Datasets:

ArmelR

/

the-pile-splitted

like 18

--- abstract: 'The background dynamical evolution of a universe filled with matter and a cosmological scalar field is analyzed employing dynamical system techniques. After the phenomenology of a canonical scalar field with exponential potential is revised, square and square root kinetic corrections to the scalar field canonical Lagrangian are considered and the resulting dynamics at cosmological distances is obtained and studied. These noncanonical cosmological models imply new interesting phenomenology including early time matter dominated solutions, cosmological scaling solutions and late time phantom dominated solutions with dynamical crossing of the phantom barrier. Stability and viability issues for these scalar fields are presented and discussed.' author: - Nicola Tamanini title: Dynamics of cosmological scalar fields --- Introduction ============ Since the Nobel Prize winning discovery of a current cosmological phase of accelerated expansion was made in 1998 [@Riess:1998cb; @Perlmutter:1998np], the theoretical models advanced to describe this phenomenon quickly multiplied in the literature. The simplest among these is the straightforward addition of a positive cosmological constant to the Einstein field equations. Although this model fits all astronomical observations, it is in tension with particle physics prediction (the cosmological constant problem [@Weinberg:1988cp; @Martin:2012bt]) and cosmological considerations (the coincidence problem [@Zlatev:1998tr]). A way to alleviate these problems consists in letting the cosmological constant to be dynamical. This implies the introduction of some cosmological field capable of reproducing the late time accelerated behavior mimiking in this way the effects of a cosmological constant. Any physical entity which at cosmic distances provide an accelerated expansion at late times is commonly called [*dark energy*]{}. The simplest field having these properties is a canonical scalar field with a potential. Dark energy models of this kind go under the name of [*quintessence*]{} and have been largely studied in the literature [@Copeland:2006wr; @Tsujikawa:2013fta]. Scalar fields play an important role in cosmology since they are sufficiently simple to handle and sufficiently complicated to produce non-trivial dynamics. They are not only employed to model dark energy, but also to characterize inflation [@Liddle:2000cg], dark matter [@Magana:2012ph], unified dark models [@Bertacca:2010ct] and other cosmological features. For dark matter phenomenology it is usually required a vanishing pressure and that the speed of sound of adiabatic perturbations is sufficiently small to allow the formation of clusters. Scalar field models proposed to unify dark matter and dark energy must thus have a dynamical equation of state evolving from dust-like to dark energy-like behavior, which can be achieved with a non canonical scalar field [@Bertacca:2010ct]. Generalizations and modifications of the canonical scalar field Lagrangian can also lead to more complex cosmological predictions. Extended models where the scalar field Lagrangian is a general function of both the scalar field $\phi$ and its kinetic term, are known as $k$-essence theories [@ArmendarizPicon:2000ah]. Within this framework it is possible to obtain not only the standard dark energy evolution but also the so called [*phantom*]{} regime and quintessence to phantom transition, though fatal problems always arise at the level of perturbations [@Vikman:2004dc; @Zhao:2005vj; @Caldwell:2005ai]. A phantom scalar field is identified by an equation of state (EoS) with a negative pressure bigger than the energy density. In other words, for a scalar field EoS $p_\phi=w_\phi\rho_\phi$, the phantom regime is identified by the condition $w_\phi<-1$, which seems to be slightly favoured by astronomical observations even after Planck [@Xia:2013dea; @Novosyadlyj:2013nya]. The first and simplest model capable of achiving such condition consists in flipping the sign of the kinetic term of a canonical scalar field [@Caldwell:1999ew]. However in this model the scalar field EoS parameter is never greater than $-1$ creating problems at early times where dark matter with vanishing pressure must dominate. Scalar fields which can cross the phantom barrier at $w_\phi=-1$ are usually dubbed [*quintom*]{} models and imply either the use of extended Lagrangians, generally instable, or of two different scalar fields [@Caldwell:2005ai; @Cai:2009zp]. The present work is devoted to study the background dynamical evolution of different scalar field models. Dynamical system techniques are employed to fully determine the solutions of the cosmological equations. Suitable dimensionless variables are introduced following [@Copeland:1997et] and the phase space dynamics is analyzed using numerical methods. Canonical and noncanonical scalar field Lagrangians are presented and their cosmological implications are discussed. The complete cosmological background dynamics of two specific noncanonical scalar fields is obtained showing that interesting phenomenology, such as early time matter dominated solutions, scaling solutions, late time phantom acceleration, super-stiff and phantom transition eras, can be achieved. The paper has the following structure. In Sec. \[sec:canonical\] the canonical scalar field will be largely discussed. Its features and cosmological dynamics will be presented and analyzed in depth and the notation and conventions needed for the following sections will be introduced. In Sec. \[sec:noncanonical\] noncanonical scalar field Lagrangians will be considered. Perturbations instabilities will be examined and the analysis will focus on models where dynamical system techniques can be sucessfully applied. Sec. \[sec:square\] and Sec. \[sec:phantom\] will then be dedicated to the study of square and square root kinetic corrections to the canonical scalar field Lagrangian. For these simple models the full dynamical features can be obtained and the background cosmological evolution can be determined for any initial condition. The analysis of these two sections will show that a rich phenomenology can be obtained with these extended scalar fields. Finally results and conclusions will be discussed in Sec. \[sec:concl\] The Canonical Scalar Field {#sec:canonical} ========================== In this section we review the cosmological dynamics of a canonical scalar field following the analysis first performed in [@Copeland:1997et]. This will serve as an introduction to the dynamical system techniques one can apply in order to completely determine the cosmological evolution of specific models. Moreover this section will be helpful in defining notation and conventions. The action of a minimally coupled canonical scalar field is given by $$\begin{aligned} S = \int d^4x\sqrt{-g} \left[\frac{R}{2\kappa^2}+\mathcal{L}_\phi +\mathcal{L}_m\right] \,, \label{014}\end{aligned}$$ where $g$ is the determinant of the metric, $R$ is the Ricci scalar, $\kappa^2=8\pi G/c^4$, $\mathcal{L}_m$ is the matter Lagrangian, and the scalar field Lagrangian is defined as $$\begin{aligned} \mathcal{L}_\phi = -\frac{1}{2}\partial\phi^2-V(\phi) \,, \label{012}\end{aligned}$$ with $\partial\phi^2=\partial_\mu\phi\partial^\mu\phi$ and $V$ a general potential for $\phi$. The variation with respect to $g_{\mu\nu}$ produces the following gravitational equations $$\begin{aligned} G_{\mu\nu} = \kappa^2 \left(T_{\mu\nu}+\partial_\mu\phi\partial_\nu\phi -\frac{1}{2}g_{\mu\nu}\partial\phi^2 -g_{\mu\nu}V\right) \,, \label{001}\end{aligned}$$ where $G_{\mu\nu}=R_{\mu\nu}-1/2g_{\mu\nu}R$ is the Einstein tensor and $T_{\mu\nu}$ the matter energy momentum tensor. The variation with respect to $\phi$ gives the Klein-Gordon equation $$\begin{aligned} \Box\phi-\frac{\partial V}{\partial\phi}=0\,, \label{002}\end{aligned}$$ with $\Box\phi=\nabla_\mu\nabla^\mu\phi$. In what follows we will analyze the background cosmological evolution of this model. The metric tensor will be assumed to be of the Friedmann-Robertson-Walker (FRW) type with vanishing spatial curvature $$\begin{aligned} g_{\mu\nu} = \mbox{diag}(-1,a(t)^2,a(t)^2,a(t)^2) \,, \label{015}\end{aligned}$$ with $a(t)$ the scale factor, while the scalar field is taken to be spatially homogeneous $\phi=\phi(t)$. The matter energy-momentum tensor will be of the perfect fluid form with $\rho(t)$ and $p(t)$ its energy density and pressure, respectively. A linear EoS $p=w\rho$, with $w$ the EoS parameter ranging from $0$ (dust) to $1/3$ (radiation), will be assumed. With these assumptions, from the gravitational equations (\[001\]) we obtain the Friedmann constraint $$\begin{aligned} 3H^2=\kappa^2 \left(\rho +\frac{1}{2}\dot\phi^2+V\right) \,, \label{003}\end{aligned}$$ and the acceleration equation $$\begin{aligned} 2 \dot H +3H^2 = -\kappa^2 \left(p+\frac{1}{2}\dot\phi^2-V\right) \,, \label{005}\end{aligned}$$ where $H=\dot a/a$ is the Hubble parameter and an overdot denotes differentiation with respect to the time $t$. On the other hand the scalar field equation (\[002\]) gives $$\begin{aligned} \ddot\phi +3H\dot\phi + \frac{\partial V}{\partial\phi} = 0 \,. \label{006}\end{aligned}$$ The energy density and pressure of the canonical scalar field are given by $$\begin{aligned} \rho_\phi &= \frac{1}{2}\dot\phi^2+V \,,\label{052}\\ p_\phi &= \frac{1}{2}\dot\phi^2-V \,,\end{aligned}$$ and its EoS parameter, defined as the ratio between its pressure and energy density, is $$\begin{aligned} w_\phi = \frac{p_\phi}{\rho_\phi} =\frac{\frac{1}{2}\dot\phi^2-V}{\frac{1}{2}\dot\phi^2+V} \,. \label{032}\end{aligned}$$ For $V\gg \dot\phi^2$ this approaches a cosmological constant EoS with $w_\phi=-1$, while for $V\ll \dot\phi^2$ this describe a stiff fluid with $w_\phi=1$. At this point, following [@Copeland:1997et], we introduce new dimensionless variables as $$\begin{aligned} x^2 = \frac{\kappa^2 \dot\phi^2}{6H^2}\,,\quad y^2=\frac{\kappa^2V}{3H^2}\,, \quad \sigma^2=\frac{\kappa^2\rho}{3H^2} \,. \label{020}\end{aligned}$$ These variables are largely employed in scalar field cosmology since not only allow to rewrite Eqs. (\[003\])-(\[006\]) as an autonomous system of equations, but can also be generalized in different contexts such as nonminimally coupled scalar fields [@Amendola:1999qq; @Hrycyna:2013yia; @Wei:2011yr; @Xu:2012jf], tachyons [@Aguirregabiria:2004xd], Galileons [@Leon:2012mt], phantom and quintom cosmology [@Chen:2008ft; @Lazkoz:2006pa], $k$-essence [@Yang:2010vv; @DeSantiago:2012nk], modified gravity [@Tamanini:2013ltp], three-form cosmology [@Koivisto:2009fb; @Boehmer:2011tp] and dark energy models coupled to dark matter [@Billyard:2000bh; @Boehmer:2008av; @Boehmer:2009tk]. With the variables (\[020\]) the Friedmann constraint (\[003\]) becomes $$\begin{aligned} 1=\Omega_m+\Omega_\phi=\sigma^2+x^2+y^2 \,, \label{022}\end{aligned}$$ where the relative energy densities are defined as $$\begin{aligned} \Omega_m = \frac{\kappa^2 \rho}{3H^2} \quad\mbox{and}\quad \Omega_\phi=\frac{\kappa^2 \rho_\phi}{3H^2} \,. \label{055}\end{aligned}$$ Eq. (\[022\]) can be used to replace $\sigma^2$ in favour of $x^2$ and $y^2$. This implies that the only dynamical variables of the system of equations will be $x$ and $y$. Also, since $\sigma^2\geq 0$ due to the assumption $\rho\geq 0$, the constraint $$\begin{aligned} x^2+y^2\leq 1 \,,\end{aligned}$$ will always hold. If in addition one assumes the potential energy $V$ to be greater than zero, then $y\geq 0$ and the phase space of the variables $(x,y)$ reduces to the upper half unit disk. At this point it is possible to convert the cosmological equations into an autonomous system of equations if one further specifies the potential $V$. If this is exponential, for example $$\begin{aligned} V(\phi) = V_0 e^{-\lambda\kappa\phi} \,, \label{004}\end{aligned}$$ with $V_0>0$ and $\lambda$ arbitrary parameters, then the phase space will remain two dimensional. If instead one choses a power-law potential, then the phase space becomes three dimensional and the new variable $$\begin{aligned} z = -\frac{1}{\kappa V} \frac{\partial V}{\partial\phi} \,,\end{aligned}$$ needs to be introduced. In this work we will only consider exponential potential of the kind (\[004\]). With this assumption the acceleration equation (\[005\]) and the scalar field equation (\[006\]) lead to the two dimensional autonomous system $$\begin{aligned} x'&=\frac{3}{2} \left[\sqrt{\frac{2}{3}} \lambda y^2- (w-1) x^3- x (w+1) \left( y^2-1\right)\right]\,,\label{010} \\ y'&=-\frac{3}{2} y \left[(w-1) x^2+(w+1) \left(y^2-1\right)+\sqrt{\frac{2}{3}} \lambda x\right]\,, \label{011}\end{aligned}$$ where a prime denotes differentiation with respect to $d\eta=Hdt$ and the variables $x$ and $y$ are functions of the dimensionless time parameter $\eta=\ln a$. Note that the dynamical system (\[010\])-(\[011\]) is invariant under the transformation $y\mapsto -y$, so even if we drop the $V>0$ assumption the dynamics on the negative $y$ half-plane would be a copy of the positive $y$ region. Note also that we are assuming $H>0$ in order to describe an expanding universe. However the dynamics of a contracting universe ($H<0$) would have the same features of our analysis in the negative $y$ plane switching the direction of time because of the $y\mapsto-y$ symmetry. On the other hand the dynamical system (\[010\])-(\[011\]) is also invariant under the symultaneous transformation $$\begin{aligned} \lambda\mapsto-\lambda \quad\mbox{and}\quad x\mapsto-x \,, \label{031}\end{aligned}$$ which shows that opposite values of $\lambda$ lead to the same dynamics after a reflection over the $y$-axis. The acceleration equation (\[005\]) gives also $$\begin{aligned} \frac{\dot H}{H^2}=\frac{3}{2} \left[(w-1) x^2+(w+1) \left(y^2-1\right)\right] \label{008} \,,\end{aligned}$$ which at any fixed point $(x_*,y_*)$ of the phase space can be solved for $a$ to give $$\begin{aligned} a\propto (t-t_0)^{2/[3(w+1)(1-x_*^2-y_*^2)+2x_*^2]} \,, \label{007}\end{aligned}$$ where $t_0$ is a constant of integration. This corresponds to a power law solution, i.e. a solution for which the scale factor $a$ evolves as a power of the cosmological time $t$. If $x=0$ and $y=0$ the universe is matter dominated and its evolution coincides with the standard $w$-dependent scaling solution. If $x=0$ and $y=1$ the denominator of (\[007\]) vanishes and the universe undergoes a de Sitter expansion as can be seen from (\[008\]) which forces $H$ to be constant. An effective EoS parameter $w_{\rm eff}$ can now be defined rewriting (\[007\]) as[^1] $$\begin{aligned} a\propto (t-t_0)^{2/[3(1+w_{\rm eff})]} \,, \label{009}\end{aligned}$$ and corresponds to the EoS parameter of an effective fluid sourcing the gravitational equations, or in other words to an effective matter energy-momentum tensor. Comparing with (\[007\]) we find $$\begin{aligned} w_{\rm eff} = x_*^2-y_*^2+w(1-x_*^2-y_*^2) \,. \label{030}\end{aligned}$$ Whenever $w_{\rm eff}<-1/3$ solution (\[009\]) describes a universe undergoing an accelerated phase of expansion. This kind of evolution is useful to model both the inflationary early universe and the late time dark energy dominated universe. We can also have a look at how $w_\phi$ can be rewritten in terms of the variables (\[020\]), $$\begin{aligned} w_\phi = \frac{x^2-y^2}{x^2+y^2} \,. \label{056}\end{aligned}$$ This espression tells us the equation of state of the scalar field at any given point of the phase space. The first step one should make in order to analyze the dynamical system (\[010\])-(\[011\]) is to compute the critical/fixed points of the system. These are the phase space points $(x,y)$ that satisfy the conditions $$\begin{aligned} x'=0 \,, \qquad y'=0 \,.\end{aligned}$$ If the system happens to be in one of these points, then there is no dynamical evolution and the universe expand according to (\[007\]). Their existence is satisfied only if their coordinates are real and lie inside the phase space, i.e. the upper unit half-disk in the present case. The stability conditions are computed linearizing the equations around the critical point under consideration which leads to the analysis of the eigenvalues of the Jacobian matrix $$\begin{aligned} \mathcal{M} = \left( \begin{array}{cc} \frac{\partial f_x}{\partial x} & \frac{\partial f_x}{\partial y} \\ \frac{\partial f_y}{\partial x} & \frac{\partial f_y}{\partial y} \end{array} \right) \,, \label{050}\end{aligned}$$ evaluated at the critical point. Here $x'=f_x(x,y)$ and $y'=f_y(x,y)$ is a compact notation for the system (\[010\])-(\[011\]). If the real part of both the eigenvalues is positive then the point is an unstable point, if they have different signs the point is a saddle point and if they are both negative the point is a stable point. Point $x$ $y$ Existence $w_{\rm eff}$ Acceleration $\Omega_\phi$ Stability ------------- ----------------------------------------------- -------------------------------------------- ------------------------------ ------------------------------- --------------------- ---------------------------------- ------------------------------------ \*[$O$]{} \*[0]{} \*[0]{} \*[$\forall\;\lambda,w$]{} \*[$w$]{} \*[No]{} \*[0]{} \*[Saddle]{} \*[$A_-$]{} \*[-1]{} \*[0]{} \*[$\forall\;\lambda,w$]{} \*[1]{} \*[No]{} \*[1]{} Unstable if $\lambda\geq-\sqrt{6}$ Saddle if $\lambda<-\sqrt{6}$ \*[$A_+$]{} \*[1]{} \*[0]{} \*[$\forall\;\lambda,w$]{} \*[1]{} \*[No]{} \*[1]{} Unstable if $\lambda\leq \sqrt{6}$ Saddle if $\lambda>\sqrt{6}$ \*[$B$]{} \*[$\sqrt{\frac{3}{2}}\frac{1+w}{\lambda}$]{} \*[$\sqrt{\frac{3(1-w^2)}{2\lambda^2}}$]{} \*[$\lambda^2\geq 3(1+w)$]{} \*[$w$]{} \*[No]{} \*[$\frac{3(1+w)}{\lambda^2}$]{} \*[Stable]{} \*[$C$]{} \*[$\frac{\lambda}{\sqrt{6}}$]{} \*[$\sqrt{1-\frac{\lambda^2}{6}}$]{} \*[$\lambda^2<6$]{} \*[$\frac{\lambda^2}{3}-1$]{} \*[$\lambda^2<2$]{} \*[1]{} Stable if $\lambda^2<3(1+w)$ Saddle if $3(1+w)\leq\lambda^2<6$ The critical points of the system (\[010\])-(\[011\]) are shown in Table \[tab:01\]. There can be up to five critical points according to the value of $\lambda$: - [*Point $O$*]{}. The origin of the phase space corresponds to a matter dominated universe ($\Omega_m=1$) and exists for all values of $\lambda$. Of course the effective EoS matches the matter EoS, $w_{\rm eff}=w$, and thus for physically admissible values of $w$ there is no acceleration. This point is always a saddle point attracting trajectories along the $x$-axis and repelling in any other direction. - [*Point $A_\pm$*]{}. In these two points the universe is dominated by the scalar field kinetic energy ($\Omega_\phi=1$) and thus the effective EoS reduces to a stiff fluid with $w_{\rm eff}=w_\phi=1$ and no acceleration. Their existence is always garanteed and they never represent stable points. They are unstable or saddle points depending on the absolute value of $\lambda$ being greater or smaller than $\sqrt{6}$. - [*Point $B$*]{}. This point is the so called scaling solution where the effective EoS matches the matter EoS, but the scalar field energy density does not vanish. In other words we always have both $0<\Omega_\phi=3(1+w)/\lambda^2<1$ and $0<\Omega_m=1-\Omega_\phi<1$, obtaining also $w_\phi=w$. This means that the universe evolves under both the matter and scalar field influence, but it expands as if it was completely matter dominated. This solution is of great physical interest for the coincidence problem since according to it a scalar field can or could be present in the universe hiding its effects on cosmological scales. However, since we have $w_{\rm eff}=w$ there cannot be accelerated expansion. When this point exists, i.e. for $\lambda^2\geq 3(1+w)$, it always represents a stable point attracting all the phase space trajectories. - [*Point $C$*]{}. The last point stands for the cosmological solution where the universe is completely scalar field dominated. This implies $\Omega_m=\sigma^2=0$ and $\Omega_\phi=x^2+y^2=1$ meaning that Point $C$ will always lie on the unit circle. It exists for $\lambda^2<6$ and it is a stable attractor for $\lambda^2<3(1+w)$ (i.e. when Point $B$ does not appear) and a saddle point for $3(1+w)\leq\lambda^2<6$. The effective EoS parameter assumes the value $w_{\rm eff}=w_\phi=\lambda^2/3-1$ which implies an accelerating universe for $\lambda^2<2$. This point represents the well-known cosmological accelerated expansion driven by a sufficiently flat scalar field potential. The physical applications abound in both the early and late time universe stages. In the limit $\lambda\rightarrow 0$ this solution reduces to a de Sitter expansion dominated by a cosmological constant. The qualitative behavior of the phase space can be divided into three regions according to the value of $\lambda^2$. In what follows we will only consider positive values for $\lambda$. The dynamics for negative values coincides with the positive one after a reflection around the $y$ axis because of (\[031\]). If $\lambda^2<3(1+w)$ there are four critical points. Points $A_\pm$ are both unstable node, while Point $O$ is a saddle point. The global attractor is Point $C$ which represents an inflationary cosmological solution if $\lambda^2<2$. The portrait of the phase space is depicted in Fig. \[fig:01\] where the values $\lambda=1$ and $w=0$ have been choosen. The yellow/shaded region delimits the zone of the phase space where the universe undergoes an accelerated expansion. Point $C$ always lies on the unit circle and it happens to be outside the acceleration region if $\lambda^2>2$. ![Phase space with $\lambda=1$ and $w=0$. The global attractor is Point $C$ which represents an accelerating solution. For values $\lambda^2>2$ Point $C$ would lie outside the acceleration region (yellow/shaded) and would not be an inflationary solution.[]{data-label="fig:01"}](Standard_1){width="\columnwidth"} In the range $3(1+w)\leq\lambda^2<6$ there are five critical points in the phase space. Points $A_\pm$ and $O$ still behaves as unstable nodes and saddle point respectively. The global attractor is now Point $B$ and Point $C$ becomes a saddle point. The phase space portrait is drawn in Fig. \[fig:02\]. Point $B$ always lies outside the acceleration region (yellow/shaded) and thus never describe an inflationary solution. However the effective EoS parameter at this point coincides with the matter EoS parameter and thus the universe experience a matter-like expansion even if it is not completely matter dominated. This is the so called scaling solution where the scalar field energy density fills part of the universe but the resulting cosmological evolution still assumes the behavior of a matter dominated expansion. ![Phase space with $\lambda=2$ and $w=0$. The global attractor is Point $B$ where the universe expands as it was completely matter dominated, while Point $C$ is a saddle point.[]{data-label="fig:02"}](Standard_2){width="\columnwidth"} Finally if $\lambda^2\geq 6$ there are again only four critical points. Point $A_-$ is the only unstable node, while Points $A_+$ and $O$ behave as saddle points. Point $C$ does not appear anymore and the global attractor is still Point $B$, which again represents a scaling solution with $w_{\rm eff}=w$. The phase space dynamics is depicted in Fig. \[fig:03\]. ![Phase space with $\lambda=3$ and $w=0$. Point $B$ is the global attractor describing a scaling solution with $w_{\rm eff}=w$.[]{data-label="fig:03"}](Standard_3){width="\columnwidth"} The cosmological dynamics of the canonical scalar field is interesting because of the appearence of late time accelerated solutions which can be employed to model dark energy and inflation. The scaling solutions are also important since allow a scalar field to hide its presence during the cosmological evolution. This situation can be used to postulate a scalar field which gives no contribution at early times but becomes relevant at late times. Unfortunately there are strong observational constraints from nucleosyntesis which force the parameter $\lambda$ to satisfy the relation $\lambda\gtrsim 9$ at early times [@Bean:2001wt]. Since for a late time accelerating solution a sufficiently flat potential is needed ($\lambda^2<2$), it is impossible to achieve both the scaling and accelerating regimes with a canonical scalar field and an exponential potential. Moreover with a canonical scalar field we always have that unstable nodes of the phase space, possibly representing very early time behaviors, are associated with scalar field kinetic dominated universe. This solutions are characterized by an effective EoS approaching the stiff regime where $w_{\rm eff}=1$. Strictly speaking this value of $w_{\rm eff}$ is not physically viable at the classical level. However since these solutions appear to be relevant only at very early times in physical applications, this feature is usually ignored. As we will see in Sec. \[sec:square\], with a non-canonical scalar field an early time matter dominated solution can always be obtained. The noncanonical scalar field {#sec:noncanonical} ============================= In this section we will generalize the scalar field Lagrangian $\mathcal{L}_\phi$. The canonical choice for $\mathcal{L}_\phi$ has been given in (\[012\]) and its dynamics on cosmological scales has been investigated in full details in the previous section. In order to simplify the following equations we define $$\begin{aligned} X = -\frac{1}{2}\partial\phi^2 = -\frac{1}{2} g^{\mu\nu}\partial_\mu\phi\partial_\nu\phi \,.\end{aligned}$$ The canonical choice for the scalar field Lagrangian corresponds then to $\mathcal{L}_\phi=X-V$. The most general Lagrangian containing $X$ and $\phi$ is given by $\mathcal{L}_\phi=P(X,\phi)$ where $P$ is an arbitrary function in both the variables. In cosmology such theories are known under the name of $k$-essence. They received a considerable amount of attention during the last few years because of their applications to dark energy, dark matter and inflation. In order to reduce our analysis we will focus only on scalar field Lagrangians generally defined by $$\begin{aligned} \mathcal{L}_\phi= V f(B)\quad \mbox{with}\quad B=\frac{X}{V} \,, \label{013}\end{aligned}$$ and $f$ an arbitrary function. This includes the canonical choice if one considers $f(B)=B-1$. In [@Piazza:2004df; @Tsujikawa:2004dp] it has been shown that, within general relativity, the most general Lagrangian leading to cosmological scaling solutions with an exponential potential can be written as[^2] (\[013\]). In addition the dimensionless variables (\[020\]) turn out to be of great advantage if a scalar field Lagrangian is assumed as in (\[013\]). As we will see they will permit to completely determine the cosmological dynamics of such a scalar field. As before we will only consider the exponential potential case $V=V_0\exp(-\lambda\kappa\phi)$. The variation of the gravitational action (\[014\]) with the scalar field Lagrangian (\[013\]) leads to the gravitational equations $$\begin{aligned} \frac{1}{\kappa^2}G_{\mu\nu} =T_{\mu\nu}+ g_{\mu\nu}\,V f+ \frac{\partial f}{\partial B} \, \partial_\mu\phi\partial_\nu\phi \,,\end{aligned}$$ while the variation with respect to the scalar field $\phi$ gives $$\begin{aligned} \nabla_\mu\left( \frac{\partial f}{\partial B} \partial^\mu\phi \right) +\frac{\partial V}{\partial\phi} \left(f-\frac{\partial f}{\partial B}\,B\right) =0 \,, \label{016}\end{aligned}$$ Notice that these equations reduces to (\[001\]) and (\[002\]) for the canonical choice of the scalar field Lagrangian. As before, the cosmological equations can be found employing the FRW metric (\[015\]) and assuming an homogeneuos scalar field $\phi=\phi(t)$. The Friedmann constraint becomes $$\begin{aligned} \frac{3H^2}{\kappa^2} = \rho - V f +\frac{\partial f}{\partial B}\dot\phi^2 \,, \label{017}\end{aligned}$$ while the acceleration equations generalizes to $$\begin{aligned} 2\dot H+3H^2 = -\kappa^2 \left(p+Vf\right) \,. \label{018}\end{aligned}$$ On the other hand the same assumptions reduce (\[016\]) to $$\begin{gathered} \left(\frac{\partial f}{\partial B}+2B\frac{\partial^2f}{\partial B^2}\right) \ddot\phi +\frac{\partial f}{\partial B}3H\dot\phi \\ -\left(f-B\frac{\partial f}{\partial B}+2B^2\frac{\partial^2 f}{\partial B^2}\right)\frac{\partial V}{\partial\phi} = 0 \,, \label{019}\end{gathered}$$ where now $B=\dot\phi^2/(2V)$. Some remarks can now be made on Eqs. (\[017\])-(\[019\]). First of all we notice again that choosing $f=B-1$ reduces these equations to (\[003\])-(\[006\]) as expected. It is interesting to find the particular form of the function $f$ for which the contribution of the scalar field in (\[017\]) completely disappears. This is realized for $f=\sqrt{B}$ or, in other words, for the Lagrangian $\mathcal{L}_\phi=\sqrt{XV}$. Unfortunately this particular choice also makes the first and last terms in (\[019\]) to vanish. This implies that $\phi$ has no dynamics at all and becomes simply a constant. In any case it is worth noting that adding the $\sqrt{B}$ term to any other function $f$ does not modify the Friedmann constraint (\[017\]) and adds a simple term $3H\sqrt{V/2}$ to the scalar field equation (\[019\]). This features will be analyzed in more details in Sec. \[sec:phantom\]. Note also that it is impossible to find a scalar field Lagrangian whose contribution in the acceleration equation (\[018\]) vanishes. This is due to the fact that $\mathcal{L}_\phi=V\,f$ and thus the vanishing of the scalar field contribution in (\[018\]) would correspond to a zero scalar field Lagrangian. At this point is useful to see what the Friedmann constraint (\[017\]) looks like in terms of the variables (\[020\]). We obtain $$\begin{aligned} 1=\sigma^2 - y^2 f +2x^2\frac{\partial f}{\partial B} \,, \label{021}\end{aligned}$$ where we also have that $$\begin{aligned} B=\frac{x^2}{y^2} \,.\end{aligned}$$ The Friedmann constraint (\[021\]) determines the boundaries of the phase space described by the variables $x$ and $y$. In the canonical case (\[012\]) this reduces to (\[022\]) and the phase space is simply the upper-half unit circle. However if we choose a different function $f$ the phase space arising from (\[021\]) can be considerably different from the canonical one. As a consequence we can even lose the compactness of the phase space. In the next sections we will study what happens with different choices for the function $f$. In particular we will look for functions for which the phase space remains compact. The expression of (\[018\]) in terms of the $x$ and $y$ variables is given by $$\begin{aligned} \frac{\dot H}{H^2}= -\frac{3}{2} \left[(1+w)(1+y^2f)-2wx^2\frac{\partial f}{\partial B}\right] \,. \label{024}\end{aligned}$$ From this we can extract the effective EoS parameter of the universe as $$\begin{aligned} w_{\rm eff} = w + (w+1)y^2f-2wx^2 \frac{\partial f}{\partial B} \,.\end{aligned}$$ We can also find the EoS of the scalar field. The energy density and pressure of $\phi$ are given by $$\begin{aligned} \rho_\phi&= 2X\frac{\partial f}{\partial B}-Vf \,, \label{036} \\ p_\phi&=\mathcal{L}_\phi= Vf \,.\end{aligned}$$ Note that $\rho_\phi$ coincides with the expression appearing in the Friedmann constraint (\[017\]) as expected. To be physically viable at the quantum level $\phi$ must satisfy $\rho_\phi>0$. The EoS parameter of the scalar field is then $$\begin{aligned} w_\phi = \frac{p_\phi}{\rho_\phi} = \left(2\frac{X}{V}\frac{1}{f}\frac{\partial f}{\partial B}-1\right)^{-1} \,.\end{aligned}$$ In the canonical case $f=B-1$ this reduces to (\[032\]), while if $f$ is constant this simply becomes $w_\phi=-1$ describing a cosmological constant. Finally another useful quantity in scalar field cosmology is the so called speed of sound of adiabatic perturbations. This is defined as $$\begin{aligned} c_s^2 = \frac{\partial p_\phi}{\partial X}/\frac{\partial\rho_\phi}{\partial X} =\left(1+2\frac{X}{V}\frac{\partial^2 f}{\partial X^2}/\frac{\partial f}{\partial X}\right)^{-1} \,, \label{034}\end{aligned}$$ and for physically viable cosmological models we must require $c_s^2>0$. If this condition is dropped then instabilities arise at the level of perturbations of the scalar fluid. For the canonical scalar field we find $c_s^2=1$ which implies that perturbations propagates at the speed of light. At this point, in order to completely determine the dynamics of a specific model of noncanonical scalar field, we need to choose the form of the function $f$. Ideally we would like both to find interesting phenomenology at cosmological scales and to satisfy the physical conditions $\rho_\phi>0$ and $c_s^2>0$. A possible attempt could be $f=-\exp(B)$. This choice seems indeed to yield some interesting features as one can realize looking at the EoS parameter for the scalar field which reads $$\begin{aligned} w_\phi = \frac{-V}{2X+V} \quad\mbox{with}\quad \mathcal{L}_\phi=-Ve^{-X/V} \,. \label{033}\end{aligned}$$ Whenever $X\gg V$ we have $w_\phi\simeq 0$, while if $X\ll V$ we get $w_\phi\simeq-1$. The scalar field (\[033\]) can thus be used to characterize a dust fluid at early times and a cosmological constant at late times. Such a field could even be used to build a unified model of dark energy and dark matter, though the effects of the scalar field when $X\gg V$ would result really small since in this limit $\mathcal{L}_\phi\simeq 0$. Another drawback of this model is given by the speed of sound (\[034\]) which results to be $$\begin{aligned} c_s^2=\left(1-\frac{2X}{V}\right)^{-1} \quad\mbox{with}\quad \mathcal{L}_\phi=-Ve^{-X/V} \,.\end{aligned}$$ As we can note, as soon as $V<2X$ we obtain $c_s^2<0$ which gives rise to instabilities at the level of perturbations. Also when $V=2X$ the speed of sound diverges. We could overlook this problem for the sake of finding interesting phenomenology for the background evolution of the universe. However, as we shall see in Sec. \[sec:phantom\], dropping this assumption can actually lead to a much richer cosmology if one chooses a different model. As we can realize the exponential Lagrangian (\[033\]) reduces to the canonical one at first order in $X/V$, so whenever this quantity is small, which usually happens at late times in cosmology, the exponential model is well approximated by the canonical scalar field. In this case the first corrections at second order would be determined by the term $-X^2/(2V)$. However, as we have seen above, the exponential Lagrangian (\[033\]) leads to instabilities at the perturbation level. Of course there could be another form for the function $f$ which does not introduce such problems and which reduces to the canonical case when $X/V$ is small. Corrections to the canonical Lagrangian will then be given by higher order power-law kinetic terms. Thus instead of guessing a specific form for the function $f$, we can take a starting point based on higher order (kinetic) corrections to the canonical Lagrangian. This will allow us to analyze models which both resemble the canonical scalar field at late times and are sufficiently simple to handle so that one can determine the complete cosmological dynamics of the scalar field. We will then consider models which gives (kinetic) power-law corrections to the canonical case characterized by the function $f=B-1+\xi B^n$ with $\xi$ and $n$ two real parameters. The corresponding Lagrangian reads $$\begin{aligned} \mathcal{L}_\phi = X-V+\xi\,V \left(\frac{X}{V}\right)^n \,, \label{035}\end{aligned}$$ which is well defined only considering $n>0$. The corrections to the canonical case are defined by the parameter $n$. For example, if $n=2$ we have that the next-to-first order corrections are of the square type, while if $n=3$ these are of the cubic type. If instead $n<1$ then we get corrections also at late times and the model does not reduce to a canonical scalar field. The energy density (\[036\]) and speed of sound (\[034\]) for this model become $$\begin{aligned} \rho_\phi &=V+X+(2 n-1) \xi\,V \left(\frac{X}{V}\right)^n \,, \label{038}\\ c_s^2&= \frac{X+\xi n V \left(\frac{X}{V}\right)^n}{X+\xi n (2 n-1) V \left(\frac{X}{V}\right)^n} \,. \label{039}\end{aligned}$$ If we assume $n> 1/2$ and $\xi\ge 0$ these quantities are always positive and finite and thus physically viable. The case $n=1/2$ is of particular interest and will be treated in Sec. \[sec:phantom\], while the value $\xi=0$ yields back the canonical case. The scalar field EoS parameter is given by $$\begin{aligned} w_\phi = \frac{X-V+\xi\, V \left(\frac{X}{V}\right)^n}{X+V+\xi \, (2 n-1)\,V \left(\frac{X}{V}\right)^n} \,, \label{040}\end{aligned}$$ and reduces to $-1$ for $V\gg X$ and to $1/(2n-1)$ for $V\ll X$ given $n>1$. This model allows for a late time cosmological constant-like EoS while the early time value of (\[040\]) is determined by the parameter $n$. Note that for $n>1$ the scalar field EoS at early times is always positive and smaller than $1$. In the next sections we will focus on the cases $n=2$ and $n=1/2$. The first one follows the philosophy of recovering a canonical scalar field at late times and will be studied in Sec. \[sec:square\]. The second one will introduce modifications at both early and late times and the phenomenology at cosmological scales will result much different and richer than the canonical one as we will see in Sec. \[sec:phantom\]. Square kinetic corrections {#sec:square} ========================== This section will be devoted to the dynamical analysis of background cosmologies arising from a scalar field described by Lagrangian (\[035\]) with $n=2$: $$\begin{aligned} \mathcal{L}_\phi = X-V+\xi\,\frac{X^2}{V} \,. \label{049}\end{aligned}$$ The parameter $\xi$ will be allowed to take any real value. If $\xi=-1/2$ this model approximates the exponential model (\[033\]) at second order in the late time small quantity $X/V$. However $\xi$ must be positive for physically viable models. In fact the energy density (\[038\]) and sound speed (\[039\]) reduce to $$\begin{aligned} \rho_\phi &= X+V+3 \xi \frac{X^2}{V} \,,\\ c_s^2 &= \frac{V+2 \xi X}{V+6 \xi X} \,.\end{aligned}$$ These two quantities are always positive, for all values of $X$ and $V$, only provided $\xi>0$. Moreover we notice that the speed of sound of adiabatic perturbations reduces to one when $V\gg X$ and to $1/3$ when $V\ll X$. At early times the perturbations travels at one third of the speed of light. The EoS parameter of the scalar field (\[040\]) becomes $$\begin{aligned} w_\phi = \frac{X V-V^2+\xi X^2}{X V+V^2+3 \xi X^2} \,.\end{aligned}$$ Interestingly this reduces to $-1$ when $V\gg X$ and to $1/3$ when $V\ll X$, implying that the scalar field acts as relativistic matter at early times and as an effective cosmological constant at late times. This feature signals that the model we are working with can lead to a physically sensible phenomenology at cosmological scales. The cosmological equations (\[017\]) and (\[018\]) for this model are given by $$\begin{aligned} \frac{3 H^2}{\kappa ^2}&=\rho +V+\frac{1}{2}\dot\phi^2+\frac{3\xi}{4} \frac{\dot\phi^4}{V} \,,\label{037}\\ 3 H^2+2 \dot H&=- \kappa ^2\left(p+\frac{1}{2} \dot\phi^2- V+\frac{\xi}{4} \frac{\dot\phi^4}{V}\right) \,, \label{042}\end{aligned}$$ while the scalar field equations (\[019\]) becomes $$\begin{aligned} \left(1+3 \xi\frac{\dot\phi^2}{V}\right)\left(\ddot\phi+3 H \dot\phi\right)+\left(1-\frac{3 \xi}{4} \frac{\dot\phi^4}{V^2}\right)\frac{\partial V}{\partial\phi}=0 \,. \label{043}\end{aligned}$$ To determine the complete dynamics of these equations, we now employ the dimensionless variables (\[020\]). The Friedmann constraint (\[037\]) can then be written as $$\begin{aligned} \sigma^2+x^2+y^2+3 \xi\frac{x^4}{y^2}=1 \,, \label{041}\end{aligned}$$ where now the relative scalar field energy density is given by $$\begin{aligned} \Omega_\phi = x^2+y^2+3\xi\frac{x^4}{y^4} \,. \label{054}\end{aligned}$$ The Friedmann constraint (\[041\]) can again be used to replace $\sigma^2$ in all the other cosmological equations. This will permit us to write the dynamical equations as an autonomous system in the variable $x$ and $y$, exactly as we did for the canonical field in Sec. \[sec:canonical\]. In addition, given that $\sigma^2>0$ due to the assumption $\rho>0$, the Friedmann constraint (\[041\]) reduces the phase space to be compact and delimitated by the close geometric curve defined by $x^2+y^2+3 \xi x^4/y^2=1$. The boundary of the phase space now depends on the parameter $\xi$: the larger is $\xi$ the smaller is the phase space. The shape of the allowed phase space region can be observed in Fig. \[fig:11\], where the right column shows the phase space for different values of $\xi$. The boundary of the region always presents an edge in the origin, but otherwise is smooth. The phase space becomes the upper half unit disk if $\xi\rightarrow 0$ as expected, while it reduces to the $y$-axis for $\xi\rightarrow \infty$. With the dimensionless variables (\[020\]) we can now write Eqs. (\[042\]) and (\[043\]) as the dynamical system $$\begin{aligned} x'&= \frac{1}{2 y^2 \left(6 \xi x^2+y^2\right)}\Big[18 \xi ^2 (1-3 w) x^7 \nonumber\\ &\quad -3 x y^4 \left(6 \xi x^2 (w+1)+(w-1)\left(x^2-1\right)\right) \nonumber\\ &\quad +3 \xi x^3 y^2 \left((7-9 w) x^2+6 w-\sqrt{6} \lambda x+2\right) \nonumber \\ &\quad +y^6\left(\sqrt{6} \lambda -3 (w+1) x\right)\Big] \,,\label{044}\\ y'&= \frac{1}{2 y}\Big[3y^2 (w+1)\left(1-y^2\right)+3 \xi (1-3 w) x^4 \nonumber\\ &\quad -xy^2 \left(\sqrt{6} \lambda +3 (w-1) x\right) \Big] \,. \label{045}\end{aligned}$$ Note that this system is invariant under the relation $y\mapsto-y$, which implies that the dynamics on the negative $y$ half plane is symmetric to the one in the upper half plane, exaclty as it happens for the canonical scalar field. This again means that even if one drops the $V>0$ assumption, the dynamics of the whole system can be determined by just the $y>0$ analysis. Of course one always has to assume $V\neq 0$ in order for the model to not become singular. Also the transformation (\[031\]) leaves the system (\[044\])-(\[045\]) unchanged, meaning that opposite values of $\lambda$ lead to the same dynamics after a reflection over the $y$-axis, again as it was in the canonical case. Note also that given $y>0$ and $\xi>0$ the system (\[044\])-(\[045\]) is never singular. Moreover the origin can be taken to be part of the phase space since in the limit $x,y\rightarrow 0$ the system remain well defined as can be proved in polar coordinates[^3]. From Eq. (\[042\]) we can also obtain $$\begin{gathered} \frac{\dot H}{H^2} = \frac{3}{2 y^2} \Big[\xi (3 w-1) x^4+(w-1) x^2 y^2 \\ +(w+1) y^2 \left(y^2-1\right)\Big] \,,\end{gathered}$$ from which we can obtain the effective EoS parameter at any critical point $(x_*,y_*)$ as $$\begin{aligned} w_{\rm eff} = w-(w-1) x_*^2-(w+1) y_*^2-\xi (3 w-1) \frac{x_*^4}{y_*^2} \,. \label{048}\end{aligned}$$ In the origin this reduces to the matter EoS parameter and the scalar field has no effects on the cosmological evolution. On the other side if $x=0$ and $y=1$ this becomes $w_{\rm eff} =-1$ and the universe undergoes a de Sitter expansion. Point $\, x\,$ $y$ Existence $w_{\rm eff}$ Accel. $\Omega_\phi$ Stability ----------- ---------- -------------------------------- -------------------------------- ----------------------- ------------ ---------------------- ---------------- \*[$O$]{} \*[0]{} \*[0]{} \*[$\forall\;\lambda,\xi,w$]{} \*[$w$]{} \*[No]{} \*[0]{} \*[Unstable]{} \*[$B$]{} \*[Fig. \[fig:10\]]{} \*[$w$]{} \*[No]{} \*[App.]{} \*[Stable]{} \*[$C$]{} \*[$\forall\;\lambda,\xi,w$]{} \*[App.]{} \*[Fig. \[fig:10\]]{} \*[1]{} \*[Fig \[fig:10\]]{} : Critical points of the system (\[044\])-(\[045\]) and their properties. The coordinates of Points $B$ and $C$ are given in the Appendix.[]{data-label="tab:03"} The critical points for the dynamical system (\[044\])-(\[045\]), together with their properties, are listed in Table \[tab:03\], while existence and stability are explained in Fig. \[fig:10\]. Assuming the origin is part of the phase space because of the considerations above, there are now only up to three critical points. Due to the high powers in $x$ and $y$ of the system (\[044\])-(\[045\]) the coordinate values of the critical points result quite lengthy and complicated. For this reason their explicit expression is given only in the Appendix. - [*Point $O$*]{}. As we already noticed the origin can be taken to be part of the phase space since the system (\[044\])-(\[045\]) is regular at this point, as can be proved in polar coordinates. Of course it represents a matter dominated universe where the effective EoS parameter equals $w$ and $\Omega_m=1$. Interestingly the origin is now always an unstable node meaning that a completely matter dominated universe results unstable and eventually evolves to other configurations. - [*Point $B$*]{}. This point represents again a scaling solutions where $w_{\rm eff}=w_\phi=w$ but the scalar field energy density does not vanish: $0<\Omega_\phi<1$. It exists only when the parameters $\lambda$ and $\xi$ lie inside region III of Fig. \[fig:10\] and it is always the global attractor of the phase space when it appears. Again since the universe evolves as it was matter dominated, this point will never characterizes an accelerating solution. - [*Point $C$*]{}. This point stands again for a completely scalar field dominated universe. In fact it always lies on the border of the phase space where $\Omega_m=\sigma=0$, $\Omega_\phi=1$ and $w_{\rm eff}=w_\phi$. However, in contrast with the canonical case, it appears in the phase space for all possible values of $\lambda$, $\xi$ and $w$. For $\lambda\rightarrow\pm\infty$ this point moves along the border of the phase space eventually approaching the origin. Regions I and II in Fig. \[fig:10\] shows the values of $\lambda$ and $\xi$ where Point $C$ is the global attractor of the phase space, while in region III it behaves as a saddle point being Point $B$ the global attractor. Region I of Fig. \[fig:10\] represents the area in the $(\lambda,\xi)$-plane where Point $C$ characterizes an inflationary/accelerating universe. In regions II and III instead the effective EoS parameter in Point $C$ is bigger than $-1/3$ and the universe undergoes a decelerating expansion. The first feature that one notices in this model, once a comparison with the canonical case is done, is that the kinetic scalar field dominated solutions appearing as critical Points $A_\pm$ of the system (\[010\])-(\[011\]), now are never part of the phase space. They are replaced by the matter dominated origin which now acts as the early time unstable solution. In this model thus, instead of having a nasty kinetic dominated solutions with $w_{\rm eff}=1$ at early time, we obtain a much more physical matter universe where $w_{\rm eff}=w$. In other words, in this model a matter dominated universe results unstable and eventually evolves to a configurations where the energy density of the scalar field does not vanish. Notice also that Point $B$ and $C$ reduce to their correspondent canonical ones in the limit $\xi\rightarrow 0$. We can now have a look at the complete phase space dynamics for the three different regions of Fig. \[fig:10\]. This has been drawn in Fig. \[fig:11\] where figures $(a)$ and $(b)$ represents region I, figures $(c)$ and $(d)$ region II and $(e)$ and $(f)$ region III. The left column shows how the dynamics of the phase space changes as the value of $\lambda$ changes, while the right column shows how it changes as the values of $\xi$ changes. As it is clear from Fig. \[fig:11\], different values of $\lambda$ do not change the shape of the phase space, while the value of $\xi$ determines the boundary, and thus the shape, of the phase space. This is of course due to the Friedmann constraint (\[041\]) which depends on $\xi$ as we already discussed above. The yellow/shaded region represents again the zone of the phase space where the universe undergoes an accelerated expansion. In Figs. \[fig:11\] $(a)$ and $(b)$ the phase spaces for the values $\lambda=1$, $\xi=1$ and $\lambda=2$, $\xi=4$ have been depicted. The only critical point appearing beside the origin is Point $C$, which, being inside the yellow region, characterizes an accelerating solution. All the trajectories evolves from the unstable matter dominated solution at the origin towards the scalar field dominated solution at Point $C$ which acts as the global attractor. This dynamics well suits the phenomenology of our universe since with this parameter choice we can have a decelerated to accelerated trasition describing the dominance of dark energy over dark matter at late times and the reverse situation at early times. The phase space dynamics for region II of Fig. \[fig:10\] has been drawn in Figs. \[fig:11\] $(c)$ and $(d)$ where the values $\lambda=2$, $\xi=1$ and $\lambda=2$, $\xi=1/2$ has been choosen respectively. The only two critical points in the phase space are again the origin (early time unstable solution) and Point $C$ (late time attractor) which now lies outside the yellow region and thus describes a decelerating scalar field dominated universe. Note that, depending on initial conditions, some trajectories will still experience a stage of accelerated expansion before ending in Point $C$. This particular evolution can thus be used to model universes with a transient inflationary era. Finally the dynamics characterized by region III of Fig. \[fig:10\] has been delineated in Figs. \[fig:11\] $(e)$ and $(f)$ where the values $\lambda=4$, $\xi=1$ and $\lambda=2$, $\xi=1/2$ has been choosen respectively. There are now all three critical points in the phase space. The origin is again the early time unstable node, the global attractor is Point $B$ representing a scaling solution and Point $C$ is now a saddle point. Depending on initial conditions we can again have a transient acceleration era before ending at Point $B$ with a matter-like cosmological evolution. This dynamics can be employed to build models of inflation where after the inflationary phase one obtains a graceful exit to the scaling solution. \ \ To conclude this section we compare this model with the canonical scalar field of Sec. \[sec:canonical\]. Both models present cosmological scaling solutions and late time inflationary attractors. They mainly differ in the early time dynamics where instead of having kinetic scalar field dominated solutions, in the noncanonical case only the matter dominated solution appears. This feature can be used to better motivate the phenomenology of dark energy. In fact with the model presented in this section a matter dominated universe is always unstable and eventually evolves to either a scaling or a scalar field dominated solution. For the right values of the parameters $\lambda$ and $\xi$ (see Fig. \[fig:10\]) the late time attractor characterizes an accelerated cosmological expansion implying a dynamics describing a transition from matter to dark energy domination in accordance with the current astronomical observations. Of course, being the model (\[049\]) a subclass of (\[013\]), we also obtain cosmological scaling solutions, identified with Point $B$ in Fig. \[fig:11\]. As we commented in Sec. \[sec:canonical\] these solutions are of great physical interest since can hide the scalar field effects on the background cosmological evolution. In the canonical case however there are strong constraints on the scalar field energy density obtained from nucleosyntesis observations which eventually impose $\lambda\gtrsim 9$. One could hope that for the scalar field (\[049\]) the constraint on $\lambda$ would relax. Unfortunately the introduction of the square kinetic corrections, parametrize by $\xi$, does not help in this situation. As can be realized from Fig. \[fig:12\], the allowed region of the $(\lambda,\xi)$-space for a viable scaling solution at early times, when $w=1/3$, is well separated from the late time acceleration region. The model (\[049\]) thus presents the same difficulties of the canonical case for hiding a scalar field at early times which eventually becomes relevant for dark energy phenomenology at late times. ![Allowed values (purple/darker region) for early time cosmological scaling solutions of the scalar field (\[049\]) (region III in Fig. \[fig:10\] (b)) permitted by nucleosyntesis observations. The same region appears for negative values of $\lambda$ due to symmetry (\[031\]).[]{data-label="fig:12"}](Nucleo){width="\columnwidth"} The addition of the square kinetic correction (\[049\]) to the canonical Lagrangian complicates the resulting cosmological background equations. As a consequence the autonomous system of equations (\[044\])-(\[045\]) contains powerlaw terms in $x$ and $y$ up to fifth order, in contrast with (\[010\])-(\[011\]) where the highest order is the third. This implies that the critical points of the system are much more complex and difficult to find as it is shown in the Appendix. If one considers cubic or higher order kinetic corrections to the canonical scalar field Lagrangian the corresponding cosmological dynamical system becomes almost impossible to analyze even with numerical techniques. The model (\[049\]) with square kinetic corrections is sufficiently simple to study and sufficiently different from the canonical case to present new phenomenology at cosmic distances, especially at early times where the scalar field kinetic dominated solutions no longer appear. If $\xi\ll 1$ the dynamics of the scalar field (\[049\]) approaches the corresponding canonical one. However as long as $\xi\neq 0$ the kinetic dominated solutions, corresponding to $x=\pm 1$ and $y=0$, will never appear in the phase space. Interestingly if $\xi$ is almost zero these points effectively behave as saddle points, but for a non vanishing $\xi$, no matter how small, they never characterize kinetic dominated solutions because we obtain $w_\phi=1/3$ at that points. This implies a radiation-like evolution meaning that the scalar field behaves as relativistic matter. Square root kinetic corrections {#sec:phantom} =============================== In this section we will consider square root kinetic corrections to the canonical scalar field Lagrangian. In other words we will study the model (\[035\]) with $n=1/2$, which is simple to analyze and capable of providing interesting phenomenology at cosmological scales. The scalar field Lagrangian of this model is then $$\begin{aligned} \mathcal{L}_\phi = X - V + \xi\, \sqrt{XV} \,, \label{023}\end{aligned}$$ where $\xi$ is again a free parameter. Recall that $V$ has been assumed to be positive and of the exponential form (\[004\]), which means there are no inconsinstencies with the square root appearing in (\[023\]). Before one proceeds with the analisys of the dynamical equations, a subtlety must be taken into account. The square root in (\[023\]) will provide terms containing $\sqrt{\dot\phi^2}$, or equivalently $\sqrt{x^2}$. This are strictly positive quantities which should be replaced by $|\dot\phi|$ or $|x|$. However this would prevent the autonomous system of equations to be differentiable in $x=0$ and the whole dynamical system analysis would be impossible since the function $f_x$ and $f_y$ of (\[050\]) must be differentiable. For this reason we will replace $\sqrt{\dot\phi^2}$ with $\dot\phi$ in what follows, or equivalently $\sqrt{x^2}\mapsto x$. This operation can be made mathematically rigorous assuming that $\xi\mapsto s\,\xi$ with $s$ the sign of $\dot\phi$, i.e. $s=1$ if $\dot\phi>0$ and $s=-1$ if $\dot\phi<0$. Note that we could have choose the opposite branch, i.e. $\sqrt{\dot\phi^2}\mapsto-\dot\phi$ or $\sqrt{x^2}\mapsto-x$, however due to the symmetry (\[028\]) of the resulting dynamical system both the choices would have result in the same cosmological dynamics. With the scalar field (\[023\]), the energy density (\[038\]) and speed of sound (\[039\]) reduce to $$\begin{aligned} \rho_\phi &= X+V \,,\label{051}\\ c_s^2 &= 1 + s\frac{\xi}{2}\sqrt{\frac{V}{X}} = 1+\frac{\xi}{2}\frac{y}{x} \,. \label{053} \end{aligned}$$ Of course the energy density (\[051\]) corresponds to the canonical energy density (\[052\]). As we already noticed in Sec. \[sec:noncanonical\], we can always add a $\sqrt{B}$ term to the function $f$ on the noncanonical scalar field (\[013\]) leaving the Friedmann constraint unmodified. For this reason the choice (\[023\]) yields nothing but the Friedmann constraint $$\begin{aligned} 1=\sigma^2+x^2+y^2 \,,\end{aligned}$$ which corresponds to the one arising in the canonical case, i.e. (\[022\]). This equals to say that the (gravitating) energy density of the scalar field (\[023\]) is the same as the canonical one and that the square root term does not give any energy contribution. The scalar field relative energy density will thus be $$\begin{aligned} \Omega_\phi = x^2+y^2 \,,\end{aligned}$$ which equals (\[055\]). The speed of sound (\[053\]) prevents the scalar field (\[023\]) to be physically viable. In fact it is easy to see that whenever $X=0$, or $x=0$, the speed of sound (\[053\]) diverges giving an infinite velocity of propagation for adiabatic perturbations. Moreover if $\xi x<0$ we always obtain $c_s^2<0$ in some region of the phase space in which the scalar field will present instabilities at the perturbation level. The model (\[023\]) results thus to be theoretically instable and non viable. However, despite all these drawbacks, in what follows we will ignore all the problems arising form Eq. (\[053\]). We will go on in analyzing the cosmological background dynamics implied by the scalar field (\[023\]) showing that it is capable of producing phenomenology which cannot be obtained with the canonical scalar field and which is slightly favoured by astronomical observations. The cosmological equations (\[018\]) and (\[019\]) now become $$\begin{aligned} 2\dot H+3H^2=-\kappa^2 \left(p+\frac{1}{2}\dot\phi^2-V+\xi\,\dot\phi\,\sqrt{V}\right) \,,\label{058}\\ \ddot\phi+3H\dot\phi+3H\,\xi\,\sqrt{V}+\frac{\partial V}{\partial\phi} = 0 \,.\end{aligned}$$ Notice that though the scalar field energy density is the same, its pressure changes due to the $\xi$ term. This is a peculiar feature of the square root term (\[023\]) which, despite having no gravitating energy, it yields a non zero pressure term into the acceleration equation (\[058\]). Moreover in the scalar field equation of motion the only modification due to $\xi$ is a new term directly coupling $H$ and the potential $V$. From these equations we obtain the following dynamical system $$\begin{aligned} %\begin{multlined} x'&= \frac{1}{2} \Big[-3 (w-1) x^3-3 x \left[(w+1) y^2-w+1\right]\nonumber\\ &\qquad\quad +3 \sqrt{2}\, \xi\, x^2 y+\sqrt{2}\, y \left(\sqrt{3}\, \lambda\, y-3\, \xi \right)\Big] \,, \label{025} \\ %\end{multlined} %\begin{multlined} y'&= -\frac{1}{2} y \Big[3 (w-1) x^2+3 (w+1) \left(y^2-1\right)\nonumber\\ &\qquad\qquad +x \left(\sqrt{6} \lambda -3 \sqrt{2} \xi y\right)\Big] \,, \label{026} %\end{multlined}\end{aligned}$$ which generalizes the system (\[010\])-(\[011\]) with the terms containing $\xi$. Note that Eqs. (\[025\]) and (\[026\]) are invariant under the symultaneous replacement $$\begin{aligned} \lambda\mapsto-\lambda \,,\quad \xi\mapsto-\xi \,,\quad x\mapsto-x \,, \label{028}\end{aligned}$$ which implies that the phase space is symmetric around the $y$-axis for opposite values of the parameters $\lambda$ and $\xi$. In the $\xi\rightarrow 0$ limit the symmetry (\[028\]) becomes (\[031\]). The system is also invariant under the following transformation $$\begin{aligned} \xi\mapsto-\xi \,,\quad y\mapsto-y \,,\end{aligned}$$ which tells us that the dynamics in the $y<0$ half phase space equals the one in the upper half space after a redefinition of $\xi$. In the $\xi\rightarrow 0$ limit this reduces to the $y\mapsto-y$ symmetry of the canonical case. Eq. (\[024\]) now reduces to $$\begin{aligned} \frac{\dot H}{H^2} = \frac{3}{2} \left[(w-1)\, x^2+(w+1) \left(y^2-1\right)-\sqrt{2}\, \xi\, x y\right] \,,\end{aligned}$$ and implies the following effective EoS parameter at any critical point $(x_*,y_*)$ $$\begin{aligned} w_{\rm eff} = x_*^2-y_*^2+w \left(1-x_*^2-y_*^2\right)+\sqrt{2}\, \xi\, x_* y_* \,. \label{029}\end{aligned}$$ Exactly as in (\[048\]), the fact that now the parameter $\xi$ is non zero leads to new interesting phenomenology in comparison with the cosmology of the standard scalar field. The EoS parameter of the scalar field now reads $$\begin{aligned} w_\phi=\frac{x^2-y^2+\sqrt{2}\xi xy}{x^2+y^2} \,,\end{aligned}$$ and differs from the canonical (\[056\]) only by the $\xi$-term in the numerator. Point $x$ $y$ Existence $w_{\rm eff}$ Acceleration $\Omega_\phi$ Stability --------------- ------------------------------------------------------------------------------------------ ----------------------------------------------------------------------------------- -------------------------------- --------------- ----------------------- ----------------------- ------------------------------------ \*[$O$]{} \*[0]{} \*[0]{} \*[$\forall\;\lambda,\xi,w$]{} \*[$w$]{} \*[No]{} \*[0]{} \*[Saddle]{} \*[$A_-$]{} \*[-1]{} \*[0]{} \*[$\forall\;\lambda,\xi,w$]{} \*[1]{} \*[No]{} \*[1]{} Unstable if $\lambda\geq-\sqrt{6}$ Saddle if $\lambda<-\sqrt{6}$ \*[$A_+$]{} \*[1]{} \*[0]{} \*[$\forall\;\lambda,\xi,w$]{} \*[1]{} \*[No]{} \*[1]{} Unstable if $\lambda\leq \sqrt{6}$ Saddle if $\lambda>\sqrt{6}$ \*[$B_\pm$]{} \*[$\sqrt{\frac{3}{2}}\frac{(w+1)}{\lambda }$]{} \*[$\frac{\sqrt{3}}{2 \lambda } \left(\xi \pm\sqrt{\xi ^2-2 w^2+2}\right)$]{} \*[Fig. \[fig:04\]]{} \*[$w$]{} \*[No]{} \*[$\Omega_\phi^B$]{} \*[Stable]{} \*[$C_-$]{} \*[$\frac{\sqrt{2} \lambda -\xi \sqrt{3 \xi ^2+6- \lambda ^2}}{\sqrt{3} (\xi ^2+2)}$]{} \*[$\frac{ \lambda \xi+\sqrt{6 \xi ^2+12-2\lambda ^2} }{\sqrt{3} (\xi ^2+2)}$]{} \*[Fig. \[fig:04\]]{} \*[$Q_-$]{} \*[Fig. \[fig:04\]]{} \*[1]{} \*[Fig. \[fig:04\]]{} \*[$C_+$]{} \*[$\frac{\sqrt{2} \lambda+\xi \sqrt{3\xi^2+6- \lambda ^2} }{\sqrt{3} (\xi ^2+2)}$]{} \*[$\frac{\lambda \xi -\sqrt{6 \xi ^2+12-2\lambda ^2}}{\sqrt{3}( \xi ^2+2)}$]{} \*[Fig. \[fig:04\]]{} \*[$Q_+$]{} \*[No]{} \*[1]{} \*[Unstable]{} The critical points of the system (\[025\])-(\[026\]) are listed in Table \[tab:02\], while their existence and stability properties are expalined in Fig. \[fig:04\]. There are now seven possible critical points and up to six of them can appear in the phase space at the same time. - [*Point $O$*]{}. Again the origin of the phase space formally corresponds to a matter dominated universe where $w_{\rm eff}=w$ and $\Omega_m=1$. Its properties are unmodified being always a saddle point and existing for all values of the parameters. - [*Points $A_\pm$*]{}. Also the two kinetic dominated solutions ($w_{\rm eff}=w_\phi=1$ and $\Omega_\phi=1$), labeled by Points $A_\pm$, still appear in the phase space presenting their standard behavior. In particular they are always saddle or unstable nodes depending on the absolute value of $\lambda$ being smaller or greater than $\sqrt{6}$. - [*Points $B_\pm$*]{}. These two points describe scaling solutions since in both of them $w_{\rm eff}=w_\phi=w$ and the scalar field energy density does not vanish. In fact the relative energy density of the scalar field is $$\begin{aligned} \Omega_\phi^B = \frac{3}{2 \lambda ^2} \left(\xi ^2\pm\xi \sqrt{\xi ^2-2 w^2+2}+2 w+2\right) \,, \label{057}\end{aligned}$$ which is always between 0 and 1 when Point $B$ exists. Their existence is given by regions III$_\pm$ in Fig. \[fig:04\] and depends also on the matter EoS parameter. The smaller the value of $w$ the bigger the existence region in the $(\lambda,\xi)$ parameter space, as can be seen comparing the left and right panels of Fig. \[fig:04\]. Whenever these points are present they always represent the global attractor of the phase space, but never describe accelerating solutions. - [*Points $C_\pm$*]{}. These two points represent scalar field dominated solutions and thus always lie on the unit circle being $\Omega_\phi=1$. In Fig. \[fig:04\] the existence of Point $C_+$ is given by the disconnected region II, while Point $C_-$ exists in both zones I and II. Point $C_+$ is always an unstable node, while Point $C_+$ is always the global attractor but inside the cross regions I/III$_\pm$ and II/III$_\pm$ where it behaves as a saddle point. The effective EoS is given by $w_{\rm eff}=w_\phi =Q_\pm$, where $$\begin{aligned} Q_\pm = \frac{2 \lambda ^2-3( \xi ^2+2)\pm\lambda \xi \sqrt{6 \xi ^2+12-2 \lambda ^2}}{3 (\xi ^2+2)} \,. \label{027}\end{aligned}$$ This desrcibes an inflationary solution in the connected region delineated by the dashed lines as drawn in Fig. \[fig:04\]. Unfortunately, for positive values of $w$, this accelerating region never overlaps the existence zones of Points $B_\pm$ meaning that inflating and scaling solutions cannot live in the same phase space. This features appeared also in the standard case. Whenever both Points $B$ and $C$ were present, the latter never described an inflationary solution as one can see from Fig. \[fig:02\]. Finally note that, depending on the choice of parameters $\lambda$ and $\xi$, Point $C_-$ can also describe a phantom dominated universe where $w_{\rm eff}=w_\phi<-1$. The regions in the parameter space where this happens are delimated by the two dotted line crossing the origin in Fig. \[fig:04\]. The top right and bottom left parts denotes phantom solutions for Point $C_-$, while in the rest of zones I and II we always find $w_{\rm eff}>-1$. Note that critical Points $B_\pm$ and $C_\pm$ reduce to Points $B$ and $C$ of the canonical case of Sec. \[sec:canonical\] in the limit $\xi\rightarrow 0$. We will now have a look at the dynamics of the phase space for values of the parameters $\lambda$ and $\xi$ representing the different zones in Fig. \[fig:04\]. We will restric our analysis to $w=0$ since the qualitative dynamical features do not change with other values of the matter EoS parameter. Moreover because of the symmetry (\[028\]) we need only to study half of the parameter space, say $\lambda>0$. The remaining half will describe identical phase spaces but for a reflection $x\mapsto-x$. In Figs. \[fig:05\], \[fig:06\], \[fig:07\], \[fig:08\] and \[fig:09\] the yellow/shaded region identifies the part of the phase space where the EoS parameter of stationary points is smaller than $-1/3$, implying an accelerated cosmological solution. The blue/dark part inside the yellow/shaded region delimitates the zone where the universe undergoes a phantom acceleration, i.e. where $w_{\rm eff}<-1$. Finally the green/shaded region denotes the area of the phase space where the effective equations of state takes super-stiff values, i.e. where $w_{\rm eff}>1$. ![Phase space with $w=0$, $\lambda=1$ and $\xi=1$ (region I in Fig. \[fig:04\]). Point $C_-$ is the global attractor describing a phantom accelerating solution.[]{data-label="fig:05"}](circle_11){width="\columnwidth"} We start considering zone I. If we choose $\lambda=1$ and $\xi=1$ the phase space looks like the one drawn in Fig. \[fig:05\]. Points $A_\pm$ are unstable nodes while the origin $O$ represents the matter dominated saddle point. The global attractor is Point $C_-$ which happens to be inside the accelerated region and thus describes an inflationary solution with $w_{\rm eff}=-11/9$. Being also inside the phantom region this value is clearly smaller than $-1$. Moreover since it lies on the unit circle it characterizes a universe completely dominated by the scalar field. If we had choosen the parameters $\lambda$ and $\xi$ to be outside the connected region delimitaded by the dashed lines in Fig. 4 but still inside zone I, then Point $C_-$ would have been outside the accelerated region, though still on the circle. In that case it would have described a decelerating universe dominated by the scalar field. On the other hand if we had choosen parameters inside the acceleration region, but outside the phantom region then we would have obtained an acceleration with $w_{\rm eff}>-1$. Note how the accelerated region is now different from the one in the standard case, Figs. \[fig:01\], \[fig:02\] and \[fig:03\]. This is due to the difference between the two EoS parameters (\[030\]) and (\[029\]). Moreover because in Eq. (\[029\]) there is a dependence on $\xi$, the acceleration region will change whenever $\xi$ is different, as in the next examples. ![Phase space with $w=0$, $\lambda=2$ and $\xi=-2$ (cross region I/III$_+$ in Fig. \[fig:04\]). Point $C_-$ is a saddle point while Point $B_+$ represents the global attractor describing a scaling solution with $w_{\rm eff}=w$.[]{data-label="fig:06"}](circle_2-2){width="\columnwidth"} The second zone we analyze in Fig. \[fig:04\] is the superposition region between zone I and zone III$_+$. Choosing the values $\lambda=2$ and $\xi=-2$ the phase space can be depicted as in Fig. \[fig:06\]. Points $A_\pm$ and $O$ are again unstable nodes and a saddle point respectively. Point $C_-$ is now a saddle point and the global attractor is Point $B_+$ describing a cosmological scaling solution with effective EoS parameter matching the matter one. Point $B_+$ clearly lies outside the accelerated region which happens to be modified with respect to the one in Fig. \[fig:05\], as we discussed above. ![Phase space with $w=0$, $\lambda=3$ and $\xi=-1$ (region III$_+$ in Fig. \[fig:04\]). Point $B_+$ is the global attractor describing a scaling solution with $w_{\rm eff}=w$.[]{data-label="fig:07"}](circle_3-1){width="\columnwidth"} The phase space characterized by zone III$_+$ of Fig. \[fig:04\] is depicted in Fig. \[fig:07\] where the values $\lambda=3$ and $\xi=-1$ have been choosen. Point $B_+$ is again the global attractor representing a cosmological scaling solution. Point $A_-$ is still an unstable node, while Point $A_+$ is now a saddle point exactly as the origin $O$. ![Phase space with $w=0$, $\lambda=3$ and $\xi=2$ (region II in Fig. \[fig:04\]). Point $C_-$ is the global attractor and represents a phantom accelerated solution. Point $C_+$ is an unstable node characterizing a super-stiff ($w_{\rm eff}>1$) solution.[]{data-label="fig:08"}](circle_32){width="\columnwidth"} In Fig. \[fig:08\] the portrait of the phase space for zone II has been drawn. Now both Points $C_\pm$ appear, one describing an unstable node ($C_+$) and the other one representing the global attractor ($C_-$) which can lie inside the phantom ($w_{\rm eff}<-1$), accelerated ($-1<w_{\rm eff}<-1/3$) or decelerated ($w_{\rm eff}>-1/3$) regions depending on the values of $\lambda$ and $\xi$. Points $A_\pm$ are unstable nodes and Point $O$ is a saddle point. ![Phase space with $w=0$, $\lambda=3.5$ and $\xi=1.5$ (cross region II/III$_+$ in Fig. \[fig:04\]). Point $B_+$ is the global attractor scaling solution, while Points $C_\pm$ represent a saddle and unstable node respectively.[]{data-label="fig:09"}](circle_3515){width="\columnwidth"} Finally the phase space of the region where zones II and III$_+$ superpose has been depicted in Fig. \[fig:09\]. Now Points $C_\pm$ appear together with Point $B_+$ representing a scaling solution. The situation is now similar to the one in Fig. \[fig:06\] (crossing zone I/III$_+$): Point $B_+$ is always the global attractor while Point $C_-$ is a saddle point. The only difference is now Point $C_+$ which acts as an unstable node. Points $C_\pm$, representing the scalar field dominated solutions, always lie outside the accelerated region and thus never describes an inflationary solution. However, as it is clear from Fig. \[fig:09\], before ending in Point $B_+$ several trajectories pass through the accelerated (phantom) region, meaning that the universe undergoes a stage of accelerated (phantom) expansion before scaling as a matter dominated solution. It is now interesting to compare the results we obtain from the model (\[023\]) with the ones following from the canonical scalar field of Sec. \[sec:canonical\]. The square root term in (\[023\]) leads to a much richer phenomenology at cosmic distances which includes phantom late time solutions, scaling solutions, new early time unstable solutions, super-stiff behavior and dynamical crossing of the phantom barrier at $w_{\rm eff}=-1$. Within this model one can not only achieve a matter to phantom transition at late times, but also phantom and super-stiff transient eras. This can be easily seen from Figs. \[fig:05\] to \[fig:09\] where, depending on initial conditions, some trajectories of the phase space will cross the blue and green regions representing phantom and super-stiff behavior respectively. Thus the scalar field (\[023\]) can describe a universe which is phantom dominated at late times instead of being only dark energy dominated as it happens in the canonical case. Despite the problems at the level of cosmological perturbations arising from Eq. (\[053\]), the scalar field model (\[023\]) is actually better in agreement with the latest astronomical observations which favour a value $w_\phi<-1$ at present, though the minus one value still lies inside the 2-sigma confidence limit [@Xia:2013dea; @Novosyadlyj:2013nya]. The scalar field (\[023\]) can thus characterize a quintom scenario where the crossing of the phantom barrier happens at late times with the universe being nowadays dark energy dominated ($w_{\rm eff}>-1$) but evolving through a final phantom era ($w_{\rm eff}<-1$). However in order to render this a viable model of our universe one must first solve the problems arising at the level of cosmological perturbations. ![Allowed values (purple region) for early time cosmological scaling solutions of the scalar field (\[023\]) (region III$_+$ in Fig. \[fig:04\] (b)) permitted by nucleosyntesis observations.[]{data-label="fig:13"}](nucleo_circle){width="\columnwidth"} Finally, being the scalar field (\[023\]) a subclass of (\[013\]), we obtain again cosmological scaling solutions (Points $B_\pm$) which, as we said, can be of great phenomenological interest. The scalar field can in fact hiding its presence at early times letting the cosmological evolution scaling as a matter dominated universe. For this to happen however we need to satisfy the constraints derived from nucleosystesis observations [@Bean:2001wt]. In Fig. \[fig:13\] the region allowed by these constraints for Point $B_+$ in the $(\lambda,\xi)$-space is shown. For Point $B_-$ the same region appears at opposite values of $\lambda$ and $\xi$ due to the symmetry (\[028\]). The introduction of the square root term in (\[023\]) does not help in relaxing the $\lambda\gtrsim 9$ constraint of the canonical case. In fact, as can be realized from Fig. \[fig:13\], the allowed region is well separated from the acceleration region of Point $C_-$, meaning that scaling and late time accelerated solution cannot appear in the same phase space. The same happens in both the canonical case and the model of Sec. \[sec:square\] where the allowed region results to be much more constrained as can be understood comparing Fig. \[fig:13\] and Fig. \[fig:12\]. Conclusion {#sec:concl} ========== In the present work the cosmological background evolution characterized by different scalar field models has been studied. The use of dynamical system techniques has allowed to completely determine the cosmological features of canonical and noncanonical scalar fields. After the canonical model has been reviewed in Sec. \[sec:canonical\], extended scalar field Lagrangians have been presented and discussed in Sec. \[sec:noncanonical\]. The analysis has then focused to models whose dynamics can be completely parametrized by the dimensionless variable (\[020\]) and which always leads to scaling solutions. In Sec. \[sec:square\] a scalar field with square kinetic corrections to the canonical Lagrangian has been examined. Its late time evolution qualitatively corresponds to the canonical situations with scaling and scalar field dominated solutions, while the early time features result modified. In particular the scalar field kinetic dominated solutions no longer appear in the phase space of this model. The early time behavior is now characterized by a matter dominated solution, which is better in agreement with a radiation/dark matter dominated epoch as required by observations. The model can thus be used to describe a universe where dark energy becomes important only at late times while dark matter dominates at early times. It also happens to be safe at the level of perturbations once the condition $\xi>0$ is assumed. Furthermore the phase space boundaries of the model presented in Sec. \[sec:square\] differ from the canonical ones. The phase space ceases to be the upper half unit disk in the $(x,y)$-plane and, remaining compact, assumes a form depending on the parameter $\xi$ as can be seen from Fig. \[fig:11\]. This is an interesting mathematical feature which implies that the variables (\[020\]) can lead to different phase space boundaries depending on the scalar field Lagrangian one chooses. Sec. \[sec:phantom\] has been devoted to the study of the cosmological consequences of a scalar field models with square root kinetic corrections to the canonical Lagrangian. The background dynamics of this model presents a richer phenomenology with respect to the canonical case. The early time behavior results similar to the canonical one, though super-stiff ($w_{\rm eff}>1$) transient regions always appear in the phase space. What changes more is the late time evolution where phantom dominated solutions, dynamical crossings of the phantom barrier and new scaling solutions emerge in the phase space. This model can thus be used to describe a late time dark energy dominated universe capable of dynamically crossing the phantom barrier ($w_{\rm eff}=-1$) as the astronomical observations slightly favour. Moreover we can achieve transient periods of super-acceleration ($\dot H>0$) where the universe expands only for a finite amount of time. These solutions are characterized by the trajectories that cross the phantom region in Figs. \[fig:05\] to \[fig:09\] and can be employed to build phantom models of inflation. The drawbacks arise of course at the level of pertubations where instabilities of the scalar field always appear. Until these problems are unsolved the scalar field model of Sec. \[sec:phantom\] cannot be seriously employed to describe physical universes. The author would like to thank Christian Böhmer, Marco Bruni, Emmanuel Saridakis and David Wands for useful discussions and comments on the paper. In this appendix we will provide the coordinate values of the critical points of the system (\[044\])-(\[045\]). For the sake of simplicity we will assume $w=0$ in what follows. Point $B$ is identified by the coordinates $$\begin{aligned} x_B&=\frac{1}{\lambda}\sqrt{\frac{3}{2}}\,, \\ y_B&=\frac{\sqrt{3}}{4|\lambda|}\left(1+\sqrt{4 \xi +1}\right)^{1/2} \,,\end{aligned}$$ while Point $C$ assumes the complicate values $$\begin{aligned} x_C&=\frac{\Delta ^{2/3}+4 \lambda \Delta^{1/3} -36 \xi \left(\lambda ^2+4\right)+7 \lambda ^2-36}{3 \sqrt{6}(4 \xi +1) \Delta^{1/3} }\,,\label{046} \\ y_C&=\frac{1}{\sqrt{6}}\Bigg[3 (x_C^2+1)-\sqrt{6} \lambda x_C \nonumber\\ &\qquad\quad+\sqrt{36 \xi x_C^4+\left(3 x_C^2-\sqrt{6} \lambda x_C+3\right)^2}\Bigg]^{1/2} \,,\label{047}\end{aligned}$$ where $$\begin{gathered} \Delta=54 \left(48 \xi ^2+8 \xi -1\right) \lambda +(10-216 \xi )\lambda ^3 \\ +9 (4 \xi +1) \Big[36 \xi \left(\lambda ^6-12 \lambda ^4+24 \lambda ^2+64\right) \\ +5184 \xi ^2 \lambda ^2-\left(\lambda ^2-6\right)^2 \left(3 \lambda^2-16\right)\Big]^{1/2} \,.\end{gathered}$$ Note the complexity of the coordinates of Point $C$ where the best expression one can obtain for $y_C$ is only in terms of $x_C$. Finally to obtain the effective EoS parameter and $\Omega_\phi$ for Point $C$ one should insert expressions (\[046\])-(\[047\]) into (\[048\]) and (\[054\]) respectively. These values has not been displayed due to their lenght. [99]{} A. G. Riess [*et al.*]{} \[Supernova Search Team Collaboration\], Astron. J.  [**116**]{}, 1009 (1998) \[astro-ph/9805201\]. S. Perlmutter [*et al.*]{} \[Supernova Cosmology Project Collaboration\], Astrophys. J.  [**517**]{}, 565 (1999) \[astro-ph/9812133\]. S. Weinberg, Rev. Mod. Phys.  [**61**]{}, 1 (1989). J. Martin, Comptes Rendus Physique [**13**]{}, 566 (2012) \[arXiv:1205.3365 \[astro-ph.CO\]\]. I. Zlatev, L. -M. Wang and P. J. Steinhardt, Phys. Rev. Lett.  [**82**]{}, 896 (1999) \[astro-ph/9807002\]. E. J. Copeland, M. Sami and S. Tsujikawa, Int. J. Mod. Phys. D [**15**]{}, 1753 (2006) \[hep-th/0603057\]. S. Tsujikawa, Class. Quant. Grav.  [**30**]{}, 214003 (2013) \[arXiv:1304.1961 \[gr-qc\]\]. A. R. Liddle and D. H. Lyth, Cambridge, UK: Univ. Pr. (2000) 400 p J. Magana and T. Matos, J. Phys. Conf. Ser.  [**378**]{}, 012012 (2012) \[arXiv:1201.6107 \[astro-ph.CO\]\]. D. Bertacca, N. Bartolo and S. Matarrese, Adv. Astron.  [**2010**]{}, 904379 (2010) \[arXiv:1008.0614 \[astro-ph.CO\]\]. C. Armendariz-Picon, V. F. Mukhanov and P. J. Steinhardt, Phys. Rev. D [**63**]{}, 103510 (2001) \[astro-ph/0006373\]. A. Vikman, Phys. Rev. D [**71**]{}, 023515 (2005) \[astro-ph/0407107\]. G. -B. Zhao, J. -Q. Xia, M. Li, B. Feng and X. Zhang, Phys. Rev. D [**72**]{}, 123515 (2005) \[astro-ph/0507482\]. R. R. Caldwell and M. Doran, Phys. Rev. D [**72**]{}, 043527 (2005) \[astro-ph/0501104\]. J. -Q. Xia, H. Li and X. Zhang, Phys. Rev. D [**88**]{}, 063501 (2013) \[arXiv:1308.0188 \[astro-ph.CO\]\]. B. Novosyadlyj, O. Sergijenko, R. Durrer and V. Pelykh, arXiv:1312.6579 \[astro-ph.CO\]. R. R. Caldwell, Phys. Lett. B [**545**]{}, 23 (2002) \[astro-ph/9908168\]. Y. -F. Cai, E. N. Saridakis, M. R. Setare and J. -Q. Xia, Phys. Rept.  [**493**]{}, 1 (2010) \[arXiv:0909.2776 \[hep-th\]\]. E. J. Copeland, A. RLiddle and D. Wands, Phys. Rev. D [**57**]{}, 4686 (1998) \[gr-qc/9711068\]. L. Amendola, Phys. Rev. D [**60**]{}, 043501 (1999) \[astro-ph/9904120\]. O. Hrycyna and M. Szyd?owski, JCAP [**1312**]{}, 016 (2013) \[arXiv:1310.1961 \[gr-qc\]\]. H. Wei, Phys. Lett. B [**712**]{}, 430 (2012) \[arXiv:1109.6107 \[gr-qc\]\]. C. Xu, E. N. Saridakis and G. Leon, JCAP [**1207**]{}, 005 (2012) \[arXiv:1202.3781 \[gr-qc\]\]. J. M. Aguirregabiria and R. Lazkoz, Phys. Rev. D [**69**]{}, 123502 (2004) \[hep-th/0402190\]. G. Leon and E. N. Saridakis, JCAP [**1303**]{}, 025 (2013) \[arXiv:1211.3088 \[astro-ph.CO\]\]. X. -m. Chen, Y. -g. Gong and E. N. Saridakis, JCAP [**0904**]{}, 001 (2009) \[arXiv:0812.1117 \[gr-qc\]\]. R. Lazkoz and G. Leon, Phys. Lett. B [**638**]{}, 303 (2006) \[astro-ph/0602590\]. R. -J. Yang and G. Xiang-Ting, Class. Quant. Grav.  [**28**]{}, 065012 (2011) \[arXiv:1006.4986 \[gr-qc\]\]. J. De-Santiago, J. L. Cervantes-Cota and D. Wands, Phys. Rev. D [**87**]{}, 023502 (2013) \[arXiv:1204.3631 \[gr-qc\]\]. N. Tamanini and C. G. Böhmer, Phys. Rev. D [**87**]{}, no. 8, 084031 (2013) \[arXiv:1302.2355 \[gr-qc\]\]. T. S. Koivisto and N. J. Nunes, Phys. Rev. D [**80**]{}, 103509 (2009) \[arXiv:0908.0920 \[astro-ph.CO\]\]. C. G. Boehmer, N. Chan and R. Lazkoz, Phys. Lett. B [**714**]{}, 11 (2012) \[arXiv:1111.6247 \[gr-qc\]\]. A. P. Billyard and A. A. Coley, Phys. Rev. D [**61**]{}, 083503 (2000) \[astro-ph/9908224\]. C. G. Boehmer, G. Caldera-Cabral, R. Lazkoz and R. Maartens, Phys. Rev. D [**78**]{}, 023505 (2008) \[arXiv:0801.1565 \[gr-qc\]\]. C. G. Boehmer, G. Caldera-Cabral, N. Chan, R. Lazkoz and R. Maartens, Phys. Rev. D [**81**]{}, 083003 (2010) \[arXiv:0911.3089 \[gr-qc\]\]. R. Bean, S. H. Hansen and A. Melchiorri, Phys. Rev. D [**64**]{}, 103508 (2001) \[astro-ph/0104162\]. F. Piazza and S. Tsujikawa, JCAP [**0407**]{} (2004) 004 \[hep-th/0405054\]. S. Tsujikawa and M. Sami, Phys. Lett. B [**603**]{} (2004) 113 \[hep-th/0409212\]. [^1]: If $w_{\rm eff}<-1$ then the physical solution for the scale factor in Eq. (\[007\]) should be $a\propto (t_0-t)^{2/[3(1+w_{\rm eff})]}$, which implies a Big Rip at $t=t_0$. [^2]: In [@Piazza:2004df; @Tsujikawa:2004dp] this Lagrangian was writtes as $\mathcal{L}_\phi= X\,f(B)$, however a simple redefinition of the function $f$ can bring this in the form (\[013\]). [^3]: Defining $x=r\cos\theta$ and $y=r\sin\theta$ the limit $r\rightarrow 0$ always well-behaves but for the angles $\theta=0,\pi$ which however, corresponding to $y=0$, never happen to be part of the phase space.

ArXiv

-2.5cm -0.2cm P-and-T–VIOLATION TESTS WITH POLARIZED RESONANCE NEUTRONS V. E. Bunakov, Y. Novikov\ Petersburg Nuclear Physics Institute, 188350 Gatchina, Russia\ e-mail: [email protected] PACS numbers: 11.30.Er, 13.88.+e, 14.20.Dh [**Keywords:**]{} CP-violation, resonance neutrons\ [**Abstract**]{}: The enhancements of CP-violating effects in resonance neutron transmission through polarized targets are studied for 2 possible versions of experiment. The importance is stressed of error analysis and of pseudomagnetic effects’ compensation. [**1. Introduction.**]{} It was shown \[1-3\] about 15 years ago that CP-violation effects in transmission of polarized neutrons through the polarized target might be enhanced in the vicinity of p-resonances by 5-6 orders of magnitude. Originally it was suggested to measure the difference in transmission of neutrons with spins parallel ($N_+$) and antiparallel ($N_+$) to the vector $\vec{k}_n\times \vec{I}$ ($\vec{k}_n$ and $\vec{I}$ are the neutron momentum and the target spin):$$\eta_T=\frac{N_+-N_-}{N_++N_-}\approx 2\frac{\sigma_+-\sigma_-} {\sigma_++\sigma_-}$$ Here $N_+$ and $N_-$ are the numbers of neutrons with the corresponding helicities transmitted through the polarized target sample, $\sigma_+$ and $\sigma_-$ are the corresponding total neutron cross-sections. However, it was pointed \[4\] that without the special precautions the nuclear pseudo-magnetic precession of neutron spin together with the precession induced by the P-violating interactions would give rise to numerous effects camouflaging the CP-violating ones. As a possible remedy of this nuisance it was suggested \[4\] to compensate the nuclear pseudo-magnetic field by the external magnetic field in order to nullify the neutron spin rotation angle $\phi$. However, in order to measure the CP-violating interaction with the reasonable accuracy (about $10^{-4}$ of the P-violating one) it was necessary to check the spin rotation angle with the precision of about $ 10^{-7}$ rad \[4\]. In order to circumvent the above difficulties Stodolsky \[5\] suggested to measure the difference $N_{+-}-N_{-+}$, where $N$ is the number of neutrons transmitted through the target and the subscript indices mean the neutron helicity before and after the transmission. Consider the polarized neutron scattering amplitude of the form:$$f = A + p_tB\cdot(\vec{s}_n \cdot \vec{I}) + C\cdot(\vec{s}_n \cdot \vec{k}_n) + p_tD\cdot \vec{s}_n \cdot [\vec{k}_n \times \vec{I}]$$ where $\vec{s}_n$ is neutron spin, $p_t$ is the target degree of polarization, $A$ and $B$ are the spin-independent and spin-dependent parts of the strong interaction amplitude, $D$ is the P- and CP-violating interaction amplitude, respectively. The term $C$ contains contributions from both weak P-violating and strong interaction (from the term of the type $(\vec{s}_n\cdot \vec{k}_n)(\vec{k}_n \cdot \vec{I})$ in scattering amplitude -see e.g. \[6\]). Stodolsky demonstrated that the difference $$N_{+-}-N_{-+} \sim Im(DB^*)$$ is free from the above camouflaging effects. It is well-known that in order to improve the accuracy it is preferable to measure the relative values, i.e. to normalize the above difference. Although Stodolsky never bothered to introduce this normalization, it seems natural to consider the ratio:$$T = \frac{N_{+-} - N_{-+}}{N_{+-} + N_{-+}}$$ A few years later Serebrov \[7\] suggested to measure the quantity:$$X = \frac{(N_{++}-N_{--}) + (N_{+-} - N_{-+})} {(N_{++}-N_{--}) - (N_{+-} - N_{-+})}$$ One can easily see that $$X = 1 + 2 \frac{N_{+-} - N_{-+}}{(N_{++}-N_{--}) - (N_{+-} - N_{-+})}$$ The actual CP-violating effect causes the deviation of $X$ from unity. Therefore the actually measured quantity $\tilde{X}$ $$\tilde{X} = \frac{N_{+-} - N_{-+}}{(N_{++}-N_{--}) - (N_{+-} - N_{-+})}$$ is simply the one suggested by Stodolsky, but normalized in a rather odd manner. The main point is that up to now nobody cared to do the analysis of the energy dependence of the quantities $T$ or $\tilde{X}$ in the manner it was done for the originally considered CP-violating quantity $\eta_T$ in refs. \[1-3\]. Indeed , all the quantities in the numerators and denominators of $T$ and $\tilde{X}$ contain various combinations of real and imaginary parts of all the four amplitudes ($A$, $B$, $C$ and $D$) in Eq. (2). Most of them show a rather complicated energy dependence (see e.g. \[3,6,8\]) in the resonance region. Some of them not only vary in magnitude, but even change their sign. This means that up to now one does not know whether the suggested values $T$ and $\tilde{X}$ are really enhanced and what is the magnitude of this enhancement, if any. Investigation of these problems is the main point of our present publication. For the time being we are not going to consider the false effects arising from the difference of the polarizing and the analyzing power of polarizer and analyzer. We shall also restrict ourselves with cases of “ideal geometry” when the incident beam polarization is either parallel or anti-parallel to the neutron momentum.\ [**2. Analysis of $T$.**]{} In order to obtain the expressions for the relative quantities of interest in terms of the energy-dependent complex amplitudes $A$, $B$, $C$ and $D$ of Eq. (2), one might use the method developed in ref. \[9\]. Introducing the spin density matrix and the evolution operators of ref. \[9\], one obtains the expression for $T$:$$T = 2 \frac{Im(DB^*)}{|D|^2 + |B|^2}$$ The expressions for complex amplitudes $D(E)$ and $B(E)$ are obtained using the methods of ref. \[3\] (see also \[6\], \[8\]). The main contribution to the T-noninvariant amplitude $D$ in the vicinity of the $p_{I+1/2}$ -resonance comes from the term coupling this resonance with the corresponding $s_{I+1/2}$-resonance:$$D \approx \frac{\gamma^{s}_{I+1/2}}{(E-E^s_{I+1/2}) + \imath \Gamma_s/2} V_T \frac{\gamma^{p}_{I+1/2}}{(E-E^p_{I+1/2}) + \imath \Gamma_p/2}.$$ In the optimal cases (like $La$ target) these $s_{I+1/2}$ and $ p_{I+1/2}$ resonances contribute equally to the strong amplitude $B$ in this energy region. Taking into account all the other resonances would only lead to some numerical changes, while the general qualitative picture would be the same. Therefore we consider:$$B \approx \frac{\gamma^s_{I + 1/2} \gamma^s_{I + 1/2}} {(E-E^s_{I + 1/2}) + \imath \Gamma_s/2} + \frac{\gamma^p_{I + 1/2} \gamma^p_{I + 1/2}}{(E-E^p_{I + 1/2}) + \imath \Gamma_p/2},\nonumber \\$$ Inserting these expressions into eq. (8) we see that the quantity $T(E)$ in the vicinity of the p-wave resonance energy $E_p$ is:$$T(E)\approx -2\frac{\gamma^p_{I+1/2}}{\gamma^s_{I+1/2}} \cdot \frac{V_T \cdot \Gamma_p}{(E-E_p)^2+\Gamma_p ^2/4}$$ Here $\Gamma_p$ stands for the p-resonance total width, while $V_T$ is the matrix element of CP-violating interaction causing the transition between the p- and s-resonance states. Further on in our numerical calculations we shall assume the ratio of the CP-violating interaction strength to the P-violating one to be $10^{-4}$ (i.e. $V_T/V_P=10^{-4}$). The quantities $\gamma^{s,p}_J$ stand for the neutron width amplitudes of the s- and p-resonances with spin $J=I+1/2$. The sign of the effect is defined by the signs of $\gamma$’s and $V_T$. For the sake of simplicity we shall choose them in our numerical calculations so that the net effect is positive. We observe in Eq. (11) the resonance enhancement of the effect typical for all the symmetry-breaking effects in nuclear reactions (see \[3,8\]). In order to see explicitly the “dynamical enhancement”, which is also typical for these effects, one might cast the value of $T$ in this maximum in the following form:$$T \left( E = E^p_{I+1/2} \right) \approx \frac{\gamma^p_{I+1/2}}{\gamma^s_{I+1/2}} \cdot \frac{V_T}{d} \cdot \frac{d}{\Gamma}$$ Here $d$ and $\Gamma$ stands for the average resonance spacing and total width. It is instructive to remind that the corresponding expression for the maximal value of the quantity $\eta_T$ obtained in \[1-3\] was: $$\eta_T(E=E_p)\approx\frac{\gamma^s}{\gamma^p} \cdot \frac{V_T}{d} \cdot (\frac{d}{\Gamma})^2$$ Comparing Eqs.(12) and (13), one can see in both cases the presence of the dynamical enhancement factors $V_T/d\approx F_T\cdot 10^3$ ($F_T$ is the strength of the CP-violating interaction relative to the strong interaction one) and of the resonance enhancement factors $d/\Gamma\approx 10^3$ coming from the fact that the effect is proportional to the time $\tau\sim(1/\Gamma)$ spent by the incident neutron in the CP-violating field of the target. We also see the presence of the “entrance channel hindrance” factor (see \[3,8\]) $\gamma^p/\gamma^s\approx 10^{-3}$ typical for all the low energy scattering experiments with P-violation. However, the resonance enhancement factor enters the quantity $\eta_T$ quadratically, while $T$ contains it only linearly. Therefore the net enhancement of the $T$ quantity is only by a factor of $10^3$ instead of the $10^5\div 10^6$ factor in $\eta_T$. These conclusions are illustrated in Fig. 1a, where the energy behavior of the quantity $T(E)$ is shown for the particular case of the famous $La$ p-resonance at $E_p=0.75$ eV. Consider now a very important problem of the optimal choice of the target sample thickness. One should mind that in the case of $\eta_T$ value, likewise in the case of the longitudinal polarization $P$ caused by the P-violating weak interaction, the correct expression for the experimentally measured ratio can be written as follows (see e.g. \[10\], \[3\] and \[11\]):$$\frac{N_+-N_-}{N_++N_-}\approx\frac{\sigma_+-\sigma_-}{2}\cdot x\cdot\rho$$ where $x$ is the target sample thickness and $\rho$ is the density of nuclei in this sample. Since the experimentally observed effect is linear in target thickness, it seems that one should choose the thickest target possible. However, the neutron countings $N_{\pm}$ decrease exponentially with $x$. Therefore the statistical relative error of measuring each $N$ value $$\frac{\delta N}{N}=\frac{1}{\sqrt{N}}=\frac{1}{\sqrt{N_0}}e^{x\rho\sigma/2}$$ also increases exponentially with $x$ ($N_0$ here stands for the number of polarized neutrons incident on the target). In order to find the optimal target thickness $x_0$ one should estimate the relative error $\sigma_{\eta_T} /\eta_T$ of the quantity in the l.h.s. of Eq. (14) and define its minimum (by equating the x-derivative of the relative error to zero). In this way one obtains that the optimal thickness in the case of $\eta_T$ quantity is $2\lambda=2/\sigma\rho$ (here $\lambda$ stands for the mean free path of the neutron in the target sample). It is only by choosing the optimal $x_0$ that one obtains the last line in Eq. (1). The relative error of the quantity $T$ looks more complicated. One can easily see that the main contribution to it comes from the relative error of the numerator in $T$: $$\frac{\sigma_T}{T} \approx \frac{e^{\frac{Im(A)}{Im(f)} \frac{x}{\lambda}}}{\sqrt{2N_0}} \cdot \frac{\left| q \right|}{\sin(\theta)} \cdot \frac{1}{\sqrt{ ch^2 \left( \frac{Im(q)}{Im(f)} \frac{x}{\lambda} \right) - \cos^2 \left( \frac{Re(q)}{Im(f)} \frac{x}{\lambda} \right) }} \cdot \sqrt{\frac{|D|^2+|B|^2}{Im^2(DB^*)}}$$ Here $\lambda$ is the neutron mean free path and the (complex) quantity $q$ is defined as:$$q = \sqrt{ \frac14 \sin^2(\theta) B^2 + \frac14 \sin^2(\theta) D^2 + \frac14 (C + \cos(\theta) B)^2 }$$ The angle between the target polarization and neutron momentum vectors is denoted as $\theta$. The $\sin(\theta)$ behaviour of Eqs. (16) reflects the fact that the CP-violating term in the amplitude (2) is proportional to $\sin(\theta)$. Therefore irrespective of the value of $D$ the CP-violating effects disappear for $\theta\approx 0$ and the relative error goes to infinity. The dependence of Eq. (16) on the target thickness $x$ is complicated by the periodic $\cos^2$ oscillations. The physical origin of those oscillations is the pseudo-magnetic neutron spin rotation, discussed in ref. \[4\] - the neutron spin performs about a hundred rotations per mean free path in the target sample. The explicit dependence of the relative error (16) on the target thickness is shown in Fig. 1b for the case of the same p-resonance in $La$. The total number of the polarized neutrons $N_0$ incident on the target was somewhat arbitrary chosen to be $10^{18}$. One can see from Fig. 1b that the first minimum of the relative error is located about $x\approx 10^{-2}\lambda$. However, a slight change of $x$ increases the relative error by orders of magnitude, which makes the analysis of the experimental results practically impossible. This forces us to return to our initial idea \[4\] of compensating the pseudomagnetic precession by the external magnetic field. This field can be formally taken into account by substituting $Re (B)$ in the initial Eq. (2) by:$$Re(B') = Re(B) - H$$ Here $H$ stands for the value of the external magnetic field. Since the “pseudo-magnetic” amplitude $B(E)$ is energy-dependent, we can do the compensation by, say, putting $Re B'(E)=0$ at $E=E_p+\Gamma_p/2$. Fig. 1d shows the dependence of relative error on $x$ with this compensation. As expected, all the oscillations of Fig. 1b disappear and the relative error shows a minimum at around $x\approx 2.5\lambda$. However, the effect $T$ itself depends on the value of $Re B(E)$ - see Eq. (8). Without the compensation $Re( B)\gg Im (B)$ (approximately by 3 orders of magnitude) and the dominant contribution to the denominators and numerators of Eq. (8) comes from it. If we do the above compensation, then $Im(B){\raisebox{-.5ex}{\mbox{$\, \stackrel{>}{\sim}$}}}Re(B)$ and the effect in the vicinity of p-resonance ($|E-E_p| \leq \Gamma_p$) can be expressed as:$$T' \approx -2 \frac{Re(D)}{Im(B)}$$ Taking into account the energy dependence of the amplitudes, we get:$$T'\approx 4\frac{\gamma^p}{\gamma^s} \cdot \frac{V_T}{\Gamma_s} \cdot \frac{d(E-E_p)}{(E-E_p)^2+\Gamma^2/4}$$ Therefore the effect now changes sign at around the resonance energy $E_p$ and reaches at the points $E\approx E_p\pm\Gamma_p$ its maximal value:$$T' \approx \frac{\gamma^p}{\gamma^s}\cdot \frac{V_T}{\Gamma} \cdot \frac{d}{\Gamma}$$ Comparing this result with Eqs. (11), (12), we see that the compensating magnetic field, besides removing the oscillations of the relative error, also produced a important increase of the value of $T$ itself, giving an extra resonance enhancement factor $ d/\Gamma\sim 10^3$. It also radically changed the energy-dependence of the effect. By comparing Eqs. (8) and (16) we see that the relative error in the presence of compensation decreases by the same 3 orders of magnitude. These conclusions are confirmed by the results of calculating the effect under conditions of complete compensation $Re(B'(E_p+\Gamma_p/2))=0$ - see Fig. 1c. Thus we see, that our initial idea \[4\] of compensation the pseudomagetism turns out still to be quite productive. The only remaining point is to estimate the practically necessary accuracy of this compensation. Following \[4\], we still think that the practical way of controlling this accuracy is by measuring the neutron spin rotation angle $\phi=2Re(B)/Im(f)\cdot x/\lambda$ around $\vec{I}$ after its transmission through the target sample. Fig. 2a shows the dependence of the effect $T(E=E_p+\Gamma_p/2)$ on the spin rotation angle (which serves as a measure of the applied compensating magnetic field). Fig. 2b shows the same dependence for the relative error. We see that both the effect and its relative error are optimal for practically complete compensation ($\phi\approx 0$). The slight shift of optima to small positive $\phi$ is caused by the interference of contributions to the effect from the pseudo-magnetic rotation and rotation caused by the T-noninvariant field $D$. However, the relative error changes only by a factor of 2 - 3 when the rotation angle varies from $0^0$ to $200^0$. Thus the limitations on the accuracy of compensation are quite moderate from this point. A more essential limitation might come from the fact that the energy dependence of the effect (and, to somewhat less extend, its maximal value) changes rapidly with increasing $\phi$. In order to see this, one might compare the curves in Fig. 1c (corresponding to $\phi=0^0$) and Fig. 1a, calculated without compensation. Therefore we decided to formulate the problem of the compensation accuracy in a slightly different way: We assume that a reasonable value for the experimental energy resolution is $\Delta E\approx 10^{-2}$ eV and consider the practically reasonable accuracy $\Delta \phi$ of measuring $\phi$ as a free parameter. Then the rotation angle $\phi$ (and thus the compensating field $H$) should be chosen in such a way that energy maximum of the effect $T(E)$ should be shifted by less than $\Delta E$ while varying the rotation angle in the interval from $\phi -\Delta\phi$ to $\phi +\Delta\phi$. On performing a good deal of “computer experiments” we can state, that the accuracy $\Delta\phi=5^0$ is quite sufficient from this point of view. Thus we see, that the limitations on the accuracy of measuring the rotation angle in order to check the compensation of pseudo-magnetic rotation are quite tolerable.\ [**3. Analysis of $\tilde{X}$.**]{} Consider now the quantity $\tilde{X}$. As already mentioned, it differs from $T$ only by the normalization factor. Therefore it is also enhanced in the vicinity of the p-wave resonance. However the new normalization makes the effect itself (and not only its relative error) dependent both on the angle $\theta$ and on the target thickness $x$. Moreover, the rapid energy oscillations are superimposed on the resonance behaviour of the effect. The character of these oscillations depend on the target thickness $x$ in a very complicated way. For the sake of illustration we show in Fig. 3 the energy dependence of $\tilde{X}$, calculated for $x \approx 2.5\lambda$. All this considerably complicates the analysis. It is difficult even to find a reasonable analytical approximation for $\tilde{X}$. In the case of thick target (for $La$ resonance this means $x{\raisebox{-.5ex}{\mbox{$\, \stackrel{>}{\sim}$}}}15\lambda$) one can write:$$\tilde{X} \approx - \frac{\sin^2(\theta) Im(DB^*)} {\sin^2(\theta) Im(DB^*) + 2Re\left( q (C^* + \cos(\theta)B^*)\right)}$$ Eq. (22) shows that in the thick target limit the rapid oscillations of the effect disappear. This makes the analysis of its energy and $\theta$ dependence much easier. Consider now the $\theta$ dependence of the numerator and the denominator in $\tilde{X}$ separately. The whole interval of $\theta$ values can be separated into two regions. In the first region one can neglect all the contributions to the denominator besides $\ 2 \cos(\theta)Re(q B^*)$. In this region $q \approx \frac12 B$ and $$\tilde{X} \approx -\frac{\sin^2(\theta)}{\cos(\theta)} \frac{Im(DB^*)}{|B|^2}.$$ One can see that in this region $\tilde{X} \sim -T$. Consider now the relative error of $\tilde{X}$ in this range of $\theta$. In analogy to the above $T$ case, the main contribution to this error comes from the numerator. Therefore $$\frac{\sigma_{\tilde{X}}}{\tilde{X}} \approx \frac{\sigma_T}{T},$$ This conclusion turns out to be valid even without the thick target approximation. Therefore the relative error of $\tilde{X}$ strongly oscillates with the variation of the target thickness. The necessity of the compensating external magnetic field is again obvious. Introducing this compensation, we again observe that the oscillations of $\sigma_{\tilde{X}}/\tilde{X} (x/\lambda)$ disappear, and it is possible to find the optimal target thickness (which is obviously $x \approx 2.5 \lambda$). In the first region of angles the value of the effect increases when $\theta$ approaches the critical point where the denominator of the effect equals zero. The relative error remains more or less constant. This does not mean, however, that it is better to make measurements closer to this critical point, because the absolute value of the error also increases. Therefore the accuracy of experimental observations remains practically the same. The second region of angles is characterized by the inequality $\cos(\theta)Re(q B^*) \ll Re(qC^*)$ and is located in the vicinity of $\theta = \pi/2$. The width of this region depends on the incident neutron energy and on the magnitude of the compensating magnetic field. Without the compensation this width is $10^{-4} \div 10^{-7}$ rad. In case of full compensation it increases to a few degrees. It is important to note that this region contains the value of $\theta$ which turns the denominator into zero, while the effect formally increases to infinity. In this second region the denominator’s contribution to the relative error dominates. Therefore the relative error continuously increases and becomes infinite in point where the denominator equals to zero. This means that one should not come too close to the values $\theta = \frac{\pi}{2}$ because of the finite angular divergence of any experimental beam. It is obviously practically impossible to estimate the actual accuracy of the measurements carried close to $\theta = \frac{\pi}{2}$. In order to estimate the necessary accuracy of compensation by the external magnetic field, one should perform the same kind of analysis as in the case of $T$. It is obvious that the results of such an analysis would be essentially the same: the main limitation on the accuracy should again come from the rapid change of the effect’s energy dependence. As in the case of $T$, the resulting limitations on $\Delta\phi$ accuracy are quite reasonable.\ [**4. Summary**]{}\ We can draw the following conclusions: Analysis of the CP-violating effect’s relative error is by no means less essential than analyzing the effect itself. One can always normalize the CP-noninvariant difference (3) dividing it by a very small quantity. However such a normalization would not increase the accuracy of the measurement. The necessity to compensate the pseudomagnetic precession is caused essentially by the fact that without such a compensation the accuracy of measurement varies with target thickness in a practically uncontrollable way. The compensation of the pseudomagnetic precession increases by 3 orders of magnitude not only the effect itself but also the accuracy of its measurement. The net enhancement in the vicinity of p-wave resonance with compensation reaches 6 orders of magnitude. The energy dependence of the effect changes drastically in the presence of compensation. As a practical way to control the degree of compensation we suggest, following \[3\], to measure the rotation angle $\phi$ of neutron polarization around the target polarization vector. When $\phi$ varies between $0^0$ and $200^0$ the maximal value of the effect and its relative error varies not more than by a factor of $2\div 3$. The most stringent restriction on the accuracy of measurement of this angle comes from the fact that the energy dependence of the measured effect strongly depends on the value of the compensating magnetic field. With this restriction in mind it seems sufficient to fix $\phi$ with the accuracy about $5^0$. The CP-noninvariant quantity $X$ suggested for measurement in ref. \[7\] is shown to differ from Stodolsky’s CP-noninvariant difference (3) practically only by the choice of normalization factor. This factor becomes zero in the vicinity of $\theta=\pi/2$ (the beam polarization orthogonal to the target one). Although the value of thus normalized effect tends to infinity, its relative error also tends to infinity in this range of $\theta$ values. In the remaining range of $\theta$, where the normalizing factor exceeds the CP-noninvariant difference (3), the relative error depends on $\theta$ angle as $1/\sin(\theta)$. Although the value of the effect in this range of $\theta$ angles behaves as $\tan(\theta)$ and strongly increases approaching the value $\theta\approx \pi/2$, the relative accuracy of its measurements (besides the small vicinity of $\theta\approx 0$) remains practically the same. We acknowledge the support of the Russian Fund of Fundamental Studies (grant No. 97-02-16803). [99]{} V. E. Bunakov, V. P. Gudkov, JETP Lett. [**36**]{} (1982) 38 V. E. Bunakov, V. P. Gudkov, Z.Phys. [**308**]{} (1982) 363 V. E. Bunakov, V. P. Gudkov, Nucl.Phys. A[**401**]{} (1983) 93 V. E. Bunakov, V. P. Gudkov, J.Phys.(Paris) [**45**]{} C3 (1984) 77 L. Stodolsky, Phys.Lett. B[**172**]{} (1986) 5 A. L. Barabanov, Nucl.Phys. A[**614**]{} (1997) 1 A. P. Serebrov, JETP Lett. [**58**]{} (1993) 14 V. E. Bunakov, Elementary Particles and Nuclear Physics [**26**]{} (1995) 287 S. K. Lamoreaux, R. Golub, Phys.Rev. D[**50**]{} (1994) 5632 V. P. Alfimenkov, Nucl.Phys. [**383**]{} (1983) 93 V. E. Bunakov, L. B. Pikelner, Prog.Part.Nucl.Phys. [**39**]{} (1997) 387

ArXiv

--- abstract: 'Domain adaptation is transfer learning which aims to generalize a learning model across training and testing data with different distributions. Most previous research tackle this problem in seeking a shared feature representation between source and target domains while reducing the mismatch of their data distributions. In this paper, we propose a close yet discriminative domain adaptation method, namely CDDA, which generates a latent feature representation with two *interesting* properties. First, the discrepancy between the source and target domain, measured in terms of both marginal and conditional probability distribution via Maximum Mean Discrepancy is minimized so as to *attract* two domains close to each other. More importantly, we also design a repulsive force term, which maximizes the distances between each label dependent sub-domain to all others so as to *drag* different class dependent sub-domains far away from each other and thereby increase the discriminative power of the adapted domain. Moreover, given the fact that the underlying data manifold could have complex geometric structure, we further propose the constraints of label smoothness and geometric structure consistency for label propagation. Extensive experiments are conducted on 36 cross-domain image classification tasks over four public datasets. The Comprehensive results show that the proposed method consistently outperforms the state-of-the-art methods with significant margins.' author: - | Lingkun Luo$^{1\ast}$, Xiaofang Wang $^{2}$ [^1], Shiqiang Hu $^{1}$, Chao Wang $^{1}$, Yuxing Tang $^{2}$, Liming Chen $^{2}$\ $^1 $ School of Aeronautics and Astronautics, Shanghai Jiao Tong University, Shanghai, China.\ $\{lolinkun1988, sqhu, wang\_chao\}[email protected]\ $^2$ LIRIS, CNRS UMR 5205, École Centrale de Lyon, 36 avenue Guy de Collongue, Écully, F-69134, France.\ $\{xiaofang.wang, yuxing.tang, liming.chen\}[email protected] bibliography: - 'egbib.bib' title: Close Yet Discriminative Domain Adaptation --- [UTF8]{}[gkai]{} Introduction ============ Thanks to deep networks, recent years have witnessed impressive progress in an increasing number of machine learning and computer vision tasks, *e.g.*, image classification[@ILSVRC15; @He2015], object detection [@Everingham15; @girshickICCV15fastrcnn], semantic segmentation [@Cordts2016Cityscapes; @Everingham15; @zhao2016pspnet]. However, these impressive progress have been made possible only when massive amount of labeled training data are available and such a requirement hampers their adoption to a number of real-life applications where labeled training data don’t exist or not enough in quantity. On the other hand, manual annotation of large training data could be extremely tedious and prohibitive for a given application. An interesting solution to this problem is transfer learning through *domain adaptation* [@pan2010survey]), which aims to leverage abundant existing labeled data from a different but related domain (source domain) and generalize a predictive model learned from the source domain to unlabeled target data (target domain) despite the discrepancy between the source and target data distributions. \[fig:diff\] {width="1\linewidth"} The core idea of most proposed methods for domain adaptation is to reduce the discrepancy between domains and learn a domain-invariant predictive model from data. State of the art has so far featured two mainstream algorithms in reducing data distribution discrepancy: (1) feature representation transfer, which aims to find “good” feature representations to minimize domain differences and the error of classification or regression models; and (2) instance transfer, which attempts to re-weight some “good” data from source domain, which may be useful for the target domain. It minimizes the distribution differences by re-weighting the source domain data and then trains a predictive model on the re-weighted source data. In this paper, we are interested in feature representation transfer which seeks a domain invariant latent space, while preserving at the same time important structure of original data, *e.g.*, data variance or geometry. Early methods, *e.g.*, [@blitzer2006domain], propose a structural correspondence learning (SCL), which first defines a set of pivot features and then identifies correspondences among features from different domains by modeling their correlations with the pivot features. Later, transfer learning problems are approached via dimensionality reduction. [@pan2011domain] learns a novel feature representation across domains in a Reproducing Kernel Hilbert Space with the Maximum Mean Discrepancy (MMD) measure [@borgwardt2006integrating], through the so-called transfer component analysis (TCA). TCA [@pan2011domain] is an extension of [@pan2008transfer], with the purpose to reduce computational burden. [@long2013transfer] goes one step further and remarks that both marginal and conditional distribution could be different between the source and target domains. As a result, Joint Distribution Adaptation (JDA) is proposed to jointly minimize the mismatches of marginal and conditional probability distributions. The previous research has thus so far only focused on matching marginal and/or conditional distributions for transfer learning while ignoring the discriminative properties to be reinforced between different classes in the adapted domain. In this paper, we propose to extract a latent shared feature space underlying the domains where the discrepancy between domains is reduced but more importantly, the original discriminative information between classes is simultaneously reinforced. Specifically, not only we seek to find a shared feature space in minimizing the discrepancy of both marginal and conditional probability distributions as in JDA [@long2013transfer], but also introduce a discriminative model, called subsequently as *repulsive force*, in light of the Fisher’s linear discriminant analysis (FLDA) [@fisher1936use]. This repulsive force *drags* the sub-domains with different labels far away from each other in maximizing their distances measured in terms of *Maximum Mean Discrepancy* (MMD), thereby making more discriminative data from different sub-domains. This is in clear contrast to the previous approaches as illustrated in Fig.\[fig:diff\]. Most previous works, *e.g.*,JDA, only seek to align marginal or conditional distributions between the source and target domain and the resultant latent subspace therefore falls short in terms of discrimination power as illustrated in the lower part of the green ellipse of Fig.\[fig:diff\](a), where samples of different labels are all mixed up. In contrast, as can be seen in the lower part of the purple ellipse of Fig.\[fig:diff\](b), the proposed method unifies the decrease of data distribution discrepancy and the increase of the discriminative property between classes into a same framework and finds a novel latent subspace where samples with same label are put close to each other while samples with different labels are well separated. Moreover, given the fact that the manifold of both source and target data in the shared latent feature space could have complex geometric structure, we further propose label propagation based on the respect of two constraints, namely label smoothness consistency (LSC) and geometric structure consistency (GSC), for the prediction of target data labels. That is, a good label propagation should well preserve the label information(constraint LSC) and not change too much from the shared data manifold (constraint GSC). To sum up, the contributions in this paper are threefold: - A novel repulsive force is proposed to increase the discriminative power of the shared latent subspace, aside of decreasing both the marginal and conditional distributions between the source and target domains. - Unlike a number of domain adaptation methods, *e.g.*, JDA [@long2013transfer], which use Nearest Neighbor(NN) with Euclidean distance to predict labels in target domain, the prediction in the proposed model, is deduced via label propagation in respect of the underlying data manifold geometric structure. - Extensive experiments are conducted on comprehensive datasets, and verify the effectiveness of the proposed method which outperforms state-of-the-art domain adaptation algorithms with a significant margin. The rest of the paper is organized as follows. In Section 2, we discuss previous works related to ours and highlight their differences. In Section 3, first we describe the problem and preliminaries of domain adaptation and then we present our proposed method. Experiment results and discussions are presented in Section 4 and finally we draw the conclusion in Section 5. Related Work ============ In this section, we discuss previous works which are related to our method and analyze their differences. In machine learning, domain adaptation is transfer learning which aims to learn an effective predictive model for a target domain without labeled data in leveraging abundant existing labeled data of a different but related source domain. Because the collection of large labeled data as needed in traditional machine learning is often prohibitive for many real-life applications, there is an increasing interest on this *young* yet *hot* topic [@pan2010survey][@weiss2016survey]. According to the taxonomy made in recent surveys [@pan2010survey][@weiss2016survey] [@long2013transfer], the proposed method falls down into the feature representation category. Recent popular methods embrace the dimensionality reduction to seek a latent shared feature space between the source and the target domain. Its core idea is to project the original data into a low-dimensional latent space with preserving important structure of original data. However, [@pan2008transfer] points out that direct application of Principal Component Analysis (PCA) can not guarantee the preservation of discriminative data structures. Their proposed remedy is to maximize the variance of the embedded data. Another interesting idea in [@pan2008transfer] is the use of a nonparametric criterion, namely *Maximum Mean Discrepancy* (MMD), based on Reproducing Hilbert Space (RKHS) [@borgwardt2006integrating], to estimate the distance between two distributions. Later, [@pan2011domain] further improves [@pan2008transfer] in terms of computational efficiency. With JDA, [@long2013transfer] goes one step further and propose not only to minimize the mismatch of the cross-domains marginal probability distributions but also their conditional probability distributions based on the framework of [@pan2008transfer; @pan2011domain]. The proposed framework in this paper can be considered as an extension of JDA with two major differences. First, we seek not only for a latent subspace which minimizes the mismatch of both the marginal and conditional probability distributions across domains, but also reinforces the discriminative structure of sub-domains in original data. We achieve this goal in introducing a novel term which acts as repulsive force to drag away different sub-domains both in source and target domain, respectively. Note that we do not discuss the line of work in the literature on transfer learning which is embedded into deep convolutional neural network as the features used in this work are not deep features; Nevertheless we have noticed their impressive performance, thanks to the combination of the latest advances in transfer learning discussed above with the cutting-edge understanding on the transferability [@glorot2011domain] of state-of-the-art deep neural networks, *e.g.*, Deep Adaptation Network(DAN) [@long2015learning], *etc.* Mixing seamlessly our proposed transfer knowledge model with state-of-the-art deep networks will be the subject of our upcoming investigation. Close Yet Discriminative Domain Adaptation ========================================== In this section, we present in detail the proposed Close yet Discriminative Domain Adaptation (CDDA) method. Problem Statement ----------------- We begin with the definitions of notations and concepts most of which we borrow directly from [@long2013transfer]. A domain $D$ is defined as an m-dimensional feature space $\chi$ and a marginal probability distribution $P(x)$, *i.e.*, $\mathcal{D}=\{\chi,P(x)\}$ with $x\in \chi$. Given a specific domain $D$, a task $T$ is composed of a C-cardinality label set $\mathcal{Y}$ and a classifier $f(x)$, *i.e.*, $T = \{\mathcal{Y},f(x)\}$, where $f({x}) = \mathcal{Q}( y |x)$ which can be interpreted as the class conditional probability distribution for each input sample $x$. In unsupervised domain adaptation, we are given a source domain $\mathcal{D_S}=\{x_{i}^{s},y_{i}^{s}\}_{i=1}^{n_s}$ with $n_s$ labeled samples, and a unlabeled target domain $\mathcal{D_T}=\{x_{j}^{t}\}_{j=1}^{n_t}$ with $n_t$ unlabeled samples with the assumption that source domain $\mathcal{D_S}$ and target domain $\mathcal{D_T}$ are different, *i.e.*, $\mathcal{\chi}_S=\mathcal{{\chi}_T}$, $\mathcal{Y_S}=\mathcal{Y_T}$, $\mathcal{P}(\mathcal{\chi_S}) \neq \mathcal{P}(\mathcal{\chi_T})$, $\mathcal{Q}(\mathcal{Y_S}|\mathcal{\chi_{S}}) \neq \mathcal{Q}(\mathcal{Y_T}|\mathcal{\chi_{T}})$. We also define the notion of sub-domain, denoted as ${\cal D}_{\cal S}^{(c)}$, representing the set of samples in ${{\cal D}_{\cal S}}$ with label $c$. Similarly, a sub-domain ${\cal D}_{\cal T}^{(c)}$ can be defined for the target domain as the set of samples in ${{\cal D}_{\cal T}}$ with label $c$. However, as ${{\cal D}_{\cal T}}$is the target domain with unlabeled samples, a basic classifier, *e.g.*, NN, is needed to attribute pseudo labels for samples in ${{\cal D}_{\cal T}}$. The aim of the Close yet Discriminative Domain Adaptation (CDDA) is to learn a latent feature space with following properties: 1) the distances of both marginal and conditional probability of source and target domains are reduced; 2) The distances between each sub-domain to the others, are increased in order to push them far away from each other; 3) The deduction of label prediction is imposed via two constraints, *i.e.*, label consistency and geometric structure of label space. Latent Feature Space with Dimensionality Reduction -------------------------------------------------- The finding of a latent feature space with dimensionality reduction has been demonstrated useful in several previous works, *e.g.*, [@pan2008transfer; @pan2011domain; @long2013transfer], for domain adaptation. One of its important properties is that original data is projected to a lower dimensional space which is considered as *principal* structure of data. In the proposed method, we also apply the Principal Component Analysis (PCA). Mathematically, given with an input data matrix $\boldsymbol{X} = [{\mathcal{D_S}},\mathcal{D_T}]$, $\boldsymbol{X} \in {\mathbb{R}^{m\times({n_s} + {n_t})}}$, the centering matrix is defined as $\boldsymbol{H} = \boldsymbol{I} - \frac{1}{n_s+n_t}\boldsymbol{1}$, where $\boldsymbol{1}$ is the $(n_s+n_t) \times (n_s+n_t)$ matrix of ones. The optimization of PCA is to find a projection space $\boldsymbol{A}$ which maximizes the embedded data variance. $$\label{eq:pca} \begin{array}{c} \mathop {\max}\limits_{\boldsymbol{A^TA} = \boldsymbol{I}} tr(\boldsymbol{A}^T\boldsymbol{ XH}\boldsymbol{X}^T \boldsymbol{A}) \end{array}$$ where $tr(\mathord{\cdot})$ denotes the trace of a matrix, $\boldsymbol{XH}\boldsymbol{X}^T$ is the data covariance matrix, and $\bf A \in \mathbb{R}^{m \times k}$ with $m$ the feature dimension and $k$ the dimension of the projected subspace. The optimal solution is calculated by solving an eigendecomposition problem: $\boldsymbol{XH}\boldsymbol{X}^T=\boldsymbol{A\Phi}$, where $\boldsymbol{\Phi}=diag(\phi_1,\dots, \phi_k )$ are the $k$ largest eigenvalues. Finally, the original data $\boldsymbol{X}$ is projected into the optimal $k$-dimensional subspace using $\boldsymbol{Z} = \boldsymbol{A}^T\boldsymbol{X}$. Closer: Marginal and Conditional Distribution Domain Adaptation ---------------------------------------------------------------- However, the feature space calculated via PCA is not sufficiently *good* enough for our problem of domain adaptation problem, for PCA only seeks to maximize the variance of the projected data from the two domains and does not explicitly reduce their distribution mismatch [@long2013transfer; @long2015learning]. Since the distance of data distributions across domain can also be empirically measured , we explicitly leverage the nonparametric distance measurement MMD in RKHS [@borgwardt2006integrating] to compute the distance between expectations of source domain and target domain, once the original data projected into a low-dimensional feature space via. Formally, the empirical distance of the two domains is defined as: $$\label{eq:marginal} \begin{array}{l} Dis{t^{marginal}}({{\cal D}_{\cal S}},{{\cal D}_{\cal T}}) =\\ {\left\| {\frac{1}{{{n_s}}}\sum\limits_{i = 1}^{{n_s}} {{{\bf{A}}^T}{x_i} - } \frac{1}{{{n_t}}}\sum\limits_{j = {n_s} + 1}^{{n_s} + {n_t}} {{{\bf{A}}^T}{x_j}} } \right\|^2} = tr({{\bf{A}}^T}\bf{X}{\bf{M_0}}\bf{{X^T}A}) \end{array}$$ where ${{\bf{M}}_0}$ represents the marginal distribution between ${{\cal D}_{\cal S}}$ and ${{\cal D}_{\cal T}}$ and its calculation is obtained by: $$\label{eq:M0} \begin{array}{l} {({{\bf{M}}_0})_{ij}} = \left\{ \begin{array}{l} \frac{1}{{{n_s}{n_s}}},\;\;\;{x_i},{x_j} \in {D_{\cal S}}\\ \frac{1}{{{n_t}{n_t}}},\;\;\;{x_i},{x_j} \in {D_{\cal T}}\\ 0,\;\;\;\;\;\;\;\;\;\;\;\;otherwise \end{array} \right. \end{array}$$ where ${x_i},{x_j} \in (\mathcal{D_S} \cup \mathcal{D_T})$. The difference between the marginal distributions $\mathcal{P}(\mathcal{X_S})$ and $\mathcal{P}(\mathcal{X_T})$ is reduced in minimizing [$Dis{t^{marginal}}({{\cal D}_{\cal S}},{{\cal D}_{\cal T}})$]{}. Similarly, the distance of conditional probability distributions is defined as the sum of the empirical distances over the class labels between the sub-domains of a same label in the source and target domain: $$\label{eq:conditional} \begin{array}{c} \begin{array}{l} Dis{t^{conditional}}\sum\limits_{c = 1}^C {({{\cal D}_{\cal S}}^c,{{\cal D}_{\cal T}}^c)} = \\ {\left\| {\frac{1}{{n_s^{(c)}}}\sum\limits_{{x_i} \in {{\cal D}_{\cal S}}^{(c)}} {{{\bf{A}}^T}{x_i}} - \frac{1}{{n_t^{(c)}}}\sum\limits_{{x_j} \in {{\cal D}_{\cal T}}^{(c)}} {{{\bf{A}}^T}{x_j}} } \right\|^2}\\ = tr({{\bf{A}}^T}{\bf{X}}{{\bf{M}}_c}{{\bf{X}}^{\bf{T}}}{\bf{A}}) \end{array} \end{array}$$ where $C$ is the number of classes, $\mathcal{D_S}^{(c)} = \{ {x_i}:{x_i} \in \mathcal{D_S} \wedge y({x_i} = c)\} $ represents the ${c^{th}}$ sub-domain in the source domain, $n_s^{(c)} = {\left\| {\mathcal{D_S}^{(c)}} \right\|_0}$ is the number of samples in the ${c^{th}}$ [source]{} sub-domain. $\mathcal{D_T}^{(c)}$ and $n_t^{(c)}$ are defined similarly for the target domain. Finally, $\bf M_c$ represents the conditional distribution between sub-domains in ${{\cal D}_{\cal S}}$ and ${{\cal D}_{\cal T}}$ and it is defined as: $$\label{eq:marginal} \begin{array}{*{20}{c}} {{{({{\bf{M}}_c})}_{ij}} = \left\{ {\begin{array}{*{20}{l}} {\frac{1}{{n_s^{(c)}n_s^{(c)}}},\;\;\;{x_i},{x_j} \in {D_{\cal S}}^{(c)}}\\ {\frac{1}{{n_t^{(c)}n_t^{(c)}}},\;\;\;{x_i},{x_j} \in {D_{\cal T}}^{(c)}}\\ {\frac{{ - 1}}{{n_s^{(c)}n_t^{(c)}}},\;\;\;\left\{ {\begin{array}{*{20}{l}} {{x_i} \in {D_{\cal S}}^{(c)},{x_j} \in {D_{\cal T}}^{(c)}}\\ {{x_i} \in {D_{\cal T}}^{(c)},{x_j} \in {D_{\cal S}}^{(c)}} \end{array}} \right.}\\ {0,\;\;\;\;\;\;\;\;\;\;\;\;otherwise} \end{array}} \right.} \end{array}$$ In minimizing ${Dis{t^{conditional}}\sum\limits_{c = 1}^C {({D_{\cal S}}^c,{D_{\cal T}}^c)} }$, the mismatch of conditional distributions between ${{D_{\cal S}}^c}$ and ${{D_{\cal T}}^c}$ is reduced. More discriminative:Repulsive Force Domain Adaptation ----------------------------------------------------- The latent feature subspace obtained by the joint marginal and conditional domain adaptation as in JDA, is to reduce the differences between the source and target domain. As such, two spaces of data are *attracted* to be close to each other. However, their model has ignored an important property for the elaboration of an effective predictor, *i.e.*, the preservation or reinforcement of discriminative information related to sub-domains. In this paper, we introduce a novel *repulsive force* domain adaption, which aims to increase the distances of sub-domains with different labels, so as to improve the discriminative power of the latent shared features and thereby making it possible better predictive model for the target domain. To sum up, we aim to generate a latent feature space where the discrepancy between domains is reduced while simultaneously the distances between sub-domains of different labels are increased for an reinforced discriminative power of the underlying latent feature space. Specifically, the repulsive force domain adaptation is defined as: $Dis{t^{repulsive}} = Dist_{{\cal S} \to {\cal T}}^{repulsive} + Dist_{{\cal T} \to {\cal S}}^{repulsive}$, where ${{\cal S} \to {\cal T}}$ and ${{\cal T} \to {\cal S}}$ index the distances computed from ${D_{\cal S}}$ to ${D_{\cal T}}$ and ${D_{\cal T}}$ to ${D_{\cal S}}$, respectively. $Dist_{{\cal S} \to {\cal T}}^{repulsive}$ represents the sum of the distances between each source sub-domain ${D_{\cal S}}^{(c)}$ and all the target sub-domains ${D_{\cal T}}^{(r);\;r \in \{ \{ 1...C\} - \{ c\} \} }$ except the one with the label $c$. The sum of these distances is explicitly defined as: $$\label{eq:StoT} \resizebox{1\hsize}{!}{ ${Dist}_{{\cal S} \to {\cal T}}^{repulsive} = \sum\limits_{c = 1}^C \begin{array}{l} {\left\| {\frac{1}{{n_s^{(c)}}}\sum\limits_{{x_i} \in {D_{\cal S}}^{(c)}} {{{\bf{A}}^T}{x_i}} - \frac{1}{{\sum\limits_{r \in \{ \{ 1...C\} - \{ c\} \} } {n_t^{(r)}} }}\sum\limits_{{x_j} \in D_{\cal T}^{(r)}} {{{\bf{A}}^T}{x_j}} } \right\|^2}\\ = \sum\limits_{c = 1}^C {tr({{\bf{A}}^T}{\bf{X}}{{\bf{M}}_{{\cal S} \to {\cal T}}}{{\bf{X}}^{\bf{T}}}{\bf{A}})} \end{array} $}$$ where ${{\bf{M}}_{{\cal S} \to {\cal T}}}$ is defined as $$\label{eq:mstot} \begin{array}{c} (\bf M_{{{\cal S} \to {\cal T}}})_{ij} = \left\{ {\begin{array}{*{20}{l}} {\frac{1}{{n_s^{(c)}n_s^{(c)}}},\;\;\;{x_i},{x_j} \in {D_{\cal S}}^{(c)}}\\ {\frac{1}{{n_t^{(r)}n_t^{(r)}}},\;\;\;{x_i},{x_j} \in {D_{\cal T}}^{(r)}}\\ {\frac{{ - 1}}{{n_s^{(c)}n_t^{(r)}}},\;\;\;\left\{ {\begin{array}{*{20}{l}} {{x_i} \in {\cal D_{\cal S}}^{(c)},{x_j} \in {D_{\cal T}}^{(r)}}\\ {{x_i} \in {\cal D_{\cal T}}^{(r)},{x_j} \in {\cal D_{\cal S}}^{(c)}} \end{array}} \right.}\\ {0,\;\;\;\;\;\;\;\;\;\;\;\;otherwise} \end{array}} \right. \end{array}$$ Symmetrically, $Dist_{{\cal T} \to {\cal S}}^{repulsive}$ represents the sum of the distances from each target sub-domain ${D_{\cal T}}^{(c)}$ to all the the source sub-domains ${D_{\cal S}}^{(r);\;r \in \{ \{ 1...C\} - \{ c\} \} }$ except the source sub-domain with the label $c$. Similarly, the sum of these distances is explicitly defined as: $$\label{eq:TtoS} \resizebox{1\hsize}{!}{ $Dist_{T \to S}^{repulsive} = \sum\limits_{c = 1}^C \begin{array}{l} {\left\| {\frac{1}{{n_s^{(c)}}}\sum\limits_{{x_i} \in {D_T}^{(c)}} {{{\bf{A}}^T}{x_i}} - \frac{1}{{\sum\limits_{r \in \{ \{ 1...C\} - \{ c\} \} } {n_t^{(r)}} }}\sum\limits_{{x_j} \in D_S^{(r)}} {{{\bf{A}}^T}{x_j}} } \right\|^2}\\ = \sum\limits_{c = 1}^C {tr({{\bf{A}}^T}{\bf{X}}{{\bf{M}}_{T \to S}}{{\bf{X}}^{\bf{T}}}{\bf{A}})} \end{array} $}$$ where ${{\bf{M}}_{{\cal T} \to {\cal S}}}$ is defined as $$\label{eq:mstot} \begin{array}{c} (\bf M_{{{\cal T} \to {\cal S}}})_{ij} = \left\{ {\begin{array}{*{20}{l}} {\frac{1}{{n_t^{(c)}n_t^{(c)}}},\;\;\;{x_i},{x_j} \in {D_{\cal T}}^{(c)}}\\ {\frac{1}{{n_s^{(r)}n_s^{(r)}}},\;\;\;{x_i},{x_j} \in {D_{\cal S}}^{(r)}}\\ {\frac{{ - 1}}{{n_t^{(c)}n_s^{(r)}}},\;\;\;\left\{ {\begin{array}{*{20}{l}} {{x_i} \in {\cal D_{\cal T}}^{(c)},{x_j} \in {D_{\cal S}}^{(r)}}\\ {{x_i} \in {\cal D_{\cal S}}^{(r)},{x_j} \in {\cal D_{\cal T}}^{(c)}} \end{array}} \right.}\\ {0,\;\;\;\;\;\;\;\;\;\;\;\;otherwise} \end{array}} \right. \end{array}$$ Finally, we obtain $$\label{eq:repulsive} \resizebox{0.9\hsize}{!}{ ${Dist}^{repulsive} = \sum\limits_{c = 1}^C {tr({{\bf{A}}^T}{\bf{X}}({{\bf{M}}_{S \to T}} + {{\bf{M}}_{T \to S}}){{\bf{X}}^{\bf{T}}}{\bf{A}})} $}$$ We define ${{\bf{M}}_{\hat c}} = {{\bf{M}}_{S \to T}} + {{\bf{M}}_{T \to S}}$ as the *repulsive force* constraint matrix.While the minimization of Eq.(\[eq:marginal\]) and Eq.(\[eq:conditional\]) makes closer both marginal and conditional distributions between source and target, the maximization of Eq.(\[eq:repulsive\]) increases the distances between source and target sub-domains with different labels, thereby improve the discriminative power of the underlying latent feature space. Label Deduction ---------------- In a number of domain adaptation methods, *e.g.*,[@pan2008transfer; @pan2011domain; @long2013transfer; @4967588], the simple Nearest Neighbor (NN) classifier is applied for label deduction. In JDA, NN-based label deduction is applied twice at each iteration. NN is first applied to the target domain in order to generate the *pseudo* labels of the target data and enable the computation of the conditional probability distance as defined in section 3.3. Once the optimized latent subspace NN identified, NN is then applied once again at the end of an iteration for the label prediction of the target domain. However, NN could not be a good classifier, given the fact that it is usually based on a $L2$ or $L1$ distance. It could fall short to measure the similarity of source and target domain data which may be embedded into a manifold with complex data structure. Furthermore, the cross-domain discrepancy still exists, even within a reduced latent feature space. To respect the underlying data manifold structure and better bridge the mismatch between the source and target domain distributions, we further propose in this paper two consistency constraints, namely *label smoothness consistency* and *geometric structure consistency* for both the *pseudo* and final label prediction. **Label Smoothness Consistency (LSC)** is defined as: $$\label{eq:labelconsitency} \begin{array}{c} Dis{t^{lable}} = \sum\limits_{j = 1}^C {\mathop \sum \limits_{i = 1}^{{n_s} + {n_t}} } \left\| {\bf{Y}}^{(T)}_{i,j}-{\bf{Y}}^{(0)}_{i,j} \right\| \end{array}$$ where ${\bf{Y}} = {{\bf{Y}}_{\cal S}} \cup {{\bf{Y}}_{\cal T}}$, ${\bf{Y}}_{i,j}^{(T)}$ is the probability of ${i_{th}}$ data belonging to ${j_{th}}$ class after ${T_{th}}$ iteration. ${\bf{Y}}_{i,j}^{(0)}$ is the initial prediction, and is defined as: $$\label{eq:labelconsistency1} \begin{array}{*{20}{l}} {{\bf{Y}}_{{{\cal S}_{(ij)}}}^{(0)} = \left\{ {\begin{array}{*{20}{l}} {y_{{{\cal S}_{(ij)}}}^{(0)} = 1\;(1 \le i \le {n_s}),j = c,{y_{ij}} \in D_{\cal S}^{(c)}}\\ {0\;\;\;\;\;\;\;\;\;\;\;else} \end{array}} \right.}\\ {{\bf{Y}}_{{{\cal T}_{(ij)}}}^{(0)} = \left\{ {\begin{array}{*{20}{l}} \begin{array}{l} y_{{{\cal T}_{(ij)}}}^{(0)} = 1\;(({n_s} + 1) \le i \le {n_s} + {n_t}),j = c,\\ {y_{ij}} \in D_{\cal T}^{(c)} \end{array}\\ {0\;\;\;\;\;\;\;\;\;\;\;else} \end{array}} \right.} \end{array}$$ **Geometric Structure Consistency (GSC)** is defined as: $$\label{eq:YLY} \begin{array}{c} \begin{array}{l} \begin{array}{l} {{\bf{Y}}^T}{\bf{L}}{\bf{Y}} = {{\bf{Y}}^T}({\bf{I}} - {{\bf{D}}^{ - \frac{1}{2}}}{\bf{W}}{{\bf{D}}^{ - \frac{1}{2}}}){\bf{Y}} = \\ \sum\limits_{i = 1}^{{n_s} + {n_t}} {{d_{ii}}{{\left( {\frac{{{y_i}}}{{\sqrt {{{\bf{d}}_{ii}}} }}} \right)}^2}} - \sum\limits_{i,j = 1}^{{n_s} + {n_t}} {{{\bf{d}}_{ii}}{{\left( {\frac{{{y_i}}}{{\sqrt {{{\bf{d}}_i}} }}\frac{{{y_j}}}{{\sqrt {{{\bf{d}}_j}} }}} \right)}^2}} {{\bf{w}}_{ij}}\;\\ = \frac{1}{2}\sum\limits_{i,j = 1}^{{n_s} + {n_t}} {{{\bf{w}}_{ij}}{{\left( {\frac{{{y_i}}}{{\sqrt {{{\bf{d}}_{ii}}} }} - \frac{{{y_j}}}{{\sqrt {{{\bf{d}}_{jj}}} }}} \right)}^2}} \end{array} \end{array}, \end{array}$$ where ${\bf{W}} = {[{w_{ij}}]_{({n_s} + {n_t}) \times ({n_s} + {n_t})}}$ is an affinity matrix [@NIPS2001_2092], with ${w_{ij}}$ giving the affinity between two samples $i$ and $j$ and defined as ${w_{ij}} = \exp ( - \frac{{{{\left\| {{x_i} - {x_j}} \right\|}^2}}}{{2{\sigma ^2}}})$ if $i \ne j$ and ${w_{ii}} = 0$ otherwise, ${\bf{D}} = diag\{ {d_{11}}...{d_{({n_s} + {n_t}),({n_s} + {n_t})}}\} $ is the degree matrix with ${d_{ii}} = \sum\nolimits_j {{w_{ij}}} $. When Eq.(\[eq:YLY\]) is minimized, the geometric structure consistency ensures that the label space does not change too much between nearby data. Learning Algorithm ------------------ Our proposed domain adaptation integrates the marginal and conditional distribution and repulsive force, as well as the final label prediction using both label smoothness and geometric structure consistencies. Our model is defined as: $$\label{eq:ours_physical} \resizebox{0.96\hsize}{!}{ $\begin{array}{*{20}{l}} {\min (Dis{t^{marginal}} + Dis{t^{conditional}} + Dis{t^{label}} + {{\bf{Y}}^T}L{\bf{Y}})}\\ + \max (Dist^{repulsive}) \end{array}$}$$ It can be re-written mathematically as: $$\label{eq:ours_math} \resizebox{1\hsize}{!}{$\begin{array}{*{20}{l}} {\begin{array}{*{20}{l}} {\mathop {\min }\limits_{{{\bf{A}}^T}{\bf{XH}}{{\bf{X}}^T}{\bf{A}} = {\bf{I}}} \left( {\begin{array}{*{20}{l}} {\sum\limits_{c = 0}^C {tr({{\bf{A}}^T}{\bf{X}}{{\bf{M}}_c}{{\bf{X}}^T}A)} + \lambda \left\| {\bf{A}} \right\|_F^2}\\ { + \sum\limits_{j = 1}^C {\sum\limits_{i = 1}^{{n_s} + {n_t}} {\left\| {{\bf{Y}}_{ij}^{(T)} - {\bf{Y}}_{ij}^{(0)}} \right\|} } + {{\bf{Y}}^T}{\bf{LY}}} \end{array}} \right)}\\ { + \mathop {\max }\limits_{{{\bf{A}}^T}{\bf{XH}}{{\bf{X}}^T}{\bf{A}} = {\bf{I}}} tr({{\bf{A}}^T}{\bf{X}}{{\bf{M}}_{{\bf{\hat c}}}}{{\bf{X}}^T}{\bf{A}})} \end{array}} \end{array}$}$$ Direct solution to this problem is nontrivial. We divide it into two sub-problems: (1) $ \mathop {\min }\limits_{{{\bf{A}}^T}{\bf{XH}}{{\bf{X}}^T}{\bf{A}} = {\bf{I}}} \left( {\sum\limits_{c = 0}^C {tr({{\bf{A}}^T}{\bf{X}}{{\bf{M}}_{cyd}}{{\bf{X}}^T}A)} + \lambda \left\| {\bf{A}} \right\|_F^2} \right) $, where ${{\bf{M}}_{cyd}} = \sum\limits_{c = 0}^C {{{\bf{M}}_c} - {{\bf{M}}_{{\bf{\hat c}}}}} $ and (2) $\mathop {\min }\limits_{{{\bf{A}}^T}{\bf{XH}}{{\bf{X}}^T}{\bf{A}} = {\bf{I}}} \left( {\sum\limits_{j = 1}^C {\sum\limits_{i = 1}^{{n_s} + {n_t}} {\left\| {{\bf{Y}}_{ij}^{(T)} - {\bf{Y}}_{ij}^{(0)}} \right\|} } + {{\bf{Y}}^T}{\bf{L}}{\bf{Y}}} \right)$. These two sub-problems are then iteratively optimized. The first sub-problem, as explained in JDA, amounts to solving the generalized eigendecomposition problem,*i.e.*, $({\bf{X}}{\bf{M_{cyd}}}{{\bf{X}}^T} + \lambda {\bf{I}}){\bf{A}} = {\bf{XH}}{{\bf{X}}^T}{\bf{A}}\Phi $. Then, we obtain the adaptation matrix $\boldsymbol{A}$ and the underlying embedding space $\boldsymbol{Z}$. The second sub-problem is nontrivial. Inspired by the solution proposed in [@Zhou04learningwith] [@6341755] [@6619251], the minimum is approached where the derivative of the function is zero. An approximate solution can be provided by: $$\label{eq:Y_optimal} {{\bf{Y}}^ \star } = {({\bf{D}} - \alpha {\bf{W}})^{ - 1}}{Y^{(0)}}$$ where $Y^\star$ is the probability of prediction of the target domain corresponding to different class labels. The complete learning algorithm is summarized in Algorithm 1. Experiments =========== In this section, we validate the effectiveness of our proposed domain adaptation model, *i.e.*, CDDA, on several datasets for cross-domain image classification task. Benchmarks ---------- In domain adaptation, USPS+MINIST, COIL20, PIE and office+Caltech are standard benchmarks for the purpose of evaluation and comparison with state of the art. In this paper, we follow the data preparation as most previous works. We construct 36 datasets for different image classification tasks. They are: (1) the **USPS** and **MINIST** datasets of digits, but with different distribution probabilities. We built the cross-domains as: *USPS vs MNIST* and *MNIST vs USPS*; (2) the **COIL20** dataset with 20 classes, split into *COIL1 vs COIL2* and *COIL2 vs COIL1*; (3) the **PIE** face database with different face [poses]{}, of which five subsets are selected, denoted as PIE1, PIE2, *etc*., resulting in $5 \times 4=20$ domain adaptation tasks, *i.e.*, *PIE1 vs PIE 2* $\dots$ *PIE5 vs PIE 4*; (4) **Office** and **Caltech-256**. Office contains three real-world datasets: **Amazon**(images downloaded from online merchants), **Webcam**(low resolution images) and **DSLR**( high-resolution images by digital web camera). **Caltech-256** is standard dataset for object recognition, which contains 30,607 images for 31 categories. We denote the dataset **Amazon**,**Webcam**,**DSLR**,and **Caltech-256** as **A**,**W**,**D**,and **C**, respectively. $4\times 3=12$ domain adaptation tasks can then be constructed, namely *A* $\rightarrow$ *W* $\dots$ *C* $\rightarrow$ *D*, respectively. Baseline Methods ---------------- The proposed CDDA method is compared with six methods of the literature, excluding only CNN-based works, given the fact that we are not using deep features. They are: (1)1-Nearest Neighbor Classifier(NN); (2) Principal Component Analysis (PCA) +NN; (3) Geodesic Flow Kernel(GFK) [@gong2012geodesic] + NN; (4) Transfer Component Analysis(TCA) [@pan2011domain] +NN; (5)Transfer Subspace Learning(TSL) [@4967588] +NN; (6) Joint Domain Adaptation (JDA) [@long2013transfer] +NN. Note that TCA and TSL can be viewed as special case of JDA with $C=0$, and JDA a special case of the proposed CDDA method when the *repulsive force* domain adaptation is ignored and the label generation is simply based on NN instead of the label propagation with label smoothness and geometric structure consistency constraints. All the reported performance scores of the six methods of the literature are directly collected from the authors’ publication. They are assumed to be their *best* performance. Experimental Setup ------------------ For the problem of domain adaptation, it is not possible to tune a set of optimal hyper-parameters, given the fact that the target domain has no labeled data. Following the setting of JDA, we also evaluate the proposed CDDA by empirically searching the parameter space for the *optimal* settings. Specifically, the proposed CDDA method has three hyper-parameters, *i.e.*, the subspace dimension $k$, regularization parameters $\lambda $ and $\alpha $. In our experiments, we set $k = 100$ and 1) $\lambda = 0.1$, and $\alpha = 0.99$ for **USPS**, **MNIST** and **COIL20** , 2) $\lambda = 0.1$, $\alpha = 0.2$ for **PIE**, 3) $\lambda = 1$, $\alpha = 0.99$ for **Office** and **Caltech-256**. In our experiment, [*accuracy*]{} on the test dataset is the evaluation measurement. It is widely used in literature, *e.g.*,[@pan2008transfer; @long2013transfer; @long2015learning], *etc*. $$\label{eq:accuracy} \begin{array}{c} Accuracy = \frac{{\left| {x:x \in {D_T} \wedge \hat y(x) = y(x)} \right|}}{{\left| {x:x \in {D_T}} \right|}} \end{array}$$ where ${\cal{D_T}}$ is the target domain treated as test data, ${\hat{y}(x)}$ is the predicted label and ${y(x)}$ is the ground truth label for a test data $x$. Experimental Results and Discussion ----------------------------------- The classification accuracies of the proposed CDDA method and the six baseline methods are shown in Table.1. and illustrated in Fig.1. for the clarity of comparison. \[fig:acc\] {width="1\linewidth"} \[tab:acc\] \[fig:analysis\] -------------------------------------- {width="1\linewidth"} -------------------------------------- In Table.1, the highest accuracy for each cross-domain adaptation task is highlighted in bold. For a better understanding of the proposed CDDA, we evaluate the proposed CDDA method using two settings: (1) **CDDA(a)** where simple NN is used as label predictor instead of the proposed label propagation; and (2) **CDDA(b)** where the proposed label propagation is activated for the prediction of target data labels. As CDDA is reduced to JDA when repulsive force domain adaptation and label propagation are not integrated, the setting CDDA(a) enables to quantify the contribution of adding the repulsive force domain adaptation w.r.t. JDA whereas the setting CDDA(b) makes it possible to evidence the contribution of the proposed label propagation in comparison with CDDA(a) and highlight the overall behavior of the proposed method. As can be seen in Table.1 , the proposed CDDA depicts an overall average accuracy of $\bf 60.12\%$ and $\bf 62.92\%$, respectively, with respect to the above two settings. They both outperform the six baseline algorithms with a large margin. With the repulsive force integrated and NN as label predictor, CDDA(a) outperforms JDA on 30 cross-domain tasks out of 36 and improves JDA’s overall average accuracy by roughly 3 points, thereby demonstrating the effectiveness of the proposed repulsive force domain adaptation. Now, in adopting the proposed label propagation under the constraint of both label smoothness and geometric structure consistency, CDDA(b) further improves CDDA(a) by roughly 2 points in terms of overall average accuracy and outperforms JDA by more than 4 points. Compared with the baseline methods, the proposed CDDA method consistently shows its superiority and depicts the best average accuracy over all the four datasets (USPS+MINIST, COIL20, PIE, Amazon). As can be seen in Fig.2, **CDDA(b)** as represented by the red curve is on the top of the other curves along the axis of 36 cross-domain image classification tasks. It is worth noting that the proposed CDDA depicts $\bf 99.65$ accuracy on **COIL20**; This is rather an unexpected impressive score given the unsupervised nature of the domain adaptation for the target domain. Using *COIL2 vs COIL1*, and *C* $\rightarrow$ *W* datasets, we also empirically check the convergence and the sensitivity of the proposed CDDA with respect to the hyper-parameters. Similar trends can be observed on all the other datasets. The accuracy w.r.t. $\#$iterations is shown in Fig.3 (a). As can be seen there, the performance of the proposed CDDA along with JDA becomes stable after about 10 iterations. In the experiment, CDDA have two settings: two parameters ($k$ and $\lambda$) in **CDDA(a)** and three ($k$, $\lambda$ and $\alpha$) in **CDDA(b)**. The accuracy variation w.r.t regularization parameter $\alpha$ is shown in Fig.3 (b), which indicates **CDDA(b)** achieves the best performance when $\alpha$ is close to 0.99 in COIL20 and the performance is more or less stable when $\alpha$ is less than 0.99. Given a novel dataset, we tune the parameter $\alpha$ in the range \[0.001,1\]. For instance, in the **PIE** database, we set the optimal $\alpha$ to 0.2. The other parameters, *i.e.*, $k$ and $\lambda$, also converge. Their behavior is not shown here due to space limitation. Conclusion and Future Work ========================== In this paper, we have proposed a Close yet Discriminative Domain Adaptation (CDDA) method based on feature representation. Comprehensive experiments on 36 cross-domain datasets highlight the interest of reinforcing the data discriminative properties within the model and label propagation in respect of the geometric structure of the underlying data manifold, and verify the effectiveness of proposed method compared with six baseline methods of the literature. Our future work will concentrate on embedding the proposed method in deep networks and study other vision tasks, *e.g.*, object detection, within the setting of transfer learning. [^1]: These first two authors contributed equally.

ArXiv

--- author: - Qiming Sun - 'Garnet Kin-Lic Chan' bibliography: - 'embed.bib' title: Quantum embedding theories --- Conspectus ========== In complex systems, it is often the case that the region of interest forms only one part of a much larger system. The idea of joining two different quantum simulations - a high level calculation on the active region of interest, and a low level calculation on its environment - formally defines a [*quantum embedding*]{}. While any combination of techniques constitutes an embedding, several rigorous formalisms have emerged that provide for exact feedback between the embedded system and its environment. These three formulations: [*density functional embedding*]{}, [*Green’s function embedding*]{}, and [*density matrix embedding*]{}, respectively use the single-particle density, single-particle Green’s function, and single-particle density matrix as the quantum variables of interest. Many excellent reviews exist covering these methods individually. However, a unified presentation of the different formalisms is so far lacking. Indeed, the various languages commonly used: functional equations for density functional embedding; diagrammatics for Green’s function embedding; and entanglement arguments for density matrix embedding, make the three formulations appear vastly different. In this account, we introduce the basic equations of all three formulations in such a way as to highlight their many common intellectual strands. While we focus primarily on a straightforward theoretical perspective, we also give a brief overview of recent applications, and possible future developments. The first section starts with density functional embedding, where we introduce the key embedding potential via the Euler equation. We then discuss recent work concerning the treatment of the non-additive kinetic potential, before describing mean-field density functional embedding, and wavefunction in density functional embedding. We finish the section with extensions to time-dependence and excited states. The second section is devoted to Green’s function embedding. Here, we use the Dyson equation to obtain equations that parallel as closely as possible the density functional embedding equations, with the hybridization playing the role of the embedding potential. Embedding a high-level self-energy within a low-level self-energy is treated analogously to wavefunction in density functional embedding. The numerical computation of the high-level self energy allows us to briefly introduce the bath representation in the quantum impurity problem. We then consider translationally invariant systems to bring in the important dynamical mean-field theory. Recent developments to incorporate screening and long-range interactions are discussed. The third section concerns density matrix embedding. Here, we first highlight some mathematical complications associated with a simple Euler equation derivation, arising from the open nature of fragments. This motivates the density matrix embedding theory, where we use the Schmidt decomposition to represent the entanglement through bath orbitals. The resulting impurity plus bath formulation resembles that of dynamical mean-field theory. We discuss the numerical self-consistency associated with using a high-level correlated wavefunction with a mean-field low-level treatment, and connect the resulting numerical inversion to that used in density functional embedding. We finish with perspectives on the future of all three methods. Introduction ============ Embedding theories provide a natural way to focus computation on a small region within a larger environment, such as atoms near the active site in an enzyme. In any embedding theory, the full problem is partitioned into the fragment of interest, ($A$), and its environment ($B$). (The terms “subsystem” or “impurity” are commonly used in place of “fragment”). There may be multiple fragments of interest, and the environment may also be partitioned. In any case, the purpose of the embedding is to provide a computational recipe for the properties of $A$, taking into account its environment, without the computational cost of treating the full problem. The idea is thus very general, and encompasses a wide variety of methods. For example, whenever $A$ and $B$ are treated at two different levels, e.g. when freezing orbitals in a calculation, we are formally performing embedding! What concretely distinguishes one embedding theory from another, is the particular way in which the effects of the environment are communicated to the fragment $A$, and vice-versa. Here, we will describe a family of three related, rigorous, quantum embedding theories: density functional (DFT) embedding[@Cortona1991; @Wesolowski1993; @Jacob2014; @libisch2014embedded; @wesolowski2015frozen], Green’s function embedding[@inglesfield1981method; @Inglesfield2015; @georges1992hubbard; @Kotliar2006; @Chan2011a; @Chibani2016] and density matrix embedding[@Knizia2012; @Knizia2013; @Wouters2016; @Pernal2016]. Compared to simpler embedding techniques, these stand out as they provide for non-trivial communication between the fragment and environment. In the three theories, information is communicated via the density, the single-particle Green’s function, and the single-particle density matrix, respectively. Note that we identify the kind of embedding by its functional dependence on a quantum variable, rather than intermediary computational objects. Thus, even if a density matrix is used in the calculation, if it ultimately encodes a density functional (e.g. in a Kohn-Sham theory), then the embedding is a density functional embedding. There are excellent reviews for each of the three techniques individually, for example, Refs. [@libisch2014embedded; @Jacob2014; @wesolowski2015frozen; @Inglesfield2015; @Kotliar2006; @Wouters2016]. Here, we will give an introduction to all three together, that emphasizes their common intellectual strands, and provide a summary of their strengths and weaknesses. DFT embedding ============= We will begin with the simplest quantum embedding, DFT embedding. We provide a particular perspective to make connections to the subsequent Green’s function and density matrix embedding formalisms; other perspectives can be found in the literature [@libisch2014embedded; @Jacob2014; @wesolowski2015frozen]. The formal groundwork was developed by Cortona [@Cortona1991] and Wesolowski and Warshel[@Wesolowski1993]. In the ground-state version, the density of $A$, $\rho_A$, is adjusted through an external field, $v_A$, and we view this field as coming from the environment. Further, the energy of $A$ is modified by its interactions with the environment, through the Coulomb term, and indirectly through the exclusion principle. These energetic contributions are contained in a term $\Delta E$. The DFT embedding formalism provides rigorous foundations for $v_A$ and $\Delta E$ as density functionals. We begin with the exact energy density functional for the full system, $E[\rho]$, which determines the ground-state density through the Euler equation $$\begin{aligned} \frac{\delta E[\rho]}{\delta \rho} - \mu = 0,\end{aligned}$$ where $\mu$ is the chemical potential. Partitioning the energy into its Kohn-Sham kinetic, Coulomb, external, and exchange-correlation pieces [@parryang], $$\begin{aligned} E[\rho] = T_s[\rho] + J[\rho] + V_{ext}[\rho] + E_{xc}[\rho],\end{aligned}$$ we rewrite the Euler equation $$\begin{aligned} \frac{\delta T_s[\rho]}{\delta \rho} + v_J[\rho] + v_{ext} + v_{xc}[\rho] - \mu = 0. \label{eq:euler_pieces}\end{aligned}$$ $E_{xc}[\rho]$ and $v_{xc}[\rho]$ account for the non-trivial correlation effects, and must be approximated. When necessary, we indicate the dependence of these quantities on the approximation scheme $S$ by writing them as $E_{xc}^S[\rho]$, $v_{xc}^S[\rho]$. We now derive the expressions for $v_A$ and $\Delta E$. Splitting $E[\rho]$ into the fragment energy and its remainder, $E[\rho]=E[\rho_A]+\Delta E[\rho,\rho_A]$ ($E_A = E[\rho_A]$), stationarity with respect to $\rho_A$ implies $$\begin{aligned} \frac{\delta}{\delta \rho_A} \Bigl[E[\rho_A] + \Delta E[\rho,\rho_A]\Bigl] - \mu = 0. \label{eq:dft_stationary}\end{aligned}$$ This is the same as the Euler equation for $A$ placed in an external field $v_A$, if we choose $$\begin{aligned} v_A= \frac{\delta \Delta E[\rho,\rho_A]}{\delta \rho_A} \label{eq:exact_dft_v}.\end{aligned}$$ Eq. (5) thus defines the exact embedding potential, which yields the exact subsystem density $\rho_A$. Note we use $\rho$, $\rho_A$ as the working variables. The use of a global and a fragment quantum variable is a choice common to all three embedding formalisms in this article. However, in DFT embedding, one often works instead with the subsystem and environment densities separately, $\rho_A, \rho_B$, with $\rho = \rho_A + \rho_B$. This has advantages in practice, for example, if $\rho_A$ and $\rho_B$ are ensemble $N$-representable, then so is $\rho$ [@wesolowski2008embedding]. We express $v_A$ in terms of its components as $$\begin{aligned} v_A &= \frac{\delta}{\delta \rho_A}\Bigl[ T_s[\rho]-T_s[\rho_A]\Bigl] + v_J[\rho-\rho_A] + \frac{\delta}{\delta \rho_A}\Bigl[E_{xc}[\rho]-E_{xc}[\rho_A]\Bigl]\notag\\ &= v_{s}^\Delta + v_J^\Delta + v_{xc}^\Delta \label{eq:dft_partition}\end{aligned}$$ where $v_{s}^\Delta$ is the [*non-additive kinetic potential*]{}, $v_J^\Delta$ is the environment Coulomb potential, and $v_{xc}^\Delta$ is the non-additive exchange-correlation potential. $v_{s}^\Delta$ is the largest contribution and expresses the exclusion principle which forces electrons in the fragment to occupy states orthogonal to those in the environment [@Cortona1991; @Wesolowski1993]. Commonly in DFT calculations, $E_{xc}[\rho]$ is approximated as an explicit functional of $\rho$ and its derivatives, (e.g. as is in the LDA or GGA’s). Consequently, $v_{xc}^{\Delta}[\rho]$ can be obtained by analytical functional differentiation of $E_{xc}[\rho] - E_{xc}[\rho_A]$. $T_s[\rho]$, however, is only known as an [*implicit*]{} density functional through the Kohn-Sham orbitals. One way to evaluate $v_{s}^\Delta$ is to approximate $T_s$ by an explicit density functional approximation, such as the Thomas-Fermi (or a related) functional, from which the kinetic potential can be directly derived. This approximation was widely used in the early days of DFT embedding [@Cortona1991; @Wesolowski1993], but is limited by the accuracy of the approximate kinetic energy [@Wesolowski1996; @Visscher2009]. It has found most success in applications where the fragment is weakly bound to the environment, for example in van der Waals complexes [@Wesolowski2007], highly ionic crystals [@Cortona1992], and in solvation [@Neugebauer2005; @Wesolowski2010]. A more recent strategy has been to compute $\delta T_s/\delta \rho_A$ and the non-additive kinetic potential numerically[@roncero2008inversion; @fux2010accurate; @Goodpaster2010; @Chen2011]. Since the forward computation of $v_s \to \rho$ (to determine the non-interacting ground-state density in an external potential $v_s$) is cheap ($O(N^3)$ where $N$ is the number of electrons) the inversion is tractable [@Zhao1992; @Wu2003]. In practice, there are numerical difficulties [@Scuseria2006], because large changes in $v_s$ can yield only small changes in $\rho$. This leads to unphysical oscillations in the numerically determined $v_s$ [@Scuseria2006; @Yang2007]. The above is a basic formulation of ground-state DFT embedding, and extensions exist, for example, to ensemble representations for fragments with non-integer particle number [@pernal2009orbital; @elliott2009density; @pfet], and Kohn-Sham projector formalisms to avoid the non-additive kinetic potential altogether [@Manby2012; @tamukong2014density]. However, we now consider the two common contexts in which DFT embedding is applied. The first is DFT in DFT embedding [@Wesolowski1993], where both the subsystem $A$ and environment $B$ are treated with DFT. This may not seem to result in any simplification, but there are several ways to obtain savings. For example, different levels of DFT can be used for $A$ and $B$ (e.g. functionals with or without exact exchange) [@fornace2015embedded] or the self-consistency may be approximated (e.g. by replacing the self-consistent $\rho_B$ by a frozen superposition of densities). The latter is particularly appropriate for solvent systems [@Neugebauer2005; @Wesolowski2010], and allows for very large environments, such as protein frameworks, to be considered. The second common context is where the fragment is described using a high-level wavefunction, and the environment by DFT, and embedding connects the two [@Carter1998; @wesolowski2008embedding; @Hofener2012; @Goodpaster2012; @Visscher2013b]. Wavefunction in DFT embedding was pioneered by Carter and coworkers [@Carter1998; @Chen2011] and has attracted much recent attention, because of the difficulties of DFT in treating aspects of electronic structure such as excited states and bond-breaking. In wavefunction in DFT embedding, two approximations enter for the correlations, $S_{WF}$ and $S_{DFT}$. The energy functional is defined as $$\begin{aligned} E[\rho] = E^{S_{WF}}[\rho_A] + E^{S_{DFT}}[\rho] - E^{ S_{DFT}}[\rho_A] \label{eq:totalenergyE} \end{aligned}$$ where the wavefunction energy for fragment $A$ is formally expressed through $$\begin{aligned} E^{S_{WF}}[\rho_A] = \min_{\Psi^{S_{WF}} \to \rho_A} \langle \Psi^{S_{WF}}|\hat{H}_A|\Psi^{S_{WF}}\rangle. \label{eq:constrained_dft_wf}\end{aligned}$$ Since the effect of the environment is completely contained within $v_A$, [the minimizing $\Psi$ in Eq. (8) is the ground-state of $\hat{H}_A + v_A$]{}. Any wavefunction ansatz may be used to approximate this eigenstate [@Carter1998; @Chen2011; @Hofener2012; @Goodpaster2012; @Visscher2013b], and existing quantum chemistry programs only need be modified to include the potential $v_A$, which adds to the attractiveness of the method. In the simplest scheme, $\rho_A$ and $v_A$ are defined both from the DFT density, and the DFT expression, $\Delta E^{S_{DFT}}[\rho, \rho_A]$ in Eq. (5). In a more sophisticated scheme, the wavefunction treatment of correlation is used to improve the density self-consistently. This is achieved by defining the exchange-correlation potential of the full system $$\begin{aligned} v_{xc}[\rho] = v_{xc}^{S_{WF}}[\rho_A] + v_{xc}^{S_{DFT}}[\rho] - v_{xc}^{S_{DFT}}[\rho_A ]. \label{eq:vxc_full}\end{aligned}$$ Self-consistent iteration between $v_{xc}$ and $v_A$ then allows for the DFT density $\rho_A^{DFT}$ and the WF density $\rho_A^{WF}$ to become identical. There is an increasing number of applications using wavefunction in DFT embedding, such as to molecules adsorbed on surfaces [@Carter2001; @Carter2012a] and molecular fragments embedded in larger systems [@goodpaster2014accurate]. A growing community is exploring these techniques also for excited state properties (see e.g. Ref. [@SeveroPereiraGomes2012]). The simplest way to compute excited states is to assume that the ground-state and excited-state embedding potential and energy functional are identical[@Carter2002]. This is, however, an approximation, and in principle, the ground-state formalism must be extended. There are two ways to do so. The first is a state-specific DFT embedding, where $E[\rho]$ and $v_A$ acquire an excited state dependence. State-specific DFT embedding has been explored by several workers [@Neugebauer2013a; @prager2016first], who find that state-dependence gives significant corrections, especially when the charge character of the excited state differs from the ground-state. A second way is through time-dependent DFT, where excitation energies are poles of the response. Here, the embedding potential becomes time-dependent, with a non-local dependence on the density at earlier times, $v_A[\rho(t')]$ [@casida2004generalization; @Neugebauer2007; @Neugebauer2009a; @Neugebauer2013; @pavanello2013subsystem]. Applications using this second formalism are now beginning to appear [@chulhai2016external]. Despite its conceptual simplicity, wavefunction in DFT embedding inherits limitations intrinsic to all combinations of wavefunctions with density functional approximations. For example, in the total energy expression, Eq. (7), the non-additive part of the energy is described at the DFT level, while an accurate total energy requires error cancellation between the wavefunction and DFT descriptions of the system ($E^{S_{WF}}[\rho_A] - E^{S_{DFT}}[\rho_A]$). Incomplete cancellation is sometimes referred to as “double counting”. The two sources of error are important if the interface between the fragment and environment cuts across a bond of interest, as they affect the correlation energy of the bond. Similarly, if van der Waals’ interactions between the fragment and environment are important, such contributions, omitted in many density functionals, will be missed. Both these situations can be remedied formally by increasing the system size treated with the wavefunction method, albeit at increased cost. Another disadvantage of DFT embedding is that it is difficult, through examining $\rho_A$ alone, to distinguish between a fragment which is bonded with the environment, and one which is not. This is because the density, by definition, does not contain direct information on off-diagonal density matrix correlations (i.e. coherence and entanglement). While, in principle, all effects can be captured by the exact density functional, the lack of the off-diagonal information can pose difficulties for density functional approximations used in practice. One way to surmount this is to consider embedding theories of richer quantum variables with off-diagonal correlations, such as the single-particle Green’s function or the single-particle density matrix. We now turn to these embedding formalisms. Green’s function embedding ========================== Green’s function embedding has a long history, and is widely used in condensed matter problems [@grimley1974chemisorption; @inglesfield1981method; @georges1992hubbard; @Kotliar2006; @Inglesfield2015; @Chibani2016]. We cannot cover the extensive literature on models, and will restrict ourselves to the aspects of Green’s function embedding relevant to ab-initio quantum chemistry [@Chan2011a; @Lin2011; @Chibani2016; @Turkowski2012; @Lan2015]. The zero-temperature, time-ordered, single-particle Green’s function generalizes the familiar single-particle density matrix, to carry additional information on time-dependent correlations. In a basis labelled by $i,j$, it is defined as $$\begin{aligned} i G_{ij}(t) = \mathcal{T} \langle \Psi_0 | a_i (0) a^\dag_j(t) | \Psi_0\rangle\end{aligned}$$ where $\mathcal{T}$ denotes time-ordering, $\Psi_0$ the ground-state wavefunction, and $a^{(\dag)}$ are electron annihilation (creation) operators in the Heisenberg representation. It is common also to use $\mathbf{G}(\omega)$, the Fourier transform of $\mathbf{G}(t)$. The imaginary part of the Green’s function $-\frac{1}{\pi} \mathrm{tr} \ \mathrm{Im} \ \mathbf{G}(\omega)$ is the single-particle density of states, while the equal-time Green’s function $-i \mathbf{G}(t=0_+)$ (where $0_+$ is a positive infinitesimal) is the single-particle density matrix. As an example, the real-space non-interacting Green’s function is given by $\mathbf{g}(\mathbf{r}, \mathbf{r'}, \omega) = \sum_i \phi_i^*(\mathbf{r}) \phi_i (\mathbf{r'}) (\omega - \epsilon_i + i0_+)^{-1}$, where $\phi_i, \epsilon_i$ are the single-particle orbitals and energies. Unlike in DFT, the energy is computable explicitly from the exact Green’s function as $$\begin{aligned} E=\frac{1}{2} \int_{-\infty}^{\mu} d\omega \ \mathrm{tr} \ \Bigl[ (\mathbf{h} + \omega \mathbf{1}) \ \mathrm{Im} \ \mathbf{G}(\omega)\Bigl] \label{eq:midgal}\end{aligned}$$ ($\mu$ is the chemical potential, $\mathbf{h}$ is the single-particle Hamiltonian). For approximate Green’s functions, the expression must be modified with additional terms to obtain a variational bound [@almbladh1999variational]. One example is the Luttinger-Ward functional, from which an embedding formalism can be constructed[@potthoff2003self]. However, it is complicated to specify and we do not need all its properties here. Instead, it is sufficient to consider the corresponding Euler equation, namely the Dyson equation. The Dyson equation relates the Green’s function of different Hamiltonians. For example, the Green’s function of a non-interacting system $\mathbf{g}(\omega)$ and that of the interacting system $\mathbf{G}(\omega)$ are related by $$\begin{aligned} \mathbf{G}(\omega) = \mathbf{g}(\omega) + \mathbf{G}(\omega) \boldsymbol{\Sigma}(\omega) \mathbf{g}(\omega) \label{eq:dyson}\end{aligned}$$ where $\boldsymbol{\Sigma}(\omega)$, the self-energy, accounts for interactions. Eq. (12) is the analog of the DFT Euler equation Eq. (3) where $\mathbf{g}(\omega)$ plays a similar role to $\delta T_s/\delta \rho + v_{ext}$, and the self-energy plays the part of the Coulomb plus exchange-correlation potential. In practice, the self-energy must be approximated, and it is convenient to discuss such approximations in diagrammatic terms. The exact (proper) $\boldsymbol{\Sigma}(\omega)$ is a sum over all perturbation theory diagrams, where each diagram is a connected graph of Green’s function and interaction lines, and no diagram can be cut in two by cutting a single Green’s function line. An approximate self-energy will sum over a subset of these diagrams, which can be specified in terms of the diagram skeletons ($S$), and the Green’s functions $\mathbf{G}_\Sigma(\omega)$ and interactions $V$ within them. We thus denote a self-energy approximation by $\boldsymbol{\Sigma}^S[\mathbf{G}_\Sigma, V]$. A [*self-consistent*]{} approximate self-energy is one where the Green’s function used in the self-energy diagrams satisfies the Dyson equation, i.e. $\mathbf{G} = \mathbf{G}_\Sigma$. In Green’s function embedding, the fragment Green’s function $\mathbf{G}_A (\omega)$, is adjusted through another self-energy $\boldsymbol{\Delta}_A(\omega)$, dual to $\mathbf{G}_A(\omega)$. To prevent confusion with the self-energy arising due to interactions in Eq. (12), we term $\boldsymbol{\Delta}_A(\omega)$ the hybridization. Together with the self-energy approximation $\boldsymbol{\Sigma}_A(\omega)$, this gives the fragment Dyson equation $$\begin{aligned} \mathbf{G}_A(\omega) = \mathbf{g}_A(\omega) + \mathbf{G}_A(\omega) (\boldsymbol{\Delta}_A(\omega) + \boldsymbol{\Sigma}_A(\omega)) \mathbf{g}_A(\omega), \label{eq:embed_dyson}\end{aligned}$$ where $\mathbf{{g}}_A(\omega)$ is the non-interacting Green’s function of the system computed [*in isolation*]{}, i.e. with no couplings to the environment or any interactions. [Inverting Eq. (13) gives the hybridization as $$\begin{aligned} \boldsymbol{\Delta}_A(\omega) = \mathbf{g}_A(\omega)^{-1} - \mathbf{G}_A(\omega)^{-1} - \boldsymbol{\Sigma}_A(\omega) \label{eq:hyb}. \end{aligned}$$ ]{} $\boldsymbol{\Delta}_A(\omega)$ is the analog of the DFT embedding potential $v_A$, and Eq. (14) is analogous to Eq. (6). Similarly to $v_A$, $\boldsymbol{\Delta}_A(\omega)$ contains effects from electron delocalization into the environment as well as environment Coulomb interactions. However, unlike in DFT embedding, the inversion from $\mathbf{G}(\omega) \to \boldsymbol{\Delta}(\omega)$ is explicit through Eq. (14), and no iterative technique is necessary. We now discuss two different contexts in which Green’s function embedding is applied. These correspond to different self-energy approximations for $A$ and the environment $B$, and parallel the contexts appearing in DFT embedding. The simplest is to describe both $A$ and the full system at the mean-field level. Then, $\boldsymbol{\Sigma}_A(\omega)$ and $\boldsymbol{\Sigma}(\omega)$ correspond to the respective mean-field Coulomb and exchange terms (and thus have no frequency dependence). Similarly to DFT in DFT embedding, savings result when we use different levels of mean-field for $A$ and the full-system, for example by forgoing self-consistency in the environment. A common application of the latter is to impurities in crystals, where $\mathbf{G}(\omega)$ is first computed using translational invariance in the periodic crystal, and relaxation of the environment is ignored when the impurity is introduced [@grimley1974chemisorption]. In molecular junctions, $\boldsymbol{\Delta}_A(\omega)$ is similarly obtained for a semi-infinite electrode, and assumed to be unchanged on the introduction of the bridging molecule [@brandbyge2002density]. A second context is to perform Green’s function embedding with more sophisticated approximations for $\boldsymbol{\Sigma}_A(\omega)$ and $\boldsymbol{\Sigma}(\omega)$. This is similar to wavefunction in DFT embedding, as there are now two different correlation treatments which must be bridged. We denote the “high-level” approximation $S_H$ and the “low-level” approximation $S_L$. The low-level approximation is used to construct a self-energy including all interactions in the full problem, $\boldsymbol{\Sigma}^{S_L}[\mathbf{G}, V](\omega)$, and the high-level approximation is used to construct a self-energy considering interactions only in the subsystem $A$, $\boldsymbol{\Sigma}^{S_H}_A[\mathbf{G}_A, V_A](\omega)$. The composite self-energy for the full problem is $$\begin{aligned} \boldsymbol{\Sigma}(\omega) = \boldsymbol{\Sigma}^{S_L}[ \mathbf{G}, V] + \boldsymbol{\Sigma}_A^{S_H}[ \mathbf{G}_A, V_A](\omega) - \boldsymbol{\Sigma}^{S_L}_A[ \mathbf{G}_A, V_A](\omega) \label{eq:twolevelselfenergy}\end{aligned}$$ where $\boldsymbol{\Sigma}^{S_L}_A[ \mathbf{G}_A, V_A]$ indicates the part of $\boldsymbol{\Sigma}^{S_L}[ \mathbf{G}, V]$ involving only Green’s functions and interactions within fragment $A$. Eq. (15) is analogous to the expression for the exchange-correlation potential Eq. (9) in wavefunction in DFT embedding. In fact, one common way to compute $\boldsymbol{\Sigma}^{S_H}_A[\mathbf{G}_A, V_A](\omega)$ is by carrying out a wavefunction calculation on the subsystem $A$ in the presence of additional fictitious “bath” orbitals that reproduce the effects of the hybridization $\boldsymbol{\Delta}_A$ [@Chan2011a]. The subsystem plus bath orbitals is known as a quantum impurity problem. We return to impurity problems in the context of density matrix embedding theory. Self-consistency of the Green’s functions is obtained by solving Eq. (14) and Eq. (15) for the hybridization and self-energy. Dynamical mean-field theory [@georges1992hubbard; @Kotliar2006; @Chan2011a] (DMFT) (which here refers to both single-site and cluster extensions) provides a widely used example of a higher-level self-energy embedding within a lower-level treatment. In DMFT, we divide the full problem of interest (commonly a crystal) into multiple fragments $A$ containing strongly correlated orbitals (typically transition metal $d$ and $f$ orbitals) for which a high-level self-energy approximation in each fragment, $\boldsymbol{\Sigma}_A(\omega)$, is computed. The composite self-energy in Eq. (15) becomes $$\begin{aligned} \boldsymbol{\Sigma}(\omega) = \boldsymbol{\Sigma}^{S_L}( \mathbf{G}, V) + \sum_A [\boldsymbol{\Sigma}^{S_H}_A( \mathbf{G}_A, V_A) - \boldsymbol{\Sigma}^{S_L}_A(\mathbf{G}_A, V_A)] \label{eq:dmftselfenergy}\end{aligned}$$ where the sum over $A$ reflects a summation over the fragments. For a crystal, the self-energy in each cell is identical. Self-consistency of Eq. (16) with $\boldsymbol{\Delta}_A(\omega)$ for each cell $A$ yields the cellular DMFT equations [@Biroli2002]. The low-level approximation in DMFT is often chosen to be a DFT treatment. This combination is called DFT+DMFT [@Kotliar2006], and has been widely applied to correlated materials, especially to compute the density-of-states observed in photoemission experiments [@Kotliar2006]. It has also been used, less commonly, in molecular applications, for example, to obtain correlation corrections to the conductance of a molecular junction [@Jacob2010], and qualitative features of the electronic structure in transition metal clusters and complexes [@Turkowski2012]. Unfortunately, the combination of DFT with DMFT suffers from similar problems to wavefunction in DFT embedding, such as partial double counting of interactions, and this has been a barrier to chemical accuracy. Combining DMFT with a low-level Hartree-Fock self energy avoids this issue, as the diagrams can be correctly subtracted in Eq. (16), and has recently been explored in small molecules [@Lin2011], and in a minimal basis cubic hydrogen solid [@Chan2011a]. Unfortunately, this treatment provides no description of correlations outside of the fragments, and is thus also not quantitative. Incorporating “non-local” correlations into a DMFT description is a topic of current research, and strategies include using a random-phase approximation (RPA) for the non-local correlations, which modifies both the low-level self-energy, as well as screens the Coulomb interaction $V_A \to V_A^{RPA}(\omega)$ appearing in the high-level self-energy approximation; and using a pure self-energy approximation, such as the self-consistent second-order self-energy [@nooijen1995second]. The latter has been explored by Zgid and coworkers [@Phillips2014; @Lan2015]. The strengths of Green’s function embedding are the analytic expressions for the energy and hybridization, and the diagrammatic interpretation of the self-energy approximations. However, fully ab-initio applications lag behind those of DFT embedding, because ab-initio quantum chemistry methods are primarily developed for single states (such as the ground-state) rather than the Green’s function, and computing time-dependent Green’s functions is more expensive than computing time-independent observables. Density matrix embedding ======================== The practical complexity of working with Green’s functions motivates a third formulation of embedding: density matrix embedding, where the quantity of interest is the single-particle density matrix, $\boldsymbol{\gamma}$, where $\gamma_{ij} = \langle \Psi | a^\dag_i a_j | \Psi\rangle$. It appears straightforward to formulate embedding with the density matrix, as it interpolates in complexity between the density and the Green’s function. Indeed, one can formally define a density matrix embedding to parallel DFT (and Green’s function) embedding, starting from the energy functional of the density matrix, and using stationarity to define an embedding operator $\mathbf{v}_A$, $$\begin{aligned} \mathbf{v}_A = \frac{\delta \Delta E}{\delta \boldsymbol{\gamma}_A} \label{eq:embed_1op} \end{aligned}$$ which, when substituted into the fragment Euler equation, yields the exact fragment density matrix $\boldsymbol{\gamma}_A$. Unfortunately, there is a complication that is unique to the density matrix formulation. In DFT embedding, $\Delta E$ and the embedding potential are defined with respect to a closed non-interacting system, and the latter acts to adjust the reference density to match the fragment density. (A similar statement may be made with respect to the Green’s function for Green’s function embedding). However, the [*density matrix*]{} of a closed non-interacting system must be idempotent, while the density matrix of a general (open) fragment need not be, and the two in general cannot be matched via an embedding operator. Instead, one is forced to consider more complex interacting reference systems (to represent non-idempotent fragment density matrices) or more complicated interacting density matrix functionals that minimize to non-idempotent solutions. Some of the more complex reference systems considered include ones where the ground-state is modeled by a geminal wavefunction [@Tsuchimochi2015; @Pernal2016], as well an impurity-like formulation, where the reference is the size of the full system, but with interactions restricted only to the fragment of interest [@Senjean2016]. The above complication reflects the essential physical difference between an open system, which is entangled with its environment and should be described by a mixed state, and a closed system, described by a pure state. Rather than using a more complicated pure-state to mimic a non-idempotent density matrix, one can instead model the open fragment as part of a closed system, by introducing additional bath degrees of freedom. This is the same physical idea used in the impurity representation of the hybridization in Green’s function embedding. Of course, if the bath has the same complexity as the rest of the system, nothing is gained, but in practice, the size of the bath can be significantly less than the size of the environment [@Sun2014; @Wouters2016]. A simple example is the link orbitals in QM/MM calculations, a small set of extra atomic orbitals chosen to saturate the dangling bonds of the QM fragment. Other embedding approaches, such as divide and conquer, similarly introduce buffer orbitals [@Yang1995], using orbitals spatially close to the fragment region. The choice of additional bath degrees of freedom appears to have some arbitrariness, but a construction that is provably optimal at the mean-field level has recently been provided by the DMET (density matrix embedding theory) of Knizia and Chan [@Knizia2012; @Knizia2013]. A key property of the DMET bath is that it is (at most) the same size as the fragment, and thus rigorously removes most of the degrees of freedom in the environment. DMET thus avoids the much larger baths associated with Green’s function embeddings [@Knizia2012]. To see that the environment can be compressed to this size, consider the case where one has the exact wavefunction $\ket{\Psi}$ for the full problem. This wavefunction may be rewritten (via the Schmidt decomposition) in terms of states that live solely in the fragment, and environment Hilbert spaces, $\{ \ket {\alpha_i} \}$, $\{ \ket {\beta_i} \}$, respectively $$\begin{aligned} \ket{\Psi}=\sum_i^D \lambda_i \ket{{\alpha}_i}\ket{{\beta}_i}\end{aligned}$$ The summation is over $D$ terms, the dimension of the fragment Hilbert space, and the decomposition defines $D$ exact bath states $\ket{\beta_i}$. The exact wavefunction can then be expressed within the reduced Hilbert space $\{ \ket{\alpha_i} \} \otimes \{ \ket{\beta_j} \}$. When $\ket{\Psi}$ is a Slater determinant (e.g. the ground-state of a mean-field Hamiltonian [$\hat{f}$]{}) then the Schmidt decomposition takes a particularly simple form. In particular, if the fragment has $d$ orbitals, then the bath states are spanned exactly by the Hilbert space of $d$ (partially occupied) bath orbitals, with all other orbitals either completely filled (core orbitals) or empty. The partially filled bath orbitals are the eigenvectors (with partial occupancy) of the environment block of the mean-field density matrix. This fragment plus bath representation of a mean-field wavefunction becomes the reference system in DMET, a non-interacting problem of twice the size of the fragment. Projected into this representation, the mean-field ground-state of [$\hat{f}$]{} can reproduce any fragment density matrix by augmenting with a suitable fragment operator, $\hat{f} \to \hat{f}+\hat{v}_A$. Since the bath orbitals capture the effects of embedding at the mean-field level, $\hat{v}_A$ serves to encode additional correlation effects beyond the mean-field treatment, and is thus analogous to the exchange-correlation potential in DFT, or self-energy in Green’s function embedding. Mean-field in mean-field embedding in DMET can be formulated using different levels of mean-field theory to define bath orbitals from the full problem, and to model the smaller fragment plus bath representation. However, DMET has so far mainly been applied using a correlated wavefunction description of the fragment plus bath, on top of a mean-field reference. Denoting the mean-field description by $S_L$, the DMET bath orbitals are obtained from the mean-field reference $\Psi^{S_L}$ (the ground-state of $\hat{f} + \hat{v}_A$) which defines a fragment density matrix $\boldsymbol{\gamma}_A$. The correlated energy of the [*full*]{} problem is then $$\begin{aligned} E^{S_H}[\boldsymbol{\gamma}_A] = \min_{\Psi_A^{S_H}}\langle \Psi_A^{S_H} | \hat{H} | \Psi_A^{S_H} \rangle \label{eq:dmint} \end{aligned}$$ where the correlated wavefunctions are defined in the DMET “active” space of fragment $A$ plus its bath and core orbitals. The functional dependence on $\boldsymbol{\gamma}_A$ enters in Eq. (19) through the definition of the bath, but $\boldsymbol{\gamma}_A$ is in general different from the high level density matrix $\boldsymbol{\gamma}_A^{S_H}$, obtained from $\Psi_A^{S_H}$. Self-consistency adjusts $\hat{v}_A$ such that $\boldsymbol{\gamma}_A =\boldsymbol{\gamma}^{S_H}_A$. The numerical procedure to do so involves a non-interacting inversion $\boldsymbol{\gamma}_A \to \mathbf{v}_A$, and is analogous to the inversion $\rho_A \to v_A$ in DFT embedding, with related issues of representability [@Tsuchimochi2015; @Wouters2016]. Although $E^{S_H}[\boldsymbol{\gamma_A}]$ is an energy for the full problem, correlations are only included close to fragment $A$, as the bath orbitals are typically localized close to $A$. This is acceptable for intensive quantities (such as local reactions or excitations). However, for an extensive correlated treatment (as desired in a condensed phase problem) one must embed with multiple fragments. Then, each fragment $A$ yields a separate high-level wavefunction for the full problem, $\Psi_A^{S_H}$, and expectation values must be assembled from the different fragment wavefunctions. It can be expected that expectation values for a given $\Psi_A^{S_H}$ are most accurate close to fragment $A$, and this is reflected in the partitioning of the contributions. For example, to compute the density matrix element $\gamma_{ij} = \langle a^\dag_i a_j\rangle$, we define $$\begin{aligned} \gamma_{ij} = \begin{cases} \langle \Psi_A^{S_H} | a^\dag_i a_j | \Psi_A^{S_H} \rangle & i, j \in A \\ \frac{1}{2} (\langle \Psi_A^{S_H} | a^\dag_i a_j | \Psi_A^{S_H} \rangle + \langle \Psi_{A'}^{S_H} | a^\dag_i a_j | \Psi_{A'}^{S_H} \rangle & i \in A, j \in A'. \end{cases}\end{aligned}$$ Analogous partitionings can be defined for more complicated expectation values. Because its initial development was motivated by DMFT, the majority of DMET applications have been to correlated lattice models [@bulik2014density; @Tsuchimochi2015; @Zheng2016], where the computational simplicity of the method has enabled very accurate results to be achieved using large fragments, for example, for the 2D Hubbard model [@Zheng2016]. Importantly, self-consistency in DMET, much like in DMFT, allows for non-trivial phases, such as superconductivity in repulsive systems [@Zheng2016]. Related to the physics of such lattice models are problems of strong correlation in chemical settings. The DMET bath allows the formalism to accurately treat fragments when they are bonded to their environment, even when such a bond is stretched or dissociated, as has been demonstrated through accurate calculations of the dissociation curves of molecular chains and rings[@Knizia2013; @Wouters2016]. An exciting area of application is to reduce the cost of high-level correlated wavefunction calculations in solids, particularly for small-band gap systems such as metals. Here, DMET is used to treat a unit-cell (or a small set of them) while the extra bath orbitals ameliorate the finite size effects associated with the small cell. Demonstrations on crystals in 1, 2, and 3 dimensions by Scuseria and coworkers have shown promise [@bulik2014electron]. Recent work has focused on extensions of DMET, for example, to spectra[@Booth2015], where the additional bath orbitals are modified to carry a time-dependence that reproduces the linear response of a mean-field wavefunction. This provides a rigorous way to embed excited states, which are often much more delocalized than the ground-state, and thus more sensitive to the open nature of a chemical fragment. Conclusions =========== Quantum embedding is a natural computational framework in which to think about complex systems. We have focused on three embedding approaches based on the single-particle density, Green’s function, and density matrix respectively. While we have only been able to give a short description of these approaches, we have tried to bring out their common intellectual structure. There remain many frontier methodological areas; for example, excited states and dynamics in density functional embedding, and more efficient ab-initio technology in Green’s function and density matrix embedding. New application areas are emerging, for example, in biomolecular and condensed phase simulations. In some cases it is necessary to include classical embeddings, such as through QM/MM as well. While we cannot predict the future development of the field, the growing activity strongly suggests that quantum embedding methods will remain a key part of simulating complex systems for many years to come. Biographical information ======================== Qiming Sun received his Ph. D. from Peking University. After postdoctoral work at Princeton University, he joined the California Institute of Technology as a staff scientist. He is the principal developer of the PySCF quantum chemistry package. Garnet Kin-Lic Chan received his Ph. D. from the University of Cambridge, and carried out postdoctoral work as a Junior Research Fellow at Christ’s College, Cambridge, and as a Miller Research Fellow at the University of California, Berkeley. After appointments at Cornell University and Princeton University, he joined the California Institute of Technology, where he is currently the Bren Professor of Chemistry. Acknowledgments =============== G. K.-L. Chan acknowledges support from the US National Science Foundation through grant no. NSF-CHE-1265277.

