. Here, $\delta_q$ and $\varphi_q$ with $|\delta_q| < \epsilon/2$ and $0 \leq \varphi_q < 2\pi$ are randomly chosen in an iteration of the GA and also vary randomly from iteration to iteration. Then, it turns out that $\vec{m}_q$ is a unit vector tilted from $\vec{n}$ by $\sim \epsilon$. It should be noted that, in spite of the imperfections, since the quantum state evolves without coupling to the additional environment, the qubit rotations remain unitary, keeping the normalization condition $\langle \Psi |\Psi \rangle = 1$ for any iteration number $t$. The presence of the imperfections will provide an additional coupling between the 2-dimensional H space spanned by $\left\{|x\rangle, |y\rangle\right\}$ (“computational space”) and the rest part of the total H space with $2^{n_q}$ dimensions, leading to the [*quantum leakage*]{}[@FAZ99] from the computational space as an intrinsic source of error in ideal gate operations. Typical results of the GA with imperfections are shown in Fig. 1: $\langle p_j\rangle$ and $F$ denote an ensemble-averaged probability of the target state $|j\rangle$ and an ensemble-averaged fidelity over 100 random runs, respectively, each of which is here given for $n_q =13$ and for imperfection strengths $\epsilon = 0$, 0.005, 0.01 and 0.02, respectively, as a function of iteration number $t$. Clearly, they are given by $$\langle p_j\rangle_{\epsilon}(t)\; =\; \left\langle |\langle j|\Psi(\epsilon,t)\rangle |^2 \right\rangle\;,\;\; F_{\epsilon}(t)\; =\; \left\langle |\langle \Psi(\epsilon=0,t)|\Psi(\epsilon,t)\rangle |^2 \right\rangle\,,$$ respectively, where the outer bracket represents the ensemble average. In the case of $\epsilon=0$, $\langle p_j\rangle_{0}(t)$ oscillates between 0 and 1 and reaches 1 at $t \approx (m+1/2) (\pi/2)\sqrt{N} \approx 71,\,213,\,355, \cdots$ with $m=0,1,2, \cdots$. When $\epsilon$ is non-zero, one still finds oscillating features with the same period as in the ideal case, however, with an envelope decaying nearly exponentially with time $t$. As $t$ increases, the system approaches a saturated regime, where the noise completely dominates the ideal system dynamics, and accordingly $\langle p_j\rangle_{\epsilon}(t)$ fluctuates around $1/N$. A novel feature is that the decay affects not only the shape of the upper envelope but also that of the lower envelope such that the lower envelope is not simply given by $\langle p_j\rangle_{\epsilon}(t) = 0$. This means that the probability for the system to remain at the target state is still available even at the time it originally vanishes in the ideal unperturbed system. Furthermore, $F_{\epsilon}(t)$ is found to approximately equal the upper envelope of $\langle p_j\rangle_{\epsilon}(t)$. As noted earlier, in the absence of imperfections, the wave-function of quantum register evolves within a very small part (of dimension 2) of the total H space (of dimension $2^{n_q}$). Furthermore, since the amplitudes of $|x\rangle$ and $|y\rangle$ remain real or at least keep the same phase over the time evolution, the actual relevant space is even smaller than the 2-dimensional entire H space. Let us denote the 2-dimensional H space spanned by $|x\rangle$ and $|y\rangle$ and the total H space by ${\cal H}_2$ and ${\cal H}_t$, respectively. The above results suggest that in general, the disordered GA operator yields states which are not restricted in ${\cal H}_2$ but spread over a larger space ${\cal H}_t$ (“computational leakage”). In other words, the presence of imperfections induces a probability density flow from ${\cal H}_2$ to ${\cal H}_t$ with diffusion-like nature. Then, let us define $|w_2(t)|^2$ as the probability that the state remains in ${\cal H}_2$ at time $t$ with an exponentially decaying function of $t$, $$w_2(t) = e^{-\gamma\,t}\,,$$ where $\gamma$ represents the strength of the diffusion which depends on system parameters such as the strength of imperfections and the qubit numbers. Also, the imperfections affect the dynamics of the state within ${\cal H}_2$: in general, the phases of the two amplitudes of $|x\rangle$ and $|y\rangle$ are not equal to each other, and it is reasonable to assume that random phases are introduced during each iteration. Therefore, we would now like to adopt an effective two-level model which can encapsulate the effects of imperfections in the GA living in ${\cal H}_t$. Here, the time evolution of a quantum state $|\psi(t)\rangle = c_m(t)\,|m\rangle + c_n(t)\,|n\rangle$ on the bais $\{|m\rangle, |n\rangle\}$ is described by $$\label{eq:c_iteration} \left( \begin{array}{c} c_m(t+1)\\ c_n(t+1) \end{array} \right)\; =\; e^{-\gamma}\; \hat{R}(\omega)\; \hat{U}(\phi_m,\phi_n)\; \left( \begin{array}{c} c_m(t)\\ c_n(t) \end{array} \right)\,,$$ where $\hat{U}(\phi_m,\phi_n)$ is a diagonal matrix with $U_{mm} = e^{i\phi_m}$ and $U_{nn} = e^{i\phi_n}$, and $\phi_{m}(t)$ and $\phi_{n}(t)$ are assumed to be two independent random variables without any time correlation. Let each of these phase variables be chosen from a box distribution $[-W_{\phi}/2, W_{\phi}/2]$ for a given $W_{\phi}$. The frequency $\omega = \sin^{-1}(2\sqrt{N-1}/N)$ is the same as in the GA, and the initial conditions are given by $c_m(0) = \cos \vartheta_0$ and $c_n(0) = \sin \vartheta_0$ with $\vartheta_0 = \sin^{-1}(1/\sqrt{N})$. This is a stochastic two-level model with dissipation and we refer to it as STLM hereafter. Here, we obtain, after a minor calculation, an ensemble-averaged probability of the target state $|j\rangle$ and an ensemble-averaged fidelity, respectively: $$\label{eq:stlm} \langle p_j\rangle_{W_{\phi}}^{(\gamma)}(t)\; =\; \left\langle |c_n(t)|^2 \right\rangle\,,\; F_{W_{\phi}}^{(\gamma)}(t)\; =\; \left\langle |c_m(t)\, \cos(\omega t + \vartheta_0)\, +\, c_n(t)\, \sin(\omega t + \vartheta_0)|^2 \right\rangle\,.$$ It is noteworthy to consider the difference between the STLM and the original GA with imperfections. First, the finite fraction $2^{-n_q+1}$ occupied by ${\cal H}_2$ in ${\cal H}_t$ is neglected in the STLM so that $w_2(t)$ decays to zero instead of $\sim \sqrt{2^{-n_q+1}}$. Since we are interested mainly in the regime before saturation, this is clearly not a significant difference. Secondly, the stochastic features of $\gamma$ are not considered. But, this is not critical, either, since those features will contribute a negligible correction to $\gamma$ after an ensemble-average in eq. (\[eq:stlm\]). Now, we perform a numerical simulation to obtain $\langle p_j\rangle_{W_{\phi}}^{(\gamma)}(t)$ and $F_{W_{\phi}}^{(\gamma)}(t)$, which will be compared with $\langle p_j\rangle_{\epsilon}(t)$ and $F_{\epsilon}(t)$ of the GA with imperfections, respectively. In Fig. 1, the results from the STLM are shown as solid lines: they are given by ensemble-averages over $1000$ realizations, respectively. We find that these results from the STLM provide an impressive agreement with the results of the GA after a proper adjustment of $\gamma$ and $W_{\phi}$. This suggests that the main physical ingredients of the disordered GA are correctly incorporated in the STLM. Nevertheless, the origin of the novel feature in the lower envelopes is still unclear. Without loss of generality, $(c_m(t),c_n(t))$ in the STLM during the time evolution can be written by $(e^{-\gamma\,t} \cos \vartheta(t), e^{-\gamma\,t+i\,\phi(t)} \sin \vartheta(t))$ with $\phi(t) := \phi_n - \phi_m$ up to an overall phase. In case of $\phi(t) \equiv 0$ for arbitrary $t$, the angle $\vartheta(t)$ increases by $\omega$ after each iteration and is then given just by $\omega\,t + \vartheta_0$. However, if $\phi(t)$ does not vanish, then from its stochastic nature, it follows that $\vartheta(t)- \vartheta(t-1)$ is not constant but would fluctuate around $\omega$.[@com] Now, under the assumption that $\vartheta(t)$ and $\phi(t)$ are not correlated with each other, but simply two random variables, we can find analytic expressions of $\langle p_j\rangle_{W_{\phi}}^{(\gamma)}(t)$ and $F_{W_{\phi}}^{(\gamma)}(t)$, respectively; let $\vartheta(t)$ increase by $\omega + \eta_{t-1}$ between $t-1$ and $t$ such that $$\vartheta(t)\; =\; \vartheta_0\, +\, \omega\,t\, +\, \sum_{k=0}^{t-1}\,\eta_k\,,$$ where $\eta_k$ forms a Gian distribution with mean 0 and width $\Delta_{\vartheta}$, and then $\sum_{k=0}^{t-1} \eta_k$ also satisfies a Gian distribution with mean $0$ and width $\Delta_{\vartheta} \sqrt{t}$. From this and eq. (\[eq:stlm\]), we get: $$\begin{aligned} \label{eq:p_t} \langle p_j\rangle^{(\gamma)}(t) &=& |w_2(t)|^2\, \left\langle \sin^2 \vartheta(t) \right\rangle\; =\; \frac{|w_2(t)|^2}{\Delta_{\vartheta} \sqrt{\pi t}}\, \int^{\infty}_{-\infty} \sin^2(\omega t + \vartheta_0 + x)\; e^{-x^2/(\Delta^2_{\vartheta}\,t)}\; dx\nonumber\\ &=& \frac{e^{-2\,\gamma\,t}}{2}\, \left[\,1 - \cos(2\,\omega\,t + 2\,\vartheta_0) \cdot e^{-\Delta_{\vartheta}^2\,t}\,\right]\end{aligned}$$ (note that no subindex $W_{\phi}$ appears in $\langle p_j\rangle^{(\gamma)}(t)$). If we further assume that $\phi(t)$ also is of a Gian distribution with mean 0 and width $\Delta_{\phi}$, we then arrive at $$\label{eq:F} F^{(\gamma)}(t)\; =\; \frac{e^{-2\,\gamma\,t}}{2}\, \left[\,1 + e^{-\Delta_{\vartheta}^2\,t}\,\left\{1 - \sin^2(2\,\omega\,t + 2\,\vartheta_0) \cdot \left(1 - e^{-\Delta_{\phi}^2/4}\right)\,\right\}\,\right]\,.$$ Fig. 2 shows a comparison between the results of the GA with imperfections and those of eqs. (\[eq:p\_t\]) and (\[eq:F\]). The good agreement in $\langle p_j \rangle(t)$ would provide an explanation of why its lower envelope is not simply given by $\langle p_j \rangle = 0$ in the GA with imperfections; the uncertainty in the rotation angle during a single iteration accumulates as the iteration proceeds. Then, $\vartheta(t)$ does not represent a definite direction on a 2-dimensional plane but spreads over an interval range $(-\Delta_{\vartheta} \sqrt{t},\, \Delta_{\vartheta} \sqrt{t})$. This offers an additional decay channel into the target state $|j \rangle$ after ensemble-averaging (see the term $e^{-\Delta_{\vartheta}^2\,t}$ in eq. (\[eq:p\_t\])). Also, in eq. (\[eq:F\]) with $\Delta_{\phi} = 0$ we have $F^{(\gamma)}(t) = (e^{-2\,\gamma\,t}/2)\left(1 + e^{-\Delta_{\vartheta}^2\,t}\right)$, which is given by the solid lines in Fig. 2 as the best fit of $F_{\epsilon}(t)$ of the GA with imperfections. From eq. (\[eq:F\]), it immediately follows that $F(\Delta_{\phi} \neq 0)$ is always less than $F(\Delta_{\phi} = 0)$. In summary, we have investigated imperfection effects on the time evolution of the G’s algorithm both numerically and analytically. An effective two-level model with dissipation and randomness has been introduced and the results show a good agreement with the simulation results of the disordered G’s algorithm. It turns out that the main features in the results of the disordered G’s algorithm can be understood through the diffusion-like behavior of quantum states from the original partial H space into the full H space. The two main decaying mechanisms found in this work are its direct manifestations. Our finding will provide a useful basis for study of more general imperfection effects in quantum algorithms. We would like to thank Prof. G. J. Iafrate (NC State Univ.) and Prof. G. Mahler (Univ. of Stuttgart) for critical reading of the manuscript. [1]{} Electronic address: [email protected] A. Steane, Rep. Progr. Phys. [**61**]{}, 117 (1998). P. W. Shor, in [*Proceedings of the $35$th Annual Symposium on the Foundations of Computer Science*]{}, 124 (IEEE, Los Alamitos, CA, 1994). L. K. Grover, Phys. Rev. Lett. [**79**]{}, 325 (1997). A. R. Calderbank and P. W. Shor, Phys. Rev. A [**54**]{}, 1098 (1996). A. Steane, Proc. Roy. Soc. Lond. A [**452**]{}, 2551 (1996). J. Preskill, Phys. Today, June, 24 (1999). D. Gottesman, e-print quant-ph/0004072. J. I. Cirac and P. Zoller, Phys. Rev. Lett. [**74**]{}, 4091 (1995). C. Miquel, J. P. Paz and R. Perazzo, Phys. Rev. A [**54**]{}, 2605 (1996). C. Miquel, J. P. Paz and W. H. Zurek, Phys. Rev. Lett. [**78**]{}, 3971 (1997). P. H. Song and D. L. Shepelyansky, Phys. Rev. Lett. [**86**]{}, 2162 (2001). M. Boyer, G. Brassard, P. Hoeyer and A. Tapp, Fortsch. Phys. [**46**]{} 493 (1998). R. Fazio, G. M. Palma, and J. Siewert, Phys. Rev. Lett. [**83**]{}, 5385 (1999). Here, the leakage for one- and two-qubit operations has been studied quantitatively with the Hamiltonian of the Josephson junction qubits. After some calculations from eq. (\[eq:c\_iteration\]), we actually obtain $\vartheta(t+1) = \frac{1}{2} \cos^{-1}[\cos(2\omega + 2\vartheta(t)) - \sin(2\omega) \cdot \sin(2\vartheta(t)) \cdot (\cos\phi(t) - 1)]$, which is, for general cases, clearly different from $\vartheta(t+1) = \vartheta(t) + \omega$. =3.2in =3.2in

ArXiv

--- abstract: 'We apply Gauge Theory of Arbitrage (GTA) [@hep-th/9710148] to derivative pricing. We show how the standard results of Black-Scholes analysis appear from GTA and derive correction to the Black-Scholes equation due to a virtual arbitrage and speculators’ reaction on it. The model accounts for both violation of the no-arbitrage constraint and non-Brownian price walks which resemble real financial data. The correction is nonlocal and transforms the differential Black-Scholes equation to an integro-differential one.' author: - | Kirill Ilinski$^{1,2}$ [^1]  and  Gleb Kalinin$^{1}$ [^2]\ \[1cm\] [*$^1$ IPhys Group, CAPE, 14-th line of Vasilievskii’s Island, 29*]{}\ [*St-Petersburg, 199178, Russian Federation*]{}\ \[0.2cm\]\ \[0.3cm\] [*$^2$ School of Physics and Space Research, University of Birmingham,*]{}\ [*Edgbaston B15 2TT, Birmingham, United Kingdom*]{} date: title: 'Black-Scholes equation from Gauge Theory of Arbitrage' --- =-1cm =-1cm Introduction ============ The Black-Scholes equation [@BS] for prices of derivative financial instruments is probably the most beautiful rigorous result in the theory of financial analysis. It surprises not only with a simple derivation but also with its high applicability which might be even more important. The derivation is based on several natural simplifications (see Appendix 1 for an one-page introduction to the financial derivatives). Among those the geometrical Brownian motion (or, more general, quasi-Brownian motion) model for the price of an underlying asset and the no-arbitrage constraint are most difficult to avoid. Below we stop to describe each of the assumptions in more detail. Quasi-Brownian walks as a model for price increments keep many convenient features, such as easy-to-fit parameters [@Dubofsky], a developed and handleable mathematical description [@Duffie] which results in the existence of exact solutions obeying explicit equations [@BS; @Wilmott], and indeed resemble real price movements. However, the resemblance is not perfect and many deviations have been observed [@fin; @Mantegna; @BS94; @TLF]. These studies reveal other characteristic processes which substitute for Brownian random walks. These are the Truncated Levy Flights which are regularized Stable Levy processes slowly converging to the Brownian process at large times and large prices [@SM]. Another, more popular, approach to model price movements with excess kurtosis is to use the quasi-Brownian walks but with a stochastic variance (volatility) [@Hobson]. A recent study [@Liu] shows that, at least for the S&P500 index, the volatility is distributed log-normal and follows the geometrical random walk with a good measure of accuracy. However, the last approach cannot explain scaling properties and correlations observed in financial data. A number of approaches have been developed to correct the Black-Scholes analysis for the case of deviations from a simple geometrical Brownian motion model for prices. The well-known and widely accepted way to do it is to generalize the Black-Scholes analysis to quasi-Brownian motions with the stochastic volatility. Other approaches use various versions of risk-neutral treatment which differ by the details of the averaging procedure. A recent paper [@TLF] presents an attempt to account for the Truncated Levy Flight nature of underlying asset price walks based on a Bouchaud and Sornette scheme [@BS94; @risk]. The present paper suggests the different approach to the problem which releases both no-arbitrage assumption and Brownian character of the price walks. As was shown in Ref [@ISfin1] based on the Gauge Theory of Arbitrage (GTA) picture, the active trade behavior of speculators who profit from temporary mispricing of assets may be one of the reasons for the deviation of the probability distribution function (PDF) from the log-normal law and a possible explanation of memory effects. The origin of this is the fact that for a mispriced asset there are cash flows (inflow for a undervalued asset and outflow for an overvalued asset) which shifts the price due to a kind of supply-demand mechanism. This shift is [*directed*]{} and results in damping of the mispricing which, in its own, reduces the width of the PDF peak and increases the peak height. On the other hand, the arbitrageurs create an additive noise which causes the power-law wings of the PDF [@Takayasu; @Sornette]. The investigation of the influence of the directed speculations on the distribution function of prices in a very simple gauge arbitrage model demonstrated a qualitative agreement with the observed data [@ISfin1]. Thus, it was demonstrated that the model distribution function of the price increments is similar to one observed in a real market and possesses the same scaling properties. This motivates us to apply developed formalism to derivative pricing and examine corrections to Black-Scholes equation due to arbitrageurs. Being responsible for both the deviation of the probability distribution function of the underlying asset prices and, at the same time, restoring virtually violated no-arbitrage constraints, speculators must cause corrections to the Brownian walks based arbitrage-free equation for prices of derivatives. The goal of the paper is to derive these corrections in a simple gauge arbitrage model which generalizes the GTA stock exchange model. We will not go into details of GTA here and redirect the reader to Refs [@hep-th/9710148; @ISfin1]. However, to make the exposition self-contained, the main issues of GTA are sketched below. GTA in brief ------------ Let us remind the reader of concepts and notations of GTA [@hep-th/9710148] used in our consideration. This is field-theoretical description of the virtual arbitrage possibilities and corresponding money flows. In this framework the net present value (NPV) calculation and asset exchanges are interpreted in geometrical terms as parallel transports in some sophisticated fibre bundle space. This allows us to map the capital market theory to a theory similar to electrodynamics and then use the machinery of quantum field theory. It was shown that the free quantum gauge theory is equivalent to the assumption of log-normal walks for assets prices. In general, the theory resembles electrodynamics where particles with positive charge (“securities”) and negative charges (“debts”) interact with each other through the electromagnetic field (arbitrage excess return). In the case of a local virtual arbitrage opportunity money flows into the profitable security. Entering positive charges and leaving negative ones screen the profitable fluctuation and restore the equilibrium, i.e. speculators wash out the arbitrage opportunity. The starting point for GTA construction is an understanding of Net Present Value calculations as a parallel transport in some fibre bundle. The NPV states that money has a time value. This time value has to be taken into account through so-called discounting procedure. If an amount of money $F$ is to be received in $T$ years’ time, the Present Value of that amount ($NPV(F)$) is the sum of money $P$ (principal) which, if invested today, would generate the compound amount $F$ in $T$ years’ time (for simplicity $r$ is considered constant over the $T$ years): $$NPV(F) \equiv P = \frac{F}{(1+r)^T} \ .$$ The interest rate involved in this calculation is known as the discount rate and the term $(1+r)^{-T}$ is known as T-year discount factor $D_T$: $$D_T = (1+r)^{-T} \ . \label{DT}$$ Thus NPV method shows how to compare money amounts came at different moments of time [@Lumby]. This last phrase points directly the geometrical interpretation which we use: discounting procedure plays the role of a “parallel transport” of an amount of money through time (though in fixed currency). The discounting factor (\[DT\]) is then an element of a structural group of a fibre bundle and the discount rate coincides with the time component of the connection vector field. The “space” components of the connection are related to exchange rates and prices. Indeed, exchange rates and prices are responsible for converting money in different currencies or different securities, i.e. points of discrete “space”, to the same currency (point of the space) at a fixed moment of time. They can be interpreted as elements of the structural group which “transport” the money in “space” directions and are space analogues of the discount factor. Summing up, the capital market theory has a geometrical structure which allows us to map it onto a theory of a fibre bundle. The curvature tensor of the connection field is related to the arbitrage which is an operational opportunity to make a risk-free profit [@Wilmott] with a rate of return higher than the risk-free interest rate accrued on deposit. As was derived in Refs [@hep-th/9710148], the rate of excess return on an elementary arbitrage operation (a difference between rate of return on the operation and the risk-free interest rate) is an element of the curvature tensor calculated from the connection. It can be understood keeping in mind that the curvature tensor element is related to a difference between two results of infinitesimal parallel transports performed in different order with the same initial and final points or, in other words, a gain from an arbitrage operation. Due to this geometrical interpretation it is possible to say that the rate of excess return on an elementary arbitrage operation is an analogue of the electromagnetic field. In the absence of any uncertainty and money flows, the only state that is realized is the state of zero arbitrage. However, if we introduce the uncertainty to the game, prices and rates move and some virtual nonequivalent possibilities to get more than less appear. Therefore we can say that the uncertainty play the same role in the developing theory as the quantization did for quantum gauge theory. Money flow fields appear in the theory as “matter” fields which are transported by the connection (interests and exchange rates). It means that the matter fields interact through the connection. Dilatations of money units (which do not change real wealth) play the role of gauge transformation which eliminates the effect of the dilatation by a corresponding gauge transformations of the connection in the same way as the Fisher formula does for the real interest rate in the case of an inflation [@Lumby]. [*The symmetry of the real wealth to a local dilatation of money units, security splits and the like is the gauge symmetry of the theory*]{}. An investor’s strategy is not always optimal. This is due to partially incomplete information available, partially because of an investor’s internal objectives [@Lumby]. It means that the money flows are not certain and fluctuate in the same manner as prices and rates do. So this requires a statistical description of money flows which, once again, returns us to an effective quantization of the theory. At this stage we would like to clarify the following misunderstanding which might emerge. The arbitrage itself implies a possibility to perform an operation with a risk-free rate of return which is higher than, say, a bank deposit interest rate. In this sense buying shares cannot be considered as a such operation because of assumed random walk of the share price and the corresponding risk. What do we mean then talking about the arbitrage? The randomness of the price is equivalent to a quantization as we explained in [@hep-th/9710148] and the rate of return on an (arbitrage) plaquette operation is now a quantum variable which cannot be taken as a complex number. This exactly resembles a situation with an electromagnetic field which, after the quantization, is not a number but a quantum variable. However, it does not stop us using the same name for the variable, imagining virtual quantum fluctuations and describing the influence of these fluctuations on electric charges, keeping in mind the calculation of corresponding matrix elements. In the same way we understand the arbitrage rate of return in the financial setting. Summing up, it was shown how to map the capital market to a system of particles with positive, “securities”, and negative, “debts”, charges which interact with each other through an electromagnetic field, the gauge field of the arbitrage. In the case of a local virtual arbitrage opportunity, cash flows into the region of configuration space (money go in the profitable security) while “debts” try to escape from the region. This brings in positive charges and pushes out negative ones, leading to an effective screening of the profitable fluctuation. These processes restore an equilibrium and erase the arbitrage opportunity. Formalization of the scheme drawn above leads to the lattice quantum field theory [@Creutz]. At this point the standard machinery of quantum field theory can be applied to obtain various observables such as distribution functions of the interest/exchange rates, response functions of the system and others. It may answer questions about dynamical response of a financial market, the dynamical portfolio theory and other problems. In conclusion we want to add that notions of the (stochastic) differential geometry appeared in the context of financial modelling in paper [@Hughston]. It contains several ideas which are similar to GTA. In this paper we want to apply the idea to derivative pricing. More precisely, we show how the standard results of Black-Scholes analysis appeared from GTA and derive correction to the Black-Scholes equation due to a virtual arbitrage and speculators reaction on it. This will model both violation of the no-arbitrage constraint and non-Brownian price walks which resemble recently observed data. The paper is organized as follows. In next section we formulate a GTA model for a description of derivative instruments. To do this we construct base space of the theory which, in contrast to GTA model for the simplest stock exchange [@ISfin1], contains a double ladder. This corresponds to simultaneously treating cash, shares and derivatives. Section 3 is devoted to an investigation of the classical limit of the action which leads to the Black-Scholes equation in the particular case of free gauge field dynamics. We show that the plaquette diagrams obtained in Section 2 have a very simple interpretation as an imbalance between the derivative and the Black-Scholes hedging portfolio. Section 4 is devoted to description of the money flow fields and the corresponding correction to an effective action for the derivative price coming from virtual speculations. In the classical limit it gives a correction to the Black-Scholes equation. The last section completes the paper with final remarks. In the appendix we give simple derivation of the Black-Scholes equation. GTA model of derivatives ======================== In this section we construct a GTA model for a share-cash-derivative system. To simplify the consideration we consider only one type of shares and the perfect capital market conditions are implied. The consideration of this paper is not restricted by any type of concrete derivative contracts. However, to simplify the consideration we will illustrate the GTA application by an analysis of European and American call options. Shares or derivatives can be exchanged with cash and vice versa. The corresponding exchange rates are $S_i$ and $C_i$ (one share or derivative contract is exchanged on $S_i$ or $C_i$ units of cash) at some moment $t_i$, and the reverse rates (cash to share or derivative) are $S^{-1}_i$ and $C^{-1}_i$. We consider period from starting point $t=0$ up to moment $t=T$. For the case of option we assume that $T$ is expiration time. We suppose that there exists a shortest interval of time $\Delta=T/N$ and this $\Delta$ is taken as a unit time. So, the exchange rates $S_i$ and $C_i$ are quoted on a set of equidistant times: $\{t_i\}_{i=0}^N, t_i =i\Delta$ and represent the parallel transport along legs of a double ladder base graph. The interest rate for cash is $r_b$ so that between two subsequent times $t_i$ and $t_{i+1}$ the volume of cash is increased by factor $e^{r_b \Delta}$. The shares and derivatives are characterized by rates $r_1$ and $r_2$ correspondingly. These rates realize parallel transport in time direction. Following Refs [@hep-th/9710148] we consider elementary arbitrage operations when an arbitrageur borrows one share at time $t_i$, sells it for $S_i$ units of cash, put the cash in the bank until time $t_{i+1}$ and, at time ($t_{i+1}$), closes his short position borrowing $e^{r_1\Delta}S_{i+1}$ units of cash and buying shares. The result of the operation for the arbitrageur will be $e^{r_b\Delta}S_{i}-e^{r_1\Delta}S_{i+1}$ units of cash. The excess return on this operation is $$Q^{(1)}_i = S_i e^{r_b\Delta}S^{-1}_{i+1} e^{-r_1\Delta} - 1\ . \label{npv1}$$ To get this expression we discounted the amount and converted it in shares since the operation was started in the shares. Equation (\[npv1\]) has a form of the curvature tensor element corresponding to drawing assets through the cycle. If $Q^{(1)}_i\ne0$ an arbitrageur can get excess return performing this or the reverse operation. The following quantity $$R^{(1)}_i = (S^{-1}_i e^{r_1\Delta}S_{i+1} e^{-r_b\Delta} + S_i e^{r_b\Delta}S^{-1}_{i+1} e^{-r_1\Delta} - 2)/\Delta \label{eq:1}$$ is used to measure the arbitrage (excess rate of return) on local cash-share operation. The absences of the arbitrage is equivalent to the equality $$S^{-1}_i e^{r_1\Delta}S_{i+1} e^{-r_b\Delta} = S_i e^{r_b\Delta}S^{-1}_{i+1} e^{-r_1\Delta} = 1\ .$$ The same can be done for another possible arbitrage operation. For cash-derivative plaquette it gives the following quantities $$Q^{(2)}_i = C_i e^{r_b\Delta}C^{-1}_{i+1} e^{-r_2\Delta} - 1\ , \label{npv2}$$ $$R^{(2)}_i = (C^{-1}_i e^{r_2\Delta}C_{i+1} e^{-r_b\Delta} + C_i e^{r_b\Delta}C^{-1}_{i+1} e^{-r_2\Delta} - 2)/\Delta \ . \label{eq:2}$$ Looking at (\[eq:1\],\[eq:2\]) we can conclude that the arbitrage is represented in the theory by the curvature of the connection. Precisely, in the continuous limit ($\Delta\to0$) the RHS of Eqns (\[eq:1\],\[eq:2\]) converges as usual to a square of the curvature tensor element multiplied by area of plaquette $$R^{(l)}_i = (Q^{(l)}_i)^2/\Delta\ . \label{pq}$$ Curvature is $Q^{(l)}_i/\Delta$, the area of plaquette is proportional to the shortest time interval $\Delta$ and fixed “space” length which is omitted. Being represented by the curvature tensor, the notion of the arbitrage as well as the quantity $R^{(l)}_i$ are gauge invariant. All of this allows us to say that the rate of excess return on an elementary arbitrage operation is an analogue of the electromagnetic field. In the absence of uncertainty (or, in other words, in the absence of random walks of prices, exchange and interest rates) the only state realized is the state of zero arbitrage. However, if we introduce the uncertainty, prices and the rates move and some virtual arbitrage possibilities appear. Therefore, we can say that uncertainty plays the same role in the developing theory as the quantization does for the quantum gauge theory. Further development of the consideration are based on the following assumptions about the dynamics [@hep-th/9710148]: gauge invariance, locality, correspondence principle, extremal action principle, limited rationality (uncertainty) and absence of correlation between excess returns on different plaquettes. Formally, these assumptions are summed up in the functional form of the probability $P(\{S_i,r_k\})$ to find a set of the exchange rates/interest rates $\{S_i,r_k\}$ given by the expression: $$P(\{S_i,r_k\}) \sim e^{-\sum_{i,l} \beta_l R^{(l)}_i} \sim e^{-s_{gauge}}\ , \label{P0}$$ together with the statement that the integration measure is gauge invariant too. The introduced above parameters $\beta_l$ are measures of the uncertainty of the corresponding plaquettes. The sums run over time moments $i$ and types of the plaquettes ($l=1$ for cash-share plaquettes and $l=2$ for cash-derivative ones). As it was shown in Refs[@hep-th/9710148], the assumptions which have been made are equivalent to a log-normal model for the share price walks with the distribution function $$P(S(T)|S(0)) = \frac{1}{\sigma S \sqrt{2\pi T}} e^{-(\ln (S(T)/S(0)) - (\mu - \frac12 \sigma^2)T)^2/(2\sigma^2 T)} \ . \label{log}$$ Here the volatility $\sigma=1/\sqrt{2\beta}$ and the average rate of share return $\mu=r_b-r_1$ have been introduced. There are two points to note here. The first one is that the log-normal distribution was derived in an absence of matter fields. These matter fields can significantly change the form of the distribution function and other properties of the price random walk. Indeed, in the presence of the money flows more complicated random processes for the price motion emerge which resemble real data observations [@ISfin1]. The second point concerns other types of quasi-Brownian price motions which are considered in mathematical finance. They can be also introduced in the theory, making parameters, which we keep constant for the moment, depending on price values. Let us return to gauge fixing. Since the action $s_{gauge}$ is gauge invariant it is possible to perform a gauge transformation which will not change the dynamics but will simplify further calculation. In lattice gauge theory [@Creutz] there are several standard choices of gauge fixing and axial gauge fixing is one of them. In the axial gauge an element of the structural group is taken as something chosen on links in the time direction and exchange rates along the “space” direction at some particular time. This kind of gauge fixing is convenient for the model in question. Actually this gauge has been used in deriving (\[log\]). We choose $r_b$ to be the risk free interest rate and $r_b-r_1$, $r_b-r_2$ are the average rates of return on the share and the derivative. This means that in the situation of the double ladder base the only dynamical variables are the exchange rates (prices) as a function of time and the corresponding measure of integration is the invariant measure $\frac{\mbox{d}S_i\mbox{d}C_i}{S_i C_i}$. Below we fix the price of the shares at time $t=0$ taking $S_0=S(0)$. We also fix the exchange rate of the derivative to the share at the moment of the derivative exercise. Let us note that quantities of our theory — exchange and interest rates are not gauge invariant but gauge covariant. So it is natural to choose the gauge in which the exchange and interest rates take their real value. In our gauge “rate of return” on cash takes its real value $r_b$ while average rates of return on the share and the derivative are not $r_1$ and $r_2$ but $r_b-r_1$ and $r_b-r_2$. This provides the axial gauge fixing together with fixing the price of the shares at the time $t=0$ and the exchange rate of the derivative to the share at the time of exercise. Let us return to expression (\[P0\]). It is derived under the assumption that arbitrage opportunities are uncorrelated between different “space-time” plaquettes. However, there are many important problems where the correlation between returns on elementary plaquettes operations have to be taken into account. The most important such example is the portfolio theory where an optimal portfolio is constructed using correlation between assets [@Blake; @portfolio]. Another example is derivative pricing which is studied in the paper. To account for the correlation we have to elaborate some details of the construction of the action (the definition of the probability finding particular configuration of exchange and interest rates). Of course we retain base principles such as gauge invariance. The configuration which has less arbitrage is more probable. But we look more precisely at the problem of independence of arbitrage operations. The general form of the action is $$s_{gauge} = \sum_{\xi\zeta} Q_\xi A_{\xi\zeta} Q_\zeta /(2\Delta) \label{gen-act}$$ where $Q_\xi$ are local (dependent) arbitrage plaquette quantities and matrix $A_{\xi\zeta}$ is the correlation matrix of the plaquettes. This scheme is general and can be applied to the case of many correlated assets. Expression (\[gen-act\]) is gauge invariant and naturally generalize (\[P0\]). In general, matrix $A_{\xi\zeta}$ is not diagonal due to presence of correlations . To simplify the model we make a locality assumption. It means that the virtual arbitrage opportunities emerging at different times are independent. This makes matrix $A$ diagonal in the time index. The action can be rewritten in following form: $$s_{gauge} = \sum_{ill'} Q^{(l)}_i A_{ll'} Q^{(l')}_i /(2\Delta) \label{action-general}$$ where $A_{ll'}$ is equal-time plaquette correlation matrix and $l$ and $l'$ run over all elementary plaquettes at fixed time. It is straightforward to show that in the continuous limit ($\Delta\to0$) the previous expression takes the following form $$s_{gauge} = \int_0^T\sum_{ll'} \left(\frac1S \frac{dS_l(t)}{dt} - (r_b - r_l)\right) \frac{A_{ll'}(t)}{2} \left(\frac1S \frac{dS_{l'}(t)}{dt} - (r_b - r_{l'})\right)dt \ . \label{action-general-cont}$$ This implies the following expression for the correlation matrix $A$: $$A^{-1}_{ll'}(t) = <\frac1S dS_l(t), \frac1S dS_{l'}(t) >/dt \equiv \sigma^2_{ll'}(t) \label{A}$$ (terms $(r_b-r_l)dt$ do not contribute in continuous limit). For the share-cash-derivative system we need to consider the following correlators: $<\frac{1}{S(t)} dS(t), \frac{1}{S(t)} dS(t)>/dt$, $<\frac{1}{C(t)} dC(t), \frac{1}{S(t)} dS(t)>/dt$, $<\frac{1}{C(t)} dC(t), \frac{1}{C(t)} dC(t)>/dt$. The first one is equal to $\sigma^2 \equiv 1/(2\beta_1)$. The second correlator we denote as $\alpha(t)/(2\beta_1)$ introducing the notation: $$\alpha(t) \equiv <\frac{1}{C(t)} dC(t), \frac{1}{S(t)} dS(t)>/ <\frac{1}{S(t)} dS(t), \frac{1}{S(t)} dS(t)> \ . \label{alfa}$$ To calculate the last correlator $<\frac{1}{C(t)} dC(t), \frac{1}{C(t)} dC(t)>/dt $ we consider $C(t)$ as a random function of $S$. Then it can be represented as $$<\frac{1}{C(t)} dC(t), \frac{1}{C(t)} dC(t)>/dt = \alpha^2(t)/(2\beta_1) + 1/(2\beta_2)$$ where the term $\alpha^2(t)/(2\beta_1)$ is responsable for the volatility of the derivative due to its dependence on the share price and $1/(2\beta_2)$ stands for the volatility of the derivative due to the random character of function $C$ itself. Given these parameterization with $\alpha$, $\beta_{1,2}$ we can find matrix $A$ from Eq(\[A\]) and get the following expression for the action (\[action-general-cont\]) in this particular derivative-related setting: $$\begin{array}{rl} s_{gauge} =& \beta_1 \int_0^T dt \left(\frac{1}{S(t)}\frac{dS(t)}{dt} + (r_1 - r_b) \right)^2 + \\&+ \beta_2 \int_0^T dt \left( \left(\frac{1}{C(t)}\frac{dC(t)}{dt} + (r_2 - r_b) \right) - \alpha(t) \left(\frac{1}{S(t)}\frac{dS(t)}{dt} + (r_1 - r_b) \right) \right)^2 \ . \end{array} \label{action-cont}$$ It is interesting to note that this expression gives the main order (with respect to $\Delta$) of the lattice action $$s_{gauge} = \sum_i (\beta_1 R^{(1)}_i + \beta_2 R^{(2,1)}_i) \ . \label{action-discr}$$ where $R^{(2,1)}$ is defined as $$R^{(2,1)}_i = (Q^{(2)}_i - \alpha_i Q^{(1)}_i)^2/\Delta \ . \label{p21}$$ Action (\[action-discr\]) has a very simple interpretation. As we already mentioned, local cash-share and cash-derivative arbitrage operations (at the same time) clearly are not independent and we cannot use $R^{(1)}$ and $R^{(2)}$ simultaneously to characterize the independent arbitrages. Instead, we can find statistically independent combinations of $Q^{(1)}_i$ and $Q^{(2)}_i$ and define $R^{(2,1)}$ as in (\[p21\]) to use together with $R^{(1)}_i$ as independent plaquette quantities in expression (\[P0\]). This returns us to the action (\[action-discr\]). This is the action we deal with in the paper when we consider a lattice system while we use action (\[action-cont\]) for the continuous limit. All said above is valid for any derivative contract. A particular type of derivative is reflected by details of the construction of the base graph of the theory. We consider here in more detail European and American call options. Other kinds of derivatives require other base constructions which, however, are straightforward. We start with the European call option (see Appendix 1 for the definition). As we said before, the “space”-time graph for it and the underlying share is the double ladder. The only new element here, comparing with the model of Ref [@ISfin1] adopted for two types of shares, is the additional link at the expiration date between the option and cash and the share with the direction chosen to describe the possibility to change this option and $E$ units of cash on the share. (Fixed exchange rate on this link gives missing gauge fixing condition.) The graph for the American call option and the underlying share contains the double ladder and the directed links from the option and cash to share at any intermediate times before the expiration time $T$. These links are the same as that for European option. The links which are responsible for the swap of the derivative to the share, i.e. exercising of the option, generate new plaquettes on the “space”-time graph. Corresponding plaquette quantities $R'_i$ have to be taken into account. These terms depend very much on concrete conditions of the derivative contract. We keep the terms terms aside and rewrite (\[P0\]) in the form $$P(\{S_i,r_k\}) \sim e^{-s_{gauge}-\sum_i\beta'_i R'_i}\ . \label{P'}$$ Below we do not concentrate on the primed plaquettes. They contain directed links and contribute to the boundary conditions only. However, to make complete analytical and numerical analysis they are important and have to be retained. We return to this point in section 3 where we derive the boundary conditions. In the next section we show how this formalism reproduces all the results of standard derivative pricing theory. Derivation of the Black-Scholes equation ======================================== In this section we obtain the Black-Scholes equation as the equation of the saddle-point in the quasi-classical limit of the gauge theory in absence of money flows. Let us return to Eqn(\[action-cont\]) for the action in the continuous ($\Delta\to0$) limit. The first term in the RHS of this equation corresponds to geometrical random walks and provides a background for the derivation. Being a derivative instrument from the share price, the price of the derivative has to be correlated or even defined by the share price. That is why it is natural and more convenient for our purposes to write $C(t)$ as a some unknown function of $S(t)$. Since we integrate over all $C(t)$ (or over all functions $C(t,S)$) this does not mean any loss of generality. We use the fact that $C$ is a function of $S$ and the property of the geometrical random walk of the underlying asset to obtain a compact expression for the parameter $\alpha(t)$ (\[alfa\]): $$\alpha = \frac SC\frac{\partial C}{\partial S} \ . \label{alpha}$$ To explain this we use the following fact for the geometrical Brownian motion known as Ito’s lemma: $$d f(t,S) = \frac{\partial f}{\partial t}dt + \frac{\partial f}{\partial S} dS + \frac{S^2\sigma^2}{2}\frac{\partial^2 f}{\partial S^2} dt + o(dt) \label{Ito}$$ where $f(t,S)$ is an arbitrary function of $t$ and $S$ (the function is supposed to be smooth) [@Wilmott]. Using this equality and the fact $<dS/S,dS/S>\to S^{-2}<dS,dS>$ we obtain Eqn(\[alpha\]). Now let us turn our attention to the second term in Eqn(\[action-cont\]). Using expression (\[r2\]) and Ito’s lemma (\[Ito\]) we end with the following action term: $$\beta_2 \int dt \left(\frac1C \frac{\partial C}{\partial t} + \frac1C \frac{\sigma^2}{2} S^2 \frac{\partial^2 C}{\partial S^2} - r_b (1 - \frac SC\frac{\partial C}{\partial S}) \right)^2 \ . \label{aa}$$ This term corresponds to a virtual arbitrage account. The classical limit for the action, i.e. $\beta_2\to\infty$, reduces the functional integration over functions $C(t,S)$ to [**the contribution from the classical trajectory only and this trajectory is defined by the Black-Scholes equation**]{} for the price of the financial derivative: $$\frac{\partial C}{\partial t} + \frac{\sigma^2}{2} S^2 \frac{\partial^2 C}{\partial S^2} - r_b (C - S\frac{\partial C}{\partial S}) = 0 \ , \label{BS}$$ since, as it is easy to see, the rate $r_2$ on the derivative whose price is defined by (\[aa\]) is give by the expression $$r_2 = r_1 <\frac SC \frac{\partial C}{\partial S}> \ . \label{r2}$$ The equation does not depend on a particular type of the derivative which is encoded in the boundary conditions. For a comparison we give standard no-arbitrage derivation of the Black-Scholes equation in the appendix. Boundary conditions ------------------- In this subsection we consider boundary conditions which characterize a particular derivative. Here we treat only European and American calls but it is straightforward to generalize the consideration. To this end we return to the construction of the base graph in section 2. Let us first consider in more detail the European call option. At the expiration time $T$ the option can be sold or bought for $C(T)$ units of cash but it can also be exchanged (together with $E$ units of cash) on the share. If the price of the share $S$ is less than $E$ this leads to effective scrap of the call. These opportunities create new arbitrage possibilities which we take into account by introducing the $\beta'R'_N$ term in action (\[P’\]). There exists a new arbitrage operation which is available at expiration time $T$: one can borrow the portfolio consisted of option and $E$ currency units, exchange it on the share if $S(T)>E$ (directed link on the base graph), sell share for $S(T)$ units of money and buy the portfolio again. This gives us a plaquette excess return $$Q'_N = (S(T)\theta(S(T)-E) + E\theta(E-S(T)))(C(T)+E)^{-1} - 1\ ,$$ and for the plaquette quantity $R'_N$ we get $$R'_N = \frac{E\theta(E-S(T)) + S(T)\theta(S(T)-E)}{C(T)+E} + \frac{C(T)+E}{E\theta(E-S(T)) + S(T)\theta(S(T)-E)} - 2\ . \label{boundary}$$ In the quasi-classical (no-arbitrage) limit we have $\beta'\to\infty$ which gives the following boundary condition for the European call option from Eqn(\[boundary\]): $$C(T) = (S(T)-E)\theta(S(T)-E) \ . \label{eur-bound}$$ If we neglect this arbitrage possibilities at the time $T$, we can use this equation as a gauge fixing condition and need not consider additional links and boundary plaquettes. This approximation is quite realistic and clearly simplifies the scheme. Let us turn to the American call option. It can be exchanged for the share at any time up to $T$ by paying additional $E$ units cash. In this case the base graph contains links from the option to the share (but not backwards) at all time points. It allows two possible arbitrage operations. The first operation is similar to considering above operation with a European option at exercise time $T$. Now this operation is available at any moment $t\leq T$. One can borrow the portfolio consisting of an option and $E$ currency units, exchange it on the share, sell the share for $S(t)$ units of money and buy the portfolio again. The excess return on this operation is $$Q'_t = S(t) (C_A(t)+E)^{-1} - 1\ ,$$ and for the boundary plaquette quantity we get $$R'_t = \frac{S(t)}{C_A(t)+E} + \frac{C_A(t)+E}{S(t)} - 2\ . \label{R'-am1}$$ In close analogy with the European option in the quasi-classical (the absence of the arbitrage) limit we get the following boundary condition for the American call option: $$C_A(t) = S(t)-E \ . \label{am-bound2}$$ Now we have to determine moment when American option exercised. To this end we consider another possible arbitrage operation. Let an arbitrageur have portfolio of an option and $E$ units of cash at some time $t$. He can exchange the portfolio for a share and keep it up to time $t+dt$ or hold the portfolio and exchange it on a share at time $t+dt$. The expected return on the portfolio is $(C_A(t)(r_b-r_2)+Er_b)/(C_A(t)+E)$ so at time $t+dt$ an arbitrageur will have $$[1+\frac{C_A(t)(r_b-r_2)+Er_b)}{C_A(t)+E}dt]\frac{C_A(t)+E}{C_A(t+dt)+E}$$ portfolios of an option and $E$ cash units which he can exchange for the equal number of shares. The expected return on share is $(r_b-r_1)$. Therefore the excess return on this operation is $$Q''_t = \biggl[1+\frac{C_A(t)(r_b-r_2)+Er_b)}{C_A(t)+E}dt\biggr] \frac{C_A(t)+E}{C_A(t+dt)+E} \left( [1+(r_b-r_1)dt] \frac{S(t)}{S(t+dt)} \right)^{-1} \ ,$$ and for the boundary plaquette quantity we obtain the expression: $$R''_t = Q''_t + (Q''_t)^{-1} - 2\ . \label{R''-am1}$$ In the no-arbitrage limit we have $$Q''_t -1 = 0$$ at the exercise time. This gives us second boundary condition for the American call: $$\frac{\partial C_A(t,S)}{\partial S} = 1 \ . \label{am-bound1}$$ To prove this expression we have to use the no-arbitrage condition $Q''_t -1= 0$, expression for $r_2$ (\[r2\]) $$r_2 = \frac{S}{C}\frac{\partial C}{\partial S} r_1$$ and to take into account first boundary condition (\[am-bound2\]). So finally we obtain the following boundary conditions for the American call option [@Wilmott]: $$\frac{\partial C_A(t,S)}{\partial S}=1 \ ,\qquad{C_A}(t,S) = S(t) - E \ . \label{am-bound}$$ The equations determine the time when the option is exchanged for the share and the corresponding payoff. It is easy to show from the plaquette analysis that in absence of dividends American call is never exercised early and, hence, is equivalent to European call but we do not stop for it. Summing up, we have shown that in the absence of money flows and in the quasi-classical limit (i.e. suppression of the arbitrage operations with the derivatives) the Black-Scholes equation emerges as the equation for the saddle-point which is provided with appropriate boundary conditions. Connection with Black-Scholes analysis -------------------------------------- In conclusion we want to clarify the connection with the original Black-Scholes analysis (see Appendix). Let us return back to the formula for $R^{(2,1)}$ (\[p21\]): $$R^{(2,1)}_i = (Q^{(2,1)}_i)^2/\Delta$$ with $Q^{(2,1)}_i$ defined as $$Q^{(2,1)}_i\equiv Q^{(2)}_i - Q^{(1)}_i \alpha_i \ .$$ To first order in $\Delta$ the last expression can be rewritten as $$Q^{(2,1)}_i = C_i S_i^{-\alpha_i} e^{\alpha_i r_1\Delta} e^{(1-\alpha_i)r_b\Delta} e^{-r_2\Delta} S_{i+1}^{\alpha_i} C_{i+1}^{-1} - 1 \ .$$ It is easy to give a simple interpretation of the last expression. Indeed, it is not difficult to see that it describes the following money circle: $$C \to USD \to (1-\alpha) USD \oplus \alpha S \to USD \to C \ .$$ The portfolio comprising shares and cash (in parts $\alpha$ and $1-\alpha$) emerging at the intermediate state is the Black-Scholes hedging portfolio for the derivative and the condition (\[correl\]) is exactly the hedging relation. From this point of view the plaquette $Q^{(2,1)}_i$ represents the arbitrage fluctuations in hedging portfolio-derivative plaquette. This returns us back to the treatment of the expression $$(\frac{\partial C}{\partial t} + \frac{\sigma^2}{2} S^2 \frac{\partial^2 C}{\partial^2 S} - r_b (C - S\frac{\partial C}{\partial S}))/ (C- S\frac{\partial C}{\partial S})$$ as the arbitrage excess return on the infinitesimal operation [@Wilmott]. Money flows and correction to the Black-Scholes equation ======================================================== Now let us turn to money flow fields. These fields represent cash-debt flows in the market. The importance of the cash-debt flows for our consideration is explained by their role in a stabilization of market prices. Indeed if, say, some asset price creates a possibility to get a bigger return than from other assets (with similar risk), then an effective cash flow appears, directed to these more valuable shares. This causes restoration of equilibrium due to the demand-supply mechanism. The same picture is valid for debt flows if there is a possibility for debts restructurisation. As we will see all these features find their place in the GTA framework. Following Refs [@hep-th/9710148] we can formulate the dynamics for the cash-debts flows basing on several assumptions, such as gauge invariance of the dynamics, an investor’s wish to maximize his return and his limited rationality. This gives us the following functional integral representation for the matrix element of an evolution operator in the coherent state representation of the money flows in the case of our double ladder base graph ($r_0\equiv r_b$): $$<\bar{\psi}_N,\bar{\chi}_N| \hat{U}(t+N\Delta,t)|\psi_0,\chi_0> = \int \prod_{k=0,i=1}^{2,N-1}d\bar{\psi}_{k,i} d\psi_{k,i} d\bar{\chi}_{k,i}d\chi_{k,i} e^{(s1 + s1'+ s_b)} \ , \label{int}$$ with the actions for cash and debt flows: $$s1= \sum_{k=0,i=0}^{2,N-1} ( \bar{\psi}_{k,i+1} e^{\beta r_k\Delta} \psi_{k,i} - \bar{\psi}_{k,i}\psi_{k,i}) + \sum_{i=0}^{N-1} ( (1-t_c)^{\beta}C_i^{\beta} \bar{\psi}_{0,i+1}\psi_{2,i}$$ $$+ (1-t_c)^{\beta} C_i^{-\beta} \bar{\psi}_{2,i+1} \psi_{0,i}) + (1-t_c)^{\beta}S_i^{\beta} \bar{\psi}_{0,i+1} \psi_{1,i} + (1-t_c)^{\beta} S_i^{-\beta} \bar{\psi}_{1,i+1} \psi_{0,i}) \ , \label{s1}$$ $$s1' = \sum_{k=0,i=0}^{2,N-1} (\bar{\chi}_{k,i+1} e^{-\beta r_k \Delta} \chi_{k,i} - \bar{\chi}_{k,i}\chi_{k,i}) + \sum_{i=0}^{N-1} ((1+t_c)^{-\beta}C_i^{-\beta} \bar{\chi}_{0,i+1} \chi_{2,i}$$ $$+ (1+t_c)^{-\beta}C_i^{\beta} \bar{\chi}_{2,i+1} \chi_{0,i} +(1+t_c)^{-\beta}S_i^{-\beta} \bar{\chi}_{0,i+1} \chi_{1,i} + (1+t_c)^{-\beta}S_i^{\beta} \bar{\chi}_{1,i+1} \chi_{0,i} ) \ . \label{s1'}$$ Here $t_c$ is a relative transaction cost and the action $s_b$ represents exchanges along boundary plaquettes and is contract-dependent. Using this expression it is easy to obtain the transition probability in the occupation number representation simply by integrating over $\bar{\psi}$, $\psi$, $\bar{\chi}$, $\chi$ variables. Let us consider a transition to the state with $n_0 (n'_0)$ cash units in the the long (short) position, $n_1 (n'_1)$ stock units in the long (short) position, $n_2 (n'_2)$ derivatives units in the long (short) position at time $t_1=t+N\Delta$ from the state ($m_k$,$m'_k$) at the original time $t$. The long position corresponds to possessing assets and the short position implies possessing liabilities (“debts”). The probability for this transition has been derived in [@ISfin1]: $$P((\{n_k\},\{m_k\}),(\{n'_k\},\{m'_k\}),t_1,t) = S^{\beta (m_1 -m'_1 -n_1+n'_1)}(t) C^{\beta (m_2 -m'_2 -n_2+n'_2)}(t)$$ $$\int d\psi d\bar{\psi} d\chi d\bar{\chi} <\bar{\psi}_N,\bar{\chi}_N| \hat{U}(t+N\Delta,t)|\psi_0,\chi_0> e^{-\bar{\psi}_{N}{\psi}_{N}} \prod_{k=0}^{2} \bar{\psi}^{m_k}_{k,0} \bar{\chi}^{m'_k}_{k,0} {\psi}^{n_k}_{k,N} {\chi}^{n'_k}_{k,N} /{n_k!n'_k!} \ . \label{int1}$$ Expressions (\[action-discr\],\[P’\],\[int\],\[int1\]) form the complete set of necessary equations to describe the dynamics and mutual influence of interest/price rates and money fields. We now derive the correction to the effective action of the gauge field due to the presence of money flows. Since, as it was shown in the previous section, the Black-Scholes equation appears in the quasi-classical limit of the free gauge theory, the correction to the effective action leads in the same limit to a correction to the Black-Scholes equation. Let us suppose that the initial asset configuration at time $t$ is given by the probability distribution $F(\{m_k\},\{m^{\prime}_k\})$. Then, as it follows from section 2, the correction to the effective action for the gauge field is equal to: $$\delta s_g = \ln\sum_{\{n_k\},\{n'_k\},\{m_k\},\{m'_k\}} P((\{n_k\},\{m_k\}),(\{n'_k\},\{m'_k\}),0,T) F(\{m_k\},\{m^{\prime}_k\})$$ where the function $P((\{n_k\},\{m_k\}),(\{n'_k\},\{m'_k\}),0,T)$ is given by Eqns(\[int\] - \[int1\]). The generalization of this expression to many investment horizons is straightforward but cumbersome. The correction to the effective action of the prices’ gauge field transforms the Black-Scholes equation for the classical trajectory to the following equation for the trajectory $C(t)$ minimizing the action: $$min_{C(\cdot)} \left( \int dt \left(\frac1C \frac{\partial C}{\partial t} + \frac1C \frac{\sigma^2}{2} S^2 \frac{\partial^2 C}{\partial^2 S} - r_b (1 - \frac SC\frac{\partial C}{\partial S}) \right)^2 - \frac{1}{\beta_2}\delta s_g \right) \ . \label{BS1}$$ Then the Black-Scholes equation is formally substituted by the equation: $$\frac{\delta}{\delta C(t)} \left( \int dt \left(\frac1C \frac{\partial C}{\partial t} + \frac1C \frac{\sigma^2}{2} S^2 \frac{\partial^2 C}{\partial^2 S} - r_b (1 - \frac SC\frac{\partial C}{\partial S}) \right)^2 - \frac{1}{\beta_2}\delta s_g \right) = 0 \label{BS2}$$ with the same boundary conditions. However, we think that the former form of the problem (Eqn(\[BS1\])) is more convenient for numerical solution while the latter (Eqn(\[BS2\])) cannot provide much for analytical inside since even in the more simple situation of the GTA model of stock exchange [@ISfin1] an analytical approach hardly produce valuable results. However we can state general properties of the correction which can be derived in the same way as was done for the correction to the effective action for the price of shares [@ISfin1]. It was shown there that the correction from money flows has two important properties: 1. The correction vanishes at large time intervals; 2. The correction disappears in a limit of completely noisy traders who do not consider their potential profit as a motivation for transactions, i.e. in the limit $\beta\to0$; 3. The correction is governed by a number of traders (read as money available for arbitrage operations), i.e. the correction disappears when no money is available for arbitrage operations. This results in the convergence of the price movement process to the geometrical Brownian walks at large time scale which is well-know from real data observation. “Physically” it means that for large times the arbitrage is washed out (if there are arbitrageurs and they prefer to get more than less), no-arbitrage constraint holds firm and the Black-Scholes description is correct. We state the central result of the paper as a derivation of the Black-Scholes equation and the corresponding corrections. However we want to emphasize that [*to account for both the virtual arbitrage and the money flows one needs to calculate the functional integral over money fields and prices keeping all action terms, in particular boundary plaquettes*]{}. Conclusion ========== In conclusion, in the paper we have shown how to adapt the Gauge Theory of Arbitrage to apply it to derivative pricing. This can be done by constructing base graphs for the theory which allows exchange of the derivatives and the underlying shares not only for cash but also for each other at some prespecified moments of time. The framework is general enough to describe both European and American derivatives with various payoffs. It was demonstrated that in a quasi-classical limit in absence of money flows, the treatment reproduces the Black-Scholes equation with appropriate boundary conditions. Formal expression for the corrections to the equation in the presence of money flows is also obtained. All listed above results are obtained analytically. However, any further developments give rise to some quite complicated calculations. We believe that only computer calculation can demonstrate the agreement (or disagreement) of the theory with financial data and further development of the proposed approach to the financial modelling should be based on computer simulation. Indeed, as we mentioned before, even in the much easier case of cash-and-shares system only results of numerical calculation can be successively compared with real statistics. The situation becomes more extreme with the derivative pricing. On the other hand, the numerical algorithms to be used in this context essentially the same as ones, which have been used in lattice gauge theory with matter fields. Acknowledgments {#acknowledgments .unnumbered} =============== We want to thank Alexandra Ilinskaia for many helpful discussions. We are grateful to Paul Cooper for careful reading of the manuscript and to Jeff Miller who let us know about Ref [@Hughston]. This work was supported by IPhys Group and grant UK EPSRC GR/L29156. Appendix 1 {#appendix-1 .unnumbered} ========== In this appendix we give a simple introduction to the financial derivatives and an almost rigorous derivation of the Black-Scholes equation using standard no-arbitrage arguments. A financial derivative is a financial instrument whose value is defined by values of other (underlying) financial variables. These underlying variables may be prices of stocks, bonds or other derivatives, exchange rates, market indices and so on. Some particular derivatives (as warrants) have been known for centuries, but after the introduction in 1973 of exchange-traded derivatives on stocks in US trading in derivatives became a really huge industry (for practical aspects see, for instance, [@Dubofsky; @Hull]). Most popular derivatives are call and put options, futures while swaps, forwards and other derivatives are also important for practitioners. Let us give a simple definition of these derivatives. The call option gives a right but not an obligation to buy a certain number of shares (or any prespecified underlying asset) at fixed (strike) price $E$. If the right can be used at the [*final*]{} moment of (expiration) time $T$ only, the call option is called European. In contrast, the American call option provides the right to buy the underlying asset at the strike price at [*any*]{} time before the expiration time $T$. The put options guarantee the same rights not for buying but for selling of the share. Since the American option gives an investor more freedom it is more valuable. In general, the price of the options are nonzero and depend on the price of the underlying asset. The only time when the option price can be zero is the expiration time subject to the case when the underlying asset price is less than $E$ (for call options) or more than $E$ (for put options). A futures contract is an agreement between two parties to buy or sell an asset at a certain time in the future for a certain price. This agreement is a security and can be traded. The price of the agreement clearly depends on the underlying asset price and can be positive as well as negative. The negative price means that to close the position an additional amount should be paid. There is a number of methods to estimate derivative prices and hedge them (which is even more important from practical point of view). The most standard is Black-Scholes analysis which is based on no-arbitrage condition and quasi-Brownian character of the underlying price. Below we assume that the underlying asset is a share and reproduce simple version using the geometrical Brownian process for the underlying share price: $$dS/S = \mu dt + \sigma dW$$ where $dW$ is the Wiener process, $\mu$ is an average rate of return on the share and $\sigma$ is a standard deviation of the return (so called volatility). Though our derivation is not strictly speaking rigorous, a formal derivation can be found in [@Duffie]. The derivative price $C$ is determined by the share price $S$. The share price at some moment in the future depends only on the current share price, not on the share price in the past (i.e. we assume a form of market efficiency). We also assume that the derivative price $C$ is not influenced by other factors. This means that $C(t,S)$ is nonstochastic function of the stochastic parameter $S$ which means that for given $t$ and $S$, $C(t,S)$ has a definite value. Using Ito’s lemma (\[Ito\]) for the geometrical Brownian process, we get following equation describing the derivative price movement $$d C(t,S) = \left(\frac{\partial C}{\partial t} + \frac{S^2\sigma^2}{2}\frac{\partial^2 C}{\partial S^2} \right) dt + \frac{\partial C}{\partial S} dS + o(dt) \ . \label{option-move}$$ RHS of the equation contains $dS$ so option price movement is stochastic. However, we can construct a portfolio from the derivative and the underlying shares which will be risk-free. Indeed, let us consider portfolio of the derivative and ${\partial C}/{\partial S}$ shares in the short position. Price movement of the portfolio $\Pi$ is $$d\Pi = \left(\frac{\partial C}{\partial t} + \frac{S^2\sigma^2}{2}\frac{\partial^2 C}{\partial S^2} \right) dt + o(dt) \ . \label{portfolio-move}$$ Here we omit the term $Sd[{\partial C}/{\partial S}]$ which describes a change of the portfolio structure but not the portfolio value. Since in the last equation there is no longer $dS$ term, the portfolio is risk-free and cannot grow faster than the risk-free interest rate on a bank deposit. This latter statement is known as the no-arbitrage condition (the violation of the condition we consider in the main text). This gives us the following equation: $$\frac{\partial C}{\partial t} + \frac{S^2\sigma^2}{2}\frac{\partial^2 C}{\partial^2 S} = \left(C-S\frac{\partial C}{\partial S}\right) r_b$$ which is the famous Black-Scholes equation. It describes any derivatives and the specific of a concrete derivative is encoded in boundary conditions only. For example, the boundary condition for the futures contract is $$C(S,T) = S-E$$ while for the European call it takes the form: $$C(S,T) = (S-E) \theta (S-E) \ .$$ Other types of boundary conditions and various modifications of the Black-Scholes equations can be found in [@Hull]. The last note here concerns transaction costs. As we consider continuous time limit we neglect $o(dt)$ in formula (\[option-move\]). This means that the portfolio of share and cash (hedging portfolio) have to be rearranged continuously. This leads to infinite transaction costs. In this case we have to keep finite time steps and can not neglect $o(dt)$. This results in impossibility of perfect hedging portfolio construction. [99]{} K. Ilinski, [*Physics of Finance*]{}, in: J. Kertesz $\&$ I. Kondor (Eds.): [*Econophysics: an emerging science*]{}, Dordrecht: Kluwer, 1998; available at http://xxx.lanl.gov/abs/hep-th/9710148; F. Black, M. Scholes, [*Journal of Political Economy*]{}, [**81**]{} (1973) 637; D. A. Dubofsky, [*Options and financial futures. Valuation and uses*]{}, McGraw-Hill, 1992; J.C. Hull, [*Options, futures and other derivatives*]{}, Prentice Hall International, Inc, 1997; D. Duffie, [*Dynamic Asset Pricing Theory*]{}, Princeton University Press, 1992; P. Wilmott, S. Howison and J. Dewynne, [*The Mathematics of Financial Derivatives*]{}, Cambridge University Press, 1995; B.B. Mandelbrot, [*Journal of Business*]{}, [**36**]{}, 394 (1963); B.B. Mandelbrot, [*Jouranl of Business*]{}, [**40**]{}, 40 (1967); P.K.Clark, [*Econometrica*]{}, [**41**]{} 135 (1973); E.E. Peters, [*Fractal market analysis*]{}, John Wiley and Sons, 1994; R.N. Mantegna and H.E. Stanley, [*Nature*]{}, 376, 46-49 (1995); J.P. Bouchaud, D. Sornette, [*Journal de Physique I (France)*]{}, [**4**]{} (1994) 863; [**5**]{} (1995) 219; A. Matacz, [Financial Modelling and Option Theory with the Truncated Levy Process]{}, preprint cond-mat/9710197; available at http://xxx.lanl.gov/abs/cond-mat/9710197; R.N. Mantegna and H.E. Stanley, [*Phys.Rev.Lett*]{} [**73**]{}, (1994) 2946; D.G. Hobson, [*A Review of Stochastic Volatility Models*]{}, (1996), available at http://www.maths.bath.ac.uk/ oz/papers.html P. Cizeau, Y.Liu, M.Meyer, C.-K. Peng, H.E. Stanley, [Volatility distribution function in the S]{}&[*P500 Stock Index*]{}, [*Physica*]{} [**A 245**]{}, 1997, No.3-4, pp.441; available also at http://xxx.lanl.gov/abs/cond-mat/9708143; J.P. Bouchaud, G. Iori, D. Sornette, [*Risk*]{}, [**9(3)**]{} (1996) 61; K. Ilinski and A. Stepanenko: [*Electrodynamical model of quasi-efficient financial market*]{}, [J.Complex Systems]{}, November 1998; H. Takayasu, A.-H. Sato, M. Takayasu, [*Phys.Rev.Lett*]{} [**79**]{}, (1997) 966; D.Sornette, [Multiplicative processes and power laws]{}, [*Phys.Rev*]{} [**E 57**]{}, 1998, No.4, pp.4811; available also at http://xxx.lanl.gov/abs/cond-mat/9708231; S. Lumby, [*Investment appraisal and financial decisions*]{}, Chapman $\&$ Hall, 1994; M. Creutz, [*Quarks, gluons and lattices*]{}, Cambridge University Press, 1983; L.P. Hughston, [*Stochastic Differential Geometry, Financial Modelling and Arbitrage-Free Pricing*]{}, working paper, Merrill Lynch International Limited, June 1994; D. Blake, [*Financial Market Analysis*]{}, McGraw-Hill, 1990; E.J. Elton, M.J. Gruber, [*Modern portfolio theory and investment analysis*]{}, John Wiley $\&$ Sons, 1995. [^1]: E-mail: [email protected] [^2]: E-mail: [email protected]

ArXiv

--- abstract: 'We have performed numerical simulations of dendritic growth at very low undercoolings in two spatial dimension using a phase-field model. In this regime of growth, the dendrites present sharp corners in the tip region while the trailing region is parabolic, and the corresponding side-branching structures resemble the shape of the tip. The scaling $v\rho^2\sim constant$, where $\rho$ is the tip radius of curvature from the fitting to the parabolic trailing region, still holds approximately. We find that the values of $v\rho^2$ are consistent with those given by microscopic solvability theory. The sharpness of the tip region of the dendrite can be characterized in terms of the deviation $\lambda$ with respect to an Ivantsov parabola. We observe that this length scales as $\lambda \sim \rho$, consistent with experimental measurements.' address: | Centre for the Physics of Materials and Department of Physics, McGill University,\ Rutherford Building, 3600 University Street,\ Montréal, Québec, H3A 2T8 Canada. author: - 'José-Luis Mozos and Hong Guo' title: Dendritic growth at very low undercoolings --- = 10000 Dendritic crystal growth constitutes one of the most interesting examples of pattern formation phenomena in nonequilibrium dynamical systems [@langer; @kessler; @godreche; @brener; @pelce; @qian]. While most studies of dendritic growth have concentrated on relatively high undercoolings, recent experimental measurements in dendritic growth at low undercooling[@maurer] have revealed a new feature of this nonequilibrium process, that is, the appearance of a different dendritic morphology with a faceted structure and sharp corners at the dendritic tip. This regime, being close to equilibrium, has been addressed theoretically[@benamar] to give an explanation of the self-similarity of the interfacial profile for different undercooling values. There are however still many open questions regarding the dynamical evolution of the pattern, the shapes of the side-branches, the detailed study of the interfacial roughness near the dendritic tip, and, most importantly, whether new elements should be included in the theoretical description, or if such patterns selected by nature are compatible with our current theoretical understanding[@addabedia]. Motivated by the interesting results of Refs. [@maurer] and [@benamar], in this paper we use a phase-field model to investigate the small undercooling regime of dendritic growth with an emphasis on steady-state pattern selection. Phase-field models have proven to be a breakthrough for the numerical simulation of unstable interfaces[@caginalp; @kobayashi; @wang], and so far very impressive results have been obtained[@kobayashi; @wheeler]. The method is especially useful here since the problem involves high values of interface curvature. Our numerical data show that dendrites with sharp-cornered tips [*and*]{} sharp-shaped side-branchings (see below), are the steady-state pattern at low driving force. With appropriate characterization of the size of the sharp-cornered tip (see below), certain scaling laws hold. It is also found that the relation between growth velocity at the dendrite tip and tip radius at [*large*]{} driving force, still holds here. Finally, our numerical data are consistent with experimental measurements[@maurer]. We use the phase-field model proposed for solidification in a pure liquid[@kobayashi; @wang], which has been shown to provide a reliable means of studying dendritic growth phenomena[@wheeler]. The model includes two fields. First, a dimensionless temperature field $u$ is defined as $u(x,y,t)=(T(x,y,t)-T_{S} )/(T_{S}-T_{L})$, where $T_{S}$ is the melting temperature of the solid phase and $T_{L}$ is the temperature of the liquid phase far from the interface. Second, a phase field $\phi$ is the order parameter of our system; thus $\phi =0$ and $\phi =1$ represent solid and liquid phases, respectively. The interface locus is determined by positions at which $\phi=1/2$. The dynamical equations are the following[@wheeler]: $$\begin{aligned} \frac{\epsilon^2}{m} \frac{\partial \phi}{\partial t}&=& \phi (1-\phi ) \left[ \phi - \frac{1}{2} +30 \epsilon \alpha \Delta u \phi (1-\phi ) \right] + \epsilon^2 {\cal L}(\eta (\theta ) ) \phi \label{pf1}\end{aligned}$$ and $$\begin{aligned} \frac{\partial u}{\partial t}&=& -\frac{30}{\Delta}\phi^2 (1-\phi )^2 \frac{\partial \phi}{\partial t} +\nabla^2 u, \label{pf2}\end{aligned}$$ where the generalized Laplacian operator ${\cal L}$ is defined as ${\cal L}(\eta (\theta)) = \partial_x (\eta(\theta) \eta'(\theta) \partial_y \cdot ) + \partial_y (\eta(\theta) \eta'(\theta) \partial_x \cdot ) + \nabla (\eta(\theta)^2 \nabla \cdot) )$. The parameters $\Delta$, $\alpha$ and $m$ are defined as functions of the latent heat per unit volume $L$, the specific heat $c$, the interfacial energy and mobility[@wheeler]. In particular the dimensionless undercooling $\Delta$ is defined as $\Delta=c(T_{S}-T_{L})/L$. The interfacial width is given[@mcfadden] by the length scale $\epsilon$. In the so-called thin interface limit ($\epsilon \rightarrow 0 $)[@caginalp] one indeed obtains the equivalent macroscopic equations[@langer] from the present phase-field model, as shown by McFadden [*et al.*]{}[@mcfadden]. Thus instead of solving macroscopic equations we shall focus on solving the phase-field model directly. In the above equations the function $\eta$ accounts for the symmetry of the crystalline structure of the solid, and therefore the anisotropy of the kinetic and surface tension coefficients at the interface. Here, as usual[@langer], we have chosen a fourfold symmetry, $\eta(\theta )=1+\gamma \cos (4 \theta) $, where $\gamma$ is the surface tension anisotropy parameter. We have been motivated to choose this form of anisotropy for two reasons. First, it is the simplest form and is used in the solvability analysis of dendritic growth[@langer], thus comparison to that theory maybe made at least at large driving forces or small $\gamma$’s. Second, this form is adequate for the description of the physics of missing orientations[@herring; @liu; @cahn] in [*equilibrium*]{} crystal shapes. It is well known that missing orientations lead to sharp corners in [*equilibrium*]{} crystal shapes[@liu; @shore], hence they should play a role when the growth is very slow. With this form of the anisotropy it is easy to obtain the [*equilibrium*]{} missing orientations using the Frank diagram[@frank; @liu]. We have integrated Eqs. (\[pf1\]) and (\[pf2\]) on a square lattice of width $w\in [100,200]$ and height $h \in [1000,2000]$ using an Euler algorithm with mesh size $7.5 \times 10^{-3}$ and time step $10^{-5}$. The values of the parameters were $\alpha=400$, $m=0.05$ and $\epsilon=5 \times 10^{-3}$ [@wheeler]. We have investigated the evolution of the system for values of undercooling $\Delta$ ranging from $0.02$ to $0.35$. To our knowledge no other numerical simulations at such low undercooling have been carried out so far. A large computational effort is required due to the increasingly demanding time and space factors. For instance the computational time can be estimated to be proportional to a large negative power of $\Delta$[@foot3]. We have fixed the surface tension anisotropy parameter $\gamma=0.1$; in [*equilibrium*]{} this would correspond[@liu] to missing orientations of angles $\theta$ smaller than $\pm 21.65^o$, [*i.e.*]{} surfaces with normal angles in this range are thermodynamically unstable and do not appear on the equilibrium crystal shape. Of course we are facing a non-equilibrium situation here, but this construction for [*equilibrium*]{} shapes gives us some idea of what might happen. An additive Gaussian white noise of intensity $10^{-8}$ has been added to Eq. (\[pf1\]) with the purpose only of helping to trigger the growth of side-branches, assuming the role usually played by numerical noise. However noise with such a low intensity cannot affect the tip profile. A detailed description of the stochastic effects on dendritic growth will be presented elsewhere. Our starting configuration was a small circular solid nucleus in the center at the bottom of the integration grid. We found that dynamical evolution of the unstable interface progresses as follows. First, there was a relatively long transient, in which the temperature evolved from a initial step-like profile to a diffusive one, with a diffusion length comparable to but smaller than the size of the system. The velocity of the interface was constantly adjusting until the temperature flux at the interface acquired a value compatible with the undercooling imposed. Then, the truly dendritic evolution began. We observed that at these very low undercoolings the tip was no longer parabolic, but had a polygonal structure with a sharp corner (Fig. \[figure1\]). This in some sense resembles the equilibrium crystal shape, namely a “square” with rounded edges but still sharp corners for a fourfold anisotropy[@liu]. In our simulations we constantly monitored the interfacial profile to find the steady-state tip shape, which usually appeared at the precise moment when the first side-branch started to grow. After this moment the sharp-cornered tip structure changed very little with time. We note that the only important difference between this simulation and the previous ones[@wheeler] is the very small undercooling, and thus, very small driving force. However notice that the steady-state shape including the shape of side-branches is here quite different from that which appears in the high driving force case. Here growth of side-branches is largely inhibited as their sizes stay more or less constant along the sides of the dendrite. We have checked this side-branching structure by doubling the system size and obtained the same result. The sharp-cornered tip structure can be characterized using a length $\lambda$ which corresponds to the distance between the tip and the topmost point of the fit to a parabolic shape (regions I and II in Fig. \[figure2\], respectively). We were motivated by Ref. [@maurer] to use this distance as a useful characterization since it can be measured in the laboratories[@maurer]. To perform such a fit to compute $\lambda$, a problem is that transition between the behaviors of regions I and II is rather smooth, as shown in Fig. \[figure2\]. The parabolic profile extends down from region I but is cut off because of the appearance of side-branches. To exclude any ambiguity and thus to reduce error, we have found that the quantity $\xi'_{int}(x-x_{tip})/(x-x_{tip})$, where $\xi_{int}$ is the interfacial position, provides a valuable criterion to distinguish the sharp-cornered tip from the parabola. In region I, it is essentially hyperbolic; on the other hand, in region II, it reaches a constant value equal to the curvature at the tip of the parabola. Using this the interface belonging to region II could be extracted, and the fit done in the usual way. In Fig. \[figure3\] the curvature as a function of the steady-state velocity is shown. We have obtained the same scaling behavior as that for the purely parabolic dendrites[@brener], $\rho \sim v^{-1/2}$. In our case the radius $\rho$ is computed using the fitted parabola discussed in the last paragraph. Furthermore, we have found that the values obtained for $\rho^2 v$ lie within $7\%$ of the curve corresponding to the prediction of microscopic solvability theory[@brener] for $\gamma=0.1$. In this sense the tip structure, although showing clear sharp corners, acts effectively as a parabolic one which fits the trailing region of the dendrite. This fact suggests that the relevant quantity to be taken into account for the sharp-cornered tip shape is $\rho$, rather than the real curvature at the sharp tip, which is 10 to 20 times larger than $1/\rho$ and could be strongly conditioned by the lattice discretization. The fact that the aforementioned scaling holds, at least for the range of undercooling values considered here, gives a first indication that there is still only one relevant macroscopic length in this growth regime. A further explicit check can be done. In Fig. \[figure4\] the measured length $\lambda$, as defined above, as well as the fitted radius of curvature $\rho$ for different values of undercooling is shown. As $\Delta$ decreases (for data points on the right of the figure), the sharp corner at the dendrite tip is easily distinguishable and a scaling $\lambda \sim \rho $ is obtained. This scaling is in agreement with previously reported experimental results[@maurer]. The whole tip structure is therefore self-similar including the tip region of the dendrite. For large values of undercooling, the fitted $\rho$ is comparable to the radius of curvature at the tip; thus the measure of $\lambda$ is no longer reliable, shown by the saturation of its values on the left part of Fig. \[figure4\]. The appearance of sharp-cornered tips as the steady-state velocity is reduced is not abrupt but gradual. This suggests that the transformation from a dendrite with a parabolic tip profile and a strong side-branching process in the higher velocity regime, to a sharp-cornered tip structure as well as “faceted” looking side-branches whose growth is rapidly inhibited (Fig. \[figure1\]), is a crossover behavior and not a dynamical transition. This result could be interpreted in terms of a kinetic roughening phenomenon[@godreche; @jorgenson] where the pattern with sharp-cornered tips at low driving force is kinetically roughened at large driving force, leading to the usual parabolic tips. To explore this idea, we have looked the change in morphology of the dendrite depending on the surface tension anisotropy $\gamma$, at a fixed undercooling value. The inset in Fig. \[figure4\] shows $\Theta_{tip}$ which is the corner angle of the dendrite tip, or more precisely the extrapolation of the two sides around the tip region to $x=x_{tip}$, as a function of $\gamma$. We observe a gradual increase of $\Theta_{tip}$ asymptotically to $45^o$ which is the angle that, according to the Frank diagram[@frank] and the Wulff construction[@wulff], minimizes the [*equilibrium*]{} surface free-energy when $\gamma=1$. This energetic consideration seems play a role at the high anisotropy, or similarly, lower undercooling regimes. In this regime the nonequilibrium characteristics of the evolution appear to become less important relative to the equilibrium requirement of free energy minimization. In summary, we have studied the morphology appearing in dendritic growth of a supercooled pure liquid in the low driving force regime. We found that dendrites develop sharp-cornered tips and the growth and coarsening of the side-branches are inhibited at very low undercooling. The scaling behavior for growth at large velocities, $\rho^2v\sim $constant, is still valid at low undercoolings if we use the [*fitted*]{} $\rho$ as described above. Furthermore, this value of $\rho^2v$ is quantitatively consistent with the microscopic solvability theory. We found that there is only one relevant macroscopic length scale in this slow growth regime, such as the fitted $\rho$. The length associated with the size of the sharp corner scales with this fitted $\rho$ as $\lambda\sim \rho$, and this is consistent with experimental measurements. The appearance of the sharp corner is gradual as undercooling decreases or anisotropy increases. Finally we point out that our work focuses on the sharp-cornered tip region of the dendrite, and the steady-state shapes, whereas in the experiment of Ref. [@maurer] true facets were apparent, suggested that it was carried out below the roughening transition of the material. It is not clear how to treat equilibrium roughening transition within a phase-field model where interfaces are diffuse (with thickness $\sim\epsilon$), but this is an interesting problem and should be pursued further in order to quantitatively compare with experiments[@maurer] in the very slow growth regime below the roughening transition. Other interesting directions include detailed calculations of the dynamics of dendritic growth in this low undercooling regime, and the investigation of this regime in three spatial dimensions. We hope to report on these studies in the near future. We thank Dr. Joel D. Shore for a critical reading of the manuscript and many useful criticisms and suggestions. H.G. benefited from a discussion with Dr. John Cahn. We gratefully acknowledge support from the Natural Sciences and Engineering Research Council of Canada, le Fonds pour la Formation de Chercheurs et l’Aide à la Recherche de la Province du Québec and a NATO Collaborative Research Grant CRG.931018. JLM is also supported by the Ministerio de Educación y Ciéncia (Spain). J.S. Langer, Rev. Mod. Phys. [**52**]{}, 1 (1980), and in [*Chance and Matter*]{}, J. Souleite, J. Vannimenus and R. Stora eds., North Holland, Amsterdam (1987). D.A. Kessler, J. Koplik and H. Levine, Adv. Phys. [**37**]{}, 255 (1988). , C. Godreche ed., Cambridge University Press, New York (1992). E.A. Brener and V.I. Mel’nikov, Adv. Phys. [**40**]{}, 53 (1991). , P. Pelcé ed., Perspectives in Physics, Academic Press, Boston (1988). X.W. Qian and H.Z. Cummings, Phys. Rev. Lett. [**64**]{}, 3038 (1990). J. Maurer, P. Bouissou, B. Perrin and P. Tabeling, Europhys. Lett. [**8**]{}, 67 (1989). M. Ben Amar and Y. Pomeau, Europhys. Lett. [**6**]{}, 609 (1988). M. Adda-Bedia and V. Hakim, J. Phys. I (France) [**4**]{}, 383 (1994). G. Caginalp, Phys. Rev. A [**39**]{}, 887 (1989). R. Kobayashi, Physica D [**63**]{}, 410 (1993). S.L. Wang, R.F. Sekerka, A.A. Wheeler, B.T. Murray, S.J. Coriell, R.J. Braun and G.B. McFadden, Physica D [**69**]{}, 189 (1993). A.A. Wheeler, B.T. Murray and R.J. Schaefer, Physica D [**66**]{}, 243 (1993). G.B. McFadden, A.A. Wheeler, R.J. Braun, S.J. Coriell and R.F. Sekerka, Phys. Rev. E [**48**]{}, 2016 (1993). C. Herring, Phys. Rev. [**82**]{}, 87 (1951). F. Liu and H. Metiu, Phys. Rev. B [**48**]{}, 5808 (1993). J.W. Cahn and W.C. Carter, preprint, submitted to Met. Trans. A. (1995). F.C. Frank, in [*Metal Surfaces: Structures, Energetics and Kinetics*]{}, (American Society of Metals, Metals Park, 1963). J. Shore and D.J. Bukman, Phys. Rev. Lett. [**72**]{}, 604 (1994). A rough estimate of the computational time in terms of $\Delta$ can be done. To avoid finite size effects the system size must be larger than the diffusion length in the liquid phase. This requirement, together with the scaling of the radius of curvature $\rho \sim \Delta^{-2}$ and velocity $v \sim \Delta^{4}$ [@brener], gives the width of the system $w \sim \rho \sim \Delta^{-2}$ as well as the time of evolution $t \sim \rho/v \sim \Delta^{-6}$. Thus the computational time goes as $\Delta^{-8}$. In addition, the time has to be rescaled by the mesh size, which in turn has to be fine enough to cope with the extremely high curvature. L. Jörgerson and R. Harris, Mat. Sci. Eng. A [**173**]{}, 89 (1993). M. Wortis in [*Chemistry and Physics of Solid Surfaces*]{}, Vol. 7, ed. R. Vanselow and R. Howe, Springer, Berlin (1988).

ArXiv

--- abstract: 'We show that a Boolean degree $d$ function on the slice $\binom{[n]}{k} = \{ (x_1,\ldots,x_n) \in \{0,1\} : \sum_{i=1}^n x_i = k \}$ is a junta, assuming that $k,n-k$ are large enough. This generalizes a classical result of Nisan and Szegedy on the hypercube. Moreover, we show that the maximum number of coordinates that a Boolean degree $d$ function can depend on is the same on the slice and the hypercube.' author: - 'Yuval Filmus[^1], Ferdinand Ihringer[^2]' bibliography: - 'biblio\_plus.bib' title: | Boolean constant degree functions on the slice\ are juntas --- Introduction {#sec:intro} ============ Nisan and Szegedy [@NisanSzegedy] showed that a Boolean degree $d$ function on the hypercube $\{0,1\}^n$ depends on at most $d2^{d-1}$ coordinates, and described a Boolean degree $d$ function which depends on $\Omega(2^d)$ coordinates. Let us denote the optimal bound by $\gamma(d)$. The goal of this paper is to generalize this result to the *slice* $\binom{[n]}{k}$ or *Johnson scheme* $J(n,k)$, which consists of all points in the hypercube having Hamming weight $k$: \[thm:main\] There exists a constant $C$ such that the following holds. If $C^d \leq k \leq n-C^d$ and $f\colon \binom{[n]}{k} \to \{0,1\}$ has degree $d$, then $f$ depends on at most $\gamma(d)$ coordinates. (We explain in Section \[sec:prel\] what degree $d$ means for functions on the hypercube and on the slice.) Filmus et al. [@fkmw] proved a version of Theorem \[thm:main\] (with a non-optimal bound on the number of points) when $k/n$ is bounded away from $0,1$, but their bound deteriorates as $k/n$ gets closer to $0,1$. We use their result (which we reproduce here, to keep the proof self-contained) to bootstrap our own inductive argument. The case $d = 1$ is much easier. The following folklore result is proved formally in [@FilmusIhringer1]: \[thm:degree1\] If $2 \leq k \leq n-2$ and $f\colon \binom{[n]}{k} \to \{0,1\}$ has degree $1$, then $f$ depends on at most one coordinate. The bounds on $k$ in this theorem are optimal, since every function on $\binom{[n]}{1}$ and on $\binom{[n]}{n-1}$ has degree $1$. In contrast, the bounds on $k$ in Theorem \[thm:main\] are probably not optimal, an issue we discuss in Section \[sec:discussion\]. Let us close this introduction by mentioning a recent result of Keller and Klein [@KellerKlein], which studies Boolean functions on $\binom{[n]}{k}$ which are $\epsilon$-close to being degree $d$, where distance is measured using the squared $L_2$ norm. Assuming that $k \leq n/2$, their result states that if $\epsilon < (k/n)^{O(d)}$ then $f$ is $O(\epsilon)$-close to a junta. Preliminaries {#sec:prel} ============= In this paper, we discuss Boolean functions, which are $0,1$-valued functions, on two different domains: the hypercube and the slice. We will use the notation $[n] := \{1,\ldots,n\}$. A *degree $d$ function* (in a context in which degree is defined) is a function of degree *at most* $d$. #### The hypercube. The $n$-dimensional hypercube is the domain $\{0,1\}^n$. Every function on the hypercube can be represented uniquely as a multilinear polynomial in the $n$ input arguments $x_1,\ldots,x_n$. The *degree* of a function on the hypercube is the degree of this polynomial. Alternatively, the degree of a function on the hypercube is the minimum degree of a polynomial in $x_1,\ldots,x_n$ which agrees with the function on all points of the hypercube. A function on the hypercube is an *$m$-junta* if it depends on at most $m$ inputs, that is, if there exists a set $I$ of $m$ inputs such that $f(x) = f(y)$ as long as $x_i = y_i$ for all $i \in I$; we also say that $f$ is an $I$-junta. For more information on functions on the hypercube from this perspective, consult O’Donnell’s monograph [@RyanODonnell]. #### The slice. Let $0 \leq k \leq n$. The *slice* $\binom{[n]}{k}$ is the subset of $\{0,1\}^n$ consisting of all vectors having Hamming weight $k$. The slice appears naturally in combinatorics, coding theory, and elsewhere, and is known to algebraic combinatorialists as the *Johnson scheme* $J(n,k)$. Every function on the slice can be represented uniquely as a multilinear polynomial $P$ in the $n$ input arguments $x_1,\ldots,x_n$ of degree at most $\min(k,n-k)$ which satisfies $\sum_{i=1}^n \frac{\partial P}{\partial x_i} = 0$ (the latter condition is known as *harmonicity*). The *degree* of a function on the slice is the degree of this polynomial. Alternatively, the degree of a function on the slice is the minimum degree of a polynomial in $x_1,\ldots,x_n$ (not necessarily multilinear or harmonic) which agrees with the function on all points of the slice. A function $f$ on the slice is an *$m$-junta* if there exist a function $g\colon\{0,1\}^m \to \{0,1\}$ and $m$ indices $i_1<\ldots<i_m$ such that $f(x) = g(x|_{i_1,\ldots,i_m})$, where $x|_{i_1,\ldots,i_m} = x_{i_1},\ldots,x_{i_m}$. Alternatively, $f$ is an $m$-junta if there exists a set $I$ of $m$ coordinates such that $f$ is invariant under permutation of the coordinates in $[n] \setminus I$; we also say that $f$ is an $I$-junta. Note that the set $I$ is not defined uniquely (in contrast to the hypercube case): for example, $f = \sum_{i \in I} x_i$ is both an $I$-junta and an $[n]\setminus I$-junta. The *$p$th norm* of $f$ is given by $\|f\|_p = {\mathbb{E}}[|f|^p]^{1/p}$, where the expectation is over a uniform point in the slice. In particular, $\|f\|_2^2 = {\mathbb{E}}[f^2]$. Let $f$ be a Boolean function on the slice $\binom{[n]}{k}$, and let $P$ be its unique harmonic multilinear polynomial representation. The *$d$th level* of $f$, denoted $f^{=d}$, is the homogeneous degree $d$ part of $P$ (the sum of all degree $d$ monomials with their coefficients). The different levels are orthogonal: ${\mathbb{E}}[f^{=d} f^{=e}] = 0$ if $d \neq e$. Orthogonality of the different levels implies that $$\|f\|_2^2 = \sum_{d=0}^{\deg f} \|f^{=d}\|_2^2.$$ Let $f$ be a Boolean function on the slice $\binom{[n]}{k}$, and let $i,j \in [n]$. We define $f^{(i\;j)}$ to be the function given by $f^{(i\;j)}(x) = f(x^{(i\;j)})$, where $x^{(i\;j)}$ is obtained from $x$ by switching $x_i$ and $x_j$. The *$(i,j)$th influence* of $f$ is $\operatorname{Inf}_{ij}[f] = \frac{1}{4} \Pr[f(x) \neq f(x^{(i\;j)})]$, where $x$ is chosen uniformly at random over the slice. An equivalent formula is $\operatorname{Inf}_{ij}[f] = \frac{1}{4} {\mathbb{E}}[(f-f^{(i\;j)})^2]$. Clearly $\operatorname{Inf}_{ij}[f] = 0$ if and only if $f = f^{(i\;j)}$. The *total influence* of $f$ is $\operatorname{Inf}[f] = \frac{1}{n} \sum_{1 \leq i < j \leq n} \operatorname{Inf}_{ij}[f]$. It is given by the formula $$\label{eq:influence} \operatorname{Inf}[f] = \sum_{d=0}^{\deg f} \frac{d(n+1-d)}{n} \|f^{=d}\|_2^2.$$ For a parameter $\rho \in (0,1]$, the *noise operator* $T_\rho$, mapping functions on the slice to functions on the slice, is defined by $$\label{eq:noise} T_\rho f = \sum_{d=0}^{\deg f} \rho^{d(1-(d-1)/n)} f^{=d}.$$ Alternatively, $(T_\rho f)(x)$ is the expected value of $f(y)$, where $y$ is chosen by applying $N \sim \mathrm{Po}(\frac{n-1}{2} \log (1/\rho))$ random transpositions to $x$. Lee and Yau [@LeeYau] proved a log Sobolev inequality, which together with classical results of Diaconis and Saloff-Coste [@DSC] implies that the following *hypercontractive inequality* holds for some constant $C_H>0$: $$\label{eq:hypercontractivity} \|T_\rho f\|_2 \leq \|f\|_{4/3}, \qquad \rho = \left(\frac{2k(n-k)}{n(n-1)}\right)^{C_H}.$$ For more information on functions on the slice, consult [@fm]. Main theorem {#sec:main} ============ For the rest of this section, we fix an integer $d \geq 1$. Our goal is to prove Theorem \[thm:main\] for this value of $d$. We will use the phrase *universal constant* to refer to a constant independent of $d$. The strategy of the proof is to proceed in three steps: 1. Bootstrapping: Every Boolean degree $d$ function on $\binom{[2n]}{n}$ is a $K^d$-junta. 2. Induction: If every Boolean degree $d$ function on $\binom{[n]}{k}$ is an $M$-junta, then the same holds for $\binom{[n+1]}{k}$ and $\binom{[n+1]}{k+1}$ (under certain conditions). 3. Culmination: If a Boolean degree $d$ function on $\binom{[n]}{k}$ is an $L$-junta but not an $(L-1)$-junta, then (under certain conditions) there exists a Boolean degree $d$ function on the hypercube depending on $L$ coordinates. We also show a converse to the last step: given a Boolean degree $d$ function on the hypercube depending on $L$ coordinates, we show how to construct Boolean degree $d$ functions on large enough slices that are $L$-juntas but not $(L-1)$-juntas. Bootstrapping {#sec:bootstrap} ------------- We bootstrap our approach by proving that every Boolean degree $d$ function on $\binom{[2n]}{n}$ is a junta. The proof is a simple application of hypercontractivity, and already appears in [@fkmw]. We reproduce a simplified version here in order to make the paper self-contained. The main idea behind the proof is to obtain a dichotomy on the influences of the function. \[lem:dichotomy\] There exists a universal constant $\alpha$ such that all non-zero influences of a Boolean degree $d$ function on $\binom{[2n]}{n}$ are at least $\alpha^d$. Let $f$ be a Boolean degree $d$ function on $\binom{[2n]}{n}$. Given $i,j \in [n]$, consider the function $f_{ij} = (f - f^{(i\;j)})/2$, related to the $(i,j)$th influence of $f$ by $\operatorname{Inf}_{ij}[f] = \|f_{ij}\|_2^2$. Since $f$ is $0,1$-valued, $f_{ij}$ is $0,\pm 1$-valued. Since $f$ has degree $d$, it follows that $f_{ij}$ can be written as a degree $d$ polynomial, and so has degree at most $d$. Hypercontractivity  implies that for some universal constant $\rho$, we have $$\label{eq:bootstrapping} \|T_\rho f_{ij}\|_2 \leq \|f_{ij}\|_{4/3}.$$ We can estimate the left-hand side of  using : $$\|T_\rho f_{ij}\|_2^2 = \sum_{e=0}^d \rho^{2e(1-(e-1)/n)} \|f_{ij}^{=e}\|_2^2 \geq \rho^{2d} \sum_{e=0}^d \|f_{ij}^{=e}\|_2^2 = \rho^{2d} \|f_{ij}\|_2^2 = \rho^{2d} \operatorname{Inf}_{ij}[f].$$ We can calculate the right-hand side of  using the fact that $f_{ij}$ is $0,\pm 1$-valued: $$\|f_{ij}\|_{4/3}^{4/3} = {\mathbb{E}}[|f_{ij}|^{4/3}] = {\mathbb{E}}[|f_{ij}|^2] = \|f_{ij}\|_2^2 = \operatorname{Inf}_{ij}[f].$$ Combining the estimates on both sides of , we conclude that $$\rho^{2d} \operatorname{Inf}_{ij}[f] \leq \operatorname{Inf}_{ij}[f]^{3/2}.$$ Hence either $\operatorname{Inf}_{ij}[f] = 0$ or $\operatorname{Inf}_{ij}[f] \geq \rho^{4d}$. The next step is to prove a degenerate triangle inequality (for the non-degenerate version, see Wimmer [@Wimmer Lemma 5.4] and Filmus [@filmus4 Lemma 5.1]). \[lem:triangle-inequality\] Let $f$ be a function on a slice. If $\operatorname{Inf}_{ik}[f] = \operatorname{Inf}_{jk}[f] = 0$ then $\operatorname{Inf}_{ij}[f] = 0$. If $\operatorname{Inf}_{ik}[f] = \operatorname{Inf}_{jk}[f] = 0$ then $f(x) = f(x^{(i\;k)}) = f(x^{(j\;k)})$, and so $f(x) = f(x^{(i\;k)(j\;k)(i\;k)}) = f(x^{(i\;j)})$. It follows that $\operatorname{Inf}_{ij}[f] = 0$. To complete the proof, we use formula , which implies that $\operatorname{Inf}[f] \leq d$ for any Boolean degree $d$ function $f$, together with an idea of Wimmer [@Wimmer Proposition 5.3]. \[lem:bootstrapping\] There exists a universal constant $K > 1$ such that for $n \geq 2$, every Boolean degree $d$ function on $\binom{[2n]}{n}$ is a $K^d$-junta. Let $f$ be a Boolean degree $d$ function on $\binom{[2n]}{n}$. Construct a graph $G$ on the vertex set $[2n]$ by connecting two vertices $i,j$ if $\operatorname{Inf}_{ij}[f] \geq \alpha^d$, where $\alpha$ is the constant from Lemma \[lem:dichotomy\]. Let $M$ be a maximal matching in $G$. It is well-known that the $2|M|$ vertices of $M$ form a vertex cover $V$, that is, any edge of $G$ touches one of these vertices. Therefore if $i,j \notin V$ then $\operatorname{Inf}_{ij}[f] < \alpha^d$, and so $\operatorname{Inf}_{ij}[f] = 0$ according to Lemma \[lem:dichotomy\]. In other words, $f$ is a $V$-junta. It remains to bound the size of $V$. Let $(i,j)$ be any edge of $M$, and let $k$ be any other vertex. Lemma \[lem:triangle-inequality\] shows that either $\operatorname{Inf}_{ik}[f] \neq 0$ or $\operatorname{Inf}_{jk}[f] \neq 0$, and so either $(i,k)$ or $(j,k)$ is an edge of $G$, according to Lemma \[lem:dichotomy\]. It follows that $G$ contains at least $|M|(2n-2)/2$ edges (we divided by two since some edges could be counted twice). Therefore $n\operatorname{Inf}[f] = \sum_{1 \leq i < j \leq n} \operatorname{Inf}_{ij}[f] \geq \alpha^d |M| (n-1)$. On the other hand,  shows that $$\operatorname{Inf}[f] = \sum_{e=0}^d \frac{e(n+1-e)}{n} \|f^{=e}\|_2^2 \leq d \sum_{e=0}^d \|f^{=e}\|_2^2 = d \|f\|_2^2 \leq d.$$ It follows that $$|V| = 2|M| \leq 2 \frac{n}{n-1} (1/\alpha)^d \leq (4/\alpha)^d. \qedhere$$ Induction {#sec:induction} --------- The heart of the proof is an inductive argument which shows that if Theorem \[thm:main\] holds (with a non-optimal bound on the size of the junta) for the slice $\binom{[n]}{k}$, then it also holds for the slices $\binom{[n+1]}{k}$ and $\binom{[n+1]}{k+1}$, assuming that $n$ is large enough and that $k$ is not too close to $0$ or $n$. Given a Boolean degree $d$ function $f$ on $\binom{[n+1]}{k}$ or $\binom{[n+1]}{k+1}$, the idea is to consider restrictions of $f$ obtained by fixing one of the coordinates. \[lem:restrictions\] Suppose that every Boolean degree $d$ function on $\binom{[n]}{k}$ is an $M$-junta, where $M \geq 1$. Let $f$ be a Boolean degree $d$ function on $\binom{[n+1]}{k+b}$, where $b \in \{0,1\}$. For each $i \in [n+1]$, let $f_i$ be the restriction of $f$ to vectors satisfying $x_i = b$. Then for each $i \in [n+1]$ there exists a set $S_i \subseteq [n+1] \setminus \{i\}$ of size at most $M$ and a function $g_i\colon \{0,1\}^{S_i} \to \{0,1\}$, depending on all inputs, such that $f_i(x) = g_i(x|_{S_i})$. Choose $i \in [n+1]$. The domain of $f_i$ is isomorphic to $\binom{[n]}{k}$. Moreover, since $f$ can be represented as a polynomial of degree $d$, so can $f_i$, hence $\deg f_i \leq d$. By assumption, $f_i$ is an $M$-junta, and so $f_i(x) = h_i(x|_{T_i})$ for some set $T_i$ of $M$ indices and some function $h_i\colon \{0,1\}^{T_i} \to \{0,1\}$. Let $S_i \subseteq T_i$ be the set of inputs that $h_i$ depends on. Then there exists a function $g_i\colon \{0,1\}^{S_i} \to \{0,1\}$ such that $g_i(x|_{S_i}) = h_i(x|_{T_i})$, completing the proof. Each of the sets $S_i$ individually contains at most $M$ indices. We now show that in fact they contain at most $M$ indices *in total*. \[lem:union\] Under the assumptions of Lemma \[lem:restrictions\], suppose further that $n \geq (M+1)^3$ and $M+2 \leq k \leq n+1-(M+2)$. The union of any $M+1$ of the sets $S_i$ contains at most $M$ indices. We can assume, without loss of generality, that the sets in question are $S_1,\ldots,S_{M+1}$. Denote their union by $A$, and let $B = A \cup \{1,\ldots,M+1\}$. Since $|B| \leq (M+1)^2 \leq n$, there exists a point $r \in [n+1] \setminus B$. We proceed by bounding the number of unordered pairs of distinct indices $i,j \in [n+1] \setminus \{r\}$ such that $f_r = f_r^{(i\;j)}$, which we denote by $N$. Since $f_r$ is an $M$-junta, we know that $N \geq \binom{n-M}{2}$. We will now obtain an upper bound on $N$ in terms of $|A|$ and $|B|$. Let $1 \leq \ell \leq M+1$, and suppose that $i \in S_\ell$ and $j \notin S_\ell \cup \{\ell,r\}$. We claim that $f_r^{(i\;j)} \neq f_r$. Indeed, since $g_\ell$ depends on all inputs, there are two inputs $y,z$ to $g_\ell$, differing only on the $i$th coordinate, say $y_i = b$ and $z_i = 1-b$, such that $g_\ell(y) \neq g_\ell(z)$. Since $M+2 \leq k \leq n+1-(M+2)$, we can extend $y$ to an input $x$ to $f$ satisfying additionally the constraints $x_\ell = x_r = b$ and $x_j = 1-b$. Since $x_\ell = x_r = b$, the input $x$ is in the common domain of $f_\ell$ and $f_r$. Notice that $f_r(x) = f_\ell(x) = g_\ell(y)$, whereas $f_r(x^{(i\;j)}) = f_\ell(x^{(i\;j)}) = g_\ell(z)$, since $x_i = y_i = b$ whereas $x_j = 1-b$. By construction $g_\ell(y) \neq g_\ell(z)$, and so $f_r \neq f_r^{(i\;j)}$. The preceding argument shows that if $i \in A$ and $j \notin B \cup \{r\}$ then $f_r \neq f_r^{(i\;j)}$. Therefore $\binom{n}{2} - N \geq |A|(n-|B|) \geq |A|(n - (M+1)^2)$. Combining this with the lower bound $N \geq \binom{n-M}{2}$, we deduce that $$|A| (n - (M+1)^2) \leq \binom{n}{2} - \binom{n-M}{2} = \frac{M(2n-M-1)}{2}.$$ Rearrangement shows that $$|A| \leq \frac{M(2n-M-1)}{2(n-(M+1)^2)} = \left(1 + \frac{(M+1)(2M+1)}{2n-2(M+1)^2}\right)M.$$ When $n > (M^2 + (3/2)M + 1)(M+1)$, we have $\frac{(M+1)(2M+1)}{2n-2(M+1)^2} < \frac{1}{M}$, and so $|A| < M+1$. We conclude that when $n \geq (M+1)^3$, we have $|A| \leq M$. \[cor:union\] Under the assumptions of the preceding lemma, the union of $S_1,\ldots,S_{n+1}$ contains at most $M$ indices. Suppose that the union contained at least $M+1$ indices $i_1,\ldots,i_{M+1}$. Each index $i_t$ is contained in some set $S_{j_t}$, and in particular the union of $S_{j_1},\ldots,S_{j_{M+1}}$ contains at least $M+1$ indices, contradicting the lemma. Denoting the union of all $S_i$ by $S$, it remains to show that $f$ is an $S$-junta. \[lem:induction\] Suppose that every Boolean degree $d$ function on $\binom{[n]}{k}$ is an $M$-junta, where $M \geq 1$; that $n \geq (M+1)^3$; and that $M+2 \leq k \leq n+1-(M+2)$. Any Boolean degree $d$ function on $\binom{[n+1]}{k}$ or on $\binom{[n+1]}{k+1}$ is an $M$-junta. Let $b,f_i,g_i,S_i$ be defined as in Lemma \[lem:restrictions\], and let $S$ denote the union of $S_1,\ldots,S_{n+1}$. Corollary \[cor:union\] shows that $|S| \leq M$. Since $n \geq (M+1)^3$, it follows that there exists an index $r \in [n+1] \setminus S$. We will show that $f(x) = g_r(x|_{S_r})$, and so $f$ is an $M$-junta. Consider any input $x$ to $f$. If $x_r = b$ then $x$ is in the domain of $f_r$, and so clearly $f(x) = f_r(x) = g_r(x|_{S_r})$. Suppose therefore that $x_r = 1-b$. Since $M+2 \leq k \leq n+1-(M+2)$, there exists a coordinate $s \in [n+1] \setminus S$ such that $x_s = b$, putting $x$ in the domain of $f_s$. Again since $M+2 \leq k \leq n+1-(M+2)$, there exists a coordinate $t \in [n+1] \setminus (S \cup \{s\})$ such that $x_t = b$. Since $x^{(r\;t)}$ is in the domain of $f_r$, we have $$f(x) = f_s(x) = g_s(x|_{S_s}) = g_s(x^{(r\;t)}_{S_s}) = f_s(x^{(r\;t)}) = f_r(x^{(r\;t)}) = g_r(x^{(r\;t)}|_{S_r}) = g_r(x|_{S_r}). \qedhere$$ Culmination {#sec:culmination} ----------- Combining Lemma \[lem:bootstrapping\] and Lemma \[lem:induction\], we obtain a version of Theorem \[thm:main\] with a suboptimal upper bound on the size of the junta. \[lem:main\] There exists a universal constant $C > 1$ such that whenever $C^d \leq k \leq n - C^d$, every Boolean degree $d$ function on $\binom{[n]}{k}$ is a $C^d$-junta. Let $K$ be the constant from Lemma \[lem:bootstrapping\], and let $M = K^d$. We choose $C := (K+2)^3$. Let us assume that $k \leq n/2$ (the proof for $k \geq n/2$ is very similar). Lemma \[lem:bootstrapping\] shows that every Boolean degree $d$ function on $\binom{[2k]}{k}$ is an $M$-junta. If $m \geq 2k$ then $m \geq 2k \geq (M+1)^3$ and $M+2 \leq k \leq m-(M+2)$. Therefore Lemma \[lem:induction\] shows that if every Boolean degree $d$ function on $\binom{[m]}{k}$ is an $M$-junta, then the same holds for $\binom{[m+1]}{k}$. Applying the lemma $n-2k$ times, we conclude that every Boolean degree $d$ function on $\binom{[n]}{k}$ is a $C^d$-junta. To complete the proof of the theorem, we show how to convert a Boolean degree $d$ function on the slice depending on many coordinates to a Boolean degree $d$ function on the hypercube depending on the same number of coordinates. \[lem:slice-to-cube\] Suppose that $f$ is a Boolean degree $d$ function on $\binom{[n]}{k}$ which is an $L$-junta but not an $(L-1)$-junta, where $k \leq L \leq n-k$. Then there exists a Boolean degree $d$ function $g$ on $\{0,1\}^L$ which depends on all coordinates. Without loss of generality, we can assume that $f(x) = g(x_1,\ldots,x_L)$ for some Boolean function $g$ on the $L$-dimensional hypercube. Since $f$ is not an $(L-1)$-junta, the function $g$ depends on all coordinates. Since $k \leq L \leq n-k$, as $x$ goes over all points in $\binom{[n]}{k}$, the vector $x_1,\ldots,x_L$ goes over all points in $\{0,1\}^L$. It remains to show that there is a degree $d$ polynomial agreeing with $g$ on $\{0,1\}^L$. Since $f$ has degree at most $d$, there is a degree $d$ multilinear polynomial $P$ such that $f = P$ for every point in $\binom{[n]}{k}$. If $\pi$ is any permutation of $\{L+1,\ldots,n\}$ then $f = f^\pi$, where $f^\pi(x) = f(x^\pi)$. Denoting the set of all such permutations by $\Pi$, if we define $Q := {\mathbb{E}}_{\pi \in \Pi}[P^\pi]$ then $f = Q$ for every point in $\binom{[n]}{k}$. The polynomial $Q$ is a degree $d$ polynomial which is invariant under permutations from $\Pi$. For $a_1,\ldots,a_L \in \{0,1\}^L$ summing to $a \leq d$, let $Q_{a_1,\ldots,a_L}(x_{L+1},\ldots,x_n) = Q(a_1,\ldots,a_L,x_{L+1},\ldots,x_n)$. This is a degree $d-a$ symmetric polynomial, and so a classical result of Minsky and Papert [@Minsky:1988:PEE:50066] (see also [@NisanSzegedy Lemma 3.2]) implies that there exists a degree $d-a$ univariate polynomial $R_{a_1,\ldots,a_L}$ such that $Q_{a_1,\ldots,a_L}(x_{L+1},\ldots,x_n) = R_{a_1,\ldots,a_L}(x_{L+1} + \cdots + x_n)$ for all $x_{L+1},\ldots,x_n \in \{0,1\}^{n-L}$. Since $x_{L+1} + \cdots + x_n = k - x_1 - \cdots - x_L$, it follows that for inputs $x$ in $\binom{[n]}{k}$, we have $$f(x_1,\ldots,x_n) = \sum_{\substack{a_1,\ldots,a_L \in \{0,1\}^L \\ a_1+\cdots+a_L \leq d}} \prod_{i\colon a_i=1} x_i \cdot R_{a_1,\ldots,a_L}(k - x_1 - \cdots - x_L).$$ The right-hand side is a degree $d$ polynomial in $x_1,\ldots,x_L$ which agrees with $g$ on $\{0,1\}^L$. Theorem \[thm:main\] immediately follows from combining Lemma \[lem:main\] and Lemma \[lem:slice-to-cube\]. We conclude this section by proving a converse of Lemma \[lem:cube-to-slice\]. \[lem:cube-to-slice\] Suppose that $g$ is a Boolean degree $d$ function on $\{0,1\}^L$ depending on all coordinates. Then for all $n,k$ satisfying $L \leq k \leq n-L$ there exists a Boolean degree $d$ function $f$ on $\binom{[n]}{k}$ which is an $L$-junta but not an $(L-1)$-junta. We define $f(x_1,\ldots,x_n) = g(x_1,\ldots,x_L)$. Clearly, $f$ is an $L$-junta. Since $g$ has degree at most $d$, there is a polynomial $P$ which agrees with $g$ on all points of $\{0,1\}^L$. The same polynomial also agrees with $f$ on all points of $\binom{[n]}{k}$, and so $f$ also has degree at most $d$. It remains to show that $f$ is not an $(L-1)$-junta. Suppose, for the sake of contradiction, that $f$ were an $(L-1)$-junta. Then there exists a set $S$ of size at most $L-1$ and a Boolean function $h\colon \{0,1\}^S \to \{0,1\}$ such that $f(x) = h(x|_S)$. Since $|S| < L$, there exists some coordinate $i \in \{1,\ldots,L\} \setminus S$. Since $g$ depends on all coordinates, there are two inputs $y,z$ to $g$ differing only in the $i$th coordinate, say $y_i = 0$ and $z_i = 1$, such that $g(y) \neq g(z)$. Since $n \geq 2L$, there exists a coordinate $j \in [n] \setminus (\{1,\ldots,L\} \cup S)$. Since $L \leq k \leq n-L$, we can extend $y$ to an input $\tilde{y}$ to $f$ such that $x_j = 1$. The input $\tilde{z} = \tilde{y}^{(i\;j)}$ extends $z$. Since $i,j \notin S$, the inputs $\tilde{y},\tilde{z}$ agree on all coordinates in $S$, and so $f(\tilde{y}) = h(\tilde{y}|_S) = h(\tilde{z}|_S) = f(\tilde{z})$. On the other hand, $f(\tilde{y}) = g(y) \neq g(z) = f(\tilde{z})$. This contradiction shows that $f$ cannot be an $(L-1)$-junta. Discussion {#sec:discussion} ========== #### Optimality. Lemma \[lem:cube-to-slice\] shows that the size of the junta in Theorem \[thm:main\] is optimal. However, it is not clear whether the bounds on $k$ are optimal. The theorem fails when $k \leq d$ or $k \geq n-d$, since in these cases every function has degree $d$. This prompts us to define the following two related quantities: 1. $\zeta(d)$ is the minimal value such that every Boolean degree $d$ function on $\binom{[n]}{k}$ is an $O(1)$-junta whenever $\zeta(d) \leq k \leq n-\zeta(d)$. 2. $\xi(d)$ is the minimal value such that every Boolean degree $d$ function on $\binom{[n]}{k}$ is a $\gamma(d)$-junta whenever $\xi(d) \leq k \leq n-\xi(d)$. Clearly $d < \zeta(d) \leq \xi(d)$. We can improve this to $\zeta(d) \geq \eta(d) \geq 2 \lceil \frac{d+1}{2} \rceil$, where $\eta(d)$ is the maximum integer such that there exists a non-constant univariate degree $d$ polynomial $P_d$ satisfying $P_d(0),\ldots,P_d(\eta(d)-1) \in \{0,1\}$. Given such a polynomial $P_d$, we can construct a degree $d$ function $f_d$ on $\binom{[n]}{k}$ which is not a junta: $$f_d(x_1,\ldots,x_n) = P_d(x_1+\cdots+x_{\lfloor n/2 \rfloor}).$$ When $k < \eta(d)$, the possible values of $x_1+\cdots+x_{\lfloor n/2 \rfloor}$ are such that $f_d$ is Boolean. One can check that $f_d$ is not an $L$-junta unless $L \geq \lfloor n/2 \rfloor$. The following polynomial shows that $\eta(d) \geq 2\lceil \frac{d+1}{2} \rceil$: $$P_d(\sigma) = \sum_{e=0}^d (-1)^e \binom{\sigma}{e},$$ where $\binom{\sigma}{0} = 1$. While this construction can be improved for specific $d$ (for example, $\eta(7) = 9$ and $\eta(12) = 16$), the upper bound $\eta(d) \leq 2d$ shows that this kind of construction cannot given an exponential lower bound on $\zeta(d)$. Curiously, essentially the same function appears in [@dfh2 Section 7] as an example of a degree $d$ function on the biased hypercube which is almost Boolean but somewhat far from being constant. #### Extensions. It would be interesting to extend Theorem \[thm:main\] to other domains. In recent work [@FilmusIhringer1], we explored Boolean degree $1$ functions on various domains, including various association schemes and finite groups, and the multislice (consult the work for the appropriate definitions). Inspired by these results, we make the following conjectures: 1. If $f$ is a Boolean degree $d$ function on the symmetric group then there are sets $I,J$ of $O(1)$ indices such that $f(\pi)$ depends only on $\pi(i)|_{i \in I}$ and $\pi^{-1}(j)|_{j \in J}$. 2. If $f$ is a Boolean degree $d$ function on the Grassmann scheme then there are $O(1)$ points and hyperplanes such that $f(S)$ depends only on which of the points is contained in $S$, and which of the hyperplanes contain $S$. 3. If $f$ is a Boolean degree $d$ function on the multislice $M(k_1,\ldots,k_m)$ for $k_1,\ldots,k_m \geq \exp(d)$ then $f$ is a $\gamma_m(d)$-junta, where $\gamma_m(d)$ is the maximum number of coordinates that a Boolean degree $d$ function on the Hamming scheme $H(n,m)$ can depend on. We leave it to the reader to show that $\gamma_m(d)$ exists. In fact, simple arguments show that $m^{d-1} \leq \gamma_m(d) \leq \gamma(\lceil \log_2 m \rceil d)$. [^1]: =1500 Technion — Israel Institute of Technology, Haifa, Israel. This research was funded by ISF grant 1337/16. [^2]: =1500 Einstein Institute of Mathematics, Hebrew University of Jerusalem, Israel. Department of Pure Mathematics and Computer Algebra, Ghent University, Belgium. Supported by ERC advanced grant 320924. The author is supported by a postdoctoral fellowship of the Research Foundation — Flanders (FWO).

ArXiv

--- abstract: 'IGRJ18483-0311 was discovered with  in 2003 and later classified as a supergiant fast X-ray transient. It was observed in outburst many times, but its quiescent state is still poorly known. Here we present the results of , , and  observations of 18483.  These data improved the X-ray position of the source, and provided new information on the timing and spectral properties of 18483 in quiescence. We report the detection of pulsations in the quiescent X-ray emission of this source, and give for the first time a measurement of the spin-period derivative of this source. In 18483 the measured spin-period derivative of -(1.3$\pm$0.3)$\times$10$^{-9}$ s s$^{-1}$ likely results from light travel time effects in the binary. We compare the most recent observational results of 18483 and 18, the two supergiant fast X-ray transients for which a similar orbital period has been measured.' author: - | A. Giunta$^{1,2}$ [^1], E. Bozzo$^{3,1}$, F. Bernardini$^{1,2}$, G. Israel$^{1}$, L. Stella$^{1}$, M. Falanga$^{4}$, S. Campana$^{5}$, A. Bazzano$^{6}$, A.J. Dean$^{7}$, M. Mendez$^{8}$\ $^{1}$INAF - Osservatorio Astronomico di Roma, Via Frascati 33, 00044 Rome, Italy.\ $^{2}$Dipartimento di Fisica - Università di Roma Tor Vergata, via della Ricerca Scientifica 1, 00133 Rome, Italy.\ $^{3}$ISDC, Geneva Observatory, University of Geneva, Chemin d’Ecogia 16, CH-1290 Versoix, Switzerland.\ $^{4}$ISSI, Hallerstrasse 6, CH-3012 Bern, Switzerland.\ $^{5}$INAF – Osservatorio Astronomico di Brera, via Emilio Bianchi 46, I-23807 Merate (LC), Italy.\ $^{6}$INAF - IASF, Via del Fosso del Cavaliere 100, 00133 Roma, Italy.\ $^{7}$School of Physics and Astronomy, University of Southampton, Southampton SO17 1BJ, UK.\ $^{8}$Kapteyn Astronomical Institute, University of Groningen, PO Box 800, 9700 AV Groningen, the Netherlands. date: 'Received 7 April 2009; accepted 28 May 2009.' title: 'The supergiant fast X-ray transient IGRJ18483-0311 in quiescence: , , and  observations' --- X-rays: binaries - stars: individual ( 18483, 18 ) -stars: neutron - X-rays: stars Introduction {#sec:intro} ============ 18483 was discovered in 2003 during  deep observations of the Galactic Center [@chernyakova03]. The mean source X-ray flux was $\sim$10 mCrab in the 15-40 keV, and $\sim$5 mCrab in the 40-100 keV band [@chernyakova03; @molkov04]. The 18.5 days orbital period of the system was discovered by @levine06 using [*RXTE*]{} archival data, and was later confirmed with  [@sguera07].  data also showed that 18483 usually undergoes relatively long outbursts ($\sim$3 days) that comprise several fast flares with typical timescales of a few hours. During these bright events, the broad band (3-50 keV) spectrum is best fit by an absorbed cutoff power law model (photon index $\Gamma$=1.4, cut-off energy $E_{\rm c}$=22 keV, and absorption column density $N_{\rm H}$=9$\times$10$^{22}$ cm$^{-2}$). @sguera07 further detected a periodicity at 21.0526$\pm$0.0005 s in the  data, and interpreted it as the spin-period of the neutron star (NS) hosted in 18483. The measured pulse fraction in the 4-20 keV energy band was 48$\pm$7%[^2].  /XRT observations in 2006 detected the source at a flux level of 4.2$\times$10$^{-12}$ erg/cm$^2$/s and provided a source position of $\alpha_{\rm J2000}$=18$^{\rm h}$48$^{\rm m}$17$\fs$17 and $\delta_{\rm J2000}$=-3${\degr}$10$\arcmin$15$\farcs$54 [estimated accuracy 3$\farcs$3, @sguera07]. This allowed @rahoui08 to identify the optical counterpart of the X-ray source, a B0.5Ia star at a distance of 3-4 kpc, and to estimate its mass and radius ($M_{*}$=33 $M_{\odot}$ and $R_{*}$=33.8 $R_{\odot}$, respectively). These authors also suggested that an eccentricity 0.43$\lesssim$$e$$\lesssim$0.68 could explain the 3-day duration of the outbursts [as reported by @sguera07]. Based on these results, it was concluded that 18483 most likely belongs to the class of supergiant fast X-ray transients [SFXTs, @sguera06; @sguera07; @walter07]. However, due to the longer duration of its outbursts (a few days as opposed to a few hours) and a factor of $\sim$10 lower luminosity swings between outburst and quiescence, @rahoui08 classified 18483 as an “intermediate” SFXT, rather than a standard SFXT [see @walter07 for the definition of standard and intermediate SFXTs]. In this paper we analyze a 18 ks  observation of 18483 in quiescence, and report the results of the spectral and timing analysis of this observation. We found that the pulse fraction of the source X-ray emission decreased significantly with respect to that measured while the source was in outburst, and provide for the first time an estimate of the spin-period derivative of this source. We also analyzed all the available  /XRT observations of 18483, and studied the orbital variations of the source X-ray flux. A 1 ks  observation is also analyzed and provided an improved position of the X-ray source. The results from this study are then compared with those obtained recently on 18,  the other SFXT with a similar orbital period to that of 18483.  So far, the orbital period has been measured with certainty only in other two SFXTs, i.e. IGRJ16479-4514 [3.3194 d @jain09], and IGRJ11215-5952 [@romano07][^3]. IGRJ18483-0311: data analysis and results {#sec:observation} ========================================== [*XMM-Newton*]{} data {#sec:xmm} ----------------------  (@j01) observed 18483 on 2006 October 12, and the total good exposure time was 14.4 ks (we discarded observational intervals that were affected by a high background). The observation data files (ODFs) were processed to produce calibrated event lists using the standard  Science Analysis System (SAS 8.0). We used the [epproc]{} and [emproc]{} tasks for the EPIC-PN and the two MOS cameras, respectively. Source light curves and spectra were extracted in the 0.5–10 keV band, by using a circular extraction region with a radius of 20”. Background light curves and spectra were instead extracted by using a circular region with a radius of 50”. We used the SAS [backscale]{} task and the [lcmath]{} task in [Heasoft]{} (version 6.6.1) to account for the difference in extraction areas between source and background spectra and light curves, respectively. The times of all light curves were corrected to the barycentre of the Solar System with the SAS [barycen]{} task. In all cases, owing to poor statistics, the EPIC-MOS1 and EPIC-MOS2 cameras did not contribute significant additional information on the source spectra. Therefore, in the following we discuss only the spectra from the EPIC-PN camera. All spectra were rebinned in order to have at least 25 photons for each energy bin. In Fig. \[fig:lcurve\] we report the X-ray light curves of the source in the 0.5-5 keV and 5-10 keV energy bands; the lower panel of the figure shows the hardness ratio (i.e. the ratio of the count-rate in the hard, 5-10 keV, and soft, 0.5-5 keV, energy band) versus time. We note that the source count rate was decreasing during the first 5 ks of the observation. Unfortunately, the number of counts was insufficient to carry out any detailed investigation of the spectral variability. Therefore, we extracted only the 0.5-10 keV spectrum by using the total exposure time of the observation, and performed a fit with an absorbed power law model. The best fit parameters were $N_{\rm H}$=7.7$^{+1.2}_{-0.8}$$\times$10$^{22}$ cm$^{-2}$, and $\Gamma$=2.5$\pm$0.3 ($\chi^2_{\rm red}$/d.o.f.=1.3/39; hereafter errors are at 90% confidence level, unless otherwise indicated). The absorbed (unabsorbed) flux in the 0.5-10 keV band was 9.3$\times$10$^{-13}$ erg/cm$^2$/s (5.2$\times$10$^{-12}$ erg/cm$^2$/s). Assuming a source distance of 4 kpc, the unabsorbed flux corresponds to an X-ray luminosity (0.5-10 keV) of 1.0$\times$10$^{34}$ erg/s. Figure \[fig:spectrum\] shows the Epic-PN source spectrum, together with the best fit model and the residuals from this fit. The 90% confidence upper limit to the equivalent width for narrow iron lines at 6.4 keV and 6.7 keV is 0.13 keV and 0.10 keV, respectively. Timing analysis of the  data was carried out by using barycentred event files. We searched for the 21.0526 s spin-period of the NS in 18483 by performing first a power spectrum of the  data. No significant evidence for a peak at the corresponding frequency was found. In order to investigate further the presence of pulsations in the  data, we applied the Z$^{2}_{1}$-statistic technique [@buccheri88] to the photons event distribution for trial frequencies in a small window centered on 0.0475 Hz [@sguera07]. A spin-period of 21.025$\pm$0.005 s (hereafter errors on the NS spin periods are all at 1$\sigma$ confidence level) is found with a peak power of Z$^{2}_{1}$$\sim$26. The single-trial significance of this period is 4.7 $\sigma$. Figure \[fig:zstat\] shows the power spectrum computed with the Z$^{2}_{1}$-statistic technique by using the total exposure time of the  observation. This period estimate was then refined by employing a phase fitting technique [see e.g., @osso03]. This gave our best determined spin period of $P_{\rm spin}$=21.033$\pm$0.004 s. In order to derive the significance of this result over the entire range of spin periods considered, we assumed a spin period derivative of 1.3$\times$10$^{-9}$ s/s (see Sect. \[sec:discussion\]) and multiplied the single trial significance of the Z$^{2}_{1}$-statistic for the total number of trial $D_{\rm P}$/($P_{\rm spin}^2$/2$Dt$). Here $D_{\rm P}$ is the separation in seconds between our measured spin period and that reported by @sguera07, and $Dt$=14.4 ks is the total duration of the  observation. This gave us a significance of 3.7$\sigma$. From the folded light curve of the observation (obtained with the [efold]{} task, see Fig. \[fig:phase\]), we measured a pulsed fraction of F=15$\pm$3% in the 0.5-10 keV energy band. The profile is consistent with a sinusoid. In order to investigate the energy dependence of the pulse fraction, we also extracted and folded light curves in different energy bands and time intervals. We found that the pulsed fraction decreases slightly toward higher energies, whereas no significant variation could be measured across the “A” and “B” time intervals shown in Fig \[fig:lcurve\] (see Table \[tab:pulse\]). We note that the pulsed fraction we measured in the quiescent state of 18483 is a factor of $\sim$3 lower than that reported by @sguera07 during the source outburst[^4]. [@llll@]{} Energy Band & Total Obs. & A & B\ 0.5-10 keV & (15$\pm$3)$\%$ & $<$29$\%$ $^a$ & $<$32$\%$ $^a$\ 0.5-5 keV & (23$\pm$3)$\%$ & (30$\pm$5)$\%$ & $<$33$\%$ $^a$\ 5-10 keV & $<$18$\%$ $^a$ & $<$36$\%$ $^a$ & $<$29$\%$ $^a$\ \ \[tab:pulse\] [*Swift*]{} data {#sec:swift} ----------------- [@lllllllll@]{} IGRJ18483-0311\ OBS ID & INSTR & START TIME & STOP TIME & EXP & $N_{\rm H}$ & $\Gamma$ & $F_{\rm unabs}$ & $\chi^2_{\rm red}$/d.o.f.\ & & & & ks & (10$^{22}$ cm$^{-2}$) & & (erg/cm$^{2}$/s)\ 00035093001$^{\rm c}$ & XRT/PC & 2006-02-16 01:37:19 & 2006-02-16 22:36:57 & 7.9 & 6.0$^{+1.9}_{-1.6}$ & 1.4$^{+0.5}_{-0.4}$ & 5.3$\times$10$^{-11}$ & 1.3/21\ 00035093002$^{\rm c}$ & XRT/PC & 2006-03-05 11:12:39 & 2006-03-05 17:58:56 & 5.6 & 5.0$\pm$1.0 & 1.2$\pm$0.3 & 6.4$\times$10$^{-11}$ & 1.1/51\ 00035093003 & XRT/PC & 2008-09-26 13:49:39 & 2008-09-26 15:26:38 & 2.0 & 5.0 (fixed) & 1.2 (fixed) & 1.8$\times$10$^{-12}$ & —\ SAXJ1818.6-1703\ 00036128001 & XRT/PC & 2007-11-09 17:47:09 & 2007-11-09 21:04:57 & 1.6 & 6.0 (fixed) & 1.9 (fixed) & $<$2.1$\times$10$^{-12}$ $^{a,b}$ & —\ 00036128003 & XRT/PC & 2008-04-18 14:38:56 & 2008-04-18 17:45:49 & 2.0 & 6.0 (fixed) & 1.9 (fixed) & 3.5$\times$10$^{-12}$ $^{a}$ & —\ 00037889001 & XRT/PC & 2008-07-20 02:44:06 & 2008-07-21 01:21:56 & 3.7 & 6.0 (fixed) & 1.9 (fixed) & 8.5$\times$10$^{-12}$ $^{a}$ & —\ \ \ \ \ \[tab:log\] In Table \[tab:log\] we show a log of the  observations analyzed in the present study. Note that the observations ID 00035093001 and 00035093002 were also published previously by @sguera07. We used the [xrtpipeline]{} (v.0.12.1) task to process  /XRT data [note that part of these data were published by @sguera08]. Standard event grades of 0-12 were selected for the XRT photon-counting (PC) mode; filtering and selection criteria were applied using [ftools]{} ([Heasoft]{} v.6.6.1). We created exposure maps through the [xrtexpomap]{} task, and used the latest spectral redistribution matrices in the [Heasarc]{} calibration database (v.011). Ancillary response files, accounting for different extraction regions, vignetting and PSF corrections, were generated using the [xrtmkarf]{} task. When required, we corrected PC data for pile-up, and used the [xrtlccorr]{} to account for this correction in the background subtracted light curves. ![ /XRT light curves of the observations ID. 00035093001 (upper panel), and ID. 00035093002 (lower panel). The start times of the light curves in the upper and lower panel are 53782.0695 (MJD) and 53799.4690 (MJD), respectively.[]{data-label="fig:lcurve_swift"}](lcurve_swift.ps) For each observation in Table \[tab:log\], we extracted the light curve and spectrum, and derived a mean X-ray flux by fitting this spectrum with an absorbed power-law model (we used [Xspec]{} v.12.5.0). Spectra were rebinned in order to have at least 20 photons per bin and allow for $\chi^2$ fitting. In the observation ID. 00035093003 the very low source count rate (8.6$\pm$2.7$\times$10$^{-3}$) did not allow for a detailed spectral analysis. Therefore, we estimated the source count rate of the observation with [sosta]{} ([ximage]{} V.4.4.1), and then used this count rate within [webpimms]{}[^5] in order to derive the X-ray flux (we assumed the same spectral model of the observation ID. 00035093002). In Table \[tab:log\], we report the best fit parameters obtained using an absorbed power law model to characterize the source spectra; Fig.\[fig:lcurve\_swift\] shows the source light curves from the observations ID. 00035093001 and 00035093002 (note that, owing to poor statistics, we do not report the light curve of the observation ID. 00035093003). [@lllll@]{} IGRJ18483-0311 & & &\ INSTRUMENT & PHASE & $F_{\rm unabs}$$^a$ & $L_{\rm X}$$^c$ & ENERGY\ & & (erg/cm$^{2}$/s) & (erg/s) & RANGE\ IBIS/ISGRI$^1$ & 0 & 1.1$\times$10$^{-9}$ & 2.1$\times$10$^{36}$ & 20-100\ SWIFT/XRT$^{10}$ & 0.13 & 1.8$\times$10$^{-12}$ & 3.5$\times$10$^{33}$ & 0.5-10\ CHANDRA/HRC$^{10}$ & 0.24 & 4.3$\times$10$^{-10}$ & 8.3$\times$10$^{35}$ & 0.5-10\ XMM/Epic-pn$^{10}$ & 0.52 & 5.2$\times$10$^{-12}$ & 1.0$\times$10$^{34}$ & 0.5-10\ SWIFT/XRT$^{10}$ & 0.59 & 6.4$\times$10$^{-11}$ & 1.2$\times$10$^{35}$ & 0.5-10\ SWIFT/XRT$^{10}$ & 0.67 & 5.3$\times$10$^{-11}$ & 1.0$\times$10$^{35}$ & 0.5-10\ SAXJ1818.6-1703 & & &\ INSTRUMENT & PHASE & $F_{\rm unabs}$$^a$ & $L_{\rm X}$$^d$ & ENERGY\ & & (erg/cm$^{2}$/s) & (erg/s) & RANGE\ IBIS/ISGRI$^2$ & 0 & 3.8$\times$10$^{-10}$ & 2.9$\times$10$^{35}$ & 18-60\ IBIS/ISGRI$^4$ & 0.01 & 3.0$\times$10$^{-9}$ & 2.3$\times$10$^{36}$ & 18-60\ IBIS/ISGRI$^5$ & 0.06 & 3.8$\times$10$^{-10}$ & 2.9$\times$10$^{35}$ & 18-60\ SAX/WFC$^6$ & 0.07 & 2.1$\times$10$^{-9}$ & 1.6$\times$10$^{36}$ & 2-9\ SWIFT/XRT$^{10}$ & 0.12 & 3.5$\times$10$^{-12}$ & 2.6$\times$10$^{33}$ & 0.5-10\ SWIFT/XRT$^{10}$ & 0.22 & 8.5$\times$10$^{-12}$ & 6.4$\times$10$^{33}$ & 0.5-10\ XMM/Epic-pn$^7$ & 0.51 & $<$1.1$\times$10$^{-13}$$^b$ & $<$8.3$\times$10$^{31}$$^b$ & 0.5-10\ SWIFT/XRT$^{10}$ & 0.76 & $<$2.1$\times$10$^{-12}$$^b$ & $<$1.6$\times$10$^{33}$$^b$ & 0.5-10\ CHANDRA/ACIS-S$^{8,e}$ & 0.89 & 7.5$\times$10$^{-12}$ & 5.7$\times$10$^{33}$ & 0.5-10\ IBIS/ISGRI$^3$ & 0.91 & 3.0$\times$10$^{-10}$ & 2.3$\times$10$^{35}$ & 18-45\ IBIS/ISGRI+SWIFT/BAT$^2$ & 0.98 & $<$3.8$\times$10$^{-10}$$^b$ & $<$2.9$\times$10$^{35}$$^b$ & 18-60\ IBIS/ISGRI$^2$ & 1.00 & $<$3.8$\times$10$^{-10}$$^b$ & $<$2.9$\times$10$^{35}$$^b$ & 18-60\ IBIS/ISGRI+SWIFT/BAT$^2$ & 1.00 & 9.1$\times$10$^{-10}$ & 6.9$\times$10$^{35}$ & 18-60\ SWIFT/BAT$^9$ & 1.00 & 2.2$\times$10$^{-9}$ & 1.7$\times$10$^{36}$ & 15-150\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \[tab:comparison\] [*Chandra*]{} data {#sec:chandra} -------------------  observed 18483 on 2008 February 19 for a total exposure time of 1.2 ks with the High Resolution Camera. We reduced these data using the [ciao]{} software (v 4.1.1) and the latest calibration file available. The best source position is provided by the [wavdetect]{} task at $\alpha_{\rm J2000}$=18$^{\rm h}$48$^{\rm m}$17$\fs$2 and $\delta_{\rm J2000}$=-3${\degr}$10$\arcmin$16$\farcs$8 (the position accuracy is 0$\farcs$8 at a 90% confidence level). This is perfectly in agreement with the optical position reported by @rahoui08. We also derived the source count rate (0.51$\pm$0.02 cts/s) of the observation and then used this count rate with [webpimms]{} in order to estimate the source X-ray flux. The results are given in Table \[tab:comparison\] (we assumed the same spectral model of the  observation). SAXJ1818.6-1703: data analysis and results =========================================== [*Swift data*]{} {#sec:igrj1818} ---------------- 18 is the only SFXT whit an orbital period comparable to that of 18483 [30$\pm$0.1 days, @bird09; @zurita09]. In Table \[tab:comparison\] we report all observations of this source we found in the literature, together with two recent  /XRT observations that have not yet been published (ID. 00036128001 and 00037889001). We analyzed all these  /XRT observations with the procedures described in Sect. \[sec:swift\] and reported the results in Table \[tab:log\]. Following @bird09, we measured in Table \[tab:comparison\] the orbital phase of each observation from the epoch of the outburst occurred on 53671 MJD (phase 0), so as to permit a comparison with the orbital changes of the X-ray flux in 18483.  This comparison is carried out in Sect. \[sec:discussion\]. Discussion and conclusions {#sec:discussion} ========================== In this paper we reported on all available quiescent observations of 18483,  one of the two SFXTs for which the spin and orbital periods have been measured with certainty (the other is IGRJ11215-5952 with $P_{\rm spin}$=186.78 s, but see Sect. \[sec:intro\]). We report the detection of pulsations in the quiescent X-ray emission of this source, and give for the first time a measurement of its spin-period derivative. To our knowledge, the spin-period has so far been detected unambiguously in two other SFXTs [$P_{\rm spin}$=228 s, and 4.7 s in IGRJ16465-4507 and IGRJ1841.0-0536, respectively; @bamba01; @walter07]; however, the orbital period of these sources is not known. On the contrary, in the case of 18 the orbital period is known, but the spin-period remains to be discovered. Recently it has been suggested that a measurement of the NS spin and orbital periods can be the key to distinguish between different models proposed for SFXT sources [@bozzo08]. In fact, all these models involve a NS accreting from the intense wind of its supergiant companion, but several different mechanisms have been invoked in order to explain the very large luminosity swings observed during their transitions between outburst and quiescence [@zand05; @walter07]. In particular, @bozzo08 suggested that, if very slow spinning NSs ($P_{\rm spin}$$\gtrsim$1000 s) in relatively close orbits (few tens of days) are hosted in SFXTs, then a magnetic gating mechanism can be invoked in order to explain such luminosity swings. In this case the NS magnetic field would be in the “magnetar” range [i.e. $\gtrsim$10$^{14}$-10$^{15}$ G; @duncan92]. On the contrary, faster spin-periods might indicate that the large luminosity swings of SFXTs are caused by a centrifugal rather than a magnetic gating [a similar mechanism was suggested to explain the pronounced activity of Be X-ray pulsar transient systems; @stella86]. Alternatively, the observed variations in the X-ray luminosity of SFXTs might also be caused by drastic changes in the mass accretion rate onto the NS due to an extremely clumpy wind or to large scale structure in the immediate surroundings of the supergiant companion. In these models, the orbital periods may be as high as hundreds of days [see in particular @sidoli07; @negueruela08]. In 2008, an  observation of IGRJ16479-4514 revealed that also eclipse-like events can contribute to the luminosity swings observed in SFXTs [@bozzo08b]. Therefore, besides a measurement of the NS spin and orbital period, also an in-depth monitoring of the X-ray flux and spectral changes at different orbital phases is required in order to distinguish between different models or scenarios for SFXT sources. To this aim, we presented in Table \[tab:comparison\] an analysis of the orbital changes in the X-ray flux observed from 18483 and 18, the only two SFXTs with a comparable orbital period. In the case of 18483 only few observations have been carried out in quiescence and thus the orbital monitoring of this source is far from being complete [following @sguera07 we measured the source phases from the epoch of the brightest outburst observed with  at 53844.2 MJD]. The lowest flux state of this source was caught by  /XRT at phase 0.13, i.e. relatively close to the orbital phase where the highest X-ray activity of the source has been observed in several occasions. Unfortunately, the poor statistics of this  /XRT observation prevented an accurate spectral analysis, and thus we could not investigate the origin of this low flux state. In the other two  /XRT observations a spectral analysis could be carried out, but we did not detect any indication of a significant spectral variation. Only in the  observation we measured a slight increase in the spectral power law index. This suggests that X-ray flux changes in 18483 might have occurred due to genuine variations in the mass accretion rate onto the NS, rather than eclipse-like events. Note that, the detection of pulsations in the  data are also in agreement with the accretion scenario[^6]. This suggests that SFXTs undergo low level accretion even when they are not in outburst [see also @sidoli07]. At odds with the case of 18483, Table \[tab:comparison\] shows that the different orbital phases of 18 have been fairly well monitored. Unfortunately, the X-ray spectrum of this source could be well characterized only during the outburst, whereas in quiescence only the  observation provided a measurement of the spectral parameters (see Table \[tab:log\]). In all the other observations only a rough estimate of the source flux could be obtained. Note that the source was not detected by  at the orbital phase 0.52, and the 3$\sigma$ upper limit on the source X-ray flux was at least an order of magnitude lower than the fluxes measured in any other orbital phases. Since no spectral analysis could be carried out on 18 at this orbital phase, the origin of this low flux event could not be investigated further. In case future observations of 18 reveal that this source regularly undergoes X-ray eclipses at the orbital phase $\sim$0.5, this can help clarifying the issue of the extreme flux changes in this source. More observations of 18483 and 18 at different orbital phases with high sensitivity X-ray telescopes, like  and , are clearly required in order to understand unambiguously the origin of their outburst/quiescent activity. Being these two sources the only SFXTs with a comparable orbital period, they are very well suited to test different models proposed to explain the behavior of SFXTs. We are currently investigating the results of the application of the gated accretion model to 18483 and 18 (Bozzo et al., 2009, in preparation). In this paper, besides X-ray flux changes, we also measured a spin-period variation in 18483. By using our best-determined spin-period, $P_{\rm spin}$=21.033$\pm$0.004 s, and that found previously by @sguera07, we obtained a spin-period derivative in 18483 of $\dot{P}_{\rm spin}$=-(1.3$\pm$0.3)$\times$10$^{-9}$ s s$^{-1}$. This value is comparable with the spin-period derivative measured in the case of the SFXT AXJ1841.0-0535 [-1.5$\times$10$^{-10}$ s/s @sidoli08] and those induced by accretion torques in wind-fed binaries [see e.g., @bildsten97]. However, in the present case we believe that the spin-period derivative most likely results from light travel time effects in the binary. In fact, in a binary system with an orbital period of $\sim$18.5 days, these effects can contribute to an apparent spin-period derivative of the order of $\sim$$v_{\rm orb}$/$c$=8.6$\times$10$^{-4}$ s s$^{-1}$, i.e. much larger than the spin-period derivative we reported above [here $v_{\rm orb}$ is the orbital velocity and $c$ is the light velocity; we used the mass and radius of the supergiant companion measured by @rahoui08]. Unfortunately, since a detailed orbital solution for this source is not yet available, we do not know if accretion torques acting onto the NS in 18483 might also have contributed to the observed spin-period derivative. Note that, in principle, this can be used to study the interaction between the NS and the inflowing matter from the supergiant companion [see e.g. @bozzo08]. An orbital monitoring of 18483 is required in order to understand the origin of the measured spin-period derivative. Acknowledgments {#acknowledgments .unnumbered} =============== We thank the anonymous referee for his/her many helpful comments. EB thank M. Capalbi, M. Perri, and K. Page for kind help with the  /XRT data analysis. This work was partially supported through ASI and MIUR grants. Bamba, A., Yokogawa, J., Ueno, M., Koyama, K., & Yamauchi, S. 2001, PASJ, 53, 1179 Barthelmy, S.D., Krimm, H.A., Markwardt, C.B., Palmer, D.M., Ukwatta, T.N. 2008, GCN, 7419, 1 Bird, A.J., et al. 2009, MNRAS, 393, 11 Bildsten, et al. 1997, ApJS, 113, 367 Bozzo, E., Falanga M., Stella L. 2008, ApJ, 683, 1031 Bozzo, E., Stella, L., Israel, G., Falanga, M., Campana, S. 2008b, MNRAS, 391, L108 Bozzo, E., et al. 2008c, Astr. Tel., 1493 Buccheri, R. 1983, A&A, 128, 245 Chernyakova, M., Lutovinov, A., Capitanio, F., Lund, N., Gehrels, N. 2003, Astr. Tel., 157 Duncan, R. C. & Thompson, C. 1992, ApJ, 392, L9 Grebenev, S. A.; Sunyaev, R. A. 2005, AstL, 31, 672 Grebenev, S. A.; Sunyaev, R. A. 2008, ATel, 1482, 1 in ’t Zand, J., Heise, J., Smith, M., Muller, J. M., Ubertini, P., Bazzano, A. 1998, IAUC, 6840, 2 in ’t Zand 2005, A&A, 441, L1 in’t Zand, J., Jonker, P., Mendez, M., Markwardt, C. 2006, Astr. Tel., 915 Chetana, J., Biswajit, P., Anjan, D. 2009, MNRAS, in press \[astro-ph/0903.5403\] Dall’Osso, S., Israel, G.L., Stella, L., Possenti, A., Perozzi, E. 2003, ApJ, 599, 485 Jansen, A., Lumb, D., Altieri B., et al. 2001, A&A, 365, L1 Levine, A.M., Corbet, R. 2006, Astr. Tel., 940 Masetti, N. et al. 2008, A&A, 482, 113 Molkov, S.V., Cherepashchuk, A.M., Lutovinov, A.A., Revnivtsev, M.G., Postnov, K.A., Sunyaev, R.A. 2004, AstL., 30, 534 Negueruela, I., Torrejon, J. M., Reig, P., Ribo, M., & Smith, D. M. 2008, AIPC, 1010, 252 Rahoui, F., Chaty, S., Lagage, P., Pantin, E. 2008, A&A, 484, 801 Romano, P. et al. 2007, Astr. Tel., 1151 Sguera, V., et al. 2005, A&A, 444, 221 Sguera, V., et al. 2006, ApJ, 646, 452 Sguera, V., et al. 2007, A&A, 467, 249 Sguera, V., et al. 2008, A&A, 487, 619 Sidoli, L., et al. 2007, A&A, 476, 1307 Sidoli, L., et al. 2008, ApJ, 687, 1230 Stella, L., White, N. E., & Rosner, R. 1986, ApJ, 308, 669 Walter, R., et al. 2006, A&A, 453, 133 Walter, R. & Zurita Heras, J. A. 2007, A&A, 476, 335 Zurita Heras, J.A., Chaty, S. 2009, A&A, 493, L1 [^1]: email: [email protected] [^2]: Here we defined the pulsed fraction as $F$=($I_{\rm max}$-$I_{\rm min}$)/($I_{\rm max}$+$I_{\rm min}$), where $I_{\rm max}$ and $I_{\rm min}$ are the measured count rates at the maximum and at the minimum of the folded light curve, respectively. [^3]: However note that the behavior of IGRJ11215-5952 is somewhat peculiar with respect to the other SFXTs, and thus @walter07 excluded this system from their SFXT source list. [^4]: However, note that these pulse fractions are measured in slightly different energy bands (see Sect. \[sec:intro\]). [^5]: http://heasarc.nasa.gov/Tools/w3pimms.html [^6]: Pulsations in quiescence were also reported for other two SFXT sources, i.e. IGRJ16465-4507 [@walter06] and AXJ1841.0-0535 [@sidoli08].

ArXiv

--- abstract: 'Central diffractive production of heavy states (massive dijets, Higgs boson) is studied in the exclusive mode using a new [*Hybrid Pomeron Model*]{} (HPM). Built from Hybrid Pomerons defined by the combination of one hard and one soft color exchanges, the model describes well the centrally produced diffractive dijet data at the Tevatron. Predictions for the Higgs boson and dijet exclusive production at the LHC are presented.' author: - Robi Peschanski - 'M. Rangel' - 'C. Royon' title: Hybrid Pomeron Model of exclusive central diffractive production --- [Introduction: The Hybrid Pomeron]{} ==================================== Central diffractive production of heavy objects in its $exclusive$ mode (no other particle produced in the central rapidity region) appears as a promising complementary tool for the study of new particles at the LHC, such as the Higgs boson. Indeed, for instance, the mass determination can be made quite precise, if both incident protons are detected and measured in forward detectors located at 220 and 420 m from the interaction point at the LHC [@forward; @al00]. One expects to take advantage of the absence of other particles than the decay products in the central rapidity region and some other interesting aspects such as the depletion of b-quark production due to the helicity rule specific of this production mode [@helicity]. The key problem of central diffractive production in the exclusive mode is to determine its rate as a function of $e.g.$ the Higgs boson mass or the minimum $p_T$ of the jets. Experimental results on massive dijet production at the Tevatron has shown indirect evidence for exclusive production, by comparison with models of $inclusive$ diffractive production. Inclusive models [@cdf; @Kepka:2007nr] agree to point out an excess of events over the inclusive spectrum in the kinematical region where exclusive production is expected to contribute. The experimental interest of central exclusive production, in the first place for the Higgs boson and $\gamma$ induced processes, and the preparation of concrete proposals at the LHC is a major incentive for theorists to work out reliable predictions for the production cross section, which could serve as a basis for the necessary data simulations. This task is not easy since central diffractive processes imply both hard subprocesses, related to the high mass of the centrally produced states, and soft ones which are typical of diffractive events which leave intact the initial particles -$e.g.$ protons at the LHC. In some sense one could say that central diffractive production is expected to combine the “hardest” events such as the production of massive Higgs bosons or of any high mass object (dijet, diphoton...), with the “softest” ones, since the initial particles remain totally intact (up to a loss of energy not bigger than 10 %). This reveals the potentially $hybrid$ character of central diffractive production. On the theoretical ground, different mechanisms of exclusive central diffraction have been proposed since years [@others], but we will restrict to two classes of models which are based on the exchange of colorless objects, in order to take into account the diffractive property. Indeed, any colored object would generate particle production in the whole rapidity interval[^1]. One class is based on the exchange of two Pomerons, where the Pomeron is the colorless exchange which appear in $e.g.$ elastic reactions; it can be called the Non Perturbative Model (NPM) and was based on a typical soft interaction hypothesis, which comes from the Bialas-Landshoff mechanism [@bialas1] originally proposed for central diffusive production. It has an $inclusive$ version which describes the inclusive diffractive dijet production [@us], while its $exclusive$ version has been studied in Ref.[@us1]. One another class of models is based on the exchange of two gluons at each vertex for the exclusive production [@khoze] called KMR (from the author names) in the following. For both models there exists a detailed phenomenological discussion (see $e.g. $[@Kepka:2007nr]) based on dedicated simulations. Let us recall the present status of this physically meaning discussion. The inclusive production mechanism based on the NPM [@us] gives satisfactory results when compared to Tevatron data. Using this agreement, the extraction of the exclusive component in the DPE framework becomes possible, since it appears to be necessary to include it in a well-defined region of phase-space. When comparing [@cdf; @Kepka:2007nr] the extracted dijet cross section and spectra with the models, it appears that the KMR model [@khoze] gives a better description of the results than NPM [@us]. The main reason is that it takes into account the Sudakov suppression factors preventing cross section s to include the gluon radiation normally associated with the production of a massive object. The soft Pomeron exchanges of NPM [@us] do not contain these perturbative QCD factors and give a too flat distribution as a function of the minimum transverse momentum $p_T^{min}$ of the jet [@Kepka:2007nr]. As a consequence, the prediction for the Higgs boson cross section, which was similar for both models for a light Higgs boson [@Forshaw:2005qp], has a different form as a function of the Higgs boson mass, being steeper for the KMR model [@khoze] than for NPM [@us]. It is expected that the NPM model in Ref. [@us] works at low masses (for instance for $\chi_C$ production [@chic]) whereas a model including a hard contribution may be valid at higher masses. Our motivation is to keep the Double Pomeron Exchange (DPE) hypothesis, while taking into account the fact that the diffractive production process is expected to be a mixture of soft and hard color exchanges. Indeed, the notion of a hard Pomeron (associated in QCD with the summation of ladder diagrams in the leading or next-to-leading logarithmic approximation (LLA) of the perturbative expansion) is a theoretical result of QCD [@BFKL]. Moreover it has been successfully compared with data in (hard) inclusive diffraction (see, $e.g.$ [@Bialas:1997vt; @Marquet:2007nf]) and exclusive vector-meson production [@marquet]. The example of heavy vector meson production, in particular, is well suited for our approach since it corresponds to the (quasi-) elastic production of a heavy state, which can be formulated in the framework of a hard Pomeron exchange. In the theoretical calculations, the hard Pomeron appears to correspond to ladder diagrams connecting two exchanged reggeized colored gluons [@BFKL]. However, in central diffractive production, one could expect to have two different colored exchanges, one hard and one soft. It would correspond physically to two time scales, one short corresponding to the heavy state production, and one long corresponding to the necessary color neutralization. Hence, the qualitative picture of central diffractive production which we formulate is a DPE process in which each Pomeron exchange at the vertex would correspond to $hybrid$ Pomerons with two different types of color exchanges one soft and one hard. It would correspond to an $effective$ way of summing ladder diagrams between hard and soft colored reggeized gluons, which precise calculation remains beyond our scope (and beyond the present knowledge of non perturbative QCD physics). The plan of the paper is the following. In the next section, we shall formulate the Hybrid Pomeron Model (HPM) and determine its parameters obtained from known soft and hard Pomeron processes. In section III, we will show its good description of exclusive dijet production extracted from data at the Tevatron and the prediction for Higgs Boson and dijet production at the LHC. The last section is for discussions, conclusions and an outlook. Formulation: The Hybrid Pomeron Model ===================================== The model, adopting as a starting point the idea of the original Bialas-Landshoff formulation, consists in defining effective propagators and couplings for the colored exchanges associated with central DPE processes, see Fig.\[1\]. However, by contrast with the original NPM model of Ref.[@bialas1], we introduce two types of propagators and couplings, depending of its soft or hard character. The soft propagator $D_S$ and coupling $G_S$ are exactly those which appear in the original description of the Bialas-Landshoff model [@bialas1], themselves connected to the soft Pomeron Landshoff-Nachtmann formulation [@LN] of the elastic cross section, see Fig.\[2\]. They are constrained to describe the elastic hadronic cross section, which fixes its parameters. \[htb\] [1]{} The new aspect, w.r.t. the original formulation [@bialas1], is to introduce similarly effective propagators and couplings $D_H$ and coupling $G_H$ for hard Pomeron processes. Since we are formulating a [*hybrid Pomeron Model*]{} (HPM), we would optimally need the resummation of QCD ladder diagrams corresponding to both soft and hard colored exchanges. As we see in Fig. \[1\], the hard exchange produces the heavy state object ($D_H G_H^2$) while the colorless aspect of the exchange is ensured via the emission of a soft additional gluon ($D_S G_S^2$). This means that most of the available momentum is carried away by one of the gluon, the hard one, while the soft one carries only a very small fraction of the proton momentum. The hard part of the HPM model will be based on hard physics measured at HERA (for instance the proton structure function $F_2$) while the soft part will be based on usual soft cross section measurements. In the case of simple elementary gluon exchanges, as developed in the model [@khoze], the problem is perturbatively tractable[^2], since the loop kinematics enforces a (semi-) perturbative calculation. However, when considering Pomeron ladders, the gluon loop constraint, characteristic of the mechanism of [@khoze] does not hold and thus one relies on the Bialas-Landshoff modified picture in order to include both soft and hard effective color exchanges, see Fig.\[1\]. Hence our proposal is to start with the description of hard Pomeron scattering in terms of effective hard colored propagators $D_H$ and couplings $G_H$, in the same way as for the soft color exchanges in [@bialas1]. For this sake we consider the well-known dipole-proton amplitudes which appears in the QCD description of many hard processes. They will be used to determine the effective propagators and couplings. In that sense, it is possible to fix the parameters of the model using hard physics measurements at HERA, especially from the measurements of the proton structure function and the vector meson production cross sections. In this basic process, a dipole of size $r$ experiences an elastic scattering with the proton. Since this dipole-proton amplitude, corresponding to an hard Pomeron exchange, appears in the formulation of different observables, its parameters are well determined, and thus gives the possibility to define the appropriate hard propagators $D_H$ and couplings $G_H$, in the same way as was done for the soft ones, but with the advantage that we have a theoretical control on its precise QCD formulation. A comment has to be made at this stage. The main new aspect of HPM is to introduce a formulation for hard color exchanges. Since it is a phenomenological effective description of diagrams going beyond elementary gluon exchanges, it aims at keeping the physical image of two different time scales and thus of two different types of effective propagators. Hence the virtuality associated with the hard color exchanges cannot be transferred to the other color exchange through the loop kinematics, as is the case in the model [@khoze]. On the other hand the inclusion of hard color exchanges in the DPE formulation is expected to (and indeed will, as we shall see) correct the drawbacks of the initial soft model. **Soft color exchange** ----------------------- We evaluate the non-perturbative gluon propagator from the elastic proton-proton data, see Fig. \[2\]. Following Landshoff-Nachtmann proposal [@LN], the elastic hadron-hadron scattering is represented by the contributions of elastic valence quark scattering mediated by a non-perturbative model for gluon exchange. The elastic quark-quark amplitude in terms of soft propagator and coupling writes[^3] $$A_{qq} \equiv G_S^{2}\ D_S = s^{{\alpha_{\cal P}}(t)}\ G_s^{2}\ D_S^{(0)}\ e^{\ \frac t{\mu_S^{2}}}\ , {\label}{ampli}$$ where $s$ is the total c.o.m. energy, $t$ the transfer quadrimoment squared whose dependence is approximated by an exponential slope given by $\mu_S$ and $${\alpha_{\cal P}}(t) \equiv {\alpha_{\cal P}}(0) + {\alpha_{\cal P}}' \log s = 1+{\epsilon}+ {\alpha_{\cal P}}' \log s {\label}{regge}$$ is the soft Pomeron Regge trajectory [@pom], with ${\epsilon}\sim .08$ being the Pomeron “anomalous intercept”. Note that we have incorporated the factors due to reggeization [@bialas1] in the definition of the propagators. This is required in order to take into account the different Regge parameters (and in particular the known different energy dependence) between the soft and hard Pomeron ingredients. In other terms the [*hybrid Pomeron*]{} will have an intermediate energy dependence compared to the soft and the hard Pomeron’s ones. \[t\] All in all, the differential elastic hadronic cross section, from which the relevant parameters will be obtained, is given in a suitable normalization, by $$\frac{d\sigma}{dt} \equiv \frac{1}{4\pi s^2}\left\{9 A_{qq}\right\}^2 = \left| 3\beta \right|^{4} s^{2{\alpha_{\cal P}}(0)-2} \exp \left[( 4 b + 2{\alpha_{\cal P}}\log s )\ t \right]\ , {\label}{sigma}$$ where the parameters $\beta$ can be obtained [@bialas1] from the total cross section and $b$ from the elastic form factor or equivalently, from the differential cross section s. Note that the factor $\left\{9 A_{qq}\right\}$ comes from the number of valence quark combinations, which are considered independent in the Landshoff-Nachtmann formulation. Using the effective propagator and coupling formulation of , one finally determines $$\begin{split} &\mu_{S}^{-2} = 2b + \alpha'\log s\\ &\left[G_{S}^{2}\ D_{S}^{(0)}\right]^2 = 8 \beta^2 s^{2\epsilon} \exp (4b + 2\alpha'\log s) \ . \end{split}$$ **Hard color exchange** ----------------------- For the definition of the propagator and coupling of the hard color exchange, we will use the well-known hard Pomeron for the dipole proton elastic amplitude calculated from perturbative QCD using as a starting point the Balitsky Fadin Kuraev Lipatov (BFKL) equation [@BFKL]. It is convenient to open the possibility of saturation effects, even if they are not expected to be important in the kinematical domain we are interested in. Indeed, this form of the dipole proton amplitude (eventually modified by saturation contributions) has been proven to be phenomenologically successful in the description of proton total and diffractive structure functions [@Bialas:1997vt] and, more importantly for our analysis, for structure function $F_2$ measured at HERA including the charm contribution [@soyez] for vector meson elastic differential cross section  [@marquet] and for inelastic diffraction [@Marquet:2007nf], which will be used for parameter fixing. Hence the model we will adopt for dipole-proton elastic scattering contains saturation effects and [$\mathrm{|t|}$]{}dependence [@marquet]. \[hbt\] [3]{} We start with the following amplitude in terms of the BFKL kernel. One writes $${{\cal N}}(r,Y)=\int_{\cal C} \frac{d\gamma}{2i\pi}\, {{\cal N}}_0(\gamma)\,r^{2{\gamma}} \exp \left\{ {\bar\alpha\chi(\gamma)} Y\right\}\ , {\label}{equa1}$$ where $\chi(\gamma)$ is the Mellin transform of the BFKL kernel [@BFKL]. ${{\cal N}}_0(\gamma)$ contains information on the coupling to the proton and other normalization contributions. The effect of saturation, through nonlinear damping factors, is known [@munier] to select a $critical$ value ${\gamma}_c.$ this corresponds to a “anomalous dimension” $d_c = {\gamma}_c\!-\!1$ which is characteristic of a (saturation-corrected) hard Pomeron. In a more concrete way, the authors of [@marquet] make use of a model for the dipole-proton amplitude [@Iancu:2003ge] which successfully describes the precise proton structure function data. $$\begin{aligned} {{\cal N}}(r,Y) & = & N_{0}\ (P_{H})^{{\gamma}_c} \exp\left( - \frac{\log^2(P_H)}{2\kappa \lambda Y} \right) \exp(-B|t|) \nonumber\\ P_H & = & (r /r_{S})^{2} \nonumber \\ & & \nonumber \\ r_{S}^{2} & = & r_{0}^{2} \exp(-\lambda Y)\ , {\label}{equa2}\end{aligned}$$ where $Q_S \equiv 2/r_s$ is the well-known saturation scale, the term $\exp(-B|t|)$ has been added in order to take into account the momentum transfer dependence in vector meson production [@marquet]. The exponential term in takes into account the contribution from the kernel variation around the saddle-point. Note that this amplitude works [@Iancu:2003ge] in the region $P_{H} < 1$, which is safely true in our case. The saturation corrections are expected to be negligible in that region. It is important to notice at this stage that we need to consider [ *amputated*]{} amplitudes, that is multiplying the expressions (\[equa1\],\[equa2\]) of the dipole-proton amplitude by a factor $r^{-2}.$ Indeed, we have to remove from the usual dipole proton amplitude the factor corresponding to the geometrical dimension of the dipole cross section proportional to $r^{2}$, or on other words the gluon dipole coupling. This factor has to be removed in order to define properly the couplings and propagators of the hard effective color exchanges, which would be valid for any massive state. We are interested in applying our formalism to the exclusive production of massive dijets or the Higgs boson and thus have to switch from a dipole state to the wave function coresponding to the heavy state under study. In the kinematical configuration of central diffractive production, we have at each hard color exchange vertex (see Fig. \[1\]) $$\begin{aligned} Y = -log(\xi) \ ,{\rm \ and\ } r^2 \sim \frac{A}{M^2} {\label}{Y}\end{aligned}$$ where we used a very simple[^4] relation (with $A \sim 7$ phenomenologically) between the mass $M$ of the heavy state and the corresponding dipole size $r$ to be considered. The normalization factors $N_{0}$ for the amplitude and the scale $r_{0}$ are also determined phenomenologically from HERA data. For dipole-proton scattering, we assume also a dominance of “valence” quark-quark scattering and 6 quark-quark combinations are allowed. The main characteristic feature of the hard Pomeron by contrast with the soft one is that it has a non trivial dependence on $Y$ and $r$ (translating into a non trivial $\xi$ and $M$ dependence in the central diffraction kinematics) through the anomalous dimension ${\gamma}_c$. Indeed, this perturbative QCD dependence plays the role of the Sudakov form factors in a BFKL-like model. It acquires also a different, faster, energy dependence through the dependence on $P_H$ in . Using concretely the parameters from the fit to the HERA data [@soyez; @marquet], one finds the following expression for the couplings and propagators of the hard gluon exchanges: $$\begin{aligned} \label{eq:shgluon} \mu_H &=& 0.5 \nonumber\\ \left[G_{H}^{2} D^{(0)}_{H}\right]^{2} &=& \frac{8\pi}{6\mu_H^2} \ \times \ 2\pi R_p^2 N_0 \times {r}^{-2} \times (P_{H})^{{\gamma}_C} \exp\left( \frac{\log^{2}(P_{H})}{2{\kappa}\lambda \log(\xi)} \right)\ ,\end{aligned}$$ where we have used the values for ${\kappa},\lambda,B$ (see formulae ) taken from the phenomenological analysis [@soyez; @marquet] of massive vector mesons, charm and stucture function measurements at HERA [@heradata]. The different parameters used in the model are given in Table \[param\]. The Sudakov suppression term in this model is given through the hard pomeron characteristics and the gluon radiation is thus suppressed thanks to the hybrid structure of HPM. **The central diffractive cross section** ------------------------------------------ All in all, and following the scheme depicted in Fig.\[1\], one has the following matrix element for the central exclusive diffractive production of a massive state: $$\label{eq:cs1} |M|^{2} = (D_{S} G_{S}^{2})^{2} \; \left([D_{H}G_{H}^{2}]_{1}\right)^{2} \; \left([D_{H}G_{H}^{2}]_{2}\right)^{2} \; |M_{\hat{\sigma}}|^{2} \; $$ The notation $[D_{H}G_{H}^{2}]_{i}$, i=1,2, is used to distinguish the hard colored exchanges from each vertex, see Fig.\[1\]. $M_{\hat{\sigma}}$ is the hard process matrix element for the considered produced massive state. In parallel with the approach of Ref. [@bialas1], the cross section is written as: $$\begin{aligned} \label{eq:cs2} \sigma & = & 81 \times \frac{2s}{(2\pi)^{5}} \times \left[G_{S}^{2}\ D_{S}^{(0)} \right]^2 \; \int d^{4} p_{1} d^4 p_{2}~ \delta(p_{1}^{2}) \delta(p_{2}^{2}) \; \delta((p_{a}\!+\!p_{b}\!-\!p_{1}\!-\!p_{2})^{2} - M^{2}) \times \\ \nonumber &\times&\left (\frac{s}{s_1} \right )^{2 {\alpha_{\cal P}}(t1) -2} \left (\frac{s}{s_2} \right )^{2 {\alpha_{\cal P}}(t2) -2} e^{2bt_1} e^{2bt2} \left[G_{H}^{2}\ D_{H}^{(0)} \right]_{1}^2 \left[G_{H}^{2}\ D_{H}^{(0)} \right]_{2}^2 ~|M_{\hat{\sigma}}|^{2} \ .\end{aligned}$$ Using relation [@bialas2] $$\int d^{4}p_{i} \delta(p_{i}^{2}) = -\frac{1}{2} \int d\xi_{i}\ d^{2}\vec{v_{i}} \ ;\quad \frac{s}{s_{i}} = \frac{1}{\xi_{i}} {\label}{variables}$$ where $\vec{v_{i}}$ is the transverse momentum of the final protons and changing the variable $v_{i}$ to $|t_{i}|$ using $|\vec{v_{i}}|^{2} = (1-\xi_{1})|t_{i}|,$ one finally finds: $$\begin{aligned} \label{eq:cs3} \sigma & = & \frac{81}{2(2\pi)^3} \times \left[G_{S}^{2}\ D_{S}^{(0)} \right]^2 \times \prod_{i=1,2} \left( \int \int d\xi_{i} d|t_{i}| \frac{1-\xi_{i}}{{\xi_{i}}^{2\epsilon}} \exp(-(2b+2{\alpha_{\cal P}}'\log(\frac{1}{\xi_{i}}) |t_{i}| ) ) \left[G_{H}^{2}\ D_{H}^{(0)} \right]_{i}^2 |M_{\hat{\sigma}}| \right)~.\end{aligned}$$ **Comparison with Dijet CDF data** ================================== Model implementation in FPMC ---------------------------- The model has been fully implemented in FPMC [@fpmc], using the parameters defined in the previous sections. The different parameters in the hard part of the model come mainly from a fit to HERA data (structure function $F_2$, charm and vector meson data) inspired by saturation models. By default, we take the parameters from a fit to the diffractive structure function $F_2$ measured by the H1 and ZEUS collaborations at HERA [@soyez]. The systematics uncertainties on the fit parameters define the systematic uncertainties of our model. In addition, it is possible to include heavy quarks in the model [@soyez], and compare it to the vector meson production cross section [@marquet], which leads to different parameters of the model (see Table I). The difference of the results with and without including charm effects is also a kind of systematic uncertainty in the model and will be discussed further in the paper. In addition, the parameters related to the soft exchange come from the Donnachie-Landshoff model. All parameters are given for reference in Table \[param\]. The only parameter in the model is the free normalisation which we will obtain from a fit to the CDF exclusive diffractive measurements. Implicitely, the normalisation will thus include the surival probability. Note that the ratio of the survival probabilities between the Tevatron (0.1) and the LHC (0.03) is taken into account when we predict later on the cross sections at the LHC. The implementation in FPMC [@fpmc] allows to interface the hybrid model with a jet algorithm after hadronisation performed in HERWIG [@herwig]. The standard jet algorithm [@cdfalgo] used by the CDF collaboration has been implemented so that we are able to compare directly our model with the CDF measurements of exclusive events. Parameter Central value Uncertainties Charm included ------------------------- ------------------ --------------- ---------------- [**Hard parameters**]{} $N_0$ 0.7 - 0.7 $Q_0$ 0.254 GeV 0.243-0.263 0.298 $R_p$ 3.277 GeV$^{-1}$ 3.233-3.321 3.344 $\gamma_C$ 0.6194 0.6103-0.6285 0.7376 $\kappa$ 9.9 - 9.9 $\lambda$ 0.2545 0.2494-0.2596 0.2197 $B$ 2 - 2 $\mu_H$ 0.5 - 0.5 [**Soft parameters**]{} $\alpha_P(0)$ 1.08 - 1.08 $\alpha'$ 0.06 - 0.06 $\beta$ 4 - 4 $b$ 4 3-5 4 : [*List of parameters used in the HPM.*]{} The second column give the default values used in the model, the third one the range of values used for systematics coming from the fit uncertainties to $F_2$ data, and the fourth one the values of parameters when heavy quarks are also considered in the model (see text).[]{data-label="param"} Comparison with CDF data ------------------------ To test the accuracy of the model it is useful to compare with the CDF measurements of exclusive events in the dijet channel at the Tevatron [@cdf; @Kepka:2007nr]. CDF used the dijet mass fraction to quantify the amount of exclusive events. The dijet mass fraction, namely the ratio of the dijet mass to the total mass in dijet events, is expected to peak around $1$ for exclusive events since two jets and nothing else are produced in the final state while inclusive events show lower values of the dijet mass fraction. The comparison between the CDF measurement and what is expected from inclusive diffraction based from quark and gluon densities measured at HERA (including the survival probability) leads to an estimate of the exclusive event cross section. The result is given in Fig. \[cdf1\]. Data points show the exclusive cross section for jets with a transverse momentum greater than a threshold value given in abscissa. To compare with the expectation from HPM, the FPMC Monte Carlo was interfaced with the jet cone algorithm used by the CDF collaboration at hadron level. Since the normalisation is not determined by the model, we choose to fix it using the CDF measurement. The global normalisation is obtained by fitting our predictions to the CDF measurement given in Fig. \[cdf1\]. The normalisation is found to be: $3.85 \times 10^{-4} \pm 1.89 \times 10^{-4}$ with $\chi^2=0.67$ for 5 data points and the uncertainty comes from the uncertainty on the CDF measurement. In Fig. \[cdf1\], we give the prediction of HPM in full line, and the dashed line shows the uncertainty on normalisation ($\pm 1 \sigma$) coming from the fit to the CDF data. We note that the shape of the HPM prediction describes nicely the CDF data while the normalisation comes directly from the CDF data as we mentioned previously. [cdf1]{} In Fig. \[cdf2\], we compare the predictions from the hybrid model to the dijet mass measurements in diffractive exclusive events from the CDF collaboration. As explained in the CDF paper [@cdf], this is an indirect measurement which is MC dependent due to the method used by the CDF collaboration to extract the dijet mass cross section. We follow the same method used by the CDF collaboration to compute the dijet mass cross section. Namely, we convert the measured exclusive dijet cross section from CDF presented in Fig. \[cdf1\] to a cross section versus dijet mass using the HPM. After each $E_{T_{min}}$ cut (10, 15, 20, 25, and 35 GeV), we normalise the HPM cross section to the CDF measurement. We have thus a “calibration" factor in each $E_{T_{min}}$ interval. The $M_{JJ}$ distribution coming from the hybrid model is then reweighted after applying the $E_{T_{min}}$ cut using the same calibration factors. Removing the cuts on $E_{T_ {min}}$ allows to obtain the “CDF points" given in Fig. \[cdf2\]. We followed basically the same procedure as in Ref [@cdf], but using the reweighted HPM instead of KMR. It is worth noticing that it is not strictly speaking a measurement by the CDF collaboration since it is model dependent. Nevertheless, we can now compare the “CDF measurement" to the expectation of the hybrid model and the result is shown in Fig. \[cdf2\]. The dashed line indicates the uncertainties on the model related to the normalisation. The model leads to a good description of CDF data over the full dijet mass range. [cdf2]{} Uncertainties on the model predictions -------------------------------------- In this section, we discuss the uncertainties related to the chosen values of parameters given in Table \[param\]. The first uncertainties come from the uncertainties on the parameter used to describe the hard interaction. As we mentioned already, the values of the parameters are taken from a fit to $F_2$ data coming from the HERA experiments [@soyez]. The values of the parameters found in Ref. [@soyez] where obtained with a given uncertainty coming from the fit procedure and it is worth checking the effect on the HPM predictions. There was also another kind of fits performed in Ref. [@soyez] where heavy quarks were considered and we also compare our predictions including or not the heavy quarks. The values of the parameters are given in Table \[param\] for references. It is worth noticing that we use the same values of parameters coming from a fit to HERA data to extrapolate at LHC energies, especially when we predict the exclusive Higgs boson cross section. It will be thus important to test the values of the parameters using directly LHC data when they will be available, and to study whether this assumption is valid. The effect of changing the hard parameters are given in Fig. \[models1\] for the jet $E_{T_{min}}$ and the $M_{JJ}$ distributions. The differences are found to be less than 20 %. Another systematic study we performed was to change the $b$ slope of the soft cross section responsible for the soft interaction. The uncertainty on the $b$ slope coming from soft data is quite small but we wanted to study the dependence of our model on this parameter. Modifying the $b$ parameter from 2 to 4 leads to the cross sections given in Fig. \[models3\] for the jet $E_{T_{min}}$ and the $M_{JJ}$ distributions. The difference is found to be less than 20% everywhere. It is worth noticing that the leading uncertainty in the predictions for HPM comes from the statistical uncertainties of the $E_{T_{min}}$ cross section measurement by the CDF collaboration which is of the order of 50%. The effect of taking the parameters of the fit of Ref. [@soyez] where heavy quarks are considered are given in Fig. \[models5\]. We recomputed the normalisation by fitting the $E_{T_{min}}$ distribution to the CDF data and the normalisation for the light quark only model is 6.80 $\times$ 10$^{-3}$ $\pm$ 3.46 $\times$ 10$^{-3}$ with a $\chi^2$ of 0.83 for 5 data points. We notice that the mass dependence is stronger when heavy quarks are considered, which means that the cross section at high mass is slightly smaller, and that the fit to the CDF data on $E_{T_{min}}$ is slightly worse. [models1]{} [models3]{} [models5]{} **Predictions for the LHC** --------------------------- In Fig. \[lhc1\], we show the exclusive Higgs boson cross section using the HPM. The cross section varies from 1.1 $\pm$ 0.5 fb at 120 GeV to 0.32 $\pm$ 0.15 fb at 160 GeV. Including heavy quark effects reduces this cross section by about 60%. The values are found to be slightly lower than with the KMR model but compatible within uncertainties, and we should also notice that these predictions are at LO and it is known that NLL corrections increase the cross section of typically about 20%. In Fig. \[lhc2\], we also compare the $\xi$ distributions for jet production in exclusive events for the HPM and KMR models for jets with $p_T>$50 at the LHC. The $\xi$-slope is found to be smoother at the LHC for the KMR model than for the HPM. LHC data should thus allow to distinguish between both models or to tune better the parameters of the HPM given in the previous section. [lhc1]{} [lhc2]{} Conclusion ========== In this paper, we propose a new model to describe exclusive event production at hadronic colliders. It is based on Double Pomeron Exchange. We call it “Hybrid Pomeron Model" (HPM) since one of the color exhanges is considered to be hard, taking away most of the transverse momentum available while the colorless aspect of the overall crossed channel is ensured via a soft additional color exchange. The parameters of the model come from a fit to HERA $F_2$ data using a BFKL-based model for the hard part (eventually including saturation corrections), while the parameters of the soft part come from the usual soft cross section models. The model was successfully implemented in a generator (FPMC) to be able to compare directly with the CDF measurements performed at particle level. Our predictions are found to be in very good agreement with the measurements of the exclusive cross section as a function of the minimum jet transverse momentum or the dijet mass from the CDF collaboration. The HPM predicts a Higgs boson production cross section of about 1.1 fb at the LHC for a Higgs boson mass of 120 GeV. This is in the same range and compatible with the KMR determination. The $\xi$ distribution for exclusive events is softer for KMR than for the HPM model and it will be worth measuring it at the LHC and the Tevatron to distinguish and further tune both models. As we mentionned, the parameters used in the HPM come from an extrapolation from a fit to HERA data and it will be good to cross check the values of the parameters using direct data from the LHC. We thank Cyrille Marquet for useful remarks. One of us (M.S.R.) acknowledges support from CNPq (Brazil). [99]{} C. Royon \[RP220 Collaboration\], “Project to install roman pot detectors at 220 m in ATLAS,” arXiv:0706.1796 \[physics.ins-det\];\ M. G. Albrow [*et al.*]{} \[FP420 R&D Collaboration\], “The FP420 R&D Project: Higgs and New Physics with forward protons at the LHC,” arXiv:0806.0302 \[hep-ex\]. M G. Albrow and A. Rostovtsev, “Searching for the Higgs at hadron colliders using the missing mass method,” arXiv:hep-ph/0009336. V.A. Khoze, A.D. Martin, M.G. Ryskin, [*Eur. Phy. J*]{} [**C55**]{} (2008) 363; [*Phys. Lett.*]{} [**B650**]{} (2007) 41. CDF Coll., Phys. Rev. D [**77**]{} (2008) 052004. O. Kepka and C. Royon, Phys. Rev.  D [**76**]{}, 034012 (2007). J.D. Bjorken, [*Phys. Rev.*]{} [**D47**]{} (1993) 101;\ J.R. Cudell, O.F. Hernandez, [*Nucl. Phys.*]{} [**B471**]{} (1996) 471;\ E.M. Levin, hep-ph/9912403 and references therein;\ J. Pumplin, [*Phys. Rev.*]{} [**D52**]{} (1995) 1477;\ A. Berera and J.C. Collins, [*Nucl. Phys.*]{} [**474**]{} (1996) 183. R. Enberg, G. Ingelman, A. Kissavos, N. Timneanu, [*Phys. Rev. Lett.*]{} [ **89**]{} (2002) 081801. A. Bialas and P.V. Landshoff, [*Phys. Lett.*]{} [**B 256**]{} (1991) 540. M. Boonekamp, R. Peschanski, C. Royon, [*Phys. Rev. Lett.*]{} [**87** ]{} (2001) 251806;\ M. Boonekamp, A. De Roeck, R. Peschanski, C. Royon, [*Phys. Lett.*]{} [ **B550**]{} (2002) 93;\ M. Boonekamp, R. Peschanski, C. Royon, [*Nucl. Phys.*]{} [**B669**]{} (2003) 277, Err-ibid [**B676**]{} (2004) 493;\ for a general review see C. Royon, [*Mod. Phys. Lett.*]{} [**A18**]{} (2003) 2169. M. Boonekamp, R. Peschanski and C. Royon, Phys. Lett. B [**598**]{}, 243 (2004). V.A. Khoze, A.D. Martin, M.G. Ryskin, [*Eur. Phys. J.*]{} [**C19**]{} (2001) 477, Err-ibid [**C20**]{} (2001) 599;\ V.A. Khoze, A.D. Martin, M.G. Ryskin, [*Eur. Phys. J.*]{} [**C24**]{} (2002) 581. J. R. Forshaw, “Diffractive Higgs production: Theory,” arXiv:hep-ph/0508274. V.A. Khoze, A.D. Martin, M.G. Ryskin, W. J. Stirling, [*Eur.Phys.J.*]{} [**C35**]{} (2004) 211; M. Rangel, C. Royon, G. Alves, J. Barreto, R. Peschanski, [*Nucl. Phys.*]{} [**B774**]{} (2007) 53. V. S. Fadin, E. A. Kuraev and L. N. Lipatov, Phys. Lett.  B [**60**]{}, 50 (1975);\ E. A. Kuraev, L. N. Lipatov and V. S. Fadin, Sov. Phys. JETP [**45**]{}, 199 (1977) \[Zh. Eksp. Teor. Fiz.  [**72**]{}, 377 (1977)\];\ I. I. Balitsky and L. N. Lipatov, Sov. J. Nucl. Phys.  [**28**]{}, 822 (1978) \[Yad. Fiz.  [**28**]{}, 1597 (1978)\];\ L. N. Lipatov, Sov. Phys. JETP [**63**]{}, 904 (1986) \[Zh. Eksp. Teor. Fiz.  [**90**]{}, 1536 (1986)\].\ V. S. Fadin and L. N. Lipatov, Phys. Lett.  B [**429**]{}, 127 (1998);\ M. Ciafaloni, Phys. Lett.  B [**429**]{}, 363 (1998);\ M. Ciafaloni and G. Camici, Phys. Lett.  B [**430**]{}, 349 (1998). A. Bialas, R. B. Peschanski and C. Royon, Phys. Rev.  D [**57**]{}, 6899 (1998);\ H. Navelet, R. Peschanski, C. Royon, S. Wallon, Phys. Lett. B[**385**]{} (1996) 357. C. Marquet, Phys. Rev.  D [**76**]{}, 094017 (2007). P. V. Landshoff and O. Nachtmann, Z. Phys.  C[**35**]{}, 405 (1987) C. Marquet, R. B. Peschanski and G. Soyez, Phys. Rev.  D [**76**]{}, 034011 (2007). A. Kupco, C. Royon and R. Peschanski, Phys. Lett. B [**606**]{}, 139 (2005). J. D. Bjorken, [*Phys. Rev.* ]{} [**D47**]{}, (1993) 101; E. Gotsman, E. Levin and U. Maor, [*Phys. Lett.*]{} [**B438**]{} (1998), 229;\ A. B. Kaidalov, V. A. Khoze, A. D. Martin and M. G. Ryskin, [*Eur. Phys. J.*]{} [**C21**]{} (2001) 521;\ A. Bialas, Acta Phys. Polon. B [**33**]{}, 2635 (2002);\ A. Bialas and R. Peschanski, Phys. Lett. B [**575**]{}, 30 (2003). [A. Donnachie, P. V. Landshoff, [*Phys. Lett.* ]{}[**B207**]{} (1988) 319]{}. G. Soyez, Phys. Lett. B[**655**]{} (2007) 32. S. Munier and R. B. Peschanski, Phys. Rev. Lett.  [**91**]{}, 232001 (2003), Phys. Rev.  D [**69**]{}, 034008 (2004), Phys. Rev.  D [**70**]{}, 077503 (2004). E. Iancu, K. Itakura and S. Munier, Phys. Lett.  B [**590**]{}, 199 (2004). C. Adloff et al. \[H1 Collaboration\], [*Eur. Phys. J.*]{} [**C 21**]{} (2001) 33; J. Breitweg et al. \[ZEUS Collaboration\], [*Phys. Lett.*]{} [**B 487**]{} (2000) 273;\ S. Chekanov et al. \[ZEUS Collaboration\], [*Eur. Phys. J.*]{} [**C 21**]{} (2001) 443; C. Adloff et al. \[H1 Collaboration\], [*Phys. Lett.*]{} [**B 528**]{} (2002) 199;\ A. Aktas et al. \[H1 Collaboration\], [*Eur. Phys. J.*]{} [**C 45**]{} (2006) 23;\ S. Chekanov et al. \[ZEUS Collaboration\], [*Phys. Rev.*]{} [**D 69**]{} (2004) 012004. A. Bialas, W. Szeremeta, [*Phys. Lett.*]{} [**B296**]{} (1992) 191;\ W. Szeremeta,Acta Phys. Polon.  B [**24**]{}, 1159 (1993). M. Boonekamp, V. Juranek, O. Kepka, M. Rangel, C. Royon, in preparation; see [`ttp://cern.c`/project-fpmc/ ]{}. G. Marchesini [*et al.*]{}, Comp. Phys. Comm. [**67**]{}, 465 (1992). G. C. Blazey [*et al.*]{}, “Run II jet physics,” arXiv:hep-ex/0005012. [^1]: One notable exception is the Soft Color Interaction (SCI) model [@sci] where a colorful exchange is compensated by a phenomenological soft color interaction at long distance, which generates a gap in rapidity. We do not consider this model in the further discussion since it would need modifications to describe the CDF measurement of the dijet mass fraction [@Kepka:2007nr]. [^2]: At least partly, since the considered models have to correct for the rapidity gap survival probability, corresponding to the interaction between incident particles [@sp; @kupco]. [^3]: We have incorporated the Regge factors due to reggeization [@bialas1] in the definition of the propagators. [^4]: More refined wave function analyses are straightforward extensions of our formalism.

ArXiv

--- abstract: 'With the rapid progress of recent years, techniques that generate and manipulate multimedia content can now guarantee a very advanced level of realism. The boundary between real and synthetic media has become very thin. On the one hand, this opens the door to a series of exciting applications in different fields such as creative arts, advertising, film production, video games. On the other hand, it poses enormous security threats. Software packages freely available on the web allow any individual, without special skills, to create very realistic fake images and videos. So-called deepfakes can be used to manipulate public opinion during elections, commit fraud, discredit or blackmail people. Potential abuses are limited only by human imagination. Therefore, there is an urgent need for automated tools capable of detecting false multimedia content and avoiding the spread of dangerous false information. This review paper aims to present an analysis of the methods for visual media integrity verification, that is, the detection of manipulated images and videos. Special emphasis will be placed on the emerging phenomenon of deepfakes and, from the point of view of the forensic analyst, on modern data-driven forensic methods. The analysis will help to highlight the limits of current forensic tools, the most relevant issues, the upcoming challenges, and suggest future directions for research.' author: - Luisa Verdoliva bibliography: - 'refs.bib' title: | Media Forensics and DeepFakes:\ an overview --- Digital image forensics, video forensics, deep learning, deepfakes. Conclusion ========== Fifteen years ago multimedia forensics was a niche field of practical interest only for a restricted set of players involved in law enforcement, intelligence, private investigations. Both attacks and defences had an artisan flavor, and required painstaking work and dedication. Artificial intelligence has largely changed these rules. High-quality fakes now seem to come out from an assembly line calling for an extraordinary effort on part of both scientists and policymakers. In fact, today’s multimedia forensics is in full development, major agencies are funding large research initiatives, and scientists form many different field are contributing actively, with fast advances in ideas and tools. It is difficult to forecast whether such efforts will be able to ensure information integrity in the future, or some forms of active protection will become necessary. This is an arms race, and one part is no smarter than the other. For the present time, a large arsenal of tools is being developed, and knowing them, the principles on which they rely, and their scope of application is a prerequisite to protect institutions and ordinary people. Acknowledgement =============== We gratefully acknowledge the support of this research by a Google Faculty Award. In addition, this material is based on research sponsored by the Air Force Research Laboratory and the Defense Advanced Research Projects Agency under agreement number FA8750-16-2-0204. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright notation thereon. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the Air Force Research Laboratory and the Defense Advanced Research Projects Agency or the U.S. Government.

ArXiv

--- abstract: 'We measure the angular power spectrum $C_\ell$ of radio galaxies in the NRAO VLA Sky Survey (NVSS) using two independent methods: direct spherical harmonic analysis and maximum likelihood estimation. The results are consistent and can be understood using models for the spatial matter power spectrum and for the redshift distribution of radio galaxies at mJy flux-density levels. A good fit to the angular power spectrum can only be achieved if radio galaxies possess high bias with respect to mass fluctuations; by marginalizing over the other parameters of the model we derive a $68\%$ confidence interval $1.53 < b_0 \times \sigma_8 < 1.87$, where $b_0$ is the linear bias factor for radio galaxies and $\sigma_8$ describes the normalization of the matter power spectrum. Our models indicate that the majority of the signal in the NVSS $C_\ell$ spectrum is generated at low redshifts $z \la 0.1$. Individual redshifts for the NVSS sources are thus required to alleviate projection effects and probe directly the matter power spectrum on large scales.' author: - | Chris Blake$^{\,1,2,}$, Pedro G. Ferreira$^{\,2}$ and Julian Borrill$^{\,3}$\ \ $^1$ School of Physics, University of New South Wales, Sydney, NSW 2052, Australia\ $^2$ Astrophysics, University of Oxford, Keble Road, Oxford, OX1 3RH, UK\ $^3$ National Energy Research Scientific Computing Centre, LBNL, Berkeley, CA, USA title: The angular power spectrum of NVSS radio galaxies --- large-scale structure of Universe – galaxies: active – surveys Introduction ============ Active Galactic Nuclei (AGN) mapped in radio waves are an interesting probe of large-scale structure. They can be routinely detected out to very large redshift ($z \sim 4$) over wide areas of the sky and hence delineate the largest structures and their evolution over cosmic epoch. Radio emission is insensitive to dust obscuration and radio AGN are effective tracers of mass: they are uniformly hosted by massive elliptical galaxies and have been shown to trace both clusters (Hill & Lilly 1991) and superclusters (Brand et al. 2003). The current generation of wide-area radio surveys such as Faint Images of the Radio Sky at Twenty centimetres (FIRST; Becker, White & Helfand 1995) and the NRAO VLA Sky Survey (NVSS; Condon et al. 1998) contain radio galaxies in very large numbers ($\sim 10^6$) and have allowed accurate measurements of the imprint of radio galaxy angular clustering. These patterns are considerably harder to detect in radio waves than in optical light due to the huge redshift range that is probed. Whilst this provides access to clustering on the largest scales, it also washes out much of the angular clustering signal through the superposition of unrelated redshift slices. The angular correlation function was measured for FIRST by Cress et al. (1996) and Magliocchetti et al. (1998) and for NVSS by Blake & Wall (2002a) and Overzier et al. (2003). The NVSS radio survey, covering $\sim 80$ per cent of the sky, permits the measurement of fluctuations over very large angles. Blake & Wall (2002b) detected the imprint of the cosmological velocity dipole in the NVSS surface density, in a direction consistent with the Cosmic Microwave Background (CMB) dipole. In this study we measure the angular power spectrum, $C_\ell$, of the radio galaxy distribution (Baleisis et al. 1998). This statistic represents the source surface-density field as a sum of sinusoidal angular density fluctuations of different wavelengths, using the spherical harmonic functions. The angular power spectrum is sensitive to large-angle fluctuations and hence complements the measurement of the angular correlation function, $w(\theta)$, at small angles. Measurement of the $C_\ell$ spectrum has some advantages in comparison with $w(\theta)$. Firstly, the error matrix describing correlations between multipoles $\ell$ has a very simple structure, becoming diagonal for a complete sky. This is not the case for the separation bins in a measurement of $w(\theta)$: even for a full sky, an individual galaxy appears in many separation bins, automatically inducing correlations between those bins. Secondly, there is a natural relation between the angular power spectrum and the spatial power spectrum of density fluctuations, $P(k)$. This latter quantity provides a very convenient means of describing structure in the Universe for a number of reasons. Firstly its primordial form is produced by models of inflation, which prescribe the initial pattern of density fluctuations $\delta\rho/\rho$. Furthermore, in linear theory for the growth of perturbations, fluctuations described by different wavenumbers $k$ evolve independently, enabling the model power spectrum to be easily scaled with redshift. The physics of linear perturbations are hence more naturally described in Fourier space. In contrast, the angular correlation function is more easily related to the spatial correlation function $\xi(r)$, the Fourier transform of $P(k)$. Correlation functions more naturally serve to describe the real-space profile of collapsing structures evolving out of the linear regime. We emphasize that although the two functions $C_\ell$ and $w(\theta)$ are [*theoretically*]{} equivalent – linked by a Legendre transform – this is not true in an [*observational*]{} sense. For example, $w(\theta)$ can only be successfully measured for angles up to a few degrees, but $C_\ell$ depends on $w(\theta)$ at [*all*]{} angles. We derive the angular power spectrum using two independent methods. Firstly we apply a direct spherical harmonic estimator following Peebles (1973). Secondly, we use maximum likelihood estimation, commonly employed for deriving the angular power spectra of the CMB temperature and polarization maps. These two methods are described in Section \[secmeth\]. We find that these two approaches yield very similar results (Section \[secres\]), which is unsurprising given the wide sky coverage of the NVSS. In Section \[secpk\] we interpret the NVSS angular power spectrum in terms of the underlying spatial power spectrum of mass fluctuations and the radial distribution of radio sources. Finally in Section \[secbias\] we employ these models to derive the linear bias factor of NVSS radio galaxies by marginalizing over the other model parameters. The NVSS radio survey {#secobs} ===================== The 1.4 GHz NRAO VLA Sky Survey (NVSS; Condon et al. 1998) was performed at the Very Large Array over the period 1993 to 1996 and covers the sky north of declination $-40^\circ$. The source catalogue contains $\approx 1.8 \times 10^6$ entries and is 99 per cent complete at integrated flux density $S_{\rm 1.4 \, GHz} = 3.5$ mJy. The full width at half-maximum of the synthesized beam is 45 arcsec; the majority of radio sources are thus unresolved. The relatively broad NVSS beam yields excellent surface brightness sensitivity and photometric completeness. Before analyzing the survey for large-scale structure we imposed various angular masks. Firstly we excluded catalogue entries within $5^\circ$ of the Galactic plane, many of which are Galactic in origin (mostly supernova remnants and HII regions). The contribution of foreground Galactic sources at latitudes $|b| > 5^\circ$ is negligible. We also placed 22 masks around bright local extended radio galaxies contributing large numbers of catalogue entries, as described in Blake & Wall (2002a). Figure \[fignvss\] plots the remaining sources with $S_{\rm 1.4 \, GHz} > 200$ mJy. The remaining NVSS geometry corresponds to 75 per cent of the celestial sphere. Blake & Wall (2002a) demonstrated that the NVSS suffers from systematic gradients in surface density at flux-density levels at which it is complete. These gradients – corresponding to a $\sim 5$ per cent variation in surface density at a threshold of 3 mJy – are entirely unimportant for the vast majority of applications of this catalogue. However, they have a significant influence on the faint imprint of large-scale structure. If left uncorrected a distortion of the measured angular power spectrum would result, because the harmonic coefficients would need to reproduce the systematic gradients as well as the fluctuations due to clustering. Blake & Wall (2002a) found that these surface gradients are only significant at fluxes $S_{\rm 1.4 \, GHz} < 10$ mJy. At brighter fluxes the survey is uniform to better than 1 per cent – sufficient to allow the detection of the anticipated cosmological velocity dipole (Blake & Wall 2002b). Hence we simply restricted our $C_\ell$ analysis to fluxes $S_{\rm 1.4 \, GHz} > 10$ mJy. The NVSS source surface density at this threshold is $\sigma_0 = 16.9$ deg$^{-2}$. We note that the broadness of the radio luminosity function ensures that the projected clustering properties of radio galaxies are not a strong function of flux density in the range 3 mJy $< S_{\rm 1.4 \, GHz} <$ 50 mJy, as verified by the correlation function analyses of Blake & Wall (2002a) and Overzier et al. (2003). In this flux-density range, the redshift distribution of the radio galaxies does not change significantly. Estimating the angular power spectrum: methods {#secmeth} ============================================== Definition of the angular power spectrum ---------------------------------------- A distribution of galaxies on the sky can be generated in two statistical steps. Firstly, a density field $\sigma(\theta,\phi)$ is created; this may be described in terms of its spherical harmonic coefficients $a_{\ell m}$: $$\sigma(\theta,\phi) = \sum_{\ell=0}^{\infty} \sum_{m=-\ell}^{+\ell} \, a_{\ell m} \, Y_{\ell m}(\theta,\phi) \label{eqsigylm}$$ where $Y_{\ell m}$ are the usual spherical harmonic functions. Secondly, galaxy positions are generated in a Poisson process as a (possibly biased) realization of this density field. The angular power spectrum $C_\ell$ prescribes the spherical harmonic coefficients in the first step of this model. It is defined over many realizations of the density field by $$<|a_{\ell m}|^2> = C_\ell \label{eqcldef}$$ The assumption of isotropy ensures that $<|a_{\ell m}|^2>$ is a function of only $\ell$, not $m$. The angular power spectrum $C_\ell$ is theoretically equivalent to the angular correlation function $w(\theta)$ as a description of the galaxy distribution. The two quantities are connected by the well-known relation $$C_\ell = 2 \pi \, \sigma_0^2 \int_{-1}^{+1} \, w(\theta) \, P_\ell(\cos{\theta}) \, d(\cos{\theta}) \label{eqwtocl}$$ where $\sigma_0$ is the source surface density and $P_\ell$ is the Legendre polynomial. However, the angular scales on which the signal-to-noise is highest are very different for each statistic. $w(\theta)$ can only be measured accurately at small angles up to a few degrees (Blake & Wall 2002a), beyond which Poisson noise dominates. By contrast, $C_\ell$ for galaxies has highest signal-to-noise at small $\ell$, corresponding to large angular scales $\theta \sim 180^\circ/\ell$. Hence the two statistics are complementary, the $C_\ell$ spectrum probing fluctuations on the largest angular scales. Spherical harmonic estimation of $C_\ell$ {#secestharm} ----------------------------------------- Peebles (1973) presented the formalism of spherical harmonic analysis of a galaxy distribution over an incomplete sky (for refinements see e.g. Wright et al. 1994; Wandelt, Hivon & Gorski 2000). For a partial sky, a spherical harmonic analysis is hindered by the fact that the spherical harmonics are not an orthonormal basis, which causes the measured coefficients $a_{\ell m}$ to be statistically correlated, entangling different multipoles of the underlying $C_\ell$ spectrum. However, for the case of a survey covering $\sim 80\%$ of the sky, the repercussions (discussed below) are fairly negligible, implying shifts and correlations in the derived power spectrum that are far smaller than the error bars. We employed the original method of Peebles (1973) with only one small correction for sample variance. In Section \[secmaxlik\] we compare spherical harmonic analysis with the technique of maximum likelihood estimation and find that the two methods yield results in good agreement. The spherical harmonic coefficients of the density field may be estimated by summing over the $N$ galaxy positions $(\theta_i,\phi_i)$: $$A_{\ell m} = \sum_{i=1}^{N} \, Y_{\ell m}^*(\theta_i,\phi_i) \label{eqalm}$$ For an incomplete sky, these values need to be corrected for the unsurveyed regions, so that an estimate of $C_\ell$ is $$C_{\ell m}^{\rm obs} = \frac{|A_{\ell m} - \sigma_0 \, I_{\ell m}|^2}{J_{\ell m}} - \sigma_0 \label{eqclpeeb}$$ (Peebles 1973 equation 50) where $\sigma_0 = N/\Delta\Omega$ and $$I_{\ell m} = \int_{\Delta \Omega} \, Y_{\ell m}^* \, d\Omega$$ $$J_{\ell m} = \int_{\Delta \Omega} \, |Y_{\ell m}|^2 \, d\Omega \label{eqjlm}$$ where the integrals are over the survey area $\Delta \Omega$, and are determined in our analysis by numerical integration. The final term in equation \[eqclpeeb\] corrects for the finite number of discrete sources: for a full sky ($I_{\ell m} = 0, J_{\ell m} = 1$) we expect $<|A_{\ell m}|^2> = \sigma_0$ in the absence of clustering, i.e. this is the power spectrum of the shot noise. Note that $\sigma_0$ is the apparent source density $N/\Delta \Omega$, not the average over an imagined ensemble of catalogues. We determined the angular power spectrum for the $\ell$th multipole, $C_\ell^{\rm obs}$, by averaging equation \[eqclpeeb\] over $m$. Because the density field is real rather than complex, $C_{\ell m}^{\rm obs} = C_{\ell,-m}^{\rm obs}$, resulting in $\ell+1$ independent measurements of $C_\ell$: $$C_\ell^{\rm obs} = \frac{\sum_{m=0}^\ell \, C_{\ell m}^{\rm obs}}{\ell+1} \label{eqclobs}$$ There is no need for us to use the modified weighting formula of Peebles (1973 equation 53). In our case, $J_{\ell m}$ does not vary significantly with $m$. We verified that the modified weighting formula produced indistinguishable results. One consequence of the partial sky is to “mix” the harmonic coefficients such that the measured angular power spectrum at $\ell$ depends on a range of $C_{\ell'}$ around $\ell' = \ell$: $$<C_\ell^{\rm obs}> = \sum_{\ell'} \, C_{\ell'} \, R_{\ell\ell'}$$ The angled brackets refer to an imagined averaging over many realizations of density fields generated by $C_\ell$, in accordance with equation \[eqcldef\]. Peebles showed that $\sum_{\ell'} R_{\ell\ell'} = 1$; i.e. mixing does not spuriously enhance the measured power (this is accomplished by the factor $J_{\ell m}$ in equation \[eqclpeeb\]). For a complete sky, $R_{\ell\ell'} = \delta_{\ell\ell'}$, where $\delta_{mn} = 1$ ($m = n$) or 0 ($m \ne n$). For a partial sky, the matrix $R_{\ell\ell'}$ can be computed from the geometry of the surveyed region (Hauser & Peebles 1973). Figure \[figrll\] illustrates the result for the NVSS for $\ell = 10$ (computed using Hauser & Peebles 1973 equation 12). The NVSS covers a sufficiently large fraction of the sky (75 per cent) that mixing only occurs at the $\sim 15$ per cent level and can be neglected because the underlying $C_\ell$ spectrum is smooth: $$<C_\ell^{\rm obs}> \approx C_\ell$$ We checked the NVSS $R_{\ell\ell'}$ matrix for other multipoles $\ell$ and found very similar results. This argument ensures that the measured multipoles are statistically independent to a good approximation. The statistical error on the estimator of equation \[eqclpeeb\] is $$\sigma(C_{\ell m}^{\rm obs}) \approx (\sigma_0 + C_\ell) \sqrt{1 + \delta_{m0}} \label{eqclsig}$$ (Peebles 1973 equation 81). There are two components of the error: - “Shot noise” ($\sigma_0$) because the number of discrete objects is finite and therefore does not perfectly describe the underlying density field. - “Cosmic variance” ($C_\ell$) because even with perfect sampling of the density field, there are only a finite number of harmonics associated with the $\ell$th multipole. The error for the $m = 0$ case in equation \[eqclsig\] is greater because $A_{\ell 0}$ is purely real, rather than complex. In the latter case, we are averaging over the real and imaginary parts of $A_{\ell m}$, two independent estimates of $C_\ell$, which reduces the overall statistical error by a factor $\sqrt{2}$. For a partial sky equation \[eqclsig\] is an approximation, because the variance of multipoles of given $\ell$ depends on the underlying power spectrum at $\ell' \ne \ell$. As discussed above, this effect is negligible for the NVSS. The averaging over $m$ (equation \[eqclobs\]) decreases the error in the observation. Combining the errors of equation \[eqclsig\], assuming estimates at different $m$ are statistically independent, the resulting error in $C_\ell^{\rm obs}$ is $$\sigma(C_\ell^{\rm obs}) \approx (\sigma_0 + C_\ell) \frac{\sqrt{\ell+2}}{\ell+1} \label{eqclerr}$$ We used Monte Carlo simulations to verify that equation \[eqclerr\] produced results within 5 per cent of the true error for all relevant multipoles. Our only addition to the formalism of Peebles (1973) was to increase the total variance on the estimate of $C_\ell$ by a factor $1/f_{\rm sky}$, where $f_{\rm sky} = \Delta \Omega/4\pi$ is the fraction of sky covered (i.e. multiply equation \[eqclerr\] by $1/\sqrt{f_{\rm sky}}$). This correction factor was motivated by Scott, Srednicki and White (1994) as a fundamental property of sample variance for a partial sky, and is part of the standard CMB formalism (e.g. Bond, Efstathiou & Tegmark 1997). For the NVSS geometry, $f_{\rm sky} = 0.75$, thus this correction corresponds to a $\sim 10$ per cent increase in the error. Correction for multiple-component sources {#secmult} ----------------------------------------- Radio sources have complex morphologies and large linear sizes (up to and exceeding 1 Mpc). A radio-source catalogue such as the NVSS will contain entries which are different components of the same galaxy (for example, the two radio lobes of a “classical double” radio galaxy). The broad angular resolution of the NVSS beam leaves over 90 per cent of radio sources unresolved; however, the remaining multiple-component sources have a small but measurable effect on the angular power spectrum. It is relatively simple to model the effect of multiple-component sources on the estimator for $C_\ell$ described in Section \[secestharm\]. The relevant angular scales ($\ell < 100$) are much bigger than any component separation, and equation \[eqalm\] can be replaced by $$A_{\ell m} = \sum_{i=1}^{N_{\rm gal}} c_i Y_{\ell m}^*(i)$$ where $N_{\rm gal}$ is the total number of galaxies and $c_i$ is the number of components of the $i$th galaxy. Thus the quantity $$<A_{\ell m}> = N_{\rm gal} \, \overline{c} <Y_{\ell m}> = N <Y_{\ell m}>$$ is unchanged by the presence of multiple components ($\overline{c}$ denotes the average number of components per galaxy, and $N = \overline{c} \times N_{\rm gal}$ is the total number of catalogue entries, as in equation \[eqalm\]). But $C_{\ell m}^{\rm obs}$ in equation \[eqclpeeb\] depends on $$\begin{aligned} <|A_{\ell m}|^2> \hspace{-3mm} &=& \hspace{-3mm} N_{\rm gal} \, \overline{c^2} <|Y_{\ell m}|^2> + N_{\rm gal}^2 \, \overline{c}^2 <Y_{\ell m}> <Y_{\ell m}^*> \nonumber \\ &=& \hspace{-3mm} \frac{\overline{c^2}}{\overline{c}} \, N <|Y_{\ell m}|^2> \, + \, N^2 <Y_{\ell m}> <Y_{\ell m}^*> \nonumber\end{aligned}$$ Multiple-component sources only affect the first term in this expression, producing an offset in the $C_\ell$ spectrum $$\Delta C_\ell = \frac{\Delta (<|A_{\ell m}|^2>)}{J_{\ell m}} = \left( \frac{\overline{c^2}}{\overline{c}} - 1 \right) \frac{N <|Y_{\ell m}|^2>}{J_{\ell m}}$$ But $J_{\ell m} = <|Y_{\ell m}|^2> \Delta \Omega$ from equation \[eqjlm\], and this expression simplifies to an offset independent of $\ell$: $$\Delta C_\ell = \left( \frac{\overline{c^2}}{\overline{c}} - 1 \right) \sigma_0$$ Most multiple-component sources in the NVSS catalogue are double radio sources. Let a fraction $e \ll 1$ of the radio galaxies be doubles. Then $\overline{c} = 1 + e$ and $\overline{c^2} = 1 + 3e$, thus the constant offset may be written $$\Delta C_\ell \approx 2 \, e \, \sigma_0 \label{eqcldoub}$$ We can deduce $e = 0.07 \pm 0.005$ from the form of the NVSS angular correlation function $w(\theta)$ at small angles $\theta < 0.1^\circ$, where double sources dominate the close pairs (see Blake & Wall 2002a and also Section \[secalm\]). This correction was applied to the measured NVSS $C_\ell$ spectrum and successfully removed the small systematic offset in $C_\ell$ at high $\ell$. Maximum likelihood estimation of $C_\ell$ {#secmaxlik} ----------------------------------------- A sophisticated suite of analytical tools has been developed by the CMB community for deriving the angular power spectra of the observed CMB temperature and polarization maps. These methods can also be exploited to analyze galaxy data (see for example Efstathiou & Moody 2001, Huterer, Knox & Nichol 2001 and Tegmark et al. 2002). In this approach the power spectrum is determined using an iterative maximum likelihood analysis, in contrast to the direct estimator discussed in Section \[secestharm\]. The likelihood is a fundamental statistical quantity, and this analysis method permits straightforward control of such issues as edge effects, noise correlations and systematic errors. The starting point for maximum likelihood estimation (MLE) is Bayes’ theorem $$P(\alpha|DI) \propto P(\alpha|I) \, P(D|\alpha I)$$ where $\alpha$ are the parameters one is trying determine, $D$ is the data and $I$ is the additional information describing the problem. The quantity $P(\alpha|I)$ is the likelihood, i.e. the probability of the data given a specific set of parameters, while the left-hand side is the posterior, i.e. the probability of the parameters given the data. We will assume that the sky is a realization of a stationary Gaussian process, with an angular power spectrum $C_\ell$. We assume no cosmological information about the distribution of the $C_\ell$. The rendition of the sky will be a pixelized map, created by binning the galaxy data in equal-area cells such that the count in the $i$th cell is $n_i$, effectively constructing a “temperature map” of galaxy surface density. We performed this task using the HEALPIX software package (Gorksi, Hivon & Wandelt 1999; [http://www.eso.org/science/healpix]{}). We chose the HEALPIX pixelization scheme $n_{\rm side} = 32$, which corresponds to 12,288 pixels over a full sky. The angular power spectrum may be safely extracted to multipole $\ell_{\rm max} \approx 2 \times n_{\rm side}$. We then defined a data vector: $$x_i = \frac{n_i}{\overline{n}} - 1 \label{eqdata}$$ where $\overline{n}$ is the mean count per pixel. Figure \[fighist\] demonstrates that the data vector $x_i$ for the NVSS sample is well-approximated by a Gaussian distribution, as assumed in a maximum likelihood analysis. The covariance matrix $C^T_{ij}$ due to primordial fluctuations is given by $$\langle x_i x_j \rangle \, C^T_{ij} = \sum_\ell \frac{(2\ell+1)}{4\pi} \, C_\ell \, P_\ell(\cos\theta_{ij})$$ where $P_\ell$ is the Legendre polynomial and $\theta_{ij}$ is the angle between pixel pair $(i,j)$. In order to apply a likelihood analysis we must also specify a noise covariance matrix $C^N_{ij}$. We modelled the noise as a Gaussian random process with variance $1/\overline{n}$, uncorrelated between pixels, such that $C^N_{ij} = (1/\overline{n}) \, \delta_{ij}$. The likelihood of the map, with a particular power spectrum $C_\ell$, is given by $$\ln P(C_{\ell}|{\bf x})\propto-\frac{1}{2}({\bf x}^T(C^T+C^N)^{-1}{\bf x}) +Tr[\ln(C^T+C^N)]$$ The goal of MLE is to maximize this function, and the fastest general method is to use Newton-Raphson iteration to find the zeroes of the derivatives in $\ln P(C_{\ell}|{\bf x})$ with respect to $C_\ell$. We used the MADCAP package (Borrill 1999; [http://www.nersc.gov/$\sim$borrill/cmb/madcap]{}) to derive the maximum likelihood banded angular power spectrum from the pixelized galaxy map and noise matrix. MADCAP is a parallel implementation of the Bond, Jaffe & Knox (1998) maximum-likelihood algorithms for the analysis of CMB datasets. We ran the analysis software on the supercomputer Seaborg, administered by the National Energy Research Scientific Computing Centre (NERSC) at Lawrence Berkeley National Laboratory, California. We again applied equation \[eqcldoub\] to the MADCAP results to correct the measured power spectrum for the influence of multiple-component sources. Boughn & Crittenden (2002) also performed a HEALPIX analysis of the NVSS as part of a cross-correlation analysis with the CMB searching for evidence of the Integrated Sachs-Wolfe effect. From the pixelized map they derived an angular correlation function for the NVSS, which they used to constrain theoretical models. We compare their results with ours in Section \[secbias\]. Testing the methods {#sectest} ------------------- We tested the two methods, direct spherical harmonic and maximum likelihood estimation, by generating a dipole distribution of $N \approx 10^5$ sources over the NVSS geometry, where the model dipole possessed the same amplitude and direction as that detected in the NVSS (see Blake & Wall 2002b). In order to simulate multiple components, we added companion sources to $\approx 7$ per cent of the objects, with the same separation distribution as that measured in the NVSS (Blake & Wall 2002a). The two measurements of the $C_\ell$ spectrum, corrected for multiple components using equation \[eqcldoub\], are plotted in Figure \[figcltest\] and are consistent with zero. The $C_\ell$ measurements have been averaged into bands of width $\Delta \ell = 5$, starting from $\ell = 2$. The dipole term $\ell = 1$ is of course spuriously high, but has a negligible effect on the measured harmonics at $\ell > 1$ – the galaxy dipole (unlike the CMB dipole) is only barely detectable in current surveys (Blake & Wall 2002b). The normalization convention used in Figure \[figcltest\], and the remaining power spectrum plots, is to expand the surface overdensity $\delta = (\sigma - \sigma_0)/\sigma_0$ in terms of spherical harmonics. To convert from the definition of $C_\ell$ in equations \[eqsigylm\] and \[eqcldef\] we simply divide $A_{\ell m}$ by $\sigma_0$ (in units of sr$^{-1}$) and hence $C_\ell$ by $\sigma_0^2$. Figure \[figcltest\] permits a first comparison of the two independent methods of estimating $C_\ell$. At low $\ell$, the variances are in excellent agreement. As $\ell$ approaches $2 \times n_{\rm side}$, the variance of the maximum likelihood method begins to exceed that of the spherical harmonic analysis. This occurs as the resolution of the pixelization scheme becomes important: the angular pixel size is no longer much less than the characteristic angular scale probed by the $\ell$th multipole. The variance of the high $\ell$ bins could be reduced by adopting a finer pixelization, such as $n_{\rm side} = 64$, with the penalty of a rapidly increasing requirement of supercomputer time. Given that the signal in the NVSS angular power spectrum turns out to be confined to $\ell \la 40$, our optimum pixelization remains $n_{\rm side} = 32$. Moreover, higher pixel resolution would decrease the mean pixel count $\overline{n}$ in equation \[eqdata\], rendering a Gaussian distribution a poorer approximation for $x_i$. Results {#secres} ======= The angular power spectrum $C_\ell$ {#secmeascl} ----------------------------------- Figure \[figcl\] plots the NVSS angular power spectrum measured for flux-density threshold $S_{\rm 1.4 \, GHz} = 10$ mJy using both spherical harmonic analysis and maximum likelihood estimation. The constant offset due to double sources (equation \[eqcldoub\]) has been subtracted and the measurements are averaged into bands of width $\Delta \ell = 5$, starting from $\ell_{\rm min} = 2$. Measurements are plotted up to $\ell = 100$, although note that the variance of the maximum likelihood estimation is increased by pixelization effects above $\ell \sim 50$, as discussed in Section \[sectest\]. Table \[tabcl\] lists the plotted data. Figure \[figclcmb\] displays the same data scaled by the usual CMB normalization factor, $\ell(\ell+1)/2\pi$. ------------------ ------------------ -------------------------------- -------------------------------- $\ell_{\rm min}$ $\ell_{\rm max}$ Spherical harmonic Maximum likelihood $C_\ell^{\rm obs} \times 10^5$ $C_\ell^{\rm obs} \times 10^5$ 2 6 $2.98 \pm 1.93$ $3.35 \pm 1.62$ 7 11 $1.46 \pm 0.64$ $1.19 \pm 0.64$ 12 16 $0.85 \pm 0.44$ $0.75 \pm 0.45$ 17 21 $0.68 \pm 0.33$ $0.71 \pm 0.38$ 22 26 $0.79 \pm 0.30$ $0.41 \pm 0.31$ 27 31 $0.33 \pm 0.23$ $0.54 \pm 0.29$ 32 36 $0.34 \pm 0.21$ $0.49 \pm 0.27$ 37 41 $0.12 \pm 0.18$ $0.32 \pm 0.25$ 42 46 $0.18 \pm 0.17$ $-0.07 \pm 0.21$ 47 51 $0.30 \pm 0.17$ $0.22 \pm 0.22$ 52 56 $0.21 \pm 0.16$ $0.22 \pm 0.22$ 57 61 $0.26 \pm 0.16$ $0.31 \pm 0.23$ 62 66 $0.05 \pm 0.14$ $0.21 \pm 0.22$ 67 71 $0.06 \pm 0.13$ $-0.06 \pm 0.21$ 72 76 $0.25 \pm 0.14$ $0.35 \pm 0.24$ 77 81 $0.13 \pm 0.13$ $0.17 \pm 0.24$ 82 86 $-0.03 \pm 0.11$ $0.01 \pm 0.24$ 87 91 $-0.05 \pm 0.11$ $-0.15 \pm 0.24$ 92 96 $0.08 \pm 0.11$ $0.09 \pm 0.27$ 97 101 $0.12 \pm 0.11$ $0.11 \pm 0.27$ ------------------ ------------------ -------------------------------- -------------------------------- : Data table of banded NVSS $C_\ell$ values plotted in Figure \[figcl\]. The offset due to double sources is $\Delta C_\ell = 0.24 \times 10^{-5}$.[]{data-label="tabcl"} We detect clear signal in the NVSS $C_\ell$ spectrum at multipoles $\ell \la 40$, well-fitted by a power-law $C_\ell = a \, \ell^{-b}$, where the best-fitting parameters are $a = (2.0 \pm 0.4) \times 10^{-4}$, $b = 1.20 \pm 0.06$ (fitted to the spherical harmonic analysis result) or $a = (1.6 \pm 0.4) \times 10^{-4}$, $b = 1.12 \pm 0.08$ (fitted to the maximum likelihood method result). The amplitude of the angular power spectrum for radio galaxies is two orders of magnitude smaller than that found for optically-selected galaxies (e.g. Huterer, Knox & Nichol 2001); this is readily explained by the wide redshift range of the NVSS sources, which vastly dilutes the clustering signal through the superposition of unrelated redshift slices. The NVSS signal remains $\sim 5$ orders of magnitude greater than the CMB $C_\ell$ spectrum over the same multipole range, reflecting the growth of structure since $z = 1100$. Given the incomplete sky and finite resolution, the measured $C_\ell$ values are not independent. However, it was argued in Section \[secestharm\] that the correlations between neighbouring power spectrum measurements are small. This fact was confirmed by the maximum likelihood analysis. The degree of correlation between neighbouring bins is given by the immediately off-diagonal elements of the inverse Fisher matrix. This data is generated by the MADCAP software, and inspection revealed that the size of the immediately off-diagonal matrix elements was $\sim 20$ times smaller than that of the diagonal elements. Figure \[figclflux\] compares the angular power spectra measured at flux-density thresholds 5 mJy, 10 mJy and 20 mJy using spherical harmonic analysis. The 5 mJy data may be affected by systematic surface density gradients. The results are consistent with an unchanging underlying power spectrum. This is not surprising; the redshift distribution of radio sources does not vary significantly between 5 mJy and 20 mJy, and the angular correlation function has been found not to depend on flux density in this range (Blake & Wall 2002a). Probability distribution of $|A_{\ell m}|^2$ {#secalm} -------------------------------------------- An interesting probe of the galaxy pattern is the distribution of values of $|A_{\ell m}|^2$ (see Hauser & Peebles 1973). These quantities are measured as part of our spherical harmonic analysis (Section \[secestharm\]). For a random distribution with surface density $\sigma_0$ over a full sky, the central limit theorem ensures that the real and imaginary parts of $A_{\ell m} = \sum_i Y_{\ell m}^*(i)$ are drawn independently from Gaussian distributions such that the normalization satisfies $|A_{\ell m}|^2 = \sigma_0$. It is then easy to show that $x = |A_{\ell m}|^2$ has an exponential probability distribution for $m \ne 0$: $$P(x) \, dx = \frac{\exp{(-x/\sigma_0)}}{\sigma_0} \, dx \label{eqalmdist}$$ For a partial sky, $|A_{\ell m}|^2$ is replaced by $|A_{\ell m} - \sigma_0 I_{\ell m}|^2/J_{\ell m}$ (equation \[eqclpeeb\]). Figure \[figalmdist\] plots the distribution of observed values of $|A_{\ell m} - \sigma_0 I_{\ell m}|^2/J_{\ell m}$. We restrict this plot to the multipole range $51 < \ell < 100$: for this range of $\ell$, Figure \[figcl\] demonstrates that $C_\ell \approx 0$ and thus the survey is well-described by a random distribution with additional multiple components. For each $\ell$ we included the range $1 \leq m \leq \ell$ (negative values of $m$ are not independent). Overplotted on Figure \[figalmdist\] as the solid line is the prediction of equation \[eqalmdist\]. Multiple components cause the slope of the observed exponential distribution to be shallower than this prediction. Section \[secmult\] shows that the value of $<|A_{\ell m} - \sigma_0 I_{\ell m}|^2/J_{\ell m}>$ is increased from $\sigma_0$ to $(1+2e)\sigma_0$, where $e$ is the fraction of galaxies split into double sources. Thus equation \[eqalmdist\] must be amended such that $P(x) \propto \exp{[-x/(1+2e)\sigma_0]}$. Assuming that $e = 0.07$, this corrected prediction is plotted on Figure \[figalmdist\] as the dashed line and provides a very good fit to the observed distribution. This is an independent demonstration that approximately 7 per cent of NVSS galaxies are split into multiple-component sources. Figure \[figalmdist\] also underlines the fact that the imprint of clustering on the projected radio sky is very faint. Comparison with $w(\theta)$ --------------------------- The angular correlation function $w(\theta)$ has been measured for the NVSS by Blake & Wall (2002a) and Overzier et al. (2003). It is well-described by a power-law $w(\theta) \approx (1 \times 10^{-3}) \, \theta^{-0.8}$ for angles up to a few degrees. Equation \[eqwtocl\] allows us to derive the equivalent $C_\ell$ spectrum if we assume that this power-law extends to all angular scales. In Figure \[figwth\] we overplot the resulting prediction on the measurements and find an excellent fit. This is initially surprising: the angular correlation function is only measurable at small angles $\theta < 10^\circ$, whereas low multipoles $\ell$ describe surface fluctuations on rather larger angular scales. However, in Section \[secpk\] we establish that the signal in low multipoles is actually generated at low redshift by spatial fluctuations on relatively small scales, similar to the spatial scales which produce the signal in $w(\theta)$. Of course the comparison is not entirely straight-forward because the two statistics quantify different properties of the galaxy distribution. The value of $C_\ell$ quantifies the amplitude of fluctuations on the angular scale corresponding to $\ell$. The value of $w(\theta)$ is the average of the product of the galaxy overdensity at any point with the overdensity at a point at angular separation $\theta$ – $w(\theta)$ depends on angular fluctuations on all scales. This is illustrated by the inverse of equation \[eqwtocl\]: $$w(\theta) = \frac{1}{4\pi \, \sigma_0^2} \, \sum_{\ell=1}^\infty (2\ell + 1) \, C_\ell \, P_\ell(\cos{\theta})$$ For example, a density map constructed using just one multipole $\ell$ possesses a broad angular correlation function. Relation to the spatial power spectrum {#secpk} ====================================== The purpose of this Section is to demonstrate that the observed radio galaxy $C_\ell$ spectrum can be understood in terms of the current fiducial cosmological model. In Section \[secbias\] we utilize this framework to determine the linear bias factor of the radio galaxies, marginalizing over other relevant model parameters. Theory ------ The angular power spectrum $C_\ell$ is a projection of the spatial power spectrum of mass fluctuations at different redshifts, $P(k,z)$, where $k$ is a co-moving wavenumber. In linear perturbation theory, fluctuations with different $k$ evolve independently, scaling with redshift according to the growth factor $D(z)$. Under this assumption we can simply scale the present-day matter power spectrum $P_0(k)$ back with redshift: $$P(k,z) = P_0(k) \, D(z)^2 \label{eqdz}$$ For an $\Omega_m = 1$, $\Omega_\Lambda = 0$ universe, $D(z) = (1+z)^{-1}$. For general cosmological parameters we can use the approximation of Carroll, Press & Turner (1992). If linear theory holds then the angular power spectrum $C_\ell$ can be written in terms of the present-day matter power spectrum $P_0(k)$ as $$C_\ell = \int \, P_0(k) \, W_\ell(k) \, dk \label{eqpktocl}$$ (see e.g. Huterer, Knox & Nichol 2001, Tegmark et al. 2002) where the kernel $W_\ell(k)$ is given by $$W_\ell(k) = \frac{2}{\pi} \, \left[ \int_0^\infty j_\ell(u) \, f(u/k) \, du \right]^2 \label{eqwl}$$ Here, $j_\ell$ is a spherical Bessel function and $f(x)$ is a function which depends on the radial distribution of the sources as $$f( \, x(z) \, ) = \frac{p(z) \, D(z) \, b(z)}{dx/dz} \label{eqrad}$$ assuming a flat geometry. $x(z)$ is the co-moving radial co-ordinate at redshift $z$, $p(z)$ is the redshift probability distribution of the sources (normalized such that $\int \, p(z) \, dz = 1$) and $b(z)$ is a linear bias factor which relates the clustering of galaxies to clustering of the underlying mass: $$P_{\rm gal}(k,z) = b(z)^2 P_{\rm mass}(k,z)$$ For the purposes of this analysis we assumed that the linear bias does not evolve with epoch and may be represented by a constant bias factor $b(z) = b_0$. A useful approximation for the spherical Bessel function (which gets better as $\ell$ gets larger) is $j_\ell(x) \approx (\pi/2\ell)^{1/2} \, \delta(x - \ell)$. In this approximation, $$W_\ell(k) \approx \frac{1}{\ell} \, f(\ell/k)^2 \label{eqwlapp}$$ Thus as $\ell$ is increased, the kernel just translates along the $k$-axis. Combining equations \[eqpktocl\] and \[eqwlapp\] produces the approximation $$C_\ell \approx \frac{1}{\ell} \int \, P_0(k) \, f(\ell/k)^2 \, dk \label{eqclkapp}$$ or, converting this into an integral over radial co-ordinate, $$C_\ell \approx \int \, P_0(\ell/x) \, x^{-2} \, f(x)^2 \, dx \label{eqclxapp}$$ In Section \[secmeascl\] we found that the measured NVSS $C_\ell$ spectrum was well-fit by a power-law. Inspecting equations \[eqclkapp\] and \[eqclxapp\], a power law for $C_\ell$ can arise in two ways: - If the function $f(x)$ were a power-law $f(x) \propto x^n$, then equation \[eqclkapp\] predicts that $C_\ell \propto \ell^{\, 2n-1}$, regardless of the form of $P_0(k)$. In particular, if $f(x)$ were approximately constant over the relevant scales, then $C_\ell \propto \ell^{-1}$, which is a good description of the observed NVSS clustering. - If the spatial power spectrum were a power-law $P_0(k) \propto k^n$, then equation \[eqclxapp\] predicts that $C_\ell \propto \ell^n$, regardless of the form of the radial distribution $p(z)$ (see also Table \[tabclus\]). The result $C_\ell \propto \ell^n$ differs from the case of CMB fluctuations, for which $n = 1$ corresponds to $C_\ell \propto 1/\ell(\ell+1)$ at low $\ell$ – the well-known Sachs-Wolfe effect. This is because the Sachs-Wolfe effect is sensitive to fluctuations in gravitational potential, whereas we are probing fluctuations in mass. In Section \[secnz\] we determine that the first of these interpretations is correct. ------------------------------------------------------------------------------------------ Statistic Dependence Validity ------------------------------ ------------------------------------------ ---------------- Spatial correlation function $\xi(r) \propto r^{\, -\gamma}$ All $r$ Angular correlation function $w(\theta) \propto \theta^{\, 1-\gamma}$ Small $\theta$ Spatial power spectrum $P(k) \propto k^{\, All $k$ \gamma-3}$ Angular power spectrum $C_\ell \propto High $l$ \ell^{\, \gamma-3}$ ------------------------------------------------------------------------------------------ : Functional form of common clustering statistics for pure power-law clustering parameterized by slope $\gamma$.[]{data-label="tabclus"} Modelling the present-day power spectrum {#secpkmod} ---------------------------------------- We assumed that the primordial matter power spectrum is a featureless power-law, $P_{\rm prim}(k) = A \, k^n$. On very large scales, the only alteration to this spectrum in linear theory will be an amplitude change due to the growth factor. However, during the epoch of radiation domination, growth of fluctuations on scales less than the horizon scale is suppressed by radiation pressure. This process is described by the transfer function $T(k)$, such that the present-day linear matter power spectrum is given by $$P_0(k) = P_{\rm prim}(k) \, T(k)^2 = A \, k^n \, T(k)^2 \label{eqpkmod}$$ Accurate fitting formulae have been developed for the transfer function $T(k)$ in terms of the cosmological parameters (Eisenstein & Hu 1998), which we employed in our analysis (these fitting formulae assume adiabatic perturbations). In our fiducial cosmological model, we fixed the values of the cosmological parameters at $h = H_0/(100 {\rm \; km \, s^{-1} \, Mpc^{-1}}) = 0.73$, $\Omega_m h^2 = 0.134$ and $\Omega_b/\Omega_m = 0.17$ (Spergel et al. 2003, Table 7, column 3). We also chose a primordial spectral index $n = 0.97$ (Spergel et al. 2003), which is close to the predictions of standard inflationary models. We consider the effect of variations in these parameter values in Section \[secbias\]. We assumed that the Universe is flat, with the remaining energy density provided by a cosmological constant $\Omega_\Lambda = 1 - \Omega_m$. There are at least two independent ways of estimating the amplitude $A$ in equation \[eqpkmod\]. Firstly we can use constraints on the number density of massive clusters at low redshift, expressed in terms of $\sigma_8$, the rms fluctuation of mass in spheres of radius $R = 8 \, h^{-1}$ Mpc: $$\sigma_R^2 = \frac{9}{2\pi^2 R^2} \int P_0(k) \, [j_1(kR)]^2 \, dk \label{eqsigsq}$$ For example, Viana & Liddle (1999) determined the most likely value of $\sigma_8$ using this method to be $\sigma_8 = 0.56 \, \Omega_m^{-0.47}$. Alternatively, $A$ can be expressed in terms of the amplitude of fluctuations at the Hubble radius, $\delta_H$, and constrained by measurements of CMB anisotropies on large angular scales: $$A = 2 \pi^2 \delta_H^2 \left( \frac{c}{H_0} \right)^{3+n} \label{eqdelh}$$ For example, Bunn & White (1997) give the best-fitting constraint on $\delta_H$ and $n$ for flat models ($\Omega_m + \Omega_\Lambda = 1$) based on results of the COBE DMR experiment: $$\delta_H = 1.94 \times 10^{-5} \, \Omega_m^{-0.785-0.05\ln{\Omega_m}} e^{-0.95(n-1) - 0.169(n-1)^2} \label{eqcobe}$$ with a maximum $1\sigma$ statistical uncertainty of 7 per cent. For our fiducial cosmological parameters, a (reasonably) consistent cosmology is produced if $\sigma_8 = 1$ (i.e. $\delta_H = 4.7 \times 10^{-5}$). We assumed this fiducial normalization in our model, noting that the value of $\sigma_8$ is in fact degenerate with the amplitude of a constant linear bias factor $b(z) = b_0$. Equation \[eqpkmod\] is only valid ignoring non-linear effects, which will boost the value of $P(k)$ on small scales as modes commence non-linear collapse. We incorporated non-linear corrections using the fitting formula provided by Peacock & Dodds (1996). The resulting model power spectrum is shown in Figure \[figpkmod\]. Strictly, this modification violates the assumption of linear evolution implicit in equation \[eqdz\]. However, this is not significant in our analysis because the small scales for which non-linear evolution is important are only significant in the projection at $z \approx 0$. Modelling the radial distribution of NVSS sources {#secnz} ------------------------------------------------- The projection of the spatial power spectrum onto the sky depends on the radial distribution of the sources under consideration, which may be deduced from their redshift distribution. We only need to know the probability distribution $p(z)$ of the sources (equation \[eqrad\]), the absolute normalization is not important. Unfortunately, the radial distribution of mJy radio sources is not yet accurately known. The majority of radio galaxies are located at cosmological distances ($z \sim 1$) and their host galaxies are optically very faint. However, models exist of the radio luminosity function of AGN (i.e. the co-moving space density of objects as a function of radio luminosity and redshift), from which the redshift distribution at any flux-density threshold can be inferred. Such models have been published by Dunlop & Peacock (1990) and Willott et al. (2001). These luminosity function models are by necessity constrained by relatively bright radio sources ($S_{\rm 1.4 \, GHz} > 100$ mJy) and the extrapolation to NVSS flux-density levels must be regarded as very uncertain. The Willott model is constrained by a larger number of spectroscopic redshifts, and the samples of radio sources used provide fuller coverage of the luminosity-redshift plane (thus the required extrapolation to $S_{\rm 1.4 \, GHz} = 10$ mJy is less severe). However, the Willott samples are selected at low frequency (151 and 178 MHz), necessitating a large extrapolation to the NVSS observing frequency of 1.4 GHz. The Dunlop & Peacock models are constrained at high frequencies, but treat steep-spectrum and flat-spectrum radio sources as independent populations, which is inconsistent with current ideas concerning the unification of radio AGN (e.g. Jackson & Wall 1999). In addition, they were computed for cosmological parameters $\Omega_m = 1$, $\Omega_\Lambda = 0$ rather than the currently favoured “$\Lambda$CDM” cosmology. Furthermore, none of the aforementioned luminosity function models incorporate starburst galaxies, which contribute in significant numbers to the radio galaxy population mix at $z \la 0.1$ for flux-density threshold $S_{\rm 1.4 \, GHz} = 10$ mJy. Direct measurements of $p(z)$ are currently only achievable at low redshifts ($z \la 0.2$), where comparison with large optical galaxy redshift surveys is possible (Sadler et al. 2002; Magliocchetti et al. 2002). We matched the NVSS 10 mJy catalogue with the final data release of the 2dF Galaxy Redshift Survey (2dFGRS, available online at [http://msowww.anu.edu.au/2dFGRS/]{}, also see Colless et al. 2001), in order to estimate $p(z)$ at low redshift. As our fiducial model, we fitted the resulting redshift histogram with the simplest possible function, a constant $p(z) = p_0$ over the range $0 < z < 0.15$. In Section \[secbias\] we include the effect of variations in this model. We only considered the redshift range $z < 0.15$ in this analysis because at $z \approx 0.15$, the (very luminous) optical counterparts of the 10 mJy NVSS sources begin slipping below the 2dFGRS magnitude threshold, which we verified by plotting the magnitudes of matched 2dFGRS galaxies against redshift. Having determined the value of $p_0$, we created the full redshift distribution by assigning the remaining probability $1-p_0$ over the redshift range $z > 0.15$ in proportion to the prediction of the Dunlop & Peacock (1990) luminosity function models. For this investigation we used the average of the seven models provided by Dunlop & Peacock. Assuming the Willott et al. (2001) radial distribution for $z > 0.15$ made a negligible difference to the results (because most of the contribution to the $C_\ell$ spectrum arises at low redshifts, see Section \[secclpred\]). Matching NVSS catalogue entries brighter than $S_{\rm 1.4 \, GHz} = 10$ mJy with the 2dFGRS database yielded $N_{\rm mat} = 546$ identifications with redshifts $z < 0.15$, using matching tolerance 10 arcsec (see Sadler et al. 2002). We restricted the 2dFGRS sample to “high quality” spectra ($Q \ge 3$). An estimate of the probability of an NVSS source being located at $z < 0.15$ is $${\rm Prob} = p_0 \, \Delta z = \frac{N_{\rm mat}/A_{\rm 2dF}}{\sigma_{\rm NVSS}}$$ where $\sigma_{\rm NVSS} = 16.9$ deg$^{-2}$ is the surface density of NVSS sources brighter than 10 mJy, and $A_{\rm 2dF}$ is the 2dFGRS area under consideration, which (owing to the varying angular completeness) is not trivial to calculate. We followed Sadler et al. (2002) by dividing the number of 2dFGRS galaxies contained in the NVSS geometry by the 2dFGRS surface density $\sigma_{\rm 2dF} = 180$ deg$^{-2}$, resulting in an effective area $A_{\rm 2dF} = 1214$ deg$^2$. The result, $p_0 = 0.177$ (per unit redshift), is a significant underestimate for various reasons: - Incompletenesses in the 2dFGRS input catalogue ($5\%$). - Input catalogue galaxies unable to be assigned a 2dF spectrograph fibre ($7\%$). - Observed 2dFGRS spectra with insufficient quality ($Q \le 2$) to determine a redshift ($8\%$). - Extended radio sources with catalogue entries located more than 10 arcsec from the optical counterpart ($3\%$). The estimated correction factors in brackets were obtained from Colless et al. (2001) and from Carole Jackson (priv. comm.). Multiplying these corrections implies a total incompleteness of $25\%$, and on this basis we increased the value of $p_0$ to $0.177/0.75 = 0.237$. Figure \[fignz\] plots 2dFGRS-NVSS matches in redshift bins of width $\Delta z = 0.01$, together with the low-redshift fit described above. A constant $p(z)$ is a fairly good approximation at low redshifts ($z < 0.15$). This flat distribution arises because the overall redshift distribution is a sum of that due to AGN and that due to starburst galaxies; $p(z)$ increases with $z$ for the AGN, but decreases with $z$ for the starbursters. Figure \[fignz\] also displays the predictions of the luminosity function models of Dunlop & Peacock (1990) and Willott (2001). As explained above, the large extrapolations involved render these models a poor fit to the redshift distribution at mJy flux levels, and their use without modification at low redshift would have caused significant error. Predicting the $C_\ell$ spectrum {#secclpred} -------------------------------- We used equation \[eqpktocl\] to predict the $C_\ell$ spectrum from our fiducial models of the spatial power spectrum (Section \[secpkmod\]) and the radial distribution of the sources (Section \[secnz\]). We found that a good match to the measured angular power spectrum resulted if the NVSS sources were assigned a constant linear bias factor $b_0 \approx 1.7$ (Figure \[figclmod\]); $b_0 = 1$ provides a very poor fit to the results. The bias factor of the radio galaxies is analyzed more thoroughly in Section \[secbias\]. We note that these measurements of the radio galaxy $C_\ell$ spectrum at low $\ell$ are [*not*]{} directly probing the large-scale, small $k$, region of the power spectrum $P(k)$. Investigation of the integrands of equations \[eqclkapp\] and \[eqclxapp\] revealed that the majority of the signal is built up at low redshift, $z \la 0.1$ (see Figure \[figkz\]), where small-scale spatial power is able to contribute on large angular scales (i.e. contribute to low multipoles). Higher redshift objects principally serve to dilute the clustering amplitude. This is unfortunate: the potential of radio galaxies distributed to $z \sim 1$ to probe [*directly*]{} the large-scale power spectrum is forfeited by projection effects. In order to realize this potential, we must measure redshifts for the NVSS sources. A three-dimensional map extending to $z \sim 1$ would directly yield $P(k)$ on large scales, defining the “turn-over” sketched in Figure \[figpkmod\]. Radio galaxy bias factor {#secbias} ======================== The enhanced radio galaxy bias apparent in Figure \[figclmod\] is consistent with the nature of AGN host galaxies: optically luminous ellipticals inhabiting moderate to rich environments. Similarly high radio galaxy bias factors have been inferred from measurements of the spatial power spectrum of low-redshift radio galaxies (Peacock & Dodds 1994) and from deprojection of the NVSS angular correlation function $w(\theta)$ (Blake & Wall 2002a; Overzier et al. 2003). Furthermore, Boughn & Crittenden (2002) cross-correlated NVSS and CMB overdensities in a search for the Integrated Sachs-Wolfe effect (see also Nolta et al. 2003). Their analysis included fitting theoretical models to the NVSS angular correlation function. They derived good fits for a slightly lower linear bias factor than ours, $b_0 = 1.3 \rightarrow 1.6$. This difference is due to the assumed redshift distribution. Boughn & Crittenden also used the Dunlop & Peacock (1990) average model, but we corrected this model using observational data at low redshifts $z < 0.15$. As can be seen from Figure \[fignz\], our correction reduces the number of low-redshift sources, necessitating a higher bias factor to recover the same angular correlations. In order to derive a formal confidence interval for the linear bias parameter $b_0$ we must incorporate the effects of uncertainties in the underlying model parameters (by marginalizing over those parameters). With this in mind, we assumed Gaussian priors for Hubble’s constant $h = 0.73 \, \pm \, 0.03$, for the matter density $\Omega_m h^2 = 0.134 \, \pm \, 0.006$, and for the primordial spectral index $n = 0.97 \, \pm \, 0.03$. The widths of these priors were inspired by the cosmological parameter analysis combining the WMAP satellite observations of the CMB and the 2dFGRS galaxy power spectrum (Spergel et al. 2003, Table 7, column 3) and are a good representation of our current knowledge of the cosmological model. In addition, we considered variations in the model for the radial distribution of NVSS sources at low redshift (Section \[secnz\]), using a more general fitting formula $p(z) = a + b \, z$ to describe the probability distribution for $z < 0.15$. For each pair of values of $(a,b)$ we derived the chi-squared statistic between the model and the observations, $\chi^2_{p(z)}$, which we converted into an (unnormalized) probability density $P_{p(z)} \propto \exp{(-\chi^2_{p(z)}/2)}$. Our model is thus specified by values of ($b_0$, $h$, $\Omega_m h^2$, $n$, $a$, $b$) from which we can calculate a model $C_\ell$ spectrum and hence a chi-squared statistic with the observations, $\chi^2_{C_\ell}$, corresponding to a probability density $P_{C_\ell} \propto \exp{(-\chi^2_{C_\ell}/2)}$. We used the spherical harmonic estimation of the $C_\ell$ spectrum as the observational data. After multiplying $P_{C_\ell}$ by the redshift distribution probability density $P_{N(z)}(a,b)$ and the Gaussian prior probability densities for $h$, $\Omega_m h^2$ and $n$, we derived the probability distribution for $b_0$ by integrating over each of the other parameters. We do not marginalize over the normalization of the matter power spectrum, $\sigma_8$, because this quantity is degenerate with $b_0$ (using equations \[eqrad\], \[eqclxapp\] and \[eqsigsq\]: $C_\ell \propto b_0^2 \, \sigma_8^2$). The resulting normalized probability distribution for $b_0 \times \sigma_8$ is displayed in Figure \[figb0\], from which we determined a $68\%$ confidence region $b_0 \, \sigma_8 = 1.53 \rightarrow 1.87$. When combined with the WMAP determination of $\sigma_8 = 0.9 \pm 0.1$ (Spergel et al. 2003), we infer that $b_0 = 1.89 \pm 0.27$. Conclusions =========== This investigation has measured the angular power spectrum of radio galaxies for the first time, yielding consistent results through the application of two independent methods: direct spherical harmonic analysis and maximum likelihood estimation. The NVSS covers a sufficient fraction of sky ($\sim 80$ per cent) that spherical harmonic analysis is very effective, with minimal correlations amongst different multipoles. The form of the $C_\ell$ spectrum can be reproduced by standard models for the present-day spatial power spectrum and for the radial distribution of NVSS sources – provided that this latter is modified at low redshift through comparison with optical galaxy redshift surveys. The results strongly indicate that radio galaxies possess high bias with respect to matter fluctuations. A constant linear bias $b_0 \approx 1.7$ permits a good fit, and by marginalizing over the other parameters of the model we deduce a $68\%$ confidence interval $b_0 \, \sigma_8 = 1.53 \rightarrow 1.87$ where $\sigma_8$ describes the normalization of the matter power spectrum. We find that the majority of the angular power spectrum signal is generated at low redshifts, $z \la 0.1$. Therefore, in order to exploit the potential of radio galaxies to probe spatial fluctuations on the largest scales, we require individual redshifts for the NVSS sources. Acknowledgments {#acknowledgments .unnumbered} =============== We thank Jasper Wall and Steve Rawlings for helpful comments on earlier drafts of this paper. We acknowledge valuable discussions with Carole Jackson concerning cross-matching the NVSS and 2dFGRS source catalogues. We are grateful to Sarah Bridle for useful guidance on marginalizing over the cosmological model. Baleisis A., Lahav O., Loan A.J., Wall J.V., 1998, MNRAS, 297, 545 Becker R.H., White R.L., Helfand D.J., 1995, ApJ, 450, 559 Blake C.A., Wall J.V, 2002a, MNRAS, 329, L37 Blake C.A., Wall J.V, 2002b, Nature, 416, 150 Bond J.R., Jaffe A.H., Knox L., 1998, PhRvD, 57, 2117 Bond J.R., Efstathiou G., Tegmark M., 1997, MNRAS, 291, 33 Borrill J., 1999, in [*Proceedings of the 5th European SGI/Cray MPP Workshop*]{} ([astro-ph/9911389]{}) Boughn S.P., Crittenden R.G., 2002, PhRvL, 88, 1302 Brand K., Rawlings S., Hill G.J., Lacy M., Mitchell E., Tufts J., 2003, MNRAS, 344, 283 Bunn E.F., White M., 1997, ApJ, 480, 6 Carroll S.M., Press W.H., Turner E.L., 1992, ARA&A, 30, 499 Colless M. et al., 2001, MNRAS, 328, 1039 Condon J., Cotton W., Greisen E., Yin Q., Perley R., Taylor G., Broderick J., 1998, AJ, 115, 1693 Cress C., Helfand D., Becker R., Gregg M., White R., 1996, ApJ, 473, 7 Dunlop J.S., Peacock J.A., 1990, MNRAS, 247, 19 Efstathiou G., Moody S., 2001, MNRAS, 325, 1603 Eisenstein D.J., Hu W., 1998, ApJ, 496, 605 Gorski K.M., Hivon E., Wandelt B.D., 1999, in [*Proceedings of the MPA/ESO Cosmology Conference “Evolution of Large-Scale Structure”*]{}, p.37 ([astro-ph/9812350]{}) Hauser M.G., Peebles P.J.E., 1973, ApJ, 185, 757 Hill G.J., Lilly S.J., 1991, ApJ, 367, 1 Huterer D., Knox L., Nichol R.C., 2001, ApJ, 555, 547 Jackson C.A., Wall J.V., 1999, MNRAS, 304, 160 Magliocchetti M., Maddox S., Lahav O., Wall J., 1998, MNRAS, 300, 257 Magliocchetti M. et al., 2002, MNRAS, 333, 100 Nolta et al., 2003, ApJ submitted ([astro-ph/0305097]{}) Overzier R.A., Röttgering H.J.A., Rengelink R.B., Wilman R.J., 2003, A&A, 405, 53 Peacock J.A., Dodds S.J., 1994, MNRAS, 267, 1020 Peacock J.A., Dodds S.J., 1996, MNRAS, 280, 19 Peebles P.J.E., 1973, ApJ, 185, 413 Sadler E.M. et al., 2002, MNRAS, 329, 227 Scott D., Srednicki M., White M., 1994, ApJ, 421, 5 Spergel D.N. et al. 2003, ApJS, 148, 175 Tegmark M. et al., 2002, ApJ, 571, 191 Viana P.T.P., Liddle A.R., 1999, MNRAS, 303, 535 Wandelt B.D., Hivon E., Gorski K.M., 2001, PhRvD, 64, 3003 Willott C.J., Rawlings S., Blundell K.M., Lacy M., Eales S.A., 2001, MNRAS, 322, 536 Wright E.L., Smoot G.F., Bennett C.L., Lubin P.M., 1994, ApJ, 436, 441

ArXiv

--- abstract: 'Because of their state-of-the-art performance in computer vision, CNNs are becoming increasingly popular in a variety of fields, including medicine. However, as neural networks are black box function approximators, it is difficult, if not impossible, for a medical expert to reason about their output. This could potentially result in the expert distrusting the network when he or she does not agree with its output. In such a case, explaining why the CNN makes a certain decision becomes valuable information. In this paper, we try to open the black box of the CNN by inspecting and visualizing the learned feature maps, in the field of dermatology. We show that, to some extent, CNNs focus on features similar to those used by dermatologists to make a diagnosis. However, more research is required for fully explaining their output.' author: - Pieter Van Molle - Miguel De Strooper - Tim Verbelen - | \ Bert Vankeirsbilck - Pieter Simoens - Bart Dhoedt bibliography: - 'bibtex.bib' title: | Visualizing Convolutional Neural Networks\ to Improve Decision Support\ for Skin Lesion Classification --- Introduction ============ Over the past few years, deep neural network architectures—convolutional architectures in particular—have time and again beaten state-of-the-art on large-scale image recognition tasks [@krizhevsky2012imagenet; @simonyan2014very; @szegedy2016rethinking; @he2016deep]. As a result, the application of convolutional neural networks (CNN) has become increasingly popular in a variety of fields. In medicine, deep learning is used as a tool to assist professionals of various subfields in their diagnoses, such as histopathology [@litjens2016deep], oncology [@cirecsan2013mitosis; @fakoor2013using; @wang2016deep], pulmonology [@hua2015computer; @sun2016computer], etc[^1]. In the subfield of dermatology, CNNs have been applied to the problem of skin lesion classification, based on dermoscopy images, where they set a new state-of-the-art benchmark, matching—or even surpassing—medical expert performance [@codella2017deep; @esteva2017dermatologist; @Haenssle18]. The challenge remains, however, to understand the reasoning behind the decisions made by these networks, since they are essentially black box function approximators. This poses a problem when a neural network outputs a diagnosis, different from the diagnosis made by the medical expert, as there is no human interpretable reasoning behind the neural networks’ diagnosis. In such a case, visualizations of the network could serve as a reasoning tool to the expert. In this paper, we train a CNN for binary classification on a skin lesion dataset, and inspect the features learned by the network, by visualizing its feature maps. In the next section, we first give an overview of the different visualization strategies for inspecting CNNs. Section 3 describes our CNN architecture and training procedure. In Section 4 we present and discuss the learned CNN features and we conclude the paper in Section 5. Related Work ============ In [@zeiler2014visualizing], the authors propose a visualization technique to give some insight into the function of the intermediate feature maps of a trained CNN, by attaching a deconvolutional network to each of its convolutional layers. While a CNN maps the input from the image space to a feature space, a deconvolutional network does the opposite (mapping from a feature space back to the image space), by reversing its operations. This is done by a series of unpooling, rectifying and filtering operations. The authors use a deconvolutional network to visualize the features that result in the highest activations in a given feature map. Furthermore, they evaluate the sensitivity of a feature map, to the occlusion of a certain part of the input image, and the effect it has on the class score for the correct class. Two other visualization techniques are presented in [@simonyan2013deep] that are based on optimization. The first technique iteratively generates a canonical image representing a class of interest. To generate this image, the authors start from a zero image and pass it through a trained CNN. Optimization is done by means of the back-propagation algorithm, by calculating the derivative of the class score, with respect to the image, while keeping the parameters of the network fixed. The second technique aims to visualize the image-specific class saliency. For a given input image and a class of interest, they calculate the derivative of the class score, with respect to the input image. The per-pixel derivatives of the input image give an estimate of the importance of these pixels regarding the class score. More specifically, the magnitude of the derivate indicates which pixels affect the class score the most when they are changed. Concluding, typical visualization techniques either generate a single output image, in case of the feature visualization and the generation of the class representative, or function at the pixel level of the input image, in case of the region occlusion and the image-specific class saliency visualization. However, dermatologists typically scan a lesion for the presence of different individual features, such as asymmetry, border, color and structures, i.e. the so-called ABCD-score [@nachbar1994abcd]. Therefore, we inspect and visualize the intermediary feature maps of the CNN on a per-image basis, aiming to provide more familiar insights to dermatologists. Architecture and Training ========================= A common approach is to use a CNN pre-trained on a large image database such as ImageNet and then fine-tune this on the target dataset [@Haenssle18]. The drawback is that this CNN will also contain a lot of uninformative filters (e.g. for classifying cats and dogs) for the domain at hand. Therefore we chose to train a basic CNN from scratch, but in principle our visualization approach can work for any CNN. Our CNN consists of 4 convolutional blocks, each formed by 2 convolutional layers followed by a max pooling operation. The convolutional layers in each block have a kernel size of $3 \times 3$, and have respectively 8, 16, 32 and 64 filters. This is followed by 3 fully connected layers with 2056, 1024 and 64 hidden units. All layers have rectified linear units (ReLU) as non-linearity. We use data from the publicly available ISIC Archive[^2], to compose a training set of 12,838 dermoscopy images, spread over two classes (11,910 benign lesions, 928 malignant lesions). In a preprocessing step, the images are downscaled to a resolution of $300 \times 300$ pixels, and RGB values are normalized between 0 and 1. We augment our training set by taking random crops of $224 \times 224$ pixels, and further augment each crop by rotating (angle sampled uniformly between 0 and $2 \pi$), randomly flipping horizontally and/or vertically, adjusting brightness (factor sampled uniformly between -0.5 and 0.5), contrast (factor sampled uniformly between -0.7 and 0.7), hue (factor sampled uniformly between -0.02 and 0.02) and saturation (factor sampled uniformly between 0.7 and 1.5). We have trained the network for 192 epochs, with mini-batches of size 96 and used the Adam algorithm [@kingma2014adam] to update the parameters of the network, with an initial learning rate of $10^{-4}$ and an exponential decay rate for the first and second order momentum of respectively 0.9 and 0.999. We have evaluated the performance of the resulting CNN on a hold-out test set, comprised of 600 dermoscopy images (483 benign lesions, 117 malignant lesions), achieving an AUC score of 0.75. Feature Map Visualization ========================= For each feature map of the CNN, we created a visualization by rescaling the feature map to the input size and overlaying the activations mapped to a transparent green color (darker green = higher activation). We identify each visualization by the convolutional layer number (0..7) and filter number. Next we inspected all visualizations and tried to relate these to typical features dermatologists scan for. Especially the last two convolutional layers of the CNN (6,7) give us some insights into which image regions grasp the attention of the CNN. **Borders** Irregularities in the border of a skin lesion could indicate a malignant lesion. The feature maps shown in Fig. \[fig:borders\] both have high activations on the border of a skin lesion, but on different parts of the border. The first one (a) detects the bottom border of a lesion, while the second one (b) detects the left border. **Color** The same reasoning tends to apply to the colors inside the lesion. A lesion that has a uniform color is usually benign, while major irregularities in color could be a sign of a malignant lesion. The feature maps shown in Fig. \[fig:colors\] have a high activation when a darker region is present in the lesion, implying a non-uniform color. **Skin Type** People with a lighter skin are more prone to sunburns, which can increase the development of malignant lesions on their skin. Therefore, a dermatologist takes a patient’s skin type into account when examining his or her lesions. The same goes for the feature maps shown in Fig. \[fig:skin\]. The first feature map (a) has high activations on white-pale skin. The second one (b) has high activations on a more pinkish skin with vessel-like structures. **Hair** The CNN also learns feature maps that, from a dermatologist viewpoint, have no impact on the diagnosis. For example, the feature map in Fig. \[fig:hair\] has high activations on hair-like structures. **Artifacts** We also noticed that some of the feature maps have high activations on various artifacts in the images. For example, as shown in Fig. \[fig:artifacts\], some feature maps have high activations on specular reflections, gel application, or rulers. This highlights some of the risks when using machine learning techniques, as this could impose a potential bias to the output of the network, when such artifacts are prominently present in the training images of a specific class. A more elaborate overview of the activations of different feature maps on different images is shown in Fig. \[fig:matrix\]. -- -- -- -- -- -- -- -- -- -- -- -- Conclusion ========== In this paper, we analyzed the features learned by a CNN, trained for skin lesion classification, in the field of dermatology. By visualizing the feature maps of the CNN, we see that, indeed, the high-level convolutional layers activate on similar concepts as used by dermatologists, such as lesion border, darker regions inside the lesion, surrounding skin, etc. We also found that some feature maps activate on various image artifacts, such as specular reflections, gel application, and rulers. This flags that one should be cautious when constructing a dataset for training, that such artifacts do not lead to a bias in the machine learning model. Although this paper gives some insight in the features learned by the CNN, this does not yet explain any causal relation between the detected features of the CNN and its output. Furthermore, going through the feature maps, we did not find any that precisely highlight many of the other structures that dermatologists scan for, such as globules, dots, blood vessel structures, etc. We believe more research is required in this area in order to make CNNs a better decision support tool for dermatologists. [^1]: We refer the reader to [@litjens2017survey] for an in-depth survey on deep learning in medical analysis. [^2]: <https://isic-archive.com/>

ArXiv

--- abstract: '3D object detection from a single image without LiDAR is a challenging task due to the lack of accurate depth information. Conventional 2D convolutions are unsuitable for this task because they fail to capture local object and its scale information, which are vital for 3D object detection. To better represent 3D structure, prior arts typically transform depth maps estimated from 2D images into a pseudo-LiDAR representation, and then apply existing 3D point-cloud based object detectors. However, their results depend heavily on the accuracy of the estimated depth maps, resulting in suboptimal performance. In this work, instead of using pseudo-LiDAR representation, we improve the fundamental 2D fully convolutions by proposing a new local convolutional network (LCN), termed Depth-guided Dynamic-Depthwise-Dilated LCN (D$^4$LCN), where the filters and their receptive fields can be automatically learned from image-based depth maps, making different pixels of different images have different filters. D$^4$LCN overcomes the limitation of conventional 2D convolutions and narrows the gap between image representation and 3D point cloud representation. Extensive experiments show that D$^4$LCN outperforms existing works by large margins. For example, the relative improvement of D$^4$LCN against the state-of-the-art on KITTI is 9.1% in the moderate setting. D$^4$LCN ranks $1^{\mathrm{st}}$ on KITTI monocular 3D object detection benchmark at the time of submission (car, December 2019) . The code is available at <https://github.com/dingmyu/D4LCN>.' author: - | Mingyu Ding^1,2^   Yuqi Huo ^2,5^   Hongwei Yi ^3^   Zhe Wang^4^   Jianping Shi^4^   Zhiwu Lu^2,5^   Ping Luo^1^\ ^1^The University of Hong Kong   ^2^Gaoling School of Artificial Intelligence, Renmin University of China  \ ^3^Shenzhen Graduate School, Peking University            ^4^SenseTime Research\ ^5^Beijing Key Laboratory of Big Data Management and Analysis Methods, Beijing 100872, China\ `{myding, pluo}@cs.hku.hk`    `{bohony,luzhiwu}@ruc.edu.cn`   \ `[email protected]`    `{wangzhe,shijianping}@sensetime.com`\ bibliography: - 'egbib.bib' title: 'Learning Depth-Guided Convolutions for Monocular 3D Object Detection' --- ![(a) and (b) show pseudo-LiDAR points generated by the supervised depth estimator, DORN [@fu2018deep] and the unsupervised Monodepth [@monodepth17] respectively. The green box represents ground-truth (GT) 3D box. Pseudo-LiDAR points generated by inaccurate depth as shown in (b) have large offsets comapred to the GT box. (c) and (d) show the detection results of our method and Pseudo-Lidar [@wang2019pseudo] by using a coarse depth map. The performance of [@wang2019pseudo] depends heavily on the accuracy of the estimated depth maps, while our method achieves accurate detection results when accurate depth maps are missing. []{data-label="fig:first_fig"}](first.pdf){width="0.9\columnwidth"} Introduction ============ 3D object detection is a fundamental problem and has many applications such as autonomous driving and robotics. Previous methods show promising results by utilizing LiDAR device, which produces precise depth information in terms of 3D point clouds. However, due to the high-cost and sparse output of LiDAR, it is desirable to seek cheaper alternatives like monocular cameras. This problem remains largely unsolved, though it has drawn much attention. Recent methods towards the above goal can be generally categorized into two streams as image-based approaches [@mousavian20173d; @li2019gs3d; @qin2019monogrnet; @hu2019joint; @he2019mono3d++; @chen2016monocular] and pseudo-LiDAR point-based approaches [@wang2019pseudo; @ma2019accurate; @weng2019monocular]. The image-based approaches [@chen20153d; @he2019mono3d++] typically leverage geometry constraints including object shape, ground plane, and key points. These constraints are formulated as different terms in loss function to improve detection results. The pseudo-LiDAR point-based approaches transform depth maps estimated from 2D images to point cloud representations to mimic the LiDAR signal. As shown in Figure \[fig:first\_fig\], both of these methods have drawbacks, resulting in suboptimal performance. Specifically, the image-based methods typically fail to capture meaningful local object scale and structure information, because of the following two factors. (1) Due to perspective projection, the monocular view at far and near distance would cause significant changes in object scale. It is difficult for traditional 2D convolutional kernels to process objects of different scales simultaneously (see Figure \[fig:conv\]). (2) The local neighborhood of 2D convolution is defined in the camera plane where the depth dimension is lost. In this non-metric space (the distance between pixels does not have a clear physical meaning like depth), a filter cannot distinguish objects from the background. In that case, a car area and the background area would be treated equally. Although pseudo-LiDAR point-based approaches have achieved progressive results, they still possess two key issues. (1) The performance of these approaches heavily relies on the precision of estimated depth maps (see Figure \[fig:first\_fig\]). The depth maps extracted from monocular images are often coarse (point clouds estimated using them have wrong coordinates), leading to inaccurate 3D predictions. In other words, the accuracy of the depth map limits the performance of 3D object detection. (2) Pseudo-LiDAR methods cannot effectively employ high-level semantic information extracted from RGB images, leading to many false alarms. This is because point clouds provide spatial information but lose semantic information. As a result, regions like roadblocks, electrical boxes and even dust on the road may cause false detection, but they can be easily discriminated by using RGB images. To address the above problems, we propose a novel convolutional network, termed Depth-guided Dynamic-Depthwise-Dilated local convolutional network (D$^4$LCN), where the convolutional kernels are generated from the depth map and locally applied to each pixel and channel of individual image sample, rather than learning global kernels to apply to all images. As shown in Figure \[fig:conv\], D$^4$LCN treats the depth map as guidance to learn local dynamic-depthwise-dilated kernels from RGB images, so as to fill the gap between 2D and 3D representation. More specifically, the learned kernel in D$^4$LCN is sample-wise (exemplar kernel [@he2010guided]), position-wise (local convolution [@jia2016dynamic]), and depthwise (depthwise convolution [@howard2017mobilenets]), where each kernel has its own dilation rate (different exemplar kernels have different receptive fields). ![**Comparisons** among different convolutional approaches. (a) is the traditional 2D convolution that uses a single convolutional kernel applied on each pixel to convolve the entire image. (b) applies multiple fixed convolutional kernels on different regions (slices) of an image. (c) uses the depth map to generate dynamic kernels with the same receptive fields for each pixel. (d) denotes our approach, where the filter is dynamic, depth-wise, and has adaptive receptive fields for each pixel and channel of the feature map. It can be implemented more efficiently with fewer parameters than (c). Best viewed in color. []{data-label="fig:conv"}](conv.pdf){width="0.97\columnwidth"} D$^4$LCN is carefully designed with four considerations. (1) The exemplar kernel is to learn specific scene geometry for each image. (2) The local convolution is to distinguish object and background regions for each pixel. (3) The depth-wise convolution is to learn different channel filters in a convolutional layer with different purposes and to reduce computational complexity. (4) The exemplar dilation rate is to learn different receptive fields for different filters to account for objects with diverse scales. The above delicate designs can be easily and efficiently implemented by combing linear operators of shift and element-wise product. As a result, the efficient D$^4$LCN can not only address the problem of the scale-sensitive and meaningless local structure of 2D convolutions, but also benefit from the high-level semantic information from RGB images compared with the pseudo-LiDAR representation. Our main **contributions** are three-fold. (1) A novel component for 3D object detection, D$^4$LCN, is proposed, where the depth map guides the learning of dynamic-depthwise-dilated local convolutions from a single monocular image. (2) We carefully design a single-stage 3D object detection framework based on D$^4$LCN to learn better 3D representation for reducing the gap between 2D convolutions and 3D point cloud-based operations. (3) Extensive experiments show that D$^4$LCN outperforms state-of-the-art monocular 3D detection methods and takes the first place on the KITTI benchmark [@geiger2012we]. Related Work ============ **Image-based Monocular 3D Detection.**  Previous monocular 3D detection methods [@mousavian20173d; @li2019gs3d; @qin2019monogrnet; @brazil2019m3d; @hu2019joint; @he2019mono3d++; @chen2016monocular; @xu2018multi] usually make assumptions about the scene geometry and use this as a constraint to train the 2D-to-3D mapping. Deep3DBox [@mousavian20173d] uses the camera matrix to project a predicted 3D box onto the 2D image plane, constraining each side of the 2d detection box, such that it corresponds to any of the eight corners of the 3D box. OFTNet [@roddick2018orthographic] introduces the orthographic feature transform, which maps image-based features into an orthographic 3D space. It is helpful when scale of objects varies drastically. [@jorgensen2019monocular; @liu2019deep] investigated different ways of learning the confidence to model heteroscedastic uncertainty by using a 3D intersection-over-union (IoU) loss. To introduce more prior information, [@chabot2017deep; @kundu20183d; @zeeshan2014cars; @xiang2015data] used 3D shapes as templates to get better object geometry. [@ku2019monocular] predicts a point cloud in an object-centered coordinate system and devises a projection alignment loss to learn local scale and shape information. [@manhardt2019roi] proposes a 3D synthetic data augmentation algorithm via in-painting recovered meshes directly onto the 2D scenes. However, as it is not easy for 2D image features to represent 3D structures, the above geometric constraints fail to restore accurate 3D information of objects from just a single monocular image. Therefore, our motivation is to utilize depth information, which essentially bridges gap between 2D and 3D representation, to guide learning the 2D-to-3D feature representation. **Point Cloud-based Monocular 3D Detection.**  Previous monocular methods [@wang2019pseudo; @ma2019accurate; @weng2019monocular] convert image-based depth maps to pseudo-LiDAR representations for mimicking the LiDAR signal. With this representation, existing LiDAR-based detection algorithms can be directly applied to monocular 3D object detection. For example, [@weng2019monocular] detects 2D object proposals in the input image and extracts a point cloud frustum from the pseudo-LiDAR for each proposal. [@ma2019accurate] proposes a multi-modal features fusion module to embed the complementary RGB cue into the generated point clouds representation. However, this depth-to-LiDAR transformation relies heavily on the accuracy of depth map and cannot make use of RGB information. In contrast, our method treats depth map as guidance to learn better 3D representation from RGB images. **LiDAR-based 3D Detection.**  With the development of deep learning on point sets, 3D feature learning [@qi2017pointnet; @qi2017pointnet++; @zhou2018voxelnet] is able to learn deep point-based and voxel-based features. Benefit from this, LiDAR-based methods have achieved promising results in 3D detection. For example, [@zhou2018voxelnet] divides point clouds into equally spaced 3D voxels and transforms a group of points within each voxel into a unified feature representation. [@wang2019voxel] applies the FPN technique to voxel-based detectors. [@yan2018second] investigates a sparse convolution for voxel-based networks. [@lang2019pointpillars] utilizes PointNets to learn a representation of point clouds organized in vertical columns (pillars). [@qi2018frustum] leverages mature 2D object detectors to learn directly from 3D point clouds. [@wang2019frustum] aggregates point-wise features as frustum-level feature vectors. [@shi2019pointrcnn; @chen2019fast] directly generated a small number of high-quality 3D proposals from point clouds via segmenting the point clouds of the whole scene into foreground and background. There are also some works focus on multi-sensor fusion (LIDAR as well as cameras) for 3D object detection. [@liang2018deep; @liang2019multi] proposed a continuous fusion layer that encodes both discrete-state image features as well as continuous geometric information. [@chen2017multi; @ku2018joint] used LIDAR point clouds and RGB images to generate features and encoded the sparse 3D point cloud with a compact multi-view representation. {width="91.00000%"} **Dynamic Networks.**  A number of existing techniques can be deployed to exploit the depth information for monocular 3D detection. M3D-RPN [@brazil2019m3d] proposes depth-aware convolution which uses non-shared kernels in the row-space to learn spatially-aware features. However, this rough and fixed spatial division has bias and fail to capture object scale and local structure. Dynamic filtering network [@jia2016dynamic] uses the sample-specific and position-specific filters but has heavy computational cost, and it also fails to solve the scale-sensitive problem of 2D convolutions. Trident network [@li2019scale] utilizes manually defined multi-head detectors for 2D detection. However, it needs to manually group data for different heads. Other techniques like deformable convolution [@dai2017deformable] and variants of [@jia2016dynamic] such as [@ha2016hypernetworks; @tang2019learning; @wu2018dynamic], fail to capture object scale and local structure as well. In this work, our depth-guided dynamic dilated local convolutional network is proposed to solve the two problems associated with 2D convolutions and narrow the gap between 2D convolution and point cloud-based 3D processing. Methodology =========== As a single-stage 3D detector, our framework consists of three key components: a network backbone, a depth-guided filtering module, and a 2D-3D detection head (see Figure \[fig:overview\]). Details of each component are given below. First, we give an overview of our architecture as well as backbone networks. We then detail our depth-guided filtering module which is the key component for bridging 2D convolutions and the point cloud-based 3D processing. Finally, we outline the details of our 2D-3D detection head. Backbone -------- To utilize depth maps as guidance of 2D convolutions, we formulate our backbone as a two-branch network: the first branch is the feature extraction network using RGB images, and the other is the filter generation network to generate convolutional kernels for feature extraction network using the estimated depth as input. These two networks process the two inputs separately and their outputs of each block are merged by the depth-guided filtering module. The backbone of the feature extraction network is ResNet-50 [@he2016deep] without its final FC and pooling layers, and is pre-trained on the ImageNet classification dataset [@deng2009imagenet]. To obtain a larger field-of-view and keep the network stride at 16, we find the last convolutional layer (conv5\_1, block4) that decreases resolution and set its stride to 1 to avoid signal decimation, and replace all subsequent convolutional layers with dilated convolutional layers (the dilation rate is 2). For the filter generation network, we only use the first three blocks of ResNet-50 to reduce computational costs. Note the two branches have the same number of channels of each block for the depth guided filtering module. Depth-Guided Filtering Module ----------------------------- Traditional 2D convolution kernels fail to efficiently model the depth-dependent scale variance of the objects and effectively reason about the spatial relationship between foreground and background pixels. On the other hand, pseudo-lidar representations rely too much on the accuracy of depth and lose the RGB information. To address these problems simultaneously, we propose our depth-guided filtering module. Notably, by using our module, the convolutional kernels and their receptive fields (dilation) are different for different pixels and channels of different images. Since the kernel of our feature extraction network is trained and generated by the depth map, it is sample-specific and position-specific, as in [@jia2016dynamic; @ha2016hypernetworks], and thus can capture meaningful local structures as the point-based operator in point clouds. We first introduce the idea of depth-wise convolution [@howard2017mobilenets] to the network, termed depth-wise local convolution (DLCN). Generally, depth-wise convolution (DCN) involves a set of global filters, where each filter only operates at its corresponding channel, while DLCN requires a feature volume of local filters the same size as the input feature maps. As the generated filters are actually a feature volume, a naive way to perform DLCN requires to convert the feature volume into $h_n\times w_n$ location-specific filters and then apply depth-wise and local convolutions to the feature maps, where $h_n$ and $w_n$ are the height and width of the feature maps at layer $n$. This implementation would be time-consuming as it ignores the redundant computations in neighboring pixels. To reduce the time cost, we employ the shift and element-wise product operators, in which shift [@wu2018shift] is a zero-flop zero-parameter operation, and element-wise product requires little calculation. Concretely, let $I_n \in \mathbb{R}^{h_n\times w_n\times c_n}$ and $D_n \in \mathbb{R}^{h_n\times w_n\times c_n}$ be the output of the feature extraction network and filter generation network, respectively, where $n$ is the index of the block (note that block $n$ corresponds to the layer $conv_{n+1}$ in ResNet). Let $k$ denote the kernel size of the feature extraction network. By defining a shifting grid $\{(g_i,g_j)\}, g \in (int)[1-k/2,k/2-1]$ that contains $k \cdot k$ elements, for every vector $(g_i,g_j)$, we shift the whole feature map $D$ towards the direction and step size indicated by $(g_i,g_j)$ and get the result $D^{(g_i,g_j)}$. For example, $g \in \{-1,0,1\}$ when $k=3$, and the feature map is moved towards nine directions with a horizontal or vertical step size of 0 or 1. We then use the sum and element-wise product operations to compute our filtering result: $$\begin{aligned} I' = I \odot \frac{1}{k \cdot k}\sum_{g_i,g_j}D^{(g_i,g_j)}. \label{eq:shift}\end{aligned}$$ To encourage information flow between channels of the depth-wise convolution, we further introduce a novel shift-pooling operator in the module. Considering $n_f$ as the number of channels with information flow, we shift the feature maps along the channel axis for $n_f$ times by ${1,2,..,n_f-1}$ to obtain new $n_f-1$ shifted feature maps $I^{(n_i)}_{s},n_i \in \{1,2,...,n_f-1\}$. Then we perform element-wise mean to the shifted feature maps and the original $I$ to obtain the new feature map as the input of the module. The process of this shift-pooling operation is shown in Figure \[fig:group\] ($n_f=3$). Compared to the idea ‘group’ of depth-wise convolution in [@howard2017mobilenets; @zhang2018shufflenet] which aims to group many channels into a group to perform information fusion between them, the proposed shift-pooling operator is more efficient and adds no additional parameters to the convolution. The size of our convolutional weights of each local kernel is always $k \times k \times c_n$ when applying shift-pooling, while it changes significantly in [@howard2017mobilenets] for different number of groups from $k \times k \times c_n$ to $k \times k \times c_n \times c_n$ in group convolution (assume that the convolution keeps the number of channels unchanged). Note that it is difficult for the filter generation network to generate so many kernels for the traditional convolutions $\mathcal{F}$ between all channels, and the characteristic of being position-specific dramatically increases their computational cost. With our depth-wise formulation, different kernels can have different functions. This enables us to assign different dilation rates [@yu2015multi] for each filter to address the scale-sensitive problem. Since there are huge intra-class and inter-class scale differences in an RGB image, we use $I$ to learn an adaptive dilation rate for each filter to obtain different sizes of receptive fields by an adaptive function $\mathcal{A}$. Specifically, let $d$ denote our maximum dilation rate, the adaptive function $\mathcal{A}$ consists of three layers: (1) an AdaptiveMaxPool2d layer with the output size of $d\times d$ and channel number $c$; (2) a convolutional layer with a kernel size of $d \times d$ and channel number $d \times c$; (3) a reshape and softmax layer to generate $d$ weights $\mathcal{A}^{w}(I), w \in (int)[1,d]$ with a sum of 1 for each filter. Formally, our guided filtering with adaptive dilated function (D$^4$LCN) is formulated as follows: $$\begin{aligned} I' = \frac{1}{d \cdot k \cdot k} \cdot I \odot \sum_w{\mathcal{A}^{w}(I)\sum_{g_i,g_j}D^{(g_i*w,g_j*w)}}, \label{eq:adaptive}\end{aligned}$$ For different images, our depth-guided filtering module assigns different kernels on different pixels and adaptive receptive fields (dilation) on different channels. This solves the problem of scale-sensitive and meaningless local structure of 2D convolutions, and also makes full use of RGB information compared to pseudo-LiDAR representations. 2D-3D Detection Head -------------------- In this work, we adopt a single-stage detector with prior-based 2D-3D anchor boxes [@redmon2016you; @liu2016ssd] as our base detector. ### Formulation **Inputs:** The output feature map $I_4 \in \mathbb{R}^{h_4 \times w_4}$ of our backbone network with a network stride factor of 16. Following common practice, we use a calibrated setting which assumes that per-image camera intrinsics $K \in \mathbb{R}^{3 \times 4}$ are available both at the training and test time. The 3D-to-2D projection can be written as: $$\eqnvspace \begin{bmatrix} x \\ y \\ 1 \\ \end{bmatrix}_{P} \cdot z_{3D} = K \cdot \begin{bmatrix} ~x~ \\ ~y~ \\ ~z~ \\ ~1~ \\ \end{bmatrix}_{3D} \eqnvspace \label{eqn:proj}$$ where $[x,y,z]_{3D}$ denotes the horizontal position, height and depth of the 3D point in camera coordinates, and $[x,y]_P$ is the projection of the 3D point in 2D image coordinates. **Ground Truth:** We define a ground truth (GT) box using the following parameters: the 2D bounding box $[x,y,w,h]_{2D}$, where $(x,y)$ is the center of 2D box and $w,h$ are the width and height of 2D box; the 3D center $[x,y,z]_{3D}$ represents the location of 3D center in camera coordinates; the 3D shapes $[w,h,l]_{3D}$ (3D object dimensions: height, width, length (in meters)), and the allocentric pose $\alpha_{3D}$ in 3D space (observation angle of object, ranging $[-\pi,\pi]$) [@manhardt2019roi]. Note that we use the minimum enclosing rectangle of the projected 3D box as our ground truth 2D bounding box. **Outputs:** Let $n_a$ denote the number of anchors and $n_c$ denote the number of classes. For each position $(i,j)$ of the input, the output for an anchor contains $35+n_c$ parameters: $\{[t_x,t_y,t_w,t_h]_{2D}, [t_x,t_y]_{P}, [t_z,t_w,t_h,t_l,t_{\alpha}]_{3D}$, $t_C^{(m)}, \mathbf{s}\}$, where $[t_x,t_y,t_w,t_h]_{2D}$ is the predicted 2D box; $[t_x,t_y]_{P}$ is the position of the projected 3D corner in the 2D plane, $[t_z,t_w,t_h,t_l,t_{\alpha}]_{3D}$ denotes the depth, predicted 3D shape and rotation, respectively; $t_C^{(m)} = \{[t_x^{(m)}, t_y^{(m)}]_{P}, [t_z^{(m)}]_{3D}\}, m \in \{1,2,...,8\}$ denotes 8 projected 3D corners; $\mathbf{s}$ denotes the classification score of each class. The size of the output is $h_4 \times w_4 \times n_a \times (35+n_c)$, where $(h_4,w_4)$ is the size of the input image with a down sampling factor of 16. The output is actually an anchor-based transformation of the 2D-3D box. ![An example of our shift-pooling operator of depth-wise convolution in depth-guided filtering module when $n_f$ is 3. It is efficiently implemented by shift and element-wise mean operators.[]{data-label="fig:group"}](group.pdf){width="0.7\columnwidth"} ### 2D-3D Anchor Inspired by [@brazil2019m3d], we utilize 2D-3D anchors with priors as our default anchor boxes. More specifically, a 2D-3D anchor is first defined on the 2D space as in [@liu2016ssd] and then use the corresponding priors in the training dataset to calculate the part of it in the 3D space. One template anchor is defined using parameters of both spaces: $\{[A_x, A_y, A_w, A_h]_{2D}, [A_z, A_w, A_h, A_l, A_{\alpha}]_{3D}\}$, where $[A_z, A_w, A_h, A_l, A_{\alpha}]_{3D}$ denotes the 3D anchor (depth, shape, rotation). For 2D anchors $[A_x, A_y, A_w, A_h]_{2D}$, we use 12 different scales ranging from 30 to 400 pixels in height following the power function of $30 * 1.265^{exp}, exp \in (int)[0,11]$ and aspect ratios of $[0.5, 1.0, 1.5]$ to define a total of 36 anchors. We then project all ground truth 3D boxes to the 2D space. For each projected box, we calculate its intersection over union (IoU) with each 2D anchor and assign the corresponding 3D box to the anchors that have an IoU $\geq 0.5$. For each 2D anchor, we thus use the statistics across all matching ground truth 3D boxes as its corresponding 3D anchor $[A_z, A_w, A_h, A_l, A_{\alpha}]_{3D}$. Note that we use the same anchor parameters $[A_x, A_y]_{2D}$ for the regression of $[t_x,t_y]_{2D}$ and $[t_x,t_y]_{P}$. The anchors enable our network to learn a relative value (residual) of the ground truth, which significantly reduces the difficulty of learning. ### Data Transformation We combine the output of our network which is an anchor-based transformation of the 2D-3D box and the pre-defined anchors to obtain our estimated 3D boxes: $$\begin{aligned} & [x',y']_{2D} = [A_x,A_y]_{2D} + [t_x,t_y]_{2D} * [A_w, A_h]_{2D} \notag \\ & [x',y']_{P} = [A_x,A_y]_{2D} + [t_x,t_y]_{P} * [A_w, A_h]_{2D} \notag \\ & [x'^{(m)}, y'^{(m)}]_{P} = [A_x,A_y]_{2D} + [t_x^{(m)}, t_y^{(m)}]_{P} * [A_w, A_h]_{2D} \notag \\ & [w', h']_{2D} = [A_w,A_h]_{2D} \cdot \exp([t_w,t_h]_{2D}) \notag \\ & [w', h', l']_{3D} = [A_w,A_h,A_l]_{3D} \cdot \exp([t_w,t_h,t_l]_{3D}) \notag \\ & [z', z'^{(m)},\alpha']_{3D} = [A_z, A_z, A_{\alpha}] + [t_z, t_z, t_{alpha}]_{3D}. \label{eq:tranform}\end{aligned}$$ where $[x',y']_{P}, [z', z'^{(m)},\alpha']_{3D}$ denote respectively the estimated 3D center projection in 2D plane, the depth of 3D center and eight corners, the 3D rotation by combining output of the network and the anchor. ### Losses Our overall loss contains a classification loss, a 2D regression loss, a 3D regression loss and a 2D-3D corner loss. We use the idea of focal loss [@lin2017focal] to balance the samples. Let $s_t$ and $\gamma$ denote the classification score of target class and the focusing parameter, respectively. We have: $$\begin{aligned} & L = (1-s_t)^{\gamma}(L_{class} + L_{2d} + L_{3d} + L_{corner}), \label{eq:loss}\end{aligned}$$ where $\gamma = 0.5$ in all experiments, and $L_{class}$, $L_{2d}$, $L_{3d}$, $L_{corner}$ are the classification loss, 2D regression loss, 3D regression loss and D-3D corner loss, respectively. In this work, we employ the standard cross-entropy (CE) loss for classification: $$\begin{aligned} L_{class} = -\log(s_t). \label{eq:loss_classification}\end{aligned}$$ Moreover, for both 2D and 3D regression, we simply use the SmoothL1 regression losses: $$\begin{aligned} & L_{2D} = SmoothL1([x',y',w',h']_{2D}, [x,y,w,h]_{2D}), \notag \\ & L_{3D} = SmoothL1([w',h',l',z',\alpha']_{3D}, [w,h,l,z,\alpha]_{3D}), \notag \\ & ~~~~~~~~ + SmoothL1([x',y']_{P}, [x,y]_{P}), \notag \\ & L_{corner} = \frac{1}{8} \sum SmoothL1([x'^{(m)}, y'^{(m)}]_{P}, [x^{(m)},y^{(m)}]_P) \notag \\ & ~~~~~~~~ + SmoothL1([z'^{(m)}]_{3D}, [z]_{3D}), \label{eq:loss_regression}\end{aligned}$$ where $[x^{(m)},y^{(m)}]_P$ denotes the projected corners in image coordinates of the GT 3D box and $[z]_{3D}$ is its GT depth. Experiments =========== Dataset and Setting ------------------- **KITTI Dataset.**  The KITTI 3D object detection dataset [@geiger2012we] is widely used for monocular and LiDAR-based 3D detection. It consists of 7,481 training images and 7,518 test images as well as the corresponding point clouds and the calibration parameters, comprising a total of 80,256 2D-3D labeled objects with three object classes: Car, Pedestrian, and Cyclist. Each 3D ground truth box is assigned to one out of three difficulty classes (easy, moderate, hard) according to the occlusion and truncation levels of objects. There are two train-val splits of KITTI: the split1 [@chen20153d] contains 3,712 training and 3,769 validation images, while the split2 [@xiang2015data] uses 3,682 images for training and 3,799 images for validation. The dataset includes three tasks: 2D detection, 3D detection, and Bird’s eye view, among which 3D detection is the focus of 3D detection methods. **Evaluation Metrics.**  Precision-recall curves are used for evaluation (with the IoU threshold of 0.7). Prior to Aug. 2019, 11-point Interpolated Average Precision (AP) metric $\text{AP}|_{R_{11}}$ proposed in the Pascal VOC benchmark is separately computed on each difficulty class and each object class. After that, the 40 recall positions-based metric $\text{AP}|_{R_{40}}$ is used instead of $\text{AP}|_{R_{11}}$, following [@simonelli2019disentangling]. All methods are ranked by $\text{AP}|_{R_{11}}$ of the 3D car detection in the moderate setting. **Implementation Details.**  We use our depth-guided filtering module three times on the first three blocks of ResNet, which have different network strides of 4,8,16, respectively. [@fu2018deep] is used for depth estimation. A drop-channel layer with a drop rate of 0.2 is used after each module and a dropout layer with a drop rate of 0.5 is used after the output of the network backbone. For our single-stage detector, we use two convolutional layers as our detection head. The number of channels in the first layer is 512, and $n_a * (35 + n_c)$ for the second layer, where $n_c$ is set to 4 for three object classes and the background class, and $n_a$ is set to 36. Non Maximum Suppression (NMS) with an IoU threshold of 0.4 is used on the network output in 2D space. Since the regression of the 3D rotation $\alpha$ is more difficult than other parameters, a hill-climbing post-processing step is used for optimizing $\alpha$ as in [@brazil2019m3d]. The input images are scaled to $512 \times 1760$ and horizontal flipping is the only data augmentation. $n_f$ is set to 2 and the maximum dilation rate $d$ is set to 3 in all experiments. The network is optimized by stochastic gradient descent (SGD), with a momentum of 0.9 and a weight decay of 0.0005. We take a mini-batch size of 8 on 4 Nvidia Tesla v100 GPUs (16G). We use the ‘poly’ learning rate policy and set the base learning rate to 0.01 and power to 0.9. The iteration number for the training process is set to 40,000. Comparative Results ------------------- We conduct experiments on the official test set and two splits of validation set of the KITTI dataset. Table \[tab:testing\] includes the top 14 monocular methods in the leaderboard, among which our method ranks top-1. We can observe that: (1) Our method outperforms the second-best competitor for monocular 3D car detection by a large margin (relatively 9.1% for 10.74 vs. 11.72) under the moderate setting (which is the most important setting of KITTI). (2) Most competitors, such as [@ku2019monocular; @ma2019accurate; @simonelli2019disentangling; @naiden2019shift; @weng2019monocular; @brazil2019m3d], utilize the detector (e.g. Faster-RCNN) pre-trained on COCO/KITTI or resort to multi-stage training to obtain better 2D detection and stable 3D results, while our model is trained end-to-end using the standard ImageNet pre-trained model. However, we still achieve the state-of-the-art 3D detection results, validating the effectiveness of our D$^4$LCN to learn 3D structure. (3) Recently KITTI uses $\text{AP}|_{R_{40}}$ instead of $\text{AP}|_{R_{11}}$, however, all existing methods report the results under the old metric. We thus also give results under $\text{AP}|_{R_{11}}$ on the validation set for fair comparison. It can be seen that our method outperforms all others on the two splits for 3D car detection. Our results under $\text{AP}|_{R_{40}}$ on validation set are shown in ablation study. Detailed Analysis ----------------- ### Ablation Study To conduct ablation study on our model, we make comparison among five versions of our model: (1) 3DNet: the baseline model using $L_{2D}$ and $L_{3D}$ without our depth-guided filtering module; (2) + CL: the Corner Loss is added to 3DNet; (3) + DLCN: depth-guided depth-wise local filtering is added; (4) + SP: shift-pooling operator is added (with $n_f=3$); (5) D$^4$LCN (our full model): adaptive dilation rates are added, as in Eq. \[eq:adaptive\]. From Table \[tab:ablition\], we can observe that: (1) The performance continuously increases when more components are used for 3D object detection, showing the contribution of each component. (2) Our depth-guided filtering module increases the 3D detection AP scores (moderate) from {15.57, 12.09} to {21.71, 16.20} w.r.t. the $\text{AP}|_{R_{11}}$ and $\text{AP}|_{R_{40}}$ metrics, respectively. This suggests that it is indeed effective to capture the meaningful local structure for 3D object detection. (3) The main improvement comes from our adaptive dilated convolution (2.69 and 1.76 for $\text{AP}|_{R_{11}}$ and $\text{AP}|_{R_{40}}$, respectively), which allows each channel of the feature map to have different receptive fields and thus solves the scale-sensitive problem. Note that we have tried different values of $n_f \in \{1,2,3,4,5,6\}$, and found that $n_f=3$ is the best. ### Evaluation of Depth Maps To study the impact of accuracy of depth maps on the performance of our method, we extract depth maps using four different methods [@monodepth17; @fu2018deep; @mayer2016large; @chang2018pyramid] and then apply them to 3D detection. As reported in previous works on depth estimation, the three supervised methods (PSMNet, DispNet, and DORN) significantly outperform the unsupervised method [@monodepth17]. Among the supervised methods, Stereo-based methods [@chang2018pyramid; @mayer2016large] are better than monocular-based DORN. With these conclusions, we have the following observations from Table \[tab:depth\]: (1) The accuracy of 3D detection is higher with better depth map. This is because that better depth map can provide better scene geometry and local structure. (2) As the quality of depth map increases, the growth of detection accuracy becomes slower. (3) Even with the depth maps obtained by unsupervised learning [@monodepth17], our method achieves state-of-the-art results. Compared to the pseudo-lidar based method [@ma2019accurate], our method relies less on the quality of depth maps (19.63 vs. 15.45 using MonoDepth). ### Evaluation of Convolutional Appoaches To show the effectiveness of our guided filtering module for 3D object detection, we compare it with several alternatives: Dynamic Convolution [@jia2016dynamic], Dynamic Local Filtering [@jia2016dynamic], and Deformable Convolution [@dai2017deformable]. Our method belongs to dynamic networks but yields less computation cost and stronger representation. For the first two methods, we conduct experiments using the same depth map as ours. For the third method, we apply deformable convolution on both RGB and depth branches and merge them by element-wise product. From Table \[tab:conv\], we can observe that our method performs the best. This indicates that our method can better capture 3D information from RGB images due to the special design of our D$^4$LCN. ![Visualization of active maps corresponding to different filters of block 3 of our D$^4$LCN. Each filter learns three weights representing dilation rate of 1, 2, 3, respectively. Different filters have different functions in our model to handle the scale problem adaptively. For example, filter 89 has large receptive fields for large-scale cars, while filter 70 deals with the small-scale cars.[]{data-label="fig:filter"}](filter.pdf){width="0.97\columnwidth"} ### Multi-Class 3D Detection Since a person is a non-rigid body, its shape varies and its depth information is hard to accurately estimate. For this reason, 3D detection of pedestrians and cyclists becomes particularly difficult. Note that all pseudo-LiDAR based methods [@ma2019accurate; @weng2019monocular; @wang2019pseudo] fail to detect these two categories. However, as shown in Table \[tab:multiclass\], our method still achieves satisfactory performance on 3D detection of pedestrians and cyclists. Moreover, we also show the active maps corresponding to different filters of our D$^4$LCN in Figure \[fig:filter\]. Different filters on the same layer of our model use different sizes of receptive fields to handle objects of different scales, including pedestrians (small) and cars (big), as well as distant cars (big) and nearby cars (small). Conclusion ========== In this paper, we propose a Depth-guided Dynamic-Depthwise-Dilated Local ConvNet (D$^4$LCN) for monocular 3D objection detection, where the convolutional kernels and their receptive fields (dilation rates) are different for different pixels and channels of different images. These kernels are generated dynamically conditioned on the depth map to compensate the limitations of 2D convolution and narrow the gap between 2D convolutions and the point cloud-based 3D operators. As a result, our D$^4$LCN can not only address the problem of the scale-sensitive and meaningless local structure of 2D convolutions, but also benefit from the high-level semantic information from RGB images. Extensive experiments show that our D$^4$LCN better captures 3D information and ranks $1^{\mathrm{st}}$ for monocular 3D object detection on the KITTI dataset at the time of submission. Acknowledgements ================ We would like to thank Dr. Guorun Yang for his careful proofreading. Ping Luo is partially supported by the HKU Seed Funding for Basic Research and SenseTime’s Donation for Basic Research. Zhiwu Lu is partially supported by National Natural Science Foundation of China (61976220, 61832017, and 61573363), and Beijing Outstanding Young Scientist Program (BJJWZYJH012019100020098). **APPENDIX** Definition of 3D Corners ======================== We define the eight corners of each ground truth box as follows: $$\begin{aligned} C^{(m)} = \begin{bmatrix} x^{(m)} \\ y^{(m)} \\ 1 \\ \end{bmatrix}_{P} \hspace{-0.05in}\cdot z^{(m)}_{3D} = \begin{pmatrix} r_y \cdot \begin{bmatrix} \pm w/2 \\ \pm h/2 \\ \pm l/2 \\ 0 \end{bmatrix}_{3D} \hspace{-0.05in}+ \begin{bmatrix} x \\ y \\ z \\ 1\end{bmatrix}_{3D} \end{pmatrix} \end{aligned}$$ where $m \in (int)[1,8]$ in a defined order, and $r_y$ is the egocentric rotation matrix. Note that we use allocentric pose for regression. Comparisons between Two Rotation Definitions ============================================ As shown in Figure \[fig:definition\], while egocentric poses undergo viewpoint changes towards the camera when translated, allocentric poses always exhibit the same view, independent of the object’s location. The allocentric pose $\alpha$ and the egocentric pose $r_y$ can be converted to each other according to the viewing angle $\theta$. $$\begin{aligned} \alpha = r_y - \theta \label{eq:alpha}\end{aligned}$$ ![Comparisons between egocentric ($r_y$) and allocentric ($\alpha$) poses. The car1 and car2 have the same egocentric pose, but they are observed on different sides (views). We use allocentric pose to keep the same view (car1 and car3).[]{data-label="fig:definition"}](rotation.pdf){width="1\columnwidth"} =0.13cm [|c|c|c|c|c|c|c|c|c|]{} Method & Depth & CAD & Points & Freespace & Segmentation & Pretrain/MST & End-to-end\ ------------------------------------------------------------------ Deep3DBox [@mousavian20173d], GS3D [@li2019gs3d], MonoGRNet [@qin2019monogrnet], OFTNet [@roddick2018orthographic] FQNet [@liu2019deep], SS3D [@jorgensen2019monocular] ------------------------------------------------------------------ & & & & & & &\ ROI-10D [@manhardt2019roi] & & & & & & &\ Multi-Level Fusion [@xu2018multi], **D**$\mathbf{^4}$**LCN (Ours)** & & & & & & &\ ------------------------------------------------------------------ M3D-RPN [@brazil2019m3d], MONODIS [@simonelli2019disentangling], Shift R-CNN [@naiden2019shift] ------------------------------------------------------------------ & & & & & & &\ --------------------------------------------------------------- AM3D [@ma2019accurate] Pseudo-LiDAR [@wang2019pseudo], Mono3D-PLiDAR[@weng2019monocular], MonoPSR [@ku2019monocular] --------------------------------------------------------------- & & & & & & &\ Deep-MANTA [@chabot2017deep] & & & & & &&\ 3DVP [@xiang2015data] & & & & & & &\ Mono3D [@chen2016monocular; @chen20173d] & & & & & & &\ Mono3D++ [@he2019mono3d++] & & & & & & &\ Ablative Results for Convolutional Methods ========================================== The Depth-guided filtering module in our D$^4$LCN model can be decomposed into basic convolutional components: - Traditional Convolutional Network - Depth-guided ConvNet (CN) - Depth-guided Local CN (LCN) - Depth-guided Depth-wise LCN (DLCN) - Depth-guided DLCN with Shift-pooling (SP-DLCN) - D$^4$LCN (Our full model) The ablative results for these convolutional methods are shown in Table \[tab:ablation\]. We can observe that: (1) Using the depth map to guide the convolution of each pixel brings a considerable improvement. (2) Depth-wise convolution with shift-pooling operator not only has fewer parameters (Section 3.2 of our main paper) but also gets better performance than the standard convolution. (3) The main improvement comes from our adaptive dilated convolution, which allows each channel of the feature map to have different receptive fields. Comparisons of Labeling Information and Training Strategies =========================================================== We compare the labeling information and training strategies used in different monocular detection methods, as shown in Table \[tab:methods\]. It can be seen that: (1) our model outperforms all existing methods by only using the depth map extracted from the monocular image. (2) our model can be trained in an end-to-end manner. Distributions of Different Dilation =================================== We show the average ratio of different channels with different dilation rates in three blocks of our model over the validation set of split1 (Figure \[fig:probability\]). It can be seen that: (1) For the first block with insufficient receptive field, the model tends to increase the receptive field by large dilation rate, and then it uses small receptive field for the second block. (2) In the third block, the model uses three different dilation rates evenly to deal with the object detection of different scales. We also show the active maps corresponding to different filters of the third block of our D$^4$LCN in our main paper (Figure 5). ![The average ratio of different channels with different dilation rates in three blocks.[]{data-label="fig:probability"}](dilation.pdf){width="1\columnwidth"}

ArXiv

--- abstract: 'Let $({{\mathcal{A}}}, \alpha)$ and $({{\mathcal{B}}}, \beta)$ be C\*-dynamical systems and assume that ${{\mathcal{A}}}$ is a separable simple C\*-algebra and that $\alpha$ and $\beta$ are $*$-automorphisms. Then the semicrossed products ${{\mathcal{A}}}\times_{\alpha} {{\mathbb{Z}}}^{+}$ and ${{\mathcal{B}}}\times_{\beta} {{\mathbb{Z}}}^{+}$ are isometrically isomorphic if and only if the dynamical systems $({{\mathcal{A}}}, \alpha)$ and $({{\mathcal{B}}}, \beta)$ are outer conjugate.' address: - | Pure Math. Dept.\ U. Waterloo\ Waterloo, ON N2L–3G1\ CANADA - | Dept. Mathematics\ East Carolina U.\ Greenville, NC 27858\ USA author: - 'Kenneth R. Davidson' - 'Elias G. Katsoulis' title: 'Semicrossed Products of Simple C\*-algebras.' --- [^1] [^2] [^3] introduction ============ The main objective of this paper is the classification of semicrossed products of separable simple C\*-algebras by an automorphism, up to isometric isomorphism. It is easily seen (and well-known) that outer conjugacy between automorphisms of arbitrary C\* algebras is a sufficient condition for the existence of an isometric isomorphism between the associated semicrossed products. In this paper we show that for separable simple C\*-algebras, this is also a necessary condition. This follows from the following general result: if $\alpha$ and $\beta$ are automorphisms of arbitrary C\*-algebras ${{\mathcal{A}}}$ and ${{\mathcal{B}}}$, then the presence of an isometric isomorphism from ${{\mathcal{A}}}\times_{\alpha} {{\mathbb{Z}}}^{+}$ onto ${{\mathcal{B}}}\times_{\beta} {{\mathbb{Z}}}^{+}$ implies the existence of a C\*-isomorphism $\gamma: {{\mathcal{A}}}\rightarrow {{\mathcal{B}}}$ so that $\alpha \circ \gamma^{-1}\circ \beta^{-1}\circ \gamma$ is universally weakly inner with respect to irreducible representations. The result for simple C\*-algebras follows then from a remarkable result of Kishimoto [@Kis] which shows that for a separable simple C\*-algebra, all universally weakly inner automorphisms are actually inner. Let $({{\mathcal{A}}}, \alpha)$ be a (discrete) C\*-dynamical system, i.e., a C\*-algebra ${{\mathcal{A}}}$ together with a $*$-endomorphism $\alpha$ of ${{\mathcal{A}}}$. Motivated by a construction of Arveson [@Arv], Peters [@Pet] introduced the concept of the semicrossed product ${{\mathcal{A}}}\times_{\alpha} {{\mathbb{Z}}}^{+}$. This is the universal operator algebra for contractive covariant representations of this system. In the commutative case, the semicrossed product is an algebra ${{\mathrm{C}}}_0({{\mathcal{X}}}) \times_{\sigma}{{\mathbb{Z}}}^+$ determined by a dynamical system $({{\mathcal{X}}}, \sigma)$ given by a proper continuous map $\sigma$ acting on a locally compact Hausdorff space $X$. Under the assumption that the topological spaces are compact and the maps are aperiodic, Peters [@Pet] showed that two such semicrossed products are isomorphic as algebras if and only if the corresponding dynamical systems are conjugate, thus extending an earlier classification scheme of Arveson [@Arv] and Arveson and Josephson [@ArvJ]. In spite of the subsequent interest in semicrossed products and their variants [@AlaP; @Bu; @BP; @HPW; @Lam-1; @Lam; @MM; @OS; @Pow; @Solel-1], the problem of classifying semicrossed products of the form ${{\mathrm{C}}}_0({{\mathcal{X}}}) \times_{\sigma}{{\mathbb{Z}}}^+$ remained open in the generality introduced by Peters in [@Pet] until our recent paper [@DavK], which established that the Arveson-Josephson-Peters classification scheme holds with no restrictions on either ${{\mathcal{X}}}$ or $\sigma$. It was our desire to apply the techniques of [@DavK] and [@DK2] to more general settings that motivated the research of the present paper. The present paper provides for the first time a classification scheme for semicrossed products which is valid for a broad class of C\*-algebras, without posing any restrictions on the automorphisms involved. Our result complements a similar result of Muhly and Solel [@MS Theorem 4.1] regarding semicrossed products with automorphisms having full Connes spectrum. In Theorem \[MS\], we give an alternative proof of their result using representation theory. Both results seem to indicate that outer conjugacy is a complete invariant for isometric isomorphisms between arbitrary semicrossed products. They also suggest the problem of establishing the validity of the conclusion under the weaker requirement of an *algebraic* isomorphism instead of an isometric isomorphism. Preliminaries ============= Let ${{\mathcal{A}}}$ be a C\*-algebra and $\alpha$ an endomorphism of ${{\mathcal{A}}}$. The *skew polynomial algebra* $P({{\mathcal{A}}}, \alpha)$ consists of all polynomials of the form $\sum_n U_{\alpha}^n A_n$, $A_n\in {{\mathcal{A}}}$, where the multiplication of the “coefficients” $A\in {{\mathcal{A}}}$ with the “variable” $U_{\alpha}$ obeys the rule $$AU_{\alpha} = U_{\alpha}\alpha(A)$$ Equip $P({{\mathcal{A}}}, \alpha)$ with the $l^1$-norm $$\big\| \sum_n U_{\alpha}^n A_n \big\|_1 \equiv \sum_n \| A_n\|$$ and let $l^1({{\mathcal{A}}}, \alpha)$ be the completion of $P({{\mathcal{A}}}, \alpha)$ with respect to $\|\, .\,\|_1$. An *(isometric) covariant representation* $(\pi, V)$ of $({{\mathcal{A}}}, \alpha)$ consists of a C\*-representation $\pi$ of ${{\mathcal{A}}}$ on a Hilbert space ${{\mathcal{H}}}$ and an isometry $V$ on ${{\mathcal{H}}}$ so that $\pi(A)V= V\pi(\alpha(A))$, for all $A \in {{\mathcal{A}}}$. Each covariant representation induces a representation $\pi\times V$ in an obvious way. For $P \in l^1({{\mathcal{A}}}, \alpha)$ let $$\|P\| := \sup \big\{ \|(\pi\times V) (P)\| : \pi\times V\ \mbox{ is covariant} \big\}$$ where $\pi\times V$ runs over all isometric covariant representations of $({{\mathcal{A}}}, \alpha)$. The *semicrossed product* ${{\mathcal{A}}}\times_{\alpha} {{\mathbb{Z}}}^{+}$ of ${{\mathcal{A}}}$ by $\alpha$ is the completion of $l^1({{\mathcal{A}}}, \alpha)$ with respect to this norm. For an alternative description of ${{\mathcal{A}}}\times_{\alpha} {{\mathbb{Z}}}^{+}$, one may start by obtaining a universal covariant representation $(\pi\times V)$ of $({{\mathcal{A}}}, \alpha) $ and then define ${{\mathcal{A}}}\times_{\alpha} {{\mathbb{Z}}}^{+}$ to be the non-sefadjoint operator algebra generated $\pi({{\mathcal{A}}})$ and $V$. The two constructions produce isomorphic algebras. For each covariant representation $(\pi, V)$ of $({{\mathcal{A}}}, \alpha)$, the representation $\pi\times V$ extends uniquely to a contractive representation of ${{\mathcal{A}}}\times_{\alpha} {{\mathbb{Z}}}^{+}$, which will also be denoted as $\pi\times V$. In this paper, we will exclusively work with *invertible* C\*-dynamical systems, i.e., the endomorphism will actually be an automorphism. Therefore we now drop the adjective “invertible”, and by C\*-dynamical system we will mean an invertible one. Given an (invertible) dynamical system $({{\mathcal{A}}}, \alpha)$, the unitary covariant representations for $({{\mathcal{A}}}, \alpha)$ suffice to capture the norm for ${{\mathcal{A}}}\times_{\alpha} {{\mathbb{Z}}}^{+}$. Therefore, ${{\mathcal{A}}}\times_{\alpha} {{\mathbb{Z}}}^{+}$ is a natural nonselfadjoint subalgebra of the crossed product C\*-algebra ${{\mathcal{A}}}\times_{\alpha} {{\mathbb{Z}}}$. Two C\*-dynamical systems $({{\mathcal{A}}}, \alpha)$ and $({{\mathcal{B}}}, \beta)$ are said to be *outer conjugate* if there exists a C\*-isomorphism $\gamma : {{\mathcal{A}}}\rightarrow {{\mathcal{B}}}$ and a unitary $W \in M({{\mathcal{A}}})$, the multiplier algebra of ${{\mathcal{A}}}$, so that $$\alpha = {\operatorname{ad}}_W \gamma^{-1} \circ \beta \circ \gamma .$$ The main issue in this paper is the classification of semicrossed products up to isometric isomorphism. The following elementary result shows that the outer conjugacy of automorphisms provides a sufficient condition for the existence of such an isomorphism. \[onedir\] If the dynamical systems $({{\mathcal{A}}}, \alpha)$ and $({{\mathcal{B}}}, \beta)$ are outer conjugate, then the semicrossed products ${{\mathcal{A}}}\times_{\alpha} {{\mathbb{Z}}}^{+}$ and ${{\mathcal{B}}}\times_{\beta} {{\mathbb{Z}}}^{+}$ are isometrically isomorphic. Without loss of generality assume that ${{\mathcal{A}}}= {{\mathcal{B}}}$ and $\gamma= {{\operatorname{id}}}$. Let $W \in M({{\mathcal{A}}})$ so that $\alpha(A) = W\beta(A)W^*$. Observe that $\beta$ has a unique extension to an automorphism $\bar\beta$ of $M({{\mathcal{A}}})$ such that $$\bar\beta(M) \beta(A) = \beta(MA) {\quad\text{for all}\quad}A \in {{\mathcal{A}}}{\text{ and }}M \in M({{\mathcal{A}}}) ,$$ namely, $\bar\beta(M) A = \beta(M\beta^{-1}(A))$. Therefore ${{\mathcal{A}}}\times_\beta {{\mathbb{Z}}}$ is naturally a subalgebra of $M({{\mathcal{A}}}) \times_{\bar\beta} {{\mathbb{Z}}}$ generated by ${{\mathcal{A}}}$ and the universal unitary $U_\beta$ satisfying $U_\beta MU_\beta^* = \beta(M)$ for $M \in M({{\mathcal{A}}})$. Now notice that ${\operatorname{ad}}_{WU_\beta}$ implements $\bar\alpha$, the extension of $\alpha$ to $M({{\mathcal{A}}})$ because it is an automorphism which acts as $\alpha$ on ${{\mathcal{A}}}$. Whence it carries $M({{\mathcal{A}}})$ to itself, and is the unique extension of $\alpha$ to $M({{\mathcal{A}}})$. Therefore ${\mathrm{C}^*}({{\mathcal{A}}}, WU_\beta)$ determines a representation $\sigma$ of the C\*-algebra crossed product ${{\mathcal{A}}}\times_\alpha {{\mathbb{Z}}}$ given by $\sigma|_{{\mathcal{A}}}= {{\operatorname{id}}}$ and $\sigma(U_\alpha) = U_\beta W$. Since this representation is faithful on ${{\mathcal{A}}}$, it follows from the gauge invariance uniqueness theorem [@Kat Theorem 6.4] that this is a faithful representation of the crossed product. Next observe that ${\mathrm{C}^*}({{\mathcal{A}}}, WU_\beta) = {{\mathcal{A}}}\times_\beta {{\mathbb{Z}}}$. The point is that $$\begin{aligned} A (U_\beta W)^n &= A (U_\beta W U_\beta^*) (U_\beta^2 W U_\beta^{*2})\cdots (U_\beta^n W U_\beta^{*n}) U_\beta^n \\ &= A \bar\beta(W) \bar\beta^2(W) \cdots \bar\beta^n(W) U_\beta^n \\ &= B U_\beta^n\end{aligned}$$ where $B = A \bar\beta(W) \bar\beta^2(W) \cdots \bar\beta^n(W)$ belongs to ${{\mathcal{A}}}$. It now follows that ${{\mathcal{A}}}(U_\beta W)^n = {{\mathcal{A}}}U_\beta^n$. Hence ${\mathrm{C}^*}({{\mathcal{A}}}, WU_\beta) = {{\mathcal{A}}}\times_\beta {{\mathbb{Z}}}$. Moreover, the nonself-adjoint subalgebra ${{\mathcal{A}}}\times_\beta {{\mathbb{Z}}}^+$ generated by ${{\mathcal{A}}}$ and ${{\mathcal{A}}}U_\beta$ coincides with the algebra generated by ${{\mathcal{A}}}$ and ${{\mathcal{A}}}U_\beta W$. But this latter algebra is canonically identified with ${{\mathcal{A}}}\times_\alpha {{\mathbb{Z}}}^+$ via the identification of ${\mathrm{C}^*}({{\mathcal{A}}}, WU_\beta)$ with ${{\mathcal{A}}}\times_\alpha {{\mathbb{Z}}}$. Note that $({{\mathcal{A}}}, \alpha)$ and $({{\mathcal{B}}}, \beta)$ are outer conjugate if and only if there exists a C\*-isomorphism $\gamma : {{\mathcal{A}}}\rightarrow {{\mathcal{B}}}$ so that the automorphism $a\circ \gamma^{-1}\circ\beta^{-1}\circ\gamma$ is inner. A notion weaker than that of outer congucacy arises from the concept of a universally weakly inner automorphism. We say an automorphism $a$ of ${{\mathcal{A}}}$ is *universally weakly inner* with respect to irreducible (resp. faithful) representations, if for any irreducible (resp. faithful) representation $\pi$ of ${{\mathcal{A}}}$ on ${{\mathcal{H}}}$, there exists a unitary $W \in \pi({{\mathcal{A}}})^{''}$ so that $\pi(\alpha(A))=W^*\pi(A)W$, $A \in {{\mathcal{A}}}$. The concept of a universally weakly inner automorphism with respect to faithful representations (or $\pi$-inner automorphism) was introduced by Kadison and Ringrose [@KadR] and has been studied by various authors [@Elliott; @Lance]. Here we will be making use of universally weakly inner automorphisms with respect to *irreducible* representations. A direct integral decomposition argument shows that the two concepts coincide for type I C\*-algebras. Kishimoto [@Kis] has shown that for a separable simple C\*-algebra all universally weakly inner automorphisms with respect to irreducible representations are actually inner. Therefore the two concepts coincide there as well. We also need to recall several facts for the various concepts of spectrum from C\*-algebra theory. Let $({{\mathcal{A}}}, \alpha)$ be a dynamical system and let ${\operatorname{Prim}}A$ be the primitive ideal space of ${{\mathcal{A}}}$ equipped with the hull-kernel topology. For each cardinal $i$, we fix a Hilbert space ${{\mathcal{H}}}_i$ with dimension $i$. Then, ${{\operatorname{irred}}}({{\mathcal{A}}}, {{\mathcal{H}}}_i)$ is the collection of all irreducible representations on ${{\mathcal{H}}}_i$ and ${{\operatorname{irred}}}{{\mathcal{A}}}\equiv \cup_i\, {{\operatorname{irred}}}({{\mathcal{A}}}, {{\mathcal{H}}}_i)$. If $\rho \in {{\operatorname{irred}}}{{\mathcal{A}}}$, then $[\rho]$ denotes its equivalence class, with respect to unitary equivalence $\simeq$ between representations on the same Hilbert space. Let the *spectrum* of ${{\mathcal{A}}}$ be $$\hat{{{\mathcal{A}}}}\equiv \big\{ [\rho] \mid \rho \in {{\operatorname{irred}}}{{\mathcal{A}}}\big\}$$ and consider the canonical map $$\theta : \hat{{{\mathcal{A}}}} \longrightarrow {\operatorname{Prim}}{{\mathcal{A}}}: [\rho] \longrightarrow \ker\rho.$$ In what follows we always consider $\hat{{{\mathcal{A}}}}$ equiped with the smallest topology that makes $\theta$ continuous. For any C\*-isomorphism $\gamma: {{\mathcal{A}}}\rightarrow {{\mathcal{B}}}$, we define a map $\hat{\gamma}: \hat{{{\mathcal{A}}}} \rightarrow \hat{{{\mathcal{B}}}}$ between the corresponding spectra by the formula $\hat{\gamma}([\rho])= [\rho\circ \gamma]$. The following result is straightforward. \[theta\] Let ${{\mathcal{A}}}$ be a C\*-algebra and let ${{\mathcal{X}}}\subseteq \hat{{{\mathcal{A}}}}$ have empty interior. Then, $$\bigcap_{x \in \hat{{{\mathcal{A}}}}\setminus {{\mathcal{X}}}}\theta(x)=\{0\}.$$ Another notion of spectrum is the Connes spectrum. We do not give the precise definition but instead state the fact [@OPed Theorem 10.4] that for a separable C\*-algebra, $({{\mathcal{A}}}, \alpha)$ has full Connes spectrum if and only if there is a dense $\alpha$-invariant subset $\Delta_{\alpha} \subseteq \hat{{{\mathcal{A}}}}$ on which $\hat{\alpha}$ is freely acting. This is equivalent to the fact that the periodic points of $\hat{\alpha}$ with period $n$ has no interior for any $n \ge1$. The main result =============== We begin this section with a general result about isometric isomorphisms between arbitrary operator algebras which is well-known. \[diag\] Let $\phi: {{\mathfrak{A}}}\rightarrow {{\mathfrak{B}}}$ be an isometric isomorphism between operator algebras. Then $\phi({{\mathfrak{A}}}\cap {{\mathfrak{A}}}^*)= {{\mathfrak{B}}}\cap {{\mathfrak{B}}}^*$ and $\phi\vert_{{{\mathfrak{A}}}\cap {{\mathfrak{A}}}^*}$ is a C\*-isomorphism. The unitary operators in ${{\mathfrak{A}}}\cap {{\mathfrak{A}}}^*$ are characterized as the norm 1 elements $A \in {{\mathcal{A}}}$ so that $A^{-1} \in {{\mathfrak{A}}}$ and $\| A^{-1}\|=1$. From this it follows that $\phi$ maps the unitary group of ${{\mathfrak{A}}}\cap {{\mathfrak{A}}}^*$ onto the unitary group of ${{\mathfrak{B}}}\cap {{\mathfrak{B}}}^*$, and this proves the first assertion. The second follows from the fact that $\phi$ preserves inverses of unitaries in ${{\mathfrak{A}}}\cap {{\mathfrak{A}}}^*$ and hence adjoints. In order to prove the main result, we need to use representation theory. In light of Proposition \[diag\], it suffices to consider representations that preserve the diagonal. Hence, if ${{\mathfrak{A}}}$ is an operator algebra, then ${\operatorname{rep}}({{\mathfrak{A}}}, {{\mathcal{H}}})$ will denote the collection of all contractive representations of ${{\mathfrak{A}}}$ on ${{\mathcal{H}}}$ whose restriction on the diagonal ${{\mathfrak{A}}}\cap{{\mathfrak{A}}}^*$ is a $*$-homomorphism. \[nestrepn\] If ${{\mathfrak{A}}}$ is an operator algebra and ${{\mathcal{H}}}$ a Hilbert space, then ${\operatorname{rep}}_2({{\mathfrak{A}}}, {{\mathcal{H}}})$ will denote the collection of all (contractive) representations $\rho \in {\operatorname{rep}}({{\mathfrak{A}}}, {{\mathcal{H}}}\oplus {{\mathcal{H}}})$ of the form $$\rho(A)= \begin{bmatrix} \rho_1(A) & \rho_3(A) \\ 0 & \rho_2(A) \end{bmatrix} {\quad\text{for all}\quad}A \in {{\mathfrak{A}}},$$ so that $\rho_i \vert_{{{\mathfrak{A}}}\cap {{\mathfrak{A}}}^*} \in {{\operatorname{irred}}}({{\mathfrak{A}}}\cap {{\mathfrak{A}}}^*, {{\mathcal{H}}})$ for $i=1,2$ and $\rho_3({{\mathfrak{A}}})\neq \{0\}$. \[basic\] Let $({{\mathcal{A}}}, \alpha)$ be a C\*-dynamical system and let $$\rho \in {\operatorname{rep}}_2({{\mathcal{A}}}\times_{\alpha}{{\mathbb{Z}}}^+, {{\mathcal{H}}}).$$ Then, $\rho_1\vert_{{{\mathcal{A}}}} \simeq \rho_2 \vert_{{{\mathcal{A}}}}\circ\alpha$. First note that $\rho$ is a $*$-representation on the diagonal ${{\mathcal{A}}}$. Hence for each $A \in {{\mathcal{A}}}$, $\rho(A)$ necessarily has its $(1, 2)$-entry equal to zero, i.e., $\rho({{\mathcal{A}}})$ is in diagonal form. Let$\{E_j\}_j$ be an approximate unit for ${{\mathcal{A}}}$; and let $X, Y, Z\in {{\mathcal{H}}}$ so that $\{\rho(U_{\alpha}E_j)\}_j$ converges weakly to $\left[\begin{smallmatrix} X &Y \\0&Z \end{smallmatrix}\right]$. *We claim that $Y\neq0$.* Indeed, otherwise the equality $$\rho(U_{\alpha}A) = \lim_j \rho(U_{\alpha}E_jA) = \lim_j \rho(U_{\alpha}E_j)\rho(A) {\quad\text{for all}\quad}A \in {{\mathcal{A}}}$$ would imply that $\rho(U_{\alpha}{{\mathcal{A}}})$ is diagonal; and therefore $\rho({{\mathcal{A}}}\times_{\alpha}{{\mathbb{Z}}}^+)$ is in diagonal form, contradicting the requirement $\rho_{3}({{\mathcal{A}}}\times_{\alpha}{{\mathbb{Z}}}^+)\neq \{0\}$. Now notice that for any $A \in {{\mathcal{A}}}$, we have $$\begin{aligned} \rho(A)\lim_j \rho(U_{\alpha}E_j)& = \lim_j \rho(U_{\alpha}\alpha(A)E_j) \\ &=\lim_j(U_{\alpha}E_j)\rho(\alpha(A)).\end{aligned}$$ Hence in matricial form, $$\begin{bmatrix} \rho_1(A) & 0\\ 0 & \rho_2(A) \end{bmatrix} \begin{bmatrix} X& Y \\ 0 & Z \end{bmatrix} = \begin{bmatrix} X & Y \\ 0 & Z \end{bmatrix} \begin{bmatrix} \rho_1(\alpha(A))& 0\\ 0 & \rho_2(\alpha(A)) \end{bmatrix} .$$ By multiplying and comparing $(1,2)$-entries, we obtain $$\rho_1(A)Y=Y\rho_2(\alpha(A)).$$ Since $Y\neq0$ and $\rho_1\vert_{{{\mathcal{A}}}}$ and $\rho_2\vert_{{{\mathcal{A}}}}$ are irreducible, this implies that $\rho_1\vert_{{{\mathcal{A}}}} \simeq\rho_2\vert_{{{\mathcal{A}}}}\circ\alpha$, as desired. \[itexists\] Let $({{\mathcal{A}}}, \alpha)$ be a C\*-dynamical system, and let $\sigma$ belong to ${{\operatorname{irred}}}({{\mathcal{A}}}, {{\mathcal{H}}})$. Then there exists a representation $\rho \in {\operatorname{rep}}_2({{\mathcal{A}}}\times_{\alpha}{{\mathbb{Z}}}^+, {{\mathcal{H}}}) $ so that $\rho_2\vert_{{{\mathcal{A}}}}=\sigma$. Let $$\rho(A)= \begin{bmatrix} \sigma\circ\alpha(A) & 0 \\ 0 & \sigma(A) \end{bmatrix} {\quad\text{and}\quad}\rho(U_{\alpha}A)= \begin{bmatrix} 0 & \sigma(A) \\ 0 & 0 \end{bmatrix} .$$ It is easily verified that $\big( \rho\vert_{{{\mathcal{A}}}}, \left[\begin{smallmatrix} 0 &I \\0&0 \end{smallmatrix}\right] \big)$ is a covariant representation and so $\rho$ extends to a representation of ${{\mathcal{A}}}\times_{\alpha}{{\mathbb{Z}}}^+$. Note that the representation $\rho$ in Lemma \[itexists\] satisfies $\rho_1|_{{{\mathcal{A}}}}= \mbox{$\sigma\circ \alpha$}$. This is not just an artifact of our construction. By Lemma \[basic\], any representation in $\rho \in {\operatorname{rep}}_2({{\mathcal{A}}}\times_{\alpha}{{\mathbb{Z}}}^+, {{\mathcal{H}}})$ will satisfy that property, provided that $\rho_2\vert_{{{\mathcal{A}}}}=\sigma$. The following general result relates the classification problem for semicrossed products to the study of universally weakly inner automorphisms for C\*-algebras. \[main\] Let $({{\mathcal{A}}}, \alpha)$ and $({{\mathcal{B}}}, \beta)$ be C\*-dynamical systems, and assume that the semicrossed products ${{\mathcal{A}}}\times_{\alpha} {{\mathbb{Z}}}^{+}$ and ${{\mathcal{B}}}\times_{\beta} {{\mathbb{Z}}}^{+}$ are isometrically isomorphic. Then there exists a C\*-isomorphism $\gamma: {{\mathcal{A}}}\rightarrow{{\mathcal{B}}}$ so that $\alpha\circ\gamma^{-1}\circ \beta^{-1}\circ \gamma$ is universally weakly inner with respect to irreducible representations. Assume that there exists an isometric isomorphism $$\gamma: {{\mathcal{A}}}\times_{\alpha} {{\mathbb{Z}}}^{+}\longrightarrow {{\mathcal{B}}}\times_{\beta} {{\mathbb{Z}}}^{+}.$$ By Proposition \[diag\], $\gamma\vert_{{{\mathcal{A}}}}$ is a $*$-isomorphism onto ${{\mathcal{B}}}$, which we will also denote by $\gamma$; this is the promised isomorphism. Indeed $\gamma$ establishes a correspondence $${{\operatorname{irred}}}({{\mathcal{A}}}) \ni \sigma \longrightarrow \sigma \circ \gamma^{-1}\in {{\operatorname{irred}}}({{\mathcal{B}}})$$ that preserves equivalence classes. To show that $\alpha\circ\gamma^{-1}\circ \beta^{-1}\circ \gamma$ is universally weakly inner with respect to irreducible representations, it is enough to show that $$\sigma \circ\alpha\circ\gamma^{-1} \simeq \sigma \circ \gamma^{-1} \circ \beta$$ for any $\sigma \in {{\operatorname{irred}}}({{\mathcal{A}}})$. By Lemma \[itexists\], there exists $\rho \in {\operatorname{rep}}_2({{\mathcal{A}}}\times_{\alpha}{{\mathbb{Z}}}^+, {{\mathcal{H}}})$ so that $\rho_2\vert_{{{\mathcal{A}}}}=\sigma$. But then $$\rho\circ \gamma^{-1} \in {\operatorname{rep}}_2({{\mathcal{B}}}\times_{\beta}{{\mathbb{Z}}}^+, {{\mathcal{H}}}).$$ Hence, Lemma \[nestrepn\] implies that $$(\rho\circ\gamma^{-1})_1\vert_{{{\mathcal{B}}}} \simeq (\rho\circ\gamma^{-1})_2 \vert_{{{\mathcal{B}}}} \circ \beta$$ or, once again, $\sigma\circ \alpha\circ\gamma^{-1} \simeq \sigma\circ\gamma^{-1}\circ \beta$ . We have arrived to the main result of the paper. \[Main\] Let $({{\mathcal{A}}}, \alpha)$ and $({{\mathcal{B}}}, \beta)$ be C\*-dynamical systems and assume that ${{\mathcal{A}}}$ is a separable simple C\*-algebra. The semicrossed products ${{\mathcal{A}}}\times_{\alpha} {{\mathbb{Z}}}^{+}$ and ${{\mathcal{B}}}\times_{\beta} {{\mathbb{Z}}}^{+}$ are isometrically isomorphic if and only if the dynamical systems $({{\mathcal{A}}}, \alpha)$ and $({{\mathcal{B}}}, \beta)$ are outer conjugate. One direction follows from Proposition \[onedir\]. Conversely assume that ${{\mathcal{A}}}\times_{\alpha} {{\mathbb{Z}}}^{+}$ and ${{\mathcal{B}}}\times_{\beta} {{\mathbb{Z}}}^{+}$ are isometrically isomorphic. Theorem \[main\] shows that there exists a C\*-isomorphism $\gamma: {{\mathcal{A}}}\rightarrow{{\mathcal{B}}}$ so that $\alpha\circ\gamma^{-1}\circ \beta^{-1}\circ \gamma$ is universally weakly inner with respect to irreducible representations. The conclusion follows now from Kishimoto’s result [@Kis Corollary 2.3]. Representations and dynamics on the spectrum ============================================ As we mentioned in the introduction, an earlier result of Muhly and Solel [@MS Theorem 4.1] implies the validity of Theorem \[Main\] for arbitrary ${\mathrm{C}^*}$-algebras, provided that the automorphisms $\alpha$ and $\beta$ have full Connes spectrum. In what follows we present an alternative proof of that result of Muhly and Solel, based on the ideas developed in this paper. Let $\hat{\Phi}_{\alpha}$ denote the canonical expectation from ${{\mathcal{A}}}\times_{\alpha} {{\mathbb{Z}}}$ onto ${{\mathcal{A}}}$ and let $\Phi_{\alpha}$ denote its restriction on the semicrossed product ${{\mathcal{A}}}\times_{\alpha} {{\mathbb{Z}}}^+$. The following is the key step in their proof. Let $({{\mathcal{A}}}, \alpha)$ and $({{\mathcal{B}}}, \beta)$ be separable C\*-dynamical systems and let $\gamma :{{\mathcal{A}}}\times_{\alpha} {{\mathbb{Z}}}^{+} \rightarrow {{\mathcal{B}}}\times_{\beta} {{\mathbb{Z}}}^{+}$ be an isometric isomorphism. If $\alpha$ has full Connes spectrum, then $\gamma(\ker \Phi_{\alpha})= \ker \Phi_{\beta}$. Let $\Delta_{\alpha} = \{\sigma_j\mid j \in {{\mathbb{J}}}\}$ be the dense $\alpha$-invariant set of aperiodic points described in the introduction. By Lemma \[theta\], $\oplus_{j}\sigma_j$ is a faithful representation of ${{\mathcal{A}}}$ and so $\left(\oplus_{j}\sigma_j\right) \times_{\alpha} V$ defines a faithful representation of ${{\mathcal{A}}}\times_{\alpha} {{\mathbb{Z}}}$. Now notice that the compression on the main diagonal of $\left(\oplus_j \sigma_j\right) \times_{\alpha} V$ defines an expectation from ${{\mathcal{A}}}\times_{\alpha} {{\mathbb{Z}}}$ onto ${{\mathcal{A}}}$ which coincides with $\hat{\Phi}_{\alpha}$. Similarly, $\left(\oplus_{j}\sigma_j \circ \gamma^{-1}\right) \times_{\beta} V$ defines a faithful representation of ${{\mathcal{B}}}\times_{\beta} {{\mathbb{Z}}}$ and $\hat{\Phi}_{\beta}$ is the compression on the main diagonal. Now $\gamma({{\mathcal{A}}}) = {{\mathcal{B}}}$ and $\ker \Phi_{\alpha}=U_{\alpha}\left({{\mathcal{A}}}\times_{\alpha} {{\mathbb{Z}}}^{+}\right)$. Hence it is enough to show that $\gamma(U_{\alpha}) \in \ker \Phi_{\beta}$, i.e., the diagonal entries of $\gamma(U_{\alpha})$ in the representation $\left(\oplus_{j}\sigma_j \circ \gamma^{-1}\right) \times_{\beta} V$ of ${{\mathcal{B}}}\times_{\beta} {{\mathbb{Z}}}^{+}$ are equal to 0. To verify this examine these entries in light of the covariance equation as in Lemma \[basic\]. The rest of the proof follows the same arguments as in [@MS]. \[MS\] Let $({{\mathcal{A}}}, \alpha)$ and $({{\mathcal{B}}}, \beta)$ be separable C\*-dynamical systems and assume that $\alpha$ has full Connes spectrum. The semicrossed products ${{\mathcal{A}}}\times_{\alpha} {{\mathbb{Z}}}^{+}$ and ${{\mathcal{B}}}\times_{\beta} {{\mathbb{Z}}}^{+}$ are isometrically isomorphic if and only if the dynamical systems $({{\mathcal{A}}}, \alpha)$ and $({{\mathcal{B}}}, \beta)$ are outer conjugate. Assume that there is an isometric isomorphism $\gamma$ of ${{\mathcal{A}}}\times_{\alpha} {{\mathbb{Z}}}^{+}$ onto ${{\mathcal{B}}}\times_{\beta} {{\mathbb{Z}}}^{+}$. By the previous Lemma we have $$\begin{aligned} \gamma(U_{\alpha})&= BU_{\beta}+Y, \quad\text{where }B \in {{\mathcal{B}}}{\text{ and }}Y \in U_{\beta}^2 ({{\mathcal{B}}}\times_{\beta} {{\mathbb{Z}}}^{+}) ,\\ \gamma^{-1}(U_{\beta})&= AU_{\alpha} + Z, \quad\text{where }A \in {{\mathcal{A}}}{\text{ and }}Z \in U_{\alpha}^2 ({{\mathcal{A}}}\times_{\alpha} {{\mathbb{Z}}}^{+}).\end{aligned}$$ Since both $\gamma$ and $\gamma^{-1}$ are isometries, $\|A\|, \|B\|\leq1$. Also, $$U_{\alpha}= \gamma^{-1}(\gamma(U_{\alpha})) =\gamma^{-1}(B)AU_{\beta} +\gamma^{-1}(B)Z+\gamma^{-1}(Y),$$ which by the uniqueness of the Fourier expansion implies that $\gamma^{-1}(B)A=I$. Since $A$ and $\gamma^{-1}(B)$ are contractions, they must both be unitary. Hence $B$ and $\gamma(A)$ are also unitary. Therefore $$\begin{aligned} \hat{\Phi}_{\beta}(\gamma(U_{\alpha})^*\gamma(U_{\alpha})) &= \hat{\Phi}_{\beta}(I + U_\beta^* B^*Y + Y^* B U_\beta + Y^*Y) \\&= I+ \hat{\Phi}_{\beta}(Y^*Y)\end{aligned}$$ since $$U_\beta^* B^*Y \in U_\beta^* {{\mathcal{B}}}U_{\beta}^2 ({{\mathcal{B}}}\times_{\beta} {{\mathbb{Z}}}^{+}) \subseteq U_{\beta} ({{\mathcal{B}}}\times_{\beta} {{\mathbb{Z}}}^{+}) \subseteq \ker\Phi_\beta;$$ and likewise $Y^* B U_\beta \in (\ker\Phi_\beta)^*\subseteq \ker\hat{\Phi}_\beta$. So $\hat{\Phi}_{\beta}(Y^*Y)=0$; whence $Y=0$. Hence, $\gamma(U_{\alpha})= BU_{\beta}$ which implies that $({{\mathcal{A}}}, \alpha)$ and $({{\mathcal{B}}}, \beta)$ are outer conjugate. We finish the paper with a result of independent interest that associates the fixed points of $\hat{\alpha}$ to a certain analytic structure in the space of representations ${\operatorname{rep}}({{\mathcal{A}}}\times_{\alpha} {{\mathbb{Z}}}^+, {{\mathcal{H}}})$. Let ${{\mathfrak{A}}}$ be an operator algebra. A map $$\Pi :{{\mathbb{D}}}\equiv \{z \in {{\mathbb{C}}}\mid |z|<1\} \longrightarrow ({\operatorname{rep}}({{\mathfrak{A}}}, {{\mathcal{H}}}), \text{point--{\textsc{sot}}})$$ is called *analytic* if for each $A \in {{\mathfrak{A}}}$ and $x,y \in {{\mathcal{H}}}$ the map $z \rightarrow \left< \Pi(z)(A)x\mid y\right>$, $z \in {{\mathbb{D}}}$, is analytic in the usual sense. Assume now that $({{\mathcal{A}}}, \alpha)$ is a C\*-dynamical system, and let $$\Pi :{{\mathbb{D}}}\longrightarrow {\operatorname{rep}}({{\mathcal{A}}}\times_{\alpha} {{\mathbb{Z}}}^+, {{\mathcal{H}}})$$ be an analytic map. Since $\Pi(z)$, $z \in {{\mathbb{D}}}$, is a $*$-homomorphism on the diagonal, we may write $$\Pi(z)= \pi_z \times K_z, \quad z \in {{\mathbb{D}}},$$ where $\pi_z=\Pi(z)\vert _{{{\mathcal{A}}}}$ and $K_z=\Pi(z)(U_{\alpha})$. We claim that the map $ z \longrightarrow \pi_z$, $z \in {{\mathbb{D}}}$, is constant. Indeed, consider any selfadjoint operator $A \in {{\mathcal{A}}}$ and notice that the map $${{\mathbb{D}}}\ni z \longrightarrow {\langle}\pi_z(A)x, x {\rangle}, \quad x \in {{\mathcal{H}}},$$ is real analytic and therefore constant, which proves the claim. Let ${{\mathfrak{A}}}$ be an operator algebra. An analytic map $\Pi: {{\mathbb{D}}}\rightarrow {\operatorname{rep}}({{\mathfrak{A}}}, {{\mathcal{H}}})$ is said to be *irreducible on the diagonal* if for any $z \in {{\mathbb{D}}}$, the representation $\Pi(z)\vert_{{{\mathfrak{A}}}\cap {{\mathfrak{A}}}^*}$ is irreducible. \[analytic\] Let $({{\mathcal{A}}}, \alpha)$ be a C\*-dynamical system, and $\sigma$ belong to ${{\operatorname{irred}}}({{\mathcal{A}}}, {{\mathcal{H}}})$. Then $\sigma \simeq \sigma \circ \alpha$ if and only if there exists a non-constant analytic and irreducible on the diagonal map $$\Pi: {{\mathbb{D}}}\longrightarrow {\operatorname{rep}}({{\mathcal{A}}}\times_{\alpha} {{\mathbb{Z}}}^+, {{\mathcal{H}}})$$ so that $$\Pi(z)\vert_{{{\mathcal{A}}}} = \sigma$$ for some $z \in {{\mathbb{D}}}$. First assume that $\sigma \simeq \sigma \circ \alpha$ and let $U$ be a unitary such that $\sigma(A) U = U (\sigma \circ \alpha)(A)$ for $A \in {{\mathcal{A}}}$. Then $$\Pi(z) = \sigma \times zU {\quad\text{for}\quad}z \in {{\mathbb{D}}}$$ has the desired properties. Conversely, assume that such a map $\Pi$ exists. By the discussion above $\Pi(z)\vert { {{\mathcal{A}}}}=\sigma$ for all $z \in {{\mathbb{D}}}$. Since $\Pi$ is not constant, there exists some $z \in {{\mathbb{D}}}$ so that $K := \Pi(z)(U_{\alpha}) \ne 0$. Therefore, $$\begin{aligned} \sigma(A)K&=\Pi(z)(A)\Pi(z)(U_{\alpha}) \\ &=\Pi(z)\left(U_{\alpha}\alpha(A)\right) \\ &=K\sigma(\alpha(A)) \qquad {\quad\text{for all}\quad}A \in {{\mathcal{A}}}.\end{aligned}$$ Since both $\sigma$ and $\sigma \circ \alpha$ are irreducible and $K \neq 0$, we obtain that $\sigma \simeq \sigma\circ \alpha$. Note that if $\Pi$ is as in the above Proposition, then the operator $K$ in the proof is necessarilly a scalar multiple of the (unique) unitary operator $U$ implementing the equivalence $\sigma \simeq \sigma \circ \alpha$. Therefore the range of $\Pi$ is contained in $$D_{\sigma} \equiv \{ \sigma \times zU \mid z\in \overline{{{\mathbb{D}}}}\}.$$ In this fashion, we associate with each fixed point $\sigma \in \hat{{{\mathcal{A}}}}$, a unique *maximal analytic set* $D_{\sigma}$. Any representation of ${{\mathcal{A}}}\times_{\alpha} {{\mathbb{Z}}}^+$ not belonging to the union of the maximal analytic sets is associated with a non-fixed point of $\hat{\alpha}$. [99]{} C. Akemann, G. Pedersen and J. Tomiyama, *Multipliers of C\*-Algebras*, J. Funct. Anal. **13** (1973), 277–301. M. Alaimia and J. Peters, *Semicrossed products generated by two commuting automorphisms* J. Math. Anal. Appl. **285** (2003), 128–140. W. Arveson, *Operator algebras and measure preserving automorphisms*, Acta Math. **118**, (1967), 95–109. W. Arveson and K. Josephson, *Operator algebras and measure preserving automorphisms II*, J. Funct. Anal. **4** (1969), 100–134. D. Buske, *Hilbert modules over a class of semicrossed products*, Proc. Amer. Math. Soc. **129** (2001), 1721–1726. D. Buske and J. Peters, *Semicrossed products of the disk algebra: contractive representations and maximal ideals.*, Pacific J. Math. **185** (1998), 97–113. L. DeAlba and Peters, J. *Classification of semicrossed products of finite-dimensional C\*-algebras*, Proc. Amer. Math. Soc. **95** (1985), 557–564. K. Davidson and E. Katsoulis *Isomorphisms between topological conjugacy algebras*, J. Reine Angew. Math. (Crelle), to appear. K. Davidson and E. Katsoulis *Operator algebras for multivariable dynamics*, preprint, 2007. G. Elliott, *Ideal preserving automorphisms of postliminary C\*-algebras*, Proc. Amer. Math. Soc. **27**, (1971), 107–109. D. Hadwin and T. Hoover, *Operator algebras and the conjugacy of transformations*, J. Funct. Anal. **77** (1988), 112–122. T. Hoover, *Isomorphic operator algebras and conjugate inner functions*, Michigan Math. J. **39** (1992), 229–237. T. Hoover, J. Peters and W. Wogen, *Spectral properties of semicrossed products.* Houston J. Math. **19** (1993), 649–660. R. Kadison and Ringrose, J. *Derivations and automorphisms of operator algebras*, Comm. Math. Phys. **4** (1967), 32–63. T. Katsura, *On C\*-algebras associated with C\*-correspondences*, J. Funct. Anal. **217** (2004), 366–401. A. Kishimoto, *Outer automorphisms and reduced crossed products of simple C\*-algebras*, Comm. Math. Phys. **81** (1981), 429–435. M. Lamoureux, *Nest representations and dynamical systems*, J. Funct. Anal. **114** (1993), 467–492. M. Lamoureux, *Ideals in some continuous nonselfadjoint crossed product algebras.* J. Funct. Anal. **142** (1996), 211–248. E.C. Lance, *Automorphisms of postliminal C\*-algebras*, Pacific J. Math. **23**, (1967), 547–555. M. McAsey and P. Muhly, *Representations of nonselfadjoint crossed products*, Proc. London Math. Soc. (3) **47** (1983), 128–144. P. Muhly, B. Solel, *On the Morita equivalence of tensor algebras*, Proc. London Math. Soc. (3) **81** (2000), 113–168. D. Olesen and Pedersen, G. *Applications of the Connes spectrum to C\*-dynamical systems III*, J. Funct. Anal. **45** (1982), 357–390. T. Ohwada and K. Saito, *Factorization in analytic crossed products*, J. Math. Soc. Japan **54** (2002), 21–33. J. Peters, *Semicrossed products of C\*-algebras*, J. Funct. Anal. **59** (1984), 498–534. J. Peters, *The ideal structure of certain nonselfadjoint operator algebras*, Trans. Amer. Math. Soc. **305** (1988), 333–352. S. Power, *Classification of analytic crossed product algebras*, Bull. London Math. Soc. **24** (1992), 368–372. B. Solel, *The invariant subspace structure of nonselfadjoint crossed products*, Trans. Amer. Math. Soc. **279** (1983), 825–840. [^1]: [^2]: First author partially supported by an NSERC grant. [^3]: Second author partially supported by a summer grant from ECU

ArXiv

--- abstract: 'This paper addresses network anomography, that is, the problem of inferring network-level anomalies from indirect link measurements. This problem is cast as a low-rank subspace tracking problem for normal flows under incomplete observations, and an outlier detection problem for abnormal flows. Since traffic data is large-scale time-structured data accompanied with noise and outliers under partial observations, an efficient modeling method is essential. To this end, this paper proposes an online subspace tracking of a Hankelized time-structured traffic tensor for normal flows based on the Candecomp/PARAFAC decomposition exploiting the recursive least squares (RLS) algorithm. We estimate abnormal flows as outlier sparse flows via sparsity maximization in the underlying under-constrained linear-inverse problem. A major advantage is that our algorithm estimates normal flows by low-dimensional matrices with time-directional features as well as the spatial correlation of multiple links without using the past observed measurements and the past model parameters. Extensive numerical evaluations show that the proposed algorithm achieves faster convergence per iteration of model approximation, and better volume anomaly detection performance compared to state-of-the-art algorithms.' author: - 'Hiroyuki Kasai[^1]' - 'Wolfgang Kellerer[^2]' - 'Martin Kleinsteuber[^3]' bibliography: - 'network\_analysis.bib' - 'matrix\_tensor\_completion.bib' title: | Network Volume Anomaly Detection and Identification in\ Large-scale Networks based on\ Online Time-structured Traffic Tensor Tracking --- Introduction ============ Diagnosing unusual events (called “anomalies") in a large-scale network like Internet Service Providers and enterprise networks is critical and challenging for both network operators and end users. Anomalies occur due to activity from malicious operations, or misconfigurations and failures of network equipments. This paper addresses “[*traffic volume anomaly*]{}", which means large and sudden positive or negative [*traffic volume changes*]{} due to strong variances in traffic flows. These changes are typically caused by unexpected events such as alpha events (e.g., large file transfers), outages coming from network equipment failures, network attacks like denial-of-service attacks (DoS), and traffic shifts. Diagnosing such a volume anomaly in a flow consists of three steps; detection, identification and quantification [@Lakhina_SIGCOMM_2004]. The [*detection*]{} step is to unveil time points when the network is facing an anomaly. The [*identification*]{} step consists of selecting the right anomaly type from a set of possible candidate anomalies. The identification step addressed in this paper additionally allows to identify which flow experiences such an anomaly. Finally, the [*quantification*]{} step is about to estimate the number of additional or missing bytes in the underlying traffic flow. The detection and identification problems are of interest in this paper. The quantification problem is essential but extremely challenging in the online-based approach under an incomplete observation situation that this paper addresses, and this remains an open problem. Network-wide traffic is typically expressed in the form of matrices or multidimensional arrays, i.e., tensors. In general, a traffic volume exchanged between every pair of an ingress and an egress node or PoP (Point of Presence) during a given time period forms a two-dimensional non-negative matrix, often called the [*traffic matrix*]{} or the [*flow matrix*]{}. This paper explicitly uses the term flow matrix to avoid confusion with the link matrix mentioned hereafter. A flow is referred to as an [*origin-destination*]{} (OD) flow. Here, the term [*traffic volume*]{} refers to the number of bytes, packets, or flows measured in a certain time interval at one point in the network. A noteworthy point is that one single OD flow traverses multiple links based on the routing tables, thereby a volume anomaly in one single OD flow is visible across several links simultaneously. For diagnosing, a large-scale direct collecting of [*flow measurements*]{}, which is [*flow-level data*]{}, is extremely resource intensive due to the collection of [*fine-grained*]{} data. It also requires a flow monitoring infrastructure across an entire network, and this is extremely costly. To the contrary, [*link measurements*]{}, which are [*device-level data*]{}, can be easily collected by the Simple Network Management Protocol (SNMP) that periodically collects device readings, but provides only [*coarse-grained*]{} information. The link measurements are generally expressed as the [*link matrix*]{}, which represents traffic volume of each link over time, and the link matrix is obtained by multiplying the [*routing matrix*]{} to the flow matrix. Practical limitations are that the collected link measurements are often noisy [@Roughan_JECE_2010], contain [*missing data*]{} due to using the unreliable UDP transport, and lack measurement synchronization across an entire network. It should also be noted that the link measurements are a linear combination of OD flows, and, the observed traffic on each link is the [*superposition*]{} of multiple OD flows. Consequently, this paper addresses the [*flow matrix estimation problem*]{} from the link matrix, which is referred to as [*network tomography*]{} [@Vardi_JASA_1996; @Zhang_ACM_IMC_2005], to diagnose traffic volume anomaly. This estimates a [*directly unobservable*]{} flow matrix from the directly observable link matrix by using standard SNMP per-link byte counts. We also refer to the problem of inferring anomalies from these indirect measurement of the flow matrix as [*network anomography*]{}. It should be noted that this is different from and more complex than network tomography, in the way that network anomography is performed against sequential measurements over a certain period of time, rather than from a single snapshot of measurements. Despite extensive studies for decades in this field, network tomography and anomography represent key technical issues of network management facing network evolution. The classical but key challenges that lie at the core of flow matrix estimation and volume anomaly diagnosis are as follows; the volume anomaly diagnosis problem stems from the fact that it only uses the link measurements, whereas the number of links in a network is generally much smaller than that of OD flows [@Lakhina_SIGCOMM_2004]. The problem is described as an [*under-constrained linear-inverse problem*]{}, where the solution relies on a prior model of the flow matrix (e.g., the Poisson model [@Vardi_JASA_1996], the gravity model [@Zhang_SIGMETRICS_2003; @Zhang_IEEEACMTranNW_2005]). An additional challenge comes from the superimposed OD flows on a link. While an OD flow has pronounced spikes, the spikes are dwarfed in the corresponding link traffic. In other words, clearly recognizable anomalous spikes in the flow matrix can often be covered in the link matrix through “dreadful interference" of the superimposed OD flows. Keeping these challenges in our mind, we further need to consider an efficient and robust way to handle noisy, incomplete, high-dimensional flow matrices when the network size becomes larger. To this end, our motivation and approach for the primary contribution of this paper are summarized as follows; considering that the flow and link matrices have time-directional structure such as periodicity and seasonality, a novel analysis method uncovering latent time-directional structure inside matrices is required. For this purpose, a [*Hankel matrix*]{} representation of the link matrix with the [*tensor structure*]{} plays a crucial role. This paper specifically transfers this link matrix to the [*Hankelized time-structured traffic tensor*]{}, and analyzes it. Moreover, since these matrices have high-dimensional data accompanied with noise, a [*subspace-based approach*]{}, i.e., a low-rank approximation approach, is desirable to robustly model underlying latent behavior of network traffic. Thereby, we exploit a low-rank [*tensor decomposition*]{} not only to model robustly normal flows but also to compute big link matrices more efficiently by [*thin*]{} matrices rather than an entire size of matrices. Besides, as for the detection of abnormal flows, this can be newly cast as an outlier detection problem via the [*sparsity constraint*]{} formulation. Furthermore, the flow matrix estimation requires tolerance and insensitivity to missing measurements and non-synchronized measurements due to the unreliable UDP transport, network-wide monitoring, or machine failures. This requires to interpolate missing values in the link matrix [@Roughan_IEEEACM_TranNW_2012], and motivates us to support a [*matrix completion*]{} function. Additionally, because the flow matrix and the link matrix grow infinitely as time goes by, an [*online-based*]{} approach without storing all the past measurements and model parameters is vital and effective. Finally, considering that the underlying subspace changes dramatically, and the processing speed is faster than the data acquiring speed, this motivates us to adopt a [*second-order optimization*]{} algorithm because the fast convergence property in each iteration (analysis) is preferred over convergence in computation time. In fact, sampling of the link measurements is periodical but intermittent (e.g., every 5 minutes in basic SNMP). Many efforts have been done so far in order to identify anomalies by analyzing network traffic [@Barford_ACM_IM_2002; @Krishnamurthy_IMC_2003; @Lakhina_SIGCOMM_2004; @Soule_IMC_2005; @Zhang_ACM_IMC_2005; @Ringberg_ACMSigmetrics_2007; @Roughan_IEEEACM_TranNW_2012]. They have attempt to expose anomalies by detecting deviations or errors from the constructed underlying model of normal traffic. There is, however, no single approach to satisfy the required algorithm capabilities that can handle noisy, high-dimensional, and time-series data with missing measurements in an online fashion. Consequently, this paper presents a new proposal for an online subspace tracking of the Hankelized time-structured traffic tensor for normal flows based on the Candecomp/PARAFAC tensor decomposition by exploiting the recursive least squares (RLS) method under incomplete observation situation. This paper also estimates abnormal flows as outlier sparse flows via sparsity maximization in the under-constrained linear-inverse problem by alternating direction method of multipliers (ADMM). Extensive numerical evaluations show that the proposed algorithm achieves faster convergence per iteration of model approximation, and better volume anomaly detection performance compared with state-of-the-art algorithms Related Work {#Sec:Related Work} ============ There is a rich literature in anomaly detection algorithms, and other related fields [@Modi_JNCP_2013]. They are for example based on signature profiles, artificial neural networks, fuzzy logic, association rules, support vector machines, genetic algorithms, or hybrid techniques. The efficiency of those algorithms depends on their parameters, their configurations, target anomaly types, or network types. Hence, this section describes related work for anomaly diagnosis from the viewpoint of time-directional analysis of high-dimensional data. Then, especially addressing efficient analysis means of infinite time-directional traffic data accompanied with noise, high dimension and incomplete observation, the state-of-the-art algorithms for online-based subspace tracking, which is actively studied in the machine learning field, are discussed, which are closely related to our work. Anomaly Detection and Identification {#Sec:AnomalydetectionandIdentification} ------------------------------------ The first category of anomaly diagnosis is to detect specific kinds of traffic anomalies including network attacks, network failures, and traffic shifts [@Casas_ComputerNetwork_2010]. This category operates on individual and independent time series, and analyzes traffic on a particular network link, particular device readings, or particular packet features. Classical forecasting methods (e.g., ARIMA [@Hood_IEEETransRe_1997], Holt-Winters [@Brutlag_LISA_2000], EWMA[^4], Kalman-Filter [@Soule_LSNI_2005]), and outliers analysis methods (e.g., Wavelet [@Barford_ACM_IM_2002], Fourier Transform) are used. The second category, which this article mainly addresses, is to diagnose anomalous traffic behaviors from a network-wide perspective. Here, the mechanisms collect coarse-grained SNMP link measurements to detect and isolate volume anomalies in OD flows. This category exploits spatial correlations across the time series of traffic from all the links of a network. Representative works are as follows; the Kalman-filtering approach [@Soule_IMC_2005] tracks the evolution of OD flows from SNMP measurements, and identifies anomalies as large prediction errors. The OD flows act as the [*underlying states*]{} of a network traffic system. The states evolve over time as the OD flows evolve, but the states are not directly observable. Namely, this approach estimates both spatial and temporal correlations [@Soule_IMC_2005]. However, it requires a long training period, in which direct anomaly-free OD flow measurements are used to calibrate the model. In addition, these methods do not explicitly handle noisy and higher dimensional characteristics of networks associated with a large number of nodes and links. An efficient lower dimensional approach, i.e., subspace-based approach or low-rank approximation based approach, is desirable to robustly model network traffic behavior. A PCA (Principle Component Analysis)-based approach [@Lakhina_SIGCOMM_2004; @Zhang_ACM_IMC_2005], one of the subspace-based approaches, separates SNMP measurements into a normal subspace and an anomalous subspace. The success of this approach lies on two points: link traffic is a linear combination of OD flows, and each OD flow has a low intrinsic dimensionality. This approach detects an anomaly when the magnitude of the projection onto the anomal subspace exceeds an associated PCA Q-Statistic threshold [@Jackson_Tech_1979]. A critical problem is that it faces scalability problems in large scale networks due to the expensive calculation cost of a big size of its transformation matrix. Further subspace-based approaches are proposed in [@Roughan_IEEEACM_TranNW_2012; @Mardani_IEEEJSTSP_2013; @Mardani_IEEEACMTranNW_2015]. Here, similar to the PCA-based approach, the flow matrix is approximated by a low-rank matrix and sparsity constraints (outlier detection). [@Roughan_IEEEACM_TranNW_2012] handles missing data explicitly. A robust PCA-like approach is also proposed to handle anomaly traffic as outlier traffic [@Wanga_CN_2012] based on the Relaxed Principal Component Pursuit [@Candes_JACM_2011]. Another approach is the SSA-based approach [@Tahereh_arXiv_2014]. SSA (singular spectrum analysis) is a technique of time-series analysis, and is a nonparametric spectral estimation method [@Broomhead_Physica_1986]. The time-series data is transformed into a Hankel matrix followed by the singular value decomposition to model its time structure. Multivariate SSA (M-SSA) is an extension of basic SSA into multivariate data. Nevertheless, all of these methods consider only a batch-based operation. Other than for lower dimensional approaches, which work on a snapshot of the network information, in this paper, we address the difficulty to learn [*a priori*]{} how such anomalies appear in traffic volume statistics, because large networks are affected by various types of anomalies in different ways. This motivates us to consider an online-based anomaly diagnosis algorithm. An extension of the Holt-Winters forecasting algorithm supports incremental model updating via exponential smoothing [@Brutlag_LISA_2000]. A stochastic approximation of the Expectation-Maximization (EM) algorithm for a Gaussian mixture model is proposed in [@Hajji_IEEETranNN_2005]. With respect to the PCA-based approach, although this is a purely spatial algorithm and cannot locate the anomaly temporally, an online formulation of this is proposed that uses a sliding window implementation to identify the normal and abnormal subspaces based on a previous block of time [@Lakhina_SIGMETRICS_2004]. Furthermore, a distributed algorithm has been proposed in [@Huang_NIPS_2006]. However, since the PCA-based detection algorithm is extremely sensitive to the proper determination of the associated Q-statistics threshold, a straightforward extension into an online-based algorithm is not robust [@Ahmed_INFOCOM_2007]. To solve this issue, a kernel version of the recursive least squares algorithm is proposed to construct and adapt a dictionary of features that approximately spans the subspace of normal behavior [@Ahmed_INFOCOM_2007]. This uses, however, the number of individual IP flows. A recursive estimation of a flow matrix using a Kalman filtering approach is also proposed [@Casas_TC_2009], but it uses direct OD flow measurements to calibrate the flow model. Finally, regarding tensor-based algorithms, a higher-order PCA detection algorithm is proposed based on the Higher-Order singular value decomposition and the Higher Order Orthogonal Iteration [@Kim_CAMSAP_2009]. The proposed methods outperform the normal PCA with respect to the scalability of the network size. However, this only considers the direct measurement case, and does not consider network tomography. The evolution of its subspace over time is not also considered. In addition, another anomaly detection algorithm is proposed using the higher order robust PCA with the subspace distance measurement, but this does not also consider network tomography [@Zoltowski_GlobalSIP_2014]. General Online-based Subspace Methods for High-dimensional Data Analysis ------------------------------------------------------------------------ This section details general online-based subspace learning methods that our approach falls into. They have been actively studied in machine learning field recently, and can be applied to circumvent potential issues in network analysis especially for noisy, high-dimensional and incomplete measurements. With regard to matrix-based online algorithms, a representative research is the projection approximation subspace tracking (PAST) [@Yang_IEEESP_1995]. GROUSE [@Balzano_Conf_2010] recently proposes an incremental gradient descent algorithm performed on the Grassmannian, the set of all $d$-dimensional subspaces of $\mathbb{R}^n$. The algorithm minimizes an $\ell_2$-norm cost function. GRASTA [@He_CVPR_2012] enhances robustness against outliers by exploiting an $\ell_1$-norm cost function. pROST proposes an improved GRASTA based on $\ell_0$-surrogates by using the conjugate gradient method [@Seidel_MVA_2014]. PETRELS [@Chi_IEEETransSP_2013] calculates the underlying subspace via a discounted recursive process for each row of the subspace matrix in parallel. Meanwhile, as for tensor-based online algorithms, an adaptive algorithm to obtain the Candecomp/PARAFAC decompositions [@Nion_IEEETransSP_2009] and an accelerated online tensor learning algorithm based on the Tucker decomposition [@Yu_ICML_2015] are proposed. However, they do not deal with missing data presence. Online imputation algorithms based on the Candecomp/PARAFAC decomposition are proposed for the presence of missing data [@Kasai_IEEEICASSP_2016_s; @Mardani_IEEETransSP_2015]. While [@Kasai_IEEEICASSP_2016_s] considers the RLS-based updates, it does not consider time-structured data and anomaly detection, [@Mardani_IEEETransSP_2015] considers the stochastic gradient descent (SGD) for large-scale data, and is applied to analyze network anomalies. Nevertheless, its convergence speed of [@Mardani_IEEETransSP_2015] is not fast, and the problem definition and the formation of traffic matrices in [@Mardani_IEEETransSP_2015] are not the same as ours, thereby it cannot be directly compared with our proposed algorithm. As seen above, none of these works has provided a complete and reliable solution to model and diagnose high-dimensional large scale data of network traffic data under incomplete observation in an online manner, which is the focus of our approach presented in this paper. Notations {#sec:Preliminaries} ========= Before we present our problem formulation, we summarize the notations used in the remainder of this article. We denote scalars by lower-case letters $(a, b, c, \ldots)$, vectors as bold lower-case letters $(\vec{a}, \vec{b}, \vec{c}, \ldots)$, and matrices as bold-face capitals $(\mat{A}, \mat{B}, \mat{C}, \ldots)$. An element at $(i,j)$ of a matrix is represented as $\mat{A}_{i,j}$. If $\mat{A}$ has additional index like $\mat{A}[t]$ or $\mat{A}$ is a matrix product like $\mat{A}=\mat{BC}$, we use $(\mat{A}[t])_{i,j}$ or $(\mat{BC})_{i,j}$ with parenthesis. $i$-th row vector and $j$-th column of $\mat{A}$ are represented as $\mat{A}_{i,:}$ and $\mat{A}_{:,j}$, respectively. We should particularly note that the transposed column vector of $i$-th row vector $\mat{A}_{i,:}$ is specially denoted as $\vec{a}^i$ in order to explicitly express a row vector, i.e., a horizontal vector. $\mat{A}_{i,p:q}$ represents $(\mat{A}_{i,p}, \ldots, \mat{A}_{i,q}) \in \mathbb{R}^{1 \times (q-p+1)}$. We call a multidimensional or multi-[*way*]{} (also called [*order*]{} or [*mode*]{}) array as [*tensor*]{}, which is denoted by $(\mathbfcal{A}, \mathbfcal{B}, \mathbfcal{C}, \ldots)$. Similarly, an element at $(i,j,k)$ of a third-order tensor $\mathbfcal{A}$ is expressed as $\mathbfcal{A}_{i,j,k}$. Tensor [*slice*]{} matrices are defined as two-dimensional matrices of a tensor, defined by fixing all but two indices. For example, a [*horizontal slice*]{} and a [*frontal slices*]{} of a third-order tensor $\mathbfcal{A}$ are denoted as $\mathbfcal{A}_{i,:,:}$ and $\mathbfcal{A}_{:,:,k}$, respectively. Since $\mathbfcal{A}_{:,:,k}$ is heavily used in this article, it is simply expressed as $\mat{A}_k$ using the bold-face capital font and one single subscript in order to explicitly represent its matrix form. Finally, $\vec{a}[t]$ and $\mat{A}[t]$ with the [*square bracket*]{} represent the computed $\vec{a}$ and $\mat{A}$ after performing $t$-times updates (iterations) in the online-based subspace tracking algorithm described in Section \[Sec:ProposedOnlineTrafficTensorAnomography\]. The notation [diag]{}(), where is a vector, stands for the diagonal matrix with $\{\vec{a}_i\}$ as diagonal elements. We follow the tensor notation of the review article [@Kolda_SIAMReview_2009] throughout our article and refer to it for further details. Network Anomography {#Sec:NetworkAnomography} =================== This section formally defines [*network tomography*]{} and  [*network anomography*]{} [@Zhang_ACM_IMC_2005]. For this purpose, we summarize our assumptions. First, we assume that aggregated link measurements (i.e., link matrix) are available via SNMP, and the routing information (i.e., routing matrix) at each time can be obtained from them, such as IGP link weights and the network topology information. Here, we consider the following [*generative traffic model*]{}. Let $\vec{f}_{i} \in \mathbb{R}^{1\times T}$ be $i$-th traffic for a time period of $T$ between the $i$-th ($i=\{1, \ldots, F\}$) node pair sorted by in a certain order, where $F$ represents the number of flows. This is generated by adding a normal traffic, $\vec{f}_{i}^{(no)} \in \mathbb{R}^{1\times T}$, an anomaly traffic $\vec{f}_{i}^{(ano)} \in \mathbb{R}^{1\times T}$ and a noise $\vec{f}_{i}^{(noise)} \in \mathbb{R}^{1\times T}$ as $\vec{f}_{i} = \vec{f}_{i}^{(no)} + \vec{f}_{i}^{(ano)} + \vec{f}_{i}^{(noise)}$. Then, we obtain the flow matrix as $[(\vec{f}_{1})^T : \cdots : (\vec{f}_{i})^T : \cdots :(\vec{f}_{F})^T]^T$ $\in \mathbb{R}^{F \times T}$. Especially, $\mat{F}_{:,t}$, that is $\mat{F}$ at time $t (0 \leq t \leq T)$, is $(\vec{f}_{1}(t), \ldots, \vec{f}_{F}(t))^T \in \mathbb{R}^F$. Here, without losing generality, we assume that routing paths are [*static*]{} for each pair of nodes during this time period of $T$ because they can be adopted each time in case of a dynamic case. Then, a routing matrix $\mat{R} \subset \{0,1\} \in \mathbb{R}^{L \times F}$ is $\mat{R}_{l,i}=1$ when the flow $\vec{f}_{i}$ passes $l$-th link, $\mat{R}_{l,i}=0$ otherwise, where $L$ represents the number of [*directly connected*]{} links. Subsequently, the link traffic at time $t$ is represented as $\mat{Y}_{:,t} = \mat{R} \mat{F}_{:,t},$ and, an entire link matrix $\mat{Y} \in \mathbb{R}^{L\times T}$ is represented as $\mat{Y} = \mat{R}\mat{F}$. Now we define the problem formulation of network tomography. The relationship among $\mat{F},\ $ and $\mat{Y}$ can be reformulated by considering errors as $\mat{Y} = \mat{R} \mat{F} + \mat{E}$, where is the error matrix of size $\mathbb{R}^{L \times T}$. If $\mat{Y}$ is observable measurement and the errors are assumed to be *i.i.d.* Gaussian, the ideal flow matrix $\hat{\mat{F}}$ can be modeled by minimizing the sum-of-squared errors, i.e., $$\hat{\mat{F}} = \defargmin_{\scriptsize{\mat{F}}} \frac{1}{2}\| \mat{Y} - \mat{R} \mat{F}\|_F^2. \nonumber$$ This is an under-constrained or ill-posed inverse problem because the number of OD pairs (unknown quantities), $F$, is more than that of link measurements, $L$, that is $L \ll F$. We define network anomography according to [@Zhang_ACM_IMC_2005] as follows. Assume that the flow matrix consists of [*normal flows*]{}, , and [*abnormal flows*]{}, , as $\mat{F} = \mat{X} + \mat{V}$, where $\mat{X} = [\vec{x}_1: \cdots : \vec{x}_L]$ and $\mat{V} = [\vec{v}_1 : \cdots : \vec{v}_L]$. Then, $\hat{\mat{X}}$ and $\hat{\mat{V}}$ are calculated below; $$\label{Eq:ProblemRe-formulation} \{\hat{\mat{X}}, \hat{\mat{V}}\} = \defargmin_{\scriptsize{\mat{X}, \mat{V}}} \frac{1}{2} \| \mat{Y} - \mat{R}(\mat{X}+ \mat{V}) \|_F^2.$$ If is fixed to $\hat{\mat{X}}$, $\hat{\mat{V}}$ is solved by the ill-posed linear inverse problem as $ \hat{\mat{V}} = {\rm arg\ min}_{\scriptsize{\mat{V}}} \frac{1}{2} \|\mat{D} - \mat{R} \mat{V} \|_F^2$, where $\mat{D} = \mat{Y} - \mat{R}\hat{\mat{X}}$ is a model approximation error. Here, we consider how to calculate $\mat{D}$ in existing modeling methods. Time-series analysis techniques (e.g. ARMA, ARIMA, EWMA) derive $\mat{D}$ as forecasting errors. $\mat{D}$ in signal processing based techniques (e.g. Fourier transform, Wavelet transform) is derived from high/middle frequency components by ignoring the lower frequency component. In subspace-based approaches like PCA, $\mat{D}$ is obtained from the abnormal subspace, that is the residual subspace of normal subspace projected by principle components. Regarding the abnormal flow $\hat{\mat{V}}$ estimation from signals, a greedy algorithm [@Lakhina_SIGCOMM_2004] is proposed, which finds the single largest anomaly in each time instance. Another algorithm is a linear-inverse-based algorithm [@Zhang_ACM_IMC_2005], which calculates the abnormal flow $\hat{\mat{V}}$ by solving the ill-posed linear inverse problem above. This has two types of algorithms. The Frobenius-norm minimization algorithm yields the optimal solution for the Gaussian noise assumption as $\hat{\mat{V}} = \mat{R}^{\dagger} \mat{D} = (\mat{R}^T \mat{R})^{-1} \mat{R}^T (\mat{Y} - \mat{R}\hat{\mat{X}})$, where $\mat{R}^{\dagger}$ is the pseudo-inverse of . On the other hand, the $\ell_p$-norm minimization with $0<p \leq 1$, which is called the [*sparsity maximization*]{}, considers the errors to be sparsely distributed but possibly large in magnitude. In this case, the $\tau$-th column of $\hat{\mat{V}}$ are the solutions to $$\label{Eq:SparsityMaximization} \hat{\vec{v}}_{\tau} = \defargmin_{\vec{v}_{\tau}} \| \vec{v}_{\tau} \|_p {\rm \ \ \ s.t.\ \ } \vec{y}_{\tau} -\mat{R} \hat{\vec{x}}_{\tau} = \mat{R} \vec{v}_{\tau}.$$ This article particularly focuses on the sparsity maximization algorithm for the abnormal flow estimation because this shows superior performances compared to others in [@Zhang_ACM_IMC_2005]. Proposed Online Traffic Tensor Anomography {#Sec:ProposedOnlineTrafficTensorAnomography} ========================================== This section defines the optimization problem of our proposed algorithm, and provides detailed solutions. The overall concept and procedures are summarized in Fig.\[Fig:BasicConcept\]. ![Basic architecture and procedures of the proposed algorithm.[]{data-label="Fig:BasicConcept"}](figures_final/concept.eps){width="14cm"} Derivation of Problem Formulation {#Sec:DerivationofProblemFormulation} --------------------------------- To robustly model both, the underlying latent structure of normal flows as well as abnormal outlier flows from noisy high-dimensional link measurements, we reformulate the problem (\[Eq:ProblemRe-formulation\]) by considering its low-rank constraint of and the sparsity constraint of as $$\label{Eq:Problem-LowRank-Sparsity} \min_{\scriptsize{\mat{X}, \mat{V}}} \frac{1}{2} \| \mat{Y} - \mat{R}(\mat{X}+ \mat{V}) \|_F^2 + \mu_r \cdot {\rm rank}(\mat{R}\mat{X}) + \mu_s \| \mat{V}\|_0,$$ where $\mu_r$ and $\mu_s$ control the rank constraint and the sparsity constraint, respectively. By following the literature [@Srebro_LearnTheory_2005], the rank constraint is transformed by the decomposed rank-$R$ matrices $\mat{M}^T \mat{Q}$ of $\mat{RX}$ where $\mat{M} \in \mathbb{R}^{M \times R}$ and $\mat{Q} \in \mathbb{R}^{T \times R}$. Additionally replacing the $\ell_0$-sparsity constraint with its convex $\ell_1$-surrogate, we obtain from (\[Eq:Problem-LowRank-Sparsity\]) $$\label{Eq:Problem-LowRank-Sparsity-l1} \min_{\scriptsize{\mat{X}, \mat{V}}} \frac{1}{2} \| \mat{Y} - \mat{R}(\mat{X}+\mat{V}) \|_F^2 +\mu_r ( \|\mat{M}\|^2_F + \|\mat{Q}\|^2_F ) +\ \mu_s \| \mat{V}\|_1.$$ Now, we address how to capture the time-directional structure of multiple links. Even if we analyze (\[Eq:Problem-LowRank-Sparsity-l1\]) by keeping the matrix form, we can capture the correlation between multiple links based on a similarity of those temporal variations. However, we cannot deal with the similarity between partial temporal variations inside one single link. In fact, network traffic has periodic and seasonality characteristics accompanied with relatively large noise and fluctuation. This should be taken into account to capture time-spatial correlations among multiple links. Thus, the multidimensional matrix, i.e., tensor, with the [*Hankel structure*]{} plays a crucial rule by exploiting a three-directional model against noisy and fluctuated signals. See Appendix \[Append\_Sec:hankel\] for the Hankel matrix. Although this idea is shared with the M-SSA based approach [@Tahereh_arXiv_2014], which combines multiple matrices horizontally, the tensor-based representation of multi-dimensional data can efficiently describe temporal-spatial correlations than M-SSA because the tensor-based approach stacks multiple data into different directions instead of placing them side by side onto the same direction. To this end, we first generate the Hankelized time-structured traffic tensor $\mathbfcal{Y}$, which is created by embedding a one-dimensional time-series data into multi-dimensional series. More concretely, let $\{y^l_{1}, \ldots ,y^l_{T}\} \in \mathbb{R}^{1\times T}$ be one-directional time-series traffic volume passing through $l$-th link of length $T$. Given a window length $W$, with $1<W<T$, we construct $k$-th $W$-lagged vectors $\vec{h}^l_{k} = (y^l_{k}, \ldots , y^l_{k+W-1})^T \in \mathbb{R}^W, k = 1,2,\ldots,K$, where $K=T-W+1$, and compose these vectors into the matrix $\mat{H}^l = [\vec{h}^l_{1}:\cdots :\vec{h}^l_{K}] \in \mathbb{R}^{W \times K}$ (Fig.\[Fig:BasicConcept\]-[i]{}). By plugging this Hankel matrix $\mat{H}^l$ into the $l$-th horizontal slice matrix of $\mathbfcal{Y}$, that is $\mathbfcal{Y}_{l,:,:}$, the traffic tensor $\mathbfcal{Y}$ is finally generated (Fig.\[Fig:BasicConcept\]-[ii]{}). It should be noted that the obtained traffic tensor $\mathbfcal{Y}$ results in $L \times W\times K$ size. Next, we attempt to model $\mathbfcal{Y}$ as $\mathbfcal{X}_{\mathbfcal{R}}+\mathbfcal{V}_{\mathbfcal{R}}+\mathbfcal{E}$, where $\mathbfcal{X}_{\mathbfcal{R}}$, $\mathbfcal{V}_{\mathbfcal{R}}$, and $\mathbfcal{E}$ are its constituent [*normal link tensor*]{}, [*abnormal link tensor*]{} and [*residual link tensor*]{} with the same size, respectively. Here, we model the normal link tensor $\mathbfcal{X}_{\mathbfcal{R}}$ as a low-rank subspace structure from the noisy traffic tensor $\mathbfcal{Y}$ in order to efficiently and robustly capture the change of underlying latent traffic structure of normal flows. For this purpose, particularly addressing the Candecomp/PARAFAC decomposition as a low-rank tensor approximation, we decompose the $\tau$-th frontal slice matrix of $\mathbfcal{X}_{\mathbfcal{R}}$ as $(\mathbfcal{X}_{{\mathbfcal{R}}})_{:,:,\tau} = \mat{A} {\rm diag}(\vec{b}^{\tau}) \mat{C}^T $, where $\mat{A} = [(\vec{a}^1)^T: \cdots: (\vec{a}^L)^T]^T$, $\mat{B} = [(\vec{b}^1)^T: \cdots : (\vec{b}^T)^T]^T$, and $\mat{C} = [(\vec{c}^1)^T : \cdots : (\vec{c}^W)^T]^T$ with $\{\vec{a}^l, \vec{b}^t,\vec{c}^w\} \in \mathbb{R}^{R}$ (Fig.\[Fig:BasicConcept\]-[iii]{}). See Appendix \[Append\_Sec:CPDEC\] for a brief introduction of the Candecomp/PARAFAC decomposition. Subsequently, we obtain the transformed problem formula of (\[Eq:Problem-LowRank-Sparsity-l1\]) by additionally considering the normal link tensor $\mathbfcal{X}_{\mathbfcal{R}}$ with the Hankel structure as well as the Candecomp/PARAFAC tensor decomposition as $$\label{Eq:HankelTensorProblemFormulation} \begin{split} \min_{\scriptsize{\mat{A},\mat{B},\mat{C},\mat{V}}} \ \ & \frac{1}{2} \| \underbrace{\mathbfcal{P}_{{\Omega}}}_{\rm Missing\ data} \underbrace{\bigl[{\mathbfcal{Y}} -(\mathbfcal{X}_{\mathbfcal{R}}+\mathbfcal{V}_{\mathbfcal{R}}) \bigr]}_{\rm Approximation\ error} \|_F^2 + \underbrace{\mu_r (\| \mat{A}\|_F^2 + \| \mat{B}\|_F^2 + \| \mat{C}\|_F^2)}_{\rm Frobenius-norm\ regularizer} +\!\! \underbrace{\mu_s \|\mat{V}\|_1,}_{\rm Sparsity\ regularizer}\\ {\rm s.t.}\ & \underbrace{(\mathbfcal{X}_{{\mathbfcal{R}}})_{:,:,\tau} = \mat{A} {\rm diag}(\vec{b}^{\tau}) \mat{C}^T}_{\rm Candecomp/PARAFAC\ constraint}, {\rm \ and} \!\! \underbrace{(\mathbfcal{X}_{{\mathbfcal{R}}})_{l,:,:} \in \mathcal{S}_H,}_{\rm Hankel\ structure\ constraint} \end{split}$$ where $\mathbfcal{Y} = \mathbfcal{H}(\mat{Y})$, $\mathbfcal{X}_{\mathbfcal{R}} = \mathbfcal{H}(\mat{RX})$, and $\mathbfcal{V}_{\mathbfcal{R}} = \mathbfcal{H}(\mat{RV})$, and $\mathbfcal{H}(\cdot)$ is the operator of the Hankelization of the matrix into the tensor structure. $\mathcal{S}_H$ represents the constraint of the linear subspace of all matrices with the Hankel structure. It should be noted that the Hankel structure constraint is placed on only $\mathbfcal{X}_{\mathbfcal{R}}$ in this paper for simplicity. In addition, the Frobenius-norm regularization offers a viable option for a batch-based low-rank tensor decomposition under the Candecomp/PARAFAC model [@Bazerque_IEEETransSP_2013]. Moreover, taking into account [*incomplete observation*]{} situation, the formulation in (\[Eq:HankelTensorProblemFormulation\]) additionally considers the support of the analysis with interpolating missing measurements (Fig.\[Fig:BasicConcept\]-[iv]{}), which is generally called the [*tensor completion*]{} problem. For this purpose, $\mathbfcal{P}_{{\Omega}}(\cdot)$ represents the operator to extract observation data, more precisely, $\mathbfcal{P}_{\Omega}(\mathbfcal{X})_{i_1,i_2,i_3}=\mathbfcal{X}_{i_1,i_2,i_3}$ if $(i_1, i_2, i_3) \in \Omega$ and $\mathbfcal{P}_{\Omega}(\mathbfcal{X})_{i_1,i_2,i_3}=0$ otherwise. $\Omega$ is a subset of the complete set of indices $\{(i_1, i_2, i_3): i_d \in \{1, \ldots, n_d \}, d \in \{1,2,3\}$. Finally, we consider an [*online-based*]{} setting of subspace learning, outlier learning, and anomaly detection method to prevent all measurements and model parameters in the past from being stored (Fig.\[Fig:BasicConcept\]-[v]{}). To this end, we tackle an [*online tensor completion*]{} problem. It should be noted that, as for the Hankel structure constraint of $\mathcal{S}_H$, this present paper considers [*only*]{} the two successive slice matrices to avoid model re-construction and diagonal-averaging using $\{\vec{b}^\tau\}_{\tau=t-W+1}^{\tau=t-1}$ in the past. Consequently, instead of (\[Eq:HankelTensorProblemFormulation\]), the final problem of our proposed method is further formulated to estimate the Candecomp/PARAFAC factor matrices $\{\mat{A}, \vec{b}, \mat{C}\}$ and the abnormal flow matrix $\mat{V}$ by considering the exponential weighted least squares cost function $$\label{Eq:Final_Problem_Definition} \begin{split} \min_{\scriptsize{\mat{A},\vec{b},\mat{C},\mat{V}}} \displaystyle{\frac{1}{2}}& \underbrace{\sum_{\tau=1}^t \lambda^{t-\tau}}_{\rm Online-based} \! \biggl[ \underbrace{{\Big\| \overbrace{{\rm\bf P}_{{\bf \Omega}_{\tau}}}^{\rm Missing\ data} \Bigl( \mat{Y}_{\tau} - \left(\mat{A} {\rm diag}(\vec{b}^\tau) \mat{C}^T \!+\! {\mat{V}_{{\scriptsize \mat{R}}_{\tau}}} \right) \Bigr) \Big\|_F^2}}_{\rm Approximation\ error\ by\ low-rank\ Candecomp/PARAFAC} \\ +&\ \mu_h [\tau] \Big\| \underbrace{{\rm\bf P}_{({\bf \Omega}_{\tau})_{:,1:W\!-\!1}}}_{\rm Missing\ data} \underbrace{\Bigl( (\mat{A} {\rm diag}(\vec{b}^{\tau\!-\!1}) \mat{C}^T)_{:,2:W} - \ (\mat{A} {\rm diag}(\vec{b}^\tau) \mat{C}^T)_{:,1:W\!-\!1} \Bigr) \Big\|_F^2}_{\rm Hankel\ structure\ error} \\ +& \underbrace{\bar{\mu}_r[\tau](\| \mat{A}\|_F^2 + \| \mat{C}\|_F^2) + \mu_r [\tau] \| \vec{b}^\tau \|_2^2}_{\scriptsize \rm Frobenius\ and\ \ell_2\ norm\ regularizer\ for\ \{\mat{A},\vec{b},\mat{C}\}}\ + \!\!\!\!\!\! \underbrace{\mu_s[\tau] \| \vec{v}[\tau] \|_1}_{\scriptsize \rm Sparsity\ regularizer\ for\ \mat{V}} \!\!\!\!\!\! \biggr], \end{split}$$ where ${\rm\bf P}_{{\bf \Omega}_\tau}(\cdot)$ is the matrix linear operator of the $\tau$-th frontal slice matrix of $\mathbfcal{P}_{{\Omega}}(\cdot)$. $\mat{Y}_\tau$ and $\mat{V}_{{\scriptsize \mat{R}_\tau}}$ are the $\tau$-th frontal slice matrices of $\mathbfcal{Y}$ and $\mathbfcal{V}_{\mathbfcal{R}}$, respectively. $\vec{v}[\tau]$ is the $W$-th column of $\mat{V}_{{\scriptsize \mat{R}_\tau}}$. $\bar{\mu}_r[\tau]=\mu_r[\tau]/\sum_{j=1}^\tau\lambda^{\tau-j}$, and we choose a lower $\mu_r$ when the data can be assumed to have a lower error rate, a relatively higher value otherwise. $0 < \lambda \leq 1$ is the so-termed [*forgetting factor*]{}. When $\lambda < 1$, data in the past are exponentially down-weighted, which facilitates tracking in non-stationary environments. In the case of infinite memory $\lambda = 1$, this coincides with the batch-based estimator. We finally perform the anomaly detection and identification. A larger value in $\vec{v}$ indicates that the probability of the existence of anomaly in its flow is higher. Therefore, by introducing a threshold variable $\delta_v$, the flows of which $\vec{v}$ is larger than $\delta_v$ are categorized into abnormal flows at time $t$. Solutions and Algorithm {#Sec:DetailedSolutionsandAlgorithm} ----------------------- This section gives the detailed solutions of the final minimization problem (\[Eq:Final\_Problem\_Definition\]), where unknown variables are $\vec{v}, \mat{A}, \mat{C}$, and $\vec{b}$. It is readily seen that this function is not convex. However, if $\mat{A}$, $\mat{C}$ and $\vec{b}$ are fixed, the problem becomes convex in $\vec{v}$. Similarly, if $\vec{v}$ fixed, we can refine our estimate of , and $\vec{b}$ in successive convex optimization steps. This paper uses an alternating minimization procedure to successively solve lower-dimensional convex problems by updating the unknown variables alternatively. ### Update of $[t]$ by LS (Fig.\[Fig:BasicConcept\]: I) {#Sec:UpdateOfB} We calculate $\vec{b}[t]$ via an $\ell_2$-norm regularized least squares (LS) problem, which has a closed-form solution. The estimate $\vec{b}[t]$ is obtained by calculating (\[Eq:Final\_Problem\_Definition\]) with fixed $\{ \mat{A}[t-1], \mat{C}[t-1] \}$ derived at time $t\!-\!1$, i.e., $$\begin{split} \label{eq:Problem_Definition_b_timebase} \vec{b}[t] = \defargmin_{\vec{b} \in \mathbb{R}^R} &\frac{1}{2} \Biggl[ \| {\bf \Omega}_t \circledast [ \mat{Y}_t - \mat{A}[t\!-\!1] {\rm diag}(\vec{b}) (\mat{C}[t\!-\!1])^T ] \|_F^2 \\ & + \mu_h[t] \| ({\bf \Omega}_t)_{:,1:W\!-\!1} \circledast \bigl[ (\mat{A}[t\!-\!1] {\rm diag}(\vec{b}[t\!-\!1]) \mat{C}[t\!-\!1]^T)_{:,2:W} \\ &- (\mat{A}[t\!-\!1] {\rm diag}(\vec{b}) \mat{C}[t\!-\!1]^T)_{:,1:W\!-\!1} \bigr] \|_F^2 + \mu_r[t] \| \vec{b} \|_2^2 \Biggr], \nonumber \\ \end{split}$$ where ${\bf \Omega}_t$ denotes a $L\times W$ binary $\{0,1\}$-matrix with $({\bf \Omega}_t)_{l,w}=1$ if $\mathbfcal{Y}_{l,w,t}$ is observed, and $({\bf \Omega}_t)_{l,w}=0$ otherwise. $\circledast$ represents the Hadamard product, i.e., the element-wise product of matrices. Defining $F[t]$ as the inner objective to be minimized, we obtain $\vec{b}[t]$ since $\vec{b}[t]$ satisfies $\partial F[t]/\partial \vec{b}[t] = 0$ as $$\begin{aligned} \label{Eq:b_final} \vec{b}[t] = & \biggl[ \mu_r[t] \mat{I}_R + \sum_{l=1}^L \sum_{w=1}^W ({\bf \Omega}_t)_{l,w} \vec{g}_{l,w}[t] (\vec{g}_{l,w}[t])^T + \mu_h[t] \sum_{l=1}^{L} \sum_{w=1}^{W-1} ({\bf \Omega}_t)_{l,w} (\vec{g}_{l,w}[t]) (\vec{g}_{l,w}[t])^T \biggr]^{-1} \nonumber \\ & \biggl[ \sum_{l=1}^L \sum_{w=1}^W ({\bf \Omega}_t)_{l,w} (\mat{Y}[t])_{l,w} \vec{g}_{l,w}[t] + \mu_h[t] \sum_{l=1}^{L} \sum_{w=1}^{W-1} ({\bf \Omega}_t)_{l,w} (\vec{g}_{l,w+1}[t])^T \vec{b}[t\!-\!1] \vec{g}_{l,w}[t] \biggr],\end{aligned}$$ where $\vec{g}_{l,w}[t] = \vec{a}^l[t-1] \circledast \vec{c}^w[t-1] \in \mathbb{R}^R$. ### Update of $\vec{v}[t]$ based on ADMM (Fig.\[Fig:BasicConcept\]: II) {#Sec:UpdateOfV} $\vec{v}[t]$ is solved by the [*alternating direction method of multipliers*]{} (ADMM), which solves convex optimization problems by separating them into smaller sub-problems [@Boyd_FTML_2011]. It has recently gained big attention in wide applications in a number of areas. Thus, $\vec{v}[t]$ is obtained by solving the reformulated (\[Eq:Final\_Problem\_Definition\]) by addressing only the last column of each frontal slice matrix as $$\begin{split} \label{eq:} \vec{v}[t]& = \defargmin_{\scriptsize{\vec{v}}} {\frac{1}{2} \| {\bf \Omega}_t \circledast \Bigl( \mat{Y}_t } - \mat{A}[t\!-\!1] {\rm diag}(\vec{b}[t]) (\mat{C}[t\!-\!1])^T- {\mat{V}_{\scriptsize \mat{R}}}_t \Bigr) \|_F^2 +\ \mu_s[t] \| \vec{v} \|_1 \nonumber\\ & = \defargmin_{\scriptsize{\vec{v}}} \frac{1}{2} \| \vec{q}[t] - \mat{R}_{\omega_{t, W}}\vec{v} \|_2^2 +\ \mu_s[\tau] \| \vec{v} \|_1, \end{split}$$ where $\vec{q}[t] \!=\! ({\bf \Omega}_t)_{:,W} \!\circledast\! \bigl[ (\mat{Y}_t)_{:,W} \!-\!(\mat{A}[t\!-\!1] {\rm diag}(\vec{b}[t]) (\mat{C}[t\!-\!1])^T)_{:,W}\bigr] \in \mathbb{R}^L$, and $({\bf \Omega}_t)_{:,W}$ has only the last column of ${\bf \Omega}_t$. Furthermore, the routing matrix $\mat{R}_{\omega_{t, W}} \in \mathbb{R}^{L \times F}$, which corresponds to only observed measurements, is calculated as $\mat{R}_{\omega_{t, W}}={\rm diag}(({\bf \Omega}_t )_{:,W})\mat{R}$. Thereby, this problem can be re-written as $\min f(\vec{v}) + g(\vec{z}) \ {\rm s.t.} \ \vec{v} - \vec{z} = 0$, where $f(\vec{v}) = (1/2) \| \vec{q}[t] - \mat{R}_{\omega_{t, W}}\vec{v} \|_2^2$, and $g(\vec{z}) = \mu_s[t] \| \vec{z} \|_1$. It should be also noted that the iteration index [*t*]{} is the [*outer*]{} loop index, and this is kept fixed at the ADMM loop, i.e., the [*inner*]{} loop, where a new index [*k*]{} is used instead. The augmented Lagrangian of this constrained minimization problem is expressed as $\mathcal{L}_{\xi}(\vec{v}, \vec{z}, \vec{y}) = f(\vec{v}) + g(\vec{z}) + \vec{y}^T(\vec{v} -\vec{z}) + (\xi/2) \| \vec{v} -\vec{z} \|^2_2$, where is the dual vector. Denoting $\vec{u} = (1/\xi) \vec{y}$ as the [*scaled dual variable*]{}, the sub-problems of ADMM become \[Eq:v\_final\_a\] \^[k+1]{} = ( \_[\_[t, W]{}]{}\^T \_[\_[t, W]{}]{} + \_[F]{})\^[-1]{} (\_[\_[t, W]{}]{}\^T \[t\] + (\^[k]{} -\^[k]{} ))     &\ \[Eq:v\_final\_b\] \^[k+1]{} = S\_[\_s\[t\]/]{} (\^[k+1]{} + \^[k+1]{}) &\ \[Eq:v\_final\_c\] \^[k+1]{} = \^[k]{} + \^[k+1]{} - \^[k+1]{},& where $\xi > 0$ and $S_{\kappa}(a)$ is the [*soft thresholding operator*]{} that is defined as $S_{\kappa}(a) = {\rm sign}(a)(|a|-\kappa)_{+}$, namely, $S_{\kappa}(a)=0$ if $|a|\leq \kappa$, otherwise $S_{\kappa}(a)={\rm sign}(a)(|a|-\kappa)$. Finally, we obtain $\vec{v}^{k+1}$ as $\vec{v}[t]$. The overall algorithm for $\vec{v}[t]$ is summarized in [**Algorithm \[Alg:Admm\]**]{}. $\vec{v}[t]=\vec{v}^{k+1}, \vec{z}[t]=\vec{z}^{k+1}, \vec{y}[t]=\vec{y}^{k+1}$. ### Update of $[t]$ and $[t]$ by RLS (Fig.\[Fig:BasicConcept\]: III) {#Sec:UpdateOfAandC} The calculation of $\mat{C}[t]$ requires $\mat{A}[t\!-\!1]$, and the calculation of $\mat{A}[t]$ uses $\mat{C}[t]$. This paper addresses a second-order stochastic gradient based on the RLS method with forgetting parameters, which has been widely used in tracking of time varying parameters in many fields. Its computation is efficient since we update the estimates recursively every time new data becomes available. First, the problem (\[Eq:Final\_Problem\_Definition\]) is reformulated to obtain $\mat{A}[t]$ as $$\begin{split} \label{eq:Problem_Definition_A} \min_{\scriptsize \mat{A}}& \frac{1}{2} \sum_{\tau=1}^t \lambda^{t-\tau} \biggl[ \| {\bf \Omega}_\tau \circledast \bigl[ \mat{Y}_\tau \!\!-\!\! (\mat{A} {\rm diag}(\vec{b}[\tau]) (\mat{C}[\tau\!\!-\!\!1])^T \!\!-\!\! {\mat{V}_{\scriptsize \mat{R}}}_\tau ) \bigr] \|_F^2 \\ &+ \| ({\bf \Omega}_\tau)_{:,1:W\!-\!1} \circledast \bigl[ (\mat{A} {\rm diag}(\vec{b}[\tau\!-\!1]) \mat{C}[\tau\!\!-\!\!1]^T)_{:,2:W} \\ &- (\mat{A} {\rm diag}(\vec{b}[\tau]) (\mat{C}[\tau\!\!-\!\!1])^T)_{:,1:W\!-\!1} \bigr] \|_F^2 \biggr]+ \frac{\mu_r[t]}{2} \| \mat{A}\|_F^2. \end{split}$$ The objective function in (\[eq:Problem\_Definition\_A\]) is decomposed into a parallel set of smaller problems, one for each row $\vec{a}^l \in \mathbb{R}^R$ of $\mat{A}$. By denoting $\mat{Y}_\tau-{\mat{V}_{\scriptsize \mat{R}}}_\tau$ as $\mat{Z}_\tau$, we obtain $\vec{a}^l[t]$ as $$\begin{split} \label{eq:problem_def_am} \min_{\vec{a}^l \in \mathbb{R}^R} & \frac{1}{2} \sum_{\tau=1}^t \Biggl[ \sum_{w=1}^W \lambda^{t-\tau} \!\! \left( ({\bf \Omega}_{\tau})_{l,w} \! \left((\mat{Z}_\tau)_{l,w} \!\!-\!(\vec{a}^l)^T {\rm diag} (\vec{b}[\tau]) \vec{c}^w[\tau\!\!-\!\!1]\right) \right) ^2 \\ &+\ \mu_h \sum_{w=1}^{W\!-\!1} \lambda^{t-\tau} \left(({\bf \Omega}_{\tau})_{l,w} \left( (\vec{a}^l)^T {\rm diag} (\vec{b}[\tau\!\!-\!\!1]) \vec{c}^{w+1}[\tau\!\!-\!\!1] - (\vec{a}^l)^T {\rm diag} (\vec{b}[\tau]) \vec{c}^w[\tau\!\!-\!\!1]\right)\right)^2 \Biggr] \nonumber \\ &+\ \frac{\mu_r[t]}{2}\| \vec{a}^l\|_2^2. \end{split}$$ Here, denoting ${\rm diag}(\vec{b}[\tau])\vec{c}^w[\tau\!-\!1]$ and ${\rm diag}(\vec{b}[\tau\!-\!1]) \vec{c}^{w+1}[\tau\!-\!1]-{\rm diag}(\vec{b}[\tau]) \vec{c}^w[\tau\!-\!1]$ as $\vec{\alpha}_w[\tau] \in \mathbb{R}^R$ and $\vec{\beta}_w[\tau] \in \mathbb{R}^R$, respectively, $\vec{a}^l[t]$ is obtained by setting the derivative of (\[eq:problem\_def\_am\]) with regard to $\vec{a}^l$ equal to zero. $$\begin{aligned} \label{Eq:al_final} \vec{a}^l[t] &=& \vec{a}^l[t\!-\!1] \!-\! (\mat{RA}_l[t])^{-1} \Bigl( \mu_h\sum_{w=1}^{W\!-\!1} ({\bf \Omega}_t)_{l,w+1} \vec{\beta}_w[t] (\vec{\beta}_w[t])^T + (\mu_r[t] - \lambda \mu_r[t\!-\!1]) \mat{I}_R \Bigr) \vec{a}^l[t\!-\!1] \nonumber \\ &&+\ (\mat{RA}_l[t])^{-1} \sum_{w=1}^W ({\bf \Omega}_t)_{l,w} \left((\mat{Y}_t)_{l,w} - ({\mat{V}_{\scriptsize \mat{R}}}_t)_{l,w} - (\vec{\alpha}_w[t])^T \vec{a}^l[t\!-\!1] \right) \vec{\alpha}_w[t], \end{aligned}$$ where $\mat{RA}_l[t]$ is calculated as $$\begin{array}{lll} \label{eq:Update_RA} \mat{RA}_l[t] &=& \lambda \mat{RA}_l[t\!-\!1] + \displaystyle{\sum_{w=1}^W ({\bf \Omega}_t)_{l,w+1} \vec{\alpha}_w[t] \vec{\alpha}_w[t]^T} \\ &&+\ \displaystyle{\mu_h \sum_{w=1}^{W\!-\!1} ({\bf \Omega}_t)_{l,w} \vec{\beta}_w[t] \vec{\beta}_w[t]^T + (\mu_r[t] \!-\! \lambda \mu_r[t\!-\!1] ) \mat{I}_R}. \end{array}$$ The derivations of (\[Eq:al\_final\]) and (\[eq:Update\_RA\]) are detailed in Appendix \[Append\_Sec:RLS\]. Meanwhile, as for $\mat{C}[t]$, the problem (\[Eq:Final\_Problem\_Definition\]) is reformulated as $$\begin{split} \label{Eq:problem_def_cw} \min_{\vec{c}^w \in \mathbb{R}^R} & \frac{1}{2} \sum_{\tau=1}^t \Biggl[ \sum_{l=1}^L \lambda^{t-\tau} \!\! \left(({\bf \Omega}_{\tau})_{l,w} \!\left((\mat{Z}_\tau)_{l,w} \!\!-\!\! (\vec{a}^l[\tau])^T {\rm diag} (\vec{b}[\tau]) \vec{c}^w \right) \right) ^2 \\ &+\ \mu_h \sum_{l=1}^{L} \lambda^{t-\tau} \left(({\bf \Omega}_{\tau})_{l,w} \left( (\vec{a}^l[\tau])^T {\rm diag} (\vec{b}[\tau\!\!-\!\!1]) \vec{c}^{w+1}[\tau] - (\vec{a}^l[\tau])^T {\rm diag} (\vec{b}[\tau]) \vec{c}^w \right)\! \right)^2 \Biggr] \\ &+ \frac{\mu_r[t]}{2}\| \vec{c}^w\|_2^2.\nonumber \end{split}$$ It should be emphasized that the second term representing the Hankel structure error is not included when $w=W$. In addition, since the second term needs $\vec{c}^{w+1}[\tau]$, this calculation cannot be performed in parallel, and the order of the calculations follows the [*descending order*]{} of $w$. Finally, $\vec{c}^w[t]$ can be given by denoting $(\vec{a}^l[\tau])^T {\rm diag}(\vec{b}[{\tau}])$ and $(\vec{a}^l[\tau])^T {\rm diag}(\vec{b}[\tau\!-\!1])$ as $\vec{\gamma}_{l}[\tau] \in \mathbb{R}^{1\times R}$ and $\vec{\eta}_{l}[\tau] \in \mathbb{R}^{1\times R}$, respectively, as $$\begin{aligned} \label{Eq:cw_final} \vec{c}^w[t] &=& \vec{c}^w[t\!-\!1] - (\mat{RC}_w[t])^{-1} (\mu_r[t] - \lambda \mu_r[t-1]) \vec{c}^w[t\!-\!1] \nonumber \\ & &+\ (\mat{RC}_w[t])^{-1} \sum_{l=1}^L({\bf \Omega}_t)_{l,w} \biggl( \Bigl((\mat{Y}_t)_{l,w} - ({\mat{V}_{\scriptsize \mat{R}}}_t)_{l,w} \bigr. \Biggr. \nonumber\\ & &+\ \biggl. \Bigl. \mu_h \vec{\eta}_{l}[t] \vec{c}^{w+1}[t] \Bigr) \mat{I}_{R} - (1+\mu_h)\vec{c}^w[t\!-\!1] \vec{\gamma}_{l} [t] \biggr) (\vec{\gamma}_{l}[t])^T, \end{aligned}$$ where $\mat{RC}_w[t]$ is transformed as $$\begin{aligned} \label{eq:Update_RC} \mat{RC}_w[t] & = \lambda \mat{RC}_w[t-1] +\sum_{l=1}^L ({\bf \Omega}_t)_{l,w}(1+\mu_h)\vec{\gamma}_{l} [t] (\vec{\gamma}_{l} [t])^T + (\mu_r[t] - \lambda \mu_r[t\!-\!1]) \mat{I}_R.\end{aligned}$$ The overall algorithm to solve (\[Eq:Final\_Problem\_Definition\]) is finally summarized in [**Algorithm \[Alg:overall\]**]{}. Computational Complexity Analysis {#Sec:ComputationalComplexityAnalysis} --------------------------------- This section analyzes the computational complexity per iteration of the proposed algorithm. The calculation of $\vec{b}[t]$ in Section \[Sec:UpdateOfB\] requires $\mathcal{O}(|{\bf \Omega}_t| R^2)$ in (\[Eq:b\_final\]), where $|{\bf \Omega}_t|$ is the number of known entries in ${\bf \Omega}_t$. The calculations of $\mat{A}[t]$ and $\mat{C}[t]$ in Section \[Sec:UpdateOfAandC\] require $\mathcal{O}(L R^3)$ for (\[Eq:al\_final\]) and $\mathcal{O}(W R^3)$ for (\[Eq:cw\_final\]), respectively, for the inversion of $\mat{RA}$ in (\[eq:Update\_RA\]) and $\mat{RC}$ in (\[eq:Update\_RC\]). As for the calculation of $\vec{v}[t]$ in \[Sec:UpdateOfV\], the inversion of $(\mat{R}_{\omega_{t, W}}^T \mat{R}_{\omega_{t, W}} + \xi \mat{I}_{F})$ can be done efficiently by exploiting the [*matrix inversion lemma*]{}, which states that $(\mat{P}+\xi \mat{A}^T \mat{A})^{-1} = \mat{P}^{-1} \mat{A}^T(\mat{I} + \xi \mat{A} \mat{P}^{-1} \mat{A}^T)^{-1} \mat{A}\mat{P}^{-1} $. This leads to $\mathcal{O}(|({\bf \Omega}_t)_{:,W}|F^2)$, where $|({\bf \Omega}_t)_{:,W}|$ is the number of known values at the $t$-th iteration. Then, the total calculation in \[Sec:UpdateOfV\] needs at most $\mathcal{O}((K+1)|({\bf \Omega}_t)_{:,W}|F^2)$ due to this one-time inversion and $K$-times multiplications for all inner iterations in (\[Eq:v\_final\_a\]), where $K$ is the maximum number of inner iterations in [**Algorithm 1**]{}. Thus, the total computational complexity at $t$-th iteration in [**Algorithm 2**]{} results in $\mathcal{O}(|{\bf \Omega}_t| R^2+(L+W)R^3+(K+1)|({\bf \Omega}_t)_{:,W}|F^2))$, and reveals that the number of flows, $F$, is dominant since rank $R$ is assumed to be low-rank. Numerical Evaluation {#Sec:NumericalEvaluations} ==================== We show numerical comparisons of the proposed algorithm with state-of-the-art algorithms for synthetic and real-world datasets. All the following experiments are done on a PC with 3.0 GHz Intel Core i7 CPU and 16 GB RAM. The synthetic anomalies are injected onto the synthetic and real-world datasets to be evaluated. More concretely, as for the real-world dataset, the original signal in the dataset is firstly smoothed, and the synthetic anomalies are injected onto it. Regarding the synthetic dataset, the synthetic anomalies are injected onto the synthetic signal that includes the seasonal signal, the periodic signals and the noise. We use the Receiver Operating Characteristic (ROC) and the F-measure value as the evaluation metrics. The ROC evaluates a binary classifier, which plots [*true positive rate*]{} against [*false positive rate*]{} at various [*discrimination thresholds*]{}. In this case, this thresholds correspond to $\delta_v$ in Section \[Sec:DerivationofProblemFormulation\]. F-measure effectively references the true positives to the arithmetic mean of the predicted positives and the real positives, which is calculated as the harmonic mean between precision and recall. As for the comparison algorithms, the EWMA algorithm and the Wavelet-based algorithm proposed in [@Zhang_ACM_IMC_2005] are compared with the proposed algorithm. They provide a unified frame for integrating anomaly detection approaches and the inference techniques for anomaly identifications. The former is one representative method of the time-series modeling, and the latter is one of the signal processing based methods. But the latter is a batch-based method. It should be noted that the standard methods of these do not handle [*missing data*]{}. Therefore, the latest observed data corresponding a missing datum is interpolated to perform fair comparison. In addition, three subspace-based tracking algorithms, which are GROUSE, GRASTA, and PETRELS, are compared. We use Matlab codes provided by the respective authors. It is important to note that since these algorithms do not have the abnormal flow detection function, we have newly integrated the function defined in (\[Eq:SparsityMaximization\]) onto them. For the proposed algorithm, we use $\lambda=0.9, \mu_r=\mu_h=10^{-3}$ and $\mu_s=10^{-2}\times \max({\rm abs}(\mat{R}_{\omega_{t, W}}\vec{v}))$ for synthetic datasets, and $\mu_s=10^{-1}\times \max({\rm abs}(\mat{R}_{\omega_{t, W}}\vec{v}))$ for real-world datasets, where ${\rm abs}(\vec{a})$ returns the absolute value of each element in . $K$, $\epsilon^{abs}$, and $\epsilon^{rel}$ for ADMM are $120$, $10^{-5}$, and $10^{-3}$, respectively. Evaluation Methodology and Synthetic Anomaly Injection {#Sec:SyntheticAnomalyInjection} ------------------------------------------------------ The evaluation of any anomaly detection algorithms always faces the issue how to obtain [*ground truth*]{} because no public and reliable ground truth is available in real-world datates [@Soule_IMC_2005]. One popular way is that a security expert labels anomalies by manual inspection against collected live traffic traces. However, this is very expensive and time-consuming if datasets are large. In addition, this is not a perfect solution because the operator could make mistakes that miss an anomaly or generate a false positive anomaly. Furthermore, because such traces include a limited number of anomalies, comprehensive performance evaluations, which evaluate all algorithm capabilities, are difficult to perform. Instead, an alternative approach is to [*inject synthetic anomalies*]{}, which correspond to the ground truth, onto data signals. One advantage of this is to be able to change the [*anomaly parameters*]{} assuming various anomalies. In other words, the flexible configurability of anomaly parameters as explained below allows us to simulate a wide range of anomalies that cannot be found in real-world datasets, and this shall enable us to achieve comprehensive evaluations of the anomaly detection and identification performance of algorithms. Parameters Settings Description -------------------------------------------------------- ----------------------------------------------------------------------------------- ----------------------------------- Multiplicative ratio of volume change (${\it \delta}$) $1.5 \leq \delta \leq 2.5$ DDoS, alpha event ${\it \delta} = 0$ Outrage Duration (${\it d}$) ${\it d} = \{5,10,20,30\}$ (mins) Increase ratio (${\it \gamma_{i}}$) Change time ratio against entire duration (${\it d}$): $0 \leq {\it \gamma}< 0.5$ Gradual or sudden up/down changes Decrease ratio (${\it \gamma_{d}}$) Target OD flows (Number of flows) N-1: One flow between one source (src) and one destination (dst) DDoS, alpha N-1: N flows between srcs/dess and one src/des DDoS All-ODs-one-link: All ODs passing one link Outrage : Synthetic anomaly parameters.[]{data-label="Tbl:SyntheticAnomalyInjectionCofig"} ![Illustration of the anomaly parameters.[]{data-label="Fig:SyntheticAnomalyInjectionCofig"}](figures_final/anomaly_parameters_new3.eps){width="\linewidth"} Hereafter, the anomaly parameters are detailed by referring the conceptual illustration shown in Fig.\[Fig:SyntheticAnomalyInjectionCofig\], and the summarization in Table \[Tbl:SyntheticAnomalyInjectionCofig\]. It should be noted that this paper focuses on anomalies that bring [*changes of volume patterns*]{}, e.g. DDoS attacks, alpha events, outrages, and does not assume worms and port scans. As for the [*duration of an anomaly*]{} $d$, most DDoS attacks are observed to continue between $5$ and $30$ minutes, and some outliers last less than $1$ minute and others last several days [@Moore_USS_2001]. The DDoS attacks in the Abilene Network dataset last less than $20$ minutes, and some outliers continue more than $2$ hours. The alpha events could be of any length. However, because we cannot simulate all outliers, this paper configures the duration $d$ as $\{5,10,20,30\}$ minutes. We mimic the [*traffic volume change*]{}, when anomalies occur, by introducing a multiplicative factor $\delta$ that is multiplied by the original traffic of a target OD flow. Adding $20\%$, for example, of the original OD flow volume is simulated by using $\delta=1.2$. $\delta=0$ is used to capture the outage scenarios. We can mimic a variety of either the DDoS attacks or the alpha events by allowing $1.0 \leq \delta \leq 2.0$. We do not consider $\delta > 2.5$ because such changes are clearly irregular. We also address the [*traffic shape*]{} at the beginning and ending of anomalies. The initial rise of the DDoS attacks could be simulated by a ramp shape. The outage anomalies show an almost sudden drop in volume like a square shape. The alpha events indicate either an almost sudden rise or a ramp-shape increase. Thus, we introduce the increase ratio parameter ($\gamma_{i}$) and the decrease ratio parameter ($\gamma_{d}$) to express these shapes. For both parameters, we use $0 \leq \gamma< 0.5$ against the entire duration $d$. Finally, we address the [*flow structure*]{}, i.e., [*combination of the number of sources and destinations*]{}, that has an influence on the structure of the OD flows in an entire network. We denote “1-1" as the OD flow that traverses from one single source to one single destination. This could occur with the DDoS attacks or the alpha events. “N-1" refers to the OD flows between N-sources and one single destination, which could happen with the DDoS attacks. “All-ODs-one-link" corresponds to all the OD flows that pass one particular single link. Real-world Dataset Evaluations ------------------------------ We use the Abilene Network Dataset for the evaluation of the proposed method on real data. Abilene Network is the Internet2 backbone network in the US. It has 11 Points of Presence, where there are $121$ OD flows and $30$ links. The Abilene Network Dataset samples 2016 samples per week, and 5-minutes sampled traffic matrices are collected via Netflow. Each element of a generated flow matrix corresponds to a single OD flow over time with 5 minute increments. The procedure to generate anomaly-injected real-world dataset is explained by following Section \[Sec:SyntheticAnomalyInjection\]. An example of the overall procedure is depicted in Fig.\[fig:AnomalyInjectionAbilene\], where the case of the $50$th flow of the Abilene Network Dataset is illustrated. The long-term statistical [*trend*]{} from an original OD flow (Fig.\[fig:AnomalyInjectionAbilene\]: 1st graph) is extracted by [*smoothing*]{} the original signal (2nd graph). This is achieved by approximating the extracted signal by 5-th Wavelet levels by Daubechies-5 mother wavelet with $5$ levels because these underlying trends are generally non-stationary. Next, a Gaussian noise with zero mean (3rd graph) is added onto the smoothed, i.e., de-noised signal, where the distribution variance is calculated using the first 5 detailed signals (4th graph). Finally, injecting one of the anomalies (5th graph) onto smoothed noisy signal, we obtain the final anomaly-injected noisy signal (6th graph). ![Anomaly injection procedure in Abilene (50th flow).[]{data-label="fig:AnomalyInjectionAbilene"}](figures_final/AnomalyInjectionProcess_new.eps){width="\linewidth"} We evaluate the case where the sampling frequency is every $5$ minutes, and observation ratio $\rho$ is $30$. Since the number of nodes, $N_{node}$, is $11$, $L=30$ and $F=12$. $W$ is $288$, which corresponds to $1$ day, and $T=2016$ (= $1$ week). The ratio of the anomaly-injected flows is $1.54$%. The results of the modeling residual error and the ROC curve are shown in Fig.\[fig:Real\_OR\_30\]. From the residual error in Fig.\[fig:Real\_OR\_30\](a), the error of Wavelet is much lower than others because, in the Wavelet algorithm, the signal with higher frequency signal is removed and the residual data is not produced by its constructed model. In addition, although the EWMA algorithm constructs a parametrized model, the number of model parameters are much more than that of the subspace-based algorithms. Among the subspace-based algorithms, the proposed algorithm shows the lowest errors. The convergence speed of PETREL is faster than that of GROUSE and GRASTA because PETRELS has a second-order convergence property. On the other hand, the proposed algorithm shows much faster convergence characteristics than that of PETRELS. As for the ROC carve in Fig.\[fig:Real\_OR\_30\](b), the proposed algorithm also outperforms the other algorithms. {width="0.5\hsize"} [(a) Residual error]{} \[\] {width="0.5\hsize"} [(b) ROC]{} \[\] {width="0.5\hsize"} \[fig:Real\_OR\_50\] The result of the ROC for an observation ratio $\rho$ of $50$ is also shown in Fig.\[fig:Real\_OR\_50\]. This also shows the superior performance of the proposed algorithm against the other algorithms. Synthetic Datasets Evaluations ------------------------------ As mentioned earlier, we use the synthetic datasets for comprehensive evaluations. ### Network Generation {#sec:Network Generation} The network used as an input for the simulation is generated in a random fashion. A number of nodes $N_{node}$ are randomly placed on a 2D space, and directly-connecting links among those nodes are calculated by the [*Delaunay triangulation*]{} algorithm, which maximizes the minimum angle of the triangles. Then, $F$ traffic flows are generated between $F$ pairs of nodes that are randomly selected. Each traffic flow travels its shortest path between its source and destination nodes, which is calculated by the [*Dijkstra*]{} algorithm. ### Anomaly-injected Traffic Generation We generate a virtual flow at time $t$ for the $i$-th node pair by [*injecting*]{} synthetic anomalies as $\vec{f}_{i}(t)=\vec{f}_{i}^{(po)}(t) + \vec{f}_{i}^{(sea)}(t) + \vec{f}_{i}^{(ano)}(t) + \vec{f}_{i}^{(noise)}(t)$ by following Section \[Sec:NetworkAnomography\] [@Kim_CAMSAP_2009]. It is noted that $\vec{f}_{i}^{(no)}(t)$ in Section \[Sec:NetworkAnomography\] is further decomposed into the [*periodic component*]{} $\vec{f}_{i}^{(po)}(t)$ and the [*seasonal trend component*]{} $\vec{f}_{i}^{(sea)}(t)$, where $\vec{f}_{i}^{(po)}(t)$ is generated as $\vec{f}_{i}^{(po)}(t)=A_1 \sin(\omega t)$, and three types of signals are mixed equally for $\vec{f}_{i}^{(sea)}(t)$ as $$\begin{aligned} \vec{f}_{i}^{(sea)}(t) = \left\{ \begin{array}{lr} 0 & {\rm (no\ trend)},\\ b_1 t & {\rm (linear\ trend)},\\ A_2/b_2 \sin(7\omega t) & {\rm (weekly\ sine\ wave\ trend)}.\nonumber \end{array} \right.\end{aligned}$$ $\vec{f}_{i}^{(noise)}(t)$ is created as $\vec{f}_{i}^{(noise)}(t) \sim \mathcal{N}(0, \sigma^2)$, which is a Gaussian noise with a zero-mean and $\sigma^2$ variance. Lastly, we inject the synthetic anomalies $\vec{f}_{i}^{(ano)}(t)$ onto the above $\vec{f}_{i}(t)$, i.e., $\mat{F}_{:,t}$ as explained in Section \[Sec:SyntheticAnomalyInjection\]. The final link matrix for $L$ link-pairs at each time $t$ is calculated as $\mat{Y}_{:,t} = \mat{R} \mat{F}_{:,t}$. ### Experimental Results in Small-size Network We first consider a small-size network with $N_{node}=500$, where $L$ is $2958$, and $F$ is $5 \times 10^4$. $W$ is $24$, which corresponds to $1$ day, and $T$ is $168$, i.e., $1$ week. The data is sampled every hour, and the observation ratio $\rho$ is $30$. The ratio of injected anomalies is $1.81\times 10^{-2}$%. The results are shown in Fig.\[fig:Synthetic\_SmallSize\_OR\_30\]. ![Residual error and ROC (small-size network, $\rho=30$).[]{data-label="fig:Synthetic_SmallSize_OR_30"}](results_final/{Syn_Injection-error_Mixture-500-50000-0.00010-10-30}.eps){width="0.5\hsize"} [(a) Residual error]{} ![Residual error and ROC (small-size network, $\rho=30$).[]{data-label="fig:Synthetic_SmallSize_OR_30"}](results_final/{Syn_Injection-ROC_Mixture-500-50000-0.00010-10-30}.eps){width="0.5\hsize"} [(b) ROC]{} ![Residual error and ROC (small-size network, $\rho=30$).[]{data-label="fig:Synthetic_SmallSize_OR_30"}](results_final/{Syn_Injection-f-measure_Mixture-500-50000-0.00010-10-30}.eps){width="0.5\hsize"} [(C) F-Measure]{} It should be firstly noted that the starting index of the proposed algorithm is delayed by $24$ data stream indices because it has the Hankel structure with the $W$-lagged data as seen in Fig.\[fig:Synthetic\_SmallSize\_OR\_30\]. Thus, while the first issue we observe is the convergence speed at the beginning of subspace algorithms, that of the proposed algorithm is the fastest as can be seen in Fig.\[fig:Synthetic\_SmallSize\_OR\_30\](a). The next observation is that the proposed algorithm indicates the highest ROC values in the entire range as can be seen in Fig.\[fig:Synthetic\_SmallSize\_OR\_30\](b). The F-measure value in Fig.\[fig:Synthetic\_SmallSize\_OR\_30\](c) also yields the superior performance of the proposed algorithm. Furthermore, we evaluate the case with $\rho=50$, where the parameter configurations are the same as the case in $\rho=30$. The results in Fig.\[fig:Synthetic\_SmallSize\_OR\_50\] are very similar to the results with the observation ratio $\rho$ of $30$. ![ROC (small-size network, $\rho=50$).[]{data-label="fig:Synthetic_SmallSize_OR_50"}](results_final/{Syn_Injection-ROC_Mixture-500-50000-0.00010-10-50}.eps){width="0.5\hsize"} [cc]{} ![ROC (Different window lengths).[]{data-label="fig:Synthetic_DifferentWindowLentgh"}](results_final/{Syn_Injection-ROC_N-1-500-50000-0.00010-10-30}.eps){width="\hsize"} [(a) $\rho=30$]{} \[\] ![ROC (Different window lengths).[]{data-label="fig:Synthetic_DifferentWindowLentgh"}](results_final/{Syn_Injection-ROC_N-1-500-50000-0.00010-10-50}.eps){width="\hsize"} [(b) $\rho=50$]{} \[fig:fall\] [cc]{} ![ROC (Different window lengths).[]{data-label="fig:Synthetic_DifferentWindowLentgh"}](results_final/{Syn_Injection-ROC-diffwinlen_Mixture-500-50000-0.00010-10-30}.eps){width="\hsize"} [(a) $\rho=30$]{} \[\] ![ROC (Different window lengths).[]{data-label="fig:Synthetic_DifferentWindowLentgh"}](results_final/{Syn_Injection-ROC-diffwinlen_Mixture-500-50000-0.00010-10-50}.eps){width="\hsize"} [$\rho=50$]{} \[fig:fall\] We also evaluate the behavior of the algorithms with varying types of the injected anomalies. This especially focuses on the difference between “Mixture" case and “Only N-1" case. The “Mixture" case includes “1-1", “N-1", and “All-ODs-one-link" types of abnormal flows in Table \[Tbl:SyntheticAnomalyInjectionCofig\]. To the contrary, the “Only N-1" case contains only “N-1" types of abnormal flows. The other configurations are the same as the first evaluation experiments. The results when $\rho=30$ and $\rho=50$ are shown in Figs.\[fig:Synthetic\_DifferentAnomalyType\](a) and (b), respectively. Although we cannot directly compare these with Fig.\[fig:Synthetic\_SmallSize\_OR\_30\](b) and Fig.\[fig:Synthetic\_SmallSize\_OR\_50\] because the incomplete data position and the number and positions of anomaly flows are completely different, the results with the “Only N-1" case of GROUSE, GRASTA, Wavelet and the EWMA are similar with the “Mixture" case. Meanwhile, whereas PETRELS decreases largely, the proposed algorithm remains the similar performance ($\rho=30$) or slightly decreases ($\rho=50$). This is due to the fact that the proposed algorithm can capture structure changes in particular. The abnormal flows in the “Only N-1" case have a more biased structure than those in the “Mixture" case, where the constituent flows are fairly distributed among nodes and links across the entire network. Finally, the impact on the different window lengths $W=\{12,18,24,30,36,48\}$ for the Hankel structure are evaluated in Fig.\[fig:Synthetic\_DifferentWindowLentgh\]. The other configurations are the same as the first evaluation experiments. Fig.\[fig:Synthetic\_DifferentWindowLentgh\] reveals that $W=24$ yields the best performance in both $\rho=30$ and $\rho=50$ cases because the window length matches with the traffic periodicity. ### Experimental Results in Mid-size and Large-size Networks We further consider a mid-size network and a large-size network with the same number of flows $F=10^5$. For the mid-size network, $N_{node}$ is $1000$, $L$ is $11946$, and the ratio of injected anomalies is $1.59 \times 10^{-2}$%. Meanwhile, the large-size network has $N_{node}$ is $3000$, $L$ is $17931$, and the ratio of injected anomalies is $1.93\times 10^{-2}$%. Figs.\[fig:Synthetic\_MiddleSize\_OR\_30\] and \[fig:Synthetic\_LargeSize\_OR\_30\] show the results for the mid-size and the large-size networks when $\rho$ is $30$, respectively. These figures yield the superior performance of the proposed algorithm against other algorithms. [cc]{} ![ROC and F-measure in large-size network.[]{data-label="fig:Synthetic_LargeSize_OR_30"}](results_final/{Syn_Injection-ROC_Mixture-2000-100000-0.00010-10-30}.eps){width="\hsize"} [(a) ROC]{} \[\] ![ROC and F-measure in large-size network.[]{data-label="fig:Synthetic_LargeSize_OR_30"}](results_final/{Syn_Injection-f-measure_Mixture-2000-100000-0.00010-10-30}.eps){width="\hsize"} [(b) F-measure]{} \[\] [cc]{} ![ROC and F-measure in large-size network.[]{data-label="fig:Synthetic_LargeSize_OR_30"}](results_final/{Syn_Injection-ROC_Mixture-3000-100000-0.00010-10-30}.eps){width="\hsize"} [(a) ROC]{} \[\] ![ROC and F-measure in large-size network.[]{data-label="fig:Synthetic_LargeSize_OR_30"}](results_final/{Syn_Injection-f-measure_Mixture-3000-100000-0.00010-10-30}.eps){width="\hsize"} [(b) F-measure]{} \[\] Finally, Table \[Tbl:ProcessingTime\] shows the processing time for the mid-size network and the large-size network, respectively. The table shows not only the total processing time but also its breakdown, where “A” and “B” show the processing time for subspace tracking of $\{\vec{b}, \mat{A}, \mat{C}\}$, and the processing time for calculation of abnormal flows $\vec{v}$, respectively. It should be noted that, whereas “A” depends on the implementation of each algorithm, “B” is calculated in the same implementation. The implemented code of our algorithm is not optimized. For the mid-size network, the proposed algorithm requires a longer tracking time (“A”) for multiple matrix inversions per iteration while a shorter calculation time (“B”) of sparse abnormal, , is needed compared to other algorithms. On the other hand, for the large-size network, the proposed algorithm is much faster. Since the subspace estimated by the proposed algorithm is much closer to the real subspace and is able to more efficiently capture the underlying time-series structure than others, the subsequent $\ell_1$ calculation of the proposed algorithm converges much faster than others. The results also reveals that the processing times almost match the computational complexity analysis in Section \[Sec:ComputationalComplexityAnalysis\] with respect to the size of $L$ and $F$. Network size -------------- ------------- ---------------- ---------------- ------------- ---------------- ---------------- Algorithm A    B    Total A   B    Total Proposed 393.1 [**1707.7**]{} [**2100.8**]{} 693.7 [**3585.9**]{} [**4279.7**]{} GROUSE 2.3 2339.5 2341.8 0.5 4688.5 4689.0 GRASTA 3.0 2347.3 2350.2 1.2 4676.8 4678.0 PETRELS 11.5 2341.4 2352.9 14.4 4674.9 4689.3 Wavelet 33.8 2352.8 2386.6 50.8 4667.6 4718.5 EWMA [**0.1**]{} 2336.6 2336.7 [**0.1**]{} 4626.9 4627.0 : Processing time for mid-size and large-size network.[]{data-label="Tbl:ProcessingTime"} Conclusion ========== In this paper, we have addressed the challenge of detecting volume anomalies in large-scale data communication networks in an unsupervised way, where only the link traffic can be observed consisting of superimposed flows. For this purpose, the present paper assumes that anomalies in the flow can be identified by means of deviations of the measurements from a low rank structure. The network flow is modeled by means of a third order tensor with Hankel structure along one slice matrix to represent time-directional features as well as the spatial correlation of multiple links. By exploiting this traffic tensor with the Candecomp/PARAFC decomposition, a new online subspace tracking of the underlying low rank structure is proposed for normal flows based on the recursive least squares (RLS) method under partial observation. Meanwhile, abnormal flows are estimated as outlier sparse flows via sparsity maximization in the under-constrained linear-inverse problem. An inherent shortcoming of our approach, which is shared by all unsupervised detection methods that are based on subspace tracking, is that anomalies which consist of change of the traffic at a very low frequency cannot be detected. Numerical evaluations show that the proposed algorithm detects and identifies abnormal flows more accurately and with less computation time than the state-of-the-art online algorithms for a large-scale network. As future research directions, we plan to extend our studies to the cases where even the direct flow traffic data is only partially-observable. Additionally, the convergence analysis is a challenging task of the proposed online tensor optimization, and this remains an open problem. Acknowledgement {#acknowledgement .unnumbered} =============== This work was initiated while H. Kasai was with the Department of Electrical and Computer Engineering (ECE) of Technical University of Munich, Germany. Part of this work is part of a project that has received funding from the European Research Council (ERC) under the European Unions Horizon 2020 research and innovation program (grant agreement No 647158 - FlexNets). (Corresponding author: Hiroyuki Kasai.) Appendix {#appendix .unnumbered} ======== Hankel matrix {#Append_Sec:hankel} ============= Let $\vec{y} = \{y_{1}, \ldots ,y_{N}\}$ of length $N$ be a one-dimensional data. Given a window length $W$, with $1<W<N$, we construct the $W$-lagged vectors $\vec{h}_{k} = (y_{k}, \ldots , y_{k+W-1})^T \in \mathbb{R}^{W}$, $k = 1, \ldots, K$, where $K=N-W+1$, and compose these vectors into the Hankel matrix $\mat{H} = [\vec{h}_{1}: \cdots :\vec{h}_{K}] \in \mathbb{R}^{W \times K}$ as $$\begin{aligned} \mat{H} \ =\ [\vec{h}_1:\ldots:\vec{h}_K] \ =\ \left( \begin{array}{ccccc} y_{1} \!\!& y_{2} \!\!& y_{3} &\!\! \ldots \!\!& y_{K} \\ y_{2} \!\!& y_{3} \!\!& y_{4} &\!\! \ldots \!\!& y_{K+1} \\ \vdots & \vdots & \vdots & \!\!\ddots \!\!& \vdots \\ y_{W} & y_{W+1} & y_{W+2} &\!\! \ldots \!\!& y_{N} \end{array} \right) . \nonumber\end{aligned}$$ This matrix is often called a [*trajectory matrix*]{}, which means that all the elements along the anti-diagonal are equal. Candecomp/PARAFAC tensor decomposition {#Append_Sec:CPDEC} ====================================== The Candecomp/PARAFAC decomposition decomposes a tensor into a sum of component rank-one tensors [@Kolda_SIAMReview_2009]. Let $\mathbfcal{X}$ be a third-order tensor of size $L\times W \times T$, and assume its [*rank*]{} is $R$, we approximate $\mathbfcal{X}$ as $\mathbfcal{X} \approx \sum_{r=1}^R \vec{a}_r \circ \vec{b}_r \circ \vec{c}_r=\sum_{r=1}^R \vec{c}^t(r)\vec{a}_ r \vec{b}_r^T$, where $\vec{a}_r \in \mathbb{R}^L$, $\vec{b}_r \in \mathbb{R}^W$, and $\vec{c}_r \in \mathbb{R}^T$. The symbol “$\circ$” represents the vector outer product. Fig.\[Append\_Fig:BasicConcept\] illustrates rank-one tensor decomposition of the Candecomp/PARAFAC decomposition. The [*factor matrices*]{} refer to the combination of the vectors from the rank-one components, i.e., = $[\vec{a}_1: \vec{a}_2: \cdots : \vec{a}_R] \in \mathbb{R}^{L \times R}$ and likewise for $\mat{B} \in \mathbb{R}^{W \times R}$ and $\mat{C }\in \mathbb{R}^{T \times R}$. It should be emphasized that , and can be also represented by [*row vectors*]{}, i.e., [*horizontal vectors*]{}, namely, $\mat{A} = [(\vec{a}^1)^T: \cdots: (\vec{a}^L)^T]^T$, $\mat{B} = [(\vec{b}^1)^T: \cdots : (\vec{b}^T)^T]^T$, and $\mat{C} = [(\vec{c}^1)^T : \cdots : (\vec{c}^W)^T]^T$, where $\{\vec{a}^l, \vec{b}^t,\vec{c}^w\} \in \mathbb{R}^{R}$. Thus, $\mat{X}_t\approx\mat{A}{\rm diag}(\vec{b}^t)\mat{C}^T$. ![Candecomp/PARAFAC tensor decomposition.[]{data-label="Append_Fig:BasicConcept"}](figures_final/cp.eps){width="14.5cm"} Derivation of [**A**]{}\[t\] by Recursive Least Squares {#Append_Sec:RLS} ======================================================= This appendix describes the derivation of [**A**]{}\[t\]. The derivative of (\[eq:problem\_def\_am\]) with regard to $\vec{a}^l$ is calculated as $$\begin{aligned} \frac{\partial{F(\vec{a}^l)}}{\partial{(\vec{a}^l})} & = & \sum_{\tau=1}^t \biggl[-\sum_{w=1}^W \lambda^{t-\tau} ({\bf \Omega}_{\tau})_{l,w} \left((\mat{Z}_\tau)_{l,w} - (\vec{a}^l)^T \vec{\alpha}_w[\tau] \right) \vec{\alpha}_{w} [\tau] \nonumber \\ & &+\ \mu_h \sum_{w=1}^{W\!-\!1} \lambda^{t-\tau} ({\bf \Omega}_{\tau})_{l,w} (\vec{a}^l)^T \vec{\beta}_w[\tau] (\vec{\beta}_w[\tau])^T \biggr] + \mu_r[t]\vec{a}^l.\nonumber\end{aligned}$$ Then, by setting this derivative equal to zero, we get $\vec{a}^l[t]$ as $$\begin{split} &\left( \sum_{\tau=1}^t \biggl[ \sum_{w=1}^W \lambda^{t-\tau} ({\bf \Omega}_{\tau})_{l,w} \vec{\alpha}_w[\tau] (\vec{\alpha}_w[\tau])^T + \mu_h \sum_{w=1}^{W\!-\!1} \lambda^{t-\tau} ({\bf \Omega}_{\tau})_{l,w} \vec{\beta}_w[\tau] (\vec{\beta}_w[\tau])^T \biggr]+ \mu_r[t]\mat{I}_R \right) \vec{a}^l[t]\nonumber \\ &\hspace*{2cm}= \sum_{\tau=1}^t \sum_{w=1}^W \lambda^{t-\tau} ({\bf \Omega}_{\tau})_{l,w} (\mat{Z}_\tau)_{l,w} \vec{\alpha}_w[\tau] , \end{split}$$ where $((\vec{a}^l)^T[t] \vec{\alpha}_w[\tau]) \vec{\alpha}_w[\tau]= ((\vec{\alpha}_w[\tau])^T \vec{a}^l[t]) \vec{\alpha}_w[\tau] = \vec{\alpha}_w[\tau]((\vec{\alpha}_w[\tau])^T \vec{a}^l[t])$ is used. Finally, we get the following; $$\begin{aligned} \mat{RA}_l[t] \vec{a}^l[t] & = & \vec{s}_l[t] \nonumber \\ \vec{a}^l[t] & = & (\mat{RA}_l[t])^{\dagger} \vec{s}_l[t],\nonumber\end{aligned}$$ where $\mat{RA}_l[t] \in \mathbb{R}^{R\times R}$ and $\vec{s}_l[t] \in \mathbb{R}^{R}$ are defined as $$\begin{aligned} \mat{RA}_l[t] & :=& \sum_{\tau=1}^t \biggl[ \sum_{w=1}^W \lambda^{t-\tau} ({\bf \Omega}_{\tau})_{l,w} \vec{\alpha}_w[\tau] (\vec{\alpha}_w[\tau])^T + \mu_h \sum_{w=1}^{W\!-\!1} \lambda^{t-\tau} ({\bf \Omega}_{\tau})_{l,w} \vec{\beta}_w[\tau] (\vec{\beta}_w[\tau])^T \biggr]+ \mu_r[t]\mat{I}_R, \nonumber \\ \vec{s}_l[t] & := & \sum_{\tau=1}^t \sum_{w=1}^W \lambda^{t-\tau} ({\bf \Omega}_{\tau})_{l,w} (\mat{Z}_\tau)_{l,w} \vec{\alpha}_w[\tau].\nonumber\end{aligned}$$ Here, $\mat{RA}_l[t]$ is transformed as (\[appeq:Update\_RA\]). $$\begin{aligned} \label{appeq:Update_RA} \mat{RA}_l[t] & = & \sum_{\tau=1}^{t-1} \left( \sum_{w=1}^W \lambda^{t-\tau} ({\bf \Omega}_{\tau})_{l,w} \vec{\alpha}_w[\tau] (\vec{\alpha}_w[\tau])^T + \mu_h \sum_{w=1}^{W\!-\!1} \lambda^{t-\tau} ({\bf \Omega}_{\tau})_{l,w} \vec{\beta}_w[\tau] (\vec{\beta}_w[\tau])^T \right) \nonumber \\ &&+ \sum_{w=1}^W ({\bf \Omega}_t)_{l,w} \vec{\alpha}_w[t] (\vec{\alpha}_w[t])^T + \mu_h \sum_{w=1}^{W\!-\!1} ({\bf \Omega}_t)_{l,w} \vec{\beta}_w[t] (\vec{\beta}_w[t])^T + \mu_r[t] \mat{I}_R \nonumber \\ &=& \lambda \biggl[ \sum_{\tau=1}^{t-1} \left( \sum_{w=1}^W \lambda^{t-1-\tau} ({\bf \Omega}_{\tau})_{l,w} \vec{\alpha}_w[\tau] (\vec{\alpha}_w[\tau])^T + \mu_h \sum_{w=1}^{W\!-\!1} \lambda^{t-1-\tau} ({\bf \Omega}_{\tau})_{l,w} \vec{\beta}_w[\tau] (\vec{\beta}_w[\tau])^T \right) + \mu_r[\tau] \mat{I}_R \biggr] \nonumber \\ &&+ \sum_{w=1}^W ({\bf \Omega}_t)_{l,w} \vec{\alpha}_w[t] (\vec{\alpha}_w[t])^T + \mu_h \sum_{w=1}^{W\!-\!1} ({\bf \Omega}_t)_{l,w} \vec{\beta}_w[t] (\vec{\beta}_w[t])^T + \mu_r[t] \mat{I}_R - \lambda \mu_r[t\!-\!1] \mat{I}_R \nonumber \\ &=& \lambda \mat{RA}_l[t\!-\!1] + \sum_{w=1}^W ({\bf \Omega}_t)_{l,w} \vec{\alpha}_w[t] \vec{\alpha}_w[t]^T + \mu_h \sum_{w=1}^{W\!-\!1} ({\bf \Omega}_t)_{l,w} \vec{\beta}_w[t] \vec{\beta}_w[t]^T+ (\mu_r[t] - \lambda \mu_r[t\!-\!1] ) \mat{I}_R.\end{aligned}$$ Likewise, $\vec{s}_l[t]$ is also transformed as $$\begin{aligned} \vec{s}_l[t] & = & \sum_{w=1}^W \biggl[ \lambda^{t-1} ({\bf \Omega}_1)_{l,w} (\mat{Z}_{1})_{l,w} \vec{\alpha}_{w}[1] + \cdots + \lambda^{1} ({\bf \Omega}_{t\!-\!1})_{l,w} (\mat{Z}_{t\!-\!1})_{l,w} \vec{\alpha}_{w}[t\!-\!1] + \lambda^{0} ({\bf \Omega}_t)_{l,w} (\mat{Z}_t)_{l,w} \vec{\alpha}_w[t] \biggr]\nonumber \\ & = & \lambda \vec{s}_l[t\!-\!1] + \sum_{w=1}^W ({\bf \Omega}_t)_{l,w} (\mat{Z}_t)_{l,w} \vec{\alpha}_w[t]. \nonumber \end{aligned}$$ From $\mat{RA}_l[t] \vec{a}^l[t] = \vec{s}_l[t] $, we modify this as (\[appeq:RA\_a\]). $$\begin{array}{lll} \label{appeq:RA_a} \mat{RA}_l[t] \vec{a}^l[t] &=& \lambda \mat{RA}_l[t\!-\!1] \vec{a}^l[t\!-\!1] +\displaystyle{\sum_{w=1}^W ({\bf \Omega}_t)_{l,w} (\mat{Z}_t)_{l,w} \vec{\alpha}_w[t]} \nonumber \\ &=& \left(\mat{RA}_l[t] - \displaystyle{\sum_{w=1}^W ({\bf \Omega}_t)_{l,w} \vec{\alpha}_w[t] (\vec{\alpha}_w[t])^T} \right. -\mu_h \displaystyle{\sum_{w=1}^{W\!-\!1} ({\bf \Omega}_t )_{l,w+1} \vec{\beta}_w[t] (\vec{\beta}_w[t])^T} \\ &&\Biggl.- (\mu_r[t] - \lambda \mu_r[t\!-\!1]) \mat{I}_R \Biggr) \vec{a}^l[t\!-\!1] + \!\! \displaystyle{\sum_{w=1}^W ({\bf \Omega}_t)_{l,w} (\mat{Z}_t)_{l,w} \vec{\alpha}_w[t] } \nonumber \\ &=& \Biggl(\mat{RA}_l[t] - \mu_h \displaystyle{\sum_{w=1}^{W-1} ({\bf \Omega}_t)_{l,w+1} \vec{\beta}_w[t] (\vec{\beta}_w[t])^T} \Biggr. - \Biggl. (\mu_r[t] - \lambda \mu_r[t-1]) \mat{I}_R \Biggr) \vec{a}^l[t\!-\!1] \nonumber \\ &&+ \displaystyle{\sum_{w=1}^W ({\bf \Omega}_t)_{l,w} \left((\mat{Z}_t)_{l,w} - (\vec{\alpha}_w[t])^T \vec{a}^l[t-1] \right) \vec{\alpha}_w[t]}. \end{array}$$ Finally, we obtain $\vec{a}^l[t]$ as (\[Eq:al\_final\]) by $\mat{Z}_\tau = \mat{Y}_\tau-{\mat{V}_{\scriptsize \mat{R}}}_\tau$. [^1]: H. Kasai is with the Graduate School of Informatics and Engineering, The University of Electro-Communications, 1-5-1 Chofugaoka, Chofu-shi, Tokyo, 182-8585, Japan (e-mail: [email protected], web:www.kasailab.com) [^2]: W. Kellerer is with the Department of Electrical and Computer Engineering, Technical University of Munich, Munich 80290, Germany (e-mail: [email protected], www.lkn.ei.tum.de) [^3]: M. Kleinsteuber is with the Department of Electrical and Computer Engineering, Technical University of Munich, Munich 80290, Germany (e-mail: [email protected], www.gol.ei.tum.de) [^4]: Exponentially Weighted Moving Average. This is equivalent to ARIMA(0,1,1) model.