ArXiv

--- abstract: 'To quantitatively assess the impact of an eV-mass sterile neutrino on the neutrinoless double-beta ($0\nu \beta \beta$) decays, we calculate the posterior probability distribution of the relevant effective neutrino mass $|m^\prime_{ee}|$ in the (3+1)$\nu$ mixing scenario, following the Bayesian statistical approach. The latest global-fit analysis of neutrino oscillation data, the cosmological bound on the sum of three active neutrino masses from [*Planck*]{}, and the constraints from current $0\nu\beta\beta$ decay experiments are taken into account in our calculations. Based on the resultant posterior distributions, we find that the average value of the effective neutrino mass is shifted from $\overline{|m^{}_{ee}|} = 3.37\times 10^{-3}~{\rm eV}$ (or $7.71\times 10^{-3}~{\rm eV}$) in the standard 3$\nu$ mixing scenario to $\overline{|m^{\prime}_{ee}|}=2.54\times 10^{-2}~{\rm eV}$ (or $2.56\times 10^{-2}~{\rm eV}$) in the (3+1)$\nu$ mixing scenario, with the logarithmically uniform prior on the lightest neutrino mass (or on the sum of three active neutrino masses). Therefore, a null signal from the future $0\nu\beta\beta$ decay experiment with a sensitivity to $|m^{}_{ee}| \approx \mathcal{O}(10^{-2}_{})~{\rm eV}$ will be able to set a very stringent constraint on the sterile neutrino mass and the active-sterile mixing angle.' --- [**Impact of an eV-mass sterile neutrino on the neutrinoless double-beta decays: a Bayesian analysis**]{} [**Guo-yuan Huang**]{} [^1], [^2]\ Introduction ============ Whether massive neutrinos are Majorana or Dirac particles is one of the most important problems in particle physics [@Majorana:1937vz; @Racah:1937qq; @Tanabashi:2018oca]. Quite a number of neutrinoless double-beta ($0\nu\beta\beta$) decay experiments are devoted to answering this question [@Furry:1939qr; @Rodejohann:2011mu; @Bilenky:2012qi; @Rodejohann:2012xd; @Bilenky:2014uka; @Pas:2015eia; @DellOro:2016tmg]. If massive neutrinos are Majorana particles and thus lepton number violation exists in nature, then the $0\nu\beta\beta$ decays $A(Z, N) \to A(Z+2, N-2) + 2e^-$ could take place in some even-even nuclei, namely, both the proton number $Z$ and the neutron number $N$ for the nuclear isotope $A(Z, N)$ are even. Assuming the exchange of light Majorana neutrinos to be responsible for the $0\nu\beta\beta$ decays, one can find that the half-life of the relevant nuclear isotope is given by [@Rodejohann:2011mu] $$\begin{aligned} (T^{0\nu}_{1/2})^{-1} = G^{}_{0\nu}|\mathcal{M}^{}_{0\nu}|^2 \frac{|m^{}_{ee}|^2}{m^{2}_{e}}\;, \label{eq:halflife} $$ where $G^{}_{0\nu}$ is the phase-space factor, $\mathcal{M}^{}_{0\nu}$ is the nuclear matrix element (NME), and $m^{}_{e}$ is the electron mass. In Eq. (\[eq:halflife\]), the effective neutrino mass $|m^{}_{ee}|$ collects the contributions from light Majorana neutrinos involved in the $0\nu\beta\beta$ decays. In the standard three neutrino ($3\nu$) mixing scenario, the effective neutrino mass is defined as $|m^{}_{ee}| \equiv |m^{}_{1} U^2_{e1}+m^{}_{2}U^2_{e2}+m^{}_{3}U^2_{e3}|$, where the absolute neutrino masses $m^{}_i$ and the lepton flavor mixing matrix elements $U^{}_{ei}$ (for $i = 1, 2, 3$) appear. When the conventional parametrization of the flavor mixing matrix $U$ is adopted [@Tanabashi:2018oca], i.e., $U^{}_{e1} = \cos \theta^{}_{13} \cos \theta^{}_{12} e^{{\rm i}\rho/2}$, $U^{}_{e2} = \cos \theta^{}_{13} \sin \theta^{}_{12}$ and $U^{}_{e3} = \sin \theta^{}_{13} e^{{\rm i}\sigma/2}$, we have $$\begin{aligned} m^{}_{ee} \equiv m^{}_{1} \cos^2\theta^{}_{13} \cos^2\theta^{}_{12} e^{{\rm i} \rho} + m^{}_{2} \cos^2\theta^{}_{13} \sin^2\theta^{}_{12} + m^{}_{3} \sin^2\theta^{}_{13} e^{{\rm i} \sigma} \;, \label{eq:mee} $$ where $\{\theta^{}_{12}, \theta^{}_{13}\}$ are two of three neutrino mixing angles, and $\{\rho, \sigma\}$ are the Majorana-type CP-violating phases. Note that $m^{}_2$ is nonzero no matter whether the normal neutrino mass ordering (NO) with $m^{}_1 < m^{}_2 < m^{}_3$ or the inverted neutrino mass ordering (IO) with $m^{}_3 < m^{}_1 < m^{}_2$ is considered. Therefore, such a parametrization is favorable in the discussions about the limiting case of $m^{}_1 \to 0$ (for NO) or $m^{}_3 \to 0$ (for IO), for which one of two Majorana-type CP violating phases just disappears together with the lightest neutrino mass. However, if the eV-mass sterile neutrino indeed exists as a solution to the anomalies in the short-baseline neutrino experiments [@Giunti:2019aiy; @Aguilar:2001ty; @AguilarArevalo:2008rc; @Aguilar-Arevalo:2018gpe; @Giunti:2010zu; @Mention:2011rk; @Abdurashitov:2009tn; @Kaether:2010ag], it will contribute as well to the $0\nu \beta \beta$ decays. In this case, the effective neutrino mass is given by $|m^{\prime}_{ee}| \equiv |m^{}_{1} V^2_{e1} + m^{}_{2} V^2_{e2} + m^{}_{3} V^2_{e3} + m^{}_{4} V^2_{e4}|$ with $m^{}_{4}$ being the mass of the sterile neutrino and $V^{}_{ei}$ (for $i=1,2,3,4$) being the first-row elements of the mixing matrix in the (3+1)$\nu$ mixing scenario. Adopting the standard parametrization of the mixing matrix, one can express the effective neutrino mass as $$\begin{aligned} |m^{\prime}_{ee}| & \equiv & | m^{}_{ee} \cos^2\theta^{}_{14} + m^{}_{4}\sin^2\theta^{}_{14} e^{i \omega}|\;, \label{eq:meeprime} $$ where $m^{}_{ee}$ takes the same form as in Eq. (\[eq:mee\]), $\theta^{}_{14}$ is the active-sterile neutrino mixing angle, and $\omega$ is the additional Majorana-type CP-violating phase. Using the best-fit values $\Delta m^2_{41} \equiv m^2_4 - m^2_1 = 1.7~{\rm eV}^2$ and $\sin^2 \theta^{}_{14} = 0.019$ from the global-fit analysis of the short-baseline neutrino oscillation data [@Gariazzo:2017fdh; @Dentler:2018sju], one can find that the contribution from the sterile neutrino $|m^{}_4 \sin^2 \theta^{}_{14}| \approx 2.5\times 10^{-2}~{\rm eV}$ can be comparable to that from active neutrinos $|m^{}_{ee}| \lesssim 0.1~{\rm eV}$, which is constrained by the cosmological observations [@Aghanim:2018eyx] and current $0\nu\beta\beta$ decay experiments [@Albert:2014awa; @KamLAND-Zen:2016pfg; @Agostini:2017iyd; @Alduino:2017ehq; @Aalseth:2017btx; @Agostini:2018tnm]. With a ton-scale target mass, the future $0\nu\beta\beta$ experiments will be able to probe $|m^{}_{ee}|$ to the $\mathcal{O}(10^{-2})~{\rm eV}$ level [@Agostini:2017jim], covering the whole range of $|m^{}_{ee}|$ in the IO case. However, in the NO case, the effective neutrino mass can be as small as $|m^{}_{ee}| \approx (1.6 \cdots 3.6) \times 10^{-3} ~{\rm eV}$ when the lightest neutrino mass $m^{}_{1}$ is vanishing, or even vanishing in the contrived region of parameter space when the cancellation among the contributions from different neutrino mass eigenstates occurs [@Xing:2003jf; @Xing:2015zha; @Xing:2016ymd]. Moreover, the latest global-fit analysis of neutrino oscillation data [@deSalas:2017kay; @Capozzi:2018ubv; @Esteban:2018azc] does show a preference for the NO at the $3\sigma$ confidence level (C.L.), it is worrisome that $|m^{}_{ee}|$ may be out of the reach of the next generation $0\nu \beta\beta$ decay experiments. To quantitatively assess how likely $|m^{}_{ee}|$ is small, the authors of Refs. [@Agostini:2017jim; @Caldwell:2017mqu] have carried out a Bayesian analysis and obtained the posterior distribution of $|m^{}_{ee}|$, given the neutrino oscillation data, current experimental upper bounds on $|m^{}_{ee}|$ and the cosmological bound on the sum of three neutrino masses. For the earlier relevant works, see Refs. [@Benato:2015via; @Zhang:2015kaa; @Ge:2016tfx]. Although the impact of an eV-mass sterile neutrino on the effective neutrino mass $|m^\prime_{ee}|$ has been considered in Refs.  [@Goswami:2005ng; @Goswami:2007kv; @Barry:2011wb; @Li:2011ss; @Girardi:2013zra; @Guzowski:2015saa; @Giunti:2015kza; @Ge:2017erv; @Liu:2017ago], a statistical assessment is still lacking. Therefore, we are motivated to perform a Bayesian analysis of $|m^\prime_{ee}|$ in this work by using the global-fit results of neutrino oscillation data and other available information on the absolute neutrino masses. The rest of the present paper is organized as follows. In Section 2, we describe the necessary information for the Bayesian analysis. The prior information can be extracted from the global-fit analysis of neutrino oscillation data [@Capozzi:2018ubv; @Gariazzo:2017fdh], the cosmological observations [@Aghanim:2018eyx] and the existing $0\nu\beta\beta$ decay experiments [@Albert:2014awa; @KamLAND-Zen:2016pfg; @Agostini:2017iyd; @Alduino:2017ehq]. Then, the posterior distribution of the standard effective neutrino mass $|m^{}_{ee}|$ and that of $|m^\prime_{ee}|$ are presented in Section 3. Two-dimensional posterior probability densities in the $|m^{\prime}_{ee}|$-$m^{}_{\rm L}$ plane and those in the $|m^{\prime}_{ee}|$-$\rho$ plane have also been given, where $m^{}_{\rm L}$ denotes the lightest neutrino mass. Finally, we make some concluding remarks in Section 4. The Bayesian Analysis ===================== The Bayesian analysis provides us with a reasonable statistical framework to update the probability distribution of model parameters in light of the new experimental data. The posterior distribution of model parameters can be obtained according to the Bayesian theorem [@Skilling:book] $$\begin{aligned} \label{eq:Bayesian} P(\Theta,\mathcal{H}^{}_{i}|\mathcal{D}) = \frac{\mathcal{L}(\mathcal{D}|\Theta, \mathcal{H}^{}_{i})\mathcal{\pi}(\Theta,\mathcal{H}^{}_{i})}{\sum^{}_{i}\mathcal{Z}^{}_{i}}\;, $$ where $\Theta$ denotes the set of model parameters, $\mathcal{D}$ stands for the available experimental data, and $\{\mathcal{H}^{}_{i}\}$ are the hypotheses or models with $i$ being the model index. Here $\mathcal{L}(\mathcal{D}|\Theta, \mathcal{H}^{}_{i})$ is the likelihood of the data $\mathcal{D}$, assuming the model $\mathcal{H}^{}_{i}$ with the parameters $\Theta$, $\mathcal{\pi}(\Theta,\mathcal{H}^{}_{i})$ is the prior distribution of $\Theta$, and $\mathcal{Z}^{}_{i}$ is the evidence. The evidence $\mathcal{Z}^{}_{i}$ is given by $$\begin{aligned} \label{eq:ZEvidence} \mathcal{Z}^{}_{i} = \int \mathcal{L}(\mathcal{D}|\Theta, \mathcal{H}^{}_{i})\mathcal{\pi}(\Theta,\mathcal{H}^{}_{i}) d^{N}\Theta\;, $$ which measures the compatibility of the model with the data, and $N$ is just the dimension of the parameter space. The hypotheses relevant for our analysis are $\mathcal{H}^{}_{\rm NO}$ for the NO and $\mathcal{H}^{}_{\rm IO}$ for the IO in the $3\nu$ or (3+1)$\nu$ mixing scenario. The model parameters in the (3+1)$\nu$ mixing scenario include: (i) the involved neutrino oscillation parameters $\{ \sin^2\theta^{}_{13}, \sin^2\theta^{}_{12}, \sin^2\theta^{}_{14},\Delta m^{2}_{\rm sol},\Delta m^{2}_{\rm atm},\Delta m^{2}_{41} \}$, where $\Delta m^2_{\rm sol} \equiv m^2_2 - m^2_1$ and $\Delta m^2_{\rm atm} \equiv m^2_3 - (m^2_2 + m^2_1)/2$ are two mass-squared differences of ordinary neutrinos; (ii) the lightest neutrino mass $m^{}_{\rm L}$, which is $m^{}_{1}$ for $\mathcal{H}^{}_{\rm NO}$ and $m^{}_{3}$ for $\mathcal{H}^{}_{\rm IO}$; (iii) the Majorana-type CP-violating phases $\{\rho,\sigma,\omega\}$; (iv) the phase-space factor and the nuclear matrix element $\{ G^{}_{0\nu}, |\mathcal{M}^{}_{0\nu}| \}$ for the $0\nu\beta\beta$ decays. The overall likelihood function can be constructed as $\mathcal{L} = \mathcal{L}^{}_{\rm 3\nu} \times \mathcal{L}^{}_{\rm cosmo} \times \mathcal{L}^{}_{\rm 0\nu\beta\beta} \times \mathcal{L}^{}_{\rm sterile}$, and the details of the individual likelihood function are summarized as follows. - $\mathcal{L}^{}_{\rm 3\nu}$: the likelihood function of the $3\nu$ mixing parameters $\{ \sin^2\theta^{}_{13}, \sin^2\theta^{}_{12}, \Delta m^{2}_{\rm sol}, \Delta m^{2}_{\rm atm} \}$. Given the $\Delta \chi^2$ function from the global-fit analysis in Ref. [@Capozzi:2018ubv], we can fix the likelihood function $\mathcal{L}^{}_{\rm 3\nu} = \exp(-\Delta \chi^2/2)$, where $\Delta \chi^2$ is defined as $$\begin{aligned} \Delta \chi^2 \equiv \sum^{}_{i} \frac{(\Theta^{}_{i}-\Theta^{\rm bf}_{i})^2}{\sigma^{2}_{i}}\;,\end{aligned}$$ with $\Theta^{}_{i}$ running over $\{ \sin^2\theta^{}_{13}, \sin^2 \theta^{}_{12}, \Delta m^{2}_{\rm sol}, \Delta m^{2}_{\rm atm} \}$, $\Theta^{\rm bf}_{i}$ the corresponding best-fit value from the global analysis, and $\sigma^{}_{i}$ the symmetrized $1\sigma$ error. See Table. 1 of Ref. [@Capozzi:2018ubv] for more details about the global-fit results of neutrino oscillation data. To be explicit, we list the best-fit values and the corresponding symmetrized $1\sigma$ errors as below $$\begin{aligned} && \sin^2\theta^{}_{12} = (3.04\pm 0.14) \times 10^{-1} \; , \quad \Delta m^2_{\rm sol} = (7.34 \pm 0.16) \times 10^{-5} ~{\rm eV}^2 \; , \nonumber \\ && \sin^2\theta^{}_{13} = (2.14 \pm 0.08) \times 10^{-2}\; , ~~\quad \Delta m^2_{\rm atm} = (2.455 \pm 0.034) \times 10^{-3} ~{\rm eV}^2 \; ,\hspace{0.6cm} $$ for $\mathcal{H}^{}_{\rm NO}$; and $$\begin{aligned} && \sin^2\theta^{}_{12} = (3.03\pm 0.14) \times 10^{-1} \; , \quad \Delta m^2_{\rm sol} = (7.34 \pm 0.16) \times 10^{-5} ~{\rm eV}^2 \; , \nonumber \\ && \sin^2\theta^{}_{13} = (2.18 \pm 0.08)\times 10^{-2} \; , \quad \Delta m^2_{\rm atm} = (-2.441 \pm 0.034) \times 10^{-3} ~{\rm eV}^2 \; ,\hspace{0.6cm} $$ for $\mathcal{H}^{}_{\rm IO}$. The latest neutrino oscillation data favor the NO over the IO at the $3\sigma$ level, i.e., the difference between the minima of $\chi^2$ in these two cases is $\Delta \chi^2_{\rm min} \equiv \chi^{\rm IO}_{\rm min} - \chi^{\rm NO}_{\rm min} \approx 9$. The preference for the NO arises mainly from two different data sets. First, the excess of $\nu^{}_e$-like events in the multi-GeV energy range in Super-Kamiokande atmospheric neutrino data can be accommodated by the resonant enhancement of the oscillation probability in the $\nu^{}_\mu \to \nu^{}_e$ channel, leading to $\Delta \chi^2_{\rm min} \approx 4$. Second, the running long-baseline accelerator experiments T2K and NO$\nu$A prefer the value of $\theta^{}_{13}$ that is slightly larger than the precisely measured value from reactor neutrino experiments. Such a tension between accelerator and reactor neutrino experiments will be relieved in the NO case, contributing another $\Delta \chi^2_{\rm min} \approx 4$ to the mass ordering discrimination. To be conservative, we will take $\Delta\chi^2_{\rm min} = 4$ as the preference for the NO over the IO from neutrino oscillation data. ![The likelihood function $\mathcal{L}^{}_{\rm cosmo}$ for the sum of three neutrino masses $\Sigma \equiv m^{}_1 + m^{}_2 + m^{}_3$ from cosmological observations, which has been derived by combining the $Planck ~{\rm TT}, {\rm TE}, {\rm EE} + {\rm lowE} + {\rm lensing} + {\rm BAO}$ data sets [@Aghanim:2018eyx].[]{data-label="fig:1"}](lh_sum.pdf){width="48.00000%"} - $\mathcal{L}^{}_{\rm cosmo}$: the likelihood function for the cosmological observations on the sum of three neutrino masses $\Sigma \equiv m^{}_{1}+m^{}_{2}+m^{}_{3}$. After combining several different sets of cosmological data ($Planck ~{\rm TT}, {\rm TE}, {\rm EE} + {\rm lowE} + {\rm lensing} + {\rm BAO}$), the [*Planck*]{} Collaboration has recently updated the upper limit on the sum of neutrino masses as $\Sigma < 0.12~{\rm eV}$ at the $95\%$ C.L. [@Aghanim:2018eyx]. We obtain the likelihood information by making use of the Markov chain file available from the Planck Legacy Archive (PLA) [^3]. The likelihood function of $\Sigma$ is produced and shown in Fig. \[fig:1\] by marginalizing over the other cosmological parameters. Although the sampling file given by PLA has assumed a degenerate mass spectrum of neutrinos, a more solid analysis with the realistic neutrino mass spectrum should not change the result much [@Hannestad:2016fog]. For this reason, the likelihood shown in Fig.  \[fig:1\] will be used in the following discussions. - $\mathcal{L}^{}_{\rm 0\nu\beta\beta}$: the likelihood function derived from the experimental constraints on the effective neutrino mass $|m^{}_{ee}|$ or $|m^{\prime}_{ee}|$ due to the existing searches for $0\nu\beta\beta$ decays. For simplicity, we implement the likelihood function available from Refs. [@Caldwell:2017mqu; @Alduino:2017ehq] in our analysis. Although both $\mathcal{L}^{}_{\rm 0\nu\beta\beta}$ and $\mathcal{L}^{}_{\rm cosmo}$ contain the information about the absolute scale of neutrino masses, the constraint on $|m^{}_{ee}|$ from the $0\nu\beta\beta$ decays suffers from a large theoretical uncertainty in the prediction for the NME. For instance, the tightest bound comes from the KamLAND-Zen experiment [@KamLAND-Zen:2016pfg], namely, $|m^{}_{ee}| \lesssim (61\cdots 165)~{\rm meV}$. Given further uncertainties from the mixing parameters and the unknown Majorana CP-violating phases, the $0\nu\beta\beta$ decays are not so informative about the absolute scale of neutrino masses when compared to the cosmological observations. - $\mathcal{L}^{}_{\rm sterile}$: the likelihood function encoding the global-fit analysis of sterile neutrino mass and mixing parameters $\{ \theta^{}_{14}, \Delta m^{2}_{41}\}$. In practice, we determine the likelihood function as $\mathcal{L}^{}_{\rm sterile} = \exp[ -\Delta \chi^2_{\rm sterile}(\theta^{}_{14}, \Delta m^{2}_{41}) /2]$ by using the $\Delta \chi^2$ distribution in Fig. 9 of Ref. [@Gariazzo:2017fdh]. The result of the so-called pragmatic 3+1 global fit “PrGlo17” will be utilized [@Gariazzo:2017fdh], where the tension between appearance and disappearance oscillation data can be somewhat relaxed by ignoring the excess of low-energy ${\nu}^{}_{e}$-like events observed in the MiniBooNE experiment. After having the likelihood functions constructed from various experimental observations, we need to make clear the prior probability distributions of the model parameters, which reflect our knowledge about them prior to any experimental data. First, neutrino mass-squared differences and mixing angles $\{ \sin^2\theta^{}_{13}, \sin^2\theta^{}_{12}, \sin^2\theta^{}_{14}, \Delta m^{2}_{\rm sol}, \Delta m^{2}_{\rm atm}, \Delta m^{2}_{41} \}$ are assumed to be uniformly distributed in their allowed ranges that are wide enough to cover their global fit results. Since the oscillation data are rather informative, different choices of prior distributions of these parameters do not have much impact on the final posterior distributions. Second, the Majorana CP-violating phases are completely unknown, so it is reasonable to adopt the flat priors in the range of $[0\cdots 2\pi)$. In addition, we have to mention that the prior distributions for the following relevant parameters are by no means unique but will be incorporated into our calculations for practical purposes. - As indicated in Eq. (\[eq:halflife\]), the phase-space factor $G^{}_{0\nu}$ and the NME $|\mathcal{M}^{}_{0\nu}|$ are needed when we try to translate the experimental constraint on the half-life into that on the effective neutrino mass. The phase-space factors for different nuclear isotopes have been computed in Refs. [@Rodejohann:2011mu; @Suhonen:1998ck; @Kotila:2012zza], and we use the central values from Ref. [@Kotila:2012zza], e.g., $G^{}_{0\nu}({}^{76}{\rm Ge}) = 6.15 \times 10^{-15}~{\rm yr}^{-1}$, $G^{}_{0\nu}({}^{130}{\rm Te}) = 3.70 \times 10^{-14}~{\rm yr}^{-1}$ and $G^{}_{0\nu}({}^{136}{\rm Xe}) = 3.79 \times 10^{-14}~{\rm yr}^{-1}$, which have been obtained with the axial vector coupling constant $g^{}_{\rm A} = 1.27$. We assume that $G^{}_{0\nu}$ can be described by the Gaussian distribution with the aforementioned central value and a relative error of $7\%$. On the other hand, the NME for a specific nuclear isotope encoding the information about the nuclear structure has been theoretically calculated in a variety of nuclear models. The differences among these calculations can be treated as the theoretical uncertainty. We define this uncertainty as $\sigma^{}_{\rm NME} \equiv \sum^{}_{i}(|\mathcal{M}^{i}_{0\nu}|-\overline{|\mathcal{M}^{}_{0\nu}|})^2/n^{}_{\rm NME}$, where $|\mathcal{M}^{i}_{0\nu}|$ is the NME value of the $i$th model, $\overline{|\mathcal{M}^{}_{0\nu}|}$ is the averaged NME value of all models, and $n^{}_{\rm NME}$ is the total number of models. Using the tabulated NME values in Ref. [@Guzowski:2015saa], we find that $\overline{|\mathcal{M}^{}_{0\nu}|}({}^{76}{\rm Ge},{}^{130}{\rm Te},{}^{136}{\rm Xe}) = (4.88,3.94,2.73) $ and $\sigma^{}_{\rm NME}({}^{76}{\rm Ge},{}^{130}{\rm Te},{}^{136}{\rm Xe}) = (1.14,0.90,0.80)$. Then the Gaussian distribution with the central value $\overline{|\mathcal{M}^{}_{0\nu}|}$ and the standard deviation $\sigma^{}_{\rm NME}$ is assumed for each nuclear isotope. - For the prior of the lightest neutrino mass $m^{}_{\rm L}$, a more careful study should be performed. Four kinds of prior distributions for $m^{}_{\rm L}$ are usually considered: (i) a logarithmic prior on $m^{}_{\rm L}$ with an adjustable lower cutoff that we choose to be $10^{-4}~{\rm eV}$; (ii) a logarithmic prior on $\Sigma$ with a natural lower cutoff at $0.06~{\rm eV}$ for NO or at $0.1~{\rm eV}$ for IO, as required by neutrino oscillation experiments; (iii) a flat prior on $m^{}_{\rm L}$; (iv) a flat prior on $\Sigma$. The prior probability distributions have been plotted with respect to $\log_{10}(m^{}_{\rm L}/{\rm eV})$ in the left panel of Fig. \[fig:2\], where one can see that the flat priors on $m^{}_{\rm L}$ (gray solid curve) and $\Sigma$ (gray dashed curve) lead to nearly the same distribution. After incorporating the experimental limits from [*Planck*]{} 2018 and the $0\nu\beta\beta$ decays, as shown in the right panel of Fig. \[fig:2\], we observe that the logarithmic prior on $m^{}_{\rm L}$ (red solid curve) gives rise to a posterior distribution that is very different from those in the other scenarios. This is because a large weight has been given to very small neutrino masses in the former case. In the following discussions, we focus only on two different prior distributions, i.e., the logarithmic prior on $m^{}_{\rm L}$ and the logarithmic prior on $\Sigma$, both of which are scale invariant. Since the posterior distribution of $m^{}_{\rm L}$ with logarithmic prior on $\Sigma$ is very similar to those with two flat priors, the posterior distribution of the effective neutrino mass in the former case should also be roughly applicable to those in the latter two cases. Finally, we make some comments on the current experimental hint on neutrino mass ordering by combining the data sets of neutrino oscillation experiments, cosmological observations and the $0\nu\beta\beta$ decays, for which the likelihood functions are given by $\mathcal{L}^{}_{\rm 3\nu}$, $\mathcal{L}^{}_{\rm cosmo}$ and $\mathcal{L}^{}_{\rm 0\nu\beta\beta}$, respectively. The preference odds for NO over IO can be represented by the Bayes factor, i.e., $\mathcal{B} \equiv \mathcal{Z}^{}_{\rm NO}/\mathcal{Z}^{}_{\rm IO}$. With the help of Eq. (\[eq:ZEvidence\]), one can calculate the evidences for NO and IO and thus their ratio. The dependence of $\mathcal{B}$ on the choice of the $m^{}_{\rm L}$ prior distribution is found to be very weak. Given identical prior information on both mass orderings, we consider only the cosmological observations $\mathcal{L}^{}_{\rm cosmo}$ and obtain the logarithm of the Bayes factor as $\log(\mathcal{B}^{}_{\rm cosmo}) \approx 0.85$ [^4], corresponding to ${\cal B}^{}_{\rm cosmo} \approx 2.34$, which is in concordance with the results from Refs. [@Hannestad:2016fog; @Gerbino:2016ehw; @Vagnozzi:2017ovm; @Capozzi:2017ipn]. If only the $0\nu\beta\beta$ decay experiments are considered, then we get $\log(\mathcal{B}^{}_{ 0\nu\beta\beta}) \approx 0.2$. A combination of the cosmological observations and $0\nu\beta\beta$ decay data leads to $\log(\mathcal{B}^{}_{ {\rm cosmo} + 0\nu\beta\beta}) \approx 1.1$. Regarding the three-flavor neutrino oscillation data, if we take the conservative choice of $\Delta \chi^2_{\rm min} \approx 4$ for two neutrino mass orderings, which has been used to construct ${\cal L}^{}_{3\nu}$, the logarithm of the Bayes factor turns out to be $\log(\mathcal{B}^{}_{ 3\nu})= 2$. Combining ${\cal L}^{}_{\rm cosmo}$, ${\cal L}^{}_{0\nu\beta\beta}$ and ${\cal L}^{}_{3\nu}$ together, one can find the total Bayes factor $\mathcal{B}^{}_{\rm tot} \approx 22$. As we have mentioned before, the global-fit analysis of all the neutrino oscillation data gives rise to a $3\sigma$ preference for the NO, corresponding to ${\cal B}^{}_{3\nu} \approx 90$. If such a stronger preference for the NO is implemented instead of the conservative one, the total Bayes factor from all the data sets becomes $\mathcal{B}^{}_{\rm tot} \approx 270$, showing a strong evidence for the NO according to the Jeffreys scale [@Trotta:2008qt]. The addition of $\mathcal{L}^{}_{\rm sterile}$ into the analysis does not alter the above conclusions, since the short-baseline neutrino oscillation experiments are insensitive to the mass ordering of three ordinary neutrinos. Posterior Distributions ======================= After specifying the likelihood functions for the relevant experimental data and fixing the prior probability distributions of model parameters in the previous section, we are ready to compute the posterior distributions of the derived parameters $|m^{}_{ee}|$ and $|m^{\prime}_{ee}|$ by using Eq. (\[eq:Bayesian\]). In fact, the posterior probability distribution in Eq. (\[eq:Bayesian\]) for the model parameters is calculated via the Monte Carlo sampling, which has been done with the help of the MultiNest routine [@Feroz:2007kg; @Feroz:2008xx; @Feroz:2013hea]. In Fig. \[fig:3\], we present the posterior sampling distributions in the $|m^{}_{ee}|$-$m^{}_{\rm L}$ plane for the standard 3$\nu$ mixing scenario (the upper row) or in the $|m^{\prime}_{ee}|$-$m^{}_{\rm L}$ plane for the (3+1)$\nu$ mixing scenario (the lower row). The scattered points stand for the sampling data, and one can read off the corresponding posterior probabilities from their colors. Now we explain how to practically do so. For a given point, one can first look at the color legend and find the value of its posterior density, which is denoted as $p$. Then, the posterior probability $P$ can be calculated by definition as the product of $p$ and the area $\mathcal{A}$ of a small region, in which the point is located. For instance, take a small square in the $|m^{}_{ee}|$-$m^{}_{\rm L}$ plane, and its area is thus given by $\mathrm{d}\mathcal{A} \equiv \mathrm{d} \left[\log^{}_{10}(|m^{}_{ee}|/{\rm eV})\right] \times \mathrm{d}\left[\log^{}_{10}(m^{}_{\rm L}/{\rm eV})\right]$. Notice that the total posterior probability is normalized to one for each plot. Several comments on the numerical results in Fig. \[fig:3\] are helpful. 1. In the upper-left panel, the posterior distribution in the $|m^{}_{ee}|$-$m^{}_{\rm L}$ plane is shown for the standard 3$\nu$ mixing scenario, where the logarithmic prior on $m^{}_{\rm L}$ is assumed. The results for the logarithmic prior on $\Sigma$ are plotted in the upper-right panel. In both panels, the thin dot-dashed (or dashed) curves indicate the boundaries of the effective neutrino mass $|m^{}_{ee}|$ in the IO (or NO) case, where the best-fit values of neutrino mixing angles and mass-squared differences are input. Moreover, the current limit (taken from Ref. [@KamLAND-Zen:2016pfg] for the tightest one) on or the future sensitivity (of a ton-scale $0\nu\beta\beta$ decay experiment like nEXO [@Gerbino:2016ehw]) to $|m^{}_{ee}|$ is represented by three horizontal dotted lines. The wide range between the upper and lower lines can be ascribed to the NME uncertainty. Comparing the distributions in the left and right panels, one can observe that a larger weight has been given to smaller values of $m^{}_{\rm L}$ in the assumption of a logarithmic prior on $m^{}_{\rm L}$, as already emphasized in the previous section. 2. An urgent question is how likely $|m^{}_{ee}|$ is vanishingly small in the NO case, which has been quantitatively addressed in Refs. [@Agostini:2017jim; @Caldwell:2017mqu]. In order to draw a prior-independent conclusion from the posterior distributions, we treat the scenarios with different values of $m^{}_{\rm L}$ as different models. For each fixed $m^{}_{\rm L}$, the posterior distribution of $|m^{}_{ee}|$ can be derived with the help of the likelihood $\mathcal{L}^{}_{ 3\nu}$. Then, one can calculate the probability for the true value of the effective neutrino mass to be above a certain $|m^{}_{ee}|$. The probability contours are plotted as the blue curves in Fig. \[fig:3\], where several representative values, i.e., $68\%$, $95\%$, $99\%$ and $99.7\%$, are shown. It is evident that the probability for $|m^{}_{ee}|$ to be vanishingly small, e.g., $|m^{}_{ee}| < 10^{-4}~{\rm eV}$, is tiny (less than $0.3\%$). This conclusion is independent of the priors on $m^{}_{\rm L}$, as it should be. In particular, the probability for $|m^{}_{ee}| > 10^{-3}~{\rm eV}$ is larger than $95\%$ even when $m^{}_{\rm L}$ is located in the regime where the destructive cancellation caused by the unknown Majorana CP phases occurs. 3. In the two panels in the lower row of Fig. \[fig:3\], the posterior probability distributions in the (3+1)$\nu$ mixing scenario have been presented, where the notations and conventions for the curves are the same as those in the plots in the upper row. It is straightforward to observe that the presence of the eV-mass sterile neutrino shifts the effective neutrino mass to higher values. As the future ton-scale $0\nu\beta\beta$ decay experiments are able to explore the effective neutrino mass to the level of $\mathcal{O}(10^{-2})~{\rm eV}$, the inclusion of the sterile neutrino can raise the effective mass to the level that is within the reach of the next-generation experiments even for a very small $m^{}_{\rm L}$. If the sensitivity at the $\mathcal{O}(10^{-2})~{\rm eV}$ level is achieved, more than $99.7\%$ of the region of $|m^{\prime}_{ee}|$ can be covered for $m^{}_{\rm L} \lesssim 10^{-2}~{\rm eV}$. When $m^{}_{\rm L} \gtrsim 10^{-2}~{\rm eV}$, the chance for $|m^{\prime}_{ee}|$ to fall into the cancellation region increases. However, even in this case, at least $95\%$ of the $|m^\prime_{ee}|$ range can be probed. Therefore, in the statistical sense, it is quite promising to check the (3+1)$\nu$ mixing scenario with an eV-mass sterile neutrino in the future $0\nu\beta\beta$ decay experiments. In Fig. \[fig:4\], we present the posterior distributions in the $|m^{}_{ee}|$-$\rho$ (the upper row) or $|m^\prime_{ee}|$-$\rho$ plane (the lower row) by marginalizing over the lightest neutrino mass $m^{}_{\rm L}$ instead of the Majorana CP phase $\rho$. The notations and conventions are the same as those in Fig. \[fig:3\]. The area in the $|m^{}_{ee}|$-$\rho$ plane is defined as $\mathrm{d}\mathcal{A} \equiv \mathrm{d} \left[\log^{}_{10}(|m^{}_{ee}|/{\rm eV})\right] \times \mathrm{d}\left[\rho/{\rm rad}\right]$ in the $3\nu$ mixing scenario, and likewise for the (3+1)$\nu$ mixing scenario. Now the blue solid curves in Fig. \[fig:4\] stand for the contours of the probability for the effective neutrino mass to be above a certain $|m^{}_{ee}|$ or $|m^\prime_{ee}|$. These contours become dependent on the $m^{}_{\rm L}$ priors, because the prior information of $m^{}_{\rm L}$ has been integrated into the posterior distribution. It is worthwhile to notice that the dependence of posterior distributions on $\rho$ is very weak for the (3+1)$\nu$ mixing scenario. In the $3\nu$ mixing scenario, the fine structure around $\rho \approx \pi$ due to the cancellation can be observed. Therefore, it seems difficult to determine the Majorana CP phase $\rho$ if $|m^{}_{ee}|$ takes the value far away from that in the cancellation region. As the effective neutrino mass can be directly extracted from the experimental data on $0\nu\beta\beta$ decays, it is interesting to see the posterior distribution of $|m^{}_{ee}|$ or $|m^\prime_{ee}|$, which can be obtained by marginalizing over both $m^{}_{\rm L}$ and $\rho$. The final results can be found in Fig. \[fig:5\]. For the standard $3\nu$ case in the left panel, if we choose the logarithmic prior on $m^{}_{\rm L}$ for NO (red solid curve), a large fraction (about $92\%$) of the probable range of $|m^{}_{ee}|$ is unreachable for the future ton-scale $0\nu \beta\beta$ decay experiments. With a logarithmic prior on $\Sigma$ (blue solid curve), the next generation experiments can cover about $ 41\%$ of the range. As we have observed before, adding an eV-mass sterile neutrino can greatly enhance the probability of the effective neutrino mass $|m^\prime_{ee}|$ to larger values. The future $0\nu\beta\beta$ decay experiments with a sensitivity to the effective neutrino mass of $\mathcal{O}(10^{-2})~{\rm eV}$ can cover around $99.4\%$ ($97.4\%$) of the posterior space for the logarithmic prior on $m^{}_{\rm L}$ (the logarithmic prior on $\Sigma$) in the (3+1)$\nu$ mixing scenario. According to the posterior distributions in Fig. \[fig:5\], we find that the average value of the effective neutrino mass is shifted from $\overline{|m^{}_{ee}|} = 3.37\times 10^{-3}~{\rm eV}$ (or $7.71\times 10^{-3}~{\rm eV}$) in the standard $3\nu$ mixing scenario to $\overline{|m^{\prime}_{ee}|}=2.54\times 10^{-2}~{\rm eV}$ (or $2.56\times 10^{-2}~{\rm eV}$) in the (3+1)$\nu$ mixing scenario, with the logarithmic prior on $m^{}_{\rm L}$ (or on $\Sigma$). Therefore, a null signal from the future $0\nu\beta\beta$ decay experiments will be able to set a very stringent constraint on the sterile neutrino mass and mixing angle. Concluding Remarks ================== In this short note, we have carried out a Bayesian analysis of the effective neutrino mass in the $0\nu\beta\beta$ decays in both the standard $3\nu$ mixing scenario and the (3+1)$\nu$ mixing scenario. With the latest experimental information, including the global-fit analysis of neutrino oscillation data, the cosmological observations from the [*Planck*]{} satellite and the current limits from the $0\nu\beta\beta$ decay experiments, the posterior probability distributions of the effective neutrino mass $|m^{}_{ee}|$ in the standard $3\nu$ mixing scenario and $|m^\prime_{ee}|$ in the (3+1)$\nu$ mixing scenario have been updated. Our main results of the posterior distributions have been summarized in Fig. \[fig:3\] and Fig. \[fig:5\]. Adding an eV-mass sterile neutrino slightly mixing with ordinary neutrinos is likely to enhance the effective neutrino mass to the level of ${\cal O}(10^{-2})~{\rm eV}$, which is within the reach of the next generation $0\nu\beta\beta$ decay experiments, regardless of the prior information on the absolute mass scale of ordinary neutrinos. In other words, if a null signal is observed in future ton-scale $0\nu\beta\beta$ decay experiments, we can place very strong limits on the parameter space of the (3+1)$\nu$ mixing scenario, assuming that massive neutrinos are of Majorana nature. The sensitivity of future $0\nu\beta\beta$ decay experiments to the sterile neutrino mass and mixing angle deserves a dedicated study, which will be left for the upcoming works. Acknowledgements {#acknowledgements .unnumbered} ================ This work was supported in part by the National Natural Science Foundation of China under grant No. 11775232 and No. 11835013, and by the CAS Center for Excellence in Particle Physics. [99]{} E. Majorana, “Teoria simmetrica dell’elettrone e del positrone,” Nuovo Cim.  [**14**]{}, 171 (1937). G. Racah, “On the symmetry of particle and antiparticle,” Nuovo Cim.  [**14**]{}, 322 (1937). M. Tanabashi [*et al.*]{} \[Particle Data Group\], “Review of Particle Physics,” Phys. Rev. D [**98**]{}, no. 3, 030001 (2018). W. H. Furry, “On transition probabilities in double beta-disintegration,” Phys. Rev.  [**56**]{}, 1184 (1939). W. Rodejohann, “Neutrino-less Double Beta Decay and Particle Physics,” Int. J. Mod. Phys. E [**20**]{}, 1833 (2011) \[arXiv:1106.1334\]. S. M. Bilenky and C. Giunti, “Neutrinoless double-beta decay: A brief review,” Mod. Phys. Lett. A [**27**]{}, 1230015 (2012) \[arXiv:1203.5250\]. W. Rodejohann, “Neutrinoless double beta decay and neutrino physics,” J. Phys. G [**39**]{}, 124008 (2012) \[arXiv:1206.2560\]. S. M. Bilenky and C. Giunti, “Neutrinoless Double-Beta Decay: a Probe of Physics Beyond the Standard Model,” Int. J. Mod. Phys. A [**30**]{}, no. 04n05, 1530001 (2015) \[arXiv:1411.4791\]. H. Päs and W. Rodejohann, “Neutrinoless Double Beta Decay,” New J. Phys.  [**17**]{}, no. 11, 115010 (2015) \[arXiv:1507.00170\]. S. Dell’Oro, S. Marcocci, M. Viel and F. Vissani, “Neutrinoless double beta decay: 2015 review,” Adv. High Energy Phys.  [**2016**]{}, 2162659 (2016) \[arXiv:1601.07512\]. C. Giunti and T. Lasserre, “eV-scale Sterile Neutrinos,” arXiv:1901.08330. A. Aguilar-Arevalo [*et al.*]{} \[LSND Collaboration\], “Evidence for neutrino oscillations from the observation of anti-neutrino(electron) appearance in a anti-neutrino(muon) beam,” Phys. Rev. D [**64**]{}, 112007 (2001) \[hep-ex/0104049\]. A. A. Aguilar-Arevalo [*et al.*]{} \[MiniBooNE Collaboration\], “Unexplained Excess of Electron-Like Events From a 1-GeV Neutrino Beam,” Phys. Rev. Lett.  [**102**]{}, 101802 (2009). A. A. Aguilar-Arevalo [*et al.*]{} \[MiniBooNE Collaboration\], “Significant Excess of ElectronLike Events in the MiniBooNE Short-Baseline Neutrino Experiment,” arXiv:1805.12028. C. Giunti and M. Laveder, “Statistical Significance of the Gallium Anomaly,” Phys. Rev. C [**83**]{}, 065504 (2011) \[arXiv:1006.3244\]. G. Mention, M. Fechner, T. Lasserre, T. A. Mueller, D. Lhuillier, M. Cribier and A. Letourneau, “The Reactor Antineutrino Anomaly,” Phys. Rev. D [**83**]{}, 073006 (2011) \[arXiv:1101.2755\]. J. N. Abdurashitov [*et al.*]{} \[SAGE Collaboration\], “Measurement of the solar neutrino capture rate with gallium metal. III: Results for the 2002–2007 data-taking period,” Phys. Rev. C [**80**]{}, 015807 (2009) \[arXiv:0901.2200\]. F. Kaether, W. Hampel, G. Heusser, J. Kiko and T. Kirsten, “Reanalysis of the GALLEX solar neutrino flux and source experiments,” Phys. Lett. B [**685**]{}, 47 (2010) \[arXiv:1001.2731\]. S. Gariazzo, C. Giunti, M. Laveder and Y. F. Li, “Updated Global 3+1 Analysis of Short-BaseLine Neutrino Oscillations,” JHEP [**1706**]{}, 135 (2017) \[arXiv:1703.00860\]. M. Dentler, Á. Hernández-Cabezudo, J. Kopp, P. A. N. Machado, M. Maltoni, I. Martinez-Soler and T. Schwetz, “Updated Global Analysis of Neutrino Oscillations in the Presence of eV-Scale Sterile Neutrinos,” JHEP [**1808**]{}, 010 (2018) \[arXiv:1803.10661\]. N. Aghanim [*et al.*]{} \[Planck Collaboration\], “Planck 2018 results. VI. Cosmological parameters,” arXiv:1807.06209. J. B. Albert [*et al.*]{} \[EXO-200 Collaboration\], “Search for Majorana neutrinos with the first two years of EXO-200 data,” Nature [**510**]{}, 229 (2014) \[arXiv:1402.6956\]. A. Gando [*et al.*]{} \[KamLAND-Zen Collaboration\], “Search for Majorana Neutrinos near the Inverted Mass Hierarchy Region with KamLAND-Zen,” Phys. Rev. Lett.  [**117**]{}, no. 8, 082503 (2016) Addendum: \[Phys. Rev. Lett.  [**117**]{}, no. 10, 109903 (2016)\] \[arXiv:1605.02889\]. M. Agostini [*et al.*]{}, “Background-free search for neutrinoless double-$\beta$ decay of $^{76}$Ge with GERDA,” Nature [**544**]{}, 47 (2017) \[arXiv:1703.00570\]. C. Alduino [*et al.*]{} \[CUORE Collaboration\], “First Results from CUORE: A Search for Lepton Number Violation via $0\nu\beta\beta$ Decay of $^{130}$Te,” Phys. Rev. Lett.  [**120**]{}, no. 13, 132501 (2018) \[arXiv:1710.07988\]. C. E. Aalseth [*et al.*]{} \[Majorana Collaboration\], “Search for Neutrinoless Double-β Decay in $^{76}$Ge with the Majorana Demonstrator,” Phys. Rev. Lett.  [**120**]{}, no. 13, 132502 (2018) \[arXiv:1710.11608\]. M. Agostini [*et al.*]{} \[GERDA Collaboration\], “Improved Limit on Neutrinoless Double-$\beta$ Decay of $^{76}$Ge from GERDA Phase II,” Phys. Rev. Lett.  [**120**]{}, no. 13, 132503 (2018) \[arXiv:1803.11100\]. M. Agostini, G. Benato and J. Detwiler, “Discovery probability of next-generation neutrinoless double- β decay experiments,” Phys. Rev. D [**96**]{}, no. 5, 053001 (2017) \[arXiv:1705.02996\]. Z. z. Xing, “Vanishing effective mass of the neutrinoless double beta decay?,” Phys. Rev. D [**68**]{}, 053002 (2003) \[hep-ph/0305195\]. Z. z. Xing, Z. h. Zhao and Y. L. Zhou, “How to interpret a discovery or null result of the $0\nu 2\beta$ decay,” Eur. Phys. J. C [**75**]{}, no. 9, 423 (2015) \[arXiv:1504.05820\]. Z. z. Xing and Z. h. Zhao, “The effective neutrino mass of neutrinoless double-beta decays: how possible to fall into a well,” Eur. Phys. J. C [**77**]{}, no. 3, 192 (2017) \[arXiv:1612.08538\]. P. F. de Salas, D. V. Forero, C. A. Ternes, M. Tortola and J. W. F. Valle, “Status of neutrino oscillations 2018: 3$\sigma$ hint for normal mass ordering and improved CP sensitivity,” Phys. Lett. B [**782**]{}, 633 (2018) \[arXiv:1708.01186\]. F. Capozzi, E. Lisi, A. Marrone and A. Palazzo, “Current unknowns in the three neutrino framework,” Prog. Part. Nucl. Phys.  [**102**]{}, 48 (2018) \[arXiv:1804.09678\]. I. Esteban, M. C. Gonzalez-Garcia, A. Hernandez-Cabezudo, M. Maltoni and T. Schwetz, “Global analysis of three-flavour neutrino oscillations: synergies and tensions in the determination of $\theta^{}_{23}, \delta^{}_{\rm CP}$, and the mass ordering,” arXiv:1811.05487. A. Caldwell, A. Merle, O. Schulz and M. Totzauer, “Global Bayesian analysis of neutrino mass data,” Phys. Rev. D [**96**]{}, no. 7, 073001 (2017) \[arXiv:1705.01945\]. G. Benato, “Effective Majorana Mass and Neutrinoless Double Beta Decay,” Eur. Phys. J. C [**75**]{}, no. 11, 563 (2015) \[arXiv:1510.01089\]. J. Zhang and S. Zhou, “Determination of neutrino mass ordering in future $^{76}$Ge-based neutrinoless double-beta decay experiments,” Phys. Rev. D [**93**]{}, no. 1, 016008 (2016) \[arXiv:1508.05472\]. S. F. Ge and M. Lindner, “Extracting Majorana properties from strong bounds on neutrinoless double beta decay,” Phys. Rev. D [**95**]{}, no. 3, 033003 (2017) \[arXiv:1608.01618\]. S. Goswami and W. Rodejohann, “Constraining mass spectra with sterile neutrinos from neutrinoless double beta decay, tritium beta decay and cosmology,” Phys. Rev. D [**73**]{}, 113003 (2006) \[hep-ph/0512234\]. S. Goswami and W. Rodejohann, “MiniBooNE results and neutrino schemes with 2 sterile neutrinos: Possible mass orderings and observables related to neutrino masses,” JHEP [**0710**]{}, 073 (2007) \[arXiv:0706.1462\]. J. Barry, W. Rodejohann and H. Zhang, “Light Sterile Neutrinos: Models and Phenomenology,” JHEP [**1107**]{}, 091 (2011) \[arXiv:1105.3911\]. Y. F. Li and S. s. Liu, “Vanishing effective mass of the neutrinoless double beta decay including light sterile neutrinos,” Phys. Lett. B [**706**]{}, 406 (2012) \[arXiv:1110.5795\]. I. Girardi, A. Meroni and S. T. Petcov, “Neutrinoless Double Beta Decay in the Presence of Light Sterile Neutrinos,” JHEP [**1311**]{}, 146 (2013) \[arXiv:1308.5802\]. P. Guzowski, L. Barnes, J. Evans, G. Karagiorgi, N. McCabe and S. Soldner-Rembold, “Combined limit on the neutrino mass from neutrinoless double-β decay and constraints on sterile Majorana neutrinos,” Phys. Rev. D [**92**]{}, no. 1, 012002 (2015) \[arXiv:1504.03600\]. C. Giunti and E. M. Zavanin, “Predictions for Neutrinoless Double-Beta Decay in the 3+1 Sterile Neutrino Scenario,” JHEP [**1507**]{}, 171 (2015) \[arXiv:1505.00978\]. S. F. Ge, W. Rodejohann and K. Zuber, “Half-life Expectations for Neutrinoless Double Beta Decay in Standard and Non-Standard Scenarios,” Phys. Rev. D [**96**]{}, no. 5, 055019 (2017) \[arXiv:1707.07904\]. J. H. Liu and S. Zhou, “Another look at the impact of an eV-mass sterile neutrino on the effective neutrino mass of neutrinoless double-beta decays,” Int. J. Mod. Phys. A [**33**]{}, no. 02, 1850014 (2018) \[arXiv:1710.10359\]. D. Silva and J. Skilling, “Data Analysis: A Bayesian Tutorial,” Oxford University Press, Oxford, 2006. S. Hannestad and T. Schwetz, “Cosmology and the neutrino mass ordering,” JCAP [**1611**]{}, no. 11, 035 (2016) \[arXiv:1606.04691\]. J. Suhonen and O. Civitarese, “Weak-interaction and nuclear-structure aspects of nuclear double beta decay,” Phys. Rept.  [**300**]{}, 123 (1998). J. Kotila and F. Iachello, “Phase space factors for double-$\beta$ decay,” Phys. Rev. C [**85**]{}, 034316 (2012) \[arXiv:1209.5722\]. M. Gerbino, M. Lattanzi, O. Mena and K. Freese, “A novel approach to quantifying the sensitivity of current and future cosmological datasets to the neutrino mass ordering through Bayesian hierarchical modeling,” Phys. Lett. B [**775**]{}, 239 (2017) \[arXiv:1611.07847\]. S. Vagnozzi, E. Giusarma, O. Mena, K. Freese, M. Gerbino, S. Ho and M. Lattanzi, “Unveiling $\nu$ secrets with cosmological data: neutrino masses and mass hierarchy,” Phys. Rev. D [**96**]{}, no. 12, 123503 (2017) \[arXiv:1701.08172\]. F. Capozzi, E. Di Valentino, E. Lisi, A. Marrone, A. Melchiorri and A. Palazzo, “Global constraints on absolute neutrino masses and their ordering,” Phys. Rev. D [**95**]{}, no. 9, 096014 (2017) \[arXiv:1703.04471\]. R. Trotta, “Bayes in the sky: Bayesian inference and model selection in cosmology,” Contemp. Phys.  [**49**]{}, 71 (2008) \[arXiv:0803.4089\]. F. Feroz and M. P. Hobson, “Multimodal nested sampling: an efficient and robust alternative to MCMC methods for astronomical data analysis,” Mon. Not. Roy. Astron. Soc.  [**384**]{}, 449 (2008) \[arXiv:0704.3704\]. F. Feroz, M. P. Hobson and M. Bridges, “MultiNest: an efficient and robust Bayesian inference tool for cosmology and particle physics,” Mon. Not. Roy. Astron. Soc.  [**398**]{}, 1601 (2009) \[arXiv:0809.3437\]. F. Feroz, M. P. Hobson, E. Cameron and A. N. Pettitt, “Importance Nested Sampling and the MultiNest Algorithm,” arXiv:1306.2144. [^1]: E-mail: [email protected] [^2]: E-mail: [email protected] [^3]: This is based on the observations with [*Planck*]{} (<http://www.esa.int/Planck>), an ESA science mission with instruments and contributions directly funded by ESA Member States, NASA, and Canada. [^4]: Note that the subscript of the Bayes factor ${\cal B}^{}_{\rm cosmo}$ herein refers to the cosmological data that have been used in the calculations, and likewise for the Bayes factors from other data sets and their combinations.

ArXiv

Introduction ============ Quantum spin models in low dimensionality are currently used to describe the magnetic properties of several materials including rare earths, organic compounds, transition metals and copper oxides. Available experimental data in these systems, especially magnetic resonance and neutron scattering, allow to establish the local structure as well as the most relevant features of the long range correlations. These studies generally show that magnetic materials can be accurately described by short range spin hamiltonians, at least in temperature regimes where the effects of disorder, the presence of spatial anisotropies or of dipolar interactions are negligible. The possibility to use simple spin hamiltonians to understand the physics of real systems has always been the drawing force for the development of more and more accurate methods to study the phase diagram of these models. Purely analytical techniques, like spin-waves expansions or mean field theories [@swt] have shown to be quite accurate when magnetic ordering is present. Numerical methods, based on series expansions [@series], Quantum Monte Carlo simulations [@simul] or Density Matrix Renormalization Group (DMRG)[@white] have been successfully applied also to finite temperatures, to frustrated models and to quasi one dimensional systems, where fluctuations play a key role in determining the physics of the model. A complimentary class of theoretical methods, which has been developed for investigating the effects of quantum and thermal fluctuations in many body systems, is known as the semiclassical approach and includes different techniques which have been used to provide a physical interpretation to experimental and simulation data [@chn; @tognetti; @gianinetti]. Semiclassical theories have been widely used in framework of quantum statistical mechanics since a long time. The mapping of quantum models, in particular spin systems, onto classical effective statistical models has been extremely useful in understanding many relevant features of phase diagrams and the possible occurrence of quantum critical points. For instance, the one dimensional Ising model in a transverse magnetic field maps to the (anisotropic) two dimensional Ising model: its exact solution therefore provides a simple way to investigate the critical properties of the Ising universality class [@sachdev]. The quantum to classical mapping can be usually justified on microscopic grounds only in the low energy and long wavelength limit, where the short range features of the original quantum model become irrelevant. Therefore, strictly speaking, the use of semiclassical methods in quantum statistical mechanics is restricted to regions characterized by a diverging correlation length and gapless low energy excitations. However, the microscopic quantum hamiltonian is often assumed to be faithfully represented by its classical counterpart in a large portion of the phase diagram provided the coupling constants present in the classical lagrangian are suitably [*renormalized*]{} due to quantum (short range) fluctuations [@sachdev]. This expectation has been beautifully confirmed by the study of Heisenberg models in low dimensionality. In particular, exact solutions and numerical analysis of one dimensional antiferromagnets found exponentially decaying correlations and gapped excitation spectrum in integer spin chains. Conversely, semi-integer spin chains turned out to be gapless with power-law correlations. This picture fully agrees with the conjecture by Haldane, based on the presence of a topological term in the semiclassical action of one dimensional spin chains [@haldane]. Moreover, quantitative analysis of the three dimensional Non Linear Sigma Model (NL$\sigma$M) found a phase diagram which compares favorably with experiments on two dimensional antiferromagnets [@chn; @beard; @troyer] suggesting that semiclassical approaches can be directly used for the interpretation of experimental data. The semiclassical mapping for Heisenberg antiferromagnets has been also employed to study the effects of an external magnetic field on the model. Of particular physical relevance is the role of a staggered field, which directly couples to the order parameter: Such a staggered field can be realized in certain quasi-one dimensional spin one compounds where, by lowering the temperature, rare earths magnetic ions undergo a Néel transition to a three dimensional antiferromagnet which generates an alternating magnetic field acting on the Ni$^{2+}$ chains [@maslov]. A similar mechanism has been also invoked for the interpretation of quasi one dimensional spin one-half chains [@oshikawa]. However, the effective classical action used in the literature has not been explicitly derived from the quantum hamiltonian and we believe it should be reconsidered. In order to extract quantitative information from this NL$\sigma$M a “Single Mode Approximation" (SMA) is usually adopted [@maslov] but a detailed numerical study of its accuracy in this case is still missing, even if analytical studies [@morandi] suggest that it may be justified only for the transverse channel. In this paper, we present a microscopic derivation of the semiclassical action. We obtain an effective low energy theory different from the one commonly adopted in the literature. This theory is then analyzed in the weak and strong field limit and the results are compared to Lanczos diagonalizations. Other magnetic materials where the use of semiclassical methods may be suggestive are spin-orbital systems, like C$_{60}$ compounds [@auerbach] or transitional metal oxides [@pati], in which orbital degeneracy is present. The special points where the model has an enlarged SU(4) symmetry are particularly important in order to understand the phase diagram of the model [@santoro; @li; @itoi] and deserve a detailed analysis. We therefore apply the semiclassical mapping to the two microscopic models presenting this symmetry. We show that in one case the model maps straightforwardly to the CP$^3$ NL$\sigma$M (with topological term in the one dimensional limit) while in the other case, we have been able to carry out the mapping only in the special case of one dimension, where again we find the same semiclassical action at a different effective coupling. Interestingly, both lattice hamiltonians have been exactly solved in D=1 [@martins; @sutherland] with very different results: the first model has broken symmetry and gapped excitations, while in the second case it is gapless and critical. We believe that this mapping provides an important clue to the understanding the phase diagram of CP$^n$ models, which are shown to undergo a phase transition as a function of the coupling constant. Semiclassical approach ====================== In this Section we briefly review and generalize a method, proposed few years ago [@ap], for the derivation of low energy effective actions in bipartite spin systems. The main advantage of this technique, with respect to the original procedure developed by Haldane [@haldane], is the possibility to keep track of lattice effects and to make direct contact with other useful microscopic approaches, like spin wave theory. The method will be later applied to Heisenberg models in an external field and to spin-orbital systems. Be $H$ the hamiltonian of a spin model on a bipartite lattice. For instance, the celebrated Heisenberg antiferromagnet is described by $${\cal H} =\sum_{R\in B}\sum_{\delta} {\bf S}_R \cdot {\bf S}_{R+\delta} \label{ham1}$$ where ${\bf S}$ are spin operatrs, the site index $R$ runs over the sublattice labeled by $B$ and $\delta$ is a primitive vector on the lattice connecting nearest neighbor sites. In order to evaluate the partition function $Z={\rm Tr} \exp(-\beta {\cal H})$ we adopt the usual coherent states formalism: we first split the interval $(0,\beta)$ in a large number $N$ of Trotter (time) slices, then insert at each imaginary time $t$ a resolution of unity based on coherent states defined at each lattice site $R$. Here we follow the standard $O(3)$ notation $|{\bf\Omega}(R,t)>$: In spin $S$ models the coherent states are labeled by the unit vector ${\bf\Omega}(R,t)$ and are characterized by the requirement that $<{\bf\Omega}|\,{\bf S}\,|{\bf\Omega}>=S {\bf\Omega}$. These states may be explicitly obtained by a suitable rotation in spin space of the highest eigenvector of the $S_z$ operator. The partition function can be therefore written as: $$\begin{aligned} Z&=&\int {\cal D} {\bf \Omega}(R,t) \,e^{-S_{eff}} \nonumber \\ S_{eff}&=&\int_0^\beta dt \Big [ S^2\sum_{R\in B} \sum_{\delta} {\bf \Omega}(R,t) \cdot {\bf \Omega}(R+\delta,t) - \nonumber \\ &&\sum_R <{\bf\Omega}(R,t)|\dot{\bf\Omega}(R,t)> \Big ] \label{action}\end{aligned}$$ Up to this point, the underlying lattice structure is fully present in the functional form (\[action\]) and no approximation has been introduced. However, the imaginary Wess-Zumino term $$<{\bf\Omega}|\dot{\bf\Omega}>=iS\,{({\bf\Omega}\times\dot{\bf\Omega})_z\over 1+\Omega_z}$$ prevents a simple classical interpretation of the effective action. In order to obtain a mapping onto a physically transparent classical statistical model, it is convenient to exploit the bipartite nature of the lattice by explicitly tracing out the degrees of freedom defined on sublattice $B$. This procedure can be performed analytically because the variables we are integrating out are coupled only to the classical external fields ${\bf\Omega}(R,t)$ defined on the other sublattice (sublattice $A$). Therefore this step just requires the solution of the [*single site*]{} time dependent problem: $$\begin{aligned} e^{-F[{\bf B}(R,t)]}={\rm Tr} \,U_\beta \nonumber\\ {dU_t\over dt} = - {\bf K}(R,t) \,U_t \nonumber \\ {\bf K}(R,t)={\bf S}\cdot{\bf B}(R,t) \nonumber \\ {\bf B}(R,t)=S \sum_\delta {\bf\Omega}(R+\delta,t) \label{onsite}\end{aligned}$$ In terms of the local free energy functional $F[{\bf B}(R,t)]$, the effective action becomes: $$S_{eff}= \sum_{R\in B} F[{\bf B}(R,t)] - \int_0^\beta dt\sum_{R\in A} <{\bf\Omega}(R,t)|\dot{\bf\Omega}(R,t)> \label{act}$$ Notice that now the effective action just depends on the field ${\bf\Omega}(R,t)$ on sublattice $A$. The solution of the problem defined in Eq. (\[onsite\]) can be obtained within perturbation theory for [*slowly varying*]{} field configurations. This requirement is satisfied in the (semiclassical) large $S$ limit and represents the only, [*low energy*]{} approximation we need to introduce in the evaluation of the effective action. We notice that the lattice structure is not involved in this semiclassical analysis and the [*long wavelength*]{} approximation is actually unnecessary within our method. Explicitly, we obtain the following free energy functional: $$\begin{aligned} F[{\bf B}(t)]&=& \int_0^\beta dt \left\{\epsilon_0(t) + \Gamma_{00}(t) - \int_0^t dt^\prime \Gamma_{01}(t)\Gamma_{10}(t^\prime)\right \} \nonumber\\ \Gamma_{ij}(t)&=&<u_i(t)|\dot u_j(t)>\exp\left \{ \int_0^t dt^\prime \left [ \epsilon_i(t^\prime)-\epsilon_j(t^\prime)\right ] \right \} \label{free}\end{aligned}$$ where $\epsilon_i(t)$ is the $i^{th}$ instantaneous eigenvalue of ${\bf K}(R,t)$ and $|u_i(t)>$ is the corresponding eigenstate. In order to obtain this expression we have assumed that the ground state of ${\bf K}(R,t)$ $|u_0(t)>$ is non degenerate at every time $t$. The terms shown in Eq. (\[free\]) are correct up to second order in time derivatives, as usual in semiclassical approximation. The long wavelength limit of the Berry phase term $\Gamma_{00}$ in $F[{\bf B}(t)]$ has been shown in Ref. [@ap] to exactly compensate the Wess-Zumino contribution in the effective action (\[act\]). As a result, the low energy [*and*]{} long wavelength effective action, becomes a non-linear sigma model with, possibly, a topological $\theta$ term coming from the residual space dependence of the Berry phase $\Gamma_{00}$. Now we are in the position to apply this method to specific lattice hamiltonians: The procedure we have just outlined requires the explicit solution of the instantaneous eigenvalue problem defined by the on site hamiltonian ${\bf K}(R,t)$ (\[onsite\]), the evaluation of the terms $\Gamma_{ij}(t)$ in Eq. (\[free\]) and finally the substitution of the results into the form of the action (\[act\]). Heisenberg model in an external field ===================================== It is now established that the ground state properties and the low lying excitation spectrum of antiferromagnetic Heisenberg chains strongly depend on the value of the on site magnetic moment. Semi-integer spin chains have power law correlations and gapless spectrum while, for integer spin, correlations decay exponentially and a gap to all excitations is present [@review] ($\Delta=0.41048$ for $S=1$). This different behavior was first found by Haldane[@haldane] via a semi-classical mapping onto the $O(3)$ non-linear $\sigma$-model. This conjecture was later confirmed by numerical studies based on Lanczos diagonalizations [@parkinson], DMRG [@white] and MC simulations [@night]. Remarkably, these idealized models have also an experimental counterpart: neutron scattering experiment on quasi-one dimensional materials such as [*${\rm Ni(C_2H_8N_2)_2NO_2(ClO_4)}$*]{} (NENP), confirmed[@neutron] that the essential physics of these systems is well described by a simple $S=1$ Heisenberg hamiltonian that couples neighboring spins antiferromagnetically and takes account of the single-ion anisotropy by including an on site term $D\,(S_i^z)^2$. The predicted Haldane gap has been measured with great accuracy in these experiments showing good agreement with theoretical and numerical[@sorensen; @lanczos; @mc] predictions. The effects of a magnetic external perturbation on an antiferromagnetic chain, frequently gives rise to unexpected interesting phenomena. This is the case of Cu benzoate, a quasi one dimensional $S=1/2$ antiferromagnet displaying a gapped excitation spectrum in an applied uniform field. Oshikawa and Affleck [@oshikawa] interpreted the experimental findings on the basis of an Heisenberg hamiltonian where spins are coupled to a weak effective staggered field. This microscopic model gives a field dependence for the gap which agrees with experimental data. More recently, interest has been focussed on the study of the effects of external fields on $S=1$ systems. In the case of a uniform magnetic field, the lowest triplet excitation states are split into a transverse and a longitudinal mode and the gap closes at a critical field $H_c$, where Bose condensation of magnons takes place[@s1]. The synthesis of compounds of the form ${\rm R_2BaNiO_5}$ (where R stands for a magnetic rare earth) allowed to study the effects of a [*staggered*]{} magnetic field on the quasi-one dimensional chain of spin one Ni$^{2+}$ ions. The magnetic moment of the Ni$^{2+}$ couples with the R$^{3+}$ ions that are ordered antiferromagnetically below a Néel temperature $T_N$ (typically $16 K\lesssim T_N \lesssim 80 K$). This three dimensional antiferromagnetic matrix generates an effective staggered magnetic field on the Ni$^{2+}$ chains whose intensity can be tuned by varying the temperature below $T_N$. In this way, experiments have been able to investigate the effects of a staggered field on the Haldane gap, the staggered magnetization and the susceptibility [@zhel; @ray]. Stimulated by these experiments, few analytical and numerical studies attempted a theoretical analysis of spin chains in staggered fields by use of semiclassical mappings [@maslov; @morandi] and DMRG [@yulu]. While qualitative agreement can be easily attained, some discrepancy still remains between the NL$\sigma$M approach and DMRG findings, noticeably on the form of correlation functions. This circumstance is rather surprising in light of the very nice agreement between the NL$\sigma$M predictions and numerical data for the pure Heisenberg chain [@sorensen]. The strong field limit is particularly simple because, for every $S$, the ground state can be accurately described by a Néel state with gaussian transverse fluctuations. For such a problem, spin wave theory (SWT) can be applied also in one dimension giving a transverse spectrum of the form: $$\epsilon_k=S\,\sqrt{(|{\bf H}|/S+2D)^2-4\gamma_k^2} \label{swtd}$$ where ${\bf H}$ is the staggered field, $D$ is the space dimensionality and $\gamma_k=\sum_{i=1}^D \cos k_i$. To leading order, the transverse dynamical correlations in imaginary time predicted by SWT have single mode character: $$S_{\perp}(k,\omega)=S^2{|{\bf H}|/S+2(D -\gamma_k) \over \omega^2+\epsilon_k^2}$$ The longitudinal correlations may be expressed as convolutions of the transverse ones, implying that no longitudinal branch of elementary excitations is present. According to the SWT approach, in the strong field limit, the longitudinal gap saturates at twice the transverse one. Much more subtle is the weak field case, where quantum fluctuations strongly contrast the onset of a magnetically ordered state. In order to understand this limit, we re-examine the derivation of the NL$\sigma$M for a spin $S$ chain in an external staggered field on the basis of the method sketched in Section II. The microscopic hamiltonian we consider here is just the antiferromagnetic Heisenberg model in an external field ${\bf H}$ that can be taken either uniform or staggered: $${\cal H} =\sum_{R \in B} {\bf S}_R \cdot \big[\sum_\delta{\bf S}_{R+\delta}+ {\bf H}\big]\pm{\bf H}\cdot\sum_{R\in A}{\bf S}_R \label{hamiltonian}$$ where the conventions are the same as in (\[ham1\]), and upper (lower) sign refers to uniform (staggered) applied field. Following the derivation of the previous Section, we factorize the problem in the two sublattices ($A$ and $B$) and write the effective action as: $$\begin{aligned} S_{eff}=-\int_0^\beta dt &&\sum_{R\in A} \bigg\{<{\bf\Omega}(R,t)| \dot{\bf\Omega}(R,t)> \mp S{\bf H}\cdot{\bf\Omega}(R,t)\bigg\} \nonumber\\ &&+\sum_{R\in B} F[{\bf B}(R,t)] \label{partstag} \end{aligned}$$ where now $${\bf B}(R,t)=S\sum_\delta{\bf\Omega} (R+\delta,t)+{\bf H}. \label{field}$$ Therefore, we formally reduce to the same one-body problem as stated in Eq. (\[onsite\]). This can be solved perturbatively in the case of zero temperature ($\beta\to\infty$), and slowly varying effective fields ${\bf B}$. The resulting functional $F[{\bf B}]$ is then correct up to second order in space and time derivatives. In this context we stress that this perturbative treatment is justified only in the low-energy limit which can be physically accessed in gapless systems. In one dimension such a requirement is satisfied in half-integer spin chains while, for integer spins, it holds only in the large $S$ limit. In fact, a perturbative renormalization group analysis, in zero external field, predicts the exponential dependence [@polyakov] $\Delta\sim \exp{-(\pi S)}$. This shows that the existence of the gap is a purely quantum effect: it vanishes on approaching the [*classical*]{} limit ($S\to\infty$) where it is still justified to derive the effective action perturbatively also for integer spin systems. Interestingly, in the [*strong*]{} field limit, the one body problem can be again easily solved by considering small oscillations of the vector ${\bf \Omega}$ about the direction of the magnetic field. The resulting effective action, to quadratic order in the amplitude of the oscillations, exactly reproduces all the lowest order results of SWT, including lattice effects. Specializing Eq.(\[free\]) to the form of the effective field (\[field\]) we obtain $$F[{\bf B}(R,t)]=\int_0^\beta dt\left\{ S\frac{ |{\bf \dot m}(R,t)|^2}{2\,|{\bf B}(R,t)|}-S|{\bf B}(R,t)|+ \Gamma_{00}[{\bf m}]\right\} \label{sefstag}$$ where ${\bf m}=\frac{ {\bf B}}{|{\bf B}|}$. The first term comes from the time integration of Eq.(\[free\]) and the second one is the ground state eigenvalue $\epsilon_0(t)$. Now we perform the continuum limit of the expression (\[sefstag\]) assuming that the relevant configurations ${\bf\Omega}(R,t)$ are slowly varying functions of space on the scale set by the lattice spacing $a$. To lowest order in spatial fluctuations we have ${\bf B}(R,t)=S\,q\,{\bf\Omega} (R_0,t)+{\bf H}$ where $R_0=R-\hat x$ is a reference site belonging to sublattice $A$ ($\hat x$ is the primitive vector pointing in the $x$ direction) and $q=2D$ is the number of nearest neighbors of a hypercubic lattice in dimension $D$. In the weak field limit, we just need to keep terms up to second order in $\delta {\bf B}(R,t)={\bf H}+S\sum_\delta[{\bf\Omega}(R+\delta,t)-{\bf\Omega}(R_0,t)]$. By expanding ${\bf m}(R,t)$ we obtain: $$\begin{aligned} \delta{\bf m}&(&R,t)={\bf m}(R,t)-{\bf\Omega}(R_0,t)\nonumber\\ &&=\frac{\delta {\bf B}(R,t)}{S\,q}- \frac{{\bf\Omega}(R_0,t)\cdot\delta {\bf B}(R,t)}{S\,q} +O(\delta B^2)\end{aligned}$$ This leads to an approximation of the Berry phase that, to lowest order cancels the Wess-Zumino term in the effective action leaving a residual contribution: $$\begin{aligned} \Gamma_{00}[{\bf m}] &-& \Gamma_{00}[{\bf \Omega}(R_0,t)] \nonumber\\ &\simeq&iS\delta {\bf m}(R,t)\cdot\big({\bf m}(R,t)\times{\bf\dot m}(R,t) \big) \nonumber\\ &\simeq&\frac{i}{q} {\bf H}\cdot\big({\bf \Omega}(R_0,t) \times{\bf \dot\Omega}(R_0,t)\big) \label{gammaoo}\end{aligned}$$ In one dimension, the usual topological term, arising from the spatial derivative of ${\bf\Omega}$ present in $\delta{\bf B}$, also appears besides the contributions shown in Eq. (\[gammaoo\]). By taking the continuum limit, we finally get the form of an effective action in which one half of the degree of freedom have been integrated out: $$\begin{aligned} &S_{eff}&=\int dt \int {dR\over 2a^D} \bigg [\frac{1}{2q}\left(\dot{\bf \Omega}(R,t)+ i\,{\bf H}\times{\bf\Omega}(R,t)\right )^2 + \nonumber\\ && a^2S^2 |\nabla_R {\bf \Omega}|^2 -(1 \mp 1) S{\bf H}\cdot{\bf \Omega}(R,t)\bigg ] +2\pi i S \,Q \label{acteffstag}\end{aligned}$$ The topological charge $Q$ is non trivial only in $D=1$ where: $$Q={1\over 4\pi} \int\,dx\,dt\,{\bf \Omega}\cdot(\partial_x{\bf \Omega} \times{\bf \dot\Omega})$$ As a result, we obtain an effective NL$\sigma$M describing spin $S$ chains in (weak) uniform ($-$) or staggered ($+$) field. The microscopic derivation allows to obtain explicit expressions for the [*bare*]{} spin wave velocity $c=2Sa\sqrt{D}$ and stiffness $\rho_s=S^2a^{2-D}$ which coincide with those already known at ${\bf H}=0$. While this derivation reproduces known results for a uniform field [@fisher], it differs from the effective action usually quoted in the literature, where the external field only couples to ${\bf \Omega}$ via the Zeeman term. It is instructive to give a simple interpretation to the formal result we have obtained: a staggered field ${\bf H}$ can be written as the sum of a uniform field of the same strength minus a field twice as strong acting only on one sublattice (say the sublattice $A$). In this way, by tracing out the degrees of freedom living on sublattice $B$ we obtain the same NL$\sigma$M action appropriate for a uniform field with the addition of a Zeeman term $-2{\bf H}\cdot {\bf \Omega}$: This is exactly what we formally found in Eq. (\[acteffstag\]). In order to investigate the low energy spectrum of the NL$\sigma$M previously obtained (\[acteffstag\]), we resort to a simple [*single mode approximation*]{}: the constraint ${\bf \Omega}^2=1$ may be lifted through the introduction of a Lagrange multiplier $\lambda({\bf H})$ and a linear shift of the field ${\bf\Omega}$. In this way, the dynamical correlation functions in imaginary time acquire a Lorenzian form: $$\begin{aligned} <S^z_{k,\omega}S^z_{-k,-\omega}>&=&\frac{cgS^2/a}{\omega^2+2cg\lambda+c^2k^2} \label{smsa}\\ <S^{\pm}_{k,\omega} S^{\mp}_{-k,-\omega}> &=&\frac{2cgS^2/a}{(\omega^2-{\bf H}^2)\mp 2i|{\bf H}|\omega+2cg\lambda+c^2k^2} \nonumber\end{aligned}$$ where $\lambda$ may be determined by a saddle point equation, or by fitting numerical data and $g=ca^{D-1}/\rho_s$. Here the $z$ axis identifies the direction of the external field. The poles of the correlation functions directly give the dispersion relation of the model. Recalling that the NL$\sigma$M describes the spin degrees of freedom on a single sublattice, the wave vector $k$ is defined modulo $\pi$. From expressions (\[smsa\]) it is apparent that the correlation functions have single mode character with different dispersion relations in the transverse and longitudinal channel: In particular, at each $k$ (modulo $\pi$) the transverse excitation splits into two different branches centered around the longitudinal dispersion: $$\begin{aligned} \Delta_L(k)&=&\sqrt{\Delta({\bf H})^2+c^2k^2}\nonumber\\ \Delta_T(k)&=&\Delta_L(k)\pm |\bf{H}| \label{spectra}\end{aligned}$$ This form of the energy spectrum should apply to arbitrary spin $S$ at low energy and weak staggered fields. It definitely differs from the predictions of the usual semiclassical treatments which lead, within the same single mode approximation, to spectra of the form $\sqrt{\Delta({\bf H})^2+c^2k^2}$ in both channels. Interestingly, also the exact solution of the 1D $S=1/2$ $XY$ model in a staggered field along the $z$ axis shows[@alcaraz] a similar excitation spectrum, providing some support to our semiclassical analysis. The isotropic $S=1/2$ Heisenberg model in a staggered field has been also analyzed by bosonization techniques and conformal field theory [@oshikawa; @alcaraz]. The staggered field has been shown to open a gap both in the transverse and in the longitudinal channel. This massive triplet is degenerate to leading order in the external field: $\Delta_T=\Delta_L\propto H^{2/3}$. It would be interesting to study the spliting of the gap by including subleading terms in order to compare conformal field theory results with the NL$\sigma$M approach. In order to ascertain the validity of the single mode approximation in one dimensional models, we performed Lanczos diagonalizations in $S=1$ chains. In particular we calculated the excitation spectrum by selecting, for each $k$, the excitation energy with the highest weight in the Lehmann representation of the dynamical correlation function. In Fig. \[wei\] we show the largest normalized spectral weight $Z(k)$ as a function of the momentum $k$ of the excitation, showing that its value is always very close to one for $k\sim \pi$: this implies that the sum rule ($\sum_n Z_n(k)=1$) is almost exhausted by a single excitation both in the transverse [*and*]{} in the longitudinal channel, thereby supporting the single mode approximation usually adopted. Notice that the excitation with largest weight does not always have the lowest energy, at fixed $k$. Lanczos diagonalizations also show that the matrix element giving the spectral weight of the longitudinal excitation at $k=\pi$ decreases very quickly when the staggered field is switched on: It is lowered by a factor $2.5$ when the field reaches the value $H=0.05$. This result should be compared with the modest decrease of the transverse spectral weight, which reduces just by $20\%$ in the same range. Both findings agree with the reported behavior of neutron scattering data [@ray]. In Fig. \[spe\] the energy spectra are shown for different field strength. While at low ${\bf H}$ the spectrum is markedly asymmetrical around $k=\pi/2$, symmetry is restored at larger fields where it closely approaches the form (\[swtd\]) predicted by SWT. Data comes from Lanczos diagonalizations performed on chains with 12, 14, 16 and 18 lattice sites. A specific feature of our results is the splitting of two branches in the transverse excitation spectrum when a weak staggered field is applied (\[spectra\]). Fig. \[gap\] shows Lanczos diagonalizations data on several $S=1$ chains which provide a numerical confirmation of the semiclassical predictions. For every wavevector $k$ we selected the longitudinal and the transverse excitation with the largest spectral weight $Z$. When the continuum limit is appropriate, i.e. at sufficiently small $k$ (modulo $\pi$), two distinct excitation branches clearly appear, differing by $\pm |{\bf H}|$ from the longitudinal excitation, in agreement with the NL$\sigma$M analysis. The noticeable deviations around $k=\pi/2$ are clearly due to lattice effects which are not correctly reproduced in the continuum limit. Notice that finite size corrections do not seem to affect the overall structure of the excitation spectrum of the model. In Ref.[@yulu] the DMRG technique has been applied to investigate the lowest excitations in the longitudinal and transverse channels at $k=\pi$. The comparison of our results with DMRG is particularly instructive because it clearly shows the regimes where the use of the effective action approach is justified. In fig. \[dmrg\] we plot the quantity $\gamma=(\Delta_L-\Delta_T)/H$ as a function of $ H $. The numerical results identify two different behaviors: at strong fields ($H > 0.5$) $\gamma$ saturates at $\gamma=2$, in agreement with SWT, while in the low field limit ($H < 0.1$) the splitting between the longitudinal and transverse gap quickly increases. In particular, the value $\gamma=1$ predicted by our semiclassical approach is compatible with the numerical data in the $H\to 0$ limit, although extended calculations in lower fields are required in order to validate our analysis. Spin-orbital models =================== Spin orbital models have recently attracted considerable interest in the attempt to explain the unusual magnetic properties of a class of quasi one dimensional materials, which includes C$_{60}$ compounds (eg. TDAE-C$_{60}$)[@auerbach] and few metal oxides (eg. Na$_2$Ti$_2$SbO$_2$, NaV$_2$O$_5$)[@pati]. The physical properties of these Mott insulators are largely determined by the coupling between orbital and spin degrees of freedom which may be dominated either by Hund’s rule or by dynamical Jahn-Teller effect. Possible realizations in higher dimensions are also found in fullerides [@santoro] or in LiNiO$_2$ [@li]. The low energy physics of these systems may be described by keeping only spin and orbital degrees of freedom. If the orbital degeneracy is twofold, like in the previous examples, the low energy model can be written in terms of two sets of spin-$1/2$ operators per lattice site representing respectively spin (S) and orbital (T) degrees of freedom. Usually, spin isotropy in this effective low energy hamiltonian is retained only in the physical spin variable S while terms which break rotational invariance in the pseudo-spins T are generally allowed. However, the fully isotropic hamiltonian: $${\cal H} =J\,\sum_{<i,j>} \left [ {\bf S}_i\cdot {\bf S}_j + {\bf T}_i\cdot {\bf T}_j\right ] + K \sum_{<i,j>} {\bf S}_i\cdot {\bf S}_j {\bf T}_i\cdot {\bf T}_j \label{st}$$ has been the subject of several studies, particularly in the two special cases $K=\pm 4J$. The $K=+4J$ hamiltonian can be written in terms of permutators on each lattice site and may be relevant for TDAE-C$_{60}$ while the $K=-4J$ naturally arises as the strong coupling limit of a microscopic hamiltonian appropriate when dynamical Jahn-Teller effect prevails. These two particular models have several remarkable properties: besides an obvious SU(2)$\times$SU(2) symmetry, they are both invariant by a larger SU(4) symmetry group [@santoro]. The 15 generators of the symmetry group include the total spin and pseudospin operators: $\sum_R S^\alpha_R$ and $\sum_R T^\alpha_R$ and the further 9 operators $\sum_R (\pm)^R S^\alpha_RT^\beta_R$ where the $\pm$ sign corresponds to the two models $K=\pm 4J$. Note that, in the $-$ case, the SU(4) generators do not commute with the translations by one lattice spacing although the hamiltonian does not break any symmetry of the lattice. Remarkably, this model is also non frustrated: In a valence bond basis the ground state can be shown to have positive semi-definite weights. This feature allows to perform very accurate Monte Carlo simulations on this system[@santoro2]. Both models can be exactly solved in one dimensions [@martins; @sutherland] with very different physical properties: the $K=+4J$ model ([*Sutherland model*]{}) is gapless, with power law spin correlations whose leading asymptotic behavior has a $x^{-3/2}$ decay and is characterized by oscillations with period equal to four lattice spacing [@azaria]. Instead, when $K=-4J$ ([*Valence Bond model*]{}) the system spontaneously dimerizes, the energy spectrum is gapped and correlations decay exponentially [@martins; @santoro]. Several relevant features of the ground states of these hamiltonians have been argued to be applicable to wider regions in parameter space, also outside the special SU(4) points [@itoi]. Here we will derive the effective low energy lagrangian for both models in order to understand how such a different physical behavior in one dimension may originate and to shed light on the phase diagram of the two models in higher dimensions. In fact, analytical and numerical studies in two dimensions have suggested that SU(4) symmetry is not spontaneously broken in the ground state and a spin liquid phase may emerge[@santoro2; @li]. The general method developed in Section II can be straightforwardly applied also to this class of hamiltonians defined on bipartite lattice because the interaction just couples nearest neighbor sites. Let us discuss the two cases separately. Valence Bond model ($K=-4J$) ---------------------------- The model has four orthogonal states per site in the lattice, which correspond to the four possibilities $(\pm{1\over 2},\pm{1\over 2})$ for the $z$-components of the spin and pseudospin variables. A set of coherent states is therefore labeled by a quartet of complex numbers at each site $R$: $z_\alpha(R)$ obeying the normalization conditions $\sum_\alpha |z_\alpha(R)|^2=1$. We follow the convention to indicate the amplitude of the $|\uparrow,\uparrow>$ state by $z_1$, of $|\uparrow,\downarrow>$ by $z_2$, of $|\downarrow,\uparrow>$ by $z_3$ and of $|\downarrow,\downarrow>$ by $z_4$. Using this representation, the partition function is written as $$\begin{aligned} S_{eff}&=&\int_0^\beta dt \left [ \sum_{R} \,z^*_\alpha(R,t)\dot z_\alpha(R,t) +<z(t)|H|z(t)> \right ] \nonumber \\ Z&=&\int {\cal D} \, z_\alpha(R,t) \,e^{-S_{eff}} \label{seff}\end{aligned}$$ in close analogy with Eq. (\[action\]). Summation over repeated labels $\alpha$ is understood. Again, by tracing out sublattice B, we reduce to an effective action defined only on sublattice A, formally given by: $$S_{eff}=\int_0^\beta dt \sum_{R\in A} \,z^*_\alpha(R,t)\dot z_\alpha(R,t) +\sum_{R\in B} F[{\bf K}(R,t)] \label{actionsu4}$$ The functional $ F[{\bf K}(R,t)]$ is defined by the single site problem in an external field: $$\begin{aligned} e^{-F[{\bf K}(R,t)]}={\rm Tr} \,U_\beta \nonumber\\ {dU_t\over dt} = - {\bf K}(R,t) \,U_t \label{onsitesu4}\end{aligned}$$ where ${\bf K}(R,t)$ is a $4\times 4 $ matrix whose components $K_{\mu\nu}$ depend on the classical field $z_\alpha(R^\prime,t)$ defined at nearest neighbor sites. The explicit form of the matrix ${\bf K}(R,t)$ depends on the couplings of the hamiltonian, and for the case we are examining is given by $$K_{\mu\nu}(R,t)=-J \sum_{\delta} \zeta_\mu^*(R+\delta,t) \zeta_\nu (R+\delta,t) \label{kkk}$$ where, according to our conventions, the field $\zeta_\alpha$ is simply related to the semiclassical variable $z_\alpha$ by: $\zeta_1=z_4$, $\zeta_2=-z_3$, $\zeta_3=-z_2$ and $\zeta_4=z_1$. To lowest order in the spatial derivatives, i.e. taking $z_\alpha(R,t)\sim z_\alpha(R+\delta,t)$, the instantaneous ground state of ${\bf K}(R,t)$ is non degenerate and has components $\zeta^*_\alpha(R,t)$ with eigenvalue $\epsilon_0=-2DJ$ where $D$ is the spatial dimension and we specialized to hypercubic lattices. The three other eigenvalues of ${\bf K}(R,t)$ vanish and the corresponding eigenvectors are any three four dimensional vectors orthogonal to the ground state. Therefore the procedure outlined in Section II can be straightforwardly applied and requires the evaluation of $\epsilon_0$ to second order in the lattice spacing $a$ (i.e. to second order in the spatial derivatives) which can be obtained by standard second order perturbation theory: $$\epsilon_0=-2DJ + a^2 J \, {\bf \nabla} (z_\mu^* z_\nu) \cdot {\bf \nabla} (z_\mu z_\nu^*) \label{eigst}$$ Moreover the coefficient $\Gamma_{00}$ must be evaluated to first order in $a$. The lowest order just cancels the Wess-Zumino term, while the contribution linear in $a$ is non vanishing (for smooth configurations) only in $D=1$, where it gives rise to a residual topological term: $$\Gamma_{00}=z_\alpha \dot z^*_\alpha + a\,\partial_x (z_\alpha\dot z_\alpha^*) +O(a^2) \label{topost}$$ Finally, the terms $\Gamma_{0j}(t)$ may be evaluated to lowest order in $a$. For slowly varying fields $z_\alpha(t)$, only the $t^\prime\sim t$ region does contribute to the integral in Eq. (\[free\]) leading to: $$\int_0^t dt^\prime \sum_{j\ne 0} \Gamma_{0j}(t)\Gamma_{j0}(t^\prime)= -{1\over 2DJ} \left \{ \dot z^*_\alpha \dot z_\alpha - z_\mu\dot z_\mu^* z^*_\nu\dot z_\nu \right \} \label{f2st}$$ Combining the results (\[eigst\],\[topost\],\[f2st\]) we obtain the required long wavelength limit of the effective semiclassical action for the Valence Bond SU(4) model: $$S_{eff} ={1\over 2g} \int dR \int dt \sum_{i=1}^{D+1}\sum_{\mu,\nu} |\partial_i z_\mu^* z_\nu |^2 + i\pi Q \label{cp3}$$ The label $i$ runs over the $D+1$ space-time coordinates and the continuum limit has been taken in the $D$ spatial directions while the imaginary time variable has been suitably rescaled. The coupling constant $g$ is explicitly given by $$g=a^{D-1}\sqrt{4D} \label{gst}$$ and the topological charge $Q$ is present only in $D=1$ where it reads: $$Q = {1\over 2\pi i } \int dx\, dt\, \partial_x (z_\alpha\dot z^*_\alpha) \label{topo}$$ Tracing out a sublattice is an efficient way to take into account the short range antiferromagnetic correlations present in the model leading to an effective action describing the much smoother fluctuations on a single sublattice. In order to support this interpretation of the procedure we have adopted, Fig.(\[sq\]) shows the magnetic structure factor obtained by Lanczos diagonalization on a $16$-site lattice: The sharp peak at wavevector $k=\pi$ confirms that the most relevant correlations have indeed periodicity of two lattice spacing. The formal construction of the effective action for the Valence Bond SU(4) model shows that the long wavelength and low energy physics is described by a $D+1$ dimensional NL$\sigma$M or CP$^3$ model (with topological angle $\theta=\pi$ in $D=1$). The known exact solution of the lattice model in one dimension implies that the CP$^3$ model at $g=2$ has correlations exponentially decaying in space-time. Moreover, the spin model is known to develop dimer order in the thermodynamic limit, which implies breaking of translational invariance by a lattice spacing together with breaking of parity. In going to the continuum limit, one sublattice has been traced out and then the CP$^3$ model remains translationally invariant but parity breaking should still occur. This picture is supported by Lanczos diagonalizations of the spin model showing that the quantum numbers of the two SU(4) singlet states which collapse in the thermodynamic limit correspond to momenta $P=0$ and $P=\pi$ and opposite parity (i.e. reflections through a lattice site). Clearly, the mapping we have developed neglects cut-off effects and holds only for sufficiently smooth configurations of the classical field. Therefore, the resulting estimate of the bare coupling constant $g=2$ should be taken with caution but we are confident that the qualitative behavior of the CP$^3$ model does indeed capture the physics of this lattice spin system, analogously to the familiar SU(2) case. The Valence Bond model we have considered belongs to the SU(n) class already studied by Affleck [@affleck] and Read and Sachdev [@read] by use of 1/n expansion. These analysis show that the ground state breaks parity and translational symmetry and may be described by a Valence Bond Solid (VBS) in one dimension, at least for sufficiently large n. The exact solution of the Valence Bond model [@martins] confirms that this picture holds down to n=4. The two dimensional case is more difficult: in the n$\to \infty$ limit the system has infinite degeneracy and can be represented as an arbitrary covering of the lattice by nearest neighbor valence bonds. This degeneracy is lifted at leading order in 1/n giving rise to a (plaquette) resonating valence bond solid [@santoro]. In fact, the model maps onto a dimer hamiltonian which has been studied by Monte Carlo techniques [@runge]. However, diagonalizations and Quantum Monte Carlo simulations directly performed on the hamiltonian (\[st\]) [@santoro2] suggest that the ground state might be a spin liquid at n=4 and therefore argue in favor of a phase transition between magnetically disordered phases as a function of n. Sutherland model ($K=+4J$) -------------------------- The derivation of the NL$\sigma$M appropriate for the second SU(4) point of the spin hamiltonian (\[st\]) can be carried out following the same procedure adopted before and leading to Eqs. (\[seff\]) and (\[actionsu4\]). However, the associated quantum mechanical problem (\[onsitesu4\]) is now defined by the slightly different $4\times 4$ matrix $$K_{\mu\nu}(R,t)=J \sum_{\delta} z_\mu(R+\delta,t) z^*_\nu (R+\delta,t) \label{kkk2}$$ instead of Eq. (\[kkk\]). To leading order in spatial fluctuations, we can set $z_\mu(R+\delta)\sim z_\mu(R)$. In this case, the (instantaneous) ground state of the one site problem defined by the matrix (\[kkk2\]) is threefold degenerate. As a result, we cannot carry on straightforwardly the trace over one sublattice, suggesting that this procedure is not able to eliminate the long range oscillations in spin correlations. In turn, this means that the continuum limit cannot be taken just by considering the spin configurations on a single sublattice, as in usual antiferromagnets and larger primitive cells must be taken into account. A confirmation of such an interpretation comes from Lanczos diagonalizations on this model: as shown in Fig. \[sq\], in one dimension the spin correlations displays oscillations characterized by the wavevector $k=\pi/2$, implying a four site periodicity. This suggests a generalization of the procedure sketched in Section II: instead of tracing out one sublattice, we now keep one site every four sites of the chain. Therefore we need the solution of a three site problem, defined by the hamiltonian (\[st\]), with “time" dependent boundary conditions defined by the classical field $z_\alpha(R,t)$ on the two adjacent sites. Equation (\[actionsu4\]) is basically unchanged, but now the sublattice $A$ includes only one fourth of the sites of the chain and the free energy functional $F$ is defined by Eq. (\[onsitesu4\]) in terms of a matrix ${\bf K}$ acting on a Hilbert space of dimension $64$. To lowest order in spatial fluctuations of the classical field $z_\alpha(R,t)$, the full spectrum of ${\bf K}$ can be explicitly obtained by solving the eigenvalue equation: $$J\left [z_\mu z^*_\alpha\ u_{\alpha,\nu,\lambda} +u_{\nu,\mu,\lambda} +u_{\mu,\lambda,\nu}+z_\lambda z^*_\alpha u_{\mu,\nu,\alpha}\right ] =\epsilon\,u_{\mu,\nu,\lambda} \label{3site}$$ In particular, the ground state wavefunction of the quantum problem is now non-degenerate and reads $$u_{\mu,\nu,\lambda}={1\over \sqrt{6}} z^*_\alpha\epsilon_{\alpha\mu\nu\lambda} \label{3gs}$$ where $\epsilon_{\alpha\mu\nu\lambda}$ is the fully antisymmetric Levi-Civita tensor. The corresponding eigenvalue is simply $\epsilon_0=-2J$, while the Berry phase contribution gives $$\Gamma_{00}={1\over 6} z_\alpha \epsilon_{\alpha\mu\nu\lambda} \dot z^*_\beta \epsilon_{\beta\mu\nu\lambda}= z_\alpha\dot z^*_\alpha$$ and cancels the Wess-Zumino term in the effective action. In order to obtain a non-trivial theory we have to include long wavelength fluctuations of the classical field $z(R,t)$. This can be done by use of perturbation theory in the associated quantum three site problem. The solution is sketched in Appendix where the required terms are explicitly evaluated. Here we just quote the final result, after having taken the continuum limit: The spin fluctuations on the sublattice are described by a NL$\sigma$M which coincides with the one obtained for the Valence Bond model (\[cp3\]) at a bare coupling $$g={6\over \sqrt{5}} \label{gc}$$ larger than the estimate $g=2$ obtained in one dimension for the Valence Bond model (\[gst\]). Also for this hamiltonian the topological angle $\theta$ is given by $\theta=\pi$. As previously stressed, the Sutherland model is critical in one dimension, with power law correlations characterized by oscillations of period $4a$ corresponding to a characteristic wavevector $k=\pi/2$, as also confirmed by the Lanczos diagonalization results shown in Fig. \[sq\]. If the mapping onto a CP$^3$ theory (with topological angle $\theta=\pi$) faithfully describes the long wavelength physics of the lattice model, we conclude that the CP$^3$ action in $1+1$ dimension has correlations $$<z^*_\alpha(0,0) z_\beta(0,0) z^*_\beta(x,t)z_\alpha(x,t)> \sim (x^2+t^2)^{-3/4}$$ On the other hand, the exact solution of the Valence Bond hamiltonian together with the semiclassical mapping of Section IV-A implies that the CP$^3$ model has exponentially decaying correlations at $g=2$. These two results may be reconciled if we assume that the CP$^3$ theory in $1+1$ dimension undergoes a second order phase transition as a function of the coupling $g$ and the Sutherland model describes the physics at the critical point. Another possibility would be the occurrence of a gapless [*phase*]{} in the model, like in the celebrated CP$^1$ case. This alternative explanation, however, conflicts with available analytical and numerical evidence pointing towards a massive regime at strong coupling [@cp]. A phase transition in the CP$^3$ model cannot be related to a spontaneous breaking of the continuous SU(4) symmetry while a possibility is the occurrence of parity breaking in one of the two phases. Of course the estimate (\[gc\]) we have obtained for the critical coupling will be renormalized by finite cut-off effects but the overall picture of the phase diagram emerging from the semiclassical analysis should be robust. This result may have implications in the framework of the strong CP problem in field theory. [@cp]. Conclusions =========== In this paper we discussed some application of the semiclassical approach to spin systems in low dimensionality. This technique is known to capture the qualitative features of quantum models and to provide a useful framework for the interpretation of experimental and numerical data. For Heisenberg chains in a staggered field we pointed out some difficulty and inconsistency of usual treatments, especially in the weak field limit, where the structure of the effective NL$\sigma$M turns out to be richer than expected. The general form of the excitation spectrum predicted by semiclassical approaches has been confirmed by use of Lanczos diagonalizations in finite clusters which also provide some support to the usually adopted SMA. Our results are fully compatible with existing DMRG data and show that different physical regimes occur in the phase diagram of this model. A recent DMRG investigation of a $S=2$ spin chain also pointed out a similar behavior [@capone]. The new NL$\sigma$M we derived is expected to represent only the [*low field*]{} region while, for moderate/high staggered fields, simple perturbative approaches, like spin wave theory, are fully adequate to describe the excitation spectrum of the model. It seems unlikely that a single effective action in the continuum limit might be able to encompass these two very different physical behaviors. A semiclassical analysis of spin-orbital models characterized by a SU(4) symmetry has also been performed in two different regimes. In one case we straightforwardly obtained, in the long wavelength limit, a mapping to the CP$^3$ NL$\sigma$M describing the fluctuations of spin-orbitals degrees of freedom on the same sublattice. Recent numerical simulations argued in favor of a disordered ground state in such a model for $D=2$, while in one dimension the exact solution of the lattice hamiltonian proved that spontaneous dimerization occurs. In the other SU(4) model we have examined, a smooth continuum limit requires a coarse graining over several lattice sites, suggesting that the relevant fluctuations are characterized by a wavevector $k$ different from the antiferromagnetic one. We explicitly performed the analysis only for the one dimensional model, where we again obtained the same CP$^3$ model which now describes fluctuations about $k=\pi/2$. The phase diagram of such a NL$\sigma$M has been extensively studied, particularly in $1+1$ dimension, in connection to the CP problem in field theory. Several proposals have been put forward in the literature, including spontaneous parity breaking and deconfining transition. The long wavelength mapping between spin chains and two dimensional field theories may provide a clue for the final understanding of the phase diagram of the CP$^n$ model, analogously to what has been found for the Wess Zumino Witten model. We gratefully acknowledge useful correspondence with A.Pelissetto and G. Morandi. We also thank Yu Lu and Jizhong Lou for providing DMRG data. Appendix ======== In this appendix we briefly illustrate the procedure adopted for the analytical solution of the generalized three site problem $$J\left [z_\mu z^*_\alpha\ u_{\alpha,\nu,\lambda} +u_{\nu,\mu,\lambda} +u_{\mu,\lambda,\nu}+\bar z_\lambda \bar z^*_\alpha u_{\mu,\nu,\alpha}\right ] =\epsilon\,u_{\mu,\nu,\lambda}$$ up to second order in the lattice spacing $a$. Here $\bar z_\alpha= z_\alpha + 4a \partial_x z_\alpha +8a^2\partial_x^2 z_\alpha + O(a^3)$. We employ standard second order perturbation theory which gives $$\Delta\epsilon=<u_0|\Delta {\bf K} |u_0> + \sum_{n\ne 0} { |<u_0|\Delta {\bf K} |u_n>|^2\over \epsilon_0-\epsilon_n} \label{pert}$$ The unperturbed eigenvectors $|u_n>$ are the solutions of the eigenvalue equation (\[3site\]) and $\epsilon_n$ are the corresponding eigenvalues. The ground state $|u_0>$ is explicitly given in Eq.(\[3gs\]) and $\epsilon_0=-2J$. Due to the manifest SU(4) invariance of the eigenvalue equation, the external field $z_\alpha$ may be chosen to point in the “1" direction without loss of generality: $z_\alpha=\delta_{\alpha,1}$. The lowest order term gives: $$<u_0|\Delta {\bf K} |u_0>= {16\over 3}a^2J \left [ \,\vert (z^*_\alpha\partial_x z_\alpha) \vert^2 - \partial_x z_\alpha \partial_x z^*_\alpha\,\right ]$$ while the sum over excited states which appears at second order requires the evaluation of the matrix element $$<u_0|\Delta {\bf K} |u_n>= -{4aJ\over \sqrt{6}} \epsilon_{1\alpha\mu\nu} u_{\mu\nu 1} \partial_x z_\alpha$$ for a generic excited state $u_{\mu\nu 1}$. The only solutions of the unperturbed eigenvalue equation (\[3site\]) which give a non vanishing contribution are those corresponding to the eigenvalue $\epsilon=\pm\sqrt{2}\,J$ and to $\epsilon=0$. The former states are given by $u_{123}= u_{231}=-u_{132}=-u_{321}=\pm u_{312}/\sqrt{2}=\mp u_{213}/\sqrt{2}=1/\sqrt{8}$. The latter states are $u_{123}=u_{321}=-u_{132}=-u_{231}=1/2$ Both excited states are three times degenerate: the other states being obtained by cyclic permutations of the labels $(234)$. Inserting these results into the perturbative expansion (\[pert\]) we get: $$\Delta\epsilon=-{4\over 3} J\,a^2 \left [ \,\vert (z^*_\alpha\partial_x z_\alpha) \vert^2 - \partial_x z_\alpha \partial_x z^*_\alpha\,\right ]$$ The last required step is the evaluation of the Berry phase term $\Gamma_{00}$ to first order in the lattice spacing. Again, by use of perturbation theory, we get $$\Gamma_{00}=<u_0|\dot u_0>+2i\,{\rm Im} \sum_{n\ne 0} { <u_0|\Delta {\bf K} |u_n>\over \epsilon_0-\epsilon_n} <u_n|\dot u_0>$$ The intermediate states which contribute to the sum are those corresponding to $\epsilon=\pm\sqrt{2}\, J$ which give: $$\Gamma_{00}=z_\alpha\dot z^*_\alpha -2a\,\partial_x(\dot z^*_\alpha z_\alpha)$$ This, in turn, gives the well known topological term quoted in the text (\[topo\]). [99]{} P.W. Anderson, Phys. Rev. [**86**]{}, 694 (1952). R.R.P. Singh, , 12337 (1993). A.W. Sandvik, , 5196 (1998); L. Capriotti and S. Sorella 3173 (2000). S.R. White , 2863 (1992). S.Chakravarty, B.I.Halperin, and D.R.Nelson, , 1057 (1988); S.Chakravarty, B.I.Halperin, and D.R.Nelson, , 2344 (1989). A.Cuccoli, V.Tognetti, R.Vaia, P.Verrucchi, , 3439 (1996). P. Gianinetti and A. Parola 104414 (2001). See for instance S. Sachdev [*Quantum phase transitions*]{}, Cambridge University Press (Cambridge 1999). F.D.M. Haldane, Phys. Lett. [**93 A**]{}, 464 (1983). B.B.Beard, R.J.Birgeneau, M.Greven, and U.J.Wiese ,1742 (1998). J.K. Kim and M. Troyer, , 2705 (1998). S. Maslov and A. Zheludev, 68 (1998); S. Maslov and A. Zheludev, , 5786 (1998). M. Oshikawa and I. Affleck , 2883 (1997). E. Ercolessi, G. Morandi, P. Pieri, M. Roncaglia , 14860 (2000). D.P. Arovas and A. Auerbach , 10114 (1995). S. Pati, R.R.P. Singh, D.I. Khomskii , 5406, (1998). G. Santoro, L. Guidoni, A. Parola, E. Tosatti , 16168 (1997). Y.Q. Li, M. Ma, D.N. Shi, F.C. Zhang, , 3527 (1998). C. Itoi, S. Qin, I. Affleck , 6747 (2000). M.J. Martins and B. Nienhuis , 4956 (2000). B. Sutherland , 3795 (1975). A. Parola , 7109 (1989). I. Affleck, J. Phys.: Condens. Matter, [**1**]{} 3047 (1989). J.B. Parkinson and J.C. Bonner, 4703 (1985). M.P. Nightingale and H.W.J. Blöte, , 659 (1986). P. Renard, M. Verdaguer, L.P. Regnault, W.A.C. Erkelens, J. Rossat-Mignod and W.G. Stirling, Europhys. Lett. [**3**]{}, 945 (1987). E.S. Sørensen, I. Affleck, , 15771 (1994). S. Haas, J. Riera and E. Dagotto, , 3281 (1993). M. Takahashi, , 311 (1993). I. Affleck, , 6697 (1990). A. Zheludev, J.P. Hill, D.J. Buttrey , 7216 (1996); A. Zheludev et al., , 3630 (1998). S. Raymond, T. Yokoo, A. Zheludev, S.E. Nagler, A. Wildes and J. Akimitsu, 2382 (1999). J. Lou, X. Dai, S. Qin, Z. Su, L. Yu, , 52 (1999). A.M. Polyakov, Phys. Lett. B [**59**]{}, 121 (1975). D.S. Fisher , 11783 (1989). F.C. Alcaraz and A.L. Malvezzi J.Phys. A: Math. Gen. [**28**]{}, 1521, (1995). G. Santoro, S. Sorella, L. Guidoni, A. Parola, E. Tosatti , 3065 (1999). P. Azaria, A.O. Gogolin, P. Lecheminant, A.A. Nersesyan , 624 (1999). I. Affleck , 966 (1985). N. Read and S. Sachdev, Nucl. Phys. [**B316**]{}, 609 (1989); , 4568 (1990). P.W. Leung, K.C. Chiu, K.J Runge , 12938 (1996). N. Seiberg , 637 (1984); M. Asorey and F. Falceto, , 234 (1998); J.C. Plefka and S. Samuel , 3966 (1997). M. Capone and S. Caprara, cond-mat/0107282, (2001).

ArXiv

--- author: - Jens Oliver Gutsfeld - 'Markus Müller-Olm' - Christian Dielitz bibliography: - 'sections/conclusion/paper.bib' title: Temporal Logics with Language Parameters ---

ArXiv

--- abstract: 'We present a comparison between approximated methods for the construction of mock catalogs based on the halo-bias mapping technique. To this end, we use as reference a high resolution $N$-body simulation of 3840$^3$ dark matter particles on a 400$h^{-1}\rm{Mpc}$ cube box from the Multidark suite. In particular, we explore parametric versus non-parametric bias mapping approaches and compare them at reproducing the halo distribution in terms of the two and three point statistics down to $\sim 10^8\,{\rm M}_{\odot}\,h^{-1}$ halo masses. Our findings demonstrate that the parametric approach remains inaccurate even including complex deterministic and stochastic components. On the contrary, the non-parametric one is indistinguishable from the reference $N$-body calculation in the power-spectrum beyond $k=1\,h\,{\rm Mpc}^{-1}$, and in the bispectrum for typical configurations relevant to baryon acoustic oscillation analysis. We conclude, that approaches which extract the full bias information from $N$-body simulations in a non-parametric fashion are ready for the analysis of the new generation of large scale structure surveys.' author: - - 'Raúl E. Angulo$^{1,4}$, Gustavo Yepes$^{5,6}$, Chia-Hsun Chuang$^{7}$, Guillermo Reyes-Peraza$^8$,' - | Mathieu Autefage$^9$, Mohammadjavad Vakili$^{10}$ & Cheng Zhao$^{11}$\ \ $^{1}$Donostia International Physics Centre (DIPC), Paseo Manuel de Lardizabal 4, 20018 Donostia-San Sebastian, Spain\ $^{2}$Instituto de Astrofísica de Canarias, s/n, E-38205, La Laguna, Tenerife, Spain\ $^{3}$Departamento de Astrofísica, Universidad de La Laguna, E-38206, La Laguna, Tenerife, Spain\ $^{4}$IKERBASQUE, Basque Foundation for Science, 48013, Bilbao, Spain.\ $^{5}$Departamento de Física Teórica, Universidad Autónoma de Madrid, Cantoblanco E-28049, Madrid, Spain\ $^{6}$Centro de Investigación Avanzada en Física Fundamental, Facultad de Ciencias, Universidad Autónoma de Madrid, E-28049 Madrid, Spain\ $^{7}$Kavli Institute for Particle Astrophysics and Cosmology, Stanford University, 452 Lomita Mall, Stanford, CA 94305, USA\ $^{8}$Instituto de Física Teórica, (UAM/CSIC), Universidad Autónoma de Madrid, Cantoblanco, E-28049 Madrid, Spain\ $^{9}$Télécom Physique Strasbourg Pôle API, 300 Bd Sébastien Brant, 67400 Illkirch-Graffenstaden, France\ $^{10}$Leiden Observatory, Leiden University, Niels Bohrweg 2, NL-2333 CA Leiden, The Netherlands\ $^{11}$École polytechnique fédérale de Lausanne, Route Cantonale, 1015 Lausanne, Switzerland bibliography: - 'refs.bib' title: 'The bias of dark matter tracers: assessing the accuracy of mapping techniques ' --- \[firstpage\] cosmology: – theory - large-scale structure of Universe Introduction {#sect1} ============ With the starting-gun for some of the most impressive observational campaigns in large scale structure about to be shot (e.g. Euclid, @Euclid, and DESI, @DESI), the quest for precise and accurate covariance matrices, aimed at assessing the uncertainties in the measurements of cosmological observables such as redshift space distortions [e.g. @1987MNRAS.227....1K] and baryonic acoustic oscillations [e.g. @1998ApJ...496..605E] has become a task of front line research [see e.g. @2013PhRvD..88f3537D; @2013MNRAS.432.1928T; @2015MNRAS.454.4326P]. The construction of large sets of catalogs based on $N$-body simulations has been adopted as the standard path to obtain robust estimates on the errors of cosmological observables. Such approach is nevertheless unpractical, given the considerable amount of time and/or memory requirements that a state-of-the-art $N$-body simulation requires to generate hundreds to thousands of realizations with the desired cosmological volumes and mass resolution. Several approaches have been suggested in the literature to speed-up the construction mock catalogs [e.g. @1996ApJS..103....1B; @2002MNRAS.329..629S; @2002MNRAS.331..587M; @2013JCAP...06..036T; @2016MNRAS.459.2118K; @2016MNRAS.459.2327I; @2016MNRAS.463.2273F; @2003ApJ...593....1B; @2005ApJ...630....1Z; @2013MNRAS.428.1036M; @2015MNRAS.447..437M; @2014MNRAS.437.2594W; @2015MNRAS.450.1856A; @2015MNRAS.446.2621C; @Angulo:2013gya]. The idea in some of these approaches (such as [`PATCHY` ]{}@2014MNRAS.439L..21K) is to rely on the smooth large-scale dark matter field obtained from approximate gravity solvers, and populate it with halos (or galaxies) following some bias prescriptions, in what is known as the *bias mapping technique*. The precision of this type of approach in the resulting halo catalogs is only acceptable, according to scientific requirements of forthcoming surveys, until intermediate scales ($k\sim 0.2-0.3\, h\,{\rm{Mpc}}^{-1}$ in Fourier space), [see e.g. @2016MNRAS.456.4156K; @10.1093/mnras/stz507; @10.1093/mnras/sty2964; @10.1093/mnras/sty2757]. Higher precision ($\sim 1-2\%$) towards smaller scales beyond $k=1\, h\,{\rm{Mpc}}^{-1}$ (including the Nyquist frequency) has been only reached by the recently published *Bias Assignment Method* (`BAM` hereafter, @2019arXiv190606109B [@2019MNRAS.483L..58B]). {width="5.8cm"} {width="5.8cm"} {width="5.8cm"}\ {width="5.8cm"} {width="5.8cm"} {width="5.8cm"} The goal of this paper is to assess whether bias mapping methods such as [`PATCHY` ]{}and `BAM`, already successfully applied to mass-scales of the order of $\sim 10^{12}\, {\rm M}_{\odot}h^{-1}\,$ (see e.g. @2017MNRAS.472.4144V and @2019MNRAS.483L..58B), are still capable to capture the main features of halo bias when applied to a high resolution (low mass-scales) $N$-body simulation. The [`PATCHY` ]{}method has been shown to reach high level of accuracy circumventing the limitations of the approximate gravity solvers [see e.g. @2015MNRAS.452..686C; @2015MNRAS.450.1836K; @2017MNRAS.472.4144V], being able to make large amount of precise mock galaxy catalogues [e.g. @2016MNRAS.456.4156K] with very low memory and computational requirements, while `BAM` has been shown to replicate to percent precision summary statistics of halo catalogs (with approximately same computational time and memory requirements as [`PATCHY` ]{}). It is therefore timely to assess whether these methods can be still applied to the expected galaxy catalogs from Euclid or DESI, probing lower halo mass ranges, or if a change towards non-parametric approaches forecasts better results. In order to perform this comparison, we use the [Small MultiDark Planck]{} simulation[^1] (`SMDPL` hereafter), which belongs to the series of [MultiDark]{} simulations [@Klypin:2014kpa] with Planck cosmology [@planck2015-xiii]. The simulation consists in a set of $3840^3$ particles dark matter particles in a comoving volume of $V_{s}=(400$ $h^{-1}$Mpc$)^{3}$. Dark matter halos and subhalos are identified with the `Rockstar` algorithm [@2013ApJ...762..109B], with a minimum tracer mass of $\sim 2\times10^8 {\rm M}_{\odot} h^{-1}$ at $z=0$ ($\sim 8.2\times 10^{7}$ number of halos and subhalos). Although in what follows we use the complete sample of halos and subhalos as tracers of the dark matter field, we will refer to them generically as “halos” or “tracers”. Note that this is the first time that both techniques are tested against a sample of halos and subhalos. Both the `PATCHY` method and `BAM` are based on the concept of stochastic bias [e.g. @1999ApJ...520...24D] in order to generate a halo density field from an approximated dark matter density field (ADMF hereafter). This bias is expressed as a conditional probability distribution (CPD hereafter), accounting for the probability of obtaining a given number of halos conditional to local and non local properties of the ADMF. In [`PATCHY` ]{}the CPD is characterized by a mean and a scatter around that mean, both expressed through a number of parameters describing a given model for the mean and stochasticity. In `BAM` instead, the CPD is *directly measured* from the reference simulation in combination with the ADMF. For both methods, the ADMF is obtained by the evolution of downgraded initial conditions from the `SMDPL` simulations. In our current set-up, we performed this operation starting with a white noise embedded in a $3840^3$ mesh into one of $160^3$ grid cells. These downgraded initial conditions were thereafter evolved until redshift $z=0$ using approximated gravity solvers such as `FastPM` [@2016MNRAS.463.2273F] or the Augmented Lagrangian Perturbation Theory (ALPT) [@Kitaura:2012tj] on an equally resolved mesh ($160^3$). While for the `BAM` approach the choice of gravity solver is not relevant, that is not the case for [`PATCHY` ]{}, in which the signal of the power spectrum at intermediate scales can affect the performance of the method. We therefore adopt `FastPM` on a mesh of $160^3$ cells to evolve the downgraded initial conditions up to redshift $z=0$. Although not specifically shown in this study, a different resolution of $192^3$ was used and similar results were achieved. The outline of this paper is as follows. We describe the parametric bias prescription in § \[sec:patchy\], and show the results of the fits of such prescription to the `SMDPL` simulation. We show the results from the non-parametric approach in § \[sec:bam\]. We discuss possible signatures of non local bias in § \[sec:nl\] and end-up with conclusions and discussion. Parametric bias prescription: [`PATCHY` ]{} {#sec:patchy} =========================================== In [`PATCHY` ]{}[@2014MNRAS.439L..21K], the deterministic component (or mean of the CPD) relating the halo ($\rho_{\rm{h}}$) and the dark matter ($\rho_{\rm{m}}$) density fields is written in terms of the matter overdensity $\delta \equiv \rho/\bar{\rho}-1$ as [see e.g. @Neyrinck:2013ezr; @2014MNRAS.439L..21K] $$\begin{aligned} \label{eq:deterministic_nonlocal} \langle (1+\delta_{\rm h}) \rangle & \equiv & \langle (1+\delta_{\rm h})| (1+\delta_{\rm m}) \rangle \\ & = & \big((1+\delta_{\rm m})^\alpha + c_{\rm s^{2}}s^{2}\big) \, \theta\big(\delta_{\rm m} - \delta_{\rm th}\big) \, \exp \big[-\big((1+\delta_{\rm m})/\rho_{\epsilon}\big)^{\epsilon}\big]\,. \nonumber \end{aligned}$$ In this expression $\delta_{\rm th}$ is an overdensity threshold, aimed at suppressing the formation of halos in under-dense regions. The threshold bias [e.g. @Kaiser:1984sw; @Bardeen:1985tr; @Sheth:1999su] is represented by a step function, $\theta \big(\delta_{\rm m} - \delta_{\rm th}\big)$, and an exponential cutoff, $\exp \big[-\big((1+\delta)/\rho_{\epsilon}\big)^{\epsilon}\big]$. On the other hand, the term $(1+\delta_{\rm m})^\alpha$ accounts for nonlinear behaviour. In Eq. (\[eq:deterministic\_nonlocal\]) we have introduced a second order non-local term proportional to $s^{2}=s_{ij}s^{ij}$ [see e.g. @2009JCAP...08..020M; @2018PhR...733....1D; @Abidi:2018eyd], where $s_{ij}\equiv \partial_{ij}\Phi-\delta^{\rm K}_{ij}\delta_{\rm m}/3$ ($\delta^{\rm K}_{ij}$ stands for the Kronecker delta) are the elements of the traceless tidal field tensor [see e.g. @2007MNRAS.375..489H], written in terms of the gravitational potential $\Phi$ obtained from the Poisson equation $\nabla^{2}\Phi=\delta_{\rm m}$. The stochasticity around this deterministic bias is modelled by a negative binomial distribution (@2014MNRAS.439L..21K [@2017MNRAS.472.4144V]), with expected number of halos in each cell (given by the deterministic bias component) $\lambda_{\rm h} \equiv \langle N_{\rm{h}}\rangle_{{\rm d} V} = f_N \langle 1+\delta_{\rm h} \rangle_{{\rm d}V} {\rm d}V$ and an scatter characterized by a Poisson distribution with a parameter $\beta$ allowing for deviations from the latter. Here, $f_N \equiv \bar{N}/\langle 1+\delta_h \rangle_V$ controls the halo number density (i.e. it is fixed by the number density of the reference catalog). Hence, the probability of having $N_{\rm h}$ objects in a volume element with mean $\lambda_{h}$ is given by $$P(N_{\rm h}|\lambda_{\rm h}, \beta) = \left(\frac{\lambda_{\rm h}^{N_{\rm h}}}{N_{\rm h}!}\, \right) \left(\frac{\Gamma(\beta+N_{\rm h})}{\Gamma(\beta)(\beta + \lambda_{\rm h})^{N_{\rm h}}}\right) \left(1+\frac{\lambda_{\rm h}}{\beta}\right)^{-\beta}, \label{eq:devpois}$$ where $\Gamma(\beta)$ is the Gamma function [^2]. We further test the addition of an extra dependency over $\lambda_{\rm h}$ by substituting $\beta \rightarrow \beta (\lambda_{\rm h}/\bar{N})^\gamma$, where $\bar{N}$ is the mean number of halo in the volume. [`PATCHY` ]{}uses Eq. (\[eq:devpois\]) to sample a halo density field (number counts) according the dark matter density (local) and the tidal field (non local), constrained to generate a sample with the mean number density of the reference simulation. From this sampled halo density field we can measure the summary statistics (i.e, one, two point statistics etc). We have constrained the set of parameters defining the bias models (Eqs. (\[eq:deterministic\_nonlocal\]) and (\[eq:devpois\])), $\{\delta_{\rm th}, \alpha, \rho_{\epsilon}, \epsilon, \beta, \gamma, c_{s^{2}}\}$ following the procedures of @2015MNRAS.450.1836K and @2017MNRAS.472.4144V, based on Markov-Chain likelihood analysis of the parameter space. We used the probability density function (PDF) and the power spectrum (up to $k=0.4h/{\rm{Mpc}}$) as observables, and assumed independent likelihoods with a Poisson variance for the PDF ($\sigma^2_{\rm PDF}=N_n$, where $N_n$ is the number of cells containing $n$ dark matter halos), and a Gaussian variance for the power spectrum ($\sigma^2_P = 4\pi^2P_{\rm ref}(k)/(V_{s}k^2\Delta k )$, where $P_{\rm ref}(k)$ denotes the power spectrum measured from the reference simulation). We summarise the constraints on the bias parameters in Table \[table:BestFitMock\]. {width="18cm" height="7cm"} {width="18cm" height="6.5cm"} --------------------------- ---------------------- -------------------------- $ \alpha $ $ 1.37 \pm 0.02 $ $ 0.72 \pm 0.03 $ $ \rho_{\epsilon} $ $ 1.94 \pm 0.08 $ $ 0.14 \pm 0.06 $ $ \epsilon $ $ 0.4 \pm 0.2 $ $ -1.4 \pm 0.5 $ $ \beta $ $ 5.9 \pm 0.2 $ $ 8.0 \pm 0.2 $ $ \gamma $ $ 0.38 \pm 0.09 $ $ 0.04^{+0.09}_{-0.04}$ $ c_{s^2} $ $ \rm{-} $ $(5 \pm 3)\times 10^{-4} $ --------------------------- ---------------------- -------------------------- : Mean values with $68\%$ errors of the parameters characterizing the model of halo-bias in Eqs. (\[eq:deterministic\_nonlocal\]) and (\[eq:devpois\]) in [`PATCHY` ]{}. These values are obtained from the Markov-Chain Monte Carlo analysis using as observables the PDF and the power spectrum. The threshold density is fully consistent with the value $\delta_{\rm th}\sim -1$.[]{data-label="table:BestFitMock"} The resulting PDF and power spectrum are shown in top panel of Fig. \[Pk\_patchy\_OldBias\]. The PDF is not well fitted by any of the cases under study. [`PATCHY` ]{}without non-local terms is able to fit the `SMDPL` number of cells for those with small number of halos but fails when increasing the number of halos per cell. If the non-local term is included and we still force a correct fit for the $P(k)$, the PDF is correctly fitted for cells with large number of halos but fails in the low number end. Focusing now on the $P(k)$, the local model reaches an accuracy of $\sim5\%$ up to scales of $\sim 0.55\,h\,{\rm{Mpc}}^{-1}$, while the non-local extension remains within the same accuracy up to scales of $\sim 0.65\, h\,{\rm{Mpc}}^{-1}$. Note again that for this result we used the complete halo sample of $M>2\times10^8 {\rm M}_{\odot} h^{-1}$ (which roughly corresponds to a linear bias of $\sim 0.65$). Although not explicitly shown in Fig. \[Pk\_patchy\_OldBias\], the inclusion of the parameter $\gamma$ contributes to a marginal improvement in accuracy of the PDF and $P(k)$ when only accounting for local terms (thus its non-zero value of Table \[table:BestFitMock\]). Nonetheless, when accounting for the non-local term $s^2$, the best-fit value of $\gamma$ is consistent with zero (see Table \[table:BestFitMock\]). The value of $\epsilon$ changes sign when including non-local terms, changing the behaviour of the exponential cut-off. Furthermore, when including non-local terms, a larger $\beta$ is favoured indicating closer resemblance with Poissonity. The value of $\delta_{\rm{th}}$ is consistent with $-1$ indicating that, for the current mass-limit, the step function of Eq. (\[eq:deterministic\_nonlocal\]) does not suppress the formation of halos in under-dense regions. Therefore, the full dark matter simulation is preferred over density cuts over it. In any case, we stress that neither the one- nor the two-point statistics are reproduced to the desired per-cent precision by the model with best-fit values, indicating that such approach cannot account for a realistic description of the halo-bias at such low mass scales. The third panel of Fig. \[Pk\_patchy\_OldBias\] shows the signal of the bispectrum obtained from sampling the ADMF with Eq. (\[eq:deterministic\_nonlocal\]), using the best fit parameters. The estimates are biased with respect to the reference, regardless the presence of the non-local terms in the bias model. In order to provide an estimate of the error-bars for the bispectrum, we use the expression [see e.g. @fry_article; @1998ApJ...496..586S; @Scoccimarro:2000sn; @2015JCAP...10..039A] $$\label{err_bis} \sigma_B^2 \approx s_B \frac{k_f}{V_B}P(k_1)P(k_2)P(k_3) \, ,$$ where $s_B = 6, 2, 1$ for equilateral, isosceles and scalene configurations in Fourier space, $V_B\approx 8\pi^2k_1k_2k_3\Delta k$ and the fundamental mode is $k_f = 2\pi/L_{\rm{box}}$. We include this variance to have a taste of how much is the [`PATCHY` ]{}methodology biased in terms of the cosmic variance. Using this estimates for the variance, we quantify the difference between the bispectrum generated by [`PATCHY` ]{}and the reference to $\sim70\%$ ($\sim 10\sigma_B$). We therefore conclude that the model of Eq. (\[eq:deterministic\_nonlocal\]) cannot describe the higher order statistics of a halo sample with the mass resolution of the `SMDPL`. In the next section, when mentioning statistical significance, we will be referring to this expression, though will not explicitly present the error bars in order to avoid clutter. ![Examples of halo-bias, expressed as the probability distribution $P(y|x)$ of halo overdensity $y\equiv \log (1+\delta_{h})$ in different bins of ADM densities (expressed as $x\equiv \log (1+\delta_{\rm m})$) and for the full samples (top-row) and different cosmic-web types (second to fourth row). In all cases, the discrepancies between the reference and [`PATCHY` ]{}are mildly evident on the low dark matter density bins. For high values of dark matter density, [`PATCHY` ]{}displays a significant lack of cells with low values of halo overdensities, as can be visually inferred from Fig.\[densf\]. Knots (voids and filaments) are not found on low (high) density regions shown in this plot. Note that the bias from `BAM` and the reference are statistically indistinguishable.[]{data-label="bias_pdf"}](images/bam_patchy_vs_bam_pdf_cwc.pdf){width="13cm"} Non parametric approach: the Bias Assignment Method {#sec:bam} =================================================== [`BAM` ]{}[@2019arXiv190606109B; @2019MNRAS.483L..58B] is an example of a non-parametric approach to the mapping of halo bias [for other non parametric approaches based on machine learning techniques see also @2017arXiv170602390M; @2018arXiv180804728M; @2018arXiv181106533H; @McClintock:2019sfj]. The method statistically maps the halo bias (i.e, the CPD) measured from the reference simulation and the ADMF (a `FastPM` generated density field in this case) into the latter, producing a halo sample with the PDF of the reference. This is preformed within an iterative process, in which the ADMF is convolved with a kernel computed from the ratio between the halo power spectrum of the reference simulation and the power spectrum of the halo field produced in the previous iteration. The [`BAM` ]{}kernel is subject to a Metropolis-Hasting selection criteria, which guarantees that within each iteration the power spectrum of the new halo number counts approaches that of the reference. Indeed, after some iterations (for the current set-up, typically $\sim 10$), the halo field produced by [`BAM` ]{}is such that its power spectrum differs $\sim 1\%$ with respect to that from the reference[^3]. In [`BAM` ]{}the halo-bias is measured as a function of a set of properties of the underlying ADMF such as the local density. We also use *non-local contributions* such as the *cosmic web-type* (i.e, knots, filaments, sheets and voids), classified using the eigenvalues of the tidal field [see e.g. @2007MNRAS.375..489H]. Another non-local quantity used in [`BAM` ]{}is the mass of collapsing regions. These regions are defined as sets of friend-of-friends cells (classified as knots) [e.g. @2015MNRAS.451.4266Z]. The bottom panels from Fig. \[Pk\_patchy\_OldBias\] show summary statistics obtained from the non-parametric method. In the first and second column we show the PDF and power spectrum, which, as exposed before, display percent differences by construction (regardless whether non-local properties are also included). More relevant is the information shown in the third, where we display the reduced bispectrum measured at the configuration $k_1=0.1\,h\,{\rm{Mpc}^{-1}}$ and $k_2=0.2\,h\,{\rm{Mpc}}^{-1}$. It is impressive how `BAM` reproduces (at the resolution of the `SMDPL`) the signal of the bispectrum of the reference, specially in the case in which non-local information is included in the iterative process. Given that the bispectrum is not calibrated within the procedure, such good agreement between the signals shown in Fig.\[Pk\_patchy\_OldBias\] can be regarded as a direct consequence of both the high signal-to-noise in the measurement of the CPD and the inclusion of relevant non-local dependencies such as the cosmic-web classification. Note that in this regard, @2019arXiv190606109B showed that the agreement seen in Fig. \[Pk\_patchy\_OldBias\] was not present when lower mass resolutions or larger cosmological volumes are used as a reference (using the same *non-local* properties). We shall return to this result in § \[sec:nl\]. In Fig. \[densf\] we show a slice through the halo density field as obtained from the reference and the two methods under comparison. On large scales the two methods provide a fair description of the halo number density field (as can be inferred from the measurements of power spectrum in Fig. \[Pk\_patchy\_OldBias\]). The evident differences are observed on small scales, where the [`PATCHY` ]{}method tends to break structures such as filaments and sheets and mildly populate voids, all this at the expense of overpopulating high density regions (e.g. knots). In order to quantify this, in Fig. \[biasF\] we show the halo-bias in the form of contours of the joint probability distribution, $P(x,y)$ with $x\equiv \log (1+\delta_{\rm m})$ and $y\equiv \log (1+\delta_{h})$. This figure shows that, despite the overall good agreement among the methods, the CPD from the parametric approach shows deviations with respect to the reference, which give rise to the discrepancies observed in the summary statistics showed in Fig. \[Pk\_patchy\_OldBias\]. In order further bring these differences to light, in Fig. \[bias\_pdf\] we show examples of conditional probability distribution $P(y|x)$ in three different bins of DM density (or $x$) measured for the full samples (top panels) and in different cosmic web types (from the second to the fourth row). From the first row it is clear how [`PATCHY` ]{}tends to populate cells (or regions) of high density (e.g. $\langle x \rangle = 2.1$) with more halos, as compared to the reference. Approximately the same behavior appears when the CPD is measured in knots and filaments (second and third row, respectively), while similar behaviors are observed for the set of sheets and voids. In general, the improvement gained with `BAM` over the [`PATCHY` ]{}method (at the full mass resolution) is evident, suggesting that a parametrisation of the halo-bias as introduced in § \[sec:patchy\] cannot capture the full complexity of the clustering of dark matter haloes at this mass limit. Evidently, a path to close this gap would be that of increasing the dimensionality of the parameter space to be explored. One clear extension would be that of including higher order terms of the dark matter overdensity in Eq. (\[eq:deterministic\_nonlocal\]). Nevertheless, we consider that given the performance of the non-parametric approach, in which we can capture almost all the physical information of the halo bias without any fitting parameter, makes it worth to put more efforts in pushing the latter towards better performances in view of the construction of mock catalogs for the forthcoming large-scale structure experiments. We are currently making efforts in that direction (F-S Kitaura et al., in preparation) and on the inclusion of velocities (Balaguera et al. in preparation). ![Reduced bispectrum $Q(\theta_{12})$ measured as a function of the angle between wave-vectors $\theta_{12}=\vec{k}_{1}\cdot \vec{k}_{2}$ (with $|\vec{k}_1|=0.1\,h\, {\rm{Mpc}^{-1}}$ and $|\vec{k}_2|=0.2\,h\, {\rm{Mpc}}^{-1}$), for different halo mass-cuts at $z=0$. Solid lines represents the measurements from the `SMDPL` simulation. Dotted and dashed lines correspond to the results from `BAM`, with and without cosmic web classification respectively. In all panels, the grey line represents the measurements from the reference in the lower mass cut. The grey line represents the bispectrum of the full sample.[]{data-label="bias_bam_masscuts"}](images/bispectrum_bam_SMD_z0_b.pdf){width="14cm"} A signature of non-local bias? {#sec:nl} ============================== In § \[sec:bam\] we showed that the non-parametric approach to halo-bias generates an accurate description of the three-point statistics at the mass resolution and volume of the `SMDPL` simulation. Furthermore, Fig. \[Pk\_patchy\_OldBias\] demonstrated that the inclusion of non-local properties (viz., cosmic-web classification and the mass of collapsing regions) yields a reduced bispectrum which agrees within $\sim3\%$ ($1\sigma_B$) with that of the reference. Interestingly, this difference is increased to $\sim5\%$ ($2-3\sigma_B$) when only local information is implemented in the bias mapping. We argue that this represents an smoking-gun for the signatures of non-local halo-bias [see e.g @2013PhRvD..87h3002S]. In view of coming missions such as Euclid or DESI, it is interesting to assess such signature for higher mass cuts, specially those characterizing the selection function of the aforementioned missions [^4]. To that aim, we perform the same calibration procedures in `BAM`, using as a reference the sub-samples of the full `SMDPL` simulation selected by mass-cuts. Figure \[bias\_bam\_masscuts\] shows the results, ranging from a sample with mass $ \sim 4\times 10^{8}{\rm M}_{\odot}h^{-1}$ up to $1.2\times 10^{12}{\rm M}_{\odot}h^{-1}$. `BAM` replicates the bispectrum of the reference with non-local information (dotted-lines) within $\sim3-6\%$ (or roughly within the error bars from Eq. (\[err\_bis\])), up to masses of $2.5\times 10^{10}{\rm M}_{\odot}h^{-1}$. If the non-local information is not included (dashed lines), the signal is correct within $\sim6-12\%$ ($\sim 2-3\sigma_B$ as in Fig. \[Pk\_patchy\_OldBias\]). At $10^{11}{\rm M}_{\odot}h^{-1}$, a systematic bias in the estimates of the bispectrum appears, represented by a $\sim8\%$ ($\sim 1-2\sigma_B$) difference (with non-local dependencies) and $\sim20\%$ ($\sim 3-4\sigma_B$, with only local dependencies). The discrepancy at higher mass-cuts increases although the non-local prediction is able to follow the three-point statistics amplitude of the reference. The reason for this can reside in either the low number count statistics at such high mass-cuts (passing from a Poisson signal-to-noise of $\sim 30$ from the full sample to $\sim 1$ for the highest mass-cut used in Fig. \[bias\_bam\_masscuts\]), or in the lack of other properties of the ADMF taken into account within the `BAM` procedure. ![Reduced bispectrum measured from the `SMDPL` simulation using the full halo catalog (solid line) and diluted (long-dashed line) to different mean number densities characterized by different mass cuts (shown in each panel) (using the same triangle configuration as in Fig. \[bias\_bam\_masscuts\]). The green and red dotted lines represent the bispectrum obtained from `BAM` calibrated with each diluted versions of the `SMDPL`, with and without non-local information, respectively.[]{data-label="bias_bam_downsampled"}](images/bispectrum_nmean_test_z0.pdf){width="9.6cm"} In order to verify whether the behavior captured in Fig. \[bias\_bam\_masscuts\] have physical grounds or, on the contrary are due to a low signal-to-noise number counts, we have performed the following test. We have down-sampled the full `SMDPL` simulation to different mean number densities (associated to different mass-cuts within the same volume) and verified at which level the reduced bispectrum generated by `BAM` was still compatible with that of the reference. Note again that, in each case, `BAM` is calibrated to the down-sampled `SMDPL` and not the entire `SMDPL` as in the case of the mass cuts. The results, shown in Fig. \[bias\_bam\_downsampled\] demonstrates that up to mean number densities associated to mass cuts of $\sim 10^{11}{\rm M}_{\odot}h^{-1}$, `BAM` does a good job at reproducing the reference bispectrum using non-local information. At that mass-cut, we already discussed that, using only the local information, yields a systematic bias which amounts to $\sim20\%$. Thus, this tests concludes that the mean number density associated to this mass-cut marks a threshold up to which the use of non-local information in the measurement of the CPD can reproduce the bispectrum to $<1\sigma_B$ precision. Translating this to Fig. \[bias\_bam\_masscuts\], we can theorize that the signal at the mass-cut $10^{11}{\rm M}_{\odot}h^{-1}$ is not dominated by the signal-to-noise and hence, the discrepancy (of the order of $1\sigma_B$) between the product of `BAM` and the reference can be due to lack of physical dependencies, not accounted for in the measurements of the CPD. Including such dependencies is out of the scope of this paper, and will be addressed in a forthcoming publication (F-S. Kitaura et al., in preparation). Note that Fig. \[bias\_bam\_downsampled\] also indicates that we can trust both the local and non-local calibrations for number densities corresponding to masses lower than $\sim 6\times10^{9}{\rm M}_{\odot}h^{-1}$, and thus, the discrepancy of Fig. \[Pk\_patchy\_OldBias\] between the `BAM` can be considered as a possible signature of non-locality. Conclusions =========== In this paper we have compared the performance of two calibrated methods to construct mock catalogs of dark matter haloes. The methods [`PATCHY` ]{}[@2014MNRAS.439L..21K] and `BAM` [@2019MNRAS.483L..58B] represent examples of parametric and non-parametric approaches, respectively, to the idea of mapping the halo-bias onto a dark matter density field in order to reproduce the two-point statistics of a reference simulation. We have shown that a parametric approach, in which a mean halo-bias and the scatter around that mean is modeled with a set of parameters (including more complex versions of the widely used power-law bias), cannot account for the full complexity of the halo statistics at a mass scale of the resolution of the Small MultiDark simulation ($\sim 2\times10^{8}{\rm M}_{\odot}\,h^{-1}$), used as a reference. On the other hand, the non-parametric approach (`BAM`) generates a good description of the halo distribution up to three-point statistics, thus demonstrating to be properly tailored and suited to capture the main physical properties of the halo bias. As it is clear that larger parameter spaces can be argued in favor of a parametric approach (and perhaps new functional dependencies), we conclude that `BAM` has the potential to account for high order statistics in a transparent way, given that the physical content can be understood and no tuned parameters are requested. Furthermore, we have commented on the possibility of detecting non-local bias with `BAM`, when focused mass scales typical of experiments such as DESI. This will be addressed in a forthcoming publication (F-S Kitaura et al., in preparation). In summary, we conclude that although parametric approaches to halo-bias mapping are successful at reproducing halo statistics at halo masses of $>10^{12}\,{\rm M}_{\odot}\,h^{-1}$(@2017MNRAS.472.4144V), at lower mass scales ($>10^{8}{\rm M}_{\odot}\,h^{-1}$) these models of halo bias (including stochasticity) cannot generate the same precision in the summary statistics of the halo (and sub-halo) distribution. Instead, we have shown that non-parametric approaches (in particular `BAM`), have the potential to reproduce up to the three point statistics of a halo distribution at low mass scales, where non linear-clustering and non-local dependencies are likely to be dominant. We note however that provided that our reference sample is composed of halos and sub-halos, the parametric prescriptions are not expected to be accurate, as these were initially motivated for host-haloes. Nonetheless, foreseeing applications for forthcoming galaxy surveys probing emission line galaxies (ELGs), which populate both halos and sub-halos [see e.g. @2018MNRAS.475.2530O], we believe that it is important to show that our methodologies are able to reproduce the clustering of a sample including sub-structures. A clear assessment of the applicability of both methods in such context requires a reliable ELGs reference catalog. Our efforts are currently focused towards that direction. Acknowledgements {#acknowledgements .unnumbered} ================ MPI acknowledges financial support from the Spanish Ministry of Economy and Competitiveness (MINECO) under the grant AYA2012-39702-C02-01. ABA acknowledges financial support from the Spanish Ministry of Economy and Competitiveness (MINECO) under the Severo Ochoa program SEV-2015-0548. FSK thanks support from the grants RYC2015-18693, SEV-2015-0548 and AYA2017-89891-P. GY acknowledges support from MINECO/FEDER under research grant AYA2015-63810-P and MICIU/FEDER PGC2018-094975-C21. MPI would like to thank Rodrigo Menchón and Matteo Zennaro for their useful discussions. The `SMD` simulation has been performed on SuperMUC at LRZ in Munich within the pr87yi project . The CosmoSim database [www.cosmosim.org](www.cosmosim.org) that provides access to the `SMDPL` simulation and the Rockstar halo catalogues is a service by the Leibniz Institute for Astrophysics Potsdam (AIP). We acknowledge the support of the European Research Council through grant ERC-StG/716151. [^1]: See <http://dx.doi.org/10.17876/cosmosim/smdpl/> for details about the `SMDPL` [^2]: For $\beta\rightarrow\infty$ it can can be shown that the second term in Eq. (\[eq:devpois\]) goes to one, reducing the full expression to a Poisson distribution. [^3]: For more details on the iterative process see figure 2 from [@2019arXiv190606109B] [^4]: For instance, Euclid will observe emission line galaxies residing roughly in halos of mass $\sim 10^{11}h^{-1}{\rm M}_\odot$ [see e.g. @doi:10.1093/mnras/stx1980].