ArXiv

--- abstract: 'State space models (SSMs) are a flexible approach to modeling complex time series. However, inference in SSMs is often computationally prohibitive for long time series. Stochastic gradient MCMC (SGMCMC) is a popular method for scalable Bayesian inference for large independent data. Unfortunately when applied to dependent data, such as in SSMs, SGMCMC’s stochastic gradient estimates are biased as they break crucial temporal dependencies. To alleviate this, we propose stochastic gradient estimators that control this bias by performing additional computation in a ‘buffer’ to reduce breaking dependencies. Furthermore, we derive error bounds for this bias and show a geometric decay under mild conditions. Using these estimators, we develop novel SGMCMC samplers for discrete, continuous and mixed-type SSMs with analytic message passing. Our experiments on real and synthetic data demonstrate the effectiveness of our SGMCMC algorithms compared to batch MCMC, allowing us to scale inference to long time series with millions of time points.' author: - 'Christopher Aicher[^1]' - 'Yi-An Ma[^2]' - 'Nicholas J. Foti[^3]' - 'Emily B. Fox' bibliography: - 'references.bib' nocite: '[@*]' title: Stochastic Gradient MCMC for State Space Models --- Introduction {#sec:intro} ============ State space models (SSMs) are ubiquitous in the analysis of time series in fields as diverse as biology [@wulsin2013bayesian], finance and economics [@kim1999state; @zeng2013state], and systems and control [@elliott2008hidden]. As a defining feature, SSMs augment the observed time series with a *latent state sequence* to model complex time series dynamics with a latent Markov chain dependence structure. Given a time series, inference of model parameters involves sampling or marginalizing this latent state sequence. Unfortunately, both the runtime and memory required scale with the length of the time series, which is prohibitive for long time series (e.g. high frequency stock prices [@goodhart1997high], genome sequences [@eddy1998profile], or neural impulse recordings [@davis2016mining]). In practice, given a long time series, one could ‘segment’ or ‘downsample’ to reduce length; however, this preprocessing can destroy or change important signals and computational considerations should ideally not limit scientific modeling. To help scale inference in SSMs, we consider stochastic gradient Markov chain Monte Carlo (SGMCMC), a popular method for scaling Bayesian inference to large data sets [@chen2015convergence; @ma2015complete; @welling2011bayesian]. The key idea of SGMCMC is to employ stochastic gradient estimates based on subsets or ‘minibatches’ of data, avoiding costly computation of gradients on the full dataset, such that the resulting dynamics produce samples from the posterior distribution over SSM parameters. This approach has found much success in *independent* data models, where the stochastic gradients are *unbiased* estimates of the true gradients. However, when applying SGMCMC to SSMs, naive stochastic gradients are *biased*, as subsampling the data breaks dependencies in the SSM’s latent state sequence. This bias can destroy the dynamics of SGMCMC causing it to fail when applied to SSMs. The challenge is to correct these stochastic gradients for SSMs while maintaining the computational benefits of SGMCMC. In this work, we develop computationally efficient stochastic gradient estimators for inference in general discrete-time SSMs. To control the bias of stochastic gradients, we marginalize the latent state sequence in a *buffer* around each subsequence, propagating critical information from outside each subsequence to its local gradient estimate while avoiding costly full-chain computations. Similar buffering ideas have been previously considered for belief propagation [@gonzalez2009residual], variational inference [@foti2014stochastic], and in our earlier work on SGMCMC for hidden Markov models (HMMs) [@ma2017stochastic], but all are limited to discrete latent states. Here, we present buffering as an approximation to *Fisher’s identity* [@cappe2005inference], allowing us to naturally extend buffering trick to continuous and mixed-type latent states. We further develop analytic bounds on the bias of our proposed gradient estimator that, under mild conditions, decay geometrically in the buffer size. To obtain these bounds we prove that the latent state sequence posterior distribution has an *exponential forgetting* property [@cappe2005inference; @del2010forward]. However unlike classic results which prove a geometric decay between the approximate and exact marginal posterior distributions in total variation distance, we use Wasserstein distance [@villani2008optimal] to allow analysis of continuous and mixed-type latent state SSMs. Our approach is similar to proofs of Wasserstein ergodicity in homogeneous Markov chains [@durmus2015quantitative; @madras2010quantitative; @rudolf2015perturbation]; however we extend these ideas to the *nonhomogeneous* Markov chains defined by the latent state sequence posterior distribution. These geometrically decaying bounds guarantee that we only need a small buffer size in practice, allowing scalable inference in SSMs. Although our proposed gradient estimator can be generally applied to any stochastic gradient method, here, we develop SGMCMC samplers for Bayesian inference in a variety of SSMs such as HMMs, linear Gaussian SSMs (LGSSM), and switching linear dynamical systems (SLDS) [@cappe2005inference; @fox2009bayesian]. We also derive preconditioning matrices to take advantage of information geometry, which allows for more rapid mixing and convergence of our samplers [@girolami2011riemann; @patterson2013stochastic]. Finally, we validate our algorithms and theory on a variety of synthetic and real data experiments, finding that our gradient estimator can provide orders of magnitude run-time speed ups compared to batch sampling. This paper significantly expands upon our initial work [@ma2017stochastic], by (i) connecting buffering to Fisher’s identity, simplifying its presentation and analysis, (ii) non-trivially generalizing the approach to SSMs beyond the HMM, including continuous and mixed-type latent states, (iii) developing a general framework for bounding the error of buffered gradient estimators using Wasserstein distance, and (iv) providing extensive validation on a number of real and synthetic datasets. The paper is organized as follows. First, we review background on SSMs and SGMCMC methods in Section \[sec:background\]. We then present our framework of constructing buffered gradient estimators to extend SGMCMC to SSMs in Section \[sec:framework\]. We prove the geometrically decaying bounds for our proposed buffered gradient estimate in Section \[sec:bounds\]. We apply our framework and error bounds to discrete, continuous and mixed-type latent state SSMs in Section \[sec:models\]. Finally, we investigate our algorithms on both synthetic and real data in Section \[sec:experiments\]. Background {#sec:background} ========== State Space Models for Time Series ---------------------------------- State space models (SSMs) for time series are a class of discrete-time bivariate stochastic process $\{{u}_t, y_t \}_{t \in {\mathcal{T}}}$, ${\mathcal{T}}= \{1, \ldots, T\}$, consisting of a latent state sequence ${u}:= {u}_{1:T}$ generated by a homogeneous Markov chain and an observation sequence $y := y_{1:T}$ generated independently conditioned on ${u}$ [@cappe2005inference]. Examples of state space models include: HMMs, LGSSMs, and SLDSs (see Section \[sec:models\] for details). For a generic SSM, the joint distribution of $y$ and ${u}$ factorizes as $$\label{eq:complete_data_like} p(y, {u}\, | \, \theta) = \prod_{t = 1}^T p(y_t \, | \, {u}_t, \theta) p({u}_t \, | \, {u}_{t-1}, \theta) \cdot p_0({u}_0) \enspace,$$ where $\theta$ are model-specific parameters, $p(y_t | {u}_t, \theta)$ is the *emission density*, $p({u}_t | {u}_{t-1}, \theta)$ is the *transition density*, and $p_0({u}_0)$ is a prior for the latent states. As the latent state sequence $u$ is unobserved, the likelihood of $\theta$ given only the observations $y$ (marginalizing $u$) is $$\label{eq:marginal_like} p(y \, | \, \theta) = \int \, \prod_{t = 1}^T p(y_t \, | \, {u}_t, \theta) p({u}_t \, | \, {u}_{t-1}, \theta) \cdot p_0({u}_0) \enspace d{u}\enspace,$$ Unconditionally, the observations $y$ are not independent and the graphical model of this *marginal likelihood*, Eq. , has many long term dependencies, Figure \[fig:ssm\_graphical\] (right). In contrast, when conditioned on $u$ the observations $y$ are independent and the *complete-data likelihood*, Eq. , has a simpler chain structure, Figure \[fig:ssm\_graphical\] (left). ![Graphical Model of a SSM: (left) the joint process $u, y$, Eq.  and (right) $y$ marginalizing out $u$, Eq. . The parameters $\theta$ are not shown, but connect to all nodes. []{data-label="fig:ssm_graphical"}](figures/latex_fig/ssm.pdf){width="0.99\linewidth"} ![Graphical Model of a SSM: (left) the joint process $u, y$, Eq.  and (right) $y$ marginalizing out $u$, Eq. . The parameters $\theta$ are not shown, but connect to all nodes. []{data-label="fig:ssm_graphical"}](figures/latex_fig/ssm_marginalized.pdf){width="0.99\linewidth"} To infer $\theta$ given $y$, we can maximize the marginal likelihood $p(y \, | \, \theta)$ or, given a prior $p(\theta)$, sample from the posterior $p(\theta \, | \, y) \propto p(y \, | \, \theta) p(\theta)$. However, traditional inference methods for $\theta$, such as expectation maximization (EM), variational inference, or Gibbs sampling, take advantage of the conditional independence structure in $p(y, u \, | \, \theta)$, Eq. , rather than working directly with $p(y \, | \, \theta)$, Eq.  [@beal2003variational; @scott2002bayesian]. To use $p(y, u \, | \, \theta)$ with unobserved $u$, these methods rely on sampling or taking expectations of ${u}$ from the posterior $\gamma({u}) := p({u}\, | \, y, \theta)$. As an example, gradient-based methods take advantage of *Fisher’s identity* [@cappe2005inference] $$\label{eq:fisher_identity} {\nabla}\log p(y \, | \, \theta) = {\mathbb{E} \,}_{u | y, \theta}[ {\nabla}\log p(y, {u}\, | \, \theta)] = {\mathbb{E} \,}_{u\sim\gamma}[{\nabla}\log p(y, {u}\, | \, \theta)]\enspace,$$ which allows gradients of Eq.  to be computed in terms of Eq. . To compute the posterior $\gamma(u)$, these methods use the well-known *forward-backward algorithm* [@cappe2005inference; @scott2002bayesian]. The algorithm works by recursively computing a sequence of forward messages $\alpha_t({u}_t)$ and backward messages $\beta_t({u}_t)$ which are used to compute the pairwise marginals of $\gamma$. More specifically, $$\begin{aligned} \label{eq:forward_message} \alpha_t({u}_t) &:= p({u}_t, y_{\leq t} \, | \, \theta) = \int p(y_t, {u}_t \, | \, {u}_{t-1}, \theta) \alpha_{t-1}({u}_{t-1}) \, d {u}_{t-1} \\ \label{eq:backward_message} \beta_t({u}_t) &:= p(y_{>t} \, | \, {u}_t, \theta) = \int p(y_{t+1}, {u}_{t+1} \, | \, {u}_t, \theta) \beta_{t+1}({u}_{t+1}) \, d {u}_{t+1} \\ \label{eq:pairwise_marginal} \gamma_{t-1:t}({u}_{t-1}, {u}_{t}) &:= p({u}_{t-1}, {u}_t \, | \, y, \theta) \propto \alpha_{t-1}({u}_{t-1}) p(y_t, {u}_t \, | \, {u}_{t-1}, \theta) \beta_t({u}_t) \enspace.\end{aligned}$$ When message passing is tractable (i.e., when Eqs. - involve discrete or conjugate likelihoods), the forward-backward algorithm can be calculated in closed form. When message passing is intractable, the messages can be approximated using Monte-Carlo sampling methods (e.g. blocked Gibbs sampling [@carter1994gibbs; @fox2011bayesian], particle methods [@andrieu2010particle; @briers2010smoothing; @doucet2009tutorial; @sudderth2010nonparametric]). In both cases, when the length of the time series $|{\mathcal{T}}|$ is much larger than the dimension of $\theta$, the forward-backward algorithm (running over the entire sequence) requires $O(|{\mathcal{T}}|)$ time and memory at *each iteration*. The SSM challenge is to scale inference of model parameters $\theta$ to long time series when the computation and storage per iteration $O(|{\mathcal{T}}|)$ is prohibitive. Stochastic Gradient MCMC ------------------------ One popular method for scalable Bayesian inference is *stochastic gradient* Markov chain Monte Carlo (SGMCMC) [@chen2015convergence; @ma2015complete; @welling2011bayesian]. The idea behind gradient-based MCMC is to simulate continuous dynamics for a *potential energy* function $U(\theta) \propto -\log p(y, \theta)$ such that the dynamics generate samples from the posterior distribution $p(\theta \, | \, y)$. For example, the Langevin diffusion over $U(\theta)$ is given by the stochastic differential equation (SDE) $$\label{eq:langevin_diffusion} d\theta_s = g(\theta) ds + \sqrt{2} dW_s \enspace,$$ where $dW_s$ is Brownian motion, $g(\theta) = -{\nabla}U(\theta) = {\nabla}_\theta \log p(y, \theta)$, and $s$ indexes continuous time. As $s \rightarrow \infty$, the distribution of $\theta_s$ converges to the SDE’s stationary distribution, which by the Fokker-Planck equation is the posterior $p(\theta \, | \, y)$ [@ma2015complete]. Because we cannot perfectly simulate Eq. , in practice we use a discretized numerical approximation. One straightforward approximation is the Euler-Mayurma discretization $$\label{eq:LMC} \theta^{(s+1)} \leftarrow \theta^{(s)} + h g(\theta^{(s)}) + \mathcal{N}(0, 2h) \enspace,$$ where $h$ is the stepsize and $s$ indexes discrete time steps. This recursive update defines the Langevin Monte-Carlo (LMC) algorithm. Typically, a Metropolis-Hastings correction step is added to account for the discretization error [@roberts1996exponential; @roberts1998optimal]. For large datasets, computing $g(\theta)$ at every step in Eq.  is computationally prohibitive. To alleviate this, the key ideas of *stochastic gradient* Langevin dynamics (SGLD) are to replace $g(\theta)$ with a quick-to-compute unbiased estimator $\hat{g}(\theta)$ and to use a decreasing stepsize $h^{(s)}$ to avoid costly Metropolis-Hastings correction steps [@welling2011bayesian] $$\label{eq:sgld_step} \theta^{(s+1)} \leftarrow \theta^{(s)} + h^{(s)} \hat{g}(\theta^{(s)}) + \mathcal{N}(0, 2h^{(s)}) \enspace.$$ For i.i.d. data, an example of $\hat{g}(\theta)$ is to use a random minibatch ${{\mathcal{S}}}\subset {\mathcal{T}}$, $|{{\mathcal{S}}}|\ll|{\mathcal{T}}|$ $$\label{eq:approx_potential_iid} \hat{g}(\theta) = -\frac{1}{\Pr({{\mathcal{S}}})} \sum_{t \in {{{\mathcal{S}}}}} {\nabla}\log p(y_t \, | \, \theta) - {\nabla}\log p(\theta) \enspace,$$ which only requires $O(|{{\mathcal{S}}}|)$ time to compute. When $\hat{g}(\theta)$ is unbiased and with an appropriate decreasing stepsize schedule $h^{(s)}$, the distribution of $\theta^{(s)}$ asymptotically converges to the posterior distribution [@chen2015convergence; @teh2016consistency]. However, in practice one uses a small, finite step-size for greater efficiency, which introduces a small bias [@dalalyan2017user]. A Riemannian extension of SGLD (SGRLD) simulates the Langevin diffusion over a Riemannian manifold with metric $D(\theta)^{-1}$ by preconditioning the gradient and noise of Eq.  by $D(\theta)$. By incorporating geometric information about structure of $\theta$, SGRLD aims for a diffusion which mixes more rapidly. Suggested examples of the metric $D(\theta)^{-1}$ are the Fisher information matrix $\mathcal{I}(\theta) = {\mathbb{E} \,}_{y}[{\nabla}^2\log p(y\,|\,\theta)]$ or a noisy Hessian estimate $\widehat{{\nabla}^2 \log p}(y \, | \,\theta)$  [@girolami2011riemann; @patterson2013stochastic]. Given $D(\theta)$, each step of SGRLD is $$\label{eq:sgrld_step} \theta^{(s+1)} \leftarrow \theta^{(s)} + h \left[D(\theta^{(s)}) \cdot \hat{g}(\theta^{(s)}) + \Gamma(\theta^{(s)})\right] + \mathcal{N}\left(0, 2h D(\theta^{(s)})\right)$$ where the vector $\Gamma(\theta)$ is a correction term $\Gamma(\theta)_i = \sum_j \frac{\partial D(\theta)_{ij}}{\partial \theta_{j}}$ to ensure the dynamics converge to the target posterior [@ma2015complete; @xifara2014langevin]. Many extensions to SGMCMC have been proposed such as using control variates to reduce the variance of $\hat{g}(\theta)$ [@baker2017control; @chatterji2018theory; @nagapetyan2017true] or augmented dynamics to improve mixing [@chen2015convergence; @chen2014stochastic; @ding2014bayesian; @li2016preconditioned]. Although our ideas extend to these formulations as well, we focus on the popular SGLD and SGRLD algorithms. To apply SGMCMC to SSMs, we must choose whether to use the complete-data loglikelihood or the marginal data loglikelihood in the potential $U(\theta)$. If we use the complete-data loglikelihood, then we treat $({u}, \theta)$ as the parameters. Although the observations $y$ conditioned on $({u}, \theta)$ are independent, we must calculate gradients for ${u}_{-T:T}$ at each iteration, which is prohibitive for long sequences $|{\mathcal{T}}|$ and intractable for discrete or mixed-type ${u}$. On the other hand, if we use the marginal loglikelihood, then we only need to take gradients in $\theta$. However, the observations $y$ conditioned on $\theta$ alone are *not* independent and therefore the minibatch gradient estimator Eq.  breaks crucial dependencies causing it to be biased. Our SGMCMC challenge is correcting the bias in stochastic gradient estimates ${\nabla}\widetilde U(\theta)$ when applied to SSMs. General Framework {#sec:framework} ================= We now present our framework for scalable Bayesian inference in SSMs with long observation sequences. Our approach is to extend SGMCMC to SSMs by developing a gradient estimator that ameliorates the issue of broken temporal dependencies. In particular, we develop a computationally efficient gradient estimator that uses a *buffer* to avoid breaking crucial dependencies, only breaking weak dependencies. We first present a (computationally prohibitive) unbiased estimator of $g(\theta) = {\nabla}\log p(y \, | \, \theta)$ for SSMs using Fisher’s identity. We then derive a general computationally efficient gradient estimate $\tilde{g}(\theta)$ that accounts for the dependence in observations using a buffer. We also propose preconditioning matrices for SGRLD with SSMs. Finally, we present our general SGMCMC pseudocode for SSMs. Unbiased Gradient Estimate {#sec:framework-unbiased-gradient} -------------------------- The main challenge in constructing an efficient estimate $\tilde{g}(\theta)$ of $g(\theta)$ for SSMs is handling the lack of independence (marginally) in $y$. Because the observations in SSMs are not independent, we cannot produce an unbiased estimate of $g(\theta)$ with a randomly selected subset of data points as in Eq.. For example, a naive estimate is to take the gradient of a random contiguous *subsequence* ${{\mathcal{S}}}= \{t_1, \ldots, t_{S}\} \subset {\mathcal{T}}$ with $t_i = t_{i-1}+1$ $$\hat{g}(\theta) = - \frac{1}{\Pr({{\mathcal{S}}})}{\nabla}\log p(y_{{\mathcal{S}}}\, | \, \theta) - {\nabla}\log p(\theta) \enspace,$$ where $p(y_{{\mathcal{S}}}\, | \, \theta)$ is computed with $p({u}_{t_0}) = p_0({u}_{t_0})$. This estimate only requires $O(S)$ time compared to the $O(T)$ for $g(\theta)$. However because the marginal likelihood does not factorize as in the independent observations case, this estimate is biased ${\mathbb{E} \,}_{{\mathcal{S}}}[\hat{g}(\theta)] \neq g(\theta)$. In addition, as ${{\mathcal{S}}}$ are contiguous subsequences of ${\mathcal{T}}$, the scaling factor $\Pr({{\mathcal{S}}})^{-1}$ is no longer correct as time points in the center of $\mathcal{T}$ are sampled more frequently than the endpoints; instead each time point should be scaled point-wise. To obtain an unbiased estimate for $g(\theta)$, we use Fisher’s identity Eq.  to rewrite $g(\theta)$ in terms of the complete-data loglikelihood as a sum over time points $$\begin{aligned} g(\theta) &= - {\nabla}\log p(y \, |\, \theta) - {\nabla}\log p(\theta)\\ \nonumber &= - {\mathbb{E} \,}_{{u}| y, \theta} \left[ {\nabla}\log p(y, {u}\, | \,\theta) \right] - {\nabla}\log p(\theta) \\ \nonumber &= - \sum_{t \in {\mathcal{T}}}{\mathbb{E} \,}_{{u}| y, \theta} \left[ {\nabla}\log p(y_t, {u}_t \, | \, {u}_{t-1}, \theta) \right] - {\nabla}\log p(\theta)\end{aligned}$$ From this, we straightforwardly identify an unbiased estimator for a subsequence ${{\mathcal{S}}}$ $$\label{eq:unbiased_potential_estimate} \bar{g}(\theta) = - \sum_{t \in {{\mathcal{S}}}} \frac{1}{\Pr(t \in {{\mathcal{S}}})}{\mathbb{E} \,}_{{u}| y, \theta} \left[ {\nabla}\log p(y_t, {u}_t \, | \, {u}_{t-1}, \theta) \right] - {\nabla}\log p(\theta) \enspace,$$ where $\Pr(t \in {{\mathcal{S}}})$ is the probability $t$ is in the random subsequence ${{\mathcal{S}}}$. Although Eq.  reduces the number of gradient terms to compute from $T$ to $S$, the summation terms require calculating expectations of $u \, | \, y, \theta$. More specifically, Eq.  requires expectations with respect to the pairwise marginal posteriors $p({u}_t, {u}_{t-1}\, | \, y_{\mathcal{T}})$ for $t \in {{\mathcal{S}}}$. Recall that computing these marginals take $O(T)$ time to pass messages over the entire sequence ${\mathcal{T}}$. This defeats the purpose of using a subsequence. If we instead only pass messages over the subsequence ${{\mathcal{S}}}$, then the pairwise marginals are $p({u}_t, {u}_{t-1} \, | \, y_{{\mathcal{S}}})$ and we return to a biased gradient estimator $$\label{eq:naive_grad_approx} \hat{g}(\theta) = -\sum_{t \in {{\mathcal{S}}}} \frac{1}{\Pr(t \in {{\mathcal{S}}})} {\mathbb{E} \,}_{u | y_{{\mathcal{S}}}, \theta}[{\nabla}\log p(y_t, {u}_t \, | \, {u}_{t-1}, \theta)] - {\nabla}\log p(\theta) \enspace.$$ Approximate Gradient Estimate {#sec:framework-approx-gradient} ----------------------------- ![Graphical model of a buffered subsequence with $S = 3$ and $B=2$.[]{data-label="fig:ssm_buffer"}](figures/latex_fig/buffer.pdf){width="0.99\linewidth"} We instead propose passing messages over a *buffered* subsequence ${{\mathcal{S}^*}}:= \{t_{-B}, \ldots, t_{S+B} \}$ for some positive buffer size $B$, with ${{\mathcal{S}}}\subset {{\mathcal{S}^*}}\subset {\mathcal{T}}$ (see Figure \[fig:ssm\_buffer\]). The idea is that there exists a large enough $B$ such that $p(u_{{\mathcal{S}}}\, | \, y_{{\mathcal{S}^*}}, \theta) \approx p(u_{{\mathcal{S}}}\, | \, y_{\mathcal{T}}, \theta)$. Our *buffered gradient estimator* sums only over ${{\mathcal{S}}}$, but takes expectations over $u_{{\mathcal{S}}}\, | \, y_{{\mathcal{S}^*}}, \theta$ instead of $u_{{\mathcal{S}}}\, | \, y_{\mathcal{T}}, \theta$ $$\label{eq:efficient_potential_estimate} \tilde{g}(\theta) = - \sum_{t \in {{\mathcal{S}}}}\frac{1}{\Pr(t \in {{\mathcal{S}}})} {\mathbb{E} \,}_{{u}| y_{{{\mathcal{S}^*}}}, \theta} \left[ {\nabla}\log p(y_t, {u}_t \, | \, {u}_{t-1}, \theta) \right] - {\nabla}\log p(\theta) \enspace,$$ where $p({u}_{t_{-B-1}}) = p_0({u}_{t_{-B-1}})$. When $B = 0$ this is equivalent to the biased estimator $\hat{g}(\theta)$ of Eq. . When $B = T$ this is equivalent to the unbiased estimator $\bar{g}(\theta)$ of Eq. . The trade-off between accuracy (bias) and runtime depends on the size of the buffer $B$ and current model parameters $\theta^{(s)}$. Intuitively, when $\theta^{(s)}$ produces pairwise marginals that are similar to i.i.d. data, we can use a small buffer $B$. When $\theta^{(s)}$ produces strongly dependent pairwise marginals, we must use a larger buffer $B$. In Section \[sec:bounds\], we analyze, for a fixed value of $\theta$, how quickly the bias between $\bar{g}(\theta)$ and $\tilde{g}(\theta)$ decays with increasing $B$. We show a geometric decay $$\label{eq:geom_decay_rate} {\mathbb{E} \,}_{{\mathcal{S}}}\| \bar{g}(\theta) - \tilde{g}(\theta) \|_2 \leq C_\theta \rho_\theta^{-B} \enspace, \enspace \text{ for some } C_\theta > 0 \enspace,$$ where $\rho_\theta$ is large for i.i.d. data and small for strongly dependent data. The term $C_\theta$ depends on the smoothness of $g(\theta)$ and how accurately $p_0({u}_{t_{-B-1}})$ approximates $p({u}_{t_{-B-1}} \, | \, y_{{\mathcal{T}}\backslash{{\mathcal{S}^*}}})$. For a gradient accuracy of $\epsilon$, we only need a logarithmic buffer size $O(\log\epsilon^{-1})$.[^4] Therefore our buffered gradient estimator reduces the computation time from $O(T)$ to $O(S + \log\epsilon^{-1})$. By using buffered stochastic gradients $\tilde{g}$ with an appropriate buffer size $B$ in SGMCMC (Eq.  or ), we can generate samples $\theta^{(s)}$ that are close to the samples that would be generated if we were to use the unbiased (but intractable) stochastic gradients $\bar{g}$. In our experiments (Section \[sec:experiments\]), we find that modest buffers significantly correct for bias. Our approach is similar to fixed-lag smoothing methods in the particle filter literature [@chan2016theory; @del2017biased; @olsson2008sequential], which approximate $p({u}_t \, | \, y_{1:T}, \theta)$ using a right buffer $p({u}_t \, | \, y_{1:t+B}, \theta)$ in a streaming fashion. However, our approach, Eq. , differs by using both a left and a right buffer $p({u}_t \, | \, y_{1:T}, \theta) = p({u}_t \, | \, y_{t-B:t+B})$, which allow us to avoid a full passes over the data. Preconditioning and Fisher Information {#sec:framework-sgrld} -------------------------------------- The desirable properties for the preconditioning matrix $D(\theta)$ for SGRLD are (i) the resulting dynamics takes advantage of the geometric structure of $\theta$, (ii) both $D(\theta)$ and $\Gamma(\theta)$ can be efficiently computed, and (iii) neither $D(\theta)g(\theta)$ nor $\Gamma(\theta)$ are numerically unstable. The *expected Fisher information* $\mathcal{I}_y$ is the Riemannian metric proposed in [@girolami2011riemann] $$\label{eq:preconditioning_with_fisher} D^{-1}(\theta) = \mathcal{I}_y = {\mathbb{E} \,}_{y\,|\,\theta}\left[{\nabla}^2 \log p(y \, | \, \theta) \right] \enspace.$$ Unfortunately for SSMs, the lack of independence in the marginal likelihood requires a double sum over ${\mathcal{T}}$ to compute $\mathcal{I}_y$, which is computationally intractable for long time series. We instead replace $I_{y}$ with the *complete data Fisher information* $I_{{u}, y}$ $$\mathcal{I}_{{u},y} = {\mathbb{E} \,}_{{u}, y \,|\, \theta}\left[{\nabla}^2 \log p(y, {u}\, \ \, \theta) \right] = T \cdot {\mathbb{E} \,}_{{u}, y \, | \, \theta}\left[{\nabla}^2 \log p(y_t, {u}_t \, | \, {u}_{t-1}, \theta)\right] \enspace.$$ Because $I_{{u},y}$ can be calculated analytically for the SSMs we consider (Section \[sec:models\]), we use $D(\theta) = I_{{u},y}^{-1}$ when possible or approximations of $I_{{u}, y}^{-1}$ when not (see the Supplement for details). In our experiments, we find that in practice, using preconditioning works well and outperforms vanilla SGLD. Algorithm Pseudocode {#sec:framework-algorithm} -------------------- Algorithms \[alg:sgrld\] and \[alg:noisygradient-analytic\] summarize our generic SGMCMC method for SSMs[^5]. To select the buffer size $B$ in Algorithm \[alg:noisygradient-analytic\], we choose $B$ large enough such that the error using $B$ and a larger buffer size $B^*$ is small: $$\label{eq:adaptive_B} B = \min \left\{\hat{B} \in [0, B^*] \, : \, {\mathbb{E} \,}_{{\mathcal{S}}}\|\, \tilde{g}(\theta, {{\mathcal{S}}}, \hat{B}) - \tilde{g}(\theta, {{\mathcal{S}}}, B^*) \| < \epsilon \right\}$$ where $\tilde{g}(\theta, {{\mathcal{S}}}, B) = {\mathbb{E} \,}_{{u}| y_{{{\mathcal{S}^*}}}, \theta}[ {\nabla}\log p(y_{{\mathcal{S}}}, {u}_{{\mathcal{S}}}\, | \, \theta) ]$ and the expectation over ${{\mathcal{S}}}$ is approximated with an empirical average over $N_S$ subsequences. Eq.  uses $\tilde{g}(\theta, {{\mathcal{S}}}, B^*)$ as a proxy for $\tilde{g}(\theta, {{\mathcal{S}}}, T)$. As the error decays geometrically (Section \[sec:bounds\]), we found using $B^* = 100$ was conservative in practice. Calculating $B$ using Eq.  at every iteration for a new $\theta^{(s)}$ is impractical; therefore for our experiments, we use a fixed $B$, estimated using $\theta$ from a pilot run with $B=B^*$ and $N_S = 1000$. In addition, instead of evaluating each $\hat{B}$ in $[0, B^*]$, we can estimate the required $B$ for a target error tolerance $\epsilon$ after estimating the error $\hat{\epsilon}$ of a single $\hat{B}$, by taking advantage of the geometric error scaling rate, Eq. , to obtain $B = \hat{B} + \log_{\rho_\theta}(\hat{\epsilon}/\epsilon)$ where $\rho_\theta$ is a bound on the geometric decay rate from theory. Buffered Gradient Estimator Error Bounds {#sec:bounds} ======================================== In this section, we establish a bound on the expected error between the unbiased gradient $\bar{g}(\theta)$ and our buffered gradient estimator $\tilde{g}(\theta)$ Eq. . Given such a bound, we can control the overall error in our SGLD or SGRLD scheme when the SGMCMC dynamics possess a contraction property [@johndrow2017error]. Specifically, if we can uniformly bound $\| \bar{g}(\theta) - \tilde{g}(\theta) \|_2 < \delta$, then the difference in a single step of SGMCMC, Eq. , using the unbiased and approximate gradients $\bar{g}$ and $\tilde{g}$ is bounded by $\delta h$. Therefore we can apply Theorem 1.11 of [@johndrow2017error] which states the sample average of a test function evaluated on samples of the approximate-gradient $\tilde{g}$ chain, $\sum_{i < s} \varphi(\theta^{(i)})/s$, converges to the posterior expected value of the unbiased-gradient $\bar{g}$ chain, ${\mathbb{E} \,}_\theta[\varphi(\theta)]$, plus an additional error term proportional to $\delta h$. For our analysis, we first consider the simple case of uniformly sampling a single sequence from $T/S$ separate subsequences (i.e. $\Pr(t \in {{\mathcal{S}}}) = S/T$ for all $t$) and assume the prior $p_0$ is stationary (i.e. $p_0(u_t) = \int p(u_t | u_{t-1}) p_0(u_{t-1}) du_{t-1}$). Our approach is to bound $\| \bar{g}(\theta) - \tilde{g}(\theta)\|_2$ in terms of the Wasserstein distance between the exact posterior $\gamma_t({u}_t) = p({u}_t \, | \, y_{\mathcal{T}}, \theta)$ and our approximate posterior $\widetilde{\gamma}_t({u}_t) = p({u}_t \, | \, y_{{\mathcal{S}^*}}, \theta)$ and then show this Wasserstein distance decays geometrically. To bound the Wasserstein distance, we follow existing work on bounding Markov processes in Wasserstein distance [@durmus2015quantitative; @madras2010quantitative; @rudolf2015perturbation]. However, unlike previous work that focuses on the homogeneous Markov process of the joint model $\{{u}, y \, | \, \theta\}$, we instead focus on the induced *nonhomogeneous* Markov process of the conditional model $\{{u}\, | \, y, \theta\}$. To do so, we use the forward ($f_t$) and backward ($b_t$) *random maps* of $\{{u}\, | \, y, \theta\}$ [@diaconis1999iterated] $$\begin{aligned} {u}_t \sim p({u}_t \, | \, y, \theta) \enspace &\Rightarrow \enspace (f_t({u}_t), {u}_t) \sim p({u}_{t+1}, {u}_t \, | \, y, \theta) \\ {u}_t \sim p({u}_t \, | \, y, \theta) \enspace &\Rightarrow \enspace (b_t({u}_t), {u}_t) \sim p({u}_{t-1}, {u}_t \, | \, y, \theta) \enspace,\end{aligned}$$ If $f_t$ and $b_t$ satisfy a contractive property, then we can bound the Wasserstein distance between $\gamma_t, \widetilde\gamma_t$ in terms of $\gamma_{t-1}, \widetilde\gamma_{t-1}$ and $\gamma_{t+1}, \widetilde\gamma_{t+1}$ respectively. Bounding the error of the induced nonhomogeneous Markov process has been previously studied in the SSM literature using total variation (TV) distance [@cappe2005inference; @del2010forward; @le2000exponential; @tong2012ergodicity]. These works bound the error in total variation distance by quantifying how quickly the smoothed posterior forgets the initial condition. However, these bounds typically require stringent regularity conditions, which are hard to prove outside of finite or compact spaces[^6]. In particular, these bounds are not immediately applicable for LGSSMs. In contrast, we bound the error in Wasserstein distance by proving contraction properties of $f_t$ and $b_t$, allowing us to handle continuous and mixed-type SSMs such as the LGSSM (Section \[sec:models:LGSSM-bounds\]). Our main result is that if, for each fixed $\theta$, the gradient of $\log p(y, {u}\, | \, \theta)$ satisfies a Lipschitz condition and the random maps $\{f_t, b_t\}_{t \in {{\mathcal{S}^*}}}$ all satisfy a contraction property, then the error $\| \bar{g}(\theta) - \tilde{g}(\theta)\|_2$ decays geometrically in the buffer size $B$. \[theorem:geometric\_error\_bound\] Let $\epsilon_{\rightarrow}$ and $\epsilon_{\leftarrow}$ be the 1-Wasserstein distances between $\gamma_t$ and $\widetilde{\gamma}_t$ at the left and right ends of ${{\mathcal{S}^*}}$ respectively. Let $\epsilon_1 = \max_{{{\mathcal{S}^*}}\subset {\mathcal{T}}}\{\epsilon_\rightarrow, \epsilon_\leftarrow\}$. If the gradients of $\log p(y_t, {u}_t \, | \, {u}_{t-1}, \theta)$ are all Lipschitz in ${u}_{t-1:t}$ with constant $L_U$, and random maps $f_t$ and $b_t$ are all Lipschitz[^7] in ${u}_t$ with constant $L < 1$, then we have $$\| \bar{g}(\theta) - \tilde{g}(\theta)\|_2 \leq T \cdot L_U \cdot \frac{1+L}{1-L} \cdot \frac{1-L^S}{S} \cdot L^{B} \cdot 2\epsilon_1 \enspace.$$ A similar result for when the gradient of the complete data loglikelihood is Lipschitz in $uu^T$ instead of $u$ (as needed for LGSSM) will be proved in Section \[sec:bounds-main\]. As $L < 1$, Theorem \[theorem:geometric\_error\_bound\] states that the error of the buffered gradient estimator decays geometrically as $O(L^B)$. Therefore, the required buffer size $B$ for an error tolerance of $\delta$ scales logarithmically as $O(\log \delta^{-1})$. In contrast, the error of the gradient estimator decays only linearly in the subsequence length, $O(S^{-1})$; therefore much longer subsequences, $O(\delta^{-1})$, are required to reduce bias. This agrees with the intuition that the bias is dominated by the error at the endpoints of subsequence. Theorem \[theorem:geometric\_error\_bound\] requires bounding the Lipschitz constants of the gradient of the complete data loglikelihood and the random maps $f_t, b_t$ given the parameters $\theta$ and observations $y_{\mathcal{T}}$. We show examples of these bounds for specific models in Section \[sec:models:HMM-bounds\] (HMMs) and \[sec:models:LGSSM-bounds\] (LGSSMs). Theorem \[theorem:geometric\_error\_bound\] also depends on the maximum Wasserstein distance $\epsilon_1$ between $\gamma_t$ and $\widetilde{\gamma}_t$ for all ${{\mathcal{S}^*}}\subset {\mathcal{T}}$ and $t \in {\mathcal{T}}$, which is finite. The remainder of this section is as follows. First, in Section \[sec:bounds-wasserstein\], we show how to bound the error in $\bar{g}, \tilde{g}$ in terms of Wasserstein distances between $\gamma, \widetilde\gamma$. Second, in Section \[sec:bounds-geometric-decay\], we show these Wasserstein distances decay geometrically in $B$. Finally, in Section \[sec:bounds-main\], we prove our main results: Theorems \[theorem:geometric\_error\_bound\] and \[theorem:geometric\_error\_bound2\], and discuss relaxations of the assumptions on the sampling of subsequences ${{\mathcal{S}}}$ and the prior $p_0$. To keep the presentation clean, we leave proofs of Lemmas to the Supplement. Functional Bound in terms of Wasserstein {#sec:bounds-wasserstein} ---------------------------------------- We first review the definition of Wasserstein distance. Let $\mathcal{W}_p(\gamma, \widetilde\gamma)$ be the $p$-Wasserstein distance $$\mathcal{W}_p(\gamma, \widetilde\gamma) := \left[\inf_\xi \int \| {u}- \widetilde{u}\|_2^p \, d\xi({u}, \widetilde{u}) \right]^{1/p}$$ where $\xi$ is a joint measure or *coupling* over $({u}, \widetilde{u})$ with marginals $\int_{\widetilde{{u}}} d\xi({u}, \widetilde{{u}}) = d\gamma({u}) $ and $\int_{{u}} d\xi({u}, \widetilde{{u}}) = d\widetilde{\gamma}(\widetilde{{u}})$. Wasserstein distance satisfies all the properties of a metric. A useful property of the $1$-Wasserstein distance is the following Kantorovich-Rubinstein duality formula for the difference of expectations of Lipschitz functions [@villani2008optimal] $$\label{eq:kantorovich-rubinstein} \mathcal{W}_1(\gamma, \widetilde\gamma) = \sup_{\|\psi\|_{Lip} \leq 1}\left\{ \int \psi \, d\gamma - \int \psi \, d\widetilde\gamma \right\} \enspace \Rightarrow \enspace |{\mathbb{E} \,}_\gamma[\psi] - {\mathbb{E} \,}_{\widetilde\gamma}[\psi] | \leq \| \psi \|_{Lip} \cdot \mathcal{W}_1(\gamma, \widetilde\gamma) \enspace,$$ where $\| \psi \|_{Lip}$ denotes the Lipchitz constant of $\psi$. We connect the error $\|\bar{g}-\tilde{g}\|_2$ to the Wasserstein distances between $\gamma, \widetilde\gamma$, by applying this duality formula Eq.  to the difference of Eqs.  and $$\label{eq:diff_gradient_estimates} \bar{g}(\theta) - \tilde{g}(\theta) = \frac{T}{S} \sum_{t \in {{\mathcal{S}}}} {\mathbb{E} \,}_{\gamma_{t-1:t}}\left[{\nabla}\log p(y_t, {u}_t | {u}_{t-1}, \theta )\right] - {\mathbb{E} \,}_{\widetilde\gamma_{t-1:t}}\left[{\nabla}\log p(y_t, {u}_t | {u}_{t-1}, \theta )\right] .$$ Applying the triangle inequality gives Lemma \[lemma:grad\_error\_to\_wasserstein\]. \[lemma:grad\_error\_to\_wasserstein\] If ${\nabla}\log p(y_t, {u}_t | {u}_{t-1}, \theta)$ are Lipschitz in ${u}_{t-1:t}$ with constant $L_U$, $$\| \bar{g}(\theta) - \tilde{g}(\theta) \|_2 \leq \frac{T}{S} \cdot L_U \cdot \sum_{t \in {{\mathcal{S}}}} \mathcal{W}_1(\gamma_{t-1:t}, \widetilde\gamma_{t-1:t}).$$ If ${\nabla}\log p(y_t, {u}_t \, | \, {u}_{t-1}, \theta)$ is not Lipschitz in ${u}_{t-1:t}$, but is Lipschitz in ${u}_{t-1:t}{u}_{t-1:t}^T$ (as in LGSSMs), then the following Lemma lets us bound the $1$-Wasserstein distance of ${u}{u}^T$ in terms of the $2$-Wasserstein distance of ${u}$. \[lemma:second\_wass\_moment\] Let $\gamma'$ be the distribution of ${u}{u}^T$. Let $\widetilde\gamma'$ be the distribution of $\widetilde{u}\widetilde{u}^T$. Let $M = {\mathbb{E} \,}_\gamma[\|{u}\|_2^2] < \infty$. (Note $\mathcal{W}_2(\gamma, \widetilde\gamma) < \infty$ implies ${\mathbb{E} \,}_\gamma[\|{u}\|_2^2] < \infty$.) Then, $$\mathcal{W}_1(\gamma', \widetilde\gamma') \leq (2\sqrt{M} + 1) \cdot \max\left\{\mathcal{W}_2(\gamma, \widetilde\gamma)^{1/2},\mathcal{W}_2(\gamma, \widetilde\gamma)\right\} \enspace.$$ Geometric Wasserstein Decay {#sec:bounds-geometric-decay} --------------------------- We first review why contractive random maps induce Wasserstein bounds. If two distributions $\gamma_t, \gamma_t'$ have identically distributed random maps $f_t, f_t'$, that is there exists a random function $f_t$ satisfying $$u \sim \gamma_t \text{ and } u' \sim \gamma_t' \enspace \Rightarrow \enspace f_t(u) \sim \gamma_{t+1} \text{ and } f_t(u') \sim \gamma_{t+1}' \enspace,$$ then we can bound the Wasserstein distance of $\gamma_{t+1}, \gamma_{t+1}'$ in terms of the Wasserstein distance of $\gamma_t, \gamma_t'$ given a bound on the random map’s Lipschitz constant $\| f_t \|_{Lip} < L$ $$\begin{aligned} \label{eq:wasserstein_random_map} \mathcal{W}_p(\gamma_{t+1}, \gamma_{t+1}')^p &= \inf_{\xi_{t+1}} \int \| {u}_{t+1} - {u}_{t+1}'\|_2^p \, d\xi_{t+1}({u}_{t+1}, {u}_{t+1}') \\ \nonumber &\leq \inf_{\xi_t} \int \| f_t({u}_t) - f_t({u}_t') \|_2^p \, d\xi_t({u}_t, {u}_t') df_t \\ \nonumber &\leq \inf_{\xi_t} \int L^p \cdot \| {u}_t - {u}_t' \|_2^p \, d\xi_t({u}_t, {u}_t') = L^p \cdot \mathcal{W}_p(\gamma_t, \gamma_t')^p\enspace.\end{aligned}$$ Unfortunately for SSMs, Eq.  does not apply as the random maps $f_t, b_t$ of $\gamma$ and $\widetilde{f}_t, \widetilde{b}_t$ of $\widetilde\gamma$ are *not identically* distributed. To see this, we first review the conditional probability distributions used to define $f_t, b_t$. The forward random map $f_t$ draws ${u}_{t+1} \, | \, {u}_t$ from the *forward smoothing kernel* $$\label{eq:forward_smoothing_kernel} \mathcal{F}_t({u}_{t+1} \, | \, {u}_{t}) := p({u}_{t+1} \, | \, {u}_t, y_{>t}) = p({u}_{t+1} \, | \, {u}_t) p(y_{t+1} \, | \, {u}_{t+1}) \beta_{t+1}({u}_{t+1}) / \beta_t({u}_{t})$$ and the backward random map $b_t$ draws ${u}_{t-1} \, | \, {u}_t$ from the *backward smoothing kernel* $$\label{eq:backward_smoothing_kernel} \mathcal{B}_t({u}_{t-1} \, | \, {u}_{t}) := p({u}_{t-1} \, | \, {u}_t, y_{\geq t}) = p({u}_t \, | \, {u}_{t-1})p(y_t \, | \, {u}_t) \alpha_{t-1}({u}_{t-1})/\alpha_t({u}_t) \enspace.$$ Because $\widetilde\gamma$ uses different forward and backward messages $\widetilde\alpha$, $\widetilde\beta$ in Eqs.  and , the kernels $\widetilde{\mathcal{F}}_t, \widetilde{\mathcal{B}}_t$ are not identical to $\mathcal{F}_t, \mathcal{B}_t$ (and the random maps are *not* identically distributed). This is unlike homogeneous Markov chains, where the kernels are identical at each time $t$ (and the random maps are identically distributed). Instead of connecting $\gamma$ to $\widetilde\gamma$ directly, we use the triangle inequality to connect them through an intermediate distribution $\widehat{\gamma} := p({u}\, | \, y_{t \geq t_{-B}}, \theta)$ $$\label{eq:wasserstein_triangle} \mathcal{W}_p(\gamma, \widetilde\gamma) \leq \mathcal{W}_p(\gamma, \widehat\gamma) + \mathcal{W}_p(\widehat\gamma, \widetilde\gamma) \enspace.$$ Introducing this particular intermediate distribution $\widehat{\gamma}$ is the key step for our Wasserstein bounds between $\gamma$ and $\widetilde\gamma$. Because $\widehat{\gamma}$ conditions on all $y_t$ after $y_{{\mathcal{S}^*}}$, $\widehat{\gamma}$ and $\gamma$ have identical backward messages $\beta_t$ and therefore identically distributed forward random maps $f_t$. Similarly, because $\widehat{\gamma}$ does not condition on $y_t$ before $y_{{\mathcal{S}^*}}$, $\widehat{\gamma}$ and $\widetilde{\gamma}$ have identical forward messages $\widetilde{\alpha_t}$ and identically distributed backward random maps $\widetilde{b}_t$. Therefore, we can bound $\mathcal{W}_p(\gamma, \widehat\gamma)$ using $f_t$ and bound $\mathcal{W}_p(\widehat\gamma, \widetilde\gamma)$ using $\widetilde{b}_t$ with the contraction trick Eq.  giving us Lemma \[lemma:geometric\_wasserstein\]. \[lemma:geometric\_wasserstein\] If there exists $L_f, L_b < 1$ such that for all $t \in {{\mathcal{S}^*}}$, $\| f_t\|_{Lip} < L_f$ and $\| \widetilde{b}_t \|_{Lip} < L_b$, then for all $t \in {{\mathcal{S}}}$ we have $$\begin{aligned} \label{eq:geometric_wasserstein_forward} \mathcal{W}_p(\gamma_{t-1:t}, \widehat\gamma_{t-1:t}) &\leq (1 + L_f^p)^{1/p} \cdot \mathcal{W}_p(\gamma_{t-1}, \widehat\gamma_{t-1}) \\ \nonumber &\leq (1 + L_f^p)^{1/p} \cdot L_f^{t-1-t_{-B}} \cdot \mathcal{W}_p(\gamma_{t_{-B}}, \widehat\gamma_{t_{-B}}) \\ \label{eq:geometric_wasserstein_backward} \mathcal{W}_p(\widehat\gamma_{t-1:t}, \widetilde\gamma_{t-1:t}) &\leq (1 + L_b^p)^{1/p} \cdot \mathcal{W}_p(\widehat\gamma_{t}, \widetilde\gamma_{t}) \\ \nonumber &\leq (1 + L_b^p)^{1/p} \cdot L_b^{t_{S+B}-t} \cdot \mathcal{W}_p(\widehat\gamma_{t_{S+B}}, \widetilde\gamma_{t_{S+B}})\end{aligned}$$ We show sufficient conditions for the random maps to be contractions (i.e. $L_f, L_b < 1$) for specific models in Section \[sec:models:HMM-bounds\] (HMMs) and \[sec:models:LGSSM-bounds\] (LGSSMs). Proof of Main Theorems {#sec:bounds-main} ---------------------- Putting together the results of the previous two subsections gives us our geometric error bounds: Theorem \[theorem:geometric\_error\_bound\] when the gradient terms are Lipschitz in ${u}$ and Theorem \[theorem:geometric\_error\_bound2\] when the gradient terms are Lipschitz in ${u}{u}^T$. Both theorems require the random maps of the forward and backward smoothing kernels are contractions. We first prove Theorem \[theorem:geometric\_error\_bound\]. Combining Lemmas \[lemma:grad\_error\_to\_wasserstein\] and \[lemma:geometric\_wasserstein\] with some algebra $$\begin{aligned} \| \bar{g}(\theta) - \tilde{g}(\theta)\|_2 &\leq \frac{T}{S} \cdot L_U \cdot \sum_{t \in {{\mathcal{S}}}} \mathcal{W}_1(\gamma_{t-1:t}, \widetilde\gamma_{t-1:t}) \\ &\leq \frac{T}{S} \cdot L_U \cdot \sum_{t \in {{\mathcal{S}}}} \mathcal{W}_1(\gamma_{t-1:t}, \widehat\gamma_{t-1:t}) + \mathcal{W}_1(\widehat\gamma_{t-1:t}, \widetilde\gamma_{t-1:t}) \\ &\leq \frac{T}{S} \cdot L_U \cdot \sum_{t = 1}^S (1+L_f) L_f^{B+t-1} \epsilon_1 + (1+L_b) L_b^{B+S-t}\epsilon_1 \\ &\leq T \cdot L_U \cdot \frac{1+L}{1-L} \cdot \frac{1-L^S}{S} \cdot L^{B} \cdot 2\epsilon_1 \enspace,\end{aligned}$$ where $\max_{{{\mathcal{S}^*}}\subset {\mathcal{T}}}\{\mathcal{W}_1(\gamma_{t_{-B}}, \widehat{\gamma}_{t_{-B}}), \mathcal{W}_1(\widehat\gamma_{t_{S+B}}, \widetilde{\gamma}_{t_{S+B}})\} = \max_{{{\mathcal{S}^*}}\subset {\mathcal{T}}}\{\epsilon_{\rightarrow}, \epsilon_{\leftarrow}\} = \epsilon_1$. We now prove a similar result for when ${\nabla}\log p(y, {u}_t \,|\, {u}_{t-1} \theta)$ is Lipschitz in $uu^T$. \[theorem:geometric\_error\_bound2\] Let $\epsilon_2 = \max_{{{\mathcal{S}^*}}\subset {\mathcal{T}}}\{\mathcal{W}_2(\gamma_{t_{-B}}, \widehat{\gamma}_{t_{-B}}), \, \mathcal{W}_2(\widehat\gamma_{t_{S+B}}, \widetilde{\gamma}_{t_{S+B}})\}$. If the gradients are Lipschitz in ${u}{u}^T$ with constant $L_U'$, and there exists $L_f, L_b < 1$ for Lemma \[lemma:geometric\_wasserstein\], then with $L = \max\{L_f, L_b\}$ and $L_U = (2 \sqrt{{\mathbb{E} \,}_\gamma\|{u}\|^2_2} + 1) L_U'$ $$\| \bar{g}(\theta) - \tilde{g}(\theta)\|_2 \leq T \cdot L_U \cdot \frac{\sqrt{1+L^2}}{1-L^{1/2}} \cdot \frac{1-L^{S/4}}{S/2} \cdot L^{B/2} \cdot \max_{r \in \{ 1/2, \, 1 \}} (2\epsilon_2)^r \,.$$ Similar to Theorem \[theorem:geometric\_error\_bound\], Theorem \[theorem:geometric\_error\_bound2\] states that the squared error of the buffered gradient estimator decays geometrically if the complete-data loglikelihood is Lipschitz in $uu^T$ instead of $u$. However, the price we pay is a square-root: the error decays $O(L^{B/2})$ instead of $O(L^{B})$. Applying Lemmas \[lemma:second\_wass\_moment\] and \[lemma:geometric\_wasserstein\], we have $$\begin{aligned} \| \bar{g}(\theta) - \tilde{g}(\theta)\|_2 &\leq \frac{T}{S} \cdot L_U \cdot \sum_{t \in {{\mathcal{S}}}} \max_{r \in \{1/2, \, 1\}} \left[\mathcal{W}_2(\gamma_{t-1:t}, \widehat\gamma_{t-1:t}) + \mathcal{W}_2(\widehat\gamma_{t-1:t}, \widetilde\gamma_{t-1:t})\right]^r \\ &\leq \frac{T}{S} \cdot L_U \cdot \sum_{t = 1}^S \max_{r \in \{1/2, \, 1\}} \left[(L^{B +t-1} + L^{B+S-t}) \sqrt{1+L^2} \epsilon_2\right]^r \\ &\leq \frac{T}{S} \cdot L_U \cdot \sum_{t = 1}^S L^{(B + \min\{t-1, S-t\})/2} \cdot \sqrt{1+L^2} \cdot \max_{r \in \{1/2, \, 1\}} (2 \epsilon_2)^r \\ &\leq \frac{T}{S} \cdot L_U \cdot 2\cdot\frac{1-L^{S/4}}{1-L^{1/2}} \cdot L^{B/2} \cdot \sqrt{1+L^2} \cdot \max_{r \in \{ 1/2, \, 1 \}} (2 \epsilon_2)^r\end{aligned}$$ Our error analysis (Theorems \[theorem:geometric\_error\_bound\] and \[theorem:geometric\_error\_bound2\]) indicates that only a logarithmic buffer size is required to control the bias to a fixed error tolerance $\delta$. ### Relaxations of Assumptions We now briefly discuss relaxations of the assumptions on $\Pr(t \in {{\mathcal{S}}})$ and $p_0$. If the contiguous subsequences are not sampled from a strict partition (i.e. $\Pr(t \in {{\mathcal{S}}}) \neq S/T$ for all $t$), then we can replace the factor of $T/S$ in Theorems \[theorem:geometric\_error\_bound\] and \[theorem:geometric\_error\_bound2\] with $\max_t \Pr(t \in {{\mathcal{S}}})^{-1}$. Additional details on different sampling methods for ${{\mathcal{S}}}$ can be found in the Supplement. If the initial distribution for $u_{t_{-B-1}}$ of our buffered stochastic gradient, $p_0$, is not stationary, then our approximate posterior over the latent states $\tilde{\gamma}_t(u_t)$ is not equal to $p(u_t \, | \, y_{{\mathcal{S}^*}}, \theta)$. However Theorems \[theorem:geometric\_error\_bound\] and \[theorem:geometric\_error\_bound2\] will still apply; the choice of initial distribution only affects the Wasserstein distance between $\gamma_t, \tilde{\gamma}_t$ and therefore the terms $\epsilon_1, \epsilon_2$ in the Theorems. In fact, the optimal initial distribution is $p(u_{t_{-B}} | y_{{\mathcal{T}}\backslash {{\mathcal{S}^*}}})$, which minimizes the Wasserstein distance of $\gamma, \tilde{\gamma}$. Example Models {#sec:models} ============== In this section, we provide examples of how to apply the generic framework of Section \[sec:framework\] and bounds of Section \[sec:bounds\] to common SSMs. Gaussian HMM {#sec:models:HMM} ------------ We consider discrete latent state HMMs with Gaussian emissions. The complete data likelihood of a Gaussian HMM is as follows $$\label{eq:gaussian_hmm_model} p(y, z \, | \, \theta) = \prod_{t=1}^T \Pi_{z_{t-1}, z_{t}} \cdot \mathcal{N}(y_t \, | \, \mu_{z_t} , \Sigma_{z_k}) \enspace,$$ where $y_t \in {\mathbb{R}}^m$ are the observations, ${u}_t \equiv z_t \in \{1, \ldots, K\}$ are the discrete latent variables, and $\theta = \{\Pi, \mu, \Sigma\}$ are the parameters with $\Pi_k \in \Delta^K$ (simplex over $K$ states), $\mu_k \in {\mathbb{R}}^{m}$, $\Sigma_k \in \mathbb{S}^{m}_+$ (positive definite matrices) for $k = 1,\ldots, K$. In practice, we use the *expanded mean* parameters of $\Pi$ instead of $\Pi$ (as in [@patterson2013stochastic]) and the *Cholesky decomposition* of $\Sigma^{-1}_k$ instead of $\Sigma_k$ to ensure positive definiteness. As the latent states are discrete over a finite space, the forward backward algorithm for an HMM can be done in closed-form; thus, pairwise latent marginals $\gamma_{t-1:t}(z_{t-1}, z_t)$, gradients ${\nabla}U(\theta)$ and preconditioning terms $D(\theta)$ and $\Gamma(\theta)$ are straightforward to calculate. Complete details are provided in the Supplement. ### Error Bound Coefficients {#sec:models:HMM-bounds} In the finite discrete variable case, conditions for bounding the Lipschitz coeffficient of the smoothing kernels $\mathcal{F}_t, \mathcal{B}_t$ (as needed for Section \[sec:bounds-geometric-decay\]) are equivalent to conditions for bounding their *Dobrushin coefficients* [@cappe2005inference; @del2010forward]. The Dobrushin coefficient for a transition kernel $\mathcal{Q}$ is $$\label{eq:dobrushin} \delta(\mathcal{Q}) = \sup_{z, z'} \frac{1}{2} \| \mathcal{Q}(z, \cdot) - \mathcal{Q}(z', \cdot) \|_{TV} = \frac{\| \mathcal{Q}(z, \cdot) - \mathcal{Q}(z', \cdot) \|_{TV}}{\| \delta_z - \delta_{z'}\|_{TV}} \enspace.$$ The final term of Eq.  show the connection between Dobrushin coefficients and Lipschitz coefficients: it is the ratio of the distance of between kernels $\mathcal{Q}(z, \cdot), \mathcal{Q}(z', \cdot)$ with the distance between point masses at $z$ and $z'$. Therefore for discrete latent states, $L_f = \max_t \delta(\mathcal{F}_t)$ and $L_b = \max_t \delta(\mathcal{B}_t)$. In the discrete case, sufficient conditions for $L_f, L_b < 1$ are well known (See [@cappe2005inference] Chapter 4.3). If the transition matrix $\Pi$ satisfies the *strong mixing condition*, that is, there exists constants $\sigma^{-}$ and $\sigma^{+}$ with $0 < \sigma^{-} \leq \sigma^{+}$ and a probability distribution $\kappa \in \Delta^{K}$ over $z$ such that $\sigma^{-} \kappa(z') \leq \Pi_{z, z'} \leq \sigma^{+} \kappa(z')$ and ${\mathbb{E} \,}_\kappa[p(y\, |\, z)] < \infty$, then the Dobrushin coefficients are bounded by $L = 1-\sigma^{-}/\sigma^{+}$. Relaxations of this condition can be found in [@cappe2005inference; @del2010forward]. Alternatively, we can obtain tighter bounds for HMMs via estimating the Lyapunov exponents for the underlying random dynamical systems defined by random maps $f_t$ and $b_t$ [@ye2017estimate; @ma2017stochastic]. Finally, the Lipschitz constant $L_U$ for Lemma \[lemma:grad\_error\_to\_wasserstein\] is $$L_U = \max_{t \in {{\mathcal{S}}}, z_t, z_t'} \| {\nabla}\log p(y_t, z_t \, | \, z_{t-1}, \theta) - {\nabla}\log p(y_t, z_t' \, | \, z_{t-1}', \theta) \| \enspace.$$ This is easy to compute since at each iteration $y$ and $\theta=\theta^{(s)}$ are fixed. Given these bounds on $L_U$ and $L$, we can use Theorem \[theorem:geometric\_error\_bound\] to select the buffer size $B$ to ensure approximate convergence to the stationary distribution. Autoregressive HMM {#sec:models:ARHMM} ------------------ We now consider ARHMMs, a generalization of the discrete state HMM where each observation depends not only on the latent state, but also on the last $p$ observations. Specifically, the discrete latent state $z_t$ determines which AR($p$) process models the dynamics of $y$ at time $t$. The complete data likelihood of an ARHMM is as follows $$\label{eq:arhmm_model} p(y, z \, | \, \theta) = \prod_{t=1}^T \Pi_{z_{t-1}, z_{t}} \cdot \mathcal{N}(y_t \, | \, A_{z_t} \overline{y_{t}}, Q_{z_k}) \enspace,$$ where $y_t \in {\mathbb{R}}^m$ are the observations, $\overline{y_{t}} = y_{t-1:t-p}$ are the $p$-lagged observations, ${u}_t \equiv z_t \in \{1, \ldots, K\}$ are the discrete latent variables, and $\theta = \{\Pi, A, Q\}$ are the parameters with $\Pi_k \in \Delta^K$, $A_k \in {\mathbb{R}}^{m \times mp}$, $Q_k \in \mathbb{S}^{m}_+$ for $k = 1,\ldots, K$. From Eq. , we see that the ARHMM is a time-dependent mixture of $K$ AR processes of order $p$. The pairwise latent marginals, gradients, and preconditioning terms for an ARHMM are calculated similarly to the Gaussian HMM. Further details are provided in the Supplement. The theory and constants for the error bounds of Section \[sec:bounds\] are identical to those presented for the Gaussian HMM. Linear Gaussian SSM {#sec:models:LGSSM} ------------------- A linear Gaussian SSM (LGSSM), also called a linear dynamical system (LDS), consists of a latent Gaussian (vector) autoregressive process over states ${u}_t \equiv x_t \in {\mathbb{R}}^n$ and conditionally Gaussian emissions $y_t \in {\mathbb{R}}^m$ [@bishop2006pattern; @lutkepohl2005new]. Specifically, $$p(y, x \, | \, \theta) = \prod_{t=1}^T \mathcal{N}(x_t \, | \, A x_{t-1}, Q) \cdot \mathcal{N}(y_t \, | \, C x_t, R) \enspace,$$ where $A \in {\mathbb{R}}^{n \times n}$ is the latent state transition matrix, $Q \in \mathbb{S}^{n}_+$ is the transition noise covariance, $C \in {\mathbb{R}}^{m \times n}$ is the emission matrix, and $R \in \mathbb{S}^m_+$ is the emission noise covariance. Together $A, Q, C, R$ are the model parameters $\theta$. The matrices $A$, $C$, and $Q$ are unidentifiable without additional restriction, as applying an orthonormal transformation $M$ gives an equivalent representation $\tilde{A} = MAM^{-1}$, $\tilde{C}=CM^{-1}$, $\tilde{Q} = MQM^T$. To enforce identifiability, we choose to restrict the first $\min(n,m)$ rows and columns of $C$ to the identity matrix. In practice, we use the Cholesky decompositions $\psi_Q, \psi_R$ of $Q^{-1}, R^{-1}$ (respectively) instead of $Q, R$. The recursions for the forward backward algorithm for LGSSMs is known as the Kalman smoother [@cappe2005inference; @bishop2006pattern; @fox2009bayesian]. Because the transition and emission processes are linear Gaussian, all forward messages, backward messages, and pairwise latent marginals $\gamma_{t-1:t}(x_{t-1}, x_{t})$ are Gaussian; therefore, the gradients and preconditioning matrix can be calculated analytically. Further details are provided in the Supplement. ### Error Bound Coefficients {#sec:models:LGSSM-bounds} The random maps of an LGSSM are strict contractions under mild conditions (Lemmas \[lemma:LGSSM\_forward\_maps\], \[lemma:LGSSM\_backward\_maps\]) and the gradients are Lipschitz in $xx^T$ (Lemma \[lemma:LGSSM\_Lipschitz\]). Therefore, Theorem \[theorem:geometric\_error\_bound2\] applies. \[lemma:LGSSM\_forward\_maps\] The forward random maps of an LGSSM are Gaussian linear maps. Specifically, $f_t(x_t) = F^{f}_t x_t + \zeta^{f}_t$, where $\zeta^{f}_t$ is a Gaussian random intercept and $F^{f}_t$ is a matrix function of $\theta$ and $y_{>t}$. As a linear map, the Lipschitz constant of $f_t$ is $$\|f_t \|_{Lip} = \| F^{f}_t \|_2 \leq \| A (I_n + Q C^T R^{-1} C)^{-1} \|_2 = L_f \enspace.$$ As $\| (I_n + Q C^T R^{-1} C)^{-1} \|_2 < 1$, if $\| A\|_2 < 1$, then $\| f_t \|_{Lip} \leq L_f < 1$ for all $t$. \[lemma:LGSSM\_backward\_maps\] The backward random maps of an LGSSM are Gaussian linear maps. Specifically, $b_t(x_t) = F^b_t x_t + \zeta^b_t$, where $\zeta^b_t$ is a Gaussian random intercept and $F^b_t$ is a matrix function of $\theta$ and $y_{<t}$. As a linear map, the Lipschitz constant of $b_t$ is $$\|b_t \|_{Lip} = \| F^b_t \|_2 \leq \|A (Q A^T Q^{-1} A + Q C^T R^{-1} C)^{-1} \|_2 = L_b \enspace.$$ If $\| A\|_2 < \| (Q A^T Q^{-1} A + Q C^T R^{-1} C)^{-1} \|_2$, then $\| f_t \|_{Lip} \leq L_f < 1$ for all $t$. In addition, when the variance of the prior $p_0(x)$ is less than the steady state variance $V_\infty = Q + A V_\infty A^T$ and $A$ commutes with $Q$, we obtain a tighter bound $$\|b_t \|_{Lip} = \| F^b_t \|_2 \leq \|A (I_n + Q C^T R^{-1} C)^{-1} \|_2 = L_b \enspace.$$ In this case, if $\| A\|_2 < 1$, then $\| b_t \|_{Lip} \leq L_b < 1$ for all $t$. Lemmas \[lemma:LGSSM\_forward\_maps\] and \[lemma:LGSSM\_backward\_maps\] agree with intuition, when $\|A\|_2 \approx 0$ (no connection between $x_{t-1}$ and $x_t$) or $\| Q \|_2 \gg \| R \|_2$ (transition noise is much larger than emission noise), then $L_f, L_b \approx 0$ (observations can be treated independently). Conversely, when $\|A\|_2 \approx 1$ and $\| Q \|_2 \ll \| R\|_2$, then $L_f, L_b \approx 1$ and buffering is necessary. \[lemma:LGSSM\_Lipschitz\] As $x,y$ are jointly Gaussian in the LGSSM, the gradient of the complete data loglikelihood is a quadratic form in $xx^T$ with matrices $$\begin{aligned} \Omega = \{& I_n \otimes Q^{-1}, I_n \otimes Q^{-1} A, Q^{-1/2} \otimes I_n, Q^{-1/2} A \otimes I_n, Q^{-1/2} \otimes A, Q^{-1/2}A \otimes A, \\ \nonumber & I_n \otimes R^{-1}, I_n \otimes R^{-1} C, R^{-1/2}C \otimes C \} \enspace,\end{aligned}$$ where $Q^{-1/2} = \psi_Q$ and $R^{-1/2} = \psi_R$. Therefore a bound for the Lipschitz constant is $L_U' = \max_{\omega \in \Omega} \|\omega\|_2$. This bound grows in $\|A\|, \|C\|, \|Q\|^{-1}, \|R\|^{-1}$. The proofs can be found in the Supplement. Switching Linear Dynamical System (SLDS) {#sec:models:SLDS} ---------------------------------------- Switching linear dynamical systems (SLDSs) are an example of a state space model with both discrete and continuous latent variables. The form of SLDS models that we consider is $$\label{eq:slds_complete_likelihood} p(y, x, z \, | \, \theta) = \prod_{t=1}^T \Pi_{z_{t-1}, z_{t}} \cdot \mathcal{N}(x_t \, | \, A_{z_t} x_{t-1}, Q_{z_t}) \cdot \mathcal{N}(y_t \, | \, C x_t, R) \enspace,$$ where $y_t \in {\mathbb{R}}^m$ are the observations, ${u}_t \equiv (x_t, z_t) \in {\mathbb{R}}^n \times \{1, \ldots, K\}$ are the mixed-type latent state sequence, and $\theta = \{\Pi, A, Q, C, R\}$ the model parameters with $\Pi_k \in \Delta^K$, $A_k \in {\mathbb{R}}^{n \times n}$, $Q_k \in \mathbb{S}^{n}_+$ for $k = 1,\ldots, K$, $C \in {\mathbb{R}}^{m \times n}$ and $R \in \mathbb{S}^m_+$. The SLDS of Eq.  can be viewed either as a latent AR(1)-HMM with conditional Gaussian emissions or as hidden Markov switches of a LGSSM. As an extension of the ARHMM, the latent continuous state sequence $x_t$ can *smooth* noisy observations. As an extension of the LGSSM, the latent discrete state sequence $z_t$ allows modeling of more complex dynamics by *switching* between different states (or regimes). ![Graphical Model of a SLDS.[]{data-label="fig:slds_graphical_model"}](figures/latex_fig/slds.pdf){width="0.99\linewidth"} ### Gradient Estimators {#sec:models:SLDS:gradients} Unlike previous models, the forward-backward algorithm for the latent variables $(x, z)$ in an SLDS does not have a closed form. Specifically, the transition kernel for $x$ is a Gaussian mixture, so the forward and backward messages of $x$ are Gaussian mixtures with an exponentially increasing number of components (e.g. $\alpha_t$ has $K^t$ components). Because the forward-backward algorithm is intractable for SLDSs, we rely on sampling $(x,z)$ and forming a Monte Carlo estimate of the expectation in Fisher’s identity Eq. . We consider various options of this Monte Carlo estimate below. To sample $(x,z)$, we use a blocked Gibbs scheme as in [@fox2011bayesian], detailed in the Supplement. Given a collection of $N$ samples from blocked Gibbs $\{x^{(r)}, z^{(r)}\} \sim x, z \,|\, y, \theta$, we construct three different estimators for the marginal loglikelihood. The first estimator, replaces the expectation in Eq.  with a Monte Carlo average $$\label{eq:slds_naive} {\mathbb{E} \,}_{x, z | y, \theta}[{\nabla}\log p(y, x, z \, | \, \theta)] \approx \frac{1}{N}\sum_{r = 1}^{N} {\nabla}\log p(y, x^{(r)}, z^{(r)} \, | \, \theta) \enspace.$$ We construct two additional estimators by analytically integrating out either one of the two latent variables. These estimators are the *Rao-Blackwellization* of the naive Monte Carlo estimate [@cappe2005inference]. Integrating out either $x$ or $z$, gives us $$\begin{aligned} \label{eq:slds_z_marginal} {\mathbb{E} \,}_{x, z | y, \theta}[{\nabla}\log p(y, x, z \, | \, \theta)] &= \frac{1}{N}\sum_{r = 1}^{N} {\mathbb{E} \,}_{x | y, z^{(r)}, \theta}[{\nabla}\log p(y, x, z^{(r)} \, | \, \theta)] \\ \label{eq:slds_x_marginal} {\mathbb{E} \,}_{x, z | y, \theta}[{\nabla}\log p(y, x, z \, | \, \theta)] &= \frac{1}{N}\sum_{r = 1}^{N} {\mathbb{E} \,}_{z | y, x^{(r)}, \theta}[{\nabla}\log p(y, x^{(r)}, z \, | \, \theta)] \enspace.\end{aligned}$$ Because Eq.  integrates out $x$, it has lower variance for the gradient terms involving $x$ (i.e. $A$, $Q$ $R$). Similarly, because Eq.  integrates out $z$, it has lower variance for the gradient terms involving $z$ (i.e. $\Pi$). Selecting one of the above Monte Carlo estimates of ${\nabla}U(\theta)$, we can deploy the same buffered subsampling estimator Eq. , obtaining Algorithm \[alg:noisygradient-slds\]. Algorithm \[alg:noisygradient-slds\] replaces the forward-backward subroutine in Algorithm \[alg:noisygradient-analytic\] with blocked Gibbs sampling over ${{\mathcal{S}^*}}$. Although this is more computationally costly than the exact forward-backward algorithms of the previous sections, it still provides memory saving and runtime speed ups compared to running a full blocked Gibbs sampler over ${\mathcal{T}}$. The explicit forms of Eqs. -, precondition matrix $D(\theta)$, and correction term $\Gamma(\theta)$ for SLDS used in Alg. \[alg:sgrld\] are a combination of those for ARHMMs and LGSSMs. Complete details are provided in the Supplement. ### Error Bounds There are two primary challenges for the error analysis of the SLDS: (i) the forward and backward smoothing kernels for the SLDS are mixtures and (ii) the error from the finite-step blocked Gibbs sampler needs to be quantified. Conditions for contraction in the forward and backward smoothing random maps of switching models may follow from the conditions in [@cloez2015exponential]. Combining the convergence rate of the blocked Gibbs sampler with the error bound is an area we leave for future work. Our experiments in Section \[sec:experiments-SLDS\] provide empirical evidence of the potential benefits of the algorithm. Experiments {#sec:experiments} =========== We evaluate the performance of our proposed SGRLD algorithm (Section \[sec:framework\]) using both synthetic and real data. We organize our experiments by the corresponding models of Section \[sec:models\]. Our evaluation focuses on the following three topics: (1) the computational speed-up of SGMCMC over batch MCMC, (2) the effectiveness of buffering in correcting bias, and (3) the effectiveness of the complete-data Fisher information preconditioning of SGRLD over SGLD. For batch MCMC, we consider block-Gibbs sampling (Gibbs) and unadjusted Langevin Monte-Carlo – both with preconditioning (RLD) and without precondition (LD). Note that LD and RLD are SGLD and SGRLD with $S = T$. To assess the performance of our samplers, we measure the marginal loglikelihood of samples $\theta^{(s)}$ at different runtimes on a heldout test sequence. In synthetic data, where the true parameter $\theta^*$ is known, we also measure the mean-squared error (MSE) of the sample average $\hat\theta^{(s)} = \sum_{i \leq s} \theta^{(i)}/s$ to $\theta^*$. To assess the quality of our MCMC samples at approximating the posterior $\Pr(\theta \, | \, y)$, we measure the kernel Stein discrepancy (KSD) of each chain after burn-in given equal computation time [@gorham2017measuring; @liu2016kernelized], rather than effective sample size (ESS) [@brooks2011handbook; @gelman2013bayesian], as KSD accounts for bias in the samples. As with all gradient-based methods, our SGMCMC methods require a hyper-parameter search over the fixed step-size tuning parameter $h$. We present results for the best step-size as assessed via heldout loglikelihood on a validation set. As the potential $U(\theta)$ for SSMs is non-convex, initialization is important. For the HMM and ARHMM, we initialize the parameters $\Pi, A, Q$ using $z$ given from $K$-means clustering of the observations $y$ (or $y_{t-p:t}$). For the LGSSM, we initialize the parameters from the prior. For the mixed-type SLDS, we first sample $R$ from the prior and initialize $\Pi, A, Q$ using $z$ from $K$-means. Finally, in our experiments, we use flat and non-informative priors for $\theta$. For complete details see the Supplement. Gaussian HMM & ARHMM {#sec:experiments-HMM} -------------------- ### Synthetic ARHMM {#sec:experiments-ARHMM-synth} We first consider synthetic data generated from a $2$-state ARHMM in two dimensions $m=2$. The true model parameters $\theta^*$ are $$\Pi = \begin{bmatrix} 0.1 & 0.9 \\ 0.9 & 0.1 \end{bmatrix} \enspace, \enspace Q_1 = Q_2 = 0.1 \cdot \begin{bmatrix} 1 & 0 \\ 0 & 1 \end{bmatrix}$$ $$A_1 = 0.9 \cdot \begin{bmatrix} \cos(-\vartheta) & -\sin(-\vartheta) \\ \sin(-\vartheta) & \cos(-\vartheta) \end{bmatrix} \enspace, \enspace A_2 = 0.9 \cdot \begin{bmatrix} \cos(\vartheta) & -\sin(\vartheta) \\ \sin(\vartheta) & \cos(\vartheta) \end{bmatrix} \enspace .$$ The model’s two states are alternating rotations of $y \in {\mathbb{R}}^2$ with angle $\vartheta = \pi/4$ and the latent state sequence has a high transition rate $\Pr(z_t \neq z_{t-1}) = 0.9$. From this model we generate time series of length $T = 10^4$ and $10^6$. ![Stochastic gradient error ${\mathbb{E} \,}_{{\mathcal{S}}}\|\bar{g}(\theta) - \tilde{g}(\theta)\|_2$. (Left) varying subsequence length $S$ for no-buffer $B = 0$ and buffer $B=10$. (Right) varying buffer size $B$ for $S = 4$ and $S=64$ subsequence lengths. Error bars are SD over $100$ datasets. []{data-label="fig:arhmm_grad_error"}](figures/19-06-21/arphmm_fulltrans_1000/group_varyS.pdf){width="\textwidth"} ![Stochastic gradient error ${\mathbb{E} \,}_{{\mathcal{S}}}\|\bar{g}(\theta) - \tilde{g}(\theta)\|_2$. (Left) varying subsequence length $S$ for no-buffer $B = 0$ and buffer $B=10$. (Right) varying buffer size $B$ for $S = 4$ and $S=64$ subsequence lengths. Error bars are SD over $100$ datasets. []{data-label="fig:arhmm_grad_error"}](figures/19-06-21/arphmm_fulltrans_1000/group_varyB.pdf){width="\textwidth"} Figure \[fig:arhmm\_grad\_error\] are plots of the stochastic gradient error ${\mathbb{E} \,}_{{\mathcal{S}}}\|\bar{g}(\theta) - \tilde{g}(\theta)\|_2$ between the unbiased and buffered estimates evaluated at the true model parameters $\theta=\theta^*$. From Figure \[fig:arhmm\_grad\_error\] (left), we see that the error decays $O(1/S)$ and that the error in estimates without buffering $B=0$ (orange) are orders of magnitude larger than the estimates with moderate buffering $B=10$ (blue). From Figure \[fig:arhmm\_grad\_error\] (right), we see that the error decays geometrically in buffer size $O(L^B)$. ![Metrics vs Runtime on ARHMM data with $T = 10^4$ (top), $T=10^6$ (bottom), for different methods: , , and SGMCMC. For SGMCMC methods, solid () and dashed () lines indicate SGRLD and SGLD respectively. The different metrics are: (left) heldout loglikelihoood and (right) transition matrix estimation error $MSE(\hat\Pi^{(s)}, \Pi^*)$. []{data-label="fig:arhmm_synth_metrics"}](./figures/19-03-11/ARPHMM_10k/gibbs_heldloglike_vs_logruntime.pdf){width="\textwidth"} ![Metrics vs Runtime on ARHMM data with $T = 10^4$ (top), $T=10^6$ (bottom), for different methods: , , and SGMCMC. For SGMCMC methods, solid () and dashed () lines indicate SGRLD and SGLD respectively. The different metrics are: (left) heldout loglikelihoood and (right) transition matrix estimation error $MSE(\hat\Pi^{(s)}, \Pi^*)$. []{data-label="fig:arhmm_synth_metrics"}](./figures/19-03-11/ARPHMM_10k/gibbs_logmse_pi_vs_logruntime.pdf){width="\textwidth"} ![Metrics vs Runtime on ARHMM data with $T = 10^4$ (top), $T=10^6$ (bottom), for different methods: , , and SGMCMC. For SGMCMC methods, solid () and dashed () lines indicate SGRLD and SGLD respectively. The different metrics are: (left) heldout loglikelihoood and (right) transition matrix estimation error $MSE(\hat\Pi^{(s)}, \Pi^*)$. []{data-label="fig:arhmm_synth_metrics"}](./figures/19-03-11/ARPHMM_1m/gibbs_heldloglike_vs_logruntime.pdf){width="\textwidth"} ![Metrics vs Runtime on ARHMM data with $T = 10^4$ (top), $T=10^6$ (bottom), for different methods: , , and SGMCMC. For SGMCMC methods, solid () and dashed () lines indicate SGRLD and SGLD respectively. The different metrics are: (left) heldout loglikelihoood and (right) transition matrix estimation error $MSE(\hat\Pi^{(s)}, \Pi^*)$. []{data-label="fig:arhmm_synth_metrics"}](./figures/19-03-11/ARPHMM_1m/gibbs_logmse_pi_vs_logruntime.pdf){width="\textwidth"} ![Boxplot of MCMC samples for ARHMM data $T = 10^6$. (Top) comparison of all samplers, (bottom) zoom-in for top three. The half of each chain is discarded as burn-in. SGRLD with buffering in 6 hrs is comparable to RLD or Gibbs in 144 hrs.[]{data-label="fig:arhmm_boxplot"}](figures/19-03-11/ARPHMM_1m/boxplot_pi.pdf "fig:"){width="\textwidth"} ![Boxplot of MCMC samples for ARHMM data $T = 10^6$. (Top) comparison of all samplers, (bottom) zoom-in for top three. The half of each chain is discarded as burn-in. SGRLD with buffering in 6 hrs is comparable to RLD or Gibbs in 144 hrs.[]{data-label="fig:arhmm_boxplot"}](figures/19-06-21/arphmm_boxplots/boxplot_pi_zoom.pdf "fig:"){width="\textwidth"} Sampler $\pi$ $A$ $\Sigma$ -- ------------------- ------------- ------------- ------------- SGLD (No Buffer) 3.15 (0.46) 2.47 (0.51) 2.33 (0.30) SGLD (Buffer) 0.99 (0.13) 1.60 (0.20) 1.80 (0.13) LD 1.77 (0.72) 1.86 (0.32) 2.12 (0.36) SGRLD (No Buffer) 3.15 (0.39) 2.02 (0.24) 1.91 (0.24) SGRLD (Buffer) 0.89 (0.04) 1.53 (0.10) 1.60 (0.30) RLD 0.67 (0.27) 2.02 (0.14) 1.60 (0.18) Gibbs 0.36 (0.07) 1.30 (0.20) 0.61 (0.13) SGLD (No Buffer) 4.73 (0.07) 4.07 (0.22) 3.67 (0.25) SGLD (Buffer) 2.62 (0.06) 3.30 (0.20) 2.77 (0.31) LD 3.59 (0.22) 4.73 (0.33) 4.78 (0.34) SGRLD (No Buffer) 4.75 (0.15) 4.02 (0.06) 3.61 (0.12) SGRLD (Buffer) 2.27 (0.08) 3.38 (0.08) 2.89 (0.09) RLD 3.31 (0.05) 4.22 (0.12) 3.56 (0.07) Gibbs 3.17 (0.30) 4.18 (0.07) 3.30 (0.07) : $\log_{10}$(KSD) by variable of ARHMM samplers at 6 hrs. Mean and (SD) over runs in Figure \[fig:arhmm\_synth\_metrics\].[]{data-label="table:arhmm_ksd"} In Figures \[fig:arhmm\_synth\_metrics\] and \[fig:arhmm\_boxplot\], we compare subsequence-based MCMC methods: SGLD (no-buffer and buffer) and SGRLD (no-buffer and buffer), with full-sequence MCMC methods: LD, RLD, and Gibbs. We fit our samplers on one training sequence and evaluate performance on one test sequence. We consider two training sequences of lengths $T = 10^4$ and $T=10^6$ and evaluate on the same test sequence of length $T = 10^4$. For the SGMCMC methods we use a subsequence size of $S = 2$ and a buffer size of $B=0$ (no-buffer) or $B=2$ (buffer). We ran the subsequence methods for 6 hours and full-sequence methods for 144 hours. From Figure \[fig:arhmm\_synth\_metrics\], we see that our buffered SGMCMC (blue) helps convergence and mixing orders of magnitude faster than the full-sequence gradient MCMC (green). We also see that buffering is necessary to properly estimate $\Pi$ as the no-buffer SGMCMC methods (orange) do not properly learn $\Pi$. We also see that preconditioning helps convergence and mixing as SGRLD (solid) outperforms SGLD (dashed). Although Gibbs outperforms SGMCMC for $T=10^4$, Gibbs performs worse for $T=10^6$, as each iteration requires a full pass over the data set. Figure \[fig:arhmm\_boxplot\] are boxplots comparing the marginal distribution for the different methods on the synthetic ARHMM data $T = 10^6$. From Figure \[fig:arhmm\_boxplot\], we see that SGRLD with buffering in 6 hours is comparable to RLD or Gibbs in 144 hours; however, SGRLD without buffering is biased and RLD in 6 hours has not had enough time to mix. Table \[table:arhmm\_ksd\] displays the KSD of the samples to the posterior after discarding half the samples as burn-in. The standard deviation is over MCMC chains with different initializations. Although RLD and Gibbs perform well for $T = 10^4$, both perform worse for larger $T = 10^6$ due to the increased time between samples. We also see that the non-buffered methods do poorly for all $T$ due to sampling from the incorrect distribution. Although SGLD (buffer) and SGRLD (buffer) perform comparably after burn-in, Figure \[fig:arhmm\_synth\_metrics\] suggests SGRLD converges more rapidly. In the Supplement, we present a synthetic data experiment for the Gaussian HMM, and find similar results. ### Ion Channel Recordings {#sec:experiments-HMM-ion} We investigate the behavior of SGMCMC samplers on ion channel recording data. In particular, we consider a 1MHz recording of a single alamethicin channel [@Rosenstein:2013]. This data was previously investigated using a Bayesian nonparametric HMM in [@Palla:2014] and [@Tripuraneni:2015]. In that work, the authors downsample the data by a factor of $100$ and only used $10,000$ and $2,000$ observations due to the challenge of scaling computations to the full sequence. We present the results on the data without downsampling ($10$ million observations), where Gibbs sampling runs into memory issues. Figure \[fig:ion\_channel\] presents our results, after applying a log-transform and normalizing the observations. We train on the first 90% and evaluate on the last 10%. For our SGMCMC methods we use a subsequence size of $S = 10$ and a buffer size of $B=0$ (no-buffer) or $B=10$ (buffer). In addition to heldout loglikelihood, we also evaluate on 10-step ahead predictive loglikelihood $\sum_{t} \log \Pr(y_{t+10} \, | \, \theta, y_{\leq t})$, which is more sensitive to $\Pi$. We see that SGRLD quickly converges compared to SGLD. Although the buffered methods take longer to compute ($S+2B = 30$ vs $S=10$), we see that buffering is necessary to perform well. In the Supplement, we present results comparing SGMCMC methods with Gibbs sampling on a downsampled version. ![Ion Channel Recordings: (Left) heldout loglikelihood vs runtime. (Center) 10-step predictive loglikelihood $\sum_{t} \log \Pr(y_{t+10} \, | \, \theta, y_{\leq t})$ vs runtime. (Right) segmentation by SGRLD (Buffer). []{data-label="fig:ion_channel"}](figures/18-09-11/ion_full_heldout_loglike.pdf){width="\textwidth"} ![Ion Channel Recordings: (Left) heldout loglikelihood vs runtime. (Center) 10-step predictive loglikelihood $\sum_{t} \log \Pr(y_{t+10} \, | \, \theta, y_{\leq t})$ vs runtime. (Right) segmentation by SGRLD (Buffer). []{data-label="fig:ion_channel"}](figures/18-09-11/ion_full_predictive_loglike.pdf){width="\textwidth"} ![Ion Channel Recordings: (Left) heldout loglikelihood vs runtime. (Center) 10-step predictive loglikelihood $\sum_{t} \log \Pr(y_{t+10} \, | \, \theta, y_{\leq t})$ vs runtime. (Right) segmentation by SGRLD (Buffer). []{data-label="fig:ion_channel"}](figures/sgrld_gausshmm_ion-channel_fit.png){width="\textwidth"} ### Canine Seizure iEEG {#sec:experiments-ARHMM-seizure} ![ARHMM for Canine Seizure Data: (left) heldout loglikelihood vs time, (center) heldout loglikelihood vs time on log-scale (right) example segmentation of a test seizure channel by SLDS fit with SGRLD. The MCMC methods compared are , , , and . []{data-label="fig:canine"}](./figures/arhmm_seizure_HeldoutLoglike.pdf){width="\textwidth"} ![ARHMM for Canine Seizure Data: (left) heldout loglikelihood vs time, (center) heldout loglikelihood vs time on log-scale (right) example segmentation of a test seizure channel by SLDS fit with SGRLD. The MCMC methods compared are , , , and . []{data-label="fig:canine"}](./figures/arhmm_seizure_HeldoutLoglike_logscale.pdf){width="\textwidth"} ![ARHMM for Canine Seizure Data: (left) heldout loglikelihood vs time, (center) heldout loglikelihood vs time on log-scale (right) example segmentation of a test seizure channel by SLDS fit with SGRLD. The MCMC methods compared are , , , and . []{data-label="fig:canine"}](figures/sgrld_arhmm_seizure_fit.png){width="\textwidth"} We now consider applying SGMCMC samplers to intracranial EEG (iEEG) data. In particular, we consider data from a study on canines with epilepsy available at `ieeg.org` [@davis2016mining]. We focus on one canine, which over the course of 45.1 days was continuously monitored at 200Hz over 16 channels and recorded 90 seizures. This data was analyzed in prior work that compared a baseline ARHMM to nonparametric extensions using Gibbs sampling [@wulsin2013bayesian]. Following [@wulsin2013bayesian], we process the data into 4 minute windows around each seizure to focus on the seizure dynamics resulting in 90 time series of 48,000 points in ${\mathbb{R}}^{16}$. We use an ARHMM with $K=5$ latent states and $p = 5$ lags treating each channel independently. We perform an 80-20 train-test split over 90 seizures, running inference on the training set and evaluating log-likelihood on the heldout test set. We compare SGLD and SGRLD samplers with $S = 100$ and $B=10$ with the baseline Gibbs sampler on the full data set. Because of the large data size, we also consider a *subset* Gibbs sampler that only uses $10\%$ of the training set seizures. In Figure \[fig:canine\], we see that SGRLD converges much more rapidly than the other methods. As each iteration of the Gibbs sampler takes  6 hours, it takes a couple weeks for the Gibbs sampler to converge to the solution SGRLD converges to in a few hours. Although the subset Gibbs sampler is 10x faster than Gibbs, it does not converge to the full data posterior and its generalization error to the heldout test set is poorer than the other methods. From this experiment we see that SGMCMC methods provide order of magnitude improvements (compared to subsetting the data). LGSSM and SLDS {#sec:experiments-SLDS} -------------- We first validate the LGSSM (SLDS with $K=1$) on synthetic data. We then consider the SLDS sampler on a synthetic dataset and two real datasets: the seizure data of Section \[sec:experiments-ARHMM-seizure\] and a weather dataset. ### Synthetic LGSSM {#sec:experiments-LGSSM-synth} We consider synthetic data from a LGSSM with observations and latent state dimension $m = n = 2$. In particular, we consider, a rotating state sequence with noisy observations. The true model parameter $\theta^*$ are $$A = 0.7 \cdot \begin{bmatrix} \cos(\vartheta) & -\sin(\vartheta) \\ \sin(\vartheta) & \cos(\vartheta) \end{bmatrix} \enspace, \enspace Q = 0.1 \cdot \begin{bmatrix} 1 & 0 \\ 0 & 1 \end{bmatrix} \enspace, \enspace C = \begin{bmatrix} 1 & 0 \\ 0 & 1 \end{bmatrix} \enspace, \enspace R =\begin{bmatrix} 1 & 0 \\ 0 & 1 \end{bmatrix} \enspace,$$ where $\vartheta=\pi/4$. Because the transition error $Q$ is smaller than the emission error $R$, inclusion of previous and future observations is necessary to accurately infer the continuous latent state $x_t$. ![Stochastic gradient error ${\mathbb{E} \,}_{{\mathcal{S}}}\|\bar{g}(\theta) - \tilde{g}(\theta)\|_2$. (Left) varying subsequence length $S$ for no-buffer $B = 0$ and buffer $B=10$. (Right) varying buffer size $B$ for $S = 4$ and $S=64$ subsequence lengths. Error bars are SD over $100$ datasets.[]{data-label="fig:lds_grad_error"}](figures/19-06-21/lgssm_1000/group_varyS.pdf){width="\textwidth"} ![Stochastic gradient error ${\mathbb{E} \,}_{{\mathcal{S}}}\|\bar{g}(\theta) - \tilde{g}(\theta)\|_2$. (Left) varying subsequence length $S$ for no-buffer $B = 0$ and buffer $B=10$. (Right) varying buffer size $B$ for $S = 4$ and $S=64$ subsequence lengths. Error bars are SD over $100$ datasets.[]{data-label="fig:lds_grad_error"}](figures/19-06-21/lgssm_1000/group_varyB.pdf){width="\textwidth"} Figure \[fig:lds\_grad\_error\] are plots of the stochastic gradient error ${\mathbb{E} \,}_{{\mathcal{S}}}\|\bar{g}(\theta) - \tilde{g}(\theta)\|_2$ between the unbiased and buffered estimates evaluated at the true model parameters $\theta=\theta^*$. Similar to the ARPHMM, we see that the error decays $O(1/S)$ and that moderate buffering (e.g. $B=10$) deceases the error by orders of magnitude in Figure \[fig:lds\_grad\_error\] (left). And we see that the error decays geometrically in buffer size $O(L^B)$ in Figure \[fig:lds\_grad\_error\] (right). ![Metrics vs Runtime on LGSSM with $T = 10^4$ (top), $T = 10^6$ (bottom) for different methods: , , and SGMCMC. For SGMCMC methods, solid () and dashed () lines indicate SGRLD and SGLD respectively. The different metrics are: (left) heldout loglikelihoood and (right) transition matrix estimation error $MSE(\hat{A}^{(s)}, A^*)$. []{data-label="fig:lds_synth"}](./figures/19-03-11/LGSSM_10k/gibbs_heldloglike_vs_logruntime.pdf){width="\textwidth"} ![Metrics vs Runtime on LGSSM with $T = 10^4$ (top), $T = 10^6$ (bottom) for different methods: , , and SGMCMC. For SGMCMC methods, solid () and dashed () lines indicate SGRLD and SGLD respectively. The different metrics are: (left) heldout loglikelihoood and (right) transition matrix estimation error $MSE(\hat{A}^{(s)}, A^*)$. []{data-label="fig:lds_synth"}](./figures/19-03-11/LGSSM_10k/gibbs_logmse_A_vs_logruntime.pdf){width="\textwidth"} ![Metrics vs Runtime on LGSSM with $T = 10^4$ (top), $T = 10^6$ (bottom) for different methods: , , and SGMCMC. For SGMCMC methods, solid () and dashed () lines indicate SGRLD and SGLD respectively. The different metrics are: (left) heldout loglikelihoood and (right) transition matrix estimation error $MSE(\hat{A}^{(s)}, A^*)$. []{data-label="fig:lds_synth"}](./figures/19-03-11/LGSSM_1m/gibbs_heldloglike_vs_logruntime.pdf){width="\textwidth"} ![Metrics vs Runtime on LGSSM with $T = 10^4$ (top), $T = 10^6$ (bottom) for different methods: , , and SGMCMC. For SGMCMC methods, solid () and dashed () lines indicate SGRLD and SGLD respectively. The different metrics are: (left) heldout loglikelihoood and (right) transition matrix estimation error $MSE(\hat{A}^{(s)}, A^*)$. []{data-label="fig:lds_synth"}](./figures/19-03-11/LGSSM_1m/gibbs_logmse_A_vs_logruntime.pdf){width="\textwidth"} ![Boxplot of MCMC samples of transition matrix $A$ for LGSSM data $T = 10^4$. SGRLD with buffering in 6 hours is comparable to RLD or Gibbs in 144 hours. SGRLD without buffering is biased and RLD in 6 hours has not fully mixed.[]{data-label="fig:lds_boxplot"}](figures/19-03-11/LGSSM_10k/boxplot_A.pdf){width="\textwidth"} In Figures \[fig:lds\_synth\] and \[fig:lds\_boxplot\], we compare SGLD (no-buffer and buffer), SGRLD (no-buffer and buffer), LD, RLD, and a blocked Gibb sampler. We fit our samplers on one training sequence and evaluate performance on one test sequence. We consider two training sequences of lengths $T = 10^4$ and $T=10^6$ and evaluate on the same test sequence of length $T = 10^4$. For the SGMCMC methods, we use a subsequence size of $S = 20$ with $B=0$ (no buffer) and $B=10$ (buffer). We see that even with a large subsequence size, buffering is crucial for accurate inference as SGMCMC methods without buffering converge to a different stationary distribution than the posterior. In Table \[tab:lds\_ksd\], we evaluate the KSD of the different MCMC methods. We see that SGMCMC with buffering slightly outperforms the full sequence methods for $T = 10^4$ and significantly outperforms the full sequence methods for $T = 10^6$, while SGMCMC without buffering performs poorly due to bias. Sampler $A$ $Q$ $R$ -- ------------------- ------------- ------------- ------------- SGLD (No Buffer) 2.39 (0.01) 1.73 (0.03) 1.48 (0.03) SGLD (Buffer) 0.88 (0.11) 0.41 (0.11) 0.86 (0.08) LD 0.99 (0.13) 1.12 (0.19) 1.10 (0.17) SGRLD (No Buffer) 2.38 (0.01) 1.70 (0.02) 1.43 (0.02) SGRLD (Buffer) 0.85 (0.08) 0.18 (0.12) 0.77 (0.14) RLD 0.99 (0.12) 0.90 (0.19) 1.10 (0.17) Gibbs 0.74 (0.20) 0.33 (0.18) 1.06 (0.27) SGLD (No Buffer) 4.32 (0.01) 3.79 (0.02) 3.50 (0.02) SGLD (Buffer) 2.30 (0.19) 1.61 (0.18) 2.84 (0.03) LD 4.26 (0.35) 4.00 (0.39) 4.14 (0.19) SGRLD (No Buffer) 4.27 (0.01) 3.77 (0.02) 3.23 (0.03) SGRLD (Buffer) 2.17 (0.33) 1.64 (0.21) 3.03 (0.12) RLD 4.34 (0.23) 3.76 (0.25) 4.03 (0.23) Gibbs 3.46 (0.28) 3.52 (0.14) 3.50 (0.28) : $\log_{10}$(KSD) by variable of LGSSM samplers at 6 hrs. Mean and (SD) over runs in Figure \[fig:lds\_synth\].[]{data-label="tab:lds_ksd"} ### Synthetic SLDS We now consider synthetic data from a model we can view as switching extension of the LGSSM in Section \[sec:experiments-LGSSM-synth\] or as a noisy version of the ARHMM in the Supplement. The true model parameters $\theta^*$ are $$\Pi = \begin{bmatrix} 0.9 & 0.1 \\ 0.1 & 0.9 \end{bmatrix} \enspace, \enspace Q_1 = Q_2 = 0.1 \cdot \begin{bmatrix} 1 & 0 \\ 0 & 1 \end{bmatrix} \enspace, \enspace C = \begin{bmatrix} 1 & 0 \\ 0 & 1 \end{bmatrix} \enspace, \enspace R = 0.1 \cdot \begin{bmatrix} 1 & 0 \\ 0 & 1 \end{bmatrix} \enspace,$$ $$A_1 = 0.9 \cdot \begin{bmatrix} \cos(-\vartheta) & -\sin(-\vartheta) \\ \sin(-\vartheta) & \cos(-\vartheta) \end{bmatrix} \enspace, \enspace A_2 = 0.9 \cdot \begin{bmatrix} \cos(\vartheta) & -\sin(\vartheta) \\ \sin(\vartheta) & \cos(\vartheta) \end{bmatrix}\enspace,$$ where again $\vartheta=\pi/4$. We generate sequences of length $T = 10^4$ and $10^6$. ![SLDS gradient error for the different estimators Eqs. -. (Left) Boxplots of $\tilde{g}(\theta)_A - g(\theta)_A$. (Right) Boxplots of $\tilde{g}(\theta)_\Pi - g(\theta)_\Pi$. []{data-label="fig:slds_gradients"}](./figures/18-08-29/slds_lowtrans_true_grad_A.pdf){width="\textwidth"} ![SLDS gradient error for the different estimators Eqs. -. (Left) Boxplots of $\tilde{g}(\theta)_A - g(\theta)_A$. (Right) Boxplots of $\tilde{g}(\theta)_\Pi - g(\theta)_\Pi$. []{data-label="fig:slds_gradients"}](./figures/18-08-29/slds_lowtrans_true_grad_pi.pdf){width="\textwidth"} ![Stochastic gradient error ${\mathbb{E} \,}_{{\mathcal{S}}}\|\bar{g}(\theta) - \tilde{g}(\theta)\|_2$ for `z Gradient`. (Left) error varying subsequence length $S$ for no-buffer $B=0$ and buffer $B=4$. (Right) error varying buffer size $B$ for small $S = 2$ and long $S=32$ subsequences. Error bars are SD over $100$ datasets. []{data-label="fig:slds_stochastic_gradients"}](figures/19-06-21/slds_lowtrans_256/group_varyS.pdf){width="\textwidth"} ![Stochastic gradient error ${\mathbb{E} \,}_{{\mathcal{S}}}\|\bar{g}(\theta) - \tilde{g}(\theta)\|_2$ for `z Gradient`. (Left) error varying subsequence length $S$ for no-buffer $B=0$ and buffer $B=4$. (Right) error varying buffer size $B$ for small $S = 2$ and long $S=32$ subsequences. Error bars are SD over $100$ datasets. []{data-label="fig:slds_stochastic_gradients"}](figures/19-06-21/slds_lowtrans_256/group_varyB.pdf){width="\textwidth"} ![Metrics vs Runtime on SLDS data for different inference methods: , , , and . (Top) $T = 10^4$ (Bottom) $T = 10^6$. The metrics are: (left) heldout loglikelihood, (center) estimation error $MSE(\hat{A}^{(s)}, A^*)$, (right) estimation error $MSE(\hat\Pi^{(s)}, \Pi^*)$. []{data-label="fig:slds_lowtrans_metrics"}](./figures/18-09-27/slds_lowtrans_avg_loglike.pdf){width="\textwidth"} ![Metrics vs Runtime on SLDS data for different inference methods: , , , and . (Top) $T = 10^4$ (Bottom) $T = 10^6$. The metrics are: (left) heldout loglikelihood, (center) estimation error $MSE(\hat{A}^{(s)}, A^*)$, (right) estimation error $MSE(\hat\Pi^{(s)}, \Pi^*)$. []{data-label="fig:slds_lowtrans_metrics"}](./figures/18-09-27/slds_lowtrans_avg_mse_A.pdf){width="\textwidth"} ![Metrics vs Runtime on SLDS data for different inference methods: , , , and . (Top) $T = 10^4$ (Bottom) $T = 10^6$. The metrics are: (left) heldout loglikelihood, (center) estimation error $MSE(\hat{A}^{(s)}, A^*)$, (right) estimation error $MSE(\hat\Pi^{(s)}, \Pi^*)$. []{data-label="fig:slds_lowtrans_metrics"}](./figures/18-09-27/slds_lowtrans_avg_mse_pi.pdf){width="\textwidth"} ![Metrics vs Runtime on SLDS data for different inference methods: , , , and . (Top) $T = 10^4$ (Bottom) $T = 10^6$. The metrics are: (left) heldout loglikelihood, (center) estimation error $MSE(\hat{A}^{(s)}, A^*)$, (right) estimation error $MSE(\hat\Pi^{(s)}, \Pi^*)$. []{data-label="fig:slds_lowtrans_metrics"}](./figures/18-09-27/slds_lowtrans_1m_avg_loglike.pdf){width="\textwidth"} ![Metrics vs Runtime on SLDS data for different inference methods: , , , and . (Top) $T = 10^4$ (Bottom) $T = 10^6$. The metrics are: (left) heldout loglikelihood, (center) estimation error $MSE(\hat{A}^{(s)}, A^*)$, (right) estimation error $MSE(\hat\Pi^{(s)}, \Pi^*)$. []{data-label="fig:slds_lowtrans_metrics"}](./figures/18-09-27/slds_lowtrans_1m_avg_mse_A.pdf){width="\textwidth"} ![Metrics vs Runtime on SLDS data for different inference methods: , , , and . (Top) $T = 10^4$ (Bottom) $T = 10^6$. The metrics are: (left) heldout loglikelihood, (center) estimation error $MSE(\hat{A}^{(s)}, A^*)$, (right) estimation error $MSE(\hat\Pi^{(s)}, \Pi^*)$. []{data-label="fig:slds_lowtrans_metrics"}](./figures/18-09-27/slds_lowtrans_1m_avg_mse_pi.pdf){width="\textwidth"} We first compare the variance of the three difference Monte-Carlo gradient estimators for SLDS: using $(x,z)$ samples (`xz Gradient`) as in Eq. , only using $z$ samples (`z Gradient`) as in Eq. , and only using $x$ samples (`x Gradient`) as in Eq. . Figure \[fig:slds\_gradients\] presents boxplots of $\tilde{g}(\theta) - g(\theta)$ for the three different estimators at $\theta = \theta^*$. From Figure \[fig:slds\_gradients\] (left), we see that `z Gradient` (blue) has much lower variance than the other two estimators for the gradient of $A$. This also holds for the gradients of $Q$ and $R$ (see Supplement). From Figure \[fig:slds\_gradients\] (right), we see that all three estimators have similar variance for the gradient of $\Pi$ (with `x Gradient` (green) slightly better than the other two). This agrees with intuition described in Section \[sec:models:SLDS:gradients\]. Because `z Gradient` has lower variance than the other two estimators, we can use larger step-sizes, leading to faster convergence and mixing. Figure \[fig:slds\_stochastic\_gradients\] are plots of the stochastic gradient error ${\mathbb{E} \,}_{{\mathcal{S}}}\|\bar{g}(\theta) - \tilde{g}(\theta)\|_2$ between the unbiased and buffered estimates (for `z Gradient`) evaluated at the true model parameters $\theta=\theta^*$. For short buffered subsequences (e.g. small $S$ and $B$), the error decays as expected $O(L^B/S)$; however, for longer buffered subsequences the error is dominated by the Monte Carlo error in the number of Gibbs steps used in sampling $z$ for calculating $\tilde{g}$ in Eq. \[eq:slds\_z\_marginal\] . In Figure \[fig:slds\_lowtrans\_metrics\], we compare SGRLD (with buffer) using each of the gradient estimators Eqs. -, and a blocked Gibb sampler. We run our samplers on one training sequence and evaluate performance on another test sequence. For all SGRLD samplers, we used subsequence size of $S = 10$ and $B=10$. As the marginal loglikelihood is not available in closed form for SLDSs, we instead use a Monte Carlo approximation of the EM lower bound $\log \Pr(y \, | \, \theta) \geq {\mathbb{E} \,}_{x, z | y,\theta}[ \log \Pr(y, x, z \, | \, \theta) ]$ where the expectation is approximated with samples of $x,z$ drawn using blocked Gibbs for each fixed $\theta$. From Figure \[fig:slds\_lowtrans\_metrics\], we see that SGRLD methods perform similarly to Gibbs for $T = 10^4$, but vastly outperform Gibbs for $T = 10^6$. ### Canine Seizure iEEG {#canine-seizure-ieeg} Recall the data from Section \[sec:experiments-ARHMM-seizure\]. For our SLDS analysis, we set the continuous latent variable dimension to $n = 1$. The number of latent states remains $K = 5$. We again compare SGLD and SGRLD samplers with $S = 100$ and $B=10$ to Gibbs samplers on both the full data set and a $10\%$ subset of seizures. In Figure \[fig:canine-slds\], we see again that the SGRLD sampler converges much more rapidly than the other methods. In comparison to Figure \[fig:canine\], we also see that the SLDS is a better model for this data than the ARHMM (as measured by heldout likelihood). Qualitatively, the SLDS segmentations of seizures (Figure \[fig:canine-slds\] (right)) is more contiguous than the ARHMM segmentation (Figure \[fig:canine\] (right)). ![SLDS Canine Seizure Data: (left) heldout loglikelihood vs time, (center) heldout loglikelihood vs time on log-scale (right) example segmentation by ARHMM fit with SGRLD. The MCMC methods compared are , , , and . []{data-label="fig:canine-slds"}](./figures/slds_seizure_heldoutloglikelihood.pdf){width="\textwidth"} ![SLDS Canine Seizure Data: (left) heldout loglikelihood vs time, (center) heldout loglikelihood vs time on log-scale (right) example segmentation by ARHMM fit with SGRLD. The MCMC methods compared are , , , and . []{data-label="fig:canine-slds"}](./figures/slds_seizure_heldoutloglikelihood_logscale.pdf){width="\textwidth"} ![SLDS Canine Seizure Data: (left) heldout loglikelihood vs time, (center) heldout loglikelihood vs time on log-scale (right) example segmentation by ARHMM fit with SGRLD. The MCMC methods compared are , , , and . []{data-label="fig:canine-slds"}](figures/sgrld_slds_seizure_fit.png){width="\textwidth"} ### Historical Cities Weather Data We apply SGMCMC to historical city weather data from Kaggle [@kaggleweather]. The data consists of hourly temperature, pressure and humidity measurements $(m=3)$ for 20 US cities over 5 years with $T=44,000$ hourly observations per city. We fit SLDS models with $n=3$ and $K=4$ to both the hourly and daily average observations, treating the cities independently. For both sets of observations, we perform an 80-20 train-test split over 20 cities, running inference on the training set (16 cities) and evaluating loglikelihood on the test set (4 cities). ![SLDS Weather Data. (Top) daily aggregated data, (bottom) hourly data. (Left) heldout loglikelihood vs runtime, (center) Gibbs Houston fit, (right) SGRLD Houston fit.[]{data-label="fig:slds-weather"}](./figures/slds_weather_daily_heldoutloglikelihood.pdf){width="\textwidth"} ![SLDS Weather Data. (Top) daily aggregated data, (bottom) hourly data. (Left) heldout loglikelihood vs runtime, (center) Gibbs Houston fit, (right) SGRLD Houston fit.[]{data-label="fig:slds-weather"}](./figures/slds_weather_daily_gibbs_fit.pdf){width="\textwidth"} ![SLDS Weather Data. (Top) daily aggregated data, (bottom) hourly data. (Left) heldout loglikelihood vs runtime, (center) Gibbs Houston fit, (right) SGRLD Houston fit.[]{data-label="fig:slds-weather"}](./figures/slds_weather_daily_sgrld_fit.pdf){width="\textwidth"} ![SLDS Weather Data. (Top) daily aggregated data, (bottom) hourly data. (Left) heldout loglikelihood vs runtime, (center) Gibbs Houston fit, (right) SGRLD Houston fit.[]{data-label="fig:slds-weather"}](./figures/slds_weather_hourly_heldoutloglikelihood.pdf){width="\textwidth"} ![SLDS Weather Data. (Top) daily aggregated data, (bottom) hourly data. (Left) heldout loglikelihood vs runtime, (center) Gibbs Houston fit, (right) SGRLD Houston fit.[]{data-label="fig:slds-weather"}](./figures/slds_weather_hourly_gibbs_fit.pdf){width="\textwidth"} ![SLDS Weather Data. (Top) daily aggregated data, (bottom) hourly data. (Left) heldout loglikelihood vs runtime, (center) Gibbs Houston fit, (right) SGRLD Houston fit.[]{data-label="fig:slds-weather"}](./figures/slds_weather_hourly_sgrld_fit.pdf){width="\textwidth"} Figure \[fig:slds-weather\] (top-left) shows the heldout loglikelihood vs the runtime for the different samplers on the daily data. From this plot, we see that SGRLD clearly outperforms Gibbs. Although Gibbs converges quickly on the daily data, it gets stuck in local optima. In particular, the Gibbs runs converge to a suboptimal parametrization that mixes over three states, while SGRLD converges to a two state (summer-winter) solution (with the remaining states for sudden shifts or jumps). For example, Figure \[fig:slds-weather\] (top-center and right) are fits of the daily model to the Houston time series for both Gibbs and SGRLD respectively. Figure \[fig:slds-weather\] (bottom-left) shows the heldout loglikelihood vs the runtime of the different samplers for the hourly data. SGRLD again outperforms Gibbs and, for the hourly data, the Gibbs sampler is significantly slower than the SGMCMC samplers. Conclusions {#sec:conclusions} =========== In this work, we developed stochastic gradient MCMC samplers for state space models of sequential data. Our key contribution is a *buffered* gradient estimator $\tilde{g}(\theta)$ for general discrete-time SSMs based on Fisher’s identity. We developed bounds for the error of this buffered gradient estimator and showed that the error decays geometrically in the buffer size under mild conditions. Using this estimator and bound, we developed SGRLD samplers for discrete (Gaussian HMM, ARHMM), continuous (LGSSM), and mixed-type (SLDS) state space models. In our experiments, we find that our methods can provide orders of magnitude run-time speed ups compared to Gibbs sampling, control bias with modest buffer size, and converge and mix more rapidly using preconditioning. In particular, our SGRLD method only uses subsequences at each iteration and is able to take advantage of geometric structure using the complete-data Fisher information matrix. There are many interesting directions for future work. This buffered gradient estimator for sequential data could be applied to other stochastic gradient methods such as maximum likelihood estimation or variational inference [@archer2015black; @krishnan2017structured]. The approach could also be extended to non-linear continuous SSMs (e.g. stochastic volatility models) replacing message passing with particle filtering [@andrieu2010particle; @cappe2005inference; @doucet2009tutorial; @olsson2008sequential]. The buffered gradient estimator could likewise be applied to diffusions with control variates [@baker2017control; @chatterji2018theory] or with augmented dynamics, such as using momentum (SGHMC) [@chen2014stochastic] or temperature (SGNHT) [@ding2014bayesian]. In terms of analysis, the standard SGLD error analysis could be extended to analyze the optimal trade-off between buffer size and subsequence length. Acknowledgments {#acknowledgments .unnumbered} =============== We would like to thank Drausin Wulsin, Jack Baker, Chris Nemeth and other members of the Dynamode lab at UW for their helpful discussions. This work was supported in part by ONR Grant N00014-15-1-2380 and NSF CAREER Award IIS-1350133. Nicholas J. Foti was supported by a Washington Research Foundation Innovation Postdoctoral Fellowship in Neuroengineering and Data Science. Supplement for SGMCMC for State Space Models This supplement is organized as follows. In Section \[supp-sec:lemma\_proofs\], we provide the proofs of Lemmas for Section \[sec:bounds\]. In Section \[supp-sec:models\], we provide additional details for how to calculate the forward backward messages, gradients, and preconditioning terms for the models in Section \[sec:models\]. In particular, in  \[supp-sec:LGSSM:lemma\_proofs\], we provide the proofs of the error bound lemmas from Section \[sec:models:LGSSM-bounds\]. Finally, in Section \[supp-sec:experiments\], we provide additional details and figures of experiments. Proof of Lemmas in Section \[sec:bounds\] {#supp-sec:lemma_proofs} ========================================= We now provide proofs to the Lemmas in section \[sec:bounds\]. We first present a proof of Lemma \[lemma:grad\_error\_to\_wasserstein\] that relates the error in the difference of expectations in Eq.  to Wasserstein distance. Let $g_t({u}_{t-1:t}) = {\nabla}\log p(y_t, {u}_t \, | \, {u}_{t-1}, \theta)$. Recall $\| g_t({u}_{t-1:t}) \|_{Lip} \leq L_U$ for all $t$ by assumption. Then, by the Kantorovich-Rubinstein duality formula Eq. , we have $$\left\| {\mathbb{E} \,}_{\gamma_{t-1:t}}\left[g_t({u}_{t-1:t})\right] - {\mathbb{E} \,}_{\widetilde\gamma_{t-1:t}}\left[g_t({u}_{t-1:t})\right] \right\|_2 \leq L_U \cdot \mathcal{W}_1(\gamma_{t-1:t}, \widetilde\gamma_{t-1:t}) \enspace.$$ Therefore, $$\begin{aligned} \| \bar{g}(\theta) - \tilde{g}(\theta) \|_2 &\leq \left\| \frac{T}{S} \sum_{t \in {{\mathcal{S}}}} {\mathbb{E} \,}_{\gamma_{t-1:t}}\left[g_t({u}_{t-1:t})\right] - {\mathbb{E} \,}_{\widetilde\gamma_{t-1:t}}\left[{\nabla}U_t({u}_{t-1:t})\right] \right\|_2 \\ &\leq \frac{T}{S} \sum_{t \in {{\mathcal{S}}}} \left\| {\mathbb{E} \,}_{\gamma_{t-1:t}}\left[g_t({u}_{t-1:t})\right] - {\mathbb{E} \,}_{\widetilde\gamma_{t-1:t}}\left[{\nabla}U_t({u}_{t-1:t})\right] \right\|_2 \\ &\leq\frac{T}{S} \cdot L_U \cdot \sum_{t \in {{\mathcal{S}}}} \mathcal{W}_1(\gamma_{t-1:t}, \widetilde\gamma_{t-1:t}) \enspace.\end{aligned}$$ We now present the proof of Lemma \[lemma:second\_wass\_moment\] that relates the $1$-Wasserstein distance between distributons $(\gamma', \widetilde\gamma')$ of ${u}{u}^T$ to the $2$-Wasserstein distance between $(\gamma,\widetilde\gamma)$ over ${u}$. Let $\xi$ be a joint distribution over ${u}$ and $\widetilde{u}$ with marginals $\gamma$ and $\widetilde\gamma$. Let $w := \widetilde{u}- {u}$, which implies $\widetilde{u}= {u}+ w$. Then we have $$\begin{aligned} {\mathbb{E} \,}\| \widetilde{u}\widetilde{u}^T - {u}{u}^T \|_F &= {\mathbb{E} \,}\| {u}w^T + w {u}^T + w w^T \|_F \\ &\leq {\mathbb{E} \,}\| {u}w^T \|_F + {\mathbb{E} \,}\|w {u}^T\|_F + {\mathbb{E} \,}\|w w^T \|_F \\ &= 2{\mathbb{E} \,}| {u}^T w | + {\mathbb{E} \,}[\|w\|^2] \\ &\leq 2\sqrt{{\mathbb{E} \,}[\|{u}\|^2] {\mathbb{E} \,}[\|w\|^2]} + {\mathbb{E} \,}[\|w\|^2] \\ &\leq (2\sqrt{M} + 1) \max \, \{{\mathbb{E} \,}[\|w\|^2]^{1/2}, {\mathbb{E} \,}[\|w\|^2] \} \\ &= (2\sqrt{M} + 1) \max \, \{{\mathbb{E} \,}[\|\widetilde{u}- {u}\|^2]^{1/2}, {\mathbb{E} \,}[\|\widetilde{u}- {u}\|^2] \}\end{aligned}$$ where we observe $\|xx^T\|_F = \| x x^T\|_2 = \|x\|_2^2 = x^T x$ and we use Cauchy-Schwartz. Taking the infimum over all $\xi$ gives the result $$\begin{aligned} \mathcal{W}_1(\gamma', \widetilde\gamma') &= \inf_\xi {\mathbb{E} \,}\| \widetilde{u}\widetilde{u}^T - {u}{u}^T \|_F \\ &\leq \inf_\xi \left[(2\sqrt{M} + 1) \max \, \{{\mathbb{E} \,}[\|\widetilde{u}- {u}\|^2]^{1/2}, \, {\mathbb{E} \,}[\|\widetilde{u}- {u}\|^2] \} \right] \\ &= (2\sqrt{M} + 1) \max \, \{ \inf_\xi {\mathbb{E} \,}[\|\widetilde{u}- {u}\|^2]^{1/2}, \, \inf_\xi {\mathbb{E} \,}[\|\widetilde{u}- {u}\|^2] \} \\ &= (2\sqrt{M} + 1) \cdot \max_{r \in {1, 1/2}} \mathcal{W}_2(\gamma, \widetilde\gamma)^{r} \enspace.\end{aligned}$$ We now prove Lemma \[lemma:geometric\_wasserstein\] that bounds $\mathcal{W}_p(\gamma_{t-1:t}, \widetilde\gamma_{t-1:t})$ in terms of buffer size, if the forward and backward random maps $f_t, b_t$ are Lipschitz. We will first prove Eq. . Recall $f_t$ is Lipschitz with constant $L_f < 1$ for all $t \in {{\mathcal{S}^*}}$. Let $\xi_{t:t+1}$ be a joint distribution over ${u}_{t:t+1}$ and $\widehat{u}_{t:t+1}$ with marginals $\gamma_{t:t+1}$ and $\widehat\gamma_{t:t+1}$. Let $\xi_{t}$ be a joint distribution over ${u}_{t}$ and $\widehat{u}_{t}$ with marginals $\gamma_{t}$ and $\widehat\gamma_{t}$. Then for all $t \in {{\mathcal{S}}}$, we have $$\begin{aligned} \mathcal{W}_p(\gamma_{t:t+1}, \widehat\gamma_{t:t+1})^p &\leq \inf_{\xi_{t:t+1}} \int \| {u}_{t} - \hat{u}_{t}\|_2^p + \| {u}_{t+1} - \hat{u}_{t+1}\|_2^p \, d\xi_{t:t+1}({u}_{t:t+1}, \hat{u}_{t:t+1}) \\ &\leq \inf_{\xi_t} \int \| {u}_{t} - \hat{u}_{t}\|_2^p + \| f_t({u}_t) - f_t(\hat{u}_t) \|_2^p \, d\xi_t({u}_t, \hat{u}_t) df_t \\ &\leq \inf_{\xi_t} \int \| {u}_{t} - \hat{u}_{t}\|_2^p + L_f^p \cdot \| {u}_t -\hat{u}_t \|_2^p \, d\xi_t({u}_t, \hat{u}_t) \\ &\leq (1 + L_f^p) \cdot \mathcal{W}_p(\gamma_{t}, \widehat\gamma_{t})^p\end{aligned}$$ Repeatedly applying Eq.  completes the proof for Eq.  $$\begin{aligned} \mathcal{W}_p(\gamma_{t-1:t}, \widehat\gamma_{t-1:t}) &\leq (1 + L_f^p)^{1/p} \cdot \mathcal{W}_p(\gamma_{t-1}, \widehat\gamma_{t-1}) \\ &\leq (1 + L_f^p)^{1/p} \cdot L_f \cdot \mathcal{W}_p(\gamma_{t-2}, \widehat\gamma_{t-2}) \\ &\leq (1 + L_f^p)^{1/p} \cdot L_f^2 \cdot \mathcal{W}_p(\gamma_{t-3}, \widehat\gamma_{t-3}) \\ &\leq \ldots \\ &\leq (1 + L_f^p)^{1/p} \cdot L_f^{B+t-1} \cdot \mathcal{W}_p(\gamma_{-B}, \widehat\gamma_{-B}) \enspace.\end{aligned}$$ The proof of Eq.  is identical. Additional Model Details {#supp-sec:models} ======================== Gaussian HMM {#supp-sec:hmm} ------------ See Sections \[sec:models:HMM\] for notation. ### Forward Backward The forward and backward recursions (Eqs.  and ) for an HMM are $$\begin{aligned} \label{eq:hmm_forward_message} \alpha_t & := p(z_t, y_{\leq t}) = \alpha_{t-1} \cdot \Pi \cdot P_t \\ \label{eq:hmm_backward_message} \beta_t & := p(y_{>t} \, | \, z_t) = \Pi \cdot P_{t+1} \beta_{t+1} \enspace,\end{aligned}$$ where $\alpha_{-T} = \mathbf{1}/K$, $\beta_T = \mathbf{1}$, and $$P_t := \operatorname{diag}\{\mathcal{N}(y_t \, | \, \mu_k, \Sigma_k)\}_{k=1}^K \enspace.$$ Given the messages $\alpha_t, \beta_t$, the marginal and pairwise posteriors of the latent states are computed as $$\begin{aligned} \gamma_t(z_t) & := p(z_t \, | \, y) \propto \alpha_t \odot \beta_t \\ \gamma_{t:t-1}(z_{t-1}, z_t) & := p(z_{t-1}, z_t \, | \, y) \propto \operatorname{diag}(\alpha_{t-1}) \cdot \Pi \cdot P_t \cdot \operatorname{diag}(\beta_t) \enspace. \label{eq:hmm_pairwise_marginal}\end{aligned}$$ ### Gradient Estimator As stated in Sec. \[sec:models:HMM\], we use the ‘expanded mean’ parameters of $\Pi$ instead of $\Pi$ (as in [@patterson2013stochastic]) and the Cholesky decomposition of $\Sigma^{-1}_k$ instead of $\Sigma_k$ to ensure positive definiteness. The expanded mean parametrization is $\phi \in {\mathbb{R}}^{K \times K}_+$ where $\Pi_{k,\cdot} = \phi_{k,\cdot} / \sum_{k'} \phi_{k, k'}$. The Cholesky decomposition of the precision $\Sigma^{-1}_k$ is $\psi_{\Sigma_k}$ such that $\psi_{\Sigma_k} \psi_{\Sigma_k}^T = \Sigma_k^{-1}$. The gradient of the marginal loglikelihood takes the form $$\begin{aligned} \label{eq:gaussian_hmm_marginal_loglike_start} {\nabla}_{\phi_k} \log p(y \, | \, \theta) &= \sum_{t \in \mathcal{T}} {\mathbb{E} \,}_{z_t, z_{t-1} | y}[\mathbb{I}(z_{t-1}=k) \cdot \phi_k^{-1} \odot (\vec{e}_{z_t} - \Pi_{k})] \\ {\nabla}_\mu \log p(y | \theta) &= \sum_{t=1}^T {\mathbb{E} \,}_{z_t | y} \left[ \Sigma_{z_t}^{-1}(y_t - \mu_{z_t}) \right] \\ {\nabla}_{\psi_\Sigma} \log p(y \, | \, \theta)&= \sum_{t \in \mathcal{T}} {\mathbb{E} \,}_{z_t | y}\left[ \left(\Sigma_{z_t} - (y_t - \mu_{z_t})(y_t - \mu_{z_t})^T\right) {\psi_\Sigma}_{z_t} \right] \enspace. \label{eq:gaussian_hmm_marginal_loglike_end}\end{aligned}$$ As $z$ is discrete and these expectations only involve pairwise elements of $z$, they can be tractably computed as weighted average using $\gamma(z_{t}, z_{t-1})$ from forward backward. ### Preconditioning {#supp-sec:hmm_precond} For the Gaussian HMM, the complete-data Fisher information matrix is block diagonal. With some algebra, the Fisher information matrix, precondition matrices, and correction term are $$\begin{aligned} \label{eq:gaussian_hmm_precondition_start} \mathcal{I}_{\phi_k} &= (\operatorname{diag}(\Pi_k) - 11^T) \cdot (1^T\phi_k)^{-2} \, \Rightarrow \, D(\theta)_{\phi_k} = \operatorname{diag}(\phi_k) \enspace \text{and} \enspace \Gamma(\theta)_{\phi_k} = 1 \\ \mathcal{I}_{\mu_k} &= \Sigma_k^{-1} \, \Rightarrow \, D(\theta)_{\mu_k} = \Sigma_k \enspace \text{and} \enspace \Gamma(\theta)_{\mu_k} = 0 \\ \mathcal{I}_{\psi_{\Sigma_k}} &= 2 (I_m \otimes \Sigma_k) \, \Rightarrow \, D(\theta)_{\psi_{\Sigma_k}} = \frac{1}{2} (I_m \otimes \Sigma_k^{-1}) \enspace \text{and} \enspace \Gamma(\theta)_{\psi_{\Sigma_k}} = \psi_{\Sigma_k} \label{eq:gaussian_hmm_precondition_end}\end{aligned}$$ For $\phi_k$, we use $D(\theta)_{\phi_k} = \operatorname{diag}(\phi_k)$ and $\Gamma(\theta)_{\phi_k} = 1$, following past work [@patterson2013stochastic; @ma2017stochastic]. However, we observed that $\phi_k$ will be absorbed at $0$, whenever $\phi_k$ approaches to closely to $0$. To fix this we recommend adding a small identity matrix $\nu_\phi I_K$ (for some $\nu_\phi > 0$) to $D(\theta)_\phi$. An alternative solution is to use a stochastic Cox-Ingersoll-Ross process to sample $\pi$ instead [@baker2018large]. ARHMM {#supp-sec:arhmm} ----- See Section \[sec:models:ARHMM\] for notation. ### Forward Backward The forward backward recursions for the ARHMM are identical to the Gaussian HMM Eqs. -, where $P_t$ is now $$P_t := \operatorname{diag}\{\mathcal{N}(y_t \, | \, A_k \overline{y_t}, Q_k)\}_{k=1}^K \enspace.$$ ### Gradient Estimator The gradient of the marginal loglikelihood is similar to the Gaussian HMM Eqs. - with $\mu_k$ replaced with $A_k \overline{y}_t$ $$\begin{aligned} \label{eq:arhmm_marginal_loglike_start} {\nabla}_{\phi_k} \log p(y \, | \, \theta) &= \sum_{t \in \mathcal{T}} {\mathbb{E} \,}_{z_t, z_{t-1} | y}[\mathbb{I}(z_{t-1}=k) \cdot \phi_k^{-1} \odot (\vec{e}_{z_t} - \Pi_{k})] \\ {\nabla}_A \log p(y | \theta) &= \sum_{t=1}^T {\mathbb{E} \,}_{z_t | y} \left[ Q_{z_t}^{-1}(y_t - A_{z_t} \overline{y}_t) \overline{y}_t^T \right] \\ {\nabla}_{\psi_Q} \log p(y \, | \, \theta)&= \sum_{t \in \mathcal{T}} {\mathbb{E} \,}_{z_t | y}\left[ \left(Q_{z_t} - (y_t - A_{z_t} \overline{y}_t)(y_t - A_{z_t} \overline{y}_t)^T\right) {\psi_Q}_{z_t} \right] \enspace. \label{eq:arhmm_marginal_loglike_end}\end{aligned}$$ ### Preconditioning {#preconditioning} The preconditioning terms for the ARHMM is similar to the Gaussian HMM $$\begin{aligned} \label{eq:arhmm_precondition_start} \mathcal{I}_{\phi_k} &= (\operatorname{diag}(\Pi_k) - 11^T) \cdot (1^T\phi_k)^{-2} \, \Rightarrow \, D(\theta)_{\phi_k} = \operatorname{diag}(\phi_k) \enspace \text{and} \enspace \Gamma(\theta)_{\phi_k} = 1 \\ \mathcal{I}_{A_k} &= {\mathbb{E} \,}_{y, z |\theta}[\overline{y}_t \overline{y}_t^{T}] \otimes Q_k^{-1} \, \Rightarrow \, D(\theta)_{A_k} = I_m \otimes Q_k \enspace \text{and} \enspace \Gamma(\theta)_{A_k} = 0 \\ \mathcal{I}_{\psi_{Q_k}} &= 2 (I_m \otimes Q_k) \, \Rightarrow \, D(\theta)_{\psi_{Q_k}} = \frac{1}{2} (I_m \otimes Q_k^{-1}) \enspace \text{and} \enspace \Gamma(\theta)_{\psi_{Q_k}} = \psi_{Q_k} \label{eq:arhmm_precondition_end}\end{aligned}$$ The expectation ${\mathbb{E} \,}[\overline{y}_t \overline{y}_t^T]$ does not have a closed form as the expectation is over $z$ is a combinatorial sum. Therefore, we choose to replace ${\mathbb{E} \,}[\overline{y}_t \overline{y}_t^T]$ with the identity matrix $I_m$ in our preconditioning matrix $D(\theta)_A$. LGSSM {#supp-sec:LGSSM} ----- See Section \[sec:models:LGSSM\] for notation. ### Forward Backward The recursions for the forward backward algorithm for LGSSMs is known as the Kalman smoother [@cappe2005inference; @bishop2006pattern; @fox2009bayesian]. Because the transition and emission processes are linear Gaussian, all forward messages, backward messages, and pairwise latent marginals $\gamma(x_t, x_{t-1})$ are Gaussian. $$\begin{aligned} \label{eq:lds_kalman_forward} \alpha_t &:= p(x_t, y_{\leq t}) = \mathcal{N}(x_t \, | \, \mu_{\alpha_t} = \Lambda_{\alpha_t}^{-1} h_{\alpha_t}, \Sigma_{\alpha_t} = \Lambda_{\alpha_t}^{-1}) \\ \label{eq:lds_kalman_backward} \beta_t &:= p(y_{>t} \, | \, x_t) \propto \mathcal{N}(x_t \, | \, \mu_{\beta_t} = \Lambda_{\beta_t}^{-1} h_{\beta_t}, \Sigma_{\beta_t} = \Lambda_{\beta_t}^{-1}) \enspace,\end{aligned}$$ where $h_{\alpha_t}, \Lambda_{\alpha_t}$ are the Gaussian natural parameters of $\alpha$ that satisfy the recursion $$\begin{aligned} \label{eq:lds_forward_lambda} \Lambda_{\alpha_t} &= C^T R^{-1} C + (Q + A\Lambda_{\alpha_{t-1}}^{-1} A^T)^{-1} \\ \label{eq:lds_forward_h} h_{\alpha_t} &= C^T R^{-1} y_{t} + (Q + A\Lambda_{\alpha_{t-1}}^{-1}A^T)^{-1} A \Lambda_{\alpha_{t-1}}^{-1} h_{\alpha_{t-1}} \enspace,\end{aligned}$$ and $h_{\beta_t}, \Lambda_{\beta_t}$ are the Gaussian natural parameters of $\beta$ that satisfy the recursion $$\begin{aligned} \label{eq:lds_backward_lambda} \Lambda_{\beta_t} &= A^TQ^{-1} A - A^T Q^{-1}(Q^{1} + C^T R^{-1} C + \Lambda_{\beta_{t+1}})^{-1} Q^{-1} A \\ \label{eq:lds_backward_h} h_{\beta_t} &= A^T Q^{-1}(Q^{-1} + C^T R^{-1} C + \Lambda_{\beta_{t+1}})^{-1}(C^T R^{-1}y_{t+1} + h_{\beta_{t+1}}) \enspace.\end{aligned}$$ Given the messages $\alpha_t, \beta_t$ the marginal and pairwise posteriors of the latent states $x_t$ and $(x_{t-1}, x_t)$ are computed as $$\begin{aligned} \gamma_t(x_t) := p(x_t \,|\, y) &\propto \alpha_t(x_t) \beta_t(x_t) \\ \nonumber &\propto \mathcal{N}(x_t \, | \, \mu = \Sigma (h_{\alpha_t} + h_{\beta_t}), \Sigma = (\Lambda_{\alpha_t} + \Lambda_{\beta_t})^{-1}) \end{aligned}$$ $$\begin{aligned} \label{eq:lds_pairwise_marginal} \gamma_{t-1, t}(x_{t-1}, x_t) := &p(x_{t-1}, x_t \,|\, y) \propto \alpha_{t-1}(x_{t-1} p(y_t, x_t \, | \, x_{t-1}) \beta_{t}(x_t) \\ \nonumber \propto\mathcal{N}\Big( \begin{bmatrix} x_{t-1} \\ x_t \end{bmatrix} \, \Big| \, &\mu = \Sigma \cdot \begin{bmatrix} h_{\alpha_{t-1}} \\ C^T R_{-1} y_t + h_{\beta_t}\end{bmatrix}, \\ \nonumber &\Sigma = \begin{bmatrix} \Lambda_{\alpha_{t-1}} + A^T Q^{-1} A & A^T Q^{-1} \\ Q^{-1} A & C^TR^{-1} C + Q^{-1} + \Lambda_{\beta_t} \end{bmatrix}^{-1} \Big) \enspace.\end{aligned}$$ ### Gradient Estimator We compute the gradient of marginal loglikelihood via Fisher’s identity $$\begin{aligned} \label{eq:lds_marginal_loglike_start} {\nabla}_A \log p(y | \theta) &= \sum_{t=1}^T {\mathbb{E} \,}_{x | y} \left[ Q^{-1}(x_t - A x_{t-1}) x_{t-1}^T \right] \\ {\nabla}_{\psi_Q} \log p(y | \theta) &= \sum_{t=1}^T {\mathbb{E} \,}_{x | y} \left[(Q - (x_t - A x_{t-1})(x_t - A x_{t-1})^T) \psi_Q \right] \\ {\nabla}_C \log p(y | \theta) &= \sum_{t=1}^T {\mathbb{E} \,}_{x | y} \left[ R^{-1}(y_t - C x_t) x_t^T \right] \\ {\nabla}_{\psi_R} \log p(y | \theta) &= \sum_{t=1}^T {\mathbb{E} \,}_{x | y} \left[(R - (y_t - C x_t)(y_t - C x_t)^T) \psi_R \right] \label{eq:lds_marginal_loglike_end}\end{aligned}$$ Because each gradient is linear with respect to first and second order terms (e.g. $x_t$, $x_t x_t^T$ and $x_t x_{t-1}^T$), their expectation of each of these terms is easily computable given $\gamma(x_t, x_{t-1})$. Let $\gamma_{t,t-1}(x_t, x_{t-1})$ be the Gaussian pairwise marginal posterior from forward backward (see Eq. ) $$\gamma_{t-1, t}(x_{t-1}, x_t) = \mathcal{N}\left( \begin{bmatrix} x_{t-1} \\ x_t \end{bmatrix} \, \big| \, \mu = \begin{bmatrix} \mu_{t-1} \\ \mu_t \end{bmatrix}, \Sigma = \begin{bmatrix} \Sigma_{t-1,t-1} & \Sigma_{t-1,t} \\ \Sigma_{t,t-1} & \Sigma_{t,t} \end{bmatrix} \right) \enspace.$$ Let $M = \Sigma + \mu\mu^T$ be the second moment of $\gamma_{t-1,t}$, that is $M_{t,t'} := {\mathbb{E} \,}[x_t, x_{t'}^T]$. Then the expectations in the summations of Eqs. - are $$\begin{aligned} {\mathbb{E} \,}_{x|y} \left[ (x_t - A x_{t-1}) x_{t-1}^T \right] &= M_{t,t-1} - A M_{t-1,t-1} \\ {\mathbb{E} \,}_{x|y} \left[ (x_t - A x_{t-1})(x_t - A x_{t-1})^T \right] &= M_{t,t} - A M_{t-1,t} - M_{t, t-1}A^T + A M_{t-1, t-1} A^T \\ {\mathbb{E} \,}_{x|y} \left[ (y_t - C x_t) x_t^T \right] &= y_t \mu_t^T - C M_{t,t} \\ {\mathbb{E} \,}_{x | y} \left[ (y_t - C x_t)(y_t - C x_t)^T \right] &= y_t y_t^T - C \mu_t y_t^T - y_t \mu_t^T C^T + C M_{t,t} C^T \enspace.\end{aligned}$$ ### Preconditioning {#preconditioning-1} For the LGSSM, the complete data Fisher information matrix is block diagonal. With some algebra, the Fisher information matrix, precondition matrices, and correction term are $$\begin{aligned} \label{eq:lds_precondition_start} \mathcal{I}_A = {\mathbb{E} \,}[x_t x_t]^T \otimes Q^{-1} \enspace &\Rightarrow \enspace D_A = I_n \otimes Q \enspace \text{and} \enspace \Gamma(\theta)_{A} = 0 \\ \mathcal{I}_{\psi_Q} = 2 (I_n \otimes Q) \enspace &\Rightarrow \enspace D_{\psi_Q} = \frac{1}{2} (I_n \otimes Q^{-1}) \enspace \text{and} \enspace \Gamma(\theta)_{\psi_Q} = \psi_Q \\ \mathcal{I}_C = {\mathbb{E} \,}[x_t x_t]^T \otimes R^{-1} \enspace &\Rightarrow \enspace D_C = I_n \otimes R \enspace \text{and} \enspace \Gamma(\theta)_{C} = 0 \\ \mathcal{I}_{\psi_R} = 2 (I_m \otimes R) \enspace &\Rightarrow \enspace D_{\psi_R} = \frac{1}{2} (I_m \otimes R^{-1}) \enspace \text{and} \enspace \Gamma(\theta)_{\psi_R} = \psi_R \enspace, \label{eq:lds_precondition_end}\end{aligned}$$ where ${\mathbb{E} \,}[x_t x_t]^T = \sum_{s = 0}^\infty A^s Q (A^s)^T$ for the LGSSM, In our experiments we chose to replace ${\mathbb{E} \,}[x_t x_t]^T$ with the identity matrix $I_n$ to match the ARHMM setup. ### Proof of Lemmas in Section \[sec:models:LGSSM-bounds\] {#supp-sec:LGSSM:lemma_proofs} We now provide proofs to the Lemmas in section \[sec:models:LGSSM-bounds\]. Note that these bound hold pointwise for $\theta$ for all random maps conditioned on any observed sequence $Y_{1:T}$. We first present a proof of Lemma \[lemma:LGSSM\_forward\_maps\] that shows the forward random maps $f_t$ are contractions if $\| A \| < 1$. For an LGSSM, the forward smoothing kernel $\mathcal{F}_t$ takes the form $$\mathcal{F}_t(x_{t+1}, x_{t} | y) \propto \underbrace{\beta_{t+1}(x_{t+1})p(y_{t+1}\, | \, x_{t+1})}_{p(y_{>t}\,|\, x_{t+1})} p(x_{t+1} \, | \, x_{t})$$ where $\beta_{t+1}$ is the backward message at time $t+1$ given by Eq. . The recursive formula for $\beta_t$ can be extended to $p(y_{>t} \, | \, x_{t+1}) \propto \mathcal{N}(x_{t+1} \, | \, \Lambda_{t+1}^{-1}h_{t+1}, \Lambda_{t+1}^{-1})$ with $$\begin{aligned} \label{eq:lds_backward_extended_lambda} \Lambda_t &= C^T R^{-1} C + A^T (Q + \Lambda_{t+1}^{-1})^{-1} A \\ h_{t} &= C^T R^{-1} y_{t} + A^T (Q + \Lambda_{t+1}^{-1})^{-1} \Lambda_{t+1} h_{t+1} \enspace.\end{aligned}$$ With this parametrization, the forward smoothing kernel takes the form $$\mathcal{F}_t(x_{t+1} | x_{t} ) = \mathcal{N}(x_{t+1} \, | \, (Q^{-1} + \Lambda_{t+1})^{-1} (Q^{-1}A x_t + h_{t+1}), (Q^{-1} + \Lambda_{t+1})^{-1}) \enspace.$$ Therefore our random map $f_t$ is $$f_t(x_t) = \underbrace{(Q^{-1} + \Lambda_{t+1})^{-1} Q^{-1}A}_{F^f_t} x_t + \underbrace{(Q^{-1} + \Lambda_{t+1})^{-1} h_{t+1} + (Q^{-1} + \Lambda_{t+1})^{-1/2} \nu_t}_{\zeta^f_t} \enspace,$$ where $\nu_t \sim \mathcal{N}(0, I)$ makes $f_t$ a random map. The Lipschitz constant for $f_t$ with respect to $x_t$ is $\|F_t^f\| = \| (Q^{-1} + \Lambda_{t+1})^{-1} Q^{-1} A \|$. From Eq. , $\Lambda_\theta = C^T R^{-1} C$ is a lower bound on $\Lambda_{t}$ and is tight when $\Lambda_{t+1} = 0$ (at the very beginning of the recursion). Therefore we have a uniform bound on the Lipschitz constants of $f_t$ : $$\|F_t^f\| = \| (Q^{-1} + \Lambda_{t+1})^{-1} Q^{-1} A \| < \| (Q^{-1} + \Lambda_\theta)^{-1} Q^{-1} A \| = L_f \enspace.$$ We now present a proof of Lemma \[lemma:LGSSM\_backward\_maps\] that similarly shows the backward random maps $b_t$ are contractions. We first prove the bound for general prior $p_0(x)$ and then present the special case when the prior variance is less than the steady state variance $V_\infty = (Q + A V_\infty A^T) = \sum_{k=0}^\infty A^k Q (A^T)^k$ and $A$ and $Q$ commute. The backward smoothing kernel $\mathcal{B}_t$ takes the form $$\mathcal{B}_t(x_{t-1}, x_{t} | y) \propto p(x_t \, | \, x_{t-1}) \alpha_{t-1}(x_{t-1})$$ where $\alpha_{t-1}$ is the forward message at time $t-1$. Recall from Eq.  the forward messages are $\alpha_t(x_{t}) \propto \mathcal{N}(x_{t} \, | \, \Lambda_{\alpha_t}^{-1}h_{\alpha_t}, \Lambda_{\alpha_t}^{-1})$. With this parametrization, the backward smoothing kernel takes the form $$\mathcal{B}_t(x_{t-1}, x_{t} | y) = \mathcal{N}(x_{t-1} \, | \, (A^T Q^{-1}A + \Lambda_{\alpha_{t-1}})^{-1} (A^T Q^{-1} x_t + h_{\alpha_{t-1}}), (A^T Q^{-1} A + \Lambda_{\alpha_{t-1}})^{-1}).$$ Our backward random map $b_t$ is thus $$b_t(x_t) = (A^T Q^{-1} A + \Lambda_{\alpha_{t-1}})^{-1}(A^T Q^{-1} x_t + h_{\alpha_{t-1}}) + (A^T Q^{-1}A + \Lambda_{\alpha_{t-1}})^{-1/2} \nu_t \enspace,$$ where $\nu_t \sim \mathcal{N}(0, I)$ with $$\begin{aligned} F_t^b &= (A^T Q^{-1} A + \Lambda_{\alpha_{t-1}})^{-1}A^T Q^{-1} \\ \zeta^b_t &= (A^T Q^{-1} A + \Lambda_{\alpha_{t-1}})^{-1}h_{\alpha_{t-1}} + (A^T Q^{-1}A + \Lambda_{\alpha_{t-1}})^{-1/2} \nu_t \enspace.\end{aligned}$$ The Lipschitz constant for $b_t$ with respect to $x_t$ is $$\|F_t^b\| = \| (A^T Q^{-1}A + \Lambda_{\alpha_{t-1}})^{-1} A^TQ^{-1} \| \enspace.$$ From Eq. , $\Lambda_\theta = C^T R^{-1} C$ is a lower bound on $\Lambda_{\alpha_t}$ and is tight when $\Lambda_{\alpha_{t-1}} = 0$ (at the very beginning of the recursion). Therefore we have a uniform bound on the Lipschitz constants of $b_t$ for Lemma \[lemma:geometric\_wasserstein\]: $$\label{eq:lds_backward_lipbound} \|F_t^b\| = \| (A^T Q^{-1}A + \Lambda_{\alpha_{t-1}})^{-1} A^TQ^{-1} \| \leq \| (A^T Q^{-1}A + \Lambda_{\theta})^{-1} A^TQ^{-1} \| = L_b \enspace,$$ where $$L_b =\|(A^T Q^{-1} A + C^T R^{-1} C\|^{-1}_2 \| A \|_2 \| Q \|_2 = \|A (Q A^T Q^{-1} A + Q C^T R^{-1} C)^{-1} \|_2 \enspace.$$ If the prior variance is less than the steady state variance $V_\infty$, then $\Lambda_\theta = C^T R^{-1} C + V_\infty^{-1}$ is a larger lower bound on $\Lambda_{\alpha_t}$ as by induction $\Lambda_{\alpha_1} = C^T R^{-1} C + V_\infty^{-1} = \Lambda_\theta$ and from Eq.  $$\Lambda_{\alpha_t}^{-1} \leq V_\infty \, \Rightarrow \, \Lambda_{\alpha_{t+1}} = C^T R^{-1} C + (Q + A \Lambda^{-1}_{\alpha_t} A^T)^{-1} \geq C^T R^{-1} C + (Q + A V_\infty A^T)^{-1} = \Lambda_\theta \enspace.$$ If $A$ and $Q$ commute, then $V_\infty = (Q^{-1} - A^T Q^{-1} A)^{-1}$ as $$\underbrace{(Q^{-1} - A^T Q^{-1} A)^{-1}}_{V_\infty} = Q + Q A^T(Q - A Q A^T)^{-1} A Q = \underbrace{Q + A(Q^{-1} - A^T Q^{-1}A)^{-1} A^T}_{Q + A V_\infty A^T} \enspace.$$ Therefore plugging lower bound for $\Lambda_{\alpha_t} \leq C^T R^{-1} C + Q^{-1} - A^T Q^{-1} A$ into Eq.  gives we obtain $$L_b =\|(Q^{-1} + C^T R^{-1} C\|^{-1}_2 \| A \|_2 \| Q \|_2 = \|A (I_n + Q C^T R^{-1} C)^{-1} \|_2 \enspace.$$ Finally, we prove Lemma \[lemma:LGSSM\_Lipschitz\] which bounds the Lipschitz constant for the complete data loglikelihood terms. The gradient of the complete data loglikelihood for $\theta = (A, Q, C, R)$ is $$\begin{aligned} \label{eq:lgssm_complete_grad_begin} {\nabla}_A \log p(y, x_t \, |\, x_{t-1}, \theta) &= Q^{-1}(x_t - A x_{t-1}) x_{t-1}^T \\ {\nabla}_{\psi_Q} \log p(y, x_t \, |\, x_{t-1}, \theta) &= (Q - (x_t - A x_{t-1})(x_t - A x_{t-1})^T)\psi_Q \\ {\nabla}_C \log p(y, x_t \, |\, x_{t-1}, \theta) &= R^{-1}(y_t - C x_t) x_t^T \\ {\nabla}_{\psi_R} \log p(y, x_t \, |\, x_{t-1}, \theta) &= (R - (y_t - C x_t)(y_t - C y_t)^T)\psi_R \enspace. \label{eq:lgssm_complete_grad_end}\end{aligned}$$ From Eqs. - it is clear that the complete data loglikelihood are quadratic form in $xx^T$ with matrices given by $\Omega$ in Lemma \[lemma:LGSSM\_Lipschitz\]. SLDS ---- See Section \[sec:models:SLDS\] for notation. As the SLDS does not have a closed form forward-backward algorithm, we instead present the details for the blocked Gibbs sampling scheme (conditional distributions and Initialization) used in Algorithm \[alg:noisygradient-slds\]. ### Blocked Gibbs Conditional Distributions The conditional posterior distribution of $x$ given $y$ and $z$ follows a time-varying LGSSM. To sample $x$, we can use the time-varying Kalman filter [@hamilton1994time]. We first calculate the forward messages $\alpha_t(x_t)$ using the Kalman filter recursion Eq.  with $A_t = A_{z_t}$, $C_t = C$, $Q_t = Q_{z_t}$, and $R_t = R$. Given $\alpha_t(x_t) \propto \mathcal{N}(x_t \, | \, \mu_{\alpha_t}, \Sigma_{\alpha_t})$, we sample $x$ using the backward sampler (starting from $t=T$ and descending) $$\label{eq:slds_x_conditional_gibbs} x_t \, |\, x_{t-1} \sim \begin{cases} \mathcal{N}\left( x_T \, | \, \mu = \mu_{\alpha_T} \, , \, \Sigma = \Sigma_{\alpha_T}\right) \text{ if } t = T \enspace, \text{ otherwise }\\ \mathcal{N}\left( x_t \, \Big| \, \mu = \Sigma (\Sigma_{\alpha_t}^{-1} \mu_{\alpha_t} + A_{z_{t+1}}^T Q^{-1}_{z_{t+1}} x_{t+1}) \, , \, \Sigma = (\Sigma_{\alpha_t}^{-1} + A_{z_{t+1}}^T Q^{-1}_{z_{t+1}} A_{z_{t+1}})^{-1}\right) \\ \end{cases}$$ The conditional posterior distribution of $z$ given $y$ and $x$ follows the ARHMM. To sample $z$, we apply a similar sampler for the ARHMM. We first calculate the backward messages $\beta_t(z_t)$ using the ARHMM forward messages Eq. , replacing $y$ with $x$. Given $\alpha_t(z_t)$, we then sample $z$ sequentially in ascending order using the forward sampler $$\label{eq:slds_z_conditional_gibbs} p(z_t = k \, | \, z_{t-1}, x, y) \propto p(x_t, y_t \, | \, x_{t-1}, z_t = k, \theta) \odot \Pi_{z_{t-1}, k} \odot \beta_t(k)\enspace.$$ Finally, the conditional posterior distribution of $z_t$ given $y$ and $z_{\backslash t}$ can be calculated using the forward backward algorithm to marginalize $x$. Specifically, $$\label{eq:slds_zt_conditional_gibbs} p(z_t = k \, | \, z_{\backslash t}, y) \propto \Pi_{z_{t-1}, k} \Pi_{k, z_t} \cdot \int \alpha_{t-1}(x_{t-1}) p(y_t, x_t \, | \, x_{t-1}, z_t = k) \beta_{t}(x_t) \, dx_t dx_{t-1} \enspace,$$ where $\alpha_{t-1}, \beta_t$ are calculated using Eqs. - with $A_{t'} = A_{z_{t'}}, Q_{t'} = Q_{z_{t'}}$ for all $t' \in {{\mathcal{S}^*}}\backslash \{t\}$. Note that Eq  requires $O(|{{\mathcal{S}^*}}|)$ time per time point $z_t$; therefore one pass over $z_{{\mathcal{S}^*}}$ requires $O(|{{\mathcal{S}^*}}|^2)$. ### Initialization of Blocked Gibbs Sampler To sample $z$ from the filtered process, we recursively sample from the conditional distribution $z_t \, | \, y_t, z_{t-1}$ $$p(z_t = k \, | \, z_{t-1}, y_t) \propto \Pi_{z_{t-1}, k} \cdot \int \alpha_{t-1}(x_{t-1}) p(y_{t}, x_t \, | \, x_{t-1}, z_t = k) \, dx_t dx_{t-1} \enspace,$$ where $\alpha_{t-1}$ is calculated using Eq.  with $A_{t'} = A_{z_{t'}}, Q_{t'} = Q_{z_{t'}}$ for all $t' < t$. Because we do not condition on $y_{>t}$ when $z_t$ is sampled, we emphasize that this distribution is not the posterior $z \, | \, y$ (it is the *filtered* distribution, not the *smoothed* distribution). However, it provides a better initialization point than sampling $z$ from the prior. Alternatively, when $\dim(x) = n \leq \dim(y) = m$, we can initialize $z^{(0)}$ by sampling $z \, | \, x', y, \theta$ using Eq.  with $x' = y$. ### Gradient Estimator For the SLDS, the gradients are similarly a combination of those for the ARHMM Eqs. - and the LGSSM Eqs. -. ### Preconditioning {#preconditioning-2} For the SLDS, the precondition matrices are similarly a combination of those for the ARHMM Eqs. - and the LGSSM Eqs. -. $$\begin{aligned} \label{eq:slds_precondition_start} D(\theta)_{\phi_k} = \operatorname{diag}(\phi_k) \enspace &\text{and} \enspace \Gamma(\theta)_{\phi_k} = 1 \\ D(\theta)_{A_k} = I_m \otimes Q_k \enspace &\text{and} \enspace \Gamma(\theta)_{A_k} = 0 \\ D(\theta)_{\psi_{Q_k}} = \frac{1}{2} I_n \otimes Q_k^{-1} \enspace &\text{and} \enspace \Gamma(\theta)_{\psi_{Q_k}} = \psi_{Q_k} \\ D(\theta)_C = I_n \otimes R \enspace &\text{and} \enspace \Gamma(\theta)_{Q} = 0 \\ D(\theta)_{\psi_{R_k}} = \frac{1}{2} I_m \otimes R_k^{-1} \enspace &\text{and} \enspace \Gamma(\theta)_{\psi_{R_k}} = \psi_{R_k} \label{eq:slds_precondition_end} \enspace.\end{aligned}$$ Additional Experiment Details {#supp-sec:experiments} ============================= Experiment Hyperparameters -------------------------- ### Priors In our experiments, we use the following (conjugate) priors for $\theta$. For the discrete latent state sequence transition matrix $\Pi$, we use a flat-Dirichlet prior $$\Pr(\Pi_k) \propto \prod_{k'} \Pi_{k,k'}^{\alpha_{k,k'}-1} \enspace, \text{ where } \alpha_{k,k'} = 1 \enspace.$$ For the continuous transition matrix $A$, we use a matrix normal prior $$\Pr(A) \propto \exp\left(-\operatorname{tr}\left[V^{-1} (A-M)^T U^{-1}(A-M)\right]/2\right) \enspace,$$ with mean $M=0$, diagonal column covariance $V = 10^{2} \cdot I_n$, and row variance $U = Q$. For the noise covariances $Q$ and $R$, we use flat Wishart priors over $Q^{-1}$ and $R^{-1}$ $$\Pr(Q^{-1}) \propto |Q|^{(n+1-\nu)/2} e^{-\operatorname{tr}(\Psi Q^{-1})/2} \enspace, \enspace \Pr(R^{-1}) \propto |R|^{(m+1-\nu)/2} e^{-\operatorname{tr}(\Psi R^{-1})/2} \enspace,$$ where $\Psi = \nu \cdot I$ and $\nu = n+1$ or $m+1$. ### Sampling Subsequences {#supp-sec:sample_subsequences} In our experiments, we sample subsequences $\mathcal{S} = \{t_1, \ldots, t_{S}\} \subset \mathcal{T} = \{1, \ldots, T\}$ uniformly from all $T-S+1$ possible contiguous subsequences. That is, $\Pr(t_1 = t) = 1/(T-S+1)$ for $t \in \{1, \ldots, T-S+1\}$and $\Pr(t \in \mathcal{S})$ is given by $$\label{eq:naive_partition} \Pr(t \in \mathcal{S}) = \frac{\min \{t, T-t+1, S, T-S+1\}}{T-S+1} \enspace.$$ An alternative method for sampling subsequences is to sample $\mathcal{S}$ from separate partitions of $\mathcal{T}$. That is if $T/S = L$ is a whole number, then $\Pr(t_1 = t) = 1/L$ for $t \in \{1+kL \, | \, k = 0, 1, \ldots, L-1 \}$ and $$\label{eq:strict_partition} \Pr(t \in \mathcal{S}) = \frac{1}{L} = \frac{S}{T} \enspace.$$ We found both methods work well in practice, but found empirically that the former has reduced variance in the stochastic gradient estimates $\hat{g}(\theta)$; therefore we use the former in our experiments. ### List of Hyperparameters - Synthetic Gaussian HMM $T = 10^4$ - Prior: $\Pi_k$ are Dirichlet, $\mu$ is Normal, and $Q^{-1}$ are Wishart. - Initialization: using K-means on $y_t$ - Stepsizes: ----------- -------- ------ ----------- -------- ------ No-Buffer Buffer Full No-Buffer Buffer Full 0.001 0.001 0.1 0.001 0.001 1.0 ----------- -------- ------ ----------- -------- ------ - Synthetic Gaussian HMM $T = 10^6$ - Prior: $\Pi_k$ are Dirichlet, $\mu$ is Normal, and $Q^{-1}$ are Wishart. - Initialization: using K-means on $y_t$ - Stepsizes: ----------- -------- ------ ----------- -------- ------ No-Buffer Buffer Full No-Buffer Buffer Full 0.001 0.001 0.1 0.001 0.01 0.1 ----------- -------- ------ ----------- -------- ------ - Ion Channel (Full) HMM - Prior: $\Pi_k$ are Dirichlet, $\mu$ is Normal, and $Q^{-1}$ are Wishart. - Initialization: using K-means on $y_t$ - Stepsizes: ----------- -------- ----------- -------- No-Buffer Buffer No-Buffer Buffer 0.0001 0.0001 0.01 0.01 ----------- -------- ----------- -------- - Ion Channel (Subset) HMM - Prior: $\Pi_k$ are Dirichlet, $\mu$ is Normal, and $Q^{-1}$ are Wishart. - Initialization: using K-means on $y_t$ - Stepsizes: ----------- -------- ----------- -------- No-Buffer Buffer No-Buffer Buffer 0.001 0.001 0.001 0.001 ----------- -------- ----------- -------- - Synthetic ARHMM $T = 10^4$ - Prior: $\Pi_k$ are Dirichlet, $A$ is matrix Normal, and $Q^{-1}$ are Wishart. - Initialization: using K-means on $[y_t, y_{t-1}]$ - Stepsizes: ----------- -------- ------ ----------- -------- ------ No-Buffer Buffer Full No-Buffer Buffer Full 0.0001 0.0001 0.01 0.001 0.001 0.1 ----------- -------- ------ ----------- -------- ------ - Synthetic ARHMM $T = 10^6$ - Prior: $\Pi_k$ are Dirichlet, $A$ is matrix Normal, and $Q^{-1}$ are Wishart. - Initialization: using K-means on $[y_t, y_{t-1}]$ - Stepsizes: ----------- -------- ------ ----------- -------- ------ No-Buffer Buffer Full No-Buffer Buffer Full 0.0001 0.0001 0.1 0.0001 0.0001 0.1 ----------- -------- ------ ----------- -------- ------ - Canine Seizure ARHMM - Prior: $\Pi_k$ are Dirichlet, $A$ is matrix Normal, and $Q^{-1}$ are Wishart. - Initialization: using K-means on $[y_t, y_{t-1}]$ - Stepsizes: SGLD $=0.01$, SGRLD $=0.1$. - Synthetic LGSSM $T = 10^4$ - Prior: $A$ is matrix Normal and $Q^{-1}, R^{-1}$ are Wishart. - Initialization: From prior with $\nu = 4, \Psi=4\cdot I_2$ for the Wishart priors. - Stepsizes: ----------- -------- ------ ----------- -------- ------ No-Buffer Buffer Full No-Buffer Buffer Full 0.01 0.01 0.1 0.01 0.01 0.1 ----------- -------- ------ ----------- -------- ------ - Synthetic LGSSM $T = 10^6$ - Prior: $A$ is matrix Normal and $Q^{-1}, R^{-1}$ are Wishart. - Initialization: From prior with $\nu = 4, \Psi=4\cdot I_2$ for the Wishart priors. - Stepsizes: ----------- -------- ------ ----------- -------- ------ No-Buffer Buffer Full No-Buffer Buffer Full 0.01 0.01 1.0 0.01 0.01 1.0 ----------- -------- ------ ----------- -------- ------ - Synthetic SLDS $T = 10^4$ - Prior: $\Pi_k$ is Dirichlet, $A_k$ is matrix Normal and $Q_k^{-1}, R^{-1}$ are Wishart. - Initialization: $R$ from Wishart Prior, $\Pi, A, Q$ from $K$-means as in ARHMM. - Stepsizes: SGRLD X $=0.5$, SGRLD Z $=0.1$, SGRLD XZ $=0.1$. - Synthetic SLDS $T = 10^6$ - Prior: $\Pi_k$ is Dirichlet, $A_k$ is matrix Normal and $Q_k^{-1}, R^{-1}$ are Wishart. - Initialization: $R$ from Wishart Prior, $\Pi, A, Q$ from $K$-means as in ARHMM. - Stepsizes: SGRLD X $=0.5$, SGRLD Z $=0.1$, SGRLD XZ $=0.1$. - Canine Seizure SLDS - Prior: $\Pi_k$ is Dirichlet, $A_k$ is matrix Normal and $Q_k^{-1}, R^{-1}$ are Wishart. - Initialization: $R$ from Wishart Prior, $\Pi, A, Q$ from $K$-means as in ARHMM. - Stepsizes: SGRLD $=0.1$, SGLD $=0.1$. - Daily Weather SLDS - Prior: $\Pi_k$ is Dirichlet, $A_k$ is matrix Normal and $Q_k^{-1}, R^{-1}$ are Wishart. - Initialization: $R$ from Wishart Prior, $\Pi, A, Q$ from $K$-means as in ARHMM. - Stepsizes: SGRLD $=0.1$, SGLD $=0.1$. - Hourly Weather SLDS - Prior: $\Pi_k$ is Dirichlet, $A_k$ is matrix Normal and $Q_k^{-1}, R^{-1}$ are Wishart. - Initialization: $R$ from Wishart Prior, $\Pi, A, Q$ from $K$-means as in ARHMM. - Stepsizes: SGRLD $=0.1$, SGLD $=0.01$. Additional Metric Details ------------------------- To assess the ‘mixing’ rate of our MCMC samplers, we measure each sampled chain’s kernel Stein divergence (KSD) to the posterior [@liu2016kernelized; @gorham2017measuring]. Given a chain of sampled $\{\theta^{(i)}\}_{1}^N$ (after burnin and thinning), let $q(\theta)$ be the empirical distribution of the samples, that is $$q(\theta) = \frac{1}{N}\sum_{i = 1}^{N} \delta_{\theta = \theta^{(i)}} \enspace.$$ Then the KSD between $q(\theta)$ and the posterior distribution $p(\theta)$ is $$\begin{aligned} KSD(q,p) &= \sum_{d = 1}^{\text{dim}(\theta)} \sqrt{\sum_{i,i' = 1}^n \frac{k_0^d(\theta_i, \theta_{i'})}{n^2}} \enspace, \text{ where } \enspace \\ k_0^d(\theta_i, \theta_{i'}) &= {\nabla}_{\theta_d} \log p(\theta_i) k(\theta_i, \theta_{i'}) {\nabla}_{\theta_d} \log p(\theta_{i'}) + {\nabla}\log p(\theta_{i'}) {\nabla}_x k(\theta_i, \theta_{i'}) \\ \nonumber &\quad + {\nabla}\log p(\theta_i) {\nabla}_y k(\theta_i, \theta_{i'}) + {\nabla}_x {\nabla}_y k(\theta_i, \theta_{i'})\end{aligned}$$ and $k(\cdot, \cdot)$ is a valid kernel function. Following [@gorham2017measuring], we use the inverse multiquadratic kernel (IMQ) $k(x,y) = (1+\|x-y\|_2^2)^{-0.5}$ in our experiments. As full gradient evaluations ${\nabla}\log p(\theta)$ are computationally intractable for our long time series, we replace them with stochastic estimates based on Eq.  using $S = 10^4$ and $B=100$ when $T > 10^4$. To measure the recovery of discrete latent state variables $z_t$ when the true latent states are known (e.g. in synthetic experiments), we use normalized mutual information (NMI). NMI is an information theoretic measure of similarity between discrete assignments [@vinh2010information]. $$\text{NMI}(Z_i, Z_*) = \frac{I(Z_i, Z_*)}{\sqrt{H(Z_i)H(Z_*)}} \enspace, \text{ with } Z_i = (z^{(i)}_1, \ldots, z^{(i)}_T) \enspace,$$ where $I(X,Y)$ is mutual information and $H(X)$ is entropy. NMI is maximized at 1 when the assignments are equal up to a permutation and minimized at 0 when the assignments share no information. This serves as ‘clustering’ or segmentation metric for measuring the coherence between our model’s inferred latent states and the true latent states. To measure the recovery of continuous latent state variables $x_t$ when the true latent states are known, we use root mean-squared error (RMSE) $\text{RMSE}(x, x') = \sum_t \| x_t - x_t'\|_2$. Synthetic Gaussian HMM {#supp-sec:experiments-synthHMM} ---------------------- Following [@foti2014stochastic; @ma2017stochastic], we generate data from a Gaussian HMM with $K = 8$ latent states (see Figure \[fig:gausshmm\_grad\_error\] (left)) This *reversed cycles* (RC) dataset strongly transitions between two cycles over three states, each in opposite directions. ![(Left) Sample dataset; arrows indicate Markov transitions. Stochastic gradient error ${\mathbb{E} \,}_{{\mathcal{S}}}\|\bar{g}(\theta) - \tilde{g}(\theta)\|_2$: (center) varying subsequence size $S$ for no-buffer $B=0$ and buffer $B=5$, (right) varying buffer size $B$ for $S = 2$ and $S=50$. Error bars are SD over $100$ randomly generated datasets. []{data-label="fig:gausshmm_grad_error"}](figures/RC_demo.pdf){width="\textwidth"} ![(Left) Sample dataset; arrows indicate Markov transitions. Stochastic gradient error ${\mathbb{E} \,}_{{\mathcal{S}}}\|\bar{g}(\theta) - \tilde{g}(\theta)\|_2$: (center) varying subsequence size $S$ for no-buffer $B=0$ and buffer $B=5$, (right) varying buffer size $B$ for $S = 2$ and $S=50$. Error bars are SD over $100$ randomly generated datasets. []{data-label="fig:gausshmm_grad_error"}](figures/19-06-21/gauss_hmm_rc_1000/group_varyS.pdf){width="\textwidth"} ![(Left) Sample dataset; arrows indicate Markov transitions. Stochastic gradient error ${\mathbb{E} \,}_{{\mathcal{S}}}\|\bar{g}(\theta) - \tilde{g}(\theta)\|_2$: (center) varying subsequence size $S$ for no-buffer $B=0$ and buffer $B=5$, (right) varying buffer size $B$ for $S = 2$ and $S=50$. Error bars are SD over $100$ randomly generated datasets. []{data-label="fig:gausshmm_grad_error"}](figures/19-06-21/gauss_hmm_rc_1000/group_varyB.pdf){width="\textwidth"} Figure \[fig:gausshmm\_grad\_error\] (right-pair) are plots of the stochastic gradient error ${\mathbb{E} \,}_{{\mathcal{S}}}\|\bar{g}(\theta) - \tilde{g}(\theta)\|_2$ between the unbiased and buffered estimates evaluated at the true model parameters $\theta=\theta^*$. Similar to the ARPHMM and LGSSM, we see that the error decays $O(1/S)$ and that buffering deceases the error by orders of magnitude in Figure \[fig:gausshmm\_grad\_error\] (center). In Figure \[fig:gausshmm\_grad\_error\] we see that the error decays geometrically in buffer size $O(L^B)$. For this RC dataset, the geometric decay rate $L$ is very small; thus small buffers (e.g. $B=2$) reduce the error drastically. From Figure \[fig:gausshmm\_grad\_error\] (center), we see that the stochastic gradients are heavily biased without buffering (orange) for small subsequence lengths, as they fail to capture the structured transitions between states. However this bias disappears with buffering (blue). From Figure \[fig:gausshmm\_grad\_error\] (right), we see that the stochastic gradient decays quickly with increasing buffer size $B$ for small subsequence $S = 2$ (purple). The bias in the stochastic gradients of observations parameters $(\mu, \Sigma)$ is less extreme than for transition matrix $\Pi$ which is associated with the latent states; we include their error plots in the Supplement. ![Metrics vs Runtime on RC data with $T = 10^4$ (top), $T=10^6$ (bottom), for different methods: , , and SGMCMC. For SGMCMC methods, solid () and dashed () lines indicate SGRLD and SGLD respectively. The different metrics are: (left) heldout loglikelihoood and (right) transition matrix estimation error $MSE(\hat\Pi^{(s)}, \Pi^*)$. . []{data-label="fig:gausshmm_rc_metrics"}](./figures/19-03-11/GaussHMM_10k/gibbs_heldloglike_vs_logruntime.pdf){width="\textwidth"} ![Metrics vs Runtime on RC data with $T = 10^4$ (top), $T=10^6$ (bottom), for different methods: , , and SGMCMC. For SGMCMC methods, solid () and dashed () lines indicate SGRLD and SGLD respectively. The different metrics are: (left) heldout loglikelihoood and (right) transition matrix estimation error $MSE(\hat\Pi^{(s)}, \Pi^*)$. . []{data-label="fig:gausshmm_rc_metrics"}](./figures/19-03-11/GaussHMM_10k/gibbs_logmse_pi_vs_logruntime.pdf){width="\textwidth"} ![Metrics vs Runtime on RC data with $T = 10^4$ (top), $T=10^6$ (bottom), for different methods: , , and SGMCMC. For SGMCMC methods, solid () and dashed () lines indicate SGRLD and SGLD respectively. The different metrics are: (left) heldout loglikelihoood and (right) transition matrix estimation error $MSE(\hat\Pi^{(s)}, \Pi^*)$. . []{data-label="fig:gausshmm_rc_metrics"}](./figures/19-03-11/GaussHMM_1m/gibbs_heldloglike_vs_logruntime.pdf){width="\textwidth"} ![Metrics vs Runtime on RC data with $T = 10^4$ (top), $T=10^6$ (bottom), for different methods: , , and SGMCMC. For SGMCMC methods, solid () and dashed () lines indicate SGRLD and SGLD respectively. The different metrics are: (left) heldout loglikelihoood and (right) transition matrix estimation error $MSE(\hat\Pi^{(s)}, \Pi^*)$. . []{data-label="fig:gausshmm_rc_metrics"}](./figures/19-03-11/GaussHMM_1m/gibbs_logmse_pi_vs_logruntime.pdf){width="\textwidth"} ![Boxplot of MCMC samples for RC data $T = 10^6$ for select values of $\Pi$. []{data-label="fig:gausshmm_rc_boxplot"}](figures/19-03-11/GaussHMM_1m/boxplot_pi.pdf){width="\textwidth"} In Figures \[fig:gausshmm\_rc\_metrics\] and \[fig:gausshmm\_rc\_boxplot\], we compare SGLD (no-buffer and buffer), SGRLD (no-buffer and buffer), and Gibbs. We run our samplers on one training sequence and evaluate performance on another test sequence. We consider sequence lengths of $T = 10^4$ and $T=10^6$. For the SGMCMC methods, we use a subsequence size of $S = 2$ and a buffer size of $B=0$ (no-buffer) or $B=2$ (buffer). From Figure \[fig:gausshmm\_rc\_metrics\] we again see that preconditioning helps convergence and mixing as SGRLD outperforms SGLD and from figure \[fig:gausshmm\_rc\_boxplot\] that buffering is necessary to properly estimate $\Pi$. Note that for the $T = 10^4$ case, we observe that SGRLD underestimates the variance of $\Pi$ (Figure \[fig:gausshmm\_rc\_metrics\] bottom-left). This is due to the preconditioner $D(\theta)_{\Pi} = \Pi$, creating absorbing states in the discretized dynamics (see comment in Section \[supp-sec:hmm\_precond\]). Sampler $\pi$ $\mu$ $\Sigma$ -- ------------------- ------------- ------------- ------------- SGLD (No Buffer) 1.95 (0.05) 1.12 (0.06) 2.46 (0.05) SGLD (Buffer) 1.33 (0.15) 1.16 (0.17) 1.99 (0.10) LD 1.99 (0.07) 1.50 (0.39) 2.10 (0.72) SGRLD (No Buffer) 1.69 (0.01) 0.77 (0.03) 2.49 (0.03) SGRLD (Buffer) 0.81 (0.01) 0.53 (0.01) 2.09 (0.05) RLD 0.85 (0.03) 0.54 (0.06) 2.09 (0.06) Gibbs 0.77 (0.01) 0.38 (0.06) 1.74 (0.07) SGLD (No Buffer) 4.25 (0.41) 2.93 (0.52) 4.63 (0.44) SGLD (Buffer) 3.34 (0.12) 2.84 (0.47) 3.94 (0.04) LD 5.42 (0.03) 4.07 (0.35) 5.30 (0.41) SGRLD (No Buffer) 3.67 (0.01) 2.59 (0.05) 3.99 (0.09) SGRLD (Buffer) 2.07 (0.04) 2.38 (0.08) 3.78 (0.09) RLD 3.91 (0.05) 3.17 (0.11) 4.76 (0.03) Gibbs 3.11 (0.05) 3.10 (0.07) 4.65 (0.05) : $\log_{10}$(KSD) by variable of RC samplers. Mean and (SD) over runs in Figure \[fig:gausshmm\_rc\_metrics\].[]{data-label="table:gausshmm_rc_ksd"} Table \[table:gausshmm\_rc\_ksd\] shows the KSD of different sampling methods for different components of $\theta$. Although full sequence methods performs well for small $T$, they perform worse for larger $T$ due to increase time between iterations. We also see that buffered SGRLD outperforms the other SGMCMC methods on $\Pi$, as the non-buffered methods are sampling from the incorrect distribution and SGLD suffers from extreme autocorrelation. Downsampled Ion Channel Recordings ---------------------------------- We now consider a downsampled version of the ion channel recording data presented in Section \[sec:experiments-HMM-ion\]. In particular, we consider downsampling the data by a factor of $50$ (as in [@ma2017stochastic]), resulting in $|{\mathcal{T}}| = 209,634$ observations. We again train on the first 90% and evaluate on the last 10% after applying a log-transform and normalizing the observations to use Gaussian emissions. For our SGMCMC methods we again use a subsequence size of $S = 10$ and a buffer size of $B=0$ (no-buffer) or $B=10$ (buffer). Figure \[fig:down\_sampled\_ion\_channel\] presents our results including comparisons to Gibbs sampling (red). For this (shorter) downsampled data, Gibbs sampling outperforms the SGMCMC methods. We see that the performance of the SGMCMC method is similar to the full sample case (compare to Figure \[fig:ion\_channel\]) and that SGRLD with buffering quickly reaches the same mode as Gibbs. ![Ion Channel Recordings: (Left) predictive loglikelihood vs runtime. (Center) example segmentation using Gibbs (Right) example segmentation using SGRLD. []{data-label="fig:down_sampled_ion_channel"}](figures/18-10-06/ion_subset_predictive_loglike.pdf){width="\textwidth"} ![Ion Channel Recordings: (Left) predictive loglikelihood vs runtime. (Center) example segmentation using Gibbs (Right) example segmentation using SGRLD. []{data-label="fig:down_sampled_ion_channel"}](figures/gibbs_gausshmm_ion-channel_fit.png){width="\textwidth"} ![Ion Channel Recordings: (Left) predictive loglikelihood vs runtime. (Center) example segmentation using Gibbs (Right) example segmentation using SGRLD. []{data-label="fig:down_sampled_ion_channel"}](figures/sgrld_gausshmm_ion-channel_fit.png){width="\textwidth"} Additional Synthetic Experiment Plots ------------------------------------- We now present additional plots for the synthetic data experiments. These plots show the MSE for ‘other’ components of $\theta$ to the true parameters of $\theta^*$ as well as other measures of fit such as predictive loglikelihood or recovery of the latent state sequence (NMI or RMSE). ### Gaussian HMM {#gaussian-hmm} The parametrization of the RC data set is as follows: $$\Pi = \begin{bmatrix} .01 & .99 & 0 & 0 & 0 & 0 & 0 & 0 \\ 0 & .01 & .99 & 0 & 0 & 0 & 0 & 0 \\ .85 & 0 & 0 & .15 & 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 & .01 & .99 & 0 & 0 \\ 0 & 0 & 0 & 0 & 0 & .01 & .99 & 0 \\ 0 & 0 & 0 & 0 & .85 & 0 & 0 & .15\\ 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \end{bmatrix}$$ with $$\mu_{1:8} = \left\{ (-50,0); (30,-30); (30,30); (-100,-10); (40,-40); (-65,0); (40, 40); (100,10) \right\},$$ and $\Sigma_k = 20*I_2$ for all states $k = 1:K$. Figure \[fig:gausshmm\_rc\_extra\_metrics\] are plots of additional metrics for the Gaussian HMM experiment on the RC data set. We see a bigger difference between the buffered and non-buffered methods in predictive loglikelihood as it is more sensitive to $\Pi$. For RC data, there is less difference between the buffered and non-buffered methods for estimating $A$ and $Q$ (Figure \[fig:gausshmm\_rc\_extra\_metrics\] (bottom)). ![Additional Metrics vs Runtime on RC data with $T = 10^4$ (top), $T=10^6$ (bottom), for different methods: , , and SGMCMC. For SGMCMC methods, solid () and dashed () lines indicate SGRLD and SGLD respectively. The different metrics are: (left) NMI, (center) estimation error $MSE(\hat{\mu}^{(s)}, \mu^*)$ (right) estimation error $MSE(\hat{Q}^{(s)}, Q^*)$ . []{data-label="fig:gausshmm_rc_extra_metrics"}](./figures/19-03-11/GaussHMM_10k/gibbs_mean_nmiz_vs_logruntime.pdf){width="\textwidth"} ![Additional Metrics vs Runtime on RC data with $T = 10^4$ (top), $T=10^6$ (bottom), for different methods: , , and SGMCMC. For SGMCMC methods, solid () and dashed () lines indicate SGRLD and SGLD respectively. The different metrics are: (left) NMI, (center) estimation error $MSE(\hat{\mu}^{(s)}, \mu^*)$ (right) estimation error $MSE(\hat{Q}^{(s)}, Q^*)$ . []{data-label="fig:gausshmm_rc_extra_metrics"}](./figures/19-03-11/GaussHMM_10k/gibbs_mean_logmse_mu_vs_logruntime.pdf){width="\textwidth"} ![Additional Metrics vs Runtime on RC data with $T = 10^4$ (top), $T=10^6$ (bottom), for different methods: , , and SGMCMC. For SGMCMC methods, solid () and dashed () lines indicate SGRLD and SGLD respectively. The different metrics are: (left) NMI, (center) estimation error $MSE(\hat{\mu}^{(s)}, \mu^*)$ (right) estimation error $MSE(\hat{Q}^{(s)}, Q^*)$ . []{data-label="fig:gausshmm_rc_extra_metrics"}](./figures/19-03-11/GaussHMM_10k/gibbs_mean_logmse_R_vs_logruntime.pdf){width="\textwidth"} ![Additional Metrics vs Runtime on RC data with $T = 10^4$ (top), $T=10^6$ (bottom), for different methods: , , and SGMCMC. For SGMCMC methods, solid () and dashed () lines indicate SGRLD and SGLD respectively. The different metrics are: (left) NMI, (center) estimation error $MSE(\hat{\mu}^{(s)}, \mu^*)$ (right) estimation error $MSE(\hat{Q}^{(s)}, Q^*)$ . []{data-label="fig:gausshmm_rc_extra_metrics"}](./figures/19-03-11/GaussHMM_1m/gibbs_mean_nmiz_vs_logruntime.pdf){width="\textwidth"} ![Additional Metrics vs Runtime on RC data with $T = 10^4$ (top), $T=10^6$ (bottom), for different methods: , , and SGMCMC. For SGMCMC methods, solid () and dashed () lines indicate SGRLD and SGLD respectively. The different metrics are: (left) NMI, (center) estimation error $MSE(\hat{\mu}^{(s)}, \mu^*)$ (right) estimation error $MSE(\hat{Q}^{(s)}, Q^*)$ . []{data-label="fig:gausshmm_rc_extra_metrics"}](./figures/19-03-11/GaussHMM_1m/gibbs_mean_logmse_mu_vs_logruntime.pdf){width="\textwidth"} ![Additional Metrics vs Runtime on RC data with $T = 10^4$ (top), $T=10^6$ (bottom), for different methods: , , and SGMCMC. For SGMCMC methods, solid () and dashed () lines indicate SGRLD and SGLD respectively. The different metrics are: (left) NMI, (center) estimation error $MSE(\hat{\mu}^{(s)}, \mu^*)$ (right) estimation error $MSE(\hat{Q}^{(s)}, Q^*)$ . []{data-label="fig:gausshmm_rc_extra_metrics"}](./figures/19-03-11/GaussHMM_1m/gibbs_mean_logmse_R_vs_logruntime.pdf){width="\textwidth"} ### ARHMM {#arhmm} Figure \[fig:arhmm\_synth\_extra\_metrics\] are plots of additional metrics for the ARHMM. ![Additional Metrics vs Runtime on ARHMM data with $T = 10^4$ (top), $T=10^6$ (bottom), for different methods: , , and SGMCMC. For SGMCMC methods, solid () and dashed () lines indicate SGRLD and SGLD respectively. The different metrics are: (left) NMI, (center) $MSE(\hat{A}^{(s)}, A^*)$ (right) $MSE(\hat{Q}^{(s)}, Q^*)$. []{data-label="fig:arhmm_synth_extra_metrics"}](./figures/19-03-11/ARPHMM_10k/gibbs_mean_nmiz_vs_logruntime.pdf){width="\textwidth"} ![Additional Metrics vs Runtime on ARHMM data with $T = 10^4$ (top), $T=10^6$ (bottom), for different methods: , , and SGMCMC. For SGMCMC methods, solid () and dashed () lines indicate SGRLD and SGLD respectively. The different metrics are: (left) NMI, (center) $MSE(\hat{A}^{(s)}, A^*)$ (right) $MSE(\hat{Q}^{(s)}, Q^*)$. []{data-label="fig:arhmm_synth_extra_metrics"}](./figures/19-03-11/ARPHMM_10k/gibbs_mean_logmse_D_vs_logruntime.pdf){width="\textwidth"} ![Additional Metrics vs Runtime on ARHMM data with $T = 10^4$ (top), $T=10^6$ (bottom), for different methods: , , and SGMCMC. For SGMCMC methods, solid () and dashed () lines indicate SGRLD and SGLD respectively. The different metrics are: (left) NMI, (center) $MSE(\hat{A}^{(s)}, A^*)$ (right) $MSE(\hat{Q}^{(s)}, Q^*)$. []{data-label="fig:arhmm_synth_extra_metrics"}](./figures/19-03-11/ARPHMM_10k/gibbs_mean_logmse_R_vs_logruntime.pdf){width="\textwidth"} ![Additional Metrics vs Runtime on ARHMM data with $T = 10^4$ (top), $T=10^6$ (bottom), for different methods: , , and SGMCMC. For SGMCMC methods, solid () and dashed () lines indicate SGRLD and SGLD respectively. The different metrics are: (left) NMI, (center) $MSE(\hat{A}^{(s)}, A^*)$ (right) $MSE(\hat{Q}^{(s)}, Q^*)$. []{data-label="fig:arhmm_synth_extra_metrics"}](./figures/19-03-11/ARPHMM_1m/gibbs_mean_nmiz_vs_logruntime.pdf){width="\textwidth"} ![Additional Metrics vs Runtime on ARHMM data with $T = 10^4$ (top), $T=10^6$ (bottom), for different methods: , , and SGMCMC. For SGMCMC methods, solid () and dashed () lines indicate SGRLD and SGLD respectively. The different metrics are: (left) NMI, (center) $MSE(\hat{A}^{(s)}, A^*)$ (right) $MSE(\hat{Q}^{(s)}, Q^*)$. []{data-label="fig:arhmm_synth_extra_metrics"}](./figures/19-03-11/ARPHMM_1m/gibbs_mean_logmse_D_vs_logruntime.pdf){width="\textwidth"} ![Additional Metrics vs Runtime on ARHMM data with $T = 10^4$ (top), $T=10^6$ (bottom), for different methods: , , and SGMCMC. For SGMCMC methods, solid () and dashed () lines indicate SGRLD and SGLD respectively. The different metrics are: (left) NMI, (center) $MSE(\hat{A}^{(s)}, A^*)$ (right) $MSE(\hat{Q}^{(s)}, Q^*)$. []{data-label="fig:arhmm_synth_extra_metrics"}](./figures/19-03-11/ARPHMM_1m/gibbs_mean_logmse_R_vs_logruntime.pdf){width="\textwidth"} ### LGSSM {#lgssm} Figure \[fig:lds\_extra\_metrics\] are plots of additional metrics for the LGSSM synthetic data. ![Additional Metrics vs Runtime on LGSSM data with $T = 10^4$ (top), $T=10^6$ (bottom), for different methods: , , and SGMCMC. For SGMCMC methods, solid () and dashed () lines indicate SGRLD and SGLD respectively. The different metrics are: (left) root-mean squared error (RMSE) between $\hat{x}$ and $x^*$, (center) estimation error $MSE(\hat{Q}^{(s)}, Q^*)$ (right) estimation error $MSE(\hat{R}^{(s)}, R^*)$ . []{data-label="fig:lds_extra_metrics"}](./figures/19-03-11/LGSSM_10k/gibbs_mean_rmsex_vs_logruntime.pdf){width="\textwidth"} ![Additional Metrics vs Runtime on LGSSM data with $T = 10^4$ (top), $T=10^6$ (bottom), for different methods: , , and SGMCMC. For SGMCMC methods, solid () and dashed () lines indicate SGRLD and SGLD respectively. The different metrics are: (left) root-mean squared error (RMSE) between $\hat{x}$ and $x^*$, (center) estimation error $MSE(\hat{Q}^{(s)}, Q^*)$ (right) estimation error $MSE(\hat{R}^{(s)}, R^*)$ . []{data-label="fig:lds_extra_metrics"}](./figures/19-03-11/LGSSM_10k/gibbs_mean_logmse_Q_vs_logruntime.pdf){width="\textwidth"} ![Additional Metrics vs Runtime on LGSSM data with $T = 10^4$ (top), $T=10^6$ (bottom), for different methods: , , and SGMCMC. For SGMCMC methods, solid () and dashed () lines indicate SGRLD and SGLD respectively. The different metrics are: (left) root-mean squared error (RMSE) between $\hat{x}$ and $x^*$, (center) estimation error $MSE(\hat{Q}^{(s)}, Q^*)$ (right) estimation error $MSE(\hat{R}^{(s)}, R^*)$ . []{data-label="fig:lds_extra_metrics"}](./figures/19-03-11/LGSSM_10k/gibbs_mean_logmse_R_vs_logruntime.pdf){width="\textwidth"} ![Additional Metrics vs Runtime on LGSSM data with $T = 10^4$ (top), $T=10^6$ (bottom), for different methods: , , and SGMCMC. For SGMCMC methods, solid () and dashed () lines indicate SGRLD and SGLD respectively. The different metrics are: (left) root-mean squared error (RMSE) between $\hat{x}$ and $x^*$, (center) estimation error $MSE(\hat{Q}^{(s)}, Q^*)$ (right) estimation error $MSE(\hat{R}^{(s)}, R^*)$ . []{data-label="fig:lds_extra_metrics"}](./figures/19-03-11/LGSSM_1m/gibbs_mean_rmsex_vs_logruntime.pdf){width="\textwidth"} ![Additional Metrics vs Runtime on LGSSM data with $T = 10^4$ (top), $T=10^6$ (bottom), for different methods: , , and SGMCMC. For SGMCMC methods, solid () and dashed () lines indicate SGRLD and SGLD respectively. The different metrics are: (left) root-mean squared error (RMSE) between $\hat{x}$ and $x^*$, (center) estimation error $MSE(\hat{Q}^{(s)}, Q^*)$ (right) estimation error $MSE(\hat{R}^{(s)}, R^*)$ . []{data-label="fig:lds_extra_metrics"}](./figures/19-03-11/LGSSM_1m/gibbs_mean_logmse_Q_vs_logruntime.pdf){width="\textwidth"} ![Additional Metrics vs Runtime on LGSSM data with $T = 10^4$ (top), $T=10^6$ (bottom), for different methods: , , and SGMCMC. For SGMCMC methods, solid () and dashed () lines indicate SGRLD and SGLD respectively. The different metrics are: (left) root-mean squared error (RMSE) between $\hat{x}$ and $x^*$, (center) estimation error $MSE(\hat{Q}^{(s)}, Q^*)$ (right) estimation error $MSE(\hat{R}^{(s)}, R^*)$ . []{data-label="fig:lds_extra_metrics"}](./figures/19-03-11/LGSSM_1m/gibbs_mean_logmse_R_vs_logruntime.pdf){width="\textwidth"} ### SLDS Figure \[fig:slds\_extra\_metrics\] are plots of additional metrics for the SLDS data. ![Additional Metrics vs Runtime on SLDS data: (Top) $|{\mathcal{T}}| = 10^4$, (Bottom) $|{\mathcal{T}}| = 10^6$. (Left) NMI between $\hat{z}$ and $z^*$. (Center) root-mean square error (RMSE) between $\hat{x}$ and $x^*$, (Right) estimation error $\| \theta^{(s)} - \theta^* \|$. Methods: , , , and . []{data-label="fig:slds_extra_metrics"}](./figures/18-09-27/slds_lowtrans_avg_nmi.pdf){width="\textwidth"} ![Additional Metrics vs Runtime on SLDS data: (Top) $|{\mathcal{T}}| = 10^4$, (Bottom) $|{\mathcal{T}}| = 10^6$. (Left) NMI between $\hat{z}$ and $z^*$. (Center) root-mean square error (RMSE) between $\hat{x}$ and $x^*$, (Right) estimation error $\| \theta^{(s)} - \theta^* \|$. Methods: , , , and . []{data-label="fig:slds_extra_metrics"}](./figures/18-09-27/slds_lowtrans_avg_rmse.pdf){width="\textwidth"} ![Additional Metrics vs Runtime on SLDS data: (Top) $|{\mathcal{T}}| = 10^4$, (Bottom) $|{\mathcal{T}}| = 10^6$. (Left) NMI between $\hat{z}$ and $z^*$. (Center) root-mean square error (RMSE) between $\hat{x}$ and $x^*$, (Right) estimation error $\| \theta^{(s)} - \theta^* \|$. Methods: , , , and . []{data-label="fig:slds_extra_metrics"}](./figures/18-09-27/slds_lowtrans_avg_mse_other.pdf){width="\textwidth"} ![Additional Metrics vs Runtime on SLDS data: (Top) $|{\mathcal{T}}| = 10^4$, (Bottom) $|{\mathcal{T}}| = 10^6$. (Left) NMI between $\hat{z}$ and $z^*$. (Center) root-mean square error (RMSE) between $\hat{x}$ and $x^*$, (Right) estimation error $\| \theta^{(s)} - \theta^* \|$. Methods: , , , and . []{data-label="fig:slds_extra_metrics"}](./figures/18-09-25/slds_lowtrans_1m_avg_nmi.pdf){width="\textwidth"} ![Additional Metrics vs Runtime on SLDS data: (Top) $|{\mathcal{T}}| = 10^4$, (Bottom) $|{\mathcal{T}}| = 10^6$. (Left) NMI between $\hat{z}$ and $z^*$. (Center) root-mean square error (RMSE) between $\hat{x}$ and $x^*$, (Right) estimation error $\| \theta^{(s)} - \theta^* \|$. Methods: , , , and . []{data-label="fig:slds_extra_metrics"}](./figures/18-09-25/slds_lowtrans_1m_avg_rmse.pdf){width="\textwidth"} ![Additional Metrics vs Runtime on SLDS data: (Top) $|{\mathcal{T}}| = 10^4$, (Bottom) $|{\mathcal{T}}| = 10^6$. (Left) NMI between $\hat{z}$ and $z^*$. (Center) root-mean square error (RMSE) between $\hat{x}$ and $x^*$, (Right) estimation error $\| \theta^{(s)} - \theta^* \|$. Methods: , , , and . []{data-label="fig:slds_extra_metrics"}](./figures/18-09-25/slds_lowtrans_avg_1m_mse_other.pdf){width="\textwidth"} [^1]: Department of Statistics, University of Washington, WA [^2]: Department of Electrical Engineering and Computer Sciences, UC Berkeley, CA. [^3]: Paul G. Allen School of Computer Science and Engineering, University of Washington, WA. Email: `[aicherc, nfoti, ebfox]@uw.edu`, `[email protected]` [^4]: As $\epsilon \geq C_\theta \rho_\theta^{-B} \Rightarrow B \geq -\log \epsilon/\log\rho_\theta + \log C_\theta/\log\rho_\theta \Rightarrow B$ is $O(\log \epsilon^{-1})$. [^5]: Python code for our method is available at <https://github.com/aicherc/sgmcmc_ssm_code> [^6]: These bounds have been extended to non-compact spaces for the *filtered* posterior, when the SSM satisfies a multiplicative drift condition [@whiteley2013stability]. [^7]: The random mapping $\psi$ is Lipschitz with constant $L$ if ${{\mathbb{E} \,}_\psi\| \psi({u}) - \psi({u}') \|_2} \leq L {\|{u}- {u}' \|_2} \ \forall {u}, {u}'$.

ArXiv

--- bibliography: - 'refs.bib' --- Version dated: [**Generalising rate heterogeneity across sites in statistical phylogenetics**]{} [Sarah E. Heaps$^1$, Tom M. W. Nye$^1$, Richard J. Boys$^1$,\ Tom A. Williams$^2$, Svetlana Cherlin$^3$ and T. Martin Embley$^4$]{} [ *$^1$School of Mathematics & Statistics, Newcastle University, Newcastle upon Tyne, NE1 7RU, U.K.\ $^2$School of Earth Sciences, University of Bristol, Bristol, BS8 1RJ, U.K.\ $^3$Institute of Genetic Medicine, Newcastle University, Newcastle upon Tyne, NE1 3BZ, U.K.\ $^4$Institute for Cell and Molecular Biosciences, Newcastle University, Newcastle upon Tyne, NE2 4HH, U.K*]{}.\ ### Abstract {#abstract .unnumbered} In phylogenetics, alignments of molecular sequence data for a collection of species are used to learn about their phylogeny – an evolutionary tree which places these species as leaves and ancestors as internal nodes. Sequence evolution on each branch of the tree is generally modelled using a continuous time Markov process, characterised by an instantaneous rate matrix. Early models assumed the same rate matrix governed substitutions at all sites of the alignment, ignoring the variation in evolutionary constraints. Substantial improvements in phylogenetic inference and model fit were achieved by augmenting these models with a set of multiplicative random effects that allowed different sites to evolve at different rates which scaled the baseline rate matrix. Motivated by this pioneering work, we consider an extension which allows quadratic, rather than linear, site-specific transformations of the baseline rate matrix. We derive properties of the resulting process and show that when combined with a particular class of non-stationary models, we obtain one that allows sequence composition to vary across both sites of the alignment and taxa. Formulating the model in a Bayesian framework, a Markov chain Monte Carlo algorithm for posterior inference is described. We consider two applications to alignments concerning the tree of life, fitting stationary and non-stationary models. In each case we compare inferences obtained under our site-specific quadratic transformation, with those under linear and site-homogeneous models. (**Keywords:** Phylogenetics; Across site rate heterogeneity; Compositional heterogeneity; Tree of life.) \[sec:intro\]Introduction ========================= In statistical phylogenetics, the goal is to learn about the evolutionary relationships amongst a collection of species, generally using DNA or protein sequence data. These relationships are represented through a rooted, bifurcating tree called a phylogeny. Substitutions in the molecular sequence alignment are typically modelled using continuous time Markov processes, parameterised through an instantaneous rate matrix. Early phylogenetic models were simplistic, generally assuming that the evolutionary process was in its stationary distribution and that substitutions at each site of the alignment could be described by the same underlying rate matrix. Under these models, the probability of change from one character state to another was therefore independent of both organismal lineage and the biochemical function of the nucleotide or amino acid in question. These simplifying assumptions were known to be false, but were made for the sake of mathematical convenience and computational tractability given the limited computing power for model fitting available at the time. In particular, it was already clear to early molecular evolutionists that rates of evolution vary as a result of the functional and structural constraints acting on a site: important sites evolve slowly because most mutations that arise at those sites are eliminated from the population by negative selection [@FM70]. @UZ71 showed that the numbers of substitutions occuring at different sites could be modelled using a negative binomial distribution. Later, @Yan93 incorporated the idea into statistical phylogenetics by allowing different sites to evolve at different rates. These rate parameters scaled the underlying Markov process rate matrix and were modelled as multiplicative random effects, with a unit mean gamma distribution. Incorporation of across-site rate variation into standard, stationary substitution models has led to major improvements in model fit and to the accuracy of phylogenetic inference [@Yan96]. But there are other, pervasive features of molecular sequence data that these models do not accommodate. In particular, nucleotide composition is believed to vary across both the sites of the alignment and the branches of the phylogenetic tree. For example, the GC-content of ribosomal DNA genes varies from 45-74% across the known diversity of cellular life [@CFHHE08], implying that the probabilities of each of the four nucleotides can change over time. These compositional shifts might reflect changing biases in DNA repair enzymes [@Sue88] or, at least for genes encoding structural RNAs, adaptation to different growth temperatures [@GL97]. As well as variation in sequence composition across taxa, there is also compositional variation observed among the different sites within an individual protein-coding sequence: due to functional constraints, most sites can tolerate only a limited, and typically biochemically homogeneous, subset of the twenty amino acids [@FM70]. The result is that, in addition to varying in evolutionary rate, sites can also differ in sequence composition. Neither across-branch nor across-site compositional variation are accommodated by standard stationary models, even after incorporating across-site rate variation. But as with heterogeneity in evolutionary rates, failure to account for variation in composition can lead to model misspecification and, therefore, serious phylogenetic error, as demonstrated by a number of empirical studies [@ETW93; @Fos04; @LBP07; @Phi11]. Although a few models have been developed to jointly model both sources of heterogeneity [@BL08; @JWRPJ14], they have not been widely used in practice due to computational difficulties with model-fitting. In a simple phylogenetic model, evolution at all sites is controlled by a single instantaneous rate matrix. The across-site rate variation model offers greater flexibility by allowing site-specific *linear* transformations of the baseline rate matrix. The improvement afforded by this simple modification served as motivation for this paper, and we describe a natural generalisation through a model that provides site-specific *quadratic* transformations of the baseline matrix. This allows qualitatively different patterns of transitions at different sites. Further, we demonstrate that when linear or quadratic across-site transformations are combined with a class of non-stationary Markov processes, we obtain computationally tractable models that allow sequence composition to vary both across branches of the tree and across sites of the alignment. The remainder of this paper is organised as follows. Section \[sec:models\] introduces phylogenetic models of sequence evolution and the incorporation of multiplicative random effects to allow rate variation across sites. Section \[sec:quash\] describes our quadratic generalisation and its properties. In Section \[sec:ns-quash\] we combine across-site linear and quadratic transformations with a general class of non-stationary substitution models and describe the properties of the resulting Markov processes. Section \[sec:inference\] addresses the issue of inference for models incorporating our quadratic transformation. Specifically, we take a Bayesian approach to inference and describe the posterior distribution of interest and details of our numerical approach to model-fitting via Markov chain Monte Carlo sampling. In Section \[sec:apps\] we consider analyses of two biological data sets; the first involving a stationary model and the second, a non-stationary model. In each case we compare the performance of a site-homogeneous model with analogous models incorporating linear and quadratic across-site transformations of the baseline rate matrix. Finally, we summarise our conclusions in Section \[sec:discussion\]. \[sec:models\]Phylogenetic models of sequence evolution ======================================================= Denote by ${y} = (y_{i,j})$ an alignment of molecular sequence data where $y_{i,j} \in \Omega_K$ is the character at the $j$th site for taxon $i$ and $\Omega_K$ is an alphabet with $K$ characters, for example, the DNA alphabet with $\Omega_4 = \{ {\tt A}, {\tt G}, {\tt C}, {\tt T} \}$. Denote the number of sites (columns) by $M$ and the number of taxa (rows) by $N$ and let ${\bm{y}}_j = (y_{1,j},\ldots,y_{N,j}){^\mathrm{T}}$ be the $j$-th column in the alignment. Consider a rooted, bifurcating tree $\tau$, with branch lengths ${\bm{\ell}}$, representing the evolutionary relationships amongst this collection of $N$ taxa. For every site, phylogenetic models typically assume that evolution along each branch of the tree can be modelled using a continuous time Markov process $Y(t)$, characterised by an instantaneous rate matrix ${Q}=(q_{u,v})$ which has positive off-diagonal elements and rows that sum to zero. This matrix controls the dynamics of the substitution process through the matrix equation ${P}(\ell) = \{ p_{u,v}(\ell) \} = \exp (\ell {Q})$, where $p_{u,v}(\ell) = \Pr(Y(\ell) = v {|}Y(0) = u)$ for $u,v=1,\ldots,K$ is the probability of transitioning from character $u$ to character $v$ along a branch of length $\ell$. Standard phylogenetic models assume that the underlying continuous time Markov process is time reversible and in its stationary distribution\ ${\bm{\pi}} = (\pi_1,\ldots,\pi_K) \in \mathscr{S}_K$ where $\mathscr{S}_K = \{(x_{1}, \ldots, x_{K}): x_i \geq 0 \; \forall \, i, \sum x_i = 1\}$ denotes the $K$-dimensional simplex. Reversibility implies that $\pi_{u} p_{u,v}(\ell) = \pi_{v} p_{v,u}(\ell)$ for all $u,v$ and allows the rate matrix to be represented in the form ${Q} = {S} {\Pi}$, where ${\Pi} = \text{diag}({\bm{\pi}})$, and ${S}$ is a symmetric matrix whose off-diagonal elements, $\rho_{u,v}$ with $\rho_{u,v}=\rho_{v,u}$, are termed *exchangeability* parameters. The latter determine the general propensity for change between the different pairs of characters. We define a rate matrix as reversible if it permits a parameterisation of this form. The most general reversible rate matrix, with $K(K-1)/2$ distinct exchangeabilities, characterises the general time-reversible (GTR) model. Other commonly used substitution models are special cases. For example, the TN93 model is a special case for nucleotide data where $\rho_{C,T} = \rho_{T,C} = \rho_1$, $\rho_{G,A} = \rho_{A,G} = \rho_2$ and all other $\rho_{u,v}$ are equal to $\beta$. This simplification reduces the number of exchangeabilities from six to three but retains biological realism by allowing transversions (substitutions between a pyrimidine and a purine) and the two types of transitions (substitutions between pyrimidines and between purines) to occur at different rates, here $\beta$, $\rho_1$ and $\rho_2$ respectively. Classically, the sites of the alignment ${y}$ are assumed to evolve independently of each other and so the likelihood is given by $$p({y} | {Q}, \tau, {\bm{\ell}}) = \prod_{j=1}^M \Pr({\bm{Y}}_j = {\bm{y}}_j | {Q}, \tau, {\bm{\ell}}).$$ In order to prevent compensatory rescaling of the branch lengths $\ell$ and the rate matrix ${Q}$ in the transition matrix ${P}(\ell) = \exp (\ell {Q})$ it is common to impose an identifiability constraint on the rate matrix, for example by fixing one of the exchangeability parameters $\rho_{u,v}$, $u \ne v$, to be equal to one. For instance, one can fix $\beta=1$ in the TN93 model. This allows the remaining exchangeability parameters to be interpreted as relative rates of change. \[subsec:asrh\]Modelling rate heterogeneity across sites -------------------------------------------------------- It has long been recognised that selective pressures vary across sites due to the their differing roles in the structure and function of the molecular sequence. This feature is typically captured by allowing each site $j$ to evolve at its own rate $c_j > 0$ which scales the rate matrix $Q$ linearly. To enable information to be shared between sites, the rates ${\bm{c}}=(c_1,\ldots,c_M){^\mathrm{T}}$ are generally assumed follow a gamma distribution with unit mean. The likelihood can then be represented as $$p({y} | {Q}, \tau, {\bm{\ell}}, \alpha) = \prod_{j=1}^M \int_0^{\infty} p(c_j {|}\alpha) \Pr({\bm{Y}}_j = {\bm{y}}_j | {Q}_j, \tau, {\bm{\ell}}) \, d c_j,$$ where $$\label{eq:Q_j_lin} {Q}_j = c_j {Q}$$ and $p(c_j {|}\alpha)$ is the $\mathrm{Gam}(\alpha, \alpha)$ density function evaluated at $c_j$. The single parameter $\alpha$ determines the manner and extent to which the scaling factors differ across sites. We refer to models in which a baseline rate matrix is transformed according to  as *linear across site heterogeneity* (LASH) models. In order to simplify computation, the (continuous) gamma density $p(c_j {|}\alpha)$ is typically replaced by a discrete approximation with $K_c$ categories, most often $K_c=4$ [@Yan94]. In a Bayesian setting, this numerical integration strategy may seem less natural than using data augmentation during MCMC and sampling the $c_j$. However, the discretisation allows much more caching of intermediate likelihood calculations which can substantially speed up computational inference. In this model, the rate matrix at each site is simply a linearly scaled version of some underlying base matrix ${Q}$. The transformation does not affect the theoretical stationary distribution, defined as the solution of ${\bm{\pi}} {Q} = {\bm{0}}{^\mathrm{T}}$, or, in the class of reversible models, the ratios of the exchangeability parameters. In the following section we generalise this model to allow the rate matrix at each site to be a *quadratic* function of the base matrix, which depends on the values of *two* parameters. This allows the patterns of substitution, as well as the overall substitution rate, to vary between sites. \[sec:quash\]Quadratic across site heterogeneity (QuASH) models =============================================================== Consider a baseline rate matrix $Q$. At site $j$, the instantaneous rate matrix $Q_j = (q_{j,u,v})$ is given by $$\label{eq:Q_j} Q_j = c_j Q - c_j d_j Q^2$$ where $c_j \in (0, \infty)$ and $d_j \in (l(Q), u(Q))$, which reduces to the simple LASH model when $d_j=0$. We call any model in which a baseline rate matrix is transformed in this way a *quadratic across site heterogeneity* (QuASH) model. The limits $l(Q)$ and $u(Q)$ depend on $Q$ and ensure that $Q_j$ is a valid rate matrix, that is (i) all off-diagonal elements are positive: $q_{j,u,v} > 0$, $\forall u \ne v$; (ii) all row sums are zero: $\sum_{v} q_{j,u,v} = 0$ $\forall u$. Property (ii) is automatically satisfied for any $d_j \in \mathbb{R}$. The proof is as follows. The $(u,v)$-th element of $Q_j$ is given by $$q_{j,u,v} = c_j \left(q_{u,v} - d_j \sum_{w} q_{u,w} q_{w,v} \right).$$ Therefore the sum of the elements on row $u$ of $Q_j$ is $$\begin{aligned} \sum_{v} q_{j,u,v} &= c_j \left( \sum_{v} q_{u,v} - d_j \sum_{v} \sum_{w} q_{u,w} q_{w,v} \right) = c_j \left( 0 - d_j \sum_{w} q_{u,w} \sum_{v} q_{w,v} \right)\\ &= c_j (0 - d_j \times 0) = 0\end{aligned}$$ for any $d_j \in \mathbb{R}$. For property (i) to be satisfied we need $$\label{eq:lower} l(Q) = \max\{ \mathcal{L}(Q) \}, \; \mathcal{L}(Q) = \left\{ \frac{q_{u,v}}{\sum_{w} q_{u,w} q_{w,v}} \, : u \ne v \text{ \& } \sum_{w} q_{u,w} q_{w,v} < 0 \right\}$$ and $$\label{eq:upper} u(Q) = \min\{ \{ \infty \} \cap \mathcal{U}(Q) \}, \; \mathcal{U}(Q) = \left\{ \frac{q_{u,v}}{\sum_{w} q_{u,w} q_{w,v}} \, : u \ne v \text{ \& } \sum_{w} q_{u,w} q_{w,v} > 0\right\}.$$ By definition, $l(Q) \le 0$ and $u(Q) \ge 0$. Note that the set $\mathcal{L}(Q)$ cannot be empty, that is, $\mathcal{L}(Q) \ne \emptyset$. To prove this, suppose that $q_{a,b}$ is the largest off-diagonal element in $Q$. Now $$\begin{aligned} \sum_{w} q_{a,w} q_{w,b} &= q_{a,a} q_{a,b} + q_{a,b} q_{b,b} + \sum_{w \ne a,b} q_{a,w} q_{w,b}\\ &= - q_{a,b} \sum_{w \ne a} q_{a,w} + q_{a,b} q_{b,b} + \sum_{w \ne a,b} q_{a,w} q_{w,b}\\ &= - q_{a,b} \sum_{w \ne a,b} q_{a,w} - q_{a,b}^2 + q_{a,b} q_{b,b} + \sum_{w \ne a,b} q_{a,w} q_{w,b}.\end{aligned}$$ However, $q_{w,b} < q_{a,b}$ for all $w \ne a$ and so $$\sum_{w \ne a,b} q_{a,w} q_{w,b} < q_{a,b} \sum_{w \ne a,b} q_{a,w}.$$ Because $- q_{a,b}^2$ and $q_{a,b} q_{b,b}$ are strictly negative it follows that $$\sum_{w} q_{a,w} q_{w,b} = - q_{a,b} \sum_{w \ne a,b} q_{a,w} + \sum_{w \ne a,b} q_{a,w} q_{w,b} - q_{a,b}^2 + q_{a,b} q_{b,b} < 0.$$ In contrast, the set $\mathcal{U}(Q)$ can be empty. Consider, for example, the rate matrix of the Jukes Cantor model, all of whose off-diagonal elements are $\delta>0$. In this case, $\sum_{w} q_{u,w} q_{w,v} = -4\delta^2 < 0$ for all pairs $(u,v)$ with $u \ne v$. Therefore $l(Q) = - 1 / (4 \delta)$ whilst the upper limit $u(Q)$ is infinite. To allow information to be shared between sites, we continue to assume that the coefficients ${\bm{c}}=(c_1,\ldots,c_M){^\mathrm{T}}$ of the linear term are conditionally independent and identically distributed (i.i.d.) with $c_j | \alpha \sim {\mathrm{Gam}}(\alpha, \alpha)$ for some unknown hyperparameter $\alpha$. In an analogous fashion, we assume that the coefficients ${\bm{d}}=(d_1,\ldots,d_M){^\mathrm{T}}$ of the second order term are independent of ${\bm{c}}$ and conditionally i.i.d. with $d_j | Q, \beta \sim \mathcal{F}(\beta)$ for some unknown $\beta$, where the form of the distribution $\mathcal{F}$ will be discussed in Section \[subsec:random\_effects\]. The likelihood can then be represented as $$p({y} | {Q}, \tau, {\bm{\ell}}, \alpha, \beta) = \prod_{j=1}^M \int_0^{\infty} \int_{l(Q)}^{u(Q)} p(c_j {|}\alpha) p(d_j {|}Q, \beta) \Pr({\bm{Y}}_j = {\bm{y}}_j | {Q}_j, \tau, {\bm{\ell}}) \, d c_j \, d d_j$$ where ${Q}_j$ was defined in . As with the simpler LASH model, substantial gains in computational efficiency can be achieved by replacing the continuous densities $p(c_j {|}\alpha)$ and $p(d_j {|}Q, \beta)$ by discrete approximations with $K_c$ and $K_d$ categories, respectively. We choose to place point masses of probability $1/(K_c K_d)$ at locations $\{ z_{c,a}(\alpha), z_{d,a'}(Q,\beta) \}$ for $a = 1, \ldots, K_c$, $a' = 1, \ldots, K_d$ where $z_{c,a}(\alpha)$ is the $(a - 0.5)/K_c$ quantile in the distribution of $c_j | \alpha$ and $z_{d,a'}(Q,\beta)$ is the $(a' - 0.5)/K_d$ quantile in the distribution of $d_j {|}Q, \beta$. The likelihood then simplifies to $$\begin{aligned} p({y} &| {Q}, \tau, {\bm{\ell}}, \alpha, \beta) \notag \\ &\simeq \frac{1}{K_c K_d} \prod_{j=1}^M \sum_{a=1}^{K_c} \sum_{a'=1}^{K_d} \Pr\left[ {\bm{Y}}_j = {\bm{y}}_j | {Q}_j\left\{ z_{c,a}(\alpha), z_{d,a'}(Q,\beta), Q \right\}, \tau, {\bm{\ell}} \right].\label{eq:likelihood_quash1}\end{aligned}$$ \[subsec:quash\_props\]Properties of QuASH Models ------------------------------------------------- It can easily be shown that the stationary distribution of $Q_j = c_j Q - c_j d_j Q^2$ is the same as that of $Q$; see Appendix A for a proof. Of course the same is also true under the simple linear scaling, ${Q}_j = c_j {Q}$, which we recover when $d_j = 0$. In the latter case, the linear mapping can simply be regarded as a site-specific scaling of the branch lengths. In contrast, our quadratic transformation does not preserve the ratios of the instantaneous rates of change in the baseline rate matrix, allowing different patterns of substitution at different sites. This idea is most readily exemplified in the context of reversible models where the transformation results in a site-heterogeneous model in which the exchangeability parameters vary across sites. Elucidating further, it is straightforward to show that if $Q$ is reversible, then so is $Q_j$; see Appendix A for a proof. It follows that the set of GTR rate matrices is closed under our quadratic transformation. This is also true for some special cases of the GTR rate matrix including the TN93 rate matrix which was introduced in Section \[sec:models\]. In this case, suppose that $\beta$, $\rho_1$ and $\rho_2$ are the transversion and transition rates in the baseline rate matrix and that ${\bm{\pi}}=(\pi_A , \pi_G, \pi_C , \pi_T)$ is the associated stationary distribution. After applying the quadratic transformation , it follows from  in Appendix A that the transversion and transition rates in the rate matrix for site $j$ are $$\begin{aligned} \beta_j&=c_j \beta (1 + d_j \beta),\\ \rho_{1,j}&=c_j [ \rho_1 + d_j \{ \rho_1^2 - (\rho_1 - \beta)^2 \pi_R\}],\quad \rho_{2,j}=c_j [ \rho_2 + d_j \{ \rho_2^2 - (\rho_2 - \beta)^2\pi_Y\}],\end{aligned}$$ where $\pi_R = \pi_A + \pi_G$ and $\pi_Y = \pi_C + \pi_T$. As remarked in Section \[sec:models\], for the simpler LASH model with all $d_j = 0$, we often fix one of the exchangeability parameters in the baseline rate matrix $Q$ to be equal to one for parameter identifiability. For a QuASH model, scaling the baseline rate matrix $Q$ by a constant $k>0$ can no longer be compensated by scaling all branch lengths by $1/k$ and so an identifiability constraint is not strictly required. However, for parameter interpretability and to preserve the nested structure of the LASH and QuASH models, we continue to fix the scale of baseline rate matrix. If we take the distribution at the root of the tree to be the vector ${\bm{\pi}}$ satisfying ${\bm{\pi}} Q = {\bm{0}}{^\mathrm{T}}$ then the resulting Markov process is stationary and the term $\Pr({\bm{Y}}_j = {\bm{y}}_j | {Q}_j, \tau, {\bm{\ell}})$ in the likelihood  is given by $$\label{eq:likelihood_quash2} \Pr({\bm{Y}}_j = {\bm{y}}_j | {Q}_j, \tau, {\bm{\ell}}) = \sum_{X} \pi_{X(0)} \prod_{\text{edges} \hspace*{3.0pt} b = (v,w)} p_{j, X(v), X(w)}(\ell_b).$$ Here $v$ and $w$ are the vertices at the two ends of edge $b$ with length $\ell_b$, $X(u)$ is the character at vertex $u$, $u=0$ denotes the root vertex and $P_j(\ell)=\{ p_{j,u,u'}(\ell) \}$ is the transition matrix associated with an edge of length $\ell$ at site $j$. The sum is over all functions $X$ from the vertices to $\Omega_K$ such that $X(u)$ matches the data $y_j(u)$ for all leaf vertices $u$. \[subsec:random\_effects\]Random Effect Distribution ---------------------------------------------------- We model the coefficients ${\bm{d}}=(d_1,\ldots,d_M){^\mathrm{T}}$ involved in the second order term of the quadratic transformation  as conditionally i.i.d. with $d_j | Q, \beta \sim \mathcal{F}(\beta)$ for some unknown hyperparameter $\beta$. As explained earlier in this section, the distribution $\mathcal{F}$ has support on $(l(Q),u(Q))$ where $l(Q)$ is nonpositive but assumed finite whilst $u(Q)$ is nonnegative but can be infinite. This means the interval $(l(Q),u(Q))$ can be finite or semi-infinite. In order to handle the two cases in a consistent fashion, we construct the distribution of $d_j$ through a shifted, piecewise power transformation of a Beta random variable $$d_j = \begin{cases} l(Q)+w(Q)\left(1 - b_j^{1/w(Q)}\right), \quad &\text{if $u(Q)$ is finite,}\\ l(Q) - \log b_j, &\text{otherwise,} \end{cases}$$ where $w(Q)=u(Q)-l(Q)$; $b_j | Q, \beta \sim {\mathrm{Beta}}[ \beta+a(Q), \beta\{b(Q)-1\}+1 ]$; and $\beta > 0$ is unknown. The terms $a(Q)$ and $b(Q)$ depend on the baseline rate matrix $Q$ through $a(Q)=1/w(Q)$ if $u(Q)$ is finite and $a(Q)=0$ otherwise, and $b(Q)=\{w(Q)/u(Q)\}^{w(Q)}$ if $u(Q)$ is finite and $b(Q)=e^{-l(Q)}$ otherwise. This choice ensures that the mode of the distribution is zero, with finite probability density, and that the density of $d_j$ decays smoothly to zero at its end points, except in the case where $l(Q)=0$ or $u(Q)=0$. In the special case when $l(Q)=0$ and $u(Q)$ is infinite, the conditional distribution of $d_j$ reduces to the ${\mathrm{Exp}}(\beta)$ distribution. By centring the distribution on zero, we encourage shrinkage towards the nested LASH model with all $d_j = 0$. Although it may appear more natural to set the mean or median, rather than the mode, to zero, since the lower or upper end points of the support can be equal to zero, this is not possible in the general case. The hyperparameter $\beta$ can be assigned any prior with support on the positive real line. The dependence of the marginal prior for $d_j$ on that for $\beta$ and the parameters of the baseline rate matrix $Q$ is complex. However, closed form expressions for the conditional expectation and variance of $d_j$ given $\beta$, and bounds $l$ and $u$ can be computed and are given in Appendix B. For various values of $l$ and $u$ spanning the range inferred in analyses of real data, Figure \[fig:dj\_prior\] plots the conditional mean and standard deviation as a function of $\beta$. Clearly as $\beta$ gets large, the distribution of $d_j$ tends towards a point mass at zero and we recover a simple LASH model. However, as $\beta$ approaches zero, the mean and standard deviation both become large. Therefore we can allow more heterogeneity across sites by giving $\beta$ a prior which assigns reasonable density around zero. ![Conditional mean and standard deviation of $d_j$ given $\beta$, and bounds $l$ and $u$, plotted with a log-scale on the $y$-axis.[]{data-label="fig:dj_prior"}](dj_prior.pdf) \[sec:ns-quash\]Non-stationary models ===================================== The transformations characterising LASH and QuASH models allow across-site variation in the overall magnitude of the instantaneous rates of change and, for QuASH models, their relative sizes. However, the models discussed so far have been homogeneous across branches, with a single baseline rate matrix $Q$ applying to the whole tree. Furthermore, the linear and quadratic transformations  and  preserve the stationary distribution ${\bm{\pi}}$ of $Q$. Therefore if the distribution at the root of the tree ${\bm{\pi}}_{(0)}$ is equal to ${\bm{\pi}}$, then the resulting Markov process will assume the same stationary distribution at all sites. These models cannot, therefore, explain the heterogeneities in sequence composition that are commonly observed in experimental data, either across taxa or across sites. As explained in Section \[sec:intro\], the resulting model misspecification can lead to misleading phylogenetic inferences. Non-stationary models for sequence evolution can account for differences in composition across taxa by allowing the probability of being in each state (e.g. each nucleotide for DNA data) to change over time. Typically this is achieved by permitting step changes in the theoretical stationary distribution at different points on the tree. Although these changes do not have to occur at speciation events [e.g. @BL06], this assumption is often made [e.g. @YR95; @Fos04; @HNBWE14; @Che16] and we retain it here for simplicity of notation. In general, therefore, consider a rooted topology $\tau$ with $B$ branches and a model which assumes a distribution ${\bm{\pi}}_{(0)}$ at the root of the tree, with the processes on the other branches governed by rate matrices $Q_{(1)},\ldots,Q_{(B)}$, with associated theoretical stationary distributions ${\bm{\pi}}_{(1)},\ldots,{\bm{\pi}}_{(B)}$. To achieve non-stationarity we need ${\bm{\pi}}_{(b)} \ne {\bm{\pi}}_{(0)}$ for at least one $b \in \{1,\ldots,B\}$, however for some $b \ne b'$, we might fix ${\bm{\pi}}_{(b)}$ to be equal to ${\bm{\pi}}_{(b')}$. Extending the LASH and QuASH transformations to non-stationary models of this form, the rate matrix for site $j$ on branch $b$ is given by $$Q_{b,j} = c_j Q_{(b)} - c_j d_j Q_{(b)}^2$$ where $c_j \in (0, \infty)$, whilst $d_j=0$ for LASH models and $d_j \in (l, u)$ for QuASH models. In the latter case, the limits depend on all the $Q_{(b)}$, with $l = \max \{ l(Q_{(b)}): b = 1,\ldots, B\}$ and $u = \min \{ u(Q_{(b)}): b = 1,\ldots, B\}$, where $l(\cdot)$ and $u(\cdot)$ are as in  and  respectively. This ensures that all the resulting $Q_{b,j}$ are valid rate matrices. The likelihood expressions  and  for stationary QuASH models can now be modified to give $$\begin{aligned} p({y} &| {Q}_{(1)},\ldots,{Q}_{(B)},{\bm{\pi}}_{(0)}, \tau, {\bm{\ell}}, \alpha, \beta) \notag \\ &\simeq \frac{1}{K_c K_d} \prod_{j=1}^M \sum_{a=1}^{K_c} \sum_{a'=1}^{K_d} \Pr({\bm{Y}}_j = {\bm{y}}_j | {Q}_{1,j},\ldots,{Q}_{B,j},{\bm{\pi}}_{(0)}, \tau, {\bm{\ell}})\label{eq:likelihood_nsquash1}\end{aligned}$$ where ${Q}_{b,j} = {Q}_{b,j} \left\{ z_{c,a}(\alpha), z_{d,a'}({Q}_{(1)},\ldots,{Q}_{(B)},\beta), {Q}_{(b)} \right\}$ and $$\label{eq:likelihood_nsquash2} \Pr({\bm{Y}}_j = {\bm{y}}_j | {Q}_{1,j},\ldots,{Q}_{B,j},{\bm{\pi}}_{(0)}, \tau, {\bm{\ell}}) = \sum_{X} \pi_{(0),X(0)} \prod_{\text{edges} \hspace*{3.0pt} b = (v,w)} p_{b, j, X(v), X(w)}(\ell_b)$$ in which $P_{b,j}(\ell_b) = \{ p_{b, j, h,i}(\ell_b) \} = \exp (\ell_b {Q}_{b,j})$ is the transition matrix associated with edge $b$, of length $\ell_b$, and site $j$. By definition, non-stationary LASH and QuASH models allow heterogeneities in sequence composition across taxa. However, they also allow heterogeneities across sites. Consider, for example, a simple non-stationary model which allows a single step change in the stationary distribution at the root of the tree [e.g. @Che16 Chapter 4]. In such a model, a single baseline rate matrix $Q_{(1)}$, with associated stationary distribution ${\bm{\pi}}_{(1)} \not \equiv {\bm{\pi}}_{(0)}$, applies to all branches of the tree, and we denote the rate matrix associated with site $j$ by $Q_{1,j}$. If $\lambda$ is an eigenvalue of $Q_{(1)}$, it follows immediately from  that $c_j \lambda - c_j d_j \lambda^2$ is an eigenvalue of $Q_{1,j}$, with $d_j=0$ for LASH models. Denote by $\lambda_{j,1},\lambda_{j,2},\ldots,\lambda_{j,K}$ the eigenvalues of $Q_{1,j}$ ordered such that $\lambda_{j,1} = 0 > \mathrm{Re}(\lambda_{j,2}) \ge \mathrm{Re}(\lambda_{j,3}) \ge \cdots \ge \mathrm{Re}(\lambda_{j,K})$, where $\mathrm{Re}(\lambda)$ denotes the real part of the complex number $\lambda$. Under this model, it can be shown that ${P}_j(\ell) = {\bm{1}} {\bm{\pi}}_{(1)} + O(e^{-\nu_j \ell})$ as $\ell \to \infty$ where ${\bm{1}}$ is a length $K$ column vector of 1s and $\nu_j = - \mathrm{Re}(\lambda_{j,2})$; see, for example, @Kij97, Chapter 4. It follows that at sites for which $\nu_j$ is large, the rate of convergence towards the stationary distribution ${\bm{\pi}}_{(1)}$ associated with $Q_{(1)}$ will be fast, giving rise to marginal distributions at the leaves of the tree that resemble ${\bm{\pi}}_{(1)}$. In contrast, at sites for which $\nu_j$ is small, the rate of convergence will be slow, leading to marginal distributions at the leaves that are closer to the distribution at the root ${\bm{\pi}}_{(0)}$. Clearly these effects will vary across taxa according to the overall distance from the root to the different leaves. Although LASH and QuASH models both allow this kind of behaviour, in QuASH models it is managed more flexibly by two parameters, rather than one. Further, as discussed in Section \[subsec:quash\_props\], only the QuASH mapping allows the ratios of the instantaneous rates of change, and hence transition patterns, to vary across sites. In the application in Section \[subsec:apps\_nonstat\], we focus on the HB model [@HNBWE14] where each branch of the tree has its own reversible rate matrix $Q_{(b)}$ which factorises into a composition vector ${\bm{\pi}}_{(b)}$ and a set of exchangeability parameters ${\bm{\rho}}$ that are assumed constant across the tree. We use the formulation of the model from @WHCNBE15 in which the composition vector on the root edge of the underlying unrooted topology is the same as that at the root of the tree ${\bm{\pi}}_{(0)}$. To allow information to be shared between branches, the composition vectors $\{ {\bm{\pi}}_{(b)} \}$ are positively correlated *a priori*. Full details can be found in the description of Prior B in @HNBWE14 but, briefly, a greater exchange of information between neighbouring branches is admitted by adopting a first order autoregressive structure in which the composition vector on branch $b$ is conditionally independent of the composition vectors on all non-descendant branches given its parent. \[sec:inference\]Posterior inference via MCMC ============================================= Let ${\bm{\theta}}$ represent the parameters of the distribution at the root of the tree and the set of baseline rate matrices. For example, ${\bm{\theta}}=\{ {\bm{\pi}}, {\bm{\rho}} \}$ for a simple, stationary QuASH model based on a reversible rate matrix, or ${\bm{\theta}}=\{ {\bm{\pi}}_{(0)},\ldots,{\bm{\pi}}_{(B-2)}, {\bm{\rho}} \}$ for the QuASH variant of the HB model. The joint posterior distribution for all unknowns is given by $$\pi({\bm{\theta}}, \tau, {\bm{\ell}}, \alpha, \beta | {y}) \propto p({y} | {\bm{\theta}}, \tau, {\bm{\ell}}, \alpha, \beta)\, \pi({\bm{\theta}}, \tau, {\bm{\ell}}, \alpha, \beta)$$ where the likelihood function $p({y} | {\bm{\theta}}, \tau, {\bm{\ell}}, \alpha, \beta)$ was given in  and  for a simple, stationary QuASH model, or in  and  for a non-stationary QuASH model. Irrespective of the choice of prior distribution $\pi({\bm{\theta}}, \tau, {\bm{\ell}}, \alpha, \beta)$, the posterior is analytically intractable. We therefore build up a numerical approximation using a Metropolis within Gibbs sampling scheme which iterates through a series of updates for each unknown. Real valued parameters, such as branch lengths ${\bm{\ell}}$, can be updated using standard proposal distributions, for example Gaussian random walks on the log-scale. In QuASH models whose likelihood is invariant to the root position, $\tau$ represents an unrooted topology which can be updated using standard topological moves such as nearest neighbour interchange (NNI) and subtree prune and regraft (SPR); see, for example, @mrbayes. For QuASH models whose likelihood depends on the root position, $\tau$ represents a rooted topology and so proposals which attempt to move the root are also required. In the applications in Section \[sec:apps\], for example, we consider the QuASH variant of the HB model and employ the NNI, SPR and root moves described in @HNBWE14. \[sec:apps\]Applications ======================== A controversial issue in evolutionary biology concerns the structure of the *tree of life*, whose phylogeny represents the relationships amongst its three main domains: Bacteria, Archaea and eukaryotes. There are two dominant hypotheses for the underlying unrooted topology. The classic *three domains hypothesis* of @WKW90 posits that the three domains are *monophyletic*, meaning each has an ancestor that is not shared by the others. On the basis of analyses involving previously unsequenced taxa and more sophisticated evolutionary models [@WFCE13], an alternative view – the *eocyte hypothesis* – has gained considerable support over recent years. According to this conjecture, the eukaryotes emerge from within a paraphyletic Archaea, meaning the most recent common ancestor of eukaryotes and Archaea was an Archaeon. In addition to uncertainty surrounding the unrooted topology of the tree of life, opinion is also divided on the position of its root. Under the two leading hypotheses, the root is either placed on the bacterial branch [@GKDTBBMPDO89; @IKHOM89] or, with fewer proponents, within the Bacteria [@Cav06; @LSHS09]. In this section we consider applications to biological data sets that address these controversial questions. In Section \[subsec:apps\_stat\] we analyse a concatenated alignment of small and large subunit ribosomal RNAs (SSU and LSU rRNAs) sampled from across the tree of life. After alignment using MUSCLE [@Edg04] and editing to remove poorly-aligning regions, $M=1734$ sites on $N=36$ species remained. We consider three models: (S1) a stationary, reversible TN93 model, (S2) the LASH-variant of S1 and (S3) the QuASH-variant of S1. Models that are stationary and reversible give rise to likelihood functions that are invariant to the position of the root, and so these analyses only allow inference of the unrooted topology. In Section \[subsec:apps\_nonstat\] we therefore consider three non-stationary models which also allow us to learn about the root position: (NS1) the HB model with TN93 exchangeability parameters; (NS2) the LASH-variant of NS1 and (NS3) the QuASH-variant of NS1. Inference via MCMC is substantially slower for the HB model and so, for computational tractability, we consider a smaller data set with $M=1481$ sites and only $N=16$ taxa. In all analyses, mixing and convergence of the MCMC sampler was assessed by comparing the output from multiple chains, initialised at different starting points. In phlogenetics, mixing in tree-space can be problematic due to the low acceptance rates of topological moves. Therefore, in addition to considering the usual numerical and graphical diagnostic checks for continuous parameters, we also examined graphs based on relative cumulative split (Section \[subsec:apps\_stat\]) or clade (Section \[subsec:apps\_nonstat\]) frequencies of the chains over the course of the MCMC runs; see @HNBWE14 for a full description of these diagnostics. Here a *split* refers to a bipartition of the taxa at the leaves of the tree into two disjoint sets, induced by cutting a branch. On a rooted tree, one of the partition subsets of any split is a *clade* if all the taxa lie on the same side of the root. In biological terms, this corresponds to an ancestor and all its descendants. \[subsec:apps\_stat\]Stationary TN93 model ------------------------------------------ Based on our subjective assessments of the evolutionary process, for the parameters of the S1 model we chose independent gamma $\mathrm{Gam}(1, 1)$ priors for the two transition rates $\rho_1$ and $\rho_2$, a flat Dirichlet $\mathscr{D}(1,1,1,1)$ prior for the stationary distribution ${\bm{\pi}}$, independent exponential $\mathrm{Exp}(10)$ priors for the branch lengths ${\bm{\ell}}$ and a uniform prior over unrooted topologies $\tau$. In models S2 and S3 we additionally assigned a gamma $\mathrm{Gam}(10, 10)$ prior to the shape parameter $\alpha$ in the random effects distribution for the rates $c_j$ and, in model S3, a gamma $\mathrm{Gam}(1, 1)$ prior to the parameter $\beta$ in the random effects distribution for the quadratic coefficients $d_j$ of the QuASH model. The latter distribution, with mean ${\mathrm{E}}(\beta)=1$ and coefficient of variation $\mathrm{CV}(\beta)=1$, was chosen to give reasonable support to values of $\beta$ near zero. As explained in Section \[subsec:random\_effects\], this choice makes the prior for the $d_j$ reasonably diffuse. In order to check sensitivity to the prior specification for $\beta$, we repeated the analysis with model S3 using priors that had the same mean but different coefficients of variation and different behaviour near zero: $\mathrm{Gam}(10, 10)$ ($CV(\beta) = 0.316$) and $\mathrm{Gam}(0.1, 0.1)$ ($CV(\beta) = 3.16$). The phylogenetic and posterior predictive inferences reported in this section were robust against these changes. For each model the MCMC algorithm outlined in Section \[sec:inference\] was used to generate at least $110K$ draws from the posterior, after a burn-in of $100K$ samples, thinning the remaining output to retain every 100th iterate. The diagnostics checks described earlier gave no evidence of any lack of convergence. \ In phylogenetic inference, the majority-rule consensus tree is the most widely used summary of the posterior distribution over tree space. As a summary of a sample of trees, it includes only those splits which appear in over half of the samples [@Bry03], here representing those with posterior probability greater than 0.5. For the analyses under models S1, S2 and S3, the consensus trees are shown in Figure \[fig:s\_consensus\] in which the numerical labels represent the posterior probability of the associated split. To aid comparison, the trees are all visualised with the root at the midpoint of the bacterial branch. The consensus tree under S1 supports the three domains hypothesis, whilst models S2 and S3 yield eocyte trees, with eukaryotes emerging from within two archaeal clades: the Euryarchaeota and the TACK Archaea. As expected, there is a marked difference in our phylogenetic inferences as we move from the simple TN93 model (S1) to one which incorporates across-site rate heterogeneity. However, there is very little difference in the inferences obtained when extending the LASH model (S2) to the corresponding QuASH model (S3). Comparing the prior and posterior density for $\beta$ in Figure \[fig:s\_densbeta\], the posterior seems to support larger values for $\beta$ than the prior, which suggests a distribution for the quadratic coefficients $d_j$ that is more concentrated around zero. The data do not, therefore, provide much evidence that the QuASH transformation is necessary given a model that already incorporates across-site rate heterogeneity. ![Marginal prior and posterior densities for the unknown parameter $\beta$ in the random effect distribution for the quadratic coefficients $d_j$.[]{data-label="fig:s_densbeta"}](s_beta.pdf) As explained in Section \[sec:intro\], functional and structural constraints acting on a particular site can cause it to evolve very slowly. In such cases we are likely to see little or no variation in the character state at that column of the alignment. Therefore in fitting to the alignment-wide empirical compositions, models that do not allow variation in, at least, the rate of the evolutionary process across sites tend to overestimate the mean number of distinct nucleotides per column, and underestimate the associated standard deviation. Figure \[fig:s\_postpred\_bs\] shows the posterior predictive distribution for these test statistics obtained under models S1, S2 and S3, together with the observed values calculated from the alignment. As expected, model S1 markedly overestimates the number of distinct nucleotides per site and underestimates the associated standard deviation. Whilst models S2 and S3 also overestimate the mean, the discrepancies are much less marked, with the QuASH-variant of the TN93 model (S3) being most compatible with the observed data. Interestingly, models S2 and S3 overestimate the standard deviation of the number of distinct nucleotides per site, slightly more noticeably for model S2 than S3. It is possible that models allowing sequence composition to vary across sites would be required to adequately capture this feature. ![Posterior predictive densities for the mean and standard deviation of the number of distinct nucleotides per site. The observed values are indicated by vertical lines.[]{data-label="fig:s_postpred_bs"}](s_postpred.pdf) \[subsec:apps\_nonstat\]Non-stationary HB model ----------------------------------------------- For the analyses using the non-stationary models NS1, NS2 and NS3, we adopted the prior distributions outlined in Section \[subsec:apps\_stat\] for the two transition rates $\rho_1$ and $\rho_2$, the branch lengths ${\bm{\ell}}$ and the parameters $\alpha$ and $\beta$ in the random effects distributions for the linear and quadratic coefficients $c_j$ and $d_j$. As the HB model yields a likelihood function that depends on the position of the root, our topology $\tau$ is rooted. We assigned $\tau$ a prior according to the biologically-motivated Yule model of speciation, which generates a distribution in which near equal probability is assigned to root splits of all sizes: $1 \!\! : \!\! N-1$, $2 \!\! : \! \! N-2$, and so on [@CHNBWE15]. For the composition vectors ${\bm{\pi}}_b$, $b=0,\ldots,B-2$, in the baseline rate matrix we used Prior B from @HNBWE14, choosing the hyperparameters representing the autoregressive coefficient and conditional variance to be $a=0.94$ and $b=0.31$ respectively. This specification was guided by simulations from the prior predictive distribution which suggested it led to a biologically plausible degree of heterogeneity in empirical sequence composition. For each model the MCMC algorithm was used to generate at least $510K$ draws from the posterior, after a burn-in of $500K$ samples, thinning the remaining output to retain every 100th iterate. The diagnostics checks described earlier gave no evidence of any lack of convergence. \ \ The rooted majority-rule consensus trees for each model are shown in Figure \[fig:ns\_consensus\]. Our conclusions are consistent with those from Section \[subsec:apps\_stat\]. Specifically, the model NS1 supports a three-domains tree whilst models NS2 and NS3 support very similar eocyte trees with, in this case, the same rooted topology. Although the site-homogeneous HB model (NS1) and the LASH and QuASH variants (NS2 and NS3) support different conclusions with regards to the unrooted topology, they both suggest a root within the Bacteria. The marginal posterior distribution for root splits under the three models is summarised in Table \[tab:ns\_post\_root\]. Again, the differences between the inferences under NS1 and NS2 are much more marked than the differences between those under NS2 and NS3. However, in all cases the posterior probability for a root within the Bacteria is 1.0. -------------------------------------------------------------------------------------------------------------------------------- ------- ------- ------- Root Split NS1 NS2 NS3 *Rhodopirellula baltica* 0.576 0.017 0.015 *Campylobacter jejuni*, *Chlamydia trachomatis*, *Clostridium acetobutylicum*, *Escherichia coli*, *Synechocystis sp. PCC6803* 0.153 0.067 0.088 *Clostridium acetobutylicum*, *Synechocystis sp. PCC6803* 0.085 0.324 0.343 *Synechocystis sp. PCC6803* 0.108 0.197 0.172 *Campylobacter jejuni*, *Chlamydia trachomatis*, *Escherichia coli* 0.040 0.176 0.189 *Campylobacter jejuni*, *Escherichia coli* 0.031 0.176 0.157 *Campylobacter jejuni*, *Clostridium acetobutylicum*, *Escherichia coli*, *Synechocystis sp. PCC6803* 0.004 0.018 0.014 *Clostridium acetobutylicum* 0.001 0.012 0.012 -------------------------------------------------------------------------------------------------------------------------------- ------- ------- ------- : Root splits receiving posterior support of at least 0.01 under the NS1, NS2 or NS3 models, and the associated probabilities. The root split is the bipartition induced by cutting the branch edge of the associated unrooted topology; the smaller partition is shown.[]{data-label="tab:ns_post_root"} The LASH and QuASH variants of the HB model allow sequence composition, as well as the overall rate of evolution, to vary across sites. Therefore we expect these models to be better equipped to capture the number of distinct nucleotides per site. Posterior predictive densities of the across-site mean and standard deviation are plotted in Figure \[fig:ns\_postpred\_bs\]. For the mean, all three models capture the observed statistic well, with the the site-homogeneous model (NS1) offering slightly more support to larger values, as expected. As in the analysis from Section \[subsec:apps\_stat\], the site-homogeneous model very markedly underestimates the standard deviation. The posterior predictive densities under the LASH (NS2) and QuASH (NS3) variants of the HB model are very similar. Although both overestimate the standard deviation, the observed statistic is more plausible than under the NS1 model, and the overestimation seems less marked than the corresponding comparison from Section \[subsec:apps\_stat\]. The similarity in both phylogenetic and posterior predictive inferences under the LASH and QuASH models are consistent with the implications of the comparison between the prior and posterior for $\beta$ in Figure \[fig:ns\_densbeta\]. Again, the posterior seems to support larger values of $\beta$ than the prior which suggests that the QuASH transformation adds little given a model in which linear across-site heterogeneity is already included. ![Posterior predictive densities for the mean and standard deviation of the number of distinct nucleotides per site. The observed values are indicated by vertical lines.[]{data-label="fig:ns_postpred_bs"}](ns_postpred.pdf) ![Marginal prior and posterior densities for the unknown parameter $\beta$ in the random effect distribution for the quadratic coefficients $d_j$.[]{data-label="fig:ns_densbeta"}](ns_beta.pdf) \[sec:discussion\]Discussion ============================ The introduction of across-site rate heterogeneity into substitution models for sequence evolution led to substantial improvements in model fit and the credibility of phylogenetic inferences. In practice, this feature was incorporated through a set of site-specific rates, modelled as random effects with a unit mean gamma distribution, that linearly transformed a baseline rate matrix. Motivated by the advancement gained through this simple innovation, we considered a natural extension based on the incorporation of two sets of random effects, allowing site-specific quadratic transformation of the baseline rate matrix. We derived properties of QuASH-transformed rate matrices, showing that they retain the stationary distribution of the underlying baseline matrix, and that the set of reversible rate matrices is closed under our quadratic transformation. In the context of a class of non-stationary models which permit step-changes in the theoretical stationary distribution at one or more points on the tree, we demonstrated that both the LASH and QuASH transformations lead to models which allow sequence composition to vary across sites as well as across taxa. This is due to different rates of convergence towards the theoretical stationary distributions at different sites. The QuASH-transformed, non-stationary models therefore provide a parsimonious means of allowing heterogeneity in sequence composition across both alignment dimensions. We utilised our model and inferential procedures in two biological applications concerning the tree of life. In the first, we compared inferences under a stationary, reversible TN93 model, with those obtained under the LASH and QuASH extensions. In the second, to make computational inference manageable, we considered a smaller data set and compared inferences under a non-stationary HB model to those obtained under the LASH and QuASH variants. In both applications we found that the simpler site-homogeneous models supported the three domains hypothesis, with the Archaea, Bacteria and eukaryotes appearing as monophyletic groups. Conversely the more flexible LASH and QuASH models supported the eocyte hypothesis, with eukaryotes emerging from within a paraphyletic Archaea. The non-stationary models consistently supported a root within the Bacteria. The marked differences between inferences obtained under the site-homogeneous and LASH models are congruent with other results reported in the literature [@Yan96]. However, neither analysis suggested that the quadratic transformation added much value once a linear transformation was in place. We have drawn similar conclusions from applications to several other data sets not reported here. Although our analyses have reinforced the importance of allowing heterogeneity in the rate of evolution across sites, it appears that the natural extension, exploiting a quadratic transformation of the base rate matrix, adds little value. However, in the context of non-stationary models, it is worth emphasising that even the LASH transformation generates models that allow heterogeneity in sequence composition across sites as well as across taxa. To our knowledge, this is a property that has gone unnoticed in the literature. Whilst a few, more mechanistic models have been proposed to offer this flexibility [e.g. @BL08; @JWRPJ14], their complexity has made model-fitting computationally prohibitive. In contrast, non-stationary LASH and QuASH models provide a more parsimonious, data-driven alternative for which computational inference is substantially more straightforward. Acknowledgements {#acknowledgements .unnumbered} ================ S.E.H., T.A.W., S.C. and T.M.E. were supported by a European Research Council Advanced Investigator Award (ERC-2010-AdG-268701), and a Program Grant from the Wellcome Trust (number 045404) to T.M.E. References {#references .unnumbered} ========== \[app:property\_proof\]Appendix A: Proofs of Properties of QuASH Models {#appproperty_proofappendix-a-proofs-of-properties-of-quash-models .unnumbered} ======================================================================= \[app:property\_proof\_stat\]Stationary Distribution {#appproperty_proof_statstationary-distribution .unnumbered} ---------------------------------------------------- It is easy to show that the stationary distribution of $Q_j = c_j Q - c_j d_j Q^2$ is the same as that of $Q$. Suppose the row vector ${\bm{\pi}}$ is the stationary distribution of $Q$. Then ${\bm{\pi}} Q = {\bm{0}}{^\mathrm{T}}$ where ${\bm{0}}$ is a column vectors of 0s. It follows that $${\bm{\pi}} Q_j = {\bm{\pi}}(c_j Q - c_j d_j Q^2) = c_j ({\bm{\pi}} Q) - c_j d_j ({\bm{\pi}} Q) Q = {\bm{0}}{^\mathrm{T}}.$$ Hence ${\bm{\pi}}$ is the stationary distribution of $Q_j$. \[app:property\_proof\_rev\]Reversibility {#appproperty_proof_revreversibility .unnumbered} ----------------------------------------- Suppose that the base matrix $Q$ is reversible. We can therefore write $Q = S \Pi$ where $S$ is a symmetric matrix of exchangeability parameters and $\Pi = \text{diag}({\bm{\pi}})$. Since the stationary distribution of $Q_j$ is also ${\bm{\pi}}$, $Q_j$ is reversible if we can find a symmetric matrix $S_j$ such that $Q_j = S_j \Pi$. Now $Q_j = c_j Q - c_j d_j Q^2 = c_j S \Pi - c_j d_j S \Pi S \Pi = c_j (S - d_j S \Pi S) \Pi$ and so we need to check whether the matrix $(S - d_j S \Pi S)$ is symmetric. The difference of two symmetric matrices is symmetric and so this is tantamount to checking whether $S \Pi S$ is symmetric. The matrix $S$ is symmetric and we denote the $(u,v)$-th entry by $s_{\text{min}(u,v),\text{max}(u,v)}$. The matrix $S \Pi$ has $(u,v)$-th entry $s_{\text{min}(u,v),\text{max}(u,v)} \pi_v$ and so the corresponding element of $S \Pi S$ is given by $$(S \Pi S)_{u,v} = \sum_{w} (S \Pi)_{u,w} s_{\text{min}(w,v),\text{max}(w,v)} = \sum_{w} s_{\text{min}(u,w),\text{max}(u,w)} \pi_w s_{\text{min}(w,v),\text{max}(w,v)}.$$ The expression on the right-hand-side is exchangeable with respect to the indices $u$ and $v$ and so $(S \Pi S)_{u,v} = (S \Pi S)_{v,u}$. Therefore the matrix $(S - d_j S \Pi S)$ is symmetric and so $Q_j$ is the rate matrix of a reversible model with exchangeability matrix $$\label{eq:S_j} S_j = c_j (S - d_j S \Pi S).$$ \[app:exp\_var\]Appendix B: Conditional Expectation and Variance of $(d_j | \beta, l, u)$ {#appexp_varappendix-b-conditional-expectation-and-variance-of-d_j-beta-l-u .unnumbered} ========================================================================================= It is straightforward to show that the conditional expectation and variance of the quadratic coefficient $d_j$ given $\beta$, lower limit $l$ and (finite) upper limit $l$ are given by $$\begin{aligned} {\mathrm{E}}(d_j | \beta, l, u) &= l + w \left[1-\frac{\Gamma(\beta+2/w) \Gamma\{\beta (w/u)^{w}+1/w+1\}}{\Gamma(\beta+1/w) \Gamma\{\beta (w/u)^{w}+2/w+1\}} \right] \intertext{and} {\mathrm{Var}}(d_j | \beta, l, u) &= w^2 \left[ \frac{\Gamma(\beta+3/w) \Gamma\{\beta (w/u)^{w}+1/w+1\}}{\Gamma(\beta+1/w ) \Gamma\{\beta (w/u)^{w}+3/w+1\}} - \right.\\ &\qquad \qquad \qquad \left. \frac{\Gamma(\beta+3/w)^2 \Gamma\{\beta (w/u)^{w}+1/w+1\}^2}{\Gamma(\beta+1/w)^2 \Gamma\{\beta (w/u)^{w}+3/w+1\}^2} \right]\end{aligned}$$ where $w = (u-l)$. Similarly, when the upper limit $u$ is infinite, the corresponding expressions are given by $${\mathrm{E}}(d_j | \beta, l) = l - \{\Psi(\beta) - \Psi(\beta e^{-l}+1) \} \quad\text{and}\quad {\mathrm{Var}}(d_j | \beta, l) = \Psi_1(\beta) - \Psi_1(\beta e^{-l}+1),$$ where $\Psi(\beta)=d \ln \Gamma(\beta)/d \beta$ and $\Psi_1(\beta)=d^2\ln \Gamma(\beta)/d \beta^2$ are the digamma and trigamma functions.

ArXiv

--- abstract: | We show that partizan games admit canonical forms in misère play. The proof is a synthesis of the canonical form theorems for normal-play partizan games and misère-play impartial games. It is fully constructive, and algorithms readily emerge for comparing misère games and calculating their canonical forms. We use these techniques to show that there are precisely 256 games born by day 2, and to obtain a bound on the number of games born by day 3. author: - | Aaron N. Siegel\ Institute for Advanced Study\ 1 Einstein Drive\ Princeton, NJ 08540 bibliography: - 'games.bib' title: | Misère Canonical Forms of Partizan Games\ [PREPRINT]{} --- Introduction ============ The identification of a canonical form theorem is often a pivotal moment in understanding a particular theory of combinatorial games. Canonical forms provide evidence of cohesive structure and reassurance that we are not floundering about in uncharted wilderness. Among theories of finite, loopfree games in disjunctive compounds, there have been three major results in this direction. The oldest is the celebrated *Sprague–Grundy Theorem* [@grundy_1939; @sprague_1935; @sprague_1937]: every normal-play impartial game is equivalent to a Nim-heap. Equally important is Conway’s generalization to normal-play partizan games [@berlekamp_1982; @conway_1976]: every such game $G$ is equivalent to a unique game $G'$ with no dominated or reversible followers. Finally, there is an analogous theorem, also due to Conway, for misère-play impartial games [@conway_1976; @siegel_200Xg]. The existence of these three theorems leaves an obvious and glaring exception. Despite some recent progress due to Mesdal and Ottaway [@mesdal_2007], the disjunctive theory of misère-play partizan games has largely remained a mystery. In this paper, we show that such games do indeed admit canonical forms, and that they are not all that different from normal-play canonical forms. Furthermore, the proof is constructive, and in fact canonical forms are no harder to compute in misère play than in normal play. Misère Equivalence ================== In normal play, there is a simple recursive test for equality: $G \geq H$ iff no $G^R \leq H$ and $G \leq$ no $H^L$. Furthermore, canonical forms can be obtained by eliminating dominated options and bypassing reversible ones. In this paper we generalize these results to misère play. Remarkably, the definitions of dominated and reversible options, and the canonical form theorem itself, carry over to misère play without modification. However, the recursive test for $\geq$, and the associated proofs, are considerably more involved. We denote by $o^-(G)$ the misère outcome class of $G$: $$o^-(G) = \begin{cases} {\ensuremath{\mathscr{L}}}& \textrm{if Left can win no matter who plays first}; \\ {\ensuremath{\mathscr{R}}}& \textrm{if Right can win no matter who plays first}; \\ {\ensuremath{\mathscr{P}}}& \textrm{if second player (the Previous player) can win}; \\ {\ensuremath{\mathscr{N}}}& \textrm{if first player (the Next player) can win}. \end{cases}$$ The four outcome classes are naturally partially ordered by “favorability to Left,” with ${\ensuremath{\mathscr{L}}}\geq {\ensuremath{\mathscr{P}}}\geq {\ensuremath{\mathscr{R}}}$ and ${\ensuremath{\mathscr{L}}}\geq {\ensuremath{\mathscr{N}}}\geq {\ensuremath{\mathscr{R}}}$, and this induces the usual partial order on all misère games. $$G \geq H \textrm{ iff } o^-(G+X) \geq o^-(H+X) \textrm{ for all games } X.$$ We also define misère equality in the usual manner. $$G = H \textrm{ iff } G \geq H \textrm{ and } H \geq G.$$ We begin with a simple, but useful, proposition. \[prop:geqsimpler\] $G \geq H$ iff the following two conditions hold: 1. For all $X$ with $o^-(H+X) \geq {\ensuremath{\mathscr{P}}}$, we have $o^-(G+X) \geq {\ensuremath{\mathscr{P}}}$; and 2. For all $X$ with $o^-(H+X) \geq {\ensuremath{\mathscr{N}}}$, we have $o^-(G+X) \geq {\ensuremath{\mathscr{N}}}$. $\Rightarrow$ is immediate. For the converse, we must show that $o^-(G+X) \geq o^-(H+X)$, for all $X$. If $o^-(H+X) = {\ensuremath{\mathscr{R}}}$, then there is nothing to prove; if $o^-(H+X) = {\ensuremath{\mathscr{P}}}$ or ${\ensuremath{\mathscr{N}}}$, it is immediate from (i) or (ii), respectively. Finally, if $o^-(H+X) = {\ensuremath{\mathscr{L}}}$, then by (i) and (ii) we have $o^-(G+X) \geq$ both ${\ensuremath{\mathscr{P}}}$ and ${\ensuremath{\mathscr{N}}}$, whence $o^-(G+X) = {\ensuremath{\mathscr{L}}}$. Ends and Adjoints ================= In normal play, every “one-sided” game $\{G^L|~\}$ or $\{~|G^R\}$ is equivalent to an integer. In misère play, by contrast, such games represent significant pathologies and are the source of much complication. We will sometimes use a dot to indicate “no moves,” when it is useful to clarify the notation; for example, we may write $\Game{0}{\cdot}$ in place of $\Game{0}{}$. $G$ is a *Left (Right) end* if $G$ has no Left (Right) option. In normal play, it is always the case that $G + (-G)$ is equal to zero. This is emphatically false in misère play. Nonetheless, we can give an explicit example of a game $G^\circ$ such that $G + G^\circ$ is always a ¶-position. Readers familiar with the impartial theory will recognize it as the partizan analogue of Conway’s *mate*. \[def:adjoint\] The *adjoint* of $G$, denoted $G^\circ$, is given by $$G^\circ = \begin{cases} {\ast}& \textrm{if $G = 0$} \\ \Game{(G^R)^\circ}{0} & \textrm{if $G \neq 0$ and $G$ is a Left end} \\ \Game{0}{(G^L)^\circ} & \textrm{if $G \neq 0$ and $G$ is a Right end} \\ \Game{(G^R)^\circ}{(G^L)^\circ} & \textrm{otherwise} \end{cases}$$ \[prop:gplusgadj\] $G + G^\circ$ is a ¶-position. By symmetry, it suffices to show that Left can win $G + G^\circ$ moving second. By definition, $G^\circ$ is not a Right end, so Right must have a move. If Right moves to $G^R + G^\circ$ or $G + (G^L)^\circ$, Left makes the mirror image move on the other component, which wins by induction. Finally, if $G$ is a Left end and Right moves to $G + 0$, then Left has no move, and so wins *a priori*. \[theorem:dualdistinguishability\] If $G \not\geq H$, then: 1. There is some $T$ such that $o^-(G+T) \leq {\ensuremath{\mathscr{P}}}$ but $o^-(H+T) \geq {\ensuremath{\mathscr{N}}}$; and 2. There is some $U$ such that $o^-(G+U) \leq {\ensuremath{\mathscr{N}}}$ but $o^-(H+U) \geq {\ensuremath{\mathscr{P}}}$. We know by Proposition \[prop:geqsimpler\] that one of (a) or (b) must hold, so it suffices to show that (a) $\Rightarrow$ (b) and (b) $\Rightarrow$ (a). The arguments are identical, so we will show that (a) $\Rightarrow$ (b). Fix $T$ so that $o^-(G+T) \leq {\ensuremath{\mathscr{P}}}$ and $o^-(H+T) \geq {\ensuremath{\mathscr{N}}}$, and put $$U = \Game{\left(H^R\right)^\circ}{T}.$$ Now from $G+U$, Right has a winning move, to $G+T$. Therefore $o^-(G+U) \leq {\ensuremath{\mathscr{N}}}$. Likewise, consider $H+U$. It is certainly not a Right end, since Right has a move from $U$ to $T$. Now if Right moves to $H^R+U$, Left has a winning response to $H^R+(H^R)^\circ$. If instead Right moves to $H+T$, then since $o^-(H+T) \geq {\ensuremath{\mathscr{N}}}$, Left wins *a priori*. Therefore $o^-(H+U) \geq {\ensuremath{\mathscr{P}}}$, as needed. Recently Mesdal and Ottaway [@mesdal_2007] showed that $G \neq 0$ unless $G$ is identically zero. The following Lemma generalizes that result, and it proves to be a crucial piece of the analysis. \[lemma:leftendnotleq\] If $H$ is a Left end and $G$ is not, then $G \not\geq H$. Put $$T = \Big\{\left(H^R\right)^\circ~\Big|\Big|~\cdot~\Big|~\left(G^L\right)^\circ\Big\}.$$ Consider $H+T$. If Right moves to $H^R + T$, then Left can respond to $H^R + (H^R)^\circ$, winning; if Right moves to $H + \Game{\cdot}{(G^L)^\circ}$, then Left wins outright, since he has no further move. Therefore $o^-(H+T) \geq {\ensuremath{\mathscr{P}}}$. Now consider $G+T$. Right has a move to $G + \Game{\cdot}{(G^L)^\circ}$. Left’s only response is to $G^L + \Game{\cdot}{(G^L)^\circ}$, which must exist since $G$ is not a Left end. But Right may then respond to $G^L+(G^L)^\circ$, which wins. Therefore $o^-(G+T) \leq {\ensuremath{\mathscr{N}}}$. This shows that $o^-(G+T) \not\geq o^-(H+T)$, so in fact $G \not\geq H$. Dominated and Reversible Options ================================ It is a remarkable fact that the definitions of dominated and reversible options are *exactly* the same as in normal play. Let $G$ be a game. 1. A Left option $G^L$ is said to be *dominated* if $G^{L'} \geq G^L$ for some other Left option $G^{L'}$. 2. A Right option $G^R$ is said to be *dominated* if $G^{R'} \leq G^R$ for some other Right option $G^{R'}$. 3. A Left option $G^L$ is said to be *reversible* if $G^{LR} \leq G$ for some Right option $G^{LR}$. 4. A Right option $G^R$ is said to be *reversible* if $G^{RL} \geq G$ for some Left option $G^{RL}$. Suppose $G^{L_1}$ is dominated by $G^{L_2}$, and let $G'$ be the game obtained by eliminating $G^{L_1}$ from $G$. Then $G = G'$. Since the Left options of $G'$ are a subset of those of $G$, and since $G'$ still has at least one Left option (namely, $G^{L_2}$), we trivially have $G' \leq G$. Thus it suffices to show that $G' \geq G$. So fix $X$, and suppose that Left can win $G+X$ playing first (or second). He follows exactly the same strategy on $G'+X$, except when it recommends a move from some $G+Y$ to $G^{L_1}+Y$. In that case, we have $o^-(G^{L_1}+Y) \geq {\ensuremath{\mathscr{P}}}$. Since $G^{L_2} \geq G^{L_1}$, this necessarily implies $o^-(G^{L_2}+Y) \geq {\ensuremath{\mathscr{P}}}$. So Left can win by moving from $G'+Y$ to $G^{L_2}+Y$. Suppose $G^{L_1}$ is reversible through $G^{L_1R_1}$, and let $G'$ be the game obtained by bypassing $G^{L_1}$: $$G' = \Game{G^{L_1R_1L},G^{L'}}{G^R},$$ where $G^{L'}$ is understood to range over all Left options of $G$ except $G^{L_1}$. Then $G = G'$. First suppose Left can win playing first (or second) on $G + X$. Fix a winning strategy for Left, and assume that it recommends a move on the $X$ component unless the only winning move is on $G$. Left follows exactly the same strategy on $G' + X$ except when it recommends a move from $G$ to $G^{L_1}$. In that case the position must be $G' + Y$, with $o^-(G^{L_1}+Y) \geq {\ensuremath{\mathscr{P}}}$. Thus $o^-(G^{L_1R_1}+Y) \geq {\ensuremath{\mathscr{N}}}$. Now $G$ is not a Left end (since it has $G^{L_1}$ as an option), so by Lemma \[lemma:leftendnotleq\], neither is $G^{L_1R_1}$. Therefore Left must have a winning move from $G^{L_1R_1}+Y$. It cannot be to $G^{L_1R_1}+Y^L$, since this would imply $o^-(G+Y^L) \geq {\ensuremath{\mathscr{P}}}$, contradicting Left’s choice of strategy. Therefore Left’s move to $G^{L_1R_1L} + Y$ must be winning, and he can make this move directly from $G' + Y$. Now suppose Right can win playing first (or second) on $G + X$. Fix a winning strategy for Right. She follows exactly the same strategy on $G' + X$ except when Left moves from $G'$ to $G^{L_1R_1L}$. In that case, the position must be $G^{L_1R_1L} + Y$, with $o^-(G+Y) \leq {\ensuremath{\mathscr{P}}}$. Since $G^{L_1R_1} \leq G$, we have $o^-(G^{L_1R_1}+Y) \leq {\ensuremath{\mathscr{P}}}$, so $G^{L_1R_1L} + Y$ is necessarily a win for Right moving first. Finally, since $G^{L_1R_1}$ is not a Left end (as noted above), neither is $G'$. Thus if Right follows this strategy, Left can never run out of moves prematurely. The Canonical Form Theorem ========================== \[theorem:canonicalform\] Suppose $G = H$, and assume that neither $G$ nor $H$ has any dominated or reversible options. Then for every $H^L$ there is a $G^L$ such that $G^L = H^L$, and vice versa; and likewise for Right options. In order to prove Theorem \[theorem:canonicalform\], we must generalize some machinery from Conway’s proof that *impartial* games admit misère canonical forms. \[def:linked\]   1. $G$ is *downlinked to $H$ (by $T$)* iff $o^-(G+T) \leq {\ensuremath{\mathscr{P}}}$ and $o^-(H+T) \geq {\ensuremath{\mathscr{P}}}$. 2. $G$ is *uplinked to $H$ (by $T$)* iff $o^-(G+T) \geq {\ensuremath{\mathscr{P}}}$ and $o^-(H+T) \leq {\ensuremath{\mathscr{P}}}$. \[theorem:geqiffnolink\] $G \geq H$ iff the following four conditions hold. 1. $G$ is downlinked to no $H^L$; 2. No $G^R$ is downlinked to $H$; 3. If $H$ is a Left end, then so is $G$; 4. If $G$ is a Right end, then so is $H$. For $\Rightarrow$ (i), fix any game $T$. If $o^-(G+T) \leq {\ensuremath{\mathscr{P}}}$, then necessarily $o^-(H+T) \leq {\ensuremath{\mathscr{P}}}$ as well. Therefore $o^-(H^L+T) \leq {\ensuremath{\mathscr{N}}}$, so $T$ cannot downlink $G$ to $H^L$. $\Rightarrow$ (ii) is similar, and $\Rightarrow$ (iii) and (iv) are just restatements of Lemma \[lemma:leftendnotleq\] (and its mirror image). We now prove $\Leftarrow$. First fix $T$ such that $o^-(H+T) \geq {\ensuremath{\mathscr{P}}}$, and suppose (for contradiction) that $o^-(G+T) \leq {\ensuremath{\mathscr{N}}}$. Then either $o^-(G+T^R) \leq {\ensuremath{\mathscr{P}}}$, or $o^-(G^R+T) \leq {\ensuremath{\mathscr{P}}}$, or else $G+T$ is a Right end. If $o^-(G+T^R) \leq {\ensuremath{\mathscr{P}}}$, then by induction on the birthday of $T$ we may assume that $o^-(H+T^R) \leq {\ensuremath{\mathscr{P}}}$, contradicting the assumption that $o^-(H+T) \geq {\ensuremath{\mathscr{P}}}$. If $o^-(G^R+T) \leq {\ensuremath{\mathscr{P}}}$, then $T$ downlinks $G^R$ to $H$, contradicting (ii). Finally, if $G+T$ is a Right end, then in particular $G$ is a Right end, so by (iv) $H$ is a Right end. Therefore $H+T$ is a Right end, contradicting the assumption that $o^-(H+T) \geq {\ensuremath{\mathscr{P}}}$. Finally, we must show that if $o^-(H+T) \geq {\ensuremath{\mathscr{N}}}$, then $o^-(G+T) \geq {\ensuremath{\mathscr{N}}}$. The proof is identical, with (i) and (iii) in place of (ii) and (iv). \[theorem:linkiffnogeq\] $G$ is downlinked to $H$ iff no $G^L \geq H$ and $G \geq$ no $H^R$. Suppose $T$ downlinks $G$ to $H$, so that $o^-(G+T) \leq {\ensuremath{\mathscr{P}}}$ and $o^-(H+T) \geq {\ensuremath{\mathscr{P}}}$. Then necessarily $o^-(G^L+T) \leq {\ensuremath{\mathscr{N}}}$ and $o^-(H^R+T) \geq {\ensuremath{\mathscr{N}}}$, so $T$ witnesses both $G^L \not\geq H$ and $G \not\geq H^R$. Conversely, suppose that no $G^L \geq H$ and $G \geq$ no $H^R$. Then for each $G^L_i$, Theorem \[theorem:dualdistinguishability\] yields an $X_i$ such that $$o^-(G^L_i+X_i) \leq {\ensuremath{\mathscr{P}}}\textrm{ and } o^-(H+X_i) \geq {\ensuremath{\mathscr{N}}}.$$ Likewise, for each $H^R_j$, there is some $Y_j$ such that $$o^-(G+Y_j) \leq {\ensuremath{\mathscr{N}}}\textrm{ and } o^-(H^R_j+Y_j) \geq {\ensuremath{\mathscr{P}}}.$$ Put $$T = \begin{cases} {\ast}& \textrm{if $G = H = 0$} \\ \Game{0}{(H^L)^\circ} & \textrm{if $G = 0$ and $H$ is a nonzero Right end} \\ \Game{(G^R)^\circ}{0} & \textrm{if $H = 0$ and $G$ is a nonzero Left end} \\ \Game{Y_j,(G^R)^\circ}{X_i,(H^L)^\circ} & \textrm{otherwise} \end{cases}$$ We claim that $G$ is downlinked to $H$ by $T$. We will show that $o^-(G+T) \leq {\ensuremath{\mathscr{P}}}$; the proof that $o^-(H+T) \geq {\ensuremath{\mathscr{P}}}$ is identical. We first show that $G+T$ has a Left option. If $G$ has a Left option, this is automatic. If $G$ or $H$ has a Right option, then $T$ necessarily has a Left option. This exhausts every case except when $G = 0$ and $H$ is a Right end; but then Left’s move to $0$ is built into the definition of $T$. Thus $G+T$ is not a Left end, and it therefore suffices to show that every Left option is losing. If Left moves to $G^L_i+T$, Right can respond to $G^L_i+X_i$, which wins by choice of $X_i$. If Left moves to $G+(G^R)^\circ$, Right can respond to $G^R+(G^R)^\circ$, which wins by Proposition \[prop:gplusgadj\]. Left’s move to $G+Y_j$ loses automatically, by choice of $Y_j$. The only remaining possibility is Left’s additional move to $0$ in the first two cases of the definition of $T$. But that move is only available when $G = 0$, so it ends the game immediately. Fix $H^L$. Since $G \geq H$, Theorem \[theorem:geqiffnolink\] implies that $G$ is not downlinked to $H^L$. By Theorem \[theorem:linkiffnogeq\], either $G^L \geq H^L$ or $G \geq H^{LR}$. The latter would imply that $H \geq H^{LR}$, contradicting the assumption that $H$ has no reversible options. So we must have $G^L \geq H^L$. An identical argument, using the fact that $H \geq G$, shows that $H^{L'} \geq G^L$ for some $H^{L'}$. Therefore $$H^{L'} \geq G^L \geq H^L.$$ Since $H$ has no dominated options, we must have $H^{L'} = H^L$, so that $$H^{L'} = G^L = H^L.$$ The same argument suffices for the remaining cases. Games Born by Day 2 =================== There are four games born by day $1$; and they are familiar from the normal-play theory: $$0 = \Game{\cdot}{\cdot} \qquad {\ast}= \Game{0}{0} \qquad 1 = \Game{0}{\cdot} \qquad {\overline{1}}= \Game{\cdot}{0}$$ Remarkably, they are pairwise incomparable. \[prop:day1incomparable\] The four games $0$, ${\ast}$, $1$, and ${\overline{1}}$ are pairwise incomparable. Theorem \[theorem:geqiffnolink\](iii) and (iv) immediately yield ${\ast},1 \not\geq 0,{\overline{1}}$ and $0,1 \not\geq {\ast},{\overline{1}}$. Since $0$ is downlinked to $0$ (by ${\ast}$), (i) and (ii) furthermore imply $0 \not\geq 1$ and ${\overline{1}}\not\geq 0$. Now as a trivial consequence of Theorem \[theorem:linkiffnogeq\], we have that ${\overline{1}}$ is downlinked to $0$ and $0$ is downlinked to $1$. It therefore follows from Theorem \[theorem:geqiffnolink\](i) that ${\overline{1}}\not\geq 1,{\ast}$ and from (ii) that ${\overline{1}},{\ast}\not\geq 1$. This exhausts all possibilities. There are 256 games born by day 2. There are 16 subsets of $\{0,{\ast},1,{\overline{1}}\}$. This gives 256 isomorphism types for games born by day 2, so it suffices to show that every (formal) game born by day 2 is canonical. So fix such a game $G$. By Proposition \[prop:day1incomparable\], $G$ has no dominated options, so it suffices to show that $G$ has no reversible options. Consider some $G^{LR}$. Since $G$ is born by day 2, $G^{LR}$ is born by day $0$, so necessarily $G^{LR} = 0$. But $G^{LR}$ is a Left end and $G$ is not (since it has $G^L$ as a Left option), so by Lemma \[lemma:leftendnotleq\] $G \not\geq G^{LR}$. Let $\mathbb{P}$ denote the set of games born by day 2. We next describe the partial order structure of $\mathbb{P}$. Define $$\begin{array}{lcl} \mathbb{P}^+ & = & \{G \in \mathbb{P} : G \textrm{ is a nonzero Left end}\} \\ \mathbb{P}^- & = & \{G \in \mathbb{P} : G \textrm{ is a nonzero Right end}\} \\ \mathbb{P}^0 & = & \{G \in \mathbb{P} : G \textrm{ is not an end}\} \end{array}$$ Then we can write $\mathbb{P}$ as a disjoint union $$\mathbb{P} = \mathbb{P}^+ \cup \mathbb{P}^- \cup \mathbb{P}^0 \cup \{0\}.$$ Now let $\mathbb{B}_4$ denote the complete Boolean lattice of dimension 4. Let $\mathbb{B}_4^+$ be the partial order obtained by removing the largest element from $\mathbb{B}_4$, and likewise delete the smallest element to obtain $\mathbb{B}_4^-$. We will show that $$\label{eq:poisotypes} \mathbb{P}^+ \cong \mathbb{B}_4^+; \qquad \mathbb{P}^- \cong \mathbb{B}_4^-; \qquad \mathbb{P}^0 \cong \mathbb{B}_4^+ \times \mathbb{B}_4^-. \tag{\dag}$$ In order to characterize the structure of $\mathbb{P}$, it will then suffice to describe the interaction between components. First note that there are certain trivial order relations, described by the following definition. \[def:trivialpo\] We say that *$G$ trivially exceeds $H$*, and write $G \geq_T H$, iff: 1. The Left options of $G$ form a superset of those of $H$; 2. The Right options of $G$ form a subset of those of $H$; 3. $G$ is a Left end iff $H$ is a Left end; and 4. $G$ is a Right end iff $H$ is a Right end. As the terminology suggests, it is a trivial fact that if $G \geq_T H$, then necessarily $G \geq H$. We will now show that on day 2, the converse holds with only a few exceptions. \[theorem:poisusuallytrivial\] Fix $G$ and $H$ satisfying Definition \[def:trivialpo\](iii) and (iv), and assume that $G$ and $H$ are both born by day 2. If $G \geq H$, then $G \geq_T H$. We show that every Left option of $H$ is a Left option of $G$; the argument that every Right option of $G$ is a Right option of $H$ is identical. So fix an $H^L$; by Theorem \[theorem:geqiffnolink\], $G$ is not downlinked to $H^L$. By Theorem \[theorem:linkiffnogeq\], either $G^L \geq H^L$ for some $G^L$, or else $G \geq H^{LR}$. Now since $H$ has the Left option $H^L$, it is not a Left end, whence by assumption neither is $G$. Since $H$ is born by day 2, we know that every $H^{LR} = 0$, so by Lemma \[lemma:leftendnotleq\] we cannot have $G \geq H^{LR}$. Therefore $G^L \geq H^L$ for some $G^L$. But $G^L$ and $H^L$ are both born by day 1. By Proposition \[prop:day1incomparable\], this implies $G^L = H^L$. Now if $G$ and $H$ are in the same component of $\mathbb{P}$, then they necessarily satisfy Definition \[def:trivialpo\](iii) and (iv). Therefore, *on each component*, the relations $\geq$ and $\geq_T$ coincide. But this immediately establishes (\[eq:poisotypes\]). For example, for the isomorphism $\mathbb{P}^+ \to \mathbb{B}_4^+$, we can regard $\mathbb{B}_4$ as the powerset lattice of $\{0,{\ast},1,{\overline{1}}\}$; then each $G$ maps to its set of Right options. To complete the picture of $\mathbb{P}$, we must characterize the interaction between the four components. We are concerned specifically with the case where $H$ is a Right end, but $G$ is not; or where $G$ is a Left end, but $H$ is not (the converses are ruled out by Lemma \[lemma:leftendnotleq\]). \[theorem:pointerrelationships\] The ordering of $\mathbb{P}$ is generated by its restrictions to $\mathbb{P}^+$, $\mathbb{P}^-$, and $\mathbb{P}^0$, together with the following four relations and their mirror images. $$\{|{\ast},1\} \geq 0 \qquad \{{\ast}|{\ast},1\} \geq \{{\ast}|\} \qquad \{{\overline{1}}|{\ast},1\} \geq \{{\overline{1}}|\} \qquad \{{\ast},{\overline{1}}|{\ast},1\} \geq \{{\ast},{\overline{1}}|\}$$ It is a simple matter to verify each of the four stated relations. To prove the Theorem, we must show that no further ones are possible. We first characterize those games that compare with $0$. So suppose $G \geq 0$. By Theorem \[theorem:geqiffnolink\] $G$ is necessarily a Left end, and furthermore no $G^R$ is downlinked to $0$. By Theorem \[theorem:linkiffnogeq\], the Right options of $G$ must therefore be a subset of $\{{\ast},1\}$. So either $G = 0$, or $G \geq_T \{|{\ast},1\}$. Games with $G \leq 0$ are characterized symmetrically. Now suppose $G \geq H$, $H$ is a Right end, and $G$ is not. Consider any $G^R$. By Theorem \[theorem:geqiffnolink\], $G^R$ is not downlinked to $H$. Since $H$ is a Right end, it is necessarily the case that $G^{RL} \geq H$. In particular, $G^R$ is not a Left end. Furthermore, $G$ is born by day 2, so $G^{RL} = 0$ and we have $H \leq 0$. By the previous argument, this implies that $H \leq_T \{{\ast},{\overline{1}}|\}$. Therefore the Right options of $G$ form a subset of $\{{\ast},1\}$, and the Left options of $H$ form a subset of $\{{\ast},{\overline{1}}\}$. Furthermore, the Left options of $G$ form a superset of those of $H$, just as in the proof of Theorem \[theorem:poisusuallytrivial\]. Therefore $G \geq H$ is implied by one of the three given relations, each representing one possibility for the Left options of $H$. Antichains by Day 2 {#antichains-by-day-2 .unnumbered} ------------------- Since $\mathbb{P}$ has such a clean structure, we can get a tight bound on the number of antichains. From a standard reference (such as [@sloane A000372]), we find that $\mathbb{B}_4$ has 168 antichains. This shows that $\mathbb{P}^+$ (and hence $\mathbb{P}^-$ as well) has precisely 167, since there is a unique antichain containing the largest element of $\mathbb{B}_4$. Consider $\mathbb{P}^0$. We have $\mathbb{B}_4 \times \mathbb{B}_4 \cong \mathbb{B}_8$, trivially. Again from a standard reference, we find that $\mathbb{B}_8$ has 56130437228687557907788 antichains. Since every antichain of $\mathbb{P}^0$ is an antichain of $\mathbb{B}_8$, this gives an upper bound for the number of antichains of $\mathbb{P}^0$. Finally, $\{0\}$ trivially admits just two antichains, $\emptyset$ and $\{0\}$. Since every antichain on $\mathbb{P}$ restricts to an antichain on each component, this gives an upper bound of $$M = 2 \times 167 \times 167 \times 56130437228687557907788$$ for the number of antichains of $\mathbb{P}$. Thus we obtain an upper bound of $M^2$ games born by day 3. This number is large, indeed (roughly $2^{183}$), but it is much smaller than $2^{512}$, the number of nonisomorphic game trees of height 3. Relationships to Other Theories =============================== In this section we consider how the partizan misère theory relates to other theories of combinatorial games. Denote by $o^+(G)$ the normal-play outcome class of $G$, and define $$\begin{array}{lcl} G \geq^+ H & \textrm{iff} & o^+(G+X) \geq o^+(H+X) \textrm{ for every game $X$}; \bigstrut \\ G \geq^- H & \textrm{iff} & o^-(G+X) \geq o^-(H+X) \textrm{ for every game $X$}. \end{array}$$ $\geq^-$ is the relation that we have been calling $\geq$; for this section only, we include the minus sign for clarity. $\geq^+$ is, of course, the usual Berlekamp–Conway–Guy inequality for normal-play partizan games. The following result shows that $\geq^+$ is a coarsening of $\geq^-$. If $G \geq^- H$, then $G \geq^+ H$. We must show that Left can get the last move in $G - H$. Suppose Right plays to $G^R-H$ (the argument is the same if she plays to $G - H^L$). Since $G \geq^- H$, Theorem \[theorem:geqiffnolink\] implies that $G^R$ is not downlinked to $H$. By Theorem \[theorem:linkiffnogeq\], either $G^{RL} \geq^- H$ or $G^R \geq^- H^R$. By induction, we may assume that either $G^{RL} \geq^+ H$ or $G^R \geq^+ H^R$. In the first case, Left wins by moving to $G^{RL}-H$; in the second, by moving to $G^R-H^R$. In addition to the usual equivalences $=^+$ and $=^-$, obtained by symmetrizing $\geq^+$ and $\geq^-$, we have two further equivalences when $G$ and $H$ are impartial. $$\begin{array}{lcl} G =^+_I H & \textrm{iff} & o^+(G+X) = o^+(H+X) \textrm{ for every impartial game $X$}; \bigstrut \\ G =^-_I H & \textrm{iff} & o^-(G+X) = o^-(H+X) \textrm{ for every impartial game $X$}. \end{array}$$ It is a well-known fact that $=^+_I$ is just the restriction of $=^+$ to impartial games. It is worth pointing out that the analogous statement does *not* hold for misère games. There exist impartial games $G$ and $H$ such that $G =^-_I H$ but $G \neq^- H$. It is well-known that ${\ast}+ {\ast}=^-_I 0$ (see [@conway_1976]). However, ${\ast}+ {\ast}\neq^- 0$, by Lemma \[lemma:leftendnotleq\]. Therefore $=^-_I$ is a strict coarsening of $=^-$. This highlights an interesting difference between normal and misère play: there exist impartial games $G$ and $H$ that are distinct in partizan misère play, but that are not distinguishable by any impartial game. This behavior arises in other theories, as well; for example, there exist impartial loopy games $G$ and $H$ that are distinct (in normal play), but that are not distinguishable by any impartial loopy game [@siegel_2006a]. Indeed, the coincidence of $=^+$ and $=^+_I$ appears to be an artifact of the special nature of short games in normal play: it is the exception rather than the rule. Partizan Misère Quotients ========================= Recently Thane Plambeck [@plambeck_2005] observed that, if $\mathscr{A}$ is any set of impartial games, then its misère-play structure can often be simplified by localizing the misère equivalence relation to $\mathscr{A}$. Plambeck showed that many important aspects of the theory can be generalized to the local setting, and the structure theory of such quotients has been explored in detail; see [@siegel_200Xd; @siegel_200Xe]. It is not our intention to replicate that analysis here, but merely to remark that a partizan generalization exists. The construction is exactly the same, but instead of a bipartite monoid $(\mathcal{Q},\mathcal{P})$, we now have a *tetrapartite monoid* $(\mathcal{Q},\Pi)$, where $\Pi : \mathcal{Q} \to \{{\ensuremath{\mathscr{L}}},{\ensuremath{\mathscr{R}}},{\ensuremath{\mathscr{P}}},{\ensuremath{\mathscr{N}}}\}$ is the *outcome partition* for $\mathcal{Q}$. Intriguingly, such monoids have an induced partial order structure, given by $$x \geq y \textrm{~~iff~~} \Pi(xz) \geq \Pi(yz) \textrm{ for all } z \in \mathcal{Q}.$$ If $G \geq H$, then it is certainly true that $\Phi(G) \geq \Phi(H)$. However, the quotient may also gain new order-relations that are not present in the universe of games. We include one example to illustrate the rich possibilities. In the previous section we remarked that $1$ and ${\overline{1}}$ are incomparable, and we have also seen that $1 + {\overline{1}}\neq 0$. In $\mathcal{Q}(1,{\overline{1}})$, however, the expected inequalities hold: $$\Phi({\overline{1}}) > \Phi(0) > \Phi(1); \quad \textrm{and} \quad \Phi({\overline{1}})\Phi(1) = \Phi(0);$$ and indeed we have $\mathcal{Q}(1,{\overline{1}}) \cong \mathbb{Z}$, equipped with the usual partial-order structure. We leave it to the reader to verify these assertions.

ArXiv

--- abstract: 'In this paper, we show that if the reduced Fourier-Stieltjes algebra $B_{\rho}(G)$ of a second countable locally compact group $G$ has either weak\* fixed point property or asymptotic center property, then $G$ is compact. As a result, we give affirmative answers to open problems raised by Fendler and et al. in 2013. We then conclude that a second countable group with a discrete reduced dual must be compact. This generalizes a theorem of Baggett. We also construct a compact scattered Hausdorff space $\Omega$ for which the dual of the scattered C\*-algebra $C(\Omega)$ lacks weak\* fixed point property. The C\*-algebra $C(\Omega)$ provides a negative answer to a question of Randrianantoanina in 2010. In addition, we prove a variant of Bruck’s generalized fixed point theorem for the preduals of von Neumann algebras. Furthermore, we give some examples emphasizing that the conditions in Bruck’s generalized conjecture (BGC) can not be weakened any more.' address: 'Department of Mathematical and Statistical Sciences, University of Alberta, Edmonton, Alberta, T6G 2G1, Canada.' author: - Fouad Naderi title: 'C\*-algebraic approach to fixed point theory generalizes Baggett’s theorem to groups with discrete reduced duals' --- Introduction ============ Let $K$ be a subset of a Banach space $E$. A self mapping $T$ on $K$ is said to be [*non-expansive*]{} if $\|T(x)-T(y)\|\leq \|x-y\|$ for all $x,y\in K$. We say that $E$ has the [*fixed point property (fpp)*]{} if for every bounded closed convex non-empty subset $K$ of $E$, any non-expansive self mapping on $K$ has a fixed point. $E$ is said to have the [*weak fixed point property (weak fpp)*]{} if for every weakly compact convex non-empty subset $K$ of $E$, any non-expansive self mapping on $K$ has a fixed point. For a dual Banach space $E$, [*weak\* fixed point property (weak\* fpp)*]{} is defined similarly. One can easily check that the weak fpp is the weakest notion among these properties, this is why the weak fpp is the most desirable property among them. Note that, the definition of fpp is based on bounded closed sets not on compact sets, otherwise; the fpp would be automatic by Schauder’s fixed point theorem. In recent years, Lau has launched a program in which he has successfully related various topological properties of a locally compact group $G$ to fixed point properties of Banach algebras associated to $G$ ( see [@Fendler-Lau], [@Lau-Leinert], [@Lau-Mah-Ulger], [@Lau-Ulger], and [@Randrian]). In [@Randrian], it has been shown that the Fourier-Stieltjes algebra $B(G)$ has weak fpp if and only if $G$ is an \[AU\]-group. Fendler and et al. in [@Fendler-Lau] have shown that $B(G)$ has weak\* fpp if and only if $G$ is a compact group. They also have shown that these are equivalent to saying that $B(G)$ has asymptotic center property (see Definition \[asym\]). Then, they asked if the corresponding result is true for the reduced Fourier-Stieltjes algebra $B_{\rho} (G)$ or not? In this paper we are mainly concerned with this question. It turns out that the affirmative answer to this question has many important implications. This paper is organized as follows: section 2 gives some preliminaries. In section 3, we prove that the conjecture of Fendler and et al. is true for second countable locally compact groups. We then show a second countable group with a discrete reduced dual is compact. This is a generalization of Baggett’s theorem [@Bagget]. Also, our proof gives an alternative proof to Baggett’s original theorem for a second countable locally compact group with a discrete full dual. Besides, for the compact scattered set $\Omega=\{\frac{1}{n}: n\in \mathbb{N} \}\cup \{0\} $ in real numbers, we show that the dual space of the C\*-algebra $C(\Omega)$ does not have weak\* fpp. This gives a negative answer to [@Randrian Question 3.9]. In section 4, we discuss about Bruck’s generalized fixed point theorem. Here, we show that if the predual of a von Neumann algebra has weak fpp, then it must have weak fpp for left reversible semigroups. Also, we give examples of semigroups acting separately continuous on compact sets such that every $s\in S$ has a fixed point, but $S$ still does not have a common fixed point. The problem with these examples is that they don’t enjoy one or two conditions of Bruck’s generalized conjecture. The paper concludes with some open problem arising from the current work. Some preliminaries ================== Let $G$ be locally compact group with a fixed left Haar measure $\mu$. Let $L^{1}(G)$ be the group algebra of $G$ with convolution product. We define $C^{\ast}(G)$, the group $C^{\ast}$-algebra of $G$, to be the completion of $L^{1}(G)$ with respect to the norm $$\|f\|_{*}=sup\|\pi_{f}\|$$ where the supremum is taken over all non-degenerate $\ast$-representation $\pi$ of $L^{1}(G)$ as a $\ast$-algebra of bounded operators on a Hilbert space. Let $\mathcal{B}(L^{2}(G))$ be the set of all bounded operators on the Hilbert space $L^{2}(G)$ and $\rho$ be the left regular representation of $G$, i.e., for each $f\in L^{1}(G)$, $\rho(f)$ is the bounded operator in $\mathcal{B}(L^{2}(G))$ defined by $\rho(f)(h)=f\ast h$, the convolution of $f$ and $h$ in $L^{2}(G)$. The reduced group C\*-algebra ${{C^{*}}}_{\rho}(G)$ is the closure of $\rho(L^{1}(G))$ in the norm topology of $\mathcal{B}(L^{2}(G))$ while the von Neumann algebra generated by $\rho$ , $VN(G)$, is the closure of the same set in the weak operator topology. For any other unitary representation $\pi$, the von Neumann algebra generated by $\pi$ is the closure of $\pi(L^{1}(G))$ in the weak operator topology of $\mathcal{B}({\mathcal{H}}_{\pi})$, where ${\mathcal{H}}_{\pi}$ is the Hilbert space of the representation $\pi$. Denote the set of all continuous positive definite function on $G$ by $P(G)$, and the set of all continuous function on $G$ with compact support by $C_{c}(G)$. We define the Fourier-Stieltjes algebra of $G$, denoted by $B(G)$, to be the linear span of $P(G)$. Then, $B(G)$ is a Banach algebra with pointwise multiplication and the norm of each $\phi \in B(G)$ is defined by $$\|\phi\|=sup\{ | \int f(t)\phi(t) d\mu(t) | : f\in L^{1}(G), \|f\|_{*}\leq 1 \}.$$ The Fourier algebra of $G$, denoted by $A(G)$, is defined to be the closed linear span of $P(G)\cap C_{c}(G)$. Note that, $A(G)$ is the set of all matrix elements of $\rho$ while $B(G)$ is the set of all matrix elements of all continuous unitary representations of $G$. Clearly, $A(G)=B(G)$ when $G$ is compact. It is known that $B(G)=C^{\ast}(G)^{\ast}$, where the duality is given by $\langle f,\phi\rangle = \int f(t)\phi(t) d\mu(t), f\in L^{1}(G), \phi \in B(G)$. Besides, we can define the reduced Fourier-Stieltjes algebra $ B_{\rho} (G)$ of $G$ to be the dual space of the reduced group C\*-algebra ${C^{*}}_{\rho} (G)$. The duality is pretty much the same as $B(G)=C^{\ast}(G)^{\ast}$. When $G$ is amenable (e.g., abelian and compact groups), one has ${C^{*}}_{\rho} (G)=C^{*}(G)$ and $B_{\rho} (G)=B(G)$. Furthermore, $A(G)$ and $B_{\rho}(G)$ are closed ideals in $B(G)$ and $A(G)\subseteq B_{\rho} (G)\subseteq B(G)$ (see [@Eymard] for details). The set of all unitarily equivalent classes of all irreducible representations of a C\*-algebra $\mathcal{A}$ is called the spectrum of $\mathcal{A}$ and is denoted by $\hat{\mathcal{A}}$. The dual of a locally compact group $G$, denoted by $\hat{G}$, is the spectrum of the group C\*-algebra $C^{*}(G)$. Similarly, the reduced dual of a locally compact group $G$, denoted by $\hat{G}_{\rho}$, is the spectrum of the reduced C\*-algebra ${{C^{*}}}_{\rho}(G)$. The reduced dual has many equivalent definitions (see [@Dixmier]). A von Neumann algebra $\mathcal{M}$ is called [*atomic*]{} if every nonzero (self-adjoint) projection in $\mathcal{M}$ dominates a minimal nonzero projection in $\mathcal{M}$. Let $G$ be a locally compact group. $G$ is called an \[AU\]-group if the von Neumann algebra generated by any continuous unitary representation of $G$ is atomic, i.e., every continuous unitary representation of $G$ is completely decomposable. $G$ is called an \[AR\]-group if the von Neumann algebra generated by $\rho$, $VN(G)$, is atomic. Evidently, every \[AU\]-group is an \[AR\]-group. $G$ is called an \[IN\]-group if there is a compact neighborhood of the identity in $G$ which is invariant under inner automorphisms. For these classes of groups, we have $[AU]\cap [IN]=[compact]$ (see [@Taylor] for more details). $G$ is called a \[Moore\]-group, if each irreducible unitary representation of $G$ is finite dimensional. For example, any compact group is a Moore group [@Folland Chapter 5]. The converse is false since abelian groups are also \[Moore\]-groups and can be non-compact [@Folland Chapter 4]. A Banach space $E$ is said to have [*Dunford-Pettis property (DPP),*]{} if for any Banach space $F$, every weakly compact linear operator $u:E\longrightarrow F$ sends weakly Cauchy sequences into norm convergent sequences. $E$ is said to have the [*Radon-Nikodym property (RNP)*]{} if each closed convex sub-set $D$ of $E$ is dentable, i.e., for any $\varepsilon >0$ there exists an $x$ in $D$ such that $ x\notin \overline{co}(D \setminus B_{\varepsilon}(x))$, where $B_{\varepsilon}(x)=\{y\in X: \|x-y\|<\varepsilon \}$ and $\overline{co}(K)$ is the closed convex hull of a set $K\subseteq E$. The book [@Dis] contains a complete account of these properties. A dual Banach space $E$ is said to have the [*uniform weak\* Kadec-Klee property (UKK\*)*]{} if for every $\epsilon >0$ there is $0<\delta<1$ such that for any subset $C$ of its closed unit ball and any sequence $(x_n)$ in $C$ with $sep(x_n):=inf\{\rVert\ {x_n} -{ x_m}\rVert: n\neq m\}>\epsilon$, there is an $x$ in the weak\* closure of $C$ with $\rVert x \rVert<\delta$. This property was firs introduced by van Dulst and Sims [@Dulst]. They proved that if E has property UKK\*, then $E$ has the weak\* fpp. A dual Banach space $E$ is said to have the [*weak\* Kadec-Klee property (KK\*)*]{} if the weak\* and norm convergence for sequences coincide on the unit sphere of $E$. It is well known that UKK\* implies KK\* ( see [@Lennard] and [@Lau-Mah; @88] for more details). Let $S$ be a [*semi-topological semigroup*]{}, i.e., $S$ is a semigroup with a Hausdorff topology such that for each $a \in S$, the mappings $s\mapsto sa$ and $s\mapsto as$ from $S$ into $S$ are continuous. $S$ is called [*left (right) reversible*]{} if any two closed right (left) ideals of $S$ have non-void intersection. An [*action*]{} of $S$ on a subset $K$ of a topological space $E$ is a mapping $(s,x)\mapsto s(x)$ from $ S \times K$ into $K$ such that $(st)(x)=s(t(x))$ for $ s,t\in S, x\in K$. The action is [*separately continuous*]{} if it is continuous in each variable when the other is kept fixed. When we speak of separately continuous actions, we always put the norm topology on $K$ although $K$ can be a weak or weak\* compact set. Every action of $S$ on $K$ induces a representation of $S$ as a semigroup of self-mappings on $K$ denoted by $\mathcal{S}$, and the two semigroups are usually identified. When the action is separately continuous, each member of $\mathcal{S}$ is continuous. We say that $S$ has a [*common fixed point*]{} in $K$ if there exists a point $x$ in $K$ such that $sx=x$ for all $s\in S$. When $E$ is a normed space, the action of $S$ on $K$ is [*non-expansive*]{} if $ \| s(x)-s(y)\| \leq \| x-y\|$ for all $s \in S$ and $x,y \in K$. There are also other types of action for a semi-topological semigroup (see [@Holmes] and [@Amini]). We say that a Banach space $E$ has the [*fpp for left reversible semigroups*]{} if for every bounded closed convex non-empty subset $K$ of $E$, any non-expansively separately continuous action of a left reversible semi-topological semigroup $S$ on $K$ has a fixed point. $E$ is said to have the [*weak fpp for left reversible semigroups*]{} if for every weakly compact convex non-empty subset $K$ of $E$, any non-expansively separately continuous action of a semi-topological semigroup $S$ on $K$ has a fixed point. For a dual Banach space $E$, [*weak\* fpp for left reversible semigroups*]{} is defined similarly. Let $\mathbb{T}$ be the unit circle in the complex plane which is a locally compact group under usual multiplication and Euclidean norm. Suppose that $\mathbb{Z}$ is the group of all integers under addition and discrete topology. By duality theory for abelian groups, one has $\hat{\mathbb{Z}}=\mathbb{T}$ and $\hat{\mathbb{T}}=\mathbb{Z}$ (see [@Folland Chapter 4]). Alspach [@Alspach] has shown that $L^{1}[0,1]=L^{1}(\mathbb{T})=A(\mathbb{Z})$ does not have weak fpp (for abelian semigroups). This inspired Lau and Mah in [@Lau-Mah] to show that for an \[IN\]-group $G$, $A(G)$ has weak fpp for left reversible semigroups if and only if $G$ is compact. Later, Randrianantoanina [@Randrian] showed that $A(G)$ ( resp. $B(G)$) has weak fpp for left reversible semigroups if and only if $G$ is an \[AR\]-group ( resp. \[AU\]-group). Also, Lim’s fixed point theorem [@Lim] for dual Banach spaces shows that $l^{1}(\mathbb{Z})=B(\mathbb{T})$ has weak\* fpp for left reversible semigroups. This inspired Lau and et al. in [@Fendler-Lau] and [@Lau-Mah] to show that $B(G)$ has weak\* fpp for left reversible semigroups if and only if $G$ is compact. Weak\* fixed point property of reduced Fourier-Stieltjes algebra and Baggett’s generalized theorem ================================================================================================== The following theorem answers the first problem in [@Fendler-Lau p.300] positively. Before stating it, we would like to thank Professor Gero Fendler for finding out a gap in our original proof and suggesting the idea of decomposing a separable C\*-algebra with discrete spectrum according to [@Dixmier 10.10.6]. \[main.thm\] If the reduced Fourier-Stieltjes algebra $B_{\rho} (G)$ of a second countable locally compact group has the weak\* fpp, then $G$ is compact. [**Proof.**]{} For a second countable locally compact group $G$, the reduced group C\*-algebra ${{C^{*}}}_{\rho}(G)$ is separable. Hence, the unit ball of $B_{\rho}(G)$ as the dual Banach space of the separable C\*-algebra ${{C^{*}}}_{\rho}(G)$ is metrizable in weak\* topology [@Conway p.134]. According to [@Fendler-Leinert Theorem 2.1], the reduced dual $\hat{G}_{\rho}$ must be discrete. By the separability assumption and [@Dixmier 10.10.6], one gets the following decompositions ${{C^{*}}}_{\rho}(G)=c_{0}-\oplus_{i\in I} \mathcal{K}({\mathcal{H}}_{i})$; $B_{\rho}(G)=l^{1}-\oplus_{i\in I} \mathcal{T}({\mathcal{H}}_{i})$ where each $\mathcal{H}_{i}$ is a separable Hilbert space, and $\mathcal{K}({\mathcal{H}}_{i})$ and $\mathcal{T}({\mathcal{H}}_{i})$ are compact and trace class operators on it. By a digonalization argument, one can embed $B_{\rho}(G)$ into $\mathfrak{T}=\mathcal{T}(l^{2}-\oplus_{i\in I}{\mathcal{H}}_{i})\cong {\mathcal{K}(l^{2}-\oplus_{i\in I}{\mathcal{H}}_{i})}^{*}={\mathfrak{K}}^{*}$ By [@Lennard], $\mathfrak{T}$ has UKK\*. Hence, it has KK\*. The space $\mathfrak{K}$ is separable, so the unit ball of $\mathfrak{T}$ is metrizable in the weak\* topology stemming from $\mathfrak{K}$. This alongside with the property KK\* means that the norm and weak\* topology on unit sphere of $\mathfrak{T}$ agree. On the other hand, the weak\* topology of $\mathfrak{T}$ relativized to $B_{\rho}(G)$ comes from ${{C^{*}}}_{\rho}(G)$. So, we are allowed to say that the norm and weak\* topology on unit sphere of $B_{\rho}(G)$ agree, as well. By applying [@Miao Theorem 4.2], we deduce that $G$ is compact. $\blacksquare$ The following theorem is a consequence of Theorem \[main.thm\] and generalizes a result of Baggett [@Bagget]. The proof may seem simple, but it is only after a great deal of theory that we can give such a proof. \[Bagget\] A second countable group having a discrete reduced dual is compact. [**Proof.**]{} Let $G$ be a second countable group and $\hat{G}_{\rho}$ be its discrete reduced dual. The spectrum of the reduced C\*-algebra ${C^{*}}_{\rho}(G)$ is $\hat{G}_{\rho}$. According to [@Fendler-Leinert], $ B_{\rho} (G)$ has the weak\* fpp. By Theorem \[main.thm\], $G$ is compact. $\blacksquare$ The method of the proofs of Theorems \[main.thm\] and \[Bagget\] can be used to give an alternative proof to Baggett’s theorem [@Bagget] regarding the full dual. We can also prove the converse of Theorem \[main.thm\]. Although it can be inferred from [@Fendler-Lau Theorem 4.7], we prefer a direct proof to show the sophisticated ideas of Randrianantoanina [@Randrian] in studying fixed point theory. We use a well-known fact that ${{C^{*}}}_{\rho}(G)=C^{*}(G)$ for every amenable group $G$. The discussion given here can be viewed as a simplification of the proof of [@Lau-Mah Theorem 4.2] in light of von Neumann algebraic techniques of [@Randrian]. For the sake of completeness we give some details from [@Lau-Mah]. \[converse\] If $G$ is compact, then the reduced Fourier-Stieltjes algebra $ B_{\rho} (G)$ has weak\* fpp for left reversible semi groups. [**Proof.**]{} When $G$ is compact, $G$ is amenable and every element of $\hat{G}$ is finite dimensional [@Folland Chapter 5]. So, $G$ is a Moore group and by [@Lau-Ulger Theorem 4.2], $C^{*}(G)={{C^{*}}}_{\rho}(G)$ has DPP. According to [@Lau-Ulger Lemma 4.1], $ {{C^{*}}}_{\rho}(G)$ must be a $c_0$-direct sum of finite dimensional C\*-algebras. But each finite dimensional C\*-algebra, is a direct sum of matrix algebras [@Murphy p.194], so ${{C^{*}}}_{\rho}(G)=c_{0}-\oplus_{i\in I}\mathcal{K}({\mathcal{H}}_{i})$ and $B_{\rho}(G)=l^{1}-\oplus_{i\in I}\mathcal{T}({\mathcal{H}}_{i})$ where each ${\mathcal{H}}_{i}$ is a finite dimensional Hilbert space and $\mathcal{K}({\mathcal{H}}_{i})$ and $\mathcal{T}({\mathcal{H}}_{i})$ denote compact and trace class operators on ${\mathcal{H}}_{i}$. By [@Randrian Corrolary 3.7], it is now obvious that $B_{\rho}(G)=l^{1}-\oplus_{i\in I}\mathcal{T}({\mathcal{H}}_{i})$ has the weak\* fpp for left reversible semi groups.$\blacksquare$ \[asym\] (a) Let $C$ be a non-empty subset of a Banach space $E$ and $\{ W_{\alpha} : \alpha \in A\}$ be a decreasing net of non-empty bounded subsets of $E$. For each $c\in C, \alpha \in A$ define $$r_{\alpha}(c)=sup\{ \parallel w-c\parallel: w\in W_{\alpha} \}$$ $$r(c)=inf\{r_{\alpha}(c) : \alpha \in A \}$$ $$r=inf\{ r(c) : c\in C \}$$ The set $ AC(\{W_{\alpha} :\alpha \in A\},C)=\{ c\in C : r(c)=r \}$ (the number $r$ ) will be called the [*asymptotic center (asymptotic radius)*]{} of $\{ W_{\alpha} : \alpha \in A\}$ in $C$. This notion is due to Edelstein [@Edelstein]. A dual Banach space $E$ is said to have the [*asymptotic center property*]{} if for any non-empty weak\* closed convex subset $C$ in $E$ and any decreasing net $\{ W_{\alpha} : \alpha \in A\}$ of non-empty bounded subsets of $C$, the set $ AC(\{W_{\alpha} :\alpha \in A\},C)=\{ c\in C : r(c)=r \}$ is a non-empty norm compact convex subset of $C$. This property was first used by Lim [@Lim]. \(b) A dual Banach space $E$ is said to have the [*lim-sup property*]{} if for any decreasing net $\{ W_{\alpha} : \alpha \in A\}$ of non-empty bounded subsets of $E$, and any weak\* convergent net ${(\varphi_{\mu})}_{\mu \in M}$ with weak\* limit $\varphi$, we have: $$lim sup_{\mu} \|{\varphi}_{\mu}-\varphi\|+ r(\varphi)=lim sup_{\mu} r({\varphi}_{\mu})$$ , i.e., $$lim sup_{\mu} \|{\varphi}_{\mu}-\varphi\|+ lim sup_{\alpha} \{ \|\varphi-\psi\| : \psi \in W_{\alpha} \} =lim sup_{\mu}lim sup_{\alpha} \{ \|{\varphi}_{\mu}-\psi\| : \psi \in W_{\alpha} \}.$$ This notion was first used by Lim [@Lim]. Later, Lau and et al. generalized it to arbitrary dual Banach spaces (see [@Lau-Mah Lemma 3.1] and [@Fendler-Lau Lemma 5.1 and Defenition 5.2]). The next corollary answers the second question in [@Fendler-Lau p.300]. \[asym.cpt\] If the reduced Fourier-Stieltjes algebra $ B_{\rho} (G)$ of a second countable locally compact group has asymptotic center property, then $G$ is compact. [**Proof.**]{} By [@Fendler-Lau Theorem 6.1, conditions b and c], $ B_{\rho} (G)$ has the weak\* fpp. Now, apply Theorem \[main.thm\].$\blacksquare$ The following theorem characterize the compactness of $G$ in terms of weak\* fpp and geometric properties of $ B_{\rho}(G)$. It is the reduced counterpart of [@Fendler-Lau Theorem 5.3]. Recall that a C\*-algebra $\mathcal{A}$ is called [*scattered*]{} if every positive linear functional on $\mathcal{A}$ is the sum of a sequence of pure states. Let $G$ be a second countable locally compact group. The followings are equivalent: \(a) $G$ is compact. \(b) $B_{\rho} (G)$ has the lim-sup property. \(c) $B_{\rho} (G)$ has the asymptotic center property. \(d) $ B_{\rho} (G)$ has the weak\* fpp for left reversible semi-topological semigroups. \(e) $B_{\rho}(G)$ has the weak\* fpp for non-expansive mappings. \(f) $lim sup_{\mu} \|{\varphi}_{\mu}-\varphi\| + \|\varphi\|=lim sup_{\mu} \|{\varphi}_{\mu}\|$ for any bounded net $(\varphi_{\mu})$ in $ B_{\rho} (G)$ which converges to $\varphi \in B_{\rho} (G)$ in the weak\* topology. \(g) for any net $(\varphi_{\mu})$ and any $\varphi \in B_{\rho} (G)$ we have that $\|\varphi_{\mu}-\varphi\|\rightarrow 0$ if and only if $\varphi_{\mu}\rightarrow \varphi$ in the weak\* topology and $\|\varphi_{\mu}\|\rightarrow \|\varphi\|$ \(h) On the unit sphere of $ B_{\rho} (G)$ the weak\* and norm topology are the same. \(i) Each member of $\hat{G}_{\rho}$ is finite dimensional and $B_{\rho}(G)$ has RNP. \(j) $B_{\rho}(G)$ has RNP and DPP. \(k) The C\*-algebra ${C^{*}}_{\rho} (G)$ is scattered and has DPP. \(l) The reduced dual $\hat{G}_{\rho}$ is discrete. [**Proof.**]{} (a)$\Rightarrow$(b): When $G$ is compact, $G$ is amenable, and then $ B_{\rho}(G)=B(G)$. Now apply [@Fendler-Lau Lemma 5.1]. (b)$\Rightarrow$(c)$\Rightarrow$(d)$\Rightarrow$(e) follows from [@Fendler-Lau Theorem 5.3 and Theorem 6.1]. (e)$\Longleftrightarrow$(a) follows from Theorem \[main.thm\] and Proposition \[converse\]. (b)$\Rightarrow$(f)$\Rightarrow$(g)$\Rightarrow$(h) follows from [@Fendler-Lau Theorem 5.3 and Theorem 6.1]. (h)$\Rightarrow$(a) follows from [@Miao Theorem 4.2 (iii)]. (i)$\Longleftrightarrow$(j)$\Longleftrightarrow$(k)$\Longleftrightarrow$(a) follow from [@Lau-Ulger Theorem 4.6]. (a)$\Longrightarrow$(l) by amenability of $G$ and Peter-Weyl theorem [@Folland Chapter 7]. (l)$\Longrightarrow$(a) follows from Theorem \[Bagget\]. $\blacksquare$ A locally compact Hausdorff topological space $X$ is scattered if $X$ does not contain any non-empty perfect subset. Equivalently, every non-empty subset of $X$ has an isolated point. By a theorem of Jensen [@Jensen; @78], $X$ is scattered if and only if the C\*-algebra $C_{0}(X)$ is scattered. Some authors use the term [*dispersed*]{} instead of scattered. The following example introduces some scattered spaces. \[Sctr\] For a separable Hilbert space $\mathcal{H}$, the C\*-algebra $\mathcal{K}(\mathcal{H})$ is scattered. Consider $\mathbb{N}$ with its usual topology and its one point compactification ${\mathbb{N}}^{*}=\mathbb{N} \cup \{\infty\}$. Both $\mathbb{N}$ and ${\mathbb{N}}^{*}$ are scattered. The set $\Omega=\{\frac{1}{n}: n\in \mathbb{N} \}\cup \{0\} $ of reciprocals of naturals with zero added provides another example of a compact scattered space. The set $\Omega$ has the nice property of being scattered but not discrete. We use it in our counter example. Randrianantoanina [@Randrian Corrolary 3.8] has shown that if a scattered C\*-algebra $\mathcal{A}$ is a dual space, then the Banach space ${ \mathcal{A} }^{*}$ has weak\* fpp for left reversible semigroups. Then, he asked [@Randrian Question 3.9] if this result is true without the dual assumption on ${ \mathcal{A} }$. In the following, we give a negative answer to this. As a result, the condition of being dual on $\mathcal{A}$ in [@Randrian Corrolary 3.8] cannot be removed even for abelian C\*-algebras. We remark that an abelian C\*-algebra $\mathcal{A}$ is a dual space if and only if its spectrum $\hat{A}$ is a hyperstonean space ( see [@Takesaki p.109] for more details). \[Randr\] For $\Omega=\{\frac{1}{n}: n\in \mathbb{N} \}\cup \{0\}$, the C\*-algebra $\mathcal{A}=C(\Omega)$ is scattered, but its dual Banach space ${ \mathcal{A} }^{*}$ does not have weak\* fpp for left reversible semigroups. [**Proof.**]{} Obviously by Jensen’s theorem [@Jensen; @78], $\mathcal{A}$ is scattered since $\Omega=\hat{\mathcal{A}}$ is scattered. We claim that $\mathcal{A}$ is separable. Note that, $\Omega$ is a metric space under the absolute value of real numbers and $\{\frac{1}{n}: n\in \mathbb{N} \}$ is a dense subset of it. For each $n\in \mathbb{N}$, put $f_{n}(x)=|x-\frac{1}{n}|$. By continuity of $|.|$, it is obvious that each $f_{n}$ is continuous. Also, the family $\{1\}\cup\{f_{n}: n\in \mathbb{N}\}$ of real-valued functions is a separating family in $\mathcal{A}$. By Stone-Weierstrass theorem, the algebra generated by $\{1\}\cup\{f_{n}: n\in \mathbb{N}\}$ is dense in $\mathcal{A}$. Using the rational coefficient for the linear combinations of $\{1\}\cup\{f_{n}: n\in \mathbb{N}\}$, we see that $\mathcal{A}$ is separable. Suppose to the contrary that ${ \mathcal{A} }^{*}$ has weak\* fpp for left reversible semigroups. By [@Fendler-Leinert Theorem 2.1], $\hat{\mathcal{A}}=\Omega$ must be discrete. But, this is obviously false since for each $m\neq n$, $|\frac{1}{n}-\frac{1}{m}|\neq 0$. This contradiction shows that the answer to question 3.9 in [@Randrian] is negative. $\blacksquare$ Bruck’s Generalized Conjecture for preduals of von Neumann algebras =================================================================== Bruck [@Bruck] has shown that if a Banach space $E$ has weak fpp, then it has weak fpp for abelian semigroups. We call the following statement [*Bruck’s Generalized Conjecture (BGC)*]{}. (BGC) If a Banach space $E$ has weak fpp, then it has weak fpp for any left reversible semi-topological semigroup $S$. It is a brilliant idea to pass from the private fixed points of single mappings to the common fixed point of the whole mappings. There are several techniques for this passage (for example, asymptotic center property and lim-sup property as discussed in [@Lim],[@Fendler-Lau] and[@Randrian]). But, these techniques are interwoven with the geometric and algebraic properties of the Banach space in question. On the other hand, Lim’s fixed point theorem [@Lim-wk-fpp Theorem 3] imposes a strong condition ( called normal structure) on the space in question to assure the existence of a common fixed point for left reversible semi-topological semigroups. But, many good spaces like $c_0$ lack normal structure and still have weak fpp. Meanwhile, amenability of groups has an equivalent formulation in terms of fixed points (see[@Pat p.61]). So, we suspect that the BGC would be true in general unless one is dealing with well-behaved semigroups or special spaces as in [@Bruck], [@Fendler-Lau], [@Lim], and [@Randrian]. We are going to show that the BGC is true for the preduals of von Neumann algebras. Before doing so, we recall some terminology. Let $\mathcal{M}$ be a von Neumann algebra, ${\mathcal{M}}^{'}$ be its commutant and ${ \mathcal{M}}_{prj}$ denotes the set of all projections in $\mathcal{M}$. A projection $z\in { \mathcal{M}}_{prj}$ is called [*central*]{} if $z\in \mathcal{M}\cap {\mathcal{M}}^{'}$. A projection $e\in { \mathcal{M}}_{prj}$ is called [*abelian*]{} if $e{\mathcal{M}}e=\mathcal{M}$. The von Neumann algebra $\mathcal{M}$ is called to be of [*Type I*]{} if every non-zero central projection in it dominates a non-zero abelian projection. $\mathcal{M}$ is called a [*factor*]{} if its center consists of all complex multiple of the identity, i.e., $\mathcal{M}\cap {\mathcal{M}}^{'}=\mathbb{C}\mathcal{I}$. Unfortunately, the Type I terminology conflicts with the same terminology for C\*-algebras since Type I von Neumann algebras need not be Type I C\*-algebras in general. A [*factor of Type I* ]{} is just a von Neumann algebra which is both a factor and a Type I von Neumann algebra. Consult [@Takesaki] for more details on von Neumann algebras. The proof of the following theorem is inspired by Randrianantoanina’s ideas [@Randrian] and generalizes Bruck’s fixed point theorem [@Bruck] for the preduals of von Neumann algebras. \[vn\] Let $\mathcal{M}$ be a von Neumann algebra with the predual ${\mathcal{M}}_{*}$. If ${\mathcal{M}}_{*}$ has weak fpp, then it has weak fpp for left reversible semigroups. [**Proof.**]{} By [@Fendler-Leinert Lemma 2.1], $\mathcal{M}$ is atomic, of Type I and has RNP. An application of [@Takesaki p. 299], reveals that $\mathcal{M}=l^{\infty}-\oplus_{i\in I} L^{\infty}(\mu_{i}) {\overline{\otimes}}\mathcal{B}({\mathcal{H}}_{i})$ and since $\mathcal{M}$ is atomic, every measure $\mu_{i}$ occurring in the above sum is atomic. So, without loss of generality, we write $\mathcal{M}=l^{\infty}-\oplus_{i\in I} l^{\infty}(\mu_{i}) {\overline{\otimes}}\mathcal{B}({\mathcal{H}}_{i}).$ By uniqueness of preduals of von Neumann algebras, we have $\mathcal{M}_{*}=l^{1}-\oplus_{i\in I} l^{1}(\mu_{i}) {\overline{\otimes}}\mathcal{T}({\mathcal{H}}_{i})= l^{1}-\oplus_{i\in I} l^{1}(\mu_{i} ,\mathcal{T}({\mathcal{H}}_{i})).$ Hence, by a diagonalization argument, $\mathcal{M}_{*}$ can be embedded isometrically into a closed subspace of $\mathcal{T}(l^{2}-\oplus_{i\in I} l^{2}(\mu_{i})\otimes {\mathcal{H}}_{i})$. By [@Randrian Theorem 3.6], $\mathcal{T}(\mathcal{H})$, for any Hilbert space $\mathcal{H}$, has weak fpp for left reversible semigroups. Hence, the result follows. $\blacksquare$ The following example shows one must deal with non-expansive actions in BGC. Recall that the non-amenable group $SL_{2}(\mathbb{R})$ can act on the unit sphere $\mathbb{S}$ of ${\mathbb{R}}^3$ by the Möbius transformations. Besides, the unit sphere in ${\mathbb{R}}^3$ is the one point compactification of the complex plane with south pole corresponds to zero and north pole corresponds to $\infty$. The action of the group $SL_{2}(\mathbb{R})=\{ \begin{pmatrix} a&b\\ c&d \end{pmatrix} : a, b, c, d \in \mathbb{R}; ad-bc=1 \}$ on $\mathbb{S}$ by Möbius transformations is separately continuous and fixed point free while each member of $SL_{2}(\mathbb{R})$ has a fixed point. [**Proof.**]{} The action is given by $$\begin{aligned} &SL_{2}(\mathbb{R})\times \mathbb{S}\longrightarrow \mathbb{S}\\ &{\begin{pmatrix} a&b\\ c&d \end{pmatrix}} \times {z}\longrightarrow \frac{az+b}{cz+d}\end{aligned}$$ Obviously, each member of the group defines a continuous ( not non-expansive, not affine) mapping (called Möbius transformation) on the compact set $\mathbb{S}$. So, the action is separately continuous. By Schauder’s fixed point theorem, the action of each member of the group has a fixed point. Also by the very formula of each transformation, one can find the fixed points directly. By an easy calculation, one can find two transformations without a common fixed point. Hence, the action does not have a common fixed point.$\blacksquare$ The next example, uses a semigroup which is right reversible but not left reversible. As a result, BGC is not true for right reversible semigroups. Let $S=\{f, g\}$ be the discrete semigroup with the multiplication defined by $ab=a$ for all $a,b\in S$. Obviously, $S$ is not left reversible. Let $\alpha$ and $\beta$ be two real numbers with $0<\alpha<\beta<1$. Put $K=\{(a_n): (a_n)\in c_0(\mathbb{Z}) , \alpha\leq a_0\leq\beta; \quad a_{n}=0 \quad \text{for all} \quad n\neq 0\}$. By an easy observation, we see that $K$ is convex and weakly closed set in $c_0(\mathbb{Z})$. By James’ theorem [@James], one can check that $K$ is also weakly compact. Now, consider the action $S\times K\longrightarrow K$ defined by $f((a_n))=(...,0,\alpha,0,...)$ and $g((a_n))=(...,0,\beta,0,...)$ where $\alpha$ and $\beta$ are placed at the zero location. Obviously, each member of $S$ defines a non-expansive mapping on $K$, so the action is separately continuous and non-expansive. On the other hand, $c_0(\mathbb{Z})$ has weak fpp by [@Lau-Mah-Ulger Corolarry 4.3]. Either way, $c_0(\mathbb{Z})=C^{*}(\mathbb{T})$ is a closed C\* sub-algebra of $\mathcal{K}(L^{2}(\mathbb{T}))$ and the ideal of compact operators has weak fpp by [@Dowling] and [@Garcia]. So, $c_0(\mathbb{Z})$ has weak fpp. Hence, each function in $S$ has a fixed point on $K$. But, there is no common fixed point for the action. Remember that, $S$ is not left reversible. We would like to conclude our work with the following questions: [**Question 1.**]{} Is there any counterexample of a non-expansive action to BGC? [**Question 2.**]{} Is BGC true for amenable semigroups? The norm of C\*-algebras are completely determined by the algebraic properties of the space. Inspired by Theorem \[vn\], we ask: [**Question 3.**]{} Is BGC true for non-unital C\*-algebras? From Theorem \[vn\], we know the preduals of von Neumann algebras enjoy BGC property. Now, we ask: [**Question 4.**]{} Can we characterize all Banach spaces with BGC property? [**Question 5.**]{} Can one prove Theorem \[main.thm\] for weak\* fpp without using the second countability assumption on $G$? [**Acknowledgment.**]{} The author is much indebted to Professor Gero Fendler and would like to thank him for his invaluable suggestions and remarks. He would also like to thank Professor Anthony To-Ming Lau for his instructions and hints which made this paper better.\ [1]{} D. E. Alspach, [*A fixed point free non-expansive map,* ]{} Proc. Amer. Math. Soc. 82 (3) (1981), 423-424. M. Amini, A. R. Medghalchi and F. Naderi, [*Pointwise eventually non-expansive action of semi-topological semigroups and fixed points,* ]{} J. Math. Anal. Appl., 437 (2016), 1176-1183. L. Baggett, [*A separable group having a discrete dual is compact,*]{} J. Funct. Anal. 10 (1972), 131-148. R. E. Bruck Jr., [*A common fixed point theorem for a commuting family of non-expansive mappings,*]{} Pacific J. Math. 53 (1974), 59-71. J. B. Conway, [*A Course in Functional Analysis,*]{} Springer-Verlag, New York, 1990. J. Diestel, J. J. Uhl, Jr., [*Vector Measures,*]{} Amer. Math. Soc., Providence, 1977. J. Dixmier , [*C\*-algebras,*]{} North-Holland Math., Amesterdam, 1977. P. N. Dowling, N. Randrianantoanina, and B. Turett, [*The fixed point property via dual space properties,*]{} J. Funct. Anal. 255 (3) (2008), 768-775. D. van Dulst, B. Sims, [*Fixed points of nonexpansive mappings and Chebyshev centers in Banach spaces with norms of type (KK),*]{} in: Banach Space Theory and Its Applications, Proceedings Bucharest, 1981, in: Lecture Notes in Math., vol. 991, Springer-Verlag, 1983. M. Edelstein, [*The construction of asymptotic center with a fixed point property,*]{} Bull. Amer. Math. Soc. 78 (1972), 206-208. P. Eymard, [*L’ lgebre de Fourier d’un groupe localement compact,*]{} Bull. Soc. Math. France 92 (1964) 181-236. G. Fendler, A. T. Lau, and M. Leinert, [*Weak\* fixed point property and asymptotic center for the Fourier-Stieltjes algebra of a locally compact group,*]{} J. Funct. Anal. 264 (1) (2013), 288-302. G. Fendler and M. Leinert, [*Separable C\*-algebras and weak\* fixed point property,*]{} Probab. Math. Statist. 33 (2) (2013), 233-241. G. Folland, [*A Course in Abstract Harmonic Analysis,*]{} CRC Press, 2015. J. Garcia-Falset, [*The fixed point property in Banach spaces with the NUS-property,*]{} J. Math. Anal. Appl. 215 (2) (1997) 532-542. R. D. Holmes and A. T. Lau, [*Asymptotically non-expansive actions of topological semigroups and fixed points,*]{} Bull. London. Math. Soc., 3 (1971), 343-347. R. C. James,[*Weakly compact sets,*]{} Trans. Amer. Math. Soc., 113 (1964) 129-140. H. E. Jensen, [*Scattered C∗-algebras II,*]{} Math. Scand. 43 (2) (1978), 308-310. A. T. Lau and M. Leinert, [*Fixed point property and the Fourier algebra of a Locally compact groups,*]{} Trans. Amer. Math. Soc. 360 (2) (2008), 6389-6402. A. T. Lau and P. Mah, [*Fixed point property for Banach algebras associated to Locally compact groups,*]{} J. Funct. Anal. 258 (2) (2010), 357-372. A. T. Lau and P. Mah, [*Normal structure in dual Banach spaces associated with a locally compact group,*]{} Trans. Amer. Math. Soc., 310 (1) 1988, 341-353. A. T. Lau, P. Mah and A. Ülger, [*Fixed point property and normal structure for Banach spaces associated to Locally compact groups,*]{} Proc. Amer. Math. Soc. 125 (7) (1997), 2021-2027. A. T. Lau and A. Ülger, [*Some geometric properties on the Fourier and Fourier-Stieltjes algebras of Locally compact groups, Arens Regularity and Related problems,*]{} Trans. Amer. Math. Soc. 337 (1) (1993), 321-359. C. Lennard, [*${\mathcal{C}}_{1}$ is uniformly Kadec-Klee,*]{} Proc. Amer. Math. Soc. 109 (1990), 71-77. T. C. Lim, [*Asymptotic centers and non-expansive mapping in conjugate Banach spaces,*]{} Pacific J. Math. 90 (1980), 135-143. T. C. Lim, [*Characterization of normal structure*]{} , Proc. Amer. Math. Soc. 43 (1974), 313-319. T. Miao, [*The isolated points of $\hat{G}_{\rho}$ and w\*-strongly exposed points of $P_{\rho}(G)_{0}$,*]{} J. London. Math. Soc. 65 (2) (2002), 693-704. G.J. Murphy, [*C\*-algebras and Operator Theory,*]{} Academic Press Inc., 1990. A. L. Paterson, [*Amenability*]{}, American Mathematical Society, Providence, 1988. N. Randrianantoanina, [*Fixed point properties of semigroups of non-expansive mappings,*]{} J. Funct. Anal. 258 (11) (2010), 3801–3817. M. Takesaki, [*Theory of Operator Algebras I,*]{} Springer-Verlag, New York, 2002. K. F. Taylor, [*Geometry of the Fourier algebra and Locally compact groups with atomic unitary representations,*]{} Math. Ann. 262 (1983), 183-190.

ArXiv

--- abstract: 'Computing the Newton step for a generic function ${{\mathpzc{f}}}: \mathbb{R}^N \rightarrow \mathbb{R}$ takes $O(N^{3})$ flops. In this paper, we explore avenues for reducing this bound, when the computational structure of $f$ is known beforehand. It is shown that the Newton step can be computed in time, linear in the size of the computational-graph, and cubic in its tree-width.' author: - | [**Akshay Srinivasan**]{} [^1]\ Dept. of Computer Science & Engineering,\ University of Washington\ Seattle, WA 98195\ Dept. of Computer Science & Engineering,\ University of Washington\ Seattle, WA 98195\ bibliography: - 'references.bib' title: Graphical Newton --- Introduction ============ Newton’s method forms the basis for many second-order methods in Nonlinear-optimization; it is also the core technique used in Interior point methods. It’s applicability to large-scale programming, however, is often limited due to the run-time complexity in computing the Newton step. For a generic function ${{\mathpzc{f}}} : \mathbb{R}^N \rightarrow \mathbb{R}$, computing the Hessian requires atleast $O(N^2)$ flops; further inverting the matrix requires $O(N^{\gamma})$ flops ($\gamma = 3$, in practice). This is computationally infeasible for many problems in practice. Often, however, one is also given access to the the *computational structure* of the objective. The computer routine for calculating the objective ${{\mathpzc{f}}}(\cdot)$ can be represented as a Directed Acyclic Graph \[DAG\] mapping inputs to ${{\mathpzc{f}}}(\cdot)$ *via* intermediary nodes. For instance, the objective function for the canonical optimal-control problem is given by, $$\label{eqn:optcon} \begin{aligned} &\min_{u_0, u_1, \dots, u_{n - 1}} \left[{{\mathpzc{J}}}(u_0, \dots, u_n) \triangleq \sum_{i = 0}^{n - 1} {{\mathpzc{l}}}_i(x_i, u_i) + {{\mathpzc{l}}}_n(x_n)\right],\\ &\quad \forall i, x_{i + 1} \leftarrow \mathtt{f}(x_i, u_i), \end{aligned}$$ where the dynamics and local-objectives of the system are given by $\mathtt{f}(\cdot, \cdot)$, and ${{\mathpzc{l}}}_i(\cdot, \cdot)$ respectively. The infix operator ’$\leftarrow$’ indicates that the value appearing on the right-hand side, is given the placeholder symbol present to its left; we explicitly distinguish this from the ’$=$’ operator, which is taken to represent a constraint. \[fig:optcon-directed\] =\[circle, minimum size=0pt,inner sep=0pt\] /in [0/0, 1/1, 2/2, 3/3, n-1/6]{} (U-) at (,0) [$u_{\name}$]{}; /in [1/1, 2/2, 3/3, n-1/6, n/7]{} (X-) at (, 25pt) [$x_{\name}$]{}; /in [1/2, 2/3, n-1/n]{} [ (X-) edge (X-); ]{} (X-3) – (4, 25pt); (X-3) – (4, 25pt); (5, 25pt) – (X-n-1); /in [0/1, 1/2, 2/3, n-1/n]{} [ (U-) – (X-); ]{} (U-3) – (4, 25pt); (4, 0) – (5, 0); The order of computation for the objective [(\[eqn:optcon\])]{} can be represented by a linear-chain [(Figure \[fig:optcon-directed\])]{}. Lacking constraints, the apparent sparsity in [(\[eqn:optcon\])]{}, is entirely destroyed once all the placeholders are substituted for, (u\_0, …, u\_n) = &\_0(x\_0, u\_0) +\ &\_1((x\_0, u\_0), u\_1) +\ &\_2(((x\_0, u\_0), u\_1), u\_2) + …. The Hessian of ${{\mathpzc{J}}}(\cdot)$ thus being dense, implies a run-time that is cubic in the input dimensions for the Newton step computation; computing the Hessian itself is quadratic. By contrast, once the problem [(\[eqn:optcon\])]{} is written in its constrained form (by replacing ’$\leftarrow$’ with ’$=$’), the sparsity of the resulting Karush-Kuhn-Tucker \[KKT\] system, readily allows for computing the SQP/Lagrange-Newton step in linear time [@wright1990solution]. Such a transformation, however, comes at the cost of increasing the size of the optimization problem, abandoning state feasibility, and increased implementation complexity. The question which this paper answers, is whether there exist *general* techniques, which allow exploiting the sparsity of the problem, while working solely with the input variables. Note that these are not questions merely about elimination orders, but are also verily algebraic in nature. <span style="font-variant:small-caps;">Automatic Differentiation</span>: Research on Automatic Differentiation \[AD\] has produced many techniques for exploiting the *computational structure* of generic functions. They are routinely employed for efficient calculation of gradients and Hessian vector products [@griewank2008evaluating]. The applicability of AD to second-order optimization is, however, quite limited. AD is typically used either for computing the entire Hessian matrix, or for calculating Hessian vector products for use in Nonlinear Conjugate Gradient Descent \[CG\]. Hessians are computed, by accumulating one column at a time *via* calls to the Hessian vector product routine [@griewank2008evaluating]. The sparsity of the Hessian can be exploited in reducing the number of such calls [@coleman1984estimation], but structured problems such as [(\[eqn:optcon\])]{} will not allow for any such economy. Compositional chains of functions, such as those in optimal control [(\[eqn:optcon\])]{} [(Figure \[fig:optcon-directed\])]{}, not only serve to make Hessians dense but can also lead to condition numbers, exponential in their diameters. Large condition numbers are likely to negate any computational advantages offered by methods like CG. The above techniques form the Hessian matrix, directly or indirectly, before computing the Newton step. This stands in contrast with the root-finding problem, for which there do exist methods for directly computing the Newton-Raphson step [@utke1996efficient] [@griewank2008evaluating] [@dixon2009automatic]. The root finding problem involves the inversion of the Jacobian of a function - rather than the Hessian - and these methods reduce this computation to that of inverting a sparse matrix [@utke1996efficient]. The Newton step (for optimization) can also be computed using this method by formulating it as a root-finding problem on the gradient. This, however, results in non-symmetric matrices depending on the computational graph of the gradient, as opposed to the function itself. The latter graph is transitively closed, and hence, analyses for the above Newton-Raphson AD algorithm apply to the gradient, and are difficult to extend to the underlying objective [@dixon2009automatic]. <span style="font-variant:small-caps;">Dynamic Programming</span>: The question posed earlier, has already been answered in the affirmitive, for the optimal control problem. There exists an algorithm for optimal control, based on Dynamic Programming, that exploits algebraic dependencies in [(\[eqn:optcon\])]{}, in order to compute the Newton-step in only linear time [@de1988differential] [@jacobson1970differential] [@todorov2005generalized]. The run-time of this algorithm is the direct result of the sparsity of the corresponding constrained problem [@de1988differential] [@wright1990solution][@ralph1996parallel]. The band-structure of the relevant KKT system allows for solving the system in linear time [@wright1990solution]. The relationship between computing the Newton-step (Hessian of the objective), and computing the Lagrange-Newton step (Hessian of the Lagrangian), is established by noting that there exist multiplier values such that both compute the *same* result [@de1988differential]. Such algorithms are routinely employed by practitioners for updating *control policies* in real-time, while maintaining a feasible trajectory. These algorithms have been extended to Extended Kalman Filtering (EKF), as well as various other formulations of the control problem [@toussaint2010bayesian] [@tassa2011optimal] [@bell2009inequality] [@wright1993interior]. <span style="font-variant:small-caps;">Overview</span>: We generalize such algorithms, by using Hessian vector product equations from AD, to relate the computation of Newton step and Lagrange-Newton step, for arbitrary structured objectives. We then extend this framework to structured optimization problems with equality constraints. Further, we show that solving the resultant KKT systems can be accomplished in time $\tilde{O}(\operatorname{tw}^3)$, where ’$\operatorname{tw}$’ is the tree-width of the canonical computational graph. Finally, we show results from numerical experiments. Notation ======== Let $\mathcal{G}$ be a Directed Acyclic Graph \[DAG\], and let each vertex $v \in V[\mathcal{G}]$, be associated with *state* ${\mathtt{S}_{v}} \in U_v \subset \mathbb{R}^{n_v}$, taking values in an open set. Denote by $\operatorname{\delta^{+}}(v)$, the parents of $v \in V[\mathcal{G}]$, and by $\operatorname{\delta^{-}}(v)$ its children; let ${\mathtt{S}_{A}}$ be the (labelled) concatenation of *state*s, associated with vertices in set $A \subset V[\mathcal{G}]$. Define the set of *input* nodes $X = \{x_1, x_2,\dots, x_n\} \triangleq \{v \mid \operatorname{\delta^{+}}(v) = \emptyset, v \in V[\mathcal{G}]\}$, to be the parentless vertices of $\mathcal{G}$. An objective function ${{\mathpzc{f}}} : U_{x_1} \times \dots U_{x_n} \rightarrow \mathbb{R}$, has the *computational structure* given by the tuple $(\mathcal{G}, \{\varphi_v\}, \{{{\mathpzc{l}}}_v\})$, if it can be written as the sum of local objectives ${{\mathpzc{l}}}_v : \prod_{z \in \{v\} \cup \operatorname{\delta^{+}}(v)} U_z \rightarrow \mathbb{R}$, on the graph $\mathcal{G}$, $$\label{eqn:struct} \begin{aligned} {{\mathpzc{f}}}: ({\mathtt{S}_{x_1}}, \dots, {\mathtt{S}_{x_n}})& \mapsto \sum_{v \in V[\mathcal{G}]} {{\mathpzc{l}}}_v({\mathtt{S}_{v \cup \operatorname{\delta^{+}}(v)}}),\\ {\mathtt{S}_{v}} \leftarrow \varphi_v({\mathtt{S}_{\operatorname{\delta^{+}}(v)}})&, \quad \forall v \in V[\mathcal{G}], \operatorname{\delta^{+}}(v) \neq \emptyset. \end{aligned}$$ The *state* of a non-input node $v \in V[\mathcal{G}]$ in [(\[eqn:struct\])]{}, is defined recursively as ${\mathtt{S}_{v}} \leftarrow \varphi_v({\mathtt{S}_{\operatorname{\delta^{+}}(v)}})$, for some given function $\varphi_v : \prod_{z \in \operatorname{\delta^{+}}(v)} U_z \rightarrow U_v$. It follows since $\mathcal{G}$ is a DAG, that ${\mathtt{S}_{V[\mathcal{G}]}}$ and hence ${{\mathpzc{f}}}(\cdot)$, is uniquely determined from the input ${\mathtt{S}_{X}}$, and functions $\{\varphi_v\}$. The order of computation for the objective is given by the topological ordering of $\mathcal{G}$, and the DAG $\mathcal{G}$ is called the *computational graph* of ${{\mathpzc{f}}}(\cdot)$. The computer routine for calculating any objective function, can be represented by such a structure [@griewank2008evaluating]. In the following sections, the symbolism ${\partial}_u v$ is used as a shorthand for ${{\partial}{\mathtt{S}_{v}} \over {\partial}{\mathtt{S}_{u}}} \big|_{{\mathtt{S}_{X}}}$. The derivatives of functions with respect to ${\mathtt{S}_{u}}$ are similarly denoted by the operator ${\partial}_{u}$; that with respect to a (labelled) set $A = \{v_1, v_2, \dots\} \subset V[\mathcal{G}]$ by ${\partial}_{A} \triangleq [{\partial}_{a_1}, {\partial}_{a_2}, \dots]$. Newton step {#sec:newt} =========== Consider the objective function in [(\[eqn:struct\])]{}, defined by the tuple $(\mathcal{G}, \{\varphi_v\}, \{{{\mathpzc{l}}}_v\})$. The optimization problem of interest is the following, $$\label{eqn:obj} \begin{aligned} \min_{{\mathtt{S}_{x_1}}, \dots, {\mathtt{S}_{x_n}}} &\left( {{\mathpzc{f}}} \triangleq \sum_{v \in V[\mathcal{G}]} {{\mathpzc{l}}}_v({\mathtt{S}_{v \cup \operatorname{\delta^{+}}(v)}}) \right),\\ {\mathtt{S}_{v}} \leftarrow \varphi_v({\mathtt{S}_{\operatorname{\delta^{+}}(v)}}),& \quad \forall v \in V[\mathcal{G}], \operatorname{\delta^{+}}(v) \neq \emptyset, \end{aligned}$$ and the corresponding constrained problem is obtained by replacing the operator ’$\leftarrow$’ by ’$=$’ in [(\[eqn:obj\])]{}. In the following, we consider first the constrained formulation of [(\[eqn:obj\])]{}, and define the KKT system involved in computing the Lagrange-Newton step; we then relate these to computing the Newton step. Lagrange-Newton --------------- The Lagrangian for the constrained form of (\[eqn:obj\]), is given by, $$\label{eqn:cdef} \begin{aligned} &\mathcal{L}({\mathtt{S}_{V[\mathcal{G}]}}, \lambda) \triangleq\\ &\quad\quad\sum_{v \in V[\mathcal{G}]} {{\mathpzc{l}}}_v({\mathtt{S}_{v \cup \operatorname{\delta^{+}}(v)}}) + \sum_{\substack{v \in V[\mathcal{G}],\\ \operatorname{\delta^{+}}(v) \neq \emptyset}} \lambda_v^{\mathtt{T}} h_v({\mathtt{S}_{v \cup \operatorname{\delta^{+}}(v)}}),\\ &\mbox{where,} \\ &\forall v \in V[\mathcal{G}], \operatorname{\delta^{+}}(v) \neq \emptyset, \quad h_v({\mathtt{S}_{v \cup \operatorname{\delta^{+}}(v)}}) \triangleq \varphi_v({\mathtt{S}_{\operatorname{\delta^{+}}(v)}}) - {\mathtt{S}_{v}}, \end{aligned}$$ and the vector $\lambda$ is the labelled concatenation of all $\lambda_v$’s. The necessary first order conditions for optimality of this problem are given by [@nocedal2006numerical], $$\label{eqn:lagrangefirst} {\partial}_{V} \mathcal{L}({\mathtt{S}_{V}}^*, \lambda^*) = 0,\quad h({\mathtt{S}_{V}}^*) = 0.$$ The Lagrange-Newton step for solving this system of equations, around a nominal $(\tilde{{\mathtt{S}_{V}}}, \lambda)$, entails solving the following KKT system [@nocedal2006numerical], $$\label{eqn:clnstep} \left[\begin{array}{c c} {\partial}^2_{V} \mathcal{L} & {\partial}_{V} h^{\mathtt{T}} \\ {\partial}_{V} h & 0 \end{array}\right] \left[\begin{array}{c} \delta {\mathtt{S}_{V}} \\ \delta \lambda \end{array}\right] = \left[\begin{array}{c} - {\partial}_{V} \mathcal{L} \\ -h \end{array} \right].$$ Sequential Quadratic Programming \[SQP\], involves taking a step along $(\delta {\mathtt{S}_{V}}, \delta \lambda)$ and iteratively solving for the first order conditions. In the following section, it will be shown that there exist values for Lagrange multipliers, depending only on the inputs, such that the solution to [(\[eqn:clnstep\])]{}, yields the Newton step for the unconstrained objective. Unconstrained Newton -------------------- We recollect certain defintions from AD, and then continue to present one of the central results of the paper. <span style="font-variant:small-caps;">Reverse AD</span>: The first derivatives of the objective ${{\mathpzc{f}}}(\cdot)$ can be calculated by applying the chain rule over $\mathcal{G}$, $$\label{eqn:rmad} \begin{aligned} &\forall v,\quad {\partial}_v {{\mathpzc{f}}} = \sum_{s \in v \cup \operatorname{\delta^{-}}(v)} {\partial}_v {{\mathpzc{l}}}_s + \sum_{d \in \operatorname{\delta^{-}}(v)} {\partial}_d {{\mathpzc{f}}}^{\mathtt{T}} \; {\partial}_v d;\\ &\quad\quad\quad v \in \operatorname{\delta^{+}}(d) \Rightarrow {\partial}_v d \triangleq {{\partial}\varphi_d({\mathtt{S}_{\operatorname{\delta^{+}}(d)}}) \over {\partial}{\mathtt{S}_{v}}}. \end{aligned}$$ Since $\mathcal{G}$ is a DAG, there exist child-less nodes (*i.e* $\operatorname{\delta^{-}}(v) = \emptyset$), from which the above recursion can be initialized. The recursion then proceeds backward in the depth first search order on $\mathcal{G}$. This algorithm is known as reverse-mode AD [@griewank2008evaluating]. <span style="font-variant:small-caps;">Hessian vector AD</span>: A change in the inputs $\delta {{\mathtt{S}_{X}}}$, results in the first-order change in the derivative, $\delta [{\partial}_v {{\mathpzc{f}}}] \triangleq {\partial}_{X v}^2 {{\mathpzc{f}}} \cdot \delta {\mathtt{S}_{X}}$, which is given by the Hessian vector product. Computing the Newton step is thus, equivalent to finding a $\delta {{\mathtt{S}_{X}}}$ such that, $\delta [{\partial}_X {{\mathpzc{f}}}] = - {\partial}_X {{\mathpzc{f}}}$. Applying chain-rule over the DAG $\mathcal{G}$, for all terms in [(\[eqn:rmad\])]{}, we obtain, $$\label{eqn:armad} \begin{aligned} &\forall v,\quad \delta [{\partial}_v {{\mathpzc{f}}}] =\\&\quad\quad \sum_{s \in v \cup \operatorname{\delta^{-}}(v)} \left(\sum_{a \in v \cup \operatorname{\delta^{+}}(s)} {\partial}_{av}^2 {{\mathpzc{l}}}_s \cdot \delta {\mathtt{S}_{a}} \right) +\\ &\quad\quad \sum_{d \in \operatorname{\delta^{-}}(v)} \left(\delta [{\partial}_d {{\mathpzc{f}}}]^{\mathtt{T}} {\partial}_v d + \sum_{a \in \operatorname{\delta^{+}}(d)} ({\partial}_d {{\mathpzc{f}}}^{\mathtt{T}} \; {\partial}_{av}^2 d) \cdot \delta {\mathtt{S}_{a}})\right);\\ &\forall a,\quad \delta {\mathtt{S}_{a}} = \sum_{d \in \operatorname{\delta^{+}}(a)} {\partial}_{d} a \cdot \delta {\mathtt{S}_{{d}}}. \end{aligned}$$ These equations can be solved, for a given $\delta {\mathtt{S}_{X}}$, by a forward-backward recursion similar to the one used for solving [(\[eqn:rmad\])]{} [@griewank2008evaluating]. Computing the Hessian-vector product in this manner takes time $\tilde{O}(\omega(\mathcal{\hat{G}})^2)$ [^2] [@griewank2008evaluating], where $\omega(\mathcal{\hat{G}})$ is the clique number of the moralization of $\mathcal{G}$. <span style="font-variant:small-caps;">Newton step</span>: The problem of interest is, however, the exact inverse: find a $\delta {\mathtt{S}_{X}}$, such that $\delta[{\partial}_X {{\mathpzc{f}}}] = -{\partial}_X {{\mathpzc{f}}}$. This question is answered by the following theorem. \[theo:gnewt\](Newton step) The Newton step for the objective [(\[eqn:struct\])]{} is given by the Lagrange-Newton step [(\[eqn:clnstep\])]{}, when ${\mathtt{S}_{V}}$ is feasible and when $\forall v, \lambda_v = {\partial}_v {{\mathpzc{f}}}$ as defined in [(\[eqn:rmad\])]{}. *Proof.* The second equation in [(\[eqn:armad\])]{} is equivalent to ${\partial}_V h \cdot \delta S_V = -h$, in [(\[eqn:clnstep\])]{}. Rearranging the first equation from [(\[eqn:armad\])]{}, and setting $\delta[{\partial}_v {{\mathpzc{f}}}] = -{\partial}_v {{\mathpzc{f}}}$ for all inputs, we obtain $\forall v$, $$\begin{aligned} \label{eqn:ppobj} 0 =&\\ &\sum_{\substack{s \in v \cup \operatorname{\delta^{-}}(v),\\ a \in v \cup \operatorname{\delta^{+}}(s)}} {\partial}_{va}^2 {{\mathpzc{l}}}_s \;\delta {\mathtt{S}_{a}} + \sum_{\substack{d \in \operatorname{\delta^{-}}(v),\\ a \in \operatorname{\delta^{+}}(d)}} ({\partial}_d {{\mathpzc{f}}}^{\mathtt{T}}\; {\partial}_{av}^2 d) \; \delta {\mathtt{S}_{a}} +\\ & - \left(\begin{cases} \delta [{\partial}_v {{\mathpzc{f}}}] & \operatorname{\delta^{+}}(v) \neq \emptyset\\ - {\partial}_v {{\mathpzc{f}}} & \mbox{otherwise}\end{cases}\right) + \sum_{d \in \operatorname{\delta^{-}}(v)} ({\partial}_v d)^{\mathtt{T}} \delta [{\partial}_d {{\mathpzc{f}}}]. \end{aligned}$$ Similarly, expanding the top block in [(\[eqn:clnstep\])]{} using the definitions in [(\[eqn:obj\])]{} & [(\[eqn:clnstep\])]{}, we obtain $\forall v$, $$\begin{aligned} \label{eqn:ppLexp} &-{\partial}_{v} \mathcal{L} =\\ &\sum_{\substack{s \in v \cup \operatorname{\delta^{-}}(v),\\a \in v \cup \operatorname{\delta^{+}}(s)}} {\partial}_{va}^2 {{\mathpzc{l}}}_s \; \delta {\mathtt{S}_{a}} + \sum_{\substack{d \in \operatorname{\delta^{-}}(v),\\a \in \operatorname{\delta^{+}}(d)}} (\lambda_d^{\mathtt{T}} \; {\partial}_{av}^2 d) \; \delta {\mathtt{S}_{a}} +\\ & - \left(\begin{cases} \delta \lambda_v & \operatorname{\delta^{+}}(v) \neq \emptyset \\ 0 & \mbox{otherwise}\end{cases} \right) + \sum_{d \in \operatorname{\delta^{-}}(v)} ({\partial}_v d)^{\mathtt{T}} \; \delta \lambda_d , \end{aligned}$$ where, $$\begin{aligned} \label{eqn:pLexp} &{\partial}_{v} \mathcal{L} = \\ &\begin{cases} \sum_{s \in v \cup \operatorname{\delta^{-}}(v)} {\partial}_v {{\mathpzc{l}}}_s + \sum_{d \in \operatorname{\delta^{-}}(v)} \lambda_d \cdot {\partial}_{v} d, & \operatorname{\delta^{+}}(v) = \emptyset\\ \begin{array}{l} \sum_{s \in v \cup \operatorname{\delta^{-}}(v)} {\partial}_v {{\mathpzc{l}}}_s + \sum_{d \in \operatorname{\delta^{-}}(v)} \lambda_d \cdot {\partial}_{v} d \\- \lambda_v \end{array}, & \mbox{otherwise} \\ \end{cases} \end{aligned}$$ The result follows from equations [(\[eqn:rmad\])]{}, [(\[eqn:ppobj\])]{}, [(\[eqn:ppLexp\])]{} & (\[eqn:pLexp\]). [$\hfill \square$]{} <span style="font-variant:small-caps;">Graphical Newton</span>: The above theorem immediately yields the following optimization algorithm, initial ${\mathtt{S}_{X}}$, tuple $(\mathcal{G}, \{\varphi_v\}, \{{{\mathpzc{l}}}_v\})$ Compute ${{\mathpzc{f}}}, \{{\partial}_v {{\mathpzc{f}}}\}, \{{\partial}^2 \varphi_v\}$ from (\[eqn:struct\]), (\[eqn:rmad\]). Compute the SQP step from (\[eqn:clnstep\]), with $\lambda_v = {\partial}_v {{\mathpzc{f}}}, \forall v$. Compute step-length $\eta$ *via* linesearch on inputs ${\mathtt{S}_{X}}$. Update inputs: ${\mathtt{S}_{X}} \leftarrow {\mathtt{S}_{X}} + \eta \delta {\mathtt{S}_{X}}$. The run-time of every iteration in Algorithm \[alg:gnewt\] depends crucially upon the time required to solve (\[eqn:clnstep\]). The run-time bounds for solving such KKT systems is taken up later in the paper. Extension to equality constraints --------------------------------- Consider optimization problems, which have equality constraints in addition to the structured objective from before, $$\label{eqn:cobj} \begin{aligned} \min_{{\mathtt{S}_{x_1}}, \dots, {\mathtt{S}_{x_n}}} &\left( {{\mathpzc{f}}} \triangleq \sum_{v \in V[\mathcal{G}]} {{\mathpzc{l}}}_v({\mathtt{S}_{v \cup \operatorname{\delta^{+}}(v)}}) \right),\\ {\mathtt{S}_{v}} \leftarrow \varphi_v({\mathtt{S}_{\operatorname{\delta^{+}}(v)}}),& \quad \forall v \in V[\mathcal{G}], \operatorname{\delta^{+}}(v) \neq \emptyset,\\ & c({\mathtt{S}_{C}}) = 0, \end{aligned}$$ where $c(\cdot) = 0$ is an additional equality constraint, which depends on the variables $C \subset V[\mathcal{G}]$. The Lagrangian for this problem is given by, $$\label{eqn:cdef} \hat{\mathcal{L}}({\mathtt{S}_{V[\mathcal{G}]}}, \lambda) = \mathcal{L}({\mathtt{S}_{V[\mathcal{G}]}}, \lambda_{V \backslash X}) + \lambda_h^{\mathtt{T}} c({\mathtt{S}_{C}}),$$ where $\mathcal{L}$ is as defined in [(\[eqn:cdef\])]{}, and $\lambda_{V \backslash X}$ is the corresponding set of multipliers; the variable $\lambda$, being the concatenation of $\lambda_c$ and all multipliers, $\lambda_{V\backslash X}$, appearing in [(\[eqn:cdef\])]{}. Theorem \[theo:gnewt\] can be applied to this problem by treating $\lambda_c^T c({\mathtt{S}_{C}})$ as another cost function in the objective, while also including the constraint in the KKT system [(\[eqn:clnstep\])]{}. The iteration can then proceed by solving the KKT system with $\lambda_v = {\partial}_v ({{\mathpzc{f}}} + \lambda_c^T c), \forall v$, and using a merit function for the linesearch procedure; the variables $({\mathtt{S}_{X}}, \lambda_c)$ are updated accordingly. We omit the proof for the validity of this method. Message Passing =============== The classical run-time bound for Cholesky factorization (*i.e* Gaussian Belief Propagation [^3]) [@davis2006direct] [@wainwright2008graphical], cannot be extended to problems such as [(\[eqn:clnstep\])]{}, because of the appearance of linear constraints. Such bounds for structured KKT systems, do not appear to be known within the sparse linear algebra community [@bridson2007ordering]. In this section, we provide a Message Passing algorithm for solving such KKT systems, and show that it has a run-time bound of $\tilde{O}(\operatorname{tw}^3)$ [^4], given the tree-decomposition. Hypergraph structured QPs ------------------------- For a hypergraph $\mathcal{H}$, denote the adjacency and incidence matrices by $\mathcal{A}[\mathcal{H}]$ & $\mathcal{B}[\mathcal{H}]$ respectively, $$\begin{aligned} &\mathcal{A}[\mathcal{H}] \in \mathbb{R}^{|V[\mathcal{H}]|\times|V[\mathcal{H}]|}, \quad\quad \mathcal{B}[\mathcal{H}] \in \mathbb{R}^{|E[\mathcal{H}]|\times|V[\mathcal{H}]|},\\ &\mathcal{A}[\mathcal{H}]_{uv} = \begin{cases} 1 & \exists e \in E[\mathcal{H}],\; u, v \in e\\ 0 & \mbox{otherwise}\end{cases}\\ &\mathcal{B}[\mathcal{H}]_{eu} = \begin{cases} 1 & u \in e\\ 0 & \mbox{otherwise}\end{cases} \end{aligned}$$ Given such a hypergraph $\mathcal{H}$, the family of QPs we’re interested in solving is the following, $$\label{eqn:spqpfamily} \begin{aligned} &\min_{x} \sum_{e \in E[\mathcal{H}]}{1 \over 2} {\mathtt{S}_{e}}^{\mathtt{T}} Q_e {\mathtt{S}_{e}} - b_e^{\mathtt{T}} {\mathtt{S}_{e}}, \\ &\forall e \in E[\mathcal{H}], \quad G_{e} {\mathtt{S}_{e}} = h_e. \end{aligned}$$ Assuming that the QP has a bounded solution and that the constraints are full rank, the minimizer to (\[eqn:spqpfamily\]) is given by the solution to the following KKT system, $$\label{eqn:spfamily} \begin{aligned} \left[\begin{array}{c c} Q & G^{\mathtt{T}} \\ G & 0\end{array}\right] \left[\begin{array}{c} x \\ \lambda\end{array}\right] &= \left[\begin{array}{c} b \\ h \end{array}\right],\\ x, b \in \mathbb{R}^{|V|}, \quad& \lambda, h \in \mathbb{R}^{M},\\ \end{aligned}$$ where $Q, G, \lambda, x, b$ are concatenation of terms defined in [(\[eqn:spqpfamily\])]{} respectively. The sparsity/support of (\[eqn:spfamily\]) is closely related to $\mathcal{H}$, since, &(Q) (\[\]),\ &i, e, (G\_[i, :]{}) (\[\]\_[e, :]{}). Every row of the constraint, $G_{i, :}$, has the same sparsity as some edge $e \in E[\mathcal{H}]$. <span style="font-variant:small-caps;">Tree decomposition</span>: Extending the notion of Dynamic Programming to non-trees (including Hypergraphs) requires a partitioning of the graph so as to satisfy a *lifted* notion of being a tree [@kleinberg2006algorithm]. Tree decomposition captures the essence of such graph partitions, [(Tree decomposition)]{} A tree-decomposition of a hypergraph $\mathcal{H}$ consists of a tree $\mathcal{T}$ and a map $\chi : V[\mathcal{T}] \rightarrow 2^{V[\mathcal{H}]}$, such that, 1. (*Vertex cover*) $\cup_{i \in V[\mathcal{T}]} \chi(i) = V[\mathcal{H}].$ 2. (*Edge cover*) $\forall e \in E[\mathcal{H}],\; \exists i \in V[\mathcal{T}], e \subset \chi(i).$ 3. (*Induced sub-tree*) $\forall u \in V[\mathcal{H}],\; \mathcal{T}_u \triangleq \mathcal{T}[\{i \in V[\mathcal{T}]| u \in \chi(i)\}]\; \mbox{is a non-empty subtree}$ The tree-width of a tree-decomposition $\mathcal{T}$ is defined to be $\operatorname{tw}(\mathcal{T}) = \max_{v \in V[\mathcal{T}]} |\chi(v)| - 1$. The tree-width of a graph $\mathcal{H}$ is defined to be the minimal tree-width attained by any tree-decomposition of $\mathcal{H}$. We define the vertex-induced subgraph in what follows to be $\mathcal{H}[S] \triangleq (V[\mathcal{H}], \{e \cap S, e\in E[\mathcal{H}]\})$. The following lemma ensures that such a decomposition ensures *local dependence* [@kleinberg2006algorithm]. \[lem:esep\][(Edge separation)]{} Deleting the edge $xy \in E[\mathcal{T}]$, renders $\mathcal{H}[V \backslash (\chi(x) \cap \chi(y))]$ disconnected. <span style="font-variant:small-caps;">Hypertree structured QP</span>: The tree-decomposition itself can be considered a Hypergraph, $(V[\mathcal{H}], \{\chi(u), \forall u \in V[\mathcal{T}]\})$. Such a *Hypertree*[^5] can also be thought of as a Chordal graph [@wainwright2008graphical]. We assume henceforth that the given graph $\mathcal{H}$ is a hypertree, and that $\mathcal{T}$ is its tree-decomposition. : $\mathcal{T}, \mathcal{H}, \{Q_e\}, \{b_e\}, \{G_e\}, \{h_e\}.$ GatherMessage($l, p, \mathcal{T}$) $(\tilde{Q}_l, \tilde{b}_l, \tilde{G}_l, \tilde{h}_l) \leftarrow (Q_l, b_l, G_l, h_l)$ $(Q_{c\rightarrow l}, G_{c\rightarrow l}, b_{c\rightarrow l}, h_{c\rightarrow l}) \leftarrow \mbox{GatherMessage}(c, p, \mathcal{T})$ $(\tilde{Q_l}, \tilde{b_l}) \leftarrow (\tilde{Q_l}, \tilde{b_l}) + (Q_{c\rightarrow l}, b_{c\rightarrow l})$ $\tilde{G}_l \leftarrow [\tilde{G}_l; G_{c\rightarrow l}], \tilde{h}_l \leftarrow [\tilde{h}_l; h_{c\rightarrow l}]$ Factorize($\chi(l), \chi(p), \tilde{Q}_l, \tilde{b}_l, \tilde{G}_l, \tilde{h}_l$) Factorize($\chi(l), \chi(p), \tilde{Q}, \tilde{b}, \tilde{G}, \tilde{h}$) $(\xi, \iota) \leftarrow (\chi(l) \backslash \chi(p), \chi(l) \cap \chi(p))$ $r \leftarrow \operatorname{rank}(\tilde{Q}_{\iota, \iota})$ Gaussian-BP messages from [(\[eqn:mpcore\])]{}. The gather stage of the Message Passing algorithm, is illustrated in Algorithm \[alg:gqp\]. [^6] The function, Factorize, computes the partial LU decomposition of its arguments; we describe below, its operation. Denote the vertices that are interior to $l$ by $\iota = \chi(l) \cap \chi(p)$, and those on the boundary (*i.e* common to $p, l$) by $\xi = \chi(l) \backslash \chi(p)$, and let $r = \operatorname{rank}(\tilde{Q}_{\iota, \iota})$. The function computes Gaussian-BP messages from block pivots $\{2, 3\}$ to $\{1, 4\}$ in [(\[eqn:mpcore\])]{}. Note that, unlike Gaussian-BP, the matrices in [(\[eqn:mpcore\])]{} are not necessarily positive definite, but are however invertible. $$\label{eqn:mpcore} \begin{tikzpicture} \matrix [matrix of math nodes,left delimiter={[},right delimiter={]}] (m) { \tilde{Q}_{\xi \xi} & \tilde{Q}_{\iota \xi}^{\mathtt{T}} & \tilde{G}_{:r, \xi}^{\mathtt{T}} & \tilde{G}_{r:, \iota}^{\mathtt{T}}\\ \tilde{Q}_{\iota \xi} & \tilde{Q}_{\iota \iota} & \tilde{G}_{:r, \iota}^{\mathtt{T}}& \tilde{G}_{r:, \iota}^{\mathtt{T}}\\ \tilde{G}_{:r, \xi} & \tilde{G}_{:r, \iota} & 0 & 0\\ \tilde{G}_{r:, \xi} & \tilde{G}_{r:, \iota} & 0 & 0\\ }; \draw[color=red] (m-2-2.north west) -- (m-2-3.north east) -- +(0, -1.3) -- +(-1.75, -1.3) -- (m-2-2.north west); \matrix [matrix of math nodes,left delimiter={[},right delimiter={]}] (v) at (3, 0) { {\mathtt{S}_{\xi}}\\ {\mathtt{S}_{\iota}}\\ \lambda_{:r}\\ \lambda_{r:}\\ }; \node at (4, 0) {$=$}; \matrix [matrix of math nodes,left delimiter={[},right delimiter={]}] (v) at (5, 0) { \tilde{b}_{\xi}\\ \tilde{b}_{\iota}\\ \tilde{h}_{:r}\\ \tilde{h}_{r:}\\ }; \end{tikzpicture}$$ Gaussian Belief-Propagation is essentially a restatement of LU decomposition [@davis2006direct]. Gaussian-BP consists of messages of the form [@yedidia2000generalized] [@wainwright2008graphical], $$\begin{aligned} \mu_{i \rightarrow j} &:= [J_{i \rightarrow j}, h_{i \rightarrow j}] = [J_{ii}, h_i] - \sum_{k \in \delta(i) \backslash j} J_{ik} J_{k \rightarrow i}^{-1} [J_{ki}, h_{k \rightarrow i}],\\ \mu_i & = J_{i \rightarrow j}^{-1} (h_{i \rightarrow j} - J_{i j} \mu_j), \end{aligned}$$ where $J \mu = h$ is the equation that is to be solved. These can be replaced by appropriate square-root forms to obtain instead, an LDL decomposition. \[theo:twlem\] The linear equation (\[eqn:spfamily\]) can be solved in time $\tilde{O}(\operatorname{tw}(\mathcal{H})^3)$, given the minimal tree-decomposition *via* Algorithm \[alg:gqp\]. *Proof.* The correctness of the algorithm follows from Lemma \[lem:esep\]. The bound holds trivially if, $\operatorname{rank}{\tilde{G}} \leq \operatorname{rank}{\tilde{Q}_{\iota, \iota}}$, at every step of the algorithm. Otherwise, by realizing that $\tilde{G}_{l \rightarrow p}$, can’t have rank more than $|\chi(p)|$, the proof follows. [$\hfill \square$]{} \[fig:sd\] {height="5cm"} It follows from Theorem \[theo:twlem\], that the KKT system in Algorithm \[alg:gnewt\] can be solved in time $\tilde{O}(\operatorname{tw}(\hat{\mathcal{G}})^3)$, where $\hat{\mathcal{G}}$ is the moralization of the computational graph $\mathcal{G}$. The above proof also ensures that the equivalent sparse LU/LDL decomposition [@davis2006direct], with the same pivot order, also has the same run-time. Since decompositions of indefinite systems are subject to instability, use of specialized solvers is generally preferable. Numerical Experiments ===================== In this section, we present preliminary numerical results with an implementation of Algorithm \[alg:gnewt\], using the MA57 solver [@duff2004ma57]. For ensuring convergence in constrained problems, an augmented Lagragian merit function was used [@gill1986some]. The implementation was tested on the following non-standard control problems. <span style="font-variant:small-caps;">Spring-damper limit cycle</span>: Consider the following spring-damper limit-cycle problem [@tassa2011optimal], $$\begin{aligned} &\min_{x_0, u[0, T]} \int \ell(x, \dot{x}, u) {{\mathtt{d} {t}}},\\ &\ddot{x} = -(x^3 + \dot{x}^3)/6 + u,\\ &x(0) = x(T) = x_0, \dot{x}(0) = \dot{x}(T) = \dot{x}_0, \end{aligned}$$ where, $$\ell(x, \dot{x}, u) = (1 - \operatorname{e}^{(\dot{x}_i - 2)^2} - \operatorname{e}^{- (\dot{x}_i + 2)^2}) + {1 \over 2} {\vert \vert {u_i} \vert \vert}^2_2.$$ Discretising the derivatives by finite differences, $\dot{x} \approx \Delta x_i / \Delta t = (x_i - x_{i - 1})/\Delta t$, this can be written as the following structured optimization problem, $$\label{eqn:sdd} \begin{aligned} & \quad\quad \min_{x_0, \{u_i\}_{0}^m} \sum_{1}^N \ell(x_i, \Delta x_i/\Delta t, u_i),\\ &x_{i + 1} \leftarrow x_{i} + \Delta x_{i} + (\Delta t)^2 \left[ -(x^3 + (\Delta x_{i} / \Delta t)^3)/6 + u_i\right].\\ &x_{0} = x_{N - 2}, x_{1} = x_{N - 1}. \end{aligned}$$ For $N = 100, \Delta t = 0.1$, with random initializations, the problem showed robust convergence; often taking no more than ten SQP iterations. The optimal limit cycle, and the convergence curves for one run of the algorithm are shown in (Figure \[fig:sd\]). Discussion ========== We have shown that the Newton step can be computed in time $\tilde{O}(\operatorname{tw}^3)$, where ’$\operatorname{tw}$’ is the tree-width of the computational graph. We have also derived extensions to constrained problems, and provided numerical examples. The technique presented herein, also generalizes many specialized algorithms in control. In certain control problems, the solution to the KKT system, itself can be written in *feedback form*. Given a $LU$ decomposition of the KKT system, one can replace the backsubstitution phase by $U$, with a function evaluation that uses $L$ as a *control feedback* [@jacobson1970differential]. It is unclear if such techniques can be generalized, and whether they can be made independent of the pivot-order used for solving the system. A competing method for exploiting the structure of objectives such as [(\[eqn:obj\])]{}, is by the use Hessian vector product AD routines in conjugation with CG-like methods. Computing the Hessian vector product takes time $\tilde{O}(\omega(\hat{\mathcal{G}})^2)$, where $\hat{\mathcal{G}}$ is a moralization of the computational graph [@dixon2009automatic]. By contrast, if the computational graph were chordal, then computing the Newton-step *via* Algorithm \[alg:gnewt\] is only $\tilde{O}(\omega(\hat{\mathcal{G}})^3)$. The latter is more economical when the cliques of a graph are small in comparison to the order of the graph. The ill-conditioned nature of structured objectives may also lead to bad convergence properties for CG algorithms. For problems whose tree-widths are large, the iterative method is obviously more viable. However, following the rapid advances in approximate inference in the past two decades [@wainwright2008graphical], we hope that the explicit algebraic connection to graphical models made in this paper, can be exploited in coming up with less-agnostic iterative methods. [^1]: Email:`{akshays,todorov}@cs.washington.edu` [^2]: We use $\tilde{O}(.)$ to hide factors linear in $|E| + |V|$. [^3]: Gaussian-BP, computes the LU decomposition of a matrix [^4]: The tilde hides factors linear in $|V[\mathcal{H}]|, |E[\mathcal{H}]|$. [^5]: There are multiple definitions of a *Hypertree*; we use the term to mean a maximal Hypergraph, whose tree-decomposition can be expressed in terms of its edges. [^6]: Note that the addition is performed vertex label-wise in Line 6 of Algorithm \[alg:gqp\].

ArXiv

--- abstract: 'This paper studies self-injective algebras of polynomial growth. We prove that the derived equivalence classification of weakly symmetric algebras of domestic type coincides with the classification up to stable equivalences (of Morita type). As for weakly symmetric non-domestic algebras of polynomial growth, up to some scalar problems, the derived equivalence classification coincides with the classification up to stable equivalences of Morita type. As a consequence, we get the validity of the Auslander-Reiten conjecture for stable equivalences of Morita type between weakly symmetric algebras of polynomial growth.' address: - 'Guodong Zhou Institut für Mathematik, Universität Paderborn, 33095 Paderborn, Germany' - 'Alexander Zimmermann Université de Picardie, Département de Mathématiques et LAMFA (UMR 6140 du CNRS), 33 rue St Leu, F-80039 Amiens Cedex 1, France' author: - Guodong Zhou and Alexander Zimmermann date: 'May 3, 2010' title: 'Auslander-Reiten conjecture for symmetric algebras of polynomial growth' --- Introduction {#introduction .unnumbered} ============ Drozd [@Drozd] showed that finite dimensional algebras $A$ over an algebraically closed field $K$ are either of finite representation type, or of tame representation type or of wild representation type. Here, an algebra is of finite representation type if there are only a finite number of isomorphism classes of indecomposable modules. An algebra is of wild representation type if all module categories are full subcategories of the module category of the algebra. An algebra is of tame representation type if for every positive integer $d$ there are $A-K[X]$-bimodules $M_i(d)$ for $i\in\{1,\dots,n_d\}$ so that $M_i(d)$ is free as $K[X]$-module and so that all but finite number of isomorphism classes of indecomposable $d$-dimensional $A$-modules are isomorphic to $M_i(d)\otimes_{K[X]}K[X]/(X-\lambda)$ for $\lambda\in K$ and $i\in\{1,\dots,n_d\}$. If $n_d$ is minimal with respect to $d$ satisfying this property, $A$ is of polynomial growth if there is an integer $m$ so that $\lim_{d{\rightarrow}\infty}\frac{n_d}{d^m}=0$, and $A$ is domestic if there is an integer $m$ so that $n_d\leq d\cdot m$ for all $d$. Bocian, Holm and Skowroński [@BocianHolmSkowronski2004; @BocianHolmSkowronski2007; @HolmSkowronski2006] classified all weakly symmetric algebras of domestic type up to derived equivalence and Białkowski, Holm and Skowroński [@BialkowskiHolmSkowronski2003a; @BialkowskiHolmSkowronski2003b; @HolmSkowronski2009] did the same thing for all weakly symmetric non-domestic algebras of polynomial growth. The result is a finite number of families of algebras given by quivers and relations, depending on certain parameters. In the case of domestic algebras the classification is complete, whereas in case of weakly symmetric algebras of polynomial growth it is not known if certain choices of parameters may lead to the same derived equivalence class. This question if certain parameters lead to derived equivalent algebras will be called the scalar problem in the sequel. The main result of this paper is that the derived equivalence classification of weakly symmetric algebras of polynomial growth coincides (up to some scalar problems) with the classification up to stable equivalences of Morita type as given in [@BocianHolmSkowronski2004; @BocianHolmSkowronski2007; @HolmSkowronski2006] and [@BialkowskiHolmSkowronski2003a; @BialkowskiHolmSkowronski2003b; @HolmSkowronski2009]. As a consequence, the Auslander-Reiten conjecture holds for a stable equivalence of Morita type between two weakly symmetric algebras of polynomial growth. For weakly symmetric algebras of domestic type we can show more. We prove that the derived equivalence classification coincides with the classification up to stable equivalences (no scalar problem occurs in this case) and that the Auslander-Reiten conjecture holds for stable equivalences between weakly symmetric algebras of domestic types. As mentioned above the classification of symmetric algebras of polynomial growth up to derived equivalence leaves open if for certain choices of parameters the algebras are derived equivalent or not. The same questions concerning the same choices of parameters remain open also for the classification up to stable equivalences of Morita type. Moreover, using our result [@ZhouZimmermannTameBlocks], we show that for tame symmetric algebras with periodic modules, the derived equivalence classification coincides with the classification up to stable equivalence of Morita type, up to some scalar problem, and that the Auslander-Reiten conjecture holds for stable equivalences between tame symmetric algebras with periodic modules. The paper is organised as follows. In Section \[stableinvsect\] we recall basic definitions and recall the stable invariants we use in the sequel. In Section \[weaklysymmdomesticsection\] we prove the main result Theorem \[domesticmaintheorem\] for weakly symmetric algebras of domestic representation type and in Section \[polysection\] we show the main result Theorem \[NonDomesticClassification\] for stable equivalences of Morita type for weakly symmetric algebras of polynomial growth. As soon as we use statements on representation type we shall always assume that $K$ is an algebraically closed field. Stable categories; definitions, notations and invariants {#stableinvsect} ======================================================== The definitions --------------- Let $K$ be a field and let $A$ be a finite dimensional $K$-algebra. The stable category $A-{\underline}{mod}$ has as objects all finite dimensional $A$-modules and morphisms from na $A$-module $X$ to another $A$-module $Y$ are the equivalence classes of $A$-linear homomorphisms $Hom_A(X,Y)$, where two morphisms of $A$-modules are declared to be equivalent if their difference factors through a projective $A$-module. Two algebras are called [*stably equivalent*]{} if $A-{\underline}{mod}\simeq B-{\underline}{mod}$ as additive categories. Let $M$ be an $A-B$-bimodule and let $N$ be a $B-A$-bimodule. Following Broué the couple $(M,N)$ defines a [*stable equivalence of Morita type*]{} if - $M$ is projective as $A$-module and is projective as $B$-module - $N$ is projective as $B$-module and is projective as $A$-module - $M\otimes_BN\simeq A\oplus P$ as $A-A$-bimodules, where $P$ is a projective $A-A$-bimodule. - $N\otimes_AM\simeq B\oplus Q$ as $B-B$-bimodules, where $Q$ is a projective $B-B$-bimodule. Of course, if $(M,N)$ defines a stable equivalence of Morita type, then $M\otimes_B-:B-{\underline}{mod}\simeq A-{\underline}{mod}$ is an equivalence. However, it is known that there are algebras which are stably equivalent but for which there is no stable equivalence of Morita type. Invariants under stable equivalences ------------------------------------ Auslander and Reiten conjectured that if $A$ and $B$ are stably equivalent, then the number of isomorphism classes of non-projective simple $A$-modules is equal to the number of isomorphism classes of non-projective simple $B$-modules. The conjecture is open in general, but we shall use in this article the following positive result [@Pogorzaly] of Pogorzały: If $A$ is stably equivalent to $B$ and if $A$ is selfinjective special biserial that is not Nakayama, then $B$ is self-injective special biserial as well and the Auslander-Reiten conjecture holds in this case. Krause and Zwara [@KrauseZwara] showed that stable equivalences preserve the representation type. Suppose $A$ and $B$ are two finite dimensional stably equivalent $K$-algebras. If $A$ is of domestic representation type (respectively of polynomial growth, respectively of tame representation type), then so is $B$. Reiten [@Reiten] show that being selfinjective is almost invariant under stable equivalences. More precisely, let $A$ be a selfinjective finite dimensional $K$-algebra which is stably equivalent to a finite dimensional indecomposable $K$-algebra $B$. Then $B$ is either selfinjective or a radical squared zero Nakayama algebra. Finally, if $A$ and $B$ are stably equivalent finite dimensional algebras with Loewy length at least three, then the stable Auslander-Reiten quivers are isomorphic as translation quivers. Invariants under stable equivalence of Morita type -------------------------------------------------- Much more is known for stable equivalences of Morita type. We shall cite only those properties used in the sequel. Xi has shown [@Xi2008] that if two finite dimensional $K$-algebras $A$ and $B$ are stably equivalent of Morita type, then the absolute value of the Cartan determinants are equal. Even better, the elementary divisors, including their multiplies, of the Cartan matrices different from $\pm 1$ coincide for $A$ and for $B$. Liu showed [@Liu2008] that if $A$ is an indecomposable finite dimensional $K$-algebra, and $B$ is a finite dimensional $K$-algebra stably equivalent of Morita type to $A$, then $B$ is indecomposable as well. If $K$ is an algebraically closed field of characteristic $p>0$ and $A$ a finite dimensional $K$-algebra. Define $[A,A]$ the $K$-space generated by all elements $ab-ba\in A$ for all $a,b\in A$. Then, $$\{x\in A\;|\;x^{p^n}\in[A,A]\}=:T_n(A)$$ is a $K$-subspace of $A$ containing $[A,A]$. Liu and the authors have shown [@LZZ] that $$dim_K(T_n(A)/[A,A])=dim_K(T_n(B)/[B,B])$$ for all $n\in{{\mathbb N}}$. If $A$ and $B$ are symmetric and stably equivalent of Morita type, then taking the orthogonal space with respect to the symmetrising form, observing that $[A,A]^\perp=Z(A)$, Liu, König and the first author have shown [@KLZ] that $$Z(A)/T_n(A)^\perp\simeq Z(B)/T_n(B)^\perp$$ as rings. Weakly symmetric algebras of domestic representation type {#weaklysymmdomesticsection} ========================================================== Selfinjective algebras of domestic type split into two subclasses: standard algebras and non-standard ones. A selfinjective algebra of tame representation type is called [*standard*]{} if its basic algebra admit simply connected Galois coverings. A selfinjective algebra of tame representation type is called [*nonstandard*]{} if it is not standard. Standard domestic weakly symmetric algebras ------------------------------------------- [@BocianHolmSkowronski2004 Theorem 1] \[DomesticStandardSingular\] For an algebra $A$ the following statements are equivalent: - $A$ is representation-infinite domestic weakly symmetric algebras having simply connected Galois coverings and the Cartan matrix $C_A$ is singular. - $A$ is derived equivalent to the trivial extension $T(C)$ of a canonical algebra $C$ of Euclidean type. - $A$ is stably equivalent to the trivial extension $T(C)$ of a canonical algebra $C$ of Euclidean type. Moreover, the trivial extensions $T(C)$ and $T(C')$ of two canonical algebras $C$ and $C'$ of Euclidean type are derived equivalent (respectively, stably equivalent) if and only if the algebras $C$ and $C'$ are isomorphic. From this theorem, we know that a standard weakly symmetric algebras of domestic representation type with [*singular Cartan matrice*]{} is symmetric. The class of indecomposable standard weakly symmetric algebras of domestic representation type with singular Cartan matrices is closed under stable equivalences. In fact, let $A$ be an indecomposable algebra stably equivalent to an indecomposable standard weakly symmetric algebras of domestic representation type. Then by the preceding theorem, $A$ is stably equivalent to the trivial extension $T(C)$ of a canonical algebra $C$ of Euclidean type. Again by the preceding theorem, $A$ is standard weakly symmetric algebras of domestic representation type. For standard weakly symmetric algebras of domestic representation type with [*nonsingular Cartan matrices,*]{} we have the following derived norm forms. $$\unitlength0.6cm \begin{picture}(11,6) \put(9.4,3){\circle{2.0}} \put(6.8,3){\circle{2.0}} \put(7.85,2.9){\vector(0,1){0.3}} \put(8,2.9){$\bullet$} \put(8.35,2.9){\vector(0,1){0.3}} \put(4.8,2.9){$\beta$} \put(11,2.9){$\alpha$} \put(0,1.9){$A(\lambda)$} \put(0,1){$\lambda\in K\backslash \{0\}$} \put(4.8,0.5){$\alpha^2=0, \beta^2=0, \alpha\beta=\lambda\beta\alpha$} \end{picture}$$ $$\unitlength0.5cm \begin{picture}(30,15) \put(18.1, 11.1){\vector(1,1){1.8}} \put(16.5, 11.5){$\beta_1$} \put(18, 11){\vector(-1,1){1.8}} \put(14.5, 13){$\beta_2$} \put(16, 13){\vector(-2,1){1.8}} \put(13, 13.2){$\beta_3$} \put(14, 14){\vector(-1,0){1.8}} \put(11, 13.0){$\beta_4$} \put(12, 14){\vector(-2,-1){1.8}} \put(10.7, 8.8){$\beta_{q-3}$} \put(10, 9){\vector(2,-1){1.8}} \put(12.2, 8.5){$\beta_{q-2}$} \put(12, 8){\vector(1,0){1.8}} \put(14, 8.8){$\beta_{q-1}$} \put(14, 8){\vector(2,1){1.8}} \put(16.4, 10.3){$\beta_{q}$} \put(16, 9){\vector(1,1){1.8}} \put(19.2, 11.5){$\alpha_{1}$} \put(20, 9){\vector(-1,1){1.8}} \put(21, 13){$\alpha_{2}$} \put(22, 8){\vector(-2,1){1.8}} \put(22.5, 13.5){$\alpha_{3}$} \put(24, 8){\vector(-1,0){1.8}} \put(24.5, 13){$\alpha_{4}$} \put(26, 9){\vector(-2,-1){1.8}} \put(18.8, 10.3){$\alpha_{p}$} \put(20.7, 8.8){$\alpha_{p-1}$} \put(22.4, 8.5){$\alpha_{p-2}$} \put(24.1, 8.8){$\alpha_{p-3}$} \put(20, 13){\vector(2,1){1.8}} \put(22, 14){\vector(1,0){1.8}} \put(24, 14){\vector(2,-1){1.8}} \multiput(9, 11)(0.1,0.2){10}{\circle*{0.01}} \multiput(9, 11)(0.1,-0.2){10}{\circle*{0.01}} \multiput(27, 11)(-0.1,0.2){10}{\circle*{0.01}} \multiput(27, 11)(-0.1,-0.2){10}{\circle*{0.01}} \put(1, 12){$A(p, q)$} \put(1, 10.5){$1\leq p\leq q$} \put(1, 9){$p+q\geq 3$} \put(9, 6){$\alpha_1\alpha_2\cdots \alpha_p\beta_1\beta_2\cdots \beta_q=\beta_1\beta_2\cdots \beta_q\alpha_1\alpha_2\cdots \alpha_p$} \put(9, 4.5){$ \alpha_p\alpha_1=0, \beta_q\beta_1=0$} \put(9, 3){$\alpha_i\alpha_{i+1}\cdots \alpha_p\beta_1\cdots \beta_q \alpha_1 \cdots \alpha_{i-1}\alpha_i=0, 2\leq i\leq p$} \put(9, 1.5){$\beta_j\beta_{j+1}\cdots \beta_q\alpha_1\cdots \alpha_p \beta_1 \cdots \beta_{j-1}\beta_j=0, 2\leq j\leq q$} \end{picture}$$ $$\unitlength0.5cm \begin{picture}(30,11) \multiput(9, 7)(0.1,0.2){10}{\circle*{0.01}} \multiput(9, 7)(0.1,-0.2){10}{\circle*{0.01}} \put(16.5, 7.5){$\beta_1$} \put(18, 7){\vector(-1,1){1.8}} \put(14.5, 9){$\beta_2$} \put(16, 9){\vector(-2,1){1.8}} \put(13, 9.2){$\beta_3$} \put(14, 10){\vector(-1,0){1.8}} \put(11, 9.0){$\beta_4$} \put(12, 10){\vector(-2,-1){1.8}} \put(10.7, 4.8){$\beta_{n-3}$} \put(10, 5){\vector(2,-1){1.8}} \put(12.2, 4.5){$\beta_{n-2}$} \put(12, 4){\vector(1,0){1.8}} \put(14, 4.8){$\beta_{n-1}$} \put(14, 4){\vector(2,1){1.8}} \put(16.4, 6.3){$\beta_{n}$} \put(16, 5){\vector(1,1){1.8}} \put(19.2, 7){\circle{2}} \put(19, 8){\vector(-1,0){0.01}} \put(20.5, 7){$\alpha$} \put(1, 8){$\Lambda(n)$} \put(1, 6.5){$n\geq 2$} \put(9, 2){$\alpha^2=(\beta_1\beta_2\cdots\beta_n)^2, \alpha\beta_1=0, \beta_n\alpha=0$} \put(9, 0.5){$\beta_j\beta_{j+1}\cdots \beta_n\beta_1\cdots \beta_n \beta_1 \cdots \beta_{j-1}\beta_j=0, 2\leq j\leq n$} \end{picture}$$ $$\unitlength0.5cm \begin{picture}(30,15) \multiput(9, 11)(0.1,0.2){10}{\circle*{0.01}} \multiput(9, 11)(0.1,-0.2){10}{\circle*{0.01}} \put(16.5, 11.5){$\beta_1$} \put(18, 11){\vector(-1,1){1.8}} \put(14.5, 13){$\beta_2$} \put(16, 13){\vector(-2,1){1.8}} \put(13, 13.2){$\beta_3$} \put(14, 14){\vector(-1,0){1.8}} \put(11, 13.0){$\beta_4$} \put(12, 14){\vector(-2,-1){1.8}} \put(10.7, 8.8){$\beta_{n-3}$} \put(10, 9){\vector(2,-1){1.8}} \put(12.2, 8.5){$\beta_{n-2}$} \put(12, 8){\vector(1,0){1.8}} \put(14, 8.8){$\beta_{n-1}$} \put(14, 8){\vector(2,1){1.8}} \put(16.4, 10.3){$\beta_{n}$} \put(16, 9){\vector(1,1){1.8}} \put(18.2, 11.3){\vector(1,2){1.2}} \put(19.8, 13.4){\vector(-1,-2){1.2}} \put(18.4, 10.9){\vector(1,-2){1.2}} \put(19.2, 8.4){\vector(-1,2){1.2}} \put(19.4, 9.7){$\gamma_1$} \put(19.4, 11.7){$\alpha_2$} \put(18, 9){$\gamma_2$} \put(18, 12.5){$\alpha_1$} \put(1, 12){$\Gamma(n)$} \put(1, 10.5){$n\geq 1$} \put(9, 6){$\alpha_1\alpha_2=(\beta_1\beta_2\cdots\beta_n)^2=\gamma_1\gamma_2,$} \put(9, 4.5){$\alpha_2\beta_1=0, \gamma_2\beta_1=0, \beta_n\alpha_1=0$} \put(9, 3){$\beta_n\gamma_1=0, \alpha_2\gamma_1=0, \gamma_2\alpha_1=0$} \put(9, 1.5){$\beta_j\beta_{j+1}\cdots \beta_n\beta_1\cdots \beta_n \beta_1 \cdots \beta_{j-1}\beta_j=0, 2\leq j\leq n$} \end{picture}$$ [@BocianHolmSkowronski2004 Theorem 2]\[DomesticStandardNonSingular\] For a domestic standard self-injective algebra $A$, the following statements are equivalent: - $A$ is weakly symmetric and the Cartan matrix $C_A$ is nonsingular. - $A$ is derived equivalent to an algebra of the form $A(\lambda), A(p, q), \Lambda(n), \Gamma(n)$. - $A$ is stably equivalent to an algebra of the form $A(\lambda), A(p, q), \Lambda(n), \Gamma(n)$. Moreover, two algebras of the forms $A(\lambda), A(p, q), \Lambda(n)$ or $\Gamma(n)$ are derived equivalent (respectively, stably equivalent) if and only if they are isomorphic. All these algebras except $$A(\lambda)=k\langle X, Y\rangle/(X^2, Y^2, XY-\lambda YX), \lambda\not\in\{ 0, 1\}$$ are symmetric. Remark that except $\Gamma(n)$, all algebras are special biserial algebras. Since the class of indecomposable standard weakly symmetric algebras of domestic representation type is closed under stable equivalences, we have the following \[standardweaklysymmetricstablyequiv\] Two standard weakly symmetric algebras of domestic representation type are stably equivalent if and only if they are derived equivalent. Non standard domestic weakly symmetric algebras ----------------------------------------------- Nonstandard self-injective algebras of domestic representation type are classified up to derived and stable equivalences in [@BocianHolmSkowronski2007]. [@BocianHolmSkowronski2007 Theorem 1]\[DomesticNonStandard\] Any nonstandard representation-infinite selfinjective algebra of domestic type is derived equivalent (resp. stably equivalent) to an algebra $\Omega(n)$ with $n\geq 1$. Moreover, two algebras $\Omega(n)$ and $\Omega(m)$ are derived equivalent (respectively, stably equivalent) if and only if $m = n$. The quiver with relations of $\Omega(n)$ is as follows. $$\unitlength0.5cm \begin{picture}(30,11) \multiput(9, 7)(0.1,0.2){10}{\circle*{0.01}} \multiput(9, 7)(0.1,-0.2){10}{\circle*{0.01}} \put(16.5, 7.5){$\beta_1$} \put(18, 7){\vector(-1,1){1.8}} \put(14.5, 9){$\beta_2$} \put(16, 9){\vector(-2,1){1.8}} \put(13, 9.2){$\beta_3$} \put(14, 10){\vector(-1,0){1.8}} \put(11, 9.0){$\beta_4$} \put(12, 10){\vector(-2,-1){1.8}} \put(10.7, 4.8){$\beta_{n-3}$} \put(10, 5){\vector(2,-1){1.8}} \put(12.2, 4.5){$\beta_{n-2}$} \put(12, 4){\vector(1,0){1.8}} \put(14, 4.8){$\beta_{n-1}$} \put(14, 4){\vector(2,1){1.8}} \put(16.4, 6.3){$\beta_{n}$} \put(16, 5){\vector(1,1){1.8}} \put(19.2, 7){\circle{2}} \put(19, 8){\vector(-1,0){0.01}} \put(20.5, 7){$\alpha$} \put(1, 8){$\Omega(n)$} \put(1, 6.5){$n\geq 1$} \put(3, 2){$\alpha^2=\alpha\beta_1\beta_2\cdots\beta_n, \alpha\beta_1\beta_2\cdots\beta_n+\beta_1\beta_2\cdots\beta_n\alpha=0, $} \put(3, 0.5){$\beta_n\beta_{1}=0, \beta_j\beta_{j+1}\cdots \beta_n\beta_1\cdots \beta_n \beta_1 \cdots \beta_{j-1}\beta_j=0, 2\leq j\leq n$} \end{picture}$$ Remark that the algebra $\Omega(n)$ is always weakly symmetric, but it is symmetric only when the characteristic of the base field is $2$. Note that $\Omega(n)$ is not special biserial. Domestic standard type versus domestic non standard type -------------------------------------------------------- [@HolmSkowronski2006 Theorem 1.1]\[SymmetricDomesticStandardVsNonStandard\] A standard symmetric algebra of domestic representation type cannot be derived equivalent to a non-standard symmetric one. In course of the proof of the above theorem, one needs to compare the algebras $\Omega(n)$ with $A(1, n)$ in case the characteristic of $K$ is $2$. It is proved in [@HolmSkowronski2006] that the dimensions of the centre modulo the first Külshammer ideal, for $\Omega(n)$ and $A(1, n)$, are different, but this dimension is in fact invariant under stable equivalences of Morita type ([@LZZ]). So a standard symmetric algebra of domestic representation type cannot be stably equivalent of Morita type to a non-standard symmetric one. In fact, we can even prove more. \[WeaklySymmetricDomesticStandardVsNonStandard\] A standard weakly symmetric algebra of domestic type cannot be stably equivalent to a nonstandard one. Stably equivalent algebras of Loewy length at least $3$ have isomorphic stable Auslander-Reiten quivers ([@AuslanderReiten Corollary X. 1.9]). Notice that the stable Auslander-Reiten quiver of $A(\lambda)$ for $\lambda\neq 0$ consists of an Euclidean component of type $\mathbb{Z}\tilde{A}_{1}$ and a $\mathbb{P}_1(K)$-family of homogenous tubes. Since all algebras in the above list of Theorems \[DomesticStandardSingular\], \[DomesticStandardNonSingular\], \[DomesticNonStandard\] are of Loewy length at least $3$, by comparing the shapes of the stable Auslander-Reiten quivers as in [@HolmSkowronski2006 Section 4], we know that a standard weakly symmetric algebra of domestic type $A$ in the list of Theorems \[DomesticStandardSingular\], \[DomesticStandardNonSingular\] is stably equivalent to a nonstandard one $B=\Omega(n)$ only when $A=A(1, n)$ and $B=\Omega(n)$ for $n\geq 1$. However, notice that $A(1, n)$ is special biserial and $\Omega(n)$ is not. By the result of Pogorzały [@Pogorzaly Theorem 7.3], they can never be stably equivalent. Stable equivalence classification of domestic weakly symmetric algebras ----------------------------------------------------------------------- \[ClassificationSymmetricDomestic\] 1. Two symmetric algebras of domestic representation type are derived equivalent if and only if they are stably equivalent of Morita type if and only if they are stably equivalent. 2. The class of symmetric algebras of domestic representation type is closed under stable equivalences, hence is closed under stably equivalences of Morita type and derived equivalences. Corollary \[standardweaklysymmetricstablyequiv\] shows that for standard weakly symmetric algebras the notions of derived equivalence and of stable equivalence are the same. Lemma \[WeaklySymmetricDomesticStandardVsNonStandard\] shows that standard weakly symmetric algebras cannot be stably equivalent to non standard weakly symmetric algebras. Theorem \[DomesticNonStandard\] shows that for non standard domestic weakly symmetric algebras derived and stable equivalence are the same notion. This proves (1). Now let $A$ be an indecomposable algebra stably equivalent to an indecomposable symmetric algebra $B$ of domestic representation type in the list of algebras in Theorems \[DomesticStandardSingular\], \[DomesticStandardNonSingular\] and \[DomesticNonStandard\]. Then by a result of Idun Reiten [@Reiten Theorem 2.6] (see also [@LiuPaquette Theorem]), $A$ itself is also self-injective or $A$ is isomorphic to a radical square zero algebra whose quiver is of type $A$ with linear orientation. But the latter algebra has finite representation type, and thus cannot be stably equivalent to a representation-infinite algebra. We infer that $A$ is self-injective and now by [@BocianHolmSkowronski2004 Proposition 5.2], $A$ is weakly symmetric. By [@KrauseZwara Corollary 2 and the discussions afterwards], $A$ is of domestic type. We thus proved that $A$ is weakly symmetric of domestic type. To finish the proof of the second assertion, one needs to show that $A$ is symmetric and thus to exclude the cases where $A$ is stably equivalent to $A(\lambda)$ with $\lambda\in K\backslash\{0, 1\}$ or to $\Omega(n)$ with $n\geq 1$ in case of $char K\neq 2$, since these are the only cases where non symmetric algebras may occur. If $A$ is stably equivalent to $A(\lambda)$ with $\lambda\in K\backslash\{0, 1\}$, then by [@Pogorzaly Theorem 7.3], $B$ is special biserial. Since the Auslander-Reiten conjecture is proved for self-injective special biserial algebras ([@Pogorzaly Theorem 0.1]), $B$ is also local and is thus necessary isomorphic to $A(1)$. However, $A(1)$ is never stably equivalent to $A(\lambda)$ with $\lambda\neq 1$ by an unpublished result of Rickard. If $char(K)\neq 2$ and $A$ is stably equivalent to $A=\Omega(n)$ with $n\geq 1$, then by Lemma \[WeaklySymmetricDomesticStandardVsNonStandard\], $B$ is necessarily nonstandard and $B=\Omega(n)$ for some $n$. But the fact that $B$ is symmetric implies that $char(K)=2$ which is a contradiction. In fact we can extend (at least partially) the above result to all weakly symmetric algebras of domestic representation type. \[domesticmaintheorem\] 1. Two weakly symmetric algebras of domestic representation type are derived equivalent if and only if they are stably equivalent of Morita type if and only if they are stably equivalent. 2. The class of weakly symmetric algebras of standard domestic representation type is closed under stable equivalences, hence under stably equivalences of Morita type, and derived equivalences. Since by Lemma \[WeaklySymmetricDomesticStandardVsNonStandard\], a standard weakly symmetric algebra of domestic type cannot be stably equivalent to a nonstandard one, Theorems \[DomesticStandardSingular\],  \[DomesticStandardNonSingular\] and  \[DomesticNonStandard\] imply the first assertion. For the second assertion, let $A$ be an indecomposable algebra stably equivalent to a weakly symmetric standard algebra $B$ of domestic representation type in the list of algebras in Theorems \[DomesticStandardSingular\] and  \[DomesticStandardNonSingular\]. By [@KrauseZwara Corollary 2 and the discussions afterwards], $A$ is of domestic type. As in the proof of Proposition \[ClassificationSymmetricDomestic\], $A$ is still self-injective. Since by Proposition \[ClassificationSymmetricDomestic\](2), the class of symmetric algebras of domestic type is closed under stable equivalences, one can assume that $B$ is not symmetric. Now $B=A(\lambda)$ with $\lambda\not\in\{ 0, 1\}$. Since $B=A(\lambda)$ is a local special biserial algebra, so is $A$ by [@Pogorzaly Theorems 0.1 and 7.3]. The algebra $A$ is thus weakly symmetric. 1. If the class of self-injective nonstandard algebras of domestic type is closed under under stable equivalences (hence under stably equivalences of Morita type and derived equivalences), so is the class of weakly symmetric algebras of domestic representation type. 2. The second assertion of the above theorem answers a question of [@BocianHolmSkowronski2004 Page 47]. 3. If we have a complete Morita equivalence classification of all self-injective algebras of domestic type, the classification up to stable equivalence (of Morita type) might be feasible as well. As a consequence, the Auslander-Reiten conjecture holds for this class of algebras. 1. Let $A$ be an indecomposable algebra stably equivalent to an indecomposable symmetric algebra $B$ of domestic type. Then $A$ has the same number of simple modules as $B$. 2. Let $A$ be an indecomposable algebra stably equivalent to an indecomposable weakly symmetric standard algebra $B$ of domestic type. Then $A$ has the same number of simple modules as $B$. If the class of self-injective nonstandard algebras of domestic type is closed under under stable equivalences, then Auslander-Reiten conjecture holds for a stable equivalence between two weakly symmetric algebras of domestic type. Non-domestic self-injective algebras of polynomial growth {#polysection} ========================================================= The derived equivalence classification of the standard (resp. non-standard) non-domestic symmetric algebras of polynomial growth is achieved in [@BialkowskiHolmSkowronski2003a Page 653 Theorem] (resp.[@BialkowskiHolmSkowronski2003b Theorem 3.1]). They give a complete list of representatives of derived equivalence classes in terms of quiver with relations. We recall the derived norm forms following [@HolmSkowronski2009]. We have five quivers and various relations. (15,6.5) (1.2,6)[$Q_2(1)$]{} (1,5)[$\bullet$]{} (2,5)[$\bullet$]{} (1.1,5.3)[(1,0)[.8]{}]{} (1.9,4.95)[(-1,0)[.8]{}]{} (.4,5.1) (.88,5.1)[(0,-1)[.1]{}]{} (1.4,5.4)[$\gamma$]{} (1.4,4.7)[$\beta$]{} (0,5)[$\alpha$]{} (5.2,6)[$Q_2(2)$]{} (5,5)[$\bullet$]{} (6,5)[$\bullet$]{} (5.1,5.3)[(1,0)[.8]{}]{} (5.9,4.95)[(-1,0)[.8]{}]{} (4.4,5.1) (4.88,5.1)[(0,-1)[.1]{}]{} (6.7,5.1) (6.22,5.1)[(0,1)[.1]{}]{} (5.4,5.4)[$\sigma$]{} (5.4,4.7)[$\gamma$]{} (4,5)[$\alpha$]{} (6.99,5)[$\beta$]{} (9.6,2)[$Q_3(3)$]{} (9,1)[$\bullet$]{} (10,1)[$\bullet$]{} (11,1)[$\bullet$]{} (9.1,1.3)[(1,0)[.8]{}]{} (9.9,.95)[(-1,0)[.8]{}]{} (10.1,1.3)[(1,0)[.8]{}]{} (10.9,.95)[(-1,0)[.8]{}]{} (9.4,1.4)[$\alpha$]{} (9.4,.7)[$\gamma$]{} (10.4,1.4)[$\sigma$]{} (10.4,.7)[$\beta$]{} (.6,3)[$Q_3(1)$]{} (0,1)[$\bullet$]{} (1,1)[$\bullet$]{} (2,1)[$\bullet$]{} (.1,1.3)[(1,0)[.8]{}]{} (.9,.95)[(-1,0)[.8]{}]{} (1.1,1.3)[(1,0)[.8]{}]{} (1.9,.95)[(-1,0)[.8]{}]{} (1,1.9) (1,1.48)[(1,0)[.1]{}]{} (.4,1.4)[$\beta$]{} (.4,.7)[$\gamma$]{} (1.4,1.4)[$\delta$]{} (1.4,.7)[$\sigma$]{} (1,2.1)[$\alpha$]{} (5.1,3)[$Q_3(2)$]{} (4.5,1)[$\bullet$]{} (5.5,1)[$\bullet$]{} (6.5,1)[$\bullet$]{} (4.6,1.3)[(1,0)[.8]{}]{} (5.4,.95)[(-1,0)[.8]{}]{} (5.6,1.3)[(1,0)[.8]{}]{} (6.4,.95)[(-1,0)[.8]{}]{} (5.5,2.2)[$\bullet$]{} (5.4,1.5)[(0,1)[.7]{}]{} (5.55,2.2)[(0,-1)[.7]{}]{} (4.9,1.4)[$\alpha$]{} (4.9,.7)[$\beta$]{} (5.9,1.4)[$\epsilon$]{} (5.9,.7)[$\xi$]{} (5.25,1.8)[$\delta$]{} (5.65,1.8)[$\gamma$]{} Define for any $\lambda\in K\setminus\{0,1\}$ the algebras $$\begin{aligned} \Lambda_2&:=&KQ_2(1)/(\alpha^2\gamma,\beta\alpha^2,\gamma\beta\gamma, \beta\gamma\beta,\beta\gamma-\beta\alpha\gamma,\alpha^3-\gamma\beta)\\ \Lambda_2'&:=&KQ_2(1)/(\alpha^2\gamma,\beta\alpha^2,\beta\gamma, \alpha^3-\gamma\beta)\\ \Lambda_3(\lambda)&:=&KQ_2(2)/(\alpha^4,\gamma\alpha^2,\alpha^2\sigma, \alpha^2-\sigma\gamma-\alpha^3,\lambda\beta^2-\gamma\sigma, \gamma\alpha-\beta\gamma,\sigma\beta-\alpha\sigma)\\ \Lambda_3'(\lambda)&:=&KQ_2(2)/(\alpha^2-\sigma\gamma, \lambda\beta^2-\gamma\sigma,\gamma\alpha-\beta\gamma, \sigma\beta-\alpha\sigma)\\ \\ \Lambda_5&:=&KQ_3(1)/(\alpha^2-\gamma\beta,\alpha^3-\delta\sigma, \beta\delta,\sigma\gamma,\alpha\delta,\sigma\alpha,\gamma\beta\gamma, \beta\gamma\beta,\beta\gamma-\beta\alpha\gamma)\\ \Lambda_5'&:=&KQ_3(1)/(\alpha^2-\gamma\beta,\alpha^3-\delta\sigma, \beta\delta,\sigma\gamma,\alpha\delta,\sigma\alpha,\beta\gamma)\\ \Lambda_9&:=&KQ_3(2)/(\beta\alpha+\delta\gamma+\epsilon\xi,\gamma\delta, \xi\epsilon,\alpha\beta\alpha,\beta\alpha\beta, \alpha\beta-\alpha\delta\gamma\beta)\\ \Lambda_9'&:=&KQ_3(2)/(\beta\alpha+\delta\gamma+\epsilon\xi,\gamma\delta, \xi\epsilon,\alpha\beta)\\ A_1(\lambda)&:=&KQ_3(3)/(\alpha\gamma\alpha-\alpha\sigma\beta, \beta\gamma\alpha-\lambda\cdot\beta\sigma\beta, \gamma\alpha\gamma-\sigma\beta\gamma, \gamma\alpha\sigma-\lambda\sigma\beta\sigma)\\ A_4&:=&KQ_3(2)/(\beta\alpha+\delta\gamma+\epsilon\xi, \alpha\beta,\gamma\epsilon,\xi\delta)\end{aligned}$$ Further, denote the trivial extensions of tubular canonical algebras as usual, in particular $$\Lambda(2, 2, 2, 2, \lambda)=T(C(2,2,2,\lambda))\mbox{ for }\lambda\in K \setminus \{0, 1\}$$ $$\Lambda(3, 3, 3)=T(C(3,3,3))$$ $$\Lambda(2, 4, 4)=T(C(2,4,4))$$ $$\Lambda(2, 3, 6)=T(C(2,3,6)).$$ The precise quivers with relations of the trivial extensions of tubular canonical algebras in question are displayed in [@HolmSkowronski2009]. We will refrain from presenting them here since we do not really need this information in such details. Weakly symmetric standard non-domestic polynomial growth algebras ----------------------------------------------------------------- [@BialkowskiHolmSkowronski2003a Page 653 Theorem] Let $A$ be an indecomposable standard non-domestic weakly symmetric algebra of polynomial growth. Then $A$ is derived equivalent to one of the following algebras: - two simple modules: $\Lambda_2'$ and $\Lambda_3'(\lambda), \lambda \in K \backslash\{0, 1\}$; - three simple modules: $\Lambda_5'$ and $A_1(\lambda), \lambda \in K \backslash \{0, 1\}$; - four simple modules: $\Lambda_9'$ and $A_4$; - six simple modules: $\Lambda(2, 2, 2, 2, \lambda), \lambda \in K \backslash \{0, 1\}$; - eight simple modules: $\Lambda(3, 3, 3)$; - nine simple modules: $\Lambda(2, 4, 4)$; - ten simple modules: $\Lambda(2, 3, 6)$. The above classification is complete up to the scalar problems in $\Lambda_3'(\lambda)$ and in $A_1(\lambda)$. Remark that except $\Lambda_9'$ in case the characteristic of the base field $K$ is different from $2$, all algebras are symmetric. We shall prove \[NonDomesticStandard\] The classification of indecomposable standard non-domestic weakly symmetric algebras of polynomial growth up to stable equivalences of Morita type coincide with the derived equivalence classification, modulo the scalar problems in $\Lambda_3'(\lambda), \lambda \in K \backslash\{0, 1\}$ and in $A_1(\lambda), \lambda \in K \backslash \{0, 1\}$. Since a derived equivalence between self-injective algebras induces a stable equivalence of Morita type ([@KellerVossieck], [@Rickard1989]), we only need to show that two algebras from the above list which are not derived equivalent are not stably equivalent of Morita type, either. Since the property of being symmetric is invariant under a stable equivalence of Morita type ([@Liu2008 Corollary 2.4]), in case the characteristic of the base field is different from $2$, $\Lambda_9'$ cannot be stably equivalent of Morita type to any of the other algebras. We can concentrate on the remaining symmetric algebras in the list. Now the algebras of at least six simple modules are trivial extensions of canonical algebras of tubular type $(2, 2, 2, 2)$, $ (3, 3, 3)$, $(2, 4, 4)$, $(2, 3, 6)$ respectively. They have singular Cartan matrices, while all other algebras have non-singular Cartan matrix. Since, by a result of Xi [@Xi2008 Proposition 5.1] the absolute value of the Cartan determinant is preserved by a stable equivalence of Morita type, we can consider separately those algebras of at most four simple modules and those of at least six simple modules. For trivial extension cases, by [@BialkowskiHolmSkowronski2003a Propositions 5.1 and 5.2], two algebras of this type are stably equivalent if and only if they are derived equivalent. So the number of simple modules, which is an invariant under derived equivalence, distinguishes them. Now suppose two indecomposable standard non-domestic weakly symmetric algebras of polynomial growth are stably equivalent of Morita type and have non-singular Cartan matrices. Then the algebras are algebras in the above list. The following table gives some stable invariants of Morita type which distinguish two algebras in the above list. $$\begin{array}{c||c|c|c|c|c|c} Algebra\ A & \Lambda_2' &\Lambda_3'(\lambda)& \Lambda_5'& A_1(\lambda)& \Lambda_9'& A_4\\ \hline det\ C_A& 6 & 12 & 6 & 16 & 4 & 12 \\ dim (Z(A)/R(A))& 3 & 4 & 2 &2 & 1 & 2 \end{array}$$ Selfinjective non-standard non-domestic polynomial growth algebras ------------------------------------------------------------------ Now we turn to non-domestic non-standard selfinjective algebras of polynomial growth. Recall the following (Białkowski, Holm, Skowroński [@BialkowskiHolmSkowronski2003b Theorem 3.1]) Let $A$ be an indecomposable non-standard non-domestic self-injective algebra of polynomial growth. Then we get the following. - If the base field is of characteristic $3$, then $A$ is derived equivalent to $\Lambda_2$. - else the base field is of characteristic $2$ and then - if $A$ has two simple modules, $A$ is derived equivalent to $\Lambda_3(\lambda), \lambda \in K \setminus\{0, 1\}$ - if $A$ has three simple modules, $A$ is derived equivalent to $\Lambda_5$ - if $A$ has four simple modules, $A$ is derived equivalent to $\Lambda_9$ - else $A$ has five simple modules and then $A$ is derived equivalent to $\Lambda_{10}$ If the base field is of characteristic $2$ and $\lambda\neq\lambda'$, we do not know if $\Lambda_3(\lambda)$ is derived equivalent to $\Lambda_3(\lambda')$ or not. We call this question the scalar problem. The above classification is complete up to the scalar problem in case the base field is of characteristic $2$ for the algebra $\Lambda_3(\lambda), \lambda \in K \backslash\{0, 1\}$. Remark that the algebra $\Lambda_{10}$ is not symmetric (even not weakly symmetric), so it is not stably equivalent of Morita type to any other algebra in the above list. We can now prove \[NonDomesticNonStandard\]The classification of indecomposable non-standard non-domestic self-injective algebras of polynomial growth up to stable equivalences of Morita type coincides with the derived equivalence classification, modulo the scalar problem in $\Lambda_3(\lambda)$. We only need to consider the case of characteristic two, since if the characteristic of $K$ is $3$, there is only one algebra. But notice $det\ C_{\Lambda_3(\lambda)}=12$, $det\ C_{\Lambda_5}=6$ and $det\ C_{\Lambda_9}=4$. Standard algebras versus non-standard algebras ---------------------------------------------- Now one can prove that a non-standard algebra cannot be stably equivalent to a standard one. The case of a derived equivalence is proved by Holm-Skowroński in [@HolmSkowronski2009]. [@HolmSkowronski2009 Main Theorem] Let $A$ be a standard self-injective algebra of polynomial growth and let $\Lambda$ be an indecomposable, nonstandard, non-domestic, symmetric algebra of polynomial growth. Then $A$ and $\Lambda$ are not derived equivalent. \[NonDomesticStandardVsNonStandard\] Let $A$ be an indecomposable standard weakly symmetric algebra and $\Lambda$ be an indecomposable nonstandard non-domestic self-injective algebra of polynomial growth. Then $A$ and $\Lambda$ are not stably equivalent of Morita type. Suppose that $A$ and $\Lambda$ are stably equivalent of Morita type. By Krause [@Krause Page 605 Corollary], $A$ is also non-domestic of polynomial growth. We only need to consider symmetric algebras. In fact, the only non-symmetric nonstandard, non-domestic self-injective algebra of polynomial growth $\Lambda_{10}$ is only defined when the characteristic of the base field is $2$, but in case of characteristic $2$, all standard non-domestic weakly symmetric algebras of polynomial growth are symmetric, as $\Lambda_9'$ is non-symmetric if and only if characteristic of the base field is different from $2$. Now we consider symmetric algebras. In fact, the method of the proof of [@HolmSkowronski2009 Main Theorem] works. Since two algebras which are stably equivalent of Morita type have isomorphic stable Auslander-Reiten quivers, by the shape of the stable Auslander-Reiten quivers (see the second paragraph of Section 4 of Holm-Skowroński [@HolmSkowronski2009]), we proceed by remarking the following facts. 1. If $K$ is of characteristic $3$, the algebras $\Lambda_2$ and $\Lambda_2'$ are not stably equivalent of Morita type. Indeed Holm-Skowroński [@HolmSkowronski2009] have shown that $$dim\ Z(\Lambda_2')/T_1(\Lambda_2')^{\perp}=3\neq 2 =dim\ Z(\Lambda_2)/T_1(\Lambda_2)^{\perp}.$$ 2. If $K$ is of characteristic $2$ and $\lambda,\mu \in K \setminus \{0, 1\}$, the algebras $\Lambda_3(\lambda)$ and $\Lambda_3(\mu)'$ are not stably equivalent of Morita type. Indeed Holm-Skowroński [@HolmSkowronski2009] have shown that $$dim\ Z(\Lambda_3'(\lambda))/T_1(\Lambda_3'(\lambda))^{\perp}=3\neq 2=dim\ Z(\Lambda_3(\mu))/T_1(\Lambda_3(\mu))^{\perp}.$$ 3. If $K$ is of characteristic $2$, the algebras $\Lambda_5$ and $\Lambda_5'$ are not stably equivalent of Morita type. Indeed Holm-Skowroński [@HolmSkowronski2009] have shown that $$dim Z(\Lambda_5')/T_1(\Lambda_3')^{\perp}=0\neq dim Z(\Lambda_5)/T_1(\Lambda_3)^{\perp}.$$ 4. If $K$ is of characteristic $2$, the algebras $\Lambda_9$ and $\Lambda_9'$ are not stably equivalent of Morita type. Indeed Holm-Skowroński [@HolmSkowronski2009] have shown that $$dim HH^2(\Lambda_9)=4\neq 3=dim HH^2(\Lambda_9').$$ Since $Z(A)/T_1(A)^{\perp}$ and $HH^2(A)$ are invariants under stable equivalences of Morita type, this completes the proof. Stable equivalence classification of weakly symmetric non-domestic polynomial growth algebras --------------------------------------------------------------------------------------------- Combining Propositions \[NonDomesticStandard\], \[NonDomesticNonStandard\] and \[NonDomesticStandardVsNonStandard\], we have \[NonDomesticClassification\] The classification of indecomposable non-domestic weakly symmetric algebras of polynomial growth up to stable equivalences of Morita type coincides with the derived equivalence classification, up to the above mentioned scalar problems. As a consequence, we can prove a special case of Auslander-Reiten conjecture. \[ARNonDomestic\] 1. Let $A$ be an indecomposable algebra which is stably equivalent of Morita type to an indecomposable non-domestic symmetric algebra $\Lambda$ of polynomial growth. Then $A$ and $\Lambda$ have the same number of simple modules. 2. Let $A$ and $\Lambda$ be two indecomposable algebras which are both non-domestic weakly symmetric algebra of polynomial growth or which are both non-standard self-injective algebras of polynomial growth. If they are stably equivalent of Morita type, then $A$ and $\Lambda$ have the same number of simple modules. (1). By Krause [@Krause Page 605 Corollary] and Liu [@Liu2008 Corollary 2.4], $A$ is also symmetric, non-domestic, and of polynomial growth. Then one can apply Theorem \[NonDomesticClassification\], by noticing that the existing scalar problems all occur in families of algebras for which different scalars yield algebras having the same number of simple modules. \(2) is a direct consequence of Theorem \[NonDomesticClassification\] and Proposition \[NonDomesticNonStandard\]. Periodic algebras ----------------- Recently, Karin Erdmann and Andrzej Skowroński [@ErdmannSkowronski2010] have shown that a non-simple indecomposable symmetric algebra $A$ is tame with periodic modules if and only if $A$ belongs to one of the following classes of algebras: a representation-finite symmetric algebra, a non domestic symmetric algebra of polynomial growth, or an algebra of quaternion type (in the sense of [@ErdmannLNM]). Note that these three classes of algebras are closed under stable equivalences of Morita type, by Krause [@Krause Page 605 Corollary] and Liu [@Liu2008 Corollary 2.4]. In [@Asashiba] Asashiba, in combination with Holm-Skowroński [@HolmSkowronski2006], the statement displayed in [@ErdmannSkowronskiperiodic Theorem 3.4, Theorem 3.5, Theorem 3.6] proved that for the class of representation-finite selfinjective indecomposable algebras are derived equivalent if and only if they are stable equivalent. In our recent paper [@ZhouZimmermannTameBlocks Theorem 7.1], we proved that for the class of algebras of quaternion type, derived equivalence classification coincide with the classification up to stable equivalences of Morita type (up to some scalar problems). Therefore, combining the above Theorem \[ARNonDomestic\], we showed the following 1. The class of tame symmetric algebras with periodic modules is closed under stable equivalences of Morita type, in particular under derived equivalences; 2. For tame symmetric algebras with periodic modules derived equivalence classification coincide with the classification up to stable equivalences of Morita type (up to some scalar problems); 3. The Auslander-Reiten conjecture holds for a stable equivalence of Morita type between two tame symmetric algebra with periodic modules. Concluding remarks ================== In Section \[polysection\] we classified symmetric algebras of polynomial growth only up to stable equivalences of Morita type, whereas the results in Section \[weaklysymmdomesticsection\] concern stable equivalences in general. It would be most interesting to get a stable equivalence classification for general selfinjective algebras of polynomial growth, in particular the validity of the Auslander-Reiten conjecture in general. The crucial point for this more general statement would be to first classify algebras which are stably equivalent (of Morita type) to $\Lambda_9'$ in case the characteristic of $K$ is different from $2$ or to $\Lambda_{10}$, since these are the selfinjective polynomial growth algebras which are not symmetric. The second step is to classify weakly symmetric algebras of polynomial growth up to stable equivalences. If the for the class of selfinjective algebras of polynomial growth the derived equivalence classification would coincide with the classification up to stable equivalences (of Morita type) the Auslander-Reiten conjecture for this class of algebras should follow. Another point concerns the work of Martinez-Villa ([@Martinez-Villa]) of the reduction of the Auslander-Reiten conjecture to selfinjective algebras. If we can prove that this reduction preserves representation type: domestic type, polynomial growth etc, then the conjectural result of the preceding paragraph would imply the validity of the Auslander-Reiten conjecture for algebras of polynomial growth. The results of this paper present a second occurrence of a phenomenon of the same kind: In our previous paper [@ZhouZimmermannTameBlocks] we also obtained that the classification of a class of tame symmetric algebras up to stable equivalence coincides with the derived equivalence classification. One might ask if two indecomposable tame symmetric algebras are stably equivalent of Morita type if and only if they are derived equivalent. [99]{} Hideto Asashiba, *The derived equivalence classification of representation-finite selfinjective algebras*. J. Algebra **214** (1999) 182-221. Maurice Auslander, Idun Reiten and Sverre O. Smal[ø]{}, *Representation Theory of Artin Algebras*, Cambridge Stud. Adv. Math., 36, Cambridge Univ. Press, 1995. Jerzy Białkowski, Thorsten Holm and Andrzej Skowroński, *Derived equivalences for tame weakly symmetric algebras having only periodic modules*. J. Algebra **269** (2003), no. 2, 652-668. Jerzy Białkowski, Thorsten Holm and Andrzej Skowronski, *On nonstandard tame selfinjective algebras having only periodic modules*. Colloq. Math. **97** (2003), no. 1, 33-47. Rafał Bocian, Thorsten Holm and Andrzej Skowroński, [*Derived equivalence classification of weakly symmetric algebras of Euclidean type*]{}. J. Pure Appl. Algebra **191** (2004), no. 1-2, 43-74. Rafał  Bocian, Thorsten Holm and Andrzej Skowroński, [*Derived equivalence classification of nonstandard selfinjective algebras of domestic type*]{}. Comm. Algebra **35** (2007), no. 2, 515–526. Yurij A. Drozd, *Tame and wild matrix problems. Representation theory,* (Proc. Second Internat. Conf., Carleton Univ., Ottawa, Ont., 1979), pp. 242-258, Lecture Notes in Math., 832, Springer, Berlin-New York, 1980. Karin Erdmann, [*Blocks of tame representation type and related algebras*]{}, Springer LNM 1428 (1980). Karin Erdmann and Andrzej Skowronski, [*Classification of tame symmetric algebras with periodic modules*]{}, preprint in preparation. Karin Erdmann and Andrzej Skowroński, [*Periodic algebras*]{}, in ”Trends in representation theory of algebras and related topics”, European Mathematical Society, Series of Congress Reports, (2008) pp 201-251 Thorsten Holm, [*Blocks of tame representation type and related algebras: derived equivalences and Hochschild cohomology*]{}, Habilitationsschrift (2001) Universität Magdeburg. Thorsten Holm and Andrzej Skowroński, [*Derived equivalence classification of symmetric algebras of domestic type*]{}. J. Math. Soc. Japan **58** (2006), no. 4, 1133-1149 Thorsten Holm and Andrzej Skowroński, [*Derived equivalence classification of symmetric algebras of polynomial growth*]{}. preprint (2009). Bernhard Keller and Dieter Vossieck, *Sous les catégories dérivées*. C. R. Acad. Sci. Paris Série I Math. **305** (1987), no. 6, 225–228. Steffen König, Yuming Liu and Guodong Zhou, [*Transfer maps in Hochschild (co-)homology and their applications,* ]{} preprint 2009, to appear Trans. Amer. Math. Soc. Henning Krause, [*Representation type and stable equivalence of Morita type for finite-dimensional algebras*]{}, Math. Z. **229** (1998), no. 4, 601-606. Henning Krause and Grzegorz Zwara, [*stable equivalence and generic modules*]{}, Bull. London Math. Soc. **32** (2000) 615-618. Shiping Liu and Charles Paquette, *Stably triangulated algebras*. J. Algebra **320** (2008), no. 6, 2629-2631. Yuming Liu, [*Summands of stable equivalences of Morita type*]{}. Comm. in Algebra **36**(10) (2008), 3778-3782. Yuming Liu, Guodong Zhou and Alexander Zimmermann, [*Higman ideal, stable Hochschild homology and Auslander-Reiten conjecture*]{}, preprint (2008). Roberto Martinez-Villa, [*Properties that are left invariant under stable equivalence*]{}. Comm. Algebra **18** (1990), no. 12,4141-4169. Zygmunt Pogorzały, [*Algebras stably equivalent to self-injective special biserial algebras*]{}, Comm. Algebra **22** (1994), no. 4, 1127-1160 Idun Reiten, *Stable equivalence of self-injective algebras*. J. Algebra **40** (1976), no. 1, 63-74. Jeremy Rickard, *Derived categories and stable equivalence*. J. Pure Appl. Algebra **61** (1989), no. 3, 303–317. Chang-Chang Xi, [*Stable equivalences of adjoint type*]{}. Forum Math. **20** (2008), 81-97. Guodong Zhou and Alexander Zimmermann [*Classifying tame blocks and related algebras up to stable equivalences of Morita type*]{}. preprint (2010)

ArXiv

--- abstract: 'The notions of *deleted* and *restricted* arrangements have been very useful in the study of arrangements of hyperplanes. Let ${\mathcal{A}}$ be an arrangement of hyperplanes and $x \in {\mathcal{A}}$. The deleted and restricted arrangements ${\mathcal{A}}''$ and ${\mathcal{A}}''''$ allows us to compute recursively the Poincaré polynomials of the complement space $M({\mathcal{A}})$ with the following formula : $$\operatorname{Poin}(M({\mathcal{A}}), t) = \operatorname{Poin}(M({\mathcal{A}}''), t) + t \operatorname{Poin}(M({\mathcal{A}}''''), t).$$ In this paper, we consider the extension of this formula to arbitrary subspaces arrangements. The main result is the existence of a long exact sequence connecting the rational cohomology of $M({\mathcal{A}})$, $M({\mathcal{A}}'')$ and $M({\mathcal{A}}'''')$. Using this sequence, we obtain new results connecting the Betti numbers and Poincaré polynomials of deleted and restricted arrangements.' address: | UCL, Departement de mathematique\ Chemin du Cyclotron, 2\ B-1348 Louvain-la-neuve\ Belgium author: - Gery Debongnie title: A long exact sequence in cohomology for deleted and restricted subspaces arrangements --- Introduction ============ An arrangement of subspaces is a finite set ${\mathcal{A}}= \{x_0, \dotsc, x_n\}$ of vector subspaces in ${\mathbb{C}}^l$. Unless said otherwise, we’ll always suppose that ${\mathcal{A}}$ is a complex arrangement and that there is no $x \neq y$ in ${\mathcal{A}}$ such that $x \subset y$. We consider the poset of intersections of elements of ${\mathcal{A}}$ ordered by reverse inclusion. We define two operations on $L({\mathcal{A}})$ : the meet $x \wedge y = \cap\{z \in L({\mathcal{A}}) {\mid}x \cup y \subset z\}$ and the join $x \vee y = x \cap y$. These two operations endow $L({\mathcal{A}})$ with a lattice structure. Our main interest is the study of the complement space $M({\mathcal{A}}) = {\mathbb{C}}^l \setminus \cup_{x \in {\mathcal{A}}} x$. When all the subspaces $x \in {\mathcal{A}}$ are hyperplanes ($\operatorname{codim}x = 1$), then ${\mathcal{A}}$ is an arrangement of hyperplanes. If ${\mathcal{A}}$ is such an arrangement and $x \in {\mathcal{A}}$, then we can consider the arrangements ${\mathcal{A}}' = {\mathcal{A}}\setminus \{x\}$ and ${\mathcal{A}}'' = \{x \cap y {\mid}y \in {\mathcal{A}}''\}$. The triple $({\mathcal{A}}, {\mathcal{A}}', {\mathcal{A}}'')$ has been studied (a good reference is the book of P. Orlik and H. Terao [@orl92]), and possesses some interesting properties. For example, there is a relationship between the Poincaré polynomials of $M({\mathcal{A}}), M({\mathcal{A}}')$ and $M({\mathcal{A}}'')$ : $$\operatorname{Poin}(M({\mathcal{A}}), t) = \operatorname{Poin}(M({\mathcal{A}}'), t) + t \operatorname{Poin}(M({\mathcal{A}}''), t).$$ The goal of this paper is to generalize that property to arbitrary subspaces arrangements. But before doing that, we need to define what are the deleted and restricted arrangements ${\mathcal{A}}'$ and ${\mathcal{A}}''$ in the general case. The deleted arrangement is obvious, but there is a slight difficulty when we want to extend restricted arrangements. Section \[sec:drsa\] contains a proper definition of ${\mathcal{A}}''$. It also gives a description of a simple rational model $D({\mathcal{A}})$ of $M({\mathcal{A}})$. Let $D({\mathcal{A}}), D({\mathcal{A}}')$ and $D({\mathcal{A}}'')$ be the rational models of $M({\mathcal{A}})$, $M({\mathcal{A}}')$ and $M({\mathcal{A}}'')$, as defined in section \[sec:drsa\]. From the definition, there exists an injective map (the inclusion) $j {\colon}D({\mathcal{A}}') \to D({\mathcal{A}})$. This map induces a short exact sequence of cochains complexes $$0 \to D({\mathcal{A}}') \stackrel{j}{\to} D({\mathcal{A}}) \to \frac{D({\mathcal{A}})}{D({\mathcal{A}}')} \to 0.$$ In section \[sec:quasiiso\], we will show that there is a c.d.g.a. $D(\widetilde{{\mathcal{A}}''})$ and quasi-isomorphisms $$\frac{D({\mathcal{A}})}{D({\mathcal{A}}')} \stackrel{\simeq}{\to} D(\widetilde{{\mathcal{A}}''}) \stackrel{\simeq}{\leftarrow} D({\mathcal{A}}'').$$ Together, these results give us a long exact sequence connecting the cohomology of $M({\mathcal{A}})$, $M({\mathcal{A}}')$ and $M({\mathcal{A}}'')$, which is studied in section \[sec:drt\] (see theorem \[theo:lesic\]). A few consequences are studied as well. For example, the Euler characteristic of the complement space of a complex subspaces arrangement is always 0. Under some conditions, this long exact sequence allows us to generalize the formula about the Poincaré polynomials to $$\operatorname{Poin}(M({\mathcal{A}}), t) = \operatorname{Poin}(M({\mathcal{A}}'), t) + t^{\deg(x)}\operatorname{Poin}(M({\mathcal{A}}''), t).$$ Finally, section \[sec:agl\] gives stronger results for the specific case of subspaces arrangements with a geometric lattice. Rational models and restricted subspaces arrangements {#sec:drsa} ===================================================== In [@yu05], S. Yuzvinsky and E. Feichtner described a simple rational model for the topological space $M({\mathcal{A}})$ of arbitrary subspaces arrangements. This model is a graded differential algebra denoted by $D({\mathcal{A}})$. Let ${\mathcal{A}}= \{x_0, \dotsc, x_n\}$ be a subspaces arrangement. A rational model for the space $M({\mathcal{A}})$ is given as follows : choose a linear order on ${\mathcal{A}}$ (in this paper, we’ll use $x_0 < x_1 < \dotso < x_n$). The graded vector space $D({\mathcal{A}})$ has a basis given by all subsets $\sigma \subseteq {\mathcal{A}}$. For $\sigma = \{x_{i_1}, \dotsc, x_{i_r}\}$, we define a differential by the formula $$d\sigma = \sum_{j:\vee(\sigma \setminus \{x_{i_j}\}) = \vee \sigma} (-1)^j (\sigma\setminus\{x_{i_j}\})$$ where the indexing of the elements in $\sigma$ follows the linear order imposed on ${\mathcal{A}}$. With $\deg(\sigma) = 2 \operatorname{codim}\vee \sigma - |\sigma|$, $D({\mathcal{A}})$ is a cochain complex. Finally, the multiplication is defined, for $\sigma, \tau \subseteq {\mathcal{A}}$, by $$\sigma \cdot \tau = \left\{\begin{aligned} (-1)^{\operatorname{sgn}\epsilon(\sigma, \tau)} \sigma \cup \tau &\qquad \text{ if } \operatorname{codim}\vee \sigma + \operatorname{codim}\vee \tau = \operatorname{codim}\vee(\sigma \cup \tau), \\ 0 & \qquad\text{ otherwise,} \end{aligned}\right.$$ where $\epsilon(\sigma, \tau)$ is the permutation that, applied to $ \sigma \cup \tau$ with the induced linear order, places elements of $\tau$ after elements of $\sigma$, both in the induced linear order. Since we are interested in deleted and restricted subspaces arrangements, the rest of this section will be used to define such arrangements and to better understand their models. Let ${\mathcal{A}}$ be a (non-empty) subspaces arrangements and $x_0 \in {\mathcal{A}}$. Obviously, the definition of the *deleted arrangement* is ${\mathcal{A}}' = {\mathcal{A}}\setminus \{x_0\}$. Its model $D({\mathcal{A}}')$ is a subalgebra of $D({\mathcal{A}})$ (the inclusion is an injective map). However, the restricted arrangement is a little trickier to define for subspaces arrangements. Let $\widetilde{{\mathcal{A}}''} = \{ x_0 \cap y {\mid}y \in {\mathcal{A}}' \}$. The difficulty lies in the fact that $\widetilde{{\mathcal{A}}''}$ might contain two distinct subspaces $x_0 \cap y_1$ and $x_0 \cap y_2$ such that $x_0 \cap y_1 \subsetneqq x_0 \cap y_2$. Since we are mainly interested in the complement space, it makes sense to define the *restricted arrangement* (in $x_0$) by $${\mathcal{A}}'' = \widetilde{{\mathcal{A}}''} \setminus \{ x_0 \cap y {\mid}y \in {\mathcal{A}}' \text{ and } \exists y' \in {\mathcal{A}}' \text{ with } y \neq y' \text{ and } x_0 \cap y \subset x_0 \cap y' \}.$$ It makes sense to consider the graded differential algebra $D(\widetilde{{\mathcal{A}}''})$, even if $\widetilde{{\mathcal{A}}''}$ is not a proper subspaces arrangement. Later in this section, the corollary \[cor:artilde\] states that the natural inclusion $D({\mathcal{A}}'') \hookrightarrow D(\widetilde{{\mathcal{A}}''})$ is a quasi-isomorphism. The key to this corollary \[cor:artilde\] is contained in the next proposition. \[prop:subar\] Suppose ${\mathcal{B}}$ is a subspaces arrangement with two subspaces $u$ and $v$ such that $u \subset v$. Let ${\mathcal{B}}' = {\mathcal{B}}\setminus \{u \}$. Then the inclusion $D({\mathcal{B}}') \hookrightarrow D({\mathcal{B}})$ is a quasi-isomorphism. The injective map $D({\mathcal{B}}') \to D({\mathcal{B}})$ gives us the following short exact sequence of cochain complexes : $$0 \to D({\mathcal{B}}') \to D({\mathcal{B}}) \to \frac{D({\mathcal{B}})}{D({\mathcal{B}}')} \to 0.$$ Since there is a long exact sequence in cohomology, it is sufficient to prove that $H^\star(D({\mathcal{B}})/D({\mathcal{B}}')) = 0$. The cochain complexes $D({\mathcal{B}})/D({\mathcal{B}}')$ is a vector space generated by all the elements of the form - $\sigma = \{u, x_{i_1}, \dotsc, x_{i_r}\}$ with $x_{i_j} \not \in \{u,v\}$. The differential $d\sigma$ is a sum of some $ \pm \{u, x_{i_1}, \dotsc, \widehat{x_{i_j}}, \dotsc, x_{i_r}\}$. - $\sigma = \{u, v, x_{i_1}, \dotsc, x_{i_r}\}$ with $x_{i_j} \not \in \{u, v\}$. The differential $d\sigma$ is a sum of $\{u, x_{i_1}, \dotsc, x_{i_r}\}$ and some elements $\pm \{u,v, \dotsc, \widehat{x_{i_j}}, \dotsc, x_{i_r}\}$. Let’s introduce a new grading on $D({\mathcal{B}})/D({\mathcal{B}}')$. We say that $\deg\{u, x_{i_1}, \dotsc, x_{i_r}\} = r$ and $\deg\{u,v, x_{i_1}, \dotsc, x_{i_r}\} = r + 1$. The differential decreases this degree by one, and we have a chain complex. Let’s introduce a filtration on $D({\mathcal{B}})/D({\mathcal{B}}')$ : $F_0$ is the subcomplex generated by $\{u\}$ and $\{u,v\}$, and in general, $F_r$ is generated by all the $\{u, x_{i_1}, \dotsc, x_{i_s}\}$ and $\{u,v, x_{i_1}, \dotsc, x_{i_s}\}$ with $s \leq r$. So, we have an increasing filtration $F_0 \subseteq F_1 \subseteq \dotso \subseteq F_r \subseteq \dotso$ Let’s define the filtration degree $d_0$ of $\sigma$ by the number of $x_{i_j}$ in $\sigma$, and the complementary degree is 0 if $v \not \in \sigma$ and $1$ if $v \in \sigma$. This gives us a spectral sequence. The page $E^0$ is : - $E^0_{p, 0}$ is a vector subspace generated by all the $\{u, x_{i_1}, \dotsc, x_{i_p}\}$. - $E^0_{p, 1}$ is generated by all the $\{u,v, x_{i_1}, \dotsc, x_{i_p}\}$. - $E^0_{p, q}$ is zero if $q \geq 2$. The differential $d_0{\colon}E^0_{p,q} \to E^0_{p, q-1}$ is zero for every $q \neq 1$. For $q = 1$, $d_0 {\colon}E^0_{p, 1} \to E^0_{p, 0}$ maps $\{u,v, x_{i_1}, \dotsc, x_{i_p}\}$ to $\{u, x_{i_1}, \dotsc, x_{i_p}\}$. So, it is obvious that every spaces $E^1_{p,q}$ of the next page is simply zero. Therefore, there is no homology in $D({\mathcal{B}})/D({\mathcal{B}}')$, which completes the proof. \[cor:artilde\] Let ${\mathcal{A}}$ be a subspaces arrangement. Then the inclusion $D({\mathcal{A}}'') \hookrightarrow D(\widetilde{{\mathcal{A}}''})$ is a quasi-isomorphism. It is a consequence of the definitions and proposition \[prop:subar\]. A quasi-isomorphism connecting $D({\mathcal{A}})/D({\mathcal{A}}')$ and $D(\widetilde{{\mathcal{A}}''})$ {#sec:quasiiso} ======================================================================================================== The goal of this section will be to show the existence of a quasi-isomorphism $\frac{D({\mathcal{A}})}{D({\mathcal{A}}')} \to D(\widetilde{{\mathcal{A}}''})$. This will give us a long exact sequence connecting the rational cohomology of $M({\mathcal{A}}), M({\mathcal{A}}')$ and $M({\mathcal{A}}'')$. But before that, we need to define a few things. There is an equivalence relation in ${\mathcal{A}}'$ : $x_i {\sim}x_j$ if and only if $x_0 \cap x_i = x_0 \cap x_j$. This equivalence relation gives a partition of ${\mathcal{A}}' = {\mathcal{A}}_1 \cup {\mathcal{A}}_2 \cup \dotso \cup {\mathcal{A}}_r$. In order to get the signs right, the linear order on the elements of ${\mathcal{A}}$ is chosen with the property that if $i < j$, $y \in {\mathcal{A}}_i$ and $z \in {\mathcal{A}}_j$, then $y < z$. Let $E = \{ (y,z) \in {\mathcal{A}}' \times {\mathcal{A}}' {\mid}y {\sim}z \text{ and } y \neq z \}$. Let’s define the map ${\varphi}{\colon}D({\mathcal{A}})\to D(\widetilde{{\mathcal{A}}''})$. Let $\sigma \subseteq {\mathcal{A}}$. If $x_0 \not \in \sigma$ or if $x_0 \in \sigma$ and there exists $(y,z) \in E$ such that $\{y,z\} \in \sigma$, then ${\varphi}(\sigma) = 0$. Otherwise, $\sigma = \{x_0, x_{i_1}, \dotsc, x_{i_r}\} \subseteq {\mathcal{A}}$ is sent to ${\varphi}(\sigma) = (-1)^r \{x_0 \cap x_{i_1}, \dotsc, x_0 \cap x_{i_r} \} \subseteq \widetilde{{\mathcal{A}}''}$ (in particular, ${\varphi}(\{x_0\}) = \emptyset$). This map is not multiplicative but it commutes with the differential. The map ${\varphi}$ is such that ${\varphi}d = d {\varphi}$. Let $\sigma \subseteq {\mathcal{A}}$. We need to check the following situations. - If $x_0 \not\in \sigma$, then $d{\varphi}\sigma = 0 = {\varphi}d \sigma$. - if $x_0 \in \sigma$ and there exists $(y,z) \in E$ with $\{y, z\} \in \sigma$, then $d {\varphi}\sigma = 0$ and $$\begin{gathered} d\{x_0,y,z, x_{i_1}, \dotsc, x_{i_r} \} = \alpha \left(\sigma \setminus \{x_0 \} \right) + \{x_0, z, x_{i_1}, \dotsc, x_{i_r}\} \\ - \{x_0, y, x_{i_1}, \dotsc, x_{i_r} \} + \sum_j \alpha_j \left(\sigma \setminus\{x_{i_j}\} \right), \end{gathered}$$ where $\alpha$ and the $\alpha_j$ can be $0$ or $\pm 1$. So, by definition of ${\varphi}$, we have : ${\varphi}d\sigma = 0$. - Finally, the last possibility is that $x_0 \in \sigma$ and no $(y,z) \in E$ is such that $\{y,z\} \subset \sigma$. In that case, let $\sigma = \{x_0, x_{i_1}, \dotsc, x_{i_r}\}$ and denote by $J_\sigma$ the set $\{j {\mid}\cap_{m = 1, m \neq j}^r (x_{i_m} \cap x_0) = \cap_{m=1}^r (x_{i_m} \cap x_0) \}$. We have $$\begin{aligned} {\varphi}(d \sigma) & = {\varphi}\left( \alpha \left(\sigma \setminus \{x_0\}\right) + \sum_{j \in J_\sigma} (-1)^{j+1}\sigma \setminus \{x_{i_j}\}\right) \\ &= (-1)^{r-1} \sum_{j \in J_\sigma} (-1)^{j+1} \{x_0 \cap x_{i_1}, \dotsc, \widehat{x_0 \cap x_{i_j}}, \dotsc, x_0 \cap x_{i_r} \}, \\ d {\varphi}(\sigma) &= (-1)^r d\{ x_0 \cap x_{i_1}, \dotsc, x_0 \cap x_{i_r} \} \\ &= (-1)^r \sum_{j \in J_\sigma} (-1)^j \{x_0 \cap x_{i_1}, \dotsc, \widehat{x_0 \cap x_{i_j}}, \dotsc, x_0 \cap x_{i_r} \}.\end{aligned}$$ So, the map ${\varphi}$ commutes with the differential $d$. The map ${\varphi}$ is surjective. Obviously, it satisfies ${\varphi}j = 0$ and induces a surjective map $$\bar{{\varphi}} {\colon}\frac{D({\mathcal{A}})}{D({\mathcal{A}}')} \to D(\widetilde{{\mathcal{A}}''})$$ We’ll show that this map induces an isomorphism in cohomology. If the set $E$ is empty, then $\bar{{\varphi}}$ is injective and is an isomorphism. Otherwise, for every $(u,v) \in E$, let $I_{u,v}$ be the vector subspace of $D({\mathcal{A}})/D({\mathcal{A}}')$ generated by all the elements of the form $\{x_0,u, y_{j_1}, \dotsc, y_{j_r} \} - \{x_0, v, y_{j_1}, \dotsc, y_{j_r} \}$ and $\{x_0, u, v, y_{j_1}, \dotsc, y_{j_r} \}$ with the $y_i \in {\mathcal{A}}\setminus \{x_0,u,v\}$. Let $V = \sum_{(u,v) \in E} I_{u,v}$. We have the following lemmas : \[lem:vkernel\] In $D(A)/D(A')$, we have $V = \ker \bar{{\varphi}}$. It is clear that $V \subseteq \ker \bar{{\varphi}}$. Let $u = \sum_{s \in I} \alpha_s \sigma_s \in \ker \bar{{\varphi}}$, with $\alpha_s \neq 0 \; \forall s \in I$ and $\sigma_s \neq \sigma_t$ if $s \neq t$. To show that $u \in V$, we will show that we can substract elements of $V$ from $u$ until we obtain zero. - If there exists a $s \in I$ such that $\sigma_s$ is of the form $\{x_0, u, v, y_{j_1}, \dotsc, y_{j_r}\}$ with $x_0 \cap u = x_0 \cap v$ and $u \neq v$, then $\alpha_s \sigma_s \in V$ and we can substract it from $u$. Let’s repeat this operation until we obtain a sum $\sum_{s \in I'} \alpha_s \sigma_s$ without any such $\sigma_s$. - Let $t \in I'$. The element $\sigma_t$ can only be of the form $\{x_0, x_{i_1}, \dotsc, x_{i_r}\}$ with $x_0 \cap x_{i_j} \neq x_0 \cap x_{i_k}$ for $j \neq k$. We have $$0 = \bar{{\varphi}}\left(\sum_{s \in I'} \alpha_s \sigma_s\right) = (-1)^r \alpha_t \{x_0 \cap x_{i_1}, \dotsc, x_0 \cap x_{i_r}\} + \sum_{s \in I' \setminus \{t\}} \alpha_s \bar{{\varphi}}(\sigma_s).$$ So, there exists a $s \in I' \setminus\{t\}$ such that $\bar{{\varphi}}(\sigma_s) = \bar{{\varphi}}(\sigma_t)$. This is possible only if $\sigma_s = \{x_0, y_1, \dotsc, y_r\}$ with $x_0 \cap x_{i_j} = x_0 \cap y_j$. In that case, the difference $\sigma_s - \sigma_t$ can be rewritten as $$\sum_{j {\mid}y_j \neq x_{i_j}} \{x_0, y_1, \dotsc, y_{j-1}, y_j, x_{i_{j+1}}, \dotsc, x_{i_r} \} - \{x_0, y_1, \dotsc, y_{j-1}, x_{i_j}, x_{i_{j+1}}, \dotsc, x_{i_r}\},$$ which is in $V$. We can substract $\alpha_t(\sigma_t - \sigma_s)$ from the sum and we get another sum $\sum_{s \in I''} \alpha_s \sigma_s$ such that $|I''| < |I'|$. This process can be repeated until we obtain an empty sum. Therefore, $u \in V$ and $V = \ker \bar{{\varphi}}$. Let $(u,v) \in E$. It is clear that $d(I_{u,v}) \subset I_{u,v}$, so $I_{u,v}$ form a subcomplex of $D({\mathcal{A}})/D({\mathcal{A}}')$. \[lem:iacyclic\] For every $(u,v) \in E$, the complex $I_{u,v}$ is acyclic. Let’s consider a cocycle of the form $$\alpha = \sum_i \alpha_i\left(\{x_0, u, y_{i_1}, \dotsc, y_{i_r}\} - \{x_0, v, y_{i_1}, \dotsc, y_{i_r}\} \right) + \sum_j \beta_j \{x_0, u, v, z_{j_1}, \dotsc, z_{j_{r-1}}\},$$ with the $y_{i_k}, z_{j_k} \not \in \{x_0, u, v\}$. We want to show that $\alpha$ is a coboundary. Let $$\begin{aligned} K_i &= \big\{k \in \{i_1, \dotsc, i_r\} {\mid}x_0 \cap u \cap y_{i_1} \cap \dots \cap y_{i_r} = \\ & \qquad \qquad \qquad x_0 \cap u \cap y_{i_1} \cap \dots \cap \widehat{y_{i_k}} \dots \cap y_{i_r}\big\}, \\ L_j &= \big\{ k \in \{j_1, \dotsc, j_{r-1}\} {\mid}x_0 \cap u \cap v \cap z_{j_1} \cap \dots \cap z_{j_{r-1}} = \\ & \qquad \qquad \qquad x \cap u \cap v \cap z_{j_1} \cap \dots \cap \widehat{z_{j_k}} \dots \cap z_{j_{r-1}} \big\}.\end{aligned}$$ Since $\alpha$ is a cocycle, we have : $$\begin{aligned} 0 = d\alpha =& \sum_i \alpha_i\sum_{k \in K_i} (-1)^k \big( \{x_0, u, y_{i_1}, \dotsc \widehat{y_{i_k}} \dotsc, y_{i_r}\} - \{ x_0, v, y_{i_1}, \dotsc \widehat{y_{i_k}} \dotsc, y_{i_r}\}\big) \\ & + \sum_j \beta_j\big(\{x_0, u, z_{j_1}, \dotsc, z_{j_{r-1}} \} - \{x_0, v, z_{j_1}, \dotsc, z_{j_{r-1}} \} \big) \\ & + \sum_j \beta_j \sum_{k \in L_j} (-1)^{k+1} \{x_0, u, v, z_{j_1}, \dotsc \widehat{z_{j_k}} \dotsc, z_{j_{r-1}} \}.\end{aligned}$$ It implies that $$\sum_i \alpha_i\sum_{k \in K_i} (-1)^k \{x_0, u, v, y_{i_1}, \dotsc \widehat{y_{i_k}} \dotsc, y_{i_r}\} + \sum_j \beta_j \{ x_0, u, v, z_{j_1}, \dotsc, z_{j_{r-1}} \} = 0.$$ Let’s consider $\omega = - \sum_i \alpha_i \{x_0, u, v, y_{i_1}, \dotsc, y_{i_r}\}$. We have $$\begin{aligned} d\omega &= \sum_i \alpha_i \big( \{x_0, u, y_{i_1}, \dotsc, y_{i_r}\} - \{ x_0, v, y_{i_1}, \dotsc, y_{i_r}\}\big) \\ & \qquad - \sum_i \alpha_i\sum_{k \in K_i} (-1)^k \{x_0, u, v, y_{i_1}, \dotsc \widehat{y_{i_k}} \dotsc, y_{i_r}\} \\ & = \alpha.\end{aligned}$$ Hence, $\alpha$ is a coboundary and $I_{u,v}$ is acyclic. \[lem:acyclic\] The vector space $V$ is such that $d(V) \subseteq V$ and $V$ is acyclic. It is obvious that $d(V) \subseteq V$. Showing that $V$ is acyclic is technical. We’ll do it in three steps. First, we choose $r$ elements $(u_i, v_i) \in E$ that have some nice properties. Then, using these elements, we construct a sequence of vector subspaces $I_0 \subsetneqq I_1 \subsetneqq I_r = V$. Finally, we prove by induction that all these subspaces are acyclic. Let’s remember that the linear order respects the equivalence relation ${\sim}$ on ${\mathcal{A}}'$ in the following sense : the equivalence classes are ${\mathcal{A}}' = {\mathcal{A}}_1 \cup {\mathcal{A}}_2 \cup \dotso \cup {\mathcal{A}}_r$ and if $i < j$, $y \in {\mathcal{A}}_i$ and $z \in {\mathcal{A}}_j$, then $y < z$. Let ${\mathcal{A}}_1 = \{x_{j_1}, \dotsc, x_{j_2 - 1}\}$, ${\mathcal{A}}_2 = \{x_{j_2}, \dotsc, x_{j_3 - 1}\}, \dotsc, {\mathcal{A}}_r = \{x_{j_r}, \dotsc, x_{j_{r+1} - 1}\}$. Let ${\mathcal{A}}_i$ be the first equivalence class with more than 1 element and $(u_0, v_0) = (x_{j_i}, x_{j_i + 1})\in E$, $(u_1, v_1) = (u_0, x_{j_i + 2}), \dotsc, (u_m, v_m) = (u_0, x_{j_{i+1} - 1})$. Let ${\mathcal{A}}_k$ be the next equivalence class with more than 1 element and $(u_{m+1}, v_{m+1}) = (x_{j_k}, x_{j_k+1}), (u_{m+2}, v_{m+2}) = (x_{j_k}, x_{j_k + 2}), \dotsc$ We can do this until we have a sequence $(u_i, v_i)_{0 \leq i \leq r} \subseteq E$ with the following properties : $V = \sum_{i=0}^r I_{u_i, v_i}$ and $v_n \not\in \{u_0, v_0, u_1, v_1, \dotsc, u_{n-1}, v_{n-1} \}$. Let $I_0 = I_{u_0, v_0}$ and $I_j = \sum_{i = 0}^j I_{u_i, v_i}$. The properties of the $(u_i, v_i)$ implies that we have an increasing sequence $$I_{u_0, v_0} =I_0 \subsetneqq I_1 \subsetneqq \dotso \subsetneqq I_r = V.$$ Lemma \[lem:iacyclic\] shows that $I_0$ is acyclic. Suppose $I_{n-1}$ acyclic and let’s show that $I_n = I_{n-1} + I_{u_n, v_n}$ is acyclic as well. The key observation is that every element of $I_{n-1}$ can be written as a sum $\alpha_1 + \alpha_2 + d\omega$ where $\alpha_1 \in I_{n-1}, \alpha_2 \in I_{u_n, v_n}$, $\omega \in I_{n-1}$ and there is no $v_n$ in any of the terms of $\alpha_1$. It is sufficient to prove this for every element of a generating sequence of $I_{n-1}$ : - Let $\sigma = \{x_0, u_i, v_i, \dotsc \}$. If $v_n \in \sigma$ and $u_n \in \sigma$, then $\sigma = 0 + \sigma + d(0)$. If $v_n \in \sigma$ and $u_n \not \in \sigma$, then $d(\sigma \cup \{u_n\}) = \pm \sigma \pm (\sigma \cup \{u_n\}) \setminus \{v_n\} + S$ where $S$ is a sum with $\{u_n, v_n \}$ in each term, which means that $S$ is in $I_{u_n, v_n}$. It shows that we have the required decomposition $\sigma = \pm (\sigma \cup \{u_n \}) \setminus \{v_n\} \pm S \pm d(\sigma \cup \{u_n\})$. - Let’s consider $\alpha = \{x_0, u_i, y_{j_1}, \dotsc, y_{j_r}\} - \{x_0, v_i, y_{j_1}, \dotsc, y_{j_r}\}$. If $v_n \in \{y_{j_1}, \dotsc, y_{j_r}\}$ and $u_n \not\in \{y_{j_1}, \dotsc, y_{j_r}\}$, then form $\omega = \{x_0, u_i, y_{j_1}, \dotsc, y_{j_r} \} \cup \{ u_n \} - \{x_0, v_i, y_{j_1}, \dotsc, y_{j_r}\} \cup \{u_n\}$. As in the first case, we obtain a good decomposition from $d\omega$. This observation allow us to show that every cocycle in $I_n$ is a coboundary. Let $\alpha \in I_n$ such that $d\alpha = 0$. Since $I_n = I_{n-1} + I_{u_n, v_n}$, we can write $\alpha = \alpha_1 + \alpha_2 + d\omega$ with $\alpha_1 \in I_{n-1}, \alpha_2 \in I_{u_n, v_n}$, $\omega \in I_n$ and no $v_n$ in any term of $\alpha_1$. The vector $\alpha_2$ can be written $$\alpha_2 = \sum_i \gamma_i\left(\{x_0, u_n, y_{i_1}, \dotsc, y_{i_r}\} - \{x_0, v_n, y_{i_1}, \dotsc, y_{i_r}\} \right) + \sum_j \beta_j \{x_0,u_n,v_n, z_{j_1}, \dotsc, z_{j_{r-1}}\}.$$ For appropriate sets $K_i$ and $L_j$ (as in lemma \[lem:iacyclic\]), its differential is $$\begin{aligned} d\alpha_2 =& \sum_i \gamma_i\sum_{k \in K_i} (-1)^k \big( \{x_0, u_n, y_{i_1}, \dotsc \widehat{y_{i_k}} \dotsc, y_{i_r}\} - \{ x_0, v_n, y_{i_1}, \dotsc \widehat{y_{i_k}} \dotsc, y_{i_r}\}\big) \\ & + \sum_j \beta_j\big(\{x_0,u_n, z_{j_1}, \dotsc, z_{j_{r-1}} \} - \{x_0, v_n, z_{j_1}, \dotsc, z_{j_{r-1}} \} \big) \\ & + \sum_j \beta_j \sum_{k \in L_j} (-1)^{k+1} \{x_0, u_n, v_n, z_{j_1}, \dotsc \widehat{z_{j_k}} \dotsc, z_{j_{r-1}} \}.\end{aligned}$$ But there is no term in $\alpha_1$ containing $v_n$ and $d\alpha_1 + d\alpha_2 = 0$. Therefore, $d\alpha_1 = d\alpha_2 = 0$. By the induction hypothesis, there exists a $\omega_1$ such that $d\omega_1 = \alpha_1$ and by lemma \[lem:iacyclic\], there exists a $\omega_2$ such that $d\omega_2 = \alpha_2$. So, $\alpha = d(\omega_1 + \omega_2 + \omega)$, $I_n$ is acyclic, which proves that $V$ is acyclic. Together, these lemmas give us the following proposition. \[prop:phiquasiiso\] The map $\bar{{\varphi}} {\colon}D({\mathcal{A}})/D({\mathcal{A}}') \to D(\widetilde{{\mathcal{A}}''})$ is a quasi-isomorphism. By lemma \[lem:vkernel\], we can factorize $\bar{{\varphi}}$ as shown in the following diagram. $$\xymatrix{\frac{D({\mathcal{A}})}{D({\mathcal{A}}')} \ar[r]^{\bar{{\varphi}}} \ar[d]_p & D(\widetilde{{\mathcal{A}}''}) \\ \frac{D({\mathcal{A}})}{D({\mathcal{A}}') \oplus V} \ar[ru]_{\tilde{{\varphi}}} & }$$ where $\tilde{{\varphi}}$ is the quotient map. Since $V = \ker \bar{{\varphi}}$ and $\bar{{\varphi}}$ is surjective, $\tilde{{\varphi}}$ is an isomorphism. By lemma \[lem:acyclic\], the map $p$ is a quasi-isomorphism, which implies that $\bar{{\varphi}}$ is a quasi-isomorphism as well. A long exact sequence in rational cohomology {#sec:drt} ============================================ The quasi-isomorphism $\bar{{\varphi}} {\colon}D({\mathcal{A}})/D({\mathcal{A}}') \to D(\widetilde{{\mathcal{A}}''})$, described in section \[sec:quasiiso\], decreases the degree by $2 \operatorname{codim}x_0 - 1 = \deg(x_0)$. Therefore, we have the following theorem. \[theo:lesic\] Let ${\mathcal{A}}$ be a subspaces arrangement, $x_0 \in {\mathcal{A}}$ with ${\mathcal{A}}'$ and ${\mathcal{A}}''$ the deleted and the restricted arrangements. Then, there is a long exact sequence in cohomology  : $$\begin{gathered} \dotso \to H^q(M({\mathcal{A}}'), {\mathbb{Q}}) \to H^q(M({\mathcal{A}}), {\mathbb{Q}}) \to \\ H^{q-\deg(x_0)} (M({\mathcal{A}}''), {\mathbb{Q}}) \to H^{q+1}(M({\mathcal{A}}'), {\mathbb{Q}}) \to \dotso\end{gathered}$$ It is a direct consequence from the short exact sequence $0 \to D({\mathcal{A}}') \to D({\mathcal{A}}) \to D({\mathcal{A}})/D({\mathcal{A}}') \to 0$, corollary \[cor:artilde\] and proposition \[prop:phiquasiiso\]. As a first consequence, the Euler characteristic of the complement of non empty subspaces arrangements is zero. \[cor:euler\] Let ${\mathcal{A}}$ be a subspaces arrangement in ${\mathbb{C}}^l$. Then, the Euler characteristic of $M({\mathcal{A}})$ satisfies : $\chi(M({\mathcal{A}})) = 0$ if ${\mathcal{A}}$ is non-empty, and $\chi(M({\mathcal{A}})) = 1$ if ${\mathcal{A}}$ is empty. If ${\mathcal{A}}$ is empty, then $M({\mathcal{A}}) = {\mathbb{C}}^l$ and $\chi(M({\mathcal{A}})) = 1$. If $|{\mathcal{A}}| = 1$, then $M({\mathcal{A}}) = {\mathbb{C}}^l \setminus \{x_0\}$ for some vector subspace $x_0$. So, $M({\mathcal{A}})$ has the homotopy type of an odd-dimensional sphere and $\chi(M({\mathcal{A}})) = 0$. By the universal coefficient theorems (in homology and cohomology), we have the following equality : $b_i(M({\mathcal{A}})) = \dim H^i(M({\mathcal{A}}), {\mathbb{Q}})$, where $b_i(M({\mathcal{A}}))$ is the $i$th Betti number of $M({\mathcal{A}})$. The long exact sequence in cohomology from theorem \[theo:lesic\] gives us the following formula : $\chi(M({\mathcal{A}})) = \chi(M({\mathcal{A}}')) + (-1)^{\deg(x_0)}\chi(M({\mathcal{A}}'')) = \chi(M({\mathcal{A}}')) - \chi(M({\mathcal{A}}''))$. Now, let’s prove by induction on $|{\mathcal{A}}|$ that $\chi(M({\mathcal{A}})) = 0$. We suppose that this is true for arrangements with $n$ or less subspaces. Let ${\mathcal{A}}$ be an arrangement with $n+1$ subspaces and $x_0 \in {\mathcal{A}}$. The deleted and restricted arrangements are such that $|{\mathcal{A}}'| = |{\mathcal{A}}| - 1$ and $1 \leq |{\mathcal{A}}''| < |{\mathcal{A}}|$. So, by induction, $\chi(M({\mathcal{A}})) = \chi(M({\mathcal{A}}')) - \chi(M({\mathcal{A}}'')) = 0 - 0 = 0$. The special case of arrangements of hyperplanes is well known (see [@orl92]). If ${\mathcal{A}}$ is a complex arrangement of hyperplanes, not only do we know that $\chi(M({\mathcal{A}})) = 0$, but we also know that the Poincaré polynomials of a triple $({\mathcal{A}}, {\mathcal{A}}', {\mathcal{A}}'')$ verify the relation : $\operatorname{Poin}(M({\mathcal{A}}), t) = \operatorname{Poin}(M({\mathcal{A}}'), t) + t\operatorname{Poin}(M({\mathcal{A}}''), t)$. Therefore, a natural question arises : can we extend this formula to the general case of subspaces arrangements? A generalized formula is equivalent to the splitting of the long exact sequence of theorem \[theo:lesic\] in short exact sequences $$0 \to H^q(M({\mathcal{A}}'), {\mathbb{Q}}) \to H^q(M({\mathcal{A}}), {\mathbb{Q}}) \to \\ H^{q-\deg(x_0)} (M({\mathcal{A}}''), {\mathbb{Q}}) \to 0.$$ This would give us the following formula : $$\label{eq:formula} \operatorname{Poin}(M({\mathcal{A}}), t) = \operatorname{Poin}(M({\mathcal{A}}'), t) + t^{\deg(x_0)}\operatorname{Poin}(M({\mathcal{A}}''), t).$$ But in general, the equality \[eq:formula\] does not hold, as shown by the following example. *Example 1.* In ${\mathbb{C}}^5$, let’s consider the arrangement ${\mathcal{A}}= \{h_0, h_1, h_2\}$ where $h_0, h_1$ and $h_2$ are defined by the equations $z_1 = z_5 = 0$, $z_1 = z_2 = z_3 = 0$ and $z_3 = z_4 = z_5 = 0$, respectively. A basis of the cochain complex $D({\mathcal{A}})$ is described in table \[tab:descar\]. From that, it is easy to see that $\operatorname{Poin}(M({\mathcal{A}}), t) = 1 + t^3 + 2t^5 + 2t^6$ (and that $\chi(M({\mathcal{A}})) = 0$). Let’s consider ${\mathcal{A}}' = {\mathcal{A}}\setminus \{h_0\}$ and the corresponding restricted arrangement ${\mathcal{A}}''$. Some quick computations yield $\operatorname{Poin}(M({\mathcal{A}}'), t) = 1 + 2t^5 + t^8$, $\operatorname{Poin}(M({\mathcal{A}}''), t) = 1 + 2t^3 + t^4$ and $$\operatorname{Poin}(M({\mathcal{A}}'), t) + t^3\operatorname{Poin}(M({\mathcal{A}}''), t) = 1 + t^3 + 2t^5 + 2t^6 + t^7 + t^8.$$ $\deg(x)$ $x$ $dx$ $\deg(x)$ $x$ $dx$ ----------- ----------------- ---------------------- ----------- --------------------- --------------------------------------- 0 $\{\emptyset\}$ $d\{\emptyset\} = 0$ 6 $\{h_0, h_1\}$ $d\{h_0, h_1\} = 0$ 3 $\{h_0\}$ $d\{h_0\} = 0$ 6 $\{h_0, h_2\}$ $d\{h_0, h_2\} = 0$ 5 $\{h_1\}$ $d\{h_1\} = 0$ 7 $\{h_0, h_1, h_2\}$ $d\{h_0, h_1, h_2\} = - \{h_1, h_2\}$ 5 $\{h_2\}$ $d\{h_2\} = 0$ 8 $\{h_1, h_2\}$ $d\{h_1, h_2\} = 0$ : Description of $D({\mathcal{A}})$ for example 1.[]{data-label="tab:descar"} So, the formula \[eq:formula\] isn’t always true. \[def:sep\] Let ${\mathcal{A}}$ be a subspaces arrangement in ${\mathbb{C}}^l$. We say that $x_0$ satisfies the Poincaré polynomial property (in short : PP property) if for the corresponding triple $({\mathcal{A}}, {\mathcal{A}}', {\mathcal{A}}'')$, we have $$\operatorname{Poin}(M({\mathcal{A}}), t) = \operatorname{Poin}(M({\mathcal{A}}'), t) + t^{\deg(x_0)}\operatorname{Poin}(M({\mathcal{A}}''), t).$$ In the example, $h_0$ does not satisfy the PP property. However, if we consider the triple $({\mathcal{A}}, \widehat{{\mathcal{A}}}', \widehat{{\mathcal{A}}}'')$ associated to $h_1$, then the Poincaré polynomials are : $$\begin{aligned} \operatorname{Poin}(M(\widehat{{\mathcal{A}}'}), t) &= 1 + t^3 + t^5 + t^6 \\ t^5 \operatorname{Poin}(M(\widehat{{\mathcal{A}}'}), t) &= t^5(1 + t) \\ \operatorname{Poin}(M({\mathcal{A}}), t) &= 1 + t^3 + 2t^5 + 2t^6.\end{aligned}$$ The formula \[eq:formula\] is thus satisfied for $h_1$. This example illustrates that, unlike with hyperplanes arrangements, not every subspace $x_0$ has the PP property. But the arrangement ${\mathcal{A}}$ of the example shows that even if a given subspace has not the PP property, then sometimes, we can find another subspace which possesses that property. At this point, it seems natural to ask the following questions : 1. From any (non-empty) subspaces arrangement ${\mathcal{A}}$, can we always choose a $x \in {\mathcal{A}}$ satisfying the PP property? 2. Can we find simple equivalent conditions for $x \in {\mathcal{A}}$ to have the PP property? 3. For any arrangement of hyperplanes ${\mathcal{A}}$, each hyperplane has the PP property. Can we find a larger class of subspaces arrangements for which this is true? For example, is this true for arrangements with a geometric lattice? Notice that the arrangement ${\mathcal{A}}$ of example 1 does not have a geometric lattice. To simplify the notations, for $I = \{i_1, \dotsc, i_r\} \subseteq \{0, \dotsc, n\}$, we’ll denote $x_{i_1} \cap \dotso \cap x_{i_r}$ by $x_{\{i_1, \dotsc, i_r\}}$, or sometimes simply $x_I$. The following definition is a generalization of the concept of separator in [@orl92]. Let ${\mathcal{A}}= \{x_0, \dotsc, x_n\}$ be a subspaces arrangement. We say that $x_0$ is a *separator* if $x_0 \cap x_{\{1, 2, \dotsc, n\}} > x_{\{1, 2, \dotsc, n\}}$ (in the lattice $L({\mathcal{A}})$). Now, a partial answer to question 2 is given by the following proposition. \[prop:sep\] Let ${\mathcal{A}}$ be a subspaces arrangement in ${\mathbb{C}}^l$ and $x_0 \in {\mathcal{A}}$. If $x_0$ is a separator, then $x_0$ has the PP property To prove that $x_0$ has the PP property, it is sufficient to show that the injective map $j {\colon}D({\mathcal{A}}') \to D({\mathcal{A}})$ induces an injective map $j^\star {\colon}H^\star(D({\mathcal{A}}')) \to H^\star(D({\mathcal{A}}))$, since, in that case, the long exact sequence of theorem \[theo:lesic\] splits into short exact sequences $$0 \to H^q(M({\mathcal{A}}'), {\mathbb{Q}}) \to H^q(M({\mathcal{A}}), {\mathbb{Q}}) \to \\ H^{q-\deg(x_0)} (M({\mathcal{A}}''), {\mathbb{Q}}) \to 0.$$ Let’s define a map $k {\colon}D({\mathcal{A}}) \to D({\mathcal{A}}')$ of cochain complexes by $k(\sigma) = 0$ if $x_0 \in \sigma$ and $k(\sigma) = \sigma$ if $x_0 \not\in \sigma$. This map commutes with the differential : - If $x_0 \not\in \sigma$, then $dk(\sigma) = kd(\sigma)$, - if $x_0 \in \sigma$, then $dk(\sigma) = 0$ and $kd(\sigma) = \alpha (\sigma \setminus \{x_0 \})$, with $\alpha = 0$ if and only if $\cap_{x \in {\mathcal{A}}} x \neq \cap_{y \in {\mathcal{A}}'} y$, which is the case. So, the map $k$ is a map of chain complexes and induces a map $k^\star {\colon}H^\star(D({\mathcal{A}})) \to H^\star(D{\mathcal{A}}'))$. Obviously, we have $k^\star j^\star = \operatorname{id}$, which implies that $j^\star$ is injective. Hence $x_0$ has the PP property. In the example 1, we can use the proposition \[prop:sep\] to check that $h_1$ and $h_2$ are separators. But question 2 is not completely answered : the following example is an arrangement ${\mathcal{B}}$ which has the PP property with respect to $h_0$. However, $h_0$ is not a separator. *Example 2.* In ${\mathbb{C}}^3$, let’s consider ${\mathcal{B}}= \{h_0, h_1, h_2\}$ where $h_0, h_1$ and $h_2$ are the three lines defined by the equations $z_2 = z_3 = 0$, $z_1 = z_3 = 0$ and $z_1 = z_2 = 0$, respectively. A basis for $D({\mathcal{B}})$ is $1$ in degree 0, $X_i = \{h_i\}$ in degree 3, $X_{ij} = \{h_i, h_j\}$ in degree 4 and $X_{012}$ in degree 3, with $d1 = dX_i = dX_{ij} = 0$ and $dX_{012} = -X_{12} + X_{02} - X_{01}$. From that, it is easy to compute $$\begin{aligned} \operatorname{Poin}(M({\mathcal{B}}), t) &= 1 + 3t^3 + 2t^4 = 1 + 2t^3 + t^4 + t^3(1 + t) \\ &= \operatorname{Poin}(M({\mathcal{B}}'), t) + t^{\deg(h_0)} \operatorname{Poin}(M({\mathcal{B}}''), t).\end{aligned}$$ Arrangements with a geometric lattice {#sec:agl} ===================================== When we suppose that the arrangement ${\mathcal{A}}$ has a geometric lattice, we have some stronger results. These results come from the following lemma. \[lem:prep\] Let ${\mathcal{A}}= \{x_0, \dotsc, x_n\}$ be an arrangement with a geometric lattice such that $x_0$ is a separator. Let $\{i_1, \dotsc, i_{r+1}\} \subset \{1, \dotsc, n\}$. Then we have $$x_0 \cap x_{\{i_1, \dotsc, i_r\}} = x_0 \cap x_{\{i_1, \dotsc, i_{r+1}\}} \text{ if and only if } x_{\{i_1, \dotsc, i_r\}} = x_{\{i_1, \dotsc, i_{r+1}\}}.$$ If $x_{\{i_1, \dotsc, i_r\}} = x_{\{i_1, \dotsc, i_{r+1}\}}$, then obviously, we have $x_0 \cap x_{\{i_1, \dotsc, i_r\}} = x_0 \cap x_{\{i_1, \dotsc, i_{r+1}\}}$. Suppose that $x_0 \cap x_{\{i_1, \dotsc, i_r\}} = x_0 \cap x_{\{i_1, \dotsc, i_{r+1}\}}$. From $x_0 \cap x_{\{1, \dotsc, n\}} \neq x_{\{1, \dotsc, n\}}$, we deduce that $x_0 \cap x_{\{i_1, \dotsc, i_r\}} \neq x_{\{i_1, \dotsc, i_r\}}$. Since the lattice is geometric, we obtain the maximal chain $x_{\{i_1, \dotsc, i_r\}} < x_0 \cap x_{\{i_1, \dotsc, i_r\}}$ from the maximal chain ${\mathbb{C}}^l < x_0$. Therefore, we have the two following chains of inequalities : $$\begin{aligned} x_{\{i_1, \dotsc, i_r\}} < x_0 \cap x_{\{i_1, \dotsc, i_r\}} & = x_0 \cap x_{\{i_1, \dotsc, i_{r+1}\}}, \\ x_{\{i_1, \dotsc, i_r\}} \leq x_{\{i_1, \dotsc, i_{r+1}\}} & < x_0 \cap x_{\{i_1, \dotsc, i_{r+1}\}}.\end{aligned}$$ Since the lattice is geometric, all maximal chains between two elements have the same length. The first line is a maximal chain of length 2, so the second line has the same length and we have the equality $x_{\{i_1, \dotsc, i_r\}} = x_{\{i_1, \dotsc, i_{r+1}\}}$. Suppose that the arrangement ${\mathcal{A}}$ satisfies the assumptions of lemma \[lem:prep\] and let’s define a map $\theta {\colon}D({\mathcal{A}}') \to D(\widetilde{{\mathcal{A}}''})$ by sending $\{x_{i_1}, \dotsc, x_{i_r}\}$ to $\{x_0 \cap x_{i_1}, \dotsc, x_0 \cap x_r\}$. This map is a bijection : if $x_0 \cap x_i = x_0 \cap x_j$, then $x_0 \cap x_i = x_0 \cap x_{\{i,j\}}$, and by lemma \[lem:prep\], $x_i = x_{\{i,j\}}$, which is impossible. Lemma \[lem:prep\] implies that $\theta$ commutes with the differential. So, the induced map in cohomology $H^\star \theta {\colon}H^\star(M({\mathcal{A}}'), {\mathbb{Q}}) \to H^\star(M({\mathcal{A}}''), {\mathbb{Q}})$ is an isomorphism of vector spaces. But this map $H^\star \theta$ does not preserve the multiplication, or even the grading. A simple computation shows that the element $x = \{x_{i_1}, \dotsc, x_{i_r}\}$ of degree $\deg(x)$ (in $H^\star(M({\mathcal{A}}'), {\mathbb{Q}})$) is sent to $\{x_0 \cap x_{i_1}, \dotsc, x_0 \cap x_{i_r}\}$, which has a degree $\deg(x) - 2\operatorname{codim}(x_0+x_{\{i_1, \dotsc, i_r\}})$ (in $H^\star(M({\mathcal{A}}''), {\mathbb{Q}})$). So, it decreases the degree by an even number. This proves the following proposition : Let ${\mathcal{A}}= \{x_0, \dotsc, x_n\}$ be an arrangement with a geometric lattice such that $x_0$ is a separator. Then we have the following relations : $$\begin{aligned} &\sum_{i=0}^\infty b_{2i}(M({\mathcal{A}}')) = \sum_{i=0}^\infty b_{2i}(M({\mathcal{A}}'')), && \sum_{i=0}^\infty b_{2i+1}(M({\mathcal{A}}')) = \sum_{i=0}^\infty b_{2i+1}(M({\mathcal{A}}'')), \\ &\sum_{i=0}^\infty b_{i}(M({\mathcal{A}}')) = \sum_{i=0}^\infty b_{i}(M({\mathcal{A}}'')), && \forall n \in {\mathbb{N}}, \sum_{i=0}^n b_{i}(M({\mathcal{A}}')) \leq \sum_{i=0}^n b_{i}(M({\mathcal{A}}'')).\end{aligned}$$ Finally, when the subspace $x_0$ is an hyperplane, we have a stronger result. Let ${\mathcal{A}}= \{x_0, \dotsc, x_n\}$ be an arrangement with a geometric lattice such that $x_0$ is an hyperplane and a separator, then the graded vector spaces $H^\star(M({\mathcal{A}}'), {\mathbb{Q}})$ and $H^\star(M({\mathcal{A}}''), {\mathbb{Q}})$ are isomorphic and $$\begin{aligned} \operatorname{Poin}(M({\mathcal{A}}'), t) &= \operatorname{Poin}(M({\mathcal{A}}''), t),\\ \operatorname{Poin}(M({\mathcal{A}}), t) &= (1+t) \operatorname{Poin}(M({\mathcal{A}}''), t). \end{aligned}$$ Since $x_0$ is an hyperplane and $x_0 \cap x_{\{1, \dotsc, n\}} \neq x_{\{1, \dotsc, n\}}$, we have $\operatorname{codim}(x_0+x_{\{i_1, \dotsc, i_r\}}) = 0$. Therefore, the map $H^\star \theta$ preserve the degree and is an isomorphism of graded vector spaces. This fact and proposition \[prop:sep\] implies directly the two equalities. [xx]{} <span style="font-variant:small-caps;">P. Orlik, H. Terao</span>, *Arrangements of Hyperplanes*, Springer-Verlag, 1992. <span style="font-variant:small-caps;">E. Feichtner, S. Yuzvinsky</span>, *Formality of the complements of subspace arrangements with geometric lattices*, <arXiv:math.AT/0504321>, 2005.