ArXiv

--- abstract: 'A bosonization of the quantum affine superalgebra $U_q(\widehat{sl}(M|N))$ is presented for an arbitrary level $k \in {\bf C}$. The Wakimoto realization is given by using $\xi-\eta$ system. The screening operators that commute with $U_q(\widehat{sl}(M|N))$ are presented for the level $k \neq -M+N$. New bosonization of the affine superalgebra $\widehat{sl}(M|N)$ is obtained in the limit $q \to 1$.' --- [TAKEO KOJIMA]{} Introduction {#sec:1} ============ Bosonization is a powerful method to study representation theory and its application to mathematical physics [@Frenkel]. Wakimoto realization is the bosonization that provides a bridge between representation theory of affine algebras and the geometry of the semi-infinite flag manifold. The Wakimoto realizations have been constructed for the affine Lie algebra $g=(ADE)^{(r)}~(r=1,2)$, $(BCFG)^{(1)}$ and $\widehat{sl}(M|N)$, ${osp}(2|2)^{(2)}$, $D(2,1,a)^{(1)}$ [@Wakimoto; @Feigin-Frenkel1; @Feigin-Frenkel2; @Ito-Komata; @Boer-Feher; @Szczesny; @Feher-Pusztai; @Ding-Gould-Zhang; @Yang-Zhang-Liu; @Iohara-Koga; @Shafiekhani-Chung]. They have been used to construct correlation functions of WZW models, in the study of Drinfeld-Sokolov reduction and $W$-algebras. It’s nontrivial to give quantum deformation of Wakimoto realization as the same as quantum Drinfeld-Sokolov reduction and quantum $W$-algebras. The quantum Wakimoto realizations have been constructed only for $U_q(\widehat{sl}(N))$ and $U_q(\widehat{sl}(2|1))$ [@Matsuo; @Shiraishi; @Awata-Odake-Shiraishi1; @Awata-Odake-Shiraishi2; @Zhang-Gould]. In this paper we study a higher-rank generalization of the previous works for the quantum affine superalgebra $U_q(\widehat{sl}(2|1))$. We give a bosonization of the quantum affine superalgebra $U_q(\widehat{sl}(M|N))$ for an arbitrary level $k \in {\bf C}$, and give the Wakimoto realization using $\xi-\eta$ system. We give the screening operators that commute with $U_q(\widehat{sl}(M|N))$ for the level $k \neq -M+N$. Taking the limit $q \to 1$, we obtain new bosonization of the affine superalgebra $\widehat{sl}(M|N)$. This paper is a shorter review of the papers [@Kojima1; @Kojima2; @Kojima3; @Kojima4]. Quantum affine superalgebra $U_q(\widehat{sl}(M|N))$ {#sec:2} ==================================================== In this Section we recall the definition of the quantum affine superalgebra $U_q(\widehat{sl}(M|N))$ for $M,N=1,2,3,\cdots$. Throughout this paper, $q \in {\bf C}$ is assumed to be $0<|q|<1$. For any integer $n$, define $[n]_q=\frac{q^n-q^{-n}}{q-q^{-1}}$. We set $\nu_i=+1~(1\leq i \leq M)$, $\nu_i=-1~(M+1\leq i \leq M+N)$ and $\nu_0=-1$. The Cartan matrix $(A_{i,j})_{0 \leq i, j \leq M+N-1}$ of the affine Lie superalgebra $\widehat{sl}(M|N)$ is given by $$A_{i,j}=(\nu_i+\nu_{i+1})\delta_{i,j}-\nu_i \delta_{i,j+1}-\nu_{i+1}\delta_{i+1,j}.$$ The quantum affine superalgebra $U_q(\widehat{sl}(M|N))$ [@Yamane] is the associative algebra over ${\bf C}$ with the generators $X_{m}^{\pm, i}~(i=1,2,\cdots,M+N-1, m \in {\bf Z})$, $H_{n}^i~(i=1,2,\cdots,M+N-1, n \in {\bf Z}_{\neq 0})$, $H^i~(i=1,2,\cdots,M+N-1)$, and $c$. The ${\bf Z}_2$-grading of the generators is given by $p(X_m^{\pm, M})\equiv 1 \pmod{2}$ for $m \in {\bf Z}$ and zero otherwise. The defining relations of the generators are given as follows. $$\begin{aligned} &&c : {\rm central~element},\nonumber\\ &&[H^i,H_m^j]=0,~~~[H_{m}^i,H_{n}^j]=\frac{[A_{i,j}m]_q[cm]_q}{m}\delta_{m+n,0}, \nonumber\\ &&[H^i,X^{\pm,j}(z)]=\pm A_{i,j}X^{\pm,j}(z), \nonumber\\ &&[H_{m}^i, X^{\pm,j}(z)]=\pm \frac{[A_{i,j}m]_q}{m}q^{\mp \frac{c}{2}|m|} z^m X^{\pm,j}(z), \nonumber \\ &&(z_1-q^{\pm A_{i,j}}z_2) X^{\pm,i}(z_1)X^{\pm,j}(z_2) = (q^{\pm A_{j,i}}z_1-z_2) X^{\pm,j}(z_2)X^{\pm,i}(z_1),~~{\rm for}~|A_{i,j}|\neq 0, \nonumber \\ && [X^{\pm,i}(z_1), X^{\pm,j}(z_2)]=0~~~{\rm for}~|A_{i,j}|=0, \nonumber \\ &&[X^{+,i}(z_1), X^{-,j}(z_2)] =\frac{\delta_{i,j}}{(q-q^{-1})z_1z_2} \left( \delta(q^{c}z_2/z_1)\Psi_+^i(q^{\frac{c}{2}}z_2)- \delta(q^{-c}z_2/z_1)\Psi_-^i(q^{-\frac{c}{2}}z_2)\right), \nonumber \\ && [X^{\pm,i}(z_{1}), [X^{\pm,i}(z_{2}), X^{\pm,j}(z)]_{q^{-1}}]_q+\left(z_1 \leftrightarrow z_2\right)=0 ~~~{\rm for}~|A_{i,j}|=1,~i\neq M, \nonumber\\ && [X^{\pm,M}(z_1), [X^{\pm,M+1}(w_1), [X^{\pm,M}(z_2), X^{\pm, M-1}(w_2)]_{q^{-1}}]_q ] +(z_1 \leftrightarrow z_2)=0, \nonumber\end{aligned}$$ where we use $$\begin{aligned} [X,Y]_a=XY-(-1)^{p(X)p(Y)}a YX,\nonumber\end{aligned}$$ for homogeneous elements $X,Y \in U_q(\widehat{sl}(M|N))$. For simplicity we write $[X,Y]=[X,Y]_1$. Here we set $\delta(z)=\sum_{m \in {\bf Z}}z^m$ and the generating functions $$\begin{aligned} &&X^{\pm,j}(z)= \sum_{m \in {\bf Z}}X_{m}^{\pm,j} z^{-m-1},\nonumber\\ &&\Psi_\pm^i(q^{\pm \frac{c}{2}}z)=q^{\pm h_i} \exp\left( \pm (q-q^{-1})\sum_{m>0}H_{\pm m}^i z^{\mp m} \right).\nonumber\end{aligned}$$ The multiplication rule for the tensor product is ${\bf Z}_2$-graded and is defined for homogeneous elements $X_1, X_2, Y_1, Y_2 \in U_q(\widehat{sl}(M|N))$ by $(X_1 \otimes Y_1) (X_2 \otimes Y_2)=(-1)^{p(Y_1)p(X_2)} (X_1 X_2 \otimes Y_1 Y_2)$, which extends to inhomogeneous elements through linearity. Let $\bar{\alpha}_i$, $\bar{\Lambda}_i$ $(1\leq i \leq M+N-1)$ be the classical simple roots, the classical fundamental weights, respectively. Let $(\cdot|\cdot)$ be the symmetric bilinear form satisfying $(\bar{\alpha}_i|\bar{\alpha}_j)=A_{i,j}$ and $(\bar{\Lambda}_i|\bar{\alpha}_j)=\delta_{i,j}$ for $1\leq i,j \leq M+N-1$. Let us introduce the affine weight $\Lambda_0$ and the null root $\delta$ satisfying $(\Lambda_0|\Lambda_0)=(\delta|\delta)=0$, $(\Lambda_0|\delta)=1$, and $(\Lambda_0|\bar{\alpha}_i)=(\Lambda_0|\bar{\Lambda}_i)=0$ for $1\leq i \leq M+N-1$. The other affine weights and the affine roots are given by $\Lambda_i=\bar{\Lambda}_i+\Lambda_0$, $\alpha_i=\bar{\alpha}_i$ for $1 \leq i \leq M+N-1$, and $\alpha_0=\delta-\sum_{i=1}^{M+N-1}\alpha_i$. Let $V(\lambda)$ be the highest-weight module over $U_q(\widehat{sl}(M|N))$ generated by the highest weight vector $|\lambda \rangle \neq 0$ such that $$\begin{aligned} &&H_m^i |\lambda \rangle=X_m^{\pm,i}|\lambda \rangle=0~~~(m>0),\nonumber \\ &&X_0^{+,i}|\lambda \rangle=0,~~~H^i|\lambda \rangle=l_i |\lambda \rangle,\nonumber\end{aligned}$$ where the classical part of the highest weight is $\bar{\lambda}=\sum_{i=1}^{M+N-1} l_i \bar{\Lambda}_i$. Bosonization of $U_q(\widehat{sl}(M|N))$ ======================================== In this Section we give a bosonization of $U_q(\widehat{sl}(M|N))$ for an arbitrary level $k \in {\bf C}$. Boson ----- We introduce bosons $a_m^i~(m \in {\bf Z}, 1\leq i \leq M+N-1)$, $b_m^{i,j}~(m \in {\bf Z}, 1\leq i<j \leq M+N)$, $c_m^{i,j}~(m \in {\bf Z}, 1\leq i<j \leq M+N)$, and zero mode operators $Q_a^i~(1\leq i \leq M+N-1)$, $Q_b^{i,j}~(1\leq i<j \leq M+N)$, $Q_c^{i,j}~(1\leq i<j \leq M+N)$. Their commutation relations are $$\begin{aligned} &&[a_m^i,a_n^j]=\frac{1}{m}[(k+g)m]_q[A_{i,j}m]_q\delta_{m+n,0}, ~~~[a_0^i,Q_a^j]=(k+g)A_{i,j},\nonumber \\ &&[b_m^{i,j},b_n^{i',j'}]=-\nu_i \nu_j \frac{1}{m} [m]_q^2 \delta_{i,i'}\delta_{j,j'}\delta_{m+n,0}, ~~~[b_0^{i,j},Q_b^{i',j'}]=-\nu_i \nu_j \delta_{i,i'} \delta_{j,j'}, \nonumber \\ &&[c_m^{i,j},c_n^{i',j'}]=\nu_i \nu_j \frac{1}{m} [m]_q^2 \delta_{i,i'}\delta_{j,j'}\delta_{m+n,0}, ~~~[c_0^{i,j},Q_c^{i',j'}]=\nu_i \nu_j \delta_{i,i'} \delta_{j,j'}, \nonumber\\ &&[Q_b^{i,j},Q_b^{i',j'}]=\pi \sqrt{-1}~~~(\nu_i \nu_j=\nu_{i'}\nu_{j'}=-1).\nonumber\end{aligned}$$ The remaining commutators vanish. Here $g=M-N$ stands for the dual Coxeter number. We define free boson fields $b_\pm^{i,j}(z), b^{i,j}(z)$ as follows. $$\begin{aligned} &&b_\pm^{i,j}(z)= \pm (q-q^{-1}) \sum_{m>0} b_{\pm m}^{i,j} z^{\mp m}\pm b_0^{i,j} {\rm log}q,\nonumber \\ &&b^{i,j}(z)=-\sum_{m \neq 0}\frac{b_m^{i,j}}{[m]_q}z^{-m}+Q_b^{i,j}+b_0^{i,j}{\rm log}z.\nonumber\end{aligned}$$ Free boson fields $a_\pm^i(z), c^{i,j}(z)$ are defined in the same way. We define free boson fields $(\Delta^{\varepsilon}_{L} b_\pm^{i,j})(z), (\Delta^{\varepsilon}_{R} b_\pm^{i,j})(z)$ $(\varepsilon=\pm,0)$ as follows. $$\begin{aligned} &&(\Delta^{\varepsilon}_L b_\pm^{i.j})(z)= \left\{\begin{array}{cc} b_\pm^{i+1,j}(q^{\varepsilon}z)-b_\pm^{i,j}(z)& (\varepsilon=\pm), \nonumber\\ b_\pm^{i+1,j}(z)+b_\pm^{i,j}(z)& (\varepsilon=0), \end{array}\right. \\ &&(\Delta^{\varepsilon}_R b_\pm^{i.j})(z)=\left\{\begin{array}{cc} b_\pm^{i,j+1}(q^{\varepsilon}z)-b_\pm^{i,j}(z)& (\varepsilon=\pm),\\ b_\pm^{i,j+1}(z)+b_\pm^{i,j}(z)& (\varepsilon=0). \end{array}\right. \nonumber\end{aligned}$$ We define free boson fields with parameters $L_1,\cdots,L_r, M_1,\cdots, M_r, \alpha$ as follows. $$\begin{aligned} && \left(\frac{L_1}{M_1}\frac{L_2}{M_2}\cdots \frac{L_r}{M_r}~a^i\right)(z;\alpha)\nonumber \\ &=&-\sum_{m \neq 0} \frac{[L_1 m]_q [L_2 m]_q \cdots [L_r m]_q}{[M_1 m]_q [M_2m]_q \cdots [M_r m]_q}\frac{a_m^i}{[m]_q}q^{-\alpha|m|}z^{-m} +\frac{L_1 L_2 \cdots L_r}{M_1 M_2 \cdots M_r}(Q_a^i+a_0^i {\rm log}z).\nonumber\end{aligned}$$ Normal ordering rules are defined as follows. $$\begin{aligned} &&:b_m^{i,j} b_n^{i',j'}:=:b_n^{i',j'} b_m^{i,j}:=\left\{ \begin{array}{cc} b_m^{i,j} b_n^{i',j'}& (m<0),\\ b_n^{i',j'} b_m^{i,j}& (m>0), \end{array} \right.\nonumber \\ &&:Q_b^{i,j} Q_b^{i',j'}:=:Q_b^{i',j'} Q_b^{i,j}:=Q_b^{i,j} Q_b^{i',j'}~~~(i>i'~~{\rm or}~~i=i',j>j'). \nonumber\end{aligned}$$ Normal ordering rules of $a_m^i$, $c_m^{i,j}$ and $Q_c^{i,j}$ are defined in the same way. Bosonization ------------ We define bosonic operators $\Psi_\pm^i(z)~(1\leq i \leq M+N-1)$ as follows. $$\begin{aligned} \Psi_\pm^i(q^{\pm \frac{k}{2}}z) &=& :e^{a_\pm^i(q^{\pm \frac{g}{2}}z)+\sum_{l=1}^i (\Delta_R^{\mp}b_\pm^{l,i})(q^{\pm(\frac{k}{2}+l)}z) -\sum_{l=i+1}^M (\Delta_L^{\mp}b_\pm^{i,l})(q^{\pm(\frac{k}{2}+l)}z)}\nonumber\\ &\times& e^{-\sum_{l=M+1}^{M+N}(\Delta_L^{\mp} b_\pm^{i,l})(q^{\pm(\frac{k}{2}+2M+1-l)}z)}:~~~(1\leq i \leq M-1), \label{def:psi1} \\ \Psi_\pm^M(q^{\pm \frac{k}{2}}z) &=& :e^{a_\pm^M(q^{\pm \frac{g}{2}}z)-\sum_{l=1}^{M-1} (\Delta_R^0 b_\pm^{l,M})(q^{\pm(\frac{k}{2}+l)}z)+\sum_{l=M+2}^{M+N} (\Delta_L^0 b_\pm^{M,l})(q^{\pm(\frac{k}{2}+2M+1-l)}z)}:,\nonumber \\ \label{def:psi2} \\ \Psi_\pm^i(q^{\pm \frac{k}{2}}z) &=&:e^{a_\pm^i(q^{\pm \frac{g}{2}}z) -\sum_{l=1}^M (\Delta_R^\pm b_\pm^{l,i})(q^{\pm (\frac{k}{2}+l-1)}z)-\sum_{l=M+1}^i (\Delta_R^\pm b_\pm^{l,i})(q^{\pm(\frac{k}{2}+2M-l)}z)}\nonumber\\ &\times& e^{\sum_{l=i+1}^{M+N} (\Delta_L^\pm b_\pm^{i,l})(q^{\pm(\frac{k}{2}+2M-l)}z)}:~~~(M+1\leq i \leq M+N-1). \label{def:psi3}\end{aligned}$$ We define bosonic operators $X^{\pm,i}(z)~(1\leq i \leq M+N-1)$ as follows. $$\begin{aligned} X^{+,i}(z)&=&\sum_{j=1}^i\frac{c_{i,j}}{(q-q^{-1})z}(E_{i,j}^+(z)-E_{i,j}^-(z))~~~(1\leq i \leq M-1), \label{def:X^+1}\\ X^{+,M}(z)&=&\sum_{j=1}^M c_{M,j} E_{M,j}(z), \label{def:X^+2} \\ X^{+,i}(z)&=& \sum_{j=1}^M c_{i,j} E_{i,j}(z)+\sum_{j=M+1}^i \frac{c_{i,j}}{(q-q^{-1})z}(E_{i,j}^+(z)-E_{i,j}^-(z))\nonumber\\ &&~~~~~(M+1\leq i \leq M+N-1), \label{def:X^+3}\\ X^{-,i}(z)&=& \sum_{j=1}^{i-1} \frac{d_{i,j}^1}{(q-q^{-1})z}(F_{i,j}^{1,-}(z)-F_{i,j}^{1,+}(z)) +\frac{d_{i,i}^2}{(q-q^{-1})z}(F_{i,i}^{2,-}(z)-F_{i,i}^{2,+}(z))\nonumber\\ &+&\sum_{j=i+2}^M \frac{d_{i,j}^3}{(q-q^{-1})z}(F_{i,j}^{3,-}(z)-F_{i,j}^{3,+}(z))+\sum_{j=M+1}^{M+N} d_{i,j}^3 F_{i,j}^3(z)\nonumber\\ &&~~~~~(1\leq i \leq M-1), \label{def:X^-1} \\ X^{-,M}(z) &=&\sum_{j=1}^{M-1}\frac{d_{M,j}^1}{(q-q^{-1})z}(F_{M,j}^{1,-}(z)-F_{M,j}^{1,+}(z))+\frac{d_{M,M}^2}{(q-q^{-1})z}(F_{M,M}^{2,-}(z)-F_{M,M}^{2,+}(z))\nonumber\\ &+&\sum_{j=M+2}^{M+N}\frac{d_{M,j}^3}{(q-q^{-1})z}(F_{M,j}^{3,-}(z)-F_{M,j}^{3,+}(z)), \label{def:X^-2} \\ X^{-,i}(z) &=&\sum_{j=1}^M d_{i,j}^1 F_{i,j}^{1}(z) +\sum_{j=M+1}^{i-1}\frac{d_{i,j}^1}{(q-q^{-1})z}(F_{i,j}^{1,-}(z)-F_{i,j}^{1,+}(z))\nonumber\\ &&+ \frac{d_{i,i}^2}{(q-q^{-1})z}(F_{i,i}^{2,-}(z)-F_{i,i}^{2,+}(z)) +\sum_{j=i+2}^{M+N}\frac{d_{i,j}^3}{(q-q^{-1})z}(F_{i,j}^{3,-}(z)-F_{i,j}^{3,+}(z))\nonumber\\ &&~~~~~(M+1 \leq i \leq M+N-1). \label{def:X^-3}\end{aligned}$$ We set $E_{i,j}^\pm(z)$ as follows. $$\begin{aligned} E_{i,j}^\pm (z)&=&:e^{(b+c)^{j,i}(q^{j-1}z)+b_\pm^{j,i+1}(q^{j-1}z)-(b+c)^{j,i+1}(q^{j-1\pm 1}z) +\sum_{l=1}^{j-1}(\Delta_R^- b_+^{l,i})(q^lz)}:\nonumber\\ &&~~~~~~~(1\leq j<i \leq M-1),\nonumber \\ E_{i,i}^\pm (z) &=&:e^{b_\pm^{i,i+1}(q^{i-1}z)-(b+c)^{i,i+1}(q^{i-1\pm 1}z)+\sum_{l=1}^{i-1} (\Delta_R^- b_+^{l,i})(q^lz)}:~~~(1\leq j<i \leq M-1), \nonumber\\ E_{i,i}^\pm (z) &=&:e^{-b_\pm ^{i,i+1}(q^{2M+1-i}z)-(b+c)^{i,i+1}(q^{2M+1 \mp 1-i}z)}\nonumber\\ &\times& e^{-\sum_{l=1}^M(\Delta_R^+ b_+^{l,i})(q^{l-1}z) -\sum_{l=M+1}^{i-1}(\Delta_R^+ b_+^{l,i})(q^{2M-l}z)}:~~~(M+1\leq i \leq M+N-1),\nonumber \\ E_{i,j}^\pm (z)&=&:e^{(b+c)^{j,i}(q^{2M+1-j}z)-b_\pm^{j,i+1}(q^{2M+1-j}z)-(b+c)^{j,i+1}(q^{2M+1 \mp 1-j}z)}\nonumber\\ &\times& e^{-\sum_{l=1}^M(\Delta_R^+ b_+^{l,i})(q^{l-1}z)-\sum_{l=M+1}^{j-1}(\Delta_R^+ b_+^{l,i})(q^{2M-l}z)}:~~~(M+1\leq j<i\leq M+N-1).\nonumber\end{aligned}$$ We set $E_{i,j}(z)$ as follows. $$\begin{aligned} E_{M,j}(z) &=&:e^{(b+c)^{j,M}(q^{j-1}z)+b^{j,M+1}(q^{j-1}z)-\sum_{l=1}^{j-1}(\Delta_R^0 b_+^{l,M})(q^lz)}:~~~(1\leq j \leq M-1),\nonumber \\ E_{M,M}(z) &=&:e^{b^{M,M+1}(q^{M-1}z)-\sum_{l=1}^{M-1}(\Delta_R^0 b_+^{l,M})(q^lz)}:~~~(1\leq j \leq M-1), \nonumber\\ E_{i,j}(z) &=&:e^{b_+^{j,i}(q^{j-1}z)-b^{j,i}(q^jz)+b^{j,i+1}(q^{j-1}z) -\sum_{l=1}^{j-1}(\Delta_R^+ b_+^{l,i})(q^{l-1}z)}:\nonumber\\ &&~~~~~~(M+1\leq i \leq M+N-1,1\leq j \leq M).\nonumber\end{aligned}$$ We set $F_{i,j}^{1,\pm}(z), F_{i,j}^1(z)$ as follows. $$\begin{aligned} F_{i,j}^{1,\pm}(z) &=&:e^{a_-^i(q^{-\frac{k+g}{2}}z)+(b+c)^{j,i+1}(q^{-k-j}z)-b_\pm^{j,i}(q^{-k-j}z)-(b+c)^{j,i}(q^{-k-j\mp 1}z)}\nonumber\\ &\times& e^{\sum_{l=j+1}^i (\Delta_R^+ b_-^{l,i})(q^{-k-l}z)-\sum_{l=i+1}^M (\Delta_L^+ b_-^{i,l})(q^{-k-l}z) -\sum_{l=M+1}^{M+N}(\Delta_L^+ b_-^{i,l})(q^{-k-2M-1+l}z)}: \nonumber\\ &&~~~~~(1\leq j<i\leq M-1),\nonumber\\ F_{M,j}^{1,\pm}(z)&=&:e^{a_-^M(q^{-\frac{k+g}{2}}z)-b_\pm^{j,M}(q^{-k-j}z)-(b+c)^{j,M}(q^{-k-j \mp 1}z)-b_-^{j,M+1}(q^{-k-j}z)-b^{j,M+1}(q^{-k-j+1}z)} \nonumber\\ &\times& e^{-\sum_{l=j+1}^{M-1}(\Delta_R^0 b_-^{l,M})(q^{-k-l}z) +\sum_{l=M+2}^{M+N}(\Delta_L^0 b_-^{M,l})(q^{-k-2M-1+l}z)}:~~~(1\leq j \leq M-1), \nonumber\\ F_{i,j}^1(z)&=& :e^{a_-^i(q^{-\frac{k+g}{2}}z)-b_-^{j,i+1}(q^{-k-j}z)-b^{j,i+1}(q^{-k-j+1}z)+b^{j,i}(q^{-k-j}z) -\sum_{l=j+1}^M(\Delta_R^- b_-^{l,i})(q^{-k-l+1}z)} \nonumber\\ &\times& e^{-\sum_{l=M+1}^i (\Delta_R^- b_-^{l,i})(q^{-k-2M+l}z) +\sum_{l=i+1}^{M+N}(\Delta_L^- b_-^{i,l})(q^{-k-2M+l}z)}:\nonumber\\ &&~~~~~(M+1\leq i \leq M+N-1, 1\leq j \leq M), \nonumber\\ F_{i,j}^{1,\pm}(z)&=&:e^{a_-^i(q^{-\frac{k+g}{2}}z)+(b+c)^{j,i+1}(q^{-k-2M+j}z) +b_\pm^{j,i}(q^{-k-2M+j}z)-(b+c)^{j,i}(q^{-k-2M\pm 1+j}z)}\nonumber \\ &\times& e^{-\sum_{l=j+1}^i(\Delta_R^- b_-^{l,i})(q^{-k-2M+l}z) +\sum_{l=i+1}^{M+N}(\Delta_L^- b_-^{i,l})(q^{-k-2M+l}z)}:\nonumber\\ &&~~~~~~(M+1\leq j<i \leq M+N-1).\nonumber\end{aligned}$$ We set $F_{i,i}^{2,\pm}(z)$ as follows. $$\begin{aligned} F_{i,i}^{2,\pm}(z)&=&:e^{a_\pm^i(q^{\pm \frac{k+g}{2}}z) +b_\pm^{i,i+1}(q^{\pm(k+i+1)}z)+(b+c)^{i,i+1}(q^{\pm(k+i)}z)} \nonumber\\ &\times& e^{-\sum_{l=i+2}^M(\Delta_L^\mp b_\pm^{i,l})(q^{\pm (k+l)}z) -\sum_{l=M+1}^{M+N}(\Delta_L^\mp b_\pm^{i,l})(q^{\pm (k+2M+1-l)}z)}:~~~(1\leq i \leq M-1), \nonumber \\ F_{M,M}^{2,\pm}(z) &=&:e^{a_\pm^M(q^{\pm \frac{k+g}{2}}z)-b^{M,M+1}(q^{\pm(k+M-1)}z) +\sum_{l=M+2}^{M+N}(\Delta_L^0 b_\pm^{M,l})(q^{\pm(k+2M+1-l)}z)}:, \nonumber\\ F_{i,i}^{2,\pm}(z)&=&:e^{a_\pm^i(q^{\pm \frac{k+g}{2}}z) -b_\pm^{i,i+1}(q^{\pm(k+2M-1-i)}z)+(b+c)^{i,i+1}(q^{\pm (k+2M-i)}z)}\nonumber\\ &\times& e^{\sum_{l=i+2}^{M+N}(\Delta_L^\pm b_\pm^{i,l})(q^{\pm(k+2M-l)}z)}:~~~(M+1\leq i \leq M+N-1).\nonumber\end{aligned}$$ We set $F_{i,j}^{3,\pm}(z), F_{i,j}^3(z)$ as follows. $$\begin{aligned} F_{i,j}^{3,\pm}(z)&=&:e^{a_+^i(q^{\frac{k+g}{2}}z)+ (b+c)^{i,j}(q^{k+j-1}z) +b_\pm^{i+1,j}(q^{k+j-1}z)-(b+c)^{i+1,j}(q^{k-1\pm 1+j}z)}\nonumber\\ &\times&e^{-\sum_{l=j}^M(\Delta_L^- b_+^{i,l})(q^{k+l}z) -\sum_{l=M+1}^{M+N}(\Delta_L^- b_+^{i,l})(q^{k+2M+1-l}z)}:~~~(1\leq i<j \leq M-1), \nonumber\\ F_{i,j}^{3}(z)&=&:e^{a_+^i(q^{\frac{k+g}{2}}z)-b^{i,j}(q^{k+2M-j}z)-b_+^{i+1,j}(q^{k+2M-j}z)+b^{i+1,j}(q^{k+2M+1-j}z)}\nonumber\\ &\times& e^{-\sum_{l=j+1}^{M+N}(\Delta_L^- b_+^{i,l})(q^{k+2M+1-l}z)}:~~~(1\leq i \leq M-1, M+1\leq j \leq M+N), \nonumber\\ F_{M,j}^{3,\pm}(z)&=& :e^{a_+^M(q^{\frac{k+g}{2}}z)-b^{M,j}(q^{k+2M-j}z)-b_\pm^{M+1,j}(q^{k+2M+1-j}z) -(b+c)^{M+1,j}(q^{k+2M+1 \mp 1-j}z)}\nonumber\\ &\times& e^{b_+^{M+1,j}(q^{k+2M+1-j}z)+\sum_{l=j+1}^{M+N}(\Delta_L^0 b_+^{M,l})(q^{k+2M+1-l}z)}:~~(M+2 \leq j \leq M+N), \nonumber\\ F_{i,j}^{3,\pm}(z)&=&:e^{a_+^i(q^{\frac{k+g}{2}}z)+(b+c)^{i,j}(q^{k+2M+1-j}z) -b_\pm^{i+1,j}(q^{k+2M+1-j}z)-(b+c)^{i+1,j}(q^{k+2M+1 \mp1-j}z)}\nonumber\\ &\times& e^{\sum_{l=j+1}^{M+N}(\Delta_L^+ b_+^{i,l})(q^{k+2M-l}z)}:~~~(M+1\leq i<j-1\leq M+N-1). \nonumber\end{aligned}$$ The coefficients $c_{i,j} \in {\bf C}$ and $d_{i,j}^1, d_{i,i}^2, d_{i,j}^3 \in {\bf C}$ satisfy the following conditions. $$\begin{aligned} d_{i,j}^1&=&\nu_{i+1}\frac{1}{c_{i,j}}\times \left\{ \begin{array}{cc} 1& (1 \leq i \leq M-1, 1 \leq j \leq i-1), \\ q^{j-1} & (i=M, 1\leq j \leq M-1), \\ q^{-k-1} & (M+1 \leq i \leq M+N-1, 1 \leq j \leq M), \\ 1 & (M+1 \leq i \leq M+N-1, M+1 \leq j \leq i-1), \end{array} \right.\nonumber \\ d_{i,i}^2&=& \nu_{i+1} \frac{1}{c_{i,i}} \times \left\{\begin{array}{cc} 1 & (1\leq i \neq M \leq M+N-1),\\ q^{M-1} & (i=M), \end{array} \right.\nonumber \\ d_{i,j}^3 &=&\nu_{i+1} \frac{1}{c_{i,i}}\prod_{l=1}^{j-i-1}\frac{c_{i+l,i+1}}{c_{i+l,i}} \times \left\{ \begin{array}{cc} 1& (1 \leq i \leq M-1, i+2 \leq j \leq M),\\ q^{k+3M+1-2j}& (1\leq i \leq M-1, M+1 \leq j \leq M+N),\\ q^{(M-1)(j-M)}& (i=M, M+2 \leq j \leq M+N),\\ 1 & (M+1 \leq i \leq M+N-1, i+2 \leq j \leq M+N). \end{array}\right. \nonumber\end{aligned}$$    The bosonic operators $\Psi_\pm^i(z)$ defined in (\[def:psi1\])-(\[def:psi3\]), and $X^{\pm,i}(z)$ defined in (\[def:X\^+1\])-(\[def:X\^+3\]) and (\[def:X\^-1\])-(\[def:X\^-3\]) satisfy the defining relations of the quantum affine superalgebra $U_q(\widehat{sl}(M|N))$ with the central element $c=k \in {\bf C}$. \[thm:1\] Wakimoto realization -------------------- In this Section we introduce the $\xi-\eta$ system and give the Wakimoto realization. We set the boson Fock space ${F}(p_a,p_b,p_c)$ as follows. The vacuum state $|0\rangle \neq 0$ is defined by $ a_m^i|0\rangle= b_m^{i,j}|0\rangle= c_m^{i,j}|0\rangle=0~(m \geq 0)$. Let $|p_a,p_b,p_c\rangle$ be $$\begin{aligned} &&|p_a,p_b,p_c\rangle \nonumber\\ &=&\exp\left( \sum_{i,j=1}^{M+N-1}\frac{(A^{-1})_{i,j}}{k+g}p_a^i Q_a^i-\sum_{1\leq i<j \leq M+N}\nu_i \nu_j p_b^{i,j}Q_b^{i,j}+\sum_{1\leq i<j \leq M+N \atop{\nu_i \nu_j=+1}}p_c^{i,j}Q_c^{i,j}\right)|0\rangle,\nonumber\end{aligned}$$ then $|p_a, p_b, p_c\rangle$ is the highest weight state of the boson Fock space ${F}(p_a,p_b,p_c)$. The boson Fock space $F(p_a,p_b,p_c)$ is generated by the bosons $a_m^i, b_m^{i,j}, c_m^{i,j}$ on the highest weight state $|p_a,p_b,p_c\rangle$. We set the space $F(p_a)$ by $$\begin{aligned} F(p_a)=\bigoplus_{p_b^{i,j}=-p_c^{i,j} \in {\bf Z}~(\nu_i\nu_j=+) \atop{p_b^{i,j}}\in {\bf Z}~(\nu_i\nu_j=-)}F(p_a,p_b,p_c).\nonumber\end{aligned}$$ Here we impose the restriction $p_b^{i,j}=-p_c^{i,j}~(\nu_i \nu_j=+)$, because $X_m^{\pm,i}$ change $Q_b^{i,j}+Q_c^{i,j}$. $F(p_a)$ is $U_q(\widehat{sl}(M|N))$-module. Let $|\lambda \rangle=|p_a,0,0\rangle$ where $p_a^i=l_i~(1\leq i \leq M+N-1)$. The generators $H^i, H_m^i, X_m^{\pm, i}$ act on $|\lambda\rangle$ as follows. $$\begin{aligned} &&H_m^i |\lambda \rangle=X_m^{\pm,i}|\lambda \rangle=0~~~(m>0),\nonumber\\ &&X_0^{+,i}|\lambda \rangle=0,~~~H^i|\lambda \rangle=l_i |\lambda \rangle.\nonumber\end{aligned}$$ We have the level-$k$ highest weight module $V(\lambda)$ of $U_q(\widehat{sl}(M|N))$. $$\begin{aligned} V(\lambda) \subset F(p_a).\nonumber\end{aligned}$$ Here the classical part of the highest weight is $\bar{\lambda}=\sum_{i=1}^{M+N-1} l_i \bar{\Lambda}_i$. We introduce the $\xi-\eta$ system We set bosonic operators $\xi_m^{i,j}, \eta_m^{i,j}~(\nu_i \nu_j=+1, 1\leq i<j \leq M+N)$ as follows. $$\begin{aligned} \eta^{i,j}(z)=\sum_{m \in {\bf Z}}\eta_m^{i,j} z^{-m-1}=:e^{c^{i,j}(z)}:,~~~ \xi^{i,j}(z)=\sum_{m \in {\bf Z}}\xi_m^{i,j}z^{-m}=:e^{-c^{i,j}(z)}:.\nonumber\end{aligned}$$ Fourier components $$\begin{aligned} \eta_m^{i,j}=\oint \frac{dz}{2\pi \sqrt{-1}}z^m \eta^{i,j}(z),~~~ \xi_m^{i,j}=\oint \frac{dz}{2\pi \sqrt{-1}}z^{m-1}\xi^{i,j}(z)\nonumber\end{aligned}$$ are well-defined on the module $F(p_a)$. The ${\bf Z}_2$-grading is given by $p(\xi_m^{i,j})=p(\eta_m^{i,j})=+1$. We have direct sum decomposition. $$\begin{aligned} F(p_a)=\eta_0^{i,j}\xi_0^{i,j} F(p_a) \oplus \xi_0^{i,j} \eta_0^{i,j} F(p_a),\nonumber\end{aligned}$$ where ${\rm Ker}(\eta_0^{i,j})=\eta_0^{i,j} \xi_0^{i,j} F(p_a)$, ${\rm Coker}(\eta_0^{i,j})=\xi_0^{i,j}\eta_0^{i,j} F(p_a)$. We set $$\begin{aligned} \eta_0=\prod_{1\leq i<j \leq M+N\atop{\nu_i\nu_j=+1}}\eta_0^{i,j},~~~ \xi_0=\prod_{1\leq i<j \leq M+N\atop{\nu_i\nu_j=+1}}\xi_0^{i,j}.\nonumber\end{aligned}$$ We introduce the subspace ${\cal F}(p_a)$ by $$\begin{aligned} {\cal F}(p_a)=\eta_0\xi_0 F(p_a).\nonumber\end{aligned}$$ The operators $\eta_0^{i,j}, \xi_0^{i,j}$ commute with $X^{\pm,i}(z), \Psi_\pm^i(z)$ up to sign $\pm 1$.    ${\cal F}(p_a)$ is $U_q(\widehat{sl}(M|N))$-module. We call ${\cal F}(p_a)$ the Wakimoto realization of $U_q(\widehat{sl}(M|N))$. Screening operator ================== In this Section we give the screening operators $Q_i~(1\leq i \leq M+N-1)$ that commute with $U_q(\widehat{sl}(M|N))$ for the level $c=k \neq -g$. We define bosonic operators $S_i(z)~(1\leq i \leq M+N-1)$ that we call the screening currents as follows. $$\begin{aligned} S_i(z)&=&\sum_{j=i+1}^M \frac{e_{i,j}}{(q-q^{-1})z}(S_{i,j}^-(z)-S_{i,j}^+(z))+\sum_{j=M+1}^{M+N}e_{i,j}S_{i,j}(z) \nonumber\\ &&~~~~~~(1\leq i \leq M-1),\label{def:S1} \\ S_M(z)&=&\sum_{j=M+1}^{M+N}e_{M,j} S_{M,j}(z), \label{def:S2} \\ S_i(z)&=&\sum_{j=i+1}^{M+N}\frac{e_{i,j}}{(q-q^{-1})z}(S_{i,j}^-(z)-S_{i,j}^+(z)) \nonumber\\ &&~~~~~(M+1\leq i \leq M+N-1).\label{def:S3}\end{aligned}$$ We set $S_{i,j}^\pm(z)$ as follows. $$\begin{aligned} {S}_{i,j}^\pm(z)&=&:e^{ -(\frac{1}{k+g}a^i)(z;\frac{k+g}{2}) +(b+c)^{i+1,j}(q^{M-N-j}z)-b_\pm^{i,j}(q^{M-N-j}z) -(b+c)^{i,j}(q^{M-N-j \mp 1}z)} \nonumber\\ &\times& e^{\sum_{l=j+1}^{M}(\Delta_L^+ b_-^{i,l})(q^{M-N-l}z)+\sum_{l=M+1}^{M+N}(\Delta_L^+ b_-^{i,l})(q^{-M-N+l-1}z)}: ~~~(1\leq i <j \leq M),\nonumber\\ {S}_{i,j}^\pm(z)&=&:e^{ -(\frac{1}{k+g}a^i)(z;\frac{k+g}{2}) +(b+c)^{i+1,j}(q^{-M-N+j}z)+b_\pm^{i,j}(q^{-M-N+j}z)-(b+c)^{i,j}(q^{-M-N+j\pm 1}z)}\nonumber\\ &\times& e^{-\sum_{l=j+1}^{M+N}(\Delta_L^- b_-^{i,l})(q^{-M-N+l}z)}:~~~(M+1\leq i<j \leq M+N).\nonumber\end{aligned}$$ We set $S_{i,j}(z)$ as follows. $$\begin{aligned} {S}_{i,j}(z)&=&:e^{ -(\frac{1}{k+g}a^i)(z;\frac{k+g}{2}) +b^{i,j}(q^{-M-N+j}z)+b_+^{i+1,j}(q^{-M-N+j}z)-b^{i+1,j}(q^{-M-N+j+1}z)}\nonumber\\ &\times&e^{\sum_{l=j+1}^{M+N}(\Delta_L^+ b_-^{i,l})(q^{-M-N-1+l}z)}:~~~(1\leq i \leq M-1, M+1\leq j \leq M+N), \nonumber\\ {S}_{M,j}(z)&=&:e^{ -(\frac{1}{k+g}a^i)(z;\frac{k+g}{2}) +(b+c)^{M+1,j}(q^{-M-N+j}z)+b^{M,j}(q^{-M-N+j}z)}\nonumber\\ &\times& e^{-\sum_{l=j+1}^{M+N}(\Delta_L^0 b_-^{M,l})(q^{-M-N-1+l}z)}:~~~(M+1\leq j \leq M+N).\nonumber\end{aligned}$$ Here we set $e_{i,j}$ as follows. $$\begin{aligned} e_{i,i+1}&=&\left\{\begin{array}{cc} 1/d_{i,i}^2& (1\leq i \leq M-1),\\ -q^{-N+1}/d_{M,M}^2& (i=M),\\ -1/d_{i,i}^2& (M+1\leq i \leq M+N-1), \end{array}\right. \nonumber \\ e_{i,j}&=& \left\{\begin{array}{cc} 1/d_{i,j}^3& (1\leq i \leq M-1, i+2 \leq j \leq M),\\ q^{k+1+M-N}/d_{i,j}^3& (1\leq i \leq M-1, M+1\leq j \leq M+N),\\ -q^{j-M-N}/d_{M,j}^3& (i=M, M+2\leq j \leq M+N),\\ -1/d_{i,j}^3& (M+1\leq i \leq M+N-1, i+2\leq j \leq M+N). \nonumber \end{array} \right.\end{aligned}$$ The ${\bf Z}_2$-grading of the screening current is given by $p(S_{M,j}(z))\equiv 1 \pmod{2}$ for $M+1 \leq j \leq M+N$ and zero otherwise. The Jackson integral with parameters $q \in {\bf C}$ and $s \in {\bf C}^*$ is defined by $$\begin{aligned} \int_0^{s \infty}f(w)d_q w=s(1-q)\sum_{n \in {\bf Z}}f(sq^n) q^n.\nonumber\end{aligned}$$ We define the screening operators $Q_i~(1\leq i \leq M+N-1)$ as follows, when the Jackson integrals are convergent. $$\begin{aligned} Q_i=\int_0^{s \infty}S_i(w)d_{q^{2(k+g)}}w. \label{def:S0}\end{aligned}$$    The screening operators $Q_i$ $(1\leq i \leq M+N-1)$ defined in (\[def:S1\]), (\[def:S2\]), (\[def:S3\]), (\[def:S0\]) commute with the quantum affine superalgebra $U_q(\widehat{sl}(M|N))$. $$\begin{aligned} [Q_i,U_q(\widehat{sl}(M|N))]=0.\nonumber\end{aligned}$$ Limit $q \to 1$ =============== Bosonization of the affine superalgebra $\widehat{sl}(M|N)$ for an arbitrary level $k$ have been studied in [@Ding-Gould-Zhang; @Yang-Zhang-Liu; @Iohara-Koga]. We obtain new bosonization of the affine superalgebra $\widehat{sl}(M|N)$ in the limit $q\to 1$. In what follows we set $$\begin{aligned} H^i(z)=\sum_{m \in {\bf Z}} H_m^i z^{-m-1}~~~(1\leq i \leq M+N-1).\nonumber\end{aligned}$$ We set the parameters $c_{i,j}=1$ in (\[def:X\^+1\])-(\[def:X\^+3\]), (\[def:X\^-1\])-(\[def:X\^-3\]), (\[def:S1\])-(\[def:S3\]) for simplicity. In the limit $q\to1$ we introduce operators $\alpha_i(z)~(1\leq i \leq M+N-1)$, $\beta_{i,j}(z), \widehat{\beta}_{i,j}(z), \gamma_{i,j}(z)~(1\leq i<j \leq M+N, \nu_i\nu_j=+)$, and $\psi_{i,j}(z), \psi_{i,j}^\dagger(z)~(1\leq i<j \leq M+N, \nu_i\nu_j=-)$ as follows. $$\begin{aligned} && \alpha_i(z)=\partial_z \left(a^i(z)\right), ~~~\gamma_{i,j}(z)=:e^{(b+c)^{i,j}(z)}:, \nonumber\\ && \beta_{i,j}(z)=:\partial_z \left(e^{-c^{i,j}(z)}\right) e^{-b^{i,j}(z)}:,~ \widehat{\beta}_{i,j}(z)=:\partial_z \left(e^{-b^{i,j}(z)}\right) e^{-c^{i,j}(z)}:, \nonumber\\ && \psi_{i,j}(z)=:e^{b^{i,j}(z)}:,~~~\psi_{i,j}^\dagger (z)=:e^{-b^{i,j}(z)}:. \nonumber\end{aligned}$$ They satisfy the following relations. $$\begin{aligned} && \alpha_i(z)\alpha_j(w)=\frac{(k+g)A_{i,j}}{(z-w)^2}+\cdots,\nonumber \\ && \beta_{i,j}(z)\gamma_{i',j'}(w)=\frac{\delta_{i,i'}\delta_{j,j'}}{z-w}+\cdots, ~~~ \gamma_{i,j}(z)\beta_{i',j'}(w)=-\frac{\delta_{i,i'}\delta_{j,j'}}{z-w}+\cdots, \nonumber \\ && \widehat{\beta}_{i,j}(z)\gamma_{i',j'}(w)=-\frac{\delta_{i,i'}\delta_{j,j'}}{z-w}+\cdots, ~~~ \gamma_{i,j}(z)\widehat{\beta}_{i',j'}(w)=\frac{\delta_{i,i'}\delta_{j,j'}}{z-w}+\cdots, \nonumber \\ && \psi_{i,j}(z)\psi_{i',j'}^\dagger (w)=\frac{\delta_{i,i'}\delta_{j,j'}}{z-w}+\cdots, ~~~ \psi_{i,j}^\dagger(z)\psi_{i',j'}(w)=\frac{\delta_{i,i'}\delta_{j,j'}}{z-w}+\cdots. \nonumber\end{aligned}$$ In the limit $q \to 1$ the operators $a_\pm^i(z)$, $b_\pm^{i,j}(z)$, $(\Delta_L^\epsilon b_\pm^{i,j})(z)$ and $(\Delta_R^\epsilon b_\pm^{i,j})(z)$ disappear. We obtain the following. $$\begin{aligned} H^i(z)&=&\alpha_i(z)+\sum_{j=1}^i:(\widehat{\beta}_{j,i}(z)\gamma_{j,i}(z)-\widehat{\beta}_{j,i+1}(z)\gamma_{j,i+1}(z)):\nonumber\\ &&+\sum_{j=i+1}^M:(\widehat{\beta}_{i+1,j}(z)\gamma_{i+1,j}(z)-\widehat{\beta}_{i,j}(z)\gamma_{i,j}(z)):\nonumber\\ &&+\sum_{j=M+1}^{M+N}:( (\partial_z{\psi}_{i+1,j})(z) \psi_{i+1,j}^\dagger(z)- (\partial_z {\psi}_{i,j})(z) \psi_{i,j}^\dagger(z)):~~~(1\leq i \leq M-1),\nonumber\end{aligned}$$ $$\begin{aligned} H^M(z)&=&\alpha_M(z)+\sum_{j=1}^{M-1} :((\partial_z {\psi}_{j,M+1})(z)\psi_{j,M+1}^\dagger(z) +\widehat{\beta}_{j,M}(z)\gamma_{j,M}(z)):\nonumber\\ &&-\sum_{j=M+2}^{M+N}:(\widehat{\beta}_{M+1,j}(z)\gamma_{M+1,j}(z)+(\partial_z {\psi}_{M,j})(z) \psi_{M,j}^\dagger(z)):, \nonumber \\ H^i(z)&=& \alpha_i(z)+\sum_{j=1}^M :((\partial_z {\psi}_{j,i+1})(z)\psi_{j,i+1}^\dagger(z)- (\partial_z{\psi}_{j,i})(z)\psi_{j,i}^\dagger(z)):\nonumber\\ &&+\sum_{j=M+1}^i:(\widehat{\beta}_{j,i+1}(z)\gamma_{j,i+1}(z)-\widehat{\beta}_{j,i}(z)\gamma_{j,i}(z)):\nonumber\\ &&+\sum_{j=i+1}^{M+N} :(\widehat{\beta}_{i,j}(z)\gamma_{i,j}(z)-\widehat{\beta}_{i+1,j}(z)\gamma_{i+1,j}(z)):~~(M+1\leq i \leq M+N-1). \nonumber\end{aligned}$$ $$\begin{aligned} X^{+,i}(z)&=&\sum_{j=1}^i :\beta_{j,i+1}(z)\gamma_{j,i}(z):~~~(1\leq i \leq M-1), \nonumber\\ X^{+,M}(z)&=&\sum_{j=1}^M :\gamma_{j,M}(z)\psi_{j,M+1}(z):, \nonumber\\ X^{+,i}(z)&=&\sum_{j=1}^M:\psi_{j,i+1}(z)\psi_{j,i}^\dagger(z):-\sum_{j=M+1}^i :\beta_{j,i+1}(z)\gamma_{j,i}(z):~(M+1\leq i \leq M+N-1). \nonumber\end{aligned}$$ $$\begin{aligned} X^{-,i}(z)&=&-:\alpha_i(z)\gamma_{i,i+1}(z):-\kappa_i :\partial_z \gamma_{i,i+1}(z):\nonumber\\ &&+\sum_{j=1}^{i-1}:\beta_{j,i}(z)\gamma_{j,i+1}(z): -\sum_{j=i+2}^M:\beta_{i+1,j}(z)\gamma_{i,j}(z): -\sum_{j=M+1}^{M+N}:\psi_{i+1,j}(z)\psi_{i,j}^\dagger(z):\nonumber\\ &&+\sum_{j=i+1}^M :(\widehat{\beta}_{i,j}(z)\gamma_{i,j}(z) -\widehat{\beta}_{i+1,j}(z)\gamma_{i+1,j}(z))\gamma_{i,i+1}(z): \nonumber \\ &&+\sum_{j=M+1}^{M+N} :((\partial_z {\psi}_{i,j})(z)\psi_{i,j}^\dagger(z)- (\partial_z{\psi}_{i+1,j})(z)\psi_{i+1,j}^\dagger(z))\gamma_{i,i+1}(z):\nonumber\\ &&~~~~~(1\leq i \leq M-1), \nonumber\\ X^{-,M}(z)&=&:\alpha_M(z)\psi_{M,M+1}^\dagger(z):+\kappa_M :\partial_z \psi_{M,M+1}^\dagger(z):\nonumber\\ &&-\sum_{j=1}^{M-1}:\beta_{j,M}(z)\psi_{j,M+1}^\dagger(z): -\sum_{j=M+2}^{M+N}:\beta_{M+1,j}(z)\psi_{M,j}^\dagger(z):\nonumber\\ &&-\sum_{j=M+2}^{M+N} :(\widehat{\beta}_{M+1,j}(z)\gamma_{M+1,j}^\dagger(z)+ (\partial_z {\psi}_{M,j})(z)\psi_{M,j}^\dagger(z))\psi_{M,M+1}^\dagger(z):, \nonumber\\ X^{-,i}(z)&=& :\alpha_i(z)\gamma_{i,i+1}(z):+\kappa_i:\partial_z \gamma_{i,i+1}(z):\nonumber\\ &&-\sum_{j=1}^M:\psi_{j,i}(z)\psi_{j,i+1}^\dagger(z): +\sum_{j=M+1}^{i-1}:\beta_{j,i}(z)\gamma_{j,i+1}(z): -\sum_{j=i+2}^{M+N}:\beta_{i+1,j}(z)\gamma_{i,j}(z):\nonumber\\ &&+\sum_{j=i+1}^{M+N}:(\widehat{\beta}_{i,j}(z)\gamma_{i,j}(z)- \widehat{\beta}_{i+1,j}(z)\gamma_{i+1,j}(z))\gamma_{i,j}(z):\nonumber\\ &&~~~~~(M+1\leq i \leq M+N-1).\nonumber\end{aligned}$$ Here we have set the coefficients $\kappa_i$ by $$\begin{aligned} \kappa_i=\left\{ \begin{array}{cc} k+i& (1\leq i \leq M-1)\\ k+M-1& (i=M)\\ k+2M-i& (M+1\leq i \leq M+N-1) \end{array}\right..\nonumber\end{aligned}$$ In what follows we assume $k \neq -g$. In the limit $q\to 1$ we have the following. $$\begin{aligned} S_i(z)&=&\sum_{j=i+1}^M:\tilde{s}_i(z)\beta_{i,j}(z)\gamma_{i+1,j}(z):+ \sum_{j=M+1}^{M+N}:\tilde{s}_i(z)\psi_{i,j}(z)\psi_{i+1,j}^\dagger(z):\nonumber\\ &&~~~~~~(1\leq i \leq M-1), \nonumber\\ S_M(z)&=&\sum_{j=M+1}^{M+N}:\tilde{s}_M(z)\gamma_{M+1,j}(z)\psi_{M,j}(z):, \nonumber\\ S_i(z)&=&\sum_{j=i+1}^{M+N}:\tilde{s}_i(z)\beta_{i,j}(z)\gamma_{i+1,j}(z):~~~(M+1\leq i \leq M+N-1).\nonumber\end{aligned}$$ Here we have set the boson operator $$\tilde{s}_i(z)=:e^{-\left(\frac{1}{k+g}a^i\right)(z;0)}:.$$ Our bosonization is different from [@Ding-Gould-Zhang; @Yang-Zhang-Liu; @Iohara-Koga].  \ [**Acknowledgement**]{}  This work is supported by the Grant-in-Aid for Scientific Research [**C**]{} (26400105) from Japan Society for Promotion of Science. The author would like to thank Professor Michio Jimbo and Professor Vladimir Dobrev for giving advice. The author would like to thank Professor Zengo Tsuboi, Professor Pascal Baseilhac, Professor Kouichi Takemura and Professor Kenji Iohara for discussion. The author is thankful for the kind hospitality by the organizing committee of the 10-th International Symposium “Quantum Theory and Symmetries” (QTS10) and 12-th International Workshop “Lie Theory and Its Applications in Physics” (LT12). [99.]{} E.V.Frenkel, Adv.Math. [**195**]{} (2005) 297-404. M.Wakimoto, Commun.Math.Phys.[**104**]{} (1986) 605-609. B.L.Feigin and E.V.Frenkel, [*Physics and Mathematics of Strings*]{} (World Scientific, Singapore 1980) 271-316. B.L.Feigin and E.V.Frenkel, Commun.Math.Phys. [**128**]{} (1990) 161-189. K.Ito and S.Komata, Mod.Phys.Lett.[**A6**]{} (1991) 581-589. J.de Boer and L.Fehér, Commun.Math.Phys.[**189**]{} (1997) 759-793. M.Szczesny, Math.Res.Lett.[**9**]{} (2002) 433-448. L.Fehér and B.G.Pusztai, Nucl.Phys.[**B674**]{} (2003) 509-532. X.-M.Ding, M.D.Gould and Y.-Z.Zhang, Phys.Lett.[**318**]{} (2003) 354-363. W.-L.Yang,Y.Z.Zhang and X.Liu, J.Math.Phys.[**48**]{} (2007) 053514 (pp.11). K.Iohara and Y.Koga, Math.Proc.Camb.Phil.Soc.[**132**]{} (2002) 419-433. A.Shafiekhani and W.-S.Chung, Mod.Phys.Lett.[**A13**]{} (1998) 47-57. A.Matsuo, Commun.Math.Phys.[**160**]{} (1994) 33-48. J.Shiraishi, Phys.Lett.[**A171**]{} (1992) 243-248. H.Awata, S.Odake and J.Shiraishi, Commun.Math.Phys.[**162**]{} (1994) 61-83. H.Awata, S.Odake and J.Shiraishi, Lett. Math. Phys.[**42**]{} (1997) 271-279. Y.-Z.Zhang and M.D.Gould, J.Math.Phys.[**41**]{} (2000) 5577-5291. T.Kojima, J.Math.Phys.[**53**]{} (2012) 013515 (pp.15). T.Kojima, Springer Proceedings [**111**]{} (2013) 263-276. T.Kojima, J.Math.Phys.[**53**]{} (2012) 083503 (pp.30). T.Kojima, Commun.Math.Phys.[**355**]{} (2017) 603-644. H.Yamane, Publ.Res.Inst.Math.Sci.[**35**]{} (1999) 321-390.

ArXiv

--- abstract: 'V. Guruswami and N. Resch prove that the list decodability of ${\mathbb{F}}_q$-linear rank metric codes is as good as that of random rank metric codes in [@venkat2017]. Due to the potential applications of self-orthogonal rank metric codes, we focus on list decoding of them. In this paper, we prove that with high probability, an ${\mathbb{F}}_q$-linear self-orthogonal rank metric code over ${\mathbb{F}}_q^{n\times m}$ of rate $R=(1-\tau)(1-\frac{n}{m}\tau)-\epsilon$ is shown to be list decodable up to fractional radius $\tau\in(0,1)$ and small $\epsilon\in(0,1)$ with list size depending on $\tau$ and $q$ at most $O_{\tau, q}(\frac{1}{\epsilon})$. In addition, we show that an $\mathbb{F}_{q^m}$-linear self-orthogonal rank metric code of rate up to the Gilbert-Varshamov bound is $(\tau n, \exp(O_{\tau, q}(\frac{1}{\epsilon})))$-list decodable.' author: - 'Shu Liu [^1]' title: | On the List Decodability of Self-orthogonal\ Rank Metric Codes --- Introduction ============ In the late 50’s, P. Elias [@P.E1957], [@P.E1991] and J. M. Wozencraft [@J.M.W1985] introduced list decoding. Compared with unique decoding, list decoding can output a list of codewords which contains the correct transmitted codeword rather than output a unique codeword. Consequentially, the list size of list decoding can be bigger than $1.$ In coding theory, there exists a trade-off between the fraction of errors that can be corrected and the rate. The largest list size of the decoder’s output is an important parameter in list decoding. In fact, we want to have a small list size. At least two reasons can be argued. The first reason is due to the usefulness of this list. After the output of this list, the next step is to utilize this list to decide what the original transmitted message is. This can be done, by outputting the codeword corresponding to the smallest error. If the list size is exponential, the decision step needs exponential time complexity. The second reason is that this list size provides us with a lower bound for the worst-case complexity of the decoding algorithm itself. So, if we require the decoding algorithm to be efficient, we need the list size to be as small as possible. List Decoding of Rank Metric Codes {#list-decoding-of-rank-metric-codes .unnumbered} ---------------------------------- Rank metric codes is a set of $n\times m$ matrices over a finite field ${\mathbb{F}}_q.$ By the ring isomorphism between ${\mathbb{F}}_q^m$ and ${\mathbb{F}}_{q^m},$ an $n\times m$ matrix over ${\mathbb{F}}_q$ can be defined as a vector of length $n$ over the extension field ${\mathbb{F}}_{q^m}.$ Rank metric codes have been receivced many attention because of their applications in network coding [@Sil2008], [@Koe2008], storage systems [@Roth1991], cryptography [@E.M.A.V1991], [@Ove2008], and space-time coding [@Lus2003]. Finding good list decodable rank metric codes attracts more and more researchers. In order to find the limit in which an efficient list decoding is possible, A. Wachter-Zeh provided lower and upper bounds on the list size in [@Ant2013], [@Zeh2012] and showed that the upper bound of the list size is exponential, for any decoding radius beyond half of the minimum distance. In addition, there exists an exponential list size rank metric code for any decoding radius is larger than half of the minimum distance. No efficient list decoding can be found for Gabidulin codes if decoding radius beyond the Johnson bound. Y. Ding [@Ding2015] reveals that the Singleton bound is the list decoding barrier for any rank metric code. With high probability, the decoding radius and the rate of random rank metric codes satisfy the Gilbert-Varshamov bound with constant list size. In addition, with high probability, an ${\mathbb{F}}_q$-linear rank metric code can be list decoded with list decoding radius attaining the Gilbert-Varshamov bound with exponential list size. Since efficient list decoding radius of Gabidulin codes cannot be larger than the unique decoding radius, S. Liu, C. Xing and C. Yuan show that with high probability, a random subcode of a Gabidulin code is list decodable with decoding radius far beyond the unique decoding radius in [@Liu2017]. However, for ${\mathbb{F}}_q$-linear rank metric codes, when the list decoding radius is beyond half of the minimum distance, the list size is exponential. V. Guruswami and N. Resch decrease the list size of ${\mathbb{F}}_q$-linear rank metric codes and show that it is list decodable as good as random rank metric codes in [@venkat2017]. Motivation {#motivation .unnumbered} ---------- There has been some interesting findings on the list decodability for random ${\mathbb{F}}_q$-linear rank metric codes [@venkat2017], [@Gabriele]. An interesting direction is to see whether these new results can be applied to improve results on specific ${\mathbb{F}}_q$-linear rank metric codes. Due to its potential application in many fields, the specific ${\mathbb{F}}_q$-linear rank metric codes that we are interested in is the ${\mathbb{F}}_q$-linear self-orthogonal rank metric codes. Due to the finding that from list decoding point of view, random ${\mathbb{F}}_q$-linear rank metric codes perform as well as general rank metric codes, a natural question that follows is whether the performance can still be maintained when we further restrict that the random ${\mathbb{F}}_q$-linear rank metric codes to be also self-orthogonal. Moreover, based on ${\mathbb{F}}_q$-linear case, we investigate how well one can list decode random ${\mathbb{F}}_{q^m}$-linear self-orthogonal rank metric codes. Organization {#organization .unnumbered} ------------ Firstly, we give some definitions and notations about self-orthogonal rank metric codes, list decoding and quadratic form. Then, we describe how to construct ${\mathbb{F}}_q$ and ${\mathbb{F}}_{q^m}$-linear self-orthogonal rank metric codes based on the quadratic form and analyze the list decodability of them. Ultimately, we draw a conclusion. Preliminaries ============= Defined Self-Orthogonal Rank Metric Codes ----------------------------------------- Rank metric codes can mainly be interpreted in two different representations. The first representation is to deem each codeword as matrices in ${\mathbb{F}}_q^{n\times m}.$ Alternatively, we can interpret each element of a rank metric code as a vector in ${\mathbb{F}}_{q^m}^n.$ In the first representation of codewords as matrices, linear codes are considered to be linear over ${\mathbb{F}}_q.$ On the other hand, the linearity considered when seeing a rank metric code as a set of vectors is assumed to be ${\mathbb{F}}_{q^m}$ linearity. The two different representations of rank metric codes provide us with two different ways in defining self-orthogonal rank metric codes (${\mathbb{F}}_q$-linear and ${\mathbb{F}}_{q^m}$-linear). ### ${\mathbb{F}}_q$-linear self-orthogonal rank metric codes {#mathbbf_q-linear-self-orthogonal-rank-metric-codes .unnumbered} To properly define an ${\mathbb{F}}_q$-linear self-orthogonal rank metric code, first we briefly provide the definitions and notations of matrix representation for rank metric codes. A rank metric code $\mC$ contains $n\times m$ matrices over ${\mathbb{F}}_q$ for integers $n,m$ and prime power $q.$ Through the use of matrix transpose, for simplicity, we can just assume that $n$ is at most $m.$ The rank distance between two matrices $X,Y\in{\mathbb{F}}_q^{n\times m}$ is then defined by the rank norm of the difference between the two, $d_R(X,Y):=\rank(X-Y).$ These parameters can then be used to formulate two relative parameters, namely relative minimum rank distance $\delta$ and rate $R.$ These relative parameters are then defined by $$\delta(\mC)=\frac{\min_{X\neq Y\in \mC}\{d_R(X,Y)\}-1}{n}\quad \mathrm{and} \quad R(\mC)=\frac{\log_q |\mC|}{mn}.$$ The (Delsarte) dual of $\mC$ is then defined to be $$\mC^\perp = \{X\in{\mathbb{F}}_q^{n\times m}| Tr(CX^T) = 0, \forall C\in\mC\}.$$ Based on this definition of dual, an ${\mathbb{F}}_q$-linear rank metric code $\mC$ is said to be self-orthogonal if $\mC\subseteq \mC^\perp.$ A property of ${\mathbb{F}}_q$-linear self-orthogonal rank metric code that is readily verified is that its ${\mathbb{F}}_q$-dimension should be at most $\frac{nm}{2}$ and hence its rate $R$ must be in the range $0\leq R\leq 1/2.$ ### ${\mathbb{F}}_{q^m}$-linear self-orthogonal rank metric codes {#mathbbf_qm-linear-self-orthogonal-rank-metric-codes .unnumbered} Pick ${\mathbb{F}}_{q^m},$ the extension field of ${\mathbb{F}}_q$ with degree $m.$ The ring isomorphism between ${\mathbb{F}}_{q^m}$ and ${\mathbb{F}}_q^m$ through the use of a fixed ${\mathbb{F}}_q$-basis of ${\mathbb{F}}_{q^m}$ implies the possibility to identify a rank metric codes over ${\mathbb{F}}_q^{n\times m}$ as a collection of vectors over ${\mathbb{F}}_{q^m}^n.$ For any two vectors $x=(x_1,\cdots x_n)$ and $y=(y_1,\cdots,y_n)\in{\mathbb{F}}_{q^m}^n,$ we say that they are orthogonal to each other if $\langle x,y\rangle = \sum_{i=1}^n x_iy_i=0.$ The vector $x$ is then called a self-orthogonal vector if it is orthogonal to itself. The definition of self-orthogonality can be naturally extended to a set $\{v_1,\cdots, v_t\}$ where this set is self-orthogonal if $\langle v_i,v_j\rangle=0$ for any choices of $i$ and $j,$ $1\leq i, j\leq t.$ Clearly, this suggests an alternative way to define a dual $\mC^\perp$ of a rank metric code $\mC\in{\mathbb{F}}_{q^m}^n,$ namely the collection of codewords that are orthogonal to all codewords in $\mC.$ Analogous to the previous definition, we call an ${\mathbb{F}}_{q^m}$-linear rank metric code $\mC$ to be self-orthogonal given that $\mC\subseteq \mC^\perp.$ Consequentially, an ${\mathbb{F}}_{q^m}$-linear rank metric code $\mC$ can only be self-orthogonal if its ${\mathbb{F}}_{q^m}$-dimension does not exceed $\frac{n}{2}$ or it has rate $R\in(0,1/2).$ We will devote the remainder of this section to investigate when both definition of duals can coincide. In order to do that, we first need to review some basic concepts of algebra. Recall that for the field ${\mathbb{F}}_{q^m},$ there are $n$ different ${\mathbb{F}}_q$-linear automorphisms which are the Frobenius automorphisms $x\mapsto x^{q^{i}}$ for $i=0,\cdots,n-1.$ Based on these automorphisms, we can then define the field trace $tr_{{\mathbb{F}}_{q^m}/{\mathbb{F}}_q}$ which sends $x$ to $\sum_{i=0}^{m-1} x^{q^i}.$ We also recall that for any given ${\mathbb{F}}_q$ basis $\mB=(\beta_1,\cdots, \beta_m)$ of ${\mathbb{F}}_{q^m},$ there always exists a dual bases $\mB^\ast : = (\beta_1^{\ast} , \cdots \beta_m^{\ast})$ satisfying $tr_{{\mathbb{F}}_{q^m}/{\mathbb{F}}_q}(\beta_i\beta_j^\ast)=\delta_{i,j}$ the Kronecker delta function. Now if $\beta_i$ coincides with $\beta_i^\ast$ for all $i,$ $\mB$ is called a self-dual basis. Note that the existence of a self-dual ${\mathbb{F}}_q$-basis of ${\mathbb{F}}_{q^m}$ is equivalent to the condition that $q$ is even or both $q$ and $m$ are odd [@Gabriele]. We are ready to conduct our investigations. Let $q$ and $m$ be chosen such that ${\mathbb{F}}_{q^m}$ has a self dual basis $\mB=(\beta_1,\cdots, \beta_m).$ Furthermore, pick any two rank metric codes $\mC_1$ and $\mC_2$ over ${\mathbb{F}}_q^{n\times m}.$ Then $$Tr(\mC_1\mC_2^T)=\{Tr(XY^T): X\in \mC_1, Y\in \mC_2\}=\{0\}$$ if and only if $$\langle\mC_1,\mC_2\rangle =\{\langle x,y\rangle: x\in \mC_1, y\in\mC_2\}=\{0\}.$$ Note that we assume $\mC_i$ to be in its matrix representation for the earlier equation and vector representation for the latter. Let $\mathbf{a}=(a_1,\cdots,a_n)\in\mC_1$ and $\mathbf{b}=(b_1,\cdots b_n)\in\mC_2.$ We further let $a_i=\sum_{j=1}^m a_{i,j}\beta_j$ and $b_i=\sum_{k=1}^mb_{i,k} \beta_k$ for all $i=1,2\cdots, n.$ Assuming $\langle\mC_1,\mC_2\rangle=\{0\},$ we have $$\begin{aligned} 0=\langle\mathbf{a}, \mathbf{b}\rangle=\sum_{i=1}^na_{i}b_{i}&=&\sum_{i=1}^n\left(\sum_{j=1}^ma_{i,j}\beta_{j}\right)\left(\sum_{k=1}^mb_{i,k}\beta_{k}\right)=\sum_{i=1}^n\sum_{j=1}^m\sum_{k=1}^ma_{i,j}b_{i,k}\beta_{j}\beta_{k}.\end{aligned}$$ Applying the field trace function to both sides, we have $$\begin{aligned} 0=tr\langle\mathbf{a}, \mathbf{b}\rangle=\sum_{i=1}^n\sum_{j=1}^m\sum_{k=1}^ma_{i,j}b_{i,k}tr(\beta_{j}\beta_{k})=\sum_{i=1}^n\sum_{j=1}^m\sum_{k=1}^ma_{i,j}b_{i,k}=\sum_{i=1}^n\sum_{j=1}^ma_{i,j}b_{i,j}.\end{aligned}$$ Note that by assumption, the matrices $A=(a_{i,j}), B=(b_{i,j})$ with size $n\times m$ are the matrix representations of $\mathbf{a}$ and $\mathbf{b}$ respectively. Noting that $Tr(AB^T)=\sum_{i=1}^n\sum_{j=1}^m a_{i,j}b_{i,j},$ the last equality implies that $Tr(AB^T)=0.$ Since $\mathbf{a}$ and $\mathbf{b}$ are arbitrary elements of $\mC_1$ and $\mC_2$ respectively, we have the conclusion $Tr(\mC_1\mC_2^T)=\{0\}$ from the assumption that $\langle\mC_1,\mC_2\rangle=\{0\}.$ On the other direction can be easily proved. Suppose that $Tr(AB^T)=0$ for all $A=(a_{i,j})\in\mC_1$ and $B=(b_{i,j})\in\mC_2.$ Hence, we have $B\in\mC_1^\perp$ or $\langle A,B\rangle=0$ for all $A\in\mC_1$ and $B\in\mC_2$ which ultimately implies that $\langle\mC_1,\mC_2\rangle=\{0\}.$ List Decoding ------------- Given a received word, list decoding outputs a list of codewords, if successful, it contains the correct transmitted codeword. Analogous to the Hamming ball in classical block codes, in rank metric codes, we also have the concept of rank metric ball. The formal definition is given in the following. Let $\tau\in(0,1)$ and $X\in {\mathbb{F}}_{q}^{n\times m}$. The rank metric ball centre at $X$ and radius $\tau n$ is defined by $$\mB_{R}(X,\tau n):=\{Y\in {\mathbb{F}}_{q^m}^n:d_R(X,Y)\leq \tau n\}.$$ For any $n$-dimensional vector space $V$ over ${\mathbb{F}}_q,$ we denote by ${n \brack k}_q,$ the number of subspaces of $V$ with dimension $k.$ This is called the Gaussian binomial coefficient and it has the following explicit formula, $${n\brack k}_{q}=\prod_{i=0}^{k-1}\frac{q^{n}-q^{i}}{q^{k}-q^{i}}.$$ It can be verified that this formula ${n\brack k}_q$ has the following bounds that can be used as estimation [@Gad2008], $$q^{k(n-k)}\leq{n\brack k}_{q}\leq 4 q^{k(n-k)}.$$ For an integer $L\geq 1$ and a real $\tau\in (0,1)$, a rank metric code $\mC$ is said to be $(\tau n, L)$-list decodable if for every $X\in {\mathbb{F}}_q^{n\times m}$ $$|\mB_R(X,\tau n)\cap \mC|\leq L.$$ Analogously, for $x\in{\mathbb{F}}_{q^m}^n,$ $$|\mB_R(x,\tau n)\cap \mC|\leq L.$$ Quadratic forms --------------- We say $f(x)$ is an $n$-variate quadratic form over ${\mathbb{F}}_q,$ if it is a degree $2$ homogeneous multinomial of $n$ variables with coefficients from ${\mathbb{F}}_q.$ The general formula that $f$ should follow $$f(x)=f(x_1,x_2,\cdots,x_n)=\sum_{i,j=1}^na_{ij}x_ix_j,a_{ij}\in\mathbb{F}_q.$$ Note that an $n$-variate quadratic form $f$ over ${\mathbb{F}}_q$ can be expressed as multiplication of matrices. Assuming $x=(x_1,\cdots, x_n)^T$ and $A=(a_{i,j})_{i,j=1,\cdots, n}$ over ${\mathbb{F}}_q,$ $f(x)$ can then be rewritten as $f(x) = x^T Ax.$ Two quadratic forms $f$ and $g$ of $n_1$ and $n_2$ indeterminates respectively are called equivalent provided that we can find a full rank $n_1\times n_2$ matrix $M$ over ${\mathbb{F}}_q$ satisfying $f(xM) = g(x).$ Note that equivalence implies the same number of roots. Equivalence enables two quadratic forms of different number of indeterminates to be closely related to each other. Given a non-zero quadratic form $f(x),$ the smallest number of indeterminates that a quadratic form $g(x)$ can have while still being equivalent to $f(x)$ is a parameter of $f(x)$ that is called to be its rank. By convention, we let the zero quadratic form be rank $0.$ A non-zero quadratic form $f(x)$ is said to be non-degenerate if its rank is equal to the number of its indeterminates. To aid our analysis in this paper, the number of roots a quadratic form $f(x)$ is a topic of interest. We combine several results in [@finitefields] as a lemma to consider two cases (over ${\mathbb{F}}_q$ and ${\mathbb{F}}_{q^m}$).   [@finitefields] \[lemma:4.1\] Let $f(x):=f(x_1,x_2,\cdots,x_n)$ be a quadratic form with rank $r$ over $\mathbb{F}_{q}$ (or ${\mathbb{F}}_{q^m}$). $N(f(x)=0)$ denotes the number of roots of $f(x)=0$ in $\mathbb{F}_{q}^{n\times m}$ (or ${\mathbb{F}}_{q^m}^n$). If $r=0,$ then we have $N(f(x)=0)=q^{mn}.$ If $r\geq 1,$ then we have the following results: If $f(x)$ is defined over ${\mathbb{F}}_q$ with solutions in ${\mathbb{F}}_q^{n\times m},$ $$\begin{aligned} N(f(x)=0)= \begin{cases} q^{mn-1}, &r~ is~ odd.\cr q^{mn-1}\pm(q-1)q^{mn-r/2-1}, &r~ is~ even. \end{cases}\end{aligned}$$ Alternatively, if $f(x)$ is defined over ${\mathbb{F}}_{q^m}$ with solutions in ${\mathbb{F}}_{q^m}^n,$ $$\begin{aligned} N(f(x)=0)= \begin{cases} q^{m(n-1)}, &r~ is~ odd.\cr q^{m(n-1)}\pm(q^m-1)q^{m(n-r/2-1)}, &r~ is~ even. \end{cases}\end{aligned}$$ Construction of Random Self-Orthogonal Rank Metric Codes ======================================================== Construct ${\mathbb{F}}_q$-linear Self-Orthogonal Rank Metric Codes ------------------------------------------------------------------- In this part, we construct ${\mathbb{F}}_q$-linear self-orthogonal rank metric codes based on quadratic forms. Let $A=[a_{ij}], 1\leq i\leq n, 1\leq j\leq m$ be a word. If $A$ is self-orthogonal, then $$Tr(AA^T)=\sum_{i=1}^{n}\sum_{j=1}^{m}a_{ij}a_{ij}=0.$$ Considering the standard bijection from $[n]\times [m]$ to $[nm],$ where $[n]=\{0,\cdots, n-1\}, [m]=\{0,\cdots, m-1\},$ we can rewrite the double index $(i,j)$ to a single index to obtain $$\begin{aligned} Tr(AA^T)&=&\sum_{i=1}^{n}\sum_{j=1}^{m}a_{ij}a_{ij}=\sum_{\ell=1}^{mn}a_\ell^2=0. \end{aligned}$$ [**Construction**]{} - [Step $1.$ Choose a nonzero random solution $A_1\in\mathbb{F}_{q}^{n\times m}$ of the quadratic equation $x_1^2+x_2^2+\cdots+x_{mn}^2=0.$]{} By Lemma \[lemma:4.1\] we can obtain that the above equation has at least $q^{mn-2}$ solutions, so a self-orthogonal word $A_1$ can be found. - [Step $2.$ Obtain a linearly independent set $\{A_1,A_2,\cdots,A_{k-1},A_k\}$ of random self-orthogonal matrices given $A_1,\cdots,A_{k-1}.$]{} [Firstly, we assume that a linearly independent set $\{A_1, A_2,\cdots, A_{k-1}\}$ of random self-orthogonal matrices has already been found, i.e. $Tr(A_iA_j^T)=\sum_{\ell=1}^{mn}a_{i_\ell}a_{j_\ell}=0, 1\leq i,j \leq k-1$. Then, if we want to find the $k$-th matrix $A_k,$ then we need to find a solution of the following equations $$\left\{ \begin{aligned}~\label{eq:4.1} a_{11}x_1+a_{12}x_2+\cdots+a_{1,mn}x_{mn}&=&0, \\ \vdots \\ a_{{k-1},1}x_1+a_{{k-1},2}x_2+\cdots+a_{{k-1},mn}x_{mn}&=&0,\\ x_1^2+x_2^2+\cdots+x_{mn}^2&=&0. \end{aligned} \right.$$]{} Take the first $k-1$ equations above in to the last one, we have a quadratic equation $g(x_{i_1}, x_{i_2},\cdots, x_{i_{mn-k+1}})$ of $mn-k+1$ variables. So, $N(g(x_{i_1}, x_{i_2},\cdots, x_{i_{mn-k+1}})=0)$ is the number of solutions of the equation (\[eq:4.1\]). The number of the cardinality of $\rm{span}\{A_1, A_2, \cdots, A_{k-1}\}$ is equal to $q^{mn(k-1)}.$ And, $N(g(x_{i_1}, x_{i_2},\cdots, x_{i_{mn-k+1}})=0)>q^{mn(k-1)},$ thus we can randomly choose a solution $A_k$ from (\[eq:4.1\]), it is not contained in $\rm{Span}\{A_1, A_2, \cdots, A_{k-1}\}.$ So, we can obtain a linearly independent set $\{A_1, A_2, \cdots, A_{k-1}, A_k\}$ of random self-orthogonal matrices. Moreover, by Lemma \[lemma:4.1\], the number of solution $N(g(x_{i_1}, x_{i_2},\cdots, x_{i_{mn-k+1}})=0)$ of $g(x_{i_1}, x_{i_2},\cdots, x_{i_{mn-k+1}})=0$ is at least $q^{mn-k-1}.$ Thus, [the set can always be constructed as long as]{} $k\leq(mn-1)/2.$ Construct ${\mathbb{F}}_{q^m}$-linear Self-Orthogonal Rank Metric Codes ----------------------------------------------------------------------- We study how to construct ${\mathbb{F}}_{q^m}$-linear self-orthogonal rank metric codes. The idea is similar to the construction of ${\mathbb{F}}_{q}$-linear self-orthogonal rank metric codes. Constructing a random ${\mathbb{F}}_{q^m}$-linear self-orthogonal rank metric code is equivalent to find a linearly independent set $\{x_1, x_2,\cdots, x_k\}$ of random ${\mathbb{F}}_{q^m}$-linear self-orthogonal vectors, where $x_i\in{\mathbb{F}}_{q^m}, 1\leq i\leq k.$ Choose a nonzero random solution $x_1=(x_{11}, x_{12},\cdots, x_{1n})\in\mathbb{F}_{q^m}^n$ of the quadratic equation $z_1^2+z_2^2+\cdots+z_n^2=0.$ This equation has at least $q^{m(n-2)}$ roots, so a self-orthogonal $x_1$ can be found. The same method as the construction ${\mathbb{F}}_q$-linear self-orthogonal rank metric codes can be conducted. Then, we can confirm there exists a linearly independent set $\{x_1, x_2, \cdots, x_{k-1}, x_k\}$ of ${\mathbb{F}}_{q^m}$-linear self-orthogonal vectors. In addition, by calculation we have such $x_k$ as long as $k\leq(n-1)/2.$ List Decoding Self-Orthogonal Rank Metric Codes =============================================== List Decoding ${\mathbb{F}}_{q}$-linear Self-Orthogonal Rank Metric Codes ------------------------------------------------------------------------- In this part, we investigate the list decodability of ${\mathbb{F}}_{q}$-linear self-orthogonal rank metric codes. We show that rate and decoding radius of ${\mathbb{F}}_{q}$-linear self-orthogonal rank metric codes can achieve the Gilbert-Varshamov bound. From now on, the information rate $\frac{log_q|\mC|}{mn}$ and the ratio $\frac{n}{m}$ are denoted by $R$ and $\rho,$ respectively. Our main result of list decoding of ${\mathbb{F}}_q$-linear self-orthogonal rank metric codes can be found in the Theorem \[thm:4.6\]. With the help of studying and discussing the weight distribution of certain rank metric code, we can deal with it.  [@venkat2017]\[lemma:4.7\] For all integers $n\leq m$, every $\tau\in(0,1)$ and $\ell=O(\sqrt{nm}),$ there exists a constant $C_{\tau,q}>1$ such that if $X_1,\cdots, X_\ell$ are selected independently and uniformly at random from $B_R(0,\tau n),$ then we have $$Pr[\mid \rm{span}\{X_1,\cdots, X_\ell\}\cup B_R(0,\tau n)\mid\geq C_{\tau,q}\cdot \ell]\leq q^{-(3-O(1))mn}$$ From the above lemma, it reveals that randomly picking $\ell$ words from $B_R(0, \tau n),$ there exists more than $\Omega(\ell)$ words in the span of $\ell$ words lies in the $B_R(0, \tau n)$ happens with a very small probability, where the parameter $\ell$ depends on the list size $L.$ Then, we consider the following result on the probability that a random $k$ dimension ${\mathbb{F}}_q$-linear rank metric code contains a $k-1$ dimension ${\mathbb{F}}_q$-linear self-orthogonal subcode and a given set $\{X_1,\cdots,X_\ell\}\subseteq {\mathbb{F}}_q^{n\times m}$ of linearly independent vectors. Let $\mC^*_k$ present the set of $k$ dimension ${\mathbb{F}}_q$-linear rank metric codes, where every code contains a $k-1$ dimension ${\mathbb{F}}_q$-linear self-orthogonal subcode.  [@L.F.J2015]\[lemma:4.8\] For any $\mathbb{F}_{q}$-linearly independent words $X_1, X_2,\cdots, X_\ell$ in ${\mathbb{F}}_{q}^{n\times m}$ with $\ell\leq k <mn/2,$ the probability of a random code $\mathcal{C}^*$ from $\mathcal{C}^*_k$ contains $\{X_1,X_2,\cdots, X_\ell\}$ is $$Pr_{\mathcal{C}^*\in \mathcal{C}^*_k}[\{X_1,X_2,\cdots, X_\ell\}\subseteq \mathcal{C}]\leq\left\{ \begin{aligned} {q}^{((k+\ell-mn-1)\ell+2k-1)}, ~~if~~ q~~ is~~ even;\\ {q}^{((k+\ell-mn-2)\ell+4k-2)}, ~~if~~ q~~ is~~ odd. \end{aligned} \right.$$ Thus, we have $$Pr_{\mathcal{C}^*\in \mathcal{C}^*_k}[\{X_1,X_2,\cdots, X_\ell\}\subseteq \mathcal{C}]\leq {q}^{((k+\ell-mn-2)\ell+4k-1)}.$$ Based on the Lemma \[lemma:4.7\] and Lemma \[lemma:4.8\], we prove Theorem \[thm:4.6\].  \[thm:4.6\] Let $q$ be prime power and $\tau\in(0,1).$ There exist a constant $M$ and all large enough $n$, for small $\epsilon>0,$ an ${\mathbb{F}}_q$-linear self-orthogonal rank metric code $\mC\subseteq {\mathbb{F}}_q^{n\times m}$ of rate $R=(1-\tau)(1-\rho\tau)-\epsilon$ is ${(\tau n, O_{\tau, q}(\frac{1}{\epsilon}))}$-list decodable with high probability at least $1-q^{-2mn}$. Pick $M=5C_{\tau,q},$ where $C_{\tau,q}$ is the constant in Lemma \[lemma:4.7\]. Set $L=\lceil{\frac{M}{\epsilon}}\rceil$ and $n$ to be large enough. Let $\mC$ be an ${\mathbb{F}}_{q}$-linear self-orthogonal rank metric codes with $\dim_{{\mathbb{F}}_{q}}=Rmn$ in ${\mathbb{F}}_{q}^{n\times m}$, the size $|\mC|=q^{Rmn}.$ We want to show that with high probability, $\mC$ is ${(\tau, O_{\tau,q}(\frac{1}{\epsilon}))}$-list decodable. In other words, the code $\mC$ is not ${(\tau n, O_{\tau,q}(\frac{1}{\epsilon}))}$-list decodable with low probability, i.e., $$~\label{eq:4.2} Pr_{\mC\in\mC_{Rmn}}[\exists X\in{\mathbb{F}}_{q}^{n\times m},~|B_R(X,\tau n)\cap \mC|\geq L]< q^{-2mn},$$ where $\mC_{Rmn}$ denotes the set of ${\mathbb{F}}_{q}$-linear self-orthogonal rank metric codes with dimension $Rmn.$ Let $X\in{\mathbb{F}}_{q}^{n\times m}$ be picked uniformly at random, define $$\triangle:=Pr_{{\mC\in\mC_{Rmn}}, X\in{\mathbb{F}}_{q^{n\times m}}}[|B_R(X,\tau n)\cap\mC|\geq L]$$ For (\[eq:4.2\]), it suffices to prove $$~\label{eq:4.3} \triangle< q^{-2mn}\cdot q^{-(1-R)mn}$$ The inequality (\[eq:4.3\]) is derived from (\[eq:4.2\]). For every ${\mathbb{F}}_{q}$-linear code $\mC$, due to “bad” case $X$ such that ($|B_R(X,\tau n)\cap\mC|\geq L$), there are $q^{Rmn}$ such ”bad” $X.$ Since $\mC$ is ${\mathbb{F}}_q$-linear, we have $$\begin{aligned} \triangle=&Pr_{{\mC\in\mC_{Rmn}}, {X\in{\mathbb{F}}_{q}^{n\times m}}}[|B_R(X,\tau n)\cap\mC|\geq L]\\ =&Pr_{{\mC\in\mC_{Rmn}}, {X\in{\mathbb{F}}_{q}^{n\times m}}}[|B_R(0,\tau n)\cap(\mC+X)|\geq L]\\ \leq&Pr_{{\mC\in\mC_{Rmn}}, {X\in{\mathbb{F}}_{q}^{n\times m}}}[|B_R(0,\tau n)\cap \rm Span_{{\mathbb{F}}_{q}}(\mC+X)|\geq L]\\ \leq&Pr_{{\mC^*\in\mC^*_{Rmn+1}}}[|B_R(0,\tau n)\cap \mC^*|\geq L],\end{aligned}$$ where $\mC^*$ is a $Rmn+1$ dimension random ${\mathbb{F}}_{q}$-subspace of ${\mathbb{F}}_{q}^{n\times m}$ containing $\rm Span_{{\mathbb{F}}_{q}}(\mC+X)$ ( If $X$ is not in $\mC,$ then $\mC^*=\rm Span_{{\mathbb{F}}_{q}}(\mC, X);$ otherwise $\mC^*=\rm Span_{{\mathbb{F}}_{q}}(\mC, Y),$ where word $Y$ is randomly picked from ${\mathbb{F}}_{q}^{n\times m}\backslash \mC$). For each integer $\ell \in[log_{q} L, L],$ let $\mathcal{F}_\ell$ be the set of all tuples $(X_1, X_2, \cdots, X_\ell)\in B_R(0,\tau n)^\ell$ such that $X_1, X_2, \cdots, X_\ell$ are linearly independent and $|\rm span(X_1, \cdots, X_\ell)\cap B_R(0,\tau n)|\geq L.$ Let $\mathcal{F}=\bigcup_{\ell=\lceil{loq_{q}L} \rceil}^L \mathcal{F}_\ell.$ For each $X=(X_1,\cdots, X_\ell)\in\mathcal{F},$ let $\{X\}$ and $(X)$ denote the set $\{X_1,\cdots,X_\ell\}$ and the tuple $(X_1,\cdots, X_\ell),$ respectively. Claim that if $|B_R(0,\tau n)\cap \mC^*|\geq L,$ there must exist $(X)\in\mathcal{F}$ such that $\{X\}\subseteq \mC^*.$ Indeed, let $\{H\}$ be a maximal linearly independent subset of $B_R(0,\tau n)\cap \mC^*.$ If $|\{H\}|<L,$ then we have $\{X\}=\{H\}$. Otherwise, we have $\{X\}$ to be any subset of $\{H\}$ of size $L.$ Thus, $$\begin{aligned} \triangle\leq&Pr_{{\mC^*\in\mC^*_{Rmn+1}}}[|B_R(0,\tau n)\cap \mC^*|\geq L]\\ \leq&\sum_{(X)\in\mathcal{F}_\ell}Pr_{{\mC^*\in\mC^*_{Rmn+1}}}[\{X\}\subseteq\mC^*]\\ =&\sum_{\ell=\lceil{loq_{q}L} \rceil}^L\sum_{(X)\in\mathcal{F}_\ell}Pr_{{\mC^*\in\mC^*_{Rmn+1}}}[\{X\}\subseteq\mC^*]\\ =&\sum_{\ell=\lceil{loq_{q}L} \rceil}^L|\mathcal{F}_\ell|Pr_{{\mC^*\in\mC^*_{Rmn+1}}}[\{X\}\subseteq\mC^*]\\ \leq&\sum_{\ell=\lceil{loq_{q}L} \rceil}^L|\mathcal{F}_\ell|{q}^{(((Rmn+1)+\ell-mn-2)\ell+4(Rmn+1)-1)}.$$ The last inequality is from the Lemma \[lemma:4.8\]. We want to get a good bound of our probability, so we need to take a reasonable good upper bound for $|\mathcal{F}_\ell|.$ In [@venkat2017], we bound $|\mathcal{F}_\ell|$ relying on the value of the parameter $\ell$. - [Case $1.$ $\ell<\frac{5}{\epsilon}$]{}\ In this case, we have $\frac{|\mathcal{F}_\ell|}{|B_R(0,\tau n)|^\ell}$ is a lower bound on the probability that matrices $X_1, X_2, \cdots, X_\ell$ chosen independently and uniformly at random from the rank metric ball $B_R(0,\tau n)$ are $$\mid \rm span\{X_1,\cdots, X_\ell\}\cup B_R(0,\tau n)\mid\geq L.$$ By Lemma \[lemma:4.7\], the probability is at most $q^{-2mn}$, thus $$|\mathcal{F}_\ell|\leq |B_R(0,\tau n)^\ell|\cdot q^{-2mn}\leq\left(4q^{mn(\tau+\tau \rho-\tau^2\rho)}\right)^\ell\cdot q^{-2mn}.$$ - [Case $2.$ $\ell\geq\frac{5}{\epsilon}$]{}\ We have the simple bound of $$|\mathcal{F}_\ell|\leq |B_R(0,\tau n)^\ell|\leq\left(4q^{mn(\tau+\tau \rho-\tau^2\rho)}\right)^\ell.$$ Finally, taking the value of $R=(1-\tau)(1-\rho\tau)-\epsilon$ into the below inequality, $$\begin{aligned} \triangle \leq&\sum_{\ell=\lceil{loq_{q}L} \rceil}^L|\mathcal{F}_\ell|{q}^{(((Rmn+1)+\ell-mn-2)\ell+4(Rmn+1)-1)}\\ \leq&\sum_{\ell=\lceil{loq_{q}L} \rceil}^{\lceil\frac{5}{\epsilon}\rceil-1}|\mathcal{F}_\ell|{q}^{(((Rmn+1)+\ell-mn-2)\ell+4(Rmn+1)-1)}+\sum_{\ell=\lceil\frac{5}{\epsilon}\rceil}^L|\mathcal{F}_\ell|{q}^{(((Rmn+1)+\ell-mn-2)\ell+4(Rmn+1)-1)}\\=&q^{-2mn}q^{4Rmn}\cdot\sum_{\ell=\lceil{loq_{q}L} \rceil}^{\lceil\frac{5}{\epsilon}\rceil-1} 4^\ell q^{{mn\ell}(\tau+\tau \rho-\tau^2\rho+R-1)}+q^{4Rmn}\cdot\sum_{\ell=\lceil\frac{5}{\epsilon}\rceil}^L4^\ell q^{mn\ell(\tau+\tau \rho-\tau^2\rho+R-1)}\\=&q^{-2mn}q^{4Rmn}\cdot\sum_{\ell=\lceil{loq_{q}L} \rceil}^{\lceil\frac{5}{\epsilon}\rceil-1} 4^\ell q^{{mn\ell}(-\epsilon)}+q^{4Rmn}\cdot\sum_{\ell=\lceil\frac{5}{\epsilon}\rceil}^L4^\ell q^{mn\ell(-\epsilon)}\\\leq&q^{-2mn}q^{4Rmn}\cdot\sum_{\ell=\lceil{loq_{q}L} \rceil}^{\lceil\frac{5}{\epsilon}\rceil-1} 4^\ell q^{{mn\ell}(-\epsilon)}+q^{4Rmn}\cdot\sum_{\ell=\lceil\frac{5}{\epsilon}\rceil}^L4^\ell q^{-5mn}\\ \leq&q^{-2mn}.\end{aligned}$$ Thus, an ${\mathbb{F}}_q$-linear self-orthogonal rank metric code with rate $R=(1-\tau)(1-\rho\tau)-\epsilon$ is not ${(\tau n, O_{\tau,q}(\frac{1}{\epsilon}))}$-list decodable with an exponential small probability $q^{-2mn}.$ List Decoding ${\mathbb{F}}_{q^m}$-linear Self-Orthogonal Rank Metric Codes --------------------------------------------------------------------------- We consider the probability that a random dimension $k$ $\mathbb{F}_{q^m}$-linear code contains a self-orthogonal $k-1$ dimension subcode and a given set $\{v_1,v_2,\cdots, v_\ell\}\subseteq\mathbb{F}_{q^m}^n$ of linearly independent vectors. Let $\mathcal{C}^*_k$ present the set of ${{\mathbb{F}}_{q^m}}$-linear codes in which every code contains an ${\mathbb{F}}_{q^m}$-linear dimension $k-1$ self-orthogonal subcode.  [@L.F.J2015] For any $\mathbb{F}_{q^m}$-linearly independent vectors $x_1, x_2,\cdots, x_\ell$ in $\mathbb{F}_{q^m}^n$ with $\ell\leq k <n/2,$ the probability of a random code $\mathcal{C}^*$ from $\mathcal{C}^*_k$ contains $\{x_1,x_2,\cdots, x_\ell\}$ is $$Pr_{\mathcal{C}^*\in \mathcal{C}^*_k}[\{x_1,x_2,\cdots, x_\ell\}\subseteq \mathcal{C}]\leq {q^{m}}^{((k+\ell-n-2)\ell+4k-1)}.$$ Let $q$ be prime power and a real $\tau\in(0,1).$ There exist a constant $M$ and all large enough $n$, for small $\epsilon>0,$ an ${\mathbb{F}}_{q^m}$-linear self-orthogonal rank metric code $\mC\subseteq {\mathbb{F}}_{q^m}^{n}$ of $R=(1-\tau)(1-\rho\tau)-\epsilon$ is ${(\tau n, \exp(O_{\tau, q}(\frac{1}{\epsilon})))}$-list decodable with high probablility $1-q^{-2mn}.$ Put $L=\lceil{\frac{1}{\epsilon}}\rceil,$ $n$ is large enough. Let $\mC$ be an ${\mathbb{F}}_{q^m}$-linear self-orthogonal rank metric codes with $\dim_{{\mathbb{F}}_{q^m}}=Rn$ in ${\mathbb{F}}_{q^m}^n$, the size $|\mC|=q^{Rmn}.$ We want to show that $\mC$ is not ${(\tau, \exp( O_{\tau,q}(\frac{1}{\epsilon})))}$-list decodable, i.e., $$~\label{eq:4.42} Pr_{\mC\in\mC_{Rn}}[\exists x\in{\mathbb{F}}_{q^m}^n, ~|B_R(x,\tau n)\cap \mC|\geq L]<q^{-2mn},$$ where $\mC_{Rn}$ denotes the set of ${\mathbb{F}}_{q^m}$-linear self-orthogonal rank metric codes with dimension $Rn.$ Let $x\in{\mathbb{F}}_{q^m}^n$ be picked uniformly at random, define $$\triangle:=Pr_{{\mC\in\mC_{Rn}}, {x\in{\mathbb{F}}_{q^m}^n}}[|B_R(x,\tau n)\cap\mC|\geq L]$$ To prove inequality (\[eq:4.42\]), we need to show that $$~\label{eq:4.43} \triangle<q^{-2mn}\cdot q^{-(1-R)mn}$$ The inequality (\[eq:4.43\]) is derived from (\[eq:4.42\]). For every ${\mathbb{F}}_{q^m}$-linear $\mC,$ due to “bad” case $x$ such that $|B_R(x,\tau n)\cap\mC|\geq L,$ there are $q^{Rmn}$ such “bad” $x.$ Since $\mC$ is linear, we have $$\begin{aligned} \triangle=&Pr_{{\mC\in\mC_{Rn}}, {x\in{\mathbb{F}}_{q^m}^n}}[|B_R(x,\tau n)\cap\mC|\geq L]\\ =&Pr_{{\mC\in\mC_{Rn}}, {x\in{\mathbb{F}}_{q^m}^n}}[|B_R(0,\tau n)\cap(\mC+x)|\geq L]\\ \leq&Pr_{{\mC\in\mC_{Rn}}, {x\in{\mathbb{F}}_{q^m}^n}}[|B_R(0,\tau n)\cap \rm Span_{{\mathbb{F}}_{q^m}}(\mC+x)|\geq L]\\ \leq&Pr_{{\mC^*\in\mC^*_{Rn+1}}}[|B_R(0,\tau n)\cap \mC^*|\geq L],\end{aligned}$$ where $\mC^*$ is a random $Rn+1$ dimension ${\mathbb{F}}_{q^m}$-subspace of ${\mathbb{F}}_{q^m}^n$ containing $\mC$ ( If $x\notin \mC,$then $\mC^*=\rm Span_{{\mathbb{F}}_{q^m}}(\mC, x);$ otherwise $\mC^*=\rm Span_{{\mathbb{F}}_{q^m}}(\mC, y),$ where $y$ is picked randomly from ${\mathbb{F}}_{q^m}^n\backslash \mC$). For each integer $\ell$, $\ell\in[log_{q^m} L,L].$ Let $\mathcal{F}_\ell$ be the set of all tuples $(x_1, \cdots, x_\ell)\in B_R(0,\tau n)^\ell$ such that $x_1, \cdots, x_\ell$ are linearly independent and $$|\rm span(x_1, \cdots, x_\ell)\cap B_R(0,\tau n)|\geq L.$$ Hence, $$\begin{aligned} |\mathcal{F}_t|\leq |B_R(0,\tau n)^t|\leq\left(4q^{mn(\tau+\tau b-\tau^2b)}\right)^\ell.\end{aligned}$$ Let $\mathcal{F}=\bigcup_{\ell=\lceil{loq_{q^m}L} \rceil}^L \mathcal{F}_\ell.$ For each $x=(x_1,\cdots, x_\ell)\in\mathcal{F},$ let $\{x\}$ denote the set $\{x_1,\cdots,x_\ell\}.$ We claim that if $|B_R(0,\tau n)\cap \mC^*|\geq L,$ there must exist $x\in\mathcal{F}$ such that $\{x\}\subseteq \mC^*.$ Thus, we have $$\begin{aligned} \triangle\leq&Pr_{{\mC^*\in\mC^*_{Rn+1}}}[|B_R(0,\tau n)\cap \mC^*|\geq L]\\\leq&\sum_{x\in\mathcal{F}_\ell}Pr_{{\mC^*\in\mC^*_{Rn+1}}}[\{x\}\subseteq\mC^*]\\=&\sum_{\ell=\lceil{loq_{q^m}L} \rceil}^L\sum_{x\in\mathcal{F}_\ell}Pr_{{\mC^*\in\mC^*_{Rn+1}}}[\{x\}\subseteq\mC^*]\\=&\sum_{\ell=\lceil{loq_{q^m}L} \rceil}^L|\mathcal{F}_\ell|Pr_{{\mC^*\in\mC^*_{Rn+1}}}[\{v\}\subseteq\mC^*]\\$$ By taking $R=(1-\tau)(1-\rho\tau)-\epsilon$, we can obtain $$\begin{aligned} \triangle\leq&Pr_{{\mC^*\in\mC^*_{Rn+1}}}[|B_R(0,\tau n)\cap \mC^*|\geq L]\\\leq&\sum_{\ell=\lceil{loq_{q^m}L} \rceil}^L \left(4q^{mn(\tau+\tau \rho-\tau^2\rho)}\right)^\ell Pr_{{\mC^*\in\mC^*_{Rn+1}}}[\{x\}\subseteq\mC^*]\\\leq&\sum_{\ell=\lceil{loq_{q^m}L} \rceil}^L \left(4q^{mn(\tau+\tau \rho-\tau^2\rho)}\right)^\ell q^{m((Rn+1)+\ell-n-2)\ell+4(Rn+1)-1)}\\\leq&4^\ell\cdot\sum_{\ell=\lceil{loq_{q^m}L} \rceil}^L q^{mn\ell(\tau+\tau b-\tau^2 \rho+R-1)}\\ \leq&q^{-2mn}.\end{aligned}$$ Thus, an ${\mathbb{F}}_{q^m}$-linear self-orthogonal rank metric code with rate $R=(1-\tau)(1-\rho\tau)-\epsilon$ is not ${(\tau n, \exp(O_{\tau, q}(\frac{1}{\epsilon})))}$-list decodable with an exponential small probability $q^{-2mn}.$ Conclusion ========== We investigate the list decodable ${\mathbb{F}}_q$ and ${\mathbb{F}}_{q^m}$-linear self-orthogonal rank metric codes. We show that the list decodability of ${\mathbb{F}}_q$-linear self-orthogonal rank metric codes is as good as that of general random rank metric codes as well, which can be list decoded up to the Gilbert-Varshamov bound. By using the same methods for $\mathbb{F}_{q^m}$-linear rank metric codes, our results reveal that the ${\mathbb{F}}_{q^m}$-linear self-orthogonal rank metric codes is list decodable up to $R=(1-\tau)(1-\rho\tau)-\epsilon$ with exponential list size. The list size of the codes grows polynomially in $q^m$ (rather than just $q$). It is interesting to decrease the list size of ${\mathbb{F}}_{q^m}$-linear self-orthogonal rank metric codes and ${\mathbb{F}}_{q^m}$-linear rank metric codes. [99]{} A. Wachter-Zeh, Bounds on List Decoding Gabidulin Codes, in [*Int, Workshop Alg. Combin. Coding Theory*]{}, Jun. 2012, pp. 329-334. A. Wachter-Zeh, Bounds on list decoding of rank-metric codes, [*IEEE Transactions on Information Theory*]{}, 2013(59): 7268-7277. D. Silva, F. R. Kschischang and R. Köetter, A rank-metric apporach to error control in random network coding, [*IEEE Transactions on Information Theory*]{}, 2008(54): 3951-3967. E. M. Gabidulin, A. V. Paramonov, and O. V. Tretjakov, “Ideals over a non-commutative ring and their application in cryptology,” LNCS, Vol. 573, pp. 482–489, 1991. G. Nebe, W. Willems, On self-dual MRD codes, [*Advances in Mathematics of Communication*]{}, 2016(10): 633-642. J. M. Wozencraft, List decoding, Res. Lab. Electron., MIT, Cambridge, MA, USA, Quart. Prog. Rep., 1958. L. Jin, C. Xing and X. Zhang, On the List-Decodability of Random Self-Orthogonal Codes, [*IEEE Transactions on Information Theory*]{}, Vol.61, No.2, Feb. 2015. M. Gadouleau and Z. Y. Yan, Packing and covering peoperties of rank metric codes, [*IEEE Transactions on Information Theory*]{}, 2008(54): 3873-3883. S. Liu, C. Xing and C. Yuan, List Decodability of Random Subcodes of Gabidulin Codes, [*IEEE Transactions on Information Theory*]{}, Vol.63, No.1, Jan 2017. P. Elias, List decoding for noisy channels, Res. Lab. Electron., MIT, Cambridge, MA, USA, Tech. Rep. 335, 1957. P. Elias, Error-correcting codes for list decoding, [*IEEE Transactions on Information Theory*]{}, Vol.37, No.1, pp. 5-12, Jan. 1991. P. Lusina, E. M. Gabidulin and M. Bossert, Maximum rank distance codes as space-time codes, [*IEEE Transactions on Information Theory*]{}, 2003(49): 2757-2760. R. Köetter and F. R. Kschischang, Coding for Errors and Erasures in Random Network Coding, [*IEEE Transactions on Information Theory*]{}, 2008(54): 3579-3591. R. Lidl and H. Neiderriter, [*Finite Fields,*]{} Cambridge, U.K. Cambridge University. Press, 1997. R. M. Roth, Maximum Rank Array Codes and their Application to Crisscross Error Correction, [*IEEE Transactions on Information Theory*]{}, 1991(37): 328-336. R. Overbeck, Brute-force attacks public key cryptosystem based on Gabidulin codes, [*J. Cryptography*]{}, 2008(21): 280-301. V. Guruswami, N. Resch, On the List-Decodability of Random Linear Rank Metric Codes, 2017. Y. Ding, On List-Decodability of Random Rank Metric Codes and Subspace Codes, [*IEEE Transactions on Information Theory*]{}, 2015(61): 51-59. [^1]: The author is with Division of Mathematical Sciences, School of Physical and Mathematical Sciences, Nanyang Technological University, Singapore 637371, Republic of Singapore (email: [email protected]).

ArXiv

--- abstract: 'The quantum duality principle (QDP) for homogeneous spaces gives four recipes to obtain, from a quantum homogeneous space, a dual one, in the sense of Poisson duality. One of these recipes fails (for lack of the initial ingredient) when the homogeneous space we start from is not a quasi-affine variety. In this work we solve this problem for the quantum Grassmannian, a key example of quantum projective homogeneous space, providing a suitable analogue of the QDP recipe.' --- [ ]{} -1cm **Quantum Duality Principle** **for Quantum Grassmannians** R. Fioresi$^{\,\flat,}$[[^1]]{}, F. Gavarini$^{\,\sharp}$ *$^\flat$ Dipartimento di Matematica, Università di Bologna* *piazza di Porta S. Donato, 5 — I-40127 Bologna, ITALY* [e-mail: [email protected]]{} *$^\sharp$ Dipartimento di Matematica, Università di Roma “Tor Vergata”* *via della ricerca scientifica 1 — I-00133 Roma, ITALY* [e-mail: [email protected]]{} 1,3cm [  [*Quantum Grassmann Varieties*]{}.]{} [2000 [*MSC:*]{}  Primary 20G42, 14M15; Secondary 17B37, 17B62.]{} Introduction {#intro} ============ In the theory of quantum groups, the geometrical objects that one takes into consideration are affine algebraic Poisson groups and their infinitesimal counterparts, namely Lie bialgebras. By “quantization” of either of these, one means a suitable one-parameter deformation of one of the Hopf algebras associated with them. They are respectively the algebra of regular function $ {\mathcal{O}}(G) \, $, for a Poisson group $ G $, and the universal enveloping algebra $ U({\mathfrak{g}}) $, for a Lie bialgebra $ {\mathfrak{g}}\, $. Deformations of $ {\mathcal{O}}(G) $ are called [*quantum function algebras*]{} (QFA), and are often denoted with $ {\mathcal{O}}_q(G) \, $, while deformations of $ U({\mathfrak{g}}) $ are called [*quantum universal enveloping algebras*]{} (QUEA), denoted with $ U_q({\mathfrak{g}}) \, $. The quantum duality principle (QDP), after its formulation in [@ga2; @ga3; @ga4], provides a recipe to get a QFA out of a QUEA, and vice-versa. This involves a change of the underlying geometric object, according to Poisson duality, in the following sense. Starting from a QUEA over a Lie bialgebra $ \, {\mathfrak{g}}= \text{\it Lie}\,(G) \, $, one gets a QFA for a dual Poisson group $ G^* \, $. Starting instead from a QFA over a Poisson group $ G \, $, one gets a QUEA over the dual Lie bialgebra $ {\mathfrak{g}}^* $. In [@cg], this principle is extended to the wider context of homogeneous Poisson $ G $–spaces. One describes these spaces, in global or in infinitesimal terms, using suitable subsets of $ {\mathcal{O}}(G) $ or of $ U({\mathfrak{g}}) \, $. Indeed, each homogeneous $ G $–space $ M $ can be realized as $ G \big/ K $ for some closed subgroup $ K $ of $ G $ (this amounts to fixing a point in $ M \, $: it is shown in [@cg], §1.2, how to select such a point). Thus we can deal with either the space or the subgroup. Now, $ K $ can be coded in infinitesimal terms by $ U(\mathfrak{k}) $, where $ \, \mathfrak{k} := \text{\it Lie}\,(K) \, $, and in global terms by $ \, \mathcal{I}(K) := \big\{\, \varphi \!\in\! {\mathcal{O}}(G) \,\big|\, \varphi(K) = 0 \big\} \, $, the defining ideal of $ K \, $. Instead, $ G \big/ K $ can be encoded infinitesimally by $ \, U({\mathfrak{g}}) \, \mathfrak{k} \, $ and globally by $ \, {\mathcal{O}}\big(G\big/K\big) \equiv {{\mathcal{O}}(G)}^K $, the algebra of $ K $–invariants in $ {\mathcal{O}}(G) \, $. Note that $ \, U({\mathfrak{g}}) \big/ U({\mathfrak{g}}) \, \mathfrak{k} \, $ identifies with the set of left-invariant differential operators on $ G\big/K \, $, or the set of $ K $–invariant, left-invariant differential operators on $ G \, $. These constructions [*all*]{} make sense in formal geometry, i.e. when dealing simply with formal groups and formal homogeneous spaces, as in [@cg]. Instead, if one looks for [*global*]{} geometry, then one construction might fail, namely the description of $ G \big/ K \, $ via its function algebra $ \, {\mathcal{O}}\big(G\big/K\big) = {{\mathcal{O}}(G)}^K \, $. In fact, this makes sense — i.e., $ {{\mathcal{O}}(G)}^K $ is enough to describe $ G \big/ K $ — if and only if the variety $ G \big/ K $ is [*quasi-affine*]{}. In particular, this is not the case if $ G \big/ K $ is projective, like, for instance, when $ G \big/ K $ is a Grassmann variety. By “quantization” of the homogeneous space $ G \big/ K $ one means any quantum deformation (in suitable sense) of any one of the four algebraic objects mentioned before which describe either $ G \big/ K $ or $ K \, $. Moreover one requires that given an infinitesimal or a global quantization for the group $ G $, denoted by $ U_q({\mathfrak{g}}) $ or $ {\mathcal{O}}_q(G) $ respectively, the quantization of the homogeneous space admits a $ U_q({\mathfrak{g}}) $–action or a $ {\mathcal{O}}_q(G) $–coaction respectively, which yields a quantum deformation of the algebraic counterpart of the $ G $–action on $ G \big/ K \, $. The QDP for homogeneous $ G $–spaces (cf. [@cg]) starts from an infinitesimal (global) quantization of a $ G $–space, say $ G \big/ K $, and provides a global (infinitesimal) quantization for the Poisson dual $ G^* $–space. The latter is $ \, G^* \! \big/ K^\perp $ (with $ \, \text{\it Lie}\,\big(K^\perp\big) = \mathfrak{k}^\perp \, $, the orthogonal subspace — with respect to the natural pairing between $ {\mathfrak{g}}$ and its dual space $ {\mathfrak{g}}^* $ — to $ \mathfrak{k} $ inside $ {\mathfrak{g}}^* \, $). In particular, the principle gives a concrete recipe $${\mathcal{O}}_q\big(G\big/K\big) \;\; \circ\hskip-5,3pt\relbar\relbar \relbar \relbar\joinrel\rightsquigarrow \;\; {{\mathcal{O}}_q\big(G\big/K\big)}^\vee =: U_q\big(\mathfrak{k}^\perp\big)$$ in which the right-hand side is a quantization of $ \, U \big( \mathfrak{k}^\perp \big) \, $. However, this recipe makes no sense when $ {\mathcal{O}}_q\big(G\big/K\big) $ is not available. In the non-formal setting this is the case whenever $ G \big/ K $ is not quasi-affine, e.g. when it is projective. In this paper we show how to solve this problem in the special case of the Grassmann varieties, taking $ G $ as the general linear group and $ \, K = P \, $ a maximal parabolic subgroup. We adapt the basic ideas of the original QDP recipe to these new ingredients, and we obtain a new recipe $${\mathcal{O}}_q \big( G \big/ P \big) \;\; \circ\hskip-5,3pt\relbar \relbar \relbar \relbar\joinrel\rightsquigarrow \;\; \widehat{{{\mathcal{O}}_q \big( G \big/ P \big)}^\vee}$$ which perfectly makes sense, and yields the same kind of result as predicted by the QDP for the quasi-affine case. In particular, $ \widehat{{{\mathcal{O}}_q \big( G \big/ P \big)}^\vee} $ is a quantization of $ U\big(\mathfrak{p}^\perp\big) \, $, obtained through a $ (q-1) $–adic completion process. Our construction goes as follows. First, we consider the embedding of the Grassmannian $ G \big/ P $ (where $ \, G := {GL}_n \, $ or $ \, G := {SL}_n \, $, and $ P $ is a parabolic subgroup of $ G \, $) inside a projective space, given by Pl[ü]{}cker coordinates. This will give us the first new ingredient: $${\mathcal{O}}\big( G \big/ P \big) \; := \; \hbox{ring of homogeneous coordinates on} \ G \big/ P \quad .$$ Many quantizations $ {\mathcal{O}}_q \big( G \big/ P \big) $ of $ {\mathcal{O}}\big( G \big/ P \big) $ already exist in the literature (see, e.g., [@fi1; @lr; @tt]). All these quantizations, which are equivalent, come together with a quantization of the natural $ G $–action on $ \, G/P \, $. In the original recipe (see [@cg]) $ \; {\mathcal{O}}_q \big( G\big/K \big) \, \circ\hskip-5,3pt\relbar \relbar \relbar \relbar\joinrel\rightsquigarrow {{\mathcal{O}}_q\big( G\big/K \big)}^\vee \; $ of the QDP (when $ G \big/ K $ is quasi affine) we need to look at a neighborhood of the special point $ eK $ (where $ \, e \in G \, $ is the identity), and at a quantization of it. Therefore, we shall replace the projective variety $ G \big/ P $ with such an affine neighborhood, namely the big cell of $ G \big/ P \, $. This amounts to realize the algebra of regular functions on the big cell as a “homogeneous localization” of $ {\mathcal{O}}\big( G\big/P \big) $, say $ {\mathcal{O}}^{\,\text{\it loc}} \big( G \big/ P \big) $, by inverting a suitable element. We then do the same at the quantum level, via the inversion of a suitable almost central element in $ {\mathcal{O}}_q\big( G\big/P \big) $ — which lifts the previous one in $ {\mathcal{O}}\big( G\big/P \big) \, $. The result is a quantization $ {\mathcal{O}}_q^{\,\text{\it loc}} \big( G \big/ P \big) $ of the coordinate ring of the big cell. Hence we are able to [*define*]{} $ \, {{\mathcal{O}}_q\big( G \big/ P \big)}^\vee \! := {{\mathcal{O}}_q^{\,\text{\it loc}}\big( G \big/ P \big)}^\vee $, where the right-hand side is given by the original QDP recipe applied to the big cell as an affine variety (we can forget any group action at this step). By the very construction, this $ {{\mathcal{O}}_q \big( G \big/ P \big)}^\vee $ should be a quantization of $ U \big( \mathfrak{p}^\perp \big) \, $ (as an algebra). Indeed, we prove that this is the case, so we might think at $ {{\mathcal{O}}_q \big( G \big/ P \big)}^\vee $ as a quantization (of infinitesimal type) of the variety $ G^* \big/ P^\perp \, $. On the other hand, the construction does not ensure that $ {{\mathcal{O}}_q\big( G \big/ P \big)}^\vee $ also admits a quantization of the $ G^* $–action on $ G^* \big/ P^\perp \, $ (just like the big cell is not a $ G $–space). As a last step, we look at $ \widehat{{{\mathcal{O}}_q \big( G \big/ P \big)}^\vee} $, the $ (q \! - \! 1) $–adic completion of $ {{\mathcal{O}}_q\big( G \big/ P \big)}^\vee $. Of course, it is again a quantization of $ U \big( \mathfrak{p}^\perp \big) \, $ (as an algebra). But in addition, it admits a coaction of the $ (q \! - \! 1) $–adic completion of $ {\mathcal{O}}_q(G)^\vee $ — which is a quantization of $ U({\mathfrak{g}}^*) $. This coaction yields a quantization of the infinitesimal $ G^* $–action on $ G^* \big/ P^\perp $. Therefore, in a nutshell, $ \, \widehat{{{\mathcal{O}}_q\big( G \big/ P \big)}^\vee} \, $ is a quantization of $ \, G^* \big/ P^\perp \, $ [*as a homogeneous $ G^* $–space*]{}, in the sense explained above. Notice that our arguments could be applied to any [*projective*]{} homogeneous $ G $–space $ X \, $, [*up to having the initial data to start with*]{}. Namely, one needs an embedding of $ X $ inside a projective space, a quantization (compatible with the $ G $–action) of the ring of homogeneous coordinates of $ X $ ([w.r.t.]{} such an embedding), and a quantization of a suitable open dense affine subset of $ X \, $. This program is carried out in detail in a separate work (see [@cfg]). Finally, this paper is organized as follows. In section \[quantum\] we fix the notation, and we describe the Manin deformations of the general linear group (as a Poisson group), and of its Lie bialgebra, together with its dual. In section \[grass\] we briefly recall results concerning the constructions of the quantum Grassmannian $ {\mathcal{O}}_q \big( G \big/ P \big) $ and its quantum big cell $ {\mathcal{O}}_q^{\,\text{\it loc}} \big( G \big/ P \big) \, $. These are known results, treated in detail in [@fi1; @fi2]. Finally, in section \[qdpgrass\] we extend the original QDP to build $ {{\mathcal{O}}_q \big( G \big/ P \big)}^\vee $, and we show that its $ (q - 1) $–adic completion is a quantization of the homogeneous $ G^* $–space $ G^* \big/ P^\perp $ dual to the Grassmannian $ G \big/ P \, $. **Acknowledgements** The first author wishes to thank the Dipartimento di Matematica “Tor Vergata”, and in particular Prof. V. Baldoni and Prof. E. Strickland, for the warm hospitality during the period in which this paper was written. Both authors also thank D. Parashar and M. Marcolli for their kind invitation to the workshop “Quantum Groups and Noncommutative Geometry” held at MPIM in Bonn during August 6–8, 2007. The Poisson Lie group $ {GL}_n(\Bbbk) $ and its quantum deformation {#quantum} =================================================================== Let $ \Bbbk $ be any field of characteristic zero. In this section we want to recall the construction of a quantum deformation of the Poisson Lie group $ \, {GL}_n := {GL}_n(\Bbbk) \, $. We will also describe explicitly the bialgebra structure of its Lie algebra $ \, {{\mathfrak{gl}}}_n := {{\mathfrak{gl}}}_n(\Bbbk) \, $ in a way that fits our purposes, that is to obtain a quantum duality principle for the Grassmann varieties for $ {GL}_n $ (see §\[qdpgrass\]). Let $ \, \Bbbk_q = \Bbbk\big[q,q^{-1}\big] \, $ (where $ q $ is an indeterminate), the ring of Laurent polynomials over $ q \, $, and let $ \, \Bbbk(q) \, $ be the field of rational functions in $ q \, $. \[manin\] The [*quantum matrix algebra*]{} is defined as $${\mathcal{O}}_q(M_{m \times n}) \; = \; \Bbbk_q \big\langle {\{\, x_{ij} \,\}}_{1 \leq i \leq m}^{1 \leq j \leq n} \,\big\rangle \Big/ I_M$$ where the $ x_{ij} $’s are non commutative indeterminates, and $ I_M $ is the two-sided ideal generated by the [*Manin relations*]{} $$\displaylines{ x_{ij} \, x_{ik} \; = \; q \, x_{ik} \, x_{ij} \; , \qquad x_{ji} \, x_{ki} \; = \; q \, x_{ki} \, x_{ji} \; \qquad \forall \;\; j < k \cr {\ } \qquad \qquad x_{ij} \, x_{kl} \; = \; x_{kl} \, x_{ij} \; \qquad \qquad \forall \;\; i < k \, , \, j > l \hbox{\ \ or \ } i > k \, , \, j < l \cr x_{ij} \, x_{kl} \, - \, x_{kl} \, x_{ij} \; = \; \big( q - q^{-1} \big) \, x_{kj} \, x_{il} \; \qquad \forall \;\; i < k \, , \, j < l \cr }$$ [**Warning:**]{} sometimes these relations appear with $ q $ exchanged with $ q^{-1} \, $. For simplicity we will denote $ {\mathcal{O}}_q(M_{n \times n}) $ with $ {\mathcal{O}}_q(M_{n}) \, $. There is a coalgebra structure on $ {\mathcal{O}}_q(M_{n}) \, $, given by $$\Delta(x_{ij}) \; = \; {\textstyle \sum\limits_{k=1}^n} \, x_{ik} \otimes x_{kj} \quad , \qquad \epsilon(x_{ij}) \; = \; \delta_{ij} \qquad \eqno ( \, 1 \leq i \, , j \leq n \,)$$ The [*quantum general linear group*]{} and the [*quantum special linear group*]{} are defined in the following way: $${\mathcal{O}}_q(GL_n) \! := {\mathcal{O}}_q(M_n)[T] \Big/ \! \big( T {{D_q}}- 1 \, , 1 - T {{D_q}}\big) \; , \;\; {\mathcal{O}}_q(SL_n) \! := {\mathcal{O}}_q(M_n) \Big/ \! \big( {{D_q}}- 1 \big)$$ where $ \; {{D_q}}:= \sum_{\sigma \in \mathcal{S}_n} \, (-q)^{\ell(\sigma)} \, x_{1 \, \sigma(1)} \cdots x_{n \, \sigma(n)} \; $ is a central element, called the [*quantum determinant*]{}. [**Note:**]{} We use the same letter to denote the generators $ x_{ij} $ of $ {\mathcal{O}}_q(M_{m \times n}) \, $, of $ {\mathcal{O}}_q(GL_n) $ and of $ {\mathcal{O}}_q(SL_n) \, $: the context will make clear where they sit. The algebra $ {\mathcal{O}}_q({GL}_n) $ is a quantization of the algebra $ {\mathcal{O}}({GL}_n) $ of regular functions on the affine algebraic group $ {GL}_n \, $, in the following sense: $ \; {\mathcal{O}}_q({GL}_n) \big/ (q \! - \! 1) \, {\mathcal{O}}_q({GL}_n) \; $ is isomorphic to $ {\mathcal{O}}({GL}_n) $ as a Hopf algebra (over the field $ \Bbbk \, $). Similarly, $ {\mathcal{O}}_q ({SL}_n) $ is a quantization of the algebra $ {\mathcal{O}}({SL}_n) $ of regular functions on $ {SL}_n \, $. Both $ {\mathcal{O}}_q({GL}_n) $ and $ {\mathcal{O}}_q({SL}_n) $ are Hopf algebras, that is, they also have the antipode. For more details on these constructions see for example [@cp], pg. 215. By general theory, $ {\mathcal{O}}({GL}_n) $ inherits from $ {\mathcal{O}}_q({GL}_n) $ a Poisson bracket, which makes it into a Poisson Hopf algebra, so that $ {GL}_n $ becomes a Poisson group. We want to describe now its Poisson bracket. Recall that $${\mathcal{O}}({GL}_n) \; = \; \Bbbk \big[ {\{\, \bar{x}_{ij} \,\}}_{i, j = 1, \dots, n} \,\big][t] \Big/ \big( t \, d - 1 \big)$$ where $ \; d := \text{\sl det}\, \big( \bar{x}_{i,j} \big)_{i, j = 1, \dots, n} \; $ is the usual determinant. Setting $ \, \bar{x} = \pi(x) \, $ for $ \; \pi : {\mathcal{O}}_q(GL_n) {\longrightarrow}{\mathcal{O}}(GL_n) \; $, the Poisson structure is given (as usual) by $$\big\{ \bar{a} \, , \bar{b} \,\big\} \; := \; {(q-1)}^{-1} \, ( a \, b - b \, a )\Big|_{q=1} \eqno \forall \;\; \bar{a} \, , \bar{b} \in {\mathcal{O}}({GL}_n) \;\; .$$ In terms of generators, we have $$\displaylines{ \big\{ \bar{x}_{ij} \, , \bar{x}_{ik} \big\} \, = \, \bar{x}_{ij} \, \bar{x}_{ik} \quad \forall \;\; j < k \; , \phantom{\Big|} \quad \big\{ \bar{x}_{ij} \, , \bar{x}_{\ell k} \big\} \, = \, 0 \qquad \forall \;\; i < \ell \, , k < j \cr \big\{ \bar{x}_{ij} \, , \bar{x}_{\ell j} \big\} \, = \, \bar{x}_{ij} \, \bar{x}_{\ell j} \quad \forall \;\; i < \ell \, , \phantom{\Big|} \qquad \big\{ \bar{x}_{ij} \, , \bar{x}_{\ell k} \big\} \, = \, 2 \, \bar{x}_{ij} \, \bar{x}_{\ell k} \quad \forall \;\; i < \ell \, , j < k \cr \big\{ d^{-1} , \bar{x}_{ij} \big\} \, = \,0 \; , \phantom{\Big|} \quad \big\{ d \, , \bar{x}_{ij} \big\} \,= \, 0 \qquad \forall \;\; i, j = 1, \dots, n \, . \cr }$$ As $ {GL}_n $ is a Poisson Lie group, its Lie algebra $ {{\mathfrak{gl}}}_n $ has a Lie bialgebra structure (see [@cp], [pg. ]{}24). To describe it, let us denote with $ {{E}}_{ij} $ the elementary matrices, which form a basis of $ {{\mathfrak{gl}}}_n \, $. Define ($ \, \forall \; i = 1, \dots, n-1 \, $, $ \, j = 1, \dots, n \, $) $$e_i := {{E}}_{i,i+1} \; , \quad g_j := {{E}}_{j,j} \; , \quad f_i := {{E}}_{i+1,i} \; , \quad h_i := g_{i} - g_{i+1}$$ Then $ \; \big\{\, e_i \, , \, f_i \, , \, g_j \;\big|\; i = 1, \dots, n-1, \, j = 1, \dots, n \,\big\} \; $ is a set of Lie algebra generators of $ {{\mathfrak{gl}}}_n \, $, and a Lie cobracket is defined on $ {{\mathfrak{gl}}}_n $ by $$\delta(e_i) \, = \, h_i \wedge e_i \;\; , \quad \delta(g_j) \, = \, 0 \;\; , \quad \delta(f_i) \, = \, h_i \wedge f_i \; \eqno \forall \;\, i , j .$$ This cobracket makes $ {{\mathfrak{gl}}}_n $ itself into a [*Lie bialgebra*]{}: this is the so-called [*standard*]{} Lie bialgebra structure on $ {{\mathfrak{gl}}}_n \, $. It follows immediately that $ \, U({{\mathfrak{gl}}}_n) \, $ is a co-Poisson Hopf algebra, whose co-Poisson bracket is the (unique) extension of the Lie cobracket of $ {{\mathfrak{gl}}}_n $ while the Hopf structure is the standard one. Similar constructions hold for the group $ {SL}_n \, $. One simply drops the generator $ d^{-1} \, $, imposes the relation $ \, d \! = \! 1 \, $, in the description of $ {\mathcal{O}}(SL_n) \, $, and replaces the $ g_s $’s with the $ h_i $’s ($ \, i = 1, \dots, n \, $) when describing $ \, {{\mathfrak{sl}}}_n \, $. Since $ {{\mathfrak{gl}}}_n $ is a Lie bialgebra, its dual space $ {{\mathfrak{gl}}}_n^{\,*} $ admits a Lie bialgebra structure, dual to the one of $ {{\mathfrak{gl}}}_n \, $. Let $ \, \big\{\, {{\hbox{E}}}_{ij} := {{E}}_{ij}^{\,*} \;\big|\; i, j = 1, \dots, n \,\big\} \, $ be the basis of $ \, {{\mathfrak{gl}}}_n^{\,*} \, $ dual to the basis of elementary matrices for $ {{\mathfrak{gl}}}_n \, $. As a Lie algebra, $ {{\mathfrak{gl}}}_n^{\,*} $ can be realized as the subset of $ \, {{\mathfrak{gl}}}_n \oplus {{\mathfrak{gl}}}_n \, $ of all pairs $$\left( \! \begin{pmatrix} \! -m_{11} \! & \! 0 \! & \! \cdots \! & \! 0 \! \\ \! m_{21} \! & \! -m_{22} \! & \! \cdots & \! 0 \! \\ \! \vdots \! & \! \vdots \! & \! \vdots \! & \! \vdots \! \\ \! m_{n-1,1} \! & \! m_{n-1,2} \! & \! \cdots \! & \! 0 \! \\ \! m_{n,1} \! & \! m_{n,2} \! & \! \cdots \! & \! -m_{n,n} \end{pmatrix} , \, \begin{pmatrix} m_{11} \! & \! m_{12} \! & \! \cdots \! & \! m_{1,n-1} \! & \! m_{1,n} \! \\ \! 0 \! & \! m_{22} \! & \! \cdots \! & \! m_{2,n-1} \! & \! m_{2,n} \! \\ \! \vdots \! & \! \vdots \! & \! \vdots \! & \! \vdots \! & \! \vdots \! \\ \! 0 \! & \! 0 \! & \! \cdots \! & \! m_{n-1,n-1} \! & \! m_{n-1,n} \! \\ \! 0 \! & \! 0 \! & \! \cdots \! & \! 0 \! & \! m_{n,n} \! \end{pmatrix} \! \right)$$ with its natural structure of Lie subalgebra of $ \, {{\mathfrak{gl}}}_n \oplus {{\mathfrak{gl}}}_n \, $. In fact, the elements $ \, {{\hbox{E}}}_{ij} \, $ correspond to elements in $ \, {{\mathfrak{gl}}}_n \oplus {{\mathfrak{gl}}}_n \, $ in the following way: $${{\hbox{E}}}_{ij} \cong \big( {{E}}_{ij} \, , 0 \big) \hskip5pt \forall \; i \! > \! j \, , \hskip7pt {{\hbox{E}}}_{ij} \cong \big(\! -{{E}}_{ij} \, , +{{E}}_{ij} \big) \hskip5pt \forall \; i \! = \! j \, , \hskip7pt {{\hbox{E}}}_{ij} \cong \big( 0 \, , {{E}}_{ij} \big) \hskip5pt \forall \; i < j \, .$$ Then the Lie bracket of $ \, {{\mathfrak{gl}}}_n^{\,*} \, $ is given by $$\begin{array}{clc} \big[ {{\hbox{E}}}_{i,j} \, , \, {{\hbox{E}}}_{h,k} \big] \hskip-5pt & = \, \delta_{j,h} \, {{\hbox{E}}}_{i,k} - \delta_{k,i} \, {{\hbox{E}}}_{h,j} \;\; , & \forall \;\; i \! \leq \! j \, , \, h \! \leq \! k \;\;\, \text{and} \;\; \forall \;\; i \! > \! j \, , \, h \! > \! k \\ \\ \big[ {{\hbox{E}}}_{i,j} \, , \, {{\hbox{E}}}_{h,k} \big] \hskip-5pt & = \, \delta_{k,i} \, {{\hbox{E}}}_{h,j} - \delta_{j,h} \, {{\hbox{E}}}_{i,k} \;\; , & \forall \;\; i \! = \! j \, , \, h \! > \! k \;\;\, \text{and} \;\; \forall \;\; i \! > \! j \, , \, h \! = \! k \\ \\ \big[ {{\hbox{E}}}_{i,j} \, , \, {{\hbox{E}}}_{h,k} \big] \hskip-5pt & = \, 0 \; , & \forall\;\; i \! < \! j \, , \, h \! > \! k \;\;\, \text{and} \;\; \forall \;\; i \! > \! j \, , \, h \! < \! k \end{array}$$ Note that the elements ($ \, 1 \leq i \leq n \! - \! 1 \, $, $ \, 1 \leq j \leq n \, $) $${{\hbox{e}}}_i \, = \, e_i^{\,*} \, = \, {{\hbox{E}}}_{i,i+1} \;\; , \qquad {{\hbox{f}}}_i \, = \, f_i^{\,*} \, = \, {{\hbox{E}}}_{i+1,i} \;\; , \qquad {{\hbox{g}}}_j \, = \, g_j^{\,*} \, = \, {{\hbox{E}}}_{jj}$$ are Lie algebra generators of $ {{\mathfrak{gl}}}_n^{\,*} \, $. In terms of them, the Lie bracket reads $$\big[ {{\hbox{e}}}_i \, , {{\hbox{f}}}_j \big] \, = \, 0 \; , \qquad \big[ {{\hbox{g}}}_i \, , {{\hbox{e}}}_j \big] \, = \, \delta_{ij} \, {{\hbox{e}}}_i \; , \qquad \big[ {{\hbox{g}}}_i \, , {{\hbox{f}}}_j \big] \, = \, \delta_{ij} \, {{\hbox{f}}}_j \eqno \forall \;\; i, j \; .$$ On the other hand, the Lie cobracket structure of $ {{\mathfrak{gl}}}_n^{\,*} $ is given by $$\delta\big({{\hbox{E}}}_{i,j}\big) \, = \, {\textstyle \sum\limits_{k=1}^n} \, {{\hbox{E}}}_{i,k} \wedge {{\hbox{E}}}_{k,j} \eqno \forall \;\; i, j = 1, \dots, n$$ where $ \, x \wedge y := x \otimes y - y \otimes x \; $. Finally, all these formul[æ]{} also provide a presentation of $ U\big({{\mathfrak{gl}}}_n^{\,*}\big) $ as a co-Poisson Hopf algebra. A similar description holds for $ \, {{\mathfrak{sl}}}_n^{\,*} \! = {{\mathfrak{gl}}}_n^{\,*} \Big/ Z\big({{\mathfrak{gl}}}_n^{\,*}\big) \, $, where $ \, Z \big( {{\mathfrak{gl}}}_n^{\,*} \big) \, $ is the centre of $ {{\mathfrak{gl}}}_n^{\,*} \, $, generated by $ \, \mathfrak{l}_n := {{\hbox{g}}}_1 \! + \cdots + {{\hbox{g}}}_n \, $. The construction is immediate by looking at the embedding $ \, {{\mathfrak{sl}}}_n \hookrightarrow {{\mathfrak{gl}}}_n \, $. -9pt The quantum Grassmannian and its big cell {#grass} ============================================== In this section we want to briefly recall the construction of a quantum deformation of the Grassmannian of $ r $–spaces inside an $ n $–dimensional vector space and its big cell, as they appear in [@fi1; @fi2]. The quantum Grassmannian ring will be obtained as a quantum homogeneous space, namely its deformation will come together with a deformation of the natural coaction of the function algebra of the general linear group on it. The deformation will also depend on a specific embedding (the Pl[ü]{}cker one) of the Grassmann variety into a projective space. This deformation is very natural, in fact it embeds into the deformation of its big cell ring. Let’s see explicitly these constructions. Let $ \, G := {GL}_n \, $, and let $ P $ and $ P_1 $ be the standard parabolic subgroups $$P \, := \, \left\{ \begin{pmatrix} A & B \\ 0 & C \end{pmatrix} \right\} \subset {GL}_n \quad , \qquad P_1 \, := \, P \, {\textstyle \bigcap} \, {SL}_n$$ where $ A $ is a square matrix of size $ r \, $, with $ \, 0 < r < n \, $. The [*quantum Grassmannian coordinate ring*]{} $ \, {\mathcal{O}}_q \big( G \big/ P \big) \, $ with respect to the Pl[ü]{}cker embedding is the subalgebra of $ {\mathcal{O}}_q({GL}_n) $ generated by the quantum minors (called [*quantum Pl[ü]{}cker coordinates*]{}) $$D^I \, = \, D^{i_1 \dots i_r} \, := \, {\textstyle \sum\limits_{\sigma \in \mathcal{S}_r}} \, {(-q)}^{\ell(\sigma)} \, x_{i_1 \, \sigma(1)} \, x_{i_2 \, \sigma(2)} \cdots x_{i_r \, \sigma(r)} \quad .$$ for every ordered $ r $–tuple of indices $ \; I =\{i_1 < \cdots < i_r\} \, $. [*$ \underline{\text{Remark}} $:*]{} Equivalently, $ \, {\mathcal{O}}_q \big( G\big/P \big) \, $ may be defined in the same way but with $ {\mathcal{O}}_q({SL}_n) $ instead of $ {\mathcal{O}}_q({GL}_n) \, $. The algebra $ {\mathcal{O}}_q\big(G\big/P\big) $ is a quantization of the Grassmannian $ G\big/P $ in the usual sense: the $ \Bbbk $–algebra $ \, {\mathcal{O}}_q\big( G\big/P\big) \Big/ (q-1) \, {\mathcal{O}}_q\big(G\big/P\big) \, $ is isomorphic to $ {\mathcal{O}}\big(G\big/P\big) \, $, the algebra of homogeneous coordinates of $ G\big/P $ with respect to the Pl[ü]{}cker embedding. In addition, $ {\mathcal{O}}_q\big(G\big/P\big) $ has an important property w.r.t. $ {\mathcal{O}}_q(G) \, $, given by the following result: \[firstintersection\] $${\mathcal{O}}_q\big(G\big/P\big) \,\;{\textstyle \bigcap}\;\, (q-1) \, {\mathcal{O}}_q(G) \,\; = \;\, (q-1) \; {\mathcal{O}}_q\big(G\big/P\big)$$ [[*Proof*]{}]{}. By Theorem 3.5 in [@tt], we have that certain products of minors $ \, {\{p_i\}}_{i \in I} \, $ form a basis of $ {\mathcal{O}}_q\big( G \big/ P \big) \, $ over $ \Bbbk_q \, $. Thus, a generic element in $ \, {\mathcal{O}}_q\big(G\big/P\big) \,\bigcap\, (q \! - \! 1) \, {\mathcal{O}}_q(G) \, $ can be written as $${\textstyle \sum_{i \in I}} \, \alpha_i \, p_i \; = \; (q-1) \, \phi \eqno (3.1)$$ for some $ \, \phi \in {\mathcal{O}}_q(G) \, $. Moreover, the specialization map $$\pi_G \; \colon \; {\mathcal{O}}_q(G) \; \relbar\joinrel\relbar\joinrel\relbar\joinrel\twoheadrightarrow \; {\mathcal{O}}_q(G) \Big/ (q-1) \, {\mathcal{O}}_q(G) \; = \; {\mathcal{O}}(G)$$ maps $ \, {\{p_i\}}_{i \in I} \, $ onto a basis $ \, {\big\{ \pi_G(p_i) \big\}}_{i \in I} \, $ of $ {\mathcal{O}}\big( G\big/P \big) \, $, the latter being a subalgebra of $ {\mathcal{O}}(G) \, $. Therefore, applying $ \pi_G $ to (3.1) we get $ \, \sum_{i \in I} \overline{\alpha_i} \, \pi_G(p_i) = 0 \, $, where $ \, \overline{\alpha_i} := \alpha_i \mod (q \! - \! 1) \, \Bbbk_q \, $, for all $ \, i \in I \, $. This forces $ \, \alpha_i \in (q \! - \! 1) \, \Bbbk_q \, $ for all $ i \, $, by the linear independence of the $ \pi_G(p_i) $’s, whence the claim. An immediate consequence of Proposition \[firstintersection\] is that the canonical map $${\mathcal{O}}_q\big( G\big/P \big) \Big/ (q-1) \, {\mathcal{O}}_q\big( G\big/P \big) \; \relbar\joinrel\relbar\joinrel\relbar\joinrel\longrightarrow \; {\mathcal{O}}_q(G) \Big/ (q-1) \, {\mathcal{O}}_q(G)$$ is [*injective*]{}. Therefore, the specialization map $$\pi_{G/P} \; \colon \; {\mathcal{O}}_q\big( G\big/P \big) \;\relbar\joinrel\relbar\joinrel\relbar\joinrel\twoheadrightarrow\; {\mathcal{O}}_q\big( G\big/P \big) \Big/ (q-1) \, {\mathcal{O}}_q\big( G\big/P \big)$$ coincides with the restriction to $ {\mathcal{O}}_q\big( G\big/P \big) $ of the specialization map $$\pi_G \; \colon \; {\mathcal{O}}_q(G) \; \relbar\joinrel\relbar\joinrel\relbar\joinrel\twoheadrightarrow \; {\mathcal{O}}_q(G) \Big/ (q-1) \, {\mathcal{O}}_q(G) \quad .$$ Moreover — from a geometrical point of view — the key consequence of this property is that [*$ P $ is a coisotropic subgroup*]{} of the Poisson group $ G \, $. This implies the existence of a well defined Poisson structure on the algebra $ \, {\mathcal{O}}\big( G\big/P \big) \, $, inherited from the one in $ {\mathcal{O}}(G) \, $. \[comultiplication\] The quantum deformation $ {\mathcal{O}}_q\big( G\big/P \big) $ comes naturally equipped with a coaction of $ {\mathcal{O}}_q({GL}_n) $ — or, similarly, of $ {\mathcal{O}}_q({SL}_n) $ — on it, obtained by restricting the comultiplication $ \Delta \, $. This reads $$\begin{array}{cccc} \Delta{\big|}_{{\mathcal{O}}_q(G/P)} : & {\mathcal{O}}_q\big( G \big/ P \big) & {\longrightarrow}& {\mathcal{O}}_q(G) \otimes {\mathcal{O}}_q\big( G \big/ P \big) \\ & D^I & \mapsto & \sum_K D^I_K \otimes D^K \end{array}$$ where, for any $ \, I = (i_1 \dots i_r) \, $, $ \, K = (k_1 \dots k_r) \, $, with $ \, 1 \leq i_1 < \dots < i_r \leq n \, $, $ \, 1 \leq k_1 < \dots < k_r \leq n \, $, we denote by $ D^I_K $ the [*quantum minor*]{} $$D^I_K \; \equiv \; D^{i_1 \dots i_r}_{k_1 \dots k_r} \; := \; {\textstyle \sum\limits_{\sigma \in \mathcal{S}_r}} \, {(-q)}^{\ell(\sigma)} \, x_{i_1 \, k_{\sigma(1)}} \, x_{i_2 \, k_{\sigma(2)}} \cdots x_{i_r \, k_{\sigma(r)}} \quad .$$ This provides a quantization of the natural coaction of $ {\mathcal{O}}(G) $ onto $ {\mathcal{O}}\big( G \big/ P \big) \, $. The ring $ {\mathcal{O}}_q\big( G\big/P \big) $ has been fully described in [@fi1] in terms of generators and relations. We refer the reader to this work for further details. We now turn to the construction of the quantum big cell ring. Let $ \, I_0 = (1 \dots r) \, $, $ \, D_0 := D^{I_0} \, $. Define $${\mathcal{O}}_q(G)\big[D_0^{-1}\big] \; := \; {\mathcal{O}}_q(G)[T] \Big/ \big( T \, D_0 - 1 \, , D_0 \, T - 1 \big)$$ Moreover, we define the [*big cell ring*]{} $ \, {\mathcal{O}}_q^{\,\text{\it loc}}\big( G\big/P \big) \, $ to be the $ \Bbbk_q $–subalgebra of $ \, {\mathcal{O}}_q(G)\big[D_0^{-1}\big] \, $ generated by the elements $${\ } \qquad t_{ij} \, := \, {(-q)}^{r-j} \, {D^{1 \, \dots \, \widehat{j} \, \dots \, r \, i}}\, D_0^{-1} \eqno \forall \;\;\, i \, , j \; : \; 1 \leq j \leq r < i \leq n \; .$$ See [@fi2] for more details. As in the commutative setting, we have the following result: \[classical\] $ \; \displaystyle{ {\mathcal{O}}_q^{\,\text{\it loc}}\big( G\big/P \big) \, \cong \, {\mathcal{O}}_q\big( G\big/P \big)\big[D_0^{-1}\big]_{proj} } \; $, where the right-hand side is the degree-zero component of $ \; {\mathcal{O}}_q\big( G\big/P \big)[T] \Big/ \big(T D_0 - 1 \, , D_0 \, T - 1 \big) \, $. [[*Proof*]{}]{}. In the classical setting, the analogous result is proved by this argument: one uses the so-called “straightening relations” to get rid of the extra minors (see, for example, [@dep], §2). Here the argument works essentially the same, using the [*quantum straightening [(or]{} Pl[ü]{}cker[)]{} relations*]{} (see [@fi1], §4, [@tt], formula (3.2)(c) and Note I, Note II). As before, we have that $${\mathcal{O}}_q^{\,\text{\it loc}}\big( G \big/ P \big) \,\;{\textstyle \bigcap}\;\, (q-1) \, {\mathcal{O}}_q^{\,\text{\it loc}}(G) \,\; = \;\, (q-1) \, {\mathcal{O}}_q^{\,\text{\it loc}}\big( G \big/ P \big)$$ This can be easily deduced from Proposition \[firstintersection\], taking into account Proposition \[classical\]. As a consequence, the map $${\mathcal{O}}_q^{\,\text{\it loc}}\big( G \big/ P \big) \Big/ (q-1) \, {\mathcal{O}}_q^{\,\text{\it loc}}\big( G \big/ P \big) \; \relbar\joinrel\relbar\joinrel\relbar\joinrel\longrightarrow \; {\mathcal{O}}_q^{\,\text{\it loc}}(G) \Big/ (q-1) \, {\mathcal{O}}_q^{\,\text{\it loc}}(G)$$ is [*injective*]{}, so that the specialization map $$\pi_{G/P}^{\,\text{\it loc}} \colon \, {\mathcal{O}}_q^{\,\text{\it loc}}\big( G \big/ P \big) \; \relbar\joinrel\relbar\joinrel\relbar\joinrel\twoheadrightarrow \; {\mathcal{O}}_q^{\,\text{\it loc}}\big( G \big/ P \big) \Big/ (q-1) \, {\mathcal{O}}_q^{\,\text{\it loc}}\big( G \big/ P \big)$$ coincides with the restriction of the specialization map $$\pi_G^{\,\text{\it loc}} \; \colon \; {\mathcal{O}}_q^{\,\text{\it loc}}(G) \; \relbar\joinrel\relbar\joinrel\relbar\joinrel\twoheadrightarrow \; {\mathcal{O}}_q^{\,\text{\it loc}}(G) \Big/ (q-1) \, {\mathcal{O}}_q^{\,\text{\it loc}}(G) \quad .$$ The following proposition gives a description of the algebra $ {\mathcal{O}}_q^{\,\text{\it loc}}\big( G \big/ P \big) \, $: \[bigcell\] The big cell ring is isomorphic to a matrix algebra $$\begin{array}{cccl} \qquad {\mathcal{O}}_q^{\,\text{\it loc}}\big( G\big/P \big) & {\longrightarrow}& {\mathcal{O}}_q\big(M_{(n-r) \times r}\big) & \\ \qquad t_{ij} & \mapsto & x_{ij} & \qquad \forall \;\; 1 \leq j \leq r < i \leq n \end{array}$$ i.e. the generators $ t_{ij} $’s satisfy the Manin relations. [[*Proof*]{}]{}. See [@fi2], Proposition 1.9. The Quantum Duality Principle for quantum Grassmannians {#qdpgrass} ======================================================= The quantum duality principle (QDP), originally due to Drinfeld [@dr] and later formalized in [@ga2] and extended in [@ga3; @ga4] by Gavarini, is a functorial recipe to obtain a quantum group starting from a given one. The main ingredients are the “Drinfeld functors”, which are equivalences between the category of QFA’s and the category of QUEA’s. Ciccoli and Gavarini extended this principle to the setting of homogeneous spaces. More precisely, in [@cg] they developed the QDP for homogeneous spaces in the [*local setting*]{}, i.e. for quantum groups of formal type (where topological Hopf algebras are taken into account). If one tries to find a global version of the QDP for non quasi-affine homogeneous spaces, then problems arise from the very beginning, as explained in §\[intro\]. The case of [*projective*]{} homogeneous spaces has been solved in [@cfg], where the original version of the Drinfeld-like functor for which the (global) QDP recipe should fail is suitably modified. In this section, we apply the general recipe for projective homogeneous spaces to the Grassmannian $ G/P \, $. The result is a quantization of the homogeneous space [*dual*]{} (in the sense of Poisson duality, see [@cg]) to $ G/P \, $, just as the QDP recipe predicts in the setting of [@cg]. We begin recalling the Drinfeld functor $ \; {}^\vee \, \colon \, QFA \longrightarrow QUEA \; $. Let $ G $ be an affine algebraic group over $ \Bbbk \, $, and $ {\mathcal{O}}_q(G) $ a quantization of its function algebra. Let $ J $ be the augmentation ideal of $ {\mathcal{O}}_q(G) \, $, i.e. the kernel of the counit $ \; \epsilon : {\mathcal{O}}_q(G) {\longrightarrow}\Bbbk \; $. Define $${\mathcal{O}}_q(G)^\vee \, := \, \big\langle {(q-1)}^{-1} \, J \, \big\rangle \, = \, {\textstyle \sum\limits_{n=0}^\infty} \, {(q-1)}^{-n} \, J^n \qquad \big( \subset \, {\mathcal{O}}_q(G) \otimes_{\Bbbk_{q}} \Bbbk(q) \,\big) \; .$$ It turns out that $ {\mathcal{O}}_q(G)^\vee $ is a quantization of $ U({\mathfrak{g}}^*) \, $, where $ {\mathfrak{g}}^* $ is the dual Lie bialgebra to the Lie bialgebra $ \, {\mathfrak{g}}= \text{\it Lie}\,(G) \, $. So $ {\mathcal{O}}_q(G)^\vee $ is a QUEA, and an infinitesimal quantization for any Poisson group $ G^* $ dual to $ G \, $, i.e. such that $ \, \text{\it Lie}\,\big(G^*\big) \cong {\mathfrak{g}}^* \, $ as Lie bialgebras. Moreover, the association $ \, {\mathcal{O}}_q(G) \mapsto {\mathcal{O}}_q(G)^\vee \, $ yields a functor from QFA’s to QUEA’s (see [@ga3; @ga4] for more details). \[qdp\] Let $ \, G = {GL}_n \, $. Then $ {\mathcal{O}}_q(G)^\vee $ is generated, as a unital subalgebra of $ \, {\mathcal{O}}_q(G) \otimes_{\Bbbk_q} \Bbbk(q) \, $, by the elements $${\mathcal{D}}_- \, := \, {(q-1)}^{-1} \, \big( D_q^{-1} - 1 \big) \; , \quad \hskip5pt \chi_{ij} \, := \, {(q-1)}^{-1} \, \big( x_{ij} - \delta_{ij} \big) \eqno \forall \; i, j = 1, \dots, n$$ where the $ x_{ij} $’s are the generators of $ {\mathcal{O}}_q(G) \, $. As $ \, x_{ij} = \delta_{ij} + (q-1) \, \chi_{ij} \in {\mathcal{O}}_q(G)^\vee \, $, we have an obvious embedding of $ {\mathcal{O}}_q(G) $ into $ {\mathcal{O}}_q(G)^\vee \, $. In the same spirit — mimicking the construction in [@cg] — we now want to define $ {{\mathcal{O}}_q\big( G \big/ P \big)}^\vee \, $ when $ G \big/ P $ is the Grassmannian. Let $ \, G = {GL}_n \, $, and let $ P $ be the maximal parabolic subgroup of §\[grass\]. Let $ \epsilon' $ be the natural extension to $ {\mathcal{O}}_q^{\,\text{\it loc}}(G/P) $ of the restriction to $ {\mathcal{O}}_q(G/P) $ of the counit of $ {\mathcal{O}}_q(G) \, $, and let $ \, J_{G/P}^{\,\text{\it loc}} := \text{\it Ker}\,(\epsilon'\,) \, $. We define (as a subset of $ \, {\mathcal{O}}_q^{\,\text{\it loc}}\big( G \big/ P \big) \otimes_{\Bbbk_q} \Bbbk(q) \, $) $${\mathcal{O}}_q\big( G \big/ P \big)^\vee \, := \, \big\langle \, {(q-1)}^{-1} \, J_{G/P}^{\,\text{\it loc}} \, \big\rangle \, = \, {\textstyle \sum\limits_{n=0}^\infty} \, {(q-1)}^{-n} \, {\big( J_{G/P}^{\,\text{\it loc}} \big)}^n \;\;\; .$$ It is worth pointing out that $ \, {\mathcal{O}}_q\big( G\big/P \big)^\vee \, $ is [*not*]{} a “quantum homogeneous space” for $ {\mathcal{O}}_q(G)^\vee $ in any natural way, i.e. it does not admit a coaction of $ {\mathcal{O}}_q(G)^\vee \, $. This is a consequence of the fact that there is no natural coaction of $ {\mathcal{O}}_q(G) $ on $ {\mathcal{O}}_q^{\,\text{\it loc}}\big( G\big/P \big) \, $. Now we examine this more closely. Since $ {{\mathcal{O}}_q\big( G\big/P \big)}^\vee $ is not contained in $ {{\mathcal{O}}_q(G)}^\vee \, $, we cannot have a $ {{\mathcal{O}}_q(G)}^\vee $ coaction induced by the coproduct. This would be the case if $ {{\mathcal{O}}_q\big( G \big/ P \big)}^\vee $ were a (one-sided) [*coideal*]{} of $ {{\mathcal{O}}_q(G)}^\vee \, $; but this is not true because $ {\mathcal{O}}_q^{\,\text{\it loc}}\big( G \big/ P \big) $ is not a (right) coideal of $ {\mathcal{O}}_q(G) $. This reflects the geometrical fact that the big cell of $ G/P $ is not a $ G $–space itself. Nevertheless, we shall find a way around this problem simply by [*enlarging*]{} $ {{\mathcal{O}}_q\big( G\big/P \big)}^\vee $ and $ {{\mathcal{O}}_q(G)}^\vee $, i.e. by taking their $ (q \! - \! 1) $–adic completion (which will not affect their behavior at $ \, q = 1 \, $). To begin, we provide a concrete description of $ {{\mathcal{O}}_q\big( G\big/P \big)}^\vee \, $: $${{\mathcal{O}}_q\big( G\big/P \big)}^\vee \; = \; \Bbbk_q \big\langle {\{\, \mu_{i{}j} \,\}}_{i = r+1, \dots, n}^{j = 1, \dots, r} \big\rangle \Big/ I_M$$ where $ \; \mu_{i{}j} := {(q-1)}^{-1} \, t_{i{}j} \; $ (for all $ i $ and $ j \, $), $ \, I_M \, $ is the ideal of the Manin relations among the $ \mu_{i{}j} $’s, and $ \; t_{ij} = {(-q)}^{r-j} \, {D^{1 \, \dots \, \widehat{j} \, \dots \, r \, i}}\, D_0^{-1} \, $ (for all $ i $ and $ j $). [[*Proof*]{}]{}. Trivial from definitions and Proposition \[bigcell\]. We now explain the relation between $ {{\mathcal{O}}_q\big( G\big/P \big)}^\vee $ and $ {{\mathcal{O}}_q(G)}^\vee \, $. The starting point is the following special property: \[intersection\] $${{\mathcal{O}}_q\big( G \big/ P \big)}^\vee \,\;{\textstyle \bigcap}\;\, (q-1) \, {{\mathcal{O}}_q(G)}^\vee\big[D_0^{-1}\big] \,\; = \;\, (q-1) \, {{\mathcal{O}}_q\big( G \big/ P \big)}^\vee$$ [[*Proof*]{}]{}. It is the same as for Proposition \[firstintersection\]. As a direct consequence of Proposition \[intersection\], the canonical map $${{\mathcal{O}}_q\big( G \big/ P \big)}^\vee \! \Big/ (q-1) \, {{\mathcal{O}}_q\big( G\big/P \big)}^\vee \, \relbar\joinrel\relbar\joinrel\relbar\joinrel\longrightarrow \; {{\mathcal{O}}_q(G)}^\vee\big[D_0^{-1}\big] \Big/ (q-1) \, {{\mathcal{O}}_q(G)}^\vee\big[D_0^{-1}\big]$$ is in fact [*injective*]{}: therefore, the specialization map $$\pi^\vee_{G/P} \; \colon \; {{\mathcal{O}}_q\big( G \big/ P \big)}^\vee \; \relbar\joinrel\relbar\joinrel\relbar\joinrel\twoheadrightarrow \; {{\mathcal{O}}_q\big( G \big/ P \big)}^\vee \Big/ (q-1) \, {{\mathcal{O}}_q\big( G\big/P \big)}^\vee$$ coincides with the restriction to $ {{\mathcal{O}}_q\big( G\big/P \big)}^\vee $ of the specialization map $$\pi^\vee_G \; \colon \; {{\mathcal{O}}_q(G)}^\vee\big[ D_0^{-1}\big] \; \relbar\joinrel\relbar\joinrel\relbar\joinrel\twoheadrightarrow \; {{\mathcal{O}}_q(G)}^\vee\big[D_0^{-1}\big] \Big/ (q-1) \, {{\mathcal{O}}_q(G)}^\vee\big[D_0^{-1}\big] \quad .$$ From now on, let $ \widehat{A} $ denote the $ (q-1) $–adic completion of any $ \Bbbk_q $–algebra $ A \, $. Note that $ \widehat{A} $ and $ A $ have the same specialization at $ \, q = 1 \, $, i.e. $ \; A / (q-1) \, A \; $ and $ \; \widehat{A} / (q-1) \, \widehat{A} \; $ are canonically isomorphic. When $ \, A = {\mathcal{O}}_q(G) \, $, note also that $ \, \widehat{{\mathcal{O}}_q(G)} \, $ is naturally a complete topological Hopf $ \Bbbk_q $–algebra. The next result show why it is relevant to introduce such completions. \[embedding\] $ {{\mathcal{O}}_q(G)}^\vee\big[D_0^{-1}\big] \, $ naturally embeds into $ \, \widehat{{{\mathcal{O}}_q(G)}^\vee} \, $. [[*Proof*]{}]{}. By remark \[qdp\] we have that $ {{\mathcal{O}}_q(G)}^\vee $ is generated by the elements $${\mathcal{D}}_- \, := \, {(q-1)}^{-1} \, \big( D_q^{-1} - 1 \big) \; , \quad \hskip5pt \chi_{ij} \, := \, {(q-1)}^{-1} \, \big( x_{ij} - \delta_{ij} \big) \eqno \forall \; i, j = 1, \dots, n$$ inside $ \, {\mathcal{O}}_q(G) \otimes_{\Bbbk_q} \Bbbk(q) \, $. On the other hand, observe that $$\begin{array}{c} x_{ij} \, = \, (q-1) \, \chi_{i,j} \, \in \, (q-1) \, {{\mathcal{O}}_q(G)}^\vee \, \qquad \forall \;\; i \not= j \\ \\ x_{\ell \ell} \, = \, 1 + (q-1) \, \chi_{\ell \ell} \, \in \, \big( 1 + (q-1) \, {{\mathcal{O}}_q(G)}^\vee \,\big) \, \qquad \forall \;\; \ell \; . \end{array} \leqno \text{and}$$ Then, if we expand explicitly the $ q $–determinant $ D_0 := D^{I_0} \, $, we immediately see that $ \; D_0 \in \big( 1 + (q-1) \, {{\mathcal{O}}_q(G)}^\vee \,\big) \; $ as well. Therefore $ D_0 $ is invertible in $ \widehat{{{\mathcal{O}}_q(G)}^\vee} $, and so the natural immersion $ \; {{\mathcal{O}}_q(G)}^\vee \lhook\joinrel\relbar\joinrel\longrightarrow \widehat{{{\mathcal{O}}_q(G)}^\vee} \; $ can be canonically extended to an immersion $ \; {{\mathcal{O}}_q(G)}^\vee\big[D_0^{-1}\big] \lhook\joinrel\relbar\joinrel\longrightarrow \widehat{{{\mathcal{O}}_q(G)}^\vee} \; $, q.e.d. \[specializations\] (a) The specializations at $ \, q \! = \! 1 \, $ of $ \, {{\mathcal{O}}_q(G)}^\vee \, $, $ \, {{\mathcal{O}}_q(G)}^\vee\big[D_0^{-1}\big] \, $ and $ \, \widehat{{{\mathcal{O}}_q(G)}^\vee} \, $ are canonically isomorphic. More precisely, the chain $${{\mathcal{O}}_q(G)}^\vee \; \lhook\joinrel\relbar\joinrel\longrightarrow \; {{\mathcal{O}}_q(G)}^\vee\big[D_0^{-1}\big] \; \lhook\joinrel\relbar\joinrel\longrightarrow \; \widehat{{{\mathcal{O}}_q(G)}^\vee}$$ of canonical embeddings induces at $ \, q = 1 \, $ a chain of isomorphisms. (b) $ {{\mathcal{O}}_q\big( G \big/ P \big)}^\vee \, $ embeds into $ \, \widehat{{{\mathcal{O}}_q(G)}^\vee} \, $ via the chain of embeddings $${{\mathcal{O}}_q\big( G \big/ P \big)}^\vee \; \lhook\joinrel\relbar\joinrel\longrightarrow \; {{\mathcal{O}}_q(G)}^\vee\big[D_0^{-1}\big] \; \lhook\joinrel\relbar\joinrel\longrightarrow \; \widehat{{{\mathcal{O}}_q(G)}^\vee}$$ (c) $ \displaystyle{ {{\mathcal{O}}_q\big( G \big/ P \big)}^\vee \,\;{\textstyle \bigcap}\;\, (q-1) \, \widehat{{{\mathcal{O}}_q(G)}^\vee} \,\; = \;\, (q-1) \, {{\mathcal{O}}_q\big( G \big/ P \big)}^\vee } \quad $. [[*Proof*]{}]{}. Part [*(a)*]{} and [*(b)*]{} are trivial, and [*(c)*]{} follows easily from them. Notice that part [*(c)*]{} of Corollary \[specializations\] also implies that $${{\mathcal{O}}_q\big( G \big/ P \big)}^\vee{\Big|}_{q=1} \, := \; {{\mathcal{O}}_q\big( G \big/ P \big)}^\vee \! \Big/ (q-1) \, {{\mathcal{O}}_q\big( G \big/ P \big)}^\vee$$ is a subalgebra of $$\widehat{{{\mathcal{O}}_q(G)}^\vee}{\Big|}_{q=1} \, = \; {{\mathcal{O}}_q(G)}^\vee{\Big|}_{q=1} \, := \; {{\mathcal{O}}_q(G)}^\vee \Big/ (q-1) \, {{\mathcal{O}}_q(G)}^\vee \, \cong \; U({\mathfrak{g}}^*)$$ just because the specialization map $$\pi_{G/P}^\vee \; \colon \; {{\mathcal{O}}_q\big( G\big/P \big)}^\vee \; \relbar\joinrel\relbar\joinrel\relbar\joinrel\twoheadrightarrow \; {{\mathcal{O}}_q\big( G\big/P \big)}^\vee \Big/ (q-1) \, {{\mathcal{O}}_q\big( G\big/P \big)}^\vee$$ coincides with the restriction to $ {{\mathcal{O}}_q\big( G\big/P \big)}^\vee $ of the specialization map $$\widehat{\pi_G^\vee} \; \colon \; \widehat{{{\mathcal{O}}_q(G)}^\vee} \; \relbar\joinrel\relbar\joinrel\relbar\joinrel\twoheadrightarrow \; \widehat{{{\mathcal{O}}_q(G)}^\vee} \Big/ (q-1) \, \widehat{{{\mathcal{O}}_q(G)}^\vee} \quad .$$ Now we want to see what is $ \, {{\mathcal{O}}_q\big( G\big/P \big)}^\vee {\Big|}_{q=1} $ inside $ \, U\big({{{\mathfrak{gl}}}_n}^{\!*}\big) \, $. In other words, we want to understand what is the space that $ {{\mathcal{O}}_q\big( G\big/P \big)}^\vee $ is quantizing. \[quantumspec\] $${{\mathcal{O}}_q\big( G \big/ P \big)}^\vee{\Big|}_{q=1} \, = \; U\big({{\mathfrak{p}}}^\perp\big)$$ as a subalgebra of $ \; {{\mathcal{O}}_q(G)}^\vee{\Big|}_{q=1} \! = U\big({{{\mathfrak{gl}}}_n}^{\!*}\big) \, $, where $ \, {{\mathfrak{p}}}^\perp \, $ is the orthogonal subspace to $ \, {{\mathfrak{p}}}:= \text{\it Lie}\,(P) \, $ inside $ \, {{{\mathfrak{gl}}}_n}^{\!*} \, $. [[*Proof*]{}]{}. Thanks to the previous discussion, it is enough to show that $$\pi_G^\vee\Big({{\mathcal{O}}_q\big( G \big/ P \big)}^\vee\Big) \, = \, U\big({{\mathfrak{p}}}^\perp\big) \, \subseteq \, U \big( {{{\mathfrak{gl}}}_n}^{\!*} \big) \, = \, {{\mathcal{O}}_q(G)}^\vee{\Big|}_{q=1} \quad .$$ To do this, we describe the isomorphism $ \, {{\mathcal{O}}_q(G)}^\vee {\Big|}_{q=1} \! \cong U\big({{{\mathfrak{gl}}}_n}^{\!*}\big) \, $ (cf. [@ga1]). First, recall that $ {{\mathcal{O}}_q(G)}^\vee $ is generated by the elements (see Remark \[qdp\]) $${\mathcal{D}}_- \, := \, {(q-1)}^{-1} \, \big( D_q^{-1} - 1 \big) \; , \quad \hskip5pt \chi_{ij} \, := \, {(q-1)}^{-1} \, \big( x_{ij} - \delta_{ij} \big) \eqno \forall \; i, j = 1, \dots, n$$ inside $ \, {\mathcal{O}}_q(G) \otimes_{\Bbbk_q} \Bbbk(q) \, $. In terms of these generators, the isomorphism reads $$\displaylines{ {{\mathcal{O}}_q(G)}^\vee{\Big|}_{q=1} \, \relbar\joinrel\relbar\joinrel\relbar\joinrel\longrightarrow \; U\big({{{\mathfrak{gl}}}_n}^{\!*}\big) \cr \overline{{{\mathcal{D}}}_-} \mapsto -({{\hbox{E}}}_{1,1} + \cdots + {{\hbox{E}}}_{n,n}) \; , \qquad \overline{\chi_{i,j}} \mapsto {{\hbox{E}}}_{i,j} \qquad \forall \;\; i \, , j \; . \cr }$$ where we used notation $ \; \overline{X} := X \mod (q-1) \, {{\mathcal{O}}_q(G)}^\vee \; $. Indeed, from $ \, \overline{\chi_{i,j}} \mapsto {{\hbox{E}}}_{i,j} \, $ and $ \, {(q-1)}^{-1} \, \big( D_q - 1 \big) \in {{\mathcal{O}}_q(G)}^\vee \, $, one gets $ \, \overline{D_q} \mapsto 1 \, $ and $ \, \overline{{(q-1)}^{-1} \, \big( D_q - 1 \big)} \mapsto {{\hbox{E}}}_{1,1} + \cdots + {{\hbox{E}}}_{n,n} \, $. Moreover, the relation $ \, D_q \, D_q^{-1} = 1 \, $ in $ {\mathcal{O}}_q(G) $ implies $ \; D_q \, {\mathcal{D}}_- = - {(q-1)}^{-1} \, \big( D_q - 1 \big) \; $ in $ {{\mathcal{O}}_q(G)}^\vee \, $, whence clearly $ \; \overline{{\mathcal{D}}_-} \mapsto -({{\hbox{E}}}_{1,1} + \cdots + {{\hbox{E}}}_{n,n}) \; $ as claimed (cf. [@ga1], §3, or [@ga3], §7). In other words, the specialization $ \, \pi_G^\vee \, \colon {{\mathcal{O}}_q(G)}^\vee \relbar\joinrel\relbar\joinrel\twoheadrightarrow U\big({{{\mathfrak{gl}}}_n}^{\!*}\big) \, $ is given by $$\pi_G^\vee \big( {\mathcal{D}}_- \big) \, = \, -({{\hbox{E}}}_{1,1} + \cdots + {{\hbox{E}}}_{n,n}) \; , \qquad \pi_G^\vee \big( \chi_{i,j} \big) \, = \, {{\hbox{E}}}_{i,j} \qquad \forall \;\; i \, , j \; .$$ If we look at $ \widehat{{{\mathcal{O}}_q(G)}^\vee} $, things are even simpler. Since $$D_q \, \in \, \Big( 1 + (q-1) \, {{\mathcal{O}}_q(G)}^\vee \Big) \subset \Big( 1 + (q-1) \, \widehat{{{\mathcal{O}}_q(G)}^\vee} \Big) \; ,$$ then $ \, D_q^{-1} \in \Big( 1 + (q-1) \, \widehat{{{\mathcal{O}}_q(G)}^\vee} \Big) \, $, and the generator $ \, {\mathcal{D}}_- \, $ can be dropped. The specialization map $ \widehat{\pi_{G/P}^\vee} $ of course is still described by formulæ as above. Now let’s compute $ \, \pi_{G/P}^\vee \Big( {{\mathcal{O}}_q\big( G\big/P \big)}^\vee \Big) = \widehat{\pi_G^\vee} \Big( {{\mathcal{O}}_q\big( G\big/P \big)}^\vee \Big) \, $. Recall that $ {{\mathcal{O}}_q\big( G\big/P \big)}^\vee $ is generated by the $ \mu_{i{}j} $’s, with $$\mu_{i{}j} \; := \; {(q-1)}^{-1} \, t_{i{}j} \; = \; {(q-1)}^{-1} \, {(-q)}^{r-j} \, {D^{1 \, \dots \, \widehat{j} \, \dots \, r \, i}}\, D_0^{-1}$$ for $ \, i = r+1, \dots, n \, $, and $ \, j = 1, \dots, r \, $; thus we must compute $ \, \widehat{\pi_G^\vee} \big( \mu_{i{}j} \big) \, $. By definition, for every $ \, i \not= j \, $ the element $ \, x_{i{}j} = (q-1) \, \chi_{i{}j} \, $ is mapped to 0 by $ \widehat{\pi_G^\vee} \, $. Instead, for each $ \ell $ the element $ \, x_{\ell\,\ell} = 1 + (q-1) \, \chi_{\ell\,\ell} \, $ is mapped to 1 (by $ \widehat{\pi_G^\vee} $ again). But then, expanding the $ q $–determinants one easily finds — much like in the proof of Lemma \[embedding\] — that $$\displaylines{ \widehat{\pi_G^\vee} \, \Big( {(q-1)}^{-1} \, {D^{1 \, \dots \, \widehat{j} \, \dots \, r \, i}}\Big) \; = \; \Big( {(q-1)}^{-1} \, {\textstyle \sum_{\sigma \in \, \mathcal{S}_r}} \, {(-q)}^{\ell(\sigma)} \, x_{1 \, \sigma(1)} \cdots x_{r \, \sigma(r)} \Big) \, = \hfill \cr \hfill = \; \widehat{\pi_G^\vee} \, \Big( {(q-1)}^{-1} \, {\textstyle \sum\limits_{\sigma \in \, \mathcal{S}_r}} {(-q)}^{\ell(\sigma)} \, \big( \delta_{1 \, \sigma(1)} + (q-1) \chi_{1 \, \sigma(1)}) \cdots \big( \delta_{1 \, \sigma(r)} + (q-1) \, \chi_{1 \, \sigma(r)}) \Big) \cr }$$ The only term in $ (q-1) $ in the expansion of $ {D^{1 \, \dots \, \widehat{j} \, \dots \, r \, i}}$ comes from the product $$\big( 1 + (q \! - \! 1) \chi_{1 \, 1}) \cdots \big( 1 + (q \! - \! 1) \chi_{r \,r} \big) \, (q \! - \! 1) \chi_{i \, j} \equiv (q \! - \! 1) \chi_{i \, j} \! \mod {(q \! - \! 1)}^2 {\mathcal{O}}\big(G \! \big/ \! P\big)$$ Therefore, from the previous analysis we get $$\displaylines{ \widehat{\pi_G^\vee} \, \Big( {(q-1)}^{-1} \, {D^{1 \, \dots \, \widehat{j} \, \dots \, r \, i}}\Big) \; = \; \widehat{\pi_G^\vee} \, \big( \chi_{i,j} \big) \; = \; {{\hbox{E}}}_{i,j} \cr \widehat{\pi_G^\vee} \, \big( D_0 \big) \; = \; \widehat{\pi_G^\vee} \big( 1 \big) \; = \; 1 \;\; , \qquad \widehat{\pi_G^\vee} \, \big( D_0^{-1} \big) \; = \; \widehat{\pi_G^\vee} \big(1 \big) \; = \; 1 \cr }$$ hence we conclude that $ \; \widehat{\pi_G^\vee} \big( \mu_{i{}j} \big) = {(-1)}^{r-j} \, {{\hbox{E}}}_{i,j} \; $, for all $ \, 1 \leq j \leq r < i \leq n \, $. The outcome is that $ \; \pi_{G/P}^\vee \Big( {{\mathcal{O}}_q\big( G\big/P \big)}^\vee \Big) \, = \, U(\mathfrak{h}) \; $, where $$\mathfrak{h} \, := \, \text{\it Span}\,\big( \big\{\, {{\hbox{E}}}_{i,j} \,\big|\, r+1 \leq i \leq n \, , \; 1 \leq j \leq r \,\big\} \big) \;\; .$$ On the other hand, from the very definitions and our description of $ \, {{{\mathfrak{gl}}}_n}^{\!*} \, $ one easily finds that $ \; \mathfrak{h} = {{\mathfrak{p}}}^\perp \; $, for $ \; {{\mathfrak{p}}}:= \text{\it Lie}\,(P) \; $. The claim follows. Proposition \[quantumspec\] claims that $ {{\mathcal{O}}_q \big( G\big/P \big)}^\vee $ is a quantization of $ U \big( {{\mathfrak{p}}}^\perp \big) \, $, i.e. it is a unital $ \Bbbk_q $–algebra whose semiclassical limit is $ U\big({{\mathfrak{p}}}^\perp\big) \, $. Now, the fact that $ U\big({{\mathfrak{p}}}^\perp\big) $ describes (infinitesimally) a homogeneous space for $ G^* $ is encoded in algebraic terms by the fact that it is a (left) coideal of $ U({\mathfrak{g}}^*) \, $; in other words, $ U\big({{\mathfrak{p}}}^\perp\big) $ is a (left) $ U({\mathfrak{g}}^*) $–comodule w.r.t. the restriction of the coproduct of $ U({\mathfrak{g}}^*) \, $. Thus, for $ {{\mathcal{O}}_q \big( G\big/P \big)}^\vee $ to be a quantization of $ U\big({{\mathfrak{p}}}^\perp\big) $ [*as a homogeneous space*]{} we need also a quantization of this fact: namely, we would like $ {{\mathcal{O}}_q \big( G\big/P \big)}^\vee $ to be a left coideal of $ {{\mathcal{O}}_q(G)}^\vee $, our quantization of $ U({\mathfrak{g}}^*) \, $. But this makes no sense at all, This problem leads us to enlarge a bit our quantizations $ {{\mathcal{O}}_q \big( G\big/P \big)}^\vee $ and $ {{\mathcal{O}}_q(G)}^\vee \, $: we take their $ (q \! - \! 1) $–adic completions, namely $ \widehat{{{\mathcal{O}}_q \big( G\big/P \big)}^\vee} $ and $ \widehat{{{\mathcal{O}}_q(G)}^\vee} \, $. While not affecting their behavior at $ \, q = 1 \, $ (i.e., their semiclassical limits are the same), this operation solves the problem. Indeed, $ \widehat{{{\mathcal{O}}_q(G)}^\vee} $ is big enough to contain $ {{\mathcal{O}}_q \big( G\big/P \big)}^\vee \, $, by Corollary \[specializations\][*(b)*]{}. Then, as $ \widehat{{{\mathcal{O}}_q(G)}^\vee} $ is a topological Hopf algebra, inside it we must look at the closure of $ {{\mathcal{O}}_q\big( G\big/P \big)}^\vee \, $. Thanks to Corollary \[specializations\][*(c)*]{} (which means, roughly, that an Artin-Rees lemma holds), the latter is nothing but $ \widehat{{{\mathcal{O}}_q \big( G\big/P \big)}^\vee} $. Finally, next result tells us that $ \widehat{{{\mathcal{O}}_q \big( G\big/P \big)}^\vee} $ is a left coideal of $ \widehat{{{\mathcal{O}}_q(G)}^\vee} $, as expected. $ \, \widehat{{{\mathcal{O}}_q\big( G \big/ P \big)}^\vee} \, $ is a left coideal of $ \; \widehat{{{\mathcal{O}}_q(G)}^\vee} \, $. [[*Proof*]{}]{}. Recall that the coproduct $ \widehat{\Delta} $ of $ \widehat{{{\mathcal{O}}_q(G)}^\vee} $ takes values in the [*topological*]{} tensor product $ \, \widehat{{{\mathcal{O}}_q(G)}^\vee} \,\widehat{\otimes}\, \widehat{{{\mathcal{O}}_q(G)}^\vee} \, $, which by definition is the $ (q \! - \! 1) $–adic completion of the [*algebraic*]{} tensor product $ \, \widehat{{{\mathcal{O}}_q(G)}^\vee} \otimes \widehat{{{\mathcal{O}}_q(G)}^\vee} \, $. Our purpose then is to show that this coproduct $ \widehat{\Delta} $ maps $ \widehat{{{\mathcal{O}}_q \big( G\big/P \big)}^\vee} $ in the topological tensor product $ \, \widehat{{{\mathcal{O}}_q(G)}^\vee} \,\widehat{\otimes}\, \widehat{{{\mathcal{O}}_q \big( G\big/P \big)}^\vee} \, $. By construction, the coproduct of $ {{\mathcal{O}}_q(G)}^\vee $, hence of $ \widehat{{{\mathcal{O}}_q(G)}^\vee} $ too, is induced by that of $ {\mathcal{O}}_q(G) \, $, say $ \; \Delta : {\mathcal{O}}_q(G) {\longrightarrow}{\mathcal{O}}_q(G) \otimes {\mathcal{O}}_q(G) \; $. Now, the latter can be uniquely (canonically) extended to a coassociative algebra morphism $${{\widetilde{\Delta}}}: \, {\mathcal{O}}_q(G)\big[D_{{I_0}}^{-1}\big] \relbar\joinrel{\longrightarrow}{\mathcal{O}}_q(G)\big[D_{{I_0}}^{-1}\big] \,\widetilde{\otimes}\, {\mathcal{O}}_q(G)\big[D_{{I_0}}^{-1}\big]$$ where $ \, \widetilde{\otimes}\, $ is the $ J_\otimes $–adic completion of the algebraic tensor product, with $$J_\otimes \; := \; J \otimes {\mathcal{O}}_q(G) \, + \, {\mathcal{O}}_q(G) \otimes J \;\; , \qquad J \, := \text{\it Ker}\,\big(\epsilon_{{\mathcal{O}}_q(G)}\big) \;\; .$$ In fact, since $ \; \Delta(D_0) = D_0 \otimes D_0 + {\textstyle \sum\limits_{K \neq {I_0}}} D^{I_0}_K \otimes D^K \; $, one easily computes $$\displaylines{ {{\widetilde{\Delta}}}(D_0) \; = \; \Big(\, 1 + {\textstyle \sum\limits_{K \neq {I_0}}} \, D_K^{I_0} \, D_0^{-1} \otimes D^K D_0^{-1} \,\Big) \big( D_0 \otimes D_0 \big) \cr {{\widetilde{\Delta}}}\big(D_0^{-1}\big) \; = \; {\big( D_0 \otimes D_0 \big)}^{-1} {\Big(\, 1 + {\textstyle \sum\limits_{K \neq {I_0}}} \, D_K^{I_0} \, D_0^{-1} \otimes D^K D_0^{-1} \,\Big)}^{-1} \cr \hfill = \, \Big( D_0^{-1} \otimes D_0^{-1} \Big) \; {\textstyle \sum\limits_{n \geq 0}} \, {(-1)}^n \, {\bigg(\, {\textstyle \sum\limits_{ K \neq {I_0}}} \, D_K^{I_0} \, D_0^{-1} \otimes D^K D_0^{-1} \,\bigg)}^{\!n} \cr }$$ Let’s now look at the restriction $ {{\widetilde{\Delta}}}_r $ of $ {{\widetilde{\Delta}}}$ to $ {\mathcal{O}}_q^{\,\text{\it loc}}\big( G\big/P \big) \, $. We have $$\displaylines{ {{\widetilde{\Delta}}}_r(t_{ij}) \; = \; {{\widetilde{\Delta}}}_r \big( {D^{1 \, \dots \, \widehat{j} \, \dots \, r \, i}}\, D_0^{-1} \,\big) \; = \; {{\widetilde{\Delta}}}\big({D^{1 \, \dots \, \widehat{j} \, \dots \, r \, i}}\big) \cdot {{{\widetilde{\Delta}}}\big( D_0 \big)}^{-1} \; = \hfill \cr \hfill = \, \bigg(\, {\textstyle \sum\limits_L} \, {D^{1 \, \dots \, \widehat{j} \, \dots \, r \, i}}_{\,L} \, D_0^{-1} \otimes D^L \, D_0^{-1} \,\bigg) \cdot {\textstyle \sum\limits_{n \geq 0}} \, {(-1)}^n \, {\bigg(\, {\textstyle \sum\limits_{K \neq {I_0}}} \, D_K^{I_0} \, D_0^{-1} \otimes D^K D_0^{-1} \,\bigg)}^{\!n} \cr }$$ Now, by Proposition \[classical\] we know that each product $ \, D^L \, D_{I_0}^{-1} \, $ is a combination of the $ t_{ij} $’s. Hence the formula above shows that $ {{\widetilde{\Delta}}}_r $ maps $ {\mathcal{O}}_q^{\,\text{\it loc}}\big( G\big/P \big) $ into $ \; {\mathcal{O}}_q(G)\big[D_0^{-1}\big] \,\widetilde{\otimes}\, {\mathcal{O}}_q^{\,\text{\it loc}}\big( G\big/P \big) \; $. By scalar extension, $ {{\widetilde{\Delta}}}$ uniquely extends to a map defined on the $ \Bbbk(q) $–vector space $ \, \Bbbk(q) \otimes_{\Bbbk_q} {\mathcal{O}}_q(G)\big[D_0^{-1}\big] \, $, which we still call $ {{\widetilde{\Delta}}}\, $. Its restriction to the similar scalar extension of $ {\mathcal{O}}_q^{\,\text{\it loc}}\big( G\big/P \big) $ clearly coincides with the scalar extension of $ {{\widetilde{\Delta}}}_r \, $, hence we call it $ {{\widetilde{\Delta}}}_r $ again. Finally, the restriction of $ {{\widetilde{\Delta}}}$ to $ {{\mathcal{O}}_q(G)}^\vee\big[D_0^{-1}\big] $ and of $ {{\widetilde{\Delta}}}_r $ to $ {{\mathcal{O}}_q\big( G\big/P \big)}^\vee $ both coincide — by construction — with the proper restrictions of the coproduct of $ \widehat{{{\mathcal{O}}_q(G)}^\vee} $ (cf. Corollary \[specializations\]). In the end, we are left to compute ${{\widetilde{\Delta}}}_r(\mu_{ij}) \, $. The computation above gives $$\displaylines{ \widehat{\Delta}(\mu_{ij}) \; = \; {{\widetilde{\Delta}}}_r(\mu_{ij}) \; = \; {(q \! - \! 1)}^{-1} \, {{\widetilde{\Delta}}}_r(t_{ij}) \; = \; \hfill \cr \hfill = {(q \! - \! 1)}^{-1} \, {\textstyle \sum\limits_L} \, {D^{1 \, \dots \, \widehat{j} \, \dots \, r \, i}}_{\,L} \, D_0^{-1} \otimes D^L \, D_0^{-1} \cdot {\textstyle \sum\limits_{n \geq 0}} {(-1)}^n {\bigg({\textstyle \sum\limits_{K \neq {I_0}}} \! D^{I_0}_K \, D_0^{-1} \otimes D^K D_0^{-1} \bigg)}^{\!n} \cr }$$ Now, each left-hand side factor above belongs to $ \, \widehat{{{\mathcal{O}}_q(G)}^\vee} \,\widehat{\otimes}\, \widehat{{{\mathcal{O}}_q \big( G\big/P \big)}^\vee} \, $, because either $ \, D^L \in J_{G/P}^{\,\text{\it loc}} \, $ (if $ \, L \not= I_0 \, $, with notation of §4.3), or $ \, {D^{1 \, \dots \, \widehat{j} \, \dots \, r \, i}}_{\,L} \in J \, $ (if $ \, L = I_0 \, $, with $ \, J := \text{\it Ker}\, \big(\epsilon_{{\mathcal{O}}_q(G)}\big) \, $). On right-hand side instead we have $$D^K \, \in \, J_{G/P}^{\,\text{\it loc}} \, \subseteq \, (q-1) \, {{\mathcal{O}}_q\big( G\big/P \big)}^\vee \;\; , \qquad D_K^{I_0} \, \in \, J \, \subseteq \, (q-1) \, {{\mathcal{O}}_q(G)}^\vee$$ whence — as $ \, D_0^{-1} \in \widehat{{{\mathcal{O}}_q(G)}^\vee} \, $ and $ \, D_0^{-1} \in \widehat{{{\mathcal{O}}_q\big( G\big/P \big)}^\vee} \, $ — we get $$\! {\textstyle \sum\limits_{K \neq {I_0}}} D_K^{I_0} \, D_0^{-1} \otimes D^K D_0^{-1} \; \in \; {(q-1)}^2 \, \widehat{{{\mathcal{O}}_q(G)}^\vee} \,\widehat{\otimes}\, \widehat{{{\mathcal{O}}_q\big( G\big/P \big)}^\vee}$$ so that $ \quad \displaystyle{ {\textstyle \sum\limits_{n \geq 0}} \, {(-1)}^n \, {\bigg(\! {\textstyle \sum\limits_{K \neq {I_0}}} D^{I_0}_K \, D_0^{-1} \otimes D^K D_0^{-1} \bigg)}^{\!n} \, \in \; \widehat{{{\mathcal{O}}_q(G)}^\vee} \,\widehat{\otimes}\, \widehat{{{\mathcal{O}}_q\big( G\big/P \big)}^\vee} \quad . } $ The final outcome is $ \; \widehat{\Delta}(\mu_{ij}) \in \widehat{{{\mathcal{O}}_q(G)}^\vee} \,\widehat{\otimes}\, \widehat{{{\mathcal{O}}_q\big( G\big/P \big)}^\vee} \; $ for all $ i $, $ j \, $. As the $ \mu_{ij} $’s topologically generate $ \, \widehat{{{\mathcal{O}}_q \big( G\big/P \big)}^\vee} \, $, this proves the claim. In the end, we get the main result of this paper. \[maintheorem\] $ \, \widehat{{{\mathcal{O}}_q\big( G \big/ P \big)}^\vee} \, $ is a quantum homogeneous $ \, G^* $–space, which is an infinitesimal quantization of the homogeneous $ \, G^* $–space $ {{\mathfrak{p}}}^\perp \, $. [[*Proof*]{}]{}. Just collect the previous results. By Proposition \[quantumspec\] and by the fact that $ \; \widehat{{{\mathcal{O}}_q \big( G\big/P \big)}^\vee}{\Big|}_{q=1} = \, {{\mathcal{O}}_q\big( G\big/P \big)}^\vee{\Big|}_{q=1} \; $ we have that the specialization of $ \widehat{{{\mathcal{O}}_q\big( G\big/P \big)}^\vee} $ is $ U \big( {{\mathfrak{p}}}^\perp \big) \, $. Moreover we saw that $ \widehat{{{\mathcal{O}}_q \big( G\big/P \big)}^\vee} $ is a subalgebra, and left coideal, of $ \widehat{{{\mathcal{O}}_q(G)}^\vee} \, $. Finally, we have $$\widehat{{{\mathcal{O}}_q\big( G\big/P \big)}^\vee} \,\;{\textstyle \bigcap}\;\, (q-1) \, \widehat{{{\mathcal{O}}_q(G)}^\vee} \,\; = \;\, (q-1) \, \widehat{{{\mathcal{O}}_q\big( G\big/P \big)}^\vee}$$ as an easy consequence of Corollary \[specializations\] [*(c)*]{}. Therefore, $ \widehat{{{\mathcal{O}}_q\big( G\big/P \big)}^\vee} $ is a quantum homogeneous space, in the usual sense. As $ \widehat{{{\mathcal{O}}_q(G)}^\vee} $ is a quantization of $ {\mathfrak{g}}^* \, $, we have that $ \widehat{{{\mathcal{O}}_q\big( G \big/ P \big)}^\vee} $ is in fact a quantum homogeneous space for $ G^* \, $; of course, this is a quantization of [*infinitesimal*]{} type. All these computations can be repeated, step by step, taking $ \, G = {SL}_n \, $ and $ \, P = P_1 \, $. [99]{} V. Chari, A. Pressley, [*Quantum Groups*]{}, Cambridge Univ. Press, Cambridge (1994). N. Ciccoli, R. Fioresi, F. Gavarini, [*Quantum Duality Principle for Projective Homogeneous Spaces*]{}, J. Noncommut. Geom. [**2**]{} (2008), 449–496. N. Ciccoli, F. Gavarini, [*A quantum duality principle for coisotropic subgroups and Poisson quotients*]{}, Adv. Math. [**199**]{} (2006), 104–135. C. De Concini, D. Eisenbud, C. Procesi, [*Young Diagrams and Determinantal Varieties*]{}, Invent. Math. [**56**]{} (1980), 129–165. V. G. Drinfeld, [*Quantum groups*]{}, Proc. Intern. Congress of Math. (Berkeley, 1986) (1987), 798–820. R. Fioresi, [*Quantum deformation of the Grassmannian manifold*]{}, J. Algebra [**214**]{} (1999), 418–447. R. Fioresi, [*A deformation of the big cell inside the Grassmannian manifold $G(r,n)$*]{}, Rev. Math. Phys. [**11**]{} (1999), 25–40. F. Gavarini, [*Quantum function algebras as quantum enveloping algebras*]{}, Comm. Algebra [**26**]{} (1998), 1795–1818. F. Gavarini, [*The quantum duality principle*]{}, Ann. Inst. Fourier (Grenoble) [**52**]{} (2002), 809–834. F. Gavarini, [*The global quantum duality principle: theory, examples, and applications*]{}, 120 pages, see [http:/$\!$/arxiv.org/abs/math.QA/0303019]{} (2003). F. Gavarini, [*The global quantum duality principle*]{}, J. Reine Angew. Math. [**612**]{} (2007), 17–33. V. Lakshmibai, N. Reshetikhin, [*Quantum flag and Schubert schemes*]{}, Contemp. Math. [**134**]{}, Amer. Math. Soc., Providence, RI (1992), 145–181. E. Taft, J. Towber, [*Quantum deformation of flag schemes and Grassmann schemes, I. A $q$-deformation of the shape-algebra for ${\rm GL}(n)$*]{}, J. Algebra [**142**]{} (1991), 1–36. [^1]: Partially supported by the University of Bologna, funds for selected research topics.

ArXiv

--- abstract: 'This text offers reminiscences of my personal interactions with Roman Jackiw as a way of looking back at the very fertile period in theoretical physics in the last quarter of the 20th century.' author: - | Luc Vinet\ \ title: 'Roman Jackiw: A Beacon in a Golden Period of Theoretical Physics' --- *To Roman: a bouquet of recollections as an expression of friendship.* Introduction ============ I owe much to Roman Jackiw: my postdoctoral fellowship at MIT under his supervision has shaped my scientific life and becoming friend with him and So Young Pi has been a privilege. Looking back at the last decades of the past century gives a sense without undue nostalgia, I think, that those were wonderful years for Theoretical Physics, years that have witnessed the preeminence of gauge field theories, deep interactions with modern geometry and topology, the overwhelming revival of string theory and remarkably fruitful interactions between particle and condensed matter physics as well as cosmology. Roman was a main actor in these developments and to be at his side and benefit from his guidance and insights at that time was most fortunate. Owing to his leadership and immense scholarship, also because he is a great mentor, Roman has always been surrounded by many and has thus generated a splendid network of friends and colleagues. Sometimes, with my own students, I reminisce about how it was in those days; I believe it is useful to keep a memory of the way some important ideas shaped up and were relayed. Hence as a tribute to Roman, I thought of writing the following short account of my personal connections with him in addition to the scientific hommage written in collaboration with my colleagues Nicolas Crampé and Rafael Nepomechie. I fully appreciate that my own little history is of no special interest but I am offering this text as an illustrative testimony of a vibrant intellectual period in the companionship of Roman and of other scientists who like him were larger than life. First encounters with Roman and his work as a graduate student ============================================================== The Centre de Recherches Mathématiques broadly known as the CRM, was founded in 1968. One only appreciates with hindsight which butterfly wing flaps will have a determining effect on your life. In my case one of these was the Prague Spring and its repression which occurred in 1968. As a result, two outstanding Tchech physicists, Jiri Patera and Pavel Winternitz educated in the highest Soviet scientific tradition took up positions at the CRM in the following years. They then gave to Montreal a big research impetus and developed in this city a strong school in mathematical physics. I was fortunate to join the CRM and to pursue graduate studies within their group in the mid 70s. In theoretical physics, the end of the 60s saw the advent of the Weinberg-Salam model [@Wein], [@Salam] unifying the electro-weak interactions, followed by the proof of its renormalizability[@Hooft], the development of QCD [@GW; @P], which had been preceded by the discovery of the Adler-Bell-Jackiw anomaly [@A; @BJ] (beautifully presented in [@J]). As a young student, I very much wished to get involved in those striking developments in field theory. With a much enthusiastic postdoctoral fellow then at the CRM, John Harnad, who de facto became my co-supervisor, we started a gauge theory seminar. The importance of classical solutions with topological properties such as magnetic monopoles and instantons in non-perturbative analyses was being revealed at about that time. The review on monopoles by Goddard and Olive [@GO] as well as Coleman’s Erice lectures [@C] were extremely formative. In those years, the Theory Division of the Canadian Association of Physicists was organizing Summer schools in Banff. This is where I met Roman for the first time in 1977 as he delivered lectures [@JNR] on classical solutions of the Yang-Mills theory that much impressed me. Two papers on this topic which appeared around that time had a big influence on my Ph.D. work: the first by Roman and his collaborator Rebbi [@JR] (and reviewed in [@JNR]) where the conformal $SO(5)$ invariance of the BPST one-instanton solution is identified and the second by Witten [@W] where a muti-instanton solution is obtained through the dimensional reduction under $SO(3)$ of the self-dual Yang-Mills equations to a Higgs model in curved two-dimensional space. The essential aspect in these studies is that the variance of the Yang-Mills field under the space-time transformation is compensated by a gauge transformation. These publications prompted John Harnad and myself with a geometer, Steve Shnider, then at McGill University to look systematically at the description of gauge fields that are invariant in that sense under space-time transformations. We treated this problem in the global formalism of fiber bundle theory that had gained popularity amongst physicists in part thanks to a paper by Wu and Yang [@WY]; we classified lifts of (certain) group actions on manifolds to principal bundles over such bases and then characterized the invariant connections under the lifted actions. We then applied these results to obtain solutions to the Yang-Mills equations on compactified Minkowski space. This formed the core of my thesis and the paper on the general framework remains one of my most cited. At some point in the course of these investigations, Steve made a preliminary presentation [@HSV] at a conference in Lawrence, Kansas in the Summer of 1978 which focused on the twistorial approach to instantons. Using this opportunity to take his son Nick as I recall on a road trip across America, Roman was one of the principal speakers [@J2] at this meeting which was my first occasion to present some of my work to him. Those questions regarding invariant gauge fields were timely then and in fact were addressed simultaneously by Forgacs and Manton [@FM] in a complementary fashion using infinitesimal methods. As I was completing these projects, in 1979, I had the occasion to accompany my co-supervisor Pavel Winternitz during his sabbbatical to Saclay. This is how I obtained a Doctorate from the Université Pierre et Marie Curie for work I had done separately on the classification of second order differential equations in two dimensions that are invariant under subgroups of the conformal group. Upon my return to Montreal, I was wrapping up my Ph.D. and the question came as to where I should go for my postdoctoral fellowship. To explore the possibility of becoming a Research Assistant at MIT, John Harnad extended to Roman an invitation to visit Montreal and this is how a few months later I was moving to Cambridge. We especially remember from that visit that Roman was asking if we knew the three-form whose exterior derivative would give the Chern four-form. He clearly was on his way to developing the Chern-Simons gauge theories, we still regret that at the time we did not have the culture to provide the answer and join him in these investigations. The postdoctoral years at MIT ============================= In some sense Roman is responsible for my meeting my wife. I had become desperate to find an apartment in the Boston area and thought of checking where Roman lived. As it happened I looked in an old phone book that had 808 Memorial drive as his address. It is essentially where MIT ends and where Harvard begins. As I discovered later, Roman only stayed there briefly, but I found an apartment in that building with a nice view of the Charles. The apartment number was 512. A few months later, a charming girl named Letitia rented apartment 1212. Born in Montreal, she came to take up a research engineer position at MIT’s Lincoln Lab. Meeting her has been one of the best things that have happened in my life. When I arrived in September 1980 at MIT’s Center for Theoretical Physics (CTP), Nick Manton had already been there for one year also as a postdoc and by that time had co-authored a paper [@JM] with Roman applying the theory of invariant gauge fields to the determination of constants of motion in their background. It was great to get to meet Nick, the disappointment came when Roman told me that he had achieved what he wanted to do in relation with my thesis work and was moving on to different problems. Having thus been taught that timing is of importance, I carried on and dove with excitement, delight and some trepidations in the amazing environment that the CTP was providing. It is not really possible to describe in a few words how vibrant the CTP was. The Faculty formed a truly outstanding group. Altogether they received numerous recognitions. During my postdoctoral fellowship alone in 1981, Jeffrey Goldstone obtained the Dannie Heineman Prize (which Roman got in 1995) and Viky Weisskopf was awarded the Wolf Prize. I also recall a meeting in honour of Francis Low who was the Provost of MIT on the occasion of his 60th birthday. This was the time when Alan Guth, a former student of Low who had joined the MIT Faculty in 1980 was developing an inflation cosmology [@G] and when So -Young Pi was much involved in these developments [@GP; @Pi]. And there was Roman who was attracting like a magnet a very large number of highly talented Ph.D. students and postdocs. Many of them are actually contributing to this festschrift with Antti Niemi one of the instigators of this project. It was a privilege to be part of that group which boasted incredible creativity of intensity. Under the leadership of Roman and thanks to his inspiration many important chapters of theoretical physics were developed through various collaborations involving members of that team, among these advances are the studies of the fermion-vortex system and the fermion number fractionization, the development of the Liouville field theory, the topologically massive (Chern-Simons) gauge theories and many more. It was hard to keep track with all these papers “typeset in TeX by Roger L. Gilson". The ‘brown bag" lunch seminar at the CTP certainly created great occasions for exchanges and the generation of research ideas. The interactions with Harvard were also highly stimulating. I have much benefitted for instance of the lectures on supergravity given by Steven Weinberg on one of his extended visits. The joint Harvard-MIT theoretical physics seminar in particular was a Thursday ritual not to be missed. Much appreciated also were the dinners with the speaker that followed at which the postdocs were invited. The meals at the Yenching chinese restaurant near Harvard square with Sidney Coleman placing the order bring fond memories. I keep saying that this is where I learned to use chopsticks, out of necessity, since I did not manage to eat much because of my poor dexterity the first few times I attended. In 1981, my second year at the CTP, Eric D’Hoker arrived at MIT for a three-year postdoctoral fellowship and so did Eddie Farhi to take up a Junior Faculty appointment. Eddie and Roman edited a book at that time on dynamical gauge symmetry breaking [@FJ]. Eric and I became close friends and began collaborating on problems that were often prompted by Roman. Thanks to the ‘Bruno Rossi" exchange program between the INFN and MIT and Sergio Fubini in particular, the CTP entertained close ties with Italian physicists. In a paper published in 1976, de Alfaro, Fubini and Furlan have explored the ramifications of scale invariance in one-dimensional quantum mechanics [@DFF]; building on this work in a more physically motivated context, Roman identified in 1980 the conformal symmetries of the magnetic monopole [@J3]. At some point Roman passed on to Eric and I manuscript notes from Fubini that led to the paper with Rabinovici developing superconformal quantum mechanics [@FR] and which stemmed from a lot of on-going interest in supersymmetric theories (see for example [@W1]). That led us to consider the supersymmetries of monopole systems which we did (see for instance[@DV1]) by extending the work of Roman. We carried on studies [@DV2] that built on the seminal work of Deser, Jackiw and Templetion on topologically massive theories [@DJT]; we did work then on reduction of higher dimensional Chern-Simons theories that for some reason we never published. Those postdoctoral years at MIT largely thanks to Roman were indeed defining ones. When I left Cambridge and the end of my fellowship, I promised myself to return as soon as possible. Returning to MIT as a young Faculty at Université de Montréal ============================================================= In 1982, holding University Research Fellowship from NSERC, I came back to my hometown to take on an Assistant Professor position at the Université de Montréal. I kept reaping the benefits of my sojourn at MIT: the first postdoc I hired was Haris Panagopoulos whom I had met there; I arranged for Manu Paranjape a fellow Canadian and former student of Goldstone with whom I also had become acquainted at the CTP to come to Montreal as Assistant Professor at UdeM and I kept working with Eric D’Hoker as he moved himself to Columbia University. In 1987 as I was presenting my dossier for tenure, I held the promise I had made to myself and returned to the CTP as Visiting Researcher for six months at the beginning of the year. I then shared an office with Oscar Eboli, a most charming Brazilian fellow who was working with Roman and So-Young on quantum fields out of thermal equilibrium [@EJP] and with whom I had the pleasure to reconnect in Sao Paulo in the Summer of 2018. At the time So-Young and Roman’s son Stefan was two years old. Shortly after my arrival, Roman introduced me to a brilliant young Italian physicist named Roberto Floreanini. The two were working on the functional approach to quantum field theories [@FHJ]. It clicked immediately with Roberto. I embarked in that program and co-authored a few articles (e.g. [@FV1]) with Roberto before I returned to Montreal. Once again, because of MIT, I had not only found a superb collaborator but met someone who became a dear friend. And so Roberto and I collaborated intensively until to his disappointment I suppose, I became Provost at McGill University. As always, it was truly enriching scientifically and personally to be around Roman. I was living during that stay at Longfellow Place near So-Young and Roman’s home in Beacon Hill. I recall one evening when Roman was giving me a lift in his BMW with the iconic “FFDUAL" plates, the radio was on and he was quite taken by the music saying how beautiful it was. He asked : do you know what it is? I was not completely sure but I thought I had recognized Richard Strauss and said so. A few minutes after I had passed my door, I have a phone call, it was Roman to tell me: you were right it was Ariadne auf Naxos. A cute anecdote to recall how classy Roman is. At the time in the first half of 1987, Roman was also examining issues connected to Berry potentials [@J4]. He shared with me the thought that the symmetries of a problem could determine the associated Berry connections. By then, I was back in Montreal. It was quite nice to think that the theory of invariant connections of my PhD days could be brought to bear on this current topic. I sorted this out and wrote a draft. At this point I had not co-authored papers with Roman. I then thought the moment had come to loop the loop and publish something with Roman on the topic that had initially brought me to MIT. In spite of my insistence, Roman declined arguing that I should publish the paper alone [@V]. He subsequently wrote an article [@J5] covering the question in his own way and kindly quoted my publication. Although this left me with the regret of never having had the pleasure of chiseling a text with Roman this turned out to be some kind of blessing for me and (I apologize Roman) a certain curse for him. Indeed as you are all well aware, in any grant application process one needs to suggest prestigious colleagues with whom you have never collaborated (or at least not within a certain period). I have abused of Roman in this respect but magnanimously he always obliged and has been very supportive. During the academic year 1989-1990, Eric D’Hoker kindly hosted me at UCLA where he had moved. Our first son, Jean-François, was born in LA in July 1990. Roberto Floreanini visited and this is when we launched as we were babysitting what became a vast study of the connections between quantum algebras, q-special functions and their applications. Meeting here and there ====================== The occasion for another prolonged visit at the CTP never presented itself again. Nevertheless life, family and common scientific interests provided opportunities for So-Young, Roman, Letitia and I to get together sometimes with our children and those were always very happy moments. Our friendship built over the years but I think that it certainly strengthened while we were all together in Banff in 1989. I have mentioned before that these Summer schools in the Canadian Rockies which provided my first meeting with Roman as a student, were an institution. This one entitled Physics, Geometry and Topology had a stellar group of lecturers including Roman who spoke about planar physics [@J6]. Although very successful it turned out to be the last school of that nice series because the original sources of funding disappeared. I had enjoyed them so much that I had the idea to revive them under the name *CRM Summer School in Banff* when I became director of the CRM in the nineties. This turned out to be the precursor of what is today the highly successful Banff International Research Station (BIRS) which is jointly supported by the CONACYT, the NSF and NSERC. A peculiar encounter with Roman and one that brings laughs in retrospect occurred in Kiev in 1992. There was a conference organized by the Ukrainian Academy of Sciences. Roman was participating to connect with his childhood days I believe, and I was attending because this conference was taking place immediately after a meeting in Alushta, Crimea at which I had been invited by Alexei Morozov and other friends from ITEP in Moscow. Phong who is a member of the Department of Mathematics at Columbia University and a collaborator of Eric D’Hoker was also in Crimea. Based on our understanding of what ikra meant in Russian, Phong and I were proudly thinking that we had managed to buy nice caviar. And so I arrived in Kiev with two tin cans of this great finding. My father who had wished to visit Crimea was accompanying me. Once in our room in the hotel of the Academy we noticed that there was no hot water. We got together with Roman at some point to confirm that he did not have hot water either because this time in the Summer had been chosen by the hotel management to perform plumbing work throughout the building. Facing this adversity, we told Roman about our caviar and invited him in our room with the hope of indulging in this delicacy only to find after some struggle to open the cans that they contained some dry and highly salted red fish eggs. So much for our refined party! We have also met in less exotic places like Boston or New York and at times our reunions were prompted by musical reasons. Stefan Jackiw is a magnificent violonist who is unanimously recognized as belonging to an elite group of only a few. My son Jean-François, has done musical performance training as a violist to a high level even though he chose not to pursue a professional career. The 2006 edition of the International Conference on Group Theoretical Methods in Physics (ICGTMP) was held in New York. I took along Jean-François and his younger brother Laurent under the condition that they visit one museum every day, an assignment that they fulfilled. Stefan was already living in New York and So-Young and Roman who were also staying in Manhattan at that moment very kindly arranged for the six of us to have a joyful dinner after Stefan had kindly practiced scales with Jean-François. In 2009, So-Young organized a gathering for Roman’s 70th birthday in Boston which was a lovely event and another occasion to celebrate Roman’s outstanding impact on science and people. For five years in a row, beginning in 2010, Jean-François participated in the Aspen Music Festival and School. We have always managed to spend some family time in Aspen during those Summers. In some of those years Stefan has been a guest artist. I recall in particular being subjugated one evening by an interpretation he gave of the third Brahms sonata. So-Young and Roman were also involved regularly in the Aspen Center for Physics and our stays often overlapped. We never missed the chance to get together and I vividly remember great dinners at the Pine Creek Cookhouse at the base of the Elk Mountains. Roman and the CRM and Montreal ============================== As I bring these reminiscences to a close, I want to stress how generous Roman has been with his visits to Montreal and the connections he has built with this city and its scientific organizations. I shall point at three moments in particular. In 1988, Yvan Saint-Aubin and I organized the XVIIth ICGTMP in Montreal. This international conference took place against the backdrop of perestroika which created very fortunate circumstances that allowed proeminent Soviet mathematicians and scientists such as Belavin, Faddeev, Fatteev, Manin and Zamolodchikov to attend and lecture at the meeting. For many, this was their first visit to North America. The list of plenary speakers was stunning and we were lucky to have Roman [@J7] (who also gave an additional talk [@J8]) and So-Young [@Pi2] among them. The recipient of the Wigner medal awarded during the conference was Isadore Singer from MIT whose celebrated index theorems are so intimately connected with Roman’s work. This made for a superb program; the conference was very successful and many of the MIT friends (D’Hoker, Eboli, Floreanini, Niemi, Panagopoulos, etc) attended. For a while Roman enjoyed smoking little cigars and I wondered if this was not what explained why he appreciated Montreal: cuban cigars could be found in this trendy Canadian city! Indeed whenever in town Roman would make sure to stop at the Davidoff store to stock up. There was thus a time when as a friendly gesture, I would make sure of smuggling some havanas on trips to Boston. For the good of Roman’s health and my standing with the US customs officers there is now prescription on these infractions. In 1993, I was appointed Director of the CRM (for the first time). As a central part of its activities are thematic programs. These concentration periods on topics of special interest bring specialists from all over the world around a number of workshops and conferences; these are planned with significant leadtime. The Aisenstadt Chair is CRM’s most prestigious lectureship: it is offered to distinguished scientists upon the recommendation of the CRM International Advisory Committee. The holders of the Chair deliver series of talks that are integrated within the thematic program of the semester or of the year; they are also strongly encouraged to turn their lectures into a book to be published in one of the CRM monograph series. At the beginning of my second term as Director around 1997, together with Philippe Di Francesco, Lisa Jeffrey, André Leclair and Yvan Saint-Aubin we started putting together a theme year in mathematical physics. Little did I know that I would be appointed Provost of McGill University at the beginning of July 1999. Even though I could not enjoy as much as I had intended the deployment of the scientific events that the year entailed, the program was a resounding success with Roman holding the Aisenstadt Chair. He gave his lectures in the framework of two workshops: the first on Strings, Duality and Geometry and the second in Condensed Matter and Non-Equilibrium Physics. The general topic he chose was Fluid mechanics and as you may imagine Roman offered an original and fascinating view of this broad subject from the perspective of a particle theorist [@J9]. Being Provost at McGill led to my becoming Rector (or President) of the Université de Montréal and so I went around the Mont Royal returning to the institution where I had begun my academic career. Time was at a premium in those years but as mentioned above the Jackiws and the Vinets kept meeting and I managed to maintain some research activity. I was determined not to end my term as Rector before ensuring that Roman receives recognition from us for his immense scientific accomplishments and his special relation with Montreal. This happened in 2010 when I had the great pleasure to present him with a Honorary Doctorate from the Université de Montréal. This was a touching celebration that took place within the solemn Ph.D. convocation in a packed amphitheater. Roman with his usual intellectual elegance gave an inspiring acceptance speech and generated a long and enthusiatic applause. I recall that as he was leaving the podium looking quite moved he said to me: Now I know how my son feels! Envoi ===== Dear Roman: I see you like a great artist. Writing this short and sketchy chronicle of our encounters over the years gave me the chance to reflect further on your work and its tremendous impact. Your papers and expository texts are like magnificent paintings that reveal subtle and unexpected perspectives. These paintings were much acclaimed when they were first presented and brought you fame. Apprentices came to learn from you and many emulated you from the distance. You generously shared your knowledge and craftmanship. If you had such an influence on me, we can imagine the number of people for whom this has been the case. And then new generations rediscover your work, look at it from different angles, apply it to different phenomena, other artists get inspired by it and create new movements, avatars of your past creations. You engage with that and produce new work making the wheel turn. Need it be said that we look forward to more of these inspiring pieces and to your advice and views moving ahead. Thank you for the past and thank you for the future. Here is to you from the mind and from the heart. #### Acknowledgements I wish to acknowledge that much of my connections with Roman for which I am so grateful and of which I have given only a poor glimpse here, have been made possible by the Natural Science and Engineering Research Council (NSERC) of Canada. Over the years NSERC has offered me personally postgraduate scholarships, a postdoctoral fellowship and various discovery grants for which Roman relentlessly wrote recommendations. NSERC has also continuously supported this marvellous institute for research in the mathematical sciences that the Centre de Recherches Mathématiques or CRM is and that I still have the privilege to lead; this has in particular allowed our community to interact so profitably with Roman Jackiw at various occasions. To NSERC and through this Council to the Canadian taxpayers who are making scientific journeys possible: thank you. [30]{} \[1\][\#1]{} \[1\][`#1`]{} urlstyle \[1\][doi: \#1]{} S. [Weinberg]{}, [A Model of Leptons]{}, *Physical Review Letters*, [**[19]{}**]{},[1264-1266 (1967)]{}. Abdus  [Salam]{}, [Weak and electromagnetic interactions]{}, in *Elementary Particle Theory (Nobel Symposium 8)*, edited by N. Svartholm, (Almqvist & Wiksell, 1968). G. ’t [Hooft]{}, [Renormalizable Lagrangians for massive Yang-Mills fields]{}, *Nuclear Physics B*, [**[35]{}**]{},[167-188 (1971)]{}. D. [Gross]{}, F. [Wilczek]{}, [Ultraviolet behavior of non-abelian gauge theories]{}, *Physical Review Letters*, [**[30]{}**]{},[1343-1346 (1973)]{}. H. [Politzer]{}, [Reliable perturbative results for strong interactions]{}, *Physical Review Letters*, [**[30]{}**]{},[1346-1349 (1973)]{}. S. [Adler]{}, [Axial-Vector Vertex in Spinor Electrodynamics]{}, *Physical Review*, [**[177]{}**]{},[2426-2438 (1969)]{}. J. [Bell]{}, R. [Jackiw]{}, [A PCAC puzzle: $\eta^0\rightarrow\gamma\gamma$ in the $\sigma$-model]{}, *Il Nuovo Cimento A*, [**[60]{}**]{},[47-61 (1969)]{}. R. [Jackiw]{}, [Field theoretic investigations in current algebra]{}, in *Lectures on Current Algebra and its Applications*, by S. B. Treiman, R. Jackiw and D. J. Gross (Princeton University Press, 1972). P. [Goddard]{}, D. [Olive]{}, [Magnetic monopoles in gauge field theories]{}, *Reports on Progress in Physics*, [**[41]{}**]{},[1357-1437 (1978)]{}. S. [Coleman]{}, [Aspects of Symmetry: Selected Erice Lectures of Sidney Coleman]{} (Cambridge University Press, 1985). R. [Jackiw]{}, C. [Nohl]{}, C. [Rebbi]{}, [Classical and semi-classical solutions of the Yang-Mills theory]{} in, *Particles and Fields*, edited by D. H. Boal and A. N. Kamal (Plenum Press, 1978). R. [Jackiw]{}, C. [Rebbi]{}, [Conformal properties of a Yang-Mills pseudoparticle]{}, *Physical Review*, [**[D14]{}**]{}, [517-523 (1976)]{}. E. [Witten]{}, [Some Exact Multipseudoparticle Solutions of Classical Yang-Mills Theory]{}, *Physical Review Letters*, [**[38]{}**]{}, [121-124 (1977)]{}. T. T. [Wu]{}, C. N. [Yang]{}, [Concept of nonintegrable phase factors and global formulation of gauge fields]{}, *Physical Review*, [**[D12]{}**]{}, [3845-3857 (1975)]{}. J. [Harnad]{}, S.D. [Shnider]{}, L. [Vinet]{}, [Group Actions on Principal Bundle and Invariance Conditions for Gauge Fields]{}, *Journal of Mathematical Physics*, [**[21]{}**]{}, [2719-2733, (1980\]]{} J.P. [Harnad]{}, S.D. [Shnider]{}, L.  [Vinet]{}, [Solutions to Yang-Mills Equations on $M^4$ Invariant Under Subgroups of $O(4,2)$]{} in *Complex Manifold Techniques in Theoretical Physics*, edited by D.E. Lerner, P.D. Sommers, (Pitman Press,(1979)). R.  [Jackiw]{}, [Non-linear equations in particle physics]{} in *Complex Manifold Techniques in Theoretical Physics*, edited by D.E. Lerner, P.D. Sommers, (Pitman Press,(1979)). P. [Forgacs]{}, N. S. [Manton]{}, [Space-time symmetries in gauge theories]{}, *Communications in Mathematical Physics*, [**[72]{}**]{}, [15-35 (1980)]{}. R. [Jackiw]{}, N. S. [Manton]{}, [Symmetries and conservation laws in gauge theories]{}, *Annals of Physics*, [**[127]{}**]{}, [257-273 (1980)]{}. A. H. [Guth]{}, [Inflationary universe: A possible solution to the horizon and flatness problems]{}, *Physical Review*, [**[D23]{}**]{}, [347-356 (1981)]{}. A. H. [Guth]{}, S-.Y. [Pi]{}, [Fluctuations in the New Inflationary Universe]{}, *Physical Review Letters*, [**[49]{}**]{}, [1110-1113 (1982)]{}. S-.Y. [Pi]{}, [Inflation without Tears: A Realistic Cosmological Model]{}, *Physical Review Letters*, [**[52]{}**]{}, [1725-1728 (1984)]{}. E. [Farhi]{}, R. [Jackiw]{}, [Dynamical Gauge Symmetry Breaking]{}, [World Scientific(1982)]{}. V. [de Alfaro]{}, S. [Fubini]{}, G. [Furlan]{}, [Conformal invariance in quantum mechanics]{}, *Nuovo Cimento A*, [**[34]{}**]{}, [569-612 (1976)]{}. R. [Jackiw]{}, [Dynamical symmetry of the magnetic monopole]{}, *Annals of Physics*, [**[129]{}**]{}, [183-200 (1980)]{}. S. [Fubini]{}, E. [Rabinovici]{}, [Superconformal quantum mechanics]{}, *Nuclear Physics*, [**[B245]{}**]{}, [17-44 (1984)]{}. E. [Witten]{}, [Constraints on supersymmetry breaking]{}, *Nuclear Physics*, [**[B202]{}**]{}, [253-316 (1982)]{}. E. [D’Hoker]{}, L. [Vinet]{}, [Supersymmetry of the Pauli equation in the presence of a magnetic monopole]{}, *Physics Letters*, [**[B137]{}**]{}, [72-76 (1984)]{}. E. [D’Hoker]{}, L. [Vinet]{}, [Classical solutions to topologically massive Yang-Mills theory]{}, *Annals of Physics*, [**[162]{}**]{}, [413-440 (1985)]{}. S. [Deser]{}, R. [Jackiw]{}, S. [Templeton]{}, [Topologically massive gauge theories]{}, *Annals of Physics*, [**[140]{}**]{}, [372-411 (1982)]{}. O J. P.. [Eboli]{}, R. [Jackiw]{}, S.Y. [Pi]{}, [Quantum Fields Out of Thermal Equilibrium]{}, *Physical Review*, [**[D37]{}**]{}, [3557-3581 (1988)]{}. R. [Floreanini]{}, C. T. [Hill]{}, R. [Jackiw]{}, [ Functional Representation for the lsometries of de Sitter Space]{}, *Annals of Physics*, [**[175]{}**]{}, [345-365 (1987)]{}. R. [Floreanini]{}, L. [Vinet]{}, [Vacuum State of Complex Scalar Fields in SO(2,1) Invariant Background]{}, *Physical Review*, [**[D36]{}**]{}, [1731-1739 (1987)]{}. R. [Jackiw]{}, [Berry’s Phase: Topological Ideas From Atomic, Molecular And Optical Physics]{}, *Comments on Atomic and Molecular Physics*, [**[21]{}**]{}, [71-82 (1988)]{}. L. [Vinet]{}, [Invariant Berry connections]{}, *Physical Review*, [**[D37]{}**]{}, [2369-2372(R) (1988)]{}. R. [Jackiw]{}, [Three elaborations on Berry’s connection, curvature and phase]{}, *International Journal of Modern Physics A*, [**[3]{}**]{}, [285-297 (1988)]{}. R. [Jackiw]{}, [Topics in Planar Physics]{}, in *Physics, Geometry and Topology*, [191-239]{} edited by H. C. Lee, (NATO ASI Series, vol 238, Springer, 1990). R. [Jackiw]{}, [Field Theoretic Results in the Schrödinger Representation]{}, in *Proceedings of the XVIIth International Colloquium on Group Theoretical Methods in Physics*, [90-92]{} edited by Y. Saint-Aubin and L. Vinet, (World Scientific, 1989). S.-Y. [Pi]{}, [Inflation and Non-Equilibrium Dynamics]{}, in *Proceedings of the XVIIth International Colloquium on Group Theoretical Methods in Physics*, [127-136]{} edited by Y. Saint-Aubin and L. Vinet, (World Scientific, 1989). R. [Jackiw]{}, [Quantum gravity in flatland]{}, in *Proceedings of the XVIIth International Colloquium on Group Theoretical Methods in Physics*, [331–336]{} edited by Y. Saint-Aubin and L. Vinet, (World Scientific, 1989). R. [Jackiw]{}, [Lectures on Fluid Dynamics - a Particle Theorist’s View of Supersymmetric, Non-Abelian, Noncommutative Fluid Mechanics and d-Branes]{}, *CRM Series in Mathematical Physics*, (Springer 2002)

ArXiv

--- author: - | Pedro Moreno-Sanchez, Navin Modi, Raghuvir Songhela, Aniket Kate, Sonia Fahmy\ Purdue University\ email: @purdue.edu - 'Pedro Moreno-Sanchez, Navin Modi, Raghuvir Songhela, Aniket Kate, Sonia Fahmy' bibliography: - 'main.bib' date: 'Version: ' title: - | Mind Your Credit: Assessing the Health of\ the Ripple Credit Network[^1] - | Mind Your Credit: Assessing the Health of\ the Ripple Credit Network --- [^1]: This work appears at WWW 2018. The project’s website is https://pedrorechez.github.io/Ripple-Credit-Study/

ArXiv

--- author: - Francis OGER title: '**Equivalence élémentaire entre pavages**' --- **TILINGS AND ASSOCIATED RELATIONAL STRUCTURES** Francis OGER **Abstract.** In the present paper, as we did previously in \[7\], we investigate the relations between the geometric properties of tilings and the algebraic properties of associated relational structures. Our study is motivated by the existence of aperiodic tiling systems. In \[7\], we considered tilings of the euclidean spaces $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{k}$, and isomorphism was defined up to translation. Here, we consider, more generally, tilings of a metric space, and isomorphism is defined modulo an arbitrary group of isometries. In Section 1, we define the relational structures associated to tilings. The results of Section 2 concern local isomorphism, the extraction preorder and the characterization of relational structures which can be represented by tilings of some given type. In Section 3, we show that the notions of periodicity and invariance through a translation, defined for tilings of the euclidean spaces $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{k}$, can be generalized, with appropriate hypotheses, to relational structures, and in particular to tilings of noneuclidean spaces. The results of Section 4 are valid for uniformly locally finite relational structures, and in particular tilings, which satisfy the local isomorphism property. We characterize among such structures those which are locally isomorphic to structures without nontrivial automorphism. We show that, in an euclidean space $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{k}$, this property is true for a tiling which satisfies the local isomorphism property if and only if it is not invariant through a nontrivial translation. We illustrate these results with examples, some of them concerning aperiodic tiling systems of euclidean or noneuclidean spaces. MSC: 52C23 (05B45, 52C22) Keywords: Aperiodic tiling systems; Locally isomorphic tilings; Local isomorphism property; Metric Space; Isometry; Noneuclidean; Hyperbolic plane; Periodic; Translation; Rigid. In Section 1, we consider finite systems consisting of prototiles in a metric space $E$ and local rules for assembling tiles which are equivalent to the prototiles modulo a group $G$ of isometries of $E$. With appropriate hypotheses, we associate to each such system $\Delta $ a finite relational language $\mathcal{L}_{\Delta }$ such that any $\Delta $-tiling can be interpreted as a $\mathcal{L}_{\Delta }$-structure. In \[7\], we did the same thing in a particular case: $E$ was an euclidean space $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{k}$ and $G$ consisted of the translations of $E$. In that case, $\mathcal{% L}_{\Delta }$ can be defined by considering all the possible configurations of two adjacent tiles. In the present situation, we have to consider the possible configurations for the patch consisting of all tiles within some given distance of one of them; that distance is depending on $E$, $G$ and $% \Delta $. In the following sections, we prove various properties of $\Delta $-tilings, some of them classical and others new, concerning local isomorphism, invariance through a translation... Classically, such results are shown by considering the geometrical and topological properties of the space and the tilings. The proofs given here are obtained by considering the algebraic properties of the associated $\mathcal{L}_{\Delta }$-structures. In that way, we prove the results for larger classes of tilings, and we show similar properties for relational structures which are not represented by tilings (see Section 4, Example 4). It is also natural to wonder, for a given system $\Delta $, which $\mathcal{L% }_{\Delta }$-structures can be represented by $\Delta $-tilings. Characterizations of such structures are given in Section 2. **1. Tilings and associated relational structures**. In \[7\], we considered relational structures associated to tilings of the euclidean spaces $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{k}$. The isomorphism of tiles, patches, tilings... was defined up to translation. Actually, much more various situations have been investigated. For instance, the isomorphism of tiles, patches, tilings... can be defined modulo a group of isometries of $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{k}$ which contains symmetries and/or rotations. Also, tilings of noneuclidean spaces and tilings with partially overlapping tiles are considered. The following facts appear to be common to all cases: - any tiling is a covering of a metric space $(E,\delta )$ by tiles which are obtained from a finite set of closed bounded prototiles by applying isometries belonging to some specified group; - up to isometries in the group, only finitely many configurations of some given size can appear in a tiling; - the work of a tiler can be described as follows: first he puts a small cluster of tiles somewhere in the space, then he gradually increases the patch by adding new tiles; at each step, he applies the same finite set of local rules, which only leave finitely many possibilities for adding new tiles. In order to formalize these facts, we introduce some definitions and notations. We consider a metric space $(E,\delta )$ and a group $G$ of bijective isometries of $E$. We do not suppose that $(E,\delta )$ is homogeneous, or that local isometries of $E$ can be extended to global ones. We call a *tile* any closed bounded subset $T$ of $E$. We do not suppose $T$ connected or $T$ equal to the closure of its interior. For any tiles $T,T^{\prime }$ (resp. any sets of tiles $\mathcal{E},\mathcal{E}% ^{\prime }$), we say that $T$ and $T^{\prime }$ (resp. $\mathcal{E}$ and $% \mathcal{E}^{\prime }$) are *isomorphic* if there exists $\sigma \in G$ such that $T\sigma =T^{\prime }$ (resp. $\mathcal{E}\sigma =\mathcal{E}% ^{\prime }$). We define isomorphism in the same way for the pairs $(S,T)$ and the pairs $(\mathcal{E},T)$ with $S,T$ tiles and $\mathcal{E}$ a set of tiles. **Remark.** Sometimes, tiles with drawings are also considered. A *drawing* is a map from a tile $T$ to a finite set $% \Omega $, whose elements are called *colours*. In that case, the homomorphisms that we consider must respect colours; moreover, in the definitions of configuration, tiling and patch given below, any point in the intersection of two tiles must have the same colour in each of them. The results of the present paper are proved in the same way for tiles with drawings. For each set $\mathcal{E}$ of tiles and each $T\in \mathcal{E}$, we define inductively the subsets $\mathcal{B}_{n}^{\mathcal{E}}(T)$ with $\mathcal{B}% _{0}^{\mathcal{E}}(T)=\{T\}$ and, for each $n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $, $\mathcal{B}_{n+1}^{\mathcal{E}}(T)=\{U\in \mathcal{E\mid }$ there exists $V\in \mathcal{B}_{n}^{\mathcal{E}}(T)$ such that $U\cap V\neq \emptyset \}$. For any tiles $T,T^{\prime }$ and any sets of tiles $\mathcal{E},% \mathcal{E}^{\prime }$, we write $(\mathcal{E},T)\leq (\mathcal{E}^{\prime },T^{\prime })$ if there exists $\sigma \in G$ such that $\mathcal{E}\sigma \subset \mathcal{E}^{\prime }$ and $T\sigma =T^{\prime }$. We call a *tiling* any covering $\mathcal{E}$ of $E$ by possibly overlapping tiles such that, for each $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$: 1\) for each $T\in \mathcal{E}$, $\mathcal{B}_{r}^{\mathcal{E}}(T)$ is finite and $\mathcal{B}_{r-1}^{\mathcal{E}}(T)$ is contained in the interior of the union of the tiles of $\mathcal{B}_{r}^{\mathcal{E}}(T)$; 2\) for any $S,T\in \mathcal{E}$, if $(\mathcal{B}_{r}^{\mathcal{E}% }(S),S)\leq (\mathcal{B}_{r}^{\mathcal{E}}(T),T)$, then $(\mathcal{B}_{r}^{% \mathcal{E}}(S),S)\cong (\mathcal{B}_{r}^{\mathcal{E}}(T),T)$; 3\) the pairs $(\mathcal{B}_{r}^{\mathcal{E}}(T),T)$ for $T\in \mathcal{E}$ fall in finitely many isomorphism classes. If 1) is true for $r=1$, then it is true for each $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$. If 1) is true and if 2) is true for $r=1$, then, for any $S,T\in \mathcal{E}$, each $\sigma \in G$ such that $\mathcal{B}_{1}^{\mathcal{E}% }(S)\sigma \subset \mathcal{B}_{1}^{\mathcal{E}}(T)$ and $S\sigma =T$ is an isomorphism from $(\mathcal{B}_{1}^{\mathcal{E}}(S),S)$ to $(\mathcal{B}% _{1}^{\mathcal{E}}(T),T)$. By induction, it follows that, for each $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$ and any $S,T\in \mathcal{E}$, each $\sigma \in G$ such that $% \mathcal{B}_{r}^{\mathcal{E}}(S)\sigma \subset \mathcal{B}_{r}^{\mathcal{E}% }(T)$ and $S\sigma =T$ is an isomorphism from $(\mathcal{B}_{r}^{\mathcal{E}% }(S),S)$ to $(\mathcal{B}_{r}^{\mathcal{E}}(T),T)$. Consequently, 2) is true for each $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$. We note that 2) is necessarily true if 1) is true and if the tiles of $\mathcal{E}$ are equal to the closure of their interior and nonoverlapping. It is natural to wonder whether 3) is also true for each $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$ if it is true for $r=1$. If $(E,\delta )$ is an euclidean space $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{k}$ and if $G$ consists of the translations of $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{k}$, then any pair $(S,T)\in \mathcal{E}\times \mathcal{E}$ is completely determined by $S$ and the isomorphism class of $(S,T)$. Consequently, if the pairs $(S,T)\in \mathcal{E}\times \mathcal{E}$ with $S\cap T\neq \emptyset $ fall in finitely many isomorphism classes, then, for each $n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$, the pairs $(\mathcal{B}_{n}^{\mathcal{E}}(T),T)$ for $T\in \mathcal{E}$ fall in finitely many isomorphism classes. The following example shows that the situation is different if $G$ consists of the isometries of $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{k}$: **Example 1.** Let $(E,\delta )$ be the euclidean space $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{2}$ and let $G$ consist of the isometries of $(E,\delta )$. Consider the coverings of $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{2}$ by nonoverlapping tiles isomorphic to the following prototiles: $T_{0}=\{(x,y)\in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{2}\mid x\geq 0$ and $x^{2}+y^{2}\leq 1\}$, $T_{k}=\{(x,y)\in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{2}\mid k^{2}\leq x^{2}+y^{2}\leq (k+1)^{2}\}$ for $1\leq k\leq 2n+1$, $T_{2n+2}=\{(x,y)\in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{2}\mid x^{2}+y^{2}\geq (2n+2)^{2}$ and $\sup (\left\vert x\right\vert ,\left\vert y\right\vert )\leq 2n+3\}$, with any two copies of $T_{2n+2}$ having one edge in common if they have more than one common point. Each such covering consists of squares with sides of length $4n+6$, each of them covered by two copies of $T_{0}$ and one copy of each of the tiles $% T_{1},...,T_{2n+2}$. Moreover, in each square, the two copies of $T_{0}$ can be rotated together arbitrarily. Consequently, in such a covering $\mathcal{E% }$, the pairs $(\mathcal{B}_{n+1}^{\mathcal{E}}(T),T)$ for $T\in \mathcal{E}$ do not generally fall in finitely many isomorphism classes, contrary to the pairs $(\mathcal{B}_{n}^{\mathcal{E}}(T),T)$. Now, we can ask the following question: Suppose that we specify a finite set of isomorphism classes for the pairs $(\mathcal{B}_{1}^{\mathcal{E}}(T),T)$ which can appear in a covering $\mathcal{E}$ of $E$ which satisfies 1) and 2). Is there an integer $m$ such that, if the pairs $(\mathcal{B}_{m}^{% \mathcal{E}}(T),T)$ for $T\in \mathcal{E}$ fall in finitely many isomorphism classes, then, for each $n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$, the pairs $(\mathcal{B}_{n}^{\mathcal{E}}(T),T)$ for $T\in \mathcal{E}$ fall in finitely many isomorphism classes. Example 2 below shows that this property does not necessarily hold. On the other hand, we are going to prove (see Proposition 1.10) that some natural conditions on $(E,\delta )$ and $G$, satisfied by the euclidean spaces $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{k}$, and on the pairs $(\mathcal{B}_{1}^{\mathcal{E}}(T),T)$ for $T\in \mathcal{E}$, are enough to make it true. **Example 2.** Let $E=\{(x,y,z)\in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{3}\mid x^{2}+y^{2}=1\}$ be the surface of a cylinder of infinite length. Define the distance $\delta $ by considering $E$ as a quotient of the euclidean space $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{2}$. Let $G$ consist of the isometries of $E$. Write $T_{1}=\{(x,y,z)\in E\mid x\geq 0$ and $0\leq z\leq 1\}$ and $T_{2}=\{(x,y,z)\in E\mid 0\leq z\leq 1\}$. For each $n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$, consider the coverings $\mathcal{E}$ of $E$ by nonoverlapping tiles obtained as follows: for each $a\in %TCIMACRO{\U{2124} }% %BeginExpansion \mathbb{Z} %EndExpansion $, $\{(x,y,z)\in E\mid a\leq z\leq a+1\}$ is covered by two tiles isomorphic to $T_{1}$ if $a\in (2n+2)% %TCIMACRO{\U{2124} }% %BeginExpansion \mathbb{Z} %EndExpansion $ and by one tile isomorphic to $T_{2}$ otherwise. For each $a\in (2n+2)% %TCIMACRO{\U{2124} }% %BeginExpansion \mathbb{Z} %EndExpansion $, the two tiles covering $\{(x,y,z)\in E\mid a\leq z\leq a+1\}$ can be rotated together arbitrarily. Consequently, in such a covering $\mathcal{E}$, the pairs $(\mathcal{B}_{n+1}^{\mathcal{E}}(T),T)$ for $T\in \mathcal{E}$ do not generally fall in finitely many isomorphism classes, contrary to the pairs $(\mathcal{B}_{n}^{\mathcal{E}}(T),T)$. Moreover, the pairs $(\mathcal{% B}_{1}^{\mathcal{E}}(T),T)$ for all possible choices of $\mathcal{E}$ and $T$ fall in three isomorphism classes. Now we state the definitions, the notations and the hypotheses which are necessary for our results. For each $x\in E$ and each $\eta \in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{\geq 0}$, we write $\beta (x,\eta )=\{y\in E\mid \delta (x,y)\leq \eta \}$. For each nonempty bounded $S\subset E$, we consider the $\emph{radius}$ $% \mathrm{Rad}(S)=\inf_{x\in E}(\sup_{y\in S}\delta (x,y))$ and the *diameter* $\mathrm{Diam}(S)=\sup_{x,y\in S}\delta (x,y)$; we have $\mathrm{% Diam}(S)\leq 2.\mathrm{Rad}(S)$. From now on, we suppose that $E$ is connected and that $\beta (x,\eta )$ is compact for each $x\in E$ and each $\eta \in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{>0}$. Then each bounded closed subset (i.e. tile) of $E$ is compact. For each $x\in E$ and each $\xi \in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{>0}$, we define inductively $\beta (x,\xi ,n)$ with $\beta (x,\xi ,0)=\left\{ x\right\} $ and, for each $n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $, $\beta (x,\xi ,n+1)=\cup _{y\in \beta (x,\xi ,n)}\beta (y,\xi )$. We have $\cup _{n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }\beta (x,\xi ,n)=E$ since $E$ is connected and $\cup _{n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }\beta (x,\xi ,n)$ is both open and closed. For each $n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $, we write $\omega (x,\xi ,n)=\sup \{\eta \in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{\geq 0}\mid \beta (x,\eta )\subset \beta (x,\xi ,n)\}$. Now we show that $\omega (x,\xi ,n)$ tends to infinity with $n$. Otherwise, there exists $\eta \in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{>0}$ such that $\beta (x,\eta )\nsubseteqq \beta (x,\xi ,n)$ for each $% n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $. It follows that $\beta (x,\eta )$ contains a sequence $(y_{n})_{n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }$ with $y_{n}\notin \beta (x,\xi ,n)$ for each $n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $. As $\beta (x,\eta )$ is compact, there exists a subsequence $% (z_{n})_{n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }$ which converges to a point $z\in \beta (x,\eta )$. For $n$ large enough, we have $\delta (z,z_{n})<\xi $, which implies that $z\notin \beta (x,\xi ,n-1)$ since $z_{n}\notin \beta (x,\xi ,n)$. It follows that $z\notin \cup _{n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }\beta (x,\xi ,n)$, contrary to what we proved just above. We say that $(E,\delta )$ is: - *weakly homogeneous* if $\omega (\xi ,n)=\inf_{x\in E}\omega (x,\xi ,n)$ tends to infinity with $n$; - *geodesic* if any two points of $E$ can be joined by at least one geodesic; -* transitive* if, for any $x,y\in E$, there exists an isometry $\sigma $ such that $x\sigma =y$. Any geodesic space is weakly homogeneous since $\omega (x,\xi ,n)=n\xi $ for any $x,\xi ,n$. Any transitive space is also weakly homogeneous since $% \omega (x,\xi ,n)$ does not depend on $x$. In some of our results, we shall suppose that $(E,\delta )$ is weakly homogeneous or geodesic. For each *finite* set $\mathcal{E}$ of tiles, each $T\in \mathcal{E}$ and each $p\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$, we say that $(\mathcal{E},T)$ is a $p$*-configuration* if $% \mathcal{B}_{p}^{\mathcal{E}}(T)=\mathcal{E}$ and if each $S\in \mathcal{B}% _{p-1}^{\mathcal{E}}(T)$ is contained in the interior of the union of the tiles of $\mathcal{B}_{1}^{\mathcal{E}}(S)$. Then $\mathcal{B}_{p-1}^{% \mathcal{E}}(T)$ is contained in the interior of the union of the tiles of $% \mathcal{E}$ and $(\mathcal{B}_{k}^{\mathcal{E}}(S),S)$ is a $k$-configuration for each $k\in \{1,...,p\}$ and each $S\in \mathcal{B}_{p-k}^{% \mathcal{E}}(T)$. **Lemma 1.1.** For each $1$-configuration $(\mathcal{E},T)$, there exists $\xi \in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{>0}$ such that $\cup _{x\in T}\beta (x,\xi )$ is contained in the union of the tiles of $\mathcal{E}$. **Proof.** For each $n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$, consider $x_{n}\in T$ such that $\beta (x_{n},1/n)$ is not contained in the union of the tiles of $\mathcal{E}$. As $T$ is compact, there exists a subsequence of $(x_{n})_{n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }$ which converges to a point $x\in T$, and $x$ does not belong to the interior of the union of the tiles of $\mathcal{E}$, whence a contradiction.  $\blacksquare $ **Proposition 1.2.** Let $\mathcal{E}$ be a nonempty set of tiles such that the pairs $(\mathcal{B}_{1}^{\mathcal{E}}(T),T)$ for $T\in \mathcal{E}$ are $1$-configurations and fall in finitely many isomorphism classes. Then 1\) $\mathcal{E}$ covers $E$ and $\{S\in \mathcal{E}\mid S\cap \beta (x,\eta )\neq \varnothing \}$ is finite for each $x\in E$ and each $% \eta \in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{>0}$; 2\) If $(E,\delta )$ is weakly homogeneous, then, for each $\eta \in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{>0}$, there exists $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$ such that $\delta (S,T)\leq \eta $ implies $T\in \mathcal{B}_{r}^{% \mathcal{E}}(S)$ for any $S,T\in \mathcal{E}$. **Proof.** By Lemma 1.1, there exists $\xi \in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{>0}$ such that $\mathcal{B}_{1}^{\mathcal{E}}(T)$ covers $\beta (x,\xi )$ for each $T\in \mathcal{E}$ and each $x\in T$. Consequently, the union of the tiles of $\mathcal{E}$ is both open and closed, and $\mathcal{E}$covers $E$ since $E$ is connected. Now suppose that there exist $x\in E$ and $\eta \in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{>0}$ such that $\{S\in \mathcal{E}\mid S\cap \beta (x,\eta )\neq \varnothing \}$ is infinite. Consider a sequence $(y_{n})_{n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }\subset \beta (x,\eta )$ and a sequence of distinct tiles $(T_{n})_{n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }\subset \mathcal{E}$ such that $y_{n}\in T_{n}$ for each $n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $. As $\beta (x,\eta )$ is compact, there exists a subsequence of $% (y_{n})_{n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }$ which converges to a point $y\in \beta (x,\eta )$. For each $T\in \mathcal{E}$ such that $y\in T$, $\mathcal{B}_{2}^{\mathcal{E}}(T)$contains infinitely many $T_{n}$ since $\mathcal{B}_{1}^{\mathcal{E}}(T)$covers $\beta (x,\xi )$, whence a contradiction. The second part of Proposition 1.2 is true because, for each $n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $, each $S\in \mathcal{E}$ and each $x\in S$, $\mathcal{B}_{n}^{\mathcal{E}% }(S)$ covers $\beta (x,\omega (\xi ,n))$ and $\mathcal{B}_{n+1}^{\mathcal{E% }}(S)$ contains $\{T\in \mathcal{E}\mid T\cap \beta (x,\omega (\xi ,n))\neq \varnothing \}$.  $\blacksquare $ For each $p\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$, we call a $p$*-local rule* any set $\Gamma =\{(\mathcal{C}% _{1},C_{1}),...,(\mathcal{C}_{m},C_{m})\}$ of pairwise nonisomorphic $p$-configurations such that: 1\) for any $i,j\in \{1,...,m\}$, if $(\mathcal{B}_{1}^{\mathcal{C}% _{i}}(C_{i}),C_{i})\leq (\mathcal{B}_{1}^{\mathcal{C}_{j}}(C_{j}),C_{j})$, then $(\mathcal{B}_{1}^{\mathcal{C}_{i}}(C_{i}),C_{i})\cong (\mathcal{B}% _{1}^{\mathcal{C}_{j}}(C_{j}),C_{j})$; 2\) for each $i\in \{1,...,m\}$ and each $S\in \mathcal{B}_{1}^{% \mathcal{C}_{i}}(C_{i})$, there exists $j\in \{1,...,m\}$ such that $(% \mathcal{B}_{p-1}^{\mathcal{C}_{i}}(S),S)\cong (\mathcal{B}_{p-1}^{\mathcal{C% }_{j}}(C_{j}),C_{j})$. If $\Gamma $ is a $p$-local rule, then, for each $q\in \{1,...,p-1\}$, any representatives of the isomorphism classes of $(\mathcal{% B}_{q}^{\mathcal{C}_{1}}(C_{1}),C_{1}),...,(\mathcal{B}_{q}^{\mathcal{C}% _{m}}(C_{m}),C_{m})$ form a $q$-local rule. We say that a set $\mathcal{E}$ of tiles $\emph{satisfies}$ $\Gamma $ if, for each $T\in \mathcal{E}$, the pair $(\mathcal{B}_{p}^{\mathcal{E}}(T),T)$ is isomorphic to one of the pairs $(\mathcal{C}_{i},C_{i})$. Each tiling satisfies a $p$-local rule for each $p\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$. We say that a set $\mathcal{E}$ of tiles is a *patch* if there exist no $\mathcal{F},\mathcal{G}\subset \mathcal{E}$ such that $\mathcal{F}\cup \mathcal{G}=\mathcal{E}$ and $S\cap T=\varnothing $ for each $S\in \mathcal{F}$ and each $T\in \mathcal{G}$. This property is true if and only if, for any $S,T\in \mathcal{E}$, there exists $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $ such that $T\in \mathcal{B}_{r}^{\mathcal{E}}(S)$. For each finite set $\mathcal{E}$ of tiles and each connected set $A\subset E$, if $A$ is contained in the union of the tiles of $\mathcal{E}$, and if each tile of $\mathcal{E}$ contains a point of $A$, then $\mathcal{E}$ is a patch. The union of the tiles of a patch is not necessarily connected if the tiles themselves are not connected. For each set $\mathcal{E}$ of tiles, we write $\mathrm{I}_{\Gamma }(\mathcal{% E})=\cup _{1\leq i\leq m}\{T\in \mathcal{E}\mid (\mathcal{B}_{p}^{\mathcal{E}% }(T),T)\cong (\mathcal{C}_{i},C_{i})\}$. We say that $\mathcal{E}$ is a $% \Gamma $*-patch* if: 1\) for each $T\in \mathcal{E}$, there exists $i\in \{1,...,m\}$ such that $(\mathcal{B}_{p}^{\mathcal{E}}(T),T)\leq (\mathcal{C}_{i},C_{i})$. 2) $\mathrm{I}_{\Gamma }(\mathcal{E})$ contains a patch $% \mathcal{A}$ such that $\mathcal{E}=\cup _{T\in \mathcal{A}}\mathcal{B}_{p}^{% \mathcal{E}}(T)$. For each set $\mathcal{E}$ of tiles which satisfies $\Gamma $, each $T\in \mathcal{E}$ and each integer $k\geq p$, $\mathcal{B}_{k}^{\mathcal{E}}(T)$ is a $\Gamma $-patch since $\mathcal{B}_{k-p}^{\mathcal{E}}(T)$ is connected, $\mathcal{B}_{k-p}^{\mathcal{E}}(T)\subset \mathrm{I}_{\Gamma }(% \mathcal{B}_{k}^{\mathcal{E}}(T))$ and $\mathcal{B}_{k}^{\mathcal{E}% }(T)=\cup _{S\in \mathcal{B}_{k-p}^{\mathcal{E}}(T)}\mathcal{B}_{p}^{% \mathcal{E}}(S)$. Now, for each $q\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$ and each $q$-local rule $\Delta =\{(\mathcal{D}_{1},D_{1}),...,(% \mathcal{D}_{n},D_{n})\}$ such that each $\mathcal{B}_{q-1}^{\mathcal{D}% _{i}}(D_{i})$ is fixed by no $\sigma \in G-\{\mathrm{Id}\}$, we define a finite relational language $\mathcal{L}_{\Delta }$ such that the $\Delta $-patches can be represented by $\mathcal{L}_{\Delta }$-structures. As a consequence, we show that any set of tiles which satisfies $\Delta $ is a tiling. Practically, in many examples of tilings, each $1$-configuration is fixed by no $\sigma \in G-\{\mathrm{Id}\}$, so that we can take $\Delta $ with $q=2$. For each $i\in \{1,...,n\}$, we write $\mathcal{D}_{i}=% \{D_{i,1},...,D_{i,p(i)}\}$ with $D_{i,1}=D_{i}$, and we introduce a $p(i)$-ary relational symbol $R_{i}(u_{i,1},...,u_{i,p(i)})$. We write $\mathcal{L}% _{\Delta }=\{R_{1},...,R_{n}\}$. For each set $\mathcal{E}$ of tiles, we define a $\mathcal{L}_{\Delta }$*-structure* on $\mathcal{E}$ as follows: for $1\leq i\leq n$ and $% T_{1},...,T_{p(i)}\in \mathcal{E}$, we write $R_{i}(T_{1},...,T_{p(i)})$ if $\mathcal{B}_{q}^{\mathcal{E}}(T_{1})=\{T_{1},...,T_{p(i)}\}$ and if there exists $\sigma \in G$ such that $D_{i,j}\sigma =T_{j}$ for $1\leq j\leq p(i) $. For $1\leq i,j\leq n$ and $S,S_{2},...,S_{p(i)},T_{2},...,T_{p(j)}% \in \mathcal{E}$, the relations $R_{i}(S,S_{2},...,S_{p(i)})$ and $% R_{j}(S,T_{2},...,T_{p(j)})$ imply $i=j$ and $\{S_{2},...,S_{p(i)}\}=% \{T_{2},...,T_{p(i)}\}$. Any set of tiles $\mathcal{E}$ satisfies $\Delta $ if and only if, for each $% T\in \mathcal{E}$, there exist $1\leq i\leq n$ and $T_{2},...,T_{p(i)}\in \mathcal{E}$ such that $R_{i}(T,T_{2},...,T_{p(i)})$. For any $\Delta $**-**patches $\mathcal{E}\subset \mathcal{F}$, the $\mathcal{L}_{\Delta }$-structure defined on $\mathcal{E}$ is the restriction to $\mathcal{E}$ of the $\mathcal{L}_{\Delta }$-structure defined on $\mathcal{F}$. The following theorem implies that any $\Delta $**-**patch, and in particular any set of tiles which satisfies $\Delta $, is determined up to isomorphism by the associated $\mathcal{L}_{\Delta }$-structure: **Theorem 1.3.** For any $\Delta $**-**patches $\mathcal{% E},\mathcal{F}$ and each $\mathcal{L}_{\Delta }$-homomorphism $f:\mathcal{E}% \rightarrow \mathcal{F}$, there exists a unique $\sigma \in G$ such that $% Sf=S\sigma $ for each $S\in \mathcal{E}$. **Proof.** Consider a patch $\mathcal{A}\subset \mathrm{I}% _{\Delta }(\mathcal{E})$ such that $\mathcal{E}=\cup _{T\in \mathcal{A}}% \mathcal{B}_{q}^{\mathcal{E}}(T)$. For each $T\in \mathcal{A}$, $f$ induces a bijection from $\mathcal{B}_{q}^{\mathcal{E}}(T)$ to $\mathcal{B}_{q}^{% \mathcal{F}}(Tf)$, and there exists a unique $\sigma _{T}\in G$ such that $% S\sigma _{T}=Sf$ for each $S\in \mathcal{B}_{q}^{\mathcal{E}}(T)$. It remains to be proved that $\sigma _{S}=\sigma _{T}$ for any $S,T\in \mathcal{A}$. As $\mathcal{A}$ is a patch, it suffices to show it for $% S\cap T\neq \emptyset $. Then we have $S\in \mathcal{B}_{1}^{\mathcal{E}}(T)$, and therefore $\mathcal{B}_{q-1}^{\mathcal{E}}(S)\subset \mathcal{B}_{q}^{% \mathcal{E}}(T)$. Consequently, for each $U\in \mathcal{B}_{q-1}^{\mathcal{E}% }(S)$, we have $U\sigma _{S}=Uf$ since $U\in \mathcal{B}_{q}^{\mathcal{E}% }(S) $, and $U\sigma _{T}=Uf$ since $U\in \mathcal{B}_{q}^{\mathcal{E}}(T)$. It follows $U\sigma _{S}=U\sigma _{T}$ for each $U\in \mathcal{B}_{q-1}^{% \mathcal{E}}(S)$, and therefore $\sigma _{S}=\sigma _{T}$.  $\blacksquare $ **Corollary 1.4.** For each integer $k$, there are finitely many isomorphism classes of $\Delta $**-**patches consisting of $k$ tiles. **Proof.** This result follows from Theorem 1.3 since there are finitely many isomorphism classes of $\mathcal{L}_{\Delta }$-structures consisting of $k$ elements.  $\blacksquare $ **Corollary 1.5.** For each finite $\Delta $**-**patch $% \mathcal{E}$ and each $k\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$, finitely many distinct $\Delta $**-**patches can be obtained from $\mathcal{E}$ by adding $k$ new tiles. **Proof.** Denote by $\Omega $ the set of all $\Delta $**-**patches obtained from $\mathcal{E}$ by adding $k$ new tiles. By corollary 1.4, it suffices to show that, for each $\mathcal{F}\in \Omega $, there exist finitely many pairs $(\mathcal{G},\sigma )$ with $\mathcal{G}\in \Omega $ and $\sigma :\mathcal{F}\rightarrow \mathcal{G}$ isomorphism. But any such pair is completely determined by the tiles $\sigma ^{-1}(T)$ for $% T\in \mathcal{E}$, since no $\rho \in G-\{\mathrm{Id}\}$ fixes the tiles of $% \mathcal{E}$.  $\blacksquare $ **Corollary 1.6.** For each $k\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $, $\{(\mathcal{B}_{k}^{\mathcal{E}}(T),T)\mid $ $\mathcal{E}$ satisfies $% \Delta $ and $T\in \mathcal{E}\}$ is a finite union of isomorphism classes. **Proof.** For $k\leq q$, this property is true because $% \mathcal{B}_{k}^{\mathcal{E}}(T)$ is contained in $\mathcal{B}_{q}^{% \mathcal{E}}(T)$ for each set $\mathcal{E}$ of tiles which satisfies $\Delta $ and each $T\in \mathcal{E}$. For $k\geq q+1$, it follows from Corollary 1.4 since we have a finite bound for the cardinals of the $\Delta $**-**patches $\mathcal{B}_{k}^{\mathcal{E}}(T)$ for $\mathcal{E}$ satisfying $% \Delta $ and $T\in \mathcal{E}$.  $\blacksquare $ **Corollary 1.7.** Any set of tiles which satisfies $\Delta $ is a tiling. **Proof.** Any such set $\mathcal{E}$ covers $E$ by the first part of Proposition 1.2. Moreover, for each $k\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $, $\{(\mathcal{B}_{k}^{\mathcal{E}}(T),T)\mid T\in \mathcal{E}\}$ is a finite union of isomorphism classes by Corollary 1.6.  $\blacksquare $ **Remark.** The following variant of Example 1 above shows that, in Theorem 1.3 and its corollaries, it is not enough to suppose that each $\mathcal{D}_{i}$ is fixed by no $\sigma \in G-\{\mathrm{Id}\}$. Consider the coverings $\mathcal{E}$ of $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{2}$ obtained from the following prototiles:  $T_{0}=\{(x,y)\in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{2}\mid x\geq 0$ and $x^{2}+y^{2}\leq 1\}$, $T_{k}=\{(x,y)\in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{2}\mid k^{2}\leq x^{2}+y^{2}\leq (k+1)^{2}\}$ for $1\leq k\leq 2q$, $T_{2q+1}=\{(x,y)\in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{2}\mid x^{2}+y^{2}\geq (2q+1)^{2}$ and $\sup (\left\vert x\right\vert ,\left\vert y\right\vert )\leq 2q+2\}$, with some bumps on the diameter of $T_{0}$ and on the four sides of $T_{2q+1}$ so that $T_{0}$ and $T_{2q+1}$ are fixed by no $\sigma \in G-\{\mathrm{Id}\}$. Then, in each such $\mathcal{E}$, each $\mathcal{B}_{q}^{% \mathcal{E}}(T)$ is fixed by no $\sigma \in G-\{\mathrm{Id}\}$, but the pairs $(\mathcal{B}_{q+1}^{\mathcal{E}}(T),T)$ for $T\in \mathcal{E}$ do not generally fall in finitely many isomorphism classes. We say that a group $H$ of isometries of $E$ is *discrete* if, for each $x\in E$, there exists $\eta \in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{>0}$ such that $\beta (x,\eta )\cap \{x\sigma \mid \sigma \in H\}=\{x\}$. This property is true if and only if $\beta (y,\eta )\cap \{x\sigma \mid \sigma \in H\}$ is finite for each $\eta \in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{>0}$ and any $x,y\in E$. **Proposition 1.8.** For each tiling $\mathcal{T}$ which satisfies $\Delta $, the subgroup $H=\{\sigma \in G\mid \mathcal{T}\sigma =% \mathcal{T}\}$ is discrete. **Proof.** We show that $\beta (y,\eta )\cap \{x\sigma \mid \sigma \in H\}$ is finite for each $\eta \in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{>0}$ and any $x,y\in E$. We consider $T\in \mathcal{T}$ such that $x\in T$. Any $\sigma \in H$ such that $x\sigma \in \beta (y,\eta )$ sends $T$ to a tile $U$ such that $U\cap \beta (y,\eta )\neq \varnothing $ and induces a bijection from $\mathcal{B}_{q}^{\mathcal{T}}(T)$ to $\mathcal{B}_{q}^{% \mathcal{T}}(U)$ which completely determines $\sigma $. By the first part of Proposition 1.2, $\mathcal{T}$ contains finitely many tiles $U$ such that $U\cap \beta (y,\eta )\neq \varnothing $. Moreover, for each such $U$, there exist finitely many bijections from $\mathcal{B}_{q}^{\mathcal{T}}(T)$ to $% \mathcal{B}_{q}^{\mathcal{T}}(U)$. Consequently, $\{\sigma \in H\mid x\sigma \in \beta (y,\eta )\}$ is finite.  $\blacksquare $ Now we fix $\lambda \in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{>0}$ and we state some supplementary conditions on $(E,\delta )$ and $G$ which imply that each tiling with tiles of radius $<\lambda $ satisfies a local rule with the properties stated above. We suppose that $(E,\delta )$ is geodesic and we consider the following properties: (CVX$\lambda $) For each $x\in E$ and each $\eta \in \left] 0,\lambda \right[ $, each geodesic which joins two points of $\beta (x,\eta ) $ is contained in $\cup _{\zeta <\eta }\beta (x,\zeta )$; (FIX$\lambda $) There exists $\mu \in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{>0}$ such that: (FIX$\lambda \mu $) for each $x\in E$, no $\sigma \in G-\{\mathrm{% Id}\}$ fixes the points of a set $A\subset E$ with $\beta (x,\mu )\subset \cup _{y\in A}\beta (y,\lambda )$. In euclidean or hyperbolic spaces of finite dimension, (CVX$\nu $) and (FIX$% \nu $) are true for each $\nu \in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{>0}$. On the other hand, for the surfaces of a sphere or a cylinder of infinite length, they are only true for $\nu $ small enough. We do not want to suppose from the beginning that they are true for each $\nu \in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{>0}$ since, for instance, any tiling of $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{2}$ which is invariant through a nontrivial translation induces a tiling of the surface of a cylinder of infinite length. **Lemma 1.9.** Suppose that $(E,\delta )$ is geodesic and satisfies (CVX$\lambda $). Then, for each nonempty $S\subset E$ such that $% \mathrm{Rad}(S)<\lambda $, there exists a unique $x\in E$ such that $\delta (x,y)\leq \mathrm{Rad}(S)$ for each $y\in S$, and we have $x\sigma =x$ for each isometry $\sigma $ such that $S\sigma =S$. **Proof.** We only prove the first statement, since the second one is an immediate consequence. For each $w\in E$, we have $% \sup_{y\in S}\delta (w,y)=\sup_{y\in T}\delta (w,y)$ where $T$ is the closure of $S$ in $E$. Consequently, we can suppose $S$ closed, and therefore $S$ compact. The subset $A=\{w\in E\mid \sup_{y\in S}\delta (w,y)\leq \lambda \}$ is closed, and therefore compact since it is contained in $\beta (y,\lambda )$ for each $y\in S$. Consequently, there exists $x\in A$ such that $\sup_{y\in S}\delta (x,y)=\inf_{w\in A}(\sup_{y\in S}\delta (w,y))=\mathrm{Rad}(S)$. Now suppose that there exists $x^{\prime }\neq x$ in $E$ such that $% \sup_{y\in S}\delta (x^{\prime },y)=\mathrm{Rad}(S)$, and consider $% x^{\prime \prime }\in E$ such that $\delta (x,x^{\prime \prime })=\delta (x^{\prime \prime },x^{\prime })=\delta (x,x^{\prime })/2$. By (CVX$\lambda $), we have $\delta (x^{\prime \prime },y)<\sup (\delta (x,y),\delta (x^{\prime },y))\leq \mathrm{Rad}(S)$ for each $y\in S$. It follows $% \sup_{y\in S}\delta (x^{\prime \prime },y)\leq \mathrm{Rad}(S)$, and therefore $\sup_{y\in S}\delta (x^{\prime \prime },y)=\mathrm{Rad}(S)$. As $% S $ is compact, there exists $y\in S$ such that $\delta (x^{\prime \prime },y)=\mathrm{Rad}(S)$, which contradicts (CVX$\lambda $) since $\delta (x,y)\leq \mathrm{Rad}(S)$ and $\delta (x^{\prime },y)\leq \mathrm{Rad}(S)$.  $\blacksquare $ **Proposition 1.10.** Suppose that $(E,\delta )$ is geodesic and satisfies (CVX$\lambda $). Suppose that (FIX$\lambda \mu $) is true for some $\mu \in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{>0}$. Let $\Gamma =\{(\mathcal{C}_{1},C_{1}),...,(\mathcal{C}_{m},C_{m})\}$ be a $1$-local rule defined with tiles of radius $<\lambda $. Consider $% \xi \in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{>0}$ such that, for each $i\in \{1,...,m\}$, the union of the tiles of $% \mathcal{C}_{i}$ contains $\cup _{x\in C_{i}}\beta (x,\xi )$, and $p\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$ such that $p\xi \geq \mu $. Then no $\sigma \in G-\{\mathrm{Id}\}$ stabilizes the tiles of a $p$-configuration $(\mathcal{D},D)$ which is compatible with $\Gamma $. In particular, Theorem 1.3 and its corollaries are true for each $(p+1)$-local rule $\Delta $ which is compatible with $% \Gamma $. **Proof.** Suppose that some $\sigma \in G-\{\mathrm{Id}\}$ stabilizes the tiles of a $p$-configuration $(\mathcal{D},D)$ which is compatible with $\Gamma $. As $(E,\delta )$ is geodesic and as each $(% \mathcal{B}_{1}^{\mathcal{D}}(T),T)$ with $T\in \mathcal{B}_{p-1}^{\mathcal{D% }}(D)$ is isomorphic to some $(\mathcal{C}_{i},C_{i})$, we see by induction on $0\leq k\leq p$ that $\beta (x,k\xi )$ is contained in the union of the tiles of $\mathcal{B}_{k}^{\mathcal{D}}(D)$. Consequently, $\beta (x,\mu )\subset \beta (x,p\xi )$ is contained in the union of the tiles of $% \mathcal{D}$. For each $y\in \beta (x,\mu )$ and each $T\in \mathcal{D}$ such that $y\in T$, we have $\delta (y,x_{T})\leq \mathrm{Rad}(T)<\lambda $, where $x_{T}$ is the unique point of $E$ such that $\delta (z,x_{T})\leq \mathrm{Rad}(T)$ for each $z\in T$. This contradicts (FIX$\lambda \mu $) since $x_{T}\sigma =x_{T}$ for each $T\in \mathcal{D}$ by Lemma 1.9.  $% \blacksquare $ **2. Local isomorphism and representation of relational structures by tilings.** First we define local isomorphism and the extraction preorder $\Subset $ for tilings and relational structures. The definitions for tilings are classical. We note that, by the first part of Proposition 1.2, each subpatch of a tiling is finite if and only if it is bounded. We say that a tiling $\mathcal{T}$ satisfies the* local isomorphism property* if, for each finite subpatch $\mathcal{E}$ of $\mathcal{T}$, there exists $k\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$ such that each $\mathcal{B}_{k}^{\mathcal{T}}(T)$ contains a copy of $\mathcal{E}$. Then, for each finite subpatch $\mathcal{E}$ of $\mathcal{T% }$, there exists $\rho \in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{>0}$ such that each $\beta (x,\rho )\subset E$ contains a subpatch of $% \mathcal{T}$ which is isomorphic to $\mathcal{E}$. The second part of Proposition 1.2 implies that the converse is true if $E$ is weakly homogeneous. For any tilings $\mathcal{S},\mathcal{T}$, we write $\mathcal{S}\Subset \mathcal{T}$ if each finite subpatch of $\mathcal{S}$ is isomorphic to a subpatch of $\mathcal{T}$. We say that $\mathcal{S}$ and $\mathcal{T}$ are *locally isomorphic* if $\mathcal{S}\Subset \mathcal{T}$ and $\mathcal{T% }\Subset \mathcal{S}$. Now we give the definitions and the notations for relational structures, which are similar to those in \[7, pp. 107, 112, 113\]. We consider a finite relational language $\mathcal{L}$. For each $\mathcal{L}$-structure $M$ and each $u\in M$, we define inductively the subsets $B_{M}(u,h)$ with $B_{M}(u,0)=\{u\}$ and, for $h\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $, $B_{M}(u,h+1)=B_{M}(u,h)\cup \{v\in M\mid $ there exist $% R(x_{1},...,x_{k})\in \mathcal{L}$, $u_{1},...,u_{k}\in M$ and $1\leq i,j\leq k$ such that $R(u_{1},...,u_{k})$, $u_{i}\in B_{M}(u,h)$ and $% u_{j}=v\}$. We say that $M$ is *connected* if there exists $u\in M$ such that $M=\cup _{h\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }B_{M}(u,h)$. We say that $M$ is *locally finite* if $B_{M}(u,1)$ is finite for each $% u\in M$. Then $B_{M}(u,h)$ is finite for any $u\in M$ and $h\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $. We say that $M$ is *uniformly locally finite* if there exists $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$ such that $\left\vert B_{M}(u,1)\right\vert \leq r$ for each $u\in M$. Then, for each $h\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $, there exists $s\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$ such that $\left\vert B_{M}(u,h)\right\vert \leq s$ for each $u\in M$. We say that $M$ satisfies the *local isomorphism property* if, for any $u\in M$ and $h\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $, there exists $k\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $ such that each $B_{M}(v,k)$ contains some $w$ with $(B_{M}(w,h),w)\cong (B_{M}(u,h),u)$. For any $\mathcal{L}$-structures $M,N$, we write $M\Subset N$ if, for any $% u\in M$ and $h\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $, $\left\vert \left\{ v\in M\mid (B_{M}(v,h),v)\cong (B_{M}(u,h),u)\right\} \right\vert \leq $ $\left\vert \left\{ w\in N\mid (B_{N}(w,h),w)\cong (B_{M}(u,h),u)\right\} \right\vert $ or both sets are infinite. For $M,N$ connected, we have $M\Subset N$ if and only if, for any $u\in M$ and $h\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $, there exists $v\in N$ such that $(B_{M}(u,h),u)\cong (B_{N}(v,h),v)$. We say that $M$ and $N$ are *locally isomorphic* if $M\Subset N$ and $% N\Subset M$. By \[7, Theorem 2.3\], two locally finite $\mathcal{L}$-structures $M,N$ are elementarily equivalent if and only if they are locally isomorphic. Now we consider again the metric space with bounded closed subsets $% (E,\delta )$, the group $G$ of bijective isometries of $E$, the integer $q$, the $q$-local rule $\Delta =\{(\mathcal{D}_{1},D_{1}),...,(\mathcal{D}% _{n},D_{n})\}$ such that each $\mathcal{B}_{q-1}^{\mathcal{D}_{i}}(D_{i})$ is fixed by no $\sigma \in G-\{\mathrm{Id}\}$, and the language $\mathcal{L}% _{\Delta }$, which were introduced for Theorem 1.3. We call a $\Delta $*-tiling* any tiling which satisfies $\Delta $. We define a *representation* of a $\mathcal{L}_{\Delta }$-structure $M$ as a pair $(\mathcal{E},f)$, where $\mathcal{E}$ is a $\Delta $**-**patch and $f:M\rightarrow \mathcal{E}$ is an isomorphism of $\mathcal{L}% _{\Delta }$-structures. By Theorem 1.3, the representation is unique up to isomorphism if it exists. For each $\mathcal{L}_{\Delta }$-structure $M$ and each $u\in M$, we define inductively the subsets $B_{h}^{M}(u)$ with $B_{0}^{M}(u)=\{u\}$ and, for each $h\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $, $B_{h+1}^{M}(u)=B_{h}^{M}(u)\cup \{v\in M\mid $ there exist $% 1\leq i\leq n$, $v_{1},...,v_{p(i)}\in M$ and $1\leq j,k\leq p(i)$ such that $R_{i}(v_{1},...,v_{p(i)})$, $v_{j}\in B_{h}^{M}(u)$, $v_{k}=v$ and $% D_{i,j}\cap D_{i,k}\neq \emptyset \}$ where the sets $D_{i,j}$ are the tiles used for the definition of $% \mathcal{L}_{\Delta }$. For each set $\mathcal{E}$ of tiles, each $T\in \mathcal{E}$ and each $h\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $, the $\mathcal{L}_{\Delta }$-structure $M$ defined on $\mathcal{E}$ satisfies $B_{h}^{M}(T)\subset \mathcal{B}_{h}^{\mathcal{E}}(T)$, and $% B_{h}^{M}(T)=\mathcal{B}_{h}^{\mathcal{E}}(T)$ if $\mathcal{E}$ is a $\Delta $-tiling. The notation $B_{h}^{M}(u)$ should not be confused with $% B_{M}(u,h) $. For each $\mathcal{L}_{\Delta }$-structure $M$, each $u\in M$ and each $h\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $, we have $B_{h}^{M}(u)\subset B_{M}(u,h)\subset B_{2qh}^{M}(u)$, and $% B_{M}(u,h)=B_{2qh}^{M}(u)$ if $M$ can be represented by a $\Delta $-tiling. Now we introduce something analogous to the notion of local rule considered for tilings. A *local rule* for a finite relational language $\mathcal{L% }$ is defined by specifying an integer $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$ and a finite sequence of pairs $(M_{1},x_{1}),...,(M_{k},x_{k})$ with $M_{1},...,M_{k}$ finite $\mathcal{L}$-structures and $% M_{i}=B_{M_{i}}(x_{i},r)$ for $1\leq i\leq k$. We say that a $\mathcal{L}$-structure $M$ $\emph{satisfies}$ that rule if each $(B_{M}(x,r),x)$ is isomorphic to one of the pairs $(M_{i},x_{i})$. Any such rule can be expressed by a first-order sentence. For each $s\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$, the pairs $(B_{M}(x,s),x)$ for $M$ satisfying that rule and $x\in M$ are finite and fall in finitely many isomorphism classes. In particular, any $\mathcal{L}$-structure which satisfies a local rule is uniformly locally finite. Theorem 2.1 below implies that there exists a generally infinite set of local rules which characterizes, among the connected $\mathcal{L}_{\Delta }$-structures, those which can be represented by $\Delta $-tilings. The problem of the characterization of representable $\mathcal{L}_{\Delta }$-structures by a finite set of local rules will be considered with Theorem 2.7 and the examples in Section 4. The following property of a $\mathcal{L}_{\Delta }$-structure $M$ is true, in particular, if $M$ can be represented by a $\Delta $**-**tiling: \(P) For each $u\in M$, there exist $1\leq i\leq n$ and $% u_{2},...,u_{p(i)}\in M$ such that $R_{i}(u,u_{2},...,u_{p(i)})$. **Theorem 2.1.** For each connected $\mathcal{L}_{\Delta }$-structure $M$, the following properties are equivalent: 1\) $M$ can be represented by a $\Delta $-tiling; 2\) For each $u\in M$ and each $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$, we have $(B_{M}(u,r),u)\cong (B_{\mathcal{T}}(T,r),T)$ for a tile $T$ of a $\Delta $-tiling $\mathcal{T}$; 3\) $M$ satisfies (P) and, for each $u\in M$ and each integer $% h\geq q$, there exists a representation of $B_{h}^{M}(u)$ by a $\Delta $-patch. **Proof.** The property 1) clearly implies 2). Suppose that 2) is true. Consider any element $u\in M$ and any integer $% h\geq q$. Then, for each $\Delta $-tiling $\mathcal{T}$ and each $T\in \mathcal{T}$, each isomorphism from $(B_{M}(u,h),u)\ $to $(B_{\mathcal{T}% }(T,h),T)$ induces an isomorphism from $(B_{h}^{M}(u),u)\subset (B_{M}(u,h),u)$ to $(B_{h}^{\mathcal{T}}(T),T)$. Such an isomorphism gives a representation of $B_{h}^{M}(u)$ since $B_{h}^{\mathcal{T}}(T)=\mathcal{B}% _{h}^{\mathcal{T}}(T)$ is a $\Delta $-patch. It also induces an isomorphism from $(B_{q}^{M}(u),u)$ to $(B_{q}^{\mathcal{T}}(T),T)$, which implies that there exist $1\leq i\leq n$ and $u_{2},...,u_{p(i)}\in M$ such that $R_{i}(u,u_{2},...,u_{p(i)})$. Now suppose that 3) is true and fix $u\in M$. For each integer $h\geq q$, consider a representation $(\mathcal{E}_{h},f_{h})$ of $B_{h}^{M}(u)$ by a $% \Delta $-patch. For $k\geq h\geq q$, as $f_{h}^{-1}f_{k}$ is a homomorphism of $\mathcal{L}_{\Delta }$-structures, Theorem 1.3 implies that there exists a unique $\sigma _{h,k}\in G$ such that $Sf_{h}^{-1}f_{k}=S\sigma _{h,k}$ for each $S\in \mathcal{E}_{h}$. So we can suppose that, for $k\geq h\geq q$, we have $\mathcal{E}_{h}\subset \mathcal{E}_{k}$ and $f_{h}$ is the restriction of $f_{k}$ to $B_{h}^{M}(u)$. Then $f=\cup _{h\geq q}f_{h}$ is a $\mathcal{L}_{\Delta }$**-**isomorphism from $M$ to $\mathcal{E}=\cup _{h\geq q}\mathcal{E}_{h}$. As $M$ satisfies (P), it follows that $\mathcal{E% }$ satisfies $\Delta $, and $\mathcal{E}$ is a tiling by Corollary 1.7.  $% \blacksquare $ **Corollary 2.2.** For any connected $\mathcal{L}_{\Delta }$-structures $M\Subset N$, if $N$ can be represented by a $\Delta $-tiling, then $M$ can also be represented by a $\Delta $-tiling. Similarly to \[7\], we have: **Proposition 2.3.** 1) Any $\Delta $-tiling satisfies thelocal isomorphism property as a tiling if and only if it satisfies thelocal isomorphism property as a $\mathcal{L}_{\Delta }$-structure. 2\) Two $\Delta $-tilings $\mathcal{S},\mathcal{T}$ satisfy $% \mathcal{S}\Subset \mathcal{T}$ (resp. are locally isomorphic) as tilings if and only if they satisfy $\mathcal{S}\Subset \mathcal{T}$ (resp. are locally isomorphic) as $\mathcal{L}_{\Delta }$-structures. **Proof.** The following facts will be used in the proofs of 1) and 2): a\) We have $B_{\mathcal{T}}(T,h)=B_{2qh}^{\mathcal{T}}(T)=\mathcal{% B}_{2qh}^{\mathcal{T}}(T)$ for each $\Delta $**-**tiling $\mathcal{T}$, each $T\in \mathcal{T}$ and each $h\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $. b\) Theorem 1.3 implies that, for any $\Delta $-tilings $\mathcal{S}% ,\mathcal{T}$, each $S\in \mathcal{S}$, each $T\in \mathcal{T}$ and each integer $k\geq q$, the $\Delta $-patches $\mathcal{B}_{k}^{\mathcal{S}}(S)$ and $\mathcal{B}_{k}^{\mathcal{T}}(T)$ are isomorphic as sets of tiles if and only if they are isomorphic as $\mathcal{L}_{\Delta }$-structures. c\) For any $\Delta $-tilings $\mathcal{S},\mathcal{T}$, each $S\in \mathcal{S}$, each $k\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$ and each $\sigma \in G$ such that $\mathcal{B}_{k}^{\mathcal{S}% }(S)\sigma \subset \mathcal{T}$, we have $\mathcal{B}_{k}^{\mathcal{S}% }(S)\sigma =\mathcal{B}_{k}^{\mathcal{T}}(S\sigma )$. d\) Each finite subpatch of a $\Delta $**-**tiling $\mathcal{T} $ is contained in some $\mathcal{B}_{h}^{\mathcal{T}}(T)$. First we prove 1). The facts a), b), c) above imply that $\mathcal{T}$ satisfies the local isomorphism property as a $\mathcal{L}_{\Delta }$-structure if and only if, for each $h\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$ and each $S\in \mathcal{T}$, there exists $k\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$ such that each $\mathcal{B}_{k}^{\mathcal{T}}(T)$ contains a copy of $\mathcal{B}_{h}^{\mathcal{T}}(S)$. By d), the last property is true if and only if $\mathcal{T}$ satisfies the local isomorphism property as a tiling. Now we prove 2). We only show the first statement, since the second one is an immediate consequence. The facts a), b), c) above imply that $\mathcal{S}$ and $\mathcal{T}$ satisfy $\mathcal{S}\Subset \mathcal{T}$ as $\mathcal{L}% _{\Delta }$-structures if and only if $\mathcal{T}$ contains a copy of $% \mathcal{B}_{h}^{\mathcal{S}}(S)$ for each $h\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$ and each $S\in \mathcal{S}$. By d), the last property is true if and only if $\mathcal{S}$ and $\mathcal{T}$ satisfy $\mathcal{S}\Subset \mathcal{T}$ as tilings.  $\blacksquare $ By Theorem 2.1, for each sequence $(B_{M_{i}}(u_{i},r_{i}),u_{i})_{i\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }=(B_{2qr_{i}}^{M_{i}}(u_{i}),u_{i})_{i\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }$ of pairs taken in $\mathcal{L}_{\Delta }$-structures associated to $% \Delta $**-**tilings, with $r_{i}<r_{j}$ for $i<j$, the inductive limit relative to any sequence of isomorphisms $f_{i}:(B_{M_{i}}(u_{i},r_{i}),u_{i})\rightarrow (B_{M_{i+1}}(u_{i+1},r_{i}),u_{i+1})\subset (B_{M_{i+1}}(u_{i+1},r_{i+1}),u_{i+1})$ is a pair $(M,x)$ with $M$ a $\mathcal{L}_{\Delta }$-structure associated to a $\Delta $**-**tiling. Using this fact, together with Corollary 2.2 and Proposition 2.3, we see that Corollary 2.4 (resp. 2.5, resp. 2.6) below, in the same way as \[7, Corollary 3.5\] (resp. \[7, Corollary 3.6\], resp. \[7, Corollary 3.7\]) is a consequence of \[7, Corollary 3.2\] (resp. \[7, Proposition 3.3\], resp. \[7, Proposition 3.4\]) and its proof. **Corollary 2.4.** Any $\Delta $**-**tiling is minimal for $\Subset $ if and only if it satisfies the local isomorphism property. **Corollary 2.5.** For each $\Delta $**-**tiling $% \mathcal{S}$, there exists a $\Delta $**-**tiling $\mathcal{T}\Subset \mathcal{S}$ which is minimal for $\Subset $. **Corollary 2.6.** For each $\Delta $**-**tiling $% \mathcal{S}$, we have: 1\) If there are finitely many equivalence classes of elements of $% \mathcal{S}$ modulo the isometries $\sigma \in G$ such that $\mathcal{S}% \sigma =\mathcal{S}$, then any $\Delta $**-**tiling $\mathcal{T}\Subset \mathcal{S}$ is isomorphic to $\mathcal{S}$. 2\) If $\mathcal{S}$ is minimal for $\Subset $, and if there are infinitely many equivalence classes of elements of $\mathcal{S}$ modulo the isometries $\sigma \in G$ such that $\mathcal{S}\sigma =\mathcal{S}$, then there exist $2^{\omega }$ pairwise nonisomorphic $\Delta $**-**tilings which are locally isomorphic to $\mathcal{S}$. **Remark.** Corollary 2.6 implies \[11, Theorem, p. 356\]. In \[11\], Radin and Wolff considered tilings of the euclidean spaces $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{n}$, and isomorphism was defined modulo an arbitrary group of isometries. Theorem 2.7 below is analogous to \[7, Theorem 5.2\]. Here, we suppose that $% (E,\delta )$ is weakly homogeneous and that, for any $x,y\in E$ and each $% \eta \in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{>0}$, $\beta (x,\delta (x,y))\cap \beta (y,\eta )$ is connected and contains a point $z$ with $\delta (x,z)<\delta (x,y)$. These conditions are true for each geodesic space which satisfies the properties (CVX$\nu $) considered at the end of Section 1. In each $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{k}$, they are also true for the distances defined with $\delta ((x_{1},...,x_{k}),(y_{1},...,y_{k}))=\sup (\left\vert y_{1}-x_{1}\right\vert ,...,\left\vert y_{k}-x_{k}\right\vert )$ or $\delta ((x_{1},...,x_{k}),(y_{1},...,y_{k}))=\left\vert y_{1}-x_{1}\right\vert +...+\left\vert y_{k}-x_{k}\right\vert )$. On the other hand, the connectedness condition is not true if $E$ is the surface of a cylinder of infinite length and if $\delta $ is defined by considering $E$ as a quotient of the euclidean space $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{2}$. Actually, Theorem 2.7 is not true in that case since, for some local rules $\Delta $, it is possible to find a $\mathcal{L}_{\Delta }$-structure $% M$ which can only be represented by a tiling of $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{2}$, while the substructures $B_{r}^{M}(u)$ considered in Theorem 2.7 are small enough to be represented in $(E,\delta )$. Again, we consider $\mathcal{L}_{\Delta }$-structures which satisfy the property (P) introduced for Theorem 2.1. We denote by $\lambda $ the maximum of the diameters of the tiles in $\Delta $, and $\xi $ the largest real number such that, for each $i\in \{1,...,n\}$, the union of the tiles of $% \mathcal{B}_{1}^{\mathcal{D}_{i}}(D_{i})$ contains $\cup _{x\in D_{i}}\beta (x,\xi )$. For each $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $, we write $\omega (\xi ,r)=\inf_{x\in E}\omega (x,\xi ,r)$, as in the definition of weakly homogeneous spaces. **Theorem 2.7.** Suppose that $(E,\delta )$ is weakly homogeneous and that, for any $x,y\in E$ and each $\eta \in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{>0}$, $\beta (x,\delta (x,y))\cap \beta (y,\eta )$ is connected and contains a point $z$ with $\delta (x,z)<\delta (x,y)$. Define (P), $\lambda $ and $\xi $ as above. Consider an integer $r\geq q+1$ such that $\omega (\xi ,r-q-1)\geq (4q+1)\lambda $. Let $M$ be a connected $\mathcal{L}% _{\Delta }$-structure which satisfies (P) and such that each $B_{r}^{M}(u)$ can be represented by a $\Delta $**-**patch. Then there exist a $\Delta $**-**tiling $\mathcal{T}$ and a surjective $\mathcal{L}_{\Delta }$-homomorphism $\varphi :\mathcal{T}\rightarrow M$ which induces a $\mathcal{L% }_{\Delta }$-isomorphism from $\mathcal{B}_{r}^{\mathcal{T}}(T)$ to $% B_{r}^{M}(T\varphi )$ for each $T\in \mathcal{T}$. **Remark.** The pair $(\mathcal{T},\varphi )$ satisfies $% S\varphi \neq T\varphi $ for each $S\in \mathcal{T}$ and each $T\in \mathcal{B}_{2r}^{\mathcal{T}}(S)-\{S\}$, since there exists $U\in \mathcal{T% }$ such that $S,T\in \mathcal{B}_{r}^{\mathcal{T}}(U)$. **Remark.** If $(E,\delta )$ is geodesic, we can take any integer $r\geq q+1+(4q+1)\lambda /\xi $. **Remark.** It follows that there exists a $\Delta $**-**tiling if there exists a $\mathcal{L}_{\Delta }$-structure $M$ which satisfies (P) and such that each $B_{r}^{M}(u)$ can be represented by a $% \Delta $**-**patch. **Lemma 2.7.1.** For each $u\in M$, each representation $(% \mathcal{E},f)$ of $B_{r}^{M}(u)$ and each $x\in uf$, any $T\in \mathcal{E} $ such that $T\cap \beta (x,(4q+1)\lambda )\neq \emptyset $ belongs to $% \mathrm{I}_{\Delta }(\mathcal{E})$. **Proof of Lemma 2.7.1.** As $M$ satisfies (P), for each $% v\in B_{r-q}^{M}(u)$, we have $vf\in \mathrm{I}_{\Delta }(\mathcal{E})$ and $% \cup _{y\in vf}\beta (y,\xi )$ is contained in the union of the tiles of $% B_{1}^{M}(v)f$. An induction on $k$ shows that $\beta (x,\omega (\xi ,k))$ is contained in the union of the tiles of $B_{k}^{M}(u)f$ for $0\leq k\leq r-q$. In particular, $\beta (x,(4q+1)\lambda )\subset \beta (x,\omega (\xi ,r-q-1))$ is contained in the union of the tiles of $B_{r-q-1}^{M}(u)f$. Consequently, any $T\in \mathcal{E}$ such that $T\cap \beta (x,(4q+1)\lambda )\neq \emptyset $ belongs to $B_{r-q}^{M}(u)f$, and therefore belongs to $\mathrm{I}_{\Delta }(\mathcal{E})$.  $\blacksquare $ **Proof of Theorem 2.7.** We consider an element $u\in M$, a representation $(\mathcal{G},g)$ of $B_{r}^{M}(u)$ and a point $x\in ug$. We denote by $\Omega $ the set of all pairs $(\mathcal{E},\varphi )$ such that: a\) $\mathcal{E}$ is a finite $\Delta $**-**patch, $x$ belongs to a tile of $\mathcal{E}$, any $T\in \mathcal{E}$ such that $x\in T $ belongs to $\mathrm{I}_{\Delta }(\mathcal{E})$, and $\varphi $ is a $% \mathcal{L}_{\Delta }$-homomorphism from $\mathcal{E}$ to $M$; b\) for each $S\in \mathcal{E}$, there exists $T\in \mathrm{I}% _{\Delta }(\mathcal{E})$ such that $S\in \mathcal{B}_{q}^{\mathcal{E}}(T)$ and $T\cap \beta (x,\rho _{\mathcal{E}})\neq \emptyset $, where $\rho _{% \mathcal{E}}=\sup \{\rho \in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{\geq 0}\mid $ any $U\in \mathcal{E}$ such that $U\cap \beta (x,\rho )\neq \emptyset $ belongs to $\mathrm{I}_{\Delta }(\mathcal{E})\}$. For each $(\mathcal{E},\varphi )\in \Omega $, $\beta (x,\rho _{\mathcal{E}})$ is covered by the tiles of $\mathcal{E}$: For each $y\in E$ such that $% \delta (x,y)=\rho _{\mathcal{E}}$, the second condition on $(E,\delta )$ implies that $y$ is the limit of a sequence $(y_{k})_{k\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }\subset \cup _{\zeta <\rho _{\mathcal{E}}}\beta (x,\zeta )$. As $\mathrm{I}% _{\Delta }(\mathcal{E})$ is finite, infinitely many $y_{k}$ belong to the same $T\in \mathrm{I}_{\Delta }(\mathcal{E})$ and $T$ also contains $y$. On the other hand, there exists $y\in E$ with $\delta (x,y)=\rho _{\mathcal{E% }}$ which belongs to some $S\in \mathcal{E}-\mathrm{I}_{\Delta }(\mathcal{E}% ) $: For any sequences $(y_{k})_{k\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }\subset E$ and $(S_{k})_{k\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }\subset \mathcal{E}-\mathrm{I}_{\Delta }(\mathcal{E})$ such that $y_{k}\in S_{k}$ for each $k\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $ and $\lim_{k\rightarrow +\infty }\delta (y_{k},\beta (x,\rho _{\mathcal{E}% }))=0$, there exists a subsequence of $(y_{k})_{k\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }$ which converges to a point $y\in \beta (x,\rho _{\mathcal{E}})$. As $% \mathcal{E}-\mathrm{I}_{\Delta }(\mathcal{E})$ is finite, infinitely many $% y_{k}$ in this subsequence belong to the same $S_{h}$ and $S_{h}$ also contains $y$. By Lemma 2.7.1, $\Omega $ contains $(\mathcal{E}_{0},\varphi _{0})$, where $% \mathcal{E}_{0}$ is the union of the subsets $\mathcal{B}_{q}^{\mathcal{G}% }(T)$ for $T\in \mathcal{G}$ such that $T\cap \beta (x,(4q+1)\lambda )\neq \emptyset $, and $\varphi _{0}$ is the restriction of $g^{-1}$ to $\mathcal{E% }_{0}$. We have $\rho _{\mathcal{E}_{0}}>(4q+1)\lambda $. For any $(\mathcal{E},\varphi ),(\mathcal{F},\psi )\in \Omega $, we write $(% \mathcal{E},\varphi )\leq (\mathcal{F},\psi )$ if $\mathcal{E}\subset \mathcal{F}$ and if $\varphi $ is the restriction of $\psi $ to $\mathcal{E% }$. It suffices to prove the two following claims: 1.* The conclusion of Theorem 2.7 is true if* $\Omega $*contains some strictly increasing* $(\mathcal{E}_{i},\varphi _{i})_{i\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }$. First we observe that, for each $i\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $, there exists $T\in \mathrm{I}_{\Delta }(\mathcal{E}_{i+1})-\mathrm{I}% _{\Delta }(\mathcal{E}_{i})$ such that $T\cap \beta (x,\rho _{\mathcal{E}% _{i}})\neq \emptyset $: If $\rho _{\mathcal{E}_{i+1}}>\rho _{\mathcal{E}% _{i}} $, then any $T\in \mathcal{E}_{i}-\mathrm{I}_{\Delta }(\mathcal{E}% _{i})$ such that $T\cap \beta (x,\rho _{\mathcal{E}_{i}})\neq \emptyset $belongs to$\ \mathrm{I}_{\Delta }(\mathcal{E}_{i+1})$. If $\rho _{\mathcal{E}% _{i+1}}=\rho _{\mathcal{E}_{i}}$, then, for each $S\in \mathcal{E}_{i+1}-% \mathcal{E}_{i}$,$\ $there exists $T\in \mathrm{I}_{\Delta }(\mathcal{E}% _{i+1})\ $such that $S\in \mathcal{B}_{q}^{\mathcal{E}_{i+1}}(T)$ and $T\cap \beta (x,\rho _{\mathcal{E}_{i}})=\emptyset $; any such $T$ cannot belong to $\mathrm{I}_{\Delta }(\mathcal{E}_{i})$ since $\mathcal{B}_{q}^{\mathcal{E}% _{i+1}}(T)-\mathcal{B}_{q}^{\mathcal{E}_{i}}(T)\neq \emptyset $. Then we use the function $\omega (\xi ,n)$ in order to prove that $\rho _{% \mathcal{E}_{i}}$ tends to infinity. As $(E,\delta )$ is weakly homogeneous, $\omega (\xi ,n)$ tends to infinity with $n$. Consequently, it suffices to show that, for any $i,n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $ such that $\rho _{\mathcal{E}_{i}}>\omega (\xi ,n)$, there exists $j>i$ such that $\rho _{\mathcal{E}_{j}}>\omega (\xi ,n+1)$. For each $j\geq i$ such that $\rho _{\mathcal{E}_{j}}\leq \omega (\xi ,n+1)$, we consider$\ T_{j}\in \mathrm{I}_{\Delta }(\mathcal{E}_{j+1})-\mathrm{I}% _{\Delta }(\mathcal{E}_{j})$ such that $T_{j}\cap \beta (x,\rho _{\mathcal{E}% _{j}})\neq \emptyset $, and $y_{j}\in T_{j}\cap \beta (x,\rho _{\mathcal{E}% _{j}})$. There exist $z_{j}\in \beta (x,\omega (\xi ,n))$ such that $\delta (y_{j},z_{j})\leq \xi $, and $U_{j}\in \mathcal{B}_{1}^{\mathcal{E}j}(T_{j})$ such that $z_{j}\in U_{j}$. We have $U_{j}\in \mathrm{I}_{\Delta }(\mathcal{E% }_{i})$ since $z_{j}\in \beta (x,\omega (\xi ,n))$ and $\omega (\xi ,n)<\rho _{\mathcal{E}_{i}}$. It follows $T_{j}\in \mathcal{E}_{i}$. As $% \mathcal{E}_{i}$ is finite, there exist at most finitely many such $T_{j}$, and therefore finitely many integers $j>i$ such that $\rho _{\mathcal{E}% _{j}}\leq \omega (\xi ,n+1)$. As $\rho _{\mathcal{E}_{i}}$ tends to infinity, $\mathcal{T}=\cup _{i\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }\mathcal{E}_{i}$ satisfies $\Delta $, and $\mathcal{T}$ is a $\Delta $-tiling by Corollary 1.7. The inductive limit $\varphi $ of the maps $% \varphi _{i}$ is a $\mathcal{L}_{\Delta }$-homomorphism from $\mathcal{T}$ to $M$. In particular, we have $\mathcal{B}_{k}^{\mathcal{T}}(T)\varphi \subset B_{k}^{M}(T\varphi )$ for each $T\in \mathcal{T}$ and each $k\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $. Now we show that, for each $T\in \mathcal{T}$, $\varphi $ induces a $% \mathcal{L}_{\Delta }$-isomorphism from $\mathcal{B}_{r}^{\mathcal{T}}(T)$ to $B_{r}^{M}(T\varphi )$; it follows that $\varphi $ is surjective since $% M $ is connected. We consider a representation $(\mathcal{H},h)$ of $B_{r}^{M}(T\varphi )$. As $\mathcal{B}_{r}^{\mathcal{T}}(T)$ and $\mathcal{H}$ are $\Delta $-patches, Theorem 1.3 implies that there exists $\sigma \in G$ such that $% S\sigma =S\varphi h$ for each $S\in \mathcal{B}_{r}^{\mathcal{T}}(T)$. As $% \mathcal{T}$ is a $\Delta $-tiling and $\mathcal{H}$ is a $\Delta $-patch, $% \sigma $ induces an isomorphism from $\mathcal{B}_{q}^{\mathcal{T}}(U)$ to $% \mathcal{B}_{q}^{\mathcal{H}}(U\sigma )$ for each $U\in \mathcal{B}_{r-q}^{% \mathcal{T}}(T)\subset I_{\Delta }(\mathcal{B}_{r}^{\mathcal{T}}(T))$. Consequently, we have $\mathcal{B}_{r}^{\mathcal{T}}(T)\sigma =\mathcal{B}% _{r}^{\mathcal{H}}(T\sigma )=\mathcal{B}_{r}^{\mathcal{H}}(T\varphi h)=B_{r}^{M}(T\varphi )h=\mathcal{H}$. It follows that $\sigma $ induces an isomorphism from $\mathcal{B}_{r}^{\mathcal{T}}(T)$ to $\mathcal{H}$ and $% \varphi $ induces a $\mathcal{L}_{\Delta }$-isomorphism from $\mathcal{B}% _{r}^{\mathcal{T}}(T)$ to $B_{r}^{M}(T\varphi )$. 2\. *For each* $(\mathcal{E},\varphi )\in \Omega $*, there exists* $(\mathcal{F},\psi )>(\mathcal{E},\varphi )$* in* $\Omega $. We consider $y\in E$ with $\delta (x,y)=\rho _{\mathcal{E}}$ which belongs to some $S\in \mathcal{E}-\mathrm{I}_{\Delta }(\mathcal{E})$, and $T\in \mathrm{I}_{\Delta }(\mathcal{E})\ $such that $y\in T$. There exist a representation of $B_{r}^{M}(T\varphi )$, and therefore a $\Delta $-patch $% \mathcal{E}^{\prime }$ and a $\mathcal{L}_{\Delta }$-isomorphism $\varphi ^{\prime }$ from $\mathcal{E}^{\prime }$ to $B_{r}^{M}(T\varphi )$. We have $\mathcal{B}_{q}^{\mathcal{E}}(T)\varphi \subset B_{q}^{M}(T\varphi )$ since $T$ belongs to $\mathrm{I}_{\Delta }(\mathcal{E})$, and $B_{q}^{M}(T\varphi )\varphi ^{\prime -1}\subset \mathcal{B}_{q}^{\mathcal{E}^{\prime }}(T\varphi \varphi ^{\prime -1})$ since $\varphi ^{\prime }$ is a $% \mathcal{L}_{\Delta }$-isomorphism. The map $\mathcal{B}_{q}^{\mathcal{E}% }(T)\rightarrow \mathcal{B}_{q}^{\mathcal{E}^{\prime }}(T\varphi \varphi ^{\prime -1}):U\rightarrow U\varphi \varphi ^{\prime -1}$ is a $\mathcal{L}% _{\Delta }$-homomorphism. As $\mathcal{B}_{q}^{\mathcal{E}}(T)$ and $% \mathcal{B}_{q}^{\mathcal{E}^{\prime }}(T\varphi \varphi ^{\prime -1})$ are $% \Delta $-patches, Theorem 1.3 implies that there exists $\sigma \in G$ such that $U\sigma =U\varphi \varphi ^{\prime -1}$ for each $U\in \mathcal{B}% _{q}^{\mathcal{E}}(T)$. So we can suppose for the remainder of the proof that $\mathcal{B}_{q}^{\mathcal{E}}(T)\subset \mathcal{E}^{\prime }$ and $% U\varphi =U\varphi ^{\prime }\ $for each $U\in \mathcal{B}_{q}^{\mathcal{E}% }(T)$. We denote by $\mathcal{F}\ $the union of $\mathcal{E}\ $and the subsets $% \mathcal{B}_{q}^{\mathcal{E}^{\prime }}(U)$ for $U\in \mathcal{E}^{\prime }$ such that $y\in U$. We consider the map $\psi :\mathcal{F}\rightarrow M$ with $U\psi =U\varphi $ for $U\in \mathcal{E}$ and $U\psi =U\varphi ^{\prime }$ for $U\in \mathcal{F}-\mathcal{E}$. First we prove that, for each $U\in \mathcal{F}$, we have $\mathcal{B}_{q}^{% \mathcal{F}}(U)\subset \mathcal{E}$ and $V\psi =V\varphi $ for each $V\in \mathcal{B}_{q}^{\mathcal{F}}(U)$, or $\mathcal{B}_{q}^{\mathcal{F}% }(U)\subset \mathcal{E}^{\prime }$ and $V\psi =V\varphi ^{\prime }$ for each $V\in \mathcal{B}_{q}^{\mathcal{F}}(U)$. If $U\cap \beta (y,2q\lambda )=\emptyset $, then the tiles in $\mathcal{B}% _{q}^{\mathcal{F}}(U)$ cannot belong to $\mathcal{F}-\mathcal{E}$ since they contain no point in $\beta (y,q\lambda )$. Consequently, we have $\mathcal{B% }_{q}^{\mathcal{F}}(U)\subset \mathcal{E}$ and $V\psi =V\varphi $ for each $% V\in \mathcal{B}_{q}^{\mathcal{F}}(U)$. If $U\cap \beta (y,2q\lambda )\neq \emptyset $, then we prove that $\mathcal{% B}_{q}^{\mathcal{F}}(U)\subset \mathcal{E}^{\prime }$ and $V\psi =V\varphi ^{\prime }$ for each $V\in \mathcal{B}_{q}^{\mathcal{F}}(U)$. For each $V\in \mathcal{B}_{q}^{\mathcal{F}}(U)$, we have $V\cap \beta (y,3q\lambda )\neq \emptyset $. So it suffices to show that each $V\in \mathcal{E}$ with $V\cap \beta (y,3q\lambda )\neq \emptyset $ satisfies $V\in \mathcal{E}^{\prime }$ and $V\varphi =V\varphi ^{\prime }$. We consider $W\in \mathrm{I}_{\Delta }(% \mathcal{E})$ such that $V\in \mathcal{B}_{q}^{\mathcal{E}}(W)$ and $W\cap \beta (x,\rho _{\mathcal{E}})\neq \emptyset $. As $W\cap \beta (y,4q\lambda )\neq \emptyset $, the properties $V\in \mathcal{E}^{\prime }$ and $% V\varphi =V\varphi ^{\prime }$ follow from Lemma 2.7.2 below. Consequently, $\psi $ is a $\mathcal{L}_{\Delta }$-homomorphism and, for each $U\in \mathcal{F}$, there exists $i\in \{1,...,n\}$ such that $(% \mathcal{B}_{q}^{\mathcal{F}}(U),U)\leq (\mathcal{D}_{i},D_{i})$. Moreover, the definition of $(\mathcal{F},\psi )$ implies that, for each $V\in \mathcal{F}$, there exists $U\in \mathrm{I}_{\Delta }(\mathcal{E})$ such that $V\in \mathcal{B}_{q}^{\mathcal{E}}(U)$ and $U\cap \beta (x,\rho _{% \mathcal{E}})\neq \emptyset $, or $U\in \mathrm{I}_{\Delta }(\mathcal{E}% ^{\prime })$ such that $V\in \mathcal{B}_{q}^{\mathcal{E}^{\prime }}(U)$ and $y\in U$; in both cases, we have $U\in \mathrm{I}_{\Delta }(\mathcal{F})$, $% U\cap \beta (x,\rho _{\mathcal{E}})\neq \emptyset $ and $V\in \mathcal{B}% _{q}^{\mathcal{F}}(U)$. Consequently, $\mathcal{F}$ is a finite $\Delta $**-**patch and satisfies the conditions of the definition of $\Omega $.$% ~~\blacksquare $ **Lemma 2.7.2.** For each $W\in \mathrm{I}_{\Delta }(\mathcal{% E})$ such that $W\cap \beta (x,\rho _{\mathcal{E}})\neq \emptyset $ and $% W\cap \beta (y,4q\lambda )\neq \emptyset $, we have $\mathcal{B}_{q}^{% \mathcal{E}}(W)\subset \mathcal{E}^{\prime }$ and $X\varphi =X\varphi ^{\prime }$ for each $X\in \mathcal{B}_{q}^{\mathcal{E}}(W)$. **Proof of Lemma 2.7.2.** We fix $z\in W\cap \beta (x,\rho _{% \mathcal{E}})$. We have $z\in \beta (y,(4q+1)\lambda )$. The set $A=\beta (x,\rho _{\mathcal{E}})\cap \beta (y,(4q+1)\lambda )$ is connected and contained in the union of the tiles of $\mathrm{I}_{\Delta }(\mathcal{E})$. Consequently, $\mathcal{A}=\{X\in \mathrm{I}_{\Delta }(\mathcal{E})\mid X\cap A\neq \emptyset \}$ is connected. Moreover, $T$ and $W$ belong to $% \mathcal{A}$ since $y$ and $z$ belong to $A$. We are going to prove that, for each $X\in \mathcal{A}$, we have $\mathcal{B}% _{q}^{\mathcal{E}}(X)\subset \mathcal{E}^{\prime }$ and $Z\varphi =Z\varphi ^{\prime }$ for each $Z\in \mathcal{B}_{q}^{\mathcal{E}}(X)$. As this property is true for $X=T$, it suffices to show that, if it is true for some $X\in \mathcal{A}$, then it is true for any $Y\in \mathcal{A}$ such that $% X\cap Y\neq \emptyset $. We have $Y\in \mathcal{B}_{1}^{\mathcal{E}}(X)$, and therefore $\mathcal{B}% _{q-1}^{\mathcal{E}}(Y)\subset \mathcal{E}^{\prime }$ and $Z\varphi =Z\varphi ^{\prime }$ for each $Z\in \mathcal{B}_{q-1}^{\mathcal{E}}(Y)$.Moreover, by Lemma 2.7.1, we have $Y\in \mathrm{I}_{\Delta }(\mathcal{E}% ^{\prime })$ since $Y\cap \beta (y,(4q+1)\lambda )\neq \emptyset $, and therefore $\mathcal{B}_{q}^{\mathcal{E}^{\prime }}(Y)\varphi ^{\prime }=B_{q}^{M}(Y\varphi ^{\prime })=B_{q}^{M}(Y\varphi )$ since $\varphi ^{\prime }$ is a $\mathcal{L}_{\Delta }$-isomorphism from $\mathcal{E}% ^{\prime }$ to $B_{r}^{M}(T\varphi )$. We also have $\mathcal{B}_{q}^{% \mathcal{E}}(Y)\varphi \subset B_{q}^{M}(Y\varphi )$ since $Y\in \mathrm{I}% _{\Delta }(\mathcal{E})$. Consequently, $\varphi \varphi ^{\prime -1}$ is defined on $\mathcal{B}_{q}^{\mathcal{E}}(Y)$ and stabilizes the elements of $\mathcal{B}_{q-1}^{\mathcal{E}}(Y)$.$\ $It follows $\mathcal{B}_{q}^{% \mathcal{E}}(Y)=\mathcal{B}_{q}^{\mathcal{E}^{\prime }}(Y)$ and $Z\varphi =Z\varphi ^{\prime }$ for each $Z\in \mathcal{B}_{q}^{\mathcal{E}}(Y)$.$% ~~\blacksquare $ For each relational language $\mathcal{L}$, each $\mathcal{L}$-structure $M$ and each subgroup $H$ of $\mathrm{Aut}(M)$, we define the $\mathcal{L}$-structure $M/H$ as follows: The elements of $M/H$ are the classes $xH$ for $% x\in M$. For $R(u_{1},...,u_{k})\in \mathcal{L}$ and $x_{1},...,x_{k}\in M/H$, we write $R(x_{1},...,x_{k})$ if there exist some representatives $% y_{1},...,y_{k}$ of $x_{1},...,x_{k}$ in $M$ such that $R(y_{1},...,y_{k})$. For each $i\in \{1,...,k\}$ and each representative $y_{i}$ of $x_{i}$ in $% M$, there exist some representatives $y_{1},...,y_{i-1},y_{i+1},...,y_{k}$ of $x_{1},...,x_{i-1},x_{i+1},...,x_{k}$ in $M$ such that $% R(y_{1},...,y_{k})$. The canonical surjection from $M$ to $M/H$ is a homomorphism. If $H$ is normal in $\mathrm{Aut}(M)$, then any automorphism of $M$ induces an automorphism of $M/H$. In Proposition 2.8 below, we do not use the supplementary hypotheses on $% (E,\delta )$ which were introduced for Theorem 2.7. For each $\Delta $**-**tiling $\mathcal{T}$ and each subgroup $H$ of $G$ such that $% T\sigma \in \mathcal{T}-\mathcal{B}_{2q}^{\mathcal{T}}(T)$ for each $T\in \mathcal{T}$ and each $\sigma \in H$, we denote by $\mathcal{T}/H$ the tiling of $E/H$ induced by $\mathcal{T}$. The canonical surjection $\pi :% \mathcal{T}\rightarrow \mathcal{T}/H$ is a $\mathcal{L}_{\Delta }$-homomorphism. Proposition 2.8 implies that, in Theorem 2.7, the $\mathcal{L}% _{\Delta }$-structure $M$ is isomorphic to a quotient of a $\Delta $**-**tiling. **Proposition 2.8.** Consider a $\Delta $**-**tiling $% \mathcal{T}$, a $\mathcal{L}_{\Delta }$-structure $M$ and a surjective $% \mathcal{L}_{\Delta }$-homomorphism $\varphi :\mathcal{T}\rightarrow M$ which induces a $\mathcal{L}_{\Delta }$-isomorphism from $\mathcal{B}_{q}^{% \mathcal{T}}(T)$ to $B_{q}^{M}(T\varphi )$ for each $T\in \mathcal{T}$. Then $\varphi $ induces a $\mathcal{L}_{\Delta }$-isomorphism from $\mathcal{% T}/H$ to $M$, where $H=\{\sigma \in G\mid \mathcal{T}\sigma =\mathcal{T}$ and $T\sigma \varphi =T\varphi $ for each $T\in \mathcal{T}\}$. **Proof**. First we show that $\varphi $ induces a bijective homomorphism from $\mathcal{T}/H$ to $M$. For any $S,T\in \mathcal{T}$such that $S\varphi =T\varphi $, as $\varphi $ induces some $\mathcal{L}% _{\Delta }$-isomorphisms from $\mathcal{B}_{q}^{\mathcal{T}}(S)$ and $% \mathcal{B}_{q}^{\mathcal{T}}(T)$ to $B_{q}^{M}(S\varphi )=B_{q}^{M}(T\varphi )$, Theorem 1.3 implies that there exists a unique $% \sigma _{S,T}\in G$ such that $S\sigma _{S,T}=T$ and $U\varphi =U\sigma _{S,T}\varphi $ for each $U\in \mathcal{B}_{q}^{\mathcal{T}}(S)$. For any $% S,T\in \mathcal{T}$ such that $S\varphi =T\varphi $ and for each $U\in \mathcal{B}_{1}^{\mathcal{T}}(S)$, we have $\sigma _{S,T}=\sigma _{U,V}$ where $V=U\sigma _{S,T}$, since $\sigma _{S,T}$ and $\sigma _{U,V}$ coincide on $\mathcal{B}_{q-1}^{\mathcal{T}}(S)$. Consequently, for any $S,T\in \mathcal{T}$ such that $S\varphi =T\varphi $, we have $\mathcal{T}\sigma _{S,T}=\mathcal{T}$ and $U\varphi =U\sigma _{S,T}\varphi $ for each $U\in \mathcal{T}$. It remains to be proved that, for each $R(w_{1},...,w_{k})\in \mathcal{L}% _{\Delta }$ and any $T_{1},...,T_{k}\in \mathcal{T}$ such that $M$ satisfies $R(T_{1}\varphi ,...,T_{k}\varphi )$, there exist $U_{1}\in T_{1}H,...,U_{k}\in T_{k}H$ such that $\mathcal{T}$ satisfies $% R(U_{1},...,U_{k})$. As $\varphi $ induces a $\mathcal{L}_{\Delta }$-isomorphism from $\mathcal{B}_{q}^{\mathcal{T}}(T_{1})$ to $% B_{q}^{M}(T_{1}\varphi )$, there exist $U_{2},...,U_{k}\in \mathcal{B}_{q}^{% \mathcal{T}}(T_{1})$ such that $R(T_{1},U_{2},...,U_{k})$ and $U_{2}\varphi =T_{2}\varphi ,...,U_{k}\varphi =T_{k}\varphi $. We have $U_{2}\in T_{2}H,...,U_{k}\in T_{k}H$ according to the first part of the proof.  $% \blacksquare $ **3. Periodicity, invariance through a translation.** In the present section, we consider a finite relational language $\mathcal{L} $. We generalize to uniformly locally finite $\mathcal{L}$-structures the notions of periodicity and invariance through a nontrivial translation, which are usually considered for tilings of the euclidean spaces $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{n}$. In particular, we obtain generalizations for tilings of noneuclidean spaces. The notions of mathematical logic used for Proposition 3.1 and Corollary 3.2 are defined, for instance, in \[5\]. **Proposition 3.1.** Consider a formula $\theta (u,v)$ in $% \mathcal{L}$ and two elementarily equivalent $\mathcal{L}$-structures $M,N$ with $M$ connected locally finite. Suppose that there exist an element $x\in M$ such that $\left\{ y\in M\mid \theta (x,y)\right\} $ is finite, and an automorphism $g$ of $N$ such that $\theta (y,yg)$ for each $y\in N$. Then there exists an automorphism $f$ of $M$ such that $\theta (y,yf)$ for each $% y\in M$. Moreover, for each $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $, if $\left\{ y\in N\mid yg=y\right\} $ contains no ball $B_{N}(z,r)$, then we can choose $f$ in such a way that $\left\{ y\in M\mid yf=y\right\} $ contains no ball $B_{M}(z,r)$. **Proof**. Fix $x\in M$ such that $\left\{ y\in M\mid \theta (x,y)\right\} $ is finite. For each $k\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $, as $B_{M}(x,k)$ is finite, the following property of a $\mathcal{L}$-structure $P$ can be expressed by one sentence: For each $y\in P\ $such that $(B_{P}(y,k),y)\cong (B_{M}(x,k),x)$, there exist an element $z\in P$ and an isomorphism $f:(B_{P}(y,k),y)% \rightarrow (B_{P}(z,k),z)$ such that $\theta (u,uf)$ for each $u\in B_{P}(y,k)$ (respectively $\theta (u,uf)$ for each $u\in B_{P}(y,k)$ and $% \left\{ v\in B_{P}(y,k)\mid vf=v\right\} $ contains no ball $B_{P}(u,r)$ for $u\in B_{P}(y,k-r)$). Suppose that there exists an automorphism $g$ of $N$ such that $\theta (y,yg) $ for each $y\in N$ (respectively $\theta (y,yg)$ for each $y\in N$ and $\left\{ y\in N\mid yg=y\right\} $ contains no ball $B_{N}(z,r)$). Then the sentence considered above is true in $N$, and therefore true in $M$. For each $k\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $, consider the nonempty set $A_{k}$ consisting of the pairs $(y,f)$ with $% y\in M$ and $f:(B_{M}(x,k),x)\rightarrow (B_{M}(y,k),y)$ isomorphism such that $\theta (u,uf)$ for each $u\in B_{M}(x,k)$ (respectively $\theta (u,uf)$ for each $u\in B_{M}(x,k)$ and $\left\{ v\in B_{M}(x,k)\mid vf=v\right\} $ contains no ball $B_{M}(u,r)$ for $u\in B_{M}(x,k-r)$). Then each $A_{k}$ is finite since $M$ is locally finite and $\left\{ y\in M\mid \theta (x,y)\right\} $ is finite. Moreover, for $0\leq k\leq l$, each pair $% (z,g)\in A_{l}$ gives by restriction a pair $(y,f)\in A_{k}$. Consequently, by König’s lemma, there exists a sequence $% (y_{k},f_{k})_{k\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }\in \Pi _{k\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }A_{k}$ with $f_{k}$ restriction of $f_{l}$ for $0\leq k\leq l$. The inductive limit of such a sequence gives an automorphism of $M$ which satisfies the required properties.  $\blacksquare $ By \[7, Theorem 2.3\], two locally finite $\mathcal{L}$-structures are elementarily equivalent if and only if they are locally isomorphic. Moreover, for any $r,s\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$, there exists a formula $\theta _{r,s}(u,v)$ which expresses the property $v\in B_{N}(u,r)$ in each $\mathcal{L}$-structure $N$ such that $% \left\vert B_{N}(x,r)\right\vert \leq s$ for each $x\in N$. Consequently, we have: **Corollary 3.2.** Consider two locally isomorphic uniformly locally finite $\mathcal{L}$-structures $M,N$ which $M$ connected. Let $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$ and $s\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $. Suppose that there exists an automorphism $g$ of $N$ such that $yg\in B_{N}(y,r)$ for each $y\in N$, and such that $\left\{ y\in N\mid yg=y\right\} $ contains no ball $B_{N}(z,s)$. Then, there exists an automorphism $f$ of $M$ such that $yf\in B_{M}(y,r)$ for each $y\in M$, and such that $\left\{ y\in M\mid yf=y\right\} $ contains no ball $B_{M}(z,s)$. **Remark.** In particular, for each $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$, if $N$ has an automorphism $g$ without fixed point such that $% yg\in B_{N}(y,r)$ for each $y\in N$, then $M$ has an automorphism $f$ without fixed point such that $yf\in B_{M}(y,r)$ for each $y\in M$. Now we consider the metric space $(E,\delta )$, the group $G$ of isometries of $E$ and the set $\Delta $ defined in Section 1. For each $% \Delta $-tiling $\mathcal{T}$ and each $\sigma \in G$ such that $\mathcal{T}% \sigma =\mathcal{T}$, we say that $\sigma $ is a *translation of* $% \mathcal{T}$ if there exists $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$ such that $T\sigma \in \mathcal{B}_{r}^{\mathcal{T}}(T)$ for each $T\in \mathcal{T}$. The set $\mathrm{Trans}(\mathcal{T})$ of all translations of* *$\mathcal{T}$ is a subgroup of $G$. Any $\sigma \in G$ such that $\mathcal{T}\sigma =\mathcal{T}$ is a translation of $\mathcal{T}$ if and only if there exists $s\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$ such that $T\sigma \in B_{\mathcal{T}}(T,s)$ for each $T\in \mathcal{T}$. Moreover, for each $T\in \mathcal{T}$, there exists no $% \sigma \in G-\{\mathrm{Id}\}$ such that $S\sigma =S$ for each $S\in B_{% \mathcal{T}}(T,1)=\mathcal{B}_{2q}^{\mathcal{T}}(T)$. Consequently, the result below follows from Corollary 3.2: **Corollary 3.3**. For any locally isomorphic $\Delta $-tilings $\mathcal{S},\mathcal{T}$, we have $\mathrm{Trans}(\mathcal{S}% )=\left\{ \mathrm{Id}\right\} $ if and only if $\mathrm{Trans}(\mathcal{T}% )=\left\{ \mathrm{Id}\right\} $. For each $\Delta $-tiling $\mathcal{T}$ and each $\sigma \in \mathrm{Trans}(% \mathcal{T})$, $\sup_{x\in E}\delta (x,x\sigma )$ is finite. Conversely, if $% (E,\delta )$ is weakly homogeneous, then the second part of Proposition 1.2 implies that any $\sigma \in G$ such that $\mathcal{T}\sigma =\mathcal{T}$ belongs to $\mathrm{Trans}(\mathcal{T})$ if $\sup_{x\in E}\delta (x,x\sigma ) $ is finite. For each $k\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$, if $E$ is the set $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{k}$ equipped with a distance defined from a norm, then we have $\sigma \in \mathrm{Trans}(\mathcal{T})$ if and only if $\sigma $ is a translation in the usual sense, since any surjective isometry of $E$ is affine in that case by \[2, Th. 14.1\]. The following result generalizes well known properties of the translations in the euclidean spaces $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{k}$. Here, we use the conditions (CVX$\nu $) and (FIX$\nu $) introduced at the end of Section 1. **Theorem 3.4.** Let $\mathcal{T}$ be a $\Delta $-tiling. Then, for each $\sigma \in \mathrm{Trans}(\mathcal{T})-\left\{ \mathrm{Id}% \right\} $, there exist $x,y\in E$ such that $\delta (x,x\sigma )=\inf_{z\in E}\delta (z,z\sigma )$ and $\delta (y,y\sigma )=\sup_{z\in E}\delta (z,z\sigma )$. Moreover, if $(E,\delta )$ is geodesic and satisfies (CVX$\nu $) and (FIX$\nu $) for each $\nu \in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{>0}$, then $\mathrm{Trans}(\mathcal{T})$ is torsion-free abelian and each $% \sigma \in \mathrm{Trans}(\mathcal{T})-\left\{ \mathrm{Id}\right\} $ has no fixed point. **Proof.** We fix $\sigma \in \mathrm{Trans}(\mathcal{T}% )-\left\{ \mathrm{Id}\right\} $, we write $\alpha =\inf_{z\in E}\delta (z,z\sigma )$ and we show that there exists $x\in E$ such that $\delta (x,x\sigma )=\alpha $. It can be proved in a similar way that there exists $% y\in E$ such that $\delta (y,y\sigma )=\sup_{z\in E}\delta (z,z\sigma )$. We consider $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$ such that $T\sigma \in \mathcal{B}_{r}^{\mathcal{T}}(T)$ for each $% T\in \mathcal{T}$. For each $T\in \mathcal{T}$, $\sigma $ induces a bijection $\sigma _{T}:\mathcal{B}_{q}^{\mathcal{T}}(T)\rightarrow \mathcal{B% }_{q}^{\mathcal{T}}(T\sigma )$. The triples $(\mathcal{B}_{q+r}^{\mathcal{T}% }(T),T,\sigma _{T})$ for $T\in \mathcal{T}$ fall in finitely many isomorphism classes. Consequently, there exist two sequences $(x_{k})_{k\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }\subset E$ and $(T_{k})_{k\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }\subset \mathcal{T}$ with $x_{k}\in T_{k}$ for each $k\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $ such that $\lim \delta (x_{k},x_{k}\sigma )=\alpha $ and such that all the triples $(\mathcal{B}_{q+r}^{\mathcal{T}}(T_{k}),T_{k},\sigma _{T_{k}})$ are isomorphic. For each $k\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $, we consider $\tau _{k}\in G$ which induces an isomorphism from $(% \mathcal{B}_{q+r}^{\mathcal{T}}(T_{0}),T_{0},\sigma _{T_{0}})$ to $(\mathcal{% B}_{q+r}^{\mathcal{T}}(T_{k}),T_{k},\sigma _{T_{k}})$. We have $\sigma =\tau _{k}\sigma \tau _{k}^{-1}$ for each $k\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $ since $T\sigma =T\sigma _{T_{0}}=T\tau _{k}\sigma _{T_{k}}\tau _{k}^{-1}=T\tau _{k}\sigma \tau _{k}^{-1}$ for each $T\in \mathcal{B}_{q}^{% \mathcal{T}}(T_{0})$. Consequently, the elements $y_{k}=x_{k}\tau _{k}^{-1}\in T_{0}$ satisfy $\delta (y_{k},y_{k}\sigma )=\delta (x_{k},x_{k}\sigma )$, and therefore $\lim \delta (y_{k},y_{k}\sigma )=\alpha $. As $T_{0}$ is compact, it follows that there exists $x\in T_{0}$ such that $\delta (x,x\sigma )=\alpha $. Now suppose that $(E,\delta )$ is geodesic and satisfies (CVX$\nu $) and (FIX$\nu $) for each $\nu \in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{>0}$. Consider $\sigma \in \mathrm{Trans}(\mathcal{T})$ and $s\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion -\{0,1\}$ such that $\sigma ^{s}=\mathrm{Id}$. Then, for each $x\in E$, we have $A_{x}\sigma =A_{x}$ for $A_{x}=\{x,x\sigma ,...,x\sigma ^{s-1}\}$; by Lemma 1.9, there exists a unique $w_{x}\in E$ such that $\delta (w_{x},z)\leq \mathrm{Rad}(A_{x})$ for each $z\in A_{x}$; it follows $% w_{x}\sigma =w_{x}$. The properties (FIX$\nu $) imply $\sigma =\mathrm{Id}$ since $\delta (x,w_{x})\leq \mathrm{Rad}(A_{x})\leq (s/2)\sup_{z\in E}\delta (z,z\sigma )$ for each $x\in E$. Now consider any $\sigma \in \mathrm{Trans}(\mathcal{T})$ with a fixed point $x$. Then, for each $y\in E$, there exists $s\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$ such that $y\sigma ^{s}=y$, since $\mathrm{Trans}(\mathcal{T})$ is discrete by Proposition 1.8 and the elements $y\sigma ^{k}$ for $k\in %TCIMACRO{\U{2124} }% %BeginExpansion \mathbb{Z} %EndExpansion $ all belong to $\beta (x,\delta (x,y))$. Consider $\zeta ,\eta \in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{>0}$ such that no $\tau \in G-\{\mathrm{Id}\}$ fixes the points of a set $% A\subset E$ with $\beta (y,\eta )\subset \cup _{z\in A}\beta (z,\zeta )$ for some $y\in E$, and choose a finite set $A$ with that property. Then there exists $s\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$ such that $z\sigma ^{s}=z$ for each $z\in A$, which implies $% \sigma ^{s}=\mathrm{Id}$ and $\sigma =\mathrm{Id}$. It remains to be proved that any $\sigma ,\tau \in \mathrm{Trans}(\mathcal{T}% )$ commute. It suffices to show that $\sigma $ commutes with $\tau ^{r}$ for some $r\in %TCIMACRO{\U{2124} }% %BeginExpansion \mathbb{Z} %EndExpansion ^{\ast }$, since $(\sigma ^{-1}\tau \sigma )^{r}=\sigma ^{-1}\tau ^{r}\sigma =\mathrm{Id}$ implies $\sigma ^{-1}\tau \sigma =\mathrm{Id}$. We fix $x\in E$ and we consider $\alpha \in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{>0}$ such that $\delta (y,y\sigma )\leq \alpha $ for each $y\in E$. For each $r\in %TCIMACRO{\U{2124} }% %BeginExpansion \mathbb{Z} %EndExpansion $, we have $\delta (x,x\sigma ^{-1}\tau ^{-r}\sigma \tau ^{r})\leq 2\alpha $ since $\delta (x,x\sigma ^{-1})\leq \alpha $ and $\delta (x\sigma ^{-1},x\sigma ^{-1}\tau ^{-r}\sigma \tau ^{r})=\delta (x\sigma ^{-1}\tau ^{-r},x\sigma ^{-1}\tau ^{-r}\sigma )\leq \alpha $. As $\mathrm{Trans}(\mathcal{T})$ is discrete, it follows that there exist $% r\neq s$ in $% %TCIMACRO{\U{2124} }% %BeginExpansion \mathbb{Z} %EndExpansion $ such that $x\sigma ^{-1}\tau ^{-r}\sigma \tau ^{r}=x\sigma ^{-1}\tau ^{-s}\sigma \tau ^{s}$. We have $\sigma ^{-1}\tau ^{-r}\sigma \tau ^{r}=\sigma ^{-1}\tau ^{-s}\sigma \tau ^{s}$ since the elements of $\mathrm{% Trans}(\mathcal{T})$ have no fixed point. It follows $\tau ^{-r}\sigma \tau ^{r}=\tau ^{-s}\sigma \tau ^{s}$ and $\tau ^{s-r}\sigma =\sigma \tau ^{s-r}$.  $\blacksquare $ Now we generalize the notion of periodicity to connected $\mathcal{L}$-structures. We say that such a structure $M$ is *periodic* if it contains a finite set $A$ such that $M=\cup _{\sigma \in \mathrm{Aut}% (M)}A\sigma $. **Proposition 3.5.** Let $M,N$ be locally isomorphic locally finite connected $\mathcal{L}$-structures. If $N$ is periodic then $M$ is isomorphic to $N$. **Proof.** We fix $w\in M$. As $N$ is periodic and connected, there exist $z\in N$ and $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $ such that $N=\cup _{\sigma \in \mathrm{Aut}(N)}B_{N}(z,r)\sigma $. For each $s\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $, as $M$ and $N$ are locally isomorphic, there exists $x_{s}\in N$ such that $(B_{M}(w,s),w)\cong (B_{N}(x_{s},s),x_{s})$, and therefore $y_{s}\in B_{N}(z,r)$ such that $(B_{M}(w,s),w)\cong (B_{N}(y_{s},s),y_{s})$. We consider the nonempty set $A_{s}$ consisting of the isomorphisms $\theta :(B_{M}(w,s),w)\rightarrow (B_{N}(y,s),y)$ with $y\in B_{N}(z,r)$. The sets $A_{s}$ are finite since $N$ is locally finite. Moreover, for any $% s\leq t$, the restriction of each $\theta \in A_{t}$ to $(B_{M}(w,s),w)$ belongs to $A_{s}$. Consequently, by König’s Lemma, there exists a strictly increasing sequence $(\theta _{s})_{s\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }$ with $\theta _{s}\in A_{s}$ for each $s\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $. The inductive limit is an isomorphism $\theta :(M,w)\rightarrow (N,y)$with $y\in B_{N}(z,r)$.  $\blacksquare $ We call a *period* of a $\mathcal{L}$-structure $M$ any set $A\subset M$ such that $M$ is the disjoint union of the sets $A\sigma $ for $\sigma \in \mathrm{Aut}(M)$. If $M$ is periodic and if it has some periods, then all of them have the same finite number of elements. We call it the *periodicity rank* of $M$. For each $\mathcal{L}$-structure $M$, we say that $A\subset M$ is *weakly connected* if, for each subset $B$ with $\varnothing \subsetneq B\subsetneq A$, there exist $k\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion -\{0,1\}$ and $R(u_{1},...,u_{k})\in \mathcal{L}$ which is satisfied by some $(x_{1},...,x_{k})\in M^{k}$ with $\{x_{1},...,x_{k}\}\cap B\neq \varnothing $ and $\{x_{1},...,x_{k}\}\cap (A-B)\neq \varnothing $. **Proposition 3.6.** Any connected $\mathcal{L}$-structure $M$ has a weakly connected period if the nontrivial automorphisms of $M$ have no fixed point. **Proof.** We show that $M=\cup _{\sigma \in \mathrm{Aut}% (M)}A\sigma $ for each $A\subset M$ which is maximal for the conjunction of the two properties: $A$ weakly connected and $A\cap A\sigma =\varnothing $for each $\sigma \in \mathrm{Aut}(M)-\{\mathrm{Id}\}$. Otherwise, as $M$ is connected, there exist $k\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion -\{0,1\}$, $(x_{1},...,x_{k})\in M^{k}$ which satisfies some $% R(u_{1},...,u_{k})\in \mathcal{L}$ and $1\leq i,j\leq k$ such that $x_{i}\in \cup _{\sigma \in \mathrm{Aut}(M)}A\sigma $ and $x_{j}\in M-\cup _{\sigma \in \mathrm{Aut}(M)}A\sigma $. We consider $\tau \in \mathrm{Aut}(M)$ such that $x_{i}\tau \in A$. Then $B=A$ $\cup $ $\left\{ x_{j}\tau \right\} $ is weakly connected since $(x_{1}\tau ,...,x_{k}\tau )$ satisfies $R$. As $% x_{j}\tau $ is not a fixed point of a nontrivial automorphism of $M$ and does not belong to $\cup _{\sigma \in \mathrm{Aut}(M)}A\sigma $, we have $% B\sigma \cap B=\varnothing $ for each $\sigma \in \mathrm{Aut}(M)-\{\mathrm{% Id}\}$, contrary to the maximality of $A$.  $\blacksquare $ Now we introduce some supplementary conditions which will be used for the investigation of periodic $\mathcal{L}$-structures. We say that a $\mathcal{L}$-structure $M$ is *equational* (cf. \[7, Section 4\]) if $R(x_{1},...,x_{k})$, $R(y_{1},...,y_{k})$ and $x_{i}=y_{i}$ imply $x_{j}=y_{j}$ for each $k\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$, each $R(u_{1},...,u_{k})\in \mathcal{L}$, any $% (x_{1},...,x_{k}),(y_{1},...,y_{k})\in M^{k}$ and any $i,j\in \{1,...,k\}$. Any equational $\mathcal{L}$-structure is uniformly locally finite. If it is connected, then its nontrivial automorphisms have no fixed point. For each equational $\mathcal{L}$-structure $M$, each $k\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion -\{0,1\}$, each $R(u_{1},...,u_{k})\in \mathcal{L}$, any $i,j\in \{1,...,k\}$ such that $i\neq j$ and any $x,y\in M$, we write $x(R,i,j)=y$ if there exists $(z_{1},...,z_{k})\in M^{k}$ which satisfies $R$ with $z_{i}=x$ and $z_{j}=y$. For each $n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $, each word $w=(R_{1},i_{1},j_{1})...(R_{n},i_{n},j_{n})$ and any $x,y\in M $, we write $xw=y$ if there exist $z_{0},...,z_{n}\in M$ such that $% z_{0}=x $, $z_{n}=y$ and $z_{m-1}(R_{m},i_{m},j_{m})=z_{m}$ for $1\leq m\leq n$. We denote by $\Omega _{\mathcal{L}}$ the set of all such words. For any equational $\mathcal{L}$-structures $M,N$, each $x\in M$ and each $% r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$, we call a *partial isomorphism* from $B_{M}(x,r)$ to $N$ any injective map $\rho $ such that, for each $k\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$, each $R(u_{1},...,u_{k})\in \mathcal{L}$, any $i,j\in \{1,...,k\}$ and any $y,z\in B_{M}(x,r)$, $y(R,i,j)$ exists if and only if $y\rho (R,i,j)$ exists and $y(R,i,j)=z$ if and only if $y\rho (R,i,j)=z\rho $. If $\rho $ is a partial isomorphism from $B_{M}(x,r)$ to $N$, then $\rho $is an isomorphism from $(B_{M}(x,r),x)$ to $(B_{N}(x\rho ,r),x\rho )$ and $% \rho ^{-1}$ is a partial isomorphism from $B_{N}(x\rho ,r)$ to $M$. Any isomorphism $\sigma :(B_{M}(x,r+1),x)\rightarrow (B_{N}(x\sigma ,r+1),x\sigma )$ gives by restriction a partial isomorphism from $B_{M}(x,r)$ to $N$. We say that an equational $\mathcal{L}$-structure $M$ is *commutative* if, for each $x\in M$ and each $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$, each partial isomorphism $\rho :B_{M}(x,r)\rightarrow M$ satisfies $y\rho \sigma =y\sigma \rho $ for each automorphism $\sigma $ of $% M $ and each $y\in B_{M}(x,r)$ such that $y\sigma \in B_{M}(x,r)$. We say that $M$ is *strongly commutative* if we have $xvw=xwv$ for each $x\in M$ and any $v,w\in \Omega _{\mathcal{L}}$ such that $xvw$ and $xwv$ exist. Any strongly commutative connected equational $\mathcal{L}$-structure is commutative since, for $x,y,\rho ,\sigma $ defined as above, there exist $% v,w\in \Omega _{\mathcal{L}}$ such that the equalities $y\rho =yv$ and $% y\sigma =yw$ are respectively true in $M$ and in $B_{M}(x,r)$, which implies $y\rho \sigma =yvw$ and $y\sigma \rho =ywv$. We say that a class $\mathcal{E}$ of equational $\mathcal{L}$-structures is *strongly regular* (cf. \[7, Section 4\]) if $xw=x$ and $yw=y$ are equivalent for any $M,N\in \mathcal{E}$, each $x\in M$, each $y\in N$ and each $w\in \Omega _{\mathcal{L}}$ such that $xw$ and $yw$ exist. Equationality, strong commutativity and strong regularity are local properties. They are preserved by local isomorphism. **Proposition 3.7.** Let $M$ be a connected equational commutative $\mathcal{L}$-structure. If $M$ is periodic of rank $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$, then, for each $x\in M$ and each integer $s\geq r$, each partial isomorphism from $B_{M}(x,s)$ to $M$ can be extended into a unique automorphism of $M$. **Proof.** The extension is necessarily unique since the nontrivial automorphisms of an equational $\mathcal{L}$-structure have no fixed point. It remains to be proved that it exists. By Proposition 3.6, $M$ has a weakly connected period $A$ with $\left\vert A\right\vert =r$. Replacing $A$ by its image through an appropriate automorphism of $M$, we reduce the proof to the case $x\in A$. Then $A$ is contained in $B_{M}(x,r-1)$. For each partial isomorphism $\rho :B_{M}(x,s)\rightarrow M$, we define a map $\overline{\rho }:M\rightarrow M$ as follows: for each $y\in M$, as the nontrivial automorphisms of $M$ have no fixed point, there exists a unique $\sigma \in \mathrm{Aut}(M)$ such that $y\sigma \in A$; we write $y% \overline{\rho }=y\sigma \rho \sigma ^{-1}$. For each $y\in B_{M}(x,s)$, we have $y\overline{\rho }=y\rho $ because the commutativity of $M$ implies $% y\rho \sigma =y\sigma \rho $. We observe that $\overline{\rho }^{-1}$ is defined in the same way from the partial isomorphism $\rho ^{-1}:B_{M}(x\rho ,s)\rightarrow M$ and the weakly connected period $A\rho $. Consequently, $% \overline{\rho }$ is bijective and, by reason of symmetry, it suffices to show that $\overline{\rho }$ is a homomorphism. For each $k\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$, each $R(u_{1},...,u_{k})\in \mathcal{L}$ and each $% (x_{1},...,x_{k})\in M^{k}$ which satisfies $R$, we consider $\sigma _{1},...,\sigma _{k}\in \mathrm{Aut}(M)$ such that $x_{1}\sigma _{1},...,x_{k}\sigma _{k}$ belong to $A$. Then $x_{1}\sigma _{1},...,x_{k}\sigma _{k}$ belong to $B_{M}(x,r-1)$ and $x_{2}\sigma _{1},...,x_{k}\sigma _{1}$ belong to $B_{M}(x,r)$ since $(x_{1}\sigma _{1},x_{2}\sigma _{1},...,x_{k}\sigma _{1})$ satisfies $R$. We have $x_{1}\overline{\rho }=x_{1}\sigma _{1}\rho \sigma _{1}^{-1}$. For $% 2\leq i\leq k$, we have $x_{i}\overline{\rho }=x_{i}\sigma _{i}\rho \sigma _{i}^{-1}=(x_{i}\sigma _{1})(\sigma _{1}^{-1}\sigma _{i})\rho (\sigma _{1}^{-1}\sigma _{i})^{-1}\sigma _{1}^{-1}$. As $x_{i}\sigma _{1}$ and $% (x_{i}\sigma _{1})(\sigma _{1}^{-1}\sigma _{i})=x_{i}\sigma _{i}$ belong to $% B_{M}(x,r)$, the commutativity of $M$ implies $(x_{i}\sigma _{1})(\sigma _{1}^{-1}\sigma _{i})\rho =(x_{i}\sigma _{1})\rho (\sigma _{1}^{-1}\sigma _{i})$, and therefore $x_{i}\overline{\rho }=(x_{i}\sigma _{1})(\sigma _{1}^{-1}\sigma _{i})\rho (\sigma _{1}^{-1}\sigma _{i})^{-1}\sigma _{1}^{-1}=(x_{i}\sigma _{1})\rho (\sigma _{1}^{-1}\sigma _{i})(\sigma _{1}^{-1}\sigma _{i})^{-1}\sigma _{1}^{-1}=x_{i}\sigma _{1}\rho \sigma _{1}^{-1}$. Moreover, $x_{1}\sigma _{1}\rho ,x_{2}\sigma _{1}\rho ,...,x_{k}\sigma _{1}\rho $ are all defined since $x_{1}\sigma _{1},x_{2}\sigma _{1},...,x_{k}\sigma _{1}$ belong to $B_{M}(x,r)$. Consequently, $% (x_{1}\sigma _{1}\rho ,x_{2}\sigma _{1}\rho ,...,x_{k}\sigma _{1}\rho )$ satisfies $R$ like $(x_{1}\sigma _{1},x_{2}\sigma _{1},...,x_{k}\sigma _{1})$ and $(x_{1}\overline{\rho },x_{2}\overline{\rho },...,x_{k}\overline{% \rho })=(x_{1}\sigma _{1}\rho \sigma _{1}^{-1},x_{2}\sigma _{1}\rho \sigma _{1}^{-1},...,x_{k}\sigma _{1}\rho \sigma _{1}^{-1})$ also$\ $satisfies $R$.  $\blacksquare $ **Lemma 3.8.** Let $M,N$ be equational $\mathcal{L}$-structures with $N$ connected, commutative and periodic of rank $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$. For each $x\in M$ and each integer $s\geq r$, if there exists a partial isomorphism from $B_{M}(x,s+1)$ to $N$, then each partial isomorphism from $B_{M}(x,s)$ to $N$ can be extended into a partial isomorphism from $B_{M}(x,s+1)$ to $N$. **Proof.** Let us consider a partial isomorphism $\rho :B_{M}(x,s)\rightarrow N$ and a partial isomorphism $\sigma :B_{M}(x,s+1)\rightarrow N$. Then $\rho ^{-1}\sigma $ is a partial isomorphism from $B_{N}(x\rho ,s)$ to $N$ which can be extended into a unique automorphism $\theta $ of $N$ by Proposition 3.7, and $\sigma \theta ^{-1}$ is a partial isomorphism from $B_{M}(x,s+1)$ to $N$ which extends $% \rho $.  $\blacksquare $ **Lemma 3.9.** Let $M,N$ be equational $\mathcal{L}$-structures with $\left\{ M,N\right\} $ strongly regular. Then, for each $% x\in M$ and each $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$, each partial isomorphism $\rho :B_{M}(x,r)\rightarrow N$ can be extended into a partial isomorphism from $B_{M}(x,r+1)$ to $N$ if, for each $y\in B_{M}(x,r)$, there exists a partial isomorphism $\rho _{y}:B_{M}(y,1)\rightarrow N$ such that $y\rho _{y}=y\rho $. **Proof.** We define a map $\sigma :B_{M}(x,r+1)\rightarrow N$ as follows: for each $z\in B_{M}(x,r+1)$, we consider $y\in B_{M}(x,r)$ such that $z\in B_{M}(y,1)$ and we write $z\sigma =z\rho _{y}$. For any $z_{1},z_{2}\in B_{M}(x,r+1)$ and any $y_{1},y_{2}\in B_{M}(x,r)$such that $z_{1}\in B_{M}(y_{1},1)$ and $z_{2}\in B_{M}(y_{2},1)$, there exist $v,(R_{1},i_{1},j_{1}),(R_{2},i_{2},j_{2})\in \Omega _{\mathcal{L}}$such that the equalities $y_{1}v=y_{2}$, $y_{1}(R_{1},i_{1},j_{1})=z_{1}$ and $y_{2}(R_{2},i_{2},j_{2})=z_{2}$ are respectively true in the connected $\mathcal{L}$-structures $B_{M}(x,r)$, $B_{M}(y_{1},1)$ and $B_{M}(y_{2},1)$. We have $z_{2}=z_{1}w$ for $w=(R_{1},j_{1},i_{1})v(R_{2},i_{2},j_{2})$. We also have $z_{2}\rho _{y_{2}}=z_{1}\rho _{y_{1}}w$ since the equalities $% z_{1}\rho _{y_{1}}(R_{1},j_{1},i_{1})=y_{1}\rho _{y_{1}}=y_{1}\rho $, $% y_{1}\rho v=y_{2}\rho =y_{2}\rho _{y_{2}}$ and $y_{2}\rho _{y_{2}}(R_{2},i_{2},j_{2})=z_{2}\rho _{y_{2}}$ are respectively true in $% B_{N}(y_{1}\rho _{y_{1}},1)$, $B_{N}(x\rho ,r)$ and $B_{N}(y_{2}\rho _{y_{2}},1)$. By strong regularity, the equalities $z_{2}=z_{1}w$ and $z_{2}\rho _{y_{2}}=z_{1}\rho _{y_{1}}w$ imply that $z_{1}\rho _{y_{1}}=z_{2}\rho _{y_{2}}$ if and only if $z_{1}=z_{2}$. Consequently, $\sigma $ is injective and the definition of $z\sigma $ given above does not depend on the choice of the element $y\in B_{M}(x,r)$ such that $z\in B_{M}(y,1)$. It follows that $z\sigma =z\rho $ for each $z\in B_{M}(x,r)$ and $z\sigma =z\rho _{y}$ for each $y\in B_{M}(x,r)$ and each $z\in B_{M}(y,1)$. For each $z\in B_{M}(x,r+1)$ and each $(R,i,j)\in \Omega _{\mathcal{L}}$, the element $% z\sigma (R,i,j)$ exists if and only if $z(R,i,j)$ exists, because, for each $% y\in B_{M}(x,r)$, $z\rho _{y}(R,i,j)$ exists if and only if $z(R,i,j)$ exists. Now let us consider again $y_{1},y_{2},z_{1},z_{2},w$ as above. For each$% (R,i,j)$ such that $z_{1}(R,i,j)$ exists, or equivalently such that $% z_{1}\sigma (R,i,j)$ exists, we have $z_{1}(R,i,j)w^{\prime }=z_{2}$ and $% z_{1}\sigma (R,i,j)w^{\prime }=z_{2}\sigma $ for $w^{\prime }=(R,j,i)w$. Strong regularity implies that $z_{1}(R,i,j)=z_{2}$ if and only if $% z_{1}\sigma (R,i,j)=z_{2}\sigma $.  $\blacksquare $ **Theorem 3.10.** Let $M,N$ be connected periodic $\mathcal{L} $-structures of ranks $r,s\geq 1$ with $N$ commutative and $\left\{ M,N\right\} $ strongly regular. Then, for each $x\in M$, each partial isomorphism from $B_{M}(x,r+s)$ to $N$ can be extended into an isomorphism from $M$ to $N$. **Proof.** Otherwise, there exist an integer $t\geq r+s$ and a partial isomorphism $\rho :B_{M}(x,t)\rightarrow N$ which cannot be extended into a partial isomorphism from $B_{M}(x,t+1)$ to $N$. By lemma 3.9, there exists $y\in B_{M}(x,t)-B_{M}(x,t-1)$ such that no partial isomorphism $\sigma :B_{M}(y,1)\rightarrow N$ satisfies $y\sigma =y\rho $. For each $k\in \{0,...,r\}$, we consider $z_{k}\in B_{M}(x,t-s-k)-B_{M}(x,t-s-k-1)$ such that $y\in B_{M}(z_{k},s+k)$. The restriction $\rho _{k}$ of $\rho $ to $B_{M}(z_{k},s+k)$ cannot be extended into a partial isomorphism from $B_{M}(z_{k},s+k+1)$ to $N$. As $M$ is periodic of rank $r$, there exist two integers $0\leq k_{1}<k_{2}\leq r$ and an isomorphism $\sigma :(M,z_{k_{1}})\rightarrow (M,z_{k_{2}})$. Then $\sigma \rho _{k_{2}}$ is a partial isomorphism from $% B_{M}(z_{k_{1}},s+k_{2})$ to $N$. By Lemma 3.8, it follows that $\rho _{k_{1}}$ can be extended into a partial isomorphism from $% B_{M}(z_{k_{1}},s+k_{2})$ to $N$, whence a contradiction.  $\blacksquare $ **Theorem 3.11.** For each set $\Sigma $ of local rules for $% \mathcal{L}$, if the connected $\mathcal{L}$-structures satisfying $\Sigma $ are periodic, equational and commutative, and if their class $\mathcal{E}$ is strongly regular, then $\mathcal{E}$ is a finite union of isomorphism classes. **Proof.** Otherwise, it follows from Theorem 3.10 that there exists a sequence $(N_{i})_{i\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }$ of periodic $\mathcal{L}$-structures with strictly increasing periodicity ranks which satisfy $\Sigma $. For each $q\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$, the pairs $(B_{N_{i}}(y,q),y)$ for $i\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $ and $y\in N_{i}$ fall in finitely many isomorphism classes. König’s lemma applied to the isomorphism classes of pairs $(B_{N_{i}}(y,q),y)$ for $% q\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$, $i\geq q$ and $y\in N_{i}$ implies that there exist two strictly increasing sequences of integers $(k(i))_{i\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }$ and $(q_{i})_{i\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }$, and a sequence $(y_{i})_{i\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }\in \times _{i\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }N_{k(i)}$ such that $(B_{N_{k(i)}}(y_{i},q_{i}),y_{i})\cong (B_{N_{k(j)}}(y_{j},q_{i}),y_{j})$ for $i\leq j$. Consequently, there exist a sequence $(M_{i})_{i\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }$ of periodic $\mathcal{L}$-structures with strictly increasing periodicity ranks which satisfy $\Sigma $ and a sequence $(x_{i})_{i\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }\in \times _{i\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }M_{i}$ such that $(B_{M_{i}}(x_{i},i),x_{i})\cong (B_{M_{j}}(x_{j},i),x_{j}) $ for $i\leq j$. The inductive limit of the pairs $(B_{M_{i}}(x_{i},i),x_{i})$ for these isomorphisms, which are unique and compatible because of equationality, is a pair $(M,x)$ with $M$ satisfying $\Sigma $ and $x\in M$. We denote by $r$ the periodicity rank of $% M$ and, for each $i\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $, $r_{i}$ the periodicity rank of $M_{i}$. For each $i\geq r$, we have $r_{i}>r$. Consequently, $M_{i}$ is not isomorphic to $M$. As $(B_{M_{i}}(x_{i},i),x_{i})$ and $(B_{M}(x,i),x)$ are isomorphic, there exist an integer $s_{i}\geq i-1$ and a partial isomorphism $\rho _{i}:B_{M_{i}}(x_{i},s_{i})\rightarrow M$ which cannot be extended into a partial isomorphism from $B_{M_{i}}(x_{i},s_{i}+1)$ to $M$. By Lemma 3.9, there exists $y_{i}\in B_{M_{i}}(x_{i},s_{i})-B_{M_{i}}(x_{i},s_{i}-1)$ such that no partial isomorphism $\sigma :B_{M_{i}}(y_{i},1)\rightarrow M$ satisfies $y_{i}\sigma =y_{i}\rho _{i}$. For any integers $i\geq r$ and $1\leq s\leq s_{i}$, we consider $z_{i,s}\in B_{M_{i}}(x_{i},s_{i}-s)$ such that $y_{i}\in B_{M_{i}}(z_{i,s},s)$. Then $% B_{M_{i}}(y_{i},2s)$ contains $B_{M_{i}}(z_{i,s},s)$. It follows from König’s Lemma applied to the pairs $% (B_{M_{i}}(y_{i},2s+1),y_{i})$ for $i\geq r$ and $1\leq s\leq s_{i}$ that there exist two strictly increasing sequences of integers $(k(i))_{i\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }$ and $(t_{i})_{i\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }$, with $k(i)\geq r$ and $t_{i}\leq s_{k(i)}$ for each $i\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $, such that $(B_{M_{k(i)}}(y_{k(i)},2t_{i}+1),y_{k(i)})\cong (B_{M_{k(j)}}(y_{k(j)},2t_{i}+1),y_{k(j)})$ for $i\leq j$. We denote by $% (N,y)$ the inductive limit of the pairs $% (B_{M_{k(i)}}(y_{k(i)},2t_{i}+1),y_{k(i)})$ relative to these isomorphisms, which are unique and compatible because of equationality. The $\mathcal{L}$-structure $N$ is periodic because it satisfies $\Sigma $. We denote by $t$ the periodicity rank of $N$. For each $i\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $, we have $y\in B_{N}(z_{i},t_{i})$ where $z_{i}$ is the image of $% z_{k(i),t_{i}}$ in $N$. Moreover $\rho _{k(i)}$ induces a partial isomorphism $\sigma _{i}:B_{N}(z_{i},t_{i})\rightarrow M$ such that no partial isomorphism $\tau :B_{N}(y,1)\rightarrow M$ satisfies $y\tau =y\sigma _{i}$. For $i\geq r+t-1$, this property contradicts Theorem 3.10 since we have $t_{i}\geq r+t$.  $\blacksquare $ Now we consider again the metric space $(E,\delta )$, the group $G$ of isometries of $E$ and the set $\Delta $ defined in Section 1. We say that a $\Delta $-tiling $\mathcal{T}$ is *periodic* if there exists a finite subset $\mathcal{E}$ of $\mathcal{T}$ such that $\mathcal{T}$ is the union of the subsets $\mathcal{E}\sigma $ for $\sigma \in G$ such that $\mathcal{% T}\sigma =\mathcal{T}$. By Theorem 1.3, this property is true if and only if the $\mathcal{L}_{\Delta }$-structure associated to $\mathcal{T}$ is periodic. For each $k\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$, if $(E,\delta )$ is the space $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{k}$ equipped with a distance defined from a norm and if $G$ consists of the translations of $E$, then our notion of periodicity coincides with the classical one. Consequently, the following result generalizes \[1, Th. 3.8\] which was proved for tilings of $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{2}$ by square tiles: **Theorem 3.12.** If $(E,\delta )$ is geodesic and satisfies (CVX$\lambda $) where $\lambda $ is the maximum of the radii of the tiles, if $G$ is commutative and if the elements of $G-\{\mathrm{Id}\}$ have no fixed point, then each class of periodic $\Delta $-tilings defined by local rules is a finite union of isomorphism classes. **Lemma 3.12.1.** The $\mathcal{L}_{\Delta }$-structures associated to $\Delta $-tilings are equational, strongly commutative, and form a strongly regular class. **Proof of Lemma 3.12.1.** If some $\sigma \in G$ stabilizes a tile, then $\sigma =\mathrm{Id}$ since $\sigma $ has a fixed point by Lemma 1.9. Consequently, Theorem 1.3 implies that the $\mathcal{L}_{\Delta } $-structures associated to $\Delta $-tilings are equational. Moreover, for any $\Delta $-tilings $\mathcal{S},\mathcal{T}$, each $(R,i,j)\in \Omega _{% \mathcal{L}}$, each $S\in \mathcal{S}$ and each $T\in \mathcal{T}$, if $% S(R,i,j)$ and $T(R,i,j)$ exist, then there exists a unique $\sigma \in G$such that $S\sigma =T$ and it satisfies $S(R,i,j)\sigma =T(R,i,j)$. By induction, it follows that, for each $w\in \Omega _{\mathcal{L}}$, each $% S\in \mathcal{S}$ and each $T\in \mathcal{T}$, if $Sw$ and $Tw$ exist, then there exists a unique $\sigma \in G$ such that $S\sigma =T$ and $% Sw\sigma =Tw$. In particular, we have $Sw=S$ if and only if $Tw=T$. For each $\Delta $-tiling $\mathcal{T}$, each $T\in \mathcal{T}$ and any $% v,w\in \Omega _{\mathcal{L}}$ such that $Tvw$ and $Twv$ exist, consider $% \sigma ,\tau \in G$ such that $Tv=T\sigma $, $Twv=Tw\sigma $, $Tw=T\tau $ and $Tvw=Tv\tau $. Then we have $Tvw=T\sigma \tau =T\tau \sigma =Twv$since $\sigma \tau =\tau \sigma $.  $\blacksquare $ **Proof of Theorem 3.12.** By Lemma 3.12.1, the $\mathcal{L}% _{\Delta }$-structures associated to $\Delta $-tilings are equational, commutative, and form a strongly regular class. If a class $\mathcal{C}$ of $\Delta $-tilings is defined by local rules, then, by Theorem 2.1, the same property is true for the class $K$ consisting of the associated $\mathcal{L}% _{\Delta }$-structures. The structures in $K$ are periodic if the tilings in $\mathcal{C}$ are periodic. Then Theorem 3.11 implies that $K$, and therefore $\mathcal{C}$, is a finite union of isomorphism classes.  $% \blacksquare $ **4. Local isomorphism, rigidness and aperiodicity.** We say that a tiling or a relational structure is *rigid* if it has no nontrivial automorphism. In the present section, we are interested in characterizing tilings, and more generally relational structures, which are locally isomorphic to rigid ones. Theorem 4.1 (respectively Corollary 4.2) gives a characterization for relational structures (respectively tilings) which are uniformly locally finite and satisfy the local isomorphism property. Corollary 4.3 gives a simpler characterization concerning tilings of the euclidean spaces $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{n}$, where isomorphism is defined modulo a group of isometries. In \[7\], we considered tilings of the euclidean spaces $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{n}$, and isomorphism was defined up to translation. In that case, \[7, Proposition 2.4\] implies that the relational structure $M$ associated to a tiling is rigid if and only if the tiling is not invariant through any nontrivial translation. If a connected structure $N$ is locally isomorphic to $M$, then $N$ is associated to another tiling by \[7, Corollary 5.4\]. Moreover, according to \[7, Proposition 5.1\], the tilings associated to $M$ and $N$ are invariant through the same translations of $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{n}$. It follows that $N$ is rigid if and only if $M$ is rigid. Examples 1, 2, 3, which are given after Corollary 4.3, imply that the last property is no longer true if we consider isomorphism modulo an arbitrary group of isometries of an euclidean space $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{n}$, or tilings of a noneuclidean space. Similarly, Example 4 illustrates Theorem 4.1 for relational structures which are not associated to tilings. Examples 5 and 6 are given in order to show the importance of each hypothesis in Theorem 4.1. **Theorem 4.1.** Let $\mathcal{L}$ be a finite relational language and let $M$ be a uniformly locally finite $\mathcal{L}$-structure which satisfies the local isomorphism property. Then $M$ is locally isomorphic to a connected rigid $\mathcal{L}$-structure if and only if, for each $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$, there exists $x\in M$ such that $(M,y)\cong (M,z)\ $implies $y=z$ for $y,z\in B_{M}(x,r)$. **Proof.** By König’s lemma, the following properties are equivalent for locally finite $\mathcal{L}$-structures: \(P) There exist $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$ and, for each $x\in N$, $y\neq z$ in $B_{N}(x,r)$ such that $% (N,y)\cong (N,z)$. \(Q) There exist $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$ and, for each $x\in N$ and each $s\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $, $y\neq z$ in $B_{N}(x,r)$ such that $(B_{N}(y,s),y)\cong (B_{N}(z,s),z)$. Any $\mathcal{L}$-structure $N$ which is locally isomorphic to $M$ is uniformly locally finite like $M$. If $M$ satisfies (P), and therefore satisfies (Q), then $N$ satisfies (Q) since it locally isomorphic to $M$, and therefore satisfies (P). Consequently, $N$ is not rigid. Now, let us suppose that $M$ does not satisfy (P). First we show that there exist a sequence $(x_{n})_{n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }$ in $M$ and two strictly increasing sequences $(r_{n})_{n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }$ and $(s_{n})_{n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }$ in $% %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $ such that, for each $n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $, $(B_{M}(x_{n},r_{n}+s_{n}),x_{n})\cong (B_{M}(x_{n+1},r_{n}+s_{n}),x_{n+1})$ and $B_{M}(x_{n},r_{n})$ contains no elements $y\neq z$ with $(B_{M}(y,s_{n}),y)\cong (B_{M}(z,s_{n}),z)$. We write $r_{0}=s_{0}=0$ and we take for $x_{0}$ any element of $M$. For each $n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $, supposing that $x_{n},r_{n},s_{n}$ are already defined, we define $% x_{n+1},r_{n+1},s_{n+1}$ as follows: As $M$ satisfies the local isomorphism property, there exists $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $ such that each $B_{M}(x,r)$ contains some $y$ with $% (B_{M}(y,r_{n}+s_{n}),y)\cong (B_{M}(x_{n},r_{n}+s_{n}),x_{n})$; we take $% r>r_{n}$. As $M$ does not satisfy (Q), there exist $x\in M$ and $s\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$ such that $B_{M}(x,2r)$ contains no elements $y\neq z$ with $% (B_{M}(y,s),y)\cong (B_{M}(z,s),z)$. We take for $x_{n+1}$ any $u\in B_{M}(x,r)$ such that $(B_{M}(u,r_{n}+s_{n}),u)\cong (B_{M}(x_{n},r_{n}+s_{n}),x_{n})$. Then $B_{M}(x_{n+1},r)$ contains no elements $y\neq z$ with $(B_{M}(y,s),y)\cong (B_{M}(z,s),z)$. We take $% r_{n+1}=r$ and $s_{n+1}=\sup (s_{n}+1,s)$. We consider the inductive limit $(N,x)$ of the pairs $% (B_{M}(x_{n},r_{n}+s_{n}),x_{n})$ relative to some isomorphisms $\theta _{n}:(B_{M}(x_{n},r_{n}+s_{n}),x_{n})\rightarrow (B_{M}(x_{n+1},r_{n}+s_{n}),x_{n+1})$  $\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \subset (B_{M}(x_{n+1},r_{n+1}+s_{n+1}),x_{n+1})$. As $M$ satisfies the local isomorphism property, $N$ is locally isomorphic to $M$. For $n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $ and $y\neq z$ in $B_{N}(x,r_{n})$, we have $(B_{N}(y,s_{n}),y)\ncong (B_{N}(z,s_{n}),z)$ since $B_{N}(y,s_{n})$ and $B_{N}(z,s_{n})$ are contained in$\ B_{N}(x,r_{n}+s_{n})$ and $(B_{N}(x,r_{n}+s_{n}),x)$ is isomorphic to $(B_{M}(x_{n},r_{n}+s_{n}),x_{n})$. For each nontrivial automorphism $\theta $ of $N$, there exist $n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $ and $y\neq z$ in $B_{N}(x,r_{n})$ such that $y\theta =z$. We have $% (B_{N}(y,s_{n}),y)\cong (B_{N}(z,s_{n}),z)$, whence a contradiction.  $% \blacksquare $ From now on, we consider the metric space $(E,\delta )$, the group $G$ and the set $\Delta $ defined in Section 1. As a consequence of Theorem 4.1, we have: **Corollary 4.2.** Let $\mathcal{T}$ be a $\Delta $-tiling which satisfies the local isomorphism property. Then $\mathcal{T}$ is locally isomorphic to a rigid $\Delta $-tiling if and only if, for each $% r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$, there exists $T\in \mathcal{T}$ such that no $\sigma \in G-\left\{ \mathrm{Id}\right\} $ with $\mathcal{T}\sigma =\mathcal{T}$ satisfies $T\sigma \in \mathcal{B}_{r}^{\mathcal{T}}(T)$. **Remark.** If $(E,\delta )$ is weakly homogeneous, then Corollary 4.2 and Proposition 1.2 imply that $\mathcal{T}$ is locally isomorphic to a rigid $\Delta $-tiling if and only if, for each $\alpha \in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{>0}$, there exists $x\in E$ such that no $\sigma \in G-\left\{ \mathrm{Id}% \right\} $ with $\mathcal{T}\sigma =\mathcal{T}$ satisfies $\delta (x,x\sigma )\leq \alpha $. **Proof of Corollary 4.2.** It follows from Theorem 1.3, Corollary 2.2 and Proposition 2.3 that $\mathcal{T}$ is locally isomorphic to a rigid $\Delta $-tiling if and only if the associated $\mathcal{L}% _{\Delta }$-structure is locally isomorphic to a connected rigid $\mathcal{L}% _{\Delta }$-structure, and therefore, by Theorem 4.1, if and only if, for each $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$, there exists $T\in \mathcal{T}$ such that $(\mathcal{T},U)\cong (% \mathcal{T},V)$ implies $U=V$ for any $U,V\in \mathcal{B}_{r}^{\mathcal{T}% }(T)$. Now we show that this property is true if and only if, for each $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$, there exists $T\in \mathcal{T}$ such that no $\sigma \in G-\left\{ \mathrm{Id}\right\} $ with $\mathcal{T}\sigma =\mathcal{T}$ satisfies $T\sigma \in \mathcal{B}_{r}^{\mathcal{T}}(T)$. For each $T\in \mathcal{T}$, each $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$ and each $\sigma \in G-\left\{ \mathrm{Id}\right\} $ such that $% \mathcal{T}\sigma =\mathcal{T}$, if there exist $U\neq V$ in $\mathcal{B}% _{r}^{\mathcal{T}}(T)$ such that $U\sigma =V$, then we have $T\sigma \in \mathcal{B}_{3r}^{\mathcal{T}}(T)$. Conversely, if $r\geq 2q$ for the integer $q$ of Section 1 and if $T\sigma \in \mathcal{B}_{r}^{\mathcal{T}% }(T) $, then we obtain $U\neq V$ in $\mathcal{B}_{r}^{\mathcal{T}}(T)$ such that $U\sigma =V$ as follows: we write $U=T$ and $V=T\sigma $ if $T\sigma \neq T$; otherwise, we consider any $U\in \mathcal{B}_{q}^{\mathcal{T}}(T)$such that $U\sigma \neq U$, and we write $V=U\sigma $.  $\blacksquare $ **Corollary 4.3.** Let $n\geq 1$ be an integer and let $% \mathcal{T}$ be a tiling of the euclidean space $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{n}$ which satisfies the local isomorphism property. Then $\mathcal{T}$ is locally isomorphic to a rigid tiling if and only if it is not invariant through a nontrivial translation. **Proof.** By the remark after Corollary 4.2, it suffices to show that, if $\mathcal{T}$ is not invariant through a nontrivial translation, then, for each $\alpha \in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{>0}$, there exists $x\in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{n}$ such that $\left\Vert xh-x\right\Vert >\alpha $ for each nontrivial isometry $h$ of $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{n}$ which stabilizes $\mathcal{T}$. This result follows from Theorem 4.4 below since the isometries of $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{n}$ which stabilize $\mathcal{T}$ form a discrete group by Proposition 1.8.  $\blacksquare $ **Theorem 4.4.** Let $n\geq 1$ be an integer and let $H$ be a discrete group of isometries of the euclidean space $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{n}$. Then $H$ contains no translation if and only if, for each $\alpha \in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{>0}$, there exists $x\in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{n}$ such that $\left\Vert xh-x\right\Vert >\alpha $ for each $h\in H-\left\{ \mathrm{Id}\right\} $. **Proof.** The if part is clear. For the only if part, we proceed by induction on $n$. We denote by $\mathrm{E}_{n}$ the group of all isometries of $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{n}$. For any $X,Y\subset %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{n}$, we write $\delta (X,Y)=\inf_{x\in X,y\in Y}\left\Vert y-x\right\Vert $. For each $f\in \mathrm{E}_{n}$, we consider the linear map $\overline{f}$associated to $f$, the largest affine subspace $W_{f}$ of $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{n}$ with $W_{f}f=W_{f}$ such that $f$ acts on $W_{f}$ as a translation, and the element $w_{f}\in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{n}$ such that $xf=x+w_{f}$ for each $x\in W_{f}$. We have $W_{f}=\{x\in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{n}\mid \left\Vert xf-x\right\Vert $ minimal$\}$. If $V_{f}$ is the vector subspace of $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{n}$ orthogonal to $W_{f}$ and maximal for that property, then the restriction of $\overline{f}$ to $V_{f}$ is an orthogonal transformation without nontrivial fixed point. Consequently, for each $\alpha \in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{>0}$, there exists $\beta \in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{>0}$ such that, for each $x\in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{n}$, $\delta (x,W_{f})>\beta $ implies $\left\Vert xf-x\right\Vert >\alpha $. It follows from a theorem of Bieberbach (see \[9, Theorem 1, p. 15\]) that $H$ has a normal subgroup $N$ with $N$ abelian and $H/N$ finite. As $N$ is finitely generated, the same properties remain true if we replace $N$ by $% N^{r}=\{h^{r}\mid h\in N\}$ for an integer $r\geq 2$. Consequently, we can suppose for the remainder of the proof that $N$ is torsion-free. Then we have $w_{f}\neq 0$ for each $f\in N-\left\{ \mathrm{Id}\right\} $ since $N$ is discrete like $H$. We observe that $W_{f}g=W_{f}$ for any $f,g\in \mathrm{E}_{n}$ which commute, and in particular for $f,g\in N$: We have $% (W_{f}g)f=(W_{f}f)g=W_{f}g$ and $(zg)f-(yg)f=(zf)g-(yf)g=(zf-yf)\overline{g}=(z-y)\overline{g}% =zg-yg$ for any $y,z\in W_{f}$. It follows that $f$ stabilizes $W_{f}g$ and acts on $W_{f}g$ as a translation, which implies $W_{f}g=W_{f}$. We consider $W=\cap _{f\in N}W_{f}$. In order to prove that $W$ is nonempty, it suffices to show that, for each $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$ and any $f_{1},...,f_{r+1}\in N$, if $W_{r}=W_{f_{1}}\cap ...\cap W_{f_{r}}$ is nonempty, then $W_{r+1}=W_{f_{1}}\cap ...\cap W_{f_{r+1}}$is nonempty. But we have $W_{r}f_{r+1}=W_{r}$ since $% W_{f_{i}}f_{r+1}=W_{f_{i}}$ for $1\leq i\leq r$; it follows that $W_{r+1}$ is the largest affine subspace $V$ of $W_{r}$ with $Vf_{r+1}=V$ such that $% f_{r+1}$ acts on $V$ as a translation. Now we show that $Wh=W$ for each $h\in H$. For each $f\in N$, we have $% W(hfh^{-1})=W$ since $hfh^{-1}\in N$, and therefore $Whf=W(hfh^{-1})h=Wh$. Moreover, for each $f\in N$ and any $y,z\in W$, we have $zhf-yhf=z(hfh^{-1})h-y(hfh^{-1})h=[z(hfh^{-1})-y(hfh^{-1})]% \overline{h}$ $=(z-y)\overline{h}=zh-yh$ since $hfh^{-1}\in N$. It follows that each $f\in N$ stabilizes $% Wh $ and acts on $Wh$ as a translation, which implies $Wh=W$. Now we fix $\alpha \in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{>0}$ and we prove that there exists $x\in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{n}$ such that $\left\Vert xh-x\right\Vert >\alpha $ for each $h\in H-\left\{ \mathrm{Id}\right\} $. We consider the set $\Omega $ of all affine subspaces of $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{n}$ which are orthogonal to $W$ and maximal for that property. For each $% U\in \Omega $ and each $h\in H$, we have $Uh\in \Omega $ since $Wh=W$. First we show that $\{g\in H\mid \delta (U,Ug)\leq \alpha \}$ is finite for each $U\in \Omega $. As each $h\in N$ is acting on $W$ as a translation of vector $w_{h}$, we have $\delta (V,Vh)=\left\Vert w_{h}\right\Vert $ for $% V\in \Omega $ and $h\in N$, and $w_{gh}=w_{g}+w_{h}$ for $g,h\in N$. We write $\gamma =\inf_{h\in N-\left\{ \mathrm{Id}\right\} }\left\Vert w_{h}\right\Vert $. We have $\gamma \neq 0$ since $N$ is discrete and torsion-free. We consider $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$ such that $r\gamma >2\alpha $. For each $U\in \Omega $, each $g\in H$ such that $\delta (U,Ug)\leq \alpha $ and each $h\in N-\left\{ \mathrm{% Id}\right\} $, the inequality $\delta (Ug,Ugh^{r})=\left\Vert w_{h^{r}}\right\Vert =r\left\Vert w_{h}\right\Vert \geq r\gamma >2\alpha $implies $\delta (U,Ugh^{r})>\alpha $. As $N^{r}$ has finite index in $N$and therefore in $H$, it follows that $\{g\in H\mid \delta (U,Ug)\leq \alpha \}$ is finite. Now we consider $K=\{h\in H\mid xh-x\in W$ for each $x\in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{n}\}$ and, for each $h\in K$, the restriction $h_{W}$ of $h$ to $W$. For each $x\in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{n}$, we denote by $x_{W}$ the projection of $x$ on $W$. Then $% K_{W}=\{h_{W}\mid h\in K\}$ is, like $K$, a discrete group of isometrieswithout nontrivial translation, since $xh-x=x_{W}h-x_{W}$ for each $h\in K$ and each $x\in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{n}$. The induction hypothesis applied to $W$ and $K_{W}$ implies that there exists $x\in W$ such that $\left\Vert xh_{W}-x\right\Vert >\alpha $ for each $h\in K-\left\{ \mathrm{Id}\right\} $. We consider the unique $U\in \Omega $ such that $x\in U$. We have $\delta (U,Uh)=\left\Vert xh_{W}-x\right\Vert >\alpha $ for each $h\in K-\left\{ \mathrm{Id}\right\} $. For each $h\in H-K$, we have $U\cap W_{h}\varsubsetneq U$ and $S_{h}=\{y\in U\mid \left\Vert yh-y\right\Vert \leq \alpha \}$ is contained in $A+(U\cap W_{h})$ for a bounded subset $A$ of $U$ since there exists $\beta \in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{>0}$ such that, for each $x\in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{n}$, $\delta (x,W_{h})>\beta $ implies $\left\Vert xh-x\right\Vert >\alpha $. As $S_{h}$ is empty for each $h\in H$ such that $\delta (U,Uh)>\alpha $, there exist finitely many nonempty subsets $S_{h}$, and their union cannot cover $U$.  $\blacksquare $ Now, we illustrate Corollary 4.2 and Corollary 4.3 with three examples related to aperiodicity. Several different definitions have been given for that notion (see \[4, p. 4\]). We consider the system $\Delta $ defined in Section 1 and the set $\mathcal{C% }$ of all $\Delta $-tilings which satisfy a set $\Omega $ of *local rules*, each of them saying which configurations of some given size can appear in a tiling belonging to $\mathcal{C}$. According to \[12, p. 208\], we say that $\Omega $ is *strong* if the $\Delta $-tilings in $\mathcal{% C}$ satisfy the local isomorphism property and if they are not invariant through any nontrivial translation. We use the classical definition of translation for the euclidean spaces $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{n}$, and the definition given in Section 3 for the general case. Here we do not suppose $\Omega $ finite. One reason is that some natural sets of tilings are defined by strong infinite sets of local rules (for instance we showed this property in \[8\] for the set of all complete folding sequences, and for the set of all coverings of the plane by complete folding curves which satisfy the local isomorphism property). Another reason is that, for each $\Delta $-tiling $\mathcal{T}$ which satisfies the local isomorphism property, the set of all $\Delta $-tilings which are locally isomorphic to $\mathcal{T}$ is defined by a set of local rules which can be finite as in Examples 2 and 3, or infinite as in Example 1. By Corollary 4.3, any tiling of an euclidean space of finite dimension which satisfies a strong set of local rules is locally isomorphic to a rigid tiling if and only if it is not invariant through a nontrivial translation. We do not know presently if this result can be generalized with the notion of translation that we consider. In Examples 1 and 3 below, the group $G$ consists of all isometries of $E$; in Example 2 we only consider positive isometries of $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{3}$. In each example, all the structures are uniformly locally finite and satisfy the local isomorphism property. On the other hand, they do not satisfy (P). Some of them are rigid and others have nontrivial automorphisms, but all of them are locally isomorphic. **Example 1.** For each $r\in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion -% %TCIMACRO{\U{211a} }% %BeginExpansion \mathbb{Q} %EndExpansion $ and each $s\in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion $, we consider the line $L(r,s)$ of equation $y=rx+s$. We write $\Omega (r,s)=\{(a,b)\in %TCIMACRO{\U{2124} }% %BeginExpansion \mathbb{Z} %EndExpansion \times %TCIMACRO{\U{2124} }% %BeginExpansion \mathbb{Z} %EndExpansion \mid L(r,s)\cap ([a-1/2,a+1/2[\times \lbrack b-1/2,b+1/2[)\neq \varnothing \} $. For each $a\in %TCIMACRO{\U{2124} }% %BeginExpansion \mathbb{Z} %EndExpansion $, $\left\{ a\right\} \times %TCIMACRO{\U{2124} }% %BeginExpansion \mathbb{Z} %EndExpansion $ contains $n+1$ or $n+2$ points of $\Omega (r,s)$, where $n$ is the integral part of $\left\vert r\right\vert $. We colour the point $a$ in white if $\left\{ a\right\} \times %TCIMACRO{\U{2124} }% %BeginExpansion \mathbb{Z} %EndExpansion $ contains $n+1$ points of $\Omega (r,s)$ and in black otherwise. We consider the tiling $\mathcal{T}(r,s)$ of the euclidean space $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion $ which consists of the segments $[a,a+1]$ for $a\in %TCIMACRO{\U{2124} }% %BeginExpansion \mathbb{Z} %EndExpansion $ with their endpoints coloured in white or black as above. For each $r\in %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion -% %TCIMACRO{\U{211a} }% %BeginExpansion \mathbb{Q} %EndExpansion $, the tilings $\mathcal{T}(r,s)$ are locally isomorphic and each of them satisfies the local isomorphism property. They do not satisfy (P) since they are not invariant through any nontrivial translation. They are invariant through a unique symmetry if there exist $a,b\in %TCIMACRO{\U{2124} }% %BeginExpansion \mathbb{Z} %EndExpansion $ such that $(a,b)$ or $(a+1/2,b)$ or $(a,b+1/2)$ belongs to $L(r,s)$, and rigid otherwise. The three possibilities above are respectively realized for $s\in %TCIMACRO{\U{2124} }% %BeginExpansion \mathbb{Z} %EndExpansion +r% %TCIMACRO{\U{2124} }% %BeginExpansion \mathbb{Z} %EndExpansion $, $s\in r/2+% %TCIMACRO{\U{2124} }% %BeginExpansion \mathbb{Z} %EndExpansion +r% %TCIMACRO{\U{2124} }% %BeginExpansion \mathbb{Z} %EndExpansion $, $s\in 1/2+% %TCIMACRO{\U{2124} }% %BeginExpansion \mathbb{Z} %EndExpansion +r% %TCIMACRO{\U{2124} }% %BeginExpansion \mathbb{Z} %EndExpansion $. **Example 2.** Let $(E,\delta )$ be the euclidean space $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{3}$ and let $G$ consist of the positive isometries. Let *T* be a tiling of $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{3}$ which satisfies the local isomorphism property. Suppose that the group of isometries which leave $\mathcal{T}$ globally invariant is generated by a “screwing motion” $\sigma $, which is the composition of a translation with a rotation about an axis parallel to the translation. If the angle of the rotation belongs to $\pi %TCIMACRO{\U{211a} }% %BeginExpansion \mathbb{Q} %EndExpansion $, then some nontrivial power of $\sigma $ is a translation, and $\mathcal{% T}$ satisfies (P). Otherwise, $\mathcal{T}$ does not satisfy (P), and Theorem 4.1 implies that $\mathcal{T}$ is locally isomorphic to a rigid tiling. According to \[12, Section 7.2, pp. 208-213\], examples of that situation have been given by Danzer for tilings obtained from $1$ prototile (the examples with $n$ odd must be considered). **Example 3.** In 1979, R. Penrose gave his famous example (see \[3\]) of two polygonal prototiles, the arrow and the kite, which define an aperiodic class of tilings of the euclidean space $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{2}$. There exist $2^{\omega }$ Penrose tilings. All of them are locally isomorphic and each of them satisfies the local isomorphism property. The Robinson tilings (see \[10\]) have the same properties, but they are constructed from a larger set of prototiles. Penrose asked if there exist classes of tilings of $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ^{2}$ defined from a single prototile which have these properties. The question is apparently still open for tilings with non-overlapping tiles (see \[6\]). In the hyperbolic plane, it is not difficult to construct such an example. Here, we use the representation of the hyperbolic plane by the Poincaré half-plane $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion \times %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{>0}$. Figure 1 illustrates the construction of the tilings in our example, which is a particular case of those given in \[6\]. We denote by $\Omega $ the set of all tilings constructed in that way. For any $\mathcal{S},\mathcal{T}\in \Omega $, each $S\in \mathcal{S}$ and each $T\in \mathcal{T}$, there exists a unique $\sigma \in G$ such that $S\sigma =T$, because of the arrows on the edges of the tiles. For each $\mathcal{T}\in \Omega $ and each $T\in \mathcal{T}$, we consider $% (U_{n}(T))_{n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }$ where $U_{0}(T)=T$ and, for each $n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $, $U_{n+1}(T)$ is the tile just above $U_{n}(T)$. For each $n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$, we write $a_{n}(T)=0$ if $U_{n}(T)$ is at the left of $U_{n-1}(T)$, and $a_{n}(T)=1$ otherwise. For any $\mathcal{S},\mathcal{T}\in \Omega $, each $S\in \mathcal{S}$ and each $T\in \mathcal{T}$, we have $(\mathcal{S},S)\cong (\mathcal{T},T)$ if and only if $(a_{n}(S))_{n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }}=(a_{n}(T))_{n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }}$. For each $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$, we have $(\mathcal{B}_{r}^{\mathcal{S}}(S),S)\cong (\mathcal{B}% _{r}^{\mathcal{T}}(T),T)$ if and only if $% (a_{1}(S),...,a_{r}(S))=(a_{1}(T),...,a_{r}(T))$. For any $\mathcal{S},\mathcal{T}\in \Omega $, each $S\in \mathcal{S}$, each $% T\in \mathcal{T}$ and each $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$, there exists $S^{\prime }\in \mathcal{T}$ such that $% T=U_{r}(S^{\prime })$ and $(a_{1}(S^{\prime }),...,a_{r}(S^{\prime }))=(a_{1}(S),...,a_{r}(S))$, which implies $S^{\prime }\in \mathcal{B}_{r}^{% \mathcal{T}}(T)$ and $(\mathcal{B}_{r}^{\mathcal{S}}(S),S)\cong (\mathcal{B}% _{r}^{\mathcal{T}}(S^{\prime }),S^{\prime })$. Consequently, any tiling in $% \Omega $ satisfies the local isomorphism property, and any two such tilings are locally isomorphic. For each $\mathcal{T}\in \Omega $ and any $S,T\in \mathcal{T}$, there exist $% i,j\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $ such that $U_{i}(S)=U_{j}(T)$; for $n\geq i+1$, we have $% a_{n}(S)=a_{n+k}(T)$ where $k=j-i$. Consequently, each $\mathcal{T}\in \Omega $ only realizes countably many sequences $(a_{n})_{n\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }}\in \{0,1\}^{% %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }}$. As each such sequence is realized by a tiling $\mathcal{T}\in \Omega $, it follows that $\Omega $ is the union of $2^{\omega }$isomorphism classes. This property is a particular case of Corollary 2.6 above. Now we show that, for each $\mathcal{T}\in \Omega $ which is not rigid and each $T\in \mathcal{T}$, there exists $k\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$ such that $a_{n}(T)=a_{n+k}(T)$ for $n$ large enough: We consider $\sigma \in G-\{\mathrm{Id}\}$ such that $\mathcal{T}\sigma =\mathcal{T}$ and $i,j\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $ such that $U_{i}(T)=U_{j}(T\sigma )$. We have $i\neq j$ because $% U_{i}(T)=U_{i}(T\sigma )=U_{i}(T)\sigma $ would imply $\sigma =\mathrm{Id}$. For $k=j-i$, we have $U_{n}(T)=U_{n+k}(T\sigma )$ for $n\geq i$, and therefore $a_{n}(T)=a_{n+k}(T\sigma )=a_{n+k}(T)$ for $n\geq i+1$. Conversely, for each $\mathcal{T}\in \Omega $ and each $T\in \mathcal{T}$, if $I=\{k\in %TCIMACRO{\U{2124} }% %BeginExpansion \mathbb{Z} %EndExpansion \mid a_{n}(T)=a_{n+k}(T)$ for $n$ large enough$\}$ contains a nonzero element, then $I$ is the ideal of $% %TCIMACRO{\U{2124} }% %BeginExpansion \mathbb{Z} %EndExpansion $ generated by the smallest $h\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$ which belongs to $I$. For each $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $ such that $a_{n}(T)=a_{n+h}(T)$ for $n>r$, the isometry which sends $% U_{r}(T)$ to $U_{r+h}(T)$ generates $\{\sigma \in G\mid \mathcal{T}\sigma =% \mathcal{T}\}$. **Remark.** In Example 3, similar to the case of Penrose tilings or Robinson tilings, the class of $\Delta $-tilings that we consider is defined by a local rule which describes the possible configurations of the immediate neighbours of a tile. In the case of Penrose tilings or Robinson tilings (see \[7, p. 125\]), it follows that there exists a local rule expressed by one sentence which characterizes among the connected $% \mathcal{L}_{\Delta }$-structures those which are associated to $\Delta $-tilings, because no such tiling is invariant through an infinite group of isometries. On the other hand, in Example 3, no such rule exists since any local rule satisfied by the $\mathcal{L}_{\Delta }$-structures associated to $\Delta $-tilings is also satisfied by some of their quotients. The following example generalizes the argument of Example 3, even though the relational structures that we consider are not represented by tilings: **Example 4.** We write $\mathcal{L}=\{P_{1},...,P_{k}\}$ where $P_{1},...,P_{k}$ are unary functional symbols, and we consider the nonempty $\mathcal{L}$-structures $M$ which satisfy the following properties: 1\) For each $x\in M$, there exists one and only one pair $(i,y)$ with $1\leq i\leq k$ and $y\in M$ such that $yP_{i}=x$; 2\) $xP_{i_{1}}...P_{i_{r}}=x$ implies $r=0$ for $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $, $1\leq i_{1},...,i_{r}\leq k$ and $x\in M$. Each connected such structure induces a directed tree where the pairs $(x,y)$ of consecutive vertices are characterized by the existence of a unique $i\in \{1,...,k\}$ such that $xP_{i}=y$; each vertex is the origin of $k$ edges. In order to apply the results of the present paper, it is convenient to consider $P_{1},...,P_{k}$ as binary relations. Similarly to Example 3, the nonempty $\mathcal{L}$-structures which satisfy 1) and 2) are locally isomorphic, and each of them satisfies the local isomorphism property. In fact, for any such structures $M,N$, each $x\in M$, each $y\in N$ and each $% r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion ^{\ast }$, we have $(B_{M}(x,r),x)\cong (B_{M}(z,r),z)$ for $% z=yP_{i_{r}}...P_{i_{1}}$ where $i_{1},...,i_{r}$ are the elements of $% \{1,...,k\}$ such that $xP_{i_{1}}^{-1}...P_{i_{r}}^{-1}$ exists. Now, for each nonempty connected $\mathcal{L}$-structure $M$ which satisfies 1), 2) and each $x\in M$, we consider the sequence $(i_{r}(x))_{r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }\in \{1,...,k\}^{% %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion }$ such that $xP_{i_{1}(x)}^{-1}...P_{i_{r}(x)}^{-1}$ exists for each $% r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $. Similarly to Example 3, $M$ has a nontrivial automorphism if and only if there exists an integer $k$ such that $i_{r}(x)=i_{r+k}(x)$ for $r$ large enough. In that case, the group of automorphisms of $M$ is infinite cyclic. By Theorem 4.1, it follows that the nonempty $\mathcal{L}$-structures which satisfy 1) and 2) are rigid. The last two examples are not related to tilings. They are given in order to illustrate the importance of each hypothesis in Theorem 4.1. **Example 5.** The* Cayley graph* of a group $G$ relative to a generating family $(x_{i})_{i\in I}$ is the relational structure $M$ defined on $G$ as follows: for $i\in I$ and $y,z\in G$, we write $R_{i}(y,z)$ if and only if $z=yx_{i}$. The structure $M$ is uniformly locally finite if $I$ is finite. The automorphisms of $M$ are the maps $% y\rightarrow gy$ for $g\in G$. For any $y,z\in M$, we have $(M,y)\cong (M,z)$ since there exists $g\in G$ such that $gy=z$. In particular, $M$ satisfies the local isomorphism property and $M$ is not locally isomorphic to a rigid structure. If $G$ is freely generated by the elements $x_{i}$, then $M$ is not invariant through any nontrivial translation since, for each $g\in G$ and each $r\in %TCIMACRO{\U{2115} }% %BeginExpansion \mathbb{N} %EndExpansion $, there exists $x\in M$ such that $gx\notin B_{M}(x,r)$. **Example 6.** Here, the language $\mathcal{L}$ consists of one binary relational symbol. The $\mathcal{L}$-structure $M$ shown by Figure 2 is uniformly locally finite, but it does not satisfy the local isomorphism property. The only automorphisms of $M$ are the maps $% x_{i,j}\rightarrow x_{i+k,j}$ for $k\in %TCIMACRO{\U{2124} }% %BeginExpansion \mathbb{Z} %EndExpansion $. Consequently, $M$ does not satisfy the characterization of Theorem 4.1. Anyway, each connected $\mathcal{L}$-structure $N$ locally isomorphic to $M$ is isomorphic to $M$, and therefore not rigid. **References** 1\] A. Ballier, B. Durand and E. Jeandel, Structural aspects of tilings, in STACS 2008 (25th Annual Symposium on Theoretical Aspects of Computer Science, Bordeaux, France, 2008), Dagstuhl Seminar Proceedings, Internationales Begegnungs und Forschungszentrum fuer Informatik, Schloss Dagstuhl, Germany, pp. 61-72. 2\] Y. Benyamini and J. Lindenstrauss, Geometric nonlinear functional analysis I, AMS Colloquium Publications 48, American Mathematical Society, Providence, USA, 2000. 3\] M. Gardner, Extraordinary nonperiodic tiling that enriches the theory of tiles, Scientific American, Jan. 1977, pp. 110-121. 4\] C. Goodman-Strauss, Open questions in tiling, notes available at: comp.uark.edu/cgstraus/papers. 5\] W. Hodges, Model Theory, Encyclopedia of Mathematics and its Applications, Cambridge University Press, Cambridge, 1995. 6\] G.A. Margulis and S. Mozes, Aperiodic tilings of the hyperbolic plane by convex polygons, Israel J. Math. 107 (1998), 319-325. 7\] F. Oger, Algebraic and model-theoretic properties of tilings, Theoret. Comput. Sci. 319 (2004), 103-126. 8\] F. Oger, Paperfolding sequences, paperfolding curves and local isomorphism, to appear. 9\] R.K. Oliver, On Bieberbach’s analysis of discrete euclidean groups, Proc. Amer. Math. Soc. 80 (1980), 15-21. 10\] B. Poizat, Une théorie finiment axiomatisable et superstable, Groupe d’Etude de Théories Stables 3, Université Pierre et Marie Curie, Paris, 1983. 11\] C. Radin and M. Wolff, Space tiling and local isomorphism, Geometriae Dedicata 42 (1992), 355-360. 12\] M. Senechal, Quasicrystals and Geometry, Cambridge University Press, Cambridge, GB, 1996.