ArXiv

--- abstract: 'We study the geometry of germs of definable (semialgebraic or subanalytic) sets over a $p$-adic field from the metric, differential and measure geometric point of view. We prove that the local density of such sets at each of their points does exist. We then introduce the notion of distinguished tangent cone with respect to some open subgroup with finite index in the multiplicative group of our field and show, as it is the case in the real setting, that, up to some multiplicities, the local density may be computed on this distinguished tangent cone. We also prove that these distinguished tangent cones stabilize for small enough subgroups. We finally obtain the $p$-adic counterpart of the Cauchy-Crofton formula for the density. To prove these results we use the Lipschitz decomposition of definable $p$-adic sets of [@CCL] and prove here the genericity of the regularity conditions for stratification such as $(w_f)$, $(w)$, $(a_f)$, $(b)$ and $(a)$ conditions.' address: - 'Université Lille 1, Laboratoire Painlevé, CNRS - UMR 8524, Cité Scientifique, 59655 Villeneuve d’Ascq C’edex, France, and, Katholieke Universiteit Leuven, Department of Mathematics, Celestijnenlaan 200B, B-3001 Leuven, Belgium\' - 'Laboratoire J.-A. Dieudonné, Université de Nice - Sophia Antipolis, Parc Valrose, 06108 Nice Cedex 02, France (UMR 6621 du CNRS)' - '[É]{}cole Normale Sup[é]{}rieure, D[é]{}partement de math[é]{}matiques et applications, 45 rue d’Ulm, 75230 Paris Cedex 05, France (UMR 8553 du CNRS)' author: - Raf Cluckers - Georges Comte - François Loeser title: 'Local metric properties and regular stratifications of $p$-adic definable sets' --- Introduction {#introduction .unnumbered} ============ The present paper is devoted to the study of local metric properties of definable subsets of the $p$-adic affine space, with special stress on the local density of these subsets. It contains also results on tangent cones and existence of regular stratifications. We shall start by recalling what is known in the real and complex context regarding the local density of (sub-)analytic sets. When $X_a$ is a germ at $a$ of a complex analytic subset $X$ of real dimension $d$ of the affine space $\CC^n $, the local density $\Theta_{d}(X_a)$ of $X_a$, sometimes called in this setting the Lelong number of $X_a$, has been introduced by Lelong in [@Lelong] as the limit of volumes of the intersection of a representative $X$ of the germ with suitably renormalized balls around $a$, namely $$\Theta_{d}(X_a)=\lim_{r\to 0} \frac{\mu_{d}({X}\cap B(a,r))}{\mu_{d}(B^{d}(0,r))},$$ where $ B^{d}(0,r)$ is the real $d$-dimensional ball of centre $0$ and radius $r>0$ and $\mu_d$ stands for $d$-dimensional volume. Lelong actually proved that the function $$\displaystyle r \mapsto \frac{\mu_d(X\cap B(a,r))}{r^d}$$ decreases as $r$ goes to $0$, pointing out, long before this concept has been formalized, the tame behaviour of the local normalized volume of analytic sets. Ten years after Lelong’s pioneering paper, Thie proved in [@Thie] that the local density of a complex analytic subset $X$ at a point $a$ is a positive integer by expressing it as a sum of local densities of the components, counted with multiplicities, of the tangent cone of $X$ at $a$. Finally, more than twenty years after Lelong’s definition, Draper proved [@Draper] that the local density is the algebraic multiplicity of the local ring of $X$ at $a$. The definition of local densities has been extended by Kurdyka and Raby to real subanalytic subsets of $\RR^n$ in [@KR] (see also [@KPR]). In fact, although the arguments in [@KR] and [@KPR] which prove the existence of the density in the real subanalytic case were given before the notion of definable sets in o-minimal structures emerged, they apply to the real o-minimal setting. A short proof of this existence result, again produced just before the concept of tame definable sets and involving the Cauchy-Crofton formula, may be found in the seminal paper [@Lio] for semi-pfaffian sets. In [@lk] (Theorem 1.3), one can find a proof in the real o-minimal setting of the existence of the local density viewed as the higher term of a finite sequence of localized curvature invariants involving the Cauchy-Crofton formula and the theory of regular stratifications. Of course, in the real setting, the density is in general no longer a positive integer, but a non negative real number, and Kurdyka and Raby proved an appropriate extension of Thie’s result by expressing again the local density in terms of the density of some components of the real tangent cone. The existence of the local density at each point of the closure of a subanalytic set is a manifestation of tameness of these sets near their singular points. Similarly, tameness in subanalytic geometry is also illustrated by the tame behavior of local density, viewed as a function of the base-point of the germ at which it is computed: it is actually proved in [@coliro] that this function along a given global subanalytic set is a Log-analytic function, that is, a polynomial in subanalytic functions and their logarithms (see Siu’s paper [@Siu] for similar results in the complex case). Draper’s result has been extended to the real setting by Comte in [@comte] in the following way. Recall that if $X$ is a complex analytic subset of the affine space of complex dimension $d$ at $x$, the algebraic multiplicity of the local ring of $X$ at $x$ is equal to the local degree of a generic linear projection $p: X \rightarrow \CC^d$, that is, to the number of points near $x$ in a generic fiber of $p$. Over the reals, if $X$ is of local dimension $d$ at $x$, the number of points in fibers (near $x$) of a generic linear projection $p: X \rightarrow \RR^d$ is not generically constant near $p (x)$ in general. The idea introduced in [@comte] to overcome this difficulty is to consider as a substitute for the local degree of $p$ the sum $d (p) :=\sum_{i \in \NN} i \cdot\theta_i$, where $\theta_i$ is the local density at $p(x)$ of the germ of the set of points in $\RR^d$ over which the fiber of $p$ has exactly $i$ points near $x$. The so-called [*local Cauchy-Crofton formula*]{} proved in [@comte] states that the average along all linear projections $p$ of the degrees $d (p)$ is equal to the local density of $X$ at $x$ and can be considered as the real analogue of Draper’s result. Finally, the complete multi-dimensional version of the local Cauchy-Crofton formula for real subanalytic sets is presented in [@lk] (Theorem 3.1), where the multi-dimensional substitute of the $0$-dimensional local degree $d(p)$ is obtained by considering the local Euler characteristic of generic multi-dimensional fibers, instead of the local number of points. Now let $K$ be a finite extension of $\QQ_p$ and $X$ be a definable subset (semi-algebraic or subanalytic) of $K^n$. Let $x$ be a point of $K^n$. When one tries to define the local density of $X$ at $x$ similarly to the archimedean case, one is faced to the problem, illustrated in \[fstart\], that the limit of local volumes in general no longer exists. It appears that the normalized volumes $v_n$ of $X$ in the balls $B(x,n):=\{w\in K\mid \ord (x-w)\geq n\}$ has a periodic convergence, that is to say, there exists an integer $e>0$, such that for all $c=0, \cdots, e-1$, $(v_{c+m\cdot e})_{m\in \NN} $ has a limit $v_c$ in $\QQ $ (see Proposition \[lem:pos\]), with possibly $v_c\not=v_{c'}$, for $c\not=c'$. The reader having essentially in mind the real case is thus strongly encouraged to start reading this article by the example studied in \[fstart\] that emphasizes this phenomenon. We resolve that issue by using an appropriate renormalization device that leads us to express the mean value $\displaystyle \frac{1}{e}\cdot \sum_{c=0}^{e-1} v_c$ as the local density of $X$ at $x$. Another new issue occurring in the $p$-adic setting is the lack of a natural notion of a tangent cone. Unlike the real case where only the action of the multiplicative group $\RR_+^\times$ has to be considered, in the $p$-adic case, there seems to be no preferential subgroup of $K^\times$ at hand. We remedy this by introducing, for each definable open subgroup of finite index $\Lambda$ in $K^{\times}$, a tangent cone $C_x^{\Lambda}$ at $x$ which is stable by homotheties in $\Lambda$, that is, which is a $\Lambda$-cone. One should note that such $\Lambda$-cones were already considered more than twenty years ago in the work of Heifetz on $p$-adic oscillatory integrals and wave front sets [@Heifetz]. Nevertheless, we prove in Theorem \[distinguished cone\] that, given a definable subset $X$, among these cones, some are distinguished as maximal for an inclusion property, and appear as the good tangent cones to be considered, in the sense that they capture the local geometry of our set. We are then able, by deformation to the tangent cone, to assign multiplicities to points in the tangent cone $C_x^{\Lambda}$. Our main results regarding $p$-adic local densities are Theorem \[mt\], which is a $p$-adic analogue of the result of Thie and Kurdyka-Raby, and Theorem \[lcc\], which is a $p$-adic analogue of Comte’s local Cauchy-Crofton formula. An important technical tool in our proof of Theorem \[mt\] is provided by Theorem \[Lipschitz decomposition\] which allows us to decompose our definable set into Lipschitz graphs. Such a regular decomposition has been obtained in [@CCL] and extends to the $p$-adic setting a real subanalytic result of [@KurWExp]. In section \[swf\] we prove the existence of $(w_f)$-regular stratifications for definable $p$-adic functions, and consequently the existence of Thom’s $(a_f)$-regular stratifications for definable $p$-adic functions, $(w)$-regular, or Verdier regular, stratifications, and Whitney’s $(b)$-regular stratifications for $p$-adic definable sets (Theorem \[wf\]). [During the preparation of this paper, the authors have been partially supported by grant ANR-06-BLAN-0183. We also thank the Fields Institute of Toronto, where this paper was partly written, for bringing us exceptional working conditions.]{} Preliminaries ============= Definable sets over the $p$-adics {#defpadic} --------------------------------- Let $K$ be a finite field extension of $\QQ_p$ with valuation ring $R$. We denote by $\ord$ the valuation and set $\vert x \vert := q^{- \ord (x)}$ and $\vert 0 \vert = 0$, with $q$ cardinality of the residue field of $K$. If $x = (x_i)$ is a point in $K^m$ and $n$ is an integer, we denote by $B(x,n)$ the ball in $K^m$ given by the conditions $\ord (z_i - x_i) \geq n$, $1 \leq i \leq m$. We recall the notion of (globally) subanalytic subsets of $K^n$ and of semi-algebraic subsets of $K^n$. Let $\cL_{\rm Mac}=\{0,1,+,-,\cdot,\{P_n\}_{n>0}\}$ be the language of Macintyre and $\cL_{\rm an}=\cL_{\rm Mac}\cup \{^{-1},\cup_{m>0}K\{x_1,\ldots,x_m\}\}$, where $P_n$ stands for the set of $n$th powers in $K^\times$, where $^{-1}$ stands for the field inverse extended on $0$ by $0^{-1}=0$, where $K\{x_1,\ldots,x_m\}$ is the ring of restricted power series over $K$ (that is, formal power series converging on $R^m$), and each element $f$ of $K\{x_1,\ldots,x_m\}$ is interpreted as the restricted analytic function $K^m\to K$ given by $$x\mapsto \begin{cases} f(x) & \mbox{if }x\in R^m \\ 0 & \mbox{else.} \end{cases}$$ By subanalytic we mean $\cL_{\rm an}$-definable with coefficients from $K$ and by semi-algebraic we mean $\cL_{\rm Mac}$-definable with coefficients from $K$. Note that subanalytic, resp. semi-algebraic, sets can be given by a quantifier free formula with coefficients from $K$ in the language $\cL_{\rm Mac}$, resp. $\cL_{\rm an}$. In this section we let $\cL$ be either the language $\cL_{\rm Mac}$ or $\cL_{\rm an}$ and by $\cL$-definable we will mean semi-algebraic, resp. subanalytic when $\cL$ is $\cL_{\rm Mac}$, resp. $\cL_{\rm an}$. Everything in this paper will hold for both languages and we will give appropriate references for both languages when needed. For each definable set $X\subset K^n$, let $\cC(X)$ be the $\QQ$-algebra of functions on $X$ generated by functions $|f|$ and $\ord(f)$ for all definable functions $f:X\to K^\times$. We refer to [@sd] and [@DvdD] for the definition of the dimension of $\cL$-definable sets. The $p$-adic measure {#pmes} -------------------- Suppose that $X\subset K^n$ is an $\cL$-definable set of dimension $d\geq 0$. The set $X$ contains a definable nonempty open $K$-analytic submanifold $X'\subset K^n$ such that $X\setminus X'$ has dimension $<d$, cf. [@DvdD]. There is a canonical $d$-dimensional measure $\mu_d$ on $X'$ coming from the embedding in $K^n$, which is constructed as follows, cf. [@serre]. For each $d$-element subset $J$ of $\{1,\ldots,n\}$, with $j_i<j_{i+1}$, $j_i$ in $J$, let $dx_J$ be the $d$-form $dx_{j_1}\wedge\ldots\wedge dx_{j_d}$ on $K^n$, with $x=(x_1,\ldots,x_n)$ standard global coordinates on $K^n$. Let $x_0$ be a point on $X'$ such that $x_I$ are local coordinates around $x_0$ for some $I\subset\{1,\ldots,n\}$. For each $d$-element subset $J$ of $\{1,\ldots,n\}$ let $g_J$ be the $\cL$-definable function determined at a neigborhood of $x_0$ in $X'$ by $g_Jdx_I=dx_J$. There is a unique volume form $\vert \omega_{0}\vert_{X'}$ on $X'$ which is locally equal to $(\max_J|g_J|) \vert dx_I \vert$ around every point $x_0$ in $X'$. Indeed, $\vert \omega_{0}\vert_{X'}$ is equal to $\sup_{J}\vert dx_J \vert$. The canonical $d$-dimensional measure $\mu_d$ on $X'$ (cf. [@serre] [@oesterle]), is the one induced by the volume form $\vert \omega_{0}\vert_{X'}$. We extend this measure to $X$ by zero on $X \setminus X'$ and still denote it by $\mu_d$. Adding sorts {#sr} ------------ By analogy with the motivic framework, we now expand the language $\cL$ to a three sorted language $\cL'$ having $\cL$ as language for the valued field sort, the ring language $\LL_{\rm Rings}$ for the residue field, and the Presburger language $\LL_{\rm PR}$ for the value group together with maps $\ord$ and $\ac$ as in [@cons]. By taking the product of the measure $\mu_m$ with the counting measure on $k_K^n \times \ZZ^r$ one defines a measure still denoted by $\mu_m$ on $K^m \times k_K^n \times \ZZ^r$. One defines the dimension of an $\cL'$-definable subset $X$ of $K^m\times k_K^n\times\ZZ^r$ as the dimension of its projection $ p(X) \subset K^m$. If $X$ is of dimension $d$, one defines a measure $\mu_d$ on $X$ extending the previous construction on $X$ by setting $$\mu_d (W) := \int_{ p (X)} p_! (\11_W) \mu_d$$ with $p_! (\11_W)$ the function $y \mapsto {\rm card}(p^{-1}(y) \cap W)$. For such an $X$, one defines $\cC (X)$ as the $\QQ$-algebra of functions on $X$ generated by functions $\alpha$ and $q^{- \alpha}$ with $\alpha : X \rightarrow \ZZ$ definable in $\cL'$. Note that this definition coincides with the previous one when $n = r = 0$. Since $\cL'$ is interpretable in $\cL$, the formalism developed in this section extends to $\cL'$-definable objects in a natural way. $p$-adic Cell Decomposition --------------------------- Cells are defined by induction on the number of variables \[def::cell\] An $\cL$-cell $A\subset K$ is a (nonempty) set of the form $$\{t\in K\mid |\alpha|\square_1 |t-c|\sq_2 |\beta|,\ t-c\in \lambda P_n\},$$ with constants $n>0$, $\lambda,c$ in $ K$, $\alpha,\beta$ in $K^\times$, and $\square_i$ either $<$ or no condition. An $\cL$-cell $A\subset K^{m+1}$, $m\geq0$, is a set of the form $$\label{Eq:cell:decay} \begin{array}{ll} \{(x,t)\in K^{m+1}\mid & x\in D, \ |\alpha(x)|\sq_1 |t-c( x)|\sq_2 |\beta(x)|,\\ & t-c(x)\in \lambda P_n\}, \end{array}$$ with $(x,t)=(x_1,\ldots, x_m,t)$, $n>0$, $\lambda$ in $ K$, $D=p_m(A)$ a cell where $p_m$ is the projection $K^{m+1}\to K^m$, $\cL$-definable functions $\alpha,\beta:K^m\to K^\times$ and $c:K^m\to K$, and $\square_i$ either $<$ or no condition, such that the functions $\alpha,\beta$, and $c$ are analytic on $D$. We call $c$ the center of the cell $A$ and $\lambda P_n$ the coset of $A$. In either case, if $\lambda=0$ we call $A$ a $0$-cell and if $\lambda\not=0$ we call $A$ a $1$-cell. In the $p$-adic semi-algebraic case, Cell Decomposition Theorems are due to Cohen [@cohen] and Denef [@D84], [@Dcell] and they were extended in [@Ccell] to the subanalytic setting where one can find the following version. \[thm:CellDecomp\] Let $X\subset K^{m+1}$ and $f_j:X\to K$ be $\cL$-definable for $j=1,\ldots,r$. Then there exists a finite partition of $X$ into $\cL$-cells $A_i$ with center $c_i$ and coset $\lambda_i P_{n_i}$ such that $$|f_j(x,t)|= |h_{ij}(x)|\cdot|(t-c_i(x))^{a_{ij}}\lambda_i^{-a_{ij}}|^\frac{1}{n_i},\quad \mbox{ for each }(x,t)\in A_i,$$ with $(x,t)=(x_1,\ldots, x_m,t)$, integers $a_{ij}$, and $h_{ij}:K^m\to K$ $\cL$-definable functions which are analytic on $p_m(A_i)$, $j=1,\ldots,r$. If $\lambda_i=0$, we use the convention that $a_{ij}=0$. Let us also recall the following lemma from [@Cexp]. \[prop:descrip:simple\] Let $X\subset K^{m+1}$ be $\cL$-definable and let $G_j$ be functions in $\cC(X)$ in the variables $(x_1,\ldots,x_m,t)$ for $j=1,\ldots,r$. Then there exists a finite partition of $X$ into $\cL$-cells $A_i$ with center $c_i$ and coset $\lambda_i P_{n_i}$ such that each restriction $G_j|_{A_i}$ is a finite sum of functions of the form $$|(t-c_i(x))^{a}\lambda_i^{-a}|^\frac{1}{n_i}\ord (t-c_i(x))^{s}h(x),$$ where $h$ is in $\cC(K^m)$, and $s\geq 0$ and $a$ are integers. The following $p$-adic curve selection lemma is due to van den Dries and Scowcroft [@sd] in the semi-algebraic case and to Denef and van den Dries [@DvdD] in the subanalytic case. The statement is the $p$-adic counterpart of the semi-algebraic or subanalytic curve selection lemma over the reals. \[curve\] Let $A$ be a definable subset of $K^n$ and let $x$ be in $\overline A$. Then there exists a definable function $f=(f_1,\ldots,f_n):R\to K^n$ such that the $f_i$ are given by power series (over $K$) converging on $R$, such that $f(0)=x$, and such that $f(R\setminus \{0\})\subset A$. The following is a $p$-adic analogue of a classical lemma by Whitney (see [@Whitney]). \[lemWhitney\]Let $g:R\to K^n$ be a map given by $n$ analytic power series over $K$, converging on $R$, such that the map $g$ is nonconstant. Then, the limit $\ell\in\PP^{n-1}(K)$ for $r\to 0$ of the lines $\ell_r\in\PP^{n-1}(K)$ connecting $g(0)$ with $g(r)$ exists. Also the limit $\ell'$ of the tangent lines $\ell_r':=\{g(r)+\lambda (\partial g_1/\partial r,\ldots,\partial g_1/\partial r)_{|r}\mid \lambda\in K\}$ for $r \to 0 $ exists and $\ell'=\ell$. Since $g$ is nonconstant, for $r\not=0$ close to $0$ one has $g(r)\not = g(0)$ and $(\partial g_1/\partial r,\ldots,\partial g_1/\partial r)_{|r}\not =0$, and hence, $\ell_r$ and $\ell'_r$ are well-defined for $r\not=0$ close to $0$. We may suppose that $g(0)=0$ and that each of the $g_i$ is nonconstant. Write $g_i(r)=\sum_{j\geq 0} a_{ij}r^j$ with $a_{ij}\in K$ and for each $i$, let $k_i$ be the smallest index $j$ such that $a_{ij}\not=0$. Then $k_i>0$ for each $i$ since $g(0)=0$. Let $k$ be the minimum of the $k_i$. Then clearly $\ell$ and $\ell'$ are the same line $\ell$ connecting $0$ and $(a_{1k},\ldots,a_{nk})\not=0$. Indeed, the line $\ell_r$ connects $0$ and $g(r) $ which is equivalent to connecting $0$ and $g(r)/r^k$; the point $g(r)/r^k$ converges to $(a_{1k},\ldots,a_{nk})$ and thus $\ell_r$ converges to the line $\ell$. Likewise, the line $\ell'_r$ connects $0$ and $(\partial g_1/\partial r,\ldots,\partial g_1/\partial r)_{|r}$ which is equivalent to connecting $0$ and $$\frac{1}{kr^k} (\partial g_1/\partial r,\ldots,\partial g_1/\partial r)_{|r};$$ the point $\frac{1}{kr^k} (\partial g_1/\partial r,\ldots,\partial g_1/\partial r)_{|r}$ converges to $(a_{1k},\ldots,a_{nk})$ and thus also $\ell_r'$ converges to $\ell$ when $r\to 0$. Fix two integers $d \leq m$. Let $U$ be an open definable subset of $K^d$ and let $\varphi$ be a definable analytic mapping $U \rightarrow K^{m - d}$. We view the graph $\Gamma (\varphi)$ of $\varphi$ as a definable subset of $K^m$. Let $\varepsilon$ be a positive real number. We say that $\varphi$ is $\varepsilon$-analytic, if the norm $| D \varphi | = \max_{i,j} |\partial \varphi_i/\partial x_j|$ of the differential of $\varphi$ is less or equal than $\varepsilon$ at every point of $U$. For $\varepsilon>0$, call a function $f:D\to K^{m}$ on a subset $D$ of $K^n$ $\varepsilon$-Lipschitz when for all $x, y \in D$ one has $$|f(x)-f(y)|\leq \varepsilon |x-y|.$$ The function $f$ is called locally $\varepsilon$-Lipschitz when for each $x\in D$ there exists an open subset $U$ of $K^n$ containing $x$ such that the restriction of $f$ to $U\cap D$ is $\varepsilon$-Lipschitz. \[ana-lip\] Let $U$ be open in $K^n$ and $f:U\to K^m$ a function which is $\varepsilon$-analytic. Then $f$ is locally $\varepsilon$-Lipschitz. Choose $u\in U$, and a basic neighborhood $U_u$ of $u$ in $U$ such that the component functions $f_i$ of $f$ are given by converging power series on $U_u$, where basic neighborhood means a ball of the form $c+ \lambda R^n$ with $c\in K^n$ and $\lambda \in K^\times$. We may suppose that $U_u=R^n$, that $u=0$, and that $\varepsilon=1$. We may also assume that for each $i,j$, the partial derivative $\partial f_j(x)/\partial x_i $ is bounded in norm by $1$ on $U_u$. Since $|\partial f_j(x)/\partial x_i (0)|\leq 1$, it follows that the linear term of $f_j$ in $x_i$ has a coefficient of norm $\leq 1$ for each $i,j$. By the convergence of the power series, the coefficients of the $f_j$ are bounded in norm, say by $N$, and we can put $U':=\{x\in R^n\mid |x|<1/N\}$. Clearly $U'$ contains $u=0$. By the non-archimedean property of the $p$-adic valuation, the restriction of $f$ to $U'$ is $1$-Lipschitz. For more results related to Lipschitz continuity on the $p$-adics, see [@CCL] or Theorem \[Lipschitz decomposition\] below. The following lemma is a partial converse of Lemma \[ana-lip\], especially in view of the fact that any definable function is piecewise analytic. \[ana-eps\] Let $U$ be a definable open in $K^n$ and let $f:D\to K^m$ be a definable analytic function which is locally $\varepsilon$-Lipschitz. Then $f$ is $\varepsilon$-analytic. We proceed by contradiction. Suppose that $|D f|>\varepsilon$ at $u\in U$. Choose a basic neighborhood $U_u$ of $u$ in $U$ such that the component functions $f_i$ of $f$ are given by converging power series on $U_u$, (here again by basic neighborhood we mean a ball of the form $c+\lambda R^n$ with $c\in K^n$ and $\lambda \in K^\times$). We may suppose that $U_u=R^n$, that $u=0$, and that $\varepsilon=1$. By assumption, we have for some $i,j$ that $|(\partial f_j(x)/\partial x_i)(0)|>1 $, hence, the linear term of $f_j$ in $x_i$ has a coefficient of norm strictly greater than $1$. By the convergence of the power series, the coefficients of the $f_j$ are bounded in norm, say by $N$, and thus for any $x$ in $\{x\in R^n\mid |x|<1/N\}$ one has $|f(0)-f(x)|>|x|$ which contradicts the fact that $f$ is $\varepsilon$-Lipschitz with $\varepsilon=1$. The following is the $p$-adic analogue of Proposition 1.4 of [@KR]. \[1.4\] Let $X$ be a definable subset of dimension $d$ of $K^m$. For every $\varepsilon > 0$, there exists a definable subset $Y$ of $X$ of dimension $< d$, $N (\varepsilon) \geq 0$, definable open subsets $U_i (\varepsilon)$ of $K^d$, for $1 \leq i \leq N (\varepsilon)$, definable, $\varepsilon$-analytic functions $\varphi_i (\varepsilon) : U_i (\varepsilon) \rightarrow K^{m - d}$, with graphs $\Gamma_i (\varepsilon)$, and elements $\gamma_1$, …, $\gamma_{N (\varepsilon)}$ in $\GL_m (R)$ such that the sets $\gamma_i ( \Gamma_i (\varepsilon))$ are all disjoint and contained in $X$, and $$X = \bigcup_{1 \geq i \geq N (\varepsilon)} \gamma_i (\Gamma_i (\varepsilon)) \cup Y.$$ Take a finite partition of $X$ into cells $X_i$. By neglecting cells of dimension $<d$, we may suppose that $X=X_1$ and that $X_1$ is a cell of type $(i_1,\ldots, i_m)$ with $\sum_{j=1}^m i_j=d$. By reordering the coordinates, by partitioning again into cells, and neglecting cells of dimension $<d$, we may suppose that $X$ is a cell of type $$(i_1,\ldots, i_m)=(1,\ldots,1,0,\ldots,0).$$ Let $p:K^m\to K^d$ be the projection to the first $d$ coordinates. The set $X$ itself is already a graph of a map $(f_1,\ldots,f_{m-d})=f:p(X)\to K^{m-d}$. We may suppose that the $f_i$ are analytic. Write $(x_1,\ldots,x_d)$ for the coordinates on $p(X)$. Consider the differentials $(\frac{\partial f_j}{\partial x_i})_{i=1}^d$. By partitioning further, we may suppose that for each $i=1,\ldots, d$, either $|\frac{\partial f_j}{\partial x_i}| \leq \varepsilon$ or $|\frac{\partial f_j}{\partial x_i}| > \varepsilon$ hold on $p(X)$. If $|\frac{\partial f_j}{\partial x_i}|> \varepsilon$ on $p(X)$ for some $i$ and $j$, then we may suppose that $$(x_1,\ldots,x_d)\mapsto (x_1,\ldots,x_{i-1},\frac{f_j(x_1,\ldots,x_d)}{\varepsilon},x_{i+1},\ldots,x_{d})$$ is a bi-analytic change of variables (cf. Corollary 3.7 of [@CCL]). By performing such change of variables successively, we are done by the chain rule for differentiation. The following lemma is classical, see, for example, [@DHM] for the semi-algebraic case and, for example, [@CLR1] for the subanalytic case. Let $\cL^\ast$ be the language $\cL$ together with a function symbol for the field inverse on $K^\times$ (extended by zero on zero), function symbols for each $n$ which stands for a (definable) $n$-th root picking function $\sqrt[n]{}$ on the $n$-th powers (extended by zero outside the $n$-th powers), and for each degree $n$ a Henselian root picking function $h_n$ for polynomials of degree $n$ in the $n+1$ coefficients (extended by zero if the conditions of Hensel’s Lemma are not fulfilled). \[terms\] Let $f:D\subset K^n\to K^m$ be an $\cL$-definable function. Then $D$ can be partitioned into finitely many definable pieces $D_i$ such that there are $\cL^\ast$-terms $t_i$ with $f(x)=t_i(x)$ for each $i$ and each $x\in D_i$. Local densities =============== A false start {#fstart} ------------- Let $X$ be a definable subset of $K^m$ of dimension $d$ and let $x$ be a point of $K^m$. Considering what is already known in the complex analytic and real o-minimal case, a natural way to define the local density of $X$ at $x = (x_1, \dots, x_m)$ would be to consider the limit of $q^{nd} \mu_d (X \cap B (x, n))$, as $n \to \infty$. Unfortunately this naïve attempt fails as is shown by the following example that we present in detail in order to caution the reader not to rely too heavily on intuition coming from the real setting. Take $X$ the subset of points of even valuation in $K$ and $x = 0$. Write $\pi_K$ for a uniformizer of $R$. The unit ball $B$ in $K$ being of measure $1$, the ball $\pi_K^\ell\cdot B = B(0,\ell)$ of radius $q^{-\ell}$ has volume $q^{-\ell}$ and, by consequence, the sphere $\pi_K^\ell\cdot S=\pi_K^\ell\cdot B\setminus \pi_K^{\ell+1}\cdot B$ of radius $q^{-\ell}$ has volume $q^{-\ell}(1-q^{-1}) $. For $k\in \NN$, let us first compute the volume of $X\cap B(0,2k)$. The set $X\cap B(0,2k)$ is the disjoint union of the spheres $\pi_K^{2j}\cdot S$ for $j\ge k$ and thus has as volume $$\mu_1(X\cap B(0,2k))=(1-q^{-1})(q^{-2k}+q^{-2k-2}+\cdots) =\frac{q^{-2k}}{1+q^{-1}}.$$ On the other hand, the set $X\cap B(0,2k-1)$ is also the disjoint union of the spheres $q^{2j}\cdot S$ for $j\ge k$ and thus has as volume $$\mu_1(X\cap B(0,2k-1))=\frac{q^{-2k}}{1+q^{-1}}.$$ We finally see that in this example the value of the limit $\displaystyle \lim_{\ell\to \infty} \frac{\mu_1(X\cap B(0,\ell))}{\mu_1(B(0,\ell))} $ depends on the parity of $\ell$, since $$\lim_{k \to \infty} q^{2k} \mu_1 (X \cap B (0, 2k)) = (1 + q^{- 1})^{-1}$$ and $$\lim_{k \to \infty} q^{2k - 1} \mu_1 (X \cap B (0, 2k - 1)) = (1 + q)^{-1}.$$ In our example one notices that the convergence of the ratio $\displaystyle \frac{\mu_1(X\cap B(0,\ell))}{\mu_1(B(0,\ell))} $ is $2$-periodic and that one may recover the expected local density, which should be $\displaystyle \frac{1}{2}$, by taking the average of the two limits. To obviate the kind of difficulty presented by this example (the periodic convergence), we are led to introduce a regularization device that we shall explain now. Mean value at infinity of bounded constructible functions --------------------------------------------------------- We will use the following elementary definition of the mean value at infinity of certain real valued functions on $\NN$. \[def:MV\] Say that a function $h:\NN\to \RR$ has a mean value at infinity if there exists an integer $e>0$ such that $$\lim_{{\genfrac{}{}{0pt}{}{n\to\infty}{n\equiv c\bmod e}}} h(n)$$ exists in $\RR$ for each $c=0,\ldots,e-1$ and in this case define the mean value at infinity of $h$ as the average $$\MV (h) := \frac{1}{e} \sum_{c=0}^{e-1}\lim_{{\genfrac{}{}{0pt}{}{n\to\infty}{n\equiv c\bmod e}}} h(n).$$ Clearly the value $\MV (h)$ is independent of the choice of the modulus $e>0$. Let $X$ be a definable subset of $K^m$, so that $X \times \NN$ is a definable subset of $K^m \times \ZZ$. Say that a real valued function $g$ on $X \times \NN$ is $X$-bounded if for every $x$ in $X$ the restriction of $g$ to $\{x\} \times \NN$ is bounded (in the sense that $g(\{x\} \times \NN)$ is contained in a compact subset of $\RR$). As has been indicated in the introduction and in the example of section \[fstart\], for an $X$-bounded function $\varphi$ in $\cC (X \times \NN)$ and $x\in X$, the function $\varphi_x:\NN\to\QQ:n\mapsto \varphi(x,n)$ may not have a unique limit for $n\to\infty$, but it may have a mean value at infinity $\MV(\varphi_x)$, as we will indeed show in Proposition \[lem:pos\]. We will moreover show in Proposition \[lem:pos\] that $\MV(\varphi_x)$, considered as a function in $x\in X$, lies in $\cC(X)$ and that $\MV(\varphi_x)$ can be calculated using a single integer $e$ as modulus when $x$ varies in $X$. \[lem:MV\] Let $\varphi$ be in in $\cC (X \times \NN)$. Suppose that, for each $x\in X$, the function $\varphi_x:\NN\to\QQ:n\mapsto \varphi(x,n)$ has finite image. Then $\varphi_x$ has a mean value at infinity $\MV (\varphi_x)$ for each $x$. Moreover, there exist a definable function $b:X\to \NN$ and an integer $e>0$ such that for all $c$ with $0\leq c< e$ and all $x\in X$, the rational number $d_c(x):=\varphi(x,n)$ is independent of $n$ as long as $n\geq b(x)$ and $n\equiv c \bmod e$. Thus, for each $x\in X$, one has $$\MV (\varphi_x) = \frac{1}{e} \sum_{c=0}^{e-1}d_c(x).$$ By consequence, the function $\MV (\varphi_x)$, considered as a function in $x\in X$, lies in $\cC(X)$. The lemma is a direct consequence of Lemma \[prop:descrip:simple\] and quantifier elimination in the three sorted language $\cL'$ of section \[sr\]. Indeed, for $\varphi\in\cC(X\times \NN)$ there exist, by Lemma \[prop:descrip:simple\] and quantifier elimination in $\cL'$, a definable function $b:X\to \NN$ and an integer $e>0$ such that for all $c$ with $0\leq c< e$ one has $$\label{varphixn} \varphi(x,n) = \sum_{i=1}^k n^{\ell_i} q^{a_{i}n} h_{ic}(x)$$ for all $x\in X$ and all $n$ with $n\geq b(x)$ and $n\equiv c \bmod e$, and where the $h_{ic}$ are in $\cC(X)$. Clearly, by regrouping, we may suppose that the pairs $(\ell_i,a_i)$ are mutually different. But then, since $\varphi_x$ has finite image for each $x\in X$, one must find $$\varphi(x,n) = h_{jc}(x)$$ for all $x\in X$ and all $n$ with $n\geq b(x)$ and $n\equiv c \bmod e$, where $j$ is such that $(\ell_j,a_j)=(0,0)$. Hence, one has $h_{jc}=d_c$ and we are done. \[lem:pos\] Let $\varphi$ be in $\cC (X \times \NN)$. Suppose that $\varphi$ is $X$-bounded. Then there exist $\varphi'$ in $\cC (X \times \NN)$ with $\lim _{n\to\infty}\varphi'(x,n)=0$ for all $x\in X$ and such that the function $$g_x: \NN\to\QQ:n\mapsto \varphi(x,n) - \varphi'(x,n)$$ has finite image. Clearly, the function $g:X\times \NN:(x,n)\mapsto g_x(n)$ lies in $\cC(X\times \NN)$. Hence, $\MV(g_x)$ and $\MV(\varphi_x)$ exist and are equal and the function $\MV (\varphi_x)$, considered as a function in $x\in X$, lies in $\cC(X)$. Also, if $\varphi\geq 0$ then $\MV (\varphi_x) \geq 0$ for all $x\in X$. Write again $\varphi$ as in (\[varphixn\]) for some integer $e$, where again the pairs $(\ell_i,a_i)$ are mutually different. Define $\varphi'(x,n)$ as the partial sum $$\sum_{i\in I} n^{\ell_i} q^{a_{i}n} h_{ic}(x)$$ for $x\in X$ and $n$ satisfying $n\geq b(x)$ and $n\equiv c \bmod e$, where $c=0,\ldots,e-1$, where $I$ consists of those $i$ with $a_i<0$. Extend $\varphi'$ to the whole of $X\times \NN$ by putting it equal to $\varphi$ for those $n$ with $n<b(x)$. Since $\varphi$ is $X$-bounded, one must have that $a_i\leq 0$ for all $i$, and, for those $i$ with $a_i=0$ one must have $\ell_i=0$. But then, we find $$g(x,n) = h_{jc}(x)$$ for all $x\in X$ and all $n$ with $n\geq b(x)$ and $n\equiv c \bmod e$, where $j$ is such that $(\ell_j,a_j)=(0,0)$. For $n$ with $n<b(x)$ one clearly has $g(x,n)=0$. The conclusions now follow from Lemma \[lem:pos\]. Local densities --------------- As already sketched in the introduction, we will define the local density of an $\cL$-definable set $X\subset K^m$ at a point $x$ as the mean value at infinity of the renormalized measure of the intersection of $X$ with the sphere of radius $q^{-n}$ around $x$. At our disposal to show that this is well-defined we have Proposition \[lem:pos\] and Lemma \[2.3.1\] below which guarantee the existence of the mean value at infinity. More generally, for a bounded function $\varphi$ in $\cC(X)$, we extend $\varphi$ to $K^m$ by zero outside $X$ and we will define the density of $\varphi$ at any point $x\in K^m$ by a similar procedure, replacing the measure by an integral of $\varphi$ on a small sphere around $x$. {#section-1 .unnumbered} Let $\varphi$ be a bounded function in $\cC (X)$, meaning that the image of $\varphi$ is contained in a compact subset of $\RR$. For $(x, n ) $ in $K^m \times \NN$ we set $$\label{gn} \gamma (\varphi) (x, n) := \int_{S(x,n)\cap X} \varphi (y) \mu_d,$$ where $S(x,n)$ is the sphere $\{y\in K^m\mid |x-y|=q^{-n}\}$ of radius $q^{-n}$ around $x$. Note that, by Lemma \[sph=ball\] below, one could as well work with balls around $x$ instead of spheres consequently in this section. By [@D85] for the semi-algebraic case and [@Ccell] for the subanalytic case, the function $\gamma (\varphi) : (x, n) \mapsto \gamma (\varphi) (x, n) $ lies in $\cC (K^m \times \NN)$. Suppose that $X$ is of dimension $d$. Then we renormalize $\gamma (\varphi)$ by dividing it by the volume of the $d$-dimensional sphere of corresponding radius and define the resulting function $\theta_d (\varphi)$ by $$\label{dp} \theta_d (\varphi) (x, n) := \frac{\gamma (\varphi) (x, n)}{\mu_d (S_{d}(n))},$$ where $S_{d}(n)$ is the $d$-dimensional sphere of radius $q^{-n}$, namely the set $\{w\in K^d\mid |w| = q^{-n}\}$. Note that $S_{d}(n)$ has measure equal to $(1-q^{-d}) q^{-nd}$ and thus, $\theta_d (\varphi)$ lies in $\cC (K^m \times \NN)$. The following lemma yields sufficient conditions for the mean value at infinity of $\theta_d (\varphi)$ to exist, in view of Proposition \[lem:pos\]. \[2.3.1\] Let $\varphi$ be a bounded function in $\cC (X)$. Assume $X$ is of dimension $d$. Then the function $\theta_d (\varphi)$ lies in $\cC (K^m \times \NN)$ and is $K^m$-bounded. That $\theta_d (\varphi)$ lies in $\cC (K^m \times \NN)$ is shown above, so we just have to show that $\theta_d (\varphi)$ is $K^m$-bounded. By the additivity of integrals and by cell decomposition, we may suppose that $X$ is a cell of dimension $d$. By changing the order of the coordinates if necessary and by Proposition \[jacprop\], we may suppose that $X$ projects isometrically to the first $d$ coordinates of $K^m$. If now $M>0$ is such that $\varphi(y)$ lies in the real interval $[-M,M]$ for all $y\in X$, then it is clear by construction that $\theta_d (\varphi)(x)$ also lies in $[-M,M]$ for all $x\in K^m$. It follows from Lemma \[2.3.1\] and Proposition \[lem:pos\] that if $\varphi$ is a bounded function in $\cC (X)$ one can set $$\label{Dd} \Theta_d (\varphi) := \MV \theta_d (\varphi),$$ that is, for $x\in K^m$, $\Theta_d (\varphi)(x)$ is the mean value at infinity of the function $n\mapsto \theta_d (\varphi)(x,n)$. By Proposition \[lem:pos\], the function $\Theta_d (\varphi)$ lies in $\cC (K^m)$. For $x$ in $K^m$, we call $\Theta_d (\varphi) (x)$ the local density of $\varphi$ at $x$. More generally, if $\varphi$ is bounded on a neighborhood of some $x\in K^m$, then $\Theta_d (\varphi) (x)$ can be defined by first extending $\varphi$ by zero outside of this neighborhood and calculate its local density by the above definitions which is clearly independent of the choice of the neighborhood. One should also note that $\Theta_d (\varphi) (x)$ is zero when $x$ does not belong to the closure of $X$. \[pTheta\] Let $X$ be a definable subset of $K^m$ of dimension $d$ and let $x$ be a point in $K^m$. We call the rational number $$\Theta_d (X) (x) := \Theta_d (\11_X) (x)$$ the local density of $X$ at $x$, where $\11_X$ is the characteristic function of $X$ which clearly lies in $\cC(K^m)$. Note that Definition \[pTheta\] resembles the definition of the complex and real density as given in the introduction, where instead of the limit $\lim_{r\to 0}$ one takes $\MV$. \[sph=ball\] Renormalizing with balls instead of with spheres yields the same local density functions $\Theta_d$. Precisely, for $\cL$-definable $X$ of dimension $d$ and for $\varphi$ a bounded function in $\cC (X)$ one has for $x\in K^m$ $$\Theta_d (\varphi) = \MV ( \theta'_d (\varphi) ),$$ where $$\theta'_d (\varphi) (x, n) := \frac{\gamma' (\varphi) (x, n)}{\mu_d (B_{d}(n))},$$ $$\gamma' (\varphi) (x, n) := \int_{B(x,n)\cap X} \varphi (y) \mu_d,$$ and where $B_{d}(n)$ is the $d$-dimensional ball of radius $q^{-n}$, namely $\{w\in K^d\mid |w|\leq q^{-n}\}$, and $B(x,n)$ is the ball $\{y\in K^m\mid |x-y|\leq q^{-n}\}$ around $x$ as defined in \[defpadic\]. In particular, $\theta'_d (\varphi)$ lies in $\cC(K^m\times \NN)$ and is $K^m$-bounded and thus its $\MV$ is well-defined. That $\theta'_d (\varphi)$ lies in $\cC(K^m\times \NN)$ and is $K^m$-bounded is proven as Lemma \[2.3.1\]. We have to prove that $ \Theta_d (\varphi) = \MV ( \theta'_d (\varphi) )$, that is, for $x\in K^m$, $\Theta_d (\varphi)(x)$ is the mean value at infinity of the function $n\mapsto \theta'_d (\varphi)(x,n)$. It is clear that $$\gamma (\varphi) (x, n) = \gamma' (\varphi) (x, n) - \gamma' (\varphi) (x, n+1)$$ and that $$\theta_d (\varphi) (x, n) = \frac{1}{(1-q^{-d})} \big( \theta_d' (\varphi) (x, n) - q^{-d}\theta_d' (\varphi) (x, n+1) \big) .$$ Now we are done by the following fact, which holds for any real constant $b\not=1$. If a function $f:\NN\to\RR$ has a mean value at infinity, then so does $g:\NN\to\RR:n\mapsto \frac{1}{1-b} ( f(n) - b f(n+1) )$, and their mean values at infinity are equal. Let us note that in the example of \[fstart\] of points of even valuation in $K$, one gets $\Theta_1 (X) (0) = \frac{1}{2}$. More generally, if $\Lambda$ is a definable open subgroup of finite index $r$ in $K^{\times}$ and $y$ is a point in $K^{\times}$, we have $\Theta_1 (\Lambda y) (0) = \frac{1}{r}$. Indeed, it is easily checked that $\Theta_1 (\Lambda y) (0)$ does not depend on $y$, hence if $y_1$, …, $y_r$ is a set of representative of $K^{\times} / \Lambda$, we have $1 = \Theta_1 (\cup_{1 \leq i \leq r} \Lambda y_i) (0) = r \Theta_1 (\Lambda) (0)$. Let $\varphi$ be a bounded function in $\cC (X)$ and assume $X$ is of dimension $d$. Denote by $\tilde \varphi$ the extension of $\varphi$ by zero on $K^m$. Then the support of $\tilde \varphi - \Theta_d (\varphi) $ is contained in an $\cL$-definable set of dimension $< d$. Suppose $X\subset K^m$. Since for $x$ not in $\overline X$, for all sufficiently large $n$ one has that $\theta_d (\varphi) (x, n)=0$, the support of $\Theta_d (\varphi)$ is contained in the closure $\overline X$ of $X$ in $K^n$. After removing a subset of dimension $<d$ we may assume $X$ is a smooth subvariety and $\varphi$ is locally constant (for example after an iterated application of Lemma \[prop:descrip:simple\]), in which case the result is clear. For further use we shall give some basic properties of local densities. \[bord\] Let $X$ be definable subset of dimension $d$ of $K^m$. Then $$\Theta_d (X) (x) = \Theta_d (\overline X) (x),$$ where $\overline X$ denotes the closure of $X$. Indeed, by additivity it is enough to prove that $\Theta_d (\overline X \setminus X) (x) = 0$, which follows from the fact that $\Theta_d (Y) (x) = 0$ when $Y$ is definable of dimension $< d$. \[convmon\] Let $X$ be an $\cL$-definable set of dimension $d$ and let $M>0$ be a constant. Consider a series of functions $\varphi_n:X\to\RR$, $n \in \NN$, such that the function $(x, n) \mapsto \varphi_n (x)$ lies in $\cC (X \times \NN)$ and such that $0 \leq \varphi_n\leq \varphi_{n+1} \leq \dots \leq M$ for all $n$. Then the function $\varphi$ defined as $\sup \varphi_n$ lies in $\cC(X)$ and is bounded. Moreover, $$\Theta_d (\varphi) (x) = \lim_n \Theta_d (\varphi_n) (x)$$ and $$0\leq \Theta_d (\varphi_n) (x)\leq \Theta_d (\varphi_{n+1}) (x)$$ for each $n$ and $x$. Clearly the function $\varphi$ is bounded and lies in $\cC(X)$ by Lemma \[lem:pos\]. Note that $$\gamma (\varphi)(x,m) = \lim_n \gamma (\varphi_n)(x,m)$$ for each $m$ and $x$ by the Monotone Convergence Theorem. Hence, by the definition of $\theta_d$, also $$\label{thetan} \theta_d (\varphi)(x,m) = \lim_n \theta_d (\varphi_n)(x,m)$$ for each $m$ and $x$. Clearly $$\label{nleq} 0\leq \gamma (\varphi_n)\leq \gamma (\varphi_{n+1})\ \mbox{ and }\ 0\leq \theta_d (\varphi_n)\leq \theta_d (\varphi_{n+1})$$ for all $n$, on the whole of $X$, and hence $$0\leq \Theta_d (\varphi_n) (x)\leq \Theta_d (\varphi_{n+1}) (x),$$ by the definition of $\Theta_d$. Now the equality $\Theta_d (\varphi) (x) = \lim_n \Theta_d (\varphi_n) (x)$ follows from (\[thetan\]), (\[nleq\]), and the definitions of $\MV$ and $\Theta_d$, by changing the order of limits over $n$ and over $m$. Tangent cones ============= Cones ----- We shall consider the set $\cD$ of open subgroups of $K^\times$ which are of finite index in $K^{\times}$. We order $\cD$ by inclusion. Note that for each $n>0$, the group $P_n$ of the $n$th powers in $K^\times$ lies in $\cD$, and any $\Lambda$ in $\cD$ equals, as a set, a finite disjoint union of cosets of some $P_n$, see Lemma \[cones1\], and is thus $\cL$-definable. We shall say a certain property (P) holds for $\Lambda$ small enough, if there exists $\Lambda_0$ in $\cD$ such that (P) holds for every $\Lambda\in \cD$ contained in $\Lambda_0$. Let $\Lambda$ be a subgroup of $K^\times$ in $\cD$. It acts naturally on $K^n$ by multiplicative translation $\lambda \cdot z := \lambda z$, that is, by scalar multiplication on the vector space $K^n$. By a $\Lambda$-cone in $K^n$ we mean a subset $C$ of $K^n$ which is stable under the $\Lambda$-action, that is, $\Lambda\cdot C \subset C$ (note that this implies that $\Lambda\cdot C = C$). More generally, if $x\in K^n$, by a $\Lambda$-cone with origin $x$ we mean a subset $C$ of $K^n $ such that $C-x$ is stable under the $\Lambda$-action, where $C-x=\{t\in K^n\mid t+x\in C\}$. By a local $\Lambda$-cone with origin $x$, we mean a set of the form $C \cap B (x, n)$, with $C$ a $\Lambda$-cone with origin $x$ and $n$ in $\NN$. In Lemma \[cones1\] we describe all possible $\Lambda$-cones which are subsets of $K$, which turns out to be very similar to the real situation. In section \[sec:localc\] we will show that definable sets in dimension $1$ locally look like local $\Lambda$-cones (Lemma \[coneone\]), and similarly in families of definable subsets of $K$ (Corollary \[corcone1\]). From \[sec:tangentc\] on we will define and study tangent cones and related objects, and formulate one of our main results on the relation between local densities of definable sets and of their tangent cones, viewed with multiplicities (the $p$-adic analogue of Thie’s result), see Theorem \[mt\]. \[cones1\] Let $C\subset K$ be a set. Then $C$ is a $\Lambda$-cone for some $\Lambda$ in $\cD$ if and only if it is either the empty set or it is a finite disjoint union of sets of the form $\lambda P_n$ with $n>0$ and $\lambda\in K$. Hence, any cone $C\subset K$ is a definable set. Clearly the empty set is a $\Lambda$-cone for all $\Lambda$ and $\lambda\cdot P_n$ is a $\Lambda$-cone for $\Lambda\subset P_n$, and similarly for their finite unions. Now let $C$ be a nonempty $\Lambda$-cone for some $\Lambda$ in $\cD$. Either $C=\{0\}$ and we are done, or, up to replacing $C$ by $tC$ for some nonzero $t\in K$, we may suppose that $1\in C$. But then $\Lambda\subset C$ and $C\setminus \Lambda$ is still a $\Lambda$-cone. Since the index of $\Lambda$ in $K^\times $ is finite, it follows by a finite process that $C$ consists of a finite union of sets of the form $\mu\Lambda$ with $\mu\in K$. It remains to prove that $\Lambda$ itself is a finite disjoint union of sets of the form $\lambda P_n$, for some $n\in\NN$ and some $\lambda\in K^\times$. Since $\Lambda$ has finite index in $K^\times$, it must contain an open neighborhood $U=1+\cM_K^\ell$ of $1$ for some $\ell>0$ and with $\cM_K$ the maximal ideal of $R$. Let $\pi_K$ be a uniformizer of $R$. Since $\Lambda$ has finite index in $K^\times$, there exists $n_1>0$ such that $\pi_K^{n_1}$ lies in $\Lambda$. Now let $n$ be a big enough multiple of $n_1$ such that $t^n\in U$ for all $t\in R^\times$. Then clearly $P_n\subset \Lambda$ and we are done since $P_n$ has finite index in $K^\times$ and hence also in $\Lambda$. Local conic structure of definable sets {#sec:localc} --------------------------------------- Let $X$ be a definable subset of $K^n$ and let $x$ be a point in $K^n$. We denote by $\pi_x : K^n \setminus \{x\}\rightarrow \PP^{n - 1} (K)$ the function which to a point $z \not=x$ assigns the line containing $x$ and $z$. That is, for $x=0$, $\pi_0 : K^n \setminus \{0\}\rightarrow \PP^{n - 1} (K)$ is the natural projection, and, for nonzero $x$, the map $\pi_x$ is the composition of $\pi_0$ with the translation $K^n\setminus \{x\}\to K^n \setminus \{0\}: y\mapsto y-x$. Furthermore we denote by $$\pi_x^X : X \setminus \{x\} \rightarrow \pi_x (X \setminus \{x\})$$ the restriction of $\pi_x$ to $X \setminus \{x\}$. \[coneone\] Let $Y$ be a definable subset of $K$. Then there exist $\Lambda$ in $\cD$ and a definable function $\gamma: K \to \NN$ such that $Y \cap B(y,\gamma(y))$ is a local $\Lambda$-cone with origin $y$, for all $y\in K$. If one writes $Y$ as a finite disjoint union of cells $Y_i$ with cosets $\lambda_i P_{n_i}$, then one can take $\Lambda=P_N$ with $N=\operatorname*{lcm}( n_i)_i$. The definability of $\gamma$ is not an issue by the definability of the conditions of being a local $\Lambda$-cone with origin $y$ and so on. By definition, a finite union of local $\Lambda$-cones is again a local $\Lambda$-cone. Hence, up to a finite partition using cell decomposition, we may suppose that $Y$ is a cell. Thus, $Y$ is of the form $$Y= \{t\in K\mid |\alpha|\sq_{1} |t-c|\sq_{2} |\beta|,\ t-c\in \lambda P_{n}\},$$ for some constants $n>0$, $\lambda,c$ in $K$, $\alpha,\beta$ in $K^\times$, and $\square_{i}$ either $<$ or no condition. Up to a transformation $t\mapsto t-c$ we may suppose that $c=0$. We may exclude the trivial case that $Y$ is a singleton, that is, we may suppose that $\lambda\not=0$. Then $Y$ is open, and moreover, $Y$ is closed if and only if $\sq_2$ is $<$. In the case that $\sq_2$ is no condition, then the closure of $Y$ equals $Y\cup\{0\}$. Take $y\in K$. If $y$ lies outside the closure of $Y$, then $Y\cap B(y,n)$ is empty for sufficiently large $n$, and the empty set is a $\Lambda$-cone for any $\Lambda$ in $\cD$. Also, if $y$ lies in the interior of $Y$, then $Y\cap B(y,n)$ is a ball around $y$ for sufficiently large $n$, and hence it is a local $\Lambda$-cone with origin $y$, for any $\Lambda$ in $\cD$. Finally, if $y=0$ and $y$ lies in the closure of $Y$, then $Y\cap B(y,n)= \lambda P_n \cap B(y,n)$ for sufficiently large $n$, which is clearly a local $\Lambda$-cone with origin $y$ for any $\Lambda$ contained in $P_n$. The following two corollaries of Lemma \[coneone\] are immediate. \[corcone1\] Let $Y$ be a definable subset of $K^{m+1}$. For each $x\in K^m$ write $Y_x$ for $\{t\in K\mid (x,t)\in Y\}$. Then there exist $\Lambda$ in $\cD$ and a definable function $\gamma: K^{m+1} \to \NN$ such that $Y_x \cap B(t,\gamma(x,t))$ is a local $\Lambda$-cone with origin $t$, for all $(x,t)\in K^{m+1}$. If one writes $Y$ as a finite disjoint union of cells $Y_i$ with cosets $\lambda_i P_{n_i}$, then one can take $\Lambda=P_N$ with $N=\operatorname*{lcm}( n_i)_i$. We will most often use the following variant of Corollary \[corcone1\], which can be proved by working on affine charts. \[localconic\] Let $X$ be a definable subset of $K^n$ and let $x$ be a point in $K^n$. Then there exist a definable function $\alpha_x : \PP^{n - 1} (K) \rightarrow \NN$, that is, $\alpha_x$ is definable on each affine chart of $\PP^{n - 1} (K)$, and a group $\Lambda$ in $\cD$ such that $$(\pi_x^X)^{-1} (\ell) \cap B (x, \alpha_x (\ell))$$ is a local $\Lambda$-cone with origin $x$ for every $\ell$ in $\PP^{n - 1} (K)$. Moreover, $\Lambda$ can be taken independently of $x$, and one can ensure that $(x,\ell)\mapsto \alpha_x(\ell)$ is a definable function from $K^n\times \PP^{n - 1} (K)$ to $\NN$. We shall call a subgroup $\Lambda$ in $\cD$ satisfying the first condition in Corollary \[localconic\] adapted to $(X, x)$, and if moreover $\Lambda$ is adapted to $(X,x)$ for all $x\in K^n$, then we call $\Lambda$ adapted to $X$. Tangent cones {#sec:tangentc} ------------- Now, if $X$ is a definable subset of $K^n$, $x$ a point of $K^n$, and $\Lambda$ in $\cD$, we define the tangent $\Lambda$-cone to $X$ at $x$ as $$\begin{split} C_x^{\Lambda} (X) := \Bigl\{u \in K^n ; (\forall i > 0) (\exists z \in X) (\exists \lambda \in \Lambda) \, \mbox{ such that } \\ \ord (z - x)>i \, \mbox{ and } \, \ord (\lambda (z - x) - u) >i \Bigr\}. \end{split}$$ By construction $C_x^{\Lambda} (X)$ is a closed, definable, $\Lambda$-cone, and, for any $n\in \NN$, $C_x^{\Lambda} (X) = C_x^{\Lambda} (X\cap B(x,n))$. Furthermore, for definable $X,Y\subset K^n$ and for $\Lambda' \subset \Lambda$ in $\cD$, one has $$C_x^{\Lambda} (X\cup Y)= C_x^{\Lambda} (X)\cup C_x^{\Lambda} (Y),$$ $$C_x^{\Lambda}(\overline {X}) =C_x^{\Lambda}(X),$$ $$C_x^{\Lambda'} (X) \subset C_x^{\Lambda} (X).$$ Although the previous inclusion might be strict, $\dim(C_x^\Lambda(X))$ does not depend on $\Lambda\in \cal D$ by Lemma \[dim\]. We comment some more on the previous inclusion in the following remarks. \[remarque sur les cones locaux\] Let $X$ be a local $\Lambda$-cone with origin $x$ in $K^n$. Thus, there exist $n$ in $\NN$ and $C$ a $\Lambda$-cone with origin $x$ such that $X=C\cap B(x,n)$. In this case for any $\Lambda' \subset \Lambda\in \cD$, one has $$C_x^{\Lambda'} (X) =C_x^{\Lambda} (X) \ (=C, \hbox{ when C is closed).}$$ Indeed, since $X=C\cap B(0,n)$, we have $C_x^{\Lambda'} (X)=C_x^{\Lambda'} (C)$. But we also have $C_x^{\Lambda'} (C)=C_x^{\Lambda} (C)$ ($=C$, when $C$ is closed). We indicate why $C_x^{\Lambda} (C)\subset C_x^{\Lambda'} (C)$. Assuming $x=0$ for simplicity, let $u\in C_0^{\Lambda} (C)$ and $i\in \NN$, $z\in C$, $\lambda \in \Lambda$, with $\ord(z)>i$ and $\ord(\lambda z-u)>i$. We have $z\in C$ and thus $\lambda z\in C$. Now let $\lambda'\in \Lambda'$ small enough to ensure that $\ord( \lambda'\lambda z)>i$. Then denoting $z'=\lambda'\lambda z$, one has $z' \in C$. From $\ord(z')>i$ and $\ord((1/\lambda')z'-u)>i$, we see that $u \in C_0^{\Lambda'} (C)$. Finally we indicate why $C_x^{\Lambda} (C)=C$, when $C$ is closed. As $C$ is stable by the $\Lambda$-action, the inclusion $C\subset C_x^{\Lambda} (C)$ is obvious. On the other hand, assuming again $x=0$, if $u\in C_x^{\Lambda} (C)$, for all $i\in \NN$, there exist $z\in C$ and $\lambda\in \Lambda$ such that $\ord(z)>i$ and $\ord(\lambda z-u)>i$. We can then construct a sequence of points $\lambda z\in C$ with limit $u$, this shows that $u\in C$, since $C$ is closed. When $X$ is a definable subset of $K$ and $x$ a point of $K$, by Lemma \[coneone\], $X$ is a local $\Lambda$-cone at $x$ with origin $x$, for some $\Lambda\in \cD$. By the above remark, for every $\Lambda' \subset \Lambda$, and still in the one dimensional case that $X\subset K$, one has $C_x^{\Lambda'} (X) = C_x^{\Lambda} (X)$. We cannot expect in general that for $X$ a definable subset of $K^n$, $n>1$, $X$ is a local $\Lambda$-cone for some $\Lambda\in \cD$, but one may at least ask, as it is the case for $n=1$, whether the stability property: “[*there exists $\Lambda\in \cD$ such that for any $\Lambda'\in \cD$, $\Lambda'\subset \Lambda$, one has $C_x^{\Lambda'} (X) = C_x^{\Lambda} (X)$*]{}" still holds for $n>1$. The answer to that question is yes, as we shall show in Theorem \[distinguished cone\]. More on $\varepsilon$-analytic functions ---------------------------------------- The following is the $p$-adic analogue of Proposition 1.7 of [@KR]. \[prop1.7KR\] Let $f:U\to K^{n-d}$ be a definable $\varepsilon$-analytic function on a nonempty open subset $U$ of $K^{d}$, $0\leq d\leq n$. Let $\Gamma$ be the graph of $f$ and let $z$ be in $\overline\Gamma$. Then, for any group $\Lambda$ in $\cD$ $$C_z^{\Lambda} (\Gamma) \subset \{(x,y)\in K^{d}\times K^{n-d}\mid |y|\leq \varepsilon |x|\}.$$ We may suppose that $z=0$. Choose $\Lambda$ in $\cD$. Since $$C_0^{\Lambda} (\Gamma) \subset C_0^{K^\times} (\Gamma),$$ by definition of $C_0^{\Lambda}(\cdot)$, we may assume that $\Lambda=K^\times$. We may also suppose that $\varepsilon=1$, after rescaling. Suppose by contradiction that there is $(x_0,y_0)$ in $C_0^{K^\times} (\Gamma)$ with $|y_0|>|x_0|+\delta$ for some $\delta>0$. Let $\Gamma'$ be the intersection of $\Gamma$ with the open subset $\{(x,y)\in K^{d+(n-d)}\mid |y|>|x|+\delta\}$. By our assumption on $(x_0,y_0)$ and by the definition of $C_0^{K^\times} (\Gamma)$, the set $\Gamma'$ is nonempty and $0$ lies in $\overline\Gamma'\setminus \Gamma'$. Apply the Curve Selection Lemma \[curve\] to the set $\Gamma'$ and the point $0$. This way we find power series $g_i$ over $K$ in one variable for $i=1,\ldots,n$, converging on $R$, such that $g(0)=0$ and $g(R\setminus \{0\})\subset \Gamma'\setminus\{0\}$. But this is in contradiction with Lemma \[lemWhitney\]. Indeed, the tangent line $\ell'_r$ at $r\not=0$ is of the form $g(r) + K\cdot t_r$ with some $t_r\in K^n$ satisfying $|y(t_r)|\leq |x(t_r)|$ by $\varepsilon$-analyticity of $f$ and the chain rule for differentiation, where $x(t_r)=(t_{r1},\ldots,t_{rd})$ and $y(t_r)=(t_{rd+1},\ldots,t_{rn})$. Hence, the limit $\ell'_0$ of the $\ell'_r$ for $0\not=r\ \to 0$ is of the same form $g(0) + K\cdot t_0$ for some $t_0\in K^n$ with $|y(t_0)|\leq |x(t_0)|$. On the other hand, the line $\ell_r$ for $r\not=0$ connecting $g(0)$ with $g(r)$ is of the form $g(r) + K\cdot u_r$ with some $u_r\in K^n$ satisfying $|y(u_r)| > |x(u_r)|+\delta$. Hence, the limit line $\ell_0$ of the $\ell_r$ for $r\to 0$ has the same description, which contradicts Lemma \[lemWhitney\] and the description of $\ell'_0$. \[cor1.8KR\]With the data and the notation of Proposition \[prop1.7KR\], let $x$ be in $\overline U$. Then there are only finitely many points in $\overline\Gamma$ which project to $x$ under the coordinate projection $K^{d}\times K^{n-d}\to K^d$. Suppose by contradiction that there are infinitely many such points. Then the dimension of $\overline\Gamma\cap (\{x\}\times K^{n-d})$ is $>0$. Thus, there exists $z\in \overline\Gamma$ such that $C_z^\Lambda( \overline\Gamma\cap (\{x\}\times K^{n-d}) )$ is of dimension $>0$, which is in contradiction to Proposition \[prop1.7KR\]. Deformation to the tangent cone ------------------------------- Let $X$ be a definable subset of $K^n$ and let $x$ be a point of $K^n$. Fix a subgroup $\Lambda$ in $\cD$. We consider the definable set $\cD (X, x, \Lambda)$ in $K^n \times \Lambda$ defined as $$\cD (X, x, \Lambda) := \Bigl\{ (z, \lambda) ; x + \lambda z \in X \Bigr\}$$ and its closure $$\overline{\cD (X, x, \Lambda)}$$ in $K^n \times K$. In $\overline{\cD (X, x, \Lambda)}$ one finds back the cone $C_x^{\Lambda} (X)$. Indeed, one has $ \overline{\cD (X, x, \Lambda)} \cap (K^n\times \{0\}) = C_x^{\Lambda} (X) \times \{0\}$, which we identify with $C_x^{\Lambda} (X)$. \[dim\] If $X$ is of dimension $d$, then $\cD (X, x, \Lambda)$ is of dimension $d+1$ and $C_x^{\Lambda} (X)$ is of dimension $\leq d$. Moreover, $\dim(C_x^\Lambda(X))$ does not depend on the choice of $\Lambda\in \cal D$. We may suppose that $X$ is nonempty. Consider the projection $$p : \begin{cases} \cD (X, x, \Lambda) \longrightarrow X \\ (z,\lambda)\longmapsto x+\lambda z. \end{cases}$$ Since $p$ is surjective and has fibers of dimension $1$, we get that $\cD (X, x, \Lambda)$ is of dimension $d+1$. The cone $C_x^{\Lambda} (X)$ is contained in $\{0\}\cup(\overline{\cD (X, x, \Lambda)}\setminus \cD (X, x, \Lambda))$. Hence, $C_x^{\Lambda} (X)$ is of dimension $\leq d$. The last statement follows from $$C_x^{\Lambda} (X) = \bigcup_i \mu_i C_x^{\Lambda'} (X),$$ whenever $\Lambda'\subset \Lambda$ is in $\cD$ and when one writes $\Lambda$ as a finite union of cosets $\bigcup_i \mu_i \Lambda'$ of $\Lambda'$ in $\Lambda$. Multiplicities on the tangent cones ----------------------------------- Let $X$ be a definable subset of $K^n$ of dimension $d$, let $x$ be a point of $K^n$, and let $\Lambda$ be in $\cD$. To each point $z$ on the cone $C_x^{\Lambda} (X)$ we will associate a rational number $SC_x^{\Lambda} (X)(z)$, called the multiplicity of $(X,x)$ at $z$ with respect to $\Lambda$. Define the function $$SC_x^{\Lambda} (X) : C_x^{\Lambda} (X) \to \QQ$$ as the function sending $z$ to $$[K^{\times} : \Lambda] \, \Theta_{d + 1}(\11_{\cD (X, x, \Lambda)})(z,0),$$ with $[K^{\times} : \Lambda]$ the index of $\Lambda$ in $K^\times$, and with $\11_{\cD (X, x, \Lambda)}$ the characteristic function of $\cD (X, x, \Lambda)$. The function $SC_x^{\Lambda} (X)$ is called the specialization of $X$ at $x$ with respect to $\Lambda$. The following lemma gives an indication that $SC_x^{\Lambda} (X)$ captures much local information of $(X,x)$; this principle will find a strong and precise form in Theorem \[mt\] below. \[support\] The function $SC_x^{\Lambda} (X)$ lies in $\cC( C_x^{\Lambda} (X))$. Moreover $$\Theta_{d + 1}(\11_{\cD (X, x, \Lambda)})(z,0)=0$$ for $z$ outside $C_x^{\Lambda} (X)$. The function $\11_{\cD (X, x, \Lambda)}$ is in $\cC(K^n \times K)$ since $\cD (X, x, \Lambda)$ is a definable set, and thus, also $\Theta_{d + 1}(\11_{\cD (X, x, \Lambda)})$ lies in $\cC(K^n \times K)$. For definable sets $A\subset B$, the restriction of a function in $\cC(B)$ to $A$ automatically lies in $\cC(A)$, hence, $SC_x^{\Lambda} (X)$ lies in $\cC( C_x^{\Lambda} (X))$. The second statement follows from the fact that the support of $\Theta_{d + 1}(\11_{\cD (X, x, \Lambda)})$ is contained in the closure of $\cD (X, x, \Lambda)$, which is contained in $(K^n\times K^\times) \cup (C_x^{\Lambda} (X) \times \{0\})$. More generally, if $\varphi$ is a function in $\cC (X)$ which is bounded near $x$, we define the specialization $\nu_x^{\Lambda} (\varphi)$ of $\varphi$ at $x$ with respect to $\Lambda$ in the following way. First define a function $\psi$ on $K^n \times K$ by $\psi (z, \lambda) := \varphi (x + \lambda z)$ on $\cD (X, x, \Lambda)$ and by zero elsewhere. Then one defines the function $$\nu_x^{\Lambda} (\varphi):C_x^{\Lambda} (X) \to \QQ$$ as the function sending $z$ to $[K^{\times} : \Lambda] \, \Theta_{d + 1}(\psi)(z,0)$. Note that, similarly as in Lemma \[support\], $\nu_x^{\Lambda} (\varphi)$ lies in $\cC (C_x^{\Lambda} (X))$ and that $\Theta_{d + 1}(\psi)(z,0)=0$ for $z$ outside $C_x^{\Lambda} (X)$. We recover $SC_x^{\Lambda} (X)$ since $\nu_x^{\Lambda} (\11_X)= SC_x^{\Lambda} (X) $. The following result, which will be proved in section \[secmt\], states that the local density can be computed on the tangent cone with multiplicities, for $\Lambda$ small enough. \[mt\] Let $X$ be a definable subset of $K^n$ and let $x$ be a point of $K^n$. For $\Lambda$ small enough $$\Theta_d (X) (x) = \Theta_d (SC_x^{\Lambda} (X)) (0).$$ More generally, let $\varphi$ be a function in $\cC (X)$ which is bounded near $x$. For $\Lambda$ small enough $$\Theta_d (\varphi) (x) = \Theta_d (\nu_x^{\Lambda} (\varphi)) (0).$$ Existence of $(w_f)$-regular stratifications {#swf} ============================================ In his study of stability of the topological type of mappings, R. Thom introduced the regularity condition $(a_f)$ in [@Thom], p. 274, as a relative version of condition $(a)$ of Whitney. The existence of $(a_f)$-regular stratifications was proved in the complex analytic case by H. Hironaka in [@Hi] (Corollary 1, Section 5) using resolution of singularities, under the assumption “sans éclatement" which is always satisfied for functions. One can find proofs of the existence of $(a_f)$ stratifications in the real subanalytic case in [@KR], where the Puiseux Theorem with parameters of Pawłucki (see [@Paw]) is used, and for o-minimal structures on the field of real numbers in [@Loi1]. The stronger condition $(w_f)$, the relative version of the so-called condition $(w)$ of Verdier (see [@Verdier]), was studied in the complex setting, for instance, in [@HeMeSa]. In the real subanalytic setting, it has been proved that $(w_f)$ stratifications exist by K. Bekka in [@bekka], K. Kurdyka and A. Parusiński in [@KurPar] using Puiseux Theorem with parameters, and finally by Ta Lê Loi in [@Loi2] for definable functions in some o-minimal structures over the real field (the o-minimal structure has to be polynomially bounded for the existence of $(w_f)$-regular stratifications, but need not to be so for the existence of $(a_f)$-regular stratifications). {#4.2} Let us now recall the definitions of $(w_f)$ and $(a_f)$-regular stratifications. Let $X$ be a definable subset of $K^n$, and let $(X^j)_{j\in \{1, \cdots, k\}}$ be a finite, definable and analytic stratification of $X$ satisfying the so-called frontier condition $$X^i\cap \overline{X^j}\not=\emptyset \Longrightarrow X^i\subset \overline{X^j},$$ where definable and analytic means that the strata $X^j$ are definable, $K$-analytic manifolds. Let $S$ be a definable subset of $K$ and let $f:X\to S$ be a definable continuous mapping such that for any $j\in \{1, \cdots, k\}$, $f_{\vert X^j}$ is analytic and of constant rank (being $0$ or $1$). For $j\in \{1, \cdots, k\}$ and $x\in X^j$, let us denote by $T_xX^j_f$ the tangent space at $x$ of the fiber $ f^{-1}_{\vert X^j}(f(x))$ of $f_{\vert X^j}$. Then one says that the pair of strata $(X^i,X^j)$ satisfies condition $(a_f)$ at a point $x_0\in X^i\subset \overline{X^j}$ if and only if for any sequence $(x_r)_{r\in \NN \setminus\{0\} }$ of points of $X^j$ converging to $x_0$ and such that the sequence $( T_{x_r}X^j_f )_{r\in \NN \setminus\{0\} }$ converges in the appropriate Grassmann manifold, one has $$\tag{$a_f$} \lim_{r\to \infty} \delta(T_{x_0}X^i_f,T_{x_r}X^j_f)= 0,$$ where $\delta(\cdot,\cdot) $ is a natural distance between linear subspaces of $K^n$ as defined below. Further, one says that the pair $(X^i,X^j)$ of strata satisfies condition $(w_f)$ at $x_0$ if and only if there exist a constant $C$ and a neighborhood ${\cal W}_{x_0}$ of $x_0$ in $K^n$, such that for any $x\in {\cal W}_{x_0}\cap X^i$ and any $y\in {\cal W}_{x_0}\cap X^j$, one has $$\tag{$w_f$} \delta(T_xX^i_f,T_{y}X^j_f)\le C\cdot \vert x-y\vert .$$ In both definitions, $\delta(V,V')$ denotes the distance between two linear subspaces $V$ and $V'$ of $K^n$ such that $\dim(V)\le \dim(V')$, and is defined by $$\delta(V,V')=\sup_{v\in V, \vert v\vert =1} \{\inf_{v'\in V', \vert v'\vert=1}\vert v-v' \vert \}= \sup_{v\in V, \vert v\vert =1}\hbox{dist}(v,S^{V'}(0,1)),$$ with $S^{V'}(0,1)$ the unit sphere around $0$ of $V'$. \[angle\] We have $\delta(V,V')=0$ if and only if $V\subset V'$ and for any $V''\subset V'$ such that $\dim(V)\le \dim(V'')$, $\delta(V,V'')\ge \delta(V,V')$. One says that the stratification $(X^j)_{j\in \{1, \cdots, k\}}$ is $(a_f)$-regular, respectively $(w_f)$-regular if any pair $(X^i,X^j)$ of strata is $(a_f)$-regular, respectively $(w_f)$-regular at any point of $X^i$. And finally one says that the stratification $(X^j)_{j\in \{1, \cdots, k\}}$ is $(a)$-regular, respectively $(w)$-regular, if it is $(a_f)$-regular, respectively $(w_f)$-regular, for $S$ a point in $K$. One starts the proof of the existence of $w_f$-regular stratifications with the key Lemma \[pli\] (see [@Loi2], Lemma 1.8 for its real version). But before stating this lemma, let us introduce as in [@CCL] (Definition 3.9) the notion of jacobian property for definable functions and recall from [@CCL] that this property is in a sense a generic one (see Proposition 3.10 of [@CCL] or Proposition \[jacprop\] below). This will be used in the proof of Lemma \[pli\]. Let $F:B\to B'$ a definable function with $B,B'\subset K$. We say that $F$ has the jacobian property if the following conditions hold all together: $\rm{(i)}$ $F$ is a bijection and $B,B'$ are balls, $\rm{(ii)}$ $F$ is $C^1$ on $B$, $\rm{(iii)}$ $\displaystyle \ord \Bigl(\frac{\partial F}{\partial x}\Bigr):B\to \ZZ$ is constant (and finite) on $B$, $\rm{(iv)}$ for all $x,y\in B$ with $x\not=y$, one has $$\ord\Bigl(\frac{\partial F}{\partial x}\Bigr)+\ord(x-y) =\ord(F(x)-F(y)).$$ It is proved in a much more general setting in [@CLip], Theorem 6.3.7, that the jacobian property is generic for definable mappings, which in our setting gives the following statement. \[jacprop\] Let $Y\subset K^m$ and $X\subset K\times Y$ be definable sets for some $m\in \NN$. Let $M:X\to K$ be definable. Then there exists a finite partition of $X$ into definable subsets $X_k$ such that for each $y\in Y$, the restriction $M(\cdot, y): x_1\mapsto M(x_1,y) $ of $M$ to $\{x_1\in K; (x_1,y)\in X_k\}$ is either injective or constant. Let us then assume, for simplicity, that on $X$, $M(\cdot,y)$ is injective. Then there exists a finite partition of $X$ into cells $A_k$ over $Y$ such that for each $y\in Y$ and each ball $B$ such that $B\times \{y\}$ is contained in $A_k$, there is a (unique) ball $B'$ such that the map $M_{\vert B}:B\to B': x_1\mapsto M(x_1,y)\in B'$ has the jacobian property. Now we state and prove the key lemma used in the proof of Theorem \[wf\]. \[pli\] Let $M:\Omega\to K$ be a definable and differentiable function on an open definable subset $\Omega$ of $K^m\times K$ for some $m\geq 0$. Assume that $\overline\Omega \cap (K^m\times \{0\})$ has a nonempty interior $U$ in $K^m$. Assume furthermore that $M$ is bounded on $\Omega$. Then there exist a nonempty open definable subset $V\subset U$ in $K^m$, an integer $\alpha>0$ and a constant $d\in K^\times$, such that for all $x\in V$ and all $t$ with $\ord t > \alpha$ and $(x,t)\in \Omega$ $$\Vert D_xM_{(x,t)}\Vert \le \vert d \vert\ .$$ In the above lemma and later on, $D_x M_{(x,t)}$ means $(\partial M(x,t)/\partial x_1,\ldots,\partial M(x,t)/\partial x_m )$, and analogously, $D_{x_1} M_{(x,t)}$ means $\partial M(x,t)/\partial x_1$ and so on. Let us denote $(x,t)=(x_1,\cdots, x_m,t)=(x_1,y)$ the standard coordinates on $K^m\times K=K\times K^m$, where $y=(x_2,\cdots, x_m,t)$. (We will apply cell decomposition and related results sometimes with $x_1$ and sometimes with $t$ as special variable.) By the Cell Decomposition Theorem (with special variable $t$) we can finitely partition $\Omega$ such that on each part $A$ such that $\overline A$ has nonzero intersection with $K^m\times \{0\}$ one has $$\vert D_{x_i}M_{(x,t)}\vert = \Vert D_{x}M_{(x,t)}\Vert = \vert c(x) \vert \cdot \vert \lambda t\vert^{a}$$ for some $a\in \QQ$, some $\lambda\in K^\times $, some $i\in\{1,\ldots,m\}$, and some definable function $c$. If all these exponents $a$ are nonnegative, then we are done since the $|c|$, as well as the boundary functions bounding $|t|$ from below in the cell descriptions are constant on small enough open subsets $V\subset U$. Let us assume that a particular $a$ is negative, say on a cell where $\vert D_{x_1}M_{(x,t)}\vert = \Vert D_{x}M_{(x,t)}\Vert$. By Proposition \[jacprop\] (with special variable $x_1$) applied to $ M : (x_1,y) \mapsto M(x_1,y) $ there exists a finite number of cells $A_k$ partitioning $\Omega$ such that for each $y\in K^m$ and each ball $B$ with $B\times \{y\}\subset A_k$, $B\ni x_1\mapsto M(x_1,y)$ has the jacobian property. We necessarily have one of these cells $A_k$ such that $\overline{A_k} \cap (K^m\times \{0\})$ has nonempty interior in $K^m$. We may assume by the Cell Decomposition Theorem (with special variable $t$) that $ A_k$ contains a subset $B_1\times B'\times W$ with $B_1$ an open ball in the $x_1$ line, $B'$ a Cartesian product of $m-1$ balls and $W$ an open definable subset of $K^\times$ such that $0\in \overline W$. Then, for any $y=(x_2,\cdots, x_m,t)\in B'\times W$, by the jacobian property, the one-dimensional volume $\mu_1(M(B_1\times \{y \}))$ equals $\mu_1(B_1)\cdot |c(x)|\cdot \vert t \vert^{a}$. Considering that $t$ can approach $0$ while $|c(x)|$ stays constant and that $M$ is a bounded mapping, this is a contradiction. We can finally prove our result concerning $(w_f)$-regular stratifications. \[wf\] Let $X$ be a definable subset of $K^n$, $S$ a definable subset of $K$ and $f:X\to S$ a definable continuous function. Then there exists a finite analytic definable stratification of $X$ which is $(w_f)$-regular. In particular definable subsets of $K^n$ also admit $(a_f)$, $(w)$ and $(a)$-regular definable stratifications. We proceed similarly as in [@Loi2]. Let $(X^j)_{j\in \{1, \cdots, k\}}$ be an analytic and definable stratification of $X$ such that the $f_{\vert X^j}$ are analytic and such that the rank of $f_{\vert X^j}$ is constant for all $j\in \{1, \cdots, k\}$. The set $w_f(X^i,X^j)$ of points $x\in X^i$ at which the pair $(X^i,X^j)$ is $(w_f)$-regular being a definable set, we have to show that this set is dense in $X^i$. Let us assume that the contrary holds, that is, the set $w'_f(X^i,X^j)$ of points of $X^i$ at which the pair $(X^i,X^j)$ is not $(w_f)$-regular is dense in (a nonempty open subset of) $X^i$, and let us obtain a contradiction. Up to replacing $X^i$ by a nonempty subset of $X^i$ and by the definability of $w'_f(X^i,X^j)$, we may suppose that $w'_f(X^i,X^j)$ equals $X^i$. As the condition $(w_f)$ is invariant under differentiable transformations of $K^n$ with Lipschitz continuous derivative and up to replacing $X^i$ by a nonempty open subset of $X^i$, we may assume that $X^i$ is an open definable subset of $K^m\times \{0\}^{n-m}$. (The latter transformation exists by Cell Decomposition, after shrinking $X^i$ if necessary.) Up to replacing $X^i$ by a nonempty subset, we may also assume that $f_{\vert X^i}$ is constant, equal to $0$ for simplicity. Indeed, since $w'_f(X^i,X^j)$ equals $X^i$, we can replace $X^i$ by $f_{\vert X^i}^{-1}(a)$ with $a\in f(X^i)$, since the pair $(f_{\vert X^i}^{-1}(a), X^j)$ is $(w_f)$-regular at none of the points of $f_{\vert X^i}^{-1}(a)$. Now we have two cases to consider: Case 1: $f_{\vert X^j}$ is constant (in a neighborhood of $X^i$) {#case-1-f_vert-xj-is-constant-in-a-neighborhood-of-xi .unnumbered} ---------------------------------------------------------------- Then condition $(w_f)$ is condition $(w)$. We proceed as follows. Write $X^i = U\times \{0\}^{n-m} $ with $U$ open in $K^m$. By the Cell Decomposition Theorem and the existence of definable choice functions and up to making $U$ smaller, there exists a definable $C^1$ function $\rho: U\times C \to X^j$ (called a $C^1$ wing in $X^j$ in [@Loi2]), where $C$ is a one dimensional cell in $K^\times$ with $0\in\overline C$, such that $\rho(x,t)=(x,r(x,t))$ and $\vert r(x,t)\vert < \vert t \vert$, and furthermore, $w'_f(X^i,X^j)$ being assumed equal to $X^i$, we may ask that for all $x,t$ $$\frac{\delta( K^m\times\{0\}^{n-m}, T_{\rho(x,t)}X^j)} {\vert r(x,t)\vert} \geq |t|^{-1}.$$ By Remark \[angle\], one then has $$\frac{\Vert D_x r_{(x,t)}\Vert }{\vert r(x,t)\vert } \ge \frac{\delta( K^m\times\{0\}^{n-m}, T_{\rho(x,t)}X^j)} {\vert r(x,t)\vert}.$$ By Cell Decomposition and up to replacing the function $(x,t)\mapsto r(x,t)$ by $(x,t)\mapsto r(x,t^s)$ for some integer $s>0$, we may moreover assume that on $U\times C$ $$|r(x,t)|=|a|\cdot|t|^\ell$$ for some integer $\ell>0$ and some $a\in K^\times$. But when one applies Lemma \[pli\] to $(x,t)\mapsto r(x,t)/t^\ell$, which is a bounded map, one finds a definable nonempty open subset $U'$ of $U$, and $d \in K^\times$ such that for $x\in U'$ and $t\in C$ with $\vert t \vert $ small enough, $\Vert D_xr_{(x,t)} \Vert / |r(x,t)| \le \vert d \vert$, a contradiction with the above two displayed inequalities. Case 2: $f_{\vert X^j}$ has rank 1 {#case-2-f_vert-xj-has-rank-1 .unnumbered} ----------------------------------- Write $X^i = U\times \{0\}^{n-m} $ with $U$ open in $K^m$. Clearly we may suppose that $f(x,y)\not =0$ for $(x,y)\in X^j$, $x\in U$. We further have that for each $x\in U$, $f(x,y)$ goes to $0$ when $y$ goes to zero with $(x,y)\in X^j$. Hence there exists a definable choice function $f_0:B(0,1)\to f(X^j)\cup\{0\}$ such that $f_0(t)=0$ if and only if $t=0$ and $|f_0(t)|<|t|$ for nonzero $t$. Since we assume that $w'_f(X^i,X^j)$ equals $X^i$, we may moreover assume that for each $x\in U$ and nonzero $t$ there exists $y$ satisfying $(x,y)\in X^j$, $|y|<|t|$, $f(x,y)=f_0(t)$, and $$\frac{\delta(K^m\times \{0\}^{n-m},T_{(x,y)} X^j_f)}{\vert y \vert } \geq |t|^{-1}.$$ Up to replacing $f_0$ by $t\mapsto f_0(\lambda t^s)$ for some integer $s>0$ and nonzero $\lambda\in R$, we may suppose that $f_0$ is continuous. Hence, by the existence of definable choice functions there exists a continuous definable map $\varphi : U\times B(0,1) \to K^{n-m}$ which is $C^1$ on $U\times (B(0,1)\setminus\{0\})$ and such that, for all $x\in U$ and for all nonzero $t\in B(0,1)$, $\varphi(x,0)=0$, $(x,\varphi(x,t))\in X^j$, $$\label{1} \frac{\delta(K^m\times \{0\}^{n-m},T_{(x,\varphi(x,t))} X^j_f)}{\vert \varphi(x,t) \vert } \geq |t|^{-1}$$ and $$\label{111} f(x, \varphi(x,t) )= f_0(t).$$ It follows by (\[111\]) that the $m$-dimensional linear space $W$ spanned by the vectors $(0,\ldots,0,1,0,\ldots,0,\partial \varphi(x,t)/\partial x_i)$ for $i=1,\ldots,m$ is a subspace of $T_{(x,\varphi(x,t))}X^j_f$. Combining this with Remark \[angle\] and since $\Vert D_x\varphi_{(x,t)}\Vert \geq \delta(K^m\times \{0\}^{n-m} , W)$, it follows that $$\label{1bis} \frac{\Vert D_x\varphi_{(x,t)}\Vert}{\vert \varphi(x,t) \vert }\ge \frac{\delta(K^m\times \{0\}^{n-m},T_{(x,\varphi(x,t))} X^j_f)}{\vert \varphi(x,t) \vert }.$$ By the Cell Decomposition Theorem \[thm:CellDecomp\], by making $U$ smaller, and up to replacing the function $(x,t)\mapsto \varphi(x,t)$ by $(x,t)\mapsto \varphi(x,bt^s)$ for some integer $s>0$ and some nonzero $b\in R$, we may suppose we have on $U\times B(0,1)$ $$\label{2a} \vert \varphi(x, t) \vert= \vert a \vert \cdot \vert t\vert^\ell,$$ with $a\in K^\times$ and some integer $\ell>0$, since $\varphi$ is continuous and $\varphi(x,t)=0$ if and only if $t=0$. Applying Lemma \[pli\] to the bounded function $\varphi(x,t)/t^\ell$ yields a contradiction with $(\ref{1})$ and $(\ref{1bis})$ similarly as in case 1. {#section-4} Let $X$ be a definable subset of $K^n$, and let $(X^j)_{j\in \{1, \cdots, k\}}$ be a finite, definable and analytic stratification of $X$ satisfying the frontier condition as in \[4.2\]. Let $X^i$ and $X^j$ be strata with $X^i \subset \overline{X^j}$ and let $x_0 \in X^i$. One says $(X^i, X^j)$ satisfies condition $(b)$ at $x_0$ if for every sequences $x_m \in X^i$, $y_m \in X^j$, both converging to $x_0$ and such that the line $L_m$ containing $x_m$ and $y_m$, resp. the tangent space $T_{y_m} X^j$, both converge in the appropriate Grassmann manifold to a line $L$, resp. a subspace $T$, then $L \subset T$. Over the reals, it is well known since the seminal work of T. C. Kuo [@Kuo] (in the semi-analytic case), see also [@Verdier] (subanalytic case) and [@Loi2] (o-minimal case), that condition $(w)$ implies condition $(b)$. Note that obviously $(w)$ does not imply $(b)$ in the real differential case and that even in the real algebraic case $(b)$ does not imply $(w)$. In the present setting, we have a similar result (with a similar proof): If $(X^i, X^j)$ satisfies condition $(w)$ at $x_0$, it also satisfies condition $(b)$ at $x_0$. We set $X^i = W$ and $X^j = W'$. We may assume that $W$ is open in $K^r \simeq K^r \times \{0\} \subset K^r \times K^s = K^n$ and that $x_0 = 0$. Denote by $p$ the linear projection $K^n \to K^s$. If condition $(b)$ is not satisfied at $0$, then, by condition $(w)$ at $0$ and for some $\varepsilon >0$, one has $0 \in \overline{S} \setminus S$, with $$S = \{x \in W' \, \vert \, \delta (K p (x), T_x W') \geq 2 \varepsilon\}.$$ Use the curve selection lemma \[curve\] to find an analytic definable function $\varphi : U \subset K \to S$ with $0 \in \overline{U}$ such that $\Vert \varphi (t)\Vert \leq \vert t \vert$ for all $t$ in $U$. Write $\varphi = (a, b)$ with $a : U \to K^r$ and $b : U \to K^s$. We may assume that $\Vert a' (t) \Vert$ is bounded, that $b$ and $b'$ do not vanish, and by analyticity that $\lim_{t \to 0} \Vert b (t) \Vert / \Vert b' (t) \Vert = 0$. Since $\delta (K b'(t), K b(t)) \to 0$ for $t \to 0$ which holds by Lemma \[lemWhitney\], we have $\delta (K b'(t), T_{\varphi (t)} W') \geq \varepsilon$ for $t$ small enough. From the fact that $\varphi' (t) = a' (t) + b' (t) \in T_{\varphi (t)} W'$, it follows that $$\label{one} \frac{\Vert a' (t)\Vert}{\Vert b'(t)\Vert} \delta (K a'(t), T_{\varphi (t)} W') \geq \varepsilon.$$ Now, by condition $(w)$ at $0$, there exists $C >0$ such that $$\label{two} \delta (K a'(t), T_{\varphi (t)} W') \leq C \Vert b(t) \Vert$$ for $t$ small enough. It follows from (\[one\]) and (\[two\]), that, for $t$ small enough, $$\varepsilon \leq C \frac{\Vert b(t)\Vert}{\Vert b'(t)\Vert} \Vert a' (t) \Vert,$$ which contradicts the fact that $\Vert a' (t) \Vert$ is bounded and $\lim_{t \to 0} \Vert b (t) \Vert / \Vert b' (t) \Vert = 0$. Proof of Theorem \[mt\] and existence of distinguished tangent $\Lambda$-cones {#secmt} ============================================================================== Proof of Theorem \[mt\]: a first reduction ------------------------------------------ The statement we have to prove being additive, we may cut $X$ into finitely many definable pieces. Also note that we may assume all these pieces have dimension $d$ around $x$, since pieces of dimension $<d$ contribute to zero in both sides in the equality we have to prove. Let us prove in this subsection that we may reduce to the case were $\varphi = \11_X$. Suppose that we know the result for $\varphi = \11_X$. For the general case we may assume, by additivity and linearity, that $\varphi = (\prod_{i=1}^\ell \beta_i) \cdot q^{- \alpha}$ with $\alpha$ and the $\beta_i $ definable functions from $X$ to $\ZZ$. Further we may assume that $\varphi\geq 0$ on $X$. Write $X$ as a possibly infinite disjoint union parameterized by the values of $\alpha$ and the $\beta_i$. That is $$X=\cup_{z\in\ZZ^{\ell+1}}X_z, \mbox{ with } X_z = \{x\in X\mid (\beta_1,\ldots,\beta_\ell,\alpha)(x) = z\}.$$ Since $\varphi$ is constant on each of the $X_z$, by linearity we find for each $z$ $$\Theta_d (\varphi_z) (x) = \Theta_d (\nu_x^{\Lambda} (\varphi_z)) (0),$$ where $\varphi_z$ is the product of $\varphi$ with the characteristic function of $X_z$. By Proposition \[convmon\] one finds $$\Theta_d (\varphi) (x) = \Theta_d (\sum_z \varphi_z) (x) = \sum_z \Theta_d (\varphi_z) (x)$$ and similarly $$\Theta_d (\nu_x^{\Lambda} (\varphi)) (0) = \Theta_d (\nu_x^{\Lambda} ( \sum_z \varphi_z)) (0) = \Theta_d (\sum_z \nu_x^{\Lambda} ( \varphi_z)) (0) = \sum_z \Theta_d (\nu_x^{\Lambda} ( \varphi_z)) (0),$$ and hence $\Theta_d (\varphi) (x) = \Theta_d (\nu_x^{\Lambda} (\varphi)) (0)$ which finishes the reduction. Proof of Theorem \[mt\]: the case $d = n$ {#opencase} ----------------------------------------- In this subsection, we consider the case $d = n$. It is not difficult to see (cf. Corollary \[trivcorSC\] below) that the function $SC_x^{\Lambda} (X)$ is equal to the characteristic function of $C_x^{\Lambda} (X)$ almost everywhere. Hence it is enough to prove that $$\Theta_d (X) (x) = \Theta_d (C_x^{\Lambda} (X)) (0)$$ for $\Lambda$ small enough. The proof we shall give is quite analogous to the one of Proposition 2.1 in [@KR]. By Corollary \[localconic\], there exists a definable function $\alpha : \PP^{n - 1} (K) \rightarrow \NN$ and a subgroup $\Lambda$ in $\cD$, such that for every $\ell$ in $\PP^{n - 1} (K)$, $(\pi_x^X)^{-1} (\ell) \cap B (x, \alpha (\ell))$ is a local $\Lambda$-cone with origin $x$ in $(\pi_x^X)^{-1} (\ell)$. For every $n \geq 0$, we consider the $\Lambda$-cone $C_n (\Lambda)$ with origin $x$ generated by $X \cap B (x, n)$. Note that $$C_x^{\Lambda} (X) = \cap_n \overline{C_n (\Lambda)},$$ hence, if we set $$W := \cap_n C_n (\Lambda),$$ we have $W \subset C_x^{\Lambda} (X)$. In particular, $\Theta_d (W) (x) \leq \Theta_d (C_x^{\Lambda} (X)) (0)$. By Proposition \[convmon\], we know that $\Theta_d (W) (x) = \lim_n \Theta_d (C_n (\Lambda)) (x)$ and $\Theta_d (C_x^{\Lambda} (X)) (0) = \lim_n \Theta_d (\overline{C_n (\Lambda)}) (x)$. By Proposition \[bord\], we deduce that $$\Theta_d (W) (x) = \Theta_d (C_x^{\Lambda} (X)) (0).$$ Since we have $$\Theta_d (X) (x) = \Theta_d (X \cap B (x, n) ) (x) \leq \Theta_d (C_n (\Lambda)) (x),$$ we deduce that $$\Theta_d (X) (x) \leq \Theta_d (W) (x).$$ To prove the reverse inequality, let us set consider for $n \geq 0$ the definable subset $W_n$ of all points $w$ in $W$ such that $\alpha (\pi_x (w)) \leq n$. By definition $W_n \cap B (x, n) \subset X \cap B (x, n)$, hence $$\Theta_d (W_n) (x) = \Theta_d (W_n \cap B (x, n)) (x) \leq \Theta_d (X \cap B (x, n)) (x) = \Theta_d (X) (x).$$ Since, by Proposition \[convmon\] again, $\lim_n \Theta_d (W_n) (x) = \Theta_d (W) (x)$, we get $\Theta_d (W) (x) \leq \Theta_d (X) (x)$, as required. Graphs {#gr} ------ The main technical result in this subsection is Proposition \[3.6\], which will be used in subsection \[ept\] to conclude the proof of Theorem \[mt\] and in subsection \[edt\] to prove the existence of distinguished tangent $\Lambda$-cones. Fix two integers $0<d \leq m$. Let $U$ be an open definable subset of $K^d$ and let $\varphi$ be a definable mapping $U \rightarrow K^{m - d}$. The graph $\Gamma = \Gamma (\varphi)$ of $\varphi$ is a definable subset of $K^m$. Fix a point $u$ in the closure $\overline U $ of $U$ and $\Lambda$ adapted to $(U, u)$. We assume that $\displaystyle\lim_{x \to u}\varphi (x) = v$, by Corollary \[cor1.8KR\]. We set $w := (u, v)$. The projection to the first $d$ coordinates $K^m \rightarrow K^d$ induces a function $$\vartheta : \cD (\Gamma, w, \Lambda) \longrightarrow \cD (U, u, \Lambda).$$ Note that $\vartheta$ is an isomorphism with inverse given by $$\label{theta-1} \vartheta^{- 1}: (z, \lambda) \longmapsto (z, \lambda^{-1} (\varphi (u + \lambda z) - v), \lambda).$$ By Corollary \[localconic\], there exists a definable function $\alpha : \PP^{d - 1} (K) \rightarrow \NN\cup\{\infty\}$ such that for every $\ell$ in $\PP^{d - 1} (K)$, $(\pi_u^U)^{-1} (\ell) \cap B (u, \alpha (\ell))$ is a local $\Lambda$-cone with origin $u$ in $(\pi_u^U)^{-1} (\ell)$, with the convention that $\alpha(\ell)=\infty$ if and only if $\ell$ is such that $(\pi_u^U)^{-1} (\ell) \cap U=\emptyset$. Note that being definable, the function $\alpha$ is continuous on a dense definable open subset $\Omega_0$ in $\PP^{d - 1} (K)$, where dense means that the complement of $\Omega_0$ has strictly smaller dimension. Let $\Omega_1$ be the definable subset of $\Omega_0$ consisting of the $\ell$ such that for all neighborhoods $ \cV$ of $u$ in $K^d$, the sets $(\pi_u^U)^{-1} (\ell)\cap \cV$ are nonempty. \[nonempty\] Suppose that $C_u^{\Lambda} (U)$ is of maximal dimension. Then $\Omega_1$ contains a nonempty open subset of $\PP^{d - 1} (K)$. Let $\Omega_1^c$ be the complement of $\Omega_1$ in $\PP^{d - 1} (K)$. Clearly $\Omega_1^c$ is definable. It is enough to derive a contradiction out of the assumption that $\Omega_1^c$ is dense. Suppose thus that $\Omega_1^c$ is dense in $\PP^{d - 1} (K)$. By the definability and density of $\Omega_1^c$ and of $\Omega_0$ in $\PP^{d - 1} (K)$, it follows that $\Omega_1^c\cap \Omega_0$ is dense. Take $\ell$ in $\Omega_1^c\cap \Omega_0$. By the definition of the tangent cone, one has that $(\pi_u^U)^{-1} (\ell) \cap C_u^{\Lambda} (U)=\emptyset$. Since $(\pi_u^U)^{-1} ( \Omega_1^c\cap \Omega_0 )$ is dense and definable in $K^d$, it follows that $C_u^{\Lambda} (U)$ is contained in a definable subset of dimension $<d$, a contradiction with $C_u^{\Lambda} (U)$ being of maximal dimension, that is, of dimension $d$. For any definable subset $O$ of $\PP^{d - 1} (K)$, consider the definable subset $$\label{CO} C_u^{\Lambda, O} (U) := (\pi_u^U)^{- 1} (O) \cap C_u^{\Lambda} (U)$$ of $C_u^{\Lambda} (U)$. \[nonempty2\] Suppose that $C_u^{\Lambda} (U)$ is of maximal dimension. Let $O$ be dense open in $\Omega_1$. Then the set $C_u^{\Lambda, O} (U)$ also has maximal dimension and is dense open in $C_u^{\Lambda} (U)$. The fact that $C_u^{\Lambda, O} (U)$ is open in $C_u^{\Lambda} (U)$ follows from general topology. We only have to prove that $C_u^{\Lambda, O} (U)$ is dense in $C_u^{\Lambda} (U)$. As we have noticed in the proof of Lemma \[nonempty\], for every $\ell$ in $\Omega_1^c\cap \Omega_0$, $$(\pi_u^U)^{-1} (\ell) \cap C_u^{\Lambda} (U)=\emptyset.$$ Since moreover the sets $(\pi_u^U)^{-1} (\Omega_1 \setminus O)$ and $(\pi_u^U)^{-1} (\Omega_1^c \setminus \Omega_0)$ have dimension $<d$, the lemma follows. The next lemma ensures in particular that there definable sets $\Omega=O$ as in Lemma \[nonempty2\] such that moreover for all $z$ in $C_u^{\Lambda, \Omega}(U)$ and all small enough $\lambda$ in $\Lambda$ one has that $u + \lambda z$ lies in $U$. \[smallU\] Suppose that $C_u^{\Lambda} (U)$ is of maximal dimension. Then there is a dense open definable subset $\Omega$ of $\Omega_1$ such that for all $z$ in $C_u^{\Lambda, \Omega} (U)$ of direction $\ell$ and for all small enough $\lambda$ in $\Lambda$ one has $u+\lambda z\in \overline{(\pi_u^U)^{-1}(\ell) }$. Here, small enough can be taken to mean that $\ord (\lambda z)\geq \alpha(\ell)$, where $\alpha$ is as in the beginning of section \[gr\]. We assume $u=0$ for simplicity. For any $\ell\in \Omega$, any $z\in \ell \cap C_u^{\Lambda}(U)$ and any $\lambda\in \Lambda$, one has $\lambda z\in C_u^{\Lambda}(U)$. Hence what remains to be proved is a consequence of the inclusion $\subset$ in the equality of the following claim. For almost all $\ell$ in $\Omega_1$ and with $u=0$, one has the following equality of local $\Lambda$-cones $$C_u^{\Lambda} (U)\cap \ell \cap B (0, \alpha (\ell)) = (\overline{\pi_u^{U})^{-1} (\ell)} \cap B (0, \alpha (\ell)).$$ Since $\ell\in \Omega_1$, $ (\pi_u^U)^{-1}(\ell)\cap B(0,\alpha(\ell))$ is a local $\Lambda$-cone with origin $0$, and by Remark \[remarque sur les cones locaux\], we have $$\overline{(\pi_u^U)^{-1}(\ell)} \cap B(0,\alpha(\ell)) =C_u^\Lambda( (\pi_u^U)^{-1}(\ell)) \cap B(0,\alpha(\ell))$$ $$\subset C_u^{\Lambda} (U)\cap \ell \cap B (0, \alpha (\ell)) .$$ The inclusion $\supset$ is thus clear for all $\ell\in \Omega$. We prove the inclusion $\subset$ in the claim, for almost all $\ell$. This follows from cell decomposition. Let $X\subset (K^d\setminus\{0\})\times K$ be the definable set $$\{(x,t)\in (K^d\setminus\{0\})\times K\mid u+ t\cdot x\in U\},$$ parametrizing all $\ell\cap U$ for all lines $\ell$ through $u$. Then $X$ is a finite union of cells by Theorem \[thm:CellDecomp\]. For each $x\in K^d\setminus\{0\}$ write $X_x$ for the fiber above $x$ under the projection $X\to K^d$. For each $x$, either $0$ lies in the interior of $X_x$, either $0$ lies in the boundary $\partial X_x$ of $X_x$ or $0$ lies outside the closure of $X_x$, where $\partial X_x$ is the closure of $X$ minus the interior of $X$. In the case that $0$ lies in the interior of $X_x$, one has that $(\pi_u^U)^{-1} (\ell) \cap B (u, \alpha (\ell))= B (u, \alpha (\ell))$ hence the inclusion $\subset$ is evident. The inclusion $\subset$ holds, up to a set of direction $\ell\in \PP^{d-1}(K)$ of dimension $<d-1$, for all those $x$ such that $0$ lies in $\partial X_x$ by the almost everywhere continuity of the functions in $x$ appearing in the descriptions of the cells having $0$ in their boundary. The case that $0$ lies outside the closure of $X_x$ needs not to be considered since we suppose $\ell\in\Omega_1$. \[distinguished cone for open sets\] Let $d>0$, let $U$ be a definable nonempty open subset of $K^d$ and let $\Lambda$ given by Corollary \[localconic\]. Then for all $u\in \overline U$, with $C_u^\Lambda(U)$ of maximal dimension $d$, $C_u^{\Lambda}(U)$ is a distinguished $\Lambda$-tangent cone at $u$ for $U$, that is, for all $\Lambda'\in \cD$, $\Lambda'\subset \Lambda$ implies $C_u^{\Lambda'}(U)=C_u^{\Lambda}(U)$. As usual we assume $u=0$. Let $\Lambda$ as given by Corollary \[localconic\] and $\Lambda'\in \cD$, $\Lambda'\subset \Lambda$. We show that $C_u^{\Lambda}(U)\subset C_u^{\Lambda'}(U)$. Let $z\in C_u^{\Lambda}(U)$, denote by $\ell$ its direction and assume that $\ell \in \Omega$, with $\Omega\subset \Omega_1$ as in Lemma \[smallU\]. By Lemma \[smallU\] we then have $z\in C_u^\Lambda(\overline{U\cap \ell})= C_u^\Lambda(U\cap \ell )$. But since $U\cap \ell\cap B(0,\alpha(\ell))$ is a local $\Lambda$-cone with origin $u$, by Remark \[remarque sur les cones locaux\], we get $$z \in C_u^{\Lambda}(U\cap \ell) =C_u^{\Lambda'}(U\cap \ell)\subset C_u^{\Lambda'}(U).$$ Now since we showed $C_u^{\Lambda,\Omega}(U)=C_u^{\Lambda',\Omega}(U)$, we have $\overline{C_u^{\Lambda,\Omega}(U)}=\overline{C_u^{\Lambda',\Omega}(U)}$. But by Lemma \[nonempty2\] we obtain $\overline{C_u^{\Lambda,\Omega}(U)} =C_u^{\Lambda}(U)$. We finally remark that one also has $\overline{C_u^{\Lambda',\Omega}(U)} =C_u^{\Lambda'}(U)$, with the same proof as in Lemmas \[nonempty\] and \[nonempty2\], since any adapted (to $U$) $\Lambda'$-cone may be chosen in those lemmas. Proposition \[3.6\] below can be seen as an analogue of Proposition 3.6 of [@KR] and has for consequence the existence of distinguished $\Lambda$-tangent cones for general definable sets and the $p$-adic analogue of Thie’s formula. As usual, the main point in the $p$-adic case is to overcome the lack of connectedness and deal with all its negative consequences such as the lack of strong enough mean value theorems and so forth. To go through these difficulties we essentially use the following result, the main result of [@CCL], which is the $p$-adic analogue of the existence of the so-called $L$-decompositions of real subanalytic sets, obtained in [@KurWExp], and which will be used in the proof of Proposition \[3.6\] (c). \[Lipschitz decomposition\] Let $\varepsilon>0$ and let $\varphi:X\subset K^n\to K^m$ be a locally $\varepsilon$-Lipschitz definable mapping. Then there exist $C>0$ and a finite definable partition of $X$ into parts $X_1, \cdots, X_k$, such that the restriction of $\varphi$ to each $X_i$ is globally $C$-Lipschitz. We will also use Theorem \[wf\] in the proof of Proposition \[3.6\] in the same way that Lemma 3.7 of [@KR] is used in the proof of Proposition 3.6 of [@KR]. Where Theorem \[wf\] gives the existence of $(w_f)$-regular (and consequently $(a_f)$, $(b)$, and $(w)$-regular) stratifications for a function $f$ in the definable $p$-adic setting, we will only use the genericity of the condition $(a_f)$ in the $p$-adic definable case to prove Proposition \[3.6\]. \[3.6\] Let $\varepsilon$ be a positive real number with $\varepsilon\leq 1$. Let $U$ be an open definable subset of $K^d$ and let $\varphi$ be a definable mapping $U \rightarrow K^{m - d}$. Fix a point $u$ in $\overline U$, a subgroup $\Lambda$ adapted to $(U, u)$, choose $\Omega$ sufficiently small and as in Lemma \[smallU\], and let $C_u^{\Lambda, \Omega} (U)$ be as in \[CO\]. Assume that $\varphi$ is $\varepsilon$-analytic and that $\displaystyle\lim_{x \to u}\varphi (x) = v$ by Corollary \[cor1.8KR\]. Suppose that $C_u^{\Lambda} (U)$ has maximal dimension. Then, possibly after partitioning $U$ into finitely many open subsets, replacing $U$ successively by each one of these smaller open subsets, in such a way that $\varphi$ is globally $C$-Lipschitz on $U$ by Theorem \[Lipschitz decomposition\] and neglecting those $U$ such that $C_u^{\Lambda} (U)$ has lower dimension, the following hold 1. For $z$ in $C_u^{\Lambda, \Omega} (U)$ such that $u+\lambda z\in U$, for all small enough $\lambda\in \Lambda$ see Lemma \[smallU\], the limit $$\psi (z) := \lim_{{\genfrac{}{}{0pt}{}{\lambda \to 0}{\lambda \in \Lambda}}} \lambda^{-1} (\varphi (u + \lambda z) - v)$$ exists in $K^{m-d}$, yielding a definable function $\psi:C_u^{\Lambda, \Omega} (U)\to K^{m-d}$. 2. The function $\psi$ is locally $\varepsilon$-Lipschitz. 3. The graph of $\psi$ is dense in $C_w^{\Lambda}(\Gamma(\varphi))$. We first prove (a). Choose $z$ in $C_u^{\Lambda, \Omega} (U)$ such that $u+\lambda z\in U$, for all small enough $\lambda\in \Lambda$. We can evaluate $\varphi$ at $u+\lambda z$ for small enough $\lambda$ in $\Lambda$. After partitioning $U$ into finitely many open subsets and successively replacing $U$ by each one of these smaller open subsets, Lemma \[terms\] implies that when $\lambda \to 0$, $\lambda \in \Lambda$, either $ \lambda^{-1} (\varphi (u + \lambda z) - v)$ has a limit $\psi (z)$ or its norm goes to $\infty$. Applying the Curve Selection Lemma \[curve\] to the point $(u,v)$ and the set $$\{(u+\lambda z,\varphi (u + \lambda z) )\in K^m\mid \lambda\in \Lambda\},$$ it follows from Lemma \[lemWhitney\] and $\varepsilon$-analyticity that the limit $\psi (z)$ exists. Now assume that $z\in C_u^{\Lambda, \Omega} (U)$ is such that $u+\lambda z\in \overline{U\cap \ell}$, where $\ell$ is the (direction of the) line going through $u$ and $z$. By Lemma \[smallU\], for all $\varepsilon>0 $ there exist $z',z''\in C_u^{\Lambda, \Omega} (U)$ such that $u+\lambda z'\in U$, $u+\lambda z''\in U$ for all $\lambda\in \Lambda$ and $\vert z-z'\vert \le \varepsilon$ and $\vert z-z''\vert \le \varepsilon$. Then we have $$\vert \lambda^{-1} [ (\varphi(u+\lambda z')-v)- (\varphi(u+\lambda z'')-v) ]\vert \le C\vert \lambda^{-1} \lambda(z'-z'')\vert \le C\cdot \varepsilon.$$ This shows that one can define $\psi$ on $C_u^{\Lambda,\Omega}$ by $$\label{psizz'} \psi(z)=\displaystyle \lim_{z'\to z}\lim_{\lambda\to 0}\lambda^{-1} (\varphi(u+\lambda z')-v).$$ Let us now prove (b). We first notice that, after suitable finite partition of $U$ and neglecting those $U$ such that $C_u^{\Lambda} (U)$ has lower dimension, we may suppose that the function $\psi$ is analytic on $C_u^{\Lambda, \Omega} (U)$. To prove that $\psi$ is $\varepsilon$-analytic on $C_u^{\Lambda, \Omega} (U)$, we show that the tangent space $T_x\Gamma(\psi)$ at a point $x$ of the graph $\Gamma(\psi)$ of $\psi$, for $x$ in a dense set of $\Gamma(\psi)$, is contained in $C_\varepsilon=\{(a,b)\in K^d\times K^{m-d}; \vert b\vert \le \varepsilon \vert a\vert\}$. Since, by (\[psizz’\]), $\Gamma(\psi)\subset C_w^{\Lambda}(\Gamma(\varphi))$ and since $\dim(\Gamma(\psi))=\dim(C_w^{\Lambda}(\Gamma(\varphi)))$, it is enough to prove that at a generic point $x$ of $C_w^{\Lambda}(\Gamma(\varphi))$ one has $T_xC_w^{\Lambda}(\Gamma(\varphi))\subset C_\varepsilon$. For this we consider the deformation $h:\overline{{\cD}(\Gamma(\varphi),w, \Lambda)}\to K$ to $C_w^{\Lambda}(\Gamma(\varphi))$ defined in section $3.5$. The fiber $h^{-1}(0)$ is identified with $C_w^\Lambda(\Gamma(\varphi))$ and for $\lambda\in \Lambda$ $$h^{-1}(\lambda)=\{(z,\lambda)\in K^m\times \Lambda; w+\lambda z\in \Gamma(\varphi)\}$$ is identified with $$\{z\in K^m; w+\lambda z\in \Gamma(\varphi)\}.$$ Since $\varphi$ is $\varepsilon$-analytic, for any $\lambda\in \Lambda$ and any $y\in h^{-1}(\lambda)$, one has $T_y h^{-1}(\lambda)\subset C_\varepsilon$. Let us show at $x$ a generic point of $C_w^\Lambda(\Gamma(\varphi))$, $T_xC_w^\Lambda(\Gamma(\varphi))$ is a limit of tangents $T_{y_n} h^{-1}(\lambda_n)$. But this is exactly the genericity in $h^{-1}(0)$ of the condition $(a_h)$, which is given by Theorem \[wf\].\ We now prove (c). Let $z\in C_w^\Lambda(\Gamma(\varphi))$ and $(\lambda_n)_{n\in \NN}\in \Lambda, (w_n)_{n\in \NN}\in \Gamma(\varphi)$ be two sequences such that $w_n\to w$ and $\lambda_n(w_n-w)\to z$. Denoting by $\pi$ the projection from $\Gamma$ to $U$ and $u_n=\pi(w_n)$, the sequence $(u_n)_{n\in \NN}$ of points of $U$ going to $u$ is such that $\displaystyle\lim_{n\to \infty}\lambda_n(u_n-u)= \pi(z):=a\in C^\Lambda_u(U)$. Now fix $\varepsilon>0$ and $a'\in C_u^{\Lambda,\Omega}(U)$ with $\vert a-a' \vert\le \varepsilon $. Then $u+\lambda a'\in U $ for all small enough $\lambda\in \Lambda$ by Lemma \[smallU\]. Then we may suppose, by invoking Theorem \[Lipschitz decomposition\], that $$\vert \lambda_n(\varphi(u_n)-v)-\lambda_n( \varphi(\lambda_n^{-1}a'+u)-v) \vert \le C \vert \lambda_n (u_n-u)-a')\vert.$$ This gives $$\lim_{n \to \infty} \vert \lambda_n(w_n-w)- \lambda_n(\lambda_n^{-1}a',\varphi(\lambda_n^{-1}a'+u)) \vert \le \max(C,1)\cdot \varepsilon,$$ and, finally $$\vert z-(a',\psi(a')) \vert \le \max(C,1)\cdot \varepsilon,$$ showing that the graph of $\psi $ is dense in $C_w^\Lambda(\Gamma(\varphi))$. \[trivcorSC\] Under the hypotheses and with the notation of Proposition \[3.6\], assume moreover that $\varepsilon < 1$. Write $z$ for variables running over $K^{d}$ and $y$ for variables running over $K^{m-d}$. Then, for almost all $(z,y)\in C_w^{\Lambda} (\Gamma(\varphi))$, one has that $SC_w^{\Lambda} (\Gamma(\varphi))(z,y) = 1 = SC_u^{\Lambda} (U)(z) $, and $$\Theta_d (SC_w^{\Lambda} (\Gamma(\varphi))) (0) = \Theta_d (C_w^{\Lambda} (\Gamma(\varphi))) (0) = \Theta_d (U) (u)= \Theta_d (\Gamma(\varphi)) (w).$$ We first prove that, for almost all $z\in C_u^{\Lambda} (U)$, one has that $1 = SC_u^{\Lambda} (U)(z) $. Let $\Omega$ and $\alpha$ be as in Lemma \[smallU\]. By Lemma \[smallU\], for $z\in C_u^{\Lambda, \Omega} (U)$, there exist an open ball $B$ contained in $C_u^{\Lambda, \Omega} (U)$ and containing $z$, and a ball $B_1\subset K$ around $0$ such that $$\cD(U,u,\Lambda)\cap (B\times B_1 ) = B\times ( \Lambda\cap B_1 ).$$ Hence we can calculate $$SC_u^\Lambda (U)(z) = [K:\Lambda]\Theta_{d+1}(\cD (U,u,\Lambda))(z,0)= [K:\Lambda]\Theta_{d+1}(B \times \Lambda )(z,0)=\Theta_{d}(B )(z)$$ which equals $1$ since $z\in B$. Next we prove that $SC_w^{\Lambda} (\Gamma(\varphi))(z,y)=1$ for almost all $(z,y)\in C_w^{\Lambda} (\Gamma(\varphi))$. For this purpose, define $$\cD' := \{(z,y,\lambda)\in \cD(\Gamma(\varphi),w,\Lambda);\ |y| < |z| \},$$ and consider the natural projection $$p: \cD' \to \cD (U,u,\Lambda): (z,y,\lambda)\mapsto (z,\lambda),$$ which is in fact injective. Write $U'$ for the image of $p$. By Proposition \[prop1.7KR\] and Lemma \[curve\], one finds for all $(z,y)\in K^m$ that $$\Theta_{d+1}( \cD' )(z,y,0) = \Theta_{d+1}( \cD(\Gamma(\varphi),w,\Lambda) )(z,y,0)$$ and for almost all $z\in K^{d}$ that $$\Theta_{d+1}( U' )(z,0) = \Theta_{d+1}( \cD(U,u,\Lambda) )(z,0).$$ Since for all $(z,y,\lambda)\in \cD' $ one has $|(z,y,\lambda) | = |(z,\lambda) |$, by the bijectivity of $p:\cD'\to U'$, and by definition of $\Theta_{d+1}$, one finds $$\Theta_{d+1}( \cD' )(z,y,0) = \Theta_{d+1}( U' )(z,0).$$ This shows that $SC_w^{\Lambda} (\Gamma(\varphi))(z,y)=1$ for almost all $(z,y)\in C_w^{\Lambda} (\Gamma(\varphi))$. It also follows that $$\Theta_d (SC_w^{\Lambda} (\Gamma(\varphi))) (0) = \Theta_d (C_w^{\Lambda} (\Gamma(\varphi))) (0).$$ We proceed with similar arguments to show the remaining equalities. Assume from now on until the end of the proof, for simplicity, that $w=0$. By Proposition \[3.6\] (c) one has $\Theta_d (\Gamma (\psi)) (0) = \Theta_d (C_w^\Lambda(\Gamma ( \varphi)) (0)$. By Propositions \[prop1.7KR\] and \[3.6\] and since $\varepsilon < 1$, the map $z \mapsto (z, \psi (z))$ defined for $z$ in $C_u^{\Lambda, \Omega} (U)$ preserves the norm in the sense that $|z| =|(z,\psi(z) )|$ (recall that one uses the $\sup$-norm for tuples in an ultrametric setting). Hence, by the definition of $\Theta_d$, by Lemma \[nonempty2\] and by section \[opencase\], one has $\Theta_d (\Gamma (\psi)) (0) = \Theta_d (C_u^{\Lambda, \Omega} (U)) (0) = \Theta_d (C_u^{\Lambda} (U)) (0) = \Theta_d (U) (0)$. Combining the obtained series of equalities yields $\Theta_d (SC_w^{\Lambda} (\Gamma(\varphi))) (0) = \Theta_d (U) (0)$. Finally we prove that this value also equals $\Theta_d (\Gamma(\varphi)) (0)$, by again a similar argument. Define $U'':=\{z\in U,\ |\varphi(z)| < |z| \}$. Then, by Lemma \[prop1.7KR\] and its proof based on Lemma \[curve\], we find $C_u^{\Lambda} (U) = C_u^{\Lambda} (U'')$. Hence, $\Theta (C_u^{\Lambda} (U))(0) = \Theta_d (C_u^{\Lambda} (U'') )(0)$ which also equals $ \Theta_d (U'' )(0) $ by section \[opencase\]. Since on $U''$ the map $z\mapsto (z,\varphi(z))$ preserves the norm in the sense that $|z| = |(z,\varphi(z))|$ we find by the definition of $\Theta_d$ that $ \Theta_d(U'')(u) = \Theta_d ( \Gamma(\varphi_{|U''}) ) (w) = \Theta_d ( \Gamma(\varphi) ) (w)$ which finishes the proof. An alternative view on cones with multiplicities. ------------------------------------------------- Let $X\subset K^n$ be definable and of dimension $d$. It follows from Proposition \[3.6\] and its corollary \[trivcorSC\] that there is a finite definable partition of $X$ into parts $X_j$ which are graphs of $\varepsilon$-analytic Lipschitz functions on open subsets $U_j$, such that, for small enough $\Lambda$, $\Theta_d(X_j)(0)=\Theta_d(C^\Lambda_0(X_j))(0)$ for each $j$. It follows by additivity that $$\sum_j \Theta_d (C^\Lambda_0(X_j)) = \sum_j \Theta_d(X_j)=\Theta_d (X).$$ This common value can of course can be different from $\Theta_d(C_0^\Lambda(X))$ since $X_j$ et $X_k$ may have tangent cones which coincide on a part of dimension $d$ for different $j$, $k$, that is, there might be overlap in the union $C_0^\Lambda(X) = \cup_j C_0^\Lambda(X_j)$. Let us decompose $C_0^\Lambda(X)$ into parts $C_k$, $k\geq 1$, with the property that a line $\ell\subset C_k$ (through the origin) belongs to $C_0^\Lambda(X_j)$ for exactly $k$ different $j$. (Note that such decomposition is in general not unique.) Let us then define the function $CM_0^{\Lambda} (X)$ on $C_0^{\Lambda} (X)$, up to definable subsets of $C_0^{\Lambda} (X)$ of dimension $<d$, by $$CM_0^{\Lambda} (X)=\sum_k k\cdot \11_{C_k}.$$ Any other such decomposition of $X$ into parts $X_j$ will yield the same function $CM_0^{\Lambda} (X)$ up to a definable subset of $C_0^{\Lambda} (X)$ of dimension $<d$, as can be seen by taking common refinements and by general dimension theory of definable sets. Clearly $\Theta_d(CM_0^{\Lambda} (X))= \sum_j \Theta_d(X_j)=\Theta_d (X)$. Moreover, by additivity of $SC$ and by Corollary \[trivcorSC\], for all $z\in C_0^{\Lambda} (X)$ up to a definable set of dimension $<d$ $$SC_0^{\Lambda} (X)(z) = CM_0^{\Lambda} (X)(z).$$ In particular it follows that $SC_0^{\Lambda} (X)(z)$ is a nonnegative integer for all $z\in C_0^{\Lambda} (X)$ up to a definable set of dimension $<d$. End of proof of Theorem \[mt\] {#ept} ------------------------------ We consider a definable subset $X$ of dimension $d$ in $K^n$ and a point $x$ of $K^n$. We may assume $x$ lies in the closure of $X$. Let us fix $0 < \varepsilon < 1$. By Proposition \[1.4\] there exists a decomposition $$X = \bigcup_{1 \leq i \leq N (\varepsilon)} \gamma_i (\Gamma_i (\varepsilon)) \cup Y$$ with $Y$ a definable subset of $X$ of dimension $< d$, definable open subsets $U_i (\varepsilon)$ of $K^d$, for $1 \leq i \leq N (\varepsilon)$, definable analytic functions $\varphi_i (\varepsilon) : U_i (\varepsilon) \rightarrow K^{m - d}$ whose graphs $\Gamma_i (\varepsilon)$ are all $\varepsilon$-analytic, and elements $\gamma_1$, …, $\gamma_{N (\varepsilon)}$ in $\GL_m (R)$ such that the sets $\gamma_i ( \Gamma_i (\varepsilon))$ are all disjoint and contained in $X$. We denote by $u_i$ the image of $\gamma_i^{-1} (x)$ under the projection to $K^d$ and we fix $\Lambda$ adapted to $(X,x)$ and to $(U_i (\varepsilon), u_i)$ for every $1 \leq i \leq N (\varepsilon)$. By linearity and since the $\gamma_i$’s are isometries, we have then $$\Theta_d (X) (x) = \sum_{1 \leq i \leq N (\varepsilon)} \Theta_d (\Gamma_i (\varepsilon)) (\gamma_i^{-1} (x))$$ and $$\Theta_d (SC_x^{\Lambda}(X)) (0) = \sum_{1 \leq i \leq N (\varepsilon)} \Theta_d (SC_{\gamma_i^{-1} (x)}^{\Lambda}(\Gamma_i (\varepsilon))) (0),$$ and the result follows from Corollary \[trivcorSC\]. Existence of distinguished tangent $\Lambda$-cones {#edt} -------------------------------------------------- We deduce from Corollary \[distinguished cone for open sets\] and Proposition \[3.6\] the existence of distinguished tangent $\Lambda$-cones. \[distinguished cone\] Let $X$ be a definable subset of $K^n$. Then there exists $\Lambda\in \cal D$ such that for any $x\in \overline X$, $C_x^\Lambda(X)$ is a distinguished $\Lambda$-cone, that is to say $\Lambda'\subset \Lambda$ implies $C_x^{\Lambda'}(X)=C_x^\Lambda(X)$. We will work by induction on the dimension $d$ of $X$, where for $d=0$ the statement is trivial. We may work up to a finite partition of $X$ into definable pieces $X_k$ with distinguished $\Lambda_k$-cones $C_x^{\Lambda_k}(X_k)$ for all $x$ and for some $\Lambda_k$, since one can put $\Lambda := \cap_k \Lambda_k$ and then $C_x^\Lambda(X)=\cup_k C_x^\Lambda(X_k)$ implies that $C_x^\Lambda(X)$ is a distinguished $\Lambda$-cone for all $x$. Up to a finite partition using Proposition \[1.4\] and Theorem \[Lipschitz decomposition\], we may suppose that $X$ is the graph of some definable $C$-Lipschitz and $\varepsilon$-analytic map $\varphi:U\to K^{n-d}$, where $U$ is a definable open subset of $K^{d}$ and $d$ is the dimension of $X$. Fix $x\in \overline X$ and write $u\in \overline U$ for the projection of $x$ in $K^d$. We will construct a distinguished $\Lambda$ for this fixed $x$, with the extra property that in the construction one could as well take $x$ as a parameter running over $K^n$ and consider the analogue of the set-up in families parameterized by $x$, and then only finitely many $\Lambda$ will come up. Taking the intersection of these finitely many $\Lambda$ as above then finishes the proof. First suppose that $\varphi$ falls under the conditions of Proposition \[3.6\], that is, $C_{u}^\Lambda(U)$ has maximal dimension $d$ for some $\Lambda\in \cal D $ which is adapted to $U$. We know from Corollary \[distinguished cone for open sets\] that $\Lambda$ is distinguished for $U$, meaning that for $\Lambda'\subset \Lambda$ one has $$\label{lambdal'} C_{u}^{\Lambda'}(U) =C_{u}^{\Lambda}(U).$$ Fix $\Lambda'\subset \Lambda$ and consider $$\psi: C_{u}^{\Lambda,\Omega}(U)\to K^{n-d}$$ and $$\psi': C_{u}^{\Lambda',\Omega'}(U)\to K^{n-d}$$ (the notation being coherent with Proposition \[3.6\]). We may suppose that $\Omega = \Omega'$. But then $C_{u}^{\Lambda,\Omega}(U) = C_{u}^{\Lambda',\Omega'}(U)$ by Equation (\[lambdal’\]), and, for any $z$ in this set, we have $\psi(z)=\psi'(z)$ by Proposition \[3.6\] (a). Hence, $\psi$ and $\psi'$ are the same function. Taking the closures of the graph of this function, Proposition \[3.6\] (c) now yields that $C_x^{\Lambda'}(X) = C_x^{\Lambda}(X)$ and we are done in this case. Let us finally consider the case that $C^\Lambda_x(X)$ has dimension $<d$ for some $\Lambda$ (which happens if and only if $C^\Lambda_u(U)$ has dimension $<d$). We will construct a definable $Y\subset X$ such that $\dim(Y)=\dim(C^\Lambda_x(X))$ and $C_x^\Lambda(Y)=C_x^\Lambda(X)$. Then we can replace $X$ by $Y$ and we are done by induction on the dimension. Let $h:\overline{\cal D(X,x,\Lambda)} \to K$ be the deformation to $C_x^\Lambda(X)$. We assume $x=0$ in what follows, though we keep the notation $x$. Let $\cal L(C^\Lambda_x(X))$ be $$C^\Lambda_x(X)\cap \displaystyle\bigcup_{i=0}^{e-1}S(0,i),$$ where $e=[K^\times:\Lambda]$. We call $\cal L(C^\Lambda_x(X))$ the $\Lambda$-link of $C^\Lambda_x(X)$. Note that the $\Lambda$-cone generated by $\cal L(C^\Lambda_x(X))$ equals $C^\Lambda_x(X)$. Let $\tilde{\cal L}(C^\Lambda_x(X))$ be ${\cal L}(C^\Lambda_x(X))\times (B(0,n)\cap \Lambda)$ for some ball $B(0,n)$ around $0$. Since there are definable choice functions, there is a map $$d: \tilde{\cal L}(C^\Lambda_x(X))\to \overline{\cal D(X,x,\Lambda)}$$ with $ d(z,\lambda)\in h^{-1}(\lambda)$ for all $\lambda$ and $\lim_{\lambda\to 0 } d(z,\lambda)=z$ for all $z$. Since we may and do suppose that $z\not = d(z,\lambda)$, the image of $d$ is of dimension $\dim (\tilde{\cal L}(C^\Lambda_x(X)) ) = \dim(\cal L(C^\Lambda_x(X)))+1=\dim(C^\Lambda_x(X))+1$. We send $d( \tilde{\cal L}(C^\Lambda_x(X)))$ into $X$ by $r(z,\lambda)=\lambda\cdot z$ and we set $Y=r(d( \tilde{\cal L}(C^\Lambda_x(X))))$. Then $Y$ is a definable subset of $X$ of dimension $\dim(C^\Lambda_x(X))$ and by construction $C^\Lambda_x(Y)=C^\Lambda_x(X)$. A local Crofton formula ======================= Local direct image ------------------ Let $p : X \rightarrow Y$ be a definable function between two definable sets of the same dimension $d$. If $\varphi$ is a function in $\cC (X)$ and $y$ is in $Y$ we set $p_! (\varphi) (y) = \sum_{x \in p^{- 1}(y)} \varphi (x)$ if $p^{- 1}(y)$ is finite and $p_! (\varphi) (y) = 0$ if it is infinite. The function $p_! (\varphi)$ lies in $\cC (Y)$, since the cardinality of $p^{- 1}(y)$ takes only finitely many values when $y$ runs over $Y$. If $X$ is a definable subset of $K^n$ and $x$ is a point of $K^n$, we define the algebra $\cC (X)_x$ of germs of constructible functions in $\cC (X)$ at $x$ to be the quotient of $\cC (X)$ by the equivalence relation $\varphi \sim \varphi'$ if $ \11_{B (x , n)} \varphi = \11_{B (x , n)} \varphi'$ for $n$ large enough. That definition is only relevant when $x$ is in the closure of $X$. Also, if $\varphi$ is in $\cC (X)_x$ is the germ of a locally bounded function $\psi$, $\Theta_d (\varphi) :=\Theta_d (\psi) (x)$ does not depend on the representative $\psi$. Let $p : K^m \rightarrow K^{d}$ be a linear projection and let $X$ and $Y$ be respectively definable subsets of $K^m$ and $K^{d}$ such that $p (X) \subset Y$. Fix $x$ in $K^m$. When the condition (\[sparks\]) is satisfied $$\label{sparks}\tag{$\ast$} \text{there exists}\, \, n \geq 0 \, \, \text{such that} \, \, p^{-1} (p (x)) \cap \overline{X \cap B (x, n)}= \{x\},$$ then, for every function $\varphi$ in $\cC (X)$, the class of $p_! (\varphi \11_{B (x, n)})$ in $\cC (Y)_{p (x)}$ does not depend on $n$ for $n$ large enough. We denote it by $p_{!, x} (\varphi)$. We also denote by $p_{!, x}$ the corresponding morphism $\cC (X)_x \rightarrow \cC (Y)_{p (x)}$. The local Crofton formula for the local density ----------------------------------------------- For $x$ a point in $K^n$ we consider $K^n$ as a vector space with origin $x$ and for $0 \leq d \leq n$, we denote by $G (n, n-d)$ the corresponding Grassmannian of $(n-d)$-dimensional vector subspaces of $K^n$. It is a compact $K$-analytic variety, endowed with a unique measure $\mu_{n, d}$ invariant under $\GL_n (R)$ and such that $\mu_{n, d} (G (n, n-d)) = 1$. For any $V$ in $G (n, n-d)$, we denote by $p_V : K^n \rightarrow K^n/V$, the canonical projection, where $K^n/V$ is identified with the $K$-vector space $K^d$. This identification enables the computation of the local density of germs in $K^n/V$. Let $X$ be a definable subset of $K^n$ of dimension $d$ and let $x$ be a point of $K^n$. By general dimension theory for definable sets there exists a dense definable open subset $\Omega(=\Omega_X)$ of $G(n, n-d)$ such that for every $V$ in $\Omega$ the projection $p_V$ satisfies the condition \[sparks\] with respect to $(X, x)$. The following statement is the $p$-adic analogue of the so-called [*local Crofton formula*]{} proved in [@comte] for real subanalytic sets and more generally in [@lk], again in the real subanalytic setting, in its multidimensional version. \[lcc\] Let $X$ be a definable subset of $K^n$ of dimension $d$ and let $x$ be a point of $K^n$. Let $\varphi$ in $\cC (X)_x$ be the germ of a locally bounded function. Then $$\Theta_d (\varphi) (x) = \int_{V\in \Omega \subset G (n, n-d)} \Theta_d (p_{V !, x} (\varphi)) \, d \mu_{n, d}(V).$$ We may assume that $X=\overline X$ by Proposition \[bord\] and that $x=0$ and $0\in X$, for if $0\not\in X$, $\Theta_d(X)(0) = \Theta_d (p_{V !, x} (\varphi))=0$ (for generic $V$) and the statement of Theorem \[lcc\] is then true. In order to emphasize the geometric-measure part of \[lcc\] we start with the following lemma, which is Theorem \[lcc\] for $X$ a definable $\Lambda$-cone of $K^n$ of dimension $d$ contained in some $d$-dimensional vector space of $K^n$ and $\varphi=\11_X$. \[lccconeplan\] Let $\Lambda \in \cD$, $\Pi\in G(n,d)$ and $X$ be a definable $\Lambda$-cone contained in $\Pi$ and with origin $0$. Then $$\Theta_d(X)(0)=\int_{V\in \Omega\subset G(n,n-d)} \Theta_d(p_V(X))(p_V(0)) \, d \mu_{n,d} .$$ For every $V\in G(n,n-d)$, by linearity of $p_V$, $p_V(X)$ is a $\Lambda$-cone of $K^n/V$ with origin $p_V(0)$, and as $\dim(\Pi)=\dim(K^n/V)$, $p_V(X)$ is isomorphic to $X$, for generic $V$ ($V\in \Omega=\Omega_\Pi$). In what follows we denote $p_V(0)$ by $0$. Take an integer $e>0$ such that $\pi_K^e\in\Lambda$, where we recall that $\pi_K$ is a uniformizer of $R$. The sets $X$ and $p_V(X)$ being $\Lambda$-cones, one has the following disjoint union relations $$ X=\displaystyle \coprod_{z\in \ZZ} \ \pi_K^{ze}\cdot ( \coprod_{i=0}^{e-1} X \cap S(0,i) ) , $$ $$\hbox{ and }\ \ p_V(X)=\displaystyle \coprod_{z\in \ZZ} \ \pi_K^{ze}\cdot (\coprod_{c=0}^{e-1}( p_V(X)) \cap S(0,c) ).$$ It follows by the definition of $\Theta_d$ that $$\label{a} \Theta_d(X)(0)=\frac{(1-q^{-d})^{-1}}{e} \sum_{i=0}^{e-1}q^{id} \mu_d(X \cap S(0,i))$$ $$\label{b} \hbox{ and }\ \ \Theta_d(p_V(X))(0)=\frac{(1-q^{-d})^{-1}}{e} \sum_{c=0}^{e-1}q^{cd} \mu_d(p_V(X)\cap S(0,c)).$$ For each $i=0,\ldots,e-1$, let $C_i$ be $$\coprod_{z\in\ZZ} \pi_K^{ze} \cdot p_V(X\cap S(0,i)).$$ One has $p_V(X) = \coprod_i C_i$ by the linearity of $p_V$, and, the $C_i$ are definable since $X$ is a $\Lambda$-cone. Define the disjoint definable sets $ A_c^i$, for $i$ and $c$ going from $0$ to $e-1$, by $$A_c^i = C_i \cap S(0,c).$$ Clearly $$\coprod_{i=0}^{e-1}A_c^i = p_V(X) \cap S(0,c).$$ Moreover, the sets $\pi_K^{i-c}\cdot A_c^i$ are disjoint by linearity of $p_V$ and by bijectivity of $p_V$ on $\Pi$. By the fact that $$q^{cd}\mu_d(A^i_c)=q^{id}\mu_d(\pi_K^{i-c}\cdot A_c^i),$$ we obtain $$\sum_{c=0}^{e-1}q^{cd} \mu_d(p_V(X)\cap S(0,c)) = \sum_{c=0}^{e-1}q^{cd} \sum_{i=0}^{e-1} \mu_d(A_c^i) = \sum_{i=0}^{e-1} q^{id}\mu_d(\coprod_{c=0}^{e-1} \pi_K^{i-c}\cdot A_c^i)$$ $$\label{c} =\sum_{i=0}^{e-1}q^{id}\mu_d(B_V^i),$$ where $B^i_V:=\displaystyle\coprod_{c=0}^{e-1} \pi_K^{i-c}\cdot A_c^i$. Let us now consider $\Phi_{V}: \Pi\setminus \{0\}\to (K^n/V)\setminus \{0\}$, defined by $$\Phi_V( x)= \pi_K^{ \ord(x) - \ord (p_V(x)) } \cdot p_V(x).$$ This map is bijective from $X\cap S(0,i)$ to $B^i_V$, since $p_V$ is bijective from $X$ to $p_V(X)$. By change of variables one obtains $$\mu_d(B_V^i)= \int_{X\cap S(0,i)} \vert \Jac ( \Phi_V) \vert \, d\mu_d.$$ Furthermore, by Fubini, $$\label{d} \int_{V\in \Omega} \mu_d(B_V^i)\, d\mu_{n,d}= \int_{x\in X\cap S(0,i)}\int_{V\in \Omega} \vert \Jac ( \Phi_V)(x) \vert \, d\mu_{n,d}(V) \, d\mu_d(x).$$ Note that, for $x\in S(0,i)$, the quantity $\kappa_i= \int_{V\in \Omega} \vert \Jac ( \Phi_V)(x) \vert \, d\mu_{n,d}(V)$ does not depend on $x$. Indeed, $\GL_n (R)$ acts transitively on $S(0,i)$, $\mu_{n,d}$ is invariant under this action and if $g\in \GL_n (R)$ and $x'=g\cdot x$ for $x,x'\in S(0,i)$, then $\Jac (\Phi_V)(x')= \Jac (\Phi_{g^{-1}\cdot V})(x)$. Moreover, by linearity of $p_V$, one has that $\kappa_i=\kappa$ is independent of $i$. It follows from (\[b\]), (\[c\]) and (\[d\]) that $$\int_{V\in \Omega} \Theta_d(p_V(X))(0)\, d\mu_{n,d}(V) =\frac{(1-q^{-d})^{-1}}{e}\sum_{i=0}^{e-1} q^{id} \int_{x\in X\cap S(0,i)} \kappa\, d\mu_d (x).$$ $$= \kappa \cdot \frac{(1-q^{-d})^{-1}}{e}\sum_{i=0}^{e-1} q^{id} \mu_d(X\cap S(0,i)).$$ Finally, by (\[a\]), we obtain $$\int_{V\in \Omega} \Theta_d(p_V(X))(0)\, d\mu_{n,d}(V) =\kappa\cdot \Theta_d(X)(0).$$ One gets $\kappa=1$ by taking $X=\Pi$ in the latter formula. Lemma \[lccconeplan\] may be viewed as the tangential formulation of the local Crofton formula for general definable $\Lambda$-cone sets and its proof captures its geometric measure content. Note that its proof still works assuming that $X$ is a definable $\Lambda$-cone of dimension $d$ in $K^n$, instead of a definable $\Lambda$-cone of dimension $d$ contained in some $d$-dimensional vector space $\Pi$. Indeed, it is essentially enough to replace, in the proof of Lemma \[lccconeplan\], $\Phi_V:\Pi\setminus \{0\}\to (K^n/V)\setminus \{0\}$ by the restriction of the mapping $ w\mapsto \Psi_V(x)= \pi_K^{ \ord(x) - \ord (p_V(x)) } p_V(x)$ on the smooth part of $X$ (the fibers of $\Psi_{V\vert X}$ being counted with multiplicity in the area formula). Hence we get the following extension of Lemma \[lccconeplan\]: \[lcccone\] Let $\Lambda \in \cD$ and $X$ be a definable $\Lambda$-cone of $K^n$ with origin $0$. Then $$\Theta_d(X)(0)=\int_{\Omega\subset G(n,n-d)} \Theta_d(p_{V!,0}(\11_X)) \ d \mu_{n,d}(V) . \qed $$ For $V\in G(n,n-d)$ and $y\in (K^n/ V)\setminus \{0\}$, let us denote by $\tilde V_y$ the fiber $\Psi_V^{-1}(\{y\})$ of $\Psi_V: K^n\setminus V\to (K^n/ V)\setminus\{0\} $, where $\Psi_V(x)= \pi_K^{ \ord(x) - \ord (p_V(x)) }p_V(x)$. Note that $\tilde V_y\subset S(0, \ord y )\setminus V$ and $GL_n(R)$ acts on $\tilde V=\{\tilde V_y;V\in G(n,n-d), y\in (K^n/V)\setminus \{0\}\}$. For $X$ a definable set of dimension $d$ in $S(0,c)$, where $c\in \ZZ$, the statement of Lemma \[lcccone\] may be reformulated as $$\mu_d(X)=\int_{V\in \Omega} \int_{ y\in S(0,c)\subset K^n/V} \#(X\cap \tilde V_y) \ d\mu_d(y)\ d\mu_{n,d}(V).$$ Now note that the mapping $(V,y)\mapsto \tilde V_y$ defined from $\{(V,y); V\in G(n,n-d), y\in K^n/V\} $ to $\tilde V$ is one-to-one and that the image of the Haar measure of $GL_n(R)$ under $g\mapsto g\cdot \tilde V_0$ (for $\tilde V_0$ fixed in $\tilde V$) gives a $GL_n(R)$-invariant measure $\nu$ on $\tilde V$ such that for $E\subset \tilde V$, $E$ subanalytic say, we have $$\nu(E)=\int_{V\in \Omega} \int_{ y\in S(0,c)\subset K^n/V} \11_{E}(\tilde V_y) \ d\mu_d(y)\ d\mu_{n,d}(V).$$ To obtain the above equality, it is enough to remark that the right hand side gives a function on subsets of $\tilde V$ which pulls back on $Gl_n(R)$ as a Haar measure. With these notations we see that Lemma \[lcccone\] is nothing else than the classical spherical Crofton formula for $X\cap S(0,c)$ (for a standard reference see [@Fe], Theorem 3.2.48 and note that the proof may be applied in our setting): \[scc\] Let $X$ be a definable set of $ S(0,0)\subset K^n$ of dimension $d$, then $$\mu_d(X)= \int_{V\in \Omega} \int_{ y\in S(0,0)\subset K^n/V} \#(X\cap \tilde V_y) \ d\mu_d(y)\ d\mu_{n,d}(V)$$ $$= \int_{\tilde v\in \tilde V} \#(X\cap \tilde v) \ d\nu(\tilde v).$$ For the general setting we will use the following auxiliary lemma. \[partition\] Let $\Lambda$ be in $\cD$ and let $X\subset K^n$ be a definable set of dimension $d$. Suppose that $p:K^n\to K^d$ is a coordinate projection which is injective on $X$. Then there exist definable sets $C_j$ of dimension $<d$ and a finite partition of $X$ into definable parts $X_j$ such that $p$ is injective on $C_0^{\Lambda}(X_j)\setminus C_j$ for each $j$. Since $C_0^{\Lambda}(X)\subset C_0^{K^\times}(X)$ for any $\Lambda$ in $\cD$, we may suppose that $\Lambda=K^\times$. We may also suppose that $0\in \overline{X}\setminus X$. Partition $C_0^{K^\times }(X)$ into finitely many definable parts $B_j$ such that $p$ is injective on each set $B_j$. By linearity of $p$ we may suppose that each $B_j$ is a $K^\times$-cone. For each $j$ let $B_j'$ be the definable subset of $K^n$ consisting of the union of all lines $\ell\in K^n$ through $0$ such that the distance between $\ell\cap S(0,0)$ and $B_j\cap S(0,0)$ is strictly smaller than the distance between $\ell\cap S(0,0)$ and $B_i\cap S(0,0)$ for all $i\not = j$. Put $X_{j}^0:= X\cap B_j'$ for each $j$, and take a finite definable partition of $X$ into parts $X_j$ satisfying $X_{j}^0\subset X_j$ for each $j$. By construction $C_0^{K^\times }(X_j) = C_0^{K^\times }(X_{j}^0) \subset \overline B_j$. Let $C_j$ be $\overline{B_j}\setminus B_j$. Then the $X_j$ and $C_j$ are as desired. We now prove Theorem \[lcc\] in its general setting, that is to say, for $X$ a given definable subset of $K^n$ of dimension $d$ instead of some definable $\Lambda$-cone of $K^n$ as in Lemma \[lcccone\]. As in the proof of Theorem \[mt\] we may assume that $\varphi=\11_X$. Up to a finite partition of $X$ into definable parts we may suppose that $X$ is the graph of an $\varepsilon$-analytic map $U\subset K^d\to K^{n-d}$ as in Corollary \[trivcorSC\] and then it follows by this corollary that $$\Theta_d(X)(0)= \Theta_d( SC_0^{\Lambda} (X) )(0) = \Theta_d ( C_0^{\Lambda}(X) )(0).$$ For $C_0^{\Lambda}(X) $ we know that Theorem \[lcc\] holds by Lemma \[lcccone\], that is $$\Theta_d ( C_0^{\Lambda}(X) ) (0) = \int_{V\in \Omega \subset G (n, n-d)} \Theta_d (p_{V !, 0} (\11_{C_0^{\Lambda}(X)} ) \, d \mu_{n, d}(V).$$ We claim that, for generic $V$, $$\Theta_d (p_{V!,0}(\11_{C_0^{\Lambda}(X)})) = \Theta_d (p_{V!,0}(\11_X))$$ which finishes the proof. We prove the claim as follows. Fix $V$. By Lemma \[partition\] we can partition $X$ into finitely many definable parts $X_j$ (depending on $V$) such that $p_V$ is injective on $X_j$ and, up to a definable set of dimension $<d$, also on $C_0^{\Lambda}(X_j)$. By additivity it is now enough to prove that $$\Theta_d (p_{V}(C_0^{\Lambda}(X_j)))(0) = \Theta_d (p_{V}(X_j))(0),$$ which follows from Theorem \[mt\] for open sets since $p_{V}(C_0^{\Lambda}(X_j)) = C_0^{\Lambda}(p_{V}(X_j)).$ [SGA]{} K. Bekka, *Regular stratification of subanalytic sets*, Bull. London Math. Soc. **25** (1993), 7–16. R. Cluckers, *Presburger sets and $p$-minimal fields*, J. Symbolic Logic **68** (2003), 153–162. R. Cluckers, *Analytic $p$-adic cell decomposition and integrals*, Trans. Amer. Math. Soc. **356** (2004), 1489–1499. R. Cluckers, *Multi-variate [I]{}gusa theory: Decay rates of $p$-adic exponential sums*, Int. Math. Res. Not. **76** (2004), 4093–4108. R. Cluckers, G. Comte and F. Loeser, *Lipschitz continuity properties for $p$-adic semi-algebraic and subanalytic functions*, to appear in GAFA, arXiv:0904.3853. R. Cluckers, L. Lipshitz and Z. Robinson, *Analytic cell decomposition and analytic motivic integration*, Ann. Sci. École Norm. Sup. **39** (2006), 535–568. R. Cluckers, L. Lipshitz, *Fields with analytic structure*, arXiv:0908.2376. R. Cluckers, F. Loeser, *Constructible motivic functions and motivic integration*, Invent. Math. **173** (2008), 23–121. P. J. Cohen, *Decision procedures for real and [$p$]{}-adic fields*, Comm. Pure Appl. Math. **22** (1969), 131–151. G. Comte, *' Equisingularité réelle : nombres de Lelong et images polaires*, Ann. Sci. [É]{}cole Norm. Sup. **33** (2000), 757–788. G. Comte, M. Merle, *Équisingularité réelle II : invariants locaux et conditions de régularité*, Ann. Sci. [É]{}cole Norm. Sup. **41** (2008), 1–48. G. Comte, J.-M. Lion, J.-P. Rolin, *Nature log-analytique du volume des sous-analytiques*, Illinois J. Math. **44** (2000), 884–888. J. Denef, *The rationality of the Poincaré series associated to the $p$-adic points on a variety*, Invent. Math. **77** (1984), 1–23. J. Denef, *On the evaluation of certain $p$-adic integrals*, Séminaire de théorie des nombres, Paris 1983–84, 25–47, Progr. Math., **59**, Birkhäuser Boston, Boston, MA, 1985. J. Denef, *$p$-adic semi-algebraic sets and cell decomposition*, J. Reine Angew. Math. **369** (1986), 154–166. J. Denef, L. van den Dries, *$p$-adic and real subanalytic sets*, Ann. of Math. **128**, (1988), 79–138. R. Draper, *Intersection theory in analytic geometry*, Math. Ann. **180** 1969, 175–204. L. van den Dries, D. Haskell and D. Macpherson, *One-dimensional $p$-adic subanalytic sets*, J. London Math. Soc. **59**(1999), 1–20. L. van den Dries, P. Scowcroft, *On the structure of semi-algebraic sets over $p$-adic fields*, J. Symbolic Logic **53** (1988), 1138–1164. , *Dimension of definable sets, algebraic boundedness and [H]{}enselian fields*, Ann. Pure Appl. Logic **45** (1989), 189–209. H. Federer, *Geometric measure theory*, Grundlehren Math. Wiss. **153**, Springer Verlag (1969). D. Heifetz, *$p$-adic oscillatory integrals and wave front sets*, Pacific J. Math. **116** (1985), 285–305. J. P. Henry, M. Merle, C. Sabbah, *Sur la condition de Thom stricte pour un morphisme analytique complexe*, Ann. Sci. [É]{}cole Norm. Sup. **17** (1984), no. 2, 227–268. H. Hironaka, *Stratification and flatness*, Real and complex singularities (Proc. Ninth Nordic Summer School/NAVF Sympos. Math., Oslo, 1976), Sijthoff and Noordhoff, Alphen aan den Rijn, (1977), 199–265. T. C. Kuo, *The ratio test for analytic Whitney stratifications* , Proceedings of Liverpool Singularities Symposium I, Lecture Notes in Math., **192**, Springer, Berlin, (1971), 141–149. K. Kurdyka, *On a subanalytic stratification satisfying a Whitney property with exponent 1*. Real algebraic geometry (Rennes, 1991), Lecture Notes in Math., **1524**, Springer, Berlin, (1992), 316–322. K. Kurdyka, A. Parusiński, *$w\sb f$-stratification of subanalytic functions and the Lojasiewicz inequality*. C. R. Acad. Sci. Paris Sér. I Math., **318**, no. 2, (1994), 129–133. K. Kurdyka, J.-B. Poly, G. Raby, *Moyennes des fonctions sous-analytiques, densité, cône tangent et tranches*. Real analytic and algebraic geometry (Trento, 1988), Lecture Notes in Math., **1420**, Springer, Berlin, (1990), 170–177. K. Kurdyka, G. Raby, *Densité des ensembles sous-analytiques*, Ann. Inst. Fourier **39** (1989), 753–771. P. Lelong, *Intégration sur un ensemble analytique complexe*, Bull. Soc. Math. France **85** (1957), 239–262. J.-M. Lion, *Densité des ensembles semi-pfaffiens*, Ann. Fac. Sci. Toulouse Math. **7** (1998), 87–92. T. L. Loi, *Thom stratifications for functions definable in o-minimal structures on $(\RR, +, \cdot)$*, C. R. Acad. Sci. Paris Sér. I Math. **324** (1997), no. 12, 1391–1394. T. L. Loi, *Verdier and strict Thom stratifications in o-minimal structures*, Illinois J. Math. **42** (1998), no. 2, 347–356. J. Oesterl[é]{}, *R[é]{}duction modulo $p^{n}$ des sous-ensembles analytiques ferm[é]{}s de ${\ZZ}^{N}_{p}$*, Invent. Math. **66** (1982), 325–341. W. Pawłucki, *Le théorème de Puiseux pour une application sous-analytique*, Bull. Polish Acad. Sci. Math. **32** (1984), no. 9-10, 555–560. J.-P. Serre, *Quelques applications du théorème de densité de Chebotarev*, Inst. Hautes [É]{}tudes Sci. Publ. Math. **54** (1981), 323–401. Y.T. Siu, *Analyticity of sets associated to Lelong numbers and the extension of closed positive currents*, Invent. Math. **27** (1974), 53–156. P. Thie, *The Lelong number of a point of a complex analytic set*, Math. Ann. **172** (1967), 269–312. R. Thom, *Ensembles et morphismes stratifiés*, Bull. Amer. Math. Soc. **75** (1969), 240–284 J. L. Verdier, *Stratifications de Whitney et théorème de Bertini-Sard*, Invent. Math **36** (1976), 295–312. H. Whitney, *Tangents to an analytic variety*, Ann. of Math. **81** (1965), 496–549.

ArXiv

--- abstract: 'We study the effect of a pure spin current on the Kondo singlet in a diluted magnetic alloy using non-local lateral spin valve structures with highly spin polarized Co$_{2}$FeSi electrodes. Temperature dependence of the non-local spin signals shows a sharp reduction with decreasing temperature, followed by a plateau corresponding to the low temperature Fermi liquid regime below the Kondo temperature ([*T*]{}$_{\rm K}$). The spin diffusion length of the Kondo alloy is found to increase with the evolution of spin accumulation. The results are in agreement with the intuitive description that the Kondo singlet cannot survive any more in sufficiently large spin accumulation even below [*T*]{}$_{\rm K}$.' author: - 'K. Hamaya' - 'T. Kurokawa' - 'S. Oki' - 'S. Yamada' - 'T. Kanashima' - 'T. Taniyama' title: Direct Evidence for Suppression of the Kondo Effect due to Pure Spin Current --- According to the Anderson model[@Yoshida], the $s$-$d$ mixing or hybridization of a localized impurity spin and the surrounding conduction electrons results in the formation of a spin singlet state in diluted magnetic alloys[@Kondo; @Franck; @Applebaum; @Loram; @Monod; @Haldane] or artificial nanostructures such as a quantum dot (QD) [@Gold; @Cronenwett; @Gold2; @Wiel]. This many-body effect, so-called Kondo effect, is currently being researched extensively due to its rich physics in condensed matter [@Sato; @Lee; @Buizert; @Yamaguchi; @Bultelaar; @OBrien; @Batley; @Martinek1; @Martinek2; @Choi; @Barnas; @Ralph; @Gossard; @Hamaya; @Hauptmann; @Heersche; @Taniyama; @Kobayashi]; a logarithmic increase in the resistivity of diluted magnetic alloys below the characteristic temperature, i.e., the Kondo temperature ([*T*]{}$_{\rm K}$), is a representative aspect of the Kondo effect. Among them, the effect of spin polarized electrons on the Kondo singlet is now receiving great interest from both theoretical [@Martinek1; @Martinek2; @Choi; @Barnas] and experimental [@Ralph; @Hamaya; @Heersche; @Hauptmann; @Taniyama] points of view. As illustrated in Fig. 1(a), the low temperature Fermi liquid Kondo regime at [*T*]{} $<$ [*T*]{}$_{\rm K}$ is a direct consequence of the spin-flip scattering between the impurity spin and conduction electron spins in a nonmagnetic host in zero magnetic field ($\Delta \epsilon_{\rm d}=\epsilon_{\rm d\uparrow}-\epsilon_{\rm d\downarrow}=0$). ![(Color online) Schematic illustrations of the density of states (DOS) in a Kondo alloy based on the Anderson model in [*T*]{} $<$ [*T*]{}$_{\rm K}$ (a) without and (b) with spin accumulation ($\Delta\mu$) in the host nonmagnetic metal. ](Fig1.eps){width="8cm"} If spin accumulation ($\Delta\mu=\mu_{\uparrow}-\mu_{\downarrow}$) defined by the difference in the chemical potentials of spin-up ($\mu_{\uparrow}$) and spin-down ($\mu_{\downarrow}$) electrons is generated in the host nonmagnetic metal[@MJ; @vanSon], the spin accumulation has an influence on the Kondo effect and even suppresses the formation of the Kondo singlet due to an additional energy cost for the spin-flip scattering (Fig. 1(b)). Similarly, as strong local exchange interaction[@Ralph; @Hamaya; @Hauptmann] or RKKY interaction[@Heersche] manifests itself in the QD coupled with ferromagnetic leads, the Kondo assisted tunneling is suppressed in the Kondo regime, leading to the splitting of the Kondo zero-bias anomaly even in zero magnetic field[@Martinek1; @Martinek2; @Choi; @Barnas]. Although such suppression of the Kondo effect has been observed by injecting spin polarized charge currents[@Taniyama], the effect of a pure spin current, which is a flow of spin angular moment without a net charge current[@Jedema; @Kimura], has not yet been elucidated in the Kondo systems because of the difficulty in detecting pure spin current transport in Kondo alloys or QDs. In this Letter, we report on direct evidence for the effect of a pure spin current in diluted magnetic Cu(Fe) Kondo alloy (the average Fe concentration of 100 $\sim$ 200 p.p.m)[@Franck; @Kondo; @Applebaum; @Monod; @Loram], by using non-local lateral spin valve (LSV) devices with highly spin polarized Co$_{2}$FeSi (CFS) electrodes. The half-metallic characteristics of CFS surely enable us to detect spin signals in the diluted magnetic Kondo alloy even at a very low signal level[@Wurmehl; @Bombor]. We find that the spin signal arising from pure spin current transport is strongly suppressed below the onset temperature ([*T*]{} $>$ [*T*]{}$_{\rm K}$) of $s$-$d$ spin-flip scattering while the spin signal does not decrease any more at [*T*]{} $<$ [*T*]{}$_{\rm K}$ because of the collective screening of the magnetic Fe impurity spins by Kondo clouds. The [*T*]{}$_{\rm K}$ estimated from the pure spin current measurement is in good agreement with that obtained from the temperature dependence of the resistivity. We also demonstrate that the Kondo effect can be tuned by generating larger spin accumulation in the diluted magnetic alloy. 25-nm-thick CFS layers were grown on non-doped FZ-Si(111) substrates by molecular beam epitaxy (MBE) at 60$^{\circ}$C, where Co, Fe and Si were co-evaporated from Knudsen cells[@Hamaya1; @Yamada]. In-situ reflection high energy electron diffraction patterns of CFS layers clearly exhibited symmetrical streaks, indicating good two-dimensional epitaxial growth[@Yamada]. The CFS layers were patterned into the submicron-sized electrodes by using conventional electron beam lithography and Ar ion milling for measurements of non-local spin signals[@Kimura1; @Hamaya2]. To fabricate the LSVs, 75-nm-thick Cu(Fe) wires bridging the two CFS electrodes with various center-to-center distances ($d$) were patterned by a conventional lift-off technique, together with bonding pads. The interface resistance between CFS and Cu(Fe) is negligible ($\leq$0.1 f$\Omega$m$^{2}$) by carefully cleaning the surface of the CFS layers with low-energy accelerated Ar ion milling[@Hamaya2]. Figure 2(a) shows a scanning electron microscopy image of a fabricated LSV. By using the cross terminal configuration, non-local spin signal measurements were carried out using a conventional current-bias lock-in technique (173 Hz) at various temperatures. {width="8.0cm"} We confirmed the Kondo effect in the Cu(Fe) wire of the LSV by measuring the temperature dependent resistivity, as shown in Fig. 2(b). The resistivity minimum ([*T*]{}$_{\rm min}$) can be seen at around 20 K, and the resistivity increases with decreasing temperature. An empirical functional model for the Kondo resistivity can be expressed as $\rho(T) = \rho_{0}$ + $AT^{2}$ + $BT^{5}$ + $\rho_{K}$/$\left[1+(2^{1/s} - 1)(T/T_{K})^{2}\right]^{s}$[@Gold2; @OBrien], where $\rho_{0}$ is the residual resistivity and the second and third terms are the electron-electron and electron-phonon scattering contributions, respectively. The last term arises from the Kondo effect[@Gold2; @OBrien]. Assuming $s \sim$ 0.22 in a spin-(1/2) Kondo system[@Gold2], a best fit curve with the above function was obtained as shown in the solid curve in Fig. 2(b), yielding [*T*]{}$_{\rm K} =$ 30 K. The values of $\rho_{\rm 0}$ and $\rho_{\rm K}$ are 1.791 $\mu\Omega$cm and 0.031 $\mu\Omega$cm, respectively. These analyses ensure that the Cu(Fe) wire we used shows a typical Kondo effect with [*T*]{}$_{\rm K} =$ 30 K[@Kondo; @Applebaum; @Monod; @Loram; @Haldane]. Figures 2(c) and 2(d) show typical non-local spin signals ($\Delta R_{\rm S}$) detected by using CFS-Cu(Fe)-CFS LSVs (Fig. 2(a)) with $d =$ 312 and 496 nm, respectively, at 5 K, where $\Delta R_{\rm S}$ is calculated by $\Delta V$/$I$. Clear hysteretic spin signals are observed even for the spin injection into Cu(Fe) Kondo alloys. This is the first experimental observation of the lateral transport of a pure spin current in a Kondo alloy. It should be noted that the spin signals, i.e., $\Delta R_{\rm S}$, was one order or two orders of magnitude smaller than those observed in typical Cu-based LSVs with a CFS spin injector and a detector in our previous work[@Kimura1; @Hamaya2]. This is caused by the small spin diffusion length of Cu(Fe) ($\lambda_{\rm Cu(Fe)}$)[@Taniyama]. Using the CFS-Cu(Fe)-CFS LSV devices with various $d$ values, we also measured temperature dependent $\Delta R_{\rm S}$ at $I =$ 1.0 mA as shown in Fig. 3(a). {width="8.0cm"} For all the LSVs, there is a clear maximum of $\Delta R_{\rm S}$ at a characteristic temperature which we refer to as [*T*]{}$_{\rm max}$. [*T*]{}$_{\rm max}$ is different for each LSV (see arrows). We find that the $\Delta R_{\rm S}$ values at 5 K are smaller than those at 300 K for all the LSVs. This feature is largely different from those observed in conventional Cu-based LSVs with CFS[@Hamaya2]. Of particular significance here is that the $\Delta R_{\rm S}$ is markedly decreased as the temperature is lowered from [*T*]{}$_{\rm max}$, and finally saturates below $\sim$ 20 K. Because spin signal $\Delta R_{\rm S}$ for a fixed terminal distance $d$ provides a measure of spin accumulation, the reduction in $\Delta R_{\rm S}$ at low temperatures is a signature of the enhanced spin-flip scattering events. In order to compare the temperature variation in the spin-flip scattering events for the LSVs with the different $d$ values, the normalized $\Delta R_{\rm S}$ values are also shown in Fig. 3(b) as a function of temperature, where the normalized $\Delta R_{\rm S}$ is defined as $\Delta R_{\rm S}$ divided by the maximum $\Delta R_{\rm S}$. Note that the normalized $\Delta R_{\rm S}$ values at [*T*]{} $\le$ [*T*]{}$_{\rm max}$ show significant $d$ dependence, in striking contrast to no $d$ dependence in the temperature region [*T*]{} $\ge$ [*T*]{}$_{\rm max}$. The large $d$ dependence of the normalized $\Delta R_{\rm S}$ at [*T*]{} $\le$ [*T*]{}$_{\rm max}$ indicates that the spin-flip scattering strongly depends on spin accumulation and it is intensively suppressed at a small $d$ due to the large spin accumulation, consistent with the Kondo effect as described before. We also define the value, ($1 -$ normalized $\Delta R_{\rm S}$), at [*T*]{} $\le$ [*T*]{}$_{\rm max}$, as shown in the inset of Fig. 3(b). The red line shows a fit with the logarithmic function, $a$\[1-$b$log($\frac{{\it T}}{{\it T}_{\rm K}}$)\], in the temperature range 40 K $\le$ [*T*]{} $\le$ [*T*]{}$_{\rm max}$, where $a$ and $b$ are the adjustable parameters. From these views, [*T*]{}$_{\rm max}$ is a critical temperature, below which the spin-flip scattering manifests itself, as discussed in the previous work[@OBrien]. Hereafter, we discuss the temperature independent behavior at [*T*]{} $\le$ 20 K in Fig. 3. The plateau of $\Delta R_{\rm S}$ at [*T*]{} $\le$ 20 K indicates that the degree of spin-flip scattering is almost constant even if the temperature is lowered. In a framework of the general one-dimensional spin diffusion model[@VFmodel; @Takahashi], the spin signals detected by LSVs with transparent interfaces are expressed as follows[@Kimura1; @Hamaya2]. $$\Delta R_{\rm S} = \frac{S_{\rm N}}{S_{\rm inj}S_{\rm det}} \times \frac{\{\frac{ P_F}{(1-P_F^2)} \rho_{\rm F} \lambda_{\rm F}\}^2}{\rho_{\rm N} \lambda_{\rm N} \sinh \left( {d}/{\lambda_{\rm N}} \right)},$$ where $S_{\rm inj}$, $S_{\rm det}$, and $S_{\rm N}$ are the areas of the junctions with a spin injector and a spin detector, the cross section of the Cu(Fe) strip, respectively, and $P_{\rm F}$ is the bulk spin polarization of the ferromagnetic electrode. $\rho_{\rm F}$ and $\lambda_{\rm F}$ are the resistivity and the spin diffusion length of the ferromagnetic electrode, and $\rho_{\rm N}$ and $\lambda_{\rm N}$ are those for the nonmagnetic wire, respectively. Since the experimentally obtained $\Delta R_{\rm S}$ is nearly constant at [*T*]{} $\le$ 20 K, the right-hand side of Eq. (1) should be a constant. Considering the half metallicity of the CFS spin injector and detector[@Wurmehl], we assume that the values of $P_{\rm F}$ of CFS ($P_{\rm CFS}$), $\rho_{\rm CFS}$, and $\lambda_{\rm CFS}$ are constant in the low temperature regime[@Bombor; @Hamaya2]. In addition, as the logarithmic increase in the resistivity of Cu(Fe) is suppressed in the presence of $\Delta\mu$[@Taniyama], $\rho_{\rm Cu(Fe)}$ can be considered to be almost constant at [*T*]{} $\le$ 20 K, leading to a constant $\lambda_{\rm Cu(Fe)}$ because the relationship $\rho_{\rm Cu(Fe)} \times$ $\lambda_{\rm Cu(Fe)} \approx$ constant is fulfilled in general[@Dubois]. The constant $\lambda_{\rm Cu(Fe)}$ is compatible with the description that the degree of spin-flip scattering events is independent of temperature at [*T*]{} $\le$ 20 K and therefore the temperature regime [*T*]{} $<$ [*T*]{}$_{\rm K}$ corresponds to the low temperature Fermi liquid Kondo singlet regime. From the results, we conclude that [*T*]{}$_{\rm K}$ can be directly characterized by measuring the pure spin current transport in a diluted magnetic alloy. As we have seen before, $d$ dependence of $\Delta R_{\rm S}$ results from the change in $\Delta\mu$ in the host Cu at [*T*]{} $\le$ 30 K ($=$ [*T*]{}$_{\rm K}$). We now discuss the effect of $\Delta\mu$ on the Kondo effect in more depth. When $\Delta\mu$ is increased, the denominator of the right-hand side of Eq. (1) does not remain and $\rho_{\rm Cu(Fe)}$ and $\lambda_{\rm Cu(Fe)}$ should be both a function of $\Delta\mu$. To investigate the influence of $\Delta\mu$ on $\lambda_{\rm Cu(Fe)}$, we measured temperature dependent $\Delta R_{\rm S}$ for various spin injection currents ($I$) by using the LSV with $d$= 312 nm, as shown in Fig. 4(a). {width="8.5cm"} With increasing $I$ from 1.0 to 5.0 mA, the magnitude of $\Delta R_{\rm S}$ is remarkably enhanced up to $\sim$1.2 m$\Omega$ only at [*T*]{} $\le$ [*T*]{}$_{\rm max}$ while the temperature independence of $\Delta R_{\rm S}$ at [*T*]{} $\le$ 30 K remains for each $I$ value. Since spin signals are generally observed to decrease with increasing $I$ due to Joule heating for conventional LSVs with a Cu channel (see right axis)[@Kimura1], the enhancement in $\Delta R_{\rm S}$ is clearly an anomalous result. We also note that similar results are observed for samples with different $d$ values as presented in Fig. S1, ensuring that the effect is intrinsic under pure spin current injection conditions. Eq. (1) now provides a rough estimate of $\lambda_{\rm Cu(Fe)}$ for various spin injection currents assuming the acceptable materials parameters; $P_{\rm CFS}\sim$ 0.6[@Kimura1; @Hamaya2], $\lambda_{\rm CFS}=4.0 - 5.0$ nm, $\rho_{\rm Cu(Fe)}=1.80$ $\mu\Omega$cm under the spin injection conditions at 5 K, $\rho_{\rm F}=35.1$ $\mu\Omega$cm which was obtained from our measurement. $\lambda_{\rm Cu(Fe)}$ obtained is shown in Fig. 4(b) as a function of $I$ at 5 K for the LSV with $d =$ 312 nm – see Fig. S2 for more details. As seen in Fig. 4(b), $\lambda_{\rm Cu(Fe)}$ monotonically increases with increasing $I$, which we attribute to the suppression of the Kondo effect, viz., when $\Delta\mu$ in the host Cu becomes more significant by spin injection, the formation of Kondo clouds is suppressed due to less spin-flip scattering. It should be noted that a conventional way of estimating $\lambda_{\rm Cu(Fe)}$[@VFmodel; @Takahashi] using a fit of the one-dimensional spin diffusion model to the $d$ dependence of $\Delta R_{\rm S}$ cannot be applicable any more since $\lambda_{\rm Cu(Fe)}$ is varied for different $d$ values. In fact, we have confirmed that the fitted curves largely deviate from the experimental data. This interesting phenomenon is a consequence of the generation of a large $\Delta\mu$ in the Kondo alloy by using a highly spin polarized spin injector and detector. We comment on the presence of the maximum in the $\Delta R_{\rm S}-T$ curves at $T_{\rm max}$. A recent study using conventional Cu-based LSVs found a similar maximum and suggested that the maximum was associated with the Kondo effect induced by Fe impurities at the Cu/ferromagnet interfaces[@OBrien; @Casanova1; @Batley; @OBrien2]. Our results using Kondo alloy-based LSVs are in qualitative agreement with their claim, and the maximum is very likely related to the Kondo spin-flip scattering for the formation of Kondo clouds[@OBrien]. For our conventional Cu-based LSVs with CFS, however, no maximum was observed as shown on the right axis of Fig. 4(a) although CFS contains Fe element. In our view, our LSVs have the robust CFS/Cu interfaces because of the high quality epitaxial CFS electrodes grown by MBE[@Yamada]. The single crystalline quality of the CFS electrodes efficiently prevents the contamination of Fe in the Cu wire. A similar feature showing no maximum has been observed in our previous work[@Hamaya2]. Also, the fact that $\Delta R_{\rm S}$ is reduced with increasing $I$ due to Joule heating in conventional Cu-based LSVs is in clear contrast to the results for the Kondo alloy based LSVs. In conclusion, we have demonstrated clear evidence for the suppression of the Kondo effect due to a pure spin current in LSV structures with Co$_{2}$FeSi/Cu(Fe) Kondo alloy interfaces. A remarkable reduction in the non-local spin signal has been observed associated with $s$-$d$ spin-flip scattering in the Cu(Fe) Kondo alloy. We have also shown that the non-local spin signal exhibits a plateau below the temperature regime corresponding to the formation of Fermi liquid Kondo singlets. With increasing pure spin current density, the reduction in the non-local spin signal becomes less significant, clearly indicating that $s$-$d$ spin-flip scattering associated with the Kondo effect is efficiently suppressed under greater spin accumulation conditions. From these results, we conclude that the Kondo effect can also be tuned by injecting a pure spin current into a Kondo alloy. This work was partly supported by JSPS KAKENHI Grant Numbers 26103003, 252460200, 26289229, 15H01014. S. O. acknowledges JSPS Research Fellowships for Young Scientists. [99]{} K. Yoshida, [*Theory of Magntism*]{}, (Springer-Verlag Berlin Heidelberg, 1996). J. Kondo, Prog. Theor. Phys. [**32**]{}, 37(1964). J. P. Franck, F. D. Manchester, and D. L. Martin, Proc. Roy. Soc. London [**A263**]{}, 494 (1961). J. A. Applebaum, J. Kondo, Phys. Rev. Lett. [**19**]{}, 906 (1967). J. W. Loram, T. E. Whall, P. J. Ford, Phys. Rev. B [**2**]{}, 857 (1970). P. Monod, Phys. Rev. Lett. [**19**]{}, 1113 (1967). F. D. M. Haldane, Phys. Rev. Lett. [**40**]{}, 416 (1978). D. Goldhaber-Gordon, H. Shtrikman, D. Mahalu, D. Abusch-Magder, U. Meirav, and M. A. Kastner, Nature (London) [**391**]{}, 156 (1998). S. M, Cronenwett, T. H. Oosterkamp, and L. P. Kouwenhoven, Science [**281**]{}, 540 (1998). D. Goldhaber-Gordon, J. Gores, M. A. Kastner, H. Shtrikman, D. Mahalu, and U. Meirav, Phys. Rev. Lett. [**81**]{}, 5225 (1998). W. G. van der Wiel, S. De Franceschi, T. Fujisawa, J. M. Elzerman, S. Tarucha, L. P. Kouwenhoven, Science [**289**]{}, 2105 (2000). M. R. Buitelaar, T. Nussbaumer, and C. Schönenberger, Phys. Rev. Lett. [**89**]{}, 256801 (2002). J. Martinek, Y. Utsumi, H. Imamura, J. Barnaś, S. Maekawa, J. König, and G. Schön, Phys. Rev. Lett. [**91**]{}, 127203 (2003). J. Martinek, M. Sindel, L. Borda, J. Barnaś, J. König, G. Schön, and J. von Delft, Phys. Rev. Lett. [**91**]{}, 247202 (2003). T. Taniyama, N. Fujiwara, Y. Kitamoto, and Y. Yamazaki, Phys. Rev. Lett. [**90**]{}, 016601 (2003). N. J. Craig, J. M. Taylor, E. A. Lester, C. M. Marcus, M. P. Hanson, A. C. Gossard, Science [**304**]{}, 565 (2004). A. N. Pasupathy, R. C. Bialczak, J. Martinek, J. E. Grose, L. A. K. Donev, P. L. McEuen, D. C. Ralph, Science [**306**]{}, 86 (2004). M. S. Choi, D. Sánchez, and R. López, Phys. Rev. Lett. [**92**]{}, 056601 (2004). M. Sato, H. Aikawa, K. Kobayashi, S. Katsumoto, and Y. Iye, Phys. Rev. Lett. [**95**]{}, 066801 (2005). H. B. Heersche, Z. de Groot, J. A. Folk, L. P. Kouwenhoven, H. S. J. van der Zant, A. A. Houck, J. Labaziewicz, and I. L. Chuang, Phys. Rev. Lett. [**96**]{}, 017205 (2006). K. I. Lee, S. J. Joo, J. H. Lee, K. Rhie, Tae-Suk Kim, W. Y. Lee, K. H. Shin, B. C. Lee, P. LeClair, J.-S. Lee, and J.-H. Park, Phys. Rev. Lett. [**98**]{}, 107202 (2007). K. Hamaya, M. Kitabatake, K. Shibata, M. Jung, M. Kawamura, K. Hirakawa, T. Machida, T. Taniyama, S. Ishida, and Y. Arakawa, Appl. Phys. Lett. [**91**]{}, 232105 (2007). C. Buizert, A. Oiwa, K. Shibata, K. Hirakawa, and S. Tarucha, Phys. Rev. Lett. [**99**]{}, 136806 (2007). J. Barnaś and I. Weymann, J. Phys.: Condens. Matter. [**20**]{} 423202 (2008). J. R. Hauptmann, J. Paaske, and P. E. Lindelof, Nature Phys. [**4**]{}, 373 (2008). T. Kobayashi, S. Tsuruta, S. Sasaki, T. Fujisawa, Y. Tokura, and T. Akazaki, Phys. Rev. Lett. [**104**]{}, 036804 (2010). Y. Yamauchi, K. Sekiguchi, K. Chida, T. Arakawa, S. Nakamura, K. Kobayashi, T. Ono, T. Fujii, and R. Sakano, Phys. Rev. Lett. [**106**]{}, 176601 (2011). L. O’Brien, M. J. Erickson, D. Spivak, H. Ambaye, R. J. Goyette, V. Lauter, P. A. Crowell, and C. Leighton, Nature Commun. [**5**]{}, 3927 (2014). J. T. Batley, M. C. Rosamond, M. Ali, E. H. Linfield, G. Burnell, and B. J. Hickey, Phys. Rev. B [**92**]{}, 220420(R) (2015). M. Johnson and R. H. Silsbee, Phys. Rev. Lett. [**55**]{}, 1790 (1985). P. C. van Son, H. van Kempen, and P. Wyder, Phys. Rev. Lett. [**58**]{}, 2271 (1987). F. J. Jedema, A. T. Filip, and B. J. van Wees, Nature [**410**]{}, 345 (2001). T. Kimura and Y. Otani, J. Phys. Condens. Matter [**19**]{}, 165216 (2007). S. Wurmehl, G. H. Fecher, H. C. Kandpal, V. Ksenofontov, C. Felser, H. -J. Lin, and J. Morais, Phys. Rev. B [**72**]{}, 184434 (2005). D. Bombor, C. G. F. Blum, O. Volkonskiy, S. Rodan, S. Wurmehl, C. Hess, and B. Büchner, Phys. Rev. Lett. [**110**]{}, 066601 (2013). K. Hamaya, H. Itoh, O. Nakatsuka, K. Ueda, K. Yamamoto, M. Itakura, T. Taniyama, T. Ono, and M. Miyao, Phys. Rev. Lett. [**102**]{}, 137204 (2009). S. Yamada, K. Hamaya, K. Yamamoto, T. Murakami, K. Mibu and M. Miyao, Appl. Phys. Lett. [**96**]{}, 082511 (2010). T. Kimura, N. Hashimoto, S. Yamada, M. Miyao, and K. Hamaya, NPG Asia Mater. [**4**]{}, e9 (2012). K. Hamaya, N. Hashimoto, S. Oki, S. Yamada, M. Miyao, and T. Kimura, Phys. Rev. B [**85**]{}, 100404(R) (2012). S. Dubois, L. Piraux, J. M. George, K. Ounadjela, J. L. Duvail, A. Fert, Phys. Rev. B [**60**]{}, 477 (1999). T. Valet, A. Fert, Phys. Rev. B [**48**]{}, 7099 (1993). S. Takahashi, S. Maekawa, Phys. Rev. B [**67**]{}, 052409 (2003). E. Villamor, M. Isasa, L. E. Hueso, F. Casanova, Phys. Rev. B [**88**]{}, 184411 (2013). L. O’Brien, D. Spivak, J. S. Jeong, K. A. Mkhoyan, P. A. Crowell, and C. Leighton, Phys. Rev. B [**93**]{}, 014413 (2016).

ArXiv

**UNIVERSIDAD DE BUENOS AIRES** **Facultad de Ciencias Exactas y Naturales** **Departamento de Matemática** **Tesis de Licenciatura** **On the renormalization group of quantum field theory\ (after R. Bocherds)** **Gabriel Alexis Palau** **Director: Estanislao Benito Herscovich Ramoneda**   Agradecimientos {#agradecimientos .unnumbered} =============== Me gustaría empezar esta tesis agradeciendo a todas las personas que de una forma u otra me incentivaron, ayudaron y/o soportaron durante el estudio de esta carrera. Empezando por mi padre Hector Palau, quien en mi infancia siempre que podía me daba algún acertijo sobre la naturaleza, por ejemplo contándome que es posible que las estrellas que uno ve en la noche ya no existan, o enseñándome a calcular la distancia a la que una tormenta se encuentra a partir del sonido de sus truenos. También quiero agradecer a mi madre, Isabel do Eyo quien desde siempre me apoyó e incentivó en este camino. A mis hermanas Cintia y Flavia, por aguantar tanta desatención en ciertos períodos de la carrera, y a mi novia Elim Garcia por lo mismo sobre todo en la etapa final de la misma y por tanta ayuda y apoyo. Agradezco también a mi abuelo Eduardo por sus charlas y a mi abuela Emilia por el apoyo. A mis compañeros y amigos por los ratos compartidos Melanie Bondorevsky, Virginia Gali y los “eje eques” y todos aquellos con quienes compartí momentos gratos. Académicamente le agradezco infinitamente a mi director de tesis Estanislao Herscovich quien estuvo tan presente durante la tesis y me ayudó tanto que nadie adivinaría que nos separaba un océano, por la enorme paciencia y la gran enseñanza que me dejó. También por brindarme el placer de tener con él una que otra charla sobre física. Agradezco también al jurado de esta tesis Gabriel Larotonda y Mauricio Leston por leer y juzgar este trabajo. Introduction {#introduction .unnumbered} ============ The aim of this thesis is to analyze part of the results presented by R.E. Borcherds in his article on renormalization and quantum field theories [@bor]. To facilitate this, the first chapter presents the basic facts on category theory, sheaf theory and the theory of distributions on a manifold. In the particular case of distributions on manifolds we differentiate between the H-distribution (introduced by Hörmander [@hor]) and the D-distributions (described in the book of Dieudonn[é]{} [@di]) and extend the product of distributions defined by Hörmander only for H-distibution, to an action from H-distributions over D-distributions. Each one of these facts will be used along the thesis in order to exhibit more detailed descriptions and explanations for the results. In Chapter 2 the concepts of spacetime, fields, Lagrangians, propagators and Feynman measure are defined. The notion of Feynman measure is introduced with a more detailed explanation of the Gaussian condition and its relation with the cut propagator. In the last chapter we define the renormalization group and immediately characterize it using the results from Chapter 1. The algebraic structure of this group is described in detail. We present a descending filtration of the renormalization group by normal subgroups, compute the corresponding quotients and commutators, and show that any element of the renormalization group can be written as an infinite product of elements each one living in one of the aforementioned subgroups. Lastly we show that an action from the renormalization group over the set of Feynman measures is well defined and prove that this action is transitive over the Feynman measures associated with a given cut local propagator. Preliminary results =================== Along this thesis we will need certain level of knowledge in category theory, sheaf theory and the theory of distributions on manifolds. Since we shall just use some specific result in each of these areas, we shall present them. The reader familiarized with these concepts can omit this chapter and eventually come back next to regard a specific result if needed. Results on category theory -------------------------- ### Basic notions For the basic definitions on category theory, we refer the reader to [@Mariano], Ch.9, or [@chris], Ch.XI.1. Moreover, the definitions of monoidal (or tensor) category, monoidal (or tensor) functor and monoidal natural transformation can be found in [@chris], Ch.XI.2. and XI.4. We shall use the letters $\mathcal{C}$, $\mathcal{D}$,... to denote a monoidal category. If $\mathcal{C}$ is a monoidal category we will denote by $\otimes_{\mathcal{C}}:\mathcal{C}\times\mathcal{C}\rightarrow\mathcal{C}$ the *tensor product* bifunctor and by $\textbf{1}_{\mathcal{C}}$ the unit object. The left and right unit morphisms will be denoted by $l_{M}:\textbf{1}_{\mathcal{C}}\otimes_{\mathcal{C}}M\to M$ and $r_{M}:M\otimes_{\mathcal{C}}\textbf{1}_{\mathcal{C}}\to M$ for any $M\in Obj(\mathcal{C})$; whereas the associativity morphism will be denoted by $a_{A,B,C}:(A\otimes B)\otimes C \to A\otimes (B\otimes C)$ for any $A,B$ and $C$ objects of $\mathcal{C}$. A strict monoidal category is a category where the associativity and unit morphism are the identities of the category, i.e. for all $X$, $Y$ and $Z$ objects of $\mathcal{C}$ we have $(X\otimes Y)\otimes Z=X\otimes(Y\otimes Z)$ and $\textbf{1}_{\mathcal{C}}\otimes X=X=X\otimes\textbf{1}_{\mathcal{C}}$. We present a well-known result about monoidal categories, whose proof can be found in [@chris] Theorem XI.5.3 Any monoidal category is monoidal (or tensor) equivalent to a strict one. This theorem implies Mac Lane´s coherence theorem which states that in a monoidal category all diagrams built with the unit, associativity or identity morphisms commute. Hence from now on we will suppose for simplicity in the exposition that all our monoidal categories are strict. Given a monoidal category $\mathcal{C}$, a triplet $(A,m,u)$ with $A\in Obj(\mathcal{C})$, $m\in Hom_{\mathcal{C}}(A\otimes A,A)$ and $u\in Hom_{\mathcal{C}}(\textbf{1},A)$ is called an *unitary algebra in $\mathcal{C}$* if we have: $$\begin{aligned} & m\circ(m\otimes Id_{A})=m \circ(Id_{A}\otimes m),\\ & m\circ(u\otimes Id_{A})=Id_{A}, \\ & m\circ(Id_{A}\otimes u)=Id_{A}. \end{aligned}$$ $\textbf{1}_{\mathcal{C}}$ with the product $l_{\textbf{1}_{\mathcal{C}}}=r_{\textbf{1}_{\mathcal{C}}}$ (see [@chris], Lemma XI.2.3) and unit $Id_{\textbf{1}_{\mathcal{C}}}$ is a unitary algebra. Notice that the natural domain of $m\otimes Id_{A}$ is $(A\otimes A)\otimes A$ and for $Id_{A}\otimes m$ it is $A\otimes (A\otimes A)$, so the only reason why the first equation of the definition makes sense is because in a strict monoidal category the two domains are equal. A similar observation applies to the second and third equations: if we were not in a strict monoidal category the first of the equations above should take the form $m\circ(m\otimes Id_{A})=m\circ(Id_{A}\otimes m)\circ a_{A,A,A}$, and the same applies to the rest. The morphism between two unitary algebras $(A,m_{A},u_{A})$ and $(B,m_{B},u_{B})$ in $\mathcal{C}$ is a morphism $f\in Hom_{\mathcal{C}}(A,B)$ such that $$\begin{aligned} u_{B}\circ f=u_{A} & & \text{and} & & m_{B}\circ(f\otimes f)=f\otimes m_{A} . \end{aligned}$$ The class of unitary algebras in $\mathcal{C}$ together with the previous morphisms form a category which we note by $Alg_{\mathcal{C}}$. Since we will be interested in unitary algebras, we will drop the adjective unitary and refer to them as algebras. Let $\mathcal{C}$ be a monoidal category and $(A,m,u)$ a unitary algebra in $\mathcal{C}$. A *right module over $A$* is a pair $(M,\rho)$ such that $M\in Obj(\mathcal{C})$ and $\rho \in Hom_{\mathcal{C}}(M\otimes A,M) $ satisfying the following equations $$\begin{aligned} & \rho \circ (Id_{M}\otimes m)=m\circ (\rho \otimes Id_{A}), \\ & \rho \circ (Id_{M}\otimes u)=Id_{M}.\end{aligned}$$ One defines left modules over an algebra $A$ analogously. A *bimodule over $A$* is a triple $(M,\lambda,\rho)$ such that $(M,\lambda)$ is a left module and $(M,\rho)$ is a right module over $A$ and $\rho \circ (\lambda\otimes Id_{A})=\lambda\circ (Id_{A}\otimes \rho)$. Note that $(A,\mu,\mu)$ is a bimodule, called *regular*. A morphism of bimodules is a morphism of left and right modules. They clearly form a category. Let $\mathcal{C}$ be a monoidal category and $(A,m,u)$ a unitary algebra in $\mathcal{C}$. An *ideal* I of $A$ is a subobject of $A$ in the category of bimodules over $(A,m,u)$. Let $A$ be a commutative algebra and $_{A}Mod$ be the category of $A$-modules. It is a symmetric monoidal category for the canonical tensor product over $A$ and the usual braiding denote by $\tau$. A *good category* $\mathcal{C}$ will be a symmetric monoidal category together with a fully faithfull monoidal functor from $\mathcal{C}$ to the symmetric monoidal category $_{A}Mod$, for some $A$. For the following definition we suppose our category satisfies Grothendieck’s axiom $Ab3^{\ast}$ (see [@wei], Appendix for the definition). Given an algebra $A$ in a monoidal category $\mathcal{C}$, and a subobject $X$ of $A$, the *ideal generated* by $X$ is the intersection of the family of ideals $I$ of $A$ such that $X$ is a subobject of $I$ (see [@mitch] Ch.I, section 9). \[coalgebra\] Given a monoidal category $\mathcal{C}$, a *counitary coalgebra in $\mathcal{C}$* is a triplet $(C,\varepsilon_{C},\Delta_{C})$ where $C\in Obj(\mathcal{C})$, $\varepsilon_{C} \in Hom(C,\textbf{1})$, $\Delta_{C}\in Hom(C, C\otimes C)$ such that $$\begin{aligned} &(\Delta_{C}\otimes Id_{C})\circ\Delta_{C}=(Id_{C}\otimes\Delta_{C})\circ \Delta_{C},\\ & Id_{C}=(\varepsilon_{C}\otimes Id_{C})\circ\Delta_{C},\\ & Id_{C}=(Id_{C}\otimes \varepsilon_{C})\circ\Delta_{C}.\end{aligned}$$ The morphism $\varepsilon_{C}$ is called the *counit* of $C$ and $\Delta_{C}$ the *coproduct*. The object $\textbf{1}_{\mathcal{C}}$ together with coproduct $l_{\textbf{1}_{\mathcal{C}}}^{-1}=r_{\textbf{1}_{\mathcal{C}}}^{-1}$ (see previous remark) and counit $Id_{\mathcal{C}}$ is a counitary coalgebra. Since we are mainly interested in counitary coalgebras, we will only refer to them as coalgebras. Similarly to the case of algebras the class of coalgebras in $\mathcal{C}$ constitute a category if we take as morphism from $(C,\varepsilon_{C},\Delta_{C})$ to $(D,\varepsilon_{D},\Delta_{D})$ the maps $g\in Hom_{\mathcal{C}}(C,D)$, satisfying $$\begin{aligned} f\circ \varepsilon_{C}=\varepsilon_{D} & & \text{and} & & (f\otimes f)\circ \Delta_{C}=\Delta_{D}\circ f.\end{aligned}$$ We denote this category by $Coalg_{\mathcal{C}}$. Similarly one defines coalgebras without counit and their morphisms. We denote this category $_{n}Coalg_{\mathcal{C}}$ We refer the reader to [@chris], Ch XIII for the definition of braided category. For each pair of objects $V$ and $W$ of the category $\mathcal{C}$ denote by $c_{V,W}\in Hom(V\otimes W,W\otimes V)$ the *braiding* isomorphism. A symmetric monoidal category is a braided monoidal category such that $c_{V,W}\circ c_{W,V}=Id_{W\otimes V}$ for all objects $V$ and $W$ in the category. Given a symmetric monoidal category $\mathcal{C}$, a unitary algebra $(A,m,u)$ in $\mathcal{C}$ is called a *commutative* if $m=m\circ c_{A,A}$. If $(A,m_{A},u_{A})$ and $(B,m_{B},u_{B})$ are two commutative unitary algebras, a morphism between them will be a map $f\in Hom_{Alg_{\mathcal{C}}}(A,B)$. The class of commutative unitary algebras together with these morphisms forms a category which we note by $Alg^{c}_{\mathcal{C}}$. \[cocomcoalgebra\]Given a symmetric monoidal category $\mathcal{C}$, a cocommutative counitary coalgebra in $\mathcal{C}$ is a coalgebra $(C,\Delta_{C},\varepsilon_{C})$ such that $\Delta_{C}=c_{C,C}\circ \Delta_{C}$. If $(C,\Delta_{C},u_{C})$ and $(D,\Delta_{D},\Delta_{D})$ are two cocommutative counitary coalgebras, a morphism between them will be a map $f\in Hom_{Coalg_{\mathcal{C}}}(C,D)$. The class of cocommutative counitary coalgebras together with these morphisms forms a category which we note by $Coalg^{c}_{\mathcal{C}}$. The basic example of a braided (non-strict) monoidal category is the category of vector spaces over a field. Note that the category of algebras in a braided monoidal category is also braided, as recalled in [@su], Ch.1. Given a monoidal category $\mathcal{C}$ and a coalgebra $(C,\Delta,\varepsilon)$ in $\mathcal{C}$ a *right comodule over $C$* is pair $(M,\delta)$ where $M\in Obj(\mathcal{C})$ and $\delta\in Hom(M,M\otimes C)$ satisfy $$\begin{aligned} \label{224} & (Id_{M}\otimes\Delta)\circ \delta=(\delta\otimes Id_{C})\circ \delta, & (Id_{M}\otimes \varepsilon)\circ \delta=Id_{M}.\end{aligned}$$ The morphism $\delta$ is called the *coaction*. We denote by $Com_{C}$ the category whose objects are right comodules over $C$ and whose morphisms are given as follows. If $(M,\delta_{M})$ and $(N,\delta_{N})$ are two right comodules over $C$, a morphism $f$ from $M$ to $N$ is a map $f\in Hom_{\mathcal{C}}(M,N)$ such that $$\label{1111} \xymatrix{ M\ar[rr]^{f} \ar[d]_{\delta_{M}} & & N \ar[d]^{\delta_{N}}\\ M\otimes C \ar[rr]^{f\otimes Id_{C}}& & N\otimes C }$$ is commutative. Since we shall work only with right comodules, we shall refer to them just as comodules. \[prop 1\] Let $\mathcal{C}$ be a monoidal category, $W\in Obj(\mathcal{C})$ and $(C,\Delta_{C},\varepsilon_{C})$ a coalgebra in $\mathcal{C}$. Then $W\otimes C$ has structure of right comodule over $C$, with coaction $\delta:=Id_{W}\otimes \Delta_{C}:W\otimes C \rightarrow (W\otimes C)\otimes C $. To see that this is a right comodule over $C$, one must check equation . First we will show that the following diagram $$\xymatrix{ W\otimes C \ar[rr]^{\delta} \ar[d]_{\delta} & & (W\otimes C)\otimes C \ar[d]^{Id_{W\otimes C}\otimes \Delta_{C}} \\ (W\otimes C)\otimes C \ar[rr]_{\delta\otimes Id_{C}} & & (W\otimes C)\otimes C \otimes C }$$ commutes. Indeed, we have the following chain of identities $$\begin{aligned} & (\delta\otimes Id_{C})\circ\delta=((Id_{W}\otimes \Delta_{C})\otimes Id_{C})\circ(Id_{W}\otimes \Delta_{C}) \hspace{2cm}\\ & =(Id_{W}\otimes (\Delta_{C}\otimes Id_{C}))\circ(Id_{W}\otimes \Delta_{C})=(Id_{W}\circ Id_{W})\otimes ((\Delta_{C}\otimes Id_{C})\circ \Delta_{C}),\end{aligned}$$ where we have used the strictness of the category in the second equality and the property $(f\otimes g)\circ(f'\otimes g')=(f\circ f')\otimes (g\circ g')$. Moreover, $$\begin{aligned} & (\delta\otimes Id_{C})\circ\delta=(Id_{W}\circ Id_{W})\otimes (( Id_{C}\otimes \Delta_{C})\circ \Delta_{C})\\ & =(Id_{W}\otimes(Id_{C}\otimes \Delta_{C}))\circ(Id_{W}\otimes \Delta_{C}) =((Id_{W}\otimes Id_{C})\otimes \Delta_{C})\circ (Id_{W}\otimes \Delta_{C})\\ & =(Id_{W\otimes C}\otimes \Delta_{C})\circ(Id_{W}\otimes \Delta_{C}) =(Id_{W\otimes C}\otimes \Delta_{C})\circ \delta,\end{aligned}$$ where in the last row we have used that $Id_{W\otimes C}=Id_{W}\otimes Id_{C}$. Finally we must show the commutation of the following diagram $$\xymatrix{ W\otimes C \ar[rr]^{\delta} \ar[rrd]_{Id_{W\otimes C}} & & (W\otimes C)\otimes C \ar[d]^{Id_{W\otimes C}\otimes \varepsilon_{C}} \\ & & (W\otimes C)\otimes \textbf{1}_{\mathcal{C}} }$$ Indeed, we have the following chain of identities $$\begin{aligned} & (Id_{W\otimes C}\otimes \varepsilon_{C})\circ \delta=(Id_{W\otimes C}\otimes \varepsilon_{C})\circ (Id_{W}\otimes \Delta_{C})\\ & =((Id_{W}\otimes Id_{C})\otimes \varepsilon_{C})\circ (Id_{W}\otimes \Delta_{C}) = (Id_{W}\otimes (Id_{C}\otimes \varepsilon_{C}))\circ (Id_{W}\otimes \Delta_{C}) \\ & =(Id_{W}\circ Id_{W})\otimes ((Id_{C}\otimes \varepsilon_{C})\circ \Delta_{C})=Id_{W}\otimes Id_{C}=Id_{W\otimes C}, \end{aligned}$$ where we have used the strictness of the category in the third equality. Given a (respectively symmetric) monoidal category $\mathcal{C}$, a *(respectively cocommutative) coaugmented coalgebra in $\mathcal{C}$* is a cocommutative coalgebra $(C,\Delta, \varepsilon)$ in $\mathcal{C}$ provided with a morphism of (respectively cocommutative) coalgebras $\eta\in Hom(\textbf{1}_{\mathcal{C}},C)$ such that $$\begin{aligned} \label{conmutacion1} %\xymatrix{ %\textbf{1}\otimes \textbf{1} \ar[r]^{\eta \otimes \eta} & C \otimes C \\ % \textbf{1} \ar[u]^{\iota} \ar[r]_{\eta} & C \ar[u]_{\Delta_{C}}\\ %}& & \xymatrix{ \textbf{1}\ar[rd]_{Id_{\textbf{1}}} \ar[rr]^{\eta} & & C \ar[dl]^{\varepsilon_{C}} \\ & \textbf{1} & \\ } \end{aligned}$$ Given two cocommutative coaugmented coalgebras $(C,\Delta_{C},\varepsilon_{C},\eta_{C})$ and\ $(D,\Delta_{D},\varepsilon_{D},\eta_{D})$, a morphism of cocommutative coaugmented coalgebras from $C$ to $D$ is an element $f\in Hom_{Coalg_{\mathcal{C}}}(C,D)$ such that $f\circ \eta_{C}=\eta_{D}$. The class of all cocommutative coaugmented coalgebras in $\mathcal{C}$ with morphisms between them form a category for the usual composition and identity morphisms, that we denote by $_{c}Coalg_{\mathcal{C}}$. If $V$ is a $K$-vector space, let $TV=\oplus_{n \geqslant 0} V^{\otimes n}$ its tensor construction. We define a counit morphism as the map $\varepsilon:TV\to K$ given by the projection on the zero degree component and a coproduct as follows. If $v=v_{1}\otimes \cdots \otimes v_{n}$ is an element of $TV$ then $$\Delta(v)=\sum_{i=1}^{n-1}v_{1}\cdots v_{i}\otimes v_{i+1}\cdots v_{n} +1_{K}\otimes v +v \otimes 1_{K}.$$ We also define $\Delta(1_{K})=1\otimes 1$. The coproduct is a linear extension of the previous definitions. The coaugmentation morphism is the map $\eta:K\to TV$ given by $\eta(1_{K})=1\in T^{0}V$. It is easy to verify that $(TV,\Delta,\varepsilon,\eta)$ is a coaugmented coalgebra in the symmetric monoidal category of vector spaces over the field $K$. ### Relation between coaugmented and noncounitary coalgebras in an abelian category {#sec111} All along this section $\mathcal{C}$ will be an abelian symmetric monoidal category. We shall define two functors $F:{_{c}Coalg_{\mathcal{C}}}\to {_{n}Coalg_{\mathcal{C}}}$ and $G:{_{n}Coalg_{\mathcal{C}}}\to {_{c}Coalg_{\mathcal{C}}}$ which are quasi-inverses of each other. If $(C,\Delta_{C},\varepsilon_{C},\eta_{C})$ is a coaugmented coalgebra in $\mathcal{C}$, we will denote $(\overline{C},\pi_{C})=Coker(\eta_{C})$. The universal property of the cokernel gives us a map $\Delta_{\overline{C}}$ such that the following diagram $$\label{diag1} \xymatrix{ & & & \overline{C}\otimes \overline{C} \\ \textbf{1} \ar_{\eta_{C}}[r] & C \ar_{\pi_{C}}[rr] \ar^{(\pi_{C}\otimes\pi_{C})\circ \Delta_{C}}[rru] & & \overline{C} \ar@{-->}_{\Delta_{\overline{C}}}[u] }$$ commutes. The morphism $\Delta_{\overline{C}}$ is the only one such that the diagram commutes, and it exists because $(\pi_{C} \otimes \pi_{C})\circ \Delta_{C}\circ \eta_{C}=0$. Indeed by the definition of cocommutative coaugmented coalgebra we have that $(\pi_{C} \otimes \pi_{C})\circ \Delta_{C}\circ \eta_{C}=(\pi_{C} \otimes \pi_{C})\circ (\eta_{C} \otimes \eta_{C}) \circ \iota=((\pi_{C}\circ \eta_{C})\otimes(\pi_{C}\circ \eta_{C}))\circ \iota $, which is zero by definition of the cokernel of $\eta_{C}$. If $(C,\Delta_{C},\varepsilon_{C},\eta_{C})\in obj({_{c}Coalg_{\mathcal{C}}})$ we define $F(C,\Delta_{C},\varepsilon_{C},\eta_{C})=(\overline{C},\Delta_{\overline{C}})$. If $f:(C,\Delta_{C},\varepsilon_{C},\eta_{C}) \rightarrow (D,\Delta_{D},\varepsilon_{D},\eta_{D})$ is a morphism in the category ${_{c}Coalg_{\mathcal{C}}}$, we see that there exists a unique map in the category $_{n}Coalg_{\mathcal{C}}$, which we call $\overline{f}$, satisfying that $$\label{diag1.5} \xymatrix{ & & & \overline{D} \\ \textbf{1} \ar_{\eta_{C}}[r] & C \ar_{\pi_{C}}[rr] \ar^{\pi_{D}\circ f}[rru] & & \overline{C} \ar@{-->}_{\overline{f}}[u] }$$ In order to guarantee the existence of a mapping $\overline{f}$ in $\mathcal{C}$ making commute, we must only see that $(\pi_{D}\circ f) \circ \eta_{C}=0$. By definition of $Hom_{{_{c}Coalg_{\mathcal{C}}}}(C,D)$ we have that $(\pi_{D}\circ f) \circ \eta_{C}=\pi_{D}\circ (f \circ \eta_{C})=\pi_{D}\circ \eta_{D}$ which is zero because $(\overline{D},\pi_{D})$ is the cokernel of $\eta_{D}$. In order to see that $\overline{f}\in Hom_{_{n}Coalg_{\mathcal{C}}}(\overline{C},\overline{D})$, it is enough to prove the diagram $$\xymatrix{ \overline{C} \ar[d]_{\Delta_{\overline{C}}}\ar[rr]^{\overline{f}}& & \overline{D} \ar[d]^{\Delta_{\overline{D}}} \\ \overline{C}\otimes\overline{C} \ar[rr]_{\overline{f}\otimes \overline{f}}& & \overline{D}\otimes \overline{D} \\ }$$ commutes. We will prove the equality $\Delta_{\overline{D}}\circ \overline{f}=(\overline{f}\otimes \overline{f}) \circ \Delta_{\overline{C}}$ by using that $\pi_{C}:C\to \overline{C}$ is an epimorphism, because it is a cokernel. So it is enough proving $\pi_{\overline{D}}\circ \overline{f}\circ \pi_{C}=(\overline{f}\otimes \overline{f}) \circ \Delta_{\overline{C}}\circ \pi_{C}$. Indeed, $$\begin{aligned} & \Delta_{\overline{D}}\circ \overline{f}\circ \pi_{C}=\Delta_{\overline{D}}\circ \pi_{D} \circ f=(\pi_{D}\otimes \pi_{D})\circ \Delta_{D}\circ f\\ & =(\pi_{D}\otimes \pi_{D})\circ (f\otimes f)\circ \Delta_{C}=(\overline{f}\otimes \overline{f})\circ (\pi_{C}\otimes \pi_{C})\circ \Delta_{C}\\ & =(\overline{f}\otimes \overline{f})\circ \Delta_{\overline{C}}\circ \pi_{C} \end{aligned}$$ where we have used that $\overline{f}\circ \pi_{C}=\pi_{D} \circ f$ in the first equality and the fact that the morphisms of coalgebras satisfy $\Delta_{D}\circ f= (f\otimes f)\circ \Delta_{C}$ in the third one together with common properties of the morphisms of monoidal categories. If $f \in Hom_{_{c}Coalg_{\mathcal{C}}}(C,D)$ we define $F(f)=\overline{f}\in Hom_{_{n}Coalg_{\mathcal{C}}}(\overline{C},\overline{D})$. To prove that this assignment is a functor we must see that - $F(g\circ f)=F(g)\circ F(f)$, - $F(Id_{C})=Id_{\overline{C}}$. In the following we suppose that $f\in Hom_{_{c}Coalg_{\mathcal{C}}}(C,D)$ and $g\in Hom_{_{c}Coalg_{\mathcal{C}}}(D,E)$. We know that $F(g\circ f)$ is the only morphism such that $F(g\circ f)\circ \pi_{C}=\pi_{E}\circ(g\circ f)$. Since $$\begin{aligned} F(g)\circ F(f)\circ\pi_{C}=F(g)\circ \overline{f}\circ \pi_{C} = \overline{g} \circ \pi_{D}\circ f=\pi_{E}\circ g\circ f, \end{aligned}$$ then, by uniqueness we conclude that $F(g\circ f)=F(g)\circ F(f)$. Diagram implies that $F(Id_{C})=Id_{\overline{C}}$. Then $F:{_{c}Coalg_{\mathcal{C}}}\to {_{n}Coalg_{\mathcal{C}}}$ is a functor. Moreover; we will prove that it is a categorical equivalence, and in order to see that we will construct its quasi-inverse functor $G:{_{n}Coalg_{\mathcal{C}}}\to {_{c}Coalg_{\mathcal{C}}}$. Given an object $(E,\Delta_{E})$ in the category ${_{n}Coalg_{\mathcal{C}}}$, we define its image under $G$ by $(E\oplus \textbf{1}_{\mathcal{C}},\Delta_{E\oplus \textbf{1}_{\mathcal{C}}},\pi_{2},i_{2})$, where $\pi_{2}:E\oplus \textbf{1}_{\mathcal{C}}\to \textbf{1}_{\mathcal{C}}$ is the canonical projection of the direct sum and $i_{2}:\textbf{1}_{\mathcal{C}}\to E\oplus \textbf{1}_{\mathcal{C}}$ is the canonical inclusion map of the direct sum. The definition of $\Delta_{E\oplus \textbf{1}_{\mathcal{C}}}$ is a little more involved. We write $\textbf{1}$ instead of $\textbf{1}_{\mathcal{C}}$ to denote the unit object. We first define the morphisms $$\begin{aligned} & \tilde{\Delta}_{E}=(i_{1}\otimes i_{1})\circ \Delta_{E}+(i_{1}\otimes i_{2})\circ r_{E}^{-1}+(i_{2}\otimes i_{1})\circ l_{E}^{-1}, \\ & \Delta_{\textbf{1}}=(i_{2}\otimes i_{2})\circ \iota^{-1} \end{aligned}$$ Hence, there is a unique morphism $$\Delta_{E\oplus \textbf{1}}\in Hom_{\mathcal{C}}(E \oplus \textbf{1},(E\oplus \textbf{1})\otimes (E\oplus\textbf{1}))$$ such that $\Delta_{E\oplus \textbf{1}}\otimes i_{1}=\tilde{\Delta}_{E}$ and $\Delta_{E\oplus \textbf{1}}\otimes i_{2}=\Delta_{\textbf{1}}$, i.e. making the following diagram $$\xymatrix{ & & & \textbf{1} \ar^{i_{2}}[d] \ar_{\Delta_{\textbf{1}}}[llld] \\ (E\oplus \textbf{1})\otimes (E\oplus \textbf{1}) & & & E\oplus 1 \ar@{-->}_{\hspace{1.2cm}\exists!\hspace{0.1cm} \Delta_{E\oplus \textbf{1}}}[lll] \\ & & & E \ar^{\tilde{\Delta}_{E}}[lllu] \ar_{i_{1}}[u] \\ }$$ commute. From now on, for simplicity we shall write $E_{+}=E\oplus \textbf{1}$. The next step is to define $G$ on morphisms. If $f$ is in $Hom_{{_{n}Coalg_{\mathcal{C}}}}(C,D)$, then we define $G(f)=i_{1}^{D}\circ f\circ \pi_{1}^{C}+i_{2}^{D}\circ\pi_{2}^{C}$, where $i_{1}^{D}:D\to D_{+}$, and $\pi_{D_{+}}:D_{+}\to \overline{(D_{+})}$ is the cokernel of $i_{2}^{D}:\textbf{1}\to D_{+}$ (i.e. the coaugmentation of $D_{+}$). We will denote $G(f)$ by $f_{+}$. If we are in the category of vector spaces over a field $K$, the previous constructions are of the form:   - $G(E)=E\oplus \textbf{1}$ will actually be $E\oplus K\cdot 1_{E_{+}}$, - The coaugmentation $\eta(1_{K})=1_{E_{+}}$, - $\Delta_{E\oplus K}(e)=\Delta_{E}(e)+1_{E_{+}}\otimes e+e\otimes 1_{E_{+}}$, - If $f\in Hom_{_{n}Coalg_{\mathcal{C}}}(E,H)$, then $G(f)(e+\lambda \cdot 1_{E_{+}})=f(e)+\lambda \cdot 1_{H_{+}}$. \[teo1\]Let $\mathcal{C}$ be a monoidal abelian category. Then, the functor $$F:{_{c}Coalg_{\mathcal{C}}}\to {_{n}Coalg_{\mathcal{C}}}$$ is a categorical equivalence whose quasi-inverse functor is $$G:{_{n}Coalg_{\mathcal{C}}}\to {_{c}Coalg_{\mathcal{C}}}.$$ If $X\in Obj(_{n}Coalg_{\mathcal{C}})$ we construct the natural map $N_{X}:X\to \overline{(X_{+})}$ given by the composition $N_{X}=\pi_{X_{+}}\circ i_{1}^{X}$. Given two objects $X, Y\in Obj(_{n}Coalg_{\mathcal{C}})$ and $f\in Hom_{_{n}Coalg_{\mathcal{C}}}(X,Y)$, we will show that $$\xymatrix{ X \ar[rr]_{f} \ar[d]_{N_{X}} & & Y \ar[d]^{N_{Y}}\\ \overline{(X_{+})}\ar[rr]^{\overline{(f_{+})}}& & \overline{(Y_{+})}}$$ is commutative. In order to do that, note that $$\begin{aligned} & f_{+}\circ i_{1}^{X}=(i_{1}^{Y}\circ f \circ \pi_{1}^{X} + i_{2}^{Y}\circ \pi_{2}^{X})\circ i_{1}^{X}\\ & =i_{1}^{Y}\circ f \circ \pi_{1}^{X}\circ i_{1}^{X} + i_{2}^{Y}\circ \pi_{2}^{X}\circ i_{1}^{X}\\ & = i_{1}^{Y}\circ f \circ Id_{X} + i_{2}^{Y}\circ 0 =i_{1}^{Y}\circ f,\end{aligned}$$ where we have used the bilinearity of the composition in an abelian category in the second equality and the facts that $\pi_{1}^{X}\circ i_{1}^{X}=Id_{X}$ and $\pi_{2}^{X}\circ i_{1}^{X}=0$ in the third one. Hence, $$N_{Y}\circ f=\pi_{Y_{+}}\circ i_{1}^{Y}\circ f=\pi_{Y_{+}}\circ f_{+}\circ i_{1}^{X}=\overline{(f_{+})} \circ \pi_{X_{+}}\circ i_{1}^{X}= \overline{(f_{+})} \circ N_{X}$$ where in the third equality we use the commutation of the diagram and in the last one the definition of the natural transformation. Then the functor $F\circ G$ is naturally isomorphic to $Id_{_{n}Coalg_{\mathcal{C}}}$. Also it is possible to construct a natural map between $Id_{\mathcal{C}_{_{c}Coalg_{\mathcal{C}}}}$ and $G\circ F$. Suppose $(X,\Delta_{X},\varepsilon_{X},\eta_{X})\in Obj(_{c}Coalg_{\mathcal{C}})$, we define $L_{X}:X\to (\overline{X})_{+}$ by $L_{X}=i_{1}^{\overline{X}}\circ \pi_{X}+i_{2}^{X}\circ \varepsilon_{X}$, where $i_{1}^{\overline{X}}:\overline{X}\to (\overline{X})_{+}$ and $i_{2}^{X}:\textbf{1} \to (\overline{X})_{+}$ are the inclusion maps, $\pi_{X}:X\to \overline{X}$ is the cokernel of $\eta_{X}$ and $\pi_{1}^{\overline{X}}:(\overline{X})_{+}\to \overline{X}$ and $\pi_{1}^{\overline{X}}:(\overline{X})_{+}\to \textbf{1}$ are the projections. Note that $$\begin{aligned} & (\overline{f})_{+}\circ L_{X}=(i_{1}^{\overline{Y}}\circ \overline{f}\circ \pi_{1}^{\overline{X}}+i_{2}^{Y}\circ \pi_{2}^{\overline{X}})\circ (i_{1}^{\overline{X}}\circ \pi_{X}+i_{2}^{X}\circ \varepsilon_{X})\\ & =i_{1}^{\overline{Y}}\circ \overline{f}\circ \pi_{1}^{\overline{X}}\circ i_{1}^{\overline{X}}\circ \pi_{X}+ i_{1}^{\overline{Y}}\circ \overline{f}\circ \pi_{1}^{\overline{X}}\circ i_{2}^{X}\circ \varepsilon_{X}\\ & +i_{2}^{Y}\circ \pi_{2}^{\overline{X}}\circ i_{1}^{\overline{X}}\circ \pi_{X} +i_{2}^{Y}\circ \pi_{2}^{\overline{X}}\circ i_{2}^{X}\circ \varepsilon_{X}, \end{aligned}$$ where we have used the bilinearity of the composition, the definition of the functor $G$ and the natural map $N_{X}$ defined above. Using that $\pi_{1}^{\overline{X}}\circ i_{1}^{\overline{X}}=Id_{\overline{X}}$, $\pi_{1}^{\overline{X}}\circ i_{2}^{X}=0$, $\pi_{2}^{\overline{X}}\circ i_{1}^{\overline{X}}=0$ and $\pi_{2}^{\overline{X}}\circ i_{2}^{\overline{X}}=Id_{\textbf{1}}$, we get that $$(\overline{f})_{+}\circ L_{X}=i_{1}^{\overline{Y}}\circ \overline{f}\circ \pi_{X}+i_{2}^{Y}\circ \varepsilon_{X}.$$ On the other hand, the computation of $N_{Y}\circ f$ gives us $$\begin{aligned} & N_{Y}\circ f= (i_{1}^{\overline{Y}}\circ \pi_{Y}+ i_{2}^{Y}\circ \varepsilon_{Y})\circ f=i_{1}^{\overline{Y}}\circ \pi_{Y}\circ f+ i_{2}^{Y}\circ \varepsilon_{Y} \circ f\\ & =i_{1}^{\overline{Y}}\circ \overline{f}\circ \pi_{X} + i_{2}^{Y}\circ \varepsilon_{X}, \end{aligned}$$ where we have used the definition of $\overline{f}$ and the fact that $f$ is a morphism of (counitary) coalgebras in the last equality. Hence $N_{Y}\circ f=(\overline{f})_{+}\circ L_{X}$ which together with the previous result implies that $F$ and $G$ are quasi-inverse functors.   Let us denote $\Delta^{(1)}=Id_{\overline{C}}$, $\Delta_{\overline{C}}^{(2)}=\Delta_{\overline{C}}$ and $\Delta_{\overline{C}}^{(n)}=(\Delta_{\overline{C}}\otimes Id^{\otimes (n-2)})\circ \Delta_{\overline{C}}^{(n-1)}$ if $n\geqslant 3$. Notice that there is a canonical morphism $Ker\Delta_{\overline{C}}^{(n)} \hookrightarrow Ker\Delta_{\overline{C}}^{(n+1)}$ for all $n\geqslant 2$. \[cocom\] A (respectively cocommutative) coaugmented coalgebra $(C,\Delta,\varepsilon, \eta)$ is *cocomplete (or conilpotent)* if $\displaystyle{\overline{C}=colim_{n\to \infty}Ker \Delta_{\overline{C}}^{(n)}}$. ### Symmetric coalgebras Let us begin this section with an example. \[SV\] Let $\mathcal{C}$ be a good symmetric monoidal category. Given $V\in Obj(\mathcal{C})$, let us consider the *tensor algebra* $$T(V)=\bigoplus_{n\in \mathbb{N}_{0}}V^{\otimes n},$$ where $V^{\otimes 0}=A$ and the product is given by concatenation. The algebra $S(V)$ is defined by $S(V)=T(V)/I$, where $I$ is the ideal in $T(V)$ generated by the image of $Id_{V}^{\otimes 2}-\tau_{V,V}$ as a subobject of $V^{\otimes 2}$ (where $\tau$ is the braiding given by the usual twist). We write $$S(V)=\bigoplus_{n\in \mathbb{N}_{0}}S^{n}V.$$ We will define a structure of cocommutative coaugmented coalgebra on it, denoted by $S^{c}V$. The coproduct is defined as follows. We set $\Delta(1_{S^{c}V})=1_{S^{c}V}\otimes 1_{S^{c}V}$, and for an element of the form $v_{1}\cdots v_{n}$ with $v_{i}\in V$ we define $$\label{coproductoSV} \Delta(v_{1}\cdots v_{n})=\sum_{I,J/I\sqcup J=\{1\cdots n\}}v_{I}\otimes_{A} v_{J},$$ where $v_{I}=v_{i_{1}}\cdots v_{i_{k}}$ if $I=\{ i_{1}<\cdots <i_{k} \}\neq \emptyset$ and $v_{\emptyset}=1_{S^{c}V}$ The counit $\varepsilon:S^{c}(V)\rightarrow A$ is given by the canonical projection onto the zeroeth component, and the coaugmentation $\eta:A\to S^{c}(V)$ is the canonical inclusion. From now on, we denote this coalgebra only by $SV$, unless we say the opposite. Note that $\overline{SV}$ defined in 1.1.2. is just $\overline{SV}=\oplus_{n\in \mathbb{N}} S^{n}V$ and $\Delta_{\overline{SV}}$ is $$\label{copSV} \Delta_{\overline{SV}}(v_{1}\cdots v_{n})=\sum_{I,J\neq \emptyset /I\sqcup J=\{1\cdots n\}}v_{I}\otimes v_{J}$$ for $n\geqslant 2$ and is zero over $V$. Notice that $SV$ is cocomplete. To prove that, it is sufficient note that $\Delta_{\overline{SV}}^{(n)}$ vanishes on elements of $\oplus_{k=1}^{n}S^{k}V$. \[symme\] Note that the construction of the coalgebra $SV$ can be done for any object $V$ in a symmetric monoidal category $\mathcal{C}$ with arbitrary direct sums such that $\otimes_{\mathcal{C}}$ commutes with $\oplus$. The Theorem 1 in Chapter XI.1 from [@mac] give us a canonical action from $\mathcal{S}_{n}$ to the tensor construction of degree $n$, $TV$. This theorem allow us to construct the symmetric coalgebra as equivalence classes of this action. Let us denote by $\pi_{I}:S^{n}V\to S^{|I|}V$ the projection map in the symmetric product with component in $I\subset\{1,\cdots,n\}$. Then we can consider the maps $\Delta_{n}: S^{n}V\to SV \otimes SV$ given by $$\Delta_{n}=\bigoplus_{I,J | I\sqcup J=\{ 1,\cdots,n\} } \pi_{I}\otimes \pi_{J},$$ and define the coproduct $\Delta:SV\to SV \otimes SV$ for $SV$ as $\Delta=\bigoplus_{n\in \mathbb{N}_{0}}\Delta_{n}$. However in [@mac] there is another result in Chapter XI, establishing that there is a strong monoidal functor between the tensor algebra construction and a free algebra in the category algebras in a monoidal category. This result is equally valid for the case of coalgebras. If $C$ is a coaugmented coalgebra in $\mathcal{C}$, we shall denote by $1_{C}$ the image $\eta(1_{K})$. Note it is a distinguished *group-like* element of $C$. \[prop3\]Given a symmetric monoidal category $\mathcal{C}$ with arbitrary direct sums such that $\otimes_{\mathcal{C}}$ commutes with $\oplus$, for any $(C,\Delta_{C},\varepsilon_{C},\eta_{C})\in {_{c}Coalg_{\mathcal{C}}^{c}}$, any $V\in Obj(\mathcal{C})$ and $p:C\rightarrow V$ morphism in $Hom_{\mathcal{C}}(C,V)$ such that $p\circ \eta_{C}=0$, there exists a unique morphism $P\in Hom_{_{c}Coalg_{\mathcal{C}}^{c}}(C,SV)$ such that the following diagram $$\xymatrix{ C \ar@{-->}[rd]_{P} \ar[rr]^{p}& & V \\ & SV\ar[ru]_{\pi} & \\ }$$ commutes, where $\pi$ is the canonical projection to the elements of degree one. The reader may suppose that $\mathcal{C}$ is a good category. If he or she is willing to, but the proof is exactly the same. If $C$ is as in the statement, by Theorem \[teo1\] we can write $C\simeq\overline{C} \oplus A$. So, it suffices to define $P$ by its restrictions to each direct summand. We first set $P$ such that $P\circ \eta_{C}=\eta_{SV}$. In the sequel we shall denote $p^{\odot \alpha}:\overline{C}^{\otimes \alpha}\to S^{\alpha}V$ to denote the symmetric product of the map $p:\overline{C}\to V$, i.e. $p^{\otimes l}$ composed with the quotient projection $T^{n}V\to S^{n}V$. Moreover, we define $\overline{P}=P|_{\overline{C}}$ by $$\overline{P}=\sum_{n\in \mathbb{N}}\frac{1}{n!}(p^{\odot n}\circ \Delta_{\overline{C}}^{(n)}),$$ where we are committing the abuse of calling $p$ the morphism $\overline{C}\to V$ induced by $p:C\to V$. This induced map exists because of the universal property of the cokernel since $p\circ \eta_{C}=0$, as depicted below $$\xymatrix{ & & & & V \\ A \ar[rr]_{\eta_{C}} & & C \ar[rr]_{\pi_{C}} \ar[rru]^{p} & & \overline{C} \ar@{-->}[u]_{\exists! p} }$$ In order to see that $\overline{P}$ is a coalgebra morphism we must prove the commutation of $$\xymatrix{ \overline{C} \ar[rr]^{\overline{P}} \ar[d]_{\Delta_{\overline{C}}} & & \overline{SV}\ar[d]^{\Delta_{ \overline{SV}}} \\ \overline{C}\otimes \overline{C}\ar[rr]_{\overline{P}\otimes\overline{P}} & & \overline{SV}\otimes \overline{SV} \\}$$ Changing some parentheses in the firsts equalities and using the definition of $\overline{P}$ and the equation in the fourth equality, we have $$\begin{aligned} & \Delta_{\overline{SV}}\circ \overline{P}=\Delta_{\overline{SV}}\circ(\sum_{n\in \mathbb{N}}\frac{1}{n!}(p^{\odot n}\circ \Delta_{\overline{C}}^{(n)}))=\sum_{n\in \mathbb{N}}\frac{1}{n!}\Delta_{\overline{SV}}\circ(p^{\odot n}\circ \Delta_{\overline{C}}^{(n)})\\ & =\sum_{n\in \mathbb{N}}\frac{1}{n!} (\Delta_{\overline{SV}}\circ p^{\odot n})\circ \Delta_{\overline{C}}^{(n)} = \sum_{n\in \mathbb{N}}\frac{1}{n!}\sum_{I,J\neq \emptyset /I\sqcup J=\{1\cdots n\}}(p^{\odot |I|}\otimes p^{\odot |J|})\circ \Delta_{\overline{C}}^{(n)}.\\ \end{aligned}$$ By rewriting $\Delta^{(n)}_{\overline{C}}$ and using that in a monoidal category the morphisms satisfy $(f\otimes g)\circ (h\otimes r)=(f\circ h)\otimes (g\circ r)$, we see that the latter sum coincides with $$\begin{aligned} & \sum_{n\in \mathbb{N}}\frac{1}{n!}\sum_{I,J\neq \emptyset /I\sqcup J=\{1\cdots n\}}(p^{\odot |I|}\otimes p^{\odot |J|})\circ (\Delta_{\overline{C}}\otimes Id)^{|I|-1}\circ(Id\otimes \Delta_{\overline{C}})^{|J|-1}\circ \Delta_{\overline{C}}\\ & =\sum_{n\in \mathbb{N}}\frac{1}{n!}\sum_{I,J\neq \emptyset /I\sqcup J=\{1\cdots n\}}(p^{\odot |I|}\circ \Delta_{\overline{C}}^{(|I|)} \otimes p^{\odot |J|})\circ(Id\otimes \Delta_{\overline{C}})^{|J|-1}\circ \Delta_{\overline{C}}\\ & =\sum_{n\in \mathbb{N}}\frac{1}{n!}\sum_{I,J\neq \emptyset /I\sqcup J=\{1\cdots n\}}(p^{\odot |I|}\circ \Delta_{\overline{C}}^{(|I|)} \otimes p^{\odot |J|}\circ \Delta_{\overline{C}}^{(|J|)})\circ \Delta_{\overline{C}}\\ & =\sum_{n\in \mathbb{N}}\frac{1}{n!}\sum_{I\subseteq \{1\cdots n\} / 1<|I|<n }(p^{\odot |I|}\circ \Delta_{\overline{C}}^{(|I|)} \otimes p^{\odot n-|I|}\circ \Delta_{\overline{C}}^{(n-|I|)})\circ \Delta_{\overline{C}}. \end{aligned}$$ Since $\odot$ is a symmetric product, any term who has $p^{\odot t}\circ \Delta_{\overline{C}}^{(t)}\otimes p^{\odot n-t}\circ \Delta_{\overline{C}}^{(n-t)}$ is equal to any other who has $p^{\odot r}\circ\Delta_{\overline{C}}^{(r)} \otimes p^{\odot n-r}\circ \Delta_{\overline{C}}^{(n-r)}$ if and only if $t=r$. Therefore for a fixed $r\in \{1,...,n-1\}$ there are $\binom{n}{r}$ terms equal to $p^{\odot r}\circ\Delta_{\overline{C}}^{(r)} \otimes p^{\odot n-r}\circ \Delta_{\overline{C}}^{(n-r)}$. Hence, $$\begin{aligned} & =\sum_{n\in \mathbb{N}}\frac{1}{n!}\sum_{1<|I|<n}\binom{n}{|I|}(p^{\odot |I|}\circ \Delta_{\overline{C}}^{(|I|)} \otimes p^{\odot n-|I|}\circ \Delta_{\overline{C}}^{(n-|I|)})\circ \Delta_{\overline{C}}\\ & =\sum_{n\in \mathbb{N}}\sum_{1<|I|<n}\frac{1}{|I|!}\frac{1}{(n-|I|)!} (p^{\odot |I|}\circ \Delta_{\overline{C}}^{(|I|)} \otimes p^{\odot n-|I|}\circ \Delta_{\overline{C}}^{(n-|I|)})\circ \Delta_{\overline{C}}\\ & =[\sum_{n\in \mathbb{N}}\sum_{1<k<n}\frac{1}{k!}\frac{1}{(n-k)!} (p^{\odot k}\circ \Delta_{\overline{C}}^{(k)} \otimes p^{\odot n-k}\circ \Delta_{\overline{C}}^{(n-k)})]\circ \Delta_{\overline{C}}\\ & =(\sum_{r\in \mathbb{N}}\frac{1}{r!}(p^{\odot r}\circ \Delta_{\overline{C}}^{(r)})\otimes \sum_{s\in \mathbb{N}}\frac{1}{s!}(p^{\odot s}\circ \Delta_{\overline{C}}^{(s)}))\circ \Delta_{\overline{C}}\\ & =(\overline{P}\otimes \overline{P})\circ \Delta_{\overline{C}}, \end{aligned}$$ which is exactly the commutation we want. \[correspondencia1\]The previous proposition implies that there is a bijection $$Hom_{_{_{c}Coalg_{\mathcal{C}}^{c}}}(C,SV) \leftrightsquigarrow \{f\in Hom_{_{n}Coalg_{\mathcal{C}}}(C,V) /f\circ \eta_{C}=0\}$$ for all cocommutative coaugmented coalgebra $(C,\Delta_{C},\varepsilon_{C},\eta_{C})\in {_{c}Coalg_{\mathcal{C}}^{c}}$ in any symmetric monoidal category $\mathcal{C}$. Since $\overline{P}$ is the restriction of $P$, we shall omit the bar in order to simplify the notation. Given $V\in Obj(\mathcal{C})$, we specialize Proposition \[prop3\] in the case $C=SV$. To every $P\in Hom_{_{_{c}Coalg_{\mathcal{C}}^{c}}}(SV,SV)$ it corresponds a unique $p=\pi\circ P|_{\overline{SV}}\in Hom_{\mathcal{C}}(\overline{SV},V)$ that will be represented by a denumerable family of maps $\pi\circ P|_{S^{n}V}$, which we denote $\{p_{n}\}_{n\in \mathbb{N}}$ and call it the *sequential representation* of $P$. If we apply the formula $P|_{\overline{SV}}=\sum_{n\in \mathbb{N}}\frac{1}{n!}(p^{\odot n}\circ \Delta_{SV}^{(n)})$ in this case, we get $$\label{3.33} P|_{S^{n}V}(v_{1}\cdots v_{n}) =\sum_{m=1}^{n}\sum_{I_{1}\cdots I_{m}\neq \emptyset; I_{1}\sqcup \cdots \sqcup I_{m}=\{1\cdots n\}}\frac{1}{m!}p_{|I_{1}|}(v_{1})\odot ... \odot p_{|I_{m}|}(v_{I_{m}})$$ where $|X|$ is the cardinality of the set $X$. Note that the sequential representation of $\pi\circ Id_{SV}$ is $\{Id_{V},0,0,\cdots \}$.\ \[prop4\] Let $\mathcal{C}$ be a symmetric monoidal category with arbitrary direct sums such that $\otimes_{\mathcal{C}}$ commutes with $\oplus$, the morphisms $P\in Hom_{_{c}Coalg_{\mathcal{C}}^{c}}(SV, SV)$ are represented by maps $ p\in Hom_{\mathcal{C}}(\overline{SV},V)$. Moreover $P$ is an isomorphism if and only if $p_{1}$ is an isomorphism, where we use the notation of sequential representation introduced previously. As before, the reader may assume that $\mathcal{C}$ is a good category if he or she is willing to, but the proof applies analogously to the general case. It is sufficient to prove the second statement, for the first is a consequence of Remark \[correspondencia1\]. We will prove that $P:SV\rightarrow SV$ is an isomorphism if and only if $\pi \circ P|_{S^{1}V}=p_{1}$ is an isomorphism. Consider $P, Q\in Hom_{_{c}Coalg_{\mathcal{C}}^{c}}(SV,SV)$. By formula , we obtain that $\pi \circ P\circ Q|_{\overline{SV}} = \sum_{n\in \mathbb{N}}h_{n}$ where $$\label{composicion} h_{n}(v_{1}\cdot v_{2}\cdots v_{n})=\sum_{m=1}^{n}\frac{1}{m!}\sum_{I_{1}\cdots I_{m}\neq \emptyset; I_{1}\sqcup \cdots \sqcup I_{m}=\{1\cdots n\} }p_{m}(q_{|I_{1}|}(v_{I_{1}})\cdots q_{|I_{m}|}(v_{I_{m}})).$$ Suppose that $P$ is bijective and take $Q=P^{-1}$. Then for $n=1$ implies that $q_{1}$ is the inverse of $p_{1}$. On the other hand assume that $p_{1}$ is bijective. We will prove that $P$ is bijective. In order to do so, we will exhibit a left and right inverses of $P$. Define recursively the family $\{q_{n}\}_{n\in \mathbb{N}}$ where $q_{n}\in Hom_{\mathcal{C}}(S^{n}V,V)$ is given as follows. Set $q_{1}=p_{1}^{-1}$ and if we have defined all the $q_{m}$’s with $m < n$ then $$\label{3.35} q_{n}(w_{1}\cdot w_{2}\cdots w_{n})=-\sum_{m=1}^{n-1}\sum_{I_{1}\cdots I_{m}\neq \emptyset ; I_{1}\sqcup \cdots \sqcup I_{m}=\{1\cdots n\} }q_{m}(p_{|I_{1}|}(\tilde{w}_{I_{1}})\cdots p_{|I_{m}|}(\tilde{w}_{I_{m}}))$$ for $n\geqslant 2$,where $\tilde{w}_{I_{k}}=q_{1}^{\otimes |I_{k}|}(w_{I_{k}})$. Let $Q:SV\to SV$ be the morphism of coalgebras whose sequential representation is $\{q_{n}\}_{n\in \mathbb{N}}$. Then by means of one can see that the sequential representation of $\pi\circ (Q\circ P)$ is $\{ Id_{V},0, \cdots \} $ and thus by uniqueness of the correspondence in Proposition \[prop3\] we get that $Q\circ P=Id_{\overline{S^{c}V}}$ and consequently $P$ is injective. Let us now prove the surjectivity of $P$. Define a family of morphisms $\{\tilde{q}_{n}\}_{n\in \mathbb{N}}$, where $\tilde{q}_{n}\in Hom_{\mathcal{C}}(S^{n}V,V) $ is defined as follows. Set $\tilde{q}_{1}=p_{1}^{-1}$ and if $\tilde{q}_{1},\cdots \tilde{q}_{n-1}$ for $n\in \mathbb{N}_{\geqslant 2}$ are defined, we fix $$\label{inv} \tilde{q}_{n}(w_{1}\cdot w_{2}\cdots w_{n})=-\sum_{m=2}^{n}\frac{1}{m!}\sum_{I_{1}\cdots I_{m}\neq \emptyset ; I_{1}\sqcup \cdots \sqcup I_{m}=\{1\cdots n\} }\tilde{q}_{1}(p_{m}(\tilde{q}_{|I_{1}|}(w_{I_{1}})\cdots \tilde{q}_{|I_{m}|}(w_{I_{m}})) )$$ for $n\geqslant 2$. This sequential representation defines a unique $\tilde{Q}\in Hom_{_{c}Coalg_{\mathcal{C}}^{c}}(SV,SV)$ for which $P\circ \tilde{Q}=Id_{SV}$. Then, $P$ is surjective. As a consequence, $P$ is bijective if and only if $p_{1}$ is bijective. \[prop5\]Let $\mathcal{C}$ be a monoidal abelian category, $(D,\Delta_{D},\varepsilon_{D})$ a coalgebra in $\mathcal{C}$ and $M$ a $D$-comodule in $\mathcal{C}$. If $W\in Obj(\mathcal{C})$, then there is a one to one correspondence $$\label{ddd} \fbox{ \xymatrix{ Hom_{Com_{D}}(M,W\otimes D) \ar[r] & Hom_{\mathcal{C}}(M,W) \ar[l] }}$$ given by $f\mapsto (Id_{W}\otimes \varepsilon_{D})\circ f$, where $W\otimes D$ is the right $D$ comodule described in Proposition \[prop 1\]. The inverse is given by sending $g$ to $\tilde{g}=(g\otimes Id_{D})\circ \rho_{M}$, where $\rho_{M}$ is the coaction of $M$. Let us show that if $h\in Hom_{\mathcal{C}}(M,W)$ then the morphism $\tilde{h}:M\to W\otimes D$ is an isomorphism of $D$-comodules, i.e. satisfies that $(\tilde{h}\otimes Id_{D})\circ \rho_{M}=\rho_{W\otimes D}\circ \tilde{h}$, where $\rho_{W\otimes D}$ is the coaction of $W\otimes D$. Indeed, we have the equations $$(\tilde{h}\otimes Id_{D})\circ \rho_{M}=(((h\otimes Id_{D})\circ \rho_{M})\otimes Id_{D})\circ \rho_{M}$$ $$\label{ygyg} =((h\otimes Id_{D})\otimes Id_{D})\circ ( \rho_{M}\otimes Id_{D} ) \circ \rho_{M}$$ $$=(h\otimes (Id_{D}\otimes Id_{D}))\circ ( \rho_{M}\otimes Id_{D} ) \circ \rho_{M},$$ where we have used the definition of $\tilde{h}$ in this first equality. Moreover, since $Id_{D}\otimes Id_{D}=Id_{D\otimes D}$, the latter member of coincides with $$\begin{aligned} &(h\otimes Id_{D\otimes D})\circ ( Id_{M}\otimes \Delta_{D} ) \circ \rho_{M}=(h\circ Id_{M}\otimes Id_{D\otimes D}\circ \Delta_{D}) \circ \rho_{M}\\ & =(h\otimes \Delta_{D})\circ \rho_{M}=(Id_{W}\otimes \Delta_{D})\circ (h\otimes Id_{D}) \circ \rho_{M}\\ & =\rho_{W\otimes D}\circ(h \otimes Id_{D})\circ \rho_{M}=\rho_{W\otimes D}\circ \tilde{h}, \end{aligned}$$ where we have used equation in the first equality. This proves that the map $g\mapsto \tilde{g}$ is well defined. To see that the correspondence is biunivocal take $f\in Hom_{Com_{D}}(M,W\otimes D)$. If we apply to $f$ we get $(Id_{W}\otimes \varepsilon_{D})\circ f$, and the allowed inverse gives us $$\begin{aligned} & (((Id_{W}\otimes \varepsilon_{D})\circ f)\otimes Id_{D})\circ \rho_{M}=((Id_{W}\otimes \varepsilon_{D})\otimes Id_{D})\circ (f\otimes Id_{D})\circ \rho_{M}\\ & =((Id_{W}\otimes \varepsilon_{D})\otimes Id_{D})\circ \rho_{W\otimes D}\circ f=(Id_{W}\otimes( \varepsilon_{D}\otimes Id_{D}))\circ (Id_{W}\otimes \Delta_{D}) \circ f\\ & =(Id_{W}\circ Id_{W}\otimes (\varepsilon_{D}\otimes Id_{D})\circ \Delta_{D})\circ f= (Id_{W}\otimes Id_{D})\circ f = Id_{W\otimes D}\circ f=f, \end{aligned}$$ where we have used the property $(f\otimes g)\circ(f'\otimes g')=(f\circ f')\otimes (g\circ g')$ in the first step, the fact that $f$ commutes with the coactions in the second equality, the definition of the coaction of $W\otimes D$ and the strictness of the category in the third equality and lastly the definition of coalgebra by mean of equation $(\varepsilon_{D}\otimes Id_{D})\circ \Delta_{D}=Id_{D}$ and the property $(f\otimes g)\circ(f'\otimes g')=(f\circ f')\otimes (g\circ g')$ again. The category of sheaves and the topology of its sections -------------------------------------------------------- For the basic definitions and general references about sheaf theory we refer the reader to [@te] chapters 1 to 4. Given a topological space $\mathcal{M}$ and a ringed space $(\mathcal{M},\mathcal{O})$, we shall denote by $_{\mathcal{O}}Mod$ the category of sheaves of $\mathcal{O}$-modules. If $F,G\in {_{\mathcal{O}}Mod}$, $F\otimes_{\mathcal{O}} G $ stands for the tensor product in the category $_{\mathcal{O}}Mod$. If $F\in {_{\mathcal{O}}Mod}$, we denote by $\tau_{\mathcal{M}}$ the topology of $\mathcal{M}$ and by $r_{V\subseteq U}:F(U)\to F(V)$ the corresponding restriction map. $\Gamma(F)$ will denote the sections of the sheaf and, $\Gamma_{c}(F)$ will be sections of compact support. Given two sheaves of $\mathcal{O}$-modules $F$ and $G$ and a morphism $f$ from $F$ to $G$, we shall denote by $f_{U}:F(U)\to G(U)$ the family of maps indexed by the open sets $U$ in $\mathcal{M}$ that form the morphism $f$. If $\mathcal{M}$ is a smooth manifold, and $\tau_{\mathcal{M}}$ its topology, then the assignment $\mathcal{C}^{\mathbb{R}}:\tau_{\mathcal{M}}\rightarrow {_{\mathbb{R}}Alg} $ such that $\mathcal{C}^{\mathbb{R}}(U)=\mathcal{C}^{\infty}(U)=\mathcal{C}^{\infty}(U,\mathbb{R})$ is an example of sheaf of $\mathbb{R}$-algebras, where the restrictions maps are the usual restrictions of functions. Moreover the pair $(\mathcal{M}, \mathcal{C}^{\mathbb{R}})$ satisfy the definition of ringed space of $\mathbb{R}$-algebras. We refer the reader to [@hir] to general references about vector bundles, adapted coordinated system and related concepts. Let $p:E\to \mathcal{M}$ be a complex finite dimensional vector bundle over a smooth manifold $\mathcal{M}$. Let us denote by $\Phi$ the assignment sending each $U\subseteq \mathcal{M}$ open to $\Phi(U)=\Gamma(U,E)=\{ \sigma\in \mathcal{C}^{\infty}(U,E) / p\circ \sigma=Id_{U} \}$. It is clear that $\Phi(U)$ is a module over $\mathcal{C}^{\infty}(U)$. Let be $U,V\in \tau_{\mathcal{M}}$ such that $V\subseteq U$. We define restriction maps $\Phi(U)\to \Phi(V)$ by $r_{V\subseteq U}(\sigma)=\sigma|_{V}$ where $\sigma\in \Phi(U)$. It is clear that they satisfies the axiom of presheaf. Moreover the restriction map $r_{V\subseteq U}:\Phi(U)\rightarrow \Phi(V)$ is a $\mathcal{C}^{\infty}(U)$-linear morphism, because it is linear and if $f\in \mathcal{C}^{\infty}(U)$ and $\sigma \in \Phi(U)$, then $r_{V\subseteq U}(f\sigma)=(f\sigma)|_{V}=f|_{V}\sigma|_{V}=r_{V\subseteq U}(f)r_{V\subseteq U}(\sigma)$, where $r_{V\subseteq U}(f)$ are the restriction maps of the aforementioned sheaf $\mathcal{C}^{\mathbb{R}}$. This proves that $\Phi$ is a presheaf of modules over the ringed space $(\mathcal{M},\mathcal{C}^{\mathbb{R}})$. Furthermore, the following is also true, whose proof is immediate. \[lemshe\] Let $\mathcal{M}$ be a smooth manifold and $p:E\to \mathcal{M}$ a finite dimensional complex vector bundle over $\mathcal{M}$. Then the above mentioned assignment\ $\Phi:\tau_{\mathcal{M}}\to {_{\mathbb{R}}Vect}$ is a sheaf of $\mathcal{C}^{\mathbb{R}}$-modules. ### Sheaf of jets and sheaf of densities {#jets} Consider a finite dimensional complex vector bundle $(E,p,\mathcal{M})$ and let $q\in \mathcal{M}$. Two local sections $\phi, \psi\in \Gamma_{q}(p)$ around $q$ are said to be *k-equivalent at $q$* if $\phi(q)=\psi(q)$ and if in some adapted coordinate system $(x^{i},u^{\alpha})$ around $\phi(p)$, we have that $$\frac{\partial^{|I|} \phi^{\alpha}}{\partial x^{I}}|_{q}=\frac{\partial^{|I|} \psi^{\alpha}}{\partial x^{I}}|_{q}$$ for each $1\leqslant |I|\leqslant k$ and $1\leqslant \alpha \leqslant Rank(E)$. This is an equivalence relation on the local sections around $q$, and the equivalent class containing $\phi$ is called the *k-jet of $\phi$ at $q$* and is denoted by $j^{k}_{q}\phi$ Given $q\in \mathcal{M}$, define the set of *k-th jets of $p$* at $q$ $$J^{k}p=\{j^{k}_{q}\phi \hspace{.2cm} | \hspace{.2cm} q\in \mathcal{M}\text{ and } \phi \text{ is a local section of }p\text{ arround } q \}$$ The *k-th jet manifold of $p$* is the disjoint union $\sqcup_{q\in \mathcal{M}}J_{q}^{k}p$. Lots of properties of this set can be found in [@sa]. The most important to us are the following two. \[See [@sa], Proposition 4.1.7\] $J^{k}p$ is a smooth finite dimensional manifold for any $k\in \mathbb{N}_{0}$. For the manifold $J^{k}p$ we define the projection $\pi:J^{k}p\to \mathcal{M}$ such that $\pi(j^{k}_{q}\phi)=q$. \[See [@sa], Proposition 6.2.13\] \[lema5\] Let $(E,p,\mathcal{M})$ be a finite dimensional vector bundle. Then $(J^{k}p,\pi_{k},\mathcal{M})$ is a finite dimensional vector bundle for $k\in \mathbb{N}_{0}$. For a fixed $k\in \mathbb{N}$, the last lemma enables us to consider sections of the finite dimensional vector bundle $(J^{k}p,\pi_{k},\mathcal{M})$ of jets of order $k$ which by Lemma \[lemshe\] is a sheaf of modules over the ringed space $(\mathcal{M}, \mathcal{C}^{\mathbb{R}})$. We shall denote this sheaf by $J^{k}\Phi$. Analogously one define the jets of infinite order over a fibre bundle $p:E\to \mathcal{M}$ (see [@sa], Chapter 7) which we denote by $J^{\infty}\Phi$ and also conform a vector bundle over $\mathcal{M}$, whose projection is called $\pi$, the only difference is that the fibres are infinite dimensional. All along this thesis when we write $J\Phi$ we will refer to $J^{k}$ for $k\in \mathbb{N}\cup \{ \infty \}$. The category of sheaves over a sheaf of commutative algebras is a symmetric monoidal category which satisfies $Ab3^{\ast}$ (even $Ab4^{\ast}$) axioms and $\otimes_{\mathcal{O}}$ commutes with $\oplus$ (see [@te], Chapter 4). Therefore, given any sheaf $F\in {_{\mathcal{O}}Mod}$, we define $SF$ the symmetric sheaf of $F$ as in Remark \[symme\]. We shall also use the notion of *sheaf of densities* of a smooth manifold. The reader is refereed to [@bo], Chapter 7, or [@ma], Chapter 1, for the definition and basic properties. Given a smooth manifold $\mathcal{M}$, we shall denote by $\omega$ the sheaf of $1$-densities over $\mathcal{M}$. We recall it is the sheaf of sections of a complex vector bundle over $\mathcal{M}$, so it has structure of $\mathcal{C}^{\mathbb{R}}$-module in the natural way.   Given a vector bundle $p:E\to \mathcal{M}$ and we have just defined the sheaf $J\Phi$ of jets and its symmetric sheaf $SJ\Phi$, also we have the sheaf of densities $\omega$ and both have structure of $\mathcal{C}^{\mathbb{R}}$-modules. Then we consider its tensor product $\omega\otimes_{\mathcal{C}^{\mathbb{R}}}SJ\Phi$ in the category of $\mathcal{C}^{\mathbb{R}}$-modules and denote it only by $\omega SJ\Phi$. ### Topology on the spaces of sections We refer the reader to [@ru] or [@he] for general references about locally convex spaces (LCS). We will now describe the topology of several spaces of sections of sheaves. We begin with $\Gamma_{c} \omega SJ\Phi$, following the steps of [@di] Chapter 17, Section 2. Let $\Omega$ be an open set of $\mathbb{R}^{n}$ and consider a fundamental sequence of compact sets $\{K_{m}\}_{m\in \mathbb{N}}$, i.e. $K_{m}\subseteq K_{m+1}^{\circ}$ for all $m\in \mathbb{N}$ and $\bigcup_{m\in\mathbb{N}}K_{m}= \Omega$. Denote by $\mathcal{C}_{m}^{\infty}(\Omega)$ the subspace of $\mathcal{C}^{\infty}(\Omega)$ formed by functions whose support is contained in $K_{m}$. One can define a family of *seminorms* on $\mathcal{C}^{\infty}(\Omega)$ by $$p_{m}(f)=\max \{ |D^{\alpha}f(x)|/x\in K_{m}, |\alpha|\leqslant m \}$$ that turns $\mathcal{C}^{\infty}(\Omega)$ into a *Fréchet* space (see [@ru], 1.46). It is easy to prove that $\mathcal{C}_{c}^{\infty}(\Omega)$ is a closed subspace of $\mathcal{C}^{\infty}(\Omega)$. We regard $\mathcal{C}_{m}^{\infty}(\Omega)$ as a topological subspace of $\mathcal{C}^{\infty}(\Omega)$ with its Fréchet topology. Finally we endow $\mathcal{C}^{\infty}_{c}(\Omega)$ with the final LCS topology from the family of inclusions $\{i_{m}:\mathcal{C}_{m}^{\infty}(\Omega) \rightarrow \mathcal{C}^{\infty}_{c}(\Omega)\}$. One can see that with this topology on $\mathcal{C}^{\infty}_{c}(\Omega)$ is a complete LCS. The space $\mathcal{C}_{c}^{\infty}(\Omega)$ is an example of *LF-space*, and we shall topologize the space of compact supported sections in a very similar way as a LF-space. We will now recall the topology on $\Gamma_{c} \omega SJ\Phi$. We will use the following diagram $$\label{diag} \xymatrix{ \omega S^{l}J\Phi \ar_{\tau}@/_4mm/[d] & \tau^{-1}(U_{\alpha}) \ar_{inc}[l] \ar^{\omega_{\alpha}^{l}}[r] \ar_{\tau |}@/_4mm/[d] & x_{\alpha}(U_{\alpha})\times S^{l}E \ar^{\pi_{1}}[ddl] \ar^{\pi_{2}}[d] \\ \mathcal{M} \ar_{u}@/_4mm/[u] & U_{\alpha} \ar^{x_{\alpha}}[d] \ar_{u|}@/_4mm/[u] \ar^{inc}[l] & S^{l}E \ar_{i_{l}}[r] & SE \\ \mathbb{R}^{n} & x_\alpha(U_{\alpha}) \ar^{inc}[l] }$$ where $\tau$ is the projection of the vector bundle $\omega S^{l}J\Phi$ and $\omega_{\alpha}^{l}$ is a local trivialization of such bundle over the domain $U_{\alpha}$ for $l\in \mathbb{N}$. By reducing the domains if necessary one can always think that the domains of the charts are trivializants for the vector bundle. #### Topology for the fibres of the vector bundle $J\Phi$ We began for define a topology on the fibres of the vector bundle $J\Phi$. This fibres are typically the space $\prod_{k\in \mathbb{N}_{0}}L^{k}_{sym}(\mathbb{C}^{n},\mathbb{C}^{m})$ where $L_{sym}^{k}$ denotes the space of symmetric $k$-multilinear maps from $\mathbb{C}^{n}$ to $\mathbb{C}^{m}$. Let us denote this fibre by $E$, as we did in the diagram (\[diag\]). In [@le], Lewis introduced a family of seminorms $\{\lambda_{r}\}_{r\in \mathbb{N}}$ on the space $\prod_{k\in \mathbb{N}_{0}}L^{k}_{sym}(\mathbb{C}^{n},\mathbb{C}^{m})$ which turns it into a Frechét space. More precisely if $A\in \prod_{k\in \mathbb{N}_{0}}L^{k}_{sym}(\mathbb{C}^{n},\mathbb{C}^{m})$ and we write $A=\prod_{k\in \mathbb{N}_{0}}A_{k}$ then $\lambda_{r}$ is given by $$\lambda_{r}(A)=\max \{ ||A_{0}||_{0}',||A_{1}||_{1}',\cdots, ||A_{r}||_{r}' \},$$ where $||-||_{j}'$ are norms on $L^{j}_{sym}(\mathbb{C}^{n},\mathbb{C}^{m})$ and we recall that, if $M\in L^{j}_{sym}(\mathbb{C}^{n},\mathbb{C}^{m})$, then $$||M||_{j}'=\sup\{ ||M(v,\cdots,v)|| / v\in \mathbb{C}^{n}, ||v||=1 \}.$$ #### Possibles topologies for the tensor product of LCS {#1.2.2} We shall use the seminorms $\lambda_{r}$ to define the *projective topology* over $T^{2}E$ following which was done at [@he] Chapter 3, Section 6. Given two LCS $F$ and $E$, the projective topology for $F\otimes E$ is the finest LCS topology such that the universal mapping $\chi: F\times E \to F \otimes E$ is continuous. We recall that in a LCS there is a correspondence between the Minkowski functionals and the zero open neighbourhoods. Indeed, given $U$ an open neighbourhood, set $p_{U}(x)=\inf \{t\in \mathbb{R}_{>0}/ x\in tU \}$, called the *associated Minkowski functional*. Conversely, given a Minkowski functional $p$, then $p^{-1}([0,1))$ is an open neighbourhood of zero. Given two LCS $E$ and $F$, the projective topology of $E\otimes F$ can be explicitly defined by its Minkowski functionals. If $\chi:E\times F\rightarrow E\otimes F$ is the universal mapping, then a base of zero neighbourhoods for $E\otimes F$ is the balanced convex hull of $$\{ \chi(U\times V) / U \text{ is a open neighbourhood of zero in } E \text{ and V in }F \}$$ which is denoted by $CH(U\otimes V)$. Moreover if $p_{U}$ is the Minkowski functional associated to $U\subseteq E$ and $p_{V}$ of $V\subseteq F$, then $$\label{semi} p_{CH(U\otimes V)}(u)=\inf \{ \sum_{i}p_{U}(x_{i})p_{V}(y_{i}) / u=\sum_{i} x_{i}\otimes y_{i}\}$$ is a seminorms on $E\otimes F$ which coincide with the Minkowski functional associated to $CH(U\otimes V)$ (see result 6.3. from [@he], chapter 3, section 6.). We denote it by $p_{U}\otimes p_{V}$. We denote by $E \otimes_{\pi} F$ the algebraic tensor product with the projective topology just described and call it the projective tensor product between $E$ and $F$. If $E$ and $F$ are LCS then also is $E \otimes_{\pi} F$. But with this topology if $E$ and $F$ are Fréchet spaces then $E \otimes_{\pi} F$ is not necessarily a Fréchet space and the same holds for $LF$ (inductive strict limit of Fréchet, whose principal example are the compactly supported sections over a manifold). Denoting by $E \hat{\otimes}_{\pi} F$ the completion of $E \otimes_{\pi} F$, we obtain a categorical definition of tensor product in the category of Fréchet spaces or LF. In both cases, if $E$ is a LCS by recursive applications of the LCS structure on the tensor products give us a family of countable seminorms on $T^{l}E$ for $l\in \mathbb{N}$ (that may be is a Fréchet topology, depending if $E$ is Fréchet and if we take $\otimes_{\pi}$ or $\hat{\otimes}_{\pi}$). We note them by $||-||_{T^{l}E,\beta}$ with $\beta \in \mathbb{N}_{0}$. #### Topologies for the symmetric product $S^{l}E$ {#jhjhjh} Let us consider the linear inclusion of $S^{l}E$ inside $T^{l}E$ given by the symmetrization map $x_{1}\cdots x_{l}\mapsto \sum_{\sigma\in \mathbb{S}_{l}} x_{\sigma(1)}\otimes\cdots\otimes x_{\sigma(l)}$. It is easy to show that it is a closed subspace of $T^{l}E$. Then, we may consider the induced LCS topology on $S^{l}E$ by means of this inclusion. We shall denote it by $(T^{l}E)^{\mathbb{S}_{l}}$. If $E$ is a Fréchet space, note that $T^{l}E$ is also a Fréchet space and the same holds for the closed subspace $(T^{l}E)^{\mathbb{S}_{l}}$, provided we are using the completion of the projective topology for the algebraic tensor product. On the other hand, since $S^{l}E$ is a quotient of $T^{l}E$, the LCS topology on the latter induces a LCS topology on the former (see [@top], section 3). If $E$ is a Fréchet space, then $S^{l}E$ is also. Note that the composition $(T^{l}E)^{\mathbb{S}_{l}}\hookrightarrow T^{l}E \twoheadrightarrow S^{l}E$ is clearly linear and continuous, because the inclusion is linear and continuous by definition of subspace topology at $(T^{l}E)^{\mathbb{S}_{l}}$ and the quotient projection is clearly linear and continuous. Also, one can see that the composition is bijective. If $E$ is Fréchet the open map theorem for Fréchet spaces (see [@ru] chapter 1), implies that the inverse map of this composition is also continuous and as a conclusion $S^{l}E$ and $(T^{l}E)^{\mathbb{S}}$ are homeomorphic. Note that the open map theorem can be used only if the LCS involved are Fréchet spaces and it is true only if we use the tensor product $\hat{\otimes}_{\pi}$. #### Adding the fibre of the density bundle Continuing the construction of the topology for the section space given the vector bundle $\tau: \omega S^{l}J\Phi \rightarrow \mathcal{M} $ we denote by $\omega_{\alpha}^{l}:\tau^{-1}(U_{\alpha})\rightarrow x_{\alpha}(U_{\alpha}) \times S^{l}E$ as its local trivialization. Notice that this vector bundle has the same local fibre as $S^{l}J\Phi$, because the tensor product on the fibres are over $\mathbb{R}$ and $\omega$ is a line bundle, the typical fibre of $S^{l}J\Phi$ is $S^{l}E$ and for $\omega S^{l}J\Phi$ is $\mathbb{R} \otimes_{\mathbb{R}}S^{l}E$. #### Topology on $\Gamma \omega SJ\Phi$ and $\Gamma_{c} \omega SJ\Phi$ For each chart $(U_{\alpha},x_{\alpha})$ of the aforementioned atlas of $\mathcal{M}$ we consider a fundamental sequence of compacts $ \{K_{m}^{\alpha}\}_{m\in \mathbb{N}}$ and denote by $\Gamma_{m} \omega S^{l}J\Phi$ the sections with compact support contained in $K_{m}$. Given $n\in \mathbb{N}$, define the following family of seminorms on $\Gamma_{m}\omega S^{l}J\Phi$ (see [@di], chapter XV, section 2) defined as $$\begin{aligned} & p_{s,\beta,\alpha,m}(u)=\sup _{x\in K_{m}\subseteq U_{\alpha}}\sum_{|j|\leqslant s,j\in \mathbb{N}_{o}^{n}}||\partial^{j}(\pi_{2}\circ \omega_{\alpha}^{l}\circ u \circ x_{\alpha}^{-1})(x)||_{S^{l}E,\beta}, \end{aligned}$$ where $s$ indicates the maximum order of derivation, $\alpha$ is the index of the chart and $\beta$ the index of the seminorm of $S^{l}E$. Note that the (partial) derivatives appearing in the previous definition are the classical ones (of any function of the form $f:W\subseteq \mathbb{R}^{n}\to L$, where $L$ is a Fréchet space). We finally consider the final LCS topology on $\Gamma_{c} \omega S^{l}J\Phi$ given by the family of inclusions $i_{m,l}:\Gamma_{m} \omega S^{l}J\Phi \hookrightarrow \Gamma_{c} \omega S^{l}J\Phi $, and on $\Gamma_{c} \omega SJ\Phi$ with the LCS topology of the direct sum, i.e. the final LCS topology given by the inclusions $i_{l}:\Gamma_{c} \omega S^{l}J\Phi\rightarrow \Gamma_{c} \omega SJ\Phi$. With this final topology $\Gamma_{c} \omega S^{l}J\Phi$ is a LF-space. Distributions on manifolds -------------------------- ### H-distributions & D-distributions We refer the reader to [@hor], [@di], [@ma] and [@mich] for general references about the theory of distributions on manifolds. During this section $X$ and $Y$ will be open sets of $\mathbb{R}^{n}$, and $\mathcal{M}$ will be a manifold of dimension $n$. We denote by $\mathcal{D}'(X)$ the space of distributions on $X\subseteq \mathbb{R}^{n}$ (see [@hor] for a definition). Finally we denote by $H_{d}(\mathcal{M})$ to the distributions with density character $0$ on $\mathcal{M}$ and by $D_{d}(\mathcal{M})$ to the distributions with *density character* $1$ over $\mathcal{M}$ (see [@mich] for a reference). We will call these two types of distributions H-distributions and D-distributions respectively and shall present a formal definition later in this subsection. We recall that a subset $V$ of $\mathbb{R}^{n}$ is said to be *conic* if for each $\xi\in V$ and $t>0$, $t\xi$ belongs to $V$. If $\mathcal{E}'(\mathbb{R}^{n})$ are the distributions of compact support then for any $u\in\mathcal{E}'(\mathbb{R}^{n})$ we define the set $\Sigma(u)\subseteq \mathbb{R}^{n}\setminus \{ \vec{0} \}$ as the points having no conic neighbourhood $V\subseteq \mathbb{R}^{n}\setminus \{ \vec{0} \}$ and positive constants $C_{N}$ such that $$\begin{aligned} & |\hat{u}(\xi)|\leqslant C_{N}(1+|\xi|)^{-N} \end{aligned}$$ holds for all $\xi \in V$ and all $N\in \mathbb{N}$, where $\hat{u}$ is the Fourier transform of $u$. With this definition $\Sigma(u)$ is clearly a closed cone in $\mathbb{R}^{n}\setminus \{ \vec{0} \}$. For any $u\in \mathcal{D}'(X)$ and $\psi \in \mathcal{C}_{c}(X)$, the distribution $\psi u$ belongs to $\mathcal{E}'(X)$. Given $u\in \mathcal{D}'(X)$, for each $x\in X$ we set $$\begin{aligned} \Sigma_{x}(u)=\bigcap \{ \Sigma(\phi u) / \phi\in \mathcal{C}_{0}^{\infty}(X), \phi(x)\neq 0 \} \end{aligned}$$ and call it the *cone of $u$ at $x$*. For a nice exposition of these concepts, and together with a definition of the wavefront set of a distribution on $\mathbb{R}^{n}$ and also the definition of wavefront set for an H-distribution on a manifold, see [@ma], Chapter 2. The objective of this section is to generalise the product of H-distributions defined by L.Hörmander in [@hor] (see [@ma], Chapter 2.3) to an action of the space of H-distributions over the space of D-distributions. It will be very useful to describe $H_{d}(\mathcal{M})$ and $D_{d}(\mathcal{M})$ as families of distributions in open sets of $\mathbb{R}^{n}$ satisfying some sort of covariance, which we briefly recall. If $f:X\rightarrow Y$ is a diffeomorphism, the pullback of $f$, noted by $f^{\ast}:\mathcal{D}'(Y)\rightarrow \mathcal{D}'(X)$, is defined as $$\label{227} f^{\ast}(w)(\phi)=w(|\det(f^{-1})'|\phi\circ f^{-1}),$$ where $w\in \mathcal{D}'(Y)$ and $\phi$ is a test function in $X$, i.e. $\phi\in \mathcal{C}^{\infty}_{c}(X)$. Take a differentiable structure on $\mathcal{M}$ given by charts $(U_{\alpha}, \varphi_{\alpha})_{\alpha\in \mathbb{N}}$. Then $\{\varphi_{\alpha}(U_{\alpha})\}_{\alpha\in \mathbb{N}}$ is a family of open sets of $\mathbb{R}^{n}$, so it makes sense to consider $\mathcal{D}'(\varphi_{\alpha}(U_{\alpha}))$. \[def12\] The space $H_{d}(\mathcal{M})$ is the set whose elements are families of the form $u=\{u_{\alpha}\}_{\alpha\in \mathbb{N}}$, where $u_{\alpha}\in \mathcal{D}'(\varphi_{\alpha}(U_{\alpha}))$, obeying the equality $$\label{228} u_{\beta}=(\varphi_{\alpha} \circ \varphi_{\beta}^{-1})^{\ast}u_{\alpha}$$ of distributions on $\varphi_{\beta}(U_{\alpha}\cap U_{\beta})$. Analogously, if $f$ is a diffeomorphism, the *semi-pullback* is given by $$\label{229} f^{\bullet}(w)(\phi)=w(\phi\circ f^{-1}),$$ where $w\in \mathcal{D}'(Y)$ and $\phi$ is a test function in $X$, i.e. $\phi\in \mathcal{C}^{\infty}_{c}(X)$. Notice that $$\label{2210} f^{\bullet}(w)=|\det(f')|f^{\ast}(w),$$ for all $w\in \mathcal{D}'(Y)$ \[def13\] The space $D_{d}(\mathcal{M})$ is the set whose elements are the families $u=\{u_{\alpha}\}_{\alpha\in \mathbb{N}}$, where $u_{\alpha}\in \mathcal{D}'(\varphi(U_{\alpha}))$, satisfying that $$\label{2211} u_{\beta}=(\varphi_{\alpha} \circ \varphi_{\beta}^{-1})^{\bullet}u_{\alpha}$$ as distributions on $\varphi_{\beta}(U_{\alpha}\cap U_{\beta})$. Note that tells us that equation can be rewritten as $$\label{212} u_{\beta}=|\det(\varphi_{\alpha} \circ \varphi_{\beta}^{-1})'|(\varphi_{\alpha} \circ \varphi_{\beta}^{-1})^{\ast}u_{\alpha}.$$ Definitions \[def12\] and \[def13\] are particular examples of what is called *distribution with density character $q$* (see [@mich] Chapter 3, Section 1, where they are also described as continuous dual spaces of sections of certain density bundles).\ We will extend the definition of wavefront set of H-distributions to D-distributions. We remark that the definition of wavefront set of H-distribution was introduced by L. Hörmander in [@hor] (see [@ma] for a short exposition). \[wed\] Let $f\in \mathcal{C}^{\infty}(W)$ positive, $W\subseteq R^{n}$ open and $\mu\in \mathcal{D}'(W)$. Hence $$\Sigma_{x}(\mu)=\Sigma_{x}(f\mu) \hspace{2cm}$$ for all $x\in W$. We take $\psi \in \mathcal{C}^{\infty}_{c}(W)$ such that $\psi$ takes the value $1$ in a closed neighbourhood of $x$ and $0$ in the complement of an open set containing that closed set. We also consider $\varphi\in \mathcal{C}^{\infty}_{c}(W)$ such that $\varphi(x)\neq 0$. Then, $\varphi \mu \in \mathcal{D}'(W)$ has compact support. If we apply [@ma], Lemma 2.5, to the compact support distribution $\varphi \mu$ and the smooth compact support function $f\psi$, we see that $\Sigma((f\psi) (\varphi \mu))\subseteq \Sigma(\varphi \mu)$. Since $\Sigma((f\psi)(\varphi \mu))= \Sigma((\varphi\psi) (f\mu))$, intersecting the above contention for all $\varphi\in \mathcal{C}^{\infty}_{c}(W)$ such that $\varphi(x)\neq 0$, we get that $\bigcap_{\varphi}\Sigma((\varphi\psi) (f \mu))\subseteq \Sigma_{x}(\mu)$. On the other hand, as $\varphi=\varphi\psi$ on a neighbourhood of $x$ and $\Sigma_{x}$ is a local property, we also conclude that $\Sigma_{x}(f\mu)\subseteq \Sigma_{x}(\mu)$. We apply the argument of the previous paragraph to the positive function $\phi=\frac{1}{f}$ and the distribution $u=f\mu$, to deduce that $$\begin{aligned} & \Sigma_{x}(\phi u)\subseteq \Sigma_{x}(u),\\ & \Sigma_{x}(\frac{1}{f} f\mu )\subseteq \Sigma_{x}(f\mu),\\ & \Sigma_{x}(\mu)\subseteq \Sigma_{x}(f\mu). \end{aligned}$$ As a consequence $\Sigma_{x}(\mu)=\Sigma_{x}(f\mu)$, which is what we want. Equation (2.133) of [@ma], tells us that $$\label{qqq} \Sigma_{x(p)}(u_{x})= \Sigma_{x(p)}(|\det(y\circ x^{-1})'|(y\circ x^{-1})^{\ast}(u_{y}))= \Sigma_{x(p)}((y\circ x^{-1})^{\ast}(u_{y})),$$ where $(x,U_{x})$ is an atlas for $\mathcal{M}$, $p\in \mathcal{M}$ and we have used equation and Lemma \[wed\]. As we can see in [@ma], equation (2.111), each diffeomorphism $f:X\to Y$ satisfies $$\label{kkk} \Sigma_{x}(f^{\ast}u)=f'(x)^{T}\Sigma_{f(x)}(u)$$ for any $x\in X$, where $f'(x)$ is the differential of $f$. Now if $u=\{u_{z}\}_{(z,U_{z})}$ is a D-distribution on $\mathcal{M}$ and $p\in \mathcal{M}$ is such that the charts $(x,U_{x})$ and $(y,U_{y})$ satisfy that $p\in U_{x}$ and $p\in U_{y}$, then using the Einstein’s summation convention, we get that $$\begin{aligned} & \{ \xi_{k}dx^{k}_{p}/ \xi\in \Sigma_{x(p)}(u_{x}) \}= \{ \xi_{k}dx^{k}_{p}/ \xi\in \Sigma_{x(p)}((y\circ x^{-1})^{\ast}u_{y}) \}\\ & = \{ ((y\circ x^{-1})'(x(p))^{T}\eta)_{k}dx^{k}_{p} / \eta\in \Sigma_{y(p)}(u_{y}) \}= \{ ((\frac{\partial y^{j}}{\partial x^{k}})_{p}\eta_{j}dx^{k}_{p} / \eta\in \Sigma_{y(p)}(u_{y})\} \\ & =\{ \eta_{j}dy^{j}_{p}/ \eta\in \Sigma_{y(p)}(u_{y}) \},\end{aligned}$$ where we have used equations and in the first equality and in the second one respectively. This allow us to introduce the following notion. \[wfd\] Let $u\in D_{d}(\mathcal{M})$. The wave front set of $u$ is the set $$WF(u)= \{ \{p\}\times \Sigma_{p}(u) / p\in \mathcal{M} \}\subseteq T^{\ast }\mathcal{M},$$ where $ \Sigma_{p}(u)=\{ \xi_{k}dx^{k}_{p}/ \xi\in \Sigma_{x(p)}(u_{x}) \}$ for any chart $(x,U_{x})$ such that $p\in U_{x}$, and $T^{\ast}\mathcal{M}$ is the cotangent bundle of the manifold. The pullback of distributions defined on open sets of $\mathbb{R}^{n}$ is also defined for maps which are not diffeomorphisms. \[hhhh\] Let $X$ and $Y$ be open subsets of $\mathbb{R}^{m}$ and $\mathbb{R}^{n}$ respectively and let $f:X\to Y$ be an smooth map. Define $$N_{f}=\{ (f(x),\nu)\in Y\times \mathbb{R}^{n} / x\in X, f'(x)^{T}\nu =\vec{0} \}.$$ Then there is a unique way of defining the pullback $$f^{\ast}:\{ u\in \mathcal{D}'(Y) / N_{f}\cap WF(u)=\emptyset \} \to \mathcal{D}'(X),$$ such that $f^{\ast}u=u\circ f$ for all $u\in \mathcal{C}^{\infty}(Y)$. We recall that given $u\in \mathcal{D}'(X)$ and $v\in \mathcal{D}'(Y)$, the exterior tensor product $u\otimes v$ is a distribution over $X\times Y$ (see for instance [@ma], Definition 1.48). The hypotheses of Theorem \[hhhh\] are satisfied for $\Delta:\mathbb{R}^{n}\rightarrow \mathbb{R}^{n} \times \mathbb{R}^{n}$ and a distribution of the form $\mu\otimes \nu$ if there is no $(x,\xi)\in WF(\mu)$ such that $(x,-\xi)\in WF(\nu)$ (as explained in [@ma], Theorem 2.167) where $\mu $ and $\nu$ are distributions over $\mathbb{R}^{n}$. The family $\{\Delta_{\alpha}^{\ast}(u_{\alpha}\otimes v_{\alpha})\}_{\alpha\in \mathbb{N}}$ will be of interest to us: these are distributions $ \Delta_{\alpha}^{\ast}(u_{\alpha}\otimes v_{\alpha})\subseteq \mathcal{D}'(\varphi_{\alpha}(U_{\alpha}))$ where $\Delta_{\alpha}:\varphi_{\alpha}(U_{\alpha})\rightarrow \varphi_{\alpha}(U_{\alpha})\times \varphi_{\alpha}(U_{\alpha})\subseteq \mathbb{R}^{2n}$ are the diagonal maps. The following will be useful. \[lema6\] Given $f:Z\rightarrow X$ and $g:W\rightarrow Y$ diffeomorphisms between open sets of $\mathbb{R}^{n}$, let $f\times g: Z\times W\rightarrow X\times Y$ be defined by $(f\times g)(z,w)=(f(z),g(w))$. Then $$(f\times g)^{\ast}(u\otimes v)(\varphi \otimes \psi )= (f^{\ast}(u)\otimes g^{\ast}(v))(\varphi \otimes \psi),$$ for all $u\in \mathcal{D}'(X)$, $v\in \mathcal{D}'(Y)$, $\varphi\in \mathcal{C}^{\infty}_{c}(Z)$ and $\psi\in \mathcal{C}^{\infty}_{c}(W)$. Since $\mathcal{C}^{\infty}_{c}(Z) \otimes \mathcal{C}^{\infty}_{c}(W)\subseteq \mathcal{C}^{\infty}_{c}(Z\times W)$ is dense and the pullback is sequentially continuous (see [@ma]), we can conclude that $$(f\times g)^{\ast}(u\otimes v)= f^{\ast}(u)\otimes g^{\ast}(v)$$ as elements of $\mathcal{D}'(Z\times W)$. First of all we take sequences $\{u_{j}\}_{j\in \mathbb{N}}$ and $\{v_{m}\}_{m\in \mathbb{N}}$ of continuous functions over $X$ and $Y$, respectively, such that $u_{n}\to u$ and $v_{n}\to v$. We recall that $\mathcal{D}'(X)$ has the weak$^{\star}$-topology induced by $\mathcal{C}^{\infty}_{c}(X)$, and the same holds for $Y$. Then, $$((f\times g)^{\ast}(u\otimes v))(\varphi\otimes\psi)=\lim_{l\to \infty}((f\times g)^{\ast}(u_{l}\otimes v_{l}))(\varphi\otimes\psi),$$ because the pullback is sequentially continuous. Hence, $$\lim_{l\to \infty}((f\times g)^{\ast}(u_{l}\otimes v_{l}))(\varphi\otimes\psi)=\lim_{l\to \infty}\int_{Z\times W}((u_{l}\otimes v_{l})\circ (f\times g))(z,w)\cdot (\varphi\otimes\psi)(z,w) dz dw,$$ for the pullback map is an extension of the precompositions from functions to distributions. Moreover, by definition we see that $$\lim_{l\to \infty}\int_{Z\times W}(u_{l}\circ f(z)\cdot \varphi(z))\cdot((v_{l}\circ g)(w)\cdot\psi(w)) dz dw,$$ which by Fubinni’s Theorem gives us $$\lim_{l\to \infty}[\int_{Z}(u_{l}\circ f(z)\cdot \varphi(z))dz\cdot\int_{W}((v_{l}\circ g)(w)\cdot\psi(w))dw]$$ $$= \lim_{l\to \infty}\int_{Z}(u_{l}\circ f(z)\cdot \varphi(z))dz\cdot \lim_{l\to \infty}\int_{W}((v_{l}\circ g)(w)\cdot\psi(w))dw$$ $$= \lim_{l\to \infty} f^{\ast}u_{l}(\varphi) \cdot \lim_{l\to \infty} g^{\ast}v_{l}(\psi)=f^{\ast}u(\varphi)\cdot g^{\ast}v(\psi)=(f^{\ast}u \otimes g^{\ast}v)(\varphi\otimes \psi).$$ The lemma is thus proved. Consider now the pullback of $\Delta_{\alpha}^{\ast}(u_{\alpha}\otimes v_{\alpha})\in \mathcal{D}'(\varphi_{\alpha}(U_{\alpha}\cap U_{\beta}))$ by $\varphi_{\alpha} \circ \varphi_{\beta}^{-1}$. Note that the following diagram $$\xymatrix{ \varphi_{\beta}(U_{\alpha}\cap U_{\beta}) \ar[rrr]^{\Delta_{\beta}} \ar[dd]_{\varphi_{\alpha} \circ \varphi_{\beta}^{-1}} & & & \varphi_{\beta}(U_{\alpha}\cap U_{\beta})\times \varphi_{\beta}(U_{\alpha}\cap U_{\beta}) \ar[dd]^{(\varphi_{\alpha} \circ \varphi_{\beta}^{-1})\times (\varphi_{\alpha} \circ \varphi_{\beta}^{-1})} \\ & & & \\ \varphi_{\alpha}(U_{\alpha}\cap U_{\beta}) \ar[rrr]^{\Delta_{\alpha}} & & & \varphi_{\alpha}(U_{\alpha}\cap U_{\beta})\times \varphi_{\alpha}(U_{\alpha}\cap U_{\beta}) }$$ is commutative. As a consequence, $$(\varphi_{\alpha} \circ \varphi_{\beta}^{-1})^{\ast}(\Delta^{\ast}_{\alpha}(u_{\alpha} \otimes v_{\alpha}))= (\varphi_{\alpha} \circ \varphi_{\beta}^{-1})^{\ast}\circ\Delta^{\ast}_{\alpha}(u_{\alpha} \otimes v_{\alpha})= (\Delta_{\alpha}\circ\varphi_{\alpha} \circ \varphi_{\beta}^{-1})^{\ast}(u_{\alpha} \otimes v_{\alpha})$$ $$\label{lll} =[((\varphi_{\alpha} \circ \varphi_{\beta}^{-1})\times (\varphi_{\alpha} \circ \varphi_{\beta}^{-1}))\circ \Delta_{\beta}]^{\ast}(u_{\alpha} \otimes v_{\alpha}) = \Delta_{\beta}^{\ast}\circ ((\varphi_{\alpha} \circ \varphi_{\beta}^{-1})\times (\varphi_{\alpha} \circ \varphi_{\beta}^{-1}))^{\ast}(u_{\alpha} \otimes v_{\alpha}),$$ where we have used Theorem 2.126 of [@ma] in the second equality, the commutation of the previous diagram and Lemma 2.130 of [@ma]. By Lemma \[lema6\], Definitions 14 and 15, and equation , the last member of coincides with $$\Delta_{\beta}^{\ast}\circ ((\varphi_{\alpha} \circ \varphi_{\beta}^{-1})^{\ast}(u_{\alpha})\otimes (\varphi_{\alpha} \circ \varphi_{\beta}^{-1})^{\ast}( v_{\alpha}))=\Delta_{\beta}^{\ast}(u_{\beta}\otimes \frac{1}{|\det(\varphi_{\alpha} \circ \varphi_{\beta}^{-1})'|} v_{\beta})$$ $$= \frac{1}{|\det(\varphi_{\alpha} \circ \varphi_{\beta}^{-1})'|} \Delta_{\beta}^{\ast}(u_{\beta}\otimes v_{\beta}).$$ Hence , by multiplying this equality by $|\det(\varphi_{\alpha} \circ \varphi_{\beta}^{-1})'|$ we obtain $$\Delta_{\beta}^{\ast}(u_{\beta}\otimes v_{\beta})=|\det(\varphi_{\alpha} \circ \varphi_{\beta}^{-1})'|(\varphi_{\alpha} \circ \varphi_{\beta}^{-1})^{\ast}(\Delta^{\ast}_{\alpha}(u_{\alpha} \otimes v_{\alpha})).$$ By equation we get that $$\Delta_{\beta}^{\ast}(u_{\beta}\otimes v_{\beta})=(\varphi_{\alpha} \circ \varphi_{\beta}^{-1})^{\bullet}(\Delta^{\ast}_{\alpha}(u_{\alpha} \otimes v_{\alpha})).$$ Hence, the family $\{\Delta_{\alpha}^{\ast}(u_{\alpha}\otimes v_{\alpha})\}_{\alpha\in \mathbb{N}}$ defines an element of $D_{d}(\mathcal{M})$ which we will denote $u\cdot v$. Indeed we have just proved the following proposition. There exist an action of $H_{d}(\mathcal{M})$ over $D_{d}(\mathcal{M})$ which we note $$\label{hd} \xymatrix{H_{d}(\mathcal{M})\times D_{d}(\mathcal{M})\ar[r] & D_{d}(\mathcal{M}) \\ (u,v) \ar[r] & u\cdot v }$$ provided that for each $\alpha \in \mathbb{N}$, there are not points $(x,\xi)$ in $WF(u_{\alpha})$ such that $(x,-\xi)$ is in $WF(v_{\alpha})$. \[rem\] Lastly we want mention that there is an alternative definition of H-distributions on a manifold which was introduced by [@hor] or [@ma], as elements of the continuous dual of $\Gamma_{c}(\omega)$ (the compact support sections of the density bundle of $\mathcal{M}$ (the manifold)) respect to a topology given by seminorms (this topology is explicitly defined in both text and was described here in the previous section) which makes of $\Gamma_{c}(\omega)$ a Fréchet space. On the other hand if we follow the text [@di] the distributions were introduced as elements of the continuous dual of $\mathcal{C}^{\infty}_{c}(\mathcal{M})$ (in [@hor] these were called distribution densities), we will call these D-distributions. The relation between this and the above definitions is clearly established in both texts, but we say that for any family $\{u_{\alpha}\}_{\alpha\in \mathbb{N}}$ which satisfies the Definition \[def12\] there exists only one element $\tilde{u}$ in the continuous dual of $\Gamma_{c}(\omega)$ such that locally works in the following way, $$\tilde{u}(\delta)=\int_{\varphi_{\alpha}(U_{\alpha})}u_{\alpha}(\varphi_{\alpha}^{-1}) ^{\ast}\delta$$ where $\delta$ is an element of $\Gamma_{c}(\omega)$ whose support is in $U_{\alpha}$. Using the Definition \[def12\] one sees that $\tilde{u}$ is well defined independently of $U_{\alpha}$. Similar considerations are valid for the D-distributions. We invite the reader to deepen on these topics in the aforementioned books. Basic facts on Feynman measures =============================== Lagrangian formulation ---------------------- We begin by expressing in mathematical terminology certain ideas that are common in physics. The manifolds will allways be $T_{2}$ and satisfy the second axiom of countability in this thesis and consequently they are paracompact and a have locally finite partition of the unity. A relation $\mathcal{R}$ in a topological space $X$ is said to be *closed* if the set $\{ (x,y)\in X^{2} / x\mathcal{R}y \} \subseteq X^{2}$ is closed for the product topology of $X^{2}$. \[space\] A *spacetime* will be a smooth finite-dimensional manifold $\mathcal{M}$, together with a closed, reflexive and transitive relation $\preceq$. Two points in the spacetime will be called *spacelike separated* if $x\npreceq y$ and $y\npreceq x$. \[ej1\] A first example of spacetime is $\mathbb{R}$ with the usual manifold structure and the classical order $\leqslant$ relation, which is obviously reflexive and transitive. The relation $\leqslant$ is also closed because the set $\{ (x,y) \in \mathbb{R}^{2} / x\leqslant y \}$ has as complement the positivity set of the continuous function $f(x,y)=x-y$. In this example there are no spacelike separated points, for $\leqslant$ is a total order. $\Box$ \[ej2\] We shall now consider the space $\mathbb{R}^{1,3}$, also denoted by $\mathbb{M}^{4}$, and called the *Minkowski spacetime*. Its underlying manifold is $\mathbb{R}^{4}$ and the relation is given as follows. Define the bilinear form $\langle - ; - \rangle:\mathbb{R}^{4}\times\mathbb{R}^{4}\to \mathbb{R}$ by $\langle z^{1},z^{2}\rangle=z^{1}_{0}z^{2}_{0}-z^{1}_{1}z^{2}_{1}-z^{1}_{2}z^{2}_{2}-z^{1}_{3}z^{2}_{3}$ where $z^{i}=(z^{i}_{0},z^{i}_{1},z^{i}_{2},z^{i}_{3})$ for $i=1,2$. Then the relation is the following, $z_{1}\preceq z_{2}$ iff $\langle z_{2}-z_{1};z_{2}-z_{1}\rangle \geqslant 0$ and $z^{2}_{0}\geqslant z^{1}_{0}$. The relation $\preceq$ is clearly reflexive. It is transitive, because if $z^{3}=(z^{3}_{0},z^{3}_{1},z^{3}_{2},z^{3}_{3})$ and $z_{1}\preceq z_{2}$ and $z_{2}\preceq z_{3}$, then: $$\begin{aligned} & (z^{3}-z^{1};z^{3}-z^{1})=(z^{3}_{0}-z^{1}_{0})^{2}-(z^{3}_{1}-z^{1}_{1})^{2}-(z^{3}_{2}-z^{1}_{2})^{2}-(z^{3}_{3}-z^{1}_{3})^{2} \\ & = (z^{3}_{0}-z^{2}_{0}+z^{2}_{0}-z^{1}_{0})^{2}-(z^{3}_{1}-z^{2}_{1}+z^{2}_{1}-z^{1}_{1})^{2}\\ & -(z^{3}_{2}-z^{2}_{2}+z^{2}_{2}-z^{1}_{2})^{2}-(z^{3}_{3}-z^{2}_{3}+z^{2}_{3}-z^{1}_{3})^{2} \\ & =(z^{3}_{0}-z^{2}_{0})^{2}+(z^{2}_{0}-z^{1}_{0})^{2}+2(z^{3}_{0}-z^{2}_{0})(z^{2}_{0}-z^{1}_{0})\\ & -||((z^{3}_{1},z^{3}_{2},z^{3}_{3})-(z^{2}_{1},z^{2}_{2},z^{2}_{3}))+((z^{2}_{1},z^{2}_{2},z^{2}_{3})-(z^{1}_{1},z^{1}_{2},z^{1}_{3}))||^{2}.\end{aligned}$$ Applying the triangle inequality we have that $$\begin{aligned} & (z^{3}-z^{1};z^{3}-z^{1})\geqslant (z^{3}_{0}-z^{2}_{0})^{2}+(z^{2}_{0}-z^{1}_{0})^{2}+2(z^{3}_{0}-z^{2}_{0})(z^{2}_{0}-z^{1}_{0})\\ & -||(z^{3}_{1},z^{3}_{2},z^{3}_{3})-(z^{2}_{1},z^{2}_{2},z^{2}_{3})||^{2}-||(z^{2}_{1},z^{2}_{2},z^{2}_{3})-(z^{1}_{1},z^{1}_{2},z^{1}_{3})||^{2} \\ & = [(z^{3}_{0}-z^{2}_{0})^{2}-||(z^{3}_{1},z^{3}_{2},z^{3}_{3})-(z^{2}_{1},z^{2}_{2},z^{2}_{3})||^{2}]\\ & +[(z^{2}_{0}-z^{1}_{0})^{2}-||(z^{2}_{1},z^{2}_{2},z^{2}_{3})-(z^{1}_{1},z^{1}_{2},z^{1}_{3})||^{2}]+2(z^{3}_{0}-z^{2}_{0})(z^{2}_{0}-z^{1}_{0})\\ & =(z^{3}-z^{2};z^{3}-z^{2})+(z^{2}-z^{1};z^{2}-z^{1})+2(z^{3}_{0}-z^{2}_{0})(z^{2}_{0}-z^{1}_{0}).\end{aligned}$$ The three last summands are $\geqslant 0$ because of the hypotheses $z_{1}\preceq z_{2}$ and $z_{2}\preceq z_{3}$. The transitivity of $\preceq$, follows from the fact that $z^{3}_{0}\geqslant z^{1}_{0}$. This relation is also closed, because the complement of $\{ (z^{1},z^{2})\in \mathbb{R}^{8}/ z^{1}\preceq z^{2} \}$ is the preimage of the open set $\{(x,y)\in \mathbb{R}^{2}/x>0 \text{ or }y>0 \}$ by the continuous function $f:\mathbb{R}^{8}\to \mathbb{R}^{2}$ defined by $f (z^{1},z^{2})=(||(z^{1}_{1},z^{1}_{2},z^{1}_{3})-(z^{2}_{1},z^{2}_{2},z^{2}_{3})||^{2}-(z^{1}_{0}-z^{2}_{0})^{2},z^{1}_{0}-z^{2}_{0})$.$\Box$ As we mentioned in Section 1.2, if $\mathcal{M}$ is a spacetime then it has an structure $(\mathcal{M}, \mathcal{C}^{\mathbb{R}})$ of ringed space over $\mathbb{R}$. Let $(E,p,\mathcal{M})$ be a finite dimensional complex vector bundle over $\mathcal{M}$ and $\Phi$ the sheaf of modules over the ringed space $(\mathcal{M}, \mathcal{C}^{\mathbb{R}})$ associated to it (see Lemma \[lemshe\]). This vector bundle and the corresponding sheaf are fixed from now on. It will be called the *sheaf of classical fields* over $\mathcal{M}$ and its local sections, i.e. elements of $\Phi(U)=\Gamma(U,E)$ where $U$ is open in $\mathcal{M}$, are called *classical fields*. (Classical mechanics) Consider the spacetime with the underlying manifold $\mathbb{R}^{3}$ and the identity relation, and the vector bundle $\pi:T\mathbb{R}^{3}\to \mathbb{R}^{3}$ given by the tangent bundle. A classical field in classical mechanics is a section of the previous vector bundle. $\Box$ \[ejcamppoescalar\] (Classical field theory) In a classical field theory, one usually considers the trivial line vector bundle over the Minkowski spacetime $\pi:\mathbb{M}^{4}\times \mathbb{R}\to\mathbb{M}^{4}$. A classical field in this situation is just a section of $\pi$. So a classical field $\Phi$ is given by a function $\phi:\mathbb{M}^{4}\to \mathbb{R}$. We recall that the d’Alembertian operator $\Box$ on a function $g:\mathbb{M}^{4}\to \mathbb{R}$ is given by $\Box g=\frac{\partial^{2}g}{\partial t^{2}}-\frac{\partial^{2}g}{\partial x_{1}^{2}}-\frac{\partial^{2}g}{\partial x_{2}^{2}}-\frac{\partial^{2}g}{\partial x_{3}^{2}}$. A *massive scalar field of mass $m$* is a classical field $\varPhi(z)=(z,\phi(z))$ such that $\phi$ satisfies the *Klein-Gordon equation* $(-\Box +m^{2})\phi=0$. $\Box$ \[electro\](Electromagnetism) Let $\mathbb{M}^{4}$ be the Minkowski spacetime and $\Omega^{2}\mathbb{M}^{4}$ be the vector bundle of antisymmetric two forms over $\mathbb{M}^{4}$. If $U\subseteq \mathbb{M}^{4}$ is an open set, a local section of $\Omega^{2}\mathbb{M}^{4}$ over $U$ is of the form $$F(t,\vec{x})=\frac{1}{2}F_{\mu \nu}(t,\vec{x})dx^{\mu}\wedge dx^{\nu}$$ and is a classical field (where the Einstein’s summation convention and the notation $x^{0}=t$, $x^{1}=x_{1}$,$x^{2}=x_{2}$ and $x^{3}=x_{3}$ were used). We define the electric $E:U\to \mathbb{R}^{3}$ and the magnetic $B:U\to \mathbb{R}^{3}$ fields by $$\begin{aligned} E_{x}(t,\vec{x})=F_{0 1}(t,\vec{x}), & E_{y}(t,\vec{x})=F_{0 2}(t,\vec{x}), & E_{z}(t,\vec{x})=F_{0 2}(t,\vec{x}),\\ B_{z}(t,\vec{x})=-F_{1 2}(t,\vec{x}), & B_{y}(t,\vec{x})=F_{1 3}(t,\vec{x}), & B_{x}(t,\vec{x})=-F_{2 3}(t,\vec{x}). \end{aligned}$$ The classical Maxwell’s equations in electromagnetism are written in the form $$\begin{aligned} & \partial_{\eta} F_{\mu \nu} +\partial_{\mu} F_{\nu \eta}+\partial_{\nu} F_{\eta \mu}=0 & & \text{ and } & & \partial_{\nu}F^{\mu \nu}=j^{\mu}, \end{aligned}$$ where $j^{o}(t,\vec{x})=\rho(t,\vec{x})$ is the charge density, $j^{i}$ the density current in the direction of $x_{i}$ for $1\leqslant i \leqslant 3$, and $F^{\mu \nu}= g^{\mu \alpha}g^{\nu \beta}F_{\alpha \beta}$ with $g$ the metric tensor $g=diag(-1,1,1,1)$. $\Box$ Given the vector bundle $(E,p,\mathcal{M})$ over the spacetime $\mathcal{M}$ and $\Phi$ the associated sheaf, the sheaf of jets $J\Phi$ constructed after Lemma \[lema5\] is called the sheaf of *derivatives of classical fields*. The sheaf $S_{\mathcal{C}^{\mathbb{R}}}J\Phi$ given by the symmetric construction of the sheaf of modules $J\Phi$ over the ringed space $(\mathcal{M},\mathcal{C}^{\mathbb{R}})$ is called the *sheaf of Lagrangians* (or *composite fields*). If $\phi$ is a massive scalar field of mass $m$ as in Example \[ejcamppoescalar\], then $$l_{1}=\frac{1}{2}\partial_{\mu}\phi\odot\partial^{\mu}\phi-\frac{1}{2}m^{2}\phi^{\odot 2}$$ is an example of Lagrangian, where we have used Einstein’s convention for the sum. We will usually omit the symmetric product to lighten the notation and just write $l_{1}=\frac{1}{2}\partial_{\mu}\phi\partial^{\mu}\phi-\frac{1}{2}m^{2}\phi^{2}$. An example of Lagrangian for Example \[electro\] is $$l_{2}= F_{\mu,\nu} F^{\mu,\nu}.$$ We will give a few more examples. If $\varphi$ and $\psi$ are classical fields over an spacetime $\mathcal{M}$ (see Example \[ejcamppoescalar\]) and $f,g$ are smooth functions over $\mathcal{M}$; then $$\begin{aligned} & l_{3}=\overline{\psi} \psi, \\ & l_{4}=f \varphi \partial_{1}\psi+g\varphi^{2} \partial_{1}\partial_{3}\psi,\\ & l_{5}=(f\varphi)\psi=\varphi(f\psi)=f(\varphi\psi),\end{aligned}$$ are examples of Lagrangians. Any combination of this Lagrangians with coefficients in the ring of smooth functions over the spacetime is also a Lagrangian. We recall that we denote by $\omega$ the sheaf of densities of the spacetime. \[def5\] The *sheaf of Lagrangian densities* is $\omega SJ\Phi$. A Lagrangian density is any of its sections. For example, if $\mathcal{M}$ is orientable of dimension $n$ and $\varphi$ is a classical field with support contained in an open set $U$ of $\mathcal{M}$, we can regard $d^{n}x$ (where $(U,x)$ is a chart) as a local section of the density bundle $\omega$ which together with any Lagrangian over $U$ can produce a Lagrangian density $\mathcal{L}$ as for example $$\mathcal{L}=(\varphi+m\varphi^{2}\partial\varphi)d^{n}x,$$ where $m$ is a constant. \[noloc\] A *non-local action* is an element of the symmetric algebra $S\Gamma \omega SJ\Phi$ of the $\mathbb{R}$-vector space of global sections of the sheaf $\omega SJ\Phi$. In order to topologize the symmetric algebra $S\Gamma \omega SJ\Phi$ of the $\mathbb{R}$-vector space of global sections of the sheaf $\omega SJ\Phi$, we must first consider the algebra $T\Gamma \omega SJ\Phi$. And as we saw in Section \[1.2.2\] there are two ways of topologize $T^{k}\Gamma \omega SJ\Phi$ ( for $k\in \mathbb{N}$) one considering the projective $\otimes_{\pi}$ topology over the tensor product and the second one considering its completation $\hat{\otimes}_{\pi}$. We can choose the topology we want for $T^{k}\Gamma \omega SJ\Phi$, but if we take the topology given by $\hat{\otimes}_{\pi}$ then the two ways of topologize $S^{k}\Gamma \omega SJ\Phi$ will be homeomorphic as we saw in Section \[jhjhjh\]. The proofs for the next facts can be found in [@con], Theorem 7.5.5 and Corollary 7.5.6. \[df\] Let $E$ and $F$ be two finite dimensional complex vector bundles $\mathcal{M}$. The $\mathcal{C}^{\infty}(\mathcal{M})$-linear map $\alpha:\Gamma(E)\otimes_{\mathcal{C}^{\infty}(\mathcal{M})}\Gamma(F) \to \Gamma(E\otimes F) $, defined by $\alpha(s\otimes t)(x)=s(x)\otimes t(x)\in E_{x}\otimes F_{x}=(E\otimes F)_{x}$, where $s\in \Gamma(E)$, $t\in \Gamma(F)$ and $x\in \mathcal{M}$, is a canonical isomorphism of $\mathcal{C}^{\infty}(\mathcal{M})$-modules. As a corollary of this theorem we have \[dm\] Given a finite dimensional complex vector bundle $E$ over $\mathcal{M}$, there exists a canonical isomorphism of $\mathcal{C}^{\infty}(\mathcal{M})$-modules between $\Gamma(S^{k}E)$ and $S^{k}_{\mathcal{C}^{\infty}(\mathcal{M})}\Gamma (E)$ for all $k\in \mathbb{N}_{0}$. In Section \[jets\] we told that $J\Phi$ denote the vector bundle $J^{k}\Phi$ for some $k\in \mathbb{N}\cup \{ \infty \}$ and unless $k=\infty$, $J^{k}\Phi$ is a finite dimensional vector bundle over $\mathcal{M}$ and we can apply the Theorem \[df\] and Corollary \[dm\] to it. We want to mention a result similar to Theorem \[df\], whose proof involves the Serre-Swan Theorem for non compact manifolds, which establishes that the $\mathcal{C}^{\infty}(\mathcal{M})$-module of global sections of a smooth vector bundle over a manifold $\mathcal{M}$ is projective over $\mathcal{C}^{\infty}(\mathcal{M})$ (see [@serre]). \[uhuhuh\] Given $E$ and $F$ two finite dimensional complex vector bundles over a manifold $\mathcal{M}$, the restriction of the morphism $\alpha$ of Theorem \[df\] establishes an isomorphism $\alpha| :\Gamma_{c}(E)\otimes_{\mathcal{C}^{\infty}(\mathcal{M})} \Gamma(F)\to \Gamma_{c}(E\otimes F)$. Given a finite dimensional vector bundle $H$ over a manifold $\mathcal{M}$, we consider the map $\chi:\mathcal{C}_{c}^{\infty}(\mathcal{M})\times \Gamma(H)\to \Gamma_{c}(H)$ defined by $\chi(f,\varphi)=f\varphi$ where $f\in \mathcal{C}_{c}^{\infty}(\mathcal{M})$, $\varphi\in \Gamma(H)$ and if $p\in \mathcal{M}$, $f\varphi(p)=f(p)\varphi(p)$, then $f\varphi$ is a compact support section of $H$. The map $\chi$ is clearly $\mathcal{C}^{\infty}(\mathcal{M})$-balanced, hence it factors throughout the tensor product as $\overline{\chi}:\mathcal{C}_{c}^{\infty}(\mathcal{M})\otimes_{\mathcal{C}^{\infty}(\mathcal{M})} \Gamma(H)\to \Gamma_{c}(H)$. The map $\overline{\chi}$ is an isomorphism. It is surjective because if $\psi\in \Gamma_{c}(H)$ then suppose $U$ is an open set containing the support of $\psi$ and $V$ is a compact set such that $supp(\psi)\subseteq U \subseteq V$, then if we take $f\in \mathcal{C}_{c}^{\infty}(\mathcal{M})$ such that $f$ takes the value $1$ on $U$ and the value zero outside $V$ and $\overline{\chi}(f\otimes \psi)=f\psi=\psi$. On the other hand the map $\overline{\chi}$ is injective because if $\overline{\chi}(f\otimes \varphi)=0\in \Gamma_{c}(H)$, then there is no point $p\in \mathcal{M}$ such that both $f(p)\neq 0$ and $\varphi(p)\neq 0$. Let us consider an elemental sequence of compact sets $C_{n}\nearrow supp(\phi)$ and using the Urysohn’s lemma a sequence of functions $g_{n}\in \mathcal{C}^{\infty}_{c}(\mathcal{M})$ such that $g\equiv 1$ over $C_{n}$ and $g\equiv 0$ over $supp(f)$. Then we have $f\otimes_{\mathcal{C}^{\infty}(\mathcal{M})} \varphi=\lim_{n\to \infty} f\otimes_{\mathcal{C}^{\infty}(\mathcal{M})}g_{n} \varphi=\lim_{n\to \infty} fg_{n}\otimes_{\mathcal{C}^{\infty}(\mathcal{M})} \varphi=0\in \mathcal{C}_{c}^{\infty}(\mathcal{M})\otimes_{\mathcal{C}^{\infty}(\mathcal{M})} \Gamma(H)$, as we want. As the module $\Gamma(F)$ is $\mathcal{C}^{\infty}_{c}(\mathcal{M})$-projective then we can apply the functor $(-)\otimes_{\mathcal{C}^{\infty}_{c}(\mathcal{M})} \Gamma(F)$ to the precedent isomorphism $$\overline{\chi}:\mathcal{C}_{c}^{\infty}(\mathcal{M})\otimes_{\mathcal{C}^{\infty}(\mathcal{M})} \Gamma(E)\to \Gamma_{c}(E)$$ taking $H=E$ and still have an isomorphism of $\mathcal{C}^{\infty}_{c}(\mathcal{M})$-modules, $$\mathcal{C}_{c}^{\infty}(\mathcal{M})\otimes_{\mathcal{C}^{\infty}(\mathcal{M})} \Gamma(E)\otimes_{\mathcal{C}^{\infty}(\mathcal{M})} \Gamma(F) \to \Gamma_{c}(E)\otimes_{\mathcal{C}^{\infty}(\mathcal{M})} \Gamma(F).$$ Moreover by applying Theorem \[df\] we have an isomorphism, $$\mathcal{C}_{c}^{\infty}(\mathcal{M})\otimes_{\mathcal{C}^{\infty}(\mathcal{M})} \Gamma(E\otimes F) \to \Gamma_{c}(E)\otimes_{\mathcal{C}^{\infty}(\mathcal{M})} \Gamma(F)$$ and by applying the above result again using $H=E\otimes F$ then we have the isomorphism $$\Gamma_{c}(E\otimes F) \to \Gamma_{c}(E)\otimes_{\mathcal{C}^{\infty}(\mathcal{M})} \Gamma(F)$$ which is what we established. \[ghgh\] In this remark we will apply the above results to the case of sections of the vector bundle $J\Phi$ (always under the hypothesis of finite order jets). Given $k\in \mathbb{N}$, the sections of the vector bundle $S^{k}J\Phi$, that is $\Gamma S^{k}J\Phi$ are isomorphic by Corollary \[dm\] to $S^{k}_{\mathcal{C}^{\infty}(\mathcal{M})}\Gamma J\Phi$. Moreover by Theorem \[df\] together with the recent comment we have an isomorphism between $\Gamma \omega S^{k}J\Phi$ and $\Gamma \omega \otimes_{\mathcal{C}^{\infty}(\mathcal{M})} S^{k}_{\mathcal{C}^{\infty}(\mathcal{M})}\Gamma J\Phi$. Lastly by Proposition \[uhuhuh\] we have an isomorphism similar to the last one, but very important if we want integrate a local action over the spacetime, $\Gamma_{c} \omega S^{k}J\Phi\simeq\Gamma_{c} \omega \otimes_{\mathcal{C}^{\infty}(\mathcal{M})} S^{k}_{\mathcal{C}^{\infty}(\mathcal{M})}\Gamma J\Phi$. As a consequence a *monomial element* of the space of non-local actions $A\in S^{n}\Gamma \omega S^{k}J\Phi$ with $n\in \mathbb{N}$ can be written $A=\prod_{i=1}^{n} \alpha_{i}\otimes (\odot_{j=1}^{k} l_{j}^{i})$, $\alpha_{i}\in \Gamma(\omega)$ and $l_{j}^{i}\in \Gamma(J\Phi)$, and $\prod$ denotes the product of $SV$ for $V=\Gamma\omega SJ\Phi$. In view of the above remarks the densities $\alpha_{i}\in \Gamma(\omega)$ can be taken of compact support, i.e. $\alpha_{i}\in \Gamma_{c}(\omega)$, if $A$ has compact support. Propagators ----------- ### Types of propagators Given a complex finite dimensional vector bundle $\pi:E\to \mathcal{M}$ over the spacetime and $\varphi \in \Gamma(J\Phi)$ we denote by $\varphi^{\ast}$ its complex conjugate section, i.e. the section such that $\varphi^{\ast}(p)=\overline{\varphi(p)}$ for all $p\in \mathcal{M}$ (the bar indicates the complex conjugate). And if $f=\delta \varphi \in \Gamma_{c}\omega J\Phi$, we denote by $f^{\ast}$ the section $\delta \varphi^{\ast}$ where $\ast$ does not work over the real densities, i.e. $f^{\ast}(p)=\delta(p)\otimes\overline{\varphi(p)}$ if $\delta$ is a real density an $p\in \mathcal{M}$. \[pro\] A *propagator* associated to the vector bundle $\pi:E\to \mathcal{M}$ (or its associated sheaf $\Phi$) is a continuous and $\mathbb{R}$-bilinear function $$\Delta: \Gamma_{c}\omega J\Phi \times \Gamma_{c}\omega J\Phi \rightarrow \mathbb{C}.$$ We denote the space of propagators associated to $\pi:E\to \mathcal{M}$ by $Prop(E)$. We say that the propagator is - *local* if $\Delta(f,g)=\Delta(g,f)$ for each $f$ and $g$ whose supports are spacelike separated (see Definition \[space\]) . - *Feynman* if it is symmetric. - *Hermitian* if $\Delta^{\ast}=\Delta$, where $\Delta^{\ast}(f^{\ast},g^{\ast}):=\overline{\Delta(g,f)}$ and $f^{\ast}$ is the section of $\Gamma_{c}\omega J\Phi$ given by a $\ast$-operation on it. - *positive* if $\Delta(f^{\ast},f)\geqslant 0$. Given two finite dimensional vector bundles $\pi_{E}:E\to M$ and $\pi_{F}:F\to N$, we consider the manifold $N\times M$ and pull-back these bundles by the projections $pr_{1}:N\times M \to N$ and $pr_{2}:N\times M \to M$ to construct the bundles, $ pr_{1}^{\ast}(\pi_{F}):pr_{1}^{\ast}(F) \to N\times M$ and $ pr_{2}^{\ast}(\pi_{E}):pr_{2}^{\ast}(E) \to N\times M$, so we have two bundles over $N\times M$ and we denote by $F\boxtimes E$ the bundle $ pr_{1}^{\ast}(F)\otimes pr_{2}^{\ast}(E) $ over $N\otimes M$. Given two finite dimensional $K$ (real or complex) vector bundles $E$ and $F$, we have an isomorphism $\Gamma(E\boxtimes F)\simeq\Gamma(E)\hat{\otimes}_{K}\Gamma(F)$ where the completion is with respect to the just described Fréchet topologies. The next proposition give us a link between propagators and distributions over $\mathcal{M}\times \mathcal{M}$. There is a linear map $\iota:Prop(E)\to \Gamma(\omega_{\mathcal{M}\times \mathcal{M}} \otimes (J\Phi \boxtimes J\Phi))'$ given by $\iota(\Delta)((\delta_{1}\otimes \delta_{2})\otimes (f\otimes g))=\Delta(\delta_{1}\otimes f,\delta_{2}\otimes g)$, for all $\delta_{1},\delta_{2}\in \Gamma(\omega)$ and $f,g\in \Gamma(J\Phi)$. Since $\Gamma(\omega_{\mathcal{M}\times \mathcal{M}}\otimes(J\Phi\boxtimes J\Phi))=\Gamma((\omega_{\mathcal{M}}\boxtimes \omega_{\mathcal{M}})\otimes (J\Phi \boxtimes J\Phi))$ for $\omega_{\mathcal{M}\times \mathcal{M}}=\omega_{\mathcal{M}}\boxtimes \omega_{\mathcal{M}}$, the previous conditions determine precisely one unique $(J\Phi\boxtimes J\Phi)^{\ast}$-valued distribution on $\mathcal{M}\times \mathcal{M}$ The space of bilinear maps with respect to $\mathcal{C}^{\infty}(M)^{\otimes 2}$ associated to a bundle $\pi:E\to \mathcal{M}$ $$\Gamma_{c}J\Phi\times \Gamma_{c}J\Phi \to \{ \text{H-distibutions of compact support on }\mathcal{M}\times \mathcal{M} \}=\Gamma(\omega_{\mathcal{M}\times \mathcal{M}})'$$ will be denoted by $Prop'(E)$ \[khkg\] There is a linear map $\xi: Prop(E)\to Prop'(E)$ satisfying that $\xi(\Delta)(A\otimes B)(f\otimes g)=\Delta(fA,gB)$ for all $A, B\in \Gamma_{c}J\Phi$ and $f,g\in \Gamma(\omega_{\mathcal{M}})$. This map is clearly injective, because if $\xi(\Delta)=\xi(\tilde{\Delta})$ then by using the Proposition \[uhuhuh\] and the definition of $\xi$ we conclude that $\Delta(A,B)=\tilde{\Delta}(A,B)$ for all $A, B \in \Gamma_{c} \omega J\Phi$, i.e. $\Delta$ is equal to $\tilde{\Delta}$. Reciprocally each map in $Prop'(E)$ induces a morphism of $\mathcal{C}^{\infty}(M)^{\otimes 2}$-modules between $\Gamma_{c}J\Phi\hat{\otimes}_{K}\Gamma_{c}J\Phi$ and $\Gamma(\omega_{\mathcal{M}\times \mathcal{M}})'$, then $\xi$ is bijective. There is an important family of propagators that we want to use. Given a vector space $V$, we say that a set $C\subset V$ is a cone if $\xi \in C$ implies that $t\xi\in C$ for all $t>0$. If we are working in a t.v.s. we can talk about closed cones. A cone which is properly contained in a semispace is called proper. Note that a semispace is not a proper cone. Given a closed proper convex cone $C$ of a t.v.s. $Z$, we can define a partial order on $Z$ of the form $x\curlyeqprec y$ if $y-x\in C$. \[defcut\] A propagator $\Delta:\Gamma_{c} \omega J\Phi \times \Gamma_{c}\omega J\Phi \to \mathbb{C}$ is called *cut*, if for each $\varphi, \psi \in\Gamma_{c}J\Phi$ and each $z\in\mathcal{M}$ the H-distribution over $\mathcal{M}\times\mathcal{M}$: $\xi(\Delta)(\varphi,\psi)$ possesses a partial order in the cotangent space defined by a proper closed convex cone $C_{z}$, such that if $(p,q)$ is in the wavefront set of $\xi(\Delta)(\varphi,\psi)$ at some point $(x,y)\in \mathcal{M}\times \mathcal{M}$ then $ p\curlyeqprec 0 $ and $0\curlyeqprec q $. Moreover if $x=y$ then $p+q=0$. \[ejadv\] In Quantum Field Theory one often encounters the so-called advanced propagator $\Delta_{(+)}$ such that its associated element in $Prop'(E)$. If it is evaluated in the scalar field, that is $\xi(\Delta_{(+)})(\phi,\phi)$, (see [@foll] chapter 6) in local coordinates is given by the function $$\Delta_{+}(x)=\frac{i}{2(2\pi)^{3}}\int_{\mathbb{R}^{3}}\frac{e^{-ip_{\mu}x^{\mu}}}{2p_{0}}dp_{1}dp_{2}dp_{3},$$ where $p_{0}=\sqrt{|\vec{p}|^{2}+m^{2}}$. This means that the H-distribution $\xi(\Delta_{(+)})(\phi,\phi)$ evaluated in two compact support densities $f(x)d^{4}x$ and $g(y)d^{4}y$ is given by $$\xi(\Delta_{(+)})(\phi,\phi)(f(x)d^{4}x,g(y)d^{4}y)=-i\int_{supp(f)\times supp(g)} \Delta_{+}(x-y)f(x)g(y)d^{4}xd^{4}y.$$ The wavefront set of this distributions was calculated in [@reed], Theorem IX.48, and it gives $$\label{wfffffff} WF(\Delta_{+})= \{ (0,-|\vec{p}|,\vec{p})/ 0\in \mathbb{R}^{4}, \vec{p}\in \mathbb{R}^{3}\setminus \{0\} \}\cup \{ (\pm |\vec{x}|,\vec{x},-\lambda |\vec{x}|,\mp\lambda\vec{x})/ \vec{x}\in\mathbb{R}^{3},\lambda >0 \}$$ By pulling back this distribution with the map $\mathbb{R}^{4}\times \mathbb{R}^{4}\to \mathbb{R}^{4}$, $(x,y)\mapsto x-y$ we obtain the wavefront set of $\xi(\Delta_{(+)})(\varphi,\varphi)$ over $\mathbb{M}^{4}\times \mathbb{M}^{4}$, which is $$\label{ggg} WF(\xi(\Delta_{(+)})(\varphi,\varphi))=\{ (x,p,y,-p)\in \mathbb{R}^{16}/ (x-y,p)\in WF(\Delta_{+})\}.$$ \[fey\] Consider the distribution $\theta_{x^{0}}$ on $\mathbb{M}^{4}$ defined by $\theta_{x^{0}}(f)=\int_{x^{0}\geqslant0} f(x) d^{3}\vec{x}$. In this case $$\label{wafff} WF(\theta_{x^{0}})= \{ (0,\vec{x},\lambda,\vec{0})/ \vec{x}\in \mathbb{R}^{3}, \lambda \in \mathbb{R}\setminus\{0\}\}$$ (see [@ma]). Since $WF(\theta_{x^{0}})$ and $WF(\Delta_{+})$ satisfy that there is no point $(x,p)\in WF(\theta_{x^{0}})$ such that $(x,-p)\in WF(\Delta_{+})$, because the only $x$ which is in both singular supports is $0$ but in this case the cones has not intersections (see formulas and ), we can define their product $\theta_{x^{0}} \Delta_{+}$ (see [@reed]). The Feynman propagator is given by $$\Delta_{F}(x)=\theta_{x^{0}} \cdot \Delta_{+}(x)+\theta_{-x^{0}} \cdot \Delta_{+}(-x).$$ The wavefront set of this distribution was computed in [@viet], Proposition 26. If we pull back the Feynman propagator by the diagonal map we obtain an H-distribution on $\mathbb{M}^{4}\times \mathbb{M}^{4}$. A propagator $\Delta:\Gamma_{c} \omega J\Phi \times \Gamma_{c}\omega J\Phi \to \mathbb{C}$ is called *polynomially smooth* if the expression in local coordinates of $\xi(\Delta)(A,B)$ is a sum of products of smooth functions and powers of logarithms of polynomials, for any $A,B\in \Gamma_{c}J\Phi$. We recall that the Feynman propagator of the massive scalar field described in Example \[fey\] is $$\Delta_{F}(t,\vec{x})=i\int\frac{e^{-i\omega_{p}|t|+i\vec{p\cdot \vec{x}}}}{2\omega_{p}}\frac{d^{3}\vec{p}}{(2\pi)^{3}},$$ where $\omega_{p}=\sqrt{|\vec{p}|^{2}+m^{2}}$. The integral can be explicitly computed (see [@foll] Chapter 6, Section 5) and it gives $$\Delta_{F}(t,\vec{x})=\frac{im}{4\pi^{2}\sqrt{|\vec{x}|^{2}-t^{2}}}K_{1}\Big( m\sqrt{|\vec{x}|^{2}-t^{2}} \Big) \hspace{.8cm} on \hspace{.8cm} \mathbb{R}^{4}\setminus \{ (t,\vec{x})/ |t|=|\vec{x}| \},$$ where $K_{1}$ is the modified Bessel function of order $1$. Then $\Delta_{F}$ is polynomially smooth. For more examples of smooth polynomial propagators see the Dirac and the Proca propagators in Chapter 6, Section 5 of [@foll]. ### Generalized propagators We now proceed to extend the propagator to a bigger domain. We recall that a propagator $\Delta$ gives a bilinear map $\xi(\Delta):\Gamma_{c}J\Phi\times \Gamma_{c}J\Phi\to\Gamma(\omega_{\mathcal{M}\times \mathcal{M}})'$ (see Remark \[khkg\]). We first define $\tilde{\Delta}:\Gamma_{c} SJ\Phi \times \Gamma_{c} SJ\Phi \rightarrow \Gamma(\omega_{\mathcal{M}\times \mathcal{M}})'$ by means of the formula $$\label{223} \tilde{\Delta}(a_{1}\odot...\odot a_{n},b_{1}\odot... \odot b_{n})=\sum_{\sigma\in S_{n}}\xi(\Delta)(a_{1},b_{\sigma{1}})\cdot \xi(\Delta)(a_{2},b_{\sigma_{2}})\cdot ...\cdot \xi(\Delta)(a_{n},b_{\sigma_{n}}),$$ where the $\cdot$ is a product of H-distributions (recall the product of H-distribution is described in [@ma], Chapter 2) and we define $\tilde{\Delta}$ to be zero if it is evaluated at different degree elements. If $V=\Gamma SJ\Phi$, then $SV$ admits two structures of coalgebra. The first one is the cocommutative cofree coaugmented symmetric coalgebra of $V$, with this structure the elements of $V$ are primitives. Let us call $\Delta_{C}$ to the coproduct of this coalgebra. For the other hand if we make the construction described in \[symme\] to the object $J\Phi$, we obtain the symmetric coalgebra $SJ\Phi$, in which the elements of $J\Phi$ are primitive elements. This structure has a unique extension to the algebra $SV$, such that $SV$ is a Hopf algebra (see [@tak]). Let us denote by $\Delta_{T}$ the coproduct of $SV$ with this structure. If $\varphi, \psi \in J\Phi$, then $\varphi\psi\in S^{2}J\Phi\subseteq S^{1}\Gamma S^{2}J\Phi\subseteq SV$ (we will denote with an underline the elements of $J\Phi$ viewed as elements of $S^{1}J\Phi$) and, $$\Delta_{C}(\underline{\varphi\psi})=1\otimes \underline{\varphi\psi}+\underline{\varphi\psi}\otimes 1$$ but $$\begin{aligned} & \Delta_{T}(\underline{\varphi\psi})=\Delta_{T}(\underline{\varphi})\cdot \Delta_{T}(\underline{\psi})\\ & =(1\otimes \underline{\varphi}+ \underline{\varphi}\otimes 1) \cdot (1\otimes \underline{\psi}+\underline{\psi}\otimes 1)\\ & =1\otimes \underline{\varphi\psi}+\underline{\psi}\otimes \underline{\varphi}+\underline{\varphi}\otimes \underline{\psi}+\underline{\varphi\psi}\otimes 1.\end{aligned}$$ In the following when we refer to the coproduct of an element in $SV$, it will be in the sense of $\Delta_{T}$. Finally, we extend $\tilde{\Delta}$ to a map $\hat{\Delta}$ from $S^{m}\Gamma_{c} SJ\Phi \times S^{n}\Gamma_{c} SJ\Phi$ to H-distributions of compact support on $\mathcal{M}^{m}\times \mathcal{M}^{n}$ by the recursive expression $$\label{magenta} \hat{\Delta}(A B,C)=\sum \hat{\Delta}(A,C')\otimes\hat{\Delta}(B,C''),$$ $$\hat{\Delta}(A,1)=\varepsilon(A),$$ where $\sum C'\otimes C''$ is the coproduct of $C$ in $S^{n}\Gamma_{c} SJ\Phi $, and $\hat{\Delta}(A,C')\otimes\hat{\Delta}(B,C'')$; an H-distribution on $\mathcal{M}^{m}\times \mathcal{M}^{n}$. Note that is precisely the definition of a Laplace pairing (see [@su], Definition 10.2). As an example suppose $A$, $B$ and $C$ are Lagrangian fields of compact support, i.e. $A,B,C\in \Gamma_{c}SJ\Phi$, then $AB\in S^{2}\Gamma_{c}SJ\Phi$ and $C\in S^{1}\Gamma_{c}SJ\Phi$. Then the distribution $\hat{\Delta}(A B,C)$ must be evaluated in densities on $\mathcal{M}^{2}\times \mathcal{M}$. Assume $\alpha,\beta,\gamma\in \Gamma \omega$, then $(\alpha\otimes \beta)\otimes \gamma$ is a density over $\mathcal{M}^{2}\times \mathcal{M}$ and the definition of $\hat{\Delta}$ indicates the following, $$\label{gfgfgf} \hat{\Delta}(A B,C)((\alpha\otimes \beta)\otimes \gamma)=\sum \tilde{\Delta}(A,C')(\alpha\otimes \gamma)\cdot\tilde{\Delta}(B,C'')(\beta\otimes \gamma)$$ where we put $\tilde{\Delta}$ in the right hand side, instead of $\hat{\Delta}$, because when we restrict to elements of degree zero or one it takes the same values, and $A,B,C',C''\in S^{\leqslant 1}\Gamma_{c}SJ\Phi$. The product $\cdot$ in , is a product of complex numbers. With this successive extensions of a given propagator $\Delta$, we define $\hat{\Delta}$ whose restrictions $\hat{\Delta}|_{ S^{m}\Gamma_{c}S J\Phi \times S^{n}\Gamma_{c} SJ\Phi}$ satisfy the following. A *generalized propagator* is a family $\hat{\Delta}=\{\Delta_{m,n}\}_{m,n\in\mathbb{N}}$ of continuous and $\mathbb{R}$-bilinear functions $$\Delta_{m,n}: S^{m}\Gamma_{c}S J\Phi \times S^{n}\Gamma_{c} SJ\Phi \longrightarrow H_{d}(\mathcal{M}^{m}\times \mathcal{M}^{n})$$ where the topology for the codomain is the weak-$\ast$ topology (see [@ru], Chapter 3, Section 11 and Chapter 6 for a general description of the weak-$\ast$ topology and Section 16 for the specific case $\mathcal{M}^{m}\times \mathcal{M}^{n}=\mathbb{R}^{k}$). \[newcut\] A generalized propagator $\Delta=\{\Delta_{m,n}\}_{m,n\in\mathbb{N}}$ is called *cut* if for all $x\in \mathcal{M}$ there is a partial order in the cotangent space of $\mathcal{M}$ at $x$ defined by a proper closed convex cone $C_{x}$, such that for all $(m,n)\in \mathbb{N}^{2}$ and $(p_{1},\cdots,p_{m},q_{1},\cdots,q_{n})$ in the wavefront set of $\Delta_{m,n}$ at the point $(x_{1},\cdots,x_{m},y_{1},\cdots,y_{n})$ all the $p_{i} \curlyeqprec 0$ and all the $0 \curlyeqprec q_{j}$. And if $(p_{1},\cdots,p_{m},q_{1},\cdots,q_{n})$ is in the wavefront set of $\Delta_{m,n}$ on the diagonal of $\mathcal{M}^{m}\times \mathcal{M}^{n}$ then $p_{1}+\cdots+p_{m}+q_{1}+\cdots+q_{n}=0$. Feynman measure --------------- We are interested in integrating expressions of the form $e^{iL_{F}}\mathcal{L}$, where $\mathcal{L}$ is a lagrangian density. Let us define a collection of maps $\chi_{n}:(S\Gamma_{c}\omega SJ\Phi)'\times S^{n}\Gamma_{c}\omega SJ\Phi\to \mathcal{C}^{\infty}(\mathcal{M}^{n})'$ for $n\in \mathbb{N}$ by $\chi_{n}(\delta,A_{1}\odot \dots \odot A_{n})(f_{1}\otimes \cdots \otimes f_{n})=\delta(f_{1}A_{1}\odot \cdots \odot f_{n}A_{n})$, where $f_{i}\in \mathcal{C}^{\infty}(\mathcal{M})$ and $A_{i}\in \Gamma_{c}\omega SJ\Phi$, $\forall i=1,\cdots, n$. Since the subspace of $\mathcal{C}^{\infty}(\mathcal{M})$ formed by sums of functions of the form $f_{1}\otimes \cdots \otimes f_{n}$ is dense, the maps $\chi_{n}$ are well-defined. Set $$\chi: (S\Gamma_{c}\omega SJ\Phi)'\times S\Gamma_{c}\omega SJ\Phi\to \oplus_{n\in \mathbb{N}}\mathcal{C}^{\infty}(\mathcal{M}^{n})'$$ the direct sum of those maps. \[suavidad en la diagonal\] Let $\delta: S\Gamma_{c}\omega S J\Phi\rightarrow \mathbb{C} $ be a continuous and linear map. We say that it is *smooth on the diagonal* if $(p_{1},...,p_{n})\in \Sigma_{(q,...,q)}\chi_{n}(\delta,A)$ implies that $p_{1}+\cdots+p_{n}=0$, for all $A\in S^{n}\Gamma_{c}\omega SJ\Phi$. We are now ready to introduce the notion of Feynman measure. \[dia\] A Feynman measure is a continuous linear map from $S\Gamma_{c}\omega S J\Phi$ to $\mathbb{C}$. Let $\delta:S\Gamma_{c}\omega S J\Phi\rightarrow \mathbb{C}$ be a continuous linear map and let $\Delta:\Gamma_{c}\omega J\Phi\times \Gamma_{c}\omega J\Phi \to \mathbb{C}$ be a propagator. Then, $\delta$ is said to be associated to the propagator $\Delta$ if it satisfies the following conditions: 1. $\delta$ is smooth in the diagonal; 2. (non-degeneracy) There is a smooth nowhere vanishing function $g$ so that $\delta(v)=\int_{\mathcal{M}}gv$ for all $v\in S^{1}\Gamma_{c}\omega S^{0}J\Phi=\Gamma_{c}\omega$; 3. (Gaussian condition or weak translational invariance) Let $A\in S^{m}\Gamma_{c}\omega S J \Phi$ and $B\in S^{n}\Gamma_{c}\omega S J \Phi$ such that there is no point in $supp(A)$ that is $\leqslant$ to some point in the $supp(B)$. This means that there is not point at $supp(A)$ which is $\leqslant$ to some point at $supp(B)$, we actually should say that there is no point in $supp(A)\subseteq \mathcal{M}^{m}$ such that any of its coordinate be $\leqslant$ to some coordinate of some point in $supp(B)\subseteq \mathcal{M}^{n}$. Then the following equality holds $$\label{gauss} \chi(\delta,A\cdot B)=\sum (\chi(\delta,A')\otimes \chi(\delta,B')) \cdot \hat{\Delta}(A'',B'')$$ where $\sum A'\otimes A''\in S\Gamma_{c}\omega SJ\Phi \otimes S\Gamma_{c}SJ\Phi$ is the image of $A$ given by the coaction $S^{m}\Gamma_{c}\omega SJ\Phi \rightarrow S^{m}\Gamma_{c}\omega SJ\Phi \otimes S^{m}\Gamma_{c}SJ\Phi$. The product in will be regarded in the following manner. Given $A''\in S^{m}\Gamma_{c}SJ\Phi$ and $B''\in S^{n}\Gamma_{c}SJ\Phi$, $\Delta(A'',B'')$ defines an element of $H_{d}(\mathcal{M}^{m}\times \mathcal{M}^{n})$ that can be multiplied by $\chi(\delta,A')\otimes \chi(\delta,B')\in D_{d}(\mathcal{M}^{m}\times \mathcal{M}^{n})$ using the product in \[hd\]. \[coac\] Let us explain the coaction involved in the Gaussian condition in more detail. By using the Theorem \[df\] and its Corollary \[dm\], $\Gamma_{c}\omega SJ\Phi$ can be written in the form $\Gamma_{c}\omega\otimes_{\mathcal{C}^{\infty}(\mathcal{M})} \Gamma SJ\Phi$. Then by using Proposition \[prop 1\] we conclude that $\Gamma_{c} \omega SJ\Phi$ is a comodule over $\Gamma SJ\Phi$ (call $\rho$ the coaction), and consequently $S\Gamma_{c} \omega SJ\Phi$ is a comodule over $S\Gamma SJ\Phi$ and then it has sense to talk about a coaction. In Sweedler notation if $\mathcal{L}\in \Gamma\omega SJ\Phi$ then $\rho(\mathcal{L})=\sum_{()}\mathcal{L}_{(0)}\otimes \mathcal{L}_{(1)}$ where $\mathcal{L}_{(0)}\in \Gamma\omega SJ\Phi$ and $\mathcal{L}_{(1)}\in \Gamma SJ\Phi$, then we can define a coaction $\rho^{n}:S^{n}\Gamma\omega SJ\Phi\to S^{n}\Gamma \omega SJ\Phi \otimes S^{n}\Gamma SJ\Phi$ that over sections works as $$\rho^{n}(\mathcal{L}_{1}\odot\cdots\odot\mathcal{L}_{n})=\sum_{()} (\mathcal{L}_{1(0)}\odot \cdots \odot \mathcal{L}_{n(0)}) \otimes (\mathcal{L}_{1(1)}\odot \cdots \odot \mathcal{L}_{n(1)})$$ where $\mathcal{L}_{j(0)}\in \Gamma \omega SJ\Phi $ for all $1\leqslant j \leqslant n$ and $\mathcal{L}_{k(1)}\in \Gamma SJ\Phi$ for all $1\leqslant k \leqslant n$, note that $\mathcal{L}_{1(1)}\odot \cdots \odot \mathcal{L}_{n(1)}$ belongs to $\Gamma SJ\Phi$. We denote by $\sigma$ the coaction induced by $\oplus_{n\in \mathbb{N}_{0}}\rho^{n}$ which makes of $S\Gamma\omega SJ\Phi$ a $\Gamma SJ\Phi$-comodule. \[obs4\] Provided our propagator $\Delta$ is cut, the smoothness on the diagonal allows us to compute the product at the right hand of . Indeed, suppose we have $(p_{1},\cdots ,p_{m},q_{1},\cdots q_{n})$ in the wave front set of $\Delta(A'',B'')$ at a point $(x,\cdots, x,y,\cdots,y)$, then $p_{i}\curlyeqprec 0$ and $0\curlyeqprec q_{j}$ for all $1\leqslant i \leqslant m$ and $1\leqslant j \leqslant n$. If $(-p_{1},\cdots ,-p_{m},-q_{1},\cdots -q_{n})$ is in the wave front set of $\chi(\delta,A')\otimes \chi(\delta,B')$ Lemma 2.175 in [@ma] implies that $q_{j}=0$ for all $1\leqslant j \leqslant n$ or $(-q_{1},\cdots ,-q_{n})$ is in the wave front set of $\chi(\delta,B')$ at $(x,\cdots,x)$. Consequently $-q_{1}-q_{2}\cdots-q_{n}=0$ by the smoothness on the diagonal condition of the Feynman measure. In other words we get that $q_{1}+\cdots +q_{n}=0$. This condition together with the previous one stating that $0 \curlyeqprec q_{i}$ (i.e. $q_{i}$ is in the proper convex closed cone $C_{x}$) implies that all the $q_{j}$ must be zero. A similar argument can be used to see that all the $p_{i}$ are zero. This is absurd, for $(p_{1},\cdots ,p_{m},q_{1},\cdots q_{n})$ is in the wave front set of $\Delta(A'',B'')$. The contradiction came from the assumption that $(-p_{1},\cdots ,-p_{m},-q_{1},\cdots -q_{n})$ was in the wave front set of $\chi_{m}(\delta,A')\otimes \chi_{n}(\delta,B')$. As a consequence, we conclude that the product between these distributions is well defined. The idea is to think of the Feynman measure in $A$ as the integral of $A$ over all the classical fields, that is $\delta(A)=\int_{\Phi} A(\varphi)\mathcal{D}\varphi$. \[obs5\] Given $\Delta$ a Feynman measure and $A\in S^{n}\Gamma_{c}\omega S J\Phi$, then $\delta(A)$ is a complex number and $\chi_{n}(\delta,A)$ is an element of $D_{d}(\mathcal{M}^{n})$. The latter can be thought of a density over $\mathcal{M}^{n}$ in the sense that the densities take functions and assign numbers (by integration over $\mathcal{M}$), i.e. $\chi_{n}(\delta,A)(f)=\delta(fA)$ for $f\in \mathcal{C}^{\infty}(\mathcal{M}^{n})$. By using the notation for the complex number $\delta(A)=\chi_{n}(\delta,A)(1)$, where $1$ is the function identically $1$ over $\mathcal{M}^{n}$, and as a consequence of this, if we want to see that $\delta(A)=\delta(B)$ for two elements $A,B\in S^{n}\Gamma_{c}\omega S J\Phi$ then we need only to examine $\chi_{n}(\delta,A)$ and $\chi_{n}(\delta,B)$ as elements of $D_{d}(\mathcal{M}^{n})$, and lastly conclude the equality by evaluating in the function $1$, $\chi_{n}(\delta,A)(1)=\chi_{n}(\delta,B)(1)$ or $\delta(A)=\delta(B)$ as complex numbers. As a conclusion, if we want to prove that $\delta(A)=\delta(B)$ for two elements $A, B\in S\Gamma_{c}\omega S J\Phi$ then it is sufficient to prove it for $\chi(\delta,A)$ and $\chi(\delta,B)$. A last comment concerning the physics picture of the Feynman measure is that the smoothness on the diagonal is exactly the conservation of the momentum, because the momentum coordinates are the coordinates of the Fourier transform of a field. Renormalization =============== Characterization of the renormalization group --------------------------------------------- As we saw in the precedent chapter the sheaf $S\omega SJ\Phi$ has a structure of comodule over $SJ\Phi$. A *renormalization* is an automorphism of the sheaf of coalgebras $S\omega SJ\Phi$ which preserves the coaction of $SJ\Phi$. The set of renormalizations is a group under composition and is called the renormalization group. The next theorem characterizes the renormalization group in a nice manner. \[teorema1\]The elements of the renormalization group are in correspondence with the elements in $Hom(S\omega SJ\Phi,\omega)$ that are zero over $S^{0}\omega SJ\Phi$ and are isomorphisms when they are restricted to $S^{1}\omega S^{0}J\Phi=\omega$. By Proposition \[prop4\] there is a correspondence between the coalgebra automorphisms $R\in Hom_{Coalg}(S\omega SJ\Phi, S\omega SJ\Phi) $ and the morphisms $r\in Hom_{\mathcal{C}^{\mathbb{R}}-Mod}(S\omega SJ\Phi,\omega SJ\Phi)$ whose sequential representation satisfies $r_{0}=0$ and $r_{1}$ is an isomorphism. At the beginning of this chapter we mentioned that $S\omega SJ\Phi$ is a comodule over $SJ\Phi$. We use Proposition \[prop5\] in the monoidal abelian category of sheafs over the ringed space $(\mathcal{M},\mathcal{C}^{\mathbb{R}})$, taking $D$ as the sheaf $SJ\Phi$, $M$ as the sheaf $S\omega SJ\Phi$ of $SJ\Phi$-comodules and $W$ as the sheaf $\omega$. It gives a correspondence between the comodule morphisms $r\in Hom_{Com_{SJ\Phi}}(S\omega SJ\Phi,\omega SJ\Phi)$ and the morphisms $\eta\in Hom_{\mathcal{C}^{\mathbb{R}}-Mod}(S\omega SJ\Phi,\omega)$. If we begin this process with an $R\in Hom_{Coalg}(S\omega SJ\Phi, S\omega SJ\Phi) $ which is a renormalization, then it is a coalgebra morphism and also a comodule morphism so we successively apply both correspondences to obtain the desired total correspondence between the renormalizations and the subset of $\eta\in Hom_{\mathcal{C}^{\mathbb{R}}-Mod}(S\omega SJ\Phi,\omega)$ such that $\eta_{0}=0$, $\eta_{1}= (Id_{\omega}\otimes \varepsilon_{SJ\Phi})\circ r_{1}|_{\omega S^{0}J\Phi}$ is an isomorphism (because $r_{1}$ is). The structure of the renormalization group ------------------------------------------ Notice that the renormalization group (which we will denote by $G$) preserves the increasing filtration $S^{\leqslant n}\omega SJ\Phi\subseteq S^{\leqslant n+1}\omega SJ\Phi$, because if we apply $P\in G$ to an element $v\in S^{\leqslant n}\omega SJ\Phi$, from equation is clear that $P(v)\in S^{\leqslant n}\omega SJ\Phi$. If $\mathcal{C}$ is a monoidal abelian subcategory of $_{A}Mod$ or $\mathcal{C}^{\mathbb{R}}$-Modules, we denote by $G'$ the group of automorphisms of the cocommutative coaugmented coalgebra $S^{c}V$ (with $V\in Obj(\mathcal{C})$) and $G'_{> a}=\{ g\in G'/g(\alpha)=\alpha, \forall \alpha \in S^{\leqslant a}V\}$, where $a\in \mathbb{N}_{0}$. Notice that $G'_{>0}$ is exactly $G'$. Given $a\in \mathbb{N}$ and $P\in G'$. Then $P\in G'_{>a}$ if and only if the sequential representation of $P$ satisfies $p_{1}=Id_{V}$ and $p_{n}=0$, $\forall 2\leqslant n \leqslant a$. The proof follows directly from equation . If we apply the formula for $n=1$ and $P\in G'_{>a}$, then we obtain: $$v=P(v)=\sum_{m=1}^{1}\sum_{I_{1}\neq \emptyset; I_{1}=\{1\} }\frac{1}{1!}p_{1}(v_{I_{1}})=p_{1}(v),$$ that is $p_{1}$ is the identity of $V$. Let us do the same for $n=2$ (supposing that $a\geqslant 2$), in this case the equation gives us, $$v_{1}\cdot v_{2}=\frac{1}{2}p_{1}(v_{1})\cdot p_{1}(v_{2})+\frac{1}{2}p_{1}(v_{2})\cdot p_{1}(v_{1})+p_{2}(v_{1}\cdot v_{2}),$$ where we use the previous result, $p_{1}(v_{1})\cdot p_{1}(v_{2})=v_{1}\cdot v_{2}$, and $P(v_{1}\cdot v_{2})=v_{1}\cdot v_{2}$ and so $p_{2}(v_{1}\cdot v_{2})=0$. Let us proceed by induction. Given $j\in \mathbb{N}$ such that $2\leqslant j\leqslant a$ and supposing that $p_{k}=0$ for all $k\in \mathbb{N}$ satisfying $2\leqslant k<j$ and $p_{1}=Id_{V}$, we write the equation (taking $n=j$), $$\begin{aligned} & v_{1}\cdots v_{j}=P(v_{1}\cdots v_{j})=\sum_{\text{all permutations}} \frac{1}{j!}p_{1}(v_{1})\cdots p_{1}(v_{j}) \\ & +\sum_{\text{terms with $p_{k}$ with $2\leqslant k<j$}}\cdots +p_{j}(v_{1}\cdots v_{j}) \\ & =j!\frac{1}{j!}v_{1}\cdots v_{j}+0+\cdots+0+p_{j}(v_{1}\cdots v_{j}),\end{aligned}$$ hence $p_{j}(v_{1}\cdots v_{j})=0$ for all $v_{1}\cdots v_{j}\in S^{j}V$.\ Reciprocally if the sequential representation of $P$ satisfies $p_{1}=Id_{V}$ and $p_{n}=0$ for all $2\leqslant n \leqslant a$ then if we apply $P$ to $v_{1}\cdots v_{n}\in S^{n}V$ by means of equation , only the terms such that $|I_{j}|=1$ survive, that is when $m=n$. Hence, $$\begin{aligned} & P(v_{1}\cdot v_{2}\cdots v_{n})=\sum_{m=1}^{n}\sum_{I_{1}\cdots I_{m}\neq \emptyset; I_{1}\sqcup \cdots \sqcup I_{m}=\{1\cdots n\} }\frac{1}{m!}p_{|I_{1}|}(v_{I_{1}})\cdots p_{|I_{m}|}(v_{I_{m}})\\ & =\sum_{I_{1}\cdots I_{n}\neq \emptyset; I_{1}\sqcup \cdots \sqcup I_{n}=\{1\cdots n\} }\frac{1}{n!}p_{1}(v_{I_{1}})\cdots p_{1}(v_{I_{n}})\\ & =\sum_{I_{1}\cdots I_{n}\neq \emptyset; I_{1}\sqcup \cdots \sqcup I_{n}=\{1\cdots n\} }\frac{1}{n!}Id_{V}(v_{I_{1}})\cdots Id_{V}(v_{I_{n}})\\ & =\sum_{I_{1}\cdots I_{n}\neq \emptyset; I_{1}\sqcup \cdots \sqcup I_{n}=\{1\cdots n\} }\frac{1}{n!}v_{1}\cdots v_{n}=\frac{n!}{n!}v_{1}\cdots v_{n}=v_{1}\cdots v_{n},\end{aligned}$$ where we use the fact that there are $n!$ forms of take intervals $I_{i}$ such that $|I_{i}|=1$ and $I_{1}\sqcup \cdots \sqcup I_{n}=\{ 1,\cdots, n \}$. If $P\in G'$, we call $p$ its sequential representation and usually say that $p\in G'$. \[bnbn\] \[sub\]$G'_{>a}$ is a subgroup of $G'$ for all $a\in \mathbb{N}$ By composing with $\pi:S^{c}V\rightarrow V$ it is clear that $1_{G'}$ has the sequential representation $\{Id_{V},0,0,\cdots \}$, then by the precedent proposition $1_{G'}\in G_{>a}$ for all $a\in \mathbb{N}$. If the sequential representation of $P$ and $Q\in G'$ are $\{p_{n}\}_{\mathbb{N}}$ and $\{q_{n}\}_{\mathbb{N}}$ respectively, then the product representation of $P\circ Q$ has sequential representation $\{p_{n}\}_{\mathbb{N}}\cdot \{q_{n}\}_{\mathbb{N}}=\{(p\cdot q)_{n}\}_{\mathbb{N}}$ given by the following formula (see equation ), $$(p\cdot q)_{n}(v_{1}\cdots v_{n})=\sum_{m=1}^{n}\frac{1}{m!}\sum_{I_{1}\cdots I_{m}\neq \emptyset; I_{1}\sqcup \cdots \sqcup I_{m}=\{1\cdots n\} }p_{m}(q_{|I_{1}|}(v_{I_{1}})\cdots q_{|I_{m}|}(v_{I_{m}})).$$ With this formula it is possible to show that $p\cdot q\in G'_{>a}$ provided that $p$ and $q$ are there. For example if $2\leqslant n \leqslant a$, having in account that $q_{k}=0$ for all $2\leqslant k\leqslant a$, we see that in the sum only the terms with all $q_{1}$ survive, i.e. $m=n$. Also as $q_{1}$ is the identity: $q_{1}(v_{I_{1}})\cdots q_{1}(v_{I_{n}})=v_{I_{1}}\cdots v_{I_{n}}$, but $v_{I_{i}}=v_{j_{i}}$ because $|I_{i}|=1$ when $m=n$, then $q_{1}(v_{I_{1}})\cdots q_{1}(v_{I_{n}})=v_{j_{1}}\cdots v_{j_{n}}$. But as $\cdot$ is a symmetric product then all the terms $v_{j_{1}}\cdots v_{j_{n}}$ are equal, so there are $n!$ summands which are equal and consequently the last equality can be written as $$(p\cdot q)_{n}(v_{1}\cdots v_{n})=n!\frac{1}{n!}p_{n}(v_{1}\cdots v_{m}).$$ This is zero because $p_{n}=0$. From the same equation , but this time taking $n=1$, we can conclude $(p\cdot q)_{1}=Id_{V}$. Finally if $p\in G'_{>a}$, we take $q=p^{-1}$ in $G'$. Suppose $q\notin G'_{>a}$, and let $n_{0}$ be the first natural $\geqslant 2$ such that $q_{n_{0}}\neq 0$. Equation show us that $q_{1}=p_{1}^{-1}$ (taking $n=1$). But $q\notin G'_{>a}$, then $n_{0}\leqslant a$. Then by using again the equation taking $n=n_{0}$ and using the fact that $\{(p\cdot q)_{n}\}_{\mathbb{N}}=\{Id_{V},0,0\cdots \}$ we obtain $0=p_{1}(q_{n_{0}}(v_{1}\cdots v_{n_{0}}))$. But as $p_{1}=Id_{V}$ the equation turns to: $0=q_{n_{0}}(v_{1}\cdots v_{n_{0}})$ i.e. $q_{n_{0}}=0$, which is absurd. Hence $q\in G'_{>a}$. Notice that the precedent proposition is trivially valid for the case $a=0$. For $a\geqslant 2$ we define the sets $$\begin{aligned} & G'_{a}=\{P\in G'/ p=\{Id_{V},0\cdots,0,p_{a},0\cdots\}\},\\ \text{and}\hspace{1cm}\\ & G'_{1}=\{P\in G'/ p=\{p_{1},0\cdots\}\}\end{aligned}$$ were we use the just described notation of sequential representation introduced in Chapter 1. An element in $P\in G'_{a}$ will be an automorphism $P\in G'$ such that when we restrict $\pi\circ P$ to $S^{n}V$ it is zero ($\pi:S^{n}V\to V$ is the canonical projection), and less $n=1$ or $n=a$. The next proposition is proved by induction on $a$. \[prop8\]Let $a\in \mathbb{N}$ and $P\in G'$. There are unique $P_{1}, P_{2},\cdots P_{a}, P'\in G'$ such that $P_{i}\in G'_{i}$, $P'\in G'_{>a}$ and $P=P_{1}\circ P_{2}\cdots P_{a}\circ P'$ The proof is deduced from the next lemma: \[lema1\]Given $a\in \mathbb{N}$ and $P\in G'_{>a}$ there exist unique $P_{a+1}$ and $P'\in G'$ such that $P_{a+1}\in G'_{a+1}$, $P'\in G'_{>a+1}$ and $P=P_{a+1}\circ P'$. We give the proof for the case $a\geqslant 2$, the case $a=1$ is similar. If the representation of $P$ is given by $\{Id_{V},0\cdots 0,p_{a+1},p_{a+2}\cdots \}$, we take $P_{a+1}$ as the only one whose sequential representation is $\tilde{p}=\{Id_{V},0\cdots p_{a+1},0 \cdots \}$ and $P'$ whose representation is $\{(\tilde{p}^{-1}\cdot p)_{n}\}_{n\in \mathbb{N}}$. By using the equation it is not difficult to see that $(\tilde{p}^{-1}\cdot p)_{1}=Id_{V}$, and as $\tilde{p},p\in G'_{>a}$ (which is a group as we saw in Proposition \[sub\]) then $(\tilde{p}^{-1}\cdot p)_{n}=0$ for all $2\leqslant n \leqslant a$. By means of and computing $\tilde{p}^{-1}$ up to degree $a+1$ with , we can see that $\tilde{p}^{-1}=\{Id_{V},0\cdots 0,-p_{a+1},\cdots\}$ and $(\tilde{p}^{-1}\cdot p)_{a+1}=0$, so $\tilde{p}^{-1}\cdot p \in G'_{>a+1}$. Let us see in detail that, if $\tilde{p}=\{Id_{V},0\cdots p_{a+1},0 \cdots \}$ then its inverse satisfies $(\tilde{p}^{-1})_{1}=(\tilde{p}_{1})^{-1}=Id_{V}^{-1}=Id_{V}$ and then by using the equation for $2\leqslant k\leqslant a$ we have $$\begin{aligned} & (\tilde{p}^{-1})_{k}(v_{1}\cdot v_{2}\cdots v_{k})\\ & =-\sum_{m=2}^{k}\frac{1}{m!}\sum_{I_{1}\cdots I_{m}\neq \emptyset ; I_{1}\sqcup \cdots \sqcup I_{m}=\{1\cdots k\} }(\tilde{p}^{-1})_{1}(\tilde{p}_{m}((\tilde{p}^{-1})_{|I_{1}|}(v_{I_{1}})\cdots (\tilde{p}^{-1})_{|I_{m}|}(v_{I_{m}}))),\\\end{aligned}$$ but as we say $(\tilde{p}^{-1})_{1}$ is the identity and then $$\begin{aligned} & (\tilde{p}^{-1})_{k}(v_{1}\cdot v_{2}\cdots v_{k})\\ & =-\sum_{m=2}^{k}\frac{1}{m!}\sum_{I_{1}\cdots I_{m}\neq \emptyset ; I_{1}\sqcup \cdots \sqcup I_{m}=\{1\cdots k\} }\tilde{p}_{m}((\tilde{p}^{-1})_{|I_{1}|}(v_{I_{1}})\cdots (\tilde{p}^{-1})_{|I_{m}|}(v_{I_{m}})),\\\end{aligned}$$ which is zero because all the $\tilde{p}_{m}$ has $m\leqslant k\leqslant a$, and then $\tilde{p}_{m}=0$. Lastly if we take $n=a+1$ in , we obtain: $$\begin{aligned} & (\tilde{p}^{-1})_{a+1}(v_{1}\cdot v_{2}\cdots v_{a+1})\\ & =-\sum_{m=2}^{a+1}\frac{1}{m!}\sum_{I_{1}\cdots I_{m}\neq \emptyset ; I_{1}\sqcup \cdots \sqcup I_{m}=\{1,\cdots, a+1\} }\tilde{p}_{m}((\tilde{p}^{-1})_{|I_{1}|}(v_{I_{1}})\cdots (\tilde{p}^{-1})_{|I_{m}|}(v_{I_{m}})).\end{aligned}$$ By the same aforementioned considerations all the summands are zero except the only ones in which appears $\tilde{p}_{a+1}$, so $$\begin{aligned} & (\tilde{p}^{-1})_{a+1}(v_{1}\cdot v_{2}\cdots v_{a+1})\\ & =-\frac{1}{(a+1)!}\sum_{I_{1}\cdots I_{a+1}\neq \emptyset ; I_{1}\sqcup \cdots \sqcup I_{a+1}=\{1,\cdots, a+1\} }\tilde{p}_{a+1}((\tilde{p}^{-1})_{|I_{1}|}(v_{I_{1}})\cdots (\tilde{p}^{-1})_{|I_{a+1}|}(v_{I_{a+1}})).\end{aligned}$$ The only way such that $a+1$ non-empty sets have disjoint union equal to $\{ 1,\cdots, a+1 \}$ is that each one of them has cardinal $1$ and then for all $1\leqslant j\leqslant a+1$ we have $(\tilde{p}^{-1})_{|I_{j}|}(v_{I_{j}})=(\tilde{p}^{-1})_{1}(v_{i_{j}})=I_{V}(v_{i_{j}})=v_{i_{j}}$. Consequently, $$\begin{aligned} & (\tilde{p}^{-1})_{a+1}(v_{1}\cdot v_{2}\cdots v_{a+1})\\ & =-\frac{1}{(a+1)!}\sum_{I_{1}\cdots I_{a+1}\neq \emptyset ; I_{1}\sqcup \cdots \sqcup I_{a+1}=\{1,\cdots, a+1\} }\tilde{p}_{a+1}(v_{i_{1}}\cdots v_{i_{a+1}}). \end{aligned}$$ Since the $\cdot$ is a symmetric product we can write, $$\begin{aligned} (\tilde{p}^{-1})_{a+1}(v_{1}\cdot v_{2}\cdots v_{a+1})=-\tilde{p}_{a+1}(v_{1}\cdots v_{a+1}).\end{aligned}$$ That is $ (\tilde{p}^{-1})_{a+1}=-\tilde{p}_{a+1}$. Now we know that $\tilde{p}^{-1}=\{Id_{V},0\cdots 0,-p_{a+1},\cdots\}$, it is no so difficult to see that $(\tilde{p}^{-1}\cdot p)_{a+1}=0$ by using the equation for $n=a+1$. The uniqueness is deduced from a similar argument, if $P=P_{a+1}\cdot P' =Q_{a+1}\cdot Q'$ with $P_{a+1}, Q_{a+1}\in G'_{a+1}$ and $P', Q'\in G'_{>a+1}$ then $P_{a+1}^{-1}\cdot Q_{a+1}=P'\cdot Q'^{-1}$. If the sequential representation of $P_{a+1}$ is $\{ Id_{V},0\cdots,\tilde{p}_{a+1},0, \cdots \}$ the element $a+1$ of the sequential representation of $P^{-1}_{a+1}$ is $-\tilde{p}_{a+1}$ and if $Q_{a+1}$ has representation $\{ Id_{V},0\cdots,\tilde{q}_{a+1},0, \cdots \}$ so the $a+1$-element of the representation of $P^{-1}_{a+1}\cdot Q_{a+1}$ is $-\tilde{p}_{a+1}+\tilde{q}_{a+1}$, but $P_{a+1}^{-1}\cdot Q_{a+1}\in G'_{>a+1}$ so $-\tilde{p}_{a+1}+\tilde{q}_{a+1}=0$ and $P_{a+1}=Q_{a+1}$, which says us that $P'=Q'$. The proof of Proposition \[prop8\] is deduced by repeated application of the precedent lemma. For Proposition \[prop8\] and Lemma \[lema1\] there exist similar statements in which the order of the compositions is the opposite. We rewrite Proposition \[prop8\] in these terms: [\[prop9\][Let $a\in \mathbb{N}$ and $P\in G'$. There are unique $P_{1}, P_{2},\cdots P_{a}, P'\in G'$ such that $P_{i}\in G'_{i}$, $P'\in G'_{>a}$ and $P=P'\circ P_{a}\circ P_{a-1}\cdots P_{1}$ ]{}]{} Thanks to Propositions \[bnbn\] and \[prop8\] it is sufficient to prove that $P^{-1}\cdot G'_{>a}\cdot P \subseteq G'_{>a}$, for all $P\in G'_{i}$ with $1\leqslant i \leqslant a$. If $i=1$ and $Q\in G'_{>a}$ then abusing the notation $Q=\{ Id_{V},0\cdots,0,q_{a+1},\cdots \}$, and with is deduced $P^{-1}\cdot Q\cdot P=\{ Id_{V},0\cdots,0, p_{1}\circ q_{a+1}\circ p_{1}^{\odot (a+1)},\cdots \}$ which is in $G'_{>a}$ (where $\odot$ denotes the symmetric product). If $2\leqslant i \leqslant a$; $Q=\{ Id_{V}, 0,\cdots,0,q_{a+1},\cdots \}$ and $P=\{ Id_{V},0,\cdots,0,p_{i},0\cdots \}$ by using obtain $Q\cdot P=\{ Id_{V},0,\cdots,0,p_{i},0, \cdots ,0,q'_{a+1},\cdots \}$. By using $P^{-1}=\{ Id_{V},0\cdots,0,-p_{i},r_{i+1},\cdots, r_{a},\cdots \}$.But when we multiply $P^{-1}\cdot Q\cdot P$ the first $a$ elements are calculated in the same manner which the first $a$ elements of $P^{-1}\cdot P$, because the elements with order bigger than $a$ do not appear. So $(\pi\circ P^{-1}\cdot Q \cdot P)|_{S^{n}V}=(\pi\circ P^{-1}\cdot P)|_{S^{n}V}$ for all $1\leqslant n \leqslant a$, which implicates that $P^{-1}\cdot Q\cdot P\in G'_{>a}$. The preceding proposition is valid in the case $a=0$ too but in this case the proof is immediate. The next corollary is a consequence of the above results. \[iso\] For all $a\in \mathbb{N}$ we have $G'/G'_{>a}\simeq G'_{\leqslant a}$. For all $a,b\in \mathbb{N}$; $[G'_{>a},G'_{>b}]\subseteq G'_{>a+b}$ Without loss of generality we can suppose $a\leqslant b$. Let $P\in G'_{>a}$ and $Q\in G'_{>b}$. Then the sequential representation of $P$ is $\{ p_{n}\}_{n\in \mathbb{N}}=\{Id_{V},0,\cdots,0,p_{a+1},p_{a+2},\cdots\}$ and similarly for $Q$, $\{ q_{n}\}_{n\in \mathbb{N}}=\{Id_{V},0,\cdots,0,q_{b+1},q_{b+2},\cdots\}$. By the composition formula we have $$\label{ñññ} (p\cdot q)_{n}(v_{1}\cdots v_{n})=\sum_{m=1}^{n}\frac{1}{m!}\sum_{I_{i}\neq \emptyset; \sqcup_{i=1}^{m}I_{i}=\{ 1,\cdots,n\}}p_{m}(q_{|I_{1}|}(v_{I_{1}})\cdots q_{|I_{m}|}(v_{I_{m}}) )$$ If $n=1$ in we have $(p\cdot q)_{1}=Id_{V}$. For $n\in \{ 2,\cdots,a \}$, in the sum over $m$ only the terms with $m=1$ survive, because $p_{j}=0$ for $2\leqslant j\leqslant n \leqslant a$, then $(p\cdot q)_{n}(v_{1}\cdots v_{n})=q_{n}(v_{1}\cdots v_{n})$ which is equal to zero because of $2\leqslant n \leqslant a\leqslant b$. For $a+1\leqslant n\leqslant b$, in only the terms with $m=1, a+1,a+2,\cdots n$ survive, so $$\label{kbkb} (p\cdot q)_{n}(v_{1}\cdots v_{n})=q_{n}(v_{1}\cdots v_{n}) +\sum_{m=a+1}^{n}\frac{1}{m!}\sum_{I_{i}\neq \emptyset; \sqcup_{i=1}^{m}I_{i}=\{ 1,\cdots,n\}}p_{m}(q_{|I_{1}|}(v_{I_{1}})\cdots q_{|I_{m}|}(v_{I_{m}}) )$$ but $q_{n}=0$. Moreover, if $m\in \{ a+1,\cdots, n-1 \}$ then there exists at least one $I_{i_{0}}$ such that $|I_{i_{0}}|\geqslant 2$ (and $|I_{i_{0}}|\leqslant b$) and consequently $q_{|I_{i_{0}}|}=0$ and becomes $$(p\cdot q)_{n}(v_{1}\cdots v_{n})=p_{n}(v_{1}\cdots v_{n}).$$ Lastly if $b+1\leqslant n \leqslant a+b$, as before in the sum over $m$ only the terms with $m=1,a+1,a+2,\cdots n$ persist. Then with similar arguments as in the previous cases we have, $$(p\cdot q)_{n}(v_{1}\cdots v_{n})=q_{n}(v_{1}\cdots v_{n})+p_{n}(v_{1}\cdots v_{n}).$$ So we have $$\label{aaa} \{ (p\cdot q)\}_{n\in \mathbb{N}}=\{ Id_{v},0,\cdots,0, p_{a+1},\cdots, p_{b},p_{b+1}+q_{b+1},\cdots,p_{b+a}+q_{b+a},\cdots \}.$$ The computation of the sequential representation of $Q\circ P$ is a bit easier and gives us the same result up to the $a+b$ term. For this composition we have the formula $$\label{zz} (q\cdot p)_{n}(v_{1}\cdots v_{n})=\sum_{m=1}^{n}\frac{1}{m!}\sum_{I_{i}\neq \emptyset; \sqcup_{i=1}^{m}I_{i}=\{ 1,\cdots,n\}}q_{m}(p_{|I_{1}|}(v_{I_{1}})\cdots p_{|I_{m}|}(v_{I_{m}}) ).$$ It is easy to check the $(q\cdot p)_{1}=Id_{V}$, and that if $n\in \{ 2,\cdots,a \}$ then $(q\cdot p)_{n}=0$, because becomes $$(q\cdot p)_{n}(v_{1}\cdots v_{n})=q_{1}(p_{|I_{1}|}(v_{I_{1}}))=p_{n}(v_{1}\cdots v_{n})=0,$$for $I_{1}=\{ 1,\cdots ,n \}$. For $n\in \{ a+1,\cdots,b \}$ we have $$(q\cdot p)_{n}(v_{1}\cdots v_{n})=q_{1}(p_{|I_{1}|}(v_{I_{1}}))=p_{n}(v_{1}\cdots v_{n}),$$but this time $p_{n}\neq 0$, because $n\geqslant a+1$. Finally for $n\in \{ b+1,\cdots , b+a \}$ in only the terms with $m=1,b+1,b+2,\cdots,n$ survive so, $$(q\cdot p)_{n}(v_{1}\cdots v_{n})=q_{1}(p_{n}(v_{1}\cdots v_{n}))+\sum_{m=b+1}^{n}\frac{1}{m!}\sum_{I_{i}\neq \emptyset; \sqcup_{i=1}^{m}I_{i}=\{ 1,\cdots,n\}}q_{m}(p_{|I_{1}|}(v_{I_{1}})\cdots p_{|I_{m}|}(v_{I_{m}}) ),$$ and for $m\in \{ b+1,\cdots,n-1 \}$ there exists at least one $I_{i_{0}}$ with more than one element and less then $a+1$, hence $p_{|I_{i_{0}}|}=0$. That is, $(q\cdot p)_{n}(v_{1}\cdots v_{n})=p_{n}(v_{1}\cdots v_{n})+q_{n}(v_{1}\cdots v_{n})$. Then the sequential representation of $Q\circ P$ is the same that the right hand side of up to the $a+b$-term. Whence $P\circ Q=Q\circ P$ in $G'_{\leqslant a+b}$, which proves what we want. We have just seen that it has sense to talk about $G'/G'_{>a}$ for all $a\in \mathbb{N}$ it is possible to see that $G'$ is the inverse limit of $G'/G'_{>n}$. If $a<b$ are natural numbers then we can define $\pi_{ab}:G'/G'_{>b}\rightarrow G'/G'_{>a}$ as $\pi_{ab}(P_{1}P_{2}\cdots P_{b})=P_{1}P_{2}\cdots P_{a}$ where we have used that each element that belongs to $G'/G'_{>b}$ can be written uniquely as a product by Lemma \[iso\]. This function clearly satisfies $\pi_{ab}\circ \pi_{bc}=\pi_{ac}$ and so we have a projective system. Moreover the morphisms $\pi_{a}:G'\rightarrow G'/G'_{>a}$ such that $\pi_{a}(P)=[P]_{G'/G'_{>a}}=P_{1}\cdots P_{a}$ commutes with $\pi_{ab}$ ($a<b$) because $\pi_{ab}\circ\pi_{b}(P)=\pi_{ab}(P_{1}\cdots P_{b})=P_{1}\cdots P_{a}=\pi_{a}(P)$. Suppose that $W$ is another group with a family of arrows $f_{a}:W \rightarrow G'/G'_{>a}$ satisfying $\pi_{ab}\circ f_{b}=f_{a}$ too, for $a<b$. We can define a unique morphism $f:W\rightarrow G'$ such that the following diagram $$\begin{aligned} \xymatrix{ G' \ar[rr]^{\pi_{b}}& & G'/G'_{>b} \ar[d]^{\pi_{ab}}\\ W \ar@{-->}[u]^{f} \ar[urr]^{f_{b}} \ar[rr]_{f_{a}} & & G'/G'_{>a} } \end{aligned}$$ commutes. We will define a topology on $G'$ (its projective topology, the finest such that $\pi_{n}$ are all continuous), where a subbase of the open neighbourhoods of zero is given by the $G'_{>n}$. Then in the above diagram we can define as $f(w)=\lim_{n\to \infty}f_{n}(w)\in G'/G'_{>n}=G'_{1}\cdots G'_{n}\subseteq G'$ which exists because $f_{n}$ commutes with $\pi_{ab}$. With this definition we have $\pi_{b}(f(w))=\pi_{b}(\lim_{n\to \infty}f_{n}(w))=\lim_{n\to \infty}\pi_{b}f_{n}(w)$ where used the continuity of $\pi_{b}$ with respect to this topology. Note that for all $n>b$, $f_{n}(w)$ has more than $b$ “elements”, and so $\pi_{b}(f_{n}(w))=f_{b}(w)$ and consequently $\pi_{b}\circ f=f_{b}$. As a major consequence of the last results we can stablish the next corollary, its proof is immediate. Given $P\in G'$, call $P_{i}$ the morphisms belonging to $G'_{i}$ proportioned by Proposition \[prop8\]. The limit $P_{1}\circ P_{2}\circ P_{3}\cdots$ (i.e. $\lim_{n\to \infty} P_{1}\cdots P_{n}$ ) where $P_{i}\in G'_{i}$ converges in $G'$ because a subbase of open sets for the projective topology in $G'$ are $G'_{>k}$ for $k\in \mathbb{N}$ and then the difference between $P$ and $P_{1}\cdots P_{n}$ belongs to any $G'_{>a}$, provided $n$ is sufficiently large. This corollary has also a version with the products in the opposite order. \[hhhhhh\] $G'$ is equal to $\cdots G'_{3}G'_{2}G'_{1}$in the sense that any element $P\in G'$ can be written as an infinite product $P=\cdots P_{3}\circ P_{2}\circ P_{1}$ of elements $P_{i}\in G'_{i}$ and conversely each of this infinite products in such conditions belongs to $G'$. It is of central interest for us the case $V=\omega SJ\Phi$. In this case $G'$ is not the renormalization group, because an element belonging to $G'$ may not preserve the coaction of $SJ\Phi$.\ The natural map $S\Gamma \omega SJ\Phi\rightarrow \Gamma S\omega SJ\Phi$ is not an isomorphism because as we mentioned, there is a difference between the symmetric product in the category of sheafs of modules over the ringed space $(\mathcal{M},\mathcal{C}^{\mathbb{R}})$ and the symmetric product in the category of $\mathbb{R}$-vector spaces. For example if we denote by $\odot$ the symmetric product in the category of sheafs of modules over the ringed space $(\mathcal{M},\mathcal{C}^{\mathbb{R}})$ and by $\cdot$ in the category of $\mathbb{R}$-vector spaces, the natural map sends $\psi \cdot \eta\rightarrow \psi\odot\eta$ where $\psi$ and $\eta$ are lagrangian densities. The difference between these product is that if $f$ is a smooth function over $\mathcal{M}$ then $f\psi \cdot \eta\neq \psi\cdot f\eta$ but $f\psi \odot \eta=\psi\odot f\eta$, hence the natural map is not injective. [\[lema7\][The induced action of a renormalization over $\Gamma S\omega SJ\Phi$ can be lifted to an action over $S\Gamma \omega SJ\Phi$ preserving the coproduct, the coaction of $\Gamma SJ\Phi $ and the product of elements with disjoint support.]{}]{} By Theorem \[df\] $\Gamma \omega SJ\Phi\simeq \Gamma\omega \otimes_{\mathcal{C}^{\infty}(\mathcal{M})}\Gamma SJ\Phi$, then applying Proposition \[prop 1\], $\Gamma\omega \otimes_{\mathcal{C}^{\infty}(\mathcal{M})}\Gamma SJ\Phi$ can be seen as a $\Gamma SJ\Phi$-comodule over $\mathcal{C}^{\infty}(\mathcal{M})$. Call $N$ the natural map $S\Gamma \omega SJ\Phi\rightarrow \Gamma S\omega SJ\Phi$, and $\varphi$ the isomorphism existent between $\Gamma S\omega SJ\Phi$ and $S_{\mathcal{C^{\infty}(\mathcal{M})}}\Gamma \omega SJ\Phi$ (see Corollary \[dm\]). In this context, given a renormalization $P$, we still call $P$ the induced map on the sections of $S\omega SJ\Phi$ and then we consider the composition $p:=\pi_{1}\circ \varphi\circ P\circ N :S\Gamma \omega SJ\Phi \rightarrow \Gamma\omega SJ\Phi$. By using Proposition \[prop4\], $p:S\Gamma \omega SJ\Phi \rightarrow \Gamma\omega SJ\Phi$ is in correspondence with a morphism which we call $P':S\Gamma \omega SJ\Phi \rightarrow S\Gamma\omega SJ\Phi$ and is written (as in Proposition \[prop3\] ) $$\label{dhdh} P'(c)=\sum_{n\in\mathbb{N}}\frac{1}{n!}p^{\odot n}\circ \Delta_{\overline{C}}^{(n)}(c),$$ where $C=S\Gamma \omega SJ\Phi$ and $\Delta^{(n)}_{\overline{C}}$ was defined in Section \[sec111\]. Now we prove that $P'$ preserves the product of elements with disjoint support, so let $\varphi=\prod_{i=1}^{k}f_{i}$ and $\psi=\prod_{l=1}^{r}g_{l}$ be elements of $S\Gamma \omega SJ\Phi$ with disjoint support, $\varphi \cdot \psi \in S\Gamma \omega SJ\Phi$, call $\chi_{i}=f_{i}$ for all $1\leqslant i\leqslant k$ and $\chi_{k+l}=g_{l}$ for all $1\leqslant l\leqslant r$. From the statement over the supports we can suppose that $supp(f_{i})\subseteq U$ and $supp(g_{l})\subseteq V$ for some $U$ and $V$ disjoint open sets in $\mathcal{M}$. Consider $$\Delta_{S\Gamma \omega SJ\Phi}^{(n)}(\varphi\cdot \psi)=\sum_{I_{1},\cdots,I_{n}\neq \emptyset; I_{1} \sqcup \cdots \sqcup I_{n}=\{1,\cdots,k+r \}} \chi_{I_{1}}\otimes \cdots \otimes \chi_{I_{n}},$$ so by applying $p^{\odot n}$ we obtain $$\label{343} p^{\odot n}\circ\Delta_{S\Gamma \omega SJ\Phi}^{(n)}(\varphi\cdot \psi)=\sum_{I_{1},\cdots,I_{n}\neq \emptyset; I_{1} \sqcup \cdots \sqcup I_{n}=\{1,\cdots,k+r \}} p(\chi_{I_{1}})\odot \cdots \odot p(\chi_{I_{n}}),$$ Let $I_{k}$ be a subset of $\{1,\cdots,k+r \}$ such that has at least one index $i_{0}\in \{1,\cdots,k\}$ and one index $j_{0}\in \{k+1,\cdots,k+r\}$, and $h\in \mathcal{C}_{0}^{\infty}(\mathcal{M})$ such that $h|_{Sup(\varphi)}\equiv 1$ and $h|_{Sup(\psi)}\equiv 0$ , this function exists because the spacetime is paracompact. Then, $$\begin{aligned} & N(\chi_{I_{k}})=N(\chi_{\alpha}\cdots \chi_{i_{0}}\cdots\chi_{j_{0}}\cdots \chi_{|I_{k}|}) \\ & =\chi_{\alpha}\odot\cdots \odot \chi_{i_{0}}\odot \cdots\odot \chi_{j_{0}}\odot\cdots\odot \chi_{|I_{k}|}\\ & =\chi_{\alpha}\odot\cdots \odot h\chi_{i_{0}}\odot \cdots\odot \chi_{j_{0}}\odot\cdots\odot \chi_{|I_{k}|}\\ & =\chi_{\alpha}\odot\cdots \odot \chi_{i_{0}}\odot \cdots\odot h\chi_{j_{0}}\odot\cdots\odot \chi_{|I_{k}|}\\ & =\chi_{\alpha}\odot\cdots \odot \chi_{i_{0}}\odot \cdots\odot 0 \odot\cdots\odot \chi_{|I_{k}|}=0\end{aligned}$$where $\alpha$ is the the first element of $I_{k}$. Then, in equation only terms such that all the $I_{j}$ are totally included in $\{1,\cdots,k\}$ or in $\{k+1,\cdots,k+r\}$ survive. We will call $I$ to the first and $J$ to the seconds. In the equation , we have terms with only one $I$ and $n-1$ subsets $J$, or terms with two $I$ and $n-2$ $J$’s, three $I$’s, four $I$’s ... $n-1$. In the following sums all the $I$’s are contained in $\{1,\cdots k\}$ and the $J$’s in $\{ k+1,\cdots k+r \}$, all of them are $\neq \emptyset$ and its disjoint union is $\{1,\cdots, k+r \}$. $$\begin{aligned} & p^{\odot n}\circ\Delta_{S\Gamma \omega SJ\Phi}^{(n)}(\varphi \cdot \psi) \\ & =\sum_{J_{1}\cdots J_{n-1}} n [p(\chi_{I_{1}})\odot p(\chi_{J_{1}})\odot \cdots \odot p(\chi_{J_{n-1}})] \\ &+\sum_{I_{1}, I_{2}; J_{1}\cdots J_{n-3}=\{ k+1,\cdots,k+r \}} \binom{n}{2} [p(\chi_{I_{1}})\odot p(\chi_{I_{2}})\odot p(\chi_{J_{1}})\odot \cdots \odot p(\chi_{J_{n-2}})] \\ &+\sum_{I_{1}\cdots I_{3}; J_{1}\cdots J_{n-3}} \binom{n}{3} [p(\chi_{I_{1}})\odot \cdots \odot p(\chi_{I_{3}})\odot p(\chi_{J_{1}})\odot \cdots \odot p(\chi_{J_{n-3}})]+\cdots ,\end{aligned}$$ where the binomial coefficients appear because of the possible reorderings of the expression. We will make the computations for $n=1,2,3,4...$ For the general case one must proceed with a combinatorial argument as in the proof of Proposition \[prop3\]. For the case $\textbf{n=1}$ the result of the above computation is zero because there are no subset $J$. Case $\textbf{n=2}$, $p^{2}\circ\Delta_{S\Gamma \omega SJ\Phi}(\varphi\cdot \psi)=2p(\chi_{I_{1}})\odot p(\chi_{J_{1}})=2p(\varphi)\odot p(\psi)$. Case $\textbf{n=3}$: $$p^{\odot 3}\circ \Delta_{S\Gamma \omega SJ\Phi}^{(3)}(\varphi\cdot\psi)=3p(\varphi)\odot[\sum_{J_{1},J_{2}}p(\chi_{J_{1}})\odot p(\chi_{J_{2}})]+3[\sum_{J_{1},J_{2}}p(\chi_{I_{1}})\odot p(\chi_{I_{2}})]\odot p(\psi)$$ $$=3p(\varphi)\odot p^{\odot 2}\circ \Delta_{S\Gamma \omega SJ\Phi}^{(2)}(\psi) +3p^{\odot 2}\circ\Delta_{S\Gamma \omega SJ\Phi}^{(2)}(\varphi)\odot p(\psi)$$ For $\textbf{n=4}$, $$\begin{aligned} & p^{\odot 4}\circ \Delta_{S\Gamma \omega SJ\Phi}^{(4)}(\varphi\cdot\psi)=4p(\varphi)\odot p^{\odot 3}\circ \Delta_{S\Gamma \omega SJ\Phi}^{(3)}(\psi)\\ & +\binom{4}{2}p^{\odot 2}\circ \Delta_{S\Gamma \omega SJ\Phi}^{(2)}(\varphi) \odot p^{\odot 2}\circ \Delta_{S\Gamma \omega SJ\Phi}^{(2)}(\psi)+4p^{\odot 3}\circ \Delta_{S\Gamma \omega SJ\Phi}^{(3)}(\varphi)\odot p(\psi)\end{aligned}$$ Multiplying the last equations by $\frac{1}{n!}$ and adding them as in equation , we can see how are the first terms of $P'(\varphi\cdot \psi)$, using easy facts as $\frac{1}{4!}\binom{4}{2}=\frac{1}{2}\frac{1}{2}$. Then we conclude, $$\begin{aligned} & P'(\varphi\cdot \psi)=p(\varphi)\odot p(\psi)+\frac{1}{2}p(\varphi)\odot p^{\odot 2}\circ \Delta_{S\Gamma \omega SJ\Phi}^{(2)}(\psi) +\frac{1}{2}p^{\odot 2}\circ\Delta_{S\Gamma \omega SJ\Phi}^{2}(\varphi)\odot p(\psi)\\ & +\frac{1}{3!}p(\varphi)\odot p^{\odot 3}\circ \Delta_{S\Gamma \omega SJ\Phi}^{(3)}(\psi) +\frac{1}{2}\frac{1}{2}p^{\odot 2}\circ \Delta_{S\Gamma \omega SJ\Phi}^{(2)}(\varphi) \odot p^{\odot 2}\circ \Delta_{S\Gamma \omega SJ\Phi}^{(2)}(\psi)\end{aligned}$$ $$\label{producto} +\frac{1}{3!}p^{\odot 3}\circ \Delta_{S\Gamma \omega SJ\Phi}^{(3)}(\varphi)\odot p(\psi)+\cdots$$ On the other hand if we compute $P'(\varphi)\odot P'(\psi)$, we have $$\left(\sum_{k\in\mathbb{N}}\frac{1}{k!}p^{\odot k}\circ \Delta_{S\Gamma \omega SJ\Phi}^{(k)}(\varphi)\right)\odot \left(\sum_{k\in\mathbb{N}}\frac{1}{s!}p^{\odot s}\circ \Delta_{S\Gamma \omega SJ\Phi}^{(s)}(\psi)\right)$$ $$=\left( p\circ Id_{S\Gamma \omega SJ\Phi}(\varphi)+\frac{1}{2}p^{\odot 2}\circ\Delta_{S\Gamma \omega SJ\Phi}^{(2)}(\varphi)+\frac{1}{3!}p^{\odot 3}\circ \Delta_{S\Gamma \omega SJ\Phi}^{(3)}(\varphi)+\cdots \right)$$ $$\odot \left( p\circ Id_{S\Gamma \omega SJ\Phi}(\psi)+\frac{1}{2}p^{\odot 2}\circ\Delta_{S\Gamma \omega SJ\Phi}^{(2)}(\psi)+\frac{1}{3!}p^{\odot 3}\circ \Delta_{S\Gamma \omega SJ\Phi}^{(3)}(\psi)+\cdots \right)$$ $$\begin{aligned} &=p(\varphi)\odot p(\psi)+\frac{1}{2}p(\varphi)\odot p^{\odot 2}\circ \Delta_{S\Gamma \omega SJ\Phi}^{(2)}(\psi) +\frac{1}{2}p^{\odot 2}\circ\Delta_{S\Gamma \omega SJ\Phi}^{2}(\varphi)\odot p(\psi)\\ & +\frac{1}{3!}p(\varphi)\odot p^{\odot 3}\circ \Delta_{S\Gamma \omega SJ\Phi}^{(3)}(\psi) +\frac{1}{2}p^{\odot 2}\circ \Delta_{S\Gamma \omega SJ\Phi}^{(2)}(\varphi) \odot \frac{1}{2} p^{\odot 2}\circ \Delta_{S\Gamma \omega SJ\Phi}^{(2)}(\psi)\\ & +\frac{1}{3!}p^{\odot 3}\circ \Delta_{S\Gamma \omega SJ\Phi}^{(3)}(\varphi)\odot p(\psi)+\cdots ,\end{aligned}$$ which is equal to $P'(\varphi\cdot\psi)$ by equation . Notice that we used the fact that $\Delta_{S\Gamma \omega SJ\Phi}=Id_{S\Gamma SJ\Phi}$ (see the definition of $\Delta^{(n)}$ in Section \[sec111\]). As an observation we can say that renormalizations do not necessarily preserve the product of elements of $S\Gamma_{c}\omega SJ\Phi$ if they have not disjoint support. Renormalizations $\curvearrowright$ Feynman measures ---------------------------------------------------- Thanks to the Lemma \[lema7\], given a renormalization $P$ we can think it as an automorphism of $S\Gamma \omega SJ\Phi$ which preserves the coaction of $\Gamma SJ\Phi$. A Feynman measure (see Definition \[dia\]) is a linear map from $S\Gamma_{c} \omega SJ\Phi$ to $\mathbb{C}$; then we can define the action from the renormalization group over the Feynman measures as follows. If $P$ is a renormalization and $\delta$ is a Feynman measure, we can lift $P$ by using Lemma \[lema7\] that is still called $P$. Then a new Feynman measure $P(\delta)$ is define by $P(\delta)(A):=\delta(P^{-1}(A))$ if $A\in S\Gamma_{c}\omega SJ\Phi$, . It will be interesting to prove that $P(\delta)$ is still a Feynman measure associated with the propagator $\Delta$ if $\delta$ is associated with $\Delta$ too. First of all we will prove the smoothness on the diagonal. Suppose $\delta$ is a Feynman measure and let $P$ the lift of some renormalization, and take $A\in S^{n}\Gamma_{c}\omega SJ\Phi$. $P^{-1}$ is an element of the renormalization group that preserves the filtration, so $P^{-1}(A)$ is a polynomial in monomials of degree at least $n$ , say $B_{1}+\cdots+B_{n}$. Given $(p_{1},\cdots,p_{k})\in \sum_{(q,\cdots,q)}\chi_{k}(\delta,B_{k})$ one can affirm that $p_{1}+\cdots +p_{k}=0$ because $\delta$ is smooth in the diagonal, as a conclusion **$P\delta$ is smooth in the diagonal**.\ To see that $P\delta$ is non-degenerate we can use Theorem \[teorema1\] to establish that if $P$ is a renormalization $P|_{\Gamma_{c}\omega}$ is an isomorphism. But the isomorphisms of sheafs (induced by vector bundles) are in correspondence whith the isomorphisms of vector bundles supported by the identity over $\mathcal{M}$ . Given two vector bundles, an isomorphism between them must be linear on each fiber, but in our case the bundle has range $1$, because it is the density bundle. Hence the isomorphism is given (on the fibers) by multiplication by a constant $\lambda\neq0$, so the bundle morphism is given globally by the multiplication by a nowhere-vanishing function $\lambda\in \mathcal{C}^{\infty}(\mathcal{M})$. Applying this to $P^{-1}$, given $v\in \Gamma_{c} \omega$, we obtain, $$P\delta (v)=\delta(P^{-1}(v))=\int_{\mathcal{M}}gP^{-1}(v),$$ where $g$ is a nowhere-vanishing function that exists because $\delta$ is non-degenerate. Moreover, $$P\delta (v)=\int_{\mathcal{M}}(g\lambda)(v),$$ where $\lambda$ is the nowhere-vanishing function described above. Then **$P\delta$ is non-degenerate**.\ Let $\delta$ be a Feynman measure, $Q$ an element of the renormalization group (whose inverse is $P$) and let $A\in S^{m}\Gamma_{c} \omega SJ\Phi$ and $B\in S^{n}\Gamma_{c} \omega SJ\Phi$ such that there is no coordinate of any element in $Sup(A)$ which is $\leqslant$ to some coordinate of any element in $Sup(B)$, this particularly implies that $Sup(A)\cap Sup(B)=\emptyset$, and then by Lemma \[lema7\], $P$ preserves the product. Then, $$Q\delta(A\cdot B)=\delta(P(A\cdot B))=\delta(P(A)\cdot P(B))$$ and as $\Delta$ is cut the products involved in the Gaussian condition are well defined (see Observation \[obs4\]), so $$\label{345} Q\delta(A\cdot B)=\sum (\chi(\delta,(P(A))')\otimes \chi(\delta,(P(B))'))\cdot \hat{\Delta}((P(A))'',(P(B))'').$$ We will now compute $(P(A))'$, $(P(A))''$, $(P(B))''$ and $(P(B))'$. Consider $$\xymatrix{ S\Gamma \omega SJ\Phi \ar[rr]^{\sigma} \ar[d]_{P} & & S\Gamma \omega SJ\Phi\otimes_{\mathcal{C}^{\infty}(\mathcal{M})} S\Gamma SJ\Phi \ar[d]^{P\otimes Id} \\ S\Gamma \omega SJ\Phi \ar[rr]^{\sigma} & & S\Gamma \omega SJ\Phi\otimes_{\mathcal{C}^{\infty}(\mathcal{M})} S\Gamma SJ\Phi \\}$$ which is commutative. Applying this maps to the element $A$, and remembering that $\sigma(A)=\sum A'\otimes A''$ we have $$A''=(P(A))''$$ and the same for $B$. Substituting the results in equation , we obtain $$Q\delta(A\cdot B)=\sum (\chi(\delta,(P(A))')\otimes \chi(\delta,(P(B))'))\cdot \hat{\Delta}(A'',B'').$$ The commutative diagram tells us that $(P(A))'=P(A')$, using this we have $$\begin{aligned} & Q\delta(A\cdot B)=\sum (\chi(\delta,P(A'))\otimes \chi(\delta,P(B')))\cdot \hat{\Delta}(A'',B'')\\ & =\sum (\chi(Q\delta,A')\otimes \chi(Q\delta,B'))\cdot \hat{\Delta}(A'',B''),\end{aligned}$$ which is exactly the Gaussian condition for $Q\delta$. As a conclusion we can say that if $P$ is a renormalization and $\delta$ a Feynman measure then $P\delta$ is a Feynman measure too. So we have an action of the renormalization group over the set of Feynman measures associated to a fixed cut propagator.\ We will need the following result. [\[lemacociente\][ Given $k\in \mathbb{N}$, if $\lambda:S^{k}\Gamma_{c}\omega SJ\Phi\rightarrow \mathbb{C} $ is a continuous function satisfying $Sup(\lambda)\subseteq \text{Diag}(\mathcal{M}^{k})$ then $\lambda$ is $\mathcal{C}^{\infty}(\mathcal{M})$-balanced.]{}]{} Let $f$ be a smooth function over $\mathcal{M}$. To prove that\ $\lambda(A_{1}\cdots fA_{i_{0}}\cdots A_{j_{0}}\cdots A_{k})=\lambda(A_{1}\cdots A_{i_{0}}\cdots fA_{j_{0}}\cdots A_{k})$ ($A_{i}$’s are Lagrangian densities of compact support) as complex numbers, we use Remark \[obs5\] and reduce it to prove the equality in $D_{d}(\mathcal{M}^{k})$. Let $g_{1}\otimes \cdots \otimes g_{k}$ be an element of $\mathcal{C}^{\infty}(\mathcal{M})^{\otimes k}$ ( recall these element are dense in $\mathcal{C}^{\infty}(\mathcal{M}^{k})$, and $\lambda$ is continuous ) whose support is included in $Diag(\mathcal{M}^{k})$, then we must compare $\chi_{k}(\lambda,A_{1}\cdots fA_{i_{0}}\cdots A_{j_{0}}\cdots A_{k})(g_{1}\otimes \cdots \otimes g_{k})$ and\ $\chi_{k}(\lambda,A_{1}\cdots A_{i_{0}}\cdots fA_{j_{0}}\cdots A_{k})(g_{1}\otimes \cdots \otimes g_{k})$, so if they are equal then $\lambda(A_{1}\cdots fA_{i_{0}}\cdots A_{j_{0}}\cdots A_{k})$ and $\lambda(A_{1}\cdots A_{i_{0}}\cdots fA_{j_{0}}\cdots A_{k})$ will also be equal. But, $$\chi_{k}(\lambda,A_{1}\cdots fA_{i_{0}}\cdots A_{j_{0}}\cdots A_{k})(g_{1}\otimes \cdots \otimes g_{k})$$ $$=\lambda(g_{1}A_{1}\cdots g_{i_{0}}fA_{i_{0}}\cdots g_{j_{0}}A_{j_{0}}\cdots g_{k}A_{k})$$ $$\label{3.46} =\chi_{k}(\lambda,A_{1}\cdots A_{i_{0}}\cdots A_{j_{0}}\cdots A_{k})(g_{1}\otimes \cdots \otimes fg_{i_{0}}\otimes\cdots \otimes g_{k})$$ using that $g_{1}\otimes \cdots \otimes fg_{i_{0}}\otimes\cdots \otimes g_{k}=g_{1}\otimes \cdots \otimes fg_{j_{0}}\otimes \cdots \otimes g_{k}$ (that is because evaluating the left hand side in a point $(x_{1},\cdots,x_{k})\in \mathcal{M}^{k}$ we obtain $g_{1}(x_{1}) \cdots f(x_{i_{0}})g_{i_{0}}(x_{i_{0}})\cdots g_{k}(x_{k})$, which is zero if there are two points $x_{i}\neq x_{j}$. But in the relevant case $(x_{1},\cdots,x_{k})\in \text{Diag}(\mathcal{M}^{k})$, that is $(x,\cdots,x)$ we obtain $g_{1}(x) \cdots f(x)g_{i_{0}}(x)\cdots g_{k}(x)=g_{1}(x) \cdots g_{i_{0}}(x) \cdots f(x)g_{j_{0}}(x) \cdots g_{k}(x)$ where we can move $f(x)$ because it is a product on $\mathbb{C}$) we conclude that the last term in Equation is equal to $$\chi_{k}(\lambda,A_{1}\cdots A_{i_{0}}\cdots A_{j_{0}}\cdots A_{k})(g_{1}\otimes \cdots \otimes fg_{j_{0}}\otimes\cdots \otimes g_{k})$$ $$\chi_{k}(\lambda,A_{1}\cdots fA_{j_{0}}\cdots A_{j_{0}}\cdots A_{k})(g_{1}\otimes \cdots \otimes g_{k}).$$ Then from Remark \[obs5\] we conclude that $\chi_{k}(\lambda,A_{1}\cdots fA_{i_{0}}\cdots A_{j_{0}}\cdots A_{k})=\chi_{k}(\lambda,A_{1}\cdots A_{i_{0}}\cdots fA_{j_{0}}\cdots A_{k})$ as element of $D_{d}(\mathcal{M}^{k})$ for all $k\in \mathbb{N}$, which implies that $\lambda(A_{1}\cdots fA_{i_{0}}\cdots A_{j_{0}}\cdots A_{k})=\lambda(A_{1}\cdots A_{i_{0}}\cdots fA_{j_{0}}\cdots A_{k})$ as complex numbers. To prove the next theorem we will use a well-known result, \[rot\] If $E\rightarrow \mathcal{M}$ is a finite rank vector bundle over a manifold, then $\Gamma_{c}(E)$ is a projective $\mathcal{C}^{\infty}(\mathcal{M})$-module. Consider the vector bundle $\omega S^{k}J\Phi$, which will be of finite rank if and only we take finite order jets $J\Phi$. Then from now on we must understand $J\Phi$ as the vector bundle of finite jets up to some fixed order, i.e. $J\Phi=J^{k}\Phi$ for some $k\in \mathbb{N}$. Another result we will use is, \[simet\] If $A$ and $B$ are projective modules over a commutative ring $R$, then the symmetric tensor product $(A\otimes_{R} B)/I$ is also a projective $R$-module ($I$ is the ideal of the vectors which give symmetry). Can be found in many books of elementary algebra as [@wei], and it is deduced from the fact that the tensor product commutes with the direct sum. With all these results we are ready to establish and prove the following theorem which gives us an important property of this action. The group of renormalizations acts transitively on the set of Feynman measures associated with a given cut local propagator. We already proved that the action is well defined in the sense that the result is also a Feynman measure associated to the same given cut local propagator. To finish the proof we must prove the transitivity. Given $\delta$ and $\delta'$ Feynman measures associated with the same cut local propagator, we want to prove the existence of a $g\in G$ ($g:S\Gamma\omega SJ\Phi\to S\Gamma\omega SJ\Phi$ ) such that $g\delta=\delta'$, expressing $g$ as $\cdots g_{3}\cdot g_{2}\cdot g_{1}$ (see Corollary \[hhhhhh\]). We proceed by induction, on the degree of the elements of $S\Gamma\omega SJ\Phi$ i.e. $S^{n}\Gamma\omega S^{k}J\Phi$ for $(n,k)\in \mathbb{N}\times \mathbb{N}_{0}$ in the lexicographic order. We begin proving the existence of $g_{1}:S\Gamma \omega S J\Phi\to S\Gamma\omega SJ\Phi$ whose sequential representation is $\{g_{1}\}_{n\in \mathbb{N}}=\{ f_{1},0,\cdots \}$ where $f_{1}:\Gamma\omega SJ\Phi\to \Gamma\omega SJ\Phi\simeq \Gamma_{c}\omega \otimes S_{\mathcal{C}^{\infty}(\mathcal{M})}\Gamma J\Phi$ (where we use Theorems \[df\] and \[dm\]). Hence by Proposition \[prop5\], there exits a map $h_{1}:\Gamma\omega SJ\Phi \to \Gamma_{c}\omega$ such that $f_{1}=(h_{1}\otimes Id_{\Gamma SJ\Phi})\circ(Id_{\Gamma_{c}\omega}\otimes \Delta_{\Gamma SJ\Phi})$. Taking $n=1$ and $k=0$ in $S^{n}\Gamma\omega S^{k}J\Phi$, we want to define $f_{1}$ over the elements $\mathcal{L}=\alpha\otimes l$ where $\alpha\in \Gamma_{c}\omega$ and $l=1 \in S^{0}\Gamma J\Phi$. We do that by means of $h_{1}$, $$\begin{aligned} & g_{1}(\mathcal{L})=f_{1}(\mathcal{L})\\ & =f_{1}(\alpha\otimes l)=(h_{1}\otimes Id_{\Gamma SJ\Phi})\circ(Id_{\Gamma_{c}\omega}\otimes \Delta_{\Gamma SJ\Phi})(\alpha\otimes 1) \\ & =(h_{1}\otimes Id_{\Gamma\omega SJ\Phi})\circ(\alpha \otimes (1\otimes 1)) \\ & =(h_{1}\otimes Id_{\Gamma\omega SJ\Phi})\circ((\alpha \otimes 1)\otimes 1) \\ & =h_{1}(\alpha\otimes 1)\otimes Id_{\Gamma SJ\Phi}(1)= h_{1}(\mathcal{L})\otimes 1 =h_{1}(\mathcal{L}).\end{aligned}$$ Applying $\delta'$ we have $$\begin{aligned} & \delta'(g_{1}(\mathcal{L}))=\delta'(f_{1}(\mathcal{L}))=\delta'(h_{1}(\mathcal{L})).\end{aligned}$$ We want $h_{1}$ to satisfy $\delta'\circ h_{1}=\delta$ over the elements with degree $n=1$ and $k=0$. In the case where $\Gamma\omega S^{0} J\Phi$ is a free $\mathcal{C}^{\infty}(\mathcal{M})$-module so it is easy to define $h_{1}|_{\Gamma\omega S^{0} J\Phi}$ over a basis such that $h_{1}|_{\Gamma\omega S^{0} J\Phi}$ satisfies $\delta'\circ h_{1}|_{\Gamma\omega S^{0} J\Phi}=\delta$ because $\delta'$ is epimorphism.\ In the general case were $\Gamma\omega S^{0} J\Phi$ (is not necessarily a free $\mathcal{C}^{\infty}(\mathcal{M})$-module) is a projective $\mathcal{C}^{\infty}(\mathcal{M})$-module, then it is a direct summand of a free $\mathcal{C}^{\infty}(\mathcal{M})$-module (call it $F$), so we apply the previous argument to define $\tilde{h}_{1}$ over $F$ and then by restricting $\tilde{h}_{1}$ to the direct summand we are interested to have $h_{1}|_{\Gamma\omega S^{0} J\Phi}$ satisfying $\delta'\circ h_{1}|_{\Gamma\omega S^{0} J\Phi}=\delta$. Hence the diagram, $$\xymatrix{ & & & \Gamma_{c}\omega \ar[d]^{\delta'} \\ S^{1}_{\mathcal{C}^{\infty}(\mathcal{M})}\Gamma\omega S^{0}J\Phi \ar[rrr]_{\delta} \ar@{-->}[rrru]^{h_{1}|_{ S^{1}_{\mathcal{C}^{\infty}(\mathcal{M})}\Gamma\omega S^{0}J\Phi}} & & & \mathbb{C} \\}$$ can be completed with a map $h_{1}|_{ S^{1}_{\mathcal{C}^{\infty}(\mathcal{M})}\Gamma\omega S^{0}J\Phi}$ or which is the same $h_{1}|_{\Gamma\omega S^{0} J\Phi}$. But $g_{1}\delta(\mathcal{L})=\delta'(\mathcal{L})$ if and only if $\delta(\mathcal{L})=\delta'(g_{1}(\mathcal{L}))$, and this is exactly what we have with the aforementioned election of $h_{1}|_{\omega S^{0}J\Phi}$. Continuing this process by defining $h_{1}$ over elements of $S^{1}_{\mathcal{C}^{\infty}(\mathcal{M})}\Gamma\omega S^{1}J\Phi$ (i.e. $n=1$ and $k=1$), we consider the Lagrangian density $\mathcal{L}=\alpha \otimes l$ where $\alpha \in \Gamma_{c}\omega$ and $l=\varphi \in S^{1}\Gamma J\Phi$ and compute $$\begin{aligned} & g_{1}(\mathcal{L})=f_{1}(\mathcal{L})\\ & =f_{1}(\alpha\otimes l)=(h_{1}\otimes Id_{\Gamma SJ\Phi})\circ(Id_{\Gamma_{c}\omega}\otimes \Delta_{\Gamma SJ\Phi})(\alpha\otimes \varphi) \\ & =(h_{1}\otimes Id_{\Gamma SJ\Phi})\circ(\alpha \otimes [1\otimes \varphi +\varphi \otimes 1]) \\ & =(h_{1}\otimes Id_{\Gamma SJ\Phi})((\alpha\otimes 1)\otimes \varphi+(\alpha\otimes \varphi)\otimes 1) \\ & =h_{1}(\alpha\otimes 1)\otimes Id_{\Gamma SJ\Phi}(\varphi)+h_{1}(\alpha\otimes \varphi)\otimes Id_{\Gamma\omega SJ\Phi}(1) \\ &= h_{1}(\alpha\otimes 1)\otimes \varphi+h_{1}(\mathcal{L}).\end{aligned}$$ Applying $\delta'$ to both sides of this equation we have, $$\delta'(g_{1}(\mathcal{L})=\delta'(h_{1}(\alpha\otimes 1)\otimes \varphi)+\delta'\circ h_{1}(\mathcal{L}).$$ But as we want $\delta'\circ g_{1}=\delta$ then we are looking for a $h_{1}|_{S^{1}_{\mathcal{C}^{\infty}(\mathcal{M})}\Gamma\omega S^{1}J\Phi}$ such that the equation $$\begin{aligned} & \delta'(h_{1}(\alpha)\otimes \varphi)+\delta'\circ h_{1}(\mathcal{L})=\delta(\mathcal{L}), \\ & \text{or } \delta'\circ h_{1}(\mathcal{L})=(\delta-\delta'\circ (h_{1}|_{S^{1}_{\mathcal{C}^{\infty}(\mathcal{M})}\Gamma\omega S^{0}J\Phi}\otimes Id_{\Gamma SJ\Phi}))(\mathcal{L})\end{aligned}$$ holds. Take a $h_{1}|_{S^{1}_{\mathcal{C}^{\infty}(\mathcal{M})}\Gamma\omega S^{1}J\Phi}$ such that the next diagram $$\xymatrix{ & & & & & & & \Gamma_{c}\omega \ar[d]^{\delta'} \\ S^{1}_{\mathcal{C}^{\infty}(\mathcal{M})}\Gamma\omega S^{1}J\Phi \ar[rrrrrrr]_{\delta-\delta'\circ (h_{1}|_{S^{1}_{\mathcal{C}^{\infty}(\mathcal{M})}\Gamma\omega S^{1}J\Phi}\otimes Id_{\Gamma SJ\Phi})} \ar@{-->}[rrrrrrru]^{h_{1}|_{ S^{1}_{\mathcal{C}^{\infty}(\mathcal{M})}\Gamma\omega S^{0}J\Phi}} & & & & & & & \mathbb{C} \\}$$ commutes. The existence of this $h_{1}|_{S^{1}_{\mathcal{C}^{\infty}(\mathcal{M})}\Gamma\omega S^{1}J\Phi}$ follows from the same arguments given for $h_{1}|_{S^{1}_{\mathcal{C}^{\infty}(\mathcal{M})}\Gamma\omega S^{0}J\Phi}$. Continuing this process we can recursively define $h_{1}|_{S^{1}_{\mathcal{C}^{\infty}(\mathcal{M})}\Gamma\omega S^{k}J\Phi}$ for all $k\in \mathbb{N}_{0}$ and so we have just defined $h_{1}:S^{1}\Gamma\omega SJ \Phi\to \Gamma_{c}\omega$. Proceeding by induction we can define $h_{n}:S^{n}\Gamma\omega SJ \Phi\to \Gamma_{c}\omega$ for each $n\in \mathbb{N}$ (using the Lemma \[rot\] and Lemma \[simet\] to affirm that $\Gamma\omega SJ \Phi$ is a projective module and so is $S\Gamma\omega SJ \Phi$. And consequently a is direct addend of a free $\mathcal{C}^{\infty}(\mathcal{M})$-module), each one of is in correspondence with a $g_{n}:S\Gamma\omega SJ \Phi\to S\Gamma\omega SJ \Phi$ whose sequential representation is $\{ 0,\cdots,0,f_{n},0,\cdots\}$. The maps $g_{n}$ define by Corollary \[hhhhhh\] a renormalization $g=...g_{3}g_{2}g_{1}$, such that $\delta(\mathcal{L})=\delta'(g(\mathcal{L}))$ for all $\mathcal{L}\in S_{\mathcal{C}^{\infty}(\mathcal{M})}\Gamma\omega SJ \Phi$. That is $g(\delta)=\delta'$. [99]{} A. Solotar, M. Farinati, M. Suárez Alvarez: *Anillos y sus categorías de representaciones* Cuadernos de Matemática y Mecánica. A. D. Lewis: *The Bundle of Infinite Jets* (2006) B. Mitchel: *Theory of categories* Academics press (1965). C. Brouder, N. Viet Dang, F. Hélein: *A smooth introduction to the wavefront set* http://lanl.arxiv.org/abs/1404.1778 G.Sardanashvily: *Remarks on the Serre-Swan theorem for non-compact manifolds* (2001) H. H. Schaefer: *Topological Vector Spaces* Springer-Verlag, New York, 1966. L. Hörmander: *The Analysis of Linear Partial Differential Operators Distribution Theory and Fourier Analysis* Springer-Verlag (1980) L. Conlon : *Differentiable Manifolds* Birkhäiuser, Berlin (2001) M.Ghasemio, M.Infusino, S.Kuhlmann, M.Marshall: *Moment problem for symmetric algebras of locally convex spaces*\ http://arxiv.org/abs/1507.06781v2 (2015) M. Grosser, M. Kunzinger, M. Obergugenberger, R. Steinbauer: *Geometric Theory of Generalized Functions with Applications to General Relativity Mathematics and Its Applications* (2001) M. Reed, B. Simon : *Fourier Analysis, Self-Adjointness (Methods of Modern Mathematical Physics, Vol. 2)* Academic Press (1975) R. L. Finney and J. Rotman: *Paracompactness of locally compact Hausdorff spaces* The Michigan Mathematical Journal, Vol 17, Issue 4. (1970) S. Lang: *Real and Functional Analysis* Springer-Verlag, New York, 1993. Takeuchi M. *Free Hopf algebras generated by coalgebras* Journal of the Mathematical society of Japan, Vol 23, Number 4, 1971. Vestberg M.: *The wave front set and oscillatory integrals* Master thesis (2015)

ArXiv

--- abstract: 'New criteria are established for upper bounds on the number of limit cycles of periodic Abel differential equations having two periodic invariant curves, one of them bounded. The criteria are applied to obtain upper bounds of either zero or one limit cycle for planar differential systems.' address: - 'Departamento de Matematicas, Universidad de Extremadura, 06071 Badajoz, Spain' - 'Departamento de Matematicas, Universidad de Extremadura, 06071 Badajoz, Spain' - 'Departamento de Matematicas, Universidad de Extremadura, 06071 Badajoz, Spain' author: - 'J.L. Bravo' - 'L.A. Calderón\*' - 'M. Fernández' title: Upper Bounds of Limit Cycles in Abel Differential Equations with Invariant Curves --- Introduction and Statements of Main Results =========================================== Bounding the number of periodic solutions of the generalized Abel equation $$\label{eq:genAbel} x'=\sum_{k=0}^n C_k(t) x^k,\qquad (t,x)\in\mathbb{R}^2,$$ where $C_k(t)$ are $T$-periodic smooth functions, is an open problem known as the Smale-Pugh problem [@S]. We shall denote by $u(t,x)$ the solution of Equation  determined by $u(0,x)=x$. Recall that a solution $u(t,x)$ is [*closed*]{} or [*periodic*]{} if $u(T,x)=x$, and that a [*limit cycle*]{} is a periodic solution which is isolated in the set of all periodic solutions. When $n=1$, Equation  is a linear equation, and it is known that there is at most one limit cycle. When $n=2$, Equation  is the well-known Ricatti equation, and has at most two limit cycles (see for example [@LN]). However, when $n=3$, A. Lins-Neto [@LN] proved that there exists no upper bound on the number of limit cycles when $C_k$ belongs to the whole family of trigonometric polynomials. At this point, in order to obtain upper bounds on the number of limit cycles of Equation , one may impose some additional constraints on its coefficients, $C_k(t)$ (see for example [@ABF2; @ABF4; @AGG; @BF13; @GG; @GL; @HL1; @HL4; @HL5; @HZ; @Ll; @P; @Pliss]). In this work, we study for $n=3$ with two invariant periodic curves of the form $a_0(t)x=b_0(t)$ and $a_1(t)x=b_1(t)$, where $a_0,b_0,a_1,b_1$ are smooth $T$-periodic functions, $a_0(t),a_1(t)\not\equiv 0$, and the first invariant curve is bounded, i.e., $a_0(t)\neq0$ for all $t\in[0,T]$. By the change of variables $x\to x-b_0(t)/a_0(t)$, we may assume the first invariant curve to be $x=0$. Therefore Equation can be written as $$x'=x(C_1(t)+C_2(t)x + C_3(t)x^2).$$ For the second invariant curve, we shall assume that $b_1(t)\equiv 1$, and that the set on which $a_1$ vanishes has null measure. If the functions $a_1,b_1$ are regular enough, these assumptions are automatically satisfied. For instance, if they are analytic, $a_1,b_1$ have a finite number of zeros in $[0,T]$ with bounded multiplicity, and, by uniqueness of solutions, the invariant curves are disjoint, so that every zero of $b_1$ is a zero of $a_1$, and this zero of $a_1$ has multiplicity greater than or equal to that of the zero of $b_1$. Hence, dividing by $b_1(t)$ if necessary, one may assume that $b_1(t)\equiv 1$. Dividing the quadratic polynomial, $C_1(t)+C_2(t)x + C_3(t)x^2$ by $a_1(t)x-1$, one gets (whenever $a_1(t)\neq 0$) $$\label{eq:Abelinv} x'=x\Big((a_1(t)x-1)(a_2(t)x-b_2(t))+ a_3(t)\Big),$$ for certain $T$-periodic smooth functions $a_2,b_2,a_3$ integrable in $[0,T]$, and possibly with poles at the zeros of $a_1$. Imposing that $a_1(t)x-1=0$ is invariant, one obtains that, for any $t$ such that $a_1(t)\neq 0$, reduces to $$\label{ode:1} \begin{split} x'=&(a_1x-1)(a_2x-b_2)x -\frac{a_1'}{a_1}x\\ =&a_1a_2 x^3-\left(a_1b_2+a_2\right)x^2+\left(b_2 - \frac{a_1'}{a_1}\right)x, \end{split}$$ where the arguments of the functions have been omitted for clarity. If $a_2(t)$ is identically null, is a Riccati equation with at most one non-null limit cycle (see for example [@LN]). So we shall assume that $a_2(t)$ is not identically null. The invariant curves divide the plane into connected components. Motivated by the applications to the planar systems, we may consider just two regions, gluing together the connected components separated by $x=\pm\infty$ (see Figure \[U\]). The objective of this paper is to obtain criteria for Equation to have either zero or at most one limit cycle with graph included in one of the regions, $U$. By the change of variables $x\to -x$, it is not restrictive to assume that $(t,x)\in U$ for $x>0$ small enough. [0.3]{} (150,220) (0,0)[![Region $U$.[]{data-label="U"}](V1.pdf "fig:"){width="110pt"}]{} (90,60)[U]{} (100,40)[t]{} [0.3]{} (150,220) (0,0)[![Region $U$.[]{data-label="U"}](V3.pdf "fig:"){width="110pt"}]{} (90,60)[U]{} (100,40)[t]{} [0.3]{} (150,220) (0,0)[![Region $U$.[]{data-label="U"}](V2.pdf "fig:"){width="110pt"}]{} (90,60)[U]{} (100,40)[t]{} Indeed, as limit cycles are bounded, we may reduce the study to $V\subset U$ defined as: - when $a_1(t)>0$ for all $t$, $$V=\{ (t,x): 0<x<1/a_1(t) \};$$ - when $a_1$ has zeros, $$V=\{ (t,x): x>0 \text{ and } 0<x<1/a_1(t) \text{ for } t \text{ with }a_1(t) >0 \};$$ - when $a_1(t)<0$ for all $t$, $$V= \{(t,x): x>0 \text{ or } x<1/a_1(t) \}.$$ Let $\bar b_1$ be a smooth $T$-periodic function such that $\bar b_1(t)\neq0$ for every $t$, and denote $$\bar a_1(t)=a_1(t) \bar b_1(t),\quad \bar a_2(t)= \frac{a_2(t)}{\bar b_1(t)}, \quad \bar b_2(t)=\frac{b_2(t)+\frac{\bar a_1'(t)}{\bar a_1}}{\bar b_1(t)}.$$ Then Equation  becomes $$\label{eq:HL} x'=(\bar a_1x-\bar b_1)(\bar a_2x-\bar b_2)x + \frac{\bar b_1'-\bar a_1' x}{\bar b_1}x.$$ Motivated by the study of limit cycles in planar systems, this equation was proposed by Huang, Liang, Llibre in [@HL5] who, by assuming that $\bar a_2,\bar b_2$ have no poles, obtained the following criterion: [@HL5 Proposition 10] \[prop:hl\] Equation has at most one limit cycle (counted with multiplicity) with graph included in $V$ if, for all $t \in [0,T]$, one of the following conditions holds: - $\bar a_1(t)\neq0$ and there exists $\eta\in\mathbb{R}$ such that $\bar a_1(t) \bar b_2(t)+\eta \bar a_2(t) \bar b_1(t)+ (\bar a_1'(t)/\bar b_1(t)) \geq0\ (\leq0)$; - $\bar a_2(t) \geq 0$ ($\leq 0$); - Either $\bar a_1(t) \bar a_2(t) \geq0$ and $\bar b_1(t)\bar b_2(t) \leq0$, or $\bar a_1(t) \bar a_2(t) \leq0$ and $\bar b_1(t)\bar b_2(t) \geq0$. The objective of the paper is to establish new criteria for Equation to have either zero or at most one limit cycle (counted with multiplicity) with graph in $V$. We say a $T$-periodic function $f$ has positive definite sign if $f(t)\geq 0$ for all $t\in[0,T]$. Analogously, $f$ has negative definite sign if $f(t)\leq 0$ for all $t\in[0,T]$ and $f$ has definite sign if it has either positive or negative definite sign. The main results are: \[criterium1\] Assume there exists $\eta\in\mathbb{R}$ such that $b_2+\eta a_1'/a_1$ has positive (negative) definite sign and for each $t\in[0,T]$ the following conditions hold: - if $a_1(t)< 0$ then $a_2(t)\leq 0$ ($\geq 0$); - if $a_1(t)> 0$ then $a_1(t)b_2(t)-a_2(t)+\eta a_1'(t) \geq 0$, ($\leq0$); with the inequalities in $(i)$ and $(ii)$ being strict for all $t\in P$, where $P$ is a positive measure set of $[0,T]$. Then Equation has no non-trivial limit cycle with graph included in $V$. \[criterium2\] Assume there exists $\eta\in\mathbb{R}$ such that $b_2+\eta a_1'/a_1$ has positive (negative) definite sign, $b_2$ is not identically null, and for each $t\in[0,T]$ the following conditions hold: - if $a_1(t)< 0$, then $a_2(t)\geq 0$ ($\leq 0$); - if $a_1(t)> 0$ then $a_1(t)b_2(t)-a_2(t)+\eta a_1'(t) \geq 0$, ($\leq0$); with the inequalities in $(i)$ and $(ii)$ being strict for all $t\in P$, where $P$ is a positive measure set of $[0,T]$. Then Equation has at most one non-trivial limit cycle with graph included in $V$. Moreover, this upper bound is attained. Note that $a_1(t)$ and $a_2(t)$ can have any number of zeros. Moreover, since $a_2,b_2$ may also have poles, $\eta$ must be chosen such that $b_2+\eta a_1'/a_1$ has only poles of even order. An important application of the results above is to bound the number of limit cycles of polynomial planar systems. Indeed, many planar systems reduce to an Abel equation after a change of variables (see [@DLP; @Lloyd] for example). In these systems, the origin is a singular point that is transformed into the periodic solution $x=0$, so that, if a second invariant curve is known, the equation is of the type and Theorems \[criterium1\] and \[criterium2\] apply. In Section 4, we shall detail this for homogeneous planar systems studied in [@CLl; @HL5] and compare the previous known criteria with the new ones obtained for this case. Preliminaries {#sec:2} ============= Consider the equation $$\label{eq:esc} x'=p(t,x),$$ where $p \colon \mathbb{R}^2 \to \mathbb{R}$ is sufficiently smooth, $T$-periodic with respect to $t$, i.e., $p(t,x)=p(t+T,x)$, and polynomial in $x$. Let $A=\{x\in\mathbb{R}: u(T,x) \text{ is defined}\}$. Denote $d(x)=u(T,x)-x$, the displacement application. Then, $x_0$ is a zero of $d(x)$ if and only if $u(t,x_0)$ is a periodic solution of . So isolated zeros of $d(x)$ are initial conditions of limit cycles of . Deriving with respect to the initial condition, one obtains $$\label{eq:d'} d'(x)=\exp\left(\int_{0}^{T}p_x(t,u(t,x))\,dt\right)-1.$$ When $d(x)=0$ and $d'(x)\neq0$, the limit cycle $u(t,x)$ is called hyperbolic. In this case the sign of $d'(x)$ determines its stability. Using elementary analysis, it is easy to prove the following result: Let $I\subset A$ be an interval, and assume that $d'(x)> 0$ (resp. $<0$) for all $x\in I$ such that $u(t,x)$ is periodic. Then there is at most one periodic solution with initial condition in $I$. Let $q(t,x)$ be a polynomial in $x$ with coefficients that are smooth, $T$-periodic with respect to $t$, functions. In this context, we say that $q(t,x)=0$ is a periodic algebraic invariant curve if there exists $Q(t,x)$ which is $T$-periodic with respect to $t$ and polynomial in $x$, called the cofactor of $q(t,x)$, such that $$q_t(t,x)+q_x(t,x)p(t,x)=q(t,x) Q(t,x).$$ Note that the coefficients of $Q$ as a polynomial in $x$ are smooth functions that might have poles at the zeros of the coefficients of $q(t,x)$. Let $u(t,x)$ be a periodic solution of . By uniqueness of solutions, either $q(t,u(t,x))\neq 0$ for all $t$ or $q(t,u(t,x))\equiv 0$. Assume that the first possibility is the case and that the set of $t$ such that the coefficients of $Q(t,x)$ have poles has null measure. Then $$\int_0^T Q(t,u) \, dt=\int_0^T \frac{ q_t(t,u)+q_x(t,u)p(t,u)}{q(t,u)}\,dt=\int_0^T \frac{d\left(\ln|q(t,u)|\right)}{dt}\,dt=0,$$ where $u=u(t,x)$. Therefore, for any $\alpha\in\mathbb{R}$, the sign of $d'(x)$ is the sign of $$\int_{0}^{T} \left(p_x(t,u) + \alpha Q(t,u) \right) \,dt.$$ Now consider . In this case, $$p_x(t,x)=3a_1a_2x^2-2\left(a_1b_2+a_2\right)x+b_2-\frac{a_1'}{a_1},$$ and the cofactors of $x=0$ and $a_1x-1=0$ are $$p_1(t,x)=(a_1 x - 1) (a_2 x - b_2) -\frac{a_1'}{a_1}$$ and $$p_2(t,x)= a_1 x (a_2 x - b_2).$$ Moreover, since $a_1$ is $T$-periodic, then $$\int_0^T \frac{a_1'}{a_1}\,dt=0.$$ For any $\alpha,\beta,\eta\in\mathbb{R}$, denote $$\label{eq:G} \begin{split} G(t,x)=& p_x(t,x)+\alpha p_1(t,x)+\beta p_2(t,x) +(1+\alpha + \eta) \frac{a'_1(t)}{a_1(t)} \\ &=(3 + \alpha + \beta) a_1 a_2 x^2 - \left((2 + \alpha) a_2 + (2 + \alpha + \beta) a_1 b_2\right) x \\ &+ (1 + \alpha)b_2 +\eta\frac{a_1'}{a_1}. \end{split}$$ A direct consequence is the following result: \[lema:principal\] Assume there exist $\alpha,\beta,\eta\in\mathbb{R}$ such that for every $(t,x)\in V$ except for $t$ in a zero-measure set $$G(t,x) \geq 0\quad \text{(resp. }\leq 0\text{)},$$ and there exists a positive measure set $P$ such that for every $t\in P$ the inequality is strict. Then there is at most one limit cycle in each connected component of $V$. Moreover, if one exists then it is hyperbolic and unstable (resp. stable). To check when $G(t,x) \geq0$ (resp. $\leq0$) for each $t\in[0,T]$, we shall use Sturm’s theorem applied to $G$ as a polynomial in $x$. Recall that the Sturm sequence for $G$ when $a_1(t),a_2(t)\neq0$ and $\alpha+\beta+3\neq 0$ is $\{q_0(x),q_1(x),q_2\}$, where $q_0(x)=G(t,x)$, $q_1(x)=G_x(t,x)$, and $$q_2=\frac{(2 + \alpha)^2 a_2 }{4 (3 + \alpha + \beta) a_1} +\frac{ (2 + \alpha + \beta)^2 a_1 b_2^2}{4 (3 + \alpha + \beta) a_2} -\frac{ (2 + \alpha^2 + \alpha (4 + \beta)) b_2}{2 (3 + \alpha + \beta)}+\eta \frac{a_1'}{a_1}.$$ Let $a\in\mathbb{R}$ with $q_0(a)\neq0$. Define $v(a)$ as the number of changes of sign in the sequence $\{q_0(a), q_1(a), q_2\}$, omitting the zeros. \[[Sturm’s theorem, see for example [@BS p. 297]]{}\] If $a<b$, $q_0(a)\neq0$, and $q_0(b)\neq0$, then the number of different zeros (counting the multiple roots just once) of $q_0(x)$ in $[a,b]$ is $v(a)-v(b)$. Thus, fixed $t\in[0,T]$, $a,b\in\mathbb{R}$ such that $G(t,a)$ and $G(t,b)$ differ from $0$, if $v(a)=v(b)$, then $G(t,x)$ has definite sign in the interval $(a,b)$. Note that Sturm’s theorem is also valid for $a=-\infty$ or $b=+\infty$. Proofs of the Main Results ========================== The key to the proof of Theorems \[criterium1\] and \[criterium2\] is to choose special values for $\alpha,\beta, \eta$ to simplify , and then to apply Sturm’s theorem so as to obtain the conditions for $G\geq 0$ ($G\leq 0$) in $V$, when necessary. In the case of $a_1(t)<0$ for all $t\in[0,T]$, our results imply a certain definite sign of $a_2$. But a more general criterion can be stated. This criterion was already obtained in [@HL3], but we include it here for completeness. \[prop:signo\_definido\] Assume that $a_2$ has definite sign in $[0,T]$, strict for some positive measure set. Then Equation has at most one non-trivial limit cycle with graph included in $V$. Firstly, take $\alpha=\beta=-1$, and $\eta=0$. Then $$G(t,x)=a_2(t)(a_1(t)x-1)x.$$ Therefore, if $a_2$ has definite sign, strict for some positive measure set, the same is the case for $G(t,x)$, and there is at most one limit cycle with graph in each of the connected components of $V$. Thus it suffices to prove that if $V$ has two connected components, i.e., if $a_1(t)<0$ for all $t$, then there is at most one limit cycle with graph in $V$. The transformation $y=a_1(t)x$ reduces to $$\label{eq:y} y'=\frac{1}{a_1(t)}y(y-1)\left(a_2(t)y-a_1(t)b_2(t)\right),$$ and $V$ to $\tilde{V}=\{(t,y):y<0 \text{ or } y>1\}$. If we assume that $a_2$ has definite sign, strict in some positive measure set, then the result follows comparing the solutions of with the solutions of the separated variables equation $y'=b_2(t) y(y-1)$. See [@HL3 Lemma 2.3] for the details. Now let us prove the first theorem. Firstly, if $a_1(t)<0$ for every $t$ then, by Proposition \[prop:signo\_definido\], we conclude there is at most one positive limit cycle with graph in $V$. So we assume $a_1$ has zeros or $a_1>0$. Choosing $\eta$ conveniently and $t$ such that $a_1(t)\neq 0$, we may write $$\begin{split} G(t,x)=& (3 + \alpha + \beta) a_1(t) a_2(t) x^2 - \left((2 + \alpha) a_2(t) + (2 + \alpha + \beta) a_1(t) b_2(t)\right) x \\&+ (1 + \alpha)\left( b_2(t) +\eta\frac{a_1'(t)}{a_1(t)}\right). \end{split}$$ Consider  with $\alpha + \beta+2=0$. Then, for any $t$ such that $a_1(t)\neq0$, $$G(t,x)= a_1(t)a_2(t) x^2-(\alpha+2) a_2(t)x +(\alpha+1)\left( b_2(t) +\eta\frac{a_1'(t)}{a_1(t)}\right).$$ Assume $b_2(t)+\eta a_1'(t)/a_1(t) \geq0$ for all $t\in[0,T]$, with the case $\leq0$ being analogous. Fix $\alpha>0$. We shall prove that $G(t,x) \geq 0$ for every $(t,x)\in V$ such that $a_1(t)\neq0$, with strict inequality for every $t$ in some positive measure set. Fix any $t\in [0,T]$ such that $a_1(t)\neq 0$. First, we shall consider the singular cases. <span style="font-variant:small-caps;">Case 1</span>. Assume $a_2(t)=0$. Then trivially $G(t,x)\geq 0$ for every $x\geq 0$. <span style="font-variant:small-caps;">Case 2</span>. Assume $b_2(t)+\eta a_1'(t)/a_1(t)=0$ and $a_2(t)\neq 0$. Then $$G(t,x)=a_1(t)a_2(t)x^2-(\alpha+2)a_2(t)x.$$ If $a_1(t)<0$ then, by hypothesis, $a_2(t)\leq 0$. Then $a_1(t)a_2(t)> 0$, and $G(t,x)\geq 0$ for every $x \geq 0$. If $a_1(t)>0$, we only need to check that $G(t,x)\geq 0$ for all $x\in (0,1/a_1(t))$. By hypothesis, $a_1(t)b_2(t)-a_2(t)+\eta a_1'(t)\geq 0$, and since $a_1(t)b_2(t)+\eta a_1'(t)=0$ then $a_2(t)< 0$. As the function can be rewritten as $$G(t,x)=a_2(t)(a_1(t)x-(\alpha+2))x,$$ and $x\geq 0$, we only need to check that the linear function $G^*(t,x)=a_2(t)(a_1(t)x-(\alpha +2))$ is positive in the interval, so we must verify that it is positive at the extrema of the interval $(0,1/a_1(t))$. But $$G^*(t,0)=-a_2(t)(\alpha+2)> 0,\quad G^*(t,1/a_1(t))=a_2(t)\left(1-(\alpha+2) \right)> 0.$$ For the remaining cases, consider the Sturm sequence $$S(x)=\{q_0(x),q_1(x),q_2\},$$ where $$\begin{aligned} q_0(x)&=G(t,x),\\ q_1(x)&=G_x(t,x)= 2 a_1(t)a_2(t)x-(\alpha+2)a_2(t),\\ q_2\ \ &= \frac{(\alpha+2)^2a_2(t)}{4 a_1(t)}+\frac{(\alpha+2)(\alpha+\beta+2)}{2(\alpha+\beta+3)}b_2(t) - (\alpha+1)\left(b_2(t)+\eta \frac{a_1'(t)}{a_1(t)} \right).\end{aligned}$$ <span style="font-variant:small-caps;">Case 3</span>. Let $t$ be such that $a_1(t)< 0$ and $a_2(t)(b_2(t)+\eta a_1'(t)/a_1(t)) \neq 0$. The Sturm sequences at $x=0$ and $x=\infty$ are $$\begin{aligned} &S(0)=\{(\alpha+1)(b_2(t)+\eta a_1'(t)/a_1(t)),-(\alpha+2)a_2(t),q_2 \}, \\ &S(\infty)=\{a_1(t)a_2(t),2a_1(t)a_2(t),q_2 \}.\end{aligned}$$ By hypothesis, $a_2(t)< 0$. Then $a_1(t)a_2(t)> 0$. Therefore, the Sturm sequences $S(0)$ and $S(\infty)$ have the same changes of sign as the sequence $$\{ 1,1,q_2\}.$$ Hence, $G(t,x)$ has no zeros in $(0,\infty)$. As $G(t,0)> 0$, one has that $G(t,x)> 0$ for all $x\in(0,\infty)$. <span style="font-variant:small-caps;">Case 4</span>. Let $t$ be such that $a_1(t)> 0$ and $a_2(t)(b_2(t)+\eta a_1'(t)/a_1(t)) \neq 0$. The Sturm sequences at $x=0$ and $x=1/a_1(t)$ are $$\begin{aligned} & S(0)=\{(\alpha+1)(b_2(t)+\eta a_1'(t)/a_1(t)),-(\alpha+2)a_2(t),q_2 \}, \\ & S\left(\frac{1}{a_1(t)}\right)=\left\{\frac{(1 + \alpha) ( a_1(t) b_2(t) -a_2(t) + \eta a_1')}{a_1(t)}, -\alpha a_2(t),q_2 \right\}.\end{aligned}$$ Since $b_2(t)+\eta a_1'(t)/a_1(t)>0$ and, by hypothesis, $a_1(t)b_2(t)-a_2(t)+\eta a_1'(t)\geq0$, if the inequality is strict then the Sturm sequences $S(0)$ and $S(1/a_1(t))$ have the same signs as the sequence $$\{1,-a_2(t),q_2\}.$$ Hence $G(t,x)$ has no zeros in the interval $(0,1/a_1(t))$. As $G(t,0)> 0$, one has that $G(t,x)>0$ for all $x\in(0,1/a_1(t))$. If $a_1(t)b_2(t)-a_2(t)+\eta a_1'(t)=0$, then $a_2(t)>0$ and $G(t,x)= a_2(t)(a_1(t)x^2-(\alpha+2)x+(\alpha+1)/a_1(t))>0$ for each $x \in (0,1/a_1(t))$, because the function $g(t,x)= G(t,x)/a_2(t)$ satisfies $g(t,0)>0, g(t,1/a_1(t))=0$, and $g_x(t,x)<0$ for each $x \in (0,1/a_1(t))$. So far we have obtained that there is at most one non-trivial limit cycle with graph included in $V$. Now we will show that in this case there is no such limit cycle. Assume that $b_2(t)+\eta a_1'(t)/a_1(t)$ differs from zero on a positive measure set. To prove that there is no limit cycle with graph in $V$, note that $$d'(0)= \exp \left(\int_{0}^{T}b_2(t)+\eta \frac{a_1'(t)}{a_1(t)}dt\right)-1 >0,$$ so that the origin is unstable. By Lemma \[lema:principal\], if there exists a limit cycle with graph in $V$ then it is unstable, so that there is no limit cycle in the connected component of $V$ contained in $x\geq 0$. If $b_2(t)+\eta a_1'(t)/a_1(t)\equiv 0$ then $d'(0)=0$ and $$d''(0)=-2\int_{0}^{T} a_2(t) \,dt >0.$$ Hence, the null solution is unstable, so that there is no such limit cycle in the connected component of $V$ contained in $x\geq 0$. Assume now that $a_1(t)<0$ for every $t$. In order to prove that there is no limit cycle contained in the connected component of $V$ contained in $x\leq 0$, note that the stability of the limit cycle $x(t)=1/a_1(t)$ is $$d'(1/a_1(0))= \exp\left(\int_0^T \left(a_1(t)b_2(t)-a_2(t) + \eta a_1'(t)/a_1(t)\right) \,dt \right)-1 >0,$$ so that the cycle $x(t)$ is unstable. Again by Lemma \[lema:principal\], if there exists a limit cycle with graph in $V$ then it is unstable, so that there is no limit cycle in the connected component of $V$ contained in $x\leq 0$. Now, let us prove the second main result. If $a_1(t)\leq0$ for every $t$, then $a_2$ has definite sign, and by Proposition \[prop:signo\_definido\] there is at most one limit cycle with graph contained in $V$. So we may assume that either $a_1(t)>0$ for every $t$ or $a_1$ has zeros. Consider  with $\alpha=-2$, $\beta=0$, and $\eta$ such that $$G(t,x)= a_1(t)a_2(t) x^2 -\left(b_2(t)+\eta\frac{a_1'(t)}{a_1(t)}\right).$$ We shall prove that $G(t,x)$ has definite sign in $V$ for every $t$ such that $a_1(t)\neq 0$. Firstly, assume that $b_2(t)+\eta a_1'(t)/a_1(t) \geq0$, with the other case being analogous. We shall prove that $G(t,x) \leq 0$ for all $(t,x)\in V$. Fix $t\in [0,T]$. If $a_1(t)< 0$ then, by hypothesis, $a_2(t)\geq 0$, so that $a_1(t)a_2(t)\leq 0$ and $G(t,x)\leq 0$ for all $x\in [0,+\infty)$. Now, if $a_1(t)> 0$, as the only extreme of $G(t,x)$ is at $x=0$, and $G(t,0)\leq 0$, we only need to prove that $G(t,1/a_1(t))\leq 0$. But, if $b_2(t)\neq 0$ then $$G\left(t,\frac{1}{a_1(t)}\right)=\frac{a_2(t)}{a_1(t)} - \left(b_2(t)+\eta\frac{a_1'(t)}{a_1(t)}\right)=\frac{a_2(t)-b_2(t)a_1(t)-\eta a_1'(t)}{a_1(t)}\leq 0.$$ Finally, to prove that this upper bound is attained, it suffices to take $a_1,a_2,b_2\in\mathbb{R}$, $a_1>0$, $0<b_2/a_2<a_1$. Planar Polynomial Differential Systems with Homogeneous Singularities ===================================================================== In this section, we show some examples of applications of the results to planar systems. Rigid systems {#rigid-systems .unnumbered} ------------- A family that is easily transformed into Abel equations (or into generalized Abel equations) is that of rigid systems, i.e., systems of the form $$\label{eq:rigid} \begin{cases} x'=-y+x p(x,y),\\ y'=\phantom{-}x+yp(x,y), \end{cases}$$ where $p$ is a polynomial. These systems have been studied in [@GPT; @GT] for instance. After a change to polar coordinates, becomes $$\label{eq:rigidabel} r'=rp(r\cos\theta,r\sin\theta),\quad \theta'=1.$$ Moreover, there is a correspondence between limit cycles of and positive limit cycles of . In order to have an Abel equation (though the methods of Section \[sec:2\] could be applied for any degree) with sufficient degrees of freedom to apply Theorem \[criterium2\], let $p$ be of the form $$p(x,y)=p_{00}+\sum_{i=0}^k p_{i,k-i} x^i y^{k-i}+\sum_{i=0}^{2k} p_{i,2k-i} x^i y^{2k-i}.$$ After the change of variables $\rho=r^k$, becomes the Abel equation $$\label{eq:rigidrho} \begin{split} \rho'=&k p_{00}\rho+k\left(\sum_{i=0}^k p_{i,k-i} \cos^i \theta \sin^{k-i} \theta\right)\rho^2\\ &+ k \left(\sum_{i=0}^{2k} p_{i,2k-i} \cos^i \theta \sin^{2k-i} \theta \right)\rho^3. \end{split}$$ If has an invariant curve with poles then the study of positive limit cycles can be reduced to the study of limit cycles with graph in $V$. Consider  with $$p(x,y)=1 - \frac{1}{2} x^4 y^2 + x^3 y^3 - \frac{5}{2} x^2 y^4 + x y^5 - 2 x^6 y^6 + 3 x^5 y^7 - x^4 y^8.$$ It is easy to check that $xy=1$ is an invariant curve for the planar system. Moreover, $x^3y^3=1$ is also an invariant curve. Then, after the change to polar coordinates and the change $\rho=r^6$, $\rho \cos^3\theta\sin^3\theta=1$ is an invariant curve of . Dividing by $a_1(\theta) \rho -1$, where $a_1(\theta)=\cos^3\theta\sin^3\theta$, one obtains $$a_2(\theta)= -12 \cos^3 \theta \sin^3\theta + 18 \cos^2 \theta \sin^4\theta - 6 \cos \theta \sin^5\theta,$$$$b_2(\theta)=6 + 3 \cot \theta - 3 \tan \theta.$$ Now let us check that Theorem \[criterium2\] is satisfied. In order for $b_2+\eta a_1'/a_1$ to have definite sign, we must choose $\eta=-1$, obtaining $b_2+\eta a_1'/a_1=6$. To verify that (i) and (ii) are satisfied, note that $a_1$ and $a_2$ are $\pi$-periodic, so that it suffices to show that they are satisfied in $(-\pi/2,\pi/2)$. Dividing by $\cos^6\theta$, one obtains $$\frac{a_1(\theta)}{\cos^6 \theta}=\tan^3 \theta,\quad \frac{a_2(\theta)}{\cos^6 \theta}=-12 \tan^3\theta + 18 \tan^4\theta - 6 \tan^5\theta.$$ Now we define the auxiliary polynomials $$p_1(t)=\frac{a_1(\arctan t)}{\cos^6 \arctan t}=t^3,\quad p_2(t)=\frac{a_2(\arctan t)}{\cos^6 \arctan t}=-12 t^3 + 18 t^4 - 6 t^5,$$$$p_3(t)=\frac{a_1(\arctan t)b_2(\arctan t)-a_2(\arctan t)-a_1'(\arctan t)}{\cos^6 \arctan t}= 6 t^3 (3 - 3 t + t^2).$$ Note that the signs of $a_1,a_2,a_1b_2-a_2+\eta a_1'$ in $\theta\in(-\pi/2,\pi/2)$ are the signs of $p_1,p_2,p_3$ in $t=\tan \theta\in \mathbb{R}$, respectively. Therefore, as $p_1(t)$ is negative for $t<0$ and positive for $t>0$, it only remains to check that $p_2(t)>0$ for $t<0$ and that $p_3(t)>0$ for $t>0$. But this is immediate as $p_2$ has the roots $0,1,2$ and $p_3$ has only the root $0$ (and two complex roots). Moreover, as $a_1$, $a_2$, and $a_1a_2$ have changes of sign, the criteria in Proposition \[prop:hl\] are not satisfied. Homogeneous systems {#homogeneous-systems .unnumbered} ------------------- Consider the homogeneous planar system studied in [@CLl; @HL5]: $$\begin{aligned} \label{sistema plano} x'&= a x -y +P_n(x,y),\\ y'&= x+a y +Q_n(x,y), \notag\end{aligned}$$ where $a\in\mathbb{R}$ and $P_n,Q_n$ are homogeneous polynomials of degree $n$. In polar coordinates, is written as $$\begin{aligned} r'&= a r+ \varphi (\theta)r^n,\\ \theta'&=1+ \psi(\theta)r^{n-1},\notag \end{aligned}$$ where $$\begin{aligned} \varphi(\theta)&=P_n(\cos\theta,\sin\theta) \cos\theta+Q_n(\cos\theta,\sin\theta)\sin\theta,\\ \psi(\theta)&= Q_n(\cos\theta,\sin\theta)\cos\theta-P_n(\cos\theta,\sin\theta)\sin\theta. \end{aligned}$$ Note that $\psi(\theta)$ has a finite number of zeros in $[0,2\pi]$. Since limit cycles of surrounding the origin do not intersect the curve in the $(r, \theta)$ plane, $1+ \psi(\theta)r^{n-1}=0$ (see [@CLl]), they can be determined through the limit cycles of the scalar equation $$\label{ode:1polar} \frac{dr}{d\theta}= \frac{a r+ \varphi (\theta)r^n}{1+ \psi(\theta)r^{n-1}}, \quad \theta \in \mathbb{R}.$$ We may assume that $\psi(\theta)$ is not identically null since otherwise would be a Ricatti equation that has at most one non-null limit cycle. Now, using Cherkas’s change of variable [@Che] $$\rho= \frac{r^{n-1}}{1+ \psi(\theta)r^{n-1}},$$ Equation transforms into the Abel equation $$\label{ode:1cherkas} \begin{split}\frac{d\rho}{d\theta}&=(\psi \rho -1)((n-1)(a\psi-\varphi)\rho-(n-1)a)\rho-\psi' \rho^2 \\ &=(n-1)(a\psi-\varphi)\psi\rho^3+((n-1)(\varphi-2a\psi)-\psi')\rho^2+ (n-1)a \rho, \end{split}$$ with invariant curves $\rho=0$ and $\psi(\theta)\rho-1=0$. Moreover, as the change of variable is equivalent to $$r^{n-1}=\frac{\rho}{1-\psi(\theta)\rho},$$ the set $\{(\theta,r) : \theta \in\mathbb{R}, r>0 \}$ is transformed into the region $V$ defined above (see [@CLl; @GL] for more details). Therefore, Equation is of the form with $$a_1= \psi, \quad a_2= (n-1)(a\psi-\varphi), \quad b_2= (n-1)a+\frac{\psi'}{\psi}.$$ In order to apply Theorems \[criterium1\] and \[criterium2\], it is necessary that $$b_2+\eta a_1'/a_1=(n-1)a-(1+\eta)\frac{\psi'}{\psi}$$ has definite sign. To allow $\psi$ to have zeros, we choose $\eta=-1$. Now, Theorems \[criterium1\] and \[criterium2\] have the following consequences, respectively: \[coro:1\] Assume that $a> 0\ (<0)$ and that for each $\theta \in[0,2\pi]$ the following conditions hold: - if $\psi(\theta)< 0$ then $a\psi(\theta)-\varphi(\theta)\leq 0\ (\geq0)$, - if $\psi(\theta)>0$ then $\varphi(\theta)\geq 0 \ (\leq0)$, with the inequalities in (i) and (ii) being strict for every $\theta$ in some positive measure set. Then Equation  has no limit cycles surrounding the origin. Assume that $a> 0\ (<0)$ and that for each $\theta \in[0,2\pi]$ the following conditions hold: - if $\psi(\theta)< 0$ then $a\psi(\theta)-\varphi(\theta)\geq 0\ (\leq0)$, - if $\psi(\theta)>0$ then $\varphi(\theta)\geq 0 \ (\leq0)$, with the inequalities in (i) and (ii) being strict for every $\theta$ in some positive measure set. Then Equation  has at most one limit cycle surrounding the origin. We shall show an example of a planar polynomial differential system with homogeneous cubic nonlinearities which satisfies conditions (i) and (ii) in Corollary \[coro:1\] but does not satisfy any of the conditions in [@HL5 Theorem 1, Theorem 3, and Corollary 4]. To that end, consider . A first comment is that when the functions $\psi$ and $\varphi$ are homogeneous trigonometric polynomials of odd degree then conditions (i) and (ii) in Corollary \[coro:1\] imply that $\varphi$ and $\psi$ have the same zeros. Indeed, in this case, the functions satisfy $\psi(\pi+\theta)=-\psi(\theta)$, $\varphi(\pi+\theta)=-\varphi(\theta)$, $\theta\in[0,2\pi]$. Assume $a>0$, with the other case being analogous. By Corollary \[coro:1\], one has: 1. If $\psi(\theta)< 0$ then $a\psi(\theta)-\varphi(\theta)\leq 0$. Moreover, as $\psi(\pi+\theta)=-\psi(\theta)>0$ then $\varphi(\theta+\pi)=-\varphi(\theta)\geq 0$. Hence $a\psi(\theta)\leq \varphi(\theta)\leq0$. 2. If $\psi(\theta)>0$ then $\varphi(\theta)\geq 0$. Since $\psi(\theta+\pi)<0$ then $a\psi(\pi+\theta)-\varphi(\theta+\pi)\leq0$, i.e., $a \psi(\theta)-\varphi(\theta) \geq0$. Hence $ a\psi(\theta)\geq \varphi(\theta)\geq0$. In particular, $\psi(\theta)$ and $\varphi(\theta)$ have the same zeros and $(a\psi-\varphi)\varphi$ has positive definite sign, and the number of limit cycles is bounded (see [@CLl; @Pliss]), so we shall look for examples with more degrees of freedom. Let us consider the cubic system $$\label{systemplanar} \begin{split} x'&= a x-y+P_3(x,y), \\ y'&=x+ay+Q_3(x,y), \end{split}$$ where $$\begin{aligned} P_3(x,y)&=p_3x^3+p_2x^2y+p_1xy^2+p_0y^3,\\ Q_3(x,y)&=q_3x^3+q_2x^2y+q_1xy^2+q_0y^3, \end{aligned}$$ with $p_0,\dots,p_3,q_0,\dots,q_3 \in\mathbb{R}$. Then $$\begin{aligned} \varphi(\theta)&=p_3 \cos^4\theta+ (p_2+q_3) \cos^3\theta \sin\theta+ (p_1+q_2) \cos^2\theta \sin^2\theta \\ &\phantom +(p_0+q_1) \cos\theta \sin^3\theta +q_0 \sin^4\theta, \\ \psi(\theta)&= q_3 \cos^4\theta+ (q_2-p_3) \cos^3\theta \sin\theta+ (q_1-p_2) \cos^2\theta \sin^2\theta \\ &\phantom +(q_0-p_1) \cos\theta \sin^3\theta -p_0 \sin^4\theta.\end{aligned}$$ Note that $\varphi(\theta)=\varphi(\theta+\pi)$ and $\psi(\theta)=\psi(\theta+\pi)$. Then it suffices to check (i) and (ii) of Corollary \[coro:1\] for $\theta \in [-\pi/2,\pi/2)$. Therefore, we can divide both $\psi$ and $\varphi$ by $\cos^4\theta$ to obtain a polynomial in $\tan \theta$. Note that this transformation does not affect the criteria. The change of variables $t=\tan\theta$ transforms the interval $(-\pi/2,\pi/2)$ into $\mathbb{R}$. Define the polynomials $$P_\psi(t)=\frac{\psi(\arctan t)}{\cos^4\arctan t}=- p_0 t^4 +(q_0 - p_1)t^3 + (q_1-p_2)t^2 +(q_2-p_3)t + q_3,$$ $$P_\varphi(t)=\frac{\varphi(\arctan t)}{\cos^4\arctan t}= q_0 t^4 + (p_0+q_1)t^3+(p_1+q_2)t^2+(p_2+q_3)t+p_3.$$ Then the signs of $\psi,\varphi,a\psi-\varphi$ in $\theta\in(-\pi/2,\pi/2)$ are the same as the signs of $P_\psi,P_\varphi,a P_\psi-P_\varphi$ in $t=\tan \theta\in\mathbb{R}$. Now choose $$\label{coefficients}\begin{split} a=1/2, \ &p_0=-1,\quad p_1=\frac{20731}{20000}, \quad p_2=\frac{-19}{1000}, \quad p_3=\frac{9}{10000},\\ & q_0=\frac{1}{2},\quad q_1= \frac{2}{5},\quad q_2=\frac{-17631}{20000}, \quad q_3=0. \end{split}$$ With this choice of coefficients, one has $$P_\psi(t)=\frac{(t-1 ) t (17649 + 9269 t + 20000 t^2)}{20000}.$$ So $\psi(t)\leq 0$ for $t\in [0,1]$, and $\psi(t)>0$ elsewhere. Also, $$P_\varphi(t)=\frac{(10 t-9) (10t-1) (1 - 10 t + 50 t^2)}{10000},$$ which has just two simple zeros, at $1/10$ and $9/10$, and is positive for $t\not\in (1/10,9/10)$, so that (ii) of Corollary \[coro:1\] is satisfied. Now, to check that (i) is also satisfied, it suffices to verify that $a P_\psi(t)-P_\varphi(t)\leq 0$ for $t\in[0,1]$. This can be done by using Sturm’s theorem or with any CAS with the appropiate command (for instance, in Mathematica the command [*CountRoots*]{}). Finally, we verify that system with coefficients given by does not satisfy the hypotheses in Theorem 1, Theorem 3, and Corollary 4 in [@HL5]. Define $$\begin{split} \omega_1(\theta)&=a \psi(\theta)-\varphi(\theta),\\ \omega_2(\theta)&= (n-1)(2a\psi(\theta)-\varphi(\theta))+ \psi'(\theta). \end{split}$$ We claim: 1. There is no linear combination $\mu_1\omega_1(\theta)+ \mu_2\omega_2(\theta)$ where $\mu_1,\mu_2 \in\mathbb{R}$ which has definite sign. Define $$P_1(t)=\frac{\omega_1(\arctan t)}{\cos^4\arctan t},\quad P_2(t)=\frac{\omega_2(\arctan t)}{\cos^4\arctan t}.$$ Recall that the signs of $P_1,P_2$ in $t\in\mathbb{R}$ are the signs of $\omega_1,\omega_2$ in $\theta=\arctan t\in(-\pi/2,\pi/2)$. Then $P_1,P_2$ are of degrees three and four, respectively, both with positive leading coefficient. Moreover, $P_1$ has a zero in $[-1/2,1/2]$, $t_0$, while $P_2(t)<0$ for all $t\in [-1/2,1/2]$. As $\omega_1$ has a change of sign, take $\mu_2\neq 0$ and consider $Q(t)=\mu_1 P_1(t)+\mu_2 P_2(t)$. The sign of $Q(t_0)$ is opposite to that of $Q(t)$ for $t$ close to $\infty$, so that $Q$ has no definite sign for any $\mu_1,\mu_2$. 2. There are no $\nu_1,\nu_2\geq0$ such that $(a\nu_1)^2+ \nu_2^2 \neq0$ and $\omega_1(\theta)(\nu_1 a \psi(\theta)-\nu_2 \varphi(\theta))\leq0$. The sign of $\omega_1(\theta)(\nu_1 a \psi(\theta)-\nu_2 \varphi(\theta))$ is the same as the sign of $P_1(t)(\nu_1 a P_\psi(t)-\nu_2 P_\varphi(t))$. In order to have definite sign, since $P_1$ is of degree three then $\nu_1 a P_\psi(t)-\nu_2 P_\varphi(t)$ must be an odd degree polynomial. But it is easy to check that this holds only for $\nu_1=\nu_2$, so that there is no $\nu_1,\nu_2>0$ such that $\omega_1(\theta)(\nu_1 a \psi(\theta)-\nu_2 \varphi(\theta))\leq 0$ unless $\nu_1=\nu_2=0$. 3. Neither $\omega_1$ nor $\omega_2$ have definite sign. This is a consequence of (1). 4. Neither $-(n-1)\omega_1+\omega_2$ nor $-2(n-1)\omega_1+\omega_2$ have definite sign. This is a consequence of (1). 5. Neither $a\omega_1\psi$ nor $ \omega_1\varphi$ have definite sign. Again the signs of $a\omega_1\psi$ and $\omega_1\varphi$ are the signs of $aP_1 P_\psi$ and $P_1P_\varphi$. Computing the roots of the polynomials, it is easy to check that they have no definite sign. [0]{} A. [Álvarez]{}, J.L., Bravo, M. Fernández, [*The number of limit cycles for generalized Abel equations with periodic coefficients of definite sign*]{}, Commun. Pure Appl. Anal. [**8-5**]{}, (2009), 1493–1501. A. [Álvarez]{}, J.L., Bravo, M. Fernández, [*Limit cycles of Abel equations of the first kind*]{}, J. Math. Anal. Appl. [**423(1)**]{}, (2015), 734–745. M.J. Álvarez, A. Gasull, H. Giacomini, [*A new uniqueness criterion for the number of periodic orbits of Abel equations*]{}, J. Differential Equations [**234**]{}, (2007), 161–176. J.L. Bravo, M. Fernández, [*Limit cycles of non-autonomous scalar ODEs with two summands*]{}, Commun. Pure Appl. Anal. [**12-2**]{} (2013) 1091–1102. J. Stoer, R. Bulirsch, [*Introduction to numerical analysis*]{}, Springer, Texts in applied mathematics, 1992. M. Carbonell, J. Llibre, [*Limit cycles of a class of polynomial systems*]{}, Proc. Roy. Soc. Edinburgh [**109A**]{}, (1988), 187–199. L.A. Cherkas, [*Number of limit cycles of an autonomous second-order system*]{}, Differentsial’nye Uravneniya [**12**]{}, (1975), 944–946. J. Devlin, N.G. Lloyd, J.M. Pearson, [*Cubic systems and Abel equations*]{}, J. Differential Equations, [**147**]{}, (1998), 435–454. A. Gasull and A. Guillamon. [*Limit cycles for generalized Abel equations*]{}, Int. J. Bif. Chaos, [**16**]{}, (2006), 3737–3745. A. Gasull, J. Llibre, [*Limit cycles for a class of Abel equations*]{}, SIAM J. Math. Anal., [**21-5**]{} (1990), 1235–1244. A. Gasull, R. Prohens, J. Torregrosa, [*Limit cycles for rigid cubic systems*]{}, J. Math. Anal. Appl. [**303**]{}, (2005), 391–404. A. Gasull, J. Torregrosa, [*Some results on rigid systems*]{}, In International Conference on Differential Equations (Equadiff-2003), World Sci. Publ., Hackensack, NJ. (2005), 340–345. J. Huang, H. Liang, [*A uniqueness criterion of limit cycles for planar polynomial systems with homogeneous nonlinearities*]{}, J. Math. Anal. Appl., [**457(1)**]{}, (2018), 498–521. J. Huang and H. Liang, [ *A geometric criterion for equation $\dot{x} = \sum_{i=0}^m a_i(t)x^i$ having at most m isolated periodic solutions*]{}, J. Differential Equations, [**268(10)**]{} (2020), 6230–6250. J. Huang, H. Liang, [*Estimate for the number of limit cycles of Abel equation via a geometric criterion on three curves*]{}, Nonlinear Differ. Equ. Appl. (2017) 24: 47. https://doi.org/10.1007/s00030-017-0469-3. J. Huang, H. Liang, J. Llibre, [*Non-existence and uniqueness of limit cycles for planar polynomial differential systems with homogeneous nonlinearities*]{}, J. Differential Equations, [**265(9)**]{} (2018), 3888–3913, J. Huang, Y. Zhao, [*Periodic solutions for equation $x'=A(t)x^m+B(t)x^n+C(t)x^l$ with $A(t)$ and $B(t)$ changing signs*]{}, J. Differential Equations [**253**]{} (2012) 73–99. N.G. Lloyd, [*A note on the number of limit cycles in certain two-dimensional systems,*]{} J. London Math. Soc. [**20**]{} (1979), 277–286. N.G. Lloyd, C. J. Christopher, J. Devlin, J. M. Pearson, [*Quadratic-like Cubic Systems*]{}, Diff. Eq. and Dyn. Systems, [**3/4**]{} (1997), 329–345. A. Lins Neto, [*On the number of solutions of the equation $\frac{d x}{dt}=\sum_{j=0}^n a_j(t)x^j$, $0\leq t\leq 1$, for which $x(0)=x(1)$*]{}, Inv. Math. [**59**]{}, (1980), 67–76. A.A. Panov, [*The number of periodic solutions of polynomial differential equations*]{}, Mathematical Notes [**64-5**]{}, (1998), 622–628. V.A. Pliss, [*Non local problems of the theory of oscillations*]{}, Academic Press, New York, 1966. S.Smale, [*Mathematical problems for the next century*]{}, Mathematical Intelligencer, [**20**]{}, (1998), 7–15.