ArXiv

--- abstract: 'Low-temperature thermodynamics of the classical frustrated ferromagnetic spin chain is studied. Using transfer-matrix method we found the behavior of the correlation function and zero-field susceptibility at the ferromagnetic-helical transition point. It is shown that the critical exponent for the susceptibility is changed from 2 to 4/3 at the transition point.' author: - 'D. V. Dmitriev' - 'V. Ya. Krivnov' title: 'Thermodynamics of classical frustrated spin chain at the ferromagnet-helimagnet transition point' --- Lately, there has been considerable interest in low-dimensional spin models that exhibit frustration. One of them is the spin chain with the ferromagnetic interaction $J_{1}$ of nearest neighbor (NN) spins and the antiferromagnetic next-nearest-neighbor (NNN) interaction $J_{2}$, so called the 1D F-AF model. Its Hamiltonian has a form $$H=J_{1}\sum \mathbf{S}_{n}\mathbf{\cdot S}_{n+1}+J_{2}\sum \mathbf{S}_{n}% \mathbf{\cdot S}_{n+2} \label{H}$$where $J_{1}<0$ and $J_{2}>0$. This model is characterized by a frustration parameter $\alpha =J_{2}/|J_{1}| $. The ground state properties of the quantum $s=1/2$ F-AF chain have been intensively studied last years [@Chubukov; @DK06; @HM; @Hikihara]. It is known that the ground state of the model is ferromagnetic for $\alpha <1/4$. At $\alpha =1/4$ the ground state phase transition to the incommensurate singlet phase with helical spin correlations takes place. Remarkably, this transition point does not depend on a spin value, including the classical limit $s=\infty $. Interesting question is the influence of the frustration on the low-temperature thermodynamics of the model especially near the transition point $\alpha =1/4$. We study this problem for the classical version of model (\[H\]). At zero temperature the classical model has long range-order (LRO) for all values of $\alpha $: the ferromagnetic LRO at $% \alpha \leq 1/4$ and the helical one at $\alpha >1/4$. At finite temperature the LRO is destroyed by thermal fluctuations and thermodynamic quantities have a singular behavior at $T\rightarrow 0$. In particular, the zero-field susceptibility $\chi $ diverges. For the 1D Heisenberg ferromagnet ($\alpha =0$) $\chi =2\left\vert J_{1}\right\vert /3T^{2}$ [@Fisher]. At $% 0<\alpha <1/4$ the susceptibility is $\chi =2(1-4\alpha )\left\vert J_{1}\right\vert /3T^{2}$. This behavior of $\chi $ is similar to that for the quantum $s=1/2$ F-AF model [@Hartel]. The prefactor in $\chi $ vanishes at the transition point indicating the change of the critical exponent. We focus our attention on the behavior of $\chi $ at the transition point. The partition function $Z$ of model (\[H\]) at $\alpha =1/4$ is $$Z=\prod_{n=1}^{N}\int d\Omega _{n}\exp \left\{ \frac{1}{T}\sum (\vec{S}% _{n}\cdot \vec{S}_{n+1}-\frac{1}{4}\vec{S}_{n}\cdot \vec{S}_{n+2})\right\} \label{Z1}$$where $\vec{S}_{n}$ is unit vector, $d\Omega _{n}$ is the volume element of the solid angle for $n$-th site, we put $\left\vert J_{1}\right\vert =1$ and the periodic boundary conditions are proposed. Our further calculations are based on the transfer matrix method and we use a version of this method adapted to the model with NNN interactions by Harada and Mikeska in [@Harada-Mikeska]. Following Ref.[@Harada-Mikeska] we represent $Z$ in a form $$Z=\prod_{n=1}^{N}\int d\Omega _{n}K(\theta _{n-1},\theta _{n};\varphi _{n}) \label{Z2}$$where $$K(\theta _{n-1},\theta _{n};\varphi _{n})=\exp \left( \frac{\cos \theta _{n-1}+\cos \theta _{n}}{2T}-\frac{(\cos \theta _{n-1}\cos \theta _{n}+\sin \theta _{n-1}\sin \theta _{n}\cos \varphi _{n})}{4T}\right) \label{K}$$where $\theta _{n}$ is the angle between $\vec{S}_{n}$ and $\vec{S}_{n+1}$ and $\varphi _{n}$ is the angle between components of $\vec{S}_{n-1}$ and $% \vec{S}_{n+1}$ projected onto $(X_{n},Y_{n})$ plane of the $n$-th local coordinate system with the $Z_{n}$ axis parallel to $\vec{S}_{n}$. Integrating Eq.(\[K\]) over $\varphi _{n}$ we obtain $Z$ in a form $$Z=\prod_{n}\int_{0}^{\pi }d\theta _{n}\sin \theta _{n}A(\theta _{n-1},\theta _{n}) \label{Z3}$$where $$A(\theta _{n-1},\theta _{n})=\frac{1}{2}I_{0}(-z)\exp \left( \frac{\cos \theta _{n-1}+\cos \theta _{n}}{2T}-\frac{\cos \theta _{n-1}\cos \theta _{n}% }{4T}-\frac{3}{4T}\right) \label{A1}$$and $I_{0}(-z)$ is the modified Bessel function of$$z=\frac{\sin \theta _{n-1}\sin \theta _{n}}{4T} \label{z}$$ Let us consider an integral equation $$\int_{0}^{\pi }A(\theta _{1},\theta _{2})\psi _{\alpha }(\theta _{2})\sin \left( \theta _{2}\right) d\theta _{2}=\lambda _{\alpha }\psi _{\alpha }(\theta _{1}) \label{I1}$$where $\psi _{\alpha }(\theta )$ satisfy normalization condition $$\int_{0}^{\pi }\psi _{\alpha }(\theta )\psi _{\beta }(\theta )\sin \theta \mathrm{d}\theta =\delta _{\alpha ,\beta } \label{norma}$$ Eigenfunctions $\psi _{\alpha }$ and eigenvalues $\lambda _{\alpha }$ can be chosen as real, since the kernel $A(\theta _{1},\theta _{2})$ is real and symmetric. Then, $$A(\theta _{1},\theta _{2})=\sum_{\alpha }\lambda _{\alpha }\psi _{\alpha }(\theta _{1})\psi _{\alpha }(\theta _{2}) \label{A2}$$ Substituting Eq.(\[A2\]) into Eq/(\[Z3\]) we obtain in the thermodynamic limit $$Z=\lambda _{0}^{N} \label{Z4}$$where $\lambda _{0}$ is the largest eigenvalue of Eq.(\[I1\]). In the low-temperature limit the angles $\theta _{n}$ are small and we can use the asymptotic expansion of the modified Bessel function $$I_{0}(-z)=\frac{e^{z}}{\sqrt{2\pi z}}\left( 1+\frac{1}{8z}+O(z^{-2})\right) \label{I0z}$$ Then, we expand the expression in the exponent of the transfer matrix to the fourth order in $\theta _{i}$ to obtain $$A(\theta _{1},\theta _{2})=\sqrt{\frac{T}{2\pi \theta _{1}\theta _{2}}}% \left( 1+\frac{T}{2\theta _{1}\theta _{2}}\right) \exp \left( -\frac{(\theta _{1}-\theta _{2})^{2}}{8T}-\frac{\theta _{1}^{2}\theta _{2}^{2}}{8T}+\frac{% (\theta _{1}-\theta _{2})^{4}}{96T}\right) \label{A3}$$ We can neglect the term $(\theta _{1}-\theta _{2})^{4}/96T$ as will be seen below. As a result integral equation (\[I1\]) reduces to $$\int_{0}^{\pi }\sqrt{\frac{T\theta _{2}}{2\pi \theta _{1}}}\left( 1+\frac{T}{% 2\theta _{1}\theta _{2}}\right) e^{-\frac{(\theta _{1}-\theta _{2})^{2}}{8T}-% \frac{\theta _{1}^{2}\theta _{2}^{2}}{8T}}\psi _{\alpha }(\theta _{2})d\theta _{2}=\lambda _{\alpha }\psi _{\alpha }(\theta _{1}) \label{I2}$$ The maximum of the expression in the exponent (saddle point) is at $\theta _{2}=\theta _{1}$ (more exactly $\theta _{2}=\theta _{1}-\theta _{1}^{3}+\ldots $, but it suffices to put $\theta _{2}=\theta _{1}$). Near this saddle point we expand $\psi _{\alpha }(\theta _{2})$ as follows $$\psi _{\alpha }(\theta _{2})=\psi _{\alpha }(\theta _{1})+(\theta _{2}-\theta _{1})\psi _{\alpha }^{\prime }(\theta _{1})+\frac{(\theta _{2}-\theta _{1})^{2}}{2}\psi _{\alpha }^{\prime \prime }(\theta _{1})+\ldots \label{psi1}$$and $$\sqrt{\frac{\theta _{2}}{\theta _{1}}}=\sqrt{1+\frac{\theta _{2}-\theta _{1}% }{\theta _{1}}}=1+\frac{\theta _{2}-\theta _{1}}{2\theta _{1}}-\frac{(\theta _{2}-\theta _{1})^{2}}{8\theta _{1}^{2}}+\ldots \label{q}$$ Let us introduce new scaled variables $$\begin{aligned} \theta _{2}-\theta _{1} &=&T^{1/2}x \nonumber \\ \theta _{1} &=&T^{1/3}r \label{xr}\end{aligned}$$ Now $\psi _{\alpha }(\theta )\rightarrow \psi _{\alpha }(r)$ and $$\psi _{\alpha }(\theta _{2})\rightarrow \psi _{\alpha }(r)+T^{1/6}x\psi _{\alpha }^{\prime }(r)+\frac{T^{1/3}x^{2}}{2}\psi _{\alpha }^{\prime \prime }(r)+\ldots \label{psi2}$$$$\sqrt{\frac{\theta _{2}}{\theta _{1}}}\rightarrow 1+\frac{T^{1/6}x}{2r}-% \frac{T^{1/3}x^{2}}{8r^{2}}+O(T^{1/2}) \label{qq}$$$$\exp \left( -\frac{(\theta _{1}-\theta _{2})^{2}}{8T}-\frac{\theta _{1}^{2}\theta _{2}^{2}}{8T}\right) \rightarrow \exp \left( -\frac{x^{2}}{8}-% \frac{T^{1/3}r^{4}}{8}\right) \label{qqq}$$ Summarizing all above we arrive at $$\int_{-r/T^{1/6}}^{\pi /\sqrt{T}}\left( 1+\frac{T^{1/6}x}{2r}-\frac{% T^{1/3}x^{2}}{8r^{2}}\right) \left( 1+\frac{T^{1/3}}{2r^{2}}\right) \left( \psi _{\alpha }(r)+T^{1/6}x\psi _{\alpha }^{\prime }(r)+\frac{T^{1/3}x^{2}}{2% }\psi _{\alpha }^{\prime \prime }(r)\right) e^{-x^{2}/8-T^{1/3}r^{4}/8}\frac{% Tdx}{\sqrt{2\pi }}=\lambda _{\alpha }\psi _{\alpha }(r) \label{I3}$$ At $T\rightarrow 0$, we can change the limits in the integral to $[-\infty ,\infty ]$, then only even powers in $x$ gives contribution, so taking into account only terms up to$~T^{1/3}$ we obtain $$\int_{-\infty }^{\infty }\left[ \left( 1-\frac{T^{1/3}r^{4}}{8}+\frac{T^{1/3}% }{2r^{2}}\right) \psi _{\alpha }+\frac{T^{1/3}x^{2}}{2}\left( \psi _{\alpha }^{\prime \prime }+\frac{1}{r}\psi _{\alpha }^{\prime }(r)-\frac{1}{4r^{2}}% \psi _{\alpha }\right) \right] e^{-x^{2}/8}\frac{Tdx}{\sqrt{2\pi }}=\lambda _{\alpha }\psi _{\alpha } \label{I4}$$ After integration over $x$ we obtain a linear differential equation $$2T\left( 1-\frac{T^{1/3}r^{4}}{8}+\frac{T^{1/3}}{2r^{2}}\right) \psi _{\alpha }+4T^{4/3}\left( \psi _{\alpha }^{\prime \prime }+\frac{1}{r}\psi _{\alpha }^{\prime }(r)-\frac{1}{4r^{2}}\psi _{\alpha }\right) =\lambda _{\alpha }\psi _{\alpha } \label{eq1}$$and, finally, $$-\psi _{\alpha }^{\prime \prime }-\frac{1}{r}\psi _{\alpha }^{\prime }+\frac{% r^{4}}{16}\psi _{\alpha }=\varepsilon _{\alpha }\psi _{\alpha } \label{eq2}$$with $$\varepsilon _{\alpha }=\frac{2T-\lambda _{\alpha }}{4T^{4/3}} \label{e}$$ Thus, we have got a Schrödinger equation for a particle with $Z$ component of the angular momentum $l_{z}=0$ in 2D potential well $% U(r)=r^{4}/16$. Normalization condition for $\psi _{\alpha }(r)$ is $$\int_{0}^{\infty }\psi _{\alpha }(r)\psi _{\beta }(r)2\pi rdr=\delta _{\alpha ,\beta } \label{norma2}$$ Numerical solution of Eq.(\[eq2\]) gives the following lowest eigenvalues (corresponding to the largest $\lambda $): $$\varepsilon _{\alpha }=0.9305;3.78;7.44\ldots \label{e_n}$$ As was shown in Ref.[@Harada-Mikeska] the two-spin correlation function can be expressed by the following integral $$\left\langle \vec{S}_{1}\cdot \vec{S}_{1+n}\right\rangle =\frac{1}{\lambda _{0}^{n-1}}\int_{0}^{\pi }d\theta _{n}\sin \theta _{n}\prod_{l=1}^{n-1}d\theta _{l}\sin \theta _{l}\psi _{0}(\theta _{l})\psi _{0}(\theta _{n})\left( \begin{array}{cc} 0 & 1% \end{array}% \right) B(\theta _{1})H(\theta _{l},\theta _{l+1})B(\theta _{n})\left( \begin{array}{c} 0 \\ 1% \end{array}% \right) \label{cor1}$$where $$B(\theta )=\left( \begin{array}{cc} \cos \theta /2 & \sin \theta /2 \\ -\sin \theta /2 & \cos \theta /2% \end{array}% \right) \label{B}$$$$H(\theta _{1},\theta _{2})=B(\theta _{1})\left( \begin{array}{cc} -\widetilde{A}(\theta _{1},\theta _{2}) & 0 \\ 0 & A(\theta _{1},\theta _{2})% \end{array}% \right) B(\theta _{2}) \label{Hm}$$and $\widetilde{A}(\theta _{1},\theta _{2})$ is given by Eq.(\[A1\]) with $% I_{0}(-z)$ replaced by $I_{1}(-z)$. Using the asymptotic expansion of the Bessel function $$I_{1}(-z)=-\frac{e^{z}}{\sqrt{2\pi z}}\left( 1-\frac{3}{8z}+O(z^{-2})\right) \label{I1z}$$we obtain $$H(\theta _{1},\theta _{2})=A_{0}(\theta _{1},\theta _{2})\left( \begin{array}{cc} 1-\frac{3T}{2\theta _{1}\theta _{2}} & \frac{\theta _{1}+\theta _{2}}{2} \\ -\frac{\theta _{1}+\theta _{2}}{2} & 1+\frac{T}{2\theta _{1}\theta _{2}}% \end{array}% \right) \label{Hm2}$$where $$A_{0}(\theta _{1},\theta _{2})=\sqrt{\frac{T}{2\pi \theta _{1}\theta _{2}}}% \exp \left( -\frac{(\theta _{1}-\theta _{2})^{2}}{8T}-\frac{\theta _{1}^{2}\theta _{2}^{2}}{8T}\right) \label{A0}$$ The matrix $H(\theta _{1},\theta _{2})$ is not symmetric. Therefore, to calculate $\left\langle \vec{S}_{1}\cdot \vec{S}_{1+n}\right\rangle $ it is necessary to solve a pair of the integral equations$$\begin{aligned} \int_{0}^{\pi }H(\theta _{1},\theta _{2})\vec{u}_{\alpha }(\theta _{2})\sin \left( \theta _{2}\right) d\theta _{2} &=&\eta _{\alpha }\vec{u}_{\alpha }(\theta _{1}) \label{in1} \\ \int_{0}^{\pi }H^{T}(\theta _{1},\theta _{2})\vec{v}_{\alpha }(\theta _{2})\sin \left( \theta _{2}\right) d\theta _{2} &=&\eta _{\alpha }\vec{v}% _{\alpha }(\theta _{1}) \label{in2}\end{aligned}$$where $H^{T}(\theta _{1},\theta _{2})$ is transposed matrix $H(\theta _{1},\theta _{2})$ and two-component vectors $\vec{u}_{\alpha }$ and $\vec{v}% _{\alpha }$ $$\vec{u}_{\alpha }=\left( \begin{array}{c} u_{1,\alpha } \\ u_{2,\alpha }% \end{array}% \right) ,\qquad \vec{v}_{\alpha }=\left( \begin{array}{c} v_{1,\alpha } \\ v_{2,\alpha }% \end{array}% \right) \label{uv}$$satisfy orthonormality relations, $$\int_{0}^{\pi }\vec{u}_{\alpha }^{T}(\theta )\vec{v}_{\beta }(\theta )\sin \left( \theta \right) d\theta =\int_{0}^{\pi }\vec{v}_{\alpha }^{T}(\theta )% \vec{u}_{\beta }(\theta )\sin \left( \theta \right) d\theta =\delta _{\alpha ,\beta } \label{norma_uv}$$ Then, the matrix $H(\theta _{1},\theta _{2})$ can be represented as $$H(\theta _{1},\theta _{2})=\sum_{\alpha }\eta _{\alpha }\vec{u}_{\alpha }(\theta _{1})\vec{v}_{\alpha }^{T}(\theta _{2}) \label{Hm3}$$ At small$\ \theta _{1}$, $\theta _{2}$ Eqs.(\[in1\]) and (\[in2\]) reduce to $$\begin{aligned} \int\limits_{0}^{\pi }A_{0}(\theta _{1},\theta _{2})\left[ \left( 1-\frac{3T% }{2\theta _{1}\theta _{2}}\right) u_{1,\alpha }(\theta _{2})+\theta _{1}u_{2,\alpha }(\theta _{2})\right] \sin \left( \theta _{2}\right) d\theta _{2} &=&\eta _{\alpha }u_{1,\alpha }(\theta _{1}) \label{in3} \\ \int\limits_{0}^{\pi }A_{0}(\theta _{1},\theta _{2})\left[ -\theta _{1}u_{1,\alpha }(\theta _{2})+\left( 1+\frac{T}{2\theta _{1}\theta _{2}}% \right) u_{2,\alpha }(\theta _{2})\right] \sin \left( \theta _{2}\right) d\theta _{2} &=&\eta _{\alpha }u_{2,\alpha }(\theta _{1}) \label{in4}\end{aligned}$$ Integrating these equations near the saddle point similar to Eqs.(\[I3\]),(\[I4\]), we get a pair of linear differential equation $$\begin{aligned} 2T\left( 1-\frac{T^{1/3}r^{4}}{8}-\frac{3T^{1/3}}{2r^{2}}\right) u_{1,\alpha }+4T^{4/3}\left( u_{1,\alpha }^{\prime \prime }+\frac{1}{r}u_{1,\alpha }^{\prime }-\frac{1}{4r^{2}}u_{1,\alpha }\right) +2T^{4/3}ru_{2,\alpha } &=&\eta _{\alpha }u_{1,\alpha } \label{eq_u1} \\ 2T\left( 1-\frac{T^{1/3}r^{4}}{8}+\frac{T^{1/3}}{2r^{2}}\right) u_{2,\alpha }+4T^{4/3}\left( u_{2,\alpha }^{\prime \prime }+\frac{1}{r}u_{2,\alpha }^{\prime }-\frac{1}{4r^{2}}u_{2,\alpha }\right) -2T^{4/3}ru_{1,\alpha } &=&\eta _{\alpha }u_{2,\alpha } \label{eq_u2}\end{aligned}$$and, finally, $$\begin{aligned} -u_{1,\alpha }^{\prime \prime }-\frac{1}{r}u_{1,\alpha }^{\prime }+\frac{1}{% r^{2}}u_{1,\alpha }+\frac{r^{4}}{16}u_{1,\alpha }+\frac{r}{2}u_{2,\alpha } &=&\mu _{\alpha }u_{1,\alpha } \label{eq_u11} \\ -u_{2,\alpha }^{\prime \prime }-\frac{1}{r}u_{2,\alpha }^{\prime }+\frac{% r^{4}}{16}u_{2,\alpha }-\frac{r}{2}u_{1,\alpha } &=&\mu _{\alpha }u_{2,\alpha } \label{eq_u12}\end{aligned}$$where $$\mu _{\alpha }=\frac{2T-\eta _{\alpha }}{4T^{4/3}} \label{mu}$$ A few lowest eigenvalues of Eqs.(\[eq\_u11\]) and (\[eq\_u12\]) are $$\mu _{\alpha }=1.4113;1.83;3.98\ldots \label{mu_n}$$ For $\vec{v}_{\alpha }$ similar procedure gives $$\begin{aligned} -v_{1,\alpha }^{\prime \prime }-\frac{1}{r}v_{1,\alpha }^{\prime }+\frac{1}{% r^{2}}v_{1,\alpha }+\frac{r^{4}}{16}v_{1,\alpha }-\frac{r}{2}v_{2,\alpha } &=&\mu _{\alpha }v_{1,\alpha } \label{eq_v1} \\ -v_{2,\alpha }^{\prime \prime }-\frac{1}{r}v_{2,\alpha }^{\prime }+\frac{% r^{4}}{16}v_{2,\alpha }+\frac{r}{2}v_{1,\alpha } &=&\mu _{\alpha }v_{2,\alpha } \label{eq_v2}\end{aligned}$$ It follows from Eqs.(\[eq\_u11\])-(\[eq\_u12\]) and (\[eq\_v1\])-([eq\_v2]{}) that the functions $\vec{v}_{\alpha }$ is connected with $\vec{u}% _{\alpha }$ by the relations $v_{1,\alpha }=-u_{1,\alpha }$, $v_{2,\alpha }=u_{2,\alpha }$ and, therefore, normalization condition (\[norma\_uv\]) transforms to $$T^{2/3}\int_{0}^{\infty }\left( u_{2,\alpha }u_{2,\beta }-u_{1,\alpha }u_{1,\beta }\right) rdr=\delta _{\alpha ,\beta } \label{norma_u}$$ Using Eqs.(\[Hm3\]) and (\[norma\_uv\]) we obtain the correlation function (\[cor1\]) in a form $$\left\langle \vec{S}_{1}\cdot \vec{S}_{1+n}\right\rangle =\sum_{\alpha }y_{\alpha }^{n-1}f_{\alpha }^{2} \label{cor2}$$where $y_{\alpha }=\eta _{\alpha }/\lambda _{0}$ and $$f_{\alpha }=\int_{0}^{\pi }\psi _{0}(\theta )u_{2,\alpha }(\theta )\sin \left( \theta \right) d\theta =T^{2/3}\int_{0}^{\infty }\psi _{0}(r)u_{2,\alpha }(r)rdr \label{f}$$ At $T\rightarrow 0$ $$y_{\alpha }=\frac{2T-4T^{4/3}\mu _{\alpha }}{2T-4T^{4/3}\varepsilon _{0}}% \approx 1-2T^{1/3}\left( \mu _{\alpha }-\varepsilon _{0}\right) \label{y}$$and the correlation function becomes$$\left\langle \vec{S}_{1}\cdot \vec{S}_{1+n}\right\rangle =\sum_{\alpha }f_{\alpha }^{2}\exp [-2T^{1/3}\left( \mu _{\alpha }-\varepsilon _{0}\right) (n-1)] \label{cor3}$$ According to Eq.(\[cor3\]) the correlation length $\xi $ at $T\rightarrow 0 $ is $$\xi =\frac{1}{2\left( \mu _{0}-\varepsilon _{0}\right) T^{1/3}}=\frac{1.04}{% T^{1/3}} \label{xi}$$ Now we are ready to calculate the magnetic susceptibility at $T\rightarrow 0$, which is$$\chi =\frac{1}{3TN}\sum_{n}\left\langle \vec{S}_{1}\cdot \vec{S}% _{1+n}\right\rangle =\frac{1}{3T}(1+2\sum_{\alpha }\frac{f_{\alpha }^{2}}{% 1-y_{\alpha }})=\frac{1}{3T}+\frac{1}{3T^{4/3}}\sum_{\alpha }\frac{f_{\alpha }^{2}}{\mu _{\alpha }-\varepsilon _{0}} \label{chi}$$ Now we see that $f_{\alpha }^{2}$ and $(\mu _{\alpha }-\varepsilon _{0})$ depends on the solutions of differential equations which are independent of $% T$. So, the sum in $\chi $ gives numerical constant $$\sum_{\alpha }\frac{f_{\alpha }^{2}}{\mu _{\alpha }-\varepsilon _{0}}=3C \label{C}$$ Therefore, the low-temperature susceptibility behaves as $$\chi =\frac{C\left\vert J_{1}^{1/3}\right\vert }{T^{4/3}} \label{chi_f}$$ Numerical calculations gives for the constant $C$ the value $C\approx 1.07$. Thus, the critical exponent for the susceptibility at the transition point is $4/3$ and that for the correlation length is $1/3$. [9]{} A. V. Chubukov, Phys. Rev. B **44**, 4693 (1991). D. V. Dmitriev and V. Ya. Krivnov, Phys. Rev. B **73**, 024402 (2006). F. Heidrich-Meisner, A. Honecker, and T. Vekua, Phys. Rev. B **74**, 020403(R) (2006). T. Hikihara, L. Kecke, T. Momoi, and A. Furusaki, Phys. Rev. B **78**, 144404 (2008). M. E. Fisher, Am. J. Phys. **32**, 343 (1964). M. Hartel, J. Richter, D. Ihle, and S.-L. Drechsler, Phys. Rev. B **78**, 174412 (2008). I. Harada and H. J. Mikeska, Z. Phys. B: Condens. Matter **72**, 391 (1988).

ArXiv

--- abstract: 'We consider a new interaction between a heavy Majorana neutrino ($N$) and a charged Higgs boson ($H^\pm$), and show that it can have drastic implications on lepton number violating (LNV) signal of same-sign dileptons at the LHC. The LNV signal of heavy Majorana neutrinos previously considered at the LHC, $pp \to \ell^+ N \to \ell^+ \ell^+ W^-$, may be overwhelmed by $pp \to \ell^+ N \to \ell^+ \ell^+ H^-$. With the subsequent decays $H^- \to \bar t b$ or $H^- \to W^- H^0$, the heavy Majorana neutrino production leads to the spectacular events of $\ell^+ \ell^+\ b \bar b$ + 2 jets. We also explore the case $m_N < m_{H^+}$, where the decay $H^+ \to \ell^+ N$ can become the dominant $N$-production mechanism at the LHC. In particular, we show that the process $g b \to t H^-$ followed by $t \to b W^+$ and $H^- \to \ell^- N \to \ell^- \ell^- W^+$ could lead to another type of spectacular events of $\ell^- \ell^-\ b$ + 4 jets.' author: - 'Shaouly Bar-Shalom$^{a,b}$' - Gad Eilam$^a$ - 'Tao Han$^{c,d}$' - Amarjit Soni$^e$ title: | Charged Higgs Boson Effects\ in the Production and Decay of a Heavy Majorana Neutrino at the LHC --- Introduction ============ The discovery of neutrino oscillations stands as the first direct evidence for physics beyond the Standard Model (SM), indicating that neutrinos are massive and that their flavors mix. The minimal realization of beyond the SM physics that can account for the observed sub-eV neutrino masses and mixings is constructed simply by adding heavy right-handed neutrino fields, $N$, to the SM Lagrangian (we will denote this minimal setup by $\nu$SM): $$\begin{aligned} {\cal L}_{\nu SM} \equiv {\cal L}_{SM} + {1\over 2} M N N + (Y_{H} L H N + h.c.), \label{nSMF}\end{aligned}$$ where $M$ is the right-handed Majorana neutrino mass scale, $L$ is an SU(2) leptonic doublet and $H$ is the SM Higgs doublet. The $\nu$SM Lagrangian gives rise to the light neutrinos mass matrix through the classic seesaw mechanism: $$\begin{aligned} m_\nu = - m_D M^{-1} m_D^T \label{seesawmass}~,\end{aligned}$$ where $m_D=v Y_H$ and $v= \langle H \rangle \sim 175$ GeV. Thus, $m_\nu \sim {\cal O}(10^{-2}$ eV) implies that either $m_N \sim 10^{15}$ GeV, $Y_H \sim {\cal O}(1)$; or $m_N \sim m_W$ and $Y_H \sim 10^{-7}$. Evidently, if there are heavy Majorana neutrinos at the electroweak (EW) scale, then the seesaw mechanism would seem to be somewhat unnatural and physics beyond the classic seesaw would be needed in order to understand the very small Yukawa couplings associated with the neutrino mass generation. In this minimal $\nu$SM framework, $N$ can interact with the SM gauge bosons and Higgs through its mixing with the light SM SU(2) $\nu_L$ fields (see e.g., [@book]): $$\begin{aligned} {\cal L}_{W} &=& -\frac{g}{\sqrt{2}} U_{\ell N} \bar \ell \gamma^\mu P_L N W^-_\mu + h.c. \label{YWln} ~, \\ {\cal L}_{Z} &=& -\frac{g}{2 c_W} U_{\ell N} \overline{\nu_\ell} \gamma^\mu P_L N Z_\mu + h.c. \label{YZNn} ~, \\ {\cal L}_{H} &=& -\frac{g}{2} \frac{m_N}{m_W} U_{\ell N} \overline{\nu_\ell} P_R N H^0 + h.c. \label{YHNn} ~,\end{aligned}$$ where $P_{L,R} \equiv (1 \mp \gamma_5)/2$ and $U_{\ell N}$ are the $\nu_{\ell,L} - N$ mixing elements. However, with no additional assumptions or symmetries imposed on the $\nu$SM Lagrangian in (\[nSMF\]), these heavy-to-light mixing elements are restricted to be vanishingly small by the seesaw mechanism itself. In particular, the seesaw relation in (\[seesawmass\]) leads to $U_{\ell N} \sim \sqrt{m_\nu/m_N}$. Therefore, we have $U_{\ell N} \to 0$ in order to successfully generate $m_{\nu}$ in the sub-eV range, irrespective of whether $m_N \sim m_W$ or $m_N \sim 10^{15}$ GeV. It follows that, within the minimal seesaw setup embedded in the $\nu$SM, the heavy Majorana neutrinos completely decouple and no signals of $N$ are expected in collider experiments. On the other hand, as argued above, naturalness (i.e., requiring the neutrino Yukawa terms to be of order 1) implies that there is new physics beyond the minimal seesaw mechanism of the $\nu$SM type if indeed $m_N \sim m_W$. In this case it is, therefore, phenomenologically viable to expect that the interactions of $N$ with the EW degrees of freedom are not necessarily suppressed, leading to very interesting lepton-number-violating (LNV) phenomenology mediated by $N$ at high-energy colliders such as the Tevatron, the CERN Large Hadron Collider (LHC) and the International Linear Collider (ILC). Indeed, collider phenomenology of heavy Majorana neutrinos has regained interest in the past decade due to their potential role in generating the observed sub-eV light neutrino masses. With the upcoming LHC and the future ILC $e^+e^-$ collider, the search for LNV signals mediated by heavy Majorana neutrinos is particularly well motivated. This had led to some extensive studies of heavy Majorana neutrinos in $pp$ and $p{\bar p}$ collisions [@delAguila1; @delAguila2; @han; @Wpapers; @oldpp], at an $e^+ e^-$ [@delAguila2; @buch1], $e^-e^-$ [@rizzo; @ourseetohmhm] and $e^- \gamma$ collisions [@delAguila2; @pilaftsis1] and at an $e p$ machine [@delAguila2; @buch1; @ali2]. In addition, LNV decays mediated by exchanges of heavy Majorana neutrinos were studied in top-quark and $W$-boson decays [@ourtWpaper] and in rare charged meson and lepton decays [@ali1]. It should be clear that, although not always explicitly stated, the assumption which underlies [*all*]{} the above studies is that at least one heavy Majorana neutrino has an unsuppressed coupling/mixing with the SM gauge bosons and Higgs and that $N$ production (and decays) at high-energy colliders is induced by this coupling. That is, that $U_{\ell N} \sim {\cal O}(1)\ - $ many orders of magnitudes larger than its naively expected size within the classic seesaw, which necessarily implies new physics beyond the minimal $\nu$SM. In general such new physics can be parametrized by corrections to the $\nu$SM Lagrangian represented as a series of effective operators ${\cal O}_i$ which are constructed using the $\nu$SM fields and whose coefficients are suppressed by powers of $1/\Lambda$, where $\Lambda$ is the scale of the new physics: $$\begin{aligned} {\cal L} = {\cal L}_{\nu SM} + \sum_{n=5}^\infty \frac{1}{\Lambda^{n-4}} \alpha_i {\cal O}_i^n \label{eff}~.\end{aligned}$$ For example, the dimension 6 operator [@our-effectiveN]: $$\begin{aligned} {\cal O}_{N e \phi} = i \left(\bar N \gamma^\mu \ell_R \right) \left(\phi^T \epsilon D_\mu \phi \right)~,\end{aligned}$$ can generate the new V+A charged interaction: $$\begin{aligned} {\cal L}_{W} &=& \alpha_{\ell N} \bar N \gamma^\mu P_R \ell W^+_\mu + h.c. \label{YWlnR} ~,\end{aligned}$$ with $\alpha_{\ell N} \sim {\cal O}(v^2/\Lambda^2)$ (i.e., when the coefficient in front of ${\cal O}_{N e \phi}$ is naturally $\alpha_{N e \phi} \sim {\cal O}(1)$) [@our-effectiveN]. Thus, if the new physics is around the TeV scale we can expect $\alpha_{\ell N} \lsim 0.1$. However, even with this generic parametrization of new physics it is hard to see how the unsuppressed SM-like $N \gamma^\mu P_L \ell W_\mu^+$ interaction in (\[YWln\]) can be generated, when the new heavy Majorana fields are right-handed. To generate such a large SM-like coupling one has to assume that other non-seesaw or seesaw-like mechanisms exist which utilize some fine-tuned relations or extra symmetries in the neutrino sector [@beyond1; @beyond2; @ma; @beyondss; @deGouvea:2006gz] Here we will take a phenomenological approach towards the V-A coupling of $\ell N W^+$, governed by the mixiing parameter $U_{\ell N}$, assumed to be of ${\cal O}(1)$ a priori. EW precision data from LEP imply that $U_{\ell N} \lsim 0.1$ if $m_N > m_Z$ [@kagan; @LEP92] and $U_{\ell N} \lsim 0.01$ if $m_N < m_Z$ [@LEP92]. Indeed, the leading $N$ signature at the LHC is traditionally assumed to be driven by the unsuppressed $N \ell^+ W^-$ interaction vertex as in Eq. (\[YWln\]) [@delAguila1; @delAguila2; @han; @Wpapers] $p p \to W^{+\star} \to \mu^+ N \to \mu^+ \mu^+ W^- \to \mu^+ \mu^+ j j$, where $j$ stands for a light-quark jet. In recent analyses [@han; @delAguila1], it was found that this process may be observed at the LHC at the $5\sigma$ level, with an integrated luminosity of 30$-$100 fb$^{-1}$, if $U_{\ell N} \lsim 0.1$ and $m_N \lsim 200-250$ GeV. The signal and background estimates in these studies only apply to the specific final state $\mu^+ \mu^+ j j$ with no missing energy. In general, however, the $N$-production and decay patterns may be drastically altered due to additional operators involving the interactions of $N$ with the other low-energy degrees of freedom of the underlying new physics. This can lead to interesting new LNV signatures which may be easier to trace. For example, the above Drell-Yan like process, $p p \to \mu^+ N$, may not be the dominant $N$-production mechanism at the LHC, in which case new strategies for $N$-discovery should be adopted. In this paper we wish to explore one specific example of beyond the $\nu$SM physics, in which $N$-production and decays may be completely altered. Since non-zero neutrino mass necessarily requires new physics beyond the SM and since the extension for the SM Higgs sector is well motivated in many theories beyond the SM, it is natural to consider the interplay of both. In particular, we will focus on a heavy Majorana neutrino potentially accessible at the LHC with a mass in the range $10~{\rm GeV} \lsim m_N \lsim 500~{\rm GeV}$, and the observation feasibility through its interactions with a generic new charge Higgs boson. The paper is organized as follows: in section II we lay out the theoretical setup. In sections III - V we discuss $N$-phenomenology at the LHC in the presence of the new $N-H^+$ interaction in the two cases $m_N > m_{H^+}$ and $m_N < m_{H^+}$, and in section VI we summarize and give our concluding remarks. The Charged Higgs Boson and a Heavy Majorana Neutrino ===================================================== We first introduce a generic new $\ell N H^+$ interaction in a relative model-independent approach $$% {\cal L}_{\ell N H^+ } = \frac{g}{2 \sqrt{2}} \xi_{\ell N} \frac{m_N}{m_W} \bar N (1-\gamma_5) \ell H^+ + h.c. \label{YHln2}~, {\cal L}_{\ell N H^+ } = \frac{g}{ \sqrt{2}} \xi_{\ell N} \frac{m_N}{m_W} \bar N P_L \ell H^+ + h.c. \label{YHln2}~,$$ where $\xi_{\ell N}$ are dimensionless parameters whose size depend on the underlying new physics. In particular, since ${\cal L}_{\ell N H^+ }$ is a typical dimension 4 Yukawa-like term, one naturally expects $\xi \sim {\cal O}(1)$ if the new physics contains EW-scale new scalar fields as well as heavy Majorana neutrinos. Alternatively, such an effective interaction can be generated in the underlying theory by exchanges of heavy gauge-bosons or heavy fermions. In this case, guided by the effective Lagrangian prescription in (\[eff\]) and by naturalness (i.e., $\alpha_i \sim {\cal O}(1)$), when these heavy particles are integrated out we expect: $$\begin{aligned} \xi \sim \frac{\sqrt{2}}{g} \frac{m_W}{m_N}\ \frac{v^2}{\Lambda^2},\end{aligned}$$ where $\Lambda$ is roughly the mass of the new heavy particle that gives rise to the $\ell N H^+$ interaction in (\[YHln2\]). Thus, when $m_N \sim m_W$ we can expect e.g., $\xi \gsim 0.1$ if $\Lambda \sim 1$ TeV. Since there is no direct experimental constraint on $\xi$ that we know of, we will take a phenomenological approach towards the new $\xi_{\ell N}$, exploring the implications of ${\cal L}_{\ell N H^+}$ for $\xi$ in the range $0 < \xi < 1$ (note that $\xi \sim {\cal O}(1)$ is also consistent with perturbative unitarity if $m_N \lsim 700$ GeV, as was noted in [@ourseetohmhm]). Then, depending on its exact size, these new $\ell N H^+$ interactions can have surprising implications on $N$-phenomenology at high-energy colliders. For instance, in [@ourseetohmhm] it was shown that such a coupling can drive a LNV same-sign charged Higgs pair-production signal, $e^- e^- \to H^- H^-$, at an observable rate at an ILC even if $m_N \sim 1000$ TeV. As for collider phenomenology of N in the presence of the interactions (\[YWln\])-(\[YHNn\]), (\[YWlnR\]) and (\[YHln2\]), we adopt a “one-coupling scheme" for simplicity, assuming that only one of the mixing angles dominates, e.g., $U_{\mu N} \gg U_{e N},\ U_{\tau N}$ and similarly, $\xi_{\mu N} \gg \xi_{e N},~\xi_{\tau N}$. These elements will be denoted hereafter by $U \equiv U_{\mu N}$ and $\xi \equiv \xi_{\mu N}$. $N$ decay in the presence of $H^\pm$ ------------------------------------ In the framework where a new $\mu N H^+$ interaction is present, $N$ will predominantly decay via the kinematically accessible channels $$N \to W^\mp \mu^\pm ,~Z \nu_\mu,~H^0 \nu_\mu,~H^\mp \mu^\pm .$$ The partial widths for these decay channels are given by $$\begin{aligned} \Gamma(N \to \mu^\pm W^\mp) &\approx& \frac{U^2 m_N^3}{16 \pi v^2}\ (1 + 2 r_W^{}) (1 - r_W^{})^2,\nonumber \\ \Gamma(N \to \nu_\mu Z) & \approx & \frac{U^2 m_N^3}{16 \pi v^2}\ (1 + 2 r_Z^{}) (1 - r_Z^{})^2, \nonumber \\ \Gamma(N \to \nu_\mu H^0) & \approx & \frac{U^2 m_N^3}{16 \pi v^2}\ (1 - r_{H^0}^{})^2, \nonumber \\ \Gamma(N \to \mu^\pm H^\mp) & \approx & \frac{\xi^2 m_N^3}{16 \pi v^2}\ (1 - r_{H^+}^{})^2, \label{partial}\end{aligned}$$ where $r_i=m_i^2/m_N^2$. A useful limit for illustrating that is $m_N^2 \gg m_W^2, m_Z^2, m_{H^0}^2, m_{H^+}^2$, in which case the total decay width of $N$ can be conveniently written as: $$\Gamma_N \approx \Gamma_N^0 \cdot \left(4U^2 + 2 \xi^2 \right),\quad \Gamma_N^0 \equiv \frac{g^2}{64 \pi} \frac{m_N^2}{m_W^2} m_N = \frac{m_N^3}{16 \pi v^2}. \label{width0}$$ Clearly then, $N$ can become “W/Z/$H^0$-phobic” if $\xi^2 \gg U^2$, since in this case its width is dominated by its decays to the new charged Higgs. To demonstrate the effect of the $\mu N H^+$ coupling $\xi$ on the pattern of the $N$-decays, we plot in Fig. \[fig12\](a) the BR’s for the decays $N \to \mu^+ W^-$ and $N \to \mu^+ H^-$ as a function of $\xi$, for $m_N=150$ GeV and $m_{H^+}=100$ GeV, and setting $U=0.1$ and $m_{H^0}=200$ GeV. We can indeed see the sharp drop of the $BR(N \to \mu^+ W^-)$ as $\xi$ increases. In particular, for $\xi \sim {\cal O}(1)$, we obtain $BR(N \to \mu^+ W^-) \sim 0.01$ while $BR(N \to \mu^+ H^-,\ \mu^- H^+)$ saturate the decay. In Figs. \[fig12\](b) we further plot the decays $N \to \mu^+ W^-$ and $N \to \mu^+ H^-$ as a function of $m_N$, for $\xi=1$ and for $m_{H^+}=150$ GeV, $m_{H^0}=200$ GeV and setting $U$ to its largest allowed values depending on $m_N$, i.e., $U=0.1$ for $m_N>m_Z$ and $U=0.01$ for $m_N<m_Z$. Below the threshold $m_N < m_{H^+},\ m_W$, the off-shell effects have been taken into account. Way above the threshold, the relative branchings for $N \to \mu^+ H^-$ and $N \to \mu^+ W^-$ is governed by $\xi^2/U^2$. Including $H^\pm$ decays ------------------------ As a concrete example for illustration, let us assume that the $H^+$-SM couplings have similar structure as the charged Higgs couplings in generic multi Higgs doublet models: $$\begin{aligned} {\cal L}_{H^+u d} &=& \frac{g}{\sqrt{2}} \bar u \left( f_v^u \frac{m_{u}}{m_W} P_L + f_v^d \frac{m_{d}}{m_W} P_R \right) d H^+ + h.c. \label{YHud} ~, \\ {\cal L}_{H^+ \nu_\ell \ell} &=& \frac{g}{\sqrt{2}} f_v^\ell \frac{m_{\ell}}{m_W} \bar\nu_\ell P_R \ell H^+ + h.c. \label{YHln} ~, \\ {\cal L}_{H^+ H^0 W^-} &=& \frac{i g}{2} f_{W} W^+ \cdot \left( H^0 \partial H^- - \partial H^0 H^- \right) + h.c. \label{YHHW} ~,\end{aligned}$$ and that the dominant $H^+$ decays channels are $$H^+ \to \tau^+ \nu,~c \bar s,~t \bar b,~\mu^+ N,~W^+ H^0,$$ when kinematically allowed. As an example, in Fig. \[fig34\](a) we show the expected BR’s for the decays $N \to \mu^+ W^-$ and $N \to \mu^+ \bar t b$ as a function of the $\mu N H^+$ coupling strength $\xi$, setting $U=0.1$, $m_N=220$ GeV and $m_{H^+}=m_{H^0}=200$ GeV. For the $H^+$ couplings we took $f_v^c=f_v^t=f_v^\tau=f_{W}=1$ and $f_v^d=0$. The channel $N \to \mu^+ j j$ from $W\to jj$ is also plotted for comparison by the dotted curve. We see that for this set of masses and couplings, the channel $N \to \mu^+ \bar t b$ dominates over the previously studied $N \to \mu^+ jj$ one when $\xi \gsim 0.5$. To further explore the potential deviations from “standard” $N$-phenomenology at future colliders, we plot in Fig. \[fig34\](b) the BR’s for the decays $N \to \mu^+ \bar t b$, $N \to \mu^+ j j$ and $N \to \mu^+ W^- H^0$ for another set of inputs and find that the $WH^0$ mode can be dominant, when $m_{H^+} > m_{H^0}$ and if the $H^+tb$ coupling is suppressed. In the next section we will show that these new decay channels of N via an intermediate $H^+$, can have interesting implication for N-phenomenology at the LHC. Same-sign di-muons signals: $pp \to \mu^+ \mu^+ W^-$ versus $pp \to \mu^+ \mu^+ H^-$ ==================================================================================== As reiterated recently, a heavy Majorana neutrino may be searched for at the LHC via the Drell-Yan production in [@delAguila1; @delAguila2; @han; @Wpapers; @oldpp] $$\begin{aligned} p p \to \mu^+ N + h.c. ~,\end{aligned}$$ with $N \to \mu^\pm W^\mp$, leading to the spectacular signal of $\mu^\pm \mu^\pm+$ 2 jets. As discussed above, the existence of the $\ell N H^+$ interaction in (\[YHln2\]) may change the $N$ decay pattern significantly, possibly resulting in $N \to \mu^\pm H^\mp$ as the dominant mode. We will study this effect in more detail below. Denoting the cross-section $\sigma_N \equiv \sigma(pp \to \mu^+ N)$, we can generically study the unknown couplings as parameterized in Eqs. (\[YWln\]) and (\[YWlnR\]), and expect $\sigma_N = U^2 \sigma_{V-A} $ or $\sigma_N = \alpha_{\mu N}^2 \sigma_{V+A} $ if $U \ll \alpha_{\mu N}$, where $U$ and $\alpha_{\mu N}$ are generic (V-A) and (V+A) $~$ $\mu N W$ couplings, respectively. Since $\sigma_{V-A} = \sigma_{V+A}$ (up to factors that linearly depend on the cosine of the center of mass scattering angle, $\cos\theta$, which vanish after symmetrically integrating over phase-space), we will generically denote $\sigma_N \propto U^2$ regardless of whether it is generated by the $V-A$ or the $V+A$ coupling. In particular, for $U \sim 0.1$ (i.e., at its upper limit value), one gets $\sigma_N \sim 100$ fb or $\sigma_N \sim 10$ fb for $m_N \sim 100$ or 200 GeV, respectively, with no cuts [@delAguila1; @delAguila2; @han; @Wpapers]. The kinematical cuts applied to the final state particles after the $N$ decays, e.g., to $\mu^+ \mu^+ j j$ when $N \to \mu^+ W^- \to \mu^+ j j$, are expected to reduce these cross-sections by less than an order of magnitude [@delAguila1; @han]. For $m_N$ in the range of several hundreds GeV, the total $N$ width is at most a few percents of $m_N$. Thus, we can take the narrow width approximation (NWA): $$\begin{aligned} \sigma(pp \to \mu^+ \mu^+ W^\mp) &\sim& \sigma(pp \to \mu^+ N) \times BR(N \to \mu^+ W^-) ~, \nonumber \\ \sigma(pp \to \mu^+ \mu^+ H^\mp) &\sim& \sigma(pp \to \mu^+ N) \times BR(N \to \mu^+ H^-) ~, \end{aligned}$$ as a good estimate of the total same-sign di-muon cross-sections above. Using the NWA, in Table \[tab1\] we give some sample results which compare between the above two signals, for the case $m_N > m_{H^+} > m_W$ and when $\xi=1$, $U=0.1$, without applying any kinematical cuts. To give a complete picture for the expected same-sign di-muons signal at the LHC, we list in Table \[tab1\] the total cross-sections summing the contributions from both the $\mu^+ \mu^+$ and the $\mu^- \mu^-$ channels. We denote these total cross-sections by $\sigma(pp \to \mu \mu W) \equiv \sigma(pp \to \mu^+ \mu^+ W^-) + \sigma(pp \to \mu^- \mu^- W^+)$ and similarly for $\sigma(pp \to \mu \mu H)$, reminding the reader that at the LHC the positively charged di-muons rate is about 1.5 times larger than the negatively charged dimuons one. As seen from Table \[tab1\], the cross-section for $\mu^+ \mu^+ H^-$ can become more than an order of magnitude larger than that of $\mu^+ \mu^+ W^-$ (similar for the $\mu^- \mu^-$ channels), if $m_N > m_{H^+}$ and $\xi \sim 1$, even if one assumes the largest possible cross-section for $\mu^+ \mu^+ W^-$, i.e., taking $U =0.1$. In such a case, heavy Majorana neutrinos should be searched for at the LHC through their decay to a charged Higgs boson. The representative cross section can be the order of $1-100$ fb, leading to $10^2-10^4$ events with an integrated luminosity of 100 fb$^{-1}$ before acceptance cuts on the final states. ---------- ----------- ------------------ ------------------ ----------------------- ----------------------- -------------------- -------------------- $m_N$ $m_{H^+}$ $pp \to \mu^+ N$ $pp \to \mu^- N$ $N \to \mu^\pm W^\mp$ $N \to \mu^\pm H^\mp$ $pp \to \mu \mu W$ $pp \to \mu \mu H$  \[GeV\] \[GeV\] $\sigma$ \[fb\] $\sigma$ \[fb\] BR BR $\sigma$ \[fb\] $\sigma$ \[fb\]               100 80 155.6 106.9 0.008 0.49 2.1 128.6               150 120 29.4 18.9 0.022 0.46 1.1 22.2               200 150 10.4 6.3 0.02 0.47 0.33 7.9               220 200 7.4 4.4 0.07 0.25 0.83 2.95               ---------- ----------- ------------------ ------------------ ----------------------- ----------------------- -------------------- -------------------- : Branching fractions (BR) for the two-body decays $N \to \mu^\pm W^\mp,~\mu^\pm H^\mp$ and cross-sections in fb for the Drell-Yan $N$-production $pp \to \mu^\pm N$ and for the total same-sign dimuon signals at the LHC: $\sigma(pp \to \mu \mu W) \equiv \sigma(pp \to \mu^+ \mu^+ W^-) + \sigma(pp \to \mu^- \mu^- W^+)$ and similarly for $\sigma(pp \to \mu \mu H)$, see also text. The cross-sections are evaluated using the CTEQ6M parton distribution functions and results are given for various sets of $N$ and $H^+$ masses, for $\xi=1$, $U=0.1$ and for $m_{H^0}=200$ GeV. No cuts are applied.[]{data-label="tab1"} Of particular interest is the situation of $H^\pm$ decay above the thresholds. For $m_{H^+} > m_t+m_b$, the single-top production in association with a pair of same-sign di-muons will become the dominant channel $$\begin{aligned} pp \to \mu^+ N \to \mu^+ \mu^+ H^- \to \mu^+ \mu^+ \bar t b \label{mmtb}~, \end{aligned}$$ with a BR to be about 25% (see Fig. \[fig34\](a)) $-$ almost an order of magnitude larger than that of $pp \to \mu^+ N \to \mu^+ \mu^+ j j$. After the top decays, this single-top LNV channel is manifest through $\mu^+ \mu^+ \bar t b \to \mu^+ \mu^+ b b j j$, i.e., a pair of same-sign di-muons in association with a pair of light jets and a pair of $b$-jets, where the two light jets reconstruct $m_W$, which combine with one $b$-jet to reconstruct the top mass. Along with the second $b$-jet, the whole $b\bar b jj$ system further reconstructs the charged Higgs mass. For $m_{H^+} > m_W+m_{H^0}$, the new channel $$\begin{aligned} pp \to \mu^+ N \to \mu^+ \mu^+ H^- \to \mu^+ \mu^+ W^- H^0 ~,\end{aligned}$$ can become even more important, having a BR at the level of 10% and dominating over the $pp \to \mu^+ N \to \mu^+ \mu^+ \bar t b$ and $pp \to \mu^+ N \to \mu^+ \mu^+ j j$ ones, if for example: $m_{H^+}=200$ GeV, $m_{H^0}=100$ GeV, $U=0.01$, $\xi=1$ and the $H^+$ couplings: $f_v^c=f_v^t=f_v^\tau=0.1$, $f_v^d=0$ and $f_{W}=1$ (see Fig. \[fig34\](b)).[^1] Similar to the single-top LNV signal of Eq. (\[mmtb\]), the LNV $\mu^+ \mu^+ W^- H^0$ channel can also induce the signature $\mu^+ \mu^+ W^- H^0 \to \mu^+ \mu^+ b b j j$, but with different kinematics of the final state, namely the pair of $b$-jets reconstruct $m_{H^0}$. As for the signal identification, we note that there are no SM processes with $\Delta L=2$. All backgrounds are due to some misidentification of certain sources. Recent studies demonstrated that with judicious acceptance cuts, the SM backgrounds to the $\mu^+ \mu^+ j j$ signal can be effectively suppressed [@han; @delAguila1] With the presence of two more $b$-jets coming from a top-quark decay or from a Higgs decay, along with a pair of same-sign di-muons in the final state, we expect to significantly improve the signal to background ratio, in comparison with the $\mu^+ \mu^+ j j$ one. We do not plan to further quantify the signal identification in this work. New contribution when $m_N < m_{H^+}$ ===================================== Although the spectacular decay of $N\to H^\pm \ell^\mp$ is kinematically not allowed for this case, the inverted process $H^+ \to \ell^+ N$ can become the leading $N$-production mechanism. In this case the leading $H^+$ production mechanism will be $g \bar b \to \bar t H^+$ (and similarly $g b \to t H^-$), with a cross-section at the LHC ranging from ${\cal O}$(1000 fb) if $m_{H^+} \sim 200$ GeV to ${\cal O}$(100 fb) if $m_{H^+} \sim 600$ GeV [@Berger:2003sm]. This channel has been considered as a promising channel for the discovery of a charged Higgs boson [@0710.1761]. Therefore, after $H^+$ decays via $H^+ \to \ell^+ N$, we expect the $tN$-associated production rate at the LHC to be: $$\begin{aligned} \sigma_N (g \bar b \to \bar t H^+ \to \bar t \mu^+ N) \sim (100 -1000 ~{\rm fb}) \times {\rm BR}(H^+ \to \mu^+ N) \label{Nt}~,\end{aligned}$$ for $m_{H^+}$ of several hundreds GeV. Then, if $\xi \sim f_v^t \sim 1$, we expect ${\rm BR}(H^+ \to \mu^+ N) \sim {\rm BR}(H^+ \to t \bar b) \sim {\cal O}(1)$, in which case the $N$-signal top associated production channel in (\[Nt\]) can have a cross-section at the order of 100 fb, even if $m_N \sim 500$ GeV. This is to be compared with the Drell-Yan $pp \to \mu^+ N$ production rate which is $\sigma(pp \to W^{+\star} \to \mu^+ N) \sim 0.1$ fb for $m_N \sim 500$ GeV. Thus, while no signal of heavy Majorana neutrinos with a mass around 500 GeV is expected at the LHC through the traditionally assumed Drell-Yan production mechanism, its production rate through $H^+$ decays can be as much as 3 orders of magnitude larger than the Drell-Yan one, possibly leading to hundreds-thousends of heavy Majorana neutrinos in association with a pair of $t \mu^+$ at the LHC. When the $N$ further decays there are several interesting signatures that arise from the above $\bar t \mu^+ N$ (or $t \mu^- N$) final state, depending on its couplings to the SM particles. Perhaps the simplest example is $pp \to \mu^+ \bar t N$ followed by $N \to \mu^+ W^-$, which will lead to the new LNV signature $$\mu^+ \mu^+ \bar t W^- \to \mu^+ \mu^+ W^- W^- b,$$ i.e., same sign di-muons accompanied by an opposite sign $W$-pair and a hard $b$-jet, which reconstructs the top-quark and the $H^+$. The hadronic decays of the $W^-W^-$ may be desirable for the confirmation of the lepton-number violation as well as for the $m_t$ and $m_{H^+}$ reconstruction. A comprehensive analysis of such new $N$-mediated LNV signal as well as the SM backgrounds is beyond the scope of this work, although once again we do not expect it too difficult for a signal identification. Clearly, though, such new LNV signals deserve to be separately addressed, since their kinematics and background completely depart from the ones expected for the widely studied $\mu^+ \mu^+ j j$ one, and, in particular, since hundreds to thousends of such events are expected at the LHC if $\xi \sim 1$. Finally, to conclude this section we note that if $m_N < m_W < m_{H^+}$, then one might expect an enhancement in the Drell-Yan process $pp \to W^+ \to \mu^+ N$ since the s-channel $W^+$ can resonate, giving rise to a peak in the invariant $\mu^+ N$ mass, which might be easier to handle due to its simplicity. However, since the limits on the $\mu W N$ coupling are more stringent in this $N$-mass range, i.e., $U \lsim 0.01$ for $m_N \lsim m_W$ [@LEP92], one expects at most (when $U \sim 0.01$) $\sigma (pp \to W^+ \to \mu^+ N) \sim 0.1 - 10$ \[fb\] for $m_N \sim 40 -80$ GeV, respectively, after applying typical LHC cuts, see e.g., [@delAguila1]. Evidently, even in the case $m_N < m_W < m_{H^+}$ we expect $H^+ \to \mu^+ N$ to be the dominant $N$-production mechanism at the LHC. Summary and discussion ====================== We have argued that naturalness of the seesaw mechanism, in the sense of having ${\cal O}(1)$ neutrino Yukawa terms, requires that either $m_N \sim 10^{16}$ GeV or that new physics beyond the classic seesaw mechanism exist in the neutrino sector if $m_N$ is close to the EW-scale. Therefore, in the latter case when e.g., $m_N \sim 100 -1000$ GeV, we expect that the new physics will generate ${\cal O}(1)$ couplings between the heavy Majorana neutrinos and the SM particles as well as between the heavy Majorana neutrinos and the other EW-scale degrees of freedom of the underlying new physics. These new couplings can then be manifest through new LNV signals, mediated by the heavy Majorana neutrinos, at future high-energy colliders such as the LHC and the ILC. In this note we have considered one fairly model-independent example of such TeV scale new physics that can drastically change what is considered to be the conventional $N$-phenomenology at the LHC. In particular, we have assumed that there is an ${\cal O}(1)$ coupling between the heavy Majorana neutrino and a new charged Higgs of the underlying theory. We then showed that such a new interaction term can have interesting new implications on LNV same-sign lepton pair signals at the LHC. For example, we found that the frequently-studied leading LNV signal of heavy Majorana neutrinos at the LHC, $pp \to W^+ \to \ell^+ N \to \ell^+ \ell^+ W^-$, can become irrelevant in the presence of a sizable $\ell N H^+$ coupling, since if ${\rm BR}(N \to \ell^+ H^-) \gg {\rm BR}(N \to \ell^+ W^-)$, the LNV signal of a pair of same-sign charged leptons in association with a charged Higgs, $pp \to \ell^+ N \to \ell^+ \ell^+ H^-$ become dominant and can lead to new LNV signatures such as $pp \to \ell^+ N \to \ell^+ \ell^+ \bar t b$ and $pp \to \ell^+ N \to \ell^+ \ell^+ W^- H^0$. We have also shown that, in the case $m_N < m_{H^+}$, the decay $H^+ \to \ell^+ N$ is expected to become the dominant $N$-production mechanism at the LHC, possibly leading to hundreds-thousends LNV events mediated by the heavy Majorana neutrinos, e.g., such as a new $N$-single top associated production $pp \to \ell^- \ell^- t W^+ \to \ell^- \ell^- W^+ W^+ b$. We did not perform any signal to background analysis, as our goal was only to emphasize that new physics associated with heavy Majorana neutrinos can be manifest in various, sometimes unexpected, signals and, as such, may call for new strategies in the search for lepton number violation at future colliders. However, due to the additional $b$-quarks and light jets on top of the spectacular $\mu^\pm \mu^\pm jj$ signal, we expect that the SM backgrounds will be even easier to deal with than commonly studied $\mu\mu jj$ channel. Acknowledgment ============== SBS thanks the hospitality of the theory group in Brookhaven National Laboratory where part of this study was performed. The work of SBS was supported in part by NSF Grants No. PHY-0653656 (UCI), PHY-0709742 (UCI) and by the Alfred P. Sloan Foundation. The work of TH is supported in part by the U.S. Department of Energy under grant No. DE-FG02-95ER40896, by the Wisconsin Alumni Research Foundation. The work at the KITP was supported in part by the National Science Foundation under Grant No. PHY05-51164. The work of AS was supported in part by US DOE Contract No. DE-AC02-98CH10886 (BNL). [99]{} R. N. Mohapatra and P. B. Pal, [Massive Neutrinos in Physics and Astrophysics]{}, World Scientific (1998); A. Datta and A. Pilaftsis, Phys. Lett.  B [**278**]{}, 162 (1992). W. Y. Keung and G. Senjanovic, Phys. Rev. Lett.  [**50**]{}, 1427 (1983); T. H. Ho, C. R. Ching and Z. J. Tao, Phys. Rev.  D [**42**]{}, 2265 (1990); D.A. Dicus and D.D. Karatas, Phys. Rev. [**D44**]{}, 2033 (1991). A. Datta, M. Guchait and D.P. Roy, Phys. Rev. [**D47**]{}, 961 (1993); A. Datta, M. Guchait and A. Pilaftsis, Phys. Rev. [**D50**]{}, 3195 (1994). T. Han and B. Zhang, Phys. Rev. Lett. [**97**]{}, 171804 (2006). F. del Aguila, J.A. Aguilar Saavedra and R. Pittau, J. Phys. Conf. Ser. [**53**]{}, 506 (2006). F.M.L. Almeida [*et al.*]{}, Phys. Rev. [**D62**]{}, 075004 (2000); A. Ferrari [*et al.*]{}, Phys. Rev. [**D62**]{}, 013001 (2000); O. Panella [*et al.*]{}, Phys. Rev. [**D65**]{}, 035005 (2002); S. Bray, J. S. Lee and A. Pilaftsis, arXiv: hep-ph/0702294. F. del Aguila, J.A. Aguilar-Saavedra and R. Pittau, JHEP 0710:047 (2007). F. del Aguila, E. Laermann, P. M. Zerwas, Nucl. Phys. [**B297**]{}, 1 (1988); W. Buchmüller and C. Greub, Nucl. Phys. [**B363**]{}, 345 (1991); J. Gluza, M. Zralek, Phys. Rev. [**D48**]{}, (1993); G. Azuelos, A. Djouadi, Z. Phys. [**C63**]{}, 327 (1994); J. Gluza [*et al.*]{}, Phys. Lett. [**B407**]{}, 45 (1997); F.M.L. Almeida [*et al.*]{}, Eur. Phys. J. [**C30**]{}, 327 (2003). T.G. Rizzo, Phys Let. [**B116**]{}, 23 (1982) and hep-ph/9501261; G. Belanger [*et al.*]{}, Phys. Rev. [**D53**]{}, 6292 (1996). D. Atwood, S. Bar-Shalom and A. Soni, Phys. Rev. [**D76**]{}, 033004 (2007). S. Bray, J.S. Lee and A. Pilaftsis, Phys. Lett. [**B628**]{}, 250 (2005). W. Buchmüller and C. Greub, Nucl. Phys. [**B256**]{}, 465 (1991); G. Ingelman, J. Rathsman, Z. Phys. [**C60**]{}, 243 (1993); M. Flanz, W. Rodejohann and K. Zuber, Phys. Lett. [**B473**]{}, 324 (2000), Erratum-[*ibid.*]{} [**B480**]{}, 418 (2000); A. Ali, A.V. Borisov and D.V. Zhuridov, hep-ph/0512005. S. Bar-Shalom [*et al.*]{}, Phys. Lett. [**B643**]{}, 342 (2006). A. Ali, A.V. Borisov and N.B. Zamorin, Eur. Phys. J. [**C21**]{}, 123 (2001); X.-G. He, G. Valencia and Y. Wili, Phys. Rev. [**D70**]{}, 113011 (2004). F. del Aguila, S. Bar-Shalom, A. Soni and J. Wudka, “Heavy Majorana Neutrinos In The Effective Lagrangian Description”, in preparation. For a recent overview on non-minimal seesaw-like models see, J. W. F. Valle, J. Phys. Conf. Ser. [**53**]{}, 473 (2006). F. del Aguila and M. Zralek, Acta Phys. Polon. [**B33**]{}, 2585 (2002); F. del Aguila, M. Masip and J. L. Padilla, Phys. Lett. [**B627**]{}, 131 (2005). E. Ma, Phys. Rev. [**D73**]{}, 077301 (2006). A. Pilaftsis, Z. Phys. [**C55**]{}, 275 (1992) (also in hep-ph/9901206); D. Tommasini [it et al]{}, Nucl. Phys. [**B444**]{}, 451 (1995); J. Gluza Acta Phys. Polon. [**B33**]{}, 1735 (2002); J. Kersten and A. Y. Smirnov, Phys. Rev.  D [**76**]{}, 073005 (2007) \[arXiv:0705.3221 \[hep-ph\]\]. A. de Gouvea, J. Jenkins and N. Vasudevan, Phys. Rev.  D [**75**]{}, 013003 (2007) \[arXiv:hep-ph/0608147\]. S. Bergmann and A. Kagan, Nucl. Phys. [**B538**]{}, 368 (1999). O. Adriani [*et al.*]{} for the L3 Collaboration, Phys. Lett. [**B295**]{}, 371 (1992). For a recent analysis of the $\beta \beta_{0\nu}$ bound see e.g., P. Benes [*et al*]{}, Phys. Rev. [**D71**]{}, 077901 (2005). See e.g., S. Komamiya, Phys. Rev. [**D38**]{}, 2158 (1988); A. Djouadi [*et al.*]{}, Z. Phys. [**C74**]{}, 93 (1997). E. L. Berger, T. Han, J. Jiang and T. Plehn, Phys. Rev.  D [**71**]{}, 115012 (2005) \[arXiv:hep-ph/0312286\]. A. Belyaev, D. Garcia, J. Guasch and J. Sola, JHEP [**0206**]{}, 059 (2002) \[arXiv:hep-ph/0203031\]; See also, M. Flechl (for the ATLAS and CMS collaborations), hep-ph/0710.1761. [^1]: Note that $f_{W}=1$ correspond, for example, to the coupling in a two Higgs doublets model (e.g., in the MSSM) when $\alpha=\beta$, where $\alpha$ and $\beta$ are the neutral Higgs mixing angles in the CP-even and CP-odd sectors, respectively.

ArXiv

--- abstract: | The <span style="font-variant:small-caps;">Pythia</span> program is a standard tool for the generation of high-energy collisions, comprising a coherent set of physics models for the evolution from a few-body hard process to a complex multihadronic final state. It contains a library of hard processes and models for initial- and final-state parton showers, multiple parton-parton interactions, beam remnants, string fragmentation and particle decays. It also has a set of utilities and interfaces to external programs. While previous versions were written in Fortran, <span style="font-variant:small-caps;">Pythia</span> 8 represents a complete rewrite in C++. The current release is the first main one after this transition, and does not yet in every respect replace the old code. It does contain some new physics aspects, on the other hand, that should make it an attractive option especially for LHC physics studies. PACS: 13.66.-a, 13.85.-t, 12.38.-t, 12.15.-y, 12.60.-i event generators, multiparticle production, parton showers, multiple interactions, hadronisation address: - 'CERN/PH, CH–1211 Geneva 23, Switzerland' - | Department of Theoretical Physics, Lund University,\ Sölvegatan 14A, SE-223 62 Lund, Sweden - 'Fermi National Accelerator Laboratory, Batavia, IL 60510, USA' author: - Torbjörn Sjöstrand - Stephen Mrenna - Peter Skands title: 'A Brief Introduction to PYTHIA 8.1' --- CERN-LCGAPP-2007-04\ LU TP 07-28\ FERMILAB-PUB-07-512-CD-T\ October 2007\ , , [**NEW VERSION PROGRAM SUMMARY**]{} [*Manuscript Title:A Brief Introduction to <span style="font-variant:small-caps;">Pythia</span> 8.1*]{}\ [*Authors:Torbjörn Sjöstrand, Stephen Mrenna, Peter Skands*]{}\ [*Program Title: <span style="font-variant:small-caps;">Pythia</span> 8.1*]{}\ [*Journal Reference:*]{}\ [*Catalogue identifier:*]{}\ [*Licensing provisions: GPL version 2*]{}\ [*Programming language: C++*]{}\ [*Computer: commodity PCs*]{}\ [*Operating systems: Linux; should also work on other systems*]{}\ [*RAM: 8*]{} megabytes\ [*Keywords: event generators, multiparticle production, parton showers, multiple interactions, hadronisation*]{}\ [*PACS: 13.66.-a, 13.85.-t, 12.38.-t, 12.15.-y, 12.60.-i*]{}\ [*Classification: 11.2 Phase Space and Event Simulation*]{}\ [*Catalogue identifier of previous version: ADNN\_v1\_0*]{}\ [*Journal reference of previous version: T. Sjöstrand, P. Edén, C. Friberg, L. Lönnblad, G. Miu, S. Mrenna and E. Norrbin, Computer Physics Commun. [**135**]{} (2001) 238*]{}\ [*Does the new version supersede the previous version?: yes, partly*]{}\ \ [*Nature of problem: high-energy collisions between elementary particles normally give rise to complex final states, with large multiplicities of hadrons, leptons, photons and neutrinos. The relation between these final states and the underlying physics description is not a simple one, for two main reasons. Firstly, we do not even in principle have a complete understanding of the physics. Secondly, any analytical approach is made intractable by the large multiplicities.* ]{}\ \ [*Solution method: complete events are generated by Monte Carlo methods. The complexity is mastered by a subdivision of the full problem into a set of simpler separate tasks. All main aspects of the events are simulated, such as hard-process selection, initial- and final-state radiation, beam remnants, fragmentation, decays, and so on. Therefore events should be directly comparable with experimentally observable ones. The programs can be used to extract physics from comparisons with existing data, or to study physics at future experiments.*]{}\ \ [*Reasons for the new version: improved and expanded physics models, transition from Fortran to C++*]{}\ \ [*Summary of revisions: new user interface, transverse-momentum-ordered showers, interleaving with multiple interactions, and much more*]{}\ \ [*Restrictions: depends on the problem studied*]{}\ \ [*Unusual features: none*]{}\ \ [*Running time: 10–1000 events per second, depending on process studied*]{}\ Introduction ============ The development of <span style="font-variant:small-caps;">Jetset</span> [@jetset], containing several of the components that later were merged with <span style="font-variant:small-caps;">Pythia</span> [@pythiaearly], was begun in 1978. Thus the current <span style="font-variant:small-caps;">Pythia</span> 6 generator [@pythiasixone; @pythiasixfour] is the product of almost thirty years of development, and some of the code has not been touched in a very long time. New options have been added, but old ones seldom removed. The basic structure has been expanded in different directions, well beyond what it was once intended for, making it rather cumbersome by now. From the onset, all code has been written in Fortran 77. For the LHC era, the experimental community has made the decision to move heavy computing completely to C++. Fortran support may be poor to non-existing, and young experimenters will not be conversant in Fortran any longer. Therefore it is logical also to migrate <span style="font-variant:small-caps;">Pythia</span> to C++, and in the process clean up and modernise various aspects. A first attempt in this direction was the <span style="font-variant:small-caps;">Pythia</span> 7 project [@pythiaseven]. However, priority came to be given to the development of a generic administrative structure, renamed <span style="font-variant:small-caps;">ThePEG</span> [@thepeg] and adopted by the <span style="font-variant:small-caps;">Herwig++</span> [@herwigpp] group, while the physics parts of <span style="font-variant:small-caps;">Pythia</span> 7 remained underdeveloped. <span style="font-variant:small-caps;">Pythia</span> 8 is a clean new start, to provide a successor to <span style="font-variant:small-caps;">Pythia</span> 6. It is a completely standalone generator, thus not relying on <span style="font-variant:small-caps;">ThePEG</span> or any other external library. Some optional hooks for links to other programs are provided, however. The version presented here is the first operational one in the <span style="font-variant:small-caps;">Pythia</span> 8 series. As such it is not yet tested and tuned enough to have reached the same level of maturity as <span style="font-variant:small-caps;">Pythia</span> 6, so we expect the two to coexist for a while. It is only by an increasing use of the new version that it will mature, however, so we encourage a critical try-out, and look forward to feedback. The intention is to release a version in time for comparisons with first LHC data. Thus some areas, like $\gamma\mathrm{p}$ and $\gamma\gamma$ physics, are not yet addressed. Further, some intended processes remain to be implemented. Actually, with the rise of automatic matrix-element code generation and phase-space sampling, input of process-level events via the Les Houches Accord (LHA) [@lha] and with Les Houches Event Files (LHEF) [@lhef] reduces the need to have an extensive process library inside <span style="font-variant:small-caps;">Pythia</span> itself. Thus emphasis is more on providing a good description of subsequent steps of the story, involving elements such as initial- and final-state parton showers, multiple parton–parton interactions, string fragmentation, and decays. The current article provides an introduction to <span style="font-variant:small-caps;">Pythia</span> 8. The programming aspects are covered in more detail in a set of interlinked HTML (or alternatively PHP) pages that comes in the same package as the program files, see below. Much of the physics aspects are unchanged relative to the <span style="font-variant:small-caps;">Pythia</span> 6.4 manual [@pythiasixfour], and so we refer to it and to other physics articles for that. Instead what we here give is an overview for potential users who already have some experience with event generators and want to understand how to get going with <span style="font-variant:small-caps;">Pythia</span> 8. Section \[sec:physics\] contains an ultra-brief summary of the physics of <span style="font-variant:small-caps;">Pythia</span> 8, with particular emphasis on aspects that are different relative to <span style="font-variant:small-caps;">Pythia</span> 6. The program structure (including flow, files, documentation, and a few important warnings) is described in section \[sec:structure\]; summaries of the main user methods, including the event record and particle classes, in section \[sec:main\]. Section \[sec:databases\] is concerned with the databases of flags, modes, parameters, processes, and particle data which exist in <span style="font-variant:small-caps;">Pythia</span> 8. Those who wish to link to external programs, e.g. to gain access to external parton distributions, standard input/output formats, and much more, will find the relevant information in section \[sec:external\]. A brief how-to on getting going is then included in section \[sec:how-to\]. Section \[sec:outlook\] rounds off with an outlook. Physics Summary \[sec:physics\] =============================== This article is not intended to provide a complete description of the physics content. For this we primarily refer to the <span style="font-variant:small-caps;">Pythia</span> 6 manual [@pythiasixfour] and associated literature. We would like to draw attention to some key points of difference, however. Further details are available on the HTML/PHP pages in the program distribution. Some new physics aspects will eventually be covered in separate articles. The physics components are controlled by many parameters. These have been assigned sensible default values, based on previous experience with <span style="font-variant:small-caps;">Pythia</span> 6 and some first studies with the new code. We look forward to more extensive tunes by the experimental community, however. Hard processes -------------- Currently the program only works with $\mathrm{p}\mathrm{p}$, $\overline{\mathrm{p}}\mathrm{p}$, $\mathrm{e}^+\mathrm{e}^-$ and $\mu^+\mu^-$ incoming beams. In particular, there is no provision for $\mathrm{e}\mathrm{p}$ collisions or for incoming photon beams, neither on their own nor as flux around an electron. The list of processes currently implemented is summarised further down; it corresponds to most of the ones in <span style="font-variant:small-caps;">Pythia</span> 6, with the exception of the Supersymmetry and Technicolor sectors, which are yet to come. The cross-section expressions should be identical, but default scale choices have been changed, so that cross sections may be somewhat different for that reason. The default parton distribution remains CTEQ 5L, but ones found in the <span style="font-variant:small-caps;">LhaPdf</span> library [@lhapdf] can easily be linked. It is now possible to use separate PDF sets for the hard interaction, on one hand, and for the subsequent showers and multiple interactions, on the other. Parton showers -------------- The initial- and final-state algorithms are based on the new $p_{\perp}$-ordered evolution introduced in <span style="font-variant:small-caps;">Pythia</span> 6.3 [@ptshowers], while the older mass-ordered ones have not been implemented. It is now additionally possible to have a branching of a photon to a fermion pair as part of the final-state evolution. Already in <span style="font-variant:small-caps;">Pythia</span> 6.3 the initial-state evolution and the multiple interactions were interleaved into one common decreasing $p_{\perp}$ sequence. Now also the final-state evolution is interleaved with the other two. In this context, some of that final-state radiation gets to be associated with dipoles stretched between a final-state parton and the “hole” left by an initial-state one, which therefore now can take a recoil. The initial-state-radiation algorithm remains unchanged in this respect, with recoils always taken by the hard scattering subsystem as a whole. Multiple interactions and beam remnants --------------------------------------- The multiple-interactions machinery as such contains the full functionality introduced in <span style="font-variant:small-caps;">Pythia</span> 6.3 [@newremnants]. Rescaled parton densities are defined after the first interaction, that take into account the nature of the previous partons extracted. Currently there is only one scenario for colour-reconnection in the final state, in which there is a certain probability for the partons of two subscatterings to have their colours interarranged in a way that reduces the total string length. (This is intermediate in character between the original strategy [@zijl] and the more recent ones [@wicke].) The description of beam remnants is based on the new framework. In addition to the standard QCD $2 \to 2$ processes, the possibility of multiple interactions producing prompt photons, charmonia and bottomonia, low-mass Drell-Yan pairs, and $t$-channel $\gamma^*/\mathrm{Z}^0/\mathrm{W}^{\pm}$ exchange is now also included. For dedicated studies of two low-rate processes in coincidence, two hard interactions can now be set in the same event. There are no Sudakov factors included for these two interactions, similarly to normal events with one hard interaction. Hadronisation ------------- Hadronisation is based solely on the Lund string fragmentation framework [@lundreview]; older alternative descriptions have been left out. Particle data have been updated in agreement with the 2006 PDG tables [@pdg]. This also includes a changed content of the scalar meson multiplet. Some updated charm and bottom decay tables have been obtained from the DELPHI and LHCb collaborations. The BE$_{32}$ model for Bose–Einstein effects [@boseeinstein] has been implemented, but is not on by default. Other program components ------------------------ Standardised procedures have been introduced to link the program to various external programs for specific tasks, see section \[sec:external\]. Finally, some of the old jet finders and other analysis routines are made available. Also included is a utility to generate, display and save simple one-dimensional histograms. Program Structure \[sec:structure\] =================================== Program flow ------------ The physics topics that have to come together in a complete event generator can crudely be subdivided into three stages: 1. The generation of a “process” that decides the nature of the event. Often it would be a “hard process”, such as $\mathrm{g} \mathrm{g} \to \mathrm{h}^0 \to \mathrm{Z}^0 \mathrm{Z}^0 \to \mu^+ \mu^- \mathrm{q} \overline{\mathrm{q}}$, that is calculated in perturbation theory, but a priori we impose no requirement that a hard scale must be involved. Only a very small set of partons/particles is defined at this level, so only the main aspects of the event structure are covered. 2. The generation of all subsequent activity on the partonic level, involving initial- and final-state radiation, multiple parton–parton interactions and the structure of beam remnants. Much of the phenomena are under an (approximate) perturbative control, but nonperturbative physics aspects are also important. At the end of this step, a realistic partonic structure has been obtained, e.g. with broadened jets and an underlying-event activity. 3. The hadronisation of this parton configuration, by string fragmentation, followed by the decays of unstable particles. This part is almost completely nonperturbative, and so requires extensive modelling and tuning or, especially for decays, parametrisations of existing data. It is only at the end of this step that realistic events are available, as they could be observed by a detector. This division of tasks is not watertight — parton distributions span and connect the two first steps, to give one example — but it still helps to focus the discussion. (430,370)(-215,10) (-215,350)(215,380)[0.9]{} (0,365)\[\][The User ($\approx$ Main Program)]{} (-215,300)(215,330)[0.9]{} (0,315)\[\][`Pythia`]{} (-215,250)(-170,280)[0.9]{} (-192.5,265)\[\][`Info`]{} (-130,250)(-20,280)[0.9]{} (-75,265)\[\][`Event  process`]{} (20,250)(215,280)[0.9]{} (105,265)\[\][`Event  event`]{} (-215,110)(-85,230)[0.9]{}(-215,200)(-85,200) (-150,215)\[\][`ProcessLevel`]{} (-150,185)\[\][`ProcessContainer`]{} (-150,165)\[\][`PhaseSpace`]{} (-150,145)\[\][`LHAinit, LHAevnt`]{} (-150,125)\[\][`ResonanceDecays`]{} (-65,110)(65,230)[0.9]{}(-65,200)(65,200) (0,215)\[\][`PartonLevel`]{} (0,185)\[\][`TimeShower`]{} (0,165)\[\][`SpaceShower`]{} (0,145)\[\][`MultipleInteractions`]{} (0,125)\[\][`BeamRemnants`]{} (85,110)(215,230)[0.9]{}(85,200)(215,200) (150,215)\[\][`HadronLevel`]{} (150,185)\[\][`StringFragmentation`]{} (150,165)\[\][`MiniStringFrag…`]{} (150,145)\[\][`ParticleDecays`]{} (150,125)\[\][`BoseEinstein`]{} (-130,60)(-20,90)[0.9]{} (-75,75)\[\][`BeamParticle`]{} (20,60)(200,90)[0.9]{} (110,75)\[\][`SigmaProcess, SigmaTotal`]{} (-215,10)(215,40)[0.9]{} (0,25)\[\][`Vec4, Rndm, Hist, Settings, ParticleDataTable, ResonanceWidths, …`]{} (0,350)(0,332) (-150,300)(-150,232) (0,300)(0,232) (150,300)(150,280) (150,280)(150,250)[4]{} (150,250)(150,232) (-192.5,230)(-192.5,248) (-192.5,280)(-192.5,300) (-192.5,300)(-192.5,330)[4]{} (-192.5,330)(-192.5,348) (-107.5,230)(-107.5,248) (-42.5,250)(-42.5,232) (-42.5,280)(-42.5,300) (-42.5,300)(-42.5,330)[4]{} (-42.5,330)(-42.5,348) (42.5,230)(42.5,248) (107.5,250)(107.5,232) (182.5,230)(182.5,248) (182.5,280)(182.5,300) (182.5,300)(182.5,330)[4]{} (182.5,330)(182.5,348) (-107.5,100)(-107.5,108) (-107.5,100)(-107.5,92) (-42.5,100)(-42.5,108) (-42.5,100)(-42.5,92) (42.5,100)(42.5,108) (42.5,100)(42.5,92) (-160,50)(-160,108) (-160,50)(0,50) (0,50)(19,59) The structure of the <span style="font-variant:small-caps;">Pythia</span> 8 generator, as illustrated in Fig. \[fig:generatorstructure\], is based on this subdivision. The main class for all user interaction is called `Pythia`. It calls on the three classes `ProcessLevel`, `PartonLevel` and `HadronLevel`, corresponding to points 1, 2 and 3 above. Each of these, in their turn, call on further classes that perform the separate kinds of physics tasks. Information is flowing between the different program elements in various ways, the most important being the event record, represented by the `Event` class. Actually, there are two objects of this class, one called `process`, that only covers the few partons of the “hard” process of point 1 above (i.e., containing information corresponding to what might be termed the “matrix element” level), and another called `event`, that covers the full story from the incoming beams to the final hadrons. A small `Info` class keeps track of useful one-of-a-kind information, such as kinematical variables of the hard process. There are also two incoming `BeamParticle`s, that keep track of the partonic content left in the beams after a number of interactions and initial-state radiations, and rescales parton distributions accordingly. The process library, as well as parametrisations of total, elastic and diffractive cross sections, are used both by the hard-process selection machinery and the multiple-interactions one. The `Settings` database keeps track of all integer, double, boolean and string variables that can be changed by the user to steer the performance of <span style="font-variant:small-caps;">Pythia</span>, except that `ParticleDataTable` is its own separate database. Finally, a number of utilities can be used just about anywhere, for Lorentz four-vectors, random numbers, jet finding and simple histograms, and for a number of other “minor” tasks. Orthogonally to the subdivision above, there is another, more technical classification, whereby the user interaction with the generator occurs in three phases: - Initialisation, where the tasks to be performed are specified. - Generation of individual events (the “event loop”). - Finishing, where final statistics is made available. Again the subdivision (and orthogonality) is not strict, with many utilities and tasks stretching across the borders, and with no finishing step required for many aspects. Nevertheless, as a rule, these three phases are represented by different methods inside the class of a specific physics task. Program files and documentation ------------------------------- The code is subdivided into a set of files, mainly by physics task. Each file typically contains one main class, but often with a few related helper classes that are not used elsewhere in the program. Normally the files come in pairs. 1. A header file, `.h` in the `include` subdirectory, where the public interface of the class is declared, and inline methods are defined. 2. A source code file, `.cc` in the `src` subdirectory, where the lengthier methods are implemented. During compilation, related dependency files, `.d`, and compiled code, `.o` are created in the `tmp` subdirectory. In part the `.xml` documentation files in the `xmldoc` subdirectory have matching names, but the match is broken by the desire to group topics more by user interaction than internal operation. These files contain information on all settings and particle data, but not in a convenient-to-read format. Instead they are translated into a corresponding set of `.html` files in the `htmldoc` subdirectory and a set of `.php` files in `phpdoc`. The former set can easily be read if you open the `htmldoc/Welcome.html` file in your favourite web browser, but offers no interactivity. The latter set must be installed under a webserver (like a homepage) to function properly, and then provides a simple Graphical User Interface if you open the `phpdoc/Welcome.php` file in a web browser. For output to the <span style="font-variant:small-caps;">HepMC</span> event record format [@hepmc], an interface is provided in the `hepmcinterface` subdirectory. There are also interfaces to allow parton distribution functions to be used from the <span style="font-variant:small-caps;">LhaPdf</span> library [@lhapdf] and hard processes from external programs. The installation procedure is described in a `README` file; it involves running a `configure` script, whereafter an ordinary `Makefile` is used. The former should be invoked with command-line arguments (or be edited) to provide the path to the <span style="font-variant:small-caps;">HepMC</span> library if this is going to be used. Compiled libraries are put in the `lib` subdirectory. Default is to build archive libraries, but optionally also shared-object ones can be built. The standard setup is intended for Linux systems, but a simplified alternative is provided for Windows users. Finally, some examples of main programs, along with input files, or “cards”, for them, are found in the `examples` subdirectory. This directory contains its own `configure` script and `Makefile` which will allow you to build executables, see the `examples/README` file. As above, command-line arguments or brute-force editing allows you to set the <span style="font-variant:small-caps;">LhaPdf</span> and <span style="font-variant:small-caps;">Pythia</span> 6.4 paths, if so required. The executables are placed in the `bin` directory, but with links from `examples`. Important warnings ------------------ Playing with the files in the `examples` subdirectory is encouraged, to familiarise oneself with the program. Modifying the `configure` files may be required during installation. For the rest, files should not be modified, at least not without careful consideration of consequences. In particular, the `.xml` files are set read-only, and should not be tampered with. Interspersed in them, there are lines beginning with `<flag`, `<mode`, `<parm`, `<word`, `<particle`, `<channel`, or `<a`. They contain instructions from which `Settings` and `ParticleDataTable` build up their respective databases of user-accessible variables, see further below. Any stupid changes here will cause difficult-to-track errors! Further, sometimes you will see two question marks, “??”, in the text or code. This is for internal usage, to indicate loose ends or preliminary thoughts. Please disregard. Main Program and Event Information \[sec:main\] =============================================== The `Pythia` class ------------------ The `Pythia` class is the main means of communication between the user and the event-generation process. We here present the key methods for the user to call, ordered by context. Firstly, at the top of the main program, the proper header file must be included:\ [ ]{}`#include Pythia.h`\ To simplify typing, it also makes sense to declare\ [ ]{}`using namespace Pythia8;`\ Given this, the first step in the main program is to create a generator object, e.g. with\ [ ]{}`Pythia pythia;`\ In the following we will assume that the `pythia` object has been created with this name, but of course you are free to pick another one. When this object is declared, the `Pythia` constructor initialises all the default values for the `Settings` and the `ParticleDataTable` data bases. These data are now present in memory and can be modified in a number of ways before the generator is initialised (see below). Most conveniently, <span style="font-variant:small-caps;">Pythia</span>’s settings and parameters can be changed by the two methods\ [ ]{}`pythia.readString(string);`\ for changing a single variable, and\ [ ]{}`pythia.readFile(fileName);`\ for changing a set of variables, one per line in the input file. The allowed form for a string/line will be explained as we consider the databases in the next section. Further, methods will be introduced to list all or only the changed settings and particle data. At this stage you can also optionally hook up with some external facilities, see section \[sec:external\]. After this, in the initialisation call all remaining details of the generation are to be specified. The `pythia.init(...)` method allows a few different input formats, so you can pick the one convenient for you:\ [ ]{}`pythia.init(idA, idB, eA, eB);`\ lets you specify the identities and energies of the two incoming beam particles, with A (B) assumed moving in the $+z$ ($-z$) direction;\ [ ]{}`pythia.init(idA, idB, eCM);`\ is similar, but you specify the CM energy, and you are assumed in the rest frame;\ [ ]{}`pythia.init(LHAinit*, LHAevnt*);`\ assumes LHA initialisation information is available in an `LHAinit` class object, and that LHA event information will be provided by the `LHAevnt` class object, see below;\ [ ]{}`pythia.init(fileName);`\ assumes that the file obeys the LHEF standard format and that information can be extracted from it accordingly; and finally\ [ ]{}`pythia.init();`\ will take its values from the beam specification stored in the `Settings` database. It is when the `init(...)` call is executed that all the settings values are propagated to the various program elements, and used to precalculate quantities that will be used at later stages of the generation. Further settings changed after the `init(...)` call will be ignored (unless methods are used to force a partial or complete re-initialisation). By contrast, the particle properties database is queried all the time, and so a later change would take effect immediately, for better or worse. The bulk of the code is concerned with the event generation proper. However, all the information on how this should be done has already been specified. Therefore only a command\ [ ]{}`pythia.next();`\ is required to generate the next event. This method would be located inside an event loop, where a required number of events are to be generated. The key output of the `pythia.next()` command is the event record found in `pythia.event`, see below. A process-level summary of the event is stored in `pythia.process`. When problems are encountered, in `init(...)` or `next()`, they can be assigned one of three degrees of severity. Abort is the highest. In that case the call could not complete its tasks, and returns the value `false`. If this happens in `init(...)` it is then not possible to generate any events at all. If it happens in `next()` only the current event must be skipped. In a few cases the abort may be predictable and desirable, e.g. when a file of LHA events comes to an end. Errors are less severe, and the program can usually work around them, e.g. by backing up one step and trying again. Should that not succeed, an abort may result. Warnings are of informative character only, and do not require any corrective actions (except, in the longer term, to find more reliable algorithms). At the end of the generation process, you can call\ [ ]{}`pythia.statistics();`\ to get some run statistics, both on cross sections for the subprocesses generated and on the number of aborts, errors and warnings issued. The event record ---------------- The `Event` class for event records is not much more than a wrapper for a vector of `Particle`s. This vector can expand to fit the event size. The index operator is overloaded, so that `event[i]` corresponds to the `i`’th particle of an `Event` object called `event`. For instance, given that the PDG identity code [@pdg] of a particle is provided by the `id()` method, `event[i].id()` returns the identity of the `i`’th particle. Line 0 is used to represent the event as a whole, with its total four-momentum and invariant mass, but does not form part of the event history, and only contains redundant information. When you translate to another event-record format where the first particle is assigned index 1, such as <span style="font-variant:small-caps;">HepMC</span>, this line should therefore be dropped so as to keep the rest of the indices synchronised. It is only with lines 1 and 2, which contain the two incoming beams, that the history tracing begins. That way unassigned mother and daughter indices can be put 0 without ambiguity. In this section, first the `Particle` methods are surveyed, and then the further aspects of the event record. ### The particle A `Particle` corresponds to one entry/slot/line in the event record. Its properties therefore mix ones belonging to a particle-as-such, like its identity code or four-momentum, and ones related to the event-as-a-whole, like which mother it has. The following properties are stored for each particle, listed by the member functions you can use to extract the information: - `id()` : the identity of a particle, according to the PDG particle codes. - `status()` : status code. The full set of codes provides information on where and why a given particle was produced. The key feature is that a particle is assigned a positive status code when it is created, which then is negated if later it branches into other particles. The mechanism of this branching can be inferred from the status code of the daughters. Thus, at any given stage of the event-generation process, the current final state consists of the particles with positive status code. - `mother1(), mother2()` : the indices in the event record where the first and last mothers are stored, if any. A few different cases are possible, to allow for one or many mothers. The `motherList(i)` method (see below) can return a vector with all the mother indices, based on this info. - `daughter1(), daughter2()` : the indices in the event record where the first and last daughters are stored, if any. A few different cases are possible, to allow for one or many daughters. The `daughterList(i)` method (see below) can return a vector with all the daughter indices, based on this info. - `col(), acol()` : the colour and anticolour tags, LHA style. - `px(), py(), pz(), e()` : the particle four-momentum components (in GeV, with $c = 1$), alternatively extracted as a `Vec4 p()`. - `m()` : the particle mass (in GeV). - `scale()` : the scale at which a parton was produced (in GeV); model-specific but relevant in the processing of an event. - `xProd(), yProd(), zProd(), tProd()` : the production vertex coordinates (in mm or mm/$c$), alternatively extracted as a `Vec4 vProd()`. - `tau()` : the proper lifetime (in mm/$c$). The same method names, with a value inserted between the brackets, set these quantities. In addition, a number of derived quantities can easily be obtained, but cannot be set, such as: - `isFinal()` : `true` for a remaining particle, i.e. one with positive status code, else `false`. - `pT(), pT2()` : (squared) transverse momentum. - `mT(), mT2()` : (squared) transverse mass. - `pAbs(), pAbs2()` : (squared) three-momentum magnitude. - `theta(), phi()` : polar and azimuthal angle (in radians). - `y(), eta()` : rapidity and pseudorapidity. - `xDec(), yDec(), zDec(), tDec()` : the decay vertex coordinates, assuming free-streaming propagation, alternatively extracted as a `Vec4 vDec()`. Each `Particle` contains a pointer to the respective `ParticleDataEntry` object in the particle data tables. This pointer gives access to properties of the particle species as such. It is there mainly for convenience, and should be thrown if an event is written to disk, to avoid any problems of object persistency. This pointer is used by member functions such as: - `name()` : the name of the particle, as a string. - `spinType()` : $2 s + 1$, or 0 where undefined spin. - `charge(), chargeType()` : charge, and three times it to make an integer. - `isCharged(), isNeutral()` : `bool`s for charged or not. - `colType()` : 0 for colour singlets, 1 for triplets, $-1$ for antitriplets and 2 for octets. - `m0()` : the nominal mass of the particle species. ### Other methods in the event record While the `Particle` vector is the key component of an `Event`, a few further methods are available. The event size can be found with `size()`, i.e. valid particles are stored in the range $0 \leq $`i`$ <$ `event.size()`. A listing of the whole event is obtained with `list()`. The basic identity, status, mother, daughter, colour, four-momentum and mass data are always given, but optional arguments can be set to provide further information, on the complete lists of mothers and daughters, and on production vertices. The user would normally be concerned with the `Event` object that is a public member `event` of the `Pythia` class. Thus `pythia.event[i].id()` would be used to return the identity of the `i`’th particle, and `pythia.event.size()` to give the size of the event record. A `Pythia` object contains a second event record for the hard process alone, similar to the LHA process specification, called `process`. This record is used as input for the generation of the complete event. Thus one may e.g. call either `pythia.process.list()` or `pythia.event.list()`. To distinguish those two rapidly at visual inspection, the “Pythia Event Listing” header is printed out differently, adding either “(hard process)” or “(complete event)”. There are also a few methods with an individual particle index `i` as input, but requiring some search operations in the event record, and therefore not possible to define as methods of the `Particle` class. The most important ones are `motherList(i)`, `daughterList(i)` and `sisterList(i)`. These return a `vector<int>` containing a list of all the mothers, daughters or sisters of a particle. This list may be empty or arbitrarily large, and is given in ascending order. One data member in an Event object is used to keep track of the largest `col()` or `acol()` tag set so far, so that new ones do not clash. The event record also contains two further sets of vectors. These are intended for the expert user only, so only a few words on each. The first is a vector of junctions, i.e. vertices where three string pieces meet. This list is often empty or else contains only a very few per event. The second is a storage area for parton indices, classified by subsystem. Such information is needed to interleave multiple interactions, initial-state showers, final-state showers and beam remnants. It can also be used in the hadronisation. Other event information ----------------------- A set of one-of-a-kind pieces of event information is stored in the `info` object (an instance of the class `Info`) in the `Pythia` class. This is mainly intended for processes generated internally, but some of the information is also available for external processes. You can use `pythia.info.method()` to extract e.g. the following information: - `list()` : list some information on the current event. - `eCM(), s()` : the cm energy and its square. - `name(), code()` : the name and code of the subprocess. - `id1(), id2()` : the identities of the two partons coming in to the hard subprocess. - `x1(), x2()` : $x$ fractions of the two partons coming in to the hard subprocess. - `pdf1(), pdf2(), QFac(), Q2Fac()` : parton densities $x \, f_i(x,Q^2 )$ evaluated for the two incoming partons, and the associated factorisation scale $Q$ and its square. - `mHat(), sHat(), tHat(), uHat()` : the invariant mass of the hard subprocess and the Mandelstam variables for $2 \to 2$ processes. - `pTHat(), thetaHat()` : transverse momentum and polar scattering angle of the hard subprocess for $2 \to 2$ processes. - `alphaS(), alphaEM(), QRen(), Q2Ren()` : $\alpha_{\mathrm{s}}$ and $\alpha_{\mathrm{em}}$ values for the hard process, and the associated renormalisation scale $Q$ and its square. - `nTried(), nAccepted(), sigmaGen(), sigmaErr()` : the number of trial and accepted events, and the resulting estimated cross section and estimated statistical error, in units of mb, summed over the included processes. In other classes there are also methods that can be called to do a sphericity or thrust analysis or search for jets with a clustering or simple cone jet finder. These take the event record as input. Databases \[sec:databases\] =========================== Inevitably one wants to be able to modify the default behaviour of a generator. Currently there are two <span style="font-variant:small-caps;">Pythia</span> 8 databases with modifiable values. One deals with general settings, the other specifically with particle data. The key method to set a new value is\ [ ]{}`pythia.readString(string);`\ The typical form of a string is\ [ ]{}`variable = value`\ where the equal sign is optional and the variable begins with a letter for settings and a digit for particle data. A string not beginning with either is considered as a comment and ignored. Therefore inserting an initial !, \#, \$, %, or another such character, is a good way to comment out a command. For non-commented strings, the match of the name to the database is case-insensitive. Strings that do begin with a letter or digit and still are not recognised cause a warning to be issued, unless a second argument `false` is used in the call. Any further text after the value is ignored, so the rest of the string can be used for any comments. For variables with an allowed range, values below the minimum or above the maximum are set at the respective border. For `bool` values, the following notation may be used interchangeably: `true` = `on` = `yes` = `ok` = 1. Everything else gives `false` (including but not limited to `false`, `off`, `no` and `0`). The `readString(...)` method is convenient for changing one or two settings, but becomes cumbersome for more extensive modifications. In addition, a recompilation and relinking of the main program is necessary for any change of values. Alternatively, the changes can therefore be collected in a file, for historical reasons often called a “card file”, where each line is a character string defined in the same manner as above (without quotation marks). The whole file can then be read and processed with a command\ [ ]{}`pythia.readFile(fileName);`\ As above, comments can be freely interspersed. Settings -------- We distinguish four kinds of user-modifiable variables, by the way they have to be stored: 1. A `Flag` is an on/off switch, and is stored as a `bool`. 2. A `Mode` corresponds to an enumeration of separate options, and is stored as an `int`. 3. A `Parm` — short for parameter — takes a continuum of values, and is stored as a `double`. 4. A `Word` is a text string (with no embedded blanks) and is stored as as a `string`. Collectively the four above kinds of variables are called settings. Not surprisingly, the class that stores them is called `Settings`. Each variable stored in `Settings` is associated with a few pieces of information. These are: - The variable name, of the form `class:name` (or `file:name`, or `task:name`, usually these agree), e.g. `TimeShower:pTmin`. - The default value, set in the original declaration, and intended to represent a reasonable choice. This value is not user modifiable. - The current value. During construction of the `Settings` object, this value is set equal to the default value. It can subsequently be modified, e.g. by the `pythia.readString()` or `pythia.readFile()` methods discussed above. During the `pythia.init()` initialisation this value will be stored as a local copy in the class(es) where it is used, and thereby also control the subsequent generation. - An allowed range of values, represented by meaningful minimum and maximum values. This has no sense for a flag or a word, is usually rather well-defined for a mode, but less so for a parameter. Either of the minimum and maximum may be left free, giving an open-ended range. Often the allowed range exaggerates the uncertainty in our current knowledge, so as not to restrict too much what the user can do. All the same, this information should not be modified by the user. Technically, the `Settings` class is implemented with the help of four separate maps, one for each kind of variable, with the name used as key. The default values are taken from the `.xml` files in the `xmldoc` subdirectory. The `Settings` class is purely static, i.e. exists only as one global copy, that you can interact with directly by `Settings::command(argument)`. However, a `settings` object is a public member of the `Pythia` class, so an alternative notation would be `pythia.settings.command(argument)`. As already mentioned, for input the `pythia.readString(...)` method is to be preferred, since it also can handle particle data. A typical example would be\ [ ]{}`pythia.readString(TimeShower:pTmin = 1.0);` You may obtain a listing of all variables in the database by calling\ [ ]{}`pythia.settings.listAll();`\ The listing is strictly alphabetical, which at least means that names in the same area are kept together, but otherwise may not be so well-structured: important and unimportant ones will appear mixed. A useful alternative is\ [ ]{}`pythia.settings.listChanged();`\ which will only print out a list of those variables that differ from their defaults. Processes --------- All internal processes available in <span style="font-variant:small-caps;">Pythia</span> 8 can be switched on and off via the ordinary settings machinery just discussed, using flags of the generic type `ProcessGroup:ProcessName`. A complete list of processes currently implemented is given in Table \[processes\]. By default all processes are off. A whole group can be turned on by a `ProcessGroup:all = on` command, then overriding the individual flags. Note that processes in the `SoftQCD` group are of a kind that cannot be input via the LHA, while essentially all other kinds could. Each process is assigned an integer code. This code is not used in the internal administration of events; it is only intended to allow a simpler user separation of different processes. Also the process name is available, as a string. For many processes it makes sense to apply phase space cuts. The ones currently available (in the `Settings` database) in particular include - `PhaseSpace:mHatMin, PhaseSpace:mHatMax` : the range of invariant masses of the scattering process. - `PhaseSpace:pTHatMin, PhaseSpace:pTHatMax` : the range of transverse momenta in the rest frame of the process for $2 \to 2$ and $2 \to 3$ processes (for each of the products). In addition, for any resonance with a Breit-Wigner mass distribution, the allowed mass range of that particle species is taken into account, both for $2 \to 1$, $2 \to 2$ and $2 \to 3$ processes, thereby providing a further cut possibility. Note that the `SoftQCD` processes do not use any cuts but generate their respective cross sections in full. Particle data ------------- The following particle properties are stored in the `ParticleDataTable` class for a given PDG particle identity code `id`, here presented by the method used to access this property: - `name(id)` : particle and antiparticle names are stored separately, the sign of `id` determines which of the two is returned, with “void” used to indicate the absence of an antiparticle. - `hasAnti(id)` : `bool` whether a distinct antiparticle exists or not. - `spinType(id)` : $2 s + 1$ for particles with defined spin, else 0. - `chargeType(id)` : three times the charge (to make it an integer); can also be read as a `double charge(id) = chargeType(id)/3`. - `colType(id)` : the colour type, with 0 uncoloured, 1 triplet, $-1$ antitriplet and 2 octet. - `m0(id)` : the nominal mass $m_0$ (in GeV). - `mWidth(id)` : the width $\Gamma$ of the Breit-Wigner mass distribution (in GeV). - `mMin(id), mMax(id)` : the allowed mass range generated by the Breit-Wigner, $m_{\mathrm{min}} < m < m_{\mathrm{max}}$ (in GeV). - `tau0(id)` : the nominal proper lifetime $\tau_0$ (in mm/$c$). - `constituentMass(id)` : the constituent mass for a quark, hardcoded as $m_{\mathrm{u}} = m_{\mathrm{d}} = 0.325$, $m_{\mathrm{s}} = 0.50$, $m_{\mathrm{c}} = 1.60$ and $m_{\mathrm{b}} = 5.0$ GeV, for a diquark the sum of quark constituent masses, and for everything else the same as the ordinary mass. - `mRun(id, massScale)` : the running mass for quarks, else the same as the nominal mass. - `mayDecay(id)` : a flag telling whether a particle species may decay or not, offering the main user switch (whether a given particle of this kind then actually will decay also depends on other flags in the `ParticleDecays` class). Similar methods can also be used to set most of these properties. Each particle kind in the `ParticleDataTable` also has a a vector of `DecayChannel`s associated with it. The following properties are stored for each decay channel: - `onMode()` : whether a channel is on (1) or off (0), or on only for particles (2) or antiparticles (3). - `bRatio()` : the branching ratio. - `meMode()` : the mode of processing this channel, possibly with matrix-element information; 0 gives isotropic phase space. - `multiplicity()` : the number of decay products in a channel, at most 8. - `product(i)` : a list of the decay products, 8 products $0 \leq $`i`$ < 8$, with trailing unused ones set to 0. The original particle data and decay table is read in from the `ParticleData.xml` file. The `ParticleDataTable` class is purely static, i.e. exists as one global copy, that you can interact directly with by `ParticleDataTable::command(argument)`. However, a `particleData` object of the `ParticleDataTable` class is a public member of the `Pythia` class, which offers an alternatively notation. As already mentioned, for input the `pythia.readString(string)` method is to be preferred, since it also can handle settings. It is only the form of the `string` that needs to be specified slightly differently than for settings, as\ [ ]{}`id:property = value`.\ The `id` part is the standard PDG particle code, i.e. a number, and `property` is one of the ones already described above, with a few minor differences: `name`, `antiName`, `spinType`, `chargeType`, `colType`, `m0`, `mWidth`, `mMin`, `mMax`, `tau0`, `mayDecay`, `isResonance`, `isVisible`, `doExternalDecay`, and `doForceWidth`. As before, several commands can be stored as separate lines in a file, say\ [ ]{}`111:name = piZero ! normal notation pi0`\ [ ]{}`3122:mayDecay = false ! Lambda0 stable`\ [ ]{}`431:tau0 = 0.15 ! D_s proper lifetime`\ and then be read with `pythia.readFile(fileName)`. For major changes of the properties of a particle, the above one-at-a-time changes can become rather cumbersome. Therefore a few extended input formats are available, where a whole set of properties can be given after the equal sign, separated by blanks and/or by commas. One line like\ [ ]{}`id:all = name antiName spinType chargeType colType m0 mWidth mMin mMax tau0`\ replaces all the current information on the particle itself, but keeps its decay channels, if any, while using `new` instead of `all` also removes any previous decay channels. (The flags `mayDecay`, `isResonance`, `isVisible`, `doExternalDecay`, and `doForceWidth` are in either case reset to their defaults and would have to be changed separately.) In order to change the decay data, the decay channel number needs to be given right after the particle number, i.e. the command form becomes\ [ ]{}`id:channel:property = value`.\ Recognised properties are `onMode`, `bRatio`, `meMode` and `products`, where the latter expects a list of all the decay products, separated by blanks, up until the end of the line, or until a non-number is encountered. The property `all` will replace all the information on the channel, i.e.\ [ ]{}`id:channel:all = onMode bRatio meMode products `\ To add a new channel at the end, use\ [ ]{}`id:addChannel = onMode bRatio meMode products `\ To remove all existing channels and force decays into one new channel, use\ [ ]{}`id:oneChannel = onMode bRatio meMode products `\ A first `oneChannel` command could be followed by several subsequent `addChannel` ones, to build up a completely new decay table for an existing particle.\ It is currently not possible to remove a channel selectively, but setting its branching ratio vanishing is as effective. Often one may want to allow only a specific subset of decay channels for a particle. This can be achieved e.g. by a repeated use of `id:channel:onMode` commands, but there also is a set of commands that initiates a loop over all decay channels and allows a matching to be carried out. The `id:onMode` command can switch `on` or `off` all channels. The `id:onIfAny` and `id:offIfAny` will switch on/off all channels that contain any of the enumerated particles. For instance\ [ ]{}`23:onMode = off`\ [ ]{}`23:onIfAny = 1 2 3 4 5`\ first switches off all $\mathrm{Z}^0$ decay modes and then switches back on any that contains one of the five lighter quarks. Other methods are `id:onIfAll` and `id:offIfAll`, and `id:onIfMatch` and `id:offIfMatch`, where all the enumerated products must be present for a decay channel to be switched on/off. The difference is that the former two allow further non-matched particles in a decay channel while the latter two do not. There are also further methods to switch on channels selectively either for the particle or for the antiparticle. When a particle is to be decayed, the branching ratios of the allowed channels are always rescaled to unit sum. There are also methods for by-hand rescaling of branching ratios. You may obtain a listing of all the particle data by calling\ [ ]{}`pythia.particleData.listAll()`.\ The listing is by increasing `id` number. To list only those particles that have been changed, instead use\ [ ]{}`pythia.particleData.listChanged()`.\ To list only one specific particle `id`, use `list(id)`. It is also possible to `list` a `vector<int>` of `id`’s. Links to external programs \[sec:external\] =========================================== While <span style="font-variant:small-caps;">Pythia</span> 8 itself is self-contained and can be run without reference to any external library, often one does want to make use of other programs that are specialised on some aspect of the generation process. The HTML/PHP documentation accompanying the code contains full information on how the different links should be set up. Here the purpose is mainly to point out the possibilities that exist. The Les Houches interface ------------------------- The Les Houches Accord for user processes (LHA) [@lha] is the standard way to input parton-level information from a matrix-elements-based generator into <span style="font-variant:small-caps;">Pythia</span>. The conventions for which information should be stored has been defined in a Fortran context, as two commonblocks. Here a C++ equivalent is defined, as two separate classes. The `LHAinit` and `LHAevnt` classes are base classes, containing reading and printout methods, plus a pure virtual method `set()` each. Derived classes have to provide these two virtual methods to do the actual work. Currently the only examples are for reading information at runtime from the respective Fortran commonblock or for reading it from a Les Houches Event File (LHEF) [@lhef]. The `LHAinit` class stores information equivalent to the `/HEPRUP/` commonblock, as required to initialise the event-generation chain. The `LHAevnt` class stores information equivalent to the `/HEPEUP/` commonblock, as required to hand in the next parton-level configuration for complete event generation. The `LHAinitFortran` and `LHAevntFortran` are two derived classes, containing `set()` members that read the respective LHA Fortran commonblock for initialisation and event information. This can be used for a runtime link to a Fortran library. As an example, an interface is provided to the <span style="font-variant:small-caps;">Pythia</span> 6.4 process library. The `LHAinitLHEF` and `LHAevntLHEF` are two other derived classes, that can read a file with initialisation and event information, assuming that the file has been written in the LHEF format. You do not need to declare these classes yourself, since a shortcut is provided by the `pythia.init(fileName)` command. If you create `LHAinit` and `LHAevnt` objects yourself, pointers to those should be handed in with the `init(...)` call, then of the form `pythia.init(LHAinit*, LHAevnt*)`. Semi-internal processes and resonances -------------------------------------- When you implement new processes via the Les Houches Accord you do all flavour, colour and phase-space selection externally, before your process-level events are input for further processing by <span style="font-variant:small-caps;">Pythia</span>. However, it is also possible to implement a new process in exactly the same way as the internal <span style="font-variant:small-caps;">Pythia</span> ones, thus making use of the internal phase-space selection machinery to sample an externally provided cross-section expression. The matrix-element information has to be put in a new class that derives from one of the existing classes, `Sigma1Process` for $2 \to 1$ processes, `Sigma2Process` for $2 \to 2$ ones, and `Sigma3Process` for $2 \to 3$ ones, which in their turn derive from the `SigmaProcess` base class. Note that `Pythia` is rather good at handling the phase space of $2 \to 1$ and $2 \to 2$ processes, is more primitive for $2 \to 3$ ones and does not at all address higher multiplicities. This limits the set of processes that you can implement in this framework. The produced particles may be resonances, however, so it is possible to end up with bigger “final” multiplicities through sequential decays, and to include further matrix-element weighting in those decays. In your new class you have to implement a number of methods. Chief among them is one to return the matrix-element weight for an already specified kinematics configuration and another one to set up the final-state flavours and colour flow of the process. Further methods exist, some of more informative character, such as providing the name of the process. Should you actually go ahead, it is strongly recommended to shop around for a similar process that has already been implemented, and to use that existing code as a template. Once a class has been written, a pointer of type `SigmaProcess*` to a `new` instance of your class needs to be created in the main program, and handed in with the `pythia.setSigmaPtr(...)` method. From there on the process will be handled on equal footing with internally implemented processes. If your new process introduces a new particle you have to add it and its decay channels to the particle database, as already explained. This only allows for a fixed width and fixed branching ratios, however, with only some minor generalisations. To obtain a dynamical calculation, where the width and the branching ratios can vary as a function of the currently chosen mass, you must also create a new class for it that derives from the `ResonanceWidths` class. In it you have to implement a method that returns the partial width for each of the possible decay channels. The structure is simpler than for the `SigmaProcess` case, but again it may be convenient to use a similar existing resonance as a template. You then hand in a pointer to an instance of this new class with the `pythia.setResonancePtr(...)` method. Parton distribution functions ----------------------------- The `PDF` class is the base class for all parton distribution function parametrisations, from which specific `PDF` classes are derived. Currently the selection of sets that comes with the program is very limited; for protons only CTEQ 5L (default) and GRV 94L are available. However, a built-in interface to the <span style="font-variant:small-caps;">LhaPdf</span> library [@lhapdf] allows a much broader selection, if only <span style="font-variant:small-caps;">LhaPdf</span> is linked together with <span style="font-variant:small-caps;">Pythia</span>. Should this not be enough, it is possible to write your own class derived from the `PDF` base class, wherein you implement the `xfUpdate(...)` member to do the actual updating of PDFs. Once you have created two distinct `PDF` objects, `pdfA` and `pdfB`, you should supply pointers to these as arguments in a `pythia.setPDFPtr(pdfA*, pdfB*)` call. A word of warning: to switch to a new PDF set implies that a complete retuning of the generator may be required, since the underlying-event activity from multiple interactions and parton showers is changed. There is an option that allows a replacement of the PDF for the hard process only, so that this is not required. Inconsistent but convenient. External decay packages ----------------------- While `Pythia` is set up to handle any particle decays, decay products are often (but not always) distributed isotropically in phase space, i.e. polarisation effects and nontrivial matrix elements usually are neglected in <span style="font-variant:small-caps;">Pythia</span>. Especially for the $\tau$ lepton and for some $\mathrm{B}$ mesons it is therefore common practice to rely on dedicated decay packages [@tauola; @evtgen]. To this end, `DecayHandler` is a base class for the external handling of decays. The user-written derived class is called if a pointer to it has been given with the `pythia.setDecayPtr(DecayHandler*, vector<int>)` method. The second argument to this method should contain the `id` codes of all the particles that should be decayed by the external program. It is up to the author of the derived class to send different of these particles on to separate packages, if so desired. The `decay(...)` method in the user-written `DecayHandler` class should do the decay, or return `false` if it fails. In the latter case `Pythia` will try to do the decay itself. Thus one may implement some decay channels externally and leave the rest for `Pythia`, assuming the `Pythia` decay tables are adjusted accordingly. User hooks ---------- Sometimes it may be convenient to step in during the generation process: to modify the built-in cross sections, to veto undesirable events or simply to collect statistics at various stages of the evolution. There is a base class `UserHooks` that gives you this access at a few selected places. This class in itself does nothing; the idea is that you should write your own derived class for your task. A few very simple derived classes come with the program, mainly as illustration. There are four distinct sets of routines. Ordered by increasing complexity, rather than by their appearance in the event-generation sequence, they are: - Ones that gives you access to the event record in between the process-level and parton-level steps, or in between the parton-level and hadron-level ones. You can study the event record and decide whether to veto this event. - Ones that allow you to set a scale at which the combined multiple-interactions, initial-state and final-state parton-shower downwards evolution in $p_{\perp}$ is temporarily interrupted, so the event can be studied and either vetoed or allowed to continue the evolution. - Similar ones that instead gives you access after the first few parton-shower branchings of the hardest subprocess. - Ones that gives you access to the properties of the trial hard process, so that you can modify the internal <span style="font-variant:small-caps;">Pythia</span> cross section by your own correction factors. Random-number generators ------------------------ `RndmEngine` is a base class for the external handling of random-number generation. The user-written derived class is called if a pointer to it has been handed in with the `pythia.setRndmEnginePtr(RndmEngine*)` method. Since the default Marsaglia-Zaman algorithm is quite good, there is absolutely no physics reason to replace it, but this may still be required for consistency with other program elements in big experimental frameworks. The <span style="font-variant:small-caps;">HepMC</span> event format -------------------------------------------------------------------- The <span style="font-variant:small-caps;">HepMC</span> event format [@hepmc] is a standard format for the storage of events in several major experiments. The translation from the <span style="font-variant:small-caps;">Pythia</span> 8 `Event` format should be done after `pythia.next()` has generated an event. Therefore there is no need for a tight linkage, but only to call the\ `HepMC::I_Pythia8::fill_next_event( pythia.event, hepmcevt )`\ conversion routine from the main program written by the user. Version 1 of <span style="font-variant:small-caps;">HepMC</span> makes use of the <span style="font-variant:small-caps;">CLHep</span> library [@clhep] for four-vectors, while version 2 is standalone; this requires some adjustments in the interface code based on which version is used. SUSY parameter input -------------------- <span style="font-variant:small-caps;">Pythia</span> 8 does not contain a machinery for calculating masses and couplings of supersymmetric particles from some small set of input parameters. Instead the SUSY Les Houches Accord (SLHA) [@slha] is used to provide this information, as calculated by some external program. You need to supply the name of the file where the SLHA information is stored, in an appropriate setting, and then the rest is taken care of automatically. (Or at least will be, once SUSY processes are implemented.) Parton showers -------------- It is possible to replace the existing timelike and/or spacelike showers in the program by your own. This is truly for experts, since it requires a rather strict adherence to a wide set of rules. These are described in detail in the HTML/PHP documentation accompanying the code. The <span style="font-variant:small-caps;">Vincia</span> program [@vincia] offers a first example of a plug-in of an external (timelike) shower. Getting Going \[sec:how-to\] ============================ After you download the `pythia8100.tgz` (or later) package from the <span style="font-variant:small-caps;">Pythia</span> webpage,\ [ ]{}`http://www.thep.lu.se/`$\sim$`torbjorn/Pythia.html`\ you can unpack it with `tar xvfz pythia8100.tgz`, into a new subdirectory `pythia8100`. The rest of the installation procedure is described in the `README` file in that directory. It is assumed you are on a Linux system; so far there is hardly any multiplatform support. After this, the main program is up to the user to write. A worksheet (found on the webpage) takes you through as step-by-step procedure, and sample main programs are provided in the `examples` subdirectory. These programs are included to serve as inspiration when starting to write your own program, by illustrating the principles involved. The information available if you open `htmldoc/Welcome.html` in your web browser will help you explore the program possibilities further. If you install the `phpdoc` subdirectory under a web server you will also get extra help to build a file of commands to the `Settings` and `ParticleDataTable` machineries, to steer the execution of your main program. Such “cards files” are separate from the main programs proper, so that minor changes can be made without any recompilation. It is then convenient to collect in the same place some run parameters, such as the number of events to generate, that could be used inside the main program. Therefore some such have been predefined, e.g.`Main:numberOfEvents`. Whether they actually are used is up to the author of a main program to decide. Outlook \[sec:outlook\] ======================= As already explained in the introduction, <span style="font-variant:small-caps;">Pythia</span> 8.1 is not yet a complete replacement of <span style="font-variant:small-caps;">Pythia</span> 6.4, but it is getting there, and already contains some new features not found elsewhere. In many cases the quality of the physics should be comparable between the two versions, but obviously the objective is that soon <span style="font-variant:small-caps;">Pythia</span> 8 should offer the overall better alternative. This will occur by further improvements of the existing framework and by the gradual addition of new features. The support and kind hospitality of the SFT group at CERN is gratefully acknowledged by TS. Mikhail Kirsanov has developed the configure files, the makefiles and the interface to <span style="font-variant:small-caps;">HepMC</span>, and made several valuable suggestions. Ben Lloyd has written the PHP webpage framework. Bertrand Bellenot has provided a simple makefile for Win32/NMAKE. Marc Montull has helped write the extended Higgs sector. SM and PS are supported by Fermi Research Alliance, LLC, under Contract No. DE-AC02-07CH11359 with the United States Department of Energy. This work was supported in part by the European Union Marie Curie Research Training Network MCnet under contract MRTN-CT-2006-035606. [99]{} T. Sjöstrand, Computer Physics Commun. [**27**]{} (1982) 243, [**28**]{} (1983) 229, [**39**]{} (1986) 347;\ T. Sjöstrand and M. Bengtsson, Computer Physics Commun. [**43**]{} (1987) 367 H.-U. Bengtsson, Computer Physics Commun. [**31**]{} (1984) 323;\ H.-U. Bengtsson and G. Ingelman, Computer Physics Commun. [**34**]{} (1985) 251;\ H.-U. Bengtsson and T. Sjöstrand, Computer Physics Commun. [**46**]{} (1987) 43;\ T. Sjöstrand, Computer Physics Commun. [**82**]{} (1994) 74 T. Sjöstrand, P. Edén, C. Friberg, L. Lönnblad, G. Miu, S. Mrenna and E. Norrbin, Computer Physics Commun. [**135**]{} (2001) 238 T. Sjöstrand, S. Mrenna and P. Skands, JHEP [**05**]{} (2006) 026 \[hep-ph/0603175\] L. Lönnblad, Computer Physics Commun. [**118**]{} (1999) 213;\ M. Bertini, L. Lönnblad and T. Sjöstrand, Computer Physics Commun. [**134**]{} (2001) 365 see webpage `http://www.thep.lu.se/ThePEG/` S. Gieseke, A. Ribon, M.H. Seymour, P. Stephens and B.R. Webber, JHEP [**02**]{} (2004) 005;\ see webpage `http://hepforge.cedar.ac.uk/herwig/` E. Boos et al., in the Proceedings of the Workshop on Physics at TeV Colliders,\ Les Houches, France, 21 May - 1 Jun 2001 \[hep-ph/0109068\] J. Alwall et al., Computer Physics Comm. [**176**]{} (2007) 300 M.R. Whalley, D. Bourilkov and R.C. Group, in ‘HERA and the LHC’,\ eds. A. De Roeck and H. Jung, CERN-2005-014, p. 575 \[hep-ph/0508110\] T. Sjöstrand and P. Skands, Eur. Phys. J [**C39**]{} (2005) 129 T. Sjöstrand and P. Skands, JHEP [**03**]{} (2004) 053 T. Sjöstrand and M. van Zijl, Phys. Rev. [**D36**]{} (1987) 2019 P. Skands and D. Wicke, Eur. Phys. J. [**C52**]{} (2007) 133 B. Andersson, G. Gustafson, G. Ingelman and T. Sjöstrand, Phys. Rep. [**97**]{} (1983) 31 Particle Data Group, W.-M. Yao et al., J. Phys. [**G33**]{} (2006) 1 L. Lönnblad and T. Sjöstrand, Eur. Phys. J. [**C2**]{} (1998) 165 M. Dobbs and J.B. Hansen, Computer Physics Comm. [**134**]{} (2001) 41 P. Skands et al., JHEP [**07**]{} (2004) 036 S. Jadach, Z. Was, R. Decker and J.H. Kühn, Computer Physics Commun. [**76**]{} (1993) 361 D.J. Lange, Nucl. Instrum. Meth. [**A462**]{} (2001) 152 see webpage `http://proj-clhep.web.cern.ch/proj-clhep/` W.T. Giele, D.A. Kosower and P.Z. Skands, arXiv:0707.3652

ArXiv

--- abstract: 'We propose a method for the approximation of solutions of PDEs with stochastic coefficients based on the direct, i.e., non-adapted, sampling of solutions. This sampling can be done by using any legacy code for the deterministic problem as a black box. The method converges in probability (with probabilistic error bounds) as a consequence of sparsity and a concentration of measure phenomenon on the empirical correlation between samples. We show that the method is well suited for truly high-dimensional problems (with slow decay in the spectrum).' address: - 'Aerospace Engineering Sciences Department, University of Colorado, Boulder, CO 80309, USA' - 'Applied & Computational Mathematics Department, California Institute of Technology, Pasadena, CA 91125, USA' author: - Alireza Doostan - Houman Owhadi bibliography: - 'AD\_bib\_v1.bib' title: 'A non-adapted sparse approximation of PDEs with stochastic inputs' --- [Polynomial chaos; Uncertainty quantification; Stochastic PDE; Compressive sampling; Sparse approximation]{} Introduction {#sec:intro} ============ Realistic analysis and design of complex engineering systems require not only a fine understanding and modeling of the underlying physics and their interactions, but also a significant recognition of intrinsic uncertainties and their influences on the quantities of interest. Uncertainty Quantification (UQ) is an emerging discipline that aims at addressing the latter issue; it aims at meaningful [characterization]{} of uncertainties in the physical models from the available measurements and efficient propagation of these uncertainties for a quantitative validation of model predictions. Despite recent growing interests in UQ of complex systems, it remains a grand challenge to efficiently propagate uncertainties through systems characterized by a large number of uncertain sources where the so-called curse-of-dimensionality is yet an unsolved problem. Additionally, development of [*non-intrusive*]{} uncertainty propagation techniques is of essence as the analysis of complex multi-disciplinary systems often requires the use of sophisticated coupled deterministic solvers in which one cannot readily intrude to set up the necessary propagation infrastructure. Sampling methods such as the Monte Carlo simulation and its several variants had been utilized for a long time as the primary scheme for uncertainty propagation. However, it is well understood that these methods are generally inefficient for large-scale systems due to their slow rate of convergence. There has been an increasing recent interest in developing alternative numerical methods that are more efficient than the Monte Carlo techniques. Most notably, the [stochastic Galerkin]{} schemes using polynomial chaos (PC) bases [@Ghanem03; @Deb01; @Xiu02; @Babuska04; @Wan05] have been successfully applied to a variety of engineering problems and are extremely useful when the number of uncertain parameters is not large. In their original form, the stochastic Galerkin schemes are [*intrusive*]{}, as one has to modify the deterministic solvers for their implementation. [*Stochastic collocation*]{} schemes [@Tatang95; @Mathelin03; @Xiu05a; @Babuska07a; @Nobile08a] belong to a different class of methods that rely upon (isotropic) sparse grid integration/interpolation in the stochastic space of the problem to reduce the curse-of-dimensionality associated with the conventional tensor-product integration/interpolation rules. As their construction is primarily based on the input parameter space, the computational cost of both stochastic Galerkin and collocation techniques increases rapidly for large number of independent input uncertainties. More recently, efforts have been made to construct [*solution-adaptive*]{} uncertainty propagation techniques that exploit any structures in the solution to decrease the computational cost. Among them are the multi-scale model reduction of [@Doostan07] and the sparse decomposition of [@Todor07a; @Bieri09a; @Bieri09b; @Bieri09c; @Blatman10] for the stochastic Galerkin technique, anisotropic and adaptive sparse grids of [@Nobile08b; @Ma09a] for the stochastic collocation scheme, and low-rank solution approximations of [@Nouy07; @Nouy08; @Doostan09]. In the present study, we are interested in cases where the quantity of interest is [*sparse*]{} at the stochastic level, i.e., it can be accurately represented with only [*few*]{} terms when linearly expanded into a stochastic, e.g., polynomial chaos, basis. Interestingly, sparsity is salient in the analysis of high-dimensional problems where the number of energetic basis functions (those with large coefficients) is small relative to the cardinality of the full basis. For instance, it has been shown in [@Todor07a; @Bieri09a] that, under some mild conditions, solutions to linear elliptic stochastic PDEs with high-dimensional random coefficients admit sparse representations with respect to the PC basis. Consequently, an approach based on a zero-dimensional algebraic stochastic problem has been proposed in [@Bieri09a] to detect the sparsity pattern, which then guides the stochastic Galerkin analysis of the original problem. Moreover, a “quasi"-best $N$-term approximation for a class of elliptic stochastic PDEs has been proposed in [@Bieri09c]. In this work, using [*concentration of measure*]{} inequalities and [*compressive sampling*]{} techniques, we derive a method for PC expansion of sparse solutions to stochastic PDEs. The proposed method is - Non-intrusive: it is based on the direct random sampling of the PDE solutions. This sampling can be done by using any legacy code for the deterministic problem as a black box. - Non-adapted: it does not tailor the sampling process to identify the important dimensions at the stochastic level - Provably convergent: we obtain probabilistic bounds on the approximation error proving the stability and convergence of the method. - [Well-suited to problems with high-dimensional random inputs.]{} [*Compressive sampling*]{} is an emerging direction in signal processing that aims at recovering [*sparse*]{} signals accurately (or even exactly) from a small number of their random projections [@Chen98; @Chen01a; @Candes06a; @Donoho06b; @Candes06b; @Candes06c; @Candes07a; @Cohen09a; @Bruckstein09]. A sparse signal is simply a signal that has only few significant coefficients when linearly expanded into a basis, e.g., $\{\psi_{\bm{\alpha}}\}$. For sufficiently sparse signals, the number of samples needed for a successful recovery is typically less than what is required by the Shannon-Nyquist sampling principle. Generally speaking, a successful signal reconstruction by compressive sampling is conditioned upon: - [*Sufficient*]{} sparsity of the signal; and - [*Incoherent*]{} random projections of the signal. A square-measurable stochastic function $u(\omega)$, defined on a suitable probability space $(\Omega,\mathcal{F},\mathcal{P})$ can be expanded into a mean-squared convergent series of the chaos polynomial bases, i.e., $u(\omega)\approx\sum_{\bm \alpha}c_{\bm \alpha}\psi_{\bm \alpha}(\omega)$, with some cardinality $P$. The stochastic function $u(\omega)$ is then sparse in PC basis $\{\psi_{\bm \alpha}\}$, if only a small fraction of coefficients $c_{\bm \alpha}$ are significant. In this case, under certain conditions, the sparse PC coefficients $\bm{c}$ may be computed accurately and robustly using only $N\ll P$ random samples of $u(\omega)$ via compressive sampling. Given $N$ random samples of $u(\omega)$, compressive sampling aims at finding the sparsest (or nearly sparsest) coefficients $\bm{c}$ from an optimization problem of the form $$\label{eqn:P1_delta_intro} (P_{s,\delta}):\qquad\min_{\bm{c}} \Vert\bm{Wc}\Vert_{s}\quad\mathrm{subject \;to}\quad \Vert\bm{\Psi}\bm{c}-\bm{u}\Vert_2\leq\delta,$$ where $\Vert\bm{Wc}\Vert_{s}$, with $s=\{0,1\}$ and some positive diagonal weight matrix $\bm W$, is a measure of the sparsity of $\bm{c}$ and $\Vert\bm{\Psi}\bm{c}-\bm{u}\Vert_2$ is a measure of the accuracy of the truncated PC expansion in estimating the $u(\omega)$ samples. The $N$-vector $\bm{u}$ contains the independent random samples of $u(\omega)$ and the rows of the $N\times P$ matrix $\bm{\Psi}$ consist of the corresponding samples of the PC basis $\{\psi_{\bm \alpha}\}$. Throughout the rest of this manuscript, we will elaborate on the formulation of the compressive sampling problem (\[eqn:P1\_delta\_intro\]) and the required conditions under which it leads to an accurate and stable approximation of an arbitrary sparse stochastic function as well as sparse solutions to linear elliptic stochastic PDEs. [Although we choose to study this particular class of stochastic PDEs, we stress that the proposed algorithms and theoretical developments are far more general and may be readily applied to recover sparse solution of other stochastic systems.]{} [ In Section \[sec:p-setup\], we describe the setup of the problem of interest. We then, in Section \[sec:stoch\_disc\], briefly overview the spectral stochastic discretization of the random functions using PC basis. The main contribution of this work on sparse approximation of stochastic PDEs using compressive sampling is then introduced in Sections \[sec:cs\], \[sec:stability\_spde\], and \[sec:sucess\_rec\]. Sections \[sec:delta\] and \[sec:solvers\] discuss some of the implementation details of the present technique. To demonstrate the accuracy and efficiency of the proposed procedures, in Section \[sec:examples\], we perform two numerical experiments on a $1$-$D$ (in space) linear elliptic stochastic PDE with high-dimensional random diffusion coefficients.]{} Problem setup {#sec:p-setup} ============= Let $(\Omega,\mathcal{F},\mathcal{P})$ be a complete probability space where $\mathcal{P}$ is a probability measure on the $\sigma-$field $\mathcal{F}$. We consider the following elliptic stochastic PDE defined on a bounded [Lipschitz continuous]{} domain $\mathcal{D}\subset \mathbb{R}^{D}$, $D=1,2,$ or $3$, with boundary $\partial \mathcal{D}$, $$\begin{aligned} \label{eqn:spde} -\nabla\cdot\left(a(\bm{x},\omega)\nabla u(\bm{x},\omega)\right)&=&f(\bm{x})\quad \bm{x}\in \mathcal{D},\\ u(\bm{x},\omega)&=&0\quad \bm{x}\in \partial\mathcal{D}, \nonumber\end{aligned}$$ $\mathcal{P}-a.s.\; \omega\in\Omega$. The diffusion coefficient $a(\bm{x},\omega)$ is a stochastic function defined on $(\Omega,\mathcal{F},\mathcal{P})$ and is the source of uncertainty in (\[eqn:spde\]). We assume that $a(\bm{x},\omega)$ is specified by a truncated Karhunen-Loève-“like" expansion $$\label{eqn:kle_a} a(\bm{x},\omega)=\bar{a}(\bm{x})+\sum_{i=1}^{d}\sqrt{\lambda_i}\phi_{i}(\bm{x})y_{i}(\omega),$$ where $(\lambda_i,\phi_i)$, $i=1,\cdots,d$, are the eigenpairs of the covariance function $C_{aa}(\bm{x}_1,\bm{x}_2)\in L_{2}(\mathcal{D}\times\mathcal{D})$ of $a(\bm{x},\omega)$ and $\bar{a}(\bm{x})$ is the mean of $a(\bm{x},\omega)$. We further assume that $a(\bm{x},\omega)$ satisfies the following conditions: [*A-I.*]{} For all $\bm{x}\in \mathcal{D}$, there exists constants $a_{\min}$ and $a_{\max}$ such that $$\label{eqn:a_positivity} 0<a_{\min}\leq a(\bm{x},\omega)\leq a_{\max}<\infty \quad \mathcal{P}-a.s.\; \omega\in\Omega$$ [*A-II.*]{} The covariance function $C_{aa}(\bm{x}_1,\bm{x}_2)$ is [*piecewise analytic*]{} on $\mathcal{D}\times\mathcal{D}$ [[@Schwab06a; @Bieri09a]]{}, implying that there exist real constants $c_1$ and $c_2$ such that for $i=1,\cdots,d$, $$\label{eqn:lambda_bound} 0\leq \lambda_i\leq c_1e^{-c_2i^{\kappa}}$$ and $$\label{eqn:phi_bound} \forall \bm{\alpha}\in\mathbb{N}^{d}:\qquad \sqrt{\lambda_{i}}\Vert\partial^{\bm{\alpha}}\phi_i\Vert_{L^{\infty}(\mathcal{D})}\leq c_1e^{-c_2i^{\kappa}},$$ where $\kappa:=1/D$ and $\bm{\alpha}\in\mathbb{N}^{d}$ is a fixed multi-index. Notice that the decay rates in Eqs. (\[eqn:lambda\_bound\]) and (\[eqn:phi\_bound\]) will be [*algebraic*]{} if $C_{aa}(\bm{x}_1,\bm{x}_2)$ has $C^{s}(\mathcal{D}\times\mathcal{D})$ regularity for some $s>0$ [[@Schwab06a]]{}.\ [*A-III.*]{} The random variables $\{y_{k}(\omega)\}_{k=1}^{d}$ are independent and uniformly distributed on $\Gamma_k:=[-1,1]$, $k=1,\cdots,d$, with probability distribution function $\rho_k(y_k)=1/2$ defined over $\Gamma_k$. The joint probability distribution function of the random vector [$\bm{y}:=(y_{1},\cdots,y_d)$]{} is then given by $\rho(\bm{y}):=\prod_{k=1}^{d}\rho_{k}(y_k)$.\ [**Remark:**]{} The algorithm proposed in the paper requires the existence of a sparse solution. The only role of assumption [*A-II*]{} is to guarantee the existence of such a sparse solution. It is not necessary for the application and the validity of the proposed algorithm. In particular, if the coefficient $a$ is only essentially bounded, the proposed algorithm will be accurate as long as a sparse approximation exists. Given the [finite-dimensional uncertainty representation in (\[eqn:kle\_a\])]{}, the solution $u(\bm{x},\omega)$ of (\[eqn:spde\]) also admits a finite-dimensional representation, i.e., $$\label{eqn:sol} u(\bm{x},\bm{y}):=u(\bm{x},y_1(\omega),\cdots,y_d(\omega)):\mathcal{D}\times\Gamma\rightarrow \mathbb{R},$$ where $\Gamma:=\prod_{k=1}^{d}\Gamma_{k}$. [In what follows, we first briefly outline the Legendre spectral stochastic discretization of $u(\bm{x},\bm{y})$ and consequently introduce our approach based on compressive sampling to obtain such a discretization.]{} Numerical approach {#sec:numerical_approach} ================== Spectral stochastic discretization {#sec:stoch_disc} ---------------------------------- [In the context of the spectral stochastic methods [[@Ghanem03; @Deb01; @Xiu02; @Babuska04]]{}, the solution $u(\bm{x},\bm{y})$ of (\[eqn:spde\]) is represented by an infinite series of the form $$\label{eqn:chaos_exact} u(\bm{x},\bm{y})=\sum_{\bm{\alpha}\in\mathbb{N}_{0}^{d}}c_{\bm{\alpha}}(\bm{x})\psi_{\bm{\alpha}}(\bm{y}),$$ where $\mathbb{N}_{0}^{d}:=\left\{(\alpha_1,\cdots,\alpha_d): \alpha_j\in \mathbb{N}\cup\{0\} \right\}$ is the set of multi-indices of size $d$ defined on non-negative integers.]{} The basis functions $\{\psi_{\bm{\alpha}}(\bm{y})\}$ are multi-dimensional Legendre polynomials, referred to as the Legendre polynomial chaos (PC), and are orthogonal with respect to the joint probability measure $\rho(\bm{y})$ of the random vector $\bm{y}$. Each basis function $\psi_{\bm{\alpha}}(\bm{y})$ is a tensor product of univariate Legendre polynomials $\psi_{\alpha_i}(y_i)$ of degree $\alpha_i\in\mathbb{N}_0^1$, i.e., $$\label{eqn:multid-univ} \psi_{\bm{\alpha}}(\bm{y})=\psi_{\alpha_1}(y_1)\psi_{\alpha_2}(y_2)\cdots\psi_{\alpha_d}(y_d)\qquad {\bm{\alpha}}\in\mathbb{N}_{0}^{d}.$$ We here assume that the univariate Legendre polynomials $\psi_{\alpha_i}(y_i)$ are also normalized [such that]{} $$\label{eqn:univ_legendre} \int_{\Gamma_i}\psi_{\alpha_i}^{2}(y_i)\rho_{i}(y_i)dy_i=1,\quad i=1,\cdots,d.$$ The [exact]{} generalized Fourier coefficients $c_{\bm{\alpha}}(\bm{x})$ in (\[eqn:chaos\_exact\]), referred to as the PC coefficients, are computed by the projection of $u(\bm{x},\bm{y})$ onto each basis function $\psi_{\bm{\alpha}}(\bm{y})$, $$\label{eqn:coefficients} c_{\bm{\alpha}}(\bm{x})=\mathbb{E}\left[u(\bm{x},\bm{y})\psi_{\bm{\alpha}}(\bm{y})\right] =\int_{\Gamma}u(\bm{x},\bm{y})\psi_{\bm{\alpha}}(\bm{y})\rho(\bm{y})d\bm{y}.$$ [Here, $\mathbb{E}$ denotes the expectation operator.]{} In practice, the expansion (\[eqn:chaos\_exact\]) is finite; that is, only a finite number of spectral modes is needed to approximate $u(\bm{x},\bm{y})$ within a given target accuracy. Traditionally, a finite-order truncation of the basis $\{\psi_{\bm{\alpha}}(\bm{y})\}$ is adopted for the approximation, i.e., $$\label{eqn:chaos_truncated} u_{p}(\bm{x},\bm{y}):=\sum_{\bm{\alpha}\in\Lambda_{p,d}}c_{\bm{\alpha}}(\bm{x})\psi_{\bm{\alpha}}(\bm{y}),$$ where the set of multi-indices $\Lambda_{p,d}$ is $$\label{eqn:multi-index} \Lambda_{p,d}:=\left\{\bm{\alpha}\in\mathbb{N}_{0}^{d}: \Vert\bm{\alpha}\Vert_{1}\leq p,\;\Vert \bm{\alpha}\Vert_{0}\leq d \right\}$$ and has the cardinality $$\label{eqn:card_basis_full} P:=\vert\Lambda_{p,d}\vert=\frac{(p+d)!}{p!d!}.$$ Here, $\Vert\bm{\alpha}\Vert_{1}=\sum_{i=1}^{d}\bm{\alpha}_i$ and $\Vert\bm{\alpha}\Vert_{0}=\#\{i:\alpha_i>0\}$ are the total order (degree) and dimensionality of the basis function $\psi_{\bm{\alpha}}(\bm{y})$, respectively. The approximation is then refined by increasing $p$ to achieve a given target accuracy. Under assumptions [*A-I*]{}, [*A-II*]{}, and [*A-III*]{} stated in Section \[sec:p-setup\], the solution $u(\bm{x},\bm{y})$ is analytic with respect to the random variables $\{y_i\}_{i=1}^{d}$ (see [@Babuska07a]), and as $p$ increases, the approximation (\[eqn:chaos\_truncated\]) converges exponentially fast in the mean-squares sense [@Babuska04; @Babuska07a; @Bieri09a]. [**Definition (Sparsity)**]{} [*The solution $u(\bm{x},\bm{y})$ is said to be (nearly) sparse if only a small fraction of coefficients $c_{\bm\alpha}(\bm{x})$ in (\[eqn:chaos\_truncated\]) are dominant and contribute to the solution statistics.*]{} As will be described in Section \[sec:cs\], a sparse solution $u(\bm{x},\bm{y})$ may be accurately recovered using $N\ll P$ random samples $\{u(\bm{x},\bm{y}_i)\}_{i=1}^{N}$ using compressive sampling techniques. This has to be compared, for instance, with the least-squares regression-type techniques, [@Hosder06], that normally require $N\gg P$ samples for an accurate recovery. Sparse recovery using compressive sampling {#sec:cs} ------------------------------------------ [*Compressive sampling*]{} is an emerging theory in the field of signal and image processing [@Chen98; @Chen01a; @Candes06a; @Donoho06b; @Candes06b; @Candes06c; @Candes07a; @Cohen09a; @Bruckstein09]. It hinges around the idea that a set of incomplete random observations of a sparse signal can be used to accurately, or even exactly, recover the signal (provided that the basis in which the signal is sparse is known). In particular, the number of such observations may be much smaller than the cardinality of the signal. In the context of problem (\[eqn:spde\]), compressive sampling may be interpreted as follows. The solution $u(\bm{x},\bm{y})$ that is sparse, in the sense of Lemma \[lem:sparsity\] defined in Section \[sec:sparse\_solution\], can be accurately recovered using $N\ll P$ random samples $\{u(\bm{x},\bm{y}_i)\}_{i=1}^{N}$, where $P$ is the cardinality of the Legendre PC basis $\{\bm{\psi}_{\bm{\alpha}}(\bm{y})\}$. We next elaborate on the above statement and address how such a sparse reconstruction is achieved and under what conditions it is successful. Let $\{u(\bm{y}_{i})\}_{i=1}^{N}$ be [*i.i.d.*]{} random samples of $u(\bm{x},\bm{y})$ for a fixed point $\bm{x}$ in $\mathcal{D}$. For the time being, let us assume that the $p$th-order PC basis $\{\bm{\psi}_{\bm{\alpha}}(\bm{y})\}$ is a complete basis to expand $u(\bm{y})$; we will relax this assumption as we proceed. Given pairs of $\{\bm{y}_{i}\}_{i=1}^{N}$ and $\{u(\bm{y}_{i})\}_{i=1}^{N}$, we write $$\label{eqn:pce_of_samples} u(\bm{y}_{i}) = \sum_{\bm{\alpha}\in\Lambda_{p,d}}c_{\bm{\alpha}}\psi_{\bm{\alpha}}(\bm{y}_i),\quad i=1,\cdots,N,$$ or equivalently, $$\label{eqn:underdetermind_system} \bm{\Psi}\bm{c}=\bm{u}.$$ Here $\bm{c}\in\mathbb{R}^{P}$ is the vector of PC coefficients $c_{\bm{\alpha}}$ to be determined, $\bm{u}\in\mathbb{R}^{N}$ is the vector of samples of $u(\bm{y})$, and each column of the [*measurement*]{} matrix $\bm{\Psi}\in \mathbb{R}^{N\times P}$ contains samples of the $j$th element of the PC basis, i.e., $$\label{eqn:Psi_matrix} \bm{\Psi}[i,j] = \psi_{\bm{\alpha}_{j}}(\bm{y}_{i}),\quad i=1,\cdots,N,\quad j=1,\cdots,P.$$ We are interested in the case that the number $N$ of solution samples is much smaller than the unknown PC coefficients $P$, i.e., $N\ll P$. Without any additional constraints on $\bm{c}$, the underdetermined linear system (\[eqn:underdetermind\_system\]) is ill-posed and, in general, has infinitely many solutions. When $\bm{c}$ is sparse; that is, only a small fraction of the coefficients $c_{\bm{\alpha}}$ are significant, the problem may be regularized to ensure a well-posed solution. Such a regularization may be achieved by seeking a solution $\bm{c}$ with the minimum number of non-zeros. This can be formulated in the optimization problem $$\label{eqn:P0} (P_0):\qquad\min_{\bm{c}} \Vert\bm{c}\Vert_{0}\quad\mathrm{subject \;to}\quad \bm{\Psi}\bm{c}=\bm{u},$$ where the semi-norm $\Vert \bm{c}\Vert_0:=\#\{\bm{\alpha}:c_{{\bm{\alpha}}}\neq 0\}$ is the number of non-zero components of $\bm{c}$. In general, the global minimum solution of $(P_0)$ is not unique and is NP-hard to compute: the cost of a global search is exponential in $P$. Further developments in compressive sampling resulted in a convex relaxation of problem $(P_0)$ by minimization of the $\ell_1$-norm of the solution $\bm{c}$ instead, i.e., $$\label{eqn:P1} (P_{1}):\qquad\min_{\bm{c}} \Vert\bm{Wc}\Vert_{1}\quad\mathrm{subject \;to}\quad \bm{\Psi}\bm{c}=\bm{u},$$ where $\bm{W}$ is a diagonal matrix whose $[j,j]$ entry is the $\ell_2$-norm of the $j$th column of $\bm{\Psi}$ and $\Vert\cdot\Vert_{1}$ denotes the $\ell_1$-norm. Notice that the $\ell_1$-norm is the closest convex function to the $\ell_0$-norm that compels the small coefficients $c_{\bm\alpha}$ to be zero, thus promoting the sparsity in the solution. The purpose of weighting the $\ell_1$ cost function with $\bm{W}$ is to prevent the optimization from biasing toward the non-zero entries in $\bm{c}$ whose corresponding columns in $\bm{\Psi}$ have large norms. The problem $(P_1)$ is entitled Basis Pursuit (BP) [@Chen98] and its solution can be obtained by linear programming. Since the $\ell_1$-norm functional $\Vert\bm{c}\Vert_{1}$ is convex, the optimization problem $(P_{1})$ admits a unique solution that coincides with the unique solution to problem $(P_0)$ for sufficiently sparse $\bm{c}$ with some constraints on the measurement matrix $\bm\Psi$; e.g., see [@Bruckstein09]. In general, the $p$th-order PC basis is not complete for the exact representation of $u(\bm{y})$; therefore, we have to account for the truncation error. This can be accommodated in $(P_0)$ and $(P_1)$ by allowing a non-zero residual in the constraint $\bm{\Psi c}=\bm{u}$. Therefore, as in Sections 3.2.1 and 3.2.3 of [@Bruckstein09], the proposed algorithms in this paper are error-tolerant versions of $(P_0)$ and $(P_1)$, with error tolerance $\delta$, i.e., $$\label{eqn:P0_delta} (P_{0,\delta}):\qquad\min_{\bm{c}} \Vert\bm{c}\Vert_{0}\quad\mathrm{subject \;to}\quad \Vert\bm{\Psi}\bm{c}-\bm{u}\Vert_2\leq\delta$$ and $$\label{eqn:P1_delta} (P_{1,\delta}):\qquad\min_{\bm{c}} \Vert\bm{Wc}\Vert_{1}\quad\mathrm{subject \;to}\quad \Vert\bm{\Psi}\bm{c}-\bm{u}\Vert_2\leq\delta,$$ respectively. The latter problem is named Basis Pursuit Denoising (BPDN) in [@Chen98] and may be solved using techniques from quadratic programming. We leave the discussion on the available algorithms for solving problems $(P_{1,\delta})$ and $(P_{0,\delta})$ to Section \[sec:solvers\]. Instead, we henceforth delineate on sufficient conditions under which the BPDN problem $(P_{1,\delta})$ leads to a successful Legendre PC expansion of a general essentially bounded sparse stochastic function $u(\bm y)$ and, subsequently, the sparse solution $u(\bm x,\bm y)$ to the problem . Our results are extensions of those in [@Donoho06a; @Bruckstein09], adapted to the case where the measurement matrix $\bm{\Psi}$ consists of random evaluations of the Legendre PC basis $\{\psi_{\bm{\alpha}}\}$. With slight differences that will be remarked accordingly, similar results hold for the case of the $(P_{0,\delta})$ problem. \[the:main0\] Let $u(\bm y)$ be an essentially bounded function of i.i.d. random variables $\bm y:=(y_1,\cdots,y_d)$ uniformly distributed on $\Gamma:=[-1,1]^d$. Define $$\label{eqn:maximum_sparsity} S_{\max}:=\frac{N}{64 P^{4c_{p,d}}(\ln P)},$$ with $$\label{eqn:cpd} c_{p,d}:=\frac{\ln 3}{2}\frac{p}{\ln\left(\frac{(p+d)!}{p!d!}\right)}.$$ Let $u_p^{1,\delta}(\bm{y}):=\sum_{\bm \alpha\in\Lambda_{p,d}}c_{\bm\alpha}^{1,\delta}\psi_{\bm\alpha}(\bm y)$ be the Legendre PC approximation of $u(\bm{y})$ with coefficients $\bm{c}^{1,\delta}$ computed from the $\ell_1$-minimization problem $(P_{1,\delta})$ with $\delta\geq 0$. If there exists a Legendre PC expansion $u_p^{0}(\bm y):=\sum_{\bm\alpha\in\Lambda_{p,d}^{\epsilon}}c_{\bm\alpha}^{0}\psi_{\bm\alpha}(\bm y)$, for some index set $\Lambda_{p,d}^{\epsilon}\subseteq\Lambda_{p,d}$ such that $\left\Vert u-u_{p}^{0}\right\Vert_{L^{\infty}(\Gamma)}\leq \epsilon$ and $$\label{eqn:sparsity_condition} S < S_{\max},$$ with $S:=|\Lambda_{p,d}^{\epsilon}|$, then with probability $$\label{eqn:prob_success_general} Prob_{1}\geq 1- 4P^{2-2S_{\max}}-P^{-8S_{\max}}-P^{-8S_{\max} P^{4c_{p,d}}},$$ (on the $N$ samples $\{u(\bm{y}_i)\}_{i=1}^{N}$) and for some constants $c_1$ and $c_2$, the solution $u_p^{1,\delta}$ must obey $$\label{eqn:stability_general} \left\Vert u-u_{p}^{1,\delta}\right\Vert_{L^{2}(\Gamma)}\leq c_1\epsilon+c_2\frac{\delta}{\sqrt{N}}.$$ In simple words, Theorem \[the:main0\] states that if an essentially bounded stochastic function $u(\bm y)$ admits a sparse Legendre PC expansion, then the $\ell_1$-minimization problem $(P_{1,\delta})$ can accurately recover it from a sufficiently large number of random solution samples. The recovery is stable under the truncation error $\Vert\bm{\Psi}\bm{c}-\bm{u}\Vert_2$ and is within a distance of the exact solution that is proportional to the error tolerance $\delta$. It is worth highlighting that no prior knowledge of the sparsity pattern of the PC coefficients $\bm c$ is needed for an accurate recovery. [**Remark:**]{} Based on the conditions and , the number $N$ of random samples has to grow like $P^{4c_{p,d}}\ln P$ and also proportional to the number of dominant coefficients $S=\vert\Lambda_{p,d}^{\epsilon}\vert$. Given any order $p$ of the PC expansion, for sufficiently high-dimensional problems, the constant $c_{p,d}<1/4$ (see Lemma \[lem:diameter\] and Fig. \[fig:c\_pd\]), thus justifying $N\ll P$. In fact, the conditions and are too pessimistic; in practice, the number of random samples required for an accurate recovery is much smaller than the theoretical value in . We will elaborate on this statement in Section \[sec:sucess\_rec\]. [**Remark:**]{} Although the BPDN reconstruction is achieved by minimizing the $\ell_1$-norm of the solution, based on , the approximation also converges to the exact solution in the mean-squares sense. [**Remark:**]{} A similar theorem holds for the case of the sparse approximation using the $\ell_0$-minimization problem $(P_{0,\delta})$ in . In this case, the condition has to be replaced with $S_{\max}:=\frac{N}{16 P^{4c_{p,d}}(\ln P)}$ which is, in theory, milder than that of the $(P_{1,\delta})$ problem. The error estimate also holds with a larger probability, but with different constants $c_1$ and $c_2$. Sparsity of the solution $u(\bm{x},\bm{y})$ {#sec:sparse_solution} ------------------------------------------- Notice that the accurate recovery of $u(\bm y)$ is conditioned upon the existence of a sparse PC expansion $u_{p}^{0}$ (see Theorem \[the:main0\]). In fact, this assumption may not hold for an arbitrary stochastic function $u(\bm y)$, as all the elements of the basis set $\{\psi_{\bm\alpha}(\bm y)\}$ may be important. In this case, our sparse approximation still converges to the actual solution but, perhaps, [not using as few as $N\ll P$]{} random solution samples. We will now summarize the results of [@Todor07a; @Bieri09a] on the sparsity of the Legendre PC expansion of the solution $u(\bm{x},\bm{y})$ to the problem (\[eqn:spde\]). Alternative to the $p$th-order truncated PC expansion of (\[eqn:chaos\_truncated\]), one may ideally seek a proper index set $\Lambda_{p,d}^{\epsilon}\subseteq \Lambda_{p,d}$, with sufficiently large $p$, such that for a given accuracy $\epsilon$ $$\label{eqn:sparse_set} \Lambda_{p,d}^{\epsilon}: = \arg\min\left\{\vert \tilde{\Lambda}_{p,d}\vert: \tilde{\Lambda}_{p,d}\subseteq\Lambda_{p,d}, \Vert u-\tilde{u}_{p}\Vert_{H_{0}^{1}\left(\mathcal{D},L^{\infty}(\Gamma)\right)}\leq \epsilon\right\},$$ in which $\tilde{u}_{p}(\bm{x},\bm{y}):=\sum_{\bm{\alpha}\in\tilde{\Lambda}_{p,d}}c_{\bm{\alpha}}(\bm{x})\psi_{\bm{\alpha}}(\bm{y})$. This will lead to the so-called [*sparse approximation*]{} of $u(\bm{x},\bm{y})$ if $$\label{eqn:def-sparsity} \vert\Lambda_{p,d}^{\epsilon}\vert\ll\vert\Lambda_{p,d}\vert=P,$$ where $\Lambda_{p,d}$ is defined in . Such a reduction in the number of basis functions in (\[eqn:def-sparsity\]) is possible as, given the accuracy $\epsilon$, the [*effective*]{} dimensionality $\nu$ of $u(\bm{x},\bm{y})$ in $\Gamma$ is potentially smaller than the apparent dimensionality $d$. More precisely, under assumptions *A-I*, *A-II*, and *A-III* stated in Section \[sec:p-setup\], the analyses of [@Todor07a; @Bieri09a] imply that the discretization of $u(\bm{x},\bm{y})$ using a sparse index set $\Lambda_{p,\nu}$, $$\label{eqn:multi-index-sparse} \Lambda_{p,\nu}:=\left\{\bm\alpha\in\mathbb{N}_{0}^{d}: \Vert\bm\alpha\Vert_{1}\leq p,\;\Vert \bm\alpha\Vert_{0}\leq \nu\leq d \right\}$$ preserves the exponential decay of the approximation error in the $H_{0}^{1}\left(\mathcal{D},L^{\infty}(\Gamma)\right)$ sense. For the sake of completeness, we cite this from [@Bieri09a] in the following lemma. \[lem:sparsity\] Given assumptions A-I, A-II, and A-III in Section \[sec:p-setup\], there exist constants $c_1,c_2,c_3,c_4 > 0$, depending only on $a(\bm{x},\omega)$ and $f(\bm{x})$ but independent of $d,p,\nu$, such that $$\label{eqn:err_estimate} \Vert u-u_{p,\nu}\Vert_{H_{0}^{1}\left(\mathcal{D},L^{\infty}(\Gamma)\right)}\leq c_1\left(e^{-c_2\nu^{1+\kappa}}+e^{c_3\nu(\ln d+\ln p)-c_4 p}\right),$$ [*for any $d,p,\nu\in\mathbb{N}$ with $\nu\leq d$ and $\kappa=1/D$.*]{} In particular, for $d\geq c_d\vert\ln\epsilon\vert^{1/\kappa}$, choosing $$\label{eqn:sparse_parameters} p_\epsilon=\lceil c_pd^{\kappa}\rceil\leq p \quad \mathrm{and}\quad \nu_\epsilon=\lceil c_\nu d^{\kappa/(\kappa+1)}\rceil\leq d,$$ leads to $$\label{eqn:err_estimate_sparse} \Vert u-u_{p_\epsilon,\nu_\epsilon}\Vert_{H_{0}^{1}\left(\mathcal{D},L^{\infty}(\Gamma)\right)}\leq\epsilon,$$ where $u_{p_\epsilon,\nu_\epsilon}$ is now defined on a sparse index set $$\label{eqn:sparse-set-epsilon} \Lambda_{p_\epsilon,\nu_{\epsilon}}:=\left\{\bm{\alpha}\in\mathbb{N}_{0}^{d}: \Vert\bm\alpha\Vert_{1}\leq p_\epsilon,\;\Vert \bm\alpha\Vert_{0}\leq \nu_\epsilon\right\}$$ with cardinality $$\label{eqn:sparse-set-reduced-dim} \vert \Lambda_{p_\epsilon,\nu_\epsilon}\vert\lesssim\epsilon^{-1/\rho},$$ for some arbitrary large $\rho>0$ and constants $c_d,c_p$, and $c_\nu$ independent of $d,p_\epsilon$, and $\nu_\epsilon$ [@Bieri09a]. In practice, the sparse set $\Lambda_{p,d}^{\epsilon}$ in (\[eqn:sparse\_set\]) (or equivalently $\Lambda_{p_\epsilon,\nu_\epsilon}$ in (\[eqn:sparse-set-epsilon\])) is not known [*a priori*]{}. In [@Bieri09a], an approach based on an algebraic purely-stochastic problem is proposed to adaptively identify $\Lambda_{p,d}^{\epsilon}$. Having done this, the coefficients of the the spectral modes are computed via the (intrusive) stochastic Galerkin scheme [@Ghanem03; @Xiu02]. Alternatively, in this work, we apply our sparse approximation using $(P_{1,\delta})$ and $(P_{0,\delta})$ to compute $u(\bm x,\bm y)$. The implementation of $(P_{1,\delta})$ and $(P_{0,\delta})$ is non-intrusive; only random samples of the solution are needed. Moreover, we do not adapt the sampling process to identify the important dimensions at the stochastic level; therefore, our constructions are non-adapted. Throughout the rest of the present paper, we focus our attention on the case of the stochastic PDE whose solution is provably sparse. The statement of Theorem \[the:main0\] can be specialized to the approximation of the sparse solution to the stochastic PDE using $(P_{1,\delta})$ (or $(P_{0,\delta})$) as follows. Stability of $(P_{1,\delta})$ for stochastic PDE {#sec:stability_spde} ------------------------------------------------- Combining Lemma \[lem:sparsity\] with Theorem \[the:main0\] leads to the following theorem. \[the:main\] [*Let $u_{p}^{1,\delta}(\bm{x},\bm y):=\sum_{\bm\alpha\in\Lambda_{p,d}}c_{\bm\alpha}^{1,\delta}(\bm x)\psi_{\bm\alpha}(\bm y)$ be the $p$th-order Legendre PC approximation of $u(\bm{x},\bm{y})$ in where the coefficients $c_{\bm\alpha}^{1,\delta}(\bm x)$ are obtained from the $\ell_1$-minimization problem $(P_{1,\delta})$ with $N$ independent samples of $u(\bm{x},\bm{y})$ and arbitrary $\delta\geq 0$. Write $\kappa:=1/D$. Let $c_{p,d}$ be defined by and let $S_{\max}$ be defined by . Let $\rho>0$ be arbitrary.*]{} [*There exists constants $c_1,c_2,c_3,c_4,c_5$ independent from $p,d,\kappa,N$ such that if $ \lceil c_2 d^{\kappa}\rceil\leq p$ and $\lceil c_3 d^{\kappa/(\kappa+1)}\rceil\leq d$,*]{} [*then with probability at least*]{} $$\label{eqn:coherence_boundse2} Prob_{1}\geq 1- 4P^{2-2S_{\max}}-P^{-8S_{\max}}-P^{-8P^{4c_{p,d}}S_{\max}},$$ [*the solution $u_{p}^{1,\delta}$ must obey*]{} $$\label{eqn:stability_convergence} \left\Vert u-u_{p}^{1,\delta}\right\Vert_{L^2(\mathcal{D}, L^2(\Gamma))} \leq c_{4} \epsilon+c_5 \frac{\delta}{\sqrt{N}}.$$ with $$\epsilon:=\max\left(\frac{1}{S_{\max}^\rho},\exp\left(-\left(\frac{d}{c_1}\right)^\kappa\right)\right)$$ Proofs and further ingredients of successful sparse approximations via $(P_{1,\delta})$ and $(P_{0,\delta})$ {#sec:sucess_rec} ------------------------------------------------------------------------------------------------------------ The ability of problems $(P_{1,\delta})$ and $(P_{0,\delta})$ in accurately approximating the sparse PC coefficients $\bm{c}$ in , hence the solution $u(\bm x,\bm{y})$, depends on two main factors: $i)$ the sparsity of the PC coefficients $\bm{c}$ and $ii)$ the [*mutual coherence*]{} of the measurement matrix $\bm{\Psi}$. In fact, the number $N$ of random solution samples required for a successful sparse approximation is dictated by these two factors. While sparsity is a characteristic of the solution of interest $u(\bm x,\bm y)$, the mutual coherence of the measurement matrix $\bm \Psi$ is universal as it only depends on the choice of PC basis $\{\psi_{\bm\alpha}(\bm y)\}$ and the sampling process from which $\bm\Psi$ is assembled. In Section \[sec:sparse\_solution\], based on the analysis of [@Bieri09a], we rationalized the sparsity of $u(\bm x,\bm y)$ with respect to the Legendre PC basis. We now give the definition of the mutual coherence of $\bm{\Psi}$ and discuss its role in our sparse approximation using $(P_{1,\delta})$ and $(P_{0,\delta})$. ### Mutual coherence of $\bm \Psi$ {#sec:def_mu} **Definition (Mutual Coherence [@Donoho06a])** [*The mutual coherence $\mu(\bm{\Psi})$ of a matrix $\bm{\Psi}\in\mathbb{R}^{N\times P}$ is the maximum of absolute normalized inner-products of its columns. Let $\bm{\psi}_{j}$ and $\bm{\psi}_{k}$ be two distinct columns of $\bm{\Psi}$. Then,*]{} $$\label{eqn:muA} \mu(\bm{\Psi}):=\max_{1\leq j,k\leq P,\;j\neq k}\frac{\vert \bm{\psi}_j^{T}\bm{\psi}_k\vert}{\Vert \bm{\psi}_j\Vert_2\Vert \bm{\psi}_k\Vert_2}.$$ In plain words, the mutual coherence is a measure of how close to orthogonal a matrix is. Clearly, for any general matrix $\bm\Psi$, $$\label{eqn:coherence_bound2} 0\leq\mu(\bm{\Psi})\leq 1,$$ where the lower bound is achieved, for instance, by unitary matrices. However, for the case of $N<P$, the mutual coherence $\mu(\bm{\Psi})$ is strictly positive. It is well understood that measurement matrices with smaller mutual coherence have a better ability to recover a sparse solution using compressive sampling techniques, e.g., see Lemma \[lem:stability\_discrete\]. Therefore, we shall proceed to examine the mutual coherence of the random measurement matrix $\bm{\Psi}$ in (\[eqn:underdetermind\_system\]). We first observe that, by the orthogonality of the Legendre PC basis, the mutual coherence $\mu(\bm{\Psi})$ converges to zero almost surely for asymptotically large random sample sizes $N$. However, it is essential for our purpose to $i)$ investigate if a desirably small $\mu(\bm{\Psi})$ can be achieved by a sample size $N\ll P$ and $ii)$ quantify how large $\mu(\bm{\Psi})$ can get for a [*finite*]{} $N$. These are addressed in the following theorem. \[the:mu\] [*Let $\bm{\Psi}\in\mathbb{R}^{N\times P}$, [as defined in (\[eqn:underdetermind\_system\])]{}, be the measurement matrix corresponding to $N$ independent random samples of the Legendre polynomial chaos basis of order $p$ in $d$ [*i.i.d.*]{} uniform random variables $\bm{y}$. There exists a positive constant $c_{p,d}:=\frac{\ln 3}{2}\frac{p}{\ln\left(\frac{(p+d)!}{p!d!}\right)}$ depending on $p$ and $d$, such that if*]{} $$\label{eqn:define_r} 0\leq r=2\sqrt{\zeta P^{4c_{p,d}}(\ln P)/N}\leq 1/2,$$ [*for some $\zeta>1$, then*]{} $$\label{eqn:coherence_bound} Prob\left[ \mu(\bm{\Psi})\geq \frac{r}{1-r}\right]\leq4P^{2-2\zeta}.$$ Figure \[fig:c\_pd\] illustrates the decay of $c_{p,d}$, for several values of $p$, as a function of $d$. Based on Theorem \[the:mu\], for cases where the number $d$ of random variables $\bm{y}$ is large enough such that $c_{p,d}<1/4$, it is sufficient to have $N\sim \mathcal{O}(16P^{4c_{p,d}}\ln P)\ll P$ to keep $\mu(\bm{\Psi})$ bounded from above with a large probability. Notice that such a requirement on $c_{p,d}$ is particularly suited to high-dimensional problems. ![Decay of $c_{p,d}$ as a function of $d$. $p=1$ ([$\square$]{}); $p=2$ ([$\circ$]{}); $p=3$ ($\triangledown$); $p=4$($\diamond$).[]{data-label="fig:c_pd"}](c_d_p.eps "fig:"){width="5.0in"} (-177,-2)[$d$]{} (-345,85) $c_{p,d}$ [**Remark:**]{} We observe that, given the choice of $r$ in (\[eqn:define\_r\]), the upper bound on $\mu(\bm{\Psi})$ in (\[eqn:coherence\_bound\]) decays like $1/\sqrt{N}$ for asymptotically large $N$, which is consistent with the Central Limit Theorem. In order to prove Theorem \[the:mu\], we first need to compute the maximum of the Legendre PC basis functions $\psi_{\bm{\alpha}}(\bm{y})$. This is given in the following lemma. \[lem:diameter\] Let $\{\psi_{\bm{\alpha}}(\bm{y})\}$ be the Legendre polynomial chaos basis of total order $p$ in $d$ i.i.d uniform random variables $\bm{y}$ (as defined in (\[eqn:multi-index\])) and with cardinality $P$. Then, $$\label{eqn:sup_norm_psi} \Vert\psi_{\bm{\alpha}}\Vert_{L^{\infty}(\Gamma)}\leq P^{c_{p,d}},$$ with a constant $$\label{eqn:c_pd} c_{p,d}:=\frac{\ln 3}{2}\frac{p}{\ln\left(\frac{(p+d)!}{p!d!}\right)}.$$ [*Proof.*]{} Given the equality $$\label{eqn:inf_norm_legendre} \Vert \psi_{\alpha_i}\Vert_{L^{\infty}(\Gamma_i)}=\sqrt{2\alpha_i+1},\nonumber$$ we have $$\label{eqn:proof_inf_norm_1-gh} \Vert \psi_{\bm{\alpha}}\Vert_{L^{\infty}(\Gamma)}=\prod_{i=1}^{d}\sqrt{2\alpha_i+1}.\nonumber$$ Using the constraint $\alpha_i\in \mathbb{N}_{0}^{1}$ and $\sum_{i=1}^d \alpha_i\leq p$, the right-hand-side is maximized when $p$ of $\alpha_i$s are equal to one and $d-p$ equal to zero (we assume $d\geq p$). We deduce that $$\label{eqn:proof_inf_norm_1} \Vert \psi_{\bm{\alpha}}\Vert_{L^{\infty}(\Gamma)}\leq 3^{\frac{p}{2}}=P^{\frac{\ln 3}{2} \frac{p}{\ln P}}=P^{c_{p,d}}.\nonumber$$ We are now ready to prove Theorem \[the:mu\]. [*Proof of Theorem \[the:mu\].*]{} The mutual coherence $\mu(\bm{\Psi})$ is $$\label{eqn:mutual_coherence_legendre} \mu(\bm{\Psi})=\max_{1\leq j,k\leq P,\;j\neq k}\frac{\left\vert \frac{1}{N}\sum_{i=1}^{N}\psi_{\bm{\alpha}_j}(\bm{y}_i)\psi_{\bm{\alpha}_k}(\bm{y}_i)\right\vert}{\left(\frac{1}{N}\sum_{i=1}^{N}\psi_{\bm{\alpha}_j}^{2}(\bm{y}_i)\right)^{1/2}\left(\frac{1}{N}\sum_{i=1}^{N}\psi_{\bm{\alpha}_k}^{2}(\bm{y}_i)\right)^{1/2}}.$$ Given the independence of samples $\{\bm{y}_i\}_{i=1}^{N}$ and using the McDiarmid’s inequality, we obtain $$\label{eqn:coherence_1a-1} Prob\left[ \left\vert \frac{1}{N}\sum_{i=1}^{N}\psi_{\bm{\alpha}_j}(\bm{y}_i)\psi_{\bm{\alpha}_k}(\bm{y}_i)\right\vert \ge r \right]\leq 2\exp\left({\frac{-2Nr^2}{(2\|\psi_{\bm{\alpha}_j}(\bm{y}_i)\psi_{\bm{\alpha}_k}(\bm{y}_i)\|_{L^\infty(\Gamma_i)})^2}}\right).$$ Using Lemma \[lem:diameter\], we may probabilistically bound the numerator in (\[eqn:mutual\_coherence\_legendre\]) as $$\label{eqn:coherence_1} Prob\left[ \left\vert \frac{1}{N}\sum_{i=1}^{N}\psi_{\bm{\alpha}_j}(\bm{y}_i)\psi_{\bm{\alpha}_k}(\bm{y}_i)\right\vert \ge r \right]\leq 2\exp\left({\frac{-Nr^2}{2P^{4c_{p,d}}}}\right),\nonumber$$ in which we exploit the orthonormality of $\psi_{\bm{\alpha}_j}(\bm{y})$ and $\psi_{\bm{\alpha}_k}(\bm{y})$, i.e., $\mathbb{E}[\psi_{\bm{\alpha}_j}\psi_{\bm{\alpha}_k}]=\delta_{jk}$. Similarly, for $j=1,\cdots,P$, we have $$\label{eqn:coherence_2} Prob\left[ \frac{1}{N}\sum_{i=1}^{N}\psi_{\bm{\alpha}_j}^{2}(\bm{y}_i) \leq 1-r \right]\leq \exp\left({\frac{-Nr^2}{P^{4c_{p,d}}}}\right)\leq \exp\left({\frac{-Nr^2}{2P^{4c_{p,d}}}}\right).\nonumber$$ Therefore, $$\label{eqn:coherence_3} Prob\left[ \frac{\left\vert \frac{1}{N}\sum_{i=1}^{N}\psi_{\bm{\alpha}_j}(\bm{y}_i)\psi_{\bm{\alpha}_k}(\bm{y}_i)\right\vert}{\left(\frac{1}{N}\sum_{i=1}^{N}\psi_{\bm{\alpha}_j}^{2}(\bm{y}_i)\right)^{1/2}\left(\frac{1}{N}\sum_{i=1}^{N}\psi_{\bm{\alpha}_k}^{2}(\bm{y}_i)\right)^{1/2}}\geq \frac{r}{1-r}\right]\leq 4\exp\left({\frac{-Nr^2}{2P^{4c_{p,d}}}}\right)$$ and $$\label{eqn:coherence_4} Prob\left[ \mu(\bm{\Psi})\geq \frac{r}{1-r}\right]\leq 4P^2\exp\left({\frac{-Nr^2}{2P^{4c_{p,d}}}}\right).$$ Taking $$\label{eqn:select_r} r=2\sqrt{\zeta P^{4c_{p,d}}(\ln P)/N},$$ for some $\zeta>1$, we arrive at the statement of the Theorem \[the:mu\]. $\square$ To summarize, we observe that with large probability, the mutual coherence $\mu(\bm \Psi)$ of the measurement matrix $\bm{\Psi}$ in (\[eqn:underdetermind\_system\]) can be arbitrarily bounded from above by increasing the number $N$ of independent random solution samples. Moreover, given the discussions of Section \[sec:sparse\_solution\], we know that the solution to problem (\[eqn:spde\]) is sparse in the Legendre PC basis. These are the two key factors affecting the stability and accuracy of our sparse approximation. Following [@Donoho06a; @Bruckstein09], we next state a [*sufficient*]{} condition on the sparsity of $u(\bm x,\bm y)$ (or, equivalently, the mutual coherence of $\bm\Psi$) such that the problem $(P_{1,\delta})$ leads to a stable and accurate sparse solution. By stability, we simply mean that the PC coefficients $\bm{c}^{1,\delta}$ recovered from the $\ell_1$-minimization problem $(P_{1,\delta})$ do not blow up in the presence of the truncation error $\delta$. We first assume that $\bm x$ is a fixed point in space and subsequently extend the analysis over the entire spatial domain $\mathcal{D}$. \[lem:stability\_discrete\] Let $u_p^{0}(\bm y):=\sum_{\bm\alpha\in\Lambda_{p,d}}c_{\bm\alpha}^{0}\psi_{\bm\alpha}(\bm y)$ with $c_{\bm\alpha}^{0}=0$ for $\bm\alpha\notin\Lambda_{p_\epsilon,\nu_\epsilon}$ be the sparse Legendre PC approximation of $u(\bm x,\bm y)$ at a spatial point $\bm x$ where the sparse index set $\Lambda_{p_\epsilon,\nu_\epsilon}$ is defined in . Assume that the vector of PC coefficients $\bm{c}^{0}$ satisfies the sparsity condition $$\label{eqn:sparsity-condition} \Vert\bm{c}^{0}\Vert_{0}=\vert\Lambda_{p_\epsilon,\nu_\epsilon}\vert<(1+1/\mu(\bm{\Psi}))/4.$$ Let $u_p^{1,\delta}(\bm{y}):=\sum_{\bm \alpha\in\Lambda_{p,d}}c_{\bm\alpha}^{1,\delta}\psi_{\bm\alpha}(\bm y)$ be the approximation of $u(\bm{y})$ with coefficients $\bm{c}^{1,\delta}$ computed from the $\ell_1$-minimization problem $(P_{1,\delta})$.Then, with probability at least $1-\exp(-\frac{N}{8 P^{4 c_{p,d}}})$ (on the $N$ samples $\{u(\bm{y}_i)\}_{i=1}^{N}$) and for all $\delta\geq 0$, the solution $u_p^{1,\delta}$ must obey $$\label{eqn:stability0} \mathbb{E}\left[\left(u_p^{0}-u_p^{1,\delta}\right)^2\right] \leq \frac{4}{3N} \frac{\left(\delta+\|\bm{\Psi}\bm{c}^0-\bm{u}\|_2 \right)^2}{1-\mu(\bm{\Psi})\left(4\Vert\bm{c}^0\Vert_{0}-1\right)}.$$ [*Proof.*]{} Using Theorem 3.1 of [@Donoho06a], we obtain that if $\bm{c}^0$ satisfies the sparsity condition $\Vert\bm{c}^0\Vert_{0}<(1+1/\mu(\bm{\Psi}))/4$, then $$\sum_{\bm\alpha} \left(c_{\bm\alpha}^{1,\delta}-c_{\bm\alpha}^{0}\right)^2 \left \Vert\bm{\psi}_{\bm\alpha}\right\Vert_{2}^2 \leq \frac{\left(\delta+\|\bm{\Psi}\bm{c}^0-\bm{u}\|_2 \right)^2}{1-\mu(\bm{\Psi})\left(4\Vert\bm{c}^0\Vert_{0}-1\right)}\nonumber$$ where $\left \Vert\bm{\psi}_{\bm\alpha}\right\Vert_{2}$ is the $\ell_2$-norm of the column of $\bm{\Psi}$ corresponding to index $\bm{\alpha}$. The presence of $\left \Vert\bm{\psi}_{\bm\alpha}\right\Vert_{2}$ is due to the fact that the columns of $\bm{\Psi}$ are not normalized. Next, using McDiarmid’s inequality and the independence of the entries $\bm{\Psi}[i,j]$ for distinct $i$s, we obtain that $$\label{eqn:stability1} Prob\left[\sum_{\bm\alpha} \left(c_{\bm\alpha}^{1,\delta}-c_{\bm\alpha}^{0}\right)^2 \left \Vert\bm{\psi}_{\bm\alpha}\right\Vert_{2}^2\leq \frac{3N}{4} \sum_{\bm\alpha} \left(c_{\bm\alpha}^{1,\delta}-c_{\bm\alpha}^{0}\right)^2 \left \Vert\bm{\psi}_{\bm\alpha}\right\Vert_{2}^2\right] \leq \exp\left(-\frac{N}{8 \|\psi_{\bm{\alpha}}\|_{L^\infty(\Gamma)}^4}\right)$$ We conclude using Lemma \[lem:diameter\] and the fact that due to the orthonormality of $\{\psi_{\bm{\alpha}}(\bm{y})\}$ we have $\mathbb{E}\left[\left(u_p^{0}-u_p^{1,\delta}\right)^2\right] =\left\Vert \bm{c}^0-\bm{c}^{1,\delta} \right\Vert_{2}^{2}$. $\square$ [**Remark:**]{} The error bound in is not tight; in fact, the actual error is significantly smaller than the upper bound given in . More importantly, according to [@Donoho06a], the sparsity condition is unnecessarily too restrictive. In practice, both far milder sparsity conditions are needed and much better actual errors are achieved. [**Remark:**]{} We will later use the sparsity condition to derive the sufficient condition (together with ) on the number $N$ of random samples needed for a successful recovery. As the condition is too restrictive, the theoretical lower bound on $N$ given in and is too pessimistic. [**Remark:**]{} According to Lemma \[lem:stability\_discrete\], we do not need to know [*a priori*]{} the sparse index set $\Lambda_{p_\epsilon,\nu_\epsilon}$; only the sparsity condition is required. [**Remark:**]{} We stated Lemma \[lem:stability\_discrete\] for the case where the sparsity of the PC expansion is due to the fact that the effective dimensionality $\nu_\epsilon$ is potentially smaller than $d$. However, as far as the stability condition is satisfied, similar stability results are valid for situations where dominant basis are defined over all the dimensions. [**Remark:**]{} With slight modifications, a similar argument as in Lemma \[lem:stability\_discrete\] may be asserted for the solution of $\ell_0$-minimization problem $(P_{0,\delta})$. Specifically, in that case, we only require a sparsity limit $\Vert\bm{c}^{0}\Vert_{0}=\vert\Lambda_{p_\epsilon,\nu_\epsilon}\vert<(1+1/\mu(\bm{\Psi}))/2$ to achieve the error estimate $\mathbb{E}\left[\left(u_p^{0}-u_p^{0,\delta}\right)^2\right] \leq \frac{4}{3N} \frac{\left(\delta+\|\bm{\Psi}\bm{c}^0-\bm{u}\|_2 \right)^2}{1-\mu(\bm{\Psi})\left(2\Vert\bm{c}^0\Vert_{0}-1\right)}$. Notice that the normalized truncation error $$\label{eqn:truncation_error} \epsilon_{N}^{2}:=\frac{\|\bm{\Psi}\bm{c}^0-\bm{u}\|_2^2}{N}$$ is the sample average estimate of the mean-squares sparse approximation error $\mathbb{E}\left[(u-u_{p}^{0})^{2}\right]$ at the point $\bm x$ and is a function of samples $\{\bm y\}_{i=1}^{N}$ in addition to the order $p_\epsilon$ and the dimensionality $\nu_\epsilon$ of the sparse PC expansion. For a given set of $N$ random independent samples $\{\bm y\}_{i=1}^{N}$ and $\{u(\bm y)\}_{i=1}^{N}$, we may bound $\epsilon_N^{2}$ in probability using McDiarmid’s inequality, i.e., $$\label{eqn:bound_truncation_error} Prob\left[\epsilon_N^2 \geq \mathbb{E}\left[(u-u_{p}^{0})^{2}\right] + r\right]\leq \exp\left(-2N \frac{r^2}{\|u - u_{p}^0\|_{L^\infty(\Gamma)}^4} \right).$$ Although our sparse approximations are point-wise in space, we are ultimately interested in deriving suitable global stability and error estimates for our sparse reconstructions. Such extensions are readily available from Lemma \[lem:stability\_discrete\] and are stated in the following corollary. \[cor:stability\_continuous\] Let $u_p^{0}(\bm x, \bm y):=\sum_{\bm\alpha\in\Lambda_{p,d}}c_{\bm\alpha}^{0}(\bm x)\psi_{\bm\alpha}(\bm y)$ with $c_{\bm\alpha}^{0}(\bm x)=0$ for $\bm\alpha\notin\Lambda_{p_\epsilon,\nu_\epsilon}$ be the sparse Legendre PC approximation of $u(\bm x,\bm y)$ where the sparse index set $\Lambda_{p_\epsilon,\nu_\epsilon}$ is defined in . Assume that the vector of PC coefficients $\bm{c}^{0}(\bm x)$ satisfies the sparsity condition $\Vert\bm{c}^{0}\Vert_{0}=\vert\Lambda_{p_\epsilon,\nu_\epsilon}\vert<(1+1/\mu(\bm{\Psi}))/4$. Let $u_p^{1,\delta}(\bm x,\bm{y}):=\sum_{\bm \alpha\in\Lambda_{p,d}}c_{\bm\alpha}^{1,\delta}(\bm x)\psi_{\bm\alpha}(\bm y)$ be the approximation of $u(\bm x,\bm{y})$ with coefficients $\bm{c}^{1,\delta}(\bm x)$ computed from the $\ell_1$-minimization problem $(P_{1,\delta})$ at any point $\bm x\in \mathcal{D}$.Then, with probability at least $1-\exp(-\frac{N}{8 P^{4 c_{p,d}}})$ (on the $N$ samples $\{u(\bm{y}_i)\}_{i=1}^{N}$) and for all $\delta\geq 0$, the solution $u_p^{1,\delta}$ must obey $$\label{eqn:stability} \left\Vert u_p^{0}-u_p^{1,\delta}\right\Vert_{L^2\left(\mathcal{D},L^2(\Gamma)\right)}^2 \leq \frac{4}{3N} \frac{\left(\delta+\left\Vert\bm{\Psi}\bm{c}^0-\bm{u}\right\Vert_{L^{2}(\mathcal{D},\ell_2(\mathbb{R}^{N}))} \right)^2}{1-\mu(\bm{\Psi})\left(4\Vert\bm{c}^0\Vert_{0}-1\right)}.$$ Furthermore, the normalized error $\epsilon_{N}^2=:\frac{\left\Vert\bm{\Psi}\bm{c}^0-\bm{u}\right\Vert_{L^{2}\left(\mathcal{D},\ell_2(\mathbb{R}^{N})\right)}^2}{N}$ is bounded from above in probability through $$\label{eqn:bound_truncation_error_global} Prob\left[\epsilon_N^2 \geq \left\Vert u- u_p^{0}\right\Vert_{L^2\left(\mathcal{D},L^2(\Gamma)\right)}^2 + r\right]\leq\exp\left(-2N \frac{r^2}{\|u - u_{p}^0\|_{L^2\left(\mathcal{D},L^\infty(\Gamma)\right)}^4} \right).$$ We have now all the necessary tools to proceed with the proof of our main result stated in Theorem \[the:main\] which is primarily a direct consequence of Lemma \[lem:sparsity\], Theorem \[the:mu\], and Corollary \[cor:stability\_continuous\]. *Proof of Theorem \[the:main\].* We first note that, given the conditions of Lemma \[lem:sparsity\], the solution to problem admits a sparse Legendre PC expansion $u_{p}$ with sparsity $S=\vert \Lambda_{p_\epsilon,\nu_\epsilon}\vert\lesssim\epsilon^{-1/\rho}$ when $\epsilon\geq \exp\left(-\left(\frac{d}{c_1}\right)^\kappa\right)$ for some constants $c_1$ and (arbitrary) $\rho>0$. Notice that the sparse approximation $u_{p}$ has an accuracy better than $\epsilon$ in the $H_{0}^{1}(\mathcal{D},L^{\infty}(\Gamma))$ sense. Based on Theorem \[the:mu\], it is sufficient to have random solution samples of size $N\geq64 P^{4c_{p,d}}(\ln P)S$, to meet the sparsity requirement $S=\vert\Lambda_{p_\epsilon,\nu_\epsilon}\vert<(1+1/\mu(\bm{\Psi}))/4$, in Corollary \[cor:stability\_continuous\], with probability at least $1-4P^{2-2S_{\max}}$ where $S_{\max}:=\frac{N}{64 P^{4c_{p,d}}(\ln P)}$. On the other hand, given $N$ random samples of solution, we require $\epsilon\geq \frac{1}{S_{\max}^\rho}$ to satisfy the sparsity condition. Given Corollary \[cor:stability\_continuous\] and using the triangular and Poincaré inequalities, with probability at least $1-P^{-8S_{\max}}$, we have $$\begin{aligned} \label{eqn:tri_poin} \left\Vert u-u_{p}^{1,\delta}\right\Vert_{L^2(\mathcal{D}, L^2(\Gamma))}&&\leq c_{\mathcal{D}}\left\Vert u-u_{p}^{0}\right\Vert_{H_0^{1}(\mathcal{D}, L^2(\Gamma))}+\left\Vert u_{p}^{0}-u_{p}^{1,\delta}\right\Vert_{L^2(\mathcal{D}, L^2(\Gamma))}\nonumber \\ &&\leq c_{\mathcal{D}}\epsilon+\frac{2}{\sqrt{3N}} \frac{\delta+\left\Vert\bm{\Psi}\bm{c}^0-\bm{u}\right\Vert_{L^{2}(\mathcal{D},\ell_2(\mathbb{R}^{N}))}}{\sqrt{1-\mu(\bm{\Psi})\left(4S-1\right)}},\nonumber\end{aligned}$$ for all $\delta>0$. Moreover, by choosing $r= \frac{1}{4}\left\Vert u- u_p^{0}\right\Vert_{L^2\left(\mathcal{D},L^\infty(\Gamma)\right)}^2$ in , $$\label{eqn:epsilon_n_bound} \epsilon_{N}^2=:\frac{\left\Vert\bm{\Psi}\bm{c}^0-\bm{u}\right\Vert_{L^{2}\left(\mathcal{D},\ell_2(\mathbb{R}^{N})\right)}^2}{N}\leq \left\Vert u- u_p^{0}\right\Vert_{L^2\left(\mathcal{D},L^2(\Gamma)\right)}^2+ \frac{1}{4}\left\Vert u- u_p^{0}\right\Vert_{L^2\left(\mathcal{D},L^\infty(\Gamma)\right)}^2\leq \frac{5}{4}c_{\mathcal{D}}^2\epsilon^2\nonumber$$ with probability at least $1-P^{-8S_{\max}P^{4c_{p,d}}}$. Finally, by taking $$\label{eqn:c1_c2} c_4:=c_{\mathcal{D}}\left(1+\frac{\sqrt{5}}{\sqrt{3}\sqrt{1-\mu(\bm{\Psi})\left(4S-1\right)}}\right)\quad \mathrm{and}\quad c_5:=\frac{2}{\sqrt{3}\sqrt{1-\mu(\bm{\Psi})\left(4S-1\right)}},\nonumber$$ we arrive at the statement of the Theorem \[the:main\]. $\square$ Choosing the truncation error tolerance $\delta$ {#sec:delta} ------------------------------------------------ An important component of the sparse approximation using $(P_{1,\delta})$ and $(P_{0,\delta})$ is the selection of the truncation error tolerance $\delta$. Although the stability bounds given in Lemma \[lem:stability\_discrete\] and Corollary \[cor:stability\_continuous\] are valid for any $\delta\geq0$, the actual error and the sparsity level of the solution to $(P_{1,\delta})$ and $(P_{0,\delta})$ depend on the choice of $\delta$. Ideally, we desire to choose $\delta\approx\Vert\bm\Psi\bm{c}^{0}-\bm u\Vert_2$; while larger values of $\delta$ deteriorate the accuracy of the approximation, as in Lemma \[lem:stability\_discrete\], smaller choices of $\delta$ may result in over-fitting the solution samples and, thus, less sparse solutions. In practice, as the exact values of the PC coefficients $\bm{c}^{0}$ are not known, the exact values of the truncation error $\Vert\bm\Psi\bm{c}^{0}-\bm u\Vert$ and, consequently, $\delta$ are not known [*a priori*]{}. Therefore, $\delta$ has to be estimated, for instance, using statistical techniques such as the cross-validation [@Boufounos07; @Ward09]. In this work, we propose a heuristic cross-validation algorithm to estimate $\delta$. We first divide the $N$ available solution samples to $N_r$ reconstruction and $N_v$ validation samples such that $N=N_r+N_v$. The idea is to repeat the solution of $(P_{1,\delta})$ (or $(P_{0,\delta})$) on the reconstruction samples and with multiple values of truncation error tolerance $\delta_r$. We then set $\delta=\sqrt{\frac{N}{N_r}}\hat{\delta}_r$ in which $\hat{\delta}_r$ is such that the corresponding truncation error on the $N_v$ validation samples is minimum. This is simply motivated by the fact that the truncation error on the validation samples is large for values of $\delta_r$ considerably larger and smaller than $\Vert\bm\Psi\bm{c}^{0}-\bm u\Vert_2$ evaluated using the reconstruction samples. While the former is expected from the upper bound on the approximation error in Lemma \[lem:stability\_discrete\], the latter is due to the over-fitting the reconstruction samples. The following exhibit outlines the estimation of $\delta$ using the above cross-validation approach:\ In the numerical experiments of Section \[sec:examples\], we repeat the above cross-validation algorithm for multiple replications of the reconstruction and validation samples. The estimate of $\delta=\sqrt{\frac{N}{N_r}}\hat{\delta}_r$ is then based on the value of $\hat{\delta}_r$ for which the average of the corresponding truncation errors $\delta_v$, over all replications of the validation samples, is minimum. This resulted in more accurate solutions in our numerical experiments. Algorithms {#sec:solvers} ---------- There are several numerical algorithms for solving problems $(P_{0,\delta})$ and $(P_{1,\delta})$ each with different optimization kernel, computational complexity, and degree of accuracy. An in-depth discussion on the performance of these algorithms is outside the scope of the present work; however, below we name some of the available options for each problem and briefly describe the algorithms that have been utilized in our numerical experiments. For comprehensive discussions on this subject, the interested reader is referred to [@Bruckstein09; @Berg08; @Figueiredo07; @Becker09; @Yang09; @Tropp10a]. [*Problem $(P_{0,\delta})$:*]{} A brute force search through all possible support sets in order to identify the correct sparsity for the solution $\bm{c}^{0,\delta}$ of $(P_{0,\delta})$ is NP-hard and not practical. Greedy pursuit algorithms form a major class of schemes to tackle the solution of $(P_{0,\delta})$ with a tractable computational cost. Instead of performing an exhaustive search for the support of the sparse solution, these solvers successively find one or more components of the solution that result in the largest improvement in the approximation. Some of the standard greedy pursuit algorithms are Orthogonal Marching Pursuit (OMP) [@Pati93; @Davis97], Regularized OMP (ROMP)[@Needell07], Stagewise OMP (StOMP) [@Donoho06c], Compressive Sampling MP (CoSaMP) [@Needell08a], Subspace Pursuit [@Dai09], and Iterative Hard Thresholding (IHT) [@Blumensath09]. Under well-defined conditions, all of the above schemes provide stable and accurate solutions to $(P_{0,\delta})$ in a reasonable time. In the present study, we employ the OMP algorithm to approximate the solution of $(P_{0,\delta})$. Starting from $\bm{c}^{0,\delta,(0)}=\bm{0}$ and an empty active column set of $\bm\Psi$, at any iteration $k$, OMP identifies only one column to be added to the active column set. The column is chosen such that the $\ell_2$-norm of the residual, $\Vert\bm\Psi\bm{c}^{0,\delta,(k)}-\bm u\Vert_2$, is maximally reduced. Having specified the active column set, a least-squares problem is solved to compute the solution $\bm{c}^{0,\delta,(k)}$. The iterations are continued until the error truncation tolerance $\delta$ is achieved. In general, the complexity of the OMP algorithm is $\mathcal{O}(S\cdot N\cdot P)$ where $S:=\Vert\bm{c}^{0,\delta}\Vert_{0}$ is number of non-zero (dominant) entries of $\bm{c}^{0,\delta}$. The following exhibit depicts an step-by-step implementation of the OMP algorithm. Although we chose OMP in our analysis, we note that further studies are needed to identify the most appropriate greedy algorithm for the purpose of this study. [*Problem $(P_{1,\delta})$:*]{} The majority of available solvers for $\ell_1$-minimization are based on alternative formulations of $(P_{1,\delta})$, such as the $\ell_1$-norm regularized least-squares problem $$\label{eqn:QP} (QP_{\lambda}):\qquad\min_{\bm{c}} \frac{1}{2}\left\Vert\bm{\Psi}\bm{c}-\bm{u}\right\Vert_2^2+\lambda\Vert\bm{Wc}\Vert_{1},$$ or the LASSO problem, [@Tibshirani94], $$\label{eqn:LASSO} (LS_{\tau}):\qquad\min_{\bm{c}} \frac{1}{2}\left\Vert\bm{\Psi}\bm{c}-\bm{u}\right\Vert_2^2\quad\mathrm{subject \;to}\quad\Vert\bm{Wc}\Vert_{1}\leq\tau.$$ It can be shown that for an appropriate choice of scalars $\delta$, $\lambda$, and $\tau$, the problems $(P_{1,\delta})$, $(QP_{\lambda})$, and $(LS_{\tau})$ share the same solution [@Berg08; @Bruckstein09; @Tropp10a]. Among others, the problem $(QP_{\lambda})$ is of particular interest as it is an unconstraint optimization problem. Numerous solvers based on the [*active set*]{} [@Osborne00; @Efron04], [*interior-point continuation*]{} [@Chen01a; @Kim07a] and [*projected gradient*]{} [@Daubechies04; @Combettes05; @Hale08; @Bredies08; @Bioucas-Dias07; @Berg08; @Beck09; @Becker09] methods have been developed for solving the above formulations of the $\ell_1$-minimization problem. In our numerical experiments, we adopt the Spectral Projected Gradient algorithm (SPGL1) proposed in [@Berg08] and implemented in the MATLAB package [SPGL1]{} [@spgl1:2007] to solve the $\ell_1$-minimization problem $(P_{1,\delta})$ in . SPGL1 is based on exploring the so-called Pareto curve, describing the tradeoff between the $\ell_2$-norm of the truncation error $\Vert\bm{\Psi}\bm{c}-\bm u\Vert_{2}$ and the (weighted) $\ell_1$-norm of the solution $\Vert\bm{Wc}\Vert_{1}$, for successive solution iterations. At each iteration, the LASSO problem is solved using the spectral projected gradient technique with a worst-case complexity of $\mathcal{O}(P\ln P)$ where $P$ is the number of columns in $\bm{\Psi}$. Given the error tolerance $\delta$, a scalar equation is solved to identify a value for $\tau$ such that the $(LS_{\tau})$ solution of is identical to that of $(P_{1,\delta})$ in . Besides being efficient for large-scale systems where $\bm{\Psi}$ may not be available explicitly, the SPGL1 algorithm is specifically effective for our application of interest as the truncation error $\Vert\bm{\Psi}\bm{c}-\bm u\Vert_{2}$ is known only approximately. In the next section, we explore some aspects of the proposed scheme through its application to a $1$-$D$ (in space) elliptic stochastic PDE with high-dimensional random diffusion coefficients. Numerical examples {#sec:examples} ================== We consider the solution of a one-dimensional, i.e., $D=1$, version of problem (\[eqn:spde\]), $$\begin{aligned} \label{eqn:1D_spde} &&-\frac{d}{dx}\left(a(x,\omega) \frac{du(x,\omega)}{dx}\right)=1,\quad x\in \mathcal{D}=(0,1),\\ &&\quad u(0,\omega)=u(1,\omega)=0,\nonumber\end{aligned}$$ where the stochastic diffusion coefficient $a(x,\omega)$ is given by the expansion $$\label{eqn:example_kle_a} a(x,\omega)=\bar{a}+\sigma_{a}\sum_{i=1}^{d}\sqrt{\lambda_i}\phi_{i}(x)y_{i}(\omega).$$ Here, $\{\lambda_i\}_{i=1}^{d}$ and $\{\phi_i(x)\}_{i=1}^{d}$ are, respectively, $d$ largest eigenvalues and the corresponding eigenfunctions of the Gaussian covariance kernel $$\label{eqn:gaussian_kernel} C_{aa}(x_1,x_2)=\exp\left[-\frac{(x_1-x_2)^2}{l_c^2}\right],$$ in which $l_c$ is the correlation length of $a(x,\omega)$ that prescribes the decay of the spectrum of $C_{aa}$ in (\[eqn:gaussian\_kernel\]). Random variables $\{y_i(\omega)\}_{i=1}^{d}$ are assumed to be independent and uniformly distributed on $[-1,1]$. The coefficient $\sigma_{a}$ controls the variability of $a(x,\omega)$. We verify the accuracy and efficiency of the present sparse approximation schemes for both moderate and high-dimensional diffusion coefficient $a(x,\omega)$. These two cases are obtained, respectively, by assuming $(l_c,d) = (1/5,14)$ and $(l_c,d) = (1/14,40)$ in (\[eqn:gaussian\_kernel\]) and (\[eqn:example\_kle\_a\]). We further assume that $\bar{a}=0.1$, $\sigma_a=0.03$ when $d=14$, and $\sigma_a=0.021$ when $d=40$. These choices ensure that all realizations of $a(x,\omega)$ are strictly positive on $\mathcal{D}=(0,1)$. Table \[tab:parameters\] summarizes the assumed parameters for the two test cases. Case $\bar{a}$ $\sigma_{a}$ $l_c$ $d$ -------------------- ----------- -------------- -------- ------ I $0.1$ $0.030$ $1/5$ $14$ lI $0.1$ $0.021$ $1/14$ $40$ \[tab:parameters\] : Choices of parameters defining the stochastic description of diffusion coefficient $a(x,\omega)$ in Eq. (\[eqn:example\_kle\_a\]). For both cases, the spatial discretization is done by the Finite Element Method using quadratic elements. A mesh convergence analysis is performed to ensure that spatial discretization errors are inconsequential. The solution statistics are computed using the conventional Monte Carlo simulation, the isotropic sparse grid stochastic collocation with the Clenshaw-Curtis abscissas [@Xiu05a; @Babuska07a], and the proposed sparse approximation techniques. We use the Basis Pursuit Denoising (BPDN) solver implemented in [SPGL1]{} [@spgl1:2007; @Berg08] to solve the $\ell_1$-minimization problem $(P_{1,\delta})$ and the Orthogonal Matching Pursuit (OMP) solver in [SparseLab]{} [@Sparselab] to approximate the $\ell_{0}$-minimization problem $(P_{0,\delta})$. We compare the errors in the mean, standard deviation, and root mean-squares of the solution error at $x=0.5$ using the above methods. The details of the analysis are reported below. Case I: d=14 {#sec:case_d_14} ------------ We consider an increasing number $N=\{29,120,200,280,360,421,600\}$ of random solution samples to evaluate the solution $u$ at $x=0.5$ and, consequently, to compute the PC coefficients of the solution using $\ell_1$- and $\ell_0$-minimization. These samples are nested in the sense that we recycle the previous samples when we perform calculations with larger sample sizes. The nested sampling property of our scheme is of paramount importance in large scale calculations where the computational cost of each solution evaluation is enormous. We note that sample sizes $N=29$ and $N=421$, respectively, correspond to the number of nested abscissas in the level $l=1$ and level $l=2$ of the isotropic stochastic collocation with the Clenshaw-Curtis rule. As elucidated in Section \[sec:sucess\_rec\], the accuracy of our sparse reconstruction depends on the mutual coherence $\mu(\bm{\Psi})$, the sample size $N$, and the truncation error $\Vert\bm{\Psi c}-\bm{u}\Vert_2$ (hence $\delta$). In order to reduce the approximation error, we need to reduce $\Vert\bm{\Psi c}-\bm{u}\Vert_2$, which may be done by increasing $p$ and, therefore, $P$. However, with a fixed number $N$ of samples, an increase in $P$ may result in a larger mutual coherence and, thus, the degradation of the reconstruction accuracy. Therefore, in practice, we start by approximating the lower order PC expansions when $N$ is small and increase $p$ when larger number of samples become available. Notice that such an adaptivity with respect to the order $p$ is a natural way of refining the accuracy of PC expansions, for instance, when the intrusive stochastic Galerkin scheme is adopted [@Ghanem03]. In particular, in this example, for sample sizes $N=\{29,120\}$, we attempt to estimate the coefficients of the $3$rd-order Legendre PC expansion, i.e. $p=3$ and $P=680$. For larger sample sizes $N$, we also include the first $320$ basis function from the $4$th-order chaos, thus resulting in $P=1000$. [Since all of the $4$th-order basis functions are not employed, we need to describe the ordering of our basis construction]{}. We sort the elements of $\{\psi_{\bm{\alpha}}(\bm{y})\}$ such that, for any given order $p$, the random variables $y_i$ with smaller indices $i$ contribute first in the basis. For each analysis, we estimate the truncation error tolerance $\delta$ based on the cross-validation algorithm described in Section \[sec:delta\]. For each $N$, we use $N_r\approx 3N/4$ of the samples (reconstruction set) to compute the PC coefficients $\bm{c}^{1,\delta_r}$ and the rest of the samples (validation set) are used to evaluate the truncation error $\delta_{v}$. The cross-validation is performed for four replications of reconstruction and validation sample sets. We then find the value $\hat{\delta}_{r}$ that minimizes the average of $\delta_{v}$ over the four replications of the cross-validation samples. Given an estimate of the truncation error tolerance $\delta\approx\sqrt{4/3}\hat{\delta}_{r}$, we then use all $N$ samples to compute the coefficients $\bm{c}^{1,\delta}$. Figure \[fig:pce\_coef\_14\] compares the ‘exact’ PC coefficients with those obtained using BPDN and OMP solvers. We only demonstrate the results corresponding to sample sizes $N=\{120,600\}$. An ‘exact’ solution is computed using the level $l=8$ stochastic collocation for which the approximation errors are negligible in our comparisons. We observe that BPDN tends to give less sparse solutions compared to OMP. This is due to the facts that $i)$ the solution is not exactly sparse, i.e., there are many non-zero (although negligible) coefficients $c_{\bm \alpha}$, $ii)$ the $\ell_1$ cost function does not impose a sufficiently large penalty on the small coefficients as does the $\ell_0$ cost function, and $iii)$ the truncation error tolerance $\delta$ may be under-estimated. To reduce this issue, a number of modifications, including the reweighted $\ell_1$-minimization [@Candes08a; @Xu09a; @Needell09; @Khajehnejad10], have been introduced in the literature that are the subjects of our future work. In contrary, OMP results in more sparse solutions as it adds basis function one-at-a-time until the residual falls below the truncation error. However, as is seen in Figs. \[fig:pce\_coef\_14\] (b) and (d), a number of small coefficients are still over-estimated. This is primarily due to under-estimation of the truncation error tolerance $\delta$ in the cross-validation algorithm. [c]{} ![Approximation of polynomial chaos (PC) coefficients $\bm c$ of $u(0.5,\bm{y})$ using BPDN and OMP for $d=14$. (a) BPDN with $N=120$ samples, (b) OMP with $N=120$ samples, (c) BPDN with $N=600$ samples, and (d) OMP with $N=600$ samples . ‘Exact’ coefficients computed from level 8 stochastic collocation with the Clenshaw-Curtis abscissas ([$\square$]{}); BPDN and OMP ($\bullet$).[]{data-label="fig:pce_coef_14"}](pce_coef_14_120_bpdn.eps "fig:"){width="5.0in"} (-345,60) $\vert c_{\bm \alpha_i}\vert$ (-300,25)[$\bm{(a)}$]{}\ ![Approximation of polynomial chaos (PC) coefficients $\bm c$ of $u(0.5,\bm{y})$ using BPDN and OMP for $d=14$. (a) BPDN with $N=120$ samples, (b) OMP with $N=120$ samples, (c) BPDN with $N=600$ samples, and (d) OMP with $N=600$ samples . ‘Exact’ coefficients computed from level 8 stochastic collocation with the Clenshaw-Curtis abscissas ([$\square$]{}); BPDN and OMP ($\bullet$).[]{data-label="fig:pce_coef_14"}](pce_coef_14_120_omp.eps "fig:"){width="5.0in"} (-345,60) $\vert c_{\bm \alpha_i}\vert$ (-300,25)[$\bm{(b)}$]{}\ ![Approximation of polynomial chaos (PC) coefficients $\bm c$ of $u(0.5,\bm{y})$ using BPDN and OMP for $d=14$. (a) BPDN with $N=120$ samples, (b) OMP with $N=120$ samples, (c) BPDN with $N=600$ samples, and (d) OMP with $N=600$ samples . ‘Exact’ coefficients computed from level 8 stochastic collocation with the Clenshaw-Curtis abscissas ([$\square$]{}); BPDN and OMP ($\bullet$).[]{data-label="fig:pce_coef_14"}](pce_coef_14_600_bpdn.eps "fig:"){width="5.0in"} (-345,60) $\vert c_{\bm \alpha_i}\vert$ (-300,25)[$\bm{(c)}$]{}\ ![Approximation of polynomial chaos (PC) coefficients $\bm c$ of $u(0.5,\bm{y})$ using BPDN and OMP for $d=14$. (a) BPDN with $N=120$ samples, (b) OMP with $N=120$ samples, (c) BPDN with $N=600$ samples, and (d) OMP with $N=600$ samples . ‘Exact’ coefficients computed from level 8 stochastic collocation with the Clenshaw-Curtis abscissas ([$\square$]{}); BPDN and OMP ($\bullet$).[]{data-label="fig:pce_coef_14"}](pce_coef_14_600_omp.eps "fig:"){width="5.0in"} (-345,60) $\vert c_{\bm \alpha_i}\vert$ (-230,-4)[Index of PC coefficients ($i$)]{} (-300,25)[$\bm{(d)}$]{} The convergence of the mean, standard deviation, and root mean-squares of the approximation error for $u(0.5,\bm{y})$ is illustrated in Figs. \[fig:mean\_sd\_mse\_cross\_14\] (a), (b), and (c), respectively. For the case of stochastic collocation, we apply sparse grid quadrature (cubature) integration rule to directly compute the mean and the standard deviation. The root mean-squares error of the Monte Carlo and the stochastic collocation solution are evaluated by estimating the corresponding PC coefficients using sampling and sparse grid quadrature integration, respectively, and then comparing them with the exact coefficients. [cc]{} ![Comparison of relative error in solution statistics at $x=0.5$ for the Monte Carlo simulation, isotropic sparse grid stochastic collocation with the Clenshaw-Curtis abscissas, and the proposed sparse approximations (BPDN and OMP) for $d=14$. Two sets of independent random samples of $u(0.5,\bm{y})$ are generated first and are used for the Monte Carlo simulation, BPDN, and OMP. The solid and dashed lines correspond to the first and second sets of samples, respectively. (a) Relative error in the mean; (b) Relative error in the standard deviation; (c) Relative root-mean-squares (rms) error; and (d) Estimation of $\delta$ using cross-validation: $\delta$ is computed from $\delta_{r}$ for which $\delta_{v}$ is minimum. (Monte Carlo simulation ([$\square$]{}); stochastic collocation ($\diamond$); BPDN ($\circ$); OMP ($\triangledown$)).[]{data-label="fig:mean_sd_mse_cross_14"}](mean_14d.eps "fig:"){width="2.67in"} (-205,35) [Rel. error in mean]{} (-95,-10)[$N$]{} (-30,130)[$\bm{(a)}$]{} & ![Comparison of relative error in solution statistics at $x=0.5$ for the Monte Carlo simulation, isotropic sparse grid stochastic collocation with the Clenshaw-Curtis abscissas, and the proposed sparse approximations (BPDN and OMP) for $d=14$. Two sets of independent random samples of $u(0.5,\bm{y})$ are generated first and are used for the Monte Carlo simulation, BPDN, and OMP. The solid and dashed lines correspond to the first and second sets of samples, respectively. (a) Relative error in the mean; (b) Relative error in the standard deviation; (c) Relative root-mean-squares (rms) error; and (d) Estimation of $\delta$ using cross-validation: $\delta$ is computed from $\delta_{r}$ for which $\delta_{v}$ is minimum. (Monte Carlo simulation ([$\square$]{}); stochastic collocation ($\diamond$); BPDN ($\circ$); OMP ($\triangledown$)).[]{data-label="fig:mean_sd_mse_cross_14"}](sd_14d.eps "fig:"){width="2.67in"} (-205,35) [Rel. error in s.d.]{} (-95,-10)[$N$]{} (-30,130)[$\bm{(b)}$]{}\ ![Comparison of relative error in solution statistics at $x=0.5$ for the Monte Carlo simulation, isotropic sparse grid stochastic collocation with the Clenshaw-Curtis abscissas, and the proposed sparse approximations (BPDN and OMP) for $d=14$. Two sets of independent random samples of $u(0.5,\bm{y})$ are generated first and are used for the Monte Carlo simulation, BPDN, and OMP. The solid and dashed lines correspond to the first and second sets of samples, respectively. (a) Relative error in the mean; (b) Relative error in the standard deviation; (c) Relative root-mean-squares (rms) error; and (d) Estimation of $\delta$ using cross-validation: $\delta$ is computed from $\delta_{r}$ for which $\delta_{v}$ is minimum. (Monte Carlo simulation ([$\square$]{}); stochastic collocation ($\diamond$); BPDN ($\circ$); OMP ($\triangledown$)).[]{data-label="fig:mean_sd_mse_cross_14"}](rmse_14d.eps "fig:"){width="2.67in"} (-205,40) [Rel. rms error]{} (-95,-10)[$N$]{} (-30,130)[$\bm{(c)}$]{} & ![Comparison of relative error in solution statistics at $x=0.5$ for the Monte Carlo simulation, isotropic sparse grid stochastic collocation with the Clenshaw-Curtis abscissas, and the proposed sparse approximations (BPDN and OMP) for $d=14$. Two sets of independent random samples of $u(0.5,\bm{y})$ are generated first and are used for the Monte Carlo simulation, BPDN, and OMP. The solid and dashed lines correspond to the first and second sets of samples, respectively. (a) Relative error in the mean; (b) Relative error in the standard deviation; (c) Relative root-mean-squares (rms) error; and (d) Estimation of $\delta$ using cross-validation: $\delta$ is computed from $\delta_{r}$ for which $\delta_{v}$ is minimum. (Monte Carlo simulation ([$\square$]{}); stochastic collocation ($\diamond$); BPDN ($\circ$); OMP ($\triangledown$)).[]{data-label="fig:mean_sd_mse_cross_14"}](delta_sigma_14d_N600.eps "fig:"){width="2.95in"} (-215,75) $\delta_{v}$ (-100,0)[$\delta_{r}$]{} (-170,140)[$N=600, p=4$]{} (-135,8)[ $\hat{\delta}_{r}$]{} (-172,8)[ $\hat{\delta}_{r}$]{} (-45,130)[$\bm{(d)}$]{} To make a meaningful comparison, for each $N$, the samples used to compute the solution statistics by the conventional Monte Carlo, BPDN, and OMP are identical. In this sense, the sparse approximation using $\ell_1$- and $\ell_0$-minimizations may be viewed as only post-processing steps in the Monte Carlo simulation. As the sample sizes are finite, the estimates of the PC coefficients $\bm c$ are sample dependent and are in fact random. To demonstrate the convergence of the algorithm with respect to different sets of samples, we repeat the analysis for two independent sets of $N$ samples and report the corresponding statistics errors with solid and dashed lines. Although for different solution samples of size $N$ the estimates of $\bm c$ and the solution statistics are not identical, the approximation converges, with large probability, for any set of samples with sufficiently large size $N$ (see Theorem \[the:main\]). Figure \[fig:mean\_sd\_mse\_cross\_14\] (d) illustrates the statistical estimation of $\delta$ using the cross-validation approach described in Section \[sec:delta\]. The estimation of $\delta$ is slightly different in BPDN and OPM, this is a consequence of different reconstruction accuracy of these techniques. Moreover, the solution of the BPDN algorithm is less sensitive to small perturbations in the truncation error $\delta$ compared to that of the OMP algorithm. This is justified by the fact that the $\ell_0$-norm is highly discontinuous. [**Remark:**]{} Despite the conventional implementation of the stochastic collocation approach where the approximation refinement requires a certain number of extra samples, the $\ell_1$- and $\ell_0$-minimizations may be implemented using arbitrary numbers of additional samples, which is an advantage, particularly, when only a limited number of samples is afforded. Case II: d=40 {#sec:case_d_40} ------------- The objective of this example is to highlight that a sparse reconstruction may lead to significant computational savings for problems with high-dimensional random inputs. Similar to the analysis of Case I described in Section \[sec:case\_d\_14\], we compute the solution statistics using multiple numbers of independent samples. More specifically, we evaluate the solution at $x=0.5$ for independent samples of size $N=\{81,200,400,600,800,1000\}$. The number of grid points in the level $l=1$ and $l=2$ of the Clenshaw-Curtis rule in dimension $d=40$ is $N=81$ and $N=3281$, respectively. To obtain a reference solution, the $3$rd order PC coefficients $\bm c$ of the solution at $x=0.5$ are computed using level $l=5$ stochastic collocation with the Clenshaw-Curtis rule. [c]{} ![Approximation of polynomial chaos (PC) coefficients $\bm c$ of $u(0.5,\bm{y})$ using BPDN and OMP for $d=40$. (a) BPDN with $N=200$ samples, (b) OMP with $N=200$ samples, (c) BPDN with $N=1000$ samples, and (d) OMP with $N=1000$ samples . ‘Exact’ coefficients computed from level 5 stochastic collocation with the Clenshaw-Curtis abscissas ([$\square$]{}); BPDN and OMP ($\bullet$).[]{data-label="fig:pce_coef_40"}](pce_coef_40_200_bpdn.eps "fig:"){width="4.8in"} (-335,60) $\vert c_{\bm \alpha_i}\vert$ (-300,25)[$\bm{(a)}$]{}\ ![Approximation of polynomial chaos (PC) coefficients $\bm c$ of $u(0.5,\bm{y})$ using BPDN and OMP for $d=40$. (a) BPDN with $N=200$ samples, (b) OMP with $N=200$ samples, (c) BPDN with $N=1000$ samples, and (d) OMP with $N=1000$ samples . ‘Exact’ coefficients computed from level 5 stochastic collocation with the Clenshaw-Curtis abscissas ([$\square$]{}); BPDN and OMP ($\bullet$).[]{data-label="fig:pce_coef_40"}](pce_coef_40_200_omp.eps "fig:"){width="4.8in"} (-335,60) $\vert c_{\bm \alpha_i}\vert$ (-300,25)[$\bm{(b)}$]{}\ ![Approximation of polynomial chaos (PC) coefficients $\bm c$ of $u(0.5,\bm{y})$ using BPDN and OMP for $d=40$. (a) BPDN with $N=200$ samples, (b) OMP with $N=200$ samples, (c) BPDN with $N=1000$ samples, and (d) OMP with $N=1000$ samples . ‘Exact’ coefficients computed from level 5 stochastic collocation with the Clenshaw-Curtis abscissas ([$\square$]{}); BPDN and OMP ($\bullet$).[]{data-label="fig:pce_coef_40"}](pce_coef_40_1000_bpdn.eps "fig:"){width="4.8in"} (-335,60) $\vert c_{\bm \alpha_i}\vert$ (-300,25)[$\bm{(c)}$]{}\ ![Approximation of polynomial chaos (PC) coefficients $\bm c$ of $u(0.5,\bm{y})$ using BPDN and OMP for $d=40$. (a) BPDN with $N=200$ samples, (b) OMP with $N=200$ samples, (c) BPDN with $N=1000$ samples, and (d) OMP with $N=1000$ samples . ‘Exact’ coefficients computed from level 5 stochastic collocation with the Clenshaw-Curtis abscissas ([$\square$]{}); BPDN and OMP ($\bullet$).[]{data-label="fig:pce_coef_40"}](pce_coef_40_1000_omp.eps "fig:"){width="4.8in"} (-335,60) $\vert c_{\bm \alpha_i}\vert$ (-300,25)[$\bm{(d)}$]{} (-230,-4)[Index of PC coefficients ($i$)]{} [cc]{} ![Comparison of relative error in solution statistics at $x=0.5$ for the Monte Carlo simulation, isotropic sparse grid stochastic collocation with the Clenshaw-Curtis abscissas, and the proposed sparse approximations (BPDN and OMP) for $d=40$. Two sets of independent random samples of $u(0.5,\bm{y})$ are generated first and are used for the Monte Carlo simulation, BPDN, and OMP. The solid and dashed lines correspond to the first and second sets of samples, respectively. (a) Relative error in mean; (b) Relative error in standard deviation; (c) Relative root-mean-squares (rms) error; and (d) Estimation of $\delta$ using cross-validation: $\delta$ is computed from $\delta_{r}$ for which $\delta_{v}$ is minimum. (Monte Carlo simulation ([$\square$]{}); stochastic collocation ($\diamond$); BPDN ($\circ$); OMP ($\triangledown$)).[]{data-label="fig:mean_sd_mse_cross_40"}](mean_40d.eps "fig:"){width="2.67in"} (-205,35) [Rel. error in mean]{} (-100,-10)[$N$]{} (-30,130)[$\bm{(a)}$]{} & ![Comparison of relative error in solution statistics at $x=0.5$ for the Monte Carlo simulation, isotropic sparse grid stochastic collocation with the Clenshaw-Curtis abscissas, and the proposed sparse approximations (BPDN and OMP) for $d=40$. Two sets of independent random samples of $u(0.5,\bm{y})$ are generated first and are used for the Monte Carlo simulation, BPDN, and OMP. The solid and dashed lines correspond to the first and second sets of samples, respectively. (a) Relative error in mean; (b) Relative error in standard deviation; (c) Relative root-mean-squares (rms) error; and (d) Estimation of $\delta$ using cross-validation: $\delta$ is computed from $\delta_{r}$ for which $\delta_{v}$ is minimum. (Monte Carlo simulation ([$\square$]{}); stochastic collocation ($\diamond$); BPDN ($\circ$); OMP ($\triangledown$)).[]{data-label="fig:mean_sd_mse_cross_40"}](sd_40d.eps "fig:"){width="2.67in"} (-205,35) [Rel. error in s.d.]{} (-100,-10)[$N$]{} (-30,130)[$\bm{(b)}$]{}\ ![Comparison of relative error in solution statistics at $x=0.5$ for the Monte Carlo simulation, isotropic sparse grid stochastic collocation with the Clenshaw-Curtis abscissas, and the proposed sparse approximations (BPDN and OMP) for $d=40$. Two sets of independent random samples of $u(0.5,\bm{y})$ are generated first and are used for the Monte Carlo simulation, BPDN, and OMP. The solid and dashed lines correspond to the first and second sets of samples, respectively. (a) Relative error in mean; (b) Relative error in standard deviation; (c) Relative root-mean-squares (rms) error; and (d) Estimation of $\delta$ using cross-validation: $\delta$ is computed from $\delta_{r}$ for which $\delta_{v}$ is minimum. (Monte Carlo simulation ([$\square$]{}); stochastic collocation ($\diamond$); BPDN ($\circ$); OMP ($\triangledown$)).[]{data-label="fig:mean_sd_mse_cross_40"}](rmse_40d.eps "fig:"){width="2.67in"} (-205,40) [Rel. rms error]{} (-100,-10)[$N$]{} (-30,130)[$\bm{(c)}$]{} & ![Comparison of relative error in solution statistics at $x=0.5$ for the Monte Carlo simulation, isotropic sparse grid stochastic collocation with the Clenshaw-Curtis abscissas, and the proposed sparse approximations (BPDN and OMP) for $d=40$. Two sets of independent random samples of $u(0.5,\bm{y})$ are generated first and are used for the Monte Carlo simulation, BPDN, and OMP. The solid and dashed lines correspond to the first and second sets of samples, respectively. (a) Relative error in mean; (b) Relative error in standard deviation; (c) Relative root-mean-squares (rms) error; and (d) Estimation of $\delta$ using cross-validation: $\delta$ is computed from $\delta_{r}$ for which $\delta_{v}$ is minimum. (Monte Carlo simulation ([$\square$]{}); stochastic collocation ($\diamond$); BPDN ($\circ$); OMP ($\triangledown$)).[]{data-label="fig:mean_sd_mse_cross_40"}](delta_sigma_40d_N1000.eps "fig:"){width="2.9in"} (-215,75) $\delta_{v}$ (-170,140)[$N=1000, p=4$]{} (-105,0)[$\delta_{r}$]{} (-115,8)[ $\hat{\delta}_{r}$]{} (-160,8)[ $\hat{\delta}_{r}$]{} (-45,135)[$\bm{(d)}$]{} For $N=\{81,200\}$ we only estimate the coefficients associated with the $2$nd-order PC expansion, i.e. $p=2$ and $P=861$. For larger sample sizes, we also include the first $639$ basis functions from the $3$th-order chaos, thus leading to $P=1500$. For each combination of $N$ and $p$, we estimate the truncation error $\delta$ using an identical cross-validation procedure described in Section \[sec:case\_d\_40\]. Figure \[fig:pce\_coef\_40\] illustrates the estimation of PC coefficients of $u(0.5,\bm{y})$ with BPDN and OMP algorithms with $N=200$ and $N=1000$. We again note that the recovered solution from the BPDN algorithm is less sparse as compared to that of the OMP approach, although the over-estimated coefficients (mostly from the second order term) are indeed small. As the samples size $N$ is increased, we are naturally able to recover more dominant coefficients on the expansion. Figure \[fig:mean\_sd\_mse\_cross\_40\] depicts the convergence of the statistics of the solution as functions of the sample size $N$ as well as one instance of the estimation of the truncation error tolerance $\delta$. The implementation details are similar to those described in Section \[sec:case\_d\_14\] for the case of $d=14$. It is worth highlighting that the computational saving of the present sparse approximations (in terms of the number of samples needed to achieve a certain accuracy) compared to the isotropic sparse grid collocation is even larger for the higher-dimensional ($d=40$) problem. This is due to the fact that the number of samples needed to recover the solution using $\ell_1$- and $\ell_0$-minimization is dictated more by the number of dominant terms in the PC expansion compared to the total number of terms $P$, as in Theorem \[the:main0\]. Conclusion {#sec:conclusion} ========== The present study proposes a [*non-intrusive*]{} and [*non-adapted*]{} approach based on the [*compressive sampling*]{} formalism for the approximation of sparse solution of stochastic PDEs. When sufficiently sparse in the polynomial chaos (PC) basis, the compressive sampling enables an accurate recovery of the solution using a number of random solution samples that is significantly smaller than the cardinality of the PC basis. Sparse PC approximations based on $\ell_0$- and $\ell_1$-minimization approaches have been introduced and implemented using the Basis Pursuit Denoising (BPDN) and the Orthogonal Matching Pursuit (OMP) algorithms, respectively. Probabilistic bounds based on the [*concentration of measure*]{} phenomenon have been derived to verify the convergence and the stability of the present sparse constructions. The performance and efficiency of the proposed techniques are explored through their application to a linear elliptic PDE with high-dimensional random diffusion coefficients where the sparsity of the solution with respect to the PC basis is guaranteed. The proposed formalism to recover the sparse PC expansion of stochastic functions is not restricted to the case of the elliptic PDEs, as its underlying applicability assumptions are universal. Although the discussions of this work have been focused on the particular case of the Legendre PC expansions, the proposed framework can be readily extended to other bases such as the Hermite PC (when the random variables $\bm y$ are standard Gaussian). Acknowledgments {#acknowledgments .unnumbered} =============== The first author acknowledges the support of the United States Department of Energy under Stanford’s Predictive Science Academic Alliance Program (PSAAP) for the preliminary stages of his work. The second author acknowledges the support of the National Science Foundation via NSF grant CMMI-092600 and of the United States Department of Energy under Caltech’s Predictive Science Academic Alliance Program (PSAAP).

ArXiv

--- address: - | Department of Physics, Northeastern University\ Boston, MA 02115, USA\ E-mail: [email protected] - | Department of Physics, Northeastern University, Boston, MA 02115, USA[^1]\ Physikalisches Institut, Universitat Bonn, Nussallee 12, D-53115 Bonn, Germany\ Max-Planck-Institute fuer Kernphysik, Saupfercheckweg 1, D-69117 Heidelberg, Germany\ E-mail: [email protected] author: - ACHILLE CORSETTI - PRAN NATH title: RECENT DEVELOPMENTS IN SUPERSYMMETRIC DARK MATTER --- =cmr8 1.5pt \#1\#2\#3\#4[[\#1]{} [**\#2**]{}, \#3 (\#4)]{} Introduction ============ Because of the recent significant activity in dark matter searches on the experimental side[@dama; @cdms; @hdms] there is renewed interest in the theoretical analyses of dark matter which are significantly more refined than in the previous years. Among the refinements is the inclusions of the effects of uncertainties in the input parameters in the theoretical predictions of event rates and of the neutralino proton cross sections as well inclusion of the effects of non-universalities, CP violating effects and the effects of coannihilation. The content of the paper is as follows: In Sec.2 we discuss the effects of uncertainties in the analyses of dark matter. These include the effects of uncertainties in the wimp velocity and in the wimp relic density, and the effects of uncertainties in the quark densities in the proton in the analyses of dark matter. In Sec.3 we give a discussion of the maximum and the minimum elastic neutralino-proton cross-sections. In Sec.4 we discuss the effects of non-universalities and specifically the non-universalities in the gaugino sector on dark matter analyses. A brief discussion of the effects of $\mu$ on the composition of the neutralino and its role in determining the nature of cold dark matter, i.e., if it is dominantly a bino, a wino or a higgsino is given in Sec.5. A comparison of the direct and the indirect detection of dark matter is given in Sec.6. In Sec.7 we give a discussion of the annual modulation effect in the direct detection of dark matter. In Sec.8 we give a brief discussion of the effects of CP phases on dark matter. Conclusions are given in Sec.9. Uncertainties in Theoretical Analyses of Dark Matter ==================================================== The direct detection of dark matter has been investigated by many authors ( see Ref.[@direct] for some of the recent works). Several types of uncertainties enter in these analyses such as the effects of variations in the local wimp density, the effects of variations in the wimp velocity range, and the effects of uncertainties in the quark densities in the nucleons. Other effects not considered here are the uncertainties in the nuclear form factors, and effects of halo models[@sikvie] on the event rates in dark matter detection. We begin with a discussion of the analyses of uncertainties in wimp density[@an1] and velocity[@roszkowski; @bottino1; @corsetti1]. The current range of local wimp density lies in the range[@gates] $(0.2-0.7) GeV cm^{-3}$. Defining $\xi=\rho_{\tilde\chi_1^0}/\rho_0$ one can parameterize the local wimp density in term of $\xi$ which for $\rho_0=0.3 GeVcm^{-3}$ gives $0.7\leq \xi\leq 2.3$. For the wimp velocity one typically assumes a Maxwellian velocity distribution for the wimps. Estimates of the rms wimp velocity range give[@knapp] $ v=270\pm 50$ km/s. The analysis is carried out within the framework of the minimal supergravity model (SUGRA)[@chams]. The soft SUSY breaking sector in the minimal version mSUGRA of this theory is given by $m_0,m_{1/2}, A_0$, and $\tan\beta$ where, $m_0$ is the universal scalar mass, $m_{1/2}$ is the universal gaugino mass, $A_0$ is the universal trilinear coupling, and $\tan\beta=<H_2>/<H_1>$ where $H_2$ gives mass to the up quarks and $H_1$ gives mass to the down quarks and the leptons. The Higgs mixing parameter $\mu$ is determined by the constraint of radiative breaking of the electro-weak symmetry. It is also useful to define a fine tuning parameter $\Phi$ so that[@ccn] $\Phi=(\mu^2/M_Z^2+1/4)^{1/2}$. The fine tuning parameter defines how heavy the SUSY spectrum gets. The effects of the uncertainly in the event rate as a function of the fine tuning parameter due to variations in the wimp relic density and in the wimp velocity are shown in Fig.1[@corsetti1]. One finds that the effects of this type of uncertainty can lead to a variation in the rates by a factor of 2-3.\ Next we discuss the uncertainties in neutralino-proton cross-section arising from errors in the quark densities in the proton[@bottino2; @efo; @corsetti2]. The basic interaction governing the $\chi -p$ scattering with CP conservation is the four Fermi interaction given by ${\cal L}_{eff}=\bar{\chi}\gamma_{\mu} \gamma_5 \chi \bar{q} \gamma^{\mu} (A P_L +B P_R) q + C\bar{\chi}\chi m_q \bar{q} q +D \bar{\chi}\gamma_5\chi m_q \bar{q}\gamma_5 q$. For heavy target materials the neutralino-nucleus scattering is dominated by the scalar interactions which is controlled by the scalar $\chi -p$ cross-section where $$\sigma_{\chi p}(scalar)=\frac{4\mu_r^2}{\pi} (\sum_{i=u,d,s}f_i^pC_i+\frac{2}{27}(1-\sum_{i=u,d,c}f_i^p) \sum_{a=c,b,t}C_a)^2$$ Here $\mu_r$ is the reduced mass and $f_i^p$ (i=u,d,s quarks) are the (u,d,s) quark densities defined by $m_pf_i^p=<p|m_{qi}\bar q_iq_i|p>$ (i=u,d,s). In the above C is the strength of the scalar interaction and consists of s channel contributions from the higgs $h^0, H^0$ exchange and t channel contributions from the sfermion exchange so that $C=C_{h^0}+C_{H^0}+C_{\tilde{f}}$. The uncertainly in the theoretical predictions of $\sigma_{\chi p}(scalar)$ is dominated by the uncertainty in the quark densities $f_i^p$. To study the uncertainties in $f_i^p$ it is best first to solve the quark densities analytically in terms of some judiciously chosen parameters. One finds[@corsetti2; @corsetti3] $$\begin{aligned} f_{(u,d)}^p=\frac{m_{(u,d)}}{m_u+m_d}(1\pm\xi)\frac{\sigma_{\pi N}}{m_p},~~~ f_s^p=\frac{m_s}{m_u+m_d}(1-x)\frac{\sigma_{\pi N}}{m_p}\end{aligned}$$ where the parameters $\xi$, x and $\sigma_{\pi N}$ are defined by $$\begin{aligned} \xi= \frac{<p|\bar uu-\bar dd|p>}{<p|\bar uu+\bar dd|p>},~~~ x= \frac{<p|\bar uu+\bar dd-2\bar ss|p>}{<p|\bar uu+\bar dd|p>}\nonumber\\ <p|2^{-1}(m_u+m_d)(\bar uu+\bar dd|p>=\sigma_{\pi N}\end{aligned}$$ Similarly one can analytically solve for the quark densities in the neutron and the analytic relations provide an interesting connection between the quark densities in the proton and in the neutron. One finds that independent of the details of any input one has[@corsetti2]$f_u^pf_d^p=f_u^nf_d^n$. One can use the analysis on the baryon mass splittings[@hcheng] to determine the ratio $\xi/x$. One finds[@corsetti2] $\xi/x=0.196$. Using the determination of x from lattice gauge analyses[@borasoy; @dong; @fukugita] one finds $\xi=0.132\pm 0.035$. Additional uncertainties can arise from the quark mass ratios. Here results from chiral perturbation theory[@gl; @saino] give $\frac{m_u}{m_d}=0.553\pm 0.043$, $\frac{m_s}{m_d}=18.9\pm 0.8$. Using the inputs above one finds $f_u^p=0.021\pm 0.004$, $f_d^p=0.029\pm 0.006$, $f_s^p=0.21\pm 0.12$ and $f_u^n=0.016\pm 0.003$, $f_d^n=0.037\pm 0.007$, $f_s^n=0.21\pm 0.12$. Maximum and Minimum Neutralino-proton Cross Sections ==================================================== We discuss now the numerical results of the neutralino proton cross-sections with the quark densities as discussed in Sec. 2. In Fig.2 $\sigma_{\chi -p}$ is plotted exhibiting the effects of the variations of the quark density as a function of $m_{\chi}$. The above analysis shows that the $\chi -p$ cross section cannot be computed to an accuracy of better than a factor of 3-5 with the current level of uncertainties in the input data. Fig.2 also gives a plot of the maximum and the minimum of $\sigma_{\chi -p}$ as a function of the neutralino mass. In the analysis of Fig.2 we have allowed $m_0$ and $m_{1/2}$ to vary in the range up to 1 TeV and the constraint from the flavor changing neutral current process $b\rightarrow s\gamma$ is imposed[@fcnc]. The analysis of this figure does not include the effects of the spin dependent contribution which could change the scatter plot. Specifically, the minimum cross sections are sensitive to the inclusion of the spin dependent part. Similar analyses can be found in other recent references[@mandic; @lahanas].\ Effects of Non-universalities on Dark Matter ============================================ mSUGRA is based on the assumption of a flat Kahler potential. However, the nature of physics at the Planck scale is not fully understood. Thus in general one should allow for the possibility of a curved Kahler potential. This would lead to non-universalities in the scalar sector. However, there are stringent constraints on the types of non-universalities allowed in the scalar sector due to the FCNC constraint. For example, the FCNC constraint in very strong on the amount of non-universality allowed in the first vs the second generation sector. However, this constraint is not so strong for the Higgs sector and for the third generation sector. Effects of these constraints have been studied in detail in Refs.[@nonuni; @accomando]. One finds that in general the presence of non-universalities can increase the cross-sections by a factor of 10 or more. In addition to modifying the Kahler potential Planck scale physics can also modify the gauge kinetic energy function. In general the gauge kinetic energy function transforms as the symmetric product of two adjoint representations. For the case of $SU(5)$ one finds that the gauge kinetic energy function $f_{\alpha\beta}$ transforms as the symmetric product of ${\bf 24\times 24}$ in SU(5) which contains the following irreducible representations of $SU(5)$: $$({\bf 24}\times {\bf 24})_{symm}={\bf 1}+{\bf 24}+{\bf 75} +{\bf 200}$$ The SU(5) singlet in the product on the right hand side of Eq.(4) leads to a universal gaugino mass while, the additional terms generate non-universalities in the gaugino masses at the GUT scale. Thus the $SU(3)_C\times SU(2)_L\times U(1)$ gaugino masses at the GUT scale are in general admixtures of all the allowed representations. This admixture leads to the following relation for the gaugino masses at the GUT scale $$\tilde m_i(0)=m_{\frac{1}{2}}(1+ \sum_r c_r n_i^r)$$ where $n_i^r$ depend on one of the representations on the right hand side in Eq.(4) and on the subgroup i. In addition to the appearance of gaugino mass non-universalities of the above type in supergravity models, one also finds quite naturally non-universalities in a broad class of string models: heterotic, Horava-Witten and as well in brane models based on TypeI/Type IIB string compactifications. Now the value of $\mu$ is in general very sensitively dependent on non-universalities. For the case of non-universalities in the scalar sector one can exhibit in an analytic fashion the dependence of $\mu$ on non-universalities in the Higgs sector and in the third generation sector. The analysis shows that $\mu$ is sensitively dependent on the non-universalities and their effects can significantly decrease the value of $\mu$ and thus affect gaugino vs higgsino composition of the neutralino. An analysis of the effects of non-universalities in the scalar sector is given in Ref.[@nonuni]. A similar phenomenon occurs for the case of gaugino sector non-universalities. Here also one can exhibit the dependence of $\mu$ on non-universalities. One finds[@corsetti2] $$\tilde \mu^2= \mu^2_{0}+\sum_r \frac{\partial\tilde\mu^2} {\partial c_r} c_r+ O(c_r^2)$$ where $\partial \mu_{24}^2/\partial c_{24}>0, \partial \mu_{75}^2/\partial c_{75}>0, \partial \mu_{200}^2/\partial c_{200}<0$. Again with an appropriate choice of the sign of $c_r$ one finds that $\mu$ becomes smaller relative to its universal value. An analysis of the effects of non-universalities in the gaugino sector on $\sigma_{\chi -p}$ is given in Ref.[@corsetti2]. Fig.3 gives a typical illustration of the effects of gaugino non-universality showing that a significant enhancement of $\sigma_{\chi -p}$ can occur in the presence of non-universalities. Bino, Higgsino, and Wino Dark Matter ==================================== The neutralino is in general a mixture of gauginos and higgsinos, $$\chi_1=X_{11} \tilde B+ X_{12}\tilde W_3 + X_{13}\tilde H_1 + X_{14} \tilde H_2$$ where $\tilde B$ is the bino, $\tilde W$ the neutral wino, and $\tilde H_1, \tilde H_2$ are higgsinos. Now an analysis of the neutralino mass matrix shows that for large values of the Higgs mixing parameter $\mu$ one finds that the neutralino is essentially a bino and this situation is realized over a large part of the SUGRA parameter space[@bino]. Thus SUGRA models in general predict a bino like cold dark matter at least over a major portion of the parameter space. However, in certain limited regions of the SUGRA parameter space $\mu$ can become small. In this case the higgsino and the wino components can become large and one may have cold dark matter which is higgsino like or wino like. This phenomenon can be easily understood by examining the $|\mu|>>M_Z$ limit of the components $X_{1n}$. Here one finds that the bino component in the neutralino is given by $ X_{11}\simeq 1-(M_Z^2/2\mu^2)sin^2\theta_W$, the wino component is of size $X_{12}\simeq (M_Z^2/2m_{\chi_1}^2\mu)sin2\theta_W sin\beta$, and the higgsino components have the sizes $ X_{13}\simeq -(M_Z/\mu)sin2\theta_W sin\beta$, and $X_{14}\simeq (M_Z/\mu)sin2\theta_W sin\beta$. The above illustrates the sensitive dependence of the bino, the wino and the higgsino components on $\mu$. One consequence of the effect of large higgsino components is that the scalar cross section which depends on the product of the gaugino and the higgsino components increases as the higgsino components increase[@nonuni]. An analysis of CDM with large higgsino components in the context of MSSM is given in Ref.[@drees] and with large higgsino and wino components in the context of anomaly mediated breaking of supersymmetry in Ref.[@wells] Comparison of Direct and Indirect Detection =========================================== Indirect detection is complementary to the direct detection in the search for dark matter. The most interesting indirect signal for dark matter arises from the capture and the subsequent annihilation of the neutralinos in the center of the Sun and the Earth. Some of the remnants in the annihilation of the neutralinos are the neutrinos which propagate and undergo charged current interactions in the rock surrounding the detector and produce upward moving muons. The outgoing muon flux can be written in the form $\Phi_{\mu}=\Gamma_{A}f$ where $\Gamma_{A}$ is the $\chi_1-\chi_1$ annihilation rate in the center of the Earth or the Sun and f is the product of remaining factors. It is $\Gamma_{A}$ which is the quantity sensitive to SUSY. One can parameterize $\Gamma_A$ by[@greist] $$\Gamma_A=\frac{C}{2} tanh^2(t/\tau)$$ Here C is the neutralino capture rate, t is the lifetime of the Earth or the Sun, and $\tau =(CC_A)^{-1/2}$ where $C_A$ is determined by the wimp annihilation cross section. An equilibrium between capture and annihilation is reached for the case $t >>\tau$ and in this case one has $\Gamma_A\sim \frac{C}{2}$. The equilibrium condition is strongly dependent on the susy parameter space. For the case of the sun the equilibrium condition is satisfied essentially over all of the parameter space while for the Earth it is satisfied only over part of the parameter space. This disparity shows up very strongly in the profile of the muon flux from the Earth and the Sun. In Fig.4 we exhibit the muon flux for the Earth and the Sun plotted against the direct detection rate for the germanium detector. The analysis of Fig.4 shows that for relatively large direct detection rates, it is the indirect detection from the Earth which is competitive with the direct detection while for relatively low direct detection rates it is the indirect detection from the Sun which is competitive with the direct detection. Thus indirect detections from the Earth and the Sun are complementary. For other recent analyses of indirect detection see Refs.[@berez; @wilczek; @edsjo]. Annual Modulation Effect in Direct Detection ============================================ The annual modulation effect in the direct detection of dark matter is a potentially important signal for the observation of WIMP like dark matter. The effect arises due to the periodicity of the velocity of the Earth $v_E$ relative to the galaxy,i.e., $v_E=v_S+v_0 cos\gamma cos\omega (t-t_0)$ where $v_0(=30km/s)$ is Earth’s orbital velocity around the Sun, $v_S (=232km/s)$ is the Sun’s velocity relative to the galaxy, $\gamma (\simeq 60^0)$ is the inclination of Earth’s orbit relative to the galactic plane, and $t_0$ is June 2. The effects of this motion can produce a modulation effect of about $7\%$ in the scattering event rates for wimps. The DAMA experiment claims to see an effect[@dama]. The annual modulation signal has been analyzed theoretically in several papers[@bottino4; @an1; @green; @vergados]. In the future one expects further data on this exciting possibility. Further, one needs to reduce the current ambiguities in the theoretical predictions of dark matter such as those discussed in Sec.2. This requires more accurate chiral perturbation and lattice gauge analyses. Effects of Large CP Phases on Dark Matter ========================================= It is well known that supersymmetry brings in new sources of CP violation since the soft SUSY parameters are in general complex. These CP phases induce additional corrections to the electric dipole moments of the electron and of the neutron. We consider here the case where the CP phases are large and the edm constraints are satisfied. (For an abbreviated set of references on large phases see Ref.[@largephases]). Large phases affect the analysis of dark matter and the detailed analyses of event rates for direct detection show that the effects of the phases can change the event rates by an order of magnitude or more[@fos; @cin]. However, with the inclusion of the emd constraints one finds that the effects are significantly reduced although they are still substantial[@cin]. Large phases also induce a mixing between the CP even and the CP odd higgs states[@pilaftsis; @in2] which can affect dark matter[@in2]. Conclusion ========== In this review we have presented the very recent developments in the theoretical analyses of supersymmetric dark matter. These include the effects of uncertainties on the wimp density and on the wimp velocity for the Milky Way wimps, effects of uncertainties in quark densities in the theoretical predictions of $\sigma_{\chi- p}$, and the effects of non-universalities in the gaugino sector. A analysis of the maximum and of the minimum cross sections for $\chi-p$ scattering was given. We have also given a comparison of the direct and of the indirect detection of dark matter and discussed briefly other recent developments for the detection of dark matter. An important topic not discussed is the subject of coannihilation[@coanni] which can significantly extend the domain of the allowed neutralino masses. In the future one expects that more sensitive detectors[@spooner; @genius] will be able to probe more deeply the parameter space of SUGRA and other SUSY models. Acknowledgments {#acknowledgments .unnumbered} =============== One of us(PN) wishes to thank the Physics Institute at the University of Bonn and the Max Planck Institute, Heidelberg, for hospitality and acknowledges support from an Alexander von Humboldt award. This research was supported in part by NSF grant PHY-9901057. References {#references .unnumbered} ========== [99]{} R. Belli et.al., Phys. Lett.[**B480**]{},23(2000), “Search for WIMP annual modulation signature: results from DAMa/NAI-3 and DAMA/NAI-4 and the global combined analysis”, DAMA collaboration preprint INFN/AE-00/01, 1 February, 2000. R. Abusaidi et.al., Phys. Rev. Lett.[**84**]{}, 5699(2000), “Exclusion Limits on WIMP-Nucleon Cross-Section from the Cryogenic Dark Matter Search”, CDMS Collaboration preprint CWRU-P5-00/UCSB-HEP-00-01 and astro-ph/0002471. L. Baudis, A. Dietz, B. Majorovits, F. Schwamm, H. Strecker and H.V. Klapdor-Kleingrothaus, astro-ph/0008339. K. Greist, Phys. Rev. [**D38**]{}, 2357(1988); J. Ellis and R. Flores, Nucl. Phys. [**B307**]{}, 833(1988); R. Barbieri, M. Frigeni and G. Giudice, Nucl. Phys. [**B313**]{}, 725(1989); A. Bottino et.al., [**B295**]{}, 330(1992); M. Drees and M.M. Nojiri, Phys. Rev.[**D48**]{},3483(1993); V.A. Bednyakov, H.V. Klapdor-Kleingothaus and S. Kovalenko, Phys. Rev.[**D50**]{}, 7128(1994); P. Nath and R. Arnowitt, Phys. Rev. Lett. [**74**]{}, 4592(1995); R. Arnowitt and P. Nath, Phys. Rev. [**D54**]{}, 2374(1996); E. Diehl, G.L. Kane, C. Kolda, J.D. Wells, Phys.Rev.[**D52**]{}, 4223(1995); L. Bergstrom and P. Gondolo, Astrop. Phys. [**6**]{}, 263(1996); H. Baer and M. Brhlik, Phys.Rev.[**D57**]{},567(1998); J.D. Vergados, Phys. Rev. [**D83**]{}, 3597(1998); J.L. Feng, K. T. Matchev, F. Wilczek, Phys.Lett. [**B482**]{}, 388(2000); M. Brhlik and G.L. Kane, hep-ph/0005158; V.A. Bednyakov and H.V. Klapdor-Kleingrothaus, hep-ph/0011233. M. Kamionkowski and A. Kinkhabwala, Phys. Rev.[**57**]{}, 3256(1998); P. Sikvie,Phys. Lett.[**B432**]{},139(1998). R. Arnowitt and P. Nath, Phys. Rev. [**D60**]{}, 044002(1999). M. Brhlik and L. Roszkowski, Phys. Lett. [**B464**]{}, 303(1999). P. Belli, R. Bernbei, A. Bottino, F. Donato, N. Fornengo, D. Prosperi, and S. Scopel, Phys. Rev. [**D61**]{}, 023512(2000). A. Corsetti and P. Nath, Int. J. Mod. Phys. [**A15**]{}, 905(2000). E. Gates, G. Gyuk and M.S. Turner, Phys. Rev. [**D53**]{}, 4138(1996). G.R.Knapp et.al., Astron. J. [**83**]{}, 1585(1978); F.J. Kerr and D. Lynden-Bell, Mon. Not. R. Astr. Soc. [**221**]{}, 1023(1986); J.A.R. Caldwell and J.M. Coulsen, Astron. J. [**93**]{}, 1090(1987). A.H. Chamseddine, R. Arnowitt, and P. Nath, Phys. Rev. Lett. [**49**]{}, 970(1982); R. Barbieri, S. Ferrara and C. Savoy, Phys. Lett.[**B119**]{}, 343(1982); L. Hall, J. Lykken and S. Weinberg, Phys. Rev.[**D27**]{}, 2359(1983); P. Nath, R. Arnowitt and A.H. Chamseddine, Nucl. Phys. [**B227**]{}, 121(1983). K.L. Chan, U. Chattopadhyay, and P. Nath, Phys. Rev. [**D60**]{}, 063505(1999). A. Bottino, F. Donato, N. Fornengo and S. Scopel, hep-ph/9909228. J. Ellis, A. Ferstl and K.A. Olive, hep-ph/0001005. A. Corsetti and P. Nath, hep-ph/0003186. A. Corsetti and P. Nath, hep-ph/0005234. H. Cheng, Phys. Lett. [**B219**]{}, 347(1989). B. Borasoy and U-G. Meissner, Phys. Lett. [**B365**]{}, 285(1996). S.J. Dong, J.-F. Lagae, K.F. Liu, Phys. Rev.[**D54**]{}, 5496(1996). M. Fukugita, Y. Kuamshi, M. Okawa and A. Ukawa, Phys. Rev. [**D51**]{}, 5319(1995). J. Gasser and H. Leutwyler, Phys. Rep.[**87**]{}, 77(1982) J. Gasser, H. Leutwyler, and M.E. Saino, Phys. Lett. [**B253**]{}, 252(1991). P. Nath and R. Arnowitt, Phys. Lett. [**B336**]{}, 395(1994); F. Borzumati, M. Drees, and M.M. Nojiri, Phys. Rev. [**D51**]{}, 341(1995); V. Barger and C. Kao, Phys. Rev. [**D57**]{}, 3131(1998); H. Baer, M. Brhlik, D. Castano, and X. Tata, Phys. Rev. [**D58**]{}, 015007(1998). V. Mandic, A. Pierce, P. Gondolo and H. Murayama, hep-ph/0008022. A.B. Lahanas, D.V. Nanopoulos, and V.C. Spanos, hep-ph/0009065. P. Nath and R. Arnowitt, Phys. Rev. [**D56**]{}, 2820(1997). E. Accomando, R. Arnowitt and B. Dutta, and Y. Santoso, hep-ph/0001019. P. Nath and R. Arnowitt, Phys. Lett. [**B289**]{},368(1992; R. Arnowitt and P. Nath, Phys. Rev. Lett. [**69** ]{}, 725(1992); R. Roberts and L. Roszkowski, Phys. Lett.[**B309**]{}, 329(1993). M. Drees, M.M. Nojiri, D.P. Roy, Y. Yamada; Phys.Rev. [**D56**]{}, 276(1997). B. Murakami and J.D. Wells, hep-ph/0011082; T. Moroi and R.Randall, Nucl. Phys.[**B570**]{}, 455(2000). K.Greist and D. Seckel, Nucl. Phys.[**B283**]{}, 681(1987). V. Berezinsky et.al., Astrop. Phys. [**5**]{}, 333(1996). J.L. Feng, K.T. Matchev, F.Wilczek, astro-ph/0008115 L. Bergstrom, J. Edsjo and P. Gondolo, Phys. Rev. [**D58**]{}, 103519(1998). A. Bottino et.al., Phys. Rev.[**D59**]{}, 095004(1999). A.M. Green, astro-ph/0008318. J.D. Vergados, astro-ph/0010536. P. Nath, Phys. Rev. Lett. [**66**]{}, 2565(1991); Y.Kizukuri and N.Oshimo, Phys. Rev. [**D46**]{}, 3025(1992); T. Ibrahim and P.Nath, Phys. Lett. [**B418**]{}, 98(1998); M. Brhlik and G.L. Kane, Phys. Lett.[**B437**]{}, 331(1998); T. Falk and K. Olive, Phys. Lett. [**B439**]{}, 71(1998); E. Accomando, R.Arnowitt and B.Dutta, Phys. Rev. [**D61**]{}, 115003(2000). T.Falk, K.Olive and M. Srednicki, Phys. Lett. [**B354**]{}, 99(1995); T. Falk, A. Ferstl and K. Olive, Phys. Rev.[**D59**]{}, 055009(1999). U. Chattopadhyay, T. Ibrahim and P. Nath, Phys. Rev. [**D60**]{},063505(1999); T. Falk, A. Ferstl and K. Olive, hep-ph/9908311; S. Khalil and Q. Shafi, Nucl. Phys. [**B564**]{}, 19(1999); K. Freese and P. Gondolo, hep-ph/9908390; S.Y. Choi, hep-ph/9908397. A.Pilaftsis and C.E.M.Wagner, Nucl. Phys. [**B553**]{}, 3(1999). T. Ibrahim and P. Nath, hep-ph/0008237. (To appear in Phys. Rev. D). J. Ellis, T. Falk, K. A. Olive, M. Srednicki, CERN-TH-99-146, hep-ph/9905481 and the references there in. N. Spooner, “Progress on the Boulby Mine Dark Matter Experiment”, Talk at the conference “Sources and Detection of Dark Matter/Energy in the Universe”, Marina Del Rey, CA, February 23-25, 2000; I.Liubarsky et.al., Nucl. Phys. Suppl. [**87**]{}, 64(2000). H.V. Klapor-Kleingrothaus, et.al., “GENIUS, A Supersensitive Germanium Detector System for Rare Events: Proposal”, MPI-H-V26-1999, hep-ph/9910205. [^1]: : Permanent address

ArXiv

--- abstract: 'We use the formalisms of Holographic Space-time (HST) and Matrix Theory[@bfss] to investigate the claim of [@amps] that old black holes contain a firewall, [*i.e.*]{} an in-falling detector encounters highly excited states at a time much shorter than the light crossing time of the Schwarzschild radius. In both formalisms there is no dramatic change in particle physics inside the horizon until a time of order the Schwarzschild radius. The Matrix Theory formalism has been shown to give rise to an S-matrix, which coincides with effective supergravity for an infinite number of low energy amplitudes. We conclude that the firewall results from an inappropriate use of quantum effective field theory to describe fine details of localized events near a black hole horizon. In both HST and Matrix Theory, the real quantum gravity Hilbert space in a localized region contains many low energy degrees of freedom that are not captured in QU(antum) E(ffective) F(ield) T(heory) and omits many of the high energy DOF in QUEFT.' author: - | Tom Banks\ Department of Physics and SCIPP\ University of California, Santa Cruz, CA 95064\ [*and*]{}\ Department of Physics and NHETC\ Rutgers University, Piscataway, NJ 08854\ E-mail: <[email protected]>\ \ Willy Fischler\ Department of Physics and Texas Cosmology Center\ University of Texas, Austin, TX 78712\ E-mail: <[email protected]> title: '[        UTTG-10-13 TCC-007-13 RUNHETC-2013-10 SCIPP 13/06]{}.5in No Firewalls in HST or Matrix Theory' --- Introduction ============ The purpose of this paper is to sharpen our argument that quasi-local holographic models of black holes do not exhibit the firewall phenomenon that the authors of [@amps] (AMPS) claim follows from a quantum field theoretic analysis of Hawking radiation in a regime in which quantum field theory can be trusted. While we have some sympathy with the arguments of [@nomura][@verlinde][@haydenharlow], that the thought experiment envisaged by AMPS cannot, for a variety of reasons, be carried out even in principle, our own view is that the problem lies primarily with the use of QFT to analyze fine grained aspects of the quantum information involved in black hole evaporation. In [@holofirewall] we presented these arguments, but we failed to emphasize sufficiently what the important points were, because our understanding had evolved in the course of writing the paper. We attempt to repair this here. In addition, we present a completely different set of arguments based on the Matrix Theory models of black holes[@bfssbh] in highly supersymmetric compactifications of M-theory. The details are somewhat different, but the conclusions are similar. There is no firewall, though the model has quantum states that share all of the familiar properties of black holes, and is manifestly unitary. There is also a large body of evidence that Matrix Theory does reproduce the correct scattering matrix for low energy effective supergravity. The problem of accounting for black hole entropy is one in which there is an evident breakdown of local quantum field theory in the low energy regime. In the absence of a black hole, local quantum field theory can account for at most $o(A^{3/4})$ (in 4 dimensions) of the entropy allowed by the covariant entropy bound in a causal diamond whose holographic screen has area $A$. In the presence of a black hole, field theory instead over-counts the entropy, and finds an infinite entropy per unit area, despite the fact that, in the vicinity of the horizon of a large black hole, the local space-time geometry of the hole is identical to that of flat space. Indeed, the same infinity is found for the entropy encountered by a highly accelerated Rindler observer in flat space. The infinite entropy per unit area comes from modes of arbitrarily short wavelength in regions that have a small space-like separation from the horizon. These modes have low energy from the point of view of an accelerated observer. However, if we consider a small causal diamond surrounding a portion of the horizon (Rindler or Schwarzschild) and insist that the state of QFT in that diamond be such that gravitational back reaction is negligible, then the short wavelength modes must be frozen into the Minkowski ground state. In that state, there is infinite entanglement entropy per unit area of the holographic screen, between DOF localized in the diamond, and DOF an infinitesimal space-like distance outside it. There is an obvious paradox here. For the Rindler or Schwarzschild observer, the entropy refers to real excited degrees of freedom, while entanglement entropy is a property of a pure state. Even in QUEFT, this paradox is resolved by noting that the Rindler and Schwarzschild observers use a different Hamiltonian from the geodesic observer in a locally Minkowski space. The Rindler case is particularly illuminating because observers with different acceleration see different temperatures. If we make the assumption that observers sharing the same causal diamond must see the same quantum state up to a unitary transformation, we are prodded in the direction of the conclusion that each accelerated trajectory must have its own Hamiltonian, since they each see a different von Neumann entropy. This is the starting point for the axioms of HST. We will not repeat the HST description of accelerated observers here, but refer the reader to [@holounruh]. However, we do pause to record the message of that paper: [*A causal diamond of finite area $\sim N^2$ in a 4-dimensional space-time has $o(N^2)$ (graded)-commuting copies of a super-algebra with a finite dimensional unitary representation. The fermionic generators of that algebra are labeled $\psi_i^A (P)$, where the explicit indices are those of an $N \times N+1$ matrix. At most $o(N^{3/2})$ of these DOF have a particle interpretation, and it is only to this subset that the rules of QUEFT apply. We call the rest, [*horizon DOF*]{}. By contrast, a Jacobsonian[@ted] TH(ermodynamic) E(ffective) F(ield) T(heory) will always encode the coarse grained hydrodynamics of any Lorentzian space-time in HST[^1]. For a geodesic observer in a causal diamond, the bulk of the degrees of freedom have a coupling of order $1/N$ to the particles. They give rise to particle interactions, but have no long term entanglement with properly prepared incoming particle states. Accelerated observers have different Hamiltonians, and experience more entanglement: the redshift of particle energies compared to those of horizon DOF couples the particles more strongly and leads, in the large $N$ limit, to the Unruh temperature.*]{} The common practice in studies of black holes using QUEFT, has been to cut off the infinity of low energy QUEFT DOF in the accelerated observers frame by putting a “brick wall" a finite space-like distance away from the horizon. Then, the remaining finite entropy per unit area is divided between a stretched horizon whose dynamics is admitted to be unknown, and a “zone" where standard QUEFT descriptions are valid. This is the description used by AMPS. We do not really have a quarrel with this prescription, if it is viewed as a relatively coarse grained model of the black hole. What we claim is that such a model cannot pretend to account for the quantum dynamics of the tiny fraction ($< o(N^{3/2})$) of the DOF, which are all that is necessary to describe particle physics inside the horizon of the black hole. Black Holes in HST ================== In HST, space-time physics is described in terms of an infinite number of quantum systems, each of which encodes the physics as seen along a particular time-like trajectory, in a proper time dependent Hamiltonian. Relations between density matrices for shared information, combined with the holographic connection between Hilbert space dimension and area, enable us to extract the causal structure and conformal factor of a Lorentzian metric from the quantum mechanics. When we first approached this problem, we believed that the main feature of HST, which avoided the firewall, had to do with the fact that different observers had different Hamiltonians. J. Polchinski and D. Harlow convinced us that the problem could be recast entirely in the Hilbert space of the observer that is called Alice in most of the firewall literature. Our resolution of the problem in [@holofirewall] in fact relied on different features of HST, but enough of our early thinking survived in the final draft of that paper, that it has obscured the issue. What follows is the description of a decaying black hole in HST, purely from the point of view of the detector $A$, formerly known as Alice. In fact $A$ belongs to a one parameter family of detectors $A(T)$, parametrized by the amount of proper time, after the formation of the horizon[^2], before the detector falls through the horizon of the partially evaporated black hole. To keep everything finite, we introduce a large time $- N$, prior to horizon formation, and designate the time of horizon formation by $0$. All of the relevant events occur within a causal diamond of area $N^2$. The following occurs in real-time: From the point of view of the A detector, the black hole evaporates for a long time before the detector crosses its “horizon". In HST we understand that the black hole space-time is an approximate, hydrodynamic description of some of the quantum degrees of freedom in the S-matrix theory of asymptotically flat space-time. It is appropriate for a description of those causal diamonds where the detector whose Hamiltonian we are describing receives signals from the black hole horizon. These signals cannot come from particle degrees of freedom inside the horizon. That is the lesson encoded in the classical geometry. In HST, we view those signals as originating from the $o(N^2)$ DOF on the black hole horizon, but, to the extent that the classical geometry tells us that particles inside the horizon experience ordinary particle physics, we consider the state of those $< o(N^{3/2})$ particle DOF, to be unentangled with the bulk of the horizon states. Note that the Hamiltonian of the detector A does not interact with these interior particle DOF until the detector’s trajectory crosses the horizon. The $o(N^2) - o(N^{3/2})$ DOF with which it [*does*]{} interact include both those that are associated with the stretched horizon, and those in the “zone", in the AMPS description. Cutting off the infinite entropy per unit area in QUEFT leaves a huge hole through which an entropy sub-leading in the area can fit. QUEFT describes the single vacuum state in a region in a single causal diamond in terms of a huge Hilbert space of “particle states" viewed by an accelerated observer. We know that in Minkowski space it is incorrect to view an area’s worth of entropy in this Hilbert space as describing actual, physically accessible states of the system. Yet this is precisely what AMPS do. Their argument is based on the idea that one can understand the restrictions on the utility of QUEFT in terms of a simple UV cutoff on local wavelengths. In fact, for a geodesic observer in a low curvature region of space time, this claim is simply false. Instead, as we have argued above, any entropy in the causal diamond that is proportional to the area, must consist of states which have very low energy according to the Hamiltonian of the geodesic observer, [*and which therefore cannot manifest as particles in the bulk of the diamond*]{}. This is true for diamonds inside the horizon, outside the horizon, or straddling the horizon. The utility of the picture of black hole states as [*partially described by QUEFT with a brick wall cutoff*]{}, is due to the ease with which one can understand Hawking radiation in this picture, and was of course the description used by Hawking in his original argument. This often leads to the erroneous claim that if we give up this picture of the Hilbert space, we will no longer understand the thermal nature of black holes and the calculation of the Hawking temperature. In fact, the thermal nature of Hawking radiation implies that the coarse grained details of Hawking’s calculation will be reproduced by [*any*]{} model of a black hole that exhibits it as an ergodic system[^3], with the correct entropy/energy/size relations. In fact, Hawking’s derivation of Hawking radiation is not valid even for strongly coupled field theories. The reason that we [*know*]{} that the thermodynamic picture of black holes is correct is that the Hartle-Hawking state of a black hole is a thermal state, no matter what the field theory is. The entropy and temperature of the black hole are encoded in its classical geometry. We think that the deepest understanding of why this is so comes from Jacobson’s observation that, apart from the cosmological constant, Einstein’s equations are the hydrodynamics of space-time, assuming that space-time emerges from a quantum system obeying the Covariant Entropy Conjecture that the entropy of the Hilbert space of quantum gravity is one quarter the area in Planck units of the holographic screens of infinitesimal causal diamonds around every point[^4]. The Jacobsonian point of view also sheds light on the following vexing question: On the one hand, a finite entropy black hole can be formed by scattering a finite number of particles in Minkowski space, and on the other hand, it seems to have a distinct space-time metric. The Jacobsonian interpretation of the black hole metric is as a hydrodynamic approximation to the behavior of a large number of DOF in the Minkowski system, [*whose behavior is not well modeled as particle physics.*]{}. Our goal in this section is to provide a statistical mechanics, which is in agreement with the thermodynamic predictions of this classical metric. For us, these include the behavior of in-falling particle systems before they hit the classical singularity. This was done in [@holofirewall], so we just summarize it here. We describe the Hamiltonian of the detector $A(T)$, which follows a trajectory that encounters the instantaneous black hole horizon a proper time $T$ after the original black hole horizon forms. The black hole metric is approximated by a time dependent sequence of Schwarzschild metrics with a radius parameter $R(t)$ that follows Hawking’s evaporation law. At any given time,$t$, we split off $\pi (R(t) M_P)^2$ black hole degrees of freedom from the vastly larger number $N^2$, which describe $A(T)$’s full causal diamond. For $t < T$, the horizon crossing time of $A(T)$ the bulk of these black hole DOF are given a Hamiltonian with a Planck scale time dependence, which is a sum of traces of the matrices $(1 - \Pi) \psi \psi^{\dagger}(1 - \Pi )$. $\Pi$ is a projection matrix on a $K(t) \times K(t)$ subspace with $K \leq (R(t) M_P)^{3/2}$. The rest of the $o(N^2)$ DOF, which are not associated with the black hole, are given a Hamiltonian, and an initial condition in the remote past that are appropriate for describing particle physics in Minkowski space. We do not have a complete description of this but a class of Hamiltonians that give the right qualitative physics was described in [@holounruh]. There are additonal constraints necessary to guarantee that in the large $N$ limit, the S-matrix becomes super-Poincare invariant, which we have not yet implemented. Our model does not try to describe the formation of the black hole from some particular incoming particle state. Although AMPS assume a black hole formed in this way, they assume nothing about the incoming state besides its purity. Note that, in addition to the rapid variation of the Hamiltonian mixing up the $ [\sqrt{\pi} R(t) M_P - K(t)]^2 $ states there is a much smaller time dependence of the Hamilton, coming from the time dependence of $R(t)$. The total Hamiltonian is $$H_{Mink} + H_{Hor\ (t)} + H_{K(t)} ,$$ where the second two terms act on the black hole DOF. As $R(t)$ varies, $H_{Hor\ (t)}$ and $H_{K(t)}$ act on fewer DOF, and we add those to the Hamiltonian $H_{Mink}$. This time dependence is an addition to the rapid time dependence of $H_{Hor\ (t)}$, which acts on $o( R(t) - K(t))^2$ DOF. The Hamiltonian $H_{K(t)}$ describes the evolution of DOF that will be experienced as particles by the detector $A(T)$. As in all HST models, the time dependent Hamiltonian of the detector splits into two pieces $$H(t) = H_{in} (t) + H_{out} (t) .$$ The Hamiltonian $H_{K(t)}$ is part of $H_{out} (t)$ until $t = T$. For $t > T$ it is included in $H_{in} (t)$. This is not really a discontinuous transition. The DOF included in $H_{K(t)}$ are, at early times just particles that have been sent in from past infinity and are initially causally separated from the detector $A(T)$. If we consider the large causal diamond of a geodesic observer that is causally connected to the particles at early times, but never falls into the black hole, then, at early times, these particles are described by the geodesic Hamiltonian and initial conditions in that diamond[@holounruh]. For simplicity, we assume that they don’t undergo any scattering before they enter the causal diamond where $A(T)$ will encounter them. The particles in $H_{K(t)}$ are those with which the detector $A(T)$ can interact, between the time it crosses the instantaneous horizon and the time its trajectory encounters the black hole singularity. Thus, for times $T < \tau \ll T + R(t)$, the state in the Hilbert space representing the entire black hole is approximately a tensor product of a state acted on by $H_{K(t)}$ and one acted on by $H_{Hor\ (t)}$. In terms of the matrix DOF $\psi_i^A$, these two Hamiltonians are functions of $\Pi \psi \psi^{\dagger} \Pi$ and $(1 - \Pi ) \psi \psi^{\dagger} (1 - \Pi )$. Interactions between the two sets of DOF via off block diagonal matrix elements, are suppressed by the large number $1/N$, as long as $\tau \ll R(t)$. It’s important to stress why we’re making these claims, especially the last one. Our aim is to construct a quantum system, whose behavior mimics the classical space-time picture of the interior of the black hole. In that picture, the particles described by the DOF in $H_{K(t)}$, behave, for a time of order $R(t)$, approximately as they would in flat space. At the time $T$ (within a tolerance $\ll R(T)$) the DOF in $H_{K(t)} $ are incorporated into $H_{in} (t)$ of the detector $A(T)$. The Hilbert space of that detector is vast, with entropy of order $N^2$. However, our model of this detector’s behavior is that it’s own components, and all the particles with which it interacts [*after crossing the instantaneous horizon*]{}, are described by the Hamiltonian $H_{K(t)}$. In addition to the slow time dependence induced by the change of $K(t)$and $R(t)$, this Hamiltonian has a time dependence which becomes extremely rapid as $t$ approaches $T + R(t)$. These time dependent terms are traces of products of the full $R(t)M_P \times R(t)M_P$ matrix $\psi \psi^{\dagger}$ . They are very small at $t = T$ and become competitive with the ordinary particle physics contributions in a time of order $R(t)$. Their effect is to mix up the particle DOF with the horizon, so that the distinction between particles and horizon is no longer meaningful. At a time of order $T + R(t)$ the detector $A(T)$ has “hit the singularity". The bulk of the Hilbert space of $A(T)$ knows nothing about this catastrophe, either before or after it happens. The DOF in this Hilbert space interact with the horizon DOF in the matrix $(1 - P) \psi \psi^{\dagger} (1 - P)$. If the infinitely intricate measurements envisioned by AMPS could actually be carried out one would find that the assumption in Page’s discussion[@Page] of information extraction from a black hole was subtly wrong at the time $T$, when $T$ is of order the Page time. That is, the black hole is not in a generic state of a Hilbert space of entropy $\pi (R(T)M_P)^2$, because a tiny tensor factor whose entropy is of order $K^2 (T) < (R(T) M_P)^{3/2}$ is not entangled with the rest of the space. If we synchronize the external clock to the proper time of the in-falling trajectory of $A(T)$, then this factor becomes entangled with the rest of the black hole Hilbert space at a time of order $T + R(T)$. Clearly, neither the thermodynamic properties of the black hole, nor the statement that the evaporation process is unitary are violated by this model. One of the assumptions of Page’s argument is modified by an amount that is thermodynamically negligible. The centerpiece of the AMPS argument, the representation of the Minkowski vacuum state in a local region of space-time as an entangled state in a certain factorized basis of the field theory Hilbert space, simply does not appear in the HST formalism. Our model clearly treats the black hole as a thermodynamic object with the correct energy-entropy-size relations. So it’s simply untrue that one needs this entangled picture to obtain Hawking radiation. We believe that the only valid criticism of our model is that we have not yet shown that it really reproduces the results of field theory in conventional situations where no black holes are involved. We argued that this was the case, to the best of our current ability, in [@holounruh], but we are aware that only a complete calculation of some scattering amplitude will really make the case. In the next section, we will argue that the Matrix Theory description of 11 dimensional Schwarzschild black holes, gives a picture of black hole evaporation consistent with the one we proposed in HST. There is no firewall, no description of the Minkowski vacuum as an entangled state, a manifestly unitary S-matrix for particle scattering, and manifest super-Galilean invariance. In addition, it has been shown that an infinite number of scattering amplitudes in this model coincide with those given by a super-Poincare invariant QUEFT - 11 dimensional supergravity. In principle, the above criticism might be applied to our claims about reproducing the properties of Hawking radiation. However, we demonstrated in [@holounruh] that the Minkowski Hamiltonian, which acts on $o([N - R(t)]^2)$ DOF, acts, as $N \rightarrow\infty$, like the kinetic term of a collection of massless particles, on a tensor factor of entropy $\leq N^{3/2}$ of its Hilbert space. We also argued that in the large $N$ limit, pure states of these particles remain pure and that the effect of interaction with the bulk of the $o(N^2)$ DOF could be encoded in particle interactions that were localized in space-time. Thus, our model does describe black hole evaporation as a sequence of quasi-equilibrium states of the black hole, interacting with a gas of relativistic particles of (potentially) much higher entropy. The dynamics is explicitly rotation invariant and there are emergent super-Poincare generators, which act on particle states in the large $N$ limit. This is enough to establish the thermal nature of the spectrum of evaporated particles, even though we have not established that the S-matrix of those particles is Poincare invariant. Thus, we certainly cannot claim that our model will reproduce the gray body factors that arise in the field theory treatment of Hawking radiation, but the thermal nature of the spectrum and the correct temperature are guaranteed. In HST itself, we still have to discuss the consistency conditions between detectors $A(T)$ with different values of $T$. In [@holofirewall] we showed that consistency between detectors with $T \sim R_S$ and $T \gg R_S$ implied that the late falling detector had to encounter a singularity[^5] in a time of order $R_S$ after it crosses the horizon. Consistency of the in-falling detector’s description with that of a supported detector implies that the description of the black hole from the supported detector’s point of view must include a tensor factor of entropy $ \sim (R(t)M_P)^{3/2}$, which is approximately unentangled with the horizon[^6]. Since $(RM_P)^{3/2} \ll (RM_P)^2 $, this does not affect the thermodynamics of black holes until they are of Planck size. Since $R(t)$ goes to zero eventually, there is no problem with unitarity of the S-matrix either. Of course, once the black hole is Planck size, the approximate descriptions in this paper lose their validity. Indeed, even in Minkowski space, the clean separation between particle and horizon DOF is impossible in small causal diamonds. Indeed, as shown in [@holounruh], this feature of the HST description is responsible for reactions that change the number of particles, and their momenta. Particles are emergent phenomena in HST, strictly speaking defined only in the limit of infinite causal diamonds. The validity of QUEFT, with its implicit assumption of infinite numbers of particle states, is even more restricted. Black Holes in Matrix Theory ============================ We will restrict attention to the Matrix Theory for String/M-theory in 11 non-compact dimensions. This is the quantum mechanics of the zero modes of maximally supersymmetric $SU(N)$ Yang-Mills theory. Similar results would be obtained for compactification of the theory on tori of dimension $1-3$. The Lorentz invariant limit is achieved by taking $N\rightarrow\infty$ and computing the S-matrix for states whose energy scales like $1/N$. However, in [@bfssbh], the authors argued that one could understand the qualitative dynamics of black holes of entropy $S$ in the model with $N\sim S \gg 1$. The construction of black hole states begins with a classical solution of the matrix equations $X_{cl} (t)$, whose variation away from the origin of transverse coordinates was bounded by a distance in Planck units of order one (in the sense of large $N$ counting). The matrices in the solution have rank $N$. For large $N$, we can think about them using the correspondence [@DHN] with light front membrane theory. The background defines a (fuzzy) toroidal membrane, whose volume is parametrized by two angles $p,q$. The matrices are functions on the phase space $[p,q]$ with commutator given, in the large $N$ approximation, by Poisson brackets. There are many such solutions, but in $11$ dimensions, this multiplicity gives rise to a sub-leading correction to black hole entropy. Now write $X = X_{cl} + \sum_i x_i M_i$, where $$M_i = \sum_{k,l} e^{- N[ (p - p_i - 2\pi k)^2 + (q - q_i - 2\pi l)^2]}.$$ The commutators of these matrices satisfy $$|[M_i , M_j]| \leq e^{ - \frac{1}{2}[ (p_i - p_j)^2 + (q_i - q_j)^2 ]} .$$ Taking a distribution of points separated by distances of order $\frac{1}{\sqrt{N}}$ (there are $o(N)$ such points on the torus) , we can make these commutators as small as we like. Thus, there’s a basis in which all the matrices $M_i$ are simultaneously block diagonal. The traces of these matrices are $o(1)$. The commutator between the classical background $X_{cl} (p,q)$ and the fluctuations $ x_i$ gives rise to a harmonic potential binding the $x_i$ to the background. The terms bilinear in different $x_i$ are of the same order as the commutator $[M_i , M_j ]$ and we drop them. The quadratic potential is $$\sum_i x_i^2 N \int\ dp\ dq\ [p^2 + q^2] (\nabla X_{cl})^2 e^{ - N (p^2 + q^2)}.$$ The factor of $N$ in front of the integral is the combination of a $1/N$ in the translation of traces of matrix commutators into integrals of Poisson brackets over the membrane, and two factors of $N$ coming from converting derivatives of $M_i$ into factors of $p$ or $q$. For a smooth classical membrane configuration, the gradient of $X_{cl}$ is $N$ independent for large $N$. The harmonic potential thus has an overall coefficient $1/N$. This is, it will turn out, negligible compared to other contributions to the energy. The latter come from integrating out off diagonal matrices between the different $x_i$ terms. If the $x_i$ velocities are small, which is verified self consistently, the effective Hamiltonian is $$H = \sum_i {\bf p_i}^2 + A G_N \sum_{i,j} \frac{\bf (p_i - p_j)^4}{|{\bf x_i - X_j}|^7} .$$ The coefficient $A$ is of order $1$. For a bound system with this Hamiltonian, and large $N$, the mean interparticle distance $R_S$, the energy per particle, and the total light front energy (which gives us the mass) may be calculated crudely by using the uncertainty principle and the virial theorem. The result is $$B G_N^{-1} R_S^9 = N,$$ $$E_{per\ particle} \sim R_S^{-2},$$ $$M \sim G_N^{-\frac{1}{9}} N^{\frac{8}{9}} .$$ Here $R_S$ is the average separation, which is also the size of the bound state. These are the expected relations for an $11$ dimensional Schwarzschild black hole, with the individual $D0-$branes having the kinematics expected for Hawking particles boosted to the light front frame, and $R_S$ the Schwarzschild radius, if $N$ is indeed the entropy of the black hole. The fact that $N$ is indeed the entropy follows from the fact that the $D0$ branes are [*tethered*]{} to different positions on the classical membrane, so that they are in fact distinguishable particles, obeying Boltzmann statistics. If we plug $x_i \sim R_S$ into the harmonic potential, we find a negligible correction to the total energy. In the second paper in[@bfssbh], we showed also that the correct Newtonian interaction between black holes is obtained, if we are careful to note that we are calculating energies averaged over the longitudinal circle of Discrete Light Cone Quantization. In the third paper we estimated the rate of Hawking evaporation, and found agreement with the expectations for a thermal system with the indicated energy and entropy. The mechanism of Hawking radiation was “snapping of the tethers": a quantum fluctuation, which momentarily sets to zero the piece of the classical configuration that provides the harmonic binding for a particular $D0-$brane coordinate $x_i$. That particle then flies out to infinity along the flat direction in the matrix potential. The black hole then re-equilibrates with one constituent fewer. In the second paper of [@bfssbh] it was pointed out that the analysis of the dynamics of the non-compact dimensions gave analogous results for black holes in all dimensions. However, at that time Matrix Theory technology required one to use a more and more complicated field theory to compactify the theory on more and more dimensions. For a $5$ dimensional toroidal compactification one was forced to go beyond field theory, and for $6$ and more dimensions the Matrix Theory proposal failed. Results could be established firmly only for compactifications on tori of dimensions $1 - 3$, and in the last of these cases the internal field theory contributed a finite fraction of the black hole entropy. Recently TB and Kehayias[@tbkehayias] have suggested an alternate definition of Matrix Theory, which is a simple quantum mechanics (not a field theory) for every compactification. It would be interesting to return to the black hole problem using this technology. We conjecture that a unified qualitative picture of Schwarzschild black hole dynamics might result. Be that as it may, the purpose of the present paper is to establish the absence of firewalls for 11 dimensional Schwarzschild black holes. Recall that Matrix Theory defines a scattering matrix for asymptotic states along the $U(k_1) \otimes \ldots \otimes U(k_n)$ flat directions of the Matrix Theory potential. The $SU(k_i)$ degrees of freedom are frozen into their unique BPS bound state, and the scattering states are manifestly those of eleven dimensional supergravitons. If we take $k_i$ to infinity, at fixed ratios $\frac{k_i}{k_j}$ then the asymptotic states support an action of the $SO(1,10)$ super-Poincare group. The S-matrix is manifestly invariant under the super-Galilean sub-group of this group, and the existence of the S-matrix in the limit is equivalent to longitudinal boost invariance. It is hard to see what kind of instability could make the manifestly unitary S-matrix fail to exist in this limit, because from the point of view of the quantum mechanics, it is a [*low*]{} energy limit. In particular, emission of states with longitudinal momentum that does not scale to infinity is forbidden by energy conservation. We also know that the S-matrix obeys an infinite number of non-renormalization theorems[@dineetal], which imply that an infinite number of terms in the low energy expansion of the hypothetical limit, actually coincide with those expected from the low energy expansion of a super-Poincare invariant effective Lagrangian. This proves that the limiting S-matrix is not the unit matrix, and strongly suggests the existence and super-Poincare invariance of all matrix elements. Finally, Matrix Theory provides a definition of finite time transition amplitudes, for processes that take place over a finite range of transverse distance. These amplitudes manifestly approach the corresponding S-matrix elements as the time and transverse distance go to infinity. Using these, we can model the experience of an apparatus falling into a black hole. The apparatus is modeled by a $K \times K$ block of the Matrix Theory variables, with $1 \ll K \ll N$. Initially, we take the transverse separation between the $K \times K$ block and the $N \times N$ block, which represents the black hole, to be very large, and set the center of mass of the $K \times K$ block moving slowly towards the transverse position of the hole. Consider an initial condition for the $SU(K)$ variables which consists of two groups $K_{1,2}$ of supergravitons coming in from a large distance. For comparison, we take $K_1 = K_2$. The first group collides at a time long before the c.m. of the block approaches the position of the black hole. What this means is that the incoming coefficients of a block diagonal matrix of block sizes $p_1 \ldots p_m$, with $\sum p_i = K_1$ become close enough so that it no longer takes a huge energy to excite the off diagonal matrices. Non-Abelian dynamics becomes important and we find finite amplitudes for going off in flat directions $q_1 \ldots q_n$, with $\sum q_i = K_1 $. We compute the amplitude from the time the initial separation is $L \ll R_S$ until a time after collision when final separations are of order $L$. During this entire period of time, the separation between the c.m. of the $K_1 \times K_1$ block and the black holes is $\gg R_S$. For the $K_2$ block, we instead time the incoming particles so that they begin to interact strongly with each other when the c.m. of the block is within $R_S$ of the center of the black hole. Since $L \ll R_S$, the typical distance between particles in the block is much less than their distance from any of the $D0-$brane constituents of the black hole. Thus, the interactions with the black hole can be considered a small perturbation of the particle interactions in flat space, over times where the transverse separation is $\ll R_S$. Over longer time scales, this is no longer true. For distance scales of order $R_S$ from the $K_2$ block, a particle we originally considered part of the $K_2$ block will suffer multiple scatterings with black hole constituents, whose transverse separation from it are smaller than or of the same order as those in the $K_2$ block. Since the number of black hole constituents is large compared to the number of particles in the original event (since $N \gg K_2$) it is plausible that the particle will come into equilibrium with the black hole constituents. That is, interactions with the constituents will tend to break it up into its individual $D0-$branes and these will equilibrate and become indistinguishable, in a coarse grained way, from constituent $D0-$branes that were in the black hole before the $K_2$ block approached it. There is probably a theorem to be proven here in the $N\rightarrow\infty$ limit, since in that limit, the constituents of the $K_2$ block can never escape from the black hole once they have come within a distance of order $R_S$ of it. We claim that this is evidence for the absence of a firewall in Matrix Theory. Particle physics over time scales smaller than the time to traverse a transverse distance $R_S$ is affected only perturbatively by the question of whether it takes place inside or very far from the Schwarzschild radius of a black hole. Note also that [*nowhere in the Matrix Theory model of a black hole does there exist any analog of the high energy particles that are supposed to constitute the firewall.*]{} Black hole constituents in Matrix Theory have the kinematic properties of typical thermal Hawking particles. It is probably significant that it’s a general property of physics in a light front frame, that the particle theory vacuum is trivial. This means that the vacuum entanglement that is claimed to be a crucial feature of Hawking radiation by AMPS cannot be a feature of physics formulated on the light front, as Matrix Theory is. The reason that this demonstration of the absence of firewalls adds to the credibility of our HST argument is that there is much more evidence that Matrix Theory is a systematic approximation to a super-Poincare invariant S-matrix theory of particles, with a low energy effective field theory expansion. Furthermore, there is an established construction of states with the properties of black holes, in a system with a time independent Hamiltonian. The evidence that HST leads to super-Poincare invariant scattering was presented in [@holounruh], and is much less extensive. Conclusions =========== Both of our models of quantum gravity contain “low energy" DOF, which are not captured by QUEFT, and are crucial to the description of the local dynamics of black holes. In neither of them is there any apparent hint of the picture of field theory with a stretched horizon cutoff, which pervades much of the literature on the black hole information problem, including the paper of AMPS. It is not our intention here to claim that the stretched horizon picture is completely wrong or useless. Rather, our position is that the question of whether an in-falling observer encounters large deviations from flat space physics on a time scale much shorter than the classical in-fall time to the singularity, involves only a tiny fraction of the DOF of the black hole. QUEFT with a stretched horizon cutoff is certainly a grossly thermodynamic description of the system, and is simply insensitive to these thermodynamically negligible DOF. To make the AMPS argument one must assume that the stretched horizon QUEFT description is an accurate accounting of the dynamics at the level of single bits. .3in [**Acknowledgments** ]{} T.B. would like to acknowledge conversations with J. Polchinski, L.Susskind and D.Harlow about firewalls. He would also like to thank the organizers of the firewall workshop at CERN in March 2013, for the invitation to speak at that conference, and the other participants for stimulating discussions. The work of T.B. was supported in part by the Department of Energy. The work of W.F. was supported in part by the TCC and by the NSF under Grant PHY-0969020 [99]{} A. Almheiri, D. Marolf, J. Polchinski and J. Sully, Black Holes: Complementarity or Firewalls?, 1207.3123; A. Almheiri, D. Marolf, J. Polchinski, D. Stanford and J. Sully, “An Apologia for Firewalls,” arXiv:1304.6483 \[hep-th\]. D. Harlow and P. Hayden, “Quantum Computation vs. Firewalls,” arXiv:1301.4504 \[hep-th\]. Y. Nomura, J. Varela and S. J. Weinberg, “Low Energy Description of Quantum Gravity and Complementarity,” arXiv:1304.0448 \[hep-th\]. E. Verlinde and H. Verlinde, “Black Hole Entanglement and Quantum Error Correction,” arXiv:1211.6913 \[hep-th\]. T. Jacobson, “Thermodynamics of space-time: The Einstein equation of state,” Phys. Rev. Lett.  [**75**]{}, 1260 (1995) \[gr-qc/9504004\]. T. Banks and W. Fischler, “Holographic Space-Time Does Not Predict Firewalls,” arXiv:1208.4757 \[hep-th\]. T. Banks and J. Kehayias, “Fuzzy Geometry via the Spinor Bundle, with Applications to Holographic Space-time and Matrix Theory,” Phys. Rev. D [**84**]{}, 086008 (2011) \[arXiv:1106.1179 \[hep-th\]\]. T. Banks, W. Fischler, I. R. Klebanov and L. Susskind, “Schwarzschild black holes from matrix theory,” Phys. Rev. Lett.  [**80**]{}, 226 (1998) \[hep-th/9709091\]. ; T. Banks, W. Fischler, I. R. Klebanov and L. Susskind, “Schwarzschild black holes in matrix theory. 2.,” JHEP [**9801**]{}, 008 (1998) \[hep-th/9711005\]. ; T. Banks, W. Fischler and I. R. Klebanov, “Evaporation of Schwarzschild black holes in matrix theory,” Phys. Lett. B [**423**]{}, 54 (1998) \[hep-th/9712236\]. B. de Wit, J. Hoppe and H. Nicolai, “On the Quantum Mechanics of Supermembranes,” Nucl. Phys. B [**305**]{}, 545 (1988). D. N. Page, “Information in black hole radiation," Phys. Rev. Lett. 71, 3743 (1993) \[hep-th/9306083\]. T. Banks, W. Fischler, S. H. Shenker and L. Susskind, “M theory as a matrix model: A Conjecture,” Phys. Rev. D [**55**]{}, 5112 (1997) \[hep-th/9610043\]. T. Banks, W. Fischler, [*Holographic Space-time, the Unruh Effect and the S-matrix*]{}, Journal of Emergent Physics [*to Emerge*]{}. M. Dine, R. Echols and J. P. Gray, “Tree level supergravity and the matrix model,” Nucl. Phys. B [**564**]{}, 225 (2000) \[hep-th/9810021\]. ; M. Dine, R. Echols and J. P. Gray, “Renormalization of higher derivative operators in the matrix model,” Phys. Lett. B [**444**]{}, 103 (1998) \[hep-th/9805007\]. ; K. Becker and M. Becker, “On graviton scattering amplitudes in M theory,” Phys. Rev. D [**57**]{}, 6464 (1998) \[hep-th/9712238\]. ; J. A. Harvey, “Spin dependence of D0-brane interactions,” Nucl. Phys. Proc. Suppl.  [**68**]{}, 113 (1998) \[hep-th/9706039\]. [^1]: Jacobson showed that, apart from the cosmological constant, Einstein’s equations follow from the first law of thermodynamics, applied to a system with an effective space-time description, and such that the entropy seen by a maximally accelerated Rindler observer, near each point, varies along the observer’s trajectory like the area transverse to a bundle of initially parallel trajectories. HST satisfies Jacobson’s criteria, and fixes the c.c. by a boundary condition relating the behavior of area vs. proper time in the limit of a diamond with large proper time. [^2]: In HST, the definition of space-like slices has to do with a convention of synchronization of clocks for space-like separated portions of different time-like trajectories. [^3]: We use the word ergodic very loosely here, not in its technical mathematical sense. There are a variety of hypotheses about the nature of a quantum system, which lead to certain aspects of thermal behavior. Any one of them will do. [^4]: Here infinitesimal means much larger than Planck scale but much smaller than the local radius of curvature of space-time. [^5]: Quantum translation: the time dependent Hamiltonian must mix the particle DOF inside $A(T)$’s horizon with the horizon DOF in a time of order $R_S$. [^6]: We thank D. Harlow for repeatedly emphasizing this point to us.