ArXiv

--- abstract: 'We study the propagation of a quantum probe light in an ensemble of “3+1”-level atoms when the atoms are coupled to two other classical control fields. First we calculate the dispersion properties, such as susceptibility and group velocity, of the probe light within such an atomic medium under the case of three-photon resonance via the dynamical algebra method of collective atomic excitations. Then we calculate the dispersion of the probe light not only under the case that two classical control fields have the same detunings to the relative atomic transitions but also under the case that they have the different detunings. Our results show in both cases the phenomenon of electromagnetically induced transparency can accur. Especially use the second case, we can find two transparency windows for the probe light.' author: - Yong Li - 'C. P. Sun' title: 'Propagation of Light in an Ensemble of “3+1”-Level Atoms' --- Introduction ============ The coherent interaction of atoms with optical fields has attracted much attention in studies of contemporary coherent and nonlinear optics [@scully2; @harris; @fleischhauer99]. One of the most interesting effects is electromagnetically induced transparency (EIT) [@EIT]. In an EIT system, the atoms are optically pumped into a [*so-called*]{} dark state which is decoupled from the original optical fields. Such an atomic medium possesses special optical properties such as cancellation of resonant absorption and slow group velocity of the reference probe light field. Generally, a conventional EIT system consists of a vapor cell with 3-level $ \Lambda $-type (or $V$-type and cascade type) atoms resonantly coupled to two classical fields [@Marangos98JMO; @Boon99PRA], which are called as the control and probe light field respectively. Now, many advanced studies have been done in the field of EIT. On the one hand, people find the EIT phenomenon can appears not only in the case of exactly one-photon resonance but also in the case of two-photon resonance [@Deng01; @Deng02; @li-sun-prar]. And people has done many studies on the EIT system involving 4-level (or multi-level) atoms [@Un; @duan-science; @Gu03PRA]. On the other hand, the quantum probe light is introduced to replace the weak classical probe light field in the EIT system [@Lukin00-ent; @Fl00-pol]. Recently, for example, an ensemble of $\Lambda $-type atoms, where the weak classical probe light is replace by a quantum probe light to form an EIT system, has been proposed [@Lukin00-ent; @Fl00-pol; @Fl00-OptCom; @Lukin-RMP] as a candidate for practical quantum memory to store and transfer the quantum information contained in photonic states by the collective atomic excitations. Some experiments [@liu; @group] have also already demonstrated the central principle of this technique, namely, the reduction of the group velocity and resonant absorbtion of light. Moreover, based on these previous work, we have studied a system with quasi-spin wave collective excitations of many $\Lambda $-type atoms fixed in “atomic crystal”. A hidden dynamical symmetry is discovered in such a system, and it is considered as a candidate for a robust quantum memory [Sun-prl]{}. It is observed that in certain cases [@Sun-quant-ph] the quantum state can be retrieved up to a non-Abelian Berry phase, i.e., a non-Abelian holonomy [@BPF; @ZW; @Zana; @duan-science; @Ekert], in such a $\Lambda $-type atomic system or a similar “3+1”-level atomic ensemble system [@li-zhang]. This observation extends the concept of quantum information storage and means that the stored state can be decoded in a purely geometric way in such a case. The above work about “3+1”-level atomic ensemble only considers the transfer (or quantum storage) of photonic state within the atomic ensemble. In order to achieve a complete process of photonic quantum state storage, generally the probe light should has a slow group velocity in order to make sure it being within the atomic ensemble during the time of state transfer. In this work, we shall calculate the dispersion properties of the quantum probe light field in a “3+1”-level atomic system given in Ref. [li-zhang]{} by means of the novel algebraic dynamics method of atomic collective excitation shown in Refs. [@Sun-prl; @li-sun-prar; @jin-prb]. By studying the susceptibility and group velocity of the quantum light, we will show in what cases this system appears as an EIT one and investigate how the group velocity depends on the detuning of the control and probe fields. The Model ========= The model we considered consists of $N$ identical “3+1”-level atoms [li-zhang,Un,duan-science]{}, where each atom is coupled to two single-mode classical control fields and a quantum probe field as shown in Fig. [fig1]{}. The atomic levels are labelled as the ground state $|b\rangle $, the excited state $|a\rangle $, and the meta-stable states $|k\rangle $ $(k=1,2)$. The atomic transition $|a\rangle \rightarrow |b\rangle $, with energy level difference $\omega _{ab}$=$\omega _{a}-\omega _{b}$, is coupled to the probe field of frequency $\omega $ $(=\omega _{ab}-\Delta _{p})$ with the coupling coefficient $g$; and the atomic transition $|a\rangle \rightarrow |k\rangle $ $(k=1,2)$, with energy level difference $\omega _{ak}$, is driven by the classical control field of frequency $\nu _{k}$ $(=\omega _{ak}-\Delta _{k})$ with Rabi-frequency $\Omega _{k}(t)$. ![“3+1”-type four-level atoms interacting with a quantum probe field (with coupling constant $g$, frequency $\protect\omega $, and the detuning $\Delta _{p}$) and two classic control fields (with frequency $\protect\nu _{k}$ , coupling Rabi frequency $\Omega _{k}$, and the detuning $\Delta _{k}=\protect\omega _{ak}-\protect\nu _{k}$, $k=1,2$).[]{data-label="fig1"}](fig1.eps){width="5cm" height="5.4cm"} Under the rotating wave approximation we can write the interaction Hamiltonian in the interaction picture as (let $\hbar =1$) [@li-zhang] $$\begin{aligned} H_{I} &=&\Delta _{p}S+g\sqrt{N}aA^{\dagger }+\Omega _{1}\exp [i(\Delta _{1}-\Delta _{p})t]T_{+}^{(1)} \notag \\ &&+\Omega _{2}\exp [i(\Delta _{2}-\Delta _{p})t]T_{+}^{(2)}+h.c., \label{hi01}\end{aligned}$$where $$\begin{aligned} S &=&\sum_{j=1}^{N}\sigma _{aa}^{(j)},\text{ }A=\frac{1}{\sqrt{N}}\sum_{j=1}^{N}\sigma _{ba}^{(j)}, \notag \\ T_{-}^{(k)} &=&\sum_{j=1}^{N}\sigma _{ka}^{(j)},\text{ }T_{+}^{(k)}=(T_{-}^{(k)})^{\dagger },\text{ }(k=1,2) \label{coll-ops}\end{aligned}$$are symmetrized collective atomic operators. Here $\sigma _{\mu \nu }^{(j)}=|\mu \rangle _{jj}\langle \nu |$ denotes the flip operator of the $j$-th atom from state $|\nu \rangle _{j}$ to $|\mu \rangle _{j}$ $(\mu ,\nu =a,b,1,2)$; $a^{\dagger }$ and $a$ the creation and annihilation operators of quantum probe field respectively. The coupling coefficients $g$ and $\Omega _{1,2}$ are real and assumed to be identical for all the atoms in the ensemble. Let us recall the dynamical symmetry as discovered in Ref. [@li-zhang] in the large $N$ limit and low excitation regime of the atomic ensemble where most of $N$ atoms stay in the ground state $|b\rangle $. It is obvious that $T_{-}^{(k)}$ and $T_{+}^{(k)}$ $(k=1,2)$ generate two mutually commuting $SU(2)$ subalgebras of $SU(3)$ [@jin-prb]. To form a closed algebra containing $SU(3)$ and $\{A,A^{\dagger }\}$ appeared in Hamiltonian (\[hi01\]), two additional collective operators $$C_{k}=\frac{1}{\sqrt{N}}\sum_{j=1}^{N}\sigma _{bk}^{(j)},\ \ (k=1,2) \label{C_k}$$along with their hermitian conjugates are introduced. These operators have the non-vanishing commutation relations $$\begin{aligned} &&C_{k}=[A,T_{+}^{(k)}],\ \ [C_{k},T_{-}^{(k)}]=A,\ \ (k=1,2); \notag \\ &&[A,A^{\dagger }]=[C_{1},C_{1}^{\dagger }]=[C_{2},C_{2}^{\dagger }]=1. \label{commuta22}\end{aligned}$$As a special case of quasi-spin wave excitation with zero varying phases, the above three mode symmetrized excitations defined by $A$ and $C_{1,2}$ behave as three independent bosons. The susceptibility of quantum probe light field =============================================== Now we will investigate the probe field group velocity from the time-dependent Hamiltonian (\[hi01\]). By means of the above dynamic algebra and commutation relation (\[commuta22\]), we can write down the Heisenberg equations of operators $A$ and $C_{1,2}$ as$$\begin{aligned} \dot{A} &=&-(\Gamma _{A}+i\Delta _{p})A-ig\sqrt{N}a \notag \\ &&-ie^{i(\Delta _{c}-\Delta _{p})t}(\Omega _{1}C_{1}+\Omega _{2}C_{2})+f_{A}(t), \\ \dot{C}_{1} &=&-\Gamma _{1}C_{1}-ie^{i(\Delta _{p}-\Delta _{c})t}\Omega _{1}A+f_{1}(t), \\ \dot{C}_{2} &=&-\Gamma _{2}C_{2}-ie^{i(\Delta _{p}-\Delta _{c})t}\Omega _{2}A+f_{2}(t).\end{aligned}$$Here, we have phenomenologically introduced the decay rates $\Gamma _{{1,2}}$ and $\Gamma _{A}$ of the states $|1\rangle $, $|2\rangle $ and $|a\rangle $, and $f_{A}(t)$ and $f_{1,2}(t)$ are the relative quantum fluctuation of operators with $\left\langle f_{\alpha }(t)f_{\alpha }(t^{\prime })\right\rangle \neq 0$, but $\left\langle f_{\alpha }(t)\right\rangle =0$, $(\alpha =A,1,2)$. To find the steady state solution for the above motion equations of atomic coherent excitation, it is convenient to remove the fast time-changing factors by making the transformation $C_{j}=\tilde{C}_{j}\exp [i(\Delta _{p}-\Delta _{j})t]$ ($j=1,2$). So the transformed equations are given as $$\begin{aligned} \dot{A} &=&-(\Gamma _{A}+i\Delta _{p})A-ig\sqrt{N}a \notag \\ &&-i(\Omega _{1}\tilde{C}_{1}+\Omega _{2}\tilde{C}_{2})+{f}_{A}(t), \\ \dot{\tilde{C}}_{1} &=&-\Gamma _{1}\tilde{C}_{1}-i(\Delta _{p}-\Delta _{1})\tilde{C}_{1}-i\Omega _{1}A+\tilde{f}_{1}(t), \\ \dot{\tilde{C}}_{2} &=&-\Gamma _{2}\tilde{C}_{2}-i(\Delta _{p}-\Delta _{2})\tilde{C}_{2}-i\Omega _{2}A+\tilde{f}_{2}(t).\end{aligned}$$As shown in Ref. [@li-sun-prar], in the steady state approach and taking the mean expressions of the above equations, we can obtain $$ig\sqrt{N}\left\langle a\right\rangle =-F(\Delta _{p})\left\langle A\right\rangle , \label{mean a A}$$where$$\begin{aligned} F(\Delta _{p}) &=&(\Gamma _{A}+i\Delta _{p})+\frac{\Omega _{1}^{2}}{\Gamma _{1}+i(\Delta _{p}-\Delta _{1})} \notag \\ &&+\frac{\Omega _{2}^{2}}{\Gamma _{2}+i(\Delta _{p}-\Delta _{2})}.\end{aligned}$$It is noticed that the single-mode probe quantum light is described by $$E(t)=\varepsilon e^{-i\omega t}+h.c.\equiv \sqrt{\frac{\omega }{2V\epsilon _{0}}}ae^{-i\omega t}+h.c., \label{E a}$$where $V$ is the effective mode volume and for simplicity is assumed to be equal to the volume of the atomic ensemble. While its corresponding polarization is $$\left\langle P\right\rangle =\left\langle p\right\rangle e^{-i\omega t}+h.c.\equiv \epsilon _{0}\chi \left\langle \varepsilon \right\rangle e^{-i\omega t}+h.c., \label{p chi}$$where $\chi =\left\langle p\right\rangle /(\left\langle \varepsilon \right\rangle \epsilon _{0})$ is the susceptibility. In terms of the average of the exciton operators $A$, the average polarization can be expressed as$$\left\langle p\right\rangle =\mu \left\langle \sum_{j=1}^{N}\sigma _{ba}^{(j)}\right\rangle /V=\frac{\mu \sqrt{N}}{V}\left\langle A\right\rangle , \label{mean p A}$$where $\mu $ is the dipole moment between state $|a\rangle $ and $|b\rangle $. It is also noted that the coupling coefficient $g$ can be written as $$g=-\mu \sqrt{\frac{\omega }{2V\epsilon _{0}}}. \label{g}$$According to the Eqs. (\[mean a A\])-(\[g\]), the susceptibility can be obtained as $$\chi =\frac{2ig^{2}N}{\omega F(\Delta _{p})}. \label{chi-f}$$The real and imaginary parts $\chi _{1}$ and $\chi _{2}$ of this complex susceptibility $\chi =\chi _{1}+i\chi _{2}$ are related to dispersion and absorption of quantum probe light field, respectively. Here the analytical solution of $\chi _{1}$ and $\chi _{2}$ is a little complicated to express, and we only give the numerical solution. First, we consider the case that the two classical light fields have the same detunings: $\Delta _{1}=\Delta _{2}\equiv \Delta $. Fig. \[figexp02\] shows $\chi _{1}$ and $\chi _{2}$ in such case versus the probe light detuning $\Delta _{p}$ under $\Delta =0,\pm 2$ and different Rabi frequencies $\Omega _{1,2}$ with the other parameters being fixed as $\Gamma _{1}=\Gamma _{2}=10^{-4}$, $g\sqrt{N}=100$ (all in normalized units of $\Gamma _{A}$). Seen from Figs. \[figexp02\], when $\Delta _{p}\rightarrow \Delta $, both $\chi _{1}$ and $\chi _{2}$ are almost equal to zero. This fact means that the medium indeed becomes transparent when driven by the two classical control fields as long as the system is prepared in the 3-photon resonance ($\Delta _{p}=\Delta _{1}=\Delta _{2}$) without the condition of exact one-photon resonance: $\Delta _{i}\equiv 0$ ($i=p,1,2$). We also notice that the width of the induced transparency window (which is determined by $\chi _{2}$ in the near domain of $\Delta _{p}=\Delta $) also depends on the concurrent interaction of Rabi frequencies $\Omega _{1,2}$. This is intuitionistic and coincident with the previous work [li-sun-prar,Scullybook]{}: under the case of 3-photon resonance, since each classical control field as well as the relative level induces a transparency window for the quantum probe field and appears the phenomenon of EIT independently (under 2-photon resonance), such a system appears EIT phenomenon too and results from the concurrent influence of these two control fields. Then, let us consider the case of the two classical control light fields having the different detunings: $\Delta _{1}\neq \Delta _{2}$. Fig. [figexp03]{} shows the dependance of $\chi _{1}$ and $\chi _{2}$ on the probe light detuning $\Delta _{p}$ in this case. There are two transparent windows for the probe light field this time. These two windows appear near the points of $\Delta _{p}=\Delta _{1}$ and $\Delta _{p}=\Delta _{2}$, with the width of windows depending on $\Omega _{1}$ and $\Omega _{2}$ respectively. When $\Delta _{1}\rightarrow \Delta _{2}$ and the Rabi frequencies are strong enough, two transparency windows will overlap (see Fig. \[figexp03\](c)) or even become one (see Fig. \[figexp03\](d)). This results from the fact that: when the quantum probe field together with one of the classical control fields satisfies the 2-photon resonance condition, a transparency window appears for the quantum probe field and so does the phenomenon of EIT, since at the same time the effect of the other control field (which does not satisfy the 2-photon resonance together with the probe field) is small and can be ignored. In the next section, we will continue to calculate the group velocity under this case and the case of 3-photon resonance. ![Real part $\protect\chi_1$ (solid) and imaginary part $\protect\chi _2$ (dashed) of the susceptibility $\protect\chi$ vs the probe light detuning $\Delta_{p}$ in normalized units of $\Gamma _{A}$ according to: (a) $\Delta =0$, $\Omega_1=1$ and $\Omega_2=0$; (b) $\Delta =0$, $\Omega_1=0$ and $\Omega_2=1$; (c) $\Delta =0$, $\Omega_1=\Omega_2=1$; (d) $\Delta =0$, $\Omega_1=\Omega_2=2$; (e,f) $\Delta =\pm 2$, $\Omega_1=\Omega_2=1$. The Other parameters are given as: $\Gamma _{1}=\Gamma _{2}=10^{-4}$, $g\protect\sqrt{N}=100$.[]{data-label="figexp02"}](figexp02.eps){width="8cm" height="7cm"} ![Real part $\protect\chi_1$ (solid) and imaginary part $\protect\chi _2$ (dashed) of the susceptibility $\protect\chi$ vs the probe light detuning $\Delta_{p}$ according to: (a) $\Delta_1 =1$, $\Delta_2 =-1$, $\Omega_1=\Omega_2=1$; (b) $\Delta_1 =1$, $\Delta_2 =-2$, $\Omega_1=2$, $\Omega_2=1/2$; (c) $\Delta_1 =0.5$, $\Delta_2 =-0.5$, $\Omega_1=2$, $\Omega_2=2$; (d) $\Delta_1 =0.05$, $\Delta_2 =-0.05$, $\Omega_1=4$, $\Omega_2=4$. The Other parameters are given as: $\Gamma _{1}=\Gamma _{2}=10^{-4}$, $g\protect\sqrt{N}=100$.[]{data-label="figexp03"}](figexp03.eps){width="8cm" height="5.5cm"} The Group Velocity of quantum probe light field =============================================== Next we consider the property of refraction and absorption of the single-mode probe light within the atomic ensemble medium in more detail. To this aim we will analyze the complex refractive index $$n(\omega )=\sqrt{\epsilon (\omega )}=\sqrt{1+\chi }\equiv n_{1}+in_{2}. \label{n}$$Where the real part $n_{1}$ of $n$ represents the refractive index of the medium and the imaginary $n_{2}$ is the associated absorption coefficient. Together with the formulae for the group velocity of the probe light $$v_{g}(\Delta _{p})=\frac{c}{\mathrm{Re}[n+\omega \frac{\mathrm{d}n}{\mathrm{d}\omega }]}=\frac{c}{n_{1}+\omega \frac{\mathrm{d}n_{1}}{\mathrm{d}\omega }} \label{vg1}$$(where $c$ is the light velocity in vacuum) depending on the frequency dispersion, one can obtain the explicit expression for the group velocity $v_{g}$ from Eqs. (\[chi-f\]-\[vg1\]) for arbitrary reasonable values of $\Delta _{p}$ and other parameters. Now, we consider the group velocity of the probe light $v_{g}$ under the case of EIT. At this time the values of $\chi _{1}$ and $\chi _{2}$ are almost zero, and we obtain approximately $$n_{1}\simeq 1+\chi _{1}/2\rightarrow 1,\ \ n_{2}\simeq \chi _{2}\rightarrow 0,$$and the group velocity of probe light is given briefly as: $$v_{g}(\Delta _{p})=\frac{c}{1-\frac{\omega }{2}\frac{\mathrm{d}\chi _{1}}{\mathrm{d}\Delta _{p}}}. \label{vg2}$$It is worth stressing that the above Eq. (\[vg2\]) is effective only under the case of EIT. ![The probe light group velocity $v_{g}$ under the case of 3-photon resonance vs: (a) the Rabi frequency $\Omega_{1}$ \[in normalized units\] for $\Delta =0$ and $\Omega_{2}$ being assumed to vary synchronously with $\Omega_{1}$ (solid line) or $\Omega_{2}$ being given as $100$ (dashed line); (b) the detuning $\Delta $ ($\equiv \Delta_p $) for $\Omega_1=\Omega_2=0.04$ (solid line), or $\Omega_1=\Omega_2=50$ (dashed line). The other parameters are given as: $\Gamma _{1}=\Gamma _{2}=10^{-4}$ and $g\protect\sqrt{N}=100$.[]{data-label="figexp04"}](fig04.eps){width="8cm" height="3.0cm"} ![The probe light group velocity $v_{g}$ (under EIT of non-3-photon resonance: $\Delta _1\neq \Delta _2)$ vs the Rabi frequency $\Omega_{1}$ ($\equiv\Omega_{2}$) for $\Delta _p=\Delta _1=2$, $\Delta _2=-2$ and the other parameters given as $\Gamma _{1}=\Gamma _{2}=10^{-4}$ and $g\protect\sqrt{N}=100$.[]{data-label="figexp05"}](fig05.eps){width="5cm" height="3cm"} According to Eq. (\[vg2\]), the group velocity $v_{g}$ of the probe light within the 3-photon-resonance atomic ensemble is shown in Fig. \[figexp04\]. In Fig. \[figexp04\](a), the solid line shows the dependance of $v_{g}$ on Rabi frequencies $\Omega _{1,2}$ (assume $\Omega _{1,2}$ vary at the same time), the dashed one shows the dependance of $v_{g}$ on Rabi frequencies $\Omega _{1}$ ($\Omega _{2}$ is fixed as $100$), with the other parameters being given as $\Delta _{p}\equiv \Delta =0$, $\Gamma _{1}=\Gamma _{2}=10^{-4}$, and $g\sqrt{N}=100$. Fig. \[figexp04\](b) shows the dependance of $v_{g}$ on detuning $\Delta $ when the other parameters are given as $\Delta _{p}\equiv \Delta $, $\Gamma _{1}=\Gamma _{2}=10^{-4}$, $g\sqrt{N}=100$, $\Omega _{1}=\Omega _{2}=50$ (dashed line) or $\Omega _{1}=\Omega _{2}=0.04$ (solid line). This implies that: when $\Omega _{1}$ *or* $\Omega _{2}$ is big (compared with $g\sqrt{N}$), $v_{g}$ is relatively fast and insensitive to the common detuning $\Delta _{p}$ ($\equiv \Delta $) (see the dashed line in Fig. \[figexp04\]); however when both $\Omega _{1}$ *and* $\Omega _{2}$ are small, $v_{g}$ is relatively slow and sensitive to the common detuning $\Delta _{p}$ ($\equiv \Delta $) (see the solid line in Fig. \[figexp04\]). We have also calculated $v_{g}$ under the 2-photon resonance (but not 3-photon resonance) EIT.  Fig. \[figexp05\] shows the dependance of $v_{g} $ on Rabi frequencies ($\Omega _{2}\equiv \Omega _{1}$ at this time), when $\Delta _{p}\equiv \Delta _{1}=2$ $\neq \Delta _{1}=-2$ with the other parameters being given as $\Delta _{p}\equiv \Delta =0$, $\Gamma _{1}=\Gamma _{2}=10^{-4}$, and $g\sqrt{N}=100$. The result shows under this case, the probe field group velocity $v_{g}$ also can been varied in the scope of ($0,c $) with $\Omega _{1,2}$ being varied. This fact, the probe field group velocity $v_{g}$ decreases dramatically with small $\Omega _{1,2}$, ensures that the technique as shown in Ref. [li-zhang]{} is effective to accomplish the storage and retrieve of the probe pulse. The storage process of such a technique is that: initially when the probe field enters into the 3-photon-resonance atomic medium, the Rabi frequency $\Omega $ is very large (relative to $g\sqrt{N}$) and $v_{g}\rightarrow c$; when $\Omega _{1,2}$ are reduced adiabatically to zero, $v_{g}$ reduces to zero accordingly and then one can store the pulse in the medium. Conversely, the retrieve process is the inverse one. That is, if one wants to retrieve the probe pulse, he only needs to increase $\Omega $ adiabatically so as to increase $v_{g}$. Conclusion ========== In this work, based on the dynamical algebra method of atomic collective excitation [@Sun-prl; @li-sun-prar], we have studied theoretically the susceptibility and group velocity of a quantum probe light in a “3+1”-level atomic ensemble under EIT. Our results show the quantum probe light group velocity can been reduced dramatically under tiny values of $\Omega _{1,2}$. This is very useful during the storage of the quantum probe light in such a “3+1”-level atomic ensemble. Moreover, our results show that two transparency windows for the probe light can occur in the case of two classical control light fields having the different detuings to the relative atomic transitions. Of course, in the practical experiment for store a quantum light, the influence of atomic spatial motion or atomic collisions, and the effects of buffer gases, should be taken into account. In the present work, all of these effects are ignored as the perturbations for we assume that the atomic ensemble is prepared under enough low temperature. *We acknowledge the support of the CNSF (grant No. 90203018), the Knowledge Innovation Program (KIP) of the Chinese Academy of Sciences and the National Fundamental Research Program of China (No. 2001CB309310). Y. L. also thanks the support of National Natural Science Foundation of China with No. 10447133.* [99]{} M. O. Scully, Phys. Rev. Lett. [**67**]{}, 1855 (1991). S. E. Harris [*et al.*]{}, Phys. Rev. Lett. [**64**]{}, 1107 (1990). M. Fleischhauer, M. D. Lukin, A. B. Matsko, and M. O. Scully, Phys. Rev. Lett. [**82**]{}, 1847 (1999). S. E. Harris, Physics Today **50**, 36 (1997). J. P. Marangos, Journal of Modern Optics [**45**]{}(3), 471 (1998). J. R. Boon, E. Zekou, D. McGloin, and M. H. Dunn, Phys. Rev. A [**59**]{}, 4675 (1999). L. Deng, E. W. Hagley, M. Kozuma, and M. G. Payne, Phys. Rev. A [**65**]{}, 051805(R) (2002). M. Kozuma [*et al.*]{}, Phys. Rev. A [**66**]{}, 031801(R) (2002). Y. Li and C. P. Sun, Phys. Rev. A [**69**]{}, 051802(R) (2004). R. G. Unanyan, B. W. Shore, and K. Bergmann, Phys. Rev. A **59**, 2910 (1999). L. M. Duan, J. I. Cirac, and P. Zoller, Science **292**, 1695 (2001). Y. Gu, Q. Sun, and Q. Gong, Phys. Rev. A [**67**]{}, 063809 (2003). M. D. Lukin, S. F. Yelin, and M. Fleischhauer, Phys. Rev. Lett. **84**, 4232 (2000). M. Fleischhauer and M. D. Lukin, Phys. Rev. Lett. **84**, 5094 (2000). D. F. Phillips, A. Fleischhauer, A. Mair, and R. L. Walsworth, M. D. Lukin, Phys. Rev. Lett. **86**, 783 (2001). M. D. Lukin, Rev. Mod. Phys. **75**, 457 (2003). C. Liu, Z. Dutton, C. H. Behroozi, and L. V. Hau, Nature **409**, 490 (2001). L. V. Hau, S. E. Harris, Z. Dutton, and C. H. Behroozi, Nature **397**, 594 (1999). C. P. Sun, Y. Li, and X. F. Liu, Phys. Rev. Lett. **91**, 147903 (2003). C. P. Sun, P. Zhang, and Y. Li, quant-ph/0311052. M. V. Berry, Proc. R. Soc. London A **392**, 45 (1984); A. Shapere and F. Wilczek (Ed.), *Geometric Phases in Physics* (World Scientific, Singapore, 1989). F. Wilczek and A. Zee, Phys. Rev. Lett. **52**, 2111 (1984). P. Zanardi and M. Rasetti, Phys. Lett. A **264**, 94 (1999); J. Pachos, P. Zanardi, M. Rasetti, Phys. Rev. A **61**, 010305(R) (2000). J. A. Jones, V. Vedral, A. Ekert, G. Castagnoli, Nature **403**, 869 (2000). Y. Li, P. Zhang, P. Zanardi, and C. P. Sun, Phys. Rev. A 70, 032330 (2004). G. R. Jin, P. Zhang, Yu-xi Liu, and C. P. Sun, Phys. Rev. B **68**, 134301 (2003). M. O. Scully and M. S. Zubairy, [*Quantum Optics*]{}, (Cambridge University Press, 1997).

ArXiv

--- abstract: 'We compute the Hochschild cohomology of any block of $q$-Schur algebras. We focus the even part of this Hochschild cohomology ring. To compute the Hochschild cohomology of $q$-Schur algebras, we prove the following two results: first, we construct two graded algebra surjections between the Hochschild cohomologies of quasi-hereditary algebras because all $q$-Schur algebras over a field are quasi-hereditary. Second, we give the graded algebra isomorphism of Hochschild cohomologies by using a certain derived equivalence.' address: 'Department of Mathematics, Graduate School of Science, Osaka City University, 3-3-138 Sugimoto, Sumiyoshi-ku, Osaka 558-8585 JAPAN' author: - Mayu Tsukamoto bibliography: - 'Bib\_File1.bib' title: 'Hochschild cohomology of $q$-Schur algebras ' --- Introduction ============ [*$q$-Schur algebras*]{} were introduced by Dipper and James [[@DJ]]{} in order to study the modular representation theory of finite general linear groups. There exists a surjection from the quantum general linear group onto the $q$-Schur algebra (for example, see [[@PW Theorem 11.3.1]]{}). It is known that the $q$-Schur algebras over a field are quasi-hereditary (cf. [[@PW Theorem 11.5.2]]{}). Let $\kk$ be a splitting field for the algebras we consider, and $q \in \kk \setminus \{0\}$. Let $\mathscr{H}_q (d)$ be the Hecke algebra of the symmetric group $\mathfrak{S}_d$ with parameter $q$ over $\kk$. $\mathscr{H}_q (d)$ is the associative algebra with generators $T_1, \ldots , T_{d-1}$ and satisfies quadratic relations $(T_i + 1)(T_i -q)=0$. Let $\mathscr{S}_q (n, d) := \End_{\mathscr{H}_q (d)}(\displaystyle\bigoplus_{\lambda \in \Lambda(n, d)} x_{\lambda}\mathscr{H}_q (d))$ be the $q$-Schur algebra associated with $\mathscr{H}_q (d)$ and $\Lambda(n, d)$ (see [[@Mt Chapter 4]]{} for the details on the above definitions), where $\Lambda(n, d)$ is the set of all sequences of non-negative integers $(\lambda_1, \cdots, \lambda_n)$ such that $\displaystyle\sum_{i=1}^n \lambda_i = d$ and $x_{\lambda} = \displaystyle\sum_{\pi \in \mathfrak{S}_{\lambda}}T_{\pi}$ where $\mathfrak{S}_{\lambda}$ is the Young subgroup corresponding to $\lambda$. $\Lambda(n, d)$ is a poset with the dominance ordering $\preceq$. Let $n<d$ and $\xi_{\lambda} :\displaystyle\bigoplus_{\lambda \in \Lambda(d, d)} x_{\lambda}\mathscr{H}_q (d) \twoheadrightarrow x_{\lambda} \mathscr{H}_q (d) \hookrightarrow \displaystyle\bigoplus_{\lambda \in \Lambda(d, d)} x_{\lambda}\mathscr{H}_q (d)$. Then $\xi_{\lambda} \in \mathscr{S}_q(d, d)$ is an idempotent. We set $\xi:= \displaystyle\sum_{\lambda \in \Lambda(n, d)} \xi_{\lambda}$. Then we obtain $\mathscr{S}_q (n, d) \cong \xi \mathscr{S}_q (d, d) \xi$, and we call the above idempotent $\xi$ the Green’s idempotent. In the case $q=1$, $\mathscr{S}_1 (n, d)$ is the Schur algebra. The theory of cohomology of associative algebras was introduced by Hochschild [[@HH]]{}. [*Hochschild cohomology*]{} of associative algebras is important in many areas of mathematics, such as ring theory, geometry, representation theory and so on. For example, it was observed that the second Hochschild cohomology group of an associative algebra $A$ controls the deformation theory of $A$ [[@G2]]{}. The Hochschild cohomology is a graded algebra with the Yoneda product. One of the most important properties of Hochschild cohomology is its invariance under derived equivalences, proved by Rickard in . In general, it is difficult to compute the Hochschild cohomology. For several kinds of algebras, the Hochschild cohomologies are calculated. For example, Benson and Erdmann [[@BE]]{} give the Hochschild cohomology of the Hecke algebra of the symmetric group at a root of unity in characteristic zero. In this paper, we compute the Hochschild cohomology of $q$-Schur algebras, following the method of [[@BE]]{}, which we give in §3.1. Thus their assumption might change to our assumption, see [[@MR2761936]]{} for more detail. The structure of this paper is as follows: First, we construct the following two graded algebra surjections between the Hochschild cohomology of quasi-hereditary algebras: 1. If $S$ is quasi-hereditary and $H$ is a heredity ideal in $S$, then there exists a graded algebra surjection from ${\rm HH}^{\ast}(S)$ onto ${\rm HH}^{\ast}(S/H)$; 2. Let $S$ be a quasi-hereditary algebra. We fix a complete set $\{ L(\lambda) \; | \; \lambda \in \Lambda \}$ of simple $S$-modules and a set of orthogonal idempotent $\{e_{\lambda} \; | \; \lambda \in \Lambda \}$ in $S$. For $\pi \subseteq \Lambda$, we put $\epsilon:= \displaystyle\sum_{\lambda \in \pi} e_{\lambda}$. If $\Lambda \setminus \pi$ is a poset ideal, then $\epsilon S \epsilon$ is a quasi-hereditary algebra and there exists a graded algebra surjection from ${\rm HH}^{\ast}(S)$ onto ${\rm HH}^{\ast}(\epsilon S \epsilon)$. Second, we compute the even part of the Hochschild cohomology of $q$-Schur algebras by using the above surjections. In particular, the Green’s idempotents of the $q$-Schur algebras $\mathscr{S}_q(n, d)$ satisfies the assumption (ii). Therefore we have the following graded algebra surjection: $${\rm HH}^{\ast}(\mathscr{S}_q (d, d)) \twoheadrightarrow {\rm HH}^{\ast} (\mathscr{S}_q (n, d)).$$ Third, we construct an explicit bimodule resolution of a certain block $A_e$ of $q$-Schur algebras, and determine the dimensions of Hochschild cohomology groups. $$\begin{aligned} {\rm dim} \;{\rm HH}^i (A_e) = \left\{ \begin{array}{lll} e, \; & ( i = 0);\\ 1, & ( 1 \leq i \leq 2(e-1)); \\ 0, & ( 2(e-1) < i). \end{array} \right.\end{aligned}$$ Finally, we describe the $\kk$-algebra structure of the even part of the Hochschild cohomology ring of $q$-Schur algebras. Hochschild cohomology of quasi-hereditary algebras ================================================== In this section, we construct the following two graded algebra surjections between the Hochschild cohomology of quasi-hereditary algebras. Hochschild cohomology --------------------- Let $R$ be an associative algebra over a commutative ring $K$. First, we recall the definition of the Hochschild cohomology of $R$ $$\begin{aligned} {\rm HH}^i(R):={\rm Ext}^i_{R^{\en}}(R, R),\end{aligned}$$ where $R^{\en}$:$=R \otimes_{K} R^{\op}$ (for example, see [@CE Chapeter XI §4]). This may be expressed in terms of the standard resolution: $$\begin{aligned} \cdots \xrightarrow{d_{3}} R^{\otimes 4} \xrightarrow{d_{2}} R^{\otimes 3} \xrightarrow{d_{1}} R^{\otimes 2} \xrightarrow{d_{0}} R \rightarrow 0.\end{aligned}$$ This is an $R^{\en}$-free resolution of $R$, where $d_{0}$ is the multiplication map and $$\begin{aligned} d_{i}(r_{0}\otimes r_{1}\otimes \cdots \otimes r_{i+1})=\sum_{n=0}^{i}(-1)^n r_{0} \otimes \cdots \otimes r_{n} r_{n+1} \otimes \cdots \otimes r_{i+1}.\end{aligned}$$ Applying ${\rm Hom}_{R^{\en}}(-,R)$ to (2-1), we have the following complex: $$\begin{aligned} 0 \rightarrow {\rm Hom}_{R^{\en}}(R^{\otimes 2}, R) &\xrightarrow{{\rm Hom}(d_{1}, R)}& {\rm Hom}_{R^{\en}}(R^{\otimes 3}, R) \\ \xrightarrow{{\rm Hom}(d_{2}, R)} {\rm Hom}_{R^{\en}}(R^{\otimes 4}, R) &\xrightarrow{{\rm Hom}(d_{3}, R)}& \cdots .\end{aligned}$$ Thus we have ${\rm HH}^i (R)= {\rm HH}^i(R, R)={\rm Ker Hom}(d_{i+1}, R)/ {\rm Im Hom}(d_{i}, R)$. We denote by ${\rm HH}^\ast(R)$:$=\displaystyle\bigoplus_{i \geq 0} {\rm HH}^i(R)$ the Hochschild cohomology ring of $R$, where the multiplication is given by the Yoneda product (cf. [@B]). We denote by $\star$ the Yoneda product in ${\rm HH}^\ast(R)$. Let $\alpha \in {\rm HH}^i(R)$ and $\beta \in {\rm HH}^j(R)$ be the elements which are represented by $\alpha \in {\rm Ker Hom}(d_{i+1}, R)$ and $\beta \in {\rm Ker Hom}(d_{j+1}, R)$, respectively. Then $\alpha \star \beta \in {\rm HH}^{i+j}(R)$ is given as follows. There exists the following commutative diagram of $R^{\en}$-modules $$\xymatrix{ \cdots \ar[r] &P_{i+j} \ar[d]_{\sigma_{i}} \ar[r]^{d_{i+j}} &\cdots \ar[r]^{d_{j+2}} &P_{j+1} \ar[d]_{\sigma_{1}} \ar[r]^{d_{j+1}} &P_{j} \ar[d]_{\sigma_{0}} \ar[dr]^{\beta} &\\ \cdots \ar[r] &P_{i} \ar[r]^{d_{i}} &\cdots \ar[r]^{d_{2}} &P_{1} \ar[r]^{d_{1}} &P_{0} \ar[r]^{d_{0}} &R \ar[r] &0, }$$ where $\sigma_{t} \; (0\leq t \leq i)$ are liftings of $\alpha$. Then we have $\alpha \star \beta= \alpha \circ \sigma_{i} \in {\rm HH}^{i+j}(R)$. It is known that $\alpha \star \beta$ is independent of choices of representatives $\alpha$, $\beta$ and liftings $\sigma_{t} \; (0 \leq t \leq i)$. Moreover, with this product, Gerstenhaber proved that ${\rm HH}^\ast(R)$ is a super commutative algebra in [[@G]]{}. That is, for homogeneous elements $\eta \in {\rm HH}^n(R)$ and $\theta \in {\rm HH}^m(R)$, we have $\eta \star \theta = (-1)^{nm} \theta \star \eta$. In particular, ${\rm HH}^{\rm ev}(R)$ is a commutative algebra, where ${\rm HH}^{\rm ev}(R):= \displaystyle\bigoplus_{i \geq 0} {\rm HH}^{2i}(R)$. ${\rm HH}^{\rm ev}(R)$ is called the even part of ${\rm HH}^{\ast}(R)$. Quasi-hereditary algebras ------------------------- We recall the definition of the quasi-hereditary algebra. This notion was first introduced by Scott [[@S]]{} to study highest weight categories in the representation theory of semisimple complex Lie algebras and algebraic groups. Cline, Parshall and Scott proved many important results in [[@CPS]]{}, see also [[@PS]]{}. In [[@DR]]{}, for a semiprimary ring, Dlab and Ringel gave another definition of quasi-hereditary by using an ideal chain. Let $S$ be a finite dimensional algebra over an algebraically closed field $\kk$. Let $J(S)$ be the Jacobson radical of $S$. We denote by $S\mod$ the category of finitely generated left $S$-modules. We denote by $S\proj$ the category of finitely generated projective left $S$-modules. Let $H$ be a two-sided ideal of $S$. If $H$ satisfies the following conditions, we call $H$ a [*heredity ideal*]{} in $S$: 1. $HH=H$; 2. ${\rm Hom}_S(H,S/H) =0$; 3. $HJ(S)H=0$. $S$ is called a [*quasi-hereditary algebra*]{} if there exists a chain of ideals $$S=H_{0}>H_{1}> \cdots >H_{n}=0$$ with $H_{i}/H_{i+1}$ heredity ideals in $S/H_{i+1}$, for $0 \leq i <n$. Such a chain of ideals is called a [*heredity chain*]{} of $S$. We fix a complete set of pairwise non-isomorphic simple $S$-module $\{ L(\lambda) \; | \; \lambda \in \Lambda \}$ and we fix a partial ordering $\leq$ on the index set $\Lambda$. For $\lambda \in \Lambda$, we write $P(\lambda)$ (resp. $I(\lambda)$) for the projective cover (resp. injective hull) of $L(\lambda)$. For $\lambda \in \Lambda$, there is a unique maximal submodule $K(\lambda)$ of $P(\lambda)$ which satisfies the following condition: If $[J(S) P(\lambda)/K(\lambda):L(\mu)] \not = 0$, then we have $\lambda > \mu$. We write $\Delta(\lambda) := P(\lambda)/ K(\lambda)$, and we call $\Delta(\lambda)$ the [*standard module*]{} corresponding to $\lambda \in \Lambda$. Similarly, for $\lambda \in \Lambda$, we define $\nabla(\lambda)$ as the maximal submodule of $I(\lambda)$ which satisfies the following condition: If $[\nabla(\lambda)/L(\lambda): L(\mu)] \not = 0$, then we have $\lambda < \mu.$ We call $\nabla(\lambda)$ the [*costandard module*]{} corresponding to $\lambda \in \Lambda$. Let $M \in S\mod$. If $M$ has a filtration $M=M_0 > M_1 > \cdots > M_i > M_{i+1} > \cdots > M_n=0$ such that $M_i/M_{i+1} \cong \Delta(\lambda)$ for some $\lambda \in \Lambda$ (resp. $\nabla(\lambda)$), for all $0 \leq i \leq n-1$, then $M$ is called [*$\Delta$-filtered module*]{} (resp. [*$\nabla$-filtered module*]{}) and $\Delta(\lambda)$ which is isomorphic to $M_i/M_{i+1}$ for some $i$ is called a filtration factor of $M$. \[cf. Donkin[[@D A.1 (7)]]{}\] For a $\Delta$-filtered module $M$, the element $[M]$ in the Grothendieck group $K_0(S)$ of $S\mod$ corresponding to $M$ can be written as $$[M]= \sum_{\lambda \in \Lambda} m_{\lambda} [\Delta(\lambda)] = \sum_{\mu \in \Lambda} (\sum_{\lambda \in \Lambda} m_{\lambda} [\Delta(\lambda): L(\mu)])[L(\mu)].$$ If $[\Delta(\lambda):L(\mu)] \not= 0$, then we have $\mu \leq \lambda$. Thus the coefficients $m_{\lambda}$ are uniquely determined. In other words, the filtration multiplicities do not depend on the choice of the $\Delta$-filtration of $M$. Similarly, we deduce that $\nabla$-filtration multiplicities do not depend on the choice of filtration. Moreover the length of $\Delta$-filtration (resp. $\nabla$-filtration) does not depend on the choice of $\Delta$-filtration (resp. $\nabla$-filtration). Thus we denote by $fl(M)$ the length of $\Delta$-filtration of $M$ and denote by $(M : \Delta (\lambda))$ (resp. $(M : \nabla (\lambda))$) the filtration multiplicity of $\Delta (\lambda)$ (resp. $\nabla(\lambda)$). \[cf. Donkin\] \[dnkn\] Let $X, Y \in S\mod$. We assume that $X$ is a $\Delta$-filtered module and $Y$ is a $\nabla$-filtered module. Then we have $$\begin{aligned} \Ext_S^i(X, Y)= \left\{ \begin{array}{ll} \mathop{\sum}\limits_{\nu \in \Lambda} (X : \Delta(\nu))(Y : \nabla(\nu)), &i=0;\\ 0, &i \neq 0. \end{array} \right.\end{aligned}$$ From the rest of this section, we assume that $S$ is a quasi-hereditary algebra and we fix an ideal $H$ of $S$ which appears in a heredity chain of $S$. We denote by $\overline{S}$ the quotient algebra of $S$ by $H$. We define $\mathbf{F}:=\overline{S} \otimes_{S} -$ and $\mathbf{F}^{\rm en}:=\overline{S}^{\rm en} \otimes_{S^{\rm en}} -$. Then we have the following lemma: \[sat\] Let $\lambda, \mu, \nu \in \Lambda$. If $\mathbf{F}(L(\lambda ))=L(\lambda )$ and $\mu < \lambda$, then we have $\mathbf{F}(L(\mu ))=L(\mu )$. If $\mathbf{F}(L(\lambda ))=0$ and $\lambda < \nu$, then we have $\mathbf{F}(L(\nu ))=0$. \[f\] Let $N \in S \mod$. We assume that $N$ is a $\Delta$-filtered module and $\mathbf{F}(\Delta(\lambda))=0$ for each filtration factor $\Delta(\lambda)$ of $N$. Then we have $\mathbf{F}(N)=0.$ We show the assertion by induction on $fl(N)$. In the case $fl(N)=1$, it is clear. We suppose $fl(N)>1$. Since $N$ is a $\Delta$-filtered module, there exists $\lambda \in \Lambda$ such that $$0 \rightarrow \Delta (\lambda ) \rightarrow N \rightarrow N/\Delta (\lambda ) \rightarrow 0$$ is a short exact sequence. By this short exact sequence, we have $$\cdots \rightarrow \mathbf{F}(\Delta (\lambda )) \rightarrow \mathbf{F}(N) \rightarrow \mathbf{F}(N/\Delta (\lambda )) \rightarrow 0.$$ Hence the proof is done by the induction hypothesis. \[F\] We fix $\lambda \in \Lambda$ such that $\mathbf{F}(L(\lambda ))=L(\lambda )$. Let $\iota: K(\lambda ) \hookrightarrow P(\lambda )$ be the inclusion map. Then $\mathbf{F}(\iota ): \mathbf{F}(K(\lambda )) \hookrightarrow \mathbf{F}(P(\lambda ))$ is an injection. We denote by $\iota |_{HK(\lambda )}$ the restriction of $\iota$ to $HK(\lambda)$. First we prove $H \Delta(\lambda )=0$ to show that $\iota |_{HK(\lambda )}$ is a surjection. It follows from the definition of standard modules that if $[\Delta (\lambda ):L(\mu )] \neq 0$, then we have $\mu \leq \lambda.$ Thus we deduce from Lemma \[sat\] that $\mathbf{F}(L(\mu))=L(\mu)$, for each composition factor of $\Delta(\lambda)$. Since we have $H\Delta(\lambda)=0$, it follows that $H P(\lambda )$ is isomorphic to $\iota(H K(\lambda))$. Therefore we have $\mathbf{F}(\iota)$ is an injection. We write $\mathbb{L}_i \mathbf{F}$ for the $i$-th left derived functor of $\mathbf{F}$. In the case $i=1$, we write $\mathbb{L} \mathbf{F}$. \[lem4\] Let $W$ be a left $\Delta$-filtered module. Then we have $\mathbb{L}_i \mathbf{F}(W)=0$ for any $i>0$. We show the statement by induction on $i$. Firstly, we show this statement in the case $i=1$ by induction on $fl(W)$. Now we consider the case $fl(W)=1$. Then there exists $\lambda \in \Lambda$ such that $\Delta(\lambda)$ is isomorphic to $W$. Thus we obtain $$0 \rightarrow K(\lambda) \rightarrow P(\lambda) \rightarrow \Delta(\lambda) \rightarrow 0.$$ By this short exact sequence (2-2), we have the following exact sequence: $$0 \rightarrow \mathbb{L} \mathbf{F}(\Delta(\lambda)) \rightarrow \mathbf{F}(K(\lambda)) \rightarrow \mathbf{F}(P(\lambda)) \rightarrow \mathbf{F}(\Delta(\lambda)) \rightarrow 0.$$ We assume that $\mathbf{F}(L(\lambda ))=0$. Then we deduce that if $(K(\lambda): \Delta(\mu)) \not=0$, then we have $\mathbf{F}(L(\lambda))=0$. Thus we obtain $\mathbf{F}(\Delta (\mu ))=0$ and we deduce from Lemma \[f\] that $\mathbf{F}(K(\lambda ))=0$. For our aim it is sufficient to show that $\mathbf{F}(L(\mu ))=0$ for $\mu \in \Lambda$, where $\mu$ satisfies $(K(\lambda ) : \Delta(\mu)) \neq 0$. Therefore we have $\mathbb{L}\mathbf{F}(\Delta(\lambda))=0$. If $\mathbf{F}(L(\lambda))=L(\lambda)$, then it follows from Lemma \[F\] that $\mathbf{F}(\iota )$ is injective. Thus we deduce $\mathbb{L} \mathbf{F}(\Delta(\lambda))=0$. We assume that $fl(W)>1$. Then there exists a standard module $\Delta(\lambda)$ and a factor module $Q$ of $W$ such that $$0 \rightarrow \Delta(\lambda) \rightarrow W \rightarrow Q \rightarrow 0$$ is a short exact sequence. Since $\mathbb{L} \mathbf{F}(\Delta(\lambda))=0$, we obtain $$0 \rightarrow \mathbb{L} \mathbf{F}(W) \rightarrow \mathbb{L} \mathbf{F}(Q) \rightarrow \mathbf{F}(\Delta(\lambda)) \rightarrow \mathbf{F}(W) \rightarrow \mathbf{F}(Q) \rightarrow 0$$ from the short exact sequence (2-3). Therefore we deduce from the induction hypothesis that $\mathbb{L}\mathbf{F}(M)=0$. Secondly, we also show the assertion in the case $i>1$ by induction on $fl(W)$. If $fl(W)=1$, then there exists $\lambda \in \Lambda$ such that $\Delta(\lambda)$ is isomorphic to $W$. Hence we have $\mathbb{L}_{i+1} \mathbf{F}(\Delta(\lambda)) \cong \mathbb{L}_i \mathbf{F}(K(\lambda))$ from the short exact sequence (2-2). Therefore the assertion follows from the induction hypothesis. If $fl(W)>1$, then we obtain $$\cdots \rightarrow \mathbb{L}_{i+1} \mathbf{F}(\Delta(\lambda)) \rightarrow \mathbb{L}_{i} \mathbf{F}(W) \rightarrow \mathbb{L}_i \mathbf{F}(Q) \rightarrow \mathbb{L}_i \mathbf{F}(\Delta(\lambda)) \rightarrow \cdots.$$ Thus we have $\mathbb{L}_{i} \mathbf{F}(W) \cong \mathbb{L}_i \mathbf{F}(Q)$ for any $i \geq 2$. Therefore the assertion follows from the induction hypothesis. \[gr\] Let $Y \in \overline{S} \mod$. Then we have $\mathbb{L}_i \mathbf{F}(Y)=0$ for any $i>0$. We show the assertion by induction on $i$. We assume that $\lambda \in \Lambda$ satisfies that $[Y:L(\lambda )] \neq 0$. For our aim, it is sufficient to show that $\mathbb{L}_i \mathbf{F}(L(\lambda ))=0$ for any $i>0$. If $i=1$, then there exists a submodule $K$ of $\Delta(\lambda )$ such that $$0 \rightarrow K \rightarrow \Delta(\lambda ) \rightarrow L(\lambda ) \rightarrow 0$$ is a short exact sequence. This short exact sequence (2-4) induces the long exact sequence as follows: $$\cdots \rightarrow \mathbb{L} \mathbf{F}(\Delta (\lambda )) \rightarrow \mathbb{L} \mathbf{F}(L(\lambda )) \rightarrow \mathbf{F}(K) \rightarrow \mathbf{F}(\Delta (\lambda )) \rightarrow 0.$$ Thus we deduce from Lemma \[F\] that $\mathbb{L} \mathbf{F}(\Delta (\lambda ))=0$. If $\mathbf{F}(L(\lambda ))=0$, then $\mathbf{F}(K)=0$. Hence we have $\mathbb{L} \mathbf{F}(L(\lambda ))=0$. Let $\mathbf{F}(L(\lambda ))=L(\lambda )$. Then we deduce that if $L(\mu)$ is a composition factor of $K$, then we have $\mu < \lambda.$ Thus we obtain from Lemma \[F\] that $\mathbf{F}(L(\mu ))=L(\mu )$. It follows from Lemma \[sat\] that $\mathbf{F}(K) = K$ and $\mathbf{F}(\Delta (\lambda )) = \Delta(\lambda )$. Therefore we deduce $\mathbb{L} \mathbf{F}(L (\lambda ))=0$. If $i>1$, then the short exact sequence (2-4) induces the long exact sequence as follows: $$\cdots \rightarrow \mathbb{L}_{i+1} \mathbf{F} (\Delta (\lambda )) \rightarrow \mathbb{L}_{i+1} \mathbf{F}(L(\lambda )) \rightarrow \mathbb{L}_i \mathbf{F}(K) \rightarrow \mathbb{L}_i \mathbf{F} (\Delta (\lambda )) \rightarrow \cdots .$$ Then it follows from Lemma \[F\] that $$\mathbb{L}_{i+1} \mathbf{F}(L(\lambda )) \cong \mathbb{L}_i \mathbf{F}(K).$$ Hence we obtain $\mathbb{L}_{i+1} \mathbf{F}(\Delta (\lambda ))=0$ by the induction hypothesis. \[lem2\] Let $X, Y \in \overline{S}\mod$. Then we have $${\rm Tor}_i^S(X,Y) \cong {\rm Tor}_i^{\overline{S}}(X,Y)$$ for any $i \geq 0$. For $X \in \mod \overline{S}$, we define $\mathbf{G}:=X \otimes_{\overline{S}} -$. Then $\mathbf{F}$ and $\mathbf{G}$ are right exact functors, and $\mathbf{F}(P)$ is $\mathbf{G}$-acyclic for any $P \in S \proj$. Hence there exists a Grothendieck spectral sequence (for example, see [@CE]) $E = (E^r_{p, q}, E_n)$ of $S\mod$ such that for each $Y \in S\mod$, the following holds: $$E_{p,q}^2 \cong {\rm Tor}^{\overline{S}}_{p}(X,\mathbb{L}_q \mathbf{F}(Y)), E_n = {\rm Tor}_{p+q}^S(X,Y).$$ Moreover we deduce from Lemma \[gr\] that $\forall i>0, \mathbb{L}_i \mathbf{F}(Y)=0$ for $i>0$. Hence the assertion holds. \[CE\] Let $\Lambda, \Gamma, \Sigma$ be algebras over $\kk$. For $A \in \Mod \Lambda \otimes \Sigma$, $B \in \Lambda \Mod \Sigma, C \in \Gamma \otimes \Sigma \Mod$, we assume that $${\rm Tor}_n^{\Lambda}(A,B) = 0 = {\rm Tor}_n^{\Sigma}(B,C)$$ for any $n>0$. Then we obtain $${\rm Tor}_i^{\Lambda \otimes \Sigma}(A \otimes_{\Lambda} B, C) \cong {\rm Tor}_i^{\Lambda \otimes \Gamma}(A, B \otimes_{\Sigma}C)$$ for any $i \geq 0$. Main result 1 ------------- \[thm2\] Let $S$ be a quasi-hereditary algebra over $\kk$. We assume that $H$ appears in a heredity chain of $S$. Then there exists a surjective graded algebra homomorphism: $$\phi :{\rm HH}^{\ast}(S) \twoheadrightarrow {\rm HH}^{\ast}(S/H).$$ We deduce from Lemma \[lem2\] and Lemma \[CE\] that $$\begin{aligned} {\rm Tor}_i^{S^{\rm en}} (\overline{S}^{\rm en}, S) &\cong & {\rm Tor}_i^S (\overline{S}, \overline{S}^{\op} \otimes_{S^{\op}} S)\\ & \cong & {\rm Tor}_i^{\overline{S}} (\overline{S}, \overline{S}^{\op} \otimes_{S^{\op}} S).\end{aligned}$$ Hence we obtain $${\rm Tor}_i^{S^{\rm en}} (\overline{S}^{\rm en}, S)=0$$ for any $i \geq 1$. Thus we can define the following map for each $i$. $$\begin{aligned} \phi_i : {\rm HH}^i(S) & \twoheadrightarrow & {\rm HH}^i(\overline{S})\\ \hspace{3em} [\alpha] & \mapsto & [\mathbf{F}^{\rm en}(\alpha )]. \end{aligned}$$ Firstly, we show that $\phi_i$ is well-defined. Let ($P_{\bullet}$, $d_{\bullet}$) be a projective resolution of $S$ as left $S^{\rm en}$-modules. We deduce from ${\rm HH}^i(S) \cong \Ker \Hom (d_{i+1}, S)/ \Im \Hom (d_i, S)$ that if $\alpha \in \Im \Hom (d_i, S)$, then there exists $\beta \in \Hom_{S^{\rm en}}(P_i, S)$ such that $\alpha=\beta \circ d_i$ and the following holds: $$\mathbf{F}^{\rm en}(\alpha )=\mathbf{F}^{\rm en}(\beta \circ d_i)=\mathbf{F}^{\rm en}(\alpha )\circ \mathbf{F}^{\rm en}(d_i).$$ Thus $\mathbf{F}^{\rm en}(\alpha ) \in \Im \Hom(\mathbf{F}^{\rm en}(d_i), \overline{S})$. Using the fact that $\mathbf{F}^{\rm en}$ is a functor and the Yoneda product is functorial, it follows that $\phi$ is a graded algebra homomorphism. It is straightforward to check that $\phi$ is a surjection. Let $S$ be a quasi-hereditary algebra with a heredity chain $S=H_{0}>H_{1}> \cdots >H_{n}=0$. Then we deduce the surjective graded algebra homomorphism from ${\rm HH}^{\ast}(S)$ to ${\rm HH}^{\ast}(S/H_{n-1})$ in a different way. Again we take a heredity ideal $H_{n-2}/H_{n-1}$ of $S/H_{n-1}$. Then we have the surjective graded algebra homomorphism from ${\rm HH}^{\ast}(S/H_{n-1})$ to ${\rm HH}^{\ast}(S/H_{n-2})$. Thus we can show Theorem \[thm2\] by repeating this process inductively. We provide an example of a finite dimensional algebra which dose not hold Theorem \[thm2\]. Let $A$ be the algebra over a field defined by the following quiver $$\xymatrix{1\ar@(dl,ul)^{\gamma_{1}} \ar@/^/[r]^{\alpha} &2\ar@(dr,ur)_{\gamma_2}\ar@/^/[l]^{\beta}}$$ with relations $\gamma_i ^2 \; (i=1, 2), \gamma_2 \alpha, \gamma_1 \beta, \alpha \beta \alpha, \beta \gamma_{2}$. Let $I:= A e_2 A$. Then there dose not exist a surjective graded algebra homomorphism form ${\rm HH}^{\ast}(A)$ to ${\rm HH}^{\ast}(A/I)$. We deduce the corollary from Theorem \[thm2\] as follows: \[cor1\] Let $S$ be a quasi-hereditary algebra. Then there exists an idempotent $\xi $ of $S$ such that $\xi S \xi$ is a quasi-hereditary algebra and $$\psi : {\rm HH}^{\ast}(S) \twoheadrightarrow {\rm HH}^{\ast}(\xi S \xi)$$ is a graded algebra surjection. Let $(S_{\xi})^{'}$ be the Ringel dual of $S_{\xi}$. Then there exists a two sided ideal $H^{'}$ of the Ringel dual $S^{'}$ of $S$ such that $H^{'}$ appears in a heredity chain of $S^{'}$, and the following isomorphism is given by the property of the Ringel duality, this is proved by Ringel in [[@Rin Appendix]]{}. $$(\xi S \xi)' \cong S'/H'.$$ Hence we deduce $$\begin{aligned} {\rm HH}^{\ast}(\xi S \xi ) &\cong& {\rm HH}^{\ast}(S'/H'), \\ {\rm HH}^{\ast}(S) &\cong& {\rm HH}^{\ast}(S').\end{aligned}$$ Moreover we have $${\rm HH}^{\ast}(S') \twoheadrightarrow {\rm HH}^{\ast}(S'/H')$$ is a graded algebra surjection from Theorem \[thm2\]. Therefore we can construct $\psi$. Hochschild cohomology of the $q$-Schur algebras =============================================== In this section we compute the Hochschild cohomology of the $q$-Schur algebras. From the rest of this paper, we denote by $\kk$ a field of characteristic $l \geq 0$. We put $$e:= {\rm inf} \{i \in \mathbb{Z}_{ \geq 1}\; | \; 1+ q + \cdots + q^{i-1} = 0 \; {\rm in}\; \kk\}.$$ $e$ is called the quantum characteristic. Preliminaries ------------- We prepare a certain derived equivalence and an explicit bimodule projective resolution to compute the Hochschild cohomology of $q$-Schur algebras. In this subsection, we use some combinatorial notion (e.g. $e$-weight, $e$-core and $e$-abacus) to describe the above derived equivalence, so see [[@W Chapter 2]]{} for more detail. Let $B$ a block of $\mathscr{S}_q(n, n)$ of $e$-weight[^1] $w \in \mathbb{Z}_{\geq 0}$ and $B^{'}$ be a block of some $\mathscr{S}_q(m, m)$ with the same $e$-weight $w$. Then $B$ and $B^{'}$ are derived equivalent [@CR]. \[block\] The blocks of $q$-Schur algebras $\mathscr{S}_q(n, n)$ are in one-one correspondence with pairs $(w, \tau)$, where $w \in \mathbb{Z}_{\geq 0}$ is an $e$-weight, $\tau$ is an $e$-core[@xdefthefnmark[\[fn\_label1\]]{}footnotemark]{} of size $n - we$. From Theorem \[block\], we denote by $\mathbf{B}_{\tau, w}$ the block of the $q$-Schur algebra $\mathscr{S}_q (n, n)$ corresponding to the pair $(w, \tau)$. \[Chuang-Rouquier [[@CR §7.6]]{}\] \[CR\] Let $\tau, \tau^{'}$ be $e$-cores. Then the following derived equivalence holds: $$\mathcal{D}^b(\mathbf{B}_{\tau, w} \mod) \simeq \mathcal{D}^b(\mathbf{B}_{\tau^{'}, w} \mod).$$ We suppose that $p,w \in \mathbb{Z}_{\geq 0}$ are fixed. A $p$-core $\rho$ is said to be a [*Rouquier $p$-core*]{} if it has a $p$-abacus[@xdefthefnmark[\[fn\_label1\]]{}footnotemark]{} presentation, on which there are at least $w-1$ more beads on runner $i$, than on runner $i-1$, for $i =1, \ldots ,p-1$. Let $w \in \mathbb{Z}_{\geq 0}$ and $\rho$ be a Rouquier core of $e$-weight $w$. We say that $\mathbf{B}_{\rho, w}$ is a [*Rouquier block*]{} of a $q$-Schur algebra. The notion of the Rouquier block give important information to us. For example, the following theorem holds. \[Chuang-Miyachi [[@MR2761936 Theorem 18]]{}\] \[CM\] Let $\mathbb{F}_l$ be a finite field of $l$ elements. We assume that $l=0$ or $w < l, q \in \mathbb{F}_l $. Then $\mathbf{B}_{\rho, w}$ and $B_0 (\mathscr{S}_q (e, e))^{\otimes w} \rtimes \kk \frks_w$ are Morita equivalent, where $B_0 (\mathscr{S}_q (e, e))$ is the principal block of the $q$-Schur algebra $\mathscr{S}_q(e,e)$. \[mori\] We use the same notation in Theorem \[CM\]. $B_0 (\mathscr{S}_q (e, e)) $ is Morita equivalent to $A_e $, where $A_e$ is the algebra over a field defined by the following quiver $$Q :=(1) \overset{\alpha(1)}{\underset{\alpha^-(1)}{\rightleftarrows}} \cdots (i-1) \overset{\alpha(i-1)}{\underset{\alpha^-(i-1)}{\rightleftarrows}} (i) \overset{\alpha(i)}{\underset{\alpha^-(i)}{\rightleftarrows}} (i+1) \cdots \overset{\alpha(e-1)}{\underset{\alpha^-(e-1)}{\rightleftarrows}} (e),$$ with relations $$\begin{aligned} &\alpha(i)\alpha(i-1), \; \alpha^-(i-1)\alpha^-(i) ,\\ &\alpha(i-1)\alpha^-(i-1)-\alpha^-(i)\alpha(i) \; (2\leq i \leq e-1), \; \alpha(e-1)\alpha^-(e-1).\end{aligned}$$ Theorem \[CR\], Theorem \[CM\] and Lemma \[mori\] implies the following theorem. \[a\] Let $w \in \mathbb{Z}_{\geq 0}$ and $\G$ be any block of $e$-weight $w$ of $q$-Schur algebra $\mathscr{S}_q (d, d)$. Then the following graded algebra isomorphism holds: $${\rm HH}^{\ast} (\G) \cong {\rm HH}^{\ast} (A_e^{\otimes w} \rtimes \kk \frks_w ).$$ In Theorem \[a\], we take a block of a $q$-Schur algebra $\mathscr{S}_q (d, d)$, but we can expand this in the following way: 1. If $n \geq d$, then $\mathscr{S}_q (n, d)$ is Morita equivalent to $\mathscr{S}_q (d, d)$. 2. If $n < d$, then we can choose the canonical idempotent (Green’s idempotent) $\epsilon \in \mathscr{S}_q (d, d)$ which induces the following isomorphism: $$\mathscr{S}_q (n, d) \cong \epsilon \mathscr{S}_q (d, d) \epsilon.$$ In this case, $\epsilon$ satisfies the assumption of Corollary \[cor1\]. Thus we deduce from Corollary \[cor1\] that the following graded algebra surjection: $${\rm HH}^{\ast}(\mathscr{S}_q (d, d)) \twoheadrightarrow {\rm HH}^{\ast} (\mathscr{S}_q (n, d)).$$ Let $(R_{\bullet}, d_{\bullet})$ be the minimal projective resolution of $A_e$ as $A_e$-bimodules. We have $$\begin{aligned} &&R_{2s} = \bigoplus_{i = s+1}^e P(i, i) \oplus \left( \bigoplus_{n=1}^s \left( \bigoplus_{j=s-n+1}^{e-2n} P(j, j+2n) \oplus P(j+2n, j)\right)\right),\\ &&R_{2s+1} = \bigoplus_{m=1}^{s+1}\left(\bigoplus_{t=s+2-m}^{e-(2m-1)} P(t, t+2m-1) \oplus P(t+2m-1, t)\right).\end{aligned}$$ The differential of $(R_{\bullet},d_{\bullet})$ is given as follows: $$\begin{aligned} d_0 : R_0 &\to& A_e,\\ (i) \otimes (i) &\mapsto& (i),\end{aligned}$$ Let $s \geq0$ and $1 \leq t \leq 2s+1$. With the above notation, for $i \geq 1$, we define the differential $d_i : R_{i} \to R_{i-1}$ recursively as follows: $$\begin{aligned} d_{4s+1} : R_{4s+1} &\to& R_{4s},\\ (t) \otimes (t+2m-1) &\mapsto& \left\{ \begin{array}{llll} (-1)^{m+1} \alpha (t-1) \otimes (t+2m-1) \\ +(-1)^m (t) \otimes \alpha^-(t+2m-2) \\ + (t)\otimes \alpha (t+2m-1) \\ + \alpha^-(t) \otimes (t+2m-1), &( 1 \leq m < 2s+1);\\ -(t) \otimes \alpha^-(t+2m-2) \\ +\alpha^-(t) \otimes (t+2m-1), &(m=2s+1); \end{array} \right. \\ (t+2m-1) \otimes (t) &\mapsto& \left\{ \begin{array}{llll} (-1)^{m+1} (t+2m-1) \otimes \alpha^-(t-1) \\ +(-1)^{m} \alpha(t+2m-2) \otimes (t) \\ + (t+2m-1)\otimes \otimes \alpha (t) \\ + \alpha^-(t+2m-1) \otimes (t), &( 1 \leq m < 2s+1);\\ (t+2m-1) \otimes \alpha(t)\\ -\alpha(t+2m-2) \otimes (t), &(m=2s+1). \end{array} \right.\end{aligned}$$ Let $s \geq0$ and $1 \leq j \leq 2s+1$. $$\begin{aligned} d_{4s+2} : R_{4s+2} &\to& R_{4s+1},\\ (i) \otimes (i) &\mapsto& \alpha(i-1) \otimes (i) -(i) \otimes \alpha^-(i-1) \\ &&-(i) \otimes \alpha(i) + \alpha^-(i) \otimes (i), \hspace{4.5em} ( 2s+1 \leq i \leq e);\\ (j) \otimes (j+2n) &\mapsto& \left\{ \begin{array}{lll} (-1)^n \alpha(j-1) \otimes (j+2n) \\ + (-1)^{n+1} (j) \otimes \alpha^-(j+2n-1) \\ +\alpha^-(j) \otimes (j+2n) \\ - (j) \otimes \alpha (j+2n), &(1 \leq n < 2s+1);\\ (j) \otimes \alpha^-(j+2n-1)\\ + \alpha^-(j) \otimes (j+2n), &(n=2s+1); \end{array} \right. \\ (j+2n) \otimes (j) &\mapsto& \left\{ \begin{array}{lll} (-1)^n \alpha(j+2n-1) \otimes (j) \\ + (-1)^{n+1} (j+2n) \otimes \alpha^-(j-1) \\ +\alpha^-(j+2n) \otimes (j) \\ - (j+2n) \otimes \alpha(j), &(1 \leq n < 2s+1);\\ - \alpha(J+2n-1) \otimes (j) \\ - (j+2n) \otimes \alpha(j), &(n=2s+1). \end{array} \right. \end{aligned}$$ Let $s \geq0$ and $1 \leq t \leq 2s+2$. $$\begin{aligned} d_{4s+3} : R_{4s+3} &\to& R_{4s+2},\\ (t) \otimes (t+2m-1) &\mapsto& \left\{ \begin{array}{llll} (-1)^{m} \alpha (t-1) \otimes (t+2m-1) \\ +(-1)^{m+1} (t) \otimes \alpha^-(t+2m-2) \\ + (t)\otimes \alpha (t+2m-1) \\ + \alpha^-(t) \otimes (t+2m-1), &(1 \leq m < 2s+2);\\ -(t) \otimes \alpha^-(t+2m-2)\\ +\alpha^-(t) \otimes (t+2m-1), &(m=2s+2); \end{array} \right. \\ (t+2m-1) \otimes (t) &\mapsto& \left\{ \begin{array}{llll} (-1)^{m} (t+2m-1) \otimes \alpha^-(t-1) \\ +(-1)^{m+1} \alpha(t+2m-2) \otimes (t) \\ + (t+2m-1)\otimes \alpha (t) \\ + \alpha^-(t+2m-1) \otimes (t), &( 1 \leq m < 2s+2);\\ (t+2m-1) \otimes \alpha(t)\\ -\alpha(t+2m-2) \otimes (t), &(m=2s+2). \end{array} \right.\end{aligned}$$ Let $s \geq0$ and $1 \leq j \leq 2s+2$. $$\begin{aligned} d_{4s+4} : R_{4s+4} &\to& R_{4s+3},\\ (i) \otimes (i) &\mapsto& -\alpha(i-1) \otimes (i) +(i) \otimes \alpha^-(i-1) \\ &&-(i) \otimes \alpha(i) + \alpha^-(i) \otimes (i), \hspace{3em} ( 2s+1 \leq i \leq e);\\ (j) \otimes (j+2n) &\mapsto& \left\{ \begin{array}{lll} (-1)^{n+1} \alpha(j-1) \otimes (j+2n) \\ + (-1)^{n} (j) \otimes \alpha^-(j+2n-1) \\ +\alpha^-(j) \otimes (j+2n) \\ - (j) \otimes \alpha (j+2n), &(1 \leq n < 2s+2);\\ (j) \otimes \alpha^-(j+2n-1) \\ + \alpha^-(j) \otimes (j+2n), &(n=2s+1); \end{array} \right. \\ (j+2n) \otimes (j) &\mapsto& \left\{ \begin{array}{lll} (-1)^{n+1} \alpha(j+2n-1) \otimes (j) \\ + (-1)^{n} (j+2n) \otimes \alpha^-(j-1) \\ +\alpha^-(j+2n) \otimes (j) \\ - (j+2n) \otimes \alpha(j), &(1 \leq n < 2s+2);\\ -(j+2n) \otimes \alpha(j)\\ - \alpha(j+2n-1) \otimes (j), &(n=2s+2). \end{array} \right. \end{aligned}$$ We construct the minimal $A_e$-bimodule projective resolution of $A_e$, that is, we construct the following exact sequence: $$R_{\bullet} : \cdots \to R_n \xrightarrow{d_n} R_{n-1} \to \cdots \to R_1 \xrightarrow{d_1} R_0 \xrightarrow{d_0} A_e \to 0,$$ where $R_n= \bigoplus P(i, j)$ and $P(i, j)$ is the projective $A_e$-bimodule $A_e (i) \otimes (j) A_e$. We obtain from [@H] that the projective module $P(i, j)$ occurs in $R_n$ as many times as ${\rm dim}\;{\rm Ext}_{A_e}(S_i, S_j)$. Hence we deduce $$\begin{aligned} &&R_{2s} = \bigoplus_{i = s+1}^e P(i, i) \oplus \left( \bigoplus_{n=1}^s \left( \bigoplus_{j=s-n+1}^{e-2n} P(j, j+2n) \oplus P(j+2n, j)\right)\right),\\ &&R_{2s+1} = \bigoplus_{m=1}^{s+1}\left(\bigoplus_{t=s+2-m}^{e-(2m-1)} P(t, t+2m-1) \oplus P(t+2m-1, t)\right).\end{aligned}$$ From this result, we can construct a differential of $A_e$ as above. It is straightforward to check that $d_{\bullet}$ define a complex. In [[@ES]]{}, they construct an explicit bimodule projective resolution of tame blocks of Hecke algebras of type $A$. Following their method, we show that $d_i$ is the differential of a minimal projective resolution of $A_e$ for any $i \geq 0$. We construct a projective resolution of $A_e/ J A_e$ as left $A_e$-modules, where $J:= J(A_e)$. We denote by $(P_{\bullet}, \delta_{\bullet})$ a projective resolution of $A_e/ J A_e$. Then we have $R_m \otimes_{A_e} A_e/ J A_e \cong P_m$ for all $m \geq 0$ and we deduce that the diagram $$\xymatrix{ \cdots \ar[r] & R_{m+1} \otimes_{A_e} A_e/J A_e \ar[d]_{\cong} \ar[r]^{d_{m+1} \otimes {\rm id}} & R_{m} \otimes_{A_e} A_e/J A_e \ar[d]^{\cong} \ar[r]& \cdots &\\ \cdots \ar[r] & P_{m+1} \ar[r]_{\delta_{m+1}} & P_{m} \ar[r] & \cdots. }$$ commutes for all $m \geq 1$. We suppose that $\Ker d_{m} \not\subseteq \Im d_{m+1}$ for some $m \geq 1$. Then there exists a non-zero map $\Ker d_m \to \Ker d_m/\Im d_{m+1}$. Therefore there exists a simple $A_e$-bimodule $S \otimes T$ such that $S$ is a left simple $A_e$-module and $T$ is a right $A_e$-module. Moreover there exists a non-zero map $f : \Ker d_{m} \to S \otimes T$. Thus we obtain $$\begin{aligned} \Im d_{m+1} \otimes_{A_e} A_e / J A_e & \cong & \Im (d_{m+1} \otimes {\rm id} )\\ & \cong & \Ker (d_m \otimes {\rm id} )\\ & \cong & \Ker d_m \otimes_{A_e} A_e/J A_e. \end{aligned}$$ Thus we have $$\begin{aligned} A_e \otimes_{A_e} A_e / J A_e & \xrightarrow{d_{m+1} \otimes {\rm id}} & \Im d_{m+1} \otimes_{A_e} A_e /J A_e \\ & \cong & \Ker d_m \otimes_{A_e} A_e / J A_e \\ & \xrightarrow{f \otimes {\rm id} }& (S \otimes T ) \otimes_{A_e} A_e / J A_e \\ & \cong & S \otimes T.\end{aligned}$$ On the other hand, we apply the functor $- \otimes_{A_e} A_e/J A_e$ to the following exact sequence: $$R_{m+1} \xrightarrow{d_{m+1}} \Ker d_m \xrightarrow{f} S \otimes T .$$ Then we have $$f \circ d_{m+1} = 0.$$ Thus we obtain $$(f \otimes {\rm id}) \circ (d_{m+1} \otimes {\rm id} ) \neq (f \circ d_{m+1} ) \otimes {\rm id}.$$ This is a contradiction. Thus we deduce that $\Ker d_{m} \subseteq \Im d_{m+1}$ for any $m \geq 1$. We give another simple and direct proof of the following Theorem by using the above explicit bimodule projective resolution. \[hhgp\] We have the dimension of the Hochschild cohomology group of $A_e$ as follows: $$\begin{aligned} {\rm dim} \; {\rm HH}^n (A_e) = \left\{ \begin{array}{lll} e, \; & ( n = 0);\\ 1, & ( 1 \leq n \leq 2(e-1)); \\ 0, & ( 2(e-1) < n). \end{array} \right.\end{aligned}$$ If $n=0$, then we have ${\rm dim} \;{\rm HH}^0 (A_e) = {\rm dim} \; Z(A_e) = e$, where $Z(A_e)$ is the center of $A_e$. If $1 \leq n \leq \gldim A_e = 2(e-1)$, then there exists a short exact sequence as follows: $$0 \to \Ker d_n \to R_n \to \Ker d_{n-1} \to 0.$$ We apply the functor $\Hom_{\A_}(-, A_e)$ to this exact sequence. Then we have $$\begin{aligned} 0 &\to& \Hom_{\A_} (\Ker d_{n-1} , A_e) \to \Hom_{\A_} (R_n, A_e) \to \Hom_{\A_} (\Ker d_n, A_e) \\ & \to & \Ext^1_{\A_} ( \Ker d_{n-1}, A_e) \to 0.\end{aligned}$$ Since $\Ext^i_{\A_} (R_n, A_e)=0$ for all $i \geq 1$, we deduce $$\Ext^1_{\A_} (\Ker d_{n-1}, A_e ) \cong \Ext^2_{\A_} ( \Ker d_{n-2}, A_e) \cong \cdots \cong \Ext^{n+1}_{\A_} (A_e, A_e) .$$ We shall determine ${\rm dim}\Hom_{\A_} (R_n, A_e)$ and ${\rm dim}\Hom_{\A_} (\Ker d_n, A_e)$ to determine ${\rm dim}\; {\rm HH}^i (A_e)$. It is straightforward to show ${\rm dim} \Hom_{\A_} (R_n, A_e) = 2e-n-1.$ Next we compute ${\rm dim}\;\Hom_{A_e^{{\rm en}}}(\Ker d_n, A_e)$. We have $$R_{n+1}/ \Ker d_{n+1} \cong \Ker d_n$$ for all $n \geq 0$. Then we have $$\begin{aligned} \Hom_{\A_} (\Ker d_n, A_e) & \cong& \Hom_{\A_} (R_{n+1}/ \Ker d_{n+1}, A_e)\\ &=& \{ \eta \in \Hom_{\A_}(R_{n+1}, A_e) \; | \; \eta (\Ker d_{n+1} ) = 0. \}.\end{aligned}$$ Thus we obtain $$\begin{aligned} {\rm dim} \Hom_{\A_} (\Ker d_n, A_e) = \left\{ \begin{array}{ll} e-2s-1, \; & ( n= 4s+1, 4s+2);\\ e-2s-2, & ( n= 4s+3, 4s+4 ). \end{array} \right.\end{aligned}$$ It is clear that ${\rm dim} \;{\rm HH}^n (A_e)=0$ for all $n > \gldim A_e$. Main result 2 ------------- From Theorem \[hhgp\], we can determine the ring structure of ${\rm HH}^{\ast} (A_e)$. \[prop\] We have the following $\mathbb{Z}$-graded $\kk$-algebra isomorphism: $${\rm HH}^{\ast}(A_e) \cong \kk [z_{1}, z_{2}, \cdots, z_{e-1}, x, y]/J,$$ where ${\rm deg}\ z_{i}=0$, ${\rm deg}\ x=1$, ${\rm deg}\ y=2$, and $$\begin{aligned} J= \left<{ \begin{aligned} z_{i}z_{j},\ z_{i}x, \ z_{i}y, \ x^2, \ xy^{e-1},\ y^e \end{aligned} } \right> .\end{aligned}$$ We obtain from Lemma \[hhgp\] that ${\rm dim}\;{\rm HH}^1 (A_e) = 1$, so we can take $$0 \neq \eta \in {\rm HH}^1 (A_e).$$ Since the Hochschild cohomology ring is super commutative [[@G]]{}, we have $$\eta^2 = 0.$$ Next, we take $$0 \neq \theta \in {\rm HH}^2 (A_e).$$ Then, for $1 \leq s \leq e-1$, we show that $$\eta \theta^{s-1}, \theta ^{s} \neq 0$$ by induction on $s$. In the case $s=2$, it is trivial. So we assume that the claim holds for $s>2$. Then we can compute the Yoneda product from an explicit projective resolution of $A_e$, and we have $\theta^{s+1} \neq 0$. We show that $\eta \theta^{2s} \neq 0$. We can also have an explicit computation of the Yoneda product, and we have $\eta \theta^{2s} \neq 0$. From Proposition \[prop\], we obtain the following corollary. The even part of ${\rm HH}^{\ast}(A_e)$ is given as follows: $${\rm HH}^{\rm ev}(A_e)\cong \kk [z_1, \ldots , z_{e-1}, y]/ \langle z_i z_j , z_k y , y^e \; | \; 1 \leq i, j \leq e-1 \rangle ,$$ where [deg]{} $ z_i =0$, [deg]{} $y =2$. We extend this result to ${\rm HH}^n(A_e^{\otimes w} \rtimes \kk \mathfrak{S}_{w})$ by using the following result. From the rest of this section, we assume that $l=0$ or $w<l$. \[[Alev-Farinati-Lambre-Solotar [@AFLS Proposition 3.1]]{}, [Etingof-Oblomkov [@EO Theorem 3.1]]{}\] \[eo\] Let $w \in \mathbb{Z}_{\geq 0}$, and $\Gamma$ be an algebra over $\kk$. Then we have an isomorphism as $\mathbb{Z}_{\geq 0}$-graded vector space over $\kk$. $$\begin{aligned} {\rm HH}^{\ast}(\Gamma^{\otimes w} \rtimes \kk \mathfrak{S}_{w}) \cong \bigoplus_{ \lambda \in P_w} \bigotimes_{ i \geq 1} ({\rm HH}^{\ast}(\Gamma)^{\otimes p_i (\lambda )})^{\mathfrak{S}_ {p_i (\lambda )}},\end{aligned}$$ where $P_w$ is a set of partitions of $w$ and $p_i (\lambda)$ is a multiplicity of occurrence of $i$ in a partition $\lambda$. In the following, we concentrate on the even part $\kk [z_1, \ldots , z_{e-1}, y]/ \langle z_i z_j , z_k y, y^e \rangle$ of ${\rm HH}^{\ast}(A_e)$. We expand this even part to ${\rm HH}^{{\rm ev}}(A_e^{\otimes w} \rtimes \kk \mathfrak{S}_{w})$ by using Proposition \[eo\]. We consider the following $\kk$-algebra homomorphism: $$\begin{aligned} \phi : \kk [z_1, \ldots , z_{e-1}, y]/ \langle z_i z_j , z_k y, y^e \rangle &\twoheadrightarrow & \kk [y]/ \langle y^e \rangle \\ y & \mapsto & y\\ z_i &\mapsto & 0.\end{aligned}$$ Then we have $\Ker \phi = \langle z_i \; |\; 1 \leq i \leq e-1 \rangle$. We consider the $w$-fold tensor product of $\phi$ as commutative ring. $$\phi ^{\otimes w} : ( \kk [z_1, \ldots , z_{e-1}, y]/ \langle z_i z_j , z_k y, y^e \rangle )^{\otimes w} \twoheadrightarrow ( \kk [y]/ \langle y^e \rangle )^{\otimes w}.$$ We can regard $\phi^{\otimes w}$ as a $\kk \frks_w$-homomorphism, and apply the functor $\Hom_{\kk \frks_w}(\kk, -)$ to $\phi^{\otimes w}$. Then we obtain $$\Hom_{\kk \frks_w}(\kk, \phi^{\otimes w} ) : ((\kk [z_1, \ldots , z_{e-1}, y]/ \langle z_i z_j , z_k y , y^e \rangle )^{\otimes w} )^{\frks_w} \twoheadrightarrow (( \kk [y]/ \langle y^e \rangle )^{\otimes w}) ^{\frks_w} .$$ Since $$((\kk [y]/ \langle y^e \rangle)^{\otimes w} )^{\frks_w} \cong (\kk [y_1, \ldots , y_w]/ \langle y_1^e, \ldots, y_w^e \rangle )^{\frks_w},$$ we consider the kernel of the following natural surjection to give generators of $((\kk [y]/ \langle y^e \rangle)^{\otimes w} )^{\frks_w}$: $$\begin{aligned} \pi : \Lambda_w &\to& (\kk [y_1, \ldots , y_w]/ \langle y_1^e, \ldots, y_w^e \rangle )^{\frks_w}\\ x_i & \mapsto & y_i ,\end{aligned}$$ where $\Lambda_w = \kk [x_1, \ldots , x_w]^{\frks_w}$ is the ring of symmetric polynomials in $w$ variables, for more detail on symmetric polynomials see [[@M]]{}. It is known that $$\Ker \pi = \langle p_{e+1}, \ldots, p_{e+w+1} \rangle ,$$ where $p_i$ is the power-sum symmetric polynomial [[[@GA Corollary 3.3]]{}]{}. Consequently, we obtain the following $\mathbb{Z}_{\geq 0}$-graded $\kk$-algebra isomorphism: $$\mathrm{HH}^{{\rm ev}}\left(A_{e}^{\otimes w}\rtimes\mathbf{k}\mathfrak{S}_{w}\right) / \Ker \Hom_{\kk \mathfrak{S}_w}(\kk, \phi^{\otimes w})\simeq\Lambda_{w}/\left\langle p_{e+1},\cdots,p_{e+w+1}\right\rangle.$$ acknowledgement {#acknowledgement .unnumbered} =============== The author is greatly indebted to Hyohe Miyachi, Akihiro Tsuchiya and Yoshiyuki Kimura for their helpful advices. This work is supported by by Grant-in-Aid for JSPS Fellowships No. H15J09492. [^1]: \[fn\_label1\][see, for example [[@W Chapter 2]]{}]{}

ArXiv

--- abstract: | The singular cochain complex of a topological space is a classical object. It is a Differential Graded algebra which has been studied intensively with a range of methods, not least within rational homotopy theory. More recently, the tools of Auslander-Reiten theory have also been applied to the singular cochain complex. One of the highlights is that by these methods, each Poincaré duality space gives rise to a Calabi-Yau category. This paper is a review of the theory. author: - Peter Jørgensen title: 'Calabi-Yau categories and Poincaré duality spaces' --- Primary 16E45, 16G70, 55P62. Algebraic topology, Auslander-Reiten quivers, Auslander-Reiten theory, compact objects, derived categories, Differential Graded algebras, Differential Graded homological algebra, Differential Graded modules, labelled quivers, representation types, Riedtmann Structure Theorem, singular cochain complexes, tame and wild, topological spaces. Introduction ============ Finite dimensional algebras over a field are classical, well studied mathematical objects. Their representation theory is a particularly large and active area which has inspired a number of powerful mathematical techniques, not least Auslander-Reiten theory which is a beautiful and effective set of tools and ideas. See Appendix \[app:AR\] and the references listed there for an introduction. It seems reasonable to look for applications of Auslander-Reiten (AR) theory to areas outside representation theory. Specifically, let $X$ be a topological space. The singular cochain complex ${\operatorname{C}}^*(X;k)$ with coefficients in a field $k$ of characteristic $0$ is a Differential Graded algebra which has been studied intensively, in particular in rational homotopy theory, see [@FHTbook]. For an introduction to Differential Graded (DG) homological algebra, see Appendix \[app:DG\] and the references listed there. The singular cohomology ${\operatorname{H}}^*(X;k)$ is defined as the cohomology of the complex ${\operatorname{C}}^*(X;k)$; it is a graded algebra. Now let $X$ be simply connected with $\dim_k {\operatorname{H}}^*(X;k) < \infty$; then ${\operatorname{C}}^*(X;k)$ is quasi-isomorphic to a DG algebra $R$ with $\dim_k R < \infty$, and it is natural to try to apply AR theory to $R$. This was the subject of [@artop], [@arquiv], and [@Schmidt], and the object of this paper is to review the results of those papers. Among the highlights is Theorem \[thm:Chain\_CY\] from which comes the title of the paper. Consider the derived category of DG left-$R$-modules, ${{\mathsf D}}(R)$, which is equivalent to ${{\mathsf D}}({\operatorname{C}}^*(X;k))$ since the two DG algebras are quasi-isomorphic. The latter category has the full subcategory ${{\mathsf D}^{\operatorname{c}}}({\operatorname{C}}^*(X;k))$ consisting of compact DG modules; these play the role of finitely generated representations. Theorem \[thm:Chain\_CY\] now says that if $k$ has characteristic $0$, then $$\begin{aligned} \label{equ:z} & \mbox{${{\mathsf D}^{\operatorname{c}}}({\operatorname{C}}^*(X;k))$ is an $n$-Calabi-Yau category} \\ \nonumber & \mbox{$\Leftrightarrow$ $X$ has $n$-dimensional Poincar\'{e} duality over $k$.}\end{aligned}$$ Let me briefly explain the terminology. A triangulated category ${\mathsf T}$, such as for instance ${{\mathsf D}^{\operatorname{c}}}({\operatorname{C}}^*(X;k))$, is called $n$-Calabi-Yau if $n$ is the smallest non-negative integer for which $\Sigma^n$, the $n$th power of the suspension functor, is a Serre functor, that is, permits natural isomorphisms $${\operatorname{Hom}}_k({\operatorname{Hom}}_{{\mathsf T}}(M,N),k) \cong {\operatorname{Hom}}_{{\mathsf T}}(N,\Sigma^n M).$$ The topological space $X$ is said to have $n$-dimensional Poincaré duality over $k$ if there is an isomorphism $${\operatorname{Hom}}_k({\operatorname{H}}^*(X;k),k) \cong \Sigma^n {\operatorname{H}}^*(X;k)$$ of graded left-${\operatorname{H}}^*(X;k)$-modules. Examples of $n$-Calabi-Yau categories are higher cluster categories, see [@KellerReiten sec. 4], and examples of spaces with $n$-dimensional Poincaré duality are compact $n$-dimensional manifolds. Equation provides a link between the currently popular theory of Calabi-Yau categories and algebraic topology. It also gives a new class of examples of Calabi-Yau categories which, so far, typically have been exemplified by higher cluster categories. The new categories appear to behave very differently from higher cluster categories, cf. Section \[sec:open\], Problem \[prob:CY\]. A number of other results are also obtained, not least on the structure of the AR quiver of ${{\mathsf D}^{\operatorname{c}}}({\operatorname{C}}^*(X;k))$ which, for a space with Poincaré duality, consists of copies of the repetitive quiver ${{\mathbb Z}}A_{\infty}$, see Theorem \[thm:Chain\_Gamma\]. In a speculative vein, the theory presented here ties in with the version of non-commutative geometry in which a DG algebra, or more generally a DG category, is viewed as a non-commutative scheme. The idea is to think of the derived category of the DG algebra or DG category as being the derived category of quasi-coherent sheaves on a non-commutative scheme (which does not actually exist). There appear so far to be no published references for this viewpoint which has been brought forward by Drinfeld and Kontsevich, but it does seem to call for a detailed study of the derived categories of DG algebras and DG categories. Auslander-Reiten theory is an obvious tool to try, and [@artop], [@arquiv], and [@Schmidt] along with this paper can, perhaps, be viewed as a first, modest step. As indicated, the paper is a review. The results were known previously, the main references being [@artop], [@arquiv], and [@Schmidt]; more details of the origin of individual results are given in the introductions to the sections. There is no claim to originality, except that some of the proofs are new. It is also the first time this material has appeared together. Most of the paper is phrased in terms of the DG algebra $R$ rather than ${\operatorname{C}}^*(X;k)$, see Setup \[set:blanket\]. This is merely a notational convenience: $R$ and ${\operatorname{C}}^*(X;k)$ are quasi-isomorphic, so have equivalent derived categories. Hence, all results about the derived category of $R$ also hold for the derived category of ${\operatorname{C}}^*(X;k)$. The paper is organized as follows: Some background on DG homological algebra and AR theory is collected in two appendices, \[app:DG\] and \[app:AR\]. Section \[sec:cochain\] gives some preliminary results on cochain DG algebras and their DG modules. The main result is Theorem \[thm:Gorenstein\] which gives a number of alternative descriptions of when $R$ is a so-called Gorenstein DG algebra. The importance of this condition is that ${\operatorname{C}}^*(X;k)$ is Gorenstein precisely when $X$ has Poincaré duality. Section \[sec:AR\] studies the existence of AR triangles in the category ${{\mathsf D}^{\operatorname{c}}}(R)$, which turns out to be equivalent to $R$ being Gorenstein by Theorem \[thm:R\]. Section \[sec:local\] considers the local structure of the AR quiver $\Gamma$ of ${{\mathsf D}^{\operatorname{c}}}(R)$. If $R$ is Gorenstein with $\dim_k {\operatorname{H}}\!R \geq 2$, then Theorem \[thm:quiver\_local\_structure\] shows that each component of $\Gamma$ is isomorphic to ${{\mathbb Z}}A_{\infty}$. Section \[sec:global\] reports on work by Karsten Schmidt. It looks at the global structure of $\Gamma$ where the results are so far less conclusive. If $\dim_k {\operatorname{H}}\!R = 2$, then $\Gamma$ has precisely $d-1$ components isomorphic to ${{\mathbb Z}}A_{\infty}$, where $d = \sup \{\, i \,|\, {\operatorname{H}}^i\!R \neq 0 \,\}$. On the other hand, if $R$ is Gorenstein with $\dim_k {\operatorname{H}}\!R \geq 3$, then $\Gamma$ has infinitely many components, and if $\dim_k {\operatorname{H}}^e\!R \geq 2$ for some $e$, then it is even possible to find families of distinct components indexed by projective manifolds, and these manifolds can be of arbitrarily high dimension. Section \[sec:topology\] makes explicit the highlights of the theory for the algebras ${\operatorname{C}}^*(X;k)$. Section \[sec:open\] is a list of open problems. [**Acknowledgement.**]{} Some of the results of this paper, not least the ones of Section \[sec:global\], are due to Karsten Schmidt. I thank him for a number of communications on his work, culminating in [@Schmidt]. I thank Henning Krause, Andrzej Skowronski, and the referee for comments to a previous version of the paper. I am grateful to Andrzej Skowronski for the very succesful organization of ICRA XII in Torun, August 2007, and for inviting me to submit this paper to the ensuing volume “Trends in Representation Theory of Algebras and Related Topics”. Cochain Differential Graded algebras {#sec:cochain} ==================================== This section provides some results on cochain Differential Graded (DG) algebras, not least on the ones which are Gorenstein. The results first appeared in [@artop], except Lemma \[lem:inf\] which is [@FJcochain lem. 1.5] and Theorem \[thm:k\] which is [@Schmidt cor. 3.12]. For background and terminology on DG algebras and their derived categories, see Appendix \[app:DG\]. \[set:blanket\] In Sections \[sec:cochain\] through \[sec:open\], $k$ is a field and $R$ is a DG algebra over $k$ which has the form $$\cdots \rightarrow 0 \rightarrow k \rightarrow 0 \rightarrow R^2 \rightarrow R^3 \rightarrow \cdots,$$ that is, $R^{<0} = 0$, $R^0 = k$, and $R^1 = 0$. It will be assumed that $\dim_k R < \infty$, and throughout, $$d = \sup R$$ where $\sup$ is as in Definition \[def:sup\_and\_inf\]. Note that either $d = 0$, in which case $R$ is quasi-isomorphic to $k$, or $d \geq 2$. If $X$ is a simply connected topological space with $\dim_k {\operatorname{H}}^*(X;k) < \infty$ and $k$ has characteristic $0$, then ${\operatorname{C}}^*(X;k)$ is quasi-isomorphic to a DG algebra $R$ satisfying the conditions of Setup \[set:blanket\] by [@FHTbook exa. 6, p.  146]. This means that the derived categories of ${\operatorname{C}}^*(X;k)$ and $R$ are equivalent, and hence, all results about the derived category of $R$ also hold for the derived category of ${\operatorname{C}}^*(X;k)$. The highlights of the theory will be made explicit for ${\operatorname{C}}^*(X;k)$ in Section \[sec:topology\]. \[pro:DcinDf\] The full subcategory ${{\mathsf D}^{\operatorname{c}}}(R)$ of compact objects of the derived category ${{\mathsf D}}(R)$ is contained in ${{\mathsf D}^{\operatorname{f}}}(R)$, the full subcategory of ${{\mathsf D}}(R)$ of objects with $\dim_k {\operatorname{H}}\!M < \infty$. The DG module ${}_{R}R$ is in ${{\mathsf D}^{\operatorname{f}}}(R)$ by assumption, and ${{\mathsf D}^{\operatorname{c}}}(R)$ consists of the DG modules which are finitely built from it, cf. Definition \[def:D\], so it follows that ${{\mathsf D}^{\operatorname{c}}}(R)$ is contained in ${{\mathsf D}^{\operatorname{f}}}(R)$. \[pro:DcandDfKrullSchmidt\] The triangulated categories ${{\mathsf D}^{\operatorname{f}}}(R)$ and ${{\mathsf D}^{\operatorname{c}}}(R)$ have finite dimensional Hom spaces and split idempotents. Consequently, ${{\mathsf D}^{\operatorname{f}}}(R)$ and ${{\mathsf D}^{\operatorname{c}}}(R)$ are Krull-Schmidt categories. If $M$ is in ${{\mathsf D}^{\operatorname{f}}}(R)$, then it is finitely built from ${}_{R}k$ in ${{\mathsf D}}(R)$, see Remark \[rmk:Df\]. So to see that ${{\mathsf D}^{\operatorname{f}}}(R)$ has finite dimensional Hom spaces, it is enough to see that ${\operatorname{Hom}}_{{{\mathsf D}}(R)}(\Sigma^i k,k)$ is finite dimensional for each $i$, where $\Sigma$ denotes the suspension functor of ${{\mathsf D}}(R)$. Let $F$ be a minimal semi-free resolution of ${}_{R}(\Sigma^i k)$; then $$\begin{aligned} {\operatorname{Hom}}_{{{\mathsf D}}(R)}(\Sigma^i k,k) & \stackrel{\rm (a)}{\cong} {\operatorname{H}}^0 {\operatorname{RHom}}_R(\Sigma^i k,k) \\ & \stackrel{\rm (b)}{\cong} {\operatorname{H}}^0 {\operatorname{Hom}}_R(F,k) \\ & \stackrel{\rm (c)}{\cong} {\operatorname{Hom}}_{R^{\natural}}(F^{\natural},k^{\natural})^0\\ & = (*)\end{aligned}$$ where (a) is by Definition \[def:Hom\_and\_Tensor\] and (b) and (c) are by Lemma \[lem:semi-free\], (2) and (5). However, Lemma \[lem:semi-free\](3) says that $F^{\natural} \cong \bigoplus_{j \leq i} \Sigma^j(R^{\natural})^{(\beta_j)}$ with the $\beta_j$ finite, where superscript $(\beta)$ indicates the direct sum of $\beta$ copies of the module, and so $$(*) \cong k^{(\beta_0)}.$$ This is finite dimensional. Since ${{\mathsf D}^{\operatorname{c}}}(R)$ is contained in ${{\mathsf D}^{\operatorname{f}}}(R)$ by Proposition \[pro:DcinDf\], it follows that ${{\mathsf D}^{\operatorname{c}}}(R)$ also has finite dimensional Hom spaces. Idempotents split in both ${{\mathsf D}^{\operatorname{f}}}(R)$ and ${{\mathsf D}^{\operatorname{c}}}(R)$ since by [@BN prop. 3.2] they split already in ${{\mathsf D}}(R)$ because this is a triangulated category with set indexed coproducts. By [@Ringel p. 52], both ${{\mathsf D}^{\operatorname{f}}}(R)$ and ${{\mathsf D}^{\operatorname{c}}}(R)$ are Krull-Schmidt categories. Recall from Definition \[def:sup\_and\_inf\] the notion of $\inf$ of a DG module, and from Definition \[def:DG\] that $R^{{\operatorname{o}}}$ is the opposite DG algebra of $R$ and that DG left-$R^{{\operatorname{o}}}$-modules can be viewed as DG right-$R$-modules. \[lem:inf\] Let $M$ be in ${{\mathsf D}^{\operatorname{f}}}(R^{{\operatorname{o}}})$ and let $N$ be in ${{\mathsf D}^{\operatorname{f}}}(R)$. Then $$\inf(M {\stackrel{\operatorname{L}}{\otimes}}_R N) = \inf M + \inf N.$$ If $M$ or $N$ is isomorphic to zero then the equation reads $\infty = \infty$, so let me assume not. Then $i = \inf M$ and $j = \inf N$ are integers. Lemma \[lem:truncations\](1) says that $M$ can be replaced with a quasi-isomorphic DG module which satisfies $M^{\ell} = 0$ for $\ell < i$. Lemma \[lem:semi-free\](3) says that $N$ has a semi-free resolution $F$ which satisfies that $F^{\natural} \cong \bigoplus_{\ell \leq -j} \Sigma^{\ell} (R^{\natural})^{(\beta_{\ell})}$, and it follows that $(M \otimes_R F)^{\natural} \cong \bigoplus_{\ell \leq -j} \Sigma^{\ell} (M^{\natural})^{(\beta_{\ell})}$. Since $M^{\ell} = 0$ for $\ell < i$, this implies that $(M \otimes_R F)^{\ell} = 0$ for $\ell < i+j$. In particular, $\inf(M \otimes_R F) \geq i+j$ whence $$\label{equ:h} \inf(M {\stackrel{\operatorname{L}}{\otimes}}_R N) \geq i+j = \inf M + \inf N.$$ Conversely, to give a morphism of DG left-$R$-modules $\Sigma^{-j}R \rightarrow N$ is the same thing as to give the image $z$ of $\Sigma^{-j}(1_R)$, and $z$ is a cycle in $N^j$. Since ${\operatorname{H}}^j(\Sigma^{-j}R) \cong {\operatorname{H}}^0\!R \cong k$, the induced map ${\operatorname{H}}^j(\Sigma^{-j}R) \rightarrow {\operatorname{H}}^j\!N$ is just the map $k \rightarrow {\operatorname{H}}^j\!N$ sending $1_k$ to the cohomology class of $z$. Hence, picking cycles $z_{\alpha}$ whose cohomology classes form a $k$-basis of ${\operatorname{H}}^j\!N$ and constructing a morphism $\Sigma^{-j}R^{(\beta)} \rightarrow N$ by sending the elements $\Sigma^{-j}(1_R)$ to the $z_{\alpha}$ gives that the induced map ${\operatorname{H}}^j(\Sigma^{-j}R^{(\beta)}) \rightarrow {\operatorname{H}}^j\!N$ is an isomorphism. Complete to a distinguished triangle $$\label{equ:i} \Sigma^{-j}R^{(\beta)} \rightarrow N \rightarrow N^{\prime\prime} \rightarrow;$$ since ${\operatorname{H}}^{j+1}(\Sigma^{-j}R^{(\beta)}) \cong {\operatorname{H}}^1(R^{(\beta)}) = 0$, the long exact cohomology sequence shows $$\label{equ:k} \inf N^{\prime\prime} \geq j+1.$$ Tensoring the distinguished triangle with $M$ gives $$\Sigma^{-j}M^{(\beta)} \rightarrow M {\stackrel{\operatorname{L}}{\otimes}}_R N \rightarrow M {\stackrel{\operatorname{L}}{\otimes}}_R N^{\prime\prime} \rightarrow$$ and the long exact cohomology sequence of this contains $$\label{equ:l} {\operatorname{H}}^{i+j-1}(M {\stackrel{\operatorname{L}}{\otimes}}_R N^{\prime\prime}) \rightarrow {\operatorname{H}}^{i+j}(\Sigma^{-j}M^{(\beta)}) \rightarrow {\operatorname{H}}^{i+j}(M {\stackrel{\operatorname{L}}{\otimes}}_R N).$$ The inequality can be applied to $M$ and $N^{\prime\prime}$; because of the inequality , this gives $\inf(M {\stackrel{\operatorname{L}}{\otimes}}_R N^{\prime\prime}) \geq i+j+1$ so the first term of the exact sequence is zero. The second term is ${\operatorname{H}}^{i+j}(\Sigma^{-j}M^{(\beta)}) \cong {\operatorname{H}}^i(M^{(\beta)})$ which is non-zero since $i = \inf M$. It follows that the third term is non-zero, so $$\inf(M {\stackrel{\operatorname{L}}{\otimes}}_R N) \leq i+j = \inf M + \inf N.$$ Combining with the inequality proves the lemma. The DG algebra $R$ is said to be Gorenstein if it satisfies the equivalent conditions of the following theorem. In the theorem, recall from Definition \[def:dual\] that ${\operatorname{D}}(-) = {\operatorname{Hom}}_k(-,k)$. \[thm:Gorenstein\] The following conditions are equivalent. 1. There are isomorphisms of $k$-vector spaces $${\operatorname{Ext}}_R^i(k,R) \; \cong \; \left\{ \begin{array}{cl} k & \mbox{for $i = d$}, \\ 0 & \mbox{otherwise} \end{array} \right\} \; \cong \; {\operatorname{Ext}}_{R^{{\operatorname{o}}}}^i(k,R).$$ 2. There are isomorphisms of graded ${\operatorname{H}}\!R$-modules $$\mbox{${}_{{\operatorname{H}}\!R}({\operatorname{D}}\!{\operatorname{H}}\!R) \cong {}_{{\operatorname{H}}\!R}(\Sigma^d {\operatorname{H}}\!R)$ \;\;and\;\; $({\operatorname{D}}\!{\operatorname{H}}\!R)_{{\operatorname{H}}\!R} \cong (\Sigma^d {\operatorname{H}}\!R)_{{\operatorname{H}}\!R}$.}$$ 3. There are isomorphisms $$\mbox{${}_{R}({\operatorname{D}}\!R) \cong {}_{R}(\Sigma^d R)$ \;in\; ${{\mathsf D}}(R)$ \;\;and\;\; $({\operatorname{D}}\!R)_R \cong (\Sigma^d R)_R$ \;in\; ${{\mathsf D}}(R^{{\operatorname{o}}})$.}$$ 4. $\dim_k {\operatorname{Ext}}_R(k,R) < \infty$ and $\dim_k {\operatorname{Ext}}_{R^{{\operatorname{o}}}}(k,R) < \infty$. 5. ${}_{R}({\operatorname{D}}\!R)$ is in ${{\mathsf D}^{\operatorname{c}}}(R)$ and $({\operatorname{D}}\!R)_R$ is in ${{\mathsf D}^{\operatorname{c}}}(R^{{\operatorname{o}}})$. (1)$\Rightarrow$(3). Let $F$ be a minimal semi-free resolution of ${}_{R}({\operatorname{D}}\!R)$. Then $$\label{equ:d} {\operatorname{Ext}}_{R^{{\operatorname{o}}}}(k,R) \stackrel{\rm (a)}{\cong} {\operatorname{Ext}}_R({\operatorname{D}}\!R,k) = {\operatorname{H}}({\operatorname{RHom}}_R({\operatorname{D}}\!R,k)) \stackrel{\rm (b)}{\cong} {\operatorname{Hom}}_{R^{\natural}}(F^{\natural},k^{\natural})$$ where (a) is by duality and (b) is by Lemma \[lem:semi-free\], (2) and (5). If the second isomorphism in (1) holds, then this implies $F^{\natural} \cong \Sigma^d R^{\natural}$. But then there is clearly only a single step in the semi-free filtration of $F$, whence $F \cong {}_{R}(\Sigma^d R)$ so ${}_{R}({\operatorname{D}}\!R) \cong {}_{R}(\Sigma^d R)$, proving the first isomorphism in (3). Likewise, the first isomorphism in (1) implies the second isomorphism in (3). (3)$\Rightarrow$(2). This follows by taking cohomology. (2)$\Rightarrow$(1). This follows from the Eilenberg-Moore spectral sequence $$E_2^{pq} = {\operatorname{Ext}}_{{\operatorname{H}}\!R}^p(k,{\operatorname{H}}\!R)^q \Rightarrow {\operatorname{Ext}}_R^{p+q}(k,R)$$ which exists by [@FHTpaper 1.3(2)], and the corresponding spectral sequence over $R^{{\operatorname{o}}}$. (4)$\Leftrightarrow$(5). Lemma \[lem:semi-free\](3) says that the semi-free resolution $F$ of ${}_{R}({\operatorname{D}}\!R)$ has $F^{\natural} \cong \bigoplus_i \Sigma^i(R^{\natural})^{(\beta_i)}$, and Equation shows that $\dim_k {\operatorname{Ext}}_{R^{{\operatorname{o}}}}(k,R)$ is the number of direct summands $\Sigma^i R^{\natural}$. By Lemma \[lem:semi-free\](4), this number is finite if and only if ${}_{R}({\operatorname{D}}\!R)$ is in ${{\mathsf D}^{\operatorname{c}}}(R)$, so the second condition in (4) is equivalent to the first condition in (5) and vice versa. (1)$\Rightarrow$(4) is clear. (4)$\Rightarrow$(1). When (4) holds, so does (5) by the previous part of the proof; hence ${}_{R}({\operatorname{D}}\!R)$ is finitely built from ${}_{R}R$. Then the canonical morphism $${\operatorname{RHom}}_R(k,R) {\stackrel{\operatorname{L}}{\otimes}}_R {\operatorname{D}}\!R \rightarrow {\operatorname{RHom}}_R(k,R {\stackrel{\operatorname{L}}{\otimes}}_R {\operatorname{D}}\!R)$$ is an isomorphism, because it clearly is if ${\operatorname{D}}\!R$ is replaced with $R$. That is, $$\label{equ:c} {\operatorname{RHom}}_R(k,R) {\stackrel{\operatorname{L}}{\otimes}}_R {\operatorname{D}}\!R \cong k.$$ Since (4) holds, ${\operatorname{RHom}}_R(k,R)$ is in ${{\mathsf D}^{\operatorname{f}}}(R^{{\operatorname{o}}})$, so Lemma \[lem:inf\] applies to the tensor product and gives $$\inf {\operatorname{RHom}}_R(k,R) + \inf {\operatorname{D}}\!R = \inf k = 0$$ which amounts to $$\inf {\operatorname{RHom}}_R(k,R) = d.$$ On the other hand, adjointness gives the first of the next isomorphisms, $${\operatorname{RHom}}_k(({\operatorname{D}}\!R) {\stackrel{\operatorname{L}}{\otimes}}_R k,k) \cong {\operatorname{RHom}}_R(k,{\operatorname{RHom}}_k({\operatorname{D}}\!R,k)) \cong {\operatorname{RHom}}_R(k,R),$$ and so $$\begin{aligned} \sup {\operatorname{RHom}}_R(k,R) & = \sup {\operatorname{RHom}}_k(({\operatorname{D}}\!R) {\stackrel{\operatorname{L}}{\otimes}}_R k,k) \\ & = - \inf(({\operatorname{D}}\!R) {\stackrel{\operatorname{L}}{\otimes}}_R k) \\ & \stackrel{\rm (c)}{=} - \inf {\operatorname{D}}\!R - \inf k \\ & = d\end{aligned}$$ where (c) is by Lemma \[lem:inf\] again. Hence the cohomology of ${\operatorname{RHom}}_R(k,R)$ is concentrated in degree $d$, and it is not hard to show that hence $${\operatorname{RHom}}_R(k,R) \cong (\Sigma^{-d}k^{(\beta)})_R$$ for some $\beta$. Inserting this into Equation shows $\beta = 1$, so $${\operatorname{RHom}}_R(k,R) \cong (\Sigma^{-d}k)_R.$$ This is equivalent to the first isomorphism in (1), and the second one follows by a symmetric argument. \[thm:k\] If $\dim_k {\operatorname{H}}\!R \geq 2$, then ${}_{R}k$ is not in ${{\mathsf D}^{\operatorname{c}}}(R)$. Recall from Definition \[def:sup\_and\_inf\] the notion of amplitude of a DG module. There is an amplitude inequality ${\operatorname{amp}}(M {\stackrel{\operatorname{L}}{\otimes}}_R N) \geq {\operatorname{amp}}M$ for $M$ in ${{\mathsf D}^{\operatorname{f}}}(R^{{\operatorname{o}}})$ and $N$ in ${{\mathsf D}^{\operatorname{c}}}(R)$. This was first stated in [@Schmidt prop. 3.11]; see [@FJcochain cor. 4.4] for an alternative proof. If ${}_{R}k$ were in ${{\mathsf D}^{\operatorname{c}}}(R)$, then this would give ${\operatorname{amp}}(R {\stackrel{\operatorname{L}}{\otimes}}_R k) \geq {\operatorname{amp}}R$, that is, $0 \geq {\operatorname{amp}}R$ contradicting $\dim_k {\operatorname{H}}\!R \geq 2$ whereby $R$ must (also) have cohomology in a degree different from $0$. Auslander-Reiten triangles over Differential\ Graded algebras {#sec:AR} ============================================= In this section, it is proved that the compact derived category ${{\mathsf D}^{\operatorname{c}}}(R)$ has Auslander-Reiten (AR) triangles if and only if $R$ is a Gorenstein DG algebra. In this case, a formula is found for the AR translation of ${{\mathsf D}^{\operatorname{c}}}(R)$. These results first appeared in [@artop]. For background on AR theory, see Appendix \[app:AR\]. In the following proposition, note that ${\operatorname{D}}\!R {\stackrel{\operatorname{L}}{\otimes}}_R P$ inherits a left-$R$-structure from the DG bi-$R$-module ${\operatorname{D}}\!R$ so ${\operatorname{D}}\!R {\stackrel{\operatorname{L}}{\otimes}}_R P$ is in ${{\mathsf D}}(R)$; see Definition \[def:Hom\_and\_Tensor\]. \[pro:AR\] Let $P$ be an indecomposable object of ${{\mathsf D}^{\operatorname{c}}}(R)$. There is an AR triangle in ${{\mathsf D}^{\operatorname{f}}}(R)$, $$\Sigma^{-1}({\operatorname{D}}\!R {\stackrel{\operatorname{L}}{\otimes}}_R P) \rightarrow N \rightarrow P \rightarrow.$$ Since $P$ is finitely built from ${}_{R}R$, there is a natural equivalence $$\label{equ:Serre} {\operatorname{D}}({\operatorname{Hom}}_{{{\mathsf D}}(R)}(P,-)) \simeq {\operatorname{Hom}}_{{{\mathsf D}}(R)}(-,{\operatorname{D}}\!R {\stackrel{\operatorname{L}}{\otimes}}_R P),$$ since there is clearly such an equivalence if $P$ is replaced with ${}_{R}R$. By [@KrauseARZ prop. 4.2], this means that the AR triangle of the present proposition exists in ${{\mathsf D}}(R)$. To complete the proof, observe that the triangle is in fact in ${{\mathsf D}^{\operatorname{f}}}(R)$: The object $P$ is in ${{\mathsf D}^{\operatorname{c}}}(R)$, so it is in ${{\mathsf D}^{\operatorname{f}}}(R)$ by Proposition \[pro:DcinDf\]. Since $R_R$ is in ${{\mathsf D}^{\operatorname{f}}}(R^{{\operatorname{o}}})$, the dual ${}_{R}({\operatorname{D}}\!R)$ is in ${{\mathsf D}^{\operatorname{f}}}(R)$, and since $P$ is finitely built from $R$, it follows that ${\operatorname{D}}\!R {\stackrel{\operatorname{L}}{\otimes}}_R P$ is also in ${{\mathsf D}^{\operatorname{f}}}(R)$. Finally, $N$ is in ${{\mathsf D}^{\operatorname{f}}}(R)$ by the long exact cohomology sequence. \[pro:AR\_triangles\_preserved\] An AR triangle in ${{\mathsf D}^{\operatorname{c}}}(R)$ is also an AR triangle in ${{\mathsf D}^{\operatorname{f}}}(R)$. By [@KrauseARZ lem. 4.3], each object in ${{\mathsf D}^{\operatorname{c}}}(R)$ is a pure injective object of ${{\mathsf D}}(R)$. Hence by [@KrauseARZ prop.3.2], each AR triangle in ${{\mathsf D}^{\operatorname{c}}}(R)$ is an AR triangle in ${{\mathsf D}}(R)$, and in particular in ${{\mathsf D}^{\operatorname{f}}}(R)$. \[pro:AR2\] 1. ${{\mathsf D}^{\operatorname{c}}}(R)$ has right AR triangles if and only if ${}_{R}({\operatorname{D}}\!R)$ is in ${{\mathsf D}^{\operatorname{c}}}(R)$. 2. ${{\mathsf D}^{\operatorname{c}}}(R)$ has left AR triangles if and only if $({\operatorname{D}}\!R)_R$ is in ${{\mathsf D}^{\operatorname{c}}}(R^{{\operatorname{o}}})$. (1). Suppose that ${{\mathsf D}^{\operatorname{c}}}(R)$ has right AR triangles. The object ${}_{R}R$ of ${{\mathsf D}^{\operatorname{c}}}(R)$ has endomorphism ring $k$ which is local, so there is an AR triangle $M \rightarrow N \rightarrow {}_{R}R \rightarrow$ in ${{\mathsf D}^{\operatorname{c}}}(R)$. By Proposition \[pro:AR\_triangles\_preserved\], it is even an AR triangle in ${{\mathsf D}^{\operatorname{f}}}(R)$. On the other hand, Proposition \[pro:AR\] gives that there is also an AR triangle $\Sigma^{-1}({}_{R}({\operatorname{D}}\!R)) \rightarrow N^{\prime} \rightarrow {}_{R}R \rightarrow$ in ${{\mathsf D}^{\operatorname{f}}}(R)$, and since the right hand terms of the two AR triangles are isomorphic, so are the left hand terms, $M \cong \Sigma^{-1}({}_{R}({\operatorname{D}}\!R))$. But $M$ is in ${{\mathsf D}^{\operatorname{c}}}(R)$, so it follows that $\Sigma^{-1}({}_{R}({\operatorname{D}}\!R))$ and hence ${}_{R}({\operatorname{D}}\!R)$ is in ${{\mathsf D}^{\operatorname{c}}}(R)$. Conversely, suppose that ${}_{R}({\operatorname{D}}\!R)$ is in ${{\mathsf D}^{\operatorname{c}}}(R)$. Given $P$ in ${{\mathsf D}^{\operatorname{c}}}(R)$, Proposition \[pro:AR\] gives an AR triangle $$\Sigma^{-1}({\operatorname{D}}\!R {\stackrel{\operatorname{L}}{\otimes}}_R P) \rightarrow N \rightarrow P \rightarrow$$ in ${{\mathsf D}^{\operatorname{f}}}(R)$. Since ${}_{R}({\operatorname{D}}\!R)$ is in ${{\mathsf D}^{\operatorname{c}}}(R)$, it is finitely built from ${}_{R}R$. The same is true for $P$, and so ${\operatorname{D}}\!R {\stackrel{\operatorname{L}}{\otimes}}_R P$ is also finitely built from ${}_{R}R$, that is, it is in ${{\mathsf D}^{\operatorname{c}}}(R)$. It follows that both outer terms of the AR triangle are in ${{\mathsf D}^{\operatorname{c}}}(R)$, and then so is $N$. That is, the AR triangle is in ${{\mathsf D}^{\operatorname{c}}}(R)$, so it is an AR triangle in that category. (2). The functors ${\operatorname{RHom}}_R(-,R)$ and ${\operatorname{RHom}}_{R^{{\operatorname{o}}}}(-,R)$ are quasi-inverse dualities between ${{\mathsf D}^{\operatorname{c}}}(R)$ and ${{\mathsf D}^{\operatorname{c}}}(R^{{\operatorname{o}}})$, so ${{\mathsf D}^{\operatorname{c}}}(R)$ has left AR triangles if and only if ${{\mathsf D}^{\operatorname{c}}}(R^{{\operatorname{o}}})$ has right AR triangles. By the right module version of part (1), this happens if and only if $({\operatorname{D}}\!R)_R$ is in ${{\mathsf D}^{\operatorname{c}}}(R^{{\operatorname{o}}})$. \[thm:R\] The following conditions are equivalent. 1. ${{\mathsf D}^{\operatorname{c}}}(R)$ has AR triangles. 2. ${{\mathsf D}^{\operatorname{c}}}(R^{{\operatorname{o}}})$ has AR triangles. 3. $R$ is Gorenstein. By Theorem \[thm:Gorenstein\](5), condition (3) is equivalent to having that ${}_{R}({\operatorname{D}}\!R)$ is in ${{\mathsf D}^{\operatorname{c}}}(R)$ and $({\operatorname{D}}\!R)_R$ is in ${{\mathsf D}^{\operatorname{c}}}(R^{{\operatorname{o}}})$. This is equivalent to condition (1) by Proposition \[pro:AR2\], and it is equivalent to condition (2) by the right module version of Proposition \[pro:AR2\]. \[rmk:tau\] Assume the situation of Theorem \[thm:R\]. Since ${{\mathsf D}^{\operatorname{c}}}(R)$ has AR triangles, [@KrauseARZ thm. 4.4] and Equation imply that $$\label{equ:Serre2} S(-) = {\operatorname{D}}\!R {\stackrel{\operatorname{L}}{\otimes}}_R -$$ is a Serre functor of ${{\mathsf D}^{\operatorname{c}}}(R)$, cf. Definition \[def:Serre\]. So the AR translation $\tau$ of ${{\mathsf D}^{\operatorname{c}}}(R)$ extends to the autoequivalence $$\label{equ:m} \Sigma^{-1}({\operatorname{D}}\!R {\stackrel{\operatorname{L}}{\otimes}}_R -)$$ of ${{\mathsf D}^{\operatorname{c}}}(R)$, cf. Theorem \[thm:Serre\]. A quasi-inverse equivalence is $$\Sigma {\operatorname{RHom}}_{R^{{\operatorname{o}}}}({\operatorname{D}}\!R,R) {\stackrel{\operatorname{L}}{\otimes}}_R -;$$ these two expressions can also be viewed as quasi-inverse autoequivalences of ${{\mathsf D}}(R)$. If $X$ is an indecomposable object of ${{\mathsf D}^{\operatorname{c}}}(R)$ then there are AR triangles in ${{\mathsf D}^{\operatorname{c}}}(R)$, $$\Sigma^{-1}({\operatorname{D}}\!R {\stackrel{\operatorname{L}}{\otimes}}_R X) \rightarrow Y \rightarrow X \rightarrow$$ and $$X \rightarrow Y^{\prime} \rightarrow \Sigma {\operatorname{RHom}}_{R^{{\operatorname{o}}}}({\operatorname{D}}\!R,R) {\stackrel{\operatorname{L}}{\otimes}}_R X \rightarrow.$$ Combining Equation with Theorem \[thm:Gorenstein\](3) which says $({\operatorname{D}}\!R)_R \cong (\Sigma^d R)_R$ gives $$\label{equ:j} {\operatorname{H}}(\tau(-)) \cong {\operatorname{H}}(\Sigma^{d-1}(-))$$ as graded $k$-vector spaces. The Auslander-Reiten quiver of a Differential Graded algebra: Local structure {#sec:local} ============================================================================= This section considers the AR quiver $\Gamma$ of the compact derived category ${{\mathsf D}^{\operatorname{c}}}(R)$. When $R$ is Gorenstein with $\dim_k {\operatorname{H}}\!R \geq 2$, it is proved that each component of $\Gamma$ is isomorphic to ${{\mathbb Z}}A_{\infty}$ as a translation quiver. The results first appeared in [@arquiv]; the methods of Karsten Schmidt [@Schmidt] have permitted some technical assumptions to be removed. \[set:local\] In this section, $R$ will be Gorenstein with $\dim_k {\operatorname{H}}\!R \geq 2$. The category ${{\mathsf D}^{\operatorname{c}}}(R)$ has AR triangles by Theorem \[thm:R\], and ${}_{R}k$ is not in ${{\mathsf D}^{\operatorname{c}}}(R)$ by Theorem \[thm:k\]. The AR quiver $\Gamma({{\mathsf D}^{\operatorname{c}}}(R))$ will be abbreviated to $\Gamma$. Then $\Gamma$ with the AR translation $\tau$ is a stable translation quiver by Proposition \[pro:stable\_translation\_quiver\]. By $C$ will be denoted a component of the translation quiver $\Gamma$. \[lem:no\_loops\] 1. No positive power $\tau^p$ of the AR translation $\tau$ has a fixed point in $\Gamma$. 2. $\Gamma$ has no loops. (1). Remark \[rmk:tau\] says $\tau(M) = \Sigma^{-1}({\operatorname{D}}\!R {\stackrel{\operatorname{L}}{\otimes}}_R M)$. Lemma \[lem:inf\] implies $$\inf \tau(M) = 1 + \inf {\operatorname{D}}\!R + \inf M = 1 - d +\inf M.$$ Since $d$ is either $0$ or $\geq 2$, it follows that each positive power $\tau^p(M)$ has $\inf$ different from $\inf M$, so no positive power is isomorphic to $M$. (2). The existence of a loop $[M] \rightarrow [M]$ would mean the existence of an irreducible morphism $M \rightarrow M$ in ${{\mathsf D}^{\operatorname{c}}}(R)$. Such a morphism would be in the radical of the finite dimensional algebra ${\operatorname{Hom}}_{{{\mathsf D}^{\operatorname{c}}}(R)}(M,M)$, and hence some power would be zero. Mimicking the proof of [@ARS lem. VII.2.5] now shows $\tau(M) = M$, but this contradicts part (1). A reference for the graph theoretical terminology of the following proposition is [@BensonI sec. 4.15]. A salient fact is that when $T$ is a directed tree, then the vertices of the repetitive quiver ${{\mathbb Z}}T$ have the form $(p,t)$ where $p$ is an integer, $t$ is a vertex of $T$. The translation of the stable translation quiver ${{\mathbb Z}}T$ is determined by $\tau(p,t) = (p+1,t)$. There exist a directed tree $T$ and an admissible group of automorphisms $\Pi$ of ${{\mathbb Z}}T$ so that $C \cong {{\mathbb Z}}T/\Pi$ as stable translation quivers. Since $\tau$ extends to an autoequivalence of ${{\mathsf D}^{\operatorname{c}}}(R)$ by Remark \[rmk:tau\], the AR translation is an automorphism of $\Gamma$ so restricts to an automorphism of $C$. By definition, $C$ has no multiple arrows, and by Lemma \[lem:no\_loops\](2), it has no loops. Hence the proposition follows from the Riedtmann Structure Theorem, see [@BensonI thm. 4.15.6]. To show that $T = A_{\infty}$ and that $\Pi$ acts trivially, the following definitions are useful. \[def:varphi\] Define a function on the objects of ${{\mathsf D}}(R)$ by $$\varphi(M) = \dim_k {\operatorname{Ext}}_R(M,k).$$ By abuse of notation, the induced function on the vertices of the AR quiver $\Gamma$ is also denoted by $\varphi$. Label the AR quiver $\Gamma$ by assigning to the arrow $[M] \stackrel{\mu}{\rightarrow} [N]$ the label $(\alpha_{\mu},\beta_{\mu})$, where $\alpha_{\mu}$ is the multiplicity of $M$ as a direct summand of $Y$ in the AR triangle $$\tau N \rightarrow Y \rightarrow N \rightarrow$$ and $\beta_{\mu}$ is the multiplicity of $N$ as a direct summand of $X$ in the AR triangle $$M \rightarrow X \rightarrow \tau^{-1}M \rightarrow.$$ The vertices of ${{\mathbb Z}}T$ have the form $(p,t)$ where $p$ is an integer, $t$ a vertex of $T$, so each vertex $t$ of $T$ gives a vertex $(0,t)$ of ${{\mathbb Z}}T$ and hence a vertex $\Pi(0,t)$ of ${{\mathbb Z}}T/\Pi$, that is, of $C$. Similarly, an arrow $t \rightarrow t^{\prime}$ in $T$ gives an arrow $\Pi(0,t) \rightarrow \Pi(0,t^{\prime})$ of $C$. Hence the function $\varphi$ and the labelling $(\alpha,\beta)$ on $\Gamma$ induce a function and a labelling on $T$. These will be denoted by $f$ and $(a,b)$. \[lem:varphi\] The function $\varphi$ and the labelling $(\alpha,\beta)$ have the following properties. 1. If $F$ is a minimal semi-free resolution of $M$ with $F^{\natural} \cong \bigoplus_i \Sigma^i (R^{\natural})^{(\beta_i)}$, then $\varphi(M)$ is equal to the number of direct summands $\Sigma^i R^{\natural}$ in $F^{\natural}$. 2. $\varphi(\tau N) = \varphi(N)$. 3. If $\tau N \rightarrow Y \rightarrow N \rightarrow$ is an AR triangle in ${{\mathsf D}^{\operatorname{c}}}(R)$, then $\varphi(Y) = \varphi(\tau N) + \varphi(N)$. 4. If there is an arrow $[M] \stackrel{\mu}{\rightarrow} [N]$ in $\Gamma$ then there is a corresponding arrow $\tau[N] \stackrel{\nu}{\rightarrow} [M]$, and $(\alpha_{\nu},\beta_{\nu}) = (\beta_{\mu},\alpha_{\mu})$. 5. If there is an arrow $[M] \stackrel{\mu}{\rightarrow} [N]$ in $\Gamma$ then there is also an arrow $\tau[M] \stackrel{\tau(\mu)}{\rightarrow} \tau[N]$, and $(\alpha_{\tau(\mu)},\beta_{\tau(\mu)}) = (\alpha_{\mu},\beta_{\mu})$. 6. $\sum_{\mu : [M] \rightarrow [N]}\alpha_{\mu}\varphi(M) = \varphi(\tau N) + \varphi(N)$, where the sum is over all arrows in $\Gamma$ with target $[N]$. (1). It holds that $$\varphi(M) = \dim_k {\operatorname{H}}({\operatorname{RHom}}_R(M,k)) \stackrel{\rm (c)}{=} \dim_k {\operatorname{H}}({\operatorname{Hom}}_R(F,k)) \stackrel{\rm (d)}{=} \dim_k {\operatorname{Hom}}_{R^{\natural}}(F^{\natural},k^{\natural}),$$ where (c) and (d) are by Lemma \[lem:semi-free\], parts (2) and (5). The right hand side is clearly equal to the number of direct summands $\Sigma^i R^{\natural}$ in $F^{\natural}$. (2). It holds that $$\begin{aligned} \varphi(\tau N) & \stackrel{\rm (a)}{=} \dim_k {\operatorname{H}}({\operatorname{RHom}}_R(\Sigma^{-1}({\operatorname{D}}\!R {\stackrel{\operatorname{L}}{\otimes}}_R N),k)) \\ & = \dim_k {\operatorname{H}}({\operatorname{RHom}}_R(N,\Sigma {\operatorname{RHom}}_R({\operatorname{D}}\!R,k))) \\ & \stackrel{\rm (b)}{=} \dim_k {\operatorname{H}}({\operatorname{RHom}}_R(N,\Sigma^{1-d}k)) \\ & = \dim_k {\operatorname{H}}({\operatorname{RHom}}_R(N,k)) \\ & = \varphi(N),\end{aligned}$$ where (a) is by Remark \[rmk:tau\] and (b) follows from Theorem \[thm:Gorenstein\](3). (3). The AR triangle of the lemma induces a long exact sequence consisting of pieces $${\operatorname{Ext}}_R^i(N,k) \rightarrow {\operatorname{Ext}}_R^i(Y,k) \rightarrow {\operatorname{Ext}}_R^i(\tau N,k),$$ and the claim will follow if the connecting maps are zero. Indeed, the AR triangle is also an AR triangle in ${{\mathsf D}^{\operatorname{f}}}(R)$ by Proposition \[pro:AR\_triangles\_preserved\]. A morphism $\tau N \rightarrow {}_{R}(\Sigma^i k)$ in ${{\mathsf D}^{\operatorname{f}}}(R)$ cannot be a split monomorphism since $\tau N$ is in ${{\mathsf D}^{\operatorname{c}}}(R)$ while ${}_{R}(\Sigma^i k)$ is not, cf. Setup \[set:local\]. It follows that each such morphism factors through $\tau N \rightarrow Y$ whence the composition $\Sigma^{-1}N \rightarrow \tau N \rightarrow \Sigma^i k$ is zero. Hence the connecting morphism ${\operatorname{Ext}}_R^i(\tau N,k) \rightarrow {\operatorname{Ext}}_R^{i+1}(N,k)$ is zero as desired. (4). Let $$\label{equ:e} \tau N \rightarrow Y \rightarrow N \rightarrow$$ be an AR triangle in ${{\mathsf D}^{\operatorname{c}}}(R)$. By the definition of the labelling of $\Gamma$, the multiplicity of $M$ as a direct summand of $Y$ is equal to both $\beta_{\nu}$ and $\alpha_{\mu}$, so $\beta_{\nu} = \alpha_{\mu}$. A similar argument shows $\alpha_{\nu} = \beta_{\mu}$, so $(\alpha_{\nu},\beta_{\nu}) = (\beta_{\mu},\alpha_{\mu})$. (5). This holds since the AR translation $\tau$ of ${{\mathsf D}^{\operatorname{c}}}(R)$ is the restriction of an equivalence of categories by Remark \[rmk:tau\]. (6). Consider the AR triangle . The object $Y$ is a direct sum of copies of the indecomposable objects of ${{\mathsf D}^{\operatorname{c}}}(R)$ which have irreducible morphisms to $N$, and the multiplicity of $M$ as a direct summand of $Y$ is $\alpha_{\mu}$ where $[M] \stackrel{\mu}{\rightarrow} [N]$ is the arrow in $\Gamma$. Hence $$\sum_{\mu:[M] \rightarrow [N]} \alpha_{\mu}\varphi(M) = \varphi(Y).$$ Now combine with part (3). \[lem:zero\] Let $M \langle 0 \rangle, \ldots, M \langle 2^p - 1 \rangle$ be indecomposable objects of ${{\mathsf D}^{\operatorname{c}}}(R)$ with $\varphi(M \langle i \rangle) \leq \frac{p}{\dim_k R}$ for each $i$. If $$M \langle 2^p - 1 \rangle \rightarrow M \langle 2^p - 2 \rangle \rightarrow \cdots \rightarrow M \langle 0 \rangle$$ are non-isomorphisms in ${{\mathsf D}^{\operatorname{c}}}(R)$, then the composition is zero. Let $F \langle i \rangle$ be a minimal semi-free resolution of $M \langle i \rangle$. Each $F \langle i \rangle$ must be indecomposable as a DG left-$R$-module, for if $F \langle i \rangle$ decomposed then it would do so into DG modules $F \langle i_{\alpha} \rangle$ with $\partial(F \langle i_{\alpha} \rangle) \subseteq R^{\geq 1} \cdot F \langle i_{\alpha} \rangle$, but this condition forces non-zero cohomology so the decomposition of $F \langle i \rangle$ as a DG module would induce a non-trivial decomposition of $M \langle i \rangle$ in ${{\mathsf D}^{\operatorname{c}}}(R)$. The morphisms in ${{\mathsf D}^{\operatorname{c}}}(R)$ between the $M \langle i \rangle$ are represented by morphisms $$\label{equ:f} F \langle 2^p - 1 \rangle \rightarrow F \langle 2^p - 2 \rangle \rightarrow \cdots \rightarrow F \langle 0 \rangle$$ of DG left-$R$-modules. These cannot be bijections, since if they were, then the morphisms in ${{\mathsf D}^{\operatorname{c}}}(R)$ between the $M \langle i \rangle$ would be isomorphisms. Now note that if $F \langle i \rangle^{\natural} = \bigoplus_j \Sigma^j(R^{\natural})^{(\beta_j)}$, then the direct sum has $\varphi(M \langle i \rangle)$ summands $\Sigma^j R^{\natural}$ by Lemma \[lem:varphi\](1). Hence $$\dim_k F \langle i \rangle = \varphi(M \langle i \rangle) \dim_k R \leq \frac{p}{\dim_k R} \dim_k R = p,$$ and it is not hard to mimick the proof of [@BensonI lem. 4.14.1] to see that hence, the composition of the morphisms in Equation is zero. This implies that the composition of the morphisms in the lemma is zero. \[lem:non-zero\] If $M \langle 0 \rangle$ is an indecomposable object of ${{\mathsf D}^{\operatorname{c}}}(R)$ and $q \geq 0$ is an integer, then there exist indecomposable objects and irreducible morphisms in ${{\mathsf D}^{\operatorname{c}}}(R)$, $$M \langle q \rangle \rightarrow M \langle q-1 \rangle \rightarrow \cdots \rightarrow M \langle 0 \rangle,$$ with non-zero composition. Let me prove a stronger statement which implies the lemma: If $M \langle 0 \rangle$ is an indecomposable object of ${{\mathsf D}^{\operatorname{c}}}(R)$ and $q \geq 0$ is an integer, then there exists $${}_{R}(\Sigma^i k) \stackrel{\kappa_q}{\rightarrow} M \langle q \rangle \stackrel{\mu_q}{\rightarrow} M \langle q-1 \rangle \stackrel{\mu_{q-1}}{\rightarrow} \cdots \stackrel{\mu_1}{\rightarrow} M \langle 0 \rangle$$ where the $M \langle i \rangle$ are indecomposable objects of ${{\mathsf D}^{\operatorname{c}}}(R)$ and the $\mu_i$ are irreducible morphisms in ${{\mathsf D}^{\operatorname{c}}}(R)$, such that $\mu_1 \circ \cdots \circ \mu_q \circ \kappa_q \neq 0$. Using induction on $q$, first let $q = 0$. Let $F$ be a minimal semi-free resolution of the dual ${\operatorname{D}}\!M \langle 0 \rangle$. Then $$\begin{aligned} {\operatorname{H}}({\operatorname{RHom}}_R(k,M \langle 0 \rangle)) & \cong {\operatorname{H}}({\operatorname{RHom}}_{R^{{\operatorname{o}}}}({\operatorname{D}}\!M \langle 0 \rangle,k)) \\ & \stackrel{\rm (a)}{\cong} {\operatorname{H}}({\operatorname{Hom}}_{R^{{\operatorname{o}}}}(F,k)) \\ & \stackrel{\rm (b)}{\cong} {\operatorname{Hom}}_{(R^{{\operatorname{o}}})^{\natural}}(F^{\natural},k^{\natural}) \\ & \stackrel{\rm (c)}{\not\cong} 0.\end{aligned}$$ Here (a) and (b) are by Lemma \[lem:semi-free\], parts (2) and (5). (c) is because $M \langle 0 \rangle$ is indecomposable hence has non-zero cohomology; this implies that ${\operatorname{D}}\!M \langle 0 \rangle$ has non-zero cohomology, and then $F$ is non-trivial semi-free whence $F^{\natural}$ is a non-trivial graded free module. It follows from the displayed formula that there is a non-zero morphism $${}_{R}(\Sigma^i k) \stackrel{\kappa_0}{\rightarrow} M \langle 0 \rangle$$ for some $i$. Now let $q \geq 1$ and suppose that $${}_{R}(\Sigma^i k) \stackrel{\kappa_{q-1}}{\rightarrow} M \langle q-1 \rangle \stackrel{\mu_{q-1}}{\rightarrow} M \langle q-2 \rangle \stackrel{\mu_{q-2}}{\rightarrow} \cdots \stackrel{\mu_1}{\rightarrow} M \langle 0 \rangle$$ has already been found with the desired properties. Let $\tau M \langle q-1 \rangle \rightarrow X \langle q \rangle \stackrel{\mu_{q}^{\prime}}{\rightarrow} M \langle q-1 \rangle \rightarrow$ be an AR triangle in ${{\mathsf D}^{\operatorname{c}}}(R)$. By Proposition \[pro:AR\_triangles\_preserved\] it is also an AR triangle in ${{\mathsf D}^{\operatorname{f}}}(R)$. Since ${}_{R}k$ is not in ${{\mathsf D}^{\operatorname{c}}}(R)$, see Setup \[set:local\], it is clear that $\kappa_{q-1}$ is not a split epimorphism, so it factors through $\mu_{q}^{\prime}$. Now I can get the situation claimed in the lemma by letting $M \langle q \rangle$ be a suitable indecomposable summand of $X \langle q \rangle$ and $\mu_q$ the restriction of $\mu^{\prime}_q$ to $M \langle q \rangle$. \[lem:varphi\_unbounded\] The function $\varphi$ is unbounded on $C$. If $\varphi$ were bounded on $C$ then Lemma \[lem:zero\] would apply to sufficiently long sequences of morphisms between indecomposable objects with vertices in $C$, but this would make impossible the situation established in Lemma \[lem:non-zero\]. Recall that the Cartan matrix $c$ of the labelled directed tree $T$ is a matrix with rows and columns indexed by the vertices of $T$. If $s$ and $t$ are vertices, then $$c_{st} = \left\{ \begin{array}{cl} 2 & \mbox{if $s=t$}, \\ -a_{\mu} & \mbox{if there is an arrow $s \stackrel{\mu}{\rightarrow} t$}, \\ -b_{\nu} & \mbox{if there is an arrow $t \stackrel{\nu}{\rightarrow} s$}, \\ 0 & \mbox{if $s \neq t$ and $s$ and $t$ are not connected by an arrow}; \end{array} \right.$$ cp. [@BensonI sec. 4.5]. The function $f$ on the vertices of $T$ is called additive if it satisfies $\sum_s c_{st}f(s) = 0$ for each $t$, that is, $$\label{equ:g} 2f(t) - \sum_{\mu : s \rightarrow t}a_{\mu}f(s) - \sum_{\nu : t \rightarrow u}b_{\nu}f(u) = 0$$ for each $t$, where the sums are over all arrows in $T$ into $t$ and out of $t$. Indeed: \[pro:f\] The function $f$ is additive and unbounded on $T$. Using Definition \[def:varphi\], the left hand side of Equation can be rewritten $$2\varphi(\Pi(0,t)) - \sum_{\mu : s \rightarrow t} \alpha_{\Pi(0,s) \rightarrow \Pi(0,t)} \varphi(\Pi(0,s)) - \sum_{\nu : t \rightarrow u} \beta_{\Pi(0,t) \rightarrow \Pi(0,u)} \varphi(\Pi(0,u)).$$ The translation of ${{\mathbb Z}}T/\Pi$ is given by $\tau(\Pi(p,t)) = \Pi(p+1,t)$. To each arrow $\Pi(0,t) \rightarrow \Pi(0,u)$ corresponds an arrow $\tau(\Pi(0,u)) \rightarrow \Pi(0,t)$, that is, $\Pi(1,u) \rightarrow \Pi(0,t)$. Lemma \[lem:varphi\](4) gives $\beta_{\Pi(0,t) \rightarrow \Pi(0,u)} = \alpha_{\Pi(1,u) \rightarrow \Pi(0,t)}$. Lemma \[lem:varphi\](2) gives $\varphi(\Pi(0,u)) = \varphi(\tau(\Pi(0,u))) = \varphi(\Pi(1,u))$, and also implies that $2\varphi(\Pi(0,t)) = \varphi(\Pi(0,t)) + \varphi(\tau(\Pi(0,t)))$. Substituting all this into the previous expression gives $$\begin{aligned} \lefteqn{\varphi(\Pi(0,t)) + \varphi(\tau(\Pi(0,t)))} & \\ & \hspace{10ex} - \sum_{\mu : s \rightarrow t} \alpha_{\Pi(0,s) \rightarrow \Pi(0,t)} \varphi(\Pi(0,s)) - \sum_{\nu : t \rightarrow u} \alpha_{\Pi(1,u) \rightarrow \Pi(0,t)} \varphi(\Pi(1,u)).\end{aligned}$$ Recall that the sums are over all the arrows in $T$ into $t$ and out of $t$. From the construction of the repetitive quiver ${{\mathbb Z}}T$, this means that between them, the sums can be viewed as being over all the arrows into $(0,t)$ in ${{\mathbb Z}}T$. However, the projection ${{\mathbb Z}}T \rightarrow {{\mathbb Z}}T/\Pi$ is a covering so induces a bijection between the arrows in ${{\mathbb Z}}T$ into $(0,t)$ and the arrows in ${{\mathbb Z}}T/\Pi$ into $\Pi(0,t)$. So in fact, the previous expression can be rewritten $$\varphi(\Pi(0,t)) + \varphi(\tau(\Pi(0,t))) - \sum_{m \rightarrow \Pi(0,t)} \alpha_{m \rightarrow \Pi(0,t)} \varphi(m)$$ where the sum is over all arrows in ${{\mathbb Z}}T/\Pi$ into $\Pi(0,t)$. But identifying ${{\mathbb Z}}T/\Pi$ and $C$, the displayed expression is zero by Lemma \[lem:varphi\](6), so $f$ is additive. Since $f(t) = \varphi(\Pi(0,t))$ by Definition \[def:varphi\] and $\varphi(\Pi(p,t)) = \varphi(\tau^p\Pi(0,t)) = \varphi(\Pi(0,t))$ by Lemma \[lem:varphi\](2), if $f$ were bounded on $T$ then $\varphi$ would be bounded on $C$. But this is false by Lemma \[lem:varphi\_unbounded\]. Recall that the graph $A_{\infty}$ is $$\def\objectstyle{\scriptstyle} \vcenter{ \xymatrix @!0 @+0.25pc { 1 \ar@{-}[rr] & & 2 \ar@{-}[rr] & & 3 \ar@{-}[rr] & & 4 \ar@{-}[rr] & & 5 \ar@{-}[rr]&&{\textstyle \cdots}\\ } },$$ where a convenient numbering of the vertices has been chosen. A quiver of type $A_{\infty}$ is an orientation of this graph. The repetitive quiver ${{\mathbb Z}}A_{\infty}$ does not depend on the orientation; with a standard numbering of the vertices it is $$\def\objectstyle{\scriptstyle} \vcenter{ \xymatrix @!0 @+0.5pc { & & & \vdots & & & & \vdots & & & \\ & *{(3,5)} \ar[dr] & & *{(2,5)} \ar[dr] & & *{(1,5)} \ar[dr] & & *{(0,5)} \ar[dr] & & *{(-1,5)} & \\ & & *{(2,4)} \ar[dr] \ar[ur] & & *{(1,4)} \ar[dr] \ar[ur] & & *{(0,4)} \ar[dr] \ar[ur] & & *{(-1,4)} \ar[dr] \ar[ur] & & \\ {\textstyle \cdots} & *{(2,3)} \ar[dr] \ar[ur] & *{} & *{(1,3)} \ar[dr] \ar[ur] & *{} & *{(0,3)} \ar[dr] \ar[ur] & *{} & *{(-1,3)} \ar[dr] \ar[ur] & *{} & *{(-2,3)} & {\textstyle \cdots}.\\ & & *{(1,2)} \ar[dr] \ar[ur] & & *{(0,2)} \ar[dr] \ar[ur] & & *{(-1,2)} \ar[dr] \ar[ur] & & *{(-2,2)} \ar[dr] \ar[ur] & & \\ & *{(1,1)} \ar[ur] & & *{(0,1)} \ar[ur] & & *{(-1,1)} \ar[ur] & & *{(-2,1)} \ar[ur] & & *{(-3,1)} & \\ } }$$ The translation acts by $\tau(p,t) = (p+1,t)$. \[thm:quiver\_local\_structure\] 1. The component $C$ of the AR quiver $\Gamma$ of ${{\mathsf D}^{\operatorname{c}}}(R)$ is isomorphic to ${{\mathbb Z}}A_{\infty}$ as a stable translation quiver. 2. Each label $(\alpha_{\mu},\beta_{\mu})$ on $\Gamma$ is equal to $(1,1)$. 3. If the function $\varphi$ has value $\varphi_1$ on the edge of $C \cong {{\mathbb Z}}A_{\infty}$, then it has value $n\varphi_1$ on the $n$’th horizontal row of vertices in $C$. By Proposition \[pro:f\], there is an additive unbounded function $f$ on the labelled tree $T$. Hence $T$ is of type $A_{\infty}$ with all labels equal to $(1,1)$ by [@BensonI thm. 4.5.8(iv)]. This proves (2), and it also means that to prove (1), it is sufficient to show that $\Pi$ acts trivially on ${{\mathbb Z}}A_{\infty}$. But if it did not, then there would exist a vertex $m$ on the edge of ${{\mathbb Z}}A_{\infty}$ and a $g$ in $\Pi$ such that $gm \neq m$. The vertex $gm$ would again be on the edge, and so it would have the form $\tau^p m$ for some $p \neq 0$. But then $m$ and $\tau^p m$ would get identified in ${{\mathbb Z}}A_{\infty}/\Pi$, and hence $\Pi m$ would be a fixed point in ${{\mathbb Z}}A_{\infty}/\Pi$ of $\tau^p$, that is, a fixed point in $C$ of $\tau^p$. But this is impossible by Lemma \[lem:no\_loops\](1). Finally, it is a standard consequence of additivity that if the function $f$ has value $f(1) = f_1$ at the first vertex of $A_{\infty}$, then it has value $f(n) = nf_1$ at the $n$th vertex. Since $\varphi(\Pi(p,n)) = \varphi(\tau^p(\Pi(0,n))) = \varphi(\Pi(0,n)) = f(n)$, the claim (3) on $\varphi$ follows. Report on work by Karsten Schmidt {#sec:global} ================================= In this section, the study of the AR quiver $\Gamma$ of ${{\mathsf D}^{\operatorname{c}}}(R)$ is continued, and some aspects of the global structure are revealed. If $\dim_k {\operatorname{H}}\!R = 2$ then $\Gamma$ has precisely $d-1$ components. On the other hand, for Gorenstein algebras with $\dim_k {\operatorname{H}}\!R \geq 3$, there are infinitely many components. Often, these even form families which are indexed by projective manifolds, and these manifolds can be of arbitrarily high dimension. With the exception of Theorem \[thm:spheres\] which is essentially in [@artop], the results of this section are due to Karsten Schmidt; see [@Schmidt thm. 4.1]. Only a sketch is given of the proof of the next theorem; for more information, see [@artop sec. 8]. \[thm:spheres\] If $\dim_k {\operatorname{H}}\!R = 2$ then ${{\mathsf D}^{\operatorname{c}}}(R)$ has AR triangles, and the AR quiver of ${{\mathsf D}^{\operatorname{c}}}(R)$ has $d-1$ components, each isomorphic to ${{\mathbb Z}}A_{\infty}$. The cohomology of $R$ in low degrees is ${\operatorname{H}}^0\!R = k$ and ${\operatorname{H}}^1\!R = 0$. Since $\dim_k {\operatorname{H}}\!R = 2$, it follows that the only other non-zero cohomology is ${\operatorname{H}}^d\!R = k$, and it is easy to check that $R$ therefore satisfies the conditions of Theorem \[thm:Gorenstein\](2) so $R$ is Gorenstein. Theorem \[thm:R\] says that ${{\mathsf D}^{\operatorname{c}}}(R)$ has AR triangles, and Theorem \[thm:quiver\_local\_structure\](1) says that each component of the AR quiver of ${{\mathsf D}^{\operatorname{c}}}(R)$ is isomorphic to ${{\mathbb Z}}A_{\infty}$. Replacing $R$ with a quasi-isomorphic truncation, it can be supposed that $R^{>d} = 0$, see Lemma \[lem:truncations\](3). Pick a cycle $x$ in $R^d$ with non-zero cohomology class. The graded algebra $k[X]/(X^2)$ with $X$ in cohomological degree $d$ can be viewed as a DG algebra with zero differential, and the map $k[X]/(X^2) \rightarrow R$ sending $X$ to $x$ is a quasi-isomorphism, so $R$ can be replaced with $k[X]/(X^2)$. Now consider the algebra $S = k[Y]$ with $Y$ in cohomological degree $-d + 1$, viewed as a DG algebra with zero differential. The DG module $k$ can be viewed as a DG right-$R$-right-$S$-module in an obvious way, and it induces adjoint functors $$\xymatrix{ {{\mathsf D}}(S^{{\operatorname{o}}}) \ar@<-1ex>[rr]_{{\operatorname{RHom}}_{S^{{\operatorname{o}}}}(k,-)} & & {{\mathsf D}}(R). \ar@<-1ex>[ll]_{k {\stackrel{\operatorname{L}}{\otimes}}_R -} }$$ The upper functor clearly sends ${}_{R}R$ to $k_S$, and by computing with a semi-free resolution it can be verified that the lower functor sends $k_S$ to ${}_{R}R$. Hence the functors restrict to quasi-inverse equivalences on the subcategories of objects which are finitely built, respectively, from $k_S$ and ${}_{R}R$. These subcategories are precisely ${{\mathsf D}^{\operatorname{f}}}(S^{{\operatorname{o}}})$ and ${{\mathsf D}^{\operatorname{c}}}(R)$. So it is enough to show that the AR quiver of ${{\mathsf D}^{\operatorname{f}}}(S^{{\operatorname{o}}})$ has $d-1$ components. However, $S$ is $k[Y]$ equipped with zero differential, so ${\operatorname{H}}\!S$ is just $k[Y]$ viewed as a graded algebra. This polynomial algebra in one variable has global dimension $1$, and this makes it possible to prove that if $M$ is a DG right-$S$-module, then $M$ is quasi-isomorphic to ${\operatorname{H}}\!M$ equipped with zero differential. This reduces the classification of objects of ${{\mathsf D}^{\operatorname{f}}}(S^{{\operatorname{o}}})$ to the classification of graded right-${\operatorname{H}}\!S$-modules. However, using again that ${\operatorname{H}}\!S = k[Y]$ is a polynomial algebra in one variable, one shows that its indecomposable finite dimensional graded right-modules are precisely $$\Sigma^j k[Y]/(Y^{m+1})$$ for $j$ in ${{\mathbb Z}}$ and $m \geq 0$. Viewing these as DG right-$S$-modules with zero differential gives the indecomposable objects of ${{\mathsf D}^{\operatorname{f}}}(S^{{\operatorname{o}}})$, and knowing the indecomposable objects, it is an exercise in AR theory to compute the AR triangles, find the AR quiver, and verify that it has $d-1$ components. In the rest of this section, the setup of Section \[sec:local\] will be kept: $R$ is Gorenstein with $\dim_k {\operatorname{H}}\!R \geq 2$. The category ${{\mathsf D}^{\operatorname{c}}}(R)$ has AR triangles by Theorem \[thm:R\], and ${}_{R}k$ is not in ${{\mathsf D}^{\operatorname{c}}}(R)$ by Theorem \[thm:k\]. The AR quiver $\Gamma({{\mathsf D}^{\operatorname{c}}}(R))$ will abbreviated to $\Gamma$. Since ${\operatorname{H}}^0\!R \cong k$ and ${\operatorname{H}}^1\!R = 0$, by Theorem \[thm:Gorenstein\](2) it must be the case that ${\operatorname{H}}^{d-1}\!R = 0$ and ${\operatorname{H}}^d\!R \cong k$. By definition, $d$ is the highest degree in which $R$ has non-zero cohomology; suppose that $e \not\in \{ 0,d \}$ is another degree with ${\operatorname{H}}^e\!R \neq 0$ and observe that then $$2 \leq e \leq d-2$$ and $d \geq 4$. Let $X$ be a minimal semi-free DG left-$R$-module whose semi-free filtration contains only finitely many copies of (de)suspensions of $R$. In particular, Lemma \[lem:semi-free\](4) says that $X$ is in ${{\mathsf D}^{\operatorname{c}}}(R)$; suppose that it is indecomposable in that category. Let $i \geq 2$ and consider the following cases. Case (1). : Suppose that $$\mbox{$\inf X = 0 \;\;$ and $\;\; \sup X = i$.}$$ A non-zero cohomology class in ${\operatorname{H}}^i\!X$ permits a non-zero morphism $\Sigma^{-i}R \stackrel{g}{\rightarrow} X$; denoting the mapping cone by $X(1)$, there is a distinguished triangle $$\label{equ:1} \Sigma^{-i}R \stackrel{g}{\rightarrow} X \rightarrow X(1) \rightarrow.$$ Case (2). : Suppose that $$\mbox{$\inf X = 0, \;\; \sup X = i, \;\;$ and ${\operatorname{H}}^{i-d+e}X \neq 0$.}$$ A non-zero cohomology class in ${\operatorname{H}}^{i-d+e}\!X$ permits a non-zero morphism $\Sigma^{-i+d-e}R \stackrel{h}{\rightarrow} X$; denoting the mapping cone by $X(2)$, there is a distinguished triangle $$\label{equ:2} \Sigma^{-i+d-e}R \stackrel{h}{\rightarrow} X \rightarrow X(2) \rightarrow.$$ Case (2${}_{\alpha}$). : In Case (2), suppose moreover that ${\operatorname{H}}^i\!X \cong k$ and that scalar multiplication induces a non-degenerate bilinear form $$\label{equ:b} {\operatorname{H}}^{d-e}(R) \times {\operatorname{H}}^{i-d+e}(X) \rightarrow {\operatorname{H}}^i(X) \cong k.$$ The morphism $\Sigma^{-i+d-e}R \stackrel{h}{\rightarrow} X$ corresponds to an element $\alpha$ in ${\operatorname{H}}^{i-d+e}\!X$; denote $h$ by $h_{\alpha}$ and $X(2)$ by $X(2_{\alpha})$. It follows from the mapping cone construction that $X(1)$, $X(2)$, and $X(2_{\alpha})$ are again minimal semi-free DG left-$R$-modules whose semi-free filtrations contain only finitely many copies of (de)suspensions of $R$. \[lem:C\] 1. In Case (1) of the above construction, the DG module $X(1)$ is indecomposable in ${{\mathsf D}^{\operatorname{c}}}(R)$. It has $$\inf X(1) = 0, \;\; \sup X(1) = i+d-1$$ and $${\operatorname{H}}^{i+e-1}(X(1)) \cong {\operatorname{H}}^e(R) \neq 0, \;\; {\operatorname{H}}^{i+d-1}(X(1)) \cong {\operatorname{H}}^d(R) \cong k.$$ It satisfies ${\operatorname{amp}}(X(1)) = {\operatorname{amp}}(X) + d - 1$ and $\varphi(X(1)) = \varphi(X) + 1$. Moreover, if the construction is applied to $X$ and $X^{\prime}$ then $X(1) \cong X^{\prime}(1)$ implies $X \cong X^{\prime}$ in ${{\mathsf D}^{\operatorname{c}}}(R)$. Finally, scalar multiplication induces a non-degenerate bilinear form $${\operatorname{H}}^{d-e}(R) \times {\operatorname{H}}^{i+e-1}(X(1)) \rightarrow {\operatorname{H}}^{i+d-1}(X(1)) \cong k.$$ 2. In Case (2), the DG module $X(2)$ is indecomposable in ${{\mathsf D}^{\operatorname{c}}}(R)$. It has $$\mbox{$\inf X(2) = 0 \;\;$ and $\;\; \sup X(2) = i+e-1$}.$$ It satisfies ${\operatorname{amp}}(X(2)) = {\operatorname{amp}}(X) + e - 1$ and $\varphi(X(2)) = \varphi(X) + 1$. Moreover, if the construction is applied to $X$ and $X^{\prime}$ then $X(2) \cong X^{\prime}(2)$ implies $X \cong X^{\prime}$ in ${{\mathsf D}^{\operatorname{c}}}(R)$. 3. In Case (2${}_{\alpha}$), if $\alpha$ and $\alpha^{\prime}$ are elements of ${\operatorname{H}}^{i-d+e}\!X$ then $$\mbox{$X(2_{\alpha}) \cong X(2_{\alpha^{\prime}})$ in ${{\mathsf D}^{\operatorname{c}}}(R)$ \; $\Leftrightarrow$ \; $\alpha = \kappa\alpha^{\prime}$ for a $\kappa$ in $k$.}$$ (1). Indecomposability will follow from [@HKR lem. 6.5] if I can show in ${{\mathsf D}^{\operatorname{c}}}(R)$ that $g$ is non-zero (clear), non-invertible (clear since $\inf \Sigma^{-i}R = i \geq 2$ but $\inf X = 0$), and that ${\operatorname{Hom}}_{{{\mathsf D}^{\operatorname{c}}}(R)}(X,\Sigma\Sigma^{-i}R) = 0$. However, $$\begin{aligned} {\operatorname{Hom}}_{{{\mathsf D}^{\operatorname{c}}}(R)}(X,\Sigma\Sigma^{-i}R) & \cong {\operatorname{Hom}}_{{{\mathsf D}^{\operatorname{c}}}(R^{{\operatorname{o}}})}({\operatorname{D}}\!\Sigma\Sigma^{-i}R,{\operatorname{D}}\!X) \\ & \cong {\operatorname{Hom}}_{{{\mathsf D}^{\operatorname{c}}}(R^{{\operatorname{o}}})}(\Sigma^{i-1}{\operatorname{D}}\!R,{\operatorname{D}}\!X) \\ & \stackrel{\rm (a)}{\cong} {\operatorname{Hom}}_{{{\mathsf D}^{\operatorname{c}}}(R^{{\operatorname{o}}})}(\Sigma^{i-1+d}R,{\operatorname{D}}\!X) \\ & \cong {\operatorname{H}}^{-i+1-d}({\operatorname{D}}\!X) \\ & \cong {\operatorname{D}}\!{\operatorname{H}}^{i-1+d}(X) \\ & \stackrel{\rm (b)}{=} 0\end{aligned}$$ where (a) is by Theorem \[thm:Gorenstein\](3) and (b) is because $\sup X = i$. The statements $\inf X(1) = 0$, $\sup X(1) = i+d-1$, ${\operatorname{H}}^{i+e-1}(X(1)) \cong {\operatorname{H}}^e(R) \neq 0$, and ${\operatorname{H}}^{i+d-1}(X(1)) \cong {\operatorname{H}}^d(R) \cong k$ follow from the long exact cohomology sequence of the distinguished triangle . The statement about the amplitude is a consequence, and $\varphi(X(1)) = \varphi(X) + 1$ because $X(1)$ is minimal semi-free with one more copy of a desuspension of $R$ in its semi-free filtration than $X$; cf. Lemma \[lem:varphi\](1). To get the statement on isomorphisms, first observe that by a computation like the one above, $${\operatorname{Hom}}_{{{\mathsf D}^{\operatorname{c}}}(R)}(X(1),\Sigma\Sigma^{-i}R) \cong {\operatorname{D}}\!{\operatorname{H}}^{i+d-1}(X(1)) \cong {\operatorname{D}}(k) \cong k.$$ Now suppose that there is an isomorphism $X(1) \stackrel{\sim}{\rightarrow} X^{\prime}(1)$ in ${{\mathsf D}^{\operatorname{c}}}(R)$. This gives a diagram between the distinguished triangles defining $X(1)$ and $X^{\prime}(1)$, $$\xymatrix{ \Sigma^{-i}R \ar[r] & X \ar[r] & X(1) \ar[r] \ar[d] & \Sigma^{-i+1}R \\ \Sigma^{-i}R \ar[r] & X^{\prime} \ar[r] & X^{\prime}(1) \ar[r] & \Sigma^{-i+1}R \lefteqn{.} }$$ The last morphism in the upper distinguished triangle is non-zero, for otherwise the triangle would be split contradicting that $X$ is indecomposable. Since ${\operatorname{Hom}}_{{{\mathsf D}^{\operatorname{c}}}(R)}(X(1),\Sigma\Sigma^{-i}R)$ is one-dimensional, there exists a morphism $\Sigma^{-i+1}R \rightarrow \Sigma^{-i+1}R$ to give a commutative square. Adding this morphism and its desuspension to the diagram gives $$\xymatrix{ \Sigma^{-i}R \ar[r] \ar[d] & X \ar[r] & X(1) \ar[r] \ar[d] & \Sigma^{-i+1}R \ar[d]\\ \Sigma^{-i}R \ar[r] & X^{\prime} \ar[r] & X^{\prime}(1) \ar[r] & \Sigma^{-i+1}R \lefteqn{,} }$$ and the two new vertical arrows are also isomorphisms since they are non-zero and since ${\operatorname{Hom}}_{{{\mathsf D}^{\operatorname{c}}}(R)}(R,R) \cong k$. By the axioms of triangulated categories, there is a vertical morphism $X \rightarrow X^{\prime}$ which completes to a commutative diagram, and this morphism is an isomorphism by the triangulated five lemma. Finally, to get the non-degenerate bilinear form, observe that $R$ is Gorenstein so by Theorem \[thm:Gorenstein\](2) scalar multiplication gives a non-degenerate bilinear form $${\operatorname{H}}^{d-e}(R) \times {\operatorname{H}}^{i+e-1}(\Sigma^{-i+1}R) \rightarrow {\operatorname{H}}^{i+d-1}(\Sigma^{-i+1}R) \cong k.$$ But $X(1)$ is a mapping cone which in degrees $\geq i+1$ is equal to $\Sigma^{-i+1}R$, so this gives a non-degenerate bilinear form $${\operatorname{H}}^{d-e}(R) \times {\operatorname{H}}^{i+e-1}(X(1)) \rightarrow {\operatorname{H}}^{i+d-1}(X(1)) \cong k$$ as claimed. \(2) follows by similar arguments. (3). $\Leftarrow$ is elementary. $\Rightarrow$: Given the isomorphism $X(2_{\alpha}) \rightarrow X(2_{\alpha^{\prime}})$, the method applied in the proof of (1) produces a diagram between the distinguished triangles defining $X(2_{\alpha})$ and $X(2_{\alpha^{\prime}})$, $$\xymatrix{ \Sigma^{-i+d-e}R \ar[r]^-{h_{\alpha}} \ar[d] & X \ar[r] \ar[d]^{\gamma} & X(2_{\alpha}) \ar[r] \ar[d] & \Sigma^{-i+d-e+1}R \ar[d]\\ \Sigma^{-i+d-e}R \ar[r]_-{h_{\alpha^{\prime}}} & X \ar[r] & X(2_{\alpha^{\prime}}) \ar[r] & \Sigma^{-i+d-e+1}R \lefteqn{,} }$$ where the vertical maps are isomorphisms. Commutativity of the first square implies $({\operatorname{H}}^{i-d+e}(\gamma))(\alpha) = \alpha^{\prime}$. Consider $x$ in ${\operatorname{H}}^{d-e}\!R$. Then $$x\alpha^{\prime} = 0 \Leftrightarrow x({\operatorname{H}}^{i-d+e}(\gamma))(\alpha) = 0 \Leftrightarrow ({\operatorname{H}}^{i-d+e}(\gamma))(x\alpha) = 0 \Leftrightarrow x\alpha = 0,$$ the last $\Leftrightarrow$ because $\gamma$ is an isomorphism in ${{\mathsf D}^{\operatorname{c}}}(R)$ whence ${\operatorname{H}}^{i-d+e}(\gamma)$ is bijective. Seeing that the bilinear form is non-degenerate, this means that $\alpha = \kappa\alpha^{\prime}$ for a $\kappa$ in $k$. Observe that it makes sense to insert $X(1)$ into either of Cases (1), (2), and (2${}_{\alpha}$). Likewise, it makes sense to insert $X(2)$ and $X(2_{\alpha})$ into Case (1). Iterating Cases (1) and (2), the following tree can be constructed. $$\label{equ:tree} \vcenter{ \xymatrix @C+1pc @R-2pc { & & & X(1,1,1) & \\ & & X(1,1) \ar@{.}[ur] \ar@{.}[dr] & & \ldots \\ & & & X(1,1,2) & \\ & X(1) \ar@{.}[uur] \ar@{.}[ddr] & & & \\ & & & X(1,2,1) & \\ & & X(1,2) \ar@{.}[ur] \ar@{.}[dr] & & \\ & & & {*} & \\ X \ar@{.}[uuuur] \ar@{.}[ddddr] & & & & \dots \\ & & & X(2,1,1) & \\ & & X(2,1) \ar@{.}[ur] \ar@{.}[dr] & & \\ & & & X(2,1,2) & \\ & X(2) \ar@{.}[uur] \ar@{.}[ddr] & & & \\ & & & {*} & \\ & & {*} \ar@{.}[ur] \ar@{.}[dr] & & \ldots \\ & & & {*} & \\ } }$$ The notation is straightforward; for instance, by $X(1,2)$ is denoted the DG module obtained by first performing the construction of Case (1), then the construction of Case (2). The rule for omitting nodes of the tree is that no $X(\cdots)$ must contain two neighbouring digits $2$. \[thm:Schmidt1\] Suppose that $\dim_k {\operatorname{H}}\!R \geq 3$. Then the AR quiver $\Gamma$ of ${{\mathsf D}^{\operatorname{c}}}(R)$ has infinitely many components. It is a standing assumption in this section that $R$ is Gorenstein, so each component $C$ of $\Gamma$ is isomorphic to ${{\mathbb Z}}A_{\infty}$ as a stable translation quiver by Theorem \[thm:quiver\_local\_structure\](1). Since $\dim_k {\operatorname{H}}\!R \geq 3$, there exists an $e \not\in \{ 0,d \}$ such that $R$ has non-zero cohomology in degree $e$, so the above constructions make sense. Start with $X = R$ and consider the tree . It follows from Lemma \[lem:C\], (1) and (2), that the function $\varphi$ is constant with value $r$ on the $r$’th column of the tree. On the other hand, by Theorem \[thm:quiver\_local\_structure\](3), the value of $\varphi$ on the $n$’th horizontal row of a component $C \cong {{\mathbb Z}}A_{\infty}$ of $\Gamma$ is $n\varphi_1$. Hence, if the vertices corresponding to two modules in the $r$’th column of the tree both belong to $C$, then they sit in the same horizontal row of vertices in $C$. Equation implies that ${\operatorname{amp}}\tau Y = {\operatorname{amp}}Y$ for each $Y$ in ${{\mathsf D}^{\operatorname{c}}}(R)$. However, on $C$, the action of $\tau$ is to move a vertex one step to the left. It follows that the amplitude is constant on each horizontal row of $C$. Combining these arguments, if the vertices corresponding to two modules in the $r$’th column of the tree both belong to $C$, then the modules have the same amplitude. On the other hand, in the construction above, Case (1) makes the amplitude grow by $d-1$ and Case (2) makes the amplitude grow by $e-1$. Let $a_1, \ldots, a_r$ be a sequence of the digits $1$ and $2$ which does not contain two neighbouring digits $2$. Suppose that the sequence contains $s$ digits $1$ and $r-s$ digits $2$. Then since ${\operatorname{amp}}X = {\operatorname{amp}}R = d$ it holds that ${\operatorname{amp}}X(a_1, \ldots, a_r) = d + s(d-1) + (r-s)(e-1)$, and since $e < d$ it is clear that this value changes when $s$ changes. So by choosing $r$ sufficiently large, a column of the tree can be achieved with an arbitrarily large number of DG modules with pairwise different amplitudes. By the first part of the proof, this results in an arbitrarily large number of different components of $\Gamma$, so $\Gamma$ has infinitely many components. \[thm:Schmidt2\] Suppose that there is an $e$ with $\dim_k {\operatorname{H}}^e\!R \geq 2$. Then the AR quiver $\Gamma$ of ${{\mathsf D}^{\operatorname{c}}}(R)$ has families of distinct components which are indexed by projective manifolds over $k$, and these manifolds can be of arbitrarily high dimension. Again, it is a standing assumption in this section that $R$ is Gorenstein, so each component $C$ of $\Gamma$ is isomorphic to ${{\mathbb Z}}A_{\infty}$ as a stable translation quiver by Theorem \[thm:quiver\_local\_structure\](1). Set $X = R$. With an obvious notation, consider $X(2_{\alpha},1,2_{\beta})$. Then an isomorphism $X(2_{\alpha},1,2_{\beta}) \cong X(2_{\alpha^{\prime}},1,2_{\beta^{\prime}})$ implies $X(2_{\alpha},1) \cong X(2_{\alpha^{\prime}},1)$ by Lemma \[lem:C\](2), and then $\beta = \lambda\beta^{\prime}$ for a $\lambda$ in $k$ by Lemma \[lem:C\](3). And $X(2_{\alpha},1) \cong X(2_{\alpha^{\prime}},1)$ implies $X(2_{\alpha}) \cong X(2_{\alpha^{\prime}})$ by Lemma \[lem:C\](1), and then $\alpha = \kappa\alpha^{\prime}$ for a $\kappa$ in $k$ by Lemma \[lem:C\](3). The $X(2_{\alpha},1,2_{\beta})$ hence give a family of pairwise non-isomorphic objects of ${{\mathsf D}^{\operatorname{c}}}(R)$ parametrized by the Cartesian product $\{\mbox{rays of $\alpha$'s}\} \times \{\mbox{rays of $\beta$'s}\}$. Now, $\sup X = d$ so the class $\alpha$ is in ${\operatorname{H}}^{d-d+e}(X)$, cf. the construction in Case (2). However, $${\operatorname{H}}^{d-d+e}(X) = {\operatorname{H}}^e(X) = {\operatorname{H}}^e(R).$$ Hence $\{\mbox{rays of $\alpha$'s}\} = {{\mathbb P}}({\operatorname{H}}^e\!R)$ where ${{\mathbb P}}$ denotes the projective space of rays in a vector space. Moreover, $\sup X(2_{\alpha},1) = d+(e-1)+(d-1) = 2d+e-2$ by Lemma \[lem:C\], (1) and (2), so the class $\beta$ is in ${\operatorname{H}}^{(2d+e-2)-d+e}(X(2_{\alpha},1))$. However, $$\begin{aligned} {\operatorname{H}}^{(2d+e-2)-d+e}(X(2_{\alpha},1)) & = {\operatorname{H}}^{d+2e-2}(X(2_{\alpha},1)) \\ & = {\operatorname{H}}^{(d+e-1)+e-1}(X(2_{\alpha},1)) \\ & \cong {\operatorname{H}}^e(R),\end{aligned}$$ where $\cong$ is by Lemma \[lem:C\](1) because $\sup X(2_{\alpha}) = d+e-1$. Hence it is also the case that $\{\mbox{rays of $\beta$'s}\} = {{\mathbb P}}({\operatorname{H}}^e\!R)$. This shows that the $X(2_{\alpha},1,2_{\beta})$ give a family of pairwise non-isomorphic objects of ${{\mathsf D}^{\operatorname{c}}}(R)$ indexed by ${{\mathbb P}}({\operatorname{H}}^e\!R) \times {{\mathbb P}}({\operatorname{H}}^e\!R)$. Note that the projective space ${{\mathbb P}}({\operatorname{H}}^{e}\!R)$ is non-trivial since $\dim_k {\operatorname{H}}^e\!R \geq 2$. By Lemma \[lem:C\], (1) and (2), all the $X(2_{\alpha},1,2_{\beta})$ have the same value of $\varphi$ (it is $4$), so if the vertices of two non-isomorphic ones belonged to the same component $C$ of $\Gamma$, then they would be different vertices in the same horizontal row of $C \cong {{\mathbb Z}}A_{\infty}$ because the value of $\varphi$ on the $n$’th row of $C$ is $n\varphi_1$ by Theorem \[thm:quiver\_local\_structure\](3). However, it follows from Equation that $\inf(\tau Y) = \inf(Y) - d + 1$, so different vertices in the $n$’th row of $C$ correpond to DG modules with different $\inf$, but the $X(2_{\alpha},1,2_{\beta})$ all have the same $\inf$ by Lemma \[lem:C\], (1) and (2) (it is $0$). Hence the vertices of two non-isomorphic $X(2_{\alpha},1,2_{\beta})$’s must belong to different components of $\Gamma$, so a family has been found of distinct components of $\Gamma$ parametrized by the projective manifold ${{\mathbb P}}({\operatorname{H}}^e\!R) \times {{\mathbb P}}({\operatorname{H}}^e\!R)$ over $k$. An analogous argument with objects of the form $X(2_{\alpha},1,2_{\beta},1,\ldots,1,2_{\gamma})$ produces families of distinct components of the AR quiver indexed by projective manifolds of arbitrarily high dimension, as claimed. Poincaré duality spaces {#sec:topology} ======================= This section makes explicit the highlights of the previous sections for DG algebras of the form ${\operatorname{C}}^*(X;k)$. The results first appeared in [@artop], [@arquiv], and [@Schmidt]. In this section, the field $k$ will have characteristic $0$. By $X$ will be denoted a simply connected topological space with $\dim_k {\operatorname{H}}^*(X;k) < \infty$. Write $$n = \sup \{\, i \,|\, {\operatorname{H}}^i(X;k) \neq 0 \,\}.$$ When the singular cochain complex ${\operatorname{C}}^*(X;k)$ and singular cohomology ${\operatorname{H}}^*(X;k)$ appear below, it is always with coefficients in $k$, so I will use the shorthands ${\operatorname{C}}^*(X)$ and ${\operatorname{H}}^*(X)$. \[rmk:A\] The singular chain complex ${\operatorname{C}}^*(X)$ is a DG algebra under cup product, and by [@FHTbook exa. 6, p. 146], it is quasi-isomorphic to a commutative DG algebra $A$ satisfying the conditions of Setup \[set:blanket\]. For $X$ to be simply connected means that it is path connected and that each closed path in $X$ can be shrinked continuously to a point. Equivalently, $X$ is path connected and its fundamental group $\pi_1(X)$ is trivial. The space $X$ is said to have Poincaré duality over $k$ if there is an isomorphism $${\operatorname{D}}\!{\operatorname{H}}^*(X) \cong \Sigma^n {\operatorname{H}}^*(X)$$ of graded left-${\operatorname{H}}^*(X)$-modules. It is a classical theorem that any compact $n$-dimensional manifold has Poincaré duality; indeed, this is one of the oldest results of algebraic topology. A consequence of Poincaré duality over $k$ is that there are isomorphisms of vector spaces $${\operatorname{D}}\!{\operatorname{H}}^i(X) \cong {\operatorname{H}}^{n-i}(X)$$ for each $i$, and hence that the singular cohomology ${\operatorname{H}}^*(X)$ with coefficients in $k$ is concentrated between dimensions $0$ and $n$ and has the same vector space dimension in degrees $i$ and $n-i$. Geometrically, this is in a sense the statement that the number of holes with $i$-dimensional boundary enclosed by $X$ is equal to the number of holes with $(n-i)$-dimensional boundary enclosed by $X$. Algebraically, spaces with Poincaré duality emulate Gorenstein algebras; see [@FHTpaper]. For the definition of $n$-Calabi-Yau categories, see Definitions \[def:Serre\] and \[def:CY\]. \[thm:Chain\_CY\] The following conditions are equivalent. 1. ${{\mathsf D}^{\operatorname{c}}}({\operatorname{C}}^*(X))$ is an $n$-Calabi-Yau category. 2. ${{\mathsf D}^{\operatorname{c}}}({\operatorname{C}}^*(X)^{{\operatorname{o}}})$ is an $n$-Calabi-Yau category. 3. $X$ has Poincaré duality over $k$. This will involve showing that the conditions of the theorem are also equivalent to the following two conditions. 4. ${{\mathsf D}^{\operatorname{c}}}({\operatorname{C}}^*(X))$ has AR triangles. 5. ${{\mathsf D}^{\operatorname{c}}}({\operatorname{C}}^*(X)^{{\operatorname{o}}})$ has AR triangles. For the proof, ${\operatorname{C}}^*(X)$ can be replaced with the commutative DG algebra $A$ by Remark \[rmk:A\]. So it is clear that (1)$\Leftrightarrow$(2) and that (4)$\Leftrightarrow$(5). Condition (3), that $X$ has Poincaré duality, means ${}_{{\operatorname{H}}\!A}({\operatorname{D}}\!{\operatorname{H}}\!A) \cong {}_{{\operatorname{H}}\!A}(\Sigma^n {\operatorname{H}}\!A)$; since $A$ is commutative, Theorem \[thm:Gorenstein\](2) implies that this is equivalent to $A$ being Gorenstein. Condition (4) is also equivalent to $A$ being Gorenstein by Theorem \[thm:R\]. It follows that (3)$\Leftrightarrow$(4). (1)$\Rightarrow$(4) holds since a Calabi-Yau category has a Serre functor and hence AR triangles, see Definition \[def:Serre\], Theorem \[thm:Serre\], and Definition \[def:CY\]. (3)$\Rightarrow$(1). The DG algebra $A$ is commutative, so Theorem \[thm:Gorenstein\](3) implies that condition (3) is equivalent to $${\operatorname{D}}\!A \cong \Sigma^n A$$ in the derived category of DG bi-$A$-modules. Inserting this into Equation shows that the Serre functor of ${{\mathsf D}^{\operatorname{c}}}(A)$ is $\Sigma^n$ so (1) holds, cf. Definition \[def:CY\]. \[thm:Chain\_Gamma\] Suppose that $X$ has Poincaré duality over $k$ and that it satisfies $\dim_k {\operatorname{H}}^*(X) \geq 2$. Then each component of the AR quiver $\Gamma$ of ${{\mathsf D}^{\operatorname{c}}}({\operatorname{C}}^*(X))$ is isomorphic to ${{\mathbb Z}}A_{\infty}$. If $\dim_k {\operatorname{H}}^*(X) = 2$, then $\Gamma$ has $n-1$ components. If $\dim_k {\operatorname{H}}^*(X) \geq 3$, then $\Gamma$ has infinitely many components. If $\dim_k {\operatorname{H}}^e(X) \geq 2$ for some $e$, then $\Gamma$ has families of distinct components which are indexed by projective manifolds over $k$, and these manifolds can be of arbitrarily high dimension. Since ${\operatorname{C}}^*(X)$ is quasi-isomorphic to $A$, the theory of the previous sections applies to ${\operatorname{C}}^*(X)$. As in the proof of Theorem \[thm:Chain\_CY\], since $X$ has Poincaré duality, ${\operatorname{C}}^*(X)$ is Gorenstein. The present theorem hence follows from Theorems \[thm:quiver\_local\_structure\], \[thm:spheres\], \[thm:Schmidt1\], and \[thm:Schmidt2\]. Theorem \[thm:Chain\_CY\] and its proof imply that if $X$ has Poincaré duality over $k$, then the AR quiver of ${{\mathsf D}^{\operatorname{c}}}({\operatorname{C}}^*(X))$ is a stable translation quiver. The AR quiver of ${{\mathsf D}^{\operatorname{c}}}({\operatorname{C}}^*(X))$ is a weak homotopy invariant of $X$. If $X$ is restricted to spaces with Poincaré duality over $k$, then the AR quiver of ${{\mathsf D}^{\operatorname{c}}}({\operatorname{C}}^*(X))$, viewed as a stable translation quiver, is a weak homotopy invariant of $X$. If $X$ and $X^{\prime}$ have the same weak homotopy type, then by [@FHTbook thm. 4.15] there exists a series of quasi-isomorphisms of DG algebras linking ${\operatorname{C}}^*(X)$ and ${\operatorname{C}}^*(X^{\prime})$. Hence ${{\mathsf D}^{\operatorname{c}}}({\operatorname{C}}^*(X))$ and ${{\mathsf D}^{\operatorname{c}}}({\operatorname{C}}^*(X^{\prime}))$ are equivalent triangulated categories, and this implies both parts of the theorem. Open problems {#sec:open} ============= Let me close the paper by proposing the following open problems. The first one is due to Karsten Schmidt, see [@Schmidt sec. 6]. Develop a theory of representation type of simply connected cochain DG algebras. What is known so far is the following. 1. By Theorem \[thm:spheres\], if $\dim_k {\operatorname{H}}\!R = 2$, then the AR quiver $\Gamma$ of ${{\mathsf D}^{\operatorname{c}}}(R)$ has a finite number of components. Suppose that $R$ is Gorenstein. 2. By Theorem \[thm:Schmidt1\], if $\dim_k {\operatorname{H}}\!R \geq 3$, then $\Gamma$ has infinitely many components. 3. By Theorem \[thm:Schmidt2\], if $\dim_k {\operatorname{H}}^e\!R \geq 2$ for some $e$, then $\Gamma$ has families of distinct components which are indexed by projective manifolds, and these manifolds can be of arbitrarily high dimension. It is tempting to interpret the DG algebras of (1) as having finite representation type, and the ones of (3) as having wild representation type. If $\dim_k {\operatorname{H}}\!R \geq 3$ but $\dim_k {\operatorname{H}}^i\!R \leq 1$ for each $i$, then it is not clear whether the infinitely many components of $\Gamma$ form discrete or continuous families, or indeed, what these words precisely mean in the context. Note that some previous work does exist on the representation type of derived categories, see [@GeissKrause], but it does not apply to the categories considered in this paper. What is the structure of the AR quiver of ${{\mathsf D}^{\operatorname{c}}}(R)$ if $R$ is not Gorenstein? Do components of a different shape than ${{\mathbb Z}}A_{\infty}$ become possible? Generalize the theory to cochain DG algebras which are not simply connected. Presently, not even the structure of ${{\mathsf D}^{\operatorname{c}}}({\operatorname{C}}^*(S^1;{{\mathbb Q}}))$ is known because $S^1$ and hence ${\operatorname{C}}^*(S^1;{{\mathbb Q}})$ is not simply connected. A generalization to the non-simply connected case may impact on non-commutative geometry for which more general cochain DG algebras are being considered as vehicles. Is there a link between the categories ${{\mathsf D}^{\operatorname{c}}}(R)$ which have AR quivers consisting of ${{\mathbb Z}}A_{\infty}$-components, and the appearence of ${{\mathbb Z}}A_{\infty}$-components in representation theory? See for instance [@BensonI thm. 4.17.4]. If a simply connected topological space $X$ has $\dim_{{{\mathbb Q}}}{\operatorname{H}}^*(X;{{\mathbb Q}}) = 2$, then it has the same rational homotopy type as a sphere of dimension $\geq 2$. Theorem \[thm:Chain\_Gamma\] implies that these are the only simply connected spaces with Poincaré duality for which the AR quiver of ${{\mathsf D}^{\operatorname{c}}}({\operatorname{C}}^*(X;{{\mathbb Q}}))$ has only finitely many components. Is this linked to any topological property which is special to these spaces? Let $X$ and $T$ be topological spaces. Suppose that $X$ is simply connected with $\dim_k {\operatorname{H}}^*(X;k) < \infty$, that $T$ has $\dim_k {\operatorname{H}}^i(T;k) < \infty$ for each $i$, and let $$F \rightarrow T \rightarrow X$$ be a fibration. The induced morphism ${\operatorname{C}}^*(X;k) \rightarrow {\operatorname{C}}^*(T;k)$ turns ${\operatorname{C}}^*(T;k)$ into a DG left-${\operatorname{C}}^*(X;k)$-module. By [@FHTbook thm. 7.5] there is a quasi-isomorphism $k {\stackrel{\operatorname{L}}{\otimes}}_{{\operatorname{C}}^*(X;k)} {\operatorname{C}}^*(T;k) \simeq {\operatorname{C}}^*(F;k)$, and this implies that if $\dim_k {\operatorname{H}}^*(F;k) < \infty$ then ${\operatorname{C}}^*(T;k)$ is an object of ${{\mathsf D}^{\operatorname{c}}}({\operatorname{C}}^*(X;k))$. Hence ${\operatorname{C}}^*(T;k)$ corresponds to a collection of vertices with multiplicities of the AR quiver $\Gamma$ of ${{\mathsf D}^{\operatorname{c}}}({\operatorname{C}}^*(X;k))$. If $X$ has Poincaré duality over $k$, then the theory of this paper gives information about the structure of $\Gamma$, both locally and globally. Does this have applications to the topological theory of fibrations? Do the structural results on $\Gamma$ correspond to structural results on topological fibrations? By considering the fibration $F \rightarrow T \rightarrow X$, looking at ${\operatorname{C}}^*(T;k)$ as a DG left-${\operatorname{C}}^*(X;k)$-module, and using the theory of this paper, one is in effect doing “AR theory with topological spaces”. Is there a way to do so directly with the spaces themselves? \[prob:CY\] If $X$ is a topological space with $\dim_k {\operatorname{H}}^*(X;k) < \infty$ and Poincaré duality over the field $k$ of characteristic $0$, then ${{\mathsf D}^{\operatorname{c}}}({\operatorname{C}}^*(X;k))$ is an $n$-Calabi-Yau category for some $n$ by Theorem \[thm:Chain\_CY\]. More generally, if $R$ is the DG algebra from setup \[set:blanket\] and $R$ is commutative and Gorenstein, then ${{\mathsf D}^{\operatorname{c}}}(R)$ is a $d$-Calabi-Yau category. These categories appear to behave quite differently from higher cluster categories which are standard examples of Calabi-Yau categories. For instance, an $m$-cluster category contains an $m$-cluster tilting object in terms of which every other object can be built in a single step; this seems to be far from true for ${{\mathsf D}^{\operatorname{c}}}({\operatorname{C}}^*(X;k))$ and ${{\mathsf D}^{\operatorname{c}}}(R)$. Which role do ${{\mathsf D}^{\operatorname{c}}}({\operatorname{C}}^*(X;k))$ and ${{\mathsf D}^{\operatorname{c}}}(R)$ play in the taxonomy of Calabi-Yau categories? In the context of Calabi-Yau categories, there is a “Morita” theorem for higher cluster categories, see [@KellerReiten thm. 4.2]. Is there also a Morita theorem for the categories ${{\mathsf D}^{\operatorname{c}}}(R)$? Differential Graded homological algebra {#app:DG} ======================================= This appendix is an introduction to Differential Graded (DG) homological algebra, written for a reader who is already familiar with the formalism of derived categories of rings. Some useful references are [@AFH], [@FHTpaper appendix], [@FHTbook chps. 3, 6, 18, 19, 20], [@KellerDG], and [@KellerICM]. Let $k$ be a commutative ring. \[DG algebras and modules\] \[def:DG\] A Differential Graded (DG) algebra $R$ over $k$ is a complex of $k$-modules equipped with a product which - turns $R$ into a ${{\mathbb Z}}$-graded $k$-algebra, and - satisfies the Leibniz rule $\partial^R(rs) = \partial^R(r)s + (-1)^i r\partial^R(s)$ when $r$ is in $R^i$. A DG left-$R$-module $M$ is a complex of $k$-modules equipped with an $R$-scalar multiplication which - turns it into a graded module over the underlying graded algebra of $R$, and - satisfies the Leibniz rule $\partial^M(rm) = \partial^R(r)m + (-1)^i r\partial^M(m)$ when $r$ is in $R^i$. DG right-$R$-modules and DG bi-modules are defined analogously. Note that $R$ itself is an important DG bi-$R$-module. Sometimes the notations ${}_{R}M$ and $N_R$ are used to emphasize that $M$ is a DG left-$R$-module, $N$ a DG right-$R$-module. The opposite DG algebra of $R$ is denoted by $R^{{\operatorname{o}}}$. Its product $\cdot$ is given by $r \cdot s = (-1)^{ij}sr$ in terms of the product of $R$, when $r$ and $s$ are elements of $R^i$ and $R^j$. DG right-$R$-modules can be viewed as DG left-$R^{{\operatorname{o}}}$-modules. \[DG homological algebra\] It is possible to do homological algebra with DG modules. A test case is when the DG algebra $R$ is concentrated in degree zero, that is, when $R^i = 0$ for $i \neq 0$. Then the zeroth component, $R^0$, is an ordinary $k$-algebra, DG left-$R$-modules are just complexes of left-$R^0$-modules, and DG homological algebra over $R$ specializes to ordinary homological algebra over $R^0$. \[inf, sup, and amp\] \[def:sup\_and\_inf\] The infimum and supremum of a DG module are $$\inf M = \inf \{\, i \,|\, {\operatorname{H}}^i\!M \neq 0 \,\}, \;\;\; \sup M = \sup \{\, i \,|\, {\operatorname{H}}^i\!M \neq 0 \,\},$$ and the amplitude is $${\operatorname{amp}}M = \sup M - \inf M.$$ Note that $\inf 0 = \infty$, $\sup 0 = -\infty$, and ${\operatorname{amp}}0 = -\infty$. \[Morphisms, suspensions, and mapping cones\] \[def:morphisms\] The notation $(-)^{\natural}$ is used for the operation of forgetting the differential. It sends DG algebras and DG modules to graded algebras and graded modules. A morphism $\rho : R \rightarrow S$ of DG algebras is a homomorphism $R^{\natural} \rightarrow S^{\natural}$ of the underlying graded algebras which respects the differentials, $\rho\partial^R = \partial^S\rho$. A morphism $\mu : M \rightarrow N$ of DG $R$-modules is a homomorphism $M^{\natural} \rightarrow N^{\natural}$ of the underlying graded $R^{\natural}$-modules which respects the differentials, $\mu\partial^M = \partial^N \mu$. The morphism $\mu$ is called null homotopic if there exists a homomorphism $\theta : M^{\natural} \rightarrow N^{\natural}$ of degree $-1$ of graded $R^{\natural}$-modules such that $\mu = \partial^N\theta + \theta\partial^M$. Morphisms $\mu$ and $\mu^{\prime}$ are called homotopic if $\mu - \mu^{\prime}$ is null homotopic. Suspension of complexes is denoted by $\Sigma$. Suspensions and mapping cone constructions of DG left-$R$-modules inherit DG left-$R$-module structures. Some sign issues are involved here as well as in other parts of the theory; I will not go into details but refer the reader to the references given. \[Cohomology\] \[def:homology\] The product on $R$ and the scalar multiplication of $R$ on $M$ induces a product on the cohomology ${\operatorname{H}}\!R$ and a scalar multiplication of ${\operatorname{H}}\!R$ on ${\operatorname{H}}\!M$, whereby ${\operatorname{H}}\!R$ becomes a graded $k$-algebra and ${\operatorname{H}}\!M$ becomes a graded ${\operatorname{H}}\!R$-module. A morphism $\mu$ of DG modules is called a quasi-isomorphism if the induced homomorphism ${\operatorname{H}}\!\mu$ of graded ${\operatorname{H}}\!R$-modules is an isomorphism. \[Homotopy and derived categories\] \[def:D\] The homotopy category ${\mathsf K}(R)$ has as objects the DG left-$R$-modules, and as morphisms the homotopy classes of morphisms of DG modules. The derived category ${{\mathsf D}}(R)$ is obtained from ${\mathsf K}(R)$ by formally inverting the quasi-isomorphisms. Both ${\mathsf K}(R)$ and ${{\mathsf D}}(R)$ are triangulated categories with distinguished triangles induced by the mapping cone construction. The categories ${\mathsf K}(R)$ and ${{\mathsf D}}(R)$ have set indexed coproducts which are given by ordinary direct sums. The categories ${\mathsf K}(R^{{\operatorname{o}}})$ and ${{\mathsf D}}(R^{{\operatorname{o}}})$ can be viewed as being the homotopy and derived categories of DG right-$R$-modules. A quasi-isomorphism $R \rightarrow S$ of DG algebras induces an equivalence of triangulated categories ${{\mathsf D}}(S) \rightarrow {{\mathsf D}}(R)$ given by change of scalars. Denote by ${{\mathsf D}^{\operatorname{f}}}(R)$ the full subcategory of ${{\mathsf D}}(R)$ consisting of DG modules $M$ with ${\operatorname{H}}\!M$ finitely presented over $k$. Denote by ${{\mathsf D}^{\operatorname{c}}}(R)$ the full subcategory of ${{\mathsf D}}(R)$ consisting of DG modules which are finitely built in ${{\mathsf D}}(R)$ from $R$ using distinguished triangles, (de)suspensions, coproducts, and direct summands; these are the so-called compact objects of ${{\mathsf D}}(R)$. \[Hom and Tensor\] \[def:Hom\_and\_Tensor\] If $M$ and $N$ are DG left-$R$-modules, then there is a graded $k$-module ${\operatorname{Hom}}_{R^{\natural}}(M^{\natural},N^{\natural})$ of graded $R^{\natural}$-homomorphisms $M^{\natural} \rightarrow N^{\natural}$ of different degrees. This can be turned into a complex ${\operatorname{Hom}}_R(M,N)$ with the differential induced by the differentials of $M$ and $N$. Note that ${\operatorname{Hom}}_R(M,N)^{\natural} = {\operatorname{Hom}}_{R^{\natural}}(M^{\natural},N^{\natural})$. If $A$ is a DG right-$R$-module and $B$ is a DG left-$R$-module, then the tensor product $A^{\natural} \otimes_{R^{\natural}} B^{\natural}$ is a graded $k$-module. It can be turned into a complex $A \otimes_R B$ with the differential induced by the differentials of $A$ and $B$. Note that $(A \otimes_R B)^{\natural} = A^{\natural} \otimes_{R^{\natural}} B^{\natural}$. These constructions induce functors between homotopy categories, and there are induced derived functors $${\operatorname{RHom}}_R(-,-) : {{\mathsf D}}(R) \times {{\mathsf D}}(R) \rightarrow {{\mathsf D}}(k)$$ and $$- {\stackrel{\operatorname{L}}{\otimes}}_R - : {{\mathsf D}}(R^{{\operatorname{o}}}) \times {{\mathsf D}}(R) \rightarrow {{\mathsf D}}(k).$$ These are often computed using resolutions. For instance, let $M$ be a DG left-$R$-module and let $P \rightarrow M$ be a K-projective resolution of $M$. This is a quasi-isomorphism of DG modules for which $P$ is K-projective, that is, ${\operatorname{Hom}}_R(P,-)$ preserves quasi-isomorphisms. Then ${\operatorname{Hom}}_R(P,-)$ is a well defined functor ${{\mathsf D}}(R) \rightarrow {{\mathsf D}}(k)$, and there is an equivalence of functors ${\operatorname{RHom}}_R(M,-) \simeq {\operatorname{Hom}}_R(P,-)$. The functor ${\operatorname{RHom}}_R$ has the useful property $${\operatorname{H}}^0 {\operatorname{RHom}}_R(M,N) \cong {\operatorname{Hom}}_{{{\mathsf D}}(R)}(M,N);$$ more generally, the notation $${\operatorname{H}}({\operatorname{RHom}}_R(M,N)) = {\operatorname{Ext}}_R(M,N)$$ is used so ${\operatorname{H}}^i {\operatorname{RHom}}_R(M,N) = {\operatorname{Ext}}_R^i(M,N)$. The functors ${\operatorname{RHom}}$ and ${\stackrel{\operatorname{L}}{\otimes}}$ are compatible with DG bi-modules. For instance, if $A$ is a DG bi-$R$-module then $A {\stackrel{\operatorname{L}}{\otimes}}_R B$ inherits a left-$R$-structure from $A$, so there is a functor $$A {\stackrel{\operatorname{L}}{\otimes}}_R - : {{\mathsf D}}(R) \rightarrow {{\mathsf D}}(R).$$ Now consider the special case of this paper: $k$ is a field and $R$ is a DG algebra over $k$ which has the form $$\cdots \rightarrow 0 \rightarrow k \rightarrow 0 \rightarrow R^2 \rightarrow R^3 \rightarrow \cdots,$$ that is, $R^{<0} = 0$, $R^0 = k$, and $R^1 = 0$. It will also be assumed that $\dim_k R < \infty$, and $d$ will be defined by $$d = \sup R.$$ \[Duality\] \[def:dual\] By ${\operatorname{D}}(-)$ will be denoted the functor ${\operatorname{Hom}}_k(-,k)$. When applied to graded objects, it is understood to be applied degreewise. It sends DG left-$R$-modules to DG right-$R$-modules and vice versa. It is well defined at the level of homotopy and derived categories. \[rmk:Df\] Over a DG algebra of the present special form, $k \cong R/R^{\geq 1}$ is a DG bi-$R$-module. Moreover, ${{\mathsf D}^{\operatorname{f}}}(R)$ is the full subcategory of ${{\mathsf D}}(R)$ consisting of objects $M$ with $\dim_k {\operatorname{H}}\!M < \infty$, and ${{\mathsf D}^{\operatorname{f}}}(R)$ consists precisely of the objects finitely built from ${}_{R}k$. This can be shown using the first two parts of the following result on truncations, the proof of which uses only linear algebra over the field $k$; see [@artop lem. 3.4] and [@FHTbook ex. 6, p. 146]. \[Truncations\] \[lem:truncations\] 1. If $M$ is a DG left-$R$-module with $\inf M$ finite, then there exists a quasi-isomorphism of DG left-$R$-modules $U \rightarrow M$ with $U^i = 0$ for $i < \inf M$. 2. If $N$ is a DG left-$R$-module with $\sup N$ finite, then there exists a quasi-isomorphism of DG left-$R$-modules $N \rightarrow V$ with $V^j = 0$ for $j > \sup N$. 3. The DG algebra $R$ is quasi-isomorphic to a quotient DG algebra $S$ with $S^{>d} = 0$. \[Semi-free DG modules\] A DG left-$R$-module $F$ is called semi-free if it permits a semi-free filtration, that is, a filtration by DG left-$R$-modules $$0 = F \langle -1 \rangle \subseteq F \langle 0 \rangle \subseteq F \langle 1 \rangle \subseteq \cdots \subseteq F$$ where $F = \bigcup_i F \langle i \rangle$ and where each $F \langle i \rangle / F \langle i-1 \rangle$ is a direct sum of (de)suspensions of ${}_{R}R$. If $\partial^F(F) \subseteq R^{\geq 1} \cdot F$, then $F$ is called minimal. If $M$ is in ${{\mathsf D}}(R)$ then a (minimal) semi-free resolution of $M$ is a quasi-isomorphism $F \rightarrow M$ where $F$ is (minimal) semi-free. The following lemma collects useful facts; for references see [@AFH], [@BN], [@FHTpaper appendix], [@FHTbook sec.  6], [@artop sec. 3], [@KellerDG sec. 3], and [@Spaltenstein]. \[Semi-free resolutions\] \[lem:semi-free\] 1. Each $M$ in ${{\mathsf D}}(R)$ has a semi-free resolution. 2. A semi-free DG module is [K]{}-projective, so if $F$ is a semi-free resolution of $M$ then ${\operatorname{RHom}}_R(M,-) \simeq {\operatorname{Hom}}_R(F,-)$ and $- {\stackrel{\operatorname{L}}{\otimes}}_R M \simeq - \otimes_R F$. 3. Each $M$ in ${{\mathsf D}^{\operatorname{f}}}(R)$ has a minimal semi-free resolution $F$, and for each such resolution there are finite numbers $\beta_i$ such that $$F^{\natural} \cong \bigoplus_{i \leq -\inf M} \Sigma^i(R^{\natural})^{(\beta_i)},$$ where $(R^{\natural})^{(\beta_i)}$ is a direct sum of $\beta_i$ copies of $R^{\natural}$. 4. Let $M$ in ${{\mathsf D}^{\operatorname{f}}}(R)$ have minimal semi-free resolution $F$. Then $M$ is in ${{\mathsf D}^{\operatorname{c}}}(R)$ if and only the numbers $\beta_i$ from part (3) satisfy $\beta_i = 0$ for $i \ll 0$. 5. If $F$ is minimal semi-free then ${\operatorname{Hom}}_R(F,k)$ has zero differential, so $${\operatorname{H}}({\operatorname{Hom}}_R(F,k)) \cong {\operatorname{Hom}}_R(F,k)^{\natural} = {\operatorname{Hom}}_{R^{\natural}}(F^{\natural},k^{\natural})$$ as graded $k$-vector spaces. Auslander-Reiten theory for triangulated categories {#app:AR} =================================================== This appendix is a brief introduction to the version of Auslander-Reiten (AR) theory used in the rest of the paper. Some useful references are [@ARS], [@BensonI], [@Happel], [@KrauseARZ], [@KrauseAR], and [@ReitenVandenBergh], with [@Happel] being the source of the theory. Let ${\mathsf T}$ be a triangulated category. The following definition is taken from [@KrauseAR def. 2.1]; it generalizes the earlier definition from [@Happel 3.1]. \[AR triangles\] An AR triangle in ${\mathsf T}$ is a distinguished triangle $$\label{equ:AR} M \stackrel{\mu}{\rightarrow} N \stackrel{\nu}{\rightarrow} P \stackrel{\pi}{\rightarrow}$$ for which - Each morphism $M \rightarrow N^{\prime}$ which is not a split monomorphism factors through $\mu$. - Each morphism $N^{\prime} \rightarrow P$ which is not a split epimorphism factors through $\nu$. - $\pi \neq 0$. In an AR triangle, the end terms determine each other up to isomorphism by [@Happel prop. 3.5(i)], so the following definition makes sense. \[AR translation\] Let $P$ be an object of ${\mathsf T}$ and suppose that there is an AR triangle $M \rightarrow N \rightarrow P \rightarrow$. Then $M$ is denoted by $\tau P$, and the operation $\tau$ which is defined up to isomorphism is called the AR translation of ${\mathsf T}$. In an AR triangle, the end terms have local endomorphism rings by [@KrauseAR lem. 2.3]; this explains the following terminology. The triangulated category ${\mathsf T}$ is said to have right AR triangles if, for each object $P$ with local endomorphism ring, there is an AR triangle . The category ${\mathsf T}$ is said to have left AR triangles if, for each object $M$ with local endomorphism ring, there is an AR triangle . The category ${\mathsf T}$ is said to have AR triangles if it has right and left AR triangles. \[The AR quiver\] A morphism in ${\mathsf T}$ is called irreducible if it is not an isomorphism, but has the property that when it is factored as $\rho\sigma$, then either $\rho$ is a split epimorphism or $\sigma$ is a split monomorphism. The AR quiver $\Gamma({\mathsf T})$ of ${\mathsf T}$ has one vertex $[M]$ for each isomorphism class of objects with local endomorphism rings, and one arrow $[M] \rightarrow [N]$ when there is an irreducible morphism $M \rightarrow N$. If ${\mathsf T}$ has right AR triangles, then the AR translation $\tau$ induces a map from the set of vertices of $\Gamma({\mathsf T})$ to itself. By abuse of notation, this map is also referred to as the AR translation and denoted by $\tau$. Now consider the special case of this paper: $k$ is a field and ${\mathsf T}$ is $k$-linear and has finite dimensional ${\operatorname{Hom}}$ spaces and split idempotents; cf. Proposition \[pro:DcandDfKrullSchmidt\]. Then ${\mathsf T}$ is a Krull-Schmidt category by [@Ringel p. 52]; that is, each indecomposable object has local endomorphism ring and each object splits into a finite direct sum of indecomposable objects which are unique up to isomorphism. The following lemma holds by [@Happel prop. 3.5]. \[lem:Happel\] Let $M \rightarrow N \rightarrow P \rightarrow$ be an AR triangle and let $N \cong \coprod_i N_i$ where each $N_i$ is indecomposable. Then the following statements are equivalent for an indecomposable object $N^{\prime}$. 1. There is an irreducible morphism $M \rightarrow N^{\prime}$. 2. There is an irreducible morphism $N^{\prime} \rightarrow P$. 3. There is an $i$ such that $N^{\prime} \cong N_i$. Hence if ${\mathsf T}$ has AR triangles, knowledge of these triangles implies knowledge of the AR quiver $\Gamma({\mathsf T})$. \[Stable translation quivers\] \[def:stable\_translation\_quiver\] A stable translation quiver is a quiver equipped with an injective map $\tau$ from the set of vertices to itself such that the number of arrows from $\tau(t)$ to $s$ is equal to the number of arrows from $s$ to $t$. The following proposition follows easily from Lemma \[lem:Happel\]. \[pro:stable\_translation\_quiver\] If ${\mathsf T}$ has AR triangles, then the AR translation $\tau$ turns the AR quiver $\Gamma({\mathsf T})$ into a stable translation quiver. \[Serre functors\] \[def:Serre\] A Serre functor for ${\mathsf T}$ is an autoequivalence $S$ for which there are natural isomorphisms $${\operatorname{D}}({\operatorname{Hom}}_{{\mathsf T}}(M,N)) \cong {\operatorname{Hom}}_{{\mathsf T}}(N,SM).$$ The following was proved in [@ReitenVandenBergh thm. I.2.4]. \[thm:Serre\] The category ${\mathsf T}$ has AR triangles if and only if it has a Serre functor $S$. If it does, then $\tau = \Sigma^{-1}S$ on indecomposable objects. This implies that if ${\mathsf T}$ has AR triangles, then the AR translation $\tau$ can be extended to the autoequivalence $\Sigma^{-1}S$. \[Calabi-Yau categories\] \[def:CY\] The category ${\mathsf T}$ is called $n$-Calabi-Yau if $n$ is the smallest non-negative integer for which $\Sigma^n$ is a Serre functor. [99]{} M. Auslander, I. Reiten, and S. Smalø, *Representation theory of Artin algebras*. Cambridge Stud. Adv. Math. 36, Cambridge University Press, Cambridge, 1997, paperback edition. L. L. Avramov, H.-B. Foxby, and S. Halperin, Differential Graded homological algebra. Preprint, version from 21 June 2006. D. J. Benson, *Representations and cohomology I*. Cambridge Stud. Adv. Math. 30, Cambridge University Press, Cambridge, 1995. M. Bökstedt and A. Neeman, Homotopy limits in triangulated categories. *Compositio Math.* **86** (1993), 209–234. Y. Félix, S. Halperin, and J.-C. Thomas, Gorenstein spaces. *Adv. Math.* **71** (1988), 92–112. Y. Félix, S. Halperin, and J.-C. Thomas, *Rational Homotopy Theory*. Grad. Texts in Math. 205, Springer, Berlin, 2001. A. J. Frankild and P. Jørgensen, Homological properties of cochain Differential Graded algebras. Preprint (2008). [math.RA/0801.1581.]{} C. Geiss and H. Krause, On the notion of derived tameness. *J. Algebra Appl.* [**1**]{} (2002), 133–157. D. Happel, On the derived category of a finite dimensional algebra. *Comment. Math. Helv.* **62** (1987), 339–389. D. Happel, B. Keller, and I. Reiten, Bounded derived categories and repetitive algebras. Preprint (2007). [ math.RT/0702302.]{} P. Jørgensen, Auslander-Reiten theory over topological spaces. *Comment. Math. Helv.* **79** (2004), 160–182. P. Jørgensen, The Auslander-Reiten quiver of a Poincaré duality space. *Algebr. Represent. Theory* **9** (2006), 323–336. B. Keller, Deriving DG categories. *Ann. Sci. École Norm. Sup. (4)* **27** (1994), 63–102. B. Keller, On Differential Graded categories. In *Proceedings of the International Congress of Mathematicians 2006, Vol. II* (edited by Marta Sanz-Solé et.al.). European Mathematical Society Publishing House, Zürich, 2007, 151–190. B. Keller and I. Reiten, Acyclic Calabi-Yau categories, with an appendix by Michel Van den Bergh. Preprint (2006). [math.RT/0610594]{}. H. Krause, Auslander-Reiten triangles and a theorem of Zimmermann. *Bull. London Math. Soc.* **37** (2005), 361–372. H. Krause, Auslander-Reiten theory via Brown representability. *$K$-Theory* **20** (2000), 331–344. I. Reiten and M. Van den Bergh, Noetherian hereditary abelian categories satisfying Serre duality. *J. Amer. Math. Soc.* **15** (2002), 295–366. C. M. Ringel, *Tame algebras and quadratic forms*. Lecture Notes in Math. 1099, Springer, Berlin, 1984. K. Schmidt, Auslander-Reiten theory for simply connected differential graded algebras. Thesis, University of Paderborn, Paderborn, 2007. [math.RT/0801.0651]{}. N. Spaltenstein, Resolutions of unbounded complexes. *Compositio Math.* **65** (1988), 121–154.

ArXiv

--- abstract: | The lower-critical dimension for the existence of the Ising spin-glass phase is calculated, numerically exactly, as $d_L = 2.520$ for a family of hierarchical lattices, from an essentially exact (correlation coefficent $R^2 = 0.999999$) near-linear fit to 23 different diminishing fractional dimensions. To obtain this result, the phase transition temperature between the disordered and spin-glass phases, the corresponding critical exponent $y_T$, and the runaway exponent $y_R$ of the spin-glass phase are calculated for consecutive hierarchical lattices as dimension is lowered. PACS numbers: 75.10.Nr, 05.10.Cc, 64.60.De, 75.50.Lk author: - 'Mehmet Demirtaş$^{1}$, Asl[i]{} Tuncer$^2$, and A. Nihat Berker$^{1,3}$' title: 'Lower-Critical Spin-Glass Dimension from 23 Sequenced Hierarchical Models' --- Introduction ============ Singular phase diagram behavior as a function of spatial dimensionality $d$ compounds the interest and challenge of the phase transitions problems, as effectively pausing the “phase transition of phase transitions” problem. Most visible are the lower-critical dimensions, which are the spatial dimensional thresholds for different types of orderings. For example, the lower-critical threshold for ferromagnetic ordering in magnetic systems is $d_L=1$ for one-component (Ising) spins and $d_L=2$ for spins with more than one component. Similarly, the lower-critical dimensions for ferromagnetic ordering under quenched random fields [@Jaccarino; @Birgeneau; @Wong; @Berker84; @Aizenman; @AizenmanE; @Falicov] are respectively $d_L=2$ and $d_L=4$, for one-component spins and for spins with more than one component. A very recent experimental study [@Orbach] on Ge:Mn films has shown the spin-glass lower-critical temperature to be $2<d_L<3$. This is consistent with the earlier theoretical result of $d_L=2.5$ from replica symmetry-breaking mean-field theory.[@Parisi] A numerical fit to the spin-glass critical temperatures for integer dimensions has also suggestted $d_L=2.5$.[@Boettcher] Other theoretical work have claimed $d_L=4$ from earlier ordered-phase stability studies [@Moore1; @Sompolinsky; @Dedominicis], $2<d_L<3$ from transfer-matrix studies [@Moore2],and more recently $d_L=2$ from Monte Carlo [@Matsubara; @Houdayer] and ground-state studies [@Young]. Renormalization-group work, on in effect two hierarchical lattices [@Southern] different from ours have earlier obtained $2<d_L<3$ and on a family of hierarchical lattices [@Amoruso], again different from ours, find $d_L$ close to 2.5. The lower-critical dimensions need not be integer, in view of physical systems on fractal/hierarchical lattices and algebraic manipulations that analytically continue. In fact, it would be highly interesting to find a lower-critical dimension that is neither an integer, nor a simple fraction. Our current study indicates that this is in fact the case for the family of hierarchical lattices studied here, with $d_L=2.520$. We obtain this result from a remarkably good fit to the renormalization-group runaway exponent $y_R$ from the numerically exact renormalization-group solution of a family of 23 hierarchical models with non-integer dimensions $d =$ 2.46, 2.63, 2.77, 2.89, 3.00, 3.10, 3.18, 3.26, 3.33, 3.40, 3.46, 3.52, 3.58, 3.63, 3.68, 3.72, 3.77, 3.81, 3.85, 3.89, 3.93, 3.97, 4.00. Our result is also consistent with the results that are graphically displayed in Ref.[@Amoruso] for a family of hierarchical lattices. Lower-Critical Dimension from Sequenced Hierarchical Models =========================================================== Hierarchical models are constructed [@BerkerOstlund; @Kaufman1; @Kaufman2; @McKay; @Hinczewski1] by imbedding a graph into a bond, as examplified in Fig. 1, and repeating this procedure by self-imbedding infinitely many times. This procedure can also be done on units with more than two external vertices, *e.g.*, the layered Sierpinski gasket in Ref. [@BerkerMcKay]. When interacting systems are placed on hierarchical lattices, their renormalization-group solution proceeds in the reverse direction than the lattice build-up just described, each eliminated elementary graph generating a renormalized interaction strength for the ensuing elementary bond. Hierarchical lattices were originally introduced [@BerkerOstlund] as presenting exactly soluble models with renormalization-group recursion relations that are identical to those found in approximate position-space renormalization-group treatments of Euclidian lattices [@Migdal; @Kadanoff], identifying the latter as physically realizable approximations. However, from the above, it is clear that any graph (or graphs [@Hinczewski1]) may be chosen in the self-imbedding procedure and one need not be faithful to any approximate renormalization-group solution. Hierarchical lattices have been used to study a variety of spin-glass and other statistical mechanics problems.[@Gingras2; @Migliorini; @Gingras1; @Hinczewski; @Guven; @Ohzeki; @Ozcelik; @Gulpinar; @Ilker1; @Ilker2; @Ilker3; @Kaufman; @Barre; @Monthus; @Zhang; @Shrock; @Xu; @Herrmann1; @Herrmann2; @Hwang2013; @Garel; @Hartmann; @Fortin; @Wu; @Timonin; @Derrida; @Thorpe; @Hasegawa; @Monthus2; @Lyra; @Singh; @Xu2014; @Hirose] The length rescaling factor $b$ in a hierarchical lattice is the number of bonds on the shortest distance between the external vertices of the elementary graph which is replaced by a single bond in one scale change. The volume rescaling factor $b^d$ is the number of bonds inside the elementary graph. From these two rescaling factors, the dimensionality $d$ is extracted, as exemplified in Fig. 1. In our study, $b=3$ is used in order to treate the ferromagnetic and antiferromagnetic correlations on equal footing. The lower-critical dimension of spin-glass systems is studied here by considering a systematic family of hierarchical lattices in all its possible decreasing dimensions. The spin-glass system and the renormalization-group method ========================================================== The Ising spin-glass system is defined by the Hamiltonian $$-\beta \mathcal{H}=\sum_{\langle ij \rangle} J_{ij} s_i s_j$$ where $\beta=1/kT$, at each site $i$ of a lattice the spin $s_i = \pm 1$, and $\langle ij \rangle$ denotes that the sum runs over all nearest-neighbor pairs of sites. The bond strengths $J_{ij}$ are $+J>0$ (ferromagnetic) with probability $1-p$ and $-J$ (antiferromagnetic) with probability $p$.  The renormalization-group transformation is achieved by a decimation, $$e^{J_{im}^{(dec)}s_i s_m + G_{im}}=\sum_{s_j,s_k} e^{J_{ij} s_i s_j+J_{jk} s_j s_k +J_{km} s_k s_m},$$ where the additive constants $G_{ij}$ are unavoidably generated, followed by $n$ bond movings, $$J_{ij}^{(bm)} = \sum_{k=1}^n J_{ij}^{(k)}.$$  ------------------------------------------------------------------------ ----------- -------------- ----------- ----------- Spatial Critical Critical Runaway Dimension Temperatures Exponents Exponents $d$ $1/J_C$ $y_T$ $y_R$ 2.630930 0.519268 0.098077 0.058731 2.771244 0.747982 0.188596 0.129983 2.892789 0.890503 0.253690 0.191904 3.000000 1.001319 0.313414 0.246144 3.095903 1.091770 0.361975 0.294649 3.182658 1.168653 0.393837 0.338155 3.261860 1.237723 0.425397 0.377881 3.334718 1.298225 0.451743 0.414440 3.402174 1.354258 0.476199 0.448214 3.464974 1.404661 0.495999 0.479850 3.523719 1.452817 0.513016 0.509181 3.578902 1.496452 0.531699 0.536880 3.630930 1.538271 0.549022 0.563149 3.680144 1.577300 0.562079 0.587707 3.726833 1.613844 0.573932 0.610941 3.771244 1.649036 0.585283 0.633434 3.813588 1.682659 0.594959 0.654932 3.854050 1.714417 0.605789 0.675080 3.892789 1.745469 0.616496 0.693914 3.929947 1.774176 0.623179 0.712461 3.965647 1.802906 0.630527 0.730927 4.000000 1.829792 0.638313 0.747294 ----------- -------------- ----------- ----------- : Critical temperatures $1/J_c$ and critical exponents $y_T$ of the phase transitions between the spin-glass and paramagnetic phases as a function of dimension $d$, for the hierarchical models with antiferromagnetic bond concentration $p=0.5$. The runaway exponents $y_R$ of the spin-glass phase are also shown and give a perfect fit to $y_R = -1.30908 + 0.528513 d - 0.00354805 d^2$, leading with a small extrapolation to the lower-critical dimension $d = 2.520$ for $y_R = 0$, with a very satisfactory correlation coefficient of $R^2 = 0.999999$.[]{data-label="tab:addlabel"} ------------------------------------------------------------------------ The starting bimodal quenched probability distribution of the interactions, characterized by $p$ and described above, is not conserved under rescaling. The renormalized quenched probability distribution of the interactions is obtained by the convolution [@Andelman] $$P'(J'_{i'j'}) = \int{\left[\prod_{ij}^{i'j'}dJ_{ij} P(J_{ij})\right]} \delta(J'_{i'j'}-R(\left\{J_{ij}\right\})),$$ where $R(\left\{J_{ij}\right\})$ represents the decimation and bond moving given in Eqs.(2) and (3). For numerical practicality, the bond moving and decimation of Eqs.(2) and (3) are achieved by a sequence of pairwise combinations of interactions, each pairwise combination leading to an intermediate probability distribution resulting from a pairwise convolution as in Eq.(4). The probability distribution is represented by 200 histograms [@Migliorini; @Hinczewski; @Guven; @Ozcelik; @Gulpinar; @Ilker2], which are apportioned in $J\gtrless 0$ according to total probability weight. The histograms are distributed in the interval $J_+ \pm 2.5 \sigma_+$, where $J_+$ and $\sigma_+$ are the average and standard deviation of the $J>0$ interactions, and similarly for the $J<0$ interactions. The different thermodynamic phases of the system are identified by the different asymptotic renormalization-group flows of the quenched probability distributions. For all renormalization-group flows, inside the phases and on the phase boundaries, Eq.(4) is iterated until asymptotic behavior is reached. Thus, we are able to calculate phase transition temperatures and, by linearization around the unstable asymptotic fixed distribution of the phase boundaries, critical exponents. Similar previous studies, on other spin-glass systems, are in Refs. [@Gingras2; @Migliorini; @Gingras1; @Hinczewski; @Guven; @Ohzeki; @Ozcelik; @Gulpinar; @Ilker1; @Ilker2]. Diminishing Critical, Runaway Exponents, Critical Temperatures and the Lower-Critical Dimension of the Sequence =============================================================================================================== For our chosen sequence of hierarchical systems (Fig. 1), we have calculated, at antiferromagnetic bond concentration $p=0.5$, the phase transition temperature $1/J_C$ where the renormalization-group flows bifurcate between the disordered-phase and the spin-glass-phase attractor sinks. The spin-glass sink is characterized by an interaction probability distribution $P(J_{ij})$ that is symmetric in ferromagnetism-antiferromagnetism $(J_{ij}\gtrless 0)$ and that diverges in interaction absolute value: The average interaction strength $<|J|>$ across the system diverges as $b^{n y_R}$ where $n$ is the number of renormalization-group iterations and $y_R > 0$ is the runaway exponent. The spin-glass sink and simultaneously the spin-glass phase disappears when the runaway exponent $y_R$ reaches 0.[@Ilker2] The calculated spin-glass phase transition temperatures, critical and runaway exponents are given in Fig. 2 and in Table I as a function of spatial dimension $d$. The lattice with $d=2.46$, not having a spin-glass phase, is below the lower-critical dimension. For the 22 other consecutive lattices with a spin-glass phase, we have chosen to fit the runaway exponent values, since they gives an excellent, near-linear fit with $$y_R = -1.30908 + 0.528513 d - 0.00354805 d^2,$$ with an amazingly satisfactory correlation coefficient of $R^2 = 0.999999$. This fit gives, with a small extrapolation, $y_R = 0$ for $d = 2.520$. Note the near linearity, namely the smallness of the quadratic coefficient in Eq. (5). (In fact, a linear fit gives $y_R = 0$ for $d = 2.516$, with a little less amazingly satisfactory correlation coefficient of $R^2 = 0.999992$.) Our calculated lower-critical dimension $d_L$, where the spin-glass phase disappears at zero-temperature, is thus seen to be $d_L = 2.520$, for the sequence of hierarchical lattices studied here. It is noteworthy that $d_L$ is not an integer and not even a simple fraction, contrary previous examples of lower-critical dimensions (and even contrary to upper-critical dimensions, where mean-field behavior sets in) for other models. Another important quantity is the critical exponent $y_T = 1/\nu > 0$ of the phase transition between the disordered and spin-glass phases. This exponent is calculated from the scaling behavior of small deviations of the average interaction strength from its fixed finite value at the unstable fixed distribution of the phase transition. The calculated critical exponents are also given in Fig. 2. As the spatial dimension is lowered, $y_T$ also approaches 0. At the lower-critical dimension, $y_T$ reaches 0. The disordered-spin-glass phase transition disappears at $d_L$, where the spin-glass phase disappears. Conclusion ========== Our family of hierarchical lattices (Fig. 1) yields smooth and systematic behavior in all three quantities: the critical temperatures $1/J_C$, the critical exponents $y_C$, and, eminently fitably, the runaway exponents $y_R$. All three quantities yield the lower-critical temperature of $d_L = 2.520$. It is noteworthy that $d_L$ is not an integer and not even a simple fraction, contrary previous examples of lower-critical dimensions (and even contrary to upper-critical dimensions, where mean-field behavior sets in) for other models. We thank Prof. H. Nishimori for suggesting this calculation to us. Support by the Alexander von Humboldt Foundation, the Scientific and Technological Research Council of Turkey (TÜBITAK), and the Academy of Sciences of Turkey (TÜBA) is gratefully acknowledged. D. P. Belanger, A. R. King, and V. Jaccarino, Phys. Rev. Lett. [**48**]{}, 1050 (1982). H. Yoshizawa, R. A. Cowley, G. Shirane, R. J. Birgeneau, H. J. Guggenheim, and H. Ikeda, Phys. Rev. Lett. [**48**]{}, 438 (1982). P.-z. Wong and J. W. Cable, Phys. Rev. B [**28**]{}, 5361 (1983). A. N. Berker, Phys. Rev. B [**29**]{}, 5243 (1984). M. Aizenman and J. Wehr, Phys. Rev. Lett. [**62**]{}, 2503 (1989). M. Aizenman and J. Wehr, Phys. Rev. Lett. [**64**]{}, 1311(E) (1990). A. Falicov, A. N. Berker, and S.R. McKay, Phys. Rev. B [**51**]{}, 8266 (1995). S. Guchhait and R. Orbach, Phys. Rev. Lett. [**112**]{}, 126401 (2014). S. Franz, G. Parisi, and M.A. Virasoro, J. Physique I [**4**]{}, 1657 (1994). S. Boettcher, Phys. Rev. Lett. [**95**]{}, 197205 (2005). A. J. Bray and M. A. Moore, J. Phys. C [**12**]{}, 79 (1979). H. Sompolinsky and A. Zippelius, Phys. Rev. Lett. [**50**]{}, 1294 (1983). C. De Dominicis and I. Kondor, J. Physique Lett. [**45**]{}, L205 (1984). A. J. Bray and M. A. Moore, J. Phys. C [**17**]{}, L463 (1984). F. Matsubara, T. Shirakura, and M. Shiomi, Phys. Rev. B [**58**]{}, R11821 (1998). J. Houdayer, Eur. Phys. J. B [**22**]{}, 479 (2001). A. K. Hartmann and A. P. Young Phys. Rev. B [**64**]{}, 180404(R) (2001). C. Amoruso, E. Marinari, O. C. Martin, and A. Pagnani, Phys. Rev. Lett. [**91**]{}, 087201 (2003). B. W. Southern and A. P. Young, J. Phys. C [**10**]{}, 2179 (1977). A. N. Berker and S. Ostlund, J. Phys. C [**12**]{}, 4961 (1979). R. B. Griffiths and M. Kaufman, Phys. Rev. B [**26**]{}, 5022R (1982). M. Kaufman and R. B. Griffiths, Phys. Rev. B [**30**]{}, 244 (1984). S. R. McKay and A. N. Berker, Phys. Rev. B [**29**]{}, 1315 (1984). M. Hinczewski and A. N. Berker, Phys. Rev. E [**73**]{}, 066126 (2006). A. N. Berker and S. R. McKay, J. Stat. Phys. [**36**]{}, 787 (1984). A. A. Migdal, Zh. Eksp. Teor. Fiz. [**69**]{}, 1457 (1975) \[Sov. Phys. JETP [**42**]{}, 743 (1976)\]. L. P. Kadanoff, Ann. Phys. (N.Y.) [**100**]{}, 359 (1976). M. J. P. Gingras and E. S. S[ø]{}rensen, Phys. Rev. B. [**46**]{}, 3441 (1992). G. Migliorini and A. N. Berker, Phys. Rev. B. [**57**]{}, 426 (1998). M. J. P. Gingras and E. S. S[ø]{}rensen, Phys. Rev. B. [**57**]{}, 10264 (1998). M. Hinczewski and A.N. Berker, Phys. Rev. B [**72**]{}, 144402 (2005). C. Güven, A. N. Berker, M. Hinczewski, and H. Nishimori, Phys. Rev. E [**77**]{}, 061110 (2008). M. Ohzeki, H. Nishimori, and A. N. Berker, Phys. Rev. E [**77**]{}, 061116 (2008). V. O. Özçelik and A. N. Berker, Phys. Rev. E [**78**]{}, 031104 (2008). G. Gülpinar and A. N. Berker, Phys. Rev. E [**79**]{}, 021110 (2009). E. Ilker and A. N. Berker, Phys. Rev. E [**87**]{}, 032124 (2013). E. Ilker and A. N. Berker, Phys. Rev. E [**89**]{}, 042139 (2014). E. Ilker and A. N. Berker, Phys. Rev. E [**90**]{}, 062112 (2014). M. Kaufman and H. T. Diep, Phys. Rev. E [**84**]{}, 051106 (2011). J. Barre, J. Stat. Phys. [**146**]{}, 359 (2012). C. Monthus and T. Garel, J. Stat. Mech. - Theory and Experiment, P05002 (2012). Z. Z. Zhang, Y. B. Sheng, Z. Y. Hu, and G. R. Chen, Chaos [**22**]{}, 043129 (2012). S.-C. Chang and R. Shrock, Phys. Lett. A [**377**]{}, 671 (2013). Y.-L. Xu, L.-S. Wang, and X.-M. Kong, Phys. Rev. A [**87**]{}, 012312 (2013). S. Hwang, D.-S. Lee, and B. Kahng, Phys. Rev. E [**87**]{}, 022816 (2013). R. F. S. Andrade and H. J. Herrmann, Phys. Rev. E [**87**]{}, 042113 (2013). R. F. S. Andrade and H. J. Herrmann, Phys. Rev. E [**88**]{}, 042122 (2013). C. Monthus and T. Garel, J. Stat. Phys. - Theory and Experiment, P06007 (2013). O. Melchert and A. K. Hartmann, Eur. Phys. J. B [**86**]{}, 323 (2013). J.-Y. Fortin, J. Phys.-Condensed Matter [**25**]{}, 296004 (2013). Y. H. Wu, X. Li, Z. Z. Zhang, and Z. H. Rong, Chaos Solitons Fractals [**56**]{}, 91 (2013). P. N. Timonin, Low Temp. Phys. [**40**]{}, 36 (2014). B. Derrida and G. Giacomin, J. Stat. Phys. [**154**]{}, 286 (2014). M. F. Thorpe and R. B. Stinchcombe, Philos. Trans. Royal Soc. A - Math. Phys. Eng. Sciences [**372**]{}, 20120038 (2014). C. Monthus and T. Garel, Phys. Rev. B [**89**]{}, 184408 (2014). T. Nogawa and T. Hasegawa, Phys. Rev. E [**89**]{}, 042803 (2014). M. L. Lyra, F. A. B. F. de Moura, I. N. de Oliveira, and M. Serva, Phys. Rev. E [**89**]{}, 052133 (2014). V. Singh and S. Boettcher, Phys. Rev. E [**90**]{}, 012117 (2014). Y.-L. Xu, X. Zhang, Z.-Q. Liu, K. Xiang-Mu, and R. Ting-Qi, Eur. Phys. J. B [**87**]{}, 132 (2014). Y. Hirose, A. Oguchi, and Y. Fukumoto, J. Phys. Soc. Japan [**83**]{}, 074716 (2014). H. Nishimori, *Statistical Physics of Spin Glasses and Information Processing* (Oxford University Press, 2001). D. Andelman and A. N. Berker, Phys. Rev. B [**29**]{}, 2630 (1984).

ArXiv

--- abstract: 'In this article, we investigate a domination set problem variant on vertex-weighted graphs. In the last few years, several algorithms have been presented for solving the minimum alpha and alpha-rate domination problem (also known as the positive influence dominating sets problem) on simple graphs. We recently proposed an algorithm for alpha-rate domination on weighted graphs based on randomised rounding of the solution of a linear programming formulation of this problem. Due to the use of linear programming, such an algorithm could be relatively time consuming for larger graphs. Here, we propose a new version using the divide and conquer technique, which uses a graph’s community structure to create a solution from the solutions obtained on denser subgraphs (with some adjustments, if necessary). We also investigate greedy techniques for this problem using three different initial vertex selection strategies. We compare two different randomised rounding and three greedy algorithms on three different families of randomly generated graphs and on four real-world graphs obtained from a Twitter mentions network. Our results show that on dense random graphs the divide and conquer technique produces results comparable in total weight to the unembellished randomised rounding method, but significantly faster. For graphs with intrinsic modular structure, the divide and conquer technique actually produces better results. When the running time is prioritised over the optimality of results, two of the three explored greedy algorithms strategies perform better than the simple strategy of picking always vertices with the smallest weights. Also, greedy techniques outperform randomised algorithms on the very sparse Twitter graphs, and on the random dense graphs for high thresholds.' author: - 'Danica Vukadinović Greetham [^1]' - 'Anush Poghosyan[^2]' - 'Nathaniel Charlton[^3]' bibliography: - 'sigproc.bib' date: 'November 26, 2015' title: 'Algorithms for alpha-rate domination problems on weighted graphs ' --- [***Keywords—*** weighted graphs, alpha-rate domination, positive influence dominating sets, randomised rounding, planted $l$-partition graphs]{} Introduction {#intro} ============ Studies of interplay between networks’ structure and dynamical processes on them are becoming more and more popular due to the fact that a lot of complex systems that we interact with every day can be modelled by networks or graphs. Examples include infrastructure such as power, water or road networks, human communication networks, social media, network based behaviour change interventions [@Valente:2012] and so on. In some cases, we are interested to identify a set of vertices or edges from which we can communicate to all the vertices in a given graph, or from which we can control or dominate that graph. A set of vertices/individuals such that all other or indeed all individuals are connected to that set is called a [*dominating set*]{} in graph theory. The related optimisation problem of finding a dominating set of minimal size is NP-complete [@Garey:1979]. From the 1950s onward, different variants of this problem have been investigated. Based on problems in ad-hoc communications networks, $k$-domination was explored where each vertex not in the dominating set needs to have at least $k$ neighbours in the dominating set. Similarly, two versions having a parameter for the percentage of neighbours that need to be in a dominating set were proposed: $\alpha$ domination (where each vertex needs to have at least $\alpha*100$ percent of its neighbours in the dominating set) and $\alpha$-rate domination [@zve:upperbounds] where each vertex, including ones in the dominating set, needs to have at least $\alpha*100$ percent of its neighbours in the dominating set. Again, finding minimum cardinalities of $\alpha$ and $\alpha$-rate dominating sets is NP-complete. In this work, we investigate different algorithms for the $\alpha$-rate dominating set problem on weighted graphs, thus looking into a more general problem. Weighted graphs usually allow more realistic interpretations, and models and algorithms for weighted cases have wider application potential. Our motivation for the weighted version of the $\alpha$-rate domination problem comes mainly from the health and well-being related behaviour change context. Often, for an intervention to work for an individual, it is important to have a support network (see e.g. [@Greaves:2010]), so that positive messages can come from multiple sources. For this reason, an intervention designer might want to identify a support subset of the whole social network, which will be part of the intervention. It might be that the “best” candidates (from a structural perspective) for such a support subset are not feasible to be a part of intervention for various reasons: they do not have the desired attributes, or they do not have the capability or time to invest in the intervention. This can be represented by assigning a cost to be part of the intervention to each vertex. Then, the main task is to find the most cost effective subset from which we can control or support the network. In the next section we give an overview of the relevant previous work. In Section \[greedy\], we present three greedy algorithm variants based on different strategies for vertex selection into the weighted alpha-rate dominating set. The randomised algorithm using a linear programming formulation of the problem from [@greetham2014] is given and its new version that exploits the community structure of a graph is proposed in order to improve its running time in Section \[algRR1\]. We analyse the results obtained from the application of implemented greedy and randomised algorithms to three families of randomly generated graphs and Twitter mentions networks in Section \[comparison\]. Finally, we discuss our results and give some pointers to future work in Section \[discussion\]. Previous work ------------- Due to their suitability to a wide range of applications in networks design and control, variants of domination problems have been studied thoroughly. This includes a study of corresponding computational complexities for different variants and development of exact and approximation algorithms. The widely explored variants include the basic dominating set problem and its weighted version where weights are on vertices. The minimum weighted dominating set problem is one of the classic NP-hard optimisation problems in graph theory. Approximation algorithms for a special type of graphs, unit disk graphs with weights on vertices, were investigated in [@Zou:2011]. A generalisation of the domination set problem on vertex-weighted graphs, where the direct connections are replaced with shortest paths corresponding to some measure $f$ defined on the vertices of a graph, was explored in [@chen:approx]. The authors have used randomised rounding to prove the approximation ratio of $O(\log\Delta')$ for their algorithm, where $\Delta'$ is the maximum cardinality of the sets of vertices that can be dominated by any single vertex through the defined shortest paths. In [@Chen:2007] the maximum spanning star forest problem, the complement problem of domination set, is discussed and an $0.71$-approximation algorithm for this problem is given. In the vertex-weighted case, the ratio was $0.64$. Molnar et al. [@molnar:2014] proposed probabilistic dominating set selection strategies for large heterogeneous non-weighted graphs and explored how the structure of graphs influences performances of degree dependent probabilistic method based approximation algorithms and greedy algorithms. One generalisation of the domination problem, the so called $k$-dominating set problem, requires each vertex not in the dominating set to have at least $k$ neighbours in the set. Another variant, the so called $k$-tuple domination problem requires each vertex in the graph (even those in the dominating set) to have at least $k$ neighbours in the dominating set. These problems were relevant especially for ad-hoc and wireless networks routing (having more than one neighbour in a dominating set was providing more reliable connection). In [@Forster:2013] the author shows that a greedy approach for the minimum $k$-dominating set problem leads to an approximation ratio of $\ln(\Delta + k) + 1 < \ln(\Delta) + 1.7$, where $\Delta$ is the maximum degree of the graph. Klasing and Laforest [@Klasing:2004] proved the hardness of the $k$-tuple domination problem, even in restricted families of graphs and presented several interesting approximation algorithms. Another generalisation, the $\alpha$-domination problem, was introduced by Dunbar et al. in [@dunbar:alphadom], where each vertex not in the dominating set is required to have at least $\alpha*100$ percents of neighbours in the dominating set. Similarly, the concept of $\alpha$-rate domination [@zve:upperbounds] requires each vertex in the graph to have at least $\alpha*100$ percents of neighbours in the dominating set. Both the $\alpha$ and $\alpha$-rate domination problems are proven to be NP-complete. New upper bounds and randomised algorithms for finding the $\alpha$ and $\alpha$-rate domination sets in terms of the parameter $\alpha$ and graph vertex degrees on undirected simple finite graphs are provided by using the probabilistic method in [@zve:randomized] and [@zve:upperbounds]. Wang et al. [@Wang:2012] investigated the propagation of influence in the context of social networks. They introduced new variants of domination such as the positive influence dominating set (PIDS) and total positive influence dominating set (TPIDS). Actually, the definitions of PIDS and TPIDS problems are equivalent to $\alpha$-dominating and $\alpha$-rate dominating set problems respectively for the special case when $\alpha = 1/2$. Dinh et al. [@Dinh:2014] have generalised PIDS and TPIDS by allowing any $0<\alpha<1$, thus considering $\alpha$-dominating and $\alpha$-rate dominating set problems and presenting a linear time exact algorithm for trees, and approximation algorithms for PIDS and TPIDS within a factor $\ln \Delta + O(1)$, where $\Delta$ is the maximum degree of the graph. A smaller number of studies in domination parameters consider weighted graphs in particular. A variant of the weighted dominating set problem — the weighted minimum independent $k$-domination (WMIkD) problem — was studied by Yen in [@yen:2011]; an algorithm linear in the number of vertices of the input graph for the WMIkD problem on trees was given. In our previous work [@greetham2014], we discussed alpha-rate domination on vertex-weighted graphs. An algorithm based on randomised rounding of a linear programming formulation of the problem is given, and we have proven that its approximation ratio is $O(log_2(\Delta))$, where $\Delta$ is the maximum degree of the graph. Greedy algorithms {#greedy} ================= In this section, we consider greedy techniques for approximately solving the minimum weighted alpha-rate dominating set problem. We denote with $\bar{d}_v$ and $w_v$ the closed degree (degree plus one, including a vertex itself) and the weight of a vertex $v\in V$ respectively. All the vertices that $v$ is connected to together with $v$ are called the closed neighbourhood of $v$ and denoted with $N[v]$. The Algorithm 1 below describes a generic greedy algorithm to find a low-weighted alpha-rate dominating set $D$. A graph $G$, a real number $\alpha$, $0<\alpha\le 1$ An $\alpha$-rate (total positive influence) dominating set $D$ of $G$ Initialize $D=\emptyset$; set C=$N[v]-D$; Initialize $W=\emptyset$; compute $w_v$ the sum of weights of its closed neighbourhood; $W=W \cup \{w_v\}$; Sort $W$ using strategy S\[1,2,3\] Add the first $\lceil\alpha \bar{d}_v\rceil-r$ vertices from $W$ into $D$ $D$; As is expected with a greedy process, this does not necessarily yield the optimal solution. We consider and implement three different strategies on initial selection of vertices to be added to the $\alpha$-rate dominating set. Those are: S1: : sorting vertices in vertex weight ascending order; S2: : sorting by using a combination of degree and sum of weights of closed neighbourhoods, thus sorting in $\frac{w_v}{\bar{d}_v}$ ascending order; S3: : sorting in $\frac{w_v}{w_{N[v]}}$ ascending order, where $w_{v}$ is the weight, and $w_{N[v]}$ is the sum of weights of the neighbourhood for vertex $v$. The first strategy S1 focuses on optimising weight only. The second strategy S2 tries to balance minimising weight with minimising the size of the dominating set. Finally, the third strategy S3 is based on reasoning that it should be beneficial to take ‘light’ vertices with ‘heavy’ neighbourhoods as then less heavy neighbours will be needed in the dominating set. In all cases we need to keep track of $r=|N[v]\cap D|$ for each $v\in V(G)$ only up to $r=\lceil\alpha \bar{d}_v\rceil$. Since we may need to browse through all the neighbours of vertices in $V$, in total it can take $O(n^{2})$ steps to calculate all the necessary $|N[v]\cap D|$’s for each vertex $v\in V(G)$. Then computing and sorting a sum of weights of closed neighbourhood for each vertex can take $O(n^{2}\log n)$ steps in the worst case. Hence, in total, the set $D$ can be computed in $O(n^{2} \log n)$ steps. Randomised algorithms {#algRR1} ===================== Algorithm RR ------------ Recently, we proposed an approximation algorithm for the minimum weighted $\alpha$-rate dominating set problem [@greetham2014]. A linear programming relaxation of the original problem was used to obtain a preliminary solution. A randomised rounding of that solution was then repeated a number of times in order to obtain a feasible solution. The idea was based on the techniques used by Chen at al. [@chen:approx] for simple domination with measure functions (where adjacency may be replaced with limited length paths) on weighted graphs. Let us assume that for every vertex $v_i$, $1\le i \le n$ the variable $x_i$ has the following meaning: $x_i = 1$ if $v_i$ is contained in the $\alpha$-rate dominating set and $x_i = 0$ otherwise. We consider the following linear programming relaxation LP of an integer program IP: $$\begin{aligned} \label{lpr_weighted_dom} \min & \quad & \sum_{i=1}^n{w_ix_i}\\ \mbox{s.t.} & \quad & \sum_{v_j\in N[v_i]}x_j\ge \lceil \alpha \bar{d_{v_i}}\rceil, \quad \forall v_i \in V\\ &\quad& 0\le x_i \le 1, \quad 1\le i \le n.\end{aligned}$$ As we know $LP$ is polynomial-time solvable and we can compute an optimal solution $\{\widehat{x_i}\}_{1\le i \le n}.$ If we denote with $IP_{OPT}$ an optimal solution of the corresponding integer program IP we have that $$\label{opt_sol} IP_{OPT} \ge \sum_{i=1}^n w_i\widehat{x_i}.$$ We obtain a candidate $IP$ solution $\{x_i\}_{1\le i \le n}$ by using randomised rounding, setting $x_i = 1$ with probability $\widehat{x_i}$ and $0$ otherwise. Let $D$ be the set of vertices that are assigned ones after rounding, i.e. $D=\{v_i | x_i=1, 1\le i \le n \}.$ In the next step we estimate the probability that $D$ is a feasible solution for $IP$. For any vertex $v \in V$, with $d_v$ neighbours, let $k=\lceil \alpha \bar{d_{v_i}}\rceil$. We know that $\sum_{v_i \in N[v]} \widehat{x_i} \ge k$, and $\forall x_i, 0 \le \widehat{x_i} \le 1$. Now, the probability that $v_i$ is $\alpha$-rate dominated is equal to $$Pr(v_i \mbox{ is } \alpha \mbox{-rate dom.}) = 1- (Pr(v_i\mbox{ is not } \alpha\mbox{-rate dom.})).$$ We can look at the number of neighbours of $v$ (including $v$ itself) which are in $D$ as the sum of $\bar{d}_v$ independent trials, random processes, where the success probability of each trial $i$ is equal to $\widehat{x_i}$. Thus this sum, $|N[v]\cap D|$, follows Poisson’s binomial distribution [@Wang:2003] with parameters $\widehat{x_1},\ldots,\widehat{x_{\bar{d}_v}}$. Let $S=\{1,2,\ldots \bar{d}_v\}$, and $\mathcal{F}_k=\{A|A\subseteq S, |A|=k\}$ denote all subsets of $S$ with exactly $k$ members where we are going over all possible combinations. Then $|\mathcal{F}_k|=\binom{\bar{d}_v}{k}$. Let $A^C$ denote the complementary set, i.e. $S \setminus A$. $$\label{prob_vertex} Pr(v_i\mbox{ is not } \alpha\mbox{-rate dom.})=\sum_{l=0}^{k-1} \sum_{A\in \mathcal{F}_l}(\prod_{i \in A}\widehat{x_i})(\prod_{j \in A^c}(1-\widehat{x_j})).$$ $$Pr(v_i \mbox{ is not } \alpha\mbox{-rate dom.})<\frac{1}{2}. \label{notdominated}$$ Let the random variable $X$ be the number of neighbours that vertex $v$ has in $D$. Then $X$ follows Poisson’s binomial distribution with parameters $\widehat{x_1},\ldots,\widehat{x_{\bar{d}_v}}$: $$Pr(X = l)=\sum_{A\in \mathcal{F}_l}(\prod_{i \in A}\widehat{x_i})(\prod_{j \in A^c}(1-\widehat{x_j})).$$ Showing our goal  is equivalent to showing $$\frac{1}{2} \leq \sum_{l=k}^{\bar{d}_v} \sum_{A\in \mathcal{F}_l}(\prod_{i \in A}\widehat{x_i})(\prod_{j \in A^c}(1-\widehat{x_j})) = Pr(k \leq X). \label{notdominated2}$$ So we are looking for a minimum of the right hand side of  subject to $\sum_{i=1}^{\bar{d}_v} x_i \geq k$ (this minimum must exist by continuity and compactness). The minimum will be found when $\sum_{i=1}^{\bar{d}_v} x_i = k;$ increasing one of the $x_i$s without changing the others will clearly only increase the RHS. (Intuitively, increasing the probability of success in one of the trials, while leaving the others unchanged, can only increase the probability of getting at least $k$ successes.) Keep in mind that $0 \le \widehat{x_i} \le 1$ for all $\widehat{x_i}$. So we may assume that $$\label{constraint} \sum_{v_i \in N[v]} \widehat{x_i} = k.$$ Now we can use the result from [@Hoeffding:1956], Theorem 5, that shows that the tail distribution function of Poisson’s binomial distribution attains its minimum in the binomial distribution, i.e. when all probabilities are equal. The theorem states that for two integers $b$, and $c$ such that $0\le b\le np\le c\le n$, the probability $P(b \le X \le c)$ reaches its minimum where all the probabilities $p_1=\ldots =p_n=p$, unless $b=0$ and $c=n$. Here the $p_i$s are the probabilities (or parameters) of Poisson’s binomial distribution, and $n$ and $p$ are the parameters of the related binomial distribution. We apply that theorem taking the two integers $b$ and $c$ to be our $k$ and $\bar{d}_v$ respectively. We have that $p$, the equal probability is $\frac{k}{\bar{d}_v}$ from (\[constraint\]), whence $np$ equals our $k$. The theorem gives us $$\sum_{l=k}^{\bar{d}_v} \binom{\bar{d}_v}{l} p^l (1-p)^{\bar{d}_v-l} \quad\leq\quad Pr(k \leq X).$$ Thus, we will be done if we can show that $$\label{eqn:newgoal} \sum_{l=k}^{\bar{d}_v} \binom{\bar{d}_v}{l} p^l (1-p)^{\bar{d}_v-l}$$ is at least $\frac{1}{2}$. Let $Y$ be a random variable of binomial distribution with $\bar{d}_v$ trials each of probability $p$. Then observe that in fact $Pr(Y \geq k)$ is equal to  above. The median of $Y$ is bounded by $\lfloor \bar{d}_v p\rfloor$ and $\lceil \bar{d}_v p\rceil$ [@Kaas:1980], but $\bar{d}_v p$ is exactly the integer $k$, so $k$ is the unique median of $Y$. It follows from the defining property of medians that $Pr(Y \geq k) \geq \frac{1}{2}$, and thus $Pr(Y < k) < \frac{1}{2}$ and the proof is complete. Hence, the probability is lower bounded by $\frac 1 2$, and the feasibility follows. Let $A_i$ denote the event that vertex $v_i$ is $\alpha$-rate dominated and let $B=\cap_{i=1}^nA_i$ be the event that all vertices are dominated. We use the amplification approach (repeating randomised rounding $t=O(\log_2 \Delta)$ times), where $\Delta$ is the maximum degree of the graph as found in [@chen:approx] which results in $Pr([x_i=1])=1-(1-\hat{x_i})^t$. We obtain that the expected value of the solution resulting from randomised rounding, given that event $B$ happens, (i.e. that the solution is feasible) is $$\begin{aligned} \label{expect} E\left[\sum_{i=1}^nw_ix_i | B\right] &=& \sum_{i=1}^nw_iPr([x_i=1] | B)\\ &=& \sum_{i=1}^nw_i\frac{Pr(B|[x_i=1])}{Pr[B]} Pr(x_i=1) \\ &\le & \sum_{i=1}^nw_i \frac{1}{\prod_{j \in N[v_i]} Pr(A_j)}(1-(1-\hat{x_i})^t) \\ &\le&{\frac{1}{(1- 2^{-t})^{\Delta}}}\sum_{i=1}^n w_i (1-(1-t\hat{x_i}))\\ &\le&{\frac{t} {(1- 2^{-t})^{\Delta}}}\sum_{i=1}^n w_i \hat{x_i}\\ &\le&O(\log_2\Delta \cdot OPT).\end{aligned}$$ Hence, there exists a particular solution that is within $O(\log_2 \Delta)$ ratio to the optimal solution. Note that $C={\frac{1}{(1-\frac{1}{\Delta})^\Delta}}$ decreases monotonically down to $e$ with increasing $\Delta$ and assuming that $\Delta \ge 2$, the maximum is achieved for $\Delta = 2$, $C=4$. A simple randomised rounding algorithm AlgRR follows immediately, by first solving LP and then rounding the solutions to zero or one. This process is repeated $\lceil \log_2 \Delta \rceil$ times. All vertices with ones then create with high probability an $\alpha$-rate domination set with the sum of the weights within $O(\log_2 \Delta)$ factor of the optimal solution. Finally, if any vertex is still not $\alpha$-rate dominated, a required number of its neighbours are added to the solution. We implemented the algorithm in Python, using lpsolve55 [@lpsolve:2004] through its Python interface to solve linear programmes. We have drawn random numbers from the $[0, 0.5]$ interval because this has worked better in practice then drawing from $[0, 1]$. As solving a linear programme is running in $O(n^3)$, i.e. it is quite time-consuming for larger networks, we looked at some alternatives. Algorithm RRWC -------------- As the range and size of a network determine the size of the linear programme that needs to be solved, we investigated the following strategy. Firstly, we split a network into communities, then we solve a linear programme for each of communities, and use randomised rounding inside communities. Finally, we check if all the vertices are $\alpha$-rate dominated, and if not, we add the required number of neighbours of vertices that are still not dominated into the final solution. We implemented this algorithm in Python using NetworkX, and its module[@comms:2009] based on the Louvain method of community detection given in [@Blondel:2008]. In a way, we can look at this as a divide and conquer strategy, where we split the original problem into smaller ones, obtain the smaller problems’ solutions based on the previously shown technique, and finally take all those solutions and “patch” them globally in order to obtain a feasible solution for the whole network. We will denote this algorithm [**AlgRRWC**]{} (RRWC stands for randomised rounding with communities). A graph $G$, a real number $\alpha$, $0<\alpha\le 1$ An $\alpha$-rate (total positive influence) dominating set $D$ of $G$ Initialize $D=\emptyset$; violation$=1$ Split $G$ into communities $C_1, \ldots C_k$; solve LP; $\hat{x}=$lp.result; $r=random.uniform(0, 0.5)$ add $v_i$ to $D$ violation=0 violation$=1$ no-of-run++ add first $l$ neighbours not already in $D$ to $D$ $D$; Experimental results ==================== In order to test the performance of those three greedy and two randomised algorithms, we generated three types of random graphs and used some real-life networks obtained from Twitter. We ran the experiments on a 64-bit Windows 7 workstation with Intel i5-2400 CPU at 3.10GHz and 8GB of RAM. Random generated graphs ----------------------- We generated three different types of random graphs, with $100$ graphs of each type. They all had a similar number of vertices and edges and were created using methods from the NetworkX [@Hagberg:2008] package. Weights were assigned uniformly at random from integers between $1$ and $71$ (including the boundaries). The choice of numbers of vertices, edges and weights was informed by real Twitter networks described in Section \[Twittn\]. The weights for all the graphs listed here were created by picking uniformly a random number from $1$ to $71$[^4]. The average descriptive statistics for these networks are given in Table \[random-stats\]. ### Random graphs, ER type Normally used as a benchmark, our first type, [**ER graph**]{} is widely known as Erdös-Rényi model[@Bollobas:2001]. An ER or random graph is obtained by choosing uniformly at random from a family $\mathcal{G}(n, m)$ of all possible graphs on $n$ vertices with $m$ edges [@Bollobas:2001]. We used `dense_gnm_random_graph` method from NetworkX with parameters $n=5000$ and $m=50000$ to create those graphs and denote them with ER. ### Preferential attachment - high clustering graphs, PN type We used another NetworkX method `powerlaw_cluster_graph` to create graphs that result in approximate power-law degree distribution and high average clustering (we used parameters $n=5000, m=50000$, $0.8$ for probability of triangles)[@Holme:2002]. These graphs are denoted with *PN*. ### Planted $l$-partition graphs, PLP type Additionally, we created graphs that consisted of several interlinked modules or communities (in our case 5 communities with equal sizes of $1000$). In these graphs (also called planted $l$-partition graphs [@Condon:2001]) vertices in the same community or subgraph are interconnected with higher probability, in our case $p_{in}=0.02$ (this value provides each community similar to other types of graphs density) and vertices of different communities are connected with much smaller probability, in our case $p_{out}=0.0001$. We used `random_partition_graph` NetworkX method. This results in graphs having recognisable modular or block structure - with a lot of links inside those $5$ communities and only few links between different communities. ------- ------ ------- ---- ---------- ---------- ---------------- ----------- ----------- ----------- -- Graph V E CC $\delta$ $\Delta$ $\delta_{avg}$ K$_{min}$ K$_{max}$ K$_{avg}$ ER 5000 50000 1 6 38 20 1 71 39 PN 5000 49847 1 9 37 20 1 71 39 PLP 5000 50980 1 6 38 20 1 71 39 ------- ------ ------- ---- ---------- ---------- ---------------- ----------- ----------- ----------- -- \[random-stats\] Twitter mentions networks {#Twittn} ------------------------- We used four undirected weekly graphs from Twitter in period of December 2011 to January 2012 obtained from Datasift, where there is an edge between user A and user B if A tweeted at least one message containing “@B” during that week, and B reciprocated at least once. Those networks had around $5$k vertices and around $3.5$k edges on average. For each vertex we retrieved its *Klout* score and used it as the weight. The Klout score measures an individual’s influence based on her/his social media activity[^5]. It is a single number that represents the aggregation of multiple pieces of data about individuals’ social media activity, based on a score model which is not publicly available [@klout:2013]. The descriptive statistics of the Twitter mentions weekly graphs are given in Table \[stats\] below. -------- ------ ------ ------ ---------- ---------- ---------------- ----------- ----------- ----------- -- -- Graph V E CC $\delta$ $\Delta$ $\delta_{avg}$ K$_{min}$ K$_{max}$ K$_{avg}$ twitt1 5775 3716 2174 1 16 1.2869 10 71 33 twitt2 5537 3537 2094 1 19 1.2776 10 71 34 twitt3 5279 3434 1957 1 15 1.3010 10 71 34 twitt4 5597 3599 2093 1 16 1.2860 10 71 33 -------- ------ ------ ------ ---------- ---------- ---------------- ----------- ----------- ----------- -- -- \[stats\] Comparison ---------- \[evaluation\_greedy\] In this section we compare the performances of three initial vertex selection strategies for our greedy algorithm: S1, vertices sorted in weights ascending order (AlgG\_W); S2, vertices sorted in ascending order according to the ratio of vertex weight to vertex degree (AlgG\_W/D); S3, vertices sorted in ascending order according to the ratio of vertex weight to the sum of weights of vertex’s open neighbourhood (AlgG\_W/Wn). Tables \[AlgGreedyRandom\] and \[AlgGreedyTwitter\] contain the average results on all the random types of graphs (denoted with ER, PN, PLP) and results on *twitt1-4* obtained for AlgG\_W, AlgG\_W/D and AlgG\_W/Wn correspondingly. ![Comparison of sum of weights of $0.25$-rate domination sets for AlgG\_W/D and AlgG\_W/Wn, AlgRR and AlgRRWC for random graphs.[]{data-label="fig:weight_greedy"}](plot1.png){width="\textwidth"} \[AlgGreedyRandom\] From Table \[AlgGreedyRandom\] we can see that S3 strategy wins in all bar one cases (S2 is somewhat better for PN graphs when threshold is $0.25$. Also, we can see that while S2 and S3 are comparable, S1 is visibly inferior to both - the average running time is larger while it gives worst results.We can see from Table \[AlgRR\_random\] that as expected, in random ER and preferential PN type of graphs, especially with the lowest threshold, AlgRR performs better than the alternative version AlgRRWC. However, when the graphs have clear modular structure such as PLP graphs, we see that AlgRRWC outperforms AlgRR and it does it with much smaller average running times. When we compare these results with Table \[AlgGreedyRandom\] we see that AlgRRWC has about double running times compared with S3 greedy algorithm, but also much smaller weights of $\alpha$-rate dominating sets, for lower thresholds and especially for PLP graphs (see Fig \[fig:weight\_greedy\]). [ll|lll|lll]{} & & & AlgRR& & &AlgRRWC&\ Graph& $\alpha$& \# &W &Time(s)& \# &W &Time(s)\ ER&0.25& 2188& [**47577**]{}& 313.49& 2214& 54373& 11.96\ ER& 0.5& 3417& [**103255**]{}& 774.76& 3379& 104245& 18.04\ ER&0.75& 4570& 168394& 1081.24& 4492& [**165894**]{}& 23.44\ PN& 0.25& 1334& [**26732**]{}& 64.84& 1444& 32683& 7.72\ PN&0.5& 2573& [**69321**]{}& 147.26& 2681& 75655& 14.27\ PN&0.75& 4134& [**144827**]{}& 276.77& 4126& 146125& 21.77\ PLP&0.25& 2188& 477568& 237.02& 1921& [**37431**]{}& 10.34\ PLP&0.5& 3423& 103552& 582.64& 3116& [**87823**]{}& 16.49\ PLP&0.75& 4600& 170058& 843.21& 4334& [**154137**]{}& 22.79\ \[AlgRR\_random\] Table \[AlgGreedyTwitter\] list results for Twitter mention networks. Here S2 is more competitive comparing with S3, it wins in half of the cases, but we see from the results that S2 and S3 produce similar results and their average running times are similar. Again S1 is obviously inferior to both S2 and S3. Greedy algorithms outperform both randomised algorithms (See also Table \[AlgRR\_twitter\].) \[AlgGreedyTwitter\] [ll|lll|lll]{} & & & AlgRR& & &AlgRRWC&\ Graph& $\alpha$& \# &W &Time(s)& \# &W &Time(s)\ twitt1&0.25& 4655& 154355& 11.31& 4657& [**154411**]{}& 6.74\ twitt1&0.5& 4837& [**160658**]{}& 11.85& 4838& 160666& 7.00\ twitt1&0.75& 5665& [**191648**]{}& 12.68& 5668& 191762& 8.19\ twitt2&0.25& 4469& 148980& 11.19& 4468& [**148941**]{}& 6.32\ twitt2&0.5& 4623& 154268& 10.90& 4623& [**154267**]{}& 6.51\ twitt2&0.75& 5431& [**184481**]{}& 11.65& 5434& 184609& 7.55\ twitt3&0.25& 4220& [**141094**]{}& 9.58& 4223& 141228& 5.64\ twitt3&0.5& 4391& [**146869**]{}& 9.94& 4397& 147115& 5.86\ twitt3&0.75& 5159& [**175854**]{}& 10.54& 5164& 176040& 6.88\ twitt4&0.25& 4471&[**148289**]{}& 10.75& 4471& [**148289**]{}& 6.31\ twitt4&0.5& 4651&[**154067**]{}& 11.08& 4653& 154111& 6.56\ twitt4&0.75& 5468&[**184486**]{}& 12.02& 5470& 184544& 7.69\ \[AlgRR\_twitter\] Conclusions {#discussion} ----------- We presented two types of algorithms for solving the minimum weighted alpha-rate domination problem. Our contributions are threefold: firstly, we propose three greedy strategies for this problem; secondly we propose a new randomised algorithm which uses linear programming formulation on the parts of a graph respectively and then recombines it; and finally, we test all proposed algorithms on some sparse real and three different families of dense random graphs. We were able to identify the winning greedy strategy selecting ‘light’ vertices with ‘heavy’ neighbourhoods on dense random graphs. On the sparse Twitter mention graphs in addition to the previously mentioned, another greedy strategy - using ‘light’ vertices with large degrees performed satisfactorily. Therefore, we were able to reject a simple greedy strategy of always putting neighbours with least weights into dominating set. Instead, sorting neighbours by weight over degree or weight over sum of neighbours’ weights works better. When comparing two randomised algorithms, we have seen that solving linear programmes for the parts of graph and patching the solutions obtained to create a feasible solution works much quicker for the graphs of 5k vertices and larger, as expected. Also as expected, the obtained results were worse than for solving linear programme on a whole graph (except for the graphs with clear community structure, or for very high thresholds of neighbours that need to be in the dominating set). On the other hand, results obtained by this technique are still superior to the greedy solutions for denser random networks, for lower thresholds. For that reason, the algorithm AlgRRWC is useful as a faster alternative to the AlgRR for larger graphs, especially if they have modular structure, and as slower but more efficient alternative to greedy techniques for denser graphs and lower thersholds. For higher thresholds and very sparse graphs like the ones we obtained from Twitter, we suggest to use one of the two greedy strategies mentioned above. Regarding future directions, while we were able to prove the approximation ratio of AlgRR, we would like to do the same for AlgRRWC. In addition it would be interesting to explore in more detail the structure and modularity of graphs that result in better performance of AlgRR or AlgRRWC and the level of sparseness at which the greedy techniques outperform them. ### Acknowledgments. {#acknowledgments. .unnumbered} This work is partially funded by the RCUK Digital Economy programme via EPSRC grant EP/G065802/1 ‘The Horizon Hub’. We would like to thank Datasift for providing us with the Twitter dataset. [^1]: [email protected] [^2]: [email protected] [^3]: [email protected] [^4]: The largest weight in the Twitter graphs was $71$; we made the weights for our randomly generated graphs comparable. [^5]: In Twitter, Klout focuses on retweets of a user’s tweets, their username mentions by other users, their list memberships on other users’ curated lists, the number of followers and the number and frequency of replies i.e. how engaged they are.

ArXiv

--- abstract: 'We present conditions which guarantee a parametrization of the set of positive equilibria of a [*generalized*]{} mass-action system. Our main results state that (i) if the underlying generalized chemical reaction network has an [*effective deficiency*]{} of zero, then the set of positive equilibria coincides with the parametrized set of complex-balanced equilibria and (ii) if the network is weakly reversible and has a [*kinetic deficiency*]{} of zero, then the equilibrium set is nonempty and has a positive, typically rational, parametrization. Via the method of network translation, we apply our results to [*classical*]{} mass-action systems studied in the biochemical literature, including the EnvZ-OmpR and shuttled WNT signaling pathways. A parametrization of the set of positive equilibria of a (generalized) mass-action system is often a prerequisite for the study of multistationarity and allows an easy check for the occurrence of absolute concentration robustness (ACR), as we demonstrate for the EnvZ-OmpR pathway.' author: - 'Matthew Johnston [^1]' - 'Stefan Müller [^2]' - 'Casian Pantea [^3]' title: ' A deficiency-based approach to parametrizing positive equilibria of biochemical reaction systems ' --- **Keywords:** chemical reaction network, chemical kinetics, deficiency, equilibrium **AMS Subject Classifications:** 92C42, 34A34 Introduction {#sec:intro} ============ Networks of biochemical reactions can be represented as directed graphs where the vertices are combinations of interacting species (so-called complexes) and the edges are the reactions. Under suitable assumptions, such as spatial homogeneity and sufficient dilution, the networks follow mass-action kinetics and give rise to a system of polynomial ordinary differential equations in the species concentrations. The mathematical study of positive equilibria of mass-action systems is important for establishing the uniqueness of equilibria in invariant regions of the state space (so-called compatibility classes) or, conversely, for establishing the capacity for multistationarity (for example, in models of biological switches). Such analysis, however, is challenging due to the high-dimensionality of the dynamical system, the significant nonlinearities, and the number of (unknown) parameters. Recent work has consequently focused on developing network-based methods for parameterizing the set of positive equilibria. Monomial parametrizations (Laurent monomials) have been established in [@C-D-S-S; @M-D-S-C; @MR2014; @J2014], while rational parametrizations have been constructed in [@T-G; @G-H-R-S; @MD2017]. Based on parametrizations, the uniqueness of positive equilibria has been analyzed [@C-F-R; @MR2012; @M-F-R-C-S-D; @Conradi2017], and regions for multistationarity (in the space of rate constants) have been identified for specific models, such as phosphorylation networks [@H-F-C; @C-F-M-W2016]. In this paper, we develop a method for explicitly constructing positive, typically rational, parametrizations of the set of positive equilibria for a broad class of biochemical reaction networks. Our approach is based on an extension of deficiency theory, the concept of generalized mass-action systems, and the method of network translation. The deficiency of a chemical reaction network was introduced in [@F1; @H] in the context of sufficient conditions for weakly reversible networks to have complex-balanced equilibria [@H-J1]. The notions of deficiency and complex balancing were subsequently extended to generalized mass-action systems in [@MR2012; @MR2014]. Thereby the kinetic complex determining the reaction rate was allowed to differ from the (stoichiometric) complex determining the reaction vector. Finally, the method of network translation was introduced in [@J2014], in order to relate a mass-action system to a generalized mass-action system that is dynamically equivalent, but has a different network structure. In particular, the translated network might be weakly reversible (even when the original network is not) and have a lower deficiency. A generalized mass-action system for which the underlying network is weakly reversible and has deficiency zero is known to have an equilibrium set with a monomial parametrization [@MR2012; @MR2014; @J2014]. In this paper, we extend this framework to construct positive parametrizations for a significantly wider class of generalized networks. To this end, we introduce a new notion of deficiency called *effective deficiency* based on the *condensed network* of the generalized network. Our main results state that, if a weakly reversible generalized network has an effective deficiency and kinetic deficiency of zero, then the corresponding generalized mass-action system permits a positive parametrization of the set of positive equilibria. This parametrization can be computed by linear algebra techniques and does not require tools from algebraic geometry such as Gröbner bases. Via network translation, we can apply our results to a broad class of mass-action systems. For example, consider the following two-component signaling system, which is adapted from a histidine kinase example in [@C-F-M-W2016]: $$\label{intro-example} \begin{tikzcd} X \arrow[r, "k_1"] & X_p \\[-0.15in] X_p + Y \arrow[r, yshift=+0.5ex,"k_2"] & X + Y_p \arrow[l, yshift=-0.5ex,"k_3"] \\[-0.15in] Y_p \arrow[r, "k_4"] & Y \end{tikzcd}$$ Thereby $X$ is a histidine kinase, $Y$ is a response regulator, and $p$ is a phosphate group. The network is not weakly reversible and has deficiency one. Via network translation, the system  corresponds to the following generalized mass-action system: $$\label{example2} \begin{tikzcd} \ovalbox{$\begin{array}{c} X + Y \\ (X) \end{array}$} \arrow[r,"k_1"] & \ovalbox{$\begin{array}{c} X_p + Y \\ (X_p+Y) \end{array}$} \arrow[d,xshift=+0.5ex,"k_2"] \\ \ovalbox{$\begin{array}{c} X + Y_p \\ (Y_p) \end{array}$} \arrow[u,"k_4"] & \ovalbox{$\begin{array}{c} X + Y_p \\ (X+Y_p) \end{array}$} \arrow[u,xshift=-0.5ex,"k_3"] \arrow[l,color=red,"{\sigma}"] \end{tikzcd}$$ Thereby we put a box at each vertex of the graph with the *stoichiometric complex* at the top and the *kinetic complex* (in brackets) at the bottom. The red arrow corresponds to a *phantom edge*, that is, an edge which connects identical stoichiometric complexes. Phantom edges do not contribute to the associated system of ordinary differential equations and hence can be labeled arbitrarily. Thus the edge label ${\sigma}> 0$ can be considered a free parameter. Now, the network  is weakly reversible and, as it turns out, it has an *effective deficiency* of zero and a *kinetic deficiency* of zero. Our main results guarantee that the set of positive equilibria has a positive parametrization and, in fact, constructively yield the following rational parametrization: $$\label{intro:param} \left \{ \quad \begin{array}{ll} x = \displaystyle{\frac{k_4}{{\sigma}}} , \quad & x_p = \displaystyle{\frac{k_1(k_3+{\sigma})k_4}{k_2{\sigma}^2{\tau}}} , \\ y = {\tau}, & y_p = \displaystyle{\frac{k_1}{{\sigma}}} ,\end{array} \right.$$ where ${\sigma}, {\tau}> 0$. Note that the ‘rate constant’ ${\sigma}>0$ in the network  appears explicitly in the parametrization . Importantly, the construction of via Theorem \[main2\] depends on efficient methods from linear algebra such as generalized inverses. Our algorithm therefore represents a significant computational advantage over algebraic geometry methods such as Gröbner bases. The paper is organized as follows. In Section \[sec:background\], we review the relevant terminology regarding generalized chemical reaction networks and introduce several new notions, including effective and phantom edges, parametrized sets of equilibria, condensed networks, and effective deficiency. In Section \[sec:results\], we present the crucial Lemma \[cones\] and the main results of the paper, Theorems \[main\] and \[main2\]. In Section \[sec:applications\], we discuss the method of network translation, which allows us to apply the results of Section \[sec:results\] to networks studied in the biochemical literature, such as the EnvZ-OmpR and shuttled WNT signaling pathways. In the EnvZ-OmpR example, our parametrization immediately implies the occurrence of absolute concentration robustness (ACR). In Section \[sec:conclusions\], we summarize our findings and present avenues for future work. Throughout the paper, we use the following notation: - *Vector logarithm:* for $v \in {\mathbb{R}}^{n}_{>0}$, $\ln v = (\ln v_1, \ldots, \ln v_n)^T \in {\mathbb{R}}^{n}$. - *Vector exponential:* for $v \in {\mathbb{R}}^{n}$, $e^v = (e^{v_1}, \ldots, e^{v_n})^T \in {\mathbb{R}}^{n}$. - *Vector powers:* for $v \in {\mathbb{R}}^n_{>0}$ and $w \in {\mathbb{R}}^{n}$, $v^w = \prod_{i=1}^n v_i^{w_i} \in {\mathbb{R}}$. - *Matrix powers:* for $v \in {\mathbb{R}}^{n}_{>0}$ and $A \in {\mathbb{R}}^{n \times m}$, $v^{A^T} = (\prod_{i=1}^n v_i^{A_{i1}},\ldots,\prod_{i=1}^n v_i^{A_{im}})^T \in {\mathbb{R}}^m$. - *Hadamard product:* for $v, w \in {\mathbb{R}}^{n}$, $v \circ w = (v_1w_1, \ldots, v_nw_n)^T \in {\mathbb{R}}^{n}$. Mathematical framework {#sec:background} ====================== We give a brief introduction to the relevant terminology regarding generalized chemical reaction networks (which include classical chemical reaction networks). In particular, we distinguish between effective and phantom edges and introduce parametrized sets of equilibria. Further, we define condensed networks and the notion of effective deficiency. Finally, we introduce the helpful concept of [$V^*$-directed]{}networks. Generalized mass-action systems {#sec:gcrn} ------------------------------- A directed graph $G=(V,E)$ is given by a set of vertices $V = \{ 1, \ldots, m \}$ and a set of edges $E \subseteq V \times V$. We denote an edge $e=(i,j)\in E$ by $i \to j$ to emphasize that is directed from the source $i$ to the target $j$. We additionally define the set of *source vertices* $V_s = \{ i \mid i \to j \in E \}$, that is, the set of vertices that appear as the source of some edge. We call the connected components of a graph *linkage classes* and the strongly connected components *strong linkage classes*. If linkage classes and strong linkage classes coincide, we call the graph *weakly reversible*. A generalized chemical reaction network is essentially a graph with two embeddings of the vertices in ${\mathbb{R}}^n$. The notion was introduced in [@MR2012] and extended to the present form in [@MR2014]. \[def:gcrn\] A *generalized chemical reaction network* (GCRN) $(G,y,\tilde{y})$ is given by a directed graph $G=(V,E)$ without self-loops and two maps $y \colon V \to {\mathbb{R}}^n$ and $\tilde{y} \colon V_s \to {\mathbb{R}}^n$. Thereby, $G$ is called the abstract reaction graph, and $y(i), \tilde y(i) \in {\mathbb{R}}^n$ are called the stoichiometric and kinetic-order complexes, respectively, assigned to vertex $i$. In contrast to a classical chemical reaction network (see below), a GCRN has two complexes associated to each vertex. Thereby, the maps $y$ and $\tilde{y}$ are not required to be injective, and the same stoichiometric or kinetic-order complex may be assigned to several vertices. When considering examples, we represent complexes $y, \tilde y \in {\mathbb{R}}^n$ as formal sums of *species* (often $\{X_1, X_2, \ldots, X_n \}$). The components of the complexes correspond to the coefficients in the sums, e.g., $y = (1,0,1,0,\ldots,0)$ is represented as $y = X_1 + X_3$. \[def:gmas\] A [*generalized mass-action system*]{} (GMAS) $(G_k,y,\tilde{y})$ is given by a GCRN $(G,y,\tilde{y})$ with $G=(V,E)$ together with edge labels $k \in {\mathbb{R}}^E_{>0}$, resulting in the labeled directed graph $G_k$. That is, every edge $i \to j \in E$ is labeled with a rate constant $k_{i \to j} \in {\mathbb{R}}_{>0}$. The ODE system associated with a GMAS is given by $$\label{gmas} {\frac{\text{d} x}{\text{d} t}} = f^{G}_k(x) = \sum_{i \to j \in E} k_{i \to j} \, x^{\tilde y(i)} \, (y(j) - y(i)).$$ We can rewrite the right-hand side of the ODE as $$\label{gmas_fk} f^{G}_k(x) = Y I_E \, \text{diag}(k) (I_E^s)^T \, x^{\tilde Y} = Y A^G_k \, x^{\tilde Y} ,$$ where $Y, \tilde Y \in {\mathbb{R}}^{n \times V}$ are the matrices of stoichiometric and kinetic complexes, respectively, $I_E, I_E^s \in {\mathbb{R}}^{V \times E}$ are the [*incidence*]{} and [*source*]{} matrices of the graph $G$, and $A^G_k \in {\mathbb{R}}^{V \times V}$ is the resulting [*Laplacian matrix*]{} of the labeled directed graph $G_k$. Note that columns $\tilde y^j$ of $\tilde Y$ corresponding to non-source vertices $j \not\in V_s$ can be chosen arbitrarily since the corresponding columns $(I^s_E)^j$ of $I^s_E$ and hence the columns $(A_k^G)^j$ of $A_k^G$ are zero vectors. Notably, the change over time  lies in the [*stoichiometric subspace*]{} $S = \operatorname{im}(Y I_E)$, which suggests the definition of a [*stoichiometric compatibility class*]{} $(c' + S) \cap {\mathbb{R}}^n_{\ge 0}$ with $c' \in {\mathbb{R}}^n_{\ge 0}$. The [*stoichiometric deficiency*]{} is defined as $\delta = \dim (\ker Y \cap \operatorname{im}I_E)$. Equivalently, $\delta = m - \ell - s$, where $m = |V|$ is the number of vertices, $\ell$ is the number of linkage classes of $G$, and $s = \dim S$ is the dimension of the stoichiometric subspace (for example, see [@J2014]). If $V=V_s$, that is, if every vertex is a source, we additionally define the [*kinetic-order subspace*]{} $\tilde S = \operatorname{im}( \tilde Y I_E)$ and the [*kinetic deficiency*]{} $\tilde \delta = \dim (\ker \tilde Y \cap \operatorname{im}I_E)$. Equivalently, $\tilde \delta = m - \ell - \tilde s$, where $\tilde s = \dim \tilde S$ is the dimension of the kinetic-order subspace. \[example3\] Consider the GCRN $(G,y,\tilde y)$ with the abstract graph $G = (V,E)$ given by $$\label{example10} \begin{tikzcd} 1 \arrow[r] & 2 \arrow[d,xshift=0.5ex] \\ 4 \arrow[u] & 3 \arrow[u,xshift=-0.5ex] \arrow[l] \end{tikzcd}$$ and $y$ and $\tilde y$ defined by $$y(1) = X_1+X_2, \, y(2) = X_2 + X_3, \, y(3) = y(4) = X_1 + X_4,$$ and $$\tilde y(1) = X_1, \, \tilde y(2) = X_2 + X_3, \, \tilde y(3) = X_1 + X_4, \, \tilde y(4) = X_4 .$$ This generalized network has four vertices in one linkage class and is weakly reversible. It has a two-dimensional stoichiometric subspace ($s = 2$) and a three-dimensional kinetic-order subspace ($\tilde s = 3$). It follows that the stoichiometric deficiency is one ($\delta = 4-1-2 = 1$) while the kinetic deficiency is zero ($\tilde \delta = 4-1-3 = 0$). The corresponding generalized mass-action system $(G_k,y)$ gives rise to the following system of ODEs $$\label{ex:dyn} {\frac{\text{d} x}{\text{d} t}} = f^{G}_k(x) = Y A_k^G x^{\tilde Y} = \left[ \begin{array}{cccc} 1 & 0 & 1 & 1 \\ 1 & 1 & 0 & 0 \\ 0 & 1 & 0 & 0 \\ 0 & 0 & 1 & 1 \end{array} \right] \left[ \begin{array}{cccc} -k_{1 \to 2} & 0 & 0 & k_{4 \to 1} \\ k_{1 \to 2} & -k_{2 \to 3} & k_{3 \to 2} & 0 \\ 0 & k_{2 \to 3} & -k_{3 \to 2}-k_{3 \to 4} & 0 \\ 0 & 0 & k_{3 \to 4} & -k_{4 \to 1} \end{array} \right] \left[ \begin{array}{c} x_1 \\ x_2 x_3 \\ x_1 x_4 \\ x_4 \end{array} \right].$$ Alternatively, we represent the abstract graph  and the maps $y$ and $\tilde y$ together in one graph, $$\label{exampleXXX} \begin{tikzcd} \ovalbox{$\begin{array}{c} 1 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} X_1 + X_2 \\ (X_1) \end{array}$} \arrow[r] & \ovalbox{$\begin{array}{c} 2 \\ \\ \end{array} \Bigg\lvert\begin{array}{c} X_2 + X_3 \\ (X_2+X_3) \end{array}$} \arrow[d,xshift=+0.5ex] \\ \ovalbox{$\begin{array}{c} 4 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} X_1 + X_4 \\ (X_4) \end{array}$} \arrow[u] & \ovalbox{$\begin{array}{c} 3 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} X_1 + X_4 \\ (X_1+X_4) \end{array}$} \arrow[u,xshift=-0.5ex] \arrow[l] \end{tikzcd}$$ where at each vertex we put a box with the vertex of the abstract graph (if required) on the left, the stoichiometric complex $y$ at the top, and the kinetic complex $\tilde y$ (in brackets) at the bottom. Note that the network  in the introduction is essentially the network  with specific interpretations of the species $X_1, X_2, X_3,$ and $X_4$ and of the edge labels $k_{i \to j}$ for $i \to j \in E$. Mass-action systems ------------------- Classical chemical reaction networks and mass-action systems, which have been studied extensively in industrial chemistry and systems biology, can be considered as special cases of Definitions \[def:gcrn\] and \[def:gmas\]. \[crn\] A *chemical reaction network* (CRN) $(G,y)$ is a GCRN $(G,y,\tilde y)$ with $\tilde y = y$ and $y \colon V \mapsto {\mathbb{R}}^n$ being injective. A mass-action system (MAS) $(G_k,y)$ is given by a CRN $(G,y)$, where $G=(V,E)$, together with a vector $k \in {\mathbb{R}}^E_{>0}$, resulting in the labeled directed graph $G_k$. The ODE system associated with a MAS is given by $$\label{mas} {\frac{\text{d} x}{\text{d} t}} = f^{G}_k(x) := \sum_{i \to j \in E} k_{i \to j} \, x^{y(i)} \, (y(j) - y(i)) = Y I_E \, \text{diag}(k) (I^s_E)^T \, x^{Y} = Y A_k^G \, x^{Y}$$ where the matrices $Y$, $I_E$, $I^s_E$, and $A_k^G$ are as in . For a CRN, the stoichiometric and kinetic-order subspaces coincide (i.e. $S = \tilde S$), and the stoichiometric and kinetic deficiencies are the same (i.e. $\delta = \tilde \delta$). In fact, the *deficiency* $\delta = \dim( \ker Y \cap \operatorname{im}I_E) = m - \ell - s$ was introduced first in [@F1; @H] in the context of complex-balanced mass-action systems [@H-J1]. It has been studied extensively since then [@Fe2; @F2; @F3; @Sh-F]. In a CRN, the map $y$ is unique and vertices and complexes are in one-to-one correspondence. It is typical to write the reaction graph $G$ with the complexes as vertices. Recall the CRN $(G,y)$  from the introduction, where the species and rate constants $k \in {\mathbb{R}}_{>0}^E$ have been relabeled as follows: $$\label{example237} \begin{tikzcd} X_1 \arrow[r, "k_{1 \to 2}"] & X_3 \\[-0.15in] X_2 + X_3 \arrow[r, yshift=+0.5ex,"k_{3 \to 4}"] & X_1 + X_4 \arrow[l, yshift=-0.5ex,"k_{4 \to 3}"] \\[-0.15in] X_4 \arrow[r, "k_{5 \to 6}"] & X_2 \end{tikzcd}$$ The CRN  has six vertices in three linkage classes, and is not weakly reversible. It has a stoichiometric subspace of dimension two ($s=2$), and hence its deficiency is one ($\delta = 6 - 3 - 2 = 1$). The system of ODEs  associated with the MAS $(G_k,y)$  is $$\label{ex:example237} \left\{ \quad \begin{aligned} \dot{x}_1 & = -k_{1 \to 2} x_1 + k_{3 \to 4} x_2 x_3 - k_{4 \to 3} x_1 x_4 \\ \dot{x}_2 & = -k_{3 \to 4} x_2 x_3 + k_{4 \to 3} x_1 x_4 + k_{5 \to 6} x_4 \\ \dot{x}_3 & = k_{1 \to 2} x_1 - k_{3 \to 4} x_2 x_3 + k_{4 \to 3} x_1 x_4 \\ \dot{x}_4 & = k_{3 \to 4} x_2 x_3 - k_{4 \to 3} x_1 x_4 - k_{5 \to 6} x_4. \end{aligned} \right.$$ Notably, after expanding and relabeling the rate constants in the ODE system  arising from the GCRN , the ODE system  arising from the CRN  coincides with . Results obtained by a structural analysis of the GCRN  will consequently hold for the CRN . In particular, we will investigate existing methods for corresponding MASs and GMASs with equivalent dynamics,  and , respectively, in Section \[sec:translation\]. Effective and phantom edges and parametrized sets of equilibria --------------------------------------------------------------- For a GCRN, only edges $i \to j \in E$ with $y(j) \neq y(i)$ contribute to the right-hand side of the ODE . In Example \[example3\], $y(3)=y(4)$, and hence the rate constant $k_{3 \to 4}$ does not appear in the ODEs , even though $3 \to 4 \in E$. Consequently, we may partition the set of edges $E$ into the set of [*effective edges*]{} $$E^* = \{ i \to j \in E \mid y(i) \neq y(j) \}$$ and the set of [*phantom edges*]{} $$E^0 = \{ i \to j \in E \mid y(i) = y(j) \}.$$ Obviously, $E^* \cap E^0 = \emptyset$ and $E = E^* \cup E^0$. For a vector $k \in {\mathbb{R}}^E_{>0}$, we define $k^* = k_{E^*} \in {\mathbb{R}}^{E^*}_{>0}$ and $k^0 = k_{E^0} \in {\mathbb{R}}^{E^0}_{>0}$ so that $k=(k^*,k^0)$. Further, we introduce the effective graph $G^* = (V,E^*)$. From  it follows that $$\label{gmas_simplified} f^{G}_k(x) = \sum_{i \to j \in E} k_{i \to j} \, x^{\tilde y(i)} \, (y(j) - y(i)) = \sum_{i \to j \in E^*} k_{i \to j} \, x^{\tilde y(i)} \, (y(j) - y(i)) = f^{G^*}_{k^*}(x) .$$ That is, the GMAS $(G_k,y,\tilde y)$ gives rise to the same system of ODEs as the GMAS $(G^*_k,y,\tilde y)$, involving the effective graph $G^*$. In particular, the dynamics does not depend on $k^0$. From  and  it follow that $$\label{gmas_s} f^{G}_k(x) = f^{G}_{(k^*,k^0)}(x) = f^{G}_{(k^*,{\sigma})}(x) = Y A^{G}_{(k^*,{\sigma})} \, x^{\tilde Y} ,$$ for arbitrary ${\sigma}\in {\mathbb{R}}^{E^0}_{>0}$. That is, we may replace the rate constants $k^0$ by arbitrary parameters ${\sigma}$. For a GMAS $(G_k,y,\tilde{y})$, the set of positive equilibria is given by $${X}^G_{k} := \{ x \in {\mathbb{R}}^n_{>0} \mid f^{G}_{k}(x) = 0 \} ,$$ while the set of positive [*complex-balanced*]{} equilibria (CBE) is given by $$Z^G_k := \{ x \in {\mathbb{R}}^n_{>0} \mid A^G_k \, x^{\tilde Y} = 0 \} \subseteq {X}^G_{k} .$$ Note that ${X}^G_k = {X}^{G^*}_{k^*}$, and hence the equilibrium set ${X}^G_k$ depends on $k^*$, but not on $k^0$, while $Z_k$ depends on both $k^*$ and $k^0$. Equation  motivates another definition. For an arbitrary parameter ${\sigma}\in {\mathbb{R}}^{E^0}_{>0}$, we consider $$Z^G_{(k^*,{\sigma})} := \{ x \in {\mathbb{R}}^n_{>0} \mid A^G_{(k^*,{\sigma})} \, x^{\tilde Y} = 0 \} \subseteq {X}^G_k ,$$ which is the set of positive CBE of the GMAS $(G_{(k^*,{\sigma})},y,\tilde{y})$. The [*parameterized*]{} set of positive CBE (PCBE) is given by $$\bar Z^G_k := \bigcup_{{\sigma}\in {\mathbb{R}}^{E^0}_{>0}} Z^G_{(k^*,{\sigma})} \subseteq {X}^G_{k} ,$$ thereby varying over all ${\sigma}\in {\mathbb{R}}^{E^0}_{>0}$. For a GMAS $(G_k,y,\tilde y)$, the set $\bar Z^G_k$ need not coincide with the set ${X}^G_{k}$. In our main results, however, we give conditions on the underlying GCRN $(G,y,\tilde y)$ such that ${X}^G_k = \bar Z^G_k$ (Theorem \[main\]), and also conditions under which a positive parametrization of $\bar Z^G_k$ can be constructed (Theorem \[main2\]). \[example4\] Recall the GCRN  from Example \[example3\]. The edge set $E$ can be partitioned into effective edges $E^* = \{ 1 \to 2, \, 2 \to 3, \, 3 \to 2, \, 4 \to 1 \}$ and phantom edges $E^0 = \{ 3 \to 4 \}$. The equilibrium set ${X}_{k}$ is determined by setting the right-hand sides of the ODEs  to zero, whereas the set $Z_k$ of CBE is determined by the Laplacian matrix, $$\label{ex:cb} A_k^Gx^{\tilde Y} = \left[ \begin{array}{cccc} -k_{1 \to 2} & 0 & 0 & k_{4 \to 1} \\ k_{1 \to 2} & -k_{2 \to 3} & k_{3 \to 2} & 0 \\ 0 & k_{2 \to 3} & -k_{3 \to 2}-k_{3 \to 4} & 0 \\ 0 & 0 & k_{3 \to 4} & -k_{4 \to 1} \end{array} \right] \left[ \begin{array}{c} x_1 \\ x_2 x_3 \\ x_1 x_4 \\ x_4 \end{array} \right] = \left[ \begin{array}{c} 0 \\ 0 \\ 0 \\ 0 \end{array} \right].$$ Note that these equations depend on the rate constant $k_{3 \to 4}$, even though it does not appear in the ODEs . By replacing $k_{3 \to 4}$ with an arbitrary parameter ${\sigma}$ in , we obtain the new set of CBE $Z_{(k^*,{\sigma})}$. The set $\bar Z_{k}$ of PCBE is obtained by varying over all ${\sigma}\in {\mathbb{R}}_{>0}$. A constructive method for solving systems like for the concentrations $x_i$ will be discussed in Section \[sec:kdzt\]. Condensed networks and effective deficiency ------------------------------------------- We now consider auxiliary networks with special properties. First, we introduce a network that condenses stoichiometrically identical vertices and thereby removes phantom edges. \[def:condensed\] For the GCRN $(G,y,\tilde y)$, we define the [*condensed*]{} CRN $(G',y')$ given by the digraph $G' = (V', E')$, where 1. $V' = \{ [i] \mid i \in V \}$ with $[i] = \{ j \in V \mid y(j)=y(i) \}$ for $i \in V$ and 2. $E' = \{ [i] \to [j] \mid i \to j \in E^* \}$, and the map $y' \colon V' \to {\mathbb{R}}^n, \, y'([i]) = y(i)$. For the GCRN $(G,y,\tilde y)$, we define the [*effective*]{} deficiency as the deficiency of the condensed CRN $(G',y')$, $$\delta' = \dim(\ker Y' \cap \operatorname{im}I_{E'})$$ with the incidence matrix $I_{E'} \in {\mathbb{R}}^{V' \times E'}$ and the matrix of complexes $Y' \in {\mathbb{R}}^{n \times V'}$, as defined after in Section \[sec:gcrn\]. Equivalently, $\delta' = m' - \ell' - s$, where $m' = |V'|$ is the number of vertices and $\ell'$ is the number of linkage classes of $G'$. Thereby, we use $S' = \operatorname{im}(Y' I_{E'}) = \operatorname{im}(Y I_E) = S$ and hence $s' = \dim(S') = \dim(S) = s$. Finally, we define a section $\rho \colon V' \to V$, assigning to each equivalence class $[i] \in V'$ a representative vertex $\rho([i]) \in [i]$, that is, we define a set of representative vertices $V^* = \{ \rho([i]) \mid [i] \in V' \} \subseteq V$, containing exactly one representative vertex from each equivalence class. \[ex:1\] Recall the GCRN  from Examples \[example3\] and \[example4\], in particular, that $y(3)=y(4)=X_1+X_4$. Hence we have the equivalence classes $$[1] = \{1\}, \; [2] = \{2\}, \; [3] = [4] = \{3,4\}.$$ For the GCRN, we obtain the condensed CRN $(G',y')$, in particular, the graph $G'$ $$\label{condensed} \begin{tikzcd} \{1\} \arrow[r] & \{2\} \arrow[d,xshift=0.5ex] \\ & \{3,4\} \arrow[u,xshift=-0.5ex] \arrow[lu] \end{tikzcd}$$ and the map $y'$ with $y'(\{1\}) = X_1$, $y'(\{2\}) = X_2 + X_3$, and $y'(\{3,4\}) = X_1 + X_4$. Note that we do not associate kinetic complexes to the vertices of the condensed graph. The deficiency of  is $\delta = 3-1-2=0$, that is, the effective deficiency of the GCRN  is $\delta' = 0$. [$V^*$-directed]{}networks -------------------------- Second, we introduce a class of GCRNs which is helpful for constructing a positive parametrization of the equilibrium set. \[def:vstar\] Let $(G,y,\tilde y)$ be a GCRN with $G=(V,E)$ and condensed CRN $G'=(V',E')$. Further, let $V^* \subseteq V$ be a set of representative vertices. (That is, there is a section $\rho \colon V' \to V$ such that $V^* = \{ \rho([i]) \mid [i] \in V' \}$.) We say that $(G,y,\tilde y)$ is *[$V^*$-directed]{}* if $$j \to i \in E^* \quad \text{implies} \quad i \in V^*, \quad \text{that is,} \quad i = \rho([i]),$$ and $$\begin{aligned} E^0 = \{ i \to j \mid i \in V^* , \, j \in [i] \setminus \{i\} \} , \quad \text{that is,} \quad E^0 = \{ \rho([i]) \to j \mid [i] \in V', \, j \in [i] \setminus \{\rho([i])\} \} .\end{aligned}$$ A GCRN being [$V^*$-directed]{}guarantees that effective edges (those between equivalence classes $[i]$) enter at the representative vertex $\rho([i]) \in V^*$, and that phantom edges (those within an equivalence class $[i]$) lead from $\rho([i])$ to the other vertices in the class. The representative vertices $\rho([i]) \in V^*$ may be thought of as the hubs of the representative equivalence classes through which all directed paths must travel. The class of [$V^*$-directed]{}GCRNs may seem restrictive. The following result, however, guarantees that, for every GMAS, there is a dynamically equivalent GMAS which is [$V^*$-directed]{}, that is, the associated ODEs agree, cf. . This will be instrumental in applications, cf. Section \[sec:applications\]. \[lemma:equiv2\] Let $(G,y,\tilde y)$ be a GCRN with $G=(V,E)$ and representative vertex set $V^* \subseteq V$, and let $k \in {\mathbb{R}}_{> 0}^{E}$ be a rate vector. Then there is a GCRN $({\hat}G,y,\tilde y)$ with ${\hat}G=(V, {\hat}E)$ that is [$V^*$-directed]{}and a rate vector ${\hat}k \in {\mathbb{R}}_{>0}^{{\hat}E}$ such that the GMASs $(G_k,y,\tilde y)$ and $({\hat}G_{{\hat}k},y,\tilde y)$ are dynamically equivalent, that is, the associated ODEs agree, cf. . First we define the set ${\hat}E^0 = \{ i \to j \mid i \in V^* , \, j \in [i] \setminus \{i\} \}$ and associate an arbitrary ${\hat}k_{i \to j} > 0$ to each edge $i \to j \in {\hat}E^0$. Then we define the set ${\hat}E^* = {\hat}E^*_1 \cup {\hat}E^*_2$ as follows: 1. If $i \to j \in E^*$ and $j \in V^*$, then $i \to j \in {\hat}E^*_1$ and ${\hat}k_{i \to j} = k_{i \to j}$. 2. If $i \to j \in E^*$ and $j \not\in V^*$, then $i \to \rho([j]) \in {\hat}E^*_2$ and $\displaystyle{{\hat}k_{i \to \rho([j])} = \sum_{j' \in [j] \setminus \{\rho([j])\}} k_{i \to j'}}.$ Now we consider the GCRN $({\hat}G, y, y')$ with ${\hat}G = (V, {\hat}E)$ and ${\hat}E = {\hat}E^0 \cup {\hat}E^*$, which is [$V^*$-directed]{}by construction. With the vector ${\hat}k \in {\mathbb{R}}_{> 0}^{{\hat}E}$ constructed above, we have $$\begin{split} f^G_k(x) & = \sum_{\substack{i \to j \in E^* \\ j \in V^*}} k_{i \to j} \, x^{\tilde y(i)} \, (y(j) - y(i)) + \sum_{\substack{i \to j' \in E^* \\ j' \not\in V^*}} k_{i \to j'} \, x^{\tilde y(i)} \, (y(j')) - y(i)) \\ & = \sum_{i \to j \in {\hat}E^*_1} {\hat}k_{i \to j} \, x^{\tilde y(i)} \, (y(j) - y(i)) + \sum_{i \to \rho([j]) \in {\hat}E^*_2} \sum_{j' \in [j] \setminus \{\rho([j])\}} k_{i \to j'} \, x^{\tilde y(i)} \, (y(j')) - y(i)) \\ & = \sum_{i \to j \in {\hat}E^*_1} {\hat}k_{i \to j} \, x^{\tilde y(i)} \, (y(j) - y(i)) + \sum_{i \to \rho([j]) \in {\hat}E^*_2} {\hat}k_{i \to \rho([j])} \, x^{\tilde y(i)} \, (y(\rho([j])) - y(i)) \\ &= f^{{\hat}G}_{{\hat}k}(x) , \end{split}$$ where we have omitted the edge sets $E^0$ and $\tilde E^0$ according to . Recall the GCRN  from Examples \[example3\], \[example4\] and \[ex:1\]. Since $y(3) = y(4)$ and hence $[3]=[4]=\{3,4\}$, we have two possible sections $\rho$, that is, two possible sets of representative vertices $V^*$, namely, $V_1^* = \{1, 2, 3\}$ and $V_2^* = \{1, 2, 4\}$. For the set $V_1^*$, all edges in  enter $\{3,4\}$ at $3 = \rho(\{3,4\})$, and the phantom edge $3 \to 4$ leads from $3 = \rho(\{3,4\})$ to $4$. Hence is $V_1^*$-directed. For the set $V_2^*= \{1, 2, 4\}$, the edge $2 \to 3$ in  leads to $3 \neq \rho(\{3,4\})$. We therefore replace it by the edge $2 \to 4$ with $4 = \rho(\{3,4\})$. Further, we replace the phantom edge $3 \to 4$ by the phantom edge $4 \to 3$. This construction yields the following $V_2^*$-directed GCRN $({\hat}G,y, \tilde y)$: $$\label{example32} \begin{tikzcd} \ovalbox{$\begin{array}{c} 1 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} X_1 + X_2 \\ (X_1) \end{array}$} \arrow[r] & \ovalbox{$\begin{array}{c} 2 \\ \\ \end{array} \Bigg\lvert\begin{array}{c} X_2 + X_3 \\ (X_2+X_3) \end{array}$} \arrow[dl] \\ \ovalbox{$\begin{array}{c} 4 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} X_1 + X_4 \\ (X_4) \end{array}$} \arrow[u] \arrow[r]& \ovalbox{$\begin{array}{c} 3 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} X_1 + X_4 \\ (X_1+X_4) \end{array}$} \arrow[u] \end{tikzcd}$$ The corresponding rate vector ${\hat}k \in {\mathbb{R}}_{> 0}^{\tilde E}$ is ${\hat}k_{1 \to 2} = k_{1 \to 2}$, ${\hat}k_{2 \to 4} = k_{2 \to 3}$, ${\hat}k_{3 \to 2} = k_{3 \to 2}$, ${\hat}k_{4 \to 1} = k_{4 \to 1}$, and ${\hat}k_{4 \to 3} = k_{3 \to 4}$. Hence $f_k^G = f_{{\hat}k}^{{\hat}G}$, cf. . Main results {#sec:results} ============ In Section \[sec:edzt\], we consider GCRNs with an effective deficiency of zero ($\delta' = 0$) and present Theorem \[main\], stating that the set of positive equilibria coincides with the parametrized set of complex-balanced equilibria (PCBE). In Section \[sec:kdzt\], we consider GCRNs with a kinetic deficiency of zero ($\tilde \delta = 0$) and higher ($\tilde \delta > 0$) and present Theorem \[main2\], explicitly constructing the PCBE. Effective deficiency {#sec:edzt} -------------------- Lemma \[cones\] below is crucial for the proof of Theorem \[main\]. In the following, we write $\operatorname{cone}W \subseteq {\mathbb{R}}^n$ for the polyhedral cone generated by the columns of the matrix $W \in {\mathbb{R}}^{n \times m}$. \[cones\] Let $(G,y,\tilde y)$ be a GCRN with $G=(V,E)$ and representative vertex set $V^* \subseteq V$. In particular, let $(G,y,\tilde y)$ be [$V^*$-directed]{}and have effective deficiency $\delta'=0$. Then $$\ker Y \cap \operatorname{cone}I_{E^*} \subseteq \operatorname{cone}(-I_{E^0}).$$ Moreover, $$\ker Y \cap \operatorname{relint}(\operatorname{cone}I_{E^*}) \subseteq \operatorname{relint}(\operatorname{cone}(-I_{E^0})) .$$ Let $v \in (\ker Y \cap \operatorname{cone}I_{E^*}) \subseteq {\mathbb{R}}^V$, that is, $v = I_{E^*} \, x = \sum_{i \to j \in E^*} x_{i \to j} \, (e_j - e_i)$ with $x \in {\mathbb{R}}^{E^*}_{\ge0}$ (nonnegative weights on the effective edges $E^*$) and $$\begin{aligned} 0 = Y v & = \sum_{i \to j \in E^*} x_{i \to j} \, (y(j) - y(i)) \\ & = \sum_{[i] \to [j] \in E'} \bigg( \sum_{\substack{i' \to j' \in E^*: \\ i' \in [i], j' \in [j]}} x_{i' \to j'} \bigg) \, (y'([j]) - y'([i])) \\ &= Y' \sum_{[i] \to [j] \in E'} x'_{[i] \to [j]} \,(e_{[j]} - e_{[i]}) \\ &= Y' \, I_{E'} \, x' = Y' \, v' .\end{aligned}$$ Thereby, $(G',y')$ with $G'=(V',E')$ is the corresponding condensed CRN and $v' = I_{E'} \, x' \in {\mathbb{R}}^{V'}$ with $x' \in {\mathbb{R}}^{E'}_{\ge0}$. Clearly, $ v'_{[i]} = \sum_{i' \in [i]} v_{i'} $ for $[i] \in V'$. Now, $\delta'=\dim(\ker Y' \cap \operatorname{im}I_{E'})=0$ implies $v'=0$, that is, $$\begin{aligned} 0 = v'_{[i]} = \sum_{i' \in [i]} v_{i'} $$ for $[i] \in V'$. Using that $G$ is [$V^*$-directed]{}, reconsider $v = I_{E^*} \, x \in {\mathbb{R}}^V$ (the fluxes arising from the effective edges $E^*$). Let $i \in V^*$, that is, $i = \rho([i])$. For $i' \in [i] \setminus \{i\}$, $$v_{i'} = - \sum_{i' \to j \in E^*} x_{i' \to j} ,$$ whereas $$v_{i} = - \sum_{i' \in [i] \setminus \{i\}} v_{i'} .$$ Now, choose $\tilde x \in {\mathbb{R}}^{E^0}_{\ge0}$ (nonnegative weights on the phantom edges $E^0$) as $$\label{tilde_x} \tilde x_{i \to i'} = \sum_{i' \to j \in E^*} x_{i' \to j} ,$$ where $i' \in [i]$. Then, for $i' \in [i] \setminus \{i\}$, $$v_{i'} = - \tilde x_{i \to i'} ,$$ whereas $$v_{i} = - \sum_{i' \in [i] \setminus \{i\}} v_{i'} = \sum_{i'\in[i] \setminus \{i\}} \tilde x_{i \to i'} = \sum_{i \to i' \in E^0} \tilde x_{i \to i'} .$$ That is, $- v = I_{E^0} \, \tilde x \in {\mathbb{R}}^V$ (the fluxes arising from the phantom edges $E^0$), and hence $v \in \operatorname{cone}(-I_{E^0})$. Finally, let $v \in (\ker Y \cap \operatorname{relint}(\operatorname{cone}I_{E^*}))$, that is, $v = I_{E^*} \, x$ for some $x \in {\mathbb{R}}^{E^*}_{>0}$. Then $v = -I_{E^0} \, \tilde x\in \operatorname{relint}(\operatorname{cone}(-I_{E^0}))$ with $\tilde x \in {\mathbb{R}}^{E^0}_{>0}$ by . We now present the main result of this section, which gives conditions under which the equilibrium set $X_k^G$ coincides with the parametrized set of complex-balanced equilibria $\bar{Z}_k^G$. \[main\] Let $(G,y,\tilde y)$ be a GCRN with effective deficiency $\delta'=0$. Further, let $(G,y,\tilde y)$ be [$V^*$-directed]{}for a set of representative vertices $V^* \subseteq V$. Then, for the GMAS $(G_k,y,\tilde y)$, the set of positive equilibria agrees with the parametrized set of complex-balanced equilibria, that is, ${X}^G_{k} = \bar Z^G_k$. Let $x \in {\mathbb{R}}_{>0}^n$ be a positive equilibrium, that is, $x \in {X}^G_k$. Using $G^*=(V,E^*)$, $G^0=(V,E^0)$, and ${\sigma}\in {\mathbb{R}}^{E^0}_{>0}$, we may write $$A^G_{(k^*,{\sigma})} \, x^{\tilde Y} = A^{G^*}_{k^{\ast}} \, x^{\tilde Y} + A^{G^0}_{\sigma}x^{\tilde Y} ,$$ cf. . Now $x \in {X}^G_k = {X}^{G^*}_{k^*}$ implies $Y A^{G^*}_{k^*} x^{\tilde Y} = 0$ and hence $$A^{G^*}_{k^*} \, x^{\tilde Y} \in (\ker Y \cap \operatorname{relint}(\operatorname{cone}I_{E^{\ast}})) ,$$ cf. . Since $\delta' = 0$ and $(G,y,\tilde y)$ is [$V^*$-directed]{}, we have $ A^{G^*}_{k^{\ast}} \, x^{\tilde Y} \in \operatorname{relint}(\operatorname{cone}(-I_{E^0})) , $ by Lemma \[cones\]. That is, $$A^{G^*}_{k^{\ast}} \, x^{\tilde Y} = - \sum_{i \to j \in E^0} \alpha_{i \to j} \, (e_j - e_i)$$ for some $\alpha \in {\mathbb{R}}^{E^0}_{>0}$. On the other hand, $$A^{G^0}_{\sigma}x^{\tilde Y} = \sum_{i \to j \in E^0} {\sigma}_{i \to j} \, x^{\tilde y(i)} \, (e_j - e_{i}).$$ We choose ${\sigma}_{i \to j} = \alpha_{i \to j} / x^{\tilde y(i)}$ for $i \to j \in E^0$ such that $A^{G^*}_{k^*} x^{\tilde Y} = - A^{G^0}_{{\sigma}} x^{\tilde Y}$ and hence $A^G_{(k^*,{\sigma})} \, x^{\tilde Y} = 0$, that is, $x \in Z^G_{(k^*,{\sigma})} \subseteq \bar Z^G_k$ so ${X}^G_{k} \subseteq \bar Z^G_k$. Since $\bar Z^G_k \subseteq {X}^G_{k}$ trivially, we have ${X}^G_{k} = \bar Z^G_k$. Kinetic deficiency {#sec:kdzt} ------------------ We fix the graph $G=(V,E)$ and omit the corresponding superscript, that is, we write $A_k^G = A_k$, $Z^G_k = Z_k$, and $\bar Z_k^G = \bar Z_k$. Recall that $x \in Z_k$ is equivalent to $x^{\tilde Y} \in \ker A_{k}$. Following [@J2014; @MR2014], we discuss $\ker A_k$. First, we introduce the vector of [*tree constants*]{} $K \in {\mathbb{R}}^V_{>0}$ with entries $$K_i = \sum_{(\mathcal V, \mathcal E) \in T_i} \prod_{i' \to j' \in \mathcal E} k_{i' \to j'} , \quad i \in V ,$$ where $T_i$ is the set of directed spanning trees (of the respective linkage class) rooted at vertex $i$. Clearly, the tree constants $K$ depend on the rate constants $k \in {\mathbb{R}}^E_{>0}$, that is, $K = K(k)$. For a weakly reversible GCRN, $$\ker A_k = \operatorname{span}\{ v^1, \ldots, v^{\ell} \}$$ with nonnegative vectors $v^l \in {\mathbb{R}}_{\geq 0}^n$ (for $l = 1, \ldots, \ell$) having support on the respective linkage class $l$. In particular, $v^l_i = K_i$ if vertex $i$ is in linkage class $l$ and $v^l_i = 0$ otherwise. Now, $x^{\tilde Y} \in \ker A_k$ if and only if $$x^{\tilde Y} = \sum_{l=1}^\ell \alpha_l \, v^l$$ with $\alpha_l > 0$. For any pair of vertices $i$ and $j$ in the same linkage class, we have $$\frac{x^{\tilde y(i)}}{K_{i}} = \frac{x^{\tilde y(j)}}{K_{j}}.$$ Taking the logarithm gives $$\label{eq:loglin} (\tilde y(i) - \tilde y(j))^T \ln x = \ln \left( \frac{K_{i}}{K_{j}} \right).$$ Now we choose a spanning forest $F = (V,\mathcal E)$ for $G = (V,E)$, that is, we choose spanning trees for all linkage classes. Note that $F$ contains the same vertices as $G$, but not the same edges. Also note that, in the following results and applications, the choice of the spanning tree is arbitrary. Clearly, the spanning tree of linkage class $l$ contains $m_l$ vertices and $m_l - 1$ edges. Hence, the spanning forest $F$ contains $m$ vertices and $m-\ell$ edges. We introduce the matrix $M = \tilde Y I_\mathcal{E} \in {\mathbb{R}}^{n \times \mathcal E}$ whose $m-\ell$ columns are given by $\tilde y(j) - \tilde y(i)$ for $i \to j \in \mathcal E$. Correspondingly, we define the vector ${\kappa}\in {\mathbb{R}}^\mathcal{E}_{>0}$ whose $m-\ell$ entries are given by ${\kappa}_{i \to j} = \frac{K_i}{K_j}$ for $i \to j \in \mathcal E$. As for $K$, we note that ${\kappa}$ depends on $k$, that is, ${\kappa}={\kappa}(k)$. Hence, we can write the system of equations  as $$\label{loglinear} M^T \ln x = \ln {\kappa}.$$ Theorem 1 in [@MR2014] implies the following result. \[main2\] Let $(G,y,\tilde y)$ be a GCRN that is weakly reversible, and let $(G_k,y,\tilde y)$ be a GMAS. Further, let $M \in {\mathbb{R}}^{n \times \mathcal{E}}$ and ${\kappa}={\kappa}(k)={\kappa}(k^*,k^0) \in {\mathbb{R}}^\mathcal{E}_{>0}$ be defined as above, and let $H \in {\mathbb{R}}^{n \times \mathcal{E}}$ be a generalized inverse of $M^T$ (that is, $M^T H M^T = M^T$). Finally, define $B \in {\mathbb{R}}^{n \times (n - \tilde{s})}$ with $\operatorname{im}B = \ker M^T$ and $\ker B = \{0\}$, and $C \in {\mathbb{R}}^{\mathcal{E} \times \tilde \delta}$ with $\operatorname{im}C = \ker M$ and $\ker C = \{0\}$. 1. If the kinetic deficiency is zero ($\tilde \delta = 0$), then $\bar Z_{k} \not= \emptyset$, in particular, $\bar Z_{k}$ has the positive parametrization $$\label{eq:param} \bar Z_{k} = \left\{ {\kappa}(k^*,{\sigma})^{H^T} \circ {\tau}^{B^T} \mid {\sigma}\in {\mathbb{R}}^{E^0}_{>0}, \, {\tau}\in {\mathbb{R}}^{n - \tilde s}_{>0} \right\}.$$ 2. If the kinetic deficiency is positive ($\tilde \delta > 0$) and the $\tilde \delta$ equations $$\label{eq:condition} {\kappa}(k^*,k^0)^{C} = 1^{\tilde \delta \times 1}$$ can be solved explicitly for $\tilde \delta$ components of $k^0 \in {\mathbb{R}}^{E^0}_{>0}$ (in terms of $k^* \in {\mathbb{R}}^{E^*}_{>0}$ and the remaining components of $k^0$), that is, if there exists an explicit function $h \colon {\mathbb{R}}^{E^* \cup (E^0 \setminus \tilde E^0)}_{>0} \to {\mathbb{R}}^{\tilde E^0}_{>0}$ with $\tilde E^0 \subseteq E^0$, $|\tilde E^0| = \tilde \delta$, and $k^0 = (\tilde k^0, \cdot) \in {\mathbb{R}}^{(E^0 \setminus \tilde E^0) \cup \tilde E^0}_{>0}$ such that, for all $k^* \in {\mathbb{R}}_{> 0}^{E^*}$ and $\tilde k^0 \in {\mathbb{R}}_{> 0}^{E^0 \setminus \tilde E^0}$, $${\kappa}(k^*,(\tilde k^0, h(k^*,\tilde k^0)))^{C} = 1^{\tilde \delta \times 1} ,$$ then $\bar Z_{k} \neq \emptyset$, and $\bar Z_{k}$ has the positive parametrization $$\label{eq:param2} \bar Z_{k} = \left\{ {\kappa}(k^*,({\sigma}, h(k^*, {\sigma})))^{H^T} \circ {\tau}^{B^T} \mid {\sigma}\in {\mathbb{R}}^{E^0 \setminus \tilde E^0}_{>0}, \, {\tau}\in {\mathbb{R}}^{n - \tilde s}_{>0} \right\}.$$ Before we prove statements 1 and 2 of Theorem \[main2\], we make two remarks. - If the generalized inverse $H \in {\mathbb{R}}^{n \times \mathcal{E}}$ of $M^T$ has integer entries, then  is a rational parametrization. Common generalized inverses such as the Moore-Penrose inverse, however, rarely have this property [@Israel2003]. In applications, we construct $H$ by determining the matrix of elementary row operations $P$ that transforms $M^T$ to reduced row echelon form. That is, we find $P \in {\mathbb{R}}^{\mathcal{E} \times \mathcal{E}}$ such that $PM^T \in {\mathbb{R}}^{\mathcal{E} \times n}$ is the reduced row echelon form of $M^T$. Then we determine $Q \in \{0,1\}^{n \times \mathcal{E}}$ such that $Q P M^T = I$ and hence $\ln x = H \ln {\kappa}$ with $H = Q P \in {\mathbb{R}}^{n \times \mathcal{E}}$. That is, we perform Gaussian elimination on  and then set all free parameters to zero. - As a special case of statement 2, if $\tilde E^0 = E^0$ and equations can be solved explicitly for $k^0$ (in terms of $k^*$), that is, if there exists $h \colon {\mathbb{R}}^{E^*}_{>0} \to {\mathbb{R}}^{E^0}_{>0}$ such that $$\kappa (k^*,h(k^*))^C = 1^{\tilde \delta \times 1},$$ then we obtain the monomial parametrization $$\bar{Z}_k = \left\{ \kappa(k^*,h(k^*))^{H^T} \circ \tau^{B^T} \; | \; \tau \in {\mathbb{R}}_{>0}^{n - \tilde s} \right\}.$$ Since $(G,y,\tilde y)$ is weakly reversible, $x \in Z_k$ if and only if $\ln x$ satisfies . Now $\operatorname{im}M = \operatorname{im}(\tilde Y I_\mathcal{E}) = \operatorname{im}(\tilde Y I_E) = \tilde S$ and hence $\operatorname{rank}M = \tilde{s}$. Since the kinetic deficiency is zero, we have $\tilde \delta = m - \ell - \tilde s = 0$ and hence $\tilde s = m - \ell$. That is, $M^T$ has full rank $m-\ell$ and hence $\ln {\kappa}\in \operatorname{im}M^T$ for any ${\kappa}\in {\mathbb{R}}_{> 0}^{m - \ell}$. Equivalently, the linear system has a solution $\ln x$ for any ${\kappa}\in {\mathbb{R}}_{>0}^\mathcal{E}$. Following Proposition 3 in [@MR2014], we use the generalized inverse $H \in {\mathbb{R}}^{n \times \mathcal{E}}$ of $M^T$ and obtain $$M^T H \ln {\kappa}= M^T H M^T \ln x = M^T \ln x = \ln {\kappa}.$$ That is, $\ln x^*=H \ln {\kappa}$ is a solution of  and hence $x^* = {\kappa}^{H^T} \in Z_k$. In particular, $Z_k \neq \emptyset$. For any $x \in Z_k$, $$M^T ( \ln(x) - \ln(x^*)) = 0$$ and, since $\ker M^T = \operatorname{im}M^\perp = \tilde S^\perp$, $$\ln(x) - \ln(x^*) \in \tilde S^{\perp}.$$ We use $B \in {\mathbb{R}}^{n \times (n - \tilde s)}$ with $\operatorname{im}B = \tilde{S}^{\perp}$, $\ker B = \{0\}$ and obtain $$\ln(x) - \ln(\tilde x^*) = B \alpha$$ with $\alpha \in {\mathbb{R}}^{n-\tilde s}$ and $$x = x^* \circ {\tau}^{B^T}$$ with ${\tau}= e^\alpha \in {\mathbb{R}}^{n-\tilde s}_{>0}$. Equivalently, $$Z_k = \left\{ x^* \circ {\tau}^{B^T} \mid {\tau}\in {\mathbb{R}}^{n-\tilde s}_{>0} \right\} = \left\{ {\kappa}(k^*,k^0)^{H^T} \circ {\tau}^{B^T} \mid {\tau}\in {\mathbb{R}}^{n-\tilde s}_{> 0} \right\} .$$ Note that the matrices $M$, $H$, and $B$ do not depend on $k \in {\mathbb{R}}_{>0}^{E}$, whereas ${\kappa}= {\kappa}(k) = {\kappa}(k^*,k^0)$. Finally, $$\bar Z_{k} = \bigcup_{{\sigma}\in {\mathbb{R}}^{E^0}_{>0}} Z_{(k^*,{\sigma})} = \left\{ {\kappa}(k^*,{\sigma})^{H^T} \circ {\tau}^{B^T} \mid {\sigma}\in {\mathbb{R}}^{E^0}_{>0}, \, {\tau}\in {\mathbb{R}}^{n - \tilde s}_{>0} \right\} .$$ If the kinetic deficiency is positive ($\tilde \delta > 0$), then $M^T$ does not have full rank, and does not have a solution for all right-hand sides. We use $C \in {\mathbb{R}}^{\mathcal{E} \times \tilde \delta}$ with $\operatorname{im}C = \ker M$, $\ker C = \{0\}$ and find that  has a solution if and only if $\ln {\kappa}\in \operatorname{im}M^T = \ker M^\perp = \operatorname{im}C^\perp = \ker C^T$. Equivalently, $C^T \ln {\kappa}= 0$, that is, $${\kappa}^C = {\kappa}(k^*,k^0)^C = 1^{\tilde \delta \times 1} .$$ By assumption, these $\tilde \delta$ equations can be solved explicitly for $\tilde \delta$ components of $k^0 \in {\mathbb{R}}^{E^0}_{>0}$ (in terms of $k^* \in {\mathbb{R}}^{E^*}_{>0}$ and the remaining components of $k^0$), that is, there exists an explicit function $h \colon {\mathbb{R}}^{E^* \cup (E^0 \setminus \tilde E^0)}_{>0} \to {\mathbb{R}}^{\tilde E^0}_{>0}$ with $\tilde E^0 \subseteq E^0$, $|\tilde E^0| = \tilde \delta$, and $k^0 = (\tilde k^0, \cdot) \in {\mathbb{R}}^{(E^0 \setminus \tilde E^0) \cup \tilde E^0}_{>0}$ such that, for all $k^* \in {\mathbb{R}}_{> 0}^{E^*}$ and $\tilde k^0 \in {\mathbb{R}}_{> 0}^{E^0 \setminus \tilde E^0}$, $${\kappa}(k^*,(\tilde k^0, h(k^*,\tilde k^0))^{C} = 1^{\tilde \delta \times 1} .$$ Hence  has a solution for any $k^* \in {\mathbb{R}}^{E^*}_{>0}$ and $\tilde k^0 \in {\mathbb{R}}^{E^0 \setminus \tilde E^0}_{>0}$, and from the proof of statement 1 we obtain the positive parametrization . Applications {#sec:applications} ============ The process of network translation allows to relate a classical CRN to a GCRN with potentially stronger structural properties [@J2014]. Using this method, we can apply the main results of this paper, Theorems \[main\] and \[main2\], to a broad class of mass-action systems studied in the biochemical literature. Translated chemical reaction networks {#sec:translation} ------------------------------------- The following definition was introduced in [@J2014] in order to relate a MAS to a dynamically equivalent GMAS. \[def:translation\] Let $(G,y)$ with $G=(V,E)$ be a CRN. A GCRN $(G^{\intercal},y^{\intercal},\tilde y^{\intercal})$ with $G^{\intercal}=(V^{\intercal},E^{\intercal})$ if a *translation* of $(G,y)$ is there exists a map $g \colon E \to E^{\intercal}$ such that $g(i \to j) = i^{\intercal}\to j^{\intercal}$ with $i \to j \in E$ and $i^{\intercal}\to j^{\intercal}\in E^{\intercal}$ implies (i) $y^{\intercal}(j^{\intercal})-y^{\intercal}(i^{\intercal}) = y(j) - y(i)$ and (ii) $\tilde y^{\intercal}(i^{\intercal}) = y(i)$. In other words, a GCRN is a translation of a given CRN if there is a map between reactions of the two networks which (i) preserves reaction vectors and (ii) relates source complexes in the CRN to kinetic complexes in the GCRN. Definition \[def:translation\] is more general than Definition 6 in [@J2014]. In that work, GCRNs were defined as in [@MR2012] which required $y^{\intercal}$ and $\tilde y^{\intercal}$ to be injective. Here, GCRNs are defined as in [@MR2014] which allows $y^{\intercal}$ and $\tilde y^{\intercal}$ to be noninjective. \[lemma:equiv1\] Let $(G,y)$ be a CRN, and let $k \in {\mathbb{R}}^{E}_{> 0}$ be a rate vector. Further, let the GCRN $(G^{\intercal}, y^{\intercal}, \tilde{y}^{\intercal})$ be a translation of $(G,y)$, and let $k^{\intercal}\in {\mathbb{R}}_{> 0}^{E^{\intercal}}$ be a rate vector with $k^{\intercal}_{i^{\intercal}\to j^{\intercal}} = k_{i \to j}$ if $g(i \to j) = i^{\intercal}\to j^{\intercal}$. Then the MAS $(G_k,y)$ and the GMAS $(G^{\intercal}_{k^{\intercal}}, y^{\intercal}, \tilde y^{\intercal})$ are dynamically equivalent, that is, the associated ODEs agree, cf.  and . The ODEs associated with the MAS $(G_k,y)$ are determined by $f^G_k(x)$, cf. , whereas the ODEs associated with the GMAS $(G^{\intercal}_{k^{\intercal}},y^{\intercal},\tilde y^{\intercal})$ are determined by $f^{G^{\intercal}}_{k^{\intercal}}(x)$, cf. . By Definition \[def:translation\] and the construction of $k^{\intercal}$, we have $$f^G_{k}(x) = \sum_{i \to j \in E} k_{i \to j} \, x^{y(i)} \, (y(j) - y(i))= \sum_{i^{\intercal}\to j^{\intercal}\in E^{\intercal}} k^{\intercal}_{i^{\intercal}\to j^{\intercal}} \, x^{\tilde y^{\intercal}(i^{\intercal})} \, (y^{\intercal}(j^{\intercal}) - y^{\intercal}(i^{\intercal})) = f^{G^{\intercal}}_{k^{\intercal}}(x).$$ Lemmas \[lemma:equiv1\] and \[lemma:equiv2\] provide a framework for parametrizing the set of positive equilibria of a (classical) MAS , by applying Theorems \[main\] and \[main2\]. $$\begin{tikzcd} \mbox{\doublebox{\begin{tabular}{c} Original CRN, \\ MAS $(G_k,y)$ \end{tabular}}} \arrow[rr,"\begin{tabular}{c} Network translation \\ (Lemma~\ref{lemma:equiv1}) \end{tabular}"] & & \mbox{\doublebox{\begin{tabular}{c} Translated GCRN, \\ MAS $(G_{k^{\intercal}}^{\intercal},y^{\intercal}, \tilde y^{\intercal})$ \end{tabular}}} \arrow[ddl,"\begin{tabular}{c} Network redirection \\ (Lemma~\ref{lemma:equiv2}) \end{tabular}"] \\ & & \\ & \mbox{\doublebox{\begin{tabular}{c} {$V^*$-directed\xspace}GCRN, \\ GMAS $({\hat}G^{\intercal}_{{\hat}k^{\intercal}},y^{\intercal},\tilde y^{\intercal})$ \end{tabular}}} \arrow[luu,"\begin{tabular}{c} Parametrization \\ (Theorems \ref{main} and~\ref{main2}) \end{tabular}"] & \end{tikzcd}$$ In biochemical applications, a suitable GCRN that corresponds to a given CRN may not be apparent. In particular, in order to apply Theorem \[main\], we want the translated network to have effective deficiency zero, and to apply Theorem \[main2\], we want the kinetic deficiency to be as low as possible and the translated and [$V^*$-directed]{}network to be weakly reversible. A *translation scheme* involves the addition of linear combinations of species to each side of a reaction arrow [@J2014]. This operation preserves reaction vectors and establishes a correspondence between source complexes in the original network and kinetic complexes in the new one. For small networks, this may suffice to create a suitably well-connected translation; however, it is extremely challenging for large networks. Computational approaches to optimal network translation have been conducted in [@J2015] and [@Tonello2017]. Examples {#sec:examples} -------- The following examples are drawn from the biochemical literature. Recall the histidine kinase network  from the introduction and apply the following translation scheme: $$\begin{tikzcd} X \arrow[r, "k_1"] & X_p & (+ Y)\\[-0.2in] X_p + Y \arrow[r, yshift=+0.5ex,"k_2"] & X + Y_p \arrow[l, yshift=-0.5ex,"k_3"] & (+ 0) \\[-0.2in] Y_p \arrow[r, "k_4"] & Y & (+ X) \end{tikzcd}$$ The resulting GCRN together with an additional phantom edge yields a weakly reversible GCRN, given by the (edge labeled) graph $$\label{example233} \begin{tikzcd} \ovalbox{$\begin{array}{c} 1 \\ \\ \end{array} \Bigg\lvert \begin{array}{c} X + Y \\ (X) \end{array}$} \arrow[r,"k_1"] & \ovalbox{$\begin{array}{c} 2 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} X_p + Y \\ (X_p+Y) \end{array}$} \arrow[d,xshift=+0.5ex,"k_2"]\\ \ovalbox{$\begin{array}{c} 4 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} X + Y_p \\ (Y_p) \end{array}$} \arrow[u,"k_4"] & \ovalbox{$\begin{array}{c} 3 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} X + Y_p \\ (X+Y_p) \end{array}$} \arrow[u,xshift=-0.5ex,"k_3"] \arrow[l,color=red,"{\sigma}"] \end{tikzcd}$$ The stoichiometric complex $X+Y_p$ appears twice in , specifically, $[3]=[4]=\{3,4\}$, and the network is [$V^*$-directed]{}for $V^* = \{1,2,3\}$. The network has a stoichiometric deficiency of one ($\delta = 1$) and a kinetic deficiency of zero ($\tilde \delta = 0)$. The condensed network is given by the following graph: $$\begin{tikzcd} \{1\} \arrow[r] & \{2\} \arrow[d,xshift=0.5ex] \\ & \{3,4\} \arrow[u,xshift=-0.5ex] \arrow[lu] \end{tikzcd}$$ It has a deficiency of zero ($\delta' = 0$). Theorem \[main\] guarantees that the equilibrium set coincides with the parametrized set of complex-balanced equilibria. Furthermore, since $\tilde \delta = 0$ and  is weakly reversible, Theorem \[main2\] guarantees that there is a positive parametrization of the form . By the construction preceding Theorem \[main2\], we compute the matrix $M$ (and further $H$ and $B$). In particular, we choose a spanning forest $F = (V, \mathcal{E})$ for the graph  with edges $1 \to 2$, $1 \to 3$, and $1 \to 4$, and we compute the corresponding differences of kinetic complexes $X_p + Y - X$, $X + Y_p - X$, and $Y_p - X$: $$M = \begin{array}{c} X \\ X_p \\ Y \\ Y_p \end{array} \hspace{-1ex} \left[ \begin{array}{ccc} -1 & 0 & -1 \\ 1 & 0 & 0 \\ 1 & 0 & 0 \\ 0 & 1 & 1\end{array} \right], \hspace{0.25in} H = \left[ \begin{array}{ccc} 0 & 1 & -1 \\ 1 & 1 & -1 \\ 0 & 0 & 0 \\ 0 & 1 & 0 \end{array} \right], \hspace{0.25in} B = \left[ \begin{array}{c} 0 \\ -1 \\ 1 \\ 0 \end{array} \right] .$$ Thereby $M^T H M^T = M^T$, that is, $H$ is a generalized inverse of $M^T$, and $\operatorname{im}B = \ker M^T$. In order to determine the parametrization , it remains to compute the tree constants $K=K(k^*,{\sigma})$ of the graph  and their quotients ${\kappa}={\kappa}(k^*,{\sigma})$. We find $$K_1 = k_2k_4{\sigma}, \; \; \; K_2 = k_1(k_3+{\sigma})k_4, \; \; \; K_3 = k_1k_2k_4, \; \mbox{ and } \; K_4 = k_1k_2{\sigma}.$$ Taking the spanning forest $F = (V, \mathcal{E})$ as above gives $${\kappa}_1 = \frac{K_2}{K_1} = \frac{k_1(k_3 + {\sigma})}{k_2{\sigma}}, \hspace{0.25in} {\kappa}_2 = \frac{K_3}{K_1} = \frac{k_1}{{\sigma}}, \hspace{0.25in} {\kappa}_3 = \frac{K_4}{K_1}= \frac{k_1}{k_4}.$$ As a consequence, the rational parametrization  amounts to $$\left \{ \quad \begin{aligned} x &= {\kappa}_2^1 {\kappa}_3^{-1} \cdot 1 &&= \frac{k_4}{{\sigma}} , \\ x_p &= {\kappa}_1^1 {\kappa}_2^1 {\kappa}_3^{-1} \cdot {\tau}^{-1} &&= \frac{k_1(k_3+{\sigma})k_4}{k_2{\sigma}^2{\tau}} , \\ y &= 1 \cdot {\tau}^1 &&= {\tau}, \\ y_p &= {\kappa}_2^1 \cdot 1 &&= \frac{k_1}{{\sigma}} , \end{aligned} \right.$$ where ${\sigma},{\tau}> 0$. Consider the following EnvZ-OmpR signaling pathway, which was first proposed in [@Sh-F], together with the translation scheme proposed in [@J2014]: $$\begin{tikzcd} XD \arrow[r,yshift=+0.5ex,"k_1"] & X \arrow[l,yshift=-0.5ex,"k_2"] \arrow[r,yshift=+0.5ex,"k_3"]& XT \arrow[l,yshift=-0.5ex,"k_4"] \arrow[r,"k_5"]& X_p & (+ XD + XT + Y)\\[-0.1in] X_p + Y \arrow[r,yshift=+0.5ex,"k_6"] & X_pY \arrow[l,yshift=-0.5ex,"k_7"] \arrow[r,yshift=+0.5ex,"k_8"]& X + Y_p & & (+ XD + XT) \\[-0.1in] XD + Y_p \arrow[r,yshift=+0.5ex,"k_9"] & XDY_p \arrow[l,yshift=-0.5ex,"k_{10}"] \arrow[r,yshift=+0.5ex,"k_{11}"]& XD + Y & & (+ X + XT) \\[-0.1in] XT + Y_p \arrow[r,yshift=+0.5ex,"k_{12}"] & XTY_p \arrow[l,yshift=-0.5ex,"k_{13}"] \arrow[r,yshift=+0.5ex,"k_{14}"]& XT + Y & & (+ X + XD) \\[-0.1in] \end{tikzcd}$$ The resulting GCRN together with an additional phantom edge yields a weakly reversible GCRN, given by the (edge labeled) graph $$\label{envz_ompr_directed} \begin{tikzcd}[column sep=0.5cm] \mbox{\ovalbox{$\begin{array}{c} 1 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} 2XD + XT + Y \\ (XD) \end{array}$}} \arrow[r, rightharpoonup, yshift=+0.2ex, "k_{1}"] & \mbox{\ovalbox{$\begin{array}{c} 2 \\ \\ \end{array}\Bigg\lvert \begin{array}{c}XD + X + XT + Y \\ (X) \end{array}$}} \arrow[l, rightharpoonup, yshift=-0.2ex, "k_{2}"] \arrow[r, rightharpoonup, yshift=+0.2ex, "k_{3}"] & \mbox{\ovalbox{$\begin{array}{c} 3 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} XD + 2XT + Y \\ (XT) \end{array}$}} \arrow[l, rightharpoonup, yshift=-0.2ex, "k_{4}"] \arrow[d, "k_{5}"] & \\ \mbox{\ovalbox{$\begin{array}{c} 9 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} X + XT + XTY_p \\ (XTY_p) \end{array}$}} \arrow[ur, "k_{14}"] \arrow[rd, "k_{13}"] & \mbox{\ovalbox{$\begin{array}{c} 7 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} XD + X + XDY_p \\ (XDY_p) \end{array}$}} \arrow[u, "k_{11}"] \arrow[d, rightharpoonup, "k_{10}", xshift=+0.2ex] & \mbox{\ovalbox{$\begin{array}{c} 4 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} XD + XT + X_p + Y \\ (X_p + Y) \end{array}$}} \arrow[d, rightharpoonup, "k_{6}", xshift=+0.2ex] \\ \mbox{\ovalbox{$\begin{array}{c} 8 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} XD + X + XT + Y_p \\ (XT + Y_p) \end{array}$}} \arrow[u, rightharpoonup, "k_{12}", xshift=-0.2ex] & \mbox{\ovalbox{$\begin{array}{c} 6 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} XD + X + XT + Y_p \\ (XD + Y_p) \end{array}$}} \arrow[u, rightharpoonup, "k_{9}", xshift=-0.2ex] \arrow[l,color=red, "{\sigma}"] & \mbox{\ovalbox{$\begin{array}{c} 5 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} XD + XT + X_pY \\ (X_pY) \end{array}$}} \arrow[l, "k_8"'] \arrow[u, rightharpoonup, "k_{7}", xshift=-0.2ex] & \end{tikzcd}$$ Thereby $6 \to 8$ is the phantom edge (with label ${\sigma}>0$) since $y(6) = y(8) = XD + X + XT + Y_p$. The network is [$V^*$-directed]{}for $V^* = V\setminus\{8\}$. It can be quickly checked that the condensed graph $G'$ has deficiency zero, so that  has an effective deficiency of zero ($\delta' = 0$). It follows from Theorem \[main\] that every equilibrium point is in the parametrized set of CBE (i.e. $X_k = \bar{Z}_k$). It can also be checked that  has a kinetic deficiency of one ($\tilde \delta = 1$). Hence, in order to apply Theorem \[main2\] (statement 2), we need to first determine if there is ${\sigma}=h(k^*)$ such that ${\kappa}(k^*,h(k^*))^{C} = 1$. We choose the spanning forest $F = (V, \mathcal{E})$ for the graph  consisting of the edges $1 \to i$ for $i=2, \ldots, 9$. We compute the following matrices: $$\tiny M = \hspace{-1ex} \begin{array}{c} XD \\ X \\ XT \\ X_p \\ Y \\ X_pY \\ Y_p \\ XDY_p \\ XTY_p \end{array} \hspace{-1ex} \left[ \begin{array}{cccccccc} \hspace{-0.05in} -1 \hspace{-0.05in} & \hspace{-0.05in} -1 \hspace{-0.05in} & \hspace{-0.05in} -1 \hspace{-0.05in}& \hspace{-0.05in} -1 \hspace{-0.05in}& 0 & \hspace{-0.05in} -1 \hspace{-0.05in}& \hspace{-0.05in} -1 \hspace{-0.05in} & \hspace{-0.05in} -1 \hspace{-0.05in}\\1 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\0 & 1 & 0 & 0 & 0 & 0 & 1 & 0\\0 & 0 & 1 & 0 & 0 & 0 & 0 & 0\\0 & 0 & 1 & 0 & 0 & 0 & 0 & 0\\0 & 0 & 0 & 1 & 0 & 0 & 0 & 0\\0 & 0 & 0 & 0 & 1 & 0 & 1 & 0\\0 & 0 & 0 & 0 & 0 & 1 & 0 & 0\\0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 \end{array} \right], \; H = \left[ \begin{array}{cccccccc} 0&0&0&0&0&0&0& \hspace{-0.05in}-1 \hspace{-0.05in}\\1&0&0&0&0&0&0& \hspace{-0.05in}-1 \hspace{-0.05in}\\0&1&0&0&0&0&0& \hspace{-0.05in}-1 \hspace{-0.05in}\\0&0&1&0&0&0&0& \hspace{-0.05in}-1 \hspace{-0.05in}\\0&0&0&0&0&0&0&0\\0&0&0&1&0&0&0& \hspace{-0.05in}-1 \hspace{-0.05in}\\0&0&0&0&1&0&0&0\\0&0&0&0&0&1&0& \hspace{-0.05in}-1 \hspace{-0.05in}\\0&0&0&0&0&0&0&0 \end{array} \right], \; B = \left[ \begin{array}{cc} 0 & 1 \\ 0 & 1 \\ 0 & 1 \\ \hspace{-0.05in}-1 \hspace{-0.05in} & 1 \\ 1 & 0 \\ 0 & 1 \\ 0 & 0 \\ 0 & 1 \\ 0 & 1 \end{array} \right], \; C^T = \left[ \begin{array}{c} 0 \\ \hspace{-0.05in}-1 \hspace{-0.05in} \\ 0 \\ 0 \\ \hspace{-0.05in}-1 \hspace{-0.05in} \\ 0 \\ 1 \\ 0 \end{array} \right] $$ And we find the following tree constants: $$\small \begin{split} K_{1} &= (k_4 + k_5) (((k_9 + {\sigma}) k_{14} + k_9 k_{13}) k_{11} + {\sigma}k_{14} k_{10}) k_2 k_6 k_8 k_{12}\\ K_{2} &= (k_{4} + k_{5}) (((k_{9} + {\sigma}) k_{14} + k_{9} k_{13}) k_{11} + {\sigma}k_{14} k_{10}) k_{1} k_{6} k_{8} k_{12}\\ K_{3} &= k_{6} (((k_{9} + {\sigma}) k_{14} + k_{9} k_{13}) k_{11} + {\sigma}k_{14} k_{10}) k_{12} k_{1} k_{8} k_{3}\\ K_{4} &= (k_{7} + k_{8}) (((k_{9} + {\sigma}) k_{14} + k_{9} k_{13}) k_{11} + {\sigma}k_{14} k_{10}) k_{5} k_{1} k_{3} k_{12}\\ K_{5} &= k_{5} (((k_{9} + {\sigma}) k_{14} + k_{9} k_{13}) k_{11} + {\sigma}k_{14} k_{10}) k_{12} k_{1} k_{6} k_{3}\\ K_{6} &= (k_{10} + k_{11}) k_{12} (k_{13} + k_{14}) k_{1} k_{3} k_{5} k_{6} k_{8}\\ K_{7} &= k_{1} k_{12} k_{3} k_{5} k_{6} k_{8} k_{9} (k_{13} + k_{14})\\ K_{8} &= (k_{13} + k_{14}) k_{1} k_{3} k_{5} k_{6} k_{8} {\sigma}(k_{10} + k_{11})\\ K_{9} &= k_{1} k_{3} k_{5} k_{6} k_{8} {\sigma}(k_{10} + k_{11}) k_{12} \end{split}$$ Constructing ${\kappa}={\kappa}(k^*,{\sigma})$ according to the spanning forest $F = (V,\mathcal{E})$ as above gives the $\tilde \delta = 1$ condition $${\kappa}(k^*,{\sigma})^C = \left(\frac{K_{3}}{K_{1}}\right)^{-1}\left(\frac{K_{6}}{K_{1}}\right)^{-1} \left(\frac{K_{8}}{K_{1}}\right) = \frac{k_2(k_4+k_5){\sigma}}{k_1k_3k_{12}} = 1,$$ which can be solved explicitly for ${\sigma}$ (in terms of $k^*$), $${\sigma}= \frac{k_1k_3k_{12}}{k_2(k_4+k_5)}.$$ By Theorem \[main2\] (statement 2), we have a monomial parametrization of the form . In particular, we obtain: $$\small \begin{split} XD & = \left( \frac{((k_{2} (k_{4}+k_{5}) (k_{13}+k_{14}) k_{9}+k_{1} k_{12} k_{14} k_{3}) k_{11}+k_{1} k_{3} k_{12} k_{14} k_{10}) (k_{4}+k_{5}) k_{2}}{(k_{10}+k_{11}) k_{12} k_{5} k_{3}^2 k_{1}^2 } \right){\tau}_1\\ X & = \left( \frac{(((k_{2} (k_{4}+k_{5}) k_{9}+k_{1} k_{3} k_{12} k_{14}+k_{13} k_{2} k_{9} (k_{4}+k_{5})) k_{11}+k_{1} k_{3} k_{12} k_{14} k_{10}) (k_{4}+k_{5})}{ (k_{10}+k_{11}) k_{12} k_{1} k_{5} k_{3}^2} \right){\tau}_1\\ XT & = \left( \frac{((k_{2} (k_{4}+k_{5}) k_{9}+k_{1} k_{3} k_{12}) k_{14}+k_{13} k_{2} k_{9} (k_{4}+k_{5})) k_{11}+k_{1} k_{3} k_{12} k_{14} k_{10}}{(k_{10}+k_{11}) k_{12} k_{1} k_{3} k_{5} } \right){\tau}_1\\ X_p & = \left( \frac{(((k_{2} (k_{4}+k_{5}) k_{9}+k_{1} k_{3} k_{12}) k_{14}+k_{13} k_{2} k_{9} (k_{4}+k_{5})) k_{11}+k_{1} k_{3} k_{12} k_{14} k_{10}) (k_{7}+k_{8}) }{(k_{10}+k_{11}) k_{12} k_{3} k_{1} k_{8} k_{6}} \right)\frac{{\tau}_1 }{{\tau}_2}\\ Y & = {\tau}_2\\ X_pY & = \left( \frac{((k_{2} (k_{4}+k_{5}) k_{9}+k_{1} k_{3} k_{12}) k_{14}+k_{13} k_{2} k_{9} (k_{4}+k_{5})) k_{11}+k_{1} k_{3} k_{12} k_{14} k_{10}}{(k_{10}+k_{11}) k_{12} k_{3} k_{1} k_{8}} \right){\tau}_1\\ Y_p & = \left( \frac{((k_{2} (k_{4}+k_{5}) k_{9}+k_{1} k_{3} k_{12}) k_{14}+k_{13} k_{2} k_{9} (k_{4}+k_{5})) k_{11}+k_{1} k_{3} k_{12} k_{14} k_{10}}{k_{5} k_{3} k_{1} (k_{13}+k_{14}) (k_{10}+k_{11})} \right)\\ XDY_p & = \left( \frac{k_{2} (k_{4}+k_{5}) (k_{13}+k_{14}) k_{9}}{(k_{10}+k_{11}) k_{1} k_{3} k_{12}} \right) {\tau}_1\\ XTY_p & = {\tau}_1 \end{split}$$ over ${\tau}_1, {\tau}_2 \in {\mathbb{R}}_{>0}$. This parametrization was obtained via alternative methods in [@M-D-S-C] and [@J2014]. Note that the concentration of $Y_p$ does not depend upon either parameter ${\tau}_1$ or ${\tau}_2$. Hence it takes the same value at every positive steady state. This property has been called *absolute concentration robustness* (ACR) in the literature, and the robust steady state value of $Y_p$ has been obtained by other methods in [@Sh-F; @Karp; @Tonello2017; @M-D-S-C]. Consider the model for the Shuttled WNT signaling pathway from [@G-H-R-S], which has a deficiency of four ($\delta = 4$), taken with the following translation scheme: $$\small \label{wnt} \begin{tikzcd} X_1 \arrow[r, yshift=+0.5ex, "k_1"] & X_2 \arrow[l, yshift=-0.5ex, "k_2"] \arrow[r, yshift=0.5ex, "k_{3}"] & X_3 \arrow[l,yshift=-0.5ex, "k_{4}"] & (+ 0)\\[-0.2in] X_5 \arrow[r, yshift=+0.5ex,"k_{5}"] & X_7 \arrow[l, yshift=-0.5ex,"k_{6}"] & & (+ 0) \\[-0.2in] X_{11}+ X_{12} \arrow[r, yshift=+0.5ex,"k_{7}"] & X_{13} \arrow[l, yshift=-0.5ex,"k_{8}"] & & (+ 0) \\[-0.2in] X_{3} + X_6 \arrow[r, yshift=+0.5ex,"k_{9}"] & X_{15} \arrow[l, yshift=-0.5ex,"k_{10}"] \arrow[r, "k_{11}"] & X_3 + X_7 & (+ X_9)\\[-0.2in] X_{7} + X_9 \arrow[r, yshift=+0.5ex,"k_{12}"] & X_{17} \arrow[l, yshift=-0.5ex,"k_{13}"] \arrow[r, "k_{14}"] & X_6 + X_9 & (+ X_3)\\[-0.2in] X_{2} + X_4 \arrow[r, yshift=+0.5ex,"k_{15}"] & X_{14} \arrow[l, yshift=-0.5ex,"k_{16}"] \arrow[r, "k_{17}"] & X_2 + X_5 & (+ X_8)\\[-0.2in] X_{5} + X_8 \arrow[r, yshift=+0.5ex,"k_{18}"] & X_{16} \arrow[l, yshift=-0.5ex,"k_{19}"] \arrow[r, "k_{20}"] & X_4 + X_8 & (+ X_2)\\[-0.2in] X_{4} + X_{10} \arrow[r, yshift=+0.5ex,"k_{21}"] & X_{18} \arrow[l, yshift=-0.5ex,"k_{22}"] \arrow[r, "k_{23}"] & X_4 & (+ X_6)\\[-0.2in] X_{6} + X_{11} \arrow[r, yshift=+0.5ex,"k_{24}"] & X_{19} \arrow[l, yshift=-0.5ex,"k_{25}"] \arrow[r, "k_{26}"] & X_6 & (+ X_4) \\[-0.2in] X_{10} \arrow[rr, yshift=+0.5ex,"k_{27}"] \arrow[ddr, yshift=0.5ex,"k_{29}"] & & X_{11} \arrow[ll, yshift=-0.5ex,"k_{28}"] \arrow[ddl,"k_{31}"] & \\[-0.2in] & & & (+ X_4 + X_6)\\[-0.2in] & \emptyset \arrow[uul,yshift=-0.5ex,"k_{30}"] & & \end{tikzcd}$$ In the representation above, we have kept the indexing of the species $X_1$ through $X_{19}$ as in [@G-H-R-S], but renamed the rate constants. Via Lemmas \[lemma:equiv1\] and \[lemma:equiv2\], the network corresponds to a weakly reversible, [$V^*$-directed]{}GCRN: $$\small \begin{tikzcd}[column sep=0.5cm] \mbox{\ovalbox{$\begin{array}{c} 1 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} X_1 \\ (X_1) \end{array}$}} \arrow[r, yshift=+0.5ex, "k_1"] & \mbox{\ovalbox{$\begin{array}{c} 2 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} X_2 \\ (X_2) \end{array}$}} \arrow[l, yshift=-0.5ex, "k_2"] \arrow[r, yshift=0.5ex, "k_{3}"] & \mbox{\ovalbox{$\begin{array}{c} 3 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} X_3 \\ (X_3) \end{array}$}} \arrow[l,yshift=-0.5ex, "k_{4}"] &\\[-0.2in] \mbox{\ovalbox{$\begin{array}{c} 4 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} X_5 \\ (X_5) \end{array}$}} \arrow[r, yshift=+0.5ex,"k_{5}"] & \mbox{\ovalbox{$\begin{array}{c} 5 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} X_7 \\ (X_7) \end{array}$}} \arrow[l, yshift=-0.5ex,"k_{6}"] & \mbox{\ovalbox{$\begin{array}{c} 6 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} X_{11}+ X_{12} \\ (X_{11} + X_{12}) \end{array}$}} \arrow[r, yshift=+0.5ex,"k_{7}"] & \mbox{\ovalbox{$\begin{array}{c} 7 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} X_{13} \\ (X_{13}) \end{array}$}} \arrow[l, yshift=-0.5ex,"k_{8}"] \\ \mbox{\ovalbox{$\begin{array}{c} 8 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} X_3+X_6+X_{9} \\ (X_3 + X_{6}) \end{array}$}} \arrow[r,yshift=0.5ex, "k_{9}"] & \mbox{\ovalbox{$\begin{array}{c} 9 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} X_9 + X_{15} \\ (X_{15}) \end{array}$}} \arrow[l,yshift=-0.5ex,"k_{10}"] \arrow[d,"k_{11}"] & \mbox{\ovalbox{$\begin{array}{c} 12 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} X_2+X_4+X_{8} \\ (X_2 + X_{4}) \end{array}$}} \arrow[r,yshift=0.5ex, "k_{15}"] & \mbox{\ovalbox{$\begin{array}{c} 13 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} X_8 + X_{14} \\ (X_{14}) \end{array}$}} \arrow[l,yshift=-0.5ex,"k_{16}"] \arrow[d,"k_{17}"]\\ \mbox{\ovalbox{$\begin{array}{c} 11 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} X_3+X_{17} \\ (X_{17}) \end{array}$}} \arrow[r,yshift=-0.5ex,"k_{13}"'] \arrow[u,"k_{14}"] & \mbox{\ovalbox{$\begin{array}{c} 10 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} X_3+X_7 + X_9 \\ (X_7 + X_9) \end{array}$}} \arrow[l,yshift=0.5ex,"k_{12}"']& \mbox{\ovalbox{$\begin{array}{c} 15 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} X_2+X_{16} \\ (X_{16}) \end{array}$}} \arrow[r,yshift=-0.5ex,"k_{19}"'] \arrow[u,"k_{20}"] & \mbox{\ovalbox{$\begin{array}{c} 14 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} X_2+X_5 + X_8 \\ (X_5 + X_8) \end{array}$}} \arrow[l,yshift=0.5ex,"k_{18}"'] \end{tikzcd}$$ $$\label{mess1} \begin{tikzcd}[column sep=0.5cm] \mbox{\ovalbox{$\begin{array}{c} 16 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} X_4+X_6+X_{10} \\ (X_4 + X_{10}) \end{array}$}} \arrow[d,"k_{21}"] & \mbox{\ovalbox{$\begin{array}{c} 17 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} X_4 + X_6 + X_{10} \\ (X_{10}) \end{array}$}} \arrow[l,color=red, "{\sigma}_1"'] \arrow[r, yshift=+0.5ex, "k_{27}"] \arrow[d,xshift=0.5ex,"k_{29}"]& \mbox{\ovalbox{$\begin{array}{c} 18 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} X_4+X_6+X_{11} \\ (X_{11}) \end{array}$}} \arrow[r,color=red, "{\sigma}_{2}"] \arrow[l, yshift=-0.5ex, "k_{28}"] \arrow[dl,"k_{31}"]& \mbox{\ovalbox{$\begin{array}{c} 19 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} X_4+X_6+X_{11} \\ (X_6+X_{11}) \end{array}$}} \arrow[dl,"k_{24}"]\\ \mbox{\ovalbox{$\begin{array}{c} 20 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} X_6+X_{18} \\ (X_{18}) \end{array}$}} \arrow[r,"k_{23}"'] \arrow[ur,"k_{22}"'] & \mbox{\ovalbox{$\begin{array}{c} 21 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} X_4+X_6 \\ (0) \end{array}$}} \arrow[u,xshift=-0.5ex,"k_{30}"] & \mbox{\ovalbox{$\begin{array}{c} 22 \\ \\ \end{array}\Bigg\lvert \begin{array}{c} X_4 + X_{19} \\ (X_{19}) \end{array}$}} \arrow[l,"k_{26}"] \arrow[u,"k_{25}"'] & \end{tikzcd}$$ Thereby, $17 \to 16$ and $18 \to 19$ (with labels ${\sigma}_1> 0$ and ${\sigma}_2 > 0$) are phantom egdes since $y(16)=y(17)=X_4+X_6+X_{10}$ and $y(18)=y(19)=X_4+X_6+X_{11}$. The network is [$V^*$-directed]{}for $V^* = V \setminus \{16,19\}$. It can be quickly checked that the GCRN has a stoichiometric deficiency of two ($\delta = 2$) but effective and kinetic deficiencies of zero ($\delta' = 0$ and $\tilde \delta = 0$). By Theorems \[main\] and \[main2\] (statement 1), the equilibrium set can be parametrized by . Explicitly, we choose the spanning forest $F = (V, \mathcal{E})$ for the graph  consisting of the edges $1 \to i$ for $i \in \{2, 3 \}$, $4 \to 5$, $6 \to 7$, $8 \to i$ for $i \in \{ 9, 10, 11\}$, $12 \to i$ for $i \in \{ 13, 14, 15 \}$, and $16 \to i$ for $i \in \{ 17, \ldots, 22 \}$. Then we compute the corresponding matrix $M$: $$\tiny M = \left[ \begin{array}{cccccccccccccccc} -1&-1&0&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 1&0&0&0&0&0&0&-1&-1&-1&0&0&0&0&0&0\\ 0&1&0&0&-1&-1&-1&0&0&0&0&0&0&0&0&0\\ 0&0&0&0&0&0&0&-1&-1&-1&-1&-1&-1&-1&-1&-1\\ 0&0&-1&0&0&0&0&0&1&0&0&0&0&0&0&0\\ 0&0&0&0&-1&-1&-1&0&0&0&0&0&1&0&0&0\\ 0&0&1&0&0&1&0&0&0&0&0&0&0&0&0&0\\ 0&0&0&0&0&0&0&0&1&0&0&0&0&0&0&0\\ 0&0&0&0&0&1&0&0&0&0&0&0&0&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&-1&-1&-1&-1&-1\\ 0&0&0&-1&0&0&0&0&0&0&0&1&1&0&0&0\\ 0&0&0&-1&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&0&1&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&0&0&0&0&0&1&0&0&0&0&0&0&0&0\\ 0&0&0&0&1&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0\\ 0&0&0&0&0&0&1&0&0&0&0&0&0&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1 \end{array} \right]$$ We have rank$(M)=16$ and therefore nullity$(M^T)=19-16=3$. A matrix $B$ with with $\operatorname{im}B = \ker M^T$ and $\ker B = \{0\}$ is given by $$\tiny B^T = \left[ \begin{array}{ccccccccccccccccccc} 1 & 1 & 1 & 0 & 1 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 1 & 1 & 1 & 0 & 0 \\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1& 1 & 0 & 0 & 0 & 0 & 0 & 0 \\ 1 & 1 & 1 & 0 & 0 & 0 & 0 & 1 & 1 & 0 & 0 & 0 & 0 & 1 & 1 & 1 & 1 & 0 & 0 \end{array} \right] .$$ By reducing $M^T$ to row echelon form, we obtain the following generalized inverse of $M^T$: $$\tiny H = \left[ \begin{array}{cccccccccccccccc} 0&-1&0&0&0&-1&0&0&0&0&0&1&-1&0&0&0\\ 1&-1&0&0&0&-1&0&0&0&0&0&1&-1&0&0&0\\ 0&0&0&0&0&-1&0&0&0&0&0&1&-1&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&-1&0&0&0&0&0\\ 0&0&-1&0&0&-1&1&0&0&0&0&0&0&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&-1&1&0&0&0\\ 0&0&0&0&0&-1&1&0&0&0&0&0&0&0&0&0\\ 1&-1&1&0&0&0&-1&0&0&1&-1&1&-1&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&1&0&0&0&-1&0\\ 0&0&0&0&0&0&0&0&0&0&0&1&0&0&-1&0\\ 0&0&0&-1&0&0&0&0&0&0&0&-1&0&0&1&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 1&-1&0&0&0&-1&0&1&0&0&-1&1&-1&0&0&0\\ 0&0&0&0&1&-1&0&0&0&0&0&0&0&0&0&0\\ 1&-1&0&0&0&-1&0&0&1&0&-1&1&-1&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&1&-1&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&0&-1&1 \end{array} \right]$$ That is, $M^T H M^T = M^T$. From the graph , we obtain the tree constants $K=K(k^*,{\sigma})$: $$\tiny \begin{array}{|l|l|l|} \hline K_{1} = k_2 k_4 & K_{8} = (k_{10}+k_{11})k_{12}k_{14} & K_{15} = k_{15}k_{17}k_{18} \\ K_{2} = k_1 k_4 & K_{9} = k_9k_{12}k_{14} & K_{16} = (k_{22}+k_{23})k_{24}k_{30}((k_{28}+{\sigma}_2+k_{31})k_{26}+k_{25}(k_{28}+k_{31})){\sigma}_1\\ K_{3} = k_1 k_3 & K_{10} = k_9k_{11}(k_{13}+k_{14}) & K_{17} = k_{21}(k_{22}+k_{23})k_{24}k_{30}((k_{28}+{\sigma}_2+k_{31})k_{26}+k_{25}(k_{28}+k_{31}))\\ K_{4} = k_6 & K_{11} = k_9k_{11}k_{12} & K_{18} = k_{21}(k_{22}+k_{23})k_{24}(k_{25}+k_{26})k_{27}k_{30}\\ K_{5} = k_5 & K_{12} = (k_{16}+k_{17})k_{18}k_{20} & K_{19} = k_{21}(k_{22}+k_{23})(k_{25}+k_{26})k_{27}k_{30}{\sigma}_2\\ K_{6} = k_8 & K_{13} = k_{15}k_{18}k_{20} & K_{20} = k_{21}k_{24}k_{30}((k_{28}+{\sigma}_2+k_{31})k_{26}+k_{25}(k_{28}+k_{31})){\sigma}_1\\ K_{7} = k_7 & K_{14} = k_{15}k_{17}(k_{19}+k_{20})& K_{22} = k_{21}(k_{22}+k_{23})k_{24}k_{27}k_{30}{\sigma}_2 \\ \hline \multicolumn{3}{|l|}{\begin{array}{l} K_{21} = ((((k_{28}+{\sigma}_2+k_{31})k_{29}+({\sigma}_1+k_{27})k_{31}+{\sigma}_2k_{27}+{\sigma}_1({\sigma}_2+k_{28}))k_{26}+k_{25}((k_{28}+k_{31})k_{29}+({\sigma}_1+k_{27})k_{31}+{\sigma}_1k_{28}))k_{23}\\ \hspace{1.5in} + k_{22}(((k_{28}+{\sigma}_2+k_{31})k_{29}+k_{27}(k_{31}+{\sigma}_2))k_{26}+k_{25}((k_{28}+k_{31})k_{29}+k_{31}k_{27})))k_{24}k_{21}\end{array}} \\ \hline \end{array}$$ As a result, the parametrization  amounts to [3]{} $x_1 = \left( \frac{K_{3} K_{10} K_{19}}{K_{1} K_{8} K_{18}} \right) {\tau}_1 {\tau}_3\\ x_2 = \left( \frac{K_{3} K_{10} K_{19}}{K_{2} K_{8} K_{18} } \right) {\tau}_1 {\tau}_3\\ x_3 = \left( \frac{K_{10} K_{19}}{ K_{8} K_{18} } \right){\tau}_1 {\tau}_3\\ x_4 = \left( \frac{K_{17}}{K_{16}} \right)\\ x_5 = \left( \frac{K_{5} K_{10}}{ K_{4} K_{11} }\right) {\tau}_1\\ x_6 = \left( \frac{K_{18}}{K_{19}} \right)\\ x_7 = \left( \frac{K_{10}}{K_{11}} \right) {\tau}_1\\ x_8 = \left( \frac{K_{3} K_{4} K_{11} K_{12} K_{17} K_{19} }{ K_{2} K_{5} K_{8} K_{15} K_{18} K_{16} } \right) {\tau}_3\\ x_9 = {\tau}_3\\ x_{10} = \left( \frac{K_{21}}{K_{17}} \right)\\ x_{11} = \left( \frac{K_{21}}{K_{18}} \right)\\ x_{12} = \left( \frac{K_{7} K_{18}}{ K_{6} K_{21} }\right) {\tau}_2\\ x_{13} = {\tau}_2\\ x_{14} = \left( \frac{K_{3} K_{10} K_{12} K_{17} K_{19}}{ K_{2} K_{8} K_{13} K_{18} K_{16} } \right) {\tau}_1 {\tau}_3\\ x_{15} = \left( \frac{K_{10}}{K_{9}} \right) {\tau}_1 {\tau}_3\\ x_{16} = \left( \frac{K_{3} K_{10} K_{12} K_{17} K_{19}}{ K_{2} K_{8} K_{14} K_{18} K_{16} } \right) {\tau}_1 {\tau}_3\\ x_{17} = {\tau}_1 {\tau}_3\\ x_{18} = \left( \frac{K_{21}}{K_{20}} \right) \\ x_{19} = \left( \frac{K_{21}}{K_{22}} \right)$ with $K=K(k^*,{\sigma})$ as above and ${\sigma}_1, {\sigma}_2, {\tau}_1, {\tau}_2, {\tau}_3 > 0$. Outlook {#sec:conclusions} ======= We have presented sufficient conditions for determining whether the set of positive equilibria of a generalized mass-action system coincides with the parametrized set of complex-balanced equilibria. We have also presented sufficient conditions for guaranteeing a positive parametrization of the set of complex-balanced equilibria and for effectively constructing the parametrization. Through an extension of network translations [@J2014], we have shown how the result can be immediately applied to biochemical reaction networks, including the EnvZ-OmpR signaling pathway [@Sh-F] and shuttled WNT signaling pathway [@G-H-R-S]. A number of potential avenues for further research naturally emerge from this work. 1. Recent work on generalized mass-action systems has established sign conditions sufficient for the uniqueness of equilibrium points in compatibility classes [@BanajiPantea16; @M-F-R-C-S-D]. In particular, when the steady state set is *toric* or *complex-balanced*, uniqueness and multistationarity may be established [@MR2012; @M-D-S-C]. It is currently unclear, however, whether the extension to rational parametrizations in Theorem \[main2\] might be utilized to guarantee either uniqueness or multistationarity. 2. For GCRNs with nonzero kinetic deficiency ($\tilde \delta > 0$), statement 2 in Theorem \[main2\] guarantees that, if the parameters ${\sigma}\in {\mathbb{R}}_{> 0}^{E_0}$ can be chosen to satisfy the $\tilde \delta > 0$ conditions for complex-balancing, then the system has a monomial parametrization. It is currently unclear which conditions guarantee that a set of free parameters ${\sigma}\in {\mathbb{R}}_{> 0}^{E_0}$ may satisfy the $\tilde \delta > 0$ algebraic conditions on the rate parameters required for complex balancing. 3. Even for biochemical networks of moderate size, it is difficult to determine a translation scheme for constructing a GCRN corresponding to the original CRN. Computational approaches to network translation have been investigated in [@J2015] and [@Tonello2017]. These works, however, rely on the definitions of a GCRN in [@MR2012] and of network translation in [@J2014]. Using the more general definitions in [@MR2014] would allow to extend the applicability of the computational approaches to a significantly broader class of networks. Acknowledgments {#acknowledgments .unnumbered} --------------- This project began during the ‘SQuaRE’ *Dynamical properties of deterministic and stochastic models of reaction networks* at the American Institute of Mathematics (AIM), San Jose, California, in October 2017. SM was supported by the Austrian Science Fund (FWF), project P28406. CP was supported by NSF-DMS award 1517577. [10]{} Murad Banaji and Casian Pantea. Some results on injectivity and multistationarity in chemical reaction networks. , 15:807-869, 2016. Adi Ben-Israel and Thomas N.E. Greville. . Springer, New York, 2nd edition, 2003. Carsten Conradi, Elisenda Feliu, Maya Mincheva, and Carsten Wiuf. Identifying parameter regions for multistationarity. , 13(10):e1005751, 2016. Carsten Conradi, Dietrich Flockerzi, and Jorg Raisch. Multistationarity in the activation of a [MAPK]{}: Parametrizing the relevant region in parameter space. , 211:105–131, 2008. Carsten Conradi and Maya Mincheva. Graph-theoretic analysis of multistationarity using degree theory. , 133:76–90, 2017. Gheorghe Craciun, Alicia Dickenstein, Anne Shiu, and Bernd Sturmfels. Toric dynamical systems. , 44(11):1551–1565, 2009. Martin Feinberg. Lectures on chemical reaction networks. Unpublished written versions of lectures given at the Mathematics Research Center, University of Wisconsin, 1979. Martin Feinberg. Complex balancing in general kinetic systems. , 49:187–194, 1972. Martin Feinberg. Chemical reaction network structure and the stability of complex isothermal reactors: [I.]{} the deficiency zero and deficiency one theorems. , 42(10):2229–2268, 1987. Martin Feinberg. The existence and uniqueness of steady states for a class of chemical reaction networks. , 132:311–370, 1995. E. Gross, H.A. Harrington, Z. Rosen, and B. Sturmfels. Algebraic [S]{}ystems [B]{}iology: [A]{} [C]{}ase [S]{}tudy for the [W]{}nt [P]{}athway. , 78:21–51, 2016. Katharina Holstein, Dietrich Flockerzi, and Carsten Conradi. Multistationarity in sequentially distributed multisite phosphorylation networks. , 75(11):2028–2058, 2013. Fritz Horn. Necessary and sufficient conditions for complex balancing in chemical kinetics. , 49:172–186, 1972. Fritz Horn and Roy Jackson. General mass action kinetics. , 47:81–116, 1972. Matthew D. Johnston. Translated chemical reaction networks. , 76(5):1081–1116, 2014. Matthew D. Johnston. A computational approach to steady state correspondence of regular and generalized mass action systems. , 77(6):1065–1100, 2015. R.L. Karp, M. Pérez Millán, T. Dasgupta, A. Dickenstein, and J. Gunawardena. Complex-linear invariants of biochemical networks. , 311:130–138, 2012. Mercedes Pérez Millán and Alicia Dickenstein. The structure of [MESSI]{} biological systems. 2016. Available on the ArXiv at arxiv:1612.08763. Mercedes Pérez Millán, Alicia Dickenstein, Anne Shiu, and Carsten Conradi. Chemical reaction systems with toric steady states. , 74(5):1027–1065, 2012. Stefan Müller, Elisenda Feliu, Georg Regensburger, Carsten Conradi, Anne Shiu, and Alicia Dickenstein. Sign conditions for injectivity of generalized polynomial maps with applications to chemical reaction networks and real algebraic geometry. , 16(1):69–97, 2016. Stefan Müller and Georg Regensburger. Generalized mass action systems: Complex balancing equilibria and sign vectors of the stoichiometric and kinetic-order subspaces. , 72(6):1926–1947, 2012. Stefan Müller and Georg Regensburger. Generalized mass-action systems and positive solutions of polynomial equations with real and symbolic exponents (invited talk). In Gerdt V.P., Koepf W., Seiler W.M., and Vorozhtsov E.V., editors, [*Computer Algebra in Scientific Computing. CASC 2014. Lecture Notes in Computer Science*]{}, volume 8660, pages 302–323. Springer, 2014. Guy Shinar and Martin Feinberg. Structural sources of robustness in biochemical reaction networks. , 327(5971):1389–1391, 2010. Matthew Thomson and Jeremy Gunawardena. The rational parameterisation theorem for multisite post-translational modification systems. , 261(4):626–636, 2009. Elisa Tonello and Matthew D. Johnston. Network translation and steady state properties of chemical reaction systems. Accepted in [*Bull. Math. Biol.*]{}, 2018. [^1]: Department of Mathematics, San Jose State, One Washington Square, San Jose, CA, USA 95192 [^2]: Faculty of Mathematics, University of Vienna, Austria [^3]: Department of Mathematics, West Virginia University, Morgantown, WV, USA 26506

ArXiv

--- abstract: 'Oe stars are a subset of the O-type stars that exhibit emission lines from a circumstellar disk. The recent detection of magnetic fields in some O-type stars suggests a possible explanation for the stability of disk-like structures around Oe stars. According to this hypothesis, the wind of the star is channeled by a dipolar magnetic field producing a disc in the magnetic equatorial plane. As a test of this model, we have obtained spectropolarimetric observations of the hottest Galactic Oe star HD 155806. Here we discuss the results and implications of those observations.' date: '?? and in revised form ??' --- ![Least Squares Deconvolved profiles in June (left) and July (right). The curves are the mean Stokes I profiles (bottom), the mean Stokes V profiles (top) and the N diagnostic null profiles (middle). The bold line represents the smallest-amplitude Stokes V profile corresponding to the longitudinal field of 115 G reported by Hubrig et al. (2007), which is a pure dipole with a polar strength of 380G.[]{data-label="fig:fig1"}](VPetit2_fig1.eps){width="13cm"} Oe stars are a subset of the O-type stars that exhibit emission in spectra of their Balmer lines, indicating the presence of a circumstellar disk. It is generally believed that the low incidence of the Oe stars (only 4 are known in the Milky Way) is caused by the increasingly powerful radiation-driven winds of the O-type stars, which inhibit the formation of long-lived circumstellar structures. However, all the *bona fide* Oe stars exhibit typical wind profiles in their ultraviolet resonance lines, showing that these stars have winds similar to those of other O-type stars. The mystery is therefore not “why are there so few Oe stars?”, but “why are there any at all?”. The recent detection of magnetic fields in O-type stars suggests a possible explanation for the stability of disk-like structures around Oe stars. From MHD simulations, ud-Doula et al. (2008) have shown that under high magnetic confinement and high rotation, a “magnetically confined wind” can trap material in its magnetosphere. The circumstellar structure is composed of different material at different times, but its longevity is mediated by the magnetic field of the star. Consequently, this model neatly avoids problems associated with the disruption of the disks of Oe stars, and suggests that Oe stars might represent extreme cases of a more widespread phenomenon that depends on the balance between the strength of the stellar wind and the strength of the magnetic field. HD155806 is the Galactic Oe star with the most powerful wind. We estimated that a surface dipolar magnetic field of at least 235 G is required to confine its wind. It exhibits a large, single-peaked H$\alpha$ emission profile that implies that the disk is not being viewed “edge-on”. Encouragingly, the detection of a weak magnetic field ($115\pm37$G) in HD155806 with VLT/FORS1 has recently been reported by Hubrig et al. (2007), although this detection was not reproduced in further measurements (). In addition, a re-analysis of the FORS1 observations (available in ESO archives) as a component of this present study led to less conclusive results. New high-resolution spectropolarimetric observations of HD155806 were obtained with ESPaDOnS at CFHT in June and July 2008. No magnetic Stokes V signature is detected in our observations (Figure 1). In order to extract the surface field characteristics constrained by the observed Stokes V profiles, we are in the process of comparing them with theoretical profiles derived from a large grid of dipolar magnetic field configurations that were calculated with the polarised LTE radiative transfer code Zeeman2 ([@1988ApJ...326..967L]). With the Bayesian method described by Petit et al. (2008), we will obtain an upper limit on the strength of the dipolar magnetic field at the stellar surface that is independent of the magnetic configuration. We shall also use the FORS1 measurements as additional constraints. Although detailed modeling still needs to be performed, we estimate that the upper limit on the surface magnetic dipole will be about 200G, according to our preliminary analysis. A stronger magnetic field would need to be in a configuration that produces a weak or null Stokes V signature. For dipoles, these rare configurations have well-defined geometries and correspond to the observer looking directly at the magnetic equator. According to MHD simulations of an aligned dipole rotator ([@2008MNRAS.385...97U]), an accumulation at the equator of the magnetosphere will occur when both the confinement parameter and the rotation rate are large. Rigid-field hydrodynamics modelling of a tilted dipole ([@2007MNRAS.382..139T]) predicts that the disk will have an average inclination that lies somewhere between the rotation and magnetic equatorial planes, with plasma concentrated at the intersection between them. If the disk of HD155806 is produced by a large magnetic field hidden in a Stokes V-free configuration, we would expect to see a more edge-on disk signature in the H$\alpha$ emission line than that which is observed at the phase when the Stokes V signature is null. Since a field with a surface dipolar strength lower than 200G would not confine the wind of HD155806 sufficiently to produce a disk; and since a stronger (but undetected) field would be in a highly improbable configuration that would in any case be unable to produce the emission structure seen in H$\alpha$, we conclude that it is unlikely that the disk of HD155806 is caused by a large-scale magnetic field confining the stellar wind. , S., [Yudin]{}, R. V., [Pogodin]{}, M., [Schller]{}, M. & [Peters]{}, G. J. 2007, *AN* 328, 1133 , S., [Schöller]{}, M., [Schnerr]{}, R. S., [González]{}, J. F. et al. 2008, *A&A* 490, 793 , J. D. 1988, *ApJ* 326, 967 , V., [Wade]{}, G. A., [Drissen]{}, L., [Montmerle]{}, T. & [Alecian]{}, E. 2008, *MNRAS* 387, L23 , R. H. D., [Owocki]{}, S. P., [ud-Doula]{}, A. 2007, *MNRAS* 382, 139 , A., [Owocki]{}, S. P. & [Townsend]{}, R. H. D. 2008, *MNRAS* 385, 97

ArXiv

--- abstract: 'The secular approximation of the hierarchical three body systems has been proven to be very useful in addressing many astrophysical systems, from planets, stars to black holes. In such a system two objects are on a tight orbit, and the tertiary is on a much wider orbit. Here we study the dynamics of a system by taking the tertiary mass to zero and solve the hierarchical three body system up to the octupole level of approximation. We find a rich dynamics that the outer orbit undergoes due to gravitational perturbations from the inner binary. The nominal result of the precession of the nodes is mostly limited for the lowest order of approximation, however, when the octupole-level of approximation is introduced the system becomes chaotic, as expected, and the tertiary oscillates below and above $90^\circ$, similarly to the non-test particle flip behavior [e.g., @Naoz16]. We provide the Hamiltonian of the system and investigate the dynamics of the system from the quadrupole to the octupole level of approximations. We also analyze the chaotic and quasi-periodic orbital evolution by studying the surfaces of sections. Furthermore, including general relativity, we show case the long term evolution of individual debris disk particles under the influence of a far away interior eccentric planet. We show that this dynamics can naturally result in retrograde objects and a puffy disk after a long timescale evolution (few Gyr) for initially aligned configuration.' author: - | Smadar Naoz$^1$ & Gongjie Li$^2$\ Macarena Zanardi$^{3,4}$, Gonzalo Carlos de Elía$^{3,4}$, Romina P. Di Sisto$^{3,4}$ bibliography: - 'Kozai.bib' title: 'The Eccentric Kozai-Lidov mechanism for Outer Test Particle' --- INTRODUCTION ============ The hierarchical three body secular dynamics has been studied extensively in the literature and was shown to be very effective in addressing different astrophysical phenomena [see for review @Naoz16 and reference therein]. In this hierarchical setting the [*inner binary*]{} is orbited by a third body on a much wider orbit, the [*outer binary*]{}, such that the secular approximation can be applied (i.e., phase averaged, long-term interaction). The gravitational potential is then expanded in semi-major axis ratio ($a_1/a_2$, which, in this approximation, remains constant), where $a_1$ ($a_2$) is the semi-major axis of the inner (outer) body [@Kozai; @Lidov]. This ratio is a small parameter due to the hierarchical configuration. The lowest order of approximation, which is proportional to $(a_1/a_2)^2$ is called the quadrupole-level. Most of these studies focus [**on**]{} the gravitational perturbations that a far away perturber exerts on the inner binary. In early studies of high-inclination secular perturbations [@Kozai; @Lidov], the outer orbit was assumed to be circular and it was assumed that one of the inner binary members is a massless test particle. In this situation, the component of the inner orbit’s angular momentum along the z-axis (which is set to be parallel to the total angular momentum, i.e., the invariable plane) is conserved, and the lowest order of the approximation, the quadrupole approximation, is valid. However, relaxing either one of these assumptions leads to qualitative different behavior [e.g., @Naoz11; @LN; @Katz+11]. Considering systems beyond the test particle approximation, or a circular orbit, requires the next level of approximation, called the octupole–level of approximation, which is proportional to $(a_1/a_2)^3$ [e.g. @Har68; @Har69; @Ford00; @Bla+02]. In the octupole level of approximation, the inner orbit eccentricity may reach extreme values [@Ford00; @Naoz+11sec; @Li+13; @Tey+13]. In addition, the inner orbit can flip its orientation, with respect to the total angular momentum (i.e., z-axis), from prograde to retrograde [@Naoz11]. Here we study the secular evolution of a far away test particle orbiting an inner massive binary. In this case, the inner orbit is fixed, and effectively carries all of the annular momentum of the system, while the outer orbit undergoes a dynamical evolution. This situation has large range of applications from the gravitational perturbations of binary super massive black holes on the surrounding stellar distribution to the effects of planetary orbits on debris disks, Oort cloud and trans-neptunian objects. From N-body simulations, @Zanardi+17 analyzed the long term evolution of test particles in the presence of an interior eccentric planet. Such an study produces particles on prograde and retrograde orbits, as well as particles whose orbital plane flips from prograde to retrograde and back again along their evolution. We note that @Ziglin75 investigated the oscillations of an outer circumbinary planet in the context of the restricted elliptical three body problem. Later @Verrier+09 and @Far+10 studied the stability of high inclined planet around in this situation using a combination of numerical and perturbation theory up to the quadruple level of approximation approaches [see also, @Li+14cir; @Marcos+15]. [Furthermore]{}, @Gallardo06 and @Gallardo+12 studied the effects of the Kozai-Lidov for trans-neptunian objects near mean motion resonance with Neptune. However, here, we do not allow for mean motion resonances to allow for the double averaging process [see @Naoz+11sec appendix A2 for the canonical transformation which describes the averaging process].We provide a general treatment for the outer-test particle case, up to the octupole–level of approximation in the secular theory. The paper is organized as follow: We begin by describing the outer test particle Hamiltonian and equations of motion (§\[sec:EOM\]), and continue to discuss the quadrupole-level of approximation (§\[sec:quad\]) where we also drive the relevant timescales, and then we study the role of the octupole-level of approximations, and provide surface of section maps (§\[sec:oct\]). We also discuss the role of general relativity precession in §\[sec:GR\]. We then consider one study case in the form of the long term evolution debris disk particles in §\[sec:Kuip\]. Finally, we offer our discussions in §\[sec:dis\]. The Equations of Motion {#sec:EOM} ======================= We solve the orbit of an [*exterior*]{} massless test particle to an eccentric planet ($m_2$), both orbiting a star ($m_1$), including only secular interactions expanded to octupole order. The planet is on a fixed eccentric orbit (i.e., $e_1={\rm const}$) and the outer particle’s orbit is specified by four variables: $${e_2,\omega_2,\theta,\Omega_2} \ ,$$ where $e_2$ is the test particle eccentricity, $\theta=\cos i$ and $i$ is the inclination of the test particle with respect to the inner orbit, and $\omega_2$ and $\Omega_2$ are the argument of periapse and longitude of ascending node of the outer orbit, relative to the inner planet’s periapse [@MD00]. Specifically, We set $\varpi_1 = 0$, see Appendix \[App:Ham\] for the coordinate transformation. We kept the subscript “2" in $\omega_2$ and $\Omega_2$ for consistency with the comparable masses treatments. From $e_2$ and $\theta$ we can define the canonical specific momenta $$\begin{aligned} J_2&=&\sqrt{1-e_2^2} \\ J_{2,z}&=&\theta \sqrt{1-e_2^2}\end{aligned}$$ The Hamiltonian for which $m_3\to0$ is $$f=f_{\rm quad} + \epsilon_M f_{\rm oct} \ ,$$ where $$\epsilon_M =\frac{m_1-m_2}{m_1+m_2} \frac{a_1}{a_2}\frac{e_2}{1-e_2^2}$$ and $$\label{eq:fquad} f_{\rm quad} = \frac{ (2+3e_1^2)(3\theta^2-1)+15e_1^2(1-\theta^2)\cos(2\Omega_2)}{(1-e_2^2)^{3/2}} \ ,$$ $$\begin{aligned} \label{eq:foct} f_{\rm oct} &= & \frac{ 15 e_1}{4 (1-e_2^2)^{3/2}} \bigg[ 10(1-e_1^2)\theta (1-\theta^2) \sin \omega_2\sin\Omega_2 \nonumber \\ &+& \frac{1}{2}\{ 2+19e_1^2-5(2+5e_1^2)\theta^2-35e_1^2(1-\theta^2)\cos(2\Omega_2) \} \nonumber \\ &\times& (\theta \sin\omega_2 \sin\Omega_2 -\cos\omega_2\cos\Omega_2) \bigg]\end{aligned}$$ Note that unlike the [*inner*]{} test particle approximation $ \epsilon_M$ is not constant during the motion, and the constant parameter during the evolution is: $$\delta=\frac{m_1-m_2}{m_1+m_2} \frac{a_1}{a_2}e_1$$ and the Hamiltonian up to the octupole level of approximation can be defined as: $$\tilde{f}_{\rm oct}= \frac{e_2}{1-e_2^2}\frac{f_{\rm oct} }{e_1}$$ and $$f=f_{\rm quad} + \delta \tilde{f}_{\rm oct} \ .$$ We note that $f_{\rm quad}$ has the same functional form as the inner test particle $F_{\rm quad}$ presented in @LN up to the $(1-e_2^2)^{3/2}$ which is not constant in our case. The equations of motion may be expressed as partial derivatives of an energy function $f(e_2,\omega_2,\theta,\Omega_2)$ $$\begin{aligned} \label{eq:EOM1} \frac{dJ_2}{d\tau} &=& \frac{\partial f}{\partial \omega_2} \\ \frac{dJ_{2,z}}{d\tau} &=&\frac{\partial f}{\partial \Omega_2} \\ \frac{d\omega_2 }{d\tau} &=& \frac{\partial f}{\partial e_2}\frac{J_2}{e_2} + \frac{\partial f}{\partial \theta}\frac{\theta }{J_2} \\\label{eq:EOM4} \frac{d\Omega_2 }{d\tau} &=& - \frac{\partial f}{\partial \theta}\frac{1 }{J_2}\end{aligned}$$ where $\tau$ is proportional to the true time (see Eq. \[\[eq:time\]\]). Unlike inner orbit test particle $\epsilon_M$ is not constant while $e_1$ is constant (in other words, the angular momentum of the inner orbit is conserved). The equations of motion were tested successfully compared to the general equations of motions, presented in @Naoz+11sec. We also test the evolution compared to N-body in Appendix \[sec:Nbody\]. quadrupole-level of approximation {#sec:quad} ================================== General Analysis ---------------- ![. We show the time evolution of the inclination in the left panels and the cross section trajectory in the inclination-$\Omega_2$ plane, in the right panels. We consider two cases, the top panels are for $e_1=0.9$ and the bottom panels are for $e_1=0.3$. We show the following examples: in the circulating mode (setting initially $\Omega_2=0$): $i=20^\circ$ (cyan), $60^\circ$ (blue), and in the librating mode (setting initially $\Omega_2=90^\circ$): $i=20^\circ$ (brown) and $60^\circ$ (magenta). The separatrix corresponds to $i=90^\circ$ for the two different inner eccentricities is shown in red. Note that in the case of $e_1=0.3$ there is no librating mode for $i=20^\circ$. For consistency we adopt the following orbital parameters: $m_1=1$ M$_\odot$, $m_1=1$ M$_J$, $a_1= 3$ AU and $a_2=40$ AU, $\omega_2=90^\circ$ and $e_2=0$. []{data-label="fig:quad"}](QuadTestAnn){width="\linewidth"} The quadrupole-level of approximation is integrable and thus provides a good starting point. Unlike the quadrupole-level approximation for the inner orbit test particle, here the z-component of the particle’s angular momentum is [*not*]{} conserved, as the Hamiltonian depends on $\Omega_2$. However, at this level $J_2$ is conserved, and thus the outer orbit eccentricity $e_2$ remains constant as the Hamiltonian does not depends on $\omega_2$. The equation of motion for the inclination takes a simple form: $$\begin{aligned} \label{eq:dthetadtquad} \frac{d\theta}{d t}\bigg|_{\rm quad} & =& \\&& - \frac{15}{8} \left( \frac{a_1}{a_2}\right)^2 \frac{2\pi}{P_2} \frac{m_1m_2}{(m_1+m_2)^2} e_1^2 \frac{1-\theta^2}{(1-e_2^2)^2} \sin 2\Omega_2 \ ,\nonumber \end{aligned}$$ where we consider the time evolution and not the scaled evolution for completeness, [and $P_2$ is the period of the outer orbit]{}. We can find the maximum and minimum inclination by setting $\dot{\theta}=0$. Thus, we find that the values of the longitude of ascending nodes that satisfy this condition are $\Omega_2=n\pi/2$, where $n=0,1,2..$. In other words, $\Omega_2$ has two classes of trajectories, librating and circulating. The trajectories in the librating region are bound between two values of $\Omega_2$, while the circulation region represents trajectories where the angles are not constrained between two specific values. On circulating trajectories, at $\Omega_2=0$, the inclination ($i<90^\circ$) is largest (where the $i>90^\circ$ case is a mirror image of the $i<90^\circ$ one). The extrema points for the librating mode are located at $\Omega_2=90^\circ$. The time evolution of $\Omega_2$ for the quadrupole-level of approximation is given by: $$\begin{aligned} \label{eq:Omega2dot} \frac{d\Omega_2}{dt}\bigg|_{\rm quad} &=& \\ & & -\frac{m_1 m_2}{(m_1+m_2)^2}\frac{2\pi}{P_2 } \left( \frac{a_1}{a_2} \right)^2 \frac{3 \theta \left( 2+3e_1^2 -5 e_1^2 \cos 2\Omega_2 \right) }{8 (1-e_2^2)^2} \ . \nonumber\end{aligned}$$ In Figure \[fig:quad\] we show the evolution associated for the quadrupole-level of approximation. The two librating and circulating trajectories are considered, where we folded the $\Omega_2$ angle to be between $0-180^\circ$. The librating mode gives the nominal precession of the nodes, at which the inclination oscillates between the $i_{90}$ inclination (the inclination for which $\Omega_2=90^\circ$) and $180^\circ-i_{90}$. The precession of the nodes was noted before in the literature [e.g., @Inn+97]. From the latter Equation and Equation (\[eq:dthetadtquad\]) we have $$\label{eq:OmTh} \frac{d\Omega_2}{d\theta}=\theta \frac{ 2+3e_1^2-5e_1^2\cos 2\Omega_2}{5 e_1^2 (1-\theta^2)\sin 2 \Omega_2} \ .$$ Integrating the two sides we have: $$\label{eq:OmThInt} \int^{\Omega_a}_{\Omega_b}\frac{\sin 2 \Omega_2 }{ 2+3e_1^2-5e_1^2\cos 2\Omega_2} {d\Omega_2}=\int_{\theta_a}^{\theta_b} \frac{\theta}{ 5e_1^2 (1-\theta^2) } {d\theta} \ ,$$ where $\Omega_{a,b}$ is the longitude of ascending nodes that is associated with the inclination value of $\theta_{a,b}=\cos i_{a,b}$.   For the circulating mode, we find that, after integrating over Equation (\[eq:OmTh\]) from $i_{\rm max}$ to $i_{\rm min}$ (and $\Omega_2=0$ to $\Omega_2=90^\circ$, respectively) we get $$\label{eq:imaxmin} \sin i_{\rm min}=\sin i_{\rm max} \sqrt{\frac{1-e_1^2}{1+4e_1^2} } \ .$$ Note that this expression can be also achieved by considering the conservation of energy between the minimum and maximum cases. Setting the initial conditions for the energy, Equation (\[eq:fquad\]), we can find the extrema points as a function of the initial conditions. A special case can be considered when $\Omega_2$ is initially set to be zero, and then the maximum inclination is the initial inclination $i_0$, in other words: $$\label{eq:imim} \sin i_{\rm min}=\sin i_{0} \sqrt{\frac{(1-e_1^2)}{(1+4e_1^2)}} \quad {\rm for } \quad \Omega_{2,0}=0^\circ \ .$$ In other words we can set $i_{\rm max}\to 90^\circ$ and for a given $e_1$ find the largest $i_{\rm min}$ allowed, which corresponds to the separatrix. This relationship is also apparent in figure 14 in @Zanardi+17 numerical results. As can be seen, for the circulating mode depicted in Figure \[fig:quad\], setting initially $i=20^\circ$ corresponds to $i_{\rm min} = 16.25^\circ$ and $10.56^\circ$ for $e_1=0.3$ and $e_1=0.9$, respectively, consistent with Equation (\[eq:imim\]). A comparison between the flip criterion and the numerical results can be seen in Figure \[fig:TwoPanle\_e5\]. For the librating mode we set $i_{\rm min}$ which associated with the $\Omega_2=90^\circ$ case, and thus integrating over Equation (\[eq:OmTh\]) between $i=90^\circ$ to $i_{\rm min}$ (and $\Omega_2=\Omega_{2,{\rm min}}$ to $\Omega_2=90^\circ$, respectively), we get, $$\label{eq:Omegamin} \cos 2\Omega_{2,\rm min} = \frac{2+3e_1^2-(2+8e_1^2)\sin^2 i_{\rm min}}{5e_1^2} \ .$$ Thus, setting initially $\Omega_2=90^\circ$ as in the examples depicted in Figure \[fig:quad\] we find that the minimum value that $\Omega_2$ can achieve in the $e_1=0.9$ case is $59.23^\circ$ for the $i_{\rm min}=60^\circ$ and $15.95^\circ$ for the $i_{\rm min}=20^\circ$ example, consistent with the numerical results. Hence, the range of which $\Omega_2$ is librating on is $2\times (90^\circ-\Omega_{2,{\rm min}})$ Timescale {#sec:Time} --------- The timescale associated with the evolution can be estimated from the equation of motion for $\Omega_2$ at the quadrupole-level \[Equation (\[eq:Omega2dot\])\], for the circulating mode, by setting $d \Omega_2\to \pi$ and taking the terms in the parenthesis to be roughly order of unity (which is achieved, by setting $\Omega_2\to 0$): $$t_{\rm quad} \sim \frac{4}{3} P_2 (1-e_2^2)^2 \frac{ (m_1+m_2)^2}{m_1 m_2} \left( \frac{a_2}{a_1} \right)^2 \quad {\rm circulating.} $$ For the example system depicted in Figure \[fig:quad\], this equation gives a timescale of about $6\times 10^7$ yr, for initial inclination of $60^\circ$, which agrees with the circulating mode (although we note that different $e_1$ give slightly different timescales). We also estimate the timescale in the librating mode by setting $d \Omega_2\to 2\times (90^\circ-\Omega_{2,{\rm min}})$, which we have found earlier. Thus, $$\begin{aligned} t_{\rm quad} &\sim & 2\times \frac{ 2 (\pi/2-\Omega_{2,{\rm min}})}{2\pi} \frac{8}{3} P_2 (1-e_2^2)^2 \times \nonumber \\ && \frac{ (m_1+m_2)^2}{m_1 m_2} \left( \frac{a_2}{a_1} \right)^2 \quad {\rm librating.} $$ The $2$ pre-factor of here comes from numerical comparisons to the examples depicted in Figure \[fig:quad\]. Note the $e_1$ dependency that rises from $\Omega_{2,{\rm min}}$. This timescale is consistent with the examples depicted in Figure \[fig:quad\] by less than a factor of two. The role of the octupole level of approximation {#sec:oct} ================================================ The octupole level of approximation can significantly affect the overall dynamics of the general hierarchical three body system [see @Naoz16 and reference therein]. Specifically, in the [*inner*]{} test particle case, the inner orbit’s z component of the angular momentum is not conserved anymore and the orbit is allowed to flip [for large range of initial inclinations @LN; @Li+13; @Li+14Chaos]. In our case, the $J_{2,z}$ is not conserved at the quadrupole-level, but $J_2$ is. Thus, the octupole level of approximation in this case allows for variations of $e_2$ and introduces higher level resonances, which may result in a chaotic behavior (see below). ![We consider the system: $m_1=1$ M$_\odot$, $m_2=1$ M$_j$, $a_1=0.5$ AU, $a_2=10$ AU, $e_1=0.4$ and $e_2=0.6$. We initialize the system with $\omega_2=\Omega_2=0^\circ$ and $i=85^\circ$. We show the time evolution of the orbital parameters, i.e., argument of pericenter $\omega_2$, longitude of ascending node $\Omega_2$, inclination $i$ and the outer orbital eccentricity, $e_2$. We also consider the inclination evolution as a function of $\Omega_2$. Here both $\Omega_2$ and $\omega_2$ were folded to achieve the $0-180^\circ$ symmetry. []{data-label="fig:Ex1"}](Test7){width="\linewidth"} In Figure \[fig:quadoct\] we consider two representative example for which we compare the quadrupole (blue lines) and octupole (red lines) levels of approximation, where we consider the time evolution of the eccentricity and inclination. In both of these examples we consider a $1$ M$_\odot$ star orbited by an eccentric Jupiter, with a test particle on a far away orbit. One can consider such a setting to represent a result of a scattering event for example. On the left set of panels of Figure \[fig:quadoct\] we consider a Jupiter at $0.4$ AU with $e_1=0.65$ and a test particle at $7$ AU with $e_2=0.4$, initialized on a retrograde orbit ($i=91^\circ$). With the introduction of the octupole level of approximation to the calculation, the test particle eccentricity starts to oscillate, though in this case it never increases pass its initial value (due to choice of initial conditions here). More notably, the test particle inclination, with respect to the total angular momentum, oscillates from retrograde ($>90^\circ$ which was the initial condition) to prograde ($<90^\circ$). As in the more general case, there is no apparent associated timescale for this flipping modulation and it seems chaotic in nature (see below). While the quadrupole-level is circulatory in nature (see $i-\Omega_2$ plot) a libration behavior emerges at the octupole level. On the right set of panels of Figure \[fig:quadoct\] we consider a Jupiter at $3$ AU with $e_1=0.85$ and a test particle at $50$ AU with $e_2=0.7$, and the system is initialized on a prograde orbit ($i=20^\circ$). Here, like the previous example the outer test particle eccentricity begins to oscillate, and even grows above the initial value. However, unlike the previous example this system does not flip. The inclination does oscillate with a long scale modulation and we show the long scale evolution that captures about four octupole cycles. The system does not exhibit a chaotic behavior in this case, and it remains in a circulatory trajectory even after the inclusion of the octupole level of approximation to the calculation. In Figure \[fig:Ex1\] we zoom-in on the evolution of a different example, and also provide the time evolution of $\omega_2$ for the octupole level itself. As in the general hierarchical secular three body problem we find the short scale (associated with the quadrupole) oscillations, that are modulated by the higher level octupole approximation. The octupole modulations take place on timescales which is between few$\times t_{\rm quad}$ to few tens$\times t_{\rm quad}$. {width="\linewidth"} Its interesting to note that the inclination flips shown in Figures \[fig:quadoct\] and \[fig:Ex1\] are [*qualitatively*]{} different from the ordered, back and forth oscillation of the quadrupole-level of approximate evolution. The latter produces a simple, ordered oscillation of the inclination angle between $i_0$ and $180^\circ- i_0$, for the librating regime. However, in the presence of the octupole-level of approximation the system behaves similarly to the general flips discussed in @Naoz11, where the inclination oscillates for sometime at the prograde ($i<90^\circ$) regime, and then flips to retrograde configuration ($i>90^\circ$). The eccentricity, $e_2$, gives rise to an additional complication as in essence the outer test particle eccentricity can grow so much until the orbits will cross. We adopt the nominal stability criterion $$\epsilon = \frac{a_1}{a_2}\frac{e_2}{1-e_2^2} < 0.1$$ to guid us when the system leaves stability. We discuss this stability criterion in the context of N-body comparisons in appendix \[sec:Nbody\]. {width="\linewidth"} {width="\linewidth"} To explore the chaotic nature of the system and the different dynamical regimes we use surface of sections. The outer test particle approximation reduces the general hierarchical three body system from six degrees of freedom to four degrees of freedom. In addition, in the test particle limit, the inner orbit is stationary and reduces the system to two degrees of freedom. In this system $f$ and $\delta$ are the only conserved parameters and $\omega_2$ and $\Omega_2$ are the only coordinates that can change with time. For a two-degrees-of-freedom system, the surface of section projects a four-dimensional trajectory on a two-dimensional surface, where we select intersections of the trajectories on the surfaces when $\omega_2$ and $\Omega_2$ moves in the positive directions. For simplicity, we separate the two initial conditions into three characteristic parameters: $e_1$, the inner orbit eccentricity, which remains constant during the evolution, the energy, or initial value of the reduced Hamiltonian $f$, and $$\delta=\frac{m_1-m_2}{m_1+m_2}\frac{a_1}{a_2} e_1 \ .$$ Note that the energy depends on $\delta$ and $e_1$, so in fact, although we choose to characterize the surface of sections by three parameters, there are only two independent ones. In Figures \[fig:sur1\]-\[fig:sur3\] we consider the surface of sections for various values of the $f$, $e_1$ and $\delta$, in the $J_2-\omega_2$ plane (top row in each figure) and in $J_{2,z}-\Omega_2$ plane (bottom row in each figure). In both planes we identify the resonances at which the momenta and angles undergo bound oscillations. The trajectories in this region are quasi-periodic, and the system is in the libration mode. The circulation region represents trajectories where the angles are not constrained between two specific values. Both [ librating]{} and circulatory trajectories are mapped onto a one-dimensional manifold on the surface of section and they form lines on the section. However, chaotic trajectories are mapped onto a higher dimensional manifold and they are filling an area on the surface. We note that in some of the trajectories in the Figures, due to sampling limitation, seem as dashed lines, but they actually represent a one-dimensional manifold. In all of the maps we indicate the instability regime (light orange stripe) for which $\epsilon>0.1$. [We intersect the trajectories at $\Omega_2 = \pi/2$ to produce the surfaces in the $J_2-\omega_2$ plane, in order to capture the librating cases. The empty regions at large $J_2$ at the parameter spaces in the far left and right panels in Figure \[fig:sur2\] (i.e,. for the parameters: $e = 0.4$, $\delta = 0.02$ and $f = \pm 10$) and the far left panel in Figure \[fig:sur3\] ( for the parameters: $e = 0.9$, $\delta = 0.1$ and $f = -10$) correspond to regions with no physical solutions. The variabilities in $J_2$ are mostly small in the stable regime. We see that there are regular behavior, i.e., trajectories which [fill one-dimensional lines on the surface of section]{} in most of the stable region.]{} We find the emergence of chaos in [parts of the unstable zones, in particular when $J_2$ is low ($e_2$ is high).]{} Considering the $J_{2,z}-\Omega_2$ plane, [we intersect the trajectories at $\omega_2 = 0$.]{} The system exhibits a chaotic behavior across all parameter regime of $e_1,\delta$ and $f$. Most of the circulation region, associated with curve, non chaotic one-dimensional manifold, are typically associated with $|J_{2,z}|{\mathrel{\rlap{\lower4pt\hbox{\hskip1pt$\sim$}} \raise1pt\hbox{$>$}}}0.3$. [The outer orbits can flip ($J_{2, z}$ shifts signs) in most of the parameter space.]{} [Resonances can be easily identified in a few of the maps. Specifically, in the $J_2-\omega_2$ plots, the resonances can be found centered near $\bf \omega_2 = \pi$ (e.g., $e_1 = 0.4$, $\delta = 0.02$, $f = -10$, and $e_1 = 0.9$, $\delta = 0.1$, $f = -10$, [etc.]{}), [and $\omega_2 = \pi/2$ (e.g., $e_1 = 0.4$, $\delta = 0.02$, $f = 10$)]{}. The dynamics is quite complicated when $e_1$ is higher and when $\delta$ is larger, and higher order resonances (appearing as small liberating islands) emerge in the surface of section in the $J_2-\omega_2$ plane when $e_1 = 0.9$, $\delta = 0.1$ and $f=0$. Resonances can also be identified in the $J_{2,z}-\Omega_2$ plane, such as $e_1=0.9$, $\delta=0.1$, $f=10$. ]{} The Role of the General Relativity {#sec:GR} ================================== ![ We consider two cases, the evolution without GR (red lines) and the evolution with GR (blue lines). [**Top panel**]{}: We consider the following system: $m_1=1$ M$_\odot$, $m_2=1$ M$_j$, $a_1=0.5$ AU, $a_2=10$ AU, $e_1=0.4$ and $e_2=0.6$. We initialize the system with $\omega_2=\Omega_2=0^\circ$ and $i=85^\circ$. For this system we find that $t_{\rm GR,inner}\sim 5\times 10^6$ yr which is much shorter than the quadrupole timescale. [**Bottom panel**]{}: We consider the following system: $m_1=1$ M$_\odot$, $m_2=1$ M$_j$, $a_1=3$ AU, $a_2=40$ AU, $e_1=0.9$ and $e_2=0.65$. We initialize the system with $\omega_2=90^\circ$ and $\Omega_2=100^\circ$ and $i=20^\circ$. The GR precession timescale for this system is estimated as $t_{\rm GR,inner}\sim 4.4\times 10^8$ yr, which is longer than the quadrupole timescale. This is a typical situation to an individual debris disk particle or a icy body reservoir object (see section \[sec:Kuip\]). \[fig:GR\]](TestGRh2){width="\linewidth"} As noted previously in many studies, General Relativity (GR) precession tends to suppress the inner orbit eccentricity excitations associated with the [Eccentric Kozai-Lidov]{} (EKL) mechanism, and thus suppress the flips [e.g., @Naoz+12GR]. In our secular case, the inner orbit is massive and the outer orbit is a test particle, so practically the inner orbit does not feel the outer orbit gravitational interactions. However, the inner orbit can still precess due to GR with the nominal precession rate [e.g., @Naoz+12GR]: $$\label{eq:GR1} \frac{d\omega_1}{dt} \bigg |_{\rm GR,inner} = \frac{3 k^{3} (m_1 + m_2)^{3/2} }{a_1^{5/2} c^2 (1 - e_1^2)} \ ,$$ [where $k^2$ is the gravitational constant and $c$ is the speed of light. ]{} However, in our frame of reference, where the inner orbit carries all the angular momentum, we are basically working in the rotating frame of the inner orbit. Therefore, since we set $\omega_1=-\pi-\Omega_2$, (see Appendix \[App:Ham\]), GR precession of $\omega_1$ translates to a precession of $\Omega_2$. Thus, using our coordinate transformation we find: $$\label{eq:GR2} \frac{d\Omega_2}{dt} \bigg |_{\rm GR,\Omega_2} = -\frac{3 k^{3} (m_1 + m_2)^{3/2} }{a_1^{5/2} c^2 (1 - e_1^2)} \ ,$$ which can suppress the inclination oscillations. The timescale associated with that precession is the nominal GR one: $$\label{eq:tGR} t_{\rm GR,\Omega_2} \sim 2\pi \frac{a_1^{5/2} c^2 (1-e_1^2) }{ 3 k^3 (m_1+m_2)^{3/2}} \ .$$ In Figure \[fig:GR\] we consider two examples, one for which $t_{\rm GR,\Omega_2}\sim 5\times 10^6$ yr is smaller than $t_{\rm quad}\sim 6\times 10^6$ yr (top panel), and the other of which $t_{\rm GR,\Omega_2}\sim 4.4\times 10^8$ yr is a bit longer than the corresponding $t_{\rm quad}\sim 2\times 10^7$ yr (bottom panel). Both of these examples had a Sun size star and a Jupiter planet, orbited by a far away test particle. The Jupiter has a non-negligible eccentricity, that perhaps can be a result of either a scattering event or a high eccentricity migration. In the top panel the Jupiter was set at $a_1=0.5$ AU with $e_1=0.4$ and the test particle was set at $a_2=10$ AU and $e_2=0.6$. The system in the absence of GR was in libration mode and as noted before exhibited a chaotic nature. However, the GR precession bound the system into a circulatory regime, and suppressed the flips. In the bottom panel the Jupiter was set at $a_1=3$ AU and $e_1=0.9$, while the test particle was set at $a_2=40$ AU with $e_2=0.65$. Note that in this latter case, although the Jupiter is rather far from the host star and the GR precession timescale is longer than the secular precession timescale, the GR changes the dynamics. Specifically, before the inclusion of GR precession, the system was in a libration mode and seemed quasi-periodic, however, after the inclusion of GR the system exhibits both libration and circulation, and the emergence of chaotic behavior seems to take place. The dramatic change in dynamical behavior with the inclusion of GR precession, even if takes place on [*longer*]{} timescales than the secular timescales, was noted previously in the general and inner test particle case in @Naoz+12GR. We note that in all of our calculation below we also take into account the outer orbit GR precession [e.g., @Naoz+12GR]: $$\label{eq:GR2} \frac{d\omega_2}{dt} \bigg |_{\rm GR,outer} = \frac{3 k^{3} (m_1 + m_2)^{3/2} }{a_2^{5/2} c^2 (1 - e_2^2)} \ ,$$ which typically takes place on much larger timescales. A study case application: individual Debris disk particles {#sec:Kuip} =========================================================== ![. We consider the time evolution of a test particle located at $55$ AU from a $~$ M$_\odot$ star due to the gravitational perturbations from an eccentric Jupiter $a_1=5$ AU and $e_1=0.85$. We initialized the system with $e_2=0.5$, $\Omega_2=90^\circ$, $\omega_2=0^\circ$ and $i=20^\circ$. We consider from top to bottom the inclination, eccentricity and $\Omega_2$. The transition between librating and circulating can clearly be seen in the bottom panel. When the angle is in circulation mode, it increases in value as a function of time. []{data-label="fig:Kuip1"}](Kuiper3panles){width="\linewidth"} Debris disks mark the late end stages of planet formation, and are made of the leftover material of rocks and ices. The gravitational interactions between these particles and interior or exterior companion can leave a distinct imprint on the morphology of the disk and can cause dust production [e.g., @Rodigas2014; @Matthews2014; @Nesvold2015; @Lee2016; @Nesvold2016]. Many of these studies typically focus on few million years of integration to allow for comparison of observations which usually can detect young systems. Here we allow for longer integration timescales and investigate the evolution of a test particle under the influence of an eccentric Jupiter. {width="\linewidth"} In Figure \[fig:Kuip1\] we show an example system, where we consider an eccentric Jupiter at $5$ AU, with $0.85$ eccentricity orbiting a one solar mass star. The test particle is located at $55$ AU and initialized with an eccentricity of $0.5$. We integrate the octupole level equations of motion in the presence of GR precession for both the inner and outer orbits. As in the example depicted in Figure \[fig:GR\], which also considered an icy body or comet reservoirs analogs, the orbit switches between libration and circulation as can be seen in the bottom panel. We consider the effect of the planet’s eccentricity, $e_1$, and the test particle inclination by surveying the parameter space of $e_1$ and initial inclination for a given system, where we set $a_2=55$ AU, with $e_2=0.5$, and set the system initially with $\Omega_2=90^\circ$ and $\omega_2=0^\circ$. Many giant exoplanets have high eccentricities, specifically for giant planets ($m\sin i>0.1$ M$_J$, with separation $>0.05$ AU) the average eccentricity is $\sim 0.2$, and a maximum value of $0.97$[^1]. Thus, an eccentric Jupiter doesn’t seem like an unlikely configuration for a planetary system. In the example depicted in Figure \[fig:Kuip1\] the time evolution of the test particle’s inclination that starts with a moderate eccentricity oscillates between extreme values $20^\circ \sim 160^\circ$ without increasing its eccentricity. ![ The mutual inclination, $i$, as a function of the semi-major axis, $a_2$, after $4$ Gyrs of integration. The color code shows the longitude of ascending nodes $|\Omega_2-90^\circ|$ at that time (this presentation emphasizes the symmetry in the system). The right inset shows the histogram of the final eccentricity of the disk, while the left inset shows the histogram of the final mutual inclination of the disk. As a proof of concept we consider a narrow debris disk located between $55-65$ AU, with an interior eccentric Jupiter ($e_1=0.85$) at $5$ AU around a solar mass star. The system was initially set with a mutual inclination of $20^\circ$, $e_2=0.3$ and $\Omega_2$ and $\omega_2$ are chosen from a random uniform distribution between $0-360^\circ$. [The results depicted here achieved by integrating over the equations of motion, Eqs. (\[eq:EOM1\])-(\[eq:EOM4\]). Note that GR effects are included here as well.]{} []{data-label="fig:Puff"}](PuffedDiskNew){width="1.15\linewidth"} The system depicted in Figure \[fig:Kuip1\] is, of course, just one example for a particular choice of the orbital parameters. To study the effects of planet’s eccentricity and initial inclination we have systematically explored the $e_1$-$i$ parameter space in Figure \[fig:TwoPanle\_e5\]. We choose a Jupiter like system ($a_1=5$ AU) and for a range of eccentricities, with an outer orbit at $55$ AU. The test particle orbit was initialized with $e_2=0.5$, $\Omega_2=90^\circ$, $\omega_2=0^\circ$ and a range of inclinations. We show the maximum inclination reached during the evolution as a function of the initial inclination and eccentricity in Figure \[fig:TwoPanle\_e5\]. In the left panel we depict the initial inclination vs the initial Jupiter’s eccentricity, where the color code marks the maximum inclination reached. The solid line in the left panel follows Equation (\[eq:imaxmin\]), which is consistent that the resonance associated with the quadrupole-level of approximation is indeed the main driver for the dynamical evolution of the system. In the figure we depicted the initial inclination regime to the prograde case ($i_{\rm initial}\leq 90^\circ$) to avoid clutter. However, we show in the right panel the maximum inclination reached during the evolution as a function of the initial inclination going all the way to $180^\circ$ this time. As a proof of concept we depict in Figure \[fig:Puff\] the behavior of a narrow debris disk after $4$ Gyr of evolution. This inclination represents the instantaneous inclination at $4$ Gyr of evolution. The system, of course, continue to oscillate, and the disk of particles will remain puffed. The inclination and eccentricity of the system at this snapshot are qualitatively different from the initial conditions assumed. This hypothetical system shows the orbital configuration of a disk located between $55-65$ AU, with an interior eccentric Jupiter ($e_1=0.85$) at $5$ AU around a solar mass star. The system was set initially with a mutual inclination of $20^\circ$, $e_2=0.3$ and $\Omega_2$ and $\omega_2$ are chosen from a random uniform distribution between $0-360^\circ$. While the particles in the disk have eccentric values, the disk does not appear as a coherent eccentric ring as the values of $\Omega_2$ and $\omega_2$ are random. At the end of the integration, the particles in the disk became slightly more eccentric (with an average eccentricity of $\sim 0.34$) and there is a clear trend of $\Omega_2$ as a function of inclination. The particles with inclination above the initial $20^\circ$ have a value of $\Omega_{2}$ close to zero (or with the symmetric value $180^\circ$), while the particles with inclination below $20^\circ$ have $\Omega_2$ values closer to $90^\circ$. The behavior is singular to $\Omega_2$ and is not manifest itself in $\omega_2$ (does not depicted here to avoid clutter). We note that some retrograde particles were formed as well, in line with @Zanardi+17 numerical results of an orbital flip. These behaviors are easily understandable from the surface of section maps depicted above. Its important to note that during the evolution of the system depicted in Figure \[fig:Puff\], the maximum $\epsilon$ achieved was $0.0625$. The average value of $\epsilon$ which corresponds to the maximum $e_2$ achieved during the evolution was $\sim 0.4$. Thus, the system is kept stable during this evolution, the secular approximation holds, and we do not expect any scattering event. We also note that we have compared a debris disk particles secular and N-body evolution and found a qualitative agreement, which is similar in behavior to Figure \[fig:NB1\] left panel, and this is not shown here to avoid clutter. Discussion {#sec:dis} ========== We have studied the secular evolution of an outer test particle hierarchical system. We presented the three body, outer test body Hamiltonian up to the octupole level of approximation in the power series of the semi-major axis ratio. We showed that in the quadrupole-level of approximation, $(a_1/a_2)^2$, the system has two distinct behaviors, librating and circulating (see Figure \[fig:quad\]), where the librating mode gives the nominal precession of the nodes, results for which the inclination oscillates between the $i_{90}$ inclination (the inclination for which $\Omega_2=90^\circ$ and $180^\circ-i_{90}$). Furthermore, the bound values of the liberating mode have a simple analytical expression, Eq. (\[eq:Omegamin\]). We also found the minimum and maximum inclination that the system can reach in the circulating mode (see Equations (\[eq:imaxmin\]) and (\[eq:imim\])). These conditions are sensitive to the initial inner orbit’s eccentricity, and are nicely reproduced in numerical testing here (see Figure \[fig:TwoPanle\_e5\]) and in @Zanardi+17 numerical experiments (see their figure 14). We also estimated the timescale for oscillations for the two modes (see Section \[sec:Time\]). We then showed that introducing the octupole level of approximation allows for transition between the two libration and circulation modes (see Figures \[fig:quadoct\] and \[fig:Ex1\]). This yields that the overall dynamics of the system is similar to the behavior of the general flip behavior in the eccentric Kozai-Lidov mechanism. [In particular, the dynamics is quite chaotic for parameter regions with high $e_2$ and perpendicular mutual inclinations (when $J_2$ is low and when $J_{2,z}$ is near zero), as shown in the surface of sections Figures \[fig:sur1\]-\[fig:sur3\].]{} General relativity can play an important role in suppressing or exciting the eccentricities in the hierarchical three body problem [e.g., @Naoz+12GR]. We find here similar behavior. Specifically, the [*inclination*]{} excitation will be suppressed for systems with GR precession faster than the quadrupole precession. However, when GR precession takes place on similar (or even somewhat larger) timescales to that of the quadrupole precession, the additional precession can produce inclination excitations, in a non-regular manner (see Figure \[fig:GR\]). The dynamics of these type of systems can have a wide range of applications, from stars around supermassive black hole binaries to the evolution of individual debris disk particles. We have chosen the latter as an example and presented a typical example of the evolution of a test particle due to the gravitational perturbations from an eccentric Jupiter (see Figure \[fig:Kuip1\]). We systematically varied the Jupiter’s eccentricity and the outer orbit’s inclination, where we found an agreement between the analytical relation for crossing the $90^\circ$ threshold, and the numerical tests. This also suggests that eccentric planet can pump up the inclination of icy body or comet reservoir analogs (Figure \[fig:TwoPanle\_e5\]). This was further supported by considering the evolution of an initially narrow, thin disk of test particles exterior to an eccentric planet, with initial mutual inclination of $20^\circ$. The disk became puffed with some particles on a retrograde orbits (see Figure \[fig:Puff\]). A detailed study of the effects of eccentric planets on exterior test particle is presented in @Zanardi+17. This mechanism will also be important for example to circumbinary planetary systems [[considered first by @Ziglin75]]{} and the stars near mergers of black hole binaries, but detailed studies is beyond the scope of this paper. We thank the referee for a quick and detailed report, and especially for his/her inquiry about the surface of sections. We also thank Vladislav Sidorenko for pointing out some missing references. SN acknowledges partial support from a Sloan Foundation Fellowship. GL is supported in part by the Harvard William F. Milton Award. MZ, GdE, and RPD acknowledge the financial support given by IALP, CONICET, and Agencia de Promoción Científica, through the PIP 0436/13 and PICT 2014-1292. A. Orbital Parameters and the Scaled Time {#App:Ham} ========================================== One might have expected that $\omega_1 = {\rm const}$ when the outer particle is massless. But in truth $\Omega_1$ is undefined because the reference plane is aligned with the inner orbit. Therefore, the inner planet must only have $\omega_1+\Omega_1 = {\rm const}$, and we may choose without loss of generality the constant to equal zero. Hence, elimination of the nodes (i.e., $\Omega_1-\Omega_2=\pi$) implies $\omega_1=-\pi-\Omega_2$. Similarly to the treatment done in @LN we have rescaled the momenta by an arbitrary constant to achieve the specific angular momentum. The Hamiltonian is rescaled by the same constant, and we find that the rescaled time is: $$\label{eq:time} \tau = \frac{t}{16} \frac{m_1 m_2}{(m_1+m_2)^2} \sqrt{\frac{G_N (m_1+m_2)}{a_2^3} } \left( \frac{a_1}{a_2} \right)^2 = \frac{ t}{16} \frac{m_1 m_2}{(m_1+m_2)^2} \frac{2\pi}{P_2} \left( \frac{a_1}{a_2} \right)^2 \ ,$$ where $t$ is the true time. The numerical factor $16$ comes by taking $m_3\to 0$ in the general Hamiltonian [see @Naoz16 for the general form of the hierarchical three body double averaged hamiltonian]. There is a choice to be made, to either scale the Hamiltonian by this numerical factor or $\tau$. Here, we choose to absorb this number in $\tau$ to be consistent with the inner test particle Hamiltonian. B. Comparison with N-Body {#sec:Nbody} =========================   ![ We show the inclination, eccentricity. Red lines correspond to the secular calculation (up to the octupole-level of of approximation) and blue lines corresponds to the N-body calculation. [**Left side:**]{} we consider the following system: $m_1=1$ M$_\odot$ $m_2=1$ M$_J$, $a_1=1$ AU, $a_2=15$ AU, $e_1=0.8$, $e_2=0.8$, $\omega_2=0^\circ$, $\Omega_2=90^\circ$, and $i=16^\circ$. These parameters imply an initial $\epsilon$ of $0.148$. [**Right side:**]{} we consider the following system: $m_1=1$ M$_\odot$ $m_2=1$ M$_J$, $a_1=3$ AU, $a_2=40$ AU, $e_1=0.9$, $e_2=0.65$, $\omega_2=90^\circ$, $\Omega_2=100^\circ$, and $i=20^\circ$. The latter systems initialize with $\epsilon=0.084$ however, as time goes by the eccentricity grows and the approximation breaks. []{data-label="fig:NB1F"}](NBodyFail){width="0.5\linewidth"} In this section, we compare the secular approximation at the octupole level in the test particle limit with the N-body simulation, using Mercury code [@Mercury]. In this comparison we did not include GR. Good agreements can be reached when the apocenter distance of the inner binary is much smaller than the pericenter distance of the outer binary. In particular, we include an illustrative example here in Figure \[fig:NB1\], where we consider an eccentric ($e=0.9$ Jupiter at $3$ AU around a solar like object. The test particle is set at $40$ AU with $e_2=0.65$. The system is set initially with $i=20^\circ$, $\omega_2=90^\circ$ and $\Omega_2=100^\circ$. As shown in Figure \[fig:NB1\], the secular approximation (shown as red lines) agrees [*qualitatively*]{} well with the N-body results for the eccentricity and inclination oscillations. This is likely due to the double averaging process, but nonetheless, the maximum and minimum of the orbital parameters are conserved in both N-body and secular calculations. It is interesting to note that similarly to [@LN] the approximation holds as long as $\epsilon<0.1$. However, unlike the inner-test particle case, $e_2$ can change and increase during the evolution which may break the validity of the approximation during the evolution, this is shown in Figures \[fig:NB1F\]. [^1]: Taken from The Exoplanet Orbit Database [@Wright+11].

ArXiv

--- abstract: | We calculate the effective action for Quantum Electrodynamics (QED) in $D=2,3 $ dimensions at the quadratic approximation in the gauge fields. We analyse the analytic structure of the corresponding nonlocal boson propagators nonperturbatively in $k/m$. In two dimensions for any nonzero fermion mass , we end up with one massless pole for the gauge boson . We also calculate in $D=2$ the effective potential between two static charges separated by a distance $\,L\, $ and find it to be a linearly increasing function of $\,L\,$ in agreement with the bosonized theory (massive Sine-Gordon model). In three dimensions we find nonperturbatively in $\, k/m \, $ one massive pole in the effective bosonic action leading to screening. Fitting the numerical results we derive a simple expression for the functional dependence of the boson mass upon the dimensionless parameter $\,e^{2}/m\,$ .\ [*PACS-No.:*]{} 11.15.Bt , 11.15.-q author: - | D. Dalmazi[^1]\ [*Departament of Physics and Astronomy - SUNY*]{}\ [*Stony Brook, NY 11794 , USA* ]{}\ A. de Souza Dutra[^2] and Marcelo Hott[^3]\ [*UNESP - Campus de Guaratinguetá - DFQ*]{}\ [*Av. Dr. Ariberto Pereira da Cunha, 333*]{}\ [*CEP 12500-000 - Guaratinguetá - SP - Brazil.* ]{} title: 'Quadratic Effective Action for QED in $D=2,3$ Dimensions' --- Introduction ============ The mapping of fermionic theories into bosonic ones is a very powerful technique used to understand nonperturbative aspects of quantum field theories. This so called bosonization is exact in two dimensional theories, and it has been widely employed in this dimension \[1-3\]. In the last few years, many papers have been devoted to the study of what has been called bosonization in three dimensions \[4-8\], and even in four dimensions [@marino1]. This kind of path-integral bosonization consists of obtaining the effective action by integrating out the fermion degrees of freedom and then studying the physical properties of the resulting effective theory. In most of the works along this line one makes use of the derivative expansion [@ait1; @shif] and derives weak bosonization rules for the system [@schaposnik; @banerjee; @elcio] as well as the particle content and their masses [@schaposnik]. On the other hand in [@barci] an exact, in $k/m$, action at one-loop level is used to show that the bosonization of [@marino] and [@schaposnik2], is recovered in the large and small momentum limits respectively. In addition , the authors of [@elcio] have studied the assymptotic properties of the bosonic effective action associated with QED in three dimensions, showing the screening property of the effective potential between two static charges. In this work we apply the effective action approach of bosonization on QED in two and three dimensions. The usefulness of the two dimensional case lies on the fact that $QED_{2}$ can be bosonized via the massive Sine-Gordon model which exhibits confinement [@livroElcio; @Jackiw; @Coleman]. Thus both approaches of bosonization can be compared. Here we verify that, indeed the confining behavior also appears in the bosonization *via* effective action. It is remarkable that in two dimensions at the quadratic approximation in the gauge fields but without any expansion in $k/m$, the massive pole of the Schwinger model disappears, being replaced by a massless pole, which is in agreement with what has been observed in [@gross] , but it differs from the result obtained through perturbative ($\,m/e\,$) calculation of [@adam]. In three dimensions it is shown that there is a massive excitation which depends on the dimensionless parameter $\frac{16\pi \,m}{e^{2}}$ . We have found a simple approximated expression for this function. This in fact generalizes the calculations of [@schaposnik], which were obtained at leading order of the derivative expansion, and that of [@cesar99] carried out at a higher order in $k/m$, which in its turn is related to consistent higher derivative actions [@cesar97; @jackiw99]. In both cases we consider the one-loop effective action up to second-order in the coupling constant $e$. We may write it as $$S^{(2)}_{eff}=- \frac{1}{2} \int \; \frac{d^{D}k}{(2 \pi)^D} \; {\tilde{A}}_{\mu}(-k) \left[g^{\mu \nu} k^2 - k^{\mu} k^{\nu} - \Pi^{\mu \nu}(k) \right] {\tilde{A}}_{\nu}(k),$$ where ${\tilde{A}}_{\mu}(k)$ is the Fourier transformation of $A_{\mu}(x)$ and $$\Pi^{\mu \nu}(k) = i e^2 \int \; \frac{d^{D}p}{(2 \pi)^D} \; tr \left[ \frac{1}{\ps - m + i\epsilon} \gamma^{\mu} \frac{1}{(p\!\!\!/ + k\!\!\!/) - m + i\epsilon} \gamma^{\nu} \right]$$ is the polarization tensor obtained after integrating out the fermion fields. The space-time dimension is represented by $D$ ($\, D=2,3\,$). It is useful to expand the polarization tensor in powers of $\frac{k}{m}$ which corresponds to the derivative expansion of the effective action [@das1]. Truncating the expansion at some power of $\, k/m\,$ not only yields a local effective action but also allows one to analyse the rôle of some specific term as it is the case of the leading odd-parity term in $QED_{3}$ [@das2]. It is also important to study the phenomenological aspects of a low-energy effective action [@ait2]. The order at which the series is truncated depends on the range of energy one is interested in. Recent studies of the contributions of higher-order derivative terms in a low-energy effective gauge theory revealed the possibility of the appearence of non-physical excitations. Here we overcome this difficulty by analyzing directly the poles of the full propagator at one-loop level up to second order in the coupling constant. Effective potential in $QED_2$ ============================== In $QED_2$ the polarization tensor will be given by $$\Pi^{\mu \nu}(k) = \Pi(k^2) \left( g^{\mu \nu} -\frac{k^{\mu} k^{\nu}}{k^2} \right) ,$$ where using , e. g. , dimensional regularisation we get $$\Pi(k^2) = \frac{e^2}{\pi} \left[ 1 + \frac{1}{2} \frac{4{m^2}/k^2}{(1 - {4{m^2}/k^2})^{1/2}} \ln{\frac{(1- {4{m^2}/k^2})^{1/2} +1}{(1 - {4{m^2}/k^2})^{1/2}-1}} \right] .$$ We are interested in the interacting potential between two static charges $\, Q \, $ and $\, - Q \, $ located at $x=L/2\,$ and $\, x=-L/2 \, $. Solving[^4] the equation of motion derived from the effective action we obtain the potential produced by the positive charge $Q$ : $$A_{\mu}(x) = \int \; \frac{d^{2}k}{(2 \pi)^2}\int d^{2}x^{\prime} D_{\mu \nu}(k) e^{i k (x - x^{\prime})} J^{\nu}(x^{\prime}) ,$$ where $$J^{\nu} (x^{\prime}) = Q \delta({x_{1}}^{\prime} - {\frac{L }{2}}) \delta^{\nu 0}$$ \[current\] is the conserved current and $D_{\mu \nu}(k)$ is the “photon" propagator whose longitudinal term is given by $${D^\parallel}_{\mu \nu}(k) = \frac{1}{k^2 - \Pi(k^2)} g_{\mu \nu}.$$ Due to the specific form of the external current the only contribution to the potential will be its time-component . The corresponding integral can be easily performed on the complex plane for arbitrary masses . The resulting static potential at $x_1= - {\frac{L }{2}}$ grows linearly with the distance between the charges, namely $$A_{0}(x=-\frac{L}{2}) = \frac{Q}{2 \left[1 + \frac{2}{3 \pi} (\frac{e}{2m} )^{2} \right]} L.$$ Then we conclude that, at the quadratic approximation used here, $QED_2$ results in a [*confining*]{} potential. Because of the quadratic approximation in the gauge fields we may say that the result obtained here, [*i.e.*]{}, a linearly growing inter-fermion effective potential is due to a zero mass pole for the gauge potential. One also reaches a similar result in the usual bosonization approach when the quadratic approximation in the boson field of the massive Sine-Gordon model is used [@gross]. In other words we can say that, at the quadratic approximation, turning on a mass for the fermion field will prevent the mass generation for the gauge boson and the classical result, that is a massless pole, prevails. It is worth mentioning that , since we are only interested in the real contribution to the effective action , we have dropped in (4) the imaginary part of the polarization tensor which appears beyond the pair creation threshold $k^2> 4m^2 $. More specifically , expression (4) is correct for $k^2<0$ which is the relevant region for the calculation of the effective potential of static charges due to the factor $\delta (k_0)$ which comes from the time integral in (5). However, for the analysis of the particle content of (1) we have also calculated the polarization tensor in the region $0<k^2<4m^2$ where we found no poles except in the limit $k^2\to 0^+$ where we found an agreement with the limit $k^2\to 0^-$ of (4). Finally, it is important to notice that although we can recover the Schwinger model effective action from the $\, m\to 0 \, $ limit of (4) its static potential does not correspond to the massless limit of (8) since the integral and the limit do not commute with each other. In the Schwinger model (massless $QED_{2}$) a mass for the gauge boson is generated by the gauge anomaly and as a consequence one obtains a *screening* potential between two static charges [@swieca; @elcio2]. Massive pole in $QED_3$ ======================= Here we are not concerned with the explicit expression for the interaction potential but with the behavior of the mass generated dynamically as a function of the coupling constant and the fermion mass. In this case the polarization tensor is given by $$\Pi^{\mu \nu}(k) = \Pi_{1}(k^2) i \epsilon^{\mu \nu \rho} k_{\rho} + \Pi_{2}(k^2) \left( k^{2} g^{\mu \nu} -k^{\mu} k^{\nu} \right) ,$$ where $$\Pi_{1}(k^2) = - \frac{e^2}{8 \pi} \frac{1}{\sqrt{k^{2}/ 4{m^2}}} \ln{\frac{1+ \sqrt{k^{2}/4{m^2}}}{1 - \sqrt{k^{2}/ 4{m^2}}}}, \label{pi1}$$ and $$\Pi_{2}(k^2) = \frac{e^2}{16 \pi m} \frac{1}{k^{2}/4{m^2}} \left[ 1 - \frac{1}{2} \left( \frac{1 + {k^{2}/4{m^2}}}{\sqrt{k^{2}/4{m^2}}} \right) \ln{\frac{1+ \sqrt{k^{2}/4{m^2}}}{1 - \sqrt{k^{2}/ 4{m^2}}}} \right]. \label{pi2}$$ Equations (\[pi1\]) and (\[pi2\]) were obtained by using dimensional regularization method. Equation (\[pi1\]) is regularization dependent and our result corresponds to an equal number of Pauli-Villars regulators with positive and negative masses. Like the two dimensional case , since we are only interested in the real contribution to the effective action, we have dropped the imaginary part of the polarization tensor which appears beyond the pair creation threshold ( $k^2>4m^2 $ ). More precisely we have only given the polarization tensor (see (\[pi1\]) and (\[pi2\])) in the region $0<k^2<4m^2$ where we found a pole in the photon propagator. For $k^2<0$ the correct result correspond to replace $\frac{1}{\sqrt{u}}log((1+\sqrt{u})/(1-\sqrt{u}))$ by $\frac{-2}{\sqrt{-u}} arctan\frac{1}{\sqrt{-u}} \, $ in (\[pi1\]) and (\[pi2\]) where $u=\frac{k^2}{4m^2}$. The situation is similar to QED$_4$ , see , e.g., \[26\]. We have explicitly verified that no tachyonic poles appear without any approximation on $\frac{k^2}{4m^2}$ . Back to the region $0< k^2<4m^2 $ one can check that the “photon" propagator $${D^\parallel}_{\mu \nu}(k) = - \frac{1}{\tilde{\Pi}_2 \left[ k^2 - (\frac{\Pi_1}{\tilde{\Pi}_2})^{2} \right]} g_{\mu \nu},$$ where $$\tilde{\Pi}_{2}(k) = 1 - \Pi_{2}(k),$$ develops a massive pole $k^{2}=M^{2}$. The gauge boson mass $M$ depends on the coupling constant and fermion mass. We have carried out a numerical analysis for the behavior of this mass generated dynamically and found a very simple function , namely $$M=\frac{2\,m}{c_{1}\,+\,c_{2}\,a\,+\,c_{3}\,a^{2}},$$ where $$a=16\pi \frac{m}{e^{2}}.$$ From the numerical results for the massive pole at small $a$ we get $c_{1}=1.953331381$ . On the other hand , from the large $a$ region we have $c_{3}=1$. Finally the constant $c_2$ was adjusted by choosing the best fitting for a curve with about eleven thousand numerically calculated points, and it was found to be $c_{2}=2.253$. The maximum error of the fitted function is less than $3.8\%$, with the statistical controlling parameter $\chi ^{2}=5.5\times 10^{-6}$. In the figure \[fig:figure1\] we present the curve of the massive pole as a function of the dimensionless parameter $a$. Note that it is impossible to distinguish the exact from the adjusted curve in figure 1. For this reason we present in the figure \[fig:figure2\] the dependence of the percentual deviation of the fitted curve from the numerically obtained masses. In reference [@elcio] the derivative expansion up to second-order in $\frac{k }{m}$ is used to compute the effective potential which was also found to be of the *screening* type. We have taken into account the whole expression for the polarization tensor and found only one massive pole in the whole range of the parameter $a$. Since the pole can never be found at the origin we conclude that, nonperturbatively in $k/m\,$ the screening effect prevails in agreement with the numerical analysis carried out in reference [@elcio3] for the static potential. The truncation based on the derivative expansion only amounts to a displacement of the massive pole from its nonperturbative ( in $k/m$ ) value. Acknowledgements ================ We would like to thank professors E. Abdalla , C. de Calan and J. A. Mignaco for valuable discussions. This work was partially supported by CNPq and FAPESP. [99]{} E. Abdalla, M. C. Abdalla and K. D. Rothe, “Non-perturbative mehods in two dimensional quantum field theory”. World Scientific 1991 - Singapore. S. Coleman, R. Jackiw and L. Susskind, Ann. Phys. [**93**]{} (1975) 267. S. Coleman, Ann. Phys. [**101**]{} (1976) 239. E. Fradkin and F. A. Schaposnik, Phys. Lett. B **338** (1994) 253. G. Rossini and F. A. Schaposnik, Phys. Lett. B **338** (1994) 465. D. G. Barci, C.D. Fosco and L. E. Oxman, Phys. Lett. B **375** (1996) 367. N. Banerjee, R. Banerjee and S. Gosh, Nucl. Phys. B **481** (1996) 421. J. C. Le Guillou, C. Núñez and F. A. Schaposnik, Annals of Physics **251** (1996) 426. N. Bralic, E. Fradkin, V. Manias and F. A. Schaposnik, Nucl. Phys. B **446** (1995) 144. R. Banerjee and E. C. Marino, Phys. Rev. D **56** (1997) 3763. C.M. Fraser, Z. Phys. C **28** (1985) 101. V. A. Novikov, M. A. Shifman, A. I. Vainshtein and V. I. Zakharov, Sov. J. Nucl. Phys. **39** (1984) 77. E. Abdalla and R. Banerjee, Phys. Rev. Lett (1998) 238. E. C. Marino, Phys. Lett. B **263** (1991) 63. F. A. Schaposnik, Phys. Lett. B **356** (1995) 39. D. J. Gross, I. R. Klebanov, A. V. Matytsin and A. V. Smilga, Nucl. Phys. B **461** (1996) 109. C. Adam, Phys. Lett. B **382** (1996) 383. A. de Souza Dutra and C. P. Natividade, Mod. Phys. Lett. A **14** (1999) 307. Phys. Rev. D (1999) to appear. A. de Souza Dutra and C. P. Natividade, Mod. Phys. Lett. A **11** (1996) 775. S. Deser and R. Jackiw, Phys. Lett B **451** (1999) 73. A. Das and A. Karev, Phys. Rev. D **36** (1987) 623; *ibid* 2591. K. S. Babu, A. Das and P. Panigrahi, Phys. Rev. D **36** (1987) 2725. I. J. R. Aitchison and C. M. Fraser, Phys. Lett. B **146** (1984) 63; Phys. Rev. D **32** (1985) 190. H. J. Rothe, K. D. Rothe and J. A. Swieca, Phys. Rev. D **19** (1979) 3020. E. Abdalla, R. Mohayee and A. Zadra, Int. J. Mod. Phys. A **12** (1997) 4539. E. Abdalla, R. Banerjee and C. Molina, hep-th/9808003. N. N. Bogoliubov and D. V. Shirkov, “Introduction to the theory of quantized fields". Interscience Publishers 1959. New York. [^1]: [email protected] (On leave from UNESP- Guaratinguetá - Brazil ). [^2]: [email protected] [^3]: [email protected] [^4]: We have not taken into account the general solution of the homogeneous differential equation ( without sources ) which amounts to neglect the $\theta$-vacuum ($\, \theta =0 \, $)

ArXiv

--- abstract: 'Modern latency-critical online services often rely on composing results from a large number of server components. Hence the tail latency (e.g. the 99th percentile of response time), rather than the average, of these components determines the overall service performance. When hosted on a cloud environment, the components of a service typically co-locate with short batch jobs to increase machine utilizations, and share and contend resources such as caches and I/O bandwidths with them. The highly dynamic nature of batch jobs in terms of their workload types and input sizes causes continuously changing performance interference to individual components, hence leading to their latency variability and high tail latency. However, existing techniques either ignore such fine-grained component latency variability when managing service performance, or rely on executing redundant requests to reduce the tail latency, which adversely deteriorate the service performance when load gets heavier. In this paper, we propose PCS, a predictive and component-level scheduling framework to reduce tail latency for large-scale, parallel online services. It uses an analytical performance model to simultaneously predict the component latency and the overall service performance on different nodes. Based on the predicted performance, the scheduler identifies straggling components and conducts near-optimal component-node allocations to adapt to the changing performance interferences from batch jobs. We demonstrate that, using realistic workloads, the proposed scheduler reduces the component tail latency by an average of 67.05% and the average overall service latency by 64.16% compared with the state-of-the-art techniques on reducing tail latency.' author: - - bibliography: - 'references.bib' title: 'PCS: Predictive Component-level Scheduling for Reducing Tail Latency in Cloud Online Services' --- cloud online services; component latency variability; tail latency; predictive scheduler; Introduction ============ Providing fluid responsiveness to user requests is essential for online services: their potential profits are proportional to service latency (i.e. request response time including both the request queueing delay and the time of being processed) [@jalaparti2013speeding; @tailatScale]. In large online services such as search engines, e-commerce sites, and social networks, the processing of incoming requests consists of several sequential stages, where each stage composes responses parallelized across hundreds or thousands of server *components*. Hence the tail (e.g. the 99th percentile) of these components’ latencies, rather than the average, determines the overall service performance [@tailatScale; @kapoor2012chronos]. For example, Figure \[Fig: SearchEngine\] shows an example Nutch search engine [@nutchsearch] with three stages. Suppose that at stage 2, the request processing is parallelized into 100 components, in which 99 components can respond in 10ms but only one component gets a slow response of 1 second, the overall service performance is deteriorated by this straggling component and hence providing slow responsiveness of 1 second. ![An example of Nutch search engine[]{data-label="Fig: SearchEngine"}](SearchEngine.pdf "fig:")\ In modern cloud data centers and warehouse-scale computers, it is critical to improve machine utilizations by co-locating long-running online services and offline batch jobs (e.g. Hadoop [@HadoopWebsite] and Spark [@SparkWebsite] analytics jobs) on the same node (physical machine), while still keeping the overall latency of online services at a satisfactory level [@reiss2012heterogeneity; @yang2013bubble]. Although the components of a service are typically hosted on dedicated environments such as Xen virtual machines (VMs) or LinuX Containers (LXCs), these components still share and contend resources such as processing units, caches and I/O bandwidths with their co-running batch jobs on the same node, hence inevitably suffer from performance interference. Workload traces from Google [@reiss2012heterogeneity] and Facebook [@chen2012interactive] show that small batch jobs form a majority (over 90%) of all jobs in their data center workloads. For example, approximately 50% of Google jobs complete in 10 minutes and 94% of them complete within 3 hours. These short-term batch jobs have various workload types (e.g. CPU and I/O intensive workloads) and input data sizes (e.g. ranging from KB to GB), thus causing continuously changing performance interferences to their co-located components. This results in the **component latency variability**, which can be explained from two aspects: (a) each component’s latency (performance) varies over time, and (b) components hosted on different nodes have different changes in their latencies, hence causing high tail latency in individual components of the service. Many existing techniques have been developed to guarantee the performance of latency-critical services by mitigating the performance interference due to resource sharing and contention [@kasture2014ubik; @xu2013bobtail; @ahn2012dynamic; @xu2010mitigating]. However, these techniques only manage service performance at the coarse granularity of the entire application, ignoring fine-grained component latency variability that may come to dominate service performance at large scale. Moreover, state-of-the-art techniques reduce tail latency via request redundancy. They either create replicas for all the requests [@vulimiri2012more; @ananthanarayanan2013effective; @stewart2013zoolander] or reissue slow requests’ replicas to a different component [@jalaparti2013speeding; @tailatScale], and then use the quickest replica. Although these techniques work well under light load, they adversely deteriorate the service performance when load gets heavier [@shah2013redundant]. In this paper, we propose a new component-level service scheduler that dynamically schedules the components of a service to appropriate nodes with the assistance of cost-effective online monitors. Compared to existing latency reduction techniques, the proposed scheduler applies an analytic performance model to predict the latencies of all components and their impact on the overall service performance, and then formulates the scheduling decisions based on the predicted performance. The performance model also dynamically updates the prediction results at each scheduling interval by collecting the latest resource contention information during the service execution, thus allowing the scheduler to adapt to changes in performance interference. The concrete contributions of this work are as follows: - We build a flexible analytic performance model to accurately predict the performance of an online service. The basic model comprehensively covers some of the most representative shared resources that are likely to incur contentions and predicts each component’s service time on different nodes by taking the resource contention and performance interference into consideration. The extended model further considers the request queueing delay and estimates the latency of the whole service based on its implementation topology. We show that the proposed model can predict the latency with an average error of 2.68%. - Based on the performance model, we present a framework for component-level scheduling. At each scheduling interval, our approach efficiently identifies the straggling components of a service such that the migration of these components brings the maximum reduction in the overall service latency. The effectiveness of the proposed approach is evaluated using comparative experiments on a variety of realistic workloads publicly available from the BigDataBench suite [@opensourceBigDataBench]. The experiment results in a 100-machine cluster demonstrate that compared with the state-of-the-art techniques on mitigating tail latency, our approach reduces components’ 99th percentile latency by an average of 67.05% and the average overall service latency by 64.16%. The remainder of this paper is organized as follows: Section \[Section: Background\] introduces the background information. Section \[Section: Overview of the framework\] gives an overview of the proposed scheduling framework. Section \[Section: Performance predictor\] presents the performance model and Section \[Section: Algorithm\] explains the scheduling algorithm. Section \[Section: Evaluation\] evaluates the proposed approach. Section \[Section:Related Work\] presents the related work, and finally, Sections \[Section: Conclusion\] summarizes the work. Background {#Section: Background} ========== Sources of component latency variability {#Section:Sources of tail latency} ---------------------------------------- *Resource sharing and contention*. When deploying an online service on a cloud platform, the performance interference due to the co-located batch jobs’ resource contention is often regarded as a major cause of a component’s service time variability [@tailatScale; @leverich2014reconciling]. Some system activities including hardware activities (such as garbage collections of storage devices and energy management behaviors) and software activities (such as kernel daemons and system maintenance) also influence the component’s service time. *Queueing delay*. The component’s service time variability is significantly amplified in the request queueing delay when considering different request arrival rates. Hence the variability of service time and queueing delay work together to cause large latency variability in individual components. Dynamic performance interference of batch jobs {#Section:Managing data center applications} ---------------------------------------------- The dynamic performance interference of batch jobs are caused by their short running periods and continually changing workload characteristics, which can be explained in two aspects. *Workload type*. It has twofold meanings: (i) *Computation semantics*. Batch jobs with different computation semantics (i.e. source codes) may have different resource demands. For example, Sort is an I/O-intensive workload, Bayes classification is a CPU-intensive workload with dominated floating point operations, and Page Index has similar demands for CPU and I/O resources. (ii) *Software stacks*. Model software stacks such as Hadoop and Spark usually provide rich libraries to facilitate development of new applications, and allow a programmer focus on writing a few lines of codes to implement an application. Hence a batch job of the same computation semantic may have considerably different resource demands when implemented with different software stacks [@jia2014characterizing]. For example, Hadoop Bayes is a CPU-intensive workload but Spark Bayes is an I/O-intensive workload. *Input data size*. The resource demand of a job varies when it processes different input data sizes. For example, when running on a 12-core Xeon E5635 processor, the CPU utilizations of the WordCount workload are 31%, 61%, and 79% when its input data sizes are 500MB, 2GB, and 8GB, respectively. Overview of the framework {#Section: Overview of the framework} ========================= As shown in Figure \[Fig: Overview\], the proposed framework for predictive component-level scheduling consists of three modules: the on-line monitors, the performance predictor and the scheduling heuristic. The *on-line monitor* continuously detects two types of information in a running service, whose components are distributed on $k$ nodes of a data center. The first type of information represents the service’s workload status, i.e. its request arrival rate. The second type of information reflects the resource contention information of each component due to its co-located programs on the same node. Specifically, the monitor obtains the request arrival rate by profiling service’s running logs, collects system-level contention information (e.g. core usage and I/O bandwidths) by accessing the proc file system, and profiles micro-architectural contention information (e.g. shared cache misses) using hardware performance counters for Linux 2.6+ based systems. In our monitor, Perf [@PerfWebsite] is used to profile physical machines and Oprofile [@OprofileWebsite] is used to profile VMs. At the end of each scheduling interval, the *performance predictor* collects the monitored information and predicts the component’s latency on all $k$ nodes. This predictor also estimates the impact of individual component latencies on the performance of the whole service based on its implementation topology, and organizes the predicted values as a performance matrix. Using this matrix, the *data center scheduler* applies the scheduling algorithm to identify the straggling components and enforces the appropriate node assignment of the components for the next interval. Consequently, the framework is able to dynamically and efficiently adapt to component latency variability. Note that the proposed scheduling algorithm is not intended to replace, but rather complement the existing scaling or resource provisioning techniques (e.g. reactive scaling up [@han2012lightweight] or prediction-based resource provisioning [@calheiros2011virtual; @han2014enabling] approaches) for multi-stage online services. Specifically, the component-level scheduling is enforced only after the machines have been allocated to the service. At each scheduling interval, the component-node allocation can be conducted by calling the deployment APIs offered by existing distributed realtime computation systems such as Storm [@StormWebsite] and Drill [@DrillWebsite] to migrate the components to the available machines (e.g. VMs or LXCs) on the scheduled nodes. Note also that although this component-node allocation can be enforced by directly migrating the machines to the nodes, we prefer the former solution as it produces lower overheads on scheduling. ![The overview of the framework[]{data-label="Fig: Overview"}](Overview.pdf "fig:")\ Performance predictor {#Section: Performance predictor} ===================== Predicting a component’s latency when running on different nodes is the key step to detect straggling components in a service. This requires the performance predictor to consider all causes of latency variability discussed in Section \[Section:Sources of tail latency\]. In the presence of fine-grained heterogeneity of resource contentions on each component, the basic performance predictor is responsible for collecting the information of resource sharing and contention, and predicting the impact on individual component’s performance (Section \[Section: Basic performance model\]). The extended performance model further estimates the component’s latency by taking the current request arrival rate into account, and calculates the overall service latency based on the service implementation topology (Section \[Section: Extended Performance Model\]). With these two models, the performance predictor finally exposes the component latency variability to the scheduler as a performance matrix of reduced overall service latencies (Section \[Section: Matrix Construction\]). Table \[table: Table of notations\] lists all notations. **Symbol** **Meaning** ----------------------- ------------------------------------------------------ $n$ A node $c$ A component belonging to a service $x$ $c$’s service time $sr$ One type of shared resources $U_{sr}$ $sr$’s resource contention information $\textbf{U}$ The contention vector consists of contention information of all shared resources $RG(U_{sr})$ A basic regression model $RG_{ST}(\textbf{U})$ A combined regression model representing the predicted service time $l$ $c$’s latency $l_{stage}$ A stage’s latency $l_{overall}$ The overall service latency $\textbf{L}$ The matrix of the reduced overall service latency $\textbf{L}[i][j]$ An entry of **L**, which represents the reduced overall latency when component $c_{i}$’s is migrated from its current node to node $n_{j}$ : Table of notations \[table: Table of notations\] Basic performance model {#Section: Basic performance model} ----------------------- Given a component $c$ hosted on a node $n$, the basic performance model is developed to capture the impact of resource sharing and contention on $c$’s performance and estimate its service time $x$. Table \[table: Usage information of shared resources\] lists the contention information of shared resources. The model comprehensively considers both on-chip resources (e.g. shared processing units and caches) and off-chip resources (disk and network bandwidths) contended by different programs on node $n$. In Table \[table: Usage information of shared resources\], core usage represents the ratio of time running instructions on the cores (including private cache hits); MPKI represents the number of instruction Misses Per Kilo Instructions of shared caches including last level cache (LLC), instruction Translation Lookaside Buffer (TLB/ITLB), and data TLB (DTLB). MPKI thus indicates the stalled cycles due to cache contention. Note that the *contention* of these resources comes from $c$’s co-running programs within the same service or across other applications, and node $n$’s hardware/software activities. **Shared resources** **Contention information** ------------------------------ --------------------------------- Floating point and vector $U_{core}$=core usage processing units, pipelines, and data prefetchers LLC, ITLB, DTLB $U_{cache}$=MPKI Disk bandwidth $U_{diskBW}$=the amount of read/write data per second Network bandwidth $U_{networkBW}$=the amount of send/receive data per second : Contention information of shared resources \[table: Usage information of shared resources\] Based on the contention information, the basic performance model predicts $c$’s service time $x$ using two steps. The first step employs a regression model to describe the relationship between one contention information and $c$’s service time. The training of the regression model takes a set of $v$ samples {([$U_{sr_1}$,$x_{1}$]{}),...,([$U_{sr_k}$,$x_{v}$]{})} as input and outputs a model $RG(U_{sr})$, where $sr \in \{core, cache, diskBW, networkBW\}$ and $U_{sr_i} \in \{U_{core_i}$,$U_{cache_i}$,$U_{diskBW_i}$,$U_{networkBW_i}\}$ ($i$= 1,...,$v$). Hence $c$’s service time $x$ is predicted as $RG(U_{sr})$ when the contention information is $U_{sr}$. The training samples are obtained from profiling runs or historical running logs. During the training of regression model, the first step also calculates the relevance (i.e. weight $w_{sr}$) between the contention information of shared resource $sr$ and $c$’s service time. Suppose four regression models ($RG_{U_{core}}$, $RG_{U_{cache}}$, $RG_{U_{diskBW}}$ and $RG_{U_{networkBW}}$) and their weights ($w_{core}$,$w_{cache}$,$w_{diskBW}$ and $w_{networkBW}$) are obtained, the second step predicts $c$’s service time $x$ by producing the final regression model $RG_{ST}(\textbf{U})$ that takes a weighted combination of all the four models: $$RG_{ST}(\textbf{U})=\frac{\sum_{i=1}^4(w_{sr_i} \times RG_{U_{sr_i}})}{\sum_{i=1}^4 w_{sr_i}} \label{Equation: regressionModel}$$ where the resource contention vector $\textbf{U}=\{U_{core},U_{cache},U_{diskBW},U_{networkBW}\}$. Extended Performance Model {#Section: Extended Performance Model} -------------------------- The extended performance model further employs the queueing system to estimate individual component latency under different request arrival rates. Typically, a queueing system can be described as an $A/X/m$, where $A$ represents the distribution of interarrival time of requests; $X$ denotes the distribution of service time; and $m$ is the number of servers. The choice of M/G/1 queueing system is based on the assumption that the distribution of interarrival time of incoming requests are determined by a Poisson process (M for Markov); a component is modeled as a server in the queueing system and the distribution of its service time can follow arbitrary distributions (G for General). Let $\lambda$ be the monitored request arrival rate and $\mu$ be the service rate. Let $\bar{x}$ be the mean service time ($\bar{x}$=$1/\mu$), and $var(x)$ be the variance of service time. Component $c$’s expected latency $l$ is calculated as: $$l=\bar{x} + \frac{\lambda(1+C^{2}_{x})}{2\mu^2(1-\rho)} \label{Equation: a server's response time}$$ where $C^{2}_{x}=\frac{var(x)}{\bar{x}^2}$ is the squared coefficient of variation of service time $x$ and $\rho=\frac{\lambda}{\mu}$ is the server utilization. In many service components, when the service time follows the exponential distribution, that is, the squared coefficient of variation $C^{2}_{x}=1$, the M/G/1 queueing system equals the M/M/1 queueing system and the expected latency $l=\frac{1}{\mu-\lambda}$. At each scheduling interval, a set of resource contention vectors can be collected for each component. By substituting them into Equation \[Equation: regressionModel\], the component’s corresponding service time $x$ can be estimated, so its mean and variance can be calculated. Furthermore, the model computes the overall latency of a service based on its implementation topology. In the online services studied in this work, the processing of a request includes several sequential stages, and each stage parallelizes requests across one or multiple components to aggregate their responses. Hence the calculation of an overall service latency consists of two steps. The first step computes the latency of each stage. Suppose a stage consists of $C$ parallel components, its latency is the maximum value of these component latencies: $$l_{stage}= \max_{1 \leq i \leq C}\{l_{i}\} \label{Equation: a stage's response time}$$ where $l_{i}$ is the latency of component $c_{i}$ ($i$= 1,...,$C$). Suppose the service consists of $S$ sequential stages, the second steps calculates its overall latency: $$l_{overall}= \sum_{j=1}^S l_{stage_{j}} \label{Equation: the overall response time}$$ where $l_{stage_j}$ is the latency of the $j$th stage ($j$= 1,...,$S$). Performance Matrix {#Section: Matrix Construction} ------------------ Suppose $m$ components of a service are deployed in $k$ nodes, the $m \times k$ performance matrix $\textbf{L}$ is constructed using components as rows and nodes as columns. An entry $\textbf{L}[i][j]$ denotes the changes in the *overall service latency* $l_{overall}$ when a component $c_{i}$ is migrated from its current node $n_{current}$ to node $n_{j}$ ($1 \leq i \leq m$ and $1 \leq j \leq k$). This migration may influence all $m$ components’ contention vectors. For any component $c$ of the service, let its original resource contention vector be $\textbf{U}$ and the updated contention vector after the migration be $\textbf{U}^{'}$. Let the resource contention from $c_{i}$ itself be $\textbf{U}_{c_{i}}$ and the resource consumption from all programs on node $n_j$ be $\textbf{U}_{n_{j}}$. Four situations needs to be considered when calculating the updated contention vector $\textbf{U}^{'}$, as listed in Table \[table: Updated usage vectors in migration\]. **Type of component $c$** **Updated contention vector $\textbf{U}^{'}$** -------------------------------- ------------------------------------------------ $c_{i}$ $\textbf{U}_{n_{j}}$ Any component on $n_{current}$ $\textbf{U}$-$\textbf{U}_{c_{i}}$ Any component on $n_{j}$ $\textbf{U}$+$\textbf{U}_{c_{i}}$ Any other component $\textbf{U}$ : Calculation of the updated contention vector $\textbf{U}^{'}$ \[table: Updated usage vectors in migration\] By substituting $\textbf{U}^{'}$ into Equations \[Equation: regressionModel\] and \[Equation: a server’s response time\], $c$’s updated latency $l_i^{'}$ can be calculated. We have: (i) $c_{i}$’s latency $l_i^{'}$ decreases if $n_j$ has lighter resource contention than $n_{current}$; otherwise $l_i^{'}$ increases. (ii) All the components on node $n_{current}$ have decreased latencies because the removal of $c_{i}$ alleviates the resource contention on $n_{current}$. (iii) All the components on node $n_{j}$ have increased latencies because the addition of $c_{i}$ aggravates the resource contention on $n_{j}$. (iv) the latencies of other components keep unchanged. Furthermore, by substituting the updated latencies of all $m$ components into Equations \[Equation: a stage’s response time\] and \[Equation: the overall response time\], the updated overall service latency $l_{overall}^{'}$ can be calculated. Let the overall latency be $l_{overall}$ before the migration. The entry $\textbf{L}[i][j]$ can be calculated as: $$\textbf{L}[i][j] = l_{overall} - l_{overall}^{'} \label{Equation: changes in the overall response time}$$ Figure \[Fig: EntryInMatrix\] shows an example service with three stages, where stage 2 is parallelized into two components $c_2$ and $c_3$. After $c_2$ is migrated from node $n_2$ to $n_4$, $c_4$’s latency $l_4$ increases and $c_2$’s latency $l_2$ decreases. By considering all the updated latencies, the overall service latency before and after the migration can be calculated: $l_{overall}$=57ms and $l_{overall}^{'}$=39ms. Hence the reduced latency $\textbf{L}[2][4]$=18ms. ![An example of entry **L**\[2\]\[4\] in the performance matrix[]{data-label="Fig: EntryInMatrix"}](EntryInMatrix.pdf "fig:")\ The Component-level Scheduling Algorithm {#Section: Algorithm} ======================================== Based on the performance matrix of a service, the scheduler can conduct component-node allocations to minimize the overall service latency. Let the $m$ components {$c_1$,...,$c_m$} be deployed in $k$ nodes {$n_1$,...,$n_k$}, a naïve approach needs a time complexity $O(k^m)$ to identify the optimal component-node allocation and such exhaustive search is not scalable in practical scenarios. However, performance interference is changing overtime, hence optimizing component-node allocation for a particular dynamic scheduling is not worthwhile. The proposed approach, therefore, applies a greedy algorithm with polynomial computation complexity and the algorithm has several iterations. At each iteration, the algorithm aims to minimize the overall service latency by evaluating all possible component-node migrations and selecting the migration that would reduce the latency the most. The pseudocode of the algorithm is presented in Algorithm \[SchedulingAlgorithm\]. At each scheduling interval, the algorithm first constructs the performance matrix $\textbf{L}$ using the performance mode and the monitoring information (line 2). The initial candidate array $C[N_s]$ takes all $m$ components as its elements (line 3). The scheduling process then iteratively executes under two conditions: (a) $C[N_c]$ is not empty; (b) at least one component in $C[N_c]$ can be migrated (line 5). The second condition indicates that a migration is enforced only when the predicted maximum reduced overall latency $l_{max}$ is larger than a specified threshold $\varepsilon$. This threshold prevents inefficient migrations such that the reduced latency cannot compensate the migration cost. In each loop (line 5 to 15), the algorithm first traverses the matrix $\textbf{L}$ to identify a set $S_L$ of entries with the largest value (line 6). Any of these entries denotes the migration of a component that brings the maximal reduction in the overall service latency. If set $S_L$ contains multiple entries, the algorithm further searches $S_L$ to find the entry $\textbf{L}[cmax][nDestination]$ representing the migration that brings the largest reduction to the latency of the migrated component itself (line 7). Component $c_{cmax}$ is regarded as the straggling component and it is allocated to node $n_{nDestination}$ (line 11). The migrated component $c_{cmax}$ is then removed from the candidate array $C[N_c]$ and the matrix $\textbf{L}$ is updated after this migration. $m$: the number of components;\ $k$: the number of nodes;\ $N_c$: the number of candidate components to be migrated;\ $C[N_c]$: the index array of candidate components;\ $\varepsilon$: the migration threshold;\ $A[m]$: the component-node allocation array, where $A[i]$ represents the index of the $i$th component’s hosting node;\ $cmax$: the index of the straggling component $c_{cmax}$;\ $nOrigin$: the index of $c_{cmax}$’s original node;\ $nDestination$: the index of $c_{cmax}$’s destination node;\ $l_{i,j}^c$: A component $c$’s reduced latency when migrating from note $n_i$ to $n_j$.\ Obtain the monitoring information once every scheduling interval; Construct the performance matrix $\textbf{L}$; $C[N_c]$={$c_1$,...,$c_m$}; $l_{\max}$=$\varepsilon$+1; Find a set of entries $S_L$ in the performance matrix L with the largest value; Find the entry $\textbf{L}[cmax][nDestination]$ in $S_L$ with the largest value $l_{nOrigin,nDestination}^{c_{cmax}}$; $l_{\max}=\textbf{L}[cmax][nDestination]$; $nOrigin$ = $A[cmax]$; $A[cmax]=nDestination$; Remove $c_{cmax}$ from $C[N_c]$; UpdateMatrix($\textbf{L}$, $C[N_c]$, $A[m]$, $nOrigin$, $nDestination$); Enforce component-node allocation based on $A[m]$. The detailed matrix updating function is given in Algorithm \[UpdatingFunction\]. The migration of component $c_{cmax}$ from node $n_{nOrigin}$ to $n_{nDestination}$ alleviates the resource contention on $n_{nOrigin}$ but aggravates the resource contention on $n_{nDestination}$, hence has a twofold impact on the predicted reduction of the overall latency for the following migrations. First of all, components to *migrate to* $n_{nOrigin}$ ($n_{nDestination}$) have *increased* (*decreased*) reductions in the overall latency. Hence the entries in the nOriginth and nDestinationth columns should be updated according to Equations \[Equation: regressionModel\] to \[Equation: changes in the overall response time\] (line 1 to 5). Secondly, components to *migrate out* of $n_{nOrigin}$ ($n_{nDestination}$) have *decreased* (*increased*) reductions in the overall latency. Each of such components is hosted on either $n_{nOrigin}$ or $n_{nDestination}$ (line 3) and the component corresponds to one row in the matrix, hence the entries in this row should be updated (line 7 to 10). Note that component $c_{cmax}$ is removed from the candidate array $C[N_c]$, so all the entries related to $c_{cmax}$ are not updated. $N_r$: the number of rows to be updated;\ $R[N_r]$: the index array of rows;\ Update $\textbf{L}[i][nOrigin]$ and $\textbf{L}[i][nDestination]$; Add the $i$th row $r_i$ to $R[N_r]$; Update $\textbf{L}[j][v]$; Figure \[Fig: ServerMigration\] illustrates an example loop, in which migrating component $c_2$ to either node $n_1$ or $n_4$ can result in the maximal reduction in the overall service latency. At the same time, the reduction in $c_2$’s latency is 20ms when it is migrated to $n_1$ and 30ms when it is migrated to $n_4$, which indicates $c_2$ suffers from less performance interference when it is hosted on $n_4$. Hence the scheduling algorithm allocates $c_2$ to $n_4$, after which the entries in the second and fourth columns (representing components to migrate *to* nodes $n_2$ and $n_4$), and the fourth row (representing components to migrate *out of* $n_2$ and $n_4$) are updated. All the entries in the second row are not updated because $c_2$ is not considered in the following scheduling. Let the migration threshold be $\varepsilon$=5ms, we can see that after this loop, the scheduling process is completed because no further effective migration can be conducted: all the values of entries in the updated matrix are smaller than 5ms. ![An example loop of migrating component $c_2$ to node $n_4$[]{data-label="Fig: ServerMigration"}](ServerMigration.pdf "fig:")\ The complexity of each scheduling interval is $O(m^2 \cdot k)$. Specifically, the performance matrix can be constructed in $O(m \cdot k)$ time. The scheduling process can be completed within $m$ loops, where each loops takes $O(m \cdot k)$ to find the optimal migration and $O(m+ m \cdot k)$ to update the matrix. Evaluation {#Section: Evaluation} ========== Experiment methodology {#Section: Experiment methodology} ---------------------- **Experiment platform**. The experiments were conducted in a set of 30 nodes connected with a 1Gb ethernet network. Each node has two 6-core Xeon E5645 processors and hosts multiple VMs using Xen Virtual Machine Monitor (VMM). The operating system of both physical machines and VMs is SUSE Linux Enterprise Server (SLES)-11-SP1. The Xen, JDK versions are 4.0, 1.7.0, respectively. In addition, the versions of Nutch (search engine), Hadoop, and Spark are 1.1, 1.0.2, and 0.8.0, and the versions of Storm, Python, and Zookeeper are 0.9.2, 2.6, and 3.4.6, respectively. **Workloads**. We use representative workloads from the open-source BigDataBench workload suite [@opensourceBigDataBench]. The Nutch web search engine [@nutchsearch] represents the latency-critical online service and its online web search performance was tested. As shown in Figure \[Fig: SearchEngine\], this service has three stages and we call the components at Stage 1, 2, and 3 *segmenting components*, *searching components*, and *aggregating components*, respectively. The batch jobs involve a variety of Hadoop MapReduce and Spark jobs. Hadoop jobs include the two typical CPU-intensive workloads with float point and integer calculations (Naïve Bayes classification and WordCount) and one workload having similar demands for CPU and I/O resources (Page Index). Spark jobs are mostly I/O-intensive workloads including Naïve Bayes, WordCount and Sort. These short-running batch jobs whose execution time ranges from a few seconds to several minutes represent a large fraction of jobs in today’s data center workloads [@reiss2012heterogeneity; @chen2012interactive]. **Compared techniques**. Two classes of state-of-the-art latency reduction techniques are compared. (i) *Request redundancy* [@vulimiri2012more; @ananthanarayanan2013effective; @stewart2013zoolander]. For each request, multiple replicas are created for parallel execution and only the quickest replica is used. Two different redundancy policies, which generate three or five replicas were tested. (ii) *Request reissue* [@jalaparti2013speeding; @tailatScale]. A request is first sent to the most approximate component for execution, and a replica of this request is sent if the first one is not completed after a brief delay. The quickest replica is then used. Two reissue policies, which send a secondary request after the first has been executed for more than the 90th percentile or the 99th percentile of the expected latency for this class of requests, were tested. For simplicity, we will call the four compared techniques, *RED-3*, *RED-5*, *RI-90*, *RI-99*. We also call the basic technique without any redundancy or reissue *Basic*, and our predictive component-level scheduling approach *PCS*. **Metrics**. Two metrics are used to evaluate the performance of the search engine service. The *first* metric is the 99th percentile latency of individual components of all requests. In the case of the request redundancy and reissue techniques, this metric denotes latencies of components belonging to the quickest replica. The *second* metric is the average overall service latency of all requests. **Measurement method**. In the experiments, the monitor dynamically inspects the running service, including its request arrival rate and resource contention information listed in Table \[table: Usage information of shared resources\]. The monitor obtains the request arrival rate and the system-level contention information once every second and the micro-architectural contention information once every minute. This measurement method guarantees low overheads in monitoring and does not affect the application performance. Prediction accuracy {#Section: Prediction accuracy} ------------------- The effectiveness of the proposed scheduling approach is considerably impacted by the performance model’s accuracy. To evaluate this accuracy, we ran each searching component of the service on a VM with 1 core and 1GB memory, and used another VM with 4 core and 4GB memory co-located on the same node to run a Hadoop or Spark job of different input sizes. In each test, we trained the regression models based on the historical running information and predicted the component’s service using the constructed models. As listed in Figure \[Fig: Prediction errors of the performance model under different performance interferences\], in our evaluation, the Hadoop workloads have 20 different input sizes ranging from 50MB to 4GB, and the Spark workloads have 10 different input sizes ranging from 200MB to 7GB, thus having distinct performance interferences to the component’s latency. As shown in Figure \[Fig: Prediction errors of the performance model under different performance interferences\], the prediction errors are *smaller* than 3%, 5%, and 8% in 63.33%, 82.22%, and 96.67% of the evaluation cases, respectively. When considering all the input sizes, the average prediction error is 2.68%, indicating the performance model keeps a good track of the observed latency and it is sufficient for our scheduling heuristic to achieve a near-optimal performance. \ Service Performance {#Section: Service Performance} ------------------- **Evaluation setting**. Following the deployment settings of the previous section, we tested the performance of the Nutch search engine service whose searching components are deployed in 100 VMs. Each component co-locates with a mixed of batch jobs running on VMs of the same node. The Hadoop workloads were tested with continuously changing input data sizes ranging from 1MB to 10GB. Six request arrival rates, namely 10, 20, 50, 100, 200, 500 requests/second, were tested to compare the latency reduction techniques under online services’ diurnal variation in load. **Migration threshold**. As explained in Section \[Section: Algorithm\], the proposed scheduling algorithm employs a migration threshold to control latency reduction and throttle non-beneficial component migration. This threshold should be reasonably high to filter out most of the detrimental migrations whose overheads are larger than the possible latency reduction. On the other hand, the threshold cannot be too high to miss the opportunities for latency reduction. The major overhead of migrating a component is caused by the movement from its current VM to the destination VM. The component runs on a VM installed Storm and its migration is enforced by calling Storm’s deployment APIs. Specifically, Storm first uploads the source codes (e.g. codes for looking up indexes for documents) and the configuration information of the component to ZooKeeper [@ZooKeeperWebsite], a widely used distributed coordination system to manage application deployment. ZooKeeper then allocates them to a new component on the destination VM. At each scheduling interval, the migration of components (e.g. 10 to 20 components) can be completed within 3 seconds without interrupting the running services and only causes small consumptions of memory and I/O resources. Considering the migration cost, we find out that 5% of the accepted overall service latency (100ms) is a reasonable threshold value for the studied online services and thus the threshold in scheduling is set as 5ms. Applying an adaptive threshold to improve the service performance is possible, but it is beyond the scope of this paper. \ **Evaluation results**. Figure \[Fig: Overall service time and tail latency under different request arrival rates\] shows the comparison of service performance for six different techniques. The results show that PCS achieves the smallest tail latencies and the overall service latencies in all cases. This is because during the execution of the service, PCS dynamically enforces different component-node allocations along with the latest performance interference changes on different nodes and reduces the component latency variability by migrating the straggling components to nodes with less resource contentions. By contrast, the *request redundancy* technique just collects responses from the quickest component based on the current service deployment, missing the opportunity to migrate the components to the idlest nodes with the least performance interference. Figure \[Fig: Overall service time and tail latency under different request arrival rates\](a) and (b) show that this technique achieves some latency reduction under light workloads. However, when the arrival rate gradually increases to 500, Figure \[Fig: Overall service time and tail latency under different request arrival rates\](c) to (f) show that this technique adversely causes longer latencies compared to those of Basic. In particular, RED-5 causes the longest latencies because it produces the largest workloads, namely incurring the longest queueing delay, among all techniques. Although the redundancy technique employs the cancelation mechanism that sends messages to cancel other queuing replicas when one replica begins execution, the components still execute replicas of the same request unnecessarily. This phenomenon mainly comes from two sources: (i) all replicas of a request are sent to multiple components simultaneously, hence two components having similar performances may start executing the requests in similar time; (ii) there is a network message delay for different components to communicate each other’s status, hence two components may start executing the same request and the cancelation messages are both in the flight to each other. Moreover, the *request reissue* technique applies a conservative redundancy mechanism that only creates replicas for requests judged as outliers (i.e. requests whose execution time is larger than an expected latency). Results in Figure \[Fig: Overall service time and tail latency under different request arrival rates\] show that compared to the request redundancy technique, this conservative reissue technique causes less performance deterioration when load becomes heavier. **Results**. *Considering all the six request arrival rates, PCS achieves 67.05% reduction in the 99th component latency and 64.16% reduction in the overall service latency when comparing to the request redundancy and reissue techniques.* Scalability of scheduling {#Section: Scalability of scheduling} ------------------------- In proposed scheduling heuristic, the used performance mode is constructed based on profiling of each component. That is, only one out of all homogeneous components needs to be profiled and thereby avoiding the scalability issue associated with the service profiling. For example, in the tested search engine service, only three components (segmenting, searching and aggregating) need to be profiled. Meanwhile, the proposed scheduling algorithm estimates the service performance by analyzing the resource contention information obtained from each component, and hence the *analysis time* scales only linearly with the number of components. Another important aspect of the scalability of scheduling is to *search* the appropriate component-code allocation, and the time complexity of this is $O(m^2\cdot k)$ when allocating $m$ components to $k$ nodes. To evaluate the scheduling algorithm scalability, we measured both the analysis and searching time under different numbers of components and nodes. Figure \[Fig: Scalability of the searching algorithm\] shows that even if the number of components reaches 640 (and the number of nodes reaches 128), the algorithm takes only 551ms to complete. This time is less than 0.1% of the 600 seconds scheduling interval and hence can be ignored. For services with more components, the scheduler could apply a hierarchical strategy that divides the components into small groups of 640 components or less and finds the appropriate component-node allocation between groups and then within groups. The scheduling overhead therefore can remain low even with a large number of components. ![Scalability of the schduling algorithm[]{data-label="Fig: Scalability of the searching algorithm"}](Scalability.pdf "fig:")\ Related Work {#Section:Related Work} ============ Application-level management of service performance --------------------------------------------------- At present, two categories of techniques have been proposed to meet the performance requirement of latency-critical services by alleviate the performance degradation due to resource sharing and contention. The *first category* of techniques disallow the co-location of services with applications incurring large contention of resources such as caches [@kasture2014ubik] and CPU resources [@xu2013bobtail]. The *second category* of techniques dynamically manage applications to meet their performance requirement at run-time according to the monitored interference metrics, such as the LLC miss rate reflecting cache contentions [@ahn2012dynamic] and the bandwidths reflecting I/O resource contentions [@xu2010mitigating]. These techniques focus on addressing performance variability of applications by viewing the application as a whole, ignoring issues relating to fine-grained latency variability of its individual components. However, these components’ tail latency dominates performance of large-scale, parallel services. Tail latency reduction techniques --------------------------------- We now review four categories of reduction techniques. **Modifying hardware/software systems**. These techniques aim at solving the tail latency caused by system design issues. Those include architecture-level design that disables the power saving model to promote system performance [@wang2013impact]; OS-level design that changes the default kernel scheduler to a better scheduler (e.g. Borrowed Virtual Time (BVT)) with better support for time-sensitive requests [@leverich2014reconciling]. **Adding additional resources**. These techniques require additional resources to handle slow requests, either by increasing the parallelism degree of the request processing [@jeon2014predictive] or adding new server components [@stewart2013zoolander]. **Partially processing request**. These techniques reduce tail latency by only using a portion (e.g. 90%) of the quickest sub-requests [@jalaparti2013speeding] or a synopsis representing the entire input data at a high level of approximation [@han2015sarp], thus scarifying result correctness such as query accuracy for reducing service latency. The approach proposed in this work can work together with the above techniques to reduce tail latency, thus forming a complement to these techniques. Both this work and the fourth category of techniques, namely **request redundancy** [@vulimiri2012more; @ananthanarayanan2013effective; @stewart2013zoolander] **and reissue** [@jalaparti2013speeding; @tailatScale] explained in Section \[Section: Experiment methodology\], reduce tail latency by addressing component latency variability. The key idea of the request redundancy technique is to execute the same request on multiple components so as to reduce its latency by using the quickest one. Although these techniques work well when workloads underutilize system resources [@stewart2013zoolander], they start hurting the service performance and adversely worsen the latency when load gets heavier [@shah2013redundant] . Conclusion {#Section: Conclusion} ========== This paper presents a component-level scheduling framework that can dynamically schedule components of a service across hundreds of machines in a cloud data center. To adapt to the changing performance interferences and workloads, this framework leverages cost-efficient online monitors and an analytic performance model to simultaneously predict the components’ latency when running on different nodes. Using the predicted performance, the scheduler identifies straggling components and enforces near optimum component-node allocations. By comparing to the best well-known techniques on reducing tail latency, we demonstrate that our approach achieves significant reductions in both component tail latency and overall service latency. Acknowledgements ================ We sincerely thank Moustafa M. Ghanem and Li Guo and for their useful comments, and the anonymous reviewers for their feedback on earlier versions of this manuscript. This work is supported by Chinese 973 projects under Grants No. 2014CB340402.