ArXiv

--- abstract: | We show here that the Hamiltonian for an electronic system may be written exactly in terms of fluctuation operators that transition constituent fragments between internally correlated states, accounting rigorously for inter-fragment electron exchange and charge transfer. Familiar electronic structure approaches can be applied to the renormalized Hamiltonian. For efficiency, the basis for each fragment can be truncated, removing high-energy local arrangements of electrons from consideration, and effectively defining collective coordinates for the fragments. For a large number of problems [ (especially for non-covalently interacting fragments), ]{} this has the potential to fold the majority of electron correlation into the effective Hamiltonian, and it should provide a robust approach to incorporating difficult electronic structure problems into large systems. The number of terms in the exactly transformed Hamiltonian formally scales quartically with system size, but this can be reduced to quadratic in the mesoscopic regime, to within an arbitrary error tolerance. Finally, all but a linear-scaling number of these terms may be efficiently decomposed in terms of electrostatic interactions between a linear-scaling number of pre-computed transition densities. In a companion article, this formalism is applied to an excitonic variant of coupled-cluster theory.\ \ **Keywords:** Fragment Methods; Electron Correlation; Excitons; Renormalization; Effective Hamiltonian; Range Separation author: - | Anthony D. Dutoi$^{*}$, Yuhong Liu\ [*Department of Chemistry, University of the Pacific, Stockton, California 95211, USA*]{}\ $^*$`[email protected]` title: Systematically Improvable Excitonic Hamiltonians for Electronic Structure Theory --- Introduction ============ A persistent pursuit of the electronic structure community is the elimination of unnatural barriers to efficient, detailed simulation. Formally, the amount of information needed to fully specify the quantum state of an electronic system grows exponentially with the system size. But this is juxtaposed against the common intuition that makes possible the chemical language of atoms, molecules, and functional groups, which is that these entities do not forgo their individual properties upon interaction with others, though they may be heavily perturbed. The most straightforward approach to eliminating the unnecessary bulk of an exponentially large state space is to divide a super-system into sub-systems, or fragments [ [@Gordon:2011:FMOReview; @Richard:2012:FragUnifiedView; @Collins:2015:EnergyBasedFragMethods; @Raghavachari:2015:FragmentReview; @Huang:2008:ElecStructSolids; @Wesolowski:2015:EmbeddingReview] ]{}. [ One shortcoming of presently available fragment-based electronic structure methods is that they either do not capture inter-fragment electron correlation at all or this is treated at the level of individual electrons. ]{} For inter-fragment correlation, a reliance on integrals describing interactions between individual electrons will render any high-order correlation scheme inefficient because the structures of accompanying local relaxations are essentially recomputed for each separate interaction between the electrostatically efficacious single-electron fluctuations. [ High-order (beyond perturbative) treatments of inter-fragment electron correlation are desirable, however, in order to capture many-body interactions between fragments [@Cui:2006:InductionFourthOrder; @Lao:2016:LargeSystemMBE; @Podeszwa:2007:SAPTDFT3Body; @Ambrosetti:2016:WavelikeVdW] (a “body” is a fragment). In turn, even low-order treatments of inter-fragment electron correlation are sensitive to the description of local correlation, given the sensitivity of polarizabilities to correlation level [@Riley:2010:IntermolForceDissect]. This coupling between local and long-range correlation is missing in standard local correlation models [@Forner:1985:LocalCorrelation; @Stoll:1992:LocalCorrDiamond; @Saebo:1993:LocalCorrelation; @Schutz:2001:DomainCCSD; @Maslen:2005:MP4TRIM; @Subotnik:2006:SmoothLocalCCSD; @Li:2010:ClusterInMolecule; @Hattig:2012:PNOMollerPlesset; @Kristensen:2012:DivideExpandConsolidate; @Liakos:2015:PNOCClimits], to the detriment of the long-range dispersion that holds together large systems. ]{} An unexplored approach to fragment-based electronic structure calculations is to attempt to rewrite the super-system Hamiltonian exactly in terms of interacting fluctuations among *internally correlated* states of fragments, such that its form mimics the familiar field-operator expression for the *ab initio* Hamiltonian. The degrees of freedom in the global computation are then fluctuations of entire subunits. By making truncations in this context, high-energy local arrangements of electrons are expunged from consideration, and an effective suppression of individual degrees of freedom results. Given the relative complexity of intra-fragment electron correlation (as compared to the simple picture of inter-fragment interactions), it is reasonable to view the internal coordinates of fragments in terms of such collective motions. For the sake of being concrete, let us assert that a super-system Hamiltonian ${{\hat {\mathcal{H}}}}$ may be expressed in the following manner $$\begin{aligned} \label{Hform} {{\hat {\mathcal{H}}}} = \sum_m\sum_{i_m,j_m} H^{i_m}_{j_m} ~ {{\hat {\tau}}}^{j_m}_{i_m} ~~ + \sum_{m_1<m_2}\sum_{\substack{i_{m_1},j_{m_1} \\ i_{m_2},j_{m_2}}} H^{i_{m_1} i_{m_2}}_{j_{m_1} j_{m_2}} ~ {{\hat {\tau}}}^{j_{m_1}}_{i_{m_1}} \, {{\hat {\tau}}}^{j_{m_2}}_{i_{m_2}} ~~ + ~ \cdots \end{aligned}$$ The two-index fluctuation operators are analogues of paired field operators; the effect of ${{\hat {\tau}}}^{j_m}_{i_m}$ is to induce a fluctuation of fragment $m$ only, from some state ${|{\psi_{j_m}}\rangle}$ to another state ${|{\psi_{i_m}}\rangle}$. (The positioning of indices is addressed later.) This implies a basis of super-system states, denoted $$\begin{aligned} {|{\Psi_I}\rangle} = {|{\psi_{i_1} \psi_{i_2} \cdots }\rangle} \end{aligned}$$ which have tensor-product-like structure in terms of internally correlated fragment states ${|{\psi_{i_m}}\rangle}$. The collection of fragment-state labels into a single ket implies global antisymmetry of the electronic wavefunction. The elements $H^{i_m}_{j_m}$ in [eq. (\[Hform\])]{} build a Hamiltonian matrix for fragment $m$, and the higher-order terms are responsible for couplings between fragments ($m_1$$<$$m_2$ under the summation runs over all unique pairs). ![\[concept\] An interaction between a pair of molecules can be thought of in terms of individual electrons fluctuating among local orbitals, whereby a “primary” excitation (denoted with an asterisk) is accompanied by a number of other connected relaxations (nominal excitations). The same interaction can also be conceived of in terms of fluctuations between electronically correlated states of the fragments, which inherently contain these relaxations. Given the relatively large energetic scale of intra-fragment correlation, the relaxations that accompany local fluctuations should be largely similar across interactions [ of one fragment with many neighbors, as illustrated in the panel to the right. As such, the structure of these relaxations ]{} need only be computed once for each fragment. ](figure1.jpg){width="14.5cm"} The internally correlated fragment states will be referred to as the excitonic basis, since fluctuations among these are the conceptual site basis from which Frenkel–Davydov excitons [@Frenkel:1931:Excitons; @Davydov:1971:MolecularExcitons; @May:2011:ChgNrgXfer] are built. If the Hamiltonian can indeed be represented exactly in terms of such renormalized fluctuations, then recourse to the primitive electronic degrees of freedom can be avoided at the global scope. The intuition behind the expected efficiency of the proposed scheme is illustrated pictorially in [Fig. \[concept\]]{}. Local relaxations that accompany the primary motions of fluctuating electrons are computed once for each fragment, and these are permanently folded into the effective Hamiltonian. As the fragment bases are truncated to reduce the dimensionality of the global calculation, strong intra-fragment correlations remain constant features of the retained state space that describes low-energy phenomena. Solving for intra-fragment correlation first and then interacting the fragments in the transformed picture essentially inverts the traditional paradigm of using a fragment-based decomposition of a reference wavefunction as a starting point for handling global electron correlation. Traditional electronic structure approaches are available to treat inter-fragment interactions in the excitonic picture. [ In principle, even the covalent bonding interaction can be represented by allowing correlated charge-state fluctuations, ]{} [ though excitonic renormalization is probably most useful for weakly interacting fragments. ]{} System-wide induction can be captured at the mean-field level, but using polarizabilities from internally correlated fragments. The usual post-mean-field approaches ([*e.g.*, ]{}[perturbation theory, coupled-cluster theory]{}), can be used to handle inter-fragment electron correlation ([*i.e.*, ]{}[Van der Waals forces]{}). [ Excitonically renormalized Hamiltonians might also be used in dynamical mean-field theory calculations for the electronic structures of crystals [@Kotliar:2006:DynamicalMeanField]. ]{} The computational cost for a global calculation (after the effective Hamiltonian is obtained) will depend only on the numbers of fragments and states per fragment, not on the internal structure of those fragment states. [ Block correlated (BC) methods [@Shen:2009:BCCCexcitedstates; @Xu:2013:BCPT2] ]{} [ and the active-space decomposition (ASD) approach [@Parker:2014:ActiveSpaceDecompDMRG; @Kim:2015:ActiveSpaceCovalent] both ]{} [ use a partitioning of electron correlation that is formally similar to this work. However, BC methods were not conceived as fragment methods (rather used to isolate strong correlations in a molecule), ]{} [ and work on the ASD approach has not yet allowed for charge transfer or developed a field-operator-like Hamiltonian resolution that would be amenable to, say, coupled-cluster theory. Importantly, both BC and ASD were also only proposed for non-overlapping (explicitly orthogonalized) subsystems, such that recovering the use of the full basis assigned to any fragment would involve formal charge transfers to its neighbors. ]{} The concept of working with general fragment states in an *ab initio* method is also not generally new. Most notably, it is prominent in the formal development of symmetry-adapted perturbation theory [@Jeziorski:1994:SAPT; @Hohenstein:2010:SAPTtriples]. However, the computational implementation still proceeds in terms of a Hamiltonian described in the one-electron basis, and extension of the globally antisymmetrized working equations beyond dimers is algebraically tedious. Other recent and compelling work on *ab initio* exciton theory [@Sisto:2014:AbInitoExciton; @Morrison:2015:AbInitioExciton] has focused specifically on the excited-state regime; however, these approaches currently lack a clear and tractable scheme for systematic inclusion of ever higher levels of electron correlation, which the approach should provide. The purpose of this article is to demonstrate rigorously that the Hamiltonian for a fragment-decomposed electronic system may indeed be written exactly in the form asserted in [eq. (\[Hform\])]{}, also when fragment orbitals overlap each other. We furthermore provide practicable recipes for building such excitonic Hamiltonian resolutions. The difficultly in developing rigorous and computationally tractable definitions for the super-system basis and associated fluctuation operators for overlapping fragments lies with inter-fragment electron-exchange antisymmetry. We start by next giving a conceptual outline of the loosely deductive process that leads to the resolution of the exchange problem pursued here. The bulk of the article is then spent making those concepts rigorous, in a bottom-up procedural manner. A final section before the conclusion is dedicated to discussing the practical aspects of computing the necessary matrix elements, including the applicability of approximation schemes. Approach to the Inter-fragment Exchange Problem \[exchange\_concept\] ===================================================================== The root of the exchange problem [ ([*i.e.*, ]{}[the practical difficulty of enforcing wavefunction antisymmetry]{}) ]{} is that the one-electron spaces describing each fragment cannot be orthogonalized against each other without damaging the descriptions of the correlations internal to the fragments (and in ill-defined ways that would depend on the orthogonalization scheme). We therefore take as a premise that the one-electron spaces available to each of the fragments are not to be redefined. Due to Pauli exclusion then, the available space on one fragment will be dynamically entangled with the fluctuating, internally correlated electronic structures of its neighbors. The fact that a fluctuation on one fragment then effectively changes the states of all of its neighbors initially casts doubt on whether the Hamiltonian can be expressed in terms of a *low-order* expansion of single-fragment fluctuation operators. From a more formal perspective, although it is trivial to assert a definition of an antisymmetrized tensor product of internally correlated fragment states ${|{\Psi_I}\rangle}$, a basis of such super-system states is not generally orthonormal, a reflection of the overlap-driven exchange interaction. The lack of an orthonormal basis is naturally handled by the introduction of biorthogonal complements, denoted as ${|{\Psi^I}\rangle}$. The unique association of each complement with precisely one of the original tensor-product states can be leveraged to construct operators effecting single-fragment fluctuations between the original basis states. Although this step is relatively straightforward, the matrix elements of the consequent Hamiltonian expansion depend critically on the nature of the states used to construct them. This moves the apparent difficulty of the electron-exchange problem into the construction of suitable definitions of the complements. The biorthogonal complements are, in principle, described by the inverse of the matrix of overlaps ${\langle{\Psi_I}}{|{\Psi_J}\rangle}$. This is purely formal, however, since direct numerical inversion of such an exponentially large matrix is intractable, especially since it will not be sparse nor exactly factorizable. Let us consider more fundamental questions of structure, however, under the presumption that the biorthogonal basis is somehow available. First, the complements ${|{\Psi^I}\rangle}$ will not themselves have tensor-product-like structure in terms of the original fragment states; they will appear to contain inter-fragment correlations. Furthermore, for a given *truncation* of the fragment bases, a different effective Hamiltonian may be obtained, depending on how the complements are defined. Such truncations are critical for the efficiency reasons outlined in the introduction. The ambiguity in the definition of the complements results because the projected inverse of a matrix is not the same as the inverse of a projected matrix. Insisting that the complements live in the post-truncation model space would be different than choosing them to be the corresponding members of the set of complements to the untruncated tensor-product set, or any other set. This unsettling ambiguity would likely be tolerable as an approximation for sufficiently large model spaces, but the choice of complements also affects the formal fragment order of the Hamiltonian expansion. There is one particular choice of complements with simple structure, amenable to algebraic manipulation; these are the full-space complements of the complete tensor-product set of fragment bases (which themselves span the complete Fock spaces of their respective fragments). Given a truncated model space, the complements therefore span a different subspace, but of the same dimensionality. The nature of their simplicity is that each complement does, in fact, have tensor-product-like structure, but in terms of secondary sets of states associated with each fragment. Each of the members of the implied secondary fragment bases has unit overlap with exactly one of the original states of its associated fragment, and it is strictly orthogonal to all other fragment basis states, both on that same fragment or on any other fragment. This is the key to achieving both the biorthogonality needed to define the fluctuation operators and the fragment locality needed obtain a low-order Hamiltonian expansion [ with tractable matrix elements. ]{} The forthcoming procedure starts by systematically constructing definitions of biorthogonal one-electron, single-fragment, and super-system bases. Once the super-system bases are defined, the relatively straightforward definition of the fluctuation operators in terms of them is presented. Using the properties of the bases forwarded, the corresponding matrix elements in the excitonic Hamiltonian expansion may be deduced, showing that it has a low-fragment-order resolution. The Excitonic Hamiltonian for Electronic Systems ================================================ Notation and conventions ------------------------ The scope of this work is limited to the non-relativistic electronic Hamiltonian, projected into a basis, for fixed nuclear positions. The system is divided into $N$ enumerated fragments, which are disjoint groups of atoms. For conceptual ease, we presume that the atoms of each fragment occupy the spatial positions that they have in the super-system, but the many-electron state space associated with each is defined as though it were in isolation at that location. Lower-case latin letters will be used for integer indices, and upper-case latin letters will be used for ascending-ordered tuples of integers, for example, $I = (i_1, \cdots i_{\ell_I})$, where $\ell_I$ is the length of tuple $I$. The greek letters $\Psi$ and $\psi$ will be used to refer to general states of the super-system and fragments, respectively, and $\Phi$ and $\phi$ will refer to the respective single-determinant states from which these are built. We use $\chi$ for one-electron orbitals. On matrix-valued quantities, subscripts and superscripts will be used to index covariant and contravariant dimensions, respectively. If a matrix has both a covariant and a contravariant index, then, for purposes of matrix multiplication, the rows are to be enumerated by the contravariant (typically bra) index. Sets are abbreviated as $\{y_i\}$ to represent all $y_i$ corresponding to values of $i$ that are defined for mapping $y$. Summations implicitly run over all values of an index that are allowed by the mapping to which the index is attached. The one-electron bases ---------------------- We presume a set of linearly independent orbitals $\{{|{\chi_p}\rangle}\}$, where each orbital is associated with a specific fragment ([*e.g.*, ]{}[atomic orbitals of constituent atoms, molecular orbitals of fragments, [*etc.*]{}]{}). The index that enumerates this set is taken to be “blocked” by fragment, such that the first block of consecutive values enumerates the orbitals on fragment 1, and so forth. For ease of discussion, we take the space spanned by the one-electron basis for a given problem as our working definition of complete, consequently defining what is meant by completeness of a many-electron Hilbert space and completeness of the Fock space (having electron numbers up to the cardinality of $\{{|{\chi_p}\rangle}\}$). We let the complete set of biorthogonal complement orbitals be denoted $\{{|{\chi^p}\rangle}\}$, such that ${\langle{\chi^p}}{|{\chi_q}\rangle} = \delta_{pq}$. Should a linearly dependent set of functions be proposed, then some difficulty arises. The resolution of this (discussed [ in [appendix \[lin\_alg\]]{}) ]{} maps the linearly dependent case onto a problem of the same structure; therefore, the remainder of this discussion is general. Field-operator notation will be convenient for developing transparent rules for matrix elements, and there are some important subtleties when working with biorthogonal bases [ [@Helgaker:2002:PurpleBook; @HeadGordon:1998:Biorthogonal]. ]{} Here we let ${{\hat {c}}}_p$ and ${{\hat {c}}}^p$ denote field operators that correspond to creation of ${|{\chi_p}\rangle}$ and ${|{\chi^p}\rangle}$, respectively, within a single-determinant state. The Hermitian conjugates of these operators, denoted ${{\hat {a}}}_p$ and ${{\hat {a}}}^p$, respectively, act as annihilation operators of their *complements* within ket states. The biorthogonal field operators obey the following anticommutation relationships (among others) $$\begin{aligned} \label{bo_anticomm} \big[ {{\hat {c}}}_p , {{\hat {a}}}^q \big]_+ = \delta_{pq} \;\;\;\;\; \;\;\;\;\; \big[ {{\hat {c}}}_p , {{\hat {c}}}_q \big]_+ = 0 \;\;\;\;\; \;\;\;\;\; \big[ {{\hat {a}}}^p , {{\hat {a}}}^q \big]_+ = 0 \end{aligned}$$ allowing us to write the *ab initio* Hamiltonian as $$\begin{aligned} \label{field_opH} &~& {{\hat {\mathcal{H}}}} = \sum_{p,q} h^p_q ~ {{\hat {c}}}_p {{\hat {a}}}^q + \sum_{p,q,r,s} v^{pq}_{rs} ~ {{\hat {c}}}_p{{\hat {c}}}_q{{\hat {a}}}^s{{\hat {a}}}^r {\nonumber \\}&~& h^p_q = {\langle{\chi^p}|} {{\hat {h}}} {|{\chi_q}\rangle} \;\;\;\;\; v^{pq}_{rs} = \frac{1}{4} {\langle{\chi^p \chi^q}|}{{\hat {v}}}{|{\chi_r \chi_s}\rangle} \end{aligned}$$ where ${{\hat {h}}}$ is the combined kinetic energy and nuclear attraction operator, and ${{\hat {v}}}$ is the electron–electron repulsion operator. [ The two-electron integrals here are implicitly antisymmetrized since the bra and ket are Slater determinants. ]{} [ While both integral tensors here lack the usual symmetry with respect to permuting bra and ket indices, ]{} the anticommutation relationships for the field operators recover all of the rules that are familiar from the orthonormal case for making orbital substitutions and taking matrix elements, so long as bra states are expressed in the complement basis. Many-electron bases ------------------- ### Super-system determinant bases The complete Fock space of the super-system is spanned by the set of single-determinant configurations $\{{|{\Phi_P}\rangle}\}$ where $$\begin{aligned} \label{detdef1} {|{\Phi_P}\rangle} ~=~ {|{\chi_{p_1} \cdots \chi_{p_{n_P}}}\rangle} ~=~ {{\hat {c}}}_{p_1} \cdots {{\hat {c}}}_{p_{n_P}} {|{~}\rangle} \end{aligned}$$ The index $P = (p_1, \cdots p_{n_P})$ can be any ordered tuple of any allowed length (up to the basis size), where $n_P$ is the number of electrons in configuration ${|{\Phi_P}\rangle}$, and ${|{~}\rangle}$ is the true vacuum. For illustrative purposes, we note that ${|{\Phi_P}\rangle}$ could also be written as $$\begin{aligned} \label{detdef2} {|{\Phi_P}\rangle} = \sqrt{n_P!} \, {{\hat {\mathcal{P}}}}_\text{A} \big[ {|{\chi_{p_1}}\rangle} \otimes \cdots {|{\chi_{p_{n_P}}}\rangle} \big] \end{aligned}$$ where ${{\hat {\mathcal{P}}}}_\text{A}$ is an orthogonal projector from the space of all raw (asymmetric) orbital-tensor-product states onto the subspace of all antisymmetric states ([*i.e.*, ]{}[the complete Fock space]{}). [ The $\sqrt{n_P!}$ prefactor of [eq. (\[detdef2\])]{} is necessary so that the norm of this state matches that in [eq. (\[detdef1\])]{}, which would be unity if the orbitals were orthonormal. ]{} If the component indices of a tuple $P$ are chosen such that they all identify orbitals on a single fragment, then the use of the lower case in ${|{\phi_P}\rangle}$ will emphasize this. The set $\{{|{\phi_P}\rangle}\}$, a strict subset of $\{{|{\Phi_P}\rangle}\}$, is then the set of all possible single-fragment determinant states on all possible fragments. If we wish to indicate the fragment $m$ to which a given such determinant belongs, then the notation ${|{\phi_{P_m}}\rangle}$ is used; the subscript $m$ is thought of as placing a restriction on the possible values of the tuple. A summation over an index $P_m$ would run over only those determinant states of fragment $m$. A general tuple $P$ may be subdivided into the (potentially empty) tuples $P_1$ through $P_N$, each containing only the component orbital indices of $P$ that belong to the fragment indicated by the subscript. (It is simply convenient that the indexing of the sub-tuples of $P$ is coincident with the convention used to indicate fragment-based restrictions.) Since all tuples in this work are taken to be ordered, and since the indexing of the orbitals is blocked by fragment, $P$ is always reconstructed by simple concatenation of the sub-tuples. This then allows us to write $$\begin{aligned} \label{antisymmpdt} {|{\Phi_P}\rangle} ~=~ {|{\phi_{P_1} \cdots \phi_{P_N}}\rangle} &=& {{\hat {\phi}}}_{P_1} \cdots {{\hat {\phi}}}_{P_N} {|{~}\rangle} {\nonumber \\}&=& \sqrt{n_P! / ( n_{P_1}! \cdots n_{P_N}!)} \, {{\hat {\mathcal{P}}}}_\text{A} \big[{|{\phi_{P_1}}\rangle} \otimes \cdots {|{\phi_{P_N}}\rangle} \big] \end{aligned}$$ where ${{\hat {\phi}}}_{P_m}$ collects (in ascending order) the creation operators of the orbitals in $P$ that belong to fragment $m$. The second line of [eq. (\[antisymmpdt\])]{} is shown to be equal to ${|{\Phi_P}\rangle}$ by careful manipulation of the nested antisymmetrizations as defined in [eq. (\[detdef2\])]{}, and this writing emphasizes the connection to a tensor product of fragment states. In either notation for [eq. (\[antisymmpdt\])]{}, it is clearly meaningful to say that fragment $m$ is in state ${|{\phi_{P_m}}\rangle}$ in the globally antisymmetric super-system state ${|{\Phi_P}\rangle}$. We have the intuitive result that the complete basis $\{{|{\Phi_P}\rangle}\}$ is also described as the set of all antisymmetrized tensor products of determinants for the constituent fragments. Using the same conventions as above, but applied to the orbitals $\{{|{\chi^p}\rangle}\}$, we have the set $\{{|{\Phi^P}\rangle}\}$, whose members satisfy $$\begin{aligned} \label{detcomplement} {\langle{\Phi^P}|} ~=~ {\langle{\phi^{P_1} \cdots \phi^{P_N}}|} &=& {\langle{~}|} {{\hat {\phi}}}^{P_N} \cdots {{\hat {\phi}}}^{P_1} {\nonumber \\}&=& \sqrt{n_P! / ( n_{P_1}! \cdots n_{P_N}!)} \, \big[{\langle{\phi^{P_1}}|} \otimes \cdots {\langle{\phi^{P_N}}|} \big] {{\hat {\mathcal{P}}}}_\text{A} \end{aligned}$$ where the operator ${{\hat {\phi}}}^{P_m}$ is built from annihilation operators (with indices in descending order). We can quickly verify that $\{{|{\Phi^P}\rangle}\}$ is the set of biorthogonal complements to $\{{|{\Phi_P}\rangle}\}$ by using the aforementioned anticommutation relationships $$\begin{aligned} {\langle{\Phi^P}}{|{\Phi_Q}\rangle} = {\langle{\chi^{p_1} \cdots \chi^{p_n}}}{|{\chi_{q_1} \cdots \chi_{q_n}}\rangle} = \big\langle {{\hat {a}}}^{p_{n_P}} \cdots {{\hat {a}}}^{p_1} {{\hat {c}}}_{q_1} \cdots {{\hat {c}}}_{q_{n_Q}} \big\rangle = \delta_{PQ} \end{aligned}$$ where the angle brackets denote a vacuum expectation value. This makes use of the fact that, since the tuples are ordered, if $P \neq Q$, then they are different in composition. ### General tensor-product super-system bases We can now define a basis for the super-system Fock space, in terms of more general fragment states. Let the set $\{{|{\psi_i}\rangle}\}$ collect all such basis states for all fragments. These are defined by introduction of an invertible matrix ${\textbf{\textit{z}}}$ with elements $z^P_i$, such that $$\begin{aligned} \label{general_frag} {|{\psi_{i_m}}\rangle} ~=~ \sum_{P_m} z^{P_m}_{i_m} \, {|{\phi_{P_m}}\rangle} ~=~ \Big( \sum_{P_m} z^{P_m}_{i_m} ~ {{\hat {\phi}}}_{P_m} \Big) {|{~}\rangle} ~=~ {{\hat {\psi}}}_{i_m} {|{~}\rangle} \end{aligned}$$ The square matrix ${\textbf{\textit{z}}}$ has rows indexed by a tuple and columns indexed by an integer. The subscript on $i_m$ restricts the index to refer to one of the states of fragment $m$. As an artifact of our notation for blocking indices by fragment, ${\textbf{\textit{z}}}$ formally has elements that would mix determinants on separate fragments, but we insist that it is block-diagonal by fragment; when necessary, we let the diagonal block of ${\textbf{\textit{z}}}$ for fragment $m$ be denoted as ${\textbf{\textit{z}}}^{(m)}$. For simplicity, we will also presume that ${\textbf{\textit{z}}}$ does not mix determinants of different particle number, though some of the following could be generalized beyond this. Using the biorthogonality of the fragment-determinant bases, it is straightforward to show that ${\langle{\psi^i}}{|{\psi_j}\rangle} = \delta_{ij}$ for the set $\{{|{\psi^i}\rangle}\}$, which satisfy $$\begin{aligned} \label{general_comp} {\langle{\psi^{i_m}}|} ~=~ \sum_{P_m} \bar{z}^{i_m}_{P_m} \, {\langle{\phi^{P_m}}|} ~=~ {\langle{~}|} \Big( \sum_{P_m} \bar{z}^{i_m}_{P_m} ~ {{\hat {\phi}}}^{P_m} \Big) ~=~ {\langle{~}|} {{\hat {\psi}}}^{i_m} \end{aligned}$$ where, for convenience, we use the notation $\bar{{\textbf{\textit{z}}}} = {\textbf{\textit{z}}}^{-1}$, which has elements $\bar{z}^i_P$. Since $\bar{{\textbf{\textit{z}}}}$ is also block-diagonal by fragment the notation ${|{\psi^{i_m}}\rangle}$ to refer to a state associated with fragment $m$ is logically sensical, regardless of ambiguous physical interpretation as such. We now construct another pair of complete biorthogonal bases for the super-system Fock space, $\{{|{\Psi_I}\rangle}\}$ and $\{{|{\Psi^I}\rangle}\}$, where $$\begin{aligned} \label{fieldopstates} {|{\Psi_I}\rangle} ~=~ {|{\psi_{i_1} \cdots \psi_{i_N}}\rangle} &=& {{\hat {\psi}}}_{i_1} \cdots {{\hat {\psi}}}_{i_N} {|{~}\rangle} {\nonumber \\}&=& \sqrt{n_I! / (n_{i_1}! \cdots n_{i_N}!)} \, {{\hat {\mathcal{P}}}}_\text{A} \big[{|{\psi_{i_1}}\rangle} \otimes \cdots {|{\psi_{i_N}}\rangle} \big] {\nonumber \\}{\langle{\Psi^I}|} ~=~ {\langle{\psi^{i_1} \cdots \psi^{i_N}}|} &=& {\langle{~}|} {{\hat {\psi}}}^{i_N} \cdots {{\hat {\psi}}}^{i_1} {\nonumber \\}&=& \sqrt{n_I! / (n_{i_1}! \cdots n_{i_N}!)} \, \big[{\langle{\psi^{i_1}}|} \otimes \cdots {\langle{\psi^{i_N}}|} \big] {{\hat {\mathcal{P}}}}_\text{A} \end{aligned}$$ where $I=(i_1,\cdots i_N)$ gives the indices of the states of each of the sub-systems. The field-operator notation is an elegant manner to enforce global antisymmetry, whereas the tensor-product notation connects more directly to the state-space descriptions of the fragments. It is notable that the state-space definition of antisymmetrization for general states is the same as that applied to single-determinant states; although it does not rely on permutation operators or determinant arithmetic, it does rely on the fragment states having definite particle number. From either approach to antisymmetrization, the following connections between the two pairs of biorthogonal bases are obtained $$\begin{aligned} {|{\Psi_I}\rangle} &=& \sum_P Z^P_I \, {|{\Phi_P}\rangle} ~~;~~~~~~~~ Z^P_I = \prod_m z^{P_m}_{i_m} \label{superbasis} \\ {\langle{\Psi^I}|} &=& \sum_P \bar{Z}^I_P \, {\langle{\Phi^P}|} ~~;~~~~~~~~ \bar{Z}^I_P = \prod_m \bar{z}^{i_m}_{P_m} \label{supercompl} \end{aligned}$$ where, in the definitions of the elements of ${\textbf{\textit{Z}}}$ and $\bar{{\textbf{\textit{Z}}}}$, the index $m$ runs over all of the fragments. It is easy to verify that ${\textbf{\textit{Z}}}$ and $\bar{{\textbf{\textit{Z}}}}$ are mutually inverse, yielding ${\langle{\Psi^I}}{|{\Psi_J}\rangle} = \delta_{IJ}$. Since ${\textbf{\textit{Z}}}$ is invertible, the set $\{{|{\Psi_I}\rangle}\}$ (and also $\{{|{\Psi^I}\rangle}\}$) is a complete basis. In the most straightforward conceptualization, the orbitals on each [ individual ]{} fragment could be taken to be orthonormal among each other [ (though overlapping the orbitals of neighboring fragments), ]{} and ${\textbf{\textit{z}}}$ could be chosen to be unitary. In particular, ${|{\psi_{i_m}}\rangle}$ might (theoretically) be a full configuration-interaction energy eigenstate of fragment $m$. Overlaps between orbitals on different fragments would still require us to invoke the biorthogonal machinery, however. Later, we will also suggest an approach in which ${\textbf{\textit{z}}}$ is not unitary, opening the door to less computationally expensive parameterizations. The states described by ${\textbf{\textit{z}}}$ should also not necessarily be fragment eigenstates, but rather those that most efficiently and accurately describe interfragment correlations. Independent of the level of theory used for the fragment states, however, only a small fraction of the elements of ${\textbf{\textit{z}}}$ and $\bar{{\textbf{\textit{z}}}}$ will ever actually be computed in practice (even implicitly), due to truncations of the fragment spaces. Single-fragment fluctuation operators ------------------------------------- We now construct explicit expressions for fluctuation operators that effect transitions of single fragments, regardless of the states of the other sub-systems. [ These can be thought of as excitations, de-excitations, [[*etc.*]{}]{} ]{} We require that the action of ${{\hat {\tau}}}_{i_m}^{j_m}$ onto super-system basis state ${|{\Psi_K}\rangle}$ is as follows $$\begin{aligned} \label{action_definition} {{\hat {\tau}}}_{i_m}^{j_m}{|{\psi_{k_1}\cdots\psi_{k_m}\cdots\psi_{k_N}}\rangle} = \delta_{j_m,k_m}{|{\psi_{k_1}\cdots\psi_{i_m}\cdots\psi_{k_N}}\rangle} \end{aligned}$$ This action is reminiscent of a number-conserving pair of field operators onto a single-determinant electronic state, such that the null state results if the upper (“destruction”) index corresponds to an “empty” fragment state. As shown, the lower and upper indices must refer to (potentially identical) states of the same fragment. Operators obeying the above requirement will have the following commutation relation by definition $$\begin{aligned} \label{the_commutator} \big[{{\hat {\tau}}}^j_i,{{\hat {\tau}}}^l_k\big]_- = ~ \delta_{jk}\,{{\hat {\tau}}}^l_i ~-~ \delta_{il}\,{{\hat {\tau}}}^j_k \end{aligned}$$ This is shown from [eq. (\[action\_definition\])]{} by noting, first, that operators on different fragments commute, and, second, that a string of two operators on the same fragment gives the null state if the upper index of the left operator does not match the lower index of the right operator. This is an important property concerning manipulation of fluctuation operators with techniques analogous to those familiar from field operators. We make use of this in the companion paper, which presents an excitonic coupled-cluster method. It may seem desirable to make use of $\{{{\hat {\psi}}}_{i_m}\}$ and $\{{{\hat {\psi}}}^{i_m}\}$ directly to define state-to-state transition operators. However, these operators do not obey the requisite canonical (anti)commutation relationships with their conjugates (consider especially the case of operators for differing particle numbers). Such relationships would be necessary to generalize the notion of creation and annihilation operators [@Anderson:1994:CanonicalTrans], and these operators consequently do not transform as field operators with respect to changes of the many-electron basis. Although these apparent many-electron creation operators will be convenient later, reformulation of the coming fluctuation-operator definitions in terms of them would necessarily retain a reference to the absolute vacuum dyadic ${|{~}\rangle}{\langle{~}|}$. Since we cannot abandon the language of state dyadics for the fluctuations, we proceed entirely in a state-space notation. The following may then be regarded as a definition of a sub-system fluctuation operator on fragment $m$ $$\begin{aligned} \label{fluctopcorr} {{\hat {\tau}}}_{i_m}^{j_m} = \sum_{k_1}\cdots\sum_{k_{m-1}}\sum_{k_{m+1}}\cdots\sum_{k_N} \, {|{\psi_{k_1}\cdots\psi_{k_{m-1}}\psi_{i_m}\psi_{k_{m+1}}\cdots\psi_{k_N}}\rangle} {\langle{\psi^{k_1}\cdots\psi^{k_{m-1}}\psi^{j_m}\psi^{k_{m+1}}\cdots\psi^{k_N}}|} {\nonumber \\}\end{aligned}$$ On account of the biorthogonality of the bases $\{{|{\Psi_I}\rangle}\}$ and $\{{|{\Psi^I}\rangle}\}$, a basis state acted upon by this operator will have non-zero projection onto, at most, one bra in the summation, and that will only happen if fragment $m$ is in state ${|{\psi_{j_m}}\rangle}$, which would then give unit coefficient to the super-system basis state that simply has ${|{\psi_{j_m}}\rangle}$ replaced by ${|{\psi_{i_m}}\rangle}$, as required by [eq. (\[action\_definition\])]{}. Clearly, the choice of basis states for the single-fragment Fock spaces does not change this discussion. We may therefore introduce an analogous set of operators defined with respect to the determinant bases, denoted for convenience as the set $\{{{\hat {\sigma}}}_P^Q\}$ $$\begin{aligned} \label{fluctopdet} {{\hat {\sigma}}}_{P_m}^{Q_m} = \sum_{R_1}\cdots\sum_{R_{m-1}}\sum_{R_{m+1}}\cdots\sum_{R_N} \, {|{\phi_{R_1}\cdots\phi_{R_{m-1}}\phi_{P_m}\phi_{R_{m+1}}\cdots\phi_{R_N}}\rangle} {\langle{\phi^{R_1}\cdots\phi^{R_{m-1}}\phi^{Q_m}\phi^{R_{m+1}}\cdots\phi^{R_N}}|} {\nonumber \\}\end{aligned}$$ In [appendix \[completeness\]]{}, it is shown that either $\{{{\hat {\tau}}}_i^j\}$ or $\{{{\hat {\sigma}}}_P^Q\}$ can be used to build a complete basis for the space of all super-system Fock-space operators. In addition to abstractly assuring us that a Hamiltonian expansion in terms of fragment fluctuations is possible, any member of $\{{{\hat {\tau}}}_i^j\}$ must itself be resolvable in terms of $\{{{\hat {\sigma}}}_P^Q\}$, and vice versa. Concretely, the transformation is seen to be rather simple. Insertion of the resolutions the members of $\{{|{\psi_i}\rangle}\}$ and $\{{|{\psi^i}\rangle}\}$ in terms of the members of $\{{|{\phi_P}\rangle}\}$ and $\{{|{\phi^P}\rangle}\}$, or vice versa, into the definition of either ${{\hat {\tau}}}_{i_m}^{j_m}$ or ${{\hat {\sigma}}}_{P_m}^{Q_m}$ in [eqs. (\[fluctopcorr\]) and (\[fluctopdet\])]{} results in $N$$-$$1$ contractions of the diagonal blocks of ${\textbf{\textit{z}}}$ with $\bar{{\textbf{\textit{z}}}}$. Resolving the consequent Kronecker deltas gives $$\begin{aligned} \label{preserves_rank} {{\hat {\tau}}}_{i_m}^{j_m} &=& \sum_{P_m} \sum_{Q_m} ~ z^{P_m}_{i_m} \, \bar{z}^{j_m}_{Q_m} ~ {{\hat {\sigma}}}_{P_m}^{Q_m} {\nonumber \\}{{\hat {\sigma}}}_{P_m}^{Q_m} &=& \sum_{i_m} \sum_{j_m} ~ \bar{z}^{i_m}_{P_m} \, z^{Q_m}_{j_m} ~ {{\hat {\tau}}}_{i_m}^{j_m} \end{aligned}$$ Effectively, each index transforms separately, with one power of ${\textbf{\textit{z}}}$ or $\bar{{\textbf{\textit{z}}}}$. Inserting one of these identities into the other results in a truism. Importantly, these transformations preserve fragment rank. The Hamiltonian in terms of fragment fluctuation operators ---------------------------------------------------------- We may now resolve the Hamiltonian ${{\hat {\mathcal{H}}}}$ in terms of the single-fragment fluctuations. For simplicity, we begin with the determinant basis. In order to resolve the matrix elements ${\langle{\Phi^P}|}{{\hat {\mathcal{H}}}}{|{\Phi_Q}\rangle}$, we first decompose the *ab initio* expression for ${{\hat {\mathcal{H}}}}$ in [eq. (\[field\_opH\])]{} as $$\begin{aligned} {{\hat {\mathcal{H}}}} = {{\hat {H}}}_1 + {{\hat {H}}}_2 + {{\hat {H}}}_3 + {{\hat {H}}}_4 \end{aligned}$$ ${{\hat {H}}}_1$ collects together all terms from both the one-electron and two-electron parts of ${{\hat {\mathcal{H}}}}$ that have all indices referring to orbitals of *any* single fragment, and ${{\hat {H}}}_2$ similarly collects terms for all pairs of fragments (dimers). It will also be useful further decompose each of the ${{\hat {H}}}_M$ by separating out those terms that act on specific groups of fragments, for example, for ${{\hat {H}}}_4$, $$\begin{aligned} \label{further_decomposed} {{\hat {H}}}_4 = \sum_{m_1<m_2<m_3<m_4} {{\hat {H}}}^{(m_1,m_2,m_3,m_4)} \end{aligned}$$ where $m_1$$<$$m_2$$<$$m_3$$<$$m_4$ under the summation runs over all unique tetramers. The decomposition of ${{\hat {\mathcal{H}}}}$ truncates after ${{\hat {H}}}_4$, since there are a maximum of four orbital indices. Notably, all terms in ${{\hat {H}}}_3$ and ${{\hat {H}}}_4$ must induce an inter-fragment charge transfer somewhere in the system. [ In parallel to standard practice for matrix elements of one- and two-electron operators, it will be convenient to frame the discussion in terms of the number of fragments that have changed state in the bra, relative to the ket (henceforth, the number of *substitutions*). The logic for obtaining matrix elements closely mirrors the rules for determinant matrix elements of one- and two-electron operators. For example, on account of the anticommutation rules for the field operators, we know that a matrix element of a given ${{\hat {H}}}_M$ will be zero if the number of substitutions is greater than $M$. In addition to this, the fact that any term of any ${{\hat {H}}}_M$ operates on a maximum of two *electrons* places further restrictions on non-zero elements. For example, ${{\hat {H}}}_4$ can have no non-zero matrix elements between states that differ by less than four fragment substitutions. Similarly, ${{\hat {H}}}_3$ only has non-zero matrix elements between states that differ by either two or three substitutions (double substitutions represent a charge transfer in the average field of a third fragment). ]{} [ To denote substitutions, the number of primes on a tuple index will be used to denote the number of substitutions relative to the unprimed index, and an overbar will denote a changed value of a sub-tuple therein. For example, for two substitutions, $P'' = (P_1, \cdots \bar{P}_{m'} \cdots \bar{P}_{m''} \cdots P_N)$, where the unsubstituted tuple is $P = (P_1, \cdots P_{m'} \cdots P_{m''} \cdots P_N)$, and the fragments undergoing the substitution have been identified as $m'$ and $m''$. We will always assume $m'$$<$$m''$$<$$\cdots$. ]{} Recalling also that ${{\hat {\mathcal{H}}}}$ conserves overall particle number, we obtain the following expressions for the complete collection of non-zero matrix elements $$\begin{aligned} \label{nonzero_matelements} {\langle{\Phi^P}|}{{\hat {H}}}_1{|{\Phi_P}\rangle} &=& \sum_m {\langle{\phi^{P_m}}|} {{\hat {H}}}_1 {|{\phi_{P_m}}\rangle} {\nonumber \\}{\langle{\Phi^{P'}}|}{{\hat {H}}}_1{|{\Phi_P}\rangle} &=& {\langle{\phi^{\bar{P}_{m'}}}|} {{\hat {H}}}_1 {|{\phi_{P_{m'}}}\rangle} {\nonumber \\}{\langle{\Phi^P}|}{{\hat {H}}}_2{|{\Phi_P}\rangle} &=& \sum_{m_1<m_2} {\langle{\phi^{P_{m_1}}\phi^{P_{m_2}}}|} {{\hat {H}}}_2 {|{\phi_{P_{m_1}}\phi_{P_{m_2}}}\rangle} {\nonumber \\}{\langle{\Phi^{P'}}|}{{\hat {H}}}_2{|{\Phi_P}\rangle} &=& \sum_{m} {\langle{\phi^{\bar{P}_{m'}}\phi^{P_m}}|} {{\hat {H}}}_2 {|{\phi_{P_{m'}}\phi_{P_{m}}}\rangle} {\nonumber \\}{\langle{\Phi^{P''}}|}{{\hat {H}}}_2{|{\Phi_P}\rangle} &=& {\langle{\phi^{\bar{P}_{m'}}\phi^{\bar{P}_{m''}}}|} {{\hat {H}}}_2 {|{\phi_{P_{m'}}\phi_{P_{m''}}}\rangle} {\nonumber \\}{\langle{\Phi^{P''}}|}{{\hat {H}}}_3{|{\Phi_P}\rangle} &=& \sum_{m} (-1)^{\alpha^{P''}_m} {\langle{\phi^{\bar{P}_{m'}}\phi^{\bar{P}_{m''}}\phi^{P_m}}|} {{\hat {H}}}_3 {|{\phi_{P_{m'}}\phi_{P_{m''}}\phi_{P_m}}\rangle} {\nonumber \\}{\langle{\Phi^{P'''}}|}{{\hat {H}}}_3{|{\Phi_P}\rangle} &=& {\langle{\phi^{\bar{P}_{m'}}\phi^{\bar{P}_{m''}}\phi^{\bar{P}_{m'''}}}|} {{\hat {H}}}_3 {|{\phi_{P_{m'}}\phi_{P_{m''}}\phi_{P_{m'''}}}\rangle} {\nonumber \\}{\langle{\Phi^{P''''}}|}{{\hat {H}}}_4{|{\Phi_P}\rangle} &=& {\langle{\phi^{\bar{P}_{m'}}\phi^{\bar{P}_{m''}}\phi^{\bar{P}_{m'''}}\phi^{\bar{P}_{m''''}}}|} {{\hat {H}}}_4 {|{\phi_{P_{m'}}\phi_{P_{m''}}\phi_{P_{m'''}}\phi_{P_{m''''}}}\rangle} \end{aligned}$$ [ In the expansion of ${\langle{\Phi^{P'}}|}{{\hat {H}}}_2{|{\Phi_P}\rangle}$, for example, $m'$ denotes the single fragment whose state is different in the bra and ket; since this fluctuation occurs in the average field of all other fragments, the states of those fragments (as they are given in $P$) are summed over. ]{} These summations arise from the insertion of the decomposition in [eq. (\[further\_decomposed\])]{}, after which, orbitals on “spectator” fragments contribute only factors of their unit biorthogonal overlaps. Once the spectators are removed, the full ${{\hat {H}}}_M$ can be safely substituted for each of the terms of its prior decomposition according to [eq. (\[further\_decomposed\])]{}, simply to declutter the notation. For the sake of simplicity, the summations sometimes admit two copies of the same fragment state into a determinant; clearly, this evaluates to zero. The summations also allow the states of the fragments to appear out of order inside of a bra; this does not contradict our established notation, which only insists that *tuple* components are ordered. Only in one case does the reordering lead to a sign change, and that is for a matrix element of ${{\hat {H}}}_3$, when the summation index $m$ is between $m'$ and $m''$ (since a charge must have been transferred); the exponent ${\alpha^{P''}_m}$ is one in this case and zero otherwise. To complete the derivation, matrix elements of the form ${\langle{\phi^{P_{m_1}}\cdots\phi^{P_{m_M}}}|}{{\hat {H}}}_M{|{\phi_{Q_{m_1}}\cdots\phi_{Q_{m_M}}}\rangle}$ are multiplied with the products of $M$ single-fragment fluctuations $\{{{\hat {\sigma}}}_{P_m}^{Q_m}\}$ that effect the associated “substitutions” ($P_m$ and $Q_m$ might be equal). Sums of such products can be used generate the same matrix elements as given in [eq. (\[nonzero\_matelements\])]{}, and these can therefore be used to build an exact representation of ${{\hat {\mathcal{H}}}}$. Changing to the target basis of internally correlated states does not change the structure of this expression. The transformation of the operators to the set $\{{{\hat {\tau}}}_{i_m}^{j_m}\}$, *via* [eq. (\[preserves\_rank\])]{}, generates the contractions that define $\{{|{\psi^{i_m}}\rangle}\}$ and $\{{|{\psi_{i_m}}\rangle}\}$ in terms of $\{{|{\phi^{P_m}}\rangle}\}$ and $\{{|{\phi_{P_m}}\rangle}\}$ in [eqs. (\[general\_frag\]) and (\[general\_comp\])]{}; we then finally arrive at $$\begin{aligned} {{\hat {\mathcal{H}}}} &=& \sum_m \sum_{\substack{I=(i_m) \\ J=(j_m)}} \; {\langle{\Psi^I}|}{{\hat {H}}}_1{|{\Psi_J}\rangle} \, {{\hat {\tau}}}_{i_m}^{j_m} {\nonumber \\}&~& + \sum_{m_1<m_2} \sum_{\substack{I=(i_{m_1},i_{m_2}) \\ J=(j_{m_1},j_{m_2})}} {\langle{\Psi^I}|}{{\hat {H}}}_2{|{\Psi_J}\rangle} \, {{\hat {\tau}}}_{i_{m_1}}^{j_{m_1}} {{\hat {\tau}}}_{i_{m_2}}^{j_{m_2}} {\nonumber \\}&~& + \sum_{m_1<m_2<m_3} \sum_{\substack{I=(i_{m_1},i_{m_2},i_{m_3}) \\ J=(j_{m_1},j_{m_2},j_{m_3})}} {\langle{\Psi^I}|}{{\hat {H}}}_3{|{\Psi_J}\rangle} \, {{\hat {\tau}}}_{i_{m_1}}^{j_{m_1}} {{\hat {\tau}}}_{i_{m_2}}^{j_{m_2}} {{\hat {\tau}}}_{i_{m_3}}^{j_{m_3}} {\nonumber \\}&~& + \sum_{m_1<m_2<m_3<m_4} \sum_{\substack{I=(i_{m_1},i_{m_2},i_{m_3},i_{m_4}) \\ J=(j_{m_1},j_{m_2},j_{m_3},j_{m_4})}} {\langle{\Psi^I}|}{{\hat {H}}}_4{|{\Psi_J}\rangle} \, {{\hat {\tau}}}_{i_{m_1}}^{j_{m_1}} {{\hat {\tau}}}_{i_{m_2}}^{j_{m_2}} {{\hat {\tau}}}_{i_{m_3}}^{j_{m_3}} {{\hat {\tau}}}_{i_{m_4}}^{j_{m_4}} \end{aligned}$$ The summations run over all unique monomers, dimers, trimers, and tetramers embedded within the overall super-system. We have condensed this expression by using the super-system notation (${|{\Psi^I}\rangle}$ and ${|{\Psi_I}\rangle}$) also for states of dimer sub-systems ([*etc.*]{}), including monomers. It is interesting to note that this result can also be obtained by starting directly with the fact that each of the ${{\hat {\psi}}}_{i_m}$ commutes with any term in ${{\hat {\mathcal{H}}}}$ that does not act on fragment $m$. Though this is conceptually important, the ensuing detailed logic is essentially the same. The numerical evaluation of the excitonic matrix elements now requires explicit insertion of terms from the *ab initio* expansion of the Hamiltonian, but, conveniently, for small numbers of fragments. As discussed in [appendix \[completeness\]]{}, since the basis of all possible fluctuation products is overcomplete, this expression for the Hamiltonian is not unique, but we conjecture that it is the most compact. The formally quartic scaling of the number of matrix elements is simply a reflection of the size of the two-electron integrals tensor. One advantage of decomposing ${{\hat {\mathcal{H}}}}$ into the ${{\hat {H}}}_M$ is that the two-fragment components here do not require subtraction of any double-counted single-fragment energies, [[*etc.*]{}]{} Calculation and Approximation of Matrix Elements ================================================ The remaining nontrivial matter is the practical evaluation of matrix elements that resolve excitonic Hamiltonians for real systems. We will show that these matrix elements may be computed efficiently, using only one- and two-electron integrals for small numbers of fragments ($\leq$$4$) and low-electron-order ($\leq$$2$), single-fragment data to represent the effects of intra-fragment correlation. The single-fragment quantities are reusable and can be pre-computed in a single, linear-scaling step. Further simplifications can be made for these coupling elements when fragments fall outside of overlap radius and eventually reach the asymptotic regime. All of this is in spite of formally insisting on globally biorthogonalized bases. The applicability of approximation schemes will also briefly be addressed. The matrix elements we need are of the general form $$\begin{aligned} \label{x_mat_elem} {\langle{\Psi^I}|} {{\hat {H}}}_M {|{\Psi_J}\rangle} = {\langle{\psi^{i_{m_1}}\cdots\psi^{i_{m_M}}}|}{{\hat {H}}}^{(m_1,\cdots m_M)}{|{\psi_{j_{m_1}}\cdots\psi_{j_{m_M}}}\rangle} \end{aligned}$$ for $M$ up to 4. We recall that ${{\hat {H}}}^{(m_1,\cdots m_M)}$ collects all terms of the *ab initio* expression for ${{\hat {\mathcal{H}}}}$, where at least one field operator references an orbital on each of the fragments $m_1$ through $m_M$, and no other fragments. This can be written as a generic term-by-term expansion $$\begin{aligned} \label{decompH} {{\hat {H}}}^{(m_1,\cdots m_M)} = \sum_{l} h_l^{(m_1,\cdots m_M)} \, {{\hat {b}}}_l^{(\underline{m_1},\cdots m_M)}\cdots{{\hat {b}}}_l^{(m_1,\cdots \underline{m_M})} \end{aligned}$$ where $l$ simply enumerates the terms. The operator ${{\hat {b}}}_l^{(\underline{m_1},\cdots m_M)}$ collects together those field operators of the $l$-th term of ${{\hat {H}}}^{(m_1,\cdots m_M)}$ that reference orbitals belonging to the fragment $m_1$, whose index bears the underline (and so forth, for the other participating fragments). To within a phase of $\pm{1}$ (to account for permutations of field operators), $h_l^{(m_1,\cdots m_M)}$ denotes the respective one- or two-electron integral from the *ab initio* expansion of [eq. (\[field\_opH\])]{}. A matrix element of the excitonic Hamiltonian [ from [eq. (\[x\_mat\_elem\])]{} ]{} may then be decomposed as $$\begin{aligned} \label{buildHpdt} {\langle{\psi^{i_{m_1}}\cdots\psi^{i_{m_M}}}|}{{\hat {H}}}^{(m_1,\cdots m_M)}{|{\psi_{j_{m_1}}\cdots\psi_{j_{m_M}}}\rangle} = ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ {\nonumber \\}\sum_{l} \, (-1)^{\beta_{B_l}^{I,J}} \, h_l^{(m_1,\cdots m_M)} \, {\langle{\psi^{i_{m_1}}}|} {{\hat {b}}}_l^{(\underline{m_1},\cdots m_M)} {|{\psi_{j_{m_1}}}\rangle} \cdots {\langle{\psi^{i_{m_M}}}|} {{\hat {b}}}_l^{(m_1,\cdots \underline{m_M})} {|{\psi_{j_{m_M}}}\rangle} \end{aligned}$$ where an explanation of the phase factor therein is immediately forthcoming. To show that this is possible, we first consider generic matrix elements of the form $$\begin{aligned} \label{provefactorizes} {\langle{\phi^{P_{m_1}}\cdots\phi^{P_{m_M}}}|} {{\hat {b}}}^{(m_1)} \cdots {{\hat {b}}}^{(m_M)} {|{\phi_{Q_{m_1}}\cdots\phi_{Q_{m_M}}}\rangle} = ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ {\nonumber \\}(-1)^{\beta_B^{P,Q}} \big\langle \, \big[{{\hat {\phi}}}^{P_{m_1}} \, {{\hat {b}}}^{(m_1)} \, {{\hat {\phi}}}_{Q_{m_1}}\big] \cdots \big[{{\hat {\phi}}}^{P_{m_M}} \, {{\hat {b}}}^{(m_M)} \, {{\hat {\phi}}}_{Q_{m_M}}\big] \, \big\rangle \end{aligned}$$ where ${{\hat {b}}}^{(m)}$ can denote any string of single-electron field operators (from the set $\{{{\hat {a}}}^p\}\cup\{{{\hat {c}}}_p\}$) that pertain to fragment $m$. All of the operators that need to be permuted to reach the arrangement shown on the right-hand side of [eq. (\[provefactorizes\])]{} either commute or anticommute [ (see the definitions of ${{\hat {\phi}}}_{P_m}$ and ${{\hat {\phi}}}^{P_m}$ in [eqs. (\[antisymmpdt\]) and (\[detcomplement\])]{}). ]{} The accumulated phase exponent $\beta_B^{P,Q}$ then depends on the number of electrons on each fragment in both the bra ($P$) and in the ket ($Q$), and on the number of field operators in each of the ${{\hat {b}}}^{(m)}$, which are collected into a dependency on the tuple $B$. (Actually, it only matters whether these numbers are even or odd.) Most importantly, using the anticommutation relationships, the resulting vacuum expectation value factorizes into numbers that can each be computed individually, as vacuum expectation values of the operators for each fragment alone. [ (Consider expanding each factor ${{\hat {\phi}}}^{P_m} \, {{\hat {b}}}^{(m)} \, {{\hat {\phi}}}_{Q_m}$ as a sum of strings that are normal ordered with respect to the vacuum, such that only the constant part of each survives the expectation value.) ]{} By expanding the bra and ket on the left-hand side of [eq. (\[buildHpdt\])]{} in the determinant basis [ (see [eqs. (\[general\_frag\]) and (\[general\_comp\])]{}), ]{} manipulating each of the resulting terms after the insertion of [eq. (\[decompH\])]{} as just described [ ([*i.e.*, ]{}[[eq. (\[provefactorizes\])]{}]{}), ]{} and then contracting the factors again with the expansion coefficients for the correlated fragment states, the right-hand side of [eq. (\[buildHpdt\])]{} is obtained. The final step of this sequence does require that fragment states have definite particle number, [ in order that the phase of [eq. (\[provefactorizes\])]{} is the same for each $l$ ]{} [ ($\beta_{B_l}^{I,J} = \beta_{B_l}^{P,Q}$), ]{} [ and thus factors out; ]{} however, it need not be the same number in the bra and the ket (in the case of matrix elements for charge-transfer fluctuations). According to [eq. (\[buildHpdt\])]{}, all information about the fragment states in ${\langle{\Psi^I}|} {{\hat {H}}}_M {|{\Psi_J}\rangle}$ is encapsulated in factors of the generic form $$\begin{aligned} \label{dens_mat_elem} {\langle{\psi^{i_m}}|}{{\hat {b}}}^{(m)}{|{\psi_{j_m}}\rangle} &=& \sum_{P_m,Q_m} \bar{z}_{P_m}^{i_m} z_{j_m}^{Q_m} \big\langle {{\hat {\phi}}}^{P_m} \, {{\hat {b}}}^{(m)} \, {{\hat {\phi}}}_{Q_m} \big\rangle \end{aligned}$$ [ A crucial aspect of this is that, in spite of the general biorthogonal notation, the factor $\big\langle {{\hat {\phi}}}^{P_m} \, {{\hat {b}}}^{(m)} \, {{\hat {\phi}}}_{Q_m} \big\rangle$ is simply a vacuum expectation value of a string of field operators that obey anticommutation relationships of precisely the same form as is familiar for orthonormal sets (see [eq. (\[bo\_anticomm\])]{}). ]{} [ Although the (complement) orbitals involved are not localized to fragment $m$, their “tails” on neighboring fragments to do not change the expectation value, relative to what one would obtain if the complements were computed in the absence of neighbors (for $m$ only), since the matrix of biorthogonal overlaps remains the identity. Therefore the elements defined by [eq. (\[dens\_mat\_elem\])]{} are properties of fragment $m$ alone, which we shall interpret shortly. What does change with the inclusion of neighbors is not the value of the expression in [eq. (\[dens\_mat\_elem\])]{}, but rather the corresponding integrals tensors, with which these are contracted in [eq. (\[buildHpdt\])]{}. The entirety of the complexity of demanding a globally consistent set of monomer fluctuations while working with overlapping basis functions has been moved into transformations of the one- and two-electron integrals. ]{} [ This separation of information about the internal electronic structure of fragments and their energetic interactions is a fundamental feature of the chosen biorthogonal complements to the model-space, whose peculiar suitability we foreshadowed in [section \[exchange\_concept\]]{}. ]{} [ The quantities defined by [eq. (\[dens\_mat\_elem\])]{} can be considered as elements of generalized reduced transition-density matrices ([*i.e.*, ]{}[tensors]{}) for the individual fragments. They effectively trace out much of the detail of intra-fragment correlation, such that the computational cost of subsequent steps will not depend directly on the correlation models internal to the fragments (the contraction with the integrals will depend on the basis size, however). Conventional transition-density matrices [@Davidson:1976:DensityMatrices] involve one creation and one annihilation operator. Here, there are eight kinds of tensors, resulting from different kinds of sub-strings of the one- and two-electron operators. Each transition-density tensor needs to be evaluated between every pair of fragment states whose difference in particle number matches that of the tensor type. In the case of only one field operator, there is an interesting connection to Dyson orbitals [@FetterWalecka:2003:GFBible; @Ortiz:2004:BruecknerDyson]. The most computationally expensive tensor will be the transition-density analogue of the two-electron reduced density matrix. Since these are single-fragment properties, these tensors may be pre-computed in a formally linear-scaling step (and the most expensive tensors, for four field operators, need not be stored, since they do not contribute to couplings between fragments). ]{} [ Though the complexity of dealing with antisymmetry in an overlapping basis has now been isolated to using a biorthogonal representation of the one- and two-electron integrals, it should also be pointed out that it is also not strictly necessary to ever compute the global biorthogonalized complement orbital basis. Since the real-space resolutions of the kinetic and Coulombic operators are local, only orbitals in close proximity to a group under consideration need to be projected out of each complement.]{} [ Therefore, the computation of any given matrix element of the form given in ${eq.~(\ref{x_mat_elem})}$ will involve only information that is local to the vicinity of the fragments involved. ]{} [ Given that the renormalized matrix elements may all be computed in independent calculations (making the implementation of these computations also easily parallelizable), the only relevant question as to the scaling of this step is the number of elements to be computed.]{} With the connection to the primitive one- and two-electron integrals elucidated, we may apply a familiar analysis to determine the computational complexity of obtaining the excitonic Hamiltonian. The matrix elements for spatially localized fragments will inherit the coarse features of the *ab initio* Hamiltonian in a local basis. For example, for the tetramer terms, which must involve two disjoint charge transfers, each acceptor fragment must be in the immediate vicinity of a donor fragment; however, the two donors may be quite far from each other before the overall interaction is negligible. A thorough such analysis reveals that there are a quadratically scaling number of non-negligible terms in the excitonic Hamiltonian at the mesoscopic scale (linear in the bulk limit), for any fixed error tolerance. This quadratic scaling of the interaction kernel is consistent with expectations from classical molecular mechanics. Furthermore, as with the primitive two-electron integrals, all but a linear-scaling number of matrix elements may be cast as an electrostatic interaction between charge distributions that are outside of overlap (exchange) range of each other. These can be expressed in terms of one-electron transition densities, which involve contractions of (multi-fragment) one-electron transition-density tensors with orbital-product distributions. Such densities are importantly amenable to multipole approximations in the far field. Although constructing such densities will sometimes involve orbitals on multiple fragments, there are nevertheless a linear-scaling number of them, due to the exponential decay of these densities with distance. [ It is important to distinguish these statements from claims about associated wavefunctions, which can be arbitrarily complex in a local representation, depending on the relative magnitudes of fragment-energy gaps and inter-fragment couplings. Making excitonic Hamiltonians applicable to conductors or frustrated systems (if possible) would require additional layers of formalism. ]{} We finally turn our attention to the internally correlated fragment states themselves. With respect to approximation schemes, it suffices to simply recall that the matrices ${\textbf{\textit{z}}}^{(m)}$ for each fragment were only ever required to be invertible. Any existing method in electronic structure theory that can accomplish this and supply the necessary transition-density tensors can be applied (and the fragments are not required to all use the same method). Notably, this opens the door for using the robust and efficient equation-of-motion coupled-cluster theory [@Koch:1990:LRCC; @Stanton:1993:EOMCC] for the fragment states. In practice, the intention is to only ever (implicitly or explicitly) compute a small number of columns and rows of each ${\textbf{\textit{z}}}^{(m)}$ and $\bar{{\textbf{\textit{z}}}}^{(m)}$, respectively. Conclusion ========== We have shown that it is possible to exactly write the electronic Hamiltonian for a super-system in terms of fluctuations of fragments between internally correlated states, rigorously accounting for inter-fragment electron exchange and charge transfer (also with a linearly dependent orbital basis). This has the potential to fold the vast majority of the complexity of a wide variety of electronic structure problems into the low-scaling step of obtaining an effective Hamiltonian, [ particularly for non-covalently interacting systems. ]{} The full range of familiar ground- and excited-state electronic structure methods, and their associated approximations, are readily applied to this Hamiltonian. Explicit recipes have been given for constructing all necessary matrix elements, of which there are only an asymptotically quadratically scaling number, for a given threshold. Constructing the Hamiltonian requires independent calculations on small groups of fragments ($\leq$$4$), using one- and two-electron integrals that are transformed to reflect the presence of other fragments. All of these calculations may be cast in terms of an overall linear-scaling number of reusable single-fragment tensors. Furthermore, all but a linear-scaling number of matrix elements may be efficiently decomposed in terms of electrostatic interactions between a linear-scaling number of transition densities, which are further amenable to multipole approximations in the far field. There are many features of this framework that hold promise for building finely tunable and efficient methods to capture properties of systems interacting with a large number of other systems. First, it could potentially decouple the quality of the one-electron basis from the cost of the global calculation, to the extent that the necessitated qualities and numbers of fragment states may be separate considerations. Second, since the super-system framework is theoretically independent of the level of approximation used for the fragments, it is not inherently subject to the shortcomings of any given electron-correlation model. It is therefore immediately applicable to systems where current methods perform well for the isolated fragments. This should then provide a robust approach for incorporating difficult electronic structure problems into calculations on large systems, perhaps using multi-reference methods for small molecules undergoing reactions and less expensive methods for resolving the states of peripheral molecules, and eventually giving way to force-fields or an embedding potential. Third, pressing further into unknown territory, it may be possible to parameterize or interpolate the excitonic Hamiltonian as a function of nuclear coordinates, or even generalize the framework to handle vibronic states of the fragments. The Fock-space formulation for the fragments here allows for inter-fragment charge resonance. This could even create covalent linkages between fragments, in principle. In cases where charge resonance is clearly unimportant, the excitonic Hamiltonian can be immediately truncated after dimer terms. Since all fragment fluctuations are described by orbitals local to a fragment, it will be interesting to explore the characteristics of this method with respect to basis-set superposition error, or possibly the lack thereof. Although concrete algorithms are presently envisioned to work in overall number-conserving spaces, the ability of the formalism to handle open systems could be advantageous in the future. At a technical level, it is interesting that our approach to inter-fragment exchange has forced us to work with fragment Fock spaces, even when the model space is chosen to nominally conserve fragment charges. This requirement arises from the biorthogonalization of the one-electron bases, which implicitly introduces charge-transfer components into the super-system complement basis. This is somewhat intuitive; the fundamental ambiguity in fragment location of an electron in a nonorthogonal basis is intimately related to the concept of charge transfer. Acknowledgements {#acknowledgements .unnumbered} ================ The authors gratefully acknowledge start-up support from the Hornage Fund at the University of the Pacific, as well as equipment and travel support provided by the Dean of the College of the Pacific. Helpful conversations with Christopher D. Goff and Evgeny Epifanovsky are gratefully acknowledged. [41]{} \[1\][`#1`]{} M.S. [Gordon]{}, D.G. [Fedorov]{}, S.R. [Pruitt]{} and L.V. [Slipchenko]{}, Chem. Rev. **112**, 632 (2011). R.M. [Richard]{} and J.M. [Herbert]{}, J. Chem. Phys. **137**, 064113 (2012). M.A. [Collins]{} and R.P.A. [Bettens]{}, Chem. Rev. **115**, 5607 (2015). K. [Raghavachari]{} and A. [Saha]{}, Chem. Rev. **115**, 5643 (2015). P. [Huang]{} and E.A. [Carter]{}, Annu. Rev. Phys. Chem. **59**, 261 (2008). T.A. [Wesolowski]{}, S. [Shedge]{} and X. [Zhou]{}, Chem. Rev. **115**, 5891 (2015). J. [Cui]{}, H. [Liu]{} and K.D. [Jordan]{}, J. Phys. Chem. B **110**, 18872 (2006). K.U. [Lao]{}, K.Y. [Liu]{}, R.M. [Richard]{} and J.M. [Herbert]{}, J. Chem. Phys. **144**, 164105 (2016). R. [Podeszwa]{} and K. [Szalewicz]{}, J. Chem. Phys. **126**, 194101 (2007). A. [Ambrosetti]{}, N. [Ferri]{}, R.A. [DiStasio]{}, Jr. and A. [Tkatchenko]{}, Science **351**, 1171 (2016). K.E. [Riley]{}, M. [Pitoňák]{}, P. [Jurečka]{} and P. [Hobza]{}, Chem. Rev. **110**, 5023 (2010). W. [Förner]{}, J. [Ladik]{}, P. [Otto]{} and J. [Čízek]{}, Chem. Phys. **97**, 251 (1985). H. [Stoll]{}, Phys. Rev. B **46**, 6700 (1992). S. [Saeb[ø]{}]{} and P. [Pulay]{}, Annu. Rev. Phys. Chem. **44**, 213 (1993). M. [Schütz]{} and H.J. [Werner]{}, J. Chem. Phys. **114**, 661 (2001). P.E. [Maslen]{}, A.D. [Dutoi]{}, M.S. [Lee]{}, Y. [Shao]{} and M. [Head-Gordon]{}, Mol. Phys. **103**, 425 (2005). J.E. [Subotnik]{}, A. [Sodt]{} and M. [Head-Gordon]{}, J. Chem. Phys. **125**, 074116 (2006). W. [Li]{} and P. [Piecuch]{}, J. Phys. Chem. A **114**, 8644 (2010). C. [Hättig]{}, D.P. [Tew]{} and B. [Helmich]{}, J. Chem. Phys. **136**, 204105 (2012). K. [Kristensen]{}, I.M. [H[ø]{}yvik]{}, B. [Jansik]{}, P. [J[ø]{}rgensen]{}, T. [Kj[æ]{}rgaard]{}, S. [Reine]{} and J. [Jakowski]{}, Phys. Chem. Chem. Phys. **14**, 15706 (2012). D.G. [Liakos]{}, M. [Sparta]{}, M.K. [Kesharwani]{}, J.M.L. [Martin]{} and F. [Neese]{}, J. Chem. Theory Comput. **11**, 1525 (2015). J. [Frenkel]{}, Phys. Rev. **37**, 17 (1931). A.S. [Davydov]{}, *Theory of Molecular Excitons* (Plenum, New York, 1971). V. [May]{} and O. [Kühn]{}, *Charge and Energy Transfer Dynamics in Molecular Systems* (Wiley, Weinheim, Germany, 2011). G. [Kotliar]{}, S.Y. [Savrasov]{}, K. [Haule]{}, V.S. [Oudovenko]{}, O. [Parcollet]{} and C.A. [Marianetti]{}, Rev. Mod. Phys. **78**, 865 (2006). J. [Shen]{} and S. [Li]{}, J. Chem. Phys. **131**, 174101 (2009). E. [Xu]{} and S. [Li]{}, J. Chem. Phys. **139**, 174111 (2013). S.M. [Parker]{} and T. [Shiozaki]{}, J. Chem. Phys. **141**, 211102 (2014). I. [Kim]{}, S.M. [Parker]{} and T. [Shiozaki]{}, J. Chem. Theory Comput. **11**, 3636 (2015). B. [Jeziorski]{}, R. [Moszynski]{} and K. [Szalewicz]{}, Chem. Rev. **94**, 1887 (1994). E.G. [Hohenstein]{} and C.D. [Sherrill]{}, J. Chem. Phys. **133**, 104107 (2010). A. [Sisto]{}, D.R. [Glowacki]{} and T.J. [Martinez]{}, Acc. Chem. Res. **47**, 2857 (2014). A.F. [Morrison]{} and J.M. [Herbert]{}, J. Phys. Chem. Lett. **6**, 4390 (2015). T. [Helgaker]{}, P. [Jorgensen]{} and J. [Olsen]{}, *Molecular Electronic-Structure Theory* (Wiley, Sussex, 2002). M. [Head-Gordon]{}, P.E. [Maslen]{} and C.A. [White]{}, J. Chem. Phys. **108**, 616 (1998). A. [Anderson]{}, Ann. Phys. **232**, 292 (1994). E.R. [Davidson]{}, *Reduced Density Matrices in Quantum Chemistry* (Academic Press, New York, 1976). A.L. [Fetter]{} and J.D. [Walecka]{}, *Quantum Theory of Many Particle Systems* (Dover, Mineola, New York, 2003). J.V. [Ortiz]{}, Int. J. Quantum Chem. **100**, 1131 (2004). H. [Koch]{} and P. [J[ø]{}rgensen]{}, J. Chem. Phys. **93**, 3333 (1990). J.F. [Stanton]{} and R.J. [Bartlett]{}, J. Chem. Phys. **98**, 7029 (1993). Completeness of Fluctuation Operators \[completeness\] ====================================================== The action of any operator in a space is fully determined by the collection of its matrix elements in a linearly independent basis for that space, or, equivalently, taken with respect to biorthogonal bases. Therefore, in order to show that an arbitrary ${{\hat {O}}}$ in the super-system Fock space may be fully represented in terms of the set of all fluctuation operators $\{{{\hat {\tau}}}_i^j\}$, we need only to obtain such an expression for an operator wherein each matrix element is an independent degree of freedom. This is the same as asserting that we can use this set to construct an operator that has only a single, arbitrarily chosen non-zero matrix element. Choosing the $IJ$-th element to be non-zero, we define the operator ${{\hat {o}}}_I^J$, a component of ${{\hat {O}}}$, such that $$\begin{aligned} \label{defprimop} {\langle{\Psi^{K}}|}{{\hat {o}}}_I^J{|{\Psi_{L}}\rangle} = \delta_{IK}\delta_{JL} = \prod_m \delta_{i_m,k_m}\delta_{j_m,l_m} \end{aligned}$$ This is easily accomplished by setting $$\begin{aligned} \label{resolveprimop} {{\hat {o}}}_I^J = {|{\Psi_I}\rangle}{\langle{\Psi^J}|} = \prod_m {{\hat {\tau}}}_{i_m}^{j_m} \end{aligned}$$ Since the particular choice of fragment basis plays no role in this discussion, this conclusion applies equally well to the set $\{{{\hat {\sigma}}}_P^Q\}$. [Eq. (\[resolveprimop\])]{} essentially provides a recipe for constructing an arbitrary operator, one matrix element at a time. Although any given operator may be represented as such, it does not have a unique resolution as a linear combination of *all* products of fluctuation operators. The set of all products of $N$ fluctuations (one for each fragment) is a complete and linearly independent basis for the operator space, as just demonstrated. Adding to this set all products of lengths less than $N$ therefore builds a linearly dependent set. A product of less than $N$ fluctuations has multiple non-zero matrix elements, and this is an important point as pertains to choosing compact representations of operators. We intuitively expect the excitonic Hamiltonian to couple only small numbers of fragments simultaneously, and thereby also generate many non-zero matrix elements, analogous to the relatively short strings of field operators in the *ab initio* representation. Handling Linear Dependencies \[lin\_alg\] ========================================= [ Since the framework proposed in this work depends critically on both the ability to assign one-electron functions to specific fragments and on the linear independence of that basis, it behooves us to address the linearly dependent case. This is important, as it is certain to arise when using diffuse functions for large systems. ]{} [ The general problem is that the ambiguity in fragment location of an electron is complete if the orbital it is in can be constructed as a linear combination of orbitals solely on fragments other than the one to which it is formally assigned. We must somehow “remove” this linear dependency while simultaneously preserving the local structure of the fragment framework. However, the direct removal of (nearly) null states from a one-electron basis produces orbitals that cannot generally be assigned to specific atoms or fragments. Yet, removal of individual fragment-local or atom-local functions to resolve a linear dependency can create artificial asymmetries. With some abstraction, the derivation in the main text can be applied directly to the linearly dependent case without modifying the one-electron basis. We present here an alternate framing of that derivation, ]{} which is at a level of abstraction that allows us to handle both the linearly independent and linearly dependent cases together. This allows us to show the path necessary to obtain working expressions for the linearly dependent case without explicitly rederiving all details. For any set of many-electron states $\{{|{\Psi_I}\rangle}\}$, linear dependence notwithstanding, we may write a resolution of the identity in the following form $$\begin{aligned} \label{psiidentity} 1 = \sum_I {|{\Psi_I}\rangle}{\langle{\Psi^I}|} \end{aligned}$$ for some (potentially non-unique) choice of the set $\{{|{\Psi^I}\rangle}\}$. By inserting this resolution of the identity into the time-independent Schrödinger equation solved by ${|{\Psi_\text{eigen}}\rangle}$, we arrive at the matrix eigenvalue equation $$\begin{aligned} {\textbf{\textit{H}}}{{\boldsymbol \Psi}}_\text{eigen} = E_\text{eigen} {{\boldsymbol \Psi}}_\text{eigen} \end{aligned}$$ where ${{\boldsymbol \Psi}}_\text{eigen}$ is a column matrix with elements $\Psi^I_\text{eigen} = {\langle{\Psi^I}}{|{\Psi_\text{eigen}}\rangle}$, and the matrix ${\textbf{\textit{H}}}$ has elements $H^I_J = {\langle{\Psi^I}|}{{\hat {H}}}{|{\Psi_J}\rangle}$. In the linearly dependent case, there are clearly some redundant degrees of freedom, and multiple choices of ${{\boldsymbol \Psi}}_\text{eigen}$ can be used to represent the physical state ${|{\Psi_\text{eigen}}\rangle}$; however, for a given choice of $\{{|{\Psi^I}\rangle}\}$, the matrix ${\textbf{\textit{H}}}$ is fixed. Assuming non-zero $E_\text{eigen}$, the eigenvector of ${\textbf{\textit{H}}}$ with eigenvalue of $E_\text{eigen}$ is unique. (This can be generalized in the case of degenerate states.) The physically redundant degrees of freedom are determined by the necessity that the projection of ${{\boldsymbol \Psi}}_\text{eigen}$ into the null space of ${\textbf{\textit{H}}}$ is zero. (${\textbf{\textit{H}}}$ also has spurious eigenvectors with eigenvalue zero and no physical norm, corresponding to its null space. If the physical state ${|{\Psi_\text{eigen}}\rangle}$ happened to have eigenvalue of zero, these spurious eigenvectors could mix with it. This could be easily remedied by a scalar Hamiltonian shift, but we are generally concerned here with low-energy bound states, which have manifestly negative energy eigenvalues.) In the case where $\{{|{\Psi_I}\rangle}\}$ is linearly independent, it is straightforward to show that the members of the set $\{{|{\Psi^I}\rangle}\}$ are the biorthogonal components uniquely determined by $$\begin{aligned} {\langle{\Psi^I}|} = \sum_J \bar{S}^{IJ} {\langle{\Psi_J}|} \end{aligned}$$ where the elements $\bar{S}^{IJ}$ belong to the matrix $\bar{{\textbf{\textit{S}}}}={\textbf{\textit{S}}}^{-1}$, where ${\textbf{\textit{S}}}$ is the matrix of overlaps $S_{IJ} = {\langle{\Psi_I}}{|{\Psi_J}\rangle}$. Recalling that both ${\textbf{\textit{S}}}$ and $\bar{{\textbf{\textit{S}}}}$ are self-adjoint matrices, we then straightforwardly arrive at the expected result that ${\textbf{\textit{H}}} = \bar{{\textbf{\textit{S}}}}\tilde{{\textbf{\textit{H}}}}$, where $\tilde{{\textbf{\textit{H}}}}$ has matrix elements $\tilde{H}_{IJ} = {\langle{\Psi_I}|}{{\hat {H}}}{|{\Psi_J}\rangle}$. The remainder of the work (already completed) can then be viewed as an exercise in constructing explicit forms of the members of $\{{|{\Psi^I}\rangle}\}$ in terms of the fragment basis and underlying orbitals, such that we may write ${{\hat {H}}}$ in terms of fragment fluctuations with explicit formulas for the necessary scalar coefficients (matrix elements). In the linearly dependent case, we will find it convenient to define fluctuation operators directly in the auxiliary space of coefficients. Let the set of column matrices $\{{{\boldsymbol \Psi}}_I\}$ represent the orthonormal basis vectors of this *coefficient* space, such that $$\begin{aligned} {{\boldsymbol \Psi}}_\text{eigen} &=& \sum_I {{\boldsymbol \Psi}}_I \Psi^I_\text{eigen}{\nonumber \\}&=& \sum_{i_1} \cdots \sum_{i_N} {{\boldsymbol \Psi}}_{(i_1,\cdots i_N)} \Psi^{(i_1,\cdots i_N)}_\text{eigen} \end{aligned}$$ where, in the second line, we remind ourselves of the structure of the index $I$ as a tuple that is decomposable in terms of fragment-state labels. Let us now define a set of fluctuation matrices $\{{{\boldsymbol \tau}}^j_i\}$ with the following action $$\begin{aligned} {{\boldsymbol \tau}}^{j_m}_{i_m} {{\boldsymbol \Psi}}_{(k_1,\cdots k_m, \cdots k_N)} = \delta_{j_m,k_m} {{\boldsymbol \Psi}}_{(k_1,\cdots i_m, \cdots k_N)} \end{aligned}$$ where, as with the physical states, these fluctuation matrices may be defined as superpositions of appropriate dyads. Similarly, the proof that any matrix in the coefficient space may be written as a sum of products of fluctuation matrices proceeds along the same lines as in [appendix \[completeness\]]{}, assuring us that ${\textbf{\textit{H}}}$ may be built from such fluctuations. The precise fragment-wise structure of this expansion of ${\textbf{\textit{H}}}$ (and therefore ${{\boldsymbol \Psi}}_\text{eigen}$) again depends on determining the the elements ${\langle{\Psi^I}|} {{\hat {H}}} {|{\Psi_J}\rangle}$. The primary difference, relative to the linearly independent case, is that the choice of $\{{|{\Psi^I}\rangle}\}$ is not unique, and the null space of ${\textbf{\textit{H}}}$ depends on this choice, consequently affecting its fluctuation-matrix expansion. In spite of any linear dependencies, [eq. (\[superbasis\])]{} still provides the definition of the antisymmetrized tensor-products of correlated fragment states in terms of the (potentially linearly dependent) set of orbital configurations. Therefore, in the linearly dependent case, [eq. (\[supercompl\])]{} is still a suitable *choice* of the members of $\{{|{\Psi^I}\rangle}\}$ that satisfy [eq. (\[psiidentity\])]{}, in terms of a suitable *choice* of $\{{|{\Phi^P}\rangle}\}$ that satisfy $$\begin{aligned} \label{phiidentity} 1 = \sum_P {|{\Phi_P}\rangle}{\langle{\Phi^P}|} \end{aligned}$$ which is quickly verified by inserting these resolutions of ${|{\Psi_I}\rangle}$ and ${|{\Psi^I}\rangle}$ into [eq. (\[psiidentity\])]{}, and realizing that the invertibility of ${\textbf{\textit{Z}}}$ is independent of linear dependencies in the set of physical states. The important consequence of this is that, once formulas for the elements of a matrix ${\textbf{\textit{H}}}'$ (with elements $H^{\prime P}_Q = {\langle{\Phi^P}|}{{\hat {H}}}{|{\Phi_Q}\rangle}$) are known for a given choice of $\{{|{\Phi^P}\rangle}\}$, and their structure in terms of numbers of fragments coupled is analyzed, then the path to constructing ${\textbf{\textit{H}}}$ in terms of fluctuations $\{{{\boldsymbol \tau}}^j_i\}$ via application of fragment-local transformations is the same as presented in the main text for the linearly independent case. As with the linearly independent case, the construction of explicit forms for the members of $\{{|{\Phi^P}\rangle}\}$ can be straightforward, so long as the matrix ${\textbf{\textit{s}}}$, built from the one-electron overlaps $s_{pq} = {\langle{\chi_p}}{|{\chi_q}\rangle}$, can be partially inverted. Let the matrix $\bar{{\textbf{\textit{s}}}}$ satisfy $$\begin{aligned} {\textbf{\textit{p}}}_1 &=& {\textbf{\textit{p}}}_1\bar{{\textbf{\textit{s}}}}{\textbf{\textit{s}}} = {\textbf{\textit{s}}}\bar{{\textbf{\textit{s}}}}{\textbf{\textit{p}}}_1 {\nonumber \\}1 &=& {\textbf{\textit{p}}}_0 + {\textbf{\textit{p}}}_1 \end{aligned}$$ where ${\textbf{\textit{p}}}_0$ is the orthogonal projector onto the null space of ${\textbf{\textit{s}}}$, and ${\textbf{\textit{p}}}_1$ is the orthogonal projector onto the range (null space complement) of ${\textbf{\textit{s}}}$. Notably, only the projection of $\bar{{\textbf{\textit{s}}}}$ onto the range of ${\textbf{\textit{s}}}$ is uniquely defined. Because of this, $\bar{{\textbf{\textit{s}}}}$ is not necessarily self adjoint for any choice, but we do have ${\textbf{\textit{p}}}_1\bar{{\textbf{\textit{s}}}}{\textbf{\textit{p}}}_1 = ({\textbf{\textit{p}}}_1\bar{{\textbf{\textit{s}}}}{\textbf{\textit{p}}}_1)^\dagger$, which is equally valuable. For a given choice of $\bar{{\textbf{\textit{s}}}}$, a choice of the members of $\{{|{\Phi^P}\rangle}\}$ may be specified again by [eq. (\[detcomplement\])]{}, but using a linearly dependent set of orbitals that satisfy $$\begin{aligned} {\langle{\chi^p}|} = \sum_q \bar{s}^{pq} {\langle{\chi_q}|} \end{aligned}$$ where $\bar{s}^{pq}$ is an element of $\bar{{\textbf{\textit{s}}}}$. We may, as before, define ${{\hat {c}}}_p$ and ${{\hat {c}}}^p$ as the creation operators associated with ${|{\chi_p}\rangle}$ and ${|{\chi^p}\rangle}$, respectively, and define ${{\hat {a}}}_p$ and ${{\hat {a}}}^p$ as their respective Hermitian conjugates. The Hamiltonian again takes exactly the same form as in [eq. (\[field\_opH\])]{} in terms of these operators. (This can be shown by elimination of transformations that express the linearly dependent orbitals, their complements, and all associated field operators, in terms of some hypothetical orthonormal basis.) The difference is that there is now redundancy in the operator set, and the anticommutation rules do not follow. A given string in the field-operator resolution of the Hamiltonian will act not only on the fragments to which the indices in that string belong, but also on any fragments whose orbitals can be linearly combined to build the orbitals in question. If a given orbital is linearly independent of the rest of the basis (likely the majority), then only expected fragment is involved. Regardless, this does not increase the fragment order of the Hamiltonian, but it increases the number of fragments that could be associated with a given orbital index in the *ab initio* Hamiltonian, which is an intuitive consequence of having a linear dependency arise due to overlaps of diffuse orbitals on different fragments. A final comment is worthwhile to connect back to the abstract framework of resolutions of the identity. Similar to what we have done in the one-electron space, the matrices $\bar{{\textbf{\textit{S}}}}$ and $\bar{{\textbf{\textit{S}}}}'$ may be defined as satisfying $$\begin{aligned} {\textbf{\textit{P}}}_1\bar{{\textbf{\textit{S}}}}{\textbf{\textit{S}}} ~=~ {\textbf{\textit{S}}}\bar{{\textbf{\textit{S}}}}{\textbf{\textit{P}}}_1 &=& {\textbf{\textit{P}}}_1 ~=~ 1-{\textbf{\textit{P}}}_0 {\nonumber \\}{\textbf{\textit{P}}}'_1\bar{{\textbf{\textit{S}}}}'{\textbf{\textit{S}}}' ~=~ {\textbf{\textit{S}}}'\bar{{\textbf{\textit{S}}}}'{\textbf{\textit{P}}}'_1 &=& {\textbf{\textit{P}}}'_1 ~=~ 1-{\textbf{\textit{P}}}'_0 \end{aligned}$$ where ${\textbf{\textit{S}}}'$ has elements $S'_{PQ} = {\langle{\Phi_P}}{|{\Phi_Q}\rangle}$, and ${\textbf{\textit{P}}}_0$ and ${\textbf{\textit{P}}}'_0$ are the orthogonal projectors into the null spaces of ${\textbf{\textit{S}}}$ and ${\textbf{\textit{S}}}'$, respectively. For a given choice of $\bar{{\textbf{\textit{S}}}}$ and $\bar{{\textbf{\textit{S}}}}'$, the members of the sets $\{{|{\Psi^I}\rangle}\}$ and $\{{|{\Phi^P}\rangle}\}$ satisfy $$\begin{aligned} {\langle{\Psi^I}|} &=& \sum_J \bar{S}^{IJ} {\langle{\Psi_J}|} {\nonumber \\}{\langle{\Phi^P}|} &=& \sum_Q \bar{S}^{\prime PQ} {\langle{\Phi_Q}|} \end{aligned}$$ which coincides with the unique choice of biorthogonal complements in the linearly independent case, when $\bar{{\textbf{\textit{S}}}}={\textbf{\textit{S}}}^{-1}$ and $\bar{{\textbf{\textit{S}}}}'={\textbf{\textit{S}}}'^{-1}$. In the linearly dependent case, the use of these definitions in resolutions of the identity \[[eqs. (\[psiidentity\]) and (\[phiidentity\])]{}\] corresponds to additions of, and projections onto, null vectors. If we make the choice that the arbitrary part of $\bar{{\textbf{\textit{s}}}}$ is zero, or equivalently that the arbitrary part of $\bar{{\textbf{\textit{S}}}}'$ is zero, which is likely the most convenient and practical choice, and we furthermore insist on [eq. (\[supercompl\])]{} for our choice of the set $\{{|{\Psi^I}\rangle}\}$, then this is equivalent to choosing $$\begin{aligned} \bar{{\textbf{\textit{S}}}} = \bar{{\textbf{\textit{Z}}}}{\textbf{\textit{P}}}'_1\bar{{\textbf{\textit{S}}}}'{\textbf{\textit{P}}}'_1\bar{{\textbf{\textit{Z}}}}^\dagger \end{aligned}$$ This is clearly a valid pseudo-inverse of ${\textbf{\textit{S}}}={\textbf{\textit{Z}}}^\dagger{\textbf{\textit{S}}}'{\textbf{\textit{Z}}}$, but it has non-zero components in the arbitrary part, due the fact that some eigenvectors of ${\textbf{\textit{S}}}'$ with non-zero eigenvalue transform to vectors that lie partly in the null space of ${\textbf{\textit{S}}}$, and vice versa. The result of this is that the Hamiltonian matrix is $$\begin{aligned} {\textbf{\textit{H}}} = \bar{{\textbf{\textit{Z}}}}{\textbf{\textit{H}}}'{\textbf{\textit{Z}}} \end{aligned}$$ which is a similarity transformation of the Hamiltonian in the configuration basis, as expected, but where $$\begin{aligned} {\textbf{\textit{P}}}_0{\textbf{\textit{H}}} \neq 0 \end{aligned}$$ meaning that ${\textbf{\textit{H}}}$ does not have the same null space as ${\textbf{\textit{S}}}$. It still has a null space of the same dimension, but it corresponds to a specific, non-intuitive (but implicit and innocuous) choice of representation of the state vectors in the basis of tensor products of correlated states.

ArXiv

--- abstract: 'AM CVn systems are interacting binaries similar to cataclysmic variables (CVs), but more compact with orbital periods of less than 80 minutes. The primary is a white dwarf, whereas the nature of the secondary is not completely clear, yet. Abundances and composition of the outer layer of the secondary can be found by analysis of the accretion disk (presented by Nagel et al. these proceedings). Spectra from high-state AM CVn systems do not only show typical signatures of accretion disks, but also P Cygni line profiles, a sign of outflow being present in the system. Here we present the first quantitative spectral analysis of an accretion-disk wind in AM CVn systems. Emergent wind spectra are modeled with our 3-D Monte Carlo radiative transfer code WOMPAT. We show that P Cygni profiles can be reproduced with our wind models.' author: - 'D.-J. Kusterer, T. Nagel, and K. Werner' title: 'Accretion Disk Winds in AM CVn Binaries – a Monte Carlo Approach' --- The first detection of outflows in CVs dates back to the late 1970ies and early 1980ies when blueshifted absorption troughs and P Cygni profiles in ultraviolet (UV) resonance lines were discovered [@DJK1978Natur.275..385H; @DJK1982ApJ...260..716C]. More recent work shows that outflowing material has a strong influence on observations, not only in the “classical” UV wind lines, but also in other features. Not only in CVs, but also in their “smaller twins” the AM CVn systems, signatures of biconical outflow are found. As in CVs such an accretion-disk wind is mainly found in high-state systems, a complete model of such systems therefore has to include an outflow. Within the Tübingen group a code for modeling NLTE accretion-disk atmospheres was developed and successfully used to model spectra of CVs, AM CVn systems and ultracompact X-ray binaries . Our goal is to develop a whole package with which we are able to model a complete CV including the accretion disk, the white dwarf and the outflow. Here our first results with disk, a blackbody WD and wind for the prototype AM CVn are presented. We implemented a kinematical biconical wind model by [@DJK1993ApJ...409..372S]. Monte Carlo (MC) techniques are used to do the radiative transfer in this 3-D wind structure. Photon packets, which represent a monochromatic family of photons, are created and then followed through the wind. All parts of a photon packet’s life, like the creation, interactions, new directions of flight are determined via probabilities. Optical depths are acquired by numerical integration of local opacities along the photon’s line of flight. No Sobolev approximation is needed. Furthermore line opacities can be calculated with either Doppler or Stark broadening. Typical P Cygni wind lines, such as the C[iv]{} 1550 Å resonance line (Fig. \[DJKspec\]) are reproduced by our model. Despite the obvious difference in chemical composition and size of the accretion disks in AM CVn and normal, hydrogen rich, CVs, the wind lines in both cases are quite similar. This means that the dependence of the wind mechanism on these parameters has to be relatively small. Figure \[DJKspec\] shows a STIS spectrum of AM CVn itself, which shows several weak P Cygni features. Also included in Fig. \[DJKspec\] is a spectrum calculated with WOMPAT for an AM CVn parameter set with $M_{\mathrm{WD}} = 0.6\, \mathrm{M}_{\odot}$ and an inclination angle of $40^{\circ}$. No parameter space exploration has been done for this, therefore it is just to be taken as a very rough guidance for future progress. As a source of radiation for the photon packets a blackbody white dwarf with $T_{ \mathrm{eff}} = 18,000\, \mathrm{K}$ and the accretion disc were used. The accretion disk input consisted of twelve rings calculated with [AcDc]{} and further blackbody rings to fill up the gap to the outer edge of the disk. The combined theoretical spectrum of these twelve [AcDc]{} rings is shown in Fig. \[DJKspec\], as well. It is clearly seen that the disk alone cannot produce a C[iv]{} 1550 Å P Cygni shape.\ ![STIS spectrum of AM CVn overlaid with a theoretical disc spectrum calculated with AcDc and a Monte Carlo accretion disk wind spectrum calculated with our new code WOMPAT.[]{data-label="DJKspec"}](djkusterer1.eps){width="95.00000%"} This work is supported by DFG grant We 1312/37-1. , F. A., & [Mason]{}, K. O. 1982, ApJ, 260, 716 , S. R., Boggess, A., Holm, A., [et al.]{} 1978, Nature, 275, 385 , T., [Dreizler]{}, S., [Rauch]{}, T., & [Werner]{}, K. 2004, A&A, 428, 109 , I., & [Vitello]{}, P. 1993, ApJ, 409, 372 , K., [Nagel]{}, T., [Rauch]{}, T., [Hammer]{}, N. J., & [Dreizler]{}, S. 2004, A&A, 428, 109

ArXiv

--- abstract: | The *approximate degree* of a Boolean function $f(x_{1},x_{2},\ldots,x_{n})$ is the minimum degree of a real polynomial that approximates $f$ pointwise within $1/3$. Upper bounds on approximate degree have a variety of applications in learning theory, differential privacy, and algorithm design in general. Nearly all known upper bounds on approximate degree arise in an existential manner from bounds on quantum query complexity. We develop a first-principles, classical approach to the polynomial approximation of Boolean functions. We use it to give the first constructive upper bounds on the approximate degree of several fundamental problems: 1. $O\bigl(n^{\frac{3}{4}-\frac{1}{4(2^{k}-1)}}\bigr)$ for the $k$-element distinctness problem; 2. $O(n^{1-\frac{1}{k+1}})$ for the $k$-subset sum problem; 3. $O(n^{1-\frac{1}{k+1}})$ for any $k$-DNF or $k$-CNF formula; 4. $O(n^{3/4})$ for the surjectivity problem. In all cases, we obtain explicit, closed-form approximating polynomials that are unrelated to the quantum arguments from previous work. Our first three results match the bounds from quantum query complexity. Our fourth result improves polynomially on the $\Theta(n)$ quantum query complexity of the problem and refutes the conjecture by several experts that surjectivity has approximate degree $\Omega(n)$. In particular, we exhibit the first *natural* problem with a polynomial gap between approximate degree and quantum query complexity. author: - 'Alexander A. Sherstov' bibliography: - 'refs.bib' title: Algorithmic Polynomials --- [^1] plus 3pt minus 3pt plus 3pt minus 3pt Introduction ============ Let $f\colon X\to\zoo$ be a given Boolean function, defined on a subset $X\subseteq\zoo^{n}.$ The *$\epsilon$-approximate degree of $f$*, denoted $\deg_{\epsilon}(f),$ is the minimum degree of a multivariate real polynomial $p$ such that $|f(x)-p(x)|\le\epsilon$ for all $x\in X.$ The standard setting of the error parameter for most applications is $\epsilon=1/3$, an aesthetically motivated constant that can be replaced by any other in $(0,1/2)$ at the expense of a constant-factor increase in approximate degree. The notion of approximate degree originated 25 years ago in the pioneering work of Nisan and Szegedy [@nisan-szegedy94degree] and has since proved to be a powerful and versatile tool in theoretical computer science. Lower bounds on approximate degree have complexity-theoretic applications, whereas upper bounds are a tool in algorithm design. In the former category, the notion of approximate degree has enabled spectacular progress in circuit complexity [@paturi-saks94rational; @siu-roy-kailath94rational; @beigel91rational; @aspnes91voting; @krause94depth2mod; @KP98threshold; @sherstov07ac-majmaj; @beame-huyn-ngoc09multiparty-focs], quantum query complexity [@beals-et-al01quantum-by-polynomials; @buhrman-et-al99small-error; @aaronson-shi04distinctness; @aaronson04sdpt-for-search; @ambainis05collision; @klauck07product-thms; @BKT17poly-strikes-back], and communication complexity [@buhrman-dewolf01polynomials; @razborov02quantum; @buhrman07pp-upp; @sherstov07ac-majmaj; @sherstov07quantum; @RS07dc-dnf; @lee-shraibman08disjointness; @chatt-ada08disjointness; @dual-survey; @beame-huyn-ngoc09multiparty-focs; @sherstov12mdisj; @sherstov13directional]. On the algorithmic side, approximate degree underlies many of the strongest results obtained to date in computational learning [@tt99DNF-incl-excl; @KS01dnf; @KOS:02; @KKMS; @odonnell03degree; @ACRSZ07nand], differentially private data release [@tuv12releasing-marginals; @ctuw14release-of-marginals], and algorithm design in general [@linial-nisan90incl-excl; @kahn96incl-excl; @sherstov07inclexcl-ccc]. Despite these applications, progress in understanding approximate degree as a complexity measure has been slow and difficult. With very few exceptions [@nisan-szegedy94degree; @kahn96incl-excl; @sherstov07inclexcl-ccc; @sherstov12noisy], all known upper bounds on approximate degree arise from *quantum query algorithms*. The connection between approximate degree and quantum query complexity was discovered by Beals et al. [@beals-et-al01quantum-by-polynomials], who proved that the acceptance probability of an algorithm that makes $T$ queries is representable by a real polynomial of degree $2T$. Put another way, every quantum algorithm implies an approximating polynomial of comparable complexity for the problem in question. Since the seminal work of Beals et al., essentially all upper bounds on approximate degree have come from quantum query algorithms, e.g., [@buhrman-et-al99small-error; @de-wolf08approx-degree; @ambainis04k-distinctness; @FGG08nand; @ACRSZ07nand; @drucker-de-wolf11uniform-approx-by-quantum-algorithms; @drucker-dewolf11quantum-method; @belovs12distinctness; @mahadev-de-wolf15rational-approx-postselection]. An illustrative example is the problem of determining the approximate degree of Boolean formulas of size $n,$ posed in 2003 by O’Donnell and Servedio [@odonnell03degree]. Progress on this question was stalled for a long time until it was finally resolved by Ambainis et al. [@ACRSZ07nand], who built on the work of Farhi et al. [@FGG08nand] to give a near-optimal quantum query algorithm for any Boolean formula. While quantum query complexity has been a fruitful source of approximate degree upper bounds, the exclusive reliance on quantum techniques for the polynomial approximation of Boolean functions is problematic. For one thing, a quantum query algorithm generally does not give any information about the approximating polynomial apart from its existence. For example, converting the quantum algorithms of [@ambainis04k-distinctness; @ACRSZ07nand; @belovs12distinctness] to polynomials results in expressions so large and complicated that they are no longer meaningful. More importantly, quantum query algorithms are more constrained objects than real polynomials, and an optimal query algorithm for a given problem may be far less efficient than a polynomial constructed from scratch. Given the many unresolved questions on approximate degree, there is a compelling need for polynomial approximation techniques that go beyond quantum query complexity. In this paper, we take a fresh look at several breakthrough upper bounds for approximate degree, obtained over the years by sophisticated quantum query algorithms. In each case, we are able to construct an approximating polynomial from first principles that matches or improves on the complexity of the best quantum algorithm. All of our constructions produce explicit, closed-form polynomials that are unrelated to the corresponding quantum algorithms and are in the author’s opinion substantially simpler. In one notable instance, our construction achieves a polynomial improvement on the complexity of the best possible quantum algorithm, refuting a conjecture [@bun-thaler17adeg-ac0] on the approximate degree of that problem and exhibiting the first *natural* example of a polynomial gap between approximate degree and quantum query complexity. Our proofs, discussed shortly, contribute novel techniques to the area. *k*-Element distinctness ------------------------ The starting point in our work is the *element distinctness problem* [@bdhhmsw01element-distinctness; @aaronson-shi04distinctness; @ambainis04k-distinctness; @ambainis05collision; @kutin05collision; @belovs12distinctness], which is one of the most studied questions in quantum query complexity and a major success story of the field. The input to the problem is a list of $n$ elements from a given range of size $r,$ and the objective is to determine if the elements are pairwise distinct. A well-studied generalization of this problem is *$k$-element distinctness*, where $k$ is an arbitrary constant and the objective is to determine if some $k$-tuple of the elements are identical. Formally, the input to element distinctness and $k$-element distinctness is represented by a Boolean matrix $x\in\zoo^{n\times r}$ in which every row $i$ has precisely one “$1$” entry, corresponding to the value of the $i$th element.[^2] Aaronson and Shi [@aaronson-shi04distinctness], Ambainis [@ambainis05collision], and Kutin [@kutin05collision] showed that element distinctness has quantum query complexity $\Omega(n^{2/3}).$ In follow-up work, Ambainis [@ambainis04k-distinctness] gave a quantum algorithm for element distinctness with $O(n^{2/3})$ queries, matching the lower bound in [@aaronson-shi04distinctness; @ambainis05collision; @kutin05collision]. For the more general problem of $k$-element distinctness, Ambainis’s algorithm [@ambainis04k-distinctness] requires $O(n^{k/(k+1)})$ queries. Using a different approach, Belovs [@belovs12distinctness] gave a polynomially faster algorithm for $k$-element distinctness, with query complexity $O(n^{\frac{3}{4}-\frac{1}{4(2^{k}-1)}})$. Belovs’s algorithm is currently the fastest known. The algorithms of Ambainis [@ambainis04k-distinctness] and Belovs [@belovs12distinctness] are highly nontrivial. The former is based on a quantum walk on the Johnson graph, whereas the latter uses the framework of learning graphs. We give an elementary, closed-form construction of an approximating polynomial for $k$-element distinctness that bypasses the quantum work. Formally, let $\operatorname{ED}_{n,r,k}\colon\zoo_{\leq n}^{n\times r}\to\zoo$ be given by $$\operatorname{ED}_{n,r,k}(x)=\begin{cases} 1 & \text{if }x_{1,j}+x_{2,j}+\cdots+x_{n,j}<k\text{ for each \ensuremath{j,}}\\ 0 & \text{otherwise.} \end{cases}$$ The notation $\zoo_{\leq n}^{n\times r}$ for the domain of this function indicates that we allow *arbitrary* input matrices $x\in\zoo^{n\times r}$ of Hamming weight at most $n$, with no restriction on the placement of the “1” bits. This is of course a problem more general than $k$-element distinctness. We prove: \[thm:MAIN-ed\]Let $k\geq1$ be a fixed integer. Then for all $n,r\geq1,$ $$\begin{aligned} \deg_{1/3}(\operatorname{ED}_{n,r,k}) & =O\left(\sqrt{n}\min\{n,r\}^{\frac{1}{2}-\frac{1}{4(1-2^{-k})}}\right).\end{aligned}$$ Moreover, the approximating polynomial is given explicitly in each case. Theorem \[thm:MAIN-ed\] matches the quantum query bound of $O(n^{\frac{3}{4}-\frac{1}{4(2^{k}-1)}})\equiv O(n^{1-\frac{1}{4(1-2^{-k})}})$ due to Belovs [@belovs12distinctness] and further generalizes it to every $r\geq1.$ *k*-Subset sum, *k*-DNF and *k*-CNF formulas -------------------------------------------- Another well-studied problem in quantum query complexity is *$k$-subset sum* [@childs-eisenberg05subset-sum; @belovs-spalek13k-sum]. The input to this problem is a list of $n$ elements from a given finite Abelian group $G,$ and the objective is to determine whether there is a $k$-tuple of elements that sum to $0.$ Formally, the input is represented by a matrix $x\in\zoo^{n\times|G|}$ with precisely one “1” entry in every row. Childs and Eisenberg [@childs-eisenberg05subset-sum] contributed an alternate analysis of Ambainis’s algorithm for $k$-element distinctness [@ambainis04k-distinctness] and showed how to adapt it to compute $k$-subset sum or any other function property with $1$-certificate complexity at most $k.$ In particular, any such problem has an approximating polynomial of degree $O(n^{k/(k+1)}).$ We give a first-principles construction of an approximating polynomial for any problem in this class, using techniques that are elementary and unrelated to the quantum work of Ambainis [@ambainis04k-distinctness] and Childs and Eisenberg [@childs-eisenberg05subset-sum]. Our result is more general: \[thm:MAIN-k-dnf\]Let $k\geq0$ be a fixed integer. Let $f\colon\zoo_{\leq n}^{N}\to\zoo$ be representable on its domain by a $k$-DNF or $k$-CNF formula. Then $$\deg_{1/3}(f)=O(n^{\frac{k}{k+1}}).$$ Moreover, the approximating polynomial is given explicitly in each case. Recall that a $k$-DNF formula in Boolean variables $x_{1},x_{2},\ldots,x_{N}$ is the disjunction of an arbitrary number of terms, where each term is the conjunction of at most $k$ literals from among $x_{1},\overline{x_{1}},x_{2},\overline{x_{2}},\ldots,x_{N},\overline{x_{N}}.$ An essential aspect of Theorem \[thm:MAIN-k-dnf\] is that the approximate degree upper bound depends only on the Hamming weight $x_{1}+x_{2}+\cdots+x_{N}$ of the input and does not depend at all on the number of variables $N$, which can be arbitrarily large. Several special cases of Theorem \[thm:MAIN-k-dnf\] are worth noting. The theorem clearly applies to $k$-subset sum, which is by definition representable on its domain by a $k$-DNF formula. Moreover, in the terminology of Childs and Eisenberg [@childs-eisenberg05subset-sum], Theorem \[thm:MAIN-k-dnf\] applies to any function property with $1$-certificate complexity at most $k$. Finally, taking $N=n$ shows that Theorem \[thm:MAIN-k-dnf\] applies to any function $f\colon\zoon\to\zoo$ representable by a $k$-DNF or $k$-CNF formula. Surjectivity ------------ While our proofs of Theorems \[thm:MAIN-ed\] and \[thm:MAIN-k-dnf\] are significantly simpler than their quantum query counterparts, they do not give a quantitative improvement on previous work. This brings us to our next result. In the *surjectivity problem *[@beame-machmouchi12quantum-query-ac0], the input is a list of $n$ elements from a given range of size $r,$ where $r\leq n.$ The objective is to determine whether the input features all $r$ elements of the range. In function terminology, the input represents a mapping $\{1,2,\ldots,n\}\to\{1,2,\ldots,r\},$ and the objective is to determine whether the mapping is surjective. As usual in the quantum query literature, the input is represented by a Boolean matrix $x\in\zoo^{n\times r}$ in which every row has precisely one “$1$” entry. Beame and Machmouchi [@beame-machmouchi12quantum-query-ac0] proved that for $r=\lfloor n/2\rfloor+1,$ the surjectivity problem has the maximum possible quantum query complexity, namely, $\Theta(n).$ This led several experts to conjecture that the approximate degree of surjectivity is also $\Theta(n)$; see, e.g., [@bun-thaler17adeg-ac0]. The conjecture was significant because its resolution would give the first $\mathsf{AC}^{0}$ circuit with approximate degree $\Theta(n),$ closing a long line of research [@nisan-szegedy94degree; @aaronson-shi04distinctness; @ambainis05collision; @bun-thaler17adeg-ac0]. Surprisingly, we are able to show that surjectivity has an approximating polynomial of substantially lower degree, regardless of the range parameter $r$. Formally, let ${\operatorname{SURJ}}_{n,r}\colon\zoo_{\leq n}^{n\times r}\to\zoo$ be given by $${\operatorname{SURJ}}_{n,r}(x)=\bigwedge_{j=1}^{r}\bigvee_{i=1}^{n}x_{i,j}.$$ In keeping with our other results, our definition of ${\operatorname{SURJ}}_{n,r}$ allows arbitrary input matrices $\zoo^{n\times r}$ of Hamming weight at most $n.$ In this generalization of the surjectivity problem, the input can be thought of as an arbitrary relation rather than a function. We prove: \[thm:MAIN-surj\]For all positive integers $n$ and $r,$ $$\deg_{1/3}({\operatorname{SURJ}}_{n,r})=\begin{cases} O(\sqrt{n}\cdot r^{1/4}) & \text{if }r\leq n,\\ 0 & \text{if \ensuremath{r>n.}} \end{cases}$$ Moreover, the approximating polynomial is given explicitly in each case. In particular, the theorem gives an approximating polynomial of degree $O(n^{3/4})$ for all $r.$ This upper bound is polynomially smaller than the problem’s quantum query complexity $\Theta(n)$ for $r=\lfloor n/2\rfloor+1.$ While explicit functions with a polynomial gap between approximate degree and quantum query complexity have long been known [@ambainis06adeg-vs-query; @ABK16cheat-sheets], Theorem \[thm:MAIN-surj\] exhibits the first *natural* function with this property. The functions in previous work [@ambainis06adeg-vs-query; @ABK16cheat-sheets] were constructed with the specific purpose of separating complexity measures. Symmetric functions ------------------- Key building blocks in our proofs are symmetric functions $f\colon\zoon\to\zoo.$ A classic result due to Paturi [@paturi92approx] states that the $1/3$-approximate degree of any such function $f$ is $\Theta(\sqrt{n\ell}),$ where $\ell\in\{0,1,2,\ldots,n\}$ is the smallest number such that $f$ is constant on inputs of Hamming weight in $[\ell,n-\ell].$ When a symmetric function is used in an auxiliary role as part of a larger construction, it becomes important to have approximating polynomials for every possible setting of the error parameter, $1/2^{n}\leq\epsilon\leq1/3$. A complete characterization of the $\epsilon$-approximate degree of symmetric functions for all $\epsilon$ was obtained by de Wolf [@de-wolf08approx-degree], who sharpened previous bounds [@kahn96incl-excl; @buhrman-et-al99small-error; @sherstov07inclexcl-ccc] using an elegant quantum query algorithm. Prior to our work, no classical, first-principles proof was known for de Wolf’s characterization, which is telling in view of the basic role that $\AND_{n},\OR_{n},$ and other symmetric functions play in the area. We are able to give such a first-principles proofin fact, *three* of them. \[thm:MAIN-symmetric\]Let $f\colon\zoon\to\{0,1\}$ be a symmetric function. Let $\ell\in\{0,1,2,\ldots,n\}$ be an integer such that $f$ is constant on inputs of Hamming weight in $(\ell,n-\ell).$ Then for $1/2^{n}\leq\epsilon\leq1/3,$ $$\deg_{\epsilon}(f)=O\left(\sqrt{n\ell}+\sqrt{n\log\frac{1}{\epsilon}}\right).$$ Moreover, the approximating polynomial is given explicitly in each case. Theorem \[thm:MAIN-symmetric\] matches de Wolf’s quantum query result, tightly characterizing the $\epsilon$-approximate degree of every nonconstant symmetric function. Our techniques -------------- Our proofs use only basic tools from approximation theory, such as Chebyshev polynomials. Our constructions additionally incorporate elements of classic algorithm design, e.g., the divide-and-conquer paradigm, the inclusion-exclusion principle, and probabilistic reasoning. The title of our paper, “Algorithmic Polynomials,” is a reference to this combination of classic algorithmic methodology and approximation theory. The informal message of our work is that algorithmic polynomials are not only more powerful than quantum algorithms but also easier to construct. A detailed discussion of Theorems \[thm:MAIN-ed\]\[thm:MAIN-symmetric\] follows. ### Extension theorem. {#extension-theorem. .unnumbered} As our starting point, we prove an *extension theorem* for polynomial approximation. This theorem allows one to construct an approximant for a given function $F$ using an approximant for a restriction $f$ of $F.$ In more detail, let $f\colon\zoo_{\leq m}^{N}\to[-1,1]$ be an arbitrary function, defined on inputs $x\in\zoo^{N}$ of Hamming weight at most $m.$ Let $F_{n}\colon\zoo_{\leq n}^{N}\to[-1,1]$ be the natural extension of $f$ to inputs of Hamming weight at most $n,$ defined by $F_{n}=0$ outside the domain of $f.$ From an approximation-theoretic point of view, a fundamental question to ask is how to efficiently “extend” any approximant for $f$ to an approximant for $F_{n}.$ Unfortunately, this naïve formulation of the extension problem has no efficient solution; we describe a counterexample in Section \[sec:extension\]. We are able to show, however, that the extension problem becomes meaningful if one works with $F_{2m}$ instead of $f$. In other words, we give an efficient, explicit, black-box transformation of any approximant for the extension $F_{2m}$ into an approximant for the extension $F_{n}$, for any $n\geq2m$. This result is essentially as satisfying as the “ideal” extension theorem in that the domains of $f$ and $F_{2m}$ almost coincide and can be arbitrarily smaller than the domain of $F_{n}$. Our proof makes use of extrapolation bounds, extremal properties of Chebyshev polynomials, and ideas from rational approximation theory. ### Symmetric functions. {#symmetric-functions. .unnumbered} As mentioned earlier, we give three proofs of Theorem \[thm:MAIN-symmetric\] on the $\epsilon$-approximate degree of symmetric functions. Each of the three proofs is fully constructive. Our simplest proof uses the extension theorem and is only half-a-page long. Here, we use brute-force interpolation to compute the function $f$ of interest on inputs of small Hamming weight, and then apply the extension theorem to effortlessly extend the interpolant to the full domain of $f.$ Our second proof of Theorem \[thm:MAIN-symmetric\] is an explicit, closed-form construction that uses Chebyshev polynomials as its only ingredient. This proof is a refinement of previous, suboptimal approximants for the AND function [@kahn96incl-excl; @sherstov07inclexcl-ccc]. We eliminate the inefficiency in previous work by using Chebyshev polynomials to achieve improved control at every point of the domain. Finally, our third proof of Theorem \[thm:MAIN-symmetric\] is inspired by combinatorics rather than approximation theory. Here, we use a *sampling experiment* to construct an approximating polynomial for any symmetric function $f$ from an approximating polynomial for AND. In more detail, the experiment allows us to interpret $f$ as a linear combination of conjunctions of arbitrary degree, where the sum of the absolute values of the coefficients is reasonably small. Once such a representation is available, we simply replace every conjunction with its approximating polynomial. These substitutions increase the error of the approximation by a factor bounded by the sum of the absolute values of the coefficients in the original linear combination, which is negligible. ### k-Element distinctness, k-DNF and k-CNF formulas. ** {#k-element-distinctness-k-dnf-and-k-cnf-formulas. .unnumbered} We first establish an auxiliary result on the approximate degree of composed Boolean functions. Specifically, let $F\colon X\times\zoo_{\leq n}^{N}\to\zoo$ be given by $F(x,y)=\bigvee_{i=1}^{N}y_{i}\wedge f_{i}(x)$ for some set $X$ and some functions $f_{1},f_{2},\ldots,f_{N}\colon X\to\zoo.$ We bound the $\epsilon$-approximate degree of $F$ in terms of the approximate degree of $\bigvee_{i\in S}f_{i}$, maximized over all sets $S\subseteq\{1,2,\ldots,N\}$ of certain size. Crucially for our applications, the bound that we derive has no dependence on $N.$ The proof uses Chebyshev polynomials and the inclusion-exclusion principle. Armed with this *composition theorem*, we give a short proof of Theorem \[thm:MAIN-k-dnf\] on the approximate degree of $k$-DNF and $k$-CNF formulas. The argument proceeds by induction on $k,$ with the composition theorem invoked to implement the inductive step. The proof of Theorem \[thm:MAIN-ed\] on the approximate degree of $k$-element distinctness is more subtle. It too proceeds by induction, with the composition theorem playing a central role. This time, however, the induction is with respect to both $k$ and the range parameter $r,$ and the extension theorem is required to complete the inductive step. We note that we are able to bound the $\epsilon$-approximate degree of $k$-DNF formulas and $k$-element distinctness for every setting of the error parameter $\epsilon$, rather than just $\epsilon=1/3$ in Theorems \[thm:MAIN-ed\] and \[thm:MAIN-k-dnf\]. ### Surjectivity. {#surjectivity. .unnumbered} Our proof of Theorem \[thm:MAIN-surj\] is surprisingly short, given how improbable the statement was believed to be. As one can see from the defining equation for ${\operatorname{SURJ}}_{n,r}$, this function is the componentwise composition $\AND_{r}\circ\OR_{n}$ restricted to inputs of Hamming weight at most $n.$ With this in mind, we start with a degree-$O(\sqrt{r})$ polynomial $\widetilde{\AND}_{r}$ that approximates $\AND_{r}$ pointwise within $1/4.$ The approximant in question is simply a scaled and shifted Chebyshev polynomial. It follows that the componentwise composition $\widetilde{\AND}_{r}\circ\OR_{n}$, restricted to inputs of Hamming weight at most $n,$ approximates ${\operatorname{SURJ}}_{n,r}$ pointwise within $1/4$. We are not finished, however, because the degree of $\widetilde{\AND}_{r}\circ\OR_{n}$ is unacceptably large. Moving on, a few lines of algebra reveal that $\widetilde{\AND}_{r}\circ\OR_{n}$ is a linear combination of conjunctions in which the absolute values of the coefficients sum to $2^{O(\sqrt{r})}$. It remains to approximate each of these conjunctions pointwise within $2^{-\Omega(\sqrt{r})}$ by a polynomial of degree $O(\sqrt{n\sqrt{r}})=O(\sqrt{n}\cdot r^{1/4}),$ for which we use our explicit approximant from Theorem \[thm:MAIN-symmetric\] along with the guarantee that the input has Hamming weight at most $n.$ The proof of Theorem \[thm:MAIN-surj\] is particularly emblematic of our work in its interplay of approximation-theoretic methodology (Chebyshev polynomials, linear combinations) and algorithmic thinking (reduction of the problem to the approximation of individual conjunctions). We are pleased to report that our $O(n^{3/4})$ upper bound for the surjectivity problem has just sparked further progress in the area by Bun, Kothari, and Thaler [@BKT17poly-strikes-back], who prove tight or nearly tight lower bounds on the approximate degree of several key problems in quantum query complexity. In particular, the authors of [@BKT17poly-strikes-back] prove that our upper bound for surjectivity is tight. We are confident that the ideas of our work will inform future research as well. Preliminaries ============= We start with a review of the technical preliminaries. The purpose of this section is to make the paper as self-contained as possible, and comfortably readable by a broad audience. The expert reader may wish to skim it for the notation or skip it altogether. Notation -------- We view Boolean functions as mappings $X\to\zoo$ for some finite set $X.$ This arithmetization of the Boolean values “true” and “false” makes it possible to use Boolean operations in arithmetic expressions, as in $1-2\bigvee_{i=1}^{n}x_{i}.$ The familiar functions $\OR_{n}\colon\zoon\to\zoo$ and $\AND_{n}\colon\zoon\to\zoo$ are given by $\OR_{n}(x)=\bigvee_{i=1}^{n}x_{i}$ and $\AND_{n}(x)=\bigwedge_{i=1}^{n}x_{i}=\prod_{i=1}^{n}x_{i}.$ The negation of a Boolean function $f$ is denoted as usual by $\overline{f}=1-f.$ The composition of $f$ and $g$ is denoted $f\circ g$, with $(f\circ g)(x)=f(g(x)).$ For a string $x\in\zoon,$ we denote its Hamming weight by $|x|=x_{1}+x_{2}+\cdots+x_{n}.$ We use the following notation for strings of Hamming weight at most $k,$ greater than $k,$ and exactly $k$: $$\begin{aligned} \zoon_{\leq k} & =\{x\in\zoon:|x|\leq k\},\\ \zoon_{>k} & =\{x\in\zoon:|x|>k\},\\ \zoon_{k} & =\{x\in\zoon:|x|=k\}.\end{aligned}$$ For a string $x\in\zoon$ and a set $S\subseteq\{1,2,\ldots,n\},$ we let $x|_{S}$ denote the restriction of $x$ to the indices in $S.$ In other words, $x|_{S}=x_{i_{1}}x_{i_{2}}\ldots x_{i_{|S|}},$ where $i_{1}<i_{2}<\cdots<i_{|S|}$ are the elements of $S.$ The characteristic vector of a subset $S\subseteq\{1,2,\ldots,n\}$ is denoted $\1_{S}.$ We let $\NN=\{0,1,2,3,\ldots\}$ and $[n]=\{1,2,\ldots,n\}.$ For a set $S$ and a real number $k,$ we define $$\begin{aligned} \binom{S}{k} & =\{A\subseteq S:|A|=k\},\\ \binom{S}{\mathord{\leq}k} & =\{A\subseteq S:|A|\leq k\}.\end{aligned}$$ We analogously define $\binom{S}{\geq k},\binom{S}{<k},$ and $\binom{S}{>k}.$ We let $\ln x$ and $\log x$ stand for the natural logarithm of $x$ and the logarithm of $x$ to base $2,$ respectively. The following bound is well known [@jukna11extremal-2nd-edition Proposition 1.4]: $$\begin{aligned} \sum_{i=0}^{k}{n \choose i}\leq\left(\frac{\e n}{k}\right)^{k}, & & k=0,1,2,\dots,n,\label{eq:entropy-bound-binomial}\end{aligned}$$ where $\e=2.7182\ldots$ denotes Euler’s number. For a logical condition $C,$ we use the Iverson bracket notation $$\I[C]=\begin{cases} 1 & \text{if \ensuremath{C} holds,}\\ 0 & \text{otherwise.} \end{cases}$$ For a function $f\colon X\to\Re$ on a finite set $X,$ we use the standard norms $$\begin{aligned} & \|f\|_{\infty}=\max_{x\in X}\,|f(x)|,\\ & \|f\|_{1}=\sum_{x\in X}\,|f(x)|.\end{aligned}$$ Approximate degree ------------------ Recall that the *total degree* of a multivariate real polynomial $p\colon\Re^{n}\to\Re$, denoted $\deg p,$ is the largest degree of any monomial of $p.$ We use the terms “degree” and “total degree” interchangeably in this paper. This paper studies the approximate representation of functions of interest by polynomials. Specifically, let $f\colon X\to\Re$ be a given function, for a finite subset $X\subset\Re^{n}.$ Define $$E(f,d)=\min_{p:\deg p\leq d}\|f-p\|_{\infty},$$ where the minimum is over polynomials of degree at most $d.$ In words, $E(f,d)$ is the least error to which $f$ can be approximated by a real polynomial of degree at most $d$. For a real number $\epsilon\geq0,$ the *$\epsilon$-approximate degree* of $f$ is defined as $$\deg_{\epsilon}(f)=\min\{d:E(f,d)\leq\epsilon\}.$$ Thus, $\deg_{\epsilon}(f)$ is the least degree of a real polynomial that approximates $f$ pointwise to within $\epsilon.$ We refer to any such polynomial as a *uniform approximant for $f$ with error $\epsilon$*. In the study of Boolean functions $f$, the standard setting of the error parameter is $\epsilon=1/3$. This constant is chosen mostly for aesthetic reasons and can be replaced by any other constant in $(0,1/2)$ at the expense of a constant-factor increase in approximate degree. The following fact on the *exact* representation of functions by polynomials is well known. \[fact:trivial-upper-bound-on-deg\]For every function $f\colon\zoo_{\leq n}^{N}\to\Re,$ $$\deg_{0}(f)\leq n.$$ The proof is by induction on $n.$ The base case $n=0$ is trivial since $f$ is then a constant function. For the inductive step, let $n\geq1$ be arbitrary. By the inductive hypothesis, there is a polynomial $p_{n-1}(x)$ of degree at most $n-1$ such that $f(x)=p_{n-1}(x)$ for inputs $x\in\zoo^{N}$ of Hamming weight at most $n-1.$ Define $$p_{n}(x)=p_{n-1}(x)+\sum_{a\in\zoo_{n}^{N}}(f(a)-p_{n-1}(a))\prod_{i:a_{i}=1}x_{i}.$$ For any fixed input $x$ with $|x|\leq n-1,$ every term in the summation over $a$ evaluates to zero and therefore $p_{n}(x)=p_{n-1}(x)=f(x).$ For any fixed input $x$ with $|x|=n,$ on the other hand, the summation over $a$ contributes precisely one nonzero term, corresponding to $a=x.$ As a result, $p_{n}(x)=p_{n-1}(x)+(f(x)-p_{n-1}(x))=f(x)$ in that case. Inclusion-exclusion ------------------- All Boolean, arithmetic, and relational operations on functions in this paper are to be interpreted pointwise. For example, $\bigvee_{i=1}^{n}f_{i}$ refers to the mapping $x\mapsto\bigvee_{i=1}^{n}f_{i}(x).$ Similarly, $\prod_{i=1}^{n}f_{i}$ is the pointwise product of $f_{1},f_{2},\ldots,f_{n}$. Recall that in the case of Boolean functions, we have $\bigwedge_{i=1}^{n}f_{i}=\prod_{i=1}^{n}f_{i}.$ The well-known *inclusion-exclusion principle*, stated in terms of Boolean functions $f_{1},f_{2},\ldots,f_{n},$ asserts that $$\bigvee_{i=1}^{n}f_{i}=\sum_{\substack{S\subseteq\{1,2,\ldots,n\}\\ S\ne\varnothing } }(-1)^{|S|+1}\prod_{i\in S}f_{i}.$$ We will need the following less common form of the inclusion-exclusion principle, where the AND and OR operators are interchanged. \[fact:incl-excl-alternative\]For any $n\geq1$ and any Boolean functions $f_{1},f_{2},\ldots,f_{n}\colon X\to\zoo,$ $$\prod_{i=1}^{n}f_{i}=\sum_{\substack{S\subseteq\{1,2,\ldots,n\}\\ S\ne\varnothing } }(-1)^{|S|+1}\bigvee_{i\in S}f_{i}.$$ We have $$\begin{aligned} \prod_{i=1}^{n}f_{i} & =\prod_{i=1}^{n}(1-\overline{f_{i}})\\ & =\sum_{S\subseteq\{1,2,\ldots,n\}}(-1)^{|S|}\prod_{i\in S}\overline{f_{i}}\\ & =\sum_{S\subseteq\{1,2,\ldots,n\}}(-1)^{|S|}\left(\prod_{i\in S}\overline{f_{i}}-1\right)\\ & =\sum_{S\subseteq\{1,2,\ldots,n\}}(-1)^{|S|}\left(-\bigvee_{i\in S}f_{i}\right)\\ & =\sum_{\substack{S\subseteq\{1,2,\ldots,n\}\\ S\ne\varnothing } }(-1)^{|S|+1}\bigvee_{i\in S}f_{i},\end{aligned}$$ where the third step uses the fact that half of the subsets of $\{1,2,\ldots,n\}$ have odd cardinality and the other half have even cardinality. Symmetrization -------------- Let $S_{n}$ denote the symmetric group on $n$ elements. For a permutation $\sigma\in S_{n}$ and a string $x=(x_{1},x_{2},\ldots,x_{n}),$ we adopt the shorthand $\sigma x=(x_{\sigma(1)},x_{\sigma(2)},\ldots,x_{\sigma(n)}).$ A function $f(x_{1},x_{2},\ldots,x_{n})$ is called *symmetric* if it is invariant under permutations of the input variables: $f(x_{1},x_{2},\ldots,x_{n})\equiv f(x_{\sigma(1)},x_{\sigma(2)},\ldots,x_{\sigma(n)})$ for all $x$ and $\sigma.$ Symmetric functions on $\zoon$ are intimately related to univariate polynomials, as borne out by Minsky and Papert’s *symmetrization argument* [@minsky88perceptrons]. \[prop:minsky-papert\]Let $p\colon\zoon\to\Re$ be a polynomial of degree $d.$ Then there is a univariate polynomial $p^{*}$ of degree at most $d$ such that for all $x\in\zoon,$ $$\begin{aligned} \Exp_{\sigma\in S_{n}}p(\sigma x)=p^{*}(|x|).\end{aligned}$$ Minsky and Papert’s result generalizes to block-symmetric functions, as pointed out in [@RS07dc-dnf Prop. 2.3]: \[prop:symmetrization\] Let $n_{1},\dots,n_{k}$ be positive integers. Let $p\colon\zoo^{n_{1}}\times\cdots\times\zoo^{n_{k}}\to\Re$ be a polynomial of degree $d.$ Then there is a polynomial $p^{*}\colon\Re^{k}\to\Re$ of degree at most $d$ such that for all $x_{1}\in\zoo^{n_{1}},\ldots,x_{k}\in\zoo^{n_{k}},$ $$\begin{aligned} \Exp_{\sigma_{1}\in S_{n_{1}},\dots,\sigma_{k}\in S_{n_{k}}}p(\sigma_{1}x_{1},\dots,\sigma_{k}x_{k}) & =p^{*}(|x_{1}|,\ldots,|x_{k}|).\end{aligned}$$ Proposition \[prop:symmetrization\] follows in a straightforward manner from Proposition \[prop:minsky-papert\] by induction on the number of blocks, $k.$ \[subsec:chebyshev\]Chebyshev polynomials ----------------------------------------- Recall from Euler’s identity that $$\begin{aligned} (\cos x+{\mathbf{i}}\sin x)^{d} & =\cos dx+{\mathbf{i}}\sin dx, & & d=0,1,2,\ldots,\label{eq:de-moivre}\end{aligned}$$ where ${\mathbf{i}}$ denotes the imaginary unit. Multiplying out the left-hand side and using $\sin^{2}x=1-\cos^{2}x,$ we obtain a univariate polynomial $T_{d}$ of degree $d$ such that $$T_{d}(\cos x)=\cos dx.\label{eq:Chebyshev-definition}$$ This unique polynomial is the *Chebyshev polynomial of degree $d$*. The representation (\[eq:Chebyshev-definition\]) immediately reveals all the roots of $T_{d}$, and all the extrema of $T_{d}$ in the interval $[-1,1]$: $$\begin{aligned} & T_{d}\left(\cos\left(\frac{2i-1}{2d}\,\pi\right)\right)=0, & & i=1,2,\ldots,d,\label{eq:Chebyshev-roots}\\ & T_{d}\left(\cos\left(\frac{i}{d}\,\pi\right)\right)=(-1)^{i}, & & i=0,1,\ldots,d,\label{eq:Chebyshev-extrema}\\ \rule{0mm}{4mm} & |T_{d}(t)|\leq1, & & t\in[-1,1].\label{eqn:chebyshev-containment}\end{aligned}$$ The extremum at $1$ is of particular significance, and we note it separately: $$T_{d}(1)=1.\label{eq:chebyshev-at-1}$$ In view of (\[eq:de-moivre\]), the defining equation (\[eq:Chebyshev-definition\]) implies that $$\begin{aligned} T_{d}(\cos x) & =\sum_{i=0}^{\lfloor d/2\rfloor}\binom{d}{2i}(-1)^{i}(\sin x)^{2i}(\cos x)^{d-2i}\\ & =\sum_{i=0}^{\lfloor d/2\rfloor}\binom{d}{2i}(\cos^{2}x-1)^{i}(\cos x)^{d-2i},\end{aligned}$$ so that the leading coefficient of $T_{d}$ for $d\geq1$ is given by $\sum_{i=0}^{\lfloor d/2\rfloor}\binom{d}{2i}=2^{d-1}$. As a result, we have the factored representation $$\begin{aligned} T_{d}(t) & =2^{d-1}\prod_{i=1}^{d}\left(t-\cos\left(\frac{2i-1}{2d}\,\pi\right)\right), & & d\geq1.\label{eq:chebyshev-factored}\end{aligned}$$ By (\[eq:de-moivre\]) and (\[eq:Chebyshev-definition\]), $$\begin{aligned} T_{d}(\cos x) & =\cos dx\\ & =\frac{1}{2}(\cos x-{\mathbf{i}}\sin x)^{d}+\frac{1}{2}(\cos x+{\mathbf{i}}\sin x)^{d}\\ & =\frac{1}{2}(\cos x-{\mathbf{i}}\sqrt{1-\cos^{2}x})^{d}+\frac{1}{2}(\cos x+{\mathbf{i}}\sqrt{1-\cos^{2}x})^{d},\end{aligned}$$ whence $$\begin{aligned} T_{d}(t) & =\frac{1}{2}(t-\sqrt{t^{2}-1})^{d}+\frac{1}{2}(t+\sqrt{t^{2}-1})^{d}, & & |t|\geq1.\label{eq:chebyshev-beyond-1}\end{aligned}$$ The following fundamental fact follows from (\[eq:chebyshev-beyond-1\]) by elementary calculus. \[fact:chebyshev-derivative\]For any integer $d\geq0$ and real $t\geq1,$ $$\begin{aligned} & T_{d}'(t)\geq d^{2}.\end{aligned}$$ Together, (\[eq:chebyshev-beyond-1\]) and Fact \[fact:chebyshev-derivative\] give the following useful lower bound for Chebyshev polynomials on $[1,\infty).$ \[prop:chebyshev-beyond-1\]For any integer $d\geq1,$ $$\begin{aligned} T_{d}(1+\delta) & \geq1+d^{2}\delta, & & 0\leq\delta<\infty,\\ T_{d}(1+\delta) & \geq2^{d\sqrt{\delta}-1} & & 0\leq\delta\leq1.\end{aligned}$$ The first bound follows from the intermediate value theorem in view of (\[eq:chebyshev-at-1\]) and Fact \[fact:chebyshev-derivative\]. For the second bound, use (\[eq:chebyshev-beyond-1\]) to write $$\begin{aligned} T_{d}(1+\delta) & \geq\frac{1}{2}(1+\delta+\sqrt{(1+\delta)^{2}-1})^{d}\\ & \geq\frac{1}{2}(1+\sqrt{\delta})^{d}\\ & \geq\frac{1}{2}\cdot2^{d\sqrt{\delta}},\end{aligned}$$ where the last step uses $1+x\geq2^{x}$ for $x\in[0,1].$ Coefficient bounds for univariate polynomials --------------------------------------------- We let $P_{d}$ stand for the set of univariate polynomials of degree at most $d.$ For a univariate polynomial $p(t)=a_{d}t^{d}+a_{d-1}t^{d-1}+\cdots+a_{1}t+a_{0},$ we let ${|\!|\!| p | \! | \! |}=\sum_{i=0}^{d}|a_{i}|$ denote the sum of the absolute values of the coefficients of $p.$ Then ${|\!|\!| \cdot | \! | \! |}$ is a norm on the real linear space of polynomials, and it is in addition submultiplicative: \[fact:poly-norm\]For any polynomials $p$ and $q,$ 1. ${|\!|\!| p | \! | \! |}\geq0,$ with equality if and only if $p=0;$ 2. ${|\!|\!| \lambda p | \! | \! |}=|\lambda|\cdot{|\!|\!| p | \! | \! |}$ for any real $\lambda;$ 3. ${|\!|\!| p+q | \! | \! |}\leq{|\!|\!| p | \! | \! |}+{|\!|\!| q | \! | \! |};$ 4. \[item:poly-norm-submultiplicative\]${|\!|\!| p\cdot q | \! | \! |}\leq{|\!|\!| p | \! | \! |}\cdot{|\!|\!| q | \! | \! |}.$ All four properties follow directly from the definition. We will need a bound on the coefficients of a univariate polynomial in terms of its degree $d$ and its maximum absolute value on the interval $[0,1].$ This fundamental problem was solved in the nineteenth century by V. A. Markov [@markov-v-a1982least-dev p. 81], who proved an upper bound of $$\begin{aligned} O\left(\frac{(1+\sqrt{2})^{d}}{\sqrt{d}}\right)\label{eqn:markov}\end{aligned}$$ on the size of the coefficients of any degree-$d$ polynomial that is bounded on $[-1,1]$ in absolute value by $1.$ Markov further showed that (\[eqn:markov\]) is tight. Rather than appeal to this deep result in approximation theory, we will use the following weaker bound that suffices for our purposes. \[lem:bound-on-poly-coeffs-UNIVARIATE\] Let $p$ be a univariate polynomial of degree $d$. Then $$\begin{aligned} {|\!|\!| p | \! | \! |}\leq8^{d}\max_{i=0,1,\dots,d}\left|p\left(\frac{i}{d}\right)\right|.\label{eqn:bound-poly-coeffs}\end{aligned}$$ Lemma \[lem:bound-on-poly-coeffs-UNIVARIATE\] is a cosmetic modification of a lemma from [@sherstov12noisy], which in our notation states that ${|\!|\!| p | \! | \! |}\leq4^{d}\max_{i=0,1,\ldots,d}|p(1-\frac{2i}{d})|$ for $p\in P_{d}$. We include a detailed proof for the reader’s convenience. We use a common approximation-theoretic technique [@cheney-book; @rivlin-book] whereby one expresses $p$ as a linear combination of more structured polynomials and analyzes the latter objects. For this, define $q_{0},q_{1},\dots,q_{d}\in P_{d}$ by $$\begin{aligned} q_{j}(t)=\frac{(-1)^{d-j}d^{d}}{d!}{d \choose j}\prod_{\substack{i=0\\ i\ne j } }^{d}\left(t-\frac{i}{d}\right),\qquad j=0,1,\dots,d.\end{aligned}$$ One easily verifies that these polynomials behave like delta functions, in the sense that for $i,j=0,1,2,\dots,d,$ $$\begin{aligned} q_{j}\left(\frac{i}{d}\right)=\begin{cases} 1 & \text{if \ensuremath{i=j,}}\\ 0 & \text{otherwise.} \end{cases}\end{aligned}$$ Lagrange interpolation gives $$\begin{aligned} p=\sum_{j=0}^{d}p\left(\frac{j}{d}\right)q_{j}.\label{eqn:p-lin-comb-q}\end{aligned}$$ By Fact \[fact:poly-norm\], $$\begin{aligned} {|\!|\!| q_{j} | \! | \! |} & \leq\frac{d^{d}}{d!}{d \choose j}\prod_{\substack{i=0\\ i\ne j } }^{d}\left(1+\frac{i}{d}\right)\nonumber \\ & \leq\frac{d^{d}}{d!}{d \choose j}\prod_{i=1}^{d}\left(1+\frac{i}{d}\right)\nonumber \\ & =\frac{1}{d!}{d \choose j}\frac{(2d)!}{d!} & & \allowdisplaybreaks\nonumber \\ & ={d \choose j}\binom{2d}{d}\nonumber \\ & \leq4^{d}\binom{d}{j}, & & j=0,1,2,\ldots,d.\label{eq:qj-Pi-norm}\end{aligned}$$ Now $$\begin{aligned} {|\!|\!| p | \! | \! |} & \leq\left(\max_{j=0,1,\dots,d}\left|p\left(\frac{j}{d}\right)\right|\right)\sum_{j=0}^{d}4^{d}{d \choose j}\\ & =8^{d}\max_{j=0,1,\dots,d}\left|p\left(\frac{j}{d}\right)\right|,\end{aligned}$$ where the first step uses (\[eqn:p-lin-comb-q\]), (\[eq:qj-Pi-norm\]), and Fact \[fact:poly-norm\]. Coefficient bounds for multivariate polynomials ----------------------------------------------- Let $\phi\colon\Re^{n}\to\Re$ be a multivariate polynomial. Analogous to the univariate case, we let ${|\!|\!| \phi | \! | \! |}$ denote the sum of the absolute values of the coefficients of $\phi.$ Fact \[fact:poly-norm\] is clearly valid in this multivariate setting as well. Recall that a multivariate polynomial $\phi$ is *multilinear* if it has degree at most $1$ in each variable. The following result is an analogue of Lemma \[lem:bound-on-poly-coeffs-UNIVARIATE\]. \[lem:bound-on-poly-coeffs-MULTIVARIATE\]Let $\phi\colon\Re^{n}\to\Re$ be a symmetric multilinear polynomial. Then $${|\!|\!| \phi | \! | \! |}\leq8^{\deg\phi}\max_{x\in\zoon}|\phi(x)|.$$ Abbreviate $d=\deg\phi$ and write $$\phi(x)=\sum_{i=0}^{d}a_{i}\sum_{S\in\binom{[n]}{i}}\;\prod_{j\in S}x_{j},$$ where $a_{0},a_{1},\ldots,a_{d}$ are real coefficients. For $0\leq t\leq1,$ let $B(t)$ denote the Bernoulli distribution with success probability $t.$ Then $$\begin{aligned} {|\!|\!| \phi | \! | \! |} & =\sum_{i=0}^{d}|a_{i}|\binom{n}{i}\\ & \leq8^{d}\max_{0\leq t\leq1}\left|\sum_{i=0}^{d}a_{i}\binom{n}{i}t^{i}\right|\\ & =8^{d}\max_{0\leq t\leq1}\left|\Exp_{x_{1},x_{2},\ldots,x_{n}\sim B(t)}\phi(x)\right|\\ & \leq8^{d}\max_{x\in\zoon}|\phi(x)|,\end{aligned}$$ where the second and third steps use Lemma \[lem:bound-on-poly-coeffs-UNIVARIATE\] and multilinearity, respectively. The following lemma, due to Razborov and Sherstov [@RS07dc-dnf Lemma 3.2], bounds the value of a polynomial $p$ at a point of large Hamming weight in terms of $p$’s values at points of low Hamming weight. \[lem:extrapolation\]Let $d$ be an integer, $0\leq d\leq n-1.$ Let $\phi:\Re^{n}\to\Re$ be a polynomial of degree at most $d.$ Then $$|\phi(1^{n})|\;\leq\;2^{d}{n \choose d}\,\max_{x\in\zoo_{\leq d}^{n}}|\phi(x)|.$$ As one would expect, one can sharpen the bound of Lemma \[lem:extrapolation\] by maximizing over a larger neighborhood of the Boolean hypercube than $\zoo_{\leq d}^{n}$. The resulting bound is as follows. \[lem:extrapolation-generalized\]Fix positive integers $N>m\geq d.$ Let $\phi\colon\Re^{N}\to\Re$ be a polynomial of degree at most $d.$ Then $$\begin{aligned} |\phi(x^{*})| & \leq2^{d}\binom{\lceil|x^{*}|/\lfloor m/d\rfloor\rceil}{d}\max_{x\in\zoo_{\leq m}^{N}}|\phi(x)|, & & x^{*}\in\zoo_{>m}^{N}.\end{aligned}$$ One recovers Lemma \[lem:extrapolation\] as a special case by taking $N=n,\;m=d,\;$ and $x^{*}=1^{n}.$ Consider an arbitrary vector $x^{*}\in\zoo^{N}$ of Hamming weight $|x^{*}|>m,$ and abbreviate $n=\lceil|x^{*}|/\lfloor m/d\rfloor\rceil.$ Let $S_{1},S_{2},\ldots,S_{n}$ be a partition of $\{i:x_{i}^{*}=1\}$ such that $|S_{i}|\leq\lfloor m/d\rfloor$ for all $i.$ Observe that $$n>d.\label{eq:extrapolation-n-d}$$ Define $L\colon\zoon\to\zoo^{N}$ by $$L(z)=\sum_{i=1}^{n}z_{i}\1_{S_{i}}.$$ Then clearly $$\begin{aligned} & L(1^{n})=x^{*},\label{eq:L-x-star}\\ & |L(z)|\leq|z|\cdot\left\lfloor \frac{m}{d}\right\rfloor .\label{eq:L-hamming-weight-grows}\end{aligned}$$ Moreover, the mapping $z\mapsto\phi(L(z))$ is a real polynomial on $\zoon$ of degree at most $\deg\phi\leq d.$ As a result, $$\begin{aligned} |\phi(x^{*})| & =|\phi(L(1^{n}))|\\ & \leq2^{d}\binom{n}{d}\max_{|z|\leq d}|\phi(L(z))|\\ & \leq2^{d}\binom{n}{d}\max_{|x|\leq d\lfloor m/d\rfloor}|\phi(x)|\\ & \leq2^{d}\binom{n}{d}\max_{|x|\leq m}|\phi(x)|,\end{aligned}$$ where the first step uses (\[eq:L-x-star\]); the second step follows by (\[eq:extrapolation-n-d\]) and Lemma \[lem:extrapolation\]; and the third step is valid by (\[eq:L-hamming-weight-grows\]). The conjunction norm -------------------- Recall that a *conjunction* in Boolean variables $x_{1},x_{2},\ldots,x_{n}$ is the AND of some subset of the literals $x_{1},\overline{x_{1}},x_{2},\overline{x_{2}},\ldots,x_{n},\overline{x_{n}}.$ Analogously, a *disjunction* is the OR of some subset of $x_{1},\overline{x_{1}},x_{2},\overline{x_{2}},\ldots,x_{n},\overline{x_{n}}.$ We regard conjunctions and disjunctions as Boolean functions $\zoon\to\zoo$ and in particular as a special case of real functions $\zoon\to\Re.$ For a subset $X\subseteq\zoon$ and a function $f\colon X\to\Re$, we define the *conjunction norm* $\operatorname{\Pi}(f)$ to be the minimum $\Lambda\geq0$ such that $$\begin{aligned} f(x) & =\lambda_{1}C_{1}(x)+\lambda_{2}C_{2}(x)+\cdots+\lambda_{N}C_{N}(x) & & (x\in X)\end{aligned}$$ for some integer $N,$ some conjunctions $C_{1},C_{2},\ldots,C_{N},$ and some real coefficients $\lambda_{1},\lambda_{2},\ldots,\lambda_{N}$ with $|\lambda_{1}|+|\lambda_{2}|+\cdots+|\lambda_{N}|\leq\Lambda.$ Our choice of the symbol $\Pi,$ for “product,” is motivated by the view of conjunctions as products of literals. In particular, we have $\operatorname{\Pi}(\phi)\leq{|\!|\!| \phi | \! | \! |}$ for any multivariate polynomial $\phi\colon\zoon\to\Re.$ The next proposition shows that $\operatorname{\Pi}$ is a norm on the space of multivariate real functions and establishes other useful properties of this complexity measure. \[prop:orcomplexity\]Let $f,g\colon X\to\Re$ be given functions, for a nonempty set $X\subseteq\zoon$. Then: 1. \[item:orcomplexity-nonzero\]$\operatorname{\Pi}(f)\geq0,$ with equality if and only if $f=0;$ 2. \[item:orcomplexity-homogeneous\]$\operatorname{\Pi}(\lambda f)=|\lambda|\operatorname{\Pi}(f)$ for any real $\lambda;$ 3. \[item:orcomplexity-triangle\]$\operatorname{\Pi}(f+g)\leq\operatorname{\Pi}(f)+\operatorname{\Pi}(g);$ 4. \[item:orcomplexity-product\]$\operatorname{\Pi}(f\cdot g)\leq\operatorname{\Pi}(f)\operatorname{\Pi}(g);$ 5. \[item:orcomplexity-trivial-bound\]$\operatorname{\Pi}(f)\leq\|f\|_{1};$ 6. \[item:orcomplexity-disjunction\]$\operatorname{\Pi}(f)\leq2$ if $f$ is a disjunction$;$ 7. \[item:orcomplexity-composition-with-univariate\]$\operatorname{\Pi}(p\circ f)\leq\max\{1,\operatorname{\Pi}(f)\}^{d}\,{|\!|\!| p | \! | \! |}$ for any polynomial $p\in P_{d}.$ \[item:orcomplexity-nonzero\]\[item:orcomplexity-triangle\] Immediate from the definitions. \[item:orcomplexity-product\] Express $f$ and $g$ individually as a linear combination of conjunctions with real coefficients whose absolute values sum to $\operatorname{\Pi}(f)$ and $\Pi(g)$, respectively. Then, multiply these two linear combinations. Since the product of conjunctions is again a conjunction, the resulting representation is a linear combination of conjunctions with real coefficients whose absolute values sum to at most $\operatorname{\Pi}(f)\operatorname{\Pi}(g).$ \[item:orcomplexity-trivial-bound\] By the homogeneity \[item:orcomplexity-homogeneous\] and triangle inequality \[item:orcomplexity-triangle\], we have $$\begin{aligned} \operatorname{\Pi}(f) & =\operatorname{\Pi}\left(\sum_{a\in X}f(a)C_{a}\right)\\ & \leq\sum_{a\in X}|f(a)|\operatorname{\Pi}(C_{a})\\ & \leq\sum_{a\in X}|f(a)|\\ & =\|f\|_{1},\end{aligned}$$ where $C_{a}$ denotes the conjunction that evaluates to true on $a$ and to false on all other inputs in $\zoon.$ \[item:orcomplexity-disjunction\] We have $\operatorname{\Pi}(f)\leq\operatorname{\Pi}(f-1)+\operatorname{\Pi}(1)=\operatorname{\Pi}(1-f)+\operatorname{\Pi}(1)\leq2,$ where the first step applies the triangle inequality \[item:orcomplexity-triangle\], the second step uses the homogeneity \[item:orcomplexity-homogeneous\], and the third step uses the fact that $1$ and $1-f$ are conjunctions. \[item:orcomplexity-composition-with-univariate\] Let $p(t)=a_{d}t^{d}+a_{d-1}t^{d-1}+\cdots+a_{1}t+a_{0}$ be a given polynomial. Then $$\begin{aligned} \operatorname{\Pi}(p\circ f) & =\operatorname{\Pi}\left(\sum_{i=0}^{d}a_{i}\underbrace{f\cdot f\cdot\cdots\cdot f}_{i}\right)\\ & \leq\sum_{i=0}^{d}|a_{i}|\operatorname{\Pi}(\underbrace{f\cdot f\cdot\cdots\cdot f}_{i})\\ & \leq\sum_{i=0}^{d}|a_{i}|\operatorname{\Pi}(f)^{i}\allowdisplaybreaks\\ & \leq\max\{1,\operatorname{\Pi}(f)^{d}\}\sum_{i=0}^{d}|a_{i}|\\ & =\max\{1,\operatorname{\Pi}(f)\}^{d}\;{|\!|\!| p | \! | \! |},\end{aligned}$$ where the second step uses \[item:orcomplexity-homogeneous\] and \[item:orcomplexity-triangle\], and the third step applies \[item:orcomplexity-product\]. \[sec:extension\]The extension theorem ====================================== This section establishes an approximation-theoretic result of independent interest, the *extension theorem*, that we use several times in the rest of the paper to construct approximating polynomials. To set the stage for this result, let $f\colon\zoo_{\leq m}^{N}\to[-1,1]$ be a given function, defined on inputs of Hamming weight up to $m.$ For any integer $n>m,$ consider the extension $F_{n}$ of $f$ to inputs of Hamming weight up to $n$, given by $$F_{n}(x)=\begin{cases} f(x) & \text{if \ensuremath{|x|\leq m,}}\\ 0 & \text{otherwise.} \end{cases}$$ From the point of view of approximation theory, a fundamental question to ask is how to “extend” any approximant for $f$ to an approximant for $F_{n},$ without degrading the quality of the approximation or significantly increasing the approximant’s degree. Ideally, we would like the approximant for the extension $F_{n}$ to have degree within a small factor of the original degree, e.g., a factor of $O(n/m)^{\alpha}$ for some constant $0<\alpha<1.$ Unfortunately, the extension problem is hopeless as stated. Indeed, consider the special case of the constant function $f=1,$ so that $$F_{n}(x)=\begin{cases} 1 & \text{if }0\leq|x|\leq m,\\ 0 & \text{if }m<|x|\leq n. \end{cases}$$ In this example, $\deg_{1/3}(f)=0$ but $\deg_{1/3}(F_{n})=\Omega(\sqrt{n})$ by a well-known result of Nisan and Szegedy [@nisan-szegedy94degree]. In particular, there is no efficient way to transform an approximant for a general function $f$ into an approximant for the extension $F_{n}.$ Our contribution is to show that the extension problem becomes meaningful and efficiently solvable if one’s starting point is an approximant for $F_{2m}$ rather than for $f.$ In other words, we give an efficient, black-box transformation of an approximant for $F_{2m}$ into an approximant for any extension $F_{n},$ where $n\geq2m.$ The formal statement of our result is as follows. \[thm:extension\]Let $f\colon\zoo_{\leq m}^{N}\to[-1,1]$ be given, where $N\geq m\geq0$ are integers. For integers $n\geq m,$ define $F_{n}\colon\zoo_{\leq n}^{N}\to[-1,1]$ by $$F_{n}(x)=\begin{cases} f(x) & \text{if \ensuremath{|x|\leq m,}}\\ 0 & \text{otherwise.} \end{cases}$$ Then for some absolute constant $C>1$ and all $\epsilon,\delta\in(0,1/2)$ and $n\geq m,$ $$\deg_{\epsilon+\delta}(F_{n})\leq C\sqrt{\frac{n}{m+1}}\cdot\left(\deg_{\epsilon}(F_{2m})+\log\frac{1}{\delta}\right).\label{eq:degeps-Fn-advertise}$$ Theorem \[thm:extension\] solves the extension problem with only a factor-$\sqrt{n/m}$ increase in degree. The approximation quality of the new approximant can be made arbitrarily close to that of the original at a small additive cost in degree. This overhead in degree and error is optimal, as we will discover in applications later in this paper. We also note that the constant $2$ in this result was chosen exclusively for aesthetic reasons, and (\[eq:degeps-Fn-advertise\]) holds with $F_{2m}$ replaced by $F_{\lceil cm\rceil}$ for any constant $c>1.$ The rest of this section is devoted to the proof of Theorem \[thm:extension\]. Proof strategy -------------- In the notation of Theorem \[thm:extension\], let $p_{2m}(x)$ be an approximant for $F_{2m}(x).$ Then clearly $$F_{n}(x)\approx p_{2m}(x)\cdot\I[|x|\leq2m]\label{eq:extension-proof-strategy}$$ on the domain of $F_{n},$ where $\I[|x|\leq2m]$ is the characteristic function of the set of inputs of Hamming weight at most $2m.$ While $p_{2m}(x)$ can grow rapidly as the Hamming weight $|x|$ increases beyond $2m,$ that growth is not entirely arbitrary. Specifically, the generalized extrapolation lemma (Lemma \[lem:extrapolation-generalized\]) bounds $|p_{2m}(x)|$ in terms of the Hamming weight $|x|$ and the degree of $p_{2m}$. In particular, the approximate equality (\[eq:extension-proof-strategy\]) is preserved if $\I[|x|\leq2m]$ is replaced by a low-degree approximant. The construction of such an approximant is the crux of our proof. More precisely, we construct a low-degree *univariate* approximant to the characteristic function of any interval. To crystallize our approach, we first consider the degenerate interval $[0,0]=\{0\}.$ \[prop:ext-theorem-strategy\]For any positive integers $n$ and $d,$ there is a polynomial $p$ with $$\begin{aligned} & p(0)=1,\label{eq:or-rapid-decrease-at-0}\\ & |p(t)|\leq\frac{1}{t^{d}}, & & t\in[1,n],\label{eq:or-rapid-decrease-at-t}\\ & \deg p\leq7d\sqrt{n}.\label{eq:or-rapid-decrease-deg}\end{aligned}$$ The key property here is (\[eq:or-rapid-decrease-at-t\]), whereby the approximating polynomial gets smaller as one moves farther away from the point of interest, $0$. Reproducing this behavior in the context of a general interval is much more subtle and is the subject of Sections \[subsec:reciprocal\]\[subsec:characteristic-fn-interval\]. Define $$T(t)=\left(\prod_{i=0}^{\lceil\log n\rceil}T_{\lceil\sqrt{n/2^{i}}\rceil}\left(1+\frac{2^{i}-t}{n}\right)\right)^{d}.$$ Fix an arbitrary point $t\in[1,n]$, and let $j$ be the integer such that $t\in[2^{j},2^{j+1}).$ Then $$\begin{aligned} |T(t)| & =\prod_{i=0}^{\lceil\log n\rceil}\left|T_{\lceil\sqrt{n/2^{i}}\rceil}\left(1+\frac{2^{i}-t}{n}\right)\right|^{d}\\ & \leq\prod_{i=j+1}^{\lceil\log n\rceil}\left|T_{\lceil\sqrt{n/2^{i}}\rceil}\left(1+\frac{2^{i}-t}{n}\right)\right|^{d}\\ & \leq\prod_{i=j+1}^{\lceil\log n\rceil}\left|T_{\lceil\sqrt{n/2^{i}}\rceil}\left(1+\frac{2^{i}}{n}\right)\right|^{d}\allowdisplaybreaks\\ & =|T(0)|\;\prod_{i=0}^{j}\left|T_{\lceil\sqrt{n/2^{i}}\rceil}\left(1+\frac{2^{i}}{n}\right)\right|^{-d}\\ & \leq|T(0)|\;\prod_{i=0}^{j}2^{-d}\\ & \leq\frac{|T(0)|}{t^{d}},\end{aligned}$$ where the second step uses (\[eqn:chebyshev-containment\]), the third step follows from (\[eq:chebyshev-at-1\]) and Fact \[fact:chebyshev-derivative\], and the next-to-last step applies Proposition \[prop:chebyshev-beyond-1\]. Moreover, $$\begin{aligned} \deg T & =d\sum_{i=0}^{\lceil\log n\rceil}\left\lceil \sqrt{\frac{n}{2^{i}}}\right\rceil \\ & \leq d\sum_{i=0}^{\infty}\sqrt{\frac{n}{2^{i}}}\cdot2\\ & \leq7d\sqrt{n}.\end{aligned}$$ As a result, (\[eq:or-rapid-decrease-at-0\])(\[eq:or-rapid-decrease-deg\]) hold for $p(t)=T(t)/T(0).$ \[subsec:reciprocal\]Approximating 1/*t* ---------------------------------------- To handle actual intervals rather than singleton points, we need to develop a number of auxiliary results. To start with, we construct an approximant for the reciprocal function $1/t$ on $[1,n]$. We are specifically interested in approximation within a *multiplicative* factor close to $1$, which is a more demanding regime than pointwise approximation. \[lem:reciprocal\]For any integer $d\geq0$ and real $n>1,$ there is an $($explicitly given$)$ polynomial $p\in P_{d}$ such that $$\begin{aligned} \frac{1-\epsilon}{t} & \leq p(t)\leq\frac{1+\epsilon}{t}, & & 1\leq t\leq n,\label{eq:approximate-t-multiplicatively}\end{aligned}$$ where $$\epsilon=\frac{1}{T_{d+1}\bigl(\frac{n+1}{n-1}\bigr)}.$$ Property (\[eq:approximate-t-multiplicatively\]) can be restated as $\max_{1\leq t\leq n}|1-tp(t)|\leq\epsilon$. Thus, the existence of $p\in P_{d}$ that obeys (\[eq:approximate-t-multiplicatively\]) is equivalent to the existence of $q\in P_{d+1}$ that obeys $q(0)=1$ and $\max_{1\leq t\leq n}|q(t)|\leq\epsilon$. Now, define $q\in P_{d+1}$ by $$q(t)=\frac{T_{d+1}\left(1-2\cdot\frac{t-1}{n-1}\right)}{T_{d+1}\bigl(\frac{n+1}{n-1}\bigr)}.$$ Then $q(0)=1$ by definition. Moreover, $$\begin{aligned} \max_{1\leq t\leq n}|q(t)| & \leq\max_{-1\leq t\leq1}\frac{|T_{d+1}(t)|}{\bigl|T_{d+1}\bigl(\frac{n+1}{n-1}\bigr)\bigr|}\\ & \leq\frac{1}{\bigl|T_{d+1}\bigl(\frac{n+1}{n-1}\bigr)\bigr|}\\ & =\frac{1}{T_{d+1}\bigl(\frac{n+1}{n-1}\bigr)},\end{aligned}$$ where the last two steps use (\[eqn:chebyshev-containment\]) and (\[eq:chebyshev-beyond-1\]), respectively. It is well known [@rivlin-book Theorem 1.10] that among all polynomials of degree at most $d$ that are bounded on $[-1,1]$ in absolute value by $1,$ the Chebyshev polynomial $T_{d}$ takes on the largest possible value at every point of $[1,\infty).$ Using this fact, it is straightforward to verify that Lemma \[lem:reciprocal\] gives the best possible bound on $\epsilon$ in terms of $n$ and $d.$ \[cor:reciprocal\]For any real $n>1,$ there is an $($explicitly given$)$ univariate polynomial $p$ of degree at most $\sqrt{2(n-1)}$ such that $$\begin{aligned} & \frac{1}{2t}\leq p(t)\leq\frac{1}{t}, & & 1\leq t\leq n.\end{aligned}$$ By Proposition \[prop:chebyshev-beyond-1\], $$\begin{aligned} T_{\lfloor\sqrt{2(n-1)}\rfloor+1}\left(\frac{n+1}{n-1}\right) & \geq5.\end{aligned}$$ As a result, it suffices to invoke Lemma \[lem:reciprocal\] with $d=\lfloor\sqrt{2(n-1)}\rfloor$. Approximating 1/*t*^*i*^ ------------------------ We now construct approximants for powers of the reciprocal function, focusing this time on absolute rather than relative error. Here, we are interested only in approximation in the neighborhood of $1$. In the following construction, increasing the approximant’s degree makes the neighborhood larger and the approximation more accurate. \[lem:powers-of-reciprocal\]Let $d\geq1$ be a given integer. Then for every integer $D\geq0,$ there is an $($explicitly given$)$ polynomial $p$ with $$\begin{aligned} & \left|\frac{1}{t^{d}}-p(t)\right|\leq|1-t|^{D+1}\binom{D+d}{d}d, & & t\in\left[\frac{d}{d+D},2-\frac{d}{d+D}\right],\label{eq:inverter-error}\\ & |p(t)|\leq\binom{D+d}{d}, & & t\in[0,2],\label{eq:inverter-norm}\\ & \deg p\leq D.\label{eq:inverter-degree}\end{aligned}$$ Define $$p(t)=\sum_{i=0}^{D}\binom{i+d-1}{i}(1-t)^{i}.$$ Then (\[eq:inverter-degree\]) is immediate. For (\[eq:inverter-norm\]), it suffices to observe that $$\begin{aligned} \sum_{i=0}^{D}\binom{i+d-1}{i} & =\binom{D+d}{D}\\ & =\binom{D+d}{d},\end{aligned}$$ where the first equality is well-known and can be verified by using Pascal’s triangle or by interpreting the left-hand side as the number of ways to distribute at most $D$ identical balls into $d$ distinct bins. It remains to settle (\[eq:inverter-error\]). For $0<t<2,$ we have the Maclaurin expansion $$\begin{aligned} \frac{1}{t^{d}} & =\left(\sum_{i=0}^{\infty}(1-t)^{i}\right)^{d}\\ & =\sum_{i=0}^{\infty}\binom{i+d-1}{i}(1-t)^{i}.\end{aligned}$$ Therefore, $$\begin{aligned} \left|\frac{1}{t^{d}}-p(t)\right| & =\left|\sum_{i=D+1}^{\infty}\binom{i+d-1}{i}(1-t)^{i}\right|\\ & \leq\sum_{i=D+1}^{\infty}\binom{i+d-1}{i}|1-t|^{i}\allowdisplaybreaks\\ & \le|1-t|^{D+1}\binom{D+d}{D+1}\sum_{i=0}^{\infty}\left(\frac{D+d}{D+1}\right)^{i}|1-t|^{i}\allowdisplaybreaks\\ & \le|1-t|^{D+1}\binom{D+d}{D+1}\sum_{i=0}^{\infty}\left(\frac{D}{D+1}\right)^{i}\\ & =|1-t|^{D+1}\binom{D+d}{d}d,\end{aligned}$$ where the fourth step is legitimate in view of the range of $t$ in (\[eq:inverter-error\]). \[subsec:characteristic-fn-interval\]Approximating the characteristic function of an interval --------------------------------------------------------------------------------------------- The following lemma is the last prerequisite to our construction of a low-degree approximant for the characteristic function of an interval. Without loss of generality, it suffices to consider the interval $[0,1].$ The lemma below *almost* solves our problem except that it gives a flat bound on the approximant’s value outside the interval, not taking into account how far one is from the interval. \[lem:or-continuous\]For any reals $n\geq1$ and $0<\epsilon<1/2,$ there is an $($explicitly given$)$ univariate polynomial $p$ such that $$\begin{aligned} & |p(t)-1|\leq\epsilon, & & t\in[0,1],\label{eq:or-continuous-p-starts-large}\\ & |p(t)|\leq1, & & t\in(1,2],\label{eq:or-continuous-p-is-bounded}\\ & |p(t)|\leq\epsilon, & & t\in(2,n],\label{eq:or-continuous-p-ends-small}\\ & \deg p=O\left(\sqrt{n}\log\frac{1}{\epsilon}\right).\label{eq:or-continuous-p-degree-logeps-factor-above-chebyshev}\end{aligned}$$ For $n<2,$ the lemma holds trivially with $p=1.$ In what follows, we treat the complementary case $n\geq2.$ Consider the univariate polynomial $$q(t)=T_{\lceil\sqrt{n}\rceil}\left(1+\frac{2-t}{n}\right).$$ Using $n\geq2,$ we obtain $$\begin{aligned} q([0,n]) & \subseteq\left[-1,T_{\lceil\sqrt{n}\rceil}\left(1+\frac{2}{n}\right)\right]\nonumber \\ & \subseteq\left[-1,\left(1+\frac{2}{n}+\sqrt{\left(1+\frac{2}{n}\right)^{2}-1}\right)^{\sqrt{n}+1}\right]\nonumber \\ & \subseteq\left[-1,\left(1+\frac{2}{n}+\sqrt{\frac{6}{n}}\right)^{\sqrt{n}+1}\right]\nonumber \\ & \subset\left[-1,\exp\left(\left(\frac{2}{n}+\sqrt{\frac{6}{n}}\right)\left(\sqrt{n}+1\right)\right)\right]\nonumber \\ & \subset[-1,\e^{7}-1],\label{eq:q-global-bound}\end{aligned}$$ where the first step is legitimate in view of (\[eqn:chebyshev-containment\]), (\[eq:chebyshev-at-1\]), and Fact \[fact:chebyshev-derivative\]; and the second step uses (\[eq:chebyshev-beyond-1\]). By Proposition \[prop:chebyshev-beyond-1\], $$\begin{aligned} \min_{0\leq t\leq1}q(t) & =\min_{1\leq t\leq2}\;T_{\lceil\sqrt{n}\rceil}\left(1+\frac{t}{n}\right)\nonumber \\ & \geq2.\label{eq:q-large-on-good-set}\end{aligned}$$ By (\[eqn:chebyshev-containment\]), $$\begin{aligned} \max_{2\leq t\leq n}|q(t)| & \leq\max_{0\leq t\leq1}\;|T_{\lceil\sqrt{n}\rceil}(t)|\nonumber \\ & \leq1.\label{eq:q-small-on-bad-set}\end{aligned}$$ In view of (\[eq:q-global-bound\])(\[eq:q-small-on-bad-set\]), the normalized polynomial $q^{*}(t)=(q(t)+1)/\e^{7}$ obeys $$\begin{aligned} q^{*}([0,n]) & \subseteq[0,1],\label{eq:q-star-range}\\ q^{*}([0,1]) & \subseteq[3\e^{-7},1],\\ q^{*}([2,n]) & \subseteq[0,2\e^{-7}].\end{aligned}$$ To complete the proof, we use a technique due to Buhrman et al. [@BNRW05robust]. Consider the univariate polynomial $$B_{d}(t)=\sum_{i=\lceil2.5\,\e^{-7}d\rceil}^{d}\binom{d}{i}t^{i}(1-t)^{i}.$$ In words, $B_{d}(t)$ is the probability of observing at least $2.5\,\e^{-7}d$ heads in a sequence of $d$ independent coin flips, each coming up heads with probability $t.$ For large enough $d=O(\log(1/\epsilon)),$ the Chernoff bound guarantees that $$\begin{aligned} B_{d}([0,1]) & \subseteq[0,1],\\ B_{d}([0,2\e^{-7}]) & \subseteq[0,\epsilon],\\ B_{d}([3\e^{-7},1]) & \subseteq[1-\epsilon,1].\label{eq:B-d-large}\end{aligned}$$ Now define $p(t)=B_{d}(q^{*}(t)).$ Then the degree bound (\[eq:or-continuous-p-degree-logeps-factor-above-chebyshev\]) is immediate, whereas the remaining properties (\[eq:or-continuous-p-starts-large\])(\[eq:or-continuous-p-ends-small\]) follow from (\[eq:q-star-range\])(\[eq:B-d-large\]). Finally, we are now in a position to construct the desired approximant for the characteristic function of an interval. As mentioned above, we may without loss of generality focus on the interval $[0,1].$ \[thm:sigmoid\]For all integers $n,d\geq0$ and all $0<\epsilon<1/2,$ there is an $($explicitly given$)$ univariate polynomial $p$ such that $$\begin{aligned} & |p(t)-1|\leq\epsilon, & & t\in[0,1],\label{eq:sigmoid-starts-small}\\ & |p(t)|\leq1+\epsilon, & & t\in(1,2],\label{eq:sigmoid-bounded}\\ & |p(t)|\leq\frac{\epsilon}{t^{d}} & & t\in(2,n],\label{eq:sigmoid-ends-small}\\ & \deg p=O\left(\sqrt{n}\left(d+\log\frac{1}{\epsilon}\right)\right).\label{eq:sigmoid-deg}\end{aligned}$$ For $n<2,$ the theorem holds trivially by taking $p=1.$ In what follows, we focus on the complementary case $n\geq2.$ Corollary \[cor:reciprocal\] gives an explicit univariate polynomial $p_{1}$ such that $$\begin{aligned} & \frac{1}{2(t+1)}\leq p_{1}(t)\leq\frac{1}{t+1}, & & 0\leq t\leq n,\label{eq:reciprocal-error-restated}\\ & \deg p_{1}\leq\sqrt{2n}.\label{eq:reciprocal-deg-restated}\end{aligned}$$ Let $D$ be an integer parameter to be chosen later, $D>5d.$ Then Lemma \[lem:powers-of-reciprocal\] provides an explicit polynomial $p_{2}$ such that $$\begin{aligned} & \left|\frac{1}{t^{d}}-p_{2}(t)\right|\leq\left(\frac{5}{6}\right)^{D+1}\binom{D+d}{d}d, & & t\in\left[\frac{1}{6},1\right],\label{eq:inverter-error-restated}\\ & |p_{2}(t)|\leq\binom{D+d}{d}, & & t\in[0,2],\label{eq:inverter-norm-restated}\\ & \deg p_{2}\leq D.\label{eq:inverter-degree-restated}\end{aligned}$$ As our last building block, Lemma \[lem:or-continuous\] constructs an explicit polynomial $p_{3}$ with $$\begin{aligned} & |p_{3}(t)-1|\leq\epsilon2^{-D-d}, & & t\in[0,1],\label{eq:or-continuous-p-starts-large-restated}\\ & |p_{3}(t)|\leq1, & & t\in(1,2],\label{eq:or-continuous-p-is-bounded-restated}\\ & |p_{3}(t)|\leq\epsilon2^{-D-d}, & & t\in(2,n],\label{eq:or-continuous-p-ends-small-restated}\\ & \deg p_{3}=O\left(\sqrt{n}\left(D+d+\log\frac{1}{\epsilon}\right)\right).\label{eq:or-continuous-p-degree-logeps-factor-above-chebyshev-restated}\end{aligned}$$ In the rest of the proof, we will show that the conclusion of the theorem holds for the polynomial $$p(t)=p_{1}(t)^{d}p_{2}(p_{1}(t))p_{3}(t).$$ To begin with, $$\begin{aligned} \max_{0\leq t\leq1}|p(t)-1| & \leq\max_{0\leq t\leq1}(1+|p_{1}(t)^{d}p_{2}(p_{1}(t))-1|)\cdot(1+|1-p_{3}(t)|)-1\nonumber \\ & \leq\left(1+\frac{\epsilon}{2}\right)\max_{0\leq t\leq1}(1+|p_{1}(t)^{d}p_{2}(p_{1}(t))-1|)-1\nonumber \\ & \leq\left(1+\frac{\epsilon}{2}\right)\left(1+\max_{1/4\leq t\leq1}|t^{d}p_{2}(t)-1|\right)-1\nonumber \\ & \leq\left(1+\frac{\epsilon}{2}\right)\left(1+\max_{1/4\leq t\leq1}\left|p_{2}(t)-\frac{1}{t^{d}}\right|\right)-1\nonumber \\ & \leq\left(1+\frac{\epsilon}{2}\right)\left(1+\left(\frac{5}{6}\right)^{D+1}\binom{D+d}{d}d\right)-1,\label{eq:sigmoid-starts-small-restated}\end{aligned}$$ where the first step uses the inequality $|ab-1|\leq(1+|a-1|)(1+|b-1|)-1$ for any real $a,b;$ the second step is valid by (\[eq:or-continuous-p-starts-large-restated\]); the third applies (\[eq:reciprocal-error-restated\]); and the final step is legitimate by (\[eq:inverter-error-restated\]). Continuing, $$\begin{aligned} \max_{1\leq t\leq2}|p(t)| & =\max_{1\leq t\leq2}|p_{1}(t)^{d}p_{2}(p_{1}(t))p_{3}(t)|\nonumber \\ & \leq\max_{1\leq t\leq2}|p_{1}(t)^{d}p_{2}(p_{1}(t))|\nonumber \\ & \leq\max_{1/6\leq t\leq1/2}|t^{d}p_{2}(t)|\nonumber \\ & \leq\max_{1/6\leq t\leq1/2}|t^{d}p_{2}(t)-1|+1\nonumber \\ & \leq\max_{1/6\leq t\leq1/2}\left|p_{2}(t)-\frac{1}{t^{d}}\right|+1\nonumber \\ & \leq1+\left(\frac{5}{6}\right)^{D+1}\binom{D+d}{d}d,\label{eq:sigmoid-bounded-restated}\end{aligned}$$ where the second step uses (\[eq:or-continuous-p-is-bounded-restated\]), the third step applies (\[eq:reciprocal-error-restated\]), the fourth step is immediate from the triangle inequality, and the last step follows from (\[eq:inverter-error-restated\]). Moreover, $$\begin{aligned} \max_{2\leq t\leq n}|t^{d}p(t)| & \leq\max_{2\leq t\leq n}|t^{d}p_{1}(t)^{d}|\cdot\max_{2\leq t\leq n}|p_{2}(p_{1}(t))|\cdot\max_{2\leq t\leq n}|p_{3}(t)|\nonumber \\ & \leq\max_{2\leq t\leq n}|t^{d}p_{1}(t)^{d}|\cdot\max_{2\leq t\leq n}|p_{2}(p_{1}(t))|\cdot\epsilon2^{-D-d}\nonumber \\ & \leq\max_{2\leq t\leq n}|p_{2}(p_{1}(t))|\cdot\epsilon2^{-D-d}\nonumber \\ & \leq\max_{0\leq t\leq1/3}|p_{2}(t)|\cdot\epsilon2^{-D-d}\nonumber \\ & \le\binom{D+d}{d}\cdot\epsilon2^{-D-d}\nonumber \\ & \leq\epsilon,\label{eq:sigmoid-ends-small-restated}\end{aligned}$$ where the second step is legitimate by (\[eq:or-continuous-p-ends-small-restated\]), the third and fourth steps use (\[eq:reciprocal-error-restated\]), and the fifth step is immediate from (\[eq:inverter-norm-restated\]). Finally, (\[eq:reciprocal-deg-restated\]), (\[eq:inverter-degree-restated\]), and (\[eq:or-continuous-p-degree-logeps-factor-above-chebyshev-restated\]) imply that $$\deg p=O\left(\sqrt{n}\left(D+d+\log\frac{1}{\epsilon}\right)\right).\label{eq:sigmoid-deg-restated}$$ Now the claimed bounds (\[eq:sigmoid-starts-small\])(\[eq:sigmoid-deg\]) in the theorem statement follow immediately from (\[eq:sigmoid-starts-small-restated\])(\[eq:sigmoid-deg-restated\]) by taking $$\begin{aligned} D & =c\left\lceil d+\log\frac{1}{\epsilon}\right\rceil \end{aligned}$$ for a sufficiently large absolute constant $c>1.$ Proof of the extension theorem ------------------------------ Using the approximant constructed in Theorem \[thm:sigmoid\], we now prove the extension theorem. We restate it below for the reader’s convenience. Let $f\colon\zoo_{\leq m}^{N}\to[-1,1]$ be given, where $N\geq m\geq0$ are integers. For integers $n\geq m,$ define $F_{n}\colon\zoo_{\leq n}^{N}\to[-1,1]$ by $$F_{n}(x)=\begin{cases} f(x) & \text{if \ensuremath{|x|\leq m,}}\\ 0 & \text{otherwise.} \end{cases}$$ Then for some absolute constant $C>1$ and all $\epsilon,\delta\in(0,1/2)$ and $n\geq m,$ $$\deg_{\epsilon+\delta}(F_{n})\leq C\sqrt{\frac{n}{m+1}}\cdot\left(\deg_{\epsilon}(F_{2m})+\log\frac{1}{\delta}\right).\label{eq:degeps-Fn}$$ To simplify the presentation, we first settle two degenerate cases. For $m=0,$ consider the polynomial $$T(t)=\left(T_{\lceil\sqrt{n}\rceil}\left(1+\frac{1-t}{n}\right)\right)^{\lceil\log\frac{1}{\delta}\rceil}.$$ Then $T(0)\geq1/\delta$ by Proposition \[prop:chebyshev-beyond-1\], and $\max_{1\leq t\leq n}|T(t)|\leq1$ by (\[eqn:chebyshev-containment\]). Therefore, in this case $F_{n}$ is approximated pointwise within $\delta$ by the degree-$O(\sqrt{n}\log(1/\delta))$ polynomial $F_{n}(0^{N})T(|x|)/T(0).$ Another degenerate possibility is $n\leq2m,$ in which case $\deg_{\epsilon}(F_{n})\leq\deg_{\epsilon}(F_{2m})$ and the theorem holds trivially. In what follows, we focus on the general case when $$\begin{aligned} & m\geq1,\\ & n>2m.\end{aligned}$$ Abbreviate $d=\max\{\deg_{\epsilon}(F_{2m}),1\}$. By Fact \[fact:trivial-upper-bound-on-deg\], $$1\leq d\leq2m.\label{eq:extension-thm-d-at-most-m}$$ Fix a polynomial $\phi\colon\zoo^{N}\to\Re$ such that $$\begin{aligned} & |F_{2m}(x)-\phi(x)|\leq\epsilon, & & x\in\zoo_{\leq2m}^{N},\label{eq:extension-phi-approximates-f}\\ & \deg\phi\leq d.\label{eq:extension-phi-deg}\end{aligned}$$ Let $0<\alpha<1/2$ be a parameter to be chosen later. Then Theorem \[thm:sigmoid\] gives an explicit univariate polynomial $p$ such that $$\begin{aligned} & |p(t)-1|\leq\alpha, & & t\in[0,1],\label{eq:sigmoid-starts-small-apply}\\ & |p(t)|\leq1+\alpha, & & t\in(1,2],\label{eq:sigmoid-bounded-apply}\\ & |p(t)|\leq\frac{\alpha}{t^{d}} & & t\in\left(2,\frac{n}{m}\right],\label{eq:sigmoid-ends-small-apply}\\ & \deg p=O\left(\sqrt{\frac{n}{m}}\left(d+\log\frac{1}{\alpha}\right)\right).\label{eq:sigmoid-deg-apply}\end{aligned}$$ Consider the polynomial $\Phi\colon\zoo^{N}\to\Re$ given by $$\Phi(x)=\phi(x)\;p\!\left(\frac{|x|}{m}\right).$$ By (\[eq:extension-phi-deg\]) and (\[eq:sigmoid-deg-apply\]), $$\deg\Phi=O\left(\sqrt{\frac{n}{m}}\left(d+\log\frac{1}{\alpha}\right)\right).\label{eq:deg-Phi}$$ As the notation suggests, $\Phi$ is meant to be an extension of the approximant $\phi$ to inputs $x\in\zoo^{N}$ of Hamming weight up to $n.$ To analyze the accuracy of this new approximant, we will examine three cases depending on the Hamming weight $|x|.$ To start with, $$\begin{aligned} \max_{|x|\leq m}|F_{n}(x)-\Phi(x)| & =\max_{|x|\leq m}|F_{2m}(x)-\Phi(x)|\nonumber \\ & \leq\max_{|x|\leq m}\{|F_{2m}(x)-\phi(x)|+|\phi(x)-\Phi(x)|\}\nonumber \\ & \leq\epsilon+\max_{|x|\leq m}|\phi(x)-\Phi(x)|\nonumber \\ & \leq\epsilon+\max_{|x|\leq m}|\phi(x)|\,\max_{0\leq t\leq1}|1-p(t)|\nonumber \\ & \leq\epsilon+(1+\epsilon)\max_{0\leq t\leq1}|1-p(t)|\nonumber \\ & \leq\epsilon+(1+\epsilon)\cdot\alpha,\label{eq:Phi-approximates-on-small-inputs}\end{aligned}$$ where the third and fifth steps use (\[eq:extension-phi-approximates-f\]), and the last step uses (\[eq:sigmoid-starts-small-apply\]). Continuing, $$\begin{aligned} \max_{m<|x|\leq2m}|F_{n}(x)-\Phi(x)| & =\max_{m<|x|\leq2m}|\Phi(x)|\nonumber \\ & \leq\max_{m<|x|\leq2m}|\phi(x)|\;\max_{1<t\leq2}|p(t)|\nonumber \\ & \leq\max_{m<|x|\leq2m}(|F_{2m}(x)|+\epsilon)\,\max_{1<t\leq2}|p(t)|\nonumber \\ & \leq\epsilon\cdot(1+\alpha),\label{eq:Phi-approximation-med-inputs}\end{aligned}$$ where the last two steps use (\[eq:extension-phi-approximates-f\]) and (\[eq:sigmoid-bounded-apply\]), respectively. Finally, $$\begin{aligned} \max_{2m<|x|\leq n} & |F_{n}(x)-\Phi(x)|\nonumber \\ & =\max_{2m<|x|\leq n}|\Phi(x)|\nonumber \\ & =\max_{2m<|x|\leq n}|\phi(x)|\;p\left(\frac{|x|}{m}\right)\nonumber \\ & \leq\max_{2m<|x|\leq n}|\phi(x)|\cdot\alpha\cdot\left(\frac{m}{|x|}\right)^{d}\allowdisplaybreaks\nonumber \\ & \leq\max_{2m<|x|\leq n}\left\{ 2^{d}\binom{\lceil|x|/\lfloor2m/d\rfloor\rceil}{d}\max_{|x'|\leq2m}|\phi(x')|\cdot\alpha\cdot\left(\frac{m}{|x|}\right)^{d}\right\} \nonumber \\ & \leq\max_{2m<t\leq n}\left\{ 2^{d}\binom{\lceil t/\lfloor2m/d\rfloor\rceil}{d}(1+\epsilon)\cdot\alpha\cdot\left(\frac{m}{t}\right)^{d}\right\} \nonumber \\ & \leq(4\e)^{d}(1+\epsilon)\cdot\alpha,\label{eq:Phi-approximation-large-inputs}\end{aligned}$$ where the third step uses (\[eq:sigmoid-ends-small-apply\]), the fourth step applies (\[eq:extension-thm-d-at-most-m\]) and the generalized extrapolation lemma (Lemma \[lem:extrapolation-generalized\]), the fifth step follows from (\[eq:extension-phi-approximates-f\]), and the last step uses (\[eq:entropy-bound-binomial\]) and (\[eq:extension-thm-d-at-most-m\]). Now (\[eq:degeps-Fn\]) follows from (\[eq:deg-Phi\])(\[eq:Phi-approximation-large-inputs\]) by taking $\alpha=\delta(4\e)^{-d-1}.$ Symmetric functions =================== In this section, we study the approximation of symmetric functions. This class includes $\AND_{n}$ and $\OR_{n}$, which are fundamental building blocks of our constructions in the rest of the paper. Our result here is as follows. \[thm:SYMMETRIC\]Let $f\colon\zoon\to[-1,1]$ be an arbitrary symmetric function. Let $k$ be a nonnegative integer such that $f$ is constant on inputs of Hamming weight in $(k,n-k).$ Then for $0<\epsilon<1/2,$ $$\deg_{\epsilon}(f)=O\left(\sqrt{nk}+\sqrt{n\log\frac{1}{\epsilon}}\right).\label{eq:degeps-f-symmetric-1-1}$$ Moreover, the approximating polynomial is given explicitly in each case. Theorem \[thm:SYMMETRIC\] is tight [@sherstov07inclexcl-ccc] for every $\epsilon\in[1/2^{n},1/3]$ and every symmetric function $f\colon\zoon\to\zoo$, with the obvious exception of the constant functions $f=0$ and $f=1.$ Prior to our work, de Wolf [@de-wolf08approx-degree] proved the upper bound (\[eq:degeps-f-symmetric-1-1\]) by giving an $\epsilon$-error quantum query algorithm for any symmetric function $f.$ The novelty of Theorem \[thm:SYMMETRIC\] is the construction of an explicit, closed-form approximating polynomial that achieves de Wolf’s upper bound. We give three proofs of Theorem \[thm:SYMMETRIC\], corresponding to Sections 4.14.3 below. Approximation using the extension theorem ----------------------------------------- Our first proof of Theorem \[thm:SYMMETRIC\] is based on the extension theorem, and is the shortest of the three. The centerpiece of the proof is the following technical lemma, in which we construct a closed-form approximant for any function supported on inputs of low Hamming weight. \[lem:small-support\]Let $f\colon\zoon\to[-1,1]$ be given. Let $k$ be a nonnegative integer such that $f(x)=0$ for $|x|>k.$ Then for $0<\epsilon<1/2,$ $$\deg_{\epsilon}(f)=O\left(\sqrt{nk}+\sqrt{n\log\frac{1}{\epsilon}}\right).$$ Moreover, the approximating polynomial is given explicitly in each case. Abbreviate $$m=\left\lceil k+\log\frac{1}{\epsilon}\right\rceil .$$ If $m\geq n,$ the bound in the theorem statement follows trivially from $\deg_{0}(f)\leq n.$ In the rest of the proof, we focus on the complementary case $m<n.$ For $i\geq m,$ define $F_{i}\colon\zoon_{\leq i}\to[-1,1]$ by $$F_{i}(x)=\begin{cases} f(x) & \text{if }|x|\leq m,\\ 0 & \text{otherwise.} \end{cases}$$ Then $$\begin{aligned} \deg_{\epsilon}(f) & =\deg_{\epsilon}(F_{n})\\ & \leq\sqrt{\frac{n}{m}}\cdot O\left(\deg_{0}(F_{2m})+\log\frac{1}{\epsilon}\right)\\ & \leq\sqrt{\frac{n}{m}}\cdot O\left(2m+\log\frac{1}{\epsilon}\right)\\ & =O\left(\sqrt{nk}+\sqrt{n\log\frac{1}{\epsilon}}\right),\end{aligned}$$ where the first step uses $f=F_{n},$ the second step applies the extension theorem (Theorem \[thm:extension\]), and the third step is valid by Fact \[fact:trivial-upper-bound-on-deg\]. Moreover, the approximating polynomial is given explicitly because the extension theorem and Fact \[fact:trivial-upper-bound-on-deg\] are fully constructive. We are now in a position to prove the claimed result on the approximation of arbitrary symmetric functions. \[thm:small-support\]Let $f\colon\zoon\to[-1,1]$ be given. Let $k$ be a nonnegative integer such that $f$ is constant on inputs of Hamming weight in $(k,n-k).$ Then for $0<\epsilon<1/2,$ $$\deg_{\epsilon}(f)=O\left(\sqrt{nk}+\sqrt{n\log\frac{1}{\epsilon}}\right).$$ Moreover, the approximating polynomial is given explicitly in each case. A powerful feature of Theorem \[thm:small-support\] is that the function of interest is only assumed to be symmetric on inputs of Hamming weight in $(k,n-k).$ In particular, Theorem \[thm:small-support\] is significantly more general than Theorem \[thm:SYMMETRIC\]. If $k\geq n/2,$ the theorem follows from the trivial bound $\deg_{0}(f)\leq n.$ For the complementary case $k<n/2$, write $$f(x_{1},\ldots,x_{n})=\lambda+f'(x_{1},\ldots,x_{n})+f''(\overline{x_{1}},\ldots,\overline{x_{n}}),$$ where $\lambda\in[-1,1]$ and $f',f''\colon\zoon\to[-2,2]$ are functions that vanish on $\zoon_{>k}.$ Then $$\begin{aligned} \deg_{\epsilon}(f) & \leq\max\{\deg_{\epsilon/2}(f'),\deg_{\epsilon/2}(f'')\}\\ & \leq\max\left\{ \deg_{\epsilon/4}\!\left(\frac{f'}{2}\right),\deg_{\epsilon/4}\!\left(\frac{f''}{2}\right)\right\} \\ & =O\left(\sqrt{nk}+\sqrt{n\log\frac{1}{\epsilon}}\right),\end{aligned}$$ where the last step uses Lemma \[lem:small-support\]. Approximation from first principles ----------------------------------- We now present our second proof of Theorem \[thm:SYMMETRIC\]. This proof proceeds from first principles, using Chebyshev polynomials as its only ingredient. To convey the construction as clearly as possible, we first present an approximant for the simplest and most important symmetric function, $\AND_{n}$. For this, we adopt the strategy of previous constructions [@kahn96incl-excl; @sherstov07inclexcl-ccc], whereby one first zeroes out as many of the integer points $n-1,n-2,n-3,\ldots$ as possible and then uses a Chebyshev polynomial to approximate $\AND_{n}$ on the remaining points of $\{0,1,2,\ldots,n\}.$ We depart from the previous work in the implementation of the first step. Specifically, we produce the zeroes using a product of Chebyshev polynomials, each of which is stretched and shifted so as to obtain an extremum at $n$ and a root at one of the points $n-1,n-2,n-3,\ldots.$ The use of Chebyshev polynomials allows us to avoid explosive growth at the nonzeroes, thereby eliminating a key source of inefficiency in [@kahn96incl-excl; @sherstov07inclexcl-ccc]. The lemma below shows how to produce a single zero, at any given point $m$. \[lem:eliminate-points-close-to-n\]Let $n$ and $m$ be given integers, $0\leq m<n.$ Then there is a univariate polynomial $T_{n,m}$ such that $$\begin{aligned} & T_{n,m}(n)=1,\label{eq:Tnm-at-n}\\ & T_{n,m}(m)=0,\label{eq:Tnm-at-m}\\ & |T_{n,m}(t)|\leq1, & & 0\leq t\leq n,\label{eq:Tnm-bounded}\\ & \deg(T_{n,m})\leq\left\lceil \frac{\pi}{4}\sqrt{\frac{n}{n-m}}\right\rceil .\label{eq:deg-Tnm}\end{aligned}$$ As mentioned above, the construction involves starting with a Chebyshev polynomial and stretching and shifting it so as to move an extremum to $n$ and a root to $m.$ In more detail, let $$d=\left\lceil \frac{\pi}{4}\sqrt{\frac{n}{n-m}}\right\rceil .$$ Consider the linear map $L$ that sends $$\begin{aligned} & L(n)=1,\label{eq:L-at-n}\\ & L(m)=\cos\left(\frac{\pi}{2d}\right).\label{eq:L-at-m}\end{aligned}$$ Observe that under $L$, the length of any given interval of the real line changes by a factor of $$\begin{aligned} \frac{1}{n-m}\left(1-\cos\left(\frac{\pi}{2d}\right)\right) & \leq\frac{1}{n-m}\left(1-\left(1-\frac{\pi^{2}}{8d^{2}}\right)\right)\\ & =\frac{\pi^{2}}{8d^{2}(n-m)}\\ & \leq\frac{2}{n},\end{aligned}$$ where the first step uses $\cos x\geq1-\frac{1}{2}x^{2}$ for $x\in\Re.$ In particular, $$\begin{aligned} L([0,n]) & \subseteq\left[L(n)-\frac{2}{n}\cdot n,L(n)\right]\nonumber \\ & \subseteq[-1,1].\label{eq:L-maps-into-01}\end{aligned}$$ We now show that the sought properties (\[eq:Tnm-at-n\])(\[eq:deg-Tnm\]) hold for the polynomial $T_{n,m}(t)=T_{d}(L(t)),$ where $T_{d}$ denotes as usual the Chebyshev polynomial of degree $d.$ To start with, $$\begin{aligned} T_{n,m}(n) & =T_{d}(L(n))\\ & =T_{d}(1)\\ & =1,\end{aligned}$$ where the last two steps use (\[eq:L-at-n\]) and (\[eq:chebyshev-at-1\]), respectively. Similarly, $$\begin{aligned} T_{n,m}(m) & =T_{d}(L(m))\\ & =T_{d}\left(\cos\left(\frac{\pi}{2d}\right)\right)\\ & =\cos\left(\frac{\pi}{2}\right)\\ & =0,\end{aligned}$$ where the second and third steps follow from (\[eq:L-at-m\]) and (\[eq:Chebyshev-definition\]), respectively. Continuing, $$\begin{aligned} T_{n,m}([0,n]) & =T_{d}(L([0,n]))\\ & \subseteq T_{d}([-1,1])\\ & \subseteq[-1,1],\end{aligned}$$ where the last two steps follow from (\[eq:L-maps-into-01\]) and (\[eqn:chebyshev-containment\]), respectively. Finally, the degree bound (\[eq:deg-Tnm\]) is immediate from the choice of $d.$ We now obtain the desired approximant for AND and OR, using the two-stage approach described earlier. The reader interested exclusively in the general case may wish to skip to Theorem \[lem:orcomplexity-symmetric\]. \[thm:AND\]For some constant $c>0$ and all integers $n\geq1$ and $d\geq0,$ there is an $($explicitly given$)$ univariate polynomial $p$ such that $$\begin{aligned} & p(n)=1,\label{eq:p-at-n}\\ & |p(t)|\leq\exp\left(-\frac{cd^{2}}{n}\right), & & t=0,1,2,\ldots,n-1,\label{eq:p-at-i}\\ & |p(t)|\leq1, & & t\in[0,n],\label{eq:p-bounded}\\ & \deg p\leq d.\label{eq:deg-p}\end{aligned}$$ In particular, $$\begin{aligned} & E(\AND_{n},d)\leq\frac{1}{2}\exp\left(-\frac{c}{2}\cdot\frac{d^{2}}{n}\right), & & d=0,1,2,3,\ldots,\label{eq:E-ANDn}\\ & \deg_{\epsilon}(\AND_{n})\leq O\left(\sqrt{n\log\frac{1}{\epsilon}}\right), & & 0<\epsilon<\frac{1}{2},\label{eq:degeps-ANDn}\end{aligned}$$ and analogously $$\begin{aligned} & E(\OR_{n},d)\leq\frac{1}{2}\exp\left(-\frac{c}{2}\cdot\frac{d^{2}}{n}\right), & & d=0,1,2,3,\ldots\label{eq:E-ORn}\\ & \deg_{\epsilon}(\OR_{n})\leq O\left(\sqrt{n\log\frac{1}{\epsilon}}\right), & & 0<\epsilon<\frac{1}{2}.\label{eq:degeps-ORn}\end{aligned}$$ For $d\geq n,$ we may simply take $p(t)=t(t-1)(t-2)\cdots(t-n+1)/n!.$ In what follows, we focus on the construction of $p$ for $d<n.$ Let $\ell,r$ be integer parameters to be chosen later, where $1\leq\ell\leq n-1$ and $1\leq r\leq n.$ We define $$p(t)=\frac{T_{r}(t/(n-\ell))}{T_{r}(n/(n-\ell))}\prod_{i=n-\ell+1}^{n-1}T_{n,i}(t),$$ where $T_{n,i}$ is as constructed in Lemma \[lem:eliminate-points-close-to-n\], and $T_{r}$ stands as usual for the Chebyshev polynomial of degree $r.$ By (\[eq:Tnm-at-n\]) and (\[eq:Tnm-at-m\]), $$\begin{aligned} p(n) & =1,\label{eq:p-at-n-restated}\\ p(t) & =0, & & t=n-\ell+1,\ldots,n-1.\label{eq:p-at-t}\end{aligned}$$ Moreover, $$\begin{aligned} \max_{0\leq t\leq n-\ell}|p(t)| & =\max_{0\leq t\leq n-\ell}\;\;\frac{|T_{r}(t/(n-\ell))|}{|T_{r}(n/(n-\ell))|}\prod_{i=n-\ell+1}^{n-1}|T_{n,i}(t)|\nonumber \\ & \leq\frac{1}{|T_{r}(n/(n-\ell))|}\nonumber \\ & \leq\frac{1}{\max\left\{ 1+\frac{r^{2}\ell}{n},2^{r\sqrt{\ell/n}-1}\right\} }\nonumber \\ & \leq\frac{1}{\min\left\{ \exp\left(\frac{r^{2}\ell}{3n}\right),\exp\left(\frac{r\sqrt{\ell}}{3\sqrt{n}}\right)\right\} },\label{eq:p-intermediate}\end{aligned}$$ where the second step uses (\[eqn:chebyshev-containment\]) and (\[eq:Tnm-bounded\]); the third step follows from Proposition \[prop:chebyshev-beyond-1\]; and the last step uses $1+x\geq\exp(x/3)$ for $0\leq x\leq4,$ and $2^{\sqrt{x}-1}\geq\exp(\sqrt{x}/3)$ for $x\geq4.$ Next, $$\begin{aligned} \max_{0\leq t\leq n}|p(t)| & =\max_{0\leq t\leq n}\;\;\frac{|T_{r}(t/(n-\ell))|}{|T_{r}(n/(n-\ell))|}\prod_{i=n-\ell+1}^{n-1}|T_{n,i}(t)|\nonumber \\ & \leq\max_{0\leq t\leq n}\;\;\frac{|T_{r}(t/(n-\ell))|}{|T_{r}(n/(n-\ell))|}\nonumber \\ & \leq1,\label{eq:p-bounded-on-0-to-n}\end{aligned}$$ where the second inequality uses (\[eq:Tnm-bounded\]), and the third inequality follows from (\[eqn:chebyshev-containment\]), (\[eq:chebyshev-at-1\]), and Fact \[fact:chebyshev-derivative\]. Finally, $$\begin{aligned} \deg p & \leq r+\sum_{i=n-\ell+1}^{n-1}\deg(T_{n,i})\nonumber \\ & \leq r+\sum_{i=1}^{\ell-1}\left(\frac{\pi}{4}\sqrt{\frac{n}{i}}+1\right)\nonumber \\ & \leq r+\ell-1+\frac{\pi\sqrt{n}}{4}\int_{0}^{\ell-1}\frac{dt}{\sqrt{t}}\allowdisplaybreaks\nonumber \\ & =r+\ell-1+\frac{\pi\sqrt{n(\ell-1)}}{2}\nonumber \\ & \leq r+3\sqrt{n(\ell-1)},\label{eq:deg-p-restated}\end{aligned}$$ where the second step uses (\[eq:deg-Tnm\]). Now (\[eq:p-at-n\])(\[eq:deg-p\]) follow from (\[eq:p-at-n-restated\])(\[eq:deg-p-restated\]) by setting $r=\lceil d/2\rceil$ and $\ell=\lfloor d^{2}/(36n)\rfloor+1.$ The remaining claims in the theorem statement follow in a straightforward manner from (\[eq:p-at-n\])(\[eq:deg-p\]). For (\[eq:E-ANDn\]), we have $$\begin{aligned} E(\AND_{n},d) & \leq\max_{x\in\zoon}\left|\AND_{n}(x)-\frac{p(\sum_{i=1}^{n}x_{i})}{1+\exp(-cd^{2}/n)}\right|\\ & \leq\frac{\exp(-cd^{2}/n)}{1+\exp(-cd^{2}/n)}\\ & \leq\frac{1}{2}\exp\left(-\frac{c}{2}\cdot\frac{d^{2}}{n}\right),\end{aligned}$$ where the last step uses $a/(1+a)\leq\sqrt{a}/2$ for any $a\geq0.$ This in turn settles (\[eq:E-ORn\]) since $\OR_{n}(x)=1-\AND_{n}(1-x_{1},\ldots,1-x_{n}).$ Finally, (\[eq:degeps-ANDn\]) and (\[eq:degeps-ORn\]) are immediate from (\[eq:E-ANDn\]) and (\[eq:E-ORn\]), respectively. To generalize Theorem \[thm:AND\] to an arbitrary symmetric function $f$, it is helpful to think of $f$ as a linear combination of the characteristic functions of individual levels of the Boolean hypercube. Specifically, define $\EXACT_{n,k}\colon\zoon\to\zoo$ by $$\EXACT_{n,k}(x)=\begin{cases} 1 & \text{if }|x|=k,\\ 0 & \text{otherwise.} \end{cases}$$ In this notation, Theorem \[thm:AND\] treats the special case $\AND_{n}=\EXACT_{n,n}$. The technique of that theorem is easily adapted to yield the following more general result. \[thm:EXACT\]For any $0<\epsilon<1/2$ and any integers $n\geq m\geq k\geq0,$ there is a univariate polynomial $p$ such that $$\begin{aligned} & p(|x|)=\EXACT_{n,n-k}(x), & & |x|\leq m,\\ & p(|x|)=\EXACT_{n,n-k}(x), & & |x|\geq n-m,\\ & |p(|x|)-\EXACT_{n,n-k}(x)|\leq\epsilon, & & x\in\zoon,\\ & \deg p=O\left(\sqrt{nm}+\sqrt{n\log\frac{1}{\epsilon}}\right).\end{aligned}$$ Define $$\begin{aligned} \ell & =\left\lceil m+\log\frac{2}{\epsilon}\right\rceil ,\label{eq:arb-symm-ell}\\ r & =\left\lceil \sqrt{n\log\frac{2}{\epsilon}}\right\rceil .\label{eq:arb-symm-r}\end{aligned}$$ If $\ell\geq n/2,$ the theorem holds trivially for the degree-$n$ polynomial $$p(t)=\prod_{\substack{i=0\\ i\ne n-k } }^{n}\frac{t-i}{n-k-i}.$$ In the complementary case $\ell<n/2,$ define $$\begin{gathered} p(t)=\frac{T_{r}(t/(n-\ell))}{T_{r}((n-k)/(n-\ell))}\cdot\prod_{i=0}^{\ell}T_{n-k,i}(t)\cdot\prod_{i=n-\ell}^{n-k-1}T_{n-k,i}(t)\\ \times\prod_{i=n-k+1}^{n}(1-T_{i,n-k}(t)^{2}),\end{gathered}$$ where $T_{n-k,i}$ and $T_{i,n-k}$ are as constructed in Lemma \[lem:eliminate-points-close-to-n\], and $T_{r}$ denotes as usual the Chebyshev polynomial of degree $r.$ Then (\[eq:Tnm-at-n\]) and (\[eq:Tnm-at-m\]) imply that $$\begin{aligned} p(t) & =0, & & t\in\{0,1,\ldots,\ell\},\label{eq:arb-symm-low}\\ p(t) & =0, & & t\in\{n-\ell,\ldots,n-1,n\}\setminus\{n-k\},\label{eq:arb-symm-high}\end{aligned}$$ and $$\begin{aligned} p(n-k) & =\prod_{i=0}^{\ell}T_{n-k,i}(n-k)\cdot\prod_{i=n-\ell}^{n-k-1}T_{n-k,i}(n-k)\nonumber \\ & \qquad\qquad\qquad\qquad\qquad\qquad\qquad\times\prod_{i=n-k+1}^{n}(1-T_{i,n-k}(n-k)^{2})\nonumber \\ & =\prod_{i=0}^{\ell}1\cdot\prod_{i=n-\ell}^{n-k-1}1\cdot\prod_{i=n-k+1}^{n}(1-0^{2})\nonumber \\ & =1.\label{eq:arb-symm-at-n-k}\end{aligned}$$ Moreover, $$\begin{aligned} \max_{0\leq t\leq n-\ell}|p(t)| & =\max_{0\leq t\leq n-\ell}\;\left|\frac{T_{r}(t/(n-\ell))}{T_{r}((n-k)/(n-\ell))}\cdot\prod_{i=0}^{\ell}T_{n-k,i}(t)\right.\nonumber \\ & \qquad\qquad\qquad\times\left.\prod_{i=n-\ell}^{n-k-1}T_{n-k,i}(t)\cdot\prod_{i=n-k+1}^{n}(1-T_{i,n-k}(t)^{2})\right|\nonumber \\ & \leq\max_{0\leq t\leq n-\ell}\left|\frac{T_{r}(t/(n-\ell))}{T_{r}((n-k)/(n-\ell))}\right|\nonumber \\ & \leq\frac{1}{|T_{r}((n-k)/(n-\ell))|}\nonumber \\ & \leq2^{-r\sqrt{(\ell-k)/n}+1}\nonumber \\ & \leq\epsilon,\label{eq:arb-symm-intermediate}\end{aligned}$$ where the second step uses (\[eq:Tnm-bounded\]), the third step uses (\[eqn:chebyshev-containment\]), the fourth step applies Proposition \[prop:chebyshev-beyond-1\], and the final step substitutes the parameters (\[eq:arb-symm-ell\]) and (\[eq:arb-symm-r\]). Finally, $$\begin{aligned} \deg p & \leq r+\sum_{i=0}^{\ell}\deg(T_{n-k,i})+\sum_{i=n-\ell}^{n-k-1}\deg(T_{n-k,i})+2\sum_{i=n-k+1}^{n}\deg(T_{i,n-k})\nonumber \\ & \leq r+\sum_{i=0}^{\ell}\left(\frac{\pi}{4}\sqrt{\frac{n-k}{n-k-i}}+1\right)+\sum_{i=1}^{\ell-k}\left(\frac{\pi}{4}\sqrt{\frac{n-k}{i}}+1\right)\nonumber \\ & \qquad\qquad\qquad\qquad+2\sum_{i=1}^{k}\left(\frac{\pi}{4}\sqrt{\frac{n-k+i}{i}}+1\right)\nonumber \\ & \leq r+3\ell+\pi\sum_{i=1}^{\ell+1}\sqrt{\frac{n}{i}}\allowdisplaybreaks\nonumber \\ & \leq r+3\ell+\pi\sqrt{n}\int_{0}^{\ell+1}\frac{dt}{\sqrt{t}}\nonumber \\ & =r+3\ell+2\pi\sqrt{n(\ell+1)}\nonumber \\ & =O\left(\sqrt{nm}+\sqrt{n\log\frac{1}{\epsilon}}\right),\label{eq:arb-symm-degree}\end{aligned}$$ where the second and third steps use (\[eq:Tnm-bounded\]) and $k<\ell<n-k$, respectively. In view of (\[eq:arb-symm-low\])(\[eq:arb-symm-degree\]), the proof is complete. We are now in a position to handle arbitrary symmetric functions by expressing them as a linear combination of $\EXACT_{n,i}$ for $i=0,1,2,\ldots,n.$ This result provides a new proof of Theorem \[thm:SYMMETRIC\]. \[thm:SYMMETRIC-proof-first-principles\]Let $f\colon\zoon\to[-1,1]$ be an arbitrary symmetric function. Let $k$ be a nonnegative integer such that $f$ is constant on inputs of Hamming weight in $(k,n-k).$ Then for $0<\epsilon<1/2,$ $$\deg_{\epsilon}(f)=O\left(\sqrt{nk}+\sqrt{n\log\frac{1}{\epsilon}}\right).\label{eq:degeps-f-symmetric-first-principles}$$ More precisely, there is an $($explicitly given$)$ polynomial $\tilde{f}\colon\zoon\to\Re$ such that $$\begin{aligned} & f(x)=\tilde{f}(x), & & |x|\leq k,\label{eq:f-tilde-f-equal-on-small-Hamm-weight-first-principles}\\ & f(x)=\tilde{f}(x), & & |x|\geq n-k,\label{eq:f-tilde-f-equal-on-large-Hamm-weight-first-principles}\\ & |f(x)-\tilde{f}(x)|\leq\epsilon, & & x\in\zoon,\label{eq:f-tilde-f-close-on-med-Hamm-weight-first-principles}\\ & \deg\tilde{f}=O\left(\sqrt{nk}+\sqrt{n\log\frac{1}{\epsilon}}\right).\label{eq:tilde-f-degree-symmetric-first-principles}\end{aligned}$$ If $k\geq n/2,$ the theorem follows from the trivial bound $\deg_{0}(f)\leq n.$ For the complementary case $k<n/2,$ write $$\begin{aligned} f(x) & =\lambda+\sum_{i=0}^{k}\lambda'_{i}\cdot\EXACT_{n,i}(x)+\sum_{i=0}^{k}\lambda''_{i}\cdot\EXACT_{n,n-i}(x)\\ & =\lambda+\sum_{i=0}^{k}\lambda'_{i}\cdot\EXACT_{n,n-i}(\overline{x_{1}},\ldots,\overline{x_{n}})+\sum_{i=0}^{k}\lambda''_{i}\cdot\EXACT_{n,n-i}(x),\end{aligned}$$ where $\lambda,\lambda_{0}',\lambda_{0}'',\ldots,\lambda_{k}',\lambda_{k}''\in[-2,2]$ are fixed reals. By Theorem \[thm:EXACT\], each of the functions $\EXACT_{n,n-i}$ in this linear combination can be approximated pointwise to within $\epsilon/(2k+2)$ by a polynomial of degree $O(\sqrt{nk}+\sqrt{n\log(1/\epsilon)}).$ Moreover, the lemma guarantees that in each case, the approximation is exact on $\zoon_{\leq k}$ and $\zoon_{\geq n-k}.$ Now (\[eq:degeps-f-symmetric-first-principles\])(\[eq:f-tilde-f-close-on-med-Hamm-weight-first-principles\]) are immediate. Approximation using a sampling argument --------------------------------------- We now give a third proof of Theorem \[thm:SYMMETRIC\], inspired by combinatorics rather than approximation theory. Here, we show how to approximate an arbitrary symmetric function $f$ using an approximant for AND (cf. Theorem \[thm:AND\]) and a sampling argument. Suppose for the sake of concreteness that $f$ is supported on inputs of Hamming weight at most $k.$ Given a string $x\in\zoon,$ consider the experiment whereby one chooses $\lfloor n/k\rfloor$ bits of $x$ independently and uniformly at random, and outputs the disjunction of those bits. To approximate $f,$ we feed the expected value of the sampling experiment to a suitable univariate polynomial constructed by Lagrange interpolation. The expected value of the experiment as a function of $x$ has $\operatorname{\Pi}$-norm at most $2,$ which by Proposition \[prop:orcomplexity\] means that the overall composition has small $\Pi$-norm as well. The complete details of this construction are provided in Lemma \[lem:orcomplexity-symmetric\]. To finish the proof, we expand the composition as a linear combination of conjunctions and replace each conjunction by a corresponding approximant from Theorem \[thm:AND\]. \[lem:orcomplexity-symmetric\]Let $k\geq0$ be a given integer. Let $f\colon\zoon\to[-1,1]$ be a symmetric function that vanishes on $\zoon_{>k}.$ Then for every $0<\epsilon<1/2,$ there exists an $($explicitly given$)$ function $\tilde{f}\colon\zoon\to\Re$ such that $$\begin{aligned} & f(x)=\tilde{f}(x), & & |x|\leq k,\label{eq:f-tilde-f-equal-on-small-Hamm-weight}\\ & f(x)=\tilde{f}(x), & & |x|\geq n-k,\label{eq:f-tilde-f-equal-on-large-Hamm-weight}\\ & |f(x)-\tilde{f}(x)|\leq\epsilon, & & x\in\zoon,\label{eq:f-tilde-f-close-on-med-Hamm-weight}\\ & \operatorname{\Pi}(\tilde{f})\leq C^{k+\log(1/\epsilon)},\label{eq:tilde-f-orcomplexity}\end{aligned}$$ where $C>1$ is an absolute constant independent of $f,n,k,\epsilon.$ If $k=0,$ the only possibilities are $f(x)\equiv0$ and $f(x)=\bigwedge\overline{x_{i}},$ and therefore we may take $\tilde{f}=f$. If $k\geq n/4,$ we again may take $\tilde{f}=f$ since $\operatorname{\Pi}(f)\leq2^{n}$ by Proposition \[prop:orcomplexity\]\[item:orcomplexity-trivial-bound\]. In what follows, we treat the remaining case $$1\leq k<\frac{n}{4}.\label{eq:k-not-too-large}$$ Consider the points $0=t_{0}\leq t_{1}\leq t_{2}\leq\cdots\leq t_{n}=1,$ where $$\begin{aligned} t_{i} & =1-\left(1-\frac{i}{n}\right)^{\left\lfloor \frac{n}{2k}\right\rfloor },\qquad\qquad i=0,1,2,\ldots,n.\end{aligned}$$ The derivative of $t\mapsto1-(1-\frac{t}{n})^{\lfloor n/(2k)\rfloor}$ on $[0,2k]$ ranges in $[\frac{1}{6k},\frac{1}{2k}].$ Therefore, the mean value theorem gives $$\begin{aligned} \frac{|i-j|}{6k}\leq|t_{i}-t_{j}| & \leq\frac{|i-j|}{2k}, & & i,j=0,1,2,\ldots,2k.\label{eq:ti-tj}\end{aligned}$$ In particular, $$\begin{aligned} \frac{i}{6k} & \leq t_{i}\leq\frac{i}{2k}, & & i=0,1,2,\ldots,2k.\label{eq:ti}\end{aligned}$$ Consider the univariate polynomials $$\begin{aligned} p(t) & =(1-t)^{d}\prod_{i=n-k}^{n}(t-t_{i}),\\ q(t) & =\sum_{i=0}^{k}\frac{f(1^{i}0^{n-i})}{p(t_{i})}\prod_{\substack{j=0\\ j\ne i } }^{2k}\frac{t-t_{j}}{t_{i}-t_{j}},\end{aligned}$$ where $$d=5\left\lceil 8k+\ln\frac{1}{\epsilon}\right\rceil .\label{eq:d-defined}$$ Our definitions ensure that $p(t_{i})q(t_{i})=f(1^{i}0^{n-i})$ for $i=0,1,2,\ldots,k.$ Moreover, we have $p(t_{i})q(t_{i})=0$ for $i=\{k+1,k+2,\ldots,2k\}\cup\{n-k,n-k+1,\ldots,n\}.$ Since $f$ vanishes on inputs of Hamming weight greater than $k,$ we conclude that $$\begin{gathered} p(t_{i})q(t_{i})=f(1^{i}0^{n-i}),\\ i=\{0,1,\ldots,2k\}\cup\{n-k,n-k+1,\ldots,n\}.\qquad\label{eq:pq-equals-f-on-small-and-large-inputs}\end{gathered}$$ A routine calculation reveals the following additional properties of $p$ and $q$. \[claim:pq-close-to-f\]$|p(t_{i})q(t_{i})-f(1^{i}0^{n-i})|\leq\epsilon$ for $i\geq2k.$ \[claim:pq-norm\]${|\!|\!| p\cdot q | \! | \! |}=2^{O(k+\log(1/\epsilon))}.$ We will settle these claims once we complete the main proof. Define $\tilde{f}\colon\zoon\to\Re$ by $\tilde{f}(x)=p(t_{|x|})q(t_{|x|}).$ Then (\[eq:f-tilde-f-equal-on-small-Hamm-weight\])(\[eq:f-tilde-f-close-on-med-Hamm-weight\]) follow directly from (\[eq:pq-equals-f-on-small-and-large-inputs\]) and Claim \[claim:pq-close-to-f\]. For (\[eq:tilde-f-orcomplexity\]), observe that $$\tilde{f}(x)=p\left(\Exp_{S}\;\,\bigvee_{i\in S}x_{i}\right)q\left(\Exp_{S}\;\,\bigvee_{i\in S}x_{i}\right)$$ where the expectation is over a multiset $S$ of $\lfloor\frac{n}{2k}\rfloor$ elements that are chosen independently and uniformly at random from $\{1,2,\ldots,n\}.$ As a result, (\[eq:tilde-f-orcomplexity\]) follows from Claim \[claim:pq-norm\] and Proposition \[prop:orcomplexity\] \[item:orcomplexity-homogeneous\], \[item:orcomplexity-triangle\], \[item:orcomplexity-disjunction\], \[item:orcomplexity-composition-with-univariate\]. Fix an arbitrary point $t\in[$$t_{2k},t_{n}]=[t_{2k},1]$. Recall from (\[eq:k-not-too-large\]) that $k<n/4.$ As a result, $$\begin{aligned} |p(t)q(t)| & \leq|p(t_{2k})q(t)|\\ & \leq|p(t_{2k})|\sum_{i=0}^{k}\frac{1}{\min\{|p(t_{0})|,\ldots,|p(t_{k})|\}}\prod_{\substack{j=0\\ j\ne i } }^{2k}\frac{|1-t_{j}|}{|t_{i}-t_{j}|}\\ & \leq\frac{|p(t_{2k})|}{|p(t_{k})|}\sum_{i=0}^{k}\prod_{\substack{j=0\\ j\ne i } }^{2k}\frac{|1-t_{j}|}{|t_{i}-t_{j}|}\\ & \leq\left(\frac{1-t_{2k}}{1-t_{k}}\right)^{d}\;\sum_{i=0}^{k}\prod_{\substack{j=0\\ j\ne i } }^{2k}\frac{|1-t_{j}|}{|t_{i}-t_{j}|}\\ & =\left(1-\frac{t_{2k}-t_{k}}{1-t_{k}}\right)^{d}\;\sum_{i=0}^{k}\prod_{\substack{j=0\\ j\ne i } }^{2k}\frac{|1-t_{j}|}{|t_{i}-t_{j}|}\\ & \leq\exp\left(-\frac{t_{2k}-t_{k}}{1-t_{k}}\cdot d\right)\sum_{i=0}^{k}\prod_{\substack{j=0\\ j\ne i } }^{2k}\frac{|1-t_{j}|}{|t_{i}-t_{j}|}.\end{aligned}$$ Using the lower bounds in (\[eq:ti-tj\]) and (\[eq:ti\]), we obtain $$\begin{aligned} |p(t)q(t)| & \leq\exp\left(-\frac{(2k-k)/6k}{1-(k/6k)}\cdot d\right)\sum_{i=0}^{k}\prod_{\substack{j=0\\ j\ne i } }^{2k}\frac{1-(j/6k)}{|i-j|/6k}\\ & =\exp\left(-\frac{d}{5}\right)\sum_{i=0}^{k}\prod_{\substack{j=0\\ j\ne i } }^{2k}\frac{6k-j}{|i-j|}\\ & \leq\exp\left(-\frac{d}{5}\right)\sum_{i=0}^{k}\frac{(6k)!/(4k)!}{i!\,(2k-i)!}\\ & =\exp\left(-\frac{d}{5}\right)\sum_{i=0}^{k}\binom{6k}{4k}\binom{2k}{i}\\ & \leq\exp\left(-\frac{d}{5}\right)\binom{6k}{4k}\cdot2^{2k}\\ & \leq\exp\left(-\frac{d}{5}\right)\cdot2^{8k}\\ & \leq\epsilon,\end{aligned}$$ where the last step follows from the definition of $d$ in (\[eq:d-defined\]). Hence, $|p(t_{i})q(t_{i})-f(1^{i}0^{n-i})|=|p(t_{i})q(t_{i})|\leq\epsilon$ for $i\geq2k.$ Recall from (\[eq:k-not-too-large\]) that $k<n/4.$ As a result, $$\begin{aligned} \min_{i=0,1,\ldots,k}|p(t_{i})| & =|p(t_{k})|\nonumber \\ & =|1-t_{k}|^{d}\prod_{i=n-k}^{n}|t_{k}-t_{i}|\nonumber \\ & \geq|1-t_{k}|^{d}\cdot|t_{k}-t_{2k}|^{k+1}\nonumber \\ & \geq\frac{1}{2^{d}\cdot6^{k+1}},\label{eq:min-pt-i}\end{aligned}$$ where the last step uses the estimates in (\[eq:ti-tj\]) and (\[eq:ti\]). As a result, $$\begin{aligned} {|\!|\!| p\cdot q | \! | \! |} & \leq(1+1)^{d}\prod_{i=n-k}^{n}(1+t_{i})\cdot\sum_{i=0}^{k}\frac{|f(1^{i}0^{n-i})|}{|p(t_{i})|}\prod_{\substack{j=0\\ j\ne i } }^{2k}\frac{1+t_{j}}{|t_{i}-t_{j}|}\\ & \leq2^{d}\cdot2^{k+1}\sum_{i=0}^{k}\frac{1}{2^{-d}\cdot6^{-k-1}}\prod_{\substack{j=0\\ j\ne i } }^{2k}\frac{2}{|t_{i}-t_{j}|}\\ & \leq4^{d}\cdot12^{k+1}\sum_{i=0}^{k}\prod_{\substack{j=0\\ j\ne i } }^{2k}\frac{2\cdot6k}{|i-j|}\allowdisplaybreaks\\ & =4^{d}\cdot12^{k+1}\cdot6^{2k}\cdot\frac{(2k)^{2k}}{(2k)!}\sum_{i=0}^{k}\binom{2k}{i}\\ & \leq4^{d}\cdot12^{k+1}\cdot6^{2k}\cdot\frac{(2k)^{2k}}{(2k)!}\cdot2^{2k}\\ & =2^{O(d+k)},\end{aligned}$$ where the first step is valid by Fact \[fact:poly-norm\]; the second step uses $0\leq t_{i}\leq1$ and (\[eq:min-pt-i\]); the third step follows from the lower bound in (\[eq:ti-tj\]); and the last step is legitimate by Stirling’s approximation. In view of (\[eq:d-defined\]), the proof is complete. We have reached the promised construction of an approximating polynomial for any symmetric function. Let $f\colon\zoon\to[-1,1]$ be an arbitrary symmetric function. Let $k$ be a nonnegative integer such that $f$ is constant on inputs of Hamming weight in $(k,n-k).$ Then for $0<\epsilon<1/2,$ $$\deg_{\epsilon}(f)=O\left(\sqrt{nk}+\sqrt{n\log\frac{1}{\epsilon}}\right).\label{eq:degeps-f-symmetric}$$ Moreover, the approximating polynomial is given explicitly in each case. If $k\geq n/2,$ the theorem follows from the trivial bound $\deg_{0}(f)\leq n.$ For the complementary case $k<n/2$, write $$f(x_{1},\ldots,x_{n})=\lambda+f'(x_{1},\ldots,x_{n})+f''(\overline{x_{1}},\ldots,\overline{x_{n}}),$$ where $\lambda\in[-1,1]$ and $f',f''\colon\zoon\to[-2,2]$ are symmetric functions that vanish on $\zoon_{>k}.$ Lemma \[lem:orcomplexity-symmetric\] shows that $f'/2$ and $f''/2$ are each approximated pointwise to within $\epsilon/5$ by a linear combination of conjunctions, with real coefficients whose absolute values sum to $2^{O(k+\log(1/\epsilon))}.$ By Theorem \[thm:AND\], each such conjunction can in turn be approximated pointwise by a polynomial of degree $d$ to within $2^{-\Theta(d^{2}/n)}$. Summarizing, $$\begin{aligned} E(f,d) & \leq E(f',d)+E(f'',d)\\ & \leq2E\left(\frac{f'}{2},d\right)+2E\left(\frac{f''}{2},d\right)\\ & \leq2\left(2\cdot\frac{\epsilon}{5}+2^{O(k+\log(1/\epsilon))}\cdot2^{-\Theta(d^{2}/n)}\right),\end{aligned}$$ whence (\[eq:degeps-f-symmetric\]). Moreover, the approximating polynomial is given explicitly because Theorem \[thm:AND\] and Lemma \[lem:orcomplexity-symmetric\] provide closed-form expressions for the approximants involved. Generalizations --------------- Theorem \[thm:AND\] on the approximation of AND and OR obviously generalizes to arbitrary conjunctions and disjunctions. Somewhat less obviously, it generalizes in an optimal manner to conjunctions and disjunctions whose domain of definition is restricted to the first few levels of the hypercube. We record this generalization for later use. \[thm:DISJUNCTION\]Let $f\colon\zoo_{\leq n}^{N}\to\zoo$ be given by $$f(x)=\left(\bigvee_{i\in A}x_{i}\right)\vee\left(\bigvee_{i\in B}\overline{x_{i}}\right),$$ for some subsets $A,B\subseteq\{1,2,\ldots,N\}.$ Then $$\begin{aligned} E(f,d) & \leq\frac{1}{2}\exp\left(-\frac{cd^{2}}{n}\right), & & d=0,1,2,\ldots,\end{aligned}$$ where $c>0$ is an absolute constant. Moreover, the approximating polynomial is given explicitly in each case. If $|B|>n,$ then $f\equiv1$ on its domain of definition and hence $E(f,0)=0$. In the complementary case when $|B|\leq n,$ we have $$\begin{aligned} \sum_{i\in A}x_{i}+\sum_{i\in B}(1-x_{i}) & \in\{0,1,2,\ldots,2n\}, & & x\in\zoo_{\leq n}^{N}.\label{eq:sum-contained}\end{aligned}$$ Theorem \[thm:AND\] gives an explicit univariate polynomial $p$ of degree $d$ such that $$\begin{aligned} & p(2n)=1,\label{eq:p-at-0-restated}\\ & |p(t)|\leq\exp\left(-\frac{Cd^{2}}{n}\right), & & t=0,1,2,\ldots,2n-1,\label{eq:p-at-i-restated}\end{aligned}$$ where $C>0$ is an absolute constant. Define $$P(x)=1-\frac{1}{1+\exp(-Cd^{2}/n)}\cdot p\left(2n-\sum_{i\in A}x_{i}-\sum_{i\in B}(1-x_{i})\right).$$ Then $$\begin{aligned} \max_{x\in\zoo_{\leq n}^{N}}|f(x)-P(x)| & \leq\frac{\exp(-Cd^{2}/n)}{1+\exp(-Cd^{2}/n)}\\ & \leq\frac{1}{2}\exp\left(-\frac{Cd^{2}}{2n}\right),\end{aligned}$$ where the first step follows from (\[eq:sum-contained\])(\[eq:p-at-i-restated\]), and the second step uses $a/(1+a)\leq\sqrt{a}/2$ for any $a\geq0.$ \[cor:CONJUNCTION\]Let $f\colon\zoo_{\leq n}^{N}\to\zoo$ be given by $$f(x)=\left(\bigwedge_{i\in A}x_{i}\right)\wedge\left(\bigwedge_{i\in B}\overline{x_{i}}\right),$$ for some subsets $A,B\subseteq\{1,2,\ldots,N\}.$ Then $$\begin{aligned} E(f,d) & \leq\frac{1}{2}\exp\left(-\frac{cd^{2}}{n}\right), & & d=0,1,2,\ldots,\end{aligned}$$ where $c>0$ is an absolute constant. Moreover, the approximating polynomial is given explicitly in each case. Apply Theorem \[thm:DISJUNCTION\] to $1-f.$ *\[sec:k-DNF-and-k-CNF\]k*-DNF and *k*-CNF formulas =================================================== Recall that a *$k$-DNF formula* in Boolean variables $x_{1},x_{2},\ldots,x_{N}$ is the disjunction of zero or more *terms*, where each term is the conjunction of at most $k$ literals from among $x_{1},\overline{x_{1}},x_{2},\overline{x_{2}},\ldots,x_{N},\overline{x_{N}}.$ As a convention, we consider the constant functions $0$ and $1$ to be valid $k$-DNF formulas for every $k\geq0$. Analogously, a *$k$-CNF formula* in Boolean variables $x_{1},x_{2},\ldots,x_{N}$ is the conjunction of zero or more *clauses*, where each clause is the disjunction of at most $k$ literals from among $x_{1},\overline{x_{1}},x_{2},\overline{x_{2}},\ldots,x_{N},\overline{x_{N}}.$ Again, we consider the constant functions $0$ and $1$ to be valid $k$-CNF formulas for all $k\ge0.$ Recall that a function $f$ is representable by a $k$-DNF formula if and only if its negation $\overline{f}$ is representable by a $k$-CNF formula. Note also that the definition of $k$-DNF formulas is hereditary in the sense that a $k$-DNF formula is also a $k'$-DNF formula for any $k'\geq k,$ and analogously for CNF formulas. The contribution of this section is to settle Theorem \[thm:MAIN-k-dnf\] on the approximate degree of every $k$-DNF and $k$-CNF formula. We will in fact prove the following more precise result, for every setting of the error parameter. \[thm:k-DNF\]Let $f\colon\zoo_{\leq n}^{N}\to\zoo$ be representable on its domain by a $k$-DNF or $k$-CNF formula. Then $$\deg_{\epsilon}(f)\leq c\cdot(\sqrt{2})^{k}\,n^{\frac{k}{k+1}}\left(\log\frac{1}{\epsilon}\right)^{\frac{1}{k+1}}\label{eq:k-dnf-main}$$ for all $0<\epsilon<1/2,$ where $c>1$ is an absolute constant independent of $f,N,n,k,\epsilon.$ Moreover, the approximating polynomial is given explicitly in each case. We present the proof of this theorem in Sections \[subsec:k-dnf-key-quantities\]\[subsec:k-dnf-solving-the-recurrence\] below. Key quantities\[subsec:k-dnf-key-quantities\] --------------------------------------------- For nonnegative integers $n$ and $k$ and a real number $\Delta\geq1,$ we define $$D(n,k,\Delta)=\max_{f}\deg_{2^{-\Delta}}(f),$$ where the maximum is over all functions $f\colon\zoo_{\leq n}^{N}\to\zoo$ for some $N\geq n$ that are representable by a $k$-DNF formula. Fact \[fact:trivial-upper-bound-on-deg\] gives the upper bound $$D(n,k,\Delta)\leq n.\label{eq:D-nkr-trivial}$$ Since the only $0$-DNF formulas are the constant functions $0$ and $1,$ we obtain $$D(n,0,\Delta)=0.\label{eq:D-nkr-base}$$ We will prove Theorem \[thm:k-DNF\] by induction of $k$, with (\[eq:D-nkr-base\]) serving as the base case. A composition theorem for approximate degree -------------------------------------------- The inductive step in our analysis of $D(n,k,\Delta)$ relies on a certain general bound on approximate degree for a class of composed functions, as follows. \[lem:selector-homogeneous\]Let $F\colon X\times\zoo_{n}^{N}\to\zoo$ be given by $$F(x,y)=\bigvee_{i=1}^{N}y_{i}\wedge f_{i}(x)$$ for some functions $f_{1},f_{2},\ldots,f_{N}\colon X\to\zoo.$ Let $b$ be an integer with $b\mid n$ and $b\mid N.$ Then $$\deg_{\epsilon}(F)\leq C\sqrt{nb\log\frac{1}{\epsilon}}+\max_{\substack{S\subseteq\{1,\ldots,N\}\\ |S|\leq C\sqrt{nb\log\frac{1}{\epsilon}} } }\deg_{\epsilon\exp\left(-C\sqrt{\frac{n}{b}\log\frac{1}{\epsilon}}\right)}\left(\bigvee_{i\in S}f_{i}\right)$$ for all $0<\epsilon\leq1/2,$ where $C>1$ is an absolute constant independent of $F,N,n,b,\epsilon.$ As we will see shortly, the bound of Lemma \[lem:selector-homogeneous\] generalizes to functions $F\colon X\times\zoo_{\leq n}^{N}\to\zoo$ and to arbitrary reals $b\geq1$. It is this more general, and more natural, result on the approximate degree of composed functions that we need for our analysis of $D(n,k,\Delta).$ However, establishing Lemma \[lem:selector-homogeneous\] first considerably improves the readability and modularity of the proof. By way of notation, we remind the reader that the symbol $\bigvee_{i\in S}f_{i}$ denotes the mapping $x\mapsto\bigvee_{i\in S}f_{i}(x).$ The reader will also recall the shorthand $[n]=\{1,2,\ldots,n\}.$ In particular, $\binom{[n]}{\leq d}$ denotes the family of subsets of $\{1,2,\ldots,n\}$ of cardinality at most $d.$ The proof is constructive and uses as its building blocks two main components: an “outer” approximant (for the OR function) and “inner” approximants (for disjunctions of small sets of $f_{i}$). We first describe these components individually and then present the overall construction and error analysis. <span style="font-variant:small-caps;">Step 1: Outer approximant.</span> Theorem \[thm:AND\] provides a symmetric multilinear polynomial $\widetilde{\OR}_{n/b}\colon\zoo^{n/b}\to[0,1]$ of degree $d=O(\sqrt{n\log(1/\epsilon)/b})$ that approximates $\OR_{n/b}$ pointwise to within $\epsilon/2.$ More specifically, there are real coefficients $a_{0},a_{1},a_{2},\ldots$ such that $$\begin{aligned} \left|\bigvee_{i=1}^{n/b}z_{i}-\sum_{S\in\binom{[n/b]}{\leq d}}a_{|S|}\prod_{i\in S}z_{i}\right| & \leq\frac{\epsilon}{2}, & & z\in\zoo^{n/b},\label{eq:decomp-tilde-OR-error}\end{aligned}$$ where $$\begin{aligned} & 1\leq d\leq c\sqrt{\frac{n}{b}\log\frac{1}{\epsilon}}\label{eq:decomp-tilde-OR-degree}\end{aligned}$$ for some absolute constant $c>1$. By Lemma \[lem:bound-on-poly-coeffs-MULTIVARIATE\], $$\begin{aligned} \sum_{\ell=0}^{d}\binom{n/b}{\ell}|a_{\ell}| & \leq8^{d}.\label{eq:tilde-OR-norm}\end{aligned}$$ <span style="font-variant:small-caps;">Step 2: Inner approximants.</span> ** For a subset $S\subseteq\{1,2,\ldots,N\},$ define $f_{S}\colon X\to\zoo$ by $$f_{S}(x)=\bigvee_{i\in S}f_{i}(x).$$ Fix a polynomial $\tilde{f}_{S}\colon X\to\Re$ of the smallest possible degree such that $$\begin{aligned} & \|f_{S}-\tilde{f}_{S}\|_{\infty}\leq\frac{\epsilon}{2}\left(\sum_{\ell=0}^{d}\binom{n/b}{\ell}2^{\ell}|a_{\ell}|\right)^{-1}.\label{eq:decomp-approx-fS-error}\end{aligned}$$ To avoid notational clutter in the formulas below, we will frequently write $f_{S}$ and $\tilde{f}_{S}$ instead of $f_{S}(x)$ and $\tilde{f}_{S}(x)$, respectively, when referring to the value of these functions at a given point $x\in X.$ We have $$\begin{aligned} \deg(\tilde{f}_{S}) & \leq\deg_{\epsilon/(2\cdot16^{d})}\left(\bigvee_{i\in S}f_{i}\right)\nonumber \\ & \leq\deg_{\epsilon\exp\left(-4c\sqrt{\frac{n}{b}\log\frac{1}{\epsilon}}\right)}\left(\bigvee_{i\in S}f_{i}\right),\label{eq:tilde-fs-degree}\end{aligned}$$ where the first and second steps use (\[eq:tilde-OR-norm\]) and (\[eq:decomp-tilde-OR-degree\]), respectively. <span style="font-variant:small-caps;">Step 3: Overall approximant.</span> By appropriately composing the outer approximant with the inner approximants, we obtain an approximant for the overall function $F$. Specifically, define $\tilde{F}\colon X\times\zoo_{n}^{N}\to\Re$ by $$\begin{gathered} \tilde{F}(x,y)=a_{0}+\sum_{\ell=1}^{d}a_{\ell}\binom{n/b}{\ell}\binom{N}{b\ell}\binom{n}{b\ell}^{-1}\\ \times\Exp_{B_{1},\ldots,B_{n/b}}\left[\left(\sum_{\substack{S\subseteq\{1,2,\ldots,\ell\}\\ S\ne\varnothing } }(-1)^{|S|+1}\;\widetilde{f}_{\bigcup_{i\in S}\!B_{i}}\right)\prod_{i\in B_{1}\cup\cdots\cup B_{\ell}}y_{i}\right],\label{eq:decomp-approximant-defined}\end{gathered}$$ where expectation is taken over a uniformly random tuple of sets $B_{1},\ldots,B_{n/b}\subseteq\{1,2,\ldots,N\}$ that are pairwise disjoint and have cardinality $b$ each. Then $$\begin{aligned} \deg(\tilde{F}) & \leq\max_{B_{1},\ldots,B_{n/b}}\;\max_{S\subseteq\{1,2,\ldots,d\}}\left\{ \deg(\tilde{f}_{\bigcup_{i\in S}B_{i}})+\sum_{i=1}^{d}|B_{i}|\right\} \nonumber \\ & \leq\max_{S\in\binom{[N]}{\leq db}}\left\{ \deg(\tilde{f}_{S})\right\} +db\nonumber \\ & \leq\max_{\substack{S\subseteq\{1,\ldots,N\}\\ |S|\leq c\sqrt{nb\log\frac{1}{\epsilon}} } }\left\{ \deg_{\epsilon\exp\left(-4c\sqrt{\frac{n}{b}\log\frac{1}{\epsilon}}\right)}\left(\bigvee_{i\in S}f_{i}\right)\right\} +c\sqrt{nb\log\frac{1}{\epsilon}},\label{eq:tilde-F-degree}\end{aligned}$$ where the final step uses (\[eq:decomp-tilde-OR-degree\]) and (\[eq:tilde-fs-degree\]). <span style="font-variant:small-caps;">Step 4: Error analysis.</span> For the rest of the proof, fix $y\in\zoo_{n}^{N}$ arbitrarily. Let $L=\{i:y_{i}=1\}$. In the defining equation (\[eq:decomp-approximant-defined\]), the product $\prod_{i\in B_{1}\cup\cdots\cup B_{\ell}}y_{i}$ acts like an indicator random variable for the event that $B_{1}\cup\ldots\cup B_{\ell}\subseteq L,$ which occurs with probability precisely $$\binom{|L|}{b\ell}\binom{N}{b\ell}^{-1}=\binom{n}{b\ell}\binom{N}{b\ell}^{-1}.$$ Therefore, $$\begin{aligned} & \hspace{-5mm}\tilde{F}(x,y)\nonumber \\ & =a_{0}+\sum_{\ell=1}^{d}a_{\ell}\binom{n/b}{\ell}\Exp_{B_{1},\ldots,B_{n/b}}\left[\sum_{\substack{S\subseteq[\ell]\\ S\ne\varnothing } }(-1)^{|S|+1}\;\widetilde{f}_{\bigcup_{i\in S}\!B_{i}}\;\middle|\;B_{1},\ldots,B_{\ell}\subseteq L\right]\nonumber \\ & =a_{0}+\sum_{\ell=1}^{d}a_{\ell}\binom{n/b}{\ell}\Exp_{B_{1},\ldots,B_{n/b}}\left[\sum_{\substack{S\subseteq[\ell]\\ S\ne\varnothing } }(-1)^{|S|+1}\;\widetilde{f}_{\bigcup_{i\in S}\!B_{i}}\;\middle|\;\bigcup_{i=1}^{n/b}B_{i}=L\right]\allowdisplaybreaks\nonumber \\ & =a_{0}+\sum_{\ell=1}^{d}a_{\ell}\Exp_{B_{1},\ldots,B_{n/b}}\left[\sum_{T\in\binom{[n/b]}{\ell}}\sum_{\substack{S\subseteq T\\ S\ne\varnothing } }(-1)^{|S|+1}\;\widetilde{f}_{\bigcup_{i\in S}\!B_{i}}\;\;\middle|\;\;\bigcup_{i=1}^{n/b}B_{i}=L\right]\nonumber \\ & =\Exp_{B_{1},\ldots,B_{n/b}}\left[a_{0}+\sum_{\ell=1}^{d}a_{\ell}\sum_{T\in\binom{[n/b]}{\ell}}\sum_{\substack{S\subseteq T\\ S\ne\varnothing } }(-1)^{|S|+1}\;\widetilde{f}_{\bigcup_{i\in S}\!B_{i}}\;\;\middle|\;\;\bigcup_{i=1}^{n/b}B_{i}=L\right],\label{eq:decomp-tilde-Fxy-simplified}\end{aligned}$$ where the second step is valid because a uniformly random tuple of pairwise disjoint sets $B_{1},\ldots,B_{\ell}\subseteq L$ of cardinality $b$ each can be generated by partitioning $L$ uniformly at random into parts of size $b$ and using the first $\ell$ parts of that partition; the third step is valid in view of the symmetry of the distribution of $B_{1},\ldots,B_{n/b}$; and the last step uses the linearity of expectation. Analogously, $$\begin{aligned} F(x,y) & =\bigvee_{i=1}^{N}y_{i}\wedge f_{i}\nonumber \\ & =\bigvee_{i\in L}f_{i}\nonumber \\ & =f_{L}\nonumber \\ & =\Exp_{B_{1},\ldots,B_{n/b}}\left[f_{B_{1}\cup\cdots\cup B_{n/b}}\;\;\middle|\;\;\bigcup_{i=1}^{n/b}B_{i}=L\right].\label{eq:decomp-Fxy-simplified}\end{aligned}$$ As a result, $$\begin{aligned} & |F(x,y)-\tilde{F}(x,y)|\nonumber \\ & \qquad\leq\max_{B_{1},\ldots,B_{n/b}}\left|f_{B_{1}\cup\cdots\cup B_{n/b}}-a_{0}-\sum_{\ell=1}^{d}a_{\ell}\sum_{T\in\binom{[n/b]}{\ell}}\sum_{\substack{S\subseteq T\\ S\ne\varnothing } }(-1)^{|S|+1}\tilde{f}_{\bigcup_{i\in S}\!B_{i}}\right|\nonumber \\ & \qquad\leq\max_{B_{1},\ldots,B_{n/b}}\left|f_{B_{1}\cup\cdots\cup B_{n/b}}-a_{0}-\sum_{\ell=1}^{d}a_{\ell}\sum_{T\in\binom{[n/b]}{\ell}}\sum_{\substack{S\subseteq T\\ S\ne\varnothing } }(-1)^{|S|+1}f_{\bigcup_{i\in S}\!B_{i}}\right|\nonumber \\ & \qquad\qquad+\max_{B_{1},\ldots,B_{n/b}}\;\;\sum_{\ell=1}^{d}|a_{\ell}|\sum_{T\in\binom{[n/b]}{\ell}}\sum_{\substack{S\subseteq T\\ S\ne\varnothing } }\left|f_{\bigcup_{i\in S}B_{i}}-\tilde{f}_{\bigcup_{i\in S}\!B_{i}}\right|\allowdisplaybreaks\nonumber \\ & \qquad\leq\max_{B_{1},\ldots,B_{n/b}}\left|f_{B_{1}\cup\cdots\cup B_{n/b}}-a_{0}-\sum_{\ell=1}^{d}a_{\ell}\sum_{T\in\binom{[n/b]}{\ell}}\sum_{\substack{S\subseteq T\\ S\ne\varnothing } }(-1)^{|S|+1}f_{\bigcup_{i\in S}\!B_{i}}\right|\nonumber \\ & \qquad\qquad+\frac{\epsilon}{2}\nonumber \\ & \qquad=\max_{B_{1},\ldots,B_{n/b}}\left|\bigvee_{i=1}^{n/b}f_{B_{i}}-a_{0}-\sum_{\ell=1}^{d}a_{\ell}\sum_{T\in\binom{[n/b]}{\ell}}\sum_{\substack{S\subseteq T\\ S\ne\varnothing } }(-1)^{|S|+1}\bigvee_{i\in S}f_{B_{i}}\right|+\frac{\epsilon}{2}\nonumber \\ & \qquad=\max_{B_{1},\ldots,B_{n/b}}\left|\bigvee_{i=1}^{n/b}f_{B_{i}}-\sum_{\ell=0}^{d}a_{\ell}\sum_{T\in\binom{[n/b]}{\ell}}\prod_{i\in T}f_{B_{i}}\right|+\frac{\epsilon}{2}\nonumber \\ & \qquad\leq\epsilon,\label{eq:decomp-error}\end{aligned}$$ where the first step is immediate from (\[eq:decomp-tilde-Fxy-simplified\]) and (\[eq:decomp-Fxy-simplified\]), the second step applies the triangle inequality, the third step is valid by (\[eq:decomp-approx-fS-error\]), the fourth step is a change of notation, the fifth step uses the inclusion-exclusion formula (Fact \[fact:incl-excl-alternative\]), and the last step is justified by (\[eq:decomp-tilde-OR-error\]). By (\[eq:tilde-F-degree\]) and (\[eq:decomp-error\]), the proof of Lemma \[lem:selector-homogeneous\] is complete by taking $C=4c.$ To remove the homogeneity and divisibility assumptions in Lemma \[lem:selector-homogeneous\], we now show how to reduce the approximation of any function on $\zoo_{\leq n}^{N}$ to the approximation of a closely related function on $\zoo_{n}^{N+n}.$ This connection is surprising at first but has a short proof based on Minsky and Papert’s symmetrization argument. \[lem:homogenization\]Let $f\colon X\times\zoo_{\leq n}^{N}\to\Re$ be given. Define $f'\colon X\times\zoo_{n}^{N+n}\to\Re$ by $$f'(x,y_{1}\ldots y_{N+n})=f(x,y_{1}\ldots y_{N}).\label{eq:homogenization}$$ Then for all $\epsilon\geq0,$ $$\deg_{\epsilon}(f')=\deg_{\epsilon}(f).$$ The upper bound $\deg_{\epsilon}(f')\leq\deg_{\epsilon}(f)$ is immediate from the defining equation (\[eq:homogenization\]). For a matching lower bound, fix a polynomial $\phi'\colon X\times\Re^{N+n}\to\Re$ such that $$\begin{aligned} & |f'(x,y)-\phi'(x,y)|\leq\epsilon, & & x\in X,\;y\in\zoo_{n}^{N+n},\label{eq:homog-approx}\\ & \deg\phi'=\deg_{\epsilon}(f').\label{eq:homog-deg-tilde-f}\end{aligned}$$ Minsky and Papert’s symmetrization argument (Proposition \[prop:symmetrization\]) yields a polynomial $\phi^{*}\colon X\times\Re^{N}\times\Re\to\Re$ such that for $t=0,1,2,\ldots,n,$ $$\begin{aligned} & \phi^{*}(x,y,t)=\Exp_{z\in\zoo_{t}^{n}}\phi'(x,yz), & & x\in X,\;y\in\zoo^{N},\label{eq:homog-symm}\\ & \deg\phi^{*}\leq\deg\phi'.\label{eq:homog-deg}\end{aligned}$$ We are now in a position to construct the desired approximant for $f.$ For any $x\in X$ and $y\in\zoo_{\leq n}^{N},$ we have $$\begin{aligned} & \left|f(x,y)-\phi^{*}\!\left(x,y,n-\sum_{i=1}^{N}y_{i}\right)\right|\\ & \qquad\qquad\qquad\leq\left|f(x,y)-\Exp_{\substack{z\in\zoo_{n-|y|}^{n}} }f'(x,yz)\right|\\ & \qquad\qquad\qquad\qquad\qquad+\left|\Exp_{\substack{z\in\zoo_{n-|y|}^{n}} }f'(x,yz)-\phi^{*}\!\left(x,y,n-\sum_{i=1}^{N}y_{i}\right)\right|\\ & \qquad\qquad\qquad=\left|\Exp_{\substack{z\in\zoo_{n-|y|}^{n}} }f'(x,yz)-\phi^{*}\!\left(x,y,n-\sum_{i=1}^{N}y_{i}\right)\right|\\ & \qquad\qquad\qquad=\left|\Exp_{\substack{z\in\zoo_{n-|y|}^{n}} }[f'(x,yz)-\phi'(x,yz)]\right|\\ & \qquad\qquad\qquad\leq\epsilon,\end{aligned}$$ where the first step applies the triangle inequality, the second step is immediate from the definition of $f',$ the third step uses (\[eq:homog-symm\]), and the last step follows from (\[eq:homog-approx\]). In summary, we have shown that $\deg_{\epsilon}(f)\leq\deg\phi^{*},$ which in view of (\[eq:homog-deg\]) and (\[eq:homog-deg-tilde-f\]) completes the proof. We are now in a position to remove the divisibility assumption in Lemma \[lem:selector-homogeneous\] and additionally generalize it to the nonhomogeneous setting. \[thm:selector\]Let $F\colon X\times\zoo_{\leq n}^{N}\to\zoo$ be given by $$F(x,y)=\bigvee_{i=1}^{N}y_{i}\wedge f_{i}(x)$$ for some functions $f_{1},f_{2},\ldots,f_{N}\colon X\to\zoo.$ Then $$\begin{gathered} \deg_{\epsilon}(F)\leq C\sqrt{nb\log\frac{1}{\epsilon}}\\ +\max_{\substack{S\subseteq\{1,\ldots,N\}\\ |S|\leq C\sqrt{nb\log\frac{1}{\epsilon}} } }\deg_{\epsilon\exp\left(-C\sqrt{\frac{n}{b}\log\frac{1}{\epsilon}}\right)}\left(\bigvee_{i\in S}f_{i}\right)\qquad\label{eq:selector-general}\end{gathered}$$ for all reals $b\geq1$ and $0<\epsilon\leq1/2,$ where $C>1$ is an absolute constant independent of $F,N,n,b,\epsilon.$ We first examine the case $1\leq b\leq n.$ Consider the function $F'\colon X\times\zoo_{n'}^{N'}\to\zoo$ given by $$F'(x,y)=\bigvee_{i=1}^{N'}y_{i}\wedge f_{i}(x),$$ where $$\begin{aligned} & n'=\lfloor b\rfloor\left\lceil \frac{n}{\lfloor b\rfloor}\right\rceil ,\\ & N'=\lfloor b\rfloor\left\lceil \frac{N}{\lfloor b\rfloor}\right\rceil +\lfloor b\rfloor\left\lceil \frac{n}{\lfloor b\rfloor}\right\rceil ,\\ \rule{0mm}{4mm} & f_{N+1}=f_{N+2}=\cdots=f_{N'}=0.\end{aligned}$$ Then $$\begin{aligned} \deg_{\epsilon}(F) & \leq\deg_{\epsilon}(F')\\ & \leq c\sqrt{n'\lfloor b\rfloor\log\frac{1}{\epsilon}}+\max_{\substack{S\subseteq\{1,\ldots,N'\}\\ |S|\leq c\sqrt{n'\lfloor b\rfloor\log\frac{1}{\epsilon}} } }\deg_{\epsilon\exp\left(-c\sqrt{\frac{n'}{\lfloor b\rfloor}\log\frac{1}{\epsilon}}\right)}\left(\bigvee_{i\in S}f_{i}\right),\end{aligned}$$ for some absolute constant $c\geq1,$ where the first step uses the homogenization lemma (Lemma \[lem:homogenization\]) and the second step follows from Lemma \[lem:selector-homogeneous\]. This settles (\[eq:selector-general\]) for $C=2c.$ For the complementary case $b\geq n,$ define $F'\colon X\times\zoo_{n}^{N+n}\to\zoo$ by $$F'(x,y)=\bigvee_{i=1}^{N+n}y_{i}\wedge f_{i}(x),$$ where $f_{N+1}=f_{N+2}=\cdots=f_{N+n}=0.$ Then $$\deg_{\epsilon}(F)=\deg_{\epsilon}(F')\label{eq:case-b-large}$$ by the homogenization lemma (Lemma \[lem:homogenization\]). On the other hand, $$F'(x,y)=\sum_{S\in\binom{[N+n]}{n}}\left(\bigvee_{i\in S}f_{i}(x)\right)\prod_{i\in S}y_{i}.\label{eq:F'-exact}$$ For any input $y$ of Hamming weight $n,$ every term in this summation vanishes except for the term corresponding to $S=\{i:y_{i}=1\}.$ This means that an approximant for $F'$ with error $\epsilon$ can be obtained by replacing each disjunction in (\[eq:F’-exact\]) with a polynomial that approximates that disjunction to within $\epsilon.$ As a result, $$\deg_{\epsilon}(F')\leq n+\max_{S\in\binom{[N]}{\leq n}}\deg_{\epsilon}\left(\bigvee_{i\in S}f_{i}\right).$$ This upper bound along with (\[eq:case-b-large\]) settles (\[eq:selector-general\]) for $b\geq n.$ A recursive bound ----------------- Using Theorem \[thm:selector\] as our main tool, we now derive the promised recurrence for $D(n,k,\Delta).$ \[lem:k-dnf-recursion\]There is a constant $C\geq1$ such that for all integers $n,k\geq1$ and reals $\Delta\geq1,$ $$D(n,k,\Delta)\leq\max_{b\geq1}\left\{ C\sqrt{nb\Delta}+D\!\left(n,k-1,\Delta+C\sqrt{\frac{n\Delta}{b}}\right)\right\} .\label{eq:k-dnf-recursion}$$ Let $f\colon\zoo_{\leq n}^{N}\to\zoo$ be a $k$-DNF formula. Our objective is to bound $\deg_{2^{-\Delta}}(f)$ by the right-hand side of (\[eq:k-dnf-recursion\]). We may assume that $$f\not\equiv1,\label{eq:f-not-1}$$ since the bound holds trivially for the constant function $f=1$. Write $f=f'\vee f'',$ where $f'$ is a $k$-DNF formula in which every term has an unnegated variable, and $f''$ is a $k$-DNF formula whose terms feature only negated variables. Collecting like terms in $f',$ we immediately obtain $$f'(x)=\bigvee_{i=1}^{N}x_{i}\wedge f'_{i}(x)\label{eq:k-dnf-recursion-f'}$$ for some $(k-1)$-DNF formulas $f_{1}',f_{2}',\ldots,f_{N}'.$ We now turn to $f''$. By (\[eq:f-not-1\]), there exists $x^{*}\in\zoo_{\leq n}^{N}$ such that $f''(x^{*})=0.$ Consider the subset $I=\{i:x_{i}^{*}=1\}$, of cardinality $$|I|\leq n.\label{eq:k-dnf-recursion-I}$$ Since every occurrence of a variable in $f''(x)$ is negated, we conclude that every term in $f''(x)$ features some literal $\overline{x_{i}}$ with $i\in I.$ Collecting like terms, we obtain the representation $$f''(x)=\bigvee_{i\in I}\overline{x_{i}}\wedge f''_{i}(x),\label{eq:k-dnf-recursion-f''}$$ where each $f''_{i}$ is a $(k-1)$-DNF formula. To summarize (\[eq:k-dnf-recursion-f’\])(\[eq:k-dnf-recursion-f”\]), the function $f=f'\vee f''$ is a subfunction of some $F\colon\zoo_{\leq n}^{N}\times\zoo_{\leq2n}^{N+n}\to\zoo$ of the form $$F(x,y)=\bigvee_{i=1}^{N+n}y_{i}\wedge f_{i}(x),$$ where each $f_{i}$ is a $(k-1)$-DNF formula. Now $$\begin{aligned} \deg_{2^{-\Delta}}(f) & \leq\deg_{2^{-\Delta}}(F)\\ & \leq\max_{b\geq1}\left\{ c\sqrt{2nb\Delta}+\max_{\substack{S\subseteq\{1,2\ldots,N+n\}} }\deg_{2^{-\Delta}\exp(-c\sqrt{2n\Delta/b})}\left(\bigvee_{i\in S}f_{i}\right)\right\} \\ & \leq\max_{b\geq1}\left\{ c\sqrt{2nb\Delta}+D\!\left(n,k-1,\Delta+\frac{c}{\ln2}\sqrt{\frac{2n\Delta}{b}}\right)\right\} ,\end{aligned}$$ where the second step follows from Theorem \[thm:selector\] for a suitable absolute constant $c\geq1,$ and the third step is justified by the fact that each $\bigvee_{i\in S}f_{i}\colon\zoo_{\leq n}^{N}\to\zoo$ is a $(k-1)$-DNF formula. In conclusion, (\[eq:k-dnf-recursion\]) holds with $C=c\sqrt{2}/\ln2.$ \[subsec:k-dnf-solving-the-recurrence\]Solving the recurrence ------------------------------------------------------------- It remains to solve the recurrence for $D(n,k,\Delta)$ given by (\[eq:D-nkr-base\]) and Lemma \[lem:k-dnf-recursion\]. \[thm:D-nkr-final\]There is a constant $c\geq1$ such that for all integers $n,k\geq0$ and reals $\Delta\geq1,$ $$D(n,k,\Delta)\leq c\cdot(\sqrt{2})^{k}\,n^{\frac{k}{k+1}}\,\Delta^{\frac{1}{k+1}}.\label{eq:k-dnf-recurrence-solved}$$ This result settles Theorem \[thm:k-DNF\]. Indeed, if $f\colon\zoo_{\leq n}^{N}\to\zoo$ is representable by a $k$-DNF formula, then (\[eq:k-dnf-main\]) is immediate from (\[eq:k-dnf-recurrence-solved\]). The same bound applies to $k$-CNF formulas because they are negations of $k$-DNF formulas, and $\deg_{\epsilon}(f)=\deg_{\epsilon}(1-f)$ for any $f.$ We will prove (\[eq:k-dnf-recurrence-solved\]) for $c=2(C+1)^{2},$ where $C\geq1$ is the absolute constant from Lemma \[lem:k-dnf-recursion\]. The proof is by induction on $k.$ The base $k=0$ is valid due to (\[eq:D-nkr-base\]). For the inductive step, let $k\geq1$ be arbitrary. For $\Delta\geq n,$ the claim is immediate from (\[eq:D-nkr-trivial\]), and we focus on the complementary case $$1\leq\Delta\leq n.\label{eq:eps-bounded}$$ For every $b\geq1,$ $$\begin{aligned} D(n,k,\Delta) & \leq\min\left\{ n,C\sqrt{nb\Delta}+D\!\left(n,k-1,\Delta+C\sqrt{\frac{n\Delta}{b}}\right)\right\} \nonumber \\ & \leq\min\left\{ n,C\sqrt{nb\Delta}+2(C+1)^{2}\,2^{\frac{k-1}{2}}n^{\frac{k-1}{k}}\left(\Delta+C\sqrt{\frac{n\Delta}{b}}\right)^{\frac{1}{k}}\right\} \nonumber \\ & \leq(C+1)\sqrt{nb\Delta}+(C+1)^{2}\,2^{\frac{k+1}{2}}n^{\frac{k-1}{k}}\left((C+1)\sqrt{\frac{n\Delta}{b}}\right)^{\frac{1}{k}},\label{eq:D-nkr-solving-recurrence}\end{aligned}$$ where the first step uses (\[eq:D-nkr-trivial\]) and Lemma \[lem:k-dnf-recursion\]; the second step applies the inductive hypothesis; and the last step can be verified in a straightforward manner by examining the cases $\Delta\leq n/b$ and $\Delta\geq n/b.$ Setting $$b=(C+1)^{2}\,2^{k}\left(\frac{n}{\Delta}\right)^{1-\frac{2}{k+1}}$$ in (\[eq:D-nkr-solving-recurrence\]) now yields (\[eq:k-dnf-recurrence-solved\]), completing the inductive step. Note that our choice of parameter meets the requirement $b\geq1$, as one can see from (\[eq:eps-bounded\]). *\[sec:k-Distinctness\]k*-Element distinctness ============================================== For an integer $k$, recall that the *threshold function* ${\operatorname{THR}}_{k}\colon\zoo^{*}\to\zoo$ is given by $${\operatorname{THR}}_{k}(x)=\begin{cases} 1 & \text{if }|x|\geq k,\\ 0 & \text{otherwise.} \end{cases}$$ As a generate case, we have $${\operatorname{THR}}_{0}\equiv1.\label{eq:THR0}$$ In the *$k$-element distinctness problem*, the input is a list of $n$ integers from some range of size $r,$ and the objective is to determine whether some integer occurs at least $k$ times. Traditionally, the input to $k$-element distinctness is represented by a Boolean matrix $x\in\zoo^{n\times r}$ with precisely one nonzero entry in each row. We depart from tradition by allowing the input $x\in\zoo^{n\times r}$ to be an arbitrary matrix with at most $n$ ones. Formally, we define the $k$-element distinctness function $\operatorname{ED}_{n,r,k}\colon\zoo_{\leq n}^{nr}\to\zoo$ by $$\operatorname{ED}_{n,r,k}(x)=\neg\bigvee_{i=1}^{r}{\operatorname{THR}}_{k}(x_{1,i}x_{2,i}\ldots x_{n,i}).$$ Since our focus is on upper bounds, working with the more general domain makes our results stronger. Our main result in this section is as follows. \[thm:ED\]Let $k\geq1$ be a fixed integer. Then for all integers $n,r\geq1$ and all reals $0<\epsilon\leq1/2,$ $$\begin{gathered} \deg_{\epsilon}(\operatorname{ED}_{n,r,k})=O\left(\sqrt{n}\min\{n,r\}^{\frac{1}{2}-\frac{1}{4(1-2^{-k})}}\,\left(\log\frac{1}{\epsilon}\right)^{\frac{1}{4(1-2^{-k})}}\right)\\ +O\left(\sqrt{n\log\frac{1}{\epsilon}}\right).\end{gathered}$$ Moreover, the approximating polynomial is given explicitly in each case. Taking $\epsilon=1/3$ in this result settles Theorem \[thm:MAIN-ed\] from the introduction. To prove Theorem \[thm:ED\], we will need to consider a more general class of functions. For nonnegative integers $n,r,k$ and a real number $\Delta\geq1,$ we define $$D(n,r,k,\Delta)=\max_{F}\;\deg_{2^{-\Delta}}(F),$$ where the maximum is over all functions $F\colon\zoo_{\leq n}^{N}\to\zoo$ for some $N$ that are expressible as $$F(x)=\bigvee_{i=1}^{r}{\operatorname{THR}}_{k_{i}}(x|_{S_{i}})$$ for some pairwise disjoint sets $S_{1},S_{2},\ldots,S_{r}\subseteq\{1,2,\ldots,N\}$ and some $k_{1},k_{2},\ldots,k_{r}\in\{0,1,2,\ldots,k\}.$ The four-argument quantity $D$ that we have just defined is unrelated to the three-argument quantity $D$ from Section \[sec:k-DNF-and-k-CNF\]. We abbreviate $$D(n,\infty,k,\Delta)=\max_{r\geq1}D(n,r,k,\Delta).$$ By definition, $$\begin{aligned} \deg_{\epsilon}(\operatorname{ED}_{n,r,k}) & \leq D\left(n,r,k,\log\frac{1}{\epsilon}\right), & & 0<\epsilon\leq\frac{1}{2}.\label{eq:Dnrkd-vs-ED}\end{aligned}$$ Our analysis of $D(n,r,k,\Delta)$ proceeds by induction on $k.$ As the base cases, we have $$\begin{aligned} & D(n,\infty,0,\Delta)=0\label{eq:ed-recursive-base-case}\end{aligned}$$ by (\[eq:THR0\]), and $$D(n,\infty,1,\Delta)=C\sqrt{n\Delta}\label{eq:ed-recursive-base-case-OR}$$ by Theorem \[thm:DISJUNCTION\] for some constant $C\geq1.$ Also, Fact \[fact:trivial-upper-bound-on-deg\] implies that $$D(n,\infty,k,\Delta)\leq n.\label{eq:ed-recursive-trivial}$$ A recursive bound for small range --------------------------------- To implement the inductive step, we derive two complementary recursive bounds for $D(n,r,k,\Delta)$. The first of these bounds, presented below, is tailored to the case when $n\geq kr.$ \[lem:ed-recursive-small-range\]There is a constant $C\geq1$ such that for all positive integers $n,r,k$ and all reals $\Delta\geq1,$ $$D(n,r,k,\Delta)\leq C\cdot\sqrt{1+\frac{n}{kr}}\cdot(D(2kr,r,k,\Delta+1)+\Delta).$$ Since $D$ is monotonically increasing in every argument, the lemma holds trivially for $n<kr.$ In what follows, we consider the complementary case $$n\geq kr.\label{eq:n-kr}$$ Consider an arbitrary function $F\colon\zoo_{\leq n}^{N}\to\zoo$ of the form $$F(x)=\bigvee_{i=1}^{r}{\operatorname{THR}}_{k_{i}}(x|_{S_{i}})\label{eq:ed-recursive2-F}$$ for some pairwise disjoint sets $S_{1},S_{2},\ldots,S_{r}\subseteq\{1,2,\ldots,N\}$ and $k_{1},k_{2},\ldots,k_{r}\in\{0,1,2,\ldots,k\}.$ By discarding any irrelevant variables among $x_{1},x_{2},\ldots,x_{N}$, we may assume that $S_{1}\cup S_{2}\cup\cdots\cup S_{r}=\{1,2,\ldots,N\}.$ Then by the pigeonhole principle, any input $x$ with Hamming weight at least $kr$ satisfies at least one of the disjuncts in (\[eq:ed-recursive2-F\]). Therefore, $$\begin{aligned} F(x) & =1, & & x\in\zoo_{\geq kr}^{N}.\label{eq:F-const-beyond-kr}\end{aligned}$$ For $i\geq kr,$ define $F_{i}\colon\zoo_{\leq i}^{N}\to\zoo$ by $$F_{i}(x)=\begin{cases} F(x) & \text{if }|x|\leq kr,\\ 1 & \text{otherwise.} \end{cases}$$ Then $$\begin{aligned} \deg_{2^{-\Delta}}(F) & =\deg_{2^{-\Delta}}(F_{n})\\ & =\deg_{2^{-\Delta}}(1-F_{n})\\ & \leq c\sqrt{\frac{n}{kr}}\cdot(\deg_{2^{-\Delta-1}}(1-F_{2kr})+\Delta+1)\\ & =c\sqrt{\frac{n}{kr}}\cdot(\deg_{2^{-\Delta-1}}(F_{2kr})+\Delta+1)\\ & \leq c\sqrt{\frac{n}{kr}}\cdot(D(2kr,r,k,\Delta+1)+\Delta+1)\end{aligned}$$ for some absolute constant $c\geq1$ and all $\Delta\geq1,$ where the first and last steps use (\[eq:F-const-beyond-kr\]), and the third step applies (\[eq:n-kr\]) and the extension theorem (Theorem \[thm:extension\]) with $m=kr$ and $\epsilon=\delta=2^{-\Delta-1}$. As a result, the lemma holds with $C=2c.$ A recursive bound for large range --------------------------------- We now derive an alternate upper bound on $D(n,r,k,\Delta),$ with no dependence on the range parameter $r$. This result addresses the case of large $r$ and complements Lemma \[lem:ed-recursive-small-range\]. \[lem:ed-recursive-large-range\]There is a constant $C\geq1$ such that for all integers $n,k\geq1$ and all reals $\Delta,b\geq1,$ $$\begin{gathered} D(n,\infty,k,\Delta)\leq C\sqrt{nb\Delta}+C\left(1+\frac{1}{\sqrt{k}}\left(\frac{n}{b\Delta}\right)^{1/4}\right)\times\\ \times\left(D\left(\lfloor Ck\sqrt{nb\Delta}\rfloor,\infty,k-1,C\sqrt{\frac{n\Delta}{b}}+1\right)+\sqrt{\frac{n\Delta}{b}}\right).\label{eq:recursive1}\end{gathered}$$ Consider an arbitrary function $F\colon\zoo_{\leq n}^{N}\to\zoo$ of the form $$F(x)=\bigvee_{i=1}^{r}{\operatorname{THR}}_{k_{i}}(x|_{S_{i}})\label{eq:F-or-of-thrs}$$ for some integer $r\geq1,$ some pairwise disjoint sets $S_{1},S_{2},\ldots,S_{r}\subseteq\{1,2,\ldots,N\}$, and some $k_{1},k_{2},\ldots,k_{r}\in\{0,1,2,\ldots,k\}.$ If $k_{i}=0$ for some $i,$ then the corresponding term in (\[eq:F-or-of-thrs\]) is the constant function $1,$ resulting in $\deg_{0}(F)=0.$ In what follows, we treat the complementary case when $k_{i}\geq1$ for each $i.$ Rewriting (\[eq:F-or-of-thrs\]), $$F(x)=\bigvee_{i=1}^{r}\bigvee_{j\in S_{i}}x_{j}\wedge{\operatorname{THR}}_{k_{i}-1}(x|_{S_{i}\setminus\{j\}}).\label{eq:ed-recursive1-F}$$ As this representation suggests, our intention is to bound the approximate degree of $F$ by appeal to Theorem \[thm:selector\]. \[claim:ed-recursive1\]Fix a subset $S_{i}'\subseteq S_{i}$ for each $i=1,2,\ldots,r.$ Then for $\Delta\geq1,$ $$\begin{gathered} \deg_{2^{-\Delta}}\left(\bigvee_{i=1}^{r}\bigvee_{j\in S'_{i}}{\operatorname{THR}}_{k_{i}-1}(x|_{S_{i}\setminus\{j\}})\right)\\ \leq D\left(n,\sum_{i=1}^{r}|S_{i}'|,k-1,\Delta\right)+\sum_{i=1}^{r}|S'_{i}|.\qquad\qquad\end{gathered}$$ We will settle Claim \[claim:ed-recursive1\] once we complete the main proof. In light of this claim, the representation (\[eq:ed-recursive1-F\]) shows that $$F(x)=\bigvee_{i=1}^{N}x_{i}\wedge f_{i}(x)$$ for some functions $f_{i}$ such that $$\begin{aligned} \deg_{2^{-\Delta}}\left(\bigvee_{i\in S}f_{i}\right) & \leq D(n,|S|,k-1,\Delta)+|S|\label{eq:ors-of-thrs}\end{aligned}$$ for all $S\subseteq\{1,2,\ldots,N\}$ and all $\Delta\geq1.$ Then for some absolute constants $c',c''\geq1$ and all $\Delta\geq1$ and $b\geq1$, we have $$\begin{aligned} \deg_{2^{-\Delta}}(F) & \leq c'\sqrt{nb\Delta}+\max_{\substack{S\subseteq\{1,\ldots,N\}\\ |S|\leq c'\sqrt{nb\Delta} } }\deg_{2^{-\Delta}\exp\left(-c'\sqrt{n\Delta/b}\right)}\left(\bigvee_{i\in S}f_{i}\right)\\ & \leq2c'\sqrt{nb\Delta}+D\left(n,\lceil c'\sqrt{nb\Delta}\rceil,k-1,\Delta+\frac{c'}{\ln2}\sqrt{\frac{n\Delta}{b}}\right)\\ & \leq2c'\sqrt{nb\Delta}+c''\cdot\sqrt{1+\frac{n}{k\cdot c'\sqrt{nb\Delta}}}\times\phantom{a}\\ & \qquad\times\left(D\left(2k\lceil c'\sqrt{nb\Delta}\rceil,\infty,k-1,\Delta+\frac{c'}{\ln2}\sqrt{\frac{n\Delta}{b}}+1\right)\right.\\ & \qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\left.\phantom{a}+\Delta+\frac{c'}{\ln2}\sqrt{\frac{n\Delta}{b}}\right),\end{aligned}$$ where the first step applies Theorem \[thm:selector\], the second step uses (\[eq:ors-of-thrs\]), and the final step follows from (\[eq:ed-recursive-base-case\]) for $k=1$ and from Lemma \[lem:ed-recursive-small-range\] for $k\geq2$. This directly implies (\[eq:recursive1\]) for $\Delta\leq n/b.$ In the complementary case $\Delta>n/b,$ the right-hand side of (\[eq:recursive1\]) exceeds $n$ and therefore the bound follows trivially from (\[eq:ed-recursive-trivial\]). To start with, $$\begin{aligned} \bigvee_{i=1}^{r}\bigvee_{j\in S'_{i}}{\operatorname{THR}}_{k_{i}-1}(x|_{S_{i}\setminus\{j\}}) & =\bigvee_{i:S'_{i}\ne\varnothing}\;\bigvee_{j\in S'_{i}}{\operatorname{THR}}_{k_{i}-1}(x|_{S_{i}\setminus\{j\}})\\ & =\bigvee_{i:S'_{i}\ne\varnothing}{\operatorname{THR}}_{k_{i}-1-\min\left\{ \left|x|_{S'_{i}}\right|,\,|S_{i}'|-1\right\} }(x|_{S_{i}\setminus S_{i}'}).\\\end{aligned}$$ Considering the possible values for the Hamming weight of each $x|_{S'_{i}}$, we arrive at the representation $$\begin{gathered} \bigvee_{i=1}^{r}\;\bigvee_{j\in S'_{i}}{\operatorname{THR}}_{k_{i}-1}(x|_{S_{i}\setminus\{j\}})=\sum_{\ell_{1}=0}^{|S'_{1}|}\cdots\sum_{\ell_{r}=0}^{|S'_{r}|}\I[|x|_{S'_{i}}|=\ell_{i}\text{ for each }i]\\ \times\left(\bigvee_{i:S_{i}'\ne\varnothing}{\operatorname{THR}}_{k_{i}-1-\min\{\ell_{i},\,|S_{i}'|-1\}}(x|_{S_{i}\setminus S_{i}'})\right).\label{eq:approx-disjunction-of-thrs}\end{gathered}$$ The indicator functions in this summation are mutually exclusive in that for any given value of $x,$ precisely one of them is nonzero. As a result, the right-hand side of (\[eq:approx-disjunction-of-thrs\]) can be approximated pointwise to within $2^{-\Delta}$ by replacing each parenthesized expression with its $2^{-\Delta}$-error approximant, which by definition can be chosen to have degree at most $D(n,\sum|S_{i}'|,k-1,\Delta).$ This completes the proof since each indicator function in (\[eq:approx-disjunction-of-thrs\]) depends on only $\sum|S_{i}'|$ Boolean variables and is therefore a polynomial of degree at most $\sum|S'_{i}|.$ Solving the recurrence ---------------------- It remains to solve the newly obtained recurrences. We first solve the recurrence given by (\[eq:ed-recursive-base-case-OR\]) and Lemma \[lem:ed-recursive-large-range\], corresponding to the infinite-range case. \[thm:ed-range-independent\]There is a constant $c\geq1$ such that for all positive integers $n$ and $k,$ and all reals $\Delta\geq1,$ $$\begin{aligned} D(n,\infty,k,\Delta) & \leq c^{k}\sqrt{k!}\cdot n^{1-\frac{1}{4(1-2^{-k})}}\Delta^{\frac{1}{4(1-2^{-k})}}.\label{eq:main-recursive-large-range-k>0}\end{aligned}$$ We will prove (\[eq:main-recursive-large-range-k&gt;0\]) for $c=(4C)^{2},$ where $C\geq1$ is the larger of the constants in (\[eq:ed-recursive-base-case-OR\]) and Lemma \[lem:ed-recursive-large-range\]. The proof is by induction on $k.$ The base case $k=1$ is immediate from (\[eq:ed-recursive-base-case-OR\]). For the inductive step, let $k\geq2$ be arbitrary. When $\Delta>n,$ the right-hand side of (\[eq:main-recursive-large-range-k&gt;0\]) exceeds $n$ and therefore the bound is immediate from (\[eq:ed-recursive-trivial\]). In what follows, we assume that $$1\leq\Delta\leq n.\label{eq:ed-Delta-not-too-large}$$ Let $b\geq1$ be a parameter to be fixed later. By Lemma \[lem:ed-recursive-large-range\], $$\begin{gathered} D(n,\infty,k,\Delta)\leq C\sqrt{nb\Delta}+C\left(1+\frac{1}{\sqrt{k}}\left(\frac{n}{b\Delta}\right)^{\frac{1}{4}}\right)\left(\sqrt{\frac{n\Delta}{b}}\right.\\ \left.\phantom{a}+D\left(\lfloor Ck\sqrt{nb\Delta}\rfloor,\infty,k-1,C\sqrt{\frac{n\Delta}{b}}+1\right)\right).\end{gathered}$$ It follows that $$\begin{gathered} D(n,\infty,k,\Delta)\leq C\sqrt{knb\Delta}+2C\left(\frac{n}{kb\Delta}\right)^{\frac{1}{4}}\left(\sqrt{\frac{n\Delta}{b}}\right.\\ \left.\phantom{a}+D\left(\lfloor Ck\sqrt{nb\Delta}\rfloor,\infty,k-1,C\sqrt{\frac{n\Delta}{b}}+1\right)\right),\end{gathered}$$ as one can verify from the previous step if $n\geq kb\Delta$ and from (\[eq:ed-recursive-trivial\]) if $n<kb\Delta.$ Applying the inductive hypothesis, $$\begin{gathered} D(n,\infty,k,\Delta)\leq C\sqrt{knb\Delta}+2C\left(\frac{n}{kb\Delta}\right)^{\frac{1}{4}}\left(\sqrt{\frac{n\Delta}{b}}\right.\\ \left.\phantom{a}+c^{k-1}\sqrt{(k-1)!}\cdot(Ck\sqrt{nb\Delta})^{1-\frac{1}{4(1-2^{-k+1})}}\left(C\sqrt{\frac{n\Delta}{b}}+1\right)^{\frac{1}{4(1-2^{-k+1})}}\right).\end{gathered}$$ Now the bound $$\begin{gathered} D(n,\infty,k,\Delta)\leq C\sqrt{knb\Delta}+2C\left(\frac{n}{kb\Delta}\right)^{\frac{1}{4}}\times\\ \times4c^{k-1}\sqrt{(k-1)!}\cdot(Ck\sqrt{nb\Delta})^{1-\frac{1}{4(1-2^{-k+1})}}\left(C\sqrt{\frac{n\Delta}{b}}\right)^{\frac{1}{4(1-2^{-k+1})}}\end{gathered}$$ is immediate from the previous step if $n\geq b/\Delta$ and from (\[eq:ed-recursive-trivial\]) if $n<b/\Delta.$ Rearranging, we find that $$\begin{gathered} D(n,\infty,k,\Delta)\\ \leq C\sqrt{knb\Delta}\left(1+C\left(\frac{n}{\Delta}\right)^{\frac{1}{4}}\cdot8c^{k-1}\sqrt{(k-1)!}\;b^{-\frac{1}{4}-\frac{1}{4(1-2^{-k+1})}}\right).\label{eq:D-infinity-induction-simplified}\end{gathered}$$ The right-hand side is minimized at $$b=\left(C\left(\frac{n}{\Delta}\right)^{\frac{1}{4}}\cdot8c^{k-1}\sqrt{(k-1)!}\right)^{\frac{2^{k+1}-4}{2^{k}-1}},$$ which in view of (\[eq:ed-Delta-not-too-large\]) is a real number in $[1,\infty)$ and therefore a legitimate parameter setting. Making this substitution in (\[eq:D-infinity-induction-simplified\]), we arrive at $$\begin{aligned} D(n,\infty,k,\Delta) & \leq2C\sqrt{kn\Delta}\left(C\left(\frac{n}{\Delta}\right)^{\frac{1}{4}}\cdot8c^{k-1}\sqrt{(k-1)!}\right)^{\frac{2^{k}-2}{2^{k}-1}}\\ & \leq2C^{2}\cdot8c^{k-1}\sqrt{k!\,n\Delta\left(\frac{n}{\Delta}\right)^{\frac{2^{k-1}-1}{2^{k}-1}}}\\ & =c^{k}\sqrt{k!}\,n^{1-\frac{1}{4(1-2^{-k})}}\Delta^{\frac{1}{4(1-2^{-k})}}.\end{aligned}$$ This completes the inductive step and settles (\[eq:main-recursive-large-range-k&gt;0\]). By combining the previous result with an application of Lemma \[lem:ed-recursive-small-range\], we will now prove our main bound on $D(n,r,k,\Delta).$ \[thm:ed-range-dependent\]There is a constant $c\geq1$ such that for all positive integers $n,r,k$ and all reals $\Delta\geq1,$ $$\begin{aligned} D(n,r,k,\Delta) & \leq c^{k}\sqrt{k!}\left(\sqrt{n}\min\{n,kr\}^{\frac{1}{2}-\frac{1}{4(1-2^{-k})}}\,\Delta^{\frac{1}{4(1-2^{-k})}}+\sqrt{n\Delta}\right).\end{aligned}$$ The bound follows from Theorem \[thm:ed-range-independent\] if $kr\geq n$; and from (\[eq:ed-recursive-trivial\]) if $\Delta\geq n.$ As a result, we may assume that $$\begin{aligned} n & >kr,\label{eq:ed-range-dep-n-kr}\\ n & >\Delta.\label{eq:ed-range-dep-n-Delta}\end{aligned}$$ In what follows, let $C\geq1$ denote the larger of the constants in Lemma \[lem:ed-recursive-small-range\] and Theorem \[thm:ed-range-independent\]. Then $$\begin{aligned} & D(n,r,k,\Delta)\\ & \qquad\leq D\left(n,r+\left\lceil \frac{\Delta}{k}\right\rceil ,k,\Delta\right)\\ & \qquad\leq C\cdot\sqrt{1+\frac{n}{kr+k\lceil\Delta/k\rceil}}\cdot\left(D\left(2kr+2k\left\lceil \frac{\Delta}{k}\right\rceil ,\infty,k,\Delta+1\right)+\Delta\right)\\ & \qquad\leq2C\cdot\sqrt{\frac{n}{kr+k\lceil\Delta/k\rceil}}\cdot\left(D\left(2kr+2k\left\lceil \frac{\Delta}{k}\right\rceil ,\infty,k,\Delta+1\right)+\Delta\right)\\ & \qquad\leq2C\cdot\sqrt{\frac{n}{kr+k\lceil\Delta/k\rceil}}\\ & \qquad\qquad\qquad\times\left(C^{k}\sqrt{k!}\left(2kr+2k\left\lceil \frac{\Delta}{k}\right\rceil \right)^{1-\frac{1}{4(1-2^{-k})}}(\Delta+1)^{\frac{1}{4(1-2^{-k})}}+\Delta\right)\allowdisplaybreaks\\ & \qquad\leq2C\cdot\sqrt{\frac{n}{kr+k\lceil\Delta/k\rceil}}\\ & \qquad\qquad\qquad\times2C^{k}\sqrt{k!}\left(2kr+2k\left\lceil \frac{\Delta}{k}\right\rceil \right)^{1-\frac{1}{4(1-2^{-k})}}(\Delta+1)^{\frac{1}{4(1-2^{-k})}}\\ & \qquad=4C^{k+1}\sqrt{k!}\cdot\sqrt{2n}\left(2kr+2k\left\lceil \frac{\Delta}{k}\right\rceil \right)^{\frac{1}{2}-\frac{1}{4(1-2^{-k})}}(\Delta+1)^{\frac{1}{4(1-2^{-k})}},\end{aligned}$$ where the first step is valid because $D$ is monotonically increasing in every argument; the second step applies Lemma \[lem:ed-recursive-small-range\]; the third step uses (\[eq:ed-range-dep-n-kr\]) and (\[eq:ed-range-dep-n-Delta\]); and the fourth step applies Theorem \[thm:ed-range-independent\]. This completes the proof of the theorem for $n>kr.$ Equation (\[eq:Dnrkd-vs-ED\]) and Theorem \[thm:ed-range-dependent\] establish the main result of this section, Theorem \[thm:ED\]. We note that with a more careful analysis, the multiplicative factor $\sqrt{k!}$ in Theorems \[thm:ed-range-independent\] and \[thm:ed-range-dependent\] can be improved to a slightly smaller quantity, still of the order of $k^{O(k)}.$ Surjectivity ============ For positive integers $n$ and $r$, the *surjectivity problem* is to determine whether a given mapping $\{1,2,\ldots,n\}\to\{1,2,\ldots,r\}$ is surjective. Traditionally, the input to this problem is represented by a Boolean matrix $x\in\zoo^{n\times r}$ with precisely one nonzero entry in every row. Analogous to our work on element distinctness in the previous section, we depart from tradition by allowing arbitrary matrices $x\in\zoo^{n\times r}$ with at most $n$ ones. Specifically, we define the surjectivity function ${\operatorname{SURJ}}_{n,r}\colon\zoo_{\leq n}^{nr}\to\zoo$ by $${\operatorname{SURJ}}_{n,r}(x)=\bigwedge_{j=1}^{r}\bigvee_{i=1}^{n}x_{i,j}.$$ This formalism corresponds to determining the surjectivity of arbitrary relations on $\{1,2,\ldots,n\}\times\{1,2,\ldots,r\}$, including functions $\{1,2,\ldots,n\}\to\{1,2,\ldots,r\}$ as a special case. Since we are interested in upper bounds, working in this more general setting makes our results stronger. Approximation to 1/3 -------------------- For clarity of exposition, we first bound the approximate degree of surjectivity with the error parameter set to $\epsilon=1/3.$ This setting covers most applications of interest and allows for a shorter and simpler proof. Readers with an interest in general $\epsilon$ can skip directly to Section \[subsec:Approximation-to-any-eps\]. For all positive integers $n$ and $r,$ $$\begin{aligned} & \deg_{1/3}({\operatorname{SURJ}}_{n,r})=O(\sqrt{n}\cdot r^{1/4}) & & (r\leq n),\label{eq:surj-small-range-0.33333}\\ & \deg_{1/3}({\operatorname{SURJ}}_{n,r})=0 & & (r>n).\label{eq:surj-large-range-0.33333}\end{aligned}$$ Moreover, the approximating polynomial is given explicitly in each case. The theorem shows that $\deg_{1/3}({\operatorname{SURJ}}_{n,r})=O(n^{3/4})$ for all $r,$ disproving the conjecture of Bun and Thaler [@bun-thaler17adeg-ac0] that the $1/3$-approximate degree of ${\operatorname{SURJ}}_{n,\Omega(n)}$ is linear in $n.$ The identity ${\operatorname{SURJ}}_{n,r}\equiv0$ for $r>n$ implies (\[eq:surj-large-range-0.33333\]) directly. The proof of (\[eq:surj-small-range-0.33333\]) involves two steps. First, we construct an explicit real-valued function $\widetilde{{\operatorname{SURJ}}}_{n,r}$ that approximates ${\operatorname{SURJ}}_{n,r}$ pointwise and is representable by a linear combination of conjunctions with reasonably small coefficients. Then, we replace each conjunction in this linear combination by an approximating polynomial of low degree. In more detail, let $m\geq1$ be an integer parameter to be chosen later. Recall from (\[eqn:chebyshev-containment\]) and Proposition \[prop:chebyshev-beyond-1\] that the Chebyshev polynomial $T_{m}$ obeys $$\begin{aligned} & |T_{m}(t)|\leq1, & & -1\leq t\leq t,\\ & T_{m}\left(1+\frac{1}{r}\right)\geq1+\frac{m^{2}}{r}.\end{aligned}$$ As a result, ${\operatorname{SURJ}}_{n,r}$ is approximated pointwise within $1/(1+\frac{m^{2}}{r})$ by $$\widetilde{{\operatorname{SURJ}}}_{n,r}(x)=\frac{1}{T_{m}(1+\frac{1}{r})}\cdot T_{m}\left(\frac{1}{r}+\frac{1}{r}\sum_{j=1}^{r}\bigvee_{i=1}^{n}x_{i,j}\right).$$ Therefore, $$\begin{aligned} E({\operatorname{SURJ}}_{n,r},d) & \leq\frac{1}{1+\frac{m^{2}}{r}}+E(\widetilde{{\operatorname{SURJ}}}_{n,r},d), & & d=1,2,3,\ldots.\label{eq:approximate-surj}\end{aligned}$$ To estimate the rightmost term in (\[eq:approximate-surj\]), use the factored representation (\[eq:chebyshev-factored\]) to write $$\begin{aligned} \widetilde{{\operatorname{SURJ}}}_{n,r}(x) & =\frac{2^{m-1}}{T_{m}(1+\frac{1}{r})}\cdot\prod_{i=1}^{m}\left(\frac{1}{r}+\frac{1}{r}\sum_{j=1}^{r}\left(\bigvee_{i=1}^{n}x_{i,j}\right)-\cos\frac{(2i-1)\pi}{2m}\right)\\ & =\frac{2^{m-1}}{T_{m}(1+\frac{1}{r})}\cdot\prod_{i=1}^{m}\left(\frac{1}{r}+1-\frac{1}{r}\sum_{j=1}^{r}\prod_{i=1}^{n}\overline{x_{i,j}}-\cos\frac{(2i-1)\pi}{2m}\right).\end{aligned}$$ Multiplying out shows that $\widetilde{{\operatorname{SURJ}}}_{n,r}(x)$ is a linear combination of conjunctions with real coefficients whose absolute values sum to $2^{O(m)}.$ By Corollary \[cor:CONJUNCTION\], each of these conjunctions can be approximated by a polynomial of degree $d$ to within $2^{-\Theta(d^{2}/n)}$ pointwise. We conclude that $$E(\widetilde{{\operatorname{SURJ}}}_{n,r},d)\leq2^{O(m)}\cdot2^{-\Theta(d^{2}/n)},$$ which along with (\[eq:approximate-surj\]) gives $$E({\operatorname{SURJ}}_{n,r},d)\leq\frac{1}{1+\frac{m^{2}}{r}}+2^{O(m)}\cdot2^{-\Theta(d^{2}/n)}.$$ Now (\[eq:surj-small-range-0.33333\]) follows by taking $m=\lceil\sqrt{3r}\rceil$ and $d=\Theta(\sqrt{n}\cdot r^{1/4})$. The approximating polynomial in question is given explicitly because every stage of our proof, including the appeal to Corollary \[cor:CONJUNCTION\], is constructive. \[subsec:Approximation-to-any-eps\]Approximation to arbitrary error ------------------------------------------------------------------- We now generalize the previous theorem to arbitrary $\epsilon$. The proof closely mirrors the case of $\epsilon=1/3$ but features additional ingredients, such as Lemma \[lem:bound-on-poly-coeffs-UNIVARIATE\]. \[thm:SURJ\]For all positive integers $n$ and $r,$ and all reals $0<\epsilon<1/2,$ $$\begin{aligned} & \deg_{\epsilon}({\operatorname{SURJ}}_{n,r})=O\left(\sqrt{n}\left(r\log\frac{1}{\epsilon}\right)^{1/4}+\sqrt{n\log\frac{1}{\epsilon}}\right) & & (r\leq n),\label{eq:surj-small-range}\\ & \deg_{\epsilon}({\operatorname{SURJ}}_{n,r})=0 & & (r>n).\label{eq:surj-large-range}\end{aligned}$$ Moreover, the approximating polynomial is given explicitly in each case. As before, we need only prove (\[eq:surj-small-range\]) since ${\operatorname{SURJ}}_{n,r}\equiv0$ for $r>n.$ Theorem \[thm:AND\] provides, after rescaling, an explicit univariate polynomial $p$ such that $$\begin{aligned} & p(1)=1,\label{eq:surj-general-p-at-1}\\ & |p(t)|\leq\frac{\epsilon}{2}, & & t\in\left\{ 0,\frac{1}{r},\frac{2}{r},\ldots,\frac{r-1}{r}\right\} ,\label{eq:surj-general-p-intermediate}\\ & |p(t)|\leq1, & & t\in[0,1],\label{eq:surj-general-p-bounded}\\ & \deg p=O\left(\sqrt{r\log\frac{1}{\epsilon}}\right).\label{eq:surj-general-p-deg}\end{aligned}$$ Now define $\widetilde{{\operatorname{SURJ}}}_{n,r}\colon\zoo_{\leq n}^{nr}\to\Re$ by $$\widetilde{{\operatorname{SURJ}}}_{n,r}(x)=p\left(\frac{1}{r}\sum_{j=1}^{r}\bigvee_{i=1}^{n}x_{i,j}\right).$$ This function clearly approximates ${\operatorname{SURJ}}_{n,r}$ pointwise to $\epsilon/2.$ It follows that for any $d,$ $$\begin{aligned} E({\operatorname{SURJ}}_{n,r},d) & \leq\|{\operatorname{SURJ}}_{n,r}-\widetilde{{\operatorname{SURJ}}}_{n,r}\|_{\infty}+E(\widetilde{{\operatorname{SURJ}}}_{n,r},d)\nonumber \\ & \leq\frac{\epsilon}{2}+E(\widetilde{{\operatorname{SURJ}}}_{n,r},d).\label{eq:surj-general-intermediate-error}\end{aligned}$$ We have $$\begin{aligned} \operatorname{\Pi}(\widetilde{{\operatorname{SURJ}}}_{n,r}) & \leq\max\left\{ 1,\operatorname{\Pi}\!\left(\frac{1}{r}\sum_{j=1}^{r}\bigvee_{i=1}^{n}x_{i,j}\right)\right\} ^{\deg p}{|\!|\!| p | \! | \! |}\nonumber \\ & \leq2^{\deg p}\;{|\!|\!| p | \! | \! |}\nonumber \\ \rule{0mm}{5mm} & \leq16^{\deg p}\nonumber \\ & \leq2^{O\left(\sqrt{r\log(1/\epsilon)}\right)},\label{eq:surj-tilde-orcomplexity}\end{aligned}$$ where the first and second steps use Proposition \[prop:orcomplexity\] \[item:orcomplexity-composition-with-univariate\], \[item:orcomplexity-disjunction\]; the third step follows from (\[eq:surj-general-p-bounded\]) and Lemma \[lem:bound-on-poly-coeffs-UNIVARIATE\]; and the final step is valid by (\[eq:surj-general-p-deg\]). To restate (\[eq:surj-tilde-orcomplexity\]), we have shown that $\widetilde{{\operatorname{SURJ}}}_{n,r}$ is a linear combination of conjunctions with real coefficients whose absolute values sum to $\exp(O(\sqrt{r\log(1/\epsilon)}))$. By Corollary \[cor:CONJUNCTION\], each of these conjunctions can be approximated by a polynomial of degree $d$ to within $2^{-\Theta(d^{2}/n)}$ pointwise. We conclude that $$E(\widetilde{{\operatorname{SURJ}}}_{n,r},d)\leq2^{O\left(\sqrt{r\log(1/\epsilon)}\right)}\cdot2^{-\Theta(d^{2}/n)},$$ which along with (\[eq:surj-general-intermediate-error\]) gives $$E({\operatorname{SURJ}}_{n,r},d)\leq\frac{\epsilon}{2}+2^{O\left(\sqrt{r\log(1/\epsilon)}\right)}\cdot2^{-\Theta(d^{2}/n)}.$$ Now (\[eq:surj-small-range\]) follows by taking $$d=\Theta\left(\sqrt{n}\left(r\log\frac{1}{\epsilon}\right)^{1/4}+\sqrt{n\log\frac{1}{\epsilon}}\right).$$ Finally, the approximating polynomial in question is given explicitly because every stage of our proof, including the appeal to Theorem \[thm:AND\] and Corollary \[cor:CONJUNCTION\], is constructive. Acknowledgments {#acknowledgments .unnumbered} =============== The author is thankful to Paul Beame, Aleksandrs Belovs, Mark Bun, Robin Kothari, Justin Thaler, Emanuele Viola, and Ronald de Wolf for valuable comments on an earlier version of this paper. The author is further indebted to Mark, Robin, and Justin for stimulating discussions and for sharing a preliminary version of their manuscript [@BKT17poly-strikes-back], which inspired the title of this paper. [^1]: $^{*}$ Computer Science Department, UCLA, Los Angeles, CA 90095. [ ]{}`[email protected] `Supported by NSF CAREER award CCF-1149018 and an Alfred P. Sloan Foundation Research Fellowship. [^2]: Alternately, the input can be represented by a string of $n\lceil\log r\rceil$ bits. Switching to this more compact representation changes the complexity of the problem by a factor of at most $\lceil\log r\rceil,$ which is negligible in all settings of interest.