ArXiv

--- abstract: 'Solar cells are semiconductor devices that generate electricity through charge generation upon illumination. For optimal device efficiency, the photo-generated carriers must reach the electrical contact layers before they recombine. A deep understanding of the recombination process and transport behavior is essential to design better devices. Halide perovskite solar cells are commonly made of a polycrystalline absorber layer, but there is no consensus on the nature and role of grain boundaries. This review paper concerns theoretical approaches for the investigation of extended defects. We introduce recent computational studies on grain boundaries, and their influence on point defect distributions, in halide perovskite solar cells. We conclude the paper with discussion of future research directions.' author: - 'Ji-Sang Park' - Aron Walsh bibliography: - 'ref.bib' title: Modelling Grain Boundaries in Polycrystalline Halide Perovskite Solar Cells --- Introduction ============ Perovskite solar cells have received a lot of attention partly because of the fast optimization of the device architecture and performance, which is illustrated in the rapid increase of the power conversion efficiency from 3.8 % to 25.2 %.[@kojima2009organometal; @huang2017understanding] Both inorganic (e.g. ) and hybrid organic-inorganic (e.g. ) materials have been studied. The high performance of perovskite solar cells is due to the inherent material properties such as tunable band gap,[@ogomi2014ch3nh3sn; @noh2013chemical] efficient charge generation, long diffusion length,[@stranks2013electron; @shi2015low; @tong2019carrier] and defect tolerance.[@steirer2016defect] The solar cells are also made at relatively low temperatures,[@jeon2014solvent] leading to the production of high-quality solar cells at low cost.[@snaith2013perovskites] Nowadays even higher efficiency has been achieved by perovskite/Si tandem solar cells,[@sahli2018fully] and extensive efforts have been made to achieve large-scale solar cells with long term stability.[@li2018scalable; @jung2019efficient; @bai2019planar] Perovskite solar cells are mainly made of polycrystalline materials, which means that a substantial amount of effort should have been devoted to understanding the effects of grain boundaries.[@huang2017understanding; @wang2017scaling; @lee2018role; @tennyson2019heterogeneity; @castro2019role; @luo2019minimizing; @han2019interface; @chen2019causes] Grain boundaries are known to affect a variety of physical, chemical, and material properties, such as recombination, transport, and even degradation; however, our general knowledge of grain boundaries in halide perovskites remains far from complete. In this review, we focus our scope on the electrical and optical properties of the grain boundaries as there are many unanswered questions to be solved. These include the nature of nonradiative electron-hole recombination in halide perovskites.[@stranks2017nonradiative] Since there a range of terminology frequently used in texts without detailed explanation, we start from the basics of the grain boundary in crystals and studies in other inorganic materials. We not only outline our current understanding of grain boundaries in halide perovskites but also discuss the other extended defects and physical properties that need to be addressed in the future studies. Fundamentals of grain boundaries ================================ ![\[fig:1\] Illustration of $\Sigma$5 (120) grain boundaries. (a) A symmetrical tilt grain boundary, and (b) a twist grain boundary. In (a), the boundary plane is denoted by a dashed line. In (b), the boundary plane is in between the two overlapped planes. Circles with different colors represent the lattice points of grains. When two neighboring lattices are expanded to the other side of the boundary, one of every five lattice points overlap, resulting in the $\Sigma$ value of 5. ](Figure_SIGMA.png){width="50.00000%"} Polycrystalline materials are composed of randomly oriented grains. Grain boundaries are boundaries between such grains, and are typically two-dimensional.[@sutton2006interfaces] Grain boundaries can be categorized by the Miller indices of the grains and the rotation angle. For instance, symmetrical tilt grain boundaries, which are also known as twin boundaries, are formed by two grains with the equivalent Miller index and the zero rotation angle. On the other hand, a twist grain boundary is characterized by a non-zero rotation angle when the rotation axis is perpendicular to the boundary. A characteristic parameter widely used is the $\Sigma$ value,[@cai2016imperfections] which represents how much the two neighboring grains share coincident sites accross the lattice. Perfect materials are considered to have the $\Sigma$ value 1, and larger value indicates that fewer coincident sites form at the grain boundary. Grain boundaries can have one-dimensional or two-dimensional order in their atomic structure.[@cai2016imperfections; @yin2019ceramic] Because every material is polycrystalline in macroscopic quantities, the role of grain boundaries on the material properties has been investigated in many classes of materials. Grain boundaries have been subject of interest in metallurgy for a long time because mechanical properties of metals are highly correlated with the density and distribution of grain boundaries.[@lu2016stabilizing] In the community of thin-film solar cells, there is also growing evidence that grain boundaries can be made beneficial for transport properties. One of the well-known examples is superior photo-conversion efficiencies of polycrystalline CdTe solar cells compared to crystalline CdTe.[@visoly2006understanding; @li2013carrier] To explain this counter-intuitive result, grain boundaries in CdTe have been discussed as being beneficial.[@visoly2004polycrystalline] One hypothesis is that Cl impurities are segregated at grain boundaries, which results in local p-n junctions, resulting in better separation of charge carriers and reduced recombination.[@li2014grain] Similarly, attempts have been made in other materials to create local p-n junctions by inverting the charge carriers of grain boundaries with respect to grain interiors.[@chen2018efficiency; @xu2019defect] Besides the benefits on the electrical properties, impurities segregated at grain boundaries might form precipitates, which can lead to a lower impurity concentration in the grain interior, promoting the gettering.[@lu2003effects] Some studies show that grain boundaries can be relatively benign even though the atomic structure is far from the crystalline order. For instance, grain boundaries and dislocations in Si are relatively benign partly because the over-coordinated Si atoms at the grain boundaries do not introduce deep gap states.[@kohyama1988atomic; @chelikowsky198230] Grain boundaries, however, are generally thought to be detrimental for device performance because of faster carrier recombination and adverse band edge positions.[@klie2005enhanced; @kuo2018grain] For instance, first-principles calculations show that oxygen vacancies can be generated more at grain boundaries in (YBCO) because of the inherent strain, resulting in the lower hole concentration.[@klie2005enhanced] Other first-principles calculations have also shown that some grain boundaries in CdTe, without impurities, can introduce deep levels in the band gap.[@yan2003structure; @park2015stability] These extended defects can be passivated partly by impurities or isovalent element substitution.[@park2016effect] Although there are some examples of beneficial grain boundaries, generally we should expect them to act as recombination centers in solar cells and therefore hamper charge extraction, unless specific passivation routes have been identified and applied.[@moseley2015recombination] Models to investigate grain boundaries ====================================== Non-atomistic models -------------------- ![\[fig:2\] A schematic one-dimensional band diagram of a grain boundary in an n-type semiconductor. The space-charge region is formed due to grain boundary states. Figure is adapted with permission from Reference 42.](Figure_SCR.png){width="45.00000%"} *Stability.* Read and Shockley derived a phenomenological function that describes interfacial energy assuming that grain boundaries consist of dislocations.[@read1950dislocation] Their model indeed described the energy of grain boundaries with small misorientation angle (also known as low angle grain boundaries) well. However, it could not describe the energy of the high angle grain boundaries and the existence of local minima. *Defect segregation.* Grain boundaries are known to serve as reservoirs for point defect (e.g. vacancy, interstitial or substitutional impurity) segregation. This behavior is generally understood in terms of two contributions: elastic and electrostatic.[@sutton2006interfaces; @kliewer1965space; @desu1990interfacial; @gregori2017ion] Elastic interactions between the defects and grain boundaries can be understood as follows. If an impurity atom replaces a host atom, substitutional defects are formed and will generate stress that is proportional to the atomic size mistmatch. Grain boundaries also likely to generate pressure in their vicinity because of different atomic number density and structure compared to the perfect crystal. Electrostatic interactions can dominate when charged defects are formed. The distribution of charges and defects can be obtained through consideration of long-range electrostatics (i.e. Poisson’s equation). *Transport properties.* The function of solar cells is to extract charges generated by absorbing light into electrical contacts, and in this regard, the transport properties are of particular interest. In polycrystalline semiconductors, grain boundaries are expected to have deep trap states because of incomplete chemical bonding at the boundaries and their role as reservoirs for point defect segregation. If there is no band bending near grain boundaries, defects will start to trap free carriers, and as a result, a potential energy barrier is built that eventually inhibits transport of charge carriers from grain to grain. Several theories have been developed to explain the transport behavior of grain boundaries.[@seto1975electrical; @landsberg1984effects; @card1977electronic; @nelson2003physics] Those have successfully shown that the barrier height increases with the trap density at the grain boundaries as the space charge is increased. This results in reduced conductivity and increased grain boundary recombination. *Recombination.* The non-radiative recombination rate of a solar cell can be described by Shockley-Read-Hall (SRH) recombination statistics.[@oualid1984influence; @edmiston1996improved] Assuming a single grain boundary trap level in the gap, the SRH recombination rate under steady-state non-equilibrium condition can be represented in terms of the surface recombination velocity: $$R_{\mathrm{SRH} } = \frac{S_n S_p (np - n_i^2)} {S_n (n + n_t) + S_p (p + p_t)},$$ where $S_n$ and $S_p$ are the electron and hole recombination velocities. $n_t$ and $p_t$ are $n_i \exp{(E_t-E_i)/k_B T}$ and $n_i \exp{(E_i-E_t)/k_B T}$, respectively. $n_i$ is intrinsic carrier density, $k_B$ is the Boltzmann factor, and $T$ is temperature. $E_t$ is the trap level, and $E_i$ represents the intrinsic Fermi level. It has recently become possible to calculate the SRH rate arising from equilibrium populations of point defects from first-principles.[@kim2020upper] Atomistic simulations --------------------- In the 1970s, several methods were developed to calculate the grain boundary energy using interatomic potentials.[@hasson1970structure; @weins1972structure] These attempts are clearly different from previous phenomenological models because we can search the atomic configuration space directly. Stable configurations can be searched by minimization of the grain boundary energy. Then the grain boundary energy was calculated as a function of the misorientation angle, and found to be effective to overcome the previous problems of phenomenological models.[@read1950dislocation] Simple inter-atomic potentials such as Morse and Lennard-Jones potentials were used in early studies, but more sophisticated potentials are currently used.[@olmsted2009survey; @holm2010comparing; @restrepo2013genetic; @ratanaphan2015grain] The above approach based on structure searches using interatomic potentials were successful to predict the grain boundary atomic structure in metals; however, there was a need for a quantum mechanical description of semiconductors. Tight-binding methods were adapted to understand extended defects, and the density of states (DOS) of grain boundaries was calculated as well.[@de1980electronic; @thomson1984theoretical; @chadi1985new; @paxton1988simple] In 1986, when a first-principles method was first applied to study twin boundaries in crystals, empirical tight binding methods were employed to optimize the structures of grain boundaries in Si because of the lower computational cost.[@divincenzo1986electronic] More recently, an effective tight-binding model was developed to understand a grain boundary in $\mathrm{YBa_2Cu_3O_{7-\delta}}$ superconductor.[@graser2010grain] Motion and annihilation of grain boundaries in graphene has been investigated using a molecular dynamics tight-binding method as well.[@lee2013atomistic] First-principles simulations ---------------------------- To fully describe the stability and the electronic structure of materials, a fully quantum mechanical calculation method without empirical parameters is ideal. First-principles density functional theory (DFT) meet these needs[@kohn1999nobel] and can be used to investigate the stability and the electronic structure of grain boundaries. We note that there a number of technical challenges for halide perovskites owing to strong relativistic effects (due to Pb) and dynamic structural effects.[@whalley2017perspective] *Stability.* Since periodic boundary conditions are typically employed in simulations of crystals, a supercell model may contain two interfaces if there is no vacuum region in the supercell. Since a grain boundary is a type of interface, the method used to obtain the interface energy can be directly applicable:[@park2018quick] $$E_f(\mathrm{GB}1) + E_f(\mathrm{GB}2) = (E_{tot}(\mathrm{GB})-\Sigma_i n_i \mu_i) / A,$$ where $E_f(\mathrm{GB})$ is the formation energy of a grain boundary, $E_{tot}(\mathrm{GB})$ is the total energy of a given supercell with two grain boundaries. $n_i$ and $\mu_i$ are the number of atom of atomic species $i$ and the corresponding chemical potential. $A$ is the area of the grain boundary in the supercell. If the two interfaces are exactly the same, the formation energy becomes $$E_f(\mathrm{GB}) = (E_{tot}(\mathrm{GB})-\Sigma_i n_i \mu_i) / 2A.$$ In many cases, grain boundaries in the supercells are not identical, and therefore charges can be transferred between the grain boundaries and affect the formation energy. To obtain the formation energy of a single grain boundary, we need to employ a slab geometry that contains one interface and two surfaces. As there are two surfaces, their contribution to the formation energy should be subtracted. Park *et al.* used slab geometry and successfully obtained the formation energy of grain boundaries in CdTe.[@park2015stability] *Electronic structure.* The electronic structure of grain boundaries is often calculated with DFT using the generalized gradient approximation (GGA) exchange-correlation functional, which underestimate the band gap.[@perdew1985density] Hybrid DFT calculations,[@heyd2003hybrid; @heyd2005energy] which are commonly used to correct the band gap of semiconductor materials nowadays, are currently too computationally heavy for describing grain boundaries. Often-used strategies are to introduce the on-site Coulomb interaction[@yin2013defect; @yin2014engineering] or hybrid calculations only for analysis of the electronic structure.[@park2015stability; @park2016effect; @park2019stabilization] To further reduce the computational cost, a sparse *k*-point grid mesh can be used for the Fock exchange potential or non-self-consistent-field calculations can be performed.[@pan2018spin; @park2019stabilization; @park2020examination] *Prediction of atomic structure.* An important question is how to generate a representative three-dimensional atomistic model of a grain boundary. Structural properties of grain boundaries can be identified by electron backscattering diffraction (EBSD) at a microscopic scale.[@humphreys2001review; @moseley2015recombination] Typically various types of grain boundaries are observed. Further atomistic details can be obtained by transmission electron microscopy (TEM) measurements.[@yan2003structure; @li2014grain; @liebscher2018strain] A potential problem, however, is that information gathered from the experiment like two-dimensional images could be insufficient to construct a three-dimensional atomic structure. We also have few images than grain boundaries formed in real samples. To overcome this problem, statistical techniques such as genetic algorithms have been developed. Grain boundaries in metals have been investigated using the force field calculations, which are relatively cheaper than DFT calculations.[@olmsted2009survey; @restrepo2013genetic] Grain boundaries in semiconductors, on the other hand, are better to be investigated by the quantum mechanics code due to the importance of the electronic structure. Chua *et al.* investigated both stoichiometric and non-stoichiometric grain boundaries in .[@chua2010genetic] In their framework, thousands of trial configurations were explored using empirical interatomic potentials, and thereafter structures were refined using first-principles electronic structure methods. Similarly Park *et al.* performed DFT calculations but using the atomic orbital basis to explore the configuration space.[@park2019stabilization] Some screened structures were re-examined using plane-wave basis methods. We also note that the mirror symmetry of symmetrical tilt grain boundaries in semiconductors can be broken as a result of the rigid body translation as examined in the literature recently.[@marquis2004finite; @liebscher2018strain] Experimental findings ===================== *Beneficial grain boundaries.* The first question to be answered is whether grain boundaries in halide perovskites are beneficial or not from the device perspective. Early studies using Kelvin probe force microscopy (KPFM) and conductive atomic force microscopy (AFM) reported that grain boundaries are beneficial because charges are efficiently separated and collected through grain boundaries.[@yun2015benefit; @kim2015efficient; @li2015microscopic] Later Yun *et al.* used KPFM to detect local surface potentials caused by ion profiles in halide perovskites.[@yun2016critical] Their KPFM experiments have shown that the contact potentials difference (CPD) of grain boundaries and grain interior exhibit different trends. The grain boundary always had a lower CPD value than grain interior when there is no bias voltage. However, applying positive bias (more than 1 V) makes the grain boundaries to have higher CPD than the grain interiors, whereas negative bias exhibit the opposite effect. It also took several minutes for the CPD value to return to its initial value after the bias voltage is removed. Based on these results, the authors conclude that there were more ions at the grain boundary initially or ions migrate easily through the grain boundaries. A phenomenological model developed by the authors claims that redistribution of ions under illumination condition results in stronger band bending at grain boundaries. The contact potential difference at grain boundaries in KPFM measurements was found to be modulated by additives.[@faraji2018grain] *Neutral grain boundaries.* Some studies focused on the transport properties of grain boundaries. MacDonald *et al.* found that grain boundaries are electrically resistive, at least near the top of the film.[@macdonald2016methylammonium] Reid *et al.* observed that mobility-yield products decrease with decreasing the grain size.[@reid2016grain] Yang *et al.* constructed a kinetic model of charge transport and recombination process based on their high-resolution confocal fluorescence lifetime imaging microscopy experiments,[@yang2017grain] and pointed out that the weaker PL intensity does not necessarily mean a shorter lifetime of carriers. Snaider *et al.* also concluded that the carrier transport is slowed down by grain boundaries.[@snaider2018ultrafast] It was also discussed that long carrier lifetime can compensate for the higher resistivity at the grain boundary. Sherkar *et al.* performed device simulation modeling and found that grain boundaries become relatively inert when the charged traps become neutral after charge trapping. [@sherkar2017recombination] *Detrimental grain boundaries.* Local fluorescence lifetime imaging experiments have shown that the photoluminescence intensity is lower near the grain boundary than the center of the grain in methylammonium lead iodide ().[@de2015impact] This result indicates that grain boundaries are active for non-radiative recombination. Passivation of the boundaries (e.g. using pyridine) resulted in brighter PL. Scanning Electron Microscopy (SEM), which allows us to study surface morphology, is not sufficient to identify crystallographic information of grains and grain boundaries. Electron backscatter diffraction (EBSD) is the standard method to measure crystallographic information of grains, but its usage was hampered because of beam damage to halide perovskite samples. Adhyaksa *et al.* used a solid-state EBSD detector with better sensitivity to resolve this problem and found that grain boundaries in halide perovskites can act as recombination centers.[@adhyaksa2018understanding] ![\[fig:3\] (A) Topography map and (b) line profile data of topography and contact potential difference (CPD) under different conditions. (C) SEM, (D) fluorescence spectroscopy and (E) their composite image, showing that photoluminescence intensity spatially varies. Figures adapted with permission from Ref. 82 and 92.](Figure_EXPT.png){width="45.00000%"} First-principles studies ======================== Neutral grain boundaries ------------------------ Yin *et al.* studied two kinds of GBs, $\Sigma$3 (111) GB and $\Sigma$5 (310) GB in .[@yin2015origin] In their DFT-PBE calculations, they found that grain boundary models do not introduce deep levels in the band gap even though there are I-I bonds formed, which are not formed in perfect , as well as Pb dangling bonds. [@yin2014unusual] This is in line with the fact that iodine vacancy (Pb dangling bonds) and iodine interstitials, which form I$-$I bonds, are shallow defects in their previous study.[@yin2014unusual] Iodine anti-site defects also form I$-$I bonds and even introduce deep levels in the band gap, but those were not as stable as I interstitials. Besides these defects, Pb antisite defects created deep levels in their PBE calculations without spin-orbit coupling, and all of them had relatively high formation energy. The electronic structure of $\Sigma$3 (111) GB was more carefully examined by using the hybrid functional with spin-orbit coupling, but they were not able to find a deep level in the gap. They ascribed the origin of the deep-state-free GBs in is due to the strong *sp* coupling of the valence band maximum and to the large atomic size of . The former and the latter results in the higher band edge and the shallower defect states. Extrinsic elements such as Cl and O were found to be stable at the grain boundaries, and weaken the halogen-halogen bonds (i.e. I$-$I) at GBs and thus are able to reduce the density of shallow trap states. Guo *et al.* performed more comprehensive studies on the grain boundaries in halide perovskites ( where X = I, Br, and Cl).[@guo2017structural] Using DFT, they investigated symmetrical tilt grain boundaries having four Miller indices. Remarkably, they considered rigid body translation in the simulation to find stable geometry of the grain boundaries. Contour maps of the grain boundary energies were also reported. The grain boundary energies were obtained and based on those data, some stable structures were selected. Electronic structure calculations, performed using DFT-PBE, showed that the stable structures do not have any deep gap states, consistent with the previous study.[@yin2015origin] The electronic structure of was also examined using the same geometry, but it also had no deep gap states. ![\[fig:4\] (a) Atomic structure of a $\Sigma$5 \[130\] grain boundary in . The boundary of the cells is represented by the solid lines. The dashed lines in the middle represent the grain boundary. (b) Split-interstitial configuration of iodine interstitial (I$_i$), denoted by a green circle. (c) I$_i$ passivating under-coordinated Pb atoms, denoted by a blue pentagon. (d) I$_i$ with an iodine trimer (I-I-I) denoted by an orange rhombus. (e,f) The relative formation energy of I$_i$ in 1$-$ and 1$+$ charge state, respectively, as a function of the distance from the grain boundary. The results show the segregation of I$_i$ defects to the grain boundary. Figures adapted with permission from Ref. 97.](Figure_DFT.png){width="40.00000%"} Defect-mediated recombination ----------------------------- The atomic structure of grain boundaries differ from the bulk region, and thus defect properties can be affected. Thind *et al.* studied the grain boundaries and other planar defects that can be formed in .[@thind2019atomic] They first made nanocrystals and then fused them to make larger crystals. Various boundaries can be generated depending on how the nanocrystals are aligned. Based on the atomic structure observed experimentally, they constructed an atomic structure model for DFT calculations and investigated the electronic structure. In their study, grain boundaries cause band offsets and impact electron transport. A specific type of grain boundary ($\Sigma 5$) repels electrons and attracted holes. However, Ruddlesden-Popper faults repel both kinds of carriers. This means that the transport and optoelectronic properties of grain boundaries are greatly influenced by the atomic structure of the boundary. Interestingly, their calculations predict that the bromine vacancy could cause relatively deep levels.[@thind2019atomic] It is worth pointing that PBE describes defect properties of quite differently to hybrid functionals,[@du2015density; @meggiolaro2018iodine] which could impact the conclusions. In an early study done by Shan *et al.*, intrinsic defects were found to segregate to boundaries.[@shan2017segregation] Since they performed the calculation using PBE with spin-orbit-coupling, the band gap was underestimated and only anti-site defects were assigned to be deep traps. Later Park *et al.* re-visited iodine interstitial defects,[@park2019accumulation] which introduce deep levels in the band gap[@du2015density; @whalley2017h; @meggiolaro2018iodine] and diffuse fast.[@yang2016fast; @futscher2019quantification] Iodine interstitials were found to easily segregate at the grain boundary, whichever charge state it has. The driving force of the segregation has been attributed to the structural relaxation, which is parameterized with the distance between iodine atoms forming the interstitial defect. The results can be understood as the lower atomic density at the grain boundaries promote room for relaxation and hence energy lowering. The numerical solution of Poisson’s equation revealed that both donor and acceptor defects are heavily compensated at the grain boundaries. To investigate the effect of the environment on the defect levels, Park *et al.* assumed halide dimers and trimers embedded in a dielectric medium and found that the acceptor (I$_i^{1-}$) is expected to be shallower and the donor state (I$_i^{1+}$) deeper. The high concentration of deep traps can shorten the carrier lifetime through defect-assisted recombination at grain boundaries. Meggiolaro *et al.* also performed first-principles calculations to investigate the effect of environment on the formation energy of iodine interstitial defects.[@meggiolaro2019formation] They found that the defect formation energy at the surface was significantly lowered compared to bulk. Based on these results, they constructed a phenomenological equation to estimate defect formation energy as a function of grain size, which corresponds to the weighted average of defect formation energies corresponding to the bulk and surface. Simulation results showed that the more defects are easily formed as the grain size decreases. We note that Hentz *et al.* developed an experimental setup to measure the photoluminescence of laterally biased sample and concluded that nonradiative recombination centres migrate through grain boundaries.[@hentz2019visualizing] Among several potential defects, iodine interstitials were discussed to be the best candidate to explain the result. This is also consistent not only with the recent DFT calculation results that nonradiative iodine interstitials defects are easily accumulated at the grain boundaries[@park2019accumulation], but also with the previous experimental results of fast ion migration through grain boundaries.[@shao2016grain; @xing2016ultrafast] Band gap narrowing ------------------ Although many computational studies overlooked anion mixing, McKenna has shown that the halide composition ratio can vary spatially.[@mckenna2018electronic] According to his first-principles calculation, the {111} twin boundary in pure formamidinium lead iodide only creates a small barrier of less than 0.1 eV. However, in the mixed-cation mixed-halide perovskite, Cs and I atoms were segregated at the twin boundary. The I accumulation caused the higher valence band edge at the boundary by more than 0.2 eV than in the bulk region, indicating that the photo-generated carriers could be recombined at the twin boundary. Long *et al.* performed molecular dynamics simulations and found that a grain boundary in pure has a higher valence band edge than the bulk region.[@long2016unravelling] In their study, the reduced band gap and the higher coupling between the band edges result in the faster electron-hole recombination at grain boundaries. Cl incorporation reduced the coupling and thus the recombination became weaker. Passivation strategies ---------------------- If grain boundaries act as nonradiative recombination centers, then the origin of the deep levels should be identified, removed or passivated. Considering their importance for device efficiency and possibly lifetime, various attempts have been made to passivate the grain boundaries.[@lee2018role; @castro2019role; @chen2019causes; @luo2019minimizing] Here, we introduce some studies showing consistency with DFT calculations. On the experimental side, compositional engineering is a well-known method to enhance device efficiency.[@stranks2013electron; @jeon2015compositional] de Quilettes *et al.* found a positive correlation between the PL intensity and Cl composition by using energy dispersive x-ray spectroscopy with confocal fluorescence maps. Zheng *et al.* employed a surface model and claimed that Cl can passivate ionic point defects (e.g. Pb$_\mathrm{I}$ anti-site) accumulated at the surface, noting that the major defects at the surface were uncertain at the moment of study.[@zheng2017defect] On the computational side, Meggiolaro *et al.* has found that Br interstitials and Cl interstitials introduce shallower acceptor levels than I interstitials.[@meggiolaro2018iodine] Cl incorporation at the grain boundaries can be effective in this regard as the deeper I defects are replaced by shallower Cl defects. Another category is the passivation of surface defects by extrinsic impurities or molecules. For instance, Noel *et al*. found that Lewis bases such as thiophene and pyridine can be used to reduce nonradiative recombination in halide perovskites.[@noel2014enhanced] They suggested that the molecules can be bound to defects (vacancies) on surfaces or grain boundaries, passivating defects and improving performance accordingly. Later Shao *et al.* have claimed that PCBM molecules can also passivate grain boundaries based on experimental data,[@shao2014origin] and later Xu *et al.* also found the same conclusion based on collaboration between experiment and modelling.[@xu2015perovskite] In their DFT calculation, PCBM adsorption passivates the grain boundaries by making the deep levels of I$_{\mathrm{Pb}}$ closer to the conduction band minimum. Remaining open questions ======================== We have discussed how first-principles methods have been used to describe the structure and properties of grain boundaries in halide perovskites. Here, we highlight some of the open issues in the topic. *Twin domains.* The formation of twin domains in has been reported based on TEM and selected area electron diffraction (SAED) experiments.[@rothmann2017direct] In the TEM experiments, the striped contrast patterns (alternating bright and dark colors) were observed. Also in the SAED experiments, the split spots were observed in the striped domains. Morphology, however, was found to be not correlated with stripe contrast. It has been claimed in another study that twin defects lowers the solar conversion efficiency, which the absorption coefficient were not affected.[@Tan2019] The formation of twin boundaries was measured from the shift of the (100) *d* peak in TEM measurements. *Mixed phases.* There is growing evidence that halide perovskites are not a single phase in real samples. Kim *et al.* have reported that tetragonal and cubic can coexist at room temperature.[@kim2018self] They also observed superlattices composed of cubic and tetragonal phases in their TEM analysis. As there is no compositional change in their analysis, the superlattices were concluded to be formed as a result of intrinsic structural changes. The detailed formation mechanism, however, and their effects on the device performance are not clearly revealed by first-principles calculations. *Internal grain structure.* Using photoluminescence microscopy, Li *et al.* have reported the formation of subgrain boundaries which cannot be observed by conventional Atomic Force Microscopy (AFM) and SEM measurements.[@li2018subgrain] Those boundaries were reported to act as non-radiative recombination centers and also restrict carrier diffusion. Jones *et al.* used synchrotron scanning micro-XRD measurements with local time-resolved PL measurements to identify that lattice strain is directly associated with enhanced defect concentration and therefore non-radiative recombination.[@jones2019lattice] Jariwala *et al.* have reported that even local orientation may vary even inside a grain, exhibiting higher recombination.[@jariwala2019imaging] This finding is in contrast to a common belief that materials are aligned in a certain direction in a grain. *Dynamic properties.* Most of the studies investigated the defects in temporal or spatial average, however, time-dependent phenomena should be investigated to obtain a complete picture of the grain boundary. On experimental sides, Snaider *et al.* investigated carrier transport phenomena using Transient Absorption Microscopy (TAM).[@snaider2018ultrafast] Later Jiang *et al.* investigated carrier dynamics using SEM correlated to TAM.[@jiang2019transient] The latter study found that grain boundaries have an increased population of the sub-band-gap states than grain interior, higher quasi-Fermi energy, and faster carrier cooling rate. The origin of the shallow state was suggested to be I$-$I bonds at the grain boundaries, partly based on a previous DFT calculation.[@yin2015origin] Certainly, future studies should account for the dynamics of the photo-generated carriers. Outlook ======= We have outlined several ways to investigate grain boundaries. Early studies employed phenomenological non-atomistic methods, but the development of computer simulation methodologies and the high-performance computers have allowed us to study grain boundaries using first-principles materials modelling. There is an urgent need to study various extended defects that can be generated in halide perovskite using this methodology. Trying to narrow the gap between the calculations and experiments should be pursued as well. For instance, various techniques being developed in point defect studies should be introduced to study to extended defects that are ubiquitous in the polycrystalline thin films being used in solar cells. This work was supported by a National Research Foundation of Korea (NRF) grant funded by the Korean government (MSIT) (No. 2018R1C1B6008728). This work was supported by a National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. 2019M3D1A210410811). We are grateful to the UK Materials and Molecular Modelling Hub for computational resources used in the research discussed in this review, which is partially funded by EPSRC (EP/P020194/1).

ArXiv

--- abstract: 'We present a VLT/X-Shooter spectroscopy of the Lyman continuum (LyC) emitting galaxy [*Ion2*]{} at z=3.2121 and compare it to that of the recently discovered strongly lensed LyC$-$emitter at z=2.37, known as the [*Sunburst*]{} arc. Three main results emerge from the X-Shooter spectrum: (a) the [Ly$\alpha$]{} has three distinct peaks with the central one at the systemic redshift, indicating a ionised tunnel through which both [Ly$\alpha$]{} and LyC radiation escape; (b) the large O32 oxygen index ([\[ <span style="font-variant:small-caps;">iii</span>\]$\lambda\lambda4959,5007$]{} / [\[<span style="font-variant:small-caps;">ii</span>\]$\lambda3727,3729$]{}) of $9.18_{-1.32}^{+1.82}$ is compatible to those measured in local (z $\sim 0.4$) LyC leakers; (c) there are narrow nebular high-ionisation metal lines with $\sigma_v < 20$ , which confirms the presence of young hot, massive stars. The [<span style="font-variant:small-caps;">ii</span>$\lambda1640$]{} appears broad, consistent with a young stellar component including Wolf$-$Rayet stars. Similarly, the [*Sunburst*]{} LyC$-$emitter shows a triple$-$peaked [Ly$\alpha$]{} profile and from VLT/MUSE spectroscopy the presence of spectral features arising from young hot and massive stars. The strong lensing magnification, ($\mu > 20$), suggests that this exceptional object is a gravitationally$-$bound star cluster observed at a cosmological distance, with a stellar mass M $\lesssim 10^7$  and an effective radius smaller than $20$ pc. Intriguingly, sources like Sunburst but without lensing magnification might appear as [*Ion2*]{}$-$like galaxies, in which unresolved massive star clusters dominate the ultraviolet emission. This work supports the idea that dense young star clusters can contribute to the ionisation of the IGM through holes created by stellar feedback.' author: - | \ $^1$INAF – OAS, Osservatorio di Astrofisica e Scienza dello Spazio di Bologna, via Gobetti 93/3, I-40129 Bologna, Italy\ $^2$Kapteyn Astronomical Institute, University of Groningen, Postbus 800, 9700 AV Groningen, The Netherlands\ $^3$INAF – Osservatorio Astronomico di Trieste, via G. B. Tiepolo 11, I-34143, Trieste, Italy\ $^4$INAF – Osservatorio Astronomico di Roma, Via Frascati 33, I-00078 Monte Porzio Catone (RM), Italy\ $^5$Dipartimento di Fisica e Scienze della Terra, Università degli Studi di Ferrara, via Saragat 1, I-44122 Ferrara, Italy\ $^6$INAF – Osservatorio Astronomico di Capodimonte, Via Moiariello 16, I-80131 Napoli, Italy\ $^7$European Southern Observatory, Alonso de Cordova 3107, Casilla 19, Santiago 19001, Chile\ $^8$Dipartimento di Fisica, Università degli Studi di Milano, via Celoria 16, I-20133 Milano, Italy\ $^{9}$Astronomy Department, University of Massachusetts, Amherst, MA 01003, USA\ title: 'Ionising the Intergalactic Medium by Star Clusters: The first empirical evidence' --- galaxies: formation – galaxies: starburst – gravitational lensing: strong Introduction ============ Recently, extensive surveys attempting to identify and study galaxies emitting Lyman continuum (LyC) radiation across a large range of cosmic time have yielded several low-redshift cases whose properties are believed to be representative of the galaxies at redshift $z>7$ that contributed the radiation that has re-ionised the Universe. Since the direct detection of ionising radiation from the epoch of re-ionisation (EoR) is not possible because of the cosmic opacity, the low-redshift “analogs” of the distant galaxies play a key role in understanding the mechanisms that allow the escape of ionising radiation from star-forming galaxies. The census of LyC-galaxies is growing fast, both in the nearby Universe [@izotov18 and references therein] and at high-redshitf, i.e. $z\approx 3.5$, [@vanz18; @vanz16b; @debarros16; @shapley16; @bian17], and relevant progress has recently been made in a statistical sense by analysing dozens of high redshift galaxies with dedicated HST imaging [e.g., @felce18; @jure17] and deep spectroscopy [e.g., @steidel18; @marchi18]. In particular, a positive correlation among LyC escape and [Ly$\alpha$]{} equivalent width has been inferred, as well as an apparently higher [*fesc*]{} at fainter ultraviolet magnitudes, such that galaxies might account to more than 50% of the ionising budget at $z\sim3$ [@steidel18]. Spectral features like the profile of the escaping [Ly$\alpha$]{} line, the strength of the low-ionisation interstellar absorption lines tracing the covering fractions of neutral gas (e.g., [\[<span style="font-variant:small-caps;">ii</span>\]$\lambda1334$]{}, [\[<span style="font-variant:small-caps;">ii</span>\]$\lambda1260$]{}) the line ratios tracing the ionisation$-$ or density$-$bounded conditions in the interstellar medium (like the O32 index, [\[ <span style="font-variant:small-caps;">iii</span>\]$\lambda\lambda4959,5007$]{} / [\[<span style="font-variant:small-caps;">ii</span>\]$\lambda3727,3729$]{}), hold the promise to provide useful diagnostics of the mechanisms that govern the escape of ionising radiation, although we currently do not yet know which properties provide necessary and/or sufficient conditions for this to happen [e.g., @schaerer16; @verhamme17; @izotov18; @jaskot13; @mckinney19; @gaza18; @JC18; @reddy16; @reddy18; @steidel18; @grazian17]. ![The most relevant atomic transitions of [*Ion2*]{} in the VLT/X-Shooter UVB (R=5300), VIS (R=8900) and NIR (R=5600) arms, colour-coded in blue, green and red boxes, respectively. The expected positions at systemic redshift z=3.2121 are indicated with blue symbols (‘+’). For each transition, the corresponding rest-frame wavelength is reported. The rest-frame wavelength width of the zoomed spectra is reported at the bottom, for each arm. The significance of the reported transitions is summarised in Table \[tab\].[]{data-label="lines"}](Figure1.pdf){width="8.5cm"} While the current growing samples serve as a reference for the identification of cosmic reionisers, the physical processes that made these galaxies transparent to LyC radiation are yet to be understood. The way ionised channels are carved in the interstellar medium is under continuous investigation, especially in the local Universe [@herenz17; @micheva17; @micheva19; @bik18; @kehrig18] where the detection of LyC radiation is instrumental to catch such episodes in the act [e.g., @h11]. It remains unclear what is the spatial distribution of the escaping ionising radiation, the porosity and the kinematics of the neutral gas, and the role the internal constituents of high-z LyC$-$galaxies (namely, the star-forming complexes, OB-associations, young massive clusters and massive stars) have in carving such ionised regions. In general, the small physical scales, likely of the order of a few ten pc and in which such constituents originate, are still unreachable at cosmological distances (e.g., a typical HST pixel of 30 milli-arcsecond subtends 250-300 pc at $z\sim 2-6$, encompassing one or more star-forming complexes). This limitation is even worse when ground-based seeing-limited spectroscopy is performed. This work presents new VLT/X-Shooter spectroscopy of a LyC$-$galaxy dubbed [*Ion2*]{} [@vanz16b], and VLT/MUSE observation of the recently discovered strongly lensed LyC emitter at z=2.37 [known as Sunburst arc, @dahle16; @rivera17; @rivera19]. The similarity between the exceptional strongly$-$lensed Sunburst and [*Ion2*]{} sheds light on the possible “engine” behind the spatially unresolved high-z LyC leakers, like [*Ion2*]{}. We make use of archival VLT/MUSE and HST/ACS data targeting the Sunburst object with the aim to emphasise and explore this connection. We assume a flat cosmology with $\Omega_{M}$= 0.3, $\Omega_{\Lambda}$= 0.7 and $H_{0} = 70$ km s$^{-1}$ Mpc$^{-1}$. The Lyman continuum galaxy [*Ion2*]{} ===================================== [*Ion2*]{} is a well known LyC emitter at z=3.2121 lying in the CDFS showing an escape fraction higher than 50% [@vanz15; @vanz16b]. Here we present new VLT/X-Shooter observations that improve (in terms of depth, spectral resolution and wavelength coverage) our previous analysis [@debarros16]. X-Shooter observations ---------------------- [*Ion2*]{} was observed during November 2018 for a total integration time of 5 hours under optimal seeing conditions, typically $0.5 - 0.7$ arcsec. The slit widths were 1.0$''$, 0.9$''$, 0.9$''$, corresponding to a spectral resolution of R=5400, 8900 and 5600 in the UVB, VIS and NIR arms, respectively (Prog. 0102.A-0391(A), P.I. Vanzella). Given the good seeing, the aforementioned resolution values must be considered as lower limits. The data reduction was carried on as described in several previous works (we refer the reader to @vanz16a [@vanz17b]), in which the AB-BA sky subtraction scheme was implemented with single exposures of 915s, 946s and 900s on each of the three UVB, VIS and NIR arms, respectively. The target was dithered $2.4''$ along the slit. The spectral range from the U to the K-band covers the rest-frame wavelengths which include the [Ly$\alpha$]{} and the optical lines [\[ <span style="font-variant:small-caps;">iii</span>\]$\lambda\lambda4959,5007$]{}. The continuum is barely detected, from which a dozen of emission lines emerge with S/N ratio spanning the interval $2-50$ (see Figure \[lines\] and Table \[tab\]). A careful analysis of the statistical significance of the spectral features is reported in the Appendix \[SN\]. Results ------- The X-Shooter spectrum shows at least three features not observed with previous spectroscopy [@debarros16]: a multi-peaked [Ly$\alpha$]{} profile, the presence of narrow ultraviolet high ionisation lines and the new detection of [\[<span style="font-variant:small-caps;">ii</span>\]$\lambda3727,3729$]{} and [H$\beta$]{} optical rest-frame lines. Below we summarise these new results. $\bullet$ The [Ly$\alpha$]{} profile shows three peaks with the central one placed exactly at the systemic redshift (z=3.2121, see Figure \[ion2\]), resembling the same [Ly$\alpha$]{} structure observed in another LyC$-$galaxy discovered at z=4.0 [dubbed [*Ion3,*]{} @vanz18]. This is the fourth confirmed LyC emitter showing a [Ly$\alpha$]{}-peak emerging at the systemic velocity. The four objects are also characterised by a quite large escape fraction of ionising radiation, [*fesc*]{} &gt; 50% (@vanz18 [@rivera19; @izotov18]). Figure \[lyas\] shows a comparison of the [Ly$\alpha$]{} shapes, and includes [*Ion2*]{}, [*Ion3*]{}, a high-surface density LyC emitter at z=0.4317 with a very large escape fraction ([*fesc*]{} $>72$%) from the @izotov18 sample and Sunburst. In the LyC emitters shown in Figure \[lyas\] the positions of the peaks at the two sides of the central one are different, reflecting different kinematical features characterising each system. The same line shape has been investigated by @rivera17 performing radiative transfer (RT) calculations in the framework of three different scenarios including the expanding shell models [@DS11; @gronke15]: a density-bounded medium, picket fence medium and the presence of a ionised channel embedded in the [<span style="font-variant:small-caps;">i</span>]{} shell. Only the last case suitably reproduces the triple-peaked [Ly$\alpha$]{} profile: a significant amount of [<span style="font-variant:small-caps;">i</span>]{} gas with a perforated channel accounts for both the typical [Ly$\alpha$]{} broadening by frequency diffusion (peaks far from the resonance frequency, e.g., ‘-2,-1,+1’) and the superimposed [Ly$\alpha$]{} emission at zero velocity (peak ‘0’). Additionally, as discussed by @rivera17, the profile of the non-scattered [Ly$\alpha$]{} photons escaping through an optically thin tunnel would resemble the width of the Balmer emission lines, that would represent a proxy of the intrinsic [Ly$\alpha$]{} shape before undergoing any RT effect. The brightest Balmer emission we have in the spectrum is the [H$\beta$]{} line detected with S/N=6. Figure \[ion2\] superimposes the central [Ly$\alpha$]{} peak (‘0’) and the [H$\beta$]{}, that show compatible widths (being both marginally resolved, see Table \[tab\]). This is fully in line with what was predicted by @behrens14 (see their Figure 7). It is also worth stressing that the detection of the above narrow [Ly$\alpha$]{} features has been possible only thanks to the high spectral resolution ($R>5000$) achievable with X-Shooter (see the case R=1200 in Figure \[ion2\]), underlying the fact that the [Ly$\alpha$]{} line can be a powerful probe of optically thin media up to z=4 (and possibly up to z=6.5 in the case of transparent IGM, e.g., @matthee18). ![The top panel shows the two-dimensional [Ly$\alpha$]{} of [*Ion2*]{} (at resolution R=5300) with indicated the wavelength position of the [\[<span style="font-variant:small-caps;">iii</span>\]$\lambda5007$]{} (5$\sigma$ green contour) and the four peaks, labelled as ‘-2’, ‘-1’, ‘0’ and ‘1’. The peak ‘0’ falls exactly at the [\[<span style="font-variant:small-caps;">iii</span>\]$\lambda5007$]{} redshift. In the bottom panel the [Ly$\alpha$]{}, [\[<span style="font-variant:small-caps;">iii</span>\]$\lambda5007$]{} (rescaled for clarity) and [H$\beta$]{} line profiles are superposed in the velocity domain. The inset shows the zoomed region around the zero velocity ($-400< \Delta v <+400$ ) where the [Ly$\alpha$]{} peak ‘0’ has here been rescaled to match the peak of the [H$\beta$]{} line with the aim to emphasise the consistency among the widths of the lines. The effect of low spectral resolution is also shown (R=1200, blue dotted line). The Gaussian shape representing the spectral resolution in the NIR arm (R=5600) is also shown for comparison with a solid blue line. []{data-label="ion2"}](Figure2.pdf){width="8.0cm"} ![[Ly$\alpha$]{} profiles in the velocity space of all the LyC emitters with triple$-$peaked [Ly$\alpha$]{} emission currently known. The upper two panels have been adapted from @izotov18 and @rivera17. The insets show the two dimensional [Ly$\alpha$]{} spectra reported with their multiple peaks with indicated the emission ’0’ close to zero velocity (dotted line; see also Figure \[ion2\] for a detailed view in the case of [*Ion2*]{}). The relative escape faction ([*fesc*]{}) is also reported in each panel.[]{data-label="lyas"}](Figure3.pdf){width="8.0cm"} $\bullet$ The [<span style="font-variant:small-caps;">iv</span>$\lambda1548,1550$]{}, [<span style="font-variant:small-caps;">iii</span>\]$\lambda1661,1666$]{}, [<span style="font-variant:small-caps;">ii</span>$\lambda1640$]{} and the [<span style="font-variant:small-caps;">iii</span>\]$\lambda\lambda1907,1909$]{} high-ionisation emission lines have been detected (see Figure \[lines\] and Table \[tab\]), with the doublets being well separated. It is worth noting that all the nebular metal lines appear very narrow and possibly not resolved ($\sigma_v < 20$ ), whilst the [<span style="font-variant:small-caps;">ii</span>$\lambda1640$]{} emission, despite a relatively low S/N, is clearly broader. To emphasise such a difference, Figure \[HEII\] shows the portion of the X-Shooter two-dimensional spectrum containing the [<span style="font-variant:small-caps;">iv</span>$\lambda1548,1550$]{} doublet, the [<span style="font-variant:small-caps;">ii</span>$\lambda1640$]{} and [<span style="font-variant:small-caps;">iii</span>\]$\lambda1661,1666$]{} emission lines. The S/N of the [<span style="font-variant:small-caps;">ii</span>$\lambda1640$]{} (3.5) is sufficient to appreciate its broadness, plausibly encompassing a velocity interval up to 500 , marked in Figure \[HEII\] with a segment (see also appendix \[SN\] for more details). It is worth noting that in other cases the [<span style="font-variant:small-caps;">ii</span>$\lambda1640$]{} emission is as narrow as the other high-ionisation metal lines [e.g., @vanz16a; @vanz17c] in which the nebular origin dominates or is better captured and other cases in which both nebular and broader stellar components are measured [@erb10; @senchyna17; @senchyna19]. In the case of [*Ion2*]{} the relative contribution of nebular and stellar components is not measurable. However, the presence of a broad [<span style="font-variant:small-caps;">ii</span>$\lambda1640$]{} emission profile suggests a spectrum dominated by a young stellar population containing hot Wolf-Rayet stars with main-sequence lifetimes less than 5 Myr [@JC19]. $\bullet$ Differently from the previous analysis based on a much shallower Keck/MOSFIRE spectrum in which the optical [\[<span style="font-variant:small-caps;">ii</span>\]$\lambda3727,3729$]{} and [H$\beta$]{} lines were not detected [@debarros16], here we measure a rest-frame equivalent width (EW) of EW([H$\beta$]{}) $\simeq 100$Å, O32 $= 9.18_{-1.32}^{+1.82}$ and [\[<span style="font-variant:small-caps;">iii</span>\]$\lambda5007$]{}/ [H$\beta$]{}= $8.55_{-1.41}^{+1.96}$ (see Table \[tab\]). Such O32 value is in line with the [*necessary*]{} condition of having a large O32 index in LyC leakers [@jaskot13; @izotov18]. It is worth noting that the rest-frame EW of [\[ <span style="font-variant:small-caps;">iii</span>\]$\lambda\lambda4959,5007$]{} is 1300Å, not dissimilar from the strong oxygen emitters found at $z>7$ [e.g., @castellano17; @RB16]. This also suggests a relatively large ionising photon production efficiency ($\xi_{ion}$), defined as the production rate of -ionising photons per unit intrinsic monochromatic UV luminosity ($\xi_{ion} \simeq 25.6$, following @chevallard18).[^1] {width="17.5cm"} While the [Ly$\alpha$]{} profile further confirms [*Ion2*]{} to be a genuine LyC emitter, the detailed geometry of the LyC emission and the origin of the ionising radiation is still unknown [@vanz16b]. Not surprisingly, if the ionised channel and/or the size of the source emitting LyC radiation (namely, the region including O-type stars) is confined within a few tens pc or less (see Sect. \[ion2asymc\]), the WFC3/F336W spatial resolution would be insufficient to resolve the source (1 pix $\simeq 150$ pc). Any further detailed investigation in the rest-frame ultraviolet/optical bands would therefore be postponed to future studies with larger telescopes. Before the advent of E$-$ELT-like telescopes that will provide a spatial resolution lower than 10 mas (corresponding to $\sim $ 75 pc at the redshift of [*Ion2*]{}), the only way to address individual star-forming complexes of a few tens pc require strong gravitational lensing [e.g., @vanz19; @vanz17c; @cava18; @rigby17; @jonhson17]. Even more valuable would be the identification of strongly lensed galaxies showing escaping LyC radiation emerging from some of their internal constituents. This happened recently with the discovery of the [*Sunbust*]{} arc and is the argument of the next section. Before discussing it, it is worth stressing that the requirement of having the simultaneous alignment of the ionised channel, the observer, the presence of shot-lived O-type stars and the transparent IGM along the line of sight implies that the visibility of the LyC radiation from high redshift sources is affected by severe view-angle and l.o.s. effects (e.g., @terlevich17 [@cen15; @wise14]), not to mention the insidious foreground contamination mimicking false LyC radiation [e.g., @vanz10; @vanz12; @siana15]. Altogether, these effects make the detection of LyC$-$galaxies at high redshift still elusive and suggest that a significant fraction of them might be hidden by the aforementioned effects. Moreover, if we require that the source is also strongly magnified by an intervening gravitational lens, then the event would be extremely rare.[^2] The identification of a few LyC$-$galaxies either in non lensed or lensed fields therefore makes the current detections extremely precious, especially if we focus on the spectral similarities among these uncorrelated objects. Discussion ========== To shed more light on the nature of [*Ion2*]{}, key information might be extracted by comparing its X-Shooter spectrum to the strongly magnified LyC emitter at z=2.37, dubbed Sunburst arc [@dahle16; @rivera17; @rivera19]. In this work we highlight the similarities between such systems arguing that what is observed in Sunburst is compatible with what is currently hidden by the limited spatial resolution. {width="17.5cm"} The Sunburst arc ---------------- Sunburst is an exceptionally bright (R-band magnitude $\sim 18$) gravitationally lensed arc in which the galaxy cluster PSZ1-G311 produces a magnification of the order of $\sim 50$ or even larger [@dahle16]. Specifically, four multiple arcs are generated by the galaxy cluster, and in the most magnified ones additional amplification is generated by individual galaxy cluster members. One of the star-forming knots of the arcs has also been detected in the LyC (hereafter dubbed LyC-knot) with the unprecedented record of 12 multiple images [@rivera19] with a measured magnitude in the interval F814W = \[$21.5 - 22$\] and detected with S/N$>30$ for the most magnified ones.[^3] We present for the first time VLT/MUSE observations targeting the Sunburst and based on the DDT programme 297.A-5012(A) (PI. Aghanim). The data were acquired during May-August 2016 with seeing $0.5''-0.8''$ for a total integration of 1.2h (three exposures of 1483s each). The data reduction has been performed as described in @caminha17 [@caminha19], and we refer the reader to those works for details. A stacked MUSE spectrum obtained from seven multiple images is presented in Figure \[clusters\]. In the same figure, some high-quality spectra of local star clusters are shown for comparison (for further details, see the caption of Figure \[clusters\]). The ultraviolet emission of the LyC$-$knot resembles those of local super star clusters in which the signatures of massive stars are clearly imprinted in the spectrum, like the prominent P-Cygni profiles of the [<span style="font-variant:small-caps;">iv</span>$\lambda1548,1550$]{} doublet and a broad [<span style="font-variant:small-caps;">ii</span>$\lambda1640$]{} ascribed to the presence of Wolf-Rayet stars. The similarity with the ultraviolet spectra of a few well studied local young massive clusters is remarkable, namely R136 [@crowther16], II Z40 [@leitherer18] and cluster \#5 of NG5253 (@smith16, see also @calzetti15), as well as the analogy with the nearby star forming regions collected by @senchyna17 [@senchyna19] and showing high ionisation metal lines and in some cases broad [<span style="font-variant:small-caps;">ii</span>$\lambda1640$]{} emission of low metallicity massive stars. The same spectral features, though at slightly lower S/N ratios, have been identified and accurately modelled by @JC19, providing a stellar age of $3.0\pm0.1$ Myr and subsolar stellar metallicity Z = $0.60\pm0.05$ Z$_{\odot}$, with an inferred dust extinction E(B-V) $\simeq$ 0.15. While these are very precious quantities and independent from the lensing magnification, it is now natural to try addressing intrinsic quantities such as the stellar mass, the luminosity and the physical size of such LyC emitter. This will be the argument of the next section. The Sunburst LyC-knot as a possible gravitationally bound system ---------------------------------------------------------------- We express the most relevant physical quantities, namely the effective radius, the stellar mass, and the stellar mass surface density ($\Sigma^{\star}$), as a function of magnification $\mu_{TOT}$, adopting $R_e = 2.0$ pix, e.g., 60 mas along the tangential direction as derived by performing [*GALFIT*]{} fitting (see Appendix \[minmag\] and Figure \[profiles\]). The stellar mass is estimated assuming an instantaneous burst with the aforementioned age of 3.0 Myr and sub-solar metallicity Z = 0.6 Z$_{\odot}$, and adopting [*Starburst99*]{} models [@leitherer14] with a Salpeter initial stellar mass function (IMF, $\alpha = 2.3$), including stars with masses in the range \[1-100\] . Similarly, we derive the stellar mass using a top-heavy initial stellar mass function with a slope $\alpha = 1.6$, following @tereza17, with masses in the range \[0.1-120\] . The two IMFs should embrace two extreme cases with the aim to provide a lower and an upper limit to the stellar mass. As shown in Figure \[mass\], an effective radius ($R_e$) smaller than $20$ pc is found if $\mu_{TOT} > 25$, and it decreases below 9 pc if $\mu_{TOT} > 50$. The stellar mass ranges between $10^{6} - 10^{7}$   depending on the IMF and magnification, with a $\Sigma^{\star}$ that enters the regime of the densest objects known (e.g., the globular clusters, e.g., @hopkins10) or similar to the values of young massive star clusters [@bastian06; @ostlin07; @bastian13]. Combining $R_e$ with the age of 3.0 Myr, and the stellar mass, we can infer the dynamical age as $\Pi =$ Age/ T$_{cr}$ [where T$_{cr}$ is the crossing time: 10($R_e^{3}$/GM)$^{0.5}$, @gieles11]. Interestingly, as highlighted in Figure \[mass\], the system enters the regime of a gravitationally bound object ($\Pi > 1$) if $\mu_{TOT}>25(50)$ in the case of Salpeter(top-heavy) IMF. Such magnification values are within the expected magnification regime [e.g., @dahle16]. In particular, a [*minimum*]{} model-independent estimate of the magnification of 20 is derived and discussed in the appendix \[minmag\], based on empirical geometrical constraints. Therefore, the LyC$-$knot might be the first example of a gravitationally-bound star cluster discovered at cosmological distance. Is also the LyC$-$galaxy [*Ion2*]{} powered by star clusters? {#ion2asymc} ------------------------------------------------------------- The discovery of a very likely gravitationally bound star cluster at z=2.37 leaking LyC radiation is intriguing because it would imply that the contribution by such systems to the meta-galactic ionising background is substantial, if not dominant, depending on the UV luminosity function of such objects. Observationally, Sunburst$-$like objects in which the LyC leakage emerges from a single massive star cluster of a few pc cannot be spatially resolved even in moderately lensed fields, e.g. $\mu < 20$. Compact SF$-$clumps at high redshift, either spatially resolved or not, may be dominated by single young massive star clusters [e.g., @zanella15; @jonhson17; @rigby17]. Any seeing-limited spectrum would be the luminosity-weighted average of multiple unresolved star$-$forming complexes, as for the case of [*Ion2*]{} in which structures smaller than 200 pc cannot be resolved. However, the similitude among the spectral properties of [*Ion2*]{} and Sunburst is intriguing and might offer new clues, beyond the limitation due to the spatial resolution. The ionised channels traced by the triple$-$peaked [Ly$\alpha$]{} profiles $-$ especially the narrow peak at systemic velocity $-$ of [*Ion2*]{}, [*Ion3*]{} and Sunburst (including the local system of @izotov18) might suggest a common origin related to the presence of young massive star clusters and/or dense star-forming regions. The HST imaging of [*Ion2*]{} shows a quite nucleated morphology [the LyC emission is spatially unresolved, @vanz16b], as well as the triple$-$peaked [Ly$\alpha$]{} object of @izotov18 that shows the highest star formation rate surface density ($> 500$  yr$^{-1}$kpc$^{-2}$) in their sample. Currently, the Sunburst LyC-knot seems to be the densest stellar LyC leaker with also emergent [Ly$\alpha$]{} at systemic velocity. In addition, the broad [<span style="font-variant:small-caps;">ii</span>$\lambda1640$]{} emission observed in [*Ion2*]{} (possibly with FWHM $> 400$ , see appendix \[scan\]) and the well detected [<span style="font-variant:small-caps;">v</span>$\lambda1240$]{} P-Cygni profile of [*Ion3*]{} [@vanz18] suggest a radiation leakage through one or more channels carved by massive stars promoted by their energetic feedback might be in place, as observed in the Sunburst. ![The stellar mass, stellar mass surface density ($\Sigma^{\star}$) and the effective radius ($R_e$, on the right Y-axis) of the Sunburst LyC-knot are calculated as a function of the total magnification, $\mu_{TOT}$. Red and blue curves correspond to top-heavy and Salpeter IMF, respectively. The LyC-knot becomes compatible with a gravitationally bound star cluster if $\mu_{TOT} > 25(50)$. This regime ($\Pi > 1$) is highlighted with the thick lines, where $\Sigma^{\star}$ approaches the values measured in the densest objects known, like globular clusters and young massive clusters. The thick black horizontal arrow in the bottom marks the minimum magnification, $\mu_{TOT} > 20$, we estimate in appendix \[minmag\].[]{data-label="mass"}](Figure5.pdf){width="8.5cm"} Conluding remarks ================= In this paper we have presented new VLT/X-Shooter observations of the LyC emitting galaxy [*Ion2*]{} and VLT/MUSE spectrum of a strongly lensed LyC emitter, dubbed Sunburst. The results can be summarised as follows: [*Ion2*]{}: the spectral resolution and wavelength coverage provided by X-Shooter have improved the previous analysis presented in @debarros16, including the detection of new spectral features. First, a multi-peaked [Ly$\alpha$]{} profile is evident from the new spectrum, showing a clear emission at the systemic redshift. Second, several high ionisation ultraviolet nebular narrow lines (FWHM $<$ 50 , including [<span style="font-variant:small-caps;">iv</span>$\lambda1548,1550$]{}, [<span style="font-variant:small-caps;">iii</span>\]$\lambda1661,1666$]{}, [<span style="font-variant:small-caps;">iii</span>\]$\lambda\lambda1907,1909$]{}) have been detected for the first time, some of which with well separated doublets. Only the [<span style="font-variant:small-caps;">ii</span>$\lambda1640$]{} is detected as a broad emission (FWHM $>$ 400 ) and ascribed to the presence of Wolf-Rayet stars. Third, a large value for the O32 index of $9.18_{-1.32}^{+1.82}$ has been derived, together with the large equivalent width of 1300Å rest-frame of [\[ <span style="font-variant:small-caps;">iii</span>\]$\lambda\lambda4959,5007$]{}, as found in other LyC leakers [@jaskot13; @izotov18] and typical of systems with a large ionising photon production efficiency [e.g., @chevallard18]. [*Sunburst*]{}: we inferred for the first time the stellar mass (a few $10^{6}$ ), luminosity ($M_{UV} > -19$) and size ($R_e < 20$ pc) of the LyC-knot of the Sunburst arc, that coupled with the young age of 3 Myr [@JC19] provides constraints on its dynamical age, suggesting that the LyC-knot is a gravitationally$-$bound young massive stellar cluster at cosmological distance, whose ultraviolet spectrum is also fully comparable to those of local young clusters (Figure \[clusters\]). In addition, as the Figure \[mass\] shows, the stellar mass surface density is intriguingly large if the magnification factor exceeds 50, approaching the values observed in the densest objects, such as globular clusters and local young massive clusters [@hopkins10; @bastian06; @ostlin07; @bastian13]. Remarkably, Sunburst might also be considered a forming globular cluster, caught when the Universe was 2.7 Gyr old. This will be investigated in a future work. The LyC-knot of the Sunburst arc might very well represent the Rosetta stone of stellar ionisation at high redshift and it is an unprecedented discovery in its own right for two reasons: (1) it is a unique laboratory where the escaping LyC from a high redshift stellar system can be investigated in detail and (2) without any lensing effect, the LyC-knot (the star cluster) would have appeared like a non-spatially resolved LyC$-$emitter lying somewhere within its hosting galaxy, e.g., like [*Ion2*]{}. Interesting enough, the star cluster formation efficiency, namely the star formation occurring in gravitationally bound star clusters, increases with redshift [e.g., @pfeffer18], possibly reaching values higher than $30$% at $z>6$. This might suggest that star clusters could have played a significant role during reionisation [@ricotti02; @ricotti16], especially if the LyC leakage is more efficient for that population. The idea that such young massive star clusters (or a fraction of them) were also globular cluster precursors is currently matter of investigation (e.g., @vanz19 [@renzini17; @pozzetti19; @bouwens17; @pfeffer18; @elme18; @kruijssen19; @RC19; @calura15; @calura19; @li19]). While the direct detection of LyC radiation at $z>3$ is challenging and requires that special conditions are realised in the source, with current sensitivity, the effect that the transverse LyC leakage has on the surrounding medium might be easily detectable. Objects like [*Ion2/Ion3*]{} or Sunburst having transverse leakage of LyC radiation could induce spatially offset [Ly$\alpha$]{} or Balmer series fluorescence [e.g., @mas-ribas17]. Spatially offset [Ly$\alpha$]{} emission/nebulae routinely detected with integral field spectrographs (like MUSE) might represent a viable tool to search for possible local escaping ionising radiation around star-forming galaxies [e.g., @vanz17a; @vanz17c; @wisotzki18; @gallego18]. Finally, the prospects for future investigations of star-formation at very small scales $-$ down to single star clusters $-$ at cosmological distance appear very promising. In particular objects like Sunburst LyC-knot stretched by magnification factors larger than 30 will be probed down to FWHM(or pixel-scale) of 6(2) and 3(0.4) pc by VLT/MAVIS and ELT MAORY-MICADO, respectively. These two MCAO$-$assisted[^4] instruments will be also complementary in terms of wavelength coverage, probing the ultraviolet and optical rest-frame wavelengths. Integral field spectroscopy at VLT or ELT will also probe the spatial distribution of nebular high ionisation lines, as a signature of possible stellar mass segregation in star complexes and providing maps at pc scale opening for two-dimensional studies of feedback mechanisms and star formation processes [e.g., @james16], at cosmological distances. Line/$\lambda_{vacuum}$ Flux($\frac{S}{N}$)(FWHM)(EW) Redshift($1\sigma$) --------------------------------------------------------------------------------------- ------------------------------- --------------------- [Ly$\alpha$]{}(-2) 1215.7 1.68(8.0)($\sim$141)(5.7) 3.2044(3) [Ly$\alpha$]{}(-1) 1215.7 5.28(29.0)(146)(18.0) 3.2096(1) [Ly$\alpha$]{}(0) 1215.7 0.96(10.5)($<56$)(3.3) 3.2121(2) [Ly$\alpha$]{}(1) 1215.7 20.25(110)(298)(69.1) 3.2164(1) [Ly$\alpha$]{}(total) 28.16($-$)($-$)(97.6) [<span style="font-variant:small-caps;">v</span>]{} $\lambda 1238.82 $ $<$0.12(1.0)($-$)($<0.5$) (3.2121)fixed [<span style="font-variant:small-caps;">v</span>]{} $\lambda 1242.80$ $<$0.12(1.0)($-$)($<0.5$) (3.2121)fixed [<span style="font-variant:small-caps;">iv</span>]{} $\lambda 1548.20 $ 0.28(5.2)(Narrow)(1.5) (3.2121)fixed [<span style="font-variant:small-caps;">iv</span>]{} $\lambda 1550.78 $ 0.20(4.8)(Narrow)(1.1) (3.2121)fixed [\[<span style="font-variant:small-caps;">ii</span>\]]{} $\lambda 1640.42$ 0.45(5.0)(Broad)(2.8) (3.2121)fixed [\[<span style="font-variant:small-caps;">iii</span>\]]{} $\lambda 1660.81 $ 0.40(3.5)(Narrow)(2.5) (3.2121)fixed [\[<span style="font-variant:small-caps;">iii</span>\]]{} $\lambda 1666.15 $ 0.65(5.4)(Narrow)(4.2) (3.2121)fixed [\[<span style="font-variant:small-caps;">iii</span>\]]{} $\lambda 1882.65 $ 0.30(2.0)(Narrow)(2.4) (3.2121)fixed [\[<span style="font-variant:small-caps;">iii</span>\]]{} $\lambda 1892.03 $ 0.15(1.5)(Narrow)(1.2) (3.2121)fixed [<span style="font-variant:small-caps;">iii</span>\]]{} $\lambda 1906.68 $ 0.44(6.7)(Narrow)(4.1) 3.2127(5) [<span style="font-variant:small-caps;">iii</span>\]]{} $\lambda 1908.73 $ 0.35(4.9)(Narrow)(3.2) (3.2121)fixed [\[<span style="font-variant:small-caps;">ii</span>\]]{} $\lambda 3727-3729$ 3.5(6.0)($-$)(81) 3.2122(3) NeIII\] $\lambda 3869.81$ 1.8(4.0)($-$)(45) 3.2119(4) NeIII\] $\lambda 3968.53$ 1.2(1.5)($-$)(32) (3.2121)fixed H$\delta$ $\lambda 4102.92$ 1.1(2.0)($-$)(31) (3.2121)fixed [H$\beta$]{} $\lambda 4862.69$ 2.9(6.0)($<60$)(114) 3.2120(3) [\[<span style="font-variant:small-caps;">iii</span>\]]{} $\lambda 4960.30$ 7.4(22.0)($-$)(304) 3.2121(2) [\[<span style="font-variant:small-caps;">iii</span>\]]{} $\lambda 5008.24$ 24.8(65.0)(146)(1040) 3.2121(1) : The most relevant atomic transitions of [*Ion2*]{} are reported, showing the corresponding zoomed regions on the two-dimensional spectrum of Figure \[lines\]. The S/N ratios indicate the reliability of the lines (see also appendix \[SN\]). 1-$\sigma$ upper limits on the line fluxes are reported in the case of non detections. The [\[<span style="font-variant:small-caps;">v</span>\]$\lambda3426$]{}, [H$\gamma$]{} and the [\[<span style="font-variant:small-caps;">iii</span>\]$\lambda4364$]{} are not reported as they lie on the atmospheric absorption bands. Line fluxes are reported in units of $10^{-17}$  (no slit losses are considered) and the FWHM is expressed in ; the rest-frame equivalent width (EW) is reported in Å and calculated starting from the line fluxes and continuum derived from CANDELS photometry, adopting magnitudes $\simeq 24.45$ and $\simeq 24.1$ in the optical and near infrared arms, respectively. The comment “Narrow” means the line is not resolved, while “Broad” stands for resolved but a precise measurement is not feasible. \[tab\] Acknowledgments {#acknowledgments .unnumbered} =============== We thank the referees for providing detailed comments and suggestions about the reorganisation of the manuscript. We thank C. Leitherer, L. Smith and P. Crowther for providing the ultraviolet spectra of the star clusters shown in Figure \[clusters\]. We thank A. Renzini, A. Adamo for stimulating discussions about the possible origin of the LyC-knot. EV also thank J. Chisholm and E. Rivera-Thorsen for stimulating discussions about the Sunburst system. EV thanks T. Jerabkova for fruitful interaction about the IMF used in this work. EV also thank M. Gronke for useful discussions about the [Ly$\alpha$]{} profile. AM acknowledges funding from the INAF PRIN-SKA 2017 programme 1.05.01.88.04. KC acknowledges funding from the European Research Council through the award of the Consolidator Grant ID 681627-BUILDUP. We also acknowledge funding from the INAF for “interventi aggiuntivi a sostegno della ricerca di main-stream dell’INAF”. Bastian, N., Saglia, R. P., Goudfrooij, P., et al. 2006, , 448, 881 Bastian, N., Schweizer, F., Goudfrooij, P., Larsen, S. S., & Kissler-Patig, M. 2013, , 431, 1252 Behrens, C., Dijkstra, M., & Niemeyer, J. C. 2014, , 563,77 Bian, F., Fan, X., McGreer, I., Cai, Z., & Jiang, L. 2017, , 837, L12 Bik, A., [Ö]{}stlin, G., Menacho, V., et al. 2018, , 619, A131 Bouwens, R. J., Illingworth, G. D., Oesch, P. A., et al. 2017, arXiv:1711.02090 Calura, F., Few, C. G., Romano, D., & D’Ercole, A. 2015, , 814, L14 Calura, F., D’Ercole, A., Vesperini, E., et al. 2019, arXiv e-prints, arXiv:1906.09137 Calzetti, D., Johnson, K. E., Adamo, A., et al. 2015, , 811, 75 Caminha, G. B., Grillo, C., Rosati, P., et al. 2017, , 600, A90 Caminha, G. B., Rosati, P., Grillo, C., et al. 2019, arXiv:1903.05103 Cen, R., & Kimm, T. 2015, , 801, L25 Castellano, M., Pentericci, L., Fontana, A., et al. 2017, , 839, 73 Cava, A., Schaerer, D., Richard, J., et al. 2018, Nature Astronomy, 2, 76 Chevallard, J., Charlot, S., Senchyna, P., et al. 2018, , 479, 3264 Chisholm, J., Gazagnes, S., Schaerer, D., et al. 2018, , 616, A30 Chisholm, J., Rigby, J. R., Bayliss, M., et al. 2019, arXiv:1905.04314 Crowther, P. A., et al. 2016, , 458, 624 Dahle, H., Aghanim, N., Guennou, L., et al. 2016, , 590, L4 de Barros, S., et al. 2016, , 585, A51 Elmegreen, B. G. 2018, , 869, 119 Erb, D. K., Pettini, M., Shapley, A. E., et al. 2010, , 719, 1168 Fletcher, T. J., Robertson, B. E., Nakajima, K., et al. 2018, arXiv:1806.01741 Gallego, S. G., Cantalupo, S., Lilly, S., et al. 2018, , 475, 3854 Gazagnes, S., Chisholm, J., Schaerer, D., et al. 2018, , 616, A29 Gieles, M., & Portegies Zwart, S. F. 2011, , 410, L6 Grazian, A., Giallongo, E., Paris, D., et al. 2017, , 602, A18 Gronke, M., Bull, P., & Dijkstra, M. 2015, , 812, 123 Heckman, T. M., et al. 2011, , 730, 5. Herenz, E. C., Hayes, M., Papaderos, P., et al. 2017, , 606, L11 Hopkins, P. F., Murray, N., Quataert, E., & Thompson, T. A. 2010, , 401, L19 Li, H., & Gnedin, O. Y. 2019, , 486, 4030 Izotov, Y. I., Worseck, G., Schaerer, D., et al. 2018, , 478, 4851 James, B. L., Auger, M., Aloisi, A., Calzetti, D., & Kewley, L. 2016, , 816, 40 Japelj, J., Vanzella, E., Fontanot, F., et al. 2017, , 468, 389 Jaskot, A. E., & Oey, M. S. 2013, , 766, 91 Je[ř]{}[á]{}bkov[á]{}, T., Kroupa, P., Dabringhausen, J., Hilker, M., & Bekki, K. 2017, , 608, A53 Johnson, T. L., Rigby, J. R., Sharon, K., et al. 2017, , 843, L21 Kehrig, C., V[í]{}lchez, J. M., Guerrero, M. A., et al. 2018, , 480, 1081 Kneib, J.-P., & Natarajan, P. 2011, , 19, 47 Kruijssen, J. M. D. 2019, , 486, L20 Leitherer, C., Ekstr[ö]{}m, S., Meynet, G., et al. 2014, , 212, 14 Leitherer, C., Byler, N., Lee, J. C., & Levesque, E. M. 2018, , 865, 55 Marchi, F., Pentericci, L., Guaita, L., et al. 2018, , 614, A11 Mas-Ribas, L., Hennawi, J. F., Dijkstra, M., et al. 2017, , 846, 11 Matthee, J., Sobral, D., Gronke, M., et al. 2018, , 619, A136 McKinney, J. H., Jaskot, A. E., Oey, M. S., et al. 2019, arXiv:1902.08204 Micheva, G., Oey, M. S., Jaskot, A. E., & James, B. L. 2017, , 845, 165 Micheva, G., Herenz, E. C., Roth, M. M., [Ö]{}stlin, G., & Girichidis, P. 2019, arXiv:1902.03952 stlin, G., Cumming, R. J., & Bergvall, N. 2007, , 461, 471 Pfeffer, J., Kruijssen, J. M. D., Crain, R. A., & Bastian, N. 2018, , 475, 4309 Pozzetti, L., Maraston, C., & Renzini, A. 2019, , Reddy, N. A., Steidel, C. C., Pettini, M., et al. 2016, , 828, 107 Reddy, N. A., Shapley, A. E., Sanders, R. L., et al. 2018, , 869, 92 Reina-Campos, M., Kruijssen, J. M. D., Pfeffer, J. L., et al. 2019, , 486, 5838 Renzini, A. 2017, , 469, L63 Ricotti, M. 2002, , 336, L33 Ricotti, M., Parry, O. H., & Gnedin, N. Y. 2016, , 831, 204 Rigby, J. R., Johnson, T. L., Sharon, K., et al. 2017, , 843, 79 Rivera-Thorsen, T. E., et al. 2017, arXiv:1710.09482 Rivera-Thorsen, T. E., Dahle, H., Chisholm, J., et al. 2019, arXiv:1904.08186 Roberts-Borsani, G. W., Bouwens, R. J., Oesch, P. A., et al. 2016, , 823, 143 Schaerer, D., Hayes, M., Verhamme, A., & Teyssier, R. 2011, , 531, A12 Schaerer, D., Izotov, Y. I., Verhamme, A., et al. 2016,, 591,8 Senchyna, P., Stark, D. P., Vidal-Garc[í]{}a, A., et al. 2017, , 472, 2608 Senchyna, P., Stark, D. P., Chevallard, J., et al. 2019, arXiv:1904.01615 Shapley, A., Steidel, C. C., Strom, A. L., et al. 2016, , 826,24 Siana, B., et al. 2015, , 804, 17. Smith, L. J., Crowther, P. A., Calzetti, D., & Sidoli, F. 2016, , 823, 38 Steidel, C. C., Bogosavljevi[ć]{}, M., Shapley, A. E., et al. 2018, , 869, 123 Trebitsch, M., Blaizot, J., Rosdahl, J., Devriendt, J., & Slyz, A. 2017, , 470, 224 Vanzella, E., Siana, B., Cristiani, S., & Nonino, M. 2010, , 404, 1672 Vanzella, E., et al. 2012, , 751, 70 Vanzella, E., et al. 2015, , 576, A116 Vanzella, E., et al. 2016b, , 825, 41 Vanzella, E., De Barros, S., Cupani, G., et al. 2016a, , 821, L27 Vanzella, E., et al. 2017b, , 467, 4304 Vanzella, E., et al. 2017c, , 842, 47 Vanzella, E., et al. 2017a, , 465, 3803 Vanzella, E., et al. 2018, , 476, L15 Vanzella, E., et al. 2019, , 483, 3618 Verhamme, A., Orlitov[á]{}, I., Schaerer, D., et al. 2017, , 597,13 Wise, J. H., Demchenko, V. G., Halicek, M. T., et al. 2014, , 442, 2560 Wisotzki, L., Bacon, R., Brinchmann, J., et al. 2018, , 562, 229 Zanella, A., Daddi, E., Le Floc’h, E., et al. 2015, , 521, 54 An empirical minimum magnification for the Sunburst LyC-knot {#minmag} ============================================================ The sub-structures and star-forming knots present in the Sunburst arcs allow us to set empirical constraints on the minimum (average) magnification among the recognised multiple patterns. Figure \[fullsystem\] shows the four arcs labeled I, II, II and IV, within which 12 multiple images of the LyC-knot (indicated with ‘A’) have been discovered by @rivera19 and are marked with increasing numbers 1-12. In particular the least magnified arcs provide the less distorted version of the high-z SF complex. Starting from the smallest arc III the knots A, B, C are identified in the counter arc IV, with the inverse order (mirroring) of the SF knots as expected in strongly lensed multiple images. In this case the identification is further facilitated by the fact that the knot A has the distinctive LyC feature and the image C is the most elongated among the three objects. Such features must follow the aforementiond flipped behaviour, accordingly to the parity of the images introduced by the strong lensing [@kneib11]. Other two triplets A, B, C appear in the arc II where the mirrored groups are circled with green dotted ellipses. The triplet separated by $3.72''$ is 10 times more elongated than the triplet in the arc III, that means the average tangential magnification for the widest arc is larger than 10, being the magnification of arc III larger than 1. The situation of arc I is different because not all images (A,B,C) are present, i.e. only a part of the source is multiply lensed in this merging arc. However, six multiple images of knot A (the LyC emitter) are clearly identified (images 1-6 of Figure \[fullsystem\]). The proximity of some of the images to the critical lines, like images 2 and 3, suggest the magnification is large [see, e.g., Figure 4 of @rivera19]. In particular the measured flux ratio between images 2 (or equivalently 3) and 8 is $\simeq 2$ [@rivera19], suggesting images 2,3 have a minimum magnification of $\sim20$ ($2 \times 10$). This happens for the brightest image 10, that is 2.2 times brighter than image 8. Until this point we did not invoke any lens model. The true magnification is higher since we are assuming arc IV is not magnified ($\mu=1$). However, arc IV is subjected by strong lensing as well (being one of the multiple images of the system) and its magnification is certainly higher than 1. For example, assuming a magnification 2 for arc IV, following the aforementioned considerations the magnification for images 2,3 jumps to values of 40. This provides an average magnification, and as discussed above the proximity to the critical lines suggests the lensing amplification at the position of, e.g., images 2-3, could be be much higher, even above 50 as discussed by @dahle16. A careful modelling of the lens, including the aforementioned empirical constraints, will shed more light on this. However, the minimum magnification of 20 derived above is already relevant in our study, implying an effective radius smaller than 20 pc and a stellar mass in the range $5\times10^{6} - 2.5 \times 10^{7}$ , depending on the IMF (see Figure \[mass\]), with the system entering the range of massive, gravitationally-bound, star clusters in the case of Salpeter IMF. It is worth noting the presence of a bright and point-like object with F814W = 22.02 and at the same redshift of the Sunburst (marked with ‘Tr’ and an arrow in Figure \[fullsystem\]). Such an object is presumably a transient for two reasons: (1) it is not identified in any of the other arcs and (2) it shows unique spectral properties in the MUSE spectrum not observed in any of the other knots populating the rest of the arcs (the ongoing X-Shooter programme[^5] will investigate ‘Tr’ and the results presented elsewhere). The characterisation of such a transient (like the absolute magnitude), will provide a unique constraint for the lens model. Figure \[profiles\] shows the two multiple images “2” and “3”, each one located $0.26''$ away from the critical line, that necessarily falls between them. The giant arc$-$like shape also implies the magnification is mainly tangential, such that the total magnification is close to the tangential one, $\mu_{TOT} \simeq \mu_{tan}$ (see discussion in @vanz17b). Despite the knots lie in a region with a potentially steep tangential magnification gradient, the light profiles along the same direction are symmetric (see Figure \[profiles\], middle panel), strongly supporting the fact that the object is intrinsically compact (see also discussion by @vanz16a on another similar case). As shown in the middle panel of the same figure, the LyC$-$knot is marginally resolved in the HST/F814W image, close to the resolution in that band, in which the PSF FWHM is $\simeq 0.13''$ (though it is not a point-like source). Indeed, specific [Galfit]{}-based deconvolution analysis (as similarly performed in @vanz16a [@vanz17b]) produces effective radii ($R_e$) of the order of $1.0-2.0$ pixel, for a Sersic index in the range [*n*]{} = $0.5 - 5$. It is worth noting, however, that a large [*n*]{} and $R_e$ would produce non symmetric tangential profiles, as mentioned above. This will be fully investigated with dedicated simulations of all the knots and the emerging light profiles by placing objects with known structural parameters in the source plane close to the caustics (see appendix A of @vanz17b), once the lens model will be developed. {width="17.5cm"} ![The top panel shows the F814W zoomed image of the multiple images “2” and “3” with the rectangular aperture ($1.6'' \times 0.13''$) used to calculate the profiles reported in the middle panel, in which the 50% area along the tangential direction is highlighted (shaded regions), and the FWHM ($0.13''$) of the F814W-band is superimposed with a blue line. The expected crude position of the critical line is also indicated. The symmetry of the profiles (also outlined with the white contours in the F814W images, in the top panel) despite the vicinity of the knots to the critical line ($0.26''$) suggests the objects are quite compact. In the bottom the [Galfit]{} solution in the same band is also shown, with Sersic index 4.0 and effective radius $\lesssim 2$ pixel (see text for details).[]{data-label="profiles"}](PROFILES.pdf){width="8.5cm"} Measuring the significance of the emission lines detected in the X-Shooter spectrum {#scan} =================================================================================== The statistical significance of the emission lines reported in Table \[tab\] has been calculated performing spectral scans over the reduced spectra in the UVB, VIS and NIR arms. In particular, we describe here the case of the VIS arm, in which we explore the presence and broadness of the [<span style="font-variant:small-caps;">ii</span>$\lambda1640$]{} line. A spectral scan over the two-dimensional reduced spectrum is performed by using a widow with spatial scale [*ds*]{} and velocity [*dv*]{}. [*ds*]{} has been fixed to $0.8''$, slightly larger than the mean seeing during the observations (the spatial scale in the VIS arm is $0.16''$/pix). Thee velocity widths have been used, [*dv*]{} = 200, 450 and 900 . Small(large) [*dv*]{} captures small(large) spectral features. The scan has been performed pixel by pixel in the wavelength direction (0.2Å/pix), while the spatial direction windows do not overlap to each other (they are independent). Moreover, we distinguished between the position of the target (that lies at a fixed position) and the rest of the slit. After excluding the target and the edges of the slit we end up with 10 independent windows corresponding to each wavelength position. Figure \[SN\] shows the results of this exercise. The black line in each panel shows the scan performed at the position of the target. The scan with small [*dv*]{} recovers the narrow features and doublets discussed in the text, like the [<span style="font-variant:small-caps;">iv</span>$\lambda1548,1550$]{}, [<span style="font-variant:small-caps;">iii</span>\]$\lambda1661,1666$]{}, [<span style="font-variant:small-caps;">ii</span>$\lambda1640$]{} and the [<span style="font-variant:small-caps;">iii</span>\]$\lambda\lambda1907,1909$]{} lines. The results of the same scan avoiding the target is shown with blue dots which are more distributed around the zero value and follow the spectral pattern of the sky emission (shown with red line and rescaled as a guidance). The thick blue line represents the standard deviation of the cloud of blue points calculated over the 10 windows available at each wavelength position. The black and the thick blue line are the signal and the error, respectively. The S/N ratios reported in Table \[tab\] are inferred from this analysis. The same spectral scan has been computed by enlarging [*dv*]{}. Figure \[SN\] clearly shows that as [*dv*]{} increases the tiny spectral features disappear, hidden by the continuum fluctuation. The continuum shows a small (but wide) dip at $\lambda \sim 1810$Å, that corresponds to 7620Å at the observed frame, due to the well known sky absorption band (7600-7640Å). Two main spectral features associated to [*Ion2*]{} stand out from the continuum: the [<span style="font-variant:small-caps;">iii</span>\]$\lambda\lambda1907,1909$]{} doublet that is detected as a single entity and the broad emission identified at the position of [<span style="font-variant:small-caps;">ii</span>$\lambda1640$]{}, with a S/N $\simeq 4$. {width="17.5cm"} [^1]: Note that the [H$\alpha$]{} line is not accessible and the [H$\beta$]{} is detected at low S/N and affected by dust extinction and possibly damped by the LyC leakage. [^2]: The lensing cross section for events with magnification $\mu$ exceeding the threshold $\mu_0$ decreases rapidly with the square of the magnification itself: $\sigma_{lens} (> \mu_0) \sim \mu_{0}^{-2}$. [^3]: The magnitudes reported by @rivera19 are about half a magnitude fainter than our estimates, while the magnitude contrast among the multiple images is fully consistent with their estimates. As shown in the appendix \[minmag\] this difference is mainly due to the small aperture they used ($0.12''$ diameter), motivated by the need to measure the ratio between the ionising and non-ionising fluxes of a non-spatially resolved source. [^4]: MCAO = Multi Conjugate Adaptive Optics. [^5]: VLT/X-Shooter, P103.A-0688(A-C), PI Vanzella.

ArXiv

--- abstract: | We compute the spectral index for scalar perturbations generated in a primordial inflationary model. In this model, the transition of the inflationary phase to the radiative era is achieved through the decay of the cosmological term leading a second order phase transition and the characteristics of the model allow to implement a set of initial conditions where the perturbations display a thermal spectrum when they emerge from the horizon. The obtained value for the spectral index is equal to 2, a result that depends very weakly on the various parameters of the model and on the initial conditions used. PACS number(s): 04.20.Cv., 04.20.Me --- [UMH-MG 01-01]{} [**Scalar perturbations in a\ primordial inflationary scenario\ **]{} Júlio C. Fabris[^1] Departamento de Física, Universidade Federal do Espírito Santo, 29060-900, Vitória, Espírito Santo, Brazil and\ Philippe Spindel[^2] Mécanique et Gravitation, Université de Mons-Hainaut, 7000 Mons, Belgium Introduction ============ The big-bang standard model, based on the Friedmann-Lemaître-Robertson-Walker (FLRW) solutions of Einstein’s equations [@lfrw], suffers from some drawbacks which leaded cosmologists to propose the so-called inflationary models [@GB]. The common element to all these models is an exponential expansion phase of the size of the Universe prior to the beginning of the standard scenario that describes the adiabatic era extending from temperatures of order of $10^{-4}$ at Grand Unification era (GU) until $10^{-32}$, today’s temperature, in Planck’s units (which are the units that will be used throughout all this work). From a kinetic point of view, the exponential expansion phase corresponds to a de Sitter geometry [@HE] which later on will transform itself into a FLRW geometry. From a physical point of view, this scenario raises (at least) two questions. What is the nature of the effective cosmological constant during the de Sitter phase of the history of the Universe? What is the mechanism which leads to the end of this phase? Among the different inflationary models, the first one to be proposed [@BEG] looks specially interesting. It consists of a primordial inflationary model describing the birth and grow of the Universe as the result of a cooperative mechanism of zero total energy, generated by a quantum fluctuation occurring in a flat background. The idea that a quantum fluctuation may be the initial seed for the birth of the Universe has first been proposed by Tryon in 1973 [@TR]. It is because the total hamiltonian of gravity coupled to matter is zero (an important consequence of the invariance of general relativity with respect to time reparametrisation) and the energy contribution of the gravitational conformal modes is negative that it is possible to obtain non-trivial semi-classical solutions of Einstein equations where gravity and matter emerges from an empty flat space in a cooperative process. In the model proposed by Brout et al. [@BEG], matter is described phenomenologically through the action of a scalar field non-minimally coupled to gravity. A consistent asymptotic solution of the field equations [@BES; @BEFGNTS] describes a de Sitter geometry where the role of the cosmological constant is played by the (constant) density of the created matter which has, due to the energy conservation law, the characteristic equation of state of a cosmological constant $\rho_\Lambda + p_\Lambda = 0$. On the other hand, the consistency of the model implies that the (effective) cosmological constant $\Lambda \equiv 8\pi\rho_\Lambda$ is a free parameter of the obtained solutions. It must not necessarily be of order one, what justify a posteriori the semi-classical treatment employed. However, the mass of the scalar quanta fields is very large [@BES]: $M \simeq 6\sqrt{\pi}$, this value being almost insensible to the value of the cosmological constant [@Sp] even when the vacuum polarization effects are taken into account [@SV]; these effects play no significative role in the domain of validity of the model ($\Lambda << 1$). On the other hand, the high value of the mass of the quanta is the signal that gravitational strong coupling must play an important role in the physics of this problem[@En]. The scalar field used in the model must be interpreted as a phenomenological description of complex objects, like black holes [@CE]. This interpretation, by the way, is confirmed by the analysis of the renormalization of the gravitational constant [@En]. Indeed, the renormalized gravitational constant becomes singular [@DC] exactly for the critical value of the mass field $M = 6\sqrt{\pi}$, what is the origin of the analytical solution [@Sp2] of the semi-classical equations which describes, for this special value, the transition from a flat space to a singularity. Since the cosmological “constant” in this model is described a gas of black holes, their subsequent evaporation through the Hawking radiation mechanism may imply a time-decreasing of this effective cosmological “constant”, generating hot radiation in the Universe. For the moment, there is no consensus about how the inflationary phase ends. However, it is possible to distinguish, from the phenomenological point of view, essentially two types of transition to an adiabatic era: a first order phase transition, employed in the so-called old and new inflationary models, which deals with an effective cosmological constant given by the energy of a scalar field; a second order phase transition, as it be discussed here, based on a variable cosmological “constant”. The evaporation of the black holes or the infra-red fluctuations of the gravitational conformal modes [@AM] being responsible for the instability of the de Sitter geometry, it seems important to extract observational consequences of the model described above through the analysis of a simple scenario. In [@BS] analytical solutions of the FLRW type for the phenomenological model described above were obtained for any value of the spatial curvature ($k = 0, \pm 1$). They depend on four physical parameters: the initial value of the cosmological constant, $\Lambda_0$, its residual value $\Lambda_\infty$, the initial density of radiation in the Universe $\sigma_0$ and the characteristic time of transmutation of the cosmological constant into radiation energy $\tau$. The choice of a decaying of the cosmological constant into pure radiation is motivated in two ways. From the physical point of view, the temperatures encountered in this model permit us to ignore the masses of the usual degrees of freedom. From the mathematical point of view, the background model is exactly integrable and their perturbations easily computable. Our purpose in the present work is to study the evolution of scalar perturbations in the inflationary model described before. We will restrict ourselves to the case $\Lambda_\infty = \sigma_0 = 0 $ and $k = 0$. These hypothesis lead to the advantage that approximate solutions to the perturbed equations can be easily obtained without to oversimplify the physics. Our model differs essentially in two ways from the usual inflationary models. First (after having fixed the residual gauge freedom allowed by the coordinate transformations), the perturbation equations in the framework of this two fluid model (the effective cosmological constant and radiation) will result in a third order perturbed equation (analytically solvable in the radiation regime and that otherwise can very well be approximated by a $2 + 1$ system of decoupled equations, also analytically solvable). The second characteristic of the model we discuss here concerns the nature of the initial spectrum of the fluctuations. As it will be verified later, the fluctuations evolve freely, i.e., they are determined by the Einstein’s equations, only from the moment where the temperature of the created radiation is significative. Consequently, we will adopt as initial fluctuation spectrum a thermal spectrum, instead of a (hypothetical) spectrum inspired by quantum fluctuations of a vacuum state (a choice imposed more by metaphysical[^3] than by physical considerations; see however the discussions in [@MRS; @MBNP], for instance). This paper is organized as follows. In section 2, the field equations and the solution for the background are reminded. In section 3 the perturbed quantities are settled out, and approximate solutions presented. In section 4, the spectrum of perturbations is discussed. Section 5 is devoted to the conclusions. The field equations =================== In order to be complete, let us review briefly the solutions that will be used in our perturbative study. The geometry of the Universe is supposed to be of the type flat FLRW: $$ds^2 = - dt^2 + a^2(t)(dx^2 + dy^2 + dz^2) \quad .$$ This metric is a reasonable simplification for the primordial stages of the Universe. Moreover, in our case, due to the existence of an inflationary phase, the curvature can also be neglected in later stages of the evolution of the Universe. The unperturbed material source of the Einstein’s equations is the sum of two comoving fluids: $$T^\mu_\nu = (\stackrel{0}{\rho} + \stackrel{0}{p})\stackrel{0}{{u}^\mu} {\stackrel{0}{u}}_\nu + \stackrel{0}{p}\delta^\mu_\nu\equiv T^\mu_{(rad.)\nu}+ T^\mu_{(cos.)\nu}$$ with $\stackrel{0}{\rho} = \rho_{(rad.)} + \rho_\Lambda$, $\stackrel{0}{p} = p_{(rad.)} + p_\Lambda$. The terms representing the radiation perfect fluid ($\rho_{(rad.)}$, $p_{(rad.)}$) are supposed to represent the high energy matter degrees of freedom under the form of radiation: $$\rho_{(rad.)} = \nu\frac{\pi^2}{30}T^4$$ where $T$ is the temperature and $\nu$ is the effective number of degrees of freedom that we will take as being those of the standard model ($\nu = 106.75$), this value of $\nu$ constituting, of course, a lower bound at the energy considered. The term $\rho_\Lambda = {\Lambda(t)}/{8\pi}= - p_\Lambda$ is purely phenomenological. It represents the density of black holes created during the cosmogenesis phase and that evaporate later.We will assume it to be of the form : $$\Lambda(t) = \Lambda_0e^{-t/\tau} \qquad ,$$ where the parameter $\tau$ is the characteristic time of the black hole evaporation which is of the order of the lifetime of the created black holes : $$\tau \sim\frac{2560\pi}{\nu}M^3_{bh} \approx 10^5 \quad.$$ The equation of state of radiation, ${\rho_{(rad.)}}={3}\/ p_{(rad.)}$, permits to obtain a linear second order differential equation for the variable $Z = a^2$ (a dot representing a derivative with respect to the cosmological time $t$): $$\ddot Z - \frac{4}{3}\Lambda_0e^{-t/\tau}Z = 0 \quad .$$ The solution of this equation is given by a combination of modified Bessel functions: $$\label{bs} Z = \alpha K_0(\varpi e^{-\xi}) + \beta I_0(\varpi e^{-\xi}) \quad ,$$ where $\xi = \frac{t}{2\tau}$ and $\varpi = \sqrt{\frac{16}{3}\tau^2 \Lambda_0}$. The coefficients $\alpha$ and $\beta$ are given by $$\label{ic} \alpha = {\varpi}(I'_0(\varpi) + I_0(\varpi)) \quad , \quad \beta = - {\varpi}(K'_0(\varpi) + K_0(\varpi)) \quad ,$$ and are fixed by the conditions $Z(0) = 1$ and $\dot Z(0) = +\sqrt{4\Lambda_0/3}$. Notice that there was an error of sign in the ref. [@SV]. The scale factor written there corresponds to $\dot Z < 0$, and the plateau appearing in the graph of the entropy as function of time is due to the crossing of the singularity $Z = 0$. The graphs for the entropy and temperature as functions of time, for initial conditions (\[ic\]) corresponding to an expanding Universe, are shown on figures 1 and 2.\ \ \ \ \ The perturbed equations ======================= The perturbed equations for the scalar modes around the background solutions exhibited above are obtained through the standard methods. We suppose the metric perturbations $$g_{\mu\nu}(t, \vec x) = \stackrel{0}{g}_{\mu\nu}(t) + \stackrel{1}{g}_{\mu\nu}(t, \vec x)$$ to be synchronous, i.e., such that $\stackrel{1}{g}_{\mu0} = 0$. The perturbations in density and velocity read $$\rho(t,\vec x) = \stackrel{0}{\rho}(t) + \stackrel{1}{\rho}(t,\vec x) \quad , \quad u^\mu(t,\vec x) = \stackrel{0}{{u}\strut^\mu} + {\stackrel{1}{u}}\strut^\mu(t,\vec x) \quad ,$$ with ${\stackrel{0}{u}}\strut^\mu = (1,0,0,0)$. All perturbed quantities can be expressed through a Fourier decomposition $$\begin{aligned} \stackrel{1}{g}_{ij}(t,\vec x) &=& \frac{1}{\pi^{3/2}}\int h_{ij}(t,\vec n) e^{-i\vec n.\vec x}d^3\vec n \quad ,\\ \stackrel{1}{\rho}(t,\vec x) &=& \frac{1}{\pi^{3/2}}\int \delta \rho(t,\vec n) e^{-i\vec n.\vec x}d^3\vec n \quad \quad ,\\ {\stackrel{1}{u}}_\mu(t,\vec x) &=& \frac{1}{\pi^{3/3}}\int{\stackrel{1}{u}}_\mu(t,\vec n)e^{-i\vec n.\vec x}d^3\vec n \quad ,\end{aligned}$$ with $\vec n.\vec x = \sum_i n^ix^i$, $\vec n$ being the comoving wave vector of the perturbations. Introducing those perturbed quantities into Einstein’s equations, we obtain three coupled perturbed equations: $$\begin{aligned} \label{ep1} ({a^2\dot h})^. = a^2\delta\rho \quad ,\\ \label{ep2} (a^4\delta\rho)^. + (\rho + p)a^4\theta = 8\pi(\rho + p)a^4h \quad ,\\ \label{ep3} ((\rho + p)a^5\theta)^. = \frac{n^2}{3}a^3\delta\rho \quad ,\end{aligned}$$ where $\theta = \vec n.\vec u$ and $h = - 16\pi a^{-2}\sum_i h_{ii}$. The system of equations (\[ep1\],\[ep2\],\[ep3\]) is underdetermined since $\rho = \rho_{(rad.)} + \rho_\Lambda$ and consequently $\delta\rho = \delta\rho_{(rad.)} + \delta\rho_\Lambda$. We need a specific equation for the effective cosmological constant. Hence, we assume $$\biggr(\nabla_\mu T^{\mu\nu}_{(cos.)}\biggl)u_\nu \equiv - \nabla_\nu(\rho_\Lambda u^\nu) - p_\Lambda\nabla_\nu u^\nu = \tau^{-1}\rho_\Lambda$$ which reduces to $u^\nu\partial_\nu\rho_\Lambda = \tau^{-1}\rho_\Lambda$, since $\rho_\Lambda = - p_\Lambda = \Lambda(t)/8\pi$. For adiabatic perturbations ($\delta\rho_\Lambda + \delta p_\Lambda = 0$), it results ($\tau$ being supposed to be a constant given by the microphysics) $${\stackrel{1}{u}}\strut^\mu\partial_\mu\stackrel{0}{\rho}_\Lambda + \partial_t\stackrel{1}{\rho}_\Lambda = \tau^{-1}\stackrel{1}{\rho}_\Lambda \quad .$$ Since ${\stackrel{1}{u}}\strut^0 = 0$, we obtain $\stackrel{1}{\rho}_\Lambda = L(\vec x)e^{-t/\tau}$. But this solution is a coordinate artifact. Indeed, the coordinate conditions $g_{\mu0} = 0$ yet remain satisfied when the residual infinitesimal transformations generated by the vector fields $$\epsilon_0 = \phi(\vec x) \quad , \quad \epsilon_i = \psi_i(\vec x)a^2 - \partial_i\phi(\vec x)a^2\int \frac{dt}{a^2}$$ are performed. The choice $\phi(\vec x) = \frac{\tau L(\vec x)}{\Lambda_0}$ allows to make $\delta\Lambda(t) = 0$. Notice that, in this way, the equations (\[ep1\],\[ep2\],\[ep3\]) are now gauge invariants with respect to the remaining coordinate transformations. Indeed, residual transformations $\epsilon_0 = 0$, $\epsilon_i = \psi_ia^2$ only modify $h_{ij}(t,\vec x)$ up to a term whose time dependence is given by $a^2$; hence, $h$ is defined up to a constant, leaving $\dot h$ invariant. In other words, in the synchronous gauge $g_{0\mu} = 0$, $\delta\rho_{\Lambda}$ may be fixed to zero and all solutions of the perturbed equations (\[ep1\],\[ep2\],\[ep3\]) now acquire a physical meaning, in contrast to what happens in the framework of a one-fluid model. In this later case, the third order perturbed equation may be reduced, using the residual coordinate freedom, to a second order equation. Here, we face a third order equation whose three integration constants have physical meaning. From now on, we introduce the new set of variables $${\cal T} = (\rho_{(rad.)} + p_{(rad.)})a^5\theta \quad , \quad {\cal R} = a^4\delta\rho_{(rad.)} \quad , \quad {\cal H} = a^2\dot h \quad ,$$ which permit us to write the system of equations (\[ep1\],\[ep2\],\[ep3\]) as $$\begin{aligned} \label{ep1'} \dot{\cal H} &=& \frac{{\cal R}}{a^2} \quad ,\\ \label{ep2'} \dot{\cal R} &=& - \frac{{\cal T}}{a} + 8\pi(\rho_{(rad.)} + p_{(rad.)})a^2{\cal H} \quad , \\ \label{ep3'} \dot{\cal T} &=& \frac{n^2}{3}\frac{{\cal R}}{a} \quad .\end{aligned}$$ The solutions of this system and their cosmological implication will be the subject of the next section. Evolution of perturbations ========================== We are not interested here in the general solution of the system of equation (\[ep1’\]-\[ep3’\]) governing the evolution of the perturbations, for arbitrary values of their parameters. We shall only focus on the values of the parameters that are relevant for the cosmological model we consider here. In order that an inflationary model is acceptable, the scale factor must be amplified at least by $70$ e-folds, before the Universe reaches the GU temperature $T_{GU} = 10^{-4}$ where radiative era takes place. At this moment, the expansion of the Universe is described by the asymptotic expression $$Z_{as} = \alpha\frac{t - t_s}{2\tau} \equiv \alpha(\xi - \xi_s) \quad , \quad \xi_s \equiv \frac{t_s}{2\tau} \simeq \ln\frac{\varpi}{2} + \gamma \quad ,$$ $\gamma \approx 0.577$ being the Euler-Mascheroni constant. Using the Einstein’s equation $G_{00} = 8\pi T_{00}$, we obtain $$\xi_{GU} - \xi_s = \biggr(\frac{128\pi^3\tau^2}{90}\nu T^4_{GU}\biggl)^{-1/2}$$ Hence, $\alpha \sim \biggr(\frac{128\pi^3\tau^2}{90}\nu T^4_{GU}\biggl)^{1/2} e^{2N_e} \approx 10^{62}$ and $\beta \approx 10^{-65}$ for a e-fold number $N_e = 75$. Accordingly, unless for $\xi$ very close to zero, we may neglect the second term in (\[bs\]). From the asymptotical expression of the $K_0$ Bessel function appearing in eq.(\[ic\]) we may deduce the value of $\varpi$ (i.e. $\Lambda_0$) insuring a given number $N_e$ of e-fold : $$\varpi = 2N_e + \frac{1}{2}\ln\frac{2^{12}\pi^2\nu T^4_{GU}}{3} + \frac{1}{2}\ln\biggr[2N_e + \frac{1}{2}\ln\frac{2^{12}\pi^2 \nu T^4_{GU}}{3}\biggl] \quad .\label{varpi}$$ This implies $\Lambda_0 << 1$ for a large range of reasonable values of $N_e$, in particular $\Lambda_0 \approx 1.4\times10^{-7}$ for $N_e = 75$, what is compatible with the domain of validity of the model, while for $\Lambda_0 = 1/(192\pi^2)$, which corresponds to the assumption of thermal equilibrium between de Sitter temperature and Hawking temperature, we obtain $N_e \simeq 2400$. According to the range of the values settled by the initial conditions, we may build approximate solutions of the perturbed equations. For instance, let us suppose that ${\cal R}$ scales with $\alpha$ as $\alpha^K$. From the perturbed equations, using the conservation of the entropy of the radiation as the cosmological constant goes to zero, we obtain ${\cal H} \sim \alpha^{K-1}$ and ${\cal T} \sim \alpha^{K - 1/2}$. From the perturbed equations, it comes out that in this case $\dot{\cal R} \sim \frac{{\cal T}}{a} \sim \alpha^{K -1}$, which is negligible compared to $\frac{32\pi}{3}\rho_{(rad.)}a^2{\cal H} \sim \alpha^K$. Hence, the perturbed equations can be approximated by : $$\begin{aligned} \label{as1} \dot{\cal H} &=& \frac{{\cal R}}{a^2} \quad ,\\ \label{as2} \dot{\cal R} &=& \frac{32\pi}{3}\rho_{(rad.)}a^2{\cal H} \quad , \\ \label{as3} \dot{\cal T} &=& \frac{n^2}{3}\frac{{\cal R}}{a} \quad ,\end{aligned}$$ which can be integrated by quadratures. The general solution is : $$\begin{aligned} \label{solap1H} {\cal H} &=& {\cal H}_i\frac{Z(t_i)}{Z(t)} + \frac{\mu}{Z(t)}\int_{t_i}^tZ(t')dt' \quad , \\ \label{solap1R} {\cal R}(t) &=& Z(t)\dot{\cal H}(t) = \mu Z(t) - \dot Z(t){\cal H}(t) \quad ,\\ {\cal T} &=& {\cal T}_i + \frac{n^2}3\int_{t_i}^t\frac{{\cal R}(t')}{a(t')}dt'\end{aligned}$$ with $$\mu = \frac{1}{Z(t_i)}\biggr({\cal R}_i + {\dot Z}(t_i){\cal H}_i \biggl)\label{mu}$$ and ${\cal R}_i$, ${\cal H}_i$, ${\cal T}_i$ being the initial values of ${\cal R}$, ${\cal H}$ and ${\cal T}$.\ On the other hand, if ${\rho a^3{\cal H}}$ can be neglected in front of ${\cal T}$, the system reduces to $$\frac{d{\cal R}}{d\zeta} = - {\cal T} \quad , \quad \frac{d{\cal T}}{d\zeta} = \frac{n^2}{3}{\cal R} \quad , \quad \dot{\cal H} = \frac{{\cal R}}{a^2}$$ where $\zeta = \int_{t_i}^t \frac{dt'}{a(t')}$. Its general solution is then $$\begin{aligned} {\cal R} &=& {\cal R}_i\cos\biggr(\frac{n}{\sqrt{3}}\zeta\biggl) - \frac{\sqrt{3}}{n}{\cal T}_i \sin\biggr(\frac{n}{\sqrt{3}}\zeta\biggl) \quad ,\label{solap2R}\\ {\cal T} &=& {\cal T}_i\cos\biggr(\frac{n}{\sqrt{3}}\zeta\biggl) + \frac{n}{\sqrt{3}}{\cal R}_i\sin\biggr(\frac{n}{\sqrt{3}}\zeta\biggl) \quad ,\label{solap2T}\\ {\cal H} &=& {\cal H}_i + \int_{t_i}^t\frac{{\cal R}(t')}{\tau Z(t')}dt' \quad \label{solap2H}.\end{aligned}$$ The initial time $t_i$, where the initial values of ${\cal R}$, ${\cal T}$ and ${\cal H}$ have to be considered, depend on the size of the perturbations. Today the temperature is of the order of $T_0 \sim 2\times10^{-32}$, the cosmological observational data which are somehow free of astrophysical noises mainly concerns sizes $\lambda_0$ from $100 Mpc$ up to $3.000 Mpc$, i.e., $2\times10^{62}l_{Pl}$ up to $6\times10^{63}l_{Pl}$. At these scales the Universe can be considered as homogeneous. At the grand unification epoch, the size of these fluctuations are shortened by a factor ${T_0}/{T_{GU}} = 2\times10^{-28}$. They must still be reduced by a factor ${a(t_i)}/{a(t_{GU})}$ in order to obtain their values at the initial time. Moreover, the evolution of the fluctuations is given by equations (\[ep1\],\[ep2\],\[ep3\]) only once they grow faster than the microscopic interactions can propagate. It is only then that their behavior is no more dictated by the microphysics but only by the overall cosmological expansion. This implies that the evolution of perturbations with comoving wave numbers $n$ and present day size $\lambda_0$ will only be determined by the perturbation equations from an initial time $t_i(n)$ given by $\dot a(t_i(n)) = n$, i.e., $$\label{it} \dot a(t_i(n)) = \frac{2\pi}{\lambda_0}\frac{T_{GU}}{T_0}e^{N_e}\quad .$$ This time $t_i(n)$ may be easily estimated by noting that $\varpi e^{-\xi_{GU}}$, and thus $\varpi e^{-\xi_{n}}$ (where $\xi_i(n) = t_i(n)/2\tau$), depends very slowly on the number of e-folds. Moreover, near $t_i(n) = 0$ equation (\[it\]) easily may be linearized. Thus for each value of $\lambda_{\star}$, it is possible to determine a minimal number $N^{min}(\lambda_{\star})$ of e-folds such that fluctuations of size as $\lambda_{\star}$ today emerges from the horizon at $t = 0$: $$N^{min}(\lambda_{\star}) = \ln\biggr[\frac{\lambda_{\star}}{4\pi\tau}\frac{T_0}{T_{GU}} \ln\frac{\lambda_{\star} T_{0}}{4\pi^2\tau T_{GU}\sqrt{\xi_{GU}}}\biggl] + 1.93 \quad ,$$ and giving this number, we obtain from eq. (\[varpi\]), for a given number $N_e$ of e-folds, $$t_i(\lambda_{\star}) \approx 2\tau\ln\biggr[\frac{\varpi(N_e)}{\varpi(N^{min})}\biggl]$$ as the time where a fluctuation of size $\lambda_{\star}$ has emerged from the horizon. We restrict ourselves to positive values of $t_i$(we do not know the physics before $t < 0$). This implies that the minimum number of e-fold between $t = 0$ and the beginning of the grand unification era we will consider is at least $75$. The decay of the cosmological constant increases very rapidly the temperature (see fig. 1), but nevertheless we shall suppose that the density fluctuations are given by the laws of the equilibrium statistical mechanics [@LL5]: $$\delta\rho(t_i,n) = \nu_{eff}^{1/2}T^{5/2}\biggr(\frac{a(t_i(n))}{n}\biggl)^{3/2} \frac{1}{a^3(t_i(n))}\qquad .$$ The extra factor $a^{-3}(t_i(n))$ is the jacobian that we have to introduce when we consider the Fourier transforms with respect to the comoving variables $n$ which are related to the physical momentum $n_{phys}$ at time $t_i$ by $$n_{phys} = \frac{n}{a(t_i)} = \frac{2\pi}{\lambda_{\star}}\frac{T_{GU}}{T_0} \frac{e^{N_e}}{a(t_i)} \quad .$$ For consistency, we may check that the ratio of the proper energy density of such fluctuation to the temperature $\epsilon(n)/T_i(n)$ is approximatively equals to $1.8$ when the fluctuation emerges from the horizon, the temperature $T_i(n)$ at this moment being of the order or $10^{-3}$. We may also check that these values are very few sensitive to the values of $\tau$ and $N_e$. Moreover, we also assume the initial spectrum of $\theta(t_i(n))$ and $\dot h(t_i(n))$ being given by classical physics, i.e., they are not affected by the overall expansion of the Universe. Using standard arguments [@LL5; @LL6], we obtain $$\begin{aligned} \theta(t_i,n) &\sim& \nu_{eff}^{1/2}T_i^{-3/2}(n)n^{-1/2}a(t_i(n))^{-5/2} \quad ,\\ \dot h(t_i,n) &\sim& \nu_{eff}^{1/2}T_i^{5/2}(n)n^{-5/2}a(t_i(n))^{-1/2} \quad .\end{aligned}$$ Hence, ${\cal R}(t_i,n) = Z^2(t_i(n))\delta\rho(t_i,n)$, ${\cal T}(t_i,n) \sim n{\cal R}(t_i,n)$ and ${\cal H}(t_i,n) \sim {{\cal R}(t_i,n)}/{na(t_i,n)}$ The second relation means that the longitudinal mode $h(t_i,p)$ is determined by the gravitational fluctuations, with a characteristic time of the order of the size of the fluctuations considered (of course, one also could ask about the metaphysical content of these considerations, that we adopt in the absence of a better scenario). With these assumptions, it is possible to verify that at initial times ${\cal T}/a$ is about $2-5\times10^3$ times greater than $\frac{32\pi}{3}\rho_{(rad.)}a^2{\cal H}$ : $$\frac{32\pi}{3}\rho_ia_i^3\frac{{\cal H}_i}{{\cal T}_i} \approx 8\pi\frac{a^2(t_i)\rho_{(rad.)}(t_i)}{n^2} \approx \frac{3}{16\tau^2n^2}\varpi e^{-\xi_i/2}e^{\varpi(1 - e^{-\xi_i})} << 1 \quad ,$$ a result again almost insensitive to the values of $N_e$ and $\tau$. As a consequence, during a first short period of time we may approximate the system of the perturbation equations by the eqs (\[as1\],\[as2\],\[as3\]). Hence, one verify that the functions ${\cal H}$, ${\cal R}$ and ${\cal T}$ given by eqs (\[solap2R\],\[solap2T\],\[solap2H\]) change little in the beginning of the evolution of perturbations. But the ratio between the two terms in the right hand side of (\[ep2\]) behaves as : $$\frac{32\pi}{3}\rho_{(rad.)}a^3\frac{{\cal H}}{{\cal T}} \sim 8\pi \frac{\rho_{(rad.)}(t)a^3(t)}{n^2a(t_i)}$$ and increases very quickly. Hence, for values of $\xi$ very near $\xi_i$, the system will be in the domain of validity of the first approximate solution (\[solap1H\],\[solap1R\]). Taking into account the uncertainties on the the parameters of the problem (initial spectrum, precise duration of validity of the perturbed equations, etc.), we can adopt them as solutions of the perturbed equations. From these considerations, it comes out that at the grand unification temperature, the density contrast $\delta_{GU} = |\delta\rho(t_{GU})/\rho_{GU}|$ is given by $$\delta_{GU} = \frac{{\cal R}(t_{GU})}{\rho_{GU}Z^2_{GU}} \approx \frac{32\pi}{3}\frac{\mu e^{2N_e}}{\alpha^2} \quad ,$$ and all dependence on the wave number $n$ is contained in the value of the integration constant $\mu$. This constant is given by the sum of two terms (see eq. (\[mu\]), but the first one $Z(t_i)\delta\rho_i$ is significantly more important than the second one. Hence, the spectrum index depends only on the initial spectrum of density perturbations and on the dates where, for a given size $\lambda_0$, the fluctuations emerge from the horizon. As the temperature $T(t_i,n)$ is almost constant during the interval of time where the relevant perturbations ($100Mpc<\lambda_0<3000Mpc$) cross the horizon while $a(t_i(n)) \div n$ we immediately obtain the spectral index $$n_s = 4 + \frac{d}{d\ln n}\ln[(\delta\rho(n)/\rho)^2] \approx 2 \quad .$$ Conclusions =========== In this work, we have studied an inflationary model that joins the radiative regime by a second order phase transition. This model is inspired in an inflationary primordial scenario where the transition from a de Sitter phase to a radiative phase occurs due to the evaporation of primordial black holes. It must be considered as a source of a possible alternative to the usual inflationary models, and not as a definitive model. Indeed, the underlined physics leading to the background geometry is phenomenological and asks for more deep fundamental investigations. It offers, however, the conceptual advantage of making no appeal to a self-interacting scalar field. The kinetics of the phase transition asks also for a more rigorous argumentation. However, we believe our results not to be dependent on the details of the of the model as the spectral index obtained is very little sensible to the values of $N_e$ and $\tau$. Moreover, we have introduced hypothesis on the nature of the initial spectrum of the fluctuations. In particular, we employed a thermal spectrum for the fluctuations as they emerge from horizon. Notice that, from the conceptual point of view, the spectrum more delicate is that of the $h$ but this quantity does not play a significative role in the computation of the spectral index, since the gravitational fluctuations is dominated by those, much more important, fluctuations of matter, which is governed by ${\cal R}$. The nature of the spectrum of $\delta\rho$ affects the spectral index essentially by the function $a(t_i) \sim n$ which appears in the jacobian of density fluctuation defining $\delta\rho_i$ and on the behavior of $Z(t_i,n)$, as appears in the expression for the initial conditions. Of course the obtained value of the spectral index is not in agreement with observational data. However here we have only discussed scalar fluctuations; it remains to see if tensorial fluctuations cannot reconcile the model with the expected Harrison-Zeldovitch spectra. But such an analysis reopen the question of the choice of the initial spectrum of the gravitational perturbations, a question whose answer, in our opinion is beyond our present knowledge of physics. ACKNOWLEDGMENTS {#acknowledgments .unnumbered} =============== The authors thank Jérôme Martin for fruitful discussions and CAPES (Brazil) for partial financial support. Ph. S. thanks Robert Brout, François Englert and Marianne Rooman for numerous enlightening discussions about many aspects of the study performed here, Jean Bricmont for noticing him ref. [@Ru] and CAPES (Brazil) for partial financial support. He also thanks the Departamento de Física of UFES for its warm hospitality. [90]{} A. Friedmann, Z. Phys. [**10**]{}, 377(1922), G. Lemaître, Ann. Soc. Sci. Bruxelles, I [**A53**]{}, 51(1933); H. P. Robertson, Astrophys. J. [**82**]{}, 248(1935); A. G. Walker, Proc. R. Soc. Edinburgh [**52**]{}, 345(1932). A. H. Guth, Phys. rev. D [**23**]{}(1981) 347, S. K. Blau and A.H. Guth, in [**300 years of gravitation**]{}, edited by S. Hawking and W. Israel, Cambridge University Press, Cambridge(1987). S.W. Hawking and G.F.R. Ellis, [**The large scale structure of space-time**]{}, Cambridge University Press, Cambridge (1973); R. Brout, F. Englert and E. Gunzig, Ann. Phys. (NY) [**115**]{}, 78(1978); Gen. Rel. Grav. [**10**]{}, 1(1979). E.P. Tryon, Nature [**246**]{}, 396(1973). R. Brout, F. Englert and Ph. Spindel, Phys. Rev. Lett. [**43**]{}, 417(1979); Phys. Rev. Lett. [ **43**]{}, 890(1979). R. Brout, F. Englert, J.-M. Frère, E. Gunzig, P. Nardone, C. Truffin and Ph. Spindel, Nucl. Phys. [**107B**]{},361(1981). Ph. Spindel, Phys. Lett. [**107B**]{}, 361(1981). Ph. Spindel and P. Vandepeutte, [*Numerical investigations in cosmogenesis*]{}, in Proceedings of the 26th Liège International Astrophysical Colloquium, Université de Liège(1986). F. Englert, [*Primordial inflation*]{}, in Proceedings of the Erice International School on “Basics and Highlights in Fundamental Physics”(1999). A. Casher and F. Englert, Phys. Lett. [**B104**]{}, 117(1981). J.S. Dowker and R. Critchley, Phys. Rev. [**D12**]{}, 327(1976). Ph. Spindel, Phys. Rev. [**D37**]{}, 2092(1988). I. Antoniadis and E. Mottola, J. Math. Phys. [**32**]{}(1991). R. Brout and Ph. Spindel, Phys. Lett. [**B320**]{}, 241(1999). B. Russell, [**My Philosophical Development**]{}, London: Routledge (1995 \[1959, page 164\]). J. Martin, A. Riazuelo, M. Sakellariadou Phys. Rev. D [**61**]{} (2000) 083518, astro-ph/9904167. J. Martin and R. H. Brandenberger, [*The trans-planckian problem of inflationary cosmology*]{}, hep-th/0005209; J. C. Niemeyer, R. Parentani [*Trans-Planckian dispersion and scale-invariance of inflationary perturbations*]{}, astro-ph/0101451. L. Landau and E. Lifshitz, [**Mécanique statistique**]{}, Mir, Moscou(1967). L. Landau and E. Lifshitz, [**Mécanique des fluides**]{}, Mir, Moscou (1971). [^1]: e-mail: [email protected] [^2]: [email protected] [^3]: During the fifties, Bertrand Russell observed: “The accusation of metaphysics has become in philosophy something like being a security risk in the public service. …The only definition I have found that fits all cases is: ‘a philosophical opinion not held by the present author’." [@Ru].

ArXiv

--- abstract: - | The theory of dynamic conductivity of nanosystem is developed within the model of rectangular potentials and different effective masses of electron in open three-barrier resonance-tunnel structure in a constant homogeneous electric field. The application of this theory for the improvement of operating characteristics of quantum cascade laser active region (for the experimentally investigated In$_{0.53}$Ga$_{0.47}$As/In$_{0.52}$Al$_{0.48}$As heterosystem) proves that for a certain geometric design of nanosystem there exists such minimal magnitude of constant electric field intensity, at which the electromagnetic field radiation power together with the density of current flowing through the separate cascade of quantum laser becomes maximal. resonance-tunnel structure, dynamic conductivity 73.21.Fg, 73.90.+f, 72.30.+q, 73.63.Hs - | =3000У моделі прямокутних потенціалів і різних ефективних мас електрона у різних елементах відкритої трибар’єрної резонансно-тунельної структури, що знаходиться в постійному однорідному електричному полі, розроблена теорія динамічної провідності наносистеми. Застосування розробленої теорії для покращення робочих характеристик активної області квантового каскадного лазера (на основі експериментально досліджуваної системи In$_{0.53}$Ga$_{0.47}$As/In$_{0.52}$Al$_{0.48}$As) показало, що при заданому геометричному дизайні наносистеми існує така мінімальна величина напруженості постійного електричного поля, при якій одночасно максималізується як величина потужності електромагнітного випромінювання, так і густина струму, що проходить крізь окремий каскад квантового каскадного лазера. резонансно-тунельна стректура, динамічна провідність address: - 'Fedkovych Chernivtsi National University, 2, Kotsyubinsky St., Chernivtsi, 58012, Ukraine[^1]' - | Чернівецький національний університет імені Юрія Федьковича, вул. Коцюбинського, 2,\ 58012 Чернівці, Україна author: - 'Ju. Seti, M. Tkach, O. Voitsekhivska' - 'Ю.О. Сеті, М.В. Ткач, О.М. Войцехівська' date: 'Received June 16, 2010' title: 'Динамічна провідність симетричної трибар’єрної плоскої наносистеми у постійному електричному полі' --- Introduction ============ Recently, there has been achieved a considerable progress in the experimental fabrication of quantum cascade lasers (QCL) [@1; @2; @3; @4] and quantum cascade detectors (QCD) [@5; @6; @7; @8] of various geometric design. The investigation of these devices attracts great attention due to their operation in the actual terahertz range of electromagnetic waves. The main focus is made on the optimization of parameters of nano-devices. However, this is a rather complicated problem due to the absence of a consequent and consistent theory of physical processes taking place in open nanosystems. The active operating elements of experimental QCL or QCD are the open resonance-tunnel structures (RTS) with different number of barriers and wells. Thus, the properties of their static and dynamic conductivities determining the basic QCL parameters, i.e., regions and widths of ranges of operating parts, radiation intensity, excited current and so on, have been theoretically studied. In references [@9; @10; @11; @12; @13; @14; @15], mainly within the model of unitary effective mass and $\delta$-like potential barriers, there have been developed the theoretical approaches to the calculation of active conductivity of electrons in open RTS. Recently, in references [@16; @17] it was shown that $\delta$-barrier model with unitary electron effective yields too rough magnitudes of resonance widths of quasistationary states (ten times bigger) relatively to the realistic model of rectangular potential barriers with different effective masses of quasiparticle in different pars of RTS. The conductivity is very sensitive to the magnitudes of resonance widths of quasistationary states. Therefore, the rectangular potential barriers and different effective masses are to be taken into account within the framework of the respective model. In the majority of theoretical papers dealing with the conductivity of open RTS, the presence of constant electric field has not been taken into account at all or has been evaluated only roughly [@18]. However, the effective QCL [@1; @2; @3; @4] has been experimentally produced just at the applied constant electric field. Thus, an urgent task is to develop a consistent theory of open RTS conductivity at an applied constant electric field; the model would be deprived of the rough $\delta$-like approximating barriers and would consider different effective masses of quasiparticles in the wells and barriers. In the proposed paper, there is developed a theory of electronic conductivity of open symmetric three-barrier RTS under the applied constant homogeneous electric field within the framework of the model of different electron effective masses in component parts of a nanosystem with rectangular potential wells and barriers. For the first time, the obtained exact solutions of the equations determining the magnitude of the active conductivity at the applied constant electric field at RTS make it possible to investigate it in a weak signal one-mode approximation over the radiation field intensity. By the example of experimental nanosystem In$_{0.53}$Ga$_{0.47}$As/In$_{0.52}$Al$_{0.48}$As, it is shown that using the presented model it is possible to optimize the operating parameters of QCL depending on its geometric design and electric field intensity. Hamiltonian. Conductivity of three-barrier nanosystem ===================================================== The open symmetric three-barrier RTS with the applied constant electric field characterized by intensity $F$ (figure 1) is under study. ![Geometrical (a) and potential energy (b) schemes of RTS.[]{data-label="fig-smp1"}](fig1){width="45.00000%"} It is assumed that the monoenergetic ($E$, the energy) electron current ($n$, the concentration) is falling at the RTS from the left side, perpendicularly to its planes. The small difference between lattice constants of RTS layers-wells (the media: $0, 2, 4, 6$) and layers-barriers (the media: $1, 3, 5$) allows us to study the nanosystem within the framework of the model of effective masses and rectangular potential barriers $$\label{eq1} m(z) = {\left\{ {{\begin{array}{*{20}c} {m_{0}} \hfill \\ {m_{1}} \hfill \\ \end{array}} } \right.}{\rm ,} \quad U(z) = {\left\{ {{\begin{array}{*{20}c} {0,\,\,\,\,\,\,\,\text{in reg.}\,\,0,\,2,\,4,\,6,} \hfill \\ {U,\,\,\,\,\,\,\text{in reg.}\,\,1 ,\,3 ,\,\,5. \,\,\,\,\,\,\,\,\,\,} \hfill \\ \end{array}} } \right.}\,\,$$ The Schrodinger equation for the electron is written as $$\label{eq2} \ri\hbar {\frac{{\partial \Psi (z,t)}}{{\partial \,t}}} = [H_{0} + H(z,t)] \Psi (z,t){\rm ,}$$ where $$\label{eq3} H_{0} = - {\frac{{\hbar ^{2}}}{{2}}}{\frac{{\partial} }{{\partial z}}}{\frac{{1}}{{m\left( {z} \right)}}}{\frac{{\partial} }{{\partial z}}} + U(z) - \re F \{z [\theta(z)-\theta(z-z_{5})]+z_{5} \theta(z-z_{5})\}{\rm }$$ – the Hamiltonian of stationary problem (with constant electric field), $$\label{eq4} H(z,t) = - e \epsilon \{z [\theta(z)-\theta(z-z_{5})]+z_{5} \theta(z-z_{5})\} ( \re^{\ri \omega t} + \re^{-\ri \omega t} )$$ – the interaction Hamiltonian of electron with the electromagnetic field varying in time ($\omega$, the frequency) and amplitude of electric field intensity ($\epsilon$). The solution of equation (2.2) in one-mode approximation, assuming the amplitude of high frequency field to be small  [@12; @13; @14; @18], according to the perturbation theory, is as follows: $$\label{eq5} \Psi \left( {z,t} \right) = \Psi _{0} \left( {z} \right)\re^{-\ri \omega _{0} t} + \Psi _{ + 1} \left( {z} \right)\re^{-\ri \left( {\omega _{0} + \omega} \right)t} + \Psi _{ - 1} \left( {z} \right)\re^{-\ri \left( {\omega _{0} - \omega} \right)t}\,, \qquad \left( {\omega _{0} = {{E} \mathord{\left/ {\vphantom {{E} {\hbar} }} \right. \kern-\nulldelimiterspace} {\hbar} }} \right).$$ Here $\Psi _{0}(z)$ function is the solution of stationary Schrodinger equation $$\label{eq6} H_{0}(z)\, \Psi _{0} (z)=E\, \Psi _{0} (z).$$ Considering that the energy of electronic current in X0Y plane is negligibly small ($k_{||}=0$) it can be written as follows: $$\begin{aligned} \label{eq7} \Psi _{0} (z)&=&\Psi^{(0)} _{0}(z)\theta(-z)+\sum\limits_{p = 1}^{5} \Psi^{(p)} _{0}(z) [\theta(z-z_{p-1})-\theta(z-z_{p})]+\Psi^{(6)} _{0}(z)\theta(z-z_{5}) \nonumber\\ &=&\left(\re^{\ri k^{(0)} z} +B^{(0)} \re^{-\ri k^{(0)} z}\right)\theta(-z)+A^{(6)} \re^{\ri k^{(6)} (z-z_{5})} \theta (z-z_{5} )\nonumber\\ &&{}+ \sum\limits_{p = 1}^{5} \left[A^{(p)} \mathrm{Ai}(\xi^{(p)})+B^{(p)} \mathrm{Bi}(\xi^{(p)})\right] \left[\theta(z-z_{p-1})-\theta(z-z_{p})\right],\end{aligned}$$ where $\mathrm{Ai}(\xi), \mathrm{Bi}(\xi)$ are the Airy functions and $$\begin{aligned} \label{eq8} k^{(0)} &=&\hbar ^{-1} \sqrt{\displaystyle 2m_{0} E} , \qquad k^{(6)} =\hbar ^{-1} \sqrt{\displaystyle 2m_{0} \left (E+\re Fz_{5} \right)} ,\nonumber\\ % \xi^{(1)}&=&\xi^{(3)}=\xi^{(5)}=\displaystyle \rho^{(1)} \left[(U-E)/\re F-z \right], \qquad \rho^{(1)}=\rho^{(3)}=\rho^{(5)}=\displaystyle(2m_{1}\re F/\hbar^{2})^{1/3},\nonumber\\ % \xi^{(2)}&=&\xi^{(4)}=-\displaystyle \rho^{(2)} \left(E/\re F+z \right), \qquad \rho^{(2)}=\rho^{(4)}=\displaystyle(2m_{0}\re F/ \hbar^{2})^{1/3} .\end{aligned}$$ The unknown coefficients ($B^{(0)}, A^{(6)}, A^{(p)}, B^{(p)}\,\ (p=1,\dots ,6)$) are fixed by the fitting conditions for the wave functions and their densities of currents at all media interfaces $$\label{eq9} \Psi_{0}^{(p)} (z_{p} )=\Psi _{0}^{(p+1)} (z_{p} ), \quad \quad \displaystyle \frac{1}{m_{0(1)} } \left. \frac{\rd{\rm \,\ ; }\Psi _{0}^{(p)} }{\rd z} \right|_{z=z_{p} } =\frac{1}{m_{1(0)} } \left. \frac{\rd{\rm \; }\Psi _{0}^{(p+1)} (z)}{\rd z} \right|_{z=z_{p} }, \qquad (p=0,\dots ,5)$$ together with the normalizing condition $$\label{eq10} \int\limits_{-\infty }^{\infty }\Psi _{0}^{*} (k'z)\Psi _{0} (kz) {\rm \; }\rd z=\delta (k-k') .$$ In order to define the both terms of corrections ($\Psi_{\pm1}(z)$) to wave function, taking into account equations (2.2)–(2.6) and preserving the magnitudes of the first order, inhomogeneous equations are obtained: $$\label{eq11} [H_{0}(z)-\hbar (\omega_{0}-\omega)]\Psi_{\pm1}(z)-\re \epsilon \{ [\theta(z)-\theta(z-z_{5})]+z_{5} \theta(z-z_{5})\} \Psi_{0}(z)=0 .$$ Their solutions are the superposition of functions $$\label{eq12} \Psi_{\pm1}(z)=\Psi_{\pm}(z)+\Phi_{\pm}(z) ,$$ where $\Psi_{\pm}(z)$ are the partial solutions of homogeneous and $\Phi_{\pm}(z)$ are the partial solutions of inhomogeneous equations (2.11). The solutions of homogeneous equations are $$\begin{aligned} \label{eq13} \Psi _{\pm} (z)&=&\Psi^{(0)} _{\pm}(z)\theta(-z)+\sum\limits_{p = 1}^{5} \Psi^{(p)} _{\pm}(z) [\theta(z-z_{p-1})-\theta(z-z_{p})]+\Psi^{(6)} _{\pm}(z)\theta(z-z_{5}) \nonumber\\ &=&B^{(0)}_{\pm} \re^{-\ri k^{(0)}_{\pm} z}\theta(-z)+A^{(6)}_{\pm} \re^{\ri k^{(6)}_{\pm} (z-z_{5})} \theta (z-z_{5} )\nonumber\\ &&{}+ \sum\limits_{p = 1}^{5} \left[A^{(p)}_{\pm} \mathrm{Ai}(\xi^{(p)}_{\pm})+B^{(p)}_{\pm} \mathrm{Bi}(\xi^{(p)}_{\pm})\right] \left[\theta(z-z_{p-1})-\theta(z-z_{p})\right],\end{aligned}$$ where $$\begin{aligned} \label{eq14} k^{(0)}_{\pm} &=&\hbar ^{-1} \sqrt{\displaystyle 2m_{0} (E \pm \hbar\omega)}, \qquad k^{(6)}_{\pm} =\hbar ^{-1} \sqrt{\displaystyle 2m_{0} \left (E\pm \hbar\omega+\re Fz_{5} \right)}, \nonumber\\ % \xi^{(1)}_{\pm}&=&\xi^{(3)}_{\pm}=\xi^{(5)}_{\pm}=\displaystyle \rho^{(1)} \left\{[U-(E\pm \hbar\omega)]/\re F-z \right\}, \qquad \xi^{(2)}_{\pm}=\xi^{(4)}_{\pm}=-\displaystyle \rho^{(2)} \left[(E\pm \hbar\omega)/\re F+z \right]. \nonumber\\\end{aligned}$$ The partial solutions of inhomogeneous equations (2.11), as it is shown, are of an exact analytical form $$\begin{aligned} \label{eq15} \Phi_{\pm}(z)\!\!\!&\!=\!&\!\displaystyle \pi \frac{\epsilon}{F} \sum\limits_{p = 1}^{5} \displaystyle \left[\mathrm{Bi}(\xi^{(p)}_{\pm}) \int\limits_{1}^{\xi ^{(p)}} \left(\eta-\rho^{(p)} \frac{U(z)-E}{\re F}\right) \mathrm{Ai}\left(\eta \mp \rho^{(p)} \frac{\hbar \omega}{\re F}\right) \Psi_{0}^{(p)}(\eta)\rd\eta \right.\nonumber\\ % \!\!\! &&\left.{}-\displaystyle \mathrm{Ai}(\xi^{(p)}_{\pm}) \!\!\int\limits_{1}^{\xi ^{(p)}} \!\!\left(\eta-\rho^{(p)} \frac{U(z)-E}{\re F}\right) \!\mathrm{Bi}\left(\eta \mp \rho^{(p)} \frac{\hbar \omega}{\re F}\right) \Psi_{0}^{(p)}(\eta)\rd\eta\right]\!\! [\theta(z-z_{p-1})-\theta(z-z_{p})] \nonumber\\ % \!\!\! &&{}\mp \displaystyle \frac{\re \epsilon z_{5}}{\hbar \omega} \Psi_{0}^{(6)}(z_{5})\theta(z-z_{5}).\end{aligned}$$ The conditions for $\Psi(z,t)$ wave function continuity (2.5), at all nanosystem interfaces, lead to the boundary conditions for the $\Psi_{\pm1}(z)$ $$\label{eq16} \Psi_{\pm1}^{(p)} (z_{p} )=\Psi _{\pm1}^{(p+1)} (z_{p} ), \qquad \displaystyle \frac{1}{m_{0(1)} } \left. \frac{\rd{\rm \; }\Psi _{\pm1}^{(p)} }{\rd z} \right|_{z=z_{p} } =\frac{1}{m_{1(0)} } \left. \frac{\rd{\rm \,\ ; }\Psi _{\pm1}^{(p+1)} (z)}{\rd z} \right|_{z=z_{p} }, \qquad (p=0,\dots ,5).$$ The solution of the system of inhomogeneous equations (2.16) defines all unknown coefficients ($B^{(0)}_{\pm}$, $A^{(6)}_{\pm}$, $A^{(p)}_{\pm}$, $B^{(p)}_{\pm}$). Now, $\Psi_{\pm} (z)$ functions (2.13) and $\Psi_{\pm1} (z)$ corrections (2.15) are definitively written, since $\Psi(z,t)$ being the whole wave function, is fixed too. Introducing the energy of interaction between the electron and electromagnetic field, which can be calculated as a sum of electronic wave energies flowing from both sides of RTS due to the current, the real part of active conductivity in quasiclassic approximation [@19] is determined by densities of currents with the energies $$\label{eq17} \sigma(\omega)=\displaystyle \frac{\hbar \omega}{2 z_{5} e \epsilon^{2}} \left\{[j(E+\hbar \omega,z_{5})-j(E-\hbar \omega,z_{5})]-[j(E+\hbar \omega,0)-j(E-\hbar \omega,0)]\right\} .$$ According to the quantum mechanics, the density of current of uncoupling electrons with concentration $n$ is related to the whole wave function $$\label{eq18} j(z,t)=\displaystyle \frac{\re \hbar n}{2 m(z)} \left[\Psi(z,t) \frac{\partial}{\partial z}\Psi^{\ast}(z,t)-\Psi(z,t)\frac{\partial}{\partial z}\Psi(z,t)\right] .$$ Thus, taking into account the expressions (2.5), (2.17) and (2.18), we obtain the final expression for the real part of nanosystem active conductivity $$\label{eq19} \sigma(\omega)=\sigma_{\mathrm{l}}(\omega)+\sigma_{\mathrm{r}}(\omega) ,$$ where $$\label{eq20} \sigma_{\mathrm{l}}(\omega)=\displaystyle \frac{\hbar^{2} \omega n}{2 z_{5} m_{0} \epsilon^{2}} (k^{(0)}_{+} |B^{(0)}_{+}|^{2}-k^{(0)}_{-} |B^{(0)}_{-}|^{2}) ,$$ $$\label{eq21} \sigma_{\mathrm{r}}(\omega)=\displaystyle \frac{\hbar^{2} \omega n}{2 z_{5} m_{0} \epsilon^{2}} (k^{(6)}_{+} |A^{(6)}_{+}|^{2}-k^{(6)}_{-} |A^{(6)}_{-}|^{2}) ,$$ $\sigma_{\mathrm{r}}(\omega), \sigma_{\mathrm{l}}(\omega)$ are the conductivities, stipulated by electronic current interacting with electromagnetic field and flowing out ($\mathrm{r}$ – right side) and back ($\mathrm{l}$ – left side) of nanosystem. Analysis of the results ======================= The numeric calculations and analysis of the symmetric three-barrier RTS conductivity ($\sigma$) and function of the electric field intensity, $F$ (equal to the magnitude of energy shift $U_{0}=\re F z_{5}$) were performed for In$_{0.53}$Ga$_{0.47}$As/In$_{0.52}$Al$_{0.48}$As open nanoheterosystem with physical parameters: $m_{0}=0.089 m_{e}$, $m_{1}=0.046 m_{e}$, $a_{0}=0.5867$ nm, $a_{1}=0.5868$ nm, $U=516$ meV. The geometrical sizes of wells and barriers were taken in the ranges of values typical of the experimentally investigated nanosystems [@1; @2; @3; @4; @5; @6; @7; @8]. It was assumed that the current of monoenergetic electrons with concentration $n=10^{16}$ cm$^{-3}$ and energy $E$, corresponding to the resonance energy of the second quasistationary state ($E_{n=2}$), falls at the RTS from the left side. ![Dependences of resonance energies ($E_{n}$), widths ($\Gamma_{n}$), conductivity ($\sigma$), its terms ($\sigma_{\mathrm{l}}, \sigma_{\mathrm{r}}$) and parameter of efficiency ($\alpha$) on energy shift ($U_{0}=\re F z_{5}$) and three different magnitudes the thicknesses of outer barriers at the fixed sizes of the inner barrier ($\Delta=1.6$ nm) and the widths of wells ($b=6$ nm).[]{data-label="fig-smp2"}](fig2){width="90.00000%"} Figure 2 shows the dependences of resonance energies spectrum ($E_{n}$) and resonance widths ($\Gamma_{n}$) of electron quasistationary states and conductivity ($\sigma$) at the transition from the second to the ground quasistationary state and its terms ($\sigma_{\mathrm{r}}$ – due to the right-side current and $\sigma_{\mathrm{l}}$ – due to the left-side current) on the magnitude of electromagnetic field shift ($U_{0}$) at the fixed: $b=6$ nm (wells), $\Delta=1.6$ nm (inner barrier) and $\Delta_{1}=2.2$ nm; 3.4 nm; 4.6 nm (three different magnitudes of outer barriers). In figure 2: In$_{0.52}$Al$_{0.48}$As barrier layers are in bold, In$_{0.53}$Ga$_{0.47}$As well layers are in roman. There is also calculated and presented in the same figure an important parameter $\alpha=\sigma (\sigma_{\mathrm{r}}-\sigma_{\mathrm{l}})/\mathrm{max}[\sigma (\sigma_{\mathrm{r}}-\sigma_{\mathrm{l}})]$, characterizing the optimal efficiency RTS acting on condition that for the fixed geometric design of its parts and energy shift ($U_{0}$), both the conductivity ($\sigma$) proportional to the force of laser radiation and the magnitude ($\sigma_{\mathrm{r}}-\sigma_{\mathrm{l}}$) proportional to the current flowing through the RTS would be optimal at the same time. From figure 2 one can see the following. The magnitudes of resonance energies ($E_{n}$) are almost independent of the thicknesses of outer barriers ($\Delta_{1}$) at the fixed sizes of wells ($b$). At the $U_{0}$ increasing (equal to the increasing intensity of the electric field) all resonance levels show the linear shift into the region of smaller energies. At the same time, the resonance widths ($\Gamma_{n}$) of odd states are increasing and even decreasing. Consequently, there exist such magnitudes of shifts ($U_{0}^{(p)}$), at which the resonance widths of neighbouring states are equal. At the increasing thicknesses of RTS outer barriers, $U_{0}^{(p)}$ magnitudes do not almost vary and resonance widths exponentially increase. According to the developed theory, the intensity of laser radiation at the frequency $\omega=(E_{2}-E_{1})/\hbar$ would be almost proportional to the maximal magnitude of negative conductivity $\sigma$, arising at the transitions of electrons from the second into the ground quasistationary state. From figure 2 it is clear that $\sigma$ absolute value firstly increases till certain maximum value and further linearly decreases for the bigger $U_{0}$, independently of RTS geometrical parameters. This takes place due to the same $\sigma_{\mathrm{r}}$ bahaviuor while $\sigma_{\mathrm{l}}$ (absolute value) only decreases at $U_{0}$ increasing. Figure 2 shows that the parameter of optimal effectiveness ($\alpha$) has one maximum at such $U_{0}^{\mathrm{eff}}$ (for the system under research, $U_{0}^{\mathrm{eff}}\approx40$ meV relating to the electric field intensity $F=18$ kVcm$^{-1}; 20 $ kVcm$^{-1}; 22 $ kVcm$^{-1}$ for the three different sizes of nanosystem) the magnitude of which almost does not depend on the sizes of outer barriers at the fixed sizes of wells and outer barriers. It is obvious that $U_{0}^{\mathrm{eff}}$ magnitude almost coincides with the shift at which $\sigma_{\mathrm{r}}$ approaches its extremum. Thus, a certain intensity of constant electric field realizing the shift $U_{0}^{\mathrm{eff}}$ at which the nanosystem, as an active element of QCL, works optimally, exists for the fixed geometrical configuration of symmetrical three-barrier RTS. As far as the increase of the sizes of the outer barriers causes an exponential increase of conductivity, as it is clear from figure 2, when the thicknesses of the outer barriers become bigger than two lattice constats, the conductivity becomes about two orders bigger. Naturally, it does not mean that by increasing the sizes of outer barriers one can obtain the arbitrarily big magnitudes of conductivities because it is obvious that herein the electron lifetime also exponentially increases. For certain sizes of nanosystem, the lifetime becomes much bigger than the relaxation time of electron energy arising due to the electron-phonon, electron-electron and other types of interactions. Therefore, the model which does not take into account these types of energy relaxation becomes useless. Taking this into consideration we studied such sizes of outer barriers at which the electron lifetime in the lowest quasistationary states not bigger than one order exceed the relaxation time due to the interaction with phonons, evaluated in reference [@1] approximately as one picosecond. Conclusions =========== 1. For the first time, there is developed a theory of dynamic conductivity of symmetric three-barrier RTS with the exact accounting of the applied constant electric field. 2. It is established that for the fixed geometrical configuration of RTS there exists one minimal magnitude of electric field intensity (equivalent to the energy shift $U_{0}$), at which the active element of QCL works in an optimal regime: the intensity of electromagnetic radiation and density of current flowing through the separate cascade of quantum laser become maximal at the same time. 3. It is shown that within the framework of the used models of effective masses and rectangular potential barriers without taking into account the electron-phonon, electron-electron interactions and other relaxation processes, at the fixed sizes of inner barrier and both wells of three-barrier RTS, the effectiveness of QCL operation exponentially increases both with the increase of outer barriers thicknesses as well with the increases of the magnitude of the shift due to the constant electric field. [99]{} Gmachl C., Capasso F., Sivco D.L., Cho A.Y., Rep. Prog. Phys., 2001, **64**, 1533;\ . Scalari G. et al., Appl. Phys. Lett., 2003, **82**, 3165; . Diehl L. et al., Appl. Phys. Lett., 2006, **88**, 201115; . Qi Jie Wang et al., Appl. Phys. Lett., 2009, **94**, 011103; . Hofstetter D., Beck M., Faist J., Appl. Phys. Lett., 2002, **81**, 2683; . Gendron L. et al., Appl. Phys. Lett., 2004, **85**, 2824; . Giorgetta F.R. et al., Appl. Phys. Lett., 2007, **91**, 111115; . Hofstetter D. et al., Appl. Phys. Lett., 2008, **93**, 221106; . Elesin V.F., JETP, 2005, **126**, 131 (in Russian). Elesin V.F., Kateev I.Ju., FTP, 2008, **42**, 586 (in Russian). Elesin V.F., Kateev I.Ju., Remnev M.A., FTP, 2009, **43**, 269 (in Russian). Pashkovskii A.B., JETP Lett., 2005, **82**, 228 (in Russian). Gelvich E.A., Golant E.I., Pashkovskii A.B., JETP Lett., 2006, **32**, 13 (in Russian). Pashkovskii A.B., JETP Lett., 2009, **89**, 32 (in Russian). Tkach N.V., Seti Yu.A., Low Temp. Phys., 2009, **35**, 556; . Tkach M., Seti Ju., Voitsekhivska O. and Fartushynsky R., AIP Conf. Proc., 2009, **1198**, 174; . Tkach M., Seti Ju., Ukr. J. Phys., 2009, **54**, 614 (in Ukrainian). Golant E.I., Pashkovskii A.B., FTP, 1994, **28**, 954 (in Russian). Golant E.I., Pashkovskii A.B., Tager A.S., FTP, **28**, 740 (in Russian). [^1]: E-mail: [email protected]

ArXiv

--- abstract: 'In split-supersymmetry (split-SUSY), gluino is a metastable particle and thus can freeze out in the early universe. The late decay of such a long-life gluino into the lightest supersymmetric particle (LSP) may provide much of the cosmic dark matter content. In this work, assuming the LSP is gravitino produced from the late decay of the metastable gluino, we examine the WMAP dark matter constraints on the gluino mass. We find that to provide the full abundance of dark matter, the gluino must be heavier than about 14 TeV and thus not accessible at the CERN Large Hadron Collider (LHC).' address: |  \ [*$^1$ Institute of Theoretical Physics, Academia Sinica, Beijing 100080, China*]{}\ \[2mm\] [*$^2$ CCAST(World Laboratory), P.O.Box 8730, Beijing 100080, China*]{}  \ author: - 'Fei Wang $^1$, Wenyu Wang $^1$, Jin Min Yang $^{2,1}$' title: Gravitino dark matter from gluino late decay in split supersymmetry --- =\#1,[[Nucl. Phys. B ]{}[**\#1**]{},]{} \#1,[[Phys. Lett. B ]{}[**\#1**]{},]{} \#1,[[Eur. Phys. J. C ]{}[**\#1**]{},]{} \#1,[[Phys. Rev. D ]{}[**\#1**]{},]{} \#1,[[Phys. Rev. Lett. ]{}[**\#1**]{},]{} \#1,[[Mod. Phys. Lett. A ]{}[**\#1**]{},]{} \#1,[[J. Phys. G]{}[**\#1**]{},]{} \#1,[[Commun. Theor. Phys. ]{}[**\#1**]{},]{} \#1[[tt$\backslash$\#1]{}]{} Introduction ============ In the recently proposed split-SUSY [@split], inspired by the need of fine-tuning for the cosmological constant, the authors argued that the fine-tuning problem in particle physics does not have to be solved by SUSY. The only phenomenological constraints on split-SUSY are then from the grand unification consideration as well as the dark matter consideration. As a result, the sfermion mass scale can be very high while the gaugino/Higgsino mass scale is still around the weak scale. While the split-SUSY has the obvious virtue of naturally solving the notorious SUSY flavor problem, it predicts that no sfermions are accessible at the CERN LHC collider. Thus if split-SUSY is indeed chosen by nature, the only way to reveal SUSY at the LHC is through gaugino or Higgsino productions, especially the gluino production [@gluino]. This makes it important to pre-examine the possible mass range of these particles before the running of the LHC. It is interesting that although the gauginos and Higgsinos in split-SUSY are required to be relatively light, they are recently found not necessarily below TeV scale from the grand unification and dark matter requirements [@senatore]. Actually, the grand unification requirement can allow a heavy gaugino mass as high as 18 TeV [@senatore]. If all gauginos and Higgsinos are above TeV scale, the LHC is doomed to find no SUSY particles except a light Higgs boson if split-SUSY is true. Although the split-SUSY consequence in the dark matter issue is also considered in the literature [@senatore; @pierce; @profumo], the authors focused on neutralino LSP or the usual NLSP decaying to the LSP during the BBN era, which is severely constrained by the Big Bang Nucleosythesis (BBN) and Cosmic Microwave Background (CMB) in split-SUSY. In this work we study the dark-matter consequence of the long-life gluino in split-SUSY. In the usual weak-scale SUSY, the LSP is usually assumed to be the lightest neutralino [^1]. Gluino decays rapidly into the LSP and thus cannot freeze out to cause any dark matter consequence. Only in case that gluino is quasi-degenerate with the neutralino LSP can it have dark matter consequence through gluino-neutralino co-annihilation [@coannihilation]. In split-SUSY, however, due to its long lifetime, gluino can freeze out before decaying and then decay slowly into the LSP, providing much of the cosmic dark matter content. So the gluino late decay is one characteristic of split SUSY. In this work, assuming the LSP is the gravitino (the so-called superWIMP dark matter) produced from the late decay of the metastable gluino, we will examine the WMAP dark matter constraints on the gluino mass. Note that if the gluino lifetime is too long (as long as BBN time), the released energy from its decay may spoil the BBN success and also affect CMB as well as large scale structure formation [@arvanitaki]. Therefore, in our study we require that the gluino decays before BBN. Calculations ============ The gluino relic density from thermal production can be calculated from the Boltzmann equation $$\frac{d n}{d t}= -3 H n -<\sigma v> (n^2-n_{eq}^2) ,$$ where $H$ is the Hubble constant, $n$ is the particle number density of gluino, $n_{eq}$ is the equilibrium density, and $<\sigma v>$ is the thermal averaged cross section of gluino annihilation. We can employ the freeze-out approximation technique to calculate the relic abundance. In split-SUSY the gluino pair annihilation proceeds through the $s$-channel gluon-exchange diagram and the $t$-channel gluino-exchange diagram, as shown in Fig. 1(a-c). The squark-exchange diagrams shown in Fig. 1(d) drop out since they are suppressed by the superheavy squark masses. The perturbation annihilation cross-section reads [@baer] $$\begin{aligned} \label{width1} \sigma\left(\tilde{g}\tilde{g}\to g g \right) & = & \frac{3 \pi \alpha_{s}^2}{16 {\beta}^2 s} \left\{ \log\frac{1+\beta}{1-\beta} \left[ 21-6{\beta}^2-3 {\beta}^4 \right] \right. \nonumber \\ && \left. -33\beta+17 {\beta}^3 \right\} , \\ \label{width2} \sigma\left(\tilde{g}\tilde{g}\to q \bar{q} \right)& = & \frac{ \pi \alpha_{s}^2 \bar{\beta} }{16 {\beta} s} (3-{\beta}^2)(3-{\bar{\beta}}^2) ,\end{aligned}$$ where $\beta=\sqrt{1-4m_{\tilde{g}}^2/s}$ and $\bar{\beta}=\sqrt{1-4m_{q}^2/s}$ with $m_{\tilde{g}}$ and $m_{q}$ being the gluino mass and quark mass, respectively. Multiple gluon exchanges between interacting $\tilde{g}$ will give rise to a Sommerfeld enhancement factor [@appelquist] $$E=\frac{ C \pi \alpha_{s} }{\beta} {\left[1-\exp\left\{- \frac{ C \pi \alpha_{s} }{\beta}\right\} \right]}^{-1} .$$ The interaction between the gluino and the goldstino (spin 1/2 component of gravitino) is suppressed by $1/F$ where $F$ characterizes the SUSY breaking scale. The gluino decay width to goldstino in split-SUSY is given by [@gambino] $$\Gamma_{\tilde{G}g}=\frac{m_{\tilde{g}}^3}{2 \pi F^2} {C_{5}^{\tilde{G}}}^2 ,$$ where $C_{5}^{\tilde{G}}=-m_{\tilde g}/2 \sqrt{2}$. Since the decay width is suppressed by $1/F$, not necessarily $1/M_{pl}$, the gluino decay can be arranged to occur before BBN by choosing the value of $F$. Note the conventional superWIMP dark matter scenario (gravitino is the LSP) is severely constrained by BBN and CMB [@feng] since the late decay of NLSP to LSP is assumed to occur at $10^6\sim 10^8$ second and the released energy may spoil the success of standard BBN. In our study we avoided such severe constraints since we require the gluino decays before BBN time. Furthermore, we also require the gluino decays before QCD era. Otherwise, R-hadrons could be formed and the R-hadron annihilation could destroy gluinos [@arvanitaki]. The relic density of the gravitino LSP from the late decay of gluino is given by $$\Omega_{\tilde{g}}\frac{m_{\tilde{G}}}{m_{\tilde{g}}} ,$$ with the gravitino mass given by $$m_{\tilde{G}}=\sqrt{\frac{8 \pi}{3}}\frac{F}{M_{pl}}.$$ Note that the gravitino can also be thermally produced [@relic] at the very early universe with temperature $T\sim M_{pl}$ (or even a bit lower) and then freeze out when temperature drops. But between the time of gravitino generation and now, the universe is expected to experience an inflation. Such an inflation would dilute the thermal relic density of gravitino. So we neglect the gravitino thermal production at the very early universe. In the context of inflation, the universe is expected to be reheated after inflation. The gravitino can be generated from reheating if the reheating temperature is high enough [@reheat]. In our study, we ignore such gravitino production by assuming the reheating temperature is not high enough to generate gravitino. The cosmic non-baryonic dark matter relic density can be obtained from the Wilkinson Microwave Anisotropy Probe (WMAP) measurements [@wmap] $$\begin{aligned} \Omega_m = 0.27_{-0.04}^{+0.04} \ , ~~ \Omega_b = 0.044_{-0.004}^{+0.004} \ ,\end{aligned}$$ where $\Omega_m$ is the total matter density and $\Omega_b$ is the baryonic matter density. Requiring the gravitino dark matter abundance from the gluino late decay is within the $2\sigma$ range of the WMAP data, we obtain the allowed parameter space in the plane of $M_{LSP}$ versus $M_{\tilde g}$, as shown in Fig. 2. We see from Fig. 2 that if the gravitino LSP from the gluino late-decay is to account for the whole gravitino dark matter content, the gluino has to be heavier than about 14 TeV and gravitino has to be lighter than about 16 TeV. A few remarks are in order regarding the above results. \(1) In our analyses the gluino is essentially assumed to be the next-to-lightest supersymmetric particle (NLSP). If it is not the NLSP, it would decay dominantly to the NLSP (say the neutralino $\tilde \chi^0_1$) followed by the decay of the NLSP to the gravitino LSP. In this case, although the gluino late decay also contributes to the dark matter content, its contribution is small compared to the freeze-out of neutralino NLSP. \(2) Although the gluino is assumed to be the NLSP, the lightest neutralino $\tilde \chi^0_1$ (assumed to be heavier than gluino) can still freeze out since its decay to gluino is suppressed by heavy squark mass in split-SUSY. Of course, the neutralino freeze-out happens much earlier than gluino freeze-out since its interaction is much weaker. Depending on the lifetime of the neutralino, its dark matter consequence can be quite different. If its decay to gluino happens before the freeze-out of gluino (corresponding to the relatively light squark mass), then the relic density of gluino is from the freeze-out, as assumed in our analyses. If its decay to gluino happens after the freeze-out of gluino (corresponding to the relatively heavy squark mass), then the relic density of gluino will be mainly from the neutralino decay. In such a case, a much stronger upper bound of about 2.2 TeV on the LSP mass obtained in [@profumo] should be applicable in order not to overclose the universe (note that in our case the upper bound of 2.2 TeV is for gravitino LSP mass and the upper bound on the neutralino mass can be relaxed since now the relic density of dark matter is given by $\Omega_{\tilde \chi^0_1} \frac{m_{\tilde G}}{m_{\tilde \chi^0_1}}$). Since our results are valid only in case that the neutralino decay to gluino happens before the freeze-out of gluino, we now examine the condition of this scenario. When the gluino is as heavy as 14 TeV, its freeze-out temperature is found to be about $m_{\tilde{g}}/30$ and the freeze-out time is thus about $10^{-9} sec$. The neutralino decays into gluino via exchanging a squark and its lifetime is sensitive to the forth power of squark mass: $$\tau_{\tilde \chi^0_1}=3\times 10^{-2} sec {\left(\frac{M_S}{10^9 GeV}\right)}^4 {\left( \frac{1 TeV}{ m_{\tilde \chi^0_1}}\right)}^5 ,$$ where $M_S$ is squark mass. For a neutralino at the order of 10 TeV, in order to let its lifetime shorter than gluino freeze-out time ($\sim 10^{-9} sec$), the squark mass can be chosen to be $M_S \sim 10^8$ GeV. \(3) Our analyses showed that the gluino (and all other gauginos or Higgsinos) must be heavier than about 14 TeV in order to provide the full dark matter abundance in the scenario we considered. It is interesting to check whether or not such a scenario is consistent with the gauge couplings unification at some high energy scale. In Fig. 3 we show the one-loop running of three gauge couplings for $M_3=14$ TeV, $M_2=20$ TeV and squark mass $M_S= 10^8$ GeV. Here, $M_3$ is gluino mass and, just like in ref.[@senatore], we assumed that Bino, Wino and Higgsino are all degenerate at the scale $M_2$. From Fig. 3 we see that starting from $M_Z$ scale [^2], $\alpha_1$ and $\alpha_2$ run up to higher energy scale and finally meet at a crosspoint at $\sim 10^{16}$ GeV. From this crosspoint $\alpha_s$ runs back to $M_Z$ scale and ends at $\alpha_s(M_Z)=0.098$. This value is welcome since, as pointed in [@split], the two-loop effects will enhance $\alpha_s$ at $M_Z$ scale by about 0.022. Taking into such effects, $\alpha_s(M_Z)$ in our scenario is just within the $2-\sigma$ range $0.119\pm 0.003$ [@data] allowed by experiments. \(4) Since in our scenario the gluino is the NLSP and all other gauginos are heavier than gluino, which is phenomenologically viable so far, the gauginos spectrum is different from that predicted by some theoretically favored models like mSUGRA. In the popular mSUGRA models, for example, the colored gluino is predicted to be heavier than other gauginos at the weak scale. However, such fancy models may not be chosen by nature and phenomenologically we should not be restricted to them. \(5) If this dark matter scenario (LSP is gravitino produced from the late decay of the metastable gluino in split-SUSY) is indeed chosen by nature, then no super particles of split-SUSY can be found at the LHC except a light Higgs boson whose mass is upper bounded by about 150 GeV [@split] [^3]. Conclusion ========== The metastable gluino in split-SUSY can freeze out in the early universe and then decay slowly into the LSP, providing much of the cosmic dark matter content. If the LSP is the gravitino produced from the late decay of the metastable gluino, we found that the dark matter consideration can constrain the parameter space of the gluino mass versus the gravitino mass: in order to provide the full abundance of dark matter, the gluino must be heavier than 14 TeV. Therefore, if nature takes this choice for dark matter, no gauginos or Higgsinos are accessible at the LHC. Then no super particles of split-SUSY can be found at the LHC except a light Higgs boson. Acknowlegement {#acknowlegement .unnumbered} ============== We thank Prof. Bing-Lin Young, Dr. Guangping Gao, Guoli Liu and Fuqiang Xu for discussions. This work is supported in part by National Natural Science Foundation of China. [11]{} N. Arkani-Hamed, S. Dimopoulos, hep-th/0405159; G.F. Giudice, A. Romanino, 699, 65 (2004). N. Arkani-Hamed, S. Dimopoulos, G.F. Giudice, A. Romanino, 709, 3 (2005). J. L. Hewett, B. Lillie, M. Masip, T. G. Rizzo, JHEP 0409, 070 (2004); A. Arvanitaki, S. Dimopoulos, A. Pierce, S. Rajendran, J. Wacker, hep-ph/0506242; J. G. Gonzalez, S. Reucroft, J. Swain, hep-ph/0504260; M. Toharia, J. D. Wells, hep-ph/0503175; K. Cheung, W.-Y. Keung, 71, 015015 (2005); A. C. Kraan, 37, 91 (2004); W. Kilian, T. Plehn, P. Richardson, E. Schmidt, 39, 229 (2005). L. Senatore, 71, 103510 (2005); A. Masiero, S. Profumo, P. Ullio, 712, 86 (2005). A. Pierce, 70, 075006 (2004); A. Arvanitaki, P. W. Graham, hep-ph/0411376. A. Masiero, S. Profumo, P. Ullio, 712, 86 (2005). S. Profumo, C. E. Yaguna, 69, 115009 (2004); 70, 095004 (2004). H. Baer, K. Cheung, J. F. Gunion, 59, 075002 (1999). S. Wolfram, 82, 65 (1979); C. B. Dover, T.K. Gaisser, G. Steigman, 42, 1117 (1979); G. D. Starkman, A. Gould, R. Esmailzadeh, S. Dimopoulos, 41, 3594 (1990) . S. Raby, 56, 2852 (1997); hep-ph/9712254; R. N. Mohapatra, S. Nussinov, 57, 1940 (1997). A. Arvanitaki, [*et al.*]{}, hep-ph/0504210; A. G. Doroshkevich, M. Y. Khlopov, Sov. J. Nucl. Phys. [**39**]{}, 551 (1984); A. G. Doroshkevich, M. Y. Khlopov, M. N. Roy, Astr. Soc. [**221**]{}, 279 (1984); A. G. Doroshkevich, M. Y. Khlopov, A. A. Klypin, M. N. Roy, Astr. Soc. [**239**]{}, 923 (1989). T. Appelquist, H. D. Politzer, 34, 43 (1975); S. J. Brodsky, J. F. Gunion, D. E. Soper, 36, 2710 (1987). P. Gambino, G. F. Giudice, P. Slavich, hep-ph/0506214. J. L. Feng, A. Rajaraman, F. Takayama, 91, 011302 (2003); 68, 063504 (2003); J. L. Feng, S. Su, F. Takayama, hep-ph/0404198; hep-ph/0404231; F. Wang, J. M. Yang, 38, 129 (2004); K. Hamaguchi, Y. Kuno, T. Nakaya, M. M. Nojiri, 70, 115007 (2004). H. Pagels, J. R. Primack, 48, 223 (1982). L. M. Krauss, 227, 556 (1983); D. V. Nanopoulos, K. A. Olive, M. Srednicki, 127, 30 (1983); T. Moroi, H. Murayama, M. Yamaguchi, 303, 289 (1993); M. Y. Khlopov, A. D. Linde, 138, 265 (1984); J. R. Ellis, J. E. Kim, D. V. Nanopoulos, 145, 181 (1984); R. Juszkiewicz, J. Silk, A. Stebbins, 158, 463 (1985); M. Bolz, A. Brandenburg, W. Buchmuller, 606, 518 (2001). WMAP Collaboration, Astrophys. J. Suppl. [**148**]{}, 1 (2003); [**148**]{}, 175 (2003); [**248**]{}, 195 (2003). G. Altarelli, M.W. Grunewald, hep-ph/0404165. J. Cao, J. M. Yang, 71, 111701 (2005). [^1]: If gluino is the LSP, its relic abundance is severely constrained by the bounds from existing anomalous heavy isotope abundances [@baer; @raby]. [^2]: The starting values of $\alpha_1$ and $\alpha_2$ at $M_Z$ scale is fixed by $\alpha^{-1}(M_Z)=128.936\pm 0.0049$ and $\sin^2\theta_W(M_Z)=0.23150\pm 0.00016$[@data]. [^3]: This bound may be lowered by a few tens of GeV if right-handed neutrinos are introduced with see-saw mechanism[@cao].

ArXiv

--- abstract: 'Let $A$ be a hereditary algebra. We construct a fundamental domain for the cluster category ${\mathcal{C}_A}$ inside the category of modules over the duplicated algebra ${\overline{A}}$ of $A$. We then prove that there exists a bijection between the tilting objects in ${\mathcal{C}_A}$ and the tilting ${\overline{A}}$-modules all of whose non projective-injective indecomposable summands lie in the left part of the module category of ${\overline{A}}$.' address: - 'Département de Mathématiques, Université de Sherbrooke, Sherbrooke (Québec), J1K 2R1, Canada' - 'Département de Mathématiques, Université de Sherbrooke, Sherbrooke (Québec), J1K 2R1, Canada *and* Department of Mathematics, Bishop’s University, Lennoxville, (Québec), J1M 1Z7, Canada' - 'Department of Mathematics and Statistics, University of Massachusetts at Amherst, Amherst, MA 01003-9305, USA' - 'Department of Mathematics, Northeastern University, Boston, MA 02115, USA' author: - 'I. Assem' - 'T. Brüstle' - 'R. Schiffler' - 'G. Todorov' title: Cluster categories and duplicated algebras --- , , , , cluster category, tilting, duplicated algebra [Introduction]{} Cluster categories were introduced in [@BMRRT], and for type $A_n$ also in [@CCS], as a means for a better understanding of the cluster algebras of Fomin and Zelevinsky [@FZ1; @FZ2]. The indecomposable objects (without self-extensions) in the cluster category correspond to the cluster variables in the cluster algebra and the tilting objects in the cluster category to the clusters in the cluster algebra. Our objective in this note is to give an interpretation of the cluster category and its tilting objects in terms of modules over a finite dimensional algebra. Indeed, let $A$ be a hereditary algebra over an algebraically closed field then, by Happel’s theorem [@H], the derived category of bounded complexes over the category $\textup{mod}\, A$ of finitely generated right $A$-modules is equivalent to the stable module category over the repetitive algebra ${\hat{A}}$ of $A$ (in the sense of Hughes and Waschbüsch [@HW]). The algebra ${\hat{A}}$ is infinite dimensional but, in order to study the cluster category, it suffices to look at a finite dimensional quotient of ${\hat{A}}$, namely the duplicated algebra ${\overline{A}}$ of $A$ defined and studied in [@A; @ANS]. The resulting embedding of $\textup{mod}\, {\overline{A}}$ into $\textup{mod} \,{\hat{A}}$ induces a functor $\bar{\pi}$ from $\textup{mod} \,{\overline{A}}$ to the cluster category ${\mathcal{C}_A}$ of $A$. We prove that the functor $\bar{\pi}$ induces a one-to-one correspondence between the indecomposable objects in the cluster category and the non projective-injective ${\overline{A}}$-modules lying in the left part ${\mathcal{L}_{\overline{A}}}$ of $\textup{mod}\,{\overline{A}}$, in the sense of Happel, Reiten and Smalø  [@HRS] (we then say that ${\mathcal{L}_{\overline{A}}}$ is an [*exact fundamental domain*]{} for the functor $\bar{\pi}$). This opens the way to our main result. Let $A$ be a hereditary algebra. There exists a one-to-one correspondence between the multiplicity-free tilting objects in the cluster category ${\mathcal{C}_A}$ of $A$ and the multiplicity-free tilting ${\overline{A}}$-modules such that all non projective-injective indecomposable summands of $T$ lie in ${\mathcal{L}_{\overline{A}}}$. This correspondence is given explicitly as follows. Since any indecomposable projective-injective ${\overline{A}}$-module is necessarily a summand of $T$, then $T=T_0\oplus \overline P$, where $\overline P$ is a uniquely determined projective-injective ${\overline{A}}$-module and $T_0$ has no projective-injective summands. If all the indecomposable summands of $T_0$ lie in ${\mathcal{L}_{\overline{A}}}$, then $\bar\pi(T_0)$ is a tilting object in ${\mathcal{C}_A}$ and conversely, any tilting object in ${\mathcal{C}_A}$ is of this form. Since duplicated algebras appear as a perfect context to view (cluster-)tilting objects as actual tilting *modules*, we investigate these algebras further. In particular we show that the simply-laced Dynkin case corresponds to representation-finite duplicated algebras, which, in addition, are simply connected. In this case several techniques are known for computing the tilting modules, allowing us to find the clusters in the corresponding cluster algebra. We now describe the contents of our paper. After a brief preliminary section, devoted to fixing the notation and recalling the main facts we shall be using, the second section contains a detailed description of the left part ${\mathcal{L}_{\overline{A}}}$. In the third section, we prove that ${\mathcal{L}_{\overline{A}}}$ is an exact fundamental domain for the natural functor and we prove our main result in section four. Our final section is devoted to deduce related properties of the duplicated algebra. [Preliminaries]{} [Notation.]{} Throughout this paper, we let $A$ denote a hereditary algebra over an algebraically closed field $k$. We denote by $\textup{mod} \,A$ the category of finitely generated right $A$-modules and by ${\textup{ind}\,}A$ a full subcategory whose objects are representatives of the isomorphism classes of indecomposable objects in $\textup{mod} \,A$. The derived category of bounded complexes over $\textup{mod}\, A$ will be denoted by ${\mathcal{D}^b(\textup{mod}\, A)}$. For a vertex $x$ in the quiver ${\mathcal{Q}_A}$ of $A$, we write $e_x$ for the corresponding primitive idempotent and $S_x,P_x,I_x$, respectively, for the corresponding simple, indecomposable projective and indecomposable injective $A$-module. The functor $D={\textup{Hom}\,}_k(-,k)$ is the standard duality between $\textup{mod}\, A$ and $\textup{mod}\, A^{\textup{op}}$, and $\tau_A=D\,Tr$, $ \tau^{-1}_A=Tr\,D$ are the Auslander-Reiten translations in $\textup{mod}\, A$. We refer to [@ARS] for further facts about $\textup{mod}\, A$, and to [@R] for the tilting theory of $\textup{mod}\, A$. [The cluster category ${\mathcal{C}_A}$.]{} The cluster category ${\mathcal{C}_A}$ of $A$ is defined as follows. Let $F$ denote the endofunctor of ${\mathcal{D}^b(\textup{mod}\, A)}$ defined as the composition $\tau^{-1}[1]$, where $\tau $ is the Auslander-Reiten translation in ${\mathcal{D}^b(\textup{mod}\, A)}$ and $[1]$ is the shift functor. Then ${\mathcal{C}_A}$ is the quotient category ${\mathcal{D}^b(\textup{mod}\, A)}/F$. Its objects are the $F$-orbits of objects in ${\mathcal{D}^b(\textup{mod}\, A)}$ and the morphisms are given by $${\textup{Hom}\,}_{{\mathcal{C}_A}}(\tilde X, \tilde Y) = \oplus_{i\in \Zset} \ {\textup{Hom}\,}_{{\mathcal{D}^b(\textup{mod}\, A)}}(F^iX,Y)$$ where $X$ and $Y$ are objects in ${\mathcal{D}^b(\textup{mod}\, A)}$ and $\tilde X, \tilde Y$ are their respective $F$-orbits. It is shown in [@K] that ${\mathcal{C}_A}$ is a triangulated category. Furthermore, the canonical functor ${\mathcal{D}^b(\textup{mod}\, A)}\to {\mathcal{C}_A}$ is a functor of triangulated categories. We refer to [@BMRRT] for facts about the cluster category. [The duplicated algebra $ \overline{A} $]{} The duplicated algebra of a hereditary algebra $A$ is the matrix algebra $$\overline{A}=\left[ \begin{array}{cc} \ A\ &\ 0\ \\ DA & A \end{array}\right]= \left\{\left[\left. \begin{array}{cc} \ a\ &\ 0\ \\ q&b \end{array}\right] \ \right|\ {a,b\in A ,\atop q \in DA}\right\}$$ with the ordinary matrix addition and the multiplication induced by the bimodule structure of $DA$. Writing $1$ for the identity of $A$, and setting $$e=\left[{\begin{array}{cc} \ 1\ &\ 0\ \\0&0\end{array}}\right] \quad \textup{and} \quad e'=\left[\begin{array}{cc} \ 0\ &\ 0\ \\0&1\end{array}\right],$$ we see that ${\overline{A}}$ contains two copies of $A$ given respectively by $e {\overline{A}}e$ and by $e'{\overline{A}}e'$. In order to distinguish between these we denote the first one by $A$ and the second one by $A'$. Accordingly, ${\mathcal{Q}_A}'$ denotes the quiver of $A'$, $x'$ the vertex of ${\mathcal{Q}_A}'$ corresponding to $x\in ({\mathcal{Q}_A})_0$, and $e'_x $ the corresponding idempotent. Let $\overline{S}_x,\overline{P}_x,\overline{I}_x$ denote respectively the simple, indecomposable projective and indecomposable injective module in $\textup{mod}\, {\overline{A}}$ corresponding to $x\in({\mathcal{Q}_A}\cup{\mathcal{Q}_A}')_0$. The ordinary quiver ${\mathcal{Q}_{\overline{A}}}$ of ${\overline{A}}$ is constructed as follows. It contains ${\mathcal{Q}_A}$ and ${\mathcal{Q}_A}'$ as full convex connected subquivers and every vertex of ${\mathcal{Q}_{\overline{A}}}$ lies in either ${\mathcal{Q}_A}$ or ${\mathcal{Q}_A}'$. There is an arrow $x'\to y$ whenever ${\textup{rad}\,}(e'_x{\overline{A}}e_y)/ {\textup{rad}\,}^2(e'_x{\overline{A}}e_y) \ne 0$. Observe that $e'_x{\overline{A}}e_y=D(e_y A e_x)$ and therefore, if $e_y A e_x \ne 0$ then there is a non-zero path in ${\mathcal{Q}_{\overline{A}}}$ from $x'$ to $y$. Also, since $e'_x {\overline{A}}{\cong}D({\overline{A}}e_x)$, each $\overline{I}_x=\overline{P}_{x'}$ is projective-injective having $S_x$ as a socle and $S_{x'}$ as a top. On the other hand, each $\overline{P}_x$ has its support lying in ${\mathcal{Q}_A}$ and is therefore equal to the projective $A$-module $P_x$. Dually, $\overline{I}_{x'}$ has its support lying completely in ${\mathcal{Q}_A}'$ and equals the injective $A'$-module $I_{x'}$. For facts about the duplicated algebra, we refer to [@A; @ANS]. [The repetitive algebra ${\hat{A}}$]{}\[section\] For our purposes, another description of ${\overline{A}}$ is needed. The repetitive algebra ${\hat{A}}$ of the hereditary algebra $A$, is the infinite matrix algebra $${\hat{A}}\quad = \quad \left[ \begin{array}{cccccccccc} \ddots &&&\ 0\ \\ &\ A_{m-1}\ \\ &Q_m&\ A_m\ \\ &&Q_{m+1}&\ A_{m+1}\ \\ &\ 0\ &&&\ddots \end{array} \right]$$ where matrices have only finitely many non-zero coefficients, $A_m=A$ and $Q_m= {_A\!DA _A}$ for all $m\in \Zset$, all the remaining coefficients are zero and multiplication is induced from the canonical isomorphisms $A\otimes_A DA{\cong}{_A\!DA _A}{\cong}DA\otimes_A A$ and the zero morphism $DA\otimes_A DA \to 0$, see [@HW]. Then ${\overline{A}}$ is identified to the quotient algebra of ${\hat{A}}$ defined by the surjection $${\hat{A}}\quad \to \quad \left[ \begin{array}{cc} A_0&0\\ Q_1&A_1 \end{array}\right].$$ This identification yields an embedding functor $\textup{mod}\, {\overline{A}}{\hookrightarrow}\textup{mod}\, {\hat{A}}$. Similarly, the canonical surjection ${\overline{A}}\to e{\overline{A}}e=A$ yields an embedding functor $\textup{mod}\, A {\hookrightarrow}\textup{mod}\, {\overline{A}}$. Our first objective will be to look more closely at these embeddings. [The left part ${\mathcal{L}_{\overline{A}}}$ of the duplicated algebra ${\overline{A}}$]{} [Definitions and a preparatory lemma]{} Let $C$ be any finite dimensional $k$-algebra, and $M,N$ be two indecomposable $C$-modules. A [*path*]{} from $M$ to $N$ in ${\textup{ind}\,}C$ is a sequence of non-zero morphisms $$\label{eq*} M=M_0\stackrel{f_1\,}{\to} M_1 \stackrel{f_2\,}{\to} \cdots \stackrel{f_t\,}{\to} M_t =N$$ with all $M_i$ in ${\textup{ind}\,}C$. We denote such a path by $M\leadsto N$ and say that $M$ is a [*predecessor*]{} of $N$ (or that $N$ is a [*successor*]{} of $M$). When each $f_{i}$ in (\[eq\*\]) is irreducible, we say that (\[eq\*\]) is a [ *path of irreducible morphisms*]{}. A path (\[eq\*\]) of irreducible morphisms is [*sectional*]{} if $\tau_C\ M_{i+1}\ne M_{i-1}$ for all $i$ with $1\le i\le t$. A [*refinement*]{} of (\[eq\*\]) is a path in ${\textup{ind}\,}C$: $$M=M'_0\stackrel{f_1'\,}{\to} M_1' \stackrel{f_2'\,}{\to} \cdots \stackrel{f_s'\,}{\to} M'_s =N$$ with $s\ge t$ such that there exists an order-preserving injection ${\sigma}: \{1,\ldots,t-1\}\to \{1,\ldots , s-1\}$ verifying $M_{i} = M'_{{\sigma}(i)}$ for all $i$ with $1\leq i\le t$. A full subcategory $\mathcal{C}$ of ${\textup{ind}\,}{C}$ is called [*convex*]{} in ${\textup{ind}\,}C$ if, for any path (\[eq\*\]) from $M$ to $N$ in ${\textup{ind}\,}C$, with $M,N$ lying in $\mathcal{C}$, all the $M_i$ lie in $\mathcal{C}$. Useful examples of convex subcategories arise from the standard embeddings $\textup{mod}\, A {\hookrightarrow}\textup{mod}\, {\overline{A}}$ and $\textup{mod}\, {\overline{A}}{\hookrightarrow}\textup{mod}\, {\hat{A}}$, as seen in \[section\] above. We have the following lemma (see [@A 2.5], [@T 3.4,3.5] or [@Y 4.1]), which will be used quite often when considering $A$-modules as ${\overline{A}}$-modules or ${\hat{A}}$-modules. \[lemma1\] a) The embeddings $ \textup{mod}\, A {\hookrightarrow}\textup{mod}\, {\overline{A}}$ and $\textup{mod}\, {\overline{A}}{\hookrightarrow}\textup{mod}\, {\hat{A}}$ are full, exact and preserve indecomposable modules, almost split sequences and irreducible morphisms.\ b) Under these embeddings, ${\textup{ind}\,}A$ is a full convex subcategory of ${\textup{ind}\,}{\overline{A}}$, closed under predecessors, and ${\textup{ind}\,}{\overline{A}}$ is a full convex subcategory of ${\textup{ind}\,}{\hat{A}}$. [The left part]{} Let again $C$ be a finite dimensional algebra. Following Happel, Reiten and Smalø  [@HRS], we define the [*left part*]{} ${\mathcal{L}_{{C}}}$ of $\textup{mod}\,C$ to be the full subcategory of $\textup{mod}\,C$ consisting of all indecomposable $C$-modules such that if $L\leadsto M$, then the projective dimension ${\textup{pd}\,}L$ of $L$ is at most one. The right part ${\mathcal{R}_{{C}}}$ is defined dually. Our objective now is to compute the left part of the module category of the duplicated algebra ${\overline{A}}$ of a hereditary algebra $A$. We start by observing that, by Lemma \[lemma1\], the complete slice of the Auslander-Reiten quiver ${\Gamma}(\textup{mod}\, A)$ of $A$ consisting of the indecomposable injective $A$-modules embeds fully inside ${\Gamma}(\textup{mod}\, {\overline{A}})$. The sources in this slice are the injectives $I_a$ with $a$ a sink in ${\mathcal{Q}_A}$. For each sink $a$ in ${\mathcal{Q}_A}$, the injective $A$-module $I_a$ is the radical of the projective-injective ${\overline{A}}$-module ${\overline{I}}_a={\overline{P}}_{a'}$. We recall that for any algebra $C$ and any $L$ in $\textup{mod}\, C$, ${\textup{pd}\,}L\le 1$ if and only if ${\textup{Hom}\,}_{C}(DC,\tau_C L)=0$  (see [@ARS IX.1.7, p.319] or [@R p.79]). \[lemma2\] Let $M$ be an indecomposable ${\overline{A}}$-module. Then: 1. If $M$ belongs to ${\textup{ind}\,}A$, then $M\in {\mathcal{L}_{\overline{A}}}$ and $\tau^{-1}_{{\overline{A}}}M \in {\mathcal{L}_{\overline{A}}}$. 2. If $M$ does not belong to ${\textup{ind}\,}A$, then there exist a sink $a\in ({\mathcal{Q}_A})_0$ and a path $\overline{P}_{a'} \leadsto M$. a\) Any $A$-module $M$ admits a projective resolution in $\textup{mod}\, A$ of the form $$0\to P_1 \to P_0\to M \to 0$$ with $P_0$ and $P_1$ projective $A$-modules, hence projective ${\overline{A}}$-modules. Thus the projective dimension of $M$ as an $A$-module and also as an ${\overline{A}}$-module is at most one. This shows that ${\textup{ind}\,}A \subset {\mathcal{L}_{\overline{A}}}$, because ${\textup{ind}\,}A$ is closed under predecessors. To see that $\tau_{{\overline{A}}}^{-1}M$ is in ${\mathcal{L}_{\overline{A}}}$, notice that, since $M$ is in ${\textup{ind}\,}A$, ${\textup{Hom}\,}_{{\overline{A}}}({\overline{I}}_{x},M)= 0$ for all injective ${\overline{A}}$-modules ${\overline{I}}_{x}$. So ${\textup{pd}\,}_{{\overline{A}}}(\tau_{{\overline{A}}}^{-1}M)\le 1$ by the above remark. Furthermore, any non-projective predecessor $L$ of $\tau_{{\overline{A}}}^{-1}M$ lies in ${\textup{ind}\,}A \cup\tau_{{\overline{A}}}^{-1}({\textup{ind}\,}A)$, hence ${\textup{pd}\,}L\le 1$. b\) Assume now that $M$ is not in ${\textup{ind}\,}A$. Then there exists $b\in ({\mathcal{Q}_A})_0$ such that ${\textup{Hom}\,}_{{\overline{A}}}(\overline {P}_{b'}, M) \ne 0$. If $b$ is a sink, we are done. If not, consider the projective $A$-module $P_{b}$. Let $S_{a}$ be a simple submodule of $P_{b}$. Note that $S_{a}$ is projective since $A$ is hereditary. Therefore $S_{a}= P_{a}$ and $a$ is a sink. Then ${\textup{Hom}\,}_{A}(P_{a},P_{b})\ne0$ implies ${\textup{Hom}\,}_{A}(I_{a},I_{b})\ne0$, which induces a non-zero morphism $\overline{P}_{a'}=\overline{I}_{a}\to \overline{I}_{b}=\overline{P}_{b'}$ of ${\overline{A}}$-modules. This yields the required path $\overline{P}_{a'} \leadsto M$. [A characterization of the modules in ${\mathcal{L}_{\overline{A}}}$]{} Before stating the next proposition, we recall that, by [@AC 1.6], ${\mathcal{L}_{\overline{A}}}$ consists of all $M\in {\textup{ind}\,}{\overline{A}}$, such that, if there exists a path from an indecomposable injective module to $M$, then this path can be refined to a path of irreducible morphisms, and any such refinement is sectional. \[proposition3\] An indecomposable ${\overline{A}}$-module $M$ is in ${\mathcal{L}_{\overline{A}}}$ if and only if, whenever there exists a path $\overline{P}_{a'} \leadsto M$, with $a$ a sink in $({\mathcal{Q}_A})_0$, this path can be refined to a path of irreducible morphisms, and each such refinement is sectional. Since the necessity follows directly from the above statement, we only prove the sufficiency. Assume that $M$ satisfies the stated condition. In order to prove that $M\in{\mathcal{L}_{\overline{A}}}$, it suffices to show that, if there exists a path ${\overline{I}}_{x}\leadsto M$, with ${\overline{I}}_{x}$ injective in $\textup{mod}\, {\overline{A}}$, then this path can be refined to a path of irreducible morphisms, and any such refinement is sectional. Since ${\overline{I}}_{x}$ is not an $A$-module, it follows from Lemma \[lemma2\] b), that there exist a sink $a$ in ${\mathcal{Q}_A}$ and a path ${\overline{P}}_{a'}\leadsto{\overline{I}}_{x}$, giving a path ${\overline{P}}_{a'}\leadsto{\overline{I}}_{x}\leadsto M$. The conclusion follows at once. [Ext-injectives in ${\mathcal{L}_{\overline{A}}}$]{} We now characterize the Ext-injectives in the additive full subcategory ${\textup{add}\,}{\mathcal{L}_{\overline{A}}}$ of $\textup{mod}\, {\overline{A}}$ generated by the left part. We recall from [@AS] that, if ${\mathcal A}$ is an additive full subcategory of $\textup{mod}\, {\overline{A}}$, closed under extensions, then an indecomposable module $M$ in ${\mathcal A}$ is called an Ext-$injective$ $in$ $ \mathcal A$ if ${\textup{Ext}^1}_{{\overline{A}}}(\quad ,M)\vert_{\mathcal A} = 0$. It is known that $M$ is Ext-injective in ${\textup{add}\,}{\mathcal{L}_{\overline{A}}}$ if and only if $\tau_{{\overline{A}}}^{-1} M$ is not in ${\textup{add}\,}{\mathcal{L}_{\overline{A}}}$ (see [@AS 3.4]). We denote by ${\Sigma}$ the set of all indecomposable Ext-injectives in ${\textup{add}\,}{\mathcal{L}_{\overline{A}}}$. The following corollary says that ${\mathcal{L}_{\overline{A}}}= {\textup{ind}\,}A \cup {\Sigma}$. \[cor1\] The following are equivalent for an ${\overline{A}}$-module $M$:\ a) $M$ is in ${\Sigma}$.\ b) $M$ is in ${\mathcal{L}_{\overline{A}}}$ and $M$ is not in ${\textup{ind}\,}A$.\ c) $M$ is in ${\mathcal{L}_{\overline{A}}}$ and there exist a sink $a\in({\mathcal{Q}_A})_0$ and a path $\overline{P}_{a'}\leadsto M$.\ d) There exist a sink $a\in({\mathcal{Q}_A})_0$ and a path $\overline{P}_{a'}\leadsto M$ and any such path is refinable to a sectional path. \ a) implies b) since ${\textup{ind}\,}A \cup \tau^{-1}({\textup{ind}\,}A) \subset {\mathcal{L}_{\overline{A}}}$ by Lemma \[lemma2\] a).\ b) implies c) follows from Lemma \[lemma2\] b).\ c) implies d) follows from Proposition \[proposition3\].\ d) implies a) Proposition \[proposition3\] implies that $M$ is in ${\mathcal{L}_{\overline{A}}}$. The fact that there exist a sink $a\in({\mathcal{Q}_A})_0$ and a path $\overline{I}_{a}=\overline{P}_{a'}\leadsto M$ (hence a sectional path), implies that ${\textup{Hom}\,}_{{\overline{A}}}(\overline{I}_{a},M) \ne 0$ by [@ARS III.2.4, p.239]. By the remark before Lemma \[lemma2\], it follows that ${\textup{pd}\,}_{\overline{A}}(\tau_{{\overline{A}}}^{-1}M) \ge 2$ and therefore $\tau_{{\overline{A}}}^{-1}M$ is not in ${\mathcal{L}_{\overline{A}}}$. \[cor2\] The set ${\Sigma}$ of all indecomposable Ext-injectives in ${\textup{add}\,}{\mathcal{L}_{\overline{A}}}$ consists of all the projective-injectives lying in ${\mathcal{L}_{\overline{A}}}$ as well as all the modules of the form $\tau^{-1} I_x$ with $ x \in ({\mathcal{Q}_A})_0$, that is $${\Sigma}= \{ \tau^{-1}_{{\overline{A}}} I_{x}\mid {x\in (Q_{A})_{0}}\} \cup \{{\overline{P}}_{x'}\mid {\overline{P}}_{x'}\in {\mathcal{L}_{\overline{A}}}\}.$$ Clearly, projective-injective modules which lie in ${\mathcal{L}_{\overline{A}}}$ belong to ${\Sigma}$. Now let $x\in ({\mathcal{Q}_A})_0$ and consider $ \tau^{-1}_{{\overline{A}}} I_{x}$. Let $a$ be a sink and $I_{a}$ be a maximal indecomposable injective $A$-module such that there is an epimorphism $I_{a} \to I_{x}$. Then there is a non-zero map $\tau_{{\overline{A}}}^{-1}I_{a} \to \tau_{{\overline{A}}}^{-1}I_{x}$ and therefore a path ${\overline{I}}_{a} \to \tau_{{\overline{A}}}^{-1}I_{a} \to \tau_{{\overline{A}}}^{-1}I_{x}$. Since $\tau_{{\overline{A}}}^{-1}I_{x}$ is in ${\mathcal{L}_{\overline{A}}}$, it follows that $\tau_{{\overline{A}}}^{-1}I_{x}$ is in ${\Sigma}$ by Corollary \[cor1\] d). Conversely, suppose $X$ belongs to ${\Sigma}$, but is not a projective-injective lying in ${\mathcal{L}_{\overline{A}}}$. By Corollary \[cor1\], there exists a sink $a$ and a sectional path in ${\textup{ind}\,}{\overline{A}}$ $${\overline{P}}_{a'}={\overline{I}}_{a}= M_{0}\to M_{1} \to \ldots \to M_{t} = X$$ with $t\ge 1$ and $M_1= {\overline{I}}_a/ S_a = \tau^{-1}_{{\overline{A}}}\,I_a$. We claim that no $M_i$ (with $i\ge 1$) is a projective ${\overline{A}}$-module. Indeed, assume first that $M_i$ (with $i\ge 1$) is projective-injective. By hypothesis, $i<t$. Then $M_{i-1}={\textup{rad}\,}M_i$ and $M_{i+1}= M_i/{\textup{soc}\,}M_i = \tau^{-1}_{{\overline{A}}}\, M_{i-1}$, contradicting the sectionality of the above path. On the other hand, for any $i\le t$, ${\textup{Hom}\,}_{{\overline{A}}}({\overline{P}}_{a'},M_i)\ne 0$ hence $M_i$ is not an $A$-module, and a fortiori not projective in $\textup{mod}\, A$. This establishes our claim. We infer the existence of a sectional path in ${\textup{ind}\,}{\overline{A}}$ $$I_a=\tau_{{\overline{A}}}\, M_{1} \to \tau_{{\overline{A}}}\, M_2 \ldots \to \tau_{{\overline{A}}}\, M_{t} = \tau_{{\overline{A}}}\, X .$$ Since $X\in{\mathcal{L}_{\overline{A}}}$, then, for any $i\le t$, $M_i\in {\mathcal{L}_{\overline{A}}}$ and so ${\textup{pd}\,}M_i\le 1$ implying that ${\textup{Hom}\,}_{{\overline{A}}}({\overline{P}}_{x'},\tau_{{\overline{A}}}\, M_i)=0$ for any $x\in({\mathcal{Q}_A})_0$. This shows that the above path lies entirely in $\textup{mod}\, A$. Since $I_a$ is injective, all the modules on it are injective. In particular, there exists $x\in({\mathcal{Q}_A})_0$ such that $ \tau_{{\overline{A}}}\, X= I_x$. We now give another expression for the set of all indecomposable Ext-injectives in ${\textup{add}\,}{\mathcal{L}_{\overline{A}}}$. For this, we need to recall that, if $M$ is an ${\overline{A}}$-module, then its first cosyzygy ${\Omega}_{{\overline{A}}}^{-1}\,M$ is the cokernel of an injective envelope $M\to {\overline{I}}$ in $\textup{mod}\,{\overline{A}}$. \[prop1\] Let $x\in (Q_{A})_{0}$. Then ${\Omega}^{-1}_{{\overline{A}}} \,P_{x} {\cong}\tau^{-1}_{{\overline{A}}} I_{x}$. Consequently, $${\Sigma}= \{ {\Omega}^{-1}_{{\overline{A}}} P_{x}\mid {x\in (Q_{A})_{0}}\} \cup \{{\overline{P}}_{x'}\mid {\overline{P}}_{x'}\in {\mathcal{L}_{\overline{A}}}\}.$$ We prove this by induction on the Loewy length of the projective module $P_{x}$. Recall that the Loewy length of a module $M$ is the smallest integer $i$ with ${\textup{rad}\,}^i M =0$. Let $P_{a}= S_{a}$ be a simple projective module. Then ${\Omega}^{-1}_{{\overline{A}}} P_{a} {\cong}{\overline{I}}_{a}/ S_{a}$. On the other hand, from the almost split sequence: $$0 \to I_{a} \to {\overline{I}}_{a} \oplus I_{a}/S_{a} \to {\overline{I}}_{a}/S_{a} \to 0$$ it follows that ${\Omega}_{{\overline{A}}}^{-1}\,P_{a} {\cong}\tau_{{\overline{A}}}^{-1} I_{a}$ for any sink $a$, which proves our claim in this case. For an indecomposable non-simple projective $P_{x}$ let the radical be ${\textup{rad}\,}P_{x}= \oplus\, P_{y_{i}}$. Then there are the following isomorphisms of the injective envelopes: $I_{0}\, (P_{x}) = I_{0}\,({\textup{rad}\,}P_{x}) {\cong}\oplus\, I_{0}\,(P_{y_{i}})$. Then ${\Omega}_{{\overline{A}}}^{-1}(P_{x}) = I_{0}\,(P_{x})/ P_{x}$ and ${\Omega}_{{\overline{A}}}^{-1} ({\textup{rad}\,}P_{x}) = \oplus\, {\Omega}_{{\overline{A}}}^{-1}(P_{y_{i}}){\cong}I_{0}\,(P_{x})\, /\,(\oplus \,P_{y_{i}})$. A simple application of the snake lemma yields ${\Omega}_{{\overline{A}}}^{-1} (P_{x}) {\cong}{\Omega}_{{\overline{A}}}^{-1} ({\textup{rad}\,}P_{x}) \,/\,S_{x}$. Now, it is easy to see that there is an almost split sequence $$0 \to I_{x} \to \left( \oplus\,\left(\tau^{-1}_{{\overline{A}}} I_{y_{i}}\right)\right)\oplus I_{x}/S_{x} \to \tau^{-1}_{{\overline{A}}} I_{x} \to 0.$$ Since each morphism in this sequence is irreducible, it is either a monomorphism or an epimorphism. Since $S_x$ is the kernel of the morphism $I_{x}\to I_{x}/S_{x}$, another application of the snake lemma and the induction hypothesis $\tau^{-1}_{{\overline{A}}} (I_{y_{i}})\ {\cong}\ {\Omega}_{{\overline{A}}}^{-1}(P_{y_{i}})$ yield $$\tau^{-1}_{{\overline{A}}} I_{x}\ {\cong}\ \left.\left( \oplus\,\left(\tau^{-1}_{{\overline{A}}} I_{y_{i}}\right)\right)\right/ S_{x} \ {\cong}\ \left. \left( \oplus\,\left({\Omega}^{-1}_{{\overline{A}}} P_{y_{i}}\right)\right)\right/ S_{x}\ {\cong}\ {\Omega}_{{\overline{A}}}^{-1}(P_{x}).$$ [Fundamental domain for the cluster category.]{} [${\mathcal{L}_{\overline{A}}}$ as a subcategory of $\textup{mod}\, {\hat{A}}$ ]{} As a consequence of the above description, the left part ${\mathcal{L}_{\overline{A}}}$ is nicely embedded in $\textup{mod}\, {\overline{A}}$, and thus in $\textup{mod}\, {\hat{A}}$. The embedding ${\mathcal{L}_{\overline{A}}}{\hookrightarrow}\textup{mod}\, {\overline{A}}{\hookrightarrow}\textup{mod}\, {\hat{A}}$ is full, exact and preserves indecomposable modules, irreducible morphisms and almost split sequences. [Relation between ${\mathcal{L}_{\overline{A}}}$ and ${\mathcal{C}_A}$]{} We are now able to describe an exact fundamental domain for the cluster category ${\mathcal{C}_A}$ inside $\textup{mod}\,{\overline{A}}$, and actually inside ${\mathcal{L}_{\overline{A}}}$. Indeed, since $A$ is hereditary, and thus of finite global dimension, we have a triangulated equivalence ${\mathcal{D}^b(\textup{mod}\, A)}{\cong}\underline{\textup{mod}\,}{\hat{A}}$ (see [@H]). Let $$\hat \pi : \textup{mod}\,\, {\hat{A}}{\twoheadrightarrow}\underline{\textup{mod}\,} {\hat{A}}{\cong}{\mathcal{D}^b(\textup{mod}\, A)}{\twoheadrightarrow}{\mathcal{C}_A}$$ be the canonical functor. We define an [*exact fundamental domain*]{} for $\hat \pi$ to be a full convex subcategory of ${\textup{ind}\,}{\hat{A}}$ which contains exactly one point of each fibre $\hat \pi ^{-1}(X)$, with $X$ an indecomposable object in $ {\mathcal{C}_A}$. We recall at this point that ${\textup{ind}\,}{\overline{A}}$ is a full convex subcategory of ${\textup{ind}\,}{\hat{A}}$. \[cor5\] The functor $\hat\pi$ induces a one-to-one correspondence between the non projective-injective modules in ${\mathcal{L}_{\overline{A}}}$ and the indecomposable objects in ${\mathcal{C}_A}$. In particular, ${\mathcal{L}_{\overline{A}}}$ is an exact fundamental domain for $\hat\pi$. Since ${\mathcal{L}_{\overline{A}}}$ is a full convex subcategory of ${\textup{ind}\,}{\overline{A}}$, it is also convex inside ${\textup{ind}\,}{\hat{A}}$. Furthermore, the non projective-injective modules in ${\mathcal{L}_{\overline{A}}}$ are just the modules in ${\textup{ind}\,}A$ and those of $\{{\Omega}_{{\overline{A}}}^{-1} P_x\mid x \in ({\mathcal{Q}_A})_0\}$. The statement follows at once from the definition of ${\mathcal{C}_A}$ and from the fact that under the triangle equivalence ${\mathcal{D}^b(\textup{mod}\, A)}{\cong}\underline{\textup{mod}}\, {\hat{A}}$, the shift of ${\mathcal{D}^b(\textup{mod}\, A)}$ corresponds to ${\Omega}_{{\hat{A}}}^{-1}$ (see [@H]). [Tilting modules vs tilting objects]{} [The main theorem]{} In this section, we prove our main theorem, which compares the tilting ${\overline{A}}$-modules with the tilting objects in ${\mathcal{C}_A}$. For this purpose, we assume without loss of generality that our tilting modules and our tilting objects are multiplicity-free. We start by observing that, if $T$ is a tilting ${\overline{A}}$-module, then every indecomposable projective-injective ${\overline{A}}$-module is a direct summand of $T$. Hence $T$ decomposes uniquely as $T=T_0\oplus e'{\overline{A}}$, where $T_0$ has no projective-injective direct summands. We say that $T$ is an $\mathcal{L}$-[*tilting module*]{} if $T_0\in{\textup{add}\,}{\mathcal{L}_{\overline{A}}}$. We denote by $\bar\pi: \textup{mod}\,{\overline{A}}\to {\mathcal{C}_A}$, the composition of the inclusion $\textup{mod}\, {\overline{A}}{\hookrightarrow}\textup{mod}\,{\hat{A}}$ and the functor $\hat\pi$. By abuse of notation, the modules will be often denoted by the same letter even when considered as objects in different categories. \[mainthm\] There is a one-to-one correspondence $$\begin{aligned} \left\{ \mathcal{L}-{tilting \ modules}\right\} &\longleftrightarrow& \left\{ {Tilting\ objects\ in \ }{\mathcal{C}_A}\right\}\nonumber\\ {given\ by } \hspace{2cm} T=T_0\oplus e'\, {\overline{A}}&\longleftrightarrow& \bar\pi(T_0).\nonumber\end{aligned}$$ Let $T=T_0\oplus e'{\overline{A}}$ be an $\mathcal{L}$-tilting module and let $X=\bar\pi(T_0)$. Say $T_0=\oplus_{i=1}^n T_i$ where the $T_i$ are pairwise non-isomorphic indecomposable ${\overline{A}}$-modules. Then $X=\oplus_{i=1}^n X_i$ with $X_i=\bar\pi(T_i)$. We first notice that, clearly, the number $n$ of indecomposable summands of $T_0$ is equal to the rank of the Grothendieck group of $A$. Hence, in order to show that $X$ is a tilting object in ${\mathcal{C}_A}$, it suffices to prove that ${\textup{Ext}^1}_{{\mathcal{C}_A}}(X,X)=0$. Suppose to the contrary that there exist $i,j$ such that ${\textup{Ext}^1}_{{\mathcal{C}_A}}(X_j,X_i)\ne 0$. Since ${\textup{Ext}^1}$ is symmetric in the cluster category by [@BMRRT 1.7], we also have ${\textup{Ext}^1}_{{\mathcal{C}_A}}(X_i,X_j)\ne 0$. Thus there are non-zero morphisms $X_i\to \tau_{{\mathcal{C}_A}} X_j$ and $X_j\to \tau_{{\mathcal{C}_A}} X_i$ in ${\mathcal{C}_A}$. Let $\hat F={\Omega}_{{\hat{A}}}^{-1}\tau_{{\hat{A}}}^{-1}$. Then there exist integers $s,t\ge 0$ such that the previous morphisms lift to non-zero morphisms in ${\underline{\textup{mod}}\,}{\hat{A}}$ $$T_i\to \hat F^s\tau_{{\hat{A}}} T_j \quad \textup{and} \quad T_j\to \hat F^t\tau_{{\hat{A}}} T_i,$$ by definition of the cluster category and the triangulated structure of $\textup{mod}\,{\hat{A}}$, see [@H]. Moreover $s\ne 0$ and $t\ne 0$ since by hypothesis ${\textup{Ext}^1}_{{\overline{A}}}(T_j,T_i)={\textup{Ext}^1}_{{\overline{A}}}(T_i,T_j)=0$. Now $T_i,T_j$ are in ${\mathcal{L}_{\overline{A}}}={\textup{ind}\,}A \cup {\Sigma}$. We then have $3$ cases to consider. 1. $T_i,T_j\in {\Sigma}$. Then $X_i$ and $X_j$ lie on a slice of ${\mathcal{C}_A}$, hence ${\textup{Ext}^1}_{{\mathcal{C}_A}}(X_i,X_j)=0$, a contradiction. 2. $T_i,T_j\in {\textup{ind}\,}A$. If $s=1$, then there is a non-zero morphism $T_i\to \hat F\tau_{{\hat{A}}} T_j ={\Omega}_{{\hat{A}}}^{-1} T_j$ in ${\underline{\textup{mod}}\,}{\hat{A}}$. But this is impossible since $${\underline{\textup{Hom}}\,}_{{\hat{A}}}(T_i,{\Omega}_{{\hat{A}}}^{-1} T_j) = {\textup{Hom}\,}_{{\mathcal{D}^b(\textup{mod}\, A)}}(T_i,T_j[1]) = {\textup{Ext}^1}_A(T_i,T_j)=0$$ where we have identified the modules $T_i$ and $T_j$ with the corresponding stalk complexes in ${\mathcal{D}^b(\textup{mod}\, A)}$. Assume thus that $s\ge 2$. Now, either $\tau_{{\hat{A}}} T_j$ is an $A$-module, or $T_j$ is a projective $A$-module, and then $\tau_{{\hat{A}}} T_j[1]$ is an $A$-module. But this fact and the structure of the morphisms in the derived category (see [@H]) imply that $$\underline{{\textup{Hom}\,}}_{{\hat{A}}}(T_i,\hat F^s\tau_{{\hat{A}}} T_j) = {\textup{Hom}\,}_{{\mathcal{D}^b(\textup{mod}\, A)}}(T_i, F^s\tau_{{\hat{A}}} T_j) =0,$$ again a contradiction. 3. $T_i\in{\textup{ind}\,}A, T_j\in{\Sigma}$. Then by Proposition \[prop1\], there exists an indecomposable projective $A$-module $P_x$ such that $T_j={\Omega}^{-1}_{{\hat{A}}}\, P_x$. Since ${\hat{A}}$ is self-injective, it follows from [@ARS IV.3.7] that ${\Omega}^{-2}_{{\hat{A}}}=\nu_{{\hat{A}}}\tau^{-1}_{{\hat{A}}}$ where $\nu_{{\hat{A}}}$ is the Nakayama functor in $\textup{mod}\,{\hat{A}}$. Thus $\hat F \,\tau_{{\hat{A}}}\, T_j= {\Omega}^{-2}_{{\hat{A}}}\, P_x = \nu_{{\hat{A}}} \, \tau^{-1}_{{\hat{A}}}\, P_x$ which is an $A'$-module (unless $A$ is of Dynkin type $A_n$, linearly oriented and $P_x$ is projective-injective, in which case $\hat F \,\tau_{{\hat{A}}}\, T_j[-1] $ is an $A'$-module). Therefore the modules $T_i$ and $\hat F^s \,\tau_{{\hat{A}}}\, T_j$ have disjoint supports for any $s\ge 1$. Therefore $ \underline{{\textup{Hom}\,}}_{{\hat{A}}}(T_i,\hat F^s\tau_{{\hat{A}}} T_j) =0$ for any $s\ge 1$, contradiction. This completes the proof that $X=\hat\pi(T_0) $ is a tilting object in ${\mathcal{C}_A}$. Conversely, let $X=\oplus_{i=1}^n X_i$ be any tilting object in ${\mathcal{C}_A}$, where we assume that the objects $X_i$ are indecomposable and pairwise non-isomorphic. By Theorem \[cor5\], there exists, for each $i$ with $1\le i\le n$, a unique module $T_i\in{\mathcal{L}_{\overline{A}}}$ in the fibre $\hat\pi^{-1}(X_i)$. Let $T_0=\oplus_{i=1}^n T_i$. Then, clearly $\hat\pi(T_0)=X$. We want to show that $T=T_0\oplus e'{\overline{A}}$ is an $\mathcal{L}$-tilting ${\overline{A}}$-module. Since $T_0\in{\textup{add}\,}{\mathcal{L}_{\overline{A}}}$ by construction and, on the other hand, the number of indecomposable summands of $T_0$ is equal to the rank of the Grothendieck group of $A$, we only have to prove that ${\textup{Ext}^1}_{{\overline{A}}}(T,T)=0$. Suppose to the contrary, that there exist $i,j$ such that ${\textup{Ext}^1}_{{\overline{A}}}(T_i,T_j)\ne 0$. Then ${\textup{Hom}\,}_{{\overline{A}}}(T_j,\tau_{{\overline{A}}} T_i)\ne 0$. In particular, $T_i $ is not projective in mod${\overline{A}}$. Now, $T_i\in{\mathcal{L}_{\overline{A}}}$ implies that $\tau_{{\overline{A}}} T_i=\tau_{{\hat{A}}} T_i$. By Lemma \[lemma2\] and Corollary \[cor1\], we also have $\tau_{{\overline{A}}} T_i\in\textup{ind}\, A$. Therefore ${\textup{Hom}\,}_{{\overline{A}}}(T_j,\tau_{{\overline{A}}} T_i) \ne 0$ implies that $T_j\in\textup{ind}\, A$ (because ${\textup{ind}\,}A$ is closed under predecessors in ${\textup{ind}\,}{\overline{A}}$). Thus ${\textup{Hom}\,}_A(T_j,\tau_{{\overline{A}}} T_i)\ne 0$ and then ${\textup{Ext}^1}_{{\mathcal{C}_A}}(X_i,X_j)\ne 0$, contradiction. [Example]{} Let $A$ be given by the quiver $$\xymatrix@C=20pt {&\ar[ld]^{\alpha}2\\ 1&\ar[l]^{\beta}3\\ &\ar[lu]^{\gamma}4}.$$ Then the ordinary quiver of ${\overline{A}}$ is given by $$\xymatrix@C=20pt {&\ar[ld]^{\alpha}2&&\ar[ld]^{{\alpha}'} 2'\\ 1&\ar[l]^{\beta}3 &\ar[lu]^{\lambda}\ar[l]^{\mu}\ar[ld]^{\nu} 1&\ar[l]^{{\beta}'}3'\\ &\ar[lu]^{\gamma}4 &&\ar[lu]^{{\gamma}'}4'}$$ bound by the relations ${\lambda}{\alpha}=\mu{\beta}=\nu{\gamma}, \ {\alpha}'\mu={\alpha}'\nu={\beta}'\nu={\beta}'{\lambda}={\gamma}'{\lambda}={\gamma}'\mu=0$. The Auslander-Reiten quiver of ${\overline{A}}$ is given by $$\xymatrix@C=10pt {\save[0,0].[5,7] *[F-:]\frm{} \\ &&&&&\circ\ar[rdd]&&&&&\circ\ar[rd]&&&&&&&\\ &\diamond\ar[rd]&& \bullet\ar[rd]&& \bullet\ar[rd]&& \bullet\ar[rd]\ && \bullet\ar[rd]\ar[ru]&\circ\ar[rd]& \bullet\ar[rd]&& \bullet\ar[rd]& &\bullet \\ \ \bullet\ar[ru]\ar[r]\ar[rd]& \diamond\ar[r]& \bullet\ar[ru]\ar[r]\ar[rd]& \bullet\ar[r]& \bullet\ar[ru]\ar[r]\ar[rd]\ar[ruu]& \bullet\ar[r]& \bullet\ar[ru]\ar[r]\ar[rd]& \bullet\ \ar[r]& \bullet\ar[ru]\ar[r]\ar[rd]& \bullet\ar[r]\ar[ru]& \bullet\ar[ru]\ar[r]\ar[rd]& \bullet\ar[r]& \bullet\ar[ru]\ar[r]\ar[rd]& \bullet\ar[r]& \bullet\ar[ru]\ar[r]\ar[rd]& \bullet\\ &\bullet\ar[ru]&& \diamond\ar[ru]&& \bullet\ar[ru]&& \diamond\ \ar[ru]&& \bullet\ar[ru]\ar[rd]&& \bullet\ar[ru]&& \bullet\ar[ru]&& \bullet \\ &&&&&&&&&&\circ\ar[ru]&\\ &&&{\mathcal{L}_{\overline{A}}}\\ \restore }$$ where we have indicated the left part ${\mathcal{L}_{\overline{A}}}$. We have also indicated an $\mathcal{L}$-tilting module $T=T_0\oplus e'{\overline{A}}$, where $T_0 \in {\textup{add}\,}{\mathcal{L}_{\overline{A}}}$. The summands of $T_0$ are indicated by diamonds and the (projective-injective) summands of $e'{\overline{A}}$ by circles. [More on duplicated algebras of hereditary algebras]{} It follows from our main theorem that the duplicated algebras of hereditary algebras are quite a natural class to consider, since all the tilting objects of the cluster category correspond to the actual modules over the duplicated algebras. In this section we study other properties of these algebras, which are consequences of the description of the left part ${\mathcal{L}_{\overline{A}}}$ and the $\textup{Ext}$-injectives as done in the previous sections. We recall that a finite dimensional algebra $C$ is called [*left (or right) supported*]{} provided the class ${\textup{add}\,}{\mathcal{L}_{{C}}}$ (or ${\textup{add}\,}{\mathcal{R}_{{C}}}$) is contravariantly finite (or covariantly finite, respectively) in $\textup{mod}\,C$, see [@ACT; @ACPT]. The duplicated algebra ${\overline{A}}$ of a hereditary algebra is both left and right supported. By [@ACT 3.3], the canonical module $T=U\oplus V$ (with $U=\oplus_{X\in{\Sigma}} X$ and $V=\oplus_{{\overline{P}}_x\notin {\mathcal{L}_{\overline{A}}}} \,{\overline{P}}_x$) is a partial tilting module. Now the number of its indecomposable summands equals the number of isomorphism classes of indecomposable injective $A$-modules plus the number of isomorphism classes of indecomposable projective-injective ${\overline{A}}$-modules. Hence $T$ is a tilting module and ${\overline{A}}$ is left supported, by [@ACT thm. A]. The other statement follows by symmetry. The assumption that $A$ is a hereditary algebra is essential. If $A$ is a tilted algebra which is the endomorphism algebra of a regular tilting module, then it is easily seen that ${\overline{A}}$ is neither left nor right supported. Equivalent statements to duplicated algebras being representation-finite are given in the next corollary. We recall that an algebra $C$ is said to be a [*laura algebra*]{} [@AC] provided the class ${\textup{ind}\,}C \setminus ({\mathcal{L}_{{C}}}\cup {\mathcal{R}_{{C}}})$ contains only finitely many indecomposables. Let $A$ be a hereditary algebra. The following conditions are equivalent: $$\begin{array}{ll} (a) & {\overline{A}}\ is\ a\ laura\ algebra.\\ (b) & A \ is \ of \ Dynkin\ type. \\ (c) &{\overline{A}}\ is \ \textup{{\it representation-finite}}.\\ \end{array}$$ If this is the case, then ${\overline{A}}$ is simply connected. We denote by ${\Sigma}'$ the set of all indecomposable Ext-projectives in ${\textup{add}\,}{\mathcal{R}_{\overline{A}}}$. By Lemma \[lemma2\], Corollary \[cor1\] and their duals, the duplicated algebra ${\overline{A}}$ is laura if and only if the class $[\tau_{{\overline{A}}}\,{\Sigma}, \tau^{-1}_{{\overline{A}}}\, {\Sigma}'] $ of all the $M\in {\textup{ind}\,}{\overline{A}}$ such that there exists a path $L\leadsto M\leadsto N,$ with $\tau^{-1}_{{\overline{A}}} L\in {\Sigma}$ and $\tau_{{\overline{A}}} N \in {\Sigma}'$ consists of finitely many indecomposables. Now, by [@A 2.6] this class is an exact fundamental domain for the module category over the trivial extension $T(A)$ of $A$ by its minimal injective cogenerator $DA$. Therefore ${\overline{A}}$ is laura if and only if $T(A)$ is representation finite, or, by [@T], if and only if $A$ is of Dynkin type which, by [@A 2.6] is the case if and only if ${\overline{A}}$ is representation-finite. The last statement follows from [@A 2.7]. Assume $A$ to be representation-infinite. Then, of course, Theorems \[cor1\] and \[mainthm\] still apply. In this case as well, a good description of the module category of the duplicated algebra ${\overline{A}}$ is known (see [@A; @ANS]) and, at least in the tame case, it is possible to compute explicitly the $\mathcal{L}$-tilting modules. The authors wish to thank Raymundo Bautista for very useful discussions. I. Assem, On representations of the bimodule $DA$, [ *Tsukuba J. Math.* ]{}[**9**]{}, No.2 (1985) 217–232. I. Assem, J. Cappa, M.I. Platzeck and S. Trepode, Some characterisations of supported algebras , [*Preprint No.21, Dép. de Math., U. de Sherbrooke*]{} (2005) I. Assem, F.U. Coelho, Two-sided gluings of tilted algebras , [*J. Algebra* ]{}[**269**]{}, (2) (2003) 456–470. I. Assem, F.U. Coelho and S. Trepode, The left and right parts of a module category, [*J. Algebra* ]{}[**281**]{},(2) (2004), 518–534. I. Assem, J. Nehring and W. Schewe, Fundamental domains and duplicated algebras, [*CMS Conference Proceedings* ]{}[ **11**]{}, (1991) 25–51. M. Auslander and S.O. Smalø, Almost split sequences in subcategories [*J. Algebra* ]{} [**69**]{}, (1981) 426–454. M. Auslander, I. Reiten and S.O. Smalø, [ *Representation Theory of Artin Algebras* ]{} Cambridge Studies in Advanced Math. 36, (Cambridge University Press, 1995). A. Buan, R. Marsh, M. Reineke, I. Reiten and G. Todorov, Tilting theory and cluster combinatorics, to appear in [*Adv. Math.*]{} P. Caldero, F. Chapoton and R. Schiffler, Quivers with relations arising from clusters ($A_n$ case), to appear in [*Trans. Amer. Math. Soc.*]{} D. Hughes and J. Waschbüsch, Trivial extensions of tilted algebras [*Proc. London Math. Soc.* ]{}[**46**]{} (1983) 347–364. D. Happel, [*Triangulated Categories in the Representation Theory of Finite Dimensional Algebras* ]{} London Mathematical Society. Lecture Notes Series 119, (Cambridge University Press, 1988). D. Happel, I. Reiten and S.O. Smalø, Tilting in abelian categories and quasi tilted algebras, [*Memoirs Amer. Math. Soc.* ]{}[**575**]{}, (1996). S. Fomin and A. Zelevinsky, Cluster algebras I. Foundations, [*J. Amer. Math. Soc.* ]{}[ **15**]{}(2), (2002), 497–529 (electronic) S. Fomin and A. Zelevinsky, Cluster algebras II. Finite type classification, [*Inventiones Mathematicae* ]{}[**154**]{}(1), (2003), 63–121. B. Keller, Triangulated orbit categories, preprint (2005). C.M. Ringel, [*Tame algebras and integral quadratic forms*]{} Lecture Notes in Math. 1099, (Springer Verlag, 1984). H. Tachikawa, Representations of trivial extensions of hereditary algebras [*Proc. ICRA II* ]{}, Lecture Notes in Math. 832, (Springer Verlag 1980) 579–599. K. Yamagata, Extensions over hereditary artinian rings with self-dualities I. [*J. Algebra* ]{}[**73**]{} (2) (1981) 386–433.

ArXiv

--- abstract: 'We define the concept of *companion automorphism* of a Hopf algebra $H$ as an automorphism $\sigma:H \rightarrow H$: $\sigma^2={\mathcal{S}}^2$ –where ${\mathcal{S}}$ denotes the antipode–. A Hopf algebra is said to be *almost involutive* (AI) if it admits a companion automorphism that can be viewed as a special additional symmetry. We present examples and study some of the basic properties and constructions of AI-Hopf algebras centering the attention in the finite dimensional case. In particular we show that within the family of Hopf algebras of dimension smaller or equal than 15, only in dimension eight and twelve, there are non almost involutive Hopf algebras.' address: | Facultad de Ciencias\ Universidad de la República\ Iguá 4225\ 11400 Montevideo\ Uruguay\ author: - Andrés Abella - Walter Ferrer Santos title: Almost involutive Hopf algebras --- [^1] Introduction. ============= We start by summarizing the contents of this paper. In Section \[section:defejemplos\] we introduce the definition of an *almost involutive Hopf algebra*, show that Sweedler’s Hopf algebra is an example and observe that compact quantum groups are also examples –in the infinite dimensional case–. Then we show that some of the standardt constructions, such as that of a matched pair or in particular of the Drinfel’d double, when applied to almost involutive Hopf algebras yield as result an almost involutive Hopf algebra. Moreover, as in the case that the order of the antipode squared is odd the AI property is authomatic, we concentrate in the interesting case that is when the order of the antipode squared is even. In Section \[section:prelim\] we recall some aspects of the theory of integrals in finite dimensional Hopf algebras, the modular function, the modular element, etc. These tools will be used in the rest of the paper. In Section \[section:examples\] we present some examples and non–examples of almost involutive Hopf algebras, and show that except at dimension 8 and 12 and for only a few types, all Hopf algebras of dimension smaller or equal than 15 are almost involutive. In the Appendix, we present –a probably well known– result on square roots of linear automorphisms of Hopf algebras that yield necessary and sufficient conditions for a finite order automorphism to have a square root that is also a Hopf automorphism. This general result can be used in some examples, for example we use it to show that the Taft algebras are almost involutive. We finish this introduction with some commentaries about notations. We work over an algebraically closed field ${{\mathbb K}}$ of characteristic zero and adopt the usual Sweedler’s notation and the other conventions in the theory of Hopf algebras as they appear for example in [@kn:mont], e.g. ${\mathcal{S}}$ denotes the antipode, and $\lambda$ is an integral, . Also a non zero element $g$ in $H$ is called a *group-like* element if $\Delta(g)=g\otimes g$ and we designate as $G(H)$ and $G(H^\vee)$ the set of group-like elements in the original Hopf algebra and its dual. Moreover, if $x\in H$ is such that $\Delta(x)=x\otimes g+ h\otimes x$, where $g,h\in G(H)$, then $x$ is called a [$(g,h)$-primitive]{} element and we write as $P_{g,h}$ the space of $(g,h)$-primitive elements oh $H$. A Hopf algebra is *pointed* if its only simple subcoalgebras are the dimensional. In general we concentrate our considerations in the case that finite dimensional Hopf algebras. Main definition and general examples. {#section:defejemplos} ===================================== General considerations ---------------------- Recall that a Hopf algebra $H$ is called *involutive* if ${\mathcal{S}}^2={\ensuremath{\mathrm{id}}}$. Examples of involutive Hopf algebras are commutative, cocommutative or semisimple Hopf algebras over ${{\mathbb C}}$. We say that a Hopf algebra $H$ is *almost involutive* or that is an *AI-Hopf algebra* if there exists a Hopf algebra automorphism $\sigma:H\to H$ such that ${\mathcal{S}}^2=\sigma^2$. An automorphism $\sigma$ as above, is called a [*companion automorphism*]{} of ${\mathcal{S}}$ or simply a [*companion automorphism*]{} of $H$. It follows from Radford’s formula –see [@kn:bbt], [@kn:radfordbasic] and [@kn:schcordoba]– that ${\mathcal{S}}^2$ is a diagonalizable Hopf algebra automorphism of finite order, more precisely: $\operatorname{ord}({\mathcal{S}}^2)| 2\operatorname{ord}(a)\operatorname{ord}(\alpha)$ –where $a$ is the modular element and $\alpha$ the modular form, see Section \[section:prelim\]. It can also be proved that if $H_n$ is the coradical filtration of $H$, then $\operatorname{ord}{{\mathcal{S}}^2}|_{H_1}=\operatorname{ord}{{\mathcal{S}}^2}$ –see [@kn:radfordsch Lemma 4]–. The observation that follows shows that if the antipode squared is of finite order $m$ –in particular if $H$ is finite dimensional–, then the only case in which it is interesting to consider the AI condition is when $m$ is even. \[obse:oddtrivial\] Let $H$ be a Hopf algebra with order of ${\mathcal{S}}^2$ odd. Then $H$ is almost involutive. Indeed, if $({\mathcal{S}}^2)^{2k-1}=\operatorname{id}$, then ${\mathcal{S}}^2=({\mathcal{S}}^{2k})^2$. \[exam:first\] *Sweedler’s Hopf algebra.*\[example:sweedler\] As an illustration of an almost involutive Hopf algebra with the antipode squared of even order, we take Sweedler’s Hopf algebra $H_4$. As an algebra $H_4$ is given by generators and relations as $H_4=\langle g,x:\ g^2=1,\ x^2=0,\ xg+gx=0\rangle$. The comultiplication is given by $\Delta(g)=g\otimes g$, $\Delta(x)=x\otimes g+1\otimes x$, and the antipode is defined by ${\mathcal{S}}(g)=g$, ${\mathcal{S}}(x)=-xg$. Clearly, ${\mathcal{S}}^2(g)=g$ and ${\mathcal{S}}^2(x)=-x$ and $|{\mathcal{S}}^2|=2$. A direct verification shows that the map $\sigma$ defined by $\sigma:g \mapsto g$, $\sigma: x \mapsto ix$ and extended multiplicatively, is a companion automorphism where $i$ is a square root of $-1$. The map: $g \mapsto g,\, x \mapsto -ix$ is also a companion automorphism. The following result shows that almost involutive non involutive Hopf algebras abound. Let $H$ be a finite dimensional pointed Hopf algebra. If the order $|G(H)|$ is odd, then $H$ is almost involutive. In particular this happens if the dimension of $H$ is odd. We know $\operatorname{ord}{{\mathcal{S}}^2}=\operatorname{ord}{{\mathcal{S}}^2}|_{H_1}$, being $H_n$ the coradical filtration of $H$. As $H$ is pointed, then $H_1={{\mathbb K}}G(H)\oplus\oplus_{g,h\in G(H)}P'_{g,h}(H)$, where $P'_{g,h}(H)$ is an arbitrary linear complement of ${{\mathbb K}}(g-h) \subseteq P_{g,h}(H)$ –see [@kn:stefan Thm. 0.1]–. It is easy to see that ${\mathcal{S}}^2|_{G(H)}={\ensuremath{\mathrm{id}}}$ and if $x\in P_{g,h}$, then ${\mathcal{S}}^2(x)=gh^{-1}xg^{-1}h$. This last formula implies ${\mathcal{S}}^{2n}(x)=\left(gh^{-1}\right)^nx\left(g^{-1}h\right)^n$, for all $n$. Hence $\operatorname{ord}{{\mathcal{S}}^2}$ divides $|G(H)|$ and we can apply the observation \[obse:oddtrivial\]. The last assertion follows because $|G(H)|$ divides the dimension of $H$. If the order of $|G(H)|$ is even, then the above theorem is no longer true – see examples \[ej:8\] and \[ej:12\] below–. Concerning the infinite dimensional situation, we mention that in [@kn:andres-walter-mariana], the authors –together with M. Haim– proved the following general theorem. *\[[@kn:andres-walter-mariana], Theorem 8.7.\]* Let $H$ be a compact quantum group with antipode ${\mathcal{S}}$. Then, $H$ admits a unique companion automorphism –that in this case is denoted as ${\mathcal{S}}_+$ instead of $\sigma$–, that is a positive operator with respect to the natural inner product of $H$, and there is a linear functional $\beta: H \rightarrow \mathbb C$ such that ${\mathcal{S}}_+(x)=\sum \beta(x_1)x_2\beta^{-1}(x_3)$. In the mentioned paper the following explicit example is constructed. See Woronowicz’s [@kn:worosu2]. Let $\mu \in \mathbb C$ be a non zero real number such that $|\mu|<1$, and call $\operatorname{SU}_\mu(2,\mathbb C)$ the algebra generated by $\{\alpha,\alpha^*, \gamma, \gamma^*\}$, subject to the following relations: $\alpha^*\alpha+\mu^2 \gamma^*\gamma=1$, $\alpha \alpha^*+\mu^4 \gamma \gamma^*=1$, $\gamma^*\gamma=\gamma\gamma^*$, $\mu\gamma\alpha=\alpha\gamma$, $\mu\gamma^*\alpha=\alpha\gamma^*$, $\mu^{-1}\gamma\alpha^*=\alpha^*\gamma$, $\mu^{-1}\gamma^*\alpha^*=\alpha^*\gamma^*$. The star structure in $\operatorname{SU}_\mu(2,\mathbb C)$, is defined as being antimultiplicative, involutive, conjugate linear and defined on the generators as shown. This algebra admits a natural compatible comultiplication and becomes a *compact quantum group* –see [@kn:worosu2] for details– with antipode: ${\mathcal{S}}(\alpha)=\alpha^*, {\mathcal{S}}(\alpha^*)= \alpha, {\mathcal{S}}(\gamma) =-\mu \gamma, {\mathcal{S}}(\gamma^*)=-\mu^{-1} \gamma^*$. One shows that the companion automorphism ${\mathcal{S}}_+$ is given as: ${\mathcal{S}}_+(\alpha)=\alpha$, ${\mathcal{S}}_+(\alpha^*)=\alpha^*$, ${\mathcal{S}}_+(\gamma)=|\mu|\gamma$, ${\mathcal{S}}_+(\gamma^*)=|\mu|^{-1}\gamma^*$. 1. Recall that if $\tau:H \rightarrow H$ is a Hopf algebra homomorphism, then it commutes with ${\mathcal{S}}$. In particular in the situation above the antipode ${\mathcal{S}}$ and the companion automorphism $\sigma$ commute. 2. In general one cannot guarantee the uniqueness of $\sigma$. See the comment at the end of Example \[exam:first\]. 3. More examples and non examples will be provided in later sections, here we mention the following: involutive Hopf algebras and the quantum enveloping algebra of $\mathfrak{sl} (2)$. Constructions of almost involutive Hopf algebras. {#subsection:const} ------------------------------------------------- The following constructions always produce AI-Hopf algebras. 1. *Duals and tensor products.* If $H$ is a finite dimensional AI-Hopf algebra, so is its dual. If $H$ and $K$ are AI-Hopf algebras so is $H \otimes K$. 2. *Matched pairs.* Assume that we have a matched pair of Hopf algebras $(A,H,{\triangleleft},{\triangleright})$. If $A$ and $H$ are almost involutive with companion morphisms $\sigma_A$ and $\sigma_H$, so is the bicrossed product $A \bowtie H$ provided that the given companion automorphisms for $A$ and $H$ satisfy the following compatibility conditions: 1. $\sigma_A(x {\triangleright}a)= \sigma_H(x) {\triangleright}\sigma_A(a)$ 2. $\sigma_H(x {\triangleleft}a)= \sigma_H(x) {\triangleleft}\sigma_A(a)$. This assertion follows easily form the fact that the antipode for the bicrossed product is simply ${\mathcal{S}}(ax)={\mathcal{S}}_H(x){\mathcal{S}}_A(a)$ for $a \in A, x \in H$ –see for example [@kn:andres-walter-mariana-2] or [@kn:Majid] for a detailed description of the structure of $A \bowtie H$–. Hence if we write ${\mathcal{S}}_A^2=\sigma_A^2$ and ${\mathcal{S}}_H^2=\sigma_H^2$, the Hopf algebra morphism $\sigma(ax)=\sigma_A(a) \sigma_H(x)$ does the job. 3. *Drinfel’d double.* In particular, if $H$ is almost involutive, so is its Drinfel’d double $D(H)$. This follows from the fact that the Drinfel’d double can be viewed as a bicrossed product (see [@kn:andres-walter-mariana-2] or [@kn:Majid]) $D(H)=H^{\vee{\ensuremath{\mathrm{cop}}}} \bowtie H$ where the structure of matched pair is given by $H \stackrel{\triangleleft}{\leftarrow} H\otimes H^\vee\stackrel{\triangleright}{\to}H^\vee$ defined as: $$(x{\triangleright}\alpha)(y)=\sum \alpha\left({\mathcal{S}}^{-1}(x_2) y x_1\right),\quad x{\triangleleft}\alpha = \sum \alpha\left({\mathcal{S}}^{-1}(x_3)x_1 \right) x_2 ,\quad \forall \alpha\in H^\vee,\ x,y\in H.$$ In this situation one easily verifies that if $\sigma$ is a companion automorphism for $H$, its natural extension to $D(H)$ is a companion automorphism for the double. 4. \[item:trivial\] *Trivial extensions.* 1. Consider a Hopf algebra $H$ that can be written $H=K \oplus M$ where $K$ and $M$ satisfy the following conditions: 1. $K$ is a sub Hopf algebra of $H$. 2. $M$ is a $K$–bimodule and a $K$–bicomodule, in other words the following holds: 1. $KM + MK \subseteq M$; 2. $\Delta(M) \subseteq K \otimes M + M\otimes K$; 3. $M$ is invariant by ${\mathcal{S}}$. 4. The extension of $K$ by $M$ is trivial, in other words, $M^2=0$. If $m \in M$ we have that $\Delta(m)=\sum h_i \otimes m_i + n_i \otimes k_i$ with $h_i,k_i \in K$ and $m_i,n_i \in M$. Then $m=\sum \varepsilon(m_i)h_i + \varepsilon(k_i)n_i$. If the $h_i$ are linearly independent, we deduce that $\varepsilon(m_i)=0$ and similarly for $n_i$. Hence, if we write write the expression $\Delta(m)=\sum h_i \otimes m_i + n_i \otimes k_i$ with $h_i$ and $k_i$ linearly independent, and apply $\varepsilon \otimes \varepsilon$ to the expression obtain that $\varepsilon(m)=0$. Hence $\varepsilon(M)=0$. Assume that $K$ is almost involutive, and call $\sigma_K$ the corresponding companion automorphism. Suppose also that we can find a linear map $\sigma_M$, with the property that $\sigma_M^2={\mathcal{S}}^2|_M$ and such that: 1. $\sigma_M(hm)=\sigma_K(h)\sigma_M(m)\,,\,\sigma_M(mh)=\sigma_M(m)\sigma_K(h)$. 2. If $\Delta(m)=\sum h_i \otimes m_i + n_i \otimes k_i \in K \otimes M + M \otimes K$, then $\Delta(\sigma_M(m))=\sum \sigma_K(h_i) \otimes \sigma_M(m_i) + \sigma_M(n_i) \otimes \sigma_K(k_i)$. In that situation the map $\sigma(h + m)=\sigma_K(h)+\sigma_M(m)$ is a companion automorphism in $H$. First, it is clear that $\sigma^2={\mathcal{S}}^2$. Morover $\sigma((h+m)(k+n))=\sigma(hk+hn+mk)=\sigma_K(hk)+\sigma_M(hn+mk)= \sigma_K(h)\sigma_K(k)+ \sigma_K(h)\sigma_M(n)+\sigma_M(m)\sigma_K(k)= (\sigma_K(h)+\sigma_M(m))(\sigma_K(k)+\sigma_M(n))=\sigma(h+m)\sigma(k+n)$. Also, a direct verification shows that the map $\sigma$ is an automorphism of coalgebras. 2. Consider the particular case of a trivial extension such that $K=\left\{x\in H:\ {\mathcal{S}}^2(x)=x\right\}$. Using the techniques developped in the Appendix –see Example \[example:general\]– one can prove that in this situation, $H$ is an almost involutive Hopf algebra. Sweedler’s Hopf algebra is an example. Clearly, $H_4$ can be written in the above manner with: $K={{\mathbb K}}+{{\mathbb K}}g$ and $M={{\mathbb K}}x + {{\mathbb K}}gx$. In this situation $\sigma_K=\operatorname{id}$ and $\sigma_M=i\operatorname{id}$ satisfies all the required properties. Other examples of the above situation will appear later –see Section \[section:examples\]–. Basic results on finite dimensional Hopf algebras. {#section:prelim} ================================================== Here we recall basic results and constructions concerning Hopf algebras that will be used later in the study of some of the examples –see [@kn:Dascalescu] and [@kn:radfordbasic] for proofs and details. We concentrate on the basic properties of integrals, modular function and modular element. There exists $\lambda \in H^\vee$ and $\ell \in H$ that satisfy $\lambda(\ell)=1$, and invertible elements $a\in G(H)$ and $\alpha\in \text{Alg}(H,{{\mathbb K}})$ such that for all $x \in H$: $$\label{eq:intl} \sum x_1\lambda(x_2)=\lambda(x)1\,\quad x\ell=\varepsilon(x)\ell,\quad \sum \lambda(x_1)x_2=\lambda(x)a,\quad \ell x=\alpha(x)\ell.$$ The elements $\lambda$ and $\ell$ that are *left integrals* and the elements $a$ and $\alpha$ are called *modular element* and *modular function* respectively. 1. The modular function and the modular element are of finite order in $H^\vee$ and $H$ respectively and then $\alpha(a)$ is a root of one of order less or equal than the dimension of $H$. In fact $\operatorname{ord}(\alpha(a))|\gcd\{\operatorname{ord}(\alpha),\operatorname{ord}(a)\}$. 2. The elements $\lambda {\,{\scriptstyle \circ }\,}{\mathcal{S}}= \rho$ and ${\mathcal{S}}^{-1}(\ell)=r$ are *right integrals* such that $\rho(r)=1$. We have that for all $x \in H$: $$\label{eq:rint} \sum \rho(x_1)x_2=\rho(x)1,\quad rx=\varepsilon(x)r,\quad \sum x_1\rho(x_2)=\rho(x)a^{-1},\quad xr=\alpha^{-1}(x)r.$$ 3. The elements $\lambda$ and $\ell$ ($\rho$ and $r$) are uniquely determined up to a non zero scalar. 4. The Hopf algebra $H$ (or $H^\vee$) is unimodular, i.e. a left integral is also a right integral, if and only if $\alpha=\varepsilon$ (or $a = 1$) respectively. The following holds –see [@kn:schcordoba]–: 1. $$\label{eq:antipode} {\mathcal{S}}(x)=\sum \ell_1 \lambda(x\ell_2), \quad \forall x\in H . $$ 2. From the formula , we can easily deduce that: $$\label{eq:lambdaS} {\mathcal{S}}(\ell)= \sum \ell_1 \alpha(\ell_2) \quad\text{and}\quad \lambda\big({\mathcal{S}}(x)\big) = \lambda(xa), \quad \forall x\in H .$$ Indeed, applying $\lambda$ we have that: $\lambda({\mathcal{S}}(x))=\sum \lambda(\ell_1) \lambda(x\ell_2)=\sum \lambda(x\lambda(\ell_1)\ell_2) = \lambda(xa)$. The other formula is the dual. 3. In the situation above, we have that: $$\begin{aligned} \label{eq:bilinear} {\mathcal{S}}^2(\ell)= \alpha(a) \ell,\quad {\mathcal{S}}^{2}(r)= \alpha(a) r, \quad \lambda {\,{\scriptstyle \circ }\,}{\mathcal{S}}^2=\alpha(a) \lambda,\quad \rho {\,{\scriptstyle \circ }\,}{\mathcal{S}}^2=\alpha(a) \rho . $$ Indeed, by iteration of the formula we obtain that $\lambda\/{\,{\scriptstyle \circ }\,}\/{\mathcal{S}}^2 = a {{\rightharpoonup}}\lambda {{\leftharpoonup}}a^{-1}$. Being $\lambda {\,{\scriptstyle \circ }\,}{\mathcal{S}}^2$ another left integral, we conclude that it has to be a scalar multiple of $\lambda$. As a $(a {{\rightharpoonup}}\lambda {{\leftharpoonup}}a^{-1})(\ell)=\lambda(a^{-1}\ell a)=\varepsilon(a^{-1}) \lambda(\ell) \alpha(a)=\alpha(a)$ we conclude that $\lambda {\,{\scriptstyle \circ }\,}{\mathcal{S}}^2=\alpha(a) \lambda$. The other formulæ are proved similarly. The formulæ for right integrals are obtained by composition with ${\mathcal{S}}^{\pm 1}$. Notice that both $\ell$ and $r$ are eigenvectors of ${\mathcal{S}}^2$ with eigenvalue $\alpha(a)$. 4. In particular $\lambda({\mathcal{S}}(\ell))=\alpha(a)$ and $\lambda(r)=1$. Indeed, if we put $x=\ell$ in the equation we obtain that $\lambda({\mathcal{S}}(\ell))= \lambda(\ell a)= \lambda\big(\alpha(a)\ell\big)=\alpha(a)$. Moreover, applying the equality $\lambda \circ {\mathcal{S}}^2=\alpha(a) \lambda$ to the element $r={\mathcal{S}}^{-1}(\ell)$ we deduce that: $\lambda(r)=1$. Next we look at the behaviour of the elements above when transformed with a companion automorphism. Let $\sigma$ be a companion automorphism of $H$. 1. The following holds because $\sigma$ is a Hopf algebra map –the proof is omitted as it is standard–: $$\begin{gathered} \label{eqn:first} \sigma(\ell) =\lambda(\sigma(\ell))\ell ,\quad \sigma(r)=\rho(\sigma(r))r,\quad \lambda \circ \sigma = \lambda(\sigma(\ell))\lambda ,\quad \rho \circ \sigma = \rho(\sigma(r))\rho , \\ \label{eq:alfasigma} \sigma(a)=a,\quad \sigma\left(a^{-1}\right)=a^{-1},\quad \alpha \circ \sigma = \alpha ,\quad \alpha^{-1} \circ \sigma = \alpha^{-1}. $$ 2. From the definition of $\rho$ and $r$ and using that $\sigma$ is a Hopf algebra map we deduce $\rho(\sigma(r))=\lambda(\sigma(\ell))$. Then using and we deduce $\lambda(\sigma(\ell))^2=\alpha(a)$. Hence we have that: $$\begin{aligned} \label{eq:erresigma} \sigma(\ell)=r_\sigma \ell, \quad \sigma(r)=r_\sigma r ,\quad \lambda \circ \sigma= r_\sigma \lambda ,\quad \rho \circ \sigma= r_\sigma \rho \quad \text{with} \quad r_\sigma^2=\alpha(a). $$ 3. We consider the Sweedler’s algebra. The set $\{1,g,x,gx\}$ is a basis of $H_4$ and we will denote as $\{1^*,g^*,x^*,(gx)^*\}$ its dual basis. We get $$\ell= (1+g)x;\ r=x(1+g);\ \lambda= x^*; \ \rho=-(gx)^*;\ a=g\,\,;\,\, \alpha:H_4 \rightarrow {{\mathbb K}},\ \text{is given by } \alpha(g)=-1,\ \alpha(x)=0.$$ Recall that the map $\sigma:H_4 \rightarrow H_4$, defined as $\sigma(g)=g$ and $\sigma(x)=ix$ and extended multiplicatively, is a companion automorphism for $H_4$. As $\sigma(\ell)=i\ell$ , in this case $r_\sigma=i$. 4. Call $E_{\sigma,\nu}$ the eigenspace of $\sigma$ corresponding to the eigenvalue $\nu$. Then: $$\begin{gathered} \ell,r \in E_{\sigma,r_\sigma},\quad 1,a,a^{-1} \in E_{\sigma,1}, \nonumber\\ \bigoplus_{\nu \neq r_\sigma}E_{\sigma,\nu} \subset \operatorname{Ker}(\lambda), \nonumber \\ \label{eqn:fifth} \bigoplus_{\nu \neq 1}E_{\sigma,\nu} \subset \operatorname{Ker}(\alpha)\cap \operatorname{Ker}\left(\alpha^{-1}\right)\cap \operatorname{Ker}(\varepsilon), \\ E_{\sigma,\nu} \ell = \ell E_{\sigma,\nu} =0,\quad \nu\ne 1. \nonumber\end{gathered}$$ First observe that the equations and mean that $\ell,r \in E_{\sigma,\lambda(\sigma(\ell))}$ and $a,a^{-1} \in E_{\sigma,1}$, and is clear that $1 \in E_{\sigma,1}$. Moreover, if $x \in E_{\sigma,\nu}$, applying respectively $\lambda,\alpha,\alpha^{-1},\varepsilon$ to the equality $\sigma(x)=\nu x$, we deduce that $r_\sigma\lambda(x)= \nu \lambda(x)$, $\alpha(x)=\nu \alpha(x)$, $\alpha^{-1}(x)=\nu \alpha^{-1}(x)$ and $\varepsilon(x)=\nu\varepsilon(x)$. From the first of these equalities we deduce that if $\nu\ne r_\sigma$, then $\lambda(x)=0$, and similarly for the others. For the last assertion: if $x \in E_{\sigma,\nu}$, from the equality $x\ell=\varepsilon(x)\ell$ and we deduce that $x\ell=0$. Similarly we deduce that $\ell x=0$. Description of the almost involutive Hopf algebras up to dimension 15. {#section:examples} ====================================================================== In the following examples we will often have to deal with a Hopf algebra $H$ with a given group-like element $g$ and a $(g,1)$-primitive element $x$, then ${\mathcal{S}}(g)=g^{-1}$, ${\mathcal{S}}(x)=-xg^{-1}$, ${\mathcal{S}}^2(g)=g$ and ${\mathcal{S}}^2(x)=gxg^{-1}$. \[Example 1, in [@kn:radfordbasic]\] \[example:rad\] Let be $\omega\in{{\mathbb K}}$ a primitive root of order $n$ of 1 and $$H=\left\langle g,x,y:\ g^n=1,\ x^n=0,\ y^n=0,\ gx-\omega^{-1}xg=0,\ gy-\omega yg=0,\ xy-\omega yx=0 \right\rangle.$$ $H$ is a Hopf algebra when it is equipped with coalgebra structure given by $g$ being a group-like element and $x,y$ $(g,1)$-primitive elements. The set $\{g^rx^py^s:\ 0\leq r,p,s\leq n-1\}$ is a basis of $H$, so $\dim H=n^3$. We have ${\mathcal{S}}^2(g)=g$, ${\mathcal{S}}^2(x)=\omega^{-1} x$ and ${\mathcal{S}}^2(y)=\omega y$, hence the order of ${\mathcal{S}}^2$ is $n$. Using again the same method than before but with more labour, we can prove that $H$ is an almost involutive Hopf algebra. Here we can also obtain four companion automorphisms $\sigma$ by direct inspection, defining $\sigma(g)=g$, $\sigma(x)=\pm \nu^{-1} x$ and $\sigma(y)=\pm \nu y$, being $\nu\in{{\mathbb K}}$ such that $\nu^2=\omega$. \[ej:8\] If $H$ is a 8-dimensional not semisimple Hopf algebra, then Stefan shows in [@kn:stefan] that $H$ is isomorphic to one and only one of the following list $$A_{C_2},\quad A'_{C_4},\quad A''_{C_4},\quad A'''_{C_4,\omega},\quad A_{C_2\times C_2},\quad \left( A''_{C_4}\right)^*,$$ being 1. \[item:ac2\] $A_{C_2}=\left\langle g,x,y:\ g^2=1,\ x^2=0,\ y^2=0,\ gx + xg=0,\ gy+yg=0,\ xy+ yx=0 \right\rangle$, $g$ is a group-like element and $x,y$ are $(g,1)$-primitive elements. 2. \[item:a1c4\] $A'_{C_4}=\left\langle g,x:\ g^4=1,\ x^2=0,\ gx + xg=0 \right\rangle$, $g$ is a group-like element and $x$ is a $(g,1)$-primitive element. 3. \[item:a2c4\] $A''_{C_4}=\left\langle g,x:\ g^4=1,\ x^2=g^2-1,\ gx + xg=0 \right\rangle$, $g$ is a group-like element and $x$ is a $(g,1)$-primitive element. 4. \[item:a3c4\] $A'''_{C_4,\omega}=\left\langle g,x:\ g^4=1,\ x^2=0,\ gx- \omega xg=0 \right\rangle$, $g$ is a group-like element, $x$ is a $(g,1)$-primitive element and $\omega\in{{\mathbb K}}$ is a primitive root of unity of order 4. 5. \[item:a2c2c2\] $A_{C_2\times C_2}=\left\langle g,h,x:\ g^2=1,\ h^2=1,\ x^2=0,\ gx + xg=0,\ hx + xh=0,\ gh-hg=0 \right\rangle$, $g$ and $h$ are group-like elements and $x$ is a $(g,1)$-primitive element. The algebras $A_{C_2}, A'_{C_4}, A'''_{C_4,\omega}$ and $A_{C_2\times C_2}$ are almost involutive. We give the companion automorphism $\sigma$ by its values in the generators. - $A_{C_2}$, $\sigma(g)=g$, $\sigma(x)=ix$ and $\sigma(y)=iy$. Observe that this is the case $n=2$ in the example \[example:rad\]. - $A'_{C_4}$, $\sigma(g)=g$ and $\sigma(x)=ix$. - $A'''_{C_4,\omega}$, $\sigma(g)=g$ and $\sigma(x)=\omega x$. - $A_{C_2\times C_2}$, $\sigma(g)=g$, $\sigma(h)=h$ and $\sigma(x)=ix$. Notice, that of these cases, the situation described in , and fit into the pattern of the results appearing in Subsection \[subsection:const\], . Now we consider the algebra $A''_{C_4}$. The set $\left\{1,g,g^2,g^3,x,gx,g^2x,g^3x \right\}$ is a basis of $A''_{C_4}$ with decomposition $E_{{\mathcal{S}}^2,1}=\langle1,g,g^2,g^3\rangle_{{{\mathbb K}}}$, $E_{{\mathcal{S}}^2,-1}=\langle x,gx,g^2x,g^3x\rangle_{{{\mathbb K}}}$ and with the following normalized integrals: $$\ell=\left(1+g+g^2+g^3\right)x ,\quad r=\left(-1+g-g^2+g^3\right)x,\quad \lambda=x^*,\quad \rho=\left(g^3x\right)^*.$$ The modular element is $a=g$ and the modular function $\alpha$ is defined by $\alpha(g)=-1$ and $\alpha(x)=0$. Suppose there exists a companion automorphism $\sigma$ in $A''_{C_4}$. From we get $\sigma(g)=g$. Then the condition $gx + xg=0$ implies $\sigma(x)=b_0x+b_1gx+b_2g^2x+b_3g^3x$, for some $b_0,b_1,b_2,b_3\in{{\mathbb K}}$. Using for $\lambda$ and $\rho$ we have $\sigma(x)=r_\sigma x+b_1gx+b_2g^2x$, where $r_\sigma^2=-1$. Now using for $\ell$ and $r$ we conclude $\sigma(x)=r_\sigma x$. So $\sigma$ verifies $\sigma(g)=g$ and $\sigma(x)=r_\sigma x$, but then $\sigma$ can not preserve the relation $x^2=g^2-1$. Hence we have shown that $A''_{C_4}$ is not almost involutive. As the property of being almost involutive is preserved by duality, we deduce that $\left( A''_{C_4}\right)^*$ is also not almost involutive. Note that $A''_{C_4}$ is pointed but $\left( A''_{C_4}\right)^*$ it is not –see [@kn:stefan]–. \[ej:12\] Let $H$ be a Hopf algebra of dimension $12$. Natale shows in [@kn:natale] that if $H$ is non semisimple, then $H$ or $H^\vee$ is pointed; she also shows that if $H$ is pointed, then it is isomorphic to one and only one of the following list: - $A_{0}=\left\langle g,x:\ g^6=1,\ x^2=0,\ gx + xg=0 \right\rangle$, $g$ is a group-like element and $x$ is $(g,1)$-primitive. - $A_{1}=\left\langle g,x:\ g^6=1,\ x^2=1-g^2,\ gx + xg=0 \right\rangle$, $g$ is a group-like element and $x$ is $(g,1)$-primitive. - $B_{0}=\left\langle g,x:\ g^6=1,\ x^2=0,\ gx + xg=0 \right\rangle$, $g$ is a group-like element and $x$ is $\left(g^3,1\right)$-primitive. - $B_{1}=\left\langle g,x:\ g^6=1,\ x^2=0,\ gx -\omega xg=0 \right\rangle$, $g$ is a group-like element, $x$ is $\left(g^3,1\right)$-primitive and $\omega\in{{\mathbb K}}$ is a primitive root of unity of order 6. The Hopf algebras in this list satisfy the following: $A_0^*=B_1$, $B_0^*=B_0$ and $A_1^*$ is not pointed. Moreover, $A_0$, $B_0$ and $B_1$ are of the type appearing in Subsection \[subsection:const\], . The algebras in this list appear analogous to the ones in dimension 8, so we can expect that they have similar properties. Indeed, we have that $A_0$, $B_0$ and $B_1$ are almost involutive but $A_1$ –and so its dual– is not. The proof that $A_0$, $B_0$ and $B_1$ are almost involutive follows a similar pattern than the eight dimensional case. For $A_1$, if it has a companion morphism $\sigma$, then similarly than for $A''_{C_4}$ we obtain that $\sigma(g)=g$ and also prove the existence of scalars $a,b\in{{\mathbb K}}$ such that $\sigma(x)=r_\sigma x + a\left(gx-g^3x \right)+b\left(gx^2-g^4x \right)$. Being $\sigma$ a Hopf algebra map, then ${\mathcal{S}}(\sigma(x))=\sigma({\mathcal{S}}(x))$, and this relation implies $\sigma(x)=r_\sigma x$ and we obtain the same contradiction than for $A''_{C_4}$. The Hopf algebras of dimension 13, 14 and 15 are semisimple –see [@kn:beattie-gaston]– and the ones of dimension $n\leq 11$ and $n\ne 8$ are semisimple or Taft algebras –see [@kn:stefan]–. Semisimple Hopf algebras are involutive and in the example \[example:taft\] below we show that the Taft algebra is almost involutive. Hence, of the Hopf algebras of dimension $n\leq 15$, the only cases when there are non almost involutive examples is for $n=8$ or $n=12$. Appendix: square roots of finite order linear automorphisms. ============================================================ We start with some elementary considerations about the square root of a finite order linear automorphism $D:V \rightarrow V$ where $V$ is a finite dimensional vector space over a field $\mathbb K$. We call $m=|D|$ the order of $D$. Given such $m$, we take $r \in \mathbb K$ with the property that its order is $|r|=2m$, if $m$ is even or $|r|=m$, if $m$ is odd. We call $q=r^2$. Notice that $|q|=m$. Define $\mathcal E=\big\{0 \leq i \leq m-1: q^i \in \operatorname{Spec}(D)\big\}$; then $V=\bigoplus_{i \in \mathcal E}E_{D,q^i}$ where $E_{D,q^i}=\{x \in H: Dx=q^i x\}$. Assume that $\sigma:V \rightarrow V$ is a linear automorphism of $V$ such that $\sigma^2=D$. Any such $\sigma$ will satisfy that $\sigma^2|_{E_{D,q^i}}=q^i\operatorname{id}$ and hence, the minimal polynomial $m_{\sigma|_{E_{D,q^i}}} | (t^2-q^i)=t^2-r^{2i}=(t-r^i)(t+r^i)$. Then for all $i \in \mathcal E$ we can find two subspaces $E_{\sigma,r^i},E_{\sigma,-r^i} \subseteq E_{D,q^i}$ –one of them could be $\{0\}$– such that $E_{D,q^i}= E_{\sigma,r^i} \oplus E_{\sigma,-r^i}$ and $\sigma|_{E_{\sigma,\pm r^i}}=\pm r^i\operatorname{id}$. Conversely, if for every $i \in \mathcal E$ we are given an arbitrary direct sum decomposition of $E_{D,q^i}=V_{+,i} \oplus V_{-,i}$ as above, then we can define an operator $\sigma: V\rightarrow V$, by requiring that its restriction to each of the summands are $\pm r^i\operatorname{id}$. In other words, if we write $x \in E_{D,q^i}$ as $x=x_+ + x_-$ with $x_\pm \in V_{\pm,i}$, then $\sigma(x)=r^i x_{+} - r^i x_-$. By construction $\sigma^2=D$ on $E_{D,q^i}$ for all $i\in \mathcal E$ and then $\sigma$ is a square root of $D$ on all of $V$. For such $\sigma$ we have that for all $i \in \mathcal E$: $E_{\sigma,\pm r^i}=V_{\pm,i}$. Hence, to define a *linear transformation* that is a square root of $D$, we have to take for each eigenspace $E_{D,q^i}$ of $D$ with $i \in \mathcal E$, a decomposition on two subspaces $E_{D,q^i}=V_{+,i} \oplus V_{-,i}$. Given the decomposition, a square root is defined by the equations: $\sigma|_{V_{\pm,i}}=\pm r^i\operatorname{id}$. The case of an automorphism of Hopf algebras -------------------------------------------- Assume now that $H$ is a finite dimensional Hopf algebra and that $D:H \rightarrow H$ is an *automorphism* of Hopf algebras of order $|D|=m$. We want to find conditions for the pair of subspaces $V_{+,i}$ and $V_{-,i}$ that guarantee that the $\sigma$ thus defined is a Hopf algebra automorphism. The elementary results that we present below, are probably well known, we wrote them for the lack of an adequate reference. \[obse:Dcase\]In the situation above with $H$ a finite dimensional Hopf algebra and $D$ a linear automorphism of finite order $m$. 1. $D$ is a morphism of algebras if and only if the following holds: 1. $1 \in E_{D,1}$;\ 2. For $i,j \in \mathcal E$, $E_{D,q^i}E_{D,q^j} \subseteq \begin{cases}E_{D,q^{i+j}}\quad &\text{if}\quad i+j<m;\\ E_{D,q^{i+j-m}} &\text{if}\quad i+j \geq m.\end{cases}$ 2. $D$ is a morphism of coalgebras if and only if the following holds: 1. If $0 \neq i \in \mathcal E$, then $\varepsilon(E_{D,q^i})=0$ ; 2. For $i \in \mathcal E$: $$\Delta(E_{D,q^i}) \subseteq \bigoplus_{\{a,b \in \mathcal E: a+b=i\}}(E_{D,q^a} \otimes E_{D,q^b}) \quad \oplus \bigoplus_{\{a,b \in \mathcal E: a+b=i+m\}}(E_{D,q^a} \otimes E_{D,q^b}).$$ Observe that in the considerations above we used that if $r,s \in \mathcal E$, then $0 \leq r,s \leq m-1$, and then $0 \leq r+s \leq 2m-2$. \[theo:algebraconditions\] Let $A$ be an algebra and let $D:A \rightarrow A$ be an automorphism of algebras of finite order $m$. Define as above, $q,r$, $\mathcal E$ and $E_{D,q^i}$ –for $i \in \mathcal E$–. For each $i \in \mathcal E$ when we take an arbitrary decomposition of $E_{D,q^i}=V_{+,i}\oplus V_{-,i}$ and define a linear transformation $\sigma$ on $A$ as: $\sigma|_{V_{\pm,i}}=\pm r^i\operatorname{id}$ for all $i \in \mathcal E$, then $\sigma^2=D$. Moreover, $\sigma$ is an automorphism of algebras if and only if the following conditions hold: 1. $1 \in V_{+,0}$*;* 2. 1. If $0 \leq i+j \leq m-1$, then $V_{+,i} V_{+,j}+ V_{-,i}V_{-,j} \subseteq V_{+,i+j}$ and $V_{+,i} V_{-,j}+ V_{-,i}V_{+,j} \subseteq V_{-,i+j}$. 2. If $m \leq i+j \leq 2m-2$ then*:* 1. If $m$ is even, then $V_{+,i} V_{+,j}+ V_{-,i}V_{-,j} \subseteq V_{-,i+j-m}$ and $V_{+,i} V_{-,j}+ V_{-,i}V_{+,j} \subseteq V_{+,i+j-m}$*;* 2. If $m$ is odd, then $V_{+,i} V_{+,j}+ V_{-,i}V_{-,j} \subseteq V_{+,i+j-m}$ and $V_{+,i} V_{-,j}+ V_{-,i}V_{+,j} \subseteq V_{-,i+j-m}$. <!-- --> 1. *Conditions for the unit.* The unit element, $1 \in E_{D,q^0}=E_{D,1}$ and as we want that $\sigma(1)=1$, in the decomposition of $E_{D,1}=V_{+,0} \oplus V_{-,0}$, $1 \in V_{+,0}$. 2. *Multiplicativity.* It is enough to prove that for all $i,j \in \mathcal E, x \in V_{\pm,i}, y \in V_{\pm,j} \Rightarrow \sigma(xy)=\sigma(x)\sigma(y)$. Being $D(xy)=q^{i+j}xy$, we have two alternatives: 1. If $xy=0$, then $\sigma(x)\sigma(y)=(\pm r^ix)(\pm r^jy)=0=\sigma(xy)$. 2. If $xy \neq 0$, $q^{i+j}\in \operatorname{Spec}(D)$ and $\exists k \in \mathcal E, i+j \equiv k(\!\!\!\mod m)$. Then: $k= \begin{cases}i+j\quad &\text{for}\,\, 0 \leq i+j \leq m-1;\\ i+j-m\quad &\text{for}\,\, m \leq i+j\leq 2m-2.\end{cases}$ As $xy \in E_{D,q^k}=V_{+,k} \oplus V_{-,k}$ for $k \in \mathcal E$, we can find $(xy)_\pm \in V_{\pm,k}$, such that: $xy=(xy)_++(xy)_-\,,\, \sigma(xy)=r^k(xy)_+-r^k (xy)_{-}$.\ We consider the following alternatives.\ (A) $x \in V_{+,i}$ and $y \in V_{+,j}$ or $x \in V_{-,i}$ and $\in V_{-,j}$. In this case $\sigma(x)\sigma(y)=r^{i+j}xy=r^{i+j}(xy)_{+}+r^{i+j}(xy)_{-}$. Now, if $i+j=k \leq m-1$ then $r^{i+j}=r^k$ and in accordance with the above formulæ  the multiplicativity holds if and only if $(xy)_{-}= 0$. If $m \leq i+j=k+m$, then $r^{i+j}=r^kr^m=\begin{cases}-r^k &\quad \text{if $m$ is even}\\ \,\,\,\, r^k &\quad \text{if $m$ is odd}.\end{cases}$\ Then, the multiplicativity holds if and only if $\begin{cases}(xy)_+=0 &\,\text{if $m$ is even}\\ (xy)_-=0 &\,\text{if $m$ is odd}.\end{cases}$\ (B) $x \in V_{+,i}$ and $y \in V_{-,j}$ or $x \in V_{-,i}$ and $y \in V_{+,j}$. In this case $\sigma(x)\sigma(y)=-r^{i+j}xy=-r^{i+j}(xy)_+-r^{i+j}(xy)_-$. Now, if $i+j=k \leq m-1$ then $r^{i+j}=r^k$ and the multiplicativity holds if and only if $(xy)_+ = 0$. If $m \leq i+j=k+m$, then $r^{i+j}=r^kr^m=\begin{cases}-r^k &\quad \text{if $m$ is even}\\ \,\,\,\,r^k &\quad \text{if $m$ is odd}.\end{cases}$\ Hence, in this situation the multiplicativity holds if and only if $\begin{cases}(xy)_-=0 &\,\text{if $m$ is even}\\ (xy)_+=0 &\,\text{if $m$ is odd}.\end{cases}$ As we are dealing with finite dimensional objects, we may proceed by duality and obtain the following result: \[theo:coalgebraconditions\] Let $C$ be a coalgebra and let $D:C \rightarrow C$ be a automorphism of coalgebras of finite order $m$. Define as above, $q,r$, $\mathcal E$ and $E_{D,q^i}$ –for $i \in \mathcal E$–. For each $i \in \mathcal E$ when we take an arbitrary decomposition of $E_{D,q^i}=V_{+,i}\oplus V_{-,i}$ and define a linear transformation $\sigma$ on $C$ as: $\sigma|_{V_{\pm,i}}=\pm r^i\operatorname{id}$, then $\sigma^2=D$. Moreover, $\sigma$ is an automorphism of coalgebras if and only if the following conditions hold: 1. $\varepsilon(V_{-,0})=0$ 2. $$\Delta(V_{+,i}) \subseteq \bigoplus_{\{a,b\in \mathcal E:a+b=i\}}(V_{+,a} \otimes V_{+,b} + V_{-,a} \otimes V_{-,b}) \oplus \begin{cases}\bigoplus_{\{a,b\in \mathcal E:a+b=i+m\}}(V_{+,a} \otimes V_{-,b} + V_{-,a} \otimes V_{+,b})\,,\, \text{m even};\\\bigoplus_{\{a,b\in \mathcal E:a+b=i+m\}}(V_{+,a} \otimes V_{+,b} + V_{-,a} \otimes V_{-,b})\,,\, \text{m odd}.\end{cases}$$ $$\Delta(V_{-,i}) \subseteq \bigoplus_{\{a,b\in \mathcal E:a+b=i\}}(V_{+,a} \otimes V_{-,b} + V_{-,a} \otimes V_{+,b}) \oplus \begin{cases}\bigoplus_{\{a,b\in \mathcal E:a+b=i+m\}}(V_{+,a} \otimes V_{+,b} + V_{-,a} \otimes V_{-,b})\,,\, \text{m even};\\\bigoplus_{\{a,b\in \mathcal E:a+b=i+m\}}(V_{+,a} \otimes V_{-,b} + V_{-,a} \otimes V_{+,b})\,,\, \text{m odd}.\end{cases}$$ \[coro:Hopf-conditions\] Let $H$ be a Hopf algebra and let $D:H \rightarrow H$ be an automorphism of Hopf algebras of finite order $m$. Define as above, $q,r$, $\mathcal E$ and $E_{D,q^i}$ –for $i \in \mathcal E$–. For each $i \in \mathcal E$ we take an arbitrary decomposition of $E_{D,q^i}=V_{+,i}\oplus V_{-,i}$ and define a linear transformation $\sigma$ on $C$ as: $\sigma|_{V_{\pm,i}}=\pm r^i\operatorname{id}$. If the hypothesis of theorems \[theo:algebraconditions\] and \[theo:coalgebraconditions\] are simultaneously satisfied, then $\sigma$ is a Hopf algebra automorphism and $\sigma^2=D$. A particular situation. ----------------------- We consider the following special cases of the above Corollary \[coro:Hopf-conditions\]. Assume that the splitting of the eigenspaces $E_{D,q^i}$ is trivial: $$V_{+,i}=E_{D,q^i}\,\,\,\,\text{and}\quad V_{-,i}=0.$$ In this situation, and using the considerations of Observation \[obse:Dcase\], it is clear that some of the conditions of Corollary \[coro:Hopf-conditions\] –i.e. of Theorems \[theo:algebraconditions\] and \[theo:coalgebraconditions\]– are authomatically verified. In particular the case $m$ odd becomes conditionless. Hence, we have the following particular result that in some cases provides an answer for the existence of a square root of a Hopf automorphism that is both multiplicative and comultiplicative. \[coro:sqrpart\] Assume that we are in the situation above. If $m$ is odd, then the square root of $D$ associated to the family of subspaces $V_{+,i}=E_{D,q^i}$ and $V_{-,i}=0$ is an automorphism of Hopf algebras. If $m$ is even, it is an automorphism of Hopf algebras if and only if $E_{D,q^i} E_{D,q^j}=0 $ for all $i,j \in \mathcal E$ such that $m \leq i+j \leq 2m-2$ and $\Delta(E_{D,q^i}) \subseteq \bigoplus_{\{a,b\in \mathcal E:a+b=i\}}(E_{D,q^a} \otimes E_{D,q^b})$ for all $i\in \mathcal E$. Observe that the case in which $m$ is odd has already been treated by an elementary reasonement in Observation \[obse:oddtrivial\]. \[example:taft\] The Taft algebra $T_n$ is a generalization of the Sweedler’s algebra $H_4$. Let be $\omega\in{{\mathbb K}}$ a primitive root of order $n$ of 1 and $$T_n=\left\langle g,x:\ g^n=1,\ x^n=0,\ gx-\omega xg=0\right\rangle.$$ $T_n$ is a Hopf algebra with coalgebra structure given by $g$ being a group-like element and $x$ a $(g,1)$-primitive element. The set $\{g^rx^p:\ 0\leq r,p\leq n-1\}$ is a basis of $T_n$, so $\dim T_n=n^2$. We have ${\mathcal{S}}^2(g)=g$ and ${\mathcal{S}}^2(x)=\omega x$, therefore the order of ${\mathcal{S}}^2$ is $n$. The eigenvalues of ${\mathcal{S}}^2$ are $\{1,\omega,\cdots,\omega^{n-1}\}$, and the corresponding eigenspaces are $E_{{\mathcal{S}}^2,\omega^i}=\{x^i,gx^i,\cdots,g^{n-1}x^i\}_{\mathbb K}$. With regard of the conditions of Corollary \[coro:sqrpart\], we have that $E_{{\mathcal{S}}^2,\omega^i}E_{{\mathcal{S}}^2,\omega^j}=\{g^k \omega^{i+j}:k=0,\cdots,n-1\}$. Hence $E_{{\mathcal{S}}^2,\omega^i}E_{{\mathcal{S}}^2,\omega^j}=0$ for $i+j \geq n$ as required. Morever, as $\Delta(g^kx^i)=(g^k \otimes g^k)(x \otimes g + 1 \otimes x)^i$, it is clear that $\Delta(E_{{\mathcal{S}}^2,\omega^i}) \subseteq \sum_{a+b=i} E_{{\mathcal{S}}^2,\omega^a} \otimes E_{{\mathcal{S}}^2,\omega^b}$. Hence, in this manner we prove that $T_n$ is almost involutive. A direct verification shows that if we take $\nu\in{{\mathbb K}}$, $\nu^2=\omega$. The maps $\sigma_\pm$ defined as $\sigma_\pm(g)=g$ and $\sigma_\pm(x)=\pm \nu x$ are companion automorphisms. \[example:general\] Assume that we have a trivial extension of a finite dimensional Hopf algebra, with the additional property that $K=E_1$. The spectral decomposition of $H$ with respect to ${\mathcal{S}}^2$ becomes: $$H= E_{{\mathcal{S}}^2,1} \oplus \bigoplus_{i \in \mathcal E_M}E_{{\mathcal{S}}^{2},q^i} \quad \text{with}\quad M=\bigoplus_{i \in \mathcal E_M}E_{{\mathcal{S}}^{2},q^i}.$$ In this case it is clear that the conditions of Corollary \[coro:sqrpart\] are satisfied. Indeed, if we look at the condition regarding the product, the only cases in which the sum of exponents of the corresponding eigenvalues of two eigenvectors may be larger than $m$, is for the case that the exponents of the eigenvectors are in $\mathcal E_M$. In this case, the condition $M^2=0$, guarantees that the product of the corresponding eigenspaces is trivial. An argument along the same lines and using the condition that $M$ is a $K$–bicomodule, shows that the condition regarding the coproduct in Corollary \[coro:sqrpart\] is satisfied. [99]{} Abella, A., Ferrer, W, and Haim, M. [*Compact coalgebras, compact quantum groups and the positive antipode*]{}, São Paulo Journal of Mathematical Sciences [**3**]{}, 1, 2009, pp. 191–227. Abella, A., Ferrer, W, and Haim, M. [*Some constructions of compact quantum groups*]{}, Sao Paulo Journal of Mathematical Sciences, [**6**]{}, 1, 2012, pp. 1-40. Beattie, M., Bulacu, D. and Torrecillas, B. [*Radford’s $S^4$ formula for co-Frobenius Hopf algebras.*]{} J. Algebra, [**307**]{} , 1, (2007), pp. 330-342. Beattie, M., García, A. [*Classifying Hopf algebras of a given dimension*]{} Preprint. arXiv:1206.6529 \[math.QA\] Dascalescu, S., Nastasescu, C., Raianu, S. [*Hopf Algebras: An Introduction*]{}, Monographs and textbooks in Pure and Applied Mathematics, vol. [**235**]{}, Marcel Dekker, New York, 2001. Majid, S., [*Foundations of Quantum Group Theory*]{}, Cambridge Univ. Press, “Cambridge”, 1995. Montgomery, S. [*Hopf algebras and their actions on rings*]{}. CBMS [**28**]{}, Amer. Math. Soc.: 1993. Capítulo 7 ([*Crossed products*]{}) Natale, S. [*Hopf algebras of dimension 12.*]{} Algebr. Represent. Theory, [**5**]{} , 5, (2002), pp. 445-455. Radford, D. [*The Order of the Antipode of a Finite Dimensional Hopf Algebra is Finite*]{}. Am. Jour. of Math., Vol. 98, No. 2 (Summer, 1976), 333–355. Radford, D, and Schneider, H.[*On the even powers of the antipode of a finite dimensional Hopf algebra*]{}, Journal of algebra, Vol. 251, No. 1 (2002), 185–212. Schneider, H. [*Lectures on Hopf algebras*]{}, notes by S. Natale. Trabajos de matemática, vol 31/95, 1995. FaMAF, Córdoba, Argentina. Ştefan, D. [*Hopf algebras of low dimension.*]{} J. Algebra, [**211**]{} , 2, (1998), pp. 535-556. Woronowicz, S.L. [*Twisted $\operatorname{SU}(2)$ group, An example of non commutative differential calculus*]{}, Publ. RIMS, Kyoto Univ. Vol 23, pp 171–181. [^1]: The second author would like to thank Csic-UDELAR, Conycit-MEC, Uruguay and Anii, Uruguay.

ArXiv

--- author: - 'Erwan Faou, Benoît Grébert and Eric Paturel' title: | Birkhoff normal form and splitting methods\ for semi linear Hamiltonian PDEs.\ Part I: Finite dimensional discretization. --- Introduction ============ In this work, we consider a class of Hamiltonian partial differential equations whose Hamiltonian functionals $H=H_0+P$ can be divided into a linear unbounded operator $H_0$ with discrete spectrum and a nonlinear function $P$ having a zero of order at least $3$ at the origin of the phase space. Typical examples are given by the nonlinear wave equation or the nonlinear Schrödinger equation on the torus. We consider discretizations of this PDEs and denote by $H^{(K)}=H_0^{(K)}+P^{(K)}$ the corresponding discrete Hamiltonian, where $K$ is a discretization parameter. Typically, $K$ denotes a spectral parameter in a collocation method. Amongst all the numerical schemes that can be applied to these Hamiltonian PDEs, splitting methods entail many advantages, as they provide symplectic and explicit schemes, and can be easily implemented using fast Fourier transform if the spectrum of $H_0$ expresses easily in Fourier basis. Generally speaking, a splitting scheme is based on the approximation $$\label{E001} \varphi_{H^{(K)}}^h \simeq \varphi_{H_0^{(K)}}^h \circ \varphi_{P^{(K)}}^h$$ for small time $h$, and where $\varphi_{Q}^t$ denotes the exact flow of the Hamiltonian system associated with the Hamiltonian function $Q$. For a given time $t = nh$, $n \in {\mathbb{N}}$, the solution starting at some initial value $z^0$ is then approximated by $$\label{Ezn} \varphi_{H^{(K)}}^t(z^0) \simeq z^n = \Big(\varphi_{H_0^{(K)}}^h \circ \varphi_{P^{(K)}}^h\Big)^n (z^0).$$ The understanding of the long-time behavior of splitting methods for Hamiltonian PDEs is a fundamental ongoing challenge in the field of geometric integration, as the classical arguments of [*backward error analysis*]{} (see for instance [@HLW]) do not apply in this situation, where the frequencies of the system are arbitrary large, and where resonances phenomena are known to occur for some values of the step size. For example, considering the case of the Schrödinger equation on the one dimensional torus, the eigenvalues of $H_0^{(K)}$ range from $1$ to $K^2$ and the assumption $h K^{2} << 1$ used in the finite dimensional situation becomes drastically restrictive in practice. Recently, many progresses have been made in the understanding of the long time behaviour of numerical methods applied to Hamiltonian PDEs. A first result using normal form techniques was given by [Dujardin & Faou]{} in [@DF07] for the case of the linear Schrödinger equation with small potential. Concerning the nonlinear case, results exists by [Cohen, Hairer & Lubich]{}, see [@CHL08b; @CHL08c], for the wave equation and [Gauckler & Lubich]{}, see [@GL08a; @GL08b], for the nonlinear Schrödinger equation using the technique of modulated Fourier expansion. However to be valid these results use non-resonance conditions that are generically satisfied only under CFL conditions linking the step-size $h$ and the highest frequencies of the discretized Hamiltonian PDE. Normal form techniques have proven to be one of the most important tools for the understanding of the long time behaviour of Hamiltonian PDE (see [@Bam03; @BG06; @Greb07; @Bam07; @BDGS; @GIP]). Roughly speaking, the dynamical consequences of such results are the following: starting with an small initial value of size $\varepsilon$ in a Sobolev space $H^s$, then the solution remains small in the same norm over long time, namely for time $t \leq C_r \varepsilon^{-r}$ for arbitrary $r$ (with a constant $C_r$ depending on $r$). Such results hold under [*generic*]{} non resonance conditions on the frequencies of the underlying linear operator $H_0$ associated with the Hamiltonian PDE, that are valid in a wide number of situations (nonlinear Schrödinger equation on a torus of dimension $d$ or with Dirichlet boundary conditions, nonlinear wave equation with periodic or Dirichlet conditions in dimension 1, Klein Gordon equation on spheres or Zoll manifolds.). This work is the first of a series of two. In this paper, we consider full discretizations of the Hamiltonian PDE, with a spectral discretization parameter $K$ that is [*finite but large*]{}. We show that under the hypothesis $K \leq \varepsilon^{-\sigma}$ for some constant $\sigma$ depending on the precision degree $r$ then the [*actions*]{} of the initial value are almost preserved over a very large number of iterations $n \leq C_r \varepsilon^{-r}$, provided the initial solution is small (of order $\varepsilon$) in $L^2$ norm. These actions can be interpreted as the oscillatory energies corresponding to an eigenvalue of $H_0^{(K)}$. Moreover, the $L^2$ norm of this numerical solution remains small for this large number of iterations. The method used in this situation is by essence a finite dimensional Birkhoff normal form result (explaining why we work here essentially with the $L^2$ norm). Using a [*generic*]{} non resonance condition on the step size that turns out to be valid for many equations and that is independent on $K$, we mainly show that we can take $K$ asymptotically large without altering the nature of the classical result. Our main result is given by Theorem \[Tmain\]. Roughly speaking, the method consists in applying techniques that are now standard in normal form theory, by tracking the dependence in $K$ of the constants appearing in the estimates. The use of a non resonance condition that is independent of $K$ is however crucial, and reflects the infinite dimensional nature of the initial continuous problem without space approximation. In some sense, the second paper [@FGP2] studies the case where $K > \varepsilon^{-\sigma}$ by considering the splitting method where no discretization in space is made (i.e. $K = +\infty$). The techniques used involve the abstract framework developed in [@BG06; @Greb07; @Bam07]. However, instead of being valid for the (exact) abstract splitting , we have to consider [*rounded*]{} splitting methods of the form $$\label{E002} \Pi_{\eta,s}\circ \varphi_{H_0}^h\circ \varphi_P^h$$ where $\Pi_{\eta,s}$ puts to zero all the frequencies $\xi_j$ whose weighted energy $|j|^{2s}|\xi_j|^{2s}$ in the Sobolev space $H^s$ is smaller than a given threshold $\eta^2$. Hence, for small $\eta$, is very close to the exact splitting method . The good news is that this threshold can be taken of the order $\varepsilon^r$, making this projection $\Pi_{\eta,s}$ very close to the identity, and in any case producing an error that is far beyond the round-off error in a computer simulation (particularly for large $s$). Description of the method ========================= Before going on into the precise statements and proofs of this work, we would like to give tentative explanations of the restrictions observed in comparison with the continuous case. The method used in [@BG06] to prove the long-time conservation of Sobolev norms and the associated weighted actions for small data is to start from a Hamiltonian $H = H_0 + P$ depending on an infinite number of variable $(\xi_j,\eta_k)$, $j, k \in {\mathbb{N}}$, and for a fixed number $r$, to construct a Hamiltonian transformation $\tau$ close to the identity, and such that in the new variable, the Hamiltonian can be written $$\label{EnfBG} H_0 + Z + R$$ where $Z$ is a real Hamiltonian depending only on the action $I_j = \xi_j\eta_j$ and $R$ a real Hamiltonian having a zero of order $r$. The key for this construction is an induction process with, at each step, the resolution of an homological equation of the form $$\label{HomBG} \{ H_0, \chi\} + Z = G$$ where $G$ is a given homogeneous polynomial of order $n$, and where $Z$, depending only on the actions, and $\chi$ are unknown. Assume that $G$ is of the form $$G = G_{{{\boldsymbol{j}}}{{\boldsymbol{k}}}} \,\xi_{j_1}\cdots \xi_{j_p}\eta_{k_1}\cdots \eta_{k_q}$$ where $G_{{{\boldsymbol{j}}}{{\boldsymbol{k}}}}$ is a coefficient, ${{\boldsymbol{j}}}= (j_1,\ldots,j_p) \in {\mathbb{N}}^p$ and ${{\boldsymbol{k}}}= (k_1,\ldots,k_q) \in {\mathbb{N}}^q$. Then it is easy to see that the equation can be written $$\label{HomBG2} \Omega({{\boldsymbol{j}}},{{\boldsymbol{k}}}) \chi_{{{\boldsymbol{j}}}{{\boldsymbol{k}}}} + Z_{{{\boldsymbol{j}}}{{\boldsymbol{k}}}} = G_{{{\boldsymbol{j}}}{{\boldsymbol{k}}}}$$ where $$\Omega({{\boldsymbol{j}}},{{\boldsymbol{k}}}) = \omega_{j_1} + \cdots + \omega_{j_p} - \omega_{k_1} -\cdots - \omega_{j_q}$$ and where $Z_{{{\boldsymbol{j}}}{{\boldsymbol{k}}}}$ and $\chi_{{{\boldsymbol{j}}}{{\boldsymbol{k}}}}$ are unknown coefficients. It is clear that for ${{\boldsymbol{j}}}= {{\boldsymbol{k}}}$ (up to a permutation), we have $\Omega({{\boldsymbol{j}}},{{\boldsymbol{k}}}) = 0$ which imposes $Z_{{{\boldsymbol{j}}}{{\boldsymbol{k}}}} = G_{{{\boldsymbol{j}}}{{\boldsymbol{k}}}}$. When ${{\boldsymbol{j}}}\neq {{\boldsymbol{k}}}$ (taking into account the permutation), the solution of relies on a non resonance conditions on the small divisors $\Omega({{\boldsymbol{j}}},{{\boldsymbol{k}}})^{-1}$. In [@BG06], [Bambusi & Grébert]{} use a non resonance condition of the form $$\label{EnonresBG} \forall\, {{\boldsymbol{j}}}\neq {{\boldsymbol{k}}},\quad | \Omega({{\boldsymbol{j}}},{{\boldsymbol{k}}}) | \geq \gamma \mu({{\boldsymbol{j}}},{{\boldsymbol{k}}})^{-\alpha}$$ where $\mu({{\boldsymbol{j}}},{{\boldsymbol{k}}})$ denotes the third largest integer amongst $|j_1|,\ldots,|k_q|$. They moreover show that such a condition is guaranteed in a large number of situations (see [@BG06], [@Greb07] or [@Bam07] for precise results). Considering now the splitting method $\varphi_{H_0}^h \circ \varphi_{P}^h$, we see that we cannot work directly at the level of the Hamiltonian. To avoid this difficulty, we embed the splitting into the family of applications $$[0,1] \ni \lambda\mapsto \varphi_{H_0}^h \circ \varphi_{hP}^\lambda$$ and we derive this expression with respect to $\lambda$, in order to work in the tangent space, where it is much easier to identify real Hamiltonian than unitary flows. This explains why we deal here with time-dependent Hamiltonian. Note that we do not expand the operator $\varphi_{H_0}^h$ in powers of $h$, as this would yields positive powers of the unbounded operator $H_0$ appearing in the series. Unless a CFL condition is employed, this methods do not give the desired results (and do not explain the resonance effects observed for some specific values of $h$). Now, instead of , the Homological equation appearing for the splitting methods is given in a discrete form $$\label{HomFGP} \chi \circ \varphi_{H_0}^h - \chi + Z = G.$$ In terms of coefficients, this equations yields $$(e^{ih\Omega({{\boldsymbol{j}}},{{\boldsymbol{k}}})} - 1) \chi_{{{\boldsymbol{j}}}{{\boldsymbol{k}}}} + Z_{{{\boldsymbol{j}}}{{\boldsymbol{k}}}} = G_{{{\boldsymbol{j}}}{{\boldsymbol{k}}}}.$$ The main difference with is that we have to avoid not only the indices $({{\boldsymbol{j}}},{{\boldsymbol{k}}})$ so that $\Omega({{\boldsymbol{j}}},{{\boldsymbol{k}}}) = 0$, but all of those for which $h \Omega({{\boldsymbol{j}}},{{\boldsymbol{k}}}) = 2 m \pi$ for some (unbounded) integer $m$. In the case of a fully discretized system for which $\nabla_{z_j} P \equiv 0$ for $|j| > K$, then under the CFL-like condition of the form $h K^m\leq C$ where $m$ depends on the growth of the eigenvalues of $H_0$ and $C$ depends on $r$, then we have $|h \Omega({{\boldsymbol{j}}},{{\boldsymbol{k}}})| \leq \pi$, and hence $$\label{EnonresGL} |e^{ih\Omega({{\boldsymbol{j}}},{{\boldsymbol{k}}})} - 1| \geq h \gamma \mu({{\boldsymbol{j}}},{{\boldsymbol{k}}})^{-\alpha}$$ is then a consequence of . Under this assumption, we can apply the same techniques used in [@BG06] and draw the same conclusions. This is the kind of assumption made in [@CHL08b] and [@GL08b]. The problem with is that it is non generic in $h$ outside the CFL regime. For example, in the case of the Schrödinger equation, the frequencies of the operator $H_0$ are such that $\omega_j \simeq j^2$. Hence, for large $N$, if $(j_1,\ldots,j_p,k_1,\ldots,k_q)$ is such that $j_1 = N +1$, $k_1 = N$ and all the other are of order $1$ ($N$ is large here), we have $\Omega({{\boldsymbol{j}}},{{\boldsymbol{k}}}) \simeq (N+1)^2 - N^2 \simeq 2N$. Hence, $$|e^{ih\Omega({{\boldsymbol{j}}},{{\boldsymbol{k}}})} - 1| \simeq | e^{2ihN} - 1 |$$ cannot be assumed to be greater than $h \gamma \mu({{\boldsymbol{j}}},{{\boldsymbol{k}}})^{-\alpha} \simeq h$ for all (large) $N$. Note that a generic hypothesis on $h$ would be here that this small divisor is greater than $h \gamma N^{-\alpha}$ for some constants $\gamma$ and $\alpha$. This example shows that we cannot control the small divisors $|e^{ih\Omega({{\boldsymbol{j}}},{{\boldsymbol{k}}})} - 1|$ associated with the splitting scheme by the [*third largest*]{} integer in the multi index (which is actually of order 1 in this case), but by the [*largest*]{}. Using a generic condition on $h \leq h_0$, we prove in [@FGP2] a normal form result and show that the flow is conjugated to the flow of a Hamiltonian vector field of the form , but where $Z$ now contains terms depending only on the actions, and supplementary terms containing at least two large indices. Here, large means greater than $\varepsilon^{-\sigma}$ where $\sigma$ depends on $r$. In the case of a full discretization of the Hamiltonian PDE with a spectral discretization parameter $K$, we thus see that if $K \leq \varepsilon^{-\sigma}$ then the normal form term $Z$ actually depends only on the actions, as the high frequencies greater that $\varepsilon^{-\sigma}$ are not present. This is essentially the result of this paper. In the case where $K > \varepsilon^{-\sigma}$, the normal form result that we obtain can be interpreted as follows: the non conservation of the actions can only come from two high modes (of order greater than $\varepsilon^{-\sigma}$) interacting together and contaminating the whole spectrum. The role of the projection operator $\Pi_{\eta,s}$ is to destroy these high modes at each step. As we can take $\eta = \varepsilon^r$, the error induced is very small, and in particular, far beyond the round-off error in the numerical simulation. This is mainly the result in [@FGP2]. The differences between the present work and [@FGP2] lie in the techniques involved: In this work, the system considered are [*large*]{} but [*finite*]{} dimensional systems, and all the hypothesis made on the nonlinearity can be expressed using elementary conditions similar to those used in the finite dimensional case. In [@FGP2], we study $K = +\infty$, which requires much more elaborate technical tools in the spirit of [@BG06; @Greb07; @Bam07]. Setting of the problem ====================== Hamiltonian formalism --------------------- We set ${\mathcal{N}}= {\mathbb{Z}}^d$ or ${\mathbb{N}}^d$. For $a = (a_1,\ldots,a_d)Ê\in {\mathcal{N}}$, we set $$|a| = \max_{i = 1,\ldots,d} |a_i|.$$ Let $K \in {\mathbb{N}}$, and let ${\mathcal{N}}_K$ a finite subset of ${\mathcal{N}}$, included in the ball $\{Êa \in {\mathcal{N}}\, | \, |a| \leq K\,\}$. Typically, we can take ${\mathcal{N}}_K$ of the form $[-K,\ldots,K]^d \subset {\mathbb{Z}}^d$ or $[0,\ldots,K]^d \subset {\mathbb{N}}^d$ or a sparse set of the form (see for instance [@G07; @L08]) $${\mathcal{N}}_K= \{ \, a = (a_1,\ldots,a_d) \in {\mathbb{Z}}^d\; | \; ( 1 + |a_1|)\cdots(1 + |a_d|) \leq K\, \} \subset {\mathbb{Z}}^d.$$ We consider the set of variables $(\xi_a,\eta_b) \in {\mathbb{C}}^{{\mathcal{N}}_K} \times {\mathbb{C}}^{{\mathcal{N}}_K}$ equipped with the symplectic structure $$\label{Esymp} i \sum_{a \in {\mathcal{N}}_K} {\mathrm{d}}\xi_a \wedge {\mathrm{d}}\eta_a.$$ We define the set ${\mathcal{Z}}_K = {\mathcal{N}}_K \times \{ \pm 1\}$. For $j = (a,\delta) \in {\mathcal{Z}}_K$, we define $|j| = |a|$ and we denote by $\overline{j}$ the index $(a,-\delta)$. We then define the variables $(z_j)_{j \in {\mathcal{Z}}_K} \in {\mathbb{C}}^{{\mathcal{Z}}_K}$ by the formula $$j = (a,\delta) \in {\mathcal{Z}}_K \Longrightarrow \left\{ \begin{array}{rcll} z_{j} &=& \xi_{a}& \mbox{if}\quad \delta = 1,\\[1ex] z_j &=& \eta_a & \mbox{if}\quad \delta = - 1, \end{array} \right.$$ By abuse of notation, we often write $z = (\xi,\eta)$ to denote such an element. We set $${\|z\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}}^2 := \sum_{j \in {\mathcal{Z}}_K} |z_j|^2$$ and for any $\rho >0$, $$B_K(\rho) = \{ \, z \in {\mathbb{C}}^{{\mathcal{Z}}_K}\, | \, {\|z\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}} \leq \rho\,\}.$$ Note that in the case where $K = +\infty$, we set by convention ${\mathcal{Z}}_K = {\mathcal{Z}}= {\mathcal{N}}\times \{\pm 1 \}$ and the previous norm defines a Hilbert structure on $\ell^2_{\mathcal{Z}}$. We denote by $$\Pi_K: \ell^2_{\mathcal{Z}}\to \big({\mathbb{C}}^{{\mathcal{Z}}_K},{\|\cdot\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}}\big)$$ the natural projection. Let ${\mathcal{U}}_K$ be a an open set of ${\mathbb{C}}^{{\mathcal{Z}}_K}$. For a function $F$ in $\mathcal{C}^1({\mathcal{U}}_K,{\mathbb{C}})$, we define its gradient as $$\nabla F(z) = \left( \frac{\partial F}{\partial z_j}\right)_{j \in {\mathcal{Z}}_K}$$ where by definition, we set for $j = (a,\delta) \in {\mathcal{N}}_K \times \{ \pm 1\}$, $$\frac{\partial F}{\partial z_j} = \left\{\begin{array}{rll} \displaystyle \frac{\partial F}{\partial \xi_a} & \mbox{if}\quad\delta = 1,\\[2ex] \displaystyle \frac{\partial F}{\partial \eta_a} & \mbox{if}\quad\delta = - 1. \end{array} \right.$$ Let $H(z)$ be a function defined on ${\mathcal{U}}_K$. If $H$ is smooth enough, we can associate with this function the Hamiltonian vector field $X_H(z)$ defined as $$X_H(z) = J \nabla H(z)$$ where $J$ is the symplectic operator induced by the symplectic form . For two functions $F$ and $G$, the Poisson Bracket is defined as $$\{F,G\} = \nabla F^T J \nabla G = i \sum_{a \in {\mathcal{N}}_K} \frac{\partial F}{\partial \eta_j}\frac{\partial G}{\partial \xi_j} - \frac{\partial F}{\partial \xi_j}\frac{\partial G}{\partial \eta_j}.$$ We say that $z\in {\mathbb{C}}^{{\mathcal{Z}}_K}$ is [*real*]{} when $z_{\overline{j}} = \overline{z_j}$ for any $j\in {\mathcal{Z}}_K$. In this case, $z=(\xi,\bar\xi)$ for some $\xi_K\in {\mathbb{C}}^{{\mathcal{N}}_K}$. Further we say that a Hamiltonian function $H$ is [*real* ]{} if $H(z)$ is real for all real $z$. With a given function $H \in \mathcal{C}^{\infty}({\mathcal{U}}_K,{\mathbb{C}})$, we associate the Hamiltonian system $$\dot z = J \nabla H(z)$$ which can be written $$\label{Eham2} \left\{ \begin{array}{rcll} \dot\xi_a &=& \displaystyle - i \frac{\partial H}{\partial \eta_a}(\xi,\eta) & a \in {\mathcal{N}}_K\\[2ex] \dot\eta_a &=& \displaystyle i \frac{\partial H}{\partial \xi_a}(\xi,\eta)& a \in {\mathcal{N}}_K. \end{array} \right.$$ In this situation, we define the flow $\varphi_H^t(z)$ associated with the previous system (for times $t \geq 0$ depending on $z \in {\mathcal{U}}_K$). Note that if $z = (\xi,\bar \xi)$ and $H$ is real, the flow $(\xi^t,\eta^t) = \varphi_H^t(z)$, for all time where it is defined, satisfies the relation $\xi^t = \bar {\eta}^t$, where $\xi^t$ is solution of the equation $$\label{Eham1} \dot\xi_a = - i \frac{\partial H}{\partial \eta_a}(\xi,\bar\xi), \quad a \in {\mathcal{N}}_K.$$ In this situation, introducing the real variables $p_a$ and $q_a$ such that $$\xi_a = \frac{1}{\sqrt{2}} (p_a + i q_a)\quad \mbox{and}\quad \bar{\xi}_a = \frac{1}{\sqrt{2}} (p_a - i q_a),$$ the system is equivalent to the system $$\left\{ \begin{array}{rcll} \dot p_a &=& \displaystyle - \frac{\partial \tilde{H}}{\partial q_a}(q,p) & a \in {\mathcal{N}}_K\\[2ex] \dot q_a &=& \displaystyle \frac{\partial \tilde{H}}{\partial p_a}(q,p),& a \in {\mathcal{N}}_K. \end{array} \right.$$ where $\tilde{H}(q,p) = H(\xi,\bar\xi)$. Note that the flow $\tau^t = \varphi_\chi^t$ of a real Hamiltonian $\chi$ defines a symplectic map, i.e. satisfies for all time $t$ and all point $z$ where it is defined $$\label{Esympl} (D_z \tau^t)_z^T J (D_z\tau^t)_z = J$$ where $D_z$ denotes the derivative with respect to the initial conditions. The following result is classical: \[Lchange\] Let ${\mathcal{U}}_K$ and ${\mathcal{W}}_K$ be two domains of ${\mathbb{C}}^{{\mathcal{Z}}_K}$, and let $\tau = \varphi_\chi^1 \in \mathcal{C}^{\infty}({\mathcal{U}}_K,{\mathcal{W}}_K)$ be the flow of the real smooth Hamiltonian $\chi$. Then for $H \in \mathcal{C}^{\infty}({\mathcal{W}}_K,{\mathbb{C}})$, we have $$\forall\, z \in {\mathcal{U}}\quad X_{H \circ \tau}(z) = (D_z\tau(z))^{-1}X_H(\tau(z)).$$ Moreover, if $H$ is a real Hamiltonian, $H \circ \tau$ is a real Hamiltonian. Hypothesis ---------- We describe now the hypothesis needed on the Hamiltonian $H$. In the following, we consider an infinite set of frequencies $(\omega_a)_{a \in {\mathcal{N}}}$ satisfying $$\label{Eboundomega} \forall\, a \in {\mathcal{N}}, \quad |\omega_a| \leq C |a|^m$$ for some constants $C > 0$ and $m > 0$. Let ${\mathcal{U}}$ be an open domain of $\ell^2({\mathbb{C}}^{{\mathcal{Z}}})$ containing the origin, and let ${\mathcal{U}}_K = \Pi_K {\mathcal{U}}$ its projection onto ${\mathbb{C}}^{{\mathcal{Z}}_K}$. We consider the collection of Hamiltonian functions $$\label{Edecomp} H^{(K)} = H_0^{(K)} + P^{(K)},\quad K \geq 0,$$ with $$H_0^{(K)} = \sum_{a \in {\mathcal{N}}_K} \omega_a I_a(z)$$ where for all $a\in {\mathcal{N}}_K$, $$\label{Eaction} I_a(z) = \xi_a \eta_a$$ are the [*actions*]{} associated with $a\in {\mathcal{N}}_K$. Note that if $z = (\xi,\bar\xi)$, then $I_a(z) = |\xi_a|^2$. We moreover assume that the functions $P^{(K)} \in \mathcal{C}^{\infty}({\mathcal{U}}_K,{\mathbb{C}})$ are [*real*]{}, of [*order at least 3*]{}, and satisfy the following: For all $\ell > 1$, there exists constants $C(\ell) \geq 0$ and $\beta(\ell)\geq 0$ such that for all $K \geq 1$, $(j_1,\cdots,j_\ell) \in {\mathcal{Z}}_K^\ell$ and $z \in {\mathcal{U}}_K$, the following estimate holds: $$\label{EestP} \left| \frac{\partial P^{(K)}}{\partial z_{j_1} \cdots \partial z_{j_\ell}}(z) \right| \leq C(\ell) K^{\beta(\ell)}.$$ The Hamiltonian system can hence be written $$\label{Eham3} \left\{ \begin{array}{rcll} \dot\xi_a &=& \displaystyle - i \omega_a \xi_a - i \frac{\partial P^{(K)}}{\partial \eta_a}(\xi,\eta) & a \in {\mathcal{N}}_K\\[2ex] \dot\eta_a &=& \displaystyle i \omega_a \eta_a + i \frac{\partial P^{(K)}}{\partial \xi_a}(\xi,\eta)& a \in {\mathcal{N}}_K. \end{array} \right.$$ Denoting by $\varphi_Q^t$ the exact flow of a Hamiltonian flow, splitting methods are based on the approximation $$\varphi_{H^{(K)}}^h \simeq \varphi_{H_0^{(K)}}^h \circ \varphi_{P^{(K)}}^h$$ for a small time step $h > 0$. Note that in this case, the exact flow of $H_0^{(K)}$ is explicit and given by $$\varphi_{H_0^{(K)}}^h(\xi,\eta) = (e^{-i\omega_a h} \xi_a, e^{i \omega_a h} \eta_a)_{a \in {\mathcal{N}}_K}$$ while the calculation of $\varphi_{P^{(K)}}^h$ requires the solution of an ordinary differential equation, whose solution is often given explicitely (see the examples below). The goal of this paper is the study of the long-time behavior of the numerical solution $z^n$ given by for large number $n$ of iterations. Note that no hypothesis is made here concerning the preservation of the $L^2$ norm by the flow of . Non resonance condition {#SSak} ----------------------- In the following, for ${{\boldsymbol{j}}}= (j_1,\ldots,j_r) \in {\mathcal{Z}}_K^r$ with $r \geq 1$, we use the notation $$z_{{\boldsymbol{j}}}= z_{j_1}\cdots z_{j_r}.$$ Moreover, for ${{\boldsymbol{j}}}= (j_1,\ldots,j_r) \in {\mathcal{Z}}_K^r$ with $j_i = (a_i,\delta_i) \in {\mathcal{N}}_K \times\{ \pm 1\}$ for $i = 1,\ldots,r$, we set $$\overline {{\boldsymbol{j}}}= (\overline{j}_1,\ldots,\overline j_r)\quad\mbox{with}\quad \overline{j}_i = (a_i,-\delta_i), \quad i = 1,\ldots,r,$$ and we define $$\Omega({{\boldsymbol{j}}}) = \delta_1\omega_{a_1} + \cdots + \delta_r\omega_{a_r}.$$ We say that ${{\boldsymbol{j}}}\in {\mathcal{Z}}_K^r$ depends only of the action and we write ${{\boldsymbol{j}}}\in {\mathcal{A}}_K^r$ if $r$ is even and if we can write (up to a permutation of the indexes) $$\forall\, i = 1,\ldots r/2,\quad j_{i} = (a_i,1), \quad\mbox{and}\quad j_{i + r/2} = (a_i,-1)$$ for some $a_i \in {\mathcal{N}}_K$. Note that in this situation, $$\begin{array}{rcl} z_{{\boldsymbol{j}}}= z_{j_1}\cdots z_{j_r} &=& \xi_{a_1}\eta_{a_1} \cdots \xi_{a_{r/2}} \eta_{a_{r/2}}\\[2ex] &=& I_{a_1}(z) \cdots I_{a_{r/2}}(z) \end{array}$$ where for all $a \in {\mathcal{N}}_K$, $I_a(z)$ denote the actions associated with $a$ (see ). For odd $r$, ${\mathcal{A}}_r$ is the empty set. We will assume now that the step size $h$ satisfies the following property: \[H1\] For all $r \in {\mathbb{N}}$, there exist constants $\gamma^*$ and $\alpha^*$ such that for all $K \in {\mathbb{N}}^*$, $$\label{nonres2} (j_1,\ldots,j_r) \in {\mathcal{Z}}_K^r \backslash {\mathcal{A}}_K^r\quad \Longrightarrow \quad |1 - e^{ih\Omega({{\boldsymbol{j}}})}| \geq \frac{h \gamma^*}{K^{\alpha^*}}.$$ The following Lemma \[Lnonres\] shows that condition is [*generic*]{} in the sense that it is satisfied for a large set of $h \leq h_0$ (and in particular independently of $K$), provided that the frequencies $\omega_a$ satisfy a non resonance condition that we state now (see [@HLW; @Shan00] for similar statements): \[H2\] For all $r \in {\mathbb{N}}$, there exist constants $\gamma(r)$ and $\alpha(r)$ such that $\forall\, K \in {\mathbb{N}}^*$, $$\label{nonres} (j_1,\ldots,j_r) \in {\mathcal{Z}}_K^r \backslash {\mathcal{A}}_K^r \quad \Longrightarrow |\Omega({{\boldsymbol{j}}})| \geq \frac{ \gamma}{K^{\alpha}}.$$ In the next section, we will check that condition in different concrete cases. \[Lnonres\] Assume that Hypothesis \[H2\] holds, and let $h_0$ and $r$ be given numbers. Let $\gamma$ and $\alpha$ be such that holds and assume that $\gamma^* \leq (2/\pi) \gamma$, $\alpha^* \geq \alpha + m\sigma + r$ with $\sigma> 1$ and $m$ the constant appearing in , then we have $$\mbox{\rm meas}\{\, h < h_0\, | \, h \mbox{ does not satisfy } \eqref{nonres2}\, \} \leq C \frac{\gamma^*}{\gamma} h_0^{1+\sigma}$$ where $C$ depends on $\sigma$ and $r$. As a consequence the set $$Z(h_0) = \{\, h < h_0\, | \, h \mbox{ satisfies Hypothesis } \ref{H1}\, \}$$ is a dense open subset of $(0,h_0)$. The proof of this lemma is given in [@FGP2 Lemma 4.6]. Statement of the result and applications ======================================== Main results ------------ \[Tmain\] Assume that $P^{(K)}$ and $h < h_0$ satisfy the previous hypothesis. Let $r \in {\mathbb{N}}^*$ be fixed. There exist constants $\sigma$, $C$ and $\varepsilon_0$ depending only on $r$, $h_0$ and the constants $\beta(\ell)$ and $C(\ell)$, $\ell = 0,\ldots r$ in , such that the following holds: For all $\varepsilon < \varepsilon_0$ and $K \leq \varepsilon^{-\sigma}$, and for all $z^0$ real such that $${\|z^0\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}} \leq \varepsilon$$ if we define $$\label{Eseuil} z^n = \big(\varphi_{H_0^{(K)}}^h \circ \varphi_{P^{(K)}}^h \big)^n (z^0)$$ then for all $n$, $z^n$ is still real, and moreover $$\label{Eresnorm} {\|z^n\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}} \leq 2 \varepsilon\quad \mbox{for}\quad n \leq \frac{1}{\varepsilon^{r-1}},$$ and $$\label{Eresact} \forall\, a \in {\mathcal{N}}_K, \quad | I_a(z^n) - I_a(z^0)| \leq C\varepsilon^{5/2}\quad \mbox{for}\quad n \leq \frac{1}{\varepsilon^{r-2}}$$ The proof of this result relies on the following Birkhoff normal form result, whose proof is postponed to Section \[SProof\]: \[TNF\] Assume that that $P^{(K)}$ and $h < h_0$ satisfy hypothesis and . Let $r \in {\mathbb{N}}^*$ be fixed. Then there exists constants $\beta$ and $C$ depending on $r$, $h_0$, $\beta(\ell)$ and $C(\ell)$, $\ell = 0,\ldots r$ in and a canonical transformation $\tau_K$ from $B_K(\rho)$ into $B_K(2\rho)$ with $\rho = (CK)^{-\beta}$ satisfying for all $z \in B_K(\rho)$, $$\label{Eesttau} {\|\tau_K(z) - z\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}} \leq (CK)^\beta {\|z\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}}^2 \quad \mbox{and}\quad {\|\tau_K^{-1}(z) - z\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}} \leq (CK)^\beta {\|z\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}}^2,$$ satisfying the following result: For all $z \in B_K(\rho)$, $$\tau_{K}^{-1} \circ \varphi_{H_0^{(k)}}^h \circ \varphi_{P^{(K)}}^h \circ \tau_K(z) = \varphi_{H_0^{(K)}}^h \circ \psi_K (z)$$ where $\psi_K$ satisfies: - $\psi_K(z)$ is real if $z$ is real, - For all $z \in B_K(\rho)$, $$\label{Enorme} {\|\psi_K(z) - z\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}} \leq (CK)^\beta {\|z\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}}^{r}.$$ - For all $z \in B_K(\rho)$, $$\label{Eactions} |I_a(\psi_K(z)) - I_a(z)| \leq (CK)^\beta {\|z\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}}^{r+1}.$$ [Theorem \[Tmain\]]{} First, let us note that as the Hamiltonian functions $H_0^{(K)}$ and $P^{(K)}$ are real Hamiltonians, it is clear that there exist $\xi^n \in {\mathbb{C}}^{\mathcal{N}}$ such that for all $n$, we have $z^n = (\xi^n,\bar\xi^n)$, that is $z^n$ is real. Let $\beta$ given by Theorem \[TNF\] and let $\sigma = 1/(2\beta)$. We have for $K \leq \varepsilon^{-\sigma}$, $$(C K) ^{\beta} \leq C^\beta \varepsilon^{-1/2}.$$ Let $\tau_K$ be defined by Theorem \[TNF\], and let $y^n = \tau_K^{-1}(z^n)$. Using the property of $\tau_K$, we see that $y^n$ is real, i.e. we have $y^n = (\zeta^n,\bar\zeta^n)$ for all $n$. By definition, we have $$\label{star} \forall\, n \geq 0,\quad y^{n+1} = \big(\varphi_{H_0^{(K)}}^h \circ \psi_K\big) (y^n).$$ Using the fact that $K \leq \varepsilon^{-\sigma}$ and , the transformation $\tau_K$ in the previous Theorem satisfies the following: For all $z$ such that ${\|z\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}} \leq 2\varepsilon$, $$\label{Etransfo} {\|\tau_K^{-1}(z) - z\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}} \leq C^{\beta}\varepsilon^{-1/2} {\|z\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}}^2 \leq 4 C^\beta \varepsilon^{3/2} \leq \textstyle\frac14 \varepsilon$$ provided $\varepsilon_0$ is sufficiently small. Hence we have ${\|y^0\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}} = {\|\tau_K^{-1}(z^0)\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}} \leq \frac54 \varepsilon$. Note that we have $\rho = (CK)^{-\beta} \geq C^{-\beta}\varepsilon^{1/2} \geq 2 \varepsilon$ provided that $\varepsilon_0$ is small enough. Using we get that as long as ${\|y^n\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}} \leq 2 \varepsilon$, we have $${\|y^{n+1}\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}} \leq {\|y^n\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}} + (CK)^{\beta} {\|y^n\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}}^{r} \leq {\|y^n\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}} + 2^r C^\beta \varepsilon^{r -1/2}$$ By induction, we thus see that for $$n \leq 2^{-r-1} C^{-\beta} \varepsilon^{3/2 - r}$$ we have ${\|y^{n}\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}} \leq \frac74 \varepsilon \leq 2 \varepsilon$. Assuming that $\varepsilon_0$ is such that $2^{-r-1} C^{-\beta}\varepsilon_0^{1/2} \leq 1$, this shows that for $n \leq \varepsilon^{1-r}$ we have ${\|y^{n}\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}} \leq \frac74 \varepsilon$. Using and an inequality similar to , we conclude that $${\|z^{n}\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}} \leq 2\varepsilon,\quad \mbox{for}\quad n \leq \frac{1}{\varepsilon^{r-1}}$$ which yields to . Now using and the fact that ${\|y^n\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}} \leq 2 \varepsilon$ we see that for $n \leq \varepsilon^{1-r}$ we have $$\forall\, a \in {\mathcal{N}}_K, \quad | I_a(y^{n+1}) - I_a(y^n)| \leq 2^{r+1} C^\beta \varepsilon^{r+1/2}$$ whence $$\forall\, a \in {\mathcal{N}}_K, \quad | I_a(y^{n}) - I_a(y^0)| \leq 2^{r+1} C^\beta n \varepsilon^{r+1/2}$$ Now we have for all $a \in {\mathcal{N}}_K$ $$|ÊI_{a}(y^n) - I_a(z^n)| = \big| |\zeta_a^n|^2 - |\xi_a^n|^2 \big| = \big| |\zeta_a^n| - |\xi_a^n| \big| \times \big| |\zeta_a^n| + |\xi_a^n| \big|,$$ whence $$|ÊI_{a}(y^n) - I_a(z^n)| \leq |\zeta_a^n - \xi_a^n| ({\|y^n\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}} + {\|z^n\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}}) \leq {\|\tau_K(y^{n}) - y^n\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}}({\|y^n\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}} + {\|z^n\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}}).$$ Using we see that for all $n \leq \varepsilon^{1-r}$ and all $a \in {\mathcal{N}}_K$, $${\|\tau_K(y^{n}) - y^n\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}} \leq 4 C^{\beta} \varepsilon^{3/2}.$$ and hence, as ${\|z^n\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}} \leq 2 \varepsilon$, $$|ÊI_{a}(y^n) - I_a(z^n)| \leq 8 C^{\beta} \varepsilon^{5/2}.$$ Using , we thus see that $$\forall\, n \leq \varepsilon^{1-r}\,\quad \forall\, a \in {\mathcal{N}}_K, \quad | I_a(z^{n}) - I_a(z^0)| \leq 2^{r + 4} C^{\beta} ( \varepsilon^{5/2} + n \varepsilon^{r+ 1/2})$$ and this easily gives the result. Examples -------- In this section we present two examples, other examples like the Klein Gordon equation on the sphere (in the spirit of [@BDGS]) or the nonlinear Schrödinger operator with harmonic potential (in the spirit of [@GIP]) could also be considered with these technics. ### Schrödinger equation on the torus We first consider nonlinear Schrödinger equations of the form $$\label{E1} i \partial_t \psi = - \Delta \psi + V \star\psi + \partial_2g(\psi,\bar \psi),\quad x \in {\mathbb{T}}^d$$ where $V\in C^\infty({\mathbb{T}}^d,{\mathbb{R}})$, $g\in C^\infty({\mathcal{U}},{\mathbb{C}})$ where ${\mathcal{U}}$ is a neighborhood of the origin in ${\mathbb{C}}^2$. We assume that $g(u,\bar u) \in {\mathbb{R}}$, and that $g(u,\bar u) = \mathcal{O}(|u|^3)$. The corresponding Hamiltonian functional is given by $$H(\psi,\bar\psi) = \int_{{\mathbb{T}}^d} | \nabla \psi | ^2 + \bar\psi (V \star \psi) + g(\psi,\bar\psi) \, {\mathrm{d}}x.$$ Let $\phi_{a}(x) = e^{i a\cdot x}$, $a \in {\mathbb{Z}}^d$ be the Fourier basis on $L^2({\mathbb{T}}^d)$. With the notation $$\psi = \Big(\frac{1}{2\pi}\Big)^{d/2}\sum_{a\in {\mathbb{Z}}^d} \xi_{a} \phi_{a}(x) \quad \mbox{and}\quad \bar \psi = \Big(\frac{1}{2\pi}\Big)^{d/2}\sum_{a\in {\mathbb{Z}}^d} \eta_{a}\bar \phi_{a}(x)$$ the (abstract) Hamiltonian associated with the equation can be formally written $$\label{E2} H(\xi,\eta) = \sum_{a \in {\mathbb{Z}}^d} \omega_a \xi_a \eta_a + P(\xi,\eta).$$ Here $\omega_{a} =|a|^2 +\hat V_a$ are the eigenvalues of the operator $$\psi\mapsto - \Delta\psi + V \star\psi,$$ and we see that $\omega_{a}$ satisfy with $m = 2$. Moreover, the nonlinearity function $P(\xi,\eta)$ posesses a zero of order $3$ at the origin. In this situation, it can be shown that the Hypothesis \[H1\] is fulfilled for a large set of potential $V$ (see [@BG06] or [@Greb07]). Following [@GL08a], a space discretization of this equation using spectral collocation methods yields a problem of the form with $${\mathcal{N}}_K = [-K,\ldots,K-1]^d$$ and, with $$\label{Eschrinc} u_K = \Big(\frac{1}{2\pi}\Big)^{d/2}\sum_{a\in {\mathcal{N}}_K} \xi_{a} \phi_{a}(x) \quad \mbox{and}\quad v_K = \Big(\frac{1}{2\pi}\Big)^{d/2}\sum_{a\in {\mathcal{N}}_K} \eta_{a}\bar \phi_{a}(x)$$ the nonlinearity reads $$P^{(K)}(\xi,\eta) = \int_{{\mathbb{T}}^d} \mathcal{Q}( g(u_K,v_K) ) {\mathrm{d}}x$$ where, for a function $\psi = (\frac{1}{2\pi})^{d/2}\sum_{a\in {\mathbb{Z}}^d} \psi_{a} \phi_{a}(x)$ $$\mathcal{Q}( \psi ) = \sum_{a \in {\mathcal{N}}_K} \Big(\sum_{b \in {\mathbb{Z}}^d} \psi_{a + 2Kb} \Big)\phi_a(x)$$ is the collocation operator associated with the points $x_a = \frac{\pi}{K}\in {\mathbb{T}}^d$, $a \in {\mathcal{N}}_K$. It is easy to verify that $P^{(K)}$ satisfies for some constants $C(\ell)$ depending on $g$ and $\beta(\ell)$ depending on $g$ and the dimension $d$. Note that starting from a real initial value $u_K^0(x)$ (see ) this system reduces to solving the system of ordinary differential equation $$\forall\, a \in {\mathcal{N}}_a\, \quad i \frac{{\mathrm{d}}}{{\mathrm{d}}t} u_K(x_a,t) = \mathcal{F}_{2K} \Omega \mathcal{F}_{2K}^{-1} u_K(x_a,t) + \partial_2 g (u_K(x_a,t), \overline{u_K(x_a,t)})$$ where $\Omega$ is the matrix $(\omega_a)_{a \in {\mathcal{N}}_K}$ and $\mathcal{F}_{2K}$ the Fourier transform associated with ${\mathcal{N}}_K$. Note that in this case, the numerical solution is easily implemented: The linear part is diagonal and can be solved explicitely in the Fourier space, while the non-linear part is an ordinary differential equation with fixed parameter $x_a$ at each step. If moreover $g(u,\bar u) = G(|u|^2)$ for some real function $G$ then the solution of the nonlinear part is given explicitely by $\varphi_{P^{(K)}}^h(u) = \exp(-2ih G'(|u|^2) u$ using the fact that $|u|^2$ is constant for a fixed point $x_a$. For high dimension $d$, the previous discretization is usually replaced by a discretization on sparse grid, i.e. with $${\mathcal{N}}_K= \{ \, a = (a_1,\ldots,a_d) \in {\mathbb{Z}}^d\; | \; (1 + |a_1|)\cdots(1 +|a_d|) \leq K\, \} \subset {\mathbb{Z}}^d.$$ As explained in [@L08 Chap III.1], methods exist to write the corresponding system under the symplectic form , upon a possible loss in the approximation properties of the exact solution of by the solution of the discretized Hamiltonian $H^{(K)}$. Note that this does not influence the long time results proven here: In some sense we do not impose the nonlinearity $P^{(K)}(z)$ to approximate an exact nonlinearity $P(z)$. We give first a numerical illustration of resonance effects. We consider the equation $$i\partial_t \psi = - \Delta \psi + V \star \psi + \varepsilon^2 |\psi|^2 \psi$$ in the one dimensional torus ${\mathbb{T}}^1$, with initial value $$\psi_0(x) = \frac{2}{2 - \cos(x)}.$$ Note that this problem is equivalent to solving with a small initial value of order $\varepsilon$. We take $\varepsilon = 0.1$, $V$ with Fourier coefficients $\hat V_a = 2/(10 + 2 a^2)$ and $K = 200$ (i.e. $400$ collocation points). In Figure 1, we plot the actions of the numerical solution given by the Lie splitting algorithm in logarithmic scale. In the right we use the resonant stepsize $h = 2\pi/(\omega_7 - \omega_1) \simeq 0.17459\ldots$. In the left we plot the same result but with the non resonant stepsize $h = 0.174$. \[figure1\] In Figure 2, we show the long time almost conservation of the action in the case where $h = 0.1$ (non resonant), and $\varepsilon = 0.1$ and $\varepsilon = 0.01$ after $10^5$ iterations. \[figure2\] ### Wave equation on the circle We consider the wave equation on the circle $$u_{tt} - u_{xx} + m u = g(u), \quad x \in {\mathbb{T}}^1, \quad t \in {\mathbb{R}},$$ where $m$ is a non negative real constant and $g$ a smooth real valued function. Introducing the variable $ v = u_t$, the corresponding Hamiltonian can be written $$H(u,v) =\int_{{\mathbb{T}}} \frac12 (v^2 + u_x^2 + m u^2) + G(u) \, {\mathrm{d}}x,$$ where $G$ is such that $\partial_u G = g$. Let $A := (-\partial_{xx} + m)^{1/2}$, and define the variables $(p,q)$ by $$q := A^{1/2} u, \quad\mbox{and}\quad p = A^{-1/2}v.$$ Then the Hamiltonian can be written $$H = \frac12 \big(\langle Ap,p\rangle_{L^2} + \langle Aq,q\rangle_{L^2} \big) + \int_{{\mathbb{T}}} G(A^{-{1/2}}q) \, {\mathrm{d}}x.$$ Let $\omega_a=\sqrt{|a|^2+m}$, $a \in {\mathbb{N}}=: {\mathcal{N}}$ be the eigenvalues of the operator $A$, and $\phi_a$ the associated eigenfunctions. Plugging the decompositions $$q(x) = \sum_{a \in {\mathbb{N}}} q_a \phi_a(x) \quad \mbox{and} \quad p(x) = \sum_{a\in {\mathbb{N}}} p_a \phi_a(x)$$ into the Hamiltonian functional, we see that it takes the form $$H = \sum_{a \in {\mathbb{N}}}\omega_a \frac{p_a^2 + q_a^2}{2} + P$$ where $P$ is a function of the variables $p_a$ and $q_a$. Using the complex coordinates $$\xi_a = \frac{1}{\sqrt{2}} ( q_a + ip_a) \quad\mbox{and}\quad \eta_a = \frac{1}{\sqrt{2}} ( q_a - ip_a)$$ the Hamiltonian function can be written under the form with a nonlinearity depending on $G$. As in the previous case, it can be shown that the condition is fulfilled for a set of constant $m$ of full measure (see [@BG06; @Bam07]). A collocation discretization on equidistant points of $[0,2\pi]$ yields the same discretization as previously (with $d = 1$). In this situation, the symmetric Strang splitting scheme $$\varphi_{P^{(K)}}^{h/2} \circ \varphi_{H_0^{(K)}}^h \circ \varphi_{P^{(K)}}^{h/2}$$ corresponds to the Deuflhard’s method [@Deuf79]. If moreover we consider the Hamiltonian $$H^{(K)}(z) = H_0^{(K)}(z) + P^{(K)}(\Phi(h \Omega) z)$$ where $\Omega$ is the matrix with elements $\omega_a$, $a \in {\mathcal{N}}_K$, and $\Phi(x)$ a smooth function that is real, bounded and such that $\Phi(0) = 1$, then the splitting schemes associated with this decomposition coincide with the symplectic mollified impulse methods (see [@HLW Chap. XIII] and [@CHL08c]). Proof of the normal form result {#SProof} =============================== The rest of the paper consists in proving Theorem \[TNF\]. In the following, we denote by ${\mathcal{T}}_r$ the set of polynomial of order $r$ on ${\mathbb{C}}^{{\mathcal{Z}}_K}$ (for sake of simplicity, we do note write the dependance in $K$ in the notation ${\mathcal{T}}_r$). If $$Q = \sum_{\ell = 0}^{r} \sum_{{{\boldsymbol{j}}}\in {\mathcal{Z}}_K^\ell} Q_{{\boldsymbol{j}}}z_{{{\boldsymbol{j}}}}$$ is an element of ${\mathcal{T}}_r$, we set $${|Q|\left.\vphantom{T_{j_0}^0}\!\!\right._{{\mathcal{T}}_r}} = \max_{\ell = 0,\ldots,r} \max_{{{\boldsymbol{j}}}\in {\mathcal{Z}}_K^\ell} | Q_{{\boldsymbol{j}}}|.$$ If moreover $Q \in \mathcal{C}([0,1],{\mathcal{T}}_r)$ we set $${\|Q\|\left.\vphantom{T_{j_0}^0}\!\!\right._{{\mathcal{T}}_r}} = \max_{\lambda \in [0,1]} {|Q(\lambda)|\left.\vphantom{T_{j_0}^0}\!\!\right._{{\mathcal{T}}_r}}.$$ Using the assumptions on $P^{(K)}$, we can write a Taylor expansion of $P$ around $0$, $$P^{(K)}(z) = P_r + Q_r = \sum_{\ell = 3}^r \sum_{{{\boldsymbol{j}}}\in {\mathcal{Z}}_K^\ell} P_{{\boldsymbol{j}}}z_{{\boldsymbol{j}}}+ Q_r(z)$$ where $$|P_{{\boldsymbol{j}}}| \leq C K^{\beta_0}$$ where $C$ and $\beta_0$ depend on $\beta(\ell)$ and $C(\ell)$, $\ell = 0,\ldots,r$ in . Notice that $Q_r(z) \in \mathcal{C}^{\infty}({\mathbb{C}}^{{\mathcal{Z}}_K},{\mathbb{C}})$ admits of zero of order $r +1$ and satisfies $${\|X_{Q_r}(z)\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}} \leq C K^{\beta_0} {\|z\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}}^r$$ for $z \in {\mathcal{U}}_K$, provided $\beta_0 = \beta_0(r,d)$ is large enough. Before giving the proof of Theorem \[Tmain\], we give easy results on the flow of non autonomous polynomials Hamiltonian. \[LP1\] Let $k \geq 1$ and let $P(\lambda) \in \mathcal{C}([0,1],{\mathcal{T}}_{k+1})$ be a homogeneous polynomial of order $k+1$ depending on $\lambda \in [0,1]$. Then - There exists a constant $C$ depending on $k$ such that for all $z \in {\mathbb{C}}^{{\mathcal{Z}}_K}$ and all $\lambda \in [0,1]$, we have $$|P(\lambda, z)| \leq CK^{d(k+1)}{\|P\|\left.\vphantom{T_{j_0}^0}\!\!\right._{{\mathcal{T}}_{k+1}}}{\|z\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}}^{k+1}.$$ - There exists a constant $C$ depending on $k$ such that for any $z \in {\mathbb{C}}^{{\mathcal{Z}}_K}$ and all $\lambda \in [0,1]$, $${\|X_{P(\lambda)}(z)\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}} \leq CK^{d(k+1)} {\|P\|\left.\vphantom{T_{j_0}^0}\!\!\right._{{\mathcal{T}}_{k+1}}} {\|z\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}}^{k}.$$ Moreover, Let $k_1$ and $k_2$ two fixed integers. Let $P$ and $Q$ two homogeneous polynomials of degree $k_1+1$ and $k_2+1$ such that $P \in \mathcal{C}([0,1],{\mathcal{T}}_{k_1 + 1})$ and $Q \in \mathcal{C}([0,1],{\mathcal{T}}_{k_2 + 1})$. Then $\{P,Q\} \in \mathcal{C}([0,1],{\mathcal{T}}_{k_1 + k_2})$ and we have $${\|\{P,Q\}\|\left.\vphantom{T_{j_0}^0}\!\!\right._{{\mathcal{T}}_{k_1 + k_1}}} \leq C {\|P\|\left.\vphantom{T_{j_0}^0}\!\!\right._{{\mathcal{T}}_{k_1 + 1}}} {\|Q\|\left.\vphantom{T_{j_0}^0}\!\!\right._{{\mathcal{T}}_{k_2 + 1}}}$$ for some constant $C$ depending on $k_1$ and $k_2$. We have $$|P(\lambda, z)| \leq {\|P\|\left.\vphantom{T_{j_0}^0}\!\!\right._{{\mathcal{T}}_{k+1}}} \sum_{{{\boldsymbol{j}}}\in {\mathcal{Z}}_K^{k+1}} |z_{{\boldsymbol{j}}}|$$ where we have set for ${{\boldsymbol{j}}}= (j_1,\ldots,j_\ell) \in {\mathcal{Z}}_K^\ell$, $$|z_{{\boldsymbol{j}}}| = |z_{j_1}| \cdots | z_{j_\ell}|.$$ Using $|z_j| \leq {\|z\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}}$ we easily obtain [*(i)*]{} using $\sharp {\mathcal{Z}}_K \leq (2K +1)^{d}$. The second statement is proven similarly. The estimate on the Poisson brackets is trivial. \[LP2\] Let $r \geq 3$, $$Q(\lambda,z) = \sum_{\ell = 3}^r \sum_{{{\boldsymbol{j}}}\in {\mathcal{Z}}_K^\ell} Q_{{{\boldsymbol{j}}}}(\lambda) z_{{{\boldsymbol{j}}}}$$ be an element of $\mathcal{C}([0,1],{\mathcal{T}}_{r})$. Let $\varphi_{Q(\lambda)}^\lambda$ be the flow associated with the non autonomous real Hamiltonian $Q(\lambda)$. Then there exist a constant $C_r$ depending on $r$ such that $$\label{Eball} \rho < \mathrm{inf}\big(1/2, C_r K^{-dr}{\|Q\|\left.\vphantom{T_{j_0}^0}\!\!\right._{{\mathcal{T}}_{r}}}^{-1}\big)\quad\Longrightarrow\quad \forall\, \lambda \in [0,1], \quad \varphi_{Q(\lambda)}^\lambda(B_K(\rho)) \subset B_K(2\rho).$$ Moreover, if $F(\lambda) \in \mathcal{C}([0,1], \mathcal{C}^\infty(B_K(\rho),{\mathbb{C}}))$ has a zero of order $r$ at the origin, then $F(\lambda) \circ \varphi_{Q(\lambda)}^\lambda$ has a zero of order $r$ at the origin in $B_K(\rho)$. Let $z^\lambda =\varphi_{Q(\lambda)}^\lambda(z^0)$. Using the estimates of the previous lemma, we have $$\begin{array}{rcl} \displaystyle \frac{{\mathrm{d}}}{{\mathrm{d}}\lambda} {\|z^\lambda\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}}^2 &=& 2 \langle z^\lambda , X_{Q(\lambda)}(z^\lambda) \rangle \\[1ex] &\leq& c_r K^{dr}{\|Q\|\left.\vphantom{T_{j_0}^0}\!\!\right._{{\mathcal{T}}_{r}}}{\|z^\lambda\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}}\Big({\|z^\lambda\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}}^2 + {\|z^\lambda\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}}^{r-1}\Big) \end{array}$$ for some constant $c_r$ depending on $r$. Hence, as long as ${\|z^\lambda\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}} \leq 1$, we have $$\displaystyle\frac{{\mathrm{d}}}{{\mathrm{d}}\lambda} {\|z^\lambda\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}}^2 \leq 2 c_r K^{dr}{\|Q\|\left.\vphantom{T_{j_0}^0}\!\!\right._{{\mathcal{T}}_{r}}}{\|z^\lambda\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}}^3.$$ By a standard comparison argument, we easily get that for $z^0 \in B_K(\rho)$ we have $$\forall\, \lambda \in [0,1],\quad {\|z^\lambda\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}} \leq 2 {\|z^0\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}}.$$ This shows and the rest follows. We give now the general strategy of the proof of the normal form Theorem \[TNF\], showing in particular the need of working with non autonomous Hamiltonians and of considering the non resonance condition . We consider a fixed step size $h$ satisfying . As in this section $K$ will be considered as fixed, we denote shortly $P^{(K)}$ by $P$ and $H_0^{(K)}$ by $H_0$. We consider the propagator $$\varphi_{H_0}^h \circ \varphi_{P}^h = \varphi_{H_0}^h \circ \varphi_{hP}^1.$$ We embed this application into the family of applications $$\varphi_{H_0}^h \circ \varphi_{hP}^\lambda, \quad \lambda \in [0,1].$$ Formally, we would like to find a real Hamiltonian $\chi = \chi(\lambda)$ and a real Hamiltonian under normal form $Z = Z(\lambda)$ and such that $$\label{eq:flots} \forall\, \lambda \in [0,1]\quad \varphi_{H_0}^h \circ \varphi_{hP}^\lambda \circ \varphi_{\chi(\lambda)}^\lambda = \varphi_{\chi(\lambda)}^\lambda \circ\varphi_{H_0}^h \circ \varphi_{hZ(\lambda)}^\lambda.$$ Let $z^0 \in {\mathbb{C}}^{{\mathcal{Z}}_K}$ and $z^\lambda = \varphi_{H_0}^h \circ \varphi_{hP}^\lambda \circ \varphi_{\chi(\lambda)}^\lambda (z^0)$. Deriving the previous equation with respect to $\lambda$ yields $$\begin{gathered} \frac{{\mathrm{d}}z^\lambda}{ {\mathrm{d}}\lambda} = (D_z \varphi_{H_0}^h)_{\varphi_{H_0}^{-h}(z^\lambda)} X_{hP}( \varphi_{H_0}^{-h}(z^\lambda) ) + \\[1ex] (D_z( \varphi_{H_0}^h \circ \varphi_{hP}^\lambda ))_{ \varphi_{hP}^{-\lambda} \circ \varphi_{H_0}^{-h}(z^\lambda)} X_{\chi(\lambda) } (\varphi_{hP}^{-\lambda} \circ \varphi_{H_0}^{-h}(z^\lambda)) . \end{gathered}$$ Using Lemma \[Lchange\] that remains obviously valid for non autonomous Hamiltonian, we thus have $$\frac{{\mathrm{d}}z^\lambda}{ {\mathrm{d}}\lambda} = X_{A(\lambda)}(z^\lambda)$$ where $A(\lambda)$ it the time dependent real Hamiltonian given by $$A(\lambda) = hP \circ \varphi_{H_0}^{-h}+ \chi(\lambda) \circ \varphi_{hP}^{-\lambda} \circ \varphi_{H_0}^{-h}.$$ Using the same calculations for the right-hand side, is formally equivalent to the following equation (up to an integration constant) $$\label{eq:tg} \forall\, \lambda \in [0,1]\quad hP \circ \varphi_{H_0}^{-h} + \chi(\lambda) \circ \varphi_{hP}^{-\lambda} \circ \varphi_{H_0}^{-h} = \chi(\lambda) + hZ(\lambda) \circ \varphi^{-\lambda}_{\chi(\lambda)} \circ \varphi_{H_0}^{-h}.$$ which is equivalent to $$\label{eq:tg1} \forall\, \lambda \in [0,1]\quad \chi(\lambda) \circ \varphi_{H_0}^h - \chi(\lambda) \circ \varphi_{hP}^{-\lambda}= hP - hZ(\lambda) \circ \varphi^{-\lambda}_{\chi(\lambda)} .$$ In the following, we will solve this equation in $\chi(\lambda)$ and $Z(\lambda)$ with a remainder term of order $r+1$ in $z$. So instead of , we will solve the equation $$\label{eq:tg2} \forall\, \lambda \in [0,1]\quad \chi(\lambda) \circ \varphi_{H_0}^h - \chi(\lambda) \circ \varphi_{hP}^{-\lambda}= hP - (h Z(\lambda) + R(\lambda)) \circ \varphi^{-\lambda}_{\chi(\lambda)} .$$ where the unknown are $\chi(\lambda)$, and $Z(\lambda)$ are polynomials of order $r$, with $Z$ under normal form, and where $R(\lambda)$ possesses a zero of order $r+1$ at the origin. In the following, we formally write $$\chi(\lambda) = \sum_{\ell = 3}^r \chi_{[\ell]}(\lambda) := \sum_{\ell = 3}^r\sum_{{{\boldsymbol{j}}}\in {\mathcal{Z}}_K^\ell} \chi_{{\boldsymbol{j}}}(\lambda) z_{{{\boldsymbol{j}}}}$$ and $$Z(\lambda) = \sum_{\ell = 3}^r Z_{[\ell]}(\lambda) := \sum_{\ell = 3}^r\sum_{{{\boldsymbol{j}}}\in {\mathcal{Z}}_K^\ell} Z_{{\boldsymbol{j}}}(\lambda) z_{{{\boldsymbol{j}}}}$$ where here the coefficients $Z_{{\boldsymbol{j}}}(\lambda)$ are unknown and where we denote by $\chi_{[\ell]}(\lambda)$ and $Z_{[\ell]}(\lambda)$ the homogeneous part of degree $\ell$ in the polynomials $\chi(\lambda)$ and $Z(\lambda)$. Identifying the coefficients of degree $\ell \leq r$ in equation , we obtain $$\chi_{[\ell]}(\lambda) \circ \varphi_{H_0}^h - \chi_{[\ell]}(\lambda) = hP_{[\ell]} - h Z_{[\ell]}(\lambda) + h G_{[\ell]}(\lambda;\chi_*,P_*,Z_*) .$$ where $G$ is a real Hamiltonian homogeneous of degree $\ell$ depending on the polynomials $\chi_{[k]}$, $P_{[k]}$ and $Z_{[k]}$ for $k < \ell$. In particular, its coefficients are polynomial of order $\leq \ell$ of the coefficients $\chi_{j}$, $P_{j}$ and $Z_{j}$ for $j \in {\mathcal{Z}}_K^k$, $k < \ell$. Writing down the coefficients, this equation is equivalent to $$\forall\, {{\boldsymbol{j}}}\in {\mathcal{Z}}_K^r \quad ( e^{ih \Omega({{\boldsymbol{j}}})} - 1)\chi_{{{\boldsymbol{j}}}} = hP_{{{\boldsymbol{j}}}} - h Z_{{{\boldsymbol{j}}}} + h G_{{{\boldsymbol{j}}}}$$ and hence we see that the key is to control the small divisors $ e^{ih \Omega({{\boldsymbol{j}}})} - 1$ to solve these equations recursively. Let $\chi(\lambda)$ be an element of $\mathcal{C}([0,1],{\mathcal{T}}_{r})$. Let $\tau(\lambda) := \varphi_{\chi(\lambda)}^\lambda$ be the flow associated with the non autonomous real Hamiltonian $\chi(\lambda)$. Let $g \in \mathcal{C}([0,1],{\mathcal{T}}_r)$, then we can write for all $\sigma_0 \in [0,1]$, $$\begin{gathered} \label{Ereccomm} g(\sigma_0) \circ \tau(\sigma_0) = g(\sigma_0) \\[1ex] +\sum_{k = 0}^{r-1} \int_{0}^{\sigma_0} \cdots \int_{0}^{\sigma_k} \Big( \mathrm{Ad}_{\chi(\sigma_k)}\circ \cdots \circ \mathrm{Ad}_{\chi(\sigma_1)} g(\sigma_0)\Big) {\mathrm{d}}\sigma_1 \cdots {\mathrm{d}}\sigma_k + R(\sigma_0) \end{gathered}$$ where by definition $\mathrm{Ad}_P(Q) = \{Q,P\}$ $$\label{ERlambda} R(\sigma_0) = \int_{0}^{\sigma_0} \cdots \int_{0}^{\sigma_{r}} \Big( \mathrm{Ad}_{\chi(\sigma_{r})}\circ \cdots \circ \mathrm{Ad}_{\chi(\sigma_1)} g(\sigma_0)\Big)\circ { \tau(\sigma_{r})}\, {\mathrm{d}}\sigma_1 \cdots {\mathrm{d}}\sigma_{r}.$$ Each term in the sum in Eqn. belongs (at least) to the space $\mathcal{C}([0,1], {\mathcal{T}}_{kr})$. The term $R(\sigma_0)$ defines an element of $\mathcal{C}([0,1],\mathcal{C}^\infty({\mathbb{C}}^{{\mathcal{Z}}_K},{\mathbb{C}}))$ and has a zero of order at least $r+1$ at the origin. The proof of this lemma is given in [@FGP2]. As mentioned previously, for a given polynomial $\chi\in \mathcal{C}([0,1], {\mathcal{T}}_r)$ with $r \geq 3$, we use the following notation $$\label{Echiell} \chi(\lambda,z) = \sum_{\ell = 3}^r \chi_{[\ell]}(\lambda) = \sum_{\ell = 3}^r \sum_{{{\boldsymbol{j}}}\in {\mathcal{Z}}_K^\ell} \chi_{{{\boldsymbol{j}}}}(\lambda) z_{{{\boldsymbol{j}}}}$$ where $\chi_{[\ell]}(\lambda) \in \mathcal{C}([0,1], {\mathcal{T}}_r)$ is a homogeneous polynomial of degree $\ell$. \[Pcomposition\] Let $\chi(\lambda)$ be an element of $\mathcal{C}([0,1],{\mathcal{T}}_{r})$. Let $\varphi_{\chi(\lambda)}^\lambda$ be the flow associated with the non autonomous real Hamiltonian $\chi(\lambda)$. Let $g \in \mathcal{C}([0,1],{\mathcal{T}}_r)$, then we can write for all $\lambda \in [0,1]$, $$g(\lambda) \circ \varphi^\lambda_{\chi(\lambda)} = S^{(r)}(\lambda) + T^{(r)}(\lambda)$$ where - $S^{(r)}(\lambda) \in \mathcal{C}([0,1],{\mathcal{T}}_r)$. Moreover, if we write $$S(z) = \sum_{\ell = 3}^r S_{[\ell]}(\lambda)$$ where $S_{[\ell]}(\lambda)$ is a homogeneous polynomial of degree $\ell$, then we have for all $\ell = 3,\ldots,r$, $$S_{[\ell]}(\lambda) = g_{[\ell]}(\lambda) + G_{[\ell]}(\lambda;\chi_*,g_*)$$ where $G_{[\ell]}(\lambda;\chi_*,g_*)$ is a homogeneous polynomial depending on $\lambda$ and the coefficients $S_j$ are polynomials of order $ < \ell$ of the coefficients appearing in the decomposition of $g$ and $\chi$. Moreover, we have $$\label{Ebound1} {\|G_{[\ell]}(\lambda;\chi_*,g_*)\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}} \leq \Big( 1 + \sum_{m = 3}^{\ell - 1} {\|g_{[m]}\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}}^\ell\Big)\Big(1 + \sum_{m = 3}^{\ell - 1} {\|\chi_{[m]}\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}}^\ell\Big).$$ - $T^{(r)}(\lambda)\in \mathcal{C}([0,1],\mathcal{C}^{\infty}({\mathbb{C}}^{{\mathcal{Z}}_K}, {\mathbb{C}}))$ has a zero of order at least $r +1$ at the origin and satisfies for all $z \in B_K(1/2)$, $${\|X_{T^{(r)}(\lambda)}(z)\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}}\leq C_r K^{2rd} C_r(\chi_*,g_*) {\|z\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}}^r$$ where $$\label{Ebound2} C_r(\chi_*,g_*) \leq C \Big( 1 + \sum_{m = 3}^{r} {\|g_{[m]}\|\left.\vphantom{T_{j_0}^0}\!\!\right._{{\mathcal{T}}_r}}^r\Big)\Big(1 + \sum_{m = 3}^{r} {\|\chi_{[m]}\|\left.\vphantom{T_{j_0}^0}\!\!\right._{{\mathcal{T}}_r}}^r\Big)$$ with $C$ depending on $r$. Using the previous lemma, we define $S^{(r)}$ as the polynomial part of degree less or equal to $r$ in the expression : this polynomial part may be computed iteratively, from the homogeneity degree 3 to $r$. Actually, every Poisson bracket appearing in is taken with a polynomial $\chi(\sigma_k)$, which decomposes into homogeneous polynomials with degree 3 at least. The terms appearing in the sum in hence have an increasing valuation, and this allows the iterative computation. The remainder terms, together with the term $R(\lambda)$ in , define the term $T^{(r)}$ (which is an element of $\mathcal{C}([0,1],{\mathcal{T}}_{2r})$). The properties of $S^{(r)}(\lambda)$ and $T^{(r)}(\lambda)$ are then easily shown using Lemma \[LP1\]. The next result (Proposition \[PNF\] below) yields the construction of the [*normal form*]{} term $\psi_K$ of Theorem \[TNF\]. A polynomial $Z$ on ${\mathbb{C}}^{{\mathcal{Z}}_K}$ is said to be in normal form if we can write it $$Z = \sum_{\ell = 3}^{r} \sum_{{{\boldsymbol{j}}}\in {\mathcal{A}}_K^\ell} Z_{{\boldsymbol{j}}}z_{{{\boldsymbol{j}}}}.$$ where ${\mathcal{A}}_K^\ell$ is defined in the beginning of Subsection \[SSak\]. \[PNF\] Assume that $H := H^{(K)}$ satisfies with $P := P^{(K)}$ fulfilling and assume that $h \leq h_0$ satisfies the hypothesis . Then there exist - a polynomial $\chi \in \mathcal{C}([0,1], {\mathcal{T}}_r)$ $$\chi(\lambda) = \sum_{\ell = 3}^r \chi_{[\ell]}(\lambda) := \sum_{\ell = 3}^r\sum_{{{\boldsymbol{j}}}\in {\mathcal{Z}}_K^\ell} \chi_{{\boldsymbol{j}}}(\lambda) z_{{{\boldsymbol{j}}}}$$ - a polynomial $Z \in \mathcal{C}([0,1], {\mathcal{T}}_r)$ $$Z(\lambda) = \sum_{\ell = 3}^r Z_{[\ell]}(\lambda) := \sum_{\ell = 3}^r \sum_{{{\boldsymbol{j}}}\in {\mathcal{A}}_K^\ell} Z_j(\lambda) z_{{{\boldsymbol{j}}}}$$ in normal form, - a function $R(\lambda) \in \mathcal{C}([0,1],\mathcal{C}^\infty(B_K(\rho),{\mathbb{C}}))$ with $\rho < c_0 K^{-\beta}$ for some constant $c_0>0$ and $\beta>1$ depending on $r$ and $d$, and having a zero of order at least $r+1$ at the origin such that the following equation holds: $$\label{eq:tg22} \forall\, \lambda \in [0,1]\quad \chi(\lambda) \circ \varphi_{H_0}^h - \chi(\lambda) \circ \varphi_{hP}^{-\lambda}= hP - (h Z(\lambda) + R(\lambda)) \circ \varphi^{-\lambda}_{\chi(\lambda)} .$$ Furthermore there exists a constant $C_0$ depending on $r$ and $d$ such that $${\|\chi\|\left.\vphantom{T_{j_0}^0}\!\!\right._{{\mathcal{T}}_r}} + {\|Z\|\left.\vphantom{T_{j_0}^0}\!\!\right._{{\mathcal{T}}_r}} \leq C_0 K^{\beta}$$ and such that for all $\rho < c_0 K^{-\beta}$ and all $z \in B_K(\rho)$, we have $$\forall\, \lambda \in [0,1],\quad {\|X_{R(\lambda)}(z)\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}} \leq C_0 K^{\beta} {\|z\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}}^r.$$ Identifying the coefficients of degree $\ell \leq r$ in the equation , we get $$\chi_{[\ell]} \circ \varphi_{H_0}^h - \chi_{[\ell]} = hP_{[\ell]} - h Z_{[\ell]} + h G_{[\ell]}(\chi_*,P_*,Z_*) .$$ where $G$ is a real Hamiltonian homogeneous of degree $\ell$ depending on the polynomials $\chi_{[k]}$, $P_{[k]}$ and $Z_{[k]}$ for $k < \ell$. In particular, its coefficients are polynomial of order $\leq \ell$ of the coefficients $\chi_{{{\boldsymbol{j}}}}$, $P_{{{\boldsymbol{j}}}}$ and $Z_{{{\boldsymbol{j}}}}$ for ${{\boldsymbol{j}}}\in {\mathcal{Z}}_K^k$, $k < \ell$ and satisfy bounds like . Writing down the coefficients, this equation is equivalent to $$\forall\, {{\boldsymbol{j}}}\in {\mathcal{I}}_r \quad ( e^{ih \Omega({{\boldsymbol{j}}})} - 1)\chi_{{{\boldsymbol{j}}}} = hP_{{{\boldsymbol{j}}}} - h Z_{{{\boldsymbol{j}}}} + h G_{{{\boldsymbol{j}}}}.$$ We solve this equation by setting $$Z_{{\boldsymbol{j}}}= P_{{\boldsymbol{j}}}+ G_{{\boldsymbol{j}}}\quad \mbox{and}\quad \chi_{{\boldsymbol{j}}}= 0 \quad \mbox{for}\quad {{\boldsymbol{j}}}\in {\mathcal{A}}_K^\ell$$ and $$Z_{{\boldsymbol{j}}}= 0 \quad \mbox{and} \quad \chi_{{{\boldsymbol{j}}}} = \frac{h}{e^{ih \Omega({{\boldsymbol{j}}})} - 1} (P_{{\boldsymbol{j}}}+ G_{{\boldsymbol{j}}}) \quad \mbox{for}\quad {{\boldsymbol{j}}}\notin {\mathcal{A}}_K^\ell.$$ Using and the result of Proposition \[Pcomposition\] we get the claimed bound for some $\beta$ depending on $r$. To define $R$, we simply define it by the equation . By construction and the assumption on $P = P^{(K)}$, and using bounds of the form , it is easy to show that it satisfies the hypothesis. [Theorem \[TNF\]]{} Integrating the equation in $\lambda$, it is clear that the following equation holds: $$\forall\, \lambda \in [0,1]\quad \varphi_{H_0}^h \circ \varphi_{hP}^\lambda \circ \varphi_{\chi(\lambda)}^\lambda = \varphi_{\chi(\lambda)}^\lambda \circ\varphi_{H_0}^h \circ \varphi_{hZ(\lambda) + R(\lambda)}^\lambda.$$ Note that using Proposition \[Pcomposition\] and we show that for $z \in B_K(\rho)$ with $\rho = cK^{-\beta}$ we have $${\|\varphi^\lambda_{\chi(\lambda)}(z) - z\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}} \leq C K^{\beta} {\|z\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}}^2.$$ This implies in particular that $${\|z\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}} \leq {\|\varphi^\lambda_{\chi(\lambda)}(z)\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}} + CK^{-\beta} {\|z\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}}$$ For $K$ sufficiently large, this shows that $\varphi_{\chi(\lambda)}^\lambda$ is invertible and send $B_{K}(\rho)$ to $B_K(2\rho)$. Moreover, we have the estimate, for all $\lambda \in [0,1]$, $${\|\big(\varphi^\lambda_{\chi(\lambda)}\big)^{-1}(z) - z\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}} \leq C K^{\beta} {\|z\|\left.\vphantom{T_{j_0}^0}\!\!\right._{}}^2.$$ We then define $\tau_K = \varphi^1_{\chi(\lambda)}$ and $\psi_K = \varphi^1_{hZ(\lambda) + R(\lambda)}$ and verify that these applications satisfy the condition of the theorem for suitable constant $C$ and $\beta$. [30]{} D. Bambusi, *Birkhoff normal form for some nonlinear [PDE]{}s*, Comm. Math. Physics **234** (2003), 253–283. D. Bambusi, *A Birkhoff normal form theorem for some semilinear pdes*, Hamiltonian Dynamical Systems and Applications, Springer, 2007, pp. 213–247. D. Bambusi, J.-M. Delort, B. Gr[é]{}bert, and J. Szeftel, *Almost global existence for [H]{}amiltonian semilinear [K]{}lein-[G]{}ordon equations with small [C]{}auchy data on [Z]{}oll manifolds*, Comm. Pure Appl. Math. **60** (2007), no. 11, 1665–1690. , [*Birkhoff normal form for PDE’s with tame modulus*]{}. Duke Math. J. 135 no. 3 (2006), 507Ð-567. , [*Long-time analysis of nonlinearly perturbed wave equations via modulated Fourier expansions*]{}, Arch. Ration. Mech. Anal. 187 (2008) 341-368. , [*Normal form and long time analysis of splitting schemes for the linear Schr[ö]{}dinger equation with small potential.*]{} Numerische Mathematik 106, 2 (2007) 223–262 , [*Birkhoff normal form and abstract splitting methods for semi linear Hamiltonian PDEs.*]{} , [*Spectral semi-discretisations of weakly nonlinear wave equations over long times*]{}, Found. Comput. Math. 8 (2008) 319-334. , [*Conservation of energy, momentum and actions in numerical discretizations of nonlinear wave equations*]{}, Numerische Mathematik 110 (2008) 113–143. , [*Nonlinear Schrödinger equations and their spectral discretizations over long times*]{}, Preprint (2008). , [*Splitting integrators for nonlinear Schrödinger equations over long times*]{}, Preprint (2008). , [*Strang splitting for the time-dependent Schrödinger equation on sparse grids*]{}, SIAM J. Numer. Anal. 46 (2007), 103–123. Beno[î]{}t Gr[é]{}bert, *Birkhoff normal form and [H]{}amiltonian [PDE]{}s*, Partial differential equations and applications, Sémin. Congr., vol. 15, Soc. Math. France, Paris, 2007, pp. 1–46. , [*Birkhoff normal form and Hamiltonian PDEs*]{}, Partial differential equations and applications, Sémin. Congr., vol. 15, Soc. Math. France, Paris, 2007, pp. 1–46. , [*Long time behavior for solutions of semilinear Schrödinger equation with harmonic potential and small Cauchy data on ${\mathbb{R}}^d$.*]{} Preprint (2008) , [*A study of extrapolation methods based on multistep schemes without parasitic solutions*]{}. Z. angew. Math. Phys. 30 (1979) 177-189. , [*Geometric Numerical Integration. Structure-Preserving Algorithms for Ordinary Differential Equations*]{}. Second Edition. Springer 2006. , [*From quantum to classical molecular dynamics: reduced models and numerical analysis*]{}. ÊEuropean Math. Soc., 2008. Nonlinearity 13 (2000), 299–308.

ArXiv

[Search for the Standard Model Higgs Boson at CMS]{}\ [**Tommaso Dorigo$^{1}$ (for the CMS collaboration)**]{}\ [*(1) INFN, Italy\ *]{} ========================================================= Abstract {#abstract .unnumbered} -------- The prospects for the search of the Standard Model Higgs boson with the CMS experiment at the LHC are presented. The analyses rely on a full simulation of the detector response and emphasis is put on explicit strategies for the measurement of experimental and background systematics from data. The discovery reach is presented as a function of the Higgs boson mass. A new complete strategy is presented for the early searches and for the control of systematics at very low luminosities of $O(1 fb^{-1})$. Introduction ------------ The Standard Model (SM) requires the existence of a scalar Higgs boson to break electroweak symmetry and provide mass terms to gauge bosons and fermion fields. Indirect constraints from radiative corrections to electroweak observables indicate for the Higgs boson mass the bound $M_H<157\,GeV$ [@lepewwg], at 95% of Confidence Level (CL). Direct constraints from experimental searches at LEP II and Tevatron have determined that $M_H>114.4\,GeV$, with the exclusion of the range $160\,GeV<M_H<170\,GeV$, at 95% CL [@tev]. The SM Higgs boson can be produced in proton-proton collisions at the Large Hadron Collider (LHC) by several different mechanisms (Fig. \[fig:Prod\]). The production by gluon-gluon fusion is the most frequent, with cross sections up to several tens of picobarns; smaller is the production by vector boson fusion (VBF), which however provides a striking signature of two forward quark jets. The two main production mechanisms have been investigated to assess the chances of an early detection of the Higgs boson with the Compact Muon Solenoid (CMS) detector [@cms]. Searches in three of the main signatures of Higgs production and decay are summarized below; these target the whole favourable range of $M_H$ mass values, with the first two extending above $200\,GeV$ and the third one covering the low-mass region of $M_H<135\,GeV$: - $gg\rightarrow H\rightarrow WW^{(*)}$, with the decay of both $W$ bosons to $e\nu$ or $\mu\nu$ pairs; - $gg\rightarrow H\rightarrow ZZ^{(*)}$, with the decay of both $Z$ bosons to $e^+e^-$ or $\mu^+\mu^-$ pairs; - $qq\rightarrow qqH$, with a decay $H\rightarrow\tau\tau$ accompanied by two forward hadronic jets. The $H\rightarrow WW$ search ---------------------------- The search for the $WW^{(*)}$ decay mode at CMS [@WW] employs events containing exactly two opposite-charge leptons ($e$ or $\mu$) with transverse momenta $p_T>10 \,GeV$ and pseudorapidity $|\eta|<2.5$, with at least one of them having $p_T>20\,GeV$. The following additional pre-selection cuts are then applied: a jet veto ($N_{jet}^{E_T>15\,GeV}=0$), large missing energy ($E_T^{miss}>30\,GeV$), and a dilepton mass above resonances ($m_{ll}>12\,GeV$). \[fig:Prod\] Two separate search strategies are studied: a cut-based analysis and a Neural-Network-based analysis (NN). In both cases the selection is optimized using the azimuthal angle between the leptons, an upper cut on the dilepton mass, and requirements on lepton momenta and missing energy. The NN analysis uses additional kinematic variables to separate the signal from the main backgrounds (Fig. \[fig:WW\_NN\]). The analyses include complete techniques to determine background rates with control samples of data. The top-pair background can be sized up with events containing two additional jets, while the SM production of $WW$ pairs can be normalized using data with $m_{ll}>115 \,GeV$. The modified frequentist $CL_s$ method [@tevlhc] is used to convert the number of expected signal and background events into a significance of the observable signal, as a function of the Higgs mass. A first evidence for the Higgs boson is likely achievable in $1 fb^{-1}$ with the $WW$ final state alone in the region of best sensitivity, $155<M_H<180$ GeV. ![Output of the Neural Network for $H \to WW$ signal (black points) and backgrounds (histograms) for a Higgs boson mass of $170 \,GeV$. A further optimized cut on the NN output is used to select the final candidates.[]{data-label="fig:WW_NN"}](histo_tmva_170.ps){height="14pc"} ![Effective cross section as a function of the four-lepton invariant mass after the signal selection, in the mixed $ee\mu\mu$ channel of the $H\to ZZ$ search. Four different signal distributions (for masses of $130,\,150,\,200$ and $250\,GeV$) are compared to residual backgrounds. The latter are essentially due to SM $ZZ$ production (in purple) and $Zbb$ production (in green).[]{data-label="fig:ZZ_xsection"}](2e2mu_Mass_After_Selection.ps){height="13pc"} The $H\rightarrow ZZ$ search ---------------------------- In the $H\rightarrow ZZ^{(*)}$ search [@ZZ] events are selected to contain four charged leptons ($e^+e^-e^+e^-$, $e^+e^-\mu^+\mu^-$, or $\mu^+\mu^-\mu^+\mu^-$), with pair masses $m_{ll}>12 \,GeV$. To remove the $Zb\bar b$ and $t\bar t$ backgrounds further, CMS uses the combined isolation of the two least-isolated leptons, and the significance of their impact parameter with respect to the primary vertex. The reconstructed mass of the dilepton pairs is requested to lay in the windows $[50-100]$ and $[20-100] \,GeV$. ![Values of the log-likelihood ratio discriminant resulting from the combination of the four-lepton final states studied in the search. The dashed black curve shows the expected distribution of the discriminant in the absence of signal as a function of Higgs mass, while the red curve shows the observable value of the ratio in the presence of the Higgs boson.[]{data-label="fig:ZZ_likelihood"}](2lnQ.ps){height="13pc"} After the selection, backgrounds are almost exclusively due to SM $ZZ^{(*)}$ production. The four-body mass provides further discrimination, as shown in Fig. \[fig:ZZ\_xsection\] Mass window cuts allow to obtain a signal-to-noise ratio larger than 1.0 throughout the $120-250 \,GeV$ Higgs mass region considered in the search. The $CL_s$ method allows to estimate the significance of the extractable signal with $1 fb^{-1}$ of $14 TeV$ collision (Fig. \[fig:ZZ\_likelihood\]). A sensitivity at the $2\sigma$ level can be obtained for favourable mass values. The $qqH\rightarrow qq\tau\tau$ search -------------------------------------- A study of the observability of the VBF signature $qqH\rightarrow qq\tau\tau$, for Higgs masses between $115$ and $145 \,GeV$, has been performed [@taus] using events containing one leptonic $\tau$-decay candidate ($\tau\rightarrow e\nu\nu$ or $\tau\rightarrow \mu\nu\nu$), collected by a low-$p_T$ electron or muon trigger. A second $\tau$-lepton candidate is required to produce a narrow $E_T>30 \,GeV$ jet containing one track with $p_T>6\,GeV$ within its core. The two forward jets characteristic of VBF processes are used to reduce backgrounds, mainly coming from QCD multijet production and $Z\rightarrow \tau\tau$ decays. The mass of two forward jets with $E_T>30\,GeV$ has to exceed $400\,GeV$, and they must be separated in pseudorapidity by more than $2.5$ units. Backgrounds amount to $31.8$ events, with $0.6-1.6$ expected from the Higgs signal, depending on $M_H$. The sensitivity of this search channel is found to be insufficient to provide an independent evidence of the SM $H$ boson in early LHC data. Combination of $H\rightarrow ZZ$, $H\rightarrow WW$ searches ------------------------------------------------------------ An additional study was carried out for a combination of the $WW$ and $ZZ$ channels to determine the range of Higgs boson masses that CMS is likely to exclude at 95% C.L. in the absence of a signal, using the results of [@WW] and [@ZZ]. The combination was performed with both the $CL_s$ and a Bayesian method; in general the two methods were found to agree within 10%, which is also a measure of the typical variation in their difference. Fig. \[fig:comb\] shows the limit which can be obtained with luminosity of $1 fb^{-1}$ at $14\,TeV$ together with the result of considering a modified scenario, in which $1 fb^{-1}$ of collisions is produced at the reduced energy of $10\,TeV$. {height="15pc"} \[fig:comb\] Acknowledgements ---------------- The author wishes to thank Eleni Petrakou for her editorial help in the preparation of this paper. [99]{} TEVNHP Working Group, arXiv:hep-ex/0903.4001 (2009). CMS Collaboration, JINST 3, S08004 (2008). CMS Collaboration, CMS PAS HIG-08-006 (2009). U. Aglietti [*et al.*]{}, arXiv:hep-ph/0612172 (2007). CMS Collaboration, CMS PAS HIG-08-003 (2009). CMS Collaboration, CMS PAS HIG-08-008 (2008).

ArXiv

--- abstract: 'AR Scorpii is a close binary system containing a rotation powered white dwarf and a low mass M type companion star. This system shows non-thermal emission extending up to the X-ray energy range. We consider hybrid (lepto-hadronic) and pure hadronic models for the high energy non-thermal processes in this binary system. Relativistic electrons and hadrons are assumed to be accelerated in a strongly magnetised, turbulent region formed in collision of a rotating white dwarf magnetosphere and a magnetosphere/dense atmosphere of the M dwarf star. We propose that the non-thermal X-ray emission is produced either by the primary electrons or the secondary $e^\pm$ pairs from decay of charged pions created in collisions of hadrons with the companion star atmosphere. We show that the accompanying $\gamma$-ray emission from decay of neutral pions, that are produced by these same protons, is expected to be on the detectability level of the present and/or the future satellite and Cherenkov telescopes. The $\gamma$-ray observations of the binary system AR Sco should allow to constrain the efficiency of hadron and electron acceleration and also the details of the radiation processes.' author: - | W. Bednarek\ Department of Astrophysics, The University of Łódź, ul. Pomorska 149/153, 90-236 Łódź, Poland,\ [email protected] date: 'Accepted . Received ; in original form ' title: 'Hadronic model for the non-thermal radiation from the binary system AR Scorpii' --- \[firstpage\] white dwarfs — pulsars: general — X-rays: binaries — radiation mechanisms: non-thermal — gamma-rays: stars Introduction ============ The pulsed emission (from radio up to ultraviolet) has been recently discovered from the binary system AR Sco containing a white dwarf (WD) and a low mass M type star (Marsh et al. 2016, Buckley et al. 2017). This emission varies with the orbital period of the binary system 3.56 hr and also with the rotation period of the white dwarf 1.97 min. Moreover, the emission is accompanied by the non-thermal X-ray emission in the energy range between 0.3-10 keV (Marsh et al. 2016). X-ray emission shows modulation with a pulse fraction of 14$\%$ (Takata et al. 2017). The spin-period seems to slow down indicating that the white dwarf acts like a pulsar with the spin down luminosity $\simeq 3\times 10^{33}$ erg s$^{-1}$. This is clear evidence that not only neutron stars but also magnetized white dwarfs can act under the pulsar mechanism emitting non-thermal radiation (e.g. Paczyński 1990, Usov 1993, Zhang & Gil 2005, Ikhsanov & Biermann 2006, Malheiro et al. 2012, Bednarek 2012). Such WD pulsars are expected to originate in the initial spin-up phase due to the accretion of matter from the companion star (e.g. Beskrovnaya & Ikhsanov 2016). The inferred spin-down luminosity is sufficient to power the observed multi-wavelength emission from the system. The lack of evidence of any mass transfer from the companion star to the white dwarf (e.g. Marsh et al. 2016) suggests that the total spin down reservoir is channelled into non-thermal emission (e.g. Marsh et al. 2016, Buckley et al. 2017), which makes AR Sco unique amongst the close binaries. Most of the emission observed from AR Sco is interpreted in terms of the synchrotron process of relativistic electrons accelerated in the interaction of the white dwarf magnetosphere with the wind from the M dwarf star (Geng et al. 2016, Buckley et al. 2017). In this model, the white dwarf pulsar injects a magnetized plasma which interact with the stellar wind producing a bow shock. On the other hand, Katz (2017) proposes that the spin down power is dissipated in the direct interaction of the WD magnetic field with a dense matter of the M dwarf star atmosphere. Then, particles can be accelerated in a turbulent collision region either in the Fermi acceleration mechanism and/or the magnetic reconnection. In this paper, we assume that electrons and/or protons can be accelerated to TeV energies close to the surface of the M star in the turbulent magnetized collision region as postulated by the models mentioned above. Relativistic protons interact with the matter in the M dwarf atmosphere ($p + p\rightarrow \pi^\pm + \pi^{\rm 0}$), producing secondary $e^\pm$ pairs and high energy $\gamma$-rays via decay of pions. Primary and/or secondary leptons are responsible for the non-thermal synchrotron X-ray emission from AR Sco. The level of this non-thermal emission allows us to predict the $\gamma$-ray flux that is instantaneously produced by protons. Acceleration of particles ========================= 6.2truecm We assume that particles are accelerated in a turbulent region or a shock formed as a result of the interaction between the WD magnetosphere or the plasma from the white dwarf with the atmosphere (magnetosphere) of the companion star (see schematic representation in Fig. 1). In such turbulent magnetized plasma, electrons and hadrons can be accelerated in the second order Fermi acceleration mechanism, on small scale shocks (the first order Fermi mechanism) and possibly also in the magnetic reconnection of the turbulent magnetic field which enter the collision region from the site of the WD and the companion star. We consider the process of acceleration of particles in the turbulent plasma by parametrising their acceleration time scale by, $$\begin{aligned} \tau_{\rm acc} = R_{\rm L}/c\chi\approx 1E_{\rm TeV}/\chi_{-3}B_2~~~{\rm s}, \label{eq1}\end{aligned}$$ where $R_{\rm L}$ is the Larmor radius of the particle, $E_{\rm e,p} = 1 E_{\rm TeV}$ TeV is the particle energy, $\chi = 10^{-3}\chi_{-3}$ is the so called acceleration coefficient, $B = 100B_2$ G is the magnetic field strength in the acceleration region, and $c$ is the speed of light. The value of the acceleration coefficient, $\chi$, is related to the characteristic velocities of the scattering centres in the turbulent plasma. It can be estimated from $\chi\sim 0.1\beta^2\sim 10^{-3}$ (Malkov & Durry 2001), where $\beta = v/c$ is the relative velocity of the scattering centres in respect to the velocity of light $c$. $v$ is assumed to be smaller but of the order of the rotation velocity, $v_{\rm rot}$, of the WD magnetosphere at the location of the companion star. $v_{\rm rot}$ is equal to $2\pi d/P_{\rm WD}\approx 4.1\times 10^9$ cm s$^{-1}$, where the rotation period of the WD is $P_{\rm WD} = 117$ s and the distance between the companion stars is $d\sim 7.6\times 10^{10}$ cm (Marsh et al. 2016). Such a simple description of the acceleration process is frequently used in modelling of the radiation processes in cosmic sources due to still not well known details of the physics of complicated processes related to the injection, acceleration, and interaction of particles with the magnetized medium (e.g. see review article by Malkov & Durry 2001). The acceleration process of electrons is usually saturated by their synchrotron energy losses. The cooling time scale of electrons in the synchrotron process is, $$\begin{aligned} \tau_{\rm syn} = 3m_{\rm e}c^2/4c\sigma_{\rm T}U_{\rm B}\gamma_{\rm e}\approx 2.8\times 10^{-2}/(E_{\rm TeV}B_2^2)~~~{\rm s}, \label{eq2}\end{aligned}$$ where $\gamma_{\rm e} = E_{\rm e}/m_{\rm e}c^2$ is the Lorentz factor of electrons, $U_{\rm B} = B^2/8\pi$ is the energy density of the magnetic field, $\sigma_{\rm T}$ is the Thomson cross section, and $m_{\rm e}$ is the electron rest mass. The magnetic fields of the order of a hundred Gauss are expected to be provided to the turbulent region from the WD magnetosphere (e.g. Katz 2017, Takata et al. 2017). On the other hand, the polar surface magnetic fields of the low mass companion stars can be up to $\sim 3000$ G (e.g. Meintjes & Jurua 2006, Donati & Landstreet 2009). Therefore, these turbulent regions are expected to be strongly magnetized allowing for very efficient cooling of accelerated electrons. By comparing these two time scales, we estimate the characteristic energies of accelerated electrons, $$\begin{aligned} E_{\rm e}^{\rm max}\approx 0.17(\chi_{-3}/B_2)^{1/2}~~~{\rm TeV}. \label{eq3}\end{aligned}$$ Then, the maximum energies of synchrotron photons produced by these electrons are, $$\begin{aligned} \varepsilon_{\rm syn}\approx m_{\rm e}c^2(B/B_{\rm cr})\gamma_{e^\pm}^2\sim 100\chi_{-3}~~~{\rm keV}, \label{eq4}\end{aligned}$$ where $B_{\rm cr} = 2\pi m_{\rm e}^2c^3/he = 4.4\times 10^{13}$ G is the critical magnetic field strength, $m_{\rm e}$ is the electron rest mass, $h$ is the Planck constant, and $e$ is the charge of electron. We have calculated the synchrotron spectrum produced in the process of a complete cooling of electrons with the power law spectrum and an exponential cut-off at $E_{\rm e}^{\rm max}$, i.e. $dN_{\rm e}/dE_{\rm e}\propto E_{\rm e}^{-\alpha}\exp(-E_{\rm e}/E_{\rm e}^{\rm max})$, where $\alpha$ is the spectral index. The [*Swift*]{} X-ray spectrum is reported to have the spectral index close to -2 in the energy range 0.3$-$10 keV (Marsh et al. 2016). More recent analysis of the Swift data reports the spectral index of the pulsed component equal to 2.3$\pm$0.5 (Takata et al. 2017). The synchrotron spectrum with such spectral index is produced by a completely cooled electrons that are injected also with the spectral index close to -2. In fact, electrons with such spectral index are expected to be accelerated in the Fermi process. We also consider the possibility that the non-thermal X-ray emission is produced in the synchrotron process by leptons which are secondary products of the interactions of relativistic protons with the matter in the companion star atmosphere. Protons are accelerated in the mentioned above turbulent region of the WD magnetosphere. The acceleration of protons is limited by their escape from the turbulent region. The escape time scale is estimated as, $$\begin{aligned} \tau_{\rm esc} = 2R_\star/v_{\rm adv}\approx 50/v_{9}~~~{\rm s}, \label{eq5}\end{aligned}$$ where $v_{\rm adv} = 10^9v_{9}$ cm s$^{-1}$ is the characteristic advection velocity assumed to be smaller than the velocity of the magnetic field lines of the WD at the location of the companion star, and $R_\star = 2.5\times 10^{10}$ cm is the radius of the star. By comparing the acceleration time scale of protons with their escape time scale, we obtain the maximum energies of accelerated protons, $$\begin{aligned} E_{\rm p}\approx 50B_2\chi_{-3}/v_{9}~~~{\rm TeV}. \label{eq6}\end{aligned}$$ For the magnetic field strengths mentioned above, i.e. $\sim 100$ G, the acceleration coefficient $\xi = 10^{-3}$ and the advection velocity $v_{\rm adv} = 10^9$ cm s$^{-1}$, protons are expected to be accelerated up to $\sim 100$ TeV. Note that the Larmor radius of protons with such energies in the magnetic field of 100 G, equal to $R_{\rm L} = 3\times 10^9$ cm, is about an order of magnitude smaller than the radius of the companion star. Hadrons with the multi-TeV energies can find enough target in the atmosphere of the dwarf star suffering efficient hadronic collisions with the matter. They are expected to produce $\gamma$-rays and $e^\pm$ pairs from decay of pions ($p + p\rightarrow \pi^{\rm 0} + \pi^\pm$). We assume that the non-thermal X-ray emission, observed from the binary system AR Sco, can be produced by those secondary $e^\pm$ pairs in the synchrotron process. The energies of these $e^\pm$ pairs can be simply estimated from, $E_{e^\pm}\approx \kappa E_{\rm p}/4\mu_{pp}\sim 1$ TeV, $\kappa\approx 0.5$ is the inelasticity in the p-p interactions, $\mu_{pp}\approx 2.57\times \ln(2E_{\rm GeV}) - 6.45$ is the pion multiplicity in p-p interactions, and $E_{\rm GeV}$ is the proton energy in GeV (Orth & Buffington 1976). The $e^\pm$ pairs produce characteristic synchrotron radiation with energies, $$\begin{aligned} \varepsilon_{\rm syn}\approx m_{\rm e}c^2(B/B_{\rm cr})\gamma_{e^\pm}^2\approx 5B_2~~~{\rm MeV}. \label{eq7}\end{aligned}$$ Therefore, hadronic model predicts the synchrotron spectrum from secondary $e^\pm$ pairs extending through the hard X-ray up to the soft $\gamma$-ray energy range. Since the cooling time scale of $e^\pm$ pairs is short (see estimates above), these pairs lose energy already in the turbulent region. However, the fate of the $\gamma$-rays, produced from decay of neutral pions, can be different. A significant part of these $\gamma$-rays is expected to be directed towards the dense atmosphere of the companion star due to the geometrical effects and also due to the collimation of charged protons by the ordered component of the magnetic field of the companion star. Therefore, $\gamma$-rays can be partially absorbed in the inner dense atmosphere of the companion star since the cross section for their absorption in the matter ($\gamma + p\rightarrow e^\pm$) is of the order of the cross section for the pion production in collisions of relativistic protons with the matter (e.g. Lang 1999). Below we calculate the synchrotron spectra produced by the secondary $e^\pm$ pairs and also the $\gamma$-ray spectra from decay of pions. From the fitting of the observed X-ray spectrum by the synchrotron emission from the secondary $e^\pm$ pairs, we determine the power which should be transferred to the relativistic electrons. Based on this normalization, we predict the expected $\gamma$-ray spectrum in the GeV-TeV energy range in terms of such hadronic model. We show that the hadronic model can be tested by the observations with the HESS array, the planned Cherenkov Telescope Array (CTA) and also through the analysis of the available ${\it Fermi}$-LAT data. Non-thermal radiation ===================== The non-thermal X-ray emission from AR Sco indicates on the presence of relativistic electrons in this source. However, it is not clear whether electrons are accelerated directly in the turbulent collision region between the WD magnetosphere and the companion star atmosphere or they are the secondary products from other processes. Therefore, we consider two general scenarios. In the first one, electrons and protons are accelerated with a specific ratio of the powers described by the factor $\eta = L_{\rm e^-}/L_{\rm p}$ and the spectral indexes equal to -2. In this case, we interpret the [*Swift*]{} X-ray observations of AR Sco as due to the synchrotron emission from the primary electrons in the strong magnetic field of the turbulent region (see Fig. 2a). The obtained fit is consistent with the lower limit on the Swift X-ray spectrum. The upper limit will require flatter injection spectrum of particles than -2. In principle, it can be obtained in the acceleration process in the magnetic reconnection regions. However, the fitting of the Swift upper limit on the X-ray flux does not have significant effect on the predicted $\gamma$-ray spectrum from decay of pions since the Swift fit is obtained by the upper energy part of the synchrotron spectrum. The power transferred to relativistic electrons has to be $2.6\times 10^{31}$ erg s$^{-1}$ in order to explain the observed level of the [*Swift*]{} emission and assuming the complete cooling of electrons. This means that the WD should transfer about $1\%$ of its rotation energy to the relativistic electrons. In this scenario we postulate that the acceleration and radiation processes of primary electrons occur at a relatively rare parts of the companion star atmosphere. However, even if it occurs in a quite dense region of the companion star atmosphere, then the synchrotron process of the primary electrons likely overcomes their possible cooling on the bremsstrahlung process since $\tau_{\rm brem} = 1.3\times 10^{15}/n~[s]\gg \tau_{\rm syn}$, where $n$ is the density of matter in $cm^{-3}$ (see Eq. 2 and Lang 1999). Electrons with sub-TeV energies are captured effectively by the strong magnetic field since their Larmor radii are much lower than the radius of the companion star. Therefore, X-rays, propagating along the straight lines, are expected to escape without significant absorption in the matter of the WD atmosphere. However, relativistic protons are also likely to be accelerated in the turbulent region. Their acceleration site should be linked with the dense regions of the M dwarf companion star by its large scale magnetic field. Therefore, protons are expected to be driven towards dense regions of the stellar atmosphere. They can efficiently interact with the matter of the companion star atmosphere producing secondary $e^\pm$ pairs and $\gamma$-rays from decay of pions. We calculate the synchrotron emission from the secondary $e^\pm$ pairs and the accompanying $\gamma$-ray emission assuming that these particles are produced in the pion decay originated in a single interaction of protons (see Fig. 2a) and in the process of complete cooling of protons (Fig. 2b). If more energy is transferred to relativistic protons than to electrons, then the synchrotron spectra (from secondary $e^\pm$ pairs) and the $\gamma$-ray spectra are proportionally enhanced (see the dotted curve for the ratio of the electron to proton powers equal to $\eta = 1/3$ and the dot-dashed curve for $\eta = 1/10$ in Figs. 2a,b). 5.truecm We also consider the non-thermal processes in AR Sco under the hypothesis that the [*Swift*]{} X-ray emission is produced only by the secondary $e^\pm$ pairs, i.e. primary electrons are not accelerated to large energies in the turbulent region (see Fig. 2c). In this case we also show the $\gamma$-ray spectra that might be significantly absorbed in collisions with the matter of the atmosphere in the process $\gamma + p\rightarrow e^\pm$ (see dashed and dotted curves in Fig. 2c) for the case in which only $10\%$ and $30\%$ $\gamma$-rays escape from the companion star atmosphere). It is expected that the column density of the matter traversed by relativistic protons is much larger than that traversed by $\gamma$-rays. In fact, protons will be captured by the magnetic field on a characteristic distance scale corresponding to their Larmor radii, $R_{\rm L} = E/eB\approx 3\times 10^7 E_{\rm TeV}/B_2$ cm. Then, the column density traversed by rectilinearly propagating $\gamma$-rays should be a factor of $R_{\rm L}/a << 1$ lower. Therefore, it is expected that protons can efficiently produce $\gamma$-rays in collisions with the matter of the companion star atmosphere. On the other hand, produced $\gamma$-rays can escape without significant absorption due to the relatively low column density of matter along their straight paths. Due to the effective capturing of relativistic protons, the production of secondary leptons can already occur in a relatively rare matter allowing free escape of produced by them synchrotron X-rays without strong absorption. Note that in the optimistic case, i.e. more energy transferred to relativistic protons than to relativistic electrons ($\eta \ll 1$) and the level of $\gamma$-ray absorption is low, the $\gamma$-ray fluxes are predicted within the sensitivity limits of the present Cherenkov telescopes in the TeV energy range. In fact, the HESS array have the best conditions for the observations of AR Sco since it is located on the southern hemisphere. The spectrum predicted in the optimistic case should be detected by the HESS in 2-3 hours. This $\gamma$-ray emission should be also detected in the GeV energy range by the ${\it Fermi}$-LAT telescope in about a month. However, in the pessimistic case, i.e. the ratio of $\eta\sim 1$ (Fig.2a) or only $10\%$ of $\gamma$-rays escape from the companion atmosphere (Fig. 2c), the $\gamma$-ray emission is close to the 30 hours sensitivity of the planned CTA and also within 8 years of sensitivity of the [*Fermi*]{}-LAT telescope. Therefore, we conclude that the extensive observations of the binary system AR Sco at $\gamma$-ray energies can provide important constraints on the acceleration processes of different types of particles (leptons and hadrons) in the magnetised turbulent plasma formed in the collision of the rotating magnetosphere of the WD with the dense atmosphere of a companion star. In the above calculations we neglected the energy losses of relativistic electrons in the inverse Compton scattering (ICS) of thermal radiation emitted from the surface of the M star and the white dwarf in respect to their energy losses on the synchrotron process. These energy losses depend on the energy densities of the magnetic and radiation fields in the acceleration region of electrons. The basic parameters of these stars has been estimated on: surface temperature of M star $T_\star = 3100$ K and its radius $2.5\times 10^{10}$ cm and the white dwarf $T_\star = 9750$ K and the radius $7\times 10^{8}$ cm (Marsh et al. 2016). Then, the energy density of radiation from the M star is estimated on $\sim$0.7 erg cm$^{-3}$ and from the white dwarf on $\sim 70(R_{\rm WD}/a)^2\sim$0.07 erg cm$^{-3}$. In the case of the WD we take into account that its radiation is diluted due to the large separation of the stars within the binary system $a = 8\times 10^{10}$ cm. On the other hand, the energy density of the magnetic field with the applied strength of 100 Gs is $\sim 440$ erg cm$^{-3}$. We conclude that the IC energy losses of electrons can be safely neglected in respect to their synchrotron energy losses. The IC $\gamma$-ray spectra produced by electrons should be by a factor of $\sim 10^{-3}$ lower than the synchrotron spectra shown in Fig. 2. Conclusion ========== We consider the consequences of the hypothesis that both electrons and protons are being accelerated within the binary system AR Sco. Protons can be accelerated to the multi-TeV energies close to the dense atmosphere of the M type star. Thus, they can find enough target for efficient interaction. As a result, they produce secondary $e^\pm$ pairs and $\gamma$-rays from decay of pions. We interpret the observed non-thermal X-ray emission from this binary in terms of three scenarios. In the first one, both electrons and protons are accelerated with some ratio of the powers $\eta$. The synchrotron emission of electrons is responsible for the observed non-thermal X -ray emission. On the other hand, $\gamma$-rays originate from the decay of neutral pions that are in turn produced by protons in a single interaction with the matter of the stellar atmosphere. The $\gamma$-ray emission can be detected by the [*Fermi*]{}-LAT and future CTA provided that $\eta\sim 1$ (see Fig. 2a). The present Cherenkov telescopes (e.g. the HESS array) could detect this $\gamma$-ray emission provided that $\eta < 1/3$. In the second scenario, the X-ray emission is explained by the synchrotron emission of primary electrons (and/or secondary $e^\pm$ pairs) but protons are transported to dense regions of the M dwarf atmosphere suffering complete cooling in the hadronic collisions. Then, the $\gamma$-ray fluxes are clearly larger (see Fig. 2b). These $\gamma$-rays have a chance to be detected by the HESS array provided that $\eta < 1$ and by the [*Fermi*]{}-LAT for $\eta < 3$, i.e. the power in relativistic electrons is a factor of 3 larger than the power in relativistic protons. In the third scenario, we assumed that only protons are efficiently accelerated. Then, the non-thermal X-ray emission is explained by the synchrotron radiation from the secondary $e^\pm$ pairs. The $\gamma$-ray emission is expected to be clearly above the sensitivity limits of the [*Fermi*]{}-LAT telescope and the HESS array (see Fig. 2c). However, this emission might be additionally influenced by the absorption of $\gamma$-rays in the matter ($\gamma + p\rightarrow e^\pm$). We conclude that even if only a small part of the $\gamma$-rays can escape from the M dwarf atmosphere (e.g. $\mu = 10\%$), then their flux should be within the sensitivity of the [*Fermi*]{}-LAT telescope and the HESS array. A partial absorption of $\gamma$-rays might in principle also affect the constraints obtained above in terms of the first and the second scenarios. Therefore, in principle the limits obtained above should refer to $\eta/\mu$ but not only to $\eta$. The predictions of the GeV-TeV $\gamma$-ray emission from AR Sco, obtained in terms of the hadronic model, look quite encouraging. They could be tested by the present and future Cherenkov telescopes and also by the analysis of the available data collected by the [*Fermi*]{}-LAT telescope. Since AR Sco is the only discovered binary system in which the companion M type star is immersed directly in the inner, rigidly rotating magnetosphere of a compact object (in this case the white dwarf) similar processes are not expected to occur in other type of binary systems with compact objects such as LMXBs (containing accreting neutron stars), black widow/red-back binaries (containing millisecond pulsars) or binary systems of classical radio pulsars with massive companion stars such as PSR 1259-63/LS2883. Some of those compact binaries emit (or are expected to emit) high energy $\gamma$-rays in different geometrical and radiative scenarios. Acknowledgements {#acknowledgements .unnumbered} ================ I would like to thank the Referee for useful comments. This work is supported by the grant through the Polish Narodowe Centrum Nauki No. 2011/01/B/ST9/00411. [99]{} Aleksić, J. et al. (MAGIC Collaboration) 2012, APh 35, 435 Bednarek, W. 2012 J.Phys.G: Nucl.Part.Phys. 39, 065001 Beskrovnaya, N.G., Ikhsanov, N.R. 2016 in Proc. “Stars from collapse to collapse” (Nizhnij Arkhyz, Special Astrophysical Observatory of the Russian Acad. of Sci.), arXiv:1612.07831 Buckley, D.A.H., Meintjes, P.J., Potter, S.B., Marsh, T.R., Gänsicke, B.T. 2017 Nat.Astron. 1, 0029 Donati, J.-F., Landstreet, J.D. 2009, ARAA 47, 333 Funk, S., Hinton, J.A. (CTA Consortium) 2013, APh 43, 348 HESS Colaboration, https://www.mpi-hd.mpg.de/hfm/HESS/pages/home/proposals/sc$\_$sens.pdf Geng, J.-J., Zhang, B., Huang, Y-F. 2016 apJ 831, L10 Ikhsanov, N.R., Biermann, P.L. 2006 A&A 445, 305 Katz, J.I. 2017 ApJ 835, 150 Lang, K.R. Astrophysical Formulae, Astronomy and astrophysics Library, Springer-Verlag, 1999, p.433 Maier, G. et al. (CTA Consortium) 2017, in Proc. 35th ICRC (Busan, Korea), PoS(ICRC2017), 846 Malheiro, M., Rueda, J.A., Ruffini, R. 2012 PASJ, 64, 56 Malkov, M.A., Drury, L. O’C. 2001 Rep.Prog.Phys. 64, 429 Marsh, T.R. et al. 2016 Nature 537, 374 Meintjes, P.J., Jurua, E. 2006 MNRAS 372, 1279 Orth, C. D., Buffington, A. 1976 ApJ 206, 312 Paczyński, B. 1990 ApJL 365, L9 Takata, J., Yang, H., Cheng, K.S., 2017, ApJ, 851, 143 Usov, V.V. 1993, ApJ 410, 761 Zhang, B., Gil, J. 2005 ApJL 631, L143 \[lastpage\]

ArXiv

--- abstract: 'We give a mathematical procedure to obtain the adiabatic approximation for the generalized quantum Rabi Hamiltonian both without and with a quadratic interaction. We consider the Hamiltonian as the energy of a model describing the interaction system of a two-level artificial atom and a one-mode microwave photon in circuit QED. In the case without the quadratic interaction, we show in the adiabatic approximation that whether each bare eigenstate forms a Schrödinger-cat-like entangled state or not depends on whether the energy bias of the atom is zero or non-zero, and then, the effect of the tunnel splitting of the atom is ignored. On the other hand, in the case with the quadratic interaction, we show in the adiabatic approximation that all the physical eigenstates obtained by the (meson) pair theory form individual Schrödinger-cat-like entangled states for every energy bias. We conclude that this fact comes from the effect of the tunnel splitting.' author: - | Masao Hirokawa\ [Institute of Engineering, Hiroshima University, Higashi-Hiroshima, 739-8527, Japan]{} title: 'Schrödinger-Cat-Likeness in Adiabatic Approximation for Generalized Quantum Rabi Model without and with $A^{2}$-Term' --- Introduction {#sec:intro} ============ Quantum electrodynamics (QED) describes the interaction between light and matter. QED is a great success as a quantum field theory (QFT) for electrodynamics, which certifies that QFT is useful and excellent to explain the electromagnetic force caused by the exchanging of photons. The exchanged photon is called a virtual photon [@PRSZ]. QED enables us predict some quantities such as the Lamb shift, the difference in energy between the two energy levels of the two orbitals $2S_{1/2}$ and $2P_{1/2}$, of hydrogen atom with extreme accuracy [@feynman; @BD1; @BD2; @cohen-tannoudji; @kaku; @PS]. The Lamb shift is caused by the fact that the different orbitals interact with the vacuum fluctuations of the radiation field. The vacuum fluctuation is originated from the annihilation and creation of virtual photons; therefore, the Lamb shift results from the fact that the atom is dressed with the cloud of virtual photons. Even the ground state is a non-zero photon state; however, the photons with which it is dressed are virtual and not directly observed (cf. Complement B${}_{\mbox{\tiny III}}$.2 of Ref.[@cohen-tannoudji]). It is well known that the vacuum fluctuation due to Heisenberg’s uncertainty principle brings the generation of virtual particles from the quantum vacuum [@nori12]. The state dressed with the virtual photons is called the bare state. Some method have been considered to derive physical states, which are experimentally observable states, from the bare states [@cohen-tannoudji; @HT]. After the success of QED, some physicists developed the analogy for QED, and applied it to nuclear models. They then had to meet and straggle troubles of the strong interaction. Following Yukawa’s theory [@yukawa], nucleons are connected by a strong force, called nuclear force, and it is made by the fact that nucleons exchange $\pi$-mesons (i.e., pion). Namely, nucleon and $\pi$-meson respectively play individual roles of electron and photon in QED. In the early 1940s, (meson) pair theory were studied by Wentzel [@wentzel1; @wentzel2] to consider the nuclear forces under the strong coupling regime [@BR]. On another note, according to quantum cromodynamics (QCD), quark and gluon in QCD respectively play roles of electron and photon in QED. Hadrons are classified into mesons and baryons consisting of quarks. Thus, the well-known problem that whether we can derive Yukawa’s theory for the nuclear force from QCD arises. The recent technology of circuit QED can make a quantum simulation of cavity QED. Quantum simulation is to simulate a target quantum system by a controllable quantum system [@BN]. In particular, it enables us experimentally to demonstrate the amazingly strong interaction between a two-level artificial atom and a one-mode light on a superconducting circuit: Cavity QED has supplied us with stronger interaction than the standard QED does [@HR-RBH; @rbh01]. Experimental physicists demonstrate the interaction using a two-level atom coupled with a one-mode light in a mirror cavity. The solid-state analogue of the strong interaction in a superconducting system was theoretically proposed [@MSS; @MB], and it has been experimentally demonstrated [@Chiorescu; @Wallraff08; @Wallraff04]. The atom, the light, and the mirror resonator in cavity QED are respectively replaced by an artificial atom, a microwave, and a microwave resonator on a superconducting circuit. The artificial atom is a superconducting LC circuit based on some Josephson junctions. This replaced cavity QED is the so-called circuit QED [@YN; @nori17]. The circuit QED has been intensifying the coupling strength so that its region is beyond the strong coupling regime [@DGS; @ciuti; @Mooij; @Gross; @FD; @yoshihara1]. Yoshihara *et al*. succeeded in demonstrating the deep-strong coupling regime, and experimentally showed how the theory using the quantum Rabi model can well describe a physical set-up of circuit QED [@yoshihara1; @yoshihara2; @yoshihara3]. The set-up consists of a two-level artificial atom interacting to a one-mode photon of a microwave cavity. The notion of the deep-strong coupling regime is proposed in Ref.[@CS], and the strength of that regime is so large that it exceeds the strength of the ultra-strong coupling regime for the atom-photon interaction in circuit QED. Braak gives an analytics solution of the eigenvalue problem for the quantum Rabi model [@braak1]. In Ref.[@AN], meanwhile, Ashhab and Nori give a physical establishment of the adiabatic approximation [@irish] for the bare eigenstates of the quantum Rabi model. The adiabatically approximated eigenstates make the Schrödinger-cat-like entangled states. The adiabatic approximation is very handy to analyze the quantum Rabi model, and thus, the Schrödinger-cat-likeness is beginning to investigate [@fuse] using it. In Ref.[@HMS], we show a mathematical theory so that the adiabatic approximation is actually obtained under the strong-coupling limit in the norm resolvent sense. We are interested in a quantum simulation of some phenomena predicted in nuclear physics on superconducting circuit. In particular, this paper deals with the (meson) pair theory for the generalized quantum Rabi Hamiltonian, which is also called asymmetric quantum Rabi Hamiltonian. It consists of the two-level atom Hamiltonian, the one-mode photon Hamiltonian, and the interaction between the atom and the photon. We give our attention to the non-zero energy bias in the atom Hamiltonian. In the case where the energy bias is equal to zero, the generalized quantum Rabi model is the quantum Rabi model. The energy-bias parameter is easily tunable in experiments of circuit QED with the cutting-edge technology. Thus, we treat it as a tunable parameter. We consider the bare (physical) eigenstates of the generalized quantum Rabi Hamiltonian without (with) the quadratic interaction. The quadratic interaction is often called the $A^{2}$-term. We then show how we can mathematically obtain the adiabatic approximation for the generalized quantum Rabi Hamiltonian both without and with the $A^{2}$-term. Based on this mathematical theory, in the case without the $A^{2}$-term, we show that whether the adiabatically approximated bare eigenstates are formed as the Schrödinger-cat-like entangled states or not depends on whether the energy bias is zero or non-zero. As its result, we point out that the effect of the tunnel splitting of the two-level atom disappears. On the other hand, in the case with the $A^{2}$-term, we renormalize it using (meson) pair theory [@wentzel1; @wentzel2; @BR], and show that all the adiabatically approximated physical eigenstates are formed as the Schrödinger-cat-like entangled states for every energy bias. We realize that this fact results from the effect of the tunnel splitting of the two-level atom. Mathematical Set-Ups ==================== In this section, we prepare and recall some mathematical notations and notions to explain and consider our problem. For their details, see Refs.[@RS1; @weidmann] for instance. For a separable Hilbert space $\mathfrak{H}$ we denote its inner product by $(\,\,\, , \,\,\,)_{\mathfrak{H}}$. The norm $\|\,\,\, \|_{\mathfrak{H}}$ is naturally introduced by $\|\psi\|_{\mathfrak{H}} =\sqrt{(\psi , \psi)_{\mathfrak{H}}}$ for every vector $\psi$ in the Hilbert space $\mathfrak{H}$. An operator $A$ acting in the Hilbert space is the linear map from a linear subspace $D(A)\subset\mathfrak{H}$ to the Hilbert space $\mathfrak{H}$. The subspace $D(A)$ is called the *domain* of the operator $A$. In particular, when the operator $A$ satisfies that there is a positive constant $M$ so that the inequality, $\|A\psi\|_{\mathfrak{H}} \le M\|\psi\|_{\mathfrak{H}}$, can hold for any vector $\psi\in \mathfrak{H}$, we say the operator $A$ is *bounded*. Then, for every bounded operator $A$, the operator norm is given by $\| A\|_{\mathrm{op}}:= \sup_{\psi\in D(A)\,\mathrm{with}\,\psi\ne 0} \| A\psi\|_{\mathfrak{H}}/\|\psi\|_{\mathfrak{H}}$. The inequality, $\| A\psi\|_{\mathfrak{H}} \le \|A\|_{\mathrm{op}}\|\psi\|_{\mathfrak{H}}$, holds then. On the other hand, in the case $\| A\|_{\mathrm{op}}=\infty$, we say the operator $A$ is *unbounded*. In quantum theory, an observable $A$ corresponds to a self-adjoint operator, that is, it satisfies the domain identity, $D(A)=D(A^{*})$, and the action identity, $A\psi=A^{*}\psi$ for every vector $\psi\in D(A)$, where $A^{*}$ is the adjoint operator of the operator $A$. It is convenient to consider the resolvent $(A-iz)^{-1}$ for an unbounded self-adjoint operator $A$ for every complex number $z$ with $\Im z\ne 0$. Let $A_{n}$ be a sequence of self-adjoint operators. When there is a self-adjoint operator $A$ so that the limit, $\lim_{n\to\infty} \|(A_{n}-iz)^{-1}-(A-iz)^{-1}\|_{\mathrm{op}}=0$, holds for every complex number $z$ with $\Im z\ne 0$, the operators $A_{n}$ are said to converge to the operator $A$ in the *norm resolvent sense* [@RS1], and we often denote the convergence by $A_{n} \stackrel{\mathrm{n.r.s.}\,\,}{{-\!\!-\!\!\!}\longrightarrow} A \,\,\, \mbox{as $n\to\infty$} $ in this paper. We sometimes represent by $|E\rangle$ a vector in the state space $\mathfrak{H}$. We often use Dirac’s bra-ket notation $\langle E_{1}|E_{2}\rangle$ for the inner product $(|E_{2}\rangle , |E_{2}\rangle)_{\mathfrak{H}}$ of vectors $|E_{1}\rangle$ and $|E_{2}\rangle$, i.e., $\langle E_{1}|E_{2}\rangle:= (|E_{2}\rangle , |E_{2}\rangle)_{\mathfrak{H}}$. So, the notation $\langle E_{1}|A|E_{2}\rangle$ stands for the inner product $(|E_{2}\rangle , A|E_{2}\rangle)_{\mathfrak{H}}$ for vectors $|E_{1}\rangle$, $|E_{2}\rangle$ and an operator $A$, i.e., $\langle E_{1}|A|E_{2}\rangle:= (|E_{2}\rangle , A|E_{2}\rangle)_{\mathfrak{H}}$. Let $A_{j}$ be an operator acting in Hilbert spaces $\mathfrak{H}$, $j=1, 2$. We often omit the tensor sign $\otimes$ from the tensor product $A_{1}\otimes A_{2}$, and denote the tensor product $A_{1}\otimes A_{2}$ by $A_{1}A_{2}$, i.e., $A_{1}A_{2}=A_{1}\otimes A_{2}$. We simply write $A_{1}\otimes I_{\mathfrak{H}_{2}}$ as $A_{1}$ for the identity operator $I_{\mathfrak{H}_{2}}$, and $I_{\mathfrak{H}_{1}}\otimes A_{2}$ as $A_{2}$ for the identity matrix $I_{\mathfrak{H}_{1}}$. Correspondingly, we also omit the tensor symbol $\otimes$ form the tensor product of vectors in $\mathfrak{H}_{1} \otimes\mathfrak{H}_{2}$. The state space of the two-level atom system coupled with one-mode light is given by $\mathbb{C}^{2}\otimes L^{2}(\mathbb{R})$, where $\mathbb{C}^{2}$ is the $2$-dimensional unitary space, and $L^{2}(\mathbb{R})$ the Hilbert space consisting of the square-integrable functions. We sometimes omit the tensor sign $\otimes$ from vectors in the state space $\mathbb{C}^{2}\otimes L^{2}(\mathbb{R})$. We use the notation $|\!\!\uparrow\rangle$ for the up-spin state and the notation $|\!\!\downarrow\rangle$ for the down-spin state, which are defined by $ |\!\!\uparrow\rangle := {\scriptsize \left(\hspace*{-1.5mm} \begin{array}{cc} 1 \\ 0 \end{array} \hspace*{-1.5mm} \right) }$ and $ |\!\!\downarrow\rangle := {\scriptsize \left(\hspace*{-1.5mm} \begin{array}{cc} 0 \\ 1 \end{array} \hspace*{-1.5mm} \right) }$ in $\mathbb{C}^{2}$. We use standard notations for the Pauli matrices, $\sigma_{x}:= {\scriptsize \left(\hspace*{-1.5mm} \begin{array}{cc} 0 & 1 \\ 1 & 0 \end{array} \hspace*{-1.5mm}\right) }$, $\sigma_{y}:= {\scriptsize \left(\hspace*{-1.5mm} \begin{array}{cc} 0 & -i \\ i & 0 \end{array} \hspace*{-1.5mm}\right) }$, and $\sigma_{z}:= {\scriptsize \left(\hspace*{-1.5mm} \begin{array}{cc} 1 & 0 \\ 0 & -1 \end{array} \hspace*{-1.5mm} \right) }$. We denote by $|n\rangle$ the Fock state in $L^{2}(\mathbb{R})$ with the photon number $n=0, 1, 2, \cdots$. That is, $|0\rangle:=(w^{2}/\pi)^{1/4} \exp\left[-(wx)^{2}/2\right]$ and $|n\rangle:=\sqrt{w}\gamma_{n}H_{n}(wx)\exp\left[-(wx)^{2}/2\right] \in L^{2}(\mathbb{R})$, where $H_{n}(x)$ is the Hermite polynomial of variable $x$, $\gamma_{n}=\pi^{-1/4}(2^{n}n!)^{-1/2}$, and $w=\sqrt{m\omega/\hbar}$ for the frequency $\omega$ of a one-mode photon. We omit the tensor sign, $\otimes$, from the tensor product, $|s\rangle\otimes |n\rangle$ with $s=\uparrow, \downarrow$, $n=0, 1, 2, \cdots$, and use a compact notation, $|s\rangle|n\rangle$, for the tensor product. We respectively denote by $a$ and $a^{\dagger}$ the annihilation and creation operators of one-mode photon defined by $a|0\rangle:=0$, $a|n\rangle:=\sqrt{n}|n-1\rangle$, and $a^{\dagger}|n\rangle:=\sqrt{n+1}|n+1\rangle$. The spin-annihilation operator $\sigma_{-}$ and the spin-creation operator $\sigma_{+}$ are defined by $\sigma_{\pm}:=\frac{1}{2}\left(\sigma_{x}\pm i\sigma_{y}\right)$. The identity $2\times 2$ matrix $\sigma_{0}$ is given by $\sigma_{0}=\sigma_{+}\sigma_{-}+\sigma_{-}\sigma_{+}$. Some Reviews and Our Problem {#sec:problem} ============================ We introduce the parameters $\omega$ and $g$, respectively, playing roles of a frequency of a one-mode photon in a cavity and a coupling strength between an artificial two-level atom and the photon in the cavity. For every frequency $\omega$ and coupling strength $g$, the Hamiltonian of the generalized quantum Rabi model reads $$H_{\mbox{\tiny GQR}}(\omega,g) :=H_{\mathrm{atm}}(\varepsilon) +H_{\mathrm{ptn}}(\omega) +\hbar g\sigma_{x}\left(a+a^{\dagger}\right) \label{eq:quantum-Rabi-Hamiltonian_0}$$ with the two-level atom Hamiltonian $H_{\mathrm{atm}}(\varepsilon)$ and the one-mode photon Hamiltonian $H_{\mathrm{ptn}}(\omega)$ defined by $$H_{\mathrm{atm}}(\varepsilon) :=\frac{\hbar}{2}\left(\omega_{\mathrm{a}}\sigma_{z} -\varepsilon\sigma_{x}\right)\quad \mbox{and}\quad H_{\mathrm{ptn}}(\omega) := \hbar\omega\left(a^{\dagger}a+\frac{1}{2}\right),$$ where $\hbar\omega_{\mathrm{a}}$ and $\hbar\varepsilon$ are respectively the tunnel splitting and energy bias between the states, $|\!\!\uparrow\rangle$ and $|\!\!\downarrow\rangle$, of the two-level atom. We recall the expression of the photon annihilation operator $a$ and creation operator $a^{\dagger}$ using the position operator $x$ and momentum operator $p$: $$a=\sqrt{\frac{\omega}{2\hbar}}\, x +i\sqrt{\frac{1}{2\hbar\omega}}\, p \quad\textrm{and}\quad a^{\dagger}=\sqrt{\frac{\omega}{2\hbar}}\, x -i\sqrt{\frac{1}{2\hbar\omega}}\, p. \label{eq:expression_xp}$$ Then, we have another expression of the photon Hamiltonian $H_{\mathrm{ptn}}(\omega)$ as $$H_{\mathrm{ptn}}(\omega) =\frac{1}{2}p^{2}+\frac{\omega^{2}}{2}x^{2} \label{eq:another-representation-H_ptn}$$ using the canonical commutation relation $[x,p]=i\hbar$. In the case where the energy bias is zero (i.e., $\varepsilon=0$), the generalized quantum Rabi Hamiltonian $H_{\mbox{\tiny GQR}}(\omega,g)$ becomes the quantum Rabi Hamiltonian. We denote it by $H_{\mbox{\tiny QR}}(\omega,g)$. The quantum Rabi Hamiltonian $H_{\mbox{\tiny QR}}(\omega,g)$ has the parity symmetry, $$\left[ H_{\mbox{\tiny QR}}(\omega,g)\,\,\, ,\,\,\, \Pi\right]=0, \label{eq:parity-symmetry}$$ for the parity operator $\Pi=(-1)^{a^{\dagger}a}\sigma_{z}$. To introduce the form of the generalized quantum Rabi Hamiltonian $\mathcal{H}_{\mbox{\tiny GQR}}(\omega_{\mathrm{c}},\mathrm{g})$ that we consider in this paper, we define a unitary matrix $U_{xz}$ by $$U_{xz}:=\frac{1}{\sqrt{2}} \left(\begin{array}{cc} 1 & 1 \\ -1 & 1 \end{array}\right). \label{eq:unitary-matrix_U_xz}$$ The Hamiltonian $\mathcal{H}_{\mbox{\tiny GQR}}(\omega,g)$ is given by $$\begin{aligned} \mathcal{H}_{\mbox{\tiny GQR}}(\omega,g) &:=& U_{xz}H_{\mbox{\tiny GQR}}(\omega,g)U_{xz}^{*} \nonumber \\ &=& \mathcal{H}_{\mathrm{atm}}(\varepsilon) +H_{\mathrm{ptn}}(\omega) +\hbar g\sigma_{z}\left( a+a^{\dagger}\right) \label{eq:quantum-Rabi-Hamiltonian}\end{aligned}$$ with the atom Hamiltonian $\mathcal{H}_{\mathrm{atm}}(\varepsilon)$ given by $$\mathcal{H}_{\mathrm{atm}}(\varepsilon) :=-\frac{\hbar}{2}\left(\omega_{\mathrm{a}}\sigma_{x} +\varepsilon\sigma_{z}\right).$$ In this paper, we employ the one-mode photon frequency $\omega_{\mathrm{c}}$ and the coupling strength $\mathrm{g}$ as parameters $\omega$ and $g$, respectively, and we also call $\mathcal{H}_{\mbox{\tiny GQR}}(\omega_{\mathrm{c}},\mathrm{g}) :=U_{xz}H_{\mbox{\tiny GQR}}(\omega_{\mathrm{c}},\mathrm{g})U_{xz}^{*}$ the *generalized quantum Rabi Hamiltonian*, and $\mathcal{H}_{\mbox{\tiny QR}}(\omega_{\mathrm{c}},\mathrm{g}) :=U_{xz}H_{\mbox{\tiny QR}}(\omega_{\mathrm{c}},\mathrm{g})U_{xz}^{*}$ the *quantum Rabi Hamiltonian*. Using Eq.(\[eq:parity-symmetry\]), we have the parity symmetry in the case $\varepsilon=0$, $$0=U_{xz}\left[ H_{\mbox{\tiny QR}}(\omega_{\mathrm{c}},\mathrm{g})\,\,\, ,\,\,\, \Pi\right]U_{xz}= \left[\mathcal{H}_{\mbox{\tiny QR}}(\omega_{\mathrm{c}},\mathrm{g})\,\,\, ,\,\,\, \mathcal{P}\right], \label{eq:parity-symmetry'}$$ where $\mathcal{P}=-(-1)^{a^{\dagger}a}\sigma_{x}$. This determines the form of the eigenstates of the quantum Rabi Hamiltonian $\mathcal{H}_{\mbox{\tiny QR}}(\omega_{\mathrm{c}},\mathrm{g})$ as $$|\!\!\uparrow\rangle \biggl(|\mathrm{even}\rangle+|\mathrm{odd}\rangle\biggr) \pm |\!\!\downarrow\rangle \biggl(|\mathrm{even}\rangle-|\mathrm{odd}\rangle\biggr) \label{eq:form-eigenstates}$$ for proper states $|\mathrm{even}\rangle$ and $|\mathrm{odd}\rangle$ with the individual forms, $$|\mathrm{even}\rangle= \sum_{n:\mathrm{even}}c_{n}^{\mathrm{even}}|n\rangle\quad \mbox{and}\quad |\mathrm{odd}\rangle= \sum_{n:\mathrm{odd}}c_{n}^{\mathrm{odd}}|n\rangle.$$ Physically based on the argument in Ref.[@AN], the Schrödinger-cat-likeness appears in the *adiabatic approximation* for the quantum Rabi Hamiltonian $\mathcal{H}_{\mbox{\tiny QR}}(\omega_{\mathrm{g}},\mathrm{g})$ (i.e., $\varepsilon=0$): all the eigenstates of the quantum Rabi Hamiltonian can be approximated by the Schrödinger-cat-like states, $$\left\{ \begin{array}{ll} {\displaystyle \frac{1}{\sqrt{2}} \Biggl(|\!\!\uparrow\rangle D\left(-\mathrm{g}/\omega_{\mathrm{c}}\right) |n\rangle +|\!\!\downarrow\rangle D\left(\mathrm{g}/\omega_{\mathrm{c}}\right) |n\rangle \Biggr)}, \\ \qquad \\ {\displaystyle \frac{1}{\sqrt{2}} \Biggl( |\!\!\uparrow\rangle D\left(-\mathrm{g}/\omega_{\mathrm{c}}\right) |n\rangle -|\!\!\downarrow\rangle D\left(\mathrm{g}/\omega_{\mathrm{c}}\right) |n\rangle \Biggr)}, \end{array} \right. \label{eq:approximation_AN}$$ in the deep-strong coupling regime. Here, $D(\mathrm{g}/\omega_{\mathrm{c}})$ is the displacement operator defined by $D(\mathrm{g}/\omega_{\mathrm{c}}) :=\exp\left[\mathrm{g}\left(a^{\dagger}-a\right)/\omega_{\mathrm{c}}\right]$. Eqs.(\[eq:approximation\_AN\]) are well known as the adiabatic approximation (e.g., see Eq.(5) of Ref.[@irish]). The eigenenergies of the both eigenstates in Eq.(\[eq:approximation\_AN\]) are almost $\hbar\omega_{\mathrm{c}}(n+1/2)-\hbar\mathrm{g}^{2}/\omega_{\mathrm{c}}$; but, every true eigenstates are non-degenerate besides some cases. For instance, the adiabatically approximated eigenstates with the lowest energy $\hbar\omega_{\mathrm{c}}/2-\hbar\mathrm{g}^{2}/\omega_{\mathrm{c}}$ apparently seem to be degenerate; however, the ground state of the quantum Rabi Hamiltonian actually is unique for every coupling strength $\mathrm{g}$ [@HH]. As shown in Ref.[@HMS; @hirokawa15], the adiabatic approximation given by Eq.(\[eq:approximation\_AN\]) is mathematically justified in the following: We define the unitary operator $U(\mathrm{g}/\omega_{\mathrm{c}})$ by $U(\mathrm{g}/\omega_{\mathrm{c}}) :=\sigma_{+}\sigma_{-}D(\mathrm{g}/\omega_{\mathrm{c}}) +\sigma_{-}\sigma_{+}D(-\mathrm{g}/\omega_{\mathrm{c}})$. We note the equation, $U(\mathrm{g}/\omega_{\mathrm{c}})^{*} =U(-\mathrm{g}/\omega_{\mathrm{c}})$. Then, we obtain the unitary transformation, $$\begin{aligned} &{}& U(\mathrm{g}/\omega_{\mathrm{c}}) \left(\mathcal{H}_{\mbox{\tiny GQR}}(\omega_{\mathrm{c}},\mathrm{g}) +\hbar\mathrm{g}^{2}/\omega_{\mathrm{c}}\right) U(\mathrm{g}/\omega_{\mathrm{c}})^{*} \nonumber \\ &=&H_{\mathrm{ptn}}(\omega_{\mathrm{c}}) -\frac{\hbar}{2}\varepsilon\sigma_{z} -\frac{\hbar}{2}\omega_{\mathrm{a}} \left\{ \sigma_{+}D\left(\mathrm{g}/\omega_{\mathrm{c}}\right)^{2} +\sigma_{-}D\left(-\mathrm{g}/\omega_{\mathrm{c}}\right)^{2} \right\}, \label{eq:unitary-trans-GQR-Hamiltonian}\end{aligned}$$ where $-\hbar\mathrm{g}^{2}/\omega_{\mathrm{c}}$ is the self-energy. Taking the strong coupling limit $\mathrm{g}\to\infty$, we have the limits, $\lim_{\mathrm{g}\to\infty}\hbar\mathrm{g}^{2}/\omega_{\mathrm{c}}=\infty$ and $\lim_{\mathrm{g}\to\infty}\mathrm{g}/\omega_{\mathrm{c}}=\infty$. Thus, the energy $\hbar\mathrm{g}^{2}/\omega_{\mathrm{c}}$ plays a role of a counter term for mass renormalization (i.e., for the bare-photon divergence) in the strong coupling limit. The displacement operators $D\left(\pm\mathrm{g}/\omega_{\mathrm{c}}\right)$ decay to the zero operator in a mathematically proper sense [@HMS; @hirokawa15] in the strong-coupling limit. Developing this fact and using Theorem VIII.19(a) of Ref.[@RS1], we can prove that the unitarily transformed Hamiltonian $U(\mathrm{g}/\omega_{\mathrm{c}}) \left(\mathcal{H}_{\mbox{\tiny GQR}}(\omega_{\mathrm{c}},\mathrm{g}) +\hbar\mathrm{g}^{2}/\omega_{\mathrm{c}} \right)U(\mathrm{g}/\omega_{\mathrm{c}})^{*}$ converges to the Hamiltonian $H_{\mathrm{ptn}}(\omega_{\mathrm{c}})-\hbar\varepsilon\sigma_{z}/2$ in the norm resolvent sense: $$U(\mathrm{g}/\omega_{\mathrm{c}}) \left(\mathcal{H}_{\mbox{\tiny GQR}}(\omega_{\mathrm{c}},\mathrm{g}) +\hbar\mathrm{g}^{2}/\omega_{\mathrm{c}} \right)U(\mathrm{g}/\omega_{\mathrm{c}})^{*} \stackrel{\mathrm{n.r.s.}\,\,}{{-\!\!-\!\!\!}\longrightarrow} H_{\mathrm{ptn}}(\omega_{\mathrm{c}})-\frac{\hbar}{2}\varepsilon\sigma_{z} \,\,\, \mbox{as $\mathrm{g}\to\infty$}.$$ Thanks to Theorem VIII.23(b) of Ref.[@RS1], each eigenstate of the Hamiltonian $\mathcal{H}_{\mbox{\tiny QR}}(\omega_{\mathrm{c}},\mathrm{g})$ is well approximated by that of the Hamiltonian $U(\mathrm{g}/\omega_{\mathrm{c}})^{*} (H_{\mathrm{ptn}}(\omega_{\mathrm{c}}) -\hbar\varepsilon\sigma_{z}/2) U(\mathrm{g}/\omega_{\mathrm{c}}) -\hbar\mathrm{g}^{2}/\omega_{\mathrm{c}}$. This mathematical procedure with $\varepsilon=0$ secures the adiabatic-approximation formulas given by Eq.(\[eq:approximation\_AN\]). On the other hand, in the case where $\varepsilon\ne 0$, all the eigenstates of the generalized quantum Rabi Hamiltonian $\mathcal{H}_{\mbox{\tiny GQR}}(\omega_{\mathrm{c}},\mathrm{g})$ is well approximated by the states, $$\left\{ \begin{array}{ll} {\displaystyle |\!\!\uparrow\rangle D\left(-\mathrm{g}/\omega_{\mathrm{c}}\right) |n\rangle}, \\ \qquad \\ {\displaystyle |\!\!\downarrow\rangle D\left(\mathrm{g}/\omega_{\mathrm{c}}\right) |n\rangle }. \end{array} \right. \label{eq:approximation_0}$$ The first adiabatically approximated eigenstates $|\!\!\uparrow\rangle D\left(-\mathrm{g}/\omega_{\mathrm{c}}\right) |n\rangle$ gives the eigenenergy $\hbar\omega_{\mathrm{c}}\left(n+1/2\right)-\hbar\varepsilon/2 -\hbar\mathrm{g}^{2}/\omega_{\mathrm{c}}$, and the second one gives the eigenenergy $\hbar\omega_{\mathrm{c}}\left(n+1/2\right)+\hbar\varepsilon/2 -\hbar\mathrm{g}^{2}/\omega_{\mathrm{c}}$. Following the adiabatic-approximation formulas (\[eq:approximation\_0\]), whether the energy bias is positive or negative causes an energy level crossing. At last, we realized that i) the limit Hamiltonian, $U(\mathrm{g}/\omega_{\mathrm{c}})^{*} (H_{\mathrm{ptn}}(\omega_{\mathrm{c}})-\hbar\varepsilon\sigma_{z}/2 )U(\mathrm{g}/\omega_{\mathrm{c}}) -\hbar\mathrm{g}^{2}/\omega_{\mathrm{c}}$, as well as its eigenstates and eigenenergies does not include the tunnel splitting $\hbar\omega_{\mathrm{a}}$ of two-level atom, but the energy bias $\hbar\varepsilon$; ii) the eigenstates in Eq.(\[eq:approximation\_AN\]) are the Schrödinger-cat-like, but the eigenstates in Eq.(\[eq:approximation\_0\]) are not. Here, we make a remark on a physical role of the displacement operator $D(\pm\mathrm{g}/\omega_{\mathrm{c}})$ to introduce our problem. The appearance of the displacement operator in Eqs.(\[eq:approximation\_AN\]) and (\[eq:approximation\_0\]) makes coherent states. However, they are for bare photons; and in fact, the photon-field fluctuation $\Delta\Phi$ increases the ground-state expectation $N_{0}^{\mbox{\tiny GQR}}(\omega_{\mathrm{c}},\mathrm{g})$ of the number of photons. More precisely, Eqs.(\[eq:approximation\_AN\]) and (\[eq:approximation\_0\]) say that the ground-state expectation $N_{0}^{\mbox{\tiny GQR}}(\omega_{\mathrm{c}},\mathrm{g}) =\langle E_{0}^{\mbox{\tiny GQR}}|a^{\dagger}a |E_{0}^{\mbox{\tiny GQR}}\rangle$ increases as the coupling strength $\mathrm{g}$ grows larger, i.e., $N_{0}^{\mbox{\tiny GQR}}(\omega_{\mathrm{c}},\mathrm{g}) \sim\mathrm{g}^{2}/\omega_{\mathrm{c}}^{2}$ as $\mathrm{g}\to\infty$, where $|E_{0}^{\mbox{\tiny GQR}}\rangle$ is the ground state of the generalized quantum Rabi Hamiltonian $\mathcal{H}_{\mbox{\tiny GQR}}(\omega_{\mathrm{c}},\mathrm{g})$. This increase results from the mathematical establishment of the adiabatic approximation. Actually, it is pushed up by the fluctuation $\Delta\Phi$ of the photon field $\Phi:= (a+a^{\dagger})/\sqrt{2\omega_{\mathrm{c}}}$ in the ground state since the inequality, $$(\Delta\Phi)^{2} \le \frac{ 2N_{0}^{\mbox{\tiny GQR}}(\omega_{\mathrm{c}},\mathrm{g})+1 }{\omega_{\mathrm{c}}},$$ is obtained in the same way as in Appendix B of Ref.[@HMS]. The mathematical establishment of the adiabatic approximation also says that $(\Delta\Phi)^{2}\sim (1+4\mathrm{g}^{2}/\omega_{\mathrm{c}})/2\omega_{\mathrm{c}}$ as $\mathrm{g}\to\infty$ for $\varepsilon=0$; $(\Delta\Phi)^{2}\sim 1/2\omega_{\mathrm{c}}$ as $\mathrm{g}\to\infty$ for $\varepsilon\ne 0$. Therefore, the Schrödinger-cat-likeness is caused by bare photons, and it is not observable directly. We now try to derive physical states from the adiabatically approximated bare states given in Eqs.(\[eq:approximation\_AN\]) and (\[eq:approximation\_0\]). The Hamiltonians $H_{\pm\mbox{\tiny vH}}$ of the van Hove model for the neutral scalar field theory with a fixed sources [@van-hove] are given by $H_{\pm\mathrm{vH}}:=H_{\mathrm{ptn}}(\omega_{\mathrm{c}}) \pm\hbar\mathrm{g}(a+a^{\dagger})$. We denote by $|n_{\pm\mbox{\tiny vH}}\rangle$ the eigenstate of the van Hove Hamiltonians $H_{\pm\mathrm{vH}}$. Since each eigenstate is given by $|n_{\pm\mbox{\tiny vH}}\rangle =D(\mp\mathrm{g}/\omega_{\mathrm{c}})|n\rangle$, the ground-state expectation is calculated as $\langle 0_{\pm\mbox{\tiny vH}}|a^{\dagger}a|0_{\pm\mbox{\tiny vH}}\rangle =\mathrm{g}^{2}/\omega_{\mathrm{c}}^{2}$. It increases in association with the growth of the coupling strength $\mathrm{g}$ as it looks as it appears to be. However, we find unitary operators $U_{\pm\mbox{\tiny vH}}$ to derive physical states from bare states for the van Hove Hamiltonians $H_{\pm\mathrm{vH}}$, and then, we have the renormalized van Hove Hamiltonian given by $U_{\pm\mbox{\tiny vH}}^{*} (H_{\pm\mbox{\tiny vH}}+\hbar\mathrm{g}^{2}/\omega_{\mathrm{c}}) U_{\pm\mbox{\tiny vH}}^{\mathrm{ren}}$. Here, the energy $-\hbar\mathrm{g}^{2}/\omega_{\mathrm{c}}$ is the self-energy of the van Hove Hamiltonian, and we have to make the so-called mass renormalization [@HT]. In addition to this, Ref.[@HT] tells us that the unitary operators are given by $U_{\pm\mbox{\tiny vH}} =D(\mp\mathrm{g}/\omega_{\mathrm{c}})$, and each physical state $|n_{\pm\mbox{\tiny vH}}^{\mathrm{ren}}\rangle$ of the bare state $|n_{\pm\mbox{\tiny vH}}\rangle$ is given by $|n_{\pm\mbox{\tiny vH}}^{\mathrm{ren}}\rangle =U_{\pm\mbox{\tiny vH}}^{*}|n_{\pm\mbox{\tiny vH}}\rangle$. Eventually, the physical state $|n_{\pm\mbox{\tiny vH}}^{\mathrm{ren}}\rangle$ gets itself satisfying $|n_{\pm\mbox{\tiny vH}}^{\mathrm{ren}}\rangle=|n\rangle$. The photon in the physical state $|n_{\pm\mbox{\tiny vH}}^{\mathrm{ren}}\rangle$ is the so-called dressed photon, which sometimes called real photon. Thus, we can expect no dressed photon in the physical ground state $|0_{\pm\mbox{\tiny vH}}^{\mathrm{ren}}\rangle$, i.e., $\langle 0_{\pm\mbox{\tiny vH}}^{\mathrm{ren}}| a^{\dagger}a |0_{\pm\mbox{\tiny vH}}^{\mathrm{ren}}\rangle=0$. Therefore, $\langle 0_{\pm\mbox{\tiny vH}}|a^{\dagger}a|0_{\pm\mbox{\tiny vH}}\rangle =\mathrm{g}^{2}/\omega_{\mathrm{c}}^{2}$ is the expectation value of the number of the bare photons including virtual photons in the bare ground sate. Following the argument in Complement B${}_{\mbox{\tiny III}}$.2 of Ref.[@cohen-tannoudji], we can think that the photons in the ground state are virtual photons since the ground state expectation $N_{0}^{\mbox{\tiny GQR}}(\omega_{\mathrm{c}},\mathrm{g})$ increases as the coupling strength $\mathrm{g}$ grows larger: $N_{0}^{\mbox{\tiny GQR}}(\omega_{\mathrm{c}},\mathrm{g}) \sim \mathrm{g}^{2}/\omega_{\mathrm{c}}^{2}$ as $\mathrm{g}\to\infty$. Using the representation in Ref.[@CS], we define the annihilation operator $\alpha$ and creation operator $\alpha^{\dagger}$ by $\alpha^{\sharp}:=\sigma_{z}a^{\sharp}$. Then, we have the matrix-valued CCR, $[\alpha , \alpha^{\dagger}]=1$, and the expression, $$\mathcal{H}_{\mbox{\tiny GQR}}(\omega_{\mathrm{c}},\mathrm{g}) =\hbar\omega_{\mathrm{c}}\left(\alpha^{\dagger}\alpha+\frac{1}{2}\right) +\hbar\mathrm{g}\left(\alpha+\alpha^{\dagger}\right) +\mathcal{H}_{\mathrm{atm}}(\varepsilon).$$ In the sufficiently strong coupling regime, the atom Hamiltonian $\mathcal{H}_{\mathrm{atm}}(\varepsilon)$ can be regarded as the perturbation of the van Hove Hamiltonian $\hbar\omega_{\mathrm{c}}\left(\alpha^{\dagger}\alpha+1/2\right) +\hbar\mathrm{g}\left(\alpha+\alpha^{\dagger}\right)$. This is the very idea of the adiabatic approximation. Based on this fact, in a similar way to the van Hove model’s case, we define the unitary operator $U_{\mbox{\tiny GQR}}$ for deriving physical states from bare states by $U_{\mbox{\tiny GQR}}:=\sigma_{+}\sigma_{-}U_{+\mbox{\tiny vH}} +\sigma_{-}\sigma_{+}U_{-\mbox{\tiny vH}}$. Then, for each bare state $\psi$ of the generalized quantum Rabi Hamiltonian $\mathcal{H}_{\mbox{\tiny GQR}}(\omega_{\mathrm{c}},\mathrm{g})$, we have the physical eigenstates $\psi^{\mathrm{ren}}$ by $\psi^{\mathrm{ren}}=U_{\mbox{\tiny GQR}}^{*}\psi$. Since the renormalized Hamiltonian $\mathcal{H}_{\mbox{\tiny GQR}}^{\mathrm{ren}}(\omega_{\mathrm{c}},\mathrm{g})$ of the generalized quantum Rabi Hamiltonian $\mathcal{H}_{\mbox{\tiny GQR}}(\omega_{\mathrm{c}},\mathrm{g})$ is given by $\mathcal{H}_{\mbox{\tiny GQR}}^{\mathrm{ren}}(\omega_{\mathrm{c}},\mathrm{g}) =U_{\mbox{\tiny GQR}}^{*} \left( \mathcal{H}_{\mbox{\tiny GQR}}(\omega_{\mathrm{c}},\mathrm{g}) +\hbar\mathrm{g}^{2}/\omega_{\mathrm{c}} \right) U_{\mbox{\tiny GQR}}$ and $U_{\mbox{\tiny GQR}}=U(\mathrm{g}/\omega_{\mathrm{c}})$, we have $$\mathcal{H}_{\mbox{\tiny GQR}}^{\mathrm{ren}}(\omega_{\mathrm{c}},\mathrm{g}) = H_{\mathrm{ptn}}(\omega_{\mathrm{c}}) -\frac{\hbar}{2}\varepsilon\sigma_{z} -\frac{\hbar}{2}\omega_{\mathrm{a}} \left\{ \sigma_{+}U_{+\mbox{\tiny vH}}^{*}U_{-\mbox{\tiny vH}} +\sigma_{-}U_{-\mbox{\tiny vH}}^{*}U_{+\mbox{\tiny vH}} \right\}. \label{eq:renormalized-GQR-Hamiltonian}$$ We immediately realize that RHS of Eq.(\[eq:renormalized-GQR-Hamiltonian\]) is RHS of Eq.(\[eq:unitary-trans-GQR-Hamiltonian\]). In the case $\varepsilon=0$, using Eq.(\[eq:approximation\_AN\]), all the normalized eigenstates $\psi$ of the quantum Rabi Hamiltonian $\mathcal{H}_{\mbox{\tiny QR}}(\omega_{\mathrm{c}},\mathrm{g})$ are approximated by $(|\!\!\uparrow\rangle|n_{+\mbox{\tiny vH}}\rangle +|\!\!\downarrow\rangle|n_{-\mbox{\tiny vH}}\rangle)/\sqrt{2}$ or $(|\!\!\uparrow\rangle|n_{+\mbox{\tiny vH}}\rangle -|\!\!\downarrow\rangle|n_{-\mbox{\tiny vH}}\rangle)/\sqrt{2}$, $n=0, 1, 2, \cdots$. Thus, the physical eigenstates $\psi^{\mathrm{ren}}$ (i.e., eigenstates of the renormalized Hamiltonian $\mathcal{H}_{\mbox{\tiny QR}}^{\mathrm{ren}}(\omega_{\mathrm{c}},\mathrm{g})$ are approximated by $$\left\{ \begin{array}{ll} {\displaystyle \frac{1}{\sqrt{2}} \Biggl(|\!\!\uparrow\rangle |n_{+\mbox{\tiny vH}}^{\mathrm{ren}}\rangle +|\!\!\downarrow\rangle |n_{-\mbox{\tiny vH}}^{\mathrm{ren}}\rangle \Biggr) = \frac{1}{\sqrt{2}} \Biggl(|\!\!\uparrow\rangle +|\!\!\downarrow\rangle \Biggr) |n\rangle }, \\ \qquad \\ {\displaystyle \frac{1}{\sqrt{2}} \Biggl( |\!\!\uparrow\rangle |n_{+\mbox{\tiny vH}}^{\mathrm{ren}}\rangle -|\!\!\downarrow\rangle |n_{-\mbox{\tiny vH}}^{\mathrm{ren}}\rangle \Biggr) = \frac{1}{\sqrt{2}} \Biggl( |\!\!\uparrow\rangle -|\!\!\downarrow\rangle \Biggr) |n\rangle }. \end{array} \right. \label{eq:approximation_AN'}$$ In the same way, in the case $\varepsilon\ne 0$, using Eq.(\[eq:approximation\_0\]), we have the adiabatically approximated physical eigenstates $$\left\{ \begin{array}{ll} {\displaystyle |\!\!\uparrow\rangle |n_{+\mbox{\tiny vH}}^{\mathrm{ren}}\rangle =|\!\!\uparrow\rangle |n\rangle}, \\ \qquad \\ {\displaystyle |\!\!\downarrow\rangle |n_{-\mbox{\tiny vH}}^{\mathrm{ren}}\rangle =|\!\!\downarrow\rangle |n\rangle }, \end{array} \right. \label{eq:approximation_0'}$$ for sufficiently large coupling strength. Thus, the adiabatically approximated physical states in Eqs.(\[eq:approximation\_AN’\]) and (\[eq:approximation\_0’\]) can no longer make any coherent sate of dressed photons, and are no longer macroscopic. That is, Schrödinger-cat-likeness does not appear in those physical eigenstates. As observed above, if we employ the unitary operator $U_{\mbox{\tiny GQR}}$ to derive physical states from bare states for the (generalized) quantum Rabi model, physical eigenstates are approximated by eigenstates of the free part of the (generalized) quantum Rabi Hamiltonian (i.e., $\mathrm{g}=0$). Moreover, the adiabatic approximations given by Eqs.(\[eq:approximation\_AN’\]) and (\[eq:approximation\_0’\]) tell us that the (generalized) quantum Rabi Hamiltonian cannot make us expect any dressed photon in the physical ground state even for sufficiently large coupling strength. Here, we point out the following properties: P1) : The derivations of Eqs.(\[eq:approximation\_AN\]) and (\[eq:approximation\_0\]) tell us that whether the adiabatically approximated eigenstates of the generalized quantum Rabi Hamiltonian are formed as the Schrödinger-cat-like entangled states or not depends on whether the energy-bias parameter is zero or non-zero, in other words, whether the parity symmetry given by Eq.(\[eq:parity-symmetry’\]) is conserved or not. P2) : The limit Hamiltonian in the norm resolvent sense says that the effect of the tunnel splitting of the two-level atom can be ignored in the adiabatic approximations. P3) : Following the theory of van Hove model, the Schrödinger-cat-likeness disappears from the adiabatically approximated physical eigenstates. We are interested in a vestige of the Schrödinger-cat-likeness of bare photons in the physical eigenstates. Meanwhile, we have to consider the quadratic interaction (i.e., $A^{2}$-term) of the photon field in the case where our physical system of the two-level atom interacting with the one-mode photon field is in the very strong coupling regime such as the deep-strong coupling regime. For such a situation, we should consider the Hamiltonian $$\mathcal{H}_{A^{2}}(\varepsilon):= \mathcal{H}_{\mbox{\tiny GQR}}(\omega_{\mathrm{c}},\mathrm{g}) +\hbar\mathrm{g}C_{\mathrm{g}} \left(a+a^{\dagger}\right)^{2}, \label{eq:total-Hamiltonian}$$ where $C_{\mathrm{g}}$ is a positive function of the coupling strength $\mathrm{g}$ satisfying $\lim_{\mathrm{g}\to\infty}\mathrm{g}C_{\mathrm{g}}=\infty$. We assume the following conditions: $$\lim_{\mathrm{g}\to\infty}\mathrm{g}^{-1/3}C_{\mathrm{g}}=\infty, \label{eq:assumption}$$ and there is a non-negative constant $C_{\infty}$ so that $$\lim_{\mathrm{g}\to\infty}\mathrm{g}^{-1}C_{\mathrm{g}}=C_{\infty}. \label{eq:assumption'}$$ For instance, if we estimate $C_{\mathrm{g}}$ at $(\mathrm{const})\times\mathrm{g}$, then Eqs.(\[eq:assumption\]) and (\[eq:assumption’\]) hold. In this paper, we investigate an effect caused by the $A^{2}$-term in the physical eigenstates of our total Hamiltonian $\mathcal{H}_{A^{2}}(\varepsilon)$. To obtain the physical states, we employ the (meson) pair theory in nuclear physics [@HT; @wentzel1; @wentzel2; @BR; @KM]. Then, we have a unitary operator $U_{A^{2}}$ such that the unitarily transformed Hamiltonian $U_{A^{2}}^{*}\mathcal{H}_{A^{2}}(\varepsilon)U_{A^{2}}$ becomes the renormalized Hamiltonian for the physical eigenstates. Following (meson) pair theory, we obtain the unitary operator $U_{A^{2}}$ given by the Hopfield-Bogoliubov transformation $U_{\mbox{\tiny HB}}$, i.e., $U_{A^{2}}=U_{\mbox{\tiny HB}}$, as shown in Ref.[@HMS], so that we obtain the renormalized Hamiltonian $\mathcal{H}_{A^{2}}^{\mathrm{ren}}(\varepsilon)$ as $$\mathcal{H}_{A^{2}}^{\mathrm{ren}}(\varepsilon) := U_{A^{2}}^{*}\mathcal{H}_{A^{2}}(\varepsilon)U_{A^{2}} =\mathcal{H}_{\mathrm{atm}}(\varepsilon) +H_{\mathrm{ptn}}(\omega_{\mathrm{g}}) +\hbar\widetilde{\mathrm{g}}\sigma_{z}\left(a+a^{\dagger}\right) = \mathcal{H}_{\mbox{\tiny GQR}}(\omega_{\mathrm{g}},\widetilde{\mathrm{g}}), \qquad \label{eq:renormalized-Hamiltonian}$$ where $\omega_{\mathrm{g}}$ and $\widetilde{\mathrm{g}}$ are respectively renormalized photon frequency and the renormalized coupling strength given by $$\omega_{\mathrm{g}}= \sqrt{\omega_{\mathrm{c}}^{2}+4\omega_{\mathrm{c}}\mathrm{g}C_{\mathrm{g}}} \quad \mbox{and}\quad \widetilde{\mathrm{g}} =\mathrm{g}\sqrt{\frac{\omega_{\mathrm{c}}}{\omega_{\mathrm{g}}}}.$$ We briefly review how to obtain Eq.(\[eq:renormalized-Hamiltonian\]) in the next section. Similarly to the argument for the generalized quantum Rabi model, we introduce the unitary operator $U(\widetilde{\mathrm{g}}/\omega_{\mathrm{g}})$ by $$U(\widetilde{\mathrm{g}}/\omega_{\mathrm{g}}) :=\sigma_{+}\sigma_{-}D(\widetilde{\mathrm{g}}/\omega_{\mathrm{g}}) +\sigma_{-}\sigma_{+}D(-\widetilde{\mathrm{g}}/\omega_{\mathrm{g}}), \label{eq:unitary-operator-van_Hove}$$ where $D(\widetilde{\mathrm{g}}/\omega_{\mathrm{g}})$ is the displacement operator defined by $D(\widetilde{\mathrm{g}}/\omega_{\mathrm{g}}) :=\exp\left[\widetilde{\mathrm{g}} \left(a^{\dagger}-a\right)/\omega_{\mathrm{g}}\right]$. Then, we have the unitary transformation, $$\begin{aligned} &{}& U\left(\widetilde{\mathrm{g}}/\omega_{\mathrm{g}}\right) \left(\mathcal{H}_{A^{2}}^{\mathrm{ren}}(\varepsilon) +\hbar\widetilde{\mathrm{g}}^{2}/\omega_{\mathrm{g}}\right) U\left(\widetilde{\mathrm{g}}/\omega_{\mathrm{g}}\right)^{*} \nonumber \\ &=& H_{\mathrm{ptn}}(\omega_{\mathrm{g}}) -\frac{\hbar}{2}\varepsilon\sigma_{z} -\frac{\hbar}{2}\omega_{\mathrm{a}} \left\{ \sigma_{+}D\left(\widetilde{\mathrm{g}}/\omega_{\mathrm{g}}\right)^{2} +\sigma_{-}D\left(-\widetilde{\mathrm{g}}/\omega_{\mathrm{g}}\right)^{2} \right\}. \label{eq:unitary-trans}\end{aligned}$$ For Eq.(\[eq:unitary-trans\]), we realize the followings: Since the limit $$\lim_{\mathrm{g}\to\infty} \hbar\frac{\widetilde{\mathrm{g}}^{2}}{\omega_{\mathrm{g}}} =\frac{\hbar}{4C_{\infty}} \label{eq:limit_of_self-energy}$$ follows from Eq.(\[eq:assumption’\]), the self-energy $\hbar\widetilde{\mathrm{g}}^{2}/\omega_{\mathrm{g}}$ does not work as a counter term when we take the strong coupling limit $\mathrm{g}\to\infty$. Instead, we meet a trouble of divergence for the Hamiltonian $H_{\mathrm{ptn}}(\omega_{\mathrm{g}})$ due to the divergence $\lim_{\mathrm{g}\to\infty}\omega_{\mathrm{g}} =\infty$. In addition to this trouble, the third term of RHS of Eq.(\[eq:unitary-trans\]) does not vanish as the coupling strength $\mathrm{g}$ tends to the infinity because the displacement operators $D(\widetilde{\mathrm{g}}/\omega_{\mathrm{g}})$ and $D(\widetilde{\mathrm{g}}/\omega_{\mathrm{g}})^{*}$ do not decay to zero because of the limit, $$\lim_{\mathrm{g}\to\infty} \frac{\widetilde{\mathrm{g}}}{\,\,\omega_{\mathrm{g}}} =\lim_{\mathrm{g}\to\infty} \omega_{\mathrm{c}}^{-1/4} \left( \omega_{\mathrm{c}}\mathrm{g}^{-4/3} +4\mathrm{g}^{-1/3}C_{\mathrm{g}} \right)^{-3/4}=0,$$ by Eq.(\[eq:assumption\]). In this paper, coping with the trouble and difference, we will consider the properties corresponding to **P1**, **P2**, and **P3** for our total Hamiltonian $\mathcal{H}^{\mathrm{ren}}_{A^{2}}(\varepsilon)$. From Bare Eigenstates to Physical Eigenstates ============================================= Following (meson) pair theory [@HT; @wentzel1; @wentzel2; @BR; @KM], we obtain physical eigenstates from bare ones. We review it in brief. For more details on how to apply (meson) pair theory to our model, see the argument in Ref.[@HMS]. Our Hamiltonian $\mathcal{H}_{A^{2}}(\varepsilon)$ has the matrix representation as $$\mathcal{H}_{A^{2}}(\varepsilon) = \left( \begin{array}{cccc} H_{A^{2}}^{+}-\hbar\varepsilon/2 & -\hbar\omega_{\mathrm{a}}/2 \\ -\hbar\omega_{\mathrm{a}}/2 & H_{A^{2}}^{-}+\hbar\varepsilon/2 \end{array} \right),$$ where $$H_{A^{2}}^{\pm}= H_{\mathrm{ptn}}(\omega_{\mathrm{c}}) \pm\hbar\mathrm{g}\left( a+ a^{\dagger}\right) +\hbar C_{\mathrm{g}}\mathrm{g} \left( a+ a^{\dagger}\right)^{2}.$$ Using Eqs.(\[eq:expression\_xp\]) and (\[eq:another-representation-H\_ptn\]), we can rewrite the Hamiltonians $H_{A^{2}}^{\pm}$ as $$H_{A^{2}}^{\pm}= \frac{1}{2}p^{2}+\frac{\omega_{\mathrm{c}}^{2}}{2}x^{2} \pm\hbar\mathrm{g}\sqrt{\frac{2\omega_{\mathrm{c}}}{\hbar}}\, x +2C_{\mathrm{g}}\mathrm{g}\omega_{\mathrm{c}}x^{2} = \frac{1}{2}p^{2}+\frac{\omega_{\mathrm{g}}^{2}}{2}x^{2} \pm\hbar\widetilde{\mathrm{g}}\sqrt{\frac{2\omega_{\mathrm{g}}}{\hbar}}\, x.$$ We define new photon’s annihilation operator $b$ and creation operator $b^{\dagger}$ by $$b:=\sqrt{\frac{\omega_{\mathrm{g}}}{2\hbar}}\, x +i\sqrt{\frac{1}{2\hbar\omega_{\mathrm{g}}}}\, p \quad\textrm{and}\quad b^{\dagger}:=\sqrt{\frac{\omega_{\mathrm{g}}}{2\hbar}}\, x -i\sqrt{\frac{1}{2\hbar\omega_{\mathrm{g}}}}\, p. \label{eq:expression_xp'}$$ Then, we have expression of the Hamiltonians $H_{A^{2}}^{\pm}$ using the new annihilation and creation operators, $b$ and $b^{\dagger}$, as $$H_{A^{2}}^{\pm}= \hbar\omega_{\mathrm{g}}\left(b^{\dagger}b+\frac{1}{2}\right) \pm\hbar\widetilde{\mathrm{g}}\left(b+b^{\dagger}\right). \label{eq:key-representation}$$ Making the correspondence between the normalized eigenstates of the Hamiltonian $(1/2)p^{2}+(\omega_{\mathrm{c}}^{2}/2)x^{2}$ to those of the Hamiltonian $(1/2)p^{2}+(\omega_{\mathrm{g}}^{2}/2)x^{2}$, we eventually obtain the so-called Hopfield-Bogoliubov transformation $U_{\mbox{\tiny HB}}$, and then, reach the unitary transformation, $$\begin{cases} U_{\mbox{\tiny HB}}aU_{\mbox{\tiny HB}}^{*}=b =\frac{1}{2}(c_{1}+c_{2})a+\frac{1}{2}(c_{1}-c_{2})a^{\dagger}, \\ U_{\mbox{\tiny HB}}a^{\dagger}U_{\mbox{\tiny HB}}^{*}=b^{\dagger} =\frac{1}{2}(c_{1}-c_{2})a+\frac{1}{2}(c_{1}+c_{2})a^{\dagger}, \end{cases}$$ where $c_{1}=\sqrt{\omega_{\mathrm{g}}/\omega_{\mathrm{c}}}$ and $c_{2}=\sqrt{\omega_{\mathrm{c}}/\omega_{\mathrm{g}}}$. We note that the Hopfield-Bogoliubov transformation $U_{\mbox{\tiny HB}}$ is concretely defined by Eqs.(50) and (57) of Ref.[@HMS] or Eqs.(12.17)-(12.19) of Ref.[@HT] in (meson) pair theory. Use the Hopfield-Bogoliubov transformation $U_{\mbox{\tiny HB}}$, and we obtain the unitary transformation $$U_{\mbox{\tiny HB}}^{*}\mathcal{H}_{A^{2}}(\varepsilon)U_{\mbox{\tiny HB}} =\mathcal{H}_{\mbox{\tiny GQR}}(\omega_{\mathrm{g}},\widetilde{\mathrm{g}}). \label{eq:Hamiltonian-after-HB}$$ This is our renormalized Hamiltonian $\mathcal{H}_{A^{2}}^{\mathrm{ren}}(\varepsilon)$ in Eq.(\[eq:renormalized-Hamiltonian\]). We denote by $|\mathcal{E}_{\nu}(\varepsilon)\rangle$ eigenstates of our total Hamiltonian $\mathcal{H}_{A^{2}}(\varepsilon)$ with eigenenergy $E_{\nu}$, i.e., $\mathcal{H}_{A^{2}}(\varepsilon)|\mathcal{E}_{\nu}(\varepsilon)\rangle =E_{\nu}|\mathcal{E}_{\nu}(\varepsilon)\rangle$. We make the order of the eigenenergies as $E_{1}\le E_{2}\le \cdots\le E_{\nu}\le E_{\nu+1}\le \cdots$. These eigenstates, $|\mathcal{E}_{\nu}(\varepsilon)\rangle$, $\nu=0, 1, 2, \cdots$, are bare states. According to (meson) pair theory, we should derive the physical eigenstates $|\mathcal{E}_{\nu}^{\mathrm{ren}}(\varepsilon)\rangle$ form the bare ones by $$|\mathcal{E}_{\nu}^{\mathrm{ren}}(\varepsilon)\rangle :=U_{\mbox{\tiny HB}}^{*}|\mathcal{E}_{\nu}(\varepsilon)\rangle. \label{eq:physical-eigenstaes}$$ Then, we have $\mathcal{H}_{A^{2}}^{\mathrm{ren}}(\varepsilon) |\mathcal{E}_{\nu}^{\mathrm{ren}}(\varepsilon)\rangle =E_{\nu}|\mathcal{E}_{\nu}^{\mathrm{ren}}(\varepsilon)\rangle$. In next section, we will give the adiabatic approximation for these physical eigenstates $|\mathcal{E}_{\nu}^{\mathrm{ren}}(\varepsilon)\rangle$ by taking the strong coupling limit. Schrödinger-Cat-Likeness in Adiabatic Approximation =================================================== In this section, we show the Schrödinger-cat-likeness in the adiabatic approximation. We will give a mathematical proof of the adiabatic approximation in the next section. To make an energy renormalization for $H_{\mathrm{ptn}}(\omega_{\mathrm{g}})$ as $\mathrm{g}\to\infty$, we introduce a function $\Delta_{\mathrm{g}}$ of the coupling strength $\mathrm{g}$ such that there is a positive function $\delta_{\mathrm{g}}$ satisfying the conditions, $$\begin{aligned} &{}& |\omega_{\mathrm{c}}-\left(\omega_{\mathrm{g}}-\Delta_{\mathrm{g}}\right)| \le \delta_{\mathrm{g}}\omega_{\mathrm{c}}, \label{eq:assumption1} \\ &{}& {\displaystyle \lim_{\mathrm{g}\to\infty}\delta_{\mathrm{g}}=0}. \label{eq:assumption2}\end{aligned}$$ These conditions yield the limit $$\lim_{\mathrm{g}\to\infty} (\omega_{\mathrm{g}}-\Delta_{\mathrm{g}})=\omega_{\mathrm{c}}. \label{eq:app_omega-g}$$ For example, take $\Delta_{\mathrm{g}}$ as $\Delta_{\mathrm{g}}= \sqrt{\omega_{\mathrm{g}}^{2}-4\omega_{\mathrm{c}}\sqrt{\mathrm{g}C_{\mathrm{g}} \omega_{\mathrm{c}}}}$ for every coupling strength $\mathrm{g}$ with $\sqrt{\omega_{\mathrm{c}}/\mathrm{g}C_{\mathrm{g}}}\ne 2$. Then, we have the equations, $$\begin{aligned} &{}& \omega_{\mathrm{c}} \left\{ 1-\frac{1}{|1-\sqrt{\omega_{\mathrm{c}}/4\mathrm{g}C_{\mathrm{g}}}|} \right\} = \omega_{\mathrm{c}} \left\{ 1-\frac{1}{\sqrt{(\sqrt{\omega_{\mathrm{c}}/4\mathrm{g}C_{\mathrm{g}}})^{2} +1^{2}-\sqrt{\omega_{\mathrm{c}}/\mathrm{g}C_{\mathrm{g}}}}} \right\} \nonumber \\ &=& \omega_{\mathrm{c}} \left\{ 1-\frac{1}{\sqrt{1+(\omega_{\mathrm{c}}/4\mathrm{g}C_{\mathrm{g}}) -\sqrt{\omega_{\mathrm{c}}/\mathrm{g}C_{\mathrm{g}}}}} \right\}, \label{eq:eq-1}\end{aligned}$$ and $$\begin{aligned} &{}& \omega_{\mathrm{c}}-(\omega_{\mathrm{g}}-\Delta_{\mathrm{g}}) \nonumber \\ &=& \omega_{\mathrm{c}} \left\{ 1-\frac{2}{ \sqrt{1+(\omega_{\mathrm{c}}/4\mathrm{g}C_{\mathrm{g}})} +\sqrt{1+(\omega_{\mathrm{c}}/4\mathrm{g}C_{\mathrm{g}}) -\sqrt{\omega_{\mathrm{c}}/\mathrm{g}C_{\mathrm{g}}}} } \right\}. \label{eq:eq-2}\end{aligned}$$ Meanwhile, since we have the inequalities, $$\begin{aligned} &{}& 2\sqrt{1+(\omega_{\mathrm{c}}/4\mathrm{g}C_{\mathrm{g}}) -\sqrt{\omega_{\mathrm{c}}/\mathrm{g}C_{\mathrm{g}}}} \\ &\le& \sqrt{1+(\omega_{\mathrm{c}}/4\mathrm{g}C_{\mathrm{g}})} +\sqrt{1+(\omega_{\mathrm{c}}/4\mathrm{g}C_{\mathrm{g}}) -\sqrt{\omega_{\mathrm{c}}/\mathrm{g}C_{\mathrm{g}}}} \\ &\le& 2\sqrt{1+(\omega_{\mathrm{c}}/4\mathrm{g}C_{\mathrm{g}})}, \end{aligned}$$ we reach the inequalities, $$\begin{aligned} &{}& 1-\frac{1}{ \sqrt{1+(\omega_{\mathrm{c}}/4\mathrm{g}C_{\mathrm{g}}) -\sqrt{\omega_{\mathrm{c}}/\mathrm{g}C_{\mathrm{g}}}} } \nonumber \\ &\le& 1-\frac{2}{ \sqrt{1+(\omega_{\mathrm{c}}/4\mathrm{g}C_{\mathrm{g}})} +\sqrt{1+(\omega_{\mathrm{c}}/4\mathrm{g}C_{\mathrm{g}}) -\sqrt{\omega_{\mathrm{c}}/\mathrm{g}C_{\mathrm{g}}}} } \nonumber \\ &\le& 1-\frac{1}{ \sqrt{1+(\omega_{\mathrm{c}}/4\mathrm{g}C_{\mathrm{g}})} }. \label{eq:ineq-1}\end{aligned}$$ By Eqs.(\[eq:eq-1\]), (\[eq:eq-2\]), and (\[eq:ineq-1\]), we have the following two inequalities, $$\omega_{\mathrm{c}} \left\{ 1-\frac{1}{|1-\sqrt{\omega_{\mathrm{c}}/4\mathrm{g}C_{\mathrm{g}}}|} \right\} \le \omega_{\mathrm{c}}-(\omega_{\mathrm{g}}-\Delta_{\mathrm{g}}) \le \omega_{\mathrm{c}} \left\{ 1-\frac{1}{ \sqrt{1+(\omega_{\mathrm{c}}/4\mathrm{g}C_{\mathrm{g}})} } \right\}.$$ Here, we note the inequality, $1-\sqrt{\omega_{\mathrm{c}}/\mathrm{g}C_{\mathrm{g}}}\le 1$. These inequalities suggest us that we chose the function $\delta_{\mathrm{g}}$ as $$\delta_{\mathrm{g}}= \max \left\{ \Biggl| 1-\frac{1}{|1-\sqrt{\omega_{\mathrm{c}}/4\mathrm{g}C_{\mathrm{g}}}|} \Biggr|\,\,\, ,\,\,\, \Biggl| 1-\frac{1}{\sqrt{1+(\omega_{\mathrm{c}}/4\mathrm{g}C_{\mathrm{g}}})} \Biggr| \right\}.$$ As proved in the next section, we have the adiabatic approximation of the renormalized Hamiltonian $\mathcal{H}^{\mathrm{ren}}_{A^{2}}(\varepsilon)$: $$\mathcal{H}^{\mathrm{ren}}_{A^{2}}(\varepsilon) \approx U(\widetilde{\mathrm{g}}/\omega_{\mathrm{g}})^{*} \Biggl\{ H_{\mathrm{ptn}}(\omega_{\mathrm{g}}) +\mathcal{H}_{\mathrm{atm}}(\varepsilon) \Biggr\} U(\widetilde{\mathrm{g}}/\omega_{\mathrm{g}}) -\hbar\frac{\widetilde{\mathrm{g}}^{2}}{\omega_{\mathrm{g}}}. \label{eq:main-result}$$ We note that the whole atom energy $\mathcal{H}_{\mathrm{atm}}(\varepsilon)$ remains in the adiabatic approximation. Thanks to Theorem VIII.23(b) of Ref.[@RS1], we can obtain the adiabatic approximation of the eigenstates and their corresponding eigenenergies in the following. In the case $\varepsilon=0$, the adiabatic approximation of all the physical eigenstates $|\mathcal{E}_{\nu}^{\mathrm{ren}}(0)\rangle$ of the renormalized Hamiltonian $\mathcal{H}_{A^{2}}^{\mathrm{ren}}(0)$ are given by the same formulas of the Schrödinger-cat-like entangled sates as in Eq.(\[eq:approximation\_AN\]): $$\left\{ \begin{array}{ll} {\displaystyle \frac{1}{\sqrt{2}} \Bigg(|\!\!\uparrow\rangle D\left(-\widetilde{\mathrm{g}}/\omega_{\mathrm{g}}\right) |n\rangle +|\!\!\downarrow\rangle D\left(\widetilde{\mathrm{g}}/\omega_{\mathrm{g}}\right) |n\rangle \Biggr)}, \\ \qquad \\ {\displaystyle \frac{1}{\sqrt{2}} \Bigg( |\!\!\uparrow\rangle D\left(-\widetilde{\mathrm{g}}/\omega_{\mathrm{g}}\right) |n\rangle -|\!\!\downarrow\rangle D\left(\widetilde{\mathrm{g}}/\omega_{\mathrm{g}}\right) |n\rangle \Biggr)}. \end{array} \right. \label{eq:approximation_1}$$ We denote by $|\mathcal{E}_{n}^{\mathrm{app},+}(0)\rangle$ the first expression in Eq.(\[eq:approximation\_1\]), and by $|\mathcal{E}_{n}^{\mathrm{app},-}(0)\rangle$ the second one. The eigenenergy of the approximated eigenstate $|\mathcal{E}_{n}^{\mathrm{app},+}(0)\rangle$ is $$\hbar\omega_{\mathrm{g}}\left( n+\frac{1}{2}\right) -\frac{\hbar\omega_{\mathrm{a}}}{2} -\hbar\frac{\widetilde{\mathrm{g}}^{2}}{\omega_{\mathrm{g}}} \approx \hbar(\omega_{\mathrm{c}}+\Delta_{\mathrm{g}})n +\frac{\hbar(\omega_{\mathrm{g}}-\omega_{\mathrm{a}})}{2} -\frac{\hbar}{4C_{\infty}},$$ and that of the approximated eigenstate $|\mathcal{E}_{n}^{\mathrm{app},-}(0)\rangle$ is $$\hbar\omega_{\mathrm{g}}\left( n+\frac{1}{2}\right) +\frac{\hbar\omega_{\mathrm{a}}}{2} -\hbar\frac{\widetilde{\mathrm{g}}^{2}}{\omega_{\mathrm{g}}} \approx \hbar(\omega_{\mathrm{c}}+\Delta_{\mathrm{g}})n +\frac{\hbar(\omega_{\mathrm{g}}+\omega_{\mathrm{a}})}{2} -\frac{\hbar}{4C_{\infty}},$$ which says that the tunnel splitting of the two-level atom remains in the adiabatic approximation. In the case $\varepsilon\ne 0$, all the physical eigenstates $|\mathcal{E}_{\nu}^{\mathrm{ren}}(\varepsilon)\rangle$ of the renormalized Hamiltonian $\mathcal{H}_{A^{2}}^{\mathrm{ren}}(\varepsilon)$ have the following adiabatic approximation: $$\left\{ \begin{array}{ll} c_{+\varepsilon,\omega_{\mathrm{a}}} \Biggl( -\omega_{\mathrm{a}}|\!\!\uparrow\rangle D(-\widetilde{\mathrm{g}}/\omega_{\mathrm{g}})|n\rangle +(\varepsilon-\sqrt{\varepsilon^{2}+\omega_{\mathrm{a}}^{2}}) |\!\!\downarrow\rangle D(\widetilde{\mathrm{g}}/\omega_{\mathrm{g}})|n\rangle \Biggr),\\ \qquad \\ c_{-\varepsilon,\omega_{\mathrm{a}}} \Biggl( -\omega_{\mathrm{a}}|\!\!\uparrow\rangle D(-\widetilde{\mathrm{g}}/\omega_{\mathrm{g}})|n\rangle +(\varepsilon+\sqrt{\varepsilon^{2}+\omega_{\mathrm{a}}^{2}}) |\!\!\downarrow\rangle D(\widetilde{\mathrm{g}}/\omega_{\mathrm{g}})|n\rangle \Biggr), \end{array} \right. \label{eq:approximation_2}$$ where the positive constant $c_{\pm\varepsilon,\omega_{\mathrm{a}}}$ is given by $1/c_{\pm\varepsilon,\omega_{\mathrm{a}}}^{2}= 2(\omega_{\mathrm{a}}^{2}+\varepsilon^{2} \mp\varepsilon\sqrt{\omega_{\mathrm{a}}^{2}+\varepsilon^{2}})$. Eqs.(\[eq:approximation\_2\]) show up the Schrödinger-cat-like entangled states. We denote by $|\mathcal{E}_{n}^{\mathrm{app},+}(\varepsilon)\rangle$ the first expression in Eq.(\[eq:approximation\_2\]), and by $|\mathcal{E}_{n}^{\mathrm{app},-}(\varepsilon)\rangle$ the second one. The eigenenergy of the adiabatically approximated eigenstates $|\mathcal{E}_{n}^{\mathrm{app},\pm}(\varepsilon)\rangle$ is $$\begin{aligned} &{}& \mp\frac{\hbar}{2}\sqrt{\omega_{\mathrm{a}}^{2}+\varepsilon^{2}} +\hbar\omega_{\mathrm{g}}\left(n+\frac{1}{2}\right) -\hbar\frac{\widetilde{\mathrm{g}}^{2}}{\omega_{\mathrm{g}}} \nonumber \\ &\approx& \mp\frac{\hbar}{2}\sqrt{\omega_{\mathrm{a}}^{2}+\varepsilon^{2}} +\hbar(\omega_{\mathrm{c}}+\Delta_{\mathrm{g}})n +\frac{\hbar}{2}\omega_{\mathrm{g}} -\frac{\hbar}{4C_{\infty}}. \label{eq:eigenenergies_b}\end{aligned}$$ The energy difference between the two adiabatically approximated eigenstates, $|\mathcal{E}_{n}^{\mathrm{app},-}(\varepsilon)\rangle$ and $|\mathcal{E}_{n}^{\mathrm{app},+}(\varepsilon)\rangle$, is $\mathcal{E}_{n}^{\mathrm{app},-}(\varepsilon) -\mathcal{E}_{n}^{\mathrm{app},+}(\varepsilon) =\hbar\sqrt{\omega_{\mathrm{a}}^{2}+\varepsilon^{2}}$. Since the energy difference between $\mathcal{E}_{n}^{\mathrm{app},\pm}(\varepsilon)$ and $\mathcal{E}_{n+1}^{\mathrm{app},\pm}(\varepsilon)$ is $\mathcal{E}_{n+1}^{\mathrm{app},\pm}(\varepsilon) -\mathcal{E}_{n}^{\mathrm{app},\pm}(\varepsilon) =\hbar\left(\omega_{\mathrm{g}}+\Delta_{\mathrm{g}}\right)$, we can obtain $$0< \mathcal{E}_{n}^{\mathrm{app},-}(\varepsilon) -\mathcal{E}_{n}^{\mathrm{app},+}(\varepsilon) \le \mathcal{E}_{n+1}^{\mathrm{app},\pm}(\varepsilon) -\mathcal{E}_{n}^{\mathrm{app},\pm}(\varepsilon) \label{eq:energy-difference}$$ by controlling the parameters, $\omega_{\mathrm{a}}$, $\varepsilon$, $\omega_{\mathrm{c}}$, $\mathrm{g}$, and $C_{\mathrm{g}}$. We give an application example of the adiabatic approximation given by Eqs.(\[eq:approximation\_1\]) and (\[eq:approximation\_2\]). Take the energy-bias parameter $\varepsilon$ as the transverse axis now. The adiabatic approximation in Eq.(\[eq:approximation\_1\]) says that whether the energy bias is positive or negative causes an energy level crossing with respect to the $\varepsilon$-axis because the model does not have the $A^{2}$-term. On the other hand, the adiabatic-approximation formula, Eq.(\[eq:approximation\_2\]) with Eq.(\[eq:eigenenergies\_b\]), says that the $A^{2}$-term makes an avoided crossing with respect to the $\varepsilon$-axis. In the same way as the proof of Eq.(21) of Ref.[@HMS], we can obtain the expression of the ground-state expectation $N^{\mathrm{ren}}_{0} :=\langle\mathcal{E}^{\mathrm{ren}}_{0}|a^{\dagger}a |\mathcal{E}^{\mathrm{ren}}_{0}\rangle$ of the number of dressed photons and estimate it as $$\begin{aligned} N_{0}^{\mathrm{ren}} &=& \hbar^{2}\widetilde{\mathrm{g}}^{2} \sum_{\nu=0}^{\infty} \frac{|\langle\mathcal{E}_{\nu}^{\mathrm{ren}}| \sigma_{z}|\mathcal{E}_{0}^{\mathrm{ren}}\rangle|^{2}}{ (E_{\nu}-E_{0}+\hbar\omega_{\mathrm{g}})^{2} } \\ &\le& \frac{\widetilde{\mathrm{g}}^{2}}{\omega_{\mathrm{g}}^{2}} \sum_{\nu=0}^{\infty} |\langle\mathcal{E}_{\nu}^{\mathrm{ren}}| \sigma_{z}|\mathcal{E}_{0}^{\mathrm{ren}}\rangle|^{2} = \frac{\widetilde{\mathrm{g}}^{2}}{\omega_{\mathrm{g}}^{2}} \|\sigma_{z}|\mathcal{E}_{0}^{\mathrm{ren}}\rangle \|_{\mathbb{C}^{2}\otimes L^{2}(\mathbb{R})}^{2} \le \frac{\widetilde{\mathrm{g}}^{2}}{\omega_{\mathrm{a}}^{2}}.\end{aligned}$$ We define the approximated ground-state expectation $N_{0}^{\mathrm{app}}$ using the adiabatic approximation, i.e., $N_{0}^{\mathrm{app}}:= \langle\mathcal{E}_{0}^{\mathrm{app},+}(\varepsilon)| a^{\dagger}a|\mathcal{E}_{0}^{\mathrm{app},+}(\varepsilon)\rangle$. Then, the immediate calculation gives us the expression, $$N_{0}^{\mathrm{app}}= \frac{\widetilde{\mathrm{g}}^{2}}{\omega_{\mathrm{g}}^{2}} =\omega_{\mathrm{c}}^{-1/2} \left( \omega_{\mathrm{c}}\mathrm{g}^{-4/3} +4\mathrm{g}^{-1/3}C_{\mathrm{g}} \right)^{-3/2}, \label{eq:adiabatic-approximation_N_0}$$ and thus, we have the limit, $\lim_{\mathrm{g}\to\infty}N_{0}^{\mathrm{ren}}= \lim_{\mathrm{g}\to\infty}N_{0}^{\mathrm{app}}=0$, with $N_{0}^{\mathrm{ren}}\le N_{0}^{\mathrm{app}}$. Meanwhile, for the one-mode photon field $\Phi^{\mathrm{ren}} :=(a+a^{\dagger})/\sqrt{2\omega_{\mathrm{g}}}$ our adiabatic approximation immediately shows that the fluctuation $\Delta\Phi^{\mathrm{ren}}$ decays to zero as $\mathrm{g}\to\infty$. We have information on the dressed photon in the ground state in the following. The symbol $\sharp$ stands for one of the (in)equality symbols, $>$, $=$, $<$. Eq.(\[eq:adiabatic-approximation\_N\_0\]) says that the (in)equality, $N_{0}^{\mathrm{app}}\,\sharp\, 1$, is equivalent to the (in)equality, $$\frac{1}{4}\left(\frac{\mathrm{g}}{\omega_{\mathrm{c}}}\right)^{-1} \left\{ \left(\frac{\mathrm{g}}{\omega_{\mathrm{c}}}\right)^{4/3}-1 \right\}\,\,\, \sharp\,\,\, C_{\mathrm{g}}. \label{eq:equivalence1}$$ Particularly, in the case where the function $C_{\mathrm{g}}$ is given by $C_{\mathrm{g}}=C\mathrm{g}$ with a constant $C$, Eq.(\[eq:equivalence1\]) can be written as $$\frac{1}{4\omega_{\mathrm{c}}} \left(\frac{\mathrm{g}}{\omega_{\mathrm{c}}}\right)^{-2} \left\{ \left(\frac{\mathrm{g}}{\omega_{\mathrm{c}}}\right)^{4/3}-1 \right\}\,\,\, \sharp\,\,\, C. \label{eq:equivalence2}$$ Namely, following the (meson) pair theory, if we can make the constant $C$ so small that it satisfies the condition, $(\omega_{\mathrm{c}}/4\mathrm{g}^{2}) \{(\mathrm{g}/\omega_{\mathrm{c}})^{4/3}-1\}\ge C$, then there is a possibility that the ground state of the generalized quantum Rabi model has some dressed photons. We will explain the reason why we are interested in Eqs.(\[eq:equivalence1\]) and (\[eq:equivalence2\]) in Section \[sec:CD\]. A Proof of Adiabatic Approximation for $\mathcal{H}^{\mathrm{ren}}_{A^{2}}(\varepsilon)$ {#sec:proof} ======================================================================================== We define the modified photon Hamiltonian $\widetilde{H}_{\mathrm{ptn}}(\omega)$ for every frequency $\omega$ by removing the zero-point energy, that is, $$\widetilde{H}_{\mathrm{ptn}}(\omega) :=H_{\mathrm{ptn}}(\omega) -\hbar\omega/2 =\hbar\omega a^{\dagger}a.$$ In this section, we will take frequencies, $\omega_{\mathrm{c}}$, $\omega_{\mathrm{g}}$, $\omega_{\mathrm{c}}-(\omega_{\mathrm{g}}-\Delta_{\mathrm{g}})$, or $\omega_{\mathrm{c}}-\omega_{\mathrm{g}}$ as $\omega$. With this modification, we slightly modify the renormalized Hamiltonian $\mathcal{H}_{A^{2}}^{\mathrm{ren}}(\varepsilon)$ as $\widetilde{\mathcal{H}}_{A^{2}}^{\mathrm{ren}}(\varepsilon) :=\mathcal{H}_{A^{2}}^{\mathrm{ren}}-\hbar\omega_{\mathrm{g}}/2$. By Eq.(\[eq:renormalized-Hamiltonian\]), we have the expression, $$\widetilde{\mathcal{H}}_{A^{2}}^{\mathrm{ren}}(\varepsilon) =\mathcal{H}_{\mbox{\tiny GQR}}(\omega_{\mathrm{g}},\widetilde{\mathrm{g}}) -\hbar\omega_{\mathrm{g}}/2 =\mathcal{H}_{\mathrm{atm}}(\varepsilon) +\widetilde{H}_{\mathrm{ptn}}(\omega_{\mathrm{g}})+ \hbar\widetilde{\mathrm{g}}\sigma_{z}(a+a^{\dagger}).$$ All the eigenstates of the slightly modified Hamiltonian $\widetilde{\mathcal{H}}_{A^{2}}^{\mathrm{ren}}(\varepsilon)$ are completely same as those of the original Hamiltonian $\mathcal{H}_{A^{2}}^{\mathrm{ren}}(\varepsilon)$. Thus, we prove our desired results for the Hamiltonian $\widetilde{\mathcal{H}}_{A^{2}}^{\mathrm{ren}}(\varepsilon)$. Correspondingly, we denote the modified free Hamiltonian $\widetilde{\mathcal{H}}_{0}$ of the atom-photon system by $$\widetilde{\mathcal{H}}_{0}:=\mathcal{H}_{\mathrm{atm}}(\varepsilon) +\widetilde{H}_{\mathrm{ptn}}(\omega_{\mathrm{c}}).$$ All the eigenenergies of the atom Hamiltonian $\mathcal{H}_{\mathrm{atm}}(\varepsilon)$ are $\pm(\hbar/2)\sqrt{\omega_{\mathrm{a}}^{2}+\varepsilon^{2}}$, and thus, we have its operator norm, $\|\mathcal{H}_{\mathrm{atm}}(\varepsilon)\|_{\mathrm{op}} =(\hbar/2)\sqrt{\omega_{\mathrm{a}}^{2}+\varepsilon^{2}}$. Since we can rewrite the Hamiltonian $\widetilde{H}_{\mathrm{ptn}}(\omega_{\mathrm{c}})$ as $\widetilde{H}_{\mathrm{ptn}}(\omega_{\mathrm{c}}) =\widetilde{\mathcal{H}}_{0}-\mathcal{H}_{\mathrm{atm}}(\varepsilon)$, we have $$\begin{aligned} \|\widetilde{H}_{\mathrm{ptn}}(\omega_{\mathrm{c}})\Psi \|_{\mathbb{C}^{2}\otimes L^{2}(\mathbb{R})} &\le& \|\widetilde{\mathcal{H}}_{0}\Psi \|_{\mathbb{C}^{2}\otimes L^{2}(\mathbb{R})} +\|\mathcal{H}_{\mathrm{atm}}(\varepsilon)\|_{\mathrm{op}} \|\Psi\|_{\mathbb{C}^{2}\otimes L^{2}(\mathbb{R})} \\ &=& \|\widetilde{\mathcal{H}}_{0}\Psi \|_{\mathbb{C}^{2}\otimes L^{2}(\mathbb{R})} +\frac{\hbar}{2}\sqrt{\omega_{\mathrm{a}}^{2}+\varepsilon^{2}} \|\Psi\|_{\mathbb{C}^{2}\otimes L^{2}(\mathbb{R})} \end{aligned}$$ for every vector $\Psi\in\mathbb{C}^{2}\otimes L^{2}(\mathbb{R})$. In particular, we set $\Psi=(\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1}\Phi$ for every vector $\Phi\in\mathbb{C}^{2}\otimes L^{2}(\mathbb{R})$, and insert it into the above inequality. Then, we have the inequalities, $$\begin{aligned} &{}& \|\widetilde{H}_{\mathrm{ptn}}(\omega_{\mathrm{c}}) (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1} \Phi \|_{\mathbb{C}^{2}\otimes L^{2}(\mathbb{R})} \nonumber \\ &\le& \|\widetilde{\mathcal{H}}_{0} (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1} \Phi \|_{\mathbb{C}^{2}\otimes L^{2}(\mathbb{R})} +\frac{\hbar}{2}\sqrt{\omega_{\mathrm{a}}^{2}+\varepsilon^{2}} \| (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1} \Phi \|_{\mathbb{C}^{2}\otimes L^{2}(\mathbb{R})} \nonumber \\ &\le& \|\widetilde{\mathcal{H}}_{0} (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1}\|_{\mathrm{op}} \|\Phi \|_{\mathbb{C}^{2}\otimes L^{2}(\mathbb{R})} +\frac{\hbar}{2}\sqrt{\omega_{\mathrm{a}}^{2}+\varepsilon^{2}} \|(\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1}\|_{\mathrm{op}} \|\Phi \|_{\mathbb{C}^{2}\otimes L^{2}(\mathbb{R})}. \label{eq:proof-1_a}\end{aligned}$$ To estimate the operator norms, $\|\widetilde{\mathcal{H}}_{0} (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1}\|_{\mathrm{op}}$ and $\|(\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1}\|_{\mathrm{op}}$, we make a general argument. Let $\mathcal{H}$ be an arbitrary self-adjoint energy operator (i.e., Hamiltonian). We recall Theorem VIII.6 of Ref.[@RS1] or Theorems 7.14 and 7.17 of Ref.[@weidmann]: There is a spectral family (i.e., the set of projection-valued measures) $P_{\xi}^{\mathcal{H}}$ for the Hamiltonian $\mathcal{H}$ so that $$\mathcal{H}= \int_{-\infty}^{\infty}\xi\, dP_{\xi}^{\mathcal{H}}.$$ Using the properties of the projection-valued measures $P_{\xi}^{\mathcal{H}}$, we have the following estimates, $$\begin{aligned} \|\mathcal{H} (\mathcal{H}-i\hbar)^{-1} \Phi \|_{\mathbb{C}^{2}\otimes L^{2}(\mathbb{R})}^{2} &=& \int_{-\infty}^{\infty} \left| \frac{\xi}{\xi-i\hbar} \right|^{2} d\| P_{\xi}^{\mathcal{H}}\Phi \|_{\mathbb{C}^{2}\otimes L^{2}(\mathbb{R})}^{2} \\ &\le& \int_{-\infty}^{\infty} d\| P_{\xi}^{\mathcal{H}}\Phi \|_{\mathbb{C}^{2}\otimes L^{2}(\mathbb{R})}^{2} =\|\Phi\|_{\mathbb{C}^{2}\otimes L^{2}(\mathbb{R})}^{2}\end{aligned}$$ and $$\begin{aligned} \|(\mathcal{H}-i\hbar)^{-1} \Phi \|_{\mathbb{C}^{2}\otimes L^{2}(\mathbb{R})}^{2} &=& \int_{-\infty}^{\infty} \left| \frac{1}{\xi-i\hbar} \right|^{2} d\| P_{\xi}^{\mathcal{H}}\Phi \|_{\mathbb{C}^{2}\otimes L^{2}(\mathbb{R})}^{2} \\ &\le& \frac{1}{\hbar^{2}} \int_{-\infty}^{\infty} d\| P_{\xi}^{\mathcal{H}}\Phi \|_{\mathbb{C}^{2}\otimes L^{2}(\mathbb{R})}^{2} =\frac{1}{\hbar^{2}}\|\Phi\|_{\mathbb{C}^{2}\otimes L^{2}(\mathbb{R})}^{2}, \end{aligned}$$ for every vector $\Phi\in\mathbb{C}^{2}\otimes L^{2}(\mathbb{R})$. These estimates bring us the two operator-norm inequalities, $$\|\mathcal{H}(\mathcal{H}-i\hbar)^{-1}\|_{\mathrm{op}} \le 1\quad \mbox{and}\quad \|(\mathcal{H}-i\hbar)^{-1}\|_{\mathrm{op}} \le1/\hbar.$$ Inserting the inequalities, $\|\widetilde{\mathcal{H}}_{0} (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1}\|_{\mathrm{op}} \le 1$ and $\|(\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1}\|_{\mathrm{op}} \le1/\hbar$, into Eq.(\[eq:proof-1\_a\]), we reach the inequality $$\|\widetilde{H}_{\mathrm{ptn}}(\omega_{\mathrm{c}}) (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1} \|_{\mathrm{op}} \le 1+\frac{1}{2}\sqrt{\omega_{\mathrm{a}}^{2}+\varepsilon^{2}}. \label{eq:proof-1_b}$$ Meanwhile, since we assume Eq.(\[eq:assumption1\]), we have $$\begin{aligned} &{}& \|\widetilde{H}_{\mathrm{ptn}}\left( \omega_{\mathrm{c}} -\left(\omega_{\mathrm{g}}-\Delta_{\mathrm{g}}\right) \right) (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1} \Phi \|_{\mathbb{C}^{2}\otimes L^{2}(\mathbb{R})} \\ &=& \hbar |\omega_{\mathrm{c}} -\left(\omega_{\mathrm{g}}-\Delta_{\mathrm{g}}\right)|\, \| a^{\dagger}a (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1} \Phi \|_{\mathbb{C}^{2}\otimes L^{2}(\mathbb{R})} \\ &\le& \hbar\delta_{\mathrm{g}}\omega_{\mathrm{c}} \| a^{\dagger}a (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1} \Phi \|_{\mathbb{C}^{2}\otimes L^{2}(\mathbb{R})} \\ &=& \delta_{\mathrm{g}} \|\widetilde{H}_{\mathrm{ptn}}(\omega_{\mathrm{c}}) (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1} \Phi \|_{\mathbb{C}^{2}\otimes L^{2}(\mathbb{R})},\end{aligned}$$ which implies the operator-norm inequality, $$\|\widetilde{H}_{\mathrm{ptn}}\left( \omega_{\mathrm{c}} -\left(\omega_{\mathrm{g}}-\Delta_{\mathrm{g}}\right) \right) (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1} \|_{\mathrm{op}} \le \delta_{\mathrm{g}} \|\widetilde{H}_{\mathrm{ptn}}(\omega_{\mathrm{c}}) (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1} \|_{\mathrm{op}}. \label{eq:proof-1_c}$$ Combining Eqs.(\[eq:proof-1\_b\]) and (\[eq:proof-1\_c\]), we reach the inequality, $$\|\widetilde{H}_{\mathrm{ptn}}( \omega_{\mathrm{c}} -(\omega_{\mathrm{g}}-\Delta_{\mathrm{g}})) (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1} \|_{\mathrm{op}} \le \delta_{\mathrm{g}} \left( 1+\frac{1}{2}\sqrt{\omega_{\mathrm{a}}^{2}+\varepsilon^{2}} \right). \label{eq:proof-1}$$ Let $G$ be $\widetilde{\mathrm{g}}/\omega_{\mathrm{g}}$, i.e., $G:=\widetilde{\mathrm{g}}/\omega_{\mathrm{g}}$. Then, we immediately know that this quantity $G$ decays to the zero as the coupling strength $\mathrm{g}$ tends to the infinity, i.e., $\lim_{\mathrm{g}\to\infty}G=0$. For simplicity, we denote the displacement operators $D(\pm\widetilde{\mathrm{g}}/\omega_{\mathrm{g}})$ by $D_{\pm}(G)$, i.e., $D_{\pm}(G):= D(\pm\widetilde{\mathrm{g}}/\omega_{\mathrm{g}})$. We know the expression, $D_{\pm}(G)=e^{\pm iG\{i(a-a^{\dagger})\}}$. Since the operator $i(a-a^{\dagger})$ is self-adjoint on the domain of the photon number operator $a^{\dagger}a$, the operator $D_{\pm}(G)=e^{\pm iG\{i(a-a^{\dagger})\}}$ is a strongly continuous one-parameter unitary group by Theorem VIII.7 of Ref.[@RS1]. Namely, we have the limit, $\lim_{G\to 0}(1-D_{\pm}(G)^{2})\Psi=0$, for every vector $\Psi\in\mathbb{C}^{2}\otimes L^{2}(\mathbb{R})$. Thus, the operator $\sigma_{+}(1-D_{+}(G)^{2})+ \sigma_{-}(1-D_{-}(G)^{2})$ goes to the zero operator as $\mathrm{g}\to\infty$ in the strong operator topology, and therefore, in the weak operator topology. For details on these topologies, see §VI.1 of Ref.[@RS1]. We note that the inequality, $\|\sigma_{+}(1-D_{+}(G)^{2})+ \sigma_{-}(1-D_{-}(G)^{2})\|_{\mathrm{op}}\le 2$, holds, and that the resolvent $(\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1}$ is compact. Therefore, by applying Theorem in Appendix A of Ref.[@HMS], we obtain the limit, $$\lim_{\mathrm{g}\to\infty} \Biggl\| (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1} \Bigl\{\sigma_{+}\left(1-D_{+}(G)^{2}\right) +\sigma_{-}\left(1-D_{-}(G)^{2}\right)\Bigr\} (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1} \Biggr\|_{\mathrm{op}}=0. \label{eq:proof-2}$$ From now on, we denote the operator $\sigma_{+}D_{+}(G)^{2} +\sigma_{-}D_{-}(G)^{2}$ by $\Xi_{0}(\mathrm{g})$, and moreover, the operator $\sigma_{x}-\Xi_{0}(\mathrm{g})$ by $\Xi_{1}(\mathrm{g})$: $$\begin{aligned} \Xi_{0}(\mathrm{g})&:=& \sigma_{+}D_{+}(G)^{2}+\sigma_{-}D_{-}(G)^{2}, \\ \Xi_{1}(\mathrm{g})&:=& \sigma_{x}-\Xi_{0}(\mathrm{g}) =\sigma_{+}\left(1-D_{+}(G)^{2}\right) +\sigma_{-}\left(1-D_{-}(G)^{2}\right). \end{aligned}$$ Moreover, we define a Hamiltonian $\widetilde{\mathcal{H}}(\mathrm{g})$ by $$\widetilde{\mathcal{H}}(\mathrm{g}):= \widetilde{H}_{\mathrm{ptn}}(\omega_{\mathrm{g}}-\Delta_{\mathrm{g}}) -\frac{\hbar}{2}\varepsilon\sigma_{z} -\frac{\hbar}{2}\omega_{\mathrm{a}}\Xi_{0}(\mathrm{g}).$$ We define an operator $R$ by the difference between the resolvents of $\widetilde{\mathcal{H}}(\mathrm{g})$ and $\widetilde{\mathcal{H}}_{0}$, that is, $$R:= (\widetilde{\mathcal{H}}(\mathrm{g})-i\hbar)^{-1} - (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1}.$$ Using the $2$nd resolvent identity in Theorem 5.13(b) of Ref.[@weidmann] and the equation $\widetilde{H}_{\mathrm{ptn}}(\omega_{\mathrm{c}}) -\widetilde{H}_{\mathrm{ptn}}(\omega_{\mathrm{g}}-\Delta_{\mathrm{g}}) =\widetilde{H}_{\mathrm{ptn}}(\omega_{\mathrm{c}}- (\omega_{\mathrm{g}}-\Delta_{\mathrm{g}}))$, we can calculate the expression of the difference operator $R$ as $$\begin{aligned} R&=& (\widetilde{\mathcal{H}}(\mathrm{g})-i\hbar)^{-1} \Bigl\{ \widetilde{H}_{\mathrm{ptn}}(\omega_{\mathrm{c}}-(\omega_{\mathrm{g}}-\Delta_{\mathrm{g}})) -\frac{\hbar}{2}\omega_{\mathrm{a}}\Xi_{1}(\mathrm{g}) \Bigr\} (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1}. \label{eq:proof-a}\end{aligned}$$ Insert this into the equation $(\widetilde{\mathcal{H}}(\mathrm{g})-i\hbar)^{-1} =(\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1}+R$, then we have the equation, $$\begin{aligned} (\widetilde{\mathcal{H}}(\mathrm{g})-i\hbar)^{-1} &=& (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1} + (\widetilde{\mathcal{H}}(\mathrm{g})-i\hbar)^{-1} \widetilde{H}_{\mathrm{ptn}} (\omega_{\mathrm{c}}-(\omega_{\mathrm{g}}-\Delta_{\mathrm{g}})) (\widetilde{\mathcal{H}}_{0} -i\hbar)^{-1} \nonumber \\ &{}& -\frac{\hbar}{2}\omega_{\mathrm{a}} (\widetilde{\mathcal{H}}(\mathrm{g})-i\hbar)^{-1} \Xi_{1}(\mathrm{g}) (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1}. \label{eq:proof-b}\end{aligned}$$ Inserting Eq.(\[eq:proof-b\]) into Eq.(\[eq:proof-a\]), we have the decomposition, $$\begin{aligned} R&=& (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1} \left\{ \widetilde{H}_{\mathrm{ptn}}(\omega_{\mathrm{c}}- (\omega_{\mathrm{g}}-\Delta_{\mathrm{g}})) -\frac{\hbar\omega_{\mathrm{a}}}{2}\Xi_{1}(\mathrm{g}) \right\}(\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1} \\ &{}& +(\widetilde{\mathcal{H}}(\mathrm{g})-i\hbar)^{-1} \widetilde{H}_{\mathrm{ptn}}(\omega_{\mathrm{c}}- (\omega_{\mathrm{g}}-\Delta_{\mathrm{g}})) (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1} \\ &{}&\qquad\qquad\quad \left\{ \widetilde{H}_{\mathrm{ptn}}(\omega_{\mathrm{c}}- (\omega_{\mathrm{g}}-\Delta_{\mathrm{g}})) -\frac{\hbar\omega_{\mathrm{a}}}{2}\Xi_{1}(\mathrm{g}) \right\}(\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1} \\ &{}& -\frac{\hbar\omega_{\mathrm{a}}}{2} (\widetilde{\mathcal{H}}(\mathrm{g})-i\hbar)^{-1} \Xi_{1}(\mathrm{g}) (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1} \\ &{}&\qquad\qquad\quad \left\{ \widetilde{H}_{\mathrm{ptn}}(\omega_{\mathrm{c}}- (\omega_{\mathrm{g}}-\Delta_{\mathrm{g}})) -\frac{\hbar\omega_{\mathrm{a}}}{2}\Xi_{1}(\mathrm{g}) \right\}(\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1}. \end{aligned}$$ Eventually, we can decompose the difference operator $R$ as $$R=\sum_{j=1}^{6}I_{j},\,\,\, \mbox{and thus},\,\,\, \|R\|_{\mathrm{op}}=\sum_{j=1}^{6}\|I_{j}\|_{\mathrm{op}}, \label{eq:proof-c}$$ where $$\begin{aligned} I_{1}&=& (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1} \widetilde{H}_{\mathrm{ptn}}( \omega_{\mathrm{c}}-(\omega_{\mathrm{g}}-\Delta_{\mathrm{g}})) (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1}, \\ I_{2}&=& -\frac{\hbar\omega_{\mathrm{a}}}{2} (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1} \Xi_{1}(\mathrm{g}) (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1}, \\ I_{3}&=& (\widetilde{\mathcal{H}}(\mathrm{g})-i\hbar)^{-1} \Bigl\{ \widetilde{H}_{\mathrm{ptn}}(\omega_{\mathrm{c}}-(\omega_{\mathrm{g}}-\Delta_{\mathrm{g}})) (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1} \Bigr\}^{2}, \\ I_{4}&=& -\frac{\hbar\omega_{\mathrm{a}}}{2} (\widetilde{\mathcal{H}}(\mathrm{g})-i\hbar)^{-1} \widetilde{H}_{\mathrm{ptn}}(\omega_{\mathrm{c}}-(\omega_{\mathrm{g}}-\Delta_{\mathrm{g}})) (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1} \Xi_{1}(\mathrm{g}) (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1}, \\ I_{5}&=& -\frac{\hbar\omega_{\mathrm{a}}}{2} (\widetilde{\mathcal{H}}(\mathrm{g})-i\hbar)^{-1} \Xi_{1}(\mathrm{g}) (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1} \widetilde{H}_{\mathrm{ptn}}(\omega_{\mathrm{c}}-(\omega_{\mathrm{g}}-\Delta_{\mathrm{g}})) (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1}, \\ I_{6}&=& \left(-\, \frac{\hbar\omega_{\mathrm{a}}}{2}\right)^{2} (\widetilde{\mathcal{H}}(\mathrm{g})-i\hbar)^{-1} \Xi_{1}(\mathrm{g}) (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1} \Xi_{1}(\mathrm{g}) (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1}. \end{aligned}$$ Using Eq.(\[eq:proof-1\]) and the inequalities, $\|\widetilde{\mathcal{H}}_{0}-i\hbar\|_{\mathrm{op}} \le 1/\hbar$, $\|\widetilde{\mathcal{H}}(\mathrm{g})-i\hbar\|_{\mathrm{op}} \le 1/\hbar$, $\|\Xi_{1}(\mathrm{g})\|_{\mathrm{op}}\le 2$, individual operators $I_{j}$ are bounded from above in the following: $$\begin{aligned} \|I_{1}\|_{\mathrm{op}}&\le& \|(\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1}\|_{\mathrm{op}} \|\widetilde{H}_{\mathrm{ptn}}( \omega_{\mathrm{c}}-(\omega_{\mathrm{g}}-\Delta_{\mathrm{g}})) (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1} \|_{\mathrm{op}} \\ &\le& \frac{\delta_{\mathrm{g}}}{\hbar} \left( 1+\frac{1}{2}\sqrt{\omega_{\mathrm{a}}^{2}+\varepsilon^{2}} \right), \end{aligned}$$ $$\begin{aligned} \|I_{2}\|_{\mathrm{op}} &\le& \frac{\hbar\omega_{\mathrm{a}}}{2} \|(\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1} \Xi_{1}(\mathrm{g}) (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1} \|_{\mathrm{op}}, \end{aligned}$$ $$\begin{aligned} \|I_{3}\|_{\mathrm{op}} &\le& \|(\widetilde{\mathcal{H}}(\mathrm{g})-i\hbar)^{-1} \|_{\mathrm{op}} \| \widetilde{H}_{\mathrm{ptn}}(\omega_{\mathrm{c}}-(\omega_{\mathrm{g}}-\Delta_{\mathrm{g}})) (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1} \|_{\mathrm{op}}^{2} \\ &\le& \frac{\delta_{\mathrm{g}}^{2}}{\hbar} \left( 1+\frac{1}{2}\sqrt{\omega_{\mathrm{a}}^{2}+\varepsilon^{2}} \right)^{2}, \end{aligned}$$ $$\begin{aligned} \|I_{4}\|_{\mathrm{op}} &\le& \frac{\hbar}{2}\omega_{\mathrm{a}} \|(\widetilde{\mathcal{H}}(\mathrm{g})-i\hbar)^{-1}\|_{\mathrm{op}} \|\widetilde{H}_{\mathrm{ptn}}(\omega_{\mathrm{c}} -(\omega_{\mathrm{g}}-\Delta_{\mathrm{g}})) (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1}\|_{\mathrm{op}} \\ &{}&\qquad\qquad\times \|\Xi_{1}(\mathrm{g})\|_{\mathrm{op}} \|(\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1}\|_{\mathrm{op}} \\ &\le& \frac{\omega_{\mathrm{a}}\delta_{\mathrm{g}}}{\hbar} \left( 1+\frac{1}{2}\sqrt{\omega_{\mathrm{a}}^{2}+\varepsilon^{2}} \right), \\ \|I_{5}\|_{\mathrm{op}} &\le& \frac{\hbar}{2}\omega_{\mathrm{a}} \|(\widetilde{\mathcal{H}}(\mathrm{g})-i\hbar)^{-1}\|_{\mathrm{op}} \|\Xi_{1}(\mathrm{g})\|_{\mathrm{op}} \|(\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1}\|_{\mathrm{op}} \\ &{}&\qquad\qquad\times \|\widetilde{H}_{\mathrm{ptn}}(\omega_{\mathrm{c}} -(\omega_{\mathrm{g}}-\Delta_{\mathrm{g}})) (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1}\|_{\mathrm{op}} \\ &\le& \frac{\omega_{\mathrm{a}}\delta_{\mathrm{g}}}{\hbar} \left( 1+\frac{1}{2}\sqrt{\omega_{\mathrm{a}}^{2}+\varepsilon^{2}} \right), \\ \|I_{6}\|_{\mathrm{op}} &\le& \frac{\hbar^{2}\omega_{\mathrm{a}}^{2}}{4} \|(\widetilde{\mathcal{H}}(\mathrm{g})-i\hbar)^{-1}\|_{\mathrm{op}} \|\Xi_{1}(\mathrm{g})\|_{\mathrm{op}} \|(\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1} \Xi_{1}(\mathrm{g}) (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1}\|_{\mathrm{op}} \\ &\le& \frac{\hbar\omega_{\mathrm{a}}^{2}}{2} \|(\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1} \Xi_{1}(\mathrm{g}) (\widetilde{\mathcal{H}}_{0}-i\hbar)^{-1} \|_{\mathrm{op}}. \end{aligned}$$ Eqs.(\[eq:assumption2\]) and (\[eq:proof-2\]) tell us that all the operator norms of operators $I_{j}$ converges to zero as taking strong coupling limit, i.e., $\lim_{\mathrm{g}\to\infty}\| I_{j}\|_{\mathrm{op}}=0$. Therefore, by Eq.(\[eq:proof-c\]), we obtain our desired limit, $\lim_{\mathrm{g}\to\infty}\| R\|_{\mathrm{op}}=0$. Namely, we succeed in proving the convergence, $$\widetilde{\mathcal{H}}(\mathrm{g}) \stackrel{\mathrm{n.r.s.}\,\,}{{-\!\!-\!\!\!}\longrightarrow} \widetilde{\mathcal{H}}_{0}\,\,\, \mbox{as $\mathrm{g}\to\infty$}. \label{eq:proof-3}$$ We recall the unitary transformation given in Eq.(\[eq:unitary-trans-GQR-Hamiltonian\]), and replace the parameters $\omega_{\mathrm{c}}$ and $\mathrm{g}$ with the parameters $\omega_{\mathrm{g}}$ and $\widetilde{\mathrm{g}}$ in it, respectively. This comes up with the identity, $$\begin{aligned} &{}& U(\widetilde{\mathrm{g}}/\omega_{\mathrm{g}}) \mathcal{H}_{\mbox{\tiny GQR}}(\omega_{\mathrm{g}},\widetilde{\mathrm{g}}) U(\widetilde{\mathrm{g}}/\omega_{\mathrm{g}})^{*} -\widetilde{H}_{\mathrm{ptn}}(\Delta_{\mathrm{g}}) -\hbar\frac{\omega_{\mathrm{g}}}{2} \nonumber \\ &=& H_{\mathrm{ptn}}(\omega_{\mathrm{g}}) -\frac{\hbar}{2}\varepsilon\sigma_{z} -\frac{\hbar}{2}\omega_{\mathrm{a}}\Xi_{0}(\mathrm{g}) -\widetilde{H}_{\mathrm{ptn}}(\Delta_{\mathrm{g}}) -\hbar\frac{\omega_{\mathrm{g}}}{2} \nonumber \\ &=& \widetilde{H}_{\mathrm{ptn}}(\omega_{\mathrm{g}}) -\frac{\hbar}{2}\varepsilon\sigma_{z} -\frac{\hbar}{2}\omega_{\mathrm{a}}\Xi_{0}(\mathrm{g}) -\widetilde{H}_{\mathrm{ptn}}(\Delta_{\mathrm{g}}) \nonumber \\ &=& \widetilde{\mathcal{H}}(\mathrm{g}) -\hbar\frac{\widetilde{\mathrm{g}}^{2}}{\omega_{\mathrm{g}}}. \label{eq:proof-4}\end{aligned}$$ Applying Eq.(\[eq:proof-3\]) to the Hamiltonian $\widetilde{\mathcal{H}}(\mathrm{g})$ in RHS of Eq.(\[eq:proof-4\]) yields the limit $$U(\widetilde{\mathrm{g}}/\omega_{\mathrm{g}}) \mathcal{H}_{\mbox{\tiny GQR}}(\omega_{\mathrm{g}},\widetilde{\mathrm{g}}) U(\widetilde{\mathrm{g}}/\omega_{\mathrm{g}})^{*} -\widetilde{H}_{\mathrm{ptn}}(\Delta_{\mathrm{g}}) -\hbar\frac{\omega_{\mathrm{g}}}{2} \stackrel{\mathrm{n.r.s.}\,\,}{{-\!\!-\!\!\!}\longrightarrow} \widetilde{\mathcal{H}}_{0}-\frac{\hbar}{4C_{\infty}}\,\,\, \mbox{as $\mathrm{g}\to\infty$}. \label{eq:proof-5}$$ Since $\mathcal{H}^{\mathrm{ren}}_{A^{2}}(\varepsilon) =\mathcal{H}_{\mbox{\tiny GQR}}(\omega_{\mathrm{g}},\widetilde{\mathrm{g}})$ by Eq.(\[eq:renormalized-Hamiltonian\]), the renormalized Hamiltonian $\mathcal{H}^{\mathrm{ren}}_{A^{2}}(\varepsilon)$ can be well approximated by $$\begin{aligned} &{}& U(\widetilde{\mathrm{g}}/\omega_{\mathrm{g}})^{*} \left\{ \widetilde{\mathcal{H}}_{0} +\widetilde{H}_{\mathrm{ptn}}(\Delta_{\mathrm{g}}) +\hbar\frac{\omega_{\mathrm{g}}}{2} \right\}U(\widetilde{\mathrm{g}}/\omega_{\mathrm{g}}) -\frac{\hbar}{4C_{\infty}} \\ &=& U(\widetilde{\mathrm{g}}/\omega_{\mathrm{g}})^{*} \left\{\mathcal{H}_{\mathrm{atm}}(\varepsilon) +\widetilde{H}_{\mathrm{ptn}}(\omega_{\mathrm{c}}+\Delta_{\mathrm{g}}) +\hbar\frac{\omega_{\mathrm{g}}}{2} \right\}U(\widetilde{\mathrm{g}}/\omega_{\mathrm{g}}) -\frac{\hbar}{4C_{\infty}} \\ &\approx& U(\widetilde{\mathrm{g}}/\omega_{\mathrm{g}})^{*} \Bigl\{ \mathcal{H}_{\mathrm{atm}}(\varepsilon) +\widetilde{H}_{\mathrm{ptn}}(\omega_{\mathrm{g}}) +\hbar\frac{\omega_{\mathrm{g}}}{2} \Bigr\}U(\widetilde{\mathrm{g}}/\omega_{\mathrm{g}}) -\hbar\frac{\widetilde{\mathrm{g}}^{2}}{\omega_{\mathrm{g}}} \\ &=& U(\widetilde{\mathrm{g}}/\omega_{\mathrm{g}})^{*} \Bigl\{ \mathcal{H}_{\mathrm{atm}}(\varepsilon) +H_{\mathrm{ptn}}(\omega_{\mathrm{g}}) \Bigr\}U(\widetilde{\mathrm{g}}/\omega_{\mathrm{g}}) -\hbar\frac{\widetilde{\mathrm{g}}^{2}}{\omega_{\mathrm{g}}}, \end{aligned}$$ that is, by the limit Hamiltonian as in Eq.(\[eq:main-result\]). Here, we used $\widetilde{H}_{\mathrm{ptn}}(\omega_{\mathrm{c}}) +\widetilde{H}_{\mathrm{ptn}}(\Delta_{\mathrm{g}}) =\widetilde{H}_{\mathrm{ptn}}(\omega_{\mathrm{c}}+\Delta_{\mathrm{g}})$, and approximations, $\omega_{\mathrm{g}}\approx\omega_{\mathrm{c}}+\Delta_{\mathrm{g}}$ and $\hbar/4C_{\infty}\approx \hbar\widetilde{\mathrm{g}}^{2}/\omega_{\mathrm{g}}$, respectively secured by Eqs.(\[eq:app\_omega-g\]) and (\[eq:limit\_of\_self-energy\]). Conclusion and Discussion {#sec:CD} ========================= We have considered a mathematical establishment of the adiabatic approximation for the generalized quantum Rabi Hamiltonian both without and with the $A^{2}$-term. In the case without the $A^{2}$-term, we have shown in the adiabatic approximation that whether each bare eigenstate forms a Schrödinger-cat-like entangled state or not depends on whether the energy bias in the atom Hamiltonian is zero or non-zero. On the other hand, in the case with the $A^{2}$-term, we have renormalized the $A^{2}$-term by employing (meson) pair theory, and then, we mathematically established the adiabatic approximation for the renormalized Hamiltonian. Moreover, we have shown in the adiabatic approximation that the Schrödinger-cat-likeness appears in the both cases where the energy bias is zero and where it is non-zero. At the end of this section, we explain the reason why we take the interest in Eqs.(\[eq:equivalence1\]) and (\[eq:equivalence2\]). We showed in Ref.[@HMS] that if the $A^{2}$-term effect is sufficiently small, then the renormalized Hamiltonian of the generalized quantum Rabi model has the chance to have some dressed photons (real photons) in the ground state. Based on the adiabatic approximation given by Eqs.(\[eq:approximation\_1\]) and (\[eq:approximation\_2\]), the approximated ground-state expectation $N_{0}^{\mathrm{app}}$ can be calculated as $N_{0}^{\mathrm{app}}=\widetilde{\mathrm{g}}^{2}/\omega_{\mathrm{g}}^{2}$ as in Eq.(\[eq:adiabatic-approximation\_N\_0\]). Therefore, each of the adiabatically approximated eigenstates has the expression as $$|\mathcal{E}_{n}^{\mathrm{app},\pm}(0)\rangle =\frac{1}{\sqrt{2}} \Bigg(|\!\!\uparrow\rangle D\left(-\sqrt{N_{0}^{\mathrm{app}}}\right) |n\rangle \pm |\!\!\downarrow\rangle D\left(+\sqrt{N_{0}^{\mathrm{app}}}\right) |n\rangle \Biggr)$$ for $\varepsilon=0$ and $$|\mathcal{E}_{n}^{\mathrm{app},\pm}(\varepsilon)\rangle =c_{\pm\varepsilon,\omega_{\mathrm{a}}} \Biggl( -\omega_{\mathrm{a}}|\!\!\uparrow\rangle D(-\sqrt{N_{0}^{\mathrm{app}}})|n\rangle +(\varepsilon\mp\sqrt{\varepsilon^{2}+\omega_{\mathrm{a}}^{2}}) |\!\!\downarrow\rangle D(+\sqrt{N_{0}^{\mathrm{app}}})|n\rangle \Biggr)$$ for $\varepsilon\ne 0$. We note that the approximated ground-state expectation $N_{0}^{\mathrm{app}}$ can be expressed as $N_{0}^{\mathrm{app}} = \langle\mathcal{E}_{0}^{\mathrm{app},-}(\varepsilon)| a^{\dagger}a|\mathcal{E}_{0}^{\mathrm{app},-}(\varepsilon)\rangle$, and that the state $|\mathcal{E}_{0}^{\mathrm{app},-}(\varepsilon)\rangle$ is the $1$st excited state for sufficiently small $|\varepsilon|$ by Eq.(\[eq:energy-difference\]). Therefore, whether the Schrödinger-cat-likeness for each eigenstate can be observed depends on the number of dressed photons in the ground state or the $1$st excited state. In a sense, namely, the ‘size’ of the Schrödinger-cat-likeness is determined by the number of dressed photons in the ground state. This is the reason why we are interested in Eqs.(\[eq:equivalence1\]) and (\[eq:equivalence2\]). **Acknowledgement** This work was supported by JSPS KAKENHI Grant Number JP26310210. The author thanks Tomoko Fuse, Kouichi Semba, and Fumiki Yoshihara for the valuable discussions on their experimental results, and Qing-Hu Chen, Cristiano Ciuti, and Elinor Irish for their useful comments at the International Workshop on “Strongly Coupled Light-Matter Interactions: Models and Application” organized by Hong-Gang Luo and Jun-Hong An at Lanzhou University on July 8 and 9, 2018. He also acknowledges Franco Nori’s advice on reference literatures. [11]{} Povh B, Rith K, Scholz C and Zetsche F 2004 [*Particles and Nuclei: An Introduction to the Physical Concepts*]{} (Berlin: Springer) Feynman R 1985 [*QED: The Strange Theory of Light and Matter*]{} (Princeton: Princeton University Press) Bjorken J D and Drell S D 1964 [*Relativistic Quantum Mechanics*]{} (New York: McGraw-Hill Book Company) Bjorken J D and Drell S D 1965 [*Relativistic Quantum Fields*]{} (New York: McGraw-Hill Book Company) Cohen-Tannoudji C, Dupont-Roc J and Grynberg G 1992 [*Atom-Photon Interaction: Basic Processes and Applications*]{} (New York: John Willey & Sons, Inc) Kaku M 1993 [*Quantum Field Theory: A Modern Introduction*]{} (New York: Oxford University Press) Peskin M E and Schroeder D V 1995 [*An Introduction to Quantum Field Theory*]{} (Abingdon: Westview Press) Nation P D, Johansson J R, Blencowe M P and Nori F 2012 Rev. Mod. Phys. **84**, 1 Henley E and Thirring W 1962 [*Elementary Quantum Field Theory*]{} (New York: McGraw-Hill) Yukawa H 1935 [*Proc. Phys.-Math. Soc. Japan*]{} [**17**]{} 48 Wentzel G 1941 [*Z. Phys.*]{} [**118**]{} 277 Wentzel G 1942 [*Helv. Phys. Acta*]{} [**15**]{} 111 Brown L M and Rechenberg H 1996 [*The Origin of the Concept of Nuclear Forces*]{} (Bristol: IOP Publishing) Buluta I and Nori F 2009 [*Science*]{} **326**, 108 Haroche S and Raimond J M 2008, [*Exploring Quantum. Atoms, Cavities, and Photons*]{} (Oxford: Oxford University Press) 2008, Raimond J M, Brune M and Harohe S 2001 [*Rev. Mod. Phys.*]{} **73** 565 Makhilin Yu, Schön G and Shnirman A 2001 [*Rev. Mod. Phys.*]{} **73** 357 Marquardt F and Bruder C 2001 [*Phys. Rev. B*]{} **63** 054514 Chiorescu I, Bertet P, Semba K, Nakamura Y, Harmans C J P M and Mooij J E 2004 [*Nature*]{} **431** 159–162 Fink J M, Göppl M, Bau M, Bianchetti R, Leek P J, Blais A and Wallraff A, 2008 [*Nature*]{} **454** 315 Wallraff A, Schuster D I, Blais A, Fruzio L, Huang R -S, Majer J, Kuar S, Girvin S M and Schoelkopf R J 2004 [*Nature*]{} **431** 162 You J Q and Nori F 2011 [*Nature*]{} **474**, 589 Gu X, Kockum A F, Miranowicz A, Liu Y X and Nori F 2017 [*Phys. Rep.*]{} **718-719**, 1 Devoret M, Girvin S and Schoelkopf R 2007 [*Ann. Phys.*]{} **16** 767 Günter G, Anappara A A, Hees J, Sell A, Biasiol G, Sorba L, De Liberato S, Ciuiti C, Tredicucci A, Leitenstorfer A and Huber R 2009 [*Nature*]{} **458** 178 Forn-Díaz P, Lisenfeld J, Marcos D, García-Ripoll J J, Solano E, Harmans C J P M and Mooij J E 2010 [*Phys. Rev. Lett.*]{} **105** 237001 Niemczyk T, Deppe F, Huebl H, Menzel E P, Hocke F, Schwarz M J, Garcia-Ripoll J J, Zueco D, Hümmer T, Solano E, Marx A and Gross R 2010 [*Nature Physics*]{} **6** 772 Forn-Díaz P, J. J. García-Ripoll J J, Peropadre B, Orgiazzi J -L, Yurtalan M A, Belyansky R, Wilson C M and Lupascu A 2017 [*Nature Phys.*]{} [**13**]{} 39 Yoshihara F, Fuse T, Ashhab S, Kakuyanagi K, Saito S and Semba K 2017 [*Nature Phys.*]{} [**13**]{} 44 Yoshihara F, Fuse T, Ashhab S, Kakuyanagi K, Saito S and Semba K 2017 [*Phys. Rev. A*]{} [**95**]{} 053824 Yoshihara F, Fuse T, Ao Z, Ashhab S, Kakuyanagi K, Saito S, Aoki T, Koshino K and Semba K 2018 [*Phys. Rev. Lett.*]{} [**120**]{} 183601 Casanova J, Romero G, Lizuain I, Garćia-Ripoll J J and Solano E 2010 [*Phys. Rev. Lett.*]{} [**105**]{} 263603 Braak D 2011 [*Phys. Rev. Lett.*]{} [**107**]{} 100401 Ashhab S and Nori F 2010 [*Phys. Rev. A*]{} **81** 042311 Irish E K 2007 [*Phys. Rev. Lett.*]{} **99** 173601 Xiao Z, Fuse T, Ashhab S, Yoshihara F, Semba K, Sasaki M, Takeoka M and Dowling J P, 2018 arXiv:1807.04927 Hirokawa M, M[ø]{}ller J S and Sasaki I 2017 [*J. Phys A: Math. Theo.*]{} [**50**]{} 184003 Reed M and Simon B 1980 [*Method of Modern Mathematical Physics I. Functional Analysis*]{} (San Diego: Academic Press) Weidmann J 1980 [*Linear Operators in Hilbert Spaces*]{} (New York: Springer-Verlag) Hirokawa M and Hiroshima F 2014 [*Comm. Stoch. Anal.*]{} [**8**]{} 551 Hirokawa M 2015 [*Quantum Studies: Math. Found.*]{} [**2**]{} 379 van Hove L 1952 [*Physica*]{} [**18**]{} 145 Klein A and McCormick B H 1955 [*Phys. Rev.*]{} [**98**]{} 1428

ArXiv

--- abstract: 'In this paper we propose a novel model for unconditional audio generation based on generating one audio sample at a time. We show that our model, which profits from combining memory-less modules, namely autoregressive multilayer perceptrons, and stateful recurrent neural networks in a hierarchical structure is able to capture underlying sources of variations in the temporal sequences over very long time spans, on three datasets of different nature. Human evaluation on the generated samples indicate that our model is preferred over competing models. We also show how each component of the model contributes to the exhibited performance.' author: - | Soroush Mehri\ University of Montreal\ Kundan Kumar\ IIT Kanpur\ Ishaan Gulrajani\ University of Montreal\ Rithesh Kumar\ SSNCE\ Shubham Jain\ IIT Kanpur\ Jose Sotelo\ University of Montreal\ Aaron Courville\ University of Montreal\ CIFAR Fellow\ Yoshua Bengio\ University of Montreal\ CIFAR Senior Fellow\ bibliography: - 'iclr2017\_conference.bib' title: 'SampleRNN: An Unconditional End-to-End Neural Audio Generation Model' --- =1 Introduction ============ Audio generation is a challenging task at the core of many problems of interest, such as text-to-speech synthesis, music synthesis and voice conversion. The particular difficulty of audio generation is that there is often a very large discrepancy between the dimensionality of the the raw audio signal and that of the effective semantic-level signal. Consider the task of speech synthesis, where we are typically interested in generating utterances corresponding to full sentences. Even at a relatively low sample rate of 16kHz, on average we will have 6,000 samples per word generated. [^1] Traditionally, the high-dimensionality of raw audio signal is dealt with by first compressing it into spectral or hand-engineered features and defining the generative model over these features. However, when the generated signal is eventually decompressed into audio waveforms, the sample quality is often degraded and requires extensive domain-expert corrective measures. This results in complicated signal processing pipelines that are to adapt to new tasks or domains. Here we propose a step in the direction of replacing these handcrafted systems. In this work, we investigate the use of recurrent neural networks (RNNs) to model the dependencies in audio data. We believe RNNs are well suited as they have been designed and are suited solutions for these tasks (see @graves2013generating, @karpathy2015unreasonable, and @siegelmann1999computation). However, in practice it is a known problem of these models to not scale well at such a high temporal resolution as is found when generating acoustic signals one sample at a time, e.g., 16000 times per second. This is one of the reasons that @oord2016wavenet profits from other neural modules such as one presented by @yu2015multi to show extremely good performance. In this paper, an end-to-end unconditional audio synthesis model for raw waveforms is presented while keeping all the computations tractable.[^2] Since our model has different modules operating at different clock-rates (which is in contrast to WaveNet), we have the flexibility in allocating the amount of computational resources in modeling different levels of abstraction. In particular, we can potentially allocate very limited resource to the module responsible for sample level alignments operating at the clock-rate equivalent to sample-rate of the audio, while allocating more resources in modeling dependencies which vary very slowly in audio, for example identity of phoneme being spoken. This advantage makes our model arbitrarily flexible in handling sequential dependencies at multiple levels of abstraction. Hence, our contribution is threefold: 1. We present a novel method that utilizes RNNs at different scales to model longer term dependencies in audio waveforms while training on short sequences which results in memory efficiency during training. 2. We extensively explore and compare variants of models achieving the above effect. 3. We study and empirically evaluate the impact of different components of our model on three audio datasets. Human evaluation also has been conducted to test these generative models. SampleRNN Model =============== In this paper we propose SampleRNN (shown in Fig. \[fig:diagram\]), a density model for audio waveforms. SampleRNN models the probability of a sequence of waveform samples $X=\{x_1, x_2, \ldots, x_T\}$ (a random variable over input data sequences) as the product of the probabilities of each sample conditioned on all previous samples: $$\begin{aligned} \label{eq:px} p(X) = \prod_{i=0}^{T-1} p(x_{i+1}|x_1, \ldots , x_{i})\end{aligned}$$ RNNs are commonly used to model sequential data which can be formulated as: $$\begin{aligned} h_t &= \mathcal{H}(h_{t-1},x_{i=t})\label{eq:rnn}\\ p(x_{i+1}|x_1,\ldots,x_i) &= Softmax(MLP(h_t))\end{aligned}$$ with $\mathcal{H}$ being one of the known memory cells, Gated Recurrent Units (GRUs) [@chung2014empirical], Long Short Term Memory Units (LSTMs) [@hochreiter1997long], or their deep variations (Section \[sec:experiments\]). However, raw audio signals are challenging to model because they contain structure at very different scales: correlations exist between neighboring samples as well as between ones thousands of samples apart. SampleRNN helps to address this challenge by using a hierarchy of modules, each operating at a different temporal resolution. The lowest module processes individual samples, and each higher module operates on an increasingly longer timescale and a lower temporal resolution. Each module conditions the module below it, with the lowest module outputting sample-level predictions. The entire hierarchy is trained jointly end-to-end by backpropagation. ![Snapshot of the unrolled model at timestep $i$ with $K=3$ tiers. As a simplification only one RNN and up-sampling ratio $r=4$ is used for all tiers.[]{data-label="fig:diagram"}](figures/figure_diagram.pdf){width="\textwidth" height="\textheight"} Frame-level Modules {#sec:frame} ------------------- Rather than operating on individual samples, the higher-level modules in SampleRNN operate on *non-overlapping frames* of $FS^{(k)}$ (“Frame Size”) samples at the $k^\textrm{th}$ level up in the hierarchy at a time (frames denoted by $f^{(k)}$). Each frame-level module is a deep RNN which summarizes the history of its inputs into a conditioning vector for the next module downward. The variable number of frames we condition upon up to timestep $t-1$ is expressed by a fixed length hidden state or memory $h^{(k)}_t$ where $t$ is related to clock rate at that tier. The RNN makes a memory update at timestep $t$ as a function of the previous memory $h^{(k)}_{t-1}$ and an input $inp^{(k)}_t$. This input for top tier $k=K$ is simply the input frame. For intermediate tiers ($1<k<K$) this input is a linear combination of conditioning vector from higher tier and current input frame. See Eqs. \[eq:FLMbegin\]–\[eq:FLMend\]. Because different modules operate at different temporal resolutions, we need to upsample each vector $c$ at the output of a module into a series of $r^{(k)}$ vectors (where $r^{(k)}$ is the ratio between the temporal resolutions of the modules) before feeding it into the input of the next module downward (Eq. \[eq:upsampling\]). We do this with a set of $r^{(k)}$ separate linear projections. Here we are formalizing the frame-level module in tier $k$. Note that following equations are exclusive to tier $k$ and timestep $t$ for that specific tier. To increase the readability, unless necessary superscript ${(k)}$ is not shown for $t$, $inp^{(k)}$, $W_x^{(k)}$, $h^{(k)}$, $\mathcal{H}^{(k)}$, $W_j^{(k)}$, and $r^{(k)}$. $$\begin{aligned} inp_t &= \begin{cases}\label{eq:FLMbegin} W_x f_t^{(k)}+c_t^{(k+1)}; &1<k<K\\ f_t^{(k=K)}; &k=K \end{cases}\\ h_t &= \mathcal{H}(h_{t-1}, inp_t)\label{eq:FLMend}\\ c_{(t-1)*r + j}^{(k)} &= W_j h_t;\qquad 1\leq j \leq r\label{eq:upsampling}\end{aligned}$$ Our approach of upsampling with $r^{(k)}$ linear projections is exactly equivalent to upsampling by adding zeros and then applying a linear convolution. This is sometimes called “perforated” upsampling in the context of convolutional neural networks (CNNs). It was first demonstrated to work well in @dosovitskiy2016learning and is a fairly common upsampling technique. Sample-level Module ------------------- The lowest module (tier $k=1$; Eqs. \[eq:k1begin\]–\[eq:k1end\]) in the SampleRNN hierarchy outputs a distribution over a sample $x_{i+1}$, conditioned on the $FS^{(1)}$ *preceding samples* as well as a vector $c_i^{(k=2)}$ from the next higher module which encodes information about the sequence prior to that frame. As $FS^{(1)}$ is usually a small value and correlations in nearby samples are easy to model by a simple memoryless module, we implement it with a multilayer perceptron (MLP) rather than RNN which slightly speeds up the training. Assuming $e_i$ represents $x_i$ after passing through embedding layer (section \[sec:quantization\]), conditional distribution in Eq. \[eq:px\] can be achieved by following and for further clarity two consecutive sample-level frames are shown. In addition, $W_x$ in Eq. \[eq:k1combine\] is simply used to linearly combine a frame and conditioning vector from above. $$\begin{aligned} f_{i-1}^{(1)} &= flatten([e_{i-FS^{(1)}}, \ldots, e_{i-1}])\label{eq:k1begin}\\ \nonumber f_i^{(1)} &= flatten([e_{i-FS^{(1)}+1}, \ldots, e_i])\\ inp_i^{(1)} &= W_x^{(1)} f_i^{(1)} + c_i^{(2)}\label{eq:k1combine}\\ p(x_{i+1}|x_1, \ldots, x_i) &= Softmax(MLP(inp_i^{(1)}))\label{eq:k1end}\end{aligned}$$ We use a Softmax because we found that better results were obtained by discretizing the audio signals (also see [@van2016pixel]) and outputting a Multinoulli distribution rather than using a Gaussian or Gaussian mixture to represent the conditional density of the original real-valued signal. When processing an audio sequence, the MLP is convolved over the sequence, processing each window of $FS^{(1)}$ samples and predicting the next sample. At generation time, the MLP is run repeatedly to generate one sample at a time. Table \[table:summary\_results\] shows a considerable gap between the baseline model RNN and this model, suggesting that the proposed hierarchically structured architecture of SampleRNN makes a big difference. ### Output Quantization {#sec:quantization} The sample-level module models its output as a $q$-way discrete distribution over possible quantized values of $x_i$ (that is, the output layer of the MLP is a $q$-way Softmax). To demonstrate the importance of a discrete output distribution, we apply the same architecture on real-valued data by replacing the $q$-way Softmax with a Gaussian Mixture Models (GMM) output distribution. Table \[table:summary\_results\_gmm\] shows that our model outperforms an RNN baseline even when both models use real-valued outputs. However, samples from the real-valued model are almost indistinguishable from random noise. In this work we use linear quantization with $q=256$, corresponding to a per-sample bit depth of 8. Unintuitively, we realized that even linearly decreasing the bit depth (resolution of each audio sample) from 16 to 8 can ease the optimization procedure while generated samples still have reasonable quality and are artifact-free. In addition, early on we noticed that the model can achieve better performance and generation quality when we [*embed the quantized input values*]{} before passing them through the sample-level MLP (see Table \[table:variants\_table\]). The embedding steps maps each of the $q$ discrete values to a real-valued vector embedding. However, real-valued raw samples are still used as input to the higher modules. ### Conditionally Independent Sample Outputs To demonstrate the importance of a sample-level autoregressive module, we try replacing it with “Multi-Softmax” (see Table \[table:variants\_table\]), where the prediction of each sample $x_i$ depends only on the conditioning vector $c$ from Eq. \[eq:k1end\]. In this configuration, the model outputs an entire *frame* of $FS^{(1)}$ samples at a time, modeling all samples in a frame as conditionally independent of each other. We find that this Multi-Softmax model (which lacks a sample-level autoregressive module) scores significantly worse in terms of log-likelihood and fails to generate convincing samples. This suggests that modeling the joint distribution of the acoustic samples inside each frame is very important in order to obtain good acoustic generation. We found this to be true even when the frame size is reduced, with best results always with a frame size of 1, i.e., generating only one acoustic sample at a time. Truncated BPTT -------------- Training recurrent neural networks on long sequences can be very computationally expensive. @oord2016wavenet avoid this problem by using a stack of dilated convolutions instead of any recurrent connections. However, when they can be trained efficiently, recurrent networks have been shown to be very powerful and expressive sequence models. We enable efficient training of our recurrent model using *truncated backpropagation through time*, splitting each sequence into short subsequences and propagating gradients only to the beginning of each subsequence. We experiment with different subsequence lengths and demonstrate that we are able to train our networks, which model very long-term dependencies, despite backpropagating through relatively short subsequences. Table \[table:seqlen\] shows that by increasing the subsequence length, performance substantially increases alongside with train-time memory usage and convergence time. Yet it is noteworthy that our best models have been trained on subsequences of length 512, which corresponds to 32 milliseconds, a small fraction of the length of a single a phoneme of human speech while generated samples exhibit longer word-like structures. Despite the aforementioned fact, this generative model can mimic the existing long-term structure of the data which results in more natural and coherent samples that is preferred by human listeners. (More on this in Sections \[sec:humaneval\]–\[sec:timehorizon\].) This is due to the fast updates from TBPTT and specialized frame-level modules (Section \[sec:frame\]) with top tiers designed to model a lower resolution of signal while leaving the process of filling the details to lower tiers. Experiments and Results {#sec:experiments} ======================= In this section we are introducing three datasets which have been chosen to evaluate the proposed architecture for modeling raw acoustic sequences. The description of each dataset and their preprocessing is as follows: - **Blizzard** which is a dataset presented by @prahallad2013blizzard for speech synthesis task, contains 315 hours of a single female voice actor in English; however, for our experiments we are using only 20.5 hours. The training/validation/test split is 86%-7%-7%. - **Onomatopoeia**[^3], a relatively small dataset with 6,738 sequences adding up to 3.5 hours, is human vocal sounds like grunting, screaming, panting, heavy breathing, and coughing. Diversity of sound type and the fact that these sounds were recorded from 51 actors and many categories makes it a challenging task. To add to that, this data is extremely unbalanced. The training/validation/test split is 92%-4%-4%. - **Music** dataset is the collection of all 32 Beethoven’s piano sonatas publicly available on <https://archive.org/> amounting to 10 hours of non-vocal audio. The training/validation/test split is 88%-6%-6%. See Fig. \[fig:samples\] for a visual demonstration of examples from datasets and generated samples. For all the datasets we are using a 16 kHz sample rate and 16 bit depth. For the Blizzard and Music datasets, preprocessing simply amounts to chunking the long audio files into 8 seconds long sequences on which we will perform truncated backpropagation through time. Each sequence in the Onomatopoeia dataset is few seconds long, ranging from 1 to 11 seconds. To train the models on this dataset, zero-padding has been applied to make all the sequences in a mini-batch have the same length and corresponding cost values (for the predictions over the added 0s) would be ignored when computing the gradients. We particularly explored two gated variants of RNNs—GRUs and LSTMs. For the case of LSTMs, the forget gate bias is initialized with a large positive value of 3, as recommended by @zaremba2015empirical and @gers2001long, which has been shown to be beneficial for learning long-term dependencies. As for models that take real-valued input, e.g. the RNN-GMM and SampleRNN-GMM (with 4 components), normalization is applied per audio sample with the global mean and standard deviation obtained from the train split. For most of our experiments where the model demands discrete input, binning was applied per audio sample. All the models have been trained with teacher forcing and stochastic gradient decent (mini-batch size 128) to minimize the Negative Log-Likelihood (NLL) in bits per dimension (per audio sample). Gradients were hard-clipped to remain in \[-1, 1\] range. Update rules from the Adam optimizer [@kingma2014adam] ($\beta_1=0.9$, $\beta_2=0.999$, and $\epsilon=1e{-8}$) with an initial learning rate of 0.001 was used to adjust the parameters. For training each model, random search over hyper-parameter values [@bergstra2012random] was conducted. The initial RNN state of all the RNN-based models was always learnable. Weight Normalization [@salimans2016weight] has been used for all the linear layers in the model (except for the embedding layer) to accelerate the training procedure. Size of the embedding layer was 256 and initialized by standard normal distribution. Orthogonal weight matrices used for hidden-to-hidden connections and other weight matrices initialized similar to @he2015delving. In final model, we found GRU to work best (slightly better than LSTM). 1024 was the the number of hidden units for all GRUs (1 layer per tier for 3-tier and 3 layer for 2-tier model) and MLPs (3 fully connected layers with ReLU activation with output dimension being 1024 for first two layers and 256 for the final layer before softmax). Also $FS^{(1)}=FS^{(2)}=2$ and $FS^{(3)}=8$ were found to result in lowest NLL. ![Examples from the datasets compared to samples from our models. In the first 3 rows, 2 seconds of audio are shown. In the bottom 3 rows, 100 milliseconds of audio are shown. Rows 1 and 4 are ground truth from which one can see how the datasets look different and have complex structure in low resolution which the frame-level component of the SampleRNN is designed to capture. Samples also to some extent mimic the same global structure. At the same time, zoomed-in samples of our model shows that it can perfectly resemble the high resolution structure present in the data as well.[]{data-label="fig:samples"}](figures/samples.pdf){width="100.00000%"} [\*4c]{} & & &\ (lr)[1-4]{} RNN (Eq. \[eq:rnn\]) &1.434&2.034&1.410\ WaveNet (re-impl.)&1.480& 2.285 & 1.464\ (lr)[1-4]{} SampleRNN (2-tier) &1.392&2.026&**1.076\ SampleRNN (3-tier) &**1.387&**1.990&1.159\ (lr)[1-4]{}****** [\*2c]{} &\ (lr)[1-2]{} RNN-GMM &-2.415\ SampleRNN-GMM (2-tier) &**-2.782\ (lr)[1-2]{}** [\*6c]{} [[**Subsequence Length**]{.nodecor}]{} & 32 & 64 & 128 & 256 & 512\ (lr)[1-6]{} [[**NLL Validation**]{.nodecor}]{} &1.575 &1.468 &1.412 &1.391 &1.364 [\*2c]{} &\ (lr)[1-2]{} SampleRNN (2-tier) & 1.392 (1.369)\ Without Embedding & 1.566 (1.539)\ Multi-Softmax & 1.685 (1.656)\ (lr)[1-2]{} WaveNet Re-implementation ------------------------- We implemented the WaveNet architecture as described in [@oord2016wavenet]. Ideally, we would have liked to replicate their model exactly but owing to missing details of architecture and hyperparameters, as well as limited compute power at our disposal, we made our own design choices so that the model would fit on a single GPU while having a receptive field of around 250 milliseconds, while having a reasonable number of updates per unit time. Although our model is very similar to WaveNet, the design choices, e.g. number of convolution filters in each dilated convolution layer, length of target sequence to train on simultaneously (one can train with a single target with all samples in the receptive field as input or with target sequence length of size T with input of size receptive field + T - 1), batch-size, etc. might make our implementation different from what the authors have done in the original WaveNet model. Hence, we note here that although we did our best at exactly reproducing their results, there would very likely be different choice of hyper-parameters between our implementation and the one of the authors. For our WaveNet implementation, we have used 4 dilated convolution blocks each having 10 dilated convolution layers with dilation 1, 2, 4, 8 up to 512. Hence, our network has a receptive field of 4092 acoustic samples i.e. the parameters of multinomial distribution of sample at time step t, $p(x_i) = f_{\theta}(x_{i-1}, x_{i-2}, \ldots x_{i-4092})$ where $\theta$ is model parameters. We train on target sequence length of 1600 and use batch size of 8. Each dilated convolution filter has size 2 and the number of output channels is 64 for each dilated convolutional layer (128 filters in total due to gated non-linearity). We trained this model using Adam optimizer with a fixed global learning rate of 0.001 for Blizzard dataset and 0.0001 for Onomatopoeia and Music datasets. We trained these models for about one week on a GeForce GTX TITAN X. We dropped the learning rate in the Blizzard experiment to 0.0001 after around 3 days of training. Human Evaluation {#sec:humaneval} ---------------- Apart from reporting NLL, we conducted AB preference tests for random samples from four models trained on the Blizzard dataset. For unconditional generation of speech which at best sounds like mumbling, this type of test is the one which is more suited. Competing models were the RNN, SampleRNN (2-tier), SampleRNN (3-tier), and our implementation of WaveNet. The rest of the models were excluded as the quality of samples were definitely lower and also to keep the number of pair comparison tests manageable. We will release the samples that have been used in this test too. All the samples were set to have the same volume. Every user is then shown a set of twenty pairs of samples with one random pair at a time. Each pair had samples from two different models. The human evaluator is asked to listen to the samples and had the option of choosing between the two model or choosing not to prefer any of them. Hence, we have a quantification of preference between every pair of models. We used the online tool made publicly available by  [@waet2015]. Results in Fig. \[fig:humanevalresults\] clearly points out that SampleRNN (3-tier) is a winner by a huge margin in terms of preference by human raters, then SampleRNN (2-tier) and afterward two other models, which matches with the performance comparison in Table \[table:summary\_results\]. ![Pairwise comparison of 4 best models based on the votes from listeners conducted on samples generated from models trained on Blizzard dataset.[]{data-label="fig:humanevalresults"}](figures/figure_HE_Blizz.pdf){width="100.00000%" height="\textheight"} The same evaluation was conducted for Music dataset except for an additional filtering process of samples. Specific to only this dataset, we observed that a batch of generated samples from competing models (this time restricted to RNN, SampleRNN (2-tier), and SampleRNN (3-tier)) were either music-like or random noise. For all these models we only considered random samples that were not random noise. Fig. \[fig:humanevalresultsmusic\] is dedicated to result of human evaluation on Music dataset. ![Pairwise comparison of 3 best models based on the votes from listeners conducted on samples generated from models trained on Music dataset.[]{data-label="fig:humanevalresultsmusic"}](figures/figure_HE_Music.pdf){width="100.00000%" height="\textheight"} Quantifying Information Retention {#sec:timehorizon} --------------------------------- For the last experiment we are interested in measuring the memory span of the model. We trained our model, SampleRNN (3-tier), with best hyper-parameters on a dataset of 2 speakers reading audio books, one male and one female, respectively, with mean fundamental frequency of 125.3 and 201.8Hz. Each speaker has roughly 10 hours of audio in the dataset that has been preprocessed similar to Blizzard. We observed that it learned to stay consistent generating samples from the same speaker without having any knowledge about the speaker ID or any other conditioning information. This effect is more apparent here in comparison to the unbalanced Onomatopoeia that sometimes mixes two different categories of sounds. Another experiment was conducted to test the effect of memory and study the effective memory horizon. We inject 1 second of silence in the middle of sampling procedure in order to see if it will remember to generate from the same speaker or not. Initially when sampling we let the model generate 2 seconds of audio as it normally do. From 2 to 3 seconds instead of feeding back the generated sample at that timestep a silent token (zero amplitude) would be fed. From 3 to 5 seconds again we sample normally; feeding back the generated token. We did classification based on mean fundamental frequency of speakers for the first and last 2 seconds. In 83% of samples SampleRNN generated from the same person in two separate segments. This is in contrast to a model with fixed past window like WaveNet where injecting 16000 silent tokens (3.3 times the receptive field size) is equivalent to generating from scratch which has 50% chance (assuming each 2-second segment is coherent and not a mixed sound of two speakers). Related Work ============ Our work is related to earlier work on auto-regressive multi-layer neural networks, starting with @bengio1999modeling, then NADE [@larochelle2011neural] and more recently PixelRNN [@van2016pixel]. Similar to how they tractably model joint distribution over units of the data (e.g. words in sentences, pixels in images, etc.) through an auto-regressive decomposition, we transform the joint distribution of acoustic samples using Eq. \[eq:px\]. The idea of having part of the model running at different clock rates is related to multi-scale RNNs [@schmidhuber1992learning; @el1995hierarchical; @koutnik2014clockwork; @sordoni2015hierarchical; @serban2016building]. @chung2015recurrent also attempt to model raw audio waveforms which is in contrast to traditional approaches which use spectral features as in @tokuda2013speech, @bertrand2008unsupervised, and @lee2009unsupervised. Our work is closely related to WaveNet [@oord2016wavenet], which is why we have made the above comparisons, and makes it interesting to compare the effect of adding higher-level RNN stages working at a low resolution. Similar to this work, our models generate one acoustic sample at a time conditioned on all previously generated samples. We also share the preprocessing step of quantizing the acoustics into bins. Unlike this model, we have different modules in our models running at different clock-rates. In contrast to WaveNets, we mitigate the problem of long-term dependency with hierarchical structure and using stateful RNNs, i.e. we will always propagate hidden states to the next training sequence although the gradient of the loss will not take into account the samples in previous training sequence. Discussion and Conclusion ========================= We propose a novel model that can address unconditional audio generation in the raw acoustic domain, which typically has been done until recently with hand-crafted features. We are able to show that a hierarchy of time scales and frequent updates will help to overcome the problem of modeling extremely high-resolution temporal data. That allows us, for this particular application, to learn the data manifold directly from audio samples. We show that this model can generalize well and generate samples on three datasets that are different in nature. We also show that the samples generated by this model are preferred by human raters. Success in this application, with a general-purpose solution as proposed here, opens up room for more improvement when specific domain knowledge is applied. This method, however, proposed with audio generation application in mind, can easily be adapted to other tasks that require learning the representation of sequential data with high temporal resolution and long-range complex structure. ### Acknowledgments {#acknowledgments .unnumbered} The authors would like to thank João Felipe Santos and Kyle Kastner for insightful comments and discussion. We would like to thank the @2016arXiv160502688short[^4] and MILA staff. We acknowledge the support of the following agencies for research funding and computing support: NSERC, Calcul Québec, Compute Canada, the Canada Research Chairs and CIFAR. Jose Sotelo also thanks the Consejo Nacional de Ciencia y Tecnología (CONACyT) as well as the Secretaría de Educación Pública (SEP) for their support. This work was a collaboration with Ubisoft. Appendix A {#appendix-a .unnumbered} ========== A model variant: SampleRNN-WaveNet Hybrid {#a-model-variant-samplernn-wavenet-hybrid .unnumbered} ----------------------------------------- SampleRNN-WaveNet model has two modules operating at two different clock-rate. The slower clock-rate module (frame-level module) sees one frame (each of which has size [*FS*]{}) at a time while the faster clock-rate component(sample-level component) sees one acoustic sample at a time i.e. the ratio of clock-rates for these two modules would be the size of a single frame. Number of sequential steps for frame-level component would be [*FS*]{} times lower. We repeat the output of each step of frame-level component [*FS*]{} times so that number of time-steps for output of both the components match. The output of both these modules are concatenated for every time-step which is further operated by non-linearities for every time-step independently before generating the final output. In our experiments, we kept size of a single frame ([*FS*]{}) to be 128. We tried two variants of this model: 1. fully convolutional WaveNet and 2. RNN-WaveNet. In fully convolutional WaveNet, both modules described above are implemented using dilated convolutions as described in original WaveNet model. In RNN-WaveNet, we use high capacity RNN in the frame-level module to model the dependency between frames. The sample-level WaveNet in RNN-WaveNet has receptive field of size 509 samples from the past. Although these models are designed with the intention of combining the two models to harness their best features, preliminary experiments show that this variant is not meeting our expectations at the moment which directs us to a possible future work. [^1]: Statistics based on the average speaking rate of a set of TED talk speakers  <http://sixminutes.dlugan.com/speaking-rate/> [^2]: Code <https://github.com/soroushmehr/sampleRNN_ICLR2017> and samples <https://soundcloud.com/samplernn/sets> [^3]: Courtesy of Ubisoft [^4]: http://deeplearning.net/software/theano/

ArXiv

--- abstract: 'A scheme based on electromagnetically induced transparency is used to store light in a Bose-Einstein condensate. In this process, a photonic polarization qubit is stored in atomic Zeeman states. The performance of the storage process is characterized and optimized. The average process fidelity is $1.000 \pm 0.004$. For long storage times, temporal fluctuations of the magnetic field reduce this value, yielding a lifetime of the fidelity of $(1.1\pm0.2)$ ms. The write-read efficiency of the pulse energy can reach $0.53 \pm 0.05$.' author: - Stefan Riedl - Matthias Lettner - Christoph Vo - Simon Baur - Gerhard Rempe - Stephan Dürr title: 'Bose-Einstein condensate as a quantum memory for a photonic polarization qubit' --- Introduction ============ Optical quantum-memories [@lvovsky:09] based on electromagnetically induced transparency (EIT) [@fleischhauer:05] are a very active research area. Specifically, storage of classical light pulses [@liu:01; @turukhin:01; @phillips:01] and of single photons [@chaneliere:05; @eisaman:05] was demonstrated. In addition, different polarizations of light were stored either using extended atomic level schemes [@matsukevich:06; @tanji:09] or by converting the polarization into other degrees of freedom before storage [@choi:08; @Cho:10; @Zhang:11; @England:1112.0900]. The vast majority of applications envisioned for quantum memories requires that quantum entanglement is first generated between two or more particles and that the quantum states of one or several of these particles are subsequently stored in a quantum memory. The crucial point is that this entanglement must survive the storage. Recently, this aspect was experimentally demonstrated in three independent experiments [@saglamyurek:11; @clausen:11; @lettner:11]. Here we report in detail on the performance of the quantum memory used in one of these experiments [@lettner:11]. The experiment uses an $^{87}$Rb Bose-Einstein condensate (BEC) to realize a quantum memory for the polarization qubit of a single photon. A Raman scheme based on EIT is used to implement storage and retrieval of the photon. The atomic level-scheme is extended to allow for storage of the photonic polarization qubit in two atomic spin states. Quantum process tomography is used to determine the process fidelity which quantifies how well the polarization is maintained during storage. In addition, the decay of the process fidelity with increasing storage time is monitored. All experiments reported here use classical light pulses instead of single photons, thus profiting from count rates which are much higher than in Ref. [@lettner:11]. These increased count rates yield a more precise value of the process fidelity. Ref. [@lettner:11] reports that storage and retrieval cause no discernible deterioration of the fidelity of the entangled state within an error bar of several percent. The present experiment still observes a process fidelity compatible with unity, but now with an error bar that is an order of magnitude smaller, thus demonstrating more clearly the capabilities of the BEC as a quantum memory. In Ref. [@lettner:11], a single $^{87}$Rb atom in an optical high-finesse cavity was used to generate a triggered single photon in such a way that the photon’s polarization qubit is entangled with the spin state of the single atom. This unparalleled source is combined with the BEC that serves as a quantum memory for the single photon. The BEC is well suited for this purpose because the absence of thermal motion allows for long storage times, the large optical depth allows for high write-read efficiencies, and excellent internal-state preparation allows for high-fidelity storage of a qubit in atomic spin states. The experiment thus combines two different systems, each ideally suited for its purpose. The resulting hybrid character of the system poses an experimental challenge because the dipole traps that hold the single atom and the BEC in place have depths of several millikelvin and several microkelvin, respectively. Due to the resulting ac-Stark shifts, the single photons generated from the single atom are blue detuned by typically 70 MHz relative to the free-space atomic resonance, whereas the ac-Stark shifts experienced by the BEC are negligible on this scale. In the present paper, we therefore experimentally study the efficiency of light storage in the regime of 70 MHz single-photon detuning. The regime of much larger single-photon detunings has been studied theoretically [@gorshkov:07; @nunn:07; @mishina:08; @sheremet:10] and experimentally [@reim:10; @England:1112.0900] before, but those results are not immediately applicable to our system. The paper is structured as follows: Section \[sec-implementation\] describes the experimental implementation, Sec. \[sec-qubit\] shows how well the polarization is maintained during storage, and Sec. \[sec-efficiency\] studies the write-read efficiency. Appendix \[app\] presents a simple model for coarsely estimating the write-read efficiency. Experimental implementation {#sec-implementation} =========================== Electromagnetically induced transparency {#sub-impl-EIT} ---------------------------------------- ![ \[fig-levels\] Atomic level scheme of the $D_1$ line in $^{87}$Rb. Probe light (dashed arrows) with an arbitrary superposition of the polarizations $\sigma^+$ and $\sigma^-$ couples the initial population ($\bullet$) in the hyperfine ground state $|F,m_F\rangle=|1,0\rangle$ to the $D_1$-line excited states $|F',m_F'\rangle=|1,\pm1\rangle$. $\pi$-polarized control light (solid arrows) couples these states to the hyperfine ground states $|F,m_F\rangle=|2,\pm1\rangle$. This makes it possible to store the probe-light polarization qubit in the qubit space spanned by the atomic states $|F,m_F\rangle=|2,\pm1\rangle$. ](fig1){width="0.75\columnwidth"} EIT employs a control light beam to manipulate the propagation of a probe light beam inside a medium. Light storage in EIT-based schemes relies on the fact that the group velocity of the probe light $v_\mathrm{gr}$ can be reduced compared to the vacuum speed of light $c$ by many orders of magnitude [@hau:99] by choosing a small value for the control intensity. Upon entering the medium, the temporal duration of the probe pulse remains unchanged, whereas its spatial length is drastically reduced due to the small group velocity. A pulse which in vacuum is much longer than the medium can thus be fully compressed into the medium. Once the pulse is fully inside the medium, one can ramp the control intensity to zero in an adiabatic fashion. In our experiment, we implement an approximately linear, temporal ramp of the control intensity which lasts 30 ns. This is sufficiently adiabatic according to Ref. [@fleischhauer:02]. This ramp reduces $v_\mathrm{gr}$ all the way to zero and the pulse is stopped inside the medium. The pulse is stored for a time $t_\mathrm{store}$ which can be chosen freely. After this, we ramp the control intensity back on and the pulse resumes its propagation [@liu:01; @turukhin:01; @phillips:01; @fleischhauer:00]. During the storage time $t_\mathrm{store}$, the pulse exists in the medium in the form of an atomic spin wave. If the light pulse is compressed such that it fits inside the medium, then the spin wave stores the longitudinal and transverse shape of the light pulse. Atomic level scheme {#sub-impl-levels} ------------------- Figure \[fig-levels\] shows the atomic level scheme used in our experiment. Control and probe light for EIT are both resonant with the atomic $D_1$ line of $^{87}$Rb at a wavelength of $\lambda=795$ nm. The atoms are initially prepared in the hyperfine ground state $|F,m_F\rangle=|1,0\rangle$. The $\sigma^\pm$ polarized components of the probe light couple this population to the excited hyperfine states $|F',m_F'\rangle=|1,\pm 1\rangle$. The $\pi$-polarized control light transfers this population to the ground hyperfine states $|F,m_F\rangle=|2,\pm1\rangle$. Optical beam path {#sub-impl-beam-path} ----------------- ![ \[fig-beam-path\] Simplified scheme of the optical beam path. A detailed description is given in the text. ](fig2){width="0.95\columnwidth"} Figure \[fig-beam-path\] shows a simplified scheme of the optical beam path. An $^{87}$Rb BEC serves as a quantum memory. The BEC is illuminated by $\pi$-polarized control light propagating along the $y$ axis with a waist ($1/e^2$ radius of intensity) of $\sim 100$ $\mu$m which is much larger than the Thomas-Fermi radii of the BEC so that the control light intensity can be approximated as constant across the BEC. In addition, the BEC is illuminated by probe light propagating along the $z$ axis focused to a waist of 8 $\mu$m. This is comparable to the Thomas-Fermi radii so that the probe beam samples some fraction of the transverse inhomogeneity of the BEC. To obtain a well-defined transverse mode for the probe light before impinging onto the BEC, the light is sent through a single-mode fiber. The polarization of the probe light can be $\sigma^+$, $\sigma^-$, or any superposition thereof. After storing and retrieving the probe light, we need to measure its polarization. To this end, the beam path ends with a polarizing beam splitter (PBS) cube and two identical detectors, one in each output port of the PBS. A quarter-wave plate (QWP) followed by a half-wave plate (HWP), both placed right in front of the PBS allow for the selection of an arbitrary polarization basis. As detectors, we use photomultiplier tubes (PMTs) in this paper, instead of the avalanche photodiodes that we used in Ref. [@lettner:11]. Stray light filtering {#sub-impl-filtering} --------------------- Stray light is an issue in our setup. Much of it is eliminated using mechanical shielding and temporal gating of the detector signals. The remaining stray light is dominated by control light off-resonantly scattered from the BEC during the retrieval of the probe pulse. This stray light level would be unproblematic for the measurements presented here, but the experiments reported in Ref. [@lettner:11] required a substantial suppression. In the beam path from the BEC to the detectors, our setup therefore includes a single-mode optical fiber for transverse mode filtering and a filter cavity for spectral filtering. The spatial filtering with the single-mode fiber makes use of the fact that storage and retrieval have little effect on the transverse mode of the probe light, whereas the control light is off-resonantly scattered from all positions in the BEC and into all directions. The fiber reduces the stray light power that reaches the detector by a factor of 0.068. In the absence of the BEC, the fiber reduces the probe light power by a factor of 0.66. Hence, the fiber increases the signal-to-background ratio by one order of magnitude. Storage and retrieval in the BEC compromise the transverse mode of the probe beam slightly. This causes an additional reduction of the probe light power by a factor of 0.88 at the single-photon resonance and by a factor of 0.80 for a single-photon detuning of $\Delta_c=2\pi\times 70$ MHz. In addition, the fiber suppresses the excitation of higher transverse modes in the subsequent filter cavity. The filter cavity is a near-planar, plane-concave Fabry-Perot resonator with a finesse of 180 and a free spectral range of 40 GHz. The cavity has a transmission of 0.8 at the resonance, which is tuned to the probe light frequency. We expect that the scattered EIT control light is either elastically scattered or Raman scattered, thereby transferring an atom from $F=1$ to $F=2$. Hence, we expect the scattered light to be red detuned from the probe light by either 6.8 GHz or 13.6 GHz. Transmission through the cavity suppresses the light power for each of these frequencies by a factor of $2\times10^{-4}$. The cavity length is stabilized against long-term drift with a piezo actuator using a Pound-Drever-Hall technique with light at a wavelength of 830 nm. This light is overlapped with the probe light on a dichroic mirror in front of the mode-filtering fiber. The 830 nm light transmitted through the filter cavity is separated from the probe light using another dichroic mirror behind the filter cavity. A small fraction of the 830 nm light keeps propagating towards the detectors. This light is removed with dielectric interference filters which are not shown in Fig. \[fig-beam-path\]. In addition, the 830 nm light source is turned off during any time intervals where relevant detector signals are expected. The optical fiber for stray-light filtering is not polarization maintaining because it must work equally well for all possible input polarizations. Hence, the polarization after transmission through the fiber is related to the input polarization by a unitary transformation. As long as the fiber is not moved mechanically or exposed to temperature changes, this transformation is temporally stable. The same applies to the polarization transformations caused by the filter cavity, by the dichroic mirrors, and by other mirrors in the beam path which are not shown in Fig. \[fig-beam-path\]. We compensate the resulting overall transformation using two QWPs followed by a HWP — a combination which can generate any unitary transformation. Preparation of the BEC {#sub-impl-BEC} ---------------------- {width="90.00000%"} We produce an almost pure BEC in the hyperfine state $|F,m_F\rangle=|1,-1\rangle$, using radio-frequency (rf) induced evaporative cooling in a magnetic trap, as described in Ref. [@marte:02]. The gas is transferred into a crossed-beam optical dipole trap operated at a wavelength of 1064 nm. The measured trap frequencies are $(\omega_x,\omega_y,\omega_z)/2\pi=(70,20,20)$ Hz with gravity pointing along the $x$ axis. A magnetic hold field of $\sim 1$ G applied along the $z$ axis preserves the spin orientation of the atoms. We use two consecutive microwave pulses, each with a pulse area of $\pi$, to transfer the population into the internal state needed for our EIT level scheme. Starting from state $|1,-1\rangle$, the first pulse transfers the population into state $|2,0\rangle$. Subsequently, the second pulse transfers the population to state $|1,0\rangle$. The total process transfers $\sim 90$ % of the atoms into state $|1,0\rangle$. Atoms left in the $F=2$ hyperfine states are then removed with blast light. This is followed by temporary application of a strong magnetic field gradient which removes atoms with $m_F\neq0$ from the shallow optical dipole trap. After this procedure, the total atom number in undesired internal states lies below the detection limit of our setup which we estimate to be $\sim 200$ atoms. For the rest of the experiment, the magnetic hold field is reduced to typically 0.1 G. At this point, the BEC typically contains $N=1.2\times 10^6$ atoms. The corresponding Thomas-Fermi radii are $(R_x,R_y,R_z)=(7,25,25)$ $\mu$m. A quantum memory for the polarization qubit {#sec-qubit} =========================================== We now study how well the polarization of the probe light is maintained during storage. The probe beam propagates along the $z$ axis. Hence, an arbitrary incoming polarization state can be expanded as $c_+|\sigma^+\rangle + c_- |\sigma^-\rangle$ with coefficients $c_+$ and $c_-$. With the atomic level scheme shown in Fig. \[fig-levels\], this state is mapped onto the atomic state $c_+|2,+1\rangle + c_- |2,-1\rangle$. The retrieval process maps the atomic state back to the original polarization state. In our experiment, both mapping processes work extremely well, but magnetic field noise causes a deterioration of the state for long storage times. We apply a magnetic hold field of $B_z\sim0.1$ G and orient it along the $z$ axis. This suppresses undesired transitions between different Zeeman states caused by components of the magnetic-field noise perpendicular to the $z$ axis. In Sec. \[sub-qubit-Faraday\] we discuss the Faraday rotation caused by this hold field. In Sec. \[sub-qubit-tomography\] we use quantum state tomography to characterize the deterioration of the polarization for long storage times and show how techniques that reduce the magnetic-field noise improve the performance of the system. Faraday rotation {#sub-qubit-Faraday} ---------------- The polarization of the probe light can be characterized using the Stokes parameters [@hecht:01; @brosseau:98] \[Stokes\] $$\begin{aligned} S_0 = I_H+I_V , &&\qquad S_1 = I_H-I_V , \\ S_2 = I_D-I_A , &&\qquad S_3 = I_R-I_L ,\end{aligned}$$ where $I_H$, $I_V$, $\hdots$ denote the intensity detected after a polarizer that transmits only one polarization, namely horizontal $H$, vertical $V$, diagonal ($+45^\circ$) $D$, anti-diagonal ($-45^\circ$) $A$, right circular $R$, or left circular $L$. Here, $R$ and $L$ correspond to $\sigma^+$ and $\sigma^-$. The Stokes parameters can be regarded as the components of a four-dimensional Stokes vector. $S_0$ describes the total intensity, whereas the three-dimensional vector $$\begin{aligned} \bm u = \frac1{S_0} \left(\begin{array}{c} S_1 \\ S_2 \\ S_3 \end{array}\right)\end{aligned}$$ describes the polarization of the light and is well suited for graphical visualization, in close analogy to the Bloch vector. We call $\bm u$ the Poincaré vector. Its unit sphere is called Poincaré sphere. The magnetic hold field along the $z$ axis gives rise to a Faraday rotation of the Poincaré vector around its $z$ axis at an angular frequency $$\begin{aligned} \omega_F = \frac{\mu_B g_F \Delta m_F}\hbar B_z ,\end{aligned}$$ where $\mu_B=2\pi\hbar \times 1.40$ MHz/G is the Bohr magneton and $g_F$ is the Landé factor. For the levels used in our experiment $g_F=1/2$ and $\Delta m_F=2$. The total rotation angle of the Poincaré vector $$\begin{aligned} \label{phi} \phi = \omega_F (t_\mathrm{store} + \tau_d(L))\end{aligned}$$ has one contribution $t_\mathrm{store}$ from the storage time with the control light off and another contribution $\tau_d(L)$ which is the delay of the probe pulse caused by the propagation through the medium of length $L$ with the control light on. Note that the rotation of the polarization vector of the electric field in real space is a factor of 2 slower than the rotation of the Poincaré vector. Figure \[fig-Faraday\] shows experimental data of this Faraday rotation for a linear input polarization. The line shows a fit of a sinusoid with Gaussian damping $$\begin{aligned} \label{S-1} \frac{S_1}{S_0} = e^{-t_\mathrm{store}^2/2\sigma_\alpha^2} \cos (\phi-\phi_0) ,\end{aligned}$$ where $\sigma_\alpha$ is the $e^{-1/2}$ damping time, where $\phi$ is given by Eq. , and where $\phi_0$ represents the input polarization. The best-fit values are $\omega_F=2\pi \times 0.20$ MHz, yielding $B_z=0.14$ G, and $\sigma_\alpha=(1.1\pm0.2)$ ms. A careful inspection of the experimental data points for long times in Fig. \[fig-Faraday\] leads us to an interesting observation, revealing the physical origin of the damping. Unlike the best-fit curve, the data points, which represent a single experimental shot each, do not show a decrease of the peak-to-peak amplitude. Instead, they are noisy insofar as they do not form a smooth sinusoid. We attribute this to irreproducible, temporal fluctuations of $B_z$, which effectively scatter the data points along the horizontal axis. Only if we were to average several experimental shots to represent their mean value, would we observe a reduction of the peak-to-peak values in the experimental data. A fit of $S_1/S_0=A\cos(\phi-\phi_0)$ to the data in Fig. \[fig-Faraday\] with $t_\mathrm{store}\leq20$ $\mu$s yields a best-fit value of $A=1.02\pm0.04$. The fact that $A$ is consistent with 1 shows that temporal fluctuations of $B_z$ on a time scale of 20 $\mu$s or faster have no discernible effect. Temporal fluctuations of $B_z$ on slower time scales manifest themselves in our experiment only as shot-to-shot fluctuations of $\phi=(\mu_B g_F \Delta m_F/\hbar) \linebreak[1] \int_0^{t_\mathrm{store}+\tau_d(L)} dt \linebreak[1] B_z(t)$. In our experiment, each shot requires the preparation of a new BEC, which takes 20 s. This suggests that shot-to-shot fluctuations of $B_z$ probably yield an important contribution to the shot-to-shot fluctuations of $\phi$. Quantum process tomography {#sub-qubit-tomography} -------------------------- We now turn to a complete characterization of the effect which the storage and retrieval process has on the polarization. As the process does not conserve the total intensity, a full description of the process must use the four-dimensional Stokes vector, not just the three-dimensional Poincaré vector. We consider the regime of small probe intensity. Hence, the dependence of the Stokes parameters $S_i^\mathrm{out}$ of the retrieved probe pulse on the Stokes parameters $S_i^\mathrm{in}$ of the incoming probe pulse can be approximated as linear $$\begin{aligned} S_i^\mathrm{out} = \sum_j M_{ij} S_j^\mathrm{in} .\end{aligned}$$ $M$ is called Müller matrix [@hecht:01; @brosseau:98]. As shown in Fig. \[fig-beam-path\], we use a PBS with detectors behind both output ports. According to Eq. , a full characterization of the Stokes vector requires such measurements for 3 different settings of the wave plates in front of the PBS, which select the measurement basis. This set of 3 measurements fully characterizes the quantum state of the polarization and it can be regarded as quantum *state* tomography [@nielsen:00]. To determine $M$ experimentally, we use a set of 4 linearly independent input Stokes vectors (e.g., $H$, $D$, $R$, and $L$) and perform quantum state tomography of the output state generated for each input state. This set of 12 measurements allows for a complete determination of $M$ and it can be regarded as quantum *process* tomography [@nielsen:00]. Performing such quantum process tomography, we find that in our experiment the Müller matrix is always well approximated by $$\begin{aligned} \label{M} M = \eta \left(\begin{array}{cccc} 1 & 0 & 0 & 0 \\ 0 & \alpha \cos \phi & -\alpha \sin \phi & 0 \\ 0 & \alpha \sin \phi & \alpha \cos \phi & 0 \\ 0 & 0 & 0 & 1 \\ \end{array}\right) ,\end{aligned}$$ where $\phi$ is the angle resulting from the Faraday rotation, $\alpha$ is a damping factor, and $\eta$ is the write-read efficiency which is experimentally found to be independent of the input polarization. Above, we concluded from Fig. \[fig-Faraday\] that the polarization at short storage times is essentially pure, whereas at long storage times shot-to-shot fluctuation of $\phi$ must be taken into account. Hence, each individual shot can be described by some realization of $M$ as in Eq.  with $\alpha=1$ and with some value of $\phi$ which exhibits shot-to-shot fluctuations. We assume that the values of $\phi$ have a Gaussian distribution with root-mean-square (rms) width $\sigma_\phi$. Averaging over many shots yields Eq.  with $\alpha=\exp(-\sigma_\phi^2/2)$. To develop a simple model for the dependence of $\alpha$ on $t_\mathrm{store}$, we assume that only shot-to-shot fluctuations of $B_z$ contribute to the fluctuations of $\phi$, i.e. we approximate $B_z$ as constant during each individual shot. With this approximation, the values of $B_z$ will have a Gaussian distribution with rms width $\sigma_B$ and with $\sigma_\phi=(\mu_B g_F \Delta m_F/\hbar)\sigma_B t_\mathrm{store}$, where we neglected $\tau_d(L)\ll t_\mathrm{store}$. This yields $$\begin{aligned} \label{alpha-def} \alpha = \exp(-t_\mathrm{store}^2/2\sigma_\alpha^2)\end{aligned}$$ with $$\begin{aligned} \label{sigma-alpha} \frac1{\sigma_\alpha} = \sigma_B \frac{\partial \omega_F}{\partial B_z} .\end{aligned}$$ For a linearly polarized input state, Eqs.  and reproduce Eq. . The Faraday rotation is a unitary evolution. For any given $t_\mathrm{store}$, its effect can be compensated, e.g., using wave plates and it is therefore not much of a concern. The non-unitary damping $\alpha$, however, irreversibly deteriorates the performance of the memory. ![ \[fig-alpha\] (Color online) The damping factor $\alpha$ of Eq.  as a function of storage time. The lowest damping is obtained when synchronizing the start of the EIT write-read cycle with the 50 Hz ac line voltage ($\blacksquare$). Alternatively, we can use an open-loop feed-forward compensation ($\bullet$). Without any compensation, the damping is much stronger ($\blacktriangle$). The lines show Gaussian fits according to Eq. . ](fig4){width="0.9\columnwidth"} Experimental results for the time dependence of the damping parameter $\alpha$, as determined by quantum process tomography, are shown in Fig. \[fig-alpha\]. Data taken without any reduction of magnetic-field noise ($\blacktriangle$) yield a value of $\sigma_\alpha = 0.06$ ms for the $e^{-1/2}$ damping time, corresponding to $\sigma_B=2$ mG according to Eq. . In our experiment, the majority of this magnetic-field noise is periodic and in phase with the 50 Hz ac line voltage. We can suppress this noise drastically by synchronizing the start of the EIT write-read cycle with the ac line voltage. The corresponding data ($\blacksquare$) in Fig. \[fig-alpha\] yield a best-fit value of $\sigma_\alpha = (1.0\pm0.1)$ ms, corresponding to $\sigma_B=0.1$ mG. Evidently, the synchronization improves $\sigma_\alpha$ by a factor of $\sim 20$. The data in Fig. \[fig-Faraday\] were also taken with this synchronization and essentially reproduce the improved value of $\sigma_\alpha$. Our experiments described in Ref. [@lettner:11] required a repetition of EIT write-read cycles at a rate of 10 kHz for a total time span of several seconds. Hence, write-read cycles had to occur at essentially all possible phases of the 50 Hz ac line voltage. To reduce the noise in these measurements, we first determined the values of $B_z(t)$ for one 50 Hz period in a series of calibration measurements. We then ran a current through a coil to compensate the recorded noise with an open-loop feed-forward circuit. The corresponding data ($\bullet$) in Fig. \[fig-alpha\] yield a best-fit value of $\sigma_\alpha = (0.49\pm0.04)$ ms, corresponding to $\sigma_B=0.2$ mG. This compensation was good enough not to be the limiting factor in the overall experiment of Ref. [@lettner:11], where we observed the same lifetime but with an error bar that was a factor of four larger. The full information from the quantum process tomography is contained in $M$. To compare the overall performance of different quantum memories, one often uses the average process fidelity as a figure of merit. In terms of quantum states, the fidelity can be written as $F=\mathrm{Tr} (\rho_\mathrm{in}\rho_\mathrm{out})$, where Tr denotes the trace, $\rho$ denotes the density matrix, and we assumed that $\rho_\mathrm{in}$ represents a pure state. For polarization states, this can be rewritten as $F=(1+\bm u_\mathrm{in}\cdot \bm u_\mathrm{out})/2$. When averaging this quantity over all possible pure input states, i.e. over the surface of the Poincaré sphere, we obtain the average process fidelity $\langle F\rangle$ [@lvovsky:09; @bowdrey:02]. After compensation of the Faraday rotation, i.e. for $\phi=0$, Eq.  yields $$\begin{aligned} \langle F\rangle= \frac13(2+\alpha) .\end{aligned}$$ The synchronized data in Fig. \[fig-alpha\] yield $\langle F \rangle = 1.000 \pm 0.004$ at $t_\mathrm{store}= 1$ $\mu$s and $\langle F \rangle = 0.90\pm0.02$ at $t_\mathrm{store}= 800$ $\mu$s. The value at $t_\mathrm{store}= 1$ $\mu$s shows that the state mapping between photonic and atomic qubit states works extremely well. Write-read efficiency {#sec-efficiency} ===================== Now, we turn to another important figure of merit for light storage, namely the efficiency of a complete write-read cycle $$\begin{aligned} \eta = \frac{E_\mathrm{retr}}{E_\mathrm{in}} ,\end{aligned}$$ defined as the energy of the retrieved probe pulse $E_\mathrm{retr}$ divided by the energy of the incoming probe pulse $E_\mathrm{in}$. For our memory, $\eta$ is independent of the polarization of the probe field, as seen in Eq. . Hence, for understanding $\eta$ it suffices to consider the case where the probe polarization is fixed to $\sigma^+$. This simplifies the relevant atomic level scheme to a $\Lambda$-type three-level system. Unlike previous experiments by other groups, our work has a focus on the regime of 70 MHz detuning from the single-photon resonance. We experimentally study the dependence of $\eta$ on the intensity of the control laser in Sec. \[sub-eta-omega-C\]. In Sec. \[sub-eta-time\], we study the decay of the efficiency for long storage times. Dependence on the control intensity {#sub-eta-omega-C} ----------------------------------- ![ \[fig-efficiency-vs-power\] Write-read efficiency $\eta$ vs. Rabi frequency of the control light $\Omega_c$. All data were taken at the two-photon resonance. Data taken on the single-photon resonance $\Delta_c=0$ ($\circ$) display a clear maximum as a function of $\Omega_c$. Data taken at a single-photon detuning of $\Delta_c=2\pi\times70$ MHz ($\bullet$) display a lower maximum efficiency. The lines are fits of the simple model from appendix A to the data. ](fig5){width="0.9\columnwidth"} Figure \[fig-efficiency-vs-power\] shows the experimentally observed dependence of $\eta$ on the Rabi frequency of the control light $\Omega_c$. Data were taken for a storage time of 1 $\mu$s with an incoming probe pulse that has a Gaussian intensity profile $$\begin{aligned} \label{I-in} I_\mathrm{in}(t,z) = I_0 \exp\left(-\frac{1}{2\tau_p^2} \left(t -\frac{z}{c} \right)^2 \right) .\end{aligned}$$ Here, $I_0$ is the peak intensity and $\tau_p$ is the temporal rms width of the intensity. Data in Fig. \[fig-efficiency-vs-power\] were taken for $\tau_p=94$ ns. The control light was turned off $t_0=230$ ns after the maximum probe intensity entered the medium. Data taken at the single-photon resonance $\Delta_c=0$ ($\circ$) show a maximum of $\eta$ at $\Omega_c\sim 2\pi\times 20$ MHz. This value of $\Omega_c$ agrees fairly well with the prediction of the simple model developed in appendix A. The observed maximum efficiency of $\eta\sim 30$ %, however, is a factor of $\sim2$ lower than the expectation from the simple model. We attribute this discrepancy to the simplicity of the model and to experimental issues, such as inaccuracies in the determination of the experimental parameters. Note that because of the results of Fig. \[fig-efficiency-vs-power\], the data in Figs. \[fig-Faraday\] and \[fig-alpha\] were taken at $\Omega_c=2\pi\times 20$ MHz, where $\eta$ is maximized. We find experimentally that the write-read efficiency $\eta$ is increased if we slowly decrease the intensity of the control beam while the probe pulse enters the medium. This observation agrees with a more rigorous optimization of $\eta$ for a homogeneous medium in Ref. [@novikova:07]. Hence, all data in Fig. \[fig-efficiency-vs-power\] were taken with a linear ramp of the intensity of the control beam applied, with a ramp speed that is experimentally found to maximize $\eta$. This ramp is not included in our simple model. For storage, the control light is on for 300 ns because prior to this, there is no probe light inside the medium. The horizontal axis in the figure shows the Rabi frequency corresponding to the time-averaged value of the control intensity during this 300 ns control pulse. In a measurement independent from Fig. \[fig-efficiency-vs-power\], we achieved a write-read efficiency of $\eta=(53\pm5)$ %. The gain in efficiency compared to the data in Fig. \[fig-efficiency-vs-power\] resulted from two changes in the experimental procedure. First, we removed the filter cavity and the mode-filtering fiber after the BEC and, second, we truncated the Gaussian input probe pulse in time, such that it misses exactly that part of its falling edge that cannot be stored anyway because it reaches the BEC after $\Omega_c$ is already ramped to zero. For reasons discussed in the introduction, our experiments in Ref. [@lettner:11] had to be operated at a single-photon detuning of $\Delta_c=2\pi\times70$ MHz. An investigation of $\eta$ at this detuning was therefore necessary. Experimental results ($\bullet$) are shown in Fig. \[fig-efficiency-vs-power\]. These data display a maximum value of $\eta \sim 20$ % at $\Omega_c\sim 2\pi\times 20$ MHz. The value of $\Omega_c$ at which the maximum occurs is essentially identical to the data at the single-photon resonance, whereas the maximum efficiency is further reduced. The physical origin of this reduction is discussed in appendix \[sub-app-detuning\]. The lines in Fig. \[fig-efficiency-vs-power\] show fits to the data, based on the simple model developed in appendix A. More precisely, the dash-dotted line from Fig. \[fig-efficiency-theory\] was taken and two fit parameters were introduced, each representing a linear scaling, one for $\eta$ and one for $\Omega_c$. Dependence on the storage time {#sub-eta-time} ------------------------------ Our experiments in Ref. [@lettner:11] also required an investigation of the time scale on which $\eta$ decays during storage. Thermal motion is known to be the limiting physical effect in many experiments. Using a BEC or an optical lattice, however, thermal motion can be suppressed drastically, resulting in a very slow decay of $\eta(t_\mathrm{store})$ [@schnorrberger:09; @zhang:09; @Dudin:10]. Unlike those experiments, our experiment does not use co-propagating probe and control beams. Instead, the level scheme shown in Fig. \[fig-levels\] requires the two beams to propagate perpendicularly to each other. The resulting differential photon recoil is much larger than for co-propagating beams. In our experiment, the lifetime of $\eta(t_\mathrm{store})$ is predominantly limited by this recoil, similar to Ref. [@ginsberg:07]. As a result of the photon recoil, atoms in hyperfine states $F=1$ and $F=2$ move relative to each other. If after $t_\mathrm{store}$ these two atomic clouds do not overlap any more, the retrieval does not produce a directed beam and hardly any signal reaches the detector. In our experiment, the control and probe beams propagate along the $y$ and $z$ axes, respectively. Hence, the differential photon recoil incurred in the Raman transition is directed in the $yz$ plane, where the BEC is symmetric with Thomas-Fermi radii $R_y=R_z=25$ $\mu$m. The single-mode fiber between the BEC and the detector poses an additional constraint, also related to the photon recoil. Not only do the two atomic clouds need to overlap, the emitted light must also match the transverse mode of the single-mode fiber, resulting in a spatial filtering in the $xy$ plane. The fiber is pretty well mode matched to the incoming probe beam, which has a beam waist of $w=8$ $\mu$m, thus setting a length scale for the spatial filtering that is more stringent along $y$ than the Thomas-Fermi radius $R_y$. To obtain a simple estimate for $\eta(t_\mathrm{store})$, we approximate the medium as homogeneous, which is justified by $w\ll R_y$ and $w\ll R_z$. In this approximation, only the recoil along the $y$ axis is relevant, i.e. the photon recoil of the probe laser is irrelevant. As the mode of the optical fiber has a Gaussian transverse profile, the decay of $\eta(t_\mathrm{store})$ is expected to be Gaussian $$\begin{aligned} \label{eta-Gauss} \eta(t_\mathrm{store}) = \eta(0) e^{-t_\mathrm{store}^2/2\sigma_\eta^2}\end{aligned}$$ with a $e^{-1/2}$ time $\sigma_\eta=mw/\sqrt2 \hbar k_c=1.0$ ms, where $m$ is the atomic mass and $k_c=\omega_c/c$ is the wave vector of the control beam. Note that the factor $\sqrt2$ here has nothing to do with the modulus of the differential photon recoil. ![ \[fig-EtaVsTimePureBEC\] Decay of the write-read efficiency $\eta$ as a function of storage time in a pure BEC. The line shows a Gaussian fit according to Eq. , yielding a best-fit value $\sigma_\eta=0.48$ ms for the $e^{-1/2}$ time of $\eta$. The dominant mechanism that sets this time scale is given by spatial filtering caused by the single-mode fiber combined with photon recoil incurred during the storage process. ](fig6){width="0.9\columnwidth"} Figure \[fig-EtaVsTimePureBEC\] shows experimental data for $\eta(t_\mathrm{store})$, recorded for an essentially pure BEC and a single-photon detuning of $\Delta_c=2\pi\times 70$ MHz. The line shows a fit of Eq.  to the data, yielding a best-fit value of $\sigma_\eta=0.48$ ms. This is a factor of $\sim 2$ less than expected, probably owing to the simplicity of the model. Note that the decay of $\eta$ observed here is unproblematic for our experiment in Ref. [@lettner:11] where we took data only for $t_\mathrm{store}\leq300$ $\mu$s. Furthermore, the decay of $\alpha$ observed in Fig. \[fig-alpha\] is much slower than the decay of $\eta$ observed in Fig. \[fig-EtaVsTimePureBEC\]. This means that at long storage times, only very little light is retrieved but it still has the correct polarization. Achieving a much slower decay of $\eta$ during storage would be possible when using co-propagating or almost co-propagating beams, as mentioned above. This would be incompatible with the present atomic level scheme. But a conversion of the polarization qubit into different wave vectors of the control light, as in Ref. [@Zhang:11], could solve this problem. However, our experiments in Ref. [@lettner:11] would not immediately profit from a slower decay of $\eta$ for two reasons. First, very long storage times would drastically slow down the rate at which write-read cycles can be repeated, which would result in an unrealistically long data acquisition time for the complete experiment, due to low count rates. Second, in the setup used in Ref. [@lettner:11], the limiting factor when extending the storage time was the deterioration of the fidelity due to magnetic field noise acting on the single atom inside the high-finesse cavity. ![ \[fig-EtaVsTimeBimodal\] (Color online) Decay of the normalized write-read efficiency $\eta$ in the presence of a noticeable uncondensed fraction of the gas. Thermal motion causes a rapid initial decay of $\eta$. The decay settles to the long-lived level of $\eta$ that is caused by the BEC fraction. The data sets were taken for different BEC fractions and, correspondingly, for different temperatures. ](fig7){width="0.9\columnwidth"} To illustrate how our experiments profit from the use of a BEC, we deliberately prepare an atomic gas with a noticeable uncondensed fraction. Figure \[fig-EtaVsTimeBimodal\] shows that in this case $\eta(t_\mathrm{store})$ decays much faster. More specifically, $\eta$ is the sum of two contributions, one from the BEC and one from the uncondensed fraction. These two contributions to $\eta$ decay on quite different time scales. On the time scale shown in Fig. \[fig-EtaVsTimeBimodal\] the contribution of the uncondensed fraction decays almost completely, whereas the contribution of the BEC is essentially constant. The overall decay of $\eta$ is sensitive to the first-order spatial coherence function of the gas [@naraschewski:99; @ginsberg:07]. A bimodal decay of the first-order coherence similar to Fig. \[fig-EtaVsTimeBimodal\] was previously observed in Ref. [@bloch:00] using a different technique. Comparing the data in Fig. \[fig-EtaVsTimeBimodal\] to the temperatures extracted from the size of the uncondensed fraction in time-of-flight images, we confirm that $\lambda_\mathrm{dB}/v_\mathrm{rel}$ can be used as a coarse estimate for the time scale of the decay of $\eta$ caused by the uncondensed fraction of the gas. Here, $\bm v_\mathrm{rel}=\hbar (\bm k_p-\bm k_c)/m$ is the velocity of the $F=2$ atoms relative to the $F=1$ atoms during $t_\mathrm{store}$, $\bm k_p$ is the probe-light wave-vector, $\lambda_{dB}=\sqrt{2\pi\hbar^2/mk_BT}$ is the thermal de-Broglie wavelength, $T$ is the temperature, and $k_B$ is the Boltzmann constant. The data in Fig. \[fig-EtaVsTimeBimodal\] were taken with a waist of the probe beam of $w=30$ $\mu$m and after removing the filter cavity as well as the single-mode fiber between the BEC and the detector. Without these modifications, $\eta$ would be sensitive only to the central region of the gas, where the uncondensed fraction contributes less. Conclusion ========== To conclude, we characterized and optimized the BEC as a quantum memory and showed that a write-read efficiency above 50 % can be reached. Its decay over storage time results from the differential photon recoil in the Raman transfer combined with spatial filtering of the retrieved light. This could be mitigated using co-propagating light beams. We also showed that the mapping between photonic and atomic qubit has an impressive average process fidelity. Its decay over storage time is due to magnetic-field noise and is suppressed by appropriate techniques. We thank E. Figueroa and S. Ritter for discussions. This work was supported by the German Excellence Initiative through the Nanosystems Initiative Munich and by the Deutsche Forschungsgemeinschaft through SFB 631. Modeling the efficiency {#app} ======================= In this appendix, we derive a simple estimate for the write-read efficiency $\eta$. This estimate was used for fitting to the experimental data in Fig. \[fig-efficiency-vs-power\]. To set the stage for this calculation, Sec. \[sub-app-theory\] briefly summarizes the theoretical background from the literature. We then develop a simple, largely analytic model for $\eta$ in Sec. \[sub-app-simple\]. In Sec. \[sub-app-detuning\], we discuss why a large single-photon detuning reduces $\eta$. Theoretical background {#sub-app-theory} ---------------------- Our notation largely follows Ref. [@fleischhauer:05], except for the sign of all detunings. We denote the probe and control Rabi frequencies as $\Omega_p$ and $\Omega_c$ and the corresponding detunings as $\Delta_p=\omega_p-\omega_{p,\mathrm{res}}$ and $\Delta_c=\omega_c-\omega_{c,\mathrm{res}}$, where $\omega_p$ and $\omega_c$ are the angular frequencies of the light fields and $\omega_{p,\mathrm{res}}$ and $\omega_{c,\mathrm{res}}$ are the corresponding atomic resonances, respectively. We consider the regime of small probe intensity and neglect dephasing. The propagation of the probe light can then be described by the linear susceptibility [@fleischhauer:05] $$\begin{aligned} \label{chi} \chi = \chi_0 \frac{2\delta_2 \Gamma}{\Omega_c^2-4\delta_2(\Delta_c+\delta_2)-2i\delta_2\Gamma}\end{aligned}$$ with $$\begin{aligned} \chi_0 = \frac{n_\mathrm{gr}\Omega_c^2}{\omega_p \Gamma}\end{aligned}$$ and the group index [@fleischhauer:05] $$\begin{aligned} n_\mathrm{gr} = \frac{\Gamma_p}{\Omega_c^2} \varrho \sigma c .\end{aligned}$$ Here $\delta_2=\Delta_p-\Delta_c$ is the two-photon detuning, $\Gamma$ is the total decay rate of the excited state, $\Gamma_p$ is the partial decay rate of the excited state into the ground state involved in the probe transition, $c$ is the vacuum speed of light, $\varrho$ is the particle density, and $\sigma=3\lambda_p^2/2\pi$ is the resonant light scattering cross section for a cycling transition at wavelength $\lambda_p=2\pi c/\omega_p$. If the two-photon detuning $\delta_2$ is small, then one can expand ${\rm Re}(\chi)$ and ${\rm Im}(\chi)$ to lowest non-vanishing order in $\delta_2$, yielding \[chi-approx\] $$\begin{aligned} \label{Re-chi-approx} \mathrm{Re}(\chi) &=& \chi_0 \frac{2\Gamma}{\Omega_c^2} \delta_2 +\mathcal O(\delta_2^2) ,\\ \label{Im-chi-approx} \mathrm{Im}(\chi) &=& \chi_0 \left( \frac{2\Gamma}{\Omega_c^2} \delta_2\right)^2 +\mathcal O(\delta_2^3) .\end{aligned}$$ Note that this is independent of $\Delta_c$. The group velocity for probe light can be calculated from Eq. , yielding [@fleischhauer:05] $$\begin{aligned} \label{v-gr} v_\mathrm{gr} = \frac{c}{1+n_\mathrm{gr}} .\end{aligned}$$ We consider a medium that extends from $z=0$ to $z=L$ with a density $\varrho(z)$ which varies along the propagation direction $z$ of the probe beam. If $\Omega_c$ is constant in time, the pulse delay after propagation through the complete medium follows immediately from Eq. , yielding [@fleischhauer:05] $$\begin{aligned} \label{tau-d} \tau_d(L) = \int_0^L dz \frac{n_\mathrm{gr}(z)}{c} = \frac{\Gamma}{\Omega_c^2} d_p(L) ,\end{aligned}$$ where $$\begin{aligned} \label{d-p} d_p (L)= \int_0^L dz \frac{\Gamma_p}{\Gamma} \sigma \varrho(z)\end{aligned}$$ denotes the optical depth seen by the probe light. Irreversible absorption of the probe light inside the medium can be a serious issue. This is avoided if all relevant frequency components of the probe light are close to the two-photon resonance, $\delta_2=0$. Under this condition, Eq.  yields a Gaussian EIT intensity transmission window in frequency space with rms width $\sigma_\mathrm{trans}=\Delta\omega_\mathrm{trans}/\sqrt8$ with [@fleischhauer:05] $$\begin{aligned} \label{Delta-omega-trans} \Delta\omega_\mathrm{trans}(L) = \frac{\Omega_c^2}{\Gamma \sqrt{d_p(L)}} .\end{aligned}$$ If $\Delta \omega_p$ denotes the typical width of the frequency spectrum of the probe pulse, then the condition for small absorption reads $\Delta \omega_p\ll \Delta\omega_\mathrm{trans}(L)$. Due to the Fourier limit, the typical duration $\tau_p$ of the pulse is related to $\Delta\omega_p$ by $\tau_p \Delta\omega_p \sim 1$. Combining this with Eqs.  and , the condition for small absorption can be rewritten as $$\begin{aligned} \label{tau-tau} \frac{\tau_d(L)}{\tau_p} \ll \sqrt{d_p(L)} .\end{aligned}$$ Obviously, fully compressing the pulse longitudinally into the medium requires $\tau_d(L)/\tau_p>1$. If one simultaneously wants to avoid absorption, then according to Eq.  one needs $d_p(L)\gg 1$. The requirement of large optical depth is independent of $\Omega_c$. But in the experiment, $\Omega_c$ must be adapted to the values of $\tau_p$ and $d_p(L)$, as discussed now. Simple estimate for the efficiency {#sub-app-simple} ---------------------------------- While various numerical models for a thorough analysis of $\eta$ have been published, we find it useful to complement these elaborate models with a much simpler model that captures only part of the physics but gives a quick estimate for the efficiency. Our model assumes that the probe pulse is Gaussian in the time domain. Using Eq. , the intensity of the propagating probe pulse can be approximated as $$\begin{aligned} \label{I-t-z} I(t,z) = I_0(z) \exp\left(-\frac{1}{2\tau_p^2(z)} \left(t - \tau_d(z) -\frac{z}{c} \right)^2 \right) ,\quad\end{aligned}$$ where $I_0$ is the peak intensity, $\tau_p$ is the temporal rms width of the intensity, and $\tau_d$ is given in Eq. . Note that Eq.  is valid for all $z$. It simplifies to Eq.  for those values of $z$ where the probe light has not yet entered the medium. Our simple model separately addresses the issues of insufficient pulse compression and irreversible absorption. First, we ignore irreversible absorption. This makes $I_0$ and $\tau_p$ independent of $z$. We assume that $\Omega_c$ is switched off at a time $t_0$ and assume that the fraction of the light that is inside the medium at this moment is stored and subsequently retrieved. This yields a write-read efficiency $$\begin{aligned} \eta_\mathrm{comp} = \frac12 \left[ \mathrm{erf} \left( \frac{t_0}{\sqrt2 \tau_p} \right) - \mathrm{erf} \left( \frac{t_0-\tau_d(L)-L/c}{\sqrt2 \tau_p} \right) \right] \nonumber \\ \label{eta-comp}\end{aligned}$$ with the error function $\mathrm{erf}(x) =(2/\sqrt\pi) \int_0^x du \exp(-u^2)$. This result for the efficiency quantifies how well the pulse is compressed into the medium. Note that $n_\mathrm{gr}\gg1$ implies $\tau_d(L)\gg L/c$. Second, we turn to irreversible absorption. Here, we consider a situation in which $\Omega_c$ is constant in time, implying that no storage takes place. The fraction $\eta_\mathrm{trans}$ of the pulse energy that is transmitted through the medium is calculated easily in the frequency domain. Based on Eqs.  and , we obtain $$\begin{aligned} \label{eta-trans} \eta_\mathrm{trans} = \left( 1+ \frac{2}{(\tau_p \Delta\omega_\mathrm{trans}(L))^2} \right)^{-1/2}\end{aligned}$$ with $\Delta\omega_\mathrm{trans}$ from Eq. . This result for the efficiency expresses the issue of irreversible absorption. ![ \[fig-efficiency-theory\] (Color online) Simple theoretical estimate for the write-read efficiency $\eta$ vs. Rabi frequency of the control light $\Omega_c$. The dashed, dotted, and solid lines show $\eta_\mathrm{comp}$, $\eta_\mathrm{trans}$, and $\eta_\mathrm{comp}\times \eta_\mathrm{trans}$, respectively. Parameters are $\tau_p=94$ ns, $t_0=230$ ns, $1/\Gamma=26$ ns, and $d_p(L)=127$. The pronounced maximum in the solid line arises because for small $\Omega_c$, the EIT transmission window in frequency space is too narrow for the incoming probe pulse, whereas for large $\Omega_c$, the group velocity is not sufficiently reduced to spatially compress the complete pulse into the BEC. The dash-dotted line shows the result of averaging over the transverse profile of the BEC. ](fig8){width="0.9\columnwidth"} To obtain a simple estimate for the overall efficiency, which must take both effects into account, we simply multiply the two efficiencies from Eqs.  and . For the atomic probe transition used in our experiment, we have $\Gamma_p=\Gamma/12$. Combination with the atom number and trap frequencies quoted in Sec. \[sub-impl-BEC\] yields a peak value of $d_p(L)=127$ at $x=y=0$. Results for this optical depth and typical parameters of our experimental pulses are shown in Fig. \[fig-efficiency-theory\]. In the model, we explored $\eta$ as a function of the two-dimensional parameter space spanned by $\Omega_c$ and $t_0$. Fig. \[fig-efficiency-theory\] shows the dependence on $\Omega_c$ only for that value of $t_0$, for which the global maximum of $\eta$ is reached. A modification of the pulse duration $\tau_p$ would require a re-optimization of $\Omega_c$ and $t_0$. Considering Eqs.  and , one finds that Eqs.  and remain unchanged if the scalings $\Omega_c\propto1/\sqrt{\tau_d}$ and $t_0\propto\tau_p$ are used. As a consequence, the maximum value of $\eta(\Omega_c,t_0)$ is insensitive to a change in $\tau_d$. The transverse inhomogeneity of the BEC can be accounted for by calculating a weighted average of $\eta$, with the transverse profile of the probe light intensity as a weighting function $$\begin{aligned} \int dx dy \frac{2}{\pi w^2} e^{-2(x^2+y^2)/w^2} \eta(d_p(x,y)) . \end{aligned}$$ We assume a Thomas-Fermi parabola for $\varrho(x,y,z)$ with Thomas-Fermi radii $R_x$, $R_y$, and $R_z$. Calculation of $d_p(x,y)$ by analytic integration over $z$ is straightforward. After a transformation to new coordinates $(\tilde\rho,\varphi)$ with $x=R_x\tilde\rho\cos\varphi$ and $y=R_y\tilde\rho\sin\varphi$, the integral over $\varphi$ can also be solved analytically. The remaining integral over $\tilde\rho$ is easily computed numerically. The result is shown as a dash-dotted line in Fig. \[fig-efficiency-theory\]. This line predicts a maximum of $\eta\sim60$ % at $\Omega_c=2\pi\times15$ MHz. For this value of $\Omega_c$ and for $x=y=0$, the theory yields estimated values of $\chi_0=0.5$, $n_\mathrm{gr} = 5\times 10^6$, $\tau_d(L) = 550$ ns, and $\Delta\omega_\mathrm{trans}(L) = 2\pi\times 3.3$ MHz. ![ \[fig-chi\] (Color online) Linear susceptibility $\chi$ as a function of the two-photon detuning $\delta_2$. The solid and dotted lines show the predictions of Eq.  for the real and imaginary parts of $\chi$ at $\Omega_c = 3.3 \Gamma = 2\pi\times 20$ MHz. (a) At the single photon resonance, $\Delta_c=0$. (b) At a single photon detuning of $\Delta_c=11.4 \Gamma = 2\pi\times 70$ MHz. Dash-dotted lines (red) show the lowest-order approximations of Eq. . The range where these approximations become poor sets an upper bound for the frequency range which is useful for light storage. This range is much narrower in (b) than in (a). This explains the reduction of the efficiency observed at $\Delta_c=2\pi\times 70$ MHz. Note that the scales on the horizontal axes differ by one order of magnitude. ](fig9){width="0.9\columnwidth"} The absorption represents a filter in frequency space. Due to the Fourier limit, this causes an increase of the temporal pulse width $\tau_p$ for increasing $z$, thus complicating a more rigorous calculation of $\eta_\mathrm{comp}$. We can overestimate the effect of this increase of $\tau_p(z)$ when using the constant value $\tau_p(L)$ instead of $\tau_p(0)$ in calculating $\eta_\mathrm{comp}$. We find that for the parameters of Fig. \[fig-efficiency-theory\], this has little effect. Our model also neglects that $\tau_p(z)$ should increase due to dispersion caused by $\mathrm{Re}(\chi)$. This increase is given by $\tau_p^2(z)=\tau_p^2(0)+[\tau_d(z)/\tau_p(0)\omega_p]^2$, which is negligible for the parameters of our experiment. The simple model developed here neglects that the spatial cutting during storage broadens the frequency spectrum of the pulse, thus increasing the absorption after retrieval. The model also neglects that the transverse inhomogeneity of the medium might cause a deformation of the wavefronts, resulting in effects such as focussing of the probe beam. In addition, this model is fully based on Eq. , instead of Eq. . If the probe pulse is so broad in frequency space that Eq.  is not a good approximation, then the applicability of the model is questionable. Nonetheless, the prediction of the model agrees reasonably well with our experimental data in Fig. \[fig-efficiency-vs-power\]. Reduced efficiency at large single-photon detuning {#sub-app-detuning} -------------------------------------------------- To understand the physical origin of the reduction of $\eta$ at large single-photon detuning, we investigate the frequency range that is useful for storing light. An upper limit for this frequency range is set by the frequency range within which Eq.  is a good approximation. Using Eq. , one can easily show that the maxima of $\mathrm{Im}(\chi)$ are located at $$\begin{aligned} \delta_2 = \frac12\left(-\Delta_c \pm \sqrt{\Delta_c^2 + \Omega_c^2} \right) .\end{aligned}$$ The maximum nearest to $\delta_2=0$ clearly sets an overoptimistic upper bound for the useful frequency range for light storage. For $|\Delta_c|\gg \Omega_c$, the nearest maximum lies at $\delta_2 \approx \Omega_c^2/4 \Delta_c$. Comparison with Eq.  shows that the useful frequency range is much narrower than $\Delta\omega_\mathrm{trans}$ unless $4|\Delta_c|/\Gamma \sqrt{d_p}\ll 1$. Our experiment is operated at $4\Delta_c/\Gamma \sqrt{d_p} \approx 4$ so that this issue is obviously relevant. This reduction of the useful frequency range is to be contrasted with the delay $\tau_d$ in Eq.  which is independent of $\Delta_c$. As a result, the overall efficiency is reduced. To further illustrate this point, we show the dependence of $\chi$ on $\delta_2$ in Fig. \[fig-chi\]. Parts (a) and (b) correspond to $\Delta_c=0$ and $\Delta_c=2\pi\times 70$ MHz, respectively. Note the different scales on the horizontal axes. This figure clearly illustrates that the frequency range over which Eq.  is a good approximation differs drastically between the two cases. The light pulses that we store at $\Delta_c=2\pi\times 70$ MHz have the same spectral width as for $\Delta_c=0$. For $\Delta_c=2\pi\times 70$ MHz a considerable part of the frequency components of the light therefore samples the frequency range where Eq.  is not a good approximation. Fig. \[fig-chi\](b) shows that for components with negative $\delta_2$, the value of $d\mathrm{Re}(\chi)/d\delta_2$ is reduced, resulting in a faster group velocity, which is disadvantageous. On the other hand, for positive $\delta_2$ absorption can be substantial and $d\mathrm{Re}(\chi)/d\delta_2$ can even change sign, thus not creating slow light. These problems qualitatively explain the reduced write-read efficiency that we observe experimentally for $\Delta_c=2\pi\times 70$ MHz. [10]{} A. I. Lvovsky, B. C. Sanders, and W. Tittel, Nat. Photon. [**3**]{}, 706 (2009). M. Fleischhauer, A. Imamoglu, and J. P. Marangos, Rev. Mod. Phys. [**77**]{}, 633 (2005). C. Liu, Z. Dutton, C. H. Behroozi, and L. V. Hau, Nature [**409**]{}, 490 (2001). A. V. Turukhin, V. S. Sudarshanam, M. S. Shahriar, J. A. Musser, B. S. Ham, and P. R. Hemmer, Phys. Rev. Lett. [**88**]{}, 023602 (2001). D. F. Phillips, A. Fleischhauer, A. Mair, R. L. Walsworth, and M. D. Lukin, Phys. Rev. Lett. [**86**]{}, 783 (2001). T. Chanelière, D. N. Matsukevich, S. D. Jenkins, S.-Y. Lan, T. A. B. Kennedy, and A. Kuzmich, Nature [**438**]{}, 833 (2005). M. D. Eisaman, A. André, F. Massou, M. Fleischhauer, A. S. Zibrov, and M. D. Lukin, Nature [**438**]{}, 837 (2005). D. N. Matsukevich, T. Chanelière, S. D. Jenkins, S.-Y. Lan, T. A. B. Kennedy, and A. Kuzmich, Phys. Rev. Lett. [**96**]{}, 030405 (2006). H. Tanji, S. Ghosh, J. Simon, B. Bloom, and V. Vuletić, Phys. Rev. Lett. [**103**]{}, 043601 (2009). K. S. Choi, H. Deng, J. Laurat, and H. J. Kimble, Nature [**452**]{}, 67 (2008). Y.-W. Cho and Y.-H. Kim, Opt. Express [**18**]{}, 25786 (2010). H. Zhang, X.-M. Jin, J. Yang, H.-N. Dai, S.-J. Yang, T.-M. Zhao, J. Rui, Y. He, X. Jiang, F. Yang, G.-S. Pan, Z.-S. Yuan, Y. Deng, Z.-B. Chen, X.-H. Bao, S. Chen, B. Zhao, and J.-W. Pan, Nat. Photon. [**5**]{}, 628 (2011). D. G. England, P. S. Michelberger, T. F. M. Champion, K. F. Reim, K. C. Lee, M. R. Sprague, X.-M. Jin, N. K. Langford, W. S. Kolthammer, J. Nunn, and I. A. Walmsley, e-print <arXiv:1112.0900>. E. Saglamyurek, N. Sinclair, J. Jin, J. A. Slater, D. Oblak, F. Bussières, M. George, R. Ricken, W. Sohler, and W. Tittel, Nature [**469**]{}, 512 (2011). C. Clausen, I. Usmani, F. Bussières, N. Sangouard, M. Afzelius, H. de Riedmatten, and N. Gisin, Nature [**469**]{}, 508 (2011). M. Lettner, M. Mücke, S. Riedl, C. Vo, C. Hahn, S. Baur, J. Bochmann, S. Ritter, S. Dürr, and G. Rempe, Phys. Rev. Lett. [**106**]{}, 210503 (2011). A. V. Gorshkov, A. André, M. D. Lukin, and A. S. Sørensen, Phys. Rev. A [**76**]{}, 033805 (2007). J. Nunn, I. A. Walmsley, M. G. Raymer, K. Surmacz, F. C. Waldermann, Z. Wang, and D. Jaksch, Phys. Rev. A [**75**]{}, 011401 (2007). O. S. Mishina, N. V. Larionov, A. S. Sheremet, I. M. Sokolov, and D. V. Kupriyanov, Phys. Rev. A [**78**]{}, 042313 (2008). A. S. Sheremet, L. V. Gerasimov, I. M. Sokolov, D. V. Kupriyanov, O. S. Mishina, E. Giacobino, and J. Laurat, Phys. Rev. A [**82**]{}, 033838 (2010). K. F. Reim, J. Nunn, V. O. Lorenz, B. J. Sussman, K. C. Lee, N. K. Langford, D. Jaksch, and I. A. Walmsley, Nat. Photon. [**4**]{}, 218 (2010). L. V. Hau, S. E. Harris, Z. Dutton, and C. H. Behroozi, Nature [**397**]{}, 594 (1999). M. Fleischhauer and M. D. Lukin, Phys. Rev. A [**65**]{}, 022314 (2002). M. Fleischhauer and M. D. Lukin, Phys. Rev. Lett. [**84**]{}, 5094 (2000). A. Marte, T. Volz, J. Schuster, S. D[ü]{}rr, G. Rempe, E. G. M. van Kempen, and B. J. Verhaar, Phys. Rev. Lett. [**89**]{}, 283202 (2002). E. Hecht, [*Optics*]{} (Addison Wesley, Reading, 2001). C. Brosseau, [*Fundamentals of Polarized Light, a Statistical Optics Approach*]{} (Wiley, New York, 1998). M. A. Nielsen and I. L. Chuang, [*Quantum Computation and Quantum Information*]{} (University Press, Cambridge, 2000). M. D. Bowdrey, D. K. L. Oi, A. J. Short, K. Banaszek, and J. A. Jones, Phys. Lett. A [**294**]{}, 258 (2002). I. Novikova, A. V. Gorshkov, D. F. Phillips, A. S. Sørensen, M. D. Lukin, and R. L. Walsworth, Phys. Rev. Lett. [**98**]{}, 243602 (2007). U. Schnorrberger, J. D. Thompson, S. Trotzky, R. Pugatch, N. Davidson, S. Kuhr, and I. Bloch, Phys. Rev. Lett. [**103**]{}, 033003 (2009). R. Zhang, S. R. Garner, and L. V. Hau, Phys. Rev. Lett. [**103**]{}, 233602 (2009). Y. O. Dudin, R. Zhao, T. A. B. Kennedy, and A. Kuzmich, Phys. Rev. A [**81**]{}, 041805 (2010). N. S. Ginsberg, S. R. Garner, and L. V. Hau, Nature [**445**]{}, 623 (2007). M. Naraschewski and R. J. Glauber, Phys. Rev. A [**59**]{}, 4595 (1999). I. Bloch, T. W. Hänsch, and T. Esslinger, Nature [**403**]{}, 166 (2000).

ArXiv

--- abstract: 'We present a measurement of the abundance of Fe relative to H in the solar corona using a technique which differs from previous spectroscopic and solar wind measurements. Our method combines EUV line data from the CDS spectrometer on [*SOHO*]{} with thermal bremsstrahlung radio data from the VLA. The coronal Fe abundance is derived by equating the thermal bremsstrahlung radio emission calculated from the EUV Fe line data to that observed with the VLA, treating the Fe/H abundance as the sole unknown. We apply this technique to a compact cool active region and find Fe/H = 1.56 $\times$ 10$^{-4}$, or about 4 times its value in the solar photosphere. Uncertainties in the CDS radiometric calibration, the VLA intensity measurements, the atomic parameters, and the assumptions made in the spectral analysis yield net uncertainties $\sim$ 20%. This result implies that low first ionization potential elements such as Fe are enhanced in the solar corona relative to photospheric values.' author: - 'S. M. White, R. J. Thomas, J. W. Brosius & M. R. Kundu' nocite: '[@Fel92]' title: The Absolute Abundance of Iron in the Solar Corona --- INTRODUCTION ============ It seems remarkable that the absolute abundances (here taken to mean abundances relative to hydrogen) of trace elements in the solar corona are still a matter for debate when many measurements for much more distant stars exist. The reason for this situation is easy to understand. Classic spectroscopic techniques for measuring relative abundances require that we compare the amount of radiation emitted in spectral lines whose atomic physics we understand and which arise from the two different elements in question. In the solar corona H and He are fully ionized, and hence produce no lines. In stars with hotter coronae, thermal bremsstrahlung produces a continuum emission which can be used instead of a spectral line as a measure of the amount of H present, but the Sun’s quiescent corona is so cool that the continuum present at X–ray or EUV wavelengths is relatively weak and difficult to measure. It has been known for some time that there is a difference in the relative abundances of elements in the photosphere and the corona organized according to the first ionization potential (FIP) of the element: high–FIP elements such as O are underabundant in the corona relative to low–FIP elements such as Fe by a factor of 4 [@Mey85]. However, it has not been clear whether low–FIP elements are enhanced in the corona in an absolute sense (i.e., as a fraction of the total ion number density, dominated by H and He) or the high–FIP elements are depleted relative to H. Two main techniques have been used to infer absolute abundances in the solar corona [@FSH99]. One can observe solar flares, which produce plasma hot enough for the continuum to be measured, and compare the continuum with a suitable set of lines; or one can measure particles directly in the solar wind, and assume that the abundances there truly reflect the abundances in the solar corona. Unfortunately, these two techniques have resulted in conflicting measurements for, e.g., the absolute abundance of iron. The flare measurements suggest that it is close to the absolute abundance in the photosphere [@VeP81] or perhaps factors of 1.5 – 2 larger [@FlS99], while the solar wind particle measurements suggest that it is 4 times the photospheric ratio [@Rea92]. In this paper we use a third technique to measure the absolute abundance $N_{\rm Fe}/N_{\rm H}$ which has several desirable features. We use EUV observations of lines of Fe in order to determine the amount of Fe present in the corona, but to measure the amount of H present we use radio continuum measurements of thermal bremsstrahlung from coronal plasma. This measurement therefore pertains to the quiescent solar corona, including non-flaring weak active regions, rather than flare or solar wind plasma. Like classic spectroscopic techniques, this technique should be independent of the filling factor of plasma in the volume studied since both types of emission have the same dependence on number density. We describe the measurement technique and the observations used for this measurement in the following section. We then discuss the implications of our measurement. THE MEASUREMENT TECHNIQUE ========================= The method proposed here to measure the abundances of elements relative to H rests on the fact that the EUV line emission and radio bremsstrahlung both depend on the product of the electron number density and the number density of an element. In the case of radio bremsstrahlung the element is effectively hydrogen; in the case of an EUV line, it is the element responsible for the line. Therefore comparison of the two yields the abundance of the latter element relative to hydrogen. Specifically, since most EUV lines produced in the solar corona are collisionally excited and optically thin, the line intensity at a photon energy $E$ corresponding to a transition in ionization state Z of element X is the integral over temperature of the product $Q_{X,Z,E}\,DEM(T)\,{N_X/N_{\rm H}}$ [@FSH99], where $Q_{X,Z,E}$ is the “contribution function”: it includes the intrinsic spontaneous emission rate (the Einstein “A” coefficient) as well as the ratio $N_{X,Z}/N_X$ of the number density of the particular charge state $Z$ of the element $X$ responsible for the line to the total number density of the element. This ratio is usually determined by an ionization balance calculation appropriate to the temperature of the plasma, and $Q$ is a strong function of temperature for this reason. However, for most lines $Q$ is effectively independent of $N_X$ and $N_e$. The (hydrogen column) differential emission measure DEM(T) is the integral along the line of sight of ${d(N_e\,N_{\rm H})/dT}$. Bremsstrahlung is one of the two main mechanisms for radio emission in the solar corona and also has a significant dependence on temperature: in the optically thin limit, the flux at a frequency $f$ is $$S \ = 9.78\,\times\,10^{-3}\,{2\,k_B \over c^2} (1 + 4 {N_{\rm He} \over N_{\rm H}}) \ \int \int \ T^{-0.5}\,DEM(T)\,G(T) \ dT \,d\Omega$$ in ergs cm$^{-2}$ s$^{-1}$ Hz$^{-1}$ [@Dul85], where $G(T)$ = 24.5 - ln(${T/f}$) is the Gaunt factor, $k_B$ is the Boltzmann constant, $c$ the speed of light and the second integral is over the solid angle subtended by the source. We are assuming that in a plasma at the temperature of the solar corona the ion component is completely dominated by protons and fully–ionized helium, and we adopt $N_{\rm He}/N_{\rm H}\,=\,0.1$. It is clear from (1) that we must know the DEM in order to calculate the thermal bremsstrahlung radio flux accurately. Note that as long as $N_X/N_{\rm H}$ does not vary from place to place, (1) has exactly the same dependence on density as the EUV line emissivity and thus filling factor has exactly the same effect on the EUV line and radio continuum fluxes. We determine the abundance $N_{\rm Fe}/N_{\rm H}$ with the following steps: (i) We measure the radio flux from a source in the solar corona. This source [*must be optically thin*]{} and its radio flux [*must be due only to bremsstrahlung*]{}. The latter requirement can be tested by observing the radio spectrum of the source, which should be essentially flat for optically thin bremsstrahlung. (ii) We observe the same region with the Coronal Diagnostic Spectrometer (CDS) on [*SOHO*]{} in order to measure the intensities of lines from a range of charge states of Fe sufficient to give a reliable determination of the DEM. (iii) We use the CHIANTI atomic data package [@DLM97; @LLD99] to fit the Fe line fluxes to a DEM. (iv) We assume that the shape of the DEM for electrons and protons is the same as that for Fe, and integrate over temperature in (1) to predict the optically–thin bremsstrahlung radio flux we expect to observe, for a given absolute abundance of Fe. Comparison with the radio flux measured by the VLA yields $N_{\rm Fe}/N_{\rm H}$. Coordinated observations with CDS and the VLA[^1] were carried out on 1997 July 27, Aug. 3 and Nov. 11. We will present an analysis of all the data elsewhere; here we discuss the best target observed, a small active region AR 8105 found at about S35W15 on 1997 Nov. 11 (with the VLA in “D” configuration). This was the best target in the sense that it was compact: the VLA is an interferometer and cannot measure the flux of large sources, so radio flux measurements are more reliable for compact sources. The region was one of four observed on this day with the VLA at 1.4, 4.8 and 8.4 GHz. CDS also observed the same 4 regions alternately using the sequence FE\_IONS which reads nine 21–pixel–wide windows from the NIS1 detector of CDS containing 14 identified lines including those of Fe X, XI, XII, XIII, XIV, XV, XVI and XVII. One run of the sequence took 32 minutes, scanning with the 4 slit across a 240 wide field of view. Five sequences were acquired on AR 8105 centered at about 13:14, 16:28, 18:42, 20:56 and 23:10 UT, respectively. The VLA observations covered the period from 14:30 UT to 23:10 UT, cycling between the different frequencies rapidly in order to achieve optimal $uv$ coverage. Images of AR 8105 acquired by CDS, the VLA and EIT are shown in Figure 1. THE ABUNDANCE OF IRON ===================== It is essential for this measurement that the VLA see the same plasma that is seen in the EUV, and Fig. 1 shows that this is indeed the case: the EUV and radio images match each other extremely well, indicating that gyroresonance opacity does not contribute significantly to the radio emission and we conclude that it is due to bremsstrahlung. The higher–resolution EIT images shows that two east–west loops dominated AR 8105. The VLA observations resulted in images with resolutions of 30at 1.4 GHz, 13 at 4.8 GHz and 7 at 8.4 GHz. Using the 4.8 and 8.4 GHz images we identified an area of dimension 69 $\times$ 97 around AR 8105 which includes essentially all the flux from the region. Radio fluxes were derived by summing over this area at 4.8 and 8.4 GHz, while due to the poorer resolution at 1.4 GHz we instead fitted a Gaussian model to the feature associated with AR 8105 and determined its flux. At each frequency we have independent measurements in two separate 50 MHz bands and in the two circular polarizations, all of which are consistent with one another. The resulting fluxes are 0.25 $\pm$ 0.01 sfu at 1.4 GHz, 0.207 $\pm$ 0.004 sfu at 4.8 GHz and 0.197 $\pm$ 0.006 sfu at 8.4 GHz, which are consistent with the $f^{-0.1}$ spectrum expected for optically thin bremsstrahlung due to the Gaunt factor in (1). The quoted uncertainties are the map rms per beam times the square root of the number of beams over which integration was carried out, and are much larger than the formal uncertainties in the fluxes. Polarizations were also low, consistent with the properties of bremsstrahlung. The CDS data were calibrated using the standard reduction package in SOLARSOFT (results presented here use the version current in 1999 December). Line profiles were fitted and fluxes extracted, with uncertainties calculated according to @Tho98. We used both a recent radiometric CDS calibration derived from comparison with the SERTS rocket flight in 1997 [@TTK00] and the “version 2” CDS calibration which was in the CDS software as of 1999 December. One important step is that background subtraction was carried out on the CDS data using a linear fit to pixels to the north and south of the region studied, well away from AR 8105. This step is important because of the fact that the VLA does not measure large–scale flux and hence is insensitive to the quiet–Sun contribution. In effect we assume that a smooth quiet–Sun contribution has been subtracted from the radio fluxes, and hence it should be subtracted from the CDS data as well. This predominantly affects the cooler lines, amounting to up to 20% of the photons in the original spectrum, but being fairly uniform across the CDS field of view. The following lines (with typical line fluxes averaged over the region shown in parentheses, in ergs cm$^{-2}$ s$^{-1}$ st$^{-1}$) were then used to calculate the DEM: Fe X at 345.7 Å (55 $\pm$ 3), Fe XI at 352.7 Å (140 $\pm$ 4), Fe XII at 352.1 Å (145 $\pm$ 4) and 364.4 Å (230 $\pm$ 4), Fe XIII at 348.3 Å (95 $\pm$ 4), Fe XIV at 334.2 Å (260 $\pm$ 4), Fe XVI at 360.7 Å (380 $\pm$ 4) and Fe XVII (undetected here) at 350.5 Å. These lines were chosen to be as density–insensitive as possible while still covering the temperature range expected for the non–flaring solar corona, log T = 5.8 up to 6.8. The Fe XIII and Fe XII lines do have significant sensitivity to density in the likely range of coronal densities, 10$^9$ – 10$^{10}$ cm$^{-3}$. In practice, AR 8105 was quiet and there is no significant plasma outside the temperature range 5.9 – 6.4. The fitting procedure uses bins of width 0.1 in log T; in effect, we find that only 4 bins contain significant emission measure and the adjacent two bins are also significant, so we are fitting to 6 parameters using 8 lines from 7 different charge states. The Fe XV line at 327.1 Å was not used to determine the DEM because it was always found to be $\sim$ 30% lower than the models predicted; the atomic data for it are clearly not consistent with the strong Fe XIV and XVI lines observed. Additional details will be given in a companion paper. We adopted the ionization equilibrium calculations of @ArR92 and assume a constant pressure $nT$ of 1 $\times$ 10$^{16}$ cm$^{-3}$ K (e.g., $N$ = 5 $\times$ 10$^9$ cm$^{-3}$ at $T$ = 2 $\times$ 10$^6$ K) in determining the DEM. The line fluxes used for the fit were generally reproduced by the resulting model DEM to better than 10% (SERTS) or 5% (“version 2”). Figure 2 shows the geometric mean DEM over the 5 CDS sequences with error bars showing the standard deviation in each temperature bin. Essentially no plasma is present below log T = 6.0 (controlled by the ratio of Fe X to Fe XI) while the amount of plasma above 2 $\times$ 10$^6$ K is small (controlled by the ratio of Fe XVI to Fe XIV). These properties are consistent with those of DEMs derived for other regions of similarly low activity [@BDT96]. Integrating over each of the 5 DEMs resulted in a predicted radio flux of 0.829 $\pm$ 0.023 sfu in the 6700 arcsec$^{2}$ area at 4.8 GHz if we use the CDS calibration based on the SERTS–97 comparison and adopt $N_{\rm Fe}/N_{\rm H}$ = 3.9 $\times$ 10$^{-5}$ as in @Mey85, which is close to the photospheric value. The observed radio flux of 0.207 sfu is a factor of 4.0 smaller than this. Consequently we derive a value of 1.56 $\times$ 10$^{-4}$ for $N_{\rm Fe}/N_{\rm H}$, or $\log_{10}{A_{\rm Fe}}$ = 8.19 on the commonly used scale where $\log_{10}{A_{\rm H}}$ = 12.00. Using the “version 2” CDS calibration we predict 0.716 $\pm$ 0.013 sfu and $N_{\rm Fe}/N_{\rm H}$ = 1.35 $\times$ 10$^{-4}$ ($\log_{10}{A_{\rm Fe}}$ = 8.13). We adopt the more recent SERTS–97 calibration for our final result. The uncertainty of 3% for the predicted radio flux is the standard error in the mean of the results of the five distinct CDS observations of AR 8105 and thus incorporates the effects of both temporal variability over the 12 hours of the experiment as well as the stochastic nature of the process of fitting the line fluxes to a DEM. We adopt an uncertainty of 15% for the CDS radiometric calibration [@TTK00]. It turns out that the measured value for $N_{\rm Fe}/N_{\rm H}$ is reasonably robust to several of the assumptions made here, mostly because it depends on an integration over the DEM and is thus not very sensitive to details of the DEM. One assumption is that the result quoted uses the version of CHIANTI (0.9) used by the CDS team at NASA/GSFC as of 1999 December, which is not the most current version (2.0). For the Fe lines used here the only differences between the two versions appear to be in Fe X and Fe XI, and the newer version results in significantly poorer fits for the DEM because of these two lines: the resulting DEMs have less emission measure at low temperatures and more at high temperatures. Notwithstanding this fact, the result for the abundance differs by less than 3% with the newer version of CHIANTI. If we use the ionization equilibrium calculations of @ArR85, the DEMs show a pronounced peak at $\log{T}$ = 6.3 and the resulting abundance is about 10% larger. A change in the assumed pressure by a factor of 2 affects the results by a similar amount: it would be 8% smaller at 5.0 $\times$ 10$^{15}$ cm$^{-3}$ K and 10% larger at 2.0 $\times$ 10$^{16}$ cm$^{-3}$ K. We attempt to minimize the uncertainty due to variation in pressure across the active region by choosing a value representative of the source region for the bulk of the EUV line emission. We note also that the background subtraction carried out on the CDS data introduces some uncertainty in the final results, as does solar calibration at the VLA, which uses dedicated noise sources for solar calibration which need to be measured routinely to maintain calibration accuracy. However, the calibration at each of the three frequencies observed is independent of the others, so we take the fact that the measured spectrum is flat to be an indication that solar calibration was not in error by more than the 20% levels which all the other factors may introduce. CONCLUSIONS =========== Our measurement of an absolute abundance for Fe which is 4 times photospheric is clearly consistent with the interpretation that low–FIP elements such as Fe are enhanced relative to their photospheric abundances. Of the commonly–used solar abundance tables, Feldman’s (1992) is closest to our measurement with $\log_{10}{A_{\rm Fe}}$ = 8.10; his value is based more on the argument that H should be treated as a high–FIP element rather than any specific measurement of $N_{\rm Fe}/N_{\rm H}$. Other commonly used values for $\log_{10}{A_{\rm Fe}}$ are the photospheric results 7.60 [@All73], 7.51 [@AnG89] and 7.67 [@GrA91], the coronal results 7.59 [@Mey85], 7.83 [@FlS99] and 8.50 [@WMD94], and the solar wind measurements 7.93 [@Rea95] and 8.51 [@Rea99]. We note in passing that spectroscopic measurements of the Fe abundance in active stars with coronae much hotter than the Sun’s have tended to suggest abundances smaller than photospheric [@MKW96; @Dra96], in contrast to the solar behaviour we find here. We have reported here a single measurement of the Fe abundance. Spectroscopic studies of relative abundances suggest that they may vary considerably from flare to flare and from one region to the next [@SLM84; @Sab95; @SLB98], so it is clearly important to confirm our measurement. We have additional data sets which will be published in a longer paper elsewhere, but with larger uncertainties expected in the radio fluxes, and expect to carry out further studies of this type in the future. Acknowledgments {#acknowledgments .unnumbered} =============== This research at the University of Maryland was carried out primarily under a SOHO GI grant from NASA, NAG 5–4954. Solar radiophysics at the University of Maryland is also supported by NSF grant ATM 96-12738 and NASA grants NAG 5-7370, NAG 5-6257 and NAG 5-7901. We thank Richard Harrison, Andrzej Fludra and the CDS operators for their assistance with the project, and Barry Clark of NRAO for scheduling the VLA observations to match those of SOHO. We thank the CHIANTI consortium for their efforts to make atomic data readily accessible. [25]{} natexlab\#1[\#1]{} Allen, C. W. 1973, Astrophysical Quantities (London: Athlone Press) Anders, E. & Grevesse, N. 1989, Geochim. Cosmochim. Acta, 53, 197 Arnaud, M. & Raymond, J. 1992, ApJ, 398, 394 Arnaud, M. & Rothenflug, R. 1985, A&AS, 60, 425 Brosius, J. W., Davila, J. M., Thomas, R. J., & Monsignori-Fossi, B. C. 1996, ApJS, 106, 143 Dere, K. P., Landi, E., , Mason, H. E., [Monsignori Fossi]{}, B. C., & Young, P. R. 1997, A&AS, 125, 149 Drake, S. A. 1996, in Cosmic Abundances, ed. S. S. Holt & G. Sonneborn (San Francisco: Astron. Soc. Pacific), 215 Dulk, G. A. 1985, ARA&A, 23, 169 Feldman, U. 1992, Physica Scripta, 46, 202 Fludra, A., Saba, J. L. R., Henoux, J.-C., Murphy, R. J., Reames, D. V., Lemen, J. R., Strong, K. T., Sylwester, J., & Widing, K. G. 1999, in The Many Faces of the Sun, ed. K. T. Strong, J. L. R. Saba, B. M. Haisch, & J. T. Schmelz (Berlin: Springer Verlag), 89 Fludra, A. & Schmelz, J. T. 1999, A&A, 348, 286 Grevesse, N. & Anders, E. 1991, in Solar Interior and Atmosphere, ed. A. N. Cox, W. C. Livingston, & M. S. Matthews (Tucson: University of Arizona Press), 1227 Landi, E., Landini, M., Dere, K. P., Young, P. R., & Mason, H. E. 1999, A&AS, 135, 339 Mewe, R., Kaastra, J., White, S. M., & Pallavicini, R. 1996, A&A, 315, 170 Meyer, J.-P. 1985, ApJS, 57, 173 Reames, D. V. 1992, in Coronal Streamers, Coronal Loops and Coronal and Solar Wind Composition, ed. V. Domingo (Noordwijk: ESA SP 348), 315 Reames, D. V. 1995, Adv. Space Res., 15, 41 —. 1999, ApJ, 518, 473 Saba, J. L. R. 1995, Adv. Space Res., 15, no. 7, 13 Sylwester, J., Lemen, J. R., Bentley, R. D., Fludra, A., & Zolcinski, M. 1998, ApJ, 501, 397 Sylwester, J., Lemen, J. R., & Mewe, R. 1984, Nat., 310, 665 Thomas, R. J., Thompson, W. T., Kent, B. J., & Hollandt, J. 2000, ApJ, in preparation Thompson, W. T. 1998, CDS Software Note 49 Veck, N. J. & Parkinson, J. H. 1981, MNRAS, 197, 41 Waljeski, K., Moses, D., Dere, K. P., Saba, J. L. R., Strong, K. T., Webb, D. F., & Zarro, D. M. 1994, ApJ, 429, 909 [^1]: The Very Large Array is run by the National Radio Astronomy Observatory which is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc.

ArXiv

--- author: - 'A. H. Córsico$^{1,2}$, L. G. Althaus$^{1,2}$, and M. M. Miller Bertolami$^{1,2,3}$' title: 'Erratum: Asteroseismological constraints on the pulsating planetary nebula nucleus (PG1159-type) RXJ2117.1+3412' --- [ $^1$ Facultad de Ciencias Astronómicas y Geofísicas, Universidad Nacional de La Plata, Paseo del Bosque S/N, (1900) La Plata, Argentina.\ $^2$ Instituto de Astrofísica La Plata, IALP, CONICET-UNLP\ $^3$ Max-Planck-Institut für Astrophysik, Garching, Germany\ ]{} We correct one error which appeared in the paper [*Asteroseismological constraints on the pulsating planetary nebula nucleus (PG1159-type) RXJ2117.1+3412*]{}, A&A, 461, 1095 (2007), by A. H. Córsico, L. G. Althaus, M. M. Miller Bertolami & K. Werner (DOI: 10.1051/0004-6361:20066452), regarding the age of expansion of the planetary nebula of   ($t_{\rm PN}$). The correct value is $t_{\rm PN}\sim 2.71 \times 10^4$ yr, instead of the published value ($t_{\rm PN}\sim 5.43 \times 10^4$ yr). Note that the evolutionary timescale of the central star of  is $t_* \sim 2.5 \times 10^4$ yr. Thus, by considering the correct expansion age, we definitively solve the discrepancy between the  evolutionary timescale and the size of the nebula. The notable agreement between $t_{\rm PN}$ and $t_*$ reinforces the validity of our asterosesimological model for .

ArXiv

--- abstract: 'We discuss the statistical mechanics of a system of self-gravitating fermions in a space of dimension $D$. We plot the caloric curves of the self-gravitating Fermi gas giving the temperature as a function of energy and investigate the nature of phase transitions as a function of the dimension of space. We consider stable states (global entropy maxima) as well as metastable states (local entropy maxima). We show that for $D\ge 4$, there exists a critical temperature (for sufficiently large systems) and a critical energy below which the system cannot be found in statistical equilibrium. Therefore, for $D\ge 4$, quantum mechanics cannot stabilize matter against gravitational collapse. This is similar to a result found by Ehrenfest (1917) at the atomic level for Coulombian forces. This makes the dimension $D=3$ of our universe very particular with possible implications regarding the anthropic principle. Our study enters in a long tradition of scientific and philosophical papers who studied how the dimension of space affects the laws of physics.' author: - 'Pierre-Henri Chavanis' title: 'Statistical mechanics and thermodynamic limit of self-gravitating fermions in $D$ dimensions' --- Laboratoire de Physique Théorique, Université Paul Sabatier,\ 118 route de Narbonne 31062 Toulouse, France.\ Introduction {#sec_introduction} ============ The statistical mechanics of systems with long-range interactions is currently a topic of active research [@dauxois]. Among long-range interactions, the gravitational force plays a fundamental role. Therefore, the developement of a statistical mechanics for self-gravitating systems is of considerable interest [@paddy]. In this context, a system of self-gravitating fermions enclosed within a box provides an interesting model which can be studied in great detail [@ht; @pt]. This model incorporates an effective small-scale cut-off played by the Pauli exclusion principle and a large scale cut-off played by the confining box (other forms of confinement could also be considered). The statistical mechanics of this system is rigorously justified and presents a lot of interesting features which are of interest in statistical mechanics [@houches] and astrophysics [@chavcape]. Its detailed study is therefore important at a conceptual and practical level. In a preceding paper [@pt], we have discussed the nature of phase transitions in the self-gravitating Fermi gas in a space of dimension $D=3$. Our study was performed in both microcanonical and canonical ensembles and considered an arbitrary degree of degeneracy relative to the system size. This study completes previous investigations by Hertel & Thirring [@ht] who worked in the canonical ensemble and considered small system sizes. At high temperatures and high energies, the system is in a gaseous phase and quantum effects are completely negligible. At some transition temperature $T_{t}$ or transition energy $E_{t}$ (for sufficiently large system sizes), a first order phase transition is expected to occur and drive the system towards a condensed phase. However, gaseous states are still metastable below this transition point and gravitational collapse will rather occur at a smaller critical temperature $T_{c}$ (Jeans temperature) [@aa] or critical energy $E_{c}$ (Antonov energy) [@antonov; @lbw; @paddy] at which the metastable branch disappears (spinodal point). The end-state of the collapse is a compact object with a “core-halo” structure. Typically, it consists of a degenerate nucleus surrounded by a “vapour”. The nucleus (condensate) resembles a white dwarf star [@chandra2]. At non-zero temperature, this compact object is surrounded by a dilute atmosphere. Therefore, when quantum mechanics is properly accounted for, there exists an equilibrium state (global maximum of entropy or free energy) for each value of accessible energy and temperature. The condensate results from the balance between gravitational contraction and quantum pressure. As first noticed by Fowler [@fowler] in his classical theory of white dwarf stars, quantum mechanics is able to stabilize matter against gravitational collapse. One object of this paper is to show that this conclusion is no more valid in a space of dimension $D\ge 4$. For a system of mass $M$ enclosed within a box of radius $R$, there exists a critical temperature (for sufficiently large $R$) and a critical energy below which the system cannot be found at statistical equilibrium. This is like the Antonov instability for self-gravitating classical particles in $D=3$ [@antonov; @lbw; @paddy] but it now occurs for fermions. Therefore, quantum mechanics cannot arrest gravitational collapse in $D\ge 4$. This result is connected to our previous observation [@langevin] that a classical white dwarf star (a polytrope of index $n_{3/2}=D/2$) becomes unstable for $D\ge 4$. Interestingly, this result is similar to that of Ehrenfest [@ehrenfest] who considered the stability of atomic structures (in Bohr’s model) for different dimensions of space and concludes that $D<4$ is required for stability. In this paper, we determine the caloric curve of the self-gravitating Fermi gas for an arbitrary dimension of space and an arbitrary degree of degeneracy (or system size). We exhibit particular dimensions that play a special role in the problem. The dimension $D=2$ is critical because the results established for $D\neq 2$ cannot be directly extended to $D=2$ [@sc1]. Furthermore, in $D=2$ the radius of a white dwarf star is independent on its mass and given in terms of fundamental constants by $R=0.27 \ h m^{-3/2} G^{-1/2}$. The dimension $D=4$ is also critical because it is the dimension at which classical white dwarf stars become unstable. At this particular dimension, their mass is independent on radius and can be expressed in terms of fundamental constants as $M=1.44\ 10^{-2}h^{4}m^{-5}G^{-2}$. Mathematically, this is similar to Chandrasekhar’s limiting mass [@chandra1] for relativistic white dwarf stars in $D=3$. The dimension $D=2(1+\sqrt{2})$ is also particular because at this dimension, the white dwarf stars cease to be self-confined and have infinite mass. Finally, $D=10$ is the dimension at which the caloric curve of classical isothermal spheres loses its characteristic spiral nature [@sc1]. Although we systematically explore all dimensions of space in order to have a complete picture of the problem, only dimensions $D=1$, $D=2$ and $D=3$ are [*a priori*]{} of physical interest. The dimension $D=1$ is considered in cosmology and in connexion with shell models, and the dimension $D=2$ can be useful to describe filaments or ring structures with high aspect ratio. Two-dimensional gravity is also of interest for its properties of conformal invariance and for its relation with two-dimensional turbulence [@houches]. Non-integer dimensions can arise if the system has a fractal nature. The paper is organized as follows. In Sec. \[sec\_fermions\], we determine the thermodynamical parameters of the self-gravitating Fermi gas in dimension $D$. The Fermi-Dirac entropy is introduced from a combinatorial analysis. In Sec. \[sec\_lim\], we consider asymptotic limits corresponding to the classical self-gravitating gas and to completely degenerate structures (white dwarfs). We emphasize the importance of metastable states in astrophysics and explain how they can be taken into account in the theory (see also [@meta]). We also discuss the thermodynamic limit of the self-gravitating quantum gas and compare it with the thermodynamic limit of the self-gravitating classical gas in the dilute limit [@vs]. In Sec. \[sec\_caloric\], we provide a gallery of caloric curves of the self-gravitating Fermi gas in different dimensions of space. Rigorous mathematical results on the existence of solutions of the Fermi-Poisson equation have been obtained by Stańczy [@stanczy]. Finally, in the conclusion, we place our study in a more general perspective. We give a short historical account of scientific and philosophical papers who studied the role played by the dimension of space in determining the form of the laws of physics. These works tend to indicate that the dimension $D=3$ of our universe is very particular. This is also the result that we reach in our study. These remarks can have implications regarding the anthropic principle. Thermodynamics of self-gravitating $D$-fermions {#sec_fermions} =============================================== The Fermi-Dirac distribution {#sec_fd} ---------------------------- We consider a system of $N$ fermions interacting via Newtonian gravity in a space of dimension $D$. We assume that the mass of the configuration is sufficiently small so as to ignore relativistic effects. Let $f({\bf r},{\bf v},t)$ denote the distribution function of the system, i.e. $f({\bf r},{\bf v},t)d^{D}{\bf r} d^{D}{\bf v}$ gives the mass of particles whose position and velocity are in the cell $({\bf r},{\bf v};{\bf r}+d^{D}{\bf r},{\bf v}+d^{D}{\bf v})$ at time $t$. The integral of $f$ over the velocity determines the spatial density $$\rho=\int f d^{D}{\bf v}, \label{fd1}$$ and the total mass of the configuration is given by $$M=\int \rho d^{D}{\bf r}, \label{fd2}$$ where the integral extends over the entire domain. On the other hand, in the meanfield approximation, the total energy of the system can be expressed as $$E={1\over 2}\int fv^{2}d^{D}{\bf r}d^{D}{\bf v}+{1\over 2}\int\rho\Phi d^{D}{\bf r}=K+W, \label{fd3}$$ where $K$ is the kinetic energy and $W$ the potential energy. The gravitational potential $\Phi$ is related to the density by the Newton-Poisson equation $$\Delta\Phi=S_{D} G\rho, \label{fd4}$$ where $S_{D}=2\pi^{D/2}/\Gamma(D/2)$ is the surface of a unit sphere in a space of dimension $D$ and $G$ is the constant of gravity (which depends on the dimension of space). We now wish to determine the most probable distribution of self-gravitating fermions at statistical equilibrium. To that purpose, we divide the individual phase space $\lbrace {\bf r},{\bf v}\rbrace$ into a very large number of microcells with size $(h/m)^D$ where $h$ is the Planck constant (the mass $m$ of the particles arises because we use ${\bf v}$ instead of ${\bf p}$ as a phase space coordinate). A microcell is occupied either by $0$ or $1$ fermion (or $g=2s+1$ fermions if we account for the spin). We shall now group these microcells into macrocells each of which contains many microcells but remains nevertheless small compared to the phase-space extension of the whole system. We call $\nu$ the number of microcells in a macrocell. Consider the configuration $\lbrace n_i \rbrace$ where there are $n_1$ fermions in the $1^{\rm st}$ macrocell, $n_2$ in the $2^{\rm nd}$ macrocell etc..., each occupying one of the $\nu$ microcells with no cohabitation. The number of ways of assigning a microcell to the first element of a macrocell is $\nu$, to the second $\nu -1$ etc. Since the particles are indistinguishable, the number of ways of assigning microcells to all $n_i$ particles in a macrocell is thus $${1\over n_i!}{\times} {\nu!\over (\nu-n_i)!}. \label{fd5}$$ To obtain the number of microstates corresponding to the macrostate $\lbrace n_i \rbrace$ defined by the number of fermions $n_i$ in each macrocell (irrespective of their precise position in the cell), we need to take the product of terms such as (\[fd5\]) over all macrocells. Thus, the number of microstates corresponding to the macrostate $\lbrace n_i \rbrace$, i.e. the probability of the state $\lbrace n_i \rbrace$, is $$W(\lbrace n_i \rbrace)=\prod_i {\nu!\over n_i!(\nu-n_i)!}. \label{fd6}$$ This is the Fermi-Dirac statistics. As is customary, we define the entropy of the state $\lbrace n_i \rbrace$ by $$S(\lbrace n_i \rbrace)=\ln W(\lbrace n_i \rbrace). \label{fd7}$$ It is convenient here to return to a representation in terms of the distribution function giving the phase-space density in the $i$-th macrocell $$f_i=f({\bf r}_i,{\bf v}_i)={n_i \ m\over \nu \ ({h\over m})^D}={n_i\eta_0\over \nu}, \label{fd8}$$ where we have defined $\eta_0=m^{D+1}/h^D$, which represents the maximum value of $f$ due to Pauli’s exclusion principle. Now, using the Stirling formula, we have $$\ln W(\lbrace n_i \rbrace)\simeq \sum_i \nu (\ln\nu-1)-\nu\biggl\lbrace {f_i\over \eta_0}\biggl\lbrack \ln\biggl ({\nu f_i\over \eta_0}\biggr )-1\biggr\rbrack +\biggl (1-{f_i\over \eta_0}\biggr )\biggl\lbrack\ln\biggl\lbrace \nu\biggl (1-{f_i\over \eta_0}\biggr )\biggr\rbrace-1\biggr\rbrack\biggr\rbrace. \label{fd9}$$ Passing to the continuum limit $\nu\rightarrow 0$, we obtain the usual expression of the Fermi-Dirac entropy $$S=-k_B\int \biggl\lbrace {f\over\eta_{0}}\ln {f\over\eta_{0}}+\biggl (1- {f\over\eta_{0}}\biggr)\ln \biggl (1- {f\over\eta_{0}}\biggr)\biggr\rbrace\ {d^{D}{\bf r}d^{D}{\bf v}\over ({h\over m})^D}. \label{fd10}$$ If we take into account the spin of the particles, the above expression remains valid but the maximum value of the distribution function is now $\eta_{0}=g m^{D+1}/h^{D}$, where $g=2s+1$ is the spin multiplicity of the quantum states (the phase space element has also to be multiplied by $g$). An expression of entropy similar to (\[fd10\]), but arising for a completely different reason, has been introduced by Lynden-Bell in the context of the violent relaxation of collisionless stellar systems [@lb; @cs; @dubrovnik]. In that context, $\eta_0$ represents the maximum value of the initial distribution function and the actual distribution function (coarse-grained) must always satisfy $\overline{f}\le \eta_0$ by virtue of the Liouville theorem. This is the origin of the “effective” exclusion principle in Lynden-Bell’s theory, which has nothing to do with quantum mechanics. Since the particles (stars) are distinguishable classical objects (but subject to an exclusion principle in the collisionless regime), Lynden-Bell’s statistics corresponds to a $4$-th form of statistics (in addition to the Maxwell-Boltzmann, Fermi-Dirac and Bose-Einstein statistics). However, for a single type of phase element $\eta_0$, it leads to the same results as the Fermi-Dirac statistics. We also recall that in the non-degenerate (or classical) limit $f\ll\eta_0$, the Fermi-Dirac entropy (\[fd10\]) reduces to the Boltzmann entropy $$S=-k_B\int {f\over m}\biggl\lbrack \ln\biggl ({f h^D\over g m^{D+1}}\biggr )-1\biggr \rbrack d^{D}{\bf r}d^{D}{\bf v}. \label{fd11}$$ Now that the entropy has been precisely justified, the statistical equilibrium state (most probable state) of self-gravitating fermions is obtained by maximizing the Fermi-Dirac entropy (\[fd10\]) at fixed mass (\[fd2\]) and energy (\[fd3\]): $${\rm Max}\quad S[f]\quad | \quad E[f]=E, M[f]=M. \label{fd12}$$ Introducing Lagrange multipliers $1/T$ (inverse temperature) and $\mu$ (chemical potential) to satisfy these constraints, and writing the variational principle in the form $$\delta S-{1\over T}\ \delta E+{\mu\over T} \delta N=0, \label{fd13}$$ we find that the [*critical points*]{} of entropy correspond to the Fermi-Dirac distribution $$f={\eta_{0}\over 1+\lambda e^{\beta m ({v^{2}\over 2}+\Phi)}}, \label{fd14}$$ where $\lambda=e^{-\beta \mu}$ is a strictly positive constant (inverse fugacity) and $\beta={1\over k_B T}$ is the inverse temperature. Clearly, the distribution function satisfies $f\le \eta_{0}$, which is a consequence of Pauli’s exclusion principle. So far, we have assumed that the system is isolated so that the energy is conserved. If now the system is in contact with a thermal bath (e.g., a radiation background) fixing the temperature, the statistical equilibrium state minimizes the free energy $F=E-TS$, or maximizes the Massieu function $J=S-\beta E$, at fixed mass and temperature: $${\rm Max}\quad J[f]\quad |\quad M[f]=M. \label{fd15}$$ Introducing Lagrange multipliers and writing the variational principle in the form $$\delta J+{\mu\over T} \delta N=0, \label{fd16}$$ we find that the [*critical points*]{} of free energy are again given by the Fermi-Dirac distribution (\[fd14\]). Therefore, the critical points (first variations) of the variational problems (\[fd12\]) and (\[fd15\]) are the same. However, the stability of the system (regarding the second variations) can be different in microcanonical and canonical ensembles. When this happens, we speak of a situation of [*ensemble inequivalence*]{} [@pt]. The stability of the system can be determined by a graphical construction, by simply plotting the caloric curve/series of equilibria $\beta(E)$ and using the turning point method of Katz [@katz; @katz2]. Thermodynamical parameters {#sec_para} -------------------------- Integrating the distribution function (\[fd14\]) over velocity, we find that the density of particles is related to the gravitational potential by $$\rho={\eta_{0}S_{D}2^{{D/2}-1}\over (\beta m)^{D/2}}I_{D/2-1}(\lambda e^{\beta m\Phi}), \label{p1}$$ where $I_{n}$ denotes the Fermi integral $$I_{n}(t)=\int_{0}^{+\infty}{x^{n}\over 1+t e^{x}}dx. \label{p2}$$ We recall the identity $$I'_{n}(t)=-{n\over t}I_{n-1}(t), \qquad (n>0), \label{p3}$$ which can be established from (\[p2\]) by an integration by parts. The gravitational potential is now obtained by substituting Eq. (\[p1\]) in the Poisson equation (\[fd4\]). We introduce the rescaled distance $\xi=\lbrack {S_{D}^{2} 2^{D/2-1}G\eta_{0}/(\beta m)^{D/2-1}}\rbrack^{1/2}r$ and the variables $\psi=\beta m (\Phi-\Phi_{0})$ and $k=\lambda e^{\beta m \Phi_{0}}$, where $\Phi_{0}$ is the central potential. Thus, we get the $D$-dimensional Fermi-Poisson equation $${1\over\xi^{D-1}}{d\over d\xi}\biggl (\xi^{D-1}{d\psi\over d\xi}\biggr )=I_{D/2-1}(ke^{\psi(\xi)}), \label{p4}$$ $$\psi(0)=\psi'(0)=0. \label{p5}$$ As is well-known, self-gravitating systems at non-zero temperature have the tendency to evaporate. Therefore, there is no equilibrium state in a strict sense and the statistical mechanics of self-gravitating systems is essentially an out-of-equilibrium problem. However, the evaporation rate is small in general and the system can be found in a quasi-equilibrium state for a relatively long time. In order to describe the thermodynamics of the self-gravitating Fermi gas rigorously, we shall use an artifice and enclose the system within a spherical box of radius $R$ (the box typically represents the size of the cluster under consideration). In that case, the solution of Eq. (\[p4\]) is terminated by the box at the normalized radius $$\alpha=\biggl \lbrack {S_{D}^{2} 2^{D/2-1} G\eta_{0}\over (\beta m)^{D/2-1}}\biggr \rbrack^{1/2} R. \label{p6}$$ For a spherically symmetric configuration, the Gauss theorem can be written $${d\Phi\over dr}={GM(r)\over r^{D-1}}, \label{p7}$$ where $M(r)=\int_{0}^{r}\rho S_{D}r^{D-1}dr$ is the mass within the sphere of radius $r$. Applying this result at $r=R$ and using the variables introduced previously we get $$\eta\equiv {\beta GMm\over R^{D-2}}=\alpha\psi'_{k}(\alpha). \label{p8}$$ This equation relates the dimensionless box radius $\alpha$ and the uniformizing variable $k$ to the dimensionless inverse temperature $\eta$. According to Eqs. (\[p6\]) and (\[p8\]), $\alpha$ and $k$ are related to each other by the relation $\alpha^{2}\eta^{D/2-1}=\mu$ or, explicitly, $$\alpha^{D+2\over D-2}\psi'_{k}(\alpha)=\mu^{2\over D-2}, \label{p9}$$ where $$\mu=\eta_{0}\sqrt{S_{D}^{4}2^{D-2}G^{D}M^{D-2}R^{D(4-D)}}, \label{p10}$$ is the degeneracy parameter [@cs]. It should not be confused with the chemical potential. We shall give a physical interpretation of this parameter in Sec. \[sec\_d3\]. The calculation of the energy is a little more involved. First, we introduce the local pressure $$p={1\over D}\int f v^{2}d^{D}{\bf v}. \label{p11}$$ Using the Fermi-Dirac distribution function (\[fd14\]), we find that $$p={\eta_{0}S_{D}2^{{D/2}}\over D(\beta m)^{D/2+1}}I_{D/2}(k e^{\psi}). \label{p12}$$ The kinetic energy $K=(D/2)\int p d^{D}{\bf r}$ can thus be written $${K R^{D-2}\over GM^{2}}={\alpha^{4+4D-D^{2}\over D-2}\over \mu^{4\over D-2}}\int_{0}^{\alpha}I_{D/2}(ke^{\psi_{k}(\xi)})\xi^{D-1}d\xi. \label{p13}$$ In order to determine the potential energy, we use the $D$-dimensional version of the Virial theorem [@langevin]. For $D\neq 2$, it reads $$2K+(D-2)W=DV_{D}R^{D}p(R), \label{p14}$$ where $V_{D}=S_{D}/D$ is the volume of a hypersphere with unit radius (the case $D=2$ will be considered specifically in Sec. \[sec\_d2\]). Using the expression of the pressure (\[p12\]) at the box radius $R$, we get $${W R^{D-2}\over GM^{2}}={2\over D(D-2)}{\alpha^{2(D+2)\over D-2}\over \mu^{4\over D-2}}I_{D/2}(k e^{\psi(\alpha)})-{2KR^{D-2}\over (D-2)GM^{2}}. \label{p15}$$ Combining Eqs. (\[p13\]) and (\[p15\]), we finally obtain $$\Lambda\equiv -{E R^{D-2}\over GM^{2}}={4-D\over D-2}{\alpha^{4+4D-D^{2}\over D-2}\over \mu^{4\over D-2}}\int_{0}^{\alpha}I_{D/2}(ke^{\psi_{k}(\xi)})\xi^{D-1}d\xi-{2\over D(D-2)}{\alpha^{2(D+2)\over D-2}\over \mu^{4\over D-2}}I_{D/2}(k e^{\psi(\alpha)}). \label{p16}$$ For $D=3$, Eqs. (\[p8\]) and (\[p16\]) return the expressions derived in [@cs; @pt]. For a given value of $\mu$ and $k$, we can solve the ordinary differential equation (\[p4\]) until the value of $\alpha$ at which the condition (\[p9\]) is satisfied. Then, Eqs. (\[p8\]) and (\[p16\]) determine the temperature and the energy of the configuration. By varying the parameter $k$ (for a fixed value of the degeneracy parameter $\mu$), we can determine the full caloric curve/series of equilibria $\beta(E)$. Extending the results of [@pt] in $D$ dimensions, the entropy of each configuration, parameterized by $\alpha$, is given by $${S\over Nk_{B}}=-{4+4D-D^{2}\over D(4-D)}\Lambda\eta+\psi_{k}(\alpha)+{\eta\over D-2}+\ln k-{2(D-2)\over D^{2}(4-D)}{\alpha^{2D\over D-2}\over \mu^{2\over D-2}}I_{D/2}(ke^{\psi_{k}(\alpha)}), \label{p17}$$ and the free energy by $$F=E-TS. \label{p18}$$ In the microcanonical ensemble, a solution is stable if it corresponds to a maximum of entropy $S[f]$ at fixed mass and energy. In the canonical ensemble, the condition of stability requires that the solution be a minimum of free energy $F[f]$ at fixed mass and temperature. This meanfield approach is [*exact*]{} in a thermodynamical limit such that $N\rightarrow +\infty$ with $\mu$, $\eta$, $\Lambda$ fixed. If we fix $\eta_{0}$ (i.e. $\hbar$) and $G$, this implies that $RN^{(D-2)/(D(4-D))}$, $TN^{-4/(D(4-D))}$, $EN^{-(4D-D^{2}+4)/(D(4-D))}$, $SN^{-1}$ and $JN^{-1}$ approach a constant value for $N\rightarrow +\infty$ (the free energy $F$ scales as $N^{(4D-D^{2}+4)/(D(4-D))}$). This is the quantum thermodynamic limit (QTL) for the self-gravitating gas [@pt; @rieutord]. The usual thermodynamic limit $N,R\rightarrow +\infty$ with $N/R^{D}$ constant is clearly not relevant for inhomogeneous systems whose energy is non-additive. Asymptotic limits {#sec_lim} ================= The non degenerate limit $(\mu=\infty)$ {#sec_class} --------------------------------------- Before considering the case of an arbitrary degree of degeneracy, it may be useful to discuss first the non degenerate limit corresponding to a classical isothermal gas ($\hbar\rightarrow 0$). For $f\ll\eta_{0}$, the distribution function (\[fd14\]) reduces to the Maxwell-Boltzmann formula $$f={\eta_{0}\over\lambda}e^{-\beta m({v^{2}\over 2}+\Phi)}, \label{cc1}$$ which can be written more conveniently as $$f=\biggl ({\beta m\over 2\pi}\biggr )^{D/2}\rho({\bf r})\ e^{-\beta m {v^{2}\over 2}}. \label{cc1bis}$$ The density profile can be written $$\rho=\rho_{0}e^{-\psi(\xi)}, \label{cc2}$$ where $\rho_{0}$ is the central density, $\xi$ is the normalized distance $$\xi=(S_{D} G\beta m\rho_{0})^{1/2}r, \label{cc3}$$ and $\psi$ is the solution of the $D$-dimensional Emden equation $${1\over \xi^{D-1}}{d\over d\xi}\biggl (\xi^{D-1}{d\psi\over d\xi}\biggr )=e^{-\psi}, \label{cc4}$$ with boundary conditions $$\psi(0)=\psi'(0)=0. \label{cc5}$$ This equation can be obtained from Eq. (\[p4\]) by taking the limit $k\rightarrow +\infty$ and using the limiting form of the Fermi integral $$I_{n}(t)\sim {1\over t}\Gamma(n+1), \qquad (t\rightarrow +\infty). \label{cc6}$$ From Eq. (\[cc1bis\]), we check that the local equation of state of a classical self-gravitating isothermal gas is $p({\bf r})={\rho({\bf r})\over m}k_{B}T$ whatever the dimension of space. The thermodynamical parameters are given by $$\eta=\alpha\psi'(\alpha), \label{cc7}$$ $$\Lambda={D(4-D)\over 2(D-2)} {1\over\alpha\psi'(\alpha)}-{1\over D-2}{e^{-\psi(\alpha)}\over\psi'(\alpha)^{2}}, \label{cc8}$$ $${S-S_{0}\over Nk_{B}}=-{D-2\over 2}\ln\eta-2\ln\alpha+\psi(\alpha)+{\eta\over D-2}-2\Lambda\eta, \label{cc9}$$ $${S_{0}\over Nk_{B}}=\ln\mu+\ln\biggl ({2\pi^{D/2}\over S_{D}}\biggr )+1-{D\over 2}, \label{cc9bis}$$ where $\alpha=(S_{D} G\beta m\rho_{0})^{1/2}R$ is the normalized box radius. For $D=2$, the thermodynamical parameters can be calculated analytically [@sc1]. Introducing the pressure at the box $P=p(R)$, the global equation of state of the self-gravitating gas can be written $${PV\over Nk_{B}T}={1\over D}{\alpha^{2}\over \eta}e^{-\psi(\alpha)}. \label{cc11}$$ We recall that the foregoing expressions can be expressed in terms of the value of the Milne variables $u_{0}=u(\alpha)$ and $v_{0}=v(\alpha)$ at the normalized box radius [@aa; @grand]. The structure and the stability of classical isothermal spheres in $D$ dimensions have been studied in detail in [@sc1]. The classical thermodynamic limit (CTL) of self-gravitating systems, or dilute limit [@vs], is such that $N\rightarrow +\infty$ with $\eta$, $\Lambda$ fixed. If we take $\beta\sim 1$, this implies that $R\sim N^{1/(D-2)}$ and $E,S,J,F\sim N$. The physical distinction between the quantum thermodynamic limit (QTL) and the classical thermodynamic limit (CTL) is related to the existence of long-lived gaseous metastable states as discussed in [@rieutord; @meta]. The completely degenerate limit {#sec_deg} ------------------------------- For $\beta\rightarrow +\infty$ (i.e., $T=0$), the distribution function (\[fd14\]) reduces to a step function: $f=\eta_{0}$ if $v\le v_{F}$ and $f=0$ if $v\ge v_{F}$, where $v_{F}({\bf r})=\sqrt{2(\mu/m-\Phi)}$ is the local Fermi velocity. In that case, the density and the pressure can be explicitly evaluated: $$\rho=\int_{0}^{v_{F}}\eta_{0}S_{D}v^{D-1}dv=\eta_{0}S_{D}{v_{F}^{D}\over D}, \label{d1}$$ $$p={1\over D}\int_{0}^{v_{F}}\eta_{0}S_{D}v^{D+1}dv=\eta_{0}{S_{D}\over D}{v_{F}^{D+2}\over D+2}. \label{d2}$$ Eliminating the Fermi velocity between these two expressions, we find that the equation of state of a cold Fermi gas in $D$ dimensions is $$p=K\rho^{1+2/D}, \qquad K={1\over D+2}\biggl ({D\over\eta_{0}S_{D}}\biggr )^{2/D}. \label{d3}$$ This equation of state describes a $D$-dimensional classical white dwarf star (throughout this paper, we shall call “white dwarf star”, or “fermion ball”, a completely degenerate self-gravitating system. This terminology will be extended to any dimension of space). In $D=3$, classical white dwarf stars are equivalent to polytropes with index $n=3/2$ [@fowler]. In $D$ dimensions, classical “white dwarf stars” are equivalent to polytropes with index [@langevin]: $$n_{3/2}={D\over 2}. \label{d4}$$ The structure and the stability of polytropic spheres in $D$ dimensions have been studied in detail in [@langevin]. It is shown that a polytrope of index $n$ is self-confined for $n<n_{5}=(D+2)/(D-2)$ and stable for $n<n_{3}=D/(D-2)$. Therefore, white dwarf stars ($n=n_{3/2}=D/2$) are self-confined only for $D<2(1+\sqrt{2})$ and they are stable only for $D\le 4$. For $D>4$, quantum mechanics is not able to stabilize matter against gravitational collapse. Thus, $D=4$ is a critical dimension regarding gravitational collapse. $D=2$ is also critical [@sc1]. Therefore, the dimension of space of our universe $2<D=3<4$ lies between two critical dimensions. We now introduce dimensionless parameters associated with $n_{3/2}$ polytropes which will be useful in the sequel. Their density profile can be written $$\rho(r)=\rho_{0}\ \theta^{D/2}(\xi), \label{d5}$$ where $\rho_{0}$ is the central density, $\xi$ is the normalized distance $$\xi=\biggl \lbrack {2S_{D}G\rho_{0}^{(D-2)\over D}\over K(D+2)}\biggr \rbrack^{1/2} r, \label{d6}$$ and $\theta$ is solution of the $D$-dimensional Lane-Emden equation $${1\over \xi^{D-1}}{d\over d\xi}\biggl (\xi^{D-1}{d\theta\over d\xi}\biggr )=-\theta^{D/2}, \label{d7}$$ with boundary conditions $$\theta(0)=1, \qquad \theta'(0)=0. \label{d8}$$ This equation can be obtained from Eq. (\[p4\]) by taking the limit $k\rightarrow 0$ and using the limiting form of the Fermi integral $$I_{n}(t)\sim {(-\ln t)^{n+1}\over n+1}, \qquad (t\rightarrow 0). \label{d9}$$ For $D<2(1+\sqrt{2})$, the solution of the Lane-Emden equation (\[d7\]) vanishes at a finite distance $\xi_{1}$ defining the radius $R_{*}$ of the white dwarf star ([*complete polytrope*]{}). Using the results of [@langevin], the mass-radius relation of $D$-dimensional white dwarf stars is given by $$M^{D-2\over D}R_{*}^{4-D}={K(D+2)\over 2 G S_{D}^{2/D}}\omega_{D/2}^{D-2\over D}, \label{d10}$$ where we have defined $$\omega_{D/2}=-\xi_{1}^{D+2\over D-2}\theta'(\xi_{1}). \label{d11}$$ For $2<D<4$, the mass $M$ decreases with the radius $R_{*}$ while for $D<2$ and for $4<D<2(1+\sqrt{2})$ it increases with the radius (see Fig. \[MRD\]). The mass-radius relation (\[d10\]) exhibits the two critical dimensions of space $D=2$ and $D=4$ discussed previously. For $D=2$, the radius is independent on mass and for $D=4$, the mass is independent on radius (see Sec. \[sec\_caloric\]). The energy of a self-confined white dwarf star is $$E=-\lambda_{D/2}{ GM^{2}\over R_{*}^{D-2}}, \label{d14}$$ where $$\lambda_{D/2}={D(4-D)\over (D-2)(4+4D-D^{2})}. \label{d15}$$ We note that the energy of a white dwarf star vanishes for $D=4$. According to Poincaré’s theorem [@chandra2], this determines the onset of instability. We thus recover the fact that complete white dwarf stars are unstable for $D>4$ [@langevin]. For $D>2(1+\sqrt{2})$, the density of a $n_{3/2}$ polytrope never vanishes (as $n_{3/2}>n_{5}$) and we need to confine the system within a box of radius $R$ ([*incomplete polytrope*]{}) to avoid the infinite mass problem. In that case, the white dwarf star exerts a pressure against the box. White dwarf stars with $R_{*}>R$ when $D<2(1+\sqrt{2})$ are also incomplete. They are arrested by the box at the normalized radius $\xi=\alpha$ with $\alpha=\lbrace 2S_{D}G\rho_{0}^{(D-2)/D}/\lbrack K(D+2)\rbrack\rbrace^{1/2}R$. As shown in [@langevin], the normalized mass and the normalized energy of the configuration parameterized by $\alpha$ are given by $$\eta_{P}\equiv {M\over S_{D}}\biggl\lbrack {2S_{D}G\over K(D+2)}\biggr \rbrack^{D\over D-2}{1\over R^{D(D-4)\over D-2}}=-\alpha^{D+2\over D-2}\theta'(\alpha), \label{d17}$$ $$\Lambda\equiv -{ER^{D-2}\over GM^{2}}={-2\over D^{2}-4D-4}\biggl\lbrace {D(4-D)\over 2(D-2)}\biggl \lbrack 1+(D-2){\theta(\alpha)\over\alpha\theta'(\alpha)}\biggr \rbrack+{2-D\over 2+D}{\theta(\alpha)^{D+2\over 2}\over \theta'(\alpha)^{2}}\biggr\rbrace. \label{d18}$$ In the present context, the normalized mass $\eta_P$ is related to the degeneracy parameter $\mu$ by the relation $$\eta_{P}=\biggl ({2\mu\over D}\biggr )^{2\over D-2}. \label{d19}$$ On the other hand, using Eqs. (\[d10\]) and (\[d14\]), the normalized mass and the normalized energy of a self-confined white dwarf star with $R_{*}<R$ (complete polytrope) are given by $$\eta_{P}=\omega_{D/2}\biggl ({R_{*}\over R}\biggr )^{(D-4)D\over D-2} \label{d20}$$ $$\Lambda=\lambda_{D/2}\biggl ({R\over R_{*}}\biggr )^{D-2}. \label{d21}$$ Eliminating $R_{*}$ between these two relations, we obtain the “mass-energy” relation $$\Lambda \eta_{P}^{(D-2)^{2}\over D(D-4)}=\lambda_{D/2} (\omega_{D/2})^{(D-2)^{2}\over D(D-4)}, \label{d22}$$ which will be useful in our subsequent analysis. Caloric curves in various dimensions {#sec_caloric} ==================================== Series of equilibria and metastable states {#sec_seq} ------------------------------------------ We shall now determine the caloric curve $\beta(E)$ of the self-gravitating Fermi gas as a function of the degeneracy parameter $\mu$ for any dimension of space $D$. This study has already been performed for $D=3$ in [@pt]. The critical points of the Fermi-Dirac entropy $S[f]$ at fixed $E$ and $M$ (i.e., the distribution functions $f({\bf r},{\bf v})$ which cancel the first order variations of $S$ at fixed $E$, $M$) form a series of equilibria parameterized by the uniformizing variable $k$. At each point in the series of equilibria corresponds a temperature $\beta$ and an energy $E$ determined by Eqs. (\[p8\]) and (\[p16\]). In this approach, $\beta$ is the Lagrange multiplier associated with the conservation of energy in the variational problem (\[fd13\]). It has also the interpretation of a kinetic temperature in the Fermi-Dirac distribution (\[fd14\]). We can thus plot $\beta(E)$ along the series of equilibria. There can be several values of temperature $\beta$ for the same energy $E$ because the variational problem (\[fd12\]) can have several solutions: a local entropy maximum (metastable state), a global entropy maximum, and one or several saddle points. We shall represent all these solutions on the caloric curve because local entropy maxima (metastable states) are in general more physical than global entropy maxima for the timescales achieved in astrophysics. Indeed, the system can remain frozen in a metastable gaseous phase for a very long time. This is the case, in particular, for globular clusters and for the gaseous phase of fermionic matter (at high energy and high temperature). The time required for a metastable gaseous system to collapse is in general tremendously long and increases exponentially with the number $N$ of particles (thus, $t_{life}\rightarrow +\infty$ in the thermodynamic limit $N\rightarrow +\infty$) [@meta]. This is due to the long-range nature of the gravitational potential. Therefore, metastable states are in reality stable states. At high temperatures and high energies, the global entropy maximum is not physically relevant [@ko; @ispolatov; @grand; @rieutord]. Condensed objects (e.g., planets, stars, white dwarfs, fermion balls,...) only form below a critical energy $E_{c}$ (Antonov energy) [@antonov; @lbw; @paddy] or below a critical temperature $T_{c}$ (Jeans temperature) [@aa], when the gaseous metastable phase ceases to exist (spinodal point). The case $2<D<4$ {#sec_d3} ---------------- We start to describe the structure of the caloric curve of the self-gravitating Fermi gas for $2<D<4$ (specifically $D=3$). Let us first consider the Fermi gas at $T=0$ (white dwarf stars). The $\Lambda-\eta_{P}$ curve defined by Eqs. (\[d17\]), (\[d18\]) and (\[d22\]) is represented in Fig. \[LHpD3\]. In the present context, it gives the energy of the star as a function of its mass. Since the curve does not present turning points, all the white dwarf star configurations are stable. According to Eq. (\[d10\]), for $2<D<4$, the mass $M$ of a complete white dwarf star is a decreasing function of its radius $R_*$. Therefore, if the system is enclosed within a box, there exists a characteristic mass $$M_*(R)={\chi_{D}\over \eta_{0}^{2\over D-2}G^{D\over D-2}}R^{-{D(4-D)\over D-2}} \label{gtq1}$$ such that for $M>M_*(R)$ the star is self-confined ($R_{*}<R$) and for $M<M_*(R)$, it is restricted by the box. In terms of the dimensionless mass $\eta_P$, complete $n_{3/2}$ polytropes correspond to $\eta_P\ge \omega_{D/2}$ and incomplete $n_{3/2}$ polytropes to $\eta_P\le \omega_{D/2}$. For $2<D<4$, there exists a stable equilibrium at $T=0$ for all mass $M$. We now briefly describe the caloric curve for arbitrary temperature and energy. A more complete description is given in [@pt] for $D=3$. First, we note that, according to Eqs. (\[p10\]), (\[d3\]) and (\[d10\]), $$\mu=\mu_*(D) \biggl ({R\over R_*}\biggr )^{D(4-D)\over 2}, \label{gtq2}$$ where $$\mu_{*}(D)\equiv {D\over 2}(\omega_{D/2})^{D-2\over 2}. \label{gtq3}$$ Therefore, the degeneracy parameter $\mu$ can be seen as the ratio (with some power) between the size of the system $R$ and the size $R_{*}$ of a white dwarf star with mass $M$. Accordingly, a small value of $\mu$ corresponds to a large “effective” cut-off (played by Pauli’s exclusion principle) or, equivalently, to a small system size. Alternatively, a large value of $\mu$ corresponds to a small “effective” cut-off or a large system size. This gives a physical interpretation to the degeneracy parameter. For $\mu\rightarrow +\infty$ (i.e. $\hbar\rightarrow 0$), we recover classical isothermal spheres. In that case, the caloric curve $\beta(E)$ forms a spiral. For finite values of $\mu$, the spiral unwinds due to the influence of degeneracy and gives rise to a rich variety of caloric curves (Fig. \[LHD3\]). For large systems, the caloric curve has a $Z$-shape (“dinosaur’s neck”) and for small systems it has a $N$-shape. The phase transitions in the self-gravitating Fermi gas for $D=3$ and the notion of metastable states, spinodal points, critical points, collapse, explosion, and hysteresis are discussed in [@pt; @ispolatov; @rieutord; @meta]. Similar notions are discussed in [@stahl] for a hard spheres gas. The ground state of the self-gravitating Fermi gas ($T=0$) corresponds to a white dwarf star configuration. For given $\mu$, its structure (radius, energy) is determined by the intersection between the $\Lambda-\eta_P$ curve in Fig. \[LHpD3\] and the line defined by Eq. (\[d19\]). The “white dwarf” is complete ($R_{*}<R$) for $\mu>\mu_{*}(D)$ and incomplete ($R_{*}>R$) otherwise. For $\mu>\mu_{*}(D)$, the normalized energy of the white dwarf is given by $$\Lambda_{max}(D,\mu)=\lambda_{D/2}\biggl ({\mu\over\mu_{*}}\biggr )^{2(D-2)\over D(4-D)}. \label{gtq4}$$ This is the ground state of the self-gravitating Fermi gas corresponding to the asymptote in Fig. \[LHD3bis\] (this asymptote exists for all curves in Fig. \[LHD3\] but is outside the frame). For classical particles ($\hbar=0$), there is no equilibrium state if energy and temperature are below a critical threshold [@antonov; @lbw]. In that case, the system undergoes gravitational collapse and forms binaries (in microcanonical ensemble) or a Dirac peak (in canonical ensemble); see Appendices A and B of [@sc1] and [@grand; @rieutord; @meta]. For self-gravitating fermions, an equilibrium state exists for all values of temperature and for all accessible energies ($E\ge E_{ground}$). Gravitational collapse is arrested by quantum pressure as first realized by Fowler [@fowler]. We shall now show that this claim ceases to be true in dimension $D\ge 4$. The case $4<D<2(1+\sqrt{2})$ {#sec_d41} ---------------------------- We now consider the case $4<D<2(1+\sqrt{2})$ (specifically $D=4.1$). Let us first describe the Fermi gas at $T=0$. The $\Lambda-\eta_{P}$ curve defined by Eqs. (\[d17\]), (\[d18\]) and (\[d22\]) is represented in Fig. \[LHpD4.1\]. For $D>4$, the curves $\eta_{P}(\alpha)$ and $\Lambda(\alpha)$ associated to $n_{3/2}$ polytropes have their extrema at the same point (see Appendix C of [@langevin]). Therefore, the $\Lambda$-$\eta_{P}$ curve presents a cusp at $(\Lambda_{0},\eta_{P,c})$. Past this point in the series of equilibria, $n_{3/2}$ polytropes are unstable. According to Eq. (\[d10\]), for $D>4$, the radius $R_*$ of a self-confined white dwarf star increases with its mass. For $M<M_*(R)$ there exists self-confined white dwarf star configurations. In terms of the dimensionless mass $\eta_P$, this corresponds to $\eta_P\le \omega_{D/2}$ (see Fig. \[LHpD4.1\]). However, such configurations are unstable since they lie after the turning point [@langevin]. Therefore, only incomplete (box confined) white dwarf stars can be stable in $D>4$. Inspecting Fig. \[LHpD4.1\] again, we observe that these configurations exist only below a critical mass $$M_c(R)=\eta_{P,c}(D) S_{D}R^{D(D-4)\over D-2}\biggl\lbrack {K(D+2)\over 2S_{D}G}\biggr \rbrack^{D\over D-2}. \label{gtq5}$$ For $M>M_c(R)$, there is no equilibrium state at $T=0$ for $D>4$. In terms of the dimensionless mass $\eta_P$, equilibrium states exist only for $\eta_{P}<\eta_{P,c}(D)$. The caloric curve for arbitrary value of temperature and energy is represented in Fig. \[LHD4.1\]. For $\mu\rightarrow +\infty$, we recover the classical spiral [@sc1]. For finite values of $\mu$, there exists equilibrium solutions at all temperatures only if $\eta_P<\eta_{P,c}(D)$. Using Eq. (\[d19\]), this corresponds to $$\mu<{D\over 2}\eta_{P,c}(D)^{D-2\over 2}\equiv \mu_{c}(D). \label{gtq6}$$ If $\mu>\mu_{c}(D)$, or equivalently if $M>M_*(R)$, there exists a minimum energy $E_c=-\Lambda_c GM^2/R^{D-2}$ (which appears to be positive) and a minimum temperature $T_c=GM/(\eta_c R^{D-2})$ below which there is no equilibrium state (the values of $\eta_{c}$ and $\Lambda_{c}$ depend on $D$ and $\mu$). In that case, the system is expected to collapse. This is similar to the Antonov instability (gravothermal catastrophe) for classical particles [@antonov; @lbw]. Since we deal here with self-gravitating fermions, we could expect that quantum pressure would arrest the collapse. Our study shows that this is not the case for $D>4$. Quantum mechanics cannot stabilize matter against gravitational collapse anymore. The case $D=4$ {#sec_d4} -------------- The dimension $D=4$ is special because it is the dimension of space above which quantum pressure cannot balance gravity anymore. Therefore, $D=4$ is critical and it deserves a particular attention. First, consider the Fermi gas at $T=0$. It corresponds to a polytrope of index $n_{3/2}=n_3$ [@langevin]. The $\Lambda-\eta_{P}$ curve defined by Eqs. (\[d17\]), (\[d18\]) and (\[d22\]) is represented in Fig. \[LHpD4\]. Since the curve is monotonic the box-confined $n_{3/2}$ polytropes are stable and the complete $n_{3/2}$ polytropes are marginally stable. For $D=4$, the mass of a self-confined white dwarf star is independent on its radius, see Eq. (\[d10\]). It can be expressed in terms of fundamental constants as $$M_{limit}={\omega_{2}\over g S_{4}^{2}}{h^{4}\over m^{5}G^{2}}\simeq 1.44\ 10^{-2}\ {h^{4}\over m^{5}G^{2}}, \label{gtq7}$$ where $\omega_{2}\simeq 11.2$ (we have taken $g=2$ in the numerical application). Mathematically, this is similar to Chandrasekhar’s limiting mass for relativistic white dwarf stars equivalent to $n=3$ polytropes in $D=3$ [@chandra1]. However, it is here a purely classical (i.e. nonrelativistic) result. Relativistic effects will be considered in a forthcoming paper [@relativity]. The energy of the self-confined white dwarf stars is $E=0$. Considering Fig. \[LHpD4\] again, we see that incomplete white dwarf stars exist only for $M<M_{limit}$. In terms of the dimensionless mass $\eta_P$, this corresponds to $\eta<\eta_{P,c}=\omega_{2}\simeq 11.2$. For $M>M_{limit}$, there is no equilibrium state at $T=0$. The caloric curve for arbitrary value of temperature and energy is represented in Fig. \[LHD4\] (see an enlargement in Fig. \[LHD4ZOOM\]). Its description is similar to that of Sec \[sec\_d41\]. For $M>M_{limit}$, or equivalently $\mu\ge \mu_{c}=2\omega_{2}\simeq 22.4.$ there exists a minimum energy $E_c=-\Lambda_c GM^2/R^{2}$ and a minimum temperature $T_c=GM/(\eta_c R^{2})$ below which there is no equilibrium state. The case $D\ge 2(1+\sqrt{2})$ {#sec_d48} ----------------------------- The caloric curves for $D\ge 2(1+\sqrt{2})$ are similar to those of Secs. \[sec\_d41\] and \[sec\_d4\]. There are, however, two main differences. For $D\ge 10$, the classical spiral ceases to exist [@sc1]. Thus, the caloric curve does not wind up as $\mu\rightarrow +\infty$ contrary to Fig. \[LHD4ZOOM\]. On the other hand, for $D\ge 2(1+\sqrt{2})$, it is not possible to construct self-confined white dwarf stars [@langevin]. This is just a mathematical curiosity since complete white dwarfs stars are unstable for $D>4$ anyway. This property changes the unstable branch of the $\Lambda-\eta_{P}$ diagram without consequence on the caloric curves. The $\Lambda-\eta_{P}$ diagram is represented Figs. \[LHpDn5\] and \[LHpD5.1\]. For $D>2(1+\sqrt{2})$, it displays an infinity of cusps towards the singular solution ($\Lambda_{s}$,$\eta_{P,s}$), see Fig. \[LHpD5.1\]. For $D=2(1+\sqrt{2})$, there is just one cusp (see Fig. \[LHpDn5\]) and the Lane-Emden equation (\[d7\]) can be solved analytically. This corresponds to the $D$-dimensional Schuster solution obtained for $n=n_{5}$ [@langevin]. In that case, we find explicitly $$\theta_{5}={1\over \bigl \lbrack 1+{\xi^{2}\over 4(2+\sqrt{2})}\bigr \rbrack^{\sqrt{2}}}. \label{fw1}$$ The normalized mass and the normalized energy can be expressed as $$\eta_{P}={\alpha^{2+\sqrt{2}}\over 2(1+\sqrt{2})\bigl \lbrack 1+{\alpha^{2}\over 4(2+\sqrt{2})}\bigr \rbrack^{1+\sqrt{2}}}, \label{fw2}$$ $$\Lambda_{5}=-2(1+\sqrt{2})\biggl\lbrack 1+{\alpha^{2}\over 4(2+\sqrt{2})}\biggr\rbrack^{2(1+\sqrt{2})}{1\over \alpha^{2(2+\sqrt{2})}}\int_{0}^{\alpha}{\xi^{1+2\sqrt{2}}\over \bigl\lbrack 1+{\xi^{2}\over 4(2+\sqrt{2})}\bigr\rbrack^{2(1+\sqrt{2})}}d\xi. \label{fw3}$$ The case $D=2$ {#sec_d2} -------------- Let us now consider smaller dimensions of space. The dimension $D=2$ is critical concerning gravitational collapse as discussed in [@sc1]. For $D=2$, the relevant Fermi integrals are $I_{0}$ and $I_{1}$. By definition, $$I_{0}(t)=\int_{0}^{+\infty}{dx\over 1+t e^{x}}. \label{t1}$$ Changing variables to $y=e^{x}$, we easily find that $$I_{0}(t)=\ln\biggl (1+{1\over t}\biggr ). \label{t2}$$ Therefore, the Fermi-Poisson equation (\[p4\]) becomes $${1\over\xi}{d\over d\xi}\biggl (\xi {d\psi\over d\xi}\biggr )=\ln\bigl (1+k^{-1}e^{-\psi}\bigr ). \label{t3}$$ $$\psi(0)=\psi'(0)=0. \label{t4}$$ On the other hand, using the identity (\[p3\]), giving $$I'_{1}(t)=-{1\over t}\ln\biggl (1+{1\over t}\biggr ), \label{t5}$$ one finds that $$I_{1}(t)=-\int_{-{1/t}}^{0}{\ln(1-x)\over x}dx=-{\rm Li}_{2}\bigl (-{1\over t}\bigr ), \label{t6}$$ where ${\rm Li}_{2}$ is the dilogarithm. Consider first the Fermi gas at $T=0$. In $D=2$, a white dwarf star is equivalent to a polytrope with index $n_{3/2}=1$. The Lane-Emden equation can then be solved analytically and we obtain $\theta=J_{0}(\xi)$, where $J_{0}$ is the Bessel function of zeroth order. The density drops to zero at $\xi_{1}=\alpha_{0,1}\simeq 2.40$, the first zero of $J_{0}$. Considering the mass-radius relation (\[d10\]) in $D=2$, we see that the radius is independant on mass. Therefore, complete white dwarf stars in two dimensions all have the same radius. It can be written in terms of fundamental constants as $$R_{*}={\xi_1\over 2\pi}\biggl ({h^{2}\over g m^{3}G}\biggr )^{1/2}=0.27\ {h\over m^{3/2}G^{1/2}}. \label{t7}$$ The relation between the mass and the central density of the white dwarf star is $$M={\rho_{0}\over 4\pi^{2}}{h^{2}\over g m^{3}G}\xi_{1}|\theta'_{1}|, \label{t8}$$ where $\theta'_{1}=J_{0}'(\alpha_{0,1})\simeq -0.52$. Thus, the density profile of a two-dimensional white dwarf star can be written $$\rho(r)=\rho_{0}J_{0}\bigl ({\xi_{1}r\over R_{*}}\bigr ). \label{t9}$$ This is similar to the vorticity profile of a minimum enstrophy vortex in 2D hydrodynamics [@jfm; @gt]. The energy of a complete polytrope of index $n$ in $D=2$ is $E=-(n-1)GM^{2}/8+(1/2)GM^{2}\ln (R_{*}/R)$ with the convention $\Phi(R)=0$ [@langevin]. Therefore the energy of a 2D white dwarf star is $$E={1\over 2}GM^{2}\ln \bigl ({R_{*}\over R}\bigr ). \label{t10}$$ Two-dimensional white dwarf stars exist for any mass $M$ and they are stable. Noting that $R_{*}/R=(\mu_{*}/\mu)^{1/2}=\xi_1/\sqrt{\mu}$ where $\mu=4\pi^{2}\eta_{0}GR^{2}$, we can write the normalized energy of the self-confined white dwarf star as $$\Lambda={1\over 2}\ln\biggl ({\sqrt{\mu}\over\xi_1}\biggr ). \label{t11}$$ Let us now consider the case of incomplete white dwarf stars that are confined by the box ($R_*>R$). This corresponds to $\mu<\xi_1^2$. Using Eq. (\[d6\]), we find that $\alpha=\sqrt{\mu}$. Then, using the results of [@langevin], we find that the normalized energy of a box-confined white dwarf star in two dimensions is $$\Lambda=-{1\over 2}{J_{0}(\sqrt{\mu})\over\sqrt{\mu}J_{1}(\sqrt{\mu})}. \label{t12}$$ We now consider the self-gravitating Fermi gas at finite temperature $T\neq 0$. According to Eq. (\[p6\]) we have $\alpha=\sqrt{\mu}$. Using Eq. (\[p8\]), we obtain $$\eta\equiv \beta GMm=\sqrt{\mu}\ \psi'(\sqrt{\mu}). \label{t12bis}$$ We need to calculate the energy specifically because the expression (\[p16\]) breaks down in $D=2$. The kinetic energy $K=\int p d^{2}{\bf r}$ can be written $${K\over GM^{2}}={1\over \eta^{2}}\int_{0}^{\sqrt{\mu}}I_{1}(k e^{\psi})\xi d\xi. \label{t13}$$ On the other hand, using an integration by parts, the potential energy is given by $$W=-{1\over 4\pi G}\int (\nabla\Phi)^{2}d^{2}{\bf r}, \label{t14}$$ where we have taken $\Phi(R)=0$. Introducing the dimensionless quantities defined in Sec. \[sec\_para\], we get $${W\over GM^{2}}=-{1\over 2\eta^{2}}\int_{0}^{\sqrt{\mu}}\psi'(\xi)^{2}\xi d\xi. \label{t15}$$ Summing Eqs. (\[t13\]) and (\[t15\]), the total normalized energy of the Fermi gas in two dimensions is $$\Lambda\equiv -{E\over GM^{2}}=-{1\over \eta^{2}}\int_{0}^{\sqrt{\mu}}I_{1}(k e^{\psi})\xi d\xi+{1\over 2\eta^{2}}\int_{0}^{\sqrt{\mu}}\psi'(\xi)^{2}\xi d\xi. \label{t16}$$ The corresponding caloric curve is plotted in Fig. \[LHD2\]. For $\mu\rightarrow +\infty$, we recover the classical caloric curve displaying a critical temperature $k_{B}T_{c}=GMm/4$ [@sc1]. Below $T_{c}$, a classical gas experiences a gravitational collapse and develops a Dirac peak [@sc1]. When quantum mechanics is taken into account, the collapse stops when the system becomes degenerate. The Dirac peak is replaced by a fermion ball surrounded by a dilute halo. At $T=0$, we have a pure Fermi condensate without halo. This is the ground state of the self-gravitating Fermi gas corresponding to the vertical asymptotes in Fig. \[LHD2\]. For $\mu<\xi_1^2$ (incomplete white dwarf stars), the minimum energy is given by Eq. (\[t12\]) and for $\mu<\xi_1^2$ (complete white dwarf stars) by Eq. (\[t11\]). This discussion concerning the difference between Dirac peaks (for classical particles) and fermion balls (for quantum particles) in the canonical ensemble remains valid for $2\le D<4$. Note also that there is no collapse (gravothermal catastrophe) in the microcanonical ensemble in $D=2$ [@klb; @sc1]. The case $D<2$ {#sec_d1} -------------- We finally conclude by the case $D<2$ (specifically $D=1$). First, we consider the Fermi gas at $T=0$. The $\Lambda-\eta_{P}$ curve which gives the energy of the star as a function of its mass is represented in Fig. \[LHpD1\]. Since the curve does not present turning points, all the white dwarf star configurations are stable. According to Eq. (\[d10\]), for $D<2$, the mass $M$ of a complete white dwarf star increases with its radius $R_*$. Therefore, for $M<M_*(R)$ the star is self-confined and for $M>M_*(R)$ it is restricted by the box. There exists a stable equilibrium state at $T=0$ for all mass. In terms of the dimensionless mass $\eta_P$, complete $n_{3/2}$ polytropes correspond to $\eta_P\le \omega_{D/2}$ and incomplete $n_{3/2}$ polytropes to $\eta_P\ge \omega_{D/2}$. This situation is reversed with respect to that of Fig. \[LHpD3\]. The caloric curve for arbitrary temperature and energy is represented in Fig. \[LHD1\]. For $\mu\rightarrow +\infty$ (i.e. $\hbar\rightarrow 0$), we recover the curve obtained in [@sc1] for classical isothermal systems. The caloric curve $\beta(E)$ is monotonic. Therefore, there is no phase transition for $D<2$. Thus, the change in the caloric curve due to quantum mechanics is not very important since an equilibrium state (global maximum of entropy or free energy) already exists for any accessible energy $E$ and any temperature $T$ in classical mechanics. Quantum mechanics, however, changes the ground state of the system. The ground state of the self-gravitating Fermi gas ($T=0$) corresponds to a white dwarf star configuration. Its structure (radius, energy) is determined by the intersection between the $\Lambda-\eta_P$ curve in Fig. \[LHpD1\] and the line defined by Eq. (\[d19\]). The “white dwarf” is complete ($R_{*}<R$) for $\mu>\mu_{*}(D)$ and incomplete ($R_{*}>R$) otherwise. For $\mu>\mu_{*}(D)$, the normalized energy of the white dwarf is given by Eq. (\[gtq4\]). This is the ground state of the self-gravitating Fermi gas corresponding to the asymptote in Fig. \[LHD1\]. In $D=1$, it is possible to obtain more explicit results. Using the results of [@langevin], for $n_{3/2}=1/2$ polytropes, we have $\xi_{1}=(3\pi/4)^{1/2}\Gamma(5/3)/\Gamma(7/6)\simeq 1.49$ and $|\theta'_{1}|=2/\sqrt{3}\simeq 1.15$. Therefore, $\omega_{1/2}=0.349$ and $\mu_{*}=0.846$. For $\mu>\mu_{*}=0.846$, the normalized energy of a complete white dwarf star (ground state) is $$\Lambda_{min}=-{3\over 7}\biggl ({\mu_{*}\over \mu}\biggr )^{2/3}. \label{bff}$$ Conclusion {#sec_conclusion} ========== In this paper, we have studied how the dimension of space affects the nature of phase transitions in the self-gravitating Fermi gas. Since this model has a fundamental interest in astrophysics [@chavcape] and statistical mechanics [@houches], it is important to explore its properties thoroughly even if we sacrifice for practical applications. It is well-known in statistical mechanics that the dimension of space plays a crucial role in the problem of phase transitions. For example, concerning the Ising model, the behaviour in $D=1$ and $D\ge 2$ is radically different [@huang]. We have reached a similar conclusion for the self-gravitating Fermi gas. The solution of the problem in $D<2$ does not yield any phase transition. In $D=2$, phase transitions appear in the canonical ensemble but not in the microcanonical ensemble. In $D>2$, phase transitions appear both in microcanonical and canonical ensembles in association with gravitational collapse. The beauty of self-gravitating systems, and other systems with long-range interactions, is their simplicity since the mean-field approximation is exact in any dimension. Therefore, the mean-field theory does [*not*]{} predict any phase transition for the self-gravitating Fermi gas in $D=1$, contrary to the Ising model. At a more philosophical level, several scientists have examined the role played by the dimension of space in determining the form of the laws of physics. This question goes back to Ptolemy who argues in his treatise [*On dimensionality*]{} that no more than three spatial dimensions are possible in Nature. In the $18^{\rm th}$ century, Kant realizes the deep connection between the inverse square law of gravitation and the existence of three spatial dimensions. In the twentieth century, Ehrenfest [@ehrenfest] argues that planetary orbits, atoms and molecules would be unstable in a space of dimension $D\ge 4$. Other investigations on dimensionality are reviewed in the paper of Barrow [@barrow]. Although we ignored this literature at the begining, our study clearly enters in this type of investigations. We have found that the self-gravitating Fermi gas possesses a rich structure and displays several characteristic dimensions $D=2$, $D=4$, $D=2(1+\sqrt{2})$ and $D=10$. Moreover, as already noted in [@langevin], the dimension $D=4$ is critical because at that dimension quantum mechanics cannot stabilize matter against gravitational collapse, contrary to the situation in $D=3$. Interestingly, this result is similar to that of Ehrenfest although it applies to white dwarf stars instead of atoms. The dimension $D=2$ is also critical as found in [@sc1] and in different domains of physics. Therefore, the dimension of our (macroscopic) universe $D=3$ plays a very special role regarding the laws of physics (this is illustrated in Fig. \[MRD\]). Following the far reaching intuition of Kant, we can wonder whether the three space dimensions are a consequence of Newton’s inverse square law, rather than the opposite. We note also that extra-dimensions can appear at the microscale, an idea originating from Kaluza-Klein theory. This idea took a renaissance in modern theories of grand unification. Our approach shows that already at a simple level, the coupling between Newton’s equations (gravitation) and Fermi-Dirac statistics (quantum mechanics) reveals a rich structure as a function of $D$. Relativistic effects will be considered in a forthcoming paper [@relativity]. Finally, our study can shed light on the mathematical properties of the Vlasov-Poisson system. Indeed, there is a close connexion between collisionless stellar systems and self-gravitating fermions [@lb; @csr; @cs; @dubrovnik]. For example, the fact that the Vlasov equation does not blow up (i.e., experiences gravitational collapse) in $D=3$ for non singular initial conditions can be related to a sort of exclusion principle, as in quantum mechanics. Due to the Liouville theorem in $\mu$-space, the distribution function must remain smaller that its maximum initial value $f\le \eta_{0}$ and this prevents complete collapse [@cs; @robert], unlike for collisional stellar systems [@lbw] described by the Landau-Poisson system. Since quantum mechanics cannot arrest gravitational collapse in $D\ge 4$ (for sufficiently low energies), this suggests that the Vlasov-Poisson system can probably blow up for $D\ge 4$. This remark could be of interest for mathematicians. [**Acknowledgements**]{} I acknowledge interesting discussions with P. Biler, T. Nadzieja and R. Stańczy. I also thank B. Douçot for encouragements. [10]{}

ArXiv

--- author: - 'L. Fuhrmann,[^1] V. Karamanavis,[^2] S. Komossa,[^3] E. Angelakis, T. P. Krichbaum, R. Schulz, A. Kreikenbohm, M. Kadler, I. Myserlis, E. Ros, I. Nestoras, J. A. Zensus' bibliography: - 'references.bib' title: 'Inner jet kinematics and the viewing angle towards the $\gamma$-ray narrow-line Seyfert1 galaxy 1H0323+342 ' --- Introduction ============ auto-capitalized Narrow-line Seyfert 1 (NLS1) galaxies are a subclass of active galactic nuclei (AGNs) with extreme multi-wavelength properties. They have been defined as AGN with small widths of their broad Balmer lines (Full width at half maximum, FWHM(H$\beta$) &lt; 2000km/s) and weak \[OIII\]5007/H$\beta$ emission [e.g., @1985ApJ..297...166; @1992ApJS...80..109B; @2004AJ....127.1799G; @2012AJ....143...83X see review by @2008RMxAC..32...86K]. As a class, they are characterized by super-strong FeII emission complexes , rapid X-ray variability and soft X-ray spectra , strong outflows , host galaxies with a preference for pseudo bulges [@2011MNRAS.417.2721O; @2012ApJ...754..146M], and a deficiency of [*very*]{} radio-loud systems, even though 7% of the sources of the whole population are radio-loud [@2006AJ....132..531K]. A few of the radio-loud systems have been detected in $\gamma$-rays by the Large Area Telescope (LAT) onboard the *Fermi* Gamma-ray Space Telescope for the first time [e.g. @2009ApJ...707L.142A; @2012MNRAS.426..317D; @2015MNRAS.454L..16Y]. The majority of these are characterized by one-sided radio jets, implying that we do not see these systems edge-on [e.g. @2009ApJ...707..727A; @2011ApJ...738..126D; @2012MNRAS.426..317D; @2013MNRAS.436..191D; @2014ApJ...781...75W; @2015MNRAS.453.4037O; @2015arXiv151102631S] . Several lines of evidence suggest that near-Eddington accretion rates onto low-mass black holes are a prime driver of the multi-waveband appearance of the NLS1 population, which, however, cannot account for all of the NLS1 properties [@2008RMxAC..32...86K]. Source orientation with respect to the observer has repeatedly been considered as another important factor. Arguments in favor of, or against, the importance of orientation effects in NLS1 galaxies have been given, based on emission-line widths, line profiles, polarimetry, radio properties and other considerations . Firm conclusions have been hampered so far by the lack of a method of actually [*measuring*]{} the viewing angle towards NLS1 galaxies. In particular, the knowledge of orientation is of great importance to measure the black hole massess of NLS1 galaxies (and other AGNs) accurately, when applying scaling relations [@2011arXiv1109.4181P] in case the broad-line region (BLR) is flattened [@2006MNRAS.369..182J; @2011arXiv1109.4181P]. The focus of this paper is a viewing-angle measurement of the radio-loud NLS1 galaxy 1H0323+342. This galaxy, at redshift $z$=0.0629, was recognized as a radio-loud NLS1 galaxy, sharing the properties of NLS1 galaxies and blazars by [@2007ApJ...658L..13Z]. Its host galaxy displays a peculiar morphology in form of a one-armed spiral [@2007ApJ...658L..13Z] or ring structure . Several independent estimates imply a black hole mass of $10^7$ solar masses and an accretion rate near the Eddington rate [e.g. @2007ApJ...658L..13Z; @2015AJ....150...23Y]. 1H0323+342 is highly variable at all wavebands including the radio , optical [@2014PASJ...66..108I], X-rays [@2009AdSpR..43..889F; @2014ApJ...789..143P; @2015AJ....150...23Y] and $\gamma$ rays [@2009ApJ...707L.142A; @2011MNRAS.413.2365C; @2013ATel.5344....1C; @2015AJ....149...41P]. It is one of only a few radio-loud NLS1 galaxies detected at GeV energies by the [*Fermi Gamma-ray Space Telescope*]{} [@2009ApJ...707L.142A]. Monitoring and spectroscopy with [*Swift*]{} and [*Suzaku*]{}, along with SED modelling, suggest that the UV and X-ray emission is dominated by the accretion disk and corona [@2015AJ....150...23Y see also [@2015ApJ...798...43S]]. Radio observations of 1H0323+342 revealed a flat radio spectrum , high apparent brightness temperatures , and a core with extended structure of 15kpc observed with the VLA . At 8GHz, [@2014ApJ...781...75W] reported the presence of a compact core-jet structure at parsec (pc) scales and evidence for two slowly moving jet components ($\beta_{\rm app} \ll$ c). Aiming to constrain the viewing angle towards the source, we present multi-epoch, high-resolution VLBI images of 1H0323+342 at 15GHz [see @2015PhDT.......232K for first results] as well as single-dish radio monitoring observations obtained with the Effelsberg 100-m and IRAM 30-m telescopes during the years 2010–2014. The former are used to measure the parsec-scale jet structure and its kinematics over a time span of 3 years, and the latter to infer the brightness temperature and Doppler factor for the source. Their combination provides estimates of the viewing angle towards this NLS1 galaxy. The paper is structured as follows. Section \[sect:2\] describes the observations and data analysis procedures. In Sections \[D\_var\] and \[kinematics\], we present our results from the total flux density and VLBI monitoring, respectively. Section \[counter\_jet\_ratio\] concerns the estimation of the jet-to-counter-jet ratio for 1H0323+342, and Sect. \[sect:Combi\] presents the calculation of its viewing angle combining our single-dish and VLBI findings, along with a similar estimate for SBS0846+513. We provide a discussion of our results in Sect. \[sect:Disc\], and a short summary with conclusions in Sect. \[sect:summary\]. Throughout this paper we adopt the following cosmological parameters: ${\rm H}_{0} = 71\,{\rm km\,s}^{-1}\,{\rm Mpc}^{-1}$, $\Omega_{\rm m} = 0.27$ and $\Omega_{\Lambda} = 0.73$ [@2007ApJS..170..377S]. At the redshift of 1H0323+342 an angular separation of 1milliarcsecond (mas) corresponds to 1.196pc and a proper motion of 1masyr$^{-1}$ translates to an apparent superluminal speed of 3.89c. We also follow the convention $S \propto \nu^{+\alpha}$, where $S$ is the radio flux density, $\nu$ the observing frequency and $\alpha$ the optically thin spectral index. Observations and data analysis {#sect:2} ============================== Single-dish, total flux-density monitoring ------------------------------------------ Since July 2010, the NLS1 galaxy 1H0323+342 has been regularly observed at cm to short-mm bands in the framework of a dedicated monitoring program of NLS1 galaxies at the Effelsberg (EB) 100-m telescope and in the course of the [*Fermi*]{}-related F-GAMMA[^4] monitoring program [@2007AIPC..921..249F; @2010arXiv1006.5610A; @2014MNRAS.441.1899F] including quasi-simultaneous observations with the IRAM 30-m telescope (at Pico Veleta, PV). These observations are closely coordinated with the more general AGN flux density monitoring conducted at the IRAM 30-m telescope [e.g. @1998ASPC..144..149U]. The overall frequency range of the 1H0323+342 monitoring data spans from 2.64 to 142GHz. The EB measurements were conducted with cross-scans in the frequency range from 2.64 to 43.05GHz. The PV observations were carried out with calibrated cross-scans using the Eight MIxer Receiver (EMIR) at 86.2 and 142.3 GHz. In the data reduction process for each station, pointing offset, elevation-dependent gain, atmospheric opacity and sensitivity corrections have been applied to the data . Example light curves of 1H0323+342 at the selected frequencies of 142.3, 86.2, 32.0, 14.6 and 8.35GHz, covering the period 2010 July (2010.6; MJD55408) to 2015 January (2015.1; MJD57053), are shown in Fig. \[fgamma\_lcs\] (top panel). Earlier results of this NLS1 monitoring campaign have been presented in e.g. [@2011nlsg.confE..26F] and . The detailed multi-frequency analysis and results of the full NLS1 radio data sets (five sources) obtained between 2010 and 2014 are presented and discussed in . \ ![Single dish and VLBI radio light curves of 1H0323+342. The top panel features the F-GAMMA total flux density light curves at 142.3, 86.2, 32.0, 14.6 and 8.35GHz. In the middle panel, those at 14.6GHz from EB (purple squares) along with that of the 15GHz VLBI core (black) and the stationary component S (light blue) are shown. The bottom panel features the light curves of all individual VLBI components. Note the different scales and the presence of component S in both panels for visual comparison.[]{data-label="fgamma_lcs"}](./Plots/0323_fgamma_vlbi_LCs_2.pdf "fig:"){width="0.6\linewidth"} VLBI monitoring {#Sec:Data:Mojave} --------------- The MOJAVE program [@2009AJ....137.3718L] has been monitoring 1H0323+342 with the Very Long baseline Array (VLBA) at 15GHz between 2010 October and 2013 July, and a total of eight observing epochs of fully self-calibrated data sets are available[^5]. The brightness distribution of the source at each epoch has been parameterised by sets of two-dimensional circular Gaussian functions with the <span style="font-variant:small-caps;">DIFMAP</span> package [@1994BAAS...26..987S]. This uses a $\chi^{2}$ minimisation algorithm (<span style="font-variant:small-caps;">MODELFIT</span>) to fit the parameters of the Gaussian components – i.e. flux density, radial separation from the phase center, position angle (PA) and size – to the visibility function. First, the image of each epoch was analyzed independently by building the model iteratively, i.e. adding new components as long as the residual image shows significant flux density and the new component improved the overall fit in terms of the $\chi^{2}$ value. The addition of a component was followed by iterations of <span style="font-variant:small-caps;">MODELFIT</span> (until the fit has converged) and self-calibration of the visibility phases. In addition, a cross-check of our model was performed by comparing the position of the components with the brightness distribution in the <span style="font-variant:small-caps;">CLEAN</span> image. We adopt an uncertainty of 10% for the flux density of components [e.g. @2009AJ....137.3718L]. The positional uncertainty is set to 1/5 of the beam, if they are unresolved, or 1/5 of their FWHM for those resolved. A few data points deviate from the fitted lines, as common of such type of studies, attributable to possible epoch-to-epoch illumination of different emission regions within the same knot, and small differences in the uv-coverage. The uncertainty in PA is given by $\Delta {\rm PA} = \arctan(\Delta {\rm X}/r)$, with $\Delta {\rm X}$ the positional uncertainty and $r$ their separation from the core . We note that neither during the cleaning nor the model-fitting process, significant flux density could be associated with a putative counter-jet above the noise-level of a given epoch (see Sect. \[counter\_jet\_ratio\]). Details of the resulting multi-epoch model parameters are given in the Appendix (Table \[tab:modelfit\]). All individual <span style="font-variant:small-caps;">MODELFIT</span> maps are shown in Fig. \[vlbi\_images\]. Results from total flux density monitoring {#D_var} ========================================== The single-dish light curves[^6] displayed in Fig. \[fgamma\_lcs\] (top) show 1H0323+342 in a phase of mild, low-amplitude but rather fast variability during the period $\sim$2010 to 2012. Here, the 14.6GHz light curve, for instance, exhibits repeated low-amplitude flares with amplitudes ($\Delta S = S_{\rm{max}}-S_{\rm{min}}$) of $\sim$120–170mJy and time scales ($\Delta t=t_{\rm{max}}-t_{\rm{min}}$) of $\lesssim$40–100days. Using these values, we calculate lower limits for the apparent brightness temperature $T_{\rm{B,var}}$ associated with these events ranging between $5\times10^{10}$ and $4\times10^{11}$K. Starting in 2012, the source showed more prominent activity with a first larger mm-band flare peaking around 2012.2 and subsequently, a large outburst occurring between $\sim$2012.6 and 2013.5. The latter appears to exhibit 2–3 sub-flares with the most prominent event showing a variability amplitude of $\Delta S\sim\,570$mJy during a time period of about 50days at 14.6GHz. This yields a lower limit for $T_{\rm{B,var}}$ of about $9\times10^{11}$K. The observed brightness temperature can be further constrained using the highest significant amplitude variation occuring at adjacent times, between MJD 56296.8 and 56312.7 in the 14.6GHz light curve (Fig. \[fgamma\_lcs\]). This corresponds to a variability amplitude of 384mJy and a time scale of 15.9 days. These figures yield an apparent $T_{\rm{B,var}}$ of $5.7\times10^{12}$K. Assuming an intrinsic equipartition brightness temperature limit of $5\times10^{10}$K [@Readhead1994ApJ], we can estimate lower limits for the variability Doppler factor $D_{\rm{var}}$ required to explain the excess of $T_{\rm{B,var}}$ over the theoretical limit . The maximum observed value of $5.7\times10^{12}$K then yields $D_{\rm{var}}$$\gtrsim$5.2 for the 14.6GHz variability seen in Fig. \[fgamma\_lcs\]. Based on *a different method*, their flare decomposition method, deduce a $D_{\rm{var}}$$\gtrsim$3.6 for the flare seen at 14.6GHz, peaking at MJD $\sim$56300. For a full event-by-event treatment of the data sets until Spring 2014 see the aforementioned publication. Results from VLBI monitoring {#kinematics} ============================ Figure \[vlbi\_images\] shows the final <span style="font-variant:small-caps;">MODELFIT</span> maps of 1H0323+342 at $15$GHz. On parsec scales, the source is characterized by a one-sided jet that appears remarkably straight, oriented at a mean position angle of ${\sim} 124^{\circ}$, and extending out to almost $10$mas from its bright, unresolved core. The latter contains more than 70% of the jet’s total flux density (see Fig. \[fgamma\_lcs\]). As also seen in the VLBI maps of Fig. \[vlbi\_images\], the jet of 1H0323+342 is not undisrupted all the way until its maximum angular extent. Further downstream from the core the jet exhibits an area of reduced surface brightness (a gap of emission). On average this is visible between $3$ to $6$mas away from the core at the PA of the jet, whereafter the flow becomes visible again. This effect could be attributed to the weighting of the VLBI data, selected such that resolution is maximized at the expense of sensitivity to putative extended structure[^7]. At our observing frequency the jet can be decomposed into six to nine <span style="font-variant:small-caps;">MODELFIT</span> components. Based on the parameters of the model and their temporal evolution we are able to positively cross-identify seven components between all eight observing epochs. Starting from the outermost component – in terms of radial separation from the core – we use the letter C followed by a number between $1$ and $6$ to designate and hereafter refer to them. Apart from travelling components, our findings suggest the presence of a quasi-stationary component, referred to with the letter S, positioned very close to the core at a mean distance of ${\sim} 0.3$mas. Furthermore, at epoch $2013.52$ a new component appears to emerge, most probably ejected at some time between the latest two observing epochs; i.e. between $2012.98$ and $2013.52$ (see Sect. \[VLBI\_discussion\]). We refer to this knot as NC which stands for ‘new component’ (cf. VLBI maps in Fig. \[vlbi\_images\]). The kinematical properties of the jet were derived by identifying model components throughout our multi-epoch observations, which describe the same jet features. The kinematical parameters for each knot are obtained through a weighted linear regression fit of their positions relative to the assumed stationary core over time. In Fig. \[fig:CoreSep\] each component’s radial separation from the core at each observing epoch is shown, along with the linear fits. From the slope of each fit we deduce the component’s proper motion $\mu$ relative to the core and its apparent velocity $\beta_{\rm app}$ in units of the speed of light. The results are summarized in Table \[tab:kinematics\]. ![Eight uniformly-weighted <span style="font-variant:small-caps;">MODELFIT</span> maps of 1H0323+342 from the MOJAVE observations at 15GHz. Contour levels correspond to $-0.3$%, 0.3%, 0.6%, 1.2%, 2.4%, 4.8%, 9.6%, 19.2% and 38.4% of the highest peak flux density of 0.344 Jybeam$^{-1}$ (epoch 2011.17) for common reference. All maps are convolved with an average beam with major and minor axes of 0.77 and 0.48mas, respectively, with the major axis at PA $= -5^{\circ}$.[]{data-label="vlbi_images"}](./Plots/vlbi_maps_all.pdf){width="0.55\linewidth"} ![Temporal evolution of fitted component separation from the core. Line segments represent the best-fit linear model describing each component’s kinematical behavior. Apart from robustly-identified knots (C1–C6), open squares denote non-robust features. The new component (NC) visible in the latest epoch only, is shown as a red circle immediately after the quasi-stationary feature S.[]{data-label="fig:CoreSep"}](./Plots/0323_coreSeparation.pdf){width="0.6\linewidth"} Beyond the first mas and before the area of reduced radio emission, $\beta_{\rm app}$ shows an increasing trend. There, superluminal knots C5 and C4 travel with apparent velocities of ${\sim} 4$c and ${\sim} 7$c, respectively. Superluminal knot C4 is the fastest moving component in the relativistic flow of 1H0323+342. After the gap of emission, when the jet becomes visible again, the flow is characterized by slower speeds. Components C3, C2 and C1 are used to describe the elongated area of emission extending at a radial distance between 6 and 10mas away from the core. Their speeds are superluminal, only slightly slower than in the jet area before the gap, with $\beta_{\rm app,\,C3} \simeq 1.6$, $\beta_{\rm app,\,C2} \simeq 3$, and $\beta_{\rm app,\,C1} \simeq 4$. Ejection times of knots are derived by back-extrapolating their motion to the time of zero separation from the core and are given in the last column of Table \[tab:kinematics\]. The source appears very active in this respect, showing an ejection of a new VLBI component every ${\sim}3.1$ years on average with the earliest ejection in $\sim$1994.6 and the most recent component ejections in 2009/2010. ---- ------------------------ ------------------- ------------------- -- -- -- -- -- -- -- -- ID $\mu$ $\beta_{\rm app}$ $t_{\rm ej}$ $({\rm mas\,yr}^{-1})$ $({\rm c})$ $({\rm yr})$ C1 $1.07 \pm 0.09$ $4.38 \pm 0.35$ $2003.9 \pm 0.7$ C2 $0.74 \pm 0.15$ $3.03 \pm 0.63$ $2002.0 \pm 2.1$ C3 $0.38 \pm 0.26$ $1.58 \pm 1.06$ $1994.6 \pm 12.1$ C4 $1.69 \pm 0.07$ $6.92 \pm 0.29$ $2009.9 \pm 0.1$ C5 $1.00 \pm 0.20$ $4.12 \pm 0.82$ $2010.1 \pm 0.3$ C6 $0.23 \pm 0.06$ $0.93 \pm 0.24$ $2008.3 \pm 0.9$ S $0.02 \pm 0.01$ $0.10 \pm 0.05$ … ---- ------------------------ ------------------- ------------------- -- -- -- -- -- -- -- -- : Kinematical parameters for all components of 1H0323+342. Columns from left to right are: (1) component designation, (2) proper motion in mas/yr, (3) apparent velocity in units of c, (4) time of knot ejection.[]{data-label="tab:kinematics"} The main variable component at $15$GHz, according to our VLBI findings, appears to be the core whereas all other VLBI jet components of 1H0323+342 share a lower flux density level. Comparison between the F-GAMMA filled-aperture light curve at $14.6$GHz and the VLBI total flux density of all components (see Fig. \[fgamma\_lcs\]) indicates that VLBI flux densities account for most of the single-dish flux density, a hint that extended structure does not contribute a substancial flux density level. It in fact suggests that the total intensity flares originate in the unresolved $15$GHz VLBI core region of the source, since the core flux density is variable and accounts for more than 70% of the observed total VLBI flux density. Jet-to-counter-jet ratio {#counter_jet_ratio} ======================== Assuming intrinsic symmetry between jet and counter-jet, the jet-to-counter-jet ratio $R$ represents the ratio of flux densities of the approaching and receding jets and is given by $$R = \dfrac{S_{\rm Jet}}{S_{\rm cJet}} = \left( \dfrac{1 + \beta \cos \theta}{1 - \beta \cos \theta} \right)^{2 - \alpha} = \left(\beta_{\rm app}^{2} + D^{2} \right)^{2-\alpha}, \label{eq_Ratio}$$ where $\beta$ and $\theta$ are the intrinsic speed and jet angle to the observer’s line of sight. The index ($2-\alpha$) refers to a continuous jet flow and changes to ($3-\alpha$) for a ‘blobby’ jet [@1979Natur.277..182S]. For the NLS1 galaxy 1H0323+342 the total flux density of all components in the VLBA images is associated with the jet, while putative counter-jet emission from the source is not visible at the opposite direction of the core and to the North (see Fig. \[vlbi\_images\]; this statement assumes, that there is no contribution from counter-jet emission to the brightest radio component). Despite the fact that there is no sign of emission from the counter-jet side, we have nonetheless tried to constrain its non-observability. First, the VLBI map with the best signal-to-noise ratio was selected, i.e. epoch 2011 March 5. Using the <span style="font-variant:small-caps;">CLEAN</span> algorithm and by utilizing <span style="font-variant:small-caps;">CLEAN</span> windows to constrain the area where the algorithm searches for peaks, we have tried to place delta components at the opposite side of the jet in an attempts to ‘artificially create’ a fake counter-jet. Mostly positive and negative components with no significant flux (below the noise level of the map) were fitted in this area of the image – if at all – as expected. Another test was performed using the <span style="font-variant:small-caps;">MODELFIT</span> algorithm, this time by attempting to fit the self-calibrated (only for the phase term) visibilities with circular Gaussian components. The results were negative in this case too, with no significant components on the opposite side of the jet. Under the assumption that Doppler de-boosting is responsible for the suppression of the counter-jet one can obtain a lower limit for the jet-to-counter-jet ratio $R$. We calculate R using the <span style="font-variant:small-caps;">MODELFIT</span> map at epoch 2011 March 5. As an upper limit for the flux density of the counter-jet, $S_{\rm cJet}$, we take three times the rms integrated flux density, over a region with comparable size to and in the opposite side of the jet (rms $= 0.38$mJy). Being undetected, any counter-jet emission ought to be below this level. Consequently, through $R = S_{\rm Jet}/S_{\rm cJet} = S_{\rm Jet}/S_{3 \sigma}$ , we estimate that $R \ge 363.5$, where $S_{\rm Jet}$ is the total integrated flux density of the fitted core and jet components. A first estimate for the viewing angle towards the source can then be obtained through $$\theta \le \arccos \left( \frac{R^{\frac{1}{2-\alpha}} - 1}{R^{\frac{1}{2-\alpha}} + 1} \right) \hspace{0.2cm} \beta^{-1}.$$ Under the assumption that the intrinsic speed of the components of the pc-scale jet is close to the speed of light, i.e. letting $\beta \rightarrow 1$, and with a mean $\alpha = -0.04$ , we arrive at an upper limit for the viewing angle of $\theta \le 27^{\circ}$, for $R = 363.5$. If instead, we exclude the core flux density for this calculation, an even weaker constraint is obtained. Note that solving Eq. \[eq\_Ratio\] for $D$, this value of $R$, combined with the highest measured apparent velocity in the jet, implies a non-physical Doppler factor ($D \in \mathbb{I}$; imaginary). To remedy this and get a real-valued Doppler factor, a much higher jet-to-counter-jet ratio is needed (${\sim} 2700$), which in turn yields a viewing angle of $\theta \le 16^{\circ}$. Another way of constraining the viewing angle towards the source is the following. Having a measured value for $\beta_{\rm app}$ we calculate the minimum Lorentz factor in order to observe this apparent speed as $$\gamma_{\rm min} = \sqrt{\beta_{\rm app}^{2} + 1}. \label{gamma_min}$$ Then using Eq. \[gamma\_min\] under the assumption that this is the Lorentz factor characterizing the flow, the critical viewing angle is estimated through [@1966Natur.211..468R] $$\theta_{\rm crit} = \arcsin \left( \dfrac{1}{\gamma_{\rm min}} \right).$$ In this way and for $\beta_{\rm app} = 6.92$, the minimum Lorentz factor is $\gamma_{\rm min} = 6.99$ and correspondingly we obtain $\theta_{\rm crit} = 8.2^{\circ}$. For a discussion on the use of the critical angle as a proxy for the viewing angle see . The viewing angle towards 1H0323+342 {#sect:Combi} ==================================== A viewing angle estimate towards 1H0323+342 can be obtained based on our radio variability and imaging results. With regard to the obtained Doppler factor, we would first like to note, that there is, in principle, several methods for estimating Doppler factors in blazars, and NLS1 galaxies in particular. One approach is based on SED modelling, which has also been carried out for 1H0323+342 (e.g., [@2009ApJ...707L.142A; @2013ApJ...774L...5Z; @2014ApJ...789..143P; @2015AJ....150...23Y; @2015ApJ...798...43S] While this is a very important approach, the estimated parameters still come with a number of uncertainties, based on, for instance, non-simultaneity of some of the data points, assumptions about black holes mass and correspondingly the accretion disk contribution to the SED, and based on different assumptions of different authors on the dominant contribution to EC processes (like the BLR or the molecular torus, or both). Alternatively, the variability Doppler factor in the radio regime can be directly estimated from radio flux density monitoring data. We have carried out such a program over several years, and it is these data sets we have used here for directly estimating the variability Doppler factor (the value we find (Sect. 3) is consistent within the errors with estimates based on SED fitting [e.g., @2015AJ....150...23Y their Tab. 6]. We have performed the estimate at 15 GHz, because this is the frequency at which the VLBI data were taken. Combining the variability Doppler factor $D_{\rm var}$ estimated in Sect. \[D\_var\], based on our single-dish monitoring with the apparent jet speed $\beta_{\rm app}$ estimated in Sect. \[kinematics\], we are able to estimate the variability Lorentz factor $\gamma_{\rm var}$ and the viewing angle $\theta_{\rm var}$ towards 1H0323+342 according to [@2005AJ....130.1418J for the derivation see ] $$\gamma_{\rm var} = \dfrac{ \beta_{\rm app}^{2} + D_{\rm var}^{2} + 1 }{ 2 D_{\rm var} } ~~~ \textrm{and} \label{eq:minLorfctr}$$ $$\theta_{\rm var} = \arctan \left( \dfrac{ 2 \beta_{\rm app} }{ \beta_{\rm app}^{2} + D_{\rm var}^{2} - 1 } \right). \label{eq:thetaVar}$$ Ideally, one aims at using the apparent jet speed of the travelling VLBI knot causing the outburst for which $D_{\rm var}$ has been obtained. However, this is not possible in the case of 1H0323+342, given the available data sets (see also Sect. \[VLBI\_discussion\]). Nevertheless, we can assume the range of obtained $D_{\rm var}$ and jet speeds in Sect. \[D\_var\] and \[kinematics\] as representative, average jet characteristics to obtain a range of plausible values for the Lorentz factor as well as viewing angle of the source. Assuming a constant Doppler factor, we use Eqs. \[eq:minLorfctr\] and \[eq:thetaVar\] and combine the observed apparent velocities of the flow with the variability Doppler factors discussed in Sect. \[D\_var\]. While the most stringent constraint is obtained from the combination of the lowest observed $\beta_{\rm app} = 0.93$ with the highest Doppler factor, given the large scatter of observed speeds, we perform the same calculation using the lowest $D_{\rm var} \ge 3.6$ and the highest $\beta_{\rm app} = 6.92$. In the former case we obtain a variability Lorentz factor $\gamma_{\rm var} \ge 2.8$ and the viewing angle $\theta_{\rm var} \le 4.0^{\circ}$. In the latter, we have $\gamma_{\rm var} \ge 8.6$ and $\theta_{\rm var} \le 13.0^{\circ}$. The final range of values for the viewing angle is $\theta_{\rm var} \le 4^{\circ}$–$13^{\circ}$. Results of all calculations are summarized in Table \[tab:resultsAll\]. [l r]{} Measured quantities & Inferred parameters\ \ $R = 363.5$ & $D \in \mathbb{I}$\ $\beta \rightarrow 1$ & $\theta \leq 27^{\circ}$\ $R_{\rm req}$ = $2700$ & $D \in \mathbb{R}$\ $\beta \rightarrow 1$ & $\theta \leq 16^{\circ}$\ \ $\beta_{\rm app} = 6.92$ & $\gamma_{\rm min} = 6.99$\ & $\theta_{\rm crit} = 8.2^{\circ}$\ \ $\beta_{\rm app, min} = 0.93$ & $\gamma_{\rm var} \ge 2.8$\ $D_{\rm var} \ge 5.2$ & $\theta_{\rm var} \le 4^{\circ}$\ $\beta_{\rm app, max} = 6.92$ & $\gamma_{\rm var} \ge 8.6 $\ $D_{\rm var} \ge 3.6$ & $\theta_{\rm var} \le 13^{\circ}$\ \[tab:resultsAll\] Discussion {#sect:Disc} ========== On the variability of 1H0323+342 {#VLBI_discussion} -------------------------------- The unresolved 15GHz VLBI core flux density is variable and accounts for more than 70% of the total VLBI flux density. Consequently, it is likely the region responsible for the variability and flares seen in the single-dish total flux density light curves, however, a direct connection between total flux density outbursts and pc-scale VLBI jet features cannot be established given the poor VLBI data sampling. This is particularly true for the most dramatic outburst observed around 2013.1. In addition, given the limited time baseline of the single-dish observations, we are unable to establish a one-to-one correspondence between component ejection times and total intensity flares. This becomes clear from the estimated ejection times of components reported in Table \[tab:kinematics\], none of which appears to have separated from the core within the time baseline of our single-dish radio monitoring. The only exception appears to be a new jet component (refereed to as NC) only seen to be present in the data of the last epoch ($2013.52$, see Fig. \[fig:CoreSep\]) with an integrated flux density of ${\sim} 19$mJy. Given the flaring activity preceding its separation from the core, this knot could be tentatively associated with the above mentioned, most dramatic outburst seen in our single-dish data around epoch ${\sim} 2013.1$ (see Fig. \[fgamma\_lcs\]). On the viewing angle towards 1H0323+342 --------------------------------------- A first indication that the viewing angle towards 1H0323+342 is not exceptionally small is the remarkably straight jet. All jet components lie at a mean position angle of about 124 degrees, with minor deviations. Since the jet appears so aligned, this likely implies that the viewing angle is not too small, otherwise PA variations would be greatly enhanced and more pronounced due to relativistic aberration effects, as in the case of objects seen at very small viewing angles (${\sim} 1^{\circ}$–$2^{\circ}$). As shown in Sect. \[counter\_jet\_ratio\], the viewing angle $\theta$ towards 1H0323+342 can be loosely constrained by using the interferometric data alone. Here, the minimum Lorentz factor estimate yields a value of $8.2^{\circ}$ from $\theta_{\rm crit}$, while the jet-to-counter-jet ratio estimate provides loose upper limits of $\theta\,\le\,27^{\circ}$ and ultimately $\theta\,\le\,16^{\circ}$. The combination of VLBI kinematics and single-dish variability allows us to further constrain $\theta$ (Sect. \[sect:Combi\]). We note that the variability Doppler factors constitute only lower limits, given the fact that the emission region can be smaller than the size implied by the variability time scales obtained. presented a detailed analysis of the single-dish data sets of 1H0323+342 between 2010.6 and 2014.2 based on a multi-frequency flare decomposition method. At 14.6GHz, their study yielded $D_{{\rm var}} \ge 3.6$. Consequently, using $D_{{\rm var}}>3.6$ and $\beta_{\rm app} = 6.92$c we obtain a viewing angle of ${\le} 13^{\circ}$. Additionally, in Sect. \[D\_var\] we deduced a variability Doppler factor $D_{\rm var} \ge 5.2$. This, combined with the slowest apparent speed of the moving knot C6, yields another, more stringent upper limit for $\theta$ of ${\le}4^{\circ}$. The very small apparent speed of C6 can be reconciled with such a high Doppler factor, if projection effects are at play and the jet is pointed close to the observer’s line of sight, as seen from the inferred aspect angle. The viewing angle towards the NLS1 galaxy SBS0846+513 {#SBS} ----------------------------------------------------- Using the results presented by [@2013MNRAS.436..191D] [see also @2012MNRAS.426..317D] we can also estimate the viewing angle towards the radio and $\gamma$-ray-loud NLS1 SBS0846+513, using the same method presented in Sect. \[sect:Combi\]. The authors report an apparent jet speed for this source of 9.3c. Furthermore, we estimate $D_{\rm var}$ as in Sect. \[D\_var\] from the 15GHz variability presented in Figs. 3 and 4 of [@2012MNRAS.426..317D] to be of the order ${\ge} 6$–$7$, in agreement with a value of ${\ge} 7.1$ which we obtain from the variability characteristics given in [@2013MNRAS.436..191D] and using an intrinsic equipartition brightness temperature limit of $5\times10^{10}$K. These values for $\beta_{{\rm app}}$ and $D_{\rm var}$ then provide a viewing angle estimate of ${\le}8^{\circ}$–$9^{\circ}$ using Eq. \[eq:thetaVar\]. Comparison with classical blazars --------------------------------- Out of the six moving features populating the jet of 1H0323+342, five are superluminal and only one moves with a marginally subluminal velocity. The apparent velocities span a wide range between 0.93c up to 6.92c. Such a range of apparent speeds is often observed for the class of blazars [e.g. @2013AJ....146..120L]. The Lorentz factors obtained for 1H0323+342 and SBS0846+513 are in the observed range of blazars but at the lower end of the blazar Lorentz factor distribution and more typical for BL Lac objects . The different methods we employ constrain the viewing angle towards 1H0323+342 to $\theta \leq 4^{\circ}$–$13^{\circ}$, while $\theta \leq 8^{\circ}$–$9^{\circ}$ for SBS0846+513. These numbers are larger than the average viewing angles towards blazars, of ${\sim} 1^{\circ}$–$2^{\circ}$ , but given that out measurements are upper limits, the angles could still be consistent with smaller values. Whether this is a general characteristic of radio-loud, $\gamma$-ray-loud NLS1 galaxies needs to be answered by detailed studies of a larger sample of this source type. Based on VLA observations of the kpc-scale jets of three radio-loud but non-$\gamma$-ray detected NLS1 galaxies, the finding of [@2015ApJ...800L...8R] suggest that these sources are seen at a moderate angle to the line of sight (10$^{\circ}$–15$^{\circ}$). Our results confirm previous findings that radio-loud, $\gamma$-ray-loud NLS1 galaxies, as a class, share blazar-like properties[^8]. However, compared to typical GeV-blazars [e.g., @2012ApJ...752..157Z; @2014ApJ...788..104Z], and references therein), most radio-loud NLS1 galaxies appear as lower brightness temperature sources[^9], possibly seen at slightly larger viewing angles than classical blazars. These properties can then explain the typically small Doppler boosting factors and low flux densities of these sources . The application of the method presented here to a larger sample of radio-loud NLS1 galaxies will provide us with first robust measurements of the viewing angle towards this population of AGN. Summary and conclusions {#sect:summary} ======================= We have presented a VLBI study of the $\gamma$-ray- and radio-loud NLS1 galaxy 1H0323+342, including a viewing angle measurement towards the inner jet of this source, by combining single-dish and VLBI monitoring. The salient results of our study can be summarized as follows: 1. 1H0323+342 displays a pc-scale morphology of the core-jet type. The emission is dominated by the unresolved core which contains more than 70% of the jet’s total flux density. The jet extends straight up to about 10mas from the bright nucleus, confirming previous results at a lower frequency. The VLBI structure at 15GHz can be represented using six moving components and one quasi-stationary feature. 2. Five out of the six moving knots exhibit superluminal motion. The range of apparent velocities is 0.93c up to 6.92c. Closer to the core (${\le} 1$mas) speeds are subluminal. Between 1mas and 5mas, the highest speeds of the flow are observed with components moving at about 4c and 7c. 3. 1H0323+342 is active, showing flux density flares, of low to mild amplitude, but very rapid compared to the typical long-term variability of blazars. The behaviour is seen both in single-dish total flux density and on VLBI scales. Previous findings infer high brightness temperatures (above the equipartition limit) and a variability Doppler factor of 3.6, taking into account the parameters of full flaring episodes. Here we also deduce an extreme $T_{\rm B,\,var} = 5.7\times10^{12}$K and $D_{\rm var} = 5.2$, based on the most rapid variability of the highest amplitude. 4. Using different methods we ultimately constrain the viewing angle towards 1H0323+342 to be in the range $\theta_{\rm var} \le 4^{\circ}$–$13^{\circ}$. 5. Applying similar estimates to published results of another radio-loud NLS1 galaxy, SBS0846+513, we estimate a viewing angle of ${\le} 8^{\circ}$–$9^{\circ}$. Both sources are therefore seen at a small viewing angle. Taken at face value, inferred numbers are larger than the bulk of the blazar population, but still consistent with each other, given that our estimates are upper limits. Aspect angle measurements of a larger sample of radio-loud NLS1 galaxies will provide us with important new insight into orientation scenarios for NLS1 galaxies. We would like to thank our referee for very useful comments and suggestions. The authors thank H. Ungerechts and A. Sievers for their support in the course of the IRAM 30-m observations, and B. Boccardi for carefully reading the manuscript. V.K., I.N. and I.M. were supported for this research through a stipend from the International Max Planck Research School (IMPRS) for Astronomy and Astrophysics at the Universities of Bonn and Cologne. R.S. was supported by the Deutsche Forschungsgemeinschaft grant WI 1860/10-1. E.R. acknowledges partial support by the the Spanish MINECO project AYA2012-38491-C02-01and by the Generalitat Valenciana project PROMETEOII/2014/057. This research made use of observations with the 100-m telescope of the MPIfR (Max-Planck-Institut für Radioastronomie) at Effelsberg, and observations with the IRAM 30-m telescope. IRAM is supported by INSU/CNRS (France), MPG (Germany) and IGN (Spain). This research has made use of data from the MOJAVE database that is maintained by the MOJAVE team (Lister et al., 2009, AJ, 137, 3718) and of NASA’s Astrophysics Data System. \[appendix\] MODELFIT Results ================ [c c c c c c]{} Epoch & $S$ & $r$ & PA & FWHM & ID\ & (Jy) & (mas) & ($^{\circ}$) & (mas) &\ 2010.79 & $0.193 \pm 0.019$ & …& …& $ 0.08 \pm 0.01$ & C\ 2010.79 & $0.064 \pm 0.006$ & $ 0.29 \pm 0.03 $ & $ 131.6 \pm 6.2 $ & $ 0.16 \pm 0.02$ & S\ 2010.79 & $0.019 \pm 0.002$ & $ 0.96 \pm 0.08 $ & $ 127.0 \pm 4.8 $ & $ 0.41 \pm 0.04$ & C5\ 2010.79 & $0.011 \pm 0.001$ & $ 1.69 \pm 0.11 $ & $ 123.9 \pm 3.6 $ & $ 0.53 \pm 0.05$ & C4\ 2010.79 & $0.002 \pm 0.001$ & $ 3.11 \pm 0.19 $ & $ 123.4 \pm 3.5 $ & $ 0.94 \pm 0.09$ & un\ 2010.79 & $0.004 \pm 0.001$ & $ 5.82 \pm 0.13 $ & $ 123.5 \pm 1.3 $ & $ 0.67 \pm 0.07$ & C2\ 2010.79 & $0.011 \pm 0.001$ & $ 7.21 \pm 0.17 $ & $ 125.2 \pm 1.4 $ & $ 0.87 \pm 0.09$ & C1\ 2010.79 & $0.005 \pm 0.001$ & $ 8.81 \pm 0.24 $ & $ 125.9 \pm 1.6 $ & $ 1.21 \pm 0.12$ & C0\ 2011.17 & $0.322 \pm 0.032$ & …& …& $ 0.10 \pm 0.01$ & C\ 2011.17 & $0.048 \pm 0.005$ & $ 0.28 \pm 0.02 $ & $ 131.6 \pm 4.0 $ & $ 0.10 \pm 0.01$ & S\ 2011.17 & $0.012 \pm 0.001$ & $ 0.93 \pm 0.06 $ & $ 124.2 \pm 3.8 $ & $ 0.31 \pm 0.03$ & C5\ 2011.17 & $0.014 \pm 0.001$ & $ 1.94 \pm 0.13 $ & $ 124.4 \pm 3.8 $ & $ 0.65 \pm 0.07$ & C4\ 2011.17 & $0.003 \pm 0.001$ & $ 5.39 \pm 0.16 $ & $ 122.9 \pm 1.7 $ & $ 0.82 \pm 0.08$ & C3\ 2011.17 & $0.009 \pm 0.001$ & $ 6.85 \pm 0.13 $ & $ 124.5 \pm 1.1 $ & $ 0.66 \pm 0.07$ & C2\ 2011.17 & $0.008 \pm 0.001$ & $ 7.95 \pm 0.18 $ & $ 125.3 \pm 1.3 $ & $ 0.89 \pm 0.09$ & C1\ 2011.40 & $0.251 \pm 0.025$ & …& …& $ 0.10 \pm 0.01$ & C\ 2011.40 & $0.048 \pm 0.005$ & $ 0.34 \pm 0.02 $ & $ 135.1 \pm 3.2 $ & $ 0.10 \pm 0.01$ & S\ 2011.40 & $0.010 \pm 0.001$ & $ 0.68 \pm 0.02 $ & $ 127.5 \pm 1.6 $ & $ 0.10 \pm 0.01$ & C6\ 2011.40 & $0.006 \pm 0.001$ & $ 1.38 \pm 0.08 $ & $ 126.7 \pm 3.4 $ & $ 0.41 \pm 0.04$ & C5\ 2011.40 & $0.008 \pm 0.001$ & $ 2.45 \pm 0.17 $ & $ 124.5 \pm 4.0 $ & $ 0.85 \pm 0.09$ & C4\ 2011.40 & $0.006 \pm 0.001$ & $ 6.15 \pm 0.18 $ & $ 124.6 \pm 1.7 $ & $ 0.90 \pm 0.09$ & C3\ 2011.40 & $0.008 \pm 0.001$ & $ 7.06 \pm 0.11 $ & $ 125.0 \pm 0.9 $ & $ 0.56 \pm 0.06$ & C2\ 2011.40 & $0.005 \pm 0.001$ & $ 8.09 \pm 0.16 $ & $ 125.2 \pm 1.1 $ & $ 0.78 \pm 0.08$ & C1\ 2011.53 & $0.194 \pm 0.019$ & …& …& $ 0.10 \pm 0.01$ & C\ 2011.53 & $0.053 \pm 0.005$ & $ 0.26 \pm 0.02 $ & $ 133.4 \pm 4.2 $ & $ 0.10 \pm 0.01$ & S\ 2011.53 & $0.013 \pm 0.001$ & $ 0.70 \pm 0.02 $ & $ 129.4 \pm 1.6 $ & $ 0.10 \pm 0.01$ & C6\ 2011.53 & $0.006 \pm 0.001$ & $ 1.45 \pm 0.08 $ & $ 124.8 \pm 3.1 $ & $ 0.39 \pm 0.04$ & C5\ 2011.53 & $0.007 \pm 0.001$ & $ 2.69 \pm 0.17 $ & $ 125.8 \pm 3.6 $ & $ 0.86 \pm 0.09$ & C4\ 2011.53 & $0.005 \pm 0.001$ & $ 6.27 \pm 0.19 $ & $ 123.4 \pm 1.7 $ & $ 0.96 \pm 0.10$ & C3\ 2011.53 & $0.010 \pm 0.001$ & $ 7.19 \pm 0.11 $ & $ 124.9 \pm 0.9 $ & $ 0.55 \pm 0.06$ & C2\ 2011.53 & $0.007 \pm 0.001$ & $ 8.99 \pm 0.37 $ & $ 125.0 \pm 2.3 $ & $ 1.83 \pm 0.18$ & C1\ 2011.99 & $0.314 \pm 0.031$ & …& …& $ 0.10 \pm 0.01$ & C\ 2011.99 & $0.021 \pm 0.002$ & $ 0.34 \pm 0.02 $ & $ 129.6 \pm 3.2 $ & $ 0.10 \pm 0.01$ & S\ 2011.99 & $0.006 \pm 0.001$ & $ 0.89 \pm 0.02 $ & $ 129.7 \pm 1.2 $ & $ 0.10 \pm 0.01$ & C6\ 2011.99 & $0.008 \pm 0.001$ & $ 1.44 \pm 0.10 $ & $ 130.9 \pm 3.9 $ & $ 0.49 \pm 0.05$ & C5\ 2011.99 & $0.008 \pm 0.001$ & $ 3.53 \pm 0.34 $ & $ 121.3 \pm 5.5 $ & $ 1.71 \pm 0.17$ & C4\ 2011.99 & $0.009 \pm 0.001$ & $ 6.66 \pm 0.21 $ & $ 124.5 \pm 1.8 $ & $ 1.03 \pm 0.10$ & C3\ 2011.99 & $0.006 \pm 0.001$ & $ 7.52 \pm 0.07 $ & $ 125.0 \pm 0.6 $ & $ 0.37 \pm 0.04$ & C2\ 2011.99 & $0.006 \pm 0.001$ & $ 8.29 \pm 0.15 $ & $ 124.2 \pm 1.0 $ & $ 0.75 \pm 0.08$ & C1\ \[tab:modelfit\] [c c c c c c]{} Epoch & S & r & PA & FWHM & ID\ & (Jy) & (mas) & ($^{\circ}$) & (mas) &\ 2012.53 & $0.181 \pm 0.018$ & …& …& $ 0.10 \pm 0.01$ & C\ 2012.53 & $0.029 \pm 0.003$ & $ 0.30 \pm 0.02 $ & $ 132.8 \pm 3.7 $ & $ 0.10 \pm 0.01$ & S\ 2012.53 & $0.004 \pm 0.001$ & $ 1.02 \pm 0.05 $ & $ 128.9 \pm 2.5 $ & $ 0.23 \pm 0.02$ & C6\ 2012.53 & $0.005 \pm 0.001$ & $ 2.34 \pm 0.16 $ & $ 125.1 \pm 3.8 $ & $ 0.78 \pm 0.08$ & C5\ 2012.53 & $0.006 \pm 0.001$ & $ 4.67 \pm 0.29 $ & $ 124.0 \pm 3.6 $ & $ 1.46 \pm 0.15$ & C4\ 2012.53 & $0.004 \pm 0.001$ & $ 7.04 \pm 0.03 $ & $ 125.2 \pm 0.2 $ & $ 0.15 \pm 0.02$ & C3\ 2012.53 & $0.005 \pm 0.001$ & $ 7.69 \pm 0.10 $ & $ 124.6 \pm 0.7 $ & $ 0.50 \pm 0.05$ & C2\ 2012.53 & $0.004 \pm 0.001$ & $ 8.89 \pm 0.19 $ & $ 125.2 \pm 1.2 $ & $ 0.96 \pm 0.10$ & C1\ 2012.98 & $0.194 \pm 0.019$ & …& …& $ 0.10 \pm 0.01$ & C\ 2012.98 & $0.040 \pm 0.004$ & $ 0.33 \pm 0.02 $ & $ 125.5 \pm 3.3 $ & $ 0.10 \pm 0.01$ & S\ 2012.98 & $0.006 \pm 0.001$ & $ 0.88 \pm 0.04 $ & $ 133.0 \pm 2.9 $ & $ 0.22 \pm 0.02$ & C6\ 2012.98 & $0.004 \pm 0.001$ & $ 2.45 \pm 0.31 $ & $ 124.3 \pm 7.1 $ & $ 1.53 \pm 0.15$ & C5\ 2012.98 & $0.009 \pm 0.001$ & $ 5.03 \pm 0.38 $ & $ 123.5 \pm 4.4 $ & $ 1.92 \pm 0.19$ & C4\ 2012.98 & $0.003 \pm 0.001$ & $ 6.56 \pm 0.08 $ & $ 133.4 \pm 0.7 $ & $ 0.41 \pm 0.04$ & C3\ 2012.98 & $0.011 \pm 0.001$ & $ 7.88 \pm 0.37 $ & $ 124.9 \pm 2.7 $ & $ 1.85 \pm 0.19$ & C2\ 2013.52 & $0.139 \pm 0.014$ & …& …& $ 0.10 \pm 0.01$ & C\ 2013.52 & $0.031 \pm 0.003$ & $ 0.36 \pm 0.02 $ & $ 128.5 \pm 3.1 $ & $ 0.10 \pm 0.01$ & S\ 2013.52 & $0.019 \pm 0.002$ & $ 0.67 \pm 0.02 $ & $ 126.3 \pm 1.6 $ & $ 0.10 \pm 0.01$ & NC\ 2013.52 & $0.003 \pm 0.001$ & $ 1.26 \pm 0.08 $ & $ 132.4 \pm 3.7 $ & $ 0.41 \pm 0.04$ & C6\ 2013.52 & $0.002 \pm 0.001$ & $ 2.15 \pm 0.02 $ & $ 124.5 \pm 0.5 $ & $ 0.10 \pm 0.01$ & un\ 2013.52 & $0.003 \pm 0.001$ & $ 4.53 \pm 0.21 $ & $ 129.3 \pm 2.6 $ & $ 1.03 \pm 0.10$ & C5\ 2013.52 & $0.008 \pm 0.001$ & $ 6.19 \pm 0.16 $ & $ 124.3 \pm 1.5 $ & $ 0.79 \pm 0.08$ & C4\ 2013.52 & $0.008 \pm 0.001$ & $ 7.07 \pm 0.08 $ & $ 124.7 \pm 0.6 $ & $ 0.39 \pm 0.04$ & C3\ 2013.52 & $0.004 \pm 0.001$ & $ 8.10 \pm 0.17 $ & $ 124.5 \pm 1.2 $ & $ 0.84 \pm 0.08$ & C2\ 2013.52 & $0.002 \pm 0.001$ & $10.28 \pm 0.08 $ & $ 127.4 \pm 0.4 $ & $ 0.38 \pm 0.04$ & C1\ [^1]: E-mail: [email protected] [^2]: E-mail: [email protected] [^3]: E-mail: [email protected] [^4]: [www.mpifr-bonn.mpg.de/div/vlbi/fgamma/fgamma.html](www.mpifr-bonn.mpg.de/div/vlbi/fgamma/fgamma.html) [^5]: [www.physics.purdue.edu/MOJAVE/](www.physics.purdue.edu/MOJAVE/) [^6]: Single-dish multi-frequency data from the F-GAMMA program covering the period MJD55400–56720 are publicly available via <http://cdsarc.u-strasbg.fr/viz-bin/qcat?J/A+A/575/A55> [^7]: For a stacked image of all naturally-weighted CLEAN maps from the MOJAVE survey see www.physics.purdue.edu/astro/MOJAVE/stackedimages/ 0321+340.u.stacked.icn.gif [^8]: Note a few possible exceptions, in the following sense: while the $\gamma$-bright source PKS2004-447 shows many blazar-like characteristics, the angle of the jet to the line of sight is likely fairly large as indicated by the persistent steep radio spectrum and the diffuse emission on the counter-jet side [@2015arXiv150903735K; @2015arXiv151102631S]. Further exceptions may be the candidate $\gamma$ NLS1 galaxies B31441+476 and RXJ2314.9+2243, which both show steep radio spectra . [^9]: See [@2013MNRAS.436..191D] for an exception.

ArXiv

.5in [Cutting the cylinder into squares: The square form factor]{} 0.4in [**Juan Miguel Nieto**]{} 0.1in Departamento de Física Teórica\ Universidad Complutense de Madrid\ $28040$ Madrid, Spain\ .4in **Abstract** .1in In this article we present a method for constructing two-point functions in the spirit of the hexagon proposal, which leads us to propose a “square form factor”. Since cutting the square gives us two squares, we can write a consistency condition that heavily constrains such form factors. In particular, we are able to use this constraint to reconstruct the Gaudin through the forest expansion of the determinant appearing in its definition. We also use this procedure to compute the norm of off-shell Bethe states for some simple cases. .4in Introduction ============ A consistent method to define a quantum field theory is to construct observables starting from form factors. Form factors are matrix elements of local observables computed in the basis of asymptotic ingoing and outgoing states. Such matrix elements fulfil the *Smirnov’s Axioms* [@Smirnov], a set of restrictions on their analytical properties. This construction is non-perturbative and restricts the reconstructed theory to be local and a Wightman field theory. This construction has been successfully applied to different systems like the Sine-Gordon model [@Babujian_1999; @Babujian_2002], the Thirring model [@Nakayashiki_2002; @Takeyama_2003] and the SU(N) Lieb-Lininger model [@Babujian_2006]. Concerning the $AdS_5/CFT_4$ correspondence, we can find the first steps towards understanding form factors in [@WSFF1; @WSFF2], where the authors focus on the spin chain description and the near-plane-wave limit of the string description. Regarding direct computations of correlation functions, the Quantum Spectral Curve has already solved (at least formally) the spectral problem in $\mathcal{N}=4$ SYM [@GKLV1; @GFS1; @GKLV2; @CV1; @CV2; @GFS2; @KLV; @CV3; @CV4; @CV5] and thus the computation of two-point functions due to conformal symmetry. As a consequence, the focus has partially shifted to three-point functions. Although neither the interest nor the use of integrability are new in this context (see, for example, [@3pf1; @3pf2; @vertex1; @3pf3]), recently a very simple and successful proposal for an all-loop computation of three-point functions based on cutting the pair of pants into two separated hexagons has been put forward [@BKV; @Hex2; @Hex3]. Interestingly, these hexagons can be understood as a form factor involving a defect operator. Since the operator involved is non-local, these form factor fulfil a slightly modified version Smirnov’s axioms [@nonlocalFF]. It has been noticed that the hexagon form factors alone are enough to construct higher point functions [@Hexagonalization1; @Hexagonalization2; @Hexagonalization3; @Hexagonalization4] due to the feasibility of obtaining form factors with a larger number of sides just by gluing hexagons. In this article we argue that, together with the hexagon form factor, we can define a square “form factor”, more akin to usual form factors, related to the cutting of the cylinder (see figure \[cuttingprocedure\]). We write form factor between quotation marks because it would correspond to the form factor of the identity, which is sometimes not considered truly a form factor but still satisfy Smirnov’s axioms. However, here we will construct the square starting from states constructed using the Algebraic Bethe Ansatz instead of states constructed with Zamolodchikov-Faddeev (ZF) operators, hence it does not fulfil the usual Smirnov’s axioms but a different ones. Despite their differences there exists a relationship between both squares, we will postpone the analysis of ZF squares to future works. One of the most important differences between the square and the hexagon form factor refers to the processes of gluing and cutting: when we glue together two squares we get again a square and vice-versa when cutting. We can exploit this feature to write down restrictions on the analytic properties of squares and reconstruct the scalar product of states. In particular, we will explicitly reconstruct the Gaudin determinant, as we are going to focus mainly on a system with $SU(2)$ symmetry. (0,-0.5) – (0,1.5); (1,-0.5) – (1,1.5); (0.5,1.5) ellipse (0.5 and 0.2); (1,-0.5) arc (0:180:0.5 and 0.2); (0,-0.5) arc (180:360:0.5 and 0.2); (0.5,1.3) circle \[radius=0.04\]; (0.5,-0.7) circle \[radius=0.04\]; (0.5,1.3) – (0.5,-0.7); (1.3333333,0.4) – (1.6666666,0.4); (1.3333333,0.6) – (1.6666666,0.6); (2,0) –(2,1); (2, 0) – (3,0); (2, 1) – (3,1); (3,0) –(3,1); (2.5,1) circle \[radius=0.04\]; (2.5,0) circle \[radius=0.04\]; (2.5,0) – (2.5,1); (3.3333333,0.5) – (3.6666666,0.5); (3.5,0.3333333) – (3.5,0.66666666); (4,0) –(4,1); (4, 0) – (5,0); (4, 1) – (5,1); (5,0) –(5,1); (4.5,1) circle \[radius=0.04\]; (4.5,0) circle \[radius=0.04\]; (4.5,0) – (4.5,1); (3.96,0.46) rectangle (4.04,0.54); (4.96,0.46) rectangle (5.04,0.54); (4,0.5) – (5,0.5); (5.3333333,0.5) – (5.6666666,0.5); (5.5,0.3333333) – (5.5,0.66666666); (6,0) –(6,1); (6, 0) – (7,0); (6, 1) – (7,1); (7,0) –(7,1); (6.5,1) circle \[radius=0.04\]; (6.5,0) circle \[radius=0.04\]; (6.5,0) – (6.5,1); (5.96,0.293333333) rectangle (6.04,0.373333333); (6.96,0.293333333) rectangle (7.04,0.373333333); (5.96,0.626666666666) rectangle (6.04,0.706666666666); (6.96,0.626666666666) rectangle (7.04,0.706666666666); (6,0.333333333) – (7,0.33333333333333); (6,0.666666666) – (7,0.66666666666666); (7.3333333,0.5) – (7.6666666,0.5); (7.5,0.3333333) – (7.5,0.66666666); (7.9,0.5) circle \[radius=0.03\]; (8.1,0.5) circle \[radius=0.03\]; (8.3,0.5) circle \[radius=0.03\]; Another interesting characteristic of the square is how mirror excitations appear in the gluing procedure. In contrast with the hexagon form factor, the $\mathcal{N}=4$ SYM square “form factor” is invariant under the full $PSU(2|2)^2_{\text{c.e.}}$ symmetry. As the central extension includes the dilatation operator, some configurations have to vanish and, in particular, the number of excitations in the upper and lower edges has to be the same when no mirror excitations are present. Thus, the gluing of two squares should involve only a finite sum due to symmetry instead of the infinite sum that appears when gluing of two hexagons. This makes the square a simpler setting where we can get a better understanding of this procedure. It is worth pointing that a different square form factor from the one proposed here has appeared in the context of correlation functions of three separated operators connected by Wilson lines [@HexagonWL]. Despite the different nature of both squares, we expect them to fulfil similar consistency conditions. The article is structured as follows. In section 2 we review some concepts related to the Algebraic Bethe Ansatz. In section 3 we present the consistency condition for the square and use it to reconstruct the normalization of on-shell Bethe states of the $SU(2)$ spin chain. In section 4 we extend the analysis to compute the normalization of off-shell Bethe states for the case of two and three excitations. We also compute the three highest orders in the large length expansion of the normalization of general $SU(2)$ off-shell states. In section 5 we close with summary, conclusions and comments on some points that should be clarified in further works. In appendix A we review Kirchhoff’s matrix-tree theorem and its application to the Gaudin determinant. In appendix B we comment a little more on the form of the consistency condition for $SU(N)$ spin chains. In appendix C we explain how to recover the space-time dependence of the two point function. In appendix D we expand on the constraints imposed by crossing symmetry. In appendix E we sketch the computation of the leading mirror corrections to the squares with zero and one physical excitation. The $SU(2)$ Algebraic Bethe Ansatz ================================== In this section we present the basis of the Algebraic Bethe Ansatz (ABA) for $SU(2)$ models in order to fix the notation we will use through this article. We are going to work with the rational 6-vertex model R-matrix $$\begin{aligned} R_{1,2} (\lambda) &=\alpha (u,v) \sum_{a=1}^2 E_{aa} \otimes E_{aa} +\beta(u,v) \sum_{1\leq a\neq b \leq 2} E_{aa} \otimes E_{bb} + \sum_{1\leq a\neq b \leq 2} \gamma_{ab} (u,v) E_{ab} \otimes E_{ba} \notag \\ &=\lambda (\mathbb{I}_a \otimes \mathbb{I}_b ) +\frac{i\eta }{2} \sum_j{(\sigma^j_a \otimes \sigma^j_b)} \ ,\end{aligned}$$ but our results can be easily generalized to the trigonometric 6-vertex model by an appropriate substitution of the functions $\alpha$, $\beta$ and $\gamma$. Here $E_{ab}$ is the matrix with zeros in every position except the $a,b$ position, which contains a $1$. The monodromy matrix, written as a matrix on the auxiliary space, has the following structure $$T_a (\lambda )=\overrightarrow{\prod_{j=1}^L} \mathcal{L}_{a,j}(u)=\mathcal{L}_{a,1}(u) \mathcal{L}_{a,2}(u) \dots \mathcal{L}_{a,L}(u) =\left( \begin{array}{cc} A(\lambda ) & B(\lambda ) \\ C(\lambda ) & D(\lambda ) \end{array}\right) \ , \label{Monodromy}$$ where $\mathcal{L}_{a,j}(u)$ is the Lax matrix associated to the specific model. The action of these four operators over the reference state is given by $$\begin{aligned} A(\lambda ) {\left| 0 \right\rangle} &= a(\lambda ) {\left| 0 \right\rangle} \ , & B(\lambda ) {\left| 0 \right\rangle} &\neq 0 \ , \\ D(\lambda ) {\left| 0 \right\rangle} &= d(\lambda ) {\left| 0 \right\rangle} \ , & C(\lambda ) {\left| 0 \right\rangle} &= 0 \ ,\end{aligned}$$ while their commutation relations are fixed by the RTT relation $$R_{1,2} (\lambda - \mu ) T_1 (\lambda ) T_2 (\mu) = T_2 (\mu) T_1 (\lambda) R_{1,2} (\lambda - \mu ) \label{RTT} \ .$$ Of the relations codified in this equation, we will only need the following seven of them $$\begin{aligned} &[B(\lambda ) , B(\mu )]=[C(\lambda ) , C(\mu )]=0 \ , \label{commBB}\\ &A(\lambda ) B(\mu )=f(\mu , \lambda) B(\mu ) A(\lambda ) +g(\lambda , \mu ) B(\lambda ) A(\mu ) \ ,\label{commAB}\\ &D(\lambda ) B(\mu )=f(\lambda, \mu) B(\mu ) D(\lambda ) + g(\mu , \lambda ) B(\lambda ) D(\mu ) \ , \label{commDB} \\ &C(\mu ) A(\lambda )=f(\mu , \lambda) A(\lambda ) C(\mu ) +g(\lambda , \mu ) A(\mu ) C(\lambda ) \ ,\label{commCA}\\ &C(\mu ) D(\lambda )=f(\lambda, \mu) D(\lambda ) C(\mu ) + g(\mu , \lambda ) D(\mu ) C(\lambda ) \ , \label{commCD}\\ &[C(\lambda ) , B(\mu) ]=g(\lambda , \mu ) \left[ A(\lambda ) D(\mu ) - A(\mu ) D(\lambda ) \right] \ , \label{commCB}\end{aligned}$$ where, for convenience, we have introduced the functions [^1] $$\begin{aligned} f(\lambda , \mu) &=\frac{\alpha (\lambda , \mu)}{\beta(\lambda , \mu)} \ , & f(\lambda , \mu) &=\frac{\lambda - \mu +i}{\lambda - \mu} \ , & f(\mu , \lambda) &=\frac{\lambda - \mu -i}{\lambda - \mu} \ , \\ g(\lambda , \mu )&=\frac{\gamma_{12} (\lambda , \mu)}{\beta(\lambda , \mu)}\ , & g(\lambda , \mu )&=\frac{i}{\lambda - \mu} \ , & g(\mu , \lambda )&=\frac{-i}{\lambda - \mu} \ .\end{aligned}$$ Since the trace of the monodromy matrix, called transfer matrix, commute with itself for different values of the spectral parameter $\lambda$, its power expansion produces $L-1$ commuting operators, any of which can be chosen as Hamiltonian. We can construct the Hilbert space by applying a stack of $B$ operators over the reference state, but such space is overcomplete as not all the states created by this process are eigenstates of the transfer matrix. The states that are not eigenstates of the transfer matrix are not physical, hence we will call them *off-shell Bethe states*. Only the states constructed with B operators with some precise sets of rapidities are eigenstates of the transfer matrix and hence true physical states. We will call such states *on-shell Bethe states*. The admissible sets of rapidities are those that fulfil the Bethe Ansatz Equations $$\frac{a (u_j)}{d(u_j)}=\prod_{k\neq j} \frac{f(u_j,u_k)}{f(u_k,u_j)} \ . \label{BAE}$$ As most of the time we will work with sets of rapidities, in the remainder of the article the following shorthand notation will prove useful as it reduces the amount of products on the formulas $$\begin{aligned} f( \{\bar{\alpha}\},\{\alpha\}) &=\prod_{\substack{u_i\in \alpha \\ v_j \in \bar{\alpha}}} f(v_j,u_i) \ , & f^{\neq} (\{\alpha \} , \{ \alpha \} ) &=\prod_{\substack{u_i,u_j\in \alpha \\ u_i\neq u_j}} f(u_i , u_j) \ , & a (\{\bar{\alpha}\}) &=\prod_{v_j \in \bar{\alpha}} a (v_j) \ .\end{aligned}$$ From the Gaudin determinant to the square and back ================================================== In this section we propose the *square form factor* and a consistency condition for it. After proposing this consistency equation, we verify that the Gaudin norm fulfils it as a check for our proposal. Reversing the argument, we are able to reconstruct the Gaudin determinant from the consistency condition and some initial conditions and analytic considerations. The square bootstrap and the Gaudin determinant ----------------------------------------------- Inspired by the BKV idea of cutting the three-point correlation function into two hexagon form factors [@BKV], we apply the same procedure to two-point functions, i.e. a cylinder worldsheet. In this case we will obtain two squares after cutting two times, each with two physical sides and two mirror sides, being the last the new sides created by this process. We decided to define our square by cutting two times the cylinder instead of one time because in the later case the two mirror sides of the square are constrained to have the same number of excitations (see figure \[cuttingprocedure\]). The most general square is characterized by four sets of rapidities, one for each of the physical sides and one for each of the mirror sides. For the most part of this article the two mirror sides will be empty, so we only have to indicate the content of the two physical sides. Therefore we denote the square by $S_L ( \u \rightarrow \{ w \})$, where the subindex indicates the length of the physical sides. If both sets of rapidities are equal, we will just write $S_L ( \u )=S_L ( \u \rightarrow \{ u \})$ to alleviate notation. With this notation the relationship between the scalar product of Bethe states and the square is given by $$\Big\langle 0 \Big| \prod_i C(v_i) \prod_j B(u_j) \Big| 0 \Big\rangle =S_L (\{ u \} \rightarrow \{ v \} ) + \text{ mirror corrections} \ .$$ The superposition of two Bethe states gives us the two-point function of an operator at zero with an operator at infinity, i.e., the normalization of the two-point functions. Despite this being an unphysical quantity, it is important to compute it to correctly normalize our states. In this article we will consider the $L\gg 1$ limit, so we can ignore mirror corrections. In appendix \[mirrorappendix\] we will comment about how to include mirror corrections. Although the construction of the hexagon form factor from symmetry and crossing arguments is very simple and elegant, a similar construction for the square presents some obstructions. Since here we are addressing the problem from the weak coupling limit, we do not have access to the crossing transformation. This transformation entails shifting the rapidity of an excitation in such a way that it crosses some particular cuts in the rapidity plane. However these cuts have a width of $4g$, thus they close at weak coupling and the mirror and crossing transformations are inaccessible perturbatively [@BetheGKP]. Nevertheless, an all-loop construction of the square should behave nicely under crossing transformations, e.g. we expect that $S_L ( \{ u^{\text{cross}} \} \rightarrow \{ w^{\text{cross}} \})=S_L ( \{ w \} \rightarrow \{ u \})$. Among the several differences that exist between this square and the hexagon form factor, we are interested in the one concerning the cutting and the gluing of form factors. The cutting operation applied to the square remains well defined after we employ it once, [^2] which means that we can continue cutting a square into smaller ones. We can reverse the argument and state that gluing two squares reconstructs one square. This idea can be used as a “bootstrap procedure” to construct such square. If we consider only the asymptotic part of the square (which means directly gluing them without adding excitations to the mirror edges), the reconstruction formula will be given by $$S_L ( \{ u \} \rightarrow \{ v \} )=\sum_{\substack{\alpha \cup \bar{\alpha} =\{ u\} \\ \beta \cup \bar{\beta} =\{ v\}}} w (\{ \alpha \} , \{ \bar{\alpha} \} ,\{ \beta \} , \{ \bar{\beta} \} ) S_{l_1} ( \{ \alpha \} \rightarrow \{ \beta \} ) S_{l_2} ( \{ \bar{\alpha} \} \rightarrow \{ \bar{\beta} \} ) \ , \label{bootstrap}$$ where the weight $w (\{ \alpha \} , \{ \bar{\alpha} \} ,\{ \beta \} , \{ \bar{\beta} \} )$, which we will call *breaking factor*, takes into account the cutting of the physical states into two. We have also defined $L=l_1+l_2$. In this article we will work in the $L \sim l_1 \sim l_2 \gg 1$, so we will not consider any potential problems that could arise from the discrete nature of the states when we cut them. As we said before, this limit also allow us to ignore the mirror corrections as they are exponentially suppressed. To construct the breaking factors we proceed in a similar manner as [@composite], i.e. we write the monodromy matrix as the product of two monodromy matrices, each associated to a subchain of the original chain. Using the definition of the monodromy matrix given in equation (\[Monodromy\]) we can rewrite the product of $L$ Lax matrices into two separate products with $l_1$ and $l_2$ Lax matrices each. Equating the entries of the auxiliary space we can prove that $$T^{(L)}_a(u)=T^{(l_1)} (u) \otimes T^{(l_2)} (u) \Longrightarrow \left\{ \begin{array}{c} B_L (u) = B_{l_1} (u) \otimes D_{l_2} (u) + A_{l_1} (u) \otimes B_{l_2} (u) \\ C_L (u) = C_{l_1} (u) \otimes A_{l_2} (u) + D_{l_1} (u) \otimes C_{l_2} (u) \end{array} \right. \ . $$ These formulas provide us with the recipe to break the creation and annihilation operators for the case of one excitation. For the general square we need the expression of several creation and annihilation operators. As the RTT relations involve at most two rapidities, it is easy to argue that the structure of the weight functions can be fixed from the breaking factor for two excitations. For the case of creation operators we have $$\begin{aligned} B(u) B(v) &= B_{l_1}(u) B_{l_1}(v) \otimes D_{l_2}(u) D_{l_2}(v) + B_{l_1}(u) A_{l_1}(v) \otimes D_{l_2}(u) B_{l_2}(v) \notag \\ &+ A_{l_1}(u) B_{l_1}(v) \otimes B_{l_2}(u) D_{l_2}(v) + A_{l_1}(u) A_{l_1}(v) \otimes B_{l_2}(u) B_{l_2}(v) = \notag \\ &=B_{l_1}(u) B_{l_1}(v) \otimes D_{l_2}(u) D_{l_2}(v) + f(u,v) B_{l_1}(u) A_{l_1}(v) \otimes B_{l_2}(v) D_{l_2}(u) \notag \\ &- g(u,v) B_{l_1}(u) A_{l_1}(v) \otimes B_{l_2}(u) D_{l_2}(v) + f(v,u) B_{l_1}(v) A_{l_1}(u) \otimes B_{l_2}(u) D_{l_2}(v) \notag \\ &+ g(u,v) B_{l_1}(u) A_{l_1}(v) \otimes B_{l_2}(u) D_{l_2}(v) + A_{l_1}(u) A_{l_1}(v) \otimes B_{l_2}(u) B_{l_2}(v) \ .\end{aligned}$$ In the second equality we have used RTT equations (\[commAB\]) and (\[commDB\]) to move all $A$ and $D$ operators to the right of the $B$ operators. Applying this expression to the reference state gives us the contribution from $B$ operators to the weight in the case of two excitations. As argued before, the case of more excitations is just an extension of this computation and gives us an equivalent structure. A similar procedure can be performed for annihilation operators by instead moving the $A$ and $D$ operators to the left of the $C$ operators and applying the expression to the dual reference state. The final expression for the breaking factor is the product of the weights associated to $B$ and $C$ operators $$\begin{aligned} w (\{ \alpha \} , \{ \bar{\alpha} \} ,\{ \beta \} , \{ \bar{\beta} \} ) &= w_B (\{ \alpha \} , \{ \bar{\alpha} \} ) w_C (\{ \beta \} , \{ \bar{\beta} \} ) \label{weights} \\ w_B (\{ \alpha \} , \{ \bar{\alpha} \} ) &= f(\alpha , \bar{\alpha}) a_{l_1} (\bar{\alpha}) d_{l_2} (\alpha) \ , \notag \\ w_C (\{ \beta \} , \{ \bar{\beta} \} ) &= f(\bar{\beta} , \beta) d_{l_1} (\bar{\beta}) a_{l_2} (\beta) \ . \notag\end{aligned}$$ Despite the simplicity of the result, this computation becomes rather messy when we move to $SU(N)$ for $N>2$. A small discussion about it has been collected in appendix \[BFappendix\], together with the computation of the breaking factor for $SU(2)$ when we work with operators that satisfy the Zamolodchikov-Faddeev algebra. Let us perform now a first check of our proposal. The Gaudin determinant provides us the normalization of states of the XXX spin chain written using the Algebraic Bethe Ansatz. Its explicit expression is $$\begin{aligned} G(\{u\}) &=i^N f^{\neq} (\u , \u) \det \Phi \ , & \Phi_{ab} &=-\frac{\partial}{\partial u_b} \left( \frac{a (u_a)}{d (u_a)} \prod_{k\neq a} \frac{f(u_c , u_a)}{f(u_a , u_c)} \right) \ . \label{gaudin}\end{aligned}$$ The proof that such determinant fulfils the consistency condition we have proposed can already be found in the literature, see for example eq. (36) in [@diagonalFF]. Instead, for later convenience it proves better to use the Kirchhoff’s matrix-tree theorem [@Kirchhoff] to rewrite the determinant expression as a sum over graphs. We have summarized this construction in appendix \[Kirchhoffappendix\] and we refer to [@Tree] and references therein for more details. In this forest expansion all the dependence on the length of the chain appears only through the factors associated to the root nodes of the trees. If we break the length $L$ into $l_1+l_2$, we have to add an extra sum over trees having either $l_1 p'$ or $l_2 p'$ as root factor. Grouping terms by root nodes and rapidity labels involved in each set of trees allows us to reconstruct the two factors $S_{l_1} ( \{ \alpha \} )$ and $S_{l_2} ( \{ \bar{\alpha} \} )$, with the corresponding sum coming from the forest expansion and choices of root nodes becoming the sum of equation (\[Monodromy\]). The appearance of the breaking factor comes not from separating the determinant into trees, but from separating the prefactor $f^{\neq} (\u , \u)$. It might seem strange that the Gaudin determinant directly fulfils this relation because the square should be an off-shell construction while the states involved in the Gaudin determinant are on-shell Bethe states. This is so because the terms dependent on the Bethe Ansatz Equations appear as $$S_L (\u)= \{ \text{On-shell part}\} + \{ \text{Bethe eq.} \} \cdot \{ \text{Off-shell contributions} \} \ ,$$ so the on-shell part should also fulfils the same consistency equation. In the next section we will check for some simple examples that the square has this structure. Reconstructing the Gaudin determinant: some particular cases {#gaudinreconstruction} ------------------------------------------------------------ After checking that the Gaudin determinant fulfils the consistency condition (\[bootstrap\]), we are going to try to reconstruct it from the consistency condition. First we present the reconstruction for the particular cases from zero up to three excitations in order to clarify the procedure. After that we will provide the fully reconstructed expression for a general number of excitations, up to some functions that cannot be fixed just using the consistency condition. We will fix them using comparison with explicit ABA computation and analytic considerations. Basic idea of this reconstruction is to use the length structure in the consistency condition to impose constraints. While $S_{l_1}$ and $S_{l_2}$ depend explicitly on $l_1$ and $l_2$ respectively and the weight factors depend on those lengths implicitly though the momentum factors $a$ and $d$, $S_L$ cannot depend on these lengths but on the form $(l_1+l_2)$. Apart from the consistency condition, we will need two additional properties of the square to fix it. In particular we will need that - The square is symmetric under the exchange of two rapidities, i. e. $$S_L (u_1 , \dots , u_i , \dots , u_j , \dots )=S_L (u_1 , \dots , u_j , \dots , u_i , \dots ) \ .$$ - The limit of the norm exits and is well defined (finite and non-vanishing) $$\lim_{v_1 \rightarrow u_1} \lim_{v_2 \rightarrow u_2} \dots \, \, S_L (\{ u \} \rightarrow \{ v \}) =S_L (\u ) \ .$$ These are not the usual Smirnov’s axioms for form factors because we are constructing the consistency condition from the ABA instead of from ZF operators, so the properties satisfied by both are different. In addition, because we are working in the weak coupling regime, we will not deal in this article with the constrains imposed by crossing transformations to this square. (0,0) –(0,1); (0, 0) – (1,0); (0, 1) – (1,1); (1,0) –(1,1); (1.3,0.4) – (1.6,0.4); (1.3,0.6) – (1.6,0.6); (1.9,0) –(1.9,1); (1.9, 0) – (2.7,0); (1.9, 1) – (2.7,1); (2.7,0) –(2.7,1); (3,0) –(3,1); (3, 0) – (3.8,0); (3, 1) – (3.8,1); (3.8,0) –(3.8,1); (4.1,0.4) – (4.4,0.4); (4.1,0.6) – (4.4,0.6); (4.7,0) –(4.7,1); (4.7, 0) – (5.2,0); (4.7, 1) – (5.2,1); (5.2,0) –(5.2,1); (5.3,0) –(5.3,1); (5.3, 0) – (5.8,0); (5.3, 1) – (5.8,1); (5.8,0) –(5.8,1); (5.9,0) –(5.9,1); (5.9, 0) – (6.4,0); (5.9, 1) – (6.4,1); (6.4,0) –(6.4,1); (6.7,0.4) – (7,0.4); (6.7,0.6) – (7,0.6); (7.3,0.5) circle \[radius=0.03\]; (7.45,0.5) circle \[radius=0.03\]; (7.6,0.5) circle \[radius=0.03\]; The first case we should compute is the square with no excitations, which is fixed up to an unphysical normalization. The computation relies on the particular fact that such square can be cut as many times as we want (as long as we are far from the discrete limit) involving no other structures. As we can see from figure \[empty\], the consistency condition for it is given by $$S_L (\emptyset )=S_{l_1} (\emptyset ) S_{l_2} (\emptyset )=S_{l'_1} (\emptyset ) S_{l'_2} (\emptyset ) S_{l'_3} (\emptyset )=\dots \ ,$$ whose only physical solution is $S_L (\emptyset ) =1$. [^3] With the information about the empty square, we can move now to the computation of the square with one excitation. As we can see from its pictorial representation in figure \[one\], equation (\[bootstrap\]) reflects the fact that the excitation can end in either of the two new squares we get after we cut the original one. Applied to this case, this equation reduces to $$S_{L} (u)=S_{l_1} (u) S_{l_2} (\emptyset) a_{l_2} (u) d_{l_2} (u)+S_{l_1} (\emptyset) S_{l_2} (u) a_{l_1} (u) d_{l_1} (u) \ .$$ No other term can appear as we are not including mirror excitations. We can see that this equation simplifies by means of the rewriting $S_L (u)=\tilde{S}_L (u) a_L (u) d_L (u)$, reducing to $$\tilde{S}_L (u)=\tilde{S}_{l_1} (u) + \tilde{S}_{l_2} (u) \ .$$ The only way to solve this equation is to have $\tilde{S}_L (u) =\sigma^{(1)}_1 (u) L$, where $\sigma^{(1)}_1$ is a general function not fixed by the constraint. Note that we cannot include a term independent of $L$ as the equation automatically fixes it to zero. (0,0) –(0,1); (0, 0) – (1,0); (0, 1) – (1,1); (1,0) –(1,1); (0.5,0) circle \[radius=0.04\]; (0.5,1) circle \[radius=0.04\]; (0.5,0) – (0.5,1); (1.3,0.4) – (1.6,0.4); (1.3,0.6) – (1.6,0.6); (1.9,0) –(1.9,1); (1.9, 0) – (2.7,0); (1.9, 1) – (2.7,1); (2.7,0) –(2.7,1); (2.3,0) circle \[radius=0.04\]; (2.3,1) circle \[radius=0.04\]; (2.3,0) – (2.3,1); (2.9,0) –(2.9,1); (2.9, 0) – (3.7,0); (2.9, 1) – (3.7,1); (3.7,0) –(3.7,1); (2.9,1.05) – (2.9,1.15); (3.7,1.05) – (3.7,1.15); (2.9,1.1) – (3.7,1.1); at (3.3,1.1) [$\scriptstyle l_2$]{}; (4,0.5) – (4.3,0.5); (4.15,0.35) – (4.15,0.65); (4.6,0) –(4.6,1); (4.6, 0) – (5.4,0); (4.6, 1) – (5.4,1); (5.4,0) –(5.4,1); (4.6,1.05) – (4.6,1.15); (5.4,1.05) – (5.4,1.15); (4.6,1.1) – (5.4,1.1); at (5,1.1) [$\scriptstyle l_1$]{}; (5.6,0) –(5.6,1); (5.6, 0) – (6.4,0); (5.6, 1) – (6.4,1); (6.4,0) –(6.4,1); (6,0) circle \[radius=0.04\]; (6,1) circle \[radius=0.04\]; (6,0) – (6,1); We can fix the structure of the $\sigma^{(1)}_1 (u)$ function if we compute the case in which the rapidity of the initial (lower-edge) excitation is different from the final (upper-edge) one and impose some analytic considerations. The consistency condition in this case reads $$S_{L} (u\rightarrow v)=S_{l_1} (u\rightarrow v) S_{l_2} (\emptyset) a_{l_2} (v) d_{l_2} (u)+S_{l_1} (\emptyset) S_{l_2} (u\rightarrow v) a_{l_1} (u) d_{l_1} (v) \ .$$ To solve this equation we require the more general ansatz $$S_{L} (u\rightarrow v)= \tilde{S}_{L} (u\rightarrow v) \left[ a(u) d(v) - a(v) d(u) \right] \ ,$$ which solves the consistency equation for any function $\tilde{S}_{L} (u\rightarrow v)$. In order to get a non-vanishing norm we have to impose that $\tilde{S}_{L} (u\rightarrow v)$ has a pole of order one at $u=v$. This function can be further fixed through two different procedures. The first one is the direct comparison with the ABA matrix element ${\left\langle 0 \right|} C(v) B(u) {\left| 0 \right\rangle}$ using (\[commCB\]), which directly fixes $\tilde{S}_{L} (u\rightarrow v) =g(u,v)$. The second one is to use crossing symmetry. However, as we commented above, at weak coupling the crossing transformation is not accessible, so in order to extract the function from this symmetry we have to take the weak coupling limit after performing the transformation. As this involves dealing with the non-perturbative square, we have relegated the discussion on this topic to appendix \[CrossingSquare\]. With the information from $S_{L} (u\rightarrow v)$ we can now extract the Gaudin norm associated to the one-excitation state $$\tilde{S}_L (u)= \frac{S_{L} (u)}{a(u) d(u)}=\lim_{v\rightarrow u} \tilde{S}_{L} (u\rightarrow v) \frac{d(v)}{a(u)} (e^{ip(v) L}- e^{ip(u) L} )= \text{const.} (-i p' L) \ ,$$ where the momentum is defined as $ip(u)L=\log [a(u)/d(u)]$ and $p'=\frac{dp(u)}{du}$. We can see that the expression we have obtained agrees with the Gaudin norm (\[gaudin\]) for one excitation. (0,0) –(0,1); (0, 0) – (1,0); (0, 1) – (1,1); (1,0) –(1,1); (0.33333,0) circle \[radius=0.04\]; (0.33333,1) circle \[radius=0.04\]; (0.33333,0) – (0.333333,1); (0.66666,0) circle \[radius=0.04\]; (0.66666,1) circle \[radius=0.04\]; (0.66666,0) – (0.666666,1); (1.3,0.4) – (1.6,0.4); (1.3,0.6) – (1.6,0.6); (1.9,0) –(1.9,1); (1.9, 0) – (2.7,0); (1.9, 1) – (2.7,1); (2.7,0) –(2.7,1); (2.16666,0) circle \[radius=0.04\]; (2.16666,1) circle \[radius=0.04\]; (2.16666,0) – (2.16666,1); (2.43333,0) circle \[radius=0.04\]; (2.43333,1) circle \[radius=0.04\]; (2.43333,0) – (2.43333,1); (2.9,0) –(2.9,1); (2.9, 0) – (3.7,0); (2.9, 1) – (3.7,1); (3.7,0) –(3.7,1); (4,0.5) – (4.3,0.5); (4.15,0.35) – (4.15,0.65); (4.6,0) –(4.6,1); (4.6, 0) – (5.4,0); (4.6, 1) – (5.4,1); (5.4,0) –(5.4,1); (5.6,0) –(5.6,1); (5.6, 0) – (6.4,0); (5.6, 1) – (6.4,1); (6.4,0) –(6.4,1); (5.86666,0) circle \[radius=0.04\]; (5.86666,1) circle \[radius=0.04\]; (5.86666,0) – (5.86666,1); (6.13333,0) circle \[radius=0.04\]; (6.13333,1) circle \[radius=0.04\]; (6.13333,0) – (6.13333,1); (1.35,0.5-1.5) – (1.65,0.5-1.5); (1.5,0.35-1.5) – (1.5,0.65-1.5); (1.9,0-1.5) –(1.9,1-1.5); (1.9, 0-1.5) – (2.7,0-1.5); (1.9, 1-1.5) – (2.7,1-1.5); (2.7,0-1.5) –(2.7,1-1.5); (2.3,0-1.5) circle \[radius=0.04\]; (2.3,1-1.5) circle \[radius=0.04\]; (2.3,0-1.5) – (2.3,1-1.5); (2.9,0-1.5) –(2.9,1-1.5); (2.9, 0-1.5) – (3.7,0-1.5); (2.9, 1-1.5) – (3.7,1-1.5); (3.7,0-1.5) –(3.7,1-1.5); (3.3,0-1.5) circle \[radius=0.04\]; (3.3,1-1.5) circle \[radius=0.04\]; (3.3,0-1.5) – (3.3,1-1.5); (4,0.5-1.5) – (4.3,0.5-1.5); (4.15,0.35-1.5) – (4.15,0.65-1.5); (4.6,0-1.5) –(4.6,1-1.5); (4.6, 0-1.5) – (5.4,0-1.5); (4.6, 1-1.5) – (5.4,1-1.5); (5.4,0-1.5) –(5.4,1-1.5); (5,0-1.5) circle \[radius=0.04\]; (5,1-1.5) circle \[radius=0.04\]; (5,0-1.5) – (5,1-1.5); (5.6,0-1.5) –(5.6,1-1.5); (5.6, 0-1.5) – (6.4,0-1.5); (5.6, 1-1.5) – (6.4,1-1.5); (6.4,0-1.5) –(6.4,1-1.5); (6,0-1.5) circle \[radius=0.04\]; (6,1-1.5) circle \[radius=0.04\]; (6,0-1.5) – (6,1-1.5); Moving now to the case of two excitations, equation \[bootstrap\] (pictorially represented in fig. \[two\]) now reads $$\begin{aligned} S_{L}(u,v) &=S_{l_1}(u,v) S_{l_2} (\emptyset) a_{l_2} (u) a_{l_2} (v) d_{l_2} (u) d_{l_2} (v)+S_{l_1} (\emptyset) S_{l_2}(u,v) a_{l_1} (u) a_{l_1} (v) d_{l_1} (u) d_{l_1} (v) \notag \\ &+S_{l_1} (u) S_{l_2} (v) f(u,v) f(v,u) a_{l_1} (v) d_{l_1} (v) a_{l_2} (u) d_{l_2} (u) \notag \\ &+S_{l_1} (v) S_{l_2} (u) f(u,v) f(v,u) a_{l_1} (u) d_{l_1} (u) a_{l_2} (v) d_{l_2} (v)\ . \label{twomagnonansatz}\end{aligned}$$ One might argue that there are two terms we have not taken into account, as we have not included terms involving squares whose excitations in the upper and lower sides have different rapidities. Although this is true in general, in this section we are trying to reconstruct the Gaudin norm and these terms vanish because $S_{L} (\{u\} \rightarrow \{v\neq u\})=0$ when the Bethe equations are imposed. Following the same procedure we used for the case of one excitation, we can factor out the momentum dependence, encoded in the functions $a$ and $d$, by defining $S_{L}(u,v)=\tilde{S}_{L}(u,v) a(u) a(v) d(u) d(v)$. However, in this case the ansatz for $\tilde{S}_{L}(u,v)$ has to be generalized to $$\tilde{S}_{L}(u,v)= \sigma^{(2)}_2 (u,v) L^2 +\sigma^{(2)}_1 (u,v) L \ .$$ Note again that adding an extra term with no dependence in $L$ is forbidden by the consistency condition. Substituting this ansatz into (\[twomagnonansatz\]) fixes the first function to $$\sigma^{(2)}_2= f(u,v) f(v,u) \sigma^{(1)}_1 (u) \sigma^{(1)}_1 (v) \ ,$$ but it does not give us any information about the second one. We can reconstruct the function $\sigma^{(2)}_1$ using the same trick as before, i.e., computing the square $S (\{u,v\} \rightarrow \{w,v\} )$ and taking the limit $w\rightarrow u$. Nevertheless, we are going to postpone this calculation to section \[offshelltwoexcitations\], where we compute the off-shell square with two excitations, because such limit should be taken without imposing Bethe equations. (0,0) –(0,1); (0, 0) – (1,0); (0, 1) – (1,1); (1,0) –(1,1); (0.25,0) circle \[radius=0.04\]; (0.25,1) circle \[radius=0.04\]; (0.25,0) – (0.25,1); (0.5,0) circle \[radius=0.04\]; (0.5,1) circle \[radius=0.04\]; (0.5,0) – (0.5,1); (0.75,0) circle \[radius=0.04\]; (0.75,1) circle \[radius=0.04\]; (0.75,0) – (0.75,1); (1.3,0.4) – (1.6,0.4); (1.3,0.6) – (1.6,0.6); (1.9,0) –(1.9,1); (1.9, 0) – (2.7,0); (1.9, 1) – (2.7,1); (2.7,0) –(2.7,1); (2.1,0) circle \[radius=0.04\]; (2.1,1) circle \[radius=0.04\]; (2.1,0) – (2.1,1); (2.3,0) circle \[radius=0.04\]; (2.3,1) circle \[radius=0.04\]; (2.3,0) – (2.3,1); (2.5,0) circle \[radius=0.04\]; (2.5,1) circle \[radius=0.04\]; (2.5,0) – (2.5,1); (2.9,0) –(2.9,1); (2.9, 0) – (3.7,0); (2.9, 1) – (3.7,1); (3.7,0) –(3.7,1); (4,0.5) – (4.3,0.5); (4.15,0.35) – (4.15,0.65); (4.6,0) –(4.6,1); (4.6, 0) – (5.4,0); (4.6, 1) – (5.4,1); (5.4,0) –(5.4,1); (5,0) circle \[radius=0.04\]; (5,1) circle \[radius=0.04\]; (5,0) – (5,1); (5.6,0) –(5.6,1); (5.6, 0) – (6.4,0); (5.6, 1) – (6.4,1); (6.4,0) –(6.4,1); (5.86666,0) circle \[radius=0.04\]; (5.86666,1) circle \[radius=0.04\]; (5.86666,0) – (5.86666,1); (6.13333,0) circle \[radius=0.04\]; (6.13333,1) circle \[radius=0.04\]; (6.13333,0) – (6.13333,1); The last particular case we are going to dissect before presenting the general reconstruction is the one involving three excitations. In this case we can express equation \[bootstrap\], schematically represented in fig. \[three\], as $$\begin{aligned} S_{L}(u,v,w) &=S_{l_1}(u,v,w) S_{l_2} (\emptyset) a_{l_2} (u) a_{l_2} (v) a_{l_2} (w) d_{l_2} (u) d_{l_2} (v) d_{l_2} (w) \notag \\ &+S_{l_1} (\emptyset) S_{l_2}(u,v,w) a_{l_1} (u) a_{l_1} (v) a_{l_1} (w) d_{l_1} (u) d_{l_1} (v) d_{l_1} (w) \notag \\ &+[S_{l_1} (u) S_{l_2} (v,w) f(u,v) f(u,w) f(v,u) f(w,u) a_{l_1} (v) d_{l_1} (v) a_{l_1} (w) d_{l_1} (w) a_{l_2} (u) d_{l_2} (u) \notag \\ &+S_{l_1} (v,w) S_{l_2} (u) f(u,v) f(u,w) f(v,u) f(w,u) a_{l_1} (u) d_{l_1} (u) a_{l_2} (v) d_{l_2} (v) a_{l_2} (w) d_{l_2} (w)] \notag \\ &+[u\leftrightarrow v] + [u \leftrightarrow w] \ ,\end{aligned}$$ where the last two permutations are taken only over the terms in brackets. To solve this consistency condition we need to substitute a polynomial of degree three in $L$ as ansatz. The reason behind it is that the equation involves the product of a square with one excitation (linear in $L$) and a square with two excitations (quadratic in $L$) $$S_{L}(u,v)=\left( \sigma^{(3)}_3 (u,v) L^3 + \sigma^{(3)}_2 (u,v) L^2 +\sigma^{(3)}_1 (u,v) L \right) a(u) a(v) a(w) d(u) d(v) d(w) \ .$$ Similarly to the case of two excitations, the consistency condition fixes $\sigma^{(3)}_3$ and $\sigma^{(3)}_2$ but the function $\sigma^{(3)}_1$ remains unconstrained. In particular $$\begin{aligned} \sigma^{(3)}_3 &= f(u,v)f(u,w) f(v,w) f(w,v) f(w,u) f(v,u) \sigma^{(1)}_1 (u) \sigma^{(1)}_1 (v) \sigma^{(1)}_1 (w) \ , \\ \sigma^{(3)}_2 &= f(u,w) f(v,w) f(w,v) f(w,u)\sigma^{(2)}_1 (u,v) \sigma^{(1)}_1 (w) +[u\leftrightarrow w] + [v \leftrightarrow w] \ .\end{aligned}$$ This structure is very reminiscent of the Gaudin determinant rewritten using the Kirchhoff’s matrix-tree theorem if we understand the $\sigma^{(N)}_1$ as functions associated to spanning trees of the complete graph with $N$ nodes. The relation between our results and the forest expansion will be clarified the following subsection. Reconstructing the Gaudin determinant: the general asymptotic square -------------------------------------------------------------------- With the experience we have gathered from the previous particular cases, we are in a position to derive the full Gaudin determinant for the on-shell Bethe state with rapidities $\u=\{u_1 , \dots ,u_N\}$. The ansatz compatible with the consistency equation for the general case is $$S_{L} (\u)=a(\u) d(\u) \sum_{n=1}^{N} \sigma_n^{(N)} (\u) L^n \ .$$ After substituting this ansatz into equation \[bootstrap\], the resulting constrains can be solved by $$\sigma^{(N)}_i= f^{\neq} (\u , \u ) \sum_{\substack{\alpha_1 \cup \alpha_2 \cup \dots \cup \alpha_i =\u \\ \alpha_j \cap \alpha_k = \emptyset \\ \alpha_j \neq \emptyset}} \prod_{j=1}^i \frac{\sigma^{(|\alpha_j|)}_1 (\{\alpha_j \})}{f^{\neq} (\{\alpha_j\} , \{\alpha_j\})} \ .$$ To clarify the formula, the explicit expressions for the two highest orders in length are $$\begin{aligned} \sigma^{(N)}_N &=f^{\neq} (\u , \u ) \prod_{u_i \in \u} \sigma^{(1)}_1 (u_i) \ , \\ \sigma^{(N)}_{N-1} &= \sigma^{(N)}_N \sum_{\substack{u_i,u_j\in \u \\ u_i\neq u_j}}\frac{\sigma^{(2)}_1 (u_i,u_j) }{f(u_i,u_j) f(u_j,u_i) \sigma^{(1)}_1 (u_i) \sigma^{(1)}_1 (u_j)} \ .\end{aligned}$$ As we said in the previous subsection, we can interpret this formula as the tree expansion of the Gaudin determinant. However, to be able to make this connection we still have to make another assumption. At this point, the $\sigma^{(N)}_i$ functions are the expressions associated to dividing the complete vertex-labelled graph with $N$ vertices into $i$ disconnected objects, but the individual $\sigma^{(|\alpha_i |)}_1$ still cannot be interpreted as a tree. In order to proceed, we are going to assume that the $\sigma^{(n)}_1$ functions inherit certain factorization properties from the factorized scattering of the theory. In particular, we will assume that they can be written as the sum of the product of $n-1$ factors of $\sigma^{(2)}_1$ over non-repeated pairings of the $n$ arguments. This implies that we only have to fix the functions $\sigma^{(1)}_1$ and $\sigma^{(2)}_1$ to completely fix all squares, a task achievable just by using the ABA computations of the norm for one and two excitations as initial conditions. Even more, in the following subsection we will see that we can compute $\sigma^{(2)}_1$ just from the consistency condition, so we actually only need $\sigma^{(1)}_1$. A second consequence is that the sum over all possible pairings of the arguments can be understood as the sum over all possible spanning trees of the complete $n$-graph. The reason for that lies in the fact that the products involve $n-1$ non-repeated pairs, which is equivalent to the definition of a tree, can be mapped to the tree graph constructed with each vertex and each edge representing respectively a rapidity variable and a function $\frac{\sigma^{(2)}_1 (u,v)}{\sigma^{(1)}_1 (u) +\sigma^{(1)}_1 (v)}$ whose arguments are the vertices connected by the edge. Thus, allowing us to finally connect our formula with a forest expansion. We have chosen to divide $\sigma^{(2)}_1$ by the sum of the derivative of the momenta because it appears in the square with two magnons accompanied by a factor of the length, so it should contain a factor of the momentum and it should appear in a symmetric way. In the next section we will see that $\sigma^{(2)}_1 (u,v)\propto [p'(u) + p'(v)] \partial_{u} \log S(u,v)$. We can check from the expression (\[treegaudin\]) that our naive construction for the $\sigma^{(n)}_1$ just lacks a factor of the sum of the momenta of all excitations. [^4] We still have to reconstruct the dependence of the two-point function with the spacetime. We discuss two different approaches to it in appendix \[spacetimedependence\]. Computing the off-shell square ============================== After being able to reconstruct the Gaudin norm using the consistency condition \[bootstrap\], in this section we are going to include all the possible asymptotic terms appearing on the equation to reconstruct the norm of off-shell Bethe states. We will detail the computations for the cases of two and three excitations. We will also compute the first non-Gaudin term appearing in the square with a general number of excitations. The case of the square with one excitation require no special treatment as no new terms appear in the consistency equation, hence it remains unchanged $$S_{L} (u\rightarrow v) =g(u,v) \left[ a(u) d(v) - a(v) d(u) \right] \ , \qquad S_{L} (u)= \text{const.} (-i p' L) e^{ipL} \ .$$ The off-shell square with two excitations {#offshelltwoexcitations} ----------------------------------------- (0,0) –(0,1); (0, 0) – (1,0); (0, 1) – (1,1); (1,0) –(1,1); (0.33333,0) circle \[radius=0.04\]; (0.33333,1) circle \[radius=0.04\]; (0.33333,0) – (0.333333,1); (0.66666,0) circle \[radius=0.04\]; (0.66666,1) circle \[radius=0.04\]; (0.66666,0) – (0.666666,1); (1.3,0.4) – (1.6,0.4); (1.3,0.6) – (1.6,0.6); (1.9,0) –(1.9,1); (1.9, 0) – (2.7,0); (1.9, 1) – (2.7,1); (2.7,0) –(2.7,1); (2.16666,0) circle \[radius=0.04\]; (2.16666,1) circle \[radius=0.04\]; (2.16666,0) – (2.16666,1); (2.43333,0) circle \[radius=0.04\]; (2.43333,1) circle \[radius=0.04\]; (2.43333,0) – (2.43333,1); (2.9,0) –(2.9,1); (2.9, 0) – (3.7,0); (2.9, 1) – (3.7,1); (3.7,0) –(3.7,1); (4,0.5) – (4.3,0.5); (4.15,0.35) – (4.15,0.65); (4.6,0) –(4.6,1); (4.6, 0) – (5.4,0); (4.6, 1) – (5.4,1); (5.4,0) –(5.4,1); (5.6,0) –(5.6,1); (5.6, 0) – (6.4,0); (5.6, 1) – (6.4,1); (6.4,0) –(6.4,1); (5.86666,0) circle \[radius=0.04\]; (5.86666,1) circle \[radius=0.04\]; (5.86666,0) – (5.86666,1); (6.13333,0) circle \[radius=0.04\]; (6.13333,1) circle \[radius=0.04\]; (6.13333,0) – (6.13333,1); (1.35,0.5-1.5) – (1.65,0.5-1.5); (1.5,0.35-1.5) – (1.5,0.65-1.5); (1.9,0-1.5) –(1.9,1-1.5); (1.9, 0-1.5) – (2.7,0-1.5); (1.9, 1-1.5) – (2.7,1-1.5); (2.7,0-1.5) –(2.7,1-1.5); (2.3,0-1.5) circle \[radius=0.04\]; (2.3,1-1.5) circle \[radius=0.04\]; (2.3,0-1.5) – (2.3,1-1.5); (2.9,0-1.5) –(2.9,1-1.5); (2.9, 0-1.5) – (3.7,0-1.5); (2.9, 1-1.5) – (3.7,1-1.5); (3.7,0-1.5) –(3.7,1-1.5); (3.3,0-1.5) circle \[radius=0.04\]; (3.3,1-1.5) circle \[radius=0.04\]; (3.3,0-1.5) – (3.3,1-1.5); (4,0.5-1.5) – (4.3,0.5-1.5); (4.15,0.35-1.5) – (4.15,0.65-1.5); (4.6,0-1.5) –(4.6,1-1.5); (4.6, 0-1.5) – (5.4,0-1.5); (4.6, 1-1.5) – (5.4,1-1.5); (5.4,0-1.5) –(5.4,1-1.5); (5,0-1.5) circle \[radius=0.04\]; (5,1-1.5) circle \[radius=0.04\]; (5,0-1.5) – (5,1-1.5); (5.6,0-1.5) –(5.6,1-1.5); (5.6, 0-1.5) – (6.4,0-1.5); (5.6, 1-1.5) – (6.4,1-1.5); (6.4,0-1.5) –(6.4,1-1.5); (6,0-1.5) circle \[radius=0.04\]; (6,1-1.5) circle \[radius=0.04\]; (6,0-1.5) – (6,1-1.5); (1.35,0.5-3) – (1.65,0.5-3); (1.5,0.35-3) – (1.5,0.65-3); (1.9,0-3) –(1.9,1-3); (1.9, 0-3) – (2.7,0-3); (1.9, 1-3) – (2.7,1-3); (2.7,0-3) –(2.7,1-3); (2.3,0-3) circle \[radius=0.04\]; (2.3,1-3) circle \[radius=0.04\]; (2.3,0-3) – (2.3,1-3); (2.9,0-3) –(2.9,1-3); (2.9, 0-3) – (3.7,0-3); (2.9, 1-3) – (3.7,1-3); (3.7,0-3) –(3.7,1-3); (3.3,0-3) circle \[radius=0.04\]; (3.3,1-3) circle \[radius=0.04\]; (3.3,0-3) – (3.3,1-3); (4,0.5-3) – (4.3,0.5-3); (4.15,0.35-3) – (4.15,0.65-3); (4.6,0-3) –(4.6,1-3); (4.6, 0-3) – (5.4,0-3); (4.6, 1-3) – (5.4,1-3); (5.4,0-3) –(5.4,1-3); (5,0-3) circle \[radius=0.04\]; (5,1-3) circle \[radius=0.04\]; (5,0-3) – (5,1-3); (5.6,0-3) –(5.6,1-3); (5.6, 0-3) – (6.4,0-3); (5.6, 1-3) – (6.4,1-3); (6.4,0-3) –(6.4,1-3); (6,0-3) circle \[radius=0.04\]; (6,1-3) circle \[radius=0.04\]; (6,0-3) – (6,1-3); The consistency condition for the case of two excitations, pictorially represented in fig. \[twooff\], needs to include the two extra terms involving “no norm-like” one-excitation squares we neglected in the previous section, being its complete expression $$\begin{aligned} S_L (u,v) &= S_{l_1} (u,v) a_{l_2} (u) a_{l_2} (v) d_{l_2} (u) d_{l_2} (v)+ S_{l_2} (u,v) a_{l_1} (u) a_{l_1} (v) d_{l_1} (u) d_{l_1} (v) + \notag \\ &+S_{l_1} (u) S_{l_2} (v) f(u,v) f(v,u) a_{l_1} (v) d_{l_1} (v) a_{l_2} (u) d_{l_2} (u) \notag \\ &+S_{l_1} (v) S_{l_2} (u) f(u,v) f(v,u) a_{l_1} (u) d_{l_1} (u) a_{l_2} (v) d_{l_2} (v) \notag \\ &+S_{l_1} (v\rightarrow u) S_{l_2} (u\rightarrow v) f(v,u)^2 a_{L} (u) d_{L} (v) \notag \\ &+S_{l_1} (u\rightarrow v) S_{l_2} (v\rightarrow u) f(u,v)^2 a_{L} (v) d_{L} (u) \ .\end{aligned}$$ As the Gaudin norm solves this equation except for the last two terms, we are going to look for a solution of the form $$S_L (u,v) = \tilde{S}^{(\text{Gaudin})}_L (u,v) a(u) a(v) d(u) d(v)+\text{extra terms} \ .$$ It is important to remark that these new terms do not even have to have the same momentum structure as the on-shell terms. Indeed $$\begin{gathered} S_{l_1} (u\rightarrow v) S_{l_2} (v\rightarrow u) a_{L} (u) d_{L} (v) = -g(u,v)^2 \Big[ a_{L}(v) d_{L}(u) +a_{L}(v) d_{L}(u) \\ - a_{l_1}(u) a_{l_2}(v) d_{l_2}(u) d_{l_1}(v) - a_{l_2}(u) a_{l_1}(v) d_{l_1}(u) d_{l_2}(v) \Big] a_{L} (v) d_{L} (u) \ ,\end{gathered}$$ cannot be compensated by such momentum structure. We can upgrade our ansatz to the following one $$\begin{aligned} S_L (u,v) &= \alpha_{u,v} [a_L (u) d_L (v) ]^ 2 + \alpha_{v,u} [a_L (v) d_L (u) ]^ 2 \notag \\ &+ \left[ \sigma^{(2)}_2 L^2 +\sigma^{(2)}_1 (u,v) L +\sigma^{(2)}_0 \right] a_L (u) a_L (v) d_L (u) d_L (v) \ ,\end{aligned}$$ where the subindexes of the function $\alpha$ indicates the structure of the momentum terms accompanying them. Thus, the consistency condition can be divided into seven different equations according to the different momentum factor accompanying each term [^5] $$\begin{aligned} \alpha_{u,v}^{(L)} &= f(v,u)^2 g(u,v)^2 \ , & \alpha_{u,v}^{(l_1)} &= f(u,v)^2 g(u,v)^2 \ , & \alpha_{u,v}^{(l_2)} &= f(u,v)^2 g(u,v)^2 \ , \notag \\ \alpha_{v,u}^{(L)} &= f(u,v)^2 g(v,u)^2 \ , & \alpha_{v,u}^{(l_1)} &= f(v,u)^2 g(v,u)^2 \ , & \alpha_{v,u}^{(l_2)} &= f(v,u)^2 g(v,u)^2 \ ,\end{aligned}$$ $$\begin{aligned} &\sigma^{(2)}_2 L^2 +\sigma^{(2)}_1 L +\sigma^{(2)}_0=\sigma^{(2)}_2 l_1^2 +\sigma^{(2)}_0+\sigma^{(2)}_2 l_2^2 \sigma^{(2)}_1 l_1+\sigma^{(2)}_1 l_2+\sigma^{(2)}_0 \\ &+2l_1 l_2 \, \sigma_1^{(1)} (u) \sigma_1^{(1)} (v) f(u,v) f(v,u)+ g(u,v)^2 [f(u,v)^2 + f(v,u)^2] \ .\notag \end{aligned}$$ In contrast with the Gaudin square, here the term $\sigma_0$ is not only allowed, but needed and completely fixed. These equations are solved by $$\begin{aligned} \alpha_{u,v} &=f(v,u)^2 g(u,v)^2 \ , & \sigma^{(2)}_2=f(u,v) f(v,u) \sigma_1^{(1)} (u) \sigma_1^{(1)} (v) \ , \notag \\ \alpha_{v,u} &=f(u,v)^2 g(u,v)^2 \ , & \sigma^{(2)}_0= -g(u,v)^2 [f(u,v)^2 +f(v,u)^2] \ . \label{sigmatwomagnonoff}\end{aligned}$$ Again the function $\sigma^{(2)}_1 (u,v)$ is not fixed by this procedure. Working with off-shell states we can constrain the function $\sigma^{(2)}_1 (u,v)$ in a similar way as we constrained the function $\sigma^{(1)}_1 (u)$ in the previous section, that is, by computing the square $S_L (\{ u,v \} \rightarrow \{ w,v \} )$ and taking the limit $w\rightarrow v$. Although we only have to compute the terms proportional to $p' (v)$ to extract this function, we will compute the full square as it will appear in the consistency equation for a square with three excitations. The consistency condition for this square is $$\begin{aligned} &S_L (\{ u,v \} \rightarrow \{ w,v \} ) = S_{l_1} (\{ u,v \} \rightarrow \{ w,v \}) a_{l_2} (w) a_{l_2} (v) d_{l_2} (u) d_{l_2} (v) + S_{l_2} (\{ u,v \} \rightarrow \{ w,v \} ) \notag \\ & \cdot a_{l_1} (u) a_{l_1} (v) d_{l_1} (w) d_{l_1} (v)+S_{l_1} (u\rightarrow w) S_{l_2} (v) a_{l_2} (w) a_{l_1} (v) d_{l_2} (u) d_{l_1} (v) f(u,v) f(v,w) \notag \\ &+S_{l_1} (v) S_{l_2} (u\rightarrow w) a_{l_1} (u) a_{l_2} (v) d_{l_1} (w) d_{l_2} (v) f(w,v) f(v,u) \notag \\ &+S_{l_1} (u\rightarrow v) S_{l_2} (v\rightarrow w) a_{l_1} (v) a_{l_2} (v) d_{l_2} (u) d_{l_1} (w) f(u,v) f(w,v) \notag \\ &+S_{l_1} (v\rightarrow w) S_{l_2} (u\rightarrow v) a_{l_1} (u) a_{l_2} (w) d_{l_1} (v) d_{l_2} (v) f(v,u) f(v,w) \ .\end{aligned}$$ Here we have to choose the following ansatz $$\begin{aligned} S_L (\{ u,v \} \rightarrow \{ w,v \} ) &=\beta_u (v) a (u) a (v) d(v) d(w) +\beta_w (v) a (w) a(v) d(v) d(u) \notag \\ &+ \alpha_a a (u) a(w) [d(v)]^2 +\alpha_d [a (v)]^2 d(u) d(w) \ .\end{aligned}$$ Substituted into the consistency condition it gives us nine equations, which can be reduced to only five equations and the condition that $\alpha_a$ and $\alpha_d$ are independent of the length. These equations are solved by $$\begin{aligned} \beta_u (v) &=g(u , w) f(w,v ) f(v, u) \sigma^{(1)}_1 (v) L +g(u,v) g(v,w)- \left[ g(u,v) g(v,w) \right]^2 +k(u,v,w) \ , \notag \\ \beta_w (v) &=g(w , u) f(u, v) f(v,w) \sigma^{(1)}_1 (v) L +g(u,v) g(v,w)- \left[ g(u,v) g(v,w) \right]^2 -k(u,v,w) \ , \notag \\ \alpha_a &= -g(u,v) g(v,w) f(v,u) f(v,w) \ , \notag \\ \alpha_d &= -g(u,v) g(v,w) f(u,v) f(w,v) \ \label{beta2magnonoff}.\end{aligned}$$ Note that there is a length-independent contribution to the functions $\beta$, which we label here as $k(u,v,w)$, that cannot be fixed from the consistency condition because these relations only involves the combination $\beta_{u,l_1} (v) + \beta_{w,l_2} (v)$. We may think that going one step further and computing $S_L(\{u_1 ,u_2 \} \rightarrow \{v_1 , v_2\} )$ would give us an equation that fixes this unknown function, but only an equivalent combination appears and the function cannot be fixed. Part of $\sigma^{(2)}_1$ is extracted from this result by performing the limit $w\rightarrow u$ of the functions $\beta$ with their corresponding momentum factor, which needs some care as the $g(u,w)$ factors will transform into derivatives. Focusing only in these factors, divided by $\sigma^{(1)}_1 (v) L \, a(v) d(v)$, we have $$\begin{gathered} \lim_{w\rightarrow u}g(u,w) \Big[ f(w,v) f(v,u) a (u) d(w) - f(u,v) f(v,w) a(w) d(u) \Big]= \\ =\left( \frac{\partial f(u,v)}{\partial u} d(u) + f(u,v) \frac{\partial d(u)}{\partial u} \right) f(v,u) a(u) - \left( \frac{\partial f(v,u)}{\partial u} a(u) + f(v,u) \frac{\partial a(u)}{\partial u} \right) f(u,v) d(u) \ ,\end{gathered}$$ here the first term of each parenthesis together will generate half of $\sigma^{(2)}_1$ while the second ones generate $\sigma^{(2)}_2$. We can rewrite the two terms contributing to $\sigma^{(2)}_1$ as a single term $$\sigma^{(2)}_1 \propto [f(v,u)]^2 \frac{\partial \left( \frac{f(u,v)}{f(v,u)}\right)}{\partial u} \ ,$$ which highlights that it is just the derivative of the S-matrix, as expected from the known expression of the Gaudin norm. The other half of $\sigma^{(2)}_1$ comes from the same limit applied to the $\alpha$ terms, giving the same result but dressed with $\sigma^{(1)}_1 (u)$ instead. To end the subsection, we want to show that the non-Gaudin terms contain the two-excitation Bethe equations. This can be made explicit if we perform the rewriting $$\begin{aligned} &\left[ a(u) d(v) f(v,u) g(u,v) \right] ^2 +\left[ a(v) d(u) f(u,v) g(v,u) \right] ^2 +a(u) a (v) d(u) d(v) g(u,v) g(v,u) \notag \\ &\cdot [f(u,v)^2 +f(v,u)^2]= a(u) a (v) d(u) d(v) [g(u,v)]^2 f(u,v) f(v,u) \left[ \frac{a(u)}{d(u)} \, \frac{d(v)}{a(v)} \, \frac{f(v,u)}{f(u,v)} \right. \notag \\ &\left.- \frac{f(u,v)}{f(v,u)} -\frac{f(v,u)}{f(u,v)} +\frac{d(u)}{a(u)} \, \frac{a(v)}{d(v)} \, \frac{f(u,v)}{f(v,u)} \right] \ .\end{aligned}$$ We can see that, apart from the vanishing inherited from the one-magnon square, these terms also vanish when the two-mangon Bethe equations (\[BAE\]) $$\frac{a(u)}{d(u)}=\frac{d(v)}{a(v)}=\frac{f(u,v)}{f(v,u)} \ ,$$ are fulfilled, proving that the reconstruction is consistent with the Bethe Ansatz. The off-shell square with three excitations ------------------------------------------- The computation of the square for the case of three excitations follows the same pattern as the previous case, but the large number of different combinations of momenta obscures the consistency condition. However, our experience from isolating the different momentum contributions in the two-excitation case had shown us that most of the equations we get are redundant. If we substitute the ansatz $$\begin{aligned} &S_L (u,v,w) = \alpha_{u,v} [a_l (u) d_L (v) ]^ 2 a_L (w) d_L (w) + \alpha_{v,u} [a_L (v) d_L (v) ]^ 2 a_L (w) d_L (w)+ (w\leftrightarrow u) \notag \\ &+ (w\leftrightarrow v)+ \left[ \sigma^{(3)}_3 L^3 + \sigma^{(3)}_2 L^2 +\sigma^{(3)}_1 L +\sigma^{(3)}_0 \right] a_L (u) a_L (v) a_L (w) d_L (u) d_L (v) d_L (w) \ ,\end{aligned}$$ into the consistency condition, only three sets of equations are relevant while the others can be extracted from symmetry under exchange of rapidities or independence of the length. One set of equations we have to study is the set associated to the term $a(\u) d(\u)$, which codifies most of the length-dependent structure of the square. It is also important to study these equations because of the presence in the consistency condition of terms with the structure $\sigma^{(1)}_1 \sigma^{(2)}_0 L$, which were absent in the consistency condition for “Gaudin squares” and would make the equation inconsistent. Indeed, since the equation for $\sigma^{(3)}_1$ takes the form $$\sigma^{(3)}_1 L=\sigma^{(3)}_1 l_1 + \sigma^{(3)}_1 l_2 +h(u,v,w) L$$ a solution does not exist unless the function $h(u,v,w)$ vanishes. [^6] The extra contribution in this equation is given by $$\begin{aligned} &f(u,w) f(v,u) [f(v,w)]^2 g(v,w) \beta_v (u) +f(u,v) f(w,u) [f(w,v)]^2 g(w,v) \beta_w (u) \notag \\ &+f(u,v) f(u,w) f(v,u) f(w,u) \sigma^{(1)}_1 (u) \sigma^{(2)}_0 (v,w) + (u\leftrightarrow v) + (u\leftrightarrow w) \ ,\end{aligned}$$ which we can be shown to vanish when the explicit expression for $\beta$ and $\sigma$ from equations (\[sigmatwomagnonoff\]) and (\[beta2magnonoff\]) are substituted, as requires by consistency. Thus the equation we get just shows us that the function $\sigma_1^{(3)}$ is not fixed by the consistency condition. We should remark two additional points regarding this set of equations. First, the orders $L^3$ and $L^2$ receive no contributions from the no norm-like squares, hence $\sigma^{(3)}_3$ and $\sigma^{(3)}_2$ are not modified with respect to the Gaudin case. And second, as the equation that fixes $\sigma^{(3)}_0$ involves the function $k(u,v,w)$, appearing from the length-independent structure of the functions $\beta$, we need a larger set of inputs to completely fix the general square than for the Gaudin case. The other equations we need to solve are the ones obtained from terms involving the momentum factors $a(\u) d(\u) \frac{a(u)}{d(u)} \frac{d(v)}{a(v)}$ and $a(\u) d(\u) \frac{d(u)}{a(u)} \frac{a(v)}{d(v)}$. Any other combination of rapidities is obtained by means of the symmetry inherited from the exchange symmetry of the $B$ and $C$ operators (\[commBB\]). There equations are $$\begin{aligned} \alpha_{u,v} &= g(u,v) [f(v,u)]^2 f(w,u) f(v,w) \beta_u (w) \ , & \alpha_{v,u} &=g(u,v) [f(u,v)]^2 f(u,w) f(w,v) \beta_v (w) \ .\end{aligned}$$ Substituting the expression for the functions $\beta$ from equation (\[beta2magnonoff\]) we get $$\begin{aligned} \alpha_{u,v} &= [g(u,v) f(v,u) f(w,u) f(v,w)]^2 \sigma^{(1)}_1 (w) L - g(u,w) g(w,v) g(v,u) [f(v,u)]^2 f(w,u) f(v,w) \notag \\ &+[g(u,w) g(w,v)]^2 g(v,u) [f(v,u)]^2 f(w,u) f(v,w)+ g(u,v) [f(v,u)]^2 f(w,u) f(v,w) k(u,w,v) \ , \notag \\ \alpha_{v,u} &=-[g(u,v) f(u,v) f(u,w) f(w,v)]^2 \sigma^{(1)}_1 (w) +g(v,w) g(w,u) g(u,v) [f(u,v)]^2 f(u,w) f(w,v) \notag \\ &-[g(v,w) g(w,u)]^2 g(u,v) [f(u,v)]^2 f(u,w) f(w,v)-g(u,v) [f(u,v)]^2 f(u,w) f(w,v) k(u,w,v) \ ,\end{aligned}$$ which matches the norm computed using the commutation relations for the $B$ and $C$ operators at order $L$. The general off-shell norm-like square -------------------------------------- As we have seen in the previous case, the difficulty in computing the off-shell square and the normalization of off-shell Bethe states increases heavily with respect to the on-shell case. We are not able to give a full general expression for the off-shell square in this article, but we can at least give the exact expression for the first no norm-like contribution, which appears in the third highest order term in length. In order to do so it is more convenient to rewrite equation \[bootstrap\] as $$\begin{gathered} S_L (\{ u \} )= \sum_{\alpha \cup \beta \cup \gamma \cup \delta =\{u\}} w \Big( \{\alpha \cup \gamma\}, \{\beta \cup \delta \} , \{\alpha \cup \delta\} , \{ \beta \cup \gamma \} \Big) \\ \cdot S_{l_1} \Big( \{\alpha \cup \gamma\} \rightarrow \{ \alpha \cup \delta\} \Big) S_{l_2} \Big( \{\beta \cup \delta\} \rightarrow \{ \beta \cup \gamma\} \Big) \ .\end{gathered}$$ A factor of $L$ appears for each pair of upper- and lower-edge excitations with the same rapidity, hence in the large length limit we have $S_{l_1} \Big( \{\alpha \cup \gamma\} \rightarrow \{ \alpha \cup \delta\} \Big) \propto l_1^{|\{\alpha\}|}$ and similarly for $S_{l_2}$ and $\{\beta\}$. The terms in the consistency condition involving the highest powers of the length come from the terms of the sum with coinciding outgoing and incoming rapidities, i.e. from the terms with $\delta=\gamma=\emptyset$. As a consequence, to get the highest and second highest contributions in length we only have to take into account norm-like squares, and thus these contributions coincide with the corresponding terms appearing in Gaudin norm. The third and fourth orders get extra contributions apart from the ones coming from the Gaudin norm. In particular, these extra contributions come from the terms in the consistency condition in which both squares have one excitation different in the upper edge with respect to the lower edge ($|\delta|=|\gamma|=1$, so its highest power has two powers of $L$ less than the highest power of the Gaudin contribution, one from each square). Nevertheless, since the consistency condition involves the expression for the leading order in length of the square with one excitation different in the outgoing state with respect to the ingoing one, $S_{L} (\{u_i \cup \alpha\}\rightarrow \{u_j \cup \alpha\})$, we have to find its expression first. The correct ansatz in this case is $$S_L \left( \{u\} \rightarrow \left\{v \cup \bar{u}_1 \right\} \right)= a(\{u\} ) d(\{u\} ) \left[ \beta_{1,v} (\{ \bar{u}_1 \}) \frac{d(v)}{d(u_1)} + \beta_{v,1} (\{ \bar{u}_1 \}) \frac{a(v)}{a(u_1)} +\mathcal{O} (L^{N-1}) \right] \ ,$$ where $\{ \bar{u}_1 \}$ stands for the set $\{u\}$ without the element $u_1$. We have already computed this square for two excitations and we have seen that it is proportional to the Gaudin norm of the excitation that is present in both physical sides. Inspired by this, we are going to propose that the functions $\beta_{1,v} (\{ \bar{u}_1 \})$ and $\beta_{v,1} (\bar{u}_1)$ are proportional to $\tilde{S}^{(\text{Gaudin})} (\{ \bar{u}_1 \})$ and check it by induction. The leading order in length of the equation in the consistency condition associated to the ratio $\frac{a(v)}{a(u_1)}$ is $$\begin{gathered} \beta_{v,1}^{(L)} (\bar{u}_1) =\beta_{v,1}^{(l_1)} (\bar{u}_1) +g(v,u_1) f(u_1 ,\bar{u}_1) f(\bar{u}_1 , v) \tilde{S}_{l_2} (\bar{u}_1) \\ +\sum_{\substack{\alpha \cup \beta =\bar{u}_1 \\ \alpha \neq \emptyset}} f(u_1 , \beta) f(\beta , v) f(\alpha ,\beta ) f(\beta , \alpha) \beta^{l_1}_{v,1} (\alpha) \tilde{S}_{l_2} (\beta) \ .\end{gathered}$$ Thus, we can check that assuming the following structure for $|\alpha|<|\bar{u}_1|$ $$\begin{aligned} \beta^{(L)}_{1,v} (\alpha)&=f(\{ \alpha \}, u_1) f(v,\{ \alpha \}) g(u_1,v) \tilde{S}^{(Gaudin)}_L (\{ \alpha \}) \ , \notag \\ \beta^{(L)}_{v,1} (\alpha)&=-f(\{ \alpha \}, v) f(u_1,\{ \alpha \}) g(u_1,v) \tilde{S}^{(Gaudin)}_L (\{ \alpha \}) \ .\end{aligned}$$ gives an equivalent structure for $\beta^{(L)}_{1,v} (\bar{u}_1)$ and $\beta^{(L)}_{v,1} (\bar{u}_1)$, as the consistency condition can be recast into the consistency condition of the Gaudin square. Now we can compute the first non-Gaudin contribution to third order in length of the off-shell square. The correction is generated from the terms $S_{l_1} (\{u_i \cup \alpha\}\rightarrow \{u_j \cup \alpha\}\} \cdot S_{l_2} (\{u_j \cup \beta\}\rightarrow \{u_i \cup \beta\}\}$ and the one obtained by exchanging $l_1$ and $l_2$, which can be divided into the following three momentum contributions $$\begin{aligned} &a(u_i) d(u_j) [f(u_i,u_j)]^2 f(u_i ,\bar{u}_{ij}) f(\bar{u}_{ij} , u_j) f(\alpha , \beta) \left[ a(u_j) d(u_i) \beta^{(l_1)}_{j,i} (\alpha) \beta^{(l_2)}_{j,i} (\beta) \right. \notag \\ &+\left. a(u_i) d(u_j) \beta^{(l_1)}_{i,j} (\alpha) \beta^{(l_2)}_{i,j} (\beta) \right] +a(u_j) d(u_i) [f(u_j,u_i)]^2 f(u_j ,\bar{u}_{ij}) f(\bar{u}_{ij} , u_i) \notag \\ &\cdot \left[ a(u_i) d(u_j) \beta^{(l_1)}_{i,j} (\beta) \beta^{(l_2)}_{i,j} (\alpha) +a(u_j) d(u_j) \beta^{(l_1)}_{j,i} (\beta) \beta^{(l_2)}_{j,i} (\alpha)\right] \ .\end{aligned}$$ For each different contribution we can reconstruct the consistency condition for $\tilde{S}^{(\text{Gaudin})}$, allowing us to write $$\begin{aligned} S_L(\{u\}) &= S^{(\text{Gaudin})}_L(\{u\}) + \sum_{i\neq j} \left\{ \vphantom{\frac{a(u_i) d(u_j)}{a(u_j) d(u_i)} } \left[ g(u_i,u_j) f(u_j , u_i) f(\bar{u}_{ij} , u_j) f(u_i,\bar{u}_{ij}) \right] ^2 \right. \notag \\ &+\left[ g(u_i,u_j) f(u_i , u_j) f(\bar{u}_{ij} , u_i) f(u_j,\bar{u}_{ij}) \right] ^2 + \left[ \frac{a(u_i) d(u_j)}{a(u_j) d(u_i)} \left[ f(u_i , u_j) g(u_i , u_j) \right]^2 \right. \notag \\ &\cdot \left. \left. f(u_i , \bar{u}_{ij} ) f(u_j , \bar{u}_{ij} ) f(\bar{u}_{ij}, u_i ) f(\bar{u}_{ij}, u_j ) + i\leftrightarrow j \vphantom{\frac{a(u_i) d(u_j)}{a(u_j) d(u_i)} }\right] \right\} a(\{u\}) d(\{u\}) \tilde{S}^{(\text{Gaudin})} (\bar{u}_{ij} ) \notag \\ &+\mathcal{O} \left( L^{|\{u\}|-4} \right) \ .\end{aligned}$$ Conclusions =========== In this article we have studied how we can apply a cutting operation akin to the one used in the hexagon proposal to a cylindrical worldsheet. As the cutting operator is well defined after being applied once, we can define a square “form factor” through a consistency condition coming from subsequent applications of the cutting procedure. We have proven that, applying this construction to an integrable $SU(2)$ model, the consistency condition retrieves the Gaudin determinant. We have also computed off-shell norm of Bethe states for the $SU(2)$ spin chain for the cases of a few excitations. There exists several possible generalization of the work presented here. The most immediate are the generalizations to the full $PSU(2,2|4)$ symmetry and to general values of the ’t Hooft coupling. There also exist two computations that provide relevant checks for the validity of the square and its consistency condition. The first computation we propose is to calculate higher loops and check the proposal for the “all-loops normalization" presented in eq. (17) of [@TailoringIV], which amount to replace the Heisenberg S-matrix by the all-loops S-matrix and a modification of the $f^{\neq} (\{ u \},\{ u\})$ prefactor. The results obtained here can easily be generalized to inhomogeneous spin chains, so we can use either the $\theta$-morphism [@TailoringIV] or the inhomogeneous version of the BDS spin chain [@InhomogeneousBDS] to check it. A second computation is to study the extremal limit of the hexagon form factor. In this limit one of the physical sides of the hexagon vanishes, as it has zero length, effectively transforming it into a square. [^7] These computations will be addressed in future works. (-1/2,1.73205/2) – (0,0); (0, 0) – (1,0); (1,0) – (3/2,1.73205/2); (3/2,1.73205/2) – (1,1.73205); (1,1.73205) – (0,1.73205); (0,1.73205) –(-1/2,1.73205/2); (2/3,0) circle \[radius=0.04\]; (4/3,4/6\*1.73205) circle \[radius=0.04\]; (2/3,0) – (4/3,4/6\*1.73205); (1.65,1.73205\*0.4) – (2.05,1.73205\*0.4); (1.65,1.73205\*0.6) – (2.05,1.73205\*0.6); (2.7-1/2,1.73205/2) – (2.7+0,0); (2.7+0, 0) – (2.7+1,0); (2.7+1,0) – (2.7+3/2,1.73205/2); (2.7+3/2,1.73205/2) – (2.7+1,1.73205); (2.7+1,1.73205) – (2.7+0,1.73205); (2.7+0,1.73205) –(2.7-1/2,1.73205/2); (2.7+1.1,-0.0866) – (2.7+1.6,1.73205/2-0.0866); (2.7+1.6,1.73205/2-0.0866) – (2.7+1.6+1.73205/2,1.73205/2-0.5866); (2.7+1.6+1.73205/2,1.73205/2-0.5866) – (2.7+1.1+1.73205/2,-0.5866); (2.7+1.1+1.73205/2,-0.5866) – (2.7+1.1,-0.0866); (2.7+2/3,0) circle \[radius=0.04\]; (2.7+4/3,4/6\*1.73205) circle \[radius=0.04\]; (2.7+2/3,0) – (2.7+4/3,4/6\*1.73205); (5.166,1.73205\*0.5) – (5.566,1.73205\*0.5); (5.366,1.73205\*0.5-0.2) – (5.366,1.73205\*0.5+0.2); (6.2-1/2,1.73205/2) – (6.2+0,0); (6.2+0, 0) – (6.2+1,0); (6.2+1,0) – (6.2+3/2,1.73205/2); (6.2+3/2,1.73205/2) – (6.2+1,1.73205); (6.2+1,1.73205) – (6.2+0,1.73205); (6.2+0,1.73205) –(6.2-1/2,1.73205/2); (6.2+1.1,-0.0866) – (6.2+1.6,1.73205/2-0.0866); (6.2+1.6,1.73205/2-0.0866) – (6.2+1.6+1.73205/2,1.73205/2-0.5866); (6.2+1.6+1.73205/2,1.73205/2-0.5866) – (6.2+1.1+1.73205/2,-0.5866); (6.2+1.1+1.73205/2,-0.5866) – (6.2+1.1,-0.0866); (6.2+1.1+1.73205/4,-0.25-0.0866) circle \[radius=0.04\]; (6.2+1.6+1.73205/4,1.73205/2-0.25-0.0866) circle \[radius=0.04\]; (6.2+1.1+1.73205/4,-0.25-0.0866) – (6.2+1.6+1.73205/4,1.73205/2-0.25-0.0866); However, the most interesting line of work consists in integrating this square proposal with the hexagon proposal. In principle, we expect that the hexagon form factor follows a similar consistency condition to the one proposed here for the square, which in the asymptotic regime should be given by $$\begin{gathered} \mathcal{H} (\{ u \}, \{ v\} , \{ w \} )_{L_1,L_2,L_3}= \\ =\sum_{\substack{\alpha \cup \bar{\alpha}= \{ u \} \\ \beta \cup \bar{\beta}= \{ v \}}} w (\{ \alpha \} , \{ \bar{\alpha} \} ,\{ \beta \} , \{ \bar{\beta} \} ) \mathcal{H}_{l_1,l_2,L_3} ( \{ \alpha \} , \{ \beta \} , \{ w \} ) S_{l} ( \{ \bar{\alpha} \} \rightarrow \{ \bar{\beta} \} ) \ ,\end{gathered}$$ where $l_1+l=L_1$ and $l_2+l=L_2$, and similarly with other pairs of physical sides of the hexagon. Figure \[hexagontwo\] shows a pictorial representation of the consistency equation for the hexagon with two excitation. If we decide to include mirror contributions, the first non-trivial consistency condition we can write involves instead the hexagon with one excitation, as we can see in figure \[hexagonmirror\]. The problem regarding this consistency condition is that the square constructed here has been defined using a different operator algebra than the one used for the hexagon proposal. Here we have constructed the breaking factors from the commutation relations of the $B$ operator from the Algebraic Bethe Ansatz, while excitations in the hexagon proposal behaves according to a Zamolodchikov-Faddeev algebra. [^8] This implies that the weights associated to both constructions are different, as commented in appendix \[BFappendix\]. Despite that, we expect such consistency condition to hold when both form factors are written using the ZF algebra. (-1/2,1.73205/2) – (0,0); (0, 0) – (1,0); (1,0) – (3/2,1.73205/2); (3/2,1.73205/2) – (1,1.73205); (1,1.73205) – (0,1.73205); (0,1.73205) –(-1/2,1.73205/2); (2/3,0) circle \[radius=0.04\]; (1.65,1.73205\*0.4) – (2.05,1.73205\*0.4); (1.65,1.73205\*0.6) – (2.05,1.73205\*0.6); (2.7-1/2,1.73205/2) – (2.7+0,0); (2.7+0, 0) – (2.7+1,0); (2.7+1,0) – (2.7+3/2,1.73205/2); (2.7+3/2,1.73205/2) – (2.7+1,1.73205); (2.7+1,1.73205) – (2.7+0,1.73205); (2.7+0,1.73205) –(2.7-1/2,1.73205/2); (2.7+1.1,-0.0866) – (2.7+1.6,1.73205/2-0.0866); (2.7+1.6,1.73205/2-0.0866) – (2.7+1.6+1.73205/2,1.73205/2-0.5866); (2.7+1.6+1.73205/2,1.73205/2-0.5866) – (2.7+1.1+1.73205/2,-0.5866); (2.7+1.1+1.73205/2,-0.5866) – (2.7+1.1,-0.0866); (2.7+2/3,0) circle \[radius=0.04\]; (5.166,1.73205\*0.5) – (5.566,1.73205\*0.5); (5.366,1.73205\*0.5-0.2) – (5.366,1.73205\*0.5+0.2); (6.2-1/2,1.73205/2) – (6.2+0,0); (6.2+0, 0) – (6.2+1,0); (6.2+1,0) – (6.2+3/2,1.73205/2); (6.2+3/2,1.73205/2) – (6.2+1,1.73205); (6.2+1,1.73205) – (6.2+0,1.73205); (6.2+0,1.73205) –(6.2-1/2,1.73205/2); (6.2+1.1,-0.0866) – (6.2+1.6,1.73205/2-0.0866); (6.2+1.6,1.73205/2-0.0866) – (6.2+1.6+1.73205/2,1.73205/2-0.5866); (6.2+1.6+1.73205/2,1.73205/2-0.5866) – (6.2+1.1+1.73205/2,-0.5866); (6.2+1.1+1.73205/2,-0.5866) – (6.2+1.1,-0.0866); (6.2+1+1/4-0.04,1.73205/4-0.04) rectangle (6.2+1+1/4+0.04,1.73205/4+0.04); (6.2+1+1/4+0.1-0.04,1.73205/4-0.0866-0.04) rectangle (6.2+1+1/4+0.1+0.04,1.73205/4-0.0866+0.04); (6.2+1.1+1.73205/4,-0.25-0.0866) circle \[radius=0.04\]; (6.2+1+1/4+0.1,1.73205/4-0.0866) to \[out=-30,in=60\] (6.2+1.1+1.73205/4,-0.25-0.0866); Regarding the square described in [@HexagonWL], we expect it to fulfil a consistency condition similar to the one we found here. We denote it by $S^{(L)}_{KK}$ or $S^{(R)}_{KK}$, depending on which edge carries the contribution from the Wilson loop, to distinguish it from the square we are proposing. With this notation, we expect these two kinds of squares to satisfy the following consistency condition. $$\begin{aligned} S^{(L)}_{KK,L} (\u \rightarrow \{ v\})=\sum_{\substack{\alpha \cup \bar{\alpha} =\u \\ \beta \cup \bar{\beta} =\{ v \} }} w (\alpha , \bar{\alpha}, \beta, \bar{\beta}) S^{(L)}_{KK,l_1} (\alpha \rightarrow \beta ) S_{l_2} (\bar{\alpha} \rightarrow \bar{\beta} ) \ , \notag \\ S^{(R)}_{KK,L} (\u \rightarrow \{ v\})=\sum_{\substack{\alpha \cup \bar{\alpha} =\u \\ \beta \cup \bar{\beta} =\{ v \} }} w (\alpha , \bar{\alpha}, \beta, \bar{\beta}) S_{l_1} (\alpha \rightarrow \beta ) S^{(R)}_{KK,l_2} (\bar{\alpha} \rightarrow \bar{\beta} ) \ ,\end{aligned}$$ plus mirror corrections. In the same way that three squares $S_{KK}$ glued to the hexagon give us the three open string interaction in the semiclassical limit, we expect the gluing of $S^{(L)}_{KK}$ with $S^{(R)}_{KK}$ to be equivalent to the propagation of an open string and the gluing of $S_L$ with itself to be equivalent to the propagation of a closed string. Although only sketched here, the square could be used as a tool for understanding the gluing procedure. The appeal of the square over the hexagon and octagon form factors, for which some interesting results have already been obtained [@Hex2; @hexagon2; @octagon], is its simplicity. As sketched in appendix \[mirrorappendix\], the gluing of two squares or the gluing of a square and a hexagon form factor involve a finite amount of terms. Acknowledgements {#acknowledgements .unnumbered} ================ I would like to thank Yunfeng Jiang, Sergey Frolov and Shota Komatsu for interesting discussions and Ivan Kostov, Didina Serban, Thiago Fleury, Minkyoo Kim, Rafael Hernández and Roberto Ruiz for questions and comments on the manuscript. I thank the Institut de Physique Théorique (IPhT) for the hospitality during my visits. Kirchhoff’s matrix-tree theorem and Gaudin determinant {#Kirchhoffappendix} ====================================================== In this appendix we review the Kirchhoff’s matrix-tree theorem and the results of applying it to the Gaudin determinant. Given a graph $G$ with $n$ vertices, we can define three matrices that completely define it: the degree matrix, the adjacency matrix and the Laplacian matrix, being the last one the difference of the two previous one. The degree matrix is a diagonal matrix indicating the number of edges attaches to each vertex. The adjacency matrix is a square matrix such that its element $a_{ij}$ with $i\neq j$ is one if there exist an edge fromn vertex $i$ to vertex $j$ and zero if there is not while its elements $a_{ii}$ are equal to zero as we are not going to consider graphs with loops (edges that start and end in the same vertex). If there is a loop in vertex $i$, then we have to set $a_{ii}=2$ instead. To state the Kirchhoff’s theorem we have first to define what a spanning tree is. On the one hand, a graph $H$ with $n$ vertices is said to be a tree if there is a unique path between any two vertices. This condition is equivalent to $H$ being connected and having $n-1$ edges. The connectivity condition can be replaced with the condition of having no cycles without adding any further restriction. A directed tree is defined as a directed acyclic graph whose underlying undirected graph is a tree. A subclass of directed trees are rooted trees, which are defined as directed trees with a natural orientation away from a vertex called root. A forest graph is just an acyclic graph, so it can be understood as a graph consisting of one or more tree graphs. On the other hand, a subgraph $H$ of a graph $G$ is said to be a spanning subgraph if both have the same vertex set. Finally, a spanning rooted tree is a spanning subgraph with is also a rooted tree. *Kirchhoff’s matrix-tree theorem* states that the number of spanning trees contained in a graph $G$ is given by the determinant of any first minors of its Laplacian matrix with respect to diagonal elements, i.e., the Laplacian matrix with row and column $j$ removed for any $j$. This theorem can be generalized to higher degree minors or more complex matrices, we refer to [@Tree] and references within for a more comprehensive explanation. We are interested in the Gaudin determinant, which is the determinant of a matrix of the form $$\Phi_{ab}= D_a \delta_{ab} -K_{ab} \ ,$$ where $D_a=L \partial_{u_a} p(u_a)$ and $iK_{a,b}=\partial_{u_a} \log S(u_a ,u_b)$. The determinant of this matrix can be expressed as $$\det \Phi= \sum_{\text{Forests}} \prod_{\text{roots}} D_a \prod_{\text{edges}} K_{ab} \ . \label{treegaudin}$$ The only dependence of the determinant on the length of the spin chain is codified in the terms $D_a$, so the highest power on the length comes from the case of the forest made of isolated vertices. Furthermore, if we break the length $L$ into $l_1+l_2$, we have to add an extra sum over trees having either $l_1 p'$ or $l_2 p'$ as root factor. More on breaking factors {#BFappendix} ======================== In this section we sketch some results about breaking factors beyond rank one. We will also comment on how to modify the construction for operators obeying the Zamolodchikov-Faddeev algebra. Let us start with the breaking factor for an $SU(N)$ spin chain with R-matrix $$\begin{gathered} R_{1,2} (\lambda) =f (u,v) \sum_{a=1}^N E_{aa} \otimes E_{aa} +\sum_{1\leq a\neq b \leq N} E_{aa} \otimes E_{bb}\\ + g(u,v) \sum_{1\leq a\neq b \leq N} (\delta_{a,b+1}+\delta_{a,b-1}) E_{ab} \otimes E_{ba} \ .\end{gathered}$$ In this case the RTT relation (\[RTT\]) can be expressed in the two following ways [^9] $$\begin{aligned} \left[ T_{ij} (u) , T_{kl} (v) \right] &= g(u,v) \big( T_{il} (u) T_{kj} (v) - T_{il} (v) T_{kj} (u) \big) \ , \notag \\ \left[ T_{ij} (u) , T_{kl} (v) \right] &= g(u,v) \big( T_{kj} (v) T_{il} (u) - T_{kj} (u) T_{il} (v) \big) \ .\end{aligned}$$ Writing the monodromy matrix as the product of two monodromy matrices and equating the entries of the auxiliary space we get $$T_{ij}^{(L)}=\sum_k T_{ik}^{(l_1)} \otimes T_{kj}^{(l_2)} \ .$$ It is more convenient to divide the sum into five different sums depending on if $k$ is between $i$ and $j$, equal to one of them, or greater or lower than both. To clarify the following computations we are going to write $T_{ii}=A_i$, $T_{ij}=B_{ij}$ if $i<j$ and $T_{ij}=C_{ij}$ if $i>j$. To compute the breaking factor we have to compute the breaking of $B_{ij} (u) B_{kl} (v)$ and $C_{kl} (v) C_{ij} (u)$. In this appendix we are only going to examine the cases “$k=i$ and $l=j$” and “$j=i+1$ and $l=k+1$” for $B$ operators due to the length of the results. We have to take into account that $B$ operators only commute when both indices are equal. The breaking factor in the first case is extracted from the following expression $$\begin{aligned} &B_{ij} (v) B_{ij} (u) = A_i (v) A_i (u) \otimes B_{ij} (v) B_{ij} (u) + B_{ij} (v) B_{ij} (u) \otimes A_j (v) A_j (u) \\ &+ f(v,u) B_{ij} (v) A_i (u) \otimes B_{ij} (u) A_j (v) + f(u,v) B_{ij} (u) A_i (v) \otimes B_{ij} (v) A_j (u) \notag \\ &+ \sum_{k=i+1}^{j-1} \left[ f(u,v) B_{ik} (u) A_i (v) \otimes B_{ij} (v) B_{kj} (u) + f(v,u) B_{ik} (v) A_i (u) \otimes B_{ij} (u) B_{kj} (v) \right. \notag \\ &+ \left. f(v,u) B_{ij} (v) B_{ik} (u) \otimes B_{kj} (u) A_j (v) + f(u,v) B_{ij} (u) B_{ik} (v) \otimes B_{kj} (u) A_j (v) \right] \ . \notag\end{aligned}$$ It is worth noticing that the first three contributions are equal to the contributions we find for the $SU(2)$ case. In fact, for operators with $j=i+1$ the sum on $k$ vanishes, and we get the same result as in the $SU(2)$ case. Let us move to the breaking factor of $B_{i,i+1} (u) B_{j,j+1} (v)$. For simplicity, we are going to assume that $i<j$, but a similar computation can be done in the converse case. $$\begin{aligned} &B_{i,i+1} (v) B_{j,j+1} (u) = A_i (v) A_j (u) \otimes B_{i,i+1} (v) B_{j,j+1} (u) + B_{i,i+1} (v) B_{j,j+1} (u) \otimes A_{i+1} (v) A_{j+1} (u) \notag \\ &+ f(v,u) B_{i,i+1} (v) A_j (u) \otimes B_{j,j+1} (u) A_{i+1} (v) +f(u,v) B_{j,j+1} (u) A_i (v) \otimes B_{i,i+1} (v) A_{j+1} (u) \notag \\ &+ \delta_{i+1,j} g(v,u) B_{i,i+2} (v) A_{i+1} (u) \otimes \left[ A_{i+1} (u) A_{i+2} (v) - A_{i+1} (v) A_{i+2} (u) \right] \ .\end{aligned}$$ We should pay attention to the appearance of the composite operator $B_{i,i+2}$ in the last equation, making the construction of the weight factors (now weight matrices) more difficult. These explicit computations show that the construction of weights is more complex in the $SU(N)$ case for $N>2$, as obtaining the weights for higher number of excitations from the weights for two excitations involves summing over internal indices. To end this section we should comment about ZF operators. Sometimes it is more interesting to work with operators that form a Zamolodchikov-Faddeev algebra instead of using the creation operators from the ABA, as form factors constructed with them fulfil Smirnov’s axioms. First of all, we should fix an ordering of the operators, as extra S-matrices appear when they are exchanged. We are going to number the operators in the lower edge from left to right and vice-versa in the upper edge. As before, computing the weights for this kind of operators requires first to break one single operator into two factors, which is related to the co-product operation. The co-product applied to ZF operators gives $$\begin{aligned} \Delta \left[ Z(u) \right] &=Z(u) \otimes \mathbb{I} + H(u) e^{i p(u) l_1} \otimes Z(u) \ , \notag \\ \Delta \left[ Z^\dagger (u) \right] &=\mathbb{I} \otimes Z^\dagger(u) + Z^\dagger (u) \otimes H^{-1}(u) e^{i p(u) l_2}\end{aligned}$$ where $H$ is a well-bred operator in the sense of [@ragoucy]. In particular this kind of operators fulfil $H_1 Z_2=S_{12} Z_2 H_1$. Assuming also that $H {\left| 0 \right\rangle}={\left| 0 \right\rangle}$ we obtain the following weights $$\begin{aligned} w (\{ \alpha \} , \{ \bar{\alpha} \} ,\{ \beta \} , \{ \bar{\beta} \} ) &= w_Z (\{ \alpha \} , \{ \bar{\alpha} \} ) w_{Z^\dagger} (\{ \beta \} , \{ \bar{\beta} \} ) \label{weightsZF} \\ w_Z (\{ \alpha \} , \{ \bar{\alpha} \} ) &= e^{i p(\bar{\alpha} )l_1 }S^< (\bar{\alpha},\alpha) \ , \notag \\ w_{Z^\dagger} (\{ \beta \} , \{ \bar{\beta} \} ) &= e^{i p(\beta )l_2 }S^< (\beta, \bar{\beta}) \ . \notag\end{aligned}$$ Here $S^< (\bar{\alpha},\alpha)$ and $S^< (\beta, \bar{\beta})$ means that we are only keeping the terms $S(u_i,u_j)$ with $i<j$. A relationship between the square constructed with ZF states and ABA states can be seen already at the level of these weights. If we redefine the ZF square by extracting a factor of $f^{\neq} ({u},{u})$, we can see that the weights become the ones for the ABA square, giving us a simple connection between both squares. Reconstruction of the spacetime dependence {#spacetimedependence} ========================================== In this appendix we are going to comment on how to recover the spacetime dependence of the two-point function from two different perspectives. First we are going to make use of the string bits/spin bits construction and later we are going to apply the arguments used in the hexagon proposal. The string bits/spin bits construction was developed in [@vertex1] and [@vertex2] and the basic idea is to construct the $\mathcal{N}=4$ SYM operators using the maximal compact subgroup of $U(1) \times SU(2) \times SU(2) \subset PSU(2,2|4)$ and then rotate to the usual representation of the conformal group. Therefore in this language an operator is written as $$\begin{aligned} \hat{\mathcal{O}} (x) \left| 0 \right\rangle &= \sum_{\pi\in \mathcal{S}_L} \text{sign} (\pi) \prod_{i=1}^L \left( \hat{O}^{\pi (i)} (x) \left| 0 \right\rangle^{(i)} \right) \ , & \hat{O}^{i} (x) \left| 0 \right\rangle_i= e^{ix P} U O^i \left| 0 \right\rangle ^{(i)} \ ,\end{aligned}$$ where $\text{sign} (\pi)$ only accounts for signs arising from commutating fermions, $\left| 0 \right\rangle ^{(i)} $ represent a vacuum for each bit, $P$ is the momentum operator, $O$ is an operator constructed using such oscillator representation and $U=e^{\frac{\pi}{4} (P_0 - K_0)}$ is an rotation that relates this representation with the usual representation used for classifying fields with respect to the conformal group. This construction allow us to construct a Yangian-invariant spin vertex [@vertex2], which can be understood as the weak-coupling version of string vertex in string field theory and recovers it in the BMN limit [@vertex3]. Regarding the computation of the norm of the two-point function placed at zero and infinity, we can show that it reduces to the scalar product in the Fock space of the oscillators since the inversion of one of the states is equivalent to acting on it with $U^2$ $$\langle \hat{\mathcal{O}}_j (\infty) \hat{\mathcal{O}}_k (0) \rangle= \langle \hat{\mathcal{O}}^\dagger_j (0) U^2 \hat{\mathcal{O}}_k (0) \rangle=\frac{1}{L} \sum_{\pi, \sigma\in S_L} \text{sign} (\pi) \text{sign} (\sigma) \prod_{i=1}^L \null_i \left\langle 0 \right| O^{\dagger \pi (i)}_j O^{\sigma (i)}_k \left| 0 \right\rangle_i \ ,$$ where we have used that $U^\dagger =U$ and $U^4=1$. If the operators were placed at a finite distance, we would have to commute the exponential of the momentum through the rotation $U$, giving us the tree-level spacetime dependence. The one-loop corrections to the correlation function can be interpreted in this language as the corrections in the commutation relation between the momentum and the $U$ operator (which generate the logarithmic corrections) and from a change in the scalar product of the oscillator (which generate the correction to the mixing matrix). We refer to [@vertex1] for the proofs of these statements. Shifting now to the hexagonalization perspective [@Hexagonalization2], we have computed a square with one operator sit at zero and other at infinity, glued with a second square where we invert the points in which the operators are inserted. The relation between these two frames should be simpler than the twisted translation needed for the hexagon from factor and might impose some constraints to an all-loop construction of the square. The spacetime factor appears as an extra overall dependence in a similar fashion as in the hexagon (see eq. (35), (36) in [@Hexagonalization2] and the discussion between them), which is consistent with the way it arises in the string bit/spin bit formalism. Crossing properties of the square {#CrossingSquare} ================================= As we have commented before, we cannot apply the crossing transformation to quantities computed using the Heisenberg spin chain because it corresponds to the vanishing coupling constant limit, and thus the cuts in the rapidity plane necessary for the crossing transformation collapse. The way to bootstrap a square where we can apply the crossing transformation is to instead construct the consistency condition from the ZF algebra, as in this way we would not be making any assumption on the form of the S-matrix. If we repeat the steps followed in section \[gaudinreconstruction\] using the weights (\[weightsZF\]), we get the same functional form for the square with one excitation $$S_L (u\rightarrow v) =\bar{g} (u,v) \left( e^{ip(u) L} -e^{ip(v) L} \right) \ .$$ We expect the square from factor to have the following two properties under crossing: 1. Moving a excitation clockwise (or anti-clockwise) around the full square should be trivial (up to an S-matrix). We can write this statement as $$\begin{aligned} \bar{g}(u,v)&=\bar{g}(u^c , v)=\bar{g}(v, u^{c} ) S(u^c , v)=\bar{g}(v, u^{2c} ) S(u^c , v) \notag \\ \bar{g}(u,v)&=\bar{g}(v,u^{-c})=\bar{g}(u^{-c}, v) S(v, u^{-c})=\bar{g}(u^{-2c}, v) S(v, u^{-c}) \ . \end{aligned}$$ Where the superindex $c$ means crossing across the mirror side in the clockwise direction and $-c$ means crossing in the counter-clockwise direction. 2. The crossing of the two excitations is the same as changing the excitations, i.e., $S(u^c \rightarrow v^c )=S(v\rightarrow u)$. This imposes $$\bar{g}(u^c , v^c)=-\bar{g}(v,u) e^{i [p(u) + p(v)] L}=\bar{g}(u^{-c} , v^{-c}) \ .$$ However we are not going to attempt to solve these equations in this article. Leading mirror corrections to the empty and one excitation squares {#mirrorappendix} ================================================================== In this appendix we are going to perform computations beyond the asymptotic regime by including mirror excitations. There are two contributions we have ignored: The first one is the obvious contribution to the gluing of the mirror sides of a square to obtain the expression for the scalar product. The second one, more subtle, is the contribution to the consistency condition. We are indeed creating new mirror edges when we cut a square into two, thus we have to consider the appearance of mirror excitations in this relation. Here we will make some comments on corrections of the first kind to in the empty square and corrections of second kind to the square with one excitation. First of all, since in this section we are going to introduce excitations in the mirror sides, we need a more general notation for the square. This notation will be $S_L (\{u\} | \{v\} | \{w\} | \{x\} )$, where the different sets of rapidities are ordered starting from the lower side and continuing clockwise. (0,0) –(0,1); (0, 0) – (1,0); (0, 1) – (1,1); (1,0) –(1,1); (-0.04,0.5-0.04) rectangle (0.04,0.5+0.04); (1-0.04,0.5-0.04) rectangle (1+0.04,0.5+0.04); (0,0.5) – (1,0.5); (1.3,0.4) – (1.6,0.4); (1.3,0.6) – (1.6,0.6); (1.9,0) –(1.9,1); (1.9, 0) – (2.7,0); (1.9, 1) – (2.7,1); (2.7,0) –(2.7,1); (1.9-0.04,0.5-0.04) rectangle (1.9+0.04,0.5+0.04); (2.7-0.04,0.5-0.04) rectangle (2.7+0.04,0.5+0.04); (1.9,0.5) – (2.7,0.5); (2.9,0) –(2.9,1); (2.9, 0) – (3.7,0); (2.9, 1) – (3.7,1); (3.7,0) –(3.7,1); (2.9-0.04,0.5-0.04) rectangle (2.9+0.04,0.5+0.04); (3.7-0.04,0.5-0.04) rectangle (3.7+0.04,0.5+0.04); (2.9,0.5) – (3.7,0.5); We also have to fix our notation regarding the mirror transformation. The rules regarding this transformation can be found, for example, in appendix D of [@BKV]. In particular, we are going to define moving an excitation to the next side in the clockwise direction with the superscript $\gamma$, while moving counter-clockwise will be labelled by $-\gamma$. Applying two times this transformation corresponds to a crossing transformation, $u^{2\gamma}=u^c$, which changes the sign of momentum and energy. For simplicity and concreteness we are going to choose the spin chain model with $a(u)=e^{i p(u) L}$ and $d(u)=1$, where we have $$\begin{aligned} \tilde{a} (u) &= a( u^{-\gamma})= 1 \ , & \hat{a} (u) &=a (u^{+\gamma}) = e^{-E(u) R} \ , \\ \tilde{d} (u) &= d( u^{-\gamma})= e^{-E(u) R} \ , & \hat{d} (u) &=d (u^{+\gamma}) = 1 \ .\end{aligned}$$ Let us consider first the case of the first mirror correction to the empty square. The consistency condition, pictorially represented in figure \[emptymirror\], is $$S_L (\emptyset | u | \emptyset | w)= \int{\frac{dv}{2\pi} S_{l_1} (\emptyset | u | \emptyset | v) S_{l_2} (\emptyset | v | \emptyset | w) } \ .$$ If we substitute here the ansatz $S_L (\emptyset | u | \emptyset | w)=\hat{g}(u,w) [e^{-E(w) R} - e^{-E(u) R}]$, the equation reduces to the sum of the residues at the poles of the function $\hat{g}$. The function $\tilde{g}(u,w)\propto \frac{1}{u-w}$ we obtained for the Gaudin square gives us the correct measure and Boltzmann factor when we take the limit $w\rightarrow u$, but it is not a solution of this equation as the integrand is regular at $v=u$ and $v=w$. On the other hand, this equation can be solved if we use a square obtained from using ZF operators since, according to Smirnov’s axioms, it presents the particle-antipartice residue. This fact can be understood as the consistency condition for mirror excitations being sensible to the kind of operators we are using for constructing our squares. In particular, it forces us to use Zamolodchikov-Faddeev operators, as the form factors constructed with them fulfil the Smirnov’s axioms. (0,0) –(0,1); (0, 0) – (1,0); (0, 1) – (1,1); (1,0) –(1,1); (0.5,0) circle \[radius=0.04\]; (0.5,1) circle \[radius=0.04\]; (0.5,0) – (0.5,1); (1.3,0.4) – (1.6,0.4); (1.3,0.6) – (1.6,0.6); (1.9,0) –(1.9,1); (1.9, 0) – (2.7,0); (1.9, 1) – (2.7,1); (2.7,0) –(2.7,1); (2.3,0) circle \[radius=0.04\]; (2.3,1) circle \[radius=0.04\]; (2.3,0) – (2.3,1); (2.9,0) –(2.9,1); (2.9, 0) – (3.7,0); (2.9, 1) – (3.7,1); (3.7,0) –(3.7,1); (4,0.5) – (4.3,0.5); (4.15,0.35) – (4.15,0.65); (4.6,0) –(4.6,1); (4.6, 0) – (5.4,0); (4.6, 1) – (5.4,1); (5.4,0) –(5.4,1); (5.6,0) –(5.6,1); (5.6, 0) – (6.4,0); (5.6, 1) – (6.4,1); (6.4,0) –(6.4,1); (6,0) circle \[radius=0.04\]; (6,1) circle \[radius=0.04\]; (6,0) – (6,1); (1.35,0.5-1.5) – (1.65,0.5-1.5); (1.5,0.35-1.5) – (1.5,0.65-1.5); (1.9,0-1.5) –(1.9,1-1.5); (1.9, 0-1.5) – (2.7,0-1.5); (1.9, 1-1.5) – (2.7,1-1.5); (2.7,0-1.5) –(2.7,1-1.5); (2.3,0-1.5) circle \[radius=0.04\]; (2.7-0.04,0.5-1.5-0.04) rectangle (2.7+0.04,0.5-1.5+0.04); (2.3,0-1.5) to \[out=90,in=180\] (2.7,0.5-1.5); (2.9,0-1.5) –(2.9,1-1.5); (2.9, 0-1.5) – (3.7,0-1.5); (2.9, 1-1.5) – (3.7,1-1.5); (3.7,0-1.5) –(3.7,1-1.5); (3.3,1-1.5) circle \[radius=0.04\]; (2.9-0.04,0.5-1.5-0.04) rectangle (2.9+0.04,0.5-1.5+0.04); (3.3,1-1.5) to \[out=-90,in=0\] (2.9,0.5-1.5); (4,0.5-1.5) – (4.3,0.5-1.5); (4.15,0.35-1.5) – (4.15,0.65-1.5); (4.6,0-1.5) –(4.6,1-1.5); (4.6, 0-1.5) – (5.4,0-1.5); (4.6, 1-1.5) – (5.4,1-1.5); (5.4,0-1.5) –(5.4,1-1.5); (5,1-1.5) circle \[radius=0.04\]; (5.4-0.04,0.5-1.5-0.04) rectangle (5.4+0.04,0.5-1.5+0.04); (5,1-1.5) to \[out=-90,in=180\] (5.4,0.5-1.5); (5.6,0-1.5) –(5.6,1-1.5); (5.6, 0-1.5) – (6.4,0-1.5); (5.6, 1-1.5) – (6.4,1-1.5); (6.4,0-1.5) –(6.4,1-1.5); (6,0-1.5) circle \[radius=0.04\]; (5.6-0.04,0.5-1.5-0.04) rectangle (5.6+0.04,0.5-1.5+0.04); (6,0-1.5) to \[out=90,in=0\] (5.6,0.5-1.5); Moving now to the case of one physical excitation, figure \[onemirror\] represents the terms appearing in the consistency equation \[bootstrap\] when we allow for mirror corrections. We can write the consistency condition in the following way $$\begin{aligned} S_L (u|\emptyset |w| \emptyset ) &= S_{l_1} (u|\emptyset |w| \emptyset ) d_{l_2} (u) a_{l_2} (w) +S_{l_2} (u|\emptyset |w| \emptyset ) a_{l_1} (u) d_{l_1} (w) \notag \\ &+ \int{ \frac{dv}{2\pi} S_{l_1} (u|\emptyset | \emptyset|v ) S_{l_2} (\emptyset | v | w | \emptyset )} d_{l_2} (u) d_{l_1} (w) \notag \\ &+ \int{ \frac{dv}{2\pi} S_{l_1} (\emptyset |v| \emptyset | w ) S_{l_2} (u | v | \emptyset | \emptyset )} a_{l_1} (u) a_{l_2} (w) \ .\end{aligned}$$ In contrasts with the gluing of two hexagons, here only two new corrections appear. This is because we have no mirror excitations in the original mirror sides and only two physical excitations to annihilate the set of mirror excitations we add, so we are forced to put at most only one mirror excitation if we want a non-vanishing contribution due to symmetry. To solve the consistency equation we substitute the ansatz $$S_L (u\rightarrow w)= \alpha (u,w) a(u) a(w) + \beta(u,w) a(u) d(w) + \gamma(u,w) d(u) a(w) + \delta(u,w) d(u) d(w) \ ,$$ reducing it to $$\begin{aligned} \alpha_L &= \hat{I}_1 \ , & \beta_L &= \beta_{l_2} \ , & \gamma_L &=\gamma_{l_1} \ , & \delta_L &= \tilde{I}_4 \ , \\ \alpha_{l_1} &= -\hat{I}_3 \ , & \delta_{l_1} &= -\tilde{I}_2 \ , & \alpha_{l_2} &= -\hat{I}_2 \ , & \delta_{l_2} &= -\tilde{I}_3 \ ,\end{aligned}$$ together with $\beta_{l_1} + \gamma_{l_2}+\tilde{I}_1 + \hat{I}_4=0$. The different $I$’s are given by the following integrals $$\begin{aligned} \tilde{I}_1 &=\int{\frac{dv}{2\pi} \left[ \alpha_{l_1} \alpha_{l_2} [\tilde{a} (v)]^2 + \left( \alpha_{l_1} \gamma_{l_2} + \beta_{l_1} \alpha_{l_2} \right) \tilde{a} (v) \tilde{d} (v)+ \gamma_{l_1} \gamma_{l_2} [\tilde{d} (v)]^2 \right]} \ , \\ \tilde{I}_2 &=\int{\frac{dv}{2\pi} \left[ \gamma_{l_1} \alpha_{l_2} [\tilde{a} (v)]^2 + \left( \gamma_{l_1} \gamma_{l_2} + \delta_{l_1} \alpha_{l_2} \right) \tilde{a} (v) \tilde{d} (v)+ \delta_{l_1} \gamma_{l_2} [\tilde{d} (v)]^2 \right]} \ , \\ \tilde{I}_3 &=\int{\frac{dv}{2\pi} \left[ \alpha_{l_1} \beta_{l_2} [\tilde{a} (v)]^2 + \left( \alpha_{l_1} \delta_{l_2} + \beta_{l_1} \beta_{l_2} \right) \tilde{a} (v) \tilde{d} (v)+ \gamma_{l_1} \delta_{l_2} [\tilde{d} (v)]^2 \right]} \ , \\ \tilde{I}_4 &=\int{\frac{dv}{2\pi} \left[ \gamma_{l_1} \beta_{l_2} [\tilde{a} (v)]^2 + \left( \gamma_{l_1} \delta_{l_2} + \delta_{l_1} \beta_{l_2} \right) \tilde{a} (v) \tilde{d} (v)+ \delta_{l_1} \delta_{l_2} [\tilde{d} (v)]^2 \right]} \ , \\ \hat{I}_1 &=\int{\frac{dv}{2\pi} \left[ \alpha_{l_1} \alpha_{l_2} [\hat{a} (v)]^2 + \left( \gamma_{l_1} \alpha_{l_2} + \alpha_{l_1} \beta_{l_2} \right) \hat{a} (v) \hat{d} (v)+ \gamma_{l_1} \beta_{l_2} [\hat{d} (v)]^2 \right]} \ , \\ \hat{I}_2 &=\int{\frac{dv}{2\pi} \left[ \beta_{l_1} \alpha_{l_2} [\hat{a} (v)]^2 + \left( \delta_{l_1} \alpha_{l_2} + \beta_{l_1} \beta_{l_2} \right) \hat{a} (v) \hat{d} (v)+ \delta_{l_1} \beta_{l_2} [\hat{d} (v)]^2 \right]} \ , \\ \hat{I}_3 &=\int{\frac{dv}{2\pi} \left[ \alpha_{l_1} \gamma_{l_2} [\hat{a} (v)]^2 + \left( \gamma_{l_1} \gamma_{l_2} + \alpha_{l_1} \delta_{l_2} \right) \hat{a} (v) \hat{d} (v)+ \gamma_{l_1} \delta_{l_2} [\hat{d} (v)]^2 \right]} \ , \\ \hat{I}_4 &=\int{\frac{dv}{2\pi} \left[ \beta_{l_1} \gamma_{l_2} [\hat{a} (v)]^2 + \left( \delta_{l_1} \gamma_{l_2} + \beta_{l_1} \delta_{l_2} \right) \hat{a} (v) \hat{d} (v)+ \delta_{l_1} \delta_{l_2} [\hat{d} (v)]^2 \right]} \ .\end{aligned}$$ The arguments of the functions involved in the tilde integrals are $(u,v^{-\gamma})$ and $(v^{-\gamma},w)$ for the functions with subindexes $l_1$ and $l_2$ respectively, while for the hat integrals the arguments are $(v^{\gamma},w)$ and $(u,v^{\gamma})$ for $l_1$ and $l_2$ respectively. Note that not all these integrals are independent, as the square should be invariant under the interchange of the initial and final physical rapidities. This implies the following relations between the unknown functions of the ansatz $$\begin{aligned} \alpha (u,w) &= \alpha (w,u) \ , & \delta (u,w) &=\delta (w,u) \ , & \beta (u,w) &= \gamma (w,u) \ .\end{aligned}$$ Using these properties we can prove some consistency of this ansatz as $\alpha_{l_1}=\hat{I}_2 (u,w) = \hat{I}_3 (w,u)=\alpha_{l_2}$ and similarly for the tilde ones, which implies that both $\alpha$ and $\delta$ are independent of the length. Furthermore, as $I_1$’s and $I_4$’s transform into themselves, we can write function $\beta$ and $\gamma$ as $$\beta (u,w)-\frac{\tilde{I}_1 +\hat{I}_4}{2} = \gamma (u,w)+\frac{\tilde{I}_1 +\hat{I}_4}{2} =\tilde{g}(u,v) \ ,$$ where $\tilde{g}$ is an arbitrary function with the property $\tilde{g}(u,v)=-\tilde{g}(v,u)$. Using these properties the problem reduces to a system of coupled Fredholm integral equations of second kind, which can be solved as a Liouville-Neumann series. However, we need an initial condition to apply this method. As we have seen for the case of the mirror excitation in the empty square, the initial condition we have to use is the square computed using ZF operators. [99]{} F. Smirnov, *Form-factors in completely integrable models of quantum field theory*, Adv. Ser. Math. Phys. 14, 1 (1992). H. Babujian, A. Fring, M. Karowski and A. Zapletal, *Exact form factors in integrable quantum field theories: the sine-[G]{}ordon model*, Nucl. Phys. B 538 (1999) 535-586, . H. Babujian and M. Karowski, *Exact form factors in integrable quantum field theories: the sine-[G]{}ordon model ([II]{})*, Nucl. Phys. B 620 (2002) 407-455, . A. Nakayashiki and Y. Takeyama, *On [Form Factors of SU(2) Invariant Thirring Model]{}*, [MathPhys]{} Odyssey 2001, . Y. Takeyama, *Form factors of ${SU(N)}$ invariant [T]{}hirring model*, Publ.Res.Inst.Math.Sci.Kyoto 39 (2003) 59-116, . H. M. Babujian, A. Foerster and M. Karowski, *The [Form]{} [Factor]{} [Program]{}: a [Review]{} and [New]{} [Results]{} - the [Nested]{} ${SU}({N})$ [Off-Shell]{} [Bethe]{} [Ansatz]{}*, SIGMA 2, 082 (2006). T. Klose and T. McLoughlin, *Worldsheet Form Factors in AdS/CFT*, Phys. Rev. D [**87**]{} (2013) no.2, 026004, . T. Klose and T. McLoughlin, *Comments on World-Sheet Form Factors in AdS/CFT*, J. Phys. A [**47**]{} (2014) no.5, 055401, . N. Gromov, V. Kazakov, S. Leurent and D. Volin, *Quantum Spectral Curve for Planar $\mathcal{N} = 4$ Super-Yang-Mills Theory*, Phys. Rev. Lett.  [**112**]{} (2014) no.1, 011602, . N. Gromov, F. Levkovich-Maslyuk, G. Sizov and S. Valatka, *Quantum spectral curve at work: from small spin to strong coupling in $ \mathcal{N} $ = 4 SYM*, JHEP [**1407**]{} (2014) 156, . N. Gromov, V. Kazakov, S. Leurent and D. Volin, *Quantum spectral curve for arbitrary state/operator in AdS$_{5}$/CFT$_{4}$*, JHEP [**1509**]{} (2015) 187, 011602, . C. Marboe and D. Volin, *Quantum spectral curve as a tool for a perturbative quantum field theory*, Nucl. Phys. B [**899**]{} (2015) 810, . C. Marboe, V. Velizhanin and D. Volin,, *Six-loop anomalous dimension of twist-two operators in planar $ \mathcal{N}=4 $ SYM theory*, JHEP [**1507**]{} (2015) 084, . N. Gromov, F. Levkovich-Maslyuk and G. Sizov, *Quantum Spectral Curve and the Numerical Solution of the Spectral Problem in AdS5/CFT4*, JHEP [**1606**]{} (2016) 036, . V. Kazakov, S. Leurent and D. Volin, *T-system on T-hook: Grassmannian Solution and Twisted Quantum Spectral Curve*, JHEP [**1612**]{} (2016) 044, . C. Marboe and V. Velizhanin, *Twist-2 at seven loops in planar $ \mathcal{N} $ = 4 SYM theory: full result and analytic properties*, JHEP [**1611**]{} (2016) 013, . C. Marboe and D. Volin, *Fast analytic solver of rational Bethe equations*, J. Phys. A [**50**]{} (2017) no.20, 204002, . C. Marboe and D. Volin, *The full spectrum of AdS5/CFT4 I: Representation theory and one-loop Q-system*, J. Phys. A [**51**]{} (2018) no.16, 165401, . K. Okuyama and L. S. Tseng, *Three-point functions in N = 4 SYM theory at one-loop*, JHEP [**0408**]{} (2004) 055 . R. Roiban and A. Volovich, *Yang-Mills correlation functions from integrable spin chains*, JHEP [**0409**]{} (2004) 032 . L. F. Alday, J. R. David, E. Gava and K. S. Narain, *Towards a string bit formulation of N=4 super Yang-Mills*, JHEP [**0604**]{} (2006) 014, . J. Escobedo, N. Gromov, A. Sever and P. Vieira, *Tailoring Three-Point Functions and Integrability*, JHEP [**1109**]{} (2011) 028 . B. Basso, S. Komatsu and P. Vieira, *Structure Constants and Integrable Bootstrap in Planar N=4 SYM Theory*, . B. Basso, V. Goncalves, S. Komatsu and P. Vieira, *Gluing Hexagons at Three Loops*, Nucl. Phys. B [**907**]{} (2016) 695, . J. Caetano and T. Fleury, *Fermionic Correlators from Integrability*, JHEP [**1609**]{} (2016) 010, . J. L. Cardy, O. A. Castro-Alvaredo and B. Doyon, *Form factors of branch-point twist fields in quantum integrable models and entanglement entropy*, J. Statist. Phys.  [**130**]{} (2008) 129, . B. Eden and A. Sfondrini, *Tessellating cushions: four-point functions in $\mathcal{N} $ = 4 SYM*, JHEP [**1710**]{} (2017) 098, . T. Fleury and S. Komatsu, *Hexagonalization of Correlation Functions*, JHEP [**1701**]{} (2017) 130, . B. Basso, F. Coronado, S. Komatsu, H. T. Lam, P. Vieira and D. l. Zhong, *Asymptotic Four Point Functions*, . T. Fleury and S. Komatsu, *Hexagonalization of Correlation Functions II: Two-Particle Contributions*, JHEP [**1802**]{} (2018) 177, . M. Kim and N. Kiryu, *Structure constants of operators on the Wilson loop from integrability*, JHEP [**1711**]{} (2017) 116, . L. D. Faddeev, *How algebraic Bethe ansatz works for integrable model*, . Y. Jiang and A. Petrovskii, *Diagonal form factors and hexagon form factors*, JHEP [**1607**]{} (2016) 120, . G. Kirchhoff, *Über die Auflösung der Gleichungen, auf welche man bei der untersuchung der linearen verteilung galvanischer Ströme geführt wird*, Ann. Phys. Chem. 72, 497-508, 1847. I. Kostov, D. Serban, D. L. Vu, *TBA and tree expansion*, . B. Basso, A. Rej, *Bethe ansätze for GKP strings*, Nucl. Phys. B [**879**]{} (2014) 162, . Y. Wang, W. Yang, J. Cao, K. Shi, *Off-Diagonal Bethe Ansatz for Exactly Solvable Models*, Springer-Verlag Berlin Heidelberg, ISBN 978-3-662-46755-8. A. Hutsalyuk, A. Liashyk, S. Z. Pakuliak, E. Ragoucy and N. A. Slavnov, *Scalar products and norm of Bethe vectors for integrable models based on $U_q(\widehat{\mathfrak{gl}}_{n})$*, . J. M. Nieto, *Spinning strings and correlation functions in the AdS/CFT correspondence*, . N. Gromov and P. Vieira, *Tailoring Three-Point Functions and Integrability IV. Theta-morphism*, JHEP [**1404**]{} (2014) 068, . N. Beisert, V. Dippel and M. Staudacher, *A Novel long range spin chain and planar N=4 super Yang-Mills*, JHEP [**0407**]{} (2004) 075, . B. Basso, V. Goncalves and S. Komatsu, *Structure constants at wrapping order*, JHEP [**1705**]{} (2017) 124, . Z. Bajnok and R. A. Janik, *From the octagon to the SFT vertex — gluing and multiple wrapping*, JHEP [**1706**]{} (2017) 058, . E. Ragoucy, *Vertex operators for quantum groups and application to integrable systems*, J. Phys. A [**35**]{} (2002) 7929, . Y. Jiang, I. Kostov, A. Petrovskii and D. Serban, *String Bits and the Spin Vertex*, Nucl. Phys. B [**897**]{} (2015) 374, . Y. Jiang and A. Petrovskii, *From Spin Vertex to String Vertex*, JHEP [**1506**]{} (2015) 172, . [^1]: Using these functions instead of their explicit expressions allow us to generalize our results to the trigonometric model just by substituting their expression by the corresponding one without altering any formula. We also want to remark that we are defining $f(u,v)=1+g(u,v)$ in contrast with e.g. [@Faddeev], where $f(u,v)=1-g(u,v)$. [^2]: Although this is true for cutting the square vertically, horizontal cuts should behave differently. We will not address that kind of cuts in this article. [^3]: Actually we can choose a more general solution $S_L (\emptyset )=\exp (\rho L)$ with $\rho$ being a constant. However, we are going to choose $\rho=0$ because there is no physical parameters with inverse length units we can use for that role. [^4]: The sum over momenta appearing here can be made explicit in (\[treegaudin\]) by separating the sum over directed forests into a sum over undirected forests and the contribution from the roots, being this last sum the sum over momenta. [^5]: Note that the first six equations reduce to only two and the condition that $\alpha_{u,v}$ and $\alpha_{v,u}$ are independent of the length. [^6]: This is true for any power of $L$, as the equation $\sigma^{(n)}_j L^j=\sigma^{(n)}_j l_1^j + \sigma^{(n)}_j l_2^j +\sum_{l=1}^{j-1} \binom{j}{i} \alpha\left( \u \right) \, l_1^{j-l} l_2^l +h(u,v,w) L^j$ cannot be solved unless $h(u,v,w)$ vanishes. [^7]: The author would like to thank Ivan Kostov and Thiago Fleury for proposing this method. [^8]: For a construction of the hexagon vertex using ZF operators, see section 7.3 of [@thesis]. [^9]: The counterpart of these equations for the trigonometric case are more complex and will not be treated here. Their explicit expression can be found for example in eq. (2.4) of [@composite].

ArXiv

--- abstract: 'Active authentication refers to a new mode of identity verification in which biometric indicators are *continuously* tested to provide real-time or near real-time monitoring of an authorized access to a service or use of a device. This is in contrast to the conventional authentication systems where a single test in form of a verification token such as a password is performed. In active voice authentication (AVA), voice is the biometric modality. This paper describes an ensemble of techniques that make reliable speaker verification possible using unconventionally short voice test signals. These techniques include model adaptation and minimum verification error (MVE) training that are tailored for the extremely short training and testing requirements. A database of 25 speakers is recorded for developing this system. In our off-line evaluation on this dataset, the system achieves an average windowed-based equal error rates of 3-4% depending on the model configuration, which is remarkable considering that only 1 second of voice data is used to make every single authentication decision. On the NIST SRE 2001 Dataset, the system provides a 3.88% absolute gain over i-vector when the duration of test segment is 1 second. A real-time demonstration system has been implemented on Microsoft Surface Pro.' address: 'Georgia Institute of Technology, Atlanta, GA 30308, USA' author: - 'Zhong Meng, M Umair Bin Altaf, and Biing-Hwang (Fred) Juang' bibliography: - 'refs.bib' title: Active Voice Authentication --- Active voice authentication, Continuous speaker verification, Hidden Markov model, Minimum verification error Introduction ============ User authentication refers to the process of validating a user’s claim of identity in order to grant or deny the user access to a device or service. The prevalent method for user authentication operates in predominantly the so-called gatekeeper mode, in that the guarding system asks the user to present what he/she knows (e.g., a password), what he/she has (e.g., a key or a fob), or what he/she is (e.g., fingerprints, iris scan) for examination in order to render the decision. Once the access is granted, the device or service remains “active” until it is signed off or terminated. During the active session, no action is taken by the guarding system even though the user may have changed, resulting in security compromises. An active authentication (AA) system seeks to actively and continuously validate the identity of the person by making use of his or her unique biometric identifiers without repetitively prompting the user for credentials or requiring the user to change his/her work-flow during the use of the device or service. The AA framework differs from a conventional authentication system in that it provides a continuous and real-time monitoring of the user’s identity rather than just a one-shot authentication in form of verifying a test token as in the gatekeeper mode. Many biometric identifiers including physiological and behavioral indicators can be used as human characteristics to actively verify the identity of a user [@biometrics_define_acm]. As the intrinsic attributes, the facial appearance [@face_acm], the iris pattern [@iris_ieee], the finger-print [@fingerprint_acm], the voice pattern [@voice], the hand geometry [@hand] and body’s electric pulse response [@body_pulse_response] are widely used as the physiological identifiers while the manner people walk [@walk], write [@write], and type [@wang2012user] are commonly used as the behavioral identifiers. The target modality in this paper is the voice. The voice of a person is unique. This is because the construction of the articulatory apparatus and its use that generate and modulate the voice of a talkerthe lungs, the vocal cords, the articulators, etc.are uniquely configured for a given individual and this configuration is naturally embedded in the person’s voice characteristics. Thus, in addition to language, voice conveys the latent identity of its speaker. Human voice is ideally suited for AA as it provides contact-less authentication; it is straightforward to acquire authentication data from ubiquitous microphones available on all platforms. An active voice authentication (AVA) system uses the voice of a person to achieve AA, as the person uses the phone or any other voice application on a mobile or desktop device. The AVA system does not interfere with other active authentication methods on the device and can work in the background with already-installed voice applications, such as Skype, Voice note or the Phone to provide real-time continuous monitoring of the user’s identity. AVA can effectively strengthen the security of the existing voice assistants such as Amazon Alexa, Microsoft Cortana, Apple Siri and Google Home [@hoy2018alexa; @lopez2017alexa] and enable services which involve money transfers. On top of the initial speaker verification using the starting anchor (wake) word, AVA continues to repeatedly authenticate the user’s voice as the conversation goes on and reserves the right to overturn its initial decision at any time. Recently, voice has been successfully used to assist the other biometrics such as body-surface vibrations [@feng2017continuous], touch gestures [@peng2017continuous] and a combination of face, mouse and keystroke [@fenu2017multi] in performing continuous authentication. In these works, the voice is authenticated in form of voice commands which are first stored as audio files and are then verified through support vector machine (SVM) [@feng2017continuous; @peng2017continuous] or vector quantization [@fenu2017multi]. Although these systems perform further authentications after the session is activated, voice only plays an *auxiliary* role in protecting the system because the decisions are still made at utterance (command)-level as in conventional speaker verification with a low resolution of 3 seconds or more. The other biometrics are necessary for AA especially when the speaker is not talking. Therefore, these voice-assisted authentication systems do not meet the requirement of AVA. As with any AA system, AVA involves two phases, the registration phase and the authentication phase. During registration, the user being registered is asked to utter some standard speech material. A set of statistical models is trained a adapted to signify the user’s voice identity. At active authentication stage, once AVA system detects a valid voice signal through a voice activity detector (VAD), it starts continuously evaluating real-time confidence scores given the speech signal. Depending on the score, the system can grant or deny the user’s access to the device. If silence is detected to be long than the latency, AVA can report an authentication score that indicates impostor. AVA is significantly different from traditional speaker verification task directed and organized by NIST Speaker Recognition Evaluation (SRE). The goal of AVA is to continuously authenticate the speaker identity with the assumption that change of talker can potentially occur at any instance whereas in most SREs, such an abrupt change of talker does not happen and its goal is to produce a final decision after the entire test utterance is obtained. Because of the large distinction between AVA and the traditional speaker verification, a new design framework is necessary for the AVA system which we will elaborate in Section \[sec:real\_time\]. The AVA system integrates the techniques of sequential training and testing, maximum a posteriori (MAP) adaptation, cohort selection and minimum verification error (MVE). The major contributions of this paper are the following: - Propose a novel AVA framework that continuously verifies the speaker’s identity and instantaneously reports verification decisions. - Propose a window-based short-time sequential testing scheme to accommodate the real-time requirement of AVA. - Propose a window-based short-segment training scheme to model the short-time statistics of a speaker’s voice through an HMM and to match the real-time testing condition. - Apply MAP adaptation of an speaker-independent (SI) HMM to minimize the enrollment data needed for reliable short-time speaker modeling. - Apply MVE training to further minimize the speaker verification error on top of MAP. Propose cohort selection method to address the imbalanced target and impostor data for MVE training. AVA performs speaker verification using second-long speech signals and achieves a performance of 3-4% average window-based equal error rate (WEER), depending on the model configuration. This level of performance, being able to reasonably authenticate a talker’s claimed identity with 1 second voice, outperforms conventional techniques, as will be reported in later sections, and outstrips human capabilities based on the informal observation of our research group members. A separate talker authentication evaluation on human performance is necessary to formally establish the comparison. We also evaluate the proposed methods on NIST SRE 2001 dataset with a large number of speakers and the proposed system provides a 3.88% absolute gain over i-vector on the when the duration of test segment is 1 second. The rest of the paper is organized as follows. In Section \[sec:principle\], we briefly discuss the conventional formulation of the problem of speaker identification and verification. We explain how the differences between AVA and the traditional talker verification paradigm would call for a new design methodology. In Section \[sec:real\_time\], we introduce the challenge of real-time voice authentication, how it is performed, and why the speaker models need to be trained to match the test statistics. In Section \[sec:metric\], we use window-based EER to evaluate the performance of the proposed AVA system. In Section \[sec:ava\_data\], we discuss the registration and the data collection procedure. In Section \[sec:ava\_ivector\], we evaluate the i-vector technique for the AVA task. In Section \[sec:train\], we introduce the architecture of the training and registration modules of the AVA system and the algorithms that are applied to its major components. In Section \[sec:test\], we discuss sequential testing in the AVA system. In Section \[sec:expr\], we provide the evaluation results of the AVA system with different configurations and algorithms. System Description and Technical Issues {#sec:problem} ======================================= Conventional Voice Authentication {#sec:principle} --------------------------------- Use of a person’s voice as a biometric indicator requires processing of the signal to retain a salient representation of the speaker-specific characteristics. Traditionally, these may include the talker’s source parameters (e.g., range and dynamics of the pitch contour [@autocorrelation_pitch], stress patterns) and the tract parameters (e.g., the mean behavior of formant frequencies, vocal tract length [@cepstrum_vt; @lpc_vt]). Overall, since these biometric parameters of the voice production system represent a talker’s intrinsic articulatory characteristics, a substantial duration of the speech signal is necessary, often in tens of minutes or even hours [@long_term_feature], to support reliable estimation. With advances in statistical modeling techniques, such as the hidden Markov model (HMM), spectral features have become the dominant choice to discriminate talker-specific voice characteristics [@hmm_furui; @hmm_sarkar]. This has allowed a relative decrease in the duration of the speech material required for training and testing, though it still remains impractical for real-time monitoring applications. To address this problem, the traditional authentication approach is to use a likelihood ratio test with MAP adapted universal background models (UBM) [@spkr_id_rose; @ubm_map] which are built using Gaussian mixture models (GMMs). Adaptation techniques are used to update the parameters of a pre-trained model using the new speech signal. Further, discriminative training methods are applied to refine the speaker models with the goal of maximizing the speaker verification performance. In [@mve_speaker], MVE training is proposed to jointly estimate the target and anti-target speaker models so that the expected number of verification errors (miss detection and false alarm) on enrollment and training set are minimized. Similarly, in [@liu1995study; @angkititrakul2007discriminative], minimum classification error (MCE) criterion [@mce] is used for speaker recognition and identification. Based on these, the application of SVM in a speaker’s GMM supervector space [@svm_1; @svm_3] yields interesting results by performing a nonlinear mapping from the input space to an SVM extension space. More recently, factor analysis methods such as joint factor analysis (JFA) [@jfa_2; @jfa_3] and i-vectors [@ivector_1; @ivector_3] become the dominant approach for speaker verification. These approaches try to model the speaker and channel variability by projecting speaker dependent GMM mean supervectors onto a space of reduced dimensionality. In recent years, deep vector (d-vector) [@variani2014deep] approach has achieved state-of-the-art performance in NIST SREs, in which a deep neural network (DNN) is trained to classify speaker identities given their voice at the input. A d-vector is extracted per utterance by averaging the DNN hidden units to represent a registered speaker or a test utterance for subsequent speaker verification. Further, an end-to-end loss [@heigold2016end] and a triplet loss [@li2017deep] are introduced to learn more relevant embeddings to the speaker verification task. An attention mechanism is applied to dynamically summarize the DNN hidden units into speaker embeddings [@rahman2018attention]. To improve the noise robustness, DNN-based speaker embedding is further extended to x-vector in [@snyder2018x] by performing data augmentation. More recently, adversarial learning [@gan] with gradient reversal network [@ganin2015unsupervised] has been applied to domain adaptation [@sun2017unsupervised; @meng2017unsupervised] and domain-invariant training [@shinohara2016adversarial; @meng2018speaker; @meng2018adversarial] of the DNN acoustic model [@hinton2012deep]. Similarly, it can effectively improve the robustness of the speaker embeddings by jointly optimizing the DNN speaker classifier and an auxiliary disriminative network to mini-maximize an adversarial objective [@wang2018unsupervised; @meng2019asv]. However, these methods are specifically designed and well suited only for speaker verification tasks within the NIST SRE framework, in which long speech utterances are used as the material for single individual tests, ranging in duration from 10s to a few minutes depending on the specific task (e.g., see the NIST speaker verification tasks in the years 2000-2010 [@nist_website].). More specifically, these techniques work well only for modeling the long-term statistical characteristics of a speaker, which does not coincide with the short-time testing condition required by the AVA task. In Section \[sec:ava\_ivector\], we show that the AVA system based on i-vector achieves an excellent authentication performance when the duration of the test window is long enough. But the performance degrades rapidly as the test window duration decreases. In general, many i-vector based systems exhibit sharp performance degradation [@ivector_short_2011; @ivector_short_2012; @poddar2019quality; @ivector_short_kheder], when they are tested with short duration (below 5s) utterances. This is understandable as the covariance matrix of the i-vector is inversely proportional to the number of speech frames per test utterance and the variance of the i-vector estimate grows directly as the number of frames in the test utterance decreases [@jfa_3]. Recently, many approaches have been proposed for speaker verification with short-duration utterances. By borrowing the idea from speaker-adaptive training, the authors of [@soldi2014short] propose phone adaptive training (PAT) to learn a set of transforms that project features into a phoneme-normalized but speaker-discriminative space and use the normalized feature to improve speaker modeling given short-duration enrollment data. To alleviate the large estimation variation of i-vector due to short-duration utterances, uncertainty propagation is introduced to both the PLDA classifier [@stafylakis2013text] and the i-vector extraction [@kenny2013plda]. However, these methods only show their effectiveness on test utterances of about 3 seconds duration and can hardly meet the real-time requirement of AVA. To further overcome the mismatched prior distributions of the data used to train UBM and short-duration enrollment data, [@li2016improving] divides the speech signal into several subregions defined by speech unit and perform speaker modeling and verification within each subregion. A good improvement is achieved over GMM-UBM baseline for test utterances no longer than 2 seconds. However, these systems are rather complicated which entail large computations during testing and may lead to non-negligible delays in making real-time decisions of AVA. Challenges in Real-Time Voice Authentication {#sec:real_time} -------------------------------------------- Most speech processing systems follow the convention of the short-time analysis (STA) framework, in which segments of signal, each being called a *speech frame* with a duration (denoted by $T_f$) of $20-40$ ms, are successively extracted for analysis. The successive analysis is performed at a predefined rate, called the frame rate denoted as $r_f$, a prevalent choice of which is 100 per second. The frame rate can be converted to frame shift, $\delta_f$, which is the reciprocal of $r_f$. The continuous monitoring mode of AVA dictates that it must be *text-independent*, and it must perform *real-time* authentication sequentially, *continuously* reporting the near-instantaneous authentication results in preparation for possible breach of prior authentication at any moment. The major challenge in designing such a system is to effectively train talker-specific models, using as little enrollment speech as possible, for accurate, continuous and instantaneous *text-independent* speaker verification, with very short test signals. Since a talker change may happen abruptly, the authentication decision cannot be based on a long memory of both the signal representations and the prior decisions. However, it is well known in statistical analysis that more data means better test results. A trade-off is thus necessary in determining the duration of data, which is subject to successive authentication tests. This duration will involve multiple aforementioned frames as the typical analysis frame length of $20-40$ ms is known to be far too short for reliable hypothesis testing. We shall call the test segment a *“window”*, which is expressed in number of frames, $N_w$, and is equivalent to $(N_w - 1)\delta_f + T_f$ of signal in time. As the system slides the test “window” through the sequence of frame-based representations and obtain the corresponding test scores, the reporting interval then defines how often these scores need to be reported. In other words, the temporal resolution for authentication test may not be identical to that for reporting. Fig. \[fig:sliding\_window\] illustrates the concept of analysis frames and test windows. ![An illustration of the successive tests performed with data windows on a short-time spectrogram.[]{data-label="fig:sliding_window"}](sliding_window.png){width="0.7\columnwidth"} To make accurate decisions, we need to model the speaker characteristics within the specified short-time test windows. An unconventional speaker modeling concept is in order here due to the aforementioned short-test condition. In the usual hidden Markov modeling of speech for speaker identification or verification, the model is implicitly assumed to be a characterization of the *general statistical behavior of the signal source, without any regard to the test duration*, and the likelihood calculation is a mere accumulation of the frame likelihoods over whatever length of the given utterance might be. This means the models so trained in the conventional manner, without a definitive duration of data, will have an inherent mismatch condition in the captured statistical knowledge, and may not lead to the most reliable likelihood test results. To deal with this problem, we have to match the training and testing condition by extracting short-time speech segments with a matching duration from the training and enrollment speech feature sequence. The speech segment within a sliding window at each time will serve as a training token for HMM. In this case, the talker-specific HMM, which includes a pair of target and anti-target models, is able to model the short-time characteristics of a speaker that is required by the real-time testing process. To meet the challenge of minimal enrollment, we adapt a speaker-independent (SI) model with the MAP adaptation technique [@map] to make use of the limited adaptation data and to obtain a decent estimation of the model parameters with prior knowledge about the target model distribution. In addition, the method of MVE training [@mve_speaker] is applied so that the total verification error is minimized. MVE in [@mve_speaker] must be adapted to the current operating setup of short test signals. It means that the notion of empirical error estimate in discriminative methods must now be based on an implicitly different test sampling scheme. It is no longer an utterance-based sampling and thus the inherent test statistics must be interpreted differently. We first address these design changes from the viewpoint of performance metric in Section \[sec:metric\]. We also address the problem of data imbalance typical of MVE based systems by pre-selecting a *cohort set* consisting of the most confusing impostor data. This balances the amount of target and impostor data and expedites the time required for MVE training. Performance Metrics {#sec:metric} ------------------- In traditional speaker verification (e.g., NIST SREs), the error counts are accumulated from the *utterance-level* decisions: a “miss” occurs when a legitimate talker is denied for the entire test utterance and a “false alarm” occurs when an impostor is incorrectly accepted for the utterance. The EER is defined as the rate at which the “miss” rate and the “false alarm rate” are equal [@nist_eer]. This utterance-based error counting is obviously not suitable for the AVA task because it bears the imperative assumption that the entire test speech signal is uttered by one and only one talker. It produces a single verification decision over the entire utterance without considering the possible change of speaker identity within the test signal. As noted above, AVA has to be prepared to detect a change of talker at any moment, and a user authentication error may occur at every test window slided over the signal continuously. For AVA, we evaluate *window-based* EER (WEER) because each real-time decision about the user identity is made on a test window anchored at that time instant. A window-based miss detection error (WMDE) occurs if a “reject” decision is made while the authorized talker is actually speaking within that window. A window-based false alarm error (WFAE) occurs if an “accept” decision is made while an impostor is speaking within that window. After all the window-based testings are performed, the WMDE rate and the WFAE rate can be evaluated against a chosen testing threshold. The WEER is reached when the threshold is chosen to make the two error rates equal. Obviously, calculation of the WMDE and WFAE rates is conditioned on the voice activity detector; when there is no speech, no decision is to be included. Note that WEER differs from the conventional utterance-based EER only in that the error counts are collected from window-level decisions instead of utterance-level ones. It becomes the traditional EER when each window of speech is treated as a separate test utterance. With WEER as the performance metric, training of the models in AVA must match the short-time testing condition, particularly when discriminative modeling methods are used. The purpose of discriminative model training is to minimize the empirical error rate. For an AVA system, such an empirical error rate is calculated from a combination of the WMDEs the WFAEs (See Eq. ). All these authentication errors are based on the window-based tokens. Therefore, the sample tokens for training and enrollment must each correspond to a segment of speech signal within a test window. This is one of the crucial differences in modeling for an AVA system and for a conventional utterance based authentication system. AVA Database and Pre-Processing {#sec:ava_data} =============================== Since AVA is a different task from the conventional speaker verification directed by NIST SRE, we collect a new voice database, which we call the AVA database, from 25 volunteers (14 females, 11 males) for performance evaluation. A Microsoft Surface Pro tablet with a built-in microphone was used to record the data and the sampling rate was set to 8000 samples/s. Each talker speaks at any position relative to the device as he or she feels comfortable; we consider this a natural use configuration of the device. The data collected from each person consists of four parts: the rainbow passage [@rainbow], a user-chosen pass-phrase, 20 randomly selected sentences from the phonetically balanced Harvard sentences [@harvard] (5.5 s on average) and 30 digit pairs (each digit is randomly selected from 0 to 9). The speaker repeats the same pass-phrase 8 times. In total, the recording amounts to 2.5 hours of voice signal from all talkers. For each speaker, we choose the enrollment data from the Rainbow passage, the pass-phrases and digits while the testing data is chosen from the Harvard sentences. The enrollment and test data sets do not overlap. The duration of each test set is configured to provide at least 1000 decisions per speaker in any given configuration. In all the experiments of this paper, the audio signal is converted to the conventional 39-dimension MFCC features with frame duration $T_f=25$ ms and $\delta_f=10$ ms. For the AVA task, the enrollment or test window moves forward 10 ms each time. Successive tests are performed with each shift over a segment of the specified durations. The durations of the enrollment and test windows are equal. The cepstral mean of speech frames within each enrollment and test window is subtracted to minimize the channel variability. AVA with I-Vector {#sec:ava_ivector} ================= I-vector analysis, a new front end factor analysis technique, is the predominant tool for conventional speaker verification. In this section, we investigate if this widely applied technique can achieve satisfactory performance for AVA. The i-vector is a projection of a speech utterance onto a low-dimensional total variability space that models both the speaker and the channel variability. More specifically, it is assumed that there exists a linear dependence between the speaker adapted (SA) GMM supervectors $\boldsymbol{\mu}$ and the SI GMM supervector $\boldsymbol{m}$ [@ivector_1]. $$\begin{aligned} \boldsymbol{\mu}=\boldsymbol{m}+\boldsymbol{U}\boldsymbol{w} \label{eqn:ivector_define} \end{aligned}$$ where $\boldsymbol{U}$ is a low rank factor loading matrix representing the primary direction of variability, and $\boldsymbol{w}$ is a random vector of total factors having a standard normal distribution $\mathcal{N}(\boldsymbol{0}; \boldsymbol{I})$. The i-vector is an MAP estimate of $\boldsymbol{w}$. We first apply i-vector to the conventional speaker verification task under the assumption that each test utterance is from only one speaker. We train a GMM UBM with all the enrollment data in the AVA database. With the EM algorithm, an SI factor loading matrix $\boldsymbol{U}_{SI}$ is trained on the statistics collected from the UBM. An i-vector is then extracted for each speaker using his or her enrollment data and $\boldsymbol{U}_{SI}$. During testing, an i-vector is extracted from the each test utterance using $\boldsymbol{U}_{SI}$. The i-vector dimension is fixed at 400. A cosine distance between the i-vector of each test utterance and that of the hypothesized speaker is used as the decision score. The EER is computed with all the utterance-level decision scores. In AVA database, the i-vector achieves 0.00% EER for the utterance-based speaker verification task under all UBM configurations. We then apply i-vector for the AVA task. We adopt the same training method as in the traditional speaker verification except that the training and enrollment tokens are generated by a sliding window with a prescribed duration. During testing, a test window of the same duration is slided over the test utterance at the rate of 100 per second and an i-vector is extracted from the speech signal within each test window using $\boldsymbol{U}_{SI}$. The cosine distance between the i-vector of each test window and that of the hypothesized speaker is used as the decision score. The AVA dataset described in Section \[sec:ava\_data\] is used for the performance evaluation. We fix the duration of enrollment data at an average of 240 s per speaker and randomly select two Harvard sentences for use as the testing data for each speaker. For AVA task, the window duration ranges from 1.01 s to 3.01 s. We show the WEER results with respect to the test window duration and the number of mixtures in the UBM in Table \[table:ivector\_ava\]. \[1\][&gt;[\ ]{}m[\#1]{}]{} \[1\][&gt;[\ ]{}m[\#1]{}]{} \[1\][&gt;[\ ]{}m[\#1]{}]{} \[table:ivector\_ava\] For each UBM configuration, the i-vector based AVA system achieves &lt;1% WEER when the duration of the test window is above 3 s. The performance degrades drastically as the test window duration falls below 2 s. When the test window is 1 s, the WEER rises to 12.91%. This performance trend is consistent with what have been reported in the literature and we conclude that it is not suitable for the AVA task where accurate decisions about speaker identity need to be made instantaneously. AVA Training and Registration {#sec:train} ============================= The AVA system consists of three parts: a training module, a registration module and an authentication module. In this section, we introduce the major components of the training and registration module which train and adapt the models to the enrollment data of each speaker. Fig. \[fig:ava\_train\] shows the training and registration stages of the AVA system. First, in the training stage, a SI ergodic HMM is trained on a sufficient pool of data from a general collection of speakers in the training set. The speech signal is converted to mel-frequency cepstral coefficients (MFCCs) [@mfcc] through the front-end processing component. The parameters of the SI HMM are initialized with the K-means clustering algorithm. The final SI HMM is obtained by applying the *Baum-Welch* re-estimation algorithm in the maximum likelihood (ML) training component. Then, in the registration stage, the model adaptation component adapts the SI model parameters to the voice of the target speaker upon receipt of the new registration data and generates the SA model based on the MAP adaptation technique [@map]. Then for the target data, an equivalent and most confusing set of data is selected from the impostor set by the cohort selection component for MVE training. Finally, the MVE training component generates the MVE trained target and anti-target model by directly minimizing a combination of the WMDEs and WFAEs. We elaborate the algorithms and procedure in Section \[subsec:train\_si\], \[subsec:train\_map\], \[subsec:train\_mve\] and \[subsec:train\_cohort\]. ![The components of the AVA training and registration stages.[]{data-label="fig:ava_train"}](ava_train_register.png){width="0.8\columnwidth" height="0.80\columnwidth"} Speaker-Independent (SI) Model Training {#subsec:train_si} --------------------------------------- In the training stage, an SI ergodic HMM, also called the UBM, for use as the seed model for later adaptation is trained. The parameters of this HMM are estimated by the *Baum-Welch* re-estimation algorithm after having been initialized with the K-means clustering algorithm. Let $\gamma_j (t)$ denote the occupation probability of being in state $j$ of the ergodic HMM at time $t$ which can be calculated efficiently using the *Forward-Backward* algorithm [@hmm]. The above-mentioned short-time requirement in sequential training implies that $\gamma_j(t)$ be accumulated differently from the conventional utterance-based training approach. This is because each test in AVA involves a voice segment within a test window of duration, $N_w \delta_f$, and this condition should be matched during training. Therefore, we modify the accumulation of $\gamma_j(t)$ as follows. Let us denote an entire training utterance by $X=\{\boldsymbol{x}_1,\ldots, \boldsymbol{x}_T\}$, where $T$ is the number of frames within the training utterance. $X_{t_a,t_b}=\{\boldsymbol{x}_{t_a},\ldots,\boldsymbol{x}_{t_b}\}$ is the speech segment extracted from $X$, where $t_a$ and $t_b$ are the start and end times, respectively, and $1\le t_a \le t_b \le T$. Each $X_{t_a,t_b}$ is used as a training token for *Baum-Welch* re-estimation as is elaborated above. Assume that the number of frames within each window is $N_w$ and $s_t$ is the state that frame $\boldsymbol{x}_t$ is aligned with at time $t$. In short-time sequential training, the occupation probability $\gamma^{short}_j(t)$ of being in state $j$ at time $t$ becomes $$\begin{aligned} \gamma^{short}_j(t)&=\frac{1}{N_w}\sum_{\tau=1}^{N_w} P(s_t=j|X_{t+\tau-N_w+1,t+\tau}), \\ &=\frac{1}{N_w}\sum_{\tau=1}^{N_w} \frac{P(X_{t+\tau-N_w+1,t+\tau}|s_t=j)P(s_t=j)}{P(X_{t+\tau-N_w+1,t+\tau})}. \label{eqn:occ_win}\end{aligned}$$ In Eq. , $\gamma^{short}_j(t)$ is calculated through the average likelihood of all the speech segments of window duration $N_w\delta_f$ which include the frame $\boldsymbol{x}_t$. For the conventional utterance-based training, the state occupation is $$\begin{aligned} \gamma^{conv}_j(t)&=P(s_t=j|X_{1,T}), \\ &=\frac{P(X_{1,T}|s_t=j)P(s_t=j)}{P(X_{1,T})}. \label{eqn:occ_utt}\end{aligned}$$ In Eq.  $\gamma^{conv}_j(t)$ is computed through the likelihood of the entire utterance $X$, which is much longer than the window duration. Each term of the summation in Eq  falls back to the conventional utterance-based state occupation probability in Eq  when $N_w=T$ and $\tau=T-t$ because the window covers the duration of the entire utterance. In other words, $\gamma^{short}_j(t)$ is affected by the statistics of the window-duration speech segment which contains frame $\boldsymbol{x}_t$ while $\gamma^{conv}_j(t)$ is affected by the entire training utterance even when only a small portion of the utterance is correlated with $\boldsymbol{x}_t$ statistics. After obtaining the occupation probability, we update the HMM parameters with the average of the window-wise sufficient statistics weighted by $\gamma^{short}(t)$ in a standard way. Model Adaptation {#subsec:train_map} ---------------- When a registration procedure is initiated, the SI model is assumed to have been well-trained as described in Section \[subsec:train\_si\]. As the first step of registration, the model adaptation component in Fig. \[fig:ava\_train\] adapts the SI model to the new registration data of the authorized target speaker using MAP estimation. Assuming speech segment $X=\{\boldsymbol{x}_1,\ldots,\boldsymbol{x}_T\}$ within a sliding window from a registered user to be a training token for MAP adaptation, the likelihood of $X$ given HMM with $J$ states and parameter $\lambda=\{{\pi_j},a_{ij},\theta_j\}_{i,j=1}^J$ is $$p(X|\lambda)=\sum_{\boldsymbol{s}}\pi_{s_0}\prod_{t=1}^Ta_{s_{t-1}s_{t}}\sum_{m=1}^Kw_{s_tm}\mathcal{N}(\boldsymbol{x}_t|\boldsymbol{\mu}_{s_tm},\boldsymbol{\Sigma}_{s_tm}) % b_{s_t}(\boldsymbol{x}_t) \label{eqn:likelihood_hmm}$$ where $s=\{s_1,\ldots,s_T\}$ is the unobserved state sequence, $\pi_j$ is the initial probability of state $j$, $a_{ij}$ is the transition probability from state $i$ to state $j$, $\theta_j=\{w_{jk},\boldsymbol{\mu}_{jk},\boldsymbol{\Sigma}_{jk}\}$, $k=1,\ldots,K$, where $w_{jk}$, $\boldsymbol{\mu}_{jk}$, $\boldsymbol{\Sigma}_{jk}$ are the weight, mean vector and covariance matrix, respectively, for the $k$ th component of the Gaussian mixture which is the probability output of state $j$. The MAP estimate $\theta_{MAP}$ is aimed at maximizing the posterior probability denoted as $f(\lambda|X)$, i.e., $$\theta_{MAP}=\arg\max_\lambda f(\lambda|X)=\arg\max_\lambda p(X|\lambda)g(\lambda) \label{eqn:map}$$ where $g(\lambda)$ is the prior distribution of $\lambda$. The MAP estimate is obtained as follows. The probability of being in state $j$ at time $t$ with the $k$th mixture component accounting for $\boldsymbol{x}_t$ is $$\gamma_{j,k}(t)=\gamma_j(t)\frac{w_{jk}\mathcal{N}(\boldsymbol{x}_t|\boldsymbol{\mu}_{jk},\boldsymbol{\Sigma}_{jk})}{\sum_{m=1}^Kw_{jm}\mathcal{N}(\boldsymbol{x}_t|\boldsymbol{\mu}_{jm},\boldsymbol{\Sigma}_{jm})} \label{eqn:gamma}$$ For mixture $k$ in state $j$, the occupation likelihood and the 1st and 2nd moment of the observed adaptation data can be estimated by $$\begin{aligned} &n_{jk}=\sum_{t=1}^{T}\gamma_{j,k}(t), \quad E[{\boldsymbol{x}_t}]=\frac{1}{n_{jk}}\sum_{t=1}^T\gamma_{j,k}(t)\boldsymbol{x}_t \\ &E[{\boldsymbol{x}_t}{\boldsymbol{x}_t}^\top]=\frac{1}{n_{jk}}\sum_{t=1}^T\gamma_{j,k}(t)\boldsymbol{x}_t\boldsymbol{x}_t^\top \label{eqn:mom}\end{aligned}$$ Thus, the MAP update formula for mixture $k$ in state $j$ of an HMM is $$\begin{aligned} &\hat{w}_{jk}=\alpha_{jk}^w\frac{n_{jk}}{T}+(1-\alpha_{jk}^w)\bar{w}_{jk} \\ &\hat{\boldsymbol{\mu}}_{jk}=\alpha_{jk}^m E[\boldsymbol{x}_t]+(1-\alpha_{jk}^m)\bar{\boldsymbol{\mu}}_{jk} \\ &\hat{\boldsymbol{\Sigma}}_{jk}=\alpha_{jk}^v E[\boldsymbol{x}_t\boldsymbol{x}_t^\top]+(1-\alpha_{jk}^v)(\bar{\boldsymbol{\Sigma}}_{jk}+\bar{\boldsymbol{\mu}}_{jk}^2)-\hat{\boldsymbol{\mu}}_{jk}^2 \label{eqn:update}\end{aligned}$$ where $\{\bar{w}_{jk},\bar{\boldsymbol{\mu}}_{jk},\bar{\boldsymbol{\Sigma}}_{jk}\}$, $k=1,\ldots,K$, $j=1,\ldots,J$ are the mixture parameters of the SI HMM. The adaption coefficient $\alpha_{jk}^\rho,\rho \in\{w,m,v\}$ is defined for each mixture component in each state as $\alpha_{jk}^{\rho}=n_{jk}/(n_{jk}+\eta^\rho)$ where $\eta^\rho$ accounts for the weight of prior knowledge for $\rho$. The MAP adaptation is performed during sequential training on each training token $X$. As is explained in Section \[subsec:train\_si\], the MAP adapted model also characterizes the short-time statistics of the registration speech data since the statistic $\gamma_j(t)$ in Eq.  is accumulated through the likelihoods of the adaptation speech segments which have the same duration as the test window. MVE Training {#subsec:train_mve} ------------ As registration stage, the MVE training is performed after the speaker model adaptation. The SA HMM and the SI HMM serve as the initial target and initial anti-target model, respectively, for the MVE training. In the MVE training component in Fig. \[fig:ava\_train\], all parameters are optimized with the enrollment and training data, according to the criterion to minimize the total number of authentication errors (which is the total number of WMDEs and WFAEs) on the corpus. Let us define the enrollment data of the target speaker as the target set $D_0$, and define the training data excluding the speech of the target speaker as the impostor set $D_1$. For a window-duration MVE training token $X_n$ from either $D_0$ or $D_1$, $g(X_n|\lambda_0)$ and $g(X_n|\lambda_1)$ denote the log-likelihoods of $X_n$ given the target HMM with parameters $\lambda_0$ and the anti-target HMM with parameters $\lambda_1$, respectively. The log-likelihoods are calculated by aligning $X_n$ against the states of the target and the anti-target models using the *Viterbi* algorithm and are normalized with respect to the total number of frames $T$ within the utterance $X_n$. As the training tokens are generated by sliding a window of size $N_w$ frames every $\delta_f$ duration on the training utterance, the likelihood can be calculated more efficiently by modifying the *Viterbi* algorithm. Instead of resetting the trellis and initializing it anew each time we evaluate the log-likelihood of a new window of voice segment within the same utterance, we reset the trellis only at the beginning of a training utterance and let the trellis grow until the end of the utterance. The log-likelihood of an incoming window of voice segment is accumulated directly from the part of the fully grown trellis which starts from the very beginning of the utterance. This new implementation is equivalent to performing a partial traceback of the trellis structure within each sliding window so that the consistency is maintained in training and testing based on short window of data. The partial traceback also speeds up the MVE training procedure by a factor of $N_w$. To count the verification errors based on the log-likelihood of the tokens, the *misverification* measure is further defined for each class $$\begin{aligned} d_0(X_n|\lambda_0,\lambda_1)&=-g(X_n|\lambda_0)+g(X_n|\lambda_1), \quad \text{if } X_n \in D_0, \label{eqn:misverification_1} \\ d_1(X_n|\lambda_0,\lambda_1)&=g(X_n|\lambda_0)-g(X_n|\lambda_1), \quad \text{if } X_n \in D_1. \label{eqn:misverification_2}\end{aligned}$$ The two types of verification errors, WMDE and WFAE, can be approximated as $l_0$ and $l_1$, respectively, by embedding the two misverification measures into smooth loss functions below $$\begin{aligned} l_0(X_n|\lambda_0,\lambda_1)&=\frac{A_0}{1+\exp\left[-d_0(X_n|\lambda_0,\lambda_1)\right]}, \quad \text{if } X_n \in D_0, \label{eqn:loss_function_1}\\ l_1(X_n|\lambda_0,\lambda_1)&=\frac{A_1}{1+\exp\left[-d_1(X_n|\lambda_0,\lambda_1)\right]}, \quad \text{if } X_n \in D_0, \label{eqn:loss_function_2}\end{aligned}$$ where the weights $A_0$ and $A_1$ emphasize the respective error types. Finally, we obtain the MVE loss as an approximation of the total number of verification errors on the training and enrollment corpus as follows. $$L(\lambda_0,\lambda_1)=\sum_{X_n \in D_0} l_0(X_n |\lambda_0,\lambda_1)+\sum_{X_n \in D_1} l_1(X_n |\lambda_0,\lambda_1) \label{eqn:average_cost}$$ In Eq. , the total number of verification errors are expressed as a continuous and differentiable function of the model parameters, and hence, can be minimized with respect to all parameters by using the generalized probabilistic descent (GPD) algorithm [@gpd]. In the short-time sequential training framework, each $X_n$ is a speech segment sequentially extracted from the training utterance via a sliding window. The likelihoods of the speech segments are computed with the same duration as the test window and then utilized to calculate the gradient and the steps size of GPD update at each iteration. Therefore, a speaker model accurately matched to the testing condition is estimated through short-time sequential training. Cohort Selection {#subsec:train_cohort} ---------------- From Eqs. , ,  and , we notice that an MVE training frame from the target speaker updates the model such that the WMDE decreases and a training token from the impostor speaker updates the model such that the WFAE decreases. The impostor speaker can be any speaker other than the target speaker and the number of data tokens from the impostors in the database will obviously be greater than the number of tokens of the target speaker. Since we choose WEER as the performance indicator, we need a balanced data set of target and impostor speech. Therefore, to maintain a balance in the amount of target and impostor data without sacrificing the discriminability between the two, we pick from the impostor set $D_1$ a most confusing set of data for use in MVE training for each target speaker. This selected set is called the *cohort set*. In the cohort selection component of Fig. \[fig:ava\_train\], a screening test is run with the MAP adapted target model and the SI model to select possible cohort impostor set for MVE training. A log-likelihood ratio (LLR) is computed for each window-duration segment of the impostor data $D_1$ using Eq. , where the log-likelihoods are calculated efficiently through the modified *Viterbi* algorithm described in Section \[subsec:train\_mve\]. We further rank the segments by their LLRs in descending order and pick the top $r$ speech segments as the cohort set for subsequent MVE training, where $r$ is the number of segments in the target dataset. This does not affect the overall performance as the speech segments with lower LLRs naturally contribute less to the gradient in GPD optimization (see Eqs. , , ,  and ). With cohort selection, $D_1$ in Eq.  becomes the cohort set and the following MVE training is performed in the same way as described in Section \[subsec:train\_mve\]. Cohort selection proves useful in reducing the time needed to train the speaker models with MVE as it reduces the data that the MVE algorithm needs to process. Short-Time Sequential Testing {#sec:test} ============================= In the authentication stage of AVA, the system performs sequential testing (note: the sequential testing here is to be differentiated from the Wald’s sequential test [@wald]) and makes decisions in real-time. The testing needs both the target and anti-target models for each registered speaker that are obtained after MVE training. During operation, the sequential testing procedure continuously takes a sliding window of speech frames, accumulates the log-likelihood with respect to both target and anti-model for the speaker, and then reports the LLR confidence scores periodically to the system. Fig. \[fig:ava\_monitor\] shows the block diagram for short-time sequential testing. ![Diagram of the AVA authentication module.[]{data-label="fig:ava_monitor"}](ava_monitor.png){width="0.7\columnwidth" height="0.4\columnwidth"} The LLR scores are calculated using the following equations $$\Gamma(X)=\log p(X|\lambda_0)-\log p(X|\lambda_1) \label{eqn:llr}$$ where $X=\{\boldsymbol{x}_1,\ldots,\boldsymbol{x}_T\}$ is a window of voice frames. $\lambda_0$ and $\lambda_1$ are the parameters of the target and anti-target models defined in \[eqn:np\_log\], respectively. The likelihoods $p(X|\lambda_0)$ and $p(X|\lambda_1)$ are computed using Eq. (\[eqn:likelihood\_hmm\]) through the modified *Viterbi* algorithm described in Section \[subsec:train\_mve\]. As discussed, a speech signal inevitably contains silence gaps. These silence gaps do not contain any voice biometric information and need to be excluded from testing. We use a voice activity detector (VAD) to modulate the WEER results by ignoring the test scores from silent frames. We use the VAD algorithm suggested in the European Telecommunications Standards Institute (ETSI) Distributed Speech Recognition front-end [@etsi]. The VAD makes a binary voice/silence decision for every frame. Each VAD decision is made based on the average log mel energy of its 80 neighboring frames. We ignore the speaker authentication decision for a given testing window if the corresponding anchor frame (the frame at the middle of window) is silent according to the VAD. Experiments {#sec:expr} =========== For the performance evaluation in Sections \[subsec:expr\_win\_dur\], \[subsec:expr\_model\_config\], and \[subsec:expr\_data\_dur\], we use the enrollment and test data in AVA dataset as is described in Section \[sec:ava\_data\]. Performance with test window duration {#subsec:expr_win_dur} ------------------------------------- Our first investigation focused on the trade-offs between the duration of the test window and the authentication performance. The duration of test window directly affects the system delay and the real-time requirement. We fix the duration of enrollment data at an average of 240 s per speaker, but vary the duration of the test speech segment from $N_w\delta_f=0.1$ s to $5.01$ s corresponding to $N_w=1,\ldots, 501$ frames. We select part of the Harvard sentence set for use as the testing data for each speaker. Two Harvard sentences are randomly selected for each speaker for the window durations from 0.1 s to 1.01 s, while 4 Harvard sentences are selected for 2 s and 5 s testing windows. Fig. \[fig:win\_dur\] shows the baseline WEER for each window duration. We note that the WEER based on just 0.1 s of test data is quite poor at approximately 24%, but it improves as we increase the duration of the data for each decision epoch. As expected, the WEER performance monotonically decreases with the duration of the test data. Furthermore, a knee point can be observed at around 1 s, which can serve as a designing parameter to meet the real time requirement. We also note that the WEER with 5 s of sequential test data is 0.78%. This is quite low and in need of careful considerations as it may incorrectly imply that the system would perform flawlessly if the test window is sufficiently large. In our evaluation, we perform a test evaluation on each successive window of data which comes from the same utterance. For example, a 10 s long utterance will produce nearly 1000 test decisions; in the conventional utterance-based evaluation, it would have been just one test decision. This gives rise to the issue of statistical significance in the error probability estimate for the talker verification performance. It is fair to note that in the conventional utterance-based evaluation, the evaluation sample size tends to be rather limited, which weakens the statistical significance of the test result, while in the new window-based evaluation, the test sample size makes the error probability estimate more statistically trustworthy but it contains a sampling bias as the tests are performed on successive data windows that are not independent. This contrastive consideration, while interesting, does not affect the determination of the trade-off here. We thus choose 1 s as the nominal duration of the test data window, used in subsequent evaluations. Performance as a function of model configuration {#subsec:expr_model_config} ------------------------------------------------ We fix the duration of the enrollment data at an average of 240 s per speaker and randomly select 2 or 4 Harvard sentences for each speaker as the test set when the test window duration is 1.01 s or 2.01 s. With 25 speakers in total, more than 25,000 or 35,000 trials are generated from the 50 Harvard sentences or 100 Harvard sentences by sliding the test window. In Table \[table:w1\_t240\], we provide the performance evaluations for the average WEER (after VAD modulation) with a 1.01 second decision window duration over the various model configurations and two algorithms, MAP and MVE (the number in bold means the best performance in the column). We notice that the average WEER for MAP adapted models is 4.10% while MVE training decreases the absolute WEER to 3.00%, on average. Depending on the complexity of the model and the algorithm, the WEER ranges between 2.6-4.5%. The models with one state represented by 1024 Gaussian mixtures achieves the best performance. Table \[table:w2\_t240\] shows the evaluation results with VAD for different configurations with the decision window durations set at 2.01 seconds. \[1\][&gt;[\ ]{}m[\#1]{}]{} \[1\][&gt;[\ ]{}m[\#1]{}]{} \[1\][&gt;[\ ]{}m[\#1]{}]{} \[table:w1\_t240\] \[1\][&gt;[\ ]{}m[\#1]{}]{} \[1\][&gt;[\ ]{}m[\#1]{}]{} \[1\][&gt;[\ ]{}m[\#1]{}]{} \[table:w2\_t240\] Performance with enrollment data duration {#subsec:expr_data_dur} ----------------------------------------- It is desirable to use as little enrollment data as possible while maintaining a similar performance as in Table \[table:w1\_t240\]. In the following, we evaluate our system on the minimum amount of enrollment voice data necessary to achieve an acceptable performance which bears directly on the time it takes for a talker to register for AVA for the first time. In Table \[table:w1\_t105\] and \[table:w1\_t180\], we use 180 and 105 seconds of enrollment voice data, respectively. Two Harvard sentences from each speaker are selected to form the test set. With 25 speakers in total, more than 25,000 trials are generated from the 50 Harvard sentences by sliding a test window with a duration of 1.01 s. When we compare the results in Table \[table:w1\_t180\] with the results in Table \[table:w1\_t240\], where the average enrollment data duration is 240 seconds, we notice that using 25% less enrollment data reduces the MVE performance to an average WEER of 4.64% from 3.00%. Similarly, in Table \[table:w1\_t105\] we are using 56% less enrollment voice data than in Table \[table:w1\_t240\] which reduces the average WEER performance with MVE to 6.31%. In the case of reduced enrollment voice data, the degradations may still be acceptable as the WEER stays in the vicinity of 5-6%. \[1\][&gt;[\ ]{}m[\#1]{}]{} \[1\][&gt;[\ ]{}m[\#1]{}]{} \[1\][&gt;[\ ]{}m[\#1]{}]{} \[table:w1\_t180\] \[1\][&gt;[\ ]{}m[\#1]{}]{} \[1\][&gt;[\ ]{}m[\#1]{}]{} \[1\][&gt;[\ ]{}m[\#1]{}]{} \[table:w1\_t105\] Performance of conventional speaker verification using AVA real-time decisions {#subsec:expr_sv} ------------------------------------------------------------------------------ In the scenario of conventional speaker verification, each test utterance is assumed to include the speech of only speaker with a claimed identity. A decision is made on the speaker identity by comparing a threshold with the log-likelihood score of the entire test utterance given the claimed speaker model, while in the case of AVA, a decision needs to be made on each short-duration test window in real-time because the test utterance may undergo change of speaker at any time instant. Here, we are interested in exploring the performance of the window-based modeling and testing scheme of AVA in a conventional utterance-based speaker verification task. Under the assumption that each utterance is spoken by a single talker, we form the verification decision for the entire utterance by instituting a majority vote from the AVA short-time window-based decision sequence. Specifically, for a test utterance of duration $T$, if more than $\lfloor T/\delta_f\rfloor / 2$ of the decisions for the window-based tests are “true speaker (impostor)”, the final decision for this utterance will be “true speaker (impostor)”. With the utterance-level ground truth, EER can be computed in the way as described in Section \[sec:metric\] by varying the threshold. With the AVA real-time decisions in Section \[subsec:expr\_model\_config\], an EER of 0.00% is achieved under all HMM configurations for the utterance-based conventional speaker verification task using the AVA database. Statistical validation of performance {#subsec:expr_kfold} ------------------------------------- Cross validation is an effective statistical method to test the generalizability of a model to new or unseen data [@kfold]. In the text-independent real-time speaker verification task, in which the data is randomly divided into roughly equal K subsets and for each validation trial, one of the K subsets is used as the testing set while the rest of the K-1 subsets are put together to form a enrollment set. The enrollment set is fit to a model and predictions are made on the testing data based on the trained model. In each round, the validation is repeated K times and results are then averaged over the K validation trials or folds. To ascertain the statistical significance of the obtained performance, we use K-fold cross validation to systematically check the accuracy of the speaker models for unseen data. We set K=3 to keep the duration of the enrollment set for each speaker to 240 seconds, on average, which makes the results comparable with the evaluations in Section \[subsec:expr\_model\_config\]. We run 5 rounds of the 3-fold cross validation and average over the results for the folds and rounds. The results are shown in Table \[table:kfold\], in which we give the average WEER of the rounds and the 95% confidence range (CR) to indicate the variation of the WEER. We notice that the CR for the MAP algorithm are smaller than for MVE algorithm. Furthermore, when comparing the results in Table \[table:kfold\] with the results in Table \[table:w1\_t240\], we notice that the WEER is less than that achieved for the same model. This is because the enrollment and testing datasets, in contrast to the evaluations done in the preceding section, are more matched in terms of the material, despite being selected randomly. As is mentioned in Section \[sec:expr\], the dataset for each speaker consists of four parts: rainbow passage, 8 repeated pass-phrases, Harvard sentences and digit pairs. For 3-fold validation, we randomly select roughly a third of the utterances from each part, combine them to be the testing set and use the rest two-third as the enrollment set. This means that some of the repeated pass-phrase utterances will be shared between the enrollment and test sets. A similar sharing may occur for the digit pairs. Thus, a lower WEER is obtained in this case. \[1\][&gt;[\ ]{}m[\#1]{}]{} \[1\][&gt;[\ ]{}m[\#1]{}]{} \[1\][&gt;[\ ]{}m[\#1]{}]{} \[table:kfold\] Performance evaluation on NIST SRE ---------------------------------- The NIST SRE Training and Test Sets are widely used to evaluate the performance of speaker verification systems. In NIST SRE, the decisions are made on each test utterance based on the speaker models trained with the provided training data and the performance is evaluated with respect to the ground truth. We notice that cross-talk components exist in these datasets, i.e., even though a test utterance is labeled as coming from a certain speaker in the label, some portion of the utterance is actually from another speaker (e.g., see 2001 NIST SRE). Although these crosstalk components may not substantially affect the performance evaluation designed for NIST’s utterance-based authentication, it does not suit the evaluation of the real-time, window-based AVA system as the real identity of each sliding window of the speech signal is not known. To verify the effectiveness our method on large and publicly available datasets, we sift out the cross-talk components in NIST SRE 2001 dataset [@nist01_eval_plan] and evaluate the performance of AVA using both i-vector and the proposed method with the remaining speech signal. In NIST SRE 2001, there are 100 female target speakers and 74 male target speakers with around 2 minutes of enrollment data for each speaker. In addition, there are 2038 test segments, each of which has a duration varying between 15 to 45 seconds. Each test segment will be evaluated against 11 hypothesized speakers of the same gender as the target speaker. One of the 11 hypothesized speakers is the true speaker present in the test segment and rest of them are impostors. Since the cross-talk components in NIST SRE 2001 has significantly lower energy per frame than the speech signal from the target speaker, a decision of “cross-talk” is made for a speech frame if its log mel energy is below a certain threshold. The decision for each window of frames is then made by taking the consensus of the threshold decisions within that window. The windows labeled as “cross-talk” are eliminated in both the enrollment and test utterances and ignored in the experiments. After sifting, we kept about 75% of windows in the enrollment data and 95% of windows in the test data. We first evaluated the i-vector technique in both the utterance-based conventional speaker verification task and the AVA task in exactly the same way as described in Section \[sec:ava\_ivector\]. The WEER results with respect to the number of mixtures are listed below. \[1\][&gt;[\ ]{}m[\#1]{}]{} \[1\][&gt;[\ ]{}m[\#1]{}]{} \[1\][&gt;[\ ]{}m[\#1]{}]{} \[table:ivector\_nist01\] We then apply the proposed training and testing method described in Sections \[sec:train\] and \[sec:test\] to the AVA task on NIST SRE 2001 and obtain the WEER results below. We also compare the performance difference between MAP and MVE training. \[1\][&gt;[\ ]{}m[\#1]{}]{} \[1\][&gt;[\ ]{}m[\#1]{}]{} \[1\][&gt;[\ ]{}m[\#1]{}]{} \[table:map\_mve\_nist01\] By comparing Tables \[table:ivector\_nist01\] and \[table:map\_mve\_nist01\], we see that the proposed method achieves 3.88% absolute gain over i-vector for the AVA task when the window duration is 1.01s. For the conditional speaker verification, our i-vector based system achieves 6.67% EER, which is 1.61 % absolutely better than the UBM-GMM baseline EER 8.28 % reported in [@nist01_baseline] on NIST 2001 SRE. As a cross reference, for conventional speaker verification task, the i-vector achieves 6.02%-7.07% and 4.77%-5.15% EERs for the male and female parts, respectively, of telephone data in the core condition of NIST SRE 2008 [@nist08_telephone] (the condition most similar to NIST SRE 2001); it achieves an EER of 22.01% on NIST SRE 2008 core condition when the test utterances are truncated to 2 seconds [@ivector_short_2012]. The EER performance gain over i-vector on NIST SRE justifies the generalizability of the proposed method to the standard public datasets with large amount of speakers for the AVA task. Conclusions {#sec:conclude} =========== We present an ensemble of techniques that enable the active voice authentication. The difference between AVA and traditional speaker verification is significant: AVA makes a decision on the speaker identity at every time instant while the latter task makes a one-time decision on the speaker identity after the entire test utterance is obtained. Therefore, the major challenges for the AVA task is to train accurate speaker models using minimal amount of data for active and continuous identity authentication with very short test signals. We first show that the i-vector technique is not suitable for the AVA task since the performance degrades sharply as the duration of the test segment becomes extremely short. In our AVA system, these challenging requirements are satisfied by matching the training and testing condition, adapting the SI model to the data of each individual speaker using MAP and MVE training. We perform sequential testing with the MVE trained model. In our offline evaluation of the system on the database we recorded, the system achieves 3-4% average WEER when the testing window duration is just 1 second, which is far beyond human capabilities. Statistical validation is conducted via K-fold cross validation. From the experimental results, the WEER performance does not change too much when the total number of mixtures goes beyond 512. We use the model configuration with 1 state and 512 mixtures as it provides an acceptable trade-off between the training time of the algorithm and the WEER performance. We show that the proposed approach can be generalized to a standard public databases with large amount of speakers by showing that the proposed methods outperforms i-vector approach by 3.8% absolute on NIST SRE 2001. We decided to use about 180 seconds of voice data to train the model for a new user. We consider this amount of enrollment data to be acceptable without inducing the fatigue factor on the part of the user. It gives a good WEER performance at 4-5%. More enrollment data will further reduce the WEER although at the expense of the registering user’s time. MVE provides an average performance improvement from 0.5-1% but requires much more training time. The total time for user registration is 6-10 minutes. We select the 1.01 s window duration for the demonstration system as it provides a reasonable WEER of approximately 5% for the configuration of the HMM model and the decision latency remains within the acceptable near real-time requirement. Using a longer duration window provides a better performance but produces noticeable delays in decision. The current version of the voice authentication system assumes operation in a low-noise environment, which means its performance will be best in an indoor office locale. The model that the system builds for the user can account for some minor environmental variations but it does not take voice variability of a speaker, e.g., the Lombard effects, into account. A change in the audio path, e.g., if an external microphone is used, may require rebuilding the user model and/or the SI model. More research is needed to improve these aspects of the demonstration system. Currently, prevalent automatic voice assistants such as Google Home and Amazon Alexa are equipped with an authentication system at the front end to verify users’ identities based on their pronunciation of a fixed wake word. In the future, we will explore the optimal way that an AVA can work together with the speaker verification system after a user has obtained his/her access. We will also investigate the methods to combine the AVA and the initial one-time speaker verification scores to provide more reliable continuous monitoring of real-time identities. Acknowledgement {#acknowledgement .unnumbered} =============== The authors would like to thank Chao Weng and Antonio Moreno-Daniel for their help on the AVA system. This work was supported by Defense Advanced Research Projects Agency (DARPA) under the Active Authentication program with Li Creative Technologies as the primary contractor and Georgia Tech the sole subcontractor. The term “active voice authentication” (AVA) was used by Li Creative Technologies in the title of the DARPA supported project. References {#references .unnumbered} ========== Short Biographies {#short-biographies .unnumbered} ================= **Zhong Meng** received his Ph.D. degree from Georgia Institute of Technology in 2018. He is currently an Applied Scientist in Speech & Language Group at Microsoft Corporation. Before joining Microsoft, he worked at AT&T Labs Research, Mitsubishi Electric Research Labs and Microsoft Research as a Research Intern. He has been working on domain adaptation, end-to-end speech recognition, adversarial learning for robust speech recognition, speech enhancement and speaker verification, discriminative training, and multi-channel speech recognition. His adversarial domain adaptation work was nominated for the Best Paper Award at IEEE Automatic Speech Recognition and Understanding (ASRU) Workshop in 2017. **M. Umair Bin Altaf** is a Senior Research Scientist at Pindrop Inc. He received a Ph.D. degree in Electrical and Computer Engineering from Georgia Institute of Technology, Atlanta, GA, USA in 2015. He has been with Avaya Labs and Broadcom Inc. as a Research Intern, working on robust speech recognition and speech coding. His research interests lie in the general area of audio signal processing with special focus on speech recognition, machine learning, and environmental sound representation. **Biing-Hwang (Fred) Juang** is the Motorola Foundation Chair Professor and a Georgia Research Alliance Eminent Scholar at Georgia Institute of Technology. He received a Ph.D. degree from University of California, Santa Barbara and had conducted research at Bell Labs, including serving as Director of Acoustics and Speech Research. He joined Georgia Tech in 2002. Prof. Juang’s academic distinctions include: several best paper awards, IEEE Fellow, the Technical Achievement Award from the IEEE Signal Processing Society for contributions to the field of speech processing and communications, Bell Labs Fellow, member of the US National Academy of Engineering, Academician of Academia Sinica, the IEEE J.L. Flanagan Field Medal Award in Audio, Speech and Acoustics, and a Charter Fellow of the National Academy of Inventors.

ArXiv

--- abstract: 'In a previous work by the authors the one dimensional (doubling) renormalization operator was extended to the case of quasi-periodically forced one dimensional maps. The theory was used to explain different self-similarity and universality observed numerically in the parameter space of the Forced Logistic Maps. The extension proposed was not complete in the sense that we assumed a total of four conjectures to be true. In this paper we present numerical support for these conjectures. We also discuss the applicability of this theory to the Forced Logistic Map.' author: - Pau Rabassa - Angel Jorba - Joan Carles Tatjer title: ' Towards a renormalization theory for quasi-periodically forced one dimensional maps [III]{}. [N]{}umerical Support[^1] ' --- Introduction ============ This is the third of a series of papers (together with [@JRT11a] and [@JRT11b]) proposing an extension of the one dimensional renormalization theory for the case of quasi-periodically forced one dimensional maps. These three papers are closely related but each of them has been written to be readable independently. See also [@Rab10] for a more detailed discussion. In the previous two papers we were concerned with the theoretical part of the theory. In this paper we include different numerical computations which support the conjectures introduced for the developing of this theory. To do that we briefly review the theory developed in the previous two papers, skipping technical details and we adding some numerical computations to the discussion. The universality and self-renormalization properties in the cascade of period doubling bifurcations of families of unimodal maps is a well known phenomenon. The paradigmatic example is the Logistic map $l_\alpha(x) = \alpha x(1-x)$. Given a typical one parametric family of unimodal maps $\{f_\alpha\}_{\alpha\in I}$ one observes numerically that there exists a sequence of parameter values $\{d_n\}_{n\in {\mathbb{N}}} \subset I$ such that the attracting periodic orbit of the map undergoes a period doubling bifurcation. Between one period doubling and the next one there exists also a parameter value $s_n$, for which the critical point of $f_{s_n}$ is a periodic orbit with period $2^n$. One can observe numerically that $$\label{universal limit sumicon} \lim_{n\rightarrow \infty} \frac{d_n - d_{n-1}} {d_{n+1} - d_{n}} = \lim_{n\rightarrow \infty} \frac{s_n - s_{n-1}} {s_{n+1} - s_{n}} = {\text{\boldmath $\delta$}}= \texttt{ 4.66920...}.$$ This convergence indicates a self-similarity on the parameter space of the family. On the other hand, the constant ${\text{\boldmath $\delta$}}$ is universal, in the sense that for any family of unimodal maps with a quadratic turning point having a cascade of period doubling bifurcations, one obtains the same ratio ${\text{\boldmath $\delta$}}$. To explain these phenomena Collet and Treser ([@CT78]) and Feigenbaum ([@Fei78; @Fei79]) proposed simultaneously the renormalization operator. Their explanation was based on the existence of a hyperbolic fixed point of the operator with suitable properties. The first proof of the existence of this point and its hyperbolicity were obtained with numerical assistance [@Lan82; @EW87]. A decade later, Sullivan (see [@Sul92]) generalized the operator and provided a theoretical proof of the hyperbolicity using complex dynamics techniques. See [@Lyu99; @MvS93] for extensive summaries on the theory. In [@JRT11p2] we presented numerical evidences of self similarity and universality for families of quasi-periodically forced Logistic maps. These are maps in the cylinder where the dynamics on the periodic variable are given by a rigid rotation and the dynamics on the other variable are given by the Logistic Map plus a small perturbation which depends on both variables. This kind of maps have its origins in studies related to the existence of strange non-chaotic attractors (see [@Bje09; @FKP06; @FH10; @HH94; @Jag03; @PMR98]). Let us describe these numerical evidences with more detail. Numerical evidence of self-similarity and universality for quasi-periodic forced maps {#subsection intro numerical evidences} ------------------------------------------------------------------------------------- Consider $\{g_{\alpha,{\varepsilon}}\}_{(\alpha,{\varepsilon}) \in J\subset{\mathbb{R}}^2}$ a two parametric family of quasi-periodic maps in the cylinder ${\mathbb{T}}\times {\mathbb{R}}$ of the form $$\label{q.p- family} \left. \begin{array}{rcl} \bar{\theta} & = & \theta + \omega ,\\ \bar{x} & = & \alpha x(1-x) + {\varepsilon}h(\theta,x) , \end{array} \right\}$$ with $\omega$ a Diophantine number, $\alpha$ and ${\varepsilon}$ parameters and $h$ a periodic function with respect to $\theta$ which can also depend on $\alpha$ and ${\varepsilon}$. Recall that the Logistic map $\bar{x}=\alpha x(1-x)$ has a complete cascade of period doubling bifurcations. As before, let $\{d_n\}_{n\in {\mathbb{N}}} \subset I$ denote the parameter values where the attracting periodic orbit undergoes a period doubling bifurcation and $\{s_n\}_{n\in {\mathbb{N}}} \subset I$ the values for which the critical point of $f_{s_n}$ is a periodic orbit with period $2^n$. In [@JRT11p] we computed some bifurcation diagrams in terms of the dynamics of the attracting set. We have taken into account different properties of the attracting set, as the Lyapunov exponent and, in the case of having a periodic invariant curve, its period and its reducibility. The reducibility loss of an invariant curve is not a bifurcation in the classical sense that the attracting set of the map changes dramatically, only the spectral properties of the transfer operator associated to the continuation of that curve does (see [@JT08]). Despite of this, it can be characterized as a bifurcation (see definition 2.3 in [@JRT11p]). The numerical computations in [@JRT11p] reveal that the parameter values for which the invariant curve doubles its period are contained in regions of the parameter space where the invariant curve is reducible. These computations also reveal that from every parameter value $(\alpha,{\varepsilon}) = (s_n, 0)$ two curves of reducibility loss (of the $2^n$-periodic invariant curve) are born. This situation is sketched in the left panel of figure \[Esquema BifPar\]. Assume that these two curves can be locally expressed as $(s_n + \alpha_n'(\omega){\varepsilon}+ O({\varepsilon}^2),{\varepsilon})$ and $(s_n + \beta_n'(\omega) {\varepsilon}+ O({\varepsilon}^2),{\varepsilon})$. In [@JRT11a] we proved that these curves really exist under suitable hypothesis. We have also given explicit expressions of the slopes $\alpha_n'(\omega)$ and $\beta_n'(\omega)$ in terms of the quasi-periodic renormalization operator (introduced there). We only focus on $\alpha_n'(\omega)$, but the discussion for $\beta_n'(\omega)$ is completely analogous. The slopes $\alpha_n'(\omega)$ can be used for the numerical detection of universality and self-renormalization phenomena. If the bifurcation diagram is self-similar by an affine ratio one should have that $\alpha_n'(\omega)/ \alpha_{n-1}'(\omega)$ converges to a constant. In [@JRT11p2] we compute numerically this ratios and we show that this is not true due to the fact that when the period is doubled, the rotation number of the system also is. What we find is that there exists an affine relationship between the bifurcation diagram of the family for rotation number $\omega$ and the bifurcation diagram of the same family for rotation number $2\omega$. This is sketched in figure \[Esquema BifPar\]. Concretely, in [@JRT11p2] we observed numerically the following behavior. - [**First numerical observation:**]{} the sequence $\alpha_n'(\omega)/ \alpha_{n-1}'(\omega)$ is not convergent in $n$. But, for $\omega$ fix, one obtains the same sequence for any family of quasi-periodic forced map like (\[q.p- family\]), with a quasi-periodic forcing of the type $h(\theta,x) = f_1(x) \cos(\theta)$. - [**Second numerical observation:**]{} the sequence $\alpha_n'(\omega)/ \alpha_{n-1}'(2\omega)$ associated to maps like (\[q.p- family\]) is convergent in $n$ when we take the quasi-periodic forcing of the type $h(\theta,x) = f_1(x) \cos(\theta)$. The limit depends on $\omega$ and $f_1$. - [**Third numerical observation:**]{} the two previous observations are not true when the quasi-periodic forcing is of the type $h_\eta(\theta,x) = f_1(x) \cos(\theta) + \eta f_2(x) \cos(2\theta)$ when $\eta\neq 0$. But the sequence $\alpha_n'(\omega)/ \alpha_{n-1}'(2\omega)$ associated to the map (\[q.p- family\]) with $h = h_\eta$ is $\eta$-close to the same maps with $h= h_0 $ In [@JRT11a] we extended the renormalization operator and we obtained explicit expressions of the slopes $\alpha_n'(\omega)$ and $\beta_n'(\omega)$ in terms of the quasi-periodic renormalization operator. This is reviewed in section \[section review q-p renor\]. In [@JRT11b] we give a theoretical explanation to the numerical observations described above in terms of the dynamics of the quasi-periodic renormalization operator. This is reviewed in section \[section A not yet rigorous explanation\]. The novelty in this paper is that we present numerical support to the conjectures done in [@JRT11a; @JRT11b]. This numerical support is presented after the statement of each of the conjectures, with the exception of conjecture [**\[conjecture H2\]**]{}, which is given in section \[section applicability FLM\]. In Appendix \[section numerical computation of the specturm of L\_omega\] we describe the numerical approximation used to discretize the renormalization operator and how we use it to compute the spectrum of its derivative. [E]{}xistence of reducibility loss bifurcations {#section review q-p renor} =============================================== Consider a quasi-periodic forced map like $$\label{q.p. forced system interval} \begin{array}{rccc} F:& {\mathbb{T}}\times I &\rightarrow & {\mathbb{T}}\times I \\ & \left( \begin{array}{c} \theta \\ x \end{array}\right) & \mapsto & \left( \begin{array}{c} \theta + \omega \\ f(\theta,x) \end{array}\right), \end{array}$$ with $f\in C^r({\mathbb{T}}\times I ,I)$. To define the renormalization operator it is only necessary that $r \geq 1$, but we restrict this study to the analytic case due to technical reasons. Along section \[section definition and basic properties\] it is not necessary to require $\omega$ Diophantine, but it will be necessary in section \[chapter application\]. The definition of the operator is done in a perturbative way, in the sense that it is only applicable to maps $f(\theta,x) = g(x) + h(\theta,x)$ with $g$ renormalizable in the one dimensional sense and $h$ small. Definition of the operator and basic properties {#section definition and basic properties} ----------------------------------------------- ### Preliminary notation Let ${\mathbb{W}}$ be an open set in the complex plane containing the interval $I_\delta=[-1-\delta, 1+\delta]$ and let ${\mathbb{B}}_\rho = \{z = x + i y\in {\mathbb{C}}\text{ such that } |y| < \rho\}$. Then consider ${\mathcal{B}}={\mathcal{B}}({\mathbb{B}}_\rho,{\mathbb{W}})$ the space of functions $f: {\mathbb{B}}_\rho \times {\mathbb{W}}\rightarrow {\mathbb{C}}$ such that: 1. $f$ is holomorphic in ${\mathbb{B}}_{\rho}\times {\mathbb{W}}$ and continuous in the closure of ${\mathbb{B}}_{\rho}\times {\mathbb{W}}$. 2. $f$ is real analytic. 3. $f$ is $1$-periodic in the first variable. This space endowed with the supremum norm is Banach. Let ${\mathcal{RH}}({\mathbb{W}})$ denote the space of functions $f:{\mathbb{W}}\rightarrow {\mathbb{C}}$ such that are holomorphic in ${\mathbb{W}}$, continuous in the closure of ${\mathbb{W}}$, and send real number to real numbers. This space is also Banach with the supremum norm. Consider the operator $$\label{equation projection 0} \begin{array}{rccc} p_0:& {\mathcal{B}}&\rightarrow & {\mathcal{RH}}({\mathbb{W}}) \\ \displaystyle \rule{0pt}{3ex} & f(\theta,x) & \mapsto & \rule{0pt}{3ex} \displaystyle \int_{0}^{1} f(\theta, x) d\theta . \end{array}$$ Let ${\mathcal{B}}_0$ the natural inclusion of ${\mathcal{RH}}({\mathbb{W}}) $ into ${\mathcal{B}}$. Then we have that $p_0$ as a map from ${\mathcal{B}}$ to ${\mathcal{B}}_0$ is a projection ($(p_0)^2 = p_0$). ### Set up of the one dimensional renormalization operator. First we give a concrete definition of the one dimensional renormalization operator before extending it to the quasi-periodic case. Actually, this is a minor modification of the one given in [@Lan82]. Given a small value $\delta>0$, let ${\mathcal{M}}_\delta$ denote the subspace of ${\mathcal{RH}}({\mathbb{W}})$ formed by the even functions $\psi$ which send the interval $I_\delta=[-1-\delta,1+\delta]$ into itself, and such that $\psi(0)=1$ and $x \psi'(x) <0$ for $x\neq 0$. Set $a=\psi(1)$, $a'= (1+\delta)a$ and $b'=\psi(a')$. We can define ${\mathcal{D}}({\mathcal{R}}_\delta)$ as the set of $\psi \in {\mathcal{M}}_\delta$ such that $a<0$, $1> b'>-a'$, and $\psi(b') <- a'$. We define the renormalization operator, ${\mathcal{R}}_\delta : {\mathcal{D}}({\mathcal{R}}_\delta) \rightarrow {\mathcal{M}}_\delta$ as $$\label{renormalization operator lanford} {\mathcal{R}}_\delta(\psi) (x) = \frac{1}{a} \psi \circ \psi (a x).$$ where $a=\psi(1)$. Note that, given $\psi \in {\mathcal{D}}({\mathcal{R}}_\delta)$, one needs to ensure that $\psi\left(a{\mathbb{W}}\right)\subset {\mathbb{W}}$ in order to have ${\mathcal{R}}_\delta(\psi)$ well defined. With this aim, let us consider the following hypothesis. H0) : There exists an open set ${\mathbb{W}}\subset {\mathbb{C}}$ containing $I_\delta$ and a function $\Phi\in {\mathcal{B}}$such that $\phi= p_0(\Phi) $ is a fixed point of the renormalization operator ${\mathcal{R}}_\delta$ and such that the closure of both $a{\mathbb{W}}$ and $\phi(\Phi)(a {\mathbb{W}})$ is contained in ${\mathbb{W}}$ (with $a:=\Phi(1)$). In [@Lan92], Lanford claims that the hypothesis [**H0**]{} is satisfied by the set $$\label{set Lanford} \left\{ z\in {\mathbb{C}}\text { such that } |z^2 -1| < \frac{5}{2} \right\}.$$ This set is convenient for him because he works in the set of even holomorphic functions. In [@Lan82] Lanford introduces a discretization of the (one dimensional) renormalization operator to give a computer assisted proof of the contractivity of the operator. In the present paper we use the same techniques to discretize the quasi-periodic renormalization operator, although we do it without the use of rigorous interval arithmetics. More details on this are given in the Appendix \[section numerical computation of the specturm of L\_omega\]. We can use this discretization to check the hypothesis [**H0**]{} for a suitable set ${\mathbb{W}}$. ![ Boundaries of the sets ${\mathbb{D}}\left(\frac{1}{5},\frac{3}{2} \right)$ (solid line) and $\phi\left(a {\mathbb{D}}\left(\frac{1}{5},\frac{3}{2} \right)\right)$ (dashed line).[]{data-label="figure inclusion boundaries"}](Domains.pdf){width="7.5cm"} Our study is not restricted to the case of even functions, therefore the set (\[set Lanford\]) used by Lanford is not valid in our case. Using the method described in the Appendix \[section numerical computation of the specturm of L\_omega\] we recomputed the fixed point $\phi$ of the (one dimensional) renormalization operator ${\mathcal{R}}$. The fixed point has been computed by means of a Newton method with our discretization and then we have checked that the Taylor expansion around zero coincides with the one given in [@Lan82]. With this approximation of $\phi$ we checked (numerically) that ${\mathbb{D}}\left(\frac{1}{5},\frac{3}{2} \right)$ the disc of the complex plane centered at $\frac{1}{5}$ with radius $\frac{3}{2}$ satisfies the conditions required to the set ${\mathbb{W}}$ in hypothesis [**H0**]{}. In other words, we checked that $\phi\left( a {\mathbb{D}}\left(\frac{1}{5},\frac{3}{2} \right)\right)$ is contained inside ${\mathbb{D}}\left(\frac{1}{5},\frac{3}{2} \right)$ (recall that $a=\phi(1)$ and $a {\mathbb{D}}(z_0,\rho) = \{ z \in {\mathbb{C}}\thinspace | \thinspace az \in {\mathbb{D}}(z_0,\rho) \}$ ). Denote by $\partial {\mathbb{D}}(z_0,\rho) $ the boundary of the disk ${\mathbb{D}}(z_0,\rho) $. In figure \[figure inclusion boundaries\] we have plotted the sets $\partial {\mathbb{D}}\left(\frac{1}{5},\frac{3}{2} \right)$ and $\phi\left(a \partial {\mathbb{D}}\left(\frac{1}{5},\frac{3}{2} \right)\right)$ which give a visual evidence of the inclusion. Recall that $\phi$ is analytic, then to check that the set ${\mathbb{D}}\left(\frac{1}{5},\frac{3}{2} \right)$ is mapped inside the set delimited by $\phi\left(a \partial {\mathbb{D}}\left(\frac{1}{5},\frac{3}{2} \right)\right)$ it is enough to check that one point in the interior of ${\mathbb{D}}\left(\frac{1}{5},\frac{3}{2} \right)$ is mapped in the interior of $\phi\left(a \partial {\mathbb{D}}\left(\frac{1}{5},\frac{3}{2} \right)\right)$. Recall that $\phi(0)=1$ by hypothesis, therefore the inclusion holds. ### Definition of the renormalization operator for quasi-periodically forced maps Consider the space ${\mathcal{X}}\subset {\mathcal{B}}$ defined as: $${\mathcal{X}}= \{ f \in C^r({\mathbb{T}}\times I_\delta, I_\delta) | \thinspace p_0(f) \in {\mathcal{M}}_\delta\}.$$ Consider also the decomposition ${\mathcal{X}}={\mathcal{X}}_0 \oplus {\mathcal{X}}_0^c$ given by the projection $p_0$. In other words, we have ${\mathcal{X}}_0=\{f \in {\mathcal{X}}\thinspace |\text{ } p_0(f)=f\}$ and ${\mathcal{X}}^c_0=\{f \in {\mathcal{X}}\thinspace | \text{ } p_0(f)=0\}$. Note that from the definition of ${\mathcal{X}}$ it follows that ${\mathcal{X}}_0$ is an isomorphic copy of ${\mathcal{M}}_\delta$. Given a function $g\in {\mathcal{X}}$, we define the [**quasi-periodic renormalization**]{} of $g$ as $$\label{operator tau} [{\mathcal{T}}_\omega(g)](\theta,x) := \frac{1}{\hat{a}} g(\theta + \omega, g(\theta, \hat{a}x)),$$ where $\displaystyle \hat{a} = \int_{0}^{1} g(\theta, 1) d\theta$. Then we have that there exists a set ${\mathcal{D}}({\mathcal{T}})$, open in $\left(p_0 \circ {\mathcal{T}}_\omega\right)^{-1} ({\mathcal{M}}_\delta)$, where the operator is well defined. Moreover this set contains ${\mathcal{D}}_0({\mathcal{T}})$, the inclusion of ${\mathcal{D}}({\mathcal{R}})$ in ${\mathcal{B}}$. By definition we have that ${\mathcal{T}}_\omega$ restricted to ${\mathcal{D}}_0({\mathcal{T}})$ is isomorphically conjugate to ${\mathcal{R}}$, therefore the fixed points of ${\mathcal{R}}$ extend to fixed points of ${\mathcal{T}}_\omega$. Assume that [**H0**]{} holds and let $\Phi$ be the fixed point given by this hypothesis. Then we have that there exists $U\subset {\mathcal{D}}({\mathcal{T}})\cap{\mathcal{B}}$, an open neighborhood of $\Phi$, such that ${\mathcal{T}}_\omega :U\rightarrow {\mathcal{B}}$ is well defined. Moreover we have that ${\mathcal{T}}_\omega$ is Fréchet differentiable for any $\Psi \in U$. ### Fourier expansion of $D{\mathcal{T}}_\omega(\Psi)$. {#section The Fourier expansion of DT} Let $\Psi$ be a function in a neighborhood of $\Phi$ (given in hypothesis [**H0**]{}) where ${\mathcal{T}}_\omega$ is differentiable. Additionally, assume that $\Psi\in {\mathcal{D}}_0({\mathcal{T}}_\omega)$. Given a function $f\in {\mathcal{B}}$ we can consider its complex Fourier expansion in the periodic variable $$\label{fourier expansion complex} f(\theta,z)= \sum_{k\in {\mathbb{Z}}} c_k(z) e^{2\pi k\theta i },$$ with $$c_k(z)= \int_{0}^{1} f(\theta,z) e^{-2\pi k\theta i} d\theta.$$ Then we have that $D{\mathcal{T}}_\omega$ “diagonalizes” with respect to the complex Fourier expansion, in the sense that we have $$\label{equation Fourier expansion differential renormalization operator} \left[D{\mathcal{T}}_\omega(\Psi) f\right](\theta,z) = D{\mathcal{R}}_\delta[c_0](z) + \sum_{k\in {\mathbb{Z}}\setminus\{0\}} \left([L_1(c_k)](z) + [L_2(c_k)](z) e^{2\pi k\omega i}\right) e^{2\pi k\theta i },$$ where $$\begin{array}{rccc} L_1: & {\mathcal{RH}}({\mathbb{W}}) & \rightarrow & {\mathcal{RH}}({\mathbb{W}}) \\ & g(z) & \mapsto & \displaystyle \frac{1}{a} \psi'\circ\psi(a z) g(az), \end{array}$$ and $$\begin{array}{rccc} L_2: & {\mathcal{RH}}({\mathbb{W}}) & \rightarrow & {\mathcal{RH}}({\mathbb{W}}) \\ & g(z) & \mapsto & \displaystyle \frac{1}{a} g\circ\psi(a z), \end{array}$$ with $\psi=p_0(\Psi)$ and $a=\psi(1)$. An immediate consequence of this diagonalization is the following. Consider $$\label{equation spaces bbk} {\mathcal{B}}_k:= \big\{ f \in B |\text{ } f(\theta, x) = u(x) \cos(2\pi k \theta) + v(x) \sin(2\pi k\theta), \text{ for some } u,v\in {\mathcal{RH}}({\mathbb{W}})\big\},$$ then we have that the spaces ${\mathcal{B}}_k$ are invariant by $D{\mathcal{T}}(\Psi)$ for any $k>0$. Moreover $D{\mathcal{T}}_\omega (\Psi)$ restricted to ${\mathcal{B}}_k$ is conjugate to ${\mathcal{L}}_{k\omega}$, where ${\mathcal{L}}_\alpha$ is the defined as $$\label{equation maps L_omega} \begin{array}{rccc} {\mathcal{L}}_\alpha: & {\mathcal{RH}}({\mathbb{W}})\oplus {\mathcal{RH}}({\mathbb{W}}) & \rightarrow & {\mathcal{RH}}({\mathbb{W}})\oplus {\mathcal{RH}}({\mathbb{W}}) \\ \\ & \left( \begin{array}{c} u \\v \end{array} \right) & \mapsto & \left( \begin{array}{c} L_1(u) \\ L_1(v) \end{array} \right) + \left( \begin{array}{cc} \cos(2\pi \alpha) & - \sin(2\pi \alpha) \\ \sin(2\pi \alpha) & \cos( 2\pi \alpha) \end{array} \right) \left( \begin{array}{c} L_2(u) \\ L_2(v) \end{array} \right) . \end{array}$$ Then we have that the understanding of the derivative of the renormalization operator in ${\mathcal{B}}$ is equivalent to the study of the operator ${\mathcal{L}}_\omega$ for a any $\omega\in {\mathbb{T}}$. ### Properties of ${\mathcal{L}}_\omega$ ![Numerical approximation of the spectrum of ${\mathcal{L}}_\omega$ for $\omega \in {\mathbb{T}}$. Top: projection in the complex plane of the spectrum when $\omega$ varies in ${\mathbb{T}}$. Bottom left: evolution of the real part with respect to $\omega$. Bottom right: evolution of the imaginary part with respect to $\omega$. []{data-label="Operator Spectrum"}](SpecC.pdf "fig:"){width="12cm"} ![Numerical approximation of the spectrum of ${\mathcal{L}}_\omega$ for $\omega \in {\mathbb{T}}$. Top: projection in the complex plane of the spectrum when $\omega$ varies in ${\mathbb{T}}$. Bottom left: evolution of the real part with respect to $\omega$. Bottom right: evolution of the imaginary part with respect to $\omega$. []{data-label="Operator Spectrum"}](SpecR.pdf "fig:"){width="6cm"} ![Numerical approximation of the spectrum of ${\mathcal{L}}_\omega$ for $\omega \in {\mathbb{T}}$. Top: projection in the complex plane of the spectrum when $\omega$ varies in ${\mathbb{T}}$. Bottom left: evolution of the real part with respect to $\omega$. Bottom right: evolution of the imaginary part with respect to $\omega$. []{data-label="Operator Spectrum"}](SpecI.pdf "fig:"){width="6cm"} Given a value $\gamma \in {\mathbb{T}}$, consider the rotation $R_{\gamma}$ defined as $$\label{equation rotation rgamma} \begin{array}{rccc} R_\gamma: & {\mathcal{RH}}({\mathbb{W}})\oplus {\mathcal{RH}}({\mathbb{W}}) & \rightarrow & {\mathcal{RH}}({\mathbb{W}})\oplus {\mathcal{RH}}({\mathbb{W}}) \\ \\ & \left( \begin{array}{c} u \\v \end{array} \right) & \mapsto & \left( \begin{array}{cc} \cos( 2\pi \gamma) & - \sin(2\pi \gamma) \\ \sin(2\pi \gamma) & \cos(2\pi \gamma) \end{array} \right) \left( \begin{array}{c} u \\ v \end{array} \right) , \end{array}$$ then we have that ${\mathcal{L}}_\omega$ and $R_\gamma$ commute for any $\omega, \gamma \in {\mathbb{T}}$. This has some consequences on the spectrum of ${\mathcal{L}}_\omega$. Concretely, we have that any eigenvalue of ${\mathcal{L}}_\omega$ (different from zero) is either real with geometric multiplicity even, or a pair of complex conjugate eigenvalues. On the other hand ${\mathcal{L}}_\omega$ depends analytically on $\omega$, which (using theorems III-6.17 and VII-1.7 of [@Kat66]) imply that (as long as the eigenvalues of ${\mathcal{L}}_\omega$ are simple) the eigenvalues and their associated eigenspaces depend analytically on the parameter $\omega$. Finally, doing some minor changes on the domain of definition, we can prove the compactness of the operator ${\mathcal{L}}_\omega$. Recall that the compactness of an operator implies that its spectrum is either finite or countable with $0$ on its closure (see for instance theorem III-6.26 of [@Kat66]). In figure \[Operator Spectrum\] we have a numerical approximation of the spectrum of the operator ${\mathcal{L}}_\omega$ depending on $\omega$. We can observe that the properties described above are satisfied. The details on the numerical computations involved to approximate the spectrum are described in Appendix \[section numerical computation of the specturm of L\_omega\]. Several numerical tests on the reliability of the results are also included there. Reducibility loss and quasi-periodic renormalization {#chapter application} ---------------------------------------------------- Given a map $F$ like (\[q.p. forced system interval\]) with $f\in{\mathcal{B}}$ and $\omega\in{\mathbb{T}}$, we denote by $f^{n}:{\mathbb{T}}\times {\mathbb{R}}\rightarrow {\mathbb{R}}$ the $x$-projection of $F^n(x,\theta)$. Equivalently $f^n$ can be defined through the recurrence $$\label{definicio falan} f^n(\theta, x) = f( \theta +(n-1) \omega, f^{n-1}(\theta,x)), \quad f^0(\theta,x) \equiv x.$$ From this point on, whenever $\omega$ is used, it is assumed to be Diophantine. Denote by $\Omega= \Omega_{\gamma,\tau}$ the set of Diophantine numbers, this is the set of $\omega\in {\mathbb{T}}$ such that there exist $\gamma >0$ and $\tau \geq 1$ such that $$|q \omega - p| \geq \frac{\gamma}{|q|^{\tau}}, \quad \text{ for all } (p,q) \in {\mathbb{Z}}\times ({\mathbb{Z}}\setminus \{0\}).$$ Additionally, we will need to assume that the following conjecture is true. \[conjecture H2\] The operator ${\mathcal{T}}_{\omega}$ (for any $\omega\in\Omega$) is an injective function when restricted to the domain ${\mathcal{B}}\cap {\mathcal{D}}({\mathcal{T}})$. Moreover, there exists $U$ an open set of ${\mathcal{D}}({\mathcal{T}})$ containing $W^u(\Phi,{\mathcal{R}}) \cup W^s(\Phi,{\mathcal{R}})$[^2] where the operator ${\mathcal{T}}_\omega$ is differentiable. In [@JRT11a] we discuss the difficulties for proving this conjecture. A priory there is no way to check numerically this kind of conjecture. A posteriori we have that the results obtained assuming this conjecture are coherent with the numerical computations (see section \[section applicability FLM\]). ### Consequences for a two parametric family of maps {#Section consequences for a two parametric family of maps} Consider a two parametric families of maps $\{c(\alpha,{\varepsilon})\}_{(\alpha,{\varepsilon})\in A}$ contained in ${\mathcal{B}}$, with $A = [a,b]\times[0,c]$ and $a$, $b$ and $c$ are real numbers (with $a<b$ and $0<c$). We assume that the dependency on the parameters is analytic. Consider the following hypothesis on the family of maps. H1) : The family $\{c(\alpha,{\varepsilon})\}_{(\alpha,{\varepsilon})\in A}$ uncouples for ${\varepsilon}=0$, in the sense that the family $\{c(\alpha,0)\}_{\alpha\in[a,b]}$ does not depend on $\theta$ and it has a full cascade of period doubling bifurcations. We assume that the family $\{c(\alpha,0)\}_{\alpha\in[a,b]}$ crosses transversally the stable manifold of $\Phi$, the fixed point of the renormalization operator, and each of the manifolds $\Sigma_n$ for any $n\geq 1$, where $\Sigma_{n}$ is the inclusion in ${\mathcal{B}}$ of the set of one dimensional unimodal maps with a super-attracting $2^n$ periodic orbit. In other words, we assume that the family $c(\alpha,{\varepsilon})$ can be written as, $$c(\alpha,{\varepsilon})=c_0(\alpha) + {\varepsilon}c_1(\alpha,{\varepsilon}),$$ with $\{c_0(\alpha)\}_{\alpha\in[a,b]}\subset {\mathcal{B}}_0$ having a full cascade of period doubling bifurcations. Given a family $\{c(\alpha,{\varepsilon})\}_{(\alpha,{\varepsilon})\in A}$ satisfying the hypothesis [**H1**]{}, let $\alpha_n$ be the parameter value for which the uncoupled family $\{c(\alpha,0)\}_{\alpha\in[a,b]}$ intersects the manifold $\Sigma_n$. Note that the critical point of the map $c(\alpha_n,0)$ is a $2^n$-periodic orbit. Our main achievement in [@JRT11a] is to prove that from every parameter value $(\alpha_n,0)$ there are born two curves in the parameter space, each of them corresponding to a reducibility loss bifurcation. If we want to give a more precise statement of the result we need now to introduce some technical definitions. Let ${\mathcal{RH}}({\mathbb{B}}_\rho,{\mathbb{W}})$ denote the space of periodic real analytic maps from ${\mathbb{B}}_\rho$ to ${\mathbb{W}}$ and continuous in the closure of ${\mathbb{B}}_\rho$. Consider a map $f_0\in {\mathcal{B}}$ and $\omega \in \Omega$, such that $f$ has a periodic invariant curve $x_0$ of rotation number $\omega$ with a Lyapunov exponent less equal than certain $-K_0<0$. Using lemma 3.6 in [@JRT11a] we have that there exist a neighborhood $V\subset {\mathcal{B}}$ of $f_0$ and a map $x\in {\mathcal{RH}}({\mathbb{B}}_\rho, {\mathbb{W}})$ such that $x(f)$ is a periodic invariant curve of $f$ for any $f\in V$. Then we can define the map $G_1 $ as $$\label{equation definition G1} \begin{array}{rccc} G_1:& \Omega\times V &\rightarrow & {\mathcal{RH}}({\mathbb{B}}_\rho,{\mathbb{C}}) \\ \rule{0pt}{3ex} & (\omega,g) & \mapsto & D_x g \big(\theta+\omega,g(\theta, \left[x(\omega,g)\right](\theta))\big) D_x g \big(\theta,\left[x(\omega,g)\right](\theta)\big). \end{array}$$ On the other hand, we can consider the counterpart of the map $G_1$ in the uncoupled case. Given a map $f_0\in {\mathcal{B}}_0$, consider $U\subset {\mathcal{B}}_0$ a neighborhood of $f_0$ in the ${\mathcal{B}}_0$ topology. Assume that $f_0$ has a attracting $2$-periodic orbit $x_0\in I$. Let $x=x(f)\in {\mathbb{W}}$ be the continuation of this periodic orbit for any $f\in U$. We have that $x$ depends analytically on the map, therefore it induces a map $x:U \rightarrow {\mathbb{W}}$. Then if we take $U$ small enough we have an analytic map $x:U\rightarrow {\mathbb{W}}$ such that $x[f]$ is a periodic orbit of period 2. Now we can consider the map $$\label{equation definition widehatG1} \begin{array}{rccc} \widehat{G}_1:& U\subset {\mathcal{B}}_0 &\rightarrow & {\mathbb{C}}\\ \rule{0pt}{3ex} & f & \mapsto & D_x f \big(f(x[f])\big) D_x f \big( x[f] \big). \end{array}$$ Note that $\widehat{G}_1$ corresponds to $G_1$ restricted to the space ${\mathcal{B}}_0$ (but then $\widehat{G}_1(f)$ has to be seen as an element of ${\mathcal{RH}}({\mathbb{B}}_\rho,{\mathbb{W}})$). Consider the sequences $$\label{equation sequences corollary directions} \begin{array}{rcll} \omega_k & = & 2 \omega_{k-1}, & \text{ for } k=1,..., n-1. \\ \rule{0ex}{4ex} f^{(n)}_k & = & {\mathcal{R}}\left(f_{k-1}^{(n)}\right), & \text{ for } k=1,..., n-1. \\ \rule{0ex}{4ex} u^{(n)}_k & = & D {\mathcal{R}}\left(f^{(n)}_{k-1}\right) u^{(n)}_{k-1}, & \text{ for } k = 1, ..., n-1.\\ \rule{0ex}{4ex} v^{(n)}_k & = & D {\mathcal{T}}_{\omega_{k-1}} \left(f^{(n)}_{k-1}\right) v^{(n)}_{k-1}, & \text{ for } k = 1, ..., n-1. \end{array}$$ with $$\label{equation sequences directions initial} f^{(n)}_0 = c(\alpha_n,0), \quad u^{(n)}_0 = \partial_\alpha c(\alpha_n,0), \quad v^{(n)}_0 = \partial_{\varepsilon}c(\alpha_n,0).$$ Note that $f^{(n)}_0= \{c(\alpha,0)\}\cap \Sigma_n$, then $f^{(n)}_0$ tends to $W^s({\mathcal{R}},\Phi)$ when $n$ grow. Therefore, the sequence $\{f^{(n)}_k\}_{0\leq k <n}$ attains to $W^s({\mathcal{R}},\Phi) \cup W^u({\mathcal{R}},\Phi)$ when $n$ grows and consequently there exist $n_0$ such that $\{f^{(n)}_k\}_{0\leq k <n} \subset U$ , where $U$ is the neighborhood given in conjecture [**\[conjecture H2\]**]{}. If the conjecture is true, then the operator ${\mathcal{T}}_\omega$ is differentiable in the orbit $\{f^{(n)}_k\}_{0\leq k <n} \subset U$. Consider the following hypothesis. H2) : The family $\{c(\alpha,{\varepsilon})\}_{(\alpha,{\varepsilon})\in A}$ is such that $$D G_1 \left(\omega_{n-1}, f^{(n)}_{n-1}\right) D{\mathcal{T}}_{\omega_{n-2}}\left(f^{(n)}_{n-2}\right) \cdots D{\mathcal{T}}_{\omega_0}\left(f^{(n)}_0\right) \partial_{\varepsilon}c(\alpha_n,0),$$ has a unique non-degenerate minimum (respectively maximum) as a function from ${\mathbb{T}}$ to ${\mathbb{R}}$, for any $n\geq n_0$. Consider a family of maps $\{c(\alpha,{\varepsilon})\}_{(\alpha,{\varepsilon})\in A}$ such that the hypotheses [**H1**]{} and [**H2**]{} are satisfied and $\omega_0\in \Omega$. If the conjecture [**\[conjecture H2\]**]{} is true, then theorem 3.8 in [@JRT11a] asserts that there exists $n_0$ such that, for any $n\geq n_0$, there exist two bifurcation curves around the parameter value $(\alpha_n, 0)$, such that they correspond to a reducibility-loss bifurcation of the $2^n$-periodic invariant curve. Moreover, these curves are locally expressed as $(\alpha_n + \alpha_n'(\omega) {\varepsilon}+ o({\varepsilon}),{\varepsilon})$ and $(\alpha_n^- + \beta_n'(\omega) {\varepsilon}+ o({\varepsilon}), {\varepsilon})$ with $$\label{equation alpha n +} \alpha'_n (\omega) = - \frac{m \left( DG_1 \left(\omega_{n-1}, f^{(n)}_{n-1}\right) v_{n-1}^{(n)} \right) }{\rule{0ex}{3.5ex} D \widehat{G}_1 \left(f^{(n)}_{n-1}\right) u_{n-1}^{(n)}} ,$$ and $$\label{equation alpha n -} \beta'_n(\omega) = - \frac{M \left( D G_1 \left(\omega_{n-1}, f^{(n)}_{n-1}\right) v_{n-1}^{(n)} \right) }{\rule{0ex}{3.5ex} D \widehat{G}_1 \left(f^{(n)}_{n-1}\right) u_{n-1}^{(n)}} ,$$ where $G_1$ and $\widehat{G}_1$ are given by equations (\[equation definition G1\]) and (\[equation definition widehatG1\]), and $m$ and $M$ are the minimum and the maximum as operators, that is $$\label{equation definition of minim} \begin{array}{rccc} m:& {\mathcal{RH}}({\mathbb{B}}_\rho,{\mathbb{C}}) &\rightarrow & {\mathbb{R}}\\ \rule{0pt}{3ex} & g & \mapsto & \displaystyle \min_{\theta \in {\mathbb{T}}} g(\theta). \end{array}$$ and $$\label{equation definition of maximum} \begin{array}{rccc} M:& {\mathcal{RH}}({\mathbb{B}}_\rho,{\mathbb{C}}) &\rightarrow & {\mathbb{R}}\\ \rule{0pt}{3ex} & g & \mapsto & \displaystyle \max_{\theta \in {\mathbb{T}}} g(\theta). \end{array}$$ Let us focus again on hypothesis [**H2**]{}, which is not intuitive. We can introduce a stronger condition which much more easy to check. Moreover this conditions is automatically satisfied by maps like the Forced Logistic Map. Consider a family of maps $\{c(\alpha,{\varepsilon})\}_{(\alpha,{\varepsilon})\in A}$ as before, satisfying hypothesis [**H1**]{}. H2’) : The family $\{c(\alpha,{\varepsilon})\}_{(\alpha,{\varepsilon})\in A}$ is such that the quasi-periodic perturbation $\partial_{\varepsilon}c(\alpha,0)$ belongs to the set ${\mathcal{B}}_1$ (see equation (\[equation spaces bbk\])) for any value of $\alpha$ (with $(\alpha,0)\in A$). Then we have that [**H2’**]{} implies [**H2**]{} (see proposition 3.10 in [@JRT11a]). [A]{}symptotic behavior of reducibility loss bifurcations {#section A not yet rigorous explanation} ========================================================= Consider a two parametric family of maps $\{c(\alpha,{\varepsilon})\}_{(\alpha,{\varepsilon}) \in A}$ contained in ${\mathcal{B}}$, with $A = [a,b]\times[0,d]$ and $a$, $b$ and $d$ are real numbers (with $a<b$ and $0<d$). We assume that the dependency on the parameters is analytic and the family is such that the hypotheses [**H1**]{} and [**H2**]{} introduced in section \[chapter application\] are satisfied. Consider also the reducibility loss bifurcation curves associated to the $2^n$-periodic orbit given by (\[equation alpha n +\]). Since the value $\alpha_n'(\omega)$ depends also on the family of maps $c$ considered, we will denote it by $\alpha_n'(\omega,c)$ from now on. We omit the case concerning $\beta_n'(\omega)$ (see (\[equation alpha n -\])) because it is completely analogous to the one considered here. Rotational symmetry reduction {#subsection rotational symmetry reduction} ----------------------------- Given $\gamma\in {\mathbb{T}}$, consider the following auxiliary function $$\label{definition tgamma} \begin{array}{rccc} t_\gamma:& {\mathcal{B}}&\rightarrow & {\mathcal{B}}\\ \rule{0pt}{5ex} & v(\theta, z) & \mapsto & v(\theta + \gamma, z). \end{array}$$ Let ${\mathcal{B}}_1$ be the subspace of ${\mathcal{B}}$ defined by (\[equation spaces bbk\]) for $k=1$. The space ${\mathcal{B}}_1$ is indeed the image of the projection $\pi_1:{\mathcal{B}}\rightarrow {\mathcal{B}}$ defined as $$\label{projection pi_1} \left[\pi_1 (v)\right](\theta , x) = \left( \int_0^1 v(\theta,x) \cos(2\pi x) d\theta \right) \cos (2\pi\theta) + \left( \int_0^1 v(\theta,x) \sin(2\pi x) d\theta \right) \sin (2\pi\theta).$$ Given $x_0\in {\mathbb{W}}\cap {\mathbb{R}}$ and $\theta_0\in{\mathbb{T}}$ we can also consider the sets $${\mathcal{B}}_1' = {\mathcal{B}}_1' (\theta_0,x_0) = \{f\in {\mathcal{B}}_1 \thinspace | f(\theta_0,x_0)=0, \partial_\theta f(\theta_0,x_0)>0 \},$$ and $${\mathcal{B}}' = {\mathcal{B}}' (\theta_0,x_0) = \{f\in {\mathcal{B}}\thinspace | \pi_1(f) \in {\mathcal{B}}_1' \}.$$ Note that $ {\mathcal{B}}_1' (\theta_0,x_0)$ depends on the election of $(\theta_0,x_0)$, but for any fixed $x_0$ and $\theta_0$ the set ${\mathcal{B}}_1' (\theta_0, x_0)$ is an open subset of a codimension one linear subspace of ${\mathcal{B}}_1$. Note also that any $v\in$ then $v\neq 0$ due to the condition $\partial_\theta v(\theta_0,x_0)>0$. Moreover for any $v\in {\mathcal{B}}_1 \setminus\{0\}$ there exists a unique $\gamma_0\in {\mathbb{T}}$ such that $t_{\gamma_0}(v)\in {\mathcal{B}}_1'(\theta_0,x_0)$ (see proposition 3.1 in [@JRT11b]). Consider a two parametric family of maps $\{c(\alpha,{\varepsilon})\}_{(\alpha,{\varepsilon}) \in A}$ contained in ${\mathcal{B}}$ satisfying the hypotheses [**H1**]{} and [**H2**]{} as in section \[chapter application\]. Consider also the reducibility loss bifurcation curves associated to the $2^n$-periodic orbit with slopes given by (\[equation alpha n +\]) and (\[equation alpha n -\]). Then we have that the formulas (\[equation alpha n +\]) and (\[equation alpha n -\]) can be expressed in term of vectors in ${\mathcal{B}}_1' (\theta_0, x_0)$. Let us see this with more detail. Consider the sequences $\{\omega_k\}$, $\{f^{(n)}_k\}$ and $\{u^{(n)}_k\}$ given by (\[equation sequences corollary directions\]) and (\[equation sequences directions initial\]). Consider now the sequence $$\label{equation v_k in section} \tilde{v}^{(n)}_k = t_{\gamma\left(\tilde{v}^{(n)}_{k-1}\right)} \left( D {\mathcal{T}}_{\omega_{k-1}} \left(f^{(n)}_{k-1}\right) \tilde{v}^{(n)}_{k-1} \right) \text{ for } k = 1, ..., n-1,$$ and $$\label{equation v_0 in section} v^{(n)}_0 = t_{\gamma_0} \left(\partial_{\varepsilon}c(\alpha_n,0)\right),$$ where $\gamma(\tilde{v}^{(n)}_{k-1})$ and $\gamma_0$ are chosen such that $\tilde{v}^{(n)}_{k}$ belongs to ${\mathcal{B}}' (\theta_0, x_0)$ for $k=0,1, ..., n$. If the projection of $D{\mathcal{T}}_{\omega_{k-1}} \left(f^{(n)}_{k-1} \right) \tilde{v}^{(n)}_{k-1}$ in ${\mathcal{B}}_1$ is non zero, then $\gamma\left(\tilde{v}^{(n)}_{k-1}\right)$ is uniquely determined and the vectors $\tilde{v}^{(n)}_k$ are well defined. Moreover, if we assume that $\omega_0\in \Omega$ and that the conjecture [**\[conjecture H2\]**]{} is true then we have that $\alpha'_n (\omega)$ can be written as $$\label{equation alpha n + section} \alpha'_n (\omega) = - \frac{m \left( DG_1 \left(\omega_{n-1}, f^{(n)}_{n-1}\right) \tilde{v}_{n-1}^{(n)} \right) }{\rule{0ex}{3.5ex} D \widehat{G}_1 \left(f^{(n)}_{n-1}\right) u_{n-1}^{(n)}} ,$$ where $G_1$, $\widehat{G}_1$ and $m$ are given by equations (\[equation definition G1\]), (\[equation definition widehatG1\]) and (\[equation definition of minim\]). For more details see theorem 3.2 in [@JRT11b]. Reduction to the dynamics of the renormalization operator {#subsection reduction to the dynamics} --------------------------------------------------------- The goal of this section is to reduce the problem of describing the asymptotic behavior of ${\alpha'_n(\omega_0,c_1)}/{\alpha'_{n-1}(\omega_0,c_1)}$ to the dynamics of the quasi-periodic renormalization operator. \[definition equivalence of banach sequences\] Given two sequences $\{r_i\}_{i\in {\mathbb{Z}}_+}$ and $\{s_i\}_{i\in {\mathbb{Z}}_+}$ in a Banach space, we say that they are [**asymptotically equivalent**]{} if there exists $0<\rho <1$ and $k_0$ such that $$\| r_i - s_i\| \leq k_0 \rho^i \quad \forall i\in {\mathbb{Z}}_+ .$$ We will commit an abuse of notation and denote this equivalence relation by $s_i \sim r_i$ instead of $\{r_i\}_{i\in {\mathbb{Z}}_+} \sim \{s_i\}_{i\in {\mathbb{Z}}_+}$. Given a family $\{c(\alpha,{\varepsilon})\}_{(\alpha,{\varepsilon})\in A}$ satisfying hypotheses [**H1**]{} and [**H2**]{} and a fixed Diophantine rotation number $\omega_0$, let $\alpha^*$ denote the parameter value such that the family $\{c(\alpha,0)\}_{(\alpha,0)\in A}$ intersects with $W^s(\Phi, {\mathcal{R}})$ and $f_j^*$ denote the intersection of $W^u(\Phi, {\mathcal{R}})$ with the manifold $\Sigma_j$. Consider then $$\label{equation sequences directions simplified} \begin{array}{rcl} \rule{0ex}{2.5ex}\omega_k & = & 2 \omega_{k-1}, \text{ for } k = 1, ..., n-1, \\ \rule{0ex}{4.5ex} u_k & = &\displaystyle \left\{ \begin{array}{ll} \rule{0ex}{2.5ex} \displaystyle D {\mathcal{R}}\left(\Phi \right) u_{k-1}, & \text{ for } k = 1, \dots, [n/2]-1, \\ \rule{0ex}{2.5ex} \displaystyle D {\mathcal{R}}\left(f_{n-k}^* \right) u_{k-1}, & \text{ for } k = [n/2], \dots, n-1. \end{array} \right. \\ \\\rule{0ex}{4.5ex} v_k & = & \displaystyle \left\{ \begin{array}{ll} \rule{0ex}{2.5ex} \displaystyle t_{\gamma\left(\tilde{v}_{k-1}\right)} \left( D{\mathcal{T}}_{\omega_{k-1}} \left(\Phi \right) v_{k-1} \right), & \text{ for } k = 1, ..., [n/2]-1,\\ \rule{0ex}{2.5ex} \displaystyle t_{\gamma\left(\tilde{v}_{k-1}\right)} \left( D{\mathcal{T}}_{\omega_{k-1}} \left(f_{n-k}^*\right) v_{k-1}\right), &\text{ for } k=[n/2],\dots, n-1. \end{array} \right. \\ \end{array}$$ with $$u_0 = \partial_\alpha c(\alpha^*,0), \quad v_0 = t_{\gamma_0} \left( \partial_{\varepsilon}c(\alpha^*,0) \right),$$ and $\gamma(\tilde{v}_{s-1})$ and $\gamma_0$ are chosen such that $\tilde{v}^{(n)}_{s}$ belongs to ${\mathcal{B}}_1' (\theta_0, x_0)$ for any $s=1, ..., n$. Consider the following conjecture for the forthcoming discussion. \[conjecture H3\] For any family of maps $\{c(\alpha,{\varepsilon})\}_{(\alpha,{\varepsilon})\in A}$ satisfying [**H1** ]{} and [**H2**]{}, assume that $$\frac{\tilde{v}^{(n)}_{n-1}}{\|\tilde{v}^{(n)}_{n-1} \|} \sim \frac{v_{n-1}}{\|v_{n-1} \|},$$ with $\tilde{v}^{(n)}_{n-1}$ and $v_{n-1}$ given by (\[equation v\_k in section\]) and (\[equation sequences directions simplified\]). Also assume that there exists a constant $C>0$ such that $$\|v_{n-1}\| > C \text{ for any } n > 0 .$$ Finally assume that there exists a constant $C_0>0$ such that $$\left|m \left( DG_1\left(\omega_{n-1},f^*_1, \frac{v_{n-1}}{\| v_{n-1} \|}\right) \right) \right|> C_0,$$ for any $n\geq0$ and $\omega_0$ Diophantine, where $m$ is given by (\[equation definition of minim\]), $G_1$ by (\[equation definition G1\]) and $\{f_1^*\}= W^u({\mathcal{R}},\Phi)\cap\Sigma_1 $. In the first part of the conjecture we assume that the asymptotic behavior of the vectors $v^{(n)}_{n-1}$ is determined by the linearization of the renormalization operator in the fixed point. We have that the iterates of $f_0^{(n)}$ correspond to a passage near a saddle point. The initial point $f_0^{(n)}$ is always in $\{c(\alpha,0)\}_{(\alpha,0) \in A}$, the final point $f_{n-1}^{(n)}$ is always in $\Sigma_1$ for any $n$, and the orbit of the points spends more and more iterates in a neighborhood of $\Phi$ when $n$ is increased. Therefore it is reasonable to expect that the asymptotic behavior is governed by the linearization of the operator on the fixed point. In the second part of the conjecture we assume that the modulus of the vector does not decrease to zero. Table \[taula NumSupport conj B\] supports numerically conjecture [**\[conjecture H3\]**]{} for the Forced Logistic Map (\[FLM new set up\]) with $ \omega = \frac{\sqrt{5}}{2}$. Note that instead of computing the values $ \|v_{n-1}\| $, $ \left|m \left( DG_1\left(\omega_{n-1},f^*_1, \frac{v_{n-1}}{\| v_{n-1} \|}\right) \right) \right|$ and $ \frac{\tilde{v}^{(n)}_{n-1}}{\|\tilde{v}^{(n)}_{n-1} \|} \frac{v_{n-1}}{\|v_{n-1} \|}$ which appear in conjecture [**\[conjecture H3\]**]{} we have computed $\|v_{[n/2]-1}\|$, $\left|m \left( DG_1\left(\omega_{[n/2]-1},\Phi, \frac{v_{[n/2]-1}}{\| v_{[n/2]-1} \|}\right) \right) \right|$ and $\frac{\tilde{v}^{(n)}_{[n/2]-1}}{\left\|\tilde{v}^{(n)}_{[n/2]-1} \right\|} - \frac{v_{[n/2]-1}}{\left\|v_{[n/2]-1} \right\|}$. We have done this basically to avoid computing the maps $f_i^*= W^u(\Phi,{\mathcal{R}})\cap \Sigma_i$, which would require computing $W^u(\Phi,{\mathcal{R}})$ the unstable manifold. The point $f_i^*$ accumulate to the fixed point $\Phi$ when $i\rightarrow \infty$, then one should expect the same behavior of the sequences which appear in conjecture [**\[conjecture H3\]**]{} and the sequence computed in table \[taula NumSupport conj B\]. For more details on how this values are computed see section \[section applicability FLM\]. Finally we will need the following extension of hypothesis [**H2**]{} H3) : Consider a two parametric family of maps $\{c(\alpha,{\varepsilon})\}_{(\alpha,{\varepsilon})\in A}$ (with $A\subset{\mathbb{R}}^2$) satisfying [**H1**]{} and [**H2**]{} and a fixed Diophantine rotation number $\omega_0$. Consider also $\omega_n$ and $v_n$ given by (\[equation sequences directions simplified\]) and the point $\{f^*_{1}\} = W^u({\mathcal{R}},\Phi)\cap \Sigma_1$. We assume that $DG_1(\omega_{n-1}, f^*_1 ) v^{(n)}_{n-1}$ has a unique non-degenerate minimum for any $\omega_0\in \Omega$ and $n\geq 0$. Assume also that the projection of $D{\mathcal{T}}_{\omega_{k-1}} \left(f^{(n)}_{k-1} \right) \tilde{v}^{(n)}_{k-1}$ in ${\mathcal{B}}_1$ given by \[projection pi\_1\] is non zero. Let $\{c(\alpha,{\varepsilon})\}_{(\alpha,{\varepsilon})\in A}$ (with $A\subset{\mathbb{R}}^2$) be a two parametric family of q.p. forced maps satisfying [**H1**]{}, [**H2**]{} and [**H3**]{} and let $\omega_0$ be a Diophantine number. Consider the reducibility-loss directions $\alpha_n'(c,\omega_0)$ and the sequences $u_n$ and $v_n$ given by (\[equation sequences directions simplified\]). In theorem 3.6 of [@JRT11b] we proved that, if conjectures [**\[conjecture H2\]**]{} and [**\[conjecture H3\]**]{} are true, then $$\label{equation theorem reduction to dynamics of renormalization} \frac{\alpha_n'(c,\omega_0)}{ \alpha_{n-1}'(c,\omega_0)} \sim {\text{\boldmath $\delta$}}^{-1} \cdot \frac{ \displaystyle m \left( DG_1\left(\omega_{n-1}, f^*_{1} , \frac{v_{n-1}}{\|v_{n-1}\|} \right) \right) }{ \rule{0pt}{4ex} \displaystyle m \left( DG_1\left(\omega_{n-2}, f^*_1, \frac{v_{n-2}} {\|v_{n-2}\|} \right) \right) }\cdot \left\| D{\mathcal{T}}_{\omega_{n-2}}(\Psi) \frac{ v_{n-2}}{\|v_{n-2}\|} \right\|,$$ where $m$ is given by (\[equation definition of minim\]), $G_1$ by (\[equation definition G1\]), $\{f_1^*\}= W^u({\mathcal{R}},\Phi)\cap\Sigma_1$ is the intersection of the unstable manifold of ${\mathcal{R}}$ at the fixed point $\Phi$ with the manifold $\Sigma_1$ and ${\text{\boldmath $\delta$}}$ is the universal Feigenbaum constant. This reduces the asymptotic study of the ratios $\frac{\alpha_n'(c,\omega_0)} { \alpha_{n-1}'(c,\omega_0)}$ to the sequence of vectors $\frac{ v_{k}}{\|v_{k}\|}$, which are determined by the dynamics of the q.p. renormalization operator. Theoretical explanation to the first numerical observation {#subsection explanation 1st observation} ---------------------------------------------------------- Consider a two parametric family of maps $\{c(\alpha,{\varepsilon})\}_{(\alpha,{\varepsilon}) \in A}$ contained in ${\mathcal{B}}$ satisfying the hypotheses [**H1**]{}, [**H2**]{} and [**H3**]{}. Let $\omega_0$ be a Diophantine rotation number for the family. The values $\frac{\alpha'_n(\omega_0,c)}{\alpha'_{n-1}(\omega_0,c)}$ depend only on the sequences $\omega_n$ and $v_n$ given by (\[equation sequences directions simplified\]), with $v_0=\partial_{\varepsilon}c(\alpha^*,0) $ and $\alpha^*$ the parameter value for which the family intersects $W^s({\mathcal{R}},\Phi)$. The behavior of vectors $v_n$ is described by the dynamics of the following operator, $$\label{equation map linearized q.p. renormalization general} \begin{array}{rccc} L:& {\mathbb{T}}\times {\mathcal{B}}' &\rightarrow & {\mathbb{T}}\times {\mathcal{B}}' \\ \rule{0pt}{5ex} & (\omega, v) & \mapsto & \left(2\omega, \displaystyle \frac{t_{\gamma(v)} \left( D{\mathcal{T}}_\omega(\Phi) v\right)} {\|t_{\gamma(v)} \left(D{\mathcal{T}}_\omega(\Phi) v\right) \|} \right), \end{array}$$ where $\gamma$ is chosen such that $t_{\gamma(v)} \left( D{\mathcal{T}}_\omega(\Phi) v\right)$ belongs to ${\mathcal{B}}'$. In order to study numerically the map $L$ let us recall that the spaces ${\mathcal{B}}_k$ given by (\[equation spaces bbk\]) are invariant by $D{\mathcal{T}}_\omega(\Phi)$, moreover the restriction of $D{\mathcal{T}}_\omega(\Phi)$ to this space is equivalent to the map ${\mathcal{L}}_{k\omega}$ given by (\[equation maps L\_omega\]) with the space ${\mathcal{B}}_1$ identified with ${\mathcal{RH}}({\mathbb{W}}) \oplus {\mathcal{RH}}({\mathbb{W}})$. For more details see proposition 2.16 in [@JRT11a]. Let us consider the following maps: $$\label{equation map linearized q.p. renormalization} \begin{array}{rccc} \tilde{L}_1:& {\mathbb{T}}\times {\mathcal{B}}_1 &\rightarrow & {\mathbb{T}}\times {\mathcal{B}}_1 \\ \rule{0pt}{5ex} & (\omega, v) & \mapsto & \left(2\omega, \displaystyle \frac{{\mathcal{L}}_\omega v}{\|{\mathcal{L}}_\omega v\|} \right) , \end{array}$$ and $$\label{equation map linearized q.p. renormalization section} \begin{array}{rccc} L_1:& {\mathbb{T}}\times {\mathcal{B}}_1' &\rightarrow & {\mathbb{T}}\times {\mathcal{B}}_1' \\ \rule{0pt}{5ex} & (\omega, v) & \mapsto & \left(2\omega, \displaystyle \frac{{\mathcal{L}}_\omega' v}{\|{\mathcal{L}}_\omega'v\|} \right) , \end{array}$$ with $$\label{equation map LL_omega'} \begin{array}{rccc} {\mathcal{L}}_\omega': & {\mathcal{B}}_1' & \rightarrow & {\mathcal{B}}_1' \\ &\rule{0ex}{3ex} v & \mapsto & \displaystyle t_{\gamma(v)} \left( {\mathcal{L}}_\omega(v) \right) , \end{array}$$ where $\gamma(v)$ is chosen such that $t_{\gamma(v)} \left( {\mathcal{L}}_\omega(v)\right) \in {\mathcal{B}}_1' $. We have that $L$ restricted to ${\mathcal{B}}_1'$ is equivalent to the map $L_1$. When $L$ is restricted to ${\mathcal{B}}_k$ with $k\neq 1$, then it is equivalent to $\tilde{L_1}$. Actually we have that the map $L_1$ is equivalent to the map $\tilde{L_1}$ after applying the rotational symmetry reduction described in section \[subsection rotational symmetry reduction\]. ![Several planar projections of the attractor of the map (\[equation map linearized q.p. renormalization\]). Form left to right, and top to bottom we have the projections in the coordinates $(\omega,x_0)$, $(\omega,y_0)$, $(x_0,y_0)$, $(\omega,x_2)$, $(\omega,y_2)$, $(x_2,y_2)$, $(\omega,x_4)$, $(\omega,y_4)$ and $(x_4,y_4)$. []{data-label="figure projections of q.p. renormalization 1"}](AtrQPRen12.png "fig:"){width="5cm"} ![Several planar projections of the attractor of the map (\[equation map linearized q.p. renormalization\]). Form left to right, and top to bottom we have the projections in the coordinates $(\omega,x_0)$, $(\omega,y_0)$, $(x_0,y_0)$, $(\omega,x_2)$, $(\omega,y_2)$, $(x_2,y_2)$, $(\omega,x_4)$, $(\omega,y_4)$ and $(x_4,y_4)$. []{data-label="figure projections of q.p. renormalization 1"}](AtrQPRen13.png "fig:"){width="5cm"} ![Several planar projections of the attractor of the map (\[equation map linearized q.p. renormalization\]). Form left to right, and top to bottom we have the projections in the coordinates $(\omega,x_0)$, $(\omega,y_0)$, $(x_0,y_0)$, $(\omega,x_2)$, $(\omega,y_2)$, $(x_2,y_2)$, $(\omega,x_4)$, $(\omega,y_4)$ and $(x_4,y_4)$. []{data-label="figure projections of q.p. renormalization 1"}](AtrQPRen23.png "fig:"){width="5cm"} ![Several planar projections of the attractor of the map (\[equation map linearized q.p. renormalization\]). Form left to right, and top to bottom we have the projections in the coordinates $(\omega,x_0)$, $(\omega,y_0)$, $(x_0,y_0)$, $(\omega,x_2)$, $(\omega,y_2)$, $(x_2,y_2)$, $(\omega,x_4)$, $(\omega,y_4)$ and $(x_4,y_4)$. []{data-label="figure projections of q.p. renormalization 1"}](AtrQPRen14.png "fig:"){width="5cm"} ![Several planar projections of the attractor of the map (\[equation map linearized q.p. renormalization\]). Form left to right, and top to bottom we have the projections in the coordinates $(\omega,x_0)$, $(\omega,y_0)$, $(x_0,y_0)$, $(\omega,x_2)$, $(\omega,y_2)$, $(x_2,y_2)$, $(\omega,x_4)$, $(\omega,y_4)$ and $(x_4,y_4)$. []{data-label="figure projections of q.p. renormalization 1"}](AtrQPRen15.png "fig:"){width="5cm"} ![Several planar projections of the attractor of the map (\[equation map linearized q.p. renormalization\]). Form left to right, and top to bottom we have the projections in the coordinates $(\omega,x_0)$, $(\omega,y_0)$, $(x_0,y_0)$, $(\omega,x_2)$, $(\omega,y_2)$, $(x_2,y_2)$, $(\omega,x_4)$, $(\omega,y_4)$ and $(x_4,y_4)$. []{data-label="figure projections of q.p. renormalization 1"}](AtrQPRen45.png "fig:"){width="5cm"} ![Several planar projections of the attractor of the map (\[equation map linearized q.p. renormalization\]). Form left to right, and top to bottom we have the projections in the coordinates $(\omega,x_0)$, $(\omega,y_0)$, $(x_0,y_0)$, $(\omega,x_2)$, $(\omega,y_2)$, $(x_2,y_2)$, $(\omega,x_4)$, $(\omega,y_4)$ and $(x_4,y_4)$. []{data-label="figure projections of q.p. renormalization 1"}](AtrQPRen16.png "fig:"){width="5cm"} ![Several planar projections of the attractor of the map (\[equation map linearized q.p. renormalization\]). Form left to right, and top to bottom we have the projections in the coordinates $(\omega,x_0)$, $(\omega,y_0)$, $(x_0,y_0)$, $(\omega,x_2)$, $(\omega,y_2)$, $(x_2,y_2)$, $(\omega,x_4)$, $(\omega,y_4)$ and $(x_4,y_4)$. []{data-label="figure projections of q.p. renormalization 1"}](AtrQPRen17.png "fig:"){width="5cm"} ![Several planar projections of the attractor of the map (\[equation map linearized q.p. renormalization\]). Form left to right, and top to bottom we have the projections in the coordinates $(\omega,x_0)$, $(\omega,y_0)$, $(x_0,y_0)$, $(\omega,x_2)$, $(\omega,y_2)$, $(x_2,y_2)$, $(\omega,x_4)$, $(\omega,y_4)$ and $(x_4,y_4)$. []{data-label="figure projections of q.p. renormalization 1"}](AtrQPRen67.png "fig:"){width="5cm"} ![Several spatial projections of the attractor of the map (\[equation map linearized q.p. renormalization\]). In the left hand side of the picture we have a plot (from top to bottom) of the projections in the coordinates $(\omega,x_0,y_0)$, $(\omega,x_2,y_2)$ and $(\omega,x_4,y_4)$. In the right hand side there are displayed the image of the left side projections taking a map that embeds the solid torus in ${\mathbb{R}}^3$ (see the text for more details).[]{data-label="figure projections of q.p. renormalization 2"}](AtrQPRen231.png "fig:"){width="7.5cm"} ![Several spatial projections of the attractor of the map (\[equation map linearized q.p. renormalization\]). In the left hand side of the picture we have a plot (from top to bottom) of the projections in the coordinates $(\omega,x_0,y_0)$, $(\omega,x_2,y_2)$ and $(\omega,x_4,y_4)$. In the right hand side there are displayed the image of the left side projections taking a map that embeds the solid torus in ${\mathbb{R}}^3$ (see the text for more details).[]{data-label="figure projections of q.p. renormalization 2"}](AtrQPRen231B.png "fig:"){width="7.5cm"} ![Several spatial projections of the attractor of the map (\[equation map linearized q.p. renormalization\]). In the left hand side of the picture we have a plot (from top to bottom) of the projections in the coordinates $(\omega,x_0,y_0)$, $(\omega,x_2,y_2)$ and $(\omega,x_4,y_4)$. In the right hand side there are displayed the image of the left side projections taking a map that embeds the solid torus in ${\mathbb{R}}^3$ (see the text for more details).[]{data-label="figure projections of q.p. renormalization 2"}](AtrQPRen451.png "fig:"){width="7.5cm"} ![Several spatial projections of the attractor of the map (\[equation map linearized q.p. renormalization\]). In the left hand side of the picture we have a plot (from top to bottom) of the projections in the coordinates $(\omega,x_0,y_0)$, $(\omega,x_2,y_2)$ and $(\omega,x_4,y_4)$. In the right hand side there are displayed the image of the left side projections taking a map that embeds the solid torus in ${\mathbb{R}}^3$ (see the text for more details).[]{data-label="figure projections of q.p. renormalization 2"}](AtrQPRen451B.png "fig:"){width="7.5cm"} ![Several spatial projections of the attractor of the map (\[equation map linearized q.p. renormalization\]). In the left hand side of the picture we have a plot (from top to bottom) of the projections in the coordinates $(\omega,x_0,y_0)$, $(\omega,x_2,y_2)$ and $(\omega,x_4,y_4)$. In the right hand side there are displayed the image of the left side projections taking a map that embeds the solid torus in ${\mathbb{R}}^3$ (see the text for more details).[]{data-label="figure projections of q.p. renormalization 2"}](AtrQPRen671.png "fig:"){width="7.5cm"} ![Several spatial projections of the attractor of the map (\[equation map linearized q.p. renormalization\]). In the left hand side of the picture we have a plot (from top to bottom) of the projections in the coordinates $(\omega,x_0,y_0)$, $(\omega,x_2,y_2)$ and $(\omega,x_4,y_4)$. In the right hand side there are displayed the image of the left side projections taking a map that embeds the solid torus in ${\mathbb{R}}^3$ (see the text for more details).[]{data-label="figure projections of q.p. renormalization 2"}](AtrQPRen671B.png "fig:"){width="7.5cm"} We can use the discretization of ${\mathcal{L}}_\omega$ described in Appendix \[section numerical computation of the specturm of L\_omega\] to study numerically the maps $\tilde{L}_1$ and $L_1$. Let us focus first on the case concerning $\tilde{L}_1$. Given $v\in {\mathcal{B}}_1 = {\mathcal{RH}}({\mathbb{W}}) \oplus {\mathcal{RH}}({\mathbb{W}})$, consider the coordinates of $v=(x,y)$ given by this splitting. Following the discretization described in Appendix \[section numerical computation of the specturm of L\_omega\] each function $x\in {\mathcal{RH}}({\mathbb{W}})$ can be approximated by the vector $(x_0,x_1,x_2,...,x_N)\in {\mathbb{R}}^{N+1}$ where $x_i$ is the $i$-th coefficient of the Taylor expansion for $x$ around $0$. This also holds for $y$, the second component of $v$. With this procedure each element $v$ in ${\mathcal{B}}_1$ can be approximated by a vector $(x_0,x_1,...,x_N,y_0,...,y_N)$ in ${\mathbb{R}}^{2(N+1)}$. Following the same argument we can approximate $\tilde{L}_1: {\mathbb{T}}\times {\mathcal{B}}_1 \rightarrow {\mathbb{T}}\times {\mathcal{B}}_1$ by a map $\tilde{L}_1^{(N)} : {\mathbb{T}}\times {\mathbb{R}}^{2(N+1)} \rightarrow {\mathbb{T}}\times {\mathbb{R}}^{2(N+1)}$. We can use the discretized map $L_1^{(N)}$ to study the dynamics of $\tilde{L}_1$. Given an initial point $v_0=(x^0,y^0)\neq 0$, we have iterated the point by the map a certain transient $N_1$ and then we have plotted the following $N_2$ iterates. Figures \[figure projections of q.p. renormalization 1\] and \[figure projections of q.p. renormalization 2\] show different projections of the attracting set. The values taken for this discretization are $N=30$, $N_1=2000$ and $N_2=80000$. We have displayed the coordinates corresponding to the first even Taylor coefficients of the functions $x$ and $y$. The odd Taylor coefficients obtained were all equal to zero. This last observation suggest that the attractor is contained in the set of even functions (note that the subspace of ${\mathcal{B}}_1$ consisting of all the even functions is invariant by ${\mathcal{L}}_\omega$). The same computations have been done for bigger values of $N$ and the results are the same. This indicates that the set obtained is stable with respect to the order of discretization, therefore it is reasonable to expect that it is close to the true attracting set of the original system. Let us remark that we have not made explicit the initial values of $w_0$ and $v_0$ taken for the computations. Indeed, the results seem to be independent of these values. We have repeated this computation taking as initial value of $v_0$ all the elements of the canonical base of the discretized space ${\mathbb{R}}^{2(N+1)}$ and we have always obtained the same results. We have also repeated the computations for several values of $\omega_0$ obtaining always the same results. Given a solid torus ${\mathbb{T}}\times {\mathbb{D}}_\rho$, with ${\mathbb{D}}_\rho$ the disk of radius $\rho$ in ${\mathbb{C}}$, we have that the map $f(\theta,x+i y) = (\cos(\theta) (x+K_0), \sin(\theta)(x+K_0), y)$ embeds this torus in ${\mathbb{R}}^3$, for any $K_0>\rho$. This embedding can be used for a better visualization of the spatial projections of the attracting set of the map (\[equation map linearized q.p. renormalization section\]). In the right hand side of figure \[figure projections of q.p. renormalization 2\] we have plotted the image by the embedding of the points on the left hand side. The concrete values of $K_0$ taken are (from above to below) $2$, $3/4$ and $3/20$. The numerical approximation of the attractor displayed in figures \[figure projections of q.p. renormalization 1\] and \[figure projections of q.p. renormalization 2\] reveal the rotational symmetry of the attractor. This is the rotational symmetry described in section \[subsection rotational symmetry reduction\]. ![Several planar projections of the section of attractor of the map (\[equation map linearized q.p. renormalization section\]). Form left to right, and top to bottom we have the projections in the coordinates $(\omega,y_0)$, $(\omega,x_2)$, $(\omega,y_2)$, $(x_2,y_2)$, $(\omega,x_4)$, $(\omega,y_4)$ and $(x_4,y_4)$. []{data-label="figure projections of q.p. renormalization 3"}](AtrQPRP13.png){width="5cm"} ![Several planar projections of the section of attractor of the map (\[equation map linearized q.p. renormalization section\]). Form left to right, and top to bottom we have the projections in the coordinates $(\omega,y_0)$, $(\omega,x_2)$, $(\omega,y_2)$, $(x_2,y_2)$, $(\omega,x_4)$, $(\omega,y_4)$ and $(x_4,y_4)$. []{data-label="figure projections of q.p. renormalization 3"}](AtrQPRP14.png "fig:"){width="5cm"} ![Several planar projections of the section of attractor of the map (\[equation map linearized q.p. renormalization section\]). Form left to right, and top to bottom we have the projections in the coordinates $(\omega,y_0)$, $(\omega,x_2)$, $(\omega,y_2)$, $(x_2,y_2)$, $(\omega,x_4)$, $(\omega,y_4)$ and $(x_4,y_4)$. []{data-label="figure projections of q.p. renormalization 3"}](AtrQPRP15.png "fig:"){width="5cm"} ![Several planar projections of the section of attractor of the map (\[equation map linearized q.p. renormalization section\]). Form left to right, and top to bottom we have the projections in the coordinates $(\omega,y_0)$, $(\omega,x_2)$, $(\omega,y_2)$, $(x_2,y_2)$, $(\omega,x_4)$, $(\omega,y_4)$ and $(x_4,y_4)$. []{data-label="figure projections of q.p. renormalization 3"}](AtrQPRP45.png "fig:"){width="5cm"} ![Several planar projections of the section of attractor of the map (\[equation map linearized q.p. renormalization section\]). Form left to right, and top to bottom we have the projections in the coordinates $(\omega,y_0)$, $(\omega,x_2)$, $(\omega,y_2)$, $(x_2,y_2)$, $(\omega,x_4)$, $(\omega,y_4)$ and $(x_4,y_4)$. []{data-label="figure projections of q.p. renormalization 3"}](AtrQPRP16.png "fig:"){width="5cm"} ![Several planar projections of the section of attractor of the map (\[equation map linearized q.p. renormalization section\]). Form left to right, and top to bottom we have the projections in the coordinates $(\omega,y_0)$, $(\omega,x_2)$, $(\omega,y_2)$, $(x_2,y_2)$, $(\omega,x_4)$, $(\omega,y_4)$ and $(x_4,y_4)$. []{data-label="figure projections of q.p. renormalization 3"}](AtrQPRP17.png "fig:"){width="5cm"} ![Several planar projections of the section of attractor of the map (\[equation map linearized q.p. renormalization section\]). Form left to right, and top to bottom we have the projections in the coordinates $(\omega,y_0)$, $(\omega,x_2)$, $(\omega,y_2)$, $(x_2,y_2)$, $(\omega,x_4)$, $(\omega,y_4)$ and $(x_4,y_4)$. []{data-label="figure projections of q.p. renormalization 3"}](AtrQPRP67.png "fig:"){width="5cm"} ![Several spatial projections of the intersection of the attractor of the map (\[equation map linearized q.p. renormalization section\]). The figures in the left correspond to the projection to the coordinates $(\omega,x_2,y_2)$ (top) and $(\omega,x_4,y_4)$ (bottom). In the right hand side there are displayed the image of the left side projections taking a map that embeds the solid torus in ${\mathbb{R}}^3$ (see the text for more details).[]{data-label="figure projections of q.p. renormalization 4"}](AtrQPRP451.png "fig:"){width="7.5cm"} ![Several spatial projections of the intersection of the attractor of the map (\[equation map linearized q.p. renormalization section\]). The figures in the left correspond to the projection to the coordinates $(\omega,x_2,y_2)$ (top) and $(\omega,x_4,y_4)$ (bottom). In the right hand side there are displayed the image of the left side projections taking a map that embeds the solid torus in ${\mathbb{R}}^3$ (see the text for more details).[]{data-label="figure projections of q.p. renormalization 4"}](AtrQPRP451B.png "fig:"){width="7.5cm"} ![Several spatial projections of the intersection of the attractor of the map (\[equation map linearized q.p. renormalization section\]). The figures in the left correspond to the projection to the coordinates $(\omega,x_2,y_2)$ (top) and $(\omega,x_4,y_4)$ (bottom). In the right hand side there are displayed the image of the left side projections taking a map that embeds the solid torus in ${\mathbb{R}}^3$ (see the text for more details).[]{data-label="figure projections of q.p. renormalization 4"}](AtrQPRP671.png "fig:"){width="7.5cm"} ![Several spatial projections of the intersection of the attractor of the map (\[equation map linearized q.p. renormalization section\]). The figures in the left correspond to the projection to the coordinates $(\omega,x_2,y_2)$ (top) and $(\omega,x_4,y_4)$ (bottom). In the right hand side there are displayed the image of the left side projections taking a map that embeds the solid torus in ${\mathbb{R}}^3$ (see the text for more details).[]{data-label="figure projections of q.p. renormalization 4"}](AtrQPRP671B.png "fig:"){width="7.5cm"} We can use the discretization described in Appendix \[section numerical computation of the specturm of L\_omega\] to approximate numerically the dynamics of $L_1$, as we have just done in the case of $\tilde{L}_1$. For the numerical simulation of the operator, we have taken $\theta_0=0$ and $x_0=0$. In this case the set ${\mathcal{B}}'_1(0,0)$ is identified in ${\mathbb{R}}^{2(N+1)}$ with the half hyperplane $$\{(x,y)\in {\mathbb{R}}^{2(N+1)} \thinspace | \thinspace x_0=0 \text{ and } y_0> 0\},$$ where $x_0$ and $y_0$ are respectively the first components of $x$ and $y$. In figures \[figure projections of q.p. renormalization 3\] and \[figure projections of q.p. renormalization 4\] there are displayed different projections of the attracting set obtained iterating the map $L_1$. As before we have also considered the map $f(\theta,x+i y) = (\cos(\theta) (x+K_0), \sin(\theta)(x+K_0), y)$ which embeds the solid torus ${\mathbb{T}}\times {\mathbb{D}}_\rho$ in ${\mathbb{R}}^3$ for a better visualization of the set. This time the values of $K_0$ have been taken equal to $1/2$ (above) and $3/25$ (below). Note that the different projections of the attracting set displayed in figure \[figure projections of q.p. renormalization 3\] keep a big resemblance with the plots of the dyadic solenoid displayed in figure 5 of [@Mil85]. Indeed we believe that the attractor is the inclusion of a dyadic solenoid in ${\mathcal{B}}_1'$. For more details on the definition and the dynamics of the solenoid map see [@BT11; @KH96; @Mil85; @Sma67]. To explain this fact, let us introduce a new conjecture, this time on the operator $L_1$. ![Numerical approximation of the spectrum of ${\mathcal{L}}_\omega'$ with respect to the parameter $\omega$. From left to right we have the real part, the imaginary part and the modulus of the first eight eigenvalues of ${\mathcal{L}}_\omega'$ with respect to $\omega$. []{data-label="figure espectrum of L_omega’"}](VapsProjR.pdf "fig:"){width="5cm"} ![Numerical approximation of the spectrum of ${\mathcal{L}}_\omega'$ with respect to the parameter $\omega$. From left to right we have the real part, the imaginary part and the modulus of the first eight eigenvalues of ${\mathcal{L}}_\omega'$ with respect to $\omega$. []{data-label="figure espectrum of L_omega’"}](VapsProjI.pdf "fig:"){width="5cm"} ![Numerical approximation of the spectrum of ${\mathcal{L}}_\omega'$ with respect to the parameter $\omega$. From left to right we have the real part, the imaginary part and the modulus of the first eight eigenvalues of ${\mathcal{L}}_\omega'$ with respect to $\omega$. []{data-label="figure espectrum of L_omega’"}](VapsProjM.pdf "fig:"){width="5cm"} \[conjecture H4\] There exist an open set $V\subset {\mathcal{B}}_1'$ (independent of $\omega$) such that the second component of the map $L_1$ given by (\[equation map linearized q.p. renormalization section\]) is contractive (with the supremum norm) in the unit sphere and it maps the set $V$ into itself for any $\omega \in {\mathbb{T}}$. Additionally we will assume that the contraction is uniform for any $\omega \in {\mathbb{T}}$, in the sense that there exists a constant $0<\rho<1$ such that the Lipschitz constant associated to the second component of the map $L_1'$ is upper bounded by $\rho$ for any $\omega\in {\mathbb{T}}$. A good reason to think that conjecture [**\[conjecture H4\]**]{} is true resides in the spectrum of the operator ${\mathcal{L}}'_\omega$. In figure \[figure espectrum of L\_omega’\] we have a numerical approximation of this spectrum with respect to $\omega$ as a parameter. For this computation we have followed the same procedure that we used for the computation of the spectrum of ${\mathcal{L}}_\omega$, for details see Appendix \[section numerical computation of the specturm of L\_omega\]. Looking at figure \[figure espectrum of L\_omega’\] we can observe that there exists a dominant eigenvalue (which is plotted in a dashed line) that does not cross the rest, which varies “nicely” with respect to $\omega$. Then for each value of $\omega$, the normalization of ${\mathcal{L}}'_\omega$ is a contraction in the sphere, with the eigenvector associated to the dominant eigenvalue as a fixed point. This means that conjecture [**\[conjecture H4\]**]{} is true “point-wise”, but this is not enough because the domain of contractivity might depend on $\omega$. Let us justify now why conjecture [**\[conjecture H4\]**]{} would explain the numerical results obtained for the attractor of $L_1$. Consider the set ${\mathbb{T}}\times V \subset {\mathbb{T}}\times {\mathcal{B}}'_1(x_0,\theta_0)$. If the conjecture [**\[conjecture H4\]**]{} is true, then we would have that the set $v$ invariant by the map $L_1$. More concretely we would have that the set would be expanded to the double of its length on the ${\mathbb{T}}$ direction and contracted in the ${\mathcal{B}}_1'$ direction. Assume that this transformation is done in such a way that $L_1$ maps the set ${\mathbb{T}}\times V$ inside itself but without self intersections. When we consider the intersection of $L_1({\mathbb{T}}\times V)$ with a set of the type $\{\gamma_0\}\times V$ (for some $\gamma_0\in T$) the section is conformed by two different sets without self intersections. The subsequent images by $L_1$ we would have (for each leaf $\{\gamma_0\}\times V$) the double of components than in the previous step, each of them strictly contracted in the ${\mathcal{B}}'_1$ component and contained in the previous set. Note that the described process is completely analogous to the geometric construction of a dyadic solenoid, but this time contained inside the Banach space ${\mathcal{B}}'_1$ instead of the solid torus. Therefore conjecture [**\[conjecture H4\]**]{} would give an explanation to the numerical approximation of the attracting set of $L_1$ obtained before. Consider $\{c_1(\alpha,{\varepsilon})\}$ and $\{c_2(\beta,{\varepsilon})\}$ two different families of two parametric maps satisfying hypotheses [**H1**]{}, [**H2’**]{} and [**H3**]{}. Since the family of maps satisfy hypothesis [**H2’**]{}, we have that $\partial_{\varepsilon}c_i(\alpha^*,0)$ belongs to ${\mathcal{B}}_1$ for $i=1,2$. Therefore the dynamics of $L$ (\[equation map linearized q.p. renormalization general\]) coincide with the dynamics of $L_1$ (\[equation map linearized q.p. renormalization section\]). Theorem 3.10 in [@JRT11b] asserts the following. Consider $\{c_1(\alpha,{\varepsilon})\}$ and $\{c_2(\beta,{\varepsilon})\}$ two different families of two parametric maps satisfying hypotheses [**H1**]{} and [**H2’**]{}. Let $\alpha^*$ and $\beta^*$ be the parameter values where each family $c_1(\alpha,0)$ and $c_2(\beta,0)$ intersects $W^s({\mathcal{R}},\Phi)$, the stable manifold of the fixed point of the renormalization operator. Let $ \operatorname{Rot}(V) = \left\{ v \in {\mathcal{B}}_1 \thinspace | \thinspace t_\gamma(v) \in V \subset {\mathcal{B}}_1' \text{ for some } \gamma \in {\mathbb{T}}\right\} $ where $V$ is the set given by conjecture [**\[conjecture H4\]**]{} for some $\gamma\in{\mathbb{T}}$. If $\partial_{\varepsilon}c_i(\alpha^*,0)$ belongs to $\operatorname{Rot}(V)$ for $i=1,2$ and the conjectures [**\[conjecture H2\]**]{}, [**\[conjecture H3\]**]{} and [**\[conjecture H4\]**]{} are true, then (for any $\omega_0\in \Omega$) we have that $$\label{equation thm proof universality conjecture} \frac{\alpha_n'(\omega_0,c_1)}{ \alpha_{n-1}'(\omega_0,c_1)} \sim \frac{\alpha_n'(\omega_0,c_2)}{ \alpha_{n-1}'(\omega_0,c_2)},$$ where $\alpha_i'(\omega_0, c_i)$ are the reducibility loss directions associated to each family $c_i$ for the rotation number of the system equal to $\omega_0$. Theoretical explanation to the second numerical observation {#subsection explanation 2nd observation} ----------------------------------------------------------- As before consider a two parametric family of maps $\{c(\alpha,{\varepsilon})\}_{(\alpha,{\varepsilon}) \in A}$ satisfying hypothesis [**H1**]{}, [**H2’**]{} and [**H3’**]{} and let $\alpha'_n(\omega, c)$ denote the slope of one of the curves of the reducibility loss bifurcation associated to the $2^n$ periodic invariant curve of the family. In theorem 3.13 of [@JRT11b] we prove that the second numerical observation done in section \[subsection intro numerical evidences\] can be explained as a consequence of the universal behavior (\[equation thm proof universality conjecture\]). One of the hypothesis to prove this requires $\alpha'_i(\omega,c)/ \alpha'_i(2\omega,c)$ to be a bounded sequence. The boundedness of this sequence can be obtained on its turn as a consequence of the following conjecture on the operator ${\mathcal{L}}_\omega$. Consider ${\mathcal{L}}_\omega: {\mathcal{RH}}({\mathbb{W}}_{\rho}) \oplus {\mathcal{RH}}({\mathbb{W}}_{\rho})\rightarrow {\mathcal{RH}}({\mathbb{W}}_{\rho}) \oplus {\mathcal{RH}}({\mathbb{W}}_{\rho})$ the map given by (\[equation maps L\_omega\]) and $\omega_0 \in \Omega$. Given $v_{0,1}$ and $v_{0,2}$ two vectors in ${\mathcal{RH}}({\mathbb{W}}_{\rho}) \oplus {\mathcal{RH}}({\mathbb{W}}_{\rho})\setminus\{0\}$, consider the sequences $$\label{equation sequences directions counterexample 0} \begin{array}{rcll} \omega_k & = & 2 \omega_{k-1} & \text{ for } k = 1, ..., n-1.\\ \rule{0ex}{4ex} v_{k,1} & = & {\mathcal{L}}_{\omega_{k-1}} \left(u_{k-1,1} \right) & \text{ for } k = 1, ..., n-1.\\ \rule{0ex}{4ex} v_{k,2} & = & {\mathcal{L}}_{2\omega_{k-1}}\left(u_{k-1,2} \right) & \text{ for } k = 1, ..., n-1. \end{array}$$ \[conjecture H5\] Then, there exist constants $C_1$ and $C_2$ such that $$C_1 \frac{\|v_{0,2} \|}{\|v_{0,1}\|} \leq \frac{\|v_{n,2} \|} {\|v_{n,1}\|} \leq C_2 \frac{\|v_{0,2} \|}{\|v_{0,1}\|}$$ for any $n\geq0$. ![ We have $\frac{\|v_{0,1}\|}{\|v_{0,2} \|} \frac{\|v_{n,2} \|} {\|v_{n,1}\|}$ with respect to $n$, from $n=0$ to $n=20$ in the left and from $n=0$ to $n=2 \times 10^4$ in the right.[]{data-label="figure normes contraexemple"}](Nor_raons3.pdf "fig:"){width="7.5cm"} ![ We have $\frac{\|v_{0,1}\|}{\|v_{0,2} \|} \frac{\|v_{n,2} \|} {\|v_{n,1}\|}$ with respect to $n$, from $n=0$ to $n=20$ in the left and from $n=0$ to $n=2 \times 10^4$ in the right.[]{data-label="figure normes contraexemple"}](Nor_raons_long.pdf "fig:"){width="7.5cm"} This conjecture can be interpreted as a uniform growth condition on ${\mathcal{L}}_\omega$. To support this conjecture we have computed numerically the iterates (\[equation sequences directions counterexample 0\]), in order to estimate the growth of $\|v_{n,1} \|$ with respect to the growth of $\|v_{n,2}\|$. In figure \[figure normes contraexemple\] we have plotted the ratios $\frac{\|v_{0,1}\|}{\|v_{0,2} \|} \frac{\|v_{n,2} \|} {\|v_{n,1}\|}$ with respect to $n$ for the sequence (\[equation sequences directions counterexample 0\]) with $v_{0,1}= 1 $ and $v_{0,2} = 1$. For other initial vectors we obtain the same behavior. This suggests that conjecture [**\[conjecture H5\]**]{} is true. Given $\{c(\alpha,{\varepsilon})\}$ a two parametric maps satisfying hypotheses [**H1**]{}, [**H2’**]{} and [**H3**]{} let $\alpha^*$ be the parameter values for which the family $\{c(\alpha,0)\}_{(\alpha,0)\in A}$ intersects $W^s({\mathcal{R}},\Phi)$. Let $ \operatorname{Rot}(V) = \left\{ v \in {\mathcal{B}}_1 \thinspace | \thinspace t_\gamma(v) \in V \subset {\mathcal{B}}_1' \text{ for some } \gamma \in {\mathbb{T}}\right\} $ where $V$ is the set given by conjecture [**\[conjecture H4\]**]{} for some $\gamma\in{\mathbb{T}}$. Assume that $\partial_{\varepsilon}c(\alpha^*,0) \in \operatorname{Rot}(V)$. In corollary 3.14 of [@JRT11b] we prove that, for any $\omega_0\in \Omega$, then $$\label{equation coro univ implies renor} \lim_{n\rightarrow \infty} \frac{\alpha_n'(\omega_0,c)}{ \alpha_{n-1}'(2\omega_0,c)}$$ exists, which explains the second numerical observation of section \[subsection intro numerical evidences\]. Theoretical explanation to the third numerical observation {#subsection explanation 3rd observation} ---------------------------------------------------------- In sections \[subsection explanation 1st observation\] and \[subsection explanation 2nd observation\] we focussed the discussion on the asymptotic behavior for families satisfying hypothesis [**H2’**]{}. The map considered in the third numerical observation of section \[subsection intro numerical evidences\] is an example of a map not satisfying [**H2’**]{} for which equations (\[equation thm proof universality conjecture\]) and (\[equation coro univ implies renor\]) do not hold. Let $\{c(\alpha,{\varepsilon})\}_{(\alpha,{\varepsilon}) \in A}$ be a two parametric family of maps satisfying hypothesis [**H1**]{}, [**H2**]{} and [**H3**]{}. Let $\alpha'_n(\omega, c)$ denote slope of the reducibility loss bifurcation associated to the $2^n$ periodic invariant curve of the family. Finally consider $\omega_0$ a Diophantine rotation number for the family. Let $\alpha^*$ be the parameter value for which $\{c(\alpha,0)\}_{(\alpha,0)\in A}$ intersects $W^s(\Phi,{\mathcal{R}})$. The main difference between the example considered in the third numerical observation of section \[subsection intro numerical evidences\] and the previous two is that $$\partial_{\varepsilon}c(\alpha^*,0) = v_{0,1} + v_{0,2} \text{ with } v_{0,i} \in {\mathcal{B}}_i, \quad i=1,2,$$ where the spaces ${\mathcal{B}}_i$ are given by (\[equation spaces bbk\]). More concretely, for the numerical example cited above we have considered $$\label{eq: initial vectors} v_{0,1}= f_1(x) \cos(\theta), \quad v_{0,2} = \eta f_2(x) \cos(2\theta).$$ As this family depends on $\eta$, we denote by $c_\eta$ this concrete family. The parameter $\eta$ is considered in addition to the parameters $\alpha$ and ${\varepsilon}$ of the family. In other words, for each $\eta\geq 0$, $c_\eta$ is a two parametric family of maps. Then one has that equation (\[equation theorem reduction to dynamics of renormalization\]) still holds with $v_n= v_{n,1} + v_{n,2}$, with $v_{n,1}$ and $v_{n,2}$ the sequence given by (\[equation sequences directions counterexample 0\]) and $v_{0,1}$ and $v_{0,2}$ given by (\[eq: initial vectors\]). In this case there is no universal behavior because the sequence $v_n$ lives in a bigger invariant space, where the renormalization operator is not contractive. On the other hand, the numerical computations in [@JRT11p2] suggest that the sequence $\alpha_n'(\omega_0,c_\eta)/ \alpha_{n-1}'(\omega_0,c_\eta)$ (for $\eta>0$) is not asymptotically equivalent to $\alpha_n'(\omega_0,c_0)/ \alpha_{n-1}'(\omega_0,c_0)$, but both sequences are $\eta$-close to each other. This can be explained as a consequence conjecture [**\[conjecture H5\]**]{} where we conjecture uniform growth (in norm) of the sequences $v_{n,1}$ and $v_{n,2}$. For more details see theorem 3.20 in [@JRT11b]. Applicability to the Forced Logistic Map {#section applicability FLM} ======================================== The theory exposed in sections \[section review q-p renor\] and \[section A not yet rigorous explanation\] have been built as a response to the observations done in the study of the Forced Logistic Map (see [@JRT11p; @JRT11p2]). In this section we discuss the applicability of the quasi-periodic renormalization theory to the Forced Logistic Map. In the cited papers we considered two different version of the FLM, which correspond to the map (\[q.p- family\]) with either $h_{\alpha,{\varepsilon}}(\theta,x) = \alpha x (1-x) \cos(2\pi \theta)$ or $h_{\alpha,{\varepsilon}}(\theta,x) = \cos (2 \pi \theta)$, where the parameters $(\alpha, {\varepsilon}) \in [0,4] \times [0,1]$. Notice that these two forms of the FLM do not satisfy the requirements of the theory developed in the previous sections because the family of maps does not belong to ${\mathcal{B}}$. This problem can be easily solved as follows. For $\alpha >2$ we can consider the affine change of variables given by $y = a x + b$, with $a= \frac{4}{\alpha-2}$ and $b= - \frac{2}{\alpha-2}$. If we apply this change of variables to the family (\[q.p- family\]) when $h_{\alpha,{\varepsilon}}(\theta,x) = \alpha x (1-x) \cos(2\pi \theta)$ we obtain the following family, $$\label{FLM new set up} \left. \begin{array}{rcl} \bar{\theta} & = & \theta + \omega ,\\ \bar{y} & = & \alpha \displaystyle \left( \frac{\alpha}{\alpha -2} - \frac{\alpha(\alpha-2)}{4} y^2 \right) (1+ {\varepsilon}\cos(2\pi \theta)) - \frac{2}{\alpha -2} . \end{array} \right\}$$ If we apply the same change of variables when $h_{\alpha,{\varepsilon}}(\theta,x) = \cos(2\pi \theta)$ we obtain this other family $$\label{FLM B new set up} \left. \begin{array}{rcl} \bar{\theta} & = & \theta + \omega ,\\ \bar{y} & = & \displaystyle 1 - \frac{\alpha(\alpha-2)}{4} y^2 + \frac{4{\varepsilon}}{\alpha -2} \cos(2\pi \theta). \end{array} \right\}$$ Although the change of variables considered depends on the parameter $\alpha$, the parameter space of the maps (\[FLM new set up\]) and (\[FLM B new set up\]) is the same as the parameter space of the map (\[q.p- family\]) (for the corresponding value of $h$). Then any conclusion drawn on the parameter space of the map (\[FLM new set up\]) (respectively (\[FLM B new set up\])) extends automatically to the parameter space of (\[q.p- family\]). With this new set up, we have that both families of maps belong to ${\mathcal{B}}$ for $\alpha \in (2,4)$ and ${\varepsilon}$ small enough. One should check that the FLM satisfies hypotheses [**H1**]{} and [**H2’**]{}. To check [**H1**]{} one should check that the one dimensional Logistic Map intersects transversally $W^s(\Phi,{\mathcal{R}})$ the stable manifold of the renormalization operator. This is an implicit assumption when one uses the renormalization operator to explain the universality observed for the Logistic Map. The only proof (to our knowledge) of this fact is the one given by Lyubich (theorem 4.11 of [@Lyu99]). This proof is done in the space of quadratic-like germs, which is a smaller space than the one considered here. Hypothesis [**H2’**]{} is trivial to check for the maps (\[FLM new set up\]) and (\[FLM B new set up\])). On the other hand, note that theorem 3.8 in [@JRT11a] not only gives the existence of reducibility loss bifurcations, but it also gives an explicit expression of its slopes in term of the renormalization operator (the ones given by formulas (\[equation alpha n +\]) and (\[equation alpha n -\])). Actually we have given even a more explicit formula in corollary 3.13 in [@JRT11a]. We used these formulas to compute the reducibility directions $\alpha_n'(\omega)$ of the Forced Logistic Map (\[FLM new set up\]). The initial values $\alpha_n$ have been computed numerically, by means of a Newton method applied to their invariance equation. Differentiating on formula (\[FLM new set up\]) (respectively (\[FLM B new set up\])) it is easy to write the values of $f^{(n)}_{0}, u^{(n)}_{0}$ and $v^{(n)}_{0}$ in terms of $\alpha_n$. Then using the discretization of the operator done in Appendix \[section numerical computation of the specturm of L\_omega\], we can compute numerically the iterates $f^{(n)}_{k}, u^{(n)}_{k}$ and $v^{(n)}_{k}$, for $k=1,...,n-1$. Once we have these functions, we can evaluate them to compute the values of $\alpha'_n(\omega) $ given by formula ($48$) of [@JRT11a]. We have used this procedure to compute the values of $\alpha'_n(\omega)$ for the map (\[FLM new set up\]) (consequently also for the map (\[q.p- family\]) with $h_{\alpha,{\varepsilon}}(\theta,x) = \alpha x (1-x) \cos(2\pi \theta)$) and $\omega_0 = \frac{\sqrt{5}-1}{2}$. The results are shown in table \[taula DifDirMin\]. The values $ \alpha'_n(\omega)$ have been also computed in [@JRT11p2] via a completely different procedure, based on a continuation method with extended precision. More concretely these values are displayed in table $2$ of the cited paper. The third and fourth columns of table \[taula DifDirMin\] display the discrepancies between both computation, in absolute ($\bar{\epsilon}_a$) and relative ($\bar{\epsilon}_r$) terms. This experiment has been repeated for other values of $\omega$ and also for the map (\[FLM B new set up\]). In all cases we obtained that the slopes computed with both methods are the same up to similar accuracies to the ones displayed in table \[taula DifDirMin\]. This supports the correctness of both computations and at the same time the conjecture [**\[conjecture H2\]**]{}, which has been assumed to be true to derive the formula used for this estimation. Numerical computation of the spectrum of ${\mathcal{L}}_\omega$ {#section numerical computation of the specturm of L_omega} =============================================================== In this appendix we present the numerical method that we used to discretize ${\mathcal{L}}_\omega$ and to study its spectrum numerically. Our method is a slight modification of the one introduced by Lanford in [@Lan82] (see also [@Lan92]). Let ${\mathbb{D}}(z_0, \rho )$ be the complex disc centered on $z_0$ with radius $\rho$. Consider ${\mathcal{RH}}({\mathbb{D}}(z_0, \rho ))$ the space of real analytic functions such that they are holomorphic on ${\mathbb{D}}(z_0, \rho )$ and continuous on it closure. Given a function $f\in {\mathcal{RH}}({\mathbb{D}}(z_0, \rho ))$, we can consider the following modified Taylor expansion of $f$ around $z_0$, $$\label{modified taylor expansion} f(z) = \sum_{k=0}^{\infty} f_i \left(\frac{z-z_0}{\rho}\right)^i.$$ The truncation of a Taylor series at order $N$ induces a projection defined as $$\begin{array}{rccc} p_{(N)}: & {\mathcal{RH}}({\mathbb{D}}_{\rho}) & \rightarrow & {\mathbb{R}}^{N+1} \\ & f & \mapsto & (f_0, f_1, \dots,f_N) , \end{array}$$ On the other hand we have its pseudo-inverse by the left $$\begin{array}{rccc} i_{(N)}: & {\mathbb{R}}^{N+1} & \rightarrow & {\mathcal{RH}}({\mathbb{D}}_{\rho}) \\ & (f_0, f_1, \dots,f_N)& \mapsto & \displaystyle \sum_{k=0}^{N} f_i \left(\frac{z-z_0}{\rho}\right)^i , \end{array}$$ in the sense that $i_{(N)} \circ p_{(N)} $ is the identity on ${\mathbb{R}}^{N+1}$. Note also that both maps are linear. Given a map $T:{\mathcal{RH}}({\mathbb{D}}(z_0, \rho ))\rightarrow {\mathcal{RH}}({\mathbb{D}}(z_0, \rho ))$, we can approximate it by its discretization $T^{(N)}: {\mathbb{R}}^{N+1} \rightarrow {\mathbb{R}}^{N+1}$ defined as $T^{(N)}:= p_{(N)} \circ T \circ i_{(N)}$. More concretely if $T$ is a linear bounded operator, we can compute the eigenvalues of the operator $T^{(N)}$ in order to study the spectrum of $T$. In general the eigenvalues of $T^{(N)}$ might have nothing to do with the spectrum of $T$. For example an infinite-dimensional operator does not need to have eigenvalues, but a finite-dimensional one will always have the same number of eigenvalues (counted with multiplicity) as the dimension of the space. For this reason will do some numerical test on the results obtained with this discretization. At this point consider the map ${\mathcal{L}}_{\omega}: {\mathcal{RH}}({\mathbb{W}}) \oplus{\mathcal{RH}}({\mathbb{W}}) \rightarrow {\mathcal{RH}}({\mathbb{W}}) \oplus {\mathcal{RH}}({\mathbb{W}})$ defined by equation (\[equation maps L\_omega\]). If we set ${\mathbb{W}}={\mathbb{D}}(z_0,\rho)$ we can use the method described above in each component of ${\mathcal{RH}}({\mathbb{W}}) \oplus {\mathcal{RH}}({\mathbb{W}})$ to discretize ${\mathcal{L}}_\omega$ and approximate it by a map ${\mathcal{L}}_{\omega}^{(N)} : {\mathbb{R}}^{2(N+1)} \rightarrow {\mathbb{R}}^{2(N+1)}$. Concretely in our computation we have taken $z_0=\frac{1}{5}$ and $\rho = \frac{3}{2}$. In figure \[figure inclusion boundaries\] we include graphical evidence that the set ${\mathbb{W}}= {\mathbb{D}}\left( \frac{1}{5}, \frac{3}{2} \right)$ satisfies [**H0**]{}, in section \[section definition and basic properties\] can be found more details about this. In table \[table eigenvaules golden\] we have the first $24$ eigenvalues of ${\mathcal{L}}_{\omega}^{(N)}$ for $N=100$ and $\omega=\frac{\sqrt{5}-1}{2}$. The eigenvalues have been sorted by their modulus, from bigger to smaller. Note that the eigenvalues of the discretized operator also satisfy the properties given in section \[section The Fourier expansion of DT\]. To justify the validity of these eigenvalues we have done the following numerical tests. ![Estimation of the errors and the radii of convergence of the first twenty-four eigenvalues of ${\mathcal{L}}_{\omega}$ for $\omega=\frac{\sqrt{5}-1}{2}$ with respect to the order of the discretization. See the text for more details. []{data-label="Figure errors and radis golden"}](Err_vaps_fins24.pdf "fig:"){width="7.5cm"} ![Estimation of the errors and the radii of convergence of the first twenty-four eigenvalues of ${\mathcal{L}}_{\omega}$ for $\omega=\frac{\sqrt{5}-1}{2}$ with respect to the order of the discretization. See the text for more details. []{data-label="Figure errors and radis golden"}](Rad_vaps_fins24.pdf "fig:"){width="7.5cm"} Consider that we have a real eigenvalue of multiplicity two, or a pair of complex eigenvalues[^3] which is persistent for different values of $N$ (the order of the discretization). The first test done to the eigenvalues is to check if the distance between the associated eigenvectors decreases when $N$ is increased. In the left part of figure \[Figure errors and radis golden\] we have the distance between the eigenvectors associated to the same eigenvalue of the operators ${\mathcal{L}}^{(N)}_\omega$ and ${\mathcal{L}}^{(110)}_\omega$ as a graph of $N$, with $N$ varying from $40$ to $100$. We have plotted this distance for the first twenty-four eigenvalues. To compute the distance between eigenvectors we have estimated the supremum norm of the difference between the real function represented by each of the vectors, in other words we have computed $\|i^{(N)}(v^{(N)}) - i^{(110)}(v^{(110)}) \|_\infty$ in the interval $(z_0-\rho,z_0+\rho)={\mathbb{W}}\cap {\mathbb{R}}$. Note that, since the distance goes to zero this indicates that the eigenvectors, namely $v^{(N)}$, converge to a limit $v^*$. One should expect these eigenvalues to be in the spectrum of ${\mathcal{L}}_\omega$, but nothing ensures that $v^*$ belongs to the domain of ${\mathcal{L}}_\omega$. We have done a second test on the reliability of the approximated eigenvectors, where we check this condition. Let us remark that with the numerical computations done so far, we have only checked that the eigenvectors as elements of $ {\mathcal{RH}}({\mathbb{D}}(z_0,\rho))\oplus {\mathcal{RH}}({\mathbb{D}}(z_0,\rho)) $ to converge in the segment $(z_0-\rho,z_0+\rho)\subset {\mathbb{R}}$ instead of checking the convergence in the whole set ${\mathbb{D}}(z_0,\rho)$. Let us give evidences that approximate eigenvectors obtained with our computations have a domain of analicity containing ${\mathbb{D}}(z_0,\rho)$. Consider that we have a function $f$ holomorphic in a domain of the complex plane containing ${\mathbb{D}}(z_0,\rho)$. If the we consider the expansion of $f$ given by equation (\[modified taylor expansion\]), we have that $r$ the radius of convergence of the series around $z_0$ is given as $$r= \frac{\rho}{\limsup_{n \rightarrow \infty} \sqrt[n]{|f_n|} } .$$ With the discretization considered here we have an approximation of the terms $f_n$, hence these can be used to compute a numerical estimation of $r$. Consider $v$ an eigenvector of the operator ${\mathcal{L}}_\omega$. We have that $v=(v_1,v_2) \in {\mathcal{RH}}({\mathbb{W}})\oplus {\mathcal{RH}}({\mathbb{W}}) $. Given $v_1^{(N)}= (v_1^{(N)}, v_2^{(N)})$ a numerical approximation of the eigenvector, we can use the procedure described above to estimate the radius of convergence of each $v_1$ and $v_2$. We have done this for the eigenvectors associated to each of the first twenty-four eigenvalues of ${\mathcal{L}}_\omega$ with $\omega= \frac{\sqrt{5}-1}{2}$ (keeping only the smaller of the two radius obtained). The results are displayed on the right part of figure \[Figure errors and radis golden\], where the estimated radius has been plotted with respect to $N$, the order of the discretization. Note that the estimations give a radius bigger than $\rho=\frac{3}{2}$, which indicates that the eigenvectors are analytic in ${\mathbb{D}}(z_0,\rho)$, and continuous on its closure, for $z_0=\frac{1}{5}$ and $\rho=\frac{3}{2}$. ![Estimation of the distance between eigenvectors with the same eigenvalue (left) and estimation of the radios of convergence (right) of the first twenty-four eigenvalues of ${\mathcal{L}}_{\omega}$ with respect to $\omega$. See the text for more details. []{data-label="Figure errors and radis omega"}](Errors12_90_100.pdf "fig:"){width="7.5cm"} ![Estimation of the distance between eigenvectors with the same eigenvalue (left) and estimation of the radios of convergence (right) of the first twenty-four eigenvalues of ${\mathcal{L}}_{\omega}$ with respect to $\omega$. See the text for more details. []{data-label="Figure errors and radis omega"}](Radis24_90_100.pdf "fig:"){width="7.5cm"} Up to this point, we have considered $\omega$ fixed to $\frac{\sqrt{5}-1}{2}$, but the same computations can be done to study the spectrum of ${\mathcal{L}}_\omega$ with respect to the parameter $\omega$. In figure \[Operator Spectrum\] of section \[section The Fourier expansion of DT\] we have plotted the first twenty-four eigenvalues of the map with respect to $\omega$. The set ${\mathbb{T}}$ has been discretized in a equispaced grid of $1280$ points. Recall that the operator ${\mathcal{L}}_{\omega}$ depends analytically on $\omega$ (proposition 2.20 in [@JRT11a]), therefore the spectrum also does (as long as the eigenvalues do not collide, see theorems III-6.17 and VII-1.7 in [@Kat66]). The numerical results agree with this analytic dependence. For this computation we have also made the same test as before to the eigenvalues. The results of these tests are shown in figure \[Figure errors and radis omega\]. To estimate the convergence of the eigenvectors we have compared the eigenspaces of the eigenvalues of ${\mathcal{L}}_\omega^{(90)}$ with the eigenspaces associated to ${\mathcal{L}}_\omega^{(100)}$ for each value of $\omega$ in the cited grid of points on ${\mathbb{T}}$. The estimation of the radius of convergence has been also done with respect to $\omega$ for $N=90$. We have plotted the estimated error and convergence radius for the first twenty-four eigenvalues in the same figure. Both result indicate that the eigenvalues obtained are reliable. [40]{} K. Bjerkl[ö]{}v. S[NA]{}’s in the quasi-periodic quadratic family. , 286(1):137–161, 2009. H. W. Broer and F. Takens. , volume 172 of [*Applied Mathematical Sciences*]{}. Springer, New York, 2011. W. de Melo and S. van Strien. , volume 25 of [*Ergebnisse der Mathematik und ihrer Grenzgebiete (3) \[Results in Mathematics and Related Areas (3)\]*]{}. Springer-Verlag, Berlin, 1993. J.-P. Eckmann and P. Wittwer. A complete proof of the [F]{}eigenbaum conjectures. , 46(3-4):455–475, 1987. M. J. Feigenbaum. Quantitative universality for a class of nonlinear transformations. , 19(1):25–52, 1978. M. J. Feigenbaum. The universal metric properties of nonlinear transformations. , 21(6):669–706, 1979. U. Feudel, S. Kuznetsov, and A. Pikovsky. , volume 56 of [*World Scientific Series on Nonlinear Science. Series A: Monographs and Treatises*]{}. World Scientific Publishing Co. Pte. Ltd., Hackensack, NJ, 2006. Dynamics between order and chaos in quasiperiodically forced systems. J. Figueras and A. Haro. Computer assisted proofs of the existence of fiberwise hyperbolic invariant tori in skew products over rotations. To appear, 2010. J. F. Heagy and S. M. Hammel. The birth of strange nonchaotic attractors. , 70(1-2):140–153, 1994. T. H. Jaeger. Quasiperiodically forced interval maps with negative schwarzian derivative. , 16(4):1239–1255, 2003. A. Jorba, P. Rabassa, and J.C. Tatjer. Period doubling and reducibility in the quasi-periodically forced logistic map. Preprint available at [http://arxiv.org]{}, 2011. A. Jorba and J. C. Tatjer. A mechanism for the fractalization of invariant curves in quasi-periodically forced 1-[D]{} maps. , 10(2-3):537–567, 2008. T. Kato. . Springer Verlag, 1966. A. Katok and B. Hasselblatt. . Cambridge University Press, 1996. O. E. Lanford, III. A computer-assisted proof of the [F]{}eigenbaum conjectures. , 6(3):427–434, 1982. O. E. Lanford, III. Computer assisted proofs. In [*Computational methods in field theory ([S]{}chladming, 1992)*]{}, volume 409 of [*Lecture Notes in Phys.*]{}, pages 43–58. Springer, Berlin, 1992. M. Lyubich. Feigenbaum-[C]{}oullet-[T]{}resser universality and [M]{}ilnor’s hairiness conjecture. , 149(2):319–420, 1999. J. Milnor. On the concept of attractor. , 99(2):177–195, 1985. A. Prasad, V. Mehra, and R. Ramaskrishna. Strange nonchaotic attractors in the quasiperiodically forced logistic map. , 57(2):1576–1584, 1998. P. Rabassa. . PhD thesis, Universitat de Barcelona, 2010. P. Rabassa, A. Jorba, and J.C. Tatjer. umerical evidences of universality and self-similarity in the forced logistic map. Preprint available at [http://arxiv.org]{}, 2011. P. Rabassa, A. Jorba, and J.C. Tatjer. Towards a renormalization theory for quasi-periodically forced one dimensional maps [I]{}. [E]{}xistence of reducibily loss bifurcations. Preprint available at [http://arxiv.org]{}, 2011. P. Rabassa, A. Jorba, and J.C. Tatjer. Towards a renormalization theory for quasi-periodically forced one dimensional maps [II]{}. [A]{}symptotic behavior of reducibility loss bifurcations. Preprint available at [http://arxiv.org]{}, 2011. S. Smale. Differentiable dynamical systems. , 73:747–817, 1967. D. Sullivan. Bounds, quadratic differentials, and renormalization conjectures. In [*American [M]{}athematical [S]{}ociety centennial publications,[V]{}ol. [II]{} ([P]{}rovidence, [RI]{}, 1988)*]{}, pages 417–466. Amer. Math. Soc., Providence, RI, 1992. C. Tresser and P. Coullet. Itérations d’endomorphismes et groupe de renormalisation. , 287(7):A577–A580, 1978. [^1]: This work has been supported by the MEC grant MTM2009-09723 and the CIRIT grant 2009 SGR 67. P.R. has been partially supported by the PREDEX project, funded by the Complexity-NET: [www.complexitynet.eu]{}. [^2]: Here $W^s(\Phi,{\mathcal{R}})$ and $W^u(\Phi,{\mathcal{R}})$ are considered as the inclusion in ${\mathcal{B}}$ of the stable and the unstable manifolds of the fixed point $\Phi$ (given by [**H0**]{}) by the map ${\mathcal{R}}$ in the topology of ${\mathcal{B}}_0$ (the inclusion of one parametric maps in ${\mathcal{B}}$). [^3]: Note that there are no simple eigenvalues due to corollary 2.18 in [@JRT11a]

ArXiv

--- abstract: 'Ultraslow radiative cooling lifetimes and adiabatic detachment energies for three astrochemically relevant anions, C$_n^-$ ($n=3-5$), are measured using the Double ElectroStatic Ion Ring ExpEriment (DESIREE) infrastructure at Stockholm University. DESIREE maintains a background pressure of $\approx$10$^{-14}$mbar and temperature of $\approx$13K, allowing storage of mass-selected ions for hours and providing conditions coined a “molecular cloud in a box". Here, we construct two-dimensional (2D) photodetachment spectra for the target anions by recording photodetachment signal as a function of irradiation wavelength and ion storage time (seconds to minute timescale). Ion cooling lifetimes, which are associated with infrared radiative emission, are extracted from the 2D photodetachment spectrum for each ion by tracking the disappearance of vibrational hot-band signal with ion storage time, giving $\frac{1}{e}$ cooling lifetimes of 3.1$\pm$0.1s (C$_3^-$), 6.8$\pm$0.5s (C$_4^-$) and 24$\pm$5s (C$_5^-$). Fits of the photodetachment spectra for cold ions, i.e. those stored for at least 30s, provides adiabatic detachment energies in good agreement with values from laser photoelectron spectroscopy. Ion cooling lifetimes are simulated using a Simple Harmonic Cascade model, finding good agreement with experiment and providing a mode-by-mode understanding of the radiative cooling properties. The 2D photodetachment strategy and radiative cooling modeling developed in this study could be applied to investigate the ultraslow cooling dynamics of wide range of molecular anions.' author: - 'James N. Bull' - 'Michael S. Scholz' - Eduardo Carrascosa - 'Moa K. Kristiansson' - Gustav Eklund - Najeeb Punnakayathil - Nathalie de Ruette - Henning Zettergren - 'Henning T. Schmidt' - Henrik Cederquist - 'Mark H. Stockett' bibliography: - 'cnm.bib' title: 'Ultraslow radiative cooling of C$_n^-$ ($n=3-5$)' --- Introduction ============ Which molecular anions exist in space? What are their formation mechanisms and life cycles? These are two long-standing questions in astrochemistry.[@Larson:2012; @Millar:2017] Prior to a decade and a half ago, H$^{-}$ was the only anion thought to play a prominent role in the interstellar medium (ISM). In 2006, the first molecular anion, C$_{6}$H$^{-}$, was discovered by comparing astronomical line spectra with gas-phase action spectra recorded in the laboratory.[@C6m] Over the next four years there were five further identifications: ,[@C4hm] ,[@C8hm] ,[@CNm] ,[@C3Nm] and .[@C5Nm] Vibrationally excited was also detected alongside .[@C5Nm] However, despite increasing interest in the role of molecular anions in space, there has been a stall in new identifications. It is thought that the discovery of new molecular anions is thwarted by a lack of understanding of the formation mechanism(s) and dynamical properties of both the anions known to exist in the ISM and new anions yet to be assigned.[@Millar:2017; @herbst:2014] Dynamical properties in this context include electron capture cross-sections, electronic internal conversion efficiencies and couplings between dipole-bound and valence-localised states, cross-sections for neutralization reactions with cations, and the rates of radiative cooling.[@Carelli:2013; @Roueff:2013] As an example of the need for reliable measurements of the dynamical properties of astrochemically relevant anions, in a discussion on radiative electron attachment (which involves formation of a vibrationally excited ground state ion that must cool) Herbst[@Herbst:2009] remarked “*The discovery of molecular anions \[in space\] has generated the need to include their formation and destruction in chemical models ... the larger \[carbonaceous\] molecular anions detected ($n$ = 6, 8) have higher abundances relative to their neutral precursors because the radiative attachment rate increases with the number of degrees of freedom of the anion. However, their rate estimates are quite uncertain and experimental studies are highly welcome.*" Although gas-phase action spectroscopies can provide data on electronic transitions and detachment energies for carbonaceous anions,[@Gerlich:2006] their dynamical properties such as infrared (IR) radiative cooling lifetimes are more difficult to measure because hot anions need to be isolated (i.e. free from collision) for periods of milliseconds to minutes. These conditions are not attainable using conventional ion traps. Here, we used the Double ElectroStatic Ion Ring ExpEriment (DESIREE) infrastructure at Stockholm University to characterize the radiative cooling lifetimes and adiabatic detachment energies (ADEs) of C$_n^-$ ($n=3-5$) by monitoring the intensity of vibrational hot bands near the electron detachment threshold with ion storage time. The present investigation targeted the C$_n^-$ ($n=3-5$) species because they are likely ISM anions. In particular, they possess similar bonding and electronic structure to the molecular anions already known to exist in the ISM, they could be formed through either photodissociation or dissociative electron attachment mechanisms,[@Larson:2012; @Millar:2017] and neutral C$_{3}$ and C$_{5}$ are known interstellar molecules.[@Hinkle:1988; @Bernath:1989] Although anions are unlikely to be significant astrochemical species in ‘photon-dominated regions’ (PDRs, e.g. diffuse clouds) due to facile destruction by photodetachment with visible and ultraviolet light,[@Millar:2017; @Khamesian:2016] the abundance of anions in dark clouds (e.g. C$_{6}^{-}$ in L1527)[@Sakai:2007] has been shown to reach nearly 10% of that for the corresponding neutral molecule, suggesting that negative charge in photon-free regions of space is more likely in the form of anions than free electrons. Cold dark molecular clouds have temperatures of 10–20K.[@Bergin:2007] The normal operating temperature of DESIREE ($\approx$13K) is squarely within this range. [@Schmidt2017] In the absence of collisional quenching, the spontaneous cooling dynamics of hot molecular anions such as C$_n^-$ ($n=3-5$) can be divided into three time ($t$) regimes: - Statistical regime I ($t\leq$10$^{-3}$s): Internal energy is high, e.g. several electron-volts above the thresholds for dissociation, thermionic emission,[@Zhao:1996] and recurrent/Poincaré fluorescence.[@Leger:1988] The energy threshold for thermionic emission is the ADE, which for C$_n^-$ ($n=3-5$) is $\approx$2–4eV, and the energy threshold for dissociation can be approximated by the lowest bond dissociation energy ($\approx$3–4eV). Both thermionic emission and dissociation lead to destruction of the anion. Recurrent fluorescence, which is known for C$_4^-$ and C$_6^-$,[@Ito:2014; @Kono2015; @Ebara2016; @Kono:2018] involves inverse internal conversion to an electronic excited state (situated below the ADE, e.g. $\approx$2–3eV) followed by radiative emission.[@Shiromaru:2015] - Slow regime II ($t\approx$10$^{-3}$–1s): Internal energy is in the vicinity of the lowest thresholds for the cooling mechanisms important for Regime I, e.g. $\sim$2eV. In addition to slow cooling through these mechanisms, radiative emission due to vibrational transitions becomes important. - Ultraslow regime III ($t\gtrsim$1s): Internal energy is below the thresholds for dissociation, thermionic emission and recurrent fluorescence. Ions cool only through radiative vibrational (IR) transitions and rotational (microwave) transitions, with vibrational cooling occurring much faster than rotational cooling.[@Schmidt2017; @Meyer2017] Cooling dynamics in this regime have been explored for only a few small anions due to technical challenges associated with isolating ions for durations extending to minutes and maintaining low background temperatures.[@vh2016; @RICE2017; @Chartkunchand2016; @Connor2016; @Hansen2017; @Anderson:2018] While the target anions have been intensively studied by several groups in recent years, all studies used room-temperature electrostatic ion storage rings or beam traps and were limited to characterizing cooling dynamics occurring on sub-second timescales.[@Naaman2000; @Goto2013; @Chandrasekaran2014; @Kono2015; @Saha2017; @Kono:2018] In the present study, we have used the DESIREE infrastructure to investigate the cooling dynamics of the target anions on the ultraslow, $t\,\gtrsim$1s timescale. Our strategy involved storing ions for up to $\approx$1minute and using photodetachment spectroscopy to monitor the intensity of hot bands with ion storage time, providing an indirect characterization of cooling lifetimes. The $\frac{1}{e}$ ion cooling lifetimes, which are attributed to IR radiative emission, are well-described by a simple harmonic cascade model of this process. Fits of the cold photodetachment spectra associated with ions stored for at least 30s to the Wigner threshold law demonstrate an alternative, cryogenic method for obtaining ADE values. ![The symmetric ion storage ring in DESIREE. Neutral particles formed by photodetachment in the lower straight region (crossed-beam geometry with the optical parametric oscillator, OPO) are detected in their forward directions with the ‘Imaging Detector’. In another set of measurements on C$_3^-$, light from a cw dye laser was merged colinearly with the stored ion beam in the upper straight section. The wavelength of the cw light was tuned to be in resonance with the vibrational hot band of C$_{3}^{-}$ (620.5nm). The ring circumference is 8.6m and each straight section has a length of 0.96m.[]{data-label="fig:desiree"}](DESIREE.pdf){width="0.99\columnwidth"} Methods ======= DESIREE is a cryogenic dual electrostatic ion storage ring facility located at the Department of Physics, Stockholm University.[@Thomas2011; @Schmidt:2013] The major components constituting the so-called ‘symmetric’ storage ring are schematically illustrated in FIG.\[fig:desiree\]. The interior of the ring is cooled to $\approx$13 K by compressed helium refrigerators and is isolated from external thermal radiation by several layers of insulation.[@Schmidt2017] Vacuum is maintained at a background pressure of $\approx$10$^{-14}$mbar using cryopumping combined with turbomolecular pumps and oil-free backing pumps. These ultrahigh vacuum conditions allow storage of keV ion beams for hours.[@Backstrom2015] In the present experiments, the target anions \[C$_n^-$ ($n=3-5$)\] were produced using a caesium sputtering ion source with a graphite cathode.[@Schmidt:2013] This process generates ions with a high degree of rovibrational excitation, i.e. source-heated ions. The nascent ions were accelerated to 10 keV, selected according to their mass-to-charge ratio using a bending magnet, and injected into the symmetric ion storage ring. Transport from the source to the ring takes $\approx$100$\mu$s. The $\frac{1}{e}$ beam storage lifetimes were measured at 540$\pm$30s for C$_3^-$ and 570$\pm$30s for C$_5^-$ (see Supporting Information). Although the beam storage lifetime for C$_4^-$ was not measured in this study, we expect a similar lifetime to those for C$_3^-$ and C$_5^-$. These beam storage lifetimes are limited by loss of ions through collisions with background gas.[@Schmidt:2013] One-color experiments --------------------- In the one-color experiments, stored ions were irradiated with tunable-wavelength light from an optical parametric oscillator (OPO, EKSPLA NT342B, 10Hz) using a crossed-beam geometry through one of the straight sections of the ion storage ring (see FIG.1). Any neutral particles formed through photodetachment or photodissociation are unaffected by the ring’s electrostatic steering fields and impact on a micro-channel plate (MCP) detector (‘Imaging Detector’ in FIG.\[fig:desiree\]).[@Rosen:2007] Signal from the MCP detector was gated using a 1 $\mu$s duration pulse that was slightly delayed with respect to the OPO pulse to account for the neutral particle’s flight time from the interaction region to the detector. The purpose of the gate was to eliminate signal from scattered OPO light striking the detector and to minimize background counts from collision-induced detachment events due to the residual gas consisting of $\sim$10$^4$ H$_{2}$ molecules per cm$^{3}$. The OPO wavelengths were calibrated using an optical spectrograph (Avantes AvaSpec-3648), which was itself calibrated against a wavemeter (HighFinesse WS-8) via a diode laser (632.6 nm). The irradiation wavelength was stepped in 0.5nm increments (2nm for C$_4^-$) between ion injections for a given ion storage time, providing a two-dimensional (2D) photodetachment spectrum, i.e. a series of photodetachment spectra as a function of wavelength and ion storage time (see Refs for a similar procedure applied to rotational cooling of OH$^{-}$). For a given ion, the time evolution of the photodetachment yield at a specific wavelength or range of wavelengths can be obtained by taking a wavelength slice through the 2D photodetachment spectrum. Part of our interpretation applied Principal Component Analysis (PCA)[@pca:2002] to the 2D photodetachment spectrum for each ion. PCA is a common statistical procedure that decomposes a multi-demensional data set $X$ into a set of orthogonal principal components (PCs) which are the eigenvectors of the covariance matrix $X^TX$. The eigenvalues associated with each PC relate to the fraction of the variation in $X$ that is explained by each PC and the principal values (PVs, the projection of $X$ on its PCs) give the wieght of each PC as a function of time. In the present case, the PCs may be thought of as the underlying spectra that describe the evolution of the photodetachment spectra with ion storage time, with a time invariant background due to photodetachment signal from cold ions (or nearly time invariant because of a finite ion beam storage lifetime).[@Stockett2019] The cooling lifetimes obtained from the PVs should be considered wavelength-averaged values since each probe wavelength provides slightly different ion cooling lifetime due to a distribution of internal vibrational energies in the stored ion beam. Deplete-probe experiments with C$_3^-$ -------------------------------------- Deplete-probe experiments on source-heated C$_3^-$ were performed by adapting the procedure recently described by Schmidt*et al.*[@Schmidt2017], where the effect of the depletion laser was to preferentially photodetach rotationally excited ions and thus reduce the measured ion cooling lifetimes. Depletion involved intercepting the stored ion beam with 620.5nm light from a cw laser (Coherent 899 ring dye laser) using a merged-beam geometry in the straight section of the ion storage ring opposite the OPO light interaction region (see FIG.\[fig:desiree\]). The depletion laser wavelength (620.5nm or 1.998eV) was chosen to be close to the ADE from the present measurements (see below) because the photodetachment cross-section for vibrationally excited (hot band) C$_3^-$ is much larger than that for cold ions. Adiabatic detachment energies ----------------------------- The adiabatic detachment energy (ADE) for each anion was extracted from cold photodetachment spectrum assuming fit with the Wigner threshold law:[@Wigner:1948; @Farley:1989] $$\sigma_{PD} = (KE)^{L + \frac{1}{2}},$$ where $\sigma_{PD}$ is the photodetachment cross-section, $KE$ is the kinetic energy of the ejected electron (energy in excess of the ADE for a direct photodetachment process), and $L$ is the angular momentum of the outgoing electron. For the present systems which involves photodetachment from $\pi$ molecular orbital, we find that $L=2$ ($d$ wave photoelectron) provides best fit to the experimental data. The ADE is taken to be the energy at which the Wigner threshold law fit exceeds 3$\sigma$ of the baseline signal. We note the above expression is strictly valid for atomic species; best fit values of $L$ can deviate from integers for molecules – see example fit for C$_{3}^{-}$ in the Supporting Information. Radiative cooling lifetime modeling ----------------------------------- Spontaneous cooling in the present experiments is presumed to occur through IR radiative emission. A simple harmonic cascade (SHC) model was developed to interpret the experimental results. The model assumes vibrational density of states $\rho$ computed using the Beyer-Swinehart algorithm using anharmonic or scaled harmonic vibrational mode frequencies $\nu_{s}$ calculated at the $\omega$B97X-D//aug-cc-pVTZ level of theory with Q-Chem4.4 (see Supporting Information).[@wb97xd; @acct; @qchem] For a given mode $s$, the IR radiative cooling rate coefficient, assuming only transitions where $\Delta v_{s}=-1$ are allowed, with $v$ being the vibrational quantum number, is [@Chandrasekaran2014] $$k_s(E)=A_s^{10}\sum_{v=1}^{v\leq E/h\nu_s} \frac{\rho(E-vh\nu_s)}{\rho(E)}, \label{eq_k}$$ where $E$ is the energy of a given vibrational state, $h$ is Plank’s constant, and the summation is over $v$ ($v=0$ and $1$ are the ground and first excited vibrational states of mode $s$, respectively). The Einstein coefficients $A_s^{10}$ were calculated at the $\omega$B97X-D//aug-cc-pVTZ within the harmonic approximation (see Supporting Information). Starting from an initial Boltzmann distribution of vibrational energy $g(E,t=0)$ corresponding to 1000 K, the population in each level was recalculated at each simulation timestep. The model allowed for two treatments of intramolecular vibrational energy redistribution (IVR),[@Nesbitt:1996] i.e. statistical randomization of vibrational energy with time, $t$:\ \ (i) IVR is negligible or slow compared with radiative cooling – the population of each mode is explicitly tracked according to the expression below: $$\begin{aligned} g(E,t+dt) = \sum_s g(E,t)e^{-k_s(E)dt}\\ + \sum_s g(E+h\nu_s,t)(1-e^{-k_s(E+h\nu_s)dt}); \end{aligned}$$ \(ii) IVR is fast compared with radiative cooling – vibrational energy is statistically redistributed each simulation time step and the total energy emitted radiatively at each time step is:\ $$dE_{tot}/dt=-\int g(E,t)\sum_s h\nu_sk_s(E)dE,$$ where the total energy remaining in the ensemble as a function of time $E_{tot}(t)=\int Eg(E,t)dE$ was taken as an indicator of the progress of cooling. Given that the vibrational energy quanta are small and the number of stored ions is large, level occupation numbers were treated as continuous quantities. We expect that case (ii) should be most relevant for the present source-heated anions because ion cooling lifetimes are long (seconds timescale) compared with the expected timescale for IVR (nanoseconds to millisecond timescale). The SHC modeling starts from a hot ensemble and simulates the internal energy as a function of ion storage time. For case (i), the internal vibrational energy reached a non-zero asymptotic value because any population that was portioned to IR inactive modes is not emitted radiatively. For case (ii), because the lowest frequency vibrational modes for each anion are IR active, all vibrational energy in excess of the zero-point energy can be liberated and thus the model goes asymptotically to zero vibrational energy at long times. To compare results from the SHC model with experiment for which there is non-zero photodetachment signal at long ion storage times for wavelengths shorter than the ADE due to photodetachment from cold ions, it was necessary to add an asymptote offset equal to the value extracted from an exponential fit of the experimental data. Furthermore, it was found that the initial temperature assumed in the SHC model (e.g. 500–5000K) altered the cooling dynamics only on timescales much faster than those probed in the present experiments, e.g. milliseconds. It is worth noting that we found use of the commonly cited harmonic frequencies and intensities from Szczepanski *et al.*[@Szczepanski:1997] calculated at the B3LYP/6-31G\* level of theory within the SHC framework produced qualitatively similar results to that presented in this study, but required scaling the $A^{10}_{s}$ coefficients with factor 0.5 to best align modeled ion cooling lifetimes with experimental values. Results and discussion ====================== Tricarbon anion, C$_3^-$ ------------------------ Photodetachment spectra for C$_3^-$ as a function of ion storage time are shown in FIG.\[fig\_c3raw\], upper. The spectra have been divided into four time bins, with the 0–3s time bin corresponding to ions recently injected into the storage ring and the 30–57s time bin corresponding to ions that have been stored for at least 30s. The complete 2D photodetachment spectrum is shown in the Supporting Information. The time-binned photodetachment spectra show a broad feature over the expected ADE (1.99$\pm$0.025eV or 622$\pm$8nm from photoelectron spectroscopy)[@Arnold:1991] due to vibrationally excited C$_3^-$ ions that cool over the first $<$30s, providing a ‘cold’ photodetachment spectrum (30–57s spectrum). A fit of the cold photodetachment spectrum with the Wigner threshold law gave ADE = 1.990$\pm$0.005eV (623$\pm$0.6nm), which is within error of the earlier photoelectron spectroscopy determination. ![Upper: Time-binned photodetachment spectra for C$_3^-$, recorded by monitoring the yield of neutral particles with wavelength of light. The black bar represents the ADE and uncertainty determined for ions stored at least 30 s, and the orange bar corresponds to the photoelectron spectroscopy value from Ref.. Lower: Decay of photodetachment signal with ion storage time at selected probe wavelengths (note the log scale). Time constants for single-exponential fits in the lower panel are 4.3$\pm$0.8s (590nm), 4.6$\pm$0.7s (595nm), 3.8$\pm$0.4s (610nm), 3.2$\pm$0.2s (615nm) and 1.8$\pm$0.2s (625nm).[]{data-label="fig_c3raw"}](c3cool.pdf "fig:"){width="0.99\columnwidth"} ![Upper: Time-binned photodetachment spectra for C$_3^-$, recorded by monitoring the yield of neutral particles with wavelength of light. The black bar represents the ADE and uncertainty determined for ions stored at least 30 s, and the orange bar corresponds to the photoelectron spectroscopy value from Ref.. Lower: Decay of photodetachment signal with ion storage time at selected probe wavelengths (note the log scale). Time constants for single-exponential fits in the lower panel are 4.3$\pm$0.8s (590nm), 4.6$\pm$0.7s (595nm), 3.8$\pm$0.4s (610nm), 3.2$\pm$0.2s (615nm) and 1.8$\pm$0.2s (625nm).[]{data-label="fig_c3raw"}](dyyoffplotlog.pdf "fig:"){width="0.99\columnwidth"} The lower panel of FIG.\[fig\_c3raw\] shows photodetachment signal with ion storage time (1s time bins) at six selected wavelengths. Fits with a single-exponential decay curve gave lifetimes ranging from 4.3$\pm$0.8 s (590 nm) to 1.8$\pm$0.2 s (615 nm), demonstrating that cooling occurs more slowly for anions probed at shorter wavelength, i.e. those closer to the detachment threshold. This is readily interpreted in terms of EQN\[eq\_k\], which indicates that the cooling rate rapidly increases with vibrational excitation and the variation in ion cooling lifetime is due to the stored ion beam having a distribution of internal vibrational energies. ![Principal component analysis on C$_{3}^{-}$. Upper: Principal component (PC1) and cold photodetachment spectrum. Lower: Principal values for PC1 (denoted PV1) with ion storage time, and simple harmonic cascade (SHC) model of the cooling lifetime (note the log scale). The gradual decrease in PV1 for ion storage times longer than $\approx$20s is attributed to the beam storage lifetime in DESIREE.[]{data-label="fig_c3mod"}](c3pca_cal.pdf "fig:"){width="0.99\columnwidth"} ![Principal component analysis on C$_{3}^{-}$. Upper: Principal component (PC1) and cold photodetachment spectrum. Lower: Principal values for PC1 (denoted PV1) with ion storage time, and simple harmonic cascade (SHC) model of the cooling lifetime (note the log scale). The gradual decrease in PV1 for ion storage times longer than $\approx$20s is attributed to the beam storage lifetime in DESIREE.[]{data-label="fig_c3mod"}](time3pvt2.pdf "fig:"){width="0.99\columnwidth"} Principal component analysis (PCA) of the 2D photodetachment spectrum for C$_3^-$ suggested a single principal component (PC1 in FIG.\[fig\_c3mod\]) can describe the hot-band intensity with ion storage time. Nearly 80$\%$ of the variance in the 2D spectrum is explained by PC1, with the remaining PCs describe only statistical fluctuations with no secular time dependence. The principal values of PC1 (denoted PV1) with ion storage time are shown in the lower panel of FIG.\[fig\_c3mod\]. Fit of PV1 with a bi-exponential gave a fast lifetime of 3.1$\pm$0.1s, which, as expected, is intermediate between the lifetimes for the wavelength-selected cooling times in FIG.\[fig\_c3raw\]. The second lifetime ($>$200s) is much longer than the measurement cycle (60s for C$_{3}^{-}$) and is presumably associated with the beam storage lifetime (540$\pm$30s for C$_3^-$, see Supporting Information). The time-invariant cold spectrum (FIG.\[fig\_c3mod\], lower) was obtained by subtracting PC1, weighted by PV1, from the 2D photodetachment spectrum. This closely resembles the cold spectrum in FIG.\[fig\_c3raw\], but utilizes the entire data set rather than arbitrarily time-binned data. Fit of the cold spectrum from PCA with the Wigner threshold law gave an ADE of 1.987$\pm$0.004eV (624.0$\pm$0.6nm), which is within error of above determination using the cold, time-binned photodetachment spectrum. IR radiative cooling characteristics for C$_3^-$ from the SHC model are summarized in FIG.\[fig\_c3mod\], lower. The dashed black curve assumes the case of no IVR and the solid black curve includes IVR. Exponential fits to the SHC curves returned ion cooling lifetimes of 5.22$\pm$0.01 s (no IVR) and 3.68$\pm$0.06 s (including IVR). The latter is in reasonable agreement with the average ion cooling lifetime from PCA (3.1$\pm$0.1s). ![Cooling lifetimes for C$_3^-$ using the deplete-probe scheme. The black circles and red squares are cooling lifetimes with and without irradiation using cw laser light at 620.5 nm (dashed blue vertical line).[]{data-label="fig_dye"}](dyetimes.pdf){width="0.99\columnwidth"} The influence of the cw laser (620.5nm) on the cooling lifetimes of C$_3^-$ is shown in FIG.\[fig\_dye\]. Comparison of cw laser ON (black) with cw laser OFF (red) data at the probe wavelengths of 615, 610, 595 and 590nm show a systematic decrease of the ion cooling lifetimes by $\approx$1s because the photodetachment cross-section is larger for vibrationally excited ions than for cold ions at 620.5nm. No such effect was observed at 625nm, i.e. probe wavelength longer than that of the cw laser. These data provide a proof-of-principle measurement demonstrating a deplete-probe scheme to preferentially remove hot ions from the stored ion beam. The extent of depletion could likely be improved through better overlap of the cw beam with the ion beam and increase of cw laser power. Tetracarbon anion, C$_4^-$ -------------------------- ![Upper: Time-binned photodetachment spectra for C$_4^-$, recorded by monitoring the yield of neutral particles with wavelength of light. The black bar represents the ADE and uncertainty determined for ions stored at least 30 s, and the orange bar corresponds to the photoelectron spectroscopy value from Ref.. Middle: Principal component (PC1) and cold photodetachment spectrum from the 2D photodetachment spectrum of C$_4^-$. Bottom: Principal values for PC1 (denoted PV1) and simple harmonic cascade (SHC) model of the cooling lifetime (note the log scale).[]{data-label="fig_c4cool"}](c4cool.pdf "fig:"){width="0.99\columnwidth"} ![Upper: Time-binned photodetachment spectra for C$_4^-$, recorded by monitoring the yield of neutral particles with wavelength of light. The black bar represents the ADE and uncertainty determined for ions stored at least 30 s, and the orange bar corresponds to the photoelectron spectroscopy value from Ref.. Middle: Principal component (PC1) and cold photodetachment spectrum from the 2D photodetachment spectrum of C$_4^-$. Bottom: Principal values for PC1 (denoted PV1) and simple harmonic cascade (SHC) model of the cooling lifetime (note the log scale).[]{data-label="fig_c4cool"}](c4pca_cal.pdf "fig:"){width="0.99\columnwidth"} ![Upper: Time-binned photodetachment spectra for C$_4^-$, recorded by monitoring the yield of neutral particles with wavelength of light. The black bar represents the ADE and uncertainty determined for ions stored at least 30 s, and the orange bar corresponds to the photoelectron spectroscopy value from Ref.. Middle: Principal component (PC1) and cold photodetachment spectrum from the 2D photodetachment spectrum of C$_4^-$. Bottom: Principal values for PC1 (denoted PV1) and simple harmonic cascade (SHC) model of the cooling lifetime (note the log scale).[]{data-label="fig_c4cool"}](time4pvt2.pdf "fig:"){width="0.99\columnwidth"} Time-binned photodetachment spectra and results from PCA for C$_4^-$ are summarized in FIG.\[fig\_c4cool\] upper and middle/lower, respectively. The C$_4^-$ photodetachment data were recorded in larger wavelength increments compared with C$_3^-$ or C$_5^-$ due to substantially lower laser fluence from the OPO at the near-UV wavelengths needed to photodetach this species. The time-binned photodetachment spectra for C$_{4}^{-}$ indicate that hot-band signal has disappeared after $\approx$30s. Fit of the 30–55s time-binned spectrum with the Wigner threshold law gave an ADE of 3.83$\pm$0.03eV (323.7$\pm$2.5nm), which is consistent with the value from photoelectron spectroscopy (3.882$\pm$0.010eV).[@Arnold:1991] Application of PCA to the 2D photodetachment spectrum of C$_4^-$ again suggested that a single principal component (PC1 in FIG.\[fig\_c4cool\], middle) describes the variation in the hot band intensity with ion storage time. The principal value of PC1 with ion storage time (PV1 in FIG.\[fig\_c4cool\], lower) has a fitted lifetime of 6.8$\pm$0.5 s, which is roughly twice that for C$_{3}^{-}$ and comparable with the wavelength-binned values given above. Unfortunately, the data is of insufficient quality for a bi-exponential fit to account for the beam storage lifetime. As for C$_{3}^{-}$, shorter wavelengths are associated with longer ion cooling lifetimes – see Supporting Information for further details. IR radiative cooling lifetimes for C$_4^-$ from the SHC model are 6.74$\pm$0.01 s (no IVR) and 5.4$\pm$0.1 s (including IVR), with both models being in reasonable agreement with the PCA value of 6.8$\pm$0.5 s. Pentacarbon anion, C$_5^-$ -------------------------- Time-binned photodetachment spectra and PCA results for C$_5^-$ are summarized in FIG.\[fig\_c5cool\] upper and middle/lower, respectively. Fit of the 30–57s ‘cold’ time-binned spectrum with the Wigner threshold law gave ADE = 2.82$\pm$0.01eV (439.7$\pm$1.6nm), which agrees with the value from photoelectron spectroscopy (2.839$\pm$0.008eV).[@Arnold:1991] Intriguingly, the cooling behaviour presents a different situation compared with C$_3^-$ and C$_4^-$. Whereas hot band photodetachment signal at wavelengths longer than $\approx$435nm diminishes over the first few seconds of ion storage, there is an enhancement of photodetachment signal for wavelengths shorter than $\approx$435nm (i.e. above the ADE), which will be discussed soon. ![Upper: Time-binned photodetachment spectra for C$_5^-$, recorded by monitoring the yield of neutral particles with wavelength of light. The black bar represents the ADE and uncertainty determined for ions stored at least 30 s, and the orange bar corresponds to the photoelectron spectroscopy value from Ref.. Middle: Principal components (PC1 and PC2) and cold photodetachment spectrum extracted from the 2D photodetachment spectrum of C$_5^-$. Bottom: Principal values for PC1 and PC2 (denoted PV1 and PV2) and simple harmonic cascade (SHC) model of the cooling lifetime (note the log scale).[]{data-label="fig_c5cool"}](c5cool.pdf "fig:"){width="0.99\columnwidth"} ![Upper: Time-binned photodetachment spectra for C$_5^-$, recorded by monitoring the yield of neutral particles with wavelength of light. The black bar represents the ADE and uncertainty determined for ions stored at least 30 s, and the orange bar corresponds to the photoelectron spectroscopy value from Ref.. Middle: Principal components (PC1 and PC2) and cold photodetachment spectrum extracted from the 2D photodetachment spectrum of C$_5^-$. Bottom: Principal values for PC1 and PC2 (denoted PV1 and PV2) and simple harmonic cascade (SHC) model of the cooling lifetime (note the log scale).[]{data-label="fig_c5cool"}](c5pca_cal.pdf "fig:"){width="0.99\columnwidth"} ![Upper: Time-binned photodetachment spectra for C$_5^-$, recorded by monitoring the yield of neutral particles with wavelength of light. The black bar represents the ADE and uncertainty determined for ions stored at least 30 s, and the orange bar corresponds to the photoelectron spectroscopy value from Ref.. Middle: Principal components (PC1 and PC2) and cold photodetachment spectrum extracted from the 2D photodetachment spectrum of C$_5^-$. Bottom: Principal values for PC1 and PC2 (denoted PV1 and PV2) and simple harmonic cascade (SHC) model of the cooling lifetime (note the log scale).[]{data-label="fig_c5cool"}](time5pvt2.pdf "fig:"){width="0.99\columnwidth"} Ion cooling lifetimes at selected probe wavelengths are 22$\pm$3s (435nm), 7.7$\pm$0.8s (440nm), and 5.7$\pm$0.6s (445nm) – see Supporting Information for further details. As for C$_{3}^{-}$ and C$_{4}^{-}$, shorter wavelengths are associated with longer ion cooling lifetimes. Application of PCA to the 2D photodetachment spectrum of C$_5^-$ suggested that two principal components (PC1 and PC2 in FIG.\[fig\_c5cool\], middle) are necessary to describe the spectral variation with ion storage time. PC1 has a similar wavelength dependence and also principal value (PV1) with ion storage time when compared with PC1 for C$_3^-$ or C$_4^-$. PV1 was best fit with two exponential lifetimes of 1.7$\pm$0.3s and 24$\pm$5s, although there is also an unresolved long-lifetime decay associated with the beam storage lifetime. PC2 (FIG.\[fig\_c5cool\] middle) resembles a vibrational-like peak for wavelengths just shorter than the ADE wavelength. Fit of the principal values for PC2 with ion storage time (PV2 in FIG.\[fig\_c5cool\], lower) required both exponential decay and growth functions. The growth lifetime for PC2 is within error of the long decay lifetime associated with PC1 (i.e., after $\approx$10s PV1 + PV2 is roughly steady state), implying that hot band population associated with PC1 eventually contributes to PC2 at longer ion storage time. We assign PC2 to predominately the $^{1}\Sigma_{g}^{+}(\nu^{\prime} = 0)\leftarrow {^{2}\Pi}_{\frac{3}{2}}(\nu^{\prime\prime} = 0)$ detaching transition, which occurs at slightly shorter wavelength ($\approx$0.5nm) than the alternative $^{1}\Sigma_{g}^{+}(\nu^{\prime} = 0)\leftarrow {^{2}\Pi}_{\frac{1}{2}}(\nu^{\prime\prime} = 0)$ spin-orbit detaching transition.[@Arnold:1991; @Kitsopoulos:1991; @Hock:2012] Assuming this assignment is correct, it appears that as ions cool, the relative population of ground vibrational state anions increases and consequently the apparent photodetachment cross-section for resonant detaching transitions increases (presumably much more so than for C$_{3}^{-}$ and C$_{4}^{-}$). It follows that the long lifetime associated with PC2 is due to decay of population associated with the $^{1}\Sigma_{g}^{+}(\nu^{\prime}_{7} = 1)\leftarrow {^{2}\Pi}_{\frac{3}{2}}(\nu^{\prime\prime}_{7} = 1)$ and $^{1}\Sigma_{g}^{+}(\nu^{\prime}_{7} = 1)\leftarrow {^{2}\Pi}_{\frac{1}{2}}(\nu^{\prime\prime}_{7} = 1)$ hot band detaching transitions (see spectroscopic assignment of photodetaching hot band modes in Refs.). Note, because $\nu^{\prime\prime}_{7}$ is an IR inactive mode (see Supporting Information), decay must be due to IVR followed by radiative emission. IR radiative cooling characteristics for C$_5^-$ from the SHC model are summarized in FIG.\[fig\_c5cool\], lower. Neglect of IVR resulted in a cooling curve that was best fit with two lifetimes of 4.78$\pm$0.08 s and 21.7$\pm$0.1 s. Inclusion of IVR gave fitted lifetimes of 1.63$\pm$0.04 s and 22.0$\pm$0.2 s, which are in good agreement with values from PCA. The need for a bi-exponential fit for C$_{5}^{-}$ can be traced to mode-specific radiative emission processes. Specifically, the faster lifetime is dominated by emission from the main IR active mode $\nu_{9} \approx 1751$cm$^{-1}$ ($A_{9}^{10}\approx 1807$) and the slower lifetime attributed to emission from weaker modes $\nu_{1,2} \approx 127-141$cm$^{-1}$ ($A_{1,2}^{10}\approx 21-30$) – see mode-specific radiated power plots in the Supporting Information. Similar double lifetime cooling is not apparent for C$_3^-$ and C$_4^-$ because the majority of the cooling from the high frequency mode with a large $A^{10}$ coefficient occurs on a sub-second timescale (see Supporting Information). Summary and outlook =================== The present work has investigated the ultraslow cooling characteristics of three astrochemically relevant anions under conditions approximating a molecular cloud. Interestingly, an increase in molecular size leads to longer average ion cooling lifetimes: 3.1$\pm$0.1s for C$_3^-$, 6.8$\pm$0.5s for C$_4^-$ and 24$\pm$5s for C$_5^-$. Variation in ion cooling lifetimes across the hot band is attributed to a distribution of internal energies. These are the first known measurements on carbonaceous anions extending to the ultraslow (seconds) timescale; all previous measurements have been performed under room temperature conditions and were restricted to measuring the sub-second cooling dynamics. The increase in average ion cooling lifetime with molecular size can be understood by considering the point group symmetry (D$_{\infty h}$) of the anions and that E1 radiative transitions require a change in electric dipole moment. In particular, the high symmetry means that each of present anions have only three vibrational modes with $A^{10}$ coefficents larger than 10. Although $A^{10}$ coefficients for $^{1}\Sigma_{g}^{-}$ symmetric vibrational modes quickly increase with molecular size beyond C$_{5}^{-}$, for $n=3-5$ there are an increasing number of low frequency modes with increasing $n$ (mostly IR inactive) and the weakly IR active $\Pi_{u}$ symmetric vibrational modes become lower in frequency and have lower radiative emission rates (see EQN2 and Supporting Information). The net result is an increase in ion cooling lifetime with increasing $n$. These ion cooling dynamics would not be evident at room temperature for $n>3$ because the average thermal vibrational energy exceeds the energy of the low frequency modes: $\approx$342cm$^{-1}$ (C$_{4}^{-}$) and $\approx$587cm$^{-1}$ (C$_{5}^{-}$) at 298K assuming harmonic vibrational partition functions. We are presently applying the 2D photodetachment strategy with DESIREE to study the ultraslow cooling dynamics of larger carbonaceous anions to further explore these trends. These results will be the presented in a forthcoming paper. As part of the present study, we developed a simple harmonic cascade model that proved capable of simulating IR radiative emission using input data from conventional electronic structure calculations. With provision for IVR, the model was able to qualitatively reproduce the experimental ion cooling lifetimes and provide a mode-by-mode understanding of the cooling dynamics. The agreement between theory and experiment provides confidence for applying this model to anions for which experimental data is not available or difficult to measure. Finally, it should be noted that application of the present 2D photodetachment methodology to larger molecular anions may prove more complicated due to near-threshold resonant excitations. Specifically, if there are substantial cross-sections for photoexcitation of $\pi\pi^{*}$ states situated below the detachment threshold or for resonances situated in the detachment continuum, ensuing autodetachment and internal conversion dynamics might affect the observed ion cooling lifetimes and spectral features. For example, experiments have shown that photoexcitation followed by internal conversion to recover the ground electronic state is efficient for C$_n^{-}$ ($n>4$) and polycyclic aromatic hydrocarbon (PAH) anions,[@Zhao:1996; @Baguenard:2002; @menadione; @tetracene] with photoexcitation cross-sections for optically allowed transitions in PAH anions generally being much larger than cross-sections for direct photodetachment. If neutral formation through thermionic emission or dissociation processes takes longer than the time ions spend in the straight section of the ion storage ring after irradiation ($\approx$4–5$\mu$s), then neutrals formed outside of the straight section of the ion storage ring will not be counted. Fortunately, in DESIREE, the relative importance of delayed neutral formation can be ascertained by simultaneously measuring neutral yield on the detector on the opposite straight section of the ion storage ring (Glass Plate/MCP detector in FIG.\[fig:desiree\]). Acknowledgements {#acknowledgements .unnumbered} ================ This work was supported by the Swedish Research Council (grant numbers 2015-04990, 2016-03675, 2016-04181, 2018-04092) and the Swedish Foundation for International Collaboration in Research and Higher Education (STINT, grant number PT2017-7328 awarded to JNB, EC and MHS). We acknowledge the DESIREE infrastructure for provisioning of facilities and experimental support, and thank the operators and technical staff for their invaluable assistance. The DESIREE infrastructure receives funding from the Swedish Research Council under the grant number 2017-00621.

ArXiv

--- abstract: 'Approximate theories for the restricted primitive model electrolyte are compared in the light of Totsuji’s lower bound for the energy (an improvement over Onsager’s), Gillan’s upper bound for the free energy, and thermal stability requirements. Theories based on the Debye-Hückel (DH) approach and the mean spherical approximation (MSA), including extensions due to Bjerrum, Ebeling, Fisher and Levin, and Stell, Zhou, and Yeh (PMSA1, 2, 3) are tested. In the range ${T^{\ast}}= k_B T D a/q^2 \lesssim 10 \, {T^{\ast}}_c \simeq 0.5$, all DH-based theories satisfy Totsuji’s bound, while the MSA possesses a significant region of violation. Both DH and MSA theories violate Gillan’s bound in the critical region and below unless ion pairing [*and* ]{} the consequent free-ion depletion are incorporated. However, the PMSA theories, which recognize pairing but not depletion, fail to meet the bound. The inclusion of excluded-volume terms has only small effects in this respect. Finally, all the pairing theories exhibit negative constant-volume specific heats when ${T^{\ast}}\gtrsim 2 \, {T^{\ast}}_c \simeq 0.1$; this is attributable to the treatment of the association constant.' address: 'Institute for Physical Science and Technology, University of Maryland, College Park, Maryland 20742' author: - 'Daniel M. Zuckerman, Michael E. Fisher, and Benjamin P. Lee[^1]' date: 'May 8, 1997' title: Critique of Primitive Model Electrolyte Theories --- [2]{} Introduction ============ The liquid-gas phase transition in electrolytes is of current interest because of puzzling experiments and theoretical efforts to understand them. For recent reviews, see [@levelts; @fisher; @stellrev]. The primary model used is the restricted primitive model (RPM) consisting of two oppositely charged, but otherwise identical, sets of $N_+ = N_-$ hard spheres of diameter $a$ and charge per particle $\pm q$, immersed in a medium of dielectric constant $D$ (to represent the solvent) and volume $V$. We will restrict our attention to the RPM in $d=3$ dimensions and use the reduced temperature and density $$T^{\ast} = k_B T D a/q^2 \;\;\;\;\; \mbox{and} \;\;\;\;\; \rho^{\ast} = a^3 \rho,$$ where $\rho = (N_+ + N_-)/V \equiv N/V$; the Debye inverse screening length, $$\label{eqkapx} \kappa_D = ( 4 \pi q^2 \rho /D k_B T )^{1/2}, \;\; \mbox{with} \;\;\; x = \kappa_D a = \surd(4 \pi {\rho^{\ast}}/ {T^{\ast}}) \, ;$$ the reduced Helmholtz free energy density $${\bar{f}}(\rho,T) = - F_N(V;T)/V k_B T;$$ and the reduced configurational energy per particle, $u$, defined via $$\label{equdef} (Nq^2/Da)\,u(\rho,T) = U_N - \hbox{$\frac{3}{2}$} N k_B T,$$ where $F_N$ and $U_N = V k_B T^2 \, (\partial {\bar{f}}/ \partial T)$ denote the total free energy and (internal) energy. Recent theory[@fisher; @stellrev; @fl; @lf; @gandg; @evans; @stellassoc] has focussed on two approaches to approximating the free energy of the RPM, based on either Debye-Hückel (DH) theory[@debye] or the mean spherical approximation (MSA)[@msa; @wais; @gmsa]. Many years ago Bjerrum[@bj] proposed to improve DH theory by including ion pairing via “chemical association.” Later, Ebeling and Grigo[@ebgr] combined ion-pairing with an MSA expression for the ionic free energy; more recently, Levin and Fisher [@lf] and Stell and coworkers[@stellassoc] explored further extensions of the MSA. On the other hand, Fisher and Levin[@fl; @lf] supplemented DH theory not only with ion pairing and excluded-volume terms but also included the solvation free energy of the electrically active $(+,-)$ dipolar ion pairs. Currently, this class of DH-based theories seems to give the best, albeit semiquantitative, account of the RPM in the critical region as judged by comparison with simulations performed by various authors[@fisher; @lf]. It may be remarked that the simulation estimates for ${T^{\ast}}_c$ and ${\rho^{\ast}}_c$ have been changing at an alarming rate \[2(b)\]. Nevertheless, the MSA-based theories yield approximations for ${T^{\ast}}_c \, ( \, \gtrsim 0.073)$ that are significantly [*higher*]{} than those based on the DH approach (${T^{\ast}}_c \lesssim 0.056$), which in fact agrees much better with the simulations (${T^{\ast}}_c = 0.048 \mbox{-} 0.056$)[@fisher; @lf; @stellassoc; @orkpan]. At a purely theoretical level, however, one cannot be content since, [*a priori*]{}, there seem no clear grounds for preferring the DH-based theories — apart from their more direct and intuitive physical interpretation — rather than the more modern (and fashionable) MSA-based theories which — since they entail the pair correlation functions and the Ornstein-Zernike (OZ) relation — give the impression of being more firmly rooted in statistical mechanics. On the other hand, it has recently been shown that the DH theories yield pair correlations satisfying the OZ relation in a very natural way[@benf]. Furthermore, both theories have an essentially mean-field character despite which, in contrast to typical mean field theories for lattice systems, [*neither*]{} has any known Gibbs-Bogoliubov variational formulation or similar basis. How, then, might the two approaches be distinguished? Now Blum and his coworkers have, in various places[@blum; @bernard; @blumnew; @simblum], enthusiastically sung the praises of the MSA for the RPM, asserting that the theory “$\,$is asymptotically correct in the limit of high density and infinite charge” or “high screening parameter (Debye length going to zero).” Furthermore, “unlike the DH theory, it \[the MSA\] satisfies the exact Onsager bounds for the Helmholtz free energy and the internal energy”[@blum; @bernard] (in the same asymptotic limit) and the “internal energy of the MSA is an exact lower bound”[@blumnew]. As reported below, these claims cannot be sustained: however, they do suggest that one might usefully assess and compare the MSA and DH theories, and their various extensions, by checking their predictions against previously developed bounds for the internal energy and Helmholtz free energy. That task is undertaken here. Indeed, as discussed more fully in Sec. $\!\!$II, several bounds have been established. The well-known Onsager [*lower*]{} bound for the [*configurational energy*]{} of the RPM was derived in 1939 [@onsag]; less heralded is an improvement due to Totsuji some forty years later [@tot]. For the [*free energy*]{}, Rasaiah and Stell [@stell] proved that the hard-core free energy provides an [*upper*]{} bound, while Gillan [@gil] developed a much stronger upper bound embodying the idea of $(+,-)$ pairing into dipoles[@fisher; @fl; @lf; @bj; @ebgr]. Finally, we note that thermodynamic stability with respect to temperature requires the positivity of the specific heat at constant volume [@pippard]. We will focus particularly on the Totsuji and Gillan bounds applied in the region of the predicted gas-liquid phase transition and critical point. We find that DH theory and all its augmentations always [*satisfy*]{} Totsuji’s (and Onsager’s) bound provided ${T^{\ast}}\lesssim 10 {T^{\ast}}_c \simeq 0.5$. On the other hand, the MSA actually [*violates*]{} the Totsuji bound in a significant region of the $(\rho,T)$ plane where coexistence is predicted, unless the theory is suitably augmented. In the light cast by Gillan’s bound, the two approaches rest on a more equal footing. As already shown by Gillan [@gil], the MSA (in its usual form) fails badly for ${T^{\ast}}\lesssim 0.08$; but the same is true for the original DH theory (even when supplemented by excluded-volume terms [@fisher; @fl; @lf]). Only when both basic theories are augmented by ion-pairing contributions and by allowing for the associated depletion of the free-ion screening do they satisfy the Gillan bound. As against the hard-core electrostatic effects, included in [*both*]{} DH and MSA treatments, the presence or absence of specific excluded-volume terms has small effect numerically and does not affect the satisfaction of the bound. However, the recent PMSA (or pairing-MSA) theories of Stell and coworkers [@stellassoc] violate Gillan’s bound apparently because they do not account appropriately for the free-ion depletion. The main lesson is the crucial importance of the clustering of ions into dipolar pairs at low temperatures. Of course, this has been appreciated heuristically for a long time [@bj] and was quantitatively demonstrated in 1983 by Gillan [@gilsim] in calculations for the RPM which showed that the vapor for ${T^{\ast}}\lesssim 0.053$ consisted mainly of $(+,-)$, $(2+,2-)$, $(3+,3-)$, $\ldots$ neutral clusters and $(2+,1-)$, $(1+,2-)$, $(3+,2-)$, and $(2+,3-)$ singly charged clusters, with relatively far fewer free monopoles, $(+)$ and $(-)$. The present work, however, seems to be the first purely analytic demonstration of the thermodynamic necessity for including clustering, implicitly or, perhaps, explicitly, in approximate theories. The recognition of $(+,-)$ ion-pairing requires the specification of the corresponding [*association constant*]{}, $K(T)$. Ever since Bjerrum’s original proposal [@bj], this has been a matter of confusion and contention (see, e.g., [@fisher; @fl]). Nevertheless, in the low temperature region of principal interest here, say ${T^{\ast}}\lesssim 0.08 \simeq 1.5 \, {T^{\ast}}_c$, Bjerrum’s cutoff form and Ebeling’s more sophisticated expression agree to within 1.8% or better [@fl; @lf; @bj; @ebgr] and, along with other cutoff forms, have identical asymptotic expansions in powers of ${T^{\ast}}$ [@lf; @fuoss]. For practical purposes, therefore, $K(T)$ might be regarded as known “exactly.” At higher temperatures, where pairing should be (and is predicted to be) much weaker, it is natural to surmise that different treatments of association would prove inconsequential. However, this proves false! Indeed, for all the previous pairing theories [@fl; @lf; @evans; @stellassoc; @bj; @ebgr] we find that the constant-volume configurational specific heat becomes [*negative*]{} (violating thermodynamics [@pippard] and statistical mechanics) in the region ${T^{\ast}}= 0.1 \; \mbox{to} \; 0.5$: see Sec. IV. The source of this serious problem is found in the proposed behavior of the association constant. Initial steps towards amelioration are indicated, but the issue will be pursued in more detail elsewhere [@assocpaper]. It should be mentioned that we also examine the generalized MSA (GMSA) [@evans; @gmsa] and variants of the MSA thermodynamics derived from the (approximate) pair correlation functions by routes other than the standard energy equation [@wais]: these are discussed in Sec. III. Other even less realistic models for electrolytes exist, including the one-component plasma with hard cores [@palmer] and the corresponding “dense-point limit” \[11(c)\]; however, we address here only the RPM. The explicit comparisons of the DH and MSA theories [*without*]{} allowance for ion pairing are presented in Sec. III, below. In Sec. IV the theories that include descriptions of ion pairing are assessed, including the PMSA theories [@stellassoc]. Bounds for the Energy and Free Energy ===================================== Configurational Energy Bounds ----------------------------- The first rigorous lower bound for the configurational energy of the RPM seems to be due to Onsager [@onsag]. It is essentially a consequence of the positivity of the total electrostatic potential energy density and, with the notation of (\[equdef\]), yields $$\label{eqonsag} u(\rho,T) \geq u_{\mbox{\scriptsize Ons \normalsize}} = -1.$$ A more transparent derivation for a system with a neutralizing background has been presented by Rosenfeld and Gelbart [@onsagdemo]. Totsuji, in 1981 [@tot], improved on Onsager’s result by writing the energy as an integral over the ionic pair correlation functions and showing that the presence of the hard-core repulsions implies an upper bound on the correlation functions. He thence established $$\label{eqtot} u(\rho,T) \geq u_{\mbox{\scriptsize Tot \normalsize}} = -0.960.$$ Although the improvement is by only $4.0 \%$, it has significant consequences. As remarked by Totsuji, one may usefully compare these bounds with the electrostatic or Madelung energies of an ionic crystal; for the NaCl (sc) and CsCl (bcc) structures one has [@kittel] $$\label{eqxtal} u_{\mbox{\scriptsize NaCl \normalsize}} \simeq -0.8738 \; \; \; \; \; \mbox{and} \;\;\;\;\; u_{\mbox{\scriptsize CsCl \normalsize}} \simeq -0.8813 .$$ One may reasonably suppose that the latter represents the best possible lower bound and so we will also invoke it in testing approximate theories for the RPM. Gillan’s Free Energy Upper Bound -------------------------------- Gillan [@gil] has developed a convincing, but not fully rigorous, upper bound on the Helmholtz free energy of the RPM, which incorporates the idea of ion pairing. The pure hard-core free energy actually provides a rigorous upper bound [@stell], but Gillan’s bound is lower except for extremely low densities $({\rho^{\ast}}\lesssim 10^{-5})$ where the limiting behavior is well understood. Here we utilize only Gillan’s bound, which is derived with the aid of the Gibbs-Bogoliubov inequality by employing a sequence of truncated reference systems. The calculation finally incorporates paired $(+,-)$ ions or dipoles by using a reference system of over-sized, spherically-capped cylinders with modified Coulomb interactions. The last step of Gillan’s argument relies on a comparison of an approximate analytical expression for the pressure of a system of such spherocylinders with computer simulation estimates[@nezbeda; @rebertus]: the approximate formula appears to provide a bound on the true results. A search of the more recent literature concerning this system (e.g. Refs. [@citer1; @citer2; @levesque; @citer3]) indicates that the original simulations have withstood the test of time. (However, Frenkel [@frenkel] has observed that at high densities and for length/diameter ratios larger than needed here, the simulations — and, certainly, the analytic approximation — miss an isotropic-nematic fluid transition that is to be expected.) We thus believe that Gillan’s bound is valid. To display the bound explicitly, we write the diameter and the chosen [@gil] center-to-center distance of the spherocylinders as $a_s = (1 + \delta) a$ and put $$\lambda \equiv (5 \pi/24) \rho a_s^3 = (5 \pi/24)(1 + \delta)^3 {\rho^{\ast}}.$$ If ${\bar{f}}^{\, \mbox{\scriptsize Id \normalsize}} \!\! (\rho,T)$ is the ideal-gas free energy density, we then have [@gil] $$-{\bar{f}}(\rho,T) \leq -{\bar{f}}^{\, \mbox{\scriptsize Id \normalsize}} \!\! (\rho,T) + \hbox{$\frac{1}{2}$} \, \rho \, {\cal F}(\rho,T),$$ $${\cal F}(\rho,T) = 1 - 2 \pi {\rho^{\ast}}- \frac{1}{{T^{\ast}}} - \hbox{$\frac{18}{5}$} \, \lambda \, \frac{1 - \frac{2}{5} \lambda}{(1 - \lambda)^2} -\ln{{\cal L}(\rho,T)},$$ $${\cal L}(\rho,T) = {T^{\ast}}(1 - \lambda) \left \{ 1 - \exp{[-\delta/{T^{\ast}}(1 + \delta)]} \right \} .$$ We will adopt $\delta = 0.3$, which Gillan found optimized the bound for most values of $T$. Basic Theories for the RPM: Comparison with Bounds ================================================== DH and MSA without pairing -------------------------- ### DH theory Debye-Hückel theory[@debye] (here referred to as “pure” DH theory, since explicit dipolar pairing is not included) is the oldest theory for electrolytes still in current use. The theory entails two approximations: first, the pair correlation functions, $g_{ij}({\bf r}_i - {\bf r}_j)$, are represented by naive Boltzmann factors — with the charge, $q_j$, multiplied by the average electrostatic potential at ${\bf r}_j$ when an ion of charge $q_i$ is fixed at ${\bf r}_i$ — ignoring higher order correlation effects; and, second, these Boltzmann factors are linearized, which is valid only in the limit of low density, small charge, or high temperature. (For a modern discussion, see McQuarrie [@debye].) The thermodynamics predicted by DH theory depends only on the single parameter, $x= \kappa_D a$. The appearance of the hard core diameter, $a$, demonstrates that DH theory takes account of the electrostatic effects of the hard cores; however, the original or pure DH theory did not treat the excluded-volume effects of the hard cores (and so reduced to a theory for an ideal gas mixture in the limit of vanishing charge, $q \rightarrow 0$). Nonetheless, excluded volume contributions may be included naturally by adding to the free energy a suitably chosen pure hard-core term [@fl; @lf]; see below. In the DH critical region, such terms have a relatively small effect. ### The MSA and variants The other “basic” theory we consider, the mean spherical approximation[@msa], is defined by a closure of the Ornstein-Zernike relation in which the $g_{ij}({\bf r})$ vanish [*inside*]{} the hard core, while the direct correlation functions [*outside*]{} the hard-core exclusion zone are approximated by the Coulombic potentials. Waisman and Lebowitz[@wais] solved the MSA exactly for the RPM: that is, they determined the correlation functions which, in principle, yield the thermodynamics. The electrostatic free energy again depends only on $x = \kappa_D a$, but it and the overall free energy depend strongly on the theoretical route taken — via, in particular, the energy, pressure, or compressibility relations. Since very different results are obtained, we review them briefly. The standard MSA thermodynamics almost invariably discussed in the literature employs the energy route; but as a result, [*no excluded-volume hard-core terms are generated.*]{} Typically this problem is overcome by adding in appropriate terms “by hand,” just as for DH theory [@fl; @lf]. In light of this fact, the conceptual advantage sometimes claimed for the standard MSA in comparison to DH theory (see, e.g. \[8(b)\]), namely, that the former treats the hard cores in better fashion, seems strictly inconsequential. Note also that the density-density correlation functions, $G_{\rho \rho}({\bf r})$, and also charge-charge correlation functions, $G_{q q}({\bf r})$, that satisfy the Stillinger-Lovett second-moment-condition follow from DH theory (again contrary to \[8(b)\]) when properly generalized [@benf]. The pressure route to MSA thermodynamics (which we will denote MSpA) generates a different approximation for the electrostatic excess free energy, along with the Percus-Yevick-pressure-equation hard-core free energy. It is interesting that, like the ordinary energy-route MSA thermodynamics, the MSpA yields both a critical point and the exact DH limiting laws; early on, however, Waisman and Lebowitz \[11(c)\] dismissed it as inferior. By contrast, the compressibility route yields [*no*]{} electrostatic contribution, but generates [*only*]{} the Percus-Yevick-compressibility-equation free energy for uncharged hard spheres! Finally, note that the thermodynamics of the generalized MSA or GMSA (which is designed so that all three routes to the thermodynamics agree) [@evans; @gmsa] is identical to the ordinary, energy-route MSA combined with the Carnahan-Starling (CS) approximation for the pure hard-core free energy [@cs]. ### Hard Cores Since the RPM consists of hard spheres, it is certainly desirable to include an account of the excluded volume effects in any approximate theory. As we have seen, the two principal approximations, DH and MSA, require the insertion of hard cores terms “by hand,” and two other theories, MSpA and GMSA, entail two different hard-core approximations. For the sake of convenience and uniformity, then, we will employ the CS hard core approximation [@cs] in the calculations reported here for [*all*]{} theories that recognize excluded volume effects. The corresponding theories will be denoted DHCS, MSACS, and MSpACS, while the notation DH, MSA, and MSpA will be reserved for the “pure” (electrostatics only) theories. We have, however, checked that other approximations for the pure hard-core contributions yield qualitatively similar results. It is worth mentioning that although hard-core terms do not contribute directly to the internal energy (since their contribution to the energy of allowed configurations vanishes — as correctly reflected by the CS approximation), they [*do*]{} influence the overall internal energy picture. Specifically, for the basic theories, as we shall see, they affect internal energy isotherms by altering the coexistence curve; for the augmented, pairing theories, they enter by changing the degree of pairing. Assessment of Basic Theories ---------------------------- ### DH Configurational Energy For pure DH theory (with neither pairing nor hard-core effects) the configurational energy assumes a particularly simple form, namely, $$\label{equdh} u^{DH}({\rho^{\ast}},{T^{\ast}}) = -x/2(1+x).$$ Evidently the energy of DH theory [*violates none of the bounds*]{} for any values of $\rho$ and $T$: see (\[eqkapx\]), (\[eqonsag\]), and (\[eqtot\]). Furthermore, $u^{DH}$ remains above the crystal values (\[eqxtal\]) as is apparent in Fig. 1. The contrary statements by Blum and coworkers[@blum; @bernard; @blumnew] that $u^{DH}$ violates Onsager’s bound perhaps mistake the Debye-Hückel limiting law (DHLL) — i.e., truncation of DH theory to lowest order in $x$, which no one should take seriously for $x \gtrsim 0.3$: see Fig. 1 — for the full DH theory propounded in [@debye]. Strictly, the dependence of $u^{DH}$ on the single parameter $x$ given in (\[equdh\]) can be correct only in single-phase regions of the ($\rho,T$) plane. Below the critical temperature (as defined by the theory at hand) the energy in the coexistence region is always a weighted sum of the values in the two phases, say ${\alpha}$ and ${\beta}$. In fact, if the energies per particle are $u_{{\alpha}}$ and $u_{{\beta}}$ and the densities $\rho_{{\alpha}} = \rho_{{\alpha}}(T)$ and $\rho_{{\beta}} = \rho_{{\beta}}(T)$, one finds $$\label{equcoex} u({\rho^{\ast}},{T^{\ast}}) = \frac{ \rho_{{\alpha}} (\rho_{{\beta}} - \rho) u_{{\alpha}} + \rho_{{\beta}} ( \rho - \rho_{{\alpha}}) u_{{\beta}} }{ \rho ( \rho_{{\beta}} - \rho_{{\alpha}} ) } ,$$ so that $u$ varies linearly with $1/\rho$. Thus the main DH plot in Fig. 1 is restricted to $T \geq T_c^{DH}$, and similarly for the other theories. However, including phase coexistence according to (\[equcoex\]) cannot induce bound violation, since a weighted sum of two acceptable values also satisfies the bound: see the inset in Fig. 1 where the solid curves depict DH isotherms for $T \leq T_c^{DH}$. = Regarding the effects of hard cores, one finds that the only changes in DHCS theory occur in the two-phase regions below $T_c^{DHCS}$: the energy isotherms are shifted from those of pure DH theory since the coexistence curve differs. The dashed curves in the inset to Fig. 1 show the rather small effects: the shifts mainly reflect the expected lowered densities on the liquid branch of the coexistence curve. Naturally, these changes cannot induce any violation of Totsuji’s bound or of the crystal limits. ### MSA Configurational Energy Now Blum and Bernard [@blum; @bernard] have claimed the energy of the (pure) MSA, is “asymptotically correct.” However, as can be seen in Fig. 1, the MSA reduced excess energy, namely [@waiseq], $$\label{equmsa} u^{MSA}({\rho^{\ast}},{T^{\ast}})= - \left [ 1 + x - (1 + 2x)^{1/2} \right ]/x,$$ asymptotically approaches the Onsager bound of $-1$ but [*violates the Totsuji bound*]{} for $x \geq {x_{\mbox{\scriptsize T}}}\simeq 1200$ (as Totsuji noted originally [@tot]). Furthermore, $u^{MSA}$ lies below the crystal values for $x \geq {x_{\mbox{\scriptsize X}}}\simeq 125$. In fact, even in the absence of Totsuji’s result, it is hard to make sense of the claim [@blum; @bernard] that the MSA energy is asymptotically correct for the RPM in the limit of large $x$ by virtue of its approach to Onsager’s bound. Agreement with a bound is hardly proof of correctness [@reblum]! Furthermore, the limit $x {\rightarrow}\infty$ at fixed density implies ${T^{\ast}}\sim T/q^2 {\rightarrow}0$; but at low temperatures, one expects crystalline phases to appear for ${\rho^{\ast}}\lesssim {\rho^{\ast}}_{max} = \sqrt{2}$ (for fcc sphere packing) [@fisher] and these are not described by any of theories under consideration. It is worthwhile to interpret more explicitly the values ${x_{\mbox{\scriptsize T}}}$ and ${x_{\mbox{\scriptsize X}}}$, where violation by the pure MSA (no hard cores) occurs. On the liquid side of the coexistence curve, ${x_{\mbox{\scriptsize T}}}$ corresponds to violation when ${T^{\ast}}\leq 0.012 \simeq (0.14) T^{*MSA}_c$ and ${x_{\mbox{\scriptsize X}}}$ corresponds to ${T^{\ast}}\leq 0.035 \simeq (0.41) T^{*MSA}_c$. (The first violation temperature here is estimated with the aid of a low-temperature asymptotic analysis of the pure MSA coexistence curve [@msalowt] while the second follows directly from a numerical evaluation.) The solid curves in Fig. 2 demonstrate the effects. = The inclusion of hard-core terms (“by hand”) in the pure MSA changes the liquid-side coexistence curves more strongly than in DH theory. Thus for the MSA with CS terms or, equivalently, for the GMSA, the violations shift to much lower ratios of $T/T_c^{GMSA}$: this is clearly evidenced by the dashed coexistence isotherm shown in Fig. 2 for ${T^{\ast}}= 0.030 \simeq (0.38) T_c^{*GMSA}$ (with $T_c^{*GMSA} \simeq 0.0786$ [@evans; @tdgevans]). ### MSpA Configurational Energy The energy according to the MSpA is [@wais] $$\begin{aligned} u^{MSpA} = - \hbox{$\frac{1}{3}$} &[& 1 - ( 1 - \sqrt{1 + 2 x} )/x \\ &+& 2 \ln{\left( 1 + x + \sqrt{1 + 2 x} \right)} - 2 - \ln{4}] ,\nonumber\end{aligned}$$ which, in the single-phase region, also depends only on the parameter $x$. As evident from Fig. 1, however, this violates the Totsuji and Onsager bounds at ${x_{\mbox{\scriptsize T}}}\simeq 6.5$ and ${x_{\mbox{\scriptsize O}}}\simeq 7.1$, respectively. These results provide ample justification for a disparaging evaluation of the pressure-route thermodynamics for the MSA. For the remainder of this paper, we thus omit the MSpA. DH and MSA Free Energies {#secsimgil} ------------------------ In the pure theories (in which Bjerrum ion pairing is not explicitly included) we find that [*both*]{} DH theory and the MSA [*violate*]{} Gillan’s free energy upper bound. The entire vapor branches of both coexistence curves, as well as both sides of the DH critical region, are in violation. As shown in Fig. 3, the violations remain when hard-core excluded volume corrections are included. The DHCS and GMSA treatments exhibit very similar features, for the low densities of interest. Note that in Fig. 3 we follow the coexistence prescription for the free energy corresponding to (\[equcoex\]). Note also that non-violation on one branch of the coexistence curve (as on the GMSA liquid side) is at best a qualified virtue since the construction of the coexistence curve depends on the free energies on [*both*]{} sides. In light of these results it is clearly imperative to examine theories which allow for ion pairing. ASSESSMENT OF ION-PAIRING THEORIES {#secaug} ================================== Bjerrum and Beyond ------------------ To compensate for the effects of the DH linearization of the electrostatic Boltzmann factor, Bjerrum [@bj] postulated association of “free” ions of (residual) density $\rho_1$ into “bound” neutral dipolar pairs of density $\rho_2$ so that the overall density is $$\label{eqconserve} \rho = \rho_1 + 2 \rho_2 .$$ In terms of the ideal-gas free energy density ${\bar{f}}^{Id}_j(\rho_j,T) = \rho_j [1 - \ln(\Lambda_j^{3j} \rho_j/\zeta_j)]$ with mean thermal de Broglie wavelengths $\Lambda_j(T)$ and internal partition functions $\zeta_j(T)$ [@lf], we may then write the total free energy density as [@fl; @lf] $$\label{eqfbasic} {\bar{f}}= 2 {\bar{f}}^{Id}_1(\hbox{$\frac{1}{2}$} \rho_1) + {\bar{f}}^{Id}_2(\rho_2) + {\bar{f}}^{Ex}(\rho_1, \rho_2) ,$$ with the excess free energy density $$\label{eqfex} {\bar{f}}^{Ex}(\rho_1, \rho_2) = {\bar{f}}^{HC}(\rho_1, \rho_2) + {\bar{f}}^{Ion}(\rho_1) + {\bar{f}}^{DI}(\rho_1, \rho_2),$$ where (i) ${\bar{f}}^{HC}$ denotes the pure hard-core/excluded-volume terms, (ii) ${\bar{f}}^{Ion}$ represents the electrostatic contribution of the free ions, while (iii) ${\bar{f}}^{DI}$ denotes the dipole-ion interaction/solvation terms [@fl; @lf]. As mentioned, we take here ${\bar{f}}^{HC}$ to be of Carnahan-Starling form [@cs] with the dipoles treated as effective spheres of diameter $\sigma_2 = 2^{1/3} a$ [@benf]. = Chemical equilibrium among the $+$ and $-$ free ions and dipolar pairs is imposed via the equality $\mu_2 = 2 \mu_1$ of the chemical potentials. If the association constant is defined by $K(T) = \Lambda_+^3 \Lambda_-^3 \zeta_2 / \zeta_+ \zeta_-\Lambda_2^6 = \zeta_2$ (see [@lf]) and the reduced excess chemical potentials are $$\label{eqchempot} \overline{\mu}^{Ex}_j \equiv \mu^{Ex}_j/k_B T = \ln{\gamma_j} = - ({\partial}{\bar{f}}^{Ex}/ {\partial}\rho_j),$$ with $\rho_+ = \rho_- = \frac{1}{2} \rho_1$ and $\gamma_+ = \gamma_- = \gamma_1$, then the mass action law states $$\label{eqmassaction} \frac{\rho_2}{\rho_+ \rho_-} = \tilde{K}(T; \rho_1, \rho_2) \equiv K(T) \frac{\gamma_+ \gamma_-}{\gamma_2}.$$ The optimal expression for $K(T)$ is a matter for debate [@fl; @lf] — and will be discussed further below. For reference purposes we adopt Ebeling’s form [@lf; @ebgr; @ebeling] which guarantees an exact representation of the RPM’s electrostatic second virial coefficient when one uses DH theory or the MSA (but not the MSpA) for ${\bar{f}}^{Ion}(\rho_1)$. Note that for ${T^{\ast}}\leq 0.05 \simeq {T^{\ast}}_c$ the difference between $K^{Eb}$ and Bjerrum’s original proposal, $K^{Bj}$, is less than 0.01%; it rises to 3.0% at ${T^{\ast}}= T_c^{*MSA} = 0.0858$, in accord with the Introduction. Bjerrum’s original theory [@bj] amounts to the approximation $$\mbox{DHBj:} \;\;\; {\bar{f}}^{Ex} \simeq {\bar{f}}^{Ion} \simeq {\bar{f}}^{DH}(x_1) \;\;\;\; \mbox{with} \;\;\;\; x_1 = \kappa_1 a ,$$ where $\kappa_1^2 = 4 \pi q^2 \rho_1/D k_B T$ represents the inverse squared Debye length for the [*free ions alone*]{}, while as usual [@debye], $$\label{eqfdh} {\bar{f}}^{DH}(x) = \left [ \ln{(1 + x)} - x + \hbox{ $\frac{1}{2}$ } x^2 \right ]/ 4 \pi a^3.$$ Friedman and Larsen [@friedlar] later found that the predicted coexistence curve was unphysical. More recently, Fisher and Levin [@fisher; @fl; @lf] elucidated the peculiar “banana” shape of the DHBj coexistence curve (see Fig. 4 below) and showed it became worse when excluded-volume terms were added as, e.g., in DHBjCS theory. However, they also estimated the dipole-ion solvation term as [@lf] $$\label{eqfdi} {\bar{f}}^{DI} = \rho_2 (a a_1^2/a_2^3 {T^{\ast}}) \tilde{\omega}_2(x_2), \;\;\;\;\; x_2 = \kappa_1 a_2,$$ $$\label{eqdiomega} \tilde{\omega}_2(x) = 3 \left [ \ln{( 1 + x + \hbox{$\frac{1}{3}$} x^2 )} - x + \hbox{$\frac{1}{6}$} x^2 \right ]/x^2 \approx x^2/12,$$ where $a_1 = (1.0 \mbox{-} 1.3)a$ is the mean dipolar size, or $+/-$ ion separation, while $a_2 \simeq 1.1619_8 a$ represents the effective electrostatic exclusion radius [@lf]. (Note that all the results given here use $a_1 = a$ and $a_2 = 1.16198 a$.) The resulting DHBjDI theories lead to sensible coexistence curves (see Fig. 5 below) that agree fairly well with current simulations \[5,2(b)\]. = At an earlier stage, Ebeling and Grigo [@ebgr] combined Bjerrum pairing with the MSA by replacing ${\bar{f}}^{DH}$ by [@evans; @wais] $$\label{eqfmsa} {\bar{f}}^{MSA}(x) = \left[ 2 + 6x + 3x^2 - 2 (1 + 2x)^{3/2} \right ]/12 \pi a^3,$$ with $x \Rightarrow x_1$ again evaluated at $\rho_1$. They also added excluded-volume terms. The resulting MSABj and MSABjCS $\equiv$ EGA \[8(b),14\] theories yield fully acceptable coexistence curves [@lf] but, as mentioned, the predicted critical temperatures are significantly too high \[2(b),5\]. = Recently, Zhou, Yeh, and Stell (ZYS) [@stellassoc] have extended Ebeling’s approach by using the MSA in conjunction with a “reference cavity theory of association” [@zhoustell]. Their [*pairing mean-spherical approximations*]{} or PMSA theories may be described by $$\begin{aligned} \mbox{PMSA:}&& \;\;\; \nonumber\\ {\bar{f}}^{Ex} &=& {\bar{f}}^{MSA}(x) + {\bar{f}}^{CS}(\rho) + \rho_2(T,\rho) \ln (\gamma_+ \gamma_- / \gamma_2),\nonumber\\ &&\label{eqpmsagen}\end{aligned}$$ where $x = \kappa_D a$ is now evaluated with the [*total density*]{}, $\rho$, and ${\bar{f}}^{CS}$ represents the single-component Carnahan-Starling form, evaluated at $\rho = \rho_1 + 2 \rho_2$ (i.e., bound pairs are not treated as geometrically distinct objects). Note that $\rho_2$ is here to be determined from (\[eqmassaction\]) once $K$, $\gamma_1$, and $\gamma_2$ are specified (see below); hence $\rho_2$ is an explicit algebraic function of the arguments stated in (\[eqpmsagen\]). The use of only the total density (in place of the free ion density $\rho_1$) results in an analytically simpler, more explicit formulation; but, in the light of the original DH and Bjerrum arguments, it seems rather unphysical since neutral bound pairs cannot contribute to screening in a direct way. Furthermore, as we will see, this approach entails a significant cost in accuracy. The specification of the PMSA may be completed by first noting that ZYS also adopt Ebeling’s association constant, $K^{Eb}(T)$ [@lf; @ebgr; @ebeling]. Then, for the activity coefficients, $\gamma_+ \equiv \gamma_-$ and $\gamma_2$, ZYS propose three levels of approximation, first: $$\begin{aligned} \mbox{PMSA1:} &&\nonumber\\ \ln{\gamma_1} &=& -(\partial {\bar{f}}^{MSA}/ \partial \rho )_T \equiv \overline{\mu}^{MSA}(T,\rho), \;\;\; \gamma_2 = 1,\nonumber\\ &&\end{aligned}$$ which neglects dipole-ion contributions \[cf. (\[eqpmsagen\])\]. Second, dipole-ion interactions are introduced by replacing the approximation $\gamma_2 = 1$ by $$\begin{aligned} \mbox{PMSA2:}&&\nonumber\\ \ln{\gamma_2} & = & \left [ 2(1+x)\sqrt{1+2x} - 2 - 4x - x^2 \right ]/ {T^{\ast}}x^2, \nonumber \;\;\;\;\;\;\;\;\;\; \\ & \approx & -x^2/4 {T^{\ast}}[1 + O(x)];\end{aligned}$$ see \[8(b)\], Eq. (4.11). Finally, the dumbbell-shaped hard cores of a dipolar ion pair are incorporated \[8(a)\] by using the CS cavity-value contact function and incrementing $\ln{\gamma_2}$ by $$\label{eqpmsa3} \mbox{PMSA3:} \;\;\;\;\; \Delta \ln{\gamma_2} = \ln{[ 2(1-\eta)^3/(2-\eta) ]},$$ where $\eta = \pi {\rho^{\ast}}/ 6$. PMSA3 is the preferred theory of ZYS and yields $({T^{\ast}}_c,{\rho^{\ast}}_c) \simeq (0.0745,0.0245)$. PMSA1 and PMSA2 give $(0.0748,0.0250)$ and $(0.0733,0.0229)$, respectively. The $T_c$ values are still significantly higher \[8(b)\] than the DH-based estimates, namely, ${T^{\ast}}_c \simeq 0.052 \mbox{-} 0.057$ [@fisher; @lf; @fishlee], while the simulations suggest ${T^{\ast}}_c \simeq 0.048 \mbox{-} 0.055$ \[2(b),15\]. Pairing Theories vs. Gillan’s Bound ----------------------------------- Comparison of the pairing theories with Gillan’s free energy bound is mainly encouraging. We find that theories that incorporate association in the Bjerrum chemical picture, in which the free ion density is [*depleted*]{} by pairing (i.e., $\rho_1 = \rho - 2 \rho_2$), never violate the bound. Indeed, even the most primitive Bjerrum theories, DHBj and MSABj — which include neither hard-core nor dipole-ion interactions — satisfy Gillan’s bound for all $({\rho^{\ast}},{T^{\ast}})$ values tested: see Fig. 4. On the other hand, all three PMSA theories turn out to violate Gillan’s bound in significant regions of the $({\rho^{\ast}},{T^{\ast}})$ plane, including nearly the entire vapor branches of the coexistence curves. As regards the MSABj and DHBj theories, the more-or-less vertical “excess contour lines” in Fig. 4 reveal the magnitude of non-violation in the density-temperature plane: they are loci on which Gillan’s upper bound exceeds the corresponding approximate reduced free energy density, $-{\bar{f}}a^3$, by the indicated amounts, ranging from $6 \times 10^{-4}$ up to $0.1$. The associated coexistence curves are also shown and one may notice that the excess contours undergo a jump in curvature on entering the corresponding two-phase region: this results from the coexistence prescription analogous to (\[equcoex\]). Fig. 5 shows the effects of incorporating dipole-ion solvation (DI) and excluded-volume (CS) terms. Note that removing the excluded-volume terms from these BjDICS theories produces only slight shifts in the excess contours at high densities and low temperatures. By contrast, the solid curve in Fig. 6 marks the boundary of the region inside which the PMSA3 free energy violates Gillan’s bound. The coexistence curve is also shown. (Note, however, that the coexistence prescription was not used here to compute the violation boundary within the two-phase domain.) The region of violation found for PMSA2 is nearly identical, while that for the PMSA1 theory is slightly larger, extending [*above*]{} the corresponding critical point, $T_c^{PMSA1}$: see the dashed curve in Fig. 6. = In conclusion, the violations of Gillan’s bound found previously and seen here for the PMSA theories demonstrate convincingly that association of oppositely charged ions into dipoles [*along with*]{} a concomitant depletion of free ions and their screening effects is a crucial element in the critical-region behavior of the RPM. Gillan’s bound also serves to highlight interesting contrasts between DH- and MSA-based theories: the MSA coexistence curve shifts only slightly when pairing is added (MSABj) yet, surprisingly, violation of Gillan’s bound is still completely avoided; the unphysical DHBj “banana” coexistence curve (in Fig. 4), on the other hand, immediately points to the significance of pairing, while satisfaction of Gillan’s bound is surprising here because the coexistence curve is so unconvincing. Pairing Theories vs. Energy Bounds ---------------------------------- Testing the pairing theories against the bounds of Totsuji and Onsager yields mixed results. For a window of temperatures that includes the critical region, namely, $0.015 \lesssim {T^{\ast}}\lesssim 0.5$, all the theories embodying ion association satisfy the energy bounds. We also find a surprising level of agreement among the various theories as to the value of the critical energy per particle: see Table I. At low temperatures, however, some of the MSA-based theories violate Totsuji’s bound. Moreover, at moderate temperatures (${T^{\ast}}\gtrsim 0.5$) [*all*]{} of the pairing theories violate fundamental thermal stability requirements (as discussed in the next section); for some of the approximations, this is also accompanied by violation of the Totsuji and Onsager bounds, as explained below. Now the energy for a general pairing theory follows from (\[eqfbasic\]) via the thermodynamic relation (\[equdef\]) and the mass action law (\[eqmassaction\]), etc., which leads to $$\label{equwithpair} u(\rho,T) = \frac{a^3 T^{\ast 2}}{{\rho^{\ast}}} \frac{{\partial}}{{\partial}{T^{\ast}}} {\bar{f}}^{Ex}(\rho,T) + \frac{\rho_2}{\rho} u_2(T),$$ where $u_2(T)$ is given by $$\label{equ2} u_{2}(T) = T^{\ast2} (d \ln{K(T)}/d {T^{\ast}})$$ But this can be recognized simply as the mean energy of a [*single*]{} $(+,-)$ bound pair since the corresponding internal configurational partition function for a pair is embodied in the association constant, $K(T)$ — see the text above (\[eqchempot\]) and Levin and Fisher [@lf]. Of course, the factor $\rho_2(\rho,T)$ in (\[equwithpair\]) is also to be determined via the law of mass action (\[eqmassaction\]). For theories of the form (\[eqfex\]), one can further write $$\label{equex} u^{Ex} = (a^3 T^{\ast 2}/{\rho^{\ast}}) ({\partial}{\bar{f}}^{Ex}/ {\partial}{T^{\ast}}) = u^{Ion} + u^{DI},$$ where the “basic” expressions for the electrostatic contribution, $u^{Ion}$, are now given by the natural generalizations of (\[equdh\]) and (\[equmsa\]), namely, $$\label{equdhpair} u^{DH}(\rho,T) = -(\rho_1/\rho)x_1/2(1+x_1),$$ $$\label{equmsapair} u^{MSA}(\rho,T)= - (\rho_1/\rho) \left [ 1 + x_1 - (1 + 2x_1)^{1/2} \right ]/x_1 .$$ For reference, we also quote the explicit result for $u^{DI}$ following from the treatment of Fisher and Levin in leading order [@DInote]. Defining $a_1$ and $a_2$ as in (\[eqfdi\]) and (\[eqdiomega\]) [@lf], one finds $$u^{DI} = - \frac{aa_1^2}{2 a_2^3}\frac{\rho_2}{\rho} \frac{({\kappa a_2})^2}{ [3 + 3 {\kappa a_2}+ ({\kappa a_2})^2] } .$$ The corresponding expressions for the PMSA theories are omitted for the sake of brevity. -------------------------- ------------------------ ------------------------ ------------------------ ------------------------ ------------------------ ------------------------ $ \;\;\;\;\;\;\;\;\;\; $ DH $+$CS $+$Bj $+$BjCS $+$BjDI $+$BjDICS \[0.1cm\] ${T^{\ast}}_c$ $0.0625$ $0.061_3$ $0.0625$ $0.061_5$ $0.057_4$ $0.052_5$ $u_c$ $-0.25$ $-0.241_1$ $-0.431_5$ $-0.437_8$ $-0.444_3$ $-0.453_3$ $x_c$ $1$ $0.931_5$ $3.013_5$ $3.281_1$ $2.466_1$ $2.424_0$ $x_{1c}$ $1$ $0.931_5$ $1$ $0.938_6$ $1.122_9$ $0.931_5$ \[0.1cm\] MSA $+$CS $+$Bj $+$BjCS $+$BjDI $+$BjDICS ${T^{\ast}}_c$ $0.085_8$ $0.078_6$ $0.085_8$ $0.078_7$ $0.082_1$ $0.072_3$ $u_c$ $\;\;\;-0.414_2\;\;\;$ $\;\;\;-0.335_8\;\;\;$ $\;\;\;-0.415_7\;\;\;$ $\;\;\;-0.378_1\;\;\;$ $\;\;\;-0.444_2\;\;\;$ $\;\;\;-0.414_8\;\;\;$ $x_c$ $2.414_2$ $1.522_1$ $2.721_3$ $2.040_8$ $3.072_9$ $2.208_3$ $x_{1c}$ $2.414_2$ $1.522_1$ $2.414_2$ $1.531_9$ $2.450_9$ $1.485_0$ \[0.1cm\] PMSA1 PMSA2 PMSA3 ${T^{\ast}}_c$ $0.073_3$ $0.074_8$ $0.074_5$ $u_c$ $-0.374_0$ $-0.426_6$ $-0.426_5$ $x_c$ $1.981_4$ $2.049_4$ $2.032_9$ -------------------------- ------------------------ ------------------------ ------------------------ ------------------------ ------------------------ ------------------------ : Some critical-point parameters for various theories: ${T^{\ast}}_c$; $u_c$, the reduced energy per particle; $x_c = (4 \pi {\rho^{\ast}}_c / {T^{\ast}}_c)^{1/2}$, the (overall) Debye parameter; and, $x_{1c} = (4 \pi \rho^*_{1c} / {T^{\ast}}_c)^{1/2}$, the screening parameter. (Note that the values quoted for $x_c$ in [@lf] correspond here to $x_{1c}$ and that the Ebeling association constant [@ebgr] was used throughout.) [2]{} ### Low Temperatures: Violation in MSA Pairing Theories For ${T^{\ast}}\lesssim 0.015$, evaluation of $u(\rho,T)$ reveals violations of the Totsuji bound for most of the MSA theories. The reason turns out to be literally the same as for the pure MSA: in the corresponding Bj, BjCS, BjDI, and BjDICS theories, as well as in the PMSA1 (although [*not*]{} PMSA2 and 3) theory, the mass-action pairing predicted by (\[eqmassaction\]) becomes exponentially small as ${T^{\ast}}{\rightarrow}0$ [@lfmsapair]. As a result, all these theories revert to their ion-only form (i.e., MSA or MSACS) and violations occur: see Fig. 1. A similar depletion of pairs occurs when ${T^{\ast}}{\rightarrow}0$ in the DHBjDI/CS (but [*not*]{} DHBj/CS) theories, and so these theories revert to the corresponding [*non-violating*]{} DH/CS theories. These results are independent of whether one uses the Ebeling or Bjerrum association constant, or any other reasonable partition-function-like form, as discussed below. ### Moderately Low Temperatures In the temperature range $0.015 \lesssim {T^{\ast}}\lesssim 0.5$, which includes ${T^{\ast}}_c$, all the pairing theories described in the present study satisfy the Totsuji bound, and hence Onsager’s as well. Fig. 7 depicts energy isotherms for the pairing theories at ${T^{\ast}}= 0.07$ . The plotted isotherms have been cut off for large $x = \kappa_D a$ at the hard-core packing limit, $\rho^{\ast}_{max} = \sqrt{2}$. Fig. 7 also shows the location of the critical point of the DHBjDICS theory, which may be regarded as a reference point in reading Table I. The table lists the various critical energies and Debye parameters. As mentioned, there is a fair measure of agreement among the different pairing theories regarding the energy at criticality even though other parameters vary quite strongly. = ### Violations at Moderate and High Temperatures Violation of the energy bounds are found again, as mentioned above, at [*higher*]{} temperatures in the range ${T^{\ast}}\gtrsim 0.5$, some 6 to 10 times greater than the estimates for ${T^{\ast}}_c$. The reason for this surprising fact, however, is quite different from the cause at low temperatures: it transpires, indeed, that the form of the association constant is now crucially important. In fact, any theory with pairing governed by Bjerrum’s association constant violates both the Totsuji and Onsager bounds when ${T^{\ast}}{\rightarrow}\frac{1}{2}-$ and $\rho$ is large enough! Once noticed numerically, this behavior can be understood analytically by evaluating the factor $u_2(T)$ in (\[equwithpair\]) using (\[equ2\]) with $K = K^{Bj}(T)$. To that end recall, first, the well known fact [@lf] that ${K^{Bj}(T)}$ vanishes linearly, say as $c_{Bj}(1 - 2 {T^{\ast}})$, when ${T^{\ast}}{\rightarrow}\frac{1}{2}-$ (and remains identically zero for ${T^{\ast}}> \frac{1}{2}$). Consequently, $u_2(T)$ diverges to $-\infty$ like $-\frac{1}{2}/(1 - 2 {T^{\ast}})$ in this limit. However, the factor $\rho_2(\rho,T)$ in (\[equwithpair\]) must be evaluated via the mass action law (\[eqmassaction\]) and is proportional to ${K^{Bj}(T)}$; this gives $$\label{equpairterm} \frac{\rho_2}{\rho} u_2 = \frac{\rho_1^2 \gamma_1^2}{4 \rho \gamma_2} T^{*2} \frac{dK}{d{T^{\ast}}} \approx -\frac{c_{Bj}}{8 a^3} \frac{\gamma_1^2}{\gamma_2} {\rho^{\ast}}< 0,$$ as ${T^{\ast}}{\rightarrow}\frac{1}{2}-$, so that $\rho_2 {\rightarrow}0$ and $\rho_1 {\rightarrow}\rho$. Note that the $\gamma_i(\rho,T)$, defined via (\[eqchempot\]), depend on the theory under consideration. One finds that $c_{Bj}/8 a^3 \simeq 11.6$: this is large enough so that the pairing term (\[equpairterm\]) by itself yields a violation of Onsager’s bound when (in DHBj theory) ${\rho^{\ast}}> \rho_{\mbox{\scriptsize Ons}}^{*\mbox{\scriptsize DHBj}} \simeq 0.39$ or (for MSABj) ${\rho^{\ast}}> \rho_{\mbox{\scriptsize Ons}}^{*\mbox{\scriptsize MSABj}} \simeq 0.64$. However, as the other terms in (\[equwithpair\]) are also negative, violations must arise at even lower densities. One finds numerically, in fact, that the violations occur at or below ${\rho^{\ast}}\lesssim 0.3$ in [*all*]{} the theories with pairing governed by Bjerrum’s association constant. One expects Ebeling’s choice, ${K^{Eb}(T)}$, which provides a match to the exact RPM second virial coefficient and never vanishes [@lf; @ebgr; @ebeling] — in contrast to the singular vanishing of ${K^{Bj}(T)}$ at ${T^{\ast}}\geq \frac{1}{2}$ — to fare better. Nevertheless, Ebeling’s association constant leads to Onsager and Totsuji bound violations in the region ${T^{\ast}}\simeq 0.7 \, \mbox{-} \, 1.0$ — although only in those theories which explicitly allow for the excluded volume effects. The PMSA3 treatment, furthermore, falls into this same category of violation; however, PMSA1 and 2 do [*not*]{} because the excluded-volume terms there do not affect the degree of pairing. All the violations just described turn out to be symptoms of a more serious weakness of both the Bjerrum and Ebeling association constants, as we will now demonstrate. Violations of Thermal Stability ------------------------------- To pursue further the origins of the Totsuji and Onsager energy bound violations at ${T^{\ast}}\gtrsim 0.5$, consider the energy isochores shown in Fig. 8. The two densities ${\rho^{\ast}}= 0.03$ and $0.1$ have been chosen for display because they bracket the critical density; similar behavior is seen at higher and lower densities. For the pure DH and MSA theories, included in Fig. 8 for reference purposes, $u(\rho,T)$ rises monotonically with $T$: this implies a positive constant-volume configurational specific heat, $C_V^{\mbox{\scriptsize conf}}(\rho,T)$. (Note that outside the two-phase region these two energy isochores are identical to those for DHCS and GMSA, respectively.) Now the positivity of the [*total*]{} constant-volume specific heat is a thermodynamic necessity dictated by the Second Law [@pippard]. For a classical particle system, however, the configurational contribution must be separately nonnegative: this follows [*either*]{}, thermodynamically, by regarding the kinetic and configurational degrees of freedom as thermally distinct systems [*or*]{}, from statistical mechanics, by expressing $C_V^{\mbox{ \scriptsize conf}}(\rho,T)$ as a mean-square energy fluctuation which is necessarily positive at finite positive temperatures in any nontrivial system [@cvcon]. = However, a quick perusal of Fig. 8 shows that [*all*]{} the pairing theory isochores — the solid and dashed curves representing DH- and MSA-based theories, respectively, and the dot-dash plots for PMSA1 and 3 — display regions where $u(\rho,T)$ [*decreases*]{} as $T$ increases. In other words, all the paring theories predict negative constant-volume specific heats and violate the Second Law! The reason is not far to seek. In the limit of complete pairing (i.e., $\rho_1 = 0, \; \rho_2 = \frac{1}{2} \rho$) all the approximate theories under consideration predict, via (\[equwithpair\]), that the energy should be simply that of independent dipolar pairs: this corresponds to the plots labeled ${\frac{1}{2}}u_2^{Bj}$ and ${\frac{1}{2}}u_2^{Eb}$ in Fig. 8 which derive from (\[equ2\]) and the Bjerrum and Ebeling forms for $K(T)$. But, as evident from the figure, both $u_2^{Bj}(T)$ and $u_2^{Eb}(T)$ exhibit pronounced maxima in the interval ${T^{\ast}}= 0.12 \, \mbox{-} \, 0.13$ and then fall sharply as $T$ increases, dropping below $u_2^{Bj}(0) = u_2^{Eb}(0) = -1$ at ${T^{\ast}}= 0.22_2$ and $0.21_9$, respectively. It is this behavior that leads to the decreasing regions in the overall excess energy isochores with incomplete pairing. But such a variation of $u_2(T)$ is physically nonsensical since, clearly, the configurational energy ${\varepsilon_2}(r) = -q^2/Dr$ of a bound pair cannot fall below the contact value $-q^2/Da$ (which, in turn, can be achieved in equilibrium only at $T=0$). The problem with $u_2(T)$ arises because the defining relation (\[equ2\]) does not actually yield the physically anticipated thermodynamic mean value [@halpern], say $\langle {\varepsilon_2}(r) \rangle_K$, which in the Bjerrum picture of association would be $$\label{eqetwoav} \langle {\varepsilon_2}(r) \rangle_K = 4 \pi \int_a^R {\varepsilon_2}(r)e^{-\beta {\varepsilon_2}(r)} r^2 dr / K(T),$$ with association constant $$\label{eqkoft} K(T) = 4 \pi \int_a^R e^{-\beta {\varepsilon_2}(r)} r^2 dr .$$ The reason for the failure is simple: the Bjerrum cutoff $R$ is taken to be temperature dependent [@halpern], explicitly, $R^{Bj}(T) = a/2 {T^{\ast}}$ for ${T^{\ast}}\leq {\frac{1}{2}}$ [@lf; @bj]. In general, such temperature dependence leads to the difference $$\label{equ2diff} \frac{q^2}{Da} u_2(T) - \langle {\varepsilon_2}\rangle_K = \frac{ 4 \pi R^2 e^{-a/{T^{\ast}}R} }{K(T)} k_B T^2 \frac{dR}{dT},$$ which is negative whenever $R(T)$ decreases as $T$ rises and which diverges when $K(T) {\rightarrow}0$. The Ebeling association constant can also be written in the form (\[eqkoft\]) but with the large-$T$ asymptotic form $R^{Eb}(T) - a \approx a/12 T^{*4}$ [@lf], which is quite accurate once ${T^{\ast}}\gtrsim 0.3$. We must conclude that neither the Ebeling nor the Bjerrum association constants can be regarded as representing even an “effective” partition function for an isolated ion pair as is required by or implicitly assumed in the standard theories of association [@lf; @davidson]. As suggested by Fig. 8, the unphysically large values of $u_2(T)$ lead to negative specific heats over large regions of the $(\rho,T)$ plane when either ${K^{Eb}(T)}$ or ${K^{Bj}(T)}$ is employed. Fortunately for our primary focus on the critical region, the violations of thermal stability are confined in all cases to ${T^{\ast}}\geq 0.12 > 2 {T^{\ast}}_c$ (and for the PMSA theories to ${T^{\ast}}\gtrsim 0.35$). At densities below ${\rho^{\ast}}= 0.01 \, \mbox{-} \, 0.02 < 0.6 {\rho^{\ast}}_c$ the pairing is sufficiently weak that the predicted ${C_V^{\mbox{\scriptsize conf}}}(\rho,T)$ always remains positive — although it does display an unphysical oscillation. Once violations arise at a given $T$, moreover, they persist to the highest densities. Of course, certain features are specific to the choice of association constant. As remarked earlier, ${K^{Bj}(T)}$ “switches off” abruptly at ${T^{\ast}}= {\frac{1}{2}}$ where a nonphysical latent heat is implied for all $\rho > 0$; above ${T^{\ast}}= {\frac{1}{2}}$ pairing is lost and no violations remain. When ${K^{Eb}(T)}$ is used in DH- and MSA-based theories with excluded-volume terms, violations remain at the highest temperatures. What might be a cure for these pathologies? It is clear from (\[eqetwoav\])-(\[equ2diff\]) that the unphysical behavior of $u_2(T)$ can be avoided if one fixes the cutoff in (\[eqkoft\]) at, say $R = \lambda a$, so defining $K^{\lambda}(T)$. Furthermore, for any fixed $\lambda > 1$, the $\mbox{low-}T$ behavior of $K^{\lambda}(T)$ still matches ${K^{Eb}(T)}$ to all orders in ${T^{\ast}}$ [@lf; @fuoss]. In addition, the choice of $\lambda$ may be optimized by requiring that the deviation $|({K^{Eb}}/K^\lambda) - 1| \equiv \delta$ remain less than a specified level up to as high a temperature as possible. Thus one finds that $\lambda \simeq 3.4$ provides 1% precision ($\delta = 0.01$) up to ${T^{\ast}}\simeq 0.11$. One can then check that [*none*]{} of the pairing theories employing $K^\lambda(T)$ with $\lambda \simeq 3.4$ violates the energy bounds or thermal stability for any realizable thermodynamic state, $(\rho,T)$. In addition, the qualitative conclusions regarding the violation and nonviolation of the Gillan free energy bound remain unchanged. Indeed, using $K^{3.4}(T)$ causes only insignificant shifts of the free energy excess contours from those displayed in Figs. 4 and 5 when ${T^{\ast}}\lesssim 0.1$. Nevertheless, merely replacing ${K^{Eb}(T)}$ by $K^\lambda(T)$ leads to significant inaccuracies in the thermodynamics at [*higher*]{} temperatures. Thus a more thoughtful approach is essential to providing a reasonable approximate theory of the RPM that is valid over the full range of temperatures \[and up to moderate densities excluding, of course, the solid phase(s)\]. Such a treatment will be presented elsewhere [@assocpaper]. The second author is grateful for the stimulus provided for this work by Professor Joel L. Lebowitz and by his attendance at the meeting organized by Professor Lesser Blum at the University of Puerto Rico in March 1996 in honor of Bernard Jancovici. The interest of Professors George Stell and Harold Friedman is appreciated. The support of the National Science Foundation (through Grants CHE 93-11729 and CHE 96-14495) has been essential. J.M.H. Levelt Sengers and J.A. Given, [*Molec. Phys.*]{} [**80**]{}, 989 (1993). (a) M.E. Fisher, [*J. Stat. Phys.*]{} [**75**]{}, 1 (1994); (b) [*J. Phys. Cond. Matt.*]{} [**8**]{}, 1 (1996). G. Stell, [*J. Stat. Phys.*]{} [**78**]{}, 197 (1994). M.E. Fisher and Y. Levin, [*Phys. Rev. Lett.*]{} [**71**]{}, 3826 (1993). Y. Levin and M.E. Fisher, [*Physica A*]{} [**225**]{}, 164 (1996). The following corrections should be noted: (a) In Eq. (4.16) the coefficient $-8$ in the correction factor should be: $-4$. (b) In Eq. (4.21) the last term should be: $-\frac{1}{5}u_5 m^5 [1 + O(t)]$, i.e., insert a factor $1/5$. (c) In Eq. (5.10) the last factor $(b-4)$ should be preceded by a solidus: “$/$”. (d) The line before Eq. (7.12) should read: “... charging process (3.7).” (e) The right-hand side of Eq. (9.2) should read: $2.141213 \cdots$. (f) The first member of Eqs. (9.6) should read $\rho_{s-}/\rho_c \approx 32 T^{*3}$. (g) In the first member of Eqs. (9.7) the factor $T^{*2}$ should read $T^{*3}$. (h) See also [@msalowt] below. B. Guillot and Y. Guissani, [*Molec. Phys.*]{} [**87**]{}, 37 (1996). R.J.F. Leote de Carvalho and R. Evans, [*Molec. Phys.*]{} [**83**]{}, 619 (1994). (a)Y. Zhou, S. Yeh, and G. Stell, [*J. Chem. Phys.*]{} [**102**]{}, 5785 (1995); (b) S. Yeh, Y. Zhou, and G. Stell, [*J. Phys. Chem.*]{} [**100**]{}, 1415 (1996); and references therein. P.W. Debye and E. Hückel, [*Phys. Z.*]{} [**24**]{}, 185 (1923). For a modern account, see D. A. McQuarrie, [*Statistical Mechanics*]{} (Harper Collins, New York, 1976), Chap. 15. J.L. Lebowitz and J. Percus, [*Phys. Rev.*]{} [**144**]{}, 251 (1966). E. Waisman and J. L. Lebowitz, (a) [*J. Chem. Phys.*]{} [**52**]{}, 4307 (1970); (b)[*ibid*]{} [**56**]{}, 3086 (1972); and (c) [*ibid*]{} [**56**]{}, 3093 (1972). J.S. Høye, J. L. Lebowitz, and G. Stell, [*J. Chem. Phys*]{} [**61**]{}, 3253 (1974). N. Bjerrum, [*Kgl. Dan. Vidensk. Selsk. Mat.-fys. Medd.*]{} [**7**]{},1 (1926). W. Ebeling and M. Grigo, [*Ann. Phys. (Leipzig)*]{} [**37**]{}, 21 (1980); M. Grigo and W. Ebeling, [*J. Soln. Chem.*]{} [**13**]{}, 321 (1984); but note that these works suffer from numerical and other deficiencies: see, e.g., Ref. 8(b). G. Orkoulas and A.Z. Panagiotopoulos, [*J. Chem. Phys.*]{} [**101**]{}, 1452 (1994); J.M. Caillol, D. Levesque, and J.J. Weis, [*Phys. Rev. Lett.*]{} [**77**]{}, 4039 (1996); A.Z. Panagiotopoulos, private communication. B.P. Lee and M.E. Fisher, [*Phys. Rev. Lett.*]{} [**76**]{}, 2906 (1996) and to be published. L. Blum and O. Bernard, [*J. Stat. Phys.*]{} [**79**]{}, 569 (1995). O. Bernard and L. Blum [*J. Chem. Phys.*]{} [**104**]{}, 4746 (1996). L. Blum, M.F. Holvoko, and I.A. Protsykevych, [*J. Stat. Phys.*]{} [**84**]{}, 191 (1996). J.-P. Simonin and L. Blum, [*J. Chem. Soc. Faraday Trans.*]{} [**92**]{}, 1533 (1996). L. Onsager, [*J. Phys. Chem.*]{} [**43**]{}, 189 (1939). H. Totsuji, [*Phys. Rev. A*]{} [**24**]{}, 1077 (1981). J. Rasaiah and G. Stell, [*Molec. Phys.*]{} [**18**]{}, 249 (1970). M.J. Gillan, [*Molec. Phys.*]{} [**41**]{}, 75 (1980). (a) A.B. Pippard, [*Elements of Classical Thermodynamics for Advanced Students of Physics*]{} (Cambridge Univ. Press, 1957) Chap. 7; (b) H.B. Callen, [*Thermodynamics and an Introduction to Thermostatistics*]{}, 2nd Edn. (Wiley and Sons, New York, 1985) Chap. 8. M.J. Gillan, [*Molec. Phys.*]{} [**49**]{}, 421 (1983). An association constant with fixed upper cutoff was considered by R.M. Fuoss, [*Trans. Faraday Soc.*]{} [**30**]{}, 967 (1934), who also recognized the low-temperature insensitivity of $K(T)$ of the form (\[eqkoft\]) to the upper cutoff. D.M. Zuckerman and M.E. Fisher, to be published. R.G. Palmer and J.D. Weeks, [*J. Chem. Phys.*]{} [**58**]{}, 4171 (1973). Y. Rosenfeld and W. M. Gelbart, [*J. Chem Phys.*]{} [**81**]{}, 4574 (1984). C. Kittel, [*Introduction to Solid State Physics,*]{} 3rd Edn. (Wiley and Sons, New York, 1966). I. Nezbeda and T. Boublik, [*Czech. J. Phys. B*]{} [**28**]{}, 353 (1978). D.W. Rebertus and K.M. Sando, [*J. Chem. Phys.*]{} [**67**]{}, 2585 (1977). M.P. Allen, [*Phil. Trans. R. Soc. Lond.*]{} A [**344**]{}, 323 (1993). H.-L. Vörtler and J. Heybey, [*Molec. Phys.*]{} [**51**]{}, 73 (1984). D. Levesque, J.J. Weis, and J.P. Hansen, in: [*Monte Carlo Methods in Statistical Physics*]{}, ed. K. Binder (Springer, Berlin, 1979). P.A. Monson and M. Rigby, [*Chem. Phys. Lett.*]{} [**58**]{}, 122 (1978). D. Frenkel, [*J. Phys. Chem.*]{} [**91**]{}, 4912 (1987). See also D. Levesque and J.J. Weis in [*The Monte Carlo Method in Condensed Matter Physics*]{}, 2nd. Edn., ed. K. Binder (Springer, Berlin, 1995). N.F. Carnahan and K.E. Starling, [*J. Chem. Phys.*]{} [**51**]{}, 635 (1969); for mixtures, see (b) G. A. Mansoori [*et al.*]{}, [*J. Chem. Phys.*]{} [**54**]{}, 1523 (1971); (c) G. Jackson, [*Molec. Phys.*]{} [**72**]{}, 1365 (1991); (d) for the corresponding Helmholtz free energy, see also Ref. [@benf]. See Eq. (5) in Ref. \[11(a)\]. For completeness we mention that Refs. [@blum] and [@bernard] also cite the HNC approximation and compare the MSA favorably to it, in particular, for the one-component plasma (OCP), which consists of point particles. For the OCP, E.H. Lieb and H. Narnhofer, [*J. Stat. Phys.*]{} [**12**]{}, 291 (1975), have established a lower bound on the energy. In the strong coupling limit, $q^2 \rho^{1/3}/T {\rightarrow}\infty$, it has been shown by H. Gould, R.G. Palmer, and G.A. Estévez, [*J. Stat. Phys.*]{} [**21**]{}, 55 (1979), that the OCP energy computed via the MSA approaches this bound asymptotically; the same seems true for the HNC approximation: see K.-C. Ng, [*J. Chem. Phys.*]{} [**61**]{}, 2680 (1974). However, the relevance of these observations to the reliability or accuracy of the HNC and MSA for the OCP model again seems obscure. See Eqs. (9.8) and (9.9) of Ref. [@lf] but note the following corrections: (a) In Eq. (9.8) the coefficient $-6 \sqrt{2}$ in the correction factor should be: $-18$; (b) In Eq. (9.9) the right-hand side should read: $(e^3/81 \pi T^{*3} ) e^{-1/{T^{\ast}}}[1 - \frac{9}{2}{T^{\ast}}+ \cdots]$. M.M. Telo da Gama, R. Evans, and T.J. Sluckin, [*Molec. Phys.*]{} [**41**]{}, 1355 (1980). W. Ebeling, [*Z. Phys. Chem.*]{} (Leipzig) [**238**]{}, 400 (1968). For explicit expressions for $K^{Eb}(T)$, see [@lf] Eqs. (6.5)-(6.7). H.L. Friedman and B. Larsen, [*J. Chem. Phys.*]{} [**70**]{}, 92 (1979). Y. Zhou and G. Stell, [*J. Chem. Phys.*]{} [**96**]{}, 1504, 1507 (1992). M.E. Fisher and B.P. Lee, [*Phys. Rev. Lett.*]{} [**77**]{}, 3561 (1996). See Eqs. (7.12) and (7.13) of [@lf]. See Eqs. (9.15) - (9.18) of [@lf]. See, e.g., in Ref. \[25(b)\], Sec. 19-2. O. Halpern, [*J. Chem. Phys.*]{} [**2**]{}, 85 (1934). N. Davidson, [*Statistical Mechanics*]{} (McGraw-Hill, New York, 1962), Chap. 7. [^1]: Current address: Polymers Division, National Institute of Standards and Technology, Gaithersburg, MD 20899

ArXiv

--- abstract: 'The Stokes-Einstein relation for the self-diffusion coefficient of a spherical particle suspended in an incompressible fluid is an asymptotic result in the limit of large Schmidt number, that is, when momentum diffuses much faster than the particle. When the Schmidt number is moderate, which happens in most particle methods for hydrodynamics, deviations from the Stokes-Einstein prediction are expected. We study these corrections computationally using a recently-developed minimally-resolved method for coupling particles to an incompressible fluctuating fluid in both two and three dimensions. We find that for moderate Schmidt numbers the diffusion coefficient is reduced relative to the Stokes-Einstein prediction by an amount inversely proportional to the Schmidt number in both two and three dimensions. We find, however, that the Einstein formula is obeyed at all Schmidt numbers, consistent with linear response theory. The mismatch arises because thermal fluctuations affect the drag coefficient for a particle due to the nonlinear nature of the fluid-particle coupling. The numerical data is in good agreement with an approximate self-consistent theory, which can be used to estimate finite-Schmidt number corrections in a variety of methods. Our results indicate that the corrections to the Stokes-Einstein formula come primarily from the fact that the particle itself diffuses together with the momentum. Our study separates effects coming from corrections to no-slip hydrodynamics from those of finite separation of time scales, allowing for a better understanding of widely observed deviations from the Stokes-Einstein prediction in particle methods such as molecular dynamics.' author: - Florencio Balboa Usabiaga - Xiaoyi Xie - 'Rafael Delgado-Buscalioni' - Aleksandar Donev title: 'The Stokes-Einstein Relation at Moderate Schmidt Number' --- \#1 \#1 \#1 \#1[\#1]{} \#1[\#1]{} \#1[\#1]{} \#1[|\#1|]{} \#1[\#1]{} \#1[\#1]{} ¶ Introduction ============ The self-diffusion coefficient $\chi$ of a tracer particle suspended in a liquid (for example, a colloidal particle) is a quantity of fundamental interest, and can be predicted using the well-known Stokes-Einstein (SE) relation. This famous formula is a combination of two results. The first, by Einstein, comes from rather general linear response theory considerations and relates the diffusion coefficient to the mobility $\mu$, $\chi=k_{B}T\mu$ with $k_{B}T$ the thermal energy scale, where the mobility is defined by measuring the *average* steady velocity $\av u$ of the particle in response to a weak applied force $F$, $\mu=\lim_{F\rightarrow0}\,\av u/F$. The second part of the SE relation is the Stokes formula for the mobility of a sphere suspended in a viscous fluid, obtained using standard hydrodynamics. While there is little reason to doubt the applicability of Einstein’s relation, the Stokes-Einstein relation assumes that the drag force on a particle is not affected by thermal fluctuations, that is, it is assumed that the contribution of thermal kicks to the particle average to zero and the drag is the same as in a deterministic fluid. As we explain in this work, this is not necessarily so because of the nonlinear coupling between the *moving* particle and the fluctuating fluid. For a spherical particle of radius $a$ suspended in a three-dimensional fluid with shear viscosity $\eta$, the SE formula takes the form $$\chi=\frac{k_{B}T}{\alpha\eta a},\label{eq:SE}$$ where $\alpha$ is a coefficient that depends on the boundary conditions applicable at the surface of the sphere, equal to $6\pi$ for a stick surface and $4\pi$ for a slip surface [@Landau:Fluid]. In fact, the precise definition of $a$ and $\alpha$ are ambiguous except in certain limiting cases [@StokesEinstein_BCs; @StokesEinstein_MD], and it perhaps best to think of the product $\alpha a=6\pi R_{H}$ as a measure of the effective (stick) *hydrodynamic radius* of the particle $R_{H}$. It is not difficult to extend the SE relation to account for rotational diffusion and thus generalize it to more complicated rigid particle shapes. There is ample evidence and many theoretical calculations [@DiffusionRenormalization_III; @BilinearModeCoupling_SE] that demonstrate that (\[eq:SE\]) is asymptotically exact for a rigid sphere that is much larger and much more massive than the solvent molecules. Remarkably, the SE formula (\[eq:SE\]) is consistent with experimental and numerical measurements even for particles that are comparable in size to the solvent molecules (including a tagged fluid particle), with values of the hydrodynamic radius that are comparable to the actual physical size of the particle. Many deviations from this relation have been observed in particle simulations [@StokesEinstein_MD], however, it is virtually impossible to precisely attribute the cause of the mismatch because of the large number of violations of the assumptions that underlie (\[eq:SE\]). For example, traditional Navier-Stokes hydrodynamics may break down at the scales of the suspended particle [@GeneralizedHydrodynamics], or the appropriate boundary conditions may be different from the traditional no-slip condition [@StokesEinstein_BCs]. Even if one assumes traditional hydrodynamics applies, there are additional assumptions that enter the SE relation. One of the most important, yet often overlooked, assumptions, is that the Schmidt number is very large, $\text{Sc}=\nu/\chi\gg1$, where $\nu=\eta/\rho$ is the kinematic viscosity of the fluid. Physically, this means that momentum diffuses much faster than does the particle and the particle motion is viscous-dominated. In particular, at large Schmidt numbers the fluid velocity quickly relaxes to the solution of the steady Stokes equation as the particle barely moves [@BrownianParticle_SIBM]. This assumption can safely be made for realistic liquids. The Schmidt number for molecules in simple liquids is on the order of $10^{2}-10^{3}$, and for macromolecular or colloidal solutions it is at least an order of magnitude larger. However, there are several numerical methods used for simulating the diffusive motion of particles in flow for which this assumption cannot be safely made. This is particularly true for particle methods for hydrodynamics such as smooth particle hydrodynamics (SPH) [@Bian2012], smoothed dissipative particle dynamics (SDPD) [@SDPD_Scaling], stochastic hard-sphere dynamics (SHSD) [@SHSD_PRL], and stochastic rotation dynamics (SRD) (also called multiparticle collision dynamics (MPCD) [@MPC_LLNS]) [@SRD_Review; @MPCD_Review]. Achieving large Schmidt number in these methods requires a prohibitively small time step (collision frequency) and in many typical simulations $\text{Sc}$ is less than $10$. In several hydrodynamic (non-particle) methods such as the lattice Boltzmann (LB) method [@LB_SoftMatter_Review], the stochastic immersed boundary method [@StochasticImmersedBoundary; @SELM], or the inertial coupling method used here [@ISIBM], $\text{Sc}$ can be varied over a much broader range with comparative ease, however, $\text{Sc}$ is still limited to moderate values (e.g., $\text{Sc}\sim25-50$) by computational efficiency considerations due to the large separation of time scales between the viscous and diffusive dynamics. Practical heuristics have been developed empirically [@LBM_vs_BD_Burkhard; @BD_LB_Comparison], with findings generally consistent with the detailed investigation we perform in this work. In this work, we study the deviations of the diffusion coefficient of a tracer particle immersed in an incompressible Newtonian fluid from the Stokes-Einstein prediction at small and moderate Schmidt numbers. We utilize a recently-developed Incompressible Inertial Coupling Method [@ISIBM] method for coupling minimally-resolved particles with an incompressible fluctuating fluid solver. This allows us to eliminate compressibility yet consistently include both fluid *and* particle inertia, as well as the thermal fluctuations responsible for the diffusive Brownian motion. By changing the fluid viscosity we are able to control the Schmidt number and thus study whether the SE formula applies when the particle diffuses comparably fast to momentum. Furthermore, we can trivially change the dimensionality and study both two and three dimensional systems. There has been some confusion in the literature about the applicability of hydrodynamics to two dimensional systems, and statements to the effect that the SE relation does not apply in two dimensions have been made [@SE_2D]. As we will demonstrate, these mis-perceptions arise because finite-size effects diverge in two dimensions [@DiffusionRenormalization], and not because (fluctuating) hydrodynamics fails in two dimensions. In particular, we will demonstrate that in *finite* two dimensional systems the SE relation holds for $\text{Sc}\gg1$. In Section \[sec:BlobModel\] we describe the formulation of the minimally-resolved model we use to describe the coupled particle-fluid system. We first discuss the emergence of the Stokes-Einstein result in the limit of infinite Schmidt number, and then, in Section \[sub:FiniteSc\], we discuss finite Schmidt number corrections. In Section \[sec:Results\] we present numerical results for the velocity autocorrelation function and long-time diffusion coefficient at finite Schmidt numbers, and then offer some conclusions in Section \[sec:Conclusions\]. \[sec:BlobModel\]Brownian Particle Model ======================================== A detailed description of the fluid-particle equations that we employ, as well as a numerical algorithm to solve them, is given in Ref. [@ISIBM]. Here we briefly summarize the essential features, and then discuss in more detail the Brownian (overdamped) limit. Let us consider a particle of physical mass $m$ and size (e.g., radius) $a$ immersed in a fluid with density $\rho$. The position of the particle is denoted with $\V q(t)$ and its velocity with $\V u=\dot{\V q}$. The shape of the particle and its effective interaction with the fluid is captured through a smooth kernel function $\delta_{a}\left(\V r\right)$ that integrates to unity and whose support is localized in a region of size $a$. For example, one may choose any one-dimensional “bell-shaped” curve $\delta_{a}\left(r\right)$ with half-width of order $a$, such as a normalized Gaussian with standard deviation $a$ or a symmetric function of compact support of half-width $a$. This kernel is used to mediate two crucial operations. First, it is used to transfer (spread) the force $\V{\lambda}$ exerted on the fluid by the particle to the fluid. Second, it is used to impose a minimally-resolved form of the no-slip constraint stating that the velocity of the particle equals the local velocity of the fluid. We term this diffuse particle a *blob* for lack of better terminology (in polymer modeling the term bead is used for the same concept [@LB_SoftMatter_Review], while blob is used to denote an effective soft particle that includes internal degrees of freedom)*.* The physical volume of the blob $\D V$ is determined by the shape and width of the kernel function, $$\D V=\left[\int\delta_{a}^{2}\left(\V r\right)d\V r\right]^{-1}.\label{eq:dV_JS}$$ The equations of motion in our Inertial Coupling Method take the form [@ISIBM], $$\begin{aligned} \rho\left(\partial_{t}\V v+\V v\cdot\grad\V v\right) & = & -\grad\pi+\eta\grad^{2}\V v+\grad\cdot\left[\left(k_{B}T\eta\right)^{\frac{1}{2}}\left(\M{\mathcal{W}}+\M{\mathcal{W}}^{T}\right)\right]-\V{\lambda}\,\delta_{a}\left(\V q-\V r\right)\label{eq:v_t}\\ \grad\cdot\V v & = & 0\\ m_{e}\dot{\V u} & = & \V F\left(\V q\right)+\V{\lambda}\label{eq:u_t}\\ \V u & = & \int\delta_{a}\left(\V q-\V r\right)\V v\left(\V r,t\right)\, d\V r,\label{eq:no_slip}\end{aligned}$$ where $\rho$ is the constant fluid density, $\V v\left(\V r,t\right)$ is the incompressible fluid velocity, $\pi\left(\V r,t\right)$ is the hydrodynamic pressure, $m_{e}$ is the *excess mass* of the particle relative to the fluid, and $\V F\left(\V q\right)$ is the external force applied to the particle. Here the stochastic momentum flux is modeled using a white-noise random Gaussian tensor field $\M{\mathcal{W}}\left(\V r,t\right)$, that is, a tensor field whose components are independent (space-time) white noise processes [@FluctHydroNonEq_Book]. In this instantaneous frictionless coupling the total fluid-particle force $\V{\lambda}$ is a Lagrange multiplier that enforces the no-slip constraint (\[eq:no\_slip\]). In this work we will exclusively consider periodic systems, i.e., consider diffusion of the Brownian particle on a torus, in either two or three dimensions. The equations used here only account for the hydrodynamic contribution to the diffusion coefficient, and not any additional molecular (kinetic) contributions [@Friction_Diffusion_SE]. One can include a random slip to account for unresolved molecular effects, as described in the Appendix of Ref. [@ISIBM], in which case the additional “bare” diffusion coefficient would simply be added to the hydrodynamic contribution studied here [@StokesLaw]. In principle the equations employed here can be obtained by coarse-graining the complete microscopic dynamics with the momentum density field and the position of the particle as the relevant (slow) degrees of freedom [@SELM_Reduction]. Let us now assume that the nonlinear inertial effects in the momentum equation can be neglected (this is easy to check in our method by simply omitting the advective momentum flux term in the implementation), and that the immersed particle is neutrally-buoyant, $m_{e}=0$. In this work we will not carefully study the effects of the particle excess mass $m_{e}$, however, the results presented in Section \[sec:Results\] suggest that $m_{e}$ does not affect the long-time diffusive motion. This suggests that the dominant contribution to the effect we study here comes from the transient inertia of the fluid (i.e., the term $\rho\partial_{t}\V v$ in the momentum equation). Under these assumptions, we get the simpler equations of motion for the fluid-particle system, $$\begin{aligned} \rho\partial_{t}\V v+\grad\pi & = & \eta\grad^{2}\V v+\grad\cdot\left[\left(k_{B}T\eta\right)^{\frac{1}{2}}\left(\M{\mathcal{W}}+\M{\mathcal{W}}^{T}\right)\right]+\V F\left(\V q\right)\,\delta_{a}\left(\V q-\V r\right)\label{eq:v_t_Stokes}\\ \V u=\frac{d\V q}{dt} & = & \int\delta_{a}\left(\V q-\V r\right)\V v\left(\V r,t\right)\, d\V r,\nonumber \end{aligned}$$ which are much easier to analyze. Note, however, that even though the fluid equation is linearized, the no-slip constraint is a nonlinear constraint because of the presence of the particle position in the argument of the kernel. These semi-linear equations are also the equations used in the Stochastic Immersed Boundary Method [@StochasticImmersedBoundary] and the closely-related Stochastic Eulerian-Lagrangian Method [@SELM]. We note however that in our numerical calculations we employ the full equations (\[eq:v\_t\]-\[eq:no\_slip\]), and only use the linearized equations for theoretical analysis. The numerical method used to solve the equations relies on a finite-volume staggered discretization of the fluid equation [@LLNS_Staggered], and on the immersed boundary method [@IBM_PeskinReview] for handling the fluid-particle interaction [@ISIBM]. The mobility of a blob is easy to evaluate in the *deterministic* setting (the stochastic setting will be discussed later on). Consider applying a constant force $\V F$ on the blob and waiting for the velocity of the particle to reach a steady value. In the steady state the fluid velocity solves the steady Stokes equation and can be obtained explicitly, $$\V v\left(\V r\right)=\eta^{-1}\int\M G\left(\V r,\V r^{\prime}\right)\V F\left(\V q\right)\delta_{a}\left(\V q-\V r^{\prime}\right)\, d\V r^{\prime},$$ where $\M G$ is the Green’s function (Oseen tensor) for the steady Stokes equation [^1] with unit viscosity ($\grad\pi=\grad^{2}\V v+\V f$ subject to $\grad\cdot\V v=0$ and appropriate boundary conditions). Note that for periodic boundaries the integral of $\V f$ over the unit periodic cell of volume $V$ must vanish; we therefore take the definition of $\M G$ to include subtracting the total applied force $\M F$ as a uniform force density $-V^{-1}\M F$ on the right hand side of the momentum equation. The velocity of the blob $\V u=\M{\mu}\V F$ is determined from the no-slip constraint, giving the mobility tensor (see also Ref. [@SIBM_Brownian]) $$\M{\mu}=\eta^{-1}\int\delta_{a}\left(\V q-\V r\right)\M G\left(\V r,\V r^{\prime}\right)\delta_{a}\left(\V q-\V r^{\prime}\right)\, d\V rd\V r^{\prime}.\label{eq:mu_blob}$$ For isotropic systems the mobility tensor is a multiple of the identity tensor, $\M{\mu}=\mu\M I$. In three dimensions, the scalar mobility $\mu=\left(\alpha\eta a\right)^{-1}$ can be taken to define an effective hydrodynamic radius of the blob $R_{H}\sim a$ via the relation for a no-slip rigid sphere, $\mu\eta=\left(6\pi R_{H}\right)^{-1}$. There are well-known finite-size corrections to the mobility for periodic systems with a unit cell of volume $L^{d}$ that scale like $L^{-1}$ in three dimensions [@Mobility2D_Hasimoto; @LB_SoftMatter_Review; @DirectForcing_Balboa]. The hydrodynamic radius of the blobs that we employ in our numerical implementation has been measured computationally in Refs. [@ISIBM; @DirectForcing_Balboa]. The spatial discretization of the fluid/particle equations leads to a small violation of translational invariance and isotropy and the mobility tensor is not exactly constant or diagonal but rather depends on the precise location of the blob relative to the underlying grid used to solve the fluid equation [@ISIBM]. These discretization artifacts are small, on the order of a couple percent for the three-point Peskin kernel [@StaggeredIBM], and a fraction of a percent for the four-point Peskin kernel [@IBM_PeskinReview]. It is also possible to construct discrete kernel functions with even better translational invariance at the cost of increasing the support of the kernels and thus the computational cost of the algorithm. In two dimensions, the mobility of a blob diverges logarithmically with system size, consistent with the Stokes paradox for flow past a cylinder in an unbounded domain. For periodic system with a square unit cell with area $L^{2}$, the logarithmic divergence of the Green’s function for two-dimensional Stokes flow gives [^2] $$\mu=\left(4\pi\eta\right)^{-1}\ln\frac{L}{\alpha a},\label{eq:mu_2D}$$ where the coefficient $\alpha$ depends on the shape of the kernel $\delta_{a}$ [@DiffusionRenormalization]. This is analogous to the formula for the mobility of a periodic array of no-slip rigid cylinders of radius $R_{H}$, $\mu=\left(4\pi\eta\right)^{-1}\ln\left[L/\left(3.708\, R_{H}\right)\right]$ [@Mobility2D_Hasimoto], and can be used to define an effective hydrodynamic radius for a two-dimensional blob. It is important to note that while the mobility diverges for an infinite system, for any finite system the mobility is finite and well-defined even in two dimensions. There is, in fact, no fundamental difference between two and three dimensions; it is simply the slower decay of the Green’s function in two dimensions that changes the essential physics. \[sub:BrownianDynamics\]Brownian Dynamics Limit ----------------------------------------------- The short-time motion of particles immersed in the fluid is known to be very strongly affected by momentum diffusion and by inertial effects [@VACF_Langevin]. Since our method includes inertial effects, it is able to produce both the correct short-time and long time features of the Brownian motion of a particle, where “short” refers to time scales after sound waves have decayed. This observation was made for the Stochastic Immersed Boundary Method in Refs. [@BrownianParticle_SIBM; @SIBM_Brownian], and in Ref. [@ISIBM] we demonstrated that our Inertial Coupling Method also correctly captures the known physical effects of particle and fluid inertia on the velocity autocorrelation function (VACF) $$\M C(t)=\av{\V u(t)\otimes\V u(0)},$$ including the effect of particle excess mass on the short time behavior of the VACF, as well as the long time power-law tail of the VACF. Here we focus on the long-time diffusive motion of the particles, where “long” means that the motion is observed at a time scale $\tau$ over which momentum has diffused throughout the domain and the VACF has decayed to zero, $\tau>t_{L}=L^{2}/\nu$, where $\nu=\eta/\rho$ is the kinematic viscosity. In the long-time limit, the motion of a single free particle immersed in a fluid looks like simple Brownian motion with a diffusion coefficient that can be defined from the long-time mean square displacement of the particle. More generally, one defines a time-dependent *diffusion tensor*, either using the mean-square displacement of the particle, $$\M{\chi}_{\text{MSD}}(t)=\av{\frac{\left[\V q(t)-\V q(0)\right]\otimes\left[\V q(t)-\V q(0)\right]}{2t}},$$ or the integral of the VACF, $$\M{\chi}_{\text{VACF}}(t)=\int_{0}^{t}\M C(t^{\prime})dt^{\prime}=\frac{d}{dt}\av{\frac{\left[\V q(t)-\V q(0)\right]\otimes\left[\V q(t)-\V q(0)\right]}{2}},$$ The long-time diffusion tensor is then the asymptotic value $$\M{\chi}=\lim_{t\rightarrow\infty}\M{\chi}_{\text{MSD}}(t)=\lim_{t\rightarrow\infty}\M{\chi}_{\text{VACF}}(t).$$ While we will not demonstrate this here, it can be shown that, for very large Schmidt numbers, neither sound effects (compressibility) nor inertial effects affect the diffusion coefficient of a single particle [@VACF_Langevin]. For a single particle, $\M{\chi}$ can be obtained from the mobility tensor $\M{\mu}$ via the Einstein relation $\M{\chi}=\left(k_{B}T\right)\M{\mu}$. It is important to note that the diffusive motion of the particle is entirely determined by its coupling to the fluid and is not an input parameter to our method. This reflects the physical relationship between fluid velocity fluctuations (viscosity and temperature) and diffusion coefficient, as encoded in the Stokes-Einstein relation [@StokesLaw]. The Stokes-Einstein relation can formally be obtained starting from (\[eq:v\_t\_Stokes\]) by taking the *overdamped* limit $\nu=\eta/\rho\rightarrow\infty,$ which can be achieved by either assuming strong viscous friction $\eta\rightarrow\infty$, or no fluid inertia, $\rho\rightarrow0$. In this limit, momentum diffuses much faster than does the immersed particles, as measured by the Schmidt number $$\text{Sc}=\frac{\nu}{\chi}\gg1.$$ The overdamped limit is the formal limit $\text{Sc}\rightarrow\infty$, in which the fluid velocity becomes a fast degree of freedom that can be eliminated adiabatically, while the particle position becomes a slow degree of freedom [@SELM; @SELM_Reduction; @StokesLaw]. We emphasize that the Stokes-Einstein relation in a periodic domain of length $L\gg R_{H}$, $$\chi\rightarrow\chi_{\text{SE}}=\left(k_{B}T\right)\begin{cases} \left(4\pi\eta\right)^{-1}\ln\left[L/\left(3.708\, R_{H}\right)\right] & \mbox{ in two dimensions}\\ \left(6\pi\eta R_{H}\right)^{-1} & \mbox{ in three dimensions} \end{cases},$$ can only be justified in the limit $\text{Sc}\rightarrow\infty$. For a collection of neutrally-buoyant blobs, the multi-dimensional symmetric positive semi-definite mobility tensor $\M M\left(\V Q\right)=\left\{ \M{\mu}_{ij}\right\} $ (which is a block matrix with blocks of size $d^{2}$, where $d$ is the dimensionality) depends on the positions of all particles $\V Q=\left\{ \V q_{i}\right\} $. The diagonal block $\M{\mu}_{ii}$ corresponds to the single-particle mobility in the absence of other particles, while the block corresponding to particle pair $i$ and $j$ is the inter-particle mobility $\M{\mu}_{ij}$. For blobs this is given by a generalization of (\[eq:mu\_blob\]) [@SIBM_Brownian], $$\M{\mu}_{ij}=\M{\mu}_{ji}=\eta^{-1}\int\delta_{a}\left(\V q_{i}-\V r\right)\M G\left(\V r,\V r^{\prime}\right)\delta_{a}\left(\V q_{j}-\V r^{\prime}\right)\, d\V rd\V r^{\prime}.\label{eq:mu_blob_pair}$$ This pairwise hydrodynamic interaction between two blobs was studied numerically for blobs in Ref. [@ISIBM], and was shown to be closely-related to the well-known Rotne-Prager-Yamakawa (RPY) tensor used in Brownian dynamics simulations [@BrownianDynamics_DNA; @BD_LB_Ladd; @LBM_vs_BD_Burkhard]. In the overdamped limit, the collective diffusion of a collection of blobs can be described by the equations of Brownian dynamics [@SELM; @StokesLaw], $$\frac{d\V Q}{dt}=\M M\V{\mathcal{F}}+\sqrt{2k_{B}T}\,\M M^{\frac{1}{2}}\widetilde{\V{\mathcal{W}}}+k_{B}T\left(\frac{\partial}{\partial\V Q}\cdot\M M\right),\label{eq:brownian_dynamics}$$ where $\widetilde{\V{\mathcal{W}}}$ is a collection of independent white-noise processes and the noise is to be interpreted in the Ito sense, and $\V{\mathcal{F}}\left(\V Q\right)=\left\{ \V F_{i}\right\} $ are the forces applied on the particles. In this work we focus on the Brownian motion of a single particle, however, the effect of Schmidt number that we study here exists in multi-particle systems as well. It is important to point out that at times longer than the time $t_{L}=L^{2}/\nu$ it takes for momentum to traverse the system length $L$, the VACF starts decaying exponentially [@BrownianParticle_SIBM] and $\M{\chi}_{\text{VACF}}\left(t>t_{L}\right)\approx\mbox{\ensuremath{\M{\chi}}}$. Therefore, at times $t>t_{L}$ the diffusive motion of an isolated particle looks like simple Brownian motion and it becomes meaningful to use the long-time diffusion coefficient. Note however that the Brownian motions of *multiple* particles are *not* independent because of the correlations induced by the long-ranged (decaying as $\ln(r)$ in two dimensions and $r^{-1}$ in three dimensions) hydrodynamic interactions. The diffusion of a large collection of particles is therefore subtly but crucially different from that of a collection of independent Brownian walkers with diffusion coefficient $\chi$ [@StokesLaw]. \[sub:FiniteSc\]Finite Schmidt Numbers -------------------------------------- For finite Schmidt numbers, $\text{Sc}=O(1)$, theoretical analysis is significantly complicated by the fact that the particle moves while the velocity relaxes through viscous dissipation. We are not aware of any rigorous results regarding the diffusive motion of even a single particle, yet alone collective diffusion. The main theoretical approach in the literature are mode-mode coupling theories [@ModeModeCoupling], which are essentially a perturbative series in the strength of the thermal fluctuations. One key prediction of these theories is that the momentum diffusion coefficient $\nu=\eta/\rho$ should be augmented by the particle diffusion coefficient $\chi$ since the particle diffuses while the momentum diffuses. At the viscous time scale, $t>t_{\nu}=R_{H}^{2}/\nu$, conservation of momentum (hydrodynamics) in the fluid introduces a memory in the motion of the particle and the VACF $C(t)=d^{-1}\av{\V u(t)\cdot\V u(0)}$ decays algebraically rather than exponentially. The standard theory for the tail of the VACF (long-time behavior) [@VACF_Langevin] implicitly assumes that $\text{Sc}\gg1$, and leads to the conclusion that for an isolated particle in an infinite fluid asymptotically $C(t)\approx\left(t/t_{\nu}\right)^{-d/2}\sim\left(\nu t\right)^{-d/2}$. Self-consistent mode coupling theory predicts that at finite Schmidt numbers $C(t)\sim\left[\left(\chi+\nu\right)t\right]^{-d/2}$ [@VACF_Alder; @BilinearModeCoupling_SE]. This was confirmed to hold for blob particles numerically in Ref. [@ISIBM] for $\text{Sc}\gtrsim2$, with the caveat that $\chi$ was approximated by $\chi_{\text{SE}}$, the prediction of the Stokes-Einstein relation. It is not difficult to see that predictions of mode-mode coupling theories have to be approximate in nature since the diffusion coefficient of the particle, which is the *result* of the fluid-particle coupling, is used in the theory as an *input* to predict the corrections to the overdamped limit. Self-consistent mode-mode coupling theories are usually heuristic and thus also approximate. Based on the scaling of the tail of the VACF with Schmidt number, and the fact that the diffusion coefficient is given by the integral of the VACF, one may conjecture that, to leading order, the effect of finite Schmidt number can be captured by the modified Stokes-Einstein formula $$\chi\left(\nu+\chi\right)=\chi_{\text{SE}}\nu.\label{eq:SE_Bedeaux}$$ This relation was proposed as a self-consistent equation for $\chi$ in Refs. [@DiffusionRenormalization_I; @DiffusionRenormalization_II]. It is important to note, however, that both the short time *and* the long time part of the VACF contribute to the diffusion coefficient. Numerical results in Ref. [@ISIBM] indicate that the short-time VACF does not scale in the same way as the long-time tail, leading us to question the prediction (\[eq:SE\_Bedeaux\]). In particular, at very short times no rescaling is required to overlap the VACFs, and therefore, in principle, to fully overlap the VACFs over both short and long times, one would need to use an $\text{Sc}$-dependent non-uniform rescaling of time axes. One therefore expects the integral of the VACF to be somewhere in-between the Stokes-Einstein relation (corresponding to the short-time scaling) and (\[eq:SE\_Bedeaux\]) (corresponding to the long-time scaling). We are not aware of any detailed theory that successfully predicts the form of the corrections to the simple Stokes-Einstein prediction (\[eq:SE\]) for moderate values of $\text{Sc}$ (see Ref. [@BilinearModeCoupling_SE] for one attempt). In Ref. [@DiffusionRenormalization] a heuristic self-consistent theory of diffusion is proposed. The starting point is renormalization theory for the effective diffusion coefficient in the advection-diffusion equation for the concentration of a large number of passive tracers [^3]. The self-consistent theory suggests that in both two and three dimensions (c.f. Eqs. (29) and (30) in Ref. [@DiffusionRenormalization] in the case of no bare diffusion) $$\chi=\nu\left[1+2\frac{\chi_{\text{SE}}}{\nu}\right]^{1/2}-\nu=\nu\left[\left(1+\frac{2}{\text{Sc}^{\text{SE}}}\right)^{\frac{1}{2}}-1\right],\label{eq:chi_eff_2D_SC}$$ where we have defined a predicted Schmidt number $\text{Sc}^{\text{SE}}=\nu/\chi_{\text{SE}}$ (which is input to our simulations rather than output as is $\text{Sc}$). This prediction for the self-diffusion coefficient is the solution to the self-consistent equation $$\chi\left(\nu+\frac{\chi}{2}\right)=\chi_{\text{SE}}\nu,\label{eq:SE_self_consistent}$$ which differs from (\[eq:SE\_Bedeaux\]) in the prefactor of $1/2$. In the next section, we will compare the two predictions (\[eq:SE\_Bedeaux\]) and (\[eq:SE\_self\_consistent\]) for the leading-order correction to Stokes-Einstein’s relation with numerical results. Note that we only expect this type of relation to predict the leading-order corrections (proportional to $\text{Sc}^{-1}$) to the Stokes-Einstein relation and not the complete dependence on $\text{Sc}$. At very small Schmidt numbers there would likely be important contributions from higher-order (e.g., proportional to $\chi^{2}/\nu$) correction terms inside the parenthesis in (\[eq:SE\_self\_consistent\]). Mode-mode and renormalization theories, on which both (\[eq:SE\_Bedeaux\]) and (\[eq:SE\_self\_consistent\]) are based, are perturbative theories typically truncated at terms quadratic in the strength of the fluctuations. In the context of infinite (bulk) systems, a systematic perturbative theory that accounts for corrections of order higher than quadratic in the fluctuations has been discussed in Refs. [@DiffusionRenormalization_II; @TrilinearModeCoupling]. In three dimensions, the conclusion of such studies has been that the higher-order terms do not make a dramatic contribution. This can be attributed to the fact that large-scale modes of the fluctuating velocity make a negligible contribution to the diffusive dynamics (this is directly related to the fact that the $t^{-3/2}$ tail of the VACF is integrable, see Ref. [@DiffusionRenormalization] for additional discussion). In two dimensions, however, the logarithmic divergence of $\chi_{\text{SE}}$ with system size is a sign of the growing contribution of large-scale modes. In fact, the Schmidt number will become arbitrarily small for sufficiently large systems (keeping viscosity fixed) and the Stokes-Einstein relation will be strongly violated. In this case, however, the infinite-time diffusion coefficient $\chi$ does not describe the diffusive dynamics even at macroscopic length (e.g., system size $L$) and time scales ($L^{2}/\chi$). Specifically, the slowly-decaying $t^{-1}$ tail of the VACF means that the diffusive motion has correlations over macroscopic times and does not resemble simple Brownian motion at relevant scales. Instead, one needs to consider the time-dependent diffusion coefficients $\chi_{\text{MSD}}(t)$ or $\chi_{\text{VACF}}(t)$. Several calculations [@VACF_2D_SelfConsistentMC; @TrilinearModeCoupling; @SCModeModeCoupling2D] and numerical simulations [@VACF_2Divergence; @VACF_2D_HS] suggest that including higher-order terms changes the decay of the tail of the VACF in two dimensions. Specifically, it has been predicted that the self-consistent power-law decay for the VACF is faster than $t^{-1}$, $C(t)\sim\left(t\sqrt{\ln t}\right)^{-1}$. Hydrodynamics-based methods such as the method we use here (see also Ref. [@VACF_2Divergence] for a study using a lattice gas model) are better for studying this very long time decay of the VACF than are particle methods, since larger systems and longer time scales become computationally-feasible. Nevertheless, because of the slow logarithmic growth, extremely large systems are required to see any effects of higher-order corrections. To see this, let us assume Stokes-Einstein’s formula holds and estimate the system size when $$\text{Sc}\approx\text{Sc}^{\text{SE}}=\frac{\nu}{\chi_{\text{SE}}}=\frac{4\pi\rho\nu^{2}}{k_{B}T\ln\left[L/\left(4\, R_{H}\right)\right]}<1.$$ For a hard-disk fluid, we can estimate the viscosity $\eta$ using simple Enskog kinetic theory [@Enskog_2DHS], which has been found to be quite accurate even at high densities [@Hard_Disks_Transport]. For a hard-disk fluid at packing density (fraction) $\phi=\left(\rho/m\right)\left(\pi\sigma^{2}/4\right)=0.6$ (which is close to the liquid-solid transition), Enskog theory predicts $$\eta\approx2\frac{\sqrt{mk_{B}T}}{\sigma},$$ where $\sigma\approx2R_{H}$ denotes the particle diameter. This gives that the system size required to get Schmidt number close to unity, $$\text{Sc}^{\text{SE}}\approx\frac{65}{\ln\left[L/\left(4\, R_{H}\right)\right]}<1\quad\Rightarrow\quad L>1.5\cdot10^{29}\, R_{H}.$$ Reaching such a staggering system size is not feasible with any numerical method, and therefore in this work we focus on the more practically-relevant case of finite system size at moderately large Schmidt numbers. In addition to the nonlinearity coming from the fact that at moderate Schmidt numbers the particle moves while momentum diffuses [^4], additional nonlinearities arise because of the presence of the advective term $\rho\V v\cdot\grad\V v$ in the momentum equation. If we define a “thermal” Reynolds number based on the equilibrium magnitude of the particle thermal velocity fluctuations [@ISIBM] $$\av{u^{2}}^{\frac{1}{2}}\approx\sqrt{\frac{k_{B}T}{\rho R_{H}^{d}}},$$ in three dimensions we can estimate $$\text{Re}=\frac{\av{u^{2}}^{\frac{1}{2}}R_{H}}{\nu}\approx\sqrt{\frac{k_{B}T}{\nu\eta R_{H}}}\sim\sqrt{\frac{\chi_{\text{SE}}}{\nu}}=\left(\text{Sc}^{\text{SE}}\right)^{-\frac{1}{2}}.$$ A similar result applies in two dimensions as well, except for an additional factor of $\ln^{\frac{1}{2}}\left[L/\left(4\, R_{H}\right)\right]$. This shows that the thermal Reynolds number becomes $O(1)$ when $\text{Sc}^{\text{SE}}\sim1$, and in principle nonlinear effects arising from advection could arise. By contrast, at large Schmidt numbers the thermal Reynolds number is small and advective nonlinearities are expected to be negligible. While in realistic fluids, including particle simulations, advective nonlinearities are always present, in our numerical method we can simply turn off the term $\rho\V v\cdot\grad\V v$ to assess its importance. Numerical experiments indicate that nonlinear inertial effects play a minimal role, not noticeable within statistical accuracy for reasonable Schmidt numbers. Therefore, it appears appropriate to simply linearize the velocity equation, as is often done in the literature with somewhat hand-waving justifications. Similarly, we find that the excess inertia of the particles does not affect the long-time diffusion coefficient, consistent with traditional derivations based on linearized theory [@VACF_Langevin]. We do not attempt to provide rigorous justification for these observations in this work. \[sec:Results\]Results ====================== In this section we present numerical results obtained using the computational algorithm for solving (\[eq:v\_t\]-\[eq:no\_slip\]) developed in Ref. [@ISIBM]. In Ref. [@ISIBM] some of us presented results concerning the short-time behavior of the VACF $C(t)$ of a single blob particle in a three-dimensional periodic domain of length $L$, including a comparison to the case of a blob suspended in a compressible fluid. Excellent agreement with theoretical predictions for the variance of velocity (i.e., $C(t=0)$) was obtained, and the $t^{-3/2}$ power-law behavior at long times $L^{2}/\nu>t>t_{\nu}=R_{H}^{2}/\nu$ was confirmed. Here we briefly examine the VACF in two dimensions, and then focus on the self-diffusion of a blob in both two and three dimensions. Note that in Ref. [@ISIBM] it is demonstrated that the method correctly reproduces the static structure of a suspension of many interacting blobs. Also note that the presence of multiple interacting blobs affects the dynamics in a nontrivial way, and changes both the short-time VACF (via multi-particle inertial effects) and also the self-diffusion coefficient. We do not address multi-particle suspensions in this work. In the majority of the simulations we use the three-point discrete kernel function of Roma and Peskin [@StaggeredIBM; @DirectForcing_Balboa] to discretize the kernel $\delta_{a}$. By using the Peskin four-point kernel [@IBM_PeskinReview] instead of the three-point discrete kernel function the translational invariance of the spatial discretization can be improved, however, at a potentially significant increase in computational cost, particularly in three dimensions. The effective hydrodynamic radius $R_{H}$ for a given discrete kernel function can be obtained easily from the deterministic mobility tensor $\V{\mu}_{\text{det}}$, given by a discrete equivalent of (\[eq:mu\_blob\]). The deterministic mobility tensor can be obtained by turning off fluctuations, applying a unit force along each of the coordinate directions in turn, solving the spatially-discretized steady Stokes equation, and then calculating the resulting velocity of the particle. After accounting for finite-size effects due to the finite length of the periodic box, in three dimensions we numerically estimate [@DirectForcing_Balboa; @ISIBM] the effective hydrodynamic radius to be $R_{H}^{3pt}=\left(0.91\pm0.01\right)h$ the three-point kernel [^5], where $h$ is the grid spacing, and $R_{H}^{4pt}=\left(1.255\pm0.005\right)h$ for the 4pt kernel [@ReactiveBlobs]. In two dimensions, the effective (rigid disk) hydrodynamic radii are estimated to be $R_{H}^{3pt}=\left(0.72\pm0.01\right)h$ and $R_{H}^{4pt}=\left(1.04\pm0.005\right)h$. Note that the spatial discretization we use is not perfectly translationally invariant and there is a small variation of $R_{H}$ (quoted above as an error bar) as the particle moves relative to the underlying fixed fluid grid [@ISIBM; @ReactiveBlobs]. We use a relatively small grid of $32^{2}$ cells in two dimensions and $32^{3}$ cells in three dimensions in order to be able to perform sufficiently long runs even with the larger Schmidt numbers. In two dimensions we use a neutrally-buoyant particle ($m_{e}=0$), while in three dimensions we use a particle twice denser than the surrounding fluid ($m_{e}=\rho\D V$), in order to confirm that the excess mass (density) does not (significantly) affect the conclusions of our investigations. The advantage of using neutrally-buoyant blobs is that they are passive tracers (they do not affect the velocity equation), and therefore one can use multiple tracers in a single simulation in order to improve the statistical accuracy. This is useful in two dimensions, where the VACF has a slowly-decaying tail and therefore long runs are required to study the long-time diffusive motion of the particle. In all cases we have confirmed that the time step size $\D t$ is sufficiently small by comparing with a simulation using a twice smaller time step size. A detailed discussion and numerical results concerning the accuracy of the scheme as a function of $\D t$ are given in Ref. [@ISIBM]. We varied the viscosity in order to change the Schmidt number, but as explained earlier, the Schmidt number is the only relevant dimensionless number so one can equivalently change the temperature. Since the actual Schmidt number is an output rather than an input to our calculation, we cannot calculate the Schmidt number a priori. Therefore, in this section we estimate the true Schmidt number with $\text{Sc}\approx\text{Sc}^{\text{SE}}=\nu/\chi_{\text{SE}}$ for the purposes of captions and axes labels, which is a good approximation for moderate and large Schmidt numbers. During each simulation a discrete trajectory of the particle $\left\{ \V q\left(0\right),\,\V q\left(\D t\right),\,\V q\left(2\D t\right),\dots\right\} $ is recorded. From this data, we calculate an apparent (discrete) velocity $\tilde{\V u}_{k}=\left(\V q(k+1)-\V q(k)\right)/\D t,$ $k=1,2,\dots$, which for finite $\D t$ is in general different from the velocity of the particle $\V u\left(k\D t\right)$ calculated by the numerical scheme. We then obtain the discrete VACF $C_{k}=d^{-1}\av{\left(\tilde{\V u}_{k^{\prime}}\right)\cdot\left(\tilde{\V u}_{k^{\prime}+k}\right)}$ efficiently using a fast Fourier Transform of the apparent velocity. It is not hard to show that the time-dependent diffusion coefficients at the discrete time points in time can be obtained in linear time using $$\begin{aligned} \chi_{\text{MSD}}(k\D t) & = & \frac{\D t}{2}C_{0}+\D t\left[\sum_{i=1}^{k-1}C_{i}-\frac{1}{k}\sum_{i=1}^{k-1}i\cdot C_{i}\right]\label{eq:chi_MSD_num}\\ \chi_{\text{VACF}}(k\D t) & \approx & \frac{\D t}{2}C_{0}+\D t\,\sum_{i=1}^{k-1}C_{i}.\label{eq:chi_VACF_num}\end{aligned}$$ \[sub:VACF\_short\]VACF ----------------------- In Fig. \[fig:VACF\] we show the VACF $C(t)=d^{-1}\av{\V u(t)\cdot\V u(0)}$ for a neutrally-buoyant ($m_{e}=0$) blob in two dimensions, for several Schmidt numbers. The theoretical variance of the velocity of a neutrally-buoyant particle immersed in an incompressible two dimensional fluid gives $C(t=0)=(d-1)k_{B}T/(d\, m)=k_{B}T/\left(2\rho\D V\right)$, where $\D V$ is the volume of the blob given in Eq. (\[eq:dV\_JS\]). We see from the figure that the numerical curves are in excellent agreement with this value, confirming that the numerical method correctly captures the thermal particle velocity fluctuations. The standard theory for the tail of the VACF (long-time behavior) [@VACF_Langevin] implicitly assumes that $\text{Sc}\gg1$, and predicts that the long-time VACF has a power law decay $C(t)\sim\left(\nu t\right)^{-1}$. As discussed in Section \[sub:FiniteSc\], accounting for a finite Schmidt number leads to a predicted decay $C(t)\sim\left[\left(\nu+\chi\right)t\right]^{-1}$. This means that the tails of the VACFs for different $\text{Sc}$ values can be collapsed on one master curve if we plot them as a function of $\left(1+\text{Sc}^{-1}\right)\left(t/t_{\nu}\right)$. This was confirmed in Ref. [@ISIBM] in three dimensions, and it is confirmed in two dimensions in Fig. \[fig:VACF\]. The rescaling is not perfect at shorter times, especially for small Schmidt numbers, however, we note that the collapse is significantly poorer if one plots the VACF as a function of just $t/t_{\nu}$. ![\[fig:VACF\]Velocity autocorrelation function (VACF) of a single neutrally-buoyant blob (three-point kernel) in two dimensions. A deterministic calculation, corresponding to the limit $\text{Sc}\rightarrow\infty$, is also shown. The expected value at the origin is shown with a dotted line. The scaling of the time axes is chosen to collapse the different curves at long times. Note that the time $t_{L}=L^{2}/\nu\approx10^{3}t_{\nu}$ is off the scale of the time axis.](15_home_donev_Papers_ISIBM_graphics_BlobVACF_2D_3pt.eps){width="0.75\columnwidth"} Our numerical method becomes very inefficient as the Schmidt number becomes very large due to the large separation of scales between the momentum and particle diffusion. We will present modifications of the numerical method to handle the limit $\text{Sc}\rightarrow\infty$ in future work. The VACF in the limit $\text{Sc}\rightarrow\infty$ can, however, easily be obtained by a *deterministic* calculation. We simply give the particle a small momentum kick $\D p=F\D t$ during the first time step, and then the deterministic algorithm is used to track the subsequent decay of the velocity $u(t)$, which gives the VACF in the limit after a suitable rescaling. In particular, the time-dependent diffusion coefficient in the limit $\text{Sc}\rightarrow\infty$ is given by $$\lim_{\text{Sc}\rightarrow\infty}\chi_{\text{VACF}}(t)=\frac{k_{B}T}{\D p}\int_{0}^{t}u(t^{\prime})dt^{\prime}.$$ In Fig. \[fig:VACF\] we show also the result of such a deterministic calculation. Due to lack of statistical noise, the deterministic VACF shows very clearly the transition from a power-law to an exponential tail at long times. It is also clearly seen that the VACF at large Schmidt numbers closely matches the deterministic one, as expected. Note that, in principle, the deterministic VACF can be calculated analytically using discrete Fourier Transform techniques [@BrownianParticle_SIBM]. \[sub:MSD\]Diffusion Coefficient -------------------------------- The long-time diffusion coefficient can be obtained either from the limiting value of $\chi_{\text{MSD}}(t)$ or $\chi_{\text{VACF}}(t)$. It is preferable to use the discrete integral of the (discrete) VACF rather than the mean square displacement because $\chi_{\text{VACF}}$ converges faster to the asymptotic (long-time) value [^6], and thus has smaller statistical error. Because the position of the particle is a more fundamental quantity (in particular, in the limit $\text{Sc}\rightarrow\infty$ it becomes the only relevant variable), the mean square displacement has a more direct physical interpretation, and is directly obtained from the discrete particle trajectory. We therefore use $\chi_{\text{MSD}}(t)$ to illustrate the time dependence of the diffusion coefficient, and use $\chi_{\text{VACF}}(t)$ to obtain the long-time diffusion coefficient $\chi$. In the left panel of Fig. \[fig:MSD-3D\] we show the time-dependent diffusion coefficient $\chi_{\text{MSD}}(t)$ for a single particle in a three dimensional periodic domain at several Schmidt numbers, including the limit $\text{Sc}\rightarrow\infty$ as obtained from a deterministic simulation. We see that at short times we obtain collapse of all of the curves just by scaling the time axes by $t_{\nu}=R_{H}^{2}/\nu$, however, at long times the curves do not collapse. In particular, the diffusion coefficient is lower than the Stokes-Einstein prediction for the smaller Schmidt numbers. In order to obtain an estimate of the long-time diffusion coefficient that is essentially converged to the asymptotic value, while at the same time minimizing the statistical errors, we estimate $\chi\approx\chi_{\text{VACF}}(L^{2}/\left(4\nu\right))$. In the right panel of Fig. \[fig:MSD-3D\] we show the estimated $\chi/\chi_{\text{SE}}$ as a function of the approximate Schmidt number $\text{Sc}^{\text{SE}}$. The numerical data is compared to the two self-consistent theories, (\[eq:SE\_Bedeaux\]) and (\[eq:SE\_self\_consistent\]). We see good agreement of the numerical data with (\[eq:SE\_self\_consistent\]) to within the statistical accuracy. {width="0.5\columnwidth"}{width="0.5\columnwidth"} Our numerical results clearly demonstrate that the Stokes-Einstein relation does not apply at finite Schmidt numbers. On the other hand, the Einstein formula, which is a linear response relationship between the diffusion coefficient and the mobility, $\chi=\left(k_{B}T\right)\mu$, is quite general, and follows from straightforward statistical mechanics arguments [@FluctuationDissipation_Kubo]. However, one must be careful here in defining the mobility $\mu$. Mobility should be defined as the coefficient of proportionality between a small applied force $F_{0}$ on the particle (with the opposite force applied uniformly to the fluid to prevent center-of-mass motion) and the *average* velocity of the particle, $$\mu=\lim_{F_{0}\rightarrow0}\frac{\av{\V u}}{\V F_{0}},\label{eq:Stokes_Einstein_true}$$ in the *presence* of thermal fluctuations. Here the average is taken over time, or, equivalently, over a nonequilibrium but steady-state ensemble that is a weak perturbation of the equilibrium ensemble. Because of the nonlinearity of the fluid-particle equations, thermal fluctuations *can* affect average values. Therefore, we should not expect, in general, that the mobility measured at finite temperature is the same as the deterministic ($k_{B}T=0$) mobility. Indeed, in the right panel of Fig. \[fig:MSD-3D\] we show the predictions of (\[eq:Stokes\_Einstein\_true\]) with $\av{\V u}$ measured numerically under a small applied force (to ensure the system remains in the linear response regime). These predictions are in agreement with the direct measurements of the diffusion coefficient from the integral of the velocity autocorrelation function, as predicted by linear response theory. We therefore observe no violation of any known physical principles, so long as we recognize the fact that fluctuations cannot just be turned off. More precisely, at finite value of $k_{B}T$ the fluctuations do not necessarily average out, and can, in fact, affect the dynamics of macroscopic or deterministic observables. {width="0.5\columnwidth"}{width="0.5\columnwidth"} In Fig. \[fig:MSD-2D\] we confirm that the same conclusions apply to finite two-dimensional systems. In order to verify that the results are not affected by the specific choice of the kernel function, here we also use the four-point Peskin kernel [@IBM_PeskinReview]. The results in the right panel of Fig. \[fig:MSD-2D\] show that dimensionless relations are independent of the kernel function. This leads us to believe that (\[eq:SE\_self\_consistent\]) is a rather general relation that correctly captures the leading-order corrections to the Stokes-Einstein relation at finite Schmidt numbers in both two and three dimensions. \[sec:Conclusions\]Conclusions ============================== We presented a simple theory for the effect of Schmidt number $\text{Sc}=\nu/\chi$ on the long-time diffusion coefficient of a particle immersed in a fluid. The approximate self-consistent calculation suggests that the long-time diffusion coefficient of the particle can be determined from the condition $\chi\left(\nu+\chi/2\right)=\chi_{\text{SE}}\nu$. Here $\chi_{\text{SE}}$ is the prediction of the Stokes-Einstein relation, $\chi_{\text{SE}}=k_{B}T\mu_{det}$, where $\mu_{\text{det}}$ is the deterministic mobility of the particle, that is, the linear response to an applied force in the *absence* of thermal fluctuations. In the limit of large Schmidt numbers the Stokes-Einstein relation is obeyed, as expected, but the correction can be substantial at small Schmidt numbers and is also measurable at moderate Schmidt numbers. By contrast, the Einstein relation $\chi=k_{B}T\mu$ is always obeyed, as long as the mobility is defined by the linear response in the *presence* of fluctuations. The discrepancy between $\mu$ and $\mu_{\text{det}}$ stems from the nonlinear fluid-particle coupling, which becomes important at finite Schmidt numbers. Our numerical results are in good agreement with the self-consistent theory in both two and three dimensions. In two dimensions, for finite systems, Stokes-Einstein’s relation holds, but the Stokes-Einstein diffusion coefficient diverges logarithmically with system size. At the same time, however, as the system size diverges the Schmidt number decreases and the deviations from Stokes-Einstein’s prediction become stronger. The truly asymptotic behavior of a particle in an unbounded two-dimensional fluid (e.g., a tagged hard disk in a hard-disk fluid) remains an open question, albeit one of mostly academic interest since the Schmidt number is still rather large for system sizes feasible in present-day molecular dynamics. While the above results elucidate some important fundamental questions about the physical importance of fluctuations, it should be observed that for realistic fluids (rather than coarse-grained fluid models [@MPC_LLNS]) small $\text{Sc}$ implies that the size of the blob particle becomes comparable to the length scale at which hydrodynamics breaks down (molecular diameter for liquids, or mean free path for gases). The physical fidelity of the blob model and the fluctuating hydrodynamics equations [@MPC_LLNS] used here should therefore be questioned at small $\text{Sc}$. In Ref. [@StokesEinstein_MD] the validity of the Stokes-Einstein relation is studied using molecular dynamics. In that study the smallest Schmidt number is on the order of 15, and the corrections studied here are significant. Nevertheless, in the case of a true molecular fluid many other effects enter the picture as well, for example, the (in)validity of the no-slip condition with a fixed particle radius. Therefore, one should not try to explain the data in Ref. [@StokesEinstein_MD] using our model. In fact, the deviations from the SE prediction measured in Ref. [@StokesEinstein_MD] have a comparable magnitude but the *opposite* sign from the correction predicted here. This means that, when one takes into account that hydrodynamic theory should include the finite Schmidt-number correction studied here, the deviations between the MD simulation and hydrodynamic predictions are twice larger than reported in Ref. [@StokesEinstein_MD]. A similar “cancellation-of-errors” that leads to a better agreement with hydrodynamic theory than truthfully present is discussed in Ref. [@StokesEinstein_BCs]. Despite its somewhat limited practical utility, our model allows us to separate hydrodynamic from molecular effects, and to test the predictions of self-consistent renormalization theory [@DiffusionRenormalization]. More importantly, our results are quite relevant to the interpretation of results obtained using widely-used particle methods for fluid dynamics of complex fluids, e.g., MD, (S)DPD, MPCD/SRD, SHSD, and even Lattice Boltzmann schemes. In some of these methods, as they are typically used, the Schmidt number can be on the order of unity, and in all of them it is rarely larger than 50. This is to be contrasted with realistic liquids (e.g., water at room conditions) where the Schmidt number for tagged particles exceeds 100 and is often larger than 1000. In order to avoid the types of unphysical deviations from the Stokes-Einstein relation as we described here, one must use methods that take the limit $\text{Sc}\rightarrow\infty$ and use a fluctuating *steady* Stokes equation for the fluid (e.g., Brownian and Stokesian dynamics). In the future we will develop modifications of our method that achieves this limiting behavior. Note that increasing $\text{Sc}$ is not possible in traditional particle methods or Lattice Boltzmann methods because they (by design) avoid the solution of Poisson equations for the pressure, as is necessary in the steady Stokes limit. The primary cause of the correction to the SE prediction at finite $\text{Sc}$ is the fact that the particle diffuses with effective diffusion coefficient $\chi$ *while* the momentum diffuses, rather than nonlinear inertial effects. This leads us to suspect that the effect studied here is independent of the details of the microscopic dynamics giving rise to the particle diffusion, and the same self-consistent formula should apply, for example, in Lattice Boltzmann methods with a frictional fluid-particle coupling [@LB_SoftMatter_Review]. Note that in Stokes frictional coupling, and likely also in molecular (particle) simulations, there is an additional non-hydrodynamic or “bare” contribution to the diffusion coefficient $\chi_{0}=k_{B}T/\gamma$ coming from the dissipative friction coefficient $\gamma$ between the particle and the fluid [@Friction_Diffusion_SE; @LB_SoftMatter_Review]. A way to estimate such bare friction in molecules, from the autocorrelation of molecular forces in restraint dynamics, has been proposed in Ref. [@MoriZwanzig_ConstrainedMD]. The bare diffusion coefficient $\chi_{0}$ should be added to the (Schmidt-number dependent) hydrodynamic contribution $\chi_{H}=\chi-\chi_{0}$ studied in this work. In both two and three dimensions, the self-consistent theory proposed in Ref. [@DiffusionRenormalization] gives the prediction $$\chi=\chi_{0}+\left[1+\frac{\chi+\chi_{0}}{2\nu}\right]^{-1}\chi_{\text{SE}},$$ which can be written as a self-consistent equation for $\chi_{H}$, $$\chi_{H}\left(\nu+\chi_{0}+\frac{\chi_{H}}{2}\right)=\chi_{\text{SE}}\nu.$$ It would be interesting to perform numerical simulations in the future to test this theory numerically. We thank Burkhard Dünweg and Anthony Ladd for informative discussions. A. Donev was supported in part by the Air Force Office of Scientific Research under grant number FA9550-12-1-0356. R. Delgado-Buscalioni and F. Balboa acknowledge funding from the Spanish government FIS2010-22047-C05 and from the Comunidad de Madrid MODELICO-CM (S2009/ESP-1691). Collaboration between A. Donev and R. Delgado-Buscalioni was fostered at the Kavli Institute for Theoretical Physics in Santa Barbara, California, and supported in part by the National Science Foundation under Grant No. NSF PHY05-51164. [10]{} L.D. Landau and E.M. Lifshitz. , volume 6 of [*Course of Theoretical Physics*]{}. Pergamon Press, Oxford, England, 1959. J. T. Hynes, R. Kapral, and M. Weinberg. . , 70(3):1456, February 1979. J.R. Schmidt and J.L. Skinner. Hydrodynamic boundary conditions, the stokes–einstein law, and long-time tails in the brownian limit. , 119:8062, 2003. D. Bedeaux and P. Mazur. . , 80:189–202, 1975. T. Keyes and Irwin Oppenheim. Bilinear hydrodynamics and the stokes-einstein law. , 8:937–949, 1973. Berni J Alder and W Edward Alley. Generalized hydrodynamics. , 37:56, 1984. P. J. Atzberger. . , 351(4-5):225–230, 2006. X. Bian, S. Litvinov, R. Qian, M. Ellero, and N. A. Adams. . , 24(1):012002, 2012. A. V[á]{}zquez-Quesada, M. Ellero, and P. Espa[ñ]{}ol. . , 130:034901, 2009. A. Donev, A. L. Garcia, and B. J. Alder. . , 101:075902, 2008. Chien-Cheng Huang, Gerhard Gompper, and Roland G. Winkler. Hydrodynamic correlations in multiparticle collision dynamics fluids. , 86:056711, 2012. R. Kapral. . , 140:89, 2008. G. Gompper, T. Ihle, D. Kroll, and R. Winkler. . , pages 1–87, 2009. B. D[ü]{}nweg and A.J.C. Ladd. . , pages 89–166, 2009. P. J. Atzberger, P. R. Kramer, and C. S. Peskin. . , 224:1255–1292, 2007. P. J. Atzberger. . , 230:2821–2837, 2011. F. Balboa Usabiaga, R. Delgado-Buscalioni, B. E. Griffith, and A. Donev. . , 2013. In press, Arxiv preprint 1212.6427, code available at <https://code.google.com/p/fluam>. Tri T Pham, Ulf D Schiller, J Ravi Prakash, and Burkhard D[ü]{}nweg. Implicit and explicit solvent models for the simulation of a single polymer chain in solution: Lattice boltzmann versus brownian dynamics. , 131:164114, 2009. Anthony JC Ladd, Rahul Kekre, and Jason E Butler. . , 80(3):036704, 2009. J Courtenay Lewis. On the einstein-stokes diffusion coefficient for brownian motion in two dimensions. , 44(4):245–246, 1973. A. Donev, A. L. Garcia, Anton de la Fuente, and J. B. Bell. . , 2011:P06014, 2011. J. M. O. De Zarate and J. V. Sengers. . Elsevier Science Ltd, 2006. Song Hi Lee and Raymond Kapral. Friction and diffusion of a brownian particle in a mesoscopic solvent. , 121:11163, 2004. A. Donev, T. G. Fai, and E. Vanden-Eijnden. . Arxiv preprint 1306.3158, 2013. G. Tabak and P.J. Atzberger. Systematic stochastic reduction of inertial fluid-structure interactions subject to thermal fluctuations. , 2013. F. Balboa Usabiaga, J. B. Bell, R. Delgado-Buscalioni, A. Donev, T. G. Fai, B. E. Griffith, and C. S. Peskin. . , 10(4):1369–1408, 2012. C.S. Peskin. . , 11:479–517, 2002. P. J. Atzberger. . , 226(2):144–150, 2007. H Hasimoto. On the periodic fundamental solutions of the stokes equations and their application to viscous flow past a cubic array of spheres. , 5(02):317–328, 1959. F. Balboa Usabiaga, I. Pagonabarraga, and R. Delgado-Buscalioni. . , 235:701–722, 2013. Alexandre M Roma, Charles S Peskin, and Marsha J Berger. An adaptive version of the immersed boundary method. , 153(2):509–534, 1999. E. J. Hinch. . , 72(03):499–511, 1975. R. M. Jendrejack, J. J. de Pablo, and M. D. Graham. . , 116(17):7752–7759, 2002. Rahul Kekre, Jason E. Butler, and Anthony J. C. Ladd. . , 82:011802, 2010. Y. [Pomeau]{} and P. [R[é]{}sibois]{}. . , 19:63–139, June 1975. B. J. Alder and T. E. Wainwright. Decay of the velocity autocorrelation function. , 1(1):18–21, 1970. D. Bedeaux and P. Mazur. . , 73:431–458, 1974. P. Mazur and D. Bedeaux. . , 75:79–99, 1974. I. A. Michaels and I. Oppenheim. . , 81(4):522–534, 1975. T. E. Wainwright, B. J. Alder, and D. M. Gass. Decay of time correlations in two dimensions. , 4(1):233–237, 1971. H.H.H. Yuan and I. Oppenheim. . , 90(3-4):561–573, 1978. C. P. Lowe and D. Frenkel. . , 220(3-4):251–260, 1995. Masaharu Isobe. Long-time tail of the velocity autocorrelation function in a two-dimensional moderately dense hard-disk fluid. , 77(2):021201, 2008. David M. Gass. Enskog theory for a rigid disk fluid. , 54(5):1898–1902, 1971. R. García-Rojo, S. Luding, and J. J. Brey. Transport coefficients for dense hard-disk systems. , 74(6):061305, 2006. A. Pal Singh Bhalla, B. E. Griffith, N. A. Patankar, and A. Donev. . , 2013. In press, Arxiv preprint 1306.3159. R. Kubo. . , 29(1):255–284, 1966. C. Hij[ó]{}n, P. Espa[ñ]{}ol, E. Vanden-Eijnden, and R. Delgado-Buscalioni. Mori-zwanzig formalism as a practical computational tool. , 144:301–322, 2009. [^1]: For unbounded three-dimensional systems the Oseen tensor is $\V G\left(\V r^{\prime},\V r^{\prime\prime}\right)=\left(8\pi r\right)^{-1}\left(\M I+r^{-2}\V r\otimes\V r\right)$, where $\V r=\V r^{\prime}-\V r^{\prime\prime}$. [^2]: Note that for truly two-dimensional systems $\rho$ has units of $\mbox{kg}/\mbox{m}^{2}$, unlike three dimensions where it has units $\mbox{kg}/\mbox{m}^{3}$. This accounts for the difference in units of viscosity $\eta$ in the Stokes-Einstein relation ($\nu$ has units of $\mbox{m}^{2}/\mbox{s}$ independent of the dimension). [^3]: Note that freely diffusing neutrally-buoyant blobs are passive tracers since they simply follow the fluid but do not affect it. [^4]: In our model, (\[eq:no\_slip\]) has a non-linear coupling between fluid and particle velocities. In more traditional models the no-slip boundary condition needs to be applied on a moving surface, which is hard and not done in most models, see for example the assumptions made in Ref. [@VACF_Langevin]. [^5]: The small variation around the average value comes because the spatial discretization is only approximately translationally invariant. [^6]: This can be seen from the fact that the last term in (\[eq:chi\_MSD\_num\]) does not appear in (\[eq:chi\_VACF\_num\]).

ArXiv

--- abstract: | In this paper we study short exact sequences $ 0 \to \mathcal P \to \mathcal N \to {\mathcal I}_D(k) \to 0 $ with $ \mathcal P, \mathcal N $ torsion–free sheaves and $ D $ closed projective scheme. This is a classical way to construct and study projective schemes (e.g. see [@hart-1974], [@hart-2], [@mdp], [@serre-1960]). In particular, we give homological conditions on $ \mathcal P $ and $ \mathcal N $ that force $ D $ to be ACM, without constrains on its codimension. As last result, we prove that if $ \mathcal N $ is a higher syzygy sheaf of an ACM scheme $ X,$ the scheme $ D $ we get contains $ X.$ address: - 'S. Greco, Dipartimento di Matematica, Politecnico di Torino, I-10129 Torino, Italia' - 'Roberto Notari, Dipartimento di Matematica “Francesco Brioschi", Politecnico di Milano, I-20133 Milano, Italia' - 'M.L. Spreafico, Dipartimento di Matematica, Politecnico di Torino, I-10129 Torino, Italia' author: - 'S. Greco' - 'R. Notari' - 'M.L. Spreafico' title: 'Torsion–free sheaves and ACM schemes' --- [^1] Introduction ============ Homological methods have proved to be very useful in studying projective schemes. For example, many information on the geometry of a closed scheme $ X \subseteq {\mathbb P}^r $ are encoded in the minimal free resolution of the saturated ideal $ I_X $ of $ X.$ Homological methods are used also to construct schemes with prescribed properties. For example, in [@mdp], M. Martin–Deschamps and D. Perrin gave a homological construction of the ideal of a curve $ C $ in $ {\mathbb P}^3 $ with a prescribed Hartshorne–Rao module and of minimal degree. In more details, given a graded Artinian $ R:=K[x,y,z,w]$–module $ M $ with minimal free resolution $$0 \to L_4 \to L_3 \to L_2 \to L_1 \to L_0 \to M \to 0,$$ they show how to compute a free graded $ R$–module $ P $ such that the cokernel of a general injective map $ \gamma: P \to N := \ker(L_1 \to L_0) $ is isomorphic to the saturated ideal of a locally Cohen–Macaulay curve $ C \subset {\mathbb P}^3,$ up to a shift in grading, that is to say, they produce a short exact sequence $$\label{PNI} 0 \to P \stackrel{\gamma}{\longrightarrow} N \to I_C(k) \to 0.$$ An analogous sequence was used first by J.P.Serre in [@serre-1960] to construct subcanonical curves in $ {\mathbb P}^3.$ To this end, he considered a rank $ 2 $ vector bundle $ \mathcal N, $ a global section $ s $ whose zero–set has codimension $ 2,$ and the corresponding map $ {\mathcal O}\stackrel{s}{\to} \mathcal N. $ The image of the dual map $ \mathcal N^\vee \to {\mathcal O}$ is the ideal sheaf of a subcanonical curve $ C \subset {\mathbb P}^3.$ J.P. Serre’s construction was generalized to construct codimension $ 2 $ schemes in $ {\mathbb P}^r $ (see [@hart-1974], [@okonek], among others) and to sections, whose zero–set has codimension $ 2,$ of reflexive rank $ 2 $ sheaves on $ {\mathbb P}^3 $ by R. Hartshorne (see [@hart-2]). In the new more general setting, the constructed curves were generically locally complete intersection curves. While studying the construction of minimal curves by M. Martin–Deschamps and D. Perrin given in [@mdp], we applied it to syzygy modules of $ 0$–dimensional schemes of $ {\mathbb P}^3 $ instead of syzygy modules of graded Artinian $ R$–modules. The curves we produced were all arithmetically Cohen–Macaulay. To understand why the curves share this unexpected property, we were led to consider all the previous apparently different constructions from the same point of view, getting as result a quite general construction of arithmetically Cohen–Macaulay schemes of arbitrary codimension. For particular choices, we construct arithmetically Cohen–Macaulay schemes containing a given scheme with the same property but of larger codimension. We outline the structure of the paper. In section 2, first of all we describe some properties of torsion–free coherent sheaves, and their cohomology. Then, we get some bounds on the projective dimensions of $ \mathcal N $ and $ \mathcal P $ in terms of the codimension of $ D $ and of the cohomology of its ideal sheaf $ \mathcal I_D. $ Finally, we recall the well known result of Martin–Deschamps and Perrin, described in [@mdp], about maximal subsheaves which allows us to assure that the cokernel of a given injective map $ \mathcal P \to \mathcal N $ is an ideal sheaf. Section 3 is the heart of the paper. At first, we give some conditions on the coherent torsion–free sheaves $ \mathcal N $ and $ \mathcal P $ to assure that the short exact sequence (\[PNI\]) ends with the ideal sheaf of a closed arithmetically Cohen–Macaulay subscheme $ D $ of $ {\mathbb P}^r $ of codimension $ 2+{\operatorname{pd}}(\mathcal P),$ where $ {\operatorname{pd}}(\mathcal P) $ is the projective dimension of $ \mathcal P.$ Moreover, we show that the construction characterizes the couple $ (D, \mathcal P) $ in the sense that starting from an arithmetically Cohen–Macaulay scheme $ D $ and a torsion–free coherent sheaf $ \mathcal P,$ we can construct a sheaf $ \mathcal N $ fulfilling our conditions. In the codimension $ 2 $ case we give a geometrical description of our construction associating to any non–zero element of $ H^0(D, \omega_D(c)) $ an extension as (\[PNI\]). This is a new reading of the analogous result of [@serre-1960], for coherent torsion–free sheaves, without bounds on the rank of $ \mathcal N. $ We show also that some schemes we obtain in our setting cannot be obtained with Hartshorne’s construction, and conversely. So, the two constructions are not the same one. Section 4 is devoted to solve the problem of finding a codimension $ s $ closed scheme $ D $ containing a given codimension $ t (>s) $ scheme $ X,$ them both arithmetically Cohen–Macaulay. We end the section with some examples. Preliminary results =================== Let $ K $ be an algebraically closed field, and let $ R = K[x_0, \dots, x_r] $ be the graded polynomial ring. Let $ {\mathbb P}^r = {\operatorname{Proj}}(R) $ be the projective space of dimension $ r $ over $ K.$ If $ X \subseteq {\mathbb P}^r $ is a closed scheme, we denote by $ {\mathcal I}_X $ its ideal sheaf in $ {\mathcal O}_{{\mathbb P}^r} $ and by $ I_X $ its saturated ideal in $ R,$ and it holds $ I_X = H^0_*({\mathbb P}^r, {\mathcal I}_X).$ By $R$–module we mean “graded $ R$–module". By sheaf we mean “coherent $ {\mathcal O}_{{\mathbb P}^r}$–module". If $ F $ is a $ R$–module we denote by $ \mathcal F $ the corresponding sheaf, namely $ \mathcal F := \tilde F.$ We recall that a local ring $ A $ is Cohen–Macaulay if $ \dim(A) = {\operatorname{depth}}(A).$ A ring $ A $ is Cohen–Macaulay if $ A_{\mathfrak{M}} $ is Cohen–Macaulay for every maximal ideal $ \mathfrak{M} \subset A.$ A scheme $ X $ is Cohen–Macaulay if the ring $ {\mathcal O}_{X,x} $ is Cohen–Macaulay for every closed point $ x \in X.$ A closed scheme $ X \subseteq {\mathbb P}^r$ is arithmetically Cohen–Macaulay (ACM, for brief) if the coordinate ring $ R_X = R/I_X $ is a Cohen–Macaulay ring. This is equivalent to say that $H^i_*({\mathcal I}_X) = 0$ for $ 1 \le i \le \dim(X).$ For any finitely generated $ R$–module $ P $ we denote by $ {\operatorname{pd}}(P) $ the projective dimension of $ P,$ that is to say, the length of the minimal free resolution of $ P $ ([@ei], Theorem 19.1 and the previous Definition). Let $ D \subseteq {\mathbb P}^r $ be a closed scheme, and let $ I_D \subseteq R $ be its saturated ideal. If $$0 \to F_t \to \dots \to F_2 \to F_1 \to I_D \to 0$$ is the minimal free resolution of $ I_D,$ with $ t \leq r,$ and $ P $ is the kernel of $ F_1 \to I_D,$ then we have a short exact sequence $$0 \to P \to F_1 \to I_D \to 0$$ which is equivalent to the minimal free resolution. The $ R$–module $ P $ is a torsion–free finitely generated $ R$–module with projective dimension $ {\operatorname{pd}}(P) = {\operatorname{pd}}(I_D)-1.$ We can also consider the short exact sequence $$0 \to \mathcal P \to \mathcal F_1 \to {\mathcal I}_D \to 0$$ obtained by considering the sheaves associated to the modules in the former sequence. Of course, $ \mathcal P $ is a torsion–free sheaf, and $ \mathcal F_1 $ is dissocié, according to the following definitions. A $ R$–module $ M $ is torsion–free if every non–zero element of $ R $ is a non zero–divisor of $M.$ A sheaf $ \mathcal F $ on $ {\mathbb P}^r $ is torsion–free if $ \mathcal F(U) $ is a torsion–free $ {\mathcal O}_{{\mathbb P}^r}(U)$–module for every open subset $ U \subseteq {\mathbb P}^r.$ This is equivalent to say that $\mathcal F_x$ is torsion free over ${\mathcal O}_{{\mathbb P}^r ,x}$ for every $x \in {\mathbb P}^r$. Let $ \mathcal F $ be a sheaf on $ {\mathbb P}^r.$ We say that $ \mathcal F $ is dissocié of rank $ s $ if $$\mathcal F = \oplus_{i=1}^s {\mathcal O}_{{\mathbb P}^r}(a_i)$$ for suitable integers $ a_1, \dots, a_s.$ Of course, if $ F $ is a free $ R$–module, then $ \mathcal F = \tilde F $ is dissocié. Conversely, if $ \mathcal F $ is dissocié, then $ H^0_*(\mathcal F) $ is a free $ R$–module. Generalizations of the approach consist in relaxing the strong hypothesis “dissocié" on $ \mathcal F_1.$ Hence, let us consider the short exact sequence $$\label{pnd} 0 \to \mathcal P \stackrel{\gamma}{\longrightarrow} \mathcal N \to {\mathcal I}_D(k) \to 0$$ with $ \mathcal P $ torsion–free, and $ k \in {\mathbb Z}.$ Standard arguments allow us to prove that $ \mathcal N $ is torsion–free, as well. So, the weakest hypothesis on $ \mathcal N $ is torsion–free. On the other hand, short exact sequences are classified by $ {\operatorname{Ext}}^1_R({\mathcal I}_D, \mathcal P).$ As we are interested in sequences of sheaves, it will help to have the analogue for sheaves of the minimal free resolution and of projective dimension of a graded finitely generated module. By ([@hart], Ch. II, Corollary 5.18), we have that any sheaf $ \mathcal P $ admits a [*dissocié resolution*]{}, namely a resolution by dissocié sheaves. We need to be more precise on this point, and so we begin with some preliminaries. \[ass\] We recall some facts about associated points. For more details see e.g ([@matsumura], Ch. 3), where the case of (ungraded) modules is dealt with. Extending to sheaves is straightforward. \(i) Let $\mathcal F$ be a sheaf. A (not necessarily closed) point $y \in {\mathbb P}^r$ is associated to $\mathcal F$ if there is an open affine $U = {\operatorname{Spec}}(A) \subseteq {\mathbb P}^r$ containing $y$ such that the prime ideal of $A$ corresponding to $y$ is associated to the $A$-module $\Gamma (U,\mathcal F)$; this is equivalent to say that ${\operatorname{depth}}_{{\mathcal O}_{{\mathbb P}^r},_y} (\mathcal F_y) = 0$. \(ii) The set $Ass(\mathcal F)$ of the associated points to $\mathcal F$ is finite. \(iii) Any form $f$ of degree $n$ avoiding all elements of $Ass(\mathcal F)$ induces by multiplication an injective morphism $ \mathcal F \stackrel{\cdot f}{\longrightarrow} \mathcal F(n)$. Hence a general form of degree $n$ has this property. \(iv) ([@bs], Exercise 20.4.21) The graded $R$- module $H^0_*(\mathcal F) $ is finitely generated if and only if ${\operatorname{Ass}}(\mathcal F)$ contains no closed points, if and only if ${\operatorname{depth}}_{{\mathcal O}_{{\mathbb P}^r},_x} (\mathcal F_x) > 0$ for every (closed) $x \in {\mathbb P}^r$. Now, we can prove that every sheaf has a dissocié resolution of finite length. \[diss res\] Let $\mathcal F$ be a sheaf and let $M$ be a graded submodule of $H^0_*(\mathcal F)$. Then 1. any general linear form induces by multiplication an injective map $M \to M(1)$; 2. if M is finitely generated then $pd(M) \le r$; 3. $\mathcal F$ admits a dissocié resolution of length $\le r$. \(1) follows easily from Remark \[ass\](iii). \(2) By (1) we have ${\operatorname{depth}}(M) \ge 1$ and the conclusion follows by the Auslander-Buchsbaum formula ([@ei], Exercise 19.8). \(3) Since $\mathcal F$ is coherent we have $\mathcal F= \tilde M$, where M is a suitable finitely generated graded submodule of $H^0_*(\mathcal F)$ ([@hart], Ch. II, proof of Theorem 5.19). The conclusion follows from (2), because we get a dissocié resolution of $ \mathcal F $ by sheafifying the minimal free resolution of $ M.$ Following ([@gr], Section 2), we define the minimal dissocié resolution of a coherent sheaf. \[min res\] Let $ \mathcal P $ be a sheaf such that $ P : = H^0_*(\mathcal P)$ is finitely generated. Let $$0 \to H_d \to \dots \to H_0 \to P \to 0$$ be the minimal free resolution of the $ R$–module P . We’ ll call [*minimal dissocié resolution*]{} of $ \mathcal P $ the exact sequence $$0 \to \mathcal{H}_d \to \dots \to \mathcal{H}_0 \to \mathcal P \to 0$$ obtained by sheafifying the minimal free resolution of $ P.$ (Recall that $\tilde P = \mathcal P$ by ([@hart], Ch. II, Proposition 5.4)). Moreover, we define the [*projective dimension*]{} of $ \mathcal P $ as $ {\operatorname{pd}}(\mathcal P) := {\operatorname{pd}}(P).$ It is known that there exist many submodules of $ P = H^0_*(\mathcal P) $ whose associated sheaf is $ \mathcal P:$ in fact, it is enough that such a submodule $ M $ agrees with $ P $ for some large degree on. Of course, the sheafification of a minimal free resolution of $ M $ is still a dissocié resolution of $ \mathcal P,$ and no map is split. However, the resolution of $ M $ is longer than the minimal one. In fact, from the short exact sequence of modules $$0 \to M \to P \to P/M \to 0,$$ we get that $ {\operatorname{pd}}(M) = r,$ because $ P/M $ is an Artinian module. Hence, $ {\operatorname{pd}}(M) \geq {\operatorname{pd}}(P),$ as we claimed. \[rem-pd\] (i) Clearly ${\operatorname{pd}}( \mathcal P) = 0$ if and only if $ \mathcal P $ dissocié. \(ii) ${\operatorname{pd}}(\mathcal P) \le r$ whenever defined (Lemma \[diss res\](2) applied with $M= H^0_*(\mathcal P)$). The next Lemma gives a bound for the projective dimension of a torsion-free sheaf. \[initial lemma\] Let $ \mathcal P $ be a torsion-free coherent sheaf on $ {\mathbb P}^r,$ and let $ P= H^0_*(\mathcal P).$ Then: 1. $ P $ is finitely generated; 2. $P$ torsion free; 3. $ \mathcal P $ is a subsheaf of a coherent dissocié sheaf; 4. ${\operatorname{pd}}(P) = {\operatorname{pd}}(\mathcal P) \leq r-1$. \(1) It follows easily from Remark \[ass\](iv). \(2) Whatever non zero form $f \in R$ of degree $n$ induces, by multiplication, an injective morphism $ \mathcal P \stackrel{\cdot f}{\longrightarrow} \mathcal P (n), $ and consequently an injective homomorphism $ P \stackrel{\cdot f}{\longrightarrow} P(n).$ \(3) By (1) and (2), $ P $ is a torsion–free $R$-module and hence it is a graded submodule of a free $R$–module $L$. Then, $ \mathcal P = \tilde P $ is a subsheaf of $\mathcal L:= \tilde L $ and the claim follows. \(4) By (3) there are a sheaf $\mathcal F$ and an exact sequence $0 \to \mathcal P \to \mathcal L \to \mathcal F \to 0$, whence an exact sequence of $R$-modules: $$0\to P \to L \to M \to 0,$$ where $M$ is a graded submodule of $H^0_*(\mathcal F)$. By Lemma \[diss res\](2) we have ${\operatorname{pd}}(M) \le r$, whence ${\operatorname{pd}}(P) \le r-1$. Now by (1), Definition \[min res\] applies and the proof is complete From now on, every $ R$–module will be finitely generated, and so we shall skip this assumption. It is possible to describe the cohomology of a coherent sheaf, as we said before. \[lemma1\] Let $ r \geq 3,$ let $ P $ be a $ R$–module and let $ \mathcal P = \tilde P $ be its associated sheaf. Suppose $ d = {\operatorname{pd}}(P) < r. $ Then: 1. $ H^0_*(\mathcal P)=P;$ 2. $ H^i_*(\mathcal P)=0 $ for $ 1 \leq i \leq r-d-1;$ 3. $ H^{r-d}_*(\mathcal P) \neq 0. $ 4. If $ \mathcal P$ is any torsion-free sheaf with $d:= {\operatorname{pd}}(\mathcal P),$ then (2) and (3) hold. We prove claims (1) (2) (3) together, by induction on $ d.$ If $ d=0,$ the sheaf $ \mathcal P $ is dissocié and the claims hold by ([@hart], Ch. III, Theorem 5.1). If $d=1$ we have a non–split exact sequence $$0\to L_1 \to L_0 \to P \to 0$$ with $ L_1 $ and $ L_0 $ free. By passing to sheaves, we get a non–split exact sequence $$\label{d=1}0\to \mathcal L_1 \to \mathcal L_0 \to \mathcal P \to 0,$$ whence $ {\operatorname{Ext}}^1( \mathcal P, \mathcal L_1) \neq 0$. It follows easily that $ {\operatorname{Ext}}^1(\mathcal P, {\mathcal O}_{{\mathbb P}^r}(k)) \neq 0 $ for some $ k \in {\mathbb Z}$. On the other hand by duality and properties of Ext we get $ H^{r-1}({\mathbb P}^r, \mathcal P(-k-r-1)) \cong {\operatorname{Ext}}^1( \mathcal P(-k-r-1), \omega_{{\mathbb P}^r}) \cong {\operatorname{Ext}}^1(\mathcal P , {\mathcal O}_{{\mathbb P}^r} (k)),$ whence (3). Since (1) and (2) are immediate from the exact sequence (\[d=1\]), the statement holds for $d=1$ as well. Assume now $d\ge 2$. We have an exact sequence $$0 \to P_1 \to G \to P \to 0$$ where $ G $ is a free $R-$module and $ P_1 $ is a $ R$–module with $ {\operatorname{pd}}(P_1)=d-1.$ In fact, it is enough to consider the first short exact sequence that can be obtained from the minimal free resolution of $ P,$ as explained before. By taking the sheaves associated to each item, we get the short exact sequence of sheaves $$\label{p1} 0 \to \mathcal P_1 \to \mathcal G \to \mathcal P \to 0.$$ By induction, we may assume that $ H^0_*(\mathcal P_1)=P_1,$ $ H^i_*(\mathcal P_1)=0 $ for $ 1 \leq i \leq r-(d-1)-1=r-d,$ and that $ H^{r-d+1}_*(\mathcal P_1) \not= 0.$ By assumption, $ d < r,$ and so $ r-d \geq 1.$ In particular, $ H^1_*(\mathcal P_1)=0. $ By taking the cohomology sequence associated to (\[p1\]) and using the assumptions on the cohomology of $ \mathcal P_1 $ we get the conclusion. To prove (4) set $P:= H^0_*(\mathcal P)$. Then by definition and by Lemma \[initial lemma\] we have $d = {\operatorname{pd}}(P)<r$. Since $\tilde P= \mathcal P$ the conclusion follows by (2) and (3). The previous Lemma allows us to generalize Horrocks’ splitting criterion ([@okonek], Theorem 2.3.1) to torsion–free sheaves, with a completely different proof. A torsion–free sheaf $ \mathcal P $ over $ {\mathbb P}^r $ is dissocié precisely when $ H^i_*(\mathcal P) = 0 $ for $ i = 1, \dots, r-1.$ Assume $ H^i_*(\mathcal P) = 0$ for $ i = 1, \dots, r-1,$ and set $d:= {\operatorname{pd}}(\mathcal P).$ By Lemma \[lemma1\] (4) we have $ H^i_*(\mathcal P) = 0 $ for $ i = 1, \dots, r-1-d $ and $ H^{r-d}_*(\mathcal P) \neq 0$. This is possible only if $d=0$, i.e. if $\mathcal P$ is dissocié. The converse is clear. Now, we consider the short exact sequence (\[pnd\]). Our first result relates the codimension of $ D $ and the projective dimension of $ \mathcal P.$ Let $ D \subseteq {\mathbb P}^r $ be a closed scheme of codimension $ s,$ with $ s \geq 2,$ and let $ P $ a $ R$–module with $ {\operatorname{pd}}(P) = d.$ If $ s-2 > d,$ then $ {\operatorname{Ext}}^j_R({\mathcal I}_D(k), \mathcal{P}) = 0 $ for $ j = 1, \dots, s-2-d. $ We prove the claim by induction on $ d.$ If $ d = 0,$ that is to say $ P $ is a free module, then there exist $ a_1, \dots, a_n \in \mathbb{Z} $ such that $ P = \oplus_{i=1}^n R(-a_i).$ By using standard properties of $ {\operatorname{Ext}}$ groups, we have $$\begin{split} {\operatorname{Ext}}^j&_R({\mathcal I}_D(k), \mathcal{P}) = \oplus_{i=1}^n {\operatorname{Ext}}^j_R({\mathcal I}_D(k), {\mathcal O}_{{\mathbb P}^r}(-a_i)) = \\ & = \oplus_{i=1}^n {\operatorname{Ext}}^j_R({\mathcal I}_D(k+a_i-r-1), \omega_{{\mathbb P}^r}) \cong \\ & \cong \oplus_{i=1}^n H^{r-j}({\mathbb P}^r, {\mathcal I}_D(k+a_i-r-1)) = \\ & = \oplus_{i=1}^n H^{r-j-1}(D, {\mathcal O}_D(k+a_i-r-1)) = 0 \end{split}$$ as soon as $ r-j-1 > r-s $ by Grothendieck’s vanishing Theorem ([@hart], Ch.III, Theorem 2.7), where $ \omega_{{\mathbb P}^r} = {\mathcal O}_{{\mathbb P}^r}(-r-1) $ is the canonical sheaf of $ {\mathbb P}^r.$ Hence, $ {\operatorname{Ext}}^j({\mathcal I}_D(k), \mathcal{P}) = 0 $ for $ j = 1, \dots, s-2 $ and for every $ k \in \mathbb{Z},$ and the claim holds for $ d = 0.$ Assume $ d > 0 $ and the claim to hold for every $ R$–module with projective dimension $ d-1.$ As in the proof of Lemma \[lemma1\], we consider the short exact sequence $$0 \to P_1 \to G \to P \to 0$$ with $ G $ free and $ P_1 $ of projective dimension $ d-1.$ By applying $ {\operatorname{Hom}}({\mathcal I}_D(k), -) $ to the sheafified sequence, we get the exact sequence $${\operatorname{Ext}}^i({\mathcal I}_D(k), \mathcal{G}) \to {\operatorname{Ext}}^i({\mathcal I}_D(k), \mathcal{P}) \to {\operatorname{Ext}}^{i+1}({\mathcal I}_D(k), \mathcal{P}_1) \to {\operatorname{Ext}}^{i+1}({\mathcal I}_D(k), \mathcal{G}).$$ From the first part of the proof, we get that $ {\operatorname{Ext}}^i({\mathcal I}_D(k), \mathcal{G}) = {\operatorname{Ext}}^{i+1}({\mathcal I}_D(k), \mathcal{G}) = 0 $ for every $ k $ and for $ i = 1, \dots, s-3.$ From the induction assumption, $ {\operatorname{Ext}}^{i+1}({\mathcal I}_D(k), \mathcal{P}_1) = 0 $ for every $ k $ and for $ i = 0, \dots, s-2-d.$ Hence, $ {\operatorname{Ext}}^i({\mathcal I}_D(k), \mathcal{P}) = 0 $ for every $ k \in \mathbb{Z} $ and for $ i = 1, \dots, s-2-d $ as claimed. A direct consequence of the previous Proposition is that we can predict if $ N $ is the direct sum of $ P $ and $ I_D.$ In fact it holds: \[split\] Let $ D \subseteq {\mathbb P}^r $ be a closed scheme of codimension $ s \geq 2,$ and let $ P $ be a $ R$–module satisfying $ s-2 > {\operatorname{pd}}(P).$ Then, the only extension of $ {\mathcal I}_D(k) $ with $ \mathcal{P} $ is the trivial one, for every choice of $ k \in {\mathbb Z}.$ Consequently if there is a non–split exact sequence (\[pnd\]), we must have $s \leq {\operatorname{pd}}(P) + 2.$ The previous Proposition shows that $ {\operatorname{Ext}}^1({\mathcal I}_D(k), \mathcal{P}) = 0 $ and the claim follows. Now, we take into account the cohomology of $ D $ to get a bound on the projective dimension of $ \mathcal N.$ \[pd(N) geq pd(P)+2\] Let $ D \subset {\mathbb P}^r $ be a closed scheme, and let $ \mathcal{P}, \mathcal{N} $ be torsion–free sheaves such that the short sequence (\[pnd\]) is exact. If $ {\operatorname{pd}}(\mathcal N) \geq {\operatorname{pd}}(\mathcal P) + 2,$ then $ H^{r-{\operatorname{pd}}(\mathcal N)}_*({\mathcal I}_D) \not= 0. $ Conversely, if $ H^j_*({\mathcal I}_D) \not= 0 $ for some $ j \in \mathbb{Z} $ with $ 1 \leq j \leq r - 2 - {\operatorname{pd}}(\mathcal P),$ then $ {\operatorname{pd}}(\mathcal N) \geq {\operatorname{pd}}(\mathcal P) + 2.$ By Lemma \[initial lemma\], we have that $ {\operatorname{pd}}(\mathcal P) $ and $ {\operatorname{pd}}(\mathcal N) $ are strictly smaller than $ r.$ By taking the long exact cohomology sequence associated to (\[pnd\]), we get $$H^i_*(\mathcal{P}) \to H^i_*(\mathcal{N}) \to H^i_*({\mathcal I}_D) \to H^{i+1}_*(\mathcal{P}).$$ From Lemma \[lemma1\], we know that $ H^j_*(\mathcal{P}) = 0 $ for $ j = 1, \dots, r -{\operatorname{pd}}(\mathcal P)- 1,$ and so $ H^i_*(\mathcal{N}) \cong H^i_*({\mathcal I}_D) $ for $ i = 1, \dots, r - {\operatorname{pd}}(\mathcal P) - 2.$ If $ {\operatorname{pd}}(\mathcal P) + 2 \leq {\operatorname{pd}}(\mathcal N) < r,$ then $ 1 \leq r - {\operatorname{pd}}(\mathcal N) < r - {\operatorname{pd}}(\mathcal P) - 1.$ Hence, $ H^{r-{\operatorname{pd}}(\mathcal N)}_*(\mathcal{N}) \cong H^{r-{\operatorname{pd}}(\mathcal N)}_*({\mathcal I}_D) $ and we get the claim by Lemma \[lemma1\]. Assume now that $ H^j({\mathcal I}_D(k)) \not= 0 $ for some $ k \in \mathbb{Z} $ and some $ j $ such that $ 1 \leq j \leq r - {\operatorname{pd}}(\mathcal P) - 2.$ Hence, $ H^j_*(\mathcal{N}) \not= 0.$ Again by Lemma \[lemma1\], $ r - {\operatorname{pd}}(\mathcal N) \leq j $ and so $ {\operatorname{pd}}(\mathcal N) \geq {\operatorname{pd}}(\mathcal P) + 2.$ In the second part of the previous Proposition, the hypothesis on $ j $ implies that $ {\operatorname{pd}}(\mathcal P) \leq r-3.$ This last inequality is not automatically fulfilled. In fact, let $ D \subseteq {\mathbb P}^r $ be a locally Cohen–Macaulay curve with $ H^1_* {\mathcal I}_D \not= 0.$ Let $$0 \to G_r \to \dots \to G_2 \to G_1 \to I_D \to 0$$ be the minimal free resolution of $ I_D $ and let $ P = \ker(G_1 \to I_D).$ Then, $ {\operatorname{pd}}(P) = r-2 = {\operatorname{pd}}(I_D) - 1,$ and $ {\operatorname{pd}}(G_1) = 0.$ Hence, we cannot apply the previous Proposition to the short exact sequence $ 0 \to P \to G_1 \to I_D \to 0.$ Nevertheless, it could exist a different short exact sequence $ 0 \to Q \to N \to I_D \to 0 $ with $ {\operatorname{pd}}(Q) = r-3.$ In this case, $ {\operatorname{pd}}(N) = r-1.$ Notice that $ r-3 $ is the smallest projective dimension allowed for the first item of the sequence, because of the codimension of $ D.$ The case considered in the previous Proposition, namely $ {\operatorname{pd}}(\mathcal N) \geq {\operatorname{pd}}(\mathcal P) + 2,$ occurs in the $ \mathcal N$–type resolution of the ideal sheaf of a locally Cohen–Macaulay curve in $ {\mathbb P}^3 $ ([@mdp], Ch. II, Section 4). In that case, $ \mathcal P $ is dissocié and $ {\operatorname{pd}}(\mathcal N) = 2,$ where $ N $ is the second syzygy module of the Hartshorne–Rao module (graded Artinian $ R$–module) of the curve, up to a free summand. Now, we stress some consequences of the previous Proposition that we’ ll use in next sections. \[acm and pd\] Consider an exact sequence [(\[pnd\])]{} where $D$ has codimension $s \ge 2$. If $D$ is ACM and the sequence is non-split we have $$\label {pd(N)} {\operatorname{pd}}(\mathcal N) \leq {\operatorname{pd}}(\mathcal P) +1.$$ By Corollary \[split\] the non–splitting of the sequence (\[pnd\]) implies that $ s \leq {\operatorname{pd}}(\mathcal P) + 2.$ If $ {\operatorname{pd}}(\mathcal N) \geq {\operatorname{pd}}(\mathcal P) + 2,$ then $ H^{r-{\operatorname{pd}}(\mathcal N)}_*({\mathcal I}_D) \not= 0 $ by Proposition \[pd(N) geq pd(P)+2\]. On the other hand, $ r-{\operatorname{pd}}(\mathcal N) \leq r-s $ and so $ H^{r-{\operatorname{pd}}(\mathcal N)}_*({\mathcal I}_D) = 0 $ because $ D $ is ACM. The contradiction proves that $ {\operatorname{pd}}(\mathcal N) \leq {\operatorname{pd}}(\mathcal P) + 1.$ If $ {\operatorname{pd}}(\mathcal N) \leq {\operatorname{pd}}(\mathcal P) + 1,$ we can only prove that $ H^i_*({\mathcal I}_D) = 0 $ for $ i = 1, \dots, r-{\operatorname{pd}}(\mathcal P)-2.$ Hence, $ D $ could not be an ACM scheme if $ s < {\operatorname{pd}}(\mathcal P) + 2.$ A further problem related to the sequence (\[pnd\]) is the following: given the modules $ P $ and $ N,$ and an injective map $ \mathcal P \to \mathcal N,$ when is the cokernel an ideal sheaf? This problem was considered in [@mdp], and we resume their results. At first, we recall the definition and some properties of the maximal subsheaves, generalizing to $ {\mathbb P}^r $ the one given for sheaves on $ {\mathbb P}^3 $ ([@mdp], Ch. IV, Définition 1.1). Let $ \mathcal{M} \subset \mathcal{N} $ be $ {\mathcal O}_{{\mathbb P}^r}$–modules. $ \mathcal{M} $ is a maximal subsheaf of $ \mathcal N $ if for all subsheaves $ \mathcal{M'} \subset \mathcal{N} $ with $ {\operatorname{rank}}(\mathcal M) = {\operatorname{rank}}(\mathcal{M'}) $ such that $ \mathcal{M} \subseteq \mathcal{M'} \subseteq \mathcal{N},$ we have $ \mathcal{M} = \mathcal{M'}.$ The interest in such subsheaves lies in the following properties. \[maximal-sub-sheaf\] Let $ \mathcal{M} \subseteq \mathcal{N} $ be $ {\mathcal O}_{{\mathbb P}^r}$–modules. Consider the following properties: 1. $ \mathcal{M} $ is maximal; 2. $ \mathcal{N}/\mathcal{M} $ is torsion–free; 3. $ \mathcal{N}/\mathcal{M} $ is torsion–free in codimension $ 1;$ 4. $ \mathcal{N}/\mathcal{M} $ is locally free in codimension $ 1;$ 5. $ \mathcal{N}/\mathcal{M} $ has constant rank in codimension $ 1;$ 6. $ \mathcal{N}/\mathcal{M} $ is locally a direct summand of $ \mathcal{N} $ in codimension $ 1.$ Then, $ (1) \Leftrightarrow (2) \Rightarrow (3) \Leftrightarrow (4) \Leftrightarrow (5) \Rightarrow (6).$ Furthermore, if $ \mathcal{N} $ is torsion–free and $ \mathcal{M} $ is locally free, they all are equivalent. The statement was proved for sheaves on $ {\mathbb P}^3 $ in ([@mdp], Ch. IV, Proposition 1.2),but the proof works without changes also for sheaves on $ {\mathbb P}^r.$ Moreover, in the proof, the authors proved also the existence of maximal dissocié subsheaves of a sheaf $ \mathcal{N}.$ As explained in ([@mdp], Ch.IV, Remark 1.3(c)), in $ {\mathbb P}^3,$ if $ \mathcal{N} $ is a rank $ n+1 $ vector bundle, and $ \mathcal{M} $ is a rank $ n $ dissocié maximal subsheaf of $ \mathcal{N},$ then $ \mathcal{N}/\mathcal{M} $ is a rank $ 1 $ torsion–free sheaf, and so it is an ideal sheaf tensorized times $ \det(\mathcal{N}) \otimes \det(\mathcal{M}^{-1}).$ Moreover, if $ \mathcal N $ is not dissocié, then the ideal sheaf defines a curve. A construction of ACM schemes ============================= In this section, we consider two coherent torsion–free sheaves $ \mathcal{P} $ and $ \mathcal{N} $ and an injective map $ \gamma: \mathcal{P} \to \mathcal{N},$ and we study the scheme $ D $ whose ideal sheaf is isomorphic to $ {\operatorname{coker}}(\gamma),$ as in [@mdp]. We limit ourselves to consider only the case $ D $ has the largest codimension to have a non–split exact sequence (\[pnd\]) (see Corollary \[split\]) and $ \mathcal N $ to have the largest projective dimension to allow $ D $ to be an ACM scheme (see Corollary \[acm and pd\]). In more detail, we collect the hypotheses on $ \mathcal{P} $ and $ \mathcal{N} $ in the following $$\label{H-hyp} \tag{H} \begin{split} & (H.1) \ \mathcal{P} \mbox{ is torsion-free and } s:= {\operatorname{pd}}(\mathcal P) + 2 \le r;\\ & (H.2) \ \mathcal{N} \mbox{ is torsion-free and } {\operatorname{pd}}(\mathcal N) \leq {\operatorname{pd}}(\mathcal P) +1 ; \\ & (H.3) \ \mbox{the polynomial } \\ & \qquad \quad p(t) :=-\chi(\mathcal N(t-k)) + \chi(\mathcal P(t-k)) + \binom{t+r}r \\ & \qquad \mbox{has degree } r-s \mbox{ for some } k \in {\mathbb Z}. \end{split}$$ We remark that, in view of Definition \[min res\] and Remark \[rem-pd\], the condition about the projective dimensions required in (H.2) and (H.3) means that $ P := H^0_*(\mathcal P) $ and $ N := H^0_*(\mathcal N) $ have, respectively, minimal free resolutions $$\ 0 \to G_{s-1} \stackrel{\Delta_{s-1}}{\longrightarrow} G_{s-2} \stackrel{\Delta_{s-2}}{\longrightarrow} \dots \stackrel{\Delta_{2}}{\longrightarrow} G_1 \to P \to 0$$ and $$\ 0 \to F_{s} \stackrel{\delta_{s}}{\longrightarrow} F_{s-1} \stackrel{\delta_{s-1}}{\longrightarrow} \dots \stackrel{\delta_{2}}{\longrightarrow} F_1 \to N \to 0.$$ \[rem(H)\] (i) We allow $ F_j = 0 $ for some $ j $ in the minimal free resolution of $ N.$ In such a case, $ F_{j+h} = 0 $ for every $ h \geq 0.$ \(ii) Condition (H.3) implies that $ {\operatorname{rank}}(\mathcal N) = {\operatorname{rank}}(\mathcal P) + 1,$ because the rank of $ \mathcal F $ is equal to $ r! $ times the coefficient of $ t^r $ in $ \chi(\mathcal F(t)). $ Moreover, recalling that $ {\mathcal O}_{{\mathbb P}^r}(a) $ has degree $ a $ and that the degree is additive on exact sequences, we have that $ k = \deg(\mathcal{N}) - \deg(\mathcal{P}).$ Now, we describe the geometric properties of the schemes that can be obtained from such torsion–free sheaves. \[NtoD\] Let $ \mathcal P $ and $ \mathcal N $ be torsion–free coherent sheaves that fulfil the hypotheses (\[H-hyp\]). Assume that there exists an injective map $ \gamma: \mathcal{P} \to \mathcal{N} $ whose image is a maximal subsheaf of $ \mathcal{N}$. Then there exists a codimension $ s = 2 + {\operatorname{pd}}(\mathcal P) $ scheme $ D,$ closed and ACM, whose ideal sheaf fits into the short exact sequence (\[pnd\]) with $ k = \deg(\mathcal N)-\deg(\mathcal P)$. Moreover, the sequence $0 \to P \to N \to I_D(k) \to 0$ is exact. The cokernel of $ \gamma $ is a rank $ 1 $ torsion–free sheaf $ \mathcal{F}.$ Let $ \mathcal{F}^{\vee\vee} $ be its double dual. Since $ \mathcal{F} $ is torsion–free, the natural map $ \mathcal{F} \to \mathcal{F}^{\vee\vee} $ is injective. By ([@hart-2], Corollary 1.2 and Proposition 1.9), $ \mathcal{F}^{\vee\vee} \cong {\mathcal O}_{{\mathbb P}^r}(h) $ for some $ h \in {\mathbb Z},$ and so $ \mathcal{F} \cong {\mathcal I}_D(h) \subseteq {\mathcal O}_{{\mathbb P}^r}(h),$ i.e. we have an exact sequence $$0 \to \mathcal{P} \stackrel{\gamma}{\longrightarrow} \mathcal{N} \to {\mathcal I}_{D}(h) \to 0.$$ Clearly $ h = \deg(\mathcal N)-\deg(\mathcal P) = k $, and hence the above sequence coincides with (\[pnd\]). Now, by Remark \[rem(H)\], (ii), $k$ is the integer occurring in the polynomial p(t) of (H.3), then $ p(t) $ is the Hilbert polynomial of $D$, whence $\dim(D) = r-s $ by (H.3). Moreover, (H.2) and the second part of Proposition \[pd(N) geq pd(P)+2\] imply that $D$ is ACM. Finally, by (H1) we have ${\operatorname{pd}}(\mathcal P) \le r-2$, whence $r- {\operatorname{pd}}(\mathcal P)- 1 \ge 1$. Then Lemma \[lemma1\](4) implies that $H^1_*(\mathcal P)= 0$ and the last statement follows. \[rem3.3\] The map $ \gamma: \mathcal{P} \to \mathcal{N} $ induces a map of complexes between the minimal free resolutions of $ P $ and $ N.$ Let $ \gamma_i : G_i \to F_i $ be the induced map. Of course, $ \gamma_i \circ \Delta_{i+1} = \delta_{i+1} \circ \gamma_{i+1},$ for each $ i \geq 1.$ Hence, a resolution of $ I_D(k) $ can be obtained via mapping cone from (\[PNI\]), and it is $$\label{non-min-res-D} 0 \to \begin{array}{c} G_{s-1} \\ \oplus \\ F_s \end{array} \stackrel{\varepsilon_s}{\longrightarrow} \begin{array}{c} G_{s-2} \\ \oplus \\ F_{s-1} \end{array} \stackrel{\varepsilon_{s-1}}{\longrightarrow} \dots \stackrel{\varepsilon_3}{\longrightarrow} \begin{array}{c} G_1 \\ \oplus \\ F_2 \end{array} \stackrel{\varepsilon_2}{\longrightarrow} F_1 \to I_{D}(k) \to 0$$ where $ \varepsilon_i: G_{i-1} \oplus F_i \to G_{i-2} \oplus F_{i-1} $ is given by $$\left( \begin{array}{cc} \Delta_{i-1} & 0 \\ (-1)^i \gamma_{i-1} & \delta_i \end{array} \right), \mbox{ for } i \geq 2.$$ We remark that $\varepsilon_2 : G_{1} \oplus F_2 \to F_1 $ is represented by the matrix $ ( \gamma_1, \delta_2).$ For general results on free resolutions, it is clear that the minimal free resolution of $ I_D(k) $ can be obtained by cancelling the free modules corresponding to constant non–zero entries of any matrix representing the map $ \varepsilon_i, i = 2, \dots, s.$ If there exists an injective map $ \gamma: \mathcal{P} \to \mathcal{N} $ whose image is a maximal subsheaf of $ \mathcal{N} $ of rank $ {\operatorname{rank}}(\mathcal{P}) = {\operatorname{rank}}(\mathcal{N}) - 1,$ then the general map in $ {\operatorname{Hom}}(\mathcal{P}, \mathcal{N}) $ has the same property. Once we have constructed a closed ACM scheme $ D $ of codimension $ s $ as cokernel of a short exact sequence (\[pnd\]), we can construct the minimal free resolution of $ I_D $ and it is $$0 \to H_s \stackrel{\sigma_s}{\longrightarrow} \dots \stackrel{\sigma_2}{\longrightarrow} H_1 \stackrel{\sigma_1}{\longrightarrow} I_D \to 0$$ where $ H_i = \oplus_{n \in {\mathbb Z}} R(-n)^{h_i(n)}.$ Let $ K = \ker(\sigma_1).$ Then, the ideal sheaf $ {\mathcal I}_D $ is also the cokernel of the short exact sequence $$\label {K to H} 0 \to \mathcal{K} \stackrel{j}{\longrightarrow} \mathcal{H}_1 \stackrel{\sigma_1}{\longrightarrow} {\mathcal I}_D \to 0.$$ Now we compare the two sequences (\[pnd\]) and (\[K to H\]). \[DtoN\] Let $D \subseteq {\mathbb P}^r$ be an ACM scheme of codimension $s$ and let [(\[K to H\])]{} be as above. \(i) If there is a sequence [(\[pnd\])]{} with ${\operatorname{pd}}(\mathcal P) = s-2$ then there exists a map $ \psi: \mathcal K \to \mathcal P $ such that $ \mathcal N $ is the push–out of $ \mathcal P $ and $ \mathcal H_1.$ \(ii) Conversely, let $ \mathcal P $ be a torsion–free coherent sheaf with $ {\operatorname{pd}}(\mathcal P) = s-2.$ Then, for every map $ \psi: \mathcal K \to \mathcal P $ there exists a short exact sequence (\[pnd\]) whose third item is $ {\mathcal I}_D.$ \(i) Up to twisting the sequence (\[pnd\]), we can assume that $ k = 0.$ The minimal free resolution of $ I_D $ is $$0 \to H_s \stackrel{\sigma_s}{\longrightarrow} \dots \stackrel{\sigma_2}{\longrightarrow} H_1 \stackrel{\sigma_1}{\longrightarrow} I_D \to 0,$$ and so $ \sigma_1 $ maps the canonical bases of $ H_1 $ onto a minimal set of generators of $ I_D.$ The surjective map $ \mathcal N \to {\mathcal I}_D $ induces a surjective map $ N = H^0_*(\mathcal N) \to I_D $ because $ {\operatorname{pd}}(\mathcal P) = s-2 $ implies that $ H^1_*(\mathcal P) = 0 $ (see Lemma \[lemma1\]). Hence, we have a well defines map $ H_1 \to N $ given on the canonical bases of $ H_1 $ and extended by linearity. So, there exists a map $ \varphi: \mathcal{H}_1 \to \mathcal N.$ It is straightforward to check that $ \varphi $ maps the kernel of $ \sigma_1 $ to the image of $ \mathcal P,$ and so $ \varphi $ induces a map $ \psi: \mathcal K \to \mathcal{P}.$ At the end, there exists a commutative diagram $$\commdiag{0 \to & \mathcal{K} & \mapright\lft{j} & \mathcal{H}_1 & \mapright & {\mathcal I}_D & \to 0 \cr & \mapdown\lft{\psi} & & \mapdown\lft{\varphi} & & \bivline & \cr 0 \to & \mathcal{P} & \mapright & \mathcal{N} & \mapright & {\mathcal I}_D & \to 0 }$$ where the last map is the identity of $ {\mathcal I}_D.$ From the universal property of the push–out (see [@north], Ch. 3, Theorem 11 for the definition and the properties of the push–out), it follows that $ \mathcal N $ is the push–out of $ \mathcal H_1 $ and $ \mathcal P $ as claimed. \(ii) As soon as we fix a map $ \psi: \mathcal K \to \mathcal P,$ we can construct the same commutative diagram we considered in the first part of the proof. In more detail, let $ q: \mathcal{K} \to \mathcal{H}_1 \oplus \mathcal{P} $ be defined as $ j $ on the first summand and as $ -\psi $ on the second one. Then, $ \mathcal{N} = \mathcal{H}_1 \oplus \mathcal{P} / \mbox{im}(q).$ The sheaf $ \mathcal{N} $ is torsion–free of rank $ {\operatorname{rank}}(\mathcal{P}) + 1.$ The second row of commutative diagram above gives the short exact sequence $ 0 \to P \to N \to I_D \to 0 $ because $ H^1_*(\mathcal P) = 0 $ by Lemma \[lemma1\]. Hence, $ {\operatorname{pd}}(\mathcal N) \leq {\operatorname{pd}}({\mathcal I}_D) = {\operatorname{pd}}(\mathcal P) + 1,$ and the proof is complete. If $ \psi = 0,$ then $ \mathcal N = \mathcal P \oplus {\mathcal I}_D,$ and the sequence is not interesting. On the other hand, if $ \psi $ is an isomorphism, then $ \mathcal N \cong \mathcal H_1 $ and once again we get nothing new. Summarizing the above discussed results, we have that if we start from two sheaves $ \mathcal N $ and $ \mathcal P $ satisfying our hypotheses, we can construct codimension $ s $ ACM schemes, and conversely, given a codimension $ s $ ACM scheme $ D $ and a torsion–free sheaf $ \mathcal P,$ we can construct a sheaf $ \mathcal{N} $ fulfilling the conditions we ask. Starting from two given torsion–free sheaves $ \mathcal N $ and $ \mathcal P,$ there are constrains on the ACM schemes we can obtain. \[constrains\] In the same hypotheses as Theorem \[NtoD\], let $ D \subset {\mathbb P}^r $ be a codimension $ s $ ACM closed scheme whose ideal sheaf fits into a short exact sequence $$0 \to \mathcal{P} \to \mathcal{N} \to {\mathcal I}_D(k) \to 0$$ for some $ k \in {\mathbb Z}.$ Then, the minimal number of generators of $ I_D $ is not larger than $ {\operatorname{rank}}(\mathcal{F}_1) $ while the free modules $ H_i $ that appear in the minimal free resolution of $ I_D(k) $ are direct summands of $ F_i \oplus G_{i-1}.$ We constructed a free resolution of $ I_D(k) $ in Remark \[rem3.3\]. The minimal free resolution of $ I_D $ can be obtained from this last one by cancelling suitable summands. As a consequence of the hypotheses (\[H-hyp\]), to construct ACM schemes of codimension $ s \geq 3,$ we have to consider a torsion–free sheaf $ \mathcal P $ satisfying $ {\operatorname{pd}}(\mathcal P) > 0,$ that is to say, $ \mathcal P $ non–dissocié. On the other hand, if the codimension of $ D $ is $ 2,$ then $ \mathcal P $ is dissocié. In this case, we have a more geometric interpretation of the construction, and it can be compared with Serre’s construction (Hartshorne’s one, respectively) when $ \mathcal N $ is a rank $ 2 $ vector bundle (reflexive sheaf, respectively). \[cod2\] Let $ D \subset {\mathbb P}^r $ be a codimension $ 2 $ ACM closed scheme, and let $ c $ be an integer such that $ H^0(D, \omega_D(c)) \not= 0.$ Then, for every non–zero $ \xi \in H^0(D, \omega_D(c)) $ we can construct a short non-split exact sequence $$0 \to {\mathcal O}_{{\mathbb P}^r}(c-r-1) \to \mathcal{N} \to {\mathcal I}_D \to 0$$ with $ \mathcal{N} $ torsion–free, of rank $ 2 $ and $ {\operatorname{pd}}(\mathcal N) \leq 1.$ By Serre’s duality for $ {\mathbb P}^r $ ([@hart], Ch. III, Theorem 7.1), we get $ {\operatorname{Ext}}^1({\mathcal I}_D, {\mathcal O}_{{\mathbb P}^r}(c-r-1)) \cong H^{r-1}({\mathbb P}^r, {\mathcal I}_D(-c))'.$ From the inclusion $ D \hookrightarrow {\mathbb P}^r,$ we get $ H^{r-1}({\mathbb P}^r, {\mathcal I}_D(-c))' \cong H^{r-2}(D, {\mathcal O}_D(-c))',$ and again by Serre’s duality on $ D $ ([@hart], Ch. III, Theorem 7.6 and Proposition 6.3(c)), we have the further isomorphisms $ {\operatorname{Ext}}^1({\mathcal I}_D, {\mathcal O}_{{\mathbb P}^r}(c-r-1)) \cong {\operatorname{Hom}}({\mathcal O}_D(-c), \omega_D) \cong H^0(D, \omega_D(c)).$ Hence, every non–zero $ \xi \in H^0(D, \omega_D(c)) $ can be thought to as an extension of $ {\mathcal I}_D $ with $ {\mathcal O}_{{\mathbb P}^r}(c-r-1) $ and so as a non–split short exact sequence $$0 \to {\mathcal O}_{{\mathbb P}^r}(c-r-1) \to \mathcal{N} \to {\mathcal I}_D \to 0.$$ The sheaf $ \mathcal{N} $ has rank $ 2,$ and it is torsion–free. Moreover, if $$0 \to \mathcal{H}_2 \stackrel{\varphi}{\longrightarrow} \mathcal{H}_1 \to {\mathcal I}_D \to 0$$ is the minimal dissocié resolution of $ {\mathcal I}_D,$ there is a natural surjection $ {\operatorname{Hom}}(\mathcal{H}_2, {\mathcal O}_{{\mathbb P}^r}(c-r-1)) \to {\operatorname{Ext}}^1({\mathcal I}_D, {\mathcal O}_{{\mathbb P}^r}(c-r-1)),$ and so there exists a map $ \psi: \mathcal{H}_2 \to {\mathcal O}_{{\mathbb P}^r}(c-r-1) $ that does not factor through $ \varphi: \mathcal{H}_2 \to \mathcal{H}_1 $ whose image in $ {\operatorname{Ext}}^1({\mathcal I}_D, {\mathcal O}_{{\mathbb P}^r}(c-r-1)) $ is equal to $ \xi.$ By using standard results from homological algebra, we get that $ \mathcal{N} $ is the push–out of $ \mathcal{H}_1 $ and $ {\mathcal O}_{{\mathbb P}^r}(c-r-1) $ via $ (\varphi, -\psi).$ Hence, the resolution of $ \mathcal{N} $ with dissocié sheaves is $$0 \to \mathcal{H}_2 \stackrel{(\varphi,-\psi)}{\longrightarrow} \mathcal{H}_1 \oplus {\mathcal O}_{{\mathbb P}^r}(c-r-1) \to \mathcal{N} \to 0$$ and so $ \mathcal{N} $ has projective dimension less than or equal to $ 1.$ From the proof of the previous Proposition, we get that $ {\operatorname{pd}}(\mathcal N)=0 $ if and only if $ \mathcal{H}_2 = {\mathcal O}_{{\mathbb P}^r}(c-r-1), $ i.e. $ D $ is a complete intersection scheme. We can easily modify the proof to get sheaves $ \mathcal{N} $ of larger rank: it is enough to consider $ c_1, \dots, c_n \in {\mathbb Z}$ such that $ H^0(D, \omega_D(c_i)) \not= 0 $ for at least a $ c_i.$ As in the proof of the previous Proposition, $ \oplus_{i=1}^n H^0(D, \omega_D(c_i)) \cong {\operatorname{Ext}}^1({\mathcal I}_D, \oplus_{i=1}^n {\mathcal O}_{{\mathbb P}^r}(c_i-r-1)) $ and so a non–zero element $ \xi \in \oplus_{i=1}^n H^0(D, \omega_D(c_i)) $ can be considered as an extension of $ {\mathcal I}_D $ with $ \mathcal P = \oplus_{i=1}^n {\mathcal O}_{{\mathbb P}^r}(c_i-r-1),$ and we can construct $ \mathcal N $ as in the proof. In comparing Proposition \[cod2\] with Serre’s and Hartshorne’s constructions mentioned above, it is evident that the hypothesis on $ \mathcal N $ strongly affects the properties of the constructed scheme. For example, when $ \mathcal N $ is a rank $ 2 $ reflexive sheaf, as in Hartshorne’ s setting, the associated schemes are generically locally complete intersection. In fact, the locus where the reflexive sheaf $ \mathcal N $ is not locally free has codimension $ \geq 3 $ ([@hart-2], Corollary 1.4 and Theorem 4.1 for the case of curves in $ {\mathbb P}^3 $). The properties of the associated schemes show that the constructions are not the same one. In fact, following Proposition \[cod2\], it is possible to construct ACM schemes which are locally complete intersection at no point, while if $ \mathcal N $ is reflexive and $ D $ is the associated scheme, the locus of the points of $ D $ where $ D $ is not locally complete intersection has codimension $ \geq 1 $ in $ D.$ On the other hand, all the schemes constructed via Proposition \[cod2\] are ACM, while the ones associated to reflexive sheaves can have non–zero cohomology. Now, we show how to construct ACM codimension 2 schemes which contain the first infinitesimal neighborhood of another ACM codimension 2 scheme. They are candidates to have no points at which the scheme is locally complete intersection. Let $ Y $ be an ACM codimension 2 scheme and let $ \mathcal{N} = {\mathcal I}_Y \oplus {\mathcal I}_Y.$ Then, every codimension $ 2 $ ACM scheme $ D $ we obtain from the construction above contains the first infinitesimal neighborhood of $ Y.$ Moreover, $D$ is not locally complete intersection at any point of $Y$.In particular it is not generically locally complete intersection. For the first statement, it is enough to prove that $ I_D \subset I_Y^2.$ Let $ 0 \to \mathcal L_1 \stackrel{\varphi}{\longrightarrow} \mathcal L_0 \to \mathcal I_Y \to 0 $ be the minimal dissocié resolution of $ \mathcal I_Y. $ Let $ \varphi $ be represented by a matrix $ A. $ Hence, the maximal minors of $ A $ generate the ideal $ I_Y.$ Let $ \mathcal P = {\mathcal O}_{{\mathbb P}^r}(-m) $ and let $ \gamma: \mathcal P \to \mathcal N $ be a general map whose image is a maximal subsheaf of $ \mathcal N. $ Let $ \gamma': \mathcal P \to \mathcal L_0 \oplus \mathcal L_0 $ be a lifting of $ \gamma.$ The ideal $ I_D $ is generated by the maximal minors of the matrix $$M = \left( \begin{array}{ccc} A & O & C' \\ O & A & C'' \end{array} \right)$$ where the last column represents $ \gamma'.$ Every maximal minor of $ M $ can be computed by Laplace rule with respect to the last column, and so it is a combination of the maximal minors of the block matrix $ \left( \begin{array}{cc} A & O \\ O & A \end{array} \right),$ whose maximal minors generate the ideal $ I_Y^2.$ Let now $x \in Y$ and set $S: ={\mathcal O}_{{{\mathbb P}^r},x}$. We have an exact sequence of $S$-modules $$0 \to S \to \mathcal N_x \to {\mathcal I}_{D,x} \to 0.$$ It is easy to see that $\mathcal N_x$ needs at least four generators whence ${\mathcal I}_{D,x}$ needs at least three generators. Since $D$ has codimension $ 2 $ it cannot be a complete intersection at $x$. The easiest case we can consider is when the scheme $ Y $ is the complete intersection of two hypersurfaces. In this case, the scheme defined by $ I_Y^2 $ is ACM of codimension $ 2 $ and it can be obtained from the previous construction. A similar result holds both for the direct sum of $ s (\geq 2 ) $ copies of $ {\mathcal I}_Y,$ and for non–trivial extensions of $ {\mathcal I}_Y $ with itself or with twists of another ACM codimension $ 2 $ scheme $ Z $, but we do not state them. Now, we relate extensions associated to divisors that differ by hypersurface sections. Let $ D \subset {\mathbb P}^r $ be a codimension $ 2 $ ACM scheme. Let $ \xi \in H^0(D, \omega_D(c)) $ and $ \xi' \in H^0(D, \omega_D(c+d)) $ both non–zero, with $ d \geq 0,$ and let $$0 \to {\mathcal O}_{{\mathbb P}^r}(c-r-1) \to \mathcal{N} \to {\mathcal I}_D \to 0$$ and $$0 \to {\mathcal O}_{{\mathbb P}^r}(c+d-r-1) \to \mathcal{N}' \to {\mathcal I}_D \to 0$$ be the associated short exact sequences. Then, there exists a degree $ d $ hypersurface $ S = V(f) $ that cuts $ D $ along a codimension $ 3 $ subscheme such that $ \xi' = f \xi $ if, and only if, there exists a short exact sequence $$0 \to \mathcal{N} \to \mathcal{N}' \to {\mathcal O}_S(c+d-r-1) \to 0$$ that induces the identity on $ {\mathcal I}_D.$ In the proof of previous Proposition, we constructed the sheaf $ \mathcal{N} $ as push–out $$\commdiag{\mathcal{H}_2 & \mapright\lft{\varphi} & \hspace{.5 cm} \mathcal{H}_1 \cr \mapdown\lft{\psi} & & \hspace{.5 cm} \mapdown \cr {\mathcal O}_{{\mathbb P}^r}(c-r-1) & \mapright & \hspace{.5 cm} \mathcal{N} }.$$ Assume that $ \xi' = f \xi.$ The section $ f \xi \in H^0(D, \omega_D(c+d)) $ is the image of the map $ f \psi \in {\operatorname{Hom}}(\mathcal{H}_2, {\mathcal O}_{{\mathbb P}^r}(c+d-r-1)) $ in $ {\operatorname{Ext}}^1({\mathcal I}_D, {\mathcal O}_{{\mathbb P}^r}(c+d-r-1)) $ and so the sheaf $ \mathcal{N}' $ is the push–out of $ \mathcal{H}_1 $ and $ {\mathcal O}_{{\mathbb P}^r}(c+d-r-1) $ via $ \varphi $ and $ -f \psi.$ From the universal property of the push–out (see [@lang], pp. 62), we get the following map of complexes $$\commdiag{0 \to & {\mathcal O}_{{\mathbb P}^r}(c-r-1) & \mapright & \mathcal{N} & \mapright & {\mathcal I}_D & \to 0 \cr & \mapdown\lft{f} & & \mapdown\lft{\varepsilon} & & \bivline & \cr 0 \to & {\mathcal O}_{{\mathbb P}^r}(c+d-r-1) & \mapright & \mathcal{N}' & \mapright & {\mathcal I}_D & \to 0 }$$ and so $ \varepsilon $ is injective, and $ \mbox{coker}(\varepsilon) \cong {\mathcal O}_S(c+d-r-1),$ as claimed. Assume now that the short exact sequence $$0 \to \mathcal{N} \to \mathcal{N}' \to {\mathcal O}_S(c+d-r-1) \to 0$$ induces the identity on $ {\mathcal I}_D.$ Standard arguments allow us to lift $ \varepsilon $ to an injective map $ {\mathcal O}_{{\mathbb P}^r}(c-r-1) \to {\mathcal O}_{{\mathbb P}^r}(c+d-r-1) $ whose cokernel is isomorphic to $ {\mathcal O}_S(c+d-r-1).$ Hence, the map is the multiplication by $ f,$ and $ \mathcal{N}' $ is the push–out of $ \mathcal{H}_1 $ and $ {\mathcal O}_{{\mathbb P}^r}(c+d-r-1) $ via $ \varphi $ and $ f \psi.$ Hence, $ \xi' = f \xi,$ and the proof is complete. Let $ \xi, \xi' \in H^0(D, \omega_D(c)).$ By applying the previous Proposition, we get that $ \xi $ and $ \xi' $ are linearly dependent if and only if the sheaves $ \mathcal N $ and $ \mathcal N' $ associated to them are isomorphic. ACM schemes from ACM ones ========================= Let $ X \subset {\mathbb P}^r $ be a codimension $ t $ ACM scheme. For general choices, $ s (< t) $ hypersurfaces of large degree containing $ X $ define a complete intersection codimension $ s $ ACM scheme containing $ X.$ In this section, we discuss the related problem of finding an ACM codimension $ s $ closed scheme $ D \subset {\mathbb P}^r $ containing $ X.$ Of course, we make use of the construction described in the previous section. The main result is the following. \[XtoD\] Let $ X $ be a codimension $ t $ ACM scheme in $ {\mathbb P}^r $ with $ 3 \leq t \leq r $ and let $$\label{res_IX} 0 \to F_t \stackrel{\delta_t}{\longrightarrow} F_{t-1} \to \dots \to F_2 \stackrel{\delta_2}{\longrightarrow} F_1 \stackrel{\delta_1}{\longrightarrow} I_X \to 0$$ be the minimal free resolution of the saturated ideal that defines $ X.$ Let $ N = \ker(\delta_{t-s}) $ be the $ (t-s)$–th syzygy module of $ X,$ for some $ s \geq 2,$ and let $ P $ be a torsion–free $ R$–module of projective dimension $ s-2.$ Assume further that $ \mathcal N $ and $ \mathcal P $ satisfy the condition $ (H.3),$ and that there exists an injective map $ \gamma: \mathcal{P} \to \mathcal{N} $ such that $ \gamma(\mathcal{P}) $ is a maximal subsheaf of $ \mathcal{N}.$ Then, for every ACM codimension $ s $ closed scheme $ D $ constructed as in Theorem \[NtoD\] there is a short exact sequence $$0 \to \mathcal{E}xt^{s-2}(\mathcal{P}, \omega_{{\mathbb P}^r}) \to \omega_D(-k) \to \omega_X \to 0.$$ Moreover, $ D $ contains $ X. $ The $ R$–module $ N $ is torsion–free, and has no free summand, because it is computed from the minimal free resolution of $ I_X.$ Moreover, $$0 \to F_t \stackrel{\delta_t}{\longrightarrow} F_{t-1} \stackrel{\delta_{t-1}}{\longrightarrow} \dots \stackrel{\delta_{t-s+2}}{\longrightarrow} F_{t-s+1} \to N \to 0$$ is the minimal free resolution of $ N $ and so the projective dimension of $ N $ is $ s-1.$ Hence, $ \mathcal N $ and $ \mathcal P $ satisfy all the conditions (\[H-hyp\]). Hence, by Theorem \[NtoD\] there exists a codimension $ s $ ACM closed scheme $ D \subset {\mathbb P}^r,$ and an integer $ k $ such that $$0 \to \mathcal{P} \to \mathcal{N} \to {\mathcal I}_D(k) \to 0$$ is a short exact sequence. By applying $ \mathcal{H}om(-, \omega_{{\mathbb P}^r}) $ we get $$\begin{split} \mathcal{E}xt^{s-2}(\mathcal{N}, \omega_{{\mathbb P}^r}) & \to \mathcal{E}xt^{s-2}(\mathcal{P}, \omega_{{\mathbb P}^r}) \to \mathcal{E}xt^{s-1}({\mathcal I}_D(k), \omega_{{\mathbb P}^r}) \to \\ & \to \mathcal{E}xt^{s-1}(\mathcal{N}, \omega_{{\mathbb P}^r}) \to \mathcal{E}xt^{s-1}(\mathcal{P}, \omega_{{\mathbb P}^r}). \end{split}$$ $ \mathcal{E}xt^{s-1}(\mathcal{P}, \omega_{{\mathbb P}^r}) = 0 $ because $ {\operatorname{pd}}(\mathcal P) = s-2,$ while $ \mathcal{E}xt^{s-j}(\mathcal{N}, \omega_{{\mathbb P}^r}) = \mathcal{E}xt^{t-j}({\mathcal I}_X, \omega_{{\mathbb P}^r}) $ by definition of $ N.$ Hence, $ \mathcal{E}xt^{s-1}(\mathcal{N}, \omega_{{\mathbb P}^r}) = \omega_X,$ and $ \mathcal{E}xt^{s-2}(\mathcal{N}, \omega_{{\mathbb P}^r}) = 0 $ because $ X $ is ACM of codimension $ t $ ([@hart], Ch. III, Proposition 7.5 and Theorem 7.1). Again by ([@hart], Ch. III, Proposition 7.5), $ \mathcal{E}xt^{s-1}({\mathcal I}_D(k), \omega_{{\mathbb P}^r}) = \omega_D(-k).$ Summarizing the above arguments, the construction induces a short exact sequence $$0 \to \mathcal{E}xt^{s-2}(\mathcal{P}, \omega_{{\mathbb P}^r}) \to \omega_D(-k) \to \omega_X \to 0$$ that relates the dualizing sheaves of $ X $ and $ D.$ In particular, we can think of $ \omega_X $ as a quotient of $ \omega_D,$ up to a twist. The annihilator of $ \omega_X $ is $ {\mathcal I}_X $ (see, [@ei], Corollary 21.3), the one of $ \omega_D $ is $ {\mathcal I}_D,$ and so we get the last claim because it is evident that the annihilator of $ \omega_D(-k) $ is contained in the one of $ \omega_X.$ The previous Proposition explains our motivation in studying the exact sequences as (\[pnd\]). In fact, we applied the construction by M.Martin–Deschamps and D.Perrin to the first syzygy module $ N $ of a zero–dimensional scheme $ X $ in $ {\mathbb P}^3,$ i.e. $ {\operatorname{pd}}(N) = 1.$ The above mentioned construction provides a free module $ P $ ($ {\operatorname{pd}}(P) = 0 $) and a general injective map $ \gamma: P \to N $ whose cokernel is, up to a twist, the ideal of a curve $ D $ (and so the codimension of $ D $ is $ 2 $). Hence, the hypotheses of Proposition \[XtoD\] are fulfilled and the curve $ D $ is ACM and contains $ X.$ We rephrase Proposition \[constrains\] in the case $ \mathcal N $ is the $ (t-s)$–syzygy sheaf of an ACM scheme $ X $ of codimension $ t.$ \[constrains1\] Let $ X $ and $ D $ be schemes as in Proposition \[XtoD\]. Then, the Cohen–Macaulay type of $ X $ is not greater than the one of $ D.$ In particular, $ D $ is arithmetically Gorenstein if and only if $ X $ is such. The minimal dissocié resolution of $ \mathcal N $ agrees with the one of $ {\mathcal I}_X,$ and so $ F_t \oplus G_{s-1} $ appears in a free resolution of $ I_D(k),$ as it follows from Remark \[rem3.3\]. $ F_t $ cannot be cancelled because it maps to $ F_{t-1} $ and the resolution of $ I_X $ is minimal, and so the first claim follows. In particular, $ F_t $ is equal to the last free module in a minimal free resolution of $ I_D(k) $ if and only if $ \gamma_{s-1} : G_{s-1} \to F_{t-1} $ is split–injective, where $ \gamma_{s-1} $ is induced from $ \gamma: P \to N.$ The second statement is straightforward. For example, if $ X \subset {\mathbb P}^3 $ is a set of $ 5 $ general points, it is arithmetically Gorenstein with Pfaffian resolution $$0 \to R(-5) \to R^5(-3) \to R^5(-2) \to I_X \to 0.$$ By applying the previous construction with $ P = R^3(-3),$ we get that $ k = -1 $ and the minimal free resolution of $ I_D $ is $$0 \to R(-5) \to R^2(-3) \to I_D(-1) \to 0,$$ so $ D $ is a complete intersection curve in $ {\mathbb P}^3.$ \[4-points\] Among the ACM closed schemes $ D $ constructed in Proposition \[XtoD\] we might not find the ones of minimal degree containing $ X.$ For example, let $ X \subset {\mathbb P}^3 $ be the degree $ 4 $ reduced scheme consisting of the vertices of the unit tetrahedron. With an easy computation, we get that $ I_X $ is generated by $ xy, xz, xw, yz, yw, zw,$ and its minimal free resolution is $$0 \to R^3(-4) \to R^8(-3) \to R^6(-2) \to I_X \to 0.$$ An ACM curve $ C $ of minimal degree containing $ X $ is the union of the three lines $ V(x,y), V(y,z), V(z,w).$ The minimal free resolution $I_C$ is $$0 \to R^2(-3) \to R^3(-2) \to I_C \to 0.$$ It follows that $ C $ cannot be obtained from Proposition \[XtoD\] because the Cohen–Macaulay types of $ X $ and $ C $ are $ 3 $ and $ 2,$ respectively, and this is not possible by Corollary \[constrains1\]. In this example, we construct two ACM curves with different Cohen–Macaulay types starting from the same $ X.$ Let $ r = 3 $ and let $ X $ be a set of four general points in a plane. Of course, $ I_X $ is the complete intersection of a linear form and two quadratic forms, and so its minimal free resolution is $$0 \to R(-5) \to R^2(-3) \oplus R(-4) \to R(-1) \oplus R^2(-2) \to I_X \to 0.$$ If we choose $ P = R(-3),$ we get a complete intersection curve $ D $ whose minimal free resolution is $$0 \to R(-5) \to R(-3) \oplus R(-4) \to I_D(-2) \to 0.$$ On the other hand, if we choose $ P = R(-5),$ we get an ACM curve $ E $ whose minimal free resolution is $$0 \to R^2(-5) \to R^2(-3) \oplus R(-4) \to I_E \to 0.$$ Both curves are constructed by choosing a general injective map from $ P $ to $ R^2(-3) \oplus R(-4).$ Summarizing the obtained results, we proved that it is possible to construct a codimension $ s $ ACM closed scheme $ D $ containing a given codimension $ t $ ACM scheme $ X $ as soon as $ s < t.$ Some of the restrictions are: the number of minimal generators of $ I_D $ cannot be larger than the number of minimal generators of the $ R$–module $ N $ we used in the construction, and the last free module in a minimal free resolution of $ I_X $ is a direct summand of the last free module in a minimal free resolution of $ I_D(k).$ A consequence of the restrictions is that there are ACM schemes containing $ X $ that cannot be constructed as explained in Proposition \[XtoD\] (e.g., see Remark \[4-points\]). The last result we present in this section allows us to reconstruct an ACM scheme $ D $ from a subscheme $ X $ of $ D $ obtained by intersecting $ D $ with a complete intersection $ S.$ Let $ D \subset {\mathbb P}^r $ be a codimension $ s $ ACM scheme with minimal free resolution $$0 \to H_s \stackrel{\varepsilon_s}{\longrightarrow} H_{s-1} \stackrel{\varepsilon_{s-1}}{\longrightarrow} \dots \stackrel{\varepsilon_2}{\longrightarrow} H_1 \to I_D \to 0,$$ and let $ S = V(f_1, \dots, f_t) $ be a codimension $ t $ complete intersection scheme that cuts $ D $ along a codimension $ s+t \leq r $ scheme $ X.$ Then, $ D $ can be constructed from $ X $ as explained in Proposition \[XtoD\]. Let $ F = \oplus_{i=1}^t R(-deg(f_i)).$ Then, the minimal free resolution of $ I_S $ is given by the Koszul complex $$0 \to \wedge^t F \stackrel{\varphi_t}{\longrightarrow} \wedge^{t-1} F \stackrel{\varphi_{t-1}}{\longrightarrow} \dots \stackrel{\varphi_2}{\longrightarrow} F \stackrel{\varphi_1}{\longrightarrow} I_S \to 0$$ where $ \varphi_i = \wedge^i \varphi $ and $ \varphi:F \to R $ is defined as $ \varphi(e_i) = f_i $ for each $ i = 1, \dots, t,$ where $ e_1, \dots, e_t $ is the canonical basis of $ F.$ Let $ X = D \cap S,$ and let $ I_X \subset R $ be its saturated ideal. It is easy to prove that a free resolution of $ I_X $ can be constructed as tensor product of the resolutions of $ I_D $ and $ I_S $ (for the definition of the tensor product of complexes see Section 17.3 in [@ei]). Hence, it is equal to $$0 \to G_{s+t} \to G_{s+t-1} \to \dots \to G_1 \to I_X \to 0$$ where $$G_h = \bigoplus_{i+j=h,\ i, j \geq 0} H_i \otimes \wedge^j F$$ for $ h = 1, \dots, s+t,$ and the map $ \delta_h : G_h \to G_{h-1} $ restricted to $ H_i \otimes \wedge^j F \to (H_{i-1} \otimes \wedge^j F) \oplus (H_i \otimes \wedge^{j-1} F) $ is defined as $$\delta_i = \left( \begin{array}{c} \varepsilon_i \otimes 1 \\ (-1)^i\ 1 \otimes \varphi_j \end{array} \right).$$ In particular, $ X $ is ACM of codimension $ s+t.$ Let $ N $ be the kernel of $ \delta_t,$ and so a resolution of $ N $ is equal to $$0 \to G_{s+t} \to \dots \to G_{t+1} \to N \to 0.$$ Moreover, $ N $ is torsion–free. Now, let $ G'_{t+j} = (H_{j+1} \otimes \wedge^{t-1} F) \oplus \dots \oplus (H_s \otimes \wedge^{t+j-s} F) $ for $ j = 1, \dots, s-1.$ Of course, $ G_{t+j} = (H_j \otimes \wedge^t F) \oplus G'_{t+j}.$ Let $ \Delta_{t+j} : G'_{t+j} \to G'_{t+j-1} $ be the restriction of $ \delta_{t+j} $ to $ G'_{t+j},$ and let $ P = \mbox{coker}(\Delta_{t+2}).$ A free resolution of $ P $ is $$0 \to G'_{s+t-1} \to G'_{s+t-2} \to \dots \to G'_{t+1} \to P \to 0.$$ In fact, it is easy to prove that it is a complex. Furthermore, it is exact, because it is a sub–complex of the resolution of $ I_X.$ It is obvious that the inclusion $ G'_{t+j} \to G_{t+j} $ for $ j \geq 1,$ induces an inclusion $ P \to N.$ The resolution of the cokernel is $$0 \to H_s \otimes \wedge^t F \to H_{s-1} \otimes \wedge^t F \to \dots \to H_1 \otimes \wedge^t F \to N/P \to 0.$$ But $ \wedge^t F \cong R(-\sum_{i=1}^t \deg(f_i)) $ and the maps are $ \varepsilon_i \otimes 1.$ Hence, $ \mathcal{N}/\mathcal{P} \cong {\mathcal I}_D(k) $ where $ k = -\sum_{i=1}^t \deg(f_i).$ In particular, from Proposition 2.9 it follows that $ \mathcal{P} $ is a maximal sub–sheaf of $ \mathcal{N},$ and so the claim is proved. In the previous theorem, suppose $ D $ is a complete intersection. Then, $ X $ is a complete intersection too and $ I_X $ is generated by a regular sequence obtained by taking all the generators of $ I_D $ and $ I_S. $ Reversing this observation, we consider a complete intersection scheme $ X $ generated by a regular sequence of forms $ (f_0, \dots, f_i), $ with $ i \leq r.$ Starting from $ X $ we can obtain all the schemes $ D $ generated by a subset of generators of $ X. $ In particular, if we take $ i=r, $ and $ f_j=x_j, j=0, \dots r, $ the $(r-1)$–syzygy sheaf $ \mathcal N $ involved in the construction is a twist of the tangent sheaf $ T_{{\mathbb P}^r}.$ [99]{} M.P. Brodmann, R.Y. Sharp, *Local Cohomology: an algebraic introduction with geometric applications, Cambridge Univ. Press (1998).* D. Eisenbud, *Commutative Algebra, **GTM 150, Springer Verlag (1994).*** M. Green, *Koszul cohomology and geometry, in Lectures on Riemann Surfaces, ed. M. Cornalba, World Scientific Press (1989).* R. Hartshorne, *Varieties of small codimension in projective space, Bull. AMS **80 (1974), 1017–1032.*** R. Hartshorne, *Algebraic Geometry, **GTM 52,Springer Verlag (1977).*** R. Hartshorne, *Stable reflexive sheaves, Math. Ann. **254 (1980), 121–176.*** S. Lang, *Algebra, **GTM 211, Springer Verlag (2002).*** M. Martin-Deschamps, D. Perrin, *Sur la classification des courbes gauches, Astérisque **184-185, (1990).*** H. Matsumura, *Commutative Algebra, Math. Lecture Note Series, The Benjamin/Cummming Pub. Co., (1980).* J.C. Migliore, *Introduction to Liaison Theory and Deficiency Modules, Progress in Math., **Vol. 165, Birkh$\ddot{a}$user (1998).*** D.G. Northcott, *A first course of homological algebra, Cambridge University Press (1973).* C. Okonek, M. Schneider, H. Spindler, *Vector Bundles on Complex Projective Spaces, Progress in Math., **Vol. 3, Birkh$\ddot{a}$user (1980).*** J.P. Serre, *Sur les modules projectifs, Sém. Dubreil-Pisot, exp. 2, (1960/61), 1–16.* [^1]: All the authors have been supported by the framework of Prin 2008 “Geometria delle varietà algebriche e dei loro spazi di moduli", cofinanced by MIUR

ArXiv

--- abstract: | We consider dynamical systems generated by skew products of affine contractions on the real line over angle-multiplying maps on the circle $S^1$: $$T:S^{1}\times \R\to S^{1}\times \R,\qquad T(x,y)=(\lap x, \lambda y+f(x))$$ where $\lap\ge 2$, $0<\lambda<1$ and $f$ is a $C^{r}$ function on $S^{1}$. We show that, if $\lambda^{1+2s}\lap>1$ for some $0\leq s< r-2$, the density of the SBR measure for $T$ is contained in the Sobolev space $W^{s}(S^{1}\times \R)$ for almost all ($C^r$ generic, at least) $f$. author: - 'Artur Avila[^1], Sébastien Gouëzel[^2] and Masato Tsujii[^3]' date: 'February 24, 2005' title: Smoothness of solenoidal attractors --- Introduction ============ In this paper, we study dynamical systems generated by skew products of affine contractions on the real line over angle-multiplying maps on the circle $S^1=\R/\Z$: $$\label{eqn:Tdef} T:S^{1}\times \R\to S^{1}\times \R,\qquad T(x,y)=(\lap x, \lambda y+f(x))$$ where $\lap\ge 2$ is an integer, $0<\lambda<1$ is a real number and $f$ is a $C^{r}$ function on $S^{1}$ (for some integer $r \geq 3$). It admits a forward invariant closed subset $A$ such that $\omega(\x)=A$ for Lebesgue almost every point $\x\in \ps$. Further, there exists an ergodic invariant probability measure $\mu$ on $A$ for which Lebesgue almost every point on $\ps$ is generic. The measure $\mu$ is called the [*SBR measure*]{} for $T$. If $T$ is locally area contracting, [*i.e.*]{}, $\det DT\equiv \lambda \lap<1$, the subset $A$ is a Lebesgue null subset and hence the SBR measure is totally singular with respect to the Lebesgue measure. In [@T], the third named author studied the case where $T$ is locally area expanding, [*i.e.*]{}, $\lambda \lap >1$, and proved that the SBR measure is absolutely continuous with respect to the Lebesgue measure for $C^r$ generic $f$. In the present paper, we study the smoothness of the density of the SBR measure in more detail, and the mixing properties of $T$. \[th:main1\] If $\ell$ and $\lambda$ satisfy $\lambda^{1+2s} \lap>1$ for some $0\leq s< r-2$, the density of the SBR measure $\mu$ with respect to the Lebesgue measure is contained in the Sobolev space $W^{s}(\ps)$ for any $f$ in an open dense subset of $C^r(S^1,\R)$. Since the elements of $W^{s}(\ps)$ for $s>1$ are continuous up to modification on Lebesgue null subsets from Sobolev’s embedding theorem, it follows If $\lambda^{3} \lap>1$ and $r\ge 4$, the attractor $A$ has non-empty interior for $f$ in an open dense subset of $C^r(S^1,\R)$. Recently, Bamón, Kiwi and Rivera-Letelier announced the following result: for an open dense subset of $C^{1+ \epsilon}$ hyperbolic endomorphisms of the annulus, $\log d+\chi>0$ implies that the attractor has non-empty interior, where $d$ is the degree of the induced map in homology and $\chi$ is the negative Lyapunov exponent of the SBR measure. (See also [@B].) When $s>1$, we also obtain that the density of the SBR measure is essentially bounded. Together with the results of Rams in [@R], it gives examples of solenoids in higher dimensions for which the invariant measure is equivalent to the Hausdorff measure. The Perron-Frobenius operator $P:L^1(\ps)\to L^1(\ps)$ is defined by $$Ph(\x)=\frac{1}{\lambda\lap}\sum_{\y\in T^{-1}(\x)} h(\y),$$ and characterized by the property that $$\label{eqn:ch} \frac{dT_*\nu}{d\vol}=P\left(\frac{d\nu}{d\vol}\right)$$ for any finite measure $\nu$ which is absolutely continuous with respect to the Lebesgue measure $\vol$ on $\ps$. When $s>1/2$, we obtain a precise spectral description of $P$, which strengthens considerably Theorem \[th:main1\]. \[th:main2\] Assume that $\ell$ and $\lambda$ satisfy $\lambda^{1+2s} \lap>1$ for some $1/2< s< r-2$. Take $\gamma\in \left ((\lambda^{1+2s}\lap)^{-1/2},1 \right )$. For any $f$ in an open dense subset of $C^r(S^1,\R)$, there exists a Banach space ${\mathcal{B}}$ contained in $W^s(\ps)$ on which the transfer operator $P$ acts continuously with an essential spectral radius at most $\gamma$ (in particular, $P$ admits a spectral gap, and the correlations of $T$ decay exponentially fast). Moreover, ${\mathcal{B}}$ can be chosen to contain all functions in $C^{r-1}(\ps)$ supported in some given (fixed) compact subset of $\ps$. Since $T$ is uniformly hyperbolic, the exponential decay of correlations was already known. The novel feature of our theorem is that, when the contraction coefficient $\lambda$ tends to $1$, our estimates do *not* degenerate. In fact, the inequality $\lambda<1$ is used only to ensure that a compact subset of $S^1 \times \R$ is invariant, to get an SBR measure. Hence, our method may probably be generalized to settings with a neutral (or slightly positive) exponent on a compact space. Fix $\lap\ge 2$ and let $\D_{r,s}\subset (0,1)\times C^{r}(S^{1},\R)$ be the set of pairs $(\lambda,f)$ such that the conclusions of Theorems \[th:main1\] and \[th:main2\] hold. Let $\D^{\circ}_{r,s}$ be the interior of $\D_{r,s}$. The following result shows that Theorems \[th:main1\] and \[th:main2\] hold for “almost all” $T$, in a precise sense: \[th:main\] If $\ell$ and $\lambda$ satisfy $\lambda^{1+2s} \lap>1$ for some $0\leq s< r-2$, there exists a finite collection of $C^{\infty}$ functions $\varphi_{i}:S^{1}\to \R$, $1\le i\le m$, such that, for any $g\in C^{r}(S^{1},\R)$, the subset $$\left\{(t_{1},t_{2},\cdots,t_{m})\in \R^{m} \;\left|\;\left(\lambda,\; g(x)+\sum_{i=1}^{m} t_{i}\varphi_{i}(x)\right)\notin \D^{\circ}_{r,s}\right.\right\}$$ is a null subset with respect to the Lebesgue measure on $\R^{m}$. We proceed as follows. In the next section, we introduce some definitions related to a transversality condition on the mapping $T$, which is similar to (but slightly different from) that used in [@T]. This transversality condition is proved to be a generic one in the last section. In Section \[sec:pf\], we introduce some norms on the space of $C^r$ functions on $\ps$ and prove a Lasota-Yorke type inequality for them, imitating the argument in the recent paper [@GL] of C. Liverani and the second named author with slight modification. Section \[sec:ly\] is the core of this paper, where we prove a Lasota-Yorke inequality involving the $W^s$ norm and the norm introduced in Section \[sec:pf\]. Finally, in Section \[sec:proofs\_thms\], we show how these Lasota-Yorke inequalities imply the main results of the paper. Some definitions {#sec:def} ================ From here to the end of this paper, we fix an integer $\lap\ge 2$, real numbers $0<\lambda<1$ and $0\le s <r-2$ satisfying $\lambda^{1+2s}\lap>1$. We also fix a positive number $\kappa$ and consider the mapping $T$ for a function $f$ in $$\mathcal{U}=\mathcal{U}_\kappa= \left\{ f\in C^r(S^1,\R)\;;\; \|f\|_{C^r}:=\max_{0\le k\le r} \sup_{x\in S^1}\left|\frac{d^k}{dx^k}f(x)\right|\le \kappa \right\}.$$ Fix ${\alpha_0} =\kappa/(1-\lambda)$ and let $\ann=S^1\times [-{\alpha_0},{\alpha_0}]$. Then we have $T(\ann)\subset \ann$. Let $\P$ be the partition of $S^{1}$ into the intervals $\P(k)=[ (k-1)/\lap, k/\lap)$ for $1\le k\le \ell$. Let $\tau:S^{1}\to S^{1}$ be the map defined by $\tau(x)=\lap \cdot x$. Then the partition $\P^{n}:=\bigvee_{i=0}^{n-1}\tau^{-i}(\P)$ for $n\ge 1$ consists of the intervals $$\P(\a)=\bigcap_{i=0}^{n-1}\tau^{-i}\left(\P(a_{n-i})\right), \qquad \a=(a_{i})_{i=1}^{n}\in\sym^{n}$$ where $\sym^{n}$ denotes the space of words of length $n$ on the set $\sym=\{1,2,\cdots,\lap\}$. Notice that $\a$ is the [*reverse*]{} of the itinerary of points in $\P(\a)$. For $x\in S^{1}$ and $\a\in \sym^{n}$, there is a unique point $y\in \P(\a)$ such that $\tau^{n}(y)=x$, which is denoted by $\a (x)$. For $\a=(a_{i})_{i=1}^{n}\in \sym^{n}$, the image of the segment $\P(\a)\times\{0\}\subset S^{1}\times \R$ under the iterate $T^{n}$ is the graph of the function $S(\cdot,\a)$ defined by $$S(x,\a):=\sum_{i=1}^{n}\lambda^{i-1}f(\tau^{n-i}(\a(x))) =\sum_{i=1}^{n}\lambda^{i-1}f([\a]_{i}(x))$$ where $[\a]_{q}=(a_{i})_{i=1}^{q}$. For a word $\a=(a_{i})_{i=1}^{\infty}\in \sym^{\infty}$ of infinite length, we define $$S(x,\a)=\lim_{i\to\infty}S(x,[\a]_{i})=\sum_{i=1}^{\infty}\lambda^{i-1}f([\a]_{i}(x)).$$ For a word $\c$ of length $m$, let $\P_*(\c)$ be the union of the interval $\P(\c)$ and the two intervals in $\P^{m}$ adjacent to it. The function $S(\cdot,\a)$ for a word $\a\in \sym^n$ with $1\le n\le \infty$ may not be continuous on $\P_*(\c)$ when $\P(\c)$ has $0\in S^1$ as its end. Nevertheless the restriction of $S(\cdot,\a)$ to $\P(\c)$ can be naturally extended to $\P_*(\c)$ as a $C^{r}$ function. Indeed, letting $\tau^{-i}_{\c,\a}:\P_*(\c)\to S^1$ be the branch of the inverse of $\tau^i$ satisfying $\tau^{-i}_{\c,\a}(\P(\c))\subset \P([\a]_i)$, the extension is given by $$\label{eqn:Sc} S_{\c}(\cdot,\a) : \P_*(\c)\to\R,\quad S_{\c}(x,\a):=\sum_{i=1}^{n}\lambda^{i-1}f(\tau^{-i}_{\c,\a}(x)).$$ For any word $\a$ of finite or infinite length, we have $$\label{eqn:alpha} \sup_{x\in \P_*(\c)}\max_{0\le \nu\le r} \lap^\nu \left|\frac{d^{\nu}}{d x^{\nu}}S_{\c}(x,\a)\right|\le {\alpha_0}.$$ For $\a,\b\in \sym^{q}$ and $\c\in \sym^{p}$, we say that $\a$ and $\b$ are [*transversal*]{} on $\c$ and write $\a\pitchfork_{\c}\b$ if $$\left|\frac{d}{d x}S_{\c}(x,\a)-\frac{d}{d x}S_{\c}(y,\b)\right|>2\lambda^q\ell^{-q} {\alpha_0}$$ at all points $x,y$ in the closure of $\P_*(\c)$. We put $$\e(q,p)=\max_{\c\in\sym^{p}}\max_{\a\in \sym^{q}} \#\{\b\in\sym^{q}\mid \a\not \pitchfork_{\c} \b\} \qquad \mbox{ and}\qquad \e(q)=\lim_{p\to\infty}\e(q,p).$$ The main argument of the proof will be to construct norms which will satisfy a Lasota-Yorke inequality if $\e(q)$ is not too big for some $q$. This will readily imply the two main theorems if the norms have sufficiently good properties. To conclude, a transversality argument (similar to the arguments in [@T]) will show that, for almost all functions $f$ (in the sense of Theorem \[th:main\]), $\e(q)$ is not too big for some $q$. Henceforth, and until the end of Section \[sec:ly\], we fix a large integer $q$. By definition, there exists $p_0\geq 1$ such that $\e(q,p)=\e(q)$ for $p\ge p_0$. We also fix an integer $p\geq p_0$. Perron-Frobenius operator and the norm $\|\cdot\|^\ac_\rho$ {#sec:pf} =========================================================== Let $C^r(\ann)$ be the set of $C^r$ functions on $\ps$ whose supports are contained in $\ann$. In this section, we define preliminary norms on the space $C^r(\ann)$ and show Lasota-Yorke type inequalities for them. For the definition of the norms, we prepare a class $\Omega$ of $C^r$ curves on $\ps$. Let $\gamma:\mathcal{D}(\gamma)\to \ps$ be a continuous curve on $\ps$ whose domain of definition $\mathcal{D}(\gamma)$ is a compact interval. For $n\ge 0$, there are $\ell^n$ curves $\tilde\gamma_i:\mathcal{D}(\gamma) \to \ps$, $1\le i\le \ell^n$, such that $T^n\circ \tilde{\gamma}_i=\gamma$, each of which is called a backward image of $\gamma$ by $T^n$. From the hyperbolic properties of $T$, we can choose positive constants $c_i$, $1\le i\le r$, so that the following holds: Let $\Omega$ be the set of $C^r$ curves $\gamma:\mathcal{D}(\gamma)\to \ps$ such that - the domain of definition $\mathcal{D}(\gamma)$ is a compact interval, - $\gamma$ is written in the form $\gamma(t)=(\pi\circ \gamma(t),t)$ and - $|d^i(\pi\circ \gamma)/dt^i(s)|\le c_i$ for $1\le i\le r$ and $s\in \mathcal{D}(\gamma)$ where $\pi:\ps\to S^1$ is the projection to the first component. Then each backward image $\tilde\gamma$ of any $\gamma\in \Omega$ by $T^n$ with $n\ge 1$ is the composition $\hat{\gamma}\circ g$ of a curve $\hat{\gamma}\in \Omega$ and a $C^r$ diffeomorphism $ g:\mathcal{D}(\gamma)\to \mathcal{D}(\hat{\gamma})$. Further, we can take a positive constant $c$ so that the diffeomorphism $ g$ always satisfies $$\label{eqn:distg} \left|\frac{d^\nu}{ds^\nu}( g^{-1})(s)\right|<c\lambda^{n}\quad\mbox{ for $s\in \mathcal{D}(\hat{\gamma})$ and $1\le \nu \le r$.}$$ We henceforth fix such $c$, $c_i$, $1\le i\le r$, and $\Omega$ as above. Moreover, the cone $$\label{eq:defC} \C=\{(u,v) {\ |\ }|u|\leq \alpha_0^{-1}|v|\}$$ is invariant under $DT^{-1}$, whence we can take $c_1=\alpha_0^{-1}$. Finally, increasing the constants $c_2,\dots,c_r$ if necessary, we can assume that, whenever $I$ is a segment in $S^1\times \R$ and $J$ is a component of $T^{-q}(I)$ such that its tangent vectors are all contained in $\C$, then $J$ is the image of an element of $\Omega$ (recall that $q$ is fixed once and for all until the end of Section \[sec:ly\]). For a function $h\in C^r(\ann)$ and an integer $0\le \rho\le r-1$, we define $$\|h\|^\ac_{\rho}:=\max_{{\alpha}+\beta\le \rho}\;\sup_{\gamma\in \Omega}\;\;\sup_{ \varphi\in \cC^{{\alpha}+\beta}(\gamma)} \int \varphi(t) \cdot \partial_x^{\alpha} \partial_y^\beta h(\gamma(t)) dt$$ where $\max_{{\alpha}+\beta\le \rho}$ denotes the maximum over pairs $({\alpha},\beta)$ of non-negative integers such that ${\alpha}+\beta\le \rho$ and $\cC^{s}(\gamma)$ denotes the space of $C^s$ functions $\varphi$ on $\R$ such that $\operatorname{supp}\varphi \subset \operatorname{Int}(\mathcal{D}(\gamma))$ and $\|\varphi\|_{C^s}\le 1$. This is a norm on $C^r(\ann)$. It satisfies $$\label{ac_greater_L1} {\left\| h \right\|}_{L^1} \leq C {\left\| h \right\|}_0^\ac\leq C{\left\| h \right\|}_\rho^\ac.$$ The following lemma is the main ingredient of this section. \[lm:P\] There exists a constant $A_0$ such that $$\begin{aligned} \label{eqn:GLLY1} &\|P^{n} h\|^\ac_{\rho}\le A_0 \ell^{-\rho n} \|h\|^\ac_{\rho} +C(n)\|h\|^\ac_{\rho-1}\quad\mbox{ for $1\le \rho\le r-1$}, \intertext{and}\label{eqn:GLLY2} &\|P^{n} h\|^\ac_{0}\le A_0 \|h\|^\ac_{0} \end{aligned}$$ for $n\ge 0$ and $h\in C^r(\ann)$, where $C(n)$ may depend on $n$ but not on $h$. Note that the iterate $T^n$ for $n\ge 0$ is locally written in the form $$\label{eqn:Tloc} T^n(x,y)=(\ell^nx, \lambda^n y+S(\ell^n x))$$ where $S$ is a $C^r$ function whose derivatives up to order $r$ are bounded by ${\alpha_0}$. Consider non-negative integers $\rho$, $\alpha$, $\beta$ satisfying $1\le \rho \le r-1$ and $\alpha+\beta=\rho$. Differentiating both sides of $$P^n h(x,y)=\frac{1}{\lambda^{n}\lap^n}\sum_{(x',y')\in T^{-n}(x,y)} h(x',y')$$ by using (\[eqn:Tloc\]), we see that the differential $\partial_x^{\alpha}\partial_y^\beta P^n h(x,y)$ can be written as the sum of $$\begin{aligned} \Phi(x,y)&=\sum_{(x',y')\in T^{-n}(x,y)} \sum_{k=0}^{\alpha}Q_k(x) \frac{\partial_x^{{\alpha}-k}\partial_y^{\beta+k} h(x',y')} {\lambda^{(1+\beta+k)n}\ell^{(1+\alpha-k)n}} \intertext{and} \Psi(x,y)&=\sum_{(x',y')\in T^{-n}(x,y)}\sum_{a+b\le \rho-1}Q_{a,b}(x)\frac{\partial_x^a \partial_y^b h(x',y')} {\lambda^{(1+b)n}\ell^{(1+a)n}} \end{aligned}$$ where $Q_k(\cdot)$ and $Q_{a,b}(\cdot)$ are functions of class $C^\rho$ and $C^{a+b}$ respectively.[^4] It is easy to check that the $C^\rho$ norm of $Q_k(\cdot)$ and $C^{a+b}$ norm of $Q_{a,b}(\cdot)$ are bounded by some constant. For $\gamma\in \Omega$ and $\varphi\in \cC^\rho(\gamma)$, we estimate $$\label{eqn:p} \int\varphi(t)\partial_x^{\alpha}\partial_y^\beta P^n h(\gamma(t)) dt =\int\varphi(t)\Phi(\gamma(t)) dt+\int\varphi(t)\Psi(\gamma(t)) dt.$$ Let $\gamma_i$, $1\le i\le \ell^n$, be the backward images of the curve $\gamma$ by $T^n$ and write them as the composition $\hat{\gamma}_i\circ g_i$ of $\hat{\gamma}_i\in \Omega$ and a $C^r$ diffeomorphism $ g_i$. Then we have $$\begin{aligned} &\int \varphi(t) \Psi(\gamma(t)) dt = \sum_{1\le i\le \ell^n}\sum_{a+b\le \rho-1} \int \varphi(t) \frac{Q_{a,b}(\pi\circ\gamma(t))\cdot \partial_x^a \partial_y^b h(\gamma_i(t))} {\lambda^{(1+b)n}\ell^{(1+a)n}}dt\\ &= \sum_{1\le i\le \ell^n}\sum_{a+b\le \rho-1} \int \frac{\varphi( g_i^{-1}(s))\cdot Q_{a,b}(\pi\circ\gamma\circ g_i^{-1}(s))\cdot ( g_i^{-1})'(s)\cdot \partial_x^a \partial_y^b h(\hat{\gamma}_i(s))}{\lambda^{(1+b)n}\ell^{(1+a)n} } ds. \end{aligned}$$ Since the $C^{a+b}$ norm of the function $ s\mapsto \varphi( g_i^{-1}(s))\cdot Q_{a,b}(\pi\circ\gamma\circ g_i^{-1}(s))\cdot ( g_i^{-1})'(s)$ is bounded by some constant (depending on $n$) from (\[eqn:distg\]), we have $$\label{eqn:Psi} \left|\int \varphi(t) \Psi(\gamma(t)) dt \right|\le C(n) \|h\|^\ac_{\rho-1}$$ where $C(n)$ may depend on $n$ but not on $h$. The first integral on the right hand side of (\[eqn:p\]) is written as $$\begin{aligned} &\int \varphi(t) \Phi(\gamma(t)) dt =\sum_{1\le i\le \ell^n} \sum_{k=0}^{{\alpha}}\int \varphi(t) \frac{ Q_k(\pi\circ \gamma_i(t))\cdot \partial_x^{{\alpha}-k}\partial_y^{\beta+k} h(\gamma_i(t))} {\lambda^{(1+\beta+k)n}\ell^{(1+{\alpha}-k)n}} dt\\ &=\sum_{1\le i\le \ell^n} \sum_{k=0}^{{\alpha}}\int \frac{ \varphi( g_i^{-1}(s))\cdot Q_k(\pi\circ \gamma\circ g_i^{-1}(s))\cdot ( g_i^{-1})'(s) \cdot \partial_x^{{\alpha}-k}\partial_y^{\beta+k} h(\hat{\gamma}_i(s))} {\lambda^{(1+\beta+k)n}\ell^{(1+{\alpha}-k)n}} ds. \end{aligned}$$ For a while, we fix $1\le i\le \ell^n$. Since $$\begin{aligned} \frac{d}{ds}&(\partial_x^{{\alpha}-k}\partial_y^{\beta+k-1} h(\hat{\gamma}_i(s)))\\ &=(\pi\circ \hat{\gamma}_i)'(s) \cdot \partial_x^{{\alpha}-k+1}\partial_y^{\beta+k-1} h(\hat{\gamma}_i(s)) +\partial_x^{{\alpha}-k}\partial_y^{\beta+k} h(\hat{\gamma}_i(s)), \end{aligned}$$ integration by part yields, for any $\psi\in C^\rho(\mathcal{D}(\hat\gamma_i))$, $$\begin{aligned} \int \frac{d\psi}{ds}(s)\cdot \frac{ \partial_x^{{\alpha}-k}\partial_y^{\beta+k-1} h(\hat{\gamma}_i(s))} {\lambda^{(1+\beta+k)n}\ell^{(1+{\alpha}-k)n}} ds &= - \int \tilde{\psi}(s) \frac{\partial_x^{{\alpha}-k+1}\partial_y^{\beta+k-1} h(\hat{\gamma}_i(s)) } {\lambda^{(1+\beta+k-1)n}\ell^{(1+{\alpha}-k+1)n}}ds\\ &\qquad\qquad - \int \psi(s) \frac{\partial_x^{{\alpha}-k}\partial_y^{\beta+k} h(\hat{\gamma}_i(s)) } {\lambda^{(1+\beta+k)n}\ell^{(1+{\alpha}-k)n}}ds \end{aligned}$$ where $ \tilde{\psi}(s)=\lambda^{-n}\ell^n(\pi\circ \hat\gamma_i)'(s) \psi(s)$. This implies $$\label{eqn:rec} \begin{aligned} \left|\int \psi(s) \frac{ \partial_x^{{\alpha}-k}\partial_y^{\beta+k} h(\hat{\gamma}_i(s))} {\lambda^{(1+\beta+k)n}\ell^{(1+{\alpha}-k)n}} ds\right| &\le \left|\int \tilde{\psi}(s) \frac{\partial_x^{{\alpha}-k+1}\partial_y^{\beta+k-1} h(\hat{\gamma}_i(s)) } {\lambda^{(1+\beta+k-1)n}\ell^{(1+{\alpha}-k+1)n}}ds\right|\\ &\qquad\qquad\qquad +C(n)\|\psi\|_{C^{\rho}}\|h\|^\ac_{\rho-1} \end{aligned}$$ where $C(n)$ may depend on $n$ but not on $h$ nor $\psi$. Put $$\psi_0(s)=\varphi( g_i^{-1}(s))\cdot {Q_k(\pi\circ\gamma\circ g_i^{-1}(s))}\cdot{( g_i^{-1})'(s)}.$$ By using the last inequality repeatedly, we obtain $$\begin{aligned} \left|\int \frac {\psi_0(s)\partial_x^{\alpha-k}\partial_y^{\beta+k} h(\hat{\gamma}_i(s))} {\lambda^{(1+\beta+k)n}\ell^{(1+\alpha-k)n}} ds\right|&\le \left|\int \frac{\psi_{\beta+k}(s)\partial_x^{\rho}h(\hat{\gamma}_i(s))} {\lambda^n \ell^{(1+\rho)n}} ds\right|\\ &\qquad\qquad +\sum_{j=0}^{\beta+k-1}C(n){\left\| \psi_{j} \right\|}_{C^{\rho}}\|h\|^\ac_{\rho-1} \end{aligned}$$ where $ \psi_j(s)=\lambda^{-nj}\ell^{nj}((\pi\circ \hat\gamma_i)'(s))^j \psi_0(s) =\lambda^{-nj} ((\pi\circ \gamma\circ g_i^{-1})'(s))^j \psi_0(s)$. Since $\|\psi_j\|_{C^\rho}<C_0\lambda^n$ for $0\le j\le \beta+k$ for some constant $C_0$ from (\[eqn:distg\]), we get $$\begin{aligned} \left|\int \psi_0(s) \frac {\partial_x^{\alpha-k}\partial_y^{\beta+k} h(\hat{\gamma}_i(s))} {\lambda^{(1+\beta+k)n}\ell^{(1+\alpha-k)n}} dt\right|&\le C_0 \ell^{-(1+\rho)n}\|h\|^\ac_\rho +C(n)\|h\|^\ac_{\rho-1}. \end{aligned}$$ Summing up this inequality for $\gamma_i$, $1\le i\le \ell^n$, we obtain $$\left|\int\varphi(t)\Phi(\gamma(t)) dt \right| \le C_0 \ell^{-\rho n}\|h\|^\ac_\rho +C(n)\|h\|^\ac_{\rho-1}$$ for some constant $C_0$. This and (\[eqn:Psi\]) give (\[eqn:GLLY1\]). The proof of (\[eqn:GLLY2\]) is obtained in a similar but much simpler manner. Main Lasota-Yorke inequality {#sec:ly} ============================ In this section, we prove the following proposition. \[prop:mainLY\] There exists a constant $B_0$ independent of $q$ and a constant $C(q)$ such that, for all $\varphi\in C^{r}(\ann)$, for all integer ${\rho_0}$ with $s+1<{\rho_0}\leq r-1$, $${\left\| P^q \varphi \right\|}_{W^s}^2 \leq \frac{B_0 \e(q)}{(\lambda^{1+2s} \ell)^q} {\left\| \varphi \right\|}_{W^s}^2 + C(q) {\left\| \varphi \right\|}_{W^s} {\left\| \varphi \right\|}_{{\rho_0}}^\ac.$$ First of all, we introduce some notation and prove some elementary facts concerning the Sobolev norm ${\left\| \cdot \right\|}_{W^s}$. The Fourier transform $\F\varphi$ of $\varphi\in C^{r}(\ann)$ is a function on $\Z\times \R$ defined by $$\F\varphi(\xi,\eta)=\frac{1}{\sqrt{2\pi}}\int_{\ps} \varphi(x,y)\exp\left(-{\mathbf{i}}(2\pi \xi x+\eta y)\right) dx dy.$$ For $s\ge 0$ and for $\varphi_1, \varphi_2\in C^{r}(\ann)$, we define $$(\varphi_1, \varphi_2)_{W^s}:=(\varphi_1, \varphi_2)^*_{W^s}+(\varphi_1, \varphi_2)_{L^2}$$ where $$(\varphi_1, \varphi_2)_{W^s}^*:= \sum_{\xi=-\infty}^{\infty}\int_{\R}\F \varphi_1 (\xi, \eta) \cdot \overline{\F\varphi_2 (\xi, \eta)}\cdot ((2\pi\xi)^2+\eta^2)^{s} d\eta.$$ The Sobolev norm is defined by $\|\varphi\|_{W^s}=\sqrt{(\varphi, \varphi)_{W^s}}$. Note that we have $$\label{eqn:Sob1} (\varphi_1, \varphi_2)_{W^s}^*= \sum_{\alpha+\beta= [s]} b_{{\alpha}\beta} (\partial_x^{\alpha}\partial_y^\beta\varphi_1, \partial_x^{\alpha}\partial_y^\beta\varphi_2)^*_{W^{s-[s]}}$$ where $b_{{\alpha}\beta}$ are positive integers satisfying $ (X^2+Y^2)^{[s]}=\sum_{{\alpha},\beta}b_{\alpha\beta} X^{2\alpha}Y^{2\beta}$. Especially, if $s$ is an integer, we have $$\label{eqn:Sob2} (\varphi_1, \varphi_2)_{W^s}^*= \sum_{\alpha+\beta= s}b_{{\alpha}\beta}\int_{\ps} \partial_x^{\alpha}\partial_y^\beta\varphi_1 (x,y) \cdot \overline {\partial_x^{\alpha}\partial_y^\beta\varphi_2 (x,y)} dxdy.$$ In case $s$ is not an integer, we shall use the following formula ([@H pp 240]): there exists a constant $B>0$ that depends only on $0<\sigma<1$ such that $$\begin{aligned} \label{eqn:n} &(\varphi_1, \varphi_2)_{W^\sigma}^*=\\ &B\int_{\ps}\!\!\!\!\!dxdy \int_{\R^2} \frac{(\varphi_1(x+u,y+v)-\varphi_1(x,y))\overline {(\varphi_2(x+u,y+v)- \varphi_2(x,y))}}{(u^2+v^2)^{1+\sigma}}dudv. \notag \end{aligned}$$ \[lm:Sob\] [(1)]{} For $0\le t<s\le r$ and $\epsilon>0$, there is a constant $C(\epsilon,t,s)$ such that $$\label{eqn:ts} \|\varphi\|_{W^t}^2 \leq \epsilon\|\varphi\|_{W^s}^2+C(\epsilon, t,s)\|\varphi\|_{L^1}^2\quad \mbox{for $\varphi\in C^r(\ann)$.}$$ [(2)]{} For $\epsilon>0$, there exists a constant $C(\epsilon,s)$ with the following property: if the supports of functions $\varphi_1, \varphi_2\in C^r(\ann)$ are disjoint and the distance between them is greater than $\epsilon$, it holds $$\label{eqn:disj} \bigl| (\varphi_1, \varphi_2)_{W^s}\bigr|\le C(\epsilon,s)\|\varphi_1\|_{L^1}\|\varphi_2\|_{L^1}.$$ \(1) follows from the definition of the norm and the fact $\|\F\varphi\|_{L^\infty}\le \|\varphi\|_{L^1}$. If $s$ is an integer, (2) is trivial since $(\varphi_1, \varphi_2)_{s}=0$ by (\[eqn:Sob2\]). Suppose that $s$ is not an integer. Using (\[eqn:Sob1\]) and (\[eqn:n\]) with the assumption on the disjointness of the supports and changing variables, we can rewrite $(\varphi_1, \varphi_2)_{W^s}^*$ as $$-2B\sum_{\alpha+\beta= [s]}\int_{\ps}dxdy\int_{\R^2}\frac{b_{\alpha\beta}\cdot \partial_x^\alpha\partial_y^\beta\varphi_1(x+u, y+v)\cdot \overline {\partial_x^\alpha\partial_y^\beta\varphi_2(x,y)}} {(u^2+v^2)^{1+\sigma}}dudv$$ where $\sigma=s-[s]$. Integrating $[s]$ times by part on $(u,v)$, then changing variables and integrating again $[s]$ times by part, we obtain $$(\varphi_1, \varphi_2)_{W^s}^*=\int_{\ps}dxdy\int_{\R^2}\frac{\varphi_1(x+u, y+v) \overline {\varphi_2(x,y)} \tilde{B}(u,v)}{(u^2+v^2)^{1+\sigma+2[s]}}dudv$$ where $\tilde{B}(u,v)$ is a polynomial of $u$ and $v$ of order $2[s]$. With this and the assumption, we can conclude the inequality in (2). The norm ${\left\| \cdot \right\|}^\ac$ will be used through the following lemma. Let $\C^*$ be the cone in $\R^2$ defined by $$\C^*=\{(\xi,\eta)\in \R^2{\ |\ }|\eta|\le {\alpha_0}^{-1}|\xi|\},$$ so that $DT^*_{\x}(\C^*)\subset \C^*$ for $\x\in \ps$. \[lm:F\] Let ${\rho_0}$ be an integer with $s+1<{\rho_0}\leq r-1$. Let $\a$ and $\c$ elements of $\sym^q$ and $\sym^p$ respectively, and $\chi:\ps\to \R$ a $C^\infty$ function supported on $\P_*(\c\a)\times \R$. Take $(\xi,\eta)\in \Z\times\R\backslash\{(0,0)\}$ such that, for any $\x\in \P_*(\c\a)\times \R$, $(DT^q_{\x})^*(\xi,\eta)\in \C^*$. Then, for any $\varphi \in C^r$, $$\label{eqn:F} \left|(\xi^2+\eta^2)^{{\rho_0}/2}\F (P^q(\chi\cdot \varphi))(\xi, \eta)\right|\leq C(q,\chi) {\left\| \varphi \right\|}_{{\rho_0}}^\ac,$$ where $C(q,\chi)$ may depend on $q$ and $\chi$. Let $(\xi, \eta)$ be a vector satisfying the assumption. Let $\Gamma$ be the set of line segments on $\ps$ that are the intersection of a line normal to $(\xi, \eta)$ with the region $\P_*(\c)\times\R$. We parametrize the segments in $\Gamma$ by length. Since the support of $P^q(\chi\cdot \varphi)$ is contained in $\ann\cap (\P_*(\c)\times\R)$, the left hand side of is bounded by some constant multiple of $$\label{eqn:curve} \sup_{\gamma\in \Gamma} \int_{\gamma} \partial^{{\rho_0}}P^q(\chi\cdot \varphi) dt$$ where $\partial$ is partial derivative with respect to $x$ if $|\xi|>|\eta|$ and that with respect to $y$ otherwise. For each $\gamma\in \Gamma$, there exists a unique backward image $\tilde\gamma$ of $T^q$ that is contained in $\P_*(\c\a) \times \R$. If $\x\in \tilde \gamma$ and $u$ is tangent to $\gamma$ at $T^q(\x)$, then $$0= \langle u, (\xi,\eta) \rangle = \langle (DT^q_{\x})^{-1}u, (DT^q_{\x})^*(\xi,\eta) \rangle.$$ By assumption, $(DT^q_{\x})^*(\xi,\eta) \in \C^*$, whence $(DT^q_{\x})^{-1}u \in \C$ (by definition of $\C$). Hence, $\tilde\gamma$ is the composition $\hat\gamma\circ \psi$ of an element $\hat\gamma$ of $\Omega$ and a $C^r$ diffeomorphism $\psi$. By obvious estimates on the distortion of $T^m$ for $0\le m \le q$ and by the definition of the norm $\|\cdot\|^\ac_{{\rho_0}}$, we get that is bounded by $C {\left\| \varphi \right\|}_{{\rho_0}}^\ac$. Let $\{\chi_{\c}:S^1\to \R\}_{\c\in \sym^p}$ be a $C^\infty$ partition of unity subordinate to the covering $\{\operatorname{Int}\P_*(\c)\}_{\c\in \sym^p}$, whence $\operatorname{supp}(\chi_{\c})\subset \operatorname{Int}\P_*(\c)$. Define a function $\chi_{\c\a}$ by $\chi_{\c\a}(\tau_{\c,\a}^{-q}x) =\chi_{\c}(x)$ if $x\in \P_*(\c)$, and extend it by $0$ elsewhere. Then the functions $\chi_{\c\a}$ for $(\a,\c)\in \sym^q\times\sym^p$ are again a $C^\infty$ partition of unity. To keep the notation simple, we will still use $\chi_\c$ and $\chi_{\c\a}$ to denote $\chi_\c \circ \pi$ and $\chi_{\c\a} \circ \pi$. \[lm:sum\] There is a constant $C>0$ such that, for any $\varphi\in C^r(\ann)$, it holds $$\label{eqn:sum1} \sum_{(\a,\c)\in \sym^q\times\sym^p}\|\chi_{\c\a}\varphi\|_{W^s}^2\le 2\|\varphi\|_{W^s}^2+C\|\varphi\|_{L^1}^2$$ and $$\label{eqn:sum2} \|\varphi\|_{W^s}^2\le 7\sum_{\c\in \sym^p}\|\chi_{\c}\varphi\|_{W^s}^2 +C\|\varphi\|_{L^1}^2.$$ Since the claims are obvious when $s=0$, we assume $s>0$. Let $t$ be the largest integer that is (strictly) less than $s$. Then for every $\epsilon>0$ we have $$\sum_{(\a,\c)\in \sym^q\times\sym^p}\|\chi_{\c\a}\varphi\|_{W^s}^2\le (1+\epsilon)\|\varphi\|_{W^s}^2+C(\epsilon)\|\varphi\|_{W^{t}}^2.$$ Indeed, we can check this by using (\[eqn:Sob2\]) if $s$ is an integer and by using (\[eqn:Sob1\]) and (\[eqn:n\]) instead of (\[eqn:Sob2\]) otherwise. Hence (\[eqn:sum1\]) follows from lemma \[lm:Sob\](1). From lemma \[lm:Sob\](2), we have $(\chi_{\c}\varphi, \chi_{\c'}\varphi)_{W^s}\le C\|\chi_{\c}\varphi\|_{L^1}\| \chi_{\c'}\varphi\|_{L^1}\le C\|\varphi\|_{L^1}^2$ for some constant $C>0$ if the closures of $\P_*(\c)$ and $\P_*(\c')$ do not intersect. Also we have $(\chi_{\c}\varphi, \chi_{\c'}\varphi)_{W^s}\le (\|\chi_{\c}\varphi\|_{W^s}^2 +\|\chi_{\c'}\varphi\|_{W^s}^2)/2$ in general. Applying these to $\|\varphi\|_{W^s}^2=\sum_{(\c,\c')\in \sym^p\times\sym^p}(\chi_{\c}\varphi, \chi_{\c'}\varphi)_{W^s}$, we obtain (\[eqn:sum2\]). We start the proof of Proposition \[prop:mainLY\]. &gt;From , we have $$\begin{aligned} \|P^q(\varphi)\|_{W^s}^2&\le 7\sum_{\c\in \sym^p}\|\chi_{\c}P^q(\varphi)\|_{W^s}^2+C{\left\| \varphi \right\|}_{L^1}^2 \\& \le 7\sum_{\c\in \sym^p}\left\|\sum_{\a\in \sym^q} P^q(\chi_{\c\a}\varphi)\right\|_{W^s}^2+C{\left\| \varphi \right\|}_{L^1}^2. \end{aligned}$$ So we will estimate $$\left\|\sum_{\a\in \sym^q} P^q(\chi_{\c\a}\varphi)\right\|_{W^s}^2= \sum_{(\a, \b)\in \sym^q\times \sym^q}(P^q(\chi_{\c\a}\varphi), P^q(\chi_{\c\b}\varphi))_{W^s}$$ for $\c\in \sym^p$. Consider first a pair $(\a, \b)\in \sym^q\times \sym^q$ such that $\a\pitchfork_{\c}\b$. For any $(\xi,\eta)\in \Z\times \R\backslash \{(0,0)\}$, this implies that either $(DT_{\x}^q)^*(\xi,\eta)\in \C^*$ for all $\x\in \P_*(\c\a)\times \R$, or $(DT_{\x}^q)^*(\xi,\eta)\in \C^*$ for all $\x\in \P_*(\c\b)\times \R$. Let $U$ be the set of all $(\xi,\eta)\in \Z\times\R$ such that the first possibility holds, and $V=(\Z\times\R) \backslash U$. If $(\xi,\eta)\in U$, by Lemma \[lm:F\], there exists a constant $C>0$ such that $\left|\F P^q(\chi_{\c\a}\varphi)(\xi,\eta)\right| \leq C (\xi^2+\eta^2)^{-{\rho_0}/2} {\left\| \varphi \right\|}_{{\rho_0}}^\ac$. Moreover, $\left| \F P^q(\chi_{\c\a}\varphi)(\xi,\eta) \right| \leq C {\left\| \varphi \right\|}_{L^1}$, which is bounded by $C{\left\| \varphi \right\|}_{{\rho_0}}^\ac$ by . Hence, $\left|\F P^q(\chi_{\c\a}\varphi)(\xi,\eta)\right| \leq C (1+\xi^2+\eta^2)^{-{\rho_0}/2} {\left\| \varphi \right\|}_{{\rho_0}}^\ac$. So we have, for some constant $C$, $$\begin{aligned} &\left|\; \sum_{\xi=-\infty}^{\infty} \int \mathbf{1}_U(\xi,\eta)\cdot (1+\xi^2+\eta^2)^s \F P^q (\chi_{\c\a} \varphi) \cdot \overline{ \F P^q (\chi_{\c\b} \varphi)}\; d\eta\;\right| \\ &\quad \leq C \left(\sum_{\xi=-\infty}^{\infty} \int \mathbf{1}_U(\xi,\eta)\cdot (1+\xi^2+\eta^2)^s | \F P^q (\chi_{\c\a} \varphi) |^2 d\eta \right)^{1/2} {\left\| P^q(\chi_{\c\b} \varphi) \right\|}_{W^s} \\&\quad \leq C {\left\| \varphi \right\|}_{{\rho_0}}^\ac {\left\| \varphi \right\|}_{W^s}, \end{aligned}$$ since the function $(1+\xi^2+\eta^2)^{-{\rho_0}+s}$ is integrable by the assumption $s<{\rho_0}-1$. The same inequality holds on $V$, and we obtain $$\label{eqn:c1} \left|(P^q(\chi_{\c\a}\varphi), P^q(\chi_{\c\b}\varphi))_{W^s}\right| \leq C {\left\| \varphi \right\|}_{{\rho_0}}^\ac {\left\| \varphi \right\|}_{W^s}.$$ For the sum over $\a$ and $\b$ such that $\a\not\pitchfork_\c \b$, we have $$\begin{aligned} \sum_{\a\not\pitchfork_\c \b}(P^q(\chi_{\c\a}\varphi), P^q(\chi_{\c\b}\varphi))_{W^s} &\le \sum_{\a\not\pitchfork_\c \b} \frac{\|P^q(\chi_{\c\a}\varphi)\|^2_{W^s}+\|P^q(\chi_{\c\b} \varphi)\|^2_{W^s}}{2}\notag\\ &\le \e(q) \sum_{\a\in \sym^q} \|P^q(\chi_{\c\a}\varphi)\|^2_{W^s}.\label{eqn:c2} \end{aligned}$$ For the terms in the last sum, we have the estimate $$\label{eqn:c3} \|P^q(\chi_{\c\a}\varphi)\|^2_{W^s}\le \frac{C_0\|\chi_{\c\a}\varphi\|_{W^s}^2}{\lambda^{(1+2s)q}\ell^{q}} +C{\left\| \varphi \right\|}_{L^1}^2$$ where $C_0$ is a constant that depends only on $\lambda$, $\ell$ and $\kappa$. Indeed, we can check this by using and if $s$ is an integer and by using and instead of otherwise. &gt;From , , , and , we obtain $$\begin{aligned} \sum_{\c\in \sym^p} \left\|\sum_{\a\in \sym^q} P^q(\chi_{\c\a}\varphi)\right\|_{W^s}^2 & \le \frac{C_0\e(q)}{\lambda^{(1+2s)q}\ell^q} \sum_{(\a,\c)\in \sym^q\times \sym^p} \|\chi_{\c\a}\varphi\|_{W^s}^2 + C {\left\| \varphi \right\|}_{W^s} {\left\| \varphi \right\|}_{{\rho_0}}^\ac \\& \le \frac{2C_0\e(q)}{\lambda^{(1+2s)q}\ell^q} {\left\| \varphi \right\|}_{W^s}^2 + C {\left\| \varphi \right\|}_{W^s} {\left\| \varphi \right\|}_{{\rho_0}}^\ac, \end{aligned}$$ and hence Proposition \[prop:mainLY\]. Proof of the main theorems {#sec:proofs_thms} ========================== We will use Lemma \[lm:P\] and Proposition \[prop:mainLY\] to study the properties of $P$ acting on the space $C^r(D)$ equipped with the norms ${\left\| \cdot \right\|}_{{\rho_0}}^\ac$ and ${\left\| \cdot \right\|}_{W^s}$. \[lem:delta1\] Let $\delta\in (\ell^{-1},1)$. There exists $C>0$ such that, for integer $1\leq \rho \leq r-1$, for $n\in {\mathbb{N}}$, $${\left\| P^n h \right\|}_\rho^\ac \leq C \delta^{\rho n} {\left\| h \right\|}_\rho^\ac + C {\left\| h \right\|}_{\rho-1}^\ac.$$ We prove it by induction on $\rho$. Let $\rho\geq 1$. By Lemma \[lm:P\], there exists $N\in {\mathbb{N}}$ and $C>0$ such that $$\label{eq:iter} {\left\| P^N h \right\|}_{\rho}^\ac \leq \delta^{\rho N} {\left\| h \right\|}_{\rho}^\ac+ C {\left\| h \right\|}_{\rho-1}^\ac.$$ By the inductive assumption (and Lemma \[lm:P\] in the $\rho=1$ case), ${\left\| P^n h \right\|}_{\rho-1}^\ac \leq C {\left\| h \right\|}_{\rho-1}^\ac$. Hence, iterating gives the conclusion. \[lem:delta2\] Let $\delta\in (\ell^{-1},1)$, and let $0 \leq {\rho_1}<{\rho_0}\leq r-1$ be integers. Let $\nu({\rho_0},{\rho_1})=\sum_{j={\rho_1}+1}^{{\rho_0}} \frac{1}{j}$. There exists $C>0$ such that, for $n\in {\mathbb{N}}$, $${\left\| P^n h \right\|}_{{\rho_0}}^\ac \leq C \delta^{n/\nu({\rho_0},{\rho_1})} {\left\| h \right\|}_{{\rho_0}}^\ac + C {\left\| h \right\|}_{{\rho_1}}^\ac.$$ Let $n$ be a multiple of $(r-1)!$. Then Lemma \[lem:delta1\] implies by induction over ${\rho_1}+1 \leq \rho\leq {\rho_0}$ that $${\left\| P^{(\frac{1}{\rho}+\dots+\frac {1} {{\rho_1}+1})n} h \right\|}_{\rho}^\ac \leq C \delta^n {\left\| h \right\|}_{\rho}^\ac +C {\left\| h \right\|}_{{\rho_1}}^\ac.$$ For $\rho={\rho_0}$, we obtain ${\left\| P^{\nu({\rho_0},{\rho_1}) n}h \right\|}_{{\rho_0}}^\ac \leq C \delta^n {\left\| h \right\|}_{{\rho_0}}^\ac + C {\left\| h \right\|}_{{\rho_1}}^\ac$. \[Theorem:LY\] Assume that $\frac{B_0 \e(q)}{(\lambda^{1+2s}\ell)^q} <1$. Let $0 \leq {\rho_1}<{\rho_0}\leq r-1$ be integers with $s<{\rho_0}-1$, and let $\nu=\nu({\rho_0},{\rho_1})$ be as in the previous lemma. Let $$\gamma \in \left( \max\left( \ell^{-1/\nu}, \sqrt{\frac{(B_0 \e(q))^{1/q}}{\lambda^{1+2s}\ell}} \right), 1 \right).$$ Let ${\left\| \varphi \right\|}:= {\left\| \varphi \right\|}_{W^s}+{\left\| \varphi \right\|}_{{\rho_0}}^\ac$. There exists a constant $C$ such that, for all $n\in {\mathbb{N}}$, $${\left\| P^n \varphi \right\|} \leq C \gamma^n {\left\| \varphi \right\|} + C {\left\| \varphi \right\|}_{{\rho_1}}^\ac.$$ Since $\sqrt{a+b}\leq \sqrt{a}+\sqrt{b}$ and $\sqrt{ab} \leq \epsilon a +\epsilon^{-1} b$, Proposition \[prop:mainLY\] implies $${\left\| P^q \varphi \right\|}_{W^s} \leq \left(\frac{(B_0 \e(q))^{1/q}}{\lambda^{1+2s}\ell}\right)^{q/2}{\left\| \varphi \right\|}_{W^s} + \epsilon {\left\| \varphi \right\|}_{W^s}+ C(\epsilon) {\left\| \varphi \right\|}_{{\rho_0}}^\ac.$$ Since $\left(\frac{(B_0 \e(q))^{1/q}}{\lambda^{1+2s}\ell}\right)^{q/2}< \gamma^q$, taking $\epsilon$ small enough yields $${\left\| P^q \varphi \right\|}_{W^s} \leq \gamma^q {\left\| \varphi \right\|}_{W^s}+C {\left\| \varphi \right\|}_{{\rho_0}}^\ac.$$ Iterating this equation $K$ times gives $$\label{eq:sum1} {\left\| P^{Kq} \varphi \right\|}_{W^s} \leq \gamma^{Kq} {\left\| \varphi \right\|}_{W^s} + C(K) {\left\| \varphi \right\|}_{{\rho_0}}^\ac,$$ for some constant $C(K)$. If $K$ is large enough, the choice of $\gamma$ and Lemma \[lem:delta2\] also yield $$\label{eq:sum2} {\left\| P^{Kq} \varphi \right\|}_{{\rho_0}}^\ac \leq \frac{\gamma^{Kq}}{2} {\left\| \varphi \right\|}_{{\rho_0}}^\ac + C'(K) {\left\| \varphi \right\|}_{{\rho_1}}^\ac.$$ Fix such a $K$, and define a norm ${\left\| \varphi \right\|}^*:= {\left\| \varphi \right\|}_{W^s} + 2 C(K)\gamma^{-Kq} {\left\| \varphi \right\|}_{{\rho_0}}^\ac$. Adding and gives $${\left\| P^{Kq} \varphi \right\|}^* \leq \gamma^{Kq} {\left\| \varphi \right\|}^* + C {\left\| \varphi \right\|}_{{\rho_1}}^\ac.$$ Iterating this equation (and remembering ${\left\| P^n \varphi \right\|}_{{\rho_1}}^\ac \leq C {\left\| \varphi \right\|}_{{\rho_1}}^\ac$ for some constant $C$ independent of $n$, by Lemma \[lem:delta1\]), we obtain the conclusion of the theorem for the norm ${\left\| \cdot \right\|}^*$. Since it is equivalent to the original norm ${\left\| \cdot \right\|}$, this concludes the proof. \[cor:main1\] If $B_0 \e(q) < (\lambda^{1+2s}\ell)^q$, the conclusion of Theorem \[th:main1\] holds for the transformation $T$. Take ${\rho_0}=r-1$ and ${\rho_1}=0$. They satisfy the assumptions of Theorem \[Theorem:LY\] since $s<r-2$. We fix a non-negative function $\Psi_0\in C^r(\ann)$ such that $\int \Psi_0 d\vol=1$. Put $\nu_0=\Psi_0\cdot \vol$ and $\Psi_n=P^n\Psi_0$ for $n\ge 1$. From , the density of $T^n_*\nu_0$ is $\Psi_n$. As the sequence $T^n_*\nu_0$ converges to the SBR measure $\mu$ for $T$ weakly, we have $$\label{eqn:wc} \lim_{n\to \infty}\left(\Psi_n,\varphi\right)_{L^2}=\int \varphi d\mu$$ for any continuous function $\varphi$ on $\ps$ with compact support. By Theorem \[Theorem:LY\], the sequence $\Psi_n$ for $n\ge 1$ is bounded with respect to the norm ${\left\| \cdot \right\|}$, hence also for the norm $\|\cdot \|_{W^s}$. Then there is a subsequence $n(i)\to \infty$ such that $\Psi_{n(i)}$ converges weakly to some element $\Psi_\infty$ in the Hilbert space $W^s(\ps)$. This and imply $ \int \Psi_\infty \varphi d\vol=\int \varphi d\mu$ for any continuous function $\varphi$ on $\ps$ with compact support. Thereby the density of the SBR measure $\mu$ is $\Psi_\infty\in W^s(\ps)$. \[cor:main2\] Let $1/2<s<r-2$. Assume that $\frac{B_0 \e(q)}{(\lambda^{1+2s}\ell)^q} <1$. If $$\gamma \in \left( \sqrt{\frac{(B_0 \e(q))^{1/q}}{\lambda^{1+2s}\ell}}, 1 \right),$$ the conclusion of Theorem \[th:main2\] holds for the transformation $T$ and this $\gamma$. Let ${\rho_0}$ be the smallest integer such that $s<{\rho_0}-1$, and ${\rho_1}$ the largest integer such that ${\rho_1}<s-1/2$. They satisfy the assumptions of Theorem \[Theorem:LY\]. Moreover, $\nu({\rho_0},{\rho_1})\leq 1+\frac{1}{2}+\frac{1}{3}<2$. Hence, $\ell^{-1/\nu} < \frac{1}{\sqrt{\ell}}< \sqrt{\frac{(B_0 \e(q))^{1/q}}{\lambda^{1+2s}\ell}}$. Let ${\mathcal{B}}$ be the completion of $C^r(\ann)$ with respect to the norm ${\left\| \cdot \right\|}$. It is a Banach space included in $W^s(D)$ and containing $C^{r-1}(D)$. Theorem \[Theorem:LY\] gives a Lasota-Yorke inequality between ${\mathcal{B}}$ and the space ${\mathcal{B}}'$ obtained by completing $C^r(\ann)$ for the norm ${\left\| \cdot \right\|}_{{\rho_1}}^\ac$. Hence, the result is a standard consequence of Hennion’s Theorem [@He], if we can prove that the unit ball of ${\mathcal{B}}$ is relatively compact in ${\mathcal{B}}'$. The embedding of ${\mathcal{B}}$ in $W^s(\ann)$ is continuous. Let $t\in ({\rho_1}+1/2,s)$. The embedding of $W^s(\ann)$ in $W^t(\ann)$ is compact by Sobolev’s embedding theorem. To conclude, it is sufficient to check that the injection $W^t(\ann) \to {\mathcal{B}}'$ is continuous. Since $t>{\rho_1}+1/2$, [@A Theorem 7.58 (iii)] (applied with $p=q=2$ , $k=1$ and $n=2$) proves that, for any smooth curve ${\mathcal{C}}\subset \ann$, for any $\varphi\in W^t(\ann)$, $${\left\| \partial^\alpha_x \partial^\beta_y \varphi \right\|}_{L^2({\mathcal{C}})} \leq C({\mathcal{C}}) {\left\| \varphi \right\|}_{W^t(\ann)}$$ whenever $\alpha$ and $\beta$ are non-negative integers satisfying $\alpha+\beta \leq {\rho_1}$. The constant $C({\mathcal{C}})$ can be chosen uniformly over all curves of $\Omega$, and we obtain ${\left\| \varphi \right\|}_{{\rho_1}}^\ac \leq C {\left\| \varphi \right\|}_{W^t(\ann)}$. For $\beta>0, \kappa>0$ and $\lambda\in (0,1)$, let $$\mathcal{E}(\beta,\kappa, \lambda)=\left\{f \in \mathcal{U}_\kappa \;;\; \limsup_{q\to\infty} \frac 1 q \log \e(q)> \beta\right\}.$$ Note that this definition depends on $\kappa$ and $\lambda$ through $\e(q)$, since $\e(q)$ is defined in terms of $\alpha_0=\kappa/(1-\lambda)$. Since the quantity $\e(q)$ depends on $f\in \mathcal{U}_\kappa$ upper semi-continuously and since we can take arbitrarily large $\kappa$ in the beginning, Theorems \[th:main1\], \[th:main2\] and \[th:main\] follow from Corollaries \[cor:main1\] and \[cor:main2\] and the next proposition. \[pp:transv\] For any $\beta>0$ and $\lambda>0$, there is a finite collection of $C^{\infty}$ functions $\varphi_{i}:S^{1}\to \R$, $i=1,2,\cdots, m$ and a constant $D_0>0$ such that, for any $\kappa>D_0$ and any $C^{r}$ function $g\in \mathcal{U}_{\kappa-D_0}$, the subset $$\left\{(t_{1},t_{2},\cdots,t_{m})\in [-1,1]^m \;\left|\;g+\sum_{i=1}^{m} t_{i}\varphi_{i} \in \mathcal E(\beta,\kappa, \lambda)\right.\right\}$$ is a Lebesgue null subset on $[-1,1]^m$. This proposition has essentially been proved in [@T]. For completeness, we give a proof of it in the next section. Genericity of the transversality condition {#sec:gene} ========================================== In this section, we give a proof of Proposition \[pp:transv\]. For a $C^{2}$ function $g$ and $C^{\infty}$ functions $\varphi_{i}$, $1\le i\le m$, on $S^{1}$, we consider a family of functions $$\label{eqn:fam} f_{\t}(x)=g(x)+\sum_{i=1}^{m}t_{i}\varphi_{i}(x): S^{1}\to \R$$ and the corresponding family of maps $$\label{eqn:Tfam} T_{\t}:\ps\to\ps,\qquad T_{\t}(x,y)=(\lap x, \lambda y+f_{\t}(x))$$ with parameters $\t=(t_{1},t_{2},\cdots,t_{m})\in [-1,1]^m\subset \R^{m}$. Put $$\label{eqn:S2} S(x,\a;\t)=\sum_{i=1}^{n}\lambda^{i-1}f_{\t}([\a]_{i}(x))$$ for $\t\in[-1,1]^m$ and a word $\a\in\sym^{n}$ of length $1\le n\le \infty$. For a point $x\in S^{1}$ and a sequence $\sigma=(\a_{0},\a_{1},\cdots,\a_{k})$ of elements in $\sym^{\infty}$, we consider an affine map $ G_{x,\sigma}:\R^{m}\to\R^{k}$ defined by $$\label{gsigma} G_{x,\sigma}(\t)=\left(\frac{d}{d x}S(x,\a_{i};\t)-\frac{d}{d x}S(x,\a_{0};\t)\right)_{i=1,2,\cdots,k}.$$ If the affine map $G_{x,\sigma}$ is surjective, we define its Jacobian by $$\operatorname{\mathbf{Jac}}(G_{x,\sigma})=\frac{\vol_{k}([0,1]^k)}{\vol_{k}(G_{x,\sigma}^{-1}([0,1]^k)\cap \operatorname{\mathrm{Ker}}(G_{x,\sigma})^{\perp} )}$$ where $\vol_k$ is the $k$-dimensional Hausdorff measure and $\operatorname{\mathrm{Ker}}(G_{x,\sigma})^{\perp}$ is the orthogonal complement of the kernel of the linear part of $G_{x,\sigma}$, whence $$\label{eqn:Jac} \vol(G_{x,\sigma}^{-1}(Y)\cap [-1,1]^m)\le C_0\frac{\vol(Y)}{\operatorname{\mathbf{Jac}}(L)}\quad \mbox{for any Borel subset $Y\subset \R^k$}$$ where $C_0$ is a constant that depends only on the dimensions $m$ and $k$. For $0<\gamma\le 1$, $\delta>0$ and $n\ge 1$, we say that the family $T_{\t}^n$ is [*$(\gamma,\delta)$-generic*]{} if the following property holds: for any finite sequence $\{\a_{i}\}_{i=0}^d$ in $\sym^{\infty}$ such that $[\a_i]_n$ are mutually distinct, for any $x\in S^{1}$ and for any integer $0<k< \gamma d$, we can choose a subsequence $\sigma=(\b_{0},\b_{1},\cdots,\b_{k})$ of length $k$ among $\{\a_{i}\}_{i=0}^d$ so that $G_{x,\sigma}$ is surjective and satisfies $\operatorname{\mathbf{Jac}}(G_{x,\sigma})>\delta$. It is proved in [@T] that \[prop:para\] For given $0<\lambda<1$, $\ell\ge 2$ and $n\ge 1$, there exists a finite collection of $C^{\infty}$ functions $\varphi_{i}$, $1\le i\le m$, such that the corresponding family $T_{\t}^{n}$ is $(1/(n+1),1/2)$-generic, regardless of the $C^{2}$ function $g$. Recall that we are considering fixed $\lambda$ and $\ell$. Let $\beta>0$ be the positive number in the statement of Proposition \[pp:transv\]. We can and do take integers $N_0\ge 2$, $d_0\ge 2$ and $n_0\ge 1$ such that $$\label{eqn:condq} \lambda^{N_0-1}\ell^2<1, \quad d_0/(n_0+1)>N_0+1 \quad \mbox{and }\quad (d_0+1)\exp(-\beta n_0/2)<1/2.$$ Let $\varphi_i$, $1\le i\le m$, be the $C^\infty$ functions in the conclusion of Proposition \[prop:para\] for these $\lambda$, $\ell$ and $n=n_0$. Let $D_0=\sum_{i=1}^m {\left\| \varphi_i \right\|}_{C^r}$. Hence, if $g\in \mathcal{U}_{\kappa-D_0}$ and $(t_1,\dots,t_m)\in [-1,1]^m$, then $g+\sum t_i \varphi_i \in \mathcal{U}_\kappa$. In order to prove the conclusion of Proposition \[pp:transv\], we pick arbitrary $g\in \mathcal{U}_{\kappa- D_0}$ and consider the family $T_{\t}$ defined by and . For an integer $q$, we put $p(q)=[q\log(\ell/\lambda)/\log \ell]+1$. For a word $\c$ of finite length, let $x_{\c}$ be the left end of $\P(\c)$. We fix a word $\a_\infty\in \sym^\infty$ arbitrarily and, for any word $\a$ of finite length, we put $\bar \a=\a \a_\infty$. \[lm:E\]If $f_{\t}\in \mathcal{E}(\beta,\kappa)$, we can take arbitrarily large integer $q$ such that there exist $1+d_0$ words $\a_i$, $0\le i\le d_0$, in $\sym^q$ and a word $\c\in \sym^{p(q)}$ satisfying - $|\frac{d}{d x}S(x_{\c},\bar\a_i;\t)-\frac{d}{d x} S(x_{\c},\bar\a_j;\t)|\le 8\lambda^q\ell^{-q}{\alpha_0}$ for any $1\le i,j\le d_0$, and - $[\a_i]_{n_0}\neq [\a_j]_{n_0}$ if $i\neq j$. By assumption, we can take an arbitrarily large $\tm$ such that there exist a point $x\in S^1$ and subset $E\subset \sym^{\tm}$ such that $\# E\ge \exp(\beta \tm)$ and $$\label{eqn:E1} \left|\frac{d}{d x}S(x,\a;\t)-\frac{d}{d x}S(x,\b;\t)\right|\le 4\lambda^{\tm}\ell^{-\tm}{\alpha_0}\quad\mbox{ for $\a$ and $\b$ in $E$.}$$ For each $0\le j\le [\tm/n_0]$, we introduce an equivalence relation $\sim_j$ on $E$ such that $\a\sim_j \b$ if and only if $[\a]_{jn_0}=[\b]_{jn_0}$, and let $$\nu(j)=\max_{\a\in E}\#\{\b\in E\mid \b\sim_j \a\}.$$ Since $\nu(0)= \# E\ge \exp(\beta \tm)$ while $\nu(j)\le \ell^{\tm-jn_0}$ obviously, there exists $0\le j\le [\tm/n_0]$ such that $\nu(j+1)<\exp(-\beta n_0/2) \nu(j)$. Let $j_*$ be the minimum of such integers $j$ and put $q=\tm-n_0j_*$. Then we have $\nu(j_*)\ge \exp(\beta q)$ and $q\ge \beta \tm/(2\log\ell)$. The equivalence class $H$ w.r.t. $\sim_{j_*}$ of maximum cardinality contains at least $(d_0+1)$ non-empty equivalence classes w.r.t. $\sim_{j_*+1}$, because $$\nu(j_*)-(d_0+1)\nu(j_*+1)> \nu(j_*)-(d_0+1)\exp(-\beta n_0/2)\nu(j_*)>0$$ by (\[eqn:condq\]). So we can take $\b\in \sym^{\tm-q}$ and $\a_i\in \sym^q$, $0\le i\le d$, such that $\b\a_i\in H$ for $0\le i\le d_0$ and that (E2) holds. Put $x'=\b(x)$. It follows from (\[eqn:E1\]) that $$\label{eqn:Sdiff} \left|\frac{d}{d x}S(x',\a_i;\t)-\frac{d}{d x}S(x',\a_j;\t)\right|\le 4\lambda^q\ell^{-q}{\alpha_0}\quad\mbox{ for $0\le i,j\le d_0$}.$$ Take $\c\in \sym^{p(q)}$ such that $x'\in \P(\c)$. Since the distance between $x_{\c}$ and $x'$ is bounded by $\ell^{-p(q)}\le \lambda^q/\ell^q$, the condition (E1) follows from (\[eqn:Sdiff\]) and (\[eqn:alpha\]). Let $\mathcal{B}^q$ be the set of pairs $(\sigma, \c)$ of a sequence $\sigma=(\b_i)_{i=0}^{N_0}$ in $\sym^q$ and $\c\in \sym^{p(q)}$ such that $\operatorname{\mathbf{Jac}}(G_{x_\c,\bar \sigma})>1/2$, where $\bar\sigma=({\bar{\b}}_i)_{i=0}^{N_0}$. For $(\sigma,\c)\in \mathcal{B}^q$ with $\sigma=(\b_i)_{i=0}^{N_0}$, we put $$Y(\sigma,\c)=G_{x_\c,\bar \sigma}^{-1}([-8(\lambda/\ell)^q \alpha_0,8(\lambda/\ell)^q\alpha_0]^{N_0})$$ and $Y(q):=\bigcup_{(\sigma,\c)\in \mathcal{B}^q} Y(\sigma,\c)$. Since the family $T^{n_0}_{\t}$ is $(1/(n_0+1), 1/2)$-generic, the conclusion of Lemma \[lm:E\] and the second condition in imply that, if $f_{\t}\in \mathcal{E} (\beta,\kappa,\lambda)$, the parameter $\t$ is contained in $Y(q)$ for infinitely many $q$. Using and the simple estimate $\#\mathcal{B}^q\le \ell^{q (N_0+1)+p(q)}$, we get $$\vol\left(Y(q)\right)\le C\ell^{q (N_0+1)+p(q)}(\lambda/\ell)^{qN_0}$$ for some constant $C>0$ . By the first condition in , the left hand side converges to $0$ exponentially fast as $q\to \infty$. Therefore we obtain the conclusion of Proposition \[pp:transv\] by Borel-Cantelli lemma. [99]{} Adams, R.A. Sobolev spaces. Pure and Applied Mathematics, Vol. 65. Academic Press, 1975. Bamón, R., Kiwi, J., Rivera-Letelier, J. and Urzúa, R., [*On the topology of solenoidal attractors of the cylinder*]{}, preprint, Math. ArXiv math.DS/0403523. Gouëzel, S. and Liverani, C., Banach spaces adapted to Anosov systems, preprint, Math. ArXiv math.DS/0405278 Hörmander, L. The analysis of linear partial differential operators. I. 2nd edition, Springer-Verlag, Berlin, 1990 Hennion, H., [*Sur un théorème spectral et son application aux noyaux lipschitziens*]{}. Proc. Amer. Math. Soc. [**118**]{}, 627–634 (1993) Rams, M., [*Absolute continuity for the SBR measure for non-linear fat baker maps*]{}. Nonlinearity, [**16**]{}, 1649–1655 (2003) Tsujii, M., [*Fat solenoidal attractors*]{}. Nonlinearity, [**14**]{}, 1011–1027 (2001) [^1]: *Address:* Laboratoire de Probabilités et Modèles aléatoires, Université Pierre et Marie Curie–Boîte courrier 188, 75252–Paris Cedex 05, France; *Email:* [email protected] [^2]: *Address:* Département de Mathématiques et Applications, Ecole Normale Supérieure, 45 rue d’Ulm, Paris, France; *Email:* [email protected] [^3]: *Address:* Department of Mathematics, Hokkaido University, Kita 10 Nishi 8, Kita-ku, Sapporo, Japan; *Email:* [email protected] [^4]: Strictly speaking, the functions $Q_k$ and $Q_{a,b}$ are only defined on the $\lap^n$-fold covering of $S^1$, since their definition involves the choice of an inverse branch, but we will keep the dependence on the choice of the inverse branch implicit in the notation.

ArXiv

--- address: - | Departamento de Física e Química, Faculdade de Engenharia, Universidade Estadual Paulista, Campus da Guaratinguetá, Av. Dr. Ariberto Pereira da Cunha 333, 12500-000\ Guaratinguetá, SP, Brazil. - | Instituto de Física Teórica,\ Universidade Estadual Paulista,\ 01405-900, São Paulo, Brazil. author: - 'J. L. Tomazelli' - 'L. C. Costa' title: The Role of Mass and External Field on the Fermionic Casimir Effect ---

ArXiv

--- abstract: 'Nefarious actors on social media and other platforms often spread rumors and falsehoods through images whose metadata (e.g., captions) have been modified to provide visual substantiation of the rumor/falsehood. This type of modification is referred to as image repurposing, in which often an unmanipulated image is published along with incorrect or manipulated metadata to serve the actor’s ulterior motives. We present the Multimodal Entity Image Repurposing (MEIR) dataset, a substantially challenging dataset over that which has been previously available to support research into image repurposing detection. The new dataset includes location, person, and organization manipulations on real-world data sourced from Flickr. We also present a novel, end-to-end, deep multimodal learning model for assessing the integrity of an image by combining information extracted from the image with related information from a knowledge base. The proposed method is compared against state-of-the-art techniques on existing datasets as well as MEIR, where it outperforms existing methods across the board, with AUC improvement up to 0.23.' author: - Ekraam Sabir - Wael AbdAlmageed - Yue Wu - Prem Natarajan bibliography: - 'ImageRepurposing.bib' title: 'Deep Multimodal Image-Repurposing Detection' --- &lt;ccs2012&gt; &lt;concept&gt; &lt;concept\_id&gt;10003120.10003130.10003131.10011761&lt;/concept\_id&gt; &lt;concept\_desc&gt;Human-centered computing Social media&lt;/concept\_desc&gt; &lt;concept\_significance&gt;500&lt;/concept\_significance&gt; &lt;/concept&gt; &lt;concept&gt; &lt;concept\_id&gt;10010147.10010178.10010179&lt;/concept\_id&gt; &lt;concept\_desc&gt;Computing methodologies Natural language processing&lt;/concept\_desc&gt; &lt;concept\_significance&gt;500&lt;/concept\_significance&gt; &lt;/concept&gt; &lt;concept&gt; &lt;concept\_id&gt;10010147.10010178.10010224&lt;/concept\_id&gt; &lt;concept\_desc&gt;Computing methodologies Computer vision&lt;/concept\_desc&gt; &lt;concept\_significance&gt;500&lt;/concept\_significance&gt; &lt;/concept&gt; &lt;concept&gt; &lt;concept\_id&gt;10010147.10010257.10010258.10010262&lt;/concept\_id&gt; &lt;concept\_desc&gt;Computing methodologies Multi-task learning&lt;/concept\_desc&gt; &lt;concept\_significance&gt;500&lt;/concept\_significance&gt; &lt;/concept&gt; &lt;concept&gt; &lt;concept\_id&gt;10002951.10003227.10003251&lt;/concept\_id&gt; &lt;concept\_desc&gt;Information systems Multimedia information systems&lt;/concept\_desc&gt; &lt;concept\_significance&gt;300&lt;/concept\_significance&gt; &lt;/concept&gt; &lt;concept&gt; &lt;concept\_id&gt;10002951.10003317.10003371.10003386&lt;/concept\_id&gt; &lt;concept\_desc&gt;Information systems Multimedia and multimodal retrieval&lt;/concept\_desc&gt; &lt;concept\_significance&gt;300&lt;/concept\_significance&gt; &lt;/concept&gt; &lt;/ccs2012&gt;

ArXiv

--- abstract: 'I examine the effect of exogenous spatial heterogeneity on the coexistence of competing species using a simple model of non-hierarchical competition for site occupancy on a lattice. The sites on the lattice are divided into two types representing two different habitats or spatial resources. The model features no temporal variability, hierarchical competition, type-dependent interactions or other features traditionally known to support more competing species than there are resources. Nonetheless, stable coexistence of two habitat specialists and a generalist is observed in this model for a range of parameter values. In the spatially implicit mean field approximation of the model, such coexistence is shown to be impossible, demonstrating that it indeed arises from the explicit spatial structure.' address: 'Department of Mathematics and Statistics, University of Helsinki, Finland' author: - Ilmari Karonen bibliography: - 'coex2env.bib' date: 'October 13, 2011' title: Stable trimorphic coexistence in a lattice model of spatial competition with two site types --- 92D25 (population dynamics) ,60K35 (interacting random processes) ,coexistence mechanisms ,resource competition ,spatial ecology ,individual-based models ,lattice contact process ,heterogeneous environment Introduction ============ The coexistence of competing species, and factors promoting and limiting it, are of considerable practical and theoretical interest in ecology. A well known “rule of thumb” is the principle of competitive exclusion [@gause1932; @tilman1982; @levin1970], which states that at most $n$ mutually competing species may stably coexist on $n$ available resources. Under suitable assumptions, the competitive exclusion principle can be proven as a mathematical theorem. However, if these assumptions are violated — for example, if the resource abundances may fluctuate over time, either due to external resources or simply because the ecological dynamics tend to a cyclic or chaotic attractor — it may no longer hold (although a related concept, the essential dimensionality of the environment [@dieckmann2006; @metz2008], may still be applied to such systems). For systems which do satisfy these assumptions, the validity of the competitive exclusion principle depends fundamentally on just what we count as “a resource” [@abrams-88]. This is not as simple a matter as it sounds. Were one to ask a practical-minded ecologist what constitutes a resource, they might name examples such as water, sunlight and nutrients for plants, or prey species for animals. But in the mathematically exact form of the competitive exclusion principle, almost anything may constitute a distinct resource: a single prey individual, a square meter of land, a specific combination of nutrient concentrations, etc. Thus, even in systems which the competitive exclusion principle formally holds, the actual number of potentially coexisting competitors may be greater than one would expect by naively undercounting the resources. The effect of spatial structure on the maximum diversity a system can support is, in particular, often neglected. For example, systems consisting of several distinct types of habitats are often modelled by assuming that each habitat constitutes a homogeneous patch, within which the populations are well mixed. If competition between individuals is for suitable living space within these habitats, space in each habitat then becomes a single resource, and thus one would expect (and will, given these assumptions, mathematically discover) that at most $n$ competitors may stably persist in $n$ distinct habitats. In reality, however, even if habitats are homogeneous, they are certainly not usually well mixed. Thus, individuals living near habitat boundaries will, over time, experience a different environment than those living in the interior of habitat patches. (Even if the individuals themselves don’t move or interact with anything outside their local site, their offspring must still disperse and may end up in a different habitat.) This can create additional niches near habitat boundaries in which additional competing species might be able to coexist. To demonstrate this, I present in this paper a simple spatially explicit toy model of site occupancy competition, which supports stable coexistence of three strains – two specialists and one generalist – on two spatially segregated habitats on a lattice of sites. This model contains no other features known to promote coexistence, such as internally or externally generated temporal fluctuation [@hsu1978; @armstrong1980], hierarchical site competition [@adler2000; @tilman1994], direct strain-dependent competition terms [@murrell-law-03] or cooperative or other nonlinear interactions between individuals. Rather, the coexistence arises merely from the presence of habitat boundaries combined with passive distance-limited dispersal, which causes specialist strains to be locally maladapted near these boundaries and thereby allows more generalist strains to persist there. As far as I know, this particular mechanism of coexistence has not been previously studied in individual-based models. A similar coexistence mechanism was very recently described in a reaction–diffusion model and in a 1D stepping stone model by @debarre2011, who termed it “habitat boundary polymorphism”. The results in this paper parallel theirs, and confirm that this mechanism is robust with respect to the details of the model, provided that the basic features of habitat heterogeneity and passive distance-limited dispersal are present. A model almost identical to mine was studied by @lanchier2006, who showed analytically that it could support stable dimorphic coexistence, either of a generalist and a specialist strain or of two different specialist strains. My model differs from theirs only in that they restrict the habitat configuration to the special case of an infinite regular checkerboard pattern consisting of $n$-by-$n$ squares of each habitat type.[^1] However, while Lanchier and Neuhauser did briefly remark that “the generalists persist for a very long time along the boundaries …where the density of specialists is low” in numerical simulations, they do not seem to have investigated this possibility of trimorphic coexistence in their model further. Similarly, @snyder-chesson-03 define a model quite similar to mine (although in discrete time and one spatial dimension), and observe the enhancing effect of stable spatial heterogeneity and local dispersal on coexistence of competitors, but also restrict their analysis to two competitors. Model definition ================ I model a population of haploid, asexually reproducing sessile organisms with distance-limited offspring dispersal. The model I’ll define below belongs to the class of lattice contact processes [@harris1974; @neuhauser1992], in which the environment is taken to consist of a lattice of discrete sites, and interactions are (mainly) between nearest neighbor sites. Let $L$ be a regular two-dimensional lattice of sites (e.g. $L = {\mathbb{Z}}^2$, although for numerical simulations a finite lattice must obviously be used). To each site I assign a random, fixed habitat class (“A” or “B”), such that both classes of sites are present in $L$ in equal numbers. (I will describe the way in which the habitat classes are assigned in more detail below.) Each site in $L$ may, at a given time, be either vacant or occupied by an individual belonging to one of three strains: an A-specialist ($a$), a B-specialist ($b$) or a generalist ($g$). The two specialist strains can only occupy sites in their respective habitat class, while the generalist strain may occupy any site. Except for their different habitat adaptation, the strains are completely identical: all individuals die with rate $\mu$ and produce offspring with rate $\phi$. Offspring are randomly dispersed to the eight nearest sites surrounding the parent individual’s site (or possibly, with probability $\epsilon$, to a randomly chosen site in $L$), and will become new individuals if the site they land in is vacant and of a suitable habitat class. However, the generalists pay a cost for their ability to live in either habitat: their offspring survive only with probability $p_{{\mathrm{g}}} < 1$.[^2] The time evolution of the entire lattice $L$ can thus be considered as a continuous-time Markov process, whose state at time $t$ is a function $\eta_t : L \to \{0, {\mathrm{a}}, {\mathrm{b}}, {\mathrm{g}}\}$ mapping sites in $L$ to their occupancy state (with the state $0$ denoting a vacant site). The local transition rates of a site $x$ are then $$\begin{array}{r @{\hspace{4pt}=\hspace{4pt}} r} r(0 \to {\mathrm{a}}) & \displaystyle {\mathbf{1}_{\{H_x={\mathrm{A}}\}}} \phi \left( (1-\epsilon) \sum_{y \in E_x} \frac{{\mathbf{1}_{\{\eta_t(y)={\mathrm{a}}\}}}}{|E_y|} + \epsilon \sum_{y \in L} \frac{{\mathbf{1}_{\{\eta_t(y)={\mathrm{a}}\}}}}{|L|} \right), \\ r(0 \to {\mathrm{b}}) & \displaystyle {\mathbf{1}_{\{H_x={\mathrm{B}}\}}} \phi \left( (1-\epsilon) \sum_{y \in E_x} \frac{{\mathbf{1}_{\{\eta_t(y)={\mathrm{b}}\}}}}{|E_y|} + \epsilon \sum_{y \in L} \frac{{\mathbf{1}_{\{\eta_t(y)={\mathrm{b}}\}}}}{|L|} \right), \\ r(0 \to {\mathrm{g}}) & \displaystyle p_{{\mathrm{g}}} \phi \left( (1-\epsilon) \sum_{y \in E_x} \frac{{\mathbf{1}_{\{\eta_t(y)={\mathrm{g}}\}}}}{|E_y|} + \epsilon \sum_{y \in L} \frac{{\mathbf{1}_{\{\eta_t(y)={\mathrm{g}}\}}}}{|L|} \right), \end{array}\label{eqn-transitions}$$ and $$r({\mathrm{a}}\to 0) = r({\mathrm{b}}\to 0) = r({\mathrm{g}}\to 0) = \mu,$$ where $r(s \to s')$ is the rate at which a site in state $s$ changes to state $s'$, $H_x$ is the habitat class of site $x$ and $E_x$ is the set of sites adjacent to $x$. The behavior of the model is determined by the habitat configuration along with two parameters: the scaled baseline fecundity rate $\lambda = \phi/\mu$ (or equivalently, the scaled mortality rate $\lambda^{-1} = \mu/\phi$) and the generalist survival rate $p_{{\mathrm{g}}}$. The baseline fecundity affects the equilibrium population density of both the specialists and the generalist strain: at low $\lambda$ all strains die out, while at very high $\lambda$ almost all sites are occupied at any time. The parameter $0 < p_{{\mathrm{g}}} < 1$ determines the penalty which the generalist must pay for its ability to exploit both site types, and is (together with the habitat configuration) crucial in determining the outcome of the model. If $p_{{\mathrm{g}}}$ is too low, the generalist strain will not be viable, or, even if viable on its own, may lose in competition to the specialists. Conversely, if $p_{{\mathrm{g}}}$ is close to 1, the specialist strains gain little or no advantage over the generalist from their specialization, while paying a considerable price in being able to live in only one habitat, and can thus be expected to lose to the generalist. Landscape generation {#landscape} ==================== An issue so far overlooked in the model definition above is the way in which the lattice sites are assigned to their habitat classes. The simplest way to do so, of course, is to simply assign each site independently to either habitat with equal probability. This produces a lattice with no correlations between the habitat classes of different sites. However, real environmental features are often correlated, and it would be desirable to consider the effects of such correlations on the behavior of the model. To first order, such correlations can be characterized by the pairwise correlation probability $k = \mathrm{Pr}[ H_x=H_y | y \in E_x ]$, i.e. the probability that two randomly chosen adjacent sites have the same habitat type. With an equal number of sites in each habitat, the pairwise correlation probability on a completely random, uncorrelated lattice is $k = \frac{1}{2}$, while lattices with $\frac{1}{2} < k < 1$ are positively correlated and those with $0 < k < \frac{1}{2}$ are anticorrelated.[^3] To generate random habitat class distributions with a given pairwise correlation probability for numerical simulations, I start by randomly choosing half of the sites and assigning them to habitat A and the rest to habitat B. I then apply an iterative annealing process, which reassigns randomly chosen sites to new habitat classes with suitable weighted probabilities, until the desired value of $k$ is reached. There are many possible variations of the general annealing scheme I’ve used. The basic idea in all of them is to change the habitat classes of randomly chosen sites if this would change $k$ in the desired direction, while occasionally also allowing changes in the other direction so that the process doesn’t get stuck at a local minimum or maximum. The particular annealing algorithm I’ve used to generate habitat configurations for the simulations in this paper[^4] chooses random adjacent pairs of sites, and swaps their habitat classes with probability $$p = \frac{d^\gamma}{d^\gamma + (1-d)^\gamma}$$ if $k_{\mathrm{current}} < k_{\mathrm{target}}$, or with probability $1-p$ otherwise, where $d$ is the fraction of sites adjacent to the chosen pair which belong to the opposite habitat than their neighbor in the chosen pair. The exponent $\gamma$ is a free parameter which controls the “temperature” of the process. When $\gamma = 1$, the probability of exchanging the habitat classes of a chosen site pair is a linear function of $d$. This tends to produce fairly slow convergence and rough, jagged cluster boundaries. At high values of $\gamma$, $p$ approaches a step function, producing faster initial convergence and smoother cluster boundaries, but also increasing the risk of the process getting stuck at a local maximum or minimum of $k$. Mean field approximation ======================== Classical ecological theory predicts that the only possible outcome of three distinct strains competing for the occupancy of two habitats should be the eventual extinction of one or more of the strains, except at specific degenerate choices of the model parameters where neutral coexistence may occur. I will demonstrate below that this prediction indeed holds if the populations are assumed to be well mixed, either globally or within each habitat. However, I shall also show that, in the full model with explicit spatial structure, a region of stable trimorphic coexistence does exist for intermediate values of $p_{{\mathrm{g}}}$. Assuming that all offspring disperse uniformly over the entire lattice, i.e. that $\epsilon = 1$ in \[eqn-transitions\], the transition rates of each site are fully described by the mean population densities $n_{\mathrm{a}}$, $n_{\mathrm{b}}$ and $n_{\mathrm{g}}$ of the different strains, where $$n_s = \sum_{y \in L} \frac{{\mathbf{1}_{\{\eta(y) = s\}}}}{|L|}$$ for each strain $s$. Further letting the lattice size $|L|$ tend to infinity, one obtains a simple system of ordinary differential equations describing the time evolution of these mean population densities—the so called *mean field* approximation—which may be solved analytically. Such an approximation of an essentially equivalent model was presented by @lanchier2006, who showed that trimorphic coexistence was only possible for degenerate choices of parameter values. However, simply assuming all dispersal to be global completely neglects not only the detailed spatial habitat structure, but even the pairwise correlation parameter $k$. A better approximation, similar to the “improved mean field approximation” of @hiebeler-morin-07, is obtained by assuming well mixing only within each habitat. The resulting approximation can be interpreted as a model of a population inhabiting two well-mixed habitat patches, with a fraction $k$ of all offspring remaining within their parent’s patch and the rest dispersing to the other patch. This *two-patch approximation* takes into account the habitat correlation parameter $k$ but still retains the analytical tractability of the mean field approximation. (For $k=\frac{1}{2}$, the two approximations are exactly equivalent.) Below, I will analyze this approximation of the model defined above, and show that it also only supports non-degenerate coexistence of at most two strains. ![The outcome of the model as predicted by the two-patch approximation for different values of the normalized mortality rate $\mu/\phi$ and the generalist survival probability $p_{{\mathrm{g}}}$. The diagonal line at $\mu/\phi = p_{{\mathrm{g}}}$ and the vertical line at $\mu/\phi = k$ mark the boundaries at which the generalist and specialist strains go extinct even in the absence of competitors. In the regions marked “generalist wins” and “specialists win”, all strains can survive in the absence of competitors, but from any initial state with all three strains present, the system converges to either a monomorphic generalist-only state or a dimorphic specialist-only state. Along the white line at $p_{{\mathrm{g}}} = k$, the three strains can coexist neutrally.[]{data-label="fig-meanfield"}](diagram-meanfield.pdf){width="40.00000%"} Assume that the occupancy states of the lattice sites are independent, and that the probability of a site being occupied by a given strain $s$ is equal to the mean population density $n_{s,H}$ of that type in the site’s habitat $H$. Then, in the limit as $|L| \to \infty$, the time evolution of the population densities can be written as $$\begin{aligned} \frac{d}{dt} n_{s,{\mathrm{A}}} &= p_{s,{\mathrm{A}}} v_{{\mathrm{A}}} \phi (kn_{s,{\mathrm{A}}} + (1-k)n_{s,{\mathrm{B}}}) - \mu n_{s,{\mathrm{A}}}, \\ \frac{d}{dt} n_{s,{\mathrm{B}}} &= p_{s,{\mathrm{B}}} v_{{\mathrm{B}}} \phi (kn_{s,{\mathrm{B}}} + (1-k)n_{s,{\mathrm{A}}}) - \mu n_{s,{\mathrm{B}}},\end{aligned}$$ for $s \in \{{\mathrm{a}}, {\mathrm{b}}, {\mathrm{g}}\}$, where $v_{\mathrm{A}}$ and $v_{\mathrm{B}}$ are the vacant site densities in the two habitats and $p_{s,{\mathrm{A}}}$ and $p_{s,{\mathrm{B}}}$ are the probabilities of an offspring of type $s$ surviving in the respective habitats: $$\begin{aligned} p_{{\mathrm{a}},{\mathrm{A}}} = p_{{\mathrm{b}},{\mathrm{B}}} &= 1, \\ p_{{\mathrm{b}},{\mathrm{A}}} = p_{{\mathrm{a}},{\mathrm{B}}} &= 0, \\ p_{{\mathrm{g}},{\mathrm{A}}} = p_{{\mathrm{g}},{\mathrm{B}}} &= p_{\mathrm{g}}.\end{aligned}$$ Equivalently, this system may be written in matrix form as $$\frac{d}{dt} \bar n_s = M_s \bar n_s$$ for $s \in \{{\mathrm{a}}, {\mathrm{b}}, {\mathrm{g}}\}$, where $\bar n_s = [n_{s,{\mathrm{A}}}, n_{s,{\mathrm{B}}}]^T$ and $$M_s = \left[ \begin{array}{cc} \phi k p_{s,{\mathrm{A}}} v_{{\mathrm{A}}} - \mu & \phi (1-k) p_{s,{\mathrm{A}}} v_{{\mathrm{A}}} \\ \phi (1-k) p_{s,{\mathrm{B}}} v_{{\mathrm{B}}} & \phi k p_{s,{\mathrm{B}}} v_{{\mathrm{B}}} - \mu \end{array} \right].$$ If this system has a nontrivial interior equilibrium $\tilde n_s$, this necessarily implies that $M_s \tilde n_s = [0, 0]^T \ne \tilde n_s$, and therefore that $M_s$ must be singular, and thus have a zero determinant, for each strain $s$ present in the population. Writing out the determinant as $$\begin{aligned} |M_s| &= \phi^2 (2k-1) p_{s,{\mathrm{A}}} v_{{\mathrm{A}}} p_{s,{\mathrm{B}}} v_{{\mathrm{B}}} \\ &- \phi \mu k \left( p_{s,{\mathrm{A}}} v_{{\mathrm{A}}} + p_{s,{\mathrm{B}}} v_{{\mathrm{B}}} \right) + \mu^2 = 0, \end{aligned}$$ yields, for each $s$, an equation containing the same two unknown variables: $v_{{\mathrm{A}}}$ and $v_{{\mathrm{B}}}$. As the coefficients $p_{s,{\mathrm{A}}}$ and $p_{s,{\mathrm{B}}}$ will, in general, be different for each strain $s$, one can see that, except for degenerate choices of the parameter values, no solution will exist for more than two strains. More specifically, we can see that, in the absence of specialists, small generalist populations can grow in the two-patch approximation of this specific model if and only if $p_{\mathrm{g}}\phi > \mu$, and conversely that small populations of either specialist strain can grow in the absence of the generalist if and only if $k \phi > \mu$. Outside these regions, shown in figure \[fig-meanfield\], the respective strains are not viable and will always die out (as the per capita growth rates in this model are always maximized at vanishing population densities). Within the region where all strains are viable, the approximated model has (in general) two possibly stable equilibria: one where only the generalist is present, and one where the generalist is absent and both specialists present. (Any equilibria with only one specialist present are obviously unstable against invasion by the other specialist.) The former is stable if and only if $p_{\mathrm{g}}> k$, while the latter is stable if and only if $p_{\mathrm{g}}< k$. Only at exactly $p_{\mathrm{g}}= k$, shown as the white line in figure \[fig-meanfield\], both of the equilibria are neutrally stable, and are connected by a line of neutrally stable equilibria along which neutral coexistence can occur. Simulation results ================== Studying the dynamics of the full, unapproximated model requires numerical simulations. As such simulations tend to be computationally intensive, I have carried them out using custom, optimized programs written in the C programming language.[^5] The simulation code used for this paper includes two variants of the coupling-based simulation algorithm described in [@couplingwip], one using an occupancy list for low population densities, and another using a vacancy list for high densities, with the outer simulation loop periodically checking the population density and switching to the variant with the higher mean time step per iteration. I have also ported the basic simulation code (without the coupling technique) to Java for demonstration purposes using interactive applets. All simulation runs for this paper were done on a square $256 \times 256$ lattice with 8 neighbors per site and with the edges wrapping around to the opposite sides. In all simulation runs, time has been scaled so that the *per capita* mortality rate $\mu = 1$; in effect, I measure time in mean individual lifetimes. Habitat configurations were generated using an annealing method as described in section \[landscape\]. The “flea” pseudorandom number generator [@flearng] was used to produce random numbers, although I also carried out tests using other random number generators to check that the results did not depend on such details. ![Plots of equilibrium specialist and generalist densities at $k=0.5$ and $\phi = 8\mu$ as functions of $p_{\mathrm{g}}$ obtained from 20 numerical simulation runs. See text for details. The graph on the right has been plotted on a logarithmic scale and zoomed in to better show the coexistence region.[]{data-label="fig1sim"}](densitycurves-k_5-b8-t50000-lin.pdf "fig:"){width="48.00000%"} ![Plots of equilibrium specialist and generalist densities at $k=0.5$ and $\phi = 8\mu$ as functions of $p_{\mathrm{g}}$ obtained from 20 numerical simulation runs. See text for details. The graph on the right has been plotted on a logarithmic scale and zoomed in to better show the coexistence region.[]{data-label="fig1sim"}](densitycurves-k_5-b8-t50000-log.pdf "fig:"){width="48.00000%"} Figure \[fig1sim\] shows the equilibrium densities of generalists and specialists observed in repeated individual-based simulations of the model on a $256 \times 256$ lattice with reproduction to 8 nearest neighbors and wrapped edges with an uncorrelated habitat distribution and a moderate value of $\phi/\mu$. Each simulation run was started from a random habitat configuration and a random initial state, with half the sites occupied by generalists and half by specialists. Populations were allowed to equilibrate for $50000/\mu$ time units, after which population densities were averaged over another $50000/\mu$ time units. The specialist occupancy fractions are summed over both specialist strains. Contrary to the predictions from the mean field approximation, a non-degenerate region of the parameter space where all three strains stably coexist can be seen in figure \[fig1sim\]. This region is displayed more extensively in figures \[fig2sim\] and \[fig2bsim\], which plot the observed region of coexistence against $\mu/\phi$ and $p_{\mathrm{g}}$ for the various habitat configurations (anticorrelated, uncorrelated and two positively correlated patterns) shown in figure \[fig3sim\]. Figure \[fig2sim\] shows results for pure nearest-neighbor dispersal ($\epsilon = 0$), while in figure \[fig2bsim\], 1% of all offspring were permitted to disperse uniformly over the whole lattice ($\epsilon = 0.01$). ![Results of numerical simulations on a $256 \times 256$ site lattice as functions of $\mu/\phi$ and $p_{\mathrm{g}}$ on the habitat patterns from figure \[fig3sim\] with no global dispersal ($\epsilon = 0$). See text for details. Compare with figure \[fig2bsim\] and with the mean field predictions from figure \[fig-meanfield\].[]{data-label="fig2sim"}](gradient.pdf "fig:"){width="45.00000%"}\ \ ![Results of numerical simulations on a $256 \times 256$ site lattice as functions of $\mu/\phi$ and $p_{\mathrm{g}}$ on the habitat patterns from figure \[fig3sim\] with occasional global dispersal ($\epsilon = 0.01$). See text for details. The marks on figures \[fig2bsimB\] and \[fig2bsimD\] show the parameter values used for the invasion simulations in figure \[fig1inv\]. Compare with figure \[fig2sim\] and with the mean field predictions from figure \[fig-meanfield\].[]{data-label="fig2bsim"}](gradient.pdf "fig:"){width="45.00000%"}\ \ \ For figures \[fig2sim\] and \[fig2bsim\], each simulation run was started from a random initial population on the fixed pregenerated habitat landscapes shown in figure \[fig3sim\]. Populations were allowed to equilibrate for $20000/\mu$ time units, after which population densities were averaged over $2000/\mu$ time units. The red areas show where only the specialist strains survived, while in the blue regions only the generalist strain remained. The lighter shaded area between them shows the part of the parameter space where both specialists and generalists survived, with the color gradient shown above the figures indicating the relative average population densities. In both figures \[fig2sim\] and \[fig2bsim\], the region of stable coexistence can be seen as a more or less wedge-shaped area starting from the point where the viability boundaries of the strains intersect, which is where the two-patch approximation would predict a line of neutral coexistence (see figure \[fig-meanfield\]). The main difference between the figures can be seen in the lower left side of the coexistence region: with global dispersal, the lower boundary of the coexistence region is quite sharp, whereas with no global dispersal and high baseline fecundity $\phi/\mu$, the generalist can often survive in small numbers (shown as a light orange hue in the plots) even where the specialist dominates. This happens simply because the high fecundity allows even small isolated population clusters to survive for a long time, and because the strictly local dispersal prevents the specialists from recolonizing isolated habitat patches. If such a habitat patch happens to end up with no specialist individuals (either because all happen to die out, or because none were present initially), the patch can be colonized by generalists, which are then safe from competition there. Allowing a fraction of offspring to disperse globally lets the specialist strains recolonize such patches, eliminating this effect. It can also be seen that the addition of global dispersal generally reduces the width of the coexistence region somewhat, although (except for the isolated patch effect noted above) the reduction is not yet very large for $\epsilon = 0.01$. This is to be expected, given that at $\epsilon = 1$ the coexistence region reduces to a line, as shown by the mean field (and two-patch) approximation above. The results shown in figures \[fig1sim\], \[fig2sim\] and \[fig2bsim\] were calculated using a simulation technique based on monotone coupling [@couplingwip], which allows the system to be efficiently simulated for all values of the parameter $p_{\mathrm{g}}$ in parallel. Each line in figure \[fig1sim\] and each vertical stripe (out of 1024 per plot) in figures \[fig2sim\] and \[fig2bsim\] corresponds to one simulation run. Because the simulation technique used causes the effects of random demographic fluctuations on populations with different values of $p_{\mathrm{g}}$ within the same run to be correlated, the results show stronger correlations within each run than between runs, which should be kept in mind when interpreting the figures. Figure \[fig4sim\] contain snapshots of simulations run on lattices with different site type patterns. It can be seen that, when sites of the same type are strongly clustered, large contiguous clusters are dominated by the respective specialist strain, while the generalist strain is able to persist in areas near cluster boundaries and in isolated minor clusters too small to support a stable specialist population. On the other hand, when site types are uncorrelated, a different pattern is observed. Such lattices contain no large contiguous clusters that could be dominated by one specialist strain; instead, the two specialist strains tend to occur together in regions where the random distribution of site types happens to favor one or both of them. Through competition with the generalist strain, the two specialist strains indirectly support one another, even though there is no direct interaction between them. Mutual invasibility =================== To a skeptical mind, the results presented above may yet leave some doubt about whether the apparent coexistence observed in these simulations is indeed genuinely stable, or merely an artifact of slow convergence and insufficient simulation time. After all, if the simulation was run for long enough on a finite lattice, *eventually* one of the strains (and eventually all of them) would almost surely go extinct simply due to demographic stochasticity. Thus, it may not even be entirely clear what “stable coexistence on a finite lattice” should actually mean. On an infinite lattice, as in @lanchier2006, a set of strains may be said to coexist stably if they can all persist indefinitely long with non-zero probability. By this definition, no stable coexistence (or even just existence) is possible in a finite system. However, since there do exist known results that relate the scaling of the expected time to extinction on a finite lattice, as a function of lattice size, to the limiting behavior of the model on an infinite lattice, one might be inclined to try and use such scaling laws to extrapolate stability from small lattices to the infinite limit. This is not the approach I have taken. After all, real habitats and populations, like the simulations employed in this paper, are finite — in appealing to a definition of coexistence that only works for infinite populations, one ends up obscuring the fact that, in reality, if a population of tens of thousands of individuals persisting over equally many generations is not considered stable, it’s hard to say what should be. Rather, I wish to demonstrate the stability of the trimorphic coexistence in my model in a more direct manner, by showing that it exhibits *mutual invasibility*. That is to say, if a small number of individuals of any of the three strains are introduced into a stable population consisting solely of the other two strains, the introduced strain will, with positive probability, survive and grow in number up to its equilibrium density in the trimorphic equilibrium, with the initial part of the growth curve appearing approximately exponential. \ \ Figure \[fig1inv\] shows some simulations demonstrating mutual invasibility at six points within the coexistence region of the parameter space. With 100 initial invader individuals, the invader strain survived and established itself in all runs carried out — evidently the invasion probability of a single invader exceeds $1/100$ at all the sampled parameter values. The population density of the invading strain over time shows a distinctive sigmoid shape, with the initial growth being approximately exponential. A notable feature visible in the plots is that invasion by a specialist strain also increases the population of the other specialist strain already present; this happens because both specialists are in competition with the generalist strain. The relatively high variance seen in some of the plots during the growth phase is due to initial demographic stochasticity affecting the time until exponential growth sets in; once properly started, the shape of the growth curve is very similar in all runs. Were the coexistence observed in this model merely neutral, the population density of a newly introduced strain would be as likely to decrease as to increase as the result of stochastic fluctuations. The fact that, at the sampled parameter values, small populations of each strain instead show a clear increasing trend confirms that this model exhibits true, non-neutral coexistence. Discussion ========== In this paper I’ve demonstrated, using a simple toy model of competitive population dynamics on a lattice, that spatial heterogeneity is one of the mechanisms by which the competitive exclusion principle can be violated. The fact that this cannot occur in well-mixed populations shows that population viscosity and explicit spatial structure are essential to this mechanism. Had the model included more than two habitat types, temporal variation, hierarchical competition or nonlinear interactions between individuals, the coexistence of multiple strains would not have been at all surprising. Yet it has none of these, and can still support more than two strains in stable coexistence. All that allows such coexistence to persist in this model is the combination of environmental variation, persistent spatial structure and distance-limited dispersal; eliminating any of these reduces the model to one capable of supporting no more strains than would be predicted by a naive application of the competitive exclusion principle. Real organisms do not usually live in a completely homogeneous environment, nor do most of them disperse uniformly over their entire habitat. It is obvious and commonly acknowledged that environmental variation can increase diversity, yet the fact that, when combined with distance-limited dispersal, this increase can be more than linear seems to have attracted little attention. Yet the ubiquity of habitat edges and fragmented landscapes in nature suggests that it should be possible to find examples of this type of coexistence in nature, and indeed that such “edge effects” may contribute to the generation and maintenance of ecological diversity in many, if not most, ecosystems. I find, however, that in many ways this work has raised more questions than it has answered. For example, an obvious question would be whether the model allows the stable coexistence of more than three strains. Another natural question is whether the coexistence of three or more strains in this type of model can also be evolutionarily stable, and further, whether it might arise from a mono- or dimorphic state through evolutionary branching [@geritz1998; @magori2005]. Based on limited simulation experiments, the answer to all of these questions appears to be “yes”, although the conditions still need to be explored more thoroughly. Acknowledgments =============== I would like to thank my colleague Robert Service for his comment during a presentation which originally led me to investigate this model, and my advisor Éva Kisdi for her guidance and assistance. I am also grateful to Minus van Baalen for useful discussions and for suggestions regarding the landscape generation algorithm, and to the editor and reviewers of the previous revisions of this paper for their valuable feedback. This work was financially supported by the Finnish Graduate School in Computational Sciences (FICS). [^1]: also consider a somewhat different set of dispersal kernels, but both theirs and mine include the basic case of strict nearest-neighbor dispersal. [^2]: Equivalently, I could’ve scaled the offspring production rate of the generalists to $p_{{\mathrm{g}}} \phi$. However, the definition I’ve chosen allows a straightforward generalization to semi-specialist strains with different (non-zero) survival rates in different habitats. [^3]: On a square lattice where each site is adjacent to its 8 nearest neighbors, the smallest achievable value of $k$ is $\frac{1}{4}$. [^4]: The interactive Java applets from which the snapshots in figure \[fig4sim\] are taken use a different annealing algorithm. [^5]: Source code available from author under an open source license.

ArXiv

--- abstract: 'Let $T_1, T_2,.......,T_k$ be spanning trees in a graph $G$. If for any pair of vertices $\{u,v\}$ of $G$, the paths between $u$ and $v$ in every $T_i$( $1\leq i\leq k$) do not contain common edges and common vertices, except the vertices $u$ and $v$, then $T_1, T_2,.......,T_k$ are called completely independent spanning trees in $G$. The $n-$dimensional augmented cube, denoted as $AQ_n$, a variation of the hypercube possesses several embeddable properties that the hypercube and its variations do not possess. For $AQ_n$ ($n \geq 6$), construction of $4$ completely independent spanning trees of which two trees with diameters $2n-5$ and two trees with diameters $2n-3$ are given.' title: Construction of four completely independent spanning trees on augmented cubes --- 15truept S. A. Mane, S. A. Kandekar, B. N. Waphare\ [Center for Advanced Studies in Mathematics, Department of Mathematics,\ Savitribai Phule Pune University, Pune-411007, India.]{}\ [**Keywords:**]{} Completely independent spanning trees, Augmented cubes Introduction ============ Interconnection networks have been widely studied recently. The architecture of an interconnection networks is usually denoted as an undirected graph $G$. A graph $G$ is a triple consisting of a vertex set $V(G)$, an edge set $E(G)$, and a relation that associates with each edge two vertices called its endpoints[@we]. Many useful topologies have been proposed to balance performance and cost parameters. Among them, the hypercube $Q_n$ is one of the most popular topology and has been studied for parallel networks. Augmented cubes are derivatives of hypercubes (proposed by Choudam and Sunitha[@cs]) with good geometric features that retain some favorable properties of the hypercubes (since $Q_n \subset AQ_n$), such as vertex symmetry, maximum connectivity, routing and broadcasting procedures with linear time complexity. An $n-$dimensional augmented cube $AQ_n$ can be formed as an extension of $Q_n$ by adding some links. For any positive integer $n$, $AQ_n$ is $(2n-1)$-regular and $(2n-1)$-connected (except $n = 3$) graph with $2^n$ vertices. Moreover, $AQ_n$ possesses several embeddable properties that the hypercube and its variations do not possess. The main merit of augmented cubes is that their diameters are about half of those of the corresponding hypercubes.\ A tree $T$ is called a spanning tree of a graph $G$ if $V(T)= V(G)$. Two spanning trees $T_1$ and $T_2$ in $G$ are edge-disjoint if $E(T_1)\cap E(T_2)= \phi$. For a given tree $T$ and a given pair of vertices $u$ and $v$ of $T$, let $P_T(u,v)$ be the set of vertices in the unique path between $u$ and $v$ in $T$. Two spanning trees $T_1$ and $T_2$ are internally vertex disjoint if for any pair of vertices $u$ and $v$ of $V(G)$, $P_{T_1}(u,v)\cap P_{T_2}(u,v)= \{u,v\}$. Finally, the spanning trees $T_1, T_2,.......,T_k$ of $G$ are completely independent spanning trees (CISTs for short) if they are pairwise edge-disjoint and internally vertex disjoint.\ The study of CISTs was due to the early work of Hasunuma[@h1], he conjectured that there are $k$ CISTs in any $2k-$connected graph. P$\acute{e}$terfalvi[@pt] gave counter example to disprove Hasunuma’s conjecture. He showed that there exists a $k-$connected graph which does not contain two CISTs for each $k \geq 2$. Pai et al[@p1] showed that the results are negative to Hasunuma’s conjecture in case of hypercube of dimension $n \in \{10,12,14,20,22,24,26,28,30\}$. Many authors provided a necessary condition of CISTs[@a; @c1; @f; @h3; @hl]. For more detail work on CISTs and their diameters see [@c2; @g; @h1; @h2; @h3; @h4; @h5; @m; @w; @y].\ Constructing CISTs has many applications on interconnection networks such as fault-tolerant broadcasting and secure message distribution.In underlying graph of communication networks, we want vertices to be close together to avoid communication delays.\ Pai and Chang[@p2] provided a unified approach for constructing two CISTs in several hypercube-variant networks, in particular for an $n-$dimensional hypercube variant network, the diameters of the constructed CISTs were $2n-1$. They asked about the hypercube variant networks which they studied “how to design algorithms to construct more than two CISTs in high dimensional hypercube-variant networks with smaller diameter?”\ Motivated by this question, we provide a construction of four CISTs in augmented cube $AQ_n$ ($n \geq 6$) of which two trees with diameters $2n-3$ and two trees with diameters $2n-5$ are constructed. Also, construction of $n-1$ CISTs in augmented cube $AQ_n$ for $n = 3,4,5$ is given which pointed out that the Hasunuma’s conjecture does hold in the case of $AQ_n$ for $n = 3,4,5$.\ For undefined terminology and notation, see[@we] Preliminaries ============= The definition of the $n-$dimensional augmented cube is stated as the following. Let $n \geq 1$ be an integer. The $n$-dimensional augmented cube, denoted by $AQ_n$, is a graph with $2^n$ vertices, and each vertex $u$ can be distinctly labeled by an $n$-bit binary string, $u = u_1u_2....u_n$. $AQ_1$ is the graph $K_2$ with vertex set $\{0, 1\}$. For $n \geq 2$, $AQ_n$ can be recursively constructed by two copies of $AQ_{n-1}$, denoted by $AQ^0_{n-1}$ and $AQ^1_{n-1}$, and by adding $2^n$ edges between $AQ^0_{n-1}$ and $AQ^1_{n-1}$ as follows:\ Let $V(AQ^0_{n-1}) = \{0u_2....u_n : u_i \in \{0, 1\}, 2 \leq i \leq n\}$ and $V(AQ^1_{n-1}) = \{1v_2....v_n : v_i \in \{0, 1\}, 2 \leq i \leq n\}$. A vertex $u = 0u_2....u_n$ of $AQ^0_{n-1}$ is joined to a vertex $v = 1v_2....v_n $ of $AQ^1_{n-1}$ if and only if for every $i$, $2 \leq i \leq n$ either\ 1. $u_i = v_i$; in this case an edge $\langle u, v \rangle$ is called a hypercube edge and we say $v = u^h$, or\ 2. $u_i = \overline{v_i}$; in this case an edge $\langle u, v \rangle$ is called a complement edge and we say $v = u^c$.\ Let $E^h_n = \{\langle u, u^h \rangle : u \in V(AQ^0_{n-1})\}$ and $E^c_n = \{\langle u, u^c \rangle : u \in V(AQ^0_{n-1})\}$. See Fig.$1$.\ (136.25,55.25)(0,0) (5.5,16.25)[(0,1)[18.25]{}]{} (25,16.25)[(0,1)[18.25]{}]{} (63.75,16.25)[(0,1)[18.25]{}]{} (112,16.25)[(0,1)[18.25]{}]{} (47.75,16.25)[(0,1)[18.25]{}]{} (86.5,16.25)[(0,1)[18.25]{}]{} (134.75,16.25)[(0,1)[18.25]{}]{} (25.25,34.25)[(1,0)[22.5]{}]{} (64,34.25)[(1,0)[22.5]{}]{} (112.25,34.25)[(1,0)[22.5]{}]{} (25.25,16.75)[(1,0)[22.25]{}]{} (64,16.75)[(1,0)[22.25]{}]{} (112.25,16.75)[(1,0)[22.25]{}]{} (25,34)(.0439453125,-.0336914063)[512]{}[(1,0)[.0439453125]{}]{} (63.75,34)(.0439453125,-.0336914063)[512]{}[(1,0)[.0439453125]{}]{} (112,34)(.0439453125,-.0336914063)[512]{}[(1,0)[.0439453125]{}]{} (25.25,17)(.0439453125,.0336914063)[512]{}[(1,0)[.0439453125]{}]{} (64,17)(.0439453125,.0336914063)[512]{}[(1,0)[.0439453125]{}]{} (112.25,17)(.0439453125,.0336914063)[512]{}[(1,0)[.0439453125]{}]{} (86.75,34)[(1,0)[25.5]{}]{} (86.5,17.25)[(1,0)[25.5]{}]{} (64,34)(.097082495,-.0337022133)[497]{}[(1,0)[.097082495]{}]{} (86.5,16.75)(.0929672447,.0337186898)[519]{}[(1,0)[.0929672447]{}]{} (86.5,34)(.0942460317,-.0337301587)[504]{}[(1,0)[.0942460317]{}]{} (112,34)(-.0920303605,-.0336812144)[527]{}[(-1,0)[.0920303605]{}]{} (63.75,34.25)(101.5,49.25)(134.25,34.25) (64,16.5)(95.875,2.375)(134.25,16.75) (5.25,34.5) (5.5,15.5) (24.75,16.75) (25,33.5) (47.75,33.5) (47.75,16.25) (64.5,16.25) (63.75,34.25) (87,34.25) (112.5,34) (135,33.75) (134.5,17) (112.25,16.5) (86.5,17.25) (5.25,38.5)[(0,0)\[cc\][$\tiny{0}$]{}]{} (5.25,11)[(0,0)\[cc\][$\tiny{1}$]{}]{} (23.5,38.5)[(0,0)\[cc\][$\tiny{00}$]{}]{} (48.25,40.25)[(0,0)\[cc\][$\tiny{10}$]{}]{} (23,12.25)[(0,0)\[cc\][$\tiny{01}$]{}]{} (48.25,11.75)[(0,0)\[cc\][$\tiny{11}$]{}]{} (64.5,39.5)[(0,0)\[cc\][$\tiny{000}$]{}]{} (63.5,12.5)[(0,0)\[cc\][$\tiny{010}$]{}]{} (85.5,37.5)[(0,0)\[cc\][ $\tiny{001}$]{}]{} (85.5,12.5)[(0,0)\[cc\][$\tiny{011}$]{}]{} (112,38)[(0,0)\[cc\][$\tiny{101}$]{}]{} (136.25,38.25)[(0,0)\[cc\][$\tiny{100}$]{}]{} (136.25,12.25)[(0,0)\[cc\][$\tiny{110}$]{}]{} (111.75,13.25)[(0,0)\[cc\][$\tiny{111}$]{}]{} (5.25,6.75)[(0,0)\[cc\][$\tiny AQ_{1}$]{}]{} (35.25,6.75)[(0,0)\[cc\][$\tiny AQ_{2}$]{}]{} (96.75,5.5)[(0,0)\[cc\][$\tiny AQ_{3}$]{}]{} (70.5,.25)[(0,0)\[cc\] [Fig. 1. The augmented cubes of dimension 1, 2 and 3 ]{}]{} (52.75,25)[(0,0)\[cc\][$\tiny AQ^1_{1}$]{}]{} (21.5,26)[(0,0)\[cc\][$\tiny AQ^0_{1}$]{}]{} (67.5,25.25)[(0,0)\[cc\][$\tiny AQ^0_{2}$]{}]{} (130.75,25.5)[(0,0)\[cc\][$\tiny AQ^1_{2}$]{}]{} The following lemma is helpful to visualize edges of $AQ_n$.\ \[[@cs]\] Let $G$ be the simple graph with vertex set $V(G)= \{a_1a_2....a_n : a_i = 0$ or $ 1\}$ where two vertices $A= a_1a_2....a_n $ and $B = b_1b_2...b_n$ are joined iff there exists an integer $k: 1 \leq k \leq n$ such that either\ ($1$) $a_k = \overline b_k$ and $a_i = b_i$, for every $i$, $i \neq k$, or\ ($2$) $a_i = b_i$, for $1 \leq i \leq k-1$ and $a_i = \overline b_i$, for $k \leq i \leq n$. Construction of CISTs in augmented cubes ========================================= We need the following result given in [@h1] by Hasunuma.\ \[[@h1] \[lm1\]\] Let $k \geq 2$ be an integer, $T_1, T_2,.......,T_k$ are completely independent spanning trees in a graph $G$ if and only if they are edge-disjoint spanning trees of $G$ and for any $v \in V(G)$, there is at most one $T_i$ such that $d_{T_i}(v) > 1$. Pai and Chang[@p2] constructed two CISTs in several hypercube-variant networks, they proved the following result. \[[@p2] \[lm2\]\] Let $G_n$ be the $n-$dimensional variant hypercube for $n \geq 4$ and suppose that $T_1$ and $T_2$ are two CISTs of $G_n$. For $i \in \{1,2\} $, let $\overline T_i$ be a spanning tree of $G_{n+1}$ constructed from $T^0_i$ and $T^1_i$ by adding an edge $\langle u_i, v_i \rangle \in E(G_{n+1})$ to connect two internal vertices $u_i \in V(T^0_i)$ and $v_i \in V(T^1_i)$. Then, $\overline T_1$ and $\overline T_2$ are two CISTs of $G_{n+1}$. By using the same proof technique of above theorem one can state following corollary. \[ \[lm3\]\] Let $G_n$ be the $n-$dimensional variant hypercube for $n \geq 4$ and suppose that $T_i$, for $1 \leq i \leq k $ $(k < n)$ be $k$ CISTs of $G_n$. Let $\overline T_i$ be a spanning tree of $G_{n+1}$ constructed from $T^0_i$ and $T^1_i$ by adding an edge $\langle u_i, v_i \rangle \in E(G_{n+1})$ to connect two internal vertices $u_i \in V(T^0_i)$ and $v_i \in V(T^1_i)$. Then, $\overline T_i$ are $k$ CISTs of $G_{n+1}$ Firstly, we list out $n-1$ CISTs in $AQ_n$ for $n = 3, 4$. See Fig.$2$ and Fig.$3$.\ 1.2mm (101.75,50.75)(0,0) (23.25,9.75) (22.851,32.351) (78.25,10.851) (77.851,33.452) (22.25,32.5)(.0434783,-.0326087)[23]{}[(1,0)[.0434783]{}]{} (8.5,36.75) (23.531,45.781) (39.281,36.531) (93.281,37.781) (63.281,39.031) (9.531,15.031) (37.031,19.281) (36.281,9.781) (64.281,19.531) (92.531,19.781) (94.281,10.781) (23.5,45.5)[(0,-1)[12.75]{}]{} (23.5,32.75)[(4,1)[16]{}]{} (8.5,36.5)(.12394958,-.033613445)[119]{}[(1,0)[.12394958]{}]{} (9,15)(.081871345,-.033625731)[171]{}[(1,0)[.081871345]{}]{} (36.25,9.75)[(-1,0)[12.75]{}]{} (63,38.75)(.097315436,-.033557047)[149]{}[(1,0)[.097315436]{}]{} (93.25,37.75)(-.128151261,-.033613445)[119]{}[(-1,0)[.128151261]{}]{} (64,19.25)(.0591836735,-.0336734694)[245]{}[(1,0)[.0591836735]{}]{} (93.75,10.75)[(-1,0)[15.25]{}]{} (43.5,9.5)[(0,0)\[cc\][$110(7)$]{}]{} (44.25,20)[(0,0)\[cc\][$011(4)$]{}]{} (3.75,17.5)[(0,0)\[cc\][$101(6)$]{}]{} (2.5,32.75)[(0,0)\[cc\][$010(3)$]{}]{} (29.5,28.25)[(0,0)\[cc\][$000(1)$]{}]{} (41.25,39.75)[(0,0)\[cc\][$111(8)$]{}]{} (15.75,7.25)[(0,0)\[cc\][$100(5)$]{}]{} (70.5,9.5)[(0,0)\[cc\][$111(8)$]{}]{} (101.75,10.25)[(0,0)\[cc\][$110(7)$]{}]{} (99,21.75)[(0,0)\[cc\][$100(5)$]{}]{} (63.5,21.75)[(0,0)\[cc\][$101(6)$]{}]{} (85.25,29.5)[(0,0)\[cc\][$011(4)$]{}]{} (98.25,39.5)[(0,0)\[cc\][$001(2)$]{}]{} (62.5,42.25)[(0,0)\[cc\][$010(3)$]{}]{} (52.75,1.5)[(0,0)\[cc\][Fig$. 2. $ Two CISTs in $AQ_3$]{}]{} (78.531,48.531) (77.5,50.75)[(0,0)\[cc\][$000(1)$]{}]{} (22.25,48.25)[(0,0)\[cc\][$001(2)$]{}]{} (77.25,33.25)(.125,.03125)[8]{}[(1,0)[.125]{}]{} (78.25,33.75)[(0,1)[15.25]{}]{} (22.75,31.5)[(0,-1)[22.25]{}]{} (77.5,32.5)[(0,-1)[21.75]{}]{} (23.25,10)(.0481818182,.0336363636)[275]{}[(1,0)[.0481818182]{}]{} (78,10.5)(.0518181818,.0336363636)[275]{}[(1,0)[.0518181818]{}]{} (151.5,120.25)(0,0) (20.75,10.5) (20.271,29.771) (20.771,48.771) (20.27,68.27) (20.77,87.02) (7.5,14.75) (33.541,16.791) (7.291,29.291) (33.082,35.582) (6.79,39.791) (7.041,52.791) (33.082,54.832) (6.541,72.041) (32.582,74.082) (33.082,93.082) (20.25,29.75)[(0,-1)[19.5]{}]{} (20.25,49)[(0,-1)[19.75]{}]{} (6.75,39.75)(.0460992908,-.0336879433)[282]{}[(1,0)[.0460992908]{}]{} (74.252,10.771) (73.772,30.041) (73.772,49.042) (73.772,68.541) (74.272,87.291) (61.252,10.771) (87.042,10.812) (60.793,34.562) (86.582,34.603) (60.541,18.811) (60.793,53.812) (86.582,53.853) (60.293,73.061) (86.082,73.102) (60.793,92.061) (86.582,92.102) (73.752,30.021)[(0,-1)[19.5]{}]{} (73.752,49.27)[(0,-1)[19.75]{}]{} (134.252,11.271) (133.772,30.541) (133.772,49.54) (133.772,69.04) (134.272,87.791) (121.252,11.271) (147.04,18.061) (120.793,36.81) (146.582,36.852) (120.54,20.311) (120.793,56.06) (146.582,56.101) (120.293,75.311) (120.793,94.311) (146.582,94.351) (133.752,30.52)[(0,-1)[19.5]{}]{} (133.752,49.77)[(0,-1)[19.75]{}]{} (73.5,48.75)(.03125,4.8125)[8]{}[(0,1)[4.8125]{}]{} (60.5,18.5)(.058823529,-.033613445)[238]{}[(1,0)[.058823529]{}]{} (134,87.75)[(0,-1)[39]{}]{} (134.54,99.041) (134,99)[(0,-1)[11]{}]{} (120.25,20.25)(.0535714286,-.0337301587)[252]{}[(1,0)[.0535714286]{}]{} (20,5.25)[(0,0)\[cc\][$0100 (5)$]{}]{} (5,16.75)[(0,0)\[cc\][$0111 (8)$]{}]{} (35.75,19.75)[(0,0)\[cc\][$1100 (13)$]{}]{} (28.5,28.5)[(0,0)\[cc\][$0101 (6)$]{}]{} (40.5,37.75)[(0,0)\[cc\][$0110 (7)$]{}]{} (4.75,24.75)[(0,0)\[cc\][$0010 (3)$]{}]{} (4.25,35)[(0,0)\[cc\][$0001 (2)$]{}]{} (28.75,47)[(0,0)\[cc\][$1010 (11)$]{}]{} (37.75,57)[(0,0)\[cc\][$1101 (14)$]{}]{} (5.25,55)[(0,0)\[cc\][$1110 (15)$]{}]{} (28.75,65.5)[(0,0)\[cc\][$1011 (12)$]{}]{} (38.25,76.25)[(0,0)\[cc\][$1001 (10)$]{}]{} (5,74.75)[(0,0)\[cc\][$0011 (4)$]{}]{} (20.75,86)(-.03125,-4.75)[8]{}[(0,-1)[4.75]{}]{} (28.75,86.25)[(0,0)\[cc\][$1000 (9)$]{}]{} (38.75,95.5)[(0,0)\[cc\][$1111 (16)$]{}]{} (74,6)[(0,0)\[cc\][$1111 (16)$]{}]{} (54.5,12.75)[(0,0)\[cc\][$1011 (12)$]{}]{} (61,20.5)[(0,0)\[cc\][$1101 (14)$]{}]{} (91,13.25)[(0,0)\[cc\][$0000 (1)$]{}]{} (81.75,27.5)[(0,0)\[cc\][$0111 (8)$]{}]{} (91.75,37)[(0,0)\[cc\][$0101 (6)$]{}]{} (60,36.75)[(0,0)\[cc\][$1000 (9)$]{}]{} (82.25,45.25)[(0,0)\[cc\][$0011 (4)$]{}]{} (91.75,56.25)[(0,0)\[cc\][$1100 (13)$]{}]{} (62.25,55.5)[(0,0)\[cc\][$0100 (5)$]{}]{} (81,64.5)[(0,0)\[cc\][$0001 (2)$]{}]{} (61.75,75.5)[(0,0)\[cc\][$0010 (3)$]{}]{} (90.25,75.75)[(0,0)\[cc\][$1110 (15)$]{}]{} (81.75,83.75)[(0,0)\[cc\][$1001 (10)$]{}]{} (61.75,95)[(0,0)\[cc\][$0110 (7)$]{}]{} (90.75,95.25)[(0,0)\[cc\][$1010 (11)$]{}]{} (133.75,5.5)[(0,0)\[cc\][$0110 (7)$]{}]{} (112.25,10.25)[(0,0)\[cc\][$0001 (2)$]{}]{} (112.25,20.25)[(0,0)\[cc\][$0100 (5)$]{}]{} (151.5,19.5)[(0,0)\[cc\][$0111 (8)$]{}]{} (141.5,25.75)[(0,0)\[cc\][$1110 (15)$]{}]{} (120.25,39.75)[(0,0)\[cc\][$1001 (10)$]{}]{} (151,39.75)[(0,0)\[cc\][$1111 (16)$]{}]{} (140.25,45)[(0,0)\[cc\][$1100 (13)$]{}]{} (120.5,58.5)[(0,0)\[cc\][$1000 (9)$]{}]{} (150.75,58.75)[(0,0)\[cc\][$1011 (12)$]{}]{} (143,68.75)[(0,0)\[cc\][$1101 (14)$]{}]{} (120,78)[(0,0)\[cc\][$0101 (6)$]{}]{} (140.5,82)[(0,0)\[cc\][$0010 (3)$]{}]{} (120.25,97.25)[(0,0)\[cc\][$0000 (1)$]{}]{} (134,102.25)[(0,0)\[cc\][$0011 (4)$]{}]{} (150.75,97.75)[(0,0)\[cc\][$1010 (11)$]{}]{} (73.25,.75)[(0,0)\[cc\][Fig$. 3.$ Three CISTs in $AQ_4$]{}]{} (21.25,87)[(0,1)[0]{}]{} (8,96.25)[(0,0)\[cc\][$0000 (1)$]{}]{} (7.5,29.5)[(1,0)[12.5]{}]{} (21,87)(.063172043,.033602151)[186]{}[(1,0)[.063172043]{}]{} (20.5,68.75)(.07885906,.033557047)[149]{}[(1,0)[.07885906]{}]{} (20.5,49.25)(.076219512,.033536585)[164]{}[(1,0)[.076219512]{}]{} (20,29.5)(.069892473,.033602151)[186]{}[(1,0)[.069892473]{}]{} (33,35.75)[(0,1)[0]{}]{} (20.75,11)(.070224719,.033707865)[178]{}[(1,0)[.070224719]{}]{} (6.5,72.25)(.103174603,-.033730159)[126]{}[(1,0)[.103174603]{}]{} (7,53)(.122596154,-.033653846)[104]{}[(1,0)[.122596154]{}]{} (7.5,14.75)(.109243697,-.033613445)[119]{}[(1,0)[.109243697]{}]{} (60.75,92)(.092198582,-.033687943)[141]{}[(1,0)[.092198582]{}]{} (74,87.5)(.088652482,.033687943)[141]{}[(1,0)[.088652482]{}]{} (60,73)(.104477612,-.03358209)[134]{}[(1,0)[.104477612]{}]{} (85.75,73)(-.089552239,-.03358209)[134]{}[(-1,0)[.089552239]{}]{} (60.5,53.5)(.08557047,-.033557047)[149]{}[(1,0)[.08557047]{}]{} (86.5,53.75)(-.090425532,-.033687943)[141]{}[(-1,0)[.090425532]{}]{} (60.25,34.25)[(3,-1)[13.5]{}]{} (86.25,34.5)(-.097014925,-.03358209)[134]{}[(-1,0)[.097014925]{}]{} (120.75,94.25)(.065920398,-.03358209)[201]{}[(1,0)[.065920398]{}]{} (146.5,94.5)(-.058894231,-.033653846)[208]{}[(-1,0)[.058894231]{}]{} (120.25,75.25)(.07253886,-.033678756)[193]{}[(1,0)[.07253886]{}]{} (120.75,56)(.06865285,-.033678756)[193]{}[(1,0)[.06865285]{}]{} (146.25,56.25)(-.063432836,-.03358209)[201]{}[(-1,0)[.063432836]{}]{} (120.25,36.75)(.06840796,-.03358209)[201]{}[(1,0)[.06840796]{}]{} (146.5,36.75)(-.066062176,-.033678756)[193]{}[(-1,0)[.066062176]{}]{} (121.25,11.25)[(1,0)[12.75]{}]{} (146.75,18)(-.063471503,-.033678756)[193]{}[(-1,0)[.063471503]{}]{} (61,10.75)[(1,0)[13.25]{}]{} (86.75,11)[(-1,0)[12.25]{}]{} (6.791,94.04) (6.5,94.25)(.066105769,-.033653846)[208]{}[(1,0)[.066105769]{}]{} : To make things readable, we denote vertices of $AQ_5$ by using numbers $1,2,.....,32$. For example the vertex $00000$ will be denoted by $1$ and sometime will be written as $00000(1)$. Also, we will use short forms to denote edges of $AQ_5$, for example the edge $\langle 00000(1), 00001(2) \rangle$ will be denoted by $\langle 1, 2 \rangle$. For every $i$, $InV(T_i)$ denotes set of internal vertices of the tree $T_i$, for $1 \leq i \leq 4$. Vertex $v \in V(T_i)$ is called internal if $d_{T_i}(v) \geq 2$.\ Let $ V(AQ_5) = \{ 00000(1), 00001(2), 00010(3), 00011(4), 00100(5), 00101(6), 00110(7), 00111(8),$ $\hspace{2.3cm} 01000(9), 01001(10), 01010(11), 01011(12), 01100(13), 01101(14), 01110(15), $ $\hspace{2.3cm} 01111(16), 10000(17), 10001(18), 10010(19), 10011(20), 10100(21), 10101(22), $ $ \hspace{2.3cm} 10110(23), 10111(24), 11000(25), 11001(26), 11010(27), 11011(28), 11100(29),$ $\hspace{2.3cm} 11101(30), 11110(31), 11111(32) \}.$\ Now for $n = 5$, we construct four trees $T_1, T_2, T_3$ and $T_4$ as shown in Fig.$4(a)$, Fig.$4(b)$, Fig.$4(c)$ and Fig.$4(d)$ respectively.\ $$%TeXCAD Picture [Fig.4(a) Spanning Tree T_1 in AQ_5.pic]. Options: %\grade{\on} %\emlines{\off} %\epic{\off} %\beziermacro{\on} %\reduce{\on} %\snapping{\off} %\pvinsert{% Your \input, \def, etc. here} %\quality{8.000} %\graddiff{0.005} %\snapasp{1} %\zoom{4.0000} \unitlength 1mm % = 2.845pt \linethickness{0.4pt} \ifx\plotpoint\undefined\newsavebox{\plotpoint}\fi % GNUPLOT compatibility \begin{picture}(100.798,110.397)(0,0) \put(33.706,13.409){\circle*{2.268}} \put(33.48,29.511){\circle*{2.268}} \put(33.48,61.943){\circle*{2.268}} \put(33.706,78.954){\circle*{2.268}} \put(33.48,95.057){\circle*{2.268}} \put(33.706,111.386){\circle*{2.268}} \put(33.253,14.316){\line(0,1){15.422}} \put(33.253,29.965){\line(0,1){16.557}} \put(33.253,46.522){\line(0,1){16.102}} \put(33.253,62.624){\line(0,1){17.01}} \put(33.253,79.634){\line(0,1){16.103}} \put(33.253,95.737){\line(0,1){16.556}} \put(50.716,11.594){\circle*{1.434}} \put(50.3,19.115){\circle*{1.434}} \put(15.109,11.404){\circle*{1.434}} \put(14.692,18.926){\circle*{1.434}} \put(50.489,27.017){\circle*{1.434}} \put(50.072,34.538){\circle*{1.434}} \put(14.881,26.826){\circle*{1.434}} \put(14.465,34.348){\circle*{1.434}} \put(51.434,46.331){\circle*{1.434}} \put(13.522,57.671){\circle*{1.434}} \put(13.105,65.192){\circle*{1.434}} \put(52.34,62.434){\circle*{1.434}} \put(51.623,78.727){\circle*{2.268}} \put(33.933,78.954){\line(1,0){17.463}} \put(13.066,75.588){\circle*{1.434}} \put(12.65,83.11){\circle*{1.434}} \put(66.138,72.186){\circle*{1.434}} \put(65.268,79.028){\circle*{1.434}} \put(65.91,87.608){\circle*{1.434}} \put(49.129,91.882){\circle*{1.434}} \put(48.712,99.402){\circle*{1.434}} \put(13.522,91.691){\circle*{1.434}} \put(13.104,99.213){\circle*{1.434}} \put(51.433,118.68){\circle*{1.434}} \put(51.434,111.196){\circle*{1.434}} \put(13.785,111.877){\circle*{1.434}} %\emline(33.933,13.635)(50.49,11.821) \multiput(33.933,13.635)(.30660741,-.0336){54}{\line(1,0){.30660741}} %\end \put(33.706,13.862){\line(3,1){16.33}} %\emline(33.706,29.738)(50.262,27.017) \multiput(33.706,29.738)(.20439506,-.0336){81}{\line(1,0){.20439506}} %\end %\emline(33.48,29.965)(50.036,34.728) \multiput(33.48,29.965)(.116591549,.033543662){142}{\line(1,0){.116591549}} %\end \put(33.933,46.522){\line(1,0){17.237}} \put(33.48,62.398){\line(1,0){18.824}} %\emline(33.933,95.283)(48.902,91.882) \multiput(33.933,95.283)(.148205941,-.033679208){101}{\line(1,0){.148205941}} %\end %\emline(33.706,95.51)(48.902,99.366) \multiput(33.706,95.51)(.132132174,.033523478){115}{\line(1,0){.132132174}} %\end \put(33.933,111.613){\line(1,0){17.237}} %\emline(33.706,112.066)(51.17,118.87) \multiput(33.706,112.066)(.086455446,.033683168){202}{\line(1,0){.086455446}} %\end %\emline(33.48,95.283)(12.841,99.366) \multiput(33.48,95.283)(-.170571901,.033738843){121}{\line(-1,0){.170571901}} %\end %\emline(33.253,95.057)(13.294,91.654) \multiput(33.253,95.057)(-.197607921,-.033687129){101}{\line(-1,0){.197607921}} %\end %\emline(33.48,78.954)(12.161,82.81) \multiput(33.48,78.954)(-.185384348,.033530435){115}{\line(-1,0){.185384348}} %\end %\emline(33.48,78.727)(12.614,75.325) \multiput(33.48,78.727)(-.206590099,-.033687129){101}{\line(-1,0){.206590099}} %\end %\emline(32.799,62.17)(12.841,65.346) \multiput(32.799,62.17)(-.210088421,.033423158){95}{\line(-1,0){.210088421}} %\end %\emline(33.48,61.717)(13.294,57.408) \multiput(33.48,61.717)(-.1577,-.0336625){128}{\line(-1,0){.1577}} %\end %\emline(33.253,29.965)(13.975,34.501) \multiput(33.253,29.965)(-.142797037,.0336){135}{\line(-1,0){.142797037}} %\end %\emline(33.253,29.738)(15.109,26.79) \multiput(33.253,29.738)(-.20618182,-.03350909){88}{\line(-1,0){.20618182}} %\end %\emline(33.48,13.409)(14.429,19.078) \multiput(33.48,13.409)(-.112728994,.033547929){169}{\line(-1,0){.112728994}} %\end %\emline(33.706,13.409)(14.882,11.367) \multiput(33.706,13.409)(-.30859016,-.03346885){61}{\line(-1,0){.30859016}} %\end %\emline(51.397,79.407)(65.686,88.026) \multiput(51.397,79.407)(.055815625,.033665625){256}{\line(1,0){.055815625}} %\end \put(51.85,78.5){\line(2,-1){14.062}} \put(52.078,78.954){\line(1,0){12.927}} \put(33.026,8.192){\makebox(0,0)[cc]{$11110(31)$}} \put(60.015,11.367){\makebox(0,0)[cc]{$11111(32)$}} \put(59.201,19.306){\makebox(0,0)[cc]{$11101(30)$}} \put(6.171,10.687){\makebox(0,0)[cc]{$11100(29)$}} \put(6.398,19.078){\makebox(0,0)[cc]{$10110(23)$}} \put(25.676,24.749){\makebox(0,0)[cc]{$11010(27)$}} \put(60.108,27.017){\makebox(0,0)[cc]{$10101(22)$}} \put(59.108,34.954){\makebox(0,0)[cc]{$10010(19)$}} \put(61.15,46.294){\makebox(0,0)[cc]{$01000(9)$}} \put(12.614,36.542){\makebox(0,0)[cc]{$01010(11)$}} \put(6.171,26.563){\makebox(0,0)[cc]{$11001(26)$}} \put(61.922,62.398){\makebox(0,0)[cc]{$10111(24)$}} \put(5.398,57.862){\makebox(0,0)[cc]{$10011(20)$}} \put(10.8,66.933){\makebox(0,0)[cc]{$10001(18)$}} \put(24.542,46.522){\makebox(0,0)[cc]{$11000(25)$}} \put(41.964,58.315){\makebox(0,0)[cc]{$10000(17)$}} \put(26.449,73.057){\makebox(0,0)[cc]{$00000(1)$ }} \put(47.994,73.51){\makebox(0,0)[cc]{$01111(16)$}} \put(76.798,71.47){\makebox(0,0)[cc]{$01100(13)$}} \put(74.758,78.954){\makebox(0,0)[cc]{$01101(14)$}} \put(76.345,87.572){\makebox(0,0)[cc]{$01110(15)$}} \put(10.573,77.82){\makebox(0,0)[cc]{$00010(3)$ }} \put(10.346,84.397){\makebox(0,0)[cc]{$00001(2)$}} \put(39.15,90.52){\makebox(0,0)[cc]{$00100(5)$}} \put(58.52,92.562){\makebox(0,0)[cc]{$10100(21)$}} \put(59.067,100.046){\makebox(0,0)[cc]{$11011(28)$}} \put(11.026,93.469){\makebox(0,0)[cc]{$00011(4)$ }} \put(9.439,101.18){\makebox(0,0)[cc]{$01011(12)$}} \put(26.902,114.077){\makebox(0,0)[cc]{$00110(7)$}} \put(11.026,108.21){\makebox(0,0)[cc]{$00101(6)$}} \put(59.334,110.706){\makebox(0,0)[cc]{$00111(8)$}} \put(60.428,118.643){\makebox(0,0)[cc]{$01001(10)$}} \put(32.572,1.615){\makebox(0,0)[cc]{Fig. 4(a). Spanning Tree $T_1$ in $AQ_5$}} \put(33.253,111.84){\line(-1,0){19.505}} \put(33.134,46.134){\circle*{2.268}} \end{picture}$$ (81.25,111.665)(0,0) (51.65,9.4) (51.816,23.966) (51.816,37.366) (51.983,51.933) (52.216,67.566) (52.383,82.133) (52.383,95.533) (73.65,8.2) (73.683,15.233) (30.283,9.233) (30.316,16.266) (31.416,36.766) (15.683,30.832) (15.716,37.066) (15.882,44.433) (72.883,31.833) (72.916,38.866) (73.082,45.433) (73.082,23.833) (32.482,50.433) (32.649,56.998) (72.882,55.433) (32.683,67.632) (73.082,67.832) (73.45,77.633) (73.483,84.666) (32.283,78.466) (32.316,85.498) (32.283,95.832) (32.083,105.033) (73.283,95.832) (73.083,105.033) (51.85,23.8)[(0,-1)[14.2]{}]{} (51.85,37.2)[(0,-1)[13.4]{}]{} (51.65,52)[(0,-1)[14.2]{}]{} (52.25,67.4)[(0,-1)[16.6]{}]{} (52.45,82.4)[(0,-1)[14.6]{}]{} (52.25,95.4)[(0,-1)[13.2]{}]{} (31.45,36.6)[(0,1)[.2]{}]{} (31.05,37.2)[(1,0)[20.4]{}]{} (51.65,9.6)(.127906977,.03372093)[172]{}[(1,0)[.127906977]{}]{} (52.05,9.6)(.51428571,-.03333333)[42]{}[(1,0)[.51428571]{}]{} (51.25,9.6)(-.104950495,.033663366)[202]{}[(-1,0)[.104950495]{}]{} (51.65,24)[(1,0)[21]{}]{} (31.45,36.8)(-.068103448,.03362069)[232]{}[(-1,0)[.068103448]{}]{} (31.05,37)[(-1,0)[15.8]{}]{} (31.05,36.4)(-.090697674,-.03372093)[172]{}[(-1,0)[.090697674]{}]{} (52.25,37.6)(.11744186,-.03372093)[172]{}[(1,0)[.11744186]{}]{} (51.85,37.6)(.4952381,.03333333)[42]{}[(1,0)[.4952381]{}]{} (51.45,37.6)(.092241379,.03362069)[232]{}[(1,0)[.092241379]{}]{} (51.65,52)(-.12885906,.033557047)[149]{}[(-1,0)[.12885906]{}]{} (51.65,51.6)[(0,1)[.4]{}]{} (51.65,52)(-.40833333,-.03333333)[48]{}[(-1,0)[.40833333]{}]{} (51.85,51.8)(.184070796,.033628319)[113]{}[(1,0)[.184070796]{}]{} (52.05,67.6)[(1,0)[20.8]{}]{} (51.85,67.4)[(-1,0)[19.4]{}]{} (52.05,82.4)(.27179487,.03333333)[78]{}[(1,0)[.27179487]{}]{} (52.45,82.2)(.154744526,-.033576642)[137]{}[(1,0)[.154744526]{}]{} (52.65,82.2)(-.214736842,.033684211)[95]{}[(-1,0)[.214736842]{}]{} (52.25,82)(-.190654206,-.03364486)[107]{}[(-1,0)[.190654206]{}]{} (52.45,95.6)(.0731182796,.0336917563)[279]{}[(1,0)[.0731182796]{}]{} (52.65,95.8)[(1,0)[20.6]{}]{} (52.45,96)[(-1,0)[20.4]{}]{} (52.05,95.8)(-.0724014337,.0336917563)[279]{}[(-1,0)[.0724014337]{}]{} (51.25,4.8)[(0,0)\[cc\][$11101(30)$]{}]{} (84.05,7.8)[(0,0)\[cc\][$11111(32)$]{}]{} (83.85,16)[(0,0)\[cc\][$10010(19)$]{}]{} (18.85,7.8)[(0,0)\[cc\][$11010(27)$]{}]{} (21.25,15.6)[(0,0)\[cc\][$01101(14)$]{}]{} (82.25,24)[(0,0)\[cc\][$11110(31)$]{}]{} (43.25,34.6)[(0,0)\[cc\][$10110(23)$]{}]{} (81.85,32)[(0,0)\[cc\][$10101(22)$]{}]{} (82.25,38.6)[(0,0)\[cc\][$10001(18)$]{}]{} (82.85,45.8)[(0,0)\[cc\][$00110(7)$]{}]{} (33.25,40.2)[(0,0)\[cc\][$10100(21)$]{}]{} (6.45,30)[(0,0)\[cc\][$10000(17)$]{}]{} (6.45,36.8)[(0,0)\[cc\][$10011(20)$]{}]{} (6.25,44.4)[(0,0)\[cc\][$11011(28)$]{}]{} (62.05,50)[(0,0)\[cc\][$01001(10)$]{}]{} (82.05,55.4)[(0,0)\[cc\][$01110(15)$]{}]{} (23.65,50.2)[(0,0)\[cc\][$01000(9)$]{}]{} (23.05,56.8)[(0,0)\[cc\][$01010(11)$]{}]{} (60.25,65)[(0,0)\[cc\][$00001(2)$]{}]{} (82.45,68)[(0,0)\[cc\][$00101(6)$]{}]{} (24.65,66.8)[(0,0)\[cc\][$00010(3)$]{}]{} (44.85,77.2)[(0,0)\[cc\][$00011(4)$]{}]{} (83.25,84.8)[(0,0)\[cc\][$01011(12)$]{}]{} (83.45,77.6)[(0,0)\[cc\][$01100(13)$]{}]{} (23.65,85.4)[(0,0)\[cc\][$00000(1)$]{}]{} (23.65,78.2)[(0,0)\[cc\][$11100(29)$]{}]{} (44.45,93.2)[(0,0)\[cc\][$00111(8)$]{}]{} (82.25,105.6)[(0,0)\[cc\][$00100(5)$]{}]{} (82.25,96)[(0,0)\[cc\][$01111(16)$]{}]{} (23.25,105)[(0,0)\[cc\][$10111(24)$]{}]{} (23.25,95.4)[(0,0)\[cc\][$11000(25)$]{}]{} (50.65,1.0)[(0,0)\[cc\][Fig. 4(b). Spanning Tree $T_2$ in $AQ_5$]{}]{} (42.25,22.6)[(0,0)\[cc\][$11001(26)$]{}]{} (29.85,9.2)[(1,0)[21.6]{}]{} $$%TeXCAD Picture [Fig.4(c) Spanning Tree T_3 in AQ_5.pic]. Options: %\grade{\on} %\emlines{\off} %\epic{\off} %\beziermacro{\on} %\reduce{\on} %\snapping{\off} %\pvinsert{% Your \input, \def, etc. here} %\quality{8.000} %\graddiff{0.005} %\snapasp{1} %\zoom{4.0000} \unitlength 1mm % = 2.845pt \linethickness{0.4pt} \ifx\plotpoint\undefined\newsavebox{\plotpoint}\fi % GNUPLOT compatibility \begin{picture}(92.73,115.739)(0,0) \put(48.09,10.92){\circle*{2.036}} \put(47.668,31.558){\circle*{2.036}} \put(47.91,31.08){\line(0,-1){20.34}} \put(47.73,54.418){\circle*{2.036}} \put(47.308,75.056){\circle*{2.036}} \put(47.55,74.578){\line(0,-1){20.34}} \put(47.488,94.738){\circle*{2.036}} \put(47.55,94.44){\line(0,-1){19.08}} \put(47.73,54.12){\line(0,-1){22.5}} \put(28.408,54.418){\circle*{2.036}} \put(28.228,74.938){\circle*{2.036}} \put(68.008,74.938){\circle*{2.036}} \put(29.19,14.7){\circle*{1.138}} \put(29.22,8.07){\circle*{1.138}} \put(67.74,10.95){\circle*{1.138}} \put(67.38,31.47){\circle*{1.138}} \put(66.66,55.379){\circle*{1.138}} \put(66.689,48.749){\circle*{1.138}} \put(66.84,61.17){\circle*{1.138}} \put(11.759,53.579){\circle*{1.138}} \put(11.788,46.949){\circle*{1.138}} \put(11.939,59.37){\circle*{1.138}} \put(12.12,76.439){\circle*{1.138}} \put(12.149,69.809){\circle*{1.138}} \put(12.3,82.23){\circle*{1.138}} \put(12.66,87.63){\circle*{1.138}} \put(60.54,84.75){\circle*{1.138}} \put(85.38,75.719){\circle*{1.138}} \put(85.409,69.089){\circle*{1.138}} \put(85.56,81.51){\circle*{1.138}} \put(31.38,100.379){\circle*{1.138}} \put(31.409,93.749){\circle*{1.138}} \put(31.56,106.17){\circle*{1.138}} \put(67.02,102.539){\circle*{1.138}} \put(67.049,95.909){\circle*{1.138}} %\emline(29.19,14.88)(47.73,10.92) \multiput(29.19,14.88)(.157118644,-.033559322){118}{\line(1,0){.157118644}} %\end %\emline(29.19,7.86)(47.91,10.74) \multiput(29.19,7.86)(.21767442,.03348837){86}{\line(1,0){.21767442}} %\end \put(48.27,11.1){\line(1,0){19.26}} \put(47.73,31.44){\line(1,0){19.8}} \put(47.73,31.44){\line(-1,0){19.08}} %\emline(47.73,54.3)(66.45,48.9) \multiput(47.73,54.3)(.116273292,-.033540373){161}{\line(1,0){.116273292}} %\end %\emline(47.37,54.3)(66.45,55.02) \multiput(47.37,54.3)(.8672727,.0327273){22}{\line(1,0){.8672727}} %\end %\emline(66.63,60.96)(47.55,54.84) \multiput(66.63,60.96)(-.104835165,-.033626374){182}{\line(-1,0){.104835165}} %\end \put(28.47,54.3){\line(1,0){18.9}} %\emline(11.91,59.16)(28.29,54.66) \multiput(11.91,59.16)(.122238806,-.03358209){134}{\line(1,0){.122238806}} %\end %\emline(11.73,53.58)(27.93,54.3) \multiput(11.73,53.58)(.7363636,.0327273){22}{\line(1,0){.7363636}} %\end %\emline(11.91,47.1)(28.29,53.94) \multiput(11.91,47.1)(.080689655,.033694581){203}{\line(1,0){.080689655}} %\end %\emline(12.27,69.78)(28.83,74.64) \multiput(12.27,69.78)(.114206897,.033517241){145}{\line(1,0){.114206897}} %\end %\emline(12.09,76.26)(28.47,75) \multiput(12.09,76.26)(.43105263,-.03315789){38}{\line(1,0){.43105263}} %\end %\emline(12.27,82.2)(29.01,75.18) \multiput(12.27,82.2)(.080095694,-.033588517){209}{\line(1,0){.080095694}} %\end %\emline(12.63,87.6)(28.83,75.36) \multiput(12.63,87.6)(.0446280992,-.0337190083){363}{\line(1,0){.0446280992}} %\end \put(28.47,75){\line(1,0){19.26}} \put(47.55,75){\line(1,0){20.52}} %\emline(68.07,75.18)(85.35,81.48) \multiput(68.07,75.18)(.092406417,.03368984){187}{\line(1,0){.092406417}} %\end %\emline(67.89,75.36)(84.81,75.72) \multiput(67.89,75.36)(1.538182,.032727){11}{\line(1,0){1.538182}} %\end %\emline(67.71,75.18)(85.17,69.06) \multiput(67.71,75.18)(.095934066,-.033626374){182}{\line(1,0){.095934066}} %\end %\emline(47.37,75.18)(60.15,84.54) \multiput(47.37,75.18)(.045971223,.0336690647){278}{\line(1,0){.045971223}} %\end %\emline(47.73,94.62)(66.99,102.36) \multiput(47.73,94.62)(.08373913,.033652174){230}{\line(1,0){.08373913}} %\end %\emline(47.55,94.8)(66.99,95.7) \multiput(47.55,94.8)(.72,.0333333){27}{\line(1,0){.72}} %\end %\emline(31.53,106.32)(47.55,95.16) \multiput(31.53,106.32)(.0483987915,-.0337160121){331}{\line(1,0){.0483987915}} %\end %\emline(31.35,100.2)(47.73,94.98) \multiput(31.35,100.2)(.105677419,-.033677419){155}{\line(1,0){.105677419}} %\end %\emline(31.17,93.9)(47.19,94.62) \multiput(31.17,93.9)(.7281818,.0327273){22}{\line(1,0){.7281818}} %\end \put(49.55,7.86){\makebox(0,0)[cc]{$10010(19)$}} \put(76.55,11.1){\makebox(0,0)[cc]{$10101(22)$}} \put(20.07,14.52){\makebox(0,0)[cc]{$10110(23)$}} \put(20.53,7.5){\makebox(0,0)[cc]{$10000(17)$}} \put(40.13,28.28){\makebox(0,0)[cc]{$10011(20)$}} \put(76.47,30.36){\makebox(0,0)[cc]{$10111(24)$}} \put(20.25,30.18){\makebox(0,0)[cc]{$00011(4)$}} \put(39.51,50.96){\makebox(0,0)[cc]{$11100(29)$}} \put(76.21,61.14){\makebox(0,0)[cc]{$10100(21)$}} \put(76.57,55.38){\makebox(0,0)[cc]{$11000(25)$}} \put(76.57,48.18){\makebox(0,0)[cc]{$11101(30)$}} \put(28.83,58.08){\makebox(0,0)[cc]{$11011(28)$}} \put(2.61,59.52){\makebox(0,0)[cc]{$11001(26)$}} \put(2.97,53.4){\makebox(0,0)[cc]{$11010(27)$}} \put(2.25,46.38){\makebox(0,0)[cc]{$11111(32)$}} \put(39.33,72.48){\makebox(0,0)[cc]{$01100(13)$}} \put(69.27,85.26){\makebox(0,0)[cc]{$00100(5)$}} \put(65.01,70.94){\makebox(0,0)[cc]{$01101(14)$}} \put(94.73,81.66){\makebox(0,0)[cc]{$00010(3)$}} \put(94.65,75.36){\makebox(0,0)[cc]{$00101(6)$}} \put(94.93,68.7){\makebox(0,0)[cc]{$01001(10)$}} \put(34.27,78.42){\makebox(0,0)[cc]{$01110(15)$}} \put(3.71,87.96){\makebox(0,0)[cc]{$00001(2)$ }} \put(3.61,82.02){\makebox(0,0)[cc]{$00110(7)$}} \put(3.61,75.9){\makebox(0,0)[cc]{$10001(18)$}} \put(3.97,69.42){\makebox(0,0)[cc]{$11110(31)$}} \put(56.77,92.1){\makebox(0,0)[cc]{$01000(9)$}} \put(76.93,102.72){\makebox(0,0)[cc]{$00000(1)$ }} \put(76.29,95.7){\makebox(0,0)[cc]{$00111(8)$}} \put(23.59,106.32){\makebox(0,0)[cc]{$01010(11)$}} \put(23.41,99.84){\makebox(0,0)[cc]{$01011(12)$}} \put(23.23,93.54){\makebox(0,0)[cc]{$01111(16)$}} \put(47.01,2.46){\makebox(0,0)[cc]{Fig. 4(c). Spanning Tree $T_3$ in $AQ_5$}} \put(28.86,31.65){\circle*{1.138}} \end{picture}$$ $$%TeXCAD Picture [Fig.4(d) Spanning Tree T_4 in AQ_5.pic]. Options: %\grade{\on} %\emlines{\off} %\epic{\off} %\beziermacro{\on} %\reduce{\on} %\snapping{\off} %\pvinsert{% Your \input, \def, etc. here} %\quality{8.000} %\graddiff{0.005} %\snapasp{1} %\zoom{4.0000} \unitlength 1mm % = 2.845pt \linethickness{0.4pt} \ifx\plotpoint\undefined\newsavebox{\plotpoint}\fi % GNUPLOT compatibility \begin{picture}(100.7,110.48)(0,0) \put(56.42,11.435){\circle*{2.305}} \put(56.133,32.388){\circle*{2.305}} \put(56.6,53.808){\circle*{2.305}} \put(56.313,74.76){\circle*{2.305}} \put(55.593,95.028){\circle*{2.305}} \put(35.613,53.988){\circle*{2.305}} \put(35.433,74.148){\circle*{2.305}} \put(78.633,74.688){\circle*{2.305}} \put(72.62,18.095){\circle*{1.61}} \put(72.885,7.92){\circle*{1.61}} \put(37.065,10.62){\circle*{1.61}} \put(72.885,40.235){\circle*{1.61}} \put(73.15,30.06){\circle*{1.61}} \put(35.445,32.04){\circle*{1.61}} \put(72.705,62.015){\circle*{1.61}} \put(72.97,51.84){\circle*{1.61}} \put(22.305,39.06){\circle*{1.61}} \put(16.005,46.62){\circle*{1.61}} \put(12.945,53.28){\circle*{1.61}} \put(13.665,61.02){\circle*{1.61}} \put(19.425,67.14){\circle*{1.61}} \put(94.125,85.86){\circle*{1.61}} \put(95.565,74.88){\circle*{1.61}} \put(93.405,66.24){\circle*{1.61}} \put(77.745,104.675){\circle*{1.61}} \put(77.83,95.22){\circle*{1.61}} \put(38.865,104.675){\circle*{1.61}} \put(39.13,94.5){\circle*{1.61}} \put(17.265,78.3){\circle*{1.61}} \put(25.725,87.3){\circle*{1.61}} \put(56.24,11.615){\line(0,1){20.88}} \put(56.24,53.735){\line(0,-1){21.42}} \put(56.24,74.435){\line(0,-1){20.7}} \put(56.24,74.075){\line(0,1){0}} \put(78.74,74.615){\line(-1,0){22.68}} %\emline(56.24,53.915)(72.26,62.375) \multiput(56.24,53.915)(.0638247012,.0337051793){251}{\line(1,0){.0638247012}} %\end %\emline(56.42,53.735)(72.8,51.575) \multiput(56.42,53.735)(.252,-.03323077){65}{\line(1,0){.252}} %\end %\emline(78.56,74.435)(93.32,65.975) \multiput(78.56,74.435)(.0588047809,-.0337051793){251}{\line(1,0){.0588047809}} %\end \put(78.2,74.975){\line(1,0){17.46}} %\emline(78.56,74.975)(93.86,85.955) \multiput(78.56,74.975)(.0469325153,.0336809816){326}{\line(1,0){.0469325153}} %\end \put(55.52,95.135){\line(1,0){22.5}} %\emline(54.98,95.495)(77.66,105.215) \multiput(54.98,95.495)(.0784775087,.033633218){289}{\line(1,0){.0784775087}} %\end \put(39.32,94.415){\line(1,0){16.38}} %\emline(38.6,105.035)(55.34,95.135) \multiput(38.6,105.035)(.0569387755,-.0336734694){294}{\line(1,0){.0569387755}} %\end %\emline(25.64,87.215)(35.36,74.615) \multiput(25.64,87.215)(.033633218,-.0435986159){289}{\line(0,-1){.0435986159}} %\end %\emline(17,78.395)(34.82,74.615) \multiput(17,78.395)(.157699115,-.033451327){113}{\line(1,0){.157699115}} %\end %\emline(19.34,67.055)(35.36,54.275) \multiput(19.34,67.055)(.0422691293,-.0337203166){379}{\line(1,0){.0422691293}} %\end %\emline(13.4,60.755)(35.9,53.735) \multiput(13.4,60.755)(.107655502,-.033588517){209}{\line(1,0){.107655502}} %\end \put(12.68,53.195){\line(1,0){22.68}} %\emline(15.92,46.895)(35.72,53.015) \multiput(15.92,46.895)(.108791209,.033626374){182}{\line(1,0){.108791209}} %\end %\emline(22.22,39.155)(35.9,53.195) \multiput(22.22,39.155)(.0336945813,.0345812808){406}{\line(0,1){.0345812808}} %\end \put(35.54,53.915){\line(1,0){20.52}} \put(35.18,31.775){\line(1,0){20.34}} \put(36.98,10.715){\line(1,0){19.26}} %\emline(56.78,11.615)(72.62,8.015) \multiput(56.78,11.615)(.148037383,-.03364486){107}{\line(1,0){.148037383}} %\end %\emline(56.24,11.435)(72.44,18.095) \multiput(56.24,11.435)(.081818182,.033636364){198}{\line(1,0){.081818182}} %\end %\emline(56.06,32.675)(73.16,30.155) \multiput(56.06,32.675)(.228,-.0336){75}{\line(1,0){.228}} %\end %\emline(56.24,32.315)(73.16,40.595) \multiput(56.24,32.315)(.0687804878,.0336585366){246}{\line(1,0){.0687804878}} %\end \put(35.9,74.615){\line(1,0){20.34}} \put(55.88,94.595){\line(0,-1){19.62}} \put(54.44,6.935){\makebox(0,0)[cc]{$11111(32)$}} \put(82.1,7.475){\makebox(0,0)[cc]{$11100(29)$}} \put(82.92,18.275){\makebox(0,0)[cc]{$10000(17)$}} \put(28.78,10.175){\makebox(0,0)[cc]{$00000(1)$}} \put(47.84,28.355){\makebox(0,0)[cc]{$10111(24)$}} \put(82.56,40.595){\makebox(0,0)[cc]{$10110(23)$}} \put(82.46,30.155){\makebox(0,0)[cc]{$11000(25)$}} \put(27.5,30.695){\makebox(0,0)[cc]{$01000(9)$}} \put(82.64,51.935){\makebox(0,0)[cc]{$10100(21)$}} \put(82.38,62.555){\makebox(0,0)[cc]{$11101(30)$}} \put(7.86,67.595){\makebox(0,0)[cc]{$00001(2)$}} \put(3.82,60.935){\makebox(0,0)[cc]{$10010(19)$}} \put(3.74,53.195){\makebox(0,0)[cc]{$10011(20)$}} \put(5.26,45.635){\makebox(0,0)[cc]{$11001(26)$}} \put(11.38,38.435){\makebox(0,0)[cc]{$11110(31)$}} \put(65.94,70.915){\makebox(0,0)[cc]{$01010(11)$}} \put(41.7,76.855){\makebox(0,0)[cc]{$00010(3)$}} \put(16.16,87.755){\makebox(0,0)[cc]{$00011(4)$}} \put(8.34,78.395){\makebox(0,0)[cc]{$00110(7)$}} \put(74.42,77.855){\makebox(0,0)[cc]{$00101(6)$}} \put(102.8,85.955){\makebox(0,0)[cc]{$00100(5)$}} \put(104.7,74.975){\makebox(0,0)[cc]{$00111(8)$}} \put(102.72,66.155){\makebox(0,0)[cc]{$11010(27)$}} \put(65.64,90.815){\makebox(0,0)[cc]{$01011(12)$}} \put(29.48,104.315){\makebox(0,0)[cc]{$01001(10)$}} \put(30.02,96.235){\makebox(0,0)[cc]{$01100(13)$}} \put(87.96,105.035){\makebox(0,0)[cc]{$01111(16)$}} \put(87.96,94.595){\makebox(0,0)[cc]{$11011(28)$}} \put(48.945,84.6){\circle*{1.61}} \put(48.585,64.26){\circle*{1.61}} %\emline(48.68,84.515)(56.06,74.795) \multiput(48.68,84.515)(.03369863,-.044383562){219}{\line(0,-1){.044383562}} %\end %\emline(48.5,63.995)(56.24,74.435) \multiput(48.5,63.995)(.033652174,.045391304){230}{\line(0,1){.045391304}} %\end \put(41.6,48.335){\makebox(0,0)[cc]{$10001(18)$}} \put(61.74,48.695){\makebox(0,0)[cc]{$10101(22)$}} \put(39.2,85.415){\makebox(0,0)[cc]{$01101(14)$}} \put(38.56,63.095){\makebox(0,0)[cc]{$01110(15)$}} \put(55.88,1.175){\makebox(0,0)[cc]{Fig. 4(d). Spanning Tree $T_4$ in $AQ_5$}} \end{picture}$$ Here, we observe that\ $InV(T_1)= \{1, 5, 7, 16, 17, 25, 27, 31\}$,\ $InV(T_2) = \{2, 4, 8, 10, 21, 23, 26, 30\}$\ $InV(T_3)= \{9, 13, 14, 15, 19, 20, 28, 29\}$\ $InV(T_4)= \{3, 6, 11, 12, 18, 22, 24, 32\}$\ are such that $InV(T_i) \cap InV(T_j) = \phi $, for $i \neq j$ and $ 1 \leq i, j \leq 4$\ Also, observe\ $E(T_1) = \{\langle 1, 2 \rangle, \langle 1, 3 \rangle, \langle 1, 5 \rangle, \langle 1, 16 \rangle, \langle 1, 17 \rangle, \langle 4, 5 \rangle, \langle 5, 7 \rangle, \langle 5, 12 \rangle, \langle 5,21 \rangle, \langle 5, 28 \rangle, \langle 6, 7 \rangle, \langle 7, 8 \rangle, $ $\hspace{1cm} \langle 7, 10 \rangle, \langle 9, 25 \rangle, \langle 11, 27 \rangle, \langle 13, 16 \rangle, \langle 14, 16 \rangle, \langle 15, 16 \rangle, \langle 17. 18 \rangle, \langle 17, 20 \rangle, \langle 17, 24 \rangle, \langle 17, 25 \rangle, $ $\hspace{1cm} \langle 19, 27 \rangle, \langle 22, 27 \rangle, \langle 23, 31 \rangle, \langle 25, 27 \rangle, \langle 26, 27 \rangle, \langle 27, 31 \rangle, \langle 29, 31 \rangle, \langle 30, 31 \rangle, \langle 31, 32 \rangle \} $\ $E(T_2) = \{\langle 1, 4 \rangle, \langle 2, 3 \rangle, \langle 2, 4 \rangle, \langle 2, 6 \rangle, \langle 2, 10 \rangle, \langle 4, 8 \rangle, \langle 4, 12 \rangle, \langle 4, 13 \rangle, \langle 4, 29 \rangle, \langle 5, 8 \rangle, \langle 7, 23 \rangle, \langle 8, 16 \rangle, $ $\hspace{1cm} \langle 8, 24 \rangle, \langle 8, 25 \rangle, \langle 9, 10 \rangle, \langle 10, 11 \rangle, \langle 10, 15 \rangle, \langle 10, 23 \rangle, \langle 14, 30 \rangle, \langle 17, 21 \rangle, \langle 18, 23 \rangle, \langle 19, 30 \rangle, $ $\hspace{1cm} \langle 20, 21 \rangle, \langle 21, 23 \rangle, \langle 21, 28 \rangle, \langle 22, 23 \rangle, \langle 23, 26 \rangle, \langle 26, 30 \rangle, \langle 26, 31 \rangle, \langle 27, 30 \rangle, \langle 30, 32 \rangle \} $\ $E(T_3) = \{\langle 1, 9 \rangle, \langle 2, 15 \rangle, \langle 3, 14 \rangle, \langle 4, 20 \rangle, \langle 5, 13 \rangle, \langle 6, 14 \rangle, \langle 7, 15 \rangle, \langle 8, 9 \rangle, \langle 9, 11 \rangle, \langle 9, 12 \rangle, \langle 9, 13 \rangle, $ $ \hspace{1cm} \langle 9, 16 \rangle, \langle 10, 14 \rangle, \langle 13, 14 \rangle, \langle 13, 15 \rangle, \langle 13, 29 \rangle, \langle 15, 18 \rangle, \langle 15, 31 \rangle, \langle 17, 19 \rangle, \langle 19, 20 \rangle, \langle 19, 22 \rangle, $ $ \hspace{1cm} \langle 19, 23 \rangle, \langle 20, 24 \rangle, \langle 20, 29 \rangle, \langle 21, 29 \rangle, \langle 25, 29 \rangle, \langle 26, 28 \rangle, \langle 27, 28 \rangle, \langle 28, 29 \rangle, \langle 28, 32 \rangle, \langle 29, 30 \rangle \} $\ $E(T_4) = \{\langle 1, 32 \rangle, \langle 2, 18 \rangle, \langle 3, 4 \rangle, \langle 3, 7 \rangle, \langle 3, 11 \rangle, \langle 5, 6 \rangle, \langle 6, 8 \rangle, \langle 6, 11 \rangle, \langle 6, 27 \rangle, \langle 9, 24 \rangle, \langle 10, 12 \rangle, \langle 11, 12 \rangle,$ $ \hspace{1cm} \langle 11, 14 \rangle, \langle 11, 15 \rangle, \langle 11, 22 \rangle, \langle 12, 13 \rangle, \langle 12, 16 \rangle, \langle 12, 28 \rangle, \langle 17, 32 \rangle, \langle 18, 19 \rangle, \langle 18, 20 \rangle, $ $\hspace{1cm} \langle 18, 22 \rangle, \langle 18, 26, \rangle, \langle 18, 31 \rangle, \langle 21, 22 \rangle, \langle 22, 24 \rangle, \langle 22, 30 \rangle, \langle 23, 24 \rangle, \langle 24, 25 \rangle, \langle 24, 32 \rangle, \langle 29, 32 \rangle \} $\ are such that $ E(T_i) \cap E(T_j) = \phi$, for $i \neq j$ and $ 1 \leq i, j \leq 4$.\ According to Lemma \[lm1\], above constructed trees $T_1, T_2, T_3$ and $T_4$ are CISTs on $AQ_5$.\ Let $n\geq 6$ be an integer. There exist four completely independent spanning trees of which two are with diameter $2n-3$ and two are with diameter $2n-5$, in augmented cube $AQ_n$. By using above constructed four CISTs in $AQ_5$ and Corollary \[lm3\], we get four CISTs in $AQ_n$ for $n \geq 6$.\ As we want vertices to be close together to avoid communication delays. So, we will concentrate on diameters of above trees.\ Above constructed trees $T_1, T_2$ have diameters $8$ and $T_3, T_4$ have diameter $6$ in $AQ_5$. We first consider the construction of four CISTs in $AQ_6$. It is sufficient to show the construction of only one tree. Consider tree $T_1$ of $AQ_5$, having longest path of length $8$ and central vertex $10000(17)$. Now, by prefixing $0$ and $1$ to this vertex we get central vertex of $T^0_1$ and $T^1_1$ respectively. Means, we select $u_1 = 010000 \in V(T^0_1)$ and $v_1 = 110000\in V(T^1_1)$. Then, $\overline T_1$ is CIST with diameter $9$ on $AQ_6$. Constructing in the similar manner we get $\overline T_2$, $\overline T_3$, $\overline T_4$ CISTs with diameters $9$, $7$, $7$ respectively on $AQ_6$.\ Let $AQ_{n+1}$ ($n \geq 6)$ be decomposed into two augmented cubes say $AQ^0_n$ and $AQ^1_n$ with vertex set say $ \{x^0_i : 1 \leq i \leq 2^{n}\}$ and $\{x^1_i : 1 \leq i \leq 2^{n}\}$ respectively. Denote by $T^0_1, T^0_2, T^0_3, T^0_4$ the CISTs with diameter $2n-3, 2n-3, 2n-5, 2n-5$ respectively in $AQ^0_n$. Let the identical corresponding CISTs in $AQ^1_n$ be denoted by $T^1_1, T^1_2, T^1_3, T^1_4$.\ Now, we will prove by induction that the diameters of $\overline T_1, \overline T_2, \overline T_3, \overline T_4)$ CISTs in $AQ_{n+1}$ ($n \geq 6$) are $2n-1, 2n-1, 2n-3, 2n-3$ respectively.\ It is sufficient to prove result for a single tree say $T^0_1$.\ Let $P^0_1 = x^0_1-x^0_2-.......x^0_{n-1}-x^0_n-.....x^0_{2n-2} $ be the longest path naturally of length $2n-3$ in tree $T^0_1$. And $P^1_1 = x^1_1-x^1_2-.......x^1_{n-1}-x^1_n-.....x^1_{2n-2} $ be its corresponding path in corresponding tree $T^1_1$.\ The vertex $x^0_n$ is in the center of the path $P^0_1$ hence any vertex on tree $T^0_1$ will be within a distance $n-1$ from the vertex $x^0_n$, means any vertex $x^0_i \in V(T^0_1)$, $d_{T^0_1}(x^0_i, x^0_n ) \leq n-1$. Similarly, any vertex $x^1_i \in V(T^1_1)$, $d_{T^1_1}(x^1_i, x^1_n ) \leq n-1$.\ Now, consider tree say $\overline T_1$ be a spanning tree of $AQ_{n+1}$ constructed from $T^0_1$ and $T^1_1$ by adding an edge $\langle x^0_n, x^1_n \rangle \in E(AQ_{n+1})$ to connect two internal vertices $x^0_n \in V(T^0_1)$ and $x^1_n \in V(T^1_1)$. Then, $\overline T_1$ is with diameter $2n-1$. As $V(\overline T_1) = V(T^0_1) \cup V(T^1_1)$, $d_{\overline T_1}(x^0_i, x^1_j )= d_{T^0_1}(x^0_i, x^0_n ) + 1 + d_{T^1_1}(x^1_n, x^1_j ) \leq (n-1) + 1 + (n-1) = 2n-1$.\ Constructing in the similar manner we get $\overline T_2$, $\overline T_3$, $\overline T_4$ CISTs with diameters $2n-1$, $2n-3$, $2n-3$ respectively on $AQ_{n+1}$.\ [**Concluding remarks.**]{}\ In this paper, we have proposed a construction of four CISTs in augmented cube $AQ_n$ ($n \geq 6$) of which two trees with diameters $2n-3$ and two trees with diameters $2n-5.$ Our results provide $n-1$ CISTs in augmented cube $AQ_n$ ($n= 3,4,5$) and thus we pointed out that Hasunuma’s conjecture does hold for $AQ_n$ when $n= 3,4,5$. As connectivity of augmented cubes is comparatively higher than other variants of hypercubes, an interesting problem is whether Hasunuma’s conjecture is true for $AQ_n$ ($n\geq 6$) if so then how to derive an algorithm that construct $n-1$ CISTs in $AQ_n$ ($n\geq 6$)?\ [**Acknowledgment:**]{} The first author gratefully acknowledges the Department of Science and Technology, New Delhi, India for the award of Women Scientist Scheme for research in Basic/Applied Sciences. [99]{} , Dirac’s condition for completely independent spanning trees, J. Graph Theory $77 (2014) 171-179$. , Anote on the degree condition of completely independent spanning trees, IEICE Trans. Fundam. Electron. Commun. Comput. Sci. $98-A (2015) 2191-2193$. , Areliable broadcasting algorithm in locally twisted cubes, in: Proc. 2nd Int. Conf. on Cyber Security and Cloud Computing, $2015$, pp.$323-328$. , Augmented cubes, Networks $40(2) (2002) 71-84$. , Ore’s condition for completely independent spanning trees, Discrete Appl. Math. $177 (2014) 95-100$. , Disjoint rooted spanning trees with small depths in de Bruijn and Kautz graphs, SIAM J. Comput. $26 (1997) 79-92$. , Completely independent spanning trees in the underlying graph of a line digraph, Discrete Math. $234 (2001) 149-157$. , Completely independent spanning trees in maximal planar graphs, in: Proc. 28th Int. Workshop on Graph-Theoretic Concepts in Computer Science, WG 2002, in: Lecture Notes in Comput. Sci., vol.$2573, 2002$, pp.$235-245$. , Minimum degree conditions and optimal graphs for completely independent spanning trees, in: Proc. 26th Int. Workshop on Combina-torial Algorithms, IWOCA 2016, in: Lecture Notes in Comput. Sci., vol.$9538, 2016$, pp.$260-273$. , Completely independent spanning trees in torus networks, Networks $60 (2012) 59-69$. , Independent spanning trees with small depths in iterated line digraphs, Discrete Appl. Math. $110 (2001) 189-211$. , Degree condition for completely independent spanning trees, Inform. Process. Lett. $(2016) 644 - 648$. , Completely independent spanning trees in (partial) k-trees, Discuss. Math. Graph Theory $35 (2015) 427-437$. , Completely independent spanning trees on some interconnection networks, IEICE Trans. Inf. Syst. $97-$D $(2014) 2514-2517$. , Constructing two completely independent spanning trees in hypercube-variant networks, Theoret. Comput. Sci. $652 (2016) 28-37$. , Two counterexamples on completely independent spanning trees, Discrete Math. $312 (2012) 808-810$. , An algorithm to construct independent spanning trees on parity cubes, Theoret. Comput. Sci. $465 (2012) 61-72$. , Introduction to Graph theory, Second Edition, Prentice-Hall of India, New Delhi, $2002$. , Reducing the height of independent spanning trees in chordal rings, IEEE Trans. Parallel Distrib. Syst. $18 (2007), 644 - 657$. $$\diamondsuit\diamondsuit\diamondsuit$$

ArXiv

--- abstract: 'Wireless communication environments are unaware of the ongoing data exchange efforts within them. Moreover, their effect on the communication quality is intractable in all but the simplest cases. The present work proposes a new paradigm, where indoor scattering becomes software-defined and, subsequently, optimizable across wide frequency ranges. Moreover, the controlled scattering can surpass natural behavior, exemplary overriding Snell’s law, reflecting waves towards any custom angle (including negative ones). Thus, path loss and multi-path fading effects can be controlled and mitigated. The core technology of this new paradigm are metasurfaces, planar artificial structures whose effect on impinging electromagnetic waves is fully defined by their macro-structure. The present study contributes the software-programmable wireless environment model, consisting of several HyperSurface tiles controlled by a central, environment configuration server. HyperSurfaces are a novel class of metasurfaces whose structure and, hence, electromagnetic behavior can be altered and controlled via a software interface. Multiple networked tiles coat indoor objects, allowing fine-grained, customizable reflection, absorption or polarization overall. A central server calculates and deploys the optimal electromagnetic interaction per tile, to the benefit of communicating devices. Realistic simulations using full 3D ray-tracing demonstrate the groundbreaking potential of the proposed approach in $2.4\,GHz$ and $60\,GHz$ frequencies.' author: - | Christos Liaskos, Shuai Nie, Ageliki Tsioliaridou, Andreas Pitsillides, Sotiris Ioannidis, and Ian Akyildiz\ [Foundation for Research and Technology - Hellas (FORTH)]{}\ [Emails: {cliaskos,atsiolia,sotiris}@ics.forth.gr]{}\ [Georgia Institute of Technology, School of Electrical and Computer Engineering]{}\ [Emails: {shuainie, ian}@ece.gatech.edu]{}\ [University of Cyprus, Computer Science Department]{}\ [Email: [email protected]]{}\ title: 'Realizing Wireless Communication through Software-defined HyperSurface Environments' --- Wireless Environment, Communication-awareness, Indoor, Millimeter wave, Software control, Metasurfaces. Introduction\[sec:Intro\] ========================= Recent years have witnessed a tremendous increase in the efficiency of wireless communications. Multiple techniques have been developed to tackle the stochastic nature of the wireless channel, in an effort to fully adapt to its wide fluctuations. Indoor environments have attracted special attention, since multi-path fading accentuates due to the presence of multiple scatterers in a confined space. In such cases, techniques such as MIMO, beamforming, adaptive modulation and encoding have enabled wireless devices to rapidly adapt to the time-variant, unpredictable channel state [@pi2016millimeter]. The present work opens an unexplored research path: making the wireless environment fully controllable via software, enabling the optimization of major propagation factors between wireless devices. Thus, effects such as path loss and multi-path fading become controllable and mitigate-able. In order to understand the potential of exerting control over an environment, we first need to define its composition and its natural behavior. Indoor environments, which constitute the focus of the present work, comprise two or more communicating devicessuch as laptops, mobile phones, access points, base stations etc.and any object found in a domestic or work space that can influence their communication. At lower frequencies, walls, ceilings, floors, doors and sizable furniture act as electromagnetic (EM) wave scatterers, creating multiple paths between communicating end-points, especially in non-line-of-sight (NLOS) areas. At higher frequencies, such as millimeter wave (mm-wave) or terahertz ($THz$), which are expected to play a major role in upcoming 5G communications [@yilmaz2016millimetre], even small objects act as substantial scatterers. Furthermore, ultra-small wavelengths translate to considerable Doppler shift even at pedestrian speed [@yilmaz2016millimetre]. These factors, coupled with the natural ambient dissipation of power due to free space losses, lead to undermined NLOS performance at $2-5\,GHz$ and inability for NLOS communications at $60\,GHz$ and beyond [@yilmaz2016millimetre]. Existing proposals for multi-path and path loss mitigation can be classified as i) device-oriented, and ii) retransmitter-oriented. Device-oriented methods include massive MIMO deployments in communicating devices, to make constructive use of the multi-path phenomena [@Aijaz.2017]. Additionally, beamforming seeks to adaptively align the direction of wireless transmissions in order to avoid redundant free space losses [@kelif20163d; @REFLECTARRAYS]. Additional schemes include the on-the-fly selection of the modulation and encoding scheme that offers the best bit error rate (BER) under the current channel conditions [@huang2017multi]. Retransmission-oriented solutions advocate for the placement of amplifiers in key-positions within the indoor environment. Retransmitters can be either passive or active: Passive retransmitters are essentially conductive structures akin to antenna plates [@reflectInfocom.2017]. They passively reflect energy from and towards fixed directions, without tunability. Active retransmitters are powered electronic devices that amplify and re-transmit received signals within a given frequency band. Essentially, they attempt to combat power loss by diffusing more power within the environment. In mm-wave frequencies and beyond, retransmitters must be placed in line-of-sight (LOS) among each other, in an effort to eliminate NLOS areas within a floor plan. Device-to-device networking can also act as a retransmission solution for specific protocols and a limited capacity of served users [@chen2017promoting]. The overviewed solutions have a common trait: They constitute device-side approaches, which treat the environment as an uncontrollable factor that does not participate into the communication process. Metasurfaces are the core technology for introducing programmatically controlled wireless environments [@Zhu.2017; @Minovich.2015; @Lucyszyn.2010]. They constitute the outcome of a research direction in Physics interested in creating (rather than searching for) materials with required EM properties. In their earlier iterations, they comprised a metallic pattern, called *meta-atom*, periodically repeated over a Silicon substrate, as shown in Fig. \[fig:Mspatterns\]. The macroscopic EM behavior of a metasurface is fully defined by the meta-atom form. A certain pattern may fully absorb all impinging EM waves from a given direction of arrival (DoA), while another may fully reflect a given DoA towards another, at a negative reflection angle. Notably, metasurfaces (and their 3D counterpart, the metamaterials) offer a superset of EM behaviors with regard to regular materials. Lens functionality (concentration of reflections towards a given point rather than ambient dispersal) and negative refraction/reflection indexes are but some of the exotic EM capabilities they can exhibit [@Chen.2016]. Dynamic meta-atom designs allow for dynamic metasurfaces, as shown in Fig. \[fig:Mspatterns\]. Such designs include tunable factors, such as CMOS switches or Micro Electro-Mechanical Switches (MEMS) that can alter their stateand the EM behavior of the metasurfacevia an external bias [@Chen.2016]. The bias is commonly electronic, but thermal, light-based and mechanical approaches have been studied as well [@Chen.2016]. Thus, multi-functional metasurfaces, that can switch from one EM behavior to another (e.g., from absorbing to custom steering) are enabled. Finally, a very strong trait is that there is no known limitation to the operating metasurface frequency, which can be at the $mm$-wave and $THz$ bands [@Lee.2012]. ![\[fig:workflow\]The proposed workflow involving HyperSurface tile-coated environmental objects. The EM scattering is tailored to the needs of the communication link under optimization. Unnatural EM scattering, such as lens-like EM focus and negative reflection angles can be employed to mitigate path loss and multi-path phenomena, especially in challenging NLOS cases.](images/general_workflow){width="1\columnwidth"} The methodology proposed by the present study is to coat objects of EM significance within an indoor environment with a novel class of software-controlled metasurfaces. The study defines a unit of this metasurface class, called *HyperSurface tile*. A HyperSurface tile is a planar, rectangular structure that incorporates networked hardware control elements and adaptive meta-atom metasurfaces. Following a well-defined programming interface, a tile can receive external commands and set the states of its control elements to match the intended EM behavior. The tiles, covering walls, doors, offices, etc., form networks to facilitate the relaying of programmatic commands among them. Moreover, tiles can have environmental sensing and reporting capabilities, facilitating the discovery of communicating devices within the environment. As shown in Fig. \[fig:workflow\], a central server can receive incoming tile reports, calculate the optimal configuration per tile, and set the environment in the intended state by sending the corresponding commands. Collaboration with existing systems (e.g., localization services and Cloud computing), constitutes a strong aspect of the proposed approach, given that it enables the incorporation of the EM behavior of materials in smart control loops. The present study contributes the first model to describe programmable wireless indoor environments, detailing their hardware, networking and software components. The model includes the way for translating EM metasurface functionalities to reusable software functions, bridging physics and informatics. Moreover, the protocol specifications and programming interfaces for interacting with tiles for communication purposes are outlined. The practical procedure for deploying programmable EM environments to mm-wave indoor communication is detailed. The potential of programmable environments is evaluated via full 3D ray tracing in $2.4$ and $60\,GHz$ cases, demonstrating their ground-breaking potential in mitigating path loss and multi-path fading effects. The study also presents a way of modifying common ray-tracers to enable their use in programmable EM environment simulations. Uses of the proposed concept in other application domains are also discussed. The remainder of the text is organized as follows. Prerequisite knowledge on metasurfaces is given in Section \[sec:Backgrd\]. The HyperSurface-based wireless environment model is given in Section \[sec:arch\]. Applications to indoor wireless setups are discussed in Section \[sec:App\]. Evaluation via ray-tracing-based simulations is presented in Section \[sec:Evaluation\]. Finally, the conclusion is given in Section \[sec:Conclusion\]. Prerequisites \[sec:Backgrd\] ============================= This section provides the necessary background knowledge on metasurfaces, discussing dimensions and composition, operating principles and supported functionalities. The following concise description targets a wireless communications audience, given the topic of the present paper. A more detailed introduction can be found in [@Banerjee.2011]. A metasurface is a planar, artificial structure which comprises a repeated element, the meta-atom, over a substrate. In most usual compositions, the meta-atom is conductive and the substrate is dielectric. Common choices are copper over silicon, while silver and gold constitute other exemplary conductors [@Zhu.2017]. More exotic approaches employ graphene, in order to interact with $THz$-modulated waves [@Lee.2012]. Metasurfaces are able to control EM waves impinging on them, in a frequency span that depends on the overall dimensions. The size of the meta-atom is comparable to the intended interaction wavelength, $\lambda$, with $\nicefrac{\lambda}{2}$ constituting a common choice. The thickness of the metasurface is smaller than the interaction wavelength, ranging between $\nicefrac{\lambda}{10}\to\nicefrac{\lambda}{5}$ as a rule of a thumb. Metasurfaces usually comprise several hundreds of meta-atoms, which results into fine-grained control over the EM interaction control. In general, a minimum size of approximately $30\times30$ meta-atoms is required to yield an intended EM interaction [@Chen.2016]. ![\[fig:Mspatterns\]Split ring resonators (left) constituted a very common type of static metasurfaces, with fixed EM behavior. Novel designs (right) incorporate switch elements (MEMS, CMOS or other) to offer dynamically tunable EM behavior.](images/metasurfaceExamples_overlayText){width="1\columnwidth"} Figure \[fig:Mspatterns\]-a illustrates a well-studied metasurface design comprising split-ring resonators as the meta-atom pattern. Such classic designs that rely on a static meta-atom, naturally yield a static interaction with EM waves. The need for dynamic alteration of the EM wave control type has given rise to dynamic metasurfaces, illustrated in Fig. \[fig:Mspatterns\]-b. Dynamic meta-atoms incorporate phase switching components, such as MEMS or CMOS transistors, which can alter the structure of the meta-atom. Thus, dynamic meta-atoms allow for time-variant EM interaction, while meta-atom alterations may give rise to multi-frequency operation [@Zhu.2017]. Phase switching components can also be classified into state-preserving or not. For instance, mechanical switches may retain their state and require powering only for state transitions, while semiconductor switches require power to maintain their state. ![\[fig:MSprinciple\]The principle of metasurface functionality. Incident waves create a well-defined EM response to the unit cells. The cell response is crafted in such a way that the aggregate field follows a metasurface-wide design objective, e.g., reflection towards a custom angle $\Theta$.](images/howMetasurfacesWork){width="1\columnwidth"} The operating principle of metasurfaces is given in Fig. \[fig:MSprinciple\]. The meta-atoms, and their interconnected switch elements in the dynamic case, act as control factors over the surface currents flowing over the metasurface. The total EM response of the metasurface is then derived as the total emitted field by all surface currents, and can take completely engineered forms, such as the unnatural reflection angle shown in Fig. \[fig:MSprinciple\]. Engineering the total surface current is a complex process that must account for currents directly induced over the metasurface by the incident wave, the currents induced in a meta-atom wirelessly by other meta-atoms, as well as the currents flowing inwards or outwards from a meta-atom via the switch elements. A qualitative description of the dynamic metasurface operation can also be given: the meta-atoms can be viewed as either input or output antennas, connected in custom topologies via the switch elements. Impinging waves enter from the input antennas, get routed according to the switch element states, and exit via the output antennas, exemplary achieving customized reflection. State-of-the-art potential and manufacturing approaches ------------------------------------------------------- Metasurfaces constitute the state of the art in EM control in terms of capabilities and control granularity. A metasurface can support a wide range of EM interactions, denoted as *functions*. Common function types include [@Minovich.2015]: - Reflection of an impinging wave, with a given direction of arrival, towards a completely custom direction. - Refraction of EM waves via the metasurface towards any inwards direction. Both the reflection and refraction functions can override the outgoing directions predicted by Snell’s law. Reflection and refraction functions will jointly be referred to as wave *steering*. - Wave absorbing, i.e., ensuring minimal reflected and/or refracted power for impinging waves. - Wave polarizing, i.e., changing the oscillation orientation of the wave’s electric and magnetic field. Moreover, they can offer additional, advanced functions, such as collimation, resulting from near zero permittivity and permeability response, anisotropic response leading to hyperbolic dispersion relation, giant chirality, arbitrary wave-front shaping and frequency selective filtering [@Lucyszyn.2010]. Apart from communications, these traits have been exploited in a variety of applications, e.g., highly efficient energy harvesting photovoltaics, and thermophotovoltaics, ultra-high resolution medical imaging, sensing, quantum optics and military applications [@Iwaszczuk.2012]. The extended repertoire of EM function types, as well the exquisite degree of granularity in EM behavior control, sets metasurfaces apart from phased antennas and reflectarrays [@PHASEDANTENNAS; @REFLECTARRAYS], which can only support coarse EM steering, e.g., for beamforming applications in wireless devices [@kelif20163d]. Notice that highly fine-grained EM control is required in mm-wave setups, due to the extremely small wavelength [@yilmaz2016millimetre]. Regarding their manufacturing approaches, metasurfaces are common produced as conventional printed circuit boards (PCBs) [@Yang.2016]. The PCB approach has the advantage of relying on a mature, commercially accessible manufacturing technology. The PCB production cost is moderate (indicatively, USD 500 per $m^{2}$ [@PCBCART]). However, the PCB technology is originally intended for integrated circuits with far greater complexity than a metasurface. As described in the context of Fig. \[fig:Mspatterns\], a metasurface can be a very simple structure, comprising a set of conductive patches, diodes and conductive power/signal lines. Therefore, large area electronics (LAE) can constitute better manufacturing approaches in terms of ultra low production cost [@LAEBOOK; @LAEAPP]. LAE can be manufactured using conductive ink-based printing methods on flexible and transparent polymer films, and incorporate polymer/organic diodes [@LAEPRINTED]. Films with metasurface patterns and diodes printed on them can then be placed upon common objects (e.g., glass, doors, walls, desks), which may also act as the dielectric substrate for the metasurface. The HyperSurface-based Wireless Environment Model\[sec:arch\] ============================================================= {width="100.00000%"} This section details the HyperSurface tile hardware components, the tile inter-networking and the environment control software. A schematic overview is given in Fig. \[fig:Architecture\] and is detailed below. . The tile hardware consists of a dynamic metasurface, a set of networked, miniaturized controllers that control the switch elements of the metasurface, and a gateway that provides inter-tile and external connectivity. The controller network has a slave/master relation to the gateway. Via the gateway, the controller network reports its current state and receives commands to alter the state of the switch elements in a robust manner, making the metasurface yield an overall required EM function. A single controller is a miniaturized, addressable electronic device that can monitor and modify the state of at least one metasurface switch element. The controller design objectives are small size (to avoid significant interference to the EM function of the metasurface), low-cost (to support massive deployments in many tiles), high monitoring and actuation speed (to sustain fast EM reconfigurability of the metasurface), and the ability to create, receive and relay data packets (to enable controller networking). The avoidance of EM function disruption also refers to the wiring required to connect the controllers to the switch elements and to each other in a grid topology (cf. Fig. \[fig:Architecture\]). Therefore, the total wiring should also be kept low. The grid-networked controller approach is an option that balances wiring length and robustness to node failures. Bus connectivity for the controllers would minimize the required wiring, but would decrease the robustness against node failures. On the other hand, a star connectivity would offer maximum robustness but would also yield maximum wiring. Notice that future technologies, such as nanonetworking, may enable wireless, computationally-powerful nodes with autonomous, energy harvesting-based power supply [@Akyildiz.2008]. Thus, future tile designs may need no wiring or specific gateways. The setup presented in this study prioritizes cost-effective realizability with present manufacturing capabilities. At a logical level, a controller is modeled as a finite-state automaton, which reacts to incoming packets or switch element changes by transitioning from one state to another [@books2011automata]. A UML-standard state diagram should capture three basic controller processes: the data packet handling (including re-routing, consuming packets and sending acknowledgments), the node reporting (reacting to an incoming monitoring directivemonitor request packetby creating a new monitor data packet), and a fault detection process (either self- or neighbor-failure). The latter is required for robust data routing and for deducing the operational state of the tile as a whole. Regarding the controller addressing, it can be hardwired due to the fixed grid topology. The tile gateway stands between the tile controller network and the external world. It is incorporated to the tile fabric at a position selected to yield minimal EM interaction concerns (e.g., at the back of the tile). It provides mainstream protocol-compatible data exchange with any other system. Internally, it is connected to at least one controller, while more connections can be used for robust connectivity. Moreover the gateway acts as a power supply bridge for the tile. Limited size (e.g., \~$cm$) and energy requirements are the only significant constraints. Existing hardware, such as IoT platforms [@Verikoukis.2017], can be employed as tile gateways [@Verikoukis.2017]. The tile gateway may optionally have EM DoA sensing capabilities, to facilitate the location discovery of wireless user devices in the environment. . As tiles are placed over an environmental object, such as a wall, they click together, connecting data and power lines among the tile gateways (cf. Fig. \[fig:Architecture\]). Thus, the tiles form a wired ad hoc network in a grid topology. Once again, existing IoT communication protocols can be readily employed. The same protocol is used for connecting the tile network to any external system. At least one tiledenoted as exit/entry pointhas its gateway connect to the environment configuration server, which accumulates sensed data and diffuses EM actuation commands within the tile network. More than one tile can be used as exit/entry points at the same time, for the interest or robust and timely data delivery. . The environment control software is an application programming interface (API) that exists at the configuration server. The API serves as a strong layer of abstraction, hiding the internal complexity of the HyperSurfaces. It offers user-friendly and general purpose access to metasurface functions, without requiring knowledge of the underlying hardware and Physics. It provides software descriptions of metasurface functions, allowing a programmer to customize, deploy or retract them on-demand over tiles with appropriate callbacks. These callbacks have the following general form: $${\scriptstyle \texttt{outcome}\gets\texttt{callback(tile\_ID, action\_type, parameters)}}$$ The $\texttt{tile\_ID}$ is the unique address of the intended tile gateway in the inter-tile network (e.g., an IPv6). One EM function per tile is considered here for simplicity. The $\texttt{action\_type}$ is an identifier denoting the intended function, such as $\texttt{STEER}$ or $\texttt{ABSORB}$, as described in Section \[sec:Backgrd\]. Each action type is associated to a set of valid parameters. For instance, $\texttt{STEER}$ commands require: i) an incident DoA, $\overrightarrow{I}$, ii) an intended reflection direction, $\overrightarrow{O}$, and iii) the applicable wavelength, $\lambda$, (if more than one are supported). $\texttt{ABSORB}$ commands require no $\overrightarrow{O}$ parameter. Notice that metasurface properties can be symmetric: i.e., a $\texttt{STEER}\left(\overrightarrow{I},\overrightarrow{O}\right)$ can also result into $\texttt{STEER}\left(\overrightarrow{O},\overrightarrow{I}\right)$ [@Holloway.2012]. Once executed at the configuration server, a callback is translated to an appropriate configuration of the switch elements that should be deployed at the intended tile. The configuration is formatted as a data packet that enters the tile network via an entry/exit point, and is routed to the intended tile via the employed inter-tile routing protocol. The intended tile gateway translates the directive according to the controller network communication protocol specifications and diffuses it within the tile. Upon success, it returns an acknowledgment to the configuration server, or an error notification otherwise. In the general case, the translation of an EM function to a tile switch element configuration is accomplished via a lookup table, populated during the tile design/manufacturing process as follows. Let $\sigma$ be a single tile configuration, defined as an array with elements $s_{ij}$ describing the intended switch element state that is overlooked by controller with address $i,j$ in the tile controller network. (One-to-one controller-switch relation is assumed). In the MEMS case, $s_{ij}$ takes binary values, $1$ or $0$, denoting switch connection or disconnection. Additionally, let $\Sigma$ be the set of all possible configurations, i.e., $\sigma\in\Sigma$. Let an EM function of type $\texttt{ABSORB}$ from DoA $\overrightarrow{I}$ be of interest. Moreover, let $P_{\sigma}(\phi,\theta)$ be the power reflection pattern of the tile (in spherical coordinates), when a wave with DoA $\overrightarrow{I}$ impinges upon it and a configuration $\sigma$ is active. Then, the configuration $\sigma_{best}$ that best matches the intended function $\texttt{ABSORB}\left(\overrightarrow{I}\right)$ is defined as: $$\sigma_{best}\gets argmin_{\sigma\in\Sigma}\left\{ max_{\forall\phi,\theta}P_{\sigma}(\phi,\theta)\right\} \label{eq:fitness}$$ Existing heuristic optimization processes can solve this optimization problem for all functions of interest in an offline manner [@haupt2007genetic], using simulations or field measurements on prototypes. The configuration lookup table is thus populated. Finally, we note that analytical results for the EM function-configuration relation exist in the literature for several metasurface designs [@haupt2007genetic]. In such cases, the analytical results can be employed directly, without the need for lookup tables. Control Algorithm of HyperSurface\[sec:algorithm\] ================================================== In order to establish communication links between transmitters and receivers, the HyperSurface tiles need to be adaptively selected and optimally controlled to serve the desired receivers. Since in real-world communication scenarios, multiple users can be present in the same space, it is necessary to discuss the tile distribution and control algorithms. Applications to Mm-wave Indoor Setups\[sec:App\] ================================================ ![\[fig:WirCommExampleNegAngle\]Illustration of a customized wireless indoor environment. $\texttt{STEER}$ functions are applied to several tiles, to achieve a $\texttt{FOCUS}$ behavior of the corresponding wall as a whole. ](images/exampleWirComm_neg_angle){width="1\columnwidth"} In mm-wave setups, major factors affect the signal attenuation: i) the increased free space path loss (e.g., $\sim90\,dB$ at $10\,m$ for $60\,GHz$, instead of $60\,dB$ for $2.4\,GHz$), ii) acute multi-path fading even in LOS cases, iii) strong Doppler shift even at pedestrian speeds, iv) optical-like propagation of EM waves, limiting connectivity to LOS cases and exhibiting strong sensitivity to shadowing phenomena. Attenuation due to molecular absorption may not play a significant role in indoor casesdepending on the composition of the environmentas it corresponds to $10^{-5}\,dB/m$ loss [@pi2016millimeter]. Given the mentioned mm-wave considerations, we proceed to present mitigation measures offered by a HyperSurface-enabled environment. We consider the setup of Fig. \[fig:WirCommExampleNegAngle\], comprising a receiver (Rx)-transmitter (Tx) pair located in NLOS over a known floorplan. The walls are coated with HyperSurface tiles. Furthermore, we consider the existence of a location discovery service (e.g., [@localization60ghzCmaccuracy]), which reports the location of the user device. At first, the Rx and Tx may attempt high-power, omni-directional communication. The location discovery service pinpoints the location of the user device and sends it to the EM environment configuration server. (Without loss of generality, the location of the Tx/access point can be considered known). Tiles may sense their impinging power and report it to the server as well. The server can use this information to increase the accuracy of the discovered user device location. Subsequently, the following actions take place: - The tiles at the top-left part of Fig. \[fig:WirCommExampleNegAngle\] are set to a symmetric “negative focus” setup as shown. - The Tx and the Rx are signaled to direct their antenna patterns to the configured tiles using beamforming. Using this approach, the path loss can be even fully mitigated, since the emitted energy is focused at the communicating end-points, rather than scattering within the environment. This can also be of benefit to the user device’s battery lifetime, given that the redundantly emitted power is minimized. Concerning multi-path fading, the fine-grained EM control over the wave propagation can have as an objective the *crafting* of a power delay profile that mitigates the phenomenon, e.g., by ensuring a path with significantly more power than any other, or one that best matches the MIMO capabilities of the devices. Additionally, the focal point of the EM wave reflected by the tiled wall towards the use device can be altered in real-time, to match the velocity of the mobile user. Mobile trajectory predictions can be employed to facilitate this course of action. This provides a potential mitigation approach for Doppler phenomena. The environment optimization for multiple user pairs, or sub-spaces within the environment, may be of increased practical interest. Returning to the setup of Fig. \[fig:WirCommExampleNegAngle\], the configuration server can, e.g., set the tiles to preemptively minimize the delay spread within the whole NLOS area, while ensuring a minimum level of received power within it. In the sub-space optimization case, the best matching tile configurations can be calculated offline and be deployed upon request. This approach is evaluated in Section \[sec:Evaluation\]. Finally, it is noted that the programmable environment extends the communication distance of devices, without requiring extra dissipation of energy within the environment (e.g. by placing additional access points). This can constitute a considerable advantage for mm-wave communications, which are known to be highly absorbable by living tissue. Moreover, assuming tiles with state-preserving switch elements, the energy footprint of the programmable environment can be extremely low, especially in static or mildly changing user positions. Evaluation in $60\,GHz$ and $2.4\,GHz$ setups\[sec:Evaluation\] =============================================================== We proceed to evaluate the HyperSurface potential in mitigating the path loss and multipath fading effects, using a simulation platform. Specifically, the indoor 3D space of Fig. \[fig:WirCommExampleNegAngle\] is ported to a full-3D ray-tracing engine [@ActixLtd.2010], customized to take into account HyperSurface tile functions. The evaluation focuses on finding tile configurations that optimally mitigate the path loss and multipath fading for $12$ users within the NLOS area. We study the case of $60\,GHz$, which is of increased interest to upcoming 5G communications, as well as the $2.4\,GHz$ case due to its wide applicability, e.g., to WiFi setups [@yilmaz2016millimetre]. Concerning the simulation parameters, the space has a height of $H=3\,m$, corridor length (distance between opposite wall faces) $L=15\,m$, corridor width $W=4.5\,m$, a middle wall length of $12\,m$, and $0.5\,m$ wall thickness. Two stacked walls exist in the middle. The floor and ceiling are treated as plain, planar surfaces composed of concrete, without HyperSurface functionality. All walls are coated with HyperSurface tiles, which are square-sized with dimensions $1\times1\,m$. Thus, the 3D space comprises a total of $222$ tiles. The dynamic metasurface pattern of Fig. \[fig:Mspatterns\] is considered, using mechanical, state-preserving switches (ON/OFF states are allowed). Appropriate dimensions are assumed, for $60\,GHz$ and $2.4\,GHz$ respectively, as explained in the context of Fig. \[fig:Mspatterns\]. This pattern design has been extensively studied in literature, offering a wide range of steering and absorbing capabilities, even with switch elements only at the horizontal direction [@haupt2007genetic p. 235]. Although beyond of the present scope, it is noted that this metasurface design also exhibits tunable EM interaction frequency, yielding a particularly extended repertoire of supported tunability parameters. The considered tile functions account for EM wave steering and absorption from various DoAs. Specifically, we allow for any DoA and reflection direction resulting from the combination of $\left\{ -30^{o},\,-15{}^{o},\,0^{o},\,15^{o},\,30^{o}\right\} $ in azimuth and $\left\{ -30^{o},\,-15{}^{o},\,0^{o},\,15^{o},\,30^{o}\right\} $ in elevation planes, using the tile center as the origin. Notice that the considered angles have been shown to be commonly attainable by metasurfaces [@Yazdi.2017]. However, carefully designed, static metasurfaces have achieved nearly full angle coverage, i.e., almost $\left(-90^{o},\,90^{o}\right)$ in azimuth and elevation, which is indicative of their potential [@Albooyeh.2014]. The reflection coefficient is set to $100\%$ for each steering function [@haupt2007genetic p. 235]. Additionally, we consider an EM absorbing tile function which reduces the power of impinging waves (given DoA) by $35\,dB$ [@haupt2007genetic p. 235], scattering the remaining wave power towards the Snell law-derived reflection direction. Thus, a tile supports $26$ different function configurations in total. {width="49.00000%"}{width="52.00000%"} Existing ray-tracing engines employ common laws of optics to simulate the propagation of waves. As such, current ray-tracers do not readily allow for custom wave steering functions. (Absorbing functions, on the other hand, are readily supported). Thus, to implement steering functions we work as follows. First, the following observation is made: Assume a tile and a set of a required wave DoA and a reflection direction upon it, not abiding by Snell’s law. There exists a rotation of the tile in 3D space that makes the wave DoA and reflection direction comply with Snell’s law. Based on this Remark, the custom steering functions are implemented by tuning the tile’s spatial derivative as follows. Since a tile is a flat, square surface in a 3D space, its spatial derivative is normally an arrow perpendicular to the tile surface. In order to allow for custom EM wave steering within the ray-tracing engine, we allow for virtually rotating the spatial derivative (but not the tile itself) by proper azimuth and elevation angles. The modified spatial derivative is then used in all ray-tracing calculations. The external service is considered to know the tile specifications, i.e., the tile configuration that corresponds to each virtual angle combination. The service has obtained the direction of the impinging wave at each tile via the distributed sensing elements. Subsequently, it deploys the corresponding STEER or ABSORB commands at each tile, by applying the corresponding tile configuration. An EM transmitter (Tx) is placed at position $\left\{ 7,\,12,\,2\right\} \,m$ (with respect to the origin placed on the floor level, at the upper-left corner of Fig. \[fig:WirCommExampleNegAngle\]). It is equipped with a half-dipole antenna and transmits at a carrier frequency of $60\,GHz$ or $2.4\,GHz$ (two studies) and $25\,MHz$ bandwidth. The transmission power is set to $100\,dBmW$, a high number chosen to ensure that no propagation paths are disregarded by the ray-tracer due to its internal, minimum-allowed path loss threshold. The NLOS area is defined as $x\in\left[0,\,4\right]\,m$, $y\in\left[0,\,15\right]\,m$ and a constant height of $z=1.5\,m$. Within the NLOS area, a set $12$ receiverswith antennas identical to the transmitterare placed at a regular $2\times6$ uniform grid deployment, with $2.5\,m$ spacing. The receiver grid is centered in the NLOS area. Intermediate signal reception values, used only for illustration purposes in the ensuing Figures, are produced by means of interpolation. The evaluation scenario considers two case studies, corresponding to the path loss and multi-path fading mitigation objectives. In each case, the state of each of the $222$ tiles is treated as an input variable of an appropriate objective function which must be optimized. Given the vastness and discontinuity of the solution space (i.e., $222^{26}$ possible tile configurations, positioned at different walls) and the discrete nature of the input variables, a Genetic Algorithm (GA) is chosen as the optimization heuristic [@haupt2007genetic], using the MATLAB Optimization Toolbox implementation [@MatlabGA]. GAs are heuristics that are inspired by evolutionary biology principles. They treat the variables of an optimization problem as *genomes* which compete with each other in terms of best fitness to an optimization objective. Good solutions are combined iteratively by exchanging *genes*, i.e., variable sub-parts, producing new generations of solutions. In the problem at hand, a genome represents a complete tile configuration, i.e, an array containing the state of the $222$ tiles. A gene represents the state of each tile, i.e., the specific array elements. Two optimization cases are studied, denoted as (A) and (B), both for $60\,GHz$ and $2.4\,GHz$. These are defined as follows: - ****. This case expresses the path loss mitigation goal, and is defined as the following optimization objective: *Define the optimal tile configurations that maximize the minimum received power over the 12 receivers in the NLOS area*. - ****. The case expresses the multi-path fading mitigation goal and is defined as the following optimization objective: *Define the optimal tile configurations that minimize the maximum delay spread over the 12 receivers in the NLOS area, with the constraint of ensuring a minimum total received power (custom threshold*). For Case (B), the thresholds are set to *$1\,dBmW$* for $60\,GHz$, and *$30\,dBmW$* for $2.4\,GHz$, based on the floor-plan dimensions and the path loss levels discussed in Section \[sec:App\]. ------ ------------- ------------- ----------- ------------- HSF setup Plain setup HSF setup Plain setup Max **$34.98$** $22.63$ $0.69$ $3.6$ Mean $25.38$ $-75$ $0.0068$ $0.48$ Min $16.13$ $-250$ $0.0045$ $0.007$ ------ ------------- ------------- ----------- ------------- : <span style="font-variant:small-caps;">\[tab:T60\]Comparison of total received power (case A) and power delay profile (case B) with and without HyperSurface (HSF) Tiles at $60\,GHz$.</span> The results for the $60\,GHz$ case are shown in Fig. \[fig:maxminP60\], \[fig:minmaxDS60\] and are summarized in Table. \[tab:T60\]. Figure \[fig:maxminP60\] presents case (A) for the plain (left) and HyperSurface-enabled (right) environments. In the plain setup, the tile spatial derivatives (black arrows) are naturally perpendicular to the tile surfaces. The average received power over the $12$ NLOS area receivers is $-75\,dBmW$, while the minimum power is $-250\,dBmW$, which is the lowest level allowed by the ray-tracing engine. Thus, the bottom-left and the three top-right receivers of the NLOS area are essentially disconnected in the plain setup. The maximum total received power is $22.63\,dBmW$. The right inset of Fig. \[fig:maxminP60\] shows the corresponding results with the HyperSurface functionality enabled. Notably, the minimum power level over the NLOS area is $16.13\,dB$, which constitutes a raise by at least $266.13\,dBmW$ with regard to the plain case. Moreover, the received power becomes essentially uniform over the NLOS area, ranging between $16.13$ and $34.98\,dBmW$, with an average of $25.38\,dBmW$. The tile spatial derivatives exhibit a degree of directivity towards the previously disconnected area parts (e.g., cf. left-most wall). Moreover, the top-and bottom tiles across the height of the walls tend to focus towards the NLOS area height. The non-uniformity of the derivatives is in accordance with the nature of the Genetic Algorithm, which is a very exploratory but not gradient-ascending optimizer [@Luke.2009]. This means that there exists potential for an even better optimization result near the Genetic Algorithm-derived solution. ![\[fig:minmaxDS60\]Wireless environment optimization case study (B) for $60\,GHz$. The objective is to minimize the maximum delay spread over the NLOS area, while ensuring a minimum of $1\,dBmW$ total received power per receiver. The circled parts of the plain setup correspond to disconnected areas. (cf. Fig. \[fig:maxminP60\]-left).](images/60GHz_DealySpreadHSF_NO){width="1\columnwidth"} The case (B) results for $60\,GHz$ are shown in Fig. \[fig:minmaxDS60\]. The objective is to minimize the maximum delay spread over the $12$ NLOS receivers, under the constraint for at least $1\,dBmW$ total received power per receiver. For the plain setup, shown in the left inset, we note a maximum delay spread of approximately $3.6\,nsec$. The $1\,dBmW$ minimum power constraint is of course not satisfied, as previously shown in Fig. \[fig:maxminP60\]-left. The circled areas correspond to the under-powered/disconnected NLOS area parts. The minimum and average delay spread over the *connected* areas only are $7\,psec$ and $0.48\,nsec$ respectively. The HyperSurface-enabled setup (right inset), achieves $5.21$ times lower maximum delay spread ($0.69\,nsec$) than the plain setup, a minimum of $4.5\,psec$ delay spread ($1.5$ times lower), and an average of $6.8\,psec$ ($70$ times lower). This significant performance improvement is accompanied by considerable total power levels, in the range of $\left[7.07,\,16.93\right]\,dBmW$ (average: $10.64\,dBmW$), fulfilling the optimization constraint of $1\,dBmW$. {width="100.00000%"} ------ ------------- ------------- ----------- ------------- HSF setup Plain setup HSF setup Plain setup Max **$59.81$** $47$ $0.68$ $3.65$ Mean $51.37$ $-58$ $0.067$ $0.47$ Min $45.13$ $-250$ $0.0029$ $0.0014$ ------ ------------- ------------- ----------- ------------- : <span style="font-variant:small-caps;">\[tab:T2p4\]Comparison of total received power (case A) and power delay profile (case B) with and without HyperSurface (HSF) Tiles at $2.4\,GHz$.</span> The results for the $2.4\,GHz$ case are similar to the $60\,GHz$ in terms of improvement, and are collectively given in Fig. \[fig:all2p4\] and Table \[tab:T2p4\]. The objective in the two leftmost panels is to maximize the minimum total received power over the $12$ receivers in the NLOS area. The plain setup achieves $-250$, $-58$ and $47\,dBmW$ minimum, average and maximum total received power, respectively. The HyperSurface setup yields considerably improved results, with $45.13$, $51.37$ and $59.81\,dBmW$ minimum, average and maximum total received power, respectively. Thus, there is a gain of $295.13\,dBmW$ in minimum received power. The delay spread improvement is also significant, as shown in the two rightmost panels. The plain setup yields $1.4\,psec$, $0.47\,nsec$ and $3.65\,nsec$ minimum, average and maximum delay spread values, with $4$ disconnected receivers (circled parts, cf. first inset of Fig. \[fig:all2p4\]). The corresponding HyperSurface-enabled setup achieves $2.9\,psec$, $67\,psec$ and $0.68\,nsec$ min/average/max respectively. Moreover, it ensures a minimum total received power of $34.12\,dBmW$, successfully meeting the $30\,dBmW$ optimization constraint. Discussion\[sec:future\] ------------------------ The results of Section \[sec:Evaluation\] demonstrated the path loss and multi-path fading mitigation potential of the proposed softwarization of wireless indoor environments. Even at the highly-challenging $60\,GHz$ communications, a HyperSurface tile-coated indoor setup exhibited significant improvements in received power levels and delay spread. Such traits can benefit the communication distance of devices and their energy consumption, without dissipating more energy in thealready EM-strainedenvironments via retransmitters. This promising performance can encourage further exploration of the HyperSurface concept in additional usage domains: Multiple applications can be studied in both indoor and outdoor environments, and in the context of multiple systems, such as 5G, IoT and D2D, where ultra-low latency, high bandwidth, and support for massive numbers of devices is important [@Aijaz.2017]. Moreover, HyperSurfaces may act as an enabler for upcoming $THz$ communications. Operation in this band promises exceptional data rates and hardware size minimization at the nano-level, which can enable a wide range of groundbreaking applications [@akyildiz2015internet]. Nonetheless, the $THz$ band is susceptible to acute signal attenuation owed to molecular absorption. HyperSurfaces with graphene-based meta-atom designs could act as a smart environment for $THz$ communications [@Fan.2013], mitigating the attenuation effects and extending the communication range. Conclusion\[sec:Conclusion\] ============================ The present study proposed an indoor wireless communication paradigm where the electromagnetic propagation environment becomes aware of the ongoing communications within it. The key idea is to coat objects such as walls, doors and furniture with HyperSurface tiles, a forthcoming type of material with programmable electromagnetic behavior. HyperSurfaces can exert fine-grained control over impinging electromagnetic waves, steering them toward completely custom directions, polarizing them or fully absorbing them. HyperSurfaces have inter-networking capabilities, allowing for the first time the participation of electromagnetic properties of materials into control loops. A central server maintains a view of the communicating devices within an indoor space, and subsequently sets the tile electromagnetic configuration in accordance with any optimization objective. The HyperSurface tile concept has been evaluated in $2.4\,\text{and}\,60\,GHz$ setups, which demonstrated its high potential for path loss and multi-path fading mitigation, from microwave to mm-wave setups. Acknowledgment {#acknowledgment .unnumbered} ============== This work was funded by the European Union via the Horizon 2020: Future Emerging Topics call (FETOPEN), grant EU736876, project VISORSURF (http://www.visorsurf.eu). The concept of Programmable Wireless Environment was introduced in the viewpoint article [@cacm2018] by the respective authors. Programmable Wireless Environments are provisionally patented by the authors of [@cacm2018] as: ”A New Wireless Communication Paradigm: Realizing Wireless Environments through Software-controlled Metasurfaces”, US 62/539,831, filed 10-Aug-2017. The present work has been published in conference form in [@wowmom2018]. A journal version of the present work appears in [@wcm2018]. [10]{} \[1\][\#1]{} url@samestyle \[2\][\#2]{} \[2\][[l@\#1=l@\#1\#2]{}]{} Z. Pi, J. Choi, and R. Heath, “[Millimeter-wave gigabit broadband evolution toward 5G: fixed access and backhaul]{},” *[IEEE Communications Magazine]{}*, vol. 54, no. 4, pp. 138–144, 2016. C. Liaskos, A. Tsioliaridou, A. Pitsillides, S. Ioannidis and I. Akyildiz, “[Using any Surface to Realize a New Paradigm for Wireless Communications]{},” *[Communications of the ACM]{}*, in press, 2018. C. Liaskos, S. Nie, A. Tsioliaridou, A. Pitsillides, S. Ioannidis and I. Akyildiz, “[A New Wireless Communication Paradigm through Software-controlled Metasurfaces]{},” *[IEEE Communications Magazine]{}*, in press, 2018. C. Liaskos, S. Nie, A. Tsioliaridou, A. Pitsillides, S. Ioannidis and I. Akyildiz, “[Realizing Wireless Communication through Software-defined HyperSurface Environments]{},” *[IEEE Communications Magazine]{}*, In proceedings of the 19th IEEE International Symposium on a World of Wireless, Mobile and Multimedia Networks (IEEE WoWMoM 2018), June 12-15, 2018, Chania, Greece. T. Yilmaz and O. B. Akan, “[Millimetre-Wave Communications for 5G Wireless Networks]{},” *[Opportunities in 5G Networks: A Research and Development Perspective]{}*, pp. 425–440, 2016. A. Aijaz, M. Simsek, M. Dohler, and G. Fettweis, “[Shaping 5G for the Tactile Internet]{},” in *[5G Mobile Comm.]{}*1em plus 0.5em minus 0.4emSpringer, 2017, pp. 677–691. J.-M. Kelif *et al.*, “[A 3D beamforming analytical model for 5G wireless networks]{},” in *[14th WiOpt]{}*, 2016, pp. 1–8. S. V. Hum, M. Okoniewski, and R. J. Davies, “Modeling and design of electronically tunable reflectarrays,” *IEEE transactions on Antennas and Propagation*, vol. 55, no. 8, pp. 2200–2210, 2007. J. Huang, L. I. Wenxiang, Y. Su, and F. Wang, “[Multi-rate combination of partial information based routing and adaptive modulation and coding for space deterministic delay/disruption tolerant networks]{},” *[IET Communications]{}*, 2017. S. Han and K. G. Shin, “[Enhancing Wireless Performance Using Reflectors]{},” in *[INFOCOM 2017]{}*, pp. 1–10. Y. Chen, S. He, F. Hou, Z. Shi, and J. Chen, “[Promoting device-to-device communication in cellular networks by contract-based incentive mechanisms]{},” *[IEEE Network]{}*, vol. 31, no. 3, pp. 14–20, 2017. A. Y. Zhu *et al.*, “[Traditional and emerging materials for optical metasurfaces]{},” *[Nanophotonics]{}*, vol. 6, no. 2, 2017. A. E. Minovich, A. E. Miroshnichenko, A. Y. Bykov, T. V. Murzina, D. N. Neshev, and Y. S. Kivshar, “[Functional and nonlinear optical metasurfaces: Optical metasurfaces]{},” *[Laser [&]{} Photonics Reviews]{}*, vol. 9, no. 2, pp. 195–213, 2015. S. Lucyszyn, *[Advanced RF MEMS]{}*, ser. [The Cambridge RF and microwave engineering series]{}.1em plus 0.5em minus 0.4emNY: Cambridge Univ. Press, 2010. H.-T. Chen, A. J. Taylor, and N. Yu, “[A review of metasurfaces: physics and applications]{},” *[Reports on progress in physics. Physical Society (Great Britain)]{}*, vol. 79, no. 7, p. 076401, 2016. S. H. Lee *et al.*, “[Switching terahertz waves with gate-controlled active graphene metamaterials]{},” *[Nature Materials]{}*, vol. 11, no. 11, pp. 936–941, 2012. B. Banerjee, *[An introduction to metamaterials and waves in composites]{}*.1em plus 0.5em minus 0.4emBoca Raton, FL: CRC Press/Taylor [&]{} Francis Group, 2011. K. Iwaszczuk *et al.*, “[Flexible metamaterial absorbers for stealth applications at terahertz frequencies]{},” *[Optics Express]{}*, vol. 20, no. 1, p. 635, 2012. R. J. Mailloux, *Phased array antenna handbook*.1em plus 0.5em minus 0.4emArtech House Boston, 2005, vol. 2. H. Yang, X. Cao, F. Yang, J. Gao, S. Xu, M. Li, X. Chen, Y. Zhao, Y. Zheng, and S. Li, “A programmable metasurface with dynamic polarization, scattering and focusing control,” *Scientific reports*, vol. 6, p. 35692, 2016. PCBcart. (2017) Printed circuit board calculator. \[Online\]. Available: <https://www.pcbcart.com/quote> M. Caironi, *Large area and flexible electronics*.1em plus 0.5em minus 0.4emJohn Wiley & Sons, 2015. M. A. U. Karim, S. Chung, E. Alon, and V. Subramanian, “Fully inkjet-printed stress-tolerant microelectromechanical reed relays for large-area electronics,” *Advanced Electronic Materials*, vol. 2, no. 5, 2016. R. Parashkov, E. Becker, T. Riedl, H.-H. Johannes, and W. Kowalsky, “Large area electronics using printing methods,” *Proceedings of the IEEE*, vol. 93, no. 7, pp. 1321–1329, 2005. I. F. Akyildiz *et al.*, “[Nanonetworks: A new communication paradigm]{},” *[Computer Networks]{}*, vol. 52, no. 12, pp. 2260–2279, 2008. J. E. Hopcroft, R. Motwani, and J. D. Ullman, *Automata theory, languages, and computation*, 2006, vol. 24. C. Verikoukis *et al.*, “[Internet of Things: Part 2]{},” *[IEEE Communications Magazine]{}*, vol. 55, no. 2, pp. 114–115, 2017. C. L. Holloway *et al.*, “[An Overview of the Theory and Applications of Metasurfaces: The Two-Dimensional Equivalents of Metamaterials]{},” *[IEEE Antennas and Propagation Magazine]{}*, vol. 54, no. 2, pp. 10–35, 2012. R. L. Haupt and D. H. Werner, *[Genetic algorithms in electromagnetics]{}*.1em plus 0.5em minus 0.4emJohn Wiley [&]{} Sons, 2007. J. Chen *et al.*, “Pseudo lateration: Millimeter-wave localization using a single rf chain,” in *Wireless Communications and Networking Conference (IEE WCNC)*, 2017, pp. 1–6. , “[Radiowave Propagation Simulator SE]{},” *http://actix.com*, 2010. M. Yazdi and M. Albooyeh, “[Analysis of Metasurfaces at Oblique Incidence]{},” *[IEEE Transactions on Antennas and Propagation]{}*, vol. 65, no. 5, pp. 2397–2404, 2017. M. Albooyeh *et al.*, “Resonant metasurfaces at oblique incidence: interplay of order and disorder,” *Scientific reports*, vol. 4, p. 4484, 2014. Mathworks. (2017) Genetic algorithm: Finding global optima for highly nonlinear problems. \[Online\]. Available: <https://www.mathworks.com/discovery/genetic-algorithm.html> S. Luke, *[Essentials of metaheuristics]{}*, 1st ed.1em plus 0.5em minus 0.4em: Lulu, 2009. I. F. Akyildiz, M. Pierobon, S. Balasubramaniam, and Y. Koucheryavy, “[The internet of bio-nano things]{},” *[IEEE Communications Magazine]{}*, vol. 53, no. 3, pp. 32–40, 2015. K. Fan *et al.*, “[Optically Tunable Terahertz Metamaterials on Highly Flexible Substrates]{},” *[IEEE Transactions on Terahertz Science and Technology]{}*, vol. 3, no. 6, pp. 702–708, 2013.

ArXiv

--- author: - Patrick Hennebelle and Thierry Passot title: Influence of Alfvén waves on the Thermal Instability in the Interstellar Medium --- Introduction {#intro} ============ It is now well established that the atomic interstellar medium is a thermally bistable gas which at a pressure of about 4000 K cm$^{-3}$ can be in two different phases namely the Warm Neutral Medium (WNM) and the Cold Neutral Medium (CNM) roughly in pressure equilibrium (Kulkarni & Heiles 1987, Field et al. 1969, Wolfire et al. 1995, 2003). It is also believed that the atomic gas is strongly magnetized with magnetic intensity around 5$\mu$G (Troland & Heiles 1986, Heiles 1987, Heiles & Troland 2005). Although the structure of the magnetic field is poorly known, some observational evidences seem to indicate that the fluctuating part is at least comparable to the uniform one suggesting that magnetized waves may be of great importance for the dynamics of the ISM. Another important observational fact is the absence of correlation between the magnetic intensity and the density in the interstellar atomic gas. Various works have investigated the effect of MHD waves on the dynamics of a polytropic or nearly polytropic (non thermally unstable) magnetized gas (Dewar 1970, Goldstein 1978, McKee & Zweibel 1993, Passot et al. 1995, Falle & Hartquist 2002, Passot & Vázquez-Semadeni 2003). It is thus of great interest to investigate the simultaneous role of magnetic fields and of the bistable nature of the flow on the physics of the ISM. Field (1965) considers the effect of a transverse magnetic field on the thermal instability and generalizes the isobaric criterion. In the context of cooling flows, Loewenstein (1990) studies the thermal instability in the presence of a static field. Hennebelle & Pérault (2000) investigates the role of an initially uniform magnetic field, analytically and numerically, when the thermal condensation is dynamically triggered. They propose a mechanism based on magnetic tension to explain the thermal collapse in a magnetized flow and argue that the magnetic intensity in the WNM and in the CNM should not be very different. Piontek & Ostriker (2004, 2005) study the magneto-rotational instability in a thermally unstable gas. Here we investigate the effect of a finite amplitude circularly polarized Alfvén wave on the thermal instability, both analytically and numerically. These Alfvén waves, which are exact solutions of the MHD equations, are non dissipative and are thus very likely to be present in a magnetized gas such as the ISM. Similar waves, although at a smaller scale and in a regime affected by dispersive effects, exist in the solar wind upstream of the earth bow shock (Spangler et al. 1988). In-situ satellite observations have clearly identified circularly polarized quasi-monochromatic Alfvén wave packets probably generated by reflected protons. In Sect. 2 we present a stability analysis of a circularly polarized Alfvén wave propagating in a thermally unstable gas at thermal equilibrium and obtain a dispersion relation which generalizes the relation obtained by Field (1965). We then solve this equation numerically for various regimes and discuss the consequences for the ISM. In Sect. 3, we perform various numerical experiments to confirm the analytic prediction and to investigate the non-linear regime. Section 4 concludes the paper. Analysis {#sec:num} ======== The equations governing, in the magnetohydrodynamic (MHD) limit, the one-dimensional motion of a plasma permeated by a uniform magnetic field ${\bf B_0}$ in a slab geometry read $$\begin{aligned} &&{\partial\rho\over\partial t} +{\partial(\rho u) \over\partial x}=0 \label{eq:mhd1d1}\\ &&{\partial u\over\partial t}+u{\partial u\over\partial x} =-{1\over\rho} {\partial\over \partial x} \left ( P +\frac{|b|^2}{8\pi} \right ) \label{eq:mhd1d2}\\ &&{\partial v\over\partial t}+u{\partial v\over\partial x} = {B_x\over 4\pi\rho}{\partial b\over \partial x} \label{eq:mhd1d3}\\ &&{\partial b\over\partial t}+ {\partial\over \partial x}(ub)=B_x{\partial v\over \partial x}, \label{eq:mhd1d4}\\ && (\frac{\partial T}{\partial t}+u\frac{\partial T}{\partial x})=-(\gamma-1)T\frac{\partial u}{\partial x} - \frac{1}{C_v}{\cal L} + \frac{1}{\rho C_v} \partial _x (\kappa(T) \partial_x T )\end{aligned}$$ where all fields only depend on the coordinate $x$ and time $t$. The ambient field ${\bf B_0}$ is assumed to be oriented in the $x$ direction. The velocity field ${\bf V}$ has a component $u=V_x$ along the $x$ coordinate and two transverse components combined in the complex number $v=V_y+iV_z$. Similarly, we write the magnetic field as $b=B_y+iB_z$, the component $B_x=B_0$ remaining constant. The mass density and thermal pressure are denoted by $\rho$ and $P$ respectively and we assume a perfect gas law $\displaystyle{P=\frac{R \rho T}{\mu}}$ where $\displaystyle{R=\frac{k_B}{m_H}}$ is the universal gas constant, $k_B$ the Boltzmann constant and $\mu$ the mean molecular weight in units of the hydrogen mass $m_H$. Heating and cooling processes are combined in a single net cooling function ${\cal L}$. The parameter $\gamma$ denotes the ratio of the specific heats at constant pressure $C_p$ and at constant volume $C_v$. Note that $\displaystyle{C_v=\frac{R}{(\gamma-1)\mu}}$. Thermal diffusivity due to neutrals is the dominant one. It is isotropic in spite of the presence of the magnetic field and equal to $\kappa(T)=5/3 C_v \eta(T)$, where $\eta(T)=5.7 \; 10^{-5} (T/1 K)^{1/2}$ g cm$^{-1}$ s$^{-1}$ (Lang 1974). Throughout most of Sect. 2, thermal diffusivity will be neglected for simplicity, leading to assume a vanishingly small Field length. Its effect will briefly be addressed in Sect. 2.5 and it is explicitely taken into account in the numerical simulations of Sect. 3. Two main non-dimensional numbers can be defined, the sonic Mach number $M_s={V_0/c_s}$ ratio of a typical velocity $V_0$ with the constant sound speed $\displaystyle{c_s=\sqrt{\frac{\gamma k_B T}{\mu m_H}}}$ and the Alfvénic Mach number $M_a={V_0/c_a}$, where $c_a={B_x/(4\pi \rho_0)^{1/2}}$ is the Alfvén speed of the unperturbed system. The plasma beta is here defined by $\displaystyle{\beta={M_a^2\over M_s^2}}$. These equations have exact solutions in the form of circularly polarized plane Alfvén waves of arbitrary amplitude (Ferraro, 1955). They read $b_0=B_\perp\exp\left[ {-i\sigma (k_0 x-\omega_0 t)}\right]$ with constant density $\rho_0$ and temperature $T_0$, and zero longitudinal velocity $u_0=0$. The transverse wave velocity is related to the magnetic field perturbation by $\displaystyle{v_0=V_\perp \exp\left[ {-i\sigma (k_0 x-\omega_0 t)}\right]=-\frac{b_0}{\sqrt{4\pi\rho_0}}}$. The polarization of the wave is determined by the parameter $\sigma$, with $\sigma=+1$ ($\sigma=-1$) for a right-handed (left-handed) wave. In absence of dispersive effect we can, without restriction, take $\sigma=-1$. Moreover, $B_\perp$ can be taken real. The dispersion relation reads $\omega_0^2 =c_a^2 k_0^2$. Ambipolar diffusion and wave steepening --------------------------------------- Before considering the idealized configuration of a completely ionized medium with a circularly polarized monochromatic Alfvén wave train of infinite length, it is important to estimate the timescale of two important processes which are not considered in this study. The first one is the ambipolar diffusion which takes place in weakly ionized gas and which induces energy dissipation due to the friction between the neutrals and the ions. The second one is the steepening of the MHD waves that takes place when the wave polarization is not perfectly circular or when the wave packet is modulated. ### Timescale of the ambipolar diffusion The ambipolar diffusion time is given by (Shu 1992): $$t _{da} = { 4 \pi \gamma _{\rm da} \rho _n \rho _i L^2 \over B^2 },$$ where $\gamma _{\rm da}$ is the friction coefficient between ions of density $\rho_i$ and neutrals of density $\rho_n$, $L$ is the typical scale to be considered and $B$ is the magnetic intensity. In the case of a molecular cloud, it has been estimated to $3 \times 10 ^{13} \, {\rm cm^{3} \, g ^{-1} \, s^{-1}}$ (Shu 1992). In the case of the atomic gas, since the mass of the neutral is twice smaller, $\gamma _{\rm da}$ is twice larger, leading to $$t _{da} \simeq 5 10^{15} \left( { 5 \mu G \over B } \right) ^2 \, { \xi _i \over 10^{-3} } \, \left( { n _n \over 100 \, {\rm cm ^{-3} } } \right) ^2 \, \left(L \over { 10^{18} \, {\rm cm} } \right) ^2 {\rm s},$$ where $\xi_i$ is the gas ionization and $n_n$ the particle density of the neutrals. In the CNM the ionization is about $4 \times 10 ^{-4}$, therefore at the scale of say 0.1 pc, the ambipolar diffusion time is about $2\times 10^{14}$ s or about $10 Myr$. In the WNM, the density is about 1 $cm^{-3}$, the ionization about 0.1, therefore at a scale of say 1 pc, the ambipolar diffusion time is $5\times 10^{14}$ s. In both cases, ambipolar diffusion operates on a rather long characteristic time and can thus safely be neglected. ### Timescale of the wave steepening The equation governing the dynamics of parallel propagating Alfvén waves in the long wavelength, small amplitude limit was derived by Cohen & Kuslrud (1974). A similar equation was derived previously by Rogister (1971) from the Vlasov-Maxwell system in the context of a collisionless plasma, taking into account kinetic effects such as dispersion and Landau damping. The main point is that the evolution of the magnetic field components perpendicular to the ambient field involves a nonlinearity of the form $\alpha \partial_x \left ( |b|^2 b \right )$, where $\alpha$ is a coefficient depending on $\beta$. From this formula, one sees in particular that circularly polarized Alfvén waves, for which $|b|$ is constant, do not steepen. In the case where the wave amplitude is modulated, the characteristic time of steepening is given by (using Eq. (40) in Cohen and Kulsrud 1974) $$\tau=\frac{4}{3}\frac{1-\beta}{c_A}\frac{B_0^2}{\partial_x |b|^2}.$$ Denoting by $\lambda$ the wavelength, by $l_c$ the typical length of the amplitude modulation and by $\omega_s$ the characteristic growth rate associated with wave steepening, one gets $$\frac{\omega_s}{\omega_0}= \frac{3}{4(1-\beta)}\frac{\lambda}{l_c}\frac{b_0^2}{B_0^2}.\label{steepening}$$ It is to be noted that Cohen-Kulsrud equation is not valid for $\beta$ close to unity, a point where sound waves have the same phase speed as Alfvén waves and have thus to be retained explicitely in the nonlinear dynamics. In general however, we see that the ratio $\omega_s/\omega_0$ depends quadratically on the wave amplitude and is inversely proportional to the normalized scale of amplitude modulation. As an example, for a wave train of amplitude $b_0/B_0=0.5$ and coherence length $l_c/\lambda=5$, with $\beta=0.5$, one gets $\omega_s/\omega_0=0.075$, a value about five times smaller than the typical growth rates shown in Fig. 1. This process of wave steepening thus appears subdominant with respect to the combined thermal and decay instabilities considered in this paper. Moreover, as will be demonstrated by the numerical simulations, wave steepening does indeed occur but it does not affect our conclusions. Another process is at play in real three-dimensional situations, namely turbulent cascade, which occurs mainly in transverse directions as a result of nonlinear interaction between counter-propagating wave packets. The nonlinear eddy turnover time is expected to be smaller than the other processes mentioned above for high enough amplitude and situations where Alfvén waves propagate in both directions in roughly equal amount. When waves propagate in a privileged direction, as in the solar wind, the turbulent cascade is much less efficient, leaving enough time for the other instabilities to develop. The interaction between waves and turbulence in compressible MHD is still a very debated topic (see e.g. Cho & Lazarian 2003). Three-dimensional simulations will have to be performed to address the competition between all these phenomena. Derivation of the dispersion relation ------------------------------------- We shall now consider perturbations about the exact solution described above, namely a circularly polarized Alfvén wave propagating in a uniform gas at thermal equilibrium and determine the dispersion relation for normal modes. Such an analysis has been performed by Goldstein (1978) for an isothermal gas and by Lou (1996) for a self-gravitating gas. Let us write the perturbations as $\displaystyle{\rho = \rho _0 + \delta \rho , T= T _0 + \delta T , v = v_0 + \delta v , b= b_0 + \delta b , u = \delta u }$. The linearized equations write, denoting by a star the complex conjugate, $$\begin{aligned} \nonumber &&\partial _t \delta \rho + \rho _0 \partial _x \delta u = 0, \\ \nonumber &&\rho _0 \partial _t \delta u = - \partial _x \delta P - {1 \over 8 \pi} \partial _x (b_0 \delta b^* + b_0^* \delta b ), \\ &&\rho _0 \partial _t \delta v + \delta \rho \partial _t v_0 + \rho_0 \delta u \partial_x v_0 = {1 \over 4 \pi} B_x \partial _x \delta b, \\ \nonumber &&\partial _t \delta b + \delta u \partial _x b_0 + b_0 \partial _x \delta u- B_x \partial _x \delta v = 0, \\ \nonumber &&C _v( \partial _t \delta T + (\gamma -1) T_0 \partial _x \delta u) = - ( \Big (\frac{\partial {\cal L}}{\partial \rho}\Big )_T \delta \rho+ \Big (\frac{\partial {\cal L}}{\partial T}\Big )_{\rho} \delta T ). \end{aligned}$$ Taking perturbations in the form $$\begin{aligned} \nonumber \delta \rho &=& \widehat{ \delta \rho} \exp(-i \omega t + i k x ) + c.c. , \\ \nonumber \delta T &=& \widehat{ \delta T} \exp(-i \omega t + i k x ) + c.c., \\ \delta u &=& \widehat{ \delta u} \exp(-i \omega t +i k x ) + c.c., \\ \delta v &=& \widehat{ \delta v}^+ \exp(-i (\omega+\omega_0) t + i (k+k_0) x ) + (\widehat{ \delta v}^{-})^* \exp(i (\omega-\omega_0) t - i (k-k_0) x ) , \nonumber \\ \nonumber \delta b &=& \widehat{ \delta b}^+ \exp(-i (\omega+\omega_0) t + i (k+k_0) x + (\widehat{ \delta b}^-)^* \exp(i (\omega-\omega_0) t - i (k-k_0) x ),\end{aligned}$$ we obtain the following equations $$\begin{aligned} \nonumber -\omega \widehat{\delta \rho} + k\rho _0 \widehat{\delta u} &=& 0, \\ \nonumber -\rho _0 (\omega + \omega _0) \widehat{\delta v}^+ + (-\omega _0 \widehat{\delta \rho} + k _0\rho _0 \widehat{\delta u} ) V _\perp &=& {B_x \over 4 \pi} (k+k _0) \widehat{\delta b}^+ , \\ \nonumber -\rho _0 (\omega - \omega _0) \widehat{\delta v}^- + (\omega _0 \widehat{\delta \rho} - k_0\rho _0 \widehat{\delta u} ) V _\perp &=& {B_x \over 4 \pi} (k - k _0) \widehat{\delta b}^-, \\ -\omega\rho _0 \widehat{\delta u} + k {k _B \over\mu m_H} (\rho_0 \widehat{\delta T} + T_0 \widehat{\delta \rho}) + {B _\perp \over 8 \pi } k (\widehat{\delta b}^+ + \widehat{\delta b}^-) &=& 0 , \\ \nonumber -(\omega + \omega_0) \widehat{\delta b}^+ + B _\perp(k+k _0) \widehat{\delta u} - B _x (k +k_0) \widehat{\delta v}^+ &=& 0, \\ \nonumber -(\omega - \omega_0) \widehat{\delta b}^- +B_\perp(k- k _0)\widehat{\delta u} - B_x (k - k_0) \widehat{\delta v}^- &=& 0, \\ \nonumber i C_v ( -\omega \widehat{\delta T} + (\gamma -1) T _0 k \widehat{\delta u}) &=& - \left( \partial _\rho {\cal L} \widehat{\delta \rho} + \partial _T {\cal L} \widehat{\delta T} \right) .\end{aligned}$$ The fluctuations $\widehat{\delta T}$, $\widehat{\delta \rho}$, and $\widehat{\delta b}^{\pm}$ can easily be obtained as a function of $\widehat{\delta u}$ in the form $$\begin{aligned} \nonumber \omega \widehat{\delta \rho} &=& k \rho _0 \widehat{\delta u}, \\ \widehat{\delta T} &=& \frac{\widehat{\delta u}}{- \partial _T {\cal L} + i C_v \omega} \left( ik C_v (\gamma -1) T _0 + {k \over \omega }\rho _0 \partial _\rho {\cal L} \right), \\ \nonumber \widehat{\delta b}^\pm &=& { \widehat{\delta u} \over \omega _\pm ^2 - c_a^2 k _\pm ^2} B _\perp k _\pm { \omega _0 ^2 k \over \omega k _0} \left( { \omega _\pm \over \omega _0} {\omega \over \omega _0 } {k _0 \over k} \pm {\omega \over \omega _0} {k _0 \over k } - \pm 1 \right). \end{aligned}$$ One then gets the dispersion relation $$\begin{aligned} \nonumber \left( -\omega ^2 + k^2 {k_B \over \mu m_H} T _0 + k^2 {k_B \over\mu m_H} \left( { i C_v (\gamma -1) T_0 \omega + \rho _0\partial_\rho {\cal L} \over - \partial _T {\cal L} + i C _v \omega } \right) \right) & \left( {\omega\over\omega _0} - {k \over k_0} \right) \left( \left({\omega\over\omega_0} + {k \over k_0} \right)^2 - 4 \right) = \\ -\omega _0 ^2 \left( { B _\perp \over B_x} \right)^2 \left( {k \over k_0} \right) ^2 \left( \left({\omega \over \omega _0}\right)^3 + {k \over k _0} \left( {\omega \over \omega _0} \right)^2 -3 {\omega \over \omega _0} + {k \over k _0} \right) , & \label{dispersion}\end{aligned}$$ which can be rewritten $$\begin{aligned} \widetilde{\omega}^2-\beta \widetilde{k}^2 \frac {\widetilde {\omega} + i \widetilde {\omega_b}} {\widetilde{\omega}+ i \widetilde {\omega_c} } = \widetilde{k}^2 A^2 \frac {\widetilde{\omega}^3 + \widetilde{k} \widetilde{\omega}^2 -3 \widetilde{\omega} + \widetilde{k} } {(\widetilde{\omega} - \widetilde{k} ) ( (\widetilde{\omega} + \widetilde{k} )^2 - 4)}, \label{disp_rel}\end{aligned}$$ where $\displaystyle{\widetilde{\omega_c}=\frac{\omega_c}{\omega_0}=\frac{\partial_T {\cal L}}{C_v\omega_0}}$, $\displaystyle{\widetilde{\omega_b}=\frac{\omega_b}{\omega_0}=\frac{ T_0\partial_T {\cal L}-\rho_0\partial_{\rho} {\cal L} }{\gamma C_v T_0 \omega_0}}=\frac{1}{\gamma C_v\omega_0} \Big (\frac{\partial {\cal L}}{\partial T}\Big )_P$, $\widetilde{\omega}=\omega/\omega_0$, $\widetilde{k}=k/k_0$ and $A=B_\perp/B_x$. Note that in the limit $A=0$, Eq. (\[disp\_rel\]) becomes identical with the dispersion relation obtained by Field (1965) whereas if ${\cal L}=0$, it reduces to the dispersion relation obtained by Goldstein (1968). Asymptotic behaviors -------------------- Before numerically solving Eq. (\[disp\_rel\]), we consider various asymptotic limits. ### Static magnetic field It is possible to recover the dispersion relation in the absence of waves for a situation where the ambient field is oblique, making an angle $\theta$ with the $x$-axis. In this case $B_x=B_0\cos\theta$ and one must take the limit $\omega_0\rightarrow 0$ with $\displaystyle{\frac{\omega_0}{k_0}=c_A=\frac{B_0}{\sqrt{4\pi\rho_0}}\cos\theta}$ and $A=\tan \theta$. It follows that $$\omega^2- k^2 c_s^2 \frac{\omega+i\omega_b}{\omega+i\omega_c}= \frac{ k^2\omega^2v_A^2\sin^2\theta}{\omega^2-k^2v_A^2\cos^2\theta}, \label{bstat}$$ where we denote $\displaystyle{v_A^2=\frac{B_0^2}{4\pi\rho_0}}$. In the case $\theta=\pi/2$ and close to the threshold ($\omega \simeq 0 \ll kc_s$), we have $$\omega \simeq -i \frac{v_A^2\omega_c+c_s^2\omega_b}{v_A^2+c_s^2}. \label{cas3}$$ Since the magnetic field is purely transverse, the magnetic tension vanishes and the magnetic pressure adds up to the thermal pressure making the gas more stable. The criterion for thermal instability is simply $v_A^2\omega_c+c_s^2\omega_b \le 0$ which was first obtained by Field (1965). It shows that a purely transverse magnetic field can suppress the thermal instability. This is because, in this geometry, the magnetic pressure is proportional to the square of the density and therefore $\partial _\rho P _{\rm tot} $ can be positive even if $\partial _\rho P _{\rm therm}$ is negative. ### Instability thresholds This section addresses the neighborhood of the instability, a situation where $\Im(\omega) \rightarrow 0$. We shall first discuss the case $\Re(\omega) = 0$, corresponding to the so-called condensation or entropy mode. In the situation where $|\widetilde{\omega_b}|$ and $|\widetilde{\omega_c}|$ are smaller than $\widetilde{k}$, the characteristic cooling time is longer than the Alfvén crossing time and the dispersion relation reduces to $$\widetilde{\omega} \simeq- i\frac{\beta (4-\widetilde{k}^2)\widetilde{\omega_b}+A^2\widetilde{\omega_c}} {\beta(4-\widetilde{k}^2)+A^2}.\label{disp}$$ When $\beta$ is not too small sound waves have time to restore pressure equilibrium while the gas cools and one expects that the growth rate will be close to the isobaric one. In the opposite case where $\beta$ is very small, the growth rate should be close to the isochoric one since Alfvén waves are not accompanied by pressure or density perturbations. These conclusions are easily recovered form Eq. (\[disp\]) which for very small amplitude ($A^2/\beta(4-\widetilde{k}^2) \ll 1$) gives $$\widetilde{\omega} \simeq -i \Big(\widetilde{\omega_b}+ \frac{A^2}{\beta(4-\widetilde{k}^2)}\Big (\widetilde{\omega_c} -\widetilde{\omega_b} \Big) \Big ),$$ while, in the case where $\beta(4-\widetilde{k}^2)/A^2 \ll 1$ it rewrites $$\widetilde{\omega} \simeq -i \Big(\widetilde{\omega_c}+ \frac{\beta(4-\widetilde{k}^2)}{A^2}\Big (\widetilde{\omega_b} -\widetilde{\omega_c} \Big) \Big ).$$ In the case where $A^2/\beta \ll 1$, more likely to be met in the ISM, the effect of the Alfvén wave depends both on the sign of $\widetilde{\omega_c} -\widetilde{\omega_b}$ and of that of $\widetilde{k}- 2$. In a typical region of the ISM with T$\simeq$ 1000 K, $\widetilde{\omega_c}-\widetilde{\omega_b} \ge 0 $ and therefore the waves stabilize (destabilize) the gas if $\widetilde{k} < 2$, (respectively $\widetilde{k} \ge 2$). In the limit $\widetilde{k} \gg 1 $ and/or for $A=0$, one recovers the isobaric growth rate, the destabilizing effect of the Alfvén waves becoming asymptotically small as $\widetilde{k}$ increases. If $\widetilde{\omega_b}\ll\widetilde{k} \ll 1 $, we find a growth rate similar to that given by Eq. (\[cas3\]) except for the factor 2 that divides $v_A^2$. This is due to the magnetic tension that tends to unbend the magnetic field lines making the stabilization of the magnetic pressure less efficient. Two different limits are obtained when $\widetilde{k} \ll\widetilde{\omega_b}$. When the isochoric criterion is not verified and $\widetilde{\omega}/\widetilde{k}\equiv\alpha$ remains finite, a situation where the growth rate vanishes with $\widetilde{k}$, one gets $$(1-\alpha)(\alpha^2-\beta\frac{\widetilde{\omega_b}}{\widetilde{\omega_c}})= (1-3\alpha)\frac{A^2}{4}. \label{sk}$$ In the absence of waves, the growth rate asymptotically approaches $\displaystyle{\widetilde{\omega}= \pm\beta^{\frac{1}{2}}\widetilde{k}(\widetilde{\omega_b}/ \widetilde{\omega_c})^ {\frac{1}{2}}}$ (Meerson 1996). Equation (\[sk\]) shows in general, assuming $\alpha$ and $A$ small and thus $\displaystyle{\alpha^2\approx \beta\widetilde{\omega_b}/ \widetilde{\omega_c}+A^2/4}$, that the Alfvén waves have a stabilizing effect. When $\widetilde{\omega_c}<0$, the growth rate for $\widetilde{k} \ll|\widetilde{\omega_c}|$ remains finite, equal to $\widetilde{\omega}=-i\widetilde{\omega_c}$, independently of the presence of Alfvén waves. Thus, in the limit of very short and long wavelengths, Alfvén waves do not modify the growth rates of the thermal instability obtained by Field (1965). For completeness, we now consider the case of adiabatic perturbations, corresponding to a wave mode with $\Re(\omega) \ne 0 $. For this purpose, we restrict ourselves to the case of low amplitude waves, i.e $A \ll 1$ and we set $\widetilde{\omega} =\widetilde{\omega} _r + i \widetilde{\omega}_i$. We therefore have $\widetilde{\omega}_r \gg \widetilde{\omega}_i$. Moreover, it is assumed that $\widetilde{\omega_c}\ll 1$ and $\widetilde{\omega_b}\ll 1$ so that they can be neglected when multiplied by $\widetilde{\omega}_i$. With these assumptions, one finds that to the first order $$\begin{aligned} \widetilde{\omega} &=& \beta^{\frac{1}{2}}\widetilde{k} +\frac{i}{2}(\widetilde{\omega_b}-\widetilde{\omega_c}) +\frac{A^2}{2\beta}\Big (\frac{\beta^{\frac{1}{2}}\widetilde{k} +i\widetilde{\omega_c}}{\beta^{\frac{1}{2}} -1}\Big )\Big(\frac{(\beta^{\frac{3}{2}} +\beta)\widetilde{k}^2 +1-3\beta^{\frac{1}{2}}} {(1+\beta^{\frac{1}{2}})^2\widetilde{k}^2-4}\Big ). \label{sonic_crit}\end{aligned}$$ The real part, $\omega _r \simeq C_s k$ simply corresponds to a sonic wave. When $A=0$, one finds that the conditions for thermal stability of a sonic wave is $\widetilde{w _c} - \widetilde{w _b} \ge 0$, a criterion already obtained by Field (1965). When $A \ne 0$, the stability criterion is modified according to Eq. (\[sonic\_crit\]), showing that the effect of the waves depends in a complex way on $\beta$ and $\widetilde{k}$. Growth rate and physical discussion {#sec:rel} ----------------------------------- (0,12) (0,9)[![Growthrate as a function of the wave number. First panel is for the adiabatic case, second, third and fourth panels display results for various values of $\omega _c / \omega _0$. Solid line corresponds to the case $A=0$, dotted line to $\beta=0.9$, $A=0.5$, short dashed line to $\beta=0.5$, $A=0.5$, dot-dashed line to $\beta=0.5$, $A=1$, double dot-dashed line to $\beta=0.1$, $A=0.5$ and long dashed line to $\beta=0.1$, $A=1$. []{data-label="growthrate"}](fig1.ps "fig:"){width="7cm"}]{} (0,6)[![Growthrate as a function of the wave number. First panel is for the adiabatic case, second, third and fourth panels display results for various values of $\omega _c / \omega _0$. Solid line corresponds to the case $A=0$, dotted line to $\beta=0.9$, $A=0.5$, short dashed line to $\beta=0.5$, $A=0.5$, dot-dashed line to $\beta=0.5$, $A=1$, double dot-dashed line to $\beta=0.1$, $A=0.5$ and long dashed line to $\beta=0.1$, $A=1$. []{data-label="growthrate"}](fig2.ps "fig:"){width="7cm"}]{} (0,3)[![Growthrate as a function of the wave number. First panel is for the adiabatic case, second, third and fourth panels display results for various values of $\omega _c / \omega _0$. Solid line corresponds to the case $A=0$, dotted line to $\beta=0.9$, $A=0.5$, short dashed line to $\beta=0.5$, $A=0.5$, dot-dashed line to $\beta=0.5$, $A=1$, double dot-dashed line to $\beta=0.1$, $A=0.5$ and long dashed line to $\beta=0.1$, $A=1$. []{data-label="growthrate"}](fig3.ps "fig:"){width="7cm"}]{} (0,0)[![Growthrate as a function of the wave number. First panel is for the adiabatic case, second, third and fourth panels display results for various values of $\omega _c / \omega _0$. Solid line corresponds to the case $A=0$, dotted line to $\beta=0.9$, $A=0.5$, short dashed line to $\beta=0.5$, $A=0.5$, dot-dashed line to $\beta=0.5$, $A=1$, double dot-dashed line to $\beta=0.1$, $A=0.5$ and long dashed line to $\beta=0.1$, $A=1$. []{data-label="growthrate"}](fig4.ps "fig:"){width="7cm"}]{} We now numerically solve Eq. (\[disp\_rel\]) which can be rewritten as a sixth order polynomial containing 4 parameters to be specified, namely $\beta$ and $A$ (characterizing the Alfvén wave) and $\widetilde{\omega}_b$ and $\widetilde{\omega} _c$ (function of the thermal processes). The value of $\widetilde{\omega} _c = \omega _c / \omega _0 $ represents the ratio of the temporal period of the Alfvén wave divided by the cooling time and can be arbitrarily chosen. However, once $\widetilde{\omega} _c$ is specified, $\widetilde{\omega} _b$ depends on the thermal function. In order to estimate this parameter we use the standard cooling function of the neutral atomic ISM (Wolfire et al. 1995, 2003) which is used in Audit & Hennebelle (2005) in the thermally unstable regime ($n=3 \, $cm$^{-3}$ and $T \simeq 500$ K). To obtain the dispersion relation we integrate Eq. (\[dispersion\]) for $k/k_0$ between 0.01 and 100 using logarithmic spacing. The roots of the polynomials are obtained using the [*zroots*]{} subroutine (Press et al. 1992). Here, we restrict our attention to the unstable branch only. In order to verify our method, we have reproduced the dispersion relation for the decay instability of an isothermal gas presented in Goldstein (1978). Fig. \[growthrate\] displays the results of the numerical integration of Eq. (\[dispersion\]). First panel shows the adiabatic case for $\beta=0.9$, $A=0.5$ (dotted line), $\beta=0.5$, $A=0.5$ (short dashed line), $\beta=0.5$, $A=1$ (dot-dashed line), $\beta=0.1$, $A=0.5$ (double dot-dashed line), $\beta=0.1$, $A=1$ (long dashed line). As expected the circularly polarized Alfvén wave is unstable (decay instability) in a range of $k$ extending to a few times $k_0$. Both the growth rate and the largest unstable value of $k$, increase with $\beta^{-1}$ and $A$. The second panel shows results for $\widetilde{\omega} _c = \omega _c / \omega _0 = 1$. The full line corresponds to the hydrodynamical case whereas the others correspond to the same values of $\beta$ and $A$ as in the first panel. Various interesting features can be seen. i) When $k \rightarrow 0$ the effect of the Alfvén wave is to decrease the growth rate and therefore to stabilize these modes with respect to thermal instability. For $\beta =0.5, A=1$ and for $\beta=0.1$, the modes whose wavenumber is smaller than $k \simeq 1$ are perfectly stable. ii) The intermediate modes (i.e $k \simeq k_0$) are more unstable when $\beta$ is smaller and $A$ is higher. This is due to the decay instability that the wave undergoes for these values of $k$. iii) When $k \rightarrow \infty$, the growth rate is independent of $\beta$ and $A$. These features are in good agreement with the asymptotic limit $\omega \simeq 0$ studied in the previous section. Cases with a larger $\beta$, namely 1 and 1.5, have also been explored but are not exposed here for conciseness since they are not directly relevant for the regions of the ISM we consider in this paper. Although the decay instability disappears for $\beta \ge 1$ (for small enough amplitude), we find no qualitative difference with the cases $\beta \le 1$. In particular the intermediate wavelengths are still destabilized by the waves. This is in good agreement with the analytical study of the instability threshold (see Eq. (\[disp\])). The third panel shows results for $\widetilde{\omega} _c = 0.1$. In this case the cooling time is 10 times larger than the period of the Alfvén waves. The dispersion relation is qualitatively similar to the previous case. Quantitatively however, the intermediate wavelengths are much more unstable than the small wavelengths. This is due to the fact that in this range of parameters, the cooling time being larger than the dynamical time of the waves, the fastest growing instability is the decay instability. Note that since $\omega _0$ is 10 times larger in this case than in the case displayed in panel 2, the value of $\omega / \omega _0$ in the hydrodynamical case (full line) is 10 times lower. Fourth panel shows results for $\widetilde{\omega} _c = 10$. In this case, the growth rate of the thermal instability is much shorter than the growth rate of the decay instability. Therefore the only effect of the Alfvén wave is to stabilize the gas with respect to thermal instability. This effect increases when $\beta$ decreases and when $A$ increases. These results suggest that the presence of non-linear circularly polarized Alfvén waves in a thermally unstable medium like the neutral interstellar atomic gas can have two main effects. If the cooling time is short with respect to the temporal period of the waves, then the waves stabilize the gas. Therefore the gas can survive longer in the thermally unstable domain leading possibly in the ISM, to a larger fraction of thermally unstable gas. Since the short wavelengths are the most unstable the trend is that the CNM is very fragmented into several clouds. If the cooling time is larger than the wave period, then the decay instability makes the intermediate modes ($k \simeq k _0$) more unstable. In that case, the fraction of thermally unstable gas is not necessarily larger (depending on $\beta$ and $A$) but the CNM should be fragmented in structures having a size of about $1/(k _0 \times \xi)$, where $\xi$ is the density ratio between the CNM and the WNM. Effect of thermal diffusivity ----------------------------- (0,9) (0,6)[![Growthrate as a function of the wave number when the effect of the thermal diffusivity is taken into account. The curve styles are associated with the same values of $\beta$ and $A$ as in Fig. \[growthrate\].[]{data-label="growthrate_cond_therm"}](growingtime3.ps "fig:"){width="7cm"}]{} (0,3)[![Growthrate as a function of the wave number when the effect of the thermal diffusivity is taken into account. The curve styles are associated with the same values of $\beta$ and $A$ as in Fig. \[growthrate\].[]{data-label="growthrate_cond_therm"}](growingtime2.ps "fig:"){width="7cm"}]{} (0,0)[![Growthrate as a function of the wave number when the effect of the thermal diffusivity is taken into account. The curve styles are associated with the same values of $\beta$ and $A$ as in Fig. \[growthrate\].[]{data-label="growthrate_cond_therm"}](growingtime1.ps "fig:"){width="7cm"}]{} Here we briefly consider the effect of thermal diffusivity. When this term is taken into account, the dispersion relation (\[disp\_rel\]) becomes: $$\begin{aligned} \widetilde{\omega}^2-\beta \widetilde{k}^2 \frac {\widetilde {\omega} + i (\widetilde {\omega_b} + \widetilde{\kappa_b} \widetilde{k}^2 ) } {\widetilde{\omega}+ i (\widetilde {\omega_c} + \widetilde{\kappa_c} \widetilde{k}^2)} = \widetilde{k}^2 A^2 \frac {\widetilde{\omega}^3 + \widetilde{k} \widetilde{\omega}^2 -3 \widetilde{\omega} + \widetilde{k} } {(\widetilde{\omega} - \widetilde{k} ) ( (\widetilde{\omega} + \widetilde{k} )^2 - 4)}, \label{disp_rel_therm}\end{aligned}$$ where $\widetilde{\kappa_c} = \kappa (T_0) \omega _0 \beta / (C_v \rho_0 C_s^2)$ and $\widetilde{\kappa_b} = \widetilde{\kappa_c} / \gamma$. Fig. \[growthrate\_cond\_therm\] displays the growth rate obtained from Eq. (\[disp\_rel\_therm\]) using the fiducial value of $\kappa (T_0)$ given at the beginning of Sect. 2 and for $\omega_c=0.3, \; 1, \; 3$. The different curves are associated with the same values of $\beta$ and $A$ as in Fig. (\[growthrate\]) For small and intermediate values of $\widetilde{k}$ the shape and the values of $\widetilde{w}$ are very similar to the case of vanishing thermal conductivity. As expected however, thermal conduction introduces a cut-off at small scale (Field 1965). The value of $\widetilde{k}$ for which $\widetilde{\omega}$ vanishes depends on $\beta$ and increases (by a factor 2 to 3 for the values considered here) when $\beta$ decreases, confirming the trends inferred previously, i.e. the CNM should be more fragmented in the presence of Alfvén waves leading to smaller CNM structures. Numerical study {#sec:simu} =============== (0,13) (0,9)[![Density field for 3 snapshots illustrating the development of the thermal instability in the hydrodynamical case. The density of the perturbation increases until the gas reaches thermal equilibrium. []{data-label="hydro"}](bn0_1.ps "fig:"){width="7cm"}]{} (0,5)[![Density field for 3 snapshots illustrating the development of the thermal instability in the hydrodynamical case. The density of the perturbation increases until the gas reaches thermal equilibrium. []{data-label="hydro"}](bn0_2.ps "fig:"){width="7cm"}]{} (0,1)[![Density field for 3 snapshots illustrating the development of the thermal instability in the hydrodynamical case. The density of the perturbation increases until the gas reaches thermal equilibrium. []{data-label="hydro"}](bn0_3.ps "fig:"){width="7cm"}]{} (0,25) (0,17)[![Density and y-component of the magnetic field for 3 snapshots illustrating the development of the thermal instability in the presence of Alfvén waves of amplitude $B _\perp=1 \, \mu$G and a longitudinal magnetic field $B_x=5 \, \mu$G.[]{data-label="b5a1"}](bn5_bt1_1.ps "fig:"){width="7cm"}]{} (0,9)[![Density and y-component of the magnetic field for 3 snapshots illustrating the development of the thermal instability in the presence of Alfvén waves of amplitude $B _\perp=1 \, \mu$G and a longitudinal magnetic field $B_x=5 \, \mu$G.[]{data-label="b5a1"}](bn5_bt1_2.ps "fig:"){width="7cm"}]{} (0,1)[![Density and y-component of the magnetic field for 3 snapshots illustrating the development of the thermal instability in the presence of Alfvén waves of amplitude $B _\perp=1 \, \mu$G and a longitudinal magnetic field $B_x=5 \, \mu$G.[]{data-label="b5a1"}](bn5_bt1_3.ps "fig:"){width="7cm"}]{} (0,25) (0,17)[![Same as Fig. \[b5a1\] for $B _\perp=2.5 \, \mu$G and $B_x=5 \, \mu$G.[]{data-label="b5a2.5"}](bn5_1.ps "fig:"){width="7cm"}]{} (0,9)[![Same as Fig. \[b5a1\] for $B _\perp=2.5 \, \mu$G and $B_x=5 \, \mu$G.[]{data-label="b5a2.5"}](bn5_3.ps "fig:"){width="7cm"}]{} (0,1)[![Same as Fig. \[b5a1\] for $B _\perp=2.5 \, \mu$G and $B_x=5 \, \mu$G.[]{data-label="b5a2.5"}](bn5_4.ps "fig:"){width="7cm"}]{} (0,25) (0,17)[![Same as Fig. \[b5a1\] for $B _\perp=2.5 \, \mu$G and $B_x=2.5 \, \mu$G.[]{data-label="b2.5a2.5"}](bn2.5_1.ps "fig:"){width="7cm"}]{} (0,9)[![Same as Fig. \[b5a1\] for $B _\perp=2.5 \, \mu$G and $B_x=2.5 \, \mu$G.[]{data-label="b2.5a2.5"}](bn2.5_3.ps "fig:"){width="7cm"}]{} (0,1)[![Same as Fig. \[b5a1\] for $B _\perp=2.5 \, \mu$G and $B_x=2.5 \, \mu$G.[]{data-label="b2.5a2.5"}](bn2.5_4.ps "fig:"){width="7cm"}]{} In order to test the analytic results presented in the previous section and to investigate the non-linear regime, numerical simulations are performed in a slab geometry. For this purpose we use the 1D adaptive mesh refinement (AMR) code presented in Hennebelle & Pérault (1999, 2000). The AMR technique is very helpful to simultaneously resolve the sharp thermal fronts ($\simeq 10^{-3}$ pc) and the larger scale ($\ge$ 10 pc) involved in the problem. The code has been extensively tested and the results in the hydrodynamical case have been closely compared with high resolution simulations using a second order Godunov scheme. The growth rate for the parametric instability of an Alfvén wave in an adiabatic gas have been calculated and shown to match, within an accuracy of a few percents, the results of the first panel of Fig. \[growthrate\]. Two numerical experiments are carried out. First, we consider a situation for which the gas is initially thermally unstable and we study the development of the thermal instability in the presence of Alfvén waves. Although these initial conditions are somewhat artificial, they are simple and close to the assumption of the analytic analysis, making comparisons easier. Second, we setup more realistic initial conditions corresponding to a converging flow of thermally stable WNM. In this case the ram pressure of the flow drives the thermal collapse dynamically. Case of gas initially thermally unstable ---------------------------------------- In order to study the effect of the circularly polarized Alfvén waves on the development of the thermal instability, we start the simulation with thermally unstable gas ($n=3$ cm$^{-3}$ and $T \simeq 500$ K) with density fluctuations of amplitude 0.5. The cooling function is described in Audit & Hennebelle (2005). The computational domain, which initially contains 5000 pixels, has periodic boundary conditions and a length of $3\times 10^{19}$ cm, corresponding to 20 wavelengths of the initial Alfvén wave. Fig. \[hydro\] shows three snapshots of the density field in a purely hydrodynamic run (no MHD wave is present). The initial perturbation grows as a result of the thermal instability. At time $t=$2.55 Myr the density is about three times its initial value. At time $t=$3.64 Myr a cloud of CNM having a size of about 0.3 pc, has formed. Fig. \[b5a1\] displays the density and the y-component of the magnetic field for 3 snapshots showing the development of the thermal instability in the presence of Alfvén waves of amplitude $B _\perp=1 \, \mu$G and for $B_x=5 \, \mu$G. In this situation the value of $\omega_c / \omega _0$ is about 0.03, $A=0.2$ and $\beta=0.17$. The peak density at time $t=1.82$ Myr is about 4.9 whereas the initial peak density is 4.5. In the hydrodynamic case the peak density at the same time is about 6.2 showing that the waves have significantly slowed down the growth of the perturbation, by a factor of about $(6.2-4.5) / (4.9-4.5) \simeq 4$. Note that interestingly enough, this factor is significantly larger than what is predicted by the linear theory which predicts a difference of about 10% (in the case of a perturbation having an initial amplitude of $10^{-2}$ we verified that the growth agrees with the linear theory). After the central density has increased by a factor of about 2 (panel 2), the waves drastically change the structure of the gas and create significant density contrasts ($\simeq$10%). The resulting cloud (panel 3) contains density fluctuations of about $\simeq$ 50% its maximum value, and is therefore very different from the uniform cloud formed in the hydrodynamical case. These large fluctuations are due to magnetic pressure variations. It is worth noting that due to the contraction, the waves inside the growing perturbation (panel 2 and 3) have a larger amplitude and a shorter wavelength than the waves in the surrounding medium. According to the analysis of the preceeding sections, this effect tends to increase the influence of the waves. Fig. \[b5a2.5\] shows results for $B _\perp=2.5 \, \mu$G and for $B_x=5 \, \mu$G. The waves strongly influence the gas evolution. The initial perturbation is totally stabilized by the waves and does not develop (second panel). On the contrary, the waves trigger the formation of structures having a wave number $k \simeq k_0 $ (first and second panels). These structures keep condensing (panel 3) and finally about 12 CNM clouds form. Their size is about 0.03 pc and therefore about 10 times smaller than the size of the cloud that forms in the hydrodynamical case. It is worth noting that the formation of these clouds is significantly (50%-100%) faster than in the two previous cases. Therefore in this range of parameters, the waves destabilize the gas and accelerate the formation of CNM structures. Fig. \[b2.5a2.5\] shows results for $B _\perp=2.5 \, \mu$G and for $B_x=2.5 \, \mu$G. In this situation the value of $\omega_c / \omega _0$ is about 0.06, $A=1$ and $\beta=0.34$. As can be seen on first panel, the initial perturbation does not grow (the peak density is smaller at time $t=1.82$ Myr than the initial value). Unlike the previous case, the Alfvén waves do not efficiently trigger the formation of structures at $k \simeq k_0$ (see first panel). However since the waves are unable to stabilize small wavelengths, the initial perturbation breaks down in several structures which finally develop (panel 2). This nevertheless occurs after a significant delay showing that in this situation, the gas spent a longer amount of time in the thermally unstable state. At time $t=5.41$ Myr about 6 small structures of size $\simeq 0.03$ pc and one larger structure of size $\simeq 0.3$ pc have formed. ![ Spatial zoom showing the thermal condensations which have been induced by the large scale converging flow in the hydrodynamical case. Two snapshots are displayed.[]{data-label="converging_hy"}](bn0_multi2.ps){width="7cm"} (0,20.5) (0,10.5)[![ Same as Fig. \[converging\_hy\] in the presence of circularly polarized Alfvén waves (see text for detail). Four snapshots are displayed.[]{data-label="converging"}](bn5_multi2.ps "fig:"){width="7cm"}]{} (0,0.5) [![ Same as Fig. \[converging\_hy\] in the presence of circularly polarized Alfvén waves (see text for detail). Four snapshots are displayed.[]{data-label="converging"}](bn5_multi1.ps "fig:"){width="7cm"}]{} Case of thermal condensation dynamically induced ------------------------------------------------ We now consider a converging flow of WNM (Hennebelle & Pérault 1999) in a simulation box of length 150 pc. The peak velocities are 2.2 $C_{wnm}$ and -2 $C_{wnm}$ and the peak to peak distance is about 60 pc. Two simulations are performed. The first one is purely hydrodynamical, while the second one starts with an Alfvén waves of amplitude 2.5 $\mu$G with 100 spatial periods in the integration domain. The total magnetic intensity is 5 $\mu$G and the longitudinal one $B_x \simeq 4 \mu$G. An initial resolution of 25,000 pixels is used in order to ensure an accurate description of the Alfvén wave (it corresponds to 250 pixels per period). Fig.\[converging\_hy\] shows the hydrodynamical result. At time $t=6.09$ Myr, two clouds having a size of about 0.1 pc have formed. They present weak density gradients and have no significant internal velocity. The two clouds have a relative velocity of about 0.15 km/s and undergo a collision at time $t=6.3$ Myr. A shock-compressed layer forms with a density of about $\simeq 700$ cm$^{-3}$ and a length $\simeq 0.02$ pc. During the time of the collision, the structure presents a stiff velocity gradient. Fig.\[converging\] displays four snapshots of the longitudinal velocity field, the density and the y-component of the magnetic field in the MHD case. The results displayed in the first panel confirm the trends observed in the numerical experiments of Sect. 3.1. The gas fragments into few small CNM structures having a physical length as small as few 0.01 pc (note that due to the AMR scheme, these structures are well described). This situation is very different from the hydrodynamical case for which the structures are much larger and uniform. A comparison between the 2 times displayed in first panel of Fig.  \[converging\] reveals that all the structures do not form at the same time. The structure of the magnetic field deserves attention. It varies very rapidly in the new born condensations ($x=242$ cm and $x=243$ cm), therefore compressing them. It varies less stiffly near older structures ($x \simeq 244$ cm) since the field lines had time to unbend. Between the clouds, the magnetic field is much more uniform. This is a consequence of the fact that the Alfvén speed is about 10 times larger in the WNM than in the CNM. It is also clear that the intercloud magnetic pressure plays an important role in preventing the merging of the CNM structures therefore maintaining the complexity of the flow and organizing it into groups of structures rather than into a single cloud. In these groups of structures the longitudinal velocity dispersion is not negligible, unlike in the hydrodynamical case. It is about 0.1-0.2 km/s, i.e. few times the sound speed or the Alfvén speed of the CNM. This velocity dispersion is due to the transfer of magnetic energy into longitudinal motions because of the magnetic pressure fluctuations. We stress the fact that observing such group of structures with a low spatial resolution ($\simeq $0.2 pc) may lead to a rather different picture, namely a broad, uniform and turbulent (having a Mach number M$\simeq 1-2$) CNM structure. The second panel of Fig.\[converging\] shows two later snapshots of the same numerical experiment. The group of structures which is seen in the first panel is now located at $x \simeq 251.5$ pc. Large fluctuations ($\simeq 100$%) of density and magnetic field are still present as well as a longitudinal velocity dispersion of about $0.1-0.2$ km/s. Another smaller group of structures ($x \simeq$ 253.5 pc) has formed. At time $t=9.11$ Myr (dotted line), the first cloud undergoes a large density fluctuation ($n _{\rm max} \simeq$ 600 cm$^{-3}$) at a scale of about 0.01 pc. At the same time the cloud is compressed, because of the (magnetic) interaction with the other cloud, so that its length is divided by a factor $\simeq 2$. An interpretation based on the previous analytic results is that the amplitude and wavenumber of the Alfvén wave in the cloud increasing as a result of the compression, the decay instability is triggered leading to a larger density fluctuations. Interestingly, in contrast with the stiff and large fluctuations undergone by $B_y$, the longitudinal velocity field remains relatively smooth. Discussion and conclusion ========================= The analysis presented in Sect.\[sec:num\] as well as the numerical experiments discussed in the previous section show that even modest amplitude Alfvén waves may have a strong impact on the structure of the multiphase ISM. This study focused on circularly polarized parallel propagating Alfvén waves, mainly to allow analytic calculations. These waves, which are exact solutions of the MHD equations, are very weakly dissipative and are therefore very likely to be present in the ISM. Their effects depend on $\beta$, $\widetilde{\omega_c}$ and $A$, respectively the square ratio of the sound to the Alfvén speeds, the ratio between the wave temporal period and the cooling time and the wave amplitude. Depending on the values of these parameters, these waves may: i) stabilize the wavelengths larger than that of the Alfvén wave, $\lambda _{\rm AW}$, that would otherwise be thermally unstable, therefore enhancing the fraction of thermally unstable gas in the ISM , ii) destabilize the wavelengths comparable to $\lambda _{\rm AW}$ and thus fragment the CNM into several spatially correlated small clouds , iii) induce strong density fluctuations within preexisting CNM structures (up to 10 times the mean density), iv) maintain an Alfvénic velocity dispersion within the CNM structures by pumping their energy into longitudinal motions. Finally magnetic pressure tends to prevent the merging between CNM clouds. These effects are not observed in one-dimensional purely hydrodynamical simulations leading to larger and almost uniform structures with weak internal motions. In two dimensions (Koyama & Inutsuka 2002, Audit & Hennebelle 2005) the situation is more complex due to the role of turbulence, with coexistence of thermally unstable gas and small-scale structures, but the latter are still relatively uniform and do not contain significant velocity dispersion. According to the present analysis, MHD waves may enhance the thermal fragmentation found in these simulations, maintain an Alfvénic velocity dispersion and generate large density fluctuations within the CNM structures. Recent observational progress has revealed the presence of interesting features and of a large quantity of thermally unstable gas (e.g. Heiles 2001, Heiles & Troland 2003, Miville-Deschênes et al. 2003). Heiles (1997) summarizes and discusses observations of tiny small-scale structures and more recently Braun & Kanekar (2004, 2005) and Staminorovic & Heiles (2005) report the detection of very low column density CNM clouds. Qualitatively, the present study as well as the 2D simulations of Koyama & Inutsuka (2002) and Audit & Hennebelle (2005) show similar features including thermally unstable gas (also observed by Gazol et al. 2001 in simulations at larger scale), low column density structures (down to and even smaller than $10^{18}$ cm$^{-2}$) and large density fluctuations at small scale (up to 10$^3$ cm$^{-3}$). Small-scale structures and large density fluctuations appear to be a natural outcome of the two-phase nature of the flow. In the absence of turbulence or magnetic field, CNM structures have a typical length of about 0.1 pc (or a column density of $\simeq 10^{19}$ cm$^{-2}$) (Audit & Hennebelle 2005). In the presence of waves or turbulence, a growing structure can however fragment into smaller pieces. Moreover, since both the sound and the Alfvén speed change by a factor $\simeq$ 10 between the WNM and the CNM, with a transition occurring on a short distance of the order of a Field length, supersonic or superalfvénic CNM motions are expected to be more frequent. No systematic and quantitative study attempting to closely compare simulations and observations has been carried out yet. This is clearly a major challenge for the future. This work has received partial financial support from the French national program PCMI. Audit E., Hennebelle P., [ 2005, A&A 1, 433]{} Braun R., Kanekar N., [ 2004 in the IMF at 50 ed. E. Corbelli, F. Palla, & H. Zinnecker (Kluwer), astro-ph/0409427]{} Braun R., Kanekar N., [ 2005 A&A [*in press*]{}, astro-ph/0505055]{} Cho J., Lazarian A., [ 2003 Mon. Not. R. Astron. Soc., 345, 325]{} Cohen R.H., Kulsrud R.M., [ 1974, Phys. Fluid, 17, 2215]{} Dewar R.L., [ 1970, Phys. Fluid, 13, 2710]{} Falle S.A., Hartquist T.W., [ 2002, MNRAS 329, 195]{} Ferraro, V. C. A 1955, Proc. R. Soc. London, Ser. A, 223, 310 Field G., [ 1965, ApJ 142, 531]{} Field G., Goldsmith D., Habing H., [ 1969, ApJ Lett 155, 149]{} Gazol A., Vázquez-Semadeni E., Sánchez-Salcedo F., Scalo J., [ 2001, ApJ 557, L124]{} Goldstein M., [ 1978, ApJ 219, 700]{} Heiles C., [ 1987, [*Interstellar processes*]{}, ed. Hollenbach D., Thronson H. (Reidel)]{} Heiles C., [ 1997, ApJ 481, 193]{} Heiles C., [ 2001, ApJ 551, 105]{} Heiles C., Troland T., [ 2003, ApJ 586, 1067]{} Heiles C., Troland T., [ 2005, ApJ [*in press*]{}, astro-ph/0501482]{} Hennebelle P., Pérault M., [ 1999, A&A 351, 309]{} Hennebelle P., Pérault M., [ 2000, A&A 359, 1124]{} Koyoma H., Inutsuka S., [ 2002, ApJ 564, L97]{} Kulkarni S.R., Heiles C., [ 1987, [*Interstellar processes*]{}, ed. Hollenbach D., Thronson H. (Reidel)]{} Loewenstein M., [ 1990, ApJ 349, 471]{} Lou, Y. Q. 1996, MNRAS, 279, L67 McKee C.F., Zweibel E.G., [ 1995, ApJ 440, 686]{} Meerson B., [ 1996, Rev. Mod. Phys. 68, 215]{} Miville-Deschênes M.-A., Joncas G., Falgarone E., Boulanger F., [ 2003, A&A 411, 109 ]{} Passot T., Vázquez-Semadeni E., Pouquet A., [ 1995, ApJ 441, 702]{} Passot T., Vázquez-Semadeni E., [ 2003, A&A 398, 845]{} Piontek R.A., Ostriker E.C., [ 2004, ApJ 601, 905]{} Piontek R.A., Ostriker E.C., [ 2005, ApJ [*in press*]{}, astro-ph/0504669]{} Press W., Teukolsky S., Vetterling W., Flannery B., [ 1994, Numerical Recipies, Cambridge University Press]{} Rogister A., [ 1971, Phys. Fluids 12, 2733]{} Shu F., [ 1992, [*Gas Dymamics*]{} (University Science Books, Mill Valley CA) ]{} Spangler, S., Fuselier S., Fey A. , Anderson G. [ 1988 J. Geophys. Res., 93(A2), 845-857, 10.1029/88JA01034]{} Stanimirović S., Heiles C., [ 2005, ApJ [*in press*]{}]{} Troland T. & Heiles C. [ 1986, ApJ 301, 339]{} Wolfire M.G., Hollenbach D., McKee C.F., [ 1995, ApJ 443, 152]{} Wolfire M.G., Hollenbach D., McKee C.F., [ 2003, ApJ 587, 278]{}

ArXiv

--- abstract: | Product matrix processes are multi-level point processes formed by the singular values of random matrix products. In this paper we study such processes where the products of up to $m$ complex random matrices are no longer independent, by introducing a coupling term and potentials for each product. We show that such a process still forms a multi-level determinantal point processes, and give formulae for the relevant correlation functions in terms of the corresponding kernels. For a special choice of potential, leading to a Gaussian coupling between the $m$th matrix and the product of all previous $m-1$ matrices, we derive a contour integral representation for the correlation kernels suitable for an asymptotic analysis of large matrix size $n$. Here, the correlations between the first $m-1$ levels equal that of the product of $m-1$ independent matrices, whereas all correlations with the $m$th level are modified. In the hard edge scaling limit at the origin of the spectra of all products we find three different asymptotic regimes. The first regime corresponding to weak coupling agrees with the multi-level process for the product of $m$ independent complex Gaussian matrices for all levels, including the $m$-th. This process was introduced by one of the authors and can be understood as a multi-level extension of the Meijer $G$-kernel introduced by Kuijlaars and Zhang. In the second asymptotic regime at strong coupling the point process on level $m$ collapses onto level $m-1$, thus leading to the process of $m-1$ independent matrices. Finally, in an intermediate regime where the coupling is proportional to $n^{\frac12}$, we obtain a family of parameter dependent kernels, interpolating between the limiting processes in the weak and strong coupling regime. These findings generalise previous results of the authors and their coworkers for $m=2$. address: - 'Faculty of Physics, Bielefeld University, P.O. Box 100131, 33501 Bielefeld, Germany ' - 'Department of Mathematics, The Hebrew University of Jerusalem, Givat Ram, Jerusalem 91904, Israel' author: - Gernot Akemann - Eugene Strahov title: '**[Product matrix processes for coupled multi-matrix models and their hard edge scaling limits]{}**' --- Introduction {#SectionIntroduction} ============ By a random *multi-matrix model* one usually means a probability measure defined on a space formed by a collection of matrices of the same type. Probably the best known example is that of Hermitian matrices coupled in a chain considered by Eynard and Mehta [@EynardMehta]. This example represents the simplest case of multi-matrix models known from applications in quantum field theory. Another class of examples consists of multi-matrix models of positive Hermitian matrices subject to the Cauchy interaction. This class of multi-matrix models was introduced in Bertola, Gekhtman, and Szmigielski [@BertolaGekhtmanSzmigielski], and studied further in Bertola, Gekhtman, and Szmigielski [@BertolaGekhtmanSzmigielski1; @BertolaGekhtmanSzmigielski2], and in Bertola and Bothner [@BertolaBothner]. For other examples of multi-matrix models, and for an explanation of their relevance to quantum field theory and to statistical mechanics we refer the reader to Eynard, Kimura, and Ribault [@EynardKimuraRibault Section 2.2.], Filev and O’Connor [@FilevO'Connor], Bertola, Eynard, Harnad [@BertolaEynardHarnad], and references therein. The starting point of the present work is the observation that the problem about the distribution of singular values for a product of independent complex Gaussian matrices can be reformulated in terms of a multi-matrix model. This multi-matrix model can be defined by the probability measure $$\label{MultiMatrixGinibreProbabilityMeasure} \begin{split} \frac{1}{Z_n}&e^{-\sum\limits_{l=1}^m{\mathop{\mathrm{Tr}}}\left[G_l^*G_l\right]}\prod\limits_{l=1}^mdG_l. \end{split}$$ over the set of rectangular complex matrices $\left(G_1,\ldots,G_m\right)$, where $G_l$ is of size $\left(n+\nu_l\right)\times \left(n+\nu_{l-1}\right)$, $\nu_0=0$, $\nu_1\geq 0$, $\ldots$, $\nu_{m-1}\geq 0$, $dG_l$ is the corresponding flat complex Lebesgue measure, and $Z_n$ is a normalisation constant. Given this probability measure one can ask about the distribution of complex eigenvalues and of squared singular values of the total product matrix $Y_m=G_m\cdots G_1$. It turns out that the eigenvalues of $Y_m$ form a determinantal point process on the complex plane which can be understood as a generalisation of the classical Ginibre ensemble. This fact was first proved in Akemann and Burda [@Akemann1], see Adhikari, Kishore Reddy, Ram Reddy, and Saha [@Adhikari], Forrester [@Forrester], Ipsen [@Ipsen], Akemann, Ipsen, and Strahov [@AkemannIpsenStrahov], Forrester and Ipsen [@ForresterIpsen] for different proofs and extensions of this result. Moreover, it was shown in Akemann, Kieburg, and Wei [@AkemannKieburgWei], and in Akemann, Ipsen, and Kieburg [@AkemannIpsenKieburg] that the joint probability density function of the squared singular values of $Y_m$ forms a determinantal point process, representing a special polynomial ensemble, a notion introduced later by Kuijlaars and Stivigny [@ArnoDries]. A contour integral representation for the correlation kernel of this ensemble was derived in Kuijlaars and Zhang [@KuijlaarsZhang], which enabled a detailed analysis of different scaling limits, see Kuijlaars and Zhang [@KuijlaarsZhang], Liu, Wang, and Zhang [@LiuWangZhang]. A natural (multi-level) generalisation of the point process formed by the squared singular values of the total product matrix $Y_m$ can be constructed as follows. For each $l\in\left\{1,\ldots,m\right\}$, denote by $y_j^l$, $j=1,\ldots,n$, the squared singular values of the partial product $Y_l=G_l\cdots G_1$, that is the eigenvalues of $Y_l^*Y_l$. The configuration $$\label{configuration} \left\{\left(l,y_j^l\right)\vert l=1,\ldots,m;j=1,\ldots,n\right\}$$ of all eigenvalues forms a point process on $\left\{1,\ldots,m\right\}\times{\mathbb R}_{>0}$. This process was introduced and studied in Strahov [@StrahovD], calling it the *Ginibre product process*. It was shown in [@StrahovD] that it is a multi-level determinantal point process. Furthermore, a contour integral representation for the correlation kernel of this process was derived, and its hard edge scaling limit computed. Here, we will drop the assumption that the matrices $G_1$, $\ldots$, $G_m$ are independent as in (\[MultiMatrixGinibreProbabilityMeasure\]), but assume instead that these or their products becomes coupled. We can still construct and investigate the multi-level point process formed by configurations (\[configuration\]) and refer to it as the *product matrix process* associated with $G_1$, $\ldots$, $G_m$. It is an interesting general problem to describe statistical properties of such product matrix processes, their relevant correlation functions, and their scaling limits. Of course it is desirable that it still forms a multi-level determinantal point process, with explicit formulae for the correlation kernel and thus the correlation functions. An example for such a setup is the multi-matrix model for Hurwitz numbers [@AmbjornChekov]. In this paper we introduce and study a different multi-matrix model of statistically dependent random matrices, also satisfying these requirements. We show (see Theorem \[TheoremMostGerneralCorrelationKernel\]) that these product matrix processes generalise the Ginibre product process studied in Strahov [@StrahovD]. For a particular choice of potentials the first $m-1$ levels remain that of products of $m-1$ independent random matrices, to which the $m$-th level is coupled. For this example we derive contour integral representations for the relevant correlation kernels, see Theorem \[TheoremDoubleIntegralRepresentationExactKernel\]. In the hard edge scaling limit at large matrix sizes we distinguish three different asymptotic regimes. In the first regime corresponding to weak coupling, the limiting product matrix process is the same as for the product of $m$ independent complex Gaussian matrices, as described in Strahov [@StrahovD]. This process can be called the $m$-level Meijer $G$-kernel process, a multi-level extension of the Meijer $G$-kernel process introduced by Kuijlaars and Zhang [@KuijlaarsZhang]. In the second asymptotic regime at strong coupling the $m$-th level collapses to the point process on the $(m-1)$-th level. In other words, the correlations between the $m$-th level and the levels up to $m-2$ are the same as the respective correlations between the $(m-1)$-th and the levels up to $m-2$. The correlations between level $m$ and $m-1$ are as if they are on the same level, except at colliding points where we find a contact interaction in form of a Dirac delta function. Finally, at intermediate coupling we obtain a limiting process of $m$ levels, *interpolating* between that in the weak coupling regime and that in the strong coupling regime. All three hard edge scaling limits described above are given by Theorem \[TheoremHardEdgeGinibreCouplingProcess\] that we consider as the main achievement of the present work. We note that interpolating ensembles are of great interest in Random Matrix Theory. In the literature one can find examples interpolating between Gaussian ensembles of different symmetry classes - one classical example is that of Pandey and Mehta [@PandeyMehta] that interpolates between the Gaussian Orthogonal and Gaussian Unitary Ensemble. Determinantal processes whose edge behaviour interpolates between the Poisson process and the Airy process were considered in Moshe, Neuberger, and Shapiro [@Moshe], and studied further by Johansson [@JohanssonGumbel]. In the context of products of random matrices, a determinantal process with $m=2$ interpolating between the classical Bessel-kernel process (at $m=1$), and the Meijer $G$-kernel process for the product of $m=2$ independent Gaussian matrices was obtained by the authors in [@AkemannStrahov; @AkemannStrahov1], and further extended most recently by Liu [@DZLm2] and a joint work [@ACLS]. However, we are not aware of further examples of multi-level interpolating ensembles, with properties described by Theorem \[TheoremInterpolation\]. In particular, the results mentioned above lead to three different scaling limits for the biorthogonal ensemble formed by the squared singular values of the total product matrix $Y_m$: that of the Meijer $G$-kernel process for the product of $m$ ($m-1$) independent rectangular matrices with complex Gaussian entries at weak (strong) coupling, and a determinantal process interpolating between these correlation kernels, see Corollary \[CorollaryTheoremHardEdgeGinibreCouplingProcess\]. This paper is organised as follows. In Section \[SectionExactResults\] we present exact results at finite matrix sizes. In particular, in Section \[SectionExactResults\] we define a family of multi-matrix models, and introduce the product matrix process associated with this family. We show that the product matrix process under considerations is a multi-level determinantal process, and give formulae for the correlation kernels. In Section \[SectionHardEdgeScalingLimits\] we compute the hard edge scaling limits in our example, corresponding to different asymptotic regimes, and in Section \[SectionInterpolatingProcess\] we describe the properties of the obtained interpolating multi-level determinantal process. Sections \[SectionIntegrationFormula\]-\[SectionProofTheoremInterpolation\] contain the proofs our statements.\ **Acknowledgements.** We are very grateful to Marco Bertola, Tomasz Checinski and Mario Kieburg for discussions and useful comments. The anonymous referee is also thanked for several comments and corrections. This work was supported by the DFG through grant AK35/2-1 and CRC 1283 “Taming uncertainty and profiting from randomness and low regularity in analysis, stochastics and their applications” (G.A.). Exact results for general coupling and finite matrix size {#SectionExactResults} ========================================================= Coupled multi-matrix models with general potential functions {#SectionExactResults1} ------------------------------------------------------------ Fix $a>0$, $b>0$, and consider a multi-matrix model defined by the probability distribution $$\label{MainProbabilityMeasure} \begin{split} &\frac{1}{Z_n} \exp\left[-a\sum\limits_{l=1}^m{\mathop{\mathrm{Tr}}}\left[G_l^*G_l\right]+b{\mathop{\mathrm{Tr}}}\left[G_m\cdots G_1+\left(G_m\cdots G_1\right)^*\right] \right] \\ &\times \exp\left[ -\sum\limits_{l=1}^m{\mathop{\mathrm{Tr}}}\left[V_l\left(\left(G_{l}\cdots G_1\right)^*\left(G_{l}\cdots G_1\right)\right)\right]\right] \prod\limits_{l=1}^mdG_l, \end{split}$$ normalised by the constant $Z_n$. It depends on $m$ rectangular complex matrices $G_l$ is of size $\left(n+\nu_l\right)\times \left(n+\nu_{l-1}\right)$ for $l=1,\ldots,m$, with $\nu_0=0$, $\nu_m=0$, $\nu_1\geq 0$, $\ldots$, $\nu_{m-1}\geq 0$, and where $dG_l$ is the corresponding flat complex Lebesgue measure. Here, $V_l$ (the potentials) are some scalar positive functions which are continuous and grow fast enough at infinity to guarantee the convergence of the corresponding matrix measure (\[MainProbabilityMeasure\]). Note that the potentials and in particular the parameter $b$ introduce a *coupling* amongst the matrices: if $b=0$, and $V_l=0$ $\forall l=1,\ldots,m$, then the matrices $G_1$, $\ldots$, $G_m$ become independent Gaussian random matrices. The multi-level point process associated with the multi-matrix model (\[MainProbabilityMeasure\]) of coupled random matrices can be constructed as that for the Ginibre product process [@StrahovD] of independent matrices. Namely denote by $\{y_j^l\}_{j=1,\ldots,n}$ the set of squared singular values of matrix $Y_l=G_l\cdots G_1$, these are the eigenvalues of $Y_l^*Y_l$. The configuration of all these eigenvalues forms a point process on $\left\{1,\ldots,m\right\}\times{\mathbb R}_{>0}$. We will call this point process the *product matrix process* corresponding to the multi-matrix model (\[MainProbabilityMeasure\]). Let us indicate some particular cases of the multi-matrix model (\[MainProbabilityMeasure\]) studied previously. At zero coupling $b=0$, and $V_1=\ldots=V_m=0$, we are dealing with $m$ independent complex Ginibre matrices, and the product matrix process turns into the Ginibre product process studied in [@StrahovD] (by rescaling all matrices $G_l$ we can set $a=1$). A second example with $b=0$, $V_1=\ldots=V_{m-1}=0$ and non-vanishing potential $V_m$ was considered in [@AmbjornChekov] in order to study so-called hypergeometric Hurwitz numbers. As a further feature a non-trivial covariance matrix $\Sigma^{-1}$ is introduced there in the measure for the first matrix, replacing ${\mathop{\mathrm{Tr}}}[G_1^*G_1]$ by ${\mathop{\mathrm{Tr}}}[\Sigma^{-1}G_1^*G_1]$. While in [@AmbjornChekov] loop equations were applied we will indicate below that this model defines a multi-level point process. For $m=2$ matrices with non vanishing coupling $b\neq0$, with both potential functions equal to zero, $V_1=V_2=0$, and parameters given by $a=\frac{1+\mu}{2\mu}$, $b=\frac{1-\mu}{2\mu}$, we obtain a model of two coupled matrices introduced by Osborn [@Osborn] in the context of QCD with a baryon chemical potential $\mu$. For this model, the distribution of squared singular values of the product matrix $Y_2$ was studied by the authors in [@AkemannStrahov; @AkemannStrahov1]. By its construction in [@Osborn], the two coupled matrices can be defined in terms of linear combinations of independent Gaussian matrices, too. Namely, let $A$, $B$ be two independent matrices of sizes $n\times L$, with i.i.d. standard complex Gaussian entries. Define random matrices $G_1$ and $G_2$ by $$G_1=\frac{1}{\sqrt{2}}\left(A-i\sqrt{\mu}B\right),\;\; G_2=\frac{1}{\sqrt{2}}\left(A^*-i\sqrt{\mu}B^*\right). \label{m=2-model}$$ Assume that $L\geq n$. Then the joint distribution of $G_1$ and $G_2$ is given by probability measure (\[MainProbabilityMeasure\]) with $a=\frac{1+\mu}{2\mu}$ and $b=\frac{1-\mu}{2\mu}$ for $0<\mu<1$, $m=2$, $V_1=V_2=0$, and $\nu_1=L-n$. The same model with $m=2$ and $V_1=V_2=0$ was generalised in [@DZLm2; @ACLS] to include non-trivial covariance matrices instead of the scalar parameters $a$ and $b$. Here, again only the squared singular values of the product matrix $Y_2$ were studied. Without loss of generality we can always assume that the parameter $a$ in formula (\[MainProbabilityMeasure\]) can be set to $a=1$. Indeed, this can be achieved by defining new matrices $\widetilde{G_1}$, $\ldots$, $\widetilde{G_m}$ given by $$\widetilde{G_l}=\sqrt{a}G_l,\;\; 1\leq l\leq m,$$ together with a replacement of the parameter $b$ by ${b}\,{a^{-\frac{m}{2}}}$, and a modification of the potential functions accordingly. From now on, we restrict our considerations to the probability measures defined by expression (\[MainProbabilityMeasure\]) with $a=1$. Density of product matrix processes and related biorthogonal ensembles ---------------------------------------------------------------------- By definition, the density of the product matrix process corresponding to the multi-matrix model (\[MainProbabilityMeasure\]) is that of the vector $$\underline{y}=\left(y^m,\ldots,y^1\right)\in\left({\mathbb R}_{>0}^n\right)^m,$$ where $y^l$ is the vector of squared singular values of $G_l\cdots G_1$. Our first results gives the joint probability density of $\underline{y}$ explicitly. \[TheoremGinibreCouplingDensity\] Consider a multi-matrix model defined by the probability measure (\[MainProbabilityMeasure\]), with $a=1$. Denote by $\{y_j^l\}_{j=1,\ldots,n}$ the set of squared singular values of the matrix $Y_l=G_l\cdots G_1$ for each $l=1,\ldots,m$. Thus for each $l$ the vector $y^l=\left(y^l_1,\ldots, y^l_n\right)$ contains the eigenvalues of the matrix $Y_l^* Y_l$. The joint probability density of $\underline{y}=\left(y^m,\ldots,y^1\right)$ reads $$\label{CouplingDensityFormula} \begin{split} P_{n,m}(\underline{y})=&\frac{1}{Z_{n,m}}\det\left[\left(y_j^m\right)^{\frac{k-1}{2}}I_{k-1}\left(2b\left(y_j^m\right)^\frac{1}{2}\right)e^{-V_m\left(y_j^m\right)}\right]_{j,k=1}^n\\ &\times \prod\limits_{l=1}^{m-1}\det\left[\frac{\left(y_j^{l+1}\right)^{\nu_{l+1}}}{\left(y_k^l\right)^{\nu_{l+1}+1}}e^{-\frac{y_j^{l+1}}{y_k^l} -V_{l}\left(y_k^{l}\right)}\right]_{j,k=1}^n \det\left[(y_j^1)^{\nu_1+k-1}e^{-y_j^1}\right]_{j,k=1}^n, \end{split}$$ where $$\label{Znmgen} Z_{n,m}=\left(n!\right)^m\det\left[a_{i,j}\right]_{i,j=1}^n,$$ the matrix elements $a_{i,j}$ are given by $$\label{GeneralMatrixElements} \begin{split} &a_{i,j}=\int\limits_0^{\infty}\ldots\int\limits_0^{\infty} t_1^{\nu_1+i-1}\left(\frac{t_2}{t_1}\right)^{\nu_2}\ldots\left(\frac{t_{m-1}}{t_{m-2}}\right)^{\nu_{m-1}} \left(\frac{t_m}{t_{m-1}}\right)^{\nu_m} \left(t_m\right)^{\frac{j-1}{2}}I_{j-1}\left(2b\left(t_m\right)^{\frac{1}{2}}\right) \\ &\times e^{-t_1-\frac{t_2}{t_1}-\ldots-\frac{t_{m-1}}{t_{m-2}}-\frac{t_m}{t_{m-1}} -V_1\left(t_1\right)-V_2\left(t_2\right)-\ldots-V_{m-1}\left(t_{m-1}\right)-V_{m}\left(t_{m}\right)} \frac{dt_1}{t_1}\frac{dt_2}{t_2}\ldots \frac{dt_{m-1}}{t_{m-1}} dt_m, \end{split}$$ and where $I_{\mu}(z)$ denotes the modified Bessel function of the first kind. Recall that the modified Bessel function of the first kind $I_{\mu}(z)$ is defined by $$I_{\mu}(z)=\sum\limits_{m=0}^{\infty}\frac{1}{m!\Gamma(\mu+m+1)}\left(\frac{z}{2}\right)^{2m+\mu}, \label{I-def}$$ and that the function $y^{\frac{\mu}{2}}I_{\mu}\left(2by^{\frac{1}{2}}\right)$ has the following asymptotic behaviour $$\label{ISmallb} y^{\frac{\mu}{2}}I_{\mu}\left(2by^{\frac{1}{2}}\right)\sim\frac{\left(by\right)^{\mu}}{\Gamma(\mu+1)},\;\; b\rightarrow 0.$$ From (\[CouplingDensityFormula\]) and (\[ISmallb\]) we conclude that as $V_1=\ldots=V_m=0$, and as $b$ approaches zero, the joint probability density $P_{n,m}(\underline{y})$ becomes equal to $$\label{GinibreDensityFormula} \frac{1}{Z_{n,m}^{\operatorname{Ginibre}}} \Delta_n(\{y_j^m\}) \prod\limits_{l=1}^{m-1} \det\left[\frac{\left(y_j^{l+1}\right)^{\nu_{l+1}}}{\left(y_k^l\right)^{\nu_{l+1}+1}}e^{-\frac{y_j^{l+1}}{y_k^l}}\right]_{j,k=1}^n \det\left[(y_j^1)^{\nu_1+k-1}e^{-y_j^1}\right]_{j,k=1}^n,$$ where $$\label{Znm} Z_{n,m}^{\operatorname{Ginibre}}=\left(n!\right)^m\prod\limits_{j=1}^n\prod\limits_{l=0}^m\Gamma\left(j+\nu_l\right),$$ and we have defined the Vandermonde determinant $$\label{Vandermonde} \Delta_n(\{ y_j^l\})=\det\left[\left(y_j^l\right)^{k-1}\right]_{j,k=1}^n=\prod_{1\leq j<k\leq n}(y_k^l-y_j^l)\ .$$ This is the density for the Ginibre product process, cf. Strahov [@StrahovD Proposition 4.2] (see also [@AkemannIpsenKieburg]). Note that the last determinant in is proportional to the Vandermonde determinant $\Delta_n(\{y_j^1\})$. The vector $y^m=\left(y_1^m,\ldots,y^m_n\right)$ is that of the squared singular values of the total product matrix $Y_m=G_m\cdots G_1$. In view of the recent results on distributions of squared singular values for products of random matrices mentioned in the introduction, the probability distribution of $y^m$ is of special interest. Corollary \[Corrolary1\] from Theorem \[TheoremGinibreCouplingDensity\] gives the joint probability distribution of $y^m$ explicitly. \[Corrolary1\] Consider the multi-matrix model defined by the probability distribution (\[MainProbabilityMeasure\]), and form the total product $Y_m=G_m\cdots G_1$. The squared singular values of $Y_m$ form a biorthogonal ensemble with the joint probability density function given by $$P_{n,m}\left(y_1^m,\ldots,y_n^m\right)= \frac{1}{Z_{n,m}'}\det\left[\varphi_i\left(y_j^m\right)\right]_{i,j=1}^n \det\left[\psi_i\left(y_j^m\right)\right]_{i,j=1}^n,$$ with the normalisation $Z_{n,m}'=n!\det\left[a_{i,j}\right]_{i,j=1}^n$ resulting from . The functions $\varphi_i$, with $i=1,\ldots, n$ are given by $$\varphi_i\left(y\right)=y^{\frac{i-1}{2}}I_{i-1}\left(2by^{\frac{1}{2}}\right)e^{-V_m(y)},$$ and the functions $\psi_i$ with $=1,\ldots,n$ are given by $$\begin{split} \psi_i(y)=&\int\limits_0^{\infty}\ldots\int\limits_0^{\infty} t_1^{\nu_1+i-1}\left(\frac{t_2}{t_1}\right)^{\nu_2}\ldots\left(\frac{t_{m-1}}{t_{m-2}}\right)^{\nu_{m-1}} \left(\frac{y}{t_{m-1}}\right)^{\nu_m} \\ &\times e^{-t_1-\frac{t_2}{t_1}-\ldots-\frac{t_{m-1}}{t_{m-2}}-\frac{y}{t_{m-1}} -V_1\left(t_1\right)-V_2\left(t_2\right)-\ldots-V_{m-1}\left(t_{m-1}\right)} \frac{dt_1}{t_1}\frac{dt_2}{t_2}\ldots \frac{dt_{m-1}}{t_{m-1}}. \end{split}$$ For some particular choices of potentials $V_1$, $\ldots$, $V_m$ we can obtain explicit expressions for the functions $\psi_i(y)$, and for the normalisation constant $Z_{n,m}'$, as we will show below. First, let us suppose that we could set all potentials $V_1=\ldots=V_m=0$. Then the functions $\psi_i(y)$ could be written as Meijer $G$-functions[^1], $$\psi_i(y)=G^{m,0}_{0,m}\left(\begin{array}{cccc} & - & & \\ \nu_1+i-1, & \nu_2, & \ldots, & \nu_m \end{array} \biggl|y\right).$$ For the resulting normalisation we would obtain from Andréief’s integral identity $$\label{NormalizationConstant1} Z_{n,m}'=n!\det\left[\int\limits_0^{\infty}y^{\frac{i-1}{2}} I_{i-1}\left(2by^{\frac{1}{2}}\right) G^{m,0}_{0,m}\left(\begin{array}{cccc} & - & & \\ \nu_1+j-1, & \nu_2, & \ldots, & \nu_m \end{array} \biggl|y\right) dy\right]_{i,j=1}^{n}.$$ However, for $m>2$ the integrals inside the determinant do not converge. The reason is that the modified Bessel function of the second kind has the following asymptotic $$\label{A1} I_{i-1}\left(2by^{\frac{1}{2}}\right)\sim\frac{e^{2by^{\frac{1}{2}}}}{2\pi^{\frac{1}{2}}b^{\frac{1}{2}}y^{\frac{1}{4}}},\;\; y\rightarrow\infty,$$ c.f. [@NIST], which has to be compared to the asymptotic of the Meijer $G$-function [@Luke] $$G^{m,0}_{0,m}\left(\begin{array}{cccc} & - & & \\ \nu_1+j-1, & \nu_2, & \ldots, & \nu_m \end{array} \biggl|y\right)\sim \frac{(2\pi)^{\frac{m-1}{2}}}{m^{\frac12}} y^\theta \exp[-my^{\frac{1}{m}}] ,\;\; y\rightarrow\infty, \label{Gasympt}$$ with $\theta=\frac{1}{m}(\frac12(1-m)+\nu_1+j-1+\sum_{l=2}^m\nu_l)$. Only for $m=2$ the two exponentials in eqs. and together lead to convergent integrals for $b<1$, and we obtain the biorthogonal ensemble equivalent to that studied by the authors in [@AkemannStrahov; @AkemannStrahov1][^2]. Indeed for $m=2$ we have (recall that $\nu_2=0$ then) $$G^{2,0}_{0,2}\left(\begin{array}{cc} & - \\ \nu_1+j-1, & 0 \end{array} \biggl|y\right)=2y^{\frac{\nu_1+j-1}{2}}K_{\nu_1+j-1}\left(2y^{\frac{1}{2}}\right),$$ where $K_{\kappa}(z)$ denotes the modified Bessel function of the second kind. It can be defined by the integral $$\label{K-def1} K_{\kappa}(z)=\frac{\Gamma\left(\kappa+\frac{1}{2}\right)(2z)^{\kappa}}{\sqrt{\pi}} \int\limits_0^{\infty}\frac{\cos(t)dt}{\left(t^2+z^2\right)^{\kappa+\frac{1}{2}}}\,,\;\;\Re(\kappa)>-\frac12\,,$$ which is more convenient for complex contour integrals over $\kappa$, or alternatively [@NIST] as $$\label{K-def2} K_\kappa(2z)=\frac12 z^\kappa \int_0^\infty dt\,t^{-\kappa-1} \exp\left[-t-\frac{z^2}{t}\right]\, .$$ From the latter is is easy to obtain the following asymptotic expansion [@NIST 10.25.3], corresponding to for $m=2$: $$\label{A2} K_{\kappa}\left(2by^{\frac{1}{2}}\right)\sim\frac{\pi^{\frac{1}{2}}}{2^{\frac{1}{2}}\left(2by^{\frac{1}{2}}\right)^{\frac{1}{2}}}e^{-2by^{\frac{1}{2}}}, \;\; y\rightarrow\infty.$$ It is not difficult to see that the model for $m=2$ studied in [@AkemannStrahov] in the parametrisation given in can be mapped to our parametrisation , with $a=1$ and $b>0$ by rescaling in $G_1\to\alpha_1G_1$ and $G_2\to\alpha_2 G_2$, with $\alpha_1^2=\frac{2}{1+\mu}$ and $\alpha_2^2=\frac{1+\mu}{2\mu}$, leading to $b=\frac{1-\mu}{\sqrt{4\mu}}$. Therefore, in order to guarantee the existence of the corresponding matrix measure (\[MainProbabilityMeasure\]), for $m>2$ some non-zero potentials $V_l$ should be added for non-zero coupling $b>0$, to ensure convergence of the integrals. In this paper the case corresponding to $$\label{V-cond} V_1(t)=\ldots=V_{m-2}(t)=0\ , \ \ V_{m-1}(t)=b^2t\ , \ \ V_m(t)=0\ ,$$ will be considered in detail[^3]. Here, we are dealing with coupled matrices with the single coupling constant $b>0$. In this case, a simple contour integral representation for the correlation kernel suitable for an asymptotic analysis can be derived. Inserting the above conditions on the potentials into the matrix measure (\[MainProbabilityMeasure\]), it turns into $$\label{MainProbabilityMeasure1} \begin{split} &\frac{1}{Z_n(b)}e^{-\sum\limits_{l=1}^m{\mathop{\mathrm{Tr}}}\left[G_l^*G_l\right]+b{\mathop{\mathrm{Tr}}}\left[G_m\cdots G_1+\left(G_m\cdots G_1\right)^*\right] -b^2{\mathop{\mathrm{Tr}}}\left[\left(G_{m-1}\cdots G_1\right)^*\left(G_{m-1}\cdots G_1\right)\right]}\prod\limits_{l=1}^mdG_l\\ =& \frac{1}{Z_n(b)}e^{-\sum\limits_{l=1}^{m-1}{\mathop{\mathrm{Tr}}}\left[G_l^*G_l\right] -{\mathop{\mathrm{Tr}}}\left[\left(G_m-b(G_{m-1}\cdots G_1)^*\right) \left(G_m^*-bG_{m-1}\cdots G_1\right) \right]}\prod\limits_{l=1}^mdG_l\,. \end{split}$$ The normalised probability density of $\underline{y}=\left(y^m,\ldots,y^1\right)$ that follows from Theorem \[TheoremGinibreCouplingDensity\] by inserting into is given by $$\label{CouplingDensityFormula1} \begin{split} P_{n,m}(\underline{y};b)= &\frac{1}{Z_{n,m}(b)}\det\left[\left(y_j^m\right)^{\frac{k-1}{2}}I_{k-1}\left(2b\left(y_j^m\right)^\frac{1}{2}\right)\right]_{j,k=1}^n \det\left[\frac{1}{y_k^{m-1}}e^{-\frac{y_j^{m}}{y_k^{m-1}}-b^2y_k^{m-1}}\right]_{j,k=1}^n\\ &\times \prod\limits_{l=1}^{m-2}\det\left[\frac{\left(y_j^{l+1}\right)^{\nu_{l+1}}}{\left(y_k^l\right)^{\nu_{l+1}+1}}e^{-\frac{y_j^{l+1}}{y_k^l}}\right]_{j,k=1}^n \det\left[(y_j^1)^{\nu_1+k-1}e^{-y_j^1}\right]_{j,k=1}^n, \end{split}$$ where the normalisation constant $Z_{n,m}(b)$ following from remains to be determined, see below. In what follows we will refer to the point process formed by random configuration $\underline{y}$ with the joint probability density given by equation (\[CouplingDensityFormula1\]) as the *Ginibre product process with coupling*. The Ginibre product process with coupling is an one-parameter *deformation* of the Ginibre product process, and the coupling constant $b$ is the parameter of this deformation. The model defined by equation has the following interpretation. According to equation the matrices $G_1$, $\ldots$, $G_m$ and $G_m'=G_m-b\left(G_{m-1}\ldots G_1\right)^*$ can be understood as independent Ginibre matrices (each of the size $\left(n+\nu_l\right)\times\left(n+\nu_{l-1}\right)$, where $l=1,\ldots,m$). Therefore, we will have the Ginibre product process on the first $m-1$ levels which will be independent of $b$. As $b\rightarrow 0$, the matrix $G_m-b\left(G_{m-1}\ldots G_1\right)^*$ turns into the Ginibre matrix $G_m$ independent of $G_1$, $\ldots$, $G_{m-1}$, and we obtain the Ginibre product process on all $m$ levels. As $b\rightarrow\infty$, the matrix $G_m'$ is dominated by $\left(G_{m-1}\ldots G_1\right)^*$, so the singular values of $G_m'G_{m-1}\ldots G_1$ will approach the squares of the singular values of $G_{m-1}\ldots G_1$. If in the second line of we rescale $G_m\to b G_m$ and then send $b\to\infty$ this leads to a delta function constraint, due to $\lim_{b\to\infty}b\exp[-b^2(x-y)^2]/\sqrt{\pi}=\delta(x-y)$, enforcing $G_m=(G_{m-1}\cdots G_1)^*$ at finite matrix sizes. It is precisely this limit that leads to the collapse of the Ginibre product process with coupling having $m$ levels to a Ginibre product process having only $m-1$ levels, as we will see below on the level of the joint probability density for finite $n$, and in the large-$n$ limit in Theorem \[TheoremHardEdgeGinibreCouplingProcess\] (C). It is instructive to consider separately the distribution of the squared singular values of the total product matrix $Y_m=G_m\cdots G_1$, in the case when the joint distribution of $G_m$, $\ldots$, $G_1$ is given by formula (\[MainProbabilityMeasure1\]), with joint probability density . We obtain the following result. \[PropositionJointDensityTotalProductMatrix\] Consider the probability measure defined by equation (\[MainProbabilityMeasure1\]), and let $Y_m=G_m\cdots G_1$ be the total product matrix of the corresponding multi-matrix model. The squared singular values $y_1^m,\ldots,y_n^m$ of $Y_m$ form a determinantal point process on ${\mathbb R}_{>0}$. This determinantal point process is a biorthogonal ensemble defined by the joint probability density $P\left(y_1^m,\ldots,y_n^m;b\right)$. The explicit formula for $P\left(y_1^m,\ldots,y_n^m;b\right)$ reads $$\label{PExact} P\left(y_1^m,\ldots,y_n^m;b\right)=\frac{(n!)^{m-1}}{Z_{n,m}(b)}\det\left[\left(y_j^m\right)^{\frac{k-1}{2}}I_{k-1}\left(2b\left(y_j^m\right)^{\frac{1}{2}}\right)\right]_{k,j=1}^n \det\left[\psi_k(y_j^m)\right]_{k,j=1}^n,$$ where $$\label{PSIK} \begin{split} &\psi_k\left(y\right)=\frac{1}{2\pi i}\int\limits_{c-i\infty}^{c+i\infty}\prod\limits_{l=2}^m\Gamma\left(u+\nu_l\right) \Gamma\left(u+\nu_1+k-1\right)\frac{2\left(by^{\frac{1}{2}}\right)^uK_u\left(2by^{\frac{1}{2}}\right)}{\Gamma(u)}y^{-u}du, \end{split}$$ with $c>0$ such that the poles of the Gamma-functions are to the right of the contour. The normalisation constant is given by $$\label{Znmb} Z_{n,m}(b)=\left(n!\right)^mb^{\frac{n(n-1)}{2}}\prod\limits_{j=1}^n\prod\limits_{l=1}^{m-1}\Gamma\left(j+\nu_l\right).$$ Note that following [@NIST] we have $$\label{KSmallb} \frac{2\left(by^{\frac{1}{2}}\right)^uK_u\left(2by^{\frac{1}{2}}\right)}{\Gamma(u)}\sim1,\;\; b\rightarrow 0,\;\; \mbox{for}\;\; \Re(u)>0.$$ Taking this into account, and using the asymptotic expression $y^{\frac{j-1}{2}}I_{j-1}\left(2by^{\frac{1}{2}}\right)$ (see equation (\[ISmallb\]), contributing with an extra factor $1/\Gamma(j)$), we find that the joint probability density function $P\left(y_1,\ldots,y_n;b\right)$ defined by equation (\[PExact\]) has a limit as $b\rightarrow 0$. Namely, we have $$\begin{split} \underset{b\rightarrow 0}{\lim}&P\left(y_1^m,\ldots,y_n^m;b\right)= \frac{1}{n!\prod\limits_{i=1}^n\prod\limits_{j=0}^m\Gamma\left(j+\nu_j\right)}\\ &\times\Delta_n(\{y_j^m\}) \det\left[G^{m,0}_{0,m}\left( \begin{array}{cccc} & - & & \\ \nu_1+k-1, & \nu_2, & \ldots, & \nu_m \end{array} \biggl|y_j^m\right)\right]_{j,k=1}^n. \end{split}$$ This is the joint probability density function obtained by Akemann, Ipsen, and Kieburg [@AkemannIpsenKieburg] for the squared singular values of the ensemble of the products of $m$ independent rectangular matrices with independent complex Gaussian entries. Now let us consider the behaviour of $P\left(y_1^m,\ldots,y_n^m;b\right)$ as $b\rightarrow\infty$. Using the asymptotic formulae and of the modified Bessel functions involved, we find that as $b\rightarrow\infty$, the joint probability density function $P\left(y_1^m,\ldots,y_n^m;b\right)$ defined by equation (\[PExact\]) becomes approximately equal to $$\begin{split} &\sim\frac{1}{2^nb^{\frac{n(n+1)}{2}}n! \prod\limits_{i=1}^n\prod\limits_{j=0}^{m-1}\Gamma\left(j+\nu_j\right) \prod\limits_{j=1}^ny_j^{\frac{1}{2}}}\\ &\times\Delta_n(\{y_j^m\}) \det\left[ G^{m-1,0}_{0,m-1}\left( \begin{array}{cccc} & - & & \\ \nu_1+k-1, & \nu_2, & \ldots, & \nu_{m-1} \end{array} \biggl|\frac{1}{b}{\left(y_j^m\right)^{\frac{1}{2}}}\right)\right]_{k,j=1}^n. \end{split} \nonumber$$ Set $x_j=\frac{1}{b}(y_j^m)^{\frac{1}{2}}$. We then find that as $b\rightarrow\infty$, the joint probability density function of the new variables $x_1$, $\ldots$, $x_n$ converges to $$\frac{1}{n!\prod\limits_{i=1}^n\prod\limits_{j=0}^{m-1}\Gamma\left(j+\nu_j\right)}\\ \Delta_n(\{x_j\}) \det\left[G^{m-1,0}_{0,m-1}\left( \begin{array}{cccc} & - & & \\ \nu_1+k-1, & \nu_2, & \ldots, & \nu_{m-1} \end{array} \biggl|x_j\right)\right]_{j,k=1}^n.$$ This is the joint probability density function for the ensemble of the squared singular values of the product of $m-1$ rectangular matrices with independent complex Gaussian entries. We conclude that the biorthogonal ensemble defined by equation (\[PExact\]) is an interpolating ensemble: it interpolates between the process of squared singular values from the ensemble of the products of $m$ independent rectangular matrices with independent complex Gaussian entries, and that of $m-1$ independent rectangular matrices with independent complex Gaussian entries. As a final example we give the joint density of the product matrix process related to Hurwitz numbers via [@AmbjornChekov], as mentioned in the introduction. Compared to it reads $$\label{jpdfHurwitz} \frac{1}{Z_n} \exp\left[-{\mathop{\mathrm{Tr}}}\left[QG_1^*G_1\right] -\sum\limits_{l=2}^m{\mathop{\mathrm{Tr}}}\left[G_l^*G_l\right] -{\mathop{\mathrm{Tr}}}\left[V_m\left(\left(G_{m}\cdots G_1\right)^*G_{m}\cdots G_1\right)\right]\right] \prod\limits_{l=1}^mdG_l,$$ that is $V_1=\ldots=V_{m-1}=0$ vanish, and in addition the Gaussian distribution of matrix $G_1$ now includes a nontrivial covariance matrix $Q$ of size $n\times n$ with positive eigenvalues $q_1,\ldots q_n>0$. Consider a multi-matrix model defined by probability measure . Denote by $\left\{y_j^l\right\}_{j=1,\ldots,m}$ the set of the squared singular values of the matrix $Y_l=G_l\ldots G_1$ for each $l=1,\ldots,m$. The joint probability density of $\underline{y}=\left(y^m,\ldots,y^1\right)$ is proportional to $$\label{HurwitzDensityFormula} \det\left[\left(y_j^m\right)^{k-1}e^{-V_m(y_j^m)}\right]_{j,k=1}^n \prod\limits_{l=1}^{m-1} \det\left[\frac{\left(y_j^{l+1}\right)^{\nu_{l+1}}}{\left(y_k^l\right)^{\nu_{l+1}+1}}e^{-\frac{y_j^{l+1}}{y_k^l}}\right]_{j,k=1}^n \det\left[(y_j^1)^{\nu_1}e^{-y_j^1q_k}\right]_{j,k=1}^n.$$ The joint density of $\underline{y}$ for the model defined by can be obtained similarly to the proof of Theorem 2.1, using the standard Harish-Chandra-Itzykson-Zuber formula. In the following we will not pursue this example further. Exact formulae for the correlation kernels ------------------------------------------ Theorem \[TheoremGinibreCouplingDensity\] states that the density of the product matrix process associated with the multi-matrix model (\[MainProbabilityMeasure\]) can be written as a product of determinants. This enables us to apply the result by Eynard and Mehta [@EynardMehta], and to give a formula for all correlation functions. Namely, the Eynard-Mehta Theorem implies that the configuration of the squared singular values of all product matrices associated with the multi-matrix model (\[MainProbabilityMeasure\]) is a *determinantal point process* on $\left\{1,\ldots,m\right\}\times{\mathbb R}_{>0}$. The correlation kernel of this determinantal point process is given by the next theorem. \[TheoremMostGerneralCorrelationKernel\] Consider the multi-matrix model defined by the probability measure (\[MainProbabilityMeasure\]), with $a=1$. Denote by the set $\{y_1^l,\ldots,y_n^l\}$ the squared singular values of the matrix $Y_l=G_l\cdots G_1$, that are the eigenvalues of the matrix $Y_l^*Y_l$. The configuration of all these eigenvalues $$\left\{\left(l,y_j^l\right)\vert l=1,\ldots,m;j=1,\ldots,n\right\} \nonumber$$ forms a determinantal point process on $\left\{1,\ldots,m\right\}\times{\mathbb R}_{>0}$. The correlation kernel of this determinantal point process, $K_{n,m}^{V}(r,x;s,y)$, where $r,s\in\left\{1,\ldots,m\right\}$, and $x, y\in {\mathbb R}_{>0}$, can be written as $$\label{KnmV} K_{n,m}^{V}(r,x;s,y)=-\phi_{r,s}(x,y)+\sum\limits_{i,j=1}^n\phi_{r,m+1}(x,i)\left(A^{-1}\right)_{i,j}\phi_{0,s}(j,y).$$ Here, the elements of matrix $A=(a_{i,j})$, with $i,j=1,\ldots,n$, are defined in . The kernel depends on three sets of functions which are given as follows. *(i)* For the first set of functions $\phi_{r,s}(x,y)$ we distinguish the following cases: - for $r=1,\ldots,m-1$ we have: $$\label{PHI1rr+1} \phi_{r,r+1}(x,y)=\left(\frac{y}{x}\right)^{\nu_{r+1}}\frac{e^{-\frac{y}{x}-V_{r}(x)}}{x}.$$ - for $1\leq r\leq m-2$ and $r+2\leq s\leq m$ we have: $$\label{PHI1rs} \begin{split} \phi_{r,s}(x,y)=&\frac{e^{-V_r(x)}}{x}\int\limits_0^{\infty}\ldots\int\limits_0^{\infty} \left(\frac{t_{r+1}}{x}\right)^{\nu_{r+1}} \left(\frac{t_{r+2}}{t_{r+1}}\right)^{\nu_{r+2}} \ldots \left(\frac{y}{t_{s-1}}\right)^{\nu_{s}} \\ &\times e^{-\frac{t_{r+1}}{x}-\frac{t_{r+2}}{t_{r+1}}-\ldots-\frac{t_{s-1}}{t_{s-2}}-\frac{y}{t_{s-1}}-V_{r+1}\left(t_{r+1}\right) -V_{r+2}\left(t_{r+2}\right)\ldots-V_{s-1}\left(t_{s-1}\right)} \\ &\times \frac{dt_{r+1}}{t_{r+1}}\frac{dt_{r+2}}{t_{r+2}}\ldots \frac{dt_{s-2}}{t_{s-2}}\frac{dt_{s-1}}{t_{s-1}}. \end{split}$$ - for $1\leq s\leq r\leq m$ we have: $$\label{PHI1s<=r} \phi_{r,s}(x,y)=0.$$ *(ii)* For the second set of functions $\phi_{r,m+1}(x,j)$ we distinguish the following cases: - for $1\leq r\leq m-1$ we have : $$\label{PHI2rm+1} \begin{split} \phi_{r,m+1}(x,j)=&\frac{e^{-V_r(x)}}{x}\int\limits_0^{\infty}\ldots\int\limits_0^{\infty} \left(\frac{t_{r+1}}{x}\right)^{\nu_{r+1}} \left(\frac{t_{r+2}}{t_{r+1}}\right)^{\nu_{r+2}} \ldots \left(\frac{t_m}{t_{m-1}}\right)^{\nu_{m}} t_m^{\frac{j-1}{2}}I_{j-1}\left(2bt_m^{\frac{1}{2}}\right) \\ &\times e^{-\frac{t_{r+1}}{x}-\frac{t_{r+2}}{t_{r+1}}-\ldots-\frac{t_{m-1}}{t_{m-2}}-\frac{t_m}{t_{m-1}}-V_{r+1}\left(t_{r+1}\right) -V_{r+2}\left(t_{r+2}\right)\ldots-V_{m-1}\left(t_{m-1}\right)-V_{m}\left(t_{m}\right)}\\ &\times\frac{dt_{r+1}}{t_{r+1}}\frac{dt_{r+2}}{t_{r+2}}\ldots \frac{dt_{m-1}}{t_{m-1}}dt_m. \end{split}$$ - for $r=m$ we have: $$\label{PHI2mm+1} \phi_{m,m+1}(x,j)=x^{\frac{j-1}{2}}I_{j-1}\left(2bx^{\frac{1}{2}}\right)e^{-V_m(x)}.$$ *(iii)* For the third set of functions $\phi_{0,s}(i,y)$ we distinguish the following cases: - for $2\leq s\leq m$ we have: $$\label{PHI30s} \begin{split} \phi_{0,s}(i,y)=&\int\limits_0^{\infty}\ldots\int\limits_0^{\infty} t_1^{\nu_1+i-1}\left(\frac{t_2}{t_1}\right)^{\nu_2}\ldots\left(\frac{t_{s-1}}{t_{s-2}}\right)^{\nu_{s-1}} \left(\frac{y}{t_{s-1}}\right)^{\nu_s} \\ &\times e^{-t_1-\frac{t_2}{t_1}-\ldots-\frac{t_{s-1}}{t_{s-2}}-\frac{y}{t_{s-1}}- V_1\left(t_1\right)-V_2\left(t_2\right)-\ldots-V_{s-2}\left(t_{s-2}\right)-V_{s-1}\left(t_{s-1}\right)}\\ &\times \frac{dt_1}{t_1}\frac{dt_2}{t_2}\ldots \frac{dt_{s-2}}{t_{s-2}}\frac{dt_{s-1}}{t_{s-1}}. \end{split}$$ - for $s=1$ we have: $$\label{PHI301} \phi_{0,1}(i,y)=y^{\nu_1+i-1}e^{-y}.$$ Recall that the case corresponding to $V_1(t)=\ldots=V_{m-2}(t)=0$, $V_{m-1}(t)=b^2t$, and $V_m(t)=0$ corresponds to the Ginibre product process with coupling. Let us denote the relevant correlation kernel by $K_{n,m}(r,x;s,y;b)$, to emphasise its dependence on the coupling constant $b$. For the Ginibre product process with coupling we are able to reduce all multiple integrals in the statement of Theorem \[TheoremMostGerneralCorrelationKernel\] to at most single integrals, expressed in terms of special functions, and to find the matrix $A^{-1}$ explicitly. This leads to a contour integral representation for the correlation kernel $K_{n,m}(r,x;s,y;b)$ as follows. \[TheoremDoubleIntegralRepresentationExactKernel\] For the special case $V_1(t)=\ldots= V_{m-2}(t)=0$, $V_{m-1}(t)=b^2t$, and $V_m(t)=0$ the correlation kernel of Theorem \[TheoremMostGerneralCorrelationKernel\] can be written as $$\label{CorrelationKernelFormulaContour} \begin{split} &K_{n,m}(r,x;s,y;b)=-\phi_{r,s}(x,y;b)+S_{n,m}(r,x;s,y;b). \end{split}$$ The functions $\phi_{r,s}(x,y;b)$ are given by $$\label{PHI(x,y,b)} \begin{split} &\phi_{r,s}(x,y;b)\\ &=\left\{ \begin{array}{llll} \frac{e^{-\frac{y}{x}-b^2x}}{x}, & r=m-1, s=m,\\ \frac{1}{x} G^{s-r,0}_{0,s-r}\left(\begin{array}{cccc} - \\ \nu_{r+1},\ldots,\nu_{s} \end{array}\biggl|\frac{y}{x}\right), & 1\leq r<s\leq m-1,\\ \frac{1}{x} \int\limits_{0}^{\infty} G^{m-r-1,0}_{0,m-r-1}\left(\begin{array}{cccc} - \\ \nu_{r+1},\ldots,\nu_{m-1} \end{array}\biggl|\frac{t}{x}\right) e^{-\frac{y}{t}-b^2t}\frac{dt}{t}, & 1\leq r\leq m-2, s=m,\\ 0, & 1\leq s\leq r\leq m. \end{array} \right. \end{split}$$ The functions $S_{n,m}(r,x;s,y;b)$ can be written as $$\label{Snm} \begin{split} S_{n,m}(r,x;s,y;b)=&\frac{1}{(2\pi i)^2}\int\limits_{-\frac{1}{2}-i\infty}^{-\frac{1}{2}+i\infty}du \oint\limits_{\Sigma_n}dt\frac{\Gamma(t-n+1)}{\Gamma(u-n+1)} \frac{\prod_{j=0}^s\Gamma(u+\nu_j+1)}{\prod_{j=0}^r\Gamma(t+\nu_j+1)}\\ &\times\frac{x^tp_r\left(t,x;b\right)y^{-u-1}q_s\left(u+1,y;b\right)}{u-t}. \end{split}$$ Here, $\Sigma_n$ is a closed contour encircling $0,1,\ldots,n$ in positive direction and such that ${\mathop{\mathrm{Re}}}t>-\frac{1}{2}$ for $t\in\Sigma_n$. In the functions $p_r(t,x;b)$ are defined by $$\label{Functionp} p_r(t,x;b)=\left\{ \begin{array}{ll} 1, & r\in\left\{1,\ldots,m-1\right\}, \\ \frac{\Gamma(t+1)I_t\left(2bx^{\frac{1}{2}}\right)}{\left(bx^{\frac{1}{2}}\right)^t}, & r=m, \end{array} \right.$$ and the functions $q_s\left(u,y;b\right)$ are defined by $$\label{Functionq} q_s\left(u,y;b\right)=\left\{ \begin{array}{ll} 1, & s\in\left\{1,\ldots,m-1\right\}, \\ \frac{2\left(by^{\frac{1}{2}}\right)^uK_u\left(2by^{\frac{1}{2}}\right)}{\Gamma(u)}, & s=m. \end{array} \right.$$ \(a) As a determinantal point process between the first $m-1$ levels, that is on $\left\{1,\ldots,m-1\right\}\times{\mathbb R}_{>0}$, the kernel is independent of $b$ and agrees with kernel $K_{n,m-1}^{\operatorname{Ginibre}}(r,x;s,y)$ of the Ginibre product process *without coupling* found by Strahov [@StrahovD Prop.2.3]: $$\label{K1} \begin{split} K_{n,m-1}^{\operatorname{Ginibre}}(r,x;s,y)=&-\frac{1}{x}G^{s-r,0}_{0,s-r}\left(\begin{array}{ccc} - \\ \nu_{r+1},\ldots,\nu_s \end{array}\biggl|\frac{y}{x}\right)\mathbf{1}_{s>r}\\ &+\frac{1}{(2\pi i)^2}\int\limits_{-\frac{1}{2}-i\infty}^{-\frac{1}{2}+i\infty} du\oint\limits_{\Sigma_n}dt\frac{\prod_{j=0}^s\Gamma(u+\nu_j+1)}{\prod_{j=0}^r\Gamma(t+\nu_j+1)} \frac{\Gamma(t-n+1)}{\Gamma(u-n+1)} \frac{x^ty^{-u-1}}{u-t}, \end{split}$$ where $\mathbf{1}_{s>r}$ is the indicator function[^4]. Because this agreement holds at finite matrix size $n$ it will also hold for the limiting kernel. For $r=m$ or $s=m$ the kernel $K_{n,m}(r,x;s,y;b)$ depends on $b$ and differs from the kernel for the Ginibre product process. \(b) As $b\rightarrow 0$, also for $r=m$ or $s=m$ the correlation kernel $K_{n,m}(r,x;s,y;b)$ turns into the correlation kernel $K_{n,m}^{\operatorname{Ginibre}}(r,x;s,y)$ of the Ginibre product process (\[K1\]) (with $m-1\to m$) from [@StrahovD Prop.2.3]. As a particular case (corresponding to $r=s=m$) we obtain from Theorem \[TheoremDoubleIntegralRepresentationExactKernel\] a double contour integral representation for the correlation kernel of the biorthogonal ensemble defined in Proposition \[PropositionJointDensityTotalProductMatrix\], where only the squared singular values of the total product matrix $Y_m$ are retained. Consider the biorthogonal ensemble of Proposition \[PropositionJointDensityTotalProductMatrix\]. This biorthogonal ensemble defines a determinantal point process on ${\mathbb R}_{>0}$. The correlation kernel of this determinantal point process can be written as $$\label{KERNEL} \begin{split} K_{n,m}(x;y;b)=&\frac{1}{(2\pi i)^2}\int\limits_{-\frac{1}{2}-i\infty}^{-\frac{1}{2}+i\infty}du \oint\limits_{\Sigma_n}dt\frac{\Gamma(t-n+1)}{\Gamma(u-n+1)} \frac{\prod_{j=0}^m\Gamma(u+\nu_j+1)}{\prod_{j=0}^m\Gamma(t+\nu_j+1)}\\ &\times\frac{x^ty^{-u-1}}{u-t}\frac{2\Gamma(t+1)\left(by^{\frac12}\right)^{u+1}}{\Gamma(u+1)\left(bx^{\frac12}\right)^t} I_t\left(2bx^{\frac12}\right)K_{u+1}\left(2by^{\frac12}\right). \end{split}$$ As $b\rightarrow 0$, due to eqs. and the last factors in the second line of can be replaced by 1, and we obtain $$\begin{split} &\underset{b\rightarrow 0}{\lim}K_{n,m}(x,y;b)=\frac{1}{(2\pi i)^2}\int\limits_{-\frac{1}{2}-i\infty}^{-\frac{1}{2}+i\infty}du\oint\limits_{\Sigma_n}dt \frac{\Gamma(t-n+1)}{\Gamma(u-n+1)} \frac{\prod_{j=0}^m\Gamma(u+\nu_j+1)}{\prod_{j=0}^m\Gamma(t+\nu_j+1)}\frac{x^ty^{-u-1}}{u-t}. \end{split}$$ The expression in the right-hand side of this equation represents the correlation kernel for the ensemble of the squared singular values of $m$ rectangular matrices with independent complex Gaussian entries, see Kuijlaars and Zhang [@KuijlaarsZhang Prop.5.1]. Hierarchy of correlation kernels at finite matrix size ------------------------------------------------------ In order to summarise the results presented above let us briefly describe the hierarchy of the correlation kernels under considerations. On top of the hierarchy we have the correlation kernel $K_{n,m}^{V}\left(r,x;s,y\right)$ of Theorem \[TheoremMostGerneralCorrelationKernel\]. This kernel depends on $m$ potential functions $V_1(t)$, $\ldots$, $V_m(t)$, that couple the different levels, and lives on the space $\left\{1,\ldots,m\right\}\times{\mathbb R}_{>0}$. The correlation kernel $K_{n,m}\left(r,x;s,y;b\right)$ of Theorem \[TheoremDoubleIntegralRepresentationExactKernel\] is the specialisation of $K_{n,m}^{V}\left(r,x;s,y\right)$ to the case where $V_1(t)=\ldots=V_{m-2}(t)=0$, $V_{m-1}(t)=b^2t$, and $V_m(t)=0$. It depends on one coupling constant $b>0$, and defines the Ginibre product process with coupling. As we pointed out already, due to our choice of potentials this point process agrees with the Ginibre product process without coupling for the first $m-1$ levels, with $r,s<m$. On the other hand, taking $r=s=m$, we obtain the kernel $K_{n,m}(x,y;b)$ that depends on the parameter $b$. It describes the biorthogonal ensemble for the singular values of the total product matrix with coupling. Thus, for $r=m$ or $s=m$ the kernel $K_{n,m}\left(r,x;s,y;b\right)$ coupled to the $m$th level can be understood as a deformation of the kernel $K_{n,m}^{\operatorname{Ginibre}}(r,x;s,y)$ of the Ginibre product process. For $r=s=m$, the latter kernel $K_{n,m}^{\operatorname{Ginibre}}(m,x;m,y)$ specialises to the finite-$n$ Meijer $G$-kernel, which is the correlation kernel for the ensemble of the squared singular values of $m$ rectangular matrices with independent complex Gaussian entries. Note that the finite-$n$ Meijer $G$-kernel can be obtained from $K_{n,m}(x,y;b)$ by taking the coupling parameter $b$ to zero. Finally, if $r=s=1$, the kernel $K_{n,m}^{\operatorname{Ginibre}}(1,x;1,y)$ specialises to that of the classical Laguerre ensemble. Hard edge scaling limits of the multi-level determinantal processes {#SectionHardEdgeScalingLimits} =================================================================== The point processes considered in this paper are uniquely determined by their correlation functions. Therefore, we will say that the point processes ${\mathcal{P}}$ converges to the point process ${\mathcal{P}}'$, if all correlation functions of ${\mathcal{P}}$ converge to the corresponding correlation functions of ${\mathcal{P}}'$. Since we are dealing with determinantal point processes only, the convergence of the correlation kernel of ${\mathcal{P}}$ to the correlation kernel of ${\mathcal{P}}'$ will be considered as equivalent to the convergence of ${\mathcal{P}}$ to ${\mathcal{P}}'$. In order to discuss scaling limits of the multi-level determinantal processes it is convenient to introduce the following notation, following [@StrahovD]. Denote by $K_{\infty, m}^{\operatorname{Ginibre}}(r,x;s,y)$ the kernel given by the formula $$\label{KernelInfiniteGinibreProductProcess} \begin{split} K_{\infty, m}^{\operatorname{Ginibre}}(r,x;s,y)=&-\frac{1}{x}G^{s-r,0}_{0,s-r}\left(\begin{array}{ccc} - \\ \nu_{r+1},\ldots,\nu_s \end{array}\biggl|\frac{y}{x}\right)\mathbf{1}_{s>r}\\ &+\frac{1}{(2\pi i)^2}\int\limits_{-\frac{1}{2}-i\infty}^{-\frac{1}{2}+i\infty} du\oint\limits_{\Sigma_{\infty}}dt\frac{\prod_{j=0}^s\Gamma(u+\nu_j+1)}{\prod_{j=0}^r\Gamma(t+\nu_j+1)} \frac{\sin\pi u}{\sin\pi t} \frac{x^ty^{-u-1}}{u-t}. \end{split}$$ Here $x>0$, $y>0$, the parameters $r, s$ take values in $\left\{1,\ldots,m\right\}$, and $\Sigma_{\infty}$ is a contour starting from $+\infty$ in the upper half plane and returning to $+\infty$ in the lower half plane, leaving $-\frac{1}{2}$ on the left, and encircling $\{0,1,2,\ldots \}$. Here, we write the correlation kernel $K_{\infty,m}^{\operatorname{Ginibre}}(r,x;s,y)$ with an index $m$, only to emphasise that the variables $r$ and $s$ take values in $\left\{1,\ldots,m\right\}$. Note that for $r=s$ the kernel $K_{\infty,m}^{\operatorname{Ginibre}}(r,x;r,y)=K_{\infty,r}^{\operatorname{Ginibre}}(r,x;r,y)$ is the limiting *Meijer G-kernel* obtained by Kuijlaars and Zhang [@KuijlaarsZhang]. It describes the hard edge scaling limit for the product of $r$ independent complex Gaussian matrices and reduces to the standard Bessel kernel for $r=1$. In [@StrahovD] Strahov showed that $K_{\infty,m}^{\operatorname{Ginibre}}(r,x;s,y)$ can be understood as a multi-level extension of the infinite Meijer $G$-kernel. In the context of products of random matrices, the kernel $K_{\infty, m}^{\operatorname{Ginibre}}(r,x;s,y)$ describes the hard edge scaling limit of the Ginibre product process. Recall that the Ginibre product process is a determinantal point process on $\left\{1,\ldots,m\right\}\times{\mathbb R}_{>0}$ whose density is proportional to expression (\[GinibreDensityFormula\]), and whose correlation kernel is given by formula (\[K1\]). The following proposition was shown in [@StrahovD]: \[PropositionScalingLimitGInibreProcess\] Assume that the point configurations $\left(y_1^m,\ldots,y_n^m;\ldots;y_1^1,\ldots,y_n^1\right)$ form the Ginibre product process on $\left\{1,\ldots,m\right\}\times{\mathbb R}_{>0}$, given by (\[GinibreDensityFormula\]). As $n\rightarrow\infty$, the scaled Ginibre process formed by the point configurations $\left(ny_1^m,\ldots,ny_n^m;\ldots;ny_1^1,\ldots,ny_n^1\right)$ converges to the determinantal point process ${\mathcal{P}}^{\operatorname{Ginibre}}_{\infty,m}$ on $\left\{1,\ldots,m\right\}\times{\mathbb R}_{>0}$, whose correlation kernel, $K_{\infty, m}^{\operatorname{Ginibre}}(r,x;s,y)$, is defined by equation (\[KernelInfiniteGinibreProductProcess\]). Alternatively, we have $$\label{GinlimES} \underset{n\rightarrow\infty}{\lim}\frac{1}{n}K_{n,m}^{\operatorname{Ginibre}}\left(r,\frac{x}{n};s,\frac{y}{n}\right) =K_{\infty, m}^{\operatorname{Ginibre}}(r,x;s,y),$$ where $K_{n, m}^{\operatorname{Ginibre}}(r,x;s,y)$ is defined by equation (\[K1\]) and $K_{\infty, m}^{\operatorname{Ginibre}}(r,x;s,y)$ by equation (\[KernelInfiniteGinibreProductProcess\]). The variables $x$ and $y$ take values in a compact subset of ${\mathbb R}_{>0}$, and $1\leq r,s\leq m$. Now, let us consider the convergence of the Ginibre product process with coupling. Recall that by this we mean the multi-level determinantal process on $\left\{1,\ldots,m\right\}\times{\mathbb R}_{>0}$ whose density is given by expression (\[CouplingDensityFormula1\]) and whose correlation kernel $K_{n,m}(r,x;s,y;b)$ is given by Theorem \[TheoremDoubleIntegralRepresentationExactKernel\]. As we already pointed out previously, for $r,s\leq m-1$ this kernel agrees with that of the Ginibre product process *without* coupling, , for which the hard edge limit was already worked out in [@StrahovD], see in Proposition \[PropositionScalingLimitGInibreProcess\] above (with $m-1\to m$). Therefore, in the following theorem we will only find nontrivial results for the limit of the kernel $K_{n,m}(r,x;s,y;b)$ with $r=m$ or $s=m$, that includes the $m$-th level. In particular we will now drop the assumption that the coupling parameter $b$ is constant, and consider $b$ as a function of the number of particles on the same level, $b=b(n)$. Recall that once $b$ approaches zero, the Ginibre product process with coupling turns into that without, as discussed in Section \[SectionExactResults\]. Depending on how $b(n)$ behaves as a function of $n$, we find three different limits for the limiting kernel that includes level $m$. \[TheoremHardEdgeGinibreCouplingProcess\] Let the point configurations $\left(y_1^m,\ldots,y_n^m;\ldots;y_1^1,\ldots,y_n^1\right)$ with density (\[CouplingDensityFormula1\]) form the Ginibre product process with coupling on $\left\{1,\ldots,m\right\}\times{\mathbb R}_{>0}$, with correlation kernel $K_{n,m}(r,x;s,y;b)$ of Theorem \[TheoremDoubleIntegralRepresentationExactKernel\], equation (\[CorrelationKernelFormulaContour\]).\ **(A)** **Weak coupling regime**. Assume that ${b(n)}/{\sqrt{n}}\rightarrow 0$, as $n\rightarrow\infty$. The scaled Ginibre product process with coupling of configurations $\left(ny_1^m,\ldots,ny_n^m;\ldots;ny_1^1,\ldots,ny_n^1\right)$ converges as $n\rightarrow\infty$ to the determinantal point process ${\mathcal{P}}^{\operatorname{Ginibre}}_{\infty,m}$. Equivalently, we have the following relation to the limiting kernel of the Ginibre product process (\[KernelInfiniteGinibreProductProcess\]), for all levels $1\leq r,s\leq m$: $$\label{WeakCouplingRegimeLimit} \underset{n\rightarrow\infty}{\lim}\frac{1}{n}K_{n,m}\left(r,\frac{x}{n};s,\frac{y}{n};b(n)\right) =K_{\infty, m}^{\operatorname{Ginibre}}(r,x;s,y).$$ Here, the variables $x$ and $y$ take values in a compact subset of ${\mathbb R}_{>0}$.\ **(B)** **Interpolating regime**. Assume that ${b(n)}/{\sqrt{n}}\rightarrow \alpha$, as $n\rightarrow\infty$, where $\alpha>0$. As $n\rightarrow\infty$, the scaled Ginibre product process with coupling formed by the point configurations $\left(ny_1^m,\ldots,ny_n^m;\ldots;ny_1^1,\ldots,ny_n^1\right)$ converges to the determinantal point process ${\mathcal{P}}^{\operatorname{interpol}}_{\infty,m}(\alpha)$ on $\left\{1,\ldots,m\right\}\times{\mathbb R}_{>0}$. Its correlation kernel is given by $$\label{KernelZKExtended} \begin{split} &K_{\infty, m}^{\operatorname{interpol}}(r,x;s,y;\alpha)=-\phi_{r,s}(x,y;\alpha)\\ &+\frac{1}{(2\pi i)^2}\int\limits_{-\frac{1}{2}-i\infty}^{-\frac{1}{2}+i\infty} du\oint\limits_{\Sigma_{\infty}}dt\frac{\prod_{j=0}^s\Gamma(u+\nu_j+1)}{\prod_{j=0}^r\Gamma(t+\nu_j+1)} \frac{\sin\pi u}{\sin\pi t} \frac{x^tp_r(t,x;\alpha)q_s(u+1,y;\alpha)}{y^{u+1}(u-t)}, \end{split}$$ for $r=m$ or $s=m$, with the functions $\phi_{r,s}(x,y;\alpha)$, $p_r(t,x;\alpha)$, and $q_s(u+1,y;\alpha)$ defined by equations (\[PHI(x,y,b)\]), (\[Functionp\]), and (\[Functionq\]), respectively. For $1\leq r,s\leq m-1$ its correlation kernel is given by $K_{\infty, m-1}^{\operatorname{Ginibre}}(r,x;s,y)$. Equivalently, we have for $r=m$ or $s=m$: $$\underset{n\rightarrow\infty}{\lim}\frac{1}{n}K_{n,m}\left(r,\frac{x}{n};s,\frac{y}{n};b(n)\right) =K_{\infty, m}^{\operatorname{interpol}}(r,x;s,y;\alpha),$$ and (with $m\to m-1$ on the right-hand side) for $1\leq r,s\leq m-1$. Here, the variables $x$ and $y$ take values in a compact subset of ${\mathbb R}_{>0}$.\ **(C)** **Strong coupling regime**. Assume that ${b(n)}/{\sqrt{n}}\rightarrow\infty$, as $n\rightarrow\infty$. The scaled process Ginibre product process with coupling formed by the point configurations $$\left(\frac{n}{b(n)}\left(y_1^m\right)^{\frac{1}{2}},\ldots,\frac{n}{b(n)}\left(y_n^m\right)^{\frac{1}{2}}; ny_1^{m-1},\ldots,ny_n^{m-1}; \ldots; ny_1^1,\ldots,ny_n^1\right)$$ converges as $n\rightarrow\infty$ to the determinantal point process ${\mathcal{P}}^{\operatorname{Ginibre}}_{\infty,m-1}$ of $m-1$ levels, with the following identification of levels $m-1$ and $m$ on $\left\{1,\ldots,m\right\}\times{\mathbb R}_{>0}$. Namely, its correlation kernel is given by the following limiting relations:\ $\bullet$ For $1\leq r,s\leq m-1$ we have equation (with $m\to m-1$ on the right-hand side). $\bullet$ For $1\leq r\leq m-1$, $s=m$ we have $$\label{SKR2} \begin{split} &\underset{n\rightarrow\infty}{\lim}\frac{2^{\frac{1}{2}}b(n)y^{\frac{1}{2}}}{n^{\frac{3}{2}}} K_{n,m}\left(r,\frac{x}{n};m,\frac{b^2(n)y^2}{n^2};b(n)\right)\frac{e^{2\frac{b^2(n)}{n}y}}{2^{\frac{1}{2}}\pi^{\frac{1}{2}}}\\ &=K^{\operatorname{Ginibre}}_{\infty,m-1}(r,x;m-1,y)+\mathbf{1}_{r,m-1}\delta(x-y). \end{split}$$ $\bullet$ For $r=m$, $1\leq s\leq m-1$ we have $$\label{SKR3} \underset{n\rightarrow\infty}{\lim}\frac{2^{\frac{1}{2}}b(n)x^{\frac{1}{2}}}{n^{\frac{3}{2}}} K_{n,m}\left(m,\frac{b^2(n)x^2}{n^2};s,\frac{y}{n};b(n)\right)\frac{2^{\frac{1}{2}}\pi^{\frac{1}{2}}}{e^{2\frac{b^2(n)}{n}x}} =K^{\operatorname{Ginibre}}_{\infty,m-1}(m-1,x;s,y).$$ $\bullet$ For $r=m$ and $s=m$ we have $$\label{SKR4} \begin{split} &\underset{n\rightarrow\infty}{\lim}\frac{2b^2(n)x^{\frac{1}{2}}y^{\frac{1}{2}}}{n^{2}} K_{n,m}\left(m,\frac{b^2(n)x^2}{n^2};m,\frac{b^2(n)y^2}{n^2};b(n)\right) \frac{e^{2\frac{b^2(n)}{n}y}}{e^{2\frac{b^2(n)}{n}x}}\\ &=K^{\operatorname{Ginibre}}_{\infty,m-1}(m-1,x;m-1,y). \end{split}$$ In all these formulas the variables $x$ and $y$ take values in a compact subset of ${\mathbb R}_{>0}$. \(a) The interpolating point process ${\mathcal{P}}^{\operatorname{interpol}}_{\infty,m}(\alpha)$, that contains the correlation kernel $K_{\infty, m}^{\operatorname{interpol}}(r,x;s,y;\alpha)$ for $r$ or $s=m$, can be understood as a one-parameter [deformation]{} of the determinantal point process ${\mathcal{P}}^{\operatorname{Ginibre}}_{\infty,m}$ defined by the correlation kernel $K_{\infty, m}^{\operatorname{Ginibre}}(r,x;s,y)$. This will be made more precise in the next Section \[SectionInterpolatingProcess\].\ (b) The limiting point process obtained in the strong coupling regime is a determinantal process living on $m-1$ levels, with an infinite number of points on each level, and where the levels $m$ and $m-1$ have been identified. The correlations between particles on the first $m-1$ levels trivially agrees with that of ${\mathcal{P}}^{\operatorname{Ginibre}}_{\infty,m-1}$, the hard edge scaling limit of the Ginibre product process with $m-1$ levels. The correlations between particles only on the $m$-th level agree with those of the Ginibre product process on the $(m-1)$-th level. Furthermore, the correlations between particles on the $m$-th level and all other particles on the first $m-1$ levels are the same as if all the particles on the $m$-th level would be on the $(m-1)$-th level - unless two points on level $m$ and level $m-1$ coincide, hence the contact interaction in terms of the Dirac delta. In other words, in the asymptotic strong coupling regime the point process on the $m$-th level *collapses* to that on the $(m-1)$-th level, and the $m$-matrix model behaves statistically like an $(m-1)$-matrix model.\ (c) In particular, for $m=2$ the limiting process at strong coupling collapses to the classical Bessel-kernel process of a single level, with the collapse of the first to the second level described in the previous remark. For the biorthogonal ensemble formed by the singular values of the total product matrix $Y_m=G_m\cdots G_1$ only, Theorem \[TheoremHardEdgeGinibreCouplingProcess\] leads to the following result. \[CorollaryTheoremHardEdgeGinibreCouplingProcess\] Let $Y_m=G_m\cdots G_1$ be the total product matrix of the multi-matrix model defined by equation (\[MainProbabilityMeasure1\]), and let $\left(y_1,\ldots,y_n\right)$ be the squared singular values of $Y_m$.\ **(A)** Assume that ${b(n)}/{\sqrt{n}}\rightarrow 0$, as $n\rightarrow\infty$. Then, the point process formed by the point configurations $\left(ny_1,\ldots,ny_n\right)$ converges to the process for the product of $m$ independent Gaussian matrices, i.e. to the determinantal process on ${\mathbb R}_{>0}$ with the correlation kernel given by the Meijer $G$-kernel $K_{\infty, m}^{\operatorname{Ginibre}}(m,x;m,y)$, equation (\[KernelInfiniteGinibreProductProcess\]).\ **(B)** Assume that ${b(n)}/{\sqrt{n}}\rightarrow\alpha$, as $n\rightarrow\infty$, where $\alpha>0$. Then, the point process formed by the point configurations $\left(ny_1,\ldots,ny_n\right)$ converges to the determinantal process on ${\mathbb R}_{>0}$ defined by the correlation kernel $K_{\infty, m}^{\operatorname{interpol}}(m,x;m,y;\alpha)$, given by equation (\[KernelZKExtended\]). The limiting determinantal process can be understood as a one-parameter deformation of the Meijer $G$-kernel process for the product of $m$ independent Gaussian matrices.\ **(C)** Assume that ${b(n)}/{\sqrt{n}}\rightarrow\infty$, as $n\rightarrow\infty$. Then, the scaled process formed by the point configurations $ \left(\frac{n}{b(n)}\left(y_1^m\right)^{\frac{1}{2}},\ldots,\frac{n}{b(n)}\left(y_n^m\right)^{\frac{1}{2}}\right) $ converges to the determinantal process on ${\mathbb R}_{>0}$ with the correlation kernel given by the Meijer $G$-kernel $K_{\infty, m-1}^{\operatorname{Ginibre}}(m-1,x;m-1,y)$, the limiting process for the product of $m-1$ independent Gaussian matrices. The interpolating multi-level determinantal process {#SectionInterpolatingProcess} =================================================== Consider the determinantal process ${\mathcal{P}}^{\operatorname{interpol}}_{\infty,m}(\alpha)$ on $\left\{1,\ldots,m\right\}\times{\mathbb R}_{>0}$ in the interpolating regime. We will show that it interpolates between the hard edge scaling limit ${\mathcal{P}}^{\operatorname{Ginibre}}_{\infty,m}$ of the Ginibre product process of $m$ independent levels, and that in the strong coupling regime among $m-1$ levels. For this reason we will refer to ${\mathcal{P}}^{\operatorname{interpol}}_{\infty,m}(\alpha)$ as to the interpolating multi-level determinantal process. \[TheoremInterpolation\] Assume that the point configurations $$\left(x_1^m,x_2^m,\ldots;x_1^{m-1},x_2^{m-1},\ldots;\ldots;x_1^1,x_2^1,\ldots\right)$$ form the determinantal process ${\mathcal{P}}^{\operatorname{interpol}}_{\infty,m}(\alpha)$ on $\left\{1,\ldots,m\right\}\times{\mathbb R}_{>0}$ in the interpolating regime.\ **(A)** As $\alpha\rightarrow 0$, this determinantal process ${\mathcal{P}}^{\operatorname{interpol}}_{\infty,m}(\alpha)$ converges to the determinantal process ${\mathcal{P}}^{\operatorname{Ginibre}}_{\infty,m}$ corresponding to the hard edge scaling limit of the Ginibre product process of $m$ levels. In particular, for $r=m$ or $s=m$ we find $$\underset{\alpha\rightarrow 0}{\lim}K_{\infty, m}^{\operatorname{interpol}}(r,x;s,y;\alpha)=K_{\infty, m}^{\operatorname{Ginibre}}(r,x;s,y).$$ where the variables $x$ and $y$ take values in a compact subset of ${\mathbb R}_{>0}$.\ **(B)** Consider the scaled interpolating process formed by the point configurations $$\left(\frac{1}{\alpha}\left(x_1^m\right)^{\frac{1}{2}}, \frac{1}{\alpha}\left(x_2^m\right)^{\frac{1}{2}},\ldots;x_1^{m-1},x_2^{m-1},\ldots;\ldots;x_1^1,x_2^1,\ldots\right).$$ As $\alpha\rightarrow\infty$, this determinantal process converges to the hard edge scaling limit in the strong coupling regime. This is equivalent to the following statement:\ $\bullet$ For $1\leq r,s\leq m-1$ we have the kernel (with $m\to m-1$ on the right-hand side).\ $\bullet$ For $1\leq r\leq m-1$, $s=m$ we have $$\label{I2} \underset{\alpha\rightarrow\infty}{\lim}\left(2^{\frac{1}{2}}\alpha y^{\frac{1}{2}} K_{\infty, m}^{\operatorname{interpol}}(r,x;m,\alpha^2y^2;\alpha) \frac{e^{2\alpha^2y^2}}{2^{\frac{1}{2}}\pi^{\frac{1}{2}}} \right)= K_{\infty, m}^{\operatorname{Ginibre}}(r,x;m-1,y)+\mathbf{1}_{r,m-1}\delta(x-y).$$ $\bullet$ For $r=m$, $1\leq s\leq m-1$ we have $$\label{I3} \underset{\alpha\rightarrow\infty}{\lim}\left(2^{\frac{1}{2}}\alpha x^{\frac{1}{2}} K_{\infty, m}^{\operatorname{interpol}}(m,\alpha^2x^2;s,y;\alpha) \frac{2^{\frac{1}{2}}\pi^{\frac{1}{2}}}{e^{2\alpha^2x^2}} \right)= K_{\infty, m}^{\operatorname{Ginibre}}(m-1,x;s,y).$$ $\bullet$ Finally, for $r=m$, $s=m$ we have $$\label{I4} \underset{\alpha\rightarrow\infty}{\lim}\left(2\alpha^2x^{\frac{1}{2}}y^{\frac{1}{2}} K_{\infty, m}^{\operatorname{interpol}}(m,\alpha^2x^2;m,\alpha^2y^2;\alpha)\frac{e^{2\alpha^2y^2}}{e^{2\alpha^2x^2}}\right)= K_{\infty, m}^{\operatorname{Ginibre}}(m-1,x;m-1,y).$$ In all these formulas the variables $x$ and $y$ take values in a compact subset of ${\mathbb R}_{>0}$. In particular, Theorem \[TheoremInterpolation\] implies that the determinantal process on ${\mathbb R}_{>0}$ defined by the correlation kernel $K_{\infty, m}^{\operatorname{interpol}}(m,x;m,y;\alpha)$ is an interpolating (one-level) determinantal point process. It interpolates between the Meijer $G$-kernel process for $m$ independent matrices, and the Meijer $G$-kernel process for $m-1$ independent matrices (with contact interaction from the identification of level $m$ with level $m-1$). An integration formula for coupled matrices {#SectionIntegrationFormula} =========================================== Below we derive an integration formula related to the investigation of singular variables of coupled matrices, see Lemma \[LemmaCoupling\]. The obtained formula will be applied to multi-matrix models, in order to derive the corresponding joint densities. \[LemmaCoupling\] Assume that $\nu\geq 0$ is an integer, and $b>0$. Let $G$ be a complex matrix of size $(n+\nu)\times l$, where $l\geq n\geq 1$, and $X$ be a complex matrix of size $l\times n$, with non-zero squared singular values $x_1,\ldots, x_n$. Denote by $y(GX)=\{y_1,\ldots, y_n\}$ the set of the squared singular values of matrix $Y=GX$. The complex flat Lebesgue measure for matrix $G$ is denoted by $dG$, i.e. $dG=\prod_{j=1}^l\prod_{k=1}^{n+\nu}dG_{j,k}^RdG_{j,k}^I$, $G_{j,k}=G_{j,k}^R+iG_{j,k}^I$ denotes the sum of the real and imaginary parts of the matrix entries $G_{j,k}$, Let $f(.)$ be a continuous function on ${\mathbb R}_{>0}^n$ with compact support. In addition, we assume that $f(y)=f(y_1,\ldots,y_n)$ is symmetric with respect to permutations of $y_1$, $\ldots$, $y_n$. Finally, $V$ (the potential) is some scalar positive function, such that all the following three matrix integrals exist.\ **(A)** We have for $l=n$ $$\label{CouplingFormula} \begin{split} &\int f(y(GX))e^{-{\mathop{\mathrm{Tr}}}\left[G^*G\right]+b{\mathop{\mathrm{Tr}}}\left[GX+\left(GX\right)^*\right]-{\mathop{\mathrm{Tr}}}V\left(\left(GX\right)^*(GX)\right)}dG\\ &=\frac{c}{\Delta_n(\{x_j\})}\\ &\times\int\limits_0^{\infty}\ldots\int\limits_0^{\infty}f(y_1,\ldots,y_n) \det\left[\frac{e^{-\frac{y_j}{x_k}-V(y_j)}}{x_k}\right]_{j,k=1}^n\det\left[y_j^{\frac{k-1}{2}}I_{k-1} \left(2by_j^{\frac{1}{2}}\right)\right]_{j,k=1}^ndy_1\ldots dy_n, \end{split}$$ where we recall that the Vandermonde determinant was defined in . The constant $c$ does not depend on the set $\{x_1,\ldots,x_n\}$.\ **(B)** Without the coupling term ($b=0$) and for general $l\geq n$ we have $$\label{CouplingFormula1} \begin{split} &\int f(y(GX))e^{-{\mathop{\mathrm{Tr}}}\left[G^*G\right]-{\mathop{\mathrm{Tr}}}V\left(\left(GX\right)^*(GX)\right)}dG\\ &=\frac{c'}{\Delta_n(\{x_i\})}\int\limits_0^{\infty}\ldots\int\limits_0^{\infty}f(y_1,\ldots,y_n) \det\left[\frac{y_j^{\nu}e^{-\frac{y_j}{x_k}-V(y_j)}}{x_k^{\nu+1}}\right]_{j,k=1}^n\Delta_n\left(\{y_i\}\right)dy_1\ldots dy_n. \end{split}$$ Here the constant $c'$ does not depend on the set $\{x_1,\ldots,x_n\}$.\ **(C)** Without matrix $X$ and for $l=n$ we have: $$\label{CouplingFormula2} \begin{split} &\int f(y(G))e^{-{\mathop{\mathrm{Tr}}}\left[G^*G\right]-{\mathop{\mathrm{Tr}}}V\left(G^*G\right)}dG\\ &=c''\int\limits_0^{\infty}\ldots\int\limits_0^{\infty}f(y_1,\ldots,y_n) \Delta_n^2\left(\{y_i\}\right)\prod\limits_{j=1}^ny_j^{\nu}e^{-y_j-V(y_j)}dy_1\ldots dy_n, \end{split}$$ where here $y(G)=\{y_1,\ldots,y_n\}$ is the set of squared singular values of $G$ instead. Consider the following measure $$P(G,X)dG=e^{-{\mathop{\mathrm{Tr}}}\left[G^*G\right]+b{\mathop{\mathrm{Tr}}}\left[GX+\left(GX\right)^*\right]-{\mathop{\mathrm{Tr}}}V\left(\left(GX\right)^*(GX)\right)}dG.$$ Following the analysis of Fischman et al. [@Jonit] we set $$X=U\left(\begin{array}{c} X_0 \\ 0 \end{array} \right),$$ where $U$ is an $(n+\nu)\times (n+\nu)$ unitary matrix, $0$ is a $\nu\times n$ matrix consisting of zeros only, and $X_0$ is an $n\times n$ complex matrix. We have $$\begin{split} &P(G,X)dG\\ &=e^{-{\mathop{\mathrm{Tr}}}\left[G^*G\right]+b{\mathop{\mathrm{Tr}}}\left[GU\left(\begin{array}{c} X_0 \\ 0 \end{array} \right) +\left(\begin{array}{cc} X_0^* & 0 \end{array} \right)U^*G^*\right]-{\mathop{\mathrm{Tr}}}V\left( \left(\begin{array}{cc} X_0^* & 0 \end{array} \right)U^*G^* GU\left(\begin{array}{c} X_0\\ 0 \end{array}\right) \right)}dG. \end{split} \nonumber$$ Set $\widehat{G}=GU$, and note that $\widehat{G}$ is a matrix of size $n\times (n+\nu)$. Use the invariance of the Lebesgue measure under unitary transformations to write $$P(G,X)dG=e^{-{\mathop{\mathrm{Tr}}}\left[\widehat{G}^*\widehat{G}\right]+b{\mathop{\mathrm{Tr}}}\left[\widehat{G}\left(\begin{array}{c} X_0 \\ 0 \end{array} \right) +\left(\begin{array}{cc} X_0^* & 0 \end{array} \right)\widehat{G}^*\right] -{\mathop{\mathrm{Tr}}}V\left( \left(\begin{array}{cc} X_0^* & 0 \end{array} \right)\widehat{G}^* \widehat{G}\left(\begin{array}{c} X_0\\ 0 \end{array}\right) \right)}d\widehat{G},$$ where $d\widehat{G}=\prod_{j=1}^n\prod_{k=1}^{n+\nu}d\widehat{G}_{j,k}^Rd\widehat{G}_{j,k}^I$, and where $\widehat{G}_{j,k}=\widehat{G}_{j,k}^R+i\widehat{G}_{j,k}^I$ denotes the sum of the real and imaginary parts of the matrix entries $\widehat{G}_{j,k}$. Write $\widehat{G}$ as $\widehat{G}=\left(G_0,G_1\right)$, where $G_0$ is a matrix of size $n\times n$, and where $G_1$ is a matrix of size $n\times\nu$. We obtain the following decomposition of the measure $P(G,X)dG$ $$P(G,X)dG=\left(e^{-{\mathop{\mathrm{Tr}}}\left[G_0^*G_0\right]+b{\mathop{\mathrm{Tr}}}\left[G_0X_0 +G_0^*X_0^*\right] -{\mathop{\mathrm{Tr}}}V\left(\left(G_0X_0\right)^*\left(G_0X_0\right)\right)}dG_0\right) e^{-{\mathop{\mathrm{Tr}}}\left[G_1^*G_1\right]}dG_1.$$ We have put brackets here to emphasise that the integrals over $G_0$ and $G_1$ decouple, with the latter giving only an additional multiplicative constant. The important observation is that the matrices $X$ and $X_0$ have the same singular values, and that the matrices $GX$ and $G_0X_0$ have the same singular values. Set $Y=G_0X_0$. We have $dG_0=\det^{-n}\left[X_0^*X_0\right]dY$, which gives $$\begin{split} &P(G,X)dG\\ &=\left(e^{-{\mathop{\mathrm{Tr}}}\left[\left(YX_0^{-1}\right)^*YX_0^{-1}\right]+b{\mathop{\mathrm{Tr}}}\left[Y +Y^*\right] -{\mathop{\mathrm{Tr}}}V\left(Y^*Y\right)} {\det}^{-n}\left[X_0^*X_0\right]dY\right) e^{-{\mathop{\mathrm{Tr}}}\left[G_1^*G_1\right]}dG_1. \end{split}$$ Here, we have used that $X_0$ is invertible, which is ensured by its non-zero squared singular values. The the singular value decomposition of the matrix $Y$ can be written as $Y=\widetilde{U}\Sigma P$, where both $\widetilde{U}$, $P$ are unitary matrices of the same size $n\times n$, and where $\Sigma$ is an $n\times n$ diagonal matrix with a real matrix entries, $$\Sigma=\left(\begin{array}{cccc} \sqrt{y_1} & 0 & \ldots & 0 \\ 0 & \sqrt{y_2} & \ldots & 0 \\ \vdots & & &\\ 0 & 0 & 0 & \sqrt{y_n} \end{array} \right),$$ and $y_1,\ldots,y_n$ are the squared singular values of $Y$. The Jacobian corresponding to this decomposition is $$dY=c_1\Delta_n(\{y_j\})^2d\widetilde{U}dPdy_1\ldots dy_n,$$ where $c_1$ is some constant. We thus obtain $$\begin{split} P(G,X)dG=&c_1e^{-{\mathop{\mathrm{Tr}}}\left[P^*\Sigma^2 P\left(X_0^*X_0\right)^{-1}\right]+b{\mathop{\mathrm{Tr}}}\left[\widetilde{U}\Sigma P +P^*\Sigma \widetilde{U}^*\right] -{\mathop{\mathrm{Tr}}}V\left(\Sigma^2\right)} {\det}^{-n}\left[X_0^*X_0\right]\Delta_n(\{y_j\})^2d\widetilde{U}dV\\ &\times e^{-{\mathop{\mathrm{Tr}}}\left[G_1^*G_1\right]}dG_1. \end{split} \nonumber$$ The next step is to use the invariance of the Haar measure under left shifts by the group elements, $\widetilde{U}\rightarrow \hat{U}=P\widetilde{U}$, and to rewrite the expression above as $$\begin{split} P(G,X)dG=&c_1e^{-{\mathop{\mathrm{Tr}}}\left[P^*\Sigma^2 P\left(X_0^*X_0\right)^{-1}\right]+b{\mathop{\mathrm{Tr}}}\left[\hat{U}\Sigma +\Sigma \hat{U}^*\right]-{\mathop{\mathrm{Tr}}}V\left(\Sigma^2\right)} {\det}^{-n}\left[X_0^*X_0\right]\Delta_n(\{y_j\})^2d\hat{U}dP\\ &\times e^{-{\mathop{\mathrm{Tr}}}\left[G_1^*G_1\right]}dG_1. \end{split} \nonumber$$ The integration over $P$ can be performed using the Harish-Chandra–Itzykson-Zuber integral formula [@HC; @IZ] $$\int_{U(n)}e^{-{\mathop{\mathrm{Tr}}}\left[P^*\Sigma^2P\left(X_0^*X_0\right)^{-1}\right]}dP=\operatorname{const}\frac{\det\left[e^{-\frac{y_j}{x_k}}\right]_{j,k=1}^n}{\triangle(\{y_j\})\Delta_n(\{x_j^{-1}\})},$$ where the constant depends only on $n$. The integration over ${\hat{U}}$ can be done exploiting the following integration formula [@Brower] (sometimes called Leutwyler-Smilga formula [@LeutwylerSmilga]) $$\int_{U(n)}e^{b{\mathop{\mathrm{Tr}}}\left[\Sigma\left({\hat{U}}+{\hat{U}}^*\right)\right]}d{\hat{U}} =\operatorname{const}\frac{\det\left[y_j^{\frac{k-1}{2}}I_{k-1}\left(2by_j^{\frac{1}{2}}\right)\right]_{j,k=1}^n}{\Delta_n(\{y_j\})},$$ where $I_{\kappa}(x)$ denotes the modified Bessel function of the first kind. After integration, and after some simplifications, $\Delta_n(\{x_j^{-1}\})=(-1)^{\frac{n(n-1)}{2}}\prod_{k=1}^nx_k^{-n+1} \Delta_n(\{x_j\})$, we obtain formula (\[CouplingFormula\]) in the statement of the Lemma. Formula (\[CouplingFormula1\]) can be obtained in the same way, and formula (\[CouplingFormula2\]) is well known. \[TheoremDeformedDensity\] Let $\nu$, $X$, $G$ be as in the statement of Lemma \[LemmaCoupling\], and consider the probability measure $$P(G,X)dG=\frac{1}{Z_n}e^{-{\mathop{\mathrm{Tr}}}\left[G^*G\right]+b{\mathop{\mathrm{Tr}}}\left[GX+\left(GX\right)^*\right]}dG$$ over rectangular complex matrices $G$. Here, $dG$ is the flat complex Lebesgue measure, and $Z_n$ is a normalising constant. Then, the density of the squared singular values $y_1,\ldots,y_n$ of $Y=GX$ is $$P(y_1,\ldots,y_n)dy_1\ldots dy_n=\frac{\prod_{l=1}^n{{x_l}^{-1}e^{-b^2x_l}}}{b^{\frac{n(n-1)}{2}}}\frac{\det\left[e^{-\frac{y_j}{x_k}}\right]_{j,k=1}^n \det\left[y_j^{\frac{k-1}{2}}I_{k-1}(2by_j)\right]_{j,k=1}^n}{n!\Delta_n(\{x_j\})}.$$ By applying Lemma \[LemmaCoupling\] (A) with $V=0$ the distribution follows. We only need to compute the following integral in order to determine the normalisation constant: $$I_n=\int\limits_{0}^{\infty}\ldots\int\limits_0^{\infty}\det\left[e^{-\frac{y_j}{x_k}}\right]_{j,k=1}^n \det\left[y_j^{\frac{k-1}{2}}I_{k-1}\left(2by_j^{\frac{1}{2}}\right)\right]_{j,k=1}^ndy_1\ldots dy_n.$$ Applying the Andréief integral identity $$\label{Andre} \int\cdots\int\det\left[\varphi_i\left(y_j\right)\right]_{i,j=1}^n \det\left[\psi_i\left(y_j\right)\right]_{i,j=1}^ndy_1\cdots dy_n=n! \det\left[\int\varphi_i\left(y\right)\psi_j\left(y\right)dy\right]_{i,j=1}^n\ ,$$ where the two sets of functions $\varphi_i$ and $\psi_i$ are assumed to be such that all integrals exist, we find that $$I_n=n!\det\left[\int\limits_0^{\infty}e^{-\frac{y}{x_k}}y^{\frac{j-1}{2}}I_{j-1}\left(2by^{\frac{1}{2}}\right)dy\right]_{j,k=1}^n.$$ The integral inside the determinant can be computed explicitly. The result is [@Grad 6.631.4] $$\label{I-int1} \int\limits_0^{\infty}e^{-\frac{y}{x_k}}y^{\frac{j-1}{2}}I_{j-1}\left(2by^{\frac{1}{2}}\right)dy=(x_k)^jb^{j-1}e^{{b^2}x_k}.$$ This gives $$I_n=b^{\frac{n(n-1)}{2}}\left(\prod\limits_{k=1}^nx_ke^{b^2x_k}\right)\Delta_n(\{x_j\}).$$ The formula in the statement of Proposition \[TheoremDeformedDensity\] follows immediately. Note that if $b=0$, then Proposition \[TheoremDeformedDensity\] reduces to the following known result, cf. Kuijlaars and Stivigny [@ArnoDries Lemma 2.2] and references therein. Let $G$ be a complex Ginibre matrix of size $n\times (n+\nu)$, and let $X$ be a fixed complex matrix of size $(n+\nu)\times n$ with nonzero squared singular values $x(X)=(x_1,\ldots,x_n)$. Then the squared singular values $y_1,\ldots,y_n$ of matrix $Y=GX$ have a joint probability density proportional to $$\frac{\Delta_n(\{y_j\})}{\Delta_n(\{x_j\})}\det\left[\frac{e^{-\frac{y_j}{x_k}}}{x_k}\right]_{j,k=1}^n.$$ Proof of Theorem \[TheoremGinibreCouplingDensity\] ================================================== Set $G=G_m$, of size $n\times (n+\nu_{m-1})$ (with $\nu_m=0)$, and $X=G_{m-1}\cdots G_1$ of size $(n+\nu_{m-1})\times n$ (with $\nu_0=0)$. Denote by $y(GX)$ the vector $y^m=\left(y^m_1,\ldots,y^m_n\right)$, which is the vector of the squared singular values of the matrix $Y_m=GX=G_mG_{m-1}\cdots G_1$. Note that $y^{m-1}=\left(y_1^{m-1},\ldots,y^{m-1}_n\right)$ is the vector of the squared values of $X$. Let $$f:\underset{m\;\;\mbox{times}}{\underbrace{\left({\mathbb R}_{>0}\right)^n\times\ldots\times\left({\mathbb R}_{>0}\right)^n}}\longrightarrow{\mathbb C}$$ be a continuous function with compact support, and denote by $\operatorname{\mathbb{E}}\left[f\right]$ the expectation with respect to the probability measure defined by equation (\[MainProbabilityMeasure\]). Using this notation we can write $$\begin{split} &\operatorname{\mathbb{E}}\left[f\left(y^m;\ldots;y^1\right)\right]\\ &=\frac{1}{Z_n}\int\left(\int f(y(GX);y^{m-1};\ldots;y^1) e^{-{\mathop{\mathrm{Tr}}}\left[G^*G\right]+b{\mathop{\mathrm{Tr}}}\left[GX +\left(GX\right)^*\right]-{\mathop{\mathrm{Tr}}}V_m\left(\left(GX\right)^*\left(GX\right)\right)} dG\right)\\ &\times e^{-\sum\limits_{l=1}^{m-1}{\mathop{\mathrm{Tr}}}\left[G_l^*G_l\right] -\sum\limits_{l=1}^{m-1}{\mathop{\mathrm{Tr}}}V_l\left(\left(G_{l}\cdots G_1\right)^*\left(G_{l}\cdots G_1\right)\right)}\prod\limits_{l=1}^{m-1}dG_l. \end{split} \nonumber$$ The application of Lemma \[LemmaCoupling\] (more explicitly, of equation (\[CouplingFormula\])) to the integral over $G$ in the equation just written above gives $$\label{MainProofExpectation1} \begin{split} &\operatorname{\mathbb{E}}\left[f\left(y^m;\ldots;y^1\right)\right]\\ &=\int f_1\left(y^{m-1};\ldots;y^1\right)e^{-\sum\limits_{l=1}^{m-1}{\mathop{\mathrm{Tr}}}\left[G_l^*G_l\right] -\sum\limits_{l=1}^{m-1}{\mathop{\mathrm{Tr}}}V_l\left(\left(G_{l}\cdots G_1\right)^*\left(G_{l}\cdots G_1\right)\right)}\prod\limits_{l=1}^{m-1}dG_l, \end{split}$$ where $$\begin{split} &f_1\left(y^{m-1};\ldots;y^1\right)\\ &=\frac{c_m}{Z_n}\int\limits_{{\mathbb R}_{>0}^n} f\left(y^m; \ldots,y^{1}\right) \det\left[\frac{e^{-\frac{y_j^m}{y_k^{m-1}}-V_m(y_j^m)}}{y_k^{m-1}}\right]_{j,k=1}^n \frac{\det\left[\left(y_j^m\right)^{\frac{k-1}{2}}I_{k-1} \left(2b\left(y_j^m\right)^{\frac{1}{2}}\right)\right]_{j,k=1}^n}{\Delta_n\left(\{y_i^{m-1}\}\right)}dy^m. \end{split} \nonumber$$ Here, we denote by $dy^m=dy_1^m\cdots dy_n^m$ the integration measure over all squared singular values. Next, let us apply Lemma \[LemmaCoupling\], equation (\[CouplingFormula1\]) to equation (\[MainProofExpectation1\]), with $G=G_{m-1}$ of size $(n+\nu_{m-1})\times(n+\nu_{m-2})$, and $X=G_{m-2}\cdots G_1$ of size $(n+\nu_{m-2})\times n$, and integrate over $G_{m-1}$. The result can be written as $$\label{MainProofExpectation2} \begin{split} &\operatorname{\mathbb{E}}\left[f\left(y^m;\ldots;y^1\right)\right]\\ &=\int f_2\left(y^{m-2};\ldots;y^1\right)e^{-\sum\limits_{l=1}^{m-2}{\mathop{\mathrm{Tr}}}\left[G_l^*G_l\right] -\sum\limits_{l=1}^{m-2}{\mathop{\mathrm{Tr}}}V_l\left(\left(G_{l}\cdots G_1\right)^*\left(G_{l}\cdots G_1\right)\right)}\prod\limits_{l=1}^{m-2}dG_l, \end{split}$$ where $$\begin{split} &f_2\left(y^{m-2};\ldots;y^1\right)\\ &=\frac{c_mc_{m-1}}{Z_n}\int\limits_{{\mathbb R}_{>0}^n}\int\limits_{{\mathbb R}_{>0}^n} f\left(y^m; \ldots,y^{1}\right) \det\left[\frac{e^{-\frac{y_j^m}{y_k^{m-1}}-V_m(y_j^m)}}{y_k^{m-1}}\right]_{j,k=1}^n \\ &\times \det\left[\frac{\left(y_j^{m-1}\right)^{\nu_{m-1}}e^{-\frac{y_j^{m-1}}{y_k^{m-2}}-V_{m-1}(y_j^{m-1})}}{\left(y_k^{m-2}\right)^{\nu_{m-1}+1}}\right]_{j,k=1}^n \frac{\det\left[\left(y_j^m\right)^{\frac{k-1}{2}}I_{k-1} \left(2b\left(y_j^m\right)^{\frac{1}{2}}\right)\right]_{j,k=1}^n}{\Delta_n\left(\{y_i^{m-2}\}\right)} dy^{m-1}dy^m. \end{split} \nonumber$$ Repeating this procedure $m-3$ times, we get $$\label{MainProofExpectation3} \begin{split} &\operatorname{\mathbb{E}}\left[f\left(y^m;\ldots;y^1\right)\right]=\int f_{m-1}\left(y^1\right)e^{-{\mathop{\mathrm{Tr}}}\left[G_1^*G_1\right] -{\mathop{\mathrm{Tr}}}V_1\left(G_1^*G_1\right)}dG_1, \end{split}$$ where $$\label{MainProofExpectation4} \begin{split} &f_{m-1}\left(y^1\right)\\ &=\frac{\prod_{l=2}^mc_l}{Z_n\Delta_n\left(\{y_i^1\}\right)}\int\limits_{\left({\mathbb R}_{>0}^n\right)^{m-1}} f\left(y^m; \ldots,y^{1}\right) \det\left[\left(y_j^m\right)^{\frac{k-1}{2}}I_{k-1}\left(2b\left(y_j^m\right)^\frac{1}{2}\right)\right]_{j,k=1}^n\\ &\times \prod\limits_{l=1}^{m-1}\det\left[\frac{\left(y_j^{l+1}\right)^{\nu_{l+1}}}{\left(y_k^l\right)^{\nu_{l+1}+1}}e^{-\frac{y_j^{l+1}}{y_k^l} -V_{l+1}\left(y_j^{l+1}\right)}\right]_{j,k=1}^n dy^2\ldots dy^m. \end{split}$$ The result of Theorem \[TheoremGinibreCouplingDensity\] follows by application of formula (\[CouplingFormula2\]) to equation (\[MainProofExpectation3\]), and by taking into account equation (\[MainProofExpectation4\]). Eq. is obtained by redistributing the factors with the potentials among the determinants, including the remaining Vandermonde determinant. Finally the normalisation constant $Z_{n,m}$ in is obtained by an $m$-fold application of the Andréief formula . Proof of Proposition \[PropositionJointDensityTotalProductMatrix\] ================================================================== The integration of the density (\[CouplingDensityFormula1\]) of $\underline{y}=\left(y^m,y^{m-1},\ldots,y^1\right)$ over the sets of variables $y^{m-1}$, $\ldots$, $y^1$ gives Eq. (\[PExact\]) by applying the Andréief formula $m-1$ times, with $$\label{Psi-initial} \begin{split} \psi_k(y)=&\int\limits_{0}^{\infty}\ldots\int\limits_{0}^{\infty} t^{\nu_1+k-1}\left(\frac{t_2}{t_1}\right)^{\nu_2}\ldots\left(\frac{t_{m-1}}{t_{m-2}}\right)^{\nu_{m-1}} \left(\frac{y}{t_{m-1}}\right)^{\nu_m}\\ &\times e^{-t_1-\frac{t_2}{t_1}-\ldots-\frac{t_{m-1}}{t_{m-2}}-\frac{y}{t_{m-1}}-b^2t_{m-1}} \frac{dt_1}{t_1}\frac{dt_2}{t_2}\ldots\frac{dt_{m-1}}{t_{m-1}}. \end{split}$$ We note that the integration over $t_1$, $t_2$, $\ldots$, $t_{m-2}$ in the formula just written above results into the Meijer $G$-function (see e.g. [@AkemannIpsenKieburg]), namely $$\label{Psi1} \begin{split} \psi_k(y)=&\int\limits_{0}^{\infty} G^{m-1,0}_{0,m-1}\left(\begin{array}{cccc} & - & & \\ \nu_{1}+k-1 & \nu_2 & \ldots & \nu_{m-1} \end{array} \biggl\vert t_{m-1}\right)e^{-\frac{y}{t_{m-1}}-b^2t_{m-1}} \frac{dt_{m-1}}{t_{m-1}}. \end{split}$$ Using the contour integral representation for the Meijer $G$-function, cf. [@Luke], $$\begin{split} &G^{m-1,0}_{0,m-1}\left(\begin{array}{cccc} & - & & \\ \nu_{1}+k-1 & \nu_2 & \ldots & \nu_{m-1} \end{array} \biggl\vert t_{m-1}\right)\\ &=\frac{1}{2\pi i}\int\limits_{c-i\infty}^{c+i\infty} \Gamma\left(u+\nu_1+k-1\right)\Gamma\left(u+\nu_2\right)\ldots\Gamma\left(u+\nu_{m-1}\right)\left(t_{m-1}\right)^{-u}du, \end{split} \nonumber$$ with $c>0$, and taking into account that $$\int\limits_0^{\infty}\left(t_{m-1}\right)^{-u-1}e^{-\frac{y}{t_{m-1}}-b^2t_{m-1}}dt_{m-1}= 2\left(\frac{y}{b^2}\right)^{-\frac{u}{2}}K_{u}\left(2b\sqrt{y}\right),$$ we obtain equation (\[PSIK\]). The interchange of integrals can be justified with Fubini’s Theorem. The normalisation constant as given in can be obtained as follows. Applying the Andréief formula to once, we need to compute the determinant of the following integral $$a_{i,j}(b)=\int\limits_0^\infty y^{\frac{j-1}{2}} I_{j-1}\left( 2by^{\frac12}\right) \psi_i(y) dy \ .$$ Using the representation of the function $\psi_j(y)$ from , we observe that we can use the integral (with $x_k$ replaced by $t_{m-1}$) to obtain $$\label{aijb} a_{i,j}(b)=b^{j-1}\int\limits_{0}^{\infty}\ldots\int\limits_{0}^{\infty} t^{\nu_1+i-1}\left(\frac{t_2}{t_1}\right)^{\nu_2}\ldots\left(\frac{t_{m-1}}{t_{m-2}}\right)^{\nu_{m-1}} (t_{m-1})^{j} e^{-t_1-\frac{t_2}{t_1}-\ldots-\frac{t_{m-1}}{t_{m-2}} } \frac{dt_1}{t_1} \ldots\frac{dt_{m-1}}{t_{m-1}}.$$ The powers in $b$ can be taken out of the determinant of $a_{i,j}(b)$, and the remaining integral is the same as in the normalisation of $m-1$ product of independent Ginibre matrices. For these the determinant has been computed in [@AkemannIpsenKieburg] and we thus obtain for $$\det\left[ a_{i,j}(b)\right]_{i,j=1}^n=b^{\frac{n(n-1)}{2}}\prod_{j=1}^n\prod_{l=1}^{m-1}\Gamma(j+\nu_l)\ .$$ Together with the $n!$ from the Andréief formula we obtain . Measures given by products of determinants, Eynard-Mehta Theorem, and proof of Theorem \[TheoremMostGerneralCorrelationKernel\] {#SECTIONEynardMehta} =============================================================================================================================== The aim of this section is to prove Theorem \[TheoremMostGerneralCorrelationKernel\]. Recall that Theorem \[TheoremMostGerneralCorrelationKernel\] states that the product matrix process associated with probability measure (\[MainProbabilityMeasure\]), defined in Section \[SectionExactResults1\], is a multi-level determinantal process living on $\left\{1,\ldots,m\right\}\times{\mathbb R}_{>0}$. Moreover, Theorem \[TheoremMostGerneralCorrelationKernel\] gives a formula for the relevant correlation kernel. The starting point of the proof of Theorem 2.4 is the fact that the density of the product matrix process under considerations is given by a product of determinants, see Theorem \[TheoremGinibreCouplingDensity\]. This enables us to apply the Eynard-Mehta Theorem to the product matrix process. Let us first recall the formulation of the Eynard-Mehta Theorem. Here we follow the elegant presentation of the Eynard-Mehta Theorem in Johansson [@Johansson][^5]. Let $n, m\geq 1$ be two fixed natural numbers, and let ${\mathfrak{X}}_0$, ${\mathfrak{X}}_{m+1}$ be two given sets. Let ${\mathfrak{X}}$ be a complete separable metric space, and consider a probability measure on $({\mathfrak{X}}^n)^m$ given by $$\label{ProductDeterminantsMeasure} \begin{split} p_{n,m}(\underline{x})d\mu(\underline{x})&=\frac{1}{Z_{N,m}}\det\left[\phi_{0,1}(x_i^0,x_j^1)\right]_{i,j=1}^n\det\left[\phi_{m,m+1}(x_i^m,x_j^{m+1})\right]_{i,j=1}^n\\ & \times\prod\limits_{r=1}^{m-1}\det\left[\phi_{r,r+1}(x_i^r,x_j^{r+1})\right]_{i,j=1}^nd\mu(\underline{x}). \end{split}$$ In the formula just written above $Z_{N,m}$ is the normalisation constant, the functions $\phi_{r,r+1}: {\mathfrak{X}}\times {\mathfrak{X}}\rightarrow {\mathbb C}$, $r=1,\ldots,m-1$ are given *intermediate one-step transition functions*, $\phi_{0,1}: {\mathfrak{X}}_0\times {\mathfrak{X}}\rightarrow {\mathbb C}$ is a given *initial one-step transition function*, and $\phi_{m,m+1}: {\mathfrak{X}}\times {\mathfrak{X}}_{m+1}\rightarrow {\mathbb C}$ is a given *final one-step transition function*. Also, $$\underline{x}=\left(x^1,\ldots,x^m\right)\in\left(X^n\right)^m; \;\; x^r=\left(x^r_1,\ldots,x^r_n\right), r=1,\ldots, m,$$ the vectors $$x^0=(x^0_1,\ldots,x^0_n)\in {\mathfrak{X}}_0^n,\;\; x^{m+1}=(x^{m+1}_1,\ldots,x^{m+1}_n)\in {\mathfrak{X}}_{m+1}^n,$$ are fixed initial and final vectors, and $$d\mu(\underline{x})=\prod\limits_{r=1}^m\prod\limits_{j=1}^nd\mu(x_j^r).$$ Here, $\mu$ is a given Borel measure on ${\mathfrak{X}}$. Given two transition functions $\phi$ and $\psi$ set $$\phi\ast\psi(x,y)=\int_{{\mathfrak{X}}}\phi(x,t)\psi(t,y)d\mu(t).$$ Consider the probability measure defined by equation (\[ProductDeterminantsMeasure\]). The distribution of the vector $x^m=\left(x_1^m,\ldots,x_n^m\right)$ is given by $$\label{PmGeneral} \begin{split} &p(x_1^m,\ldots,x_n^m)d\mu(x_1^m)\ldots d\mu(x_n^m)\\ &=\frac{1}{Z_{n,m}'}\det\left[\phi_{0,m}(x_i^0,x_j^m)\right]_{i,j=1}^n \det\left[\phi_{m,m+1}(x_i^m,x_j^{m+1})\right]_{i,j=1}^nd\mu(x_1^m)\ldots d\mu(x_n^m), \end{split} \nonumber$$ where $$\phi_{0,m}(x,y)=\phi_{0,1}\ast\ldots\ast\phi_{m-1,m}(x,y),$$ for $m>1$, and where the normalisation constant is given by $$Z_{n,m}'=n!\det\left[\phi_{0,m+1}(x_i^0,x_j^{m+1})\right]_{i,j=1}^n.$$ The density of $\left(x_1^m,\ldots,x_n^m\right)$ can be obtained by subsequent integration of the measure $p_{n,m}(\underline{x})d\mu(\underline{x})$ over $x^1$, $\ldots$, $x^{m-1}$, and by application of the Andréief integral identity . Let us define the following correlation functions for the process defined by probability measure (\[ProductDeterminantsMeasure\]): $$\label{rhok-def} \varrho_{k_1,\ldots,k_m}\left(x_1^1,\ldots,x_{k_1}^1;\ldots;x_1^m,\ldots,x_{k_m}^m\right) = \prod_{j=1}^m\frac{n!}{(n-k_j)!} \int_{{\mathfrak{X}}}\cdots\int_{{\mathfrak{X}}}p_{n,m}(\underline{x})\prod\limits_{r=1}^m\prod\limits_{j=k_r+1}^nd\mu(x_j^r).$$ The following statement determines these for the point process and is often referred as the Eynard-Mehta Theorem [@EynardMehta]. \[TheoremEynardMehta\] The probability measure $p_{n,m}(\underline{x})d\mu(\underline{x})$ given by equation (\[ProductDeterminantsMeasure\]) defines a determinantal point process on $\left\{1,\ldots,m\right\}\times {\mathfrak{X}}$. The correlation kernel of this determinantal point process, $K_{n,m}(r,x;s,y)$ (where $r,s\in\left\{1,\ldots,m\right\}$, and $x, y\in {\mathfrak{X}})$, is given by the formula $$K_{n,m}(r,x;s,y)=-\phi_{r,s}(x,y)+\sum\limits_{i,j=1}^n\phi_{r,m+1}(x,x_i^{m+1})\left(A^{-1}\right)_{i,j}\phi_{0,s}(x_j^0,y).$$ The additional transition functions $\phi_{r,s}$ with $s\neq r+1$, and the matrix $A=(a_{i,j})$, with $i,j=1,\ldots,n$, are defined as follows in terms of the one-step transition functions $\phi_{r,r+1}$, with $r=0,1,\ldots,m$, of point process : $$\label{trans-def} \phi_{r,s}(x,y)=\left\{ \begin{array}{ll} \left(\phi_{r,r+1}\ast\ldots\ast\phi_{s-1,s}\right)(x,y), & 0\leq r<s\leq m+1, \\ 0, & r\geq s, \end{array} \right.$$ and $$a_{i,j}=\phi_{0,m+1}(x_i^0,x_j^{m+1}).$$ The correlation functions defined in can be written as determinants of block matrices, namely $$\begin{split} &\varrho_{k_1,\ldots,k_m}\left(x_1^1,\ldots,x_{k_1}^1;\ldots;x_1^m,\ldots,x_{k_m}^m\right)\\ &=\det\left[\begin{array}{ccc} \left(K_{n,m}(1,x_i^1;1,x_j^1)\right)_{i=1,\ldots,k_1}^{j=1,\ldots,k_1} & \ldots & \left(K_{n,m}(1,x_i^1;m,x_j^m)\right)_{i=1,\ldots,k_1}^{j=1,\ldots,k_m} \\ \vdots & & \\ \left(K_{n,m}(m,x_i^m;1,x_j^1)\right)_{i=1,\ldots,k_m}^{j=1,\ldots,k_1} & \ldots & \left(K_{n,m}(m,x_i^m;m,x_j^m)\right)_{i=1,\ldots,k_m}^{j=1,\ldots,k_m} \end{array} \right], \end{split} \label{rhoK}$$ where $1\leq k_1,\ldots,k_m\leq n$, and for $1\leq l,p\leq m$ $$\left(K_{n,m}(l,x_i^l;p,x_j^p)\right)_{i=1,\ldots,k_l}^{j=1,\ldots,k_p} =\left(\begin{array}{ccc} K_{n,m}(l,x_1^l;p,x_1^p) & \ldots & K_{n,m}(l,x_1^l;p,x_{k_p}^p)\\ \vdots & & \\ K_{n,m}(l,x_{k_l}^l;p,x_1^p) & \ldots & K_{n,m}(l,x_{k_l}^l;p,x_{k_p}^p) \end{array} \right).$$ In what follows the functions $$\phi_{0,s}(i,y),\;\; 2\leq s\leq m,$$ will be called *initial transition functions*, and the functions $$\phi_{r,m+1}(x,j),\;\; 1\leq r\leq m-1,$$ will be called *final transition functions*. In addition, the functions of the form $$\phi_{r,s}(x,y),\;\; 1\leq r\leq m-2,\;\; r+2\leq s\leq m,$$ will be called *intermediate transition functions.* Finally, the function $$\phi_{0,m+1}(i,j)$$ will be called *the total transition function*. In order to prove Theorem \[TheoremMostGerneralCorrelationKernel\] we need to rewrite the density of the product matrix process obtained in Theorem \[TheoremGinibreCouplingDensity\] as in the formulation of the Eynard-Mehta Theorem, see equation (\[ProductDeterminantsMeasure\]), and to obtain explicit expressions for the relevant transition functions. This is done below.\ $\bullet$ **One-step transition functions.** Recall that $\nu_0=\nu_m=0$. In our situation ${\mathfrak{X}}_0=\left\{1,\ldots,n\right\}$, ${\mathfrak{X}}_{m+1}=\left\{1,\ldots,n\right\}$, ${\mathfrak{X}}={\mathbb R}_{>0}$, and $d\mu$ is the Lebesgue measure on ${\mathbb R}_{>0}$. The initial given one-step transition function is defined by $$\label{phi01} \phi_{0,1}: \left\{1,\ldots,n\right\}\times{\mathbb R}_{>0}\rightarrow{\mathbb R}_{>0};\;\;\; \phi_{0,1}(i,x)=x^{\nu_1+i-1}e^{-x}.$$ The final given one-step transition function is defined by $$\label{phimm+1} \phi_{m,m+1}: {\mathbb R}_{>0}\times\left\{1,\ldots,n\right\}\rightarrow{\mathbb R}_{>0};\;\;\;\phi_{m,m+1}(x,k)=x^{\frac{k-1}{2}}I_{k-1}\left(2bx^{\frac{1}{2}}\right)e^{-V_m(x)}.$$ In addition, the intermediate given one-step transition functions $$\phi_{r,r+1}: {\mathbb R}_{>0}\times{\mathbb R}_{>0}\rightarrow{\mathbb R}_{>0},\;\;\; r=1,\ldots,m-1,$$ are defined by $$\label{phirr+1} \phi_{r,r+1}(x,y)=\left(\frac{y}{x}\right)^{\nu_{r+1}}\frac{e^{-\frac{y}{x}-V_{r}(x)}}{x},\;\;\; r=1,\ldots,m-1.$$ $\bullet$ **Initial transition functions.** The initial transition functions, $\phi_{0,s}(i,y)$, with $2\leq s\leq m$, can be written as $$\begin{split} \phi_{0,s}(i,y) &=\phi_{0,1}\ast\phi_{1,2}\ast\ldots\ast\phi_{s-2,s-1}\ast\phi_{s-1,s}(i,y)\\ &=\int\limits_0^{\infty}\ldots\int\limits_0^{\infty} \phi_{0,1}(i,t_1)\phi_{1,2}(t_1,t_2)\ldots\phi_{s-2,s-1}(t_{s-2},t_{s-1})\phi_{s-1,s}(t_{s-1},y)dt_1 \ldots dt_{s-1}, \end{split}$$ spelling out the convolution in the last line. Inserting the corresponding expressions for the one-step transition functions and we obtain $$\begin{split} \phi_{0,s}(i,y) =&\int\limits_0^{\infty}\ldots\int\limits_0^{\infty} t_1^{\nu_1+i-1}\left(\frac{t_2}{t_1}\right)^{\nu_2}\ldots\left(\frac{t_{s-1}}{t_{s-2}}\right)^{\nu_{s-1}} \left(\frac{y}{t_{s-1}}\right)^{\nu_s} \\ &\times e^{-t_1-\frac{t_2}{t_1}-\ldots-\frac{t_{s-1}}{t_{s-2}}-\frac{y}{t_{s-1}}- V_1\left(t_1\right)-\ldots -V_{s-1}\left(t_{s-1}\right)} \frac{dt_1}{t_1} \ldots \frac{dt_{s-1}}{t_{s-1}}, \end{split}$$ where $2\leq s\leq m$ (for $s=1$ see ).\ $\bullet$ **Final transition functions.** The final transition functions $\phi_{r,m+1}(x,j)$, with $1\leq r\leq m-1$, can be written in terms of the one-step transition functions as $$\begin{split} &\phi_{r,m+1}(x,j)\\ &=\phi_{r,r+1}\ast\phi_{r+1,r+2}\ast\ldots\ast\phi_{m-1,m}\ast\phi_{m,m+1}(x,j)\\ &=\int\limits_0^{\infty}\ldots\int\limits_0^{\infty} \phi_{r,r+1}(x,t_{r+1})\phi_{r+1,r+2}(t_{r+1},t_{r+2})\ldots\phi_{m-1,m}(t_{m-1},t_{m})\phi_{m,m+1}(t_{m},j)dt_{r+1} \ldots dt_{m}. \end{split}$$ Inserting the explicit formulae for the one-step transition functions and we find $$\label{phirm+1} \begin{split} \phi_{r,m+1}(x,j) =&\frac{e^{-V_r(x)}}{x}\int\limits_0^{\infty}\ldots\int\limits_0^{\infty} \left(\frac{t_{r+1}}{x}\right)^{\nu_{r+1}} \left(\frac{t_{r+2}}{t_{r+1}}\right)^{\nu_{r+2}} \ldots \left(\frac{t_m}{t_{m-1}}\right)^{\nu_{m}} \left(t_m\right)^{\frac{j-1}{2}}I_{j-1}\left(2b\left(t_m\right)^{\frac{1}{2}}\right) \\ &\times e^{-\frac{t_{r+1}}{x}-\frac{t_{r+2}}{t_{r+1}}-\ldots -\frac{t_m}{t_{m-1}}-V_{r+1}\left(t_{r+1}\right) \ldots -V_{m}\left(t_{m}\right)} \frac{dt_{r+1}}{t_{r+1}} \ldots \frac{dt_{m-1}}{t_{m-1}}dt_m, \end{split}$$ where $1\leq r\leq m-1$ (for $r=m$ see ). $\bullet$ **Intermediate transition functions.** Recall that the intermediate transition functions $\phi_{r,s}(x,y)$, $1\leq r\leq m-2$, $r+2\leq s\leq m$ are defined by $$\begin{split} \phi_{r,s}(x,y) &=\phi_{r,r+1}\ast\phi_{r+1,r+2}\ast\ldots\ast\phi_{s-1,s}(x,y)\\ &=\int\limits_0^{\infty}\ldots\int\limits_0^{\infty} \phi_{r,r+1}(x,t_{r+1})\phi_{r+1,r+2}(t_{r+1},t_{r+2})\ldots\phi_{s-1,s}(t_{s-1},y)dt_{r+1} \ldots dt_{s-1}, \end{split}$$ where $1\leq r\leq m-2$, $r+2\leq s\leq m$. Using the explicit formulae for the intermediate one-step transition functions we get $$\begin{split} \phi_{r,s}(x,y) =&\frac{e^{-V_r(x)}}{x}\int\limits_0^{\infty}\ldots\int\limits_0^{\infty} \left(\frac{t_{r+1}}{x}\right)^{\nu_{r+1}} \left(\frac{t_{r+2}}{t_{r+1}}\right)^{\nu_{r+2}} \ldots \left(\frac{y}{t_{s-1}}\right)^{\nu_{s}} \\ &\times e^{-\frac{t_{r+1}}{x}-\frac{t_{r+2}}{t_{r+1}}-\ldots-\frac{t_{s-1}}{t_{s-2}}-\frac{y}{t_{s-1}}-V_{r+1}\left(t_{r+1}\right) \ldots-V_{s-1}\left(t_{s-1}\right)} \frac{dt_{r+1}}{t_{r+1}} \ldots \frac{dt_{s-1}}{t_{s-1}}, \end{split}$$ where $1\leq r\leq m-2$, $r+2\leq s\leq m$ (for $s=r+1$ see ).\ $\bullet$ **Total transition function.** The total transition function $ \phi_{0,m+1}(i,j)=a_{i,j} $ that constitutes matrix $A$ can be written as the convolution of the one-step initial transition function $\phi_{0,1}$ from and the final transition function $\phi_{1,m+1}$ from for $r=1$: $$\phi_{0,m+1}(i,j)=\phi_{0,1}\ast\phi_{1,m+1}(i,j)=\int\limits_0^{\infty} \phi_{0,1}(i,t_{1})\phi_{1,m+1}(t_{1},j)dt_{1}.$$ This gives $$\begin{split} \phi_{0,m+1}(i,j) &=\int\limits_0^{\infty}\ldots\int\limits_0^{\infty} t_1^{\nu_1+i-1}\left(\frac{t_2}{t_1}\right)^{\nu_2}\ldots \left(\frac{t_m}{t_{m-1}}\right)^{\nu_m} \left(t_m\right)^{\frac{j-1}{2}}I_{j-1}\left(2b\left(t_m\right)^{\frac{1}{2}}\right) \\ &\times e^{-t_1-\frac{t_2}{t_1}-\ldots -\frac{t_m}{t_{m-1}} -V_1\left(t_1\right) -\ldots -V_{m}\left(t_{m}\right)} \frac{dt_1}{t_1} \ldots \frac{dt_{m-1}}{t_{m-1}} dt_m. \end{split}$$ Once all the transition functions are written explicitly, Theorem \[TheoremMostGerneralCorrelationKernel\] follows immediately from Theorem \[TheoremEynardMehta\]. Double contour integral representation for the correlation kernel ================================================================= In this section we consider the $m$-matrix model defined by probability measure (\[MainProbabilityMeasure1\]). The correlation kernel of the product matrix process associated with this model is denoted by $K_{n,m}\left(r,x;s,y;b\right)$. Our aim is to derive a double contour integral representation for $K_{n,m}\left(r,x;s,y;b\right)$, and to prove Theorem \[TheoremDoubleIntegralRepresentationExactKernel\]. Note that the multi-matrix model defined by probability measure (\[MainProbabilityMeasure\]) turns into that defined by probability measure (\[MainProbabilityMeasure1\]) if $$\label{Potentials} V_1(t)=\ldots=V_{m-2}(t)=0,\;\; V_{m-1}(t)=b^2t,\;\; V_m(t)=0.$$ Therefore, in order to derive a contour integral representation for the correlation kernel $K_{n,m}\left(r,x;s,y;b\right)$ we can exploit the formulae obtained in Theorem \[TheoremMostGerneralCorrelationKernel\] with the potential functions specified by (\[Potentials\]). \[PropositionKV\] For the specific case $V_1(t)=\ldots= V_{m-2}(t)=0$, $V_{m-1}(t)=b^2t$, $V_m(t)=0$ the correlation kernel $K_{n,m}(r,x;s,y;b)$ can be written as $$\label{CorrelationKernelGeneralFormulaGeneral} K_{n,m}(r,x;s,y;b)=-\phi_{r,s}(x,y;b)+\sum\limits_{i,j=1}^n\phi_{r,m+1}(x,i)\left(A^{-1}\right)_{i,j}\phi_{0,s}(j,y).$$ Here, the three sets of functions are obtained as follows. *(i)* The functions $\phi_{r,s}(x,y;b)$ are given by $$\begin{split} &\phi_{r,s}(x,y;b)\\ &=\left\{ \begin{array}{lll} \frac{e^{-\frac{y}{x}-b^2x}}{x}, & r=m-1, s=m,\\ \frac{1}{x} G^{s-r,0}_{0,s-r}\left(\begin{array}{cccc} - \\ \nu_{r+1},\ldots,\nu_{s} \end{array}\biggl|\frac{y}{x}\right), & 1\leq r<s\leq m-1,\\ \frac{1}{x} \int\limits_{0}^{\infty} G^{m-r-1,0}_{0,m-r-1}\left(\begin{array}{cccc} - \\ \nu_{r+1},\ldots,\nu_{m-1} \end{array}\biggl|\frac{t}{x}\right) e^{-\frac{y}{t}-b^2t}\frac{dt}{t}, & 1\leq r\leq m-2, s=m,\\ 0, & 1\leq s\leq r\leq m.\\ \end{array} \right. \end{split} \nonumber$$ *(ii)* The functions $\phi_{r,m+1}(x,i)$ are given by $$\phi_{r,m+1}(x,j)= \left\{ \begin{array}{ll} b^{j-1}x^{j-1}\Gamma(j+\nu_{r+1})\ldots\Gamma(j+\nu_{m-1}), & r\in\left\{1,\ldots,m-2\right\},\\ b^{j-1}x^{j-1}, & r=m-1, \\ x^{\frac{j-1}{2}}I_{j-1}\left(2bx^{\frac{1}{2}}\right), & r=m. \end{array} \right. \nonumber$$ *(iii)* The functions $\phi_{0,s}(i,y)$ are given by $$\label{phi0siy} \phi_{0,s}(i,y)=\left\{ \begin{array}{ll} G^{s,0}_{0,s}\left(\begin{array}{cccc} - \\ \nu_1+i-1, \nu_{2},\ldots,\nu_{s} \end{array}\biggl|y\right)& s\in\left\{1,\ldots,m-1\right\}\\ \int\limits_{0}^{\infty} G^{m-1,0}_{0,m-1}\left(\begin{array}{cccc} - \\ \nu_1+i-1, \nu_{2},\ldots,\nu_{m-1} \end{array}\biggl|t\right) e^{-\frac{y}{t}-b^2t}\frac{dt}{t} & s=m.\\ \end{array} \right.$$ Finally, for the matrix $A=\left(a_{i,j}\right)_{i,j=1}^n$, we have $$\label{aij} a_{i,j}=\phi_{0,m+1}(i,j)=b^{j-1}\Gamma(i+j-1+\nu_{1})\Gamma(j+\nu_2)\ldots\Gamma(j+\nu_{m-1}).$$ The formulae for the transition functions stated in Proposition \[PropositionKV\] can be obtained by straightforward calculations starting from the formulae obtained in Theorem \[TheoremMostGerneralCorrelationKernel\]. In the calculations we have exploited the following integral representation of the Meijer $G$-function, cf. [@AkemannIpsenKieburg], $$\label{GElementaryRepresentation} \begin{split} &G^{m,0}_{0,m}\left(\begin{array}{ccc} - \\ b_1,\ldots,b_m \end{array}\biggl|\frac{x_m}{x_0}\right)\\ &=\int\limits_0^{\infty}\ldots\int\limits_0^{\infty}\left(\frac{x_1}{x_0}\right)^{b_1}\left(\frac{x_2}{x_1}\right)^{b_2}\ldots\left(\frac{x_m}{x_{m-1}}\right)^{b_m} e^{-\frac{x_1}{x_0}-\frac{x_2}{x_1}-\ldots-\frac{x_m}{x_{m-1}}}\frac{dx_1}{x_1}\ldots\frac{dx_{m-1}}{x_{m-1}}, \end{split}$$ the fact that the Mellin transform of a Meijer $G$-function is given by $$\begin{split} \int\limits_0^{\infty}t^{u-1}G^{m,n}_{p,q}\left(\begin{array}{ccc} a_1,\ldots, a_p \\ b_1,\ldots, b_q \end{array}\biggl|tz\right)dt =z^{-u}\frac{\prod_{i=1}^m\Gamma(b_i+u)\prod_{i=1}^n\Gamma(1-a_i-u)}{\prod_{i=n+1}^p\Gamma(a_i+u)\prod_{i=m+1}^q\Gamma(1-b_i-u)}, \end{split}$$ see Luke [@Luke], and the integral involving the modified Bessel functions of the first kind In Proposition \[PropositionKV\] we have found the matrix entries of $A$ explicitly. This enables us to derive a formula for the inverse of $A$. \[PropositionInverse\] Let $C=\left(c_{j,k}\right)_{j,k=1}^n$ be the inverse of $A=\left(a_{i,j}\right)_{i,j=1}^n$, where $a_{i,j}$ is given by equation (\[aij\]). Thus the matrix elements $\left(c_{j,k}\right)_{j,k=1}^n$ are defined by the relation $$\sum\limits_{j=1}^na_{i,j}c_{j,k}=\delta_{i,k},\;\;\; i,k\in\left\{1,\ldots,n\right\}.$$ We have[^6] $$\label{cjk} \begin{split} c_{j,k}=&\frac{1}{b^{j-1}\Gamma\left(j+\nu_2\right)\ldots\Gamma\left(j+\nu_{m-1}\right)}\\ &\times\sum\limits_{p=0}^{n-1}\frac{\Gamma\left(\nu_1+p+1\right)}{\Gamma\left(\nu_1+j\right)\Gamma\left(\nu_1+k\right)p!} \frac{\left(-p\right)_{j-1}\left(-p\right)_{k-1}}{\left(\nu_1+1\right)_{j-1}\left(\nu_1+1\right)_{k-1}}\frac{1}{\left(j-1\right)!\left(k-1\right)!}. \end{split}$$ Clearly, the matrix elements $\left(c_{j,k}\right)_{j,k=1}^n$ can be written as $$c_{j,k}=\frac{C_{j,k}}{b^{j-1}\Gamma\left(j+\nu_2\right)\ldots\Gamma\left(j+\nu_{m-1}\right)},$$ where $C_{j,k}$ is defined by $$\sum\limits_{j=1}^n\Gamma\left(i+j-1+\nu_1\right)C_{j,k}=\delta_{i,k},\;\;\;i,k\in\left\{1,\ldots,n\right\}.$$ Next we use the fact that the inverse $\left(\alpha_{k,l}\right)_{k,l=0}^{N-1}$ of the Hankel matrix $$\left(h_{k+l}\right)_{k,l=0}^{N-1},\;\; h_k=\Gamma(k+\nu+1),$$ is given by $$\alpha_{k,l}=\sum\limits_{p=0}^{N-1} \frac{\Gamma(\nu+p+1)(-p)_{k}(-p)_{l}}{p!\Gamma(\nu_1+k+1)\Gamma(\nu_1+l+1)k!l!}, $$ see, for example, Akemann and Strahov [@AkemannStrahov Section 5]. Equation follows. The fact that the inverse of the matrix $A$ can be written explicitly allows us to spell out the correlation kernel $K_{n,m}(r,x;s,y;b)$ of Proposition \[PropositionKV\]. We start from equation , use Proposition \[PropositionInverse\], and obtain the formula $$\label{FirstFormulaForK} \begin{split} K_{n,m}\left(r,x;s,y;b\right)=-\phi_{r,s}\left(x,y;b\right)+\sum\limits_{p=0}^{n-1}\frac{\Gamma\left(\nu_1+p+1\right)}{\Gamma^2(\nu_1+1)p!} P_{r,p}(x)Q_{s,p}(y), \end{split}$$ where the transition functions $\phi_{r,s}\left(x,y\right)$ are defined in Proposition \[PropositionKV\] (i), $$P_{r,p}(x)=\sum\limits_{i=0}^p\phi_{r,m+1}\left(x,i+1\right)\frac{(-p)_i}{\Gamma\left(i+\nu_2+1\right)\ldots\Gamma\left(i+\nu_{m-1}+1\right) (\nu_1+1)_ii!b^{i}}$$ and $$\label{QSP} Q_{s,p}(y)=\sum\limits_{j=0}^p\phi_{0,s}\left(j+1,y\right)\frac{\left(-p\right)_j}{\left(\nu_1+1\right)_jj!}.$$ Note that the Pochhammer symbol truncates both sums in $P_{r,p}(x)$ and $Q_{s,p}(y)$ that initially ran up to $n-1$. Proposition \[PropositionKV\] gives the transition functions $\phi_{r,m+1}\left(x,i+1\right)$ and $\phi_{0,s}\left(j+1,y\right)$ explicitly. This enables us to obtain different, useful formulae for the functions $P_{r,p}(x)$ and $Q_{s,p}(y)$, involved in expression (\[FirstFormulaForK\]) for the correlation kernel $K_{n,m}\left(r,x;s,y;b\right)$.\ $\bullet$ **The functions** $P_{r,p}(x)$. Let us derive a contour integration formula for the functions $P_{r,p}(x)$. If $r\in\left\{1,\ldots,m-1\right\}$, we have $$\begin{split} P_{r,p}(x)&=\Gamma\left(\nu_1+1\right) \sum\limits_{i=0}^p\frac{x^i\left(-p\right)_i}{i!\Gamma\left(\nu_1+i+1\right) \ldots\Gamma\left(\nu_{r}+i+1\right)}\\ &=(-1)^p\Gamma\left(\nu_1+1\right)p!\sum\limits_{i=0}^p \frac{(-1)^{p-i}x^i}{i!(p-i)!\Gamma\left(\nu_1+i+1\right) \ldots\Gamma\left(\nu_r+i+1\right)}. \end{split}$$ For $r=m$, the formula for the function $P_{m,p}(x)$ is $$\label{Pmn} \begin{split} P_{m,p}(x)&=\Gamma\left(\nu_1+1\right) \sum\limits_{i=0}^p\frac{x^{\frac{i}{2}}\left(-p\right)_i}{\Gamma\left(\nu_1+i+1\right) \ldots\Gamma\left(\nu_{m-1}+i+1\right)}\frac{I_i\left(2bx^{\frac{1}{2}}\right)}{b^ii!}\\ &=(-1)^p\Gamma\left(\nu_1+1\right)p!\sum\limits_{i=0}^p \frac{(-1)^{p-i}x^{\frac{i}{2}}}{(p-i)!\Gamma\left(\nu_1+i+1\right) \ldots\Gamma\left(\nu_{m-1}+i+1\right)}\frac{I_i\left(2bx^{\frac{1}{2}}\right)}{b^ii!}. \end{split}$$ Using the expressions for the functions $P_{r,p}(x)$ just written above together with the Residue Theorem we immediately obtain a contour integral representation for these functions. Namely, for $r\in\left\{1,\ldots,m-1\right\}$ we find $$\label{PINTEGRAL1} P_{r,p}(x)=\frac{(-1)^p\Gamma\left(\nu_1+1\right)p!}{2\pi i} \oint\limits_{\Sigma_p}\frac{\Gamma\left(t-p\right)}{\prod_{j=0}^r\Gamma\left(t+\nu_j+1\right)} x^tdt,$$ and for $r=m$ we find $$\label{PINTEGRAL2} P_{m,p}(x)=\frac{(-1)^p\Gamma\left(\nu_1+1\right)p!}{2\pi i} \oint\limits_{\Sigma_p}\frac{\Gamma\left(t-p\right)x^{\frac{t}{2}}}{\prod_{j=0}^{m-1}\Gamma\left(t+\nu_j+1\right)} \frac{I_t\left(2bx^{\frac{1}{2}}\right)}{b^t}dt.$$ In the formulae just written above $\Sigma_p$ denotes a closed contour encircling $0,1,\ldots,p$ in positive direction.\ $\bullet$ **The functions $Q_{s,p}(y)$.** For $s\in\left\{1,\ldots,m-1\right\}$ the formula for the functions $Q_{s,p}(y)$ reads $$Q_{s,p}(y)= \sum\limits_{j=0}^p\frac{(-p)_jG^{s,0}_{0,s}\left(\begin{array}{cccc} - \\ \nu_1+j,\nu_2\ldots,\nu_s \end{array}\biggl|y\right)}{\left(\nu_1+1\right)_jj!}.$$ The contour integral representation of the Meijer $G$-function above is $$\label{ci} G^{s,0}_{0,s}\left(\begin{array}{cccc} - \\ \nu_1+j,\nu_2\ldots,\nu_s \end{array}\biggl|y\right)=\frac{1}{2\pi i}\int\limits_{c -i\infty}^{c +i\infty} \Gamma\left(u+\nu_1+j\right)\Gamma\left(u+\nu_2\right)\ldots\Gamma\left(u+\nu_s\right) y^{-u}du,$$ for $c>0$ and $\nu_j\geq0$ $\forall j$, leaving the poles of the Gamma-functions to the left of the contour. Therefore, we can write $$\begin{split} &Q_{s,p}(y)= \frac{1}{2\pi i}\int\limits_{c-i\infty}^{c+i\infty} \Gamma\left(u+\nu_1\right)\ldots\Gamma\left(u+\nu_s\right) \left(\sum\limits_{j=0}^p\frac{\left(-p\right)_j\left(u+\nu_1\right)_j}{\left(1+\nu_1\right)_jj!} \right) y^{-u}du. \end{split}$$ The sum inside the integral above can be written as Gauss’ hypergeometric function, $$\sum\limits_{j=0}^p\frac{\left(-p\right)_j\left(u+\nu_1\right)_j}{\left(1+\nu_1\right)_j} \frac{1}{j!}= {}_2F_1\left(-p,u+\nu_1;1+\nu_1;1\right).$$ By the Chu-Vandermonde formula for the Gauss hypergeometric functions (see, for example, Ismail [@Ismail Section 1.4]) $${}_2F_1\left(-p,u+\nu_1;1+\nu_1;1\right)=\frac{\left(1-u\right)_p}{\left(1+\nu_1\right)_p} =\left(-1\right)^p\frac{\Gamma(u)}{\Gamma(u-p)}\frac{1}{\left(1+\nu_1\right)_p}.$$ So we find $$\label{QINTEGRAL1} \begin{split} Q_{s,p}(y)=\frac{(-1)^p\Gamma(1+\nu_1)}{2\pi i\Gamma\left(1+\nu_1+p\right)} \int\limits_{c-i\infty}^{c+i\infty} \frac{\prod\limits_{j=0}^s\Gamma\left(u+\nu_j\right) }{\Gamma(u-p)}y^{-u}du, \end{split}$$ where $s\in\left\{1,\ldots,m-1\right\}$. To obtain an integral representation for the function $Q_{m,p}(y)$ we proceed as follows. Recalling the definition $$\label{Krepresentation} \int\limits_0^{+\infty}e^{-\frac{\beta}{t}-\gamma t}t^{\nu-1}dt= 2\left(\frac{\beta}{\gamma}\right)^{\frac{\nu}{2}}K_{\nu}\left(2\sqrt{\beta\gamma}\right), \;\;\; \beta>0,\;\gamma>0,$$ together with the contour integral representation for the corresponding Meijer $G$-function equation (\[ci\]), applying Fubini’s Theorem we obtain $$\begin{split} &\phi_{0,m}\left(j+1,y\right)=\frac{1}{2\pi i}\int\limits_{c-i\infty}^{c+i\infty} \Gamma\left(u+\nu_1+j\right)\Gamma\left(u+\nu_2\right)\ldots\Gamma\left(u+\nu_{m-1}\right) {2b^uy^{\frac{u}{2}}K_u\left(2by^{\frac{1}{2}}\right)}{y^{-u}}du. \end{split}$$ Inserting this into the formula for $Q_{m,p}(y)$, equation (\[QSP\]), gives $$\begin{split} &Q_{m,p}(y)=\frac{1}{2\pi i}\int\limits_{c-i\infty}^{c+i\infty} \prod\limits_{j=1}^{m-1}\Gamma\left(u+\nu_j\right) {}_2F_1\left(-p,u+\nu_1;1+\nu_1;1\right) {2b^uy^{-\frac{u}{2}}K_u\left(2by^{\frac{1}{2}}\right)} du. \end{split}$$ By the same arguments as above, this formula can be rewritten as $$\label{QINTEGRAL2} \begin{split} Q_{m,p}(y)=\frac{(-1)^p\Gamma(1+\nu_1)}{2\pi i\Gamma\left(1+\nu_1+p\right)} \int\limits_{c-i\infty}^{c+i\infty} \frac{\prod_{j=0}^m\Gamma\left(u+\nu_j\right) }{\Gamma(u-p)} \frac{2b^uy^{-\frac{u}{2}}K_u\left(2by^{\frac{1}{2}}\right)}{\Gamma(u)}du. \end{split}$$ Now we are ready to prove Theorem \[TheoremDoubleIntegralRepresentationExactKernel\]. In order to derive the formula for the correlation kernel stated in the Theorem \[TheoremDoubleIntegralRepresentationExactKernel\] it is enough to represent the sum $$S_{n,m}(r,x;s,y;b)=\sum\limits_{p=0}^{n-1}\frac{\Gamma\left(\nu_1+p+1\right)}{\Gamma^2\left(\nu_1+1\right)p!} P_{r,p}(x)Q_{s,p}(y)$$ as a double contour integral. For this purpose use the contour integral representations for the functions $P_{r,p}(x)$ and $Q_{s,p}(y)$ obtained above, equations (\[PINTEGRAL1\]), (\[PINTEGRAL2\]), (\[QINTEGRAL1\]), and (\[QINTEGRAL2\]), and the formula derived in [@KuijlaarsZhang Eq.5.3]: $$\label{SumGammaFormula} \sum\limits_{p=0}^{n-1}\frac{\Gamma(t-p)}{\Gamma(u-p)}=\frac{1}{u-t-1}\left[\frac{\Gamma(t-n+1)}{\Gamma(u-n)} -\frac{\Gamma(t+1)}{\Gamma(u)}\right].$$ Then note that by the Residue Theorem the second term in the right hand side of equation (\[SumGammaFormula\]) does not contribute to $S_{n,m}(r,x;s,y;b)$. Here, we choose $c=\frac12$, and due to the shift in argument of the function $q_s(u,y;b)$ in equation this leads to the contour as stated in Theorem \[TheoremDoubleIntegralRepresentationExactKernel\]. We would like to point out that for the special case $m=2$ and $r=s=2$ the kernel $K_{n,m}(r,x;s,y;b)$ we just determined in can be shown to agree with the correlation kernel $K_N(x,y)$ obtained by the authors in [@AkemannStrahov Theorem 3.2], for a particular choice of $b$. This agreement is not surprising, in view of the identification made after . Formula for our kernel gives $$\label{A0} K_{n,m=2}(r=2,x;s=2,y;b)=\sum\limits_{p=0}^{n-1}\frac{\Gamma\left(\nu_1+p+1\right)}{\Gamma^2\left(\nu_1+1\right)p!} P_{r=2,p}(x)Q_{s=2,p}(y).$$ Equation implies that the function $P_{r=2,p}(x)$ can be written as $$\label{A1p} P_{r=2,p}(x)=\sum\limits_{j=0}^p\frac{(-p)_j}{\left(\nu_1+1\right)_{j}}\frac{x^{\frac{j}{2}}I_j\left(2bx^{\frac{1}{2}}\right)}{b^jj!}.$$ Now let us find a convenient representation for the function $Q_{s=2,p}(x)$. Equation reads $$\label{A2q} Q_{s=2,p}(x)=\sum\limits_{j=0}^p\phi_{0,s=2}(j+1,y)\frac{(-p)_j}{\left(\nu_1+1\right)_jj!}.$$ The function $\phi_{0,s=2}(j+1,y)$ was obtained in Proposition \[PropositionKV\], see equation , and take into account that $s=m=2$. Using the representation of the exponential function in terms of the Meijer $G$-function $$\label{Gexp} x^{\nu} e^{-\frac{y}{x}}=G_{0,1}^{1,0}\left(\begin{array}{c} - \\ \nu \end{array} \biggl|{x}\right),$$ we have $$\label{A3} \phi_{0,s=2}(j+1,y)=\int\limits_{0}^{\infty}G_{0,1}^{1,0}\left(\begin{array}{c} - \\ \nu_1+j \end{array} \biggl|t\right)e^{-\frac{y}{t}-b^2t}\frac{dt}{t} =2\left(\frac{y}{b^2+1}\right)^{\frac{\nu_1+j}{2}}K_{\nu_1+j}\left(2\sqrt{\left(b^2+1\right)y}\right).$$ Here, we have used . If we insert (\[A3\]) in formula (\[A2q\]), we find $$\label{A5} Q_{s=2,p}(x)=2\sum\limits_{j=0}^p\frac{(-p)_j}{\left(\nu_1+1\right)_jj!}\left(\frac{y}{b^2+1}\right)^{\frac{\nu_1+j}{2}} K_{\nu_1+j}\left(2\sqrt{\left(1+b^2\right)y}\right).$$ Thus the kernel $K_{n,2}(2,x;2,y;b)$ is given by equation (\[A0\]), where the functions $P_{2,p}(x)$ and $Q_{2,p}(y)$ are defined by equations (\[A1p\]) and (\[A5\]), respectively. Next, we turn to the kernel $K_N(x,y)$ obtained in [@AkemannStrahov Theorem 3.2]. There, the parameters $\alpha(\mu)$ and $\delta(\mu)$ were defined as $$\alpha(\mu)=\frac{1+\mu}{2\mu},\;\;\;\delta(\mu)=\frac{1-\mu}{2\mu},$$ where $\mu$ takes values in the interval $(0,1]$. We thus have $$\frac{\alpha(\mu)^2-\delta(\mu)^2}{\delta(\mu)}=\frac{2}{1-\mu},\;\;\; \frac{\alpha(\mu)^2-\delta(\mu)^2}{\alpha(\mu)}=\frac{2}{1+\mu},\;\;\; \alpha(\mu)^2-\delta(\mu)^2=\frac{1}{\mu}.$$ Taking this into account, and identifying $\nu_1=\nu$ there, we see that the correlation kernel $K_N(x,y)$ in [@AkemannStrahov Theorem 3.2] can be written as $$K_N(x,y)=\sum\limits_{n=0}^{N-1}P_n(x)Q_n(y),$$ where $$P_n(x)=\frac{(-1)^n(\nu_1+n)!n!}{\sqrt{\mu}}\sum\limits_{k=0}^{n}\frac{(-n)_k}{(\nu_1+k)!k!}\left(\frac{2x^{\frac{1}{2}}}{1-\mu}\right)^k I_k\left(\frac{1-\mu}{\mu}x^{\frac{1}{2}}\right),$$ and $$Q_n(y)=\frac{2(-1)^n}{\sqrt{\mu}(n!)^2}\sum\limits_{l=0}^{n}\frac{(-n)_l}{(\nu_1+l)!l!}\left(\frac{2y^{\frac{1}{2}}}{1+\mu}\right)^{l+\nu_1} K_{l+\nu_1}\left(\frac{1+\mu}{\mu}y^{\frac{1}{2}}\right).$$ Setting $$\label{A6} b=\frac{1-\mu}{2\sqrt{\mu}},\;\;\; x=\frac{\zeta}{\mu},\;\;\; y=\frac{\eta}{\mu},$$ equations (\[A1p\]) and (\[A5\]) take the form $$\label{A11} P_{r=2,p}(x)=\sum\limits_{j=0}^p\frac{(-p)_j}{\left(\nu_1+1\right)_{j}j!} \left(\frac{2\zeta^{\frac{1}{2}}}{1-\mu}\right)^j I_j\left(\frac{1-\mu}{\mu}\zeta^{\frac{1}{2}}\right),$$ and $$\label{A12} Q_{s=2,p}(y)=2\sum\limits_{j=0}^p\frac{(-p)_j}{\left(\nu_1+1\right)_{j}j!} \left(\frac{2\eta^{\frac{1}{2}}}{1+\mu}\right)^{j+\nu_1} K_{j+\nu_1}\left(\frac{1+\mu}{\mu}\eta^{\frac{1}{2}}\right).$$ Here, we used that $(\nu_1+1)_j=(\nu_1+j)!/\nu_1!$. Thus we obtain the following identity $$\mu K_{n,m=2}\left(r=2,\frac{\zeta}{\mu};s=2,\frac{\eta}{\mu};\frac{1-\mu}{2\sqrt{\mu}}\right)=K_{N=n}\left(\zeta,\eta\right).$$ The extra factor of $\mu$ in front of the kernel on the left-hand side is due to the change of variables defined by equation (\[A6\]). Proof of Theorem \[TheoremHardEdgeGinibreCouplingProcess\] ========================================================== Given the double contour integral representation for correlation kernel $K_{n,m}\left(r,x;s,y;b\right)$, we are ready to study the asymptotic of the Ginibre product process with coupling, i.e. the asymptotic of the matrix product process associated with the multi-matrix model (\[MainProbabilityMeasure1\]). Recall that we consider the coupling parameter $b$ as a function of $n$, and investigate the hard edge scaling limit at the origin in different asymptotic regimes. The weak coupling regime {#SectionWCR} ------------------------ In this regime we assume that ${b(n)}/{\sqrt{n}}\rightarrow 0$ as $n\rightarrow\infty$. Let us first establish equation (\[WeakCouplingRegimeLimit\]). Consider the first term in the right-hand side of equation (\[CorrelationKernelFormulaContour\]). We will show that in the weak coupling regime the hard edge scaling limit of the first term is given by $$\label{ftl} \underset{n\rightarrow\infty}{\lim}\frac{1}{n}\phi_{r,s}\left(\frac{x}{n},\frac{y}{n};b(n)\right) =\frac{1}{x}G^{s-r,0}_{0,s-r}\left( \begin{array}{c} - \\ \nu_{r+1},\ldots,\nu_s \end{array}\biggl|\frac{y}{x} \right)\textbf{1}_{s>r}.$$ The explicit formula for $\phi_{r,s}\left(x,y;b\right)$ is given in the statement of Theorem \[TheoremDoubleIntegralRepresentationExactKernel\]. Using this explicit formula, we immediately see that equation (\[ftl\]) holds true for $1\leq r<s\leq m-1$, and for $1\leq s\leq r\leq m$. For $r=m-1$, $s=m$ the left hand side of equation (\[ftl\]) is equal to $\frac{1}{x}e^{-\frac{y}{x}}$, which can be rewritten as $$\frac{1}{x}G^{1,0}_{0,1}\left( \begin{array}{c} - \\ \nu_m \end{array}\biggl|\frac{y}{x} \right)=x^{\nu_m-1} e^{-\frac{y}{x}}$$ (recall that $\nu_m=0$). Therefore, equation (\[ftl\]) holds true for $r=m-1$, $s=m$ as well. For $1\leq r\leq m-2$, $s=m$ we can write $$\begin{split} \frac{1}{n}\phi_{r,s}\left(\frac{x}{n},\frac{y}{n};b(n)\right) =\frac{1}{x}\int\limits_0^{\infty}G^{m-r-1,0}_{0,m-r-1}\left( \begin{array}{c} - \\ \nu_{r+1},\ldots,\nu_{m-1} \end{array}\biggl|\frac{nt}{x} \right)e^{-\frac{y}{nt}-b^2(n)t}\frac{dt}{t}. \end{split} $$ Changing the integration variable as $t=\frac{\tau}{n}$, and taking into account that $e^{-\frac{b^2(n)}{n}\tau}\rightarrow 1$, in the weak coupling regime we obtain that equation (\[ftl\]) holds true for $1\leq r\leq m-2$, $s=m$. In this calculations the procedure of taking the limit inside the integral can be justified by the dominated convergence theorem. It remains to show that $$\label{WeakCouplingRegimeLimit1} \begin{split} &\underset{n\rightarrow\infty}{\lim}\left\{\frac{1}{n}S_{n,m}\left(r,\frac{x}{n};s,\frac{y}{n};b(n)\right)\right\}\\ &=\frac{1}{(2\pi i)^2}\int\limits_{-\frac{1}{2}-i\infty}^{-\frac{1}{2}+i\infty} du\oint\limits_{\Sigma_{\infty}}dt\frac{\prod_{j=0}^s\Gamma(u+\nu_j+1)}{\prod_{j=0}^r\Gamma(t+\nu_j+1)} \frac{\sin\pi u}{\sin\pi t} \frac{x^ty^{-u-1}}{u-t}, \end{split}$$ where $\Sigma_{\infty}$ is a contour starting from $+\infty$ in the upper half plane and returning to $+\infty$ in the lower half plane, leaving $-\frac{1}{2}$ on the left, and encircling $\left\{0,1,2,\ldots\right\}$. To see that equation (\[WeakCouplingRegimeLimit1\]) indeed holds true observe that for $r,s\in\left\{1,\ldots,m-1\right\}$ the functions $p_r\left(t,\frac{x}{n};b(n)\right)$ and $q_s\left(u,\frac{y}{n};b(n)\right)$ are both identically equal to $1$, and that for $r,s\in\left\{1,\ldots,m-1\right\}$ equation (\[WeakCouplingRegimeLimit\]) can be obtained by applying the same arguments as in Kuijlaars and Zhang [@KuijlaarsZhang Section 5.2.]. In addition, we have $$p_m\left(t,\frac{x}{n};b(n)\right)=\frac{\Gamma(t+1) I_t\left(2\frac{b(n)}{\sqrt{n}}x^{\frac{1}{2}}\right)}{\left(\frac{b(n)}{\sqrt{n}}x^{\frac{1}{2}}\right)^t}\simeq 1; \mbox{as}\; n\rightarrow\infty, \; \mbox{and}\;\frac{b(n)}{n^{\frac{1}{2}}}\rightarrow 0,$$ and $$q_m\left(u+1,\frac{y}{n};b(n)\right)=\frac{2\left(\frac{b(n)}{\sqrt{n}}y^{\frac{1}{2}}\right)^u K_u\left(2\frac{b(n)}{\sqrt{n}}y^{\frac{1}{2}}\right)}{\Gamma(u)}\simeq 1; \mbox{as}\; n\rightarrow\infty, \; \mbox{and}\;\frac{b(n)}{n^{\frac{1}{2}}}\rightarrow 0.$$ Therefore, equation (\[WeakCouplingRegimeLimit1\]) remains true for all $r,s\in\left\{1,\ldots,m\right\}$. In these calculations the procedure of taking the limit inside the double integral can be justified as in the proof of Theorem 5.3 in Kujlaars and Zhang [@KuijlaarsZhang] using the dominated convergence theorem, and the asymptotic properties of the involved Gamma and Bessel functions. Now, the correlation kernel of the scaled Ginibre product process with coupling, formed by the configurations $\left(ny_1^m,\ldots,ny_n^m;\ldots,ny_1^1,\ldots,ny_n^1\right)$, is related to the correlation kernel of the unscaled process with configurations $\left(y_1^m,\ldots,y_n^m;\ldots,y_1^1,\ldots,y_n^1\right)$ by replacing $K_{n,m}\left(r,x;s,y;b(n)\right)$ with $ \frac{1}{n}K_{n,m}\left(r,\frac{x}{n};s,\frac{y}{n};b(n)\right). $ Taking this into account, we see that equation (\[WeakCouplingRegimeLimit\]) implies the statement of Theorem \[TheoremHardEdgeGinibreCouplingProcess\] in the weak coupling regime. The intermediate coupling regime -------------------------------- The proof of Theorem \[TheoremHardEdgeGinibreCouplingProcess\] for the intermediate coupling regime is based on arguments similar to those presented in Section \[SectionWCR\]. The difference is that in the intermediate coupling regime we have for the transition functions that depend on $b(N)$, that is for $r=m$ or $s=m$: $$\begin{split} \underset{n\rightarrow\infty}{\lim}\left(\frac{1}{n}\phi_{r,s}\left(\frac{x}{n},\frac{y}{n};b(n)\right)\right) =\phi_{r,s}\left(x,y;\alpha\right). \end{split} \nonumber$$ This can be seen immediately from the formulae for $\phi_{r,s}\left(x,y;b(n)\right)$ in the statement of Theorem \[TheoremDoubleIntegralRepresentationExactKernel\] when $n\rightarrow\infty$ and $\frac{b(n)}{n^{\frac{1}{2}}}\rightarrow\alpha$, as well as $$\begin{split} &p_m\left(t,\frac{x}{n};b(n)\right)\rightarrow p_m\left(t,x;\alpha\right),\\ &q_m\left(u,\frac{y}{n};b(n)\right)\rightarrow q_m\left(u,y;\alpha\right). \end{split} \nonumber$$ using and . The strong coupling regime -------------------------- Let us first establish the limiting relations for the correlation kernel in the strong coupling regime, equations (\[SKR2\])-(\[SKR4\]). The kernels with levels $1\leq r,s\leq m-1$ are $b$-independent, and the limit was already established in . In order to establish equations (\[SKR2\])-(\[SKR4\]) we need to take into account that $p_m(t,x;b)$ and $q_m(u,y;b)$ are nontrivial expressions involving the modified Bessel functions of the first and of the second kind, see equations (\[Functionp\]), (\[Functionq\]). These functions have the following asymptotic as ${b(n)}/{n^{\frac{1}{2}}}\rightarrow\infty$: $$p_m\left(t,\frac{b^2(n)x^2}{n^2};b(n)\right)\sim\frac{\Gamma(t+1)}{2\pi^{\frac{1}{2}}} \frac{e^{2\frac{b^2(n)}{n}x}}{\left(\frac{b^2(n)}{n}x\right)^{t+\frac{1}{2}}},$$ and $$q_m\left(u+1,\frac{b^2(n)y^2}{n^2};b(n)\right)\sim\frac{\pi^{\frac{1}{2}}}{\Gamma(u+1)} \frac{b^{2u+1}(n)}{n^{u+\frac{1}{2}}}y^{u+\frac{1}{2}} e^{-2\frac{b^2(n)}{n}y}.$$ The asymptotic formulae just written above follow immediately from the known asymptotic of the modified Bessel function of the first kind $I_t\left(2bx^{\frac{1}{2}}\right)$, and from the known asymptotic of the modified Bessel function of the second kind $K_u\left(2by^{\frac{1}{2}}\right)$ as $b\rightarrow\infty$, see equations (\[A1\]) and (\[A2\]), respectively. In addition, we have the following ratio asymptotic of Gamma functions $$\frac{\Gamma(t-n+1)}{\Gamma(u-n+1)}=\frac{\sin\pi u}{\sin\pi t}n^{t-u}\left(1+O\left(n^{-1}\right)\right),\;\;\mbox{as}\;\; n\rightarrow\infty.$$ Using the asymptotic formulae mentioned above we obtain the following limiting relations.\ $\bullet$ For $1\leq r\leq m-1$, $s=m$ we have $$\label{SKR22} \begin{split} &\underset{n\rightarrow\infty}{\lim}\frac{2^{\frac{1}{2}}b(n)y^{\frac{1}{2}}}{n^{\frac{3}{2}}} S_{n,m}\left(r,\frac{x}{n};m,\frac{b^2(n)y^2}{n^2};b(n)\right)\frac{e^{2\frac{b^2(n)}{n}y}}{2^{\frac{1}{2}}\pi^{\frac{1}{2}}}\\ &=\frac{1}{(2\pi i)^2}\int\limits_{-\frac{1}{2}-i\infty}^{-\frac{1}{2}+i\infty} du\oint\limits_{\Sigma_{\infty}}dt\frac{\prod_{j=0}^{m-1}\Gamma(u+\nu_j+1)}{\prod_{j=0}^r\Gamma(t+\nu_j+1)} \frac{\sin\pi u}{\sin\pi t} \frac{x^ty^{-u-1}}{u-t}. \end{split}$$ $\bullet$ For $r=m$, $1\leq s\leq m-1$ we have $$\label{SKR33} \begin{split} &\underset{n\rightarrow\infty}{\lim}\frac{2^{\frac{1}{2}}b(n)x^{\frac{1}{2}}}{n^{\frac{3}{2}}} S_{n,m}\left(m,\frac{b^2(n)x^2}{n^2};s,\frac{y}{n};b(n)\right)\frac{2^{\frac{1}{2}}\pi^{\frac{1}{2}}}{e^{2\frac{b^2(n)}{n}x}}\\ &=\frac{1}{(2\pi i)^2}\int\limits_{-\frac{1}{2}-i\infty}^{-\frac{1}{2}+i\infty} du\oint\limits_{\Sigma_{\infty}}dt\frac{\prod_{j=0}^s\Gamma(u+\nu_j+1)}{\prod_{j=0}^{m-1}\Gamma(t+\nu_j+1)} \frac{\sin\pi u}{\sin\pi t} \frac{x^ty^{-u-1}}{u-t}. \end{split}$$ $\bullet$ Finally, for $r=m$ and $s=m$ the following limiting relation holds true $$\label{SKR44} \begin{split} &\underset{n\rightarrow\infty}{\lim}\frac{2b^2(n)x^{\frac{1}{2}}y^{\frac{1}{2}}}{n^{2}} S_{n,m}\left(m,\frac{b^2(n)x^2}{n^2};m,\frac{b^2(n)y^2}{n^2};b(n)\right) \frac{e^{2\frac{b^2(n)}{n}y}}{e^{2\frac{b^2(n)}{n}x}}\\ &=\frac{1}{(2\pi i)^2}\int\limits_{-\frac{1}{2}-i\infty}^{-\frac{1}{2}+i\infty} du\oint\limits_{\Sigma_{\infty}}dt\frac{\prod_{j=0}^{m-1}\Gamma(u+\nu_j+1)}{\prod_{j=0}^{m-1}\Gamma(t+\nu_j+1)} \frac{\sin\pi u}{\sin\pi t} \frac{x^ty^{-u-1}}{u-t}. \end{split}$$ In the derivation of these formulae the interchange of limits and integrals can be justified by the dominated convergence theorem in the same way as it is done in Kuijlaars and Zhang [@KuijlaarsZhang Section 5.2.]. Formulae (\[SKR22\])-(\[SKR44\]) give the scaling limits of the second term, $S_{n,m}\left(r,x;s,y;b(n)\right)$, in equation (\[CorrelationKernelFormulaContour\]) for the correlation kernel $K_{n,m}\left(r,x;s,y;b(n)\right)$. Now, let us find the asymptotic of the first term, $\phi_{r,s}(x,y;b)$, in equation (\[CorrelationKernelFormulaContour\]). Recall that we already know the asymptotic of $\phi_{r,s}(x,y;b)$ for $1\leq r,s\leq m-1$. So it is enough to consider two cases. $\bullet$ The first case corresponds to $1\leq r\leq m-2$, $s=m$. In this case we have $$\begin{split} &\frac{2^{\frac{1}{2}}b(n)y^{\frac{1}{2}}}{n^{\frac{3}{2}}} \phi_{r,m}\left(\frac{x}{n};\frac{b^2(n)y^2}{n^2};b(n)\right)\frac{e^{2\frac{b^2(n)}{n}y}}{2^{\frac{1}{2}}\pi^{\frac{1}{2}}}\\ &=\frac{2^{\frac{1}{2}}b(n)y^{\frac{1}{2}}}{n^{\frac{3}{2}}}\frac{n}{x} \left[\int\limits_0^{\infty}G_{0,m-r-1}^{m-r-1,0}\left(\begin{array}{c} - \\ \nu_{r+1},\ldots,\nu_{m-1} \end{array} \biggl|\frac{nt}{x}\right)e^{-\frac{b^2(n)y^2}{n^2t}-b^2(n)t}\frac{dt}{t}\right]\frac{e^{2\frac{b^2(n)}{n}y}}{2^{\frac{1}{2}}\pi^{\frac{1}{2}}}. \end{split}$$ The expression in the brackets can be written as $$I(\Lambda)=\int\limits_0^{\infty}G_{0,m-r-1}^{m-r-1,0}\left(\begin{array}{c} - \\ \nu_{r+1},\ldots,\nu_{m-1} \end{array} \biggl|\frac{\tau}{x}\right)e^{-\Lambda\left(\frac{y^2}{\tau}+\tau\right)} \frac{d\tau}{\tau},$$ where $\Lambda=\frac{b^2(n)}{n}$. Using Laplace’s method for asymptotic expansions of integrals (see, for example, Miller [@Miller Chapter 3.]) we obtain that the expression in the brackets has the following asymptotic $$\begin{split} &\int\limits_0^{\infty}G_{0,m-r-1}^{m-r-1,0}\left(\begin{array}{c} - \\ \nu_{r+1},\ldots,\nu_{m-1} \end{array} \biggl|\frac{nt}{x}\right)e^{-\frac{b^2(n)y^2}{n^2t}-b^2(n)t}\frac{dt}{t}\\ &\qquad\qquad\sim\frac{\pi^{\frac{1}{2}}n^{\frac{1}{2}}}{b(n)y^{\frac{1}{2}}} G_{0,m-r-1}^{m-r-1,0}\left(\begin{array}{c} - \\ \nu_{r+1},\ldots,\nu_{m-1} \end{array} \biggl|\frac{y}{x}\right)e^{-\frac{2b^2(n)y}{n}},\;\;\mbox{as}\;\; \frac{b(n)}{\sqrt{n}}\rightarrow\infty. \end{split}$$ This gives $$\label{philimit1} \underset{n\rightarrow\infty}{\lim}\frac{2^{\frac{1}{2}}b(n)y^{\frac{1}{2}}}{n^{\frac{3}{2}}} \phi_{r,m}\left(\frac{x}{n};\frac{b^2(n)y^2}{n^2};b(n)\right)\frac{e^{2\frac{b^2(n)}{n}y}}{2^{\frac{1}{2}}\pi^{\frac{1}{2}}} =\frac{1}{x} G_{0,m-r-1}^{m-r-1,0}\left(\begin{array}{c} - \\ \nu_{r+1},\ldots,\nu_{m-1} \end{array} \biggl|\frac{y}{x}\right),$$ where $1\leq r\leq m-2$. $\bullet$ For $r=m-1$, $s=m$ we can write $$\begin{split} \frac{2^{\frac{1}{2}}b(n)y^{\frac{1}{2}}}{n^{\frac{3}{2}}} \phi_{m-1,m}\left(\frac{x}{n};\frac{b^2(n)y^2}{n^2};b(n)\right)\frac{e^{2\frac{b^2(n)}{n}y}}{2^{\frac{1}{2}}\pi^{\frac{1}{2}}} &=\frac{b(n)y^{\frac{1}{2}}}{\pi^{\frac{1}{2}}n^{\frac{3}{2}}}\frac{n}{x}e^{-2\frac{b^2(n)y^2}{nx}-\frac{b^2(n)x}{n}+\frac{2b^2(n)y}{n}}\\ &=\frac{b(n)y^{\frac{1}{2}}}{\pi^{\frac{1}{2}}n^{\frac{1}{2}}x}e^{-\frac{b^2(n)}{xn}\left(y-x\right)^2}. \end{split}$$ We conclude that $$\label{philimit2} \begin{split} \underset{n\rightarrow\infty}{\lim}\frac{2^{\frac{1}{2}}b(n)y^{\frac{1}{2}}}{n^{\frac{3}{2}}} \phi_{m-1,m}\left(\frac{x}{n};\frac{b^2(n)y^2}{n^2};b(n)\right)\frac{e^{2\frac{b^2(n)}{n}y}}{2^{\frac{1}{2}}\pi^{\frac{1}{2}}}=\delta(x-y), \end{split}$$ representing the Dirac delta distribution. Taking into account equations (\[SKR22\])-(\[SKR44\]), (\[philimit1\]), and (\[philimit2\]) we obtain the limiting relations (\[SKR2\])-(\[SKR4\]) in the statement of Theorem \[TheoremHardEdgeGinibreCouplingProcess\]. In the weak and intermediate coupling regime in the previous two subsections the points on all levels were rescaled in the same way with $n$. In the strong coupling regime this is no longer so. Let us therefore explain how the different rescaling of the point configurations translates into different rescalings of the various correlation kernels. This justifies the different scalings that we have applied above. Assume that the point configurations $$\left(y_1^m,\ldots,y_n^m;\ldots; y_1^1,\ldots,y_n^1\right)$$ form a determinantal process on $\left\{1,\ldots,m\right\}\times{\mathbb R}_{>0}$ defined by the correlation kernel $K_{n,m}\left(r,x;s,y;b\right)$. Consider the scaled determinantal process on $\left\{1,\ldots,m\right\}\times{\mathbb R}_{>0}$ formed by the scaled point configurations $$\begin{split} &\left(\frac{n}{b(n)}\left(y_1^m\right)^{\frac{1}{2}},\ldots,\frac{n}{b(n)}\left(y_n^m\right)^{\frac{1}{2}}; ny_1^{m-1},\ldots,ny_n^{m-1};\ldots;ny_1^{m-1},\ldots,ny_n^{m-1}\right)\\ &=\left(u_1^m,\ldots,u_n^m;u_1^{m-1},\ldots,u_n^{m-1};\ldots; u_1^1,\ldots,u_n^1\right). \end{split} \nonumber$$ If $\varrho_{k_1,\ldots,k_m}$ is the correlation function of the original determinantal process, see their definition in , and $\widehat{\varrho}_{k_1,\ldots,k_m}$ is that of the scaled determinantal process, then we must have $$\label{rhokscale} \begin{split} &\varrho_{k_1,\ldots,k_m}\left(y_1^1,\ldots,y_{k_1}^1;\ldots;y_1^m,\ldots,y_{k_m}^m\right) \left(dy_1^1,\ldots,dy_{k_1}^1\right)\ldots\left(dy_1^m,\ldots,dy_{k_m}^m\right)\\ &=\widehat{\varrho}_{k_1,\ldots,k_m}\left(u_1^1,\ldots,u_{k_1}^1;\ldots;u_1^m,\ldots,u_{k_m}^m\right) \left(du_1^1,\ldots,du_{k_1}^1\right)\ldots\left(du_1^m,\ldots,du_{k_m}^m\right). \end{split} $$ This gives the following relation between the initial correlation function, $\varrho_{k_1,\ldots,k_m}$, and the correlation function $\widehat{\varrho}_{k_1,\ldots,k_m}$ of the scaled determinant process $$\label{RelationsBetweenCorrelationFunctions} \begin{split} &\widehat{\varrho}_{k_1,\ldots,k_m}\left(u_1^1,\ldots,u_{k_1}^1;\ldots;u_1^{m-1},\ldots,u_{k_{m-1}}^{m-1}; u_1^m,\ldots,u_{k_m}^m\right)\\ &=\frac{2^{k_m}u_1^m\ldots u_{k_m}^m}{n^{k_1+\ldots+k_{m-1}}}\left(\frac{b(n)}{n}\right)^{2k_m}\\ &\times\varrho_{k_1,\ldots,k_m}\left(\frac{u_1^1}{n},\ldots,\frac{u_{k_1}^1}{n};\ldots;\frac{u_1^{m-1}}{n},\ldots,\frac{u_{k_{m-1}}^{m-1}}{n}; \frac{b^2(n)}{n^2}\left(u_1^m\right)^2,\ldots,\frac{b^2(n)}{n^2}\left(u_{k_m}^m\right)^2\right). \end{split}$$ If $K_{n,m}\left(r,u;s,v\right)$ is the correlation kernel of the initial determinantal process, and $\widehat{K}_{n,m}\left(r,u;s,v\right)$ is the correlation kernel of the scaled determinantal process, then relation (\[RelationsBetweenCorrelationFunctions\]) between the correlation functions implies $$\label{k0} \begin{split} &\widehat{K}_{n,m}\left(r,u;s,v\right)\\ &=\left\{ \begin{array}{ll} \frac{1}{n}K_{n,m}\left(r,\frac{u}{n};s,\frac{v}{n}\right), & 1\leq r,s\leq m-1, \\ \frac{2^{\frac{1}{2}}b(n)v^{\frac{1}{2}}}{n^{\frac{3}{2}}}K_{n,m}\left(r,\frac{u}{n};m,\frac{b^2(n)}{n^2}v^2\right), & 1\leq r\leq m-1, s=m, \\ \frac{2^{\frac{1}{2}}b(n)u^{\frac{1}{2}}}{n^{\frac{3}{2}}}K_{n,m}\left(m,\frac{b^2(n)}{n^2}u^2;s,\frac{v}{n}\right), & r=m, 1\leq s\leq m-1, \\ \frac{2b^2(n)u^{\frac{1}{2}}v^{\frac{1}{2}}}{n^2}K_{n,m}\left(m,\frac{b^2(n)}{n^2}u^2;m,\frac{b^2(n)}{n^2}v^2\right), & r=m, s=m. \end{array} \right. \end{split}$$ Also, let us take into account that if we have two correlation kernels, say $K_{n,m}\left(r,x;s,y\right)$ and $K_{n,m}'\left(r,x;s,y\right)$ living on $\left\{1,\ldots,m\right\}\times{\mathbb R}_{>0}$, and related as $$K_{n,m}'\left(r,x;s,y\right)=\frac{\varphi_r(x)}{\varphi_s(y)}K_{n,m}\left(r,x;s,y\right),$$ where $\varphi_1$, $\ldots$, $\varphi_m$ are certain non-vanishing functions defined on ${\mathbb R}_{>0}$, then both $K_{n,m}\left(r,x;s,y\right)$ and $K_{n,m}'\left(r,x;s,y\right)$ give the same correlation functions, and therefore define the same determinantal point process on $\left\{1,\ldots,m\right\}\times{\mathbb R}_{>0}$. The two kernels are therefore called equivalent kernels. In other words, we have from $$\begin{split} &\det\left[\begin{array}{ccc} \left(K_{n,m}(1,x_i^1;1,x_j^1)\right)_{i=1,\ldots,k_1}^{j=1,\ldots,k_1} & \ldots & \left(K_{n,m}(1,x_i^1;m,x_j^m)\right)_{i=1,\ldots,k_1}^{j=1,\ldots,k_m} \\ \vdots & & \\ \left(K_{n,m}(m,x_i^m;1,x_j^1)\right)_{i=1,\ldots,k_m}^{j=1,\ldots,k_1} & \ldots & \left(K_{n,m}(m,x_i^m;m,x_j^m)\right)_{i=1,\ldots,k_m}^{j=1,\ldots,k_m} \end{array} \right]\\ &=\det\left[\begin{array}{ccc} \left(K_{n,m}'(1,x_i^1;1,x_j^1)\right)_{i=1,\ldots,k_1}^{j=1,\ldots,k_1} & \ldots & \left(K_{n,m}'(1,x_i^1;m,x_j^m)\right)_{i=1,\ldots,k_1}^{j=1,\ldots,k_m} \\ \vdots & & \\ \left(K_{n,m}'(m,x_i^m;1,x_j^1)\right)_{i=1,\ldots,k_m}^{j=1,\ldots,k_1} & \ldots & \left(K_{n,m}'(m,x_i^m;m,x_j^m)\right)_{i=1,\ldots,k_m}^{j=1,\ldots,k_m} \end{array} \right].\\ \end{split} \nonumber$$ Therefore, the correlation kernel $\widehat{K}'_{n,m}\left(r,x;s,y\right)$ defined by $$\begin{split} &\widehat{K}'_{n,m}\left(r,u;s,v\right)\\ &=\left\{ \begin{array}{ll} \frac{1}{n}K_{n,m}\left(r,\frac{u}{n};s,\frac{v}{n}\right), & 1\leq r,s\leq m-1, \\ \frac{2^{\frac{1}{2}}b(n)v^{\frac{1}{2}}}{n^{\frac{3}{2}}}K_{n,m}\left(r,\frac{u}{n};m,\frac{b^2(n)}{n^2}v^2\right) \frac{e^{2\frac{b^2(n)}{n}v}}{2^{\frac{1}{2}}\pi^{\frac{1}{2}}}, & 1\leq r\leq m-1, s=m, \\ \frac{2^{\frac{1}{2}}b(n)u^{\frac{1}{2}}}{n^{\frac{3}{2}}}K_{n,m}\left(m,\frac{b^2(n)}{n^2}u^2;s,\frac{v}{n}\right) \frac{2^{\frac{1}{2}}\pi^{\frac{1}{2}}}{e^{2\frac{b^2(n)}{n}u}}, & r=m, 1\leq s\leq m-1, \\ \frac{2b^2(n)u^{\frac{1}{2}}v^{\frac{1}{2}}}{n^2}K_{n,m}\left(m,\frac{b^2(n)}{n^2}u^2;m,\frac{b^2(n)}{n^2}v^2\right) \frac{e^{2\frac{b^2(n)}{n}v}}{e^{2\frac{b^2(n)}{n}u}}, & r=m, s=m, \end{array} \right. \end{split}$$ gives the same scaled determinantal process as the correlation kernel $\widehat{K}_{n,m}\left(r,x;s,y\right)$ defined by equation (\[k0\]). We conclude that the limiting relations (\[SKR2\])-(\[SKR4\]) imply that the scaled determinantal process on $\left\{1,\ldots,m\right\}\times{\mathbb R}_{>0}$ formed by the scaled point configurations $$\begin{split} \left(\frac{n}{b(n)}\left(y_1^m\right)^{\frac{1}{2}},\ldots,\frac{n}{b(n)}\left(y_n^m\right)^{\frac{1}{2}}; ny_1^{m-1},\ldots,ny_n^{m-1};\ldots;ny_1^{m-1},\ldots,ny_n^{m-1}\right) \end{split} \nonumber$$ converges as $n\rightarrow\infty$ to the determinantal point process ${\mathcal{P}}^{\operatorname{Ginibre}}_{\infty,m-1}$ on $\left\{1,\ldots,m-1\right\}\times{\mathbb R}_{>0}$ with the identification of levels $m$ and $m-1$ as given in Theorem \[TheoremHardEdgeGinibreCouplingProcess\] [(C)]{}. Proof of Theorem \[TheoremInterpolation\] {#SectionProofTheoremInterpolation} ========================================= Consider the determinantal point process on $\left\{1,\ldots,m\right\}\times{\mathbb R}_{>0}$ defined by the correlation kernel $K_{\infty,m}^{\operatorname{interpol}}(r,x;s,y;\alpha)$, see equation (\[KernelZKExtended\]). From equation (\[PHI(x,y,b)\]) we obtain that $$\underset{\alpha\rightarrow 0}{\lim}\phi_{r,s}(x,y;\alpha)=\frac{1}{x} G_{0,s-r}^{s-r,0}\left(\begin{array}{c} - \\ \nu_{r+1},\ldots,\nu_s \end{array} \biggl|\frac{y}{x}\right)\textbf{1}_{s>r}.$$ Here, we have used the fact that $\nu_m=0$, the integration formula $$\begin{split} \int\limits_0^{\infty}G_{0,m-r-1}^{m-r-1,0}\left(\begin{array}{c} - \\ \nu_{r+1},\ldots,\nu_{m-1} \end{array} \biggl|\frac{y}{t}\right) t^{\nu_m}e^{-t} \frac{dt}{t} =G_{0,m-r}^{m-r,0}\left(\begin{array}{c} - \\ \nu_{r+1},\ldots,\nu_{m} \end{array} \biggl|{y}\right), \end{split}$$ together with the representation of the exponential function in terms of the Meijer $G$-function . Thus, as $\alpha\rightarrow 0$, the first term in the right-hand side of equation (\[KernelZKExtended\]) for the correlation kernel $K_{\infty,m}^{\operatorname{interpol}}(r,x;s,y;\alpha)$ turns into the first term in the right-hand side of equation (\[KernelInfiniteGinibreProductProcess\]) for the correlation kernel $K_{\infty,m}^{\operatorname{Ginibre}}(r,x;s,y)$. Moreover, taking into account that $$\underset{\alpha\rightarrow 0}{\lim}p_m(t,x;\alpha)=1,\;\; \underset{\alpha\rightarrow 0}{\lim}q_m(u,y;\alpha)=1,$$ from equations (\[ISmallb\]) and (\[KSmallb\]), we see that the second term in formula (\[KernelZKExtended\]) for $K_{\infty,m}^{\operatorname{interpol}}(r,x;s,y;\alpha)$ turns into the second term in formula (\[KernelInfiniteGinibreProductProcess\]) for $K_{\infty,m}^{\operatorname{Ginibre}}(r,x;s,y)$. In these calculations the procedure of taking the limit inside the double integral can be justified using the dominated convergence theorem. Thus Theorem \[TheoremInterpolation\] [(A)]{} is proved. The proof of Theorem \[TheoremInterpolation\] [(B)]{} is very similar to that of Theorem \[TheoremHardEdgeGinibreCouplingProcess\] [(C)]{}. For this reason we present the sketch of the proof of Theorem \[TheoremInterpolation\] [(B)]{} only. For $1\leq r,s\leq m-1$ the kernel does not depend on $\alpha$ and agrees with that of the Ginibre point process given by equation (\[KernelInfiniteGinibreProductProcess\]). In order to check equations (\[I2\])-(\[I4\]) we use the known asymptotic formulae for the modified Bessel functions of the first and of the second kind (see equations (\[A1\]) and (\[A2\])) to deduce that $$p_m\left(t,\alpha^2x^2;\alpha\right)\sim\frac{\Gamma(t+1)}{\left(\alpha^2x\right)^t} \frac{e^{2\alpha^2x}}{2\pi^{\frac{1}{2}}\alpha x^{\frac{1}{2}}},\;\;\alpha\rightarrow\infty,$$ and that $$q_m\left(u+1,\alpha^2y^2;\alpha\right)\sim\frac{\pi^{\frac{1}{2}}}{\Gamma(u+1)} \frac{\left(\alpha^2y\right)^{u+1}}{\alpha y^{\frac{1}{2}}}e^{-2\alpha^2y},\;\;\alpha\rightarrow\infty.$$ Similar arguments as in the proof of Theorem \[TheoremHardEdgeGinibreCouplingProcess\] [(C)]{}, can be applied to show that equations (\[I2\])-(\[I4\]) ensure the convergence of the scaled interpolating process to $\mathcal{P}^{\operatorname{Ginibre}}_{\infty,m-1}$. [99]{} Adhikari, K.; Kishore Reddy, N.; Ram Reddy, T.; Saha, K. Determinantal point processes in the plane from products of random matrices. Ann. Inst. Henri Poincaré Probab. Stat. [**52**]{} (2016), no. 1, 16–46. Akemann, G.; Burda, Z. Universal microscopic correlation functions for products of independent Ginibre matrices. J. Phys. A: Math. Theor. [**45**]{} (2012) 465201. Akemann, G.; Checinski, T.; Liu, D.-Z.; Strahov, E. Finite rank perturbations in products of coupled random matrices: From one correlated to two Wishart ensembles. arXiv:1704.05224. Akemann, G.; Kieburg M.; Wei, L. Singular value correlation functions for products of Wishart random matrices. J. Phys. A. [**46**]{} (2013) 275205, 24pp. Akemann, G.; Ipsen, J.; Kieburg M. Products of rectangular random matrices: singular values and progressive scattering. Phys. Rev. E [**88**]{} (2013) 052118, 14pp. Akemann, G.; Ipsen, J.; Strahov, E. Permanental processes from products of complex and quaternionic induced Ginibre ensembles. Random Matrices Theory Appl. [**3**]{} (2014), no. 4, 1450014, 54. Akemann, G.; Strahov, E. Dropping the independence: singular values for products of two coupled random matrices. Commun. Math. Phys. [**345**]{} (2016), no. 1, 101–140. Akemann, G.; Strahov, E. Hard edge limit of the product of two strongly coupled random matrices. Nonlinearity [**29**]{} (2016) 3743-3776 Ambj[ø]{}rn, J,; Chekov, L. The matrix model for hypergeometric Hurwitz numbers. Theor. Math. Phys. [**181**]{}, Issue 3 (2014), 1486–1498. Bertola, M.; Bothner, T. Universality conjecture and results for a model of several coupled positive-definite matrices. Commun. Math. Phys. [**337**]{} (2015), no. 3, 1077–1141. Bertola, M.; Eynard, B.; Harnad, J. Duality, Biorthogonal Polynomials and Multi-Matrix Models Commun. Math. Phys. [**229**]{} (2002), no. 1, 73–120. Bertola, M.; Gekhtman, M.; Szmigielski, J. The Cauchy two-matrix model. Commun. Math. Phys. [**287**]{} (2009), no. 3, 983–1014. Bertola, M.; Gekhtman, M.; Szmigielski, J. Strong asymptotics for Cauchy biorthogonal polynomials with application to the Cauchy two-matrix model. J. Math. Phys. [**54**]{} (2013), no. 4, 043517, 25 Bertola, M.; Gekhtman, M.; Szmigielski, J. Cauchy-Laguerre two-matrix model and the Meijer-G random point field. Commun. Math. Phys. [**326**]{} (2014), no. 1, 111–144. Borodin, A. Deteminantal point processes. Chapter 11, The Oxford Handbook of Random Matrix Theory (Eds. G. Akemann, J. Baik, P. Di Francesco). Oxford Unversity Press, Oxford, (2011), 231–249. Brower, R.; Rossi, P; Tan, C.-I., The external field problem for QCD. Nuclear Phys. B [**190**]{} \[FS3\] (1981) no. 4, 699-718. Eynard, B.; Kimura, T.; Ribault, S. Random matrices. arXiv:1510.04430. Eynard, B.; Mehta, M. L. Matrices coupled in a chain. I. Eigenvalue correlations. J. Phys. A [**31**]{} (1998), no. 19, 4449-4456. Filev, V. G.; O’Connor, D. Multi-matrix models at general coupling. J. Phys. A [**46**]{} (2013), no. 47, 475403, 21 pp. Fischmann, J.; Bruzda, W.; Khoruzhenko, B.A.; Sommers, H.-J.; Zyczkowski, K. Induced Ginibre ensemble of random matrices and quantum operations. J. Phys. A [**45**]{} (2012), no. 7, 075203. Forrester, P. J. Eigenvalue statistics for product complex Wishart matrices. J. Phys. A [**47**]{} (2014), no. 34, 345202. Forrester, P.J.; Ipsen, J.R. Real eigenvalue statistics for products of asymmetric real Gaussian matrices. Lin. Alg. Appl. [**510**]{} (2016), 259-290 Gradshteyn, I.S.; Ryzhik, I.M. Table of Integrals, Series, and Products. A. Jeffrey and D. Zwillinger (eds.). Sixth edition (2000). Harish-Chandra. Differential operators on a semisimple Lie algebra, Am. J. Math. [**79**]{} (1957) 87-120. Ipsen, J. Products of independent quaternion Ginibre matrices and their correlation functions. J. Phys. A [**46**]{} (2013), no. 26, 265201. Ismail, M. E. H. Classical and quantum orthogonal polynomials in one variable. Encyclopedia of Mathematics and its Applications, 98. Cambridge University Press, Cambridge, 2009. Itzykson C. and Zuber, J.B. The Planar Approximation 2. J. Math. Phys. [**21**]{} (1980) 411. Johansson, K. Random matrices and determinantal processes. Lecture notes of the Les Houches Summer School Session LXXXIII, 2005. Mathematical Statistical Physics (Elsevier B. V, Amsterdam, 2006) 1-55. Johansson, K. From Gumbel to Tracy-Widom. Probab. Theory Related Fields [**138**]{}, no. 1-2 (2007) 75–112. Kuijlaars A.B.J.; Stivigny, D. Singular values of products of random matrices and polynomial ensembles. Random Matrices: Theory and Applications [**3**]{} (2014) 1450011. Kuijlaars, A.B.J.; Zhang, L. Singular values of products of Ginibre random matrices, multiple orthogonal polynomials and hard edge scaling limits. Commun. Math. Phys. [**332**]{} (2014) 759–781. Leutwyler, H.; Smilga, A. Spectrum of Dirac operator and role of winding number in QCD. Phys. Rev. D [**46**]{} (1992) 5607. Liu, D.-Z. Singular values for products of two coupled random matrices: hard edge phase transition. Constr. Approx. (2017). https://doi.org/10.1007/s00365-017-9389-z. Liu, D-Z.; Wang, D.; Zhang, L. Bulk and soft-edge universality for singular values of products of Ginibre random matrices. Ann. Inst. Henri Poincaré, Prob. Stat. [**52**]{}, No. 4 (2016) 1734-1762. Luke, Y.L. The special functions and their approximations. Academic Press, New York 1969. Miller, P. D. Applied Asymptotic Analysis. Graduate Studies in Mathematics. Volume 75. American Mathematical Society. Moshe, M.; Neuberger, H.; Shapiro, B. Generalized ensemble of random matrices. Phys. Rev. Lett. [**73**]{} (1994) 1497–1500. Olver, F.W.L et al. (Eds.), NIST Handbook of Mathematical Functions. Cambridge University Press, Cambridge 2010. Osborn, J.C. Universal results from an alternate random matrix model for QCD with a baryon chemical potential. Phys. Rev. Lett. [**93**]{} (2004) 222001. Pandey, A.; Mehta, M. L. Gaussian ensembles of random Hermitian matrices intermediate between orthogonal and unitary ones. Commun. Math. Phys. [**87**]{}, no. 4 (1982/83), 449–468. Strahov, E. Dynamical correlation functions for products of random matrices. Random Matrices Theory Appl. [**4**]{} (2015), no. 4, 1550020. Tracy, C. A.; Widom, H. Differential equations for Dyson processes. Commun. Math. Phys. [**252**]{} (2004), 7–41. [^1]: We refer the reader to the book by Luke [@Luke] for the definition of Meijer $G$-functions, and for their exact and asymptotic properties. [^2]: For $m=1$ the corresponding one-matrix model is convergent for any $b$. [^3]: Note that for $m=2$ the additional potential $V_1(t)=bt^2$ simply corresponds to a shift of the Gaussian term ${\mathop{\mathrm{Tr}}}[G_1^*G_1]$ which is already present. [^4]: We follow the convention here that the index $m-1$ of the kernel on the left-hand side indicates the range that the arguments $r$ and $s$ can take. [^5]: For other presentations of the Eynard-Mehta Theorem, and for different proofs we refer the reader to Borodin [@Borodin], Tracy and Widom [@TracyWidom]. [^6]: Recall that the Pochhammer symbol, $(a)_l$, is defined by $ (a)_l=a(a+1)\ldots(a+l-1). $

ArXiv

--- abstract: | Given a connected graph $G\ $of order $n$ and a nonnegative symmetric matrix $A=\left[ a_{i,j}\right] $ of order $n,$ define the function $F_{A}\left( G\right) $ as$$F_{A}\left( G\right) =\sum_{1\leq i<j\leq n}d_{G}\left( i,j\right) a_{i,j},$$ where $d_{G}\left( i,j\right) $ denotes the distance between the vertices $i$ and $j$ in $G.$ In this note it is shown that $F_{A}\left( G\right) \leq F_{A}\left( P\right) \,$for some path of order $n.$ Moreover, if each row of $A$ has at most one zero off-diagonal entry, then $F_{A}\left( G\right) <F_{A}\left( P\right) \,$for some path of order $n,$ unless $G$ itself is a path. In particular, this result implies two conjectures of Aouchiche and Hansen: - the spectral radius of the distance Laplacian of a connected graph $G$ of order $n$ is maximal if and only if $G$ is a path; - the spectral radius of the distance signless Laplacian of a connected graph $G$ of order $n$ is maximal if and only if $G$ is a path. **AMS classification:** *15A42; 05C50.* **Keywords:** *distance matrix; distance Laplacian; distance signless Laplacian; largest eigenvalue; path.* author: - 'Celso Marques da Silva Jr[^1] [^2]  and Vladimir Nikiforov[^3]' title: '**Graph functions maximized on a path** ' --- Introduction and main results ============================= The aim of the present note is to give a general approach to problems like the following conjectures of Aouchiche and Hansen [@AuHa11; @AuHa13a]: \[con1\]The largest eigenvalue of the distance Laplacian of a connected graph $G$ of order $n$ is maximal if and only if $G$ is a path. \[con2\]The largest eigenvalue of the distance signless Laplacian of a connected graph $G$ of order $n$ is maximal if and only if $G$ is a path. First, let us introduce some notation and recall a few definitions. We write $\lambda\left( A\right) $ for the largest eigenvalue of a symmetric matrix $A$. Given a connected graph $G,$ let $D\left( G\right) $ be the distance matrix of $G,$ and let $T\left( G\right) $ be the diagonal matrix of the rowsums of $D\left( G\right) .$ The matrix $D^{L}\left( G\right) =T\left( G\right) -D\left( G\right) $ is called the *distance Laplacian* of $G,$ and the matrix $D^{Q}\left( G\right) =T\left( G\right) +D\left( G\right) $ is called the *distance signless Laplacian* of $G$. The matrices $D^{L}\left( G\right) $ and $D^{Q}\left( G\right) $ have been introduced by Aouchiche and Hansen and have been intensively studied recently, see, e.g., [@AuHa11; @AuHa13a; @AuHa13; @LiLu14; @NaPa14; @XZL13]. Very recently, Lin and Lu [@LiLu14] succeeded to prove Conjecture \[con2\], but Conjecture \[con1\] seems a bit more difficult and still holds. Furthermore, Conjectures \[con1\] and \[con2\] suggest a similar problem for the distance matrix itself. As it turns out such problem has been partially solved a while ago by Ruzieh and Powers [@RuPo90], who showed that the largest eigenvalue of the distance matrix of a connected graph $G$ of order $n$ is maximal if $G$ is a path. The complete solution, however, was given more recently by Stevanović and Ilić [@StIl10]. \[**[@RuPo90],[@StIl10]**\]\[thSI\]The largest eigenvalue of the distance matrix of a connected graph $G$ of order $n$ is maximal if and only if $G$ is a path. These result are believed to belong to spectral graph theory, and their proofs involve nonnegligible amount of calculations. Our goal is to show that all these results stem from a much more general assertion that has nothing to do with eigenvalues. To this end, we shall introduce a fairly general graph function and shall study its maxima. The function $F_{A}\left( G\right) $ and its maxima ----------------------------------------------------- Let $G$ be a connected graph of order $n.$ Write $d_{G}\left( i,j\right) $ for the distance between the vertices $i$ and $j$ in $G,$ and let $A=\left[ a_{i,j}\right] $ be a nonnegative symmetric matrix of order $n.$ Define the function $F_{A}\left( G\right) $ as$$F_{A}\left( G\right) =\sum_{1\leq i<j\leq n}d_{G}\left( i,j\right) a_{i,j}.$$ Clearly $d_{G}\left( i,i\right) =0$ for any $i\in V\left( G\right) ,$ so the diagonal of $A$ is irrelevant for $F_{A}\left( G\right) $. In fact, the function $F_{A}\left( G\right) $ is quite mainstream, as it can be represented as$$F_{A}\left( G\right) =\left\Vert A\circ D\left( G\right) \right\Vert _{l_{1}},$$ where $\circ$ denotes the entrywise Hadamard product of matrices, and $\left\Vert \cdot\right\Vert _{l_{1}}$ is the $l_{1}$ norm. This viewpoint suggests a number of extensions, which we shall investigate elsewhere. Next, we focus on the extremal points of $F_{A}\left( G\right) ,$ that is to say, we want to know which connected graphs $G$ of order $n$ satisfy the condition$$F_{A}\left( G\right) =\max\left\{ F_{A}\left( H\right) :\text{ }H\text{ is a connected graph of order }n\right\} .\text{ }$$ In particular, we prove the somewhat surprising fact that for any admissible matrix $A,$ the function $F_{A}\left( G\right) $ is always maximized by a path. More precisely the following theorem holds. \[thg\]Let $G$ be a connected graph of order $n$ and let $A=\left[ a_{i,j}\right] $ be a symmetric matrix of order $n.$ If $A$ is nonnegative, then there is a path $P$ with $V\left( P\right) =V\left( G\right) $ such that $$\sum_{1\leq i<j\leq n}d_{G}\left( i,j\right) a_{i,j}\leq\sum_{1\leq i<j\leq n}d_{P}\left( i,j\right) a_{i,j}. \label{mi}$$ It is not hard to find nonnegative symmetric matrices $A$ for which $F_{A}\left( G\right) $ is maximized also by graphs other than paths. Thus, it is natural to attempt to characterize all symmetric, nonnegative matrices $A,$ for which $F_{A}\left( G\right) $ is maximal only if $G$ is a path. The complete solution of this problem seems difficult, so we shall give only a partial solution, sufficient for our goals. \[ths\]Let $G$ be a connected graph of order $n$ and let $A=\left[ a_{i,j}\right] $ be a symmetric nonnegative matrix of order $n.$ If each row of $A$ has at most one zero off-diagonal entry, and $G$ is not a path, then there is a path $P$ with $V\left( P\right) =V\left( G\right) $ such that$$\sum_{1\leq i<j\leq n}d_{G}\left( i,j\right) a_{i,j}<\sum_{1\leq i<j\leq n}d_{P}\left( i,j\right) a_{i,j}. \label{mip}$$ As yet we know of no application that exploits the full strength of Theorem \[ths\]. Indeed, to prove Conjectures \[con1\] and \[con2\], and Theorem \[thSI\], we shall use only the following simple corollary. \[cor1\] Let $G$ be a connected graph of order $n$ and let $A=\left[ a_{i,j}\right] $ be a symmetric matrix of order $n.$ If each off-diagonal entry of $A$ is positive, and $G$ is not a path, then there is a path $P$ with $V\left( P\right) =V\left( G\right) $ such that $F_{A}\left( P\right) >F_{A}\left( G\right) .$ Proofs of Conjectures \[con1\] and \[con2\], and Theorem \[thSI\] ----------------------------------------------------------------- We proceed with the proof of Conjecture \[con2\]. Let $G$ be a connected graph of order $n$ for which $\lambda\left( D^{Q}\left( G\right) \right) $ is maximal within all connected graphs of order $n$. We shall prove that $G$ is a path. Let $\mathbf{x}=\left( x_{1},\ldots,x_{n}\right) $ be a unit eigenvector to $\lambda\left( D^{Q}\left( G\right) \right) .$ Since $D^{Q}\left( G\right) $ is irreducible, the vector $\mathbf{x}$ is positive. Define an $n\times n$ matrix $A=\left[ a_{i,j}\right] $ by letting $a_{i,j}=\left( x_{i}+x_{j}\right) ^{2}.$ Clearly $A$ is symmetric and nonnegative. As is well-known, $$\lambda\left( D^{Q}\left( G\right) \right) =\left\langle D^{Q}\left( G\right) \mathbf{x},\mathbf{x}\right\rangle =\sum_{1\leq i<j\leq n}d_{G}\left( i,j\right) \left( x_{i}+x_{j}\right) ^{2}=F_{A}\left( G\right) .$$ Since each off-diagonal entry of $A$ is positive, Corollary \[cor1\] implies that either $G=P_{n}$ or there is a path $P$ with $V\left( P\right) =V\left( G\right) $ such that $F_{A}\left( P\right) >F_{A}\left( G\right) .$ The latter cannot hold as we would have $$\lambda\left( D^{Q}\left( G\right) \right) =F_{A}\left( G\right) <F_{A}\left( P\right) \leq\lambda\left( D^{Q}\left( P\right) \right) ,$$ contrary to the choice of $G.$ Hence $G=P_{n},$ completing the proof of Conjecture \[con2\]. Theorem \[thSI\] can be proved in the same way, with $A=\left[ a_{i,j}\right] $ defined by $a_{i,j}=x_{i}x_{j}.$ However, Conjecture \[con1\] requires a slightly more careful approach. Let $G$ be a connected graph of order $n$ such that $\lambda\left( D^{L}\left( G\right) \right) $ is maximal among all connected $n$ vertex graphs. We shall prove that $G$ must be a path. Let $\mathbf{x}=\left( x_{1},\ldots,x_{n}\right) $ be a unit eigenvector to $\lambda\left( D^{L}\left( G\right) \right) $ and define an $n\times n$ matrix $A=\left[ a_{i,j}\right] $ by letting $a_{i,j}=\left( x_{i}-x_{j}\right) ^{2}.$ Clearly $A$ is symmetric and nonnegative. Also, it is well-known that $$\lambda\left( D^{L}\left( G\right) \right) =\left\langle D^{L}\left( G\right) \mathbf{x},\mathbf{x}\right\rangle =\sum_{1\leq i<j\leq n}d_{G}\left( i,j\right) \left( x_{i}-x_{j}\right) ^{2}=F_{A}\left( G\right) .$$ However, at this stage we cannot rule out that $A$ has numerous zero entries, and so Corollary \[cor1\] does not apply as before. Yet Theorem \[thg\] implies that there is a path $P$ with $V\left( P\right) =V\left( G\right) $ such that $F_{A}\left( P\right) \geq F_{A}\left( G\right) ;$ hence,$$\lambda\left( D^{L}\left( G\right) \right) =F_{A}\left( G\right) \leq F_{A}\left( P\right) \leq\lambda\left( D^{L}\left( P\right) \right) .$$ Due to the choice of $G,$ equalities should hold throughout the above line, implying that $\mathbf{x}$ is an eigenvector to $P_{n}.$ But in Theorems. 4.4 and 4.6 of [@NaPa14] Nath and Paul have established that all entries of an eigenvector to $\lambda\left( D^{L}\left( P\right) \right) $ are different and so the off-diagonal entries of $A$ are positive. Now we apply Corollary \[cor1\] and finish the proof as for Conjecture \[con2\]. Proofs of the main theorems =========================== For graph notation undefined here we refer the reader to [@Bol98]. For general properties of the distance Laplacian and the distance signless Laplacian the reader is referred to [@AuHa11; @AuHa13a; @AuHa13]. Here is some notation that will be used later in the proofs: - $P_{n}$ and $C_{n}$ stand for the path and cycle of order $n;$ - $G-u$ denotes the graph obtained from $G$ by removing the vertex $u;$ - $G-\left\{ u,v\right\} $ denotes the graph obtained from $G$ by removing the vertices $u$ and $v.\medskip$ We shall assume that any graph of order $n$ is defined on the vertex set $\left[ n\right] =\left\{ 1,\ldots,n\right\} .\medskip$ The proofs of Theorems \[thg\] and \[ths\] have the same general structure, but the latter requires a lot of extra details so it will be presented separately. \[**Proof of Theorem \[thg\]**\]Note first that if $H$ is a spanning tree of $G,$ then $d_{G}\left( i,j\right) \leq d_{H}\left( i,j\right) $ for every $i,j\in V\left( G\right) ;$ hence $$\sum_{1\leq i<j\leq n}d_{G}\left( i,j\right) a_{i,j}\leq\sum_{1\leq i<j\leq n}d_{H}\left( i,j\right) a_{i,j}.$$ Therefore, we may and shall assume that $G$ is a tree itself. We carry out the proof by induction on $n.$ If $n\leq3,$ every tree of order $n$ is a path, so there is nothing to prove in this case. Assume now that $n>3$ and the assertion holds for any $n^{\prime}$ such that $n^{\prime}<n.$ Choose a vertex $u\in V\left( G\right) $ of degree $1.$ By symmetry, we assume that $u=n,$ and let $k$ be the single neighbor of $u;$ hence $G-n$ is a tree of order $n-1.$ Define a symmetric matrix $A^{\prime}=\left[ a_{ij}^{\prime}\right] $ of order $n-1$ as follows:$$a_{i,j}^{\prime}=\left\{ \begin{array} [c]{ll}a_{i,j}, & \text{if\textbf{ }}i\neq k\text{ and }j\neq k;\text{\textbf{ } }\\ a_{k,j}+a_{n,j}, & \text{if\textbf{ }}i=k;\\ a_{i,k}+a_{i,n}, & \text{if\textbf{ }}j=k. \end{array} \right.$$ Clearly $A^{\prime}$ is a symmetric nonnegative matrix. By the induction assumption there is a path $P^{\prime}$ with $V\left( P^{\prime}\right) =V\left( G-n\right) =\left[ n-1\right] $ such that $$\sum_{1\leq i<j<n}d_{G-n}\left( i,j\right) a_{i,j}^{\prime}\leq\sum_{1\leq i<j<n}d_{P^{\prime}}\left( i,j\right) a_{i,j}^{\prime}. \label{in1}$$ On the other hand, for each $j\in V\left( G-n\right) ,$ the shortest path between $n$ and $j$ contains $k,$ so$$d_{G}\left( j,n\right) =d_{G-n}\left( j,k\right) +1.$$ Hence we see that $$\begin{aligned} \sum_{1\leq i<j\leq n}d_{G}\left( i,j\right) a_{i,j} & =\sum_{j=1}^{n-1}d_{G}\left( j,n\right) a_{j,n}+\sum_{1\leq i<j<n}d_{G-n}\left( i,j\right) a_{i,j}\\ & =\sum_{j=1}^{n-1}\left( d_{G-n}\left( k,j\right) +1\right) a_{j,n}+\sum_{1\leq i<j<n}d_{G-n}\left( i,j\right) a_{i,j}\\ & =\sum_{j=1}^{n-1}a_{n,j}+\sum_{1\leq i<j<n}d_{G-n}\left( i,j\right) a_{i,j}^{\prime}.\end{aligned}$$ Now, (\[in1\]) implies that $$\sum_{1\leq i<j\leq n}d_{G}\left( i,j\right) a_{i,j}\leq\sum_{j=1}^{n-1}a_{j,n}+\sum_{1\leq i<j<n}d_{P^{\prime}}\left( i,j\right) a_{i,j}^{\prime}. \label{in2}$$ Further, write $T$ for the tree obtained form the path $P^{\prime}$ by joining $n$ to the vertex $k\in V\left( P^{\prime}\right) $. As before, we see that $$\begin{aligned} \sum_{1\leq i<j\leq n}d_{T}\left( i,j\right) a_{i,j} & =\sum_{j=1}^{n-1}d_{T}\left( j,n\right) a_{j,n}+\sum_{1\leq i<j<n}d_{T-n}\left( i,j\right) a_{i,j}\\ & =\sum_{j=1}^{n-1}\left( d_{P^{\prime}}\left( j,k\right) +1\right) a_{j,n}+\sum_{1\leq i<j<n}d_{P^{\prime}}\left( i,j\right) a_{i,j}\\ & =\sum_{j=1}^{n-1}a_{j,n}+\sum_{1\leq i<j<n}d_{P^{\prime}}\left( i,j\right) a_{i,j}^{\prime}.\end{aligned}$$ Hence, (\[in2\]) implies that$$\sum_{1\leq i<j\leq n}d_{G}\left( i,j\right) a_{i,j}\leq\sum_{1\leq i<j\leq n}d_{T}\left( i,j\right) a_{i,j}.$$ If $T=P_{n}$, there is nothing to prove, so suppose that $T\neq P_{n}.$ To complete the proof we shall show that we can join $n$ to one of the ends of $P^{\prime}$ so that $F_{A}\left( T\right) $ will not decrease. By symmetry, assume that the vertex sequence of the path $P^{\prime}$ is precisely $1,2,\ldots,n-1;$ thus the neighbor $k$ of $n$ satisfies $1<k<n-1.$ Write $A_{0}$ for the principal submatrix of $A$ in the first $n-1$ rows and note that$$F_{A}\left( T\right) =\sum_{i=1}^{k}\left( k-i+1\right) a_{i,n}+\sum_{i=k+1}^{n-1}\left( i-k+1\right) a_{i,n}+F_{A_{0}}\left( P^{\prime }\right) .$$ Next, delete the edge $\left\{ n,k\right\} $ in $T,$ add the edge $\left\{ n,1\right\} ,$ and write $T_{1}$ for the resulting path. If $F_{A}\left( T_{1}\right) >F_{A}\left( T\right) ,$ the proof is completed, so let us assume that $F_{A}\left( T_{1}\right) \leq F_{A}\left( T\right) .$ Since $$F_{A}\left( T_{1}\right) =\sum_{i=1}^{n-1}ia_{i,n}+F_{A_{0}}\left( P^{\prime}\right) ,$$ we see that, $$\sum_{i=1}^{k-1}\left( k-i+1\right) a_{i,n}+\sum_{i=k}^{n-1}\left( i-k+1\right) a_{i,n}\geq\sum_{i=1}^{n-1}ia_{i,n}$$ and so$$\sum_{i=1}^{k-1}\left( k-2i+1\right) a_{i,n}\geq\left( k-1\right) \left( a_{k,n}+\cdots+a_{n-1,n}\right) .$$ Hence,$$\left( k-1\right) \left( a_{1,n}+\cdots+a_{k-1,n}\right) \geq\left( k-1\right) \left( a_{k,n}+\cdots+a_{n-1,n}\right) . \label{in3}$$ Now, delete the edge $\left\{ n,k\right\} $ in $T,$ add the edge $\left\{ n,n-1\right\} ,$ and write $T_{2}$ for the resulting path. If $F_{A}\left( T_{2}\right) >F_{A}\left( T\right) ,$ the proof is completed, so let us assume that $F_{A}\left( T_{2}\right) \leq F_{A}\left( T\right) .$ Since $$F_{A}\left( T_{2}\right) =\sum_{i=1}^{n-1}\left( n-i\right) a_{i,n}+F_{A_{0}}\left( P^{\prime}\right) ,$$ we see that $$\sum_{i=1}^{k-1}\left( k-i+1\right) a_{i,n}+\sum_{i=k}^{n-1}\left( i-k+1\right) a_{i,n}\geq\sum_{i=1}^{n-1}\left( n-i\right) a_{i,n},$$ and so$$\sum_{i=k}^{n-1}\left( 2i-k-n+1\right) a_{i,n}\geq\left( n-k-1\right) \left( a_{1,n}+\cdots+a_{k-1,n}\right) .$$ Hence, $$\left( n-k-1\right) \left( a_{k,n}+\cdots+a_{n-1,n}\right) \geq\left( n-k-1\right) \left( a_{1,n}+\cdots+a_{k-1,n}\right) .$$ This inequality, together with (\[in3\]), implies that $$a_{k,n}+\cdots+a_{n-1,n}=a_{1,n}+\cdots+a_{k-1,n},$$ and that $F_{A}\left( T_{1}\right) =F_{A}\left( T\right) $ and $F_{A}\left( T_{2}\right) =F_{A}\left( T\right) .$ This completes the induction step and the proof of Theorem \[thg\]. Proof of Theorem \[ths\] ------------------------ Most of the proof of Theorem \[ths\] deals with the case of $G$ being a tree, so we extract this part in Theorem \[tht\] below. The general case will be deduced later by different means. For convenience write $N\left( n\right) $ for the class of all symmetric nonnegative matrix of order $n$ such that each row of $A$ has at most one zero off-diagonal entry. \[tht\]Let $G$ be a tree of order $n.$ If $A\in N\left( n\right) $ and $G\neq P_{n},$ then there exists a path $P$ with $V\left( P\right) =V\left( G\right) $ such that $F_{A}\left( G\right) <F_{A}\left( P\right) .$ Our proof is by induction on $n$ and is structured as the proof of Theorem \[thg\]. If $n\leq3,$ every tree of order $n$ is a path, so there is nothing to prove in this case. For technical reason we would like to give a direct proof for $n=4$ as well. There are two trees of order $4$ - a path and a star. Assume that $G$ is a star, and by symmetry suppose that $2$ is its center. We have $$F_{A}\left( G\right) =2a_{4,1}+a_{4,2}+2a_{4,3}+a_{1,2}+a_{2,3}+2a_{1,3}.$$ Remove the edge $\left\{ 4,2\right\} $ and add the edge $\left\{ 4,1\right\} ,$ thus obtaining a path $G_{1}.$ Assume for a contradiction that $F_{A}\left( G\right) \geq F_{A}\left( G_{1}\right) ,$ which implies that $a_{4,1}\geq a_{4,2}+a_{4,3}.$ Now, remove from $G$ the edge $\left\{ 4,2\right\} $ and add the edge $\left\{ 4,3\right\} ,$ thus obtaining a path $G_{2}.$ Assume for a contradiction that $F_{A}\left( G\right) \geq F_{A}\left( G_{2}\right) ,$ which implies that $a_{4,3}\geq a_{4,2}+a_{4,1}.$ We conclude that $a_{4,2}=0.$ By symmetry, we also get $a_{1,2}=0$ and $a_{3,2}=0;$ hence $A$ has a zero row, contradicting the hypothesis. Thus, $G$ is a path. Assume now that $n\geq5$ and the assertion of Theorem holds for any $n^{\prime}$ such that $n^{\prime}<n.$ Let $G$ be tree for which $F_{A}\left( G\right) $ attains a maximum. We shall prove that $G=P_{n}.$ Choose a vertex $u\in V\left( G\right) $ of degree $1.$ By symmetry, we assume that $u=n,$ and let $k$ be the single neighbor of $u$; hence $G-n$ is a tree of order $n-1.$ Define a symmetric matrix $A^{\prime}=\left[ a_{ij}^{\prime}\right] $ of order $n-1$ as follows$$a_{i,j}^{\prime}=\left\{ \begin{array} [c]{ll}a_{i,j}, & \text{if\textbf{ }}i\neq k\text{ and }j\neq k;\text{\textbf{ } }\\ a_{k,j}+a_{n,j}, & \text{if\textbf{ }}i=k;\\ a_{i,k}+a_{i,n}, & \text{if\textbf{ }}j=k. \end{array} \right.$$ Clearly $A^{\prime}\in N\left( n-1\right) $. Suppose that $G-n\neq P_{n-1}.$ By the induction assumption, there is a path $P^{\prime}$ with $V\left( P^{\prime}\right) =V\left( G-n\right) =\left[ n-1\right] $ such that $$\sum_{1\leq i<j<n}d_{G-n}\left( i,j\right) a_{i,j}^{\prime}<\sum_{1\leq i<j<n}d_{P^{\prime}}\left( i,j\right) a_{i,j}^{\prime}.$$ Hence, as in the proof of Theorem \[thg\], we find that $$F_{A}\left( G\right) =\sum_{j=1}^{n-1}a_{j,n}+\sum_{1\leq i<j<n}d_{G-n}\left( i,j\right) a_{i,j}^{\prime}<\sum_{j=1}^{n-1}a_{j,n}+\sum_{1\leq i<j<n}d_{P^{\prime}}\left( i,j\right) a_{i,j}^{\prime}.$$ Now, join $n$ to $k,$ and write $T$ for the obtained tree. As before, we see that $$F_{A}\left( T\right) =\sum_{j=1}^{n-1}a_{j,n}+\sum_{1\leq i<j<n}d_{P^{\prime}}\left( i,j\right) a_{i,j}^{\prime}>F_{A}\left( G\right) .$$ This contradicts the assumption that $F_{A}\left( G\right) $ is maximal. Therefore $G-n=P_{n-1}.$ By symmetry, assume that the vertex sequence of the path $G-n$ is precisely $1,2,\ldots,n-1.$ If $k=1$ or $k=n-1,$ we see that $G=P_{n},$ so let us assume that $1<k<n-1$. To complete the proof we shall show that we can join $n$ to $1$ or to $n-1$ so that $F_{A}\left( G\right) $ will increase. Write $A_{0}$ for the principal submatrix of $A$ in the first $n-1$ rows and note that$$F_{A}\left( G\right) =\sum_{i=1}^{k}\left( k-i+1\right) a_{i,n}+\sum_{i=k+1}^{n-1}\left( i-k+1\right) a_{i,n}+F_{A_{0}}\left( G-n\right) .$$ Next, delete the edge $\left\{ n,k\right\} $ in $G,$ add the edge $\left\{ n,1\right\} ,$ and write $G_{1}$ for the resulting path. Since $F_{A}\left( G\right) $ is maximal, we see that $F_{A}\left( G_{1}\right) \leq F_{A}\left( G\right) .$ From $$F_{A}\left( G_{1}\right) =\sum_{i=1}^{n-1}ia_{i,n}+F_{G-n}\left( A_{0}\right)$$ it follows that, $$\sum_{i=1}^{k-1}\left( k-i+1\right) a_{i,n}+\sum_{i=k}^{n-1}\left( i-k+1\right) a_{i,n}\geq\sum_{i=1}^{n-1}ia_{i,n},$$ and so$$\sum_{i=1}^{k-1}\left( k-2i+1\right) a_{i,n}\geq\left( k-1\right) \left( a_{k,n}+\cdots+a_{n-1,n}\right) .$$ Hence, letting$$S_{1}=-2\sum_{i=1}^{k-1}\left( i-1\right) a_{i,n}$$ we see that $$\left( k-1\right) \left( a_{1,n}+\cdots+a_{k-1,n}\right) +S_{1}\geq\left( k-1\right) \left( a_{k,n}+\cdots+a_{n-1,n}\right) \label{in4}$$ Finally, delete the edge $\left\{ n,k\right\} $ in $G,$ add the edge $\left\{ n,n-1\right\} ,$ and write $G_{2}$ for the resulting path. Since $F_{A}\left( G\right) $ is maximal, we see that $F_{A}\left( G_{2}\right) \leq F_{A}\left( G\right) .$ From $$F_{A}\left( G_{2}\right) =\sum_{i=1}^{n-1}\left( n-i\right) a_{i,n}+F_{G-n}\left( A_{0}\right)$$ it follows that $$\sum_{i=1}^{k}\left( k-i+1\right) a_{i,n}+\sum_{i=k+1}^{n-1}\left( i-k+1\right) a_{i,n}\geq\sum_{i=1}^{n-1}\left( n-i\right) a_{i,n},$$ and so $$\sum_{i=k}^{n-1}\left( 2i-k-n+1\right) a_{i,n}\geq\left( n-k-1\right) \left( a_{1,n}+\cdots+a_{k-1,n}\right) .$$ Hence, letting$$S_{2}=-2\sum_{i=k}^{n-1}\left( n-i-1\right) a_{i,n}$$$$\left( n-k-1\right) \left( a_{k,n}+\cdots+a_{n-1,n}\right) +S_{2}\geq\left( n-k-1\right) \left( a_{1,n}+\cdots+a_{k-1,n}\right) .$$ Comparing this inequality with (\[in4\]), in view of $S_{1}\leq0$ and $S_{2}\leq0,$ we find that $$a_{k,n}+\cdots+a_{n-1,n}=a_{1,n}+\cdots+a_{k-1,n}\text{ \ \ and \ \ \ }S_{1}=S_{2}=0.$$ Hence,$$a_{2,n}=\cdots=a_{k-1,n}=0\text{ \ \ and \ \ }a_{k,n}=\cdots=a_{n-2,n}=0.$$ Since $n-3\geq2$, among the off-diagonal entries of the $n$’th row of $A,$ there are two that are zero, contrary to the hypothesis. Therefore, $G=P_{n},$ completing the induction step and the proof of Theorem \[tht\]. Armed with Theorem \[tht\], we are able the complete the proof of Theorem \[ths\]. \[**Proof of Theorem \[ths\]**\]First we shall prove Theorem \[ths\] if $G$ is a unicyclic graph, i.e., if $G$ has exactly $n$ edges. Thus, let $G$ be a connected unicyclic graph of order $n\geq3.$ It is known that $G\ $contains a single cycle. If $G$ is not the cycle $C_{n\text{ }}$itself, then $G\ $contains a spanning tree $H\ $with maximum degree $\Delta\left( H\right) \geq3;$ thus $H\neq P_{n}.$ Hence, Theorem \[tht\] implies that there is a path $P$ with $V\left( P\right) =V\left( G\right) $ such that $$F_{A}\left( G\right) =\sum_{1\leq i<j\leq n}d_{G}\left( i,j\right) a_{i,j}\leq\sum_{1\leq i<j\leq n}d_{H}\left( i,j\right) a_{i,j}<\sum_{1\leq i<j\leq n}d_{P}\left( i,j\right) a_{i,j}.$$ If $G$ is the cycle $C_{n\text{ }}$itself, let $i,j,k$ be three consecutive vertices along the cycle. The removal of the edge $\left\{ i,j\right\} $ increases the distance between $i$  and $j,$ i.e., $$d_{G}\left( i,j\right) <d_{G-\left\{ i,j\right\} }\left( i,j\right)$$  and on the other hand $$F_{A}\left( G\right) \leq F_{A}\left( G-\left\{ i,j\right\} \right) .$$ If $F_{A}\left( G\right) <F_{A}\left( G-\left\{ i,j\right\} \right) ,$ the theorem is proved, otherwise $F_{A}\left( G\right) =F_{A}\left( G-\left\{ i,j\right\} \right) $ and so $a_{i,j}=0.$ By the same token we obtain $a_{j,k}=0;$ hence among the off-diagonal entries of the $k$’th row of $A$ there are two that are zero, contrary to the hypothesis. So the theorem holds for unicyclic graphs. Finally, note that any connected graph $G$ that is not a tree contains a connected unicyclic spanning subgraph $H$ or is unicyclic itself. Hence, if $G$ is not a tree, then $F_{A}\left( G\right) \leq F_{A}\left( H\right) $ for some connected unicyclic $H,$ and thus there is a path $P$ with $V\left( P\right) =V\left( G\right) $ such that $$F_{A}\left( G\right) \leq F_{A}\left( H\right) <F\left( P\right) .$$ The proof of Theorem \[ths\] is completed. Concluding remarks ================== Results similar to Theorem \[thSI\] have been known for the adjacency matrix, the Laplacian, and the signless Laplacian of a connected graph $G:$ \[[@LoPe73]\]\[t1\]The largest eigenvalue of the adjacency matrix of a connected graph $G$ of order $n$ is minimal if and only if $G$ is a path. \[[@PeGu02]\]\[t2\]The largest eigenvalue of the Laplacian of a connected graph $G$ of order $n$ is minimal if and only if $G$ is a path. \[[@Yan02]\]\[t3\]The largest eigenvalue of the signless Laplacian of a connected graph $G$ of order $n$ is minimal if and only if $G$ is a path. In the light of the present note we would like to raise the following question: **Question.** *Is there a result similar to Theorem* \[ths\] *that implies Theorems* \[t1\], \[t2\], *and* \[t3\]. [99]{} M. Aouchiche and P. Hansen, A signless Laplacian for the distance matrix of a graph, *Les Cahiers du GERAD,* G–2011–78, 2011. M. Aouchiche and P. Hansen, Some properties of the distance Laplacian eigenvalues of a graph, *Les Cahiers du GERAD,* G–2013–28, 2013. M. Aouchiche and P. Hansen, Two Laplacians for the distance matrix of a graph, *Linear Algebra Appl.* **439** (2013), 21–33. B. Bollobás, *Modern Graph Theory,* Graduate Texts in Mathematics, 184, Springer-Verlag, New York (1998). H. Lin and X.Lu, Bounds on the distance signless Laplacian spectral radius in terms of clique number, to appear in *Linear and Multilinear Algebra,* 2014. L. Lovász and J. Pelican, On the eigenvalues of trees, *Periodica Math. Hung.* **3** (1973), 175–182. M. Nath and S. Paul, On the distance Laplacian spectra of graphs, *Linear Algebra Appl.* **460** (2014), 97–110. M. Petrović, I. Gutman, The path is the tree with smallest greatest Laplacian eigenvalue, *Kragujevac J. Math.* **24** (2002), 67–70. S.N. Ruzieh and D.L. Powers, The distance spectrum of the path and the first distance eigenvector of connected graphs, *Linear and Multilinear Algebra* **28** (1990), 75–81. D. Stevanović and A. Ilić, Distance spectral radius of trees with fixed maximum degree, *Electron. J. Linear Algebra.* **20** (2010), 168–179. C. Yan, Properties of spectra of graphs and line graphs, *Appl. Math. J. Chinese Univ. Ser. B,* **17** (2002), 371-371. R. Xing, B. Zhou, and J. Li, On the distance signless Laplacian spectral radius of graphs, *Linear and Multilinear Algebra,* **62** (2014), 1377-1387. [^1]: PEP-COPPE, Universidade Federal do Rio de Janeiro, Brasil; *email: [email protected]* [^2]: Research partially supported by CNPq-Brasil [^3]: Department of Mathematical Sciences, University of Memphis, Memphis TN 38152, USA; *email: [email protected]*

ArXiv

--- abstract: 'Molecules are ubiquitous in natural phenomena and man-made products, but their use in quantum optical applications has been hampered by incoherent internal vibrations and other phononic interactions with their environment. We have now succeeded in turning an organic molecule into a coherent two-level quantum system by placing it in an optical microcavity. This allows several unprecedented observations such as 99% extinction of a laser beam by a single molecule, saturation with less than 0.5 photon, and nonclassical generation of few-photon super-bunched light. Furthermore, we demonstrate efficient interaction of the molecule-microcavity system with single photons generated by a second molecule in a distant laboratory. Our achievements pave the way for linear and nonlinear quantum photonic circuits based on organic platforms.' author: - Daqing Wang - Hrishikesh Kelkar - 'Diego Martin-Cano' - Dominik Rattenbacher - Alexey Shkarin - Tobias Utikal - Stephan Götzinger - Vahid Sandoghdar title: 'Turning a molecule into a coherent two-level quantum system' --- Introduction ============ Molecules provide very compact quantum systems that host well-defined transitions, ranging from the microwave to the ultraviolet domains associated with their rotational, vibrational and electronic states. In addition, these intrinsic mechanical and electronic degrees of freedom can be coupled through various well-defined transitions and selection rules. Indeed, molecular systems have attracted renewed attention within the community of quantum physics both in the gas[@Spaun:16; @Truppe:17] and condensed[@Lidzey:00; @Hakala:09; @Schwartz:11; @Chikkaraddy:16; @Polisseni:16; @Zhang:17; @Skoff:18] phases. In the former case, molecules possess long-lived vibrational levels and well-resolved rotational transitions, but their cooling and trapping are difficult so that access to single molecules has only very recently being explored[@Liu:18]. On the other hand, while addressing single molecules in solids has been feasible with high spatial and spectral resolutions for nearly three decades, a substantial degree of decoherence remains in this system due to phononic couplings[@SMbook]. The ground vibrational level of the electronic excited state ($|e, v=0\rangle$) in a dye molecule can couple to $|g, v=0\rangle$ and $|g, v\neq 0\rangle$ in the ground state following the Franck-Condon principle (see Fig.\[schematics\]a). When embedded in a solid, each of these transitions entails a zero-phonon line (ZPL) and a phonon wing caused by coupling to matrix phonons (Debye-Waller factor). In the case of polycyclic aromatic hydrocarbons (PAH), the ZPL connecting $|g, v=0\rangle$ and $|e, v=0\rangle$ (00ZPL) can be narrowed by about $10^5$ folds to the Fourier limit when cooled to liquid helium temperatures. Nevertheless, the decay of $|e, v=0\rangle$ via $|g, v\ne 0\rangle$ levels and the subsequent fast relaxation of the latter states give rise to decoherence, making phase sensitive and nonlinear quantum operations inefficient[@Pototschnig:11; @Maser:16]. One way to counter this decoherence is to enhance the 00ZPL in a selective manner and, thus, modify the branching ratio out of $|e, v=0\rangle$. In the past decade, there have been many efforts to enhance the radiative properties of molecules by plasmonic nanostructures[@Kuehn:06; @Chikkaraddy:16]. However, the large bandwidth of plasmon resonances does not allow for selective addressing of narrow transitions. To remedy this, one can use optical microcavities to enhance molecular ZPLs by a substantial Purcell factor, $F=\frac{3}{4\pi^2}\frac{Q\lambda^3}{V}\gg1$ [@book-Berman94]. First attempts in this direction have indeed been reported[@Norris:97; @Steiner:07; @Chizhik:09; @Wang:17], but the results fall short of notable enhancements. A successful laboratory realization needs to consider and tackle several technical issues, especially in regard to the microcavity design. While monolithic microcavities are fairly robust, they are difficult to tune and not always compatible with the material of the quantum emitter. Open Fabry-Perot resonators, on the other hand, are difficult to stabilize in the cryostat but are conveniently adjustable and can be more easily combined with various materials. In this work, we employ an open, tunable and scannable Fabry-Perot microcavity with very small mode volume ($V$) and moderate quality factor ($Q$) [@Toninelli:10; @Kelkar:15; @Wang:17]. Experimental ============ The method of choice for selecting single quantum emitters embedded in a solid is to exploit the inherent spectral inhomogeneity of such a system[@SMbook]. By operating at low enough temperatures, the homogenous linewidths of individual emitters become so narrow that they no longer overlap. Thus, a narrow-band laser beam can address the 00ZPLs of the various emitters located in the illumination volume one by one. In our experiment, we used dibenzoterrylene (DBT) of the PAH family embedded in a thin anthracene (AC) crystal (see Fig.\[schematics\]a). The 00ZPL of DBT:AC lies in the interval $783-785$nm and can become as narrow as its Fourier limit of about 40MHz at $T \lesssim 4$K [@Nicolet:07]. To produce a Fabry-Perot cavity, we fabricated a curved micromirror at the end of an optical fiber and used a planar mirror [@Kelkar:15; @Wang:17], both coated with a dielectric multilayer (see Fig.\[schematics\]b,c). The anthracene-filled microcavity exhibited an optical length of 4.7$\mu$m, cavity mode volume of $4.4\lambda^3$, finesse of 19,000, and $Q$-factor of 230,000. When placed in our helium exchange gas cryostat at 4K, residual vibrations broadened the line. This resulted in $Q=120,000$, deduced from the full width at half-maximum (FWHM) $\kappa/2\pi=3.3$GHz of a Voigt profile, which was composed of a Lorentzian resonance of FWHM=1.7GHz and a Gaussian broadening of FWHM=2.3GHz. In Fig.\[schematics\]d, we display an overview of the experimental arrangement. ![ **a,** An optical microscope image of a thin anthracene (AC) crystal, its molecular structures together with that of DBT, and the energy level scheme of the latter. **b,** Sketch of molecules embedded in a thin AC crystal placed in a microcavity. **c,** Electron microscope cross sectional image of the multilayer coating of the curved micromirror with radius of curvature of $10\,\mu$m after being cut by a focused ion beam. The apparent surface roughness is caused by the metallic coating necessary for electron microscopy and is absent on the mirror used in the measurements. **d,** Overview of the experimental setups in two different laboratories.[]{data-label="schematics"}](figures/schematics.pdf){width="12"} Results ======= Transmission and reflection of a laser beam ------------------------------------------- The presence of an emitter inside a cavity modifies the interference of the fields that result from the reflections between its mirrors. To probe the intracavity field, we exploit the cross-polarized reflection (CPR) that is generated by the birefringence of our cavity, providing an equivalent measure to a transmission recording[@Wang:17]. The blue symbols in Fig.\[resonant-spectra\]a display a CPR spectrum in the absence of coupling to a molecule, and the blue solid curve shows a fit to these data using a Voigt profile. The black symbols in this figure represent a reference CPR signal when the cavity frequency was detuned by about 20GHz. To examine the effect of a single molecule on the optical response of the microcavity, we tuned the resonance of the latter through the inhomogeneous band of DBT:AC and searched for the signature of molecular resonances directly in the cavity CPR spectrum while scanning the laser frequency. The magenta symbols in Fig.\[resonant-spectra\]a present an example, where the cavity transmission drops by 99% when it becomes resonant with a single molecule. In Fig.\[resonant-spectra\]b, we also display a direct transmission measurement recorded through the micromirror (see Fig.\[schematics\]d). All features in Figs.\[resonant-spectra\]a and \[resonant-spectra\]b agree, as confirmed by the high quality of the fits using common parameters. However, the signal-to-noise ratio (SNR) is lower in (b) because the numerical aperture of the single-mode fiber holding the micromirror does not match the cavity mode, thus, resulting in a weak signal. Next, we investigated the remarkably large effect of a single molecule on an incident laser beam further by measuring the direct reflection of the system. The blue symbols in Fig.\[resonant-spectra\]c show the central part of the cavity resonance measured in the absence of molecular coupling. We note that the cavity resonance does not dip to zero due to imperfect mode matching and remaining vibrations. Nevertheless, the magenta symbols show that when the cavity is brought into resonance with a single molecule, the reflection dip vanishes. We shall present a quantitative discussion of the spectra in Fig. \[resonant-spectra\]a-c as well as the solid theory curves used to fit them shortly. ![ **a,** Transmission spectra of the cavity with (magenta) and without (blue) the molecular contribution in units of counts per second (cps). The measurements were performed in cross-polarized mode. Black symbols show the intensity of the laser beam when the cavity resonance was detuned about 20GHz, acting as a reference level. A single molecule interrupts the cavity transmission by 99%. **b,** Same as in (a) but for the direct transmission through the micromirror fabricated at the end of the optical fiber. **c,** Reflection spectra recorded on the same molecule as in (a,b). **d,** Fluorescence excitation spectrum of the same molecule recorded far detuned from the cavity resonance. The inset shows the second-order intensity autocorrelation function, verifying that this light is antibunched. See text for the explanation of the theoretical fits (solid curves).[]{data-label="resonant-spectra"}](figures/resonant-spectra){width="7.75"} Cavity QED modifications and spectral analysis ---------------------------------------------- The data in Fig.\[resonant-spectra\]a-c let us deduce the FWHM of the molecular 00ZPL to be $604\pm21$MHz under coupling to the microcavity. To determine the linewidth of the very same molecule without the influence of the cavity, we detuned the resonance of the latter and recorded a red-shifted fluorescence signal from the molecule as a function of the excitation frequency. The outcome shown in Fig.\[resonant-spectra\]d reveals FWHM=$44\pm5$MHz, which is in the range of the values for the bulk DBT:AC system[@Nicolet:07]. The weaker signal in this case stems from the difficulty of extracting the broad fluorescence through the higher-order transverse modes of the cavity. Photon antibunching of the fluorescence signal recorded via intensity autocorrelation ($g^{(2)}(\tau)$) confirms that it originates from only one molecule (see inset of Fig.\[resonant-spectra\]d). The linewidth of an unperturbed molecule can be expressed as $\gamma^0=\gamma^0_\text{zpl}+\gamma_\text{red}$, where $\gamma^0_\text{zpl}$ stands for the decay rate of the excited state $|e, v=0\rangle$ into the 00ZPL channel, and $\gamma_\text{red}$ denotes the contributions of all red-shifted emission, including phonon wings and vibrational decay paths. When the cavity is resonant with the 00ZPL, the component of $\gamma^0_\text{zpl}$ emitted into the cavity mode is enhanced by the Purcell factor $F$, yielding $\gamma'_\text{zpl}=(1+F)\gamma^0_\text{zpl}$. Hence, considering that $\gamma_\text{red}\approx2\gamma^0_\text{zpl}$ for DBT:AC[@Trebbia:09], we can write the modified decay rate of the excited state as $\gamma'=\gamma_\text{red}+\gamma'_\text{zpl}=(3+F)\gamma^0_\text{zpl}$. We, thus, deduce from our experimental findings of $\gamma^0/2\pi= 44\pm5$ MHz and $\gamma'/2\pi = 604\pm21$ MHz a Purcell factor of $F=38\pm 5$. A very useful measure for the efficiency of emitter-cavity coupling is the $\beta$-factor defined as the ratio of the power emitted into the cavity mode and the total emitted power[@Petermann:79]. The $\beta$-factor associated with an ideal two-level atom can be computed as $\frac{F}{F+1}$ and would correspond to $97.4\pm0.3\%$ for our cavity. To assess the overall degree of coherence for the resonant interaction between a DBT molecule and an incoming light field, however, we also have to account for losses to the red-shifted channels. Therefore, we arrive at $\beta=\frac{\gamma^0_\text{zpl}F}{\gamma'}=\frac{F}{F+3}=93\%$. The strong modification of the molecular emission on the 00ZPL changes its branching ratio $\alpha$, defined as the fraction of the power in the 00ZPL to the overall emission from the excited state. Our results demonstrate a modification from $\alpha\sim33\%$ for bulk DBT:AC to $\alpha'=\frac{\gamma'_\text{zpl}}{\gamma'}=\frac{F+1}{F+3}=95\%$ in the cavity. This implies that we have successfully converted a molecule to a two-level quantum system to within 95%. The obtained high values of $\alpha$ and $\beta$ have immediate consequences for the efficiency of coherent linear and nonlinear processes at the single-molecule level[@Pototschnig:11; @Maser:16]. To investigate the radiative modifications further, we took advantage of the axial tunability of our microcavity and recorded a series of CPR spectra at different molecule-cavity frequency detunings. Figure\[detuned-spectra\]a demonstrates the evolution of the molecule-cavity spectral modifications, providing a wealth of quantitative data and a thorough comparison between experiment and theory. In a first simple approach, we fit the observed Fano-like dispersive line shapes using a generalized Lorentzian function. Figure\[detuned-spectra\]b nicely traces the linewidth of the molecular resonance (left vertical axis) and the corresponding Purcell factor (right vertical axis) as a function of the cavity frequency detuning. The data in Fig.\[detuned-spectra\]a can also be analyzed by considering a rigorous theory that treats the interaction of an incoming field with both the molecule and the cavity on the same footing[@Auffeves:07]. The solid curves in Figs.\[resonant-spectra\]a-c and \[detuned-spectra\]a show fits to the experimental data with excellent agreement using such a model. ![ **a,** Series of transmission spectra for different molecule-cavity frequency detunings. **b,** Linewidth of the molecular resonance and measured Purcell factor as a function of molecule-cavity frequency detuning. **c,** Frequency shift of the molecular resonance (the modified Lamb shift) as a function of molecule-cavity frequency detuning. **d,** Positions of the peaks in the transmission spectrum as a function of molecule-cavity frequency detuning. The solid curves show fits obtained from rigorous theoretical calculations. **e-i,** Calculated transmission (green) and fluorescence (red) spectra of a coupled system for decreasing degree of cavity loss. The legend in each figure denotes the cavity FWHM in terms of our experimental linewidth $\kappa$. The symbols in (g) represent the experimental CPR spectrum. The fit quality here is not as good because the contribution of vibrations to a Voigt profile are not taken into account in the calculated spectra. The dotted vertical lines displays the positions of the maxima in the transmission spectra.[]{data-label="detuned-spectra"}](figures/detuned-spectra){width="11"} Next, we analyzed the influence of the cavity coupling on the center frequency of the molecular resonance. Figure\[detuned-spectra\]c plots the latter as a function of the cavity-molecule frequency detuning and reveals frequency shifts by up to about $\pm150$MHz towards blue or red, depending on the sign of the detuning. This can be interpreted as a change in the contribution of vacuum fluctuations to the absolute value of the 00ZPL, i.e. of the Lamb shift[@book-Berman94; @Heinzen:87]. Perhaps somewhat nonintuitively, the correction disappears when the cavity is resonant with the molecular line and peaks at the largest slope of the resonance profile. This behavior is similar to the well-known AC stark shift proportional to the laser-atom detuning[@cohen-book]. To place our parameters in the context of weak and strong coupling regimes, in Fig.\[detuned-spectra\]e-i we present calculated transmission (green) and fluorescence (red) spectra for various cavity linewidths. While the fluorescence spectra clearly show the onset of a line splitting for higher finesse cavities, the splitting between the maxima in the transmission spectrum is nearly independent of the cavity finesse (see dotted vertical lines). The symbols in Fig.\[detuned-spectra\]g show that the experimental data correspond to the transitional regime where the spectrum changes its character from a molecular extinction dip on a broad cavity resonance to one with two split polaritonic resonance profiles at par. The fluorescence spectrum in this region is shortly before bifurcation into two maxima, a signature of strong coupling[@book-Berman94]. A quantitative measure for the onset of strong coupling can be formulated by the exceptional point, where the cavity-molecule coupling rate $g$ satisfies $g_{\rm ep}=\lvert \kappa-\gamma \rvert/4$ [@Choi:10]. To determine $g$ for our experiment, in Fig.\[detuned-spectra\]d we plot the frequencies of the two maxima that arise in the spectra of Fig.\[detuned-spectra\]a. The splitting at zero detuning directly equals $2g$, yielding $g=0.79 \pm 0.3$GHz. Comparison of this value with $g_{\rm ep}=0.82$GHz computed for our system confirms that our experiment is situated right at the onset of strong coupling. A convenient parameter that connects $g$, $\kappa$ and $\gamma$ is the cooperativity factor $C=\frac{4g^2}{\kappa\gamma}$. For a two-level atom, $C$ and $F$ are equivalent, but one has to distinguish between them when dealing with emitters that support multichannel decay. To this end, our measured value of $F=38$ reports on the enhancement of $\gamma^0_{\rm zpl}$ as a well-defined dipolar transition into a single mode of a cavity. The expression of $C$, on the other hand, reports on the degree of coherence in the interaction of the molecule as a whole with a photon in the cavity. Thus, to estimate $C$, we use the total decay rate $\gamma^0$ so as to account for the internal loss of coherence through the red-shifted emission paths, arriving at $C=12.7$. We note that, in fact, our microcavity values of $Q$ and $V$ let us expect a much higher Purcell factor $F\sim350$ and, thus, even stronger couplings than reported here. We attribute the discrepancy between the measured and predicted Purcell factors to the suboptimal position and orientation of the molecule with respect to the cavity electric field. Phase Shift ----------- The phase shift imprinted by a quantum emitter on a light beam can report on the state of the emitter in a nondestructive fashion. Previous experiments have demonstrated phase shifts of about three degrees applied to a focused laser beam by single molecules in a crystal[@Pototschnig:11]. Considering the high cooperativity and coupling efficiency of our system, we should now expect a much larger phase shift. To explore this, we examined the CPR of a laser beam from the cavity, following the protocol described in Ref.[@Pototschnig:11]. The black symbols in Fig.\[phase\] show the phase shift affected by the microcavity alone as the laser frequency was scanned across its resonance in the absence of a molecule. The red symbols in that figure display the recorded phase shift of the laser beam under the influence of a single molecule. The solid and dashed curves signify theoretical fits with and without the consideration of power broadening, respectively, allowing us to deduce phase shifts up to $\phi=\pm 66^\circ$. ![ Measured phase shifts of a laser beam after interacting with a cavity without (black symbols) and with (red symbols) a single molecule. The curves show the theoretical fits.[]{data-label="phase"}](figures/phase){width="10"} Photon Bunching and single-photon nonlinearity ---------------------------------------------- The photon statistics of a laser beam and of a quantum emitter take on very different forms, characterized by intensity autocorrelations $g^{(2)}(\tau)=1$ and $g^{(2)}(0)=0$, respectively. We now show that an efficient coupling between laser light and a molecule can result in highly nontrivial statistics of the emerging photons[@Rice:88]. Parts a,c,e, and g of Fig.\[bunching\] display $g^{(2)}(\tau)$ measurements on a laser beam after interaction with the molecule-cavity system at different frequency detunings (see Figs.\[bunching\]b,d,f). The outcome $g^{(2)}(\tau)=1$ in Fig.\[bunching\]a reveals that the lower polariton branch in this case (see Fig.\[bunching\]b) has a laser-like nature. The measurement shown in Fig.\[bunching\]c, on the other hand, presents a nontrivial case of antibunching for the molecule-like branch of the spectrum in Fig.\[bunching\]d. This antibunching results from the nonclassical interference of the molecular scattering with the intracavity field and provides evidence for the dipole quadrature squeezing, which was recently detected in cavities[@Ourjoumtsev2011] and in free space[@Schulte2015]. ![ **a, c, e,** Intensity autocorrelation $g^{(2)}{(\tau)}$ of the cross-polarized reflection (CPR) of a laser beam from the molecule-cavity composite for different molecule-cavity and laser frequencies. **b, d, f,** The CPR spectra corresponding to the measurements shown in (a), (c), (e). The orange dashed lines denote the laser frequency in each case. Molecule-cavity detuning is displayed as legend in each plot. **g,** Photon bunching corresponding to $g^{(2)}{(0)}=21$ at the molecular resonance in the situation of Fig.\[resonant-spectra\]. The red curves in (a-g) show the theoretical fits. **h,** Theoretical predictions of $g^{(2)}{(0)}$ as a function of cooperativity $C$. The dashed line depicts the experimental parameter used in (g).[]{data-label="bunching"}](figures/bunching){width="12"} Figure\[bunching\]e shows that tuning the laser frequency by a bit more than a linewidth (see Fig.\[bunching\]f) changes the behaviour completely to a *bunching* effect. This phenomenon stems from the selective scattering of single-photon components from the Poisson distribution of photons in the incident laser beam[@Rice:88], yielding a super-bunched few-photon state[@book-Ficek]. As displayed in Fig.\[bunching\]g, this effect is maximized at the centre of the resonant molecule-cavity spectrum shown in Fig.\[resonant-spectra\]a. Fitting the data by the theoretical model described in Ref.[@Carmichael:91] lets us deduce $g^{(2)}(0)=21$, which is among the largest photon bunchings reported to date for a single emitter[@Bennett:16; @Snijders:16]. In fact, calculations in Fig.\[bunching\]h show that for $C=12.7$ one expects $g^{(2)}(0)$ to reach as high as $2.5\times10^4$ if the molecule, cavity and laser frequencies coincide. The discrepancy with the measured value of 21 is due to the limited detector response time of 50ps and the residual background light. The underlying mechanism of the phenomena observed above is that the molecule responds to only one photon at a time. This feature is also responsible for the intrinsic nonlinearity of an atom or molecule, which in turn leads to saturation as the excitation power is increased. We studied the nonlinear response of the cavity-coupled molecule by examining the extinction signal[@Wrigge:08]. We find that we reach the saturation parameter of $S=1$ for a very low power of 420pW coupled to the cavity, corresponding to only 0.44 photons per excited state lifetime of 264 ps. This result indicates that the operation regime of our experiment not only provides nearly perfect coupling in the weak excitation limit, but it also opens doors for efficient few-photon nonlinear operations[@Chang:14; @Maser:16]. Single-Photon Reflection ------------------------ The ultimate frontier of light-matter interaction requires efficient coupling of a single photon and a single quantum emitter. To demonstrate such a “gedanken" experiment, we used a second molecule located in a different laboratory (see Fig.\[schematics\]d) as a source of narrow-band single photons (see Ref.[@Rezus:12] for details). The resulting stream of 30,000 photons per second with a FWHM linewidth of 41MHz was coupled to a single-mode fiber and sent to the laboratory housing the microcavity. The magenta symbols in Fig.\[single-photons\] present the reflection spectrum of this single-photon stream when the cavity was tuned to resonance with the 00ZPL of the “target" molecule. While the count rate and the shot-noise-limited SNR are lower, the signal reproduces our findings in Fig.\[resonant-spectra\]c, verifying that we also reach a high efficiency in coupling a molecule to single photons. We note in passing that one of the challenges in this experiment concerns tuning the frequency of single photons, which we realized via Stark effect on the “source" molecule. In future, the degree of mastery demonstrated here can be combined with pulsed excitation of the source molecule and extended to the coupling of two or more photons to a single molecule[@Chang:14; @Maser:16]. Such an experiment establishes a platform for nonlinear quantum optics at its most fundamental level for the realization of gates and for quantum information processing[@Kok2010]. ![ Reflection spectrum of the molecule-cavity system when single photons from another molecule in a different laboratory were impinged on it (magenta). Single photon budget: 6 kcps (kilo counts per second) out of the fiber in the microcavity lab, 3 kcps incident onto the cavity, 500 cps on the final detector. The black symbols show a spectrum recorded when the cavity resonance was detuned by 6GHz. Note that our freedom to change the frequency of the single photons via Stark effect on the “source" molecule is limited compared to the case of a laser beam in Fig.\[resonant-spectra\]. The small dip that is slightly shifted from the origin denotes the interference of the photons scattered by the target molecule with the component directly reflected at the flat mirror.[]{data-label="single-photons"}](figures/single-photons){width="9.5"} Discussion and outlook ====================== The large phase shift, nonlinearity at the single-photon level and strongly nonclassical photon statistics demonstrated in this work give access to a range of quantum functionalities such as photon sorting and gates[@Kok2010; @Ralph:15] in organic materials. In a next step, chip-based ring resonators[@Rotenberg:17] and nanoguides[@Tuerschmann:17] will apply these opportunities to nanophotonic circuits. Having reached a highly efficient level of interaction between single photons and single molecules, one can then devise novel linear and nonlinear cooperative effects and polaritonic states, where a controlled number of molecules and photons are coupled via a common photonic mode along the circuit [@Diniz:11; @Haakh:16]. The practical implementation of these concepts would particularly benefit from the use of polymer media instead of organic crystals for device fabrication. Indeed, the Purcell enhanced radiative rates achieved in our work already compete with and dominate phonon-induced dephasing rates in polymers, which lie in the range of 0.1-1 GHz[@Walser:09]. In addition to their immediate potential for large-scale organic quantum networks, we expect the selective modification of molecular rates demonstrated in this work to find applications in cooling and trapping of molecules in the gas phase, where closed transitions are desirable [@Kozyryev:17] and in control of molecular photochemical processes such as photochromic switching[@Schwartz:11]. Acknowledgments {#acknowledgments .unnumbered} =============== We thank Jan Renger, Anke Dutschke, and Eduard Butzen for the fabrication of micromirrors, Maksim Schwab for the construction of the mechanical components for the cryostats, and Lothar Meier for help with electronics to lock the cavity. This work was supported by the Max Planck Society. [46]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\ 12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} , , ,  and , eds., @noop [**]{} (, ) @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} , ed., @noop [**]{} (, , ) @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [**]{} (, , ) @noop [****,  ()]{} @noop [****,  ()]{} @noop [****, ()]{} @noop [****,  ()]{} @noop [**]{} (, ) @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****, ()]{} @noop [**]{} (, ) @noop [****,  ()]{} [****,  ()](\doibase 10.1364/OE.25.005397) @noop [****, ()]{} @noop [****,  ()]{} [****,  ()](\doibase 10.1103/PhysRevA.94.053840) @noop [****,  ()]{} [****,  ()](\doibase 10.1103/PhysRevLett.118.173201)

ArXiv