ArXiv

--- abstract: 'We consider the Landau-Khalatnikov two-fluid hydrodynamics of superfluid liquid [@Khalatnikov] as an effective theory, which provides a self-consistent analog of Einstein equations for gravity and matter.' address: | Helsinki University of Technology, Low Temperature Laboratory, P.O. Box 2200, FIN-02015 HUT, Finland\ L. D. Landau Institute for Theoretical Physics, RAS, Kosygin Str. 2, 117940 Moscow, Russia author: - 'G.E. Volovik' title: Links between gravity and dynamics of quantum liquids --- **IV International Conference “Cosmology. Relativistic Astropohysics. Cosmoparticle Physics”.** (COSMION-99) In the Honor of 80-th Birthday of Isaak M. Khalatnikov Introduction. Physical vacuum as condensed matter. ================================================== In a modern viewpoint the relativistic quantum field theory is an effective theory [@Weinberg]. It is an emergent phenomenon arising in the low energy corner of the physical fermionic vacuum – the medium, whose nature remains unknown. Also it is argued that in the low energy corner the symmetry must be enhanced [@Chadha]: If we neglect the very low energy region of electroweak scale, where some symmetries are spontaneously violated, then above this scale one can expect that the lower the energy, the better is the Lorentz invariance and other symmetries of the physical laws. The same phenomena occur in the condensed matter systems. If the spontaneous symmetry breaking at very low energy is neglected or avoided, then in the limit of low energy the symmetry of condensed matter is really enhanced. Moreover, there is one special universality class of Fermi systems, where in the low energy corner there appear almost all the symmetries, which we know today in high energy physics: Lorentz invariance, gauge invariance, elements of general covariance, etc (superfluid $^3$He-A is a representative of this class [@parallel]). The chiral fermions as well as gauge bosons and gravity field arise as fermionic and bosonic collective modes of such a system. The inhomogeneous states of the condensed matter ground state – vacuum – induce nontrivial effective metrics of the space, where the free quasiparticles move along geodesics. This conceptual similarity between condensed matter and quantum vacuum allows us to simulate many phenomena in high energy physics and cosmology, including axial anomaly, baryoproduction and magnetogenesis, event horizon and Hawking radiation, rotating vacuum, expansion of the Universe, etc., probing these phenomena in ultra-low-temperature superfluid helium, atomic Bose condensates and superconductors. Some of the experiments have been already conducted. The quantum field theory, which we have now, is incomplete due to ultraviolet diveregences at small scales, where the “microscopic” physics of vacuum becomes important. Here the analogy between quantum vacuum and condensed matter could give an insight into the transPlanckian physics. As in condensed matter system, one can expect that some or all of the known symmetries in Nature will be lost when the Planck energy scale is approached. The condensed matter analogue gives examples of the physically imposed deviations from Lorentz invariance. This is important in many different areas of high energy physics and cosmology, including possible CPT violation and black holes, where the infinite red shift at the horizon opens the route to the transPlanckian physics. The low-energy properties of different condensed matter substances (magnets, superfluids, crystals, superconductors, etc.) are robust, i.e. they do not depend much on the details of microscopic (atomic) structure of these substances. The main role is played by symmetry and topology of condensed matter: they determine the soft (low-energy) hydrodynamic variables, the effective Lagrangian describing the low-energy dynamics, and topological defects. The microscopic details provide us only with the “fundamental constants”, which enter the effective phenomenological Lagrangian, such as speed of “light” (say, the speed of sound), superfluid density, modulus of elasticity, magnetic susceptibility, etc. Apart from these fundamental constants, which can be rescaled, the systems behave similarly in the infrared limit if they belong to the same universality and symmetry classes, irrespective of their microscopic origin. The detailed information on the system is lost in such acoustic or hydrodynamic limit [@LaughlinPines]. From the properties of the low energy collective modes of the system – acoustic waves in case of crystals – one cannot reconstruct the atomic structure of the crystal since all the crystals have similar acoustic waves described by the same equations of the same effective theory, in a given case the classical theory of elasticity. The classical fields of collective modes can be quantized to obtain quanta of acoustic waves – the phonons, but this quantum field remains the effective field which does not give a detailed information on the real quantum structure of the underlying crystal. It is quite probable that in the same way the quantization of classical gravity, which is one of the infrared collective modes of quantum vacuum, will not add more to our understanding of the “microscopic” structure of the vacuum [@Hu96; @Padmanabhan; @LaughlinPines]. Indeed, according to this analogy, such properties of our world, as gravitation, gauge fields, elementary chiral fermions, etc., all arise in the low energy corner as a low-energy soft modes of the underlying “condensed matter”. At high energy (of the Planck scale) these soft modes disappear: actually they merge with the continuum of the high-energy degrees of freedom of the “Planck condensed matter” and thus cannot be separated anymore from the others. Since the gravity appears as an effective field in the infrared limit, the only output of its quantization would be the quanta of the low-energy gravitational waves – gravitons. The main advantage of the condensed matter analogy is that in principle we know the condensed matter structure at any relevant scale, including the interatomic distance, which corresponds to the Planck scale. Thus the condensed matter can suggest possible routes from our present low-energy corner of “phenomenology” to the “microscopic” physics at Planckian and trans-Planckian energies. Landau-Khalatnikov two-fluid model hydrodynamics as effective theory of gravity. ================================================================================ Superfluid vacuum and quasiparticles. ------------------------------------- Here we consider the simplest effective field theory of superfluids, where only the gravitational field appears as an effective field. The case of the fermi superfluids, where also the gauge field and chiral fermions appear in the low-energy corner together with Lorentz invariance is discussed in [@parallel; @LammiTalk]. According to Landau and Khalatnikov [@Khalatnikov] a weakly excited state of the collection of interacting $^4$He atoms can be considered as a small number of elementary excitations – quasiparticles, phonons and rotons. In addition, the state without excitation – the ground state or vacuum – can have collective degrees of freedom. The superfluid vacuum can move without friction, and inhomogeneity of the flow serves as the gravitational and/or other effective fields. The matter propagating in the presence of this background is represented by fermionic (in Fermi superfluids) or bosonic (in Bose superfluids) quasiparticles, which form the so called normal component of the liquid. Such two-fluid hydrodynamics introduced by Landau and Khalatnikov [@Khalatnikov] is the example of the effective field theory which incorporates the motion of both the superfluid background (gravitational field) and excitations (matter). This is the counterpart of the Einstein equations, which incorporate both gravity and matter. One must distinguish between the particles and quasiparticles in superfluids. The particles describes the system on a microscopic level, these are atoms of the underlying liquid ($^3$He or $^4$He atoms). The many-body system of the interacting atoms form the quantum vacuum – the ground state. The conservation laws experienced by the atoms and their quantum coherence in the superfluid state determine the low frequency dynamics – the hydrodynamics – of the collective variables of the superfluid vacuum. The quasiparticles – fermionic and bosonic – are the low energy excitations above the vacuum state. They form the normal component of the liquid which determines the thermal and kinetic low-energy properties of the liquid. Dynamics of superfluid vacuum. {#DynamicsSuperfluidVacuum} ------------------------------ In the simplest superfluid the coherent motion of the superfluid vacuum is characterized by two collective (hydrodynamic) variables: the particle density $n({\bf r},t)$ of atoms comprising the liquid and superfluid velocity $ {\bf v}_{\rm s}({\bf r},t)$ of their coherent motion. In a strict microscopic theory $n=\sum_{\bf p} n({\bf p})$, where $n({\bf p})$ is the particle distribution functions. The particle number conservation provides one of the equations of the effective theory of superfluids – the continuity equation $${\partial n\over \partial t}+ \nabla\cdot{\bf J}=0~. \label{ContinuityEquation}$$ The conserved current of atoms in monoatomic superfluid lquid is $${\bf J}= \sum_{\bf p} {{\bf p} \over m} n({\bf p})~, \label{TotalCurrent1}$$ where $m$ is the bare mass of the particle. Note that the liquids considered here are nonrelativistic and obeying the Galilean transformation law. In the Galilean system the momentum of particles and the particle current are related by Eq.(\[TotalCurrent1\]). In the effective theory the particle current has two contributions $${\bf J}= n {\bf v}_{\rm s}+ {\bf J}_{\rm q}~,~{\bf J}_{\rm q}=\sum_{\bf p} {{\bf p} \over m} f({\bf p})~. \label{TotalCurrent2}$$ The first term $n {\bf v}_{\rm s}$ is the current transferred coherently by the collective motion of superfluid vacuum with the superfluid velocity ${\bf v}_{\rm s}$. If quasiparticles are excited above the ground state, their momenta also contribute to the particle current, the second term in rhs of Eq.(\[TotalCurrent2\]), where $f({\bf p})$ is the distribution function of quasiparticles. Note that under the Galilean transformation to the coordinate system moving with the velocity ${\bf u}$ the superfluid velocity transforms as ${\bf v}_{\rm s}\rightarrow {\bf v}_{\rm s} + {\bf u}$, while the momenta of particle and quasiparticle transform differently: ${\bf p} \rightarrow {\bf p} + m {\bf u}$ for microscopic particles and ${\bf p} \rightarrow {\bf p}$ for quasiparticles. The second equation for the collective variables is the London equation for the superfluid velocity, which is curl-free in superfluid $^4$He ($\nabla\times{\bf v}_{\rm s}=0$): $$m{\partial {\bf v}_{(s)}\over \partial t} + \nabla {\delta {\cal E}\over \delta n} =0~. \label{LondonEquation}$$ Together with the kinetic equation for the quasiparticle distribution function $f({\bf p})$, the Eqs.(\[LondonEquation\]) and (\[ContinuityEquation\]) for collective fields ${\bf v}_{\rm s}$ and $n$ give the complete effective theory for the kinetics of quasiparticles (matter) and coherent motion of vacuum (gravitational field) if the energy functional ${\cal E}$ is known. In the limit of low density of quasiparticles, when the interaction between quasiparticles can be neglected, the simplest Ansatz satisfying the Galilean invariance is $${\cal E}=\int d^3r \left( {m\over 2}n{\bf v}_{\rm s}^2 + \epsilon(n) + \sum_{\bf p} \tilde E({\bf p},{\bf r}) f({\bf p},{\bf r})\right)~. \label{Energy}$$ Here $\epsilon(n)$ is the (phenomenological) vacuum energy as a function of the particle density; $\tilde E({\bf p},{\bf r})=E({\bf p},n({\bf r}))+ {\bf p}\cdot{\bf v}_{\rm s}({\bf r})$ is the Doppler shifted quasiparticle energy in the laboratory frame; $E({\bf p},n({\bf r}))$ is the quasiparticle energy measured in the frame comoving with the superfluid vacuum. The Eqs. (\[ContinuityEquation\]) and (\[LondonEquation\]) can be also obtained from the Hamiltonian formalism using Eq.(\[Energy\]) as Hamiltonian and Poisson brackets $$\left\{{\bf v}_{\rm s}({\bf r}_1),n({\bf r}_2)\right\}={1\over m}\nabla \delta({\bf r}_1-{\bf r}_2)~,~ \left\{n({\bf r}_1),n({\bf r}_2)\right\}= \left\{{\bf v}_{\rm s}({\bf r}_1),{\bf v}_{\rm s}({\bf r}_2)\right\}=0~. \label{PoissonBrackets}$$ Note that the Poisson brackets between components of superfluid velocity are zero only for curl-free superfluidity. In a general case it is $$\left\{ v_{{\rm s}i}({\bf r}_1),v_{{\rm s}j}({\bf r}_2)\right\}=-{1\over mn}e_{ijk} (\nabla\times {\bf v}_{\rm s})_k \delta( {\bf r}_1-{\bf r}_2)~. \label{PoissonBracketsVelocity}$$ In this case even at $T=0$, when the quasiparticles are absent, the Hamiltonian description of the hydrodynamics is only possible: There is no Lagrangian, which can be expressed in terms of the hydrodynamic variables ${\bf v}_{\rm s}$ and $n$. The absence of the Lagrangian in many condensed matter systems is one of the consequences of the reduction of the degrees of freedom in effective field theory, as compared with the fully microscopic description. In ferromagnets, for example, the number of the hydrodynamic variables is odd: 3 components of the magnetization vector ${\bf M}$. They thus cannot form the canonical pairs of conjugated variables. As a result one can use either the Hamiltonian description or introduce the effective action with the Wess-Zumino term, which contains an extra coordinate $\tau$: $$S_{\rm WZ}\propto \int d^3x~dt~d\tau ~ {\bf M}\cdot(\partial_t{\bf M}\times\partial_\tau{\bf M})~. \label{WessZumino}$$ Normal component – “matter”. ---------------------------- In a local thermal equilibrium the distribution of quasiparticles is characterized by local temperature $T$ and by local velocity ${\bf v}_{\rm n}$ called the normal component velocity $$f_{\cal T}({\bf p})=\left(\exp { \tilde E({\bf p})- {\bf p} {\bf v}_n\over T} \pm 1\right)^{-1}~, \label{Equilibrium}$$ where the sign + is for the fermionic quasiparticles in Fermi superfluids and the sign - is for the bosonic quasiparticles in Bose superfluids. Since $\tilde E({\bf p})=E({\bf p}) +{\bf p}\cdot {\bf v}_{\rm s}$, the equilibrium distribution is determined by the Galilean invariant quantity ${\bf v}_{\rm n} - {\bf v}_{\rm s}\equiv {\bf w}$, which is the normal component velocity measured in the frame comoving with superfluid vacuum. It is called the counterflow velocity. In the limit when the conterflow velocity ${\bf v}_{\rm n}-{\bf v}_{\rm s}$ is small, the quasiparticle (“matter”) contribution to the particle current is proportional to the counterflow velocity: $$J_{{\rm q} i} =n_{{\rm n}ik}(v_{{\rm n}k}-v_{{\rm s}k})~, \label{EquilibriumCurrent}$$ where the tensor $n_{{\rm n}ik}$ is the so called density of the normal component. In this linear regime the total current can be represented as the sum of the currents of the normal and superfluid components $$J_{i} =n_{{\rm s}ik}v_{{\rm s}k}+n_{{\rm n}ik}v_{{\rm n}k} ~, \label{TotalCurrent2}$$ where tensor $n_{{\rm s}ik}=n\delta_{ik}-n_{{\rm n}ik}$ is the so called density of superfluid component. In the isotropic superfluids, $^4$He and $^3$He-B, the normal component is isotropic tensor, $n_{{\rm n}ik}=n_{{\rm n}}\delta_{ik}$, while in anisotropic superfluid $^3$He-A the normal component density is a uniaxial tensor [@VollhardtWolfle]. At $T=0$ there the quasiparticles are frozen out and one has $n_{{\rm n}ik}=0$ and $n_{{\rm s}ik}=n\delta_{ik}$. Quasiparticle spectrum and effective metric ------------------------------------------- The structure of the quasiparticle spectrum in superfluid $^4$He becomes more and more universal the lower the energy. In the low energy corner the spectrum o f these quasiparticles, phonons, can be obtained in the framework of the effective theory. Note that the effective theory is unable to describe the high-energy part of the spectrum – rotons, which can be determined in a fully microscopic theory only. On the contrary, the spectrum of phonons is linear, $E({\bf p},n)\rightarrow c(n)|{\bf p}|$, and only the “fundamental constant” – the speed of “light” $c(n)$ – depends on the physics of the higher energy hierarchy rank. Phonons represent the quanta of the collective modes of the superfluid vacuum, sound waves, with the speed of sound obeying $c^2(n)=(n/m)(d^2\epsilon/dn^2)$. All other information on the microscopic atomic nature of the liquid is lost. Note that for the curl-free superfluids the sound waves represent the only “gravitational” degree of freedom. The Lagrangian for these “gravitational waves” propagating above the smoothly varying background is obtained by decomposition of the superfluid velocity and density into the smooth and fluctuating parts: ${\bf v}_{\rm s}= {\bf v}_{\rm s~~smooth} +\nabla \alpha$ [@unruh; @vissersonic]. The Lagrangian for the scalar field $\alpha$ is: $${\cal L}= {m\over 2}n \left( (\nabla \alpha)^2- {1\over c^2}\left(\dot\alpha + ({\bf v}_{\rm s}\cdot\nabla)\alpha\right)^2\right)\equiv {1\over 2}\sqrt{-g}g^{\mu\nu}\partial_\mu\alpha \partial_\nu\alpha~. \label{LagrangianSoundWaves}$$ Thus in the low energy corner the Lagrangian for sound waves has an enhanced symmetry – the Lorentzian form, where the effective Riemann metric experienced by the sound wave, the so called acoustic metric, is simulated by the smooth parts of the hydrodynamic fields: $$g^{00}=-{1\over mnc} ~,~ g^{0i}=-{v_{\rm s}^i\over mnc} ~,~ g^{ij}= {c^2\delta^{ij} -v_{\rm s}^i v_{\rm s}^j\over mnc} ~, \label{ContravarianAcousticMetric}$$ $$g_{00}=-{mn\over c}(c^2-{\bf v}_{\rm s}^2) ~,~ g_{0i}=-{mnv_{{\rm s} i}\over c} ~,~ g_{ij}= {mn\over c}\delta_{ij} ~,~ \sqrt{-g}={m^2n^2\over c}~. \label{CovarianAcousticMetric}$$ Here and further ${\bf v}_{\rm s}$ and $n$ mean the smooth parts of the velocity and density fields. The energy spectrum of sound wave quanta, phonons, which represent the “gravitons” in this effective gravity, is determined by $$g^{\mu\nu}p_\mu p_\nu=0~,~~{\rm or} ~~ (\tilde E-{\bf p}\cdot {\bf v}_{\rm s})^2=c^2p^2 ~. \label{PhononEnergySpectrum}$$ Effective quantum field and effective action -------------------------------------------- The effective action Eq.(\[LagrangianSoundWaves\]) for phonons formally obeys the general covariance, this is an example of how the enhanced symmetry arises in the low-energy corner. In addition, in the classical limit of Eq.(\[PhononEnergySpectrum\]) corresponding to geometrical optics (in our case this is geometrical acoustics) the propagation of phonons is invariant under the conformal transformation of metric, $g^{\mu\nu}\rightarrow \Omega^2 g^{\mu\nu}$. This symmetry is lost at the quantum level: the Eq.(\[LagrangianSoundWaves\]) is not invariant under general conformal transformations, however the reduced symmetry is still there: Eq.(\[LagrangianSoundWaves\]) is invariant under scale transformations with $\Omega={\rm Const}$. In superfluid $^3$He-A the other effective fields and new symmetries appear in the low energy corner, including also the effective $SU(2)$ gauge fields and gauge invariance. The symmetry of fermionic Lagrangian induces, after integration over the quasiparticles degrees of freedom, the corresponding symmetry of the effective action for the gauge fields. Moreover, in addition to superfluid velocity field there are appear the other gravitational degrees of freedom with the spin-2 gravitons. However, as distinct from the effective gauge fields, whose effective action is very similar to that in particle physics, the effective gravity cannot reproduce in a full scale the Einstein theory: the effective action for the metric is contaminated by the noncovariant terms, which come from the “transPlanckian” physics [@parallel]. The origin of difficulties with effective gravity in condensed matter is probably the same as the source of the problems related to quantum gravity and cosmological constant. The quantum quasiparticles interact with the classical collective fields ${\bf v}_{\rm s}$ and $n$, and with each other. In Fermi superfluid $^3$He the fermionic quasiparticles interact with many collective fields describing the multicomponent order parameter and with their quanta. That is why one obtains the interacting Fermi and Bose quantum fields, which are in many respect similar to that in particle physics. However, this field theory can be applied to a lowest orders of the perturbation theory only. The higher order diagrams are divergent and nonrenormalizable, which simply means that the effective theory is valid when only the low energy/momentum quasiparticles are involved even in their virtual states. This means that only those terms in the effective action can be derived by integration over the quasiparticle degrees of freedom, whose integral are concentrated solely in the low-energy region. For the other processes one must go beyond the effective field theory and consider the higher levels of description, such as Fermi liquid theory, or further the microscopic level of the underlying liquid with atoms and their interactions. In short, all the terms in effective action come from microscopic “Planck” physics, but only some fraction of them can be derived within the effective field theory. In Bose supefluids the fermionic degrees of freedom are absent, that is why the quantum field theory there is too restrictive, but nevertheless it is useful to consider it since it provides the simplest example of the effective theory. On the other hand the Landau-Khalatnikov scheme is rather universal and is easily extended to superfluids with more complicated order parameter and with fermionic degrees of freedom (see the book [@VollhardtWolfle]). Vacuum energy and cosmological constant {#VacuumEnergyAnd} --------------------------------------- The vacuum energy density $\epsilon(n)$ and the parameters which characterize the quasparticle energy spectrum cannot be determined by the effective theory: they are provided solely by the higher (microscopic) level of description. The microscopic calculations show that at zero pressure the vacuum energy per one atom of the liquid $^4$He is about $\epsilon(n_0)/n_0 \sim -7$K [@Woo]. It is instructive to compare this microscopic result with the estimation of the vacuum energy if we try to obtain it from the effective theory. In effective theory the vacuum energy is given by the zero point motion of phonons $$\epsilon_{\rm eff} = (1/2)\sum_{E({\bf p})<\Theta} cp ={1\over 16\pi^2} {\Theta^4\over \hbar ^3 c^3} ={1\over 16\pi^2}\sqrt{-g} \left(g^{\mu\nu}\Theta_{\mu} \Theta_{\nu}\right)^2 ~. \label{VacuumEnergy}$$ Here $c$ is the speed of sound; $\Theta \sim \hbar c/a$ is the Debye characteristic temperature with $a$ being the interatomic space, $\Theta$ plays the part of the “Planck” cutoff energy scale; $\Theta_{\mu}=(-\Theta,0,0,0)$. We wrote the Eq.(\[VacuumEnergy\]) in the form which is different from the conventional cosmological term $\Lambda \sqrt{-g}$. This is to show that both forms and the other possible forms too have the similar drawbacks. The Eq.(\[VacuumEnergy\]) is conformal invariant due to conformal invariance experienced by the quasiparticle energy spectrum in Eq.(\[PhononEnergySpectrum\]) (actually, since this term does not depend on derivatives, the conformal invariance is equivalent to invariance under multiplication of $g_{\mu\nu}$ by constant factor). However, in Eq.(\[VacuumEnergy\]) the general covariance is violated by the cutoff. On the contrary, the conventional cosmological term $\Lambda \sqrt{-g}$ obeys the general covariance, but it is not invariant under transformation $g_{\mu\nu}\rightarrow \Omega^2g_{\mu\nu}$ with constant $\Omega$. Thus both forms of the vacuum energy violate one or the other symmetry of the low-energy effective Lagrangian Eq.(\[LagrangianSoundWaves\]) for phonons, which means that the vacuum energy cannot be determined exclusively within the low-energy domain. Now on the magnitude of the vacuum energy. The Eq.(\[VacuumEnergy\]) gives $\epsilon_{\rm eff}(n_0)/n_0 \sim 10^{-2}\Theta \sim 10^{-1}$K. The magnitude of the energy is much smaller than the result obtained in the microscopic theory, but what is more important the energy has an opposite sign. This means again that the effective theory must be used with great caution, when one calculates those quantities, which crucially (non-logarithmically) depend on the “Planck” energy scale. For them the higher level “transPlanckian” physics must be used only. In a given case the many-body wave function of atoms of the underlying quantum liquid has been calculated to obtain the vacuum energy [@Woo]. The quantum fluctuations of the phonon degrees of freedom in Eq.(\[VacuumEnergy\]) are already contained in this microscopic wave function. To add the energy of this zero point motion of the effective field to the microscopically calculated energy $\epsilon(n_0)$ would be the double counting. Thus the proper regularization of the vacuum energy in the effective field theory must by equating it to exact zero. This conjecture is confirmed by consideration of the equilibrium conditions for the liquid. The equilibrium condition for the superfluid vacuum is $(d\epsilon/d n)_{n_0}=0$. Close to the equilibrium state one has $$\epsilon(n)=\epsilon(n_0) +{1\over 2} {m c^2\over n_0} (n-n_0)^2~. \label{VacuumCloseToEquil}$$ From this equation it follows that the variation of the vacuum energy over the metric determinant must be zero in equilibrium: $d\epsilon/dg|_{n_0}=(d\epsilon/d n)_{n_0}/(d g/d n)_{n_0}= 0$. This apparently shows that the vacuum energy can be neither of the form of Eq.(\[VacuumEnergy\]) nor in the form $\Lambda\sqrt{-g}$. The metric dependence of the vacuum energy consistent with the equilibrium condition and Eq.(\[VacuumCloseToEquil\]) could be only of the type $A+B(g-g_0)^2$, so that the cosmological term in Einstein equation would be $\propto (g-g_0)g_{\mu\nu}$. This means that in equilibrium, i.e. at $g=g_0$, the cosmological term is zero and thus the equilibrium vacuum is not gravitating. In relativistic theories such dependence of the Lagrangian on $g$ can occur in the models where the determinant of the metric is the variable which is not transformed under coordinate transformations, i.e. only the invariance under coordinate transformations with unit determinant represents the fundamental symmetry. This probably has some relation to the problem of the cosmological constant in Einstein theory of gravity, where the estimation in Eq.(\[VacuumEnergy\]) with $c$ being the speed of light and $\Theta=E_{\rm Planck}$ gives the cosmological term by 100 orders of magnitude higher than its upper experimental limit. The gravity is the low-frequency, and actually the classical output of all the quantum degrees of freedom of the “Planck condensed matter”. So one should not quantize the gravity again, i.e. one should not use the low energy quantization for construction of the Feynman diagrams technique with diagrams containing the integration over high momenta. In particular the effective field theory is not appropriate for the calculation of the vacuum energy and thus of the cosmological constant. Moreover, one can argue that, whatever the real “microscopic” energy of the vacuum is, the energy of the equilibrium vacuum is not gravitating: The diverging energy of quantum fluctuations of the effective fields and thus the cosmological term must be regularized to zero as we discussed above, since these fluctuations are already contained in the “microscopic wave function” of the vacuum. This however does not exclude the Casimir effect, which appears if the vacuum is not homogeneous. The smooth deviations from the homogeneous equilibrium vacuum are within the low-energy domain: they can be successfully described by the effective field theory, and their energy can gravitate. Einstein action and higher derivative terms {#EinsteinActionAnd} ------------------------------------------- In principle, there are the nonhydrodynamic terms in the effective action, which are not written in Eq.(\[Energy\]) since they contain space and time derivatives of the hydrodynamics variable, $n$ and ${\bf v}_{\rm s}$, and thus are relatively small. Only part of them can be obtained using the effective theory. As in the case of Sakharov effective gravity [@Sakharov], the standard integration over the massless scalar field $\alpha$ propagating in inhomogeneous $n$ and ${\bf v}_{\rm s}$ fields, which provide the effective metric, gives the curvature term in Einstein action. It can also be written in two ways. The form which respects the general covariance of the phononic Lagrangian for $\alpha$ field in Eq.(\[LagrangianSoundWaves\]) is: $${\cal L}_{\rm Einstein}=-{1 \over 16\pi G } \sqrt{-g} R ~~, \label{EinsteinAction}$$ This form does not obey the invariance under multiplication of $g_{\mu\nu}$ by constant factor, which shows its dependence on the “Planck” physics. The gravitational Newton constant $G$ is expressed in terms of the “Planck” cutoff: $G^{-1}\sim \Theta^2$. Another form, which explicitly contains the “Planck” cutoff, $${\cal L}_{\rm Einstein}=-{1 \over 16\pi } \sqrt{-g} R g^{\mu\nu}\Theta_{\mu}\Theta_{\nu} ~~, \label{EinsteinActionModified}$$ is equally bad: the action is invariant under the scale transformation of the metric, but the general covariance is violated. Such incompatibility of different low-energy symmetries is the hallmark of the effective theories. To give an impression on the relative magnitude of the Einstein action let us express the Ricci scalar in terms of the superfluid velocity field only $$\sqrt{-g} R ={1\over c^3}\left(2\partial_t \nabla\cdot {\bf v}_{\rm s} + \nabla^2(v_{\rm s}^2)\right) ~. \label{RicciScalar}$$ In superfluids the Einstein action is small compared to the dominating kinetic energy term $mn{\bf v}_{\rm s}^2/2$ in Eq.(\[Energy\]) by factor $a^2/l^2$, where $a$ is again the atomic (“Planck”) length scale and $l$ is the characteristic macroscopic length at which the velocity field changes. That is why it can be neglected in the hydrodynamic limit, $a/l \rightarrow 0$. Moreover, there are many terms of the same order in effective actions which do not display the general covariance, such as $(\nabla\cdot {\bf v}_{\rm s})^2$. They are provided by microscopic physics, and there is no rule in superfluids according to which these noncovariant terms must be smaller than the Eq.(\[EinsteinAction\]). But in principle, if the gravity field as collective field arises from the other degrees of freedom, different from the condensate motion, the Einstein action can be dominating. There are the higher order derivative terms, which are quadratic in the Riemann tensor, such as $$\sqrt{-g}R^2 ~\ln \left( {g^{\mu\nu}\Theta_{\mu}\Theta_{\nu}\over R}\right)~. \label{SquareRicciScalar}$$ They only logarithmically depend on the cut-off and thus their calculation in the framework of the effective theory is possible. Because of the logarithmic divergence (they are of the relative order $(a/l)^4~\ln (l/a)$) these terms dominate over the noncovariant terms of order $(a/l)^4$, which are obtained in fully microscopic calculations. Being determined essentially by the phononic Lagrangian in Eq.(\[LagrangianSoundWaves\]), these terms respect (with logarithmic accuracy) all the symmetries of this Lagrangian including the general covariance and the invariance under rescaling the metric. That is why they are the most appropriate terms for the self-consistent effective theory of gravity. The logarithmic terms also appear in the effective action for the effective gauge fields, which take place in superfluid $^3$He-A [@LammiTalk]. These terms in superfluid $^3$He-A have been obtained first in microscopic calculations, however it appeared that their physics can be completely determined by the low energy tail and thus they can be calculated using the effective theory. This is well known in particle physics as running coupling constants and zero charge effect. Unfortunately in effective gravity of superfluids these logarithmic terms are small compared with the main terms – the vacuum energy and the kinetic energy of the vacuum flow. This means that the superfluid liquid is not the best condensed matter for simulation of Einstein gravity. In $^3$He-A there are other components of the order parameter, which also give rise to the effective gravity, but superfluidity of $^3$He-A remains to be an obstacle. One must try to construct the non-superfluid condensed matter system which belongs to the same universality class as $^3$He-A, and thus contains the effective Einstein gravity as emergent phenomenon, which is not contaminated by the superfluidity. “Relativistic” energy-momentum tensor for “matter” moving in “gravitational” superfluid background in two fluid hydrodynamics ============================================================================================================================= Kinetic equation for quasiparticles (matter) -------------------------------------------- The distribution function $f$ of the quasiparticles is determined by the kinetic equation: $$\dot f - {\partial \tilde E\over \partial {\bf r}} \cdot {\partial f\over \partial {\bf p}}+ {\partial \tilde E\over \partial {\bf p}} \cdot {\partial f\over \partial {\bf r}}={\cal J}_{coll}~. \label{KineticEq}$$ The collision integral conserves the momentum and the energy of quasiparticles, i.e. $$\sum_{\bf p} {\bf p} {\cal J}_{coll}=\sum_{\bf p} \tilde E ({\bf p}) {\cal J}_{coll}=\sum_{\bf p} E({\bf p}) {\cal J}_{coll}=0~, \label{ConservationCollision}$$ but not necessarily the number of quasiparticle: as a rule the quasiparticle number is not conserved in superfluids. Momentum exchange between superfluid vacuum and quasiparticles -------------------------------------------------------------- From the Eq.(\[ConservationCollision\]) and from the two equations for the superfluid vacuum, Eqs.(\[ContinuityEquation\],\[LondonEquation\]), one obtains the time evolution of the momentum density for each of two subsystems: the superfluid background (vacuum) and quasiparticles (matter). The momentum evolution of the superfluid vacuum is $$m\partial_t (n {\bf v}_{{\rm s}}) =- m\nabla_i(J_i{\bf v}_{\rm s}) -n \nabla\left( {\partial \epsilon\over \partial n} +\sum_{\bf p} f {\partial E \over \partial n} \right) + P_i \nabla v_{{\rm s}i}~. \label{SuperfluidMomentumEq}$$ where ${\bf P}=m{\bf J}_{\rm q}$ is the momentum of lquid carried by quasiparticles (see Eq.(\[TotalCurrent2\])), while the evolution of the momentum density of quasiparticles: $$\partial_t {\bf P} =\sum_{\bf p} {\bf p} \partial_t f = - \nabla_i(v_{{\rm s}i}{\bf P}) -\nabla_i \left( \sum_{\bf p} {\bf p}f {\partial E\over \partial p_i}\right) -\sum_{\bf p}f \nabla E - P_i \nabla v_{{\rm s}i}~. \label{QuasiparticleMomentumEq}$$ Though the momentum of each subsystem is not conserved because of the interaction with the other subsystem, the total momentum density of the system, superfluid vacuum + quasiparticles, is conserved: $$m\partial_t J_i = \partial_t (mn v_{{\rm s}i} +P_i)= -\nabla_i \Pi_{ik}~, \label{TotalMomentumEq}$$ with the stress tensor $$\Pi_{ik} = m J_i v_{{\rm s}k} + v_{{\rm s}i} P_k+ \sum_{\bf p} p_k f {\partial E\over \partial p_i} + \delta_{ik} \left(n\left({\partial \epsilon\over \partial n}+ \sum_{\bf p} f{\partial E\over \partial n}\right) - \epsilon\right) ~. \label{StressTensor}$$ Covariance vs conservation. --------------------------- The same happens with the energy. Energy and momentum can be exchanged between the two subsystems of quasiparticles and superfluid vacuum in a way similar to the exchange of energy and momentum between matter and the gravitational field. In the low energy limit, when the quasiparticles are “relativistic”, this exchange must be described in the general relativistic covariant form. The Eq.(\[QuasiparticleMomentumEq\]) for the momentum density of quasiparticles as well as the corresponding equation for the quasiparticle energy density can be represented as $$T^\mu{}_{\nu;\mu}=0 ~~,~~{\rm or}~~ {1\over \sqrt{-g}}\partial_\mu \left(T^\mu{}_\nu \sqrt{-g}\right) - {1\over 2}T^{\alpha\beta} \partial_\nu g_{\alpha\beta}= 0 \,. \label{CovariantConservation}$$ This result does not depend on the dynamics of the superfluid condensate (gravity field), which is not “relativistic”. The Eq.(\[CovariantConservation\]) follows solely from the “relativistic” spectrum of quasiparticles. The Eq.(\[CovariantConservation\]) does not represent any conservation in a strict sense, since the covariant derivative is not a total derivative. The extra term, which is not the total derivative, describes the force acting on quasiparticles (matter) from the superfluid condensate (an effective gravitational field). Since the dynamics of the superfluid background is not covariant, it is impossible to find such total energy momentum tensor, $T^\mu{}_\nu({\rm total}) = T^\mu{}_\nu({\rm quasiparticles}) +T^\mu{}_\nu({\rm background})$, which could have a covariant form and simultaneously satisfy the real conservation law $\partial_\mu T^\mu{}_\nu({\rm total})=0$. The total stress tensor in Eq.(\[StressTensor\]) is evidently noncovariant. But this is impossible even in the fully covariant Einstein gravity, where one has an energy momentum pseudotensor for the gravitational background. Probably this is an indication that the Einstein gravity is really an effective theory. As we mentioned above, effective theories in condensed matter are full of such contradictions related to incompatible symmetries. In a given case the general covariance is incompatible with the conservation law; in cases of the vacuum energy (Sec.\[VacuumEnergyAnd\]) and the Einstein action (Sec.\[EinsteinActionAnd\]) the general covariance is incompatible with the scale invariance; in the case of an axial anomaly, which is also reproduced in condensed matter (see e.g.[@LammiTalk]), the conservation of the baryonic charge is incompatible with quantum mechanics; the action of the Wess-Zumino type, which cannot be written in 3+1 dimension in the covariant form (as we discussed at the end of Sec.\[DynamicsSuperfluidVacuum\], Eq.(\[WessZumino\])), is almost typical phenomenon in various condensed matter systems; the momentum density determined as variation of the hydrodynamic energy over ${\bf v}_{\rm s}$ does not coincide with the canonical momentum in many condensed matter systems; etc., there are many other examples of apparent inconsistencies in the effective theories of condensed matter. All such paradoxes arise due to reduction of the degrees of freedom in effective theory, and they disappear completely (together with some symmetries of the low-energy physics) on the fundamental level, i.e. in a fully microscopic description, where all degrees of freedom are taken into account. Energy-momentum tensor for “matter”. ------------------------------------ Let us specify the tensor $T^\mu{}_\nu$ which enters Eq.(\[CovariantConservation\]) for the simplest case, when the gravity is simulated by the superflow only, i.e. we neglect the space-time dependence of the density $n$ and of the speed of sound $c$. Then the constant factor $mnc$ can be removed from the metric in Eqs.(\[ContravarianAcousticMetric\]-\[CovarianAcousticMetric\]) and the effective metric is simplified: $$g^{00}=-1 ~,~ g^{0i}=-v_{\rm s}^i ~,~ g^{ij}= c^2\delta^{ij} -v_{\rm s}^i v_{\rm s}^j ~, \label{ContravarianAcousticMetricReduced}$$ $$g_{00}=-\left(1-{{\bf v}_{\rm s}^2\over c^2}\right) ~,~ g_{0i}=-{v_{{\rm s} i}\over c^2} ~,~ g_{ij}= {1\over c^2}\delta_{ij} ~,~ \sqrt{-g}={1\over c^3}~. \label{CovarianAcousticMetricReduced}$$ Then the energy-momentum tensor of quasiparticles can be represented as [@FischerVolovik] $$\sqrt{-g} T^\mu{}_\nu=\sum_{\bf p} f v_g^\mu p_\nu\,,\qquad v_g^\mu v_{g\mu} = -1 +\frac1{c^2} \frac{\partial E }{\partial p_i} \frac{\partial E }{\partial p_i}\,,$$ where $p_0= - \tilde E$, $p^0= E$; the group four velocity is defined as $$\begin{aligned} v_g^i& =& \frac{\partial \tilde E}{\partial p_i}\,,\quad v_g^0=1\,,\quad v_{gi}= \frac1{c^2} \frac{\partial E }{\partial p_i}\,, \quad v_{g0}=-\left( 1+\frac{{v_{\rm s}^i }}{c^2} \frac{\partial E }{\partial p_i}\right)\,.\end{aligned}$$ Space-time indices are throughout assumed to be raised and lowered by the metric in Eqs.(\[ContravarianAcousticMetricReduced\]-\[CovarianAcousticMetricReduced\] ). The group four velocity is null in the relativistic domain of the spectrum only: $v_g^\mu v_{g\mu}=0$ if $E =cp$. The relevant components of the energy-momentum tensor are: $$\begin{aligned} \sqrt{-g} T^0{}_i & = & \sum_{\bf p} f {p}_i = P_i\qquad \mbox{\sf momentum density in either frame}, \nonumber\\ \quad -\sqrt{-g} T^0{}_0& = & \sum_{\bf p} f \tilde E\qquad \mbox{\sf energy density in laboratory frame}, \nonumber\\ \sqrt{-g} T^k{}_i & = & \sum_{\bf p} f p_i v_g^k \qquad \mbox{\sf momentum flux in laboratory frame}, \nonumber\\ -\sqrt{-g} T^i{}_0 & = & -\sum_{\bf p} f \tilde E\frac{\partial E}{\partial p_i} = \sum_{\bf p} f \tilde E v_g^i\qquad \mbox{\sf energy flux in laboratory frame}, \nonumber\\ \sqrt{-g} T^{00} & = & \sum_{\bf p} f p^0 = \sum_{\bf p} f E \qquad \mbox{\sf energy density in comoving frame}.\end{aligned}$$ With this definition of the momentum-energy tensor the covariant conservation law in Eq.(\[CovariantConservation\]) acquires the form: $$(\sqrt{-g}T^\mu{}_\nu)_{,\mu}=\sum_{\bf p} f \partial_\nu \tilde E= P_i \partial_\nu v_{\rm s}^i + \sum_{\bf p} f |{\bf p}|\partial_\nu c \,.$$ The right-hand side represents “gravitational” forces acting on the “matter” from the superfluid vacuum. Local thermodynamic equilibrium. -------------------------------- Local thermodynamic equilibrium is characterized by the local temperature $T$ and local normal component velocity ${\bf v}_{\rm n}$ in Eq.(\[Equilibrium\]). In local thermodynamic equilibrium the components of energy-momentum for the quasiparticle system (matter) are determined by the generic thermodynamic potential (the pressure), which has the form $$\Omega=\mp T\frac1{(2\pi\hbar)^3} \sum_{s} \int d^3p~{\rm ln}(1\mp f)\,, \label{Pressure}$$ with the upper sign for fermions and lower sign for bosons. For phonons one has $$\Omega= \frac{\pi^2}{30\hbar^3} T_{\rm eff}^4\,,\qquad T_{\rm eff} = \frac{T}{\sqrt{1-w^2}}~\,, \label{EffectiveT}$$ where the renormalized effective temperature $T_{\rm eff}$ absorbs all the dependence on two velocities of liquid. The components of the energy momentum tensor are given as $$T^{\mu\nu} = (\varepsilon + \Omega )u^\mu u^\nu+\Omega g^{\mu\nu}\,,\qquad \varepsilon=-\Omega +T{\partial \Omega \over\partial T}=3\Omega\,,\qquad T^\mu{}_\mu=0~. \label{QuasipStressTensorRel2}$$ where the four velocity of the “matter”, $u^\alpha$ and $u_\alpha=g_{\alpha\beta}u^{\beta}$, which satisfies the normalization equation $u_\alpha u^\alpha=-1$, is expressed in terms of superfluid and normal component velocities as $$u^0={1\over \sqrt{ 1 - w^2}}\,,\qquad u^i={v_{(n)}^i\over \sqrt{ 1 - w^2}}\,,\qquad u_i= { w_i \over \sqrt{{ 1 - w^2}}}\,,\qquad u_0=-{1+{{\bf w} \cdot{\bf v}_{{\rm s}}} \over \sqrt{ { 1 - w^2}}}\,. \label{4Velocity}$$ Global thermodynamic equilibrium. Tolman temperature. ----------------------------------------------------- The distribution of quasiparticles in local equilibrium in Eq.(\[Equilibrium\]) can be expressed via the temperature four-vector $\beta^\mu $ and thus via the effective temperature $T_{\rm eff}$: $$f_{\cal T} = {1\over 1+\exp[-\beta^\mu p_\mu]}\,,\qquad \beta^\mu ={u^\mu\over T_{\rm eff}} =\left({1\over T}, {{\bf v}_n \over T }\right)\,,\qquad \beta^\mu\beta_\mu=-T_{\rm eff}^{-2}~. \label{4Temperature}$$ For the relativistic system, the true equilibrium with vanishing entropy production is established if $\beta^\mu$ is a timelike Killing vector: $$\beta_{\mu;\nu}+ \beta_{\nu;\mu}=0~,\;\;{\rm or}\qquad \beta^\alpha\partial_\alpha g_{\mu\nu}+ (g_{\mu\alpha}\partial_\nu +g_{\nu\alpha}\partial_\mu)\beta^\alpha=0 ~. \label{EquilibriumConditions}$$ For a time independent, space dependent situation the condition $0=\beta_{0;0}=\beta^i\partial_i g_{00}$ gives $\beta^i=0$, while the other conditions are satisfied when $\beta^0={\rm constant}$. Hence true equilibrium requires that ${\bf v}_{\rm n}=0$ in the frame where the superfluid velocity field is time independent (i.e. in the frame where $\partial_t{\bf v}_{\rm s}=0$), and $T={\rm constant}$. These are just the global equilibrium conditions in superfluids, at which no dissipation occurs. From the equilibrium conditions $T={\rm constant}$ and ${\bf v}_{\rm n}=0$ it follows that under the global equilibrium the effective temperature in Eqs.(\[EffectiveT\]) is space dependent according to $$T_{\rm eff}={T\over \sqrt{ 1 - v_{\rm s}^2}}={T\over \sqrt{-g_{00}}}\,. \label{TolmanLaw}$$ According to Eq.(\[4Temperature\]) the effective temperature $T_{\rm eff}$ corresponds to the “covariant relativistic” temperature in general relativity. It is an apparent temperature as measured by the local observer, who “lives” in superfluid vacuum and uses sound for communication as we use the light signals. The Eq.(\[TolmanLaw\]) corresponds to Tolman’s law in general relativity. Note that in condensed matter the Tolman temperature is the real temperature $T$ of the liquid. Horizons, ergoregions, degenerate metric, vacuum instability and all that. ========================================================================== Landau critical velocity, event horizon and ergoregion ------------------------------------------------------ If the superfluid velocity exceeds the Landau critical value $$v_L={\rm min}{E({\bf p})\over p} \label{LandauVelocity1}$$ the energy $\tilde E({\bf p})$ of some excitations, as measured in the laboratory frame, becomes negative. This allows for excitations to be nucleated from the vacuum. For a superfluid velocity field which is time-independent in the laboratory frame, the surface $v_{\rm s}({\bf r})=v_L$, which bounds the region where quasiparticles can have negative energy, the ergoregion, is called the ergosurface. The behavior of the system depends crucially on the dispersion of the spectrum at higher energy. There are two possible cases. The spectrum bents upwards at high energy, i.e. $ E({\bf p})=cp +\gamma p^3$ with $\gamma >0$. Such dispersion is realized for the fermionic quasiparticles in $^3$He-A. They are “relativistic” in the low energy corner but become “superluminal” at higher energy[@grishated]. In this case the Landau critical velocity coincides with the “speed of light”, $v_L=c$, so that the ergosurface is determined by $v_{\rm s}({\bf r})=c$. In the Lorentz invariant limit of the energy much below the “Planck” scale, i.e. at $p^2 \ll \gamma/c$, this corresponds to the ergosurface at $g_{00}({\bf r})=0$, which is just the definition of the ergosurface in gravity. In case of radial flow of the superfluid vacuum towards the origin, the ergosurface also represents the horizon in the Lorentz invariant limit, and the region inside the horizon simulates a black hole for low energy phonons. Strictly speaking this is not a true horizon for phonons: Due to the nonlinear dispersion, their group velocity $v_g=dE/d p =c+3\gamma p^2>c$, and thus the high energy quasiparticles are allowed to leave the black hole region. It is, hence, a horizon only for quasiparticles living exclusively in the very low energy corner: they are not aware of the possibility of “superluminal” motion. Nevertheless, the mere possibility to exchange the information across the horizon allows us to construct the thermal state on both sides of the horizon (see Sec.\[ModTolmanLaw\] below) and to investigate its thermodynamics, including the entropy related to the horizon [@FischerVolovik]. In superfluid $^4$He the negative dispersion is realized, with the group velocity $v_g=dE/d p <c$. In such superfluids the “relativistic” ergosurface $v_{\rm s}({\bf r})=c$ does not coincide with the true ergosurface, which is determined by $v_{\rm s}({\bf r})=v_L< c$. In superfluid $^4$He, the Landau velocity is related to the roton part of the spectrum, and is about four times less than $c$. In case of radial flow inward, the ergosphere occurs at $v_{\rm s}(r)=v_L<c$, while the inner surface $v_{\rm s}(r)=c$ still marks the horizon. This is in contrast to relativistically invariant systems, for which ergosurface and horizon coincide for purely radial gravitational field. The surface $v_{\rm s}(r)=c$ stays a horizon even for excitations with very high momenta up to some critical value, at which the group velocity of quasiparticle again approaches $c$. Painlevé-Gullstrand metric in effective gravity in superfluids. --------------------------------------------------------------- Let us consider the spherically symmetric radial flow of the superfluid vacuum, which is time-independent in the laboratory frame. Then dynamics of the phonon, propagating in this velocity field, is given by the line element provided by the effective metric in Eq.(\[CovarianAcousticMetricReduced\]): $$ds^2= -\left(1-{v_{\rm s}^2(r)\over c^2}\right)dt^2 + 2{v_{\rm s}(r)\over c^2}drdt + {1\over c^2} (dr^2 + r^2 d\Omega^2 )\,. \label{PaileveInterval}$$ This equation corresponds to the Painlevé-Gullstrand line elements. It describes a black hole horizon if the superflow is inward (see refs.[@unruh; @vissersonic], on the pedagogical review of Panlevé-Gullstrand metric see [@Martel]). If $v_{\rm s}(r)=-c(r_S/r)^{1/2}$ the flow simulates the black hole in general relativity. For the outward superflow with, say, $v_{\rm s}(r)=+c(r_S/r)^{1/2}$ the white hole is reproduced. For the general radial dependence of the superfluid velocity, the Schwarzschild radius $r_S$ is determined as $v_{\rm s}(r_S)=\pm c$; the “surface gravity” at the Schwarzschild radius is $\kappa_S =(1/2)dv^2_{\rm s}/dr|_{r_S}$; and the Hawking temperature $T_{\rm H}=\hbar\kappa_S/2\pi$. Vacuum resistance to formation of horizon. ------------------------------------------ It is not easy to create the flow with the horizon in the Bose liquid because of the hydrodynamic instability which takes place behind the horizon (see [@Liberati]). From Eqs.(\[ContinuityEquation\]) and (\[LondonEquation\]) of superfluid hydrodynamics at $T=0$ (which correspond to conventional hydrodynamics of ideal curl-free liquid) it follows that for stationary motion of the liquid one has the relation between $n$ and $v_{\rm s}$ along the streamline [@FluidMechanics]: $${\partial(nv_{\rm s})\over \partial v_{\rm s}}=n\left(1-{v_{\rm s}^2\over c^2}\right) \,. \label{nvVSv}$$ The current $J= nv_{\rm s}$ has a maximal value just at the horizon and thus it must decrease behind the horizon, where $1-(v_{\rm s}^2/c^2)$ is negative. This is, however, impossible in the radial flow since, according to the continuity equation (\[ContinuityEquation\]), one has $nv_{\rm s}=Const/r^2$ and thus the current must monotonically increase across the horizon. This marks the hydrodynamic instability behind the horizon and shows that it is impossible to construct the time-independent flow with the horizon without the fine-tuning of an external force acting on the liquid [@Liberati]. Thus the liquid itself resists to the formation of the horizon. Would the quantum vacuum always resist to formation of the horizon? Fortunately, not. In the considered case of superfluid $^4$He, the same “speed of light” $c$, which describes the quasipartilces (acoustic waves) and thus determines the value of the superfluid velocity at horizon, also enters the hydrodynamic equations that establish the flow pattern of the “black hole”. In $^3$He-A these two speeds are well separated. The “speed of light” $c$ for quasiparticles, which determines the velocity of liquid flow at the horizon, is much less than the speed of sound, which determines the hydrodynamic instabilities of the liquid. That is why there are no severe hydrodynamic constraints on the flow pattern, the hydrodynamic instability is never reached and the surface gravity at such horizons is always finite. However, even in such superfluids another instability can develop due to the presence of a horizon [@KopninVolovik1998]. Usually the “speed of light” $c$ for “relativistic ” quasiparticles coincides with the critical velocity, at which the superfluid state of the liquid becomes unstable towards the normal state of the liquid. When the superfluid velocity with respect to the normal component or to the container walls exceeds $c$, the slope $\partial J/ \partial v_{\rm s}$ becomes negative the superflow is locally unstable. Such superfluid instability, however, can be avoided if the container walls are properly isolated [@SimulationPainleve]. Then the reference frame imposed by the container walls is lost and the “inner” observer living in the superfluid does not know that the superfluid exceeded the Landau velocity and thus the treshold of instability. Formally this means that the superfluid instability is regulated not by the superfluid velocity field $v_{\rm s}$ (as in the case of the hydrodynamic instability discussed above) , but by the velocity ${\bf w}={\bf v}_{\rm n}-{\bf v}_{\rm s}$ of the counterflow between the normal and superfluid subsystems. A stable superfluid vacuum can be determined as the limit $T\rightarrow 0$ at fixed “subluminal” counterflow velocity $w<c$, even if the superfluid velocity itself is “superluminal”. This can be applied also to quasiequilibrium vacuum state across the horizon, which is locally the vacuum state as observed by comoving “inner” observer. The superfluid motion in this state is locally stable, though slowly decelerates due to the quantum friction caused by Hawking radiation and other processes related to the horizon and ergoregion [@grishated]. Modified Tolman’s law across the horizon. {#ModTolmanLaw} ----------------------------------------- The realization of the quasiequilibrium state across the horizon at nonzero $T$ can be found in Ref.[@FischerVolovik] for 1+1 case. In this state the superfluid velocity is “superluminal” behind the horizon, $v_{\rm s}>c$, but the counterflow is everywhere “subluminal”: the counterflow velocity $w$ reaches maximim value $w=c$ at the horizon with $w<c$ both outside and inside the horizon. The local equilibrium with the effective temperature $T_{\rm eff}$ in Eq.(\[EffectiveT\]) is thus determined on both sides of the horizon. It is interesting that one has the modified form of the Tolman’s law, which is valid on both sides of the horizon: $$T_{\rm eff}={T_\infty\over \sqrt{ \big|1 - {v_{\rm s}^2\over c^2}\big|}}={T_\infty\over \sqrt{|g_{00}|}}\,. \label{ModifiedTolmanLaw}$$ Here $T_\infty$ is the temperature at infinity. The effective temperature $T_{\rm eff}$, which determines the local “relativistic” thermodynamics, becomes infinite at the horizon, with the cutoff determined by the nonlinear dispersion of the quasiparticle spectrum at high energy, $\gamma>0$. The real temperature $T$ of the liquid is continuous across the horizon: $$T= T_\infty~~~~{\rm at}~~v_{\rm s}^2< c^2~~~~~,~~~~ T= T_\infty {c\over |v_{\rm s}|}~~~~{\rm at}~~v_{\rm s}^2> c^2 \,. \label{ModifiedTolmanLaw2}$$ One more vacuum instability: Painlevé-Gullstrand vs Schwarzschild metric in effective gravity. ---------------------------------------------------------------------------------------------- In the effective theory of gravity, which occurs in condensed matter systems, the primary quantity is the contravariant metric tensor $g^{\mu\nu}$ describing the energy spectrum. Due to this the two seemingly equivalent representations of the black hole metric, in terms of either the Schwarzschild or the Painlevé-Gullstrand line elements, are in fact not equivalent in terms of the required stability of the underlying superfluid vacuum. An “equivalent” representation of the black or white hole metric is given by the Schwarzschild line element, which in terms of the same superfluid velocity reads $$ds^2=-\left(1- v_{\rm s}^2/c^2\right)d\tilde t^2+{dr^2\over c^2- v_{\rm s}^2} +(r^2/c^2)\,d\Omega^2\,. \label{Schwarzschild}$$ The Eqs.(\[Schwarzschild\]) and (\[PaileveInterval\]) are related by the coordinate transformation. Let us for simplicity consider the abstract flow with the velocity exactly simulating the Schwarzschild metric, i.e. $v^2_{\rm s}(r)=r_S/r$ and we put $c=1$. Then the coordinate transformation is $$\tilde t(r,t)=t + \left({2\over v_{\rm s}(r)} + {\rm ln}~ {1- v_{\rm s}(r)\over 1+v_s(r)}\right) ~,~d\tilde t=dt +{v_{\rm s}\over 1-v_{\rm s}^2}dr. \label{Transformation}$$ What is the difference between the Schwarzschild and Painlevé-Gullstrand space-times in the effective gravity? The Painlevé-Gullstrand metric is determined in the “absolute” Newton’s space-time $(t,{\bf r})$ of the laboratory frame, i.e. as is measured by the external experimentalist, who lives in the real world of the laboratory and investigates the dynamics of quasipartcles using the physical laws obeying the Galilean invariance of the absolute space-time. The effective Painlevé-Gullstrand metric, which describes the quasiparticle dynamics in the inhomogeneous liquid, originates from the quasiparticle spectrum $$E =v_{\rm s}(r)p_r \pm cp~, \label{SpectrumInSuperflow2}$$ or $$(E -v_{\rm s}(r)p_r)^2 = c^2p^2~, \label{SpectrumInSuperflow1}$$ which determines the contravariant components of the metric. Thus the energy spectrum in the low-energy corner is the primary quantity, which determines the effective metric for the low-energy quasiparticles. The time $\tilde t$ in the Schwarzschild line element is the time as measured by the “inner” observer at “infinity” (i.e. far from the hole). The “inner” means that this observer “ lives” in the superfluid background and uses “relativistic” massless quasiparticles (phonons in $^4$He or “relativistic” fermionic quasiparticles in $^3$He-A) as a light for communication and for sinchronization clocks. The inner observer at some point $R\gg 1$ sends quasiparticles pulse at the moment $t_1$ which arrives at point $r$ at $t=t_1+\int_{r}^Rdr/|v_-|$ of the absolute (laboratory) time, where $v_+$ and $v_-$ are absolute (laboratory) velocities of radially propagating quasiparticles, moving outward and inward respectively $$v_\pm ={dr\over dt}={dE\over dp_r}=\pm 1 +v_{\rm s}~. \label{RadialVelocity}$$ Since from the point of view of the inner observer the speed of light (i.e. the speed of quasiparticles) is invariant quantity and does not depend on direction of propagation, for him the moment of arrival of pulse to $r$ is not $t$ but $\tilde t =(t_1+t_2)/2$, where $t_2$ is the time when the pulse reflected from $r$ returns to observer at $R$. Since $t_2-t_1=\int_{r}^Rdr/|v_-|+\int_{r}^Rdr/|v_+|$, one obtains for the time measured by inner observer as $$\tilde t(r,t)={t_1+t_2\over 2}=t + {1\over 2} \left(\int^R_r{dr\over v_+}+ \int^R_r{dr\over v_-}\right)=t + \left({2\over v_{\rm s}(r)} + {\rm ln}~ {1- v_s(r)\over 1+v_{\rm s}(r)}\right)- \left({2\over v_{\rm s}(R)} + {\rm ln}~ {1- v_s(R)\over 1+v_{\rm s}(R)}\right) ~, \label{InnerTime}$$ which is just the Eq.(\[Transformation\]) up to a constant shift. In the complete absolute physical space-time of the laboratory the external observer can detect quasiparticles radially propagating into (but not out of) the black hole or out of (but not into) the white hole. The energy spectrum of the quasiparticles remains to be well determined both outside and inside the horizon. Quasiparticles cross the black hole horizon with the absolute velocity $v_-= -1-v_{\rm s}=-2$ i.e. with the double speed of light: $r(t) = 1 - 2(t-t_0)$. In case of a white hole horizon one has $r(t) = 1 + 2(t-t_0)$. On the contrary, from the point of view of the inner observer the horizon cannot be reached and crossed: the horizon can be approached only asymptptically for infinite time: $r(\tilde t)=1+(r_0-1)\exp(-\tilde t)$. Such incompetence of the local observer, who “lives” in the curved world of superfluid vacuum, happens because he is limited in his observations by the “speed of light”, so that the coordinate frame he uses is seriously crippled in the presence of the horizon and becomes incomplete. The Schwarzschild metric naturally arises for the inner observer, if the Painlevé-Gullstrand metric is an effective metric for quasiparticles in superfluids, but not vice versa. The Schwarzschild metric Eq.(\[Schwarzschild\]) can in principle arise as an effective metric in absolute space-time; however, in the presense of a horizon such metric indicates an instability of the underlying medium. To obtain a line element of Schwarzschild metric as an effective metric for quasiparticles, the quasiparticle energy spectrum in the laboratory frame has to be $$E^2=c^2\left(1-{r_S\over r}\right)^2 p_r^2+c^2\left(1-{r_S\over r}\right) p_\perp^2~. \label{SpectrumInSchwarzschild}$$ In the presence of a horizon such spectrum has sections of the transverse momentum $p_\perp$ with $E^2<0$. The imaginary frequency of excitations signals the instability of the superfluid vacuum if this vacuum exhibits the Schwarzschild metric as an effective metric for excitations: Quasiparticle perturbations may grow exponentially without bound in laboratory (Killing) time, as $e^{t~{\rm Im} E}$, destroying the superfluid vacuum. Nothing of this kind happens in the case of the Painlevé-Gullstrand line element, for which the quasiparticle energy is real even behind the horizon. Thus the main difference between Painlevé-Gullstrand and Schwarzschild metrics as effective metrics is: The first metric leads to the slow process of the quasiparticle radiation from the vacuum at the horizon (Hawking radiation), while the second one indicates a crucial instability of the vacuum behind the horizon. In general relativity it is assumed that the two metrics can be converted to each other by the coordinate transformation in Eq.(\[Transformation\]). In condensed matter the coordinate transformation leading from one metric to another is not that innocent if an event horizon is present. The reason why the physical behaviour implied by the choice of metric representation changes drastically is that the transformation between the two line elements, $t\rightarrow t +\int^r dr~v_{\rm s}/(c^2-v_{\rm s}^2)$, is singular on the horizon, and thus it can be applied only to a part of the absolute space-time. In condensed matter, only such effective metrics are physical which are determined everywhere in the real physical space-time. The two representations of the “same” metric cannot be strictly equivalent metrics, and we have different classes of equivalence, which cannot be transformed to each other by everywhere regular coordinate transformation. Painlevé-Gullstrand metrics for black and white holes are determined everywhere, but belong to two different classes. The transition between these two metrics occurs via the singular transformation $t\rightarrow t +2\int^r dr~v_{\rm s}/(c^2-v_{\rm s}^2)$ or via the Schwarzschild line element, which is prohibited in condensed matter physics, as explained above, since it is pathological in the presence of a horizon: it is not determined in the whole space-time and it is singular at horizon. Incompleteness of space-time in effective gravity. -------------------------------------------------- It is also important that in the effective theory there is no need for the additional extension of space-time to make it geodesically complete. The effective space time is always incomplete (open) in the presence of horizon, since it exists only in the low energy “relativistic” corner and quasiparticles escape this space-time to a nonrelativistic domain when their energy increase beyond the relativistic linear approximation regime [@grishated]. Another example of the incomplete space-time in effective gravity is provided by vierbein walls, or walls with the degenerate metric. The physical origin of such walls with the degenerate metric $g^{\mu\nu}$ in general relativity has been discussed by Starobinsky at COSMION-99 [@Starobinsky]. They can arise after inflation, if the inflaton field has a $Z_2$ degenerate vacuum. The domain walls separates the domains with 2 diferent vacua of the inflaton field. The metric $g^{\mu\nu}$ can everywhere satisfy the Einstein equations in vacuum, but at the considered surfaces the metric $g^{\mu\nu}$ cannot be diagonalized as $g^{\mu\nu}={\rm diag}(-1,1,1,1)$. Instead, on such surface the metric is diagonalized as $g^{\mu\nu}={\rm diag}(-1,0,1,1)$ and thus cannot be inverted. Though the space-time can be flat everywhere, the coordinate transformation cannot remove such a surface: it can only move the surface to infinity. Thus the system of such vierbein domain walls divides the space-time into domains which cannot communicate with each other. Each domain is flat and infinite as viewed by a local observer living in a given domain. In principle, the domains can have different space-time topology, as is emphasized by Starobinsky [@Starobinsky]. In $^3$He-A such walls appear in a film of the $^3$He-A, which simulates the 2+1 vacuum. The wall is the topological solitons on which one of the vectors (say, ${\bf e}_1$) of the order parameter playing the part of the vierbein in general relativity, changes sign across the wall [@VierbeinWalls]: $${\bf e}_1(x) =\hat{\bf x} c_0\tanh {x} ~,~{\bf e}_2 =\hat{\bf y} c_0~. \label{VierbeinWall}$$ The corresponding 2+1 effective metric experienced by quasiparticles, is $$ds^2= -dt^2 +{1\over c_0^2} \left(dx^2 \tanh^{-2} {x} +dy^2\right) ~. \label{LineElement}$$ Here $c_0$ is “speed of light” at infinity. The speed of “light” propagating along the axis $x$ becomes zero at $x=0$, and thus $g_{xx}(x=0)=\infty$. This indicates that the low-energy quasiparticles cannot propagate across the wall. The coordinate singularity at $x=0$ cannot be removed by the coordinate transformation. If at $x>0$ one introduces a new coordinate $\tilde x=\int dx/\tanh {x} $, then the line element acquires the standard flat form $$ds^2=-dt^2 + d\tilde x^2 +dy^2 ~. \label{LineElementFlat}$$ This means that for the “inner” observer, who measures the time and distances using the quasiparticles, his space-time is flat and infinite. But this is only half of the real (absolute) space-time: the other domain – the left half-space at $x<0$ – which is removed by the coordinate transformation, remains completely unknown to the observer living in the right half-space. The situation is thus the same as discussed by Starobinsky for the domain wall in the inflaton field [@Starobinsky]. Thus the vierbein wall divides the bulk liquid into two classically separated flat “worlds”, when viewed by the local “inner” observers who use the low energy “relativistic” quasiparticles for communication. Such quasiparticles cannot cross the wall in the classical limit, so that the observers living on different sides of the wall cannot communicate with each other. However, at the “Planck scale” the quasiparticles have a superluminal dispersion in $^3$He-A, so that quasiparticles with high enough energy can cross the wall. This is an example of the situation, when the effective space-time which is complete from the point of view of the low energy observer appears to be only a part of the more fundamental underlying space-time. It is interesting that when the chiral fermionic quasiparticles of $^3$He-A crosses the wall, its chirality changes to the opposite [@VierbeinWalls]: the lefthanded particle viewed by the observer in one world becomes the righthanded particle in the hidden neighbouring world. Conclusion. =========== We considered here only small part of the problems which arise in the effective gravity of superfluids. There are, for example, some other interesting effective metrics, which must be exploited. Quantized vortices with circulating superfluid velocity around them simulate the spinning cosmic strings, which experience the gravitational Aharonov-Bohm effect measured in superfluids as Iordanskii force acting on the vortex [@SpinningString]; the superfluid vacuum around the rotating cylinder simulates [@parallel] Zel’dovich-Starobinsky effect of radiation by the dielectric object or black hole rotating in quantum vacuum [@Zeldovich2; @Starobinskii]. The expanding Bose condensate in the laser manipulated traps, where the speed of sound varyes in time, may simulate the inflation. The practical realization of the analogue of event horizon, the observation of the Hawking radiation and measurement of the Bekenstein entropy still remain a challenge for the condensed matter physics (see Ref.[@JacobsonTalk] for review of different proposals). However, even the theoretical consideration of the effective gravity in condensed matter can give insight into many unsolved problems in quantum field theory. We can expect that the analysis of the condensed matter analogues of the effective gravity, in particular, of the Landau-Khalatnikov two-fluid hydrodynamics [@Khalatnikov] and its extensions will allow us to solve the longstanding problem of the cosmological constant. [999]{} I.M. Khalatnikov: [*An Introduction to the Theory of Superfluidity*]{}, (Benjamin, New York, 1965). S. Weinberg, What is quantum field theory, and what did we think it is? hep-th/9702027 S. Chadha and H.B. Nielsen, Lorentz Invariance as a Low-Energy Phenomenon, Nucl. Phys. [**B 217**]{}, 125–144 (1983). G.E. Volovik, Field theory in superfluid $^3$He: What are the lessons for particle physics, gravity and high-temperature superconductivity?, Proc. Natl. Acad. Sci. USA [**96**]{}, 6042 - 6047 (1999); G.E. Volovik, $^3$He and Universe parallelism, in “Topological Defects and the Non-Equilibrium Dynamics of Symmetry Breaking Phase Transitions”, Yu. M. Bunkov, H. Godfrin (Eds.), pp. 353-387 (Kluwer, 2000); cond-mat/9902171. R.B. Laughlin and D. Pines, The Theory of Everything, unpublished. B.L. Hu, General Relativity as Geometro-Hydrodynamics, Expanded version of an invited talk at 2nd International Sakharov Conference on Physics, Moscow, 20 - 23 May 1996, e-Print Archive: gr-qc/9607070 T. Padmanabhan, Conceptual issues in combining general relativity and quantum theory, hep-th/9812018. G.E. Volovik, Axial anomaly in $^3$He-A: Simulation of baryogenesis and generation of primordial magnetic field in Manchester and Helsinki, Physica [**B 255**]{}, 86-107 (1998). D. Vollhardt, and P. Wölfle, [*The superfluid phases of helium 3*]{}, Taylor and Francis, London - New York - Philadelphia, 1990. W. G. Unruh, Experimental Black-Hole Evaporation?, Phys. Rev. Lett. [**46**]{}, 1351-1354 (1981); Sonic analogue of black holes and the effects of high frequencies on black hole evaporation, Phys. Rev. D [**51**]{}, 2827-2838 (1995). M. Visser, Acoustic black holes: horizons, ergospheres, and Hawking radiation, Class. Quantum Grav. [**15**]{}, 1767-1791 (1998). C.W. Woo, Microscopic calculations for condensed phases of helium, in: [*The Physics of Liquid and Solid Helium*]{}, Part I, eds. K.H. Bennemann and J.B. Ketterson (John Wiley & Sons, New York, 1976) . A. D. Sakharov, Vacuum Quantum Fluctuations in Curved Space and the Theory of Gravitation, Sov. Phys. Dokl. [**12**]{}, 1040-41 (1968). U.R. Fischer, and G.E. Volovik, Thermal quasi-equilibrium states across Landau horizons in the effective gravity of superfluids, gr-qc/0003017. R.C. Tolman, [*Relativity, Thermodynamics and Cosmology*]{} (Clarendon Press, Oxford, 1934). T. A. Jacobson and G. E. Volovik, Event horizons and ergoregions in $^3$He, Phys. Rev. D [**58**]{}, 064021 (1998); Effective space-time and Hawking radiation from a moving domain wall in a thin film of $^3$He-A, JETP Lett. [**68**]{}, 874-880 (1998). K. Martel and E. Poisson, Regular coordinate systems for Schwarzschild and other spherical spacetmes, gr-qc/0001069. S. Liberati, S. Sonego and M. Visser, Unexpectedly large surface gravities for acoustic horizons? gr-qc/0003105 L.D. Landau and E.M. Lifshitz, Fluid Mechanics, p. 317, Pergamon Press, 1989. N.B. Kopnin and G.E. Volovik, Critical velocity and event horizon in pair-correlated systems with “relativistic” fermionic quasiparticles, JETP Lett. [**67**]{}, 140-145 (1998). G. E. Volovik, Simulation of Painlevé-Gullstrand black hole in thin $^3$He-A film, JETP Lett. [**69**]{}, 705-713 (1999). A. Starobinsky, Plenary talk at Cosmion-99, Moscow, 17-24 October, 1999. G.E. Volovik, Vierbein walls in condensed matter, JETP Lett. [**70**]{}, 711-716 (1999). G.E. Volovik, Vortex vs spinning string: Iordanskii force and gravitational Aharonov-Bohm effect, JETP Letters [**67**]{}, 881 - 887 (1998). Ya.B. Zel’dovich, Amplification of cylindrical clectromagnetic waves reflected from rotating body, Sov. Phys. JETP [**35**]{}, 1085 (1971). A.A. Starobinskii, Amplification of waves during reflection from a rotating “black hole” Sov. Phys. JETP, [**37**]{}, 28 (1973). T.A. Jacobson, Trans-Planckian redshifts and the substance of the space-time river, hep-th/0001085.

ArXiv

--- abstract: 'Recently, Valiviita et al. (2008) have reported a large-scale early-time instability in coupled dark energy and dark matter models. We take the same form of energy-momentum exchange and specialise to the case when the interaction rate is proportional to Hubble’s parameter and the dark energy density only. Provided the coupling is made small enough for a given equation of state parameter, we show that the instability can be avoided. Expressions are derived for non-adiabatic modes on super-horizon scales in both the radiation and matter dominated regimes. We also examine the growth of dark matter perturbations in the sub-horizon limit. There we find that the coupling has almost no effect upon the growth of structure before dark energy begins to dominate. Once the universe begins to accelerate, the relative dark matter density fluctuations not only cease to grow as in uncoupled models, but actually decay as the universe continues to expand.' author: - Brendan M Jackson - Andy Taylor - Arjun Berera bibliography: - 'refs.bib' title: 'On the large-scale instability in interacting dark energy and dark matter fluids' --- Introduction ============ There is good evidence to believe the present day energy density of the universe is mostly in the form of dark energy [@2008WMAP; @2005BAO], the properties of which remain relatively unknown. Furthermore, observations of Type Ia supernovae [@1999Perl; @1998Riess] leave little doubt that the expansion of the universe is accelerating. Viable models of cosmology now require a large dark energy component, capable of producing the negative pressure required for accelerated expansion. By far the simplest model of dark energy is Einstein’s cosmological constant, $\Lambda$. The cosmological constant($\Lambda$) and cold dark matter (CDM) model, with values of today’s density parameter for the dark energy $\Omega_x \approx 0.7$ and dark matter $\Omega_m \approx 0.3$ is the current prevailing paradigm. But while consistent with observational constraints, the standard model is in many ways unsatisfactory. One such example is the ‘coincidence problem’: why are the energy densities in the dark energy and dark matter comparable today, when the redshift dependence of each is so different? Motivated to explain the coincidence problem while deviating as little as possible from the successful $\Lambda$CDM model, a coupling between dark energy and dark matter has often been considered. An energy exchange modifies the background evolution of the dark sector, and explaining the coincidence problem can be reduced to tuning a coupling parameter to an appropriate value. The coupling enters via the continuity equations. With energy exchange rate $Q$ between the dark energy (subscript $x$) and the cold dark matter (subscript $c$), the dark energy obeys the continuity equation in conformal time $$\dot{\rho_x} + 3\mathcal{H}(1+w_x)\rho_x = -Q,$$ while the dark matter obeys $$\dot{\rho_c} + 3\mathcal{H}\rho_c = Q.$$ Here we have introduced the equation of state parameter $w_A$ that gives the ratio of the pressure $P_A$ to the energy density $\rho_A$ of a fluid, $$w_A = \frac{P_A}{\rho_A}.$$ We have also used $\mathcal{H} = aH$, where $a(t)$ is the expansion scale-factor and $H$ the Hubble parameter. Acceleration of the expansion rate requires the energy density of the universe to dominated by a fluid with an effective equation of state parameter $w_{\rm{eff}} < -1/3$. We do not allow the phantom case of $w < -1$ in this work. Simple solutions for the background exist for couplings of the form $Q = \alpha \mathcal{H} \rho_x +\beta \mathcal{H} \rho_c$. These were initially investigated by Chimento [@1997Ch] and then expanded upon by Barrow and Clifton [@2006B], who provided general solutions for any cosmology with two components exchanging energy in such a fashion, provided the components were modelled as cosmological fluids with constant $w$. Quartin et al. [@2008Q] examined the observational constraints upon such a class of models, significantly limiting the available parameter space. Again, the equation of state parameter was treated as fixed. Non-zero values of $\beta$ were found to reduce the required fine-tuning of the initial energy density, as well as increase the observationally allowed values of $\alpha$ [@2008Q]. A coupling would influence more than just the background dynamics of the universe. In particular, the growth of perturbations in the coupled fluids would be affected. Recent work by Valiviita et al. [@Valiviita2008] has shown that couplings of the simple form described above, with constant $w$, exhibit extremely rapid growth of dark energy fluctuations on super-horizon scales in the early universe. In fact, the perturbations in the dark energy become unstable for any model with non-zero $\beta$, no matter how small this parameter is made. While this would appear to rule out all couplings of the above form and with constant $w$, the explicit examples in [@Valiviita2008] included no cases where the interaction rate was proportional to the density of dark energy and not of the dark matter, i.e. with $\beta = 0$ and $\alpha \neq 0$. Here we look at just such a scenario. Background evolution {#background} ==================== Friedmann’s equation relates the evolution of the scale-factor $a(t)$ to the background energy density $\rho$. We make use of conformal time, $\tau$, which is related to cosmic time via $dt = a d\tau$. Overdots indicate derivatives with respect to conformal time. Friedmann’s equation reads $$\mathcal{H}^2 \equiv \left(\frac{a^{\prime}}{a} \right)^2 = \frac{8 \pi G}{3} \rho a^2.$$ With the choice of $Q = \alpha \mathcal{H} \rho_x$, the continuity equations can be solved to yield [@2006B; @1997Ch] $$\rho_x = \rho_{x,0} \; a^{-(3(1+w)+\alpha)},$$ $$\begin{aligned} \rho_c &=& -\frac{\alpha \rho_{x,0}}{3w+\alpha} \; a^{-(3(1+w) +\alpha)} \nonumber \\ &+& \left(\rho_{c,0} + \frac{\alpha \rho_{x,0}}{3w+\alpha}\right) \; a^{-3}.\end{aligned}$$ We follow the standard notation where a subscript zero indicates today’s value. We normalise the scale-factor so that $a_0 = 1$. The ratio of dark energy to dark matter density $r$ can then be written $$\frac{1}{r} = \frac{\rho_{c}}{\rho_{x}} = \left( \frac{\rho_{c,0}}{\rho_{x,0}} + \frac{\alpha}{3w+\alpha} \right) a^{3w+\alpha} - \frac{\alpha}{3w+\alpha}$$ With $|3w| < \alpha$, the dark energy and dark matter approach a constant ratio as the universe expands. The coincidence problem can be said to be solved if this ratio is of order unity, but this requires a value of $\alpha$ already observationally excluded [@2008Q]. Nevertheless, as argued in [@2008Q], non-zero values of $\alpha$ can still be said to alleviate the problem. We restrict ourselves to positive values of $\alpha$. Perturbed FRW Cosmology {#perturbed} ======================= We assume a flat FRW cosmology and work in Newtonian gauge, $$-ds^2 = dt^2(1+2\Psi) - a^2 (1-2\Phi)\delta_{i j}dx^i dx^j,$$ with metric signature $(+,-,-,-)$. We work in Fourier space, using comoving Fourier wave-vectors $k^i = k_i$, so that $\partial_i \partial^i \to k^2/a^2$. The four-velocity of fluid $A$ is given by $$U^\mu_{(A)} = \left((1-\Psi), a^{-1} v^i_{(A)} \right).$$ The peculiar velocity three-vector $v^i=v_i$ are small. We define the velocity perturbation $\theta \equiv \partial_i v^i$. Energy-momentum tensors ----------------------- The energy-momentum tensor for fluid $A$ is given by: $$T^{\mu (A)}_{\; \; \nu} = \left( \rho^{(A)} + P^{(A)} \right) U^{(A)\mu} U^{(A)}_\nu - \delta^\mu_{\,\nu} P^{(A)}.$$ The total energy-momentum tensor is simply the sum of the components, $$T^{\mu}_{\; \; \nu} = \sum_A T^{\mu (A)}_{\; \; \nu}.$$ We define the density perturbation in fluid $\delta_A$ using $\rho_A \equiv (1+\delta_A)\bar{\rho}_A$. An overbar denotes the background quantity, though we will usually leave this implicit. Energy and momentum conservation for fluid $A$ implies $$\nabla_{\mu} T^{\mu (A)}_{\; \; \nu} = Q^{(A)}_\nu, \nonumber$$ and conservation for the entire system requires $$\sum_A Q^\mu _{(A)} = 0.$$ The four-vector $Q^\mu _{(A)}$ governs the energy exchange between components, and it is to this we now turn our attention. Covariant energy exchange ------------------------- The energy exchange in the background does not determine a fully covariant form of energy exchange [@K2005; @Valiviita2008]. Instead, an energy exchange four-vector must be specified. We adopt the approach of [@Valiviita2008] and consider two scenarios; aligning the four-vector with the dark energy four-velocity, $$\label{dark energy align} Q^\mu_{(A)} = Q_A U^\mu_x,$$ or with the four-vector of the dark matter four-velocity, $$\label{dark matter align} Q^\mu_{(A)} = Q_A U^\mu_c.$$ These choices produce slightly different outcomes, and the differences are noted as we proceed. To produce the desired changes to the continuity equations, we see that $aQ_c = -aQ_x = \alpha \mathcal{H} \rho_x$ in both cases. We also make the common assumption that $\alpha \mathcal{H}$ gives an interaction rate that has no spatial dependence. We therefore perturb only $\rho_x$, not $\mathcal{H}$, in the coupling. Sound speed of dark energy -------------------------- The speed of sound of a fluid or scalar field $A$ is denoted by $c_{s A} $. For a barotropic fluid with a constant value of $w_A$, then $c_{s A}^2 = w_A$. This leads to an imaginary speed of sound for the dark energy ($c_{s x}^2 = w_x < 0$). An imaginary sound speed leads to instabilities in the dark energy; the problem is commonly remedied by imposing a real sound speed by hand. A common choice (and the one we make here) is the scalar field value of $c_{s x} = 1$. This choice leads to an intrinsic non-adiabatic pressure perturbation in the dark energy. This contains a term, highlighted recently in [@Valiviita2008], that arises due to the coupling between dark energy and dark matter. We include this term, and refer the interested reader to [@Valiviita2008]. Perturbation equations of motion -------------------------------- Conservation of the energy-momentum tensor, combined with results of the previous sections and our choice of energy exchange four-vector, implies the following. For the dark energy density perturbation: $$\begin{aligned} \label{delta x} \delta_x^\prime &+& 3\mathcal{H} (1-w_x) \delta_x + (1+w_x) \theta_x + 9\mathcal{H}^2 (1-w_x^2)\frac{\theta_x}{k^2} \nonumber \\&-& 3(1+w_x)\Phi^\prime= -\alpha \mathcal{H} \left[\Psi + 3\mathcal{H}(1-w_x) \frac{\theta_x}{k^2} \right].\end{aligned}$$ For the dark energy velocity perturbation, the right-hand side differs slightly depending on our choice of energy exchange four-vector. $$\begin{aligned} \label{theta x} \theta_x^\prime - 2\mathcal{H}\theta_x &-& \frac{k^2}{1+w_x} \delta_x - k^2\Psi \nonumber \\ &=& \begin{array}{ll} {\displaystyle \frac{(1+b)\alpha \mathcal{H}}{1+w_x} \theta_x}, \end{array}\end{aligned}$$ where $$b = \left\{ \begin{array}{rl} 0 & \text{if } Q^\mu_{(A)} = Q_A U^\mu_x, \\ 1 & \text{if } Q^\mu_{(A)} = Q_A U^\mu_c. \end{array} \right.$$ For the dark matter, the density perturbation obeys $$\label{delta c} \delta_c ^\prime + \theta_c -3\Phi^{\prime} = \alpha \mathcal{H}\frac{\rho_x}{\rho_c}\left[ \delta_x-\delta_c \right],$$ while the velocity perturbation is governed by $$\begin{aligned} \label{theta c} \theta^\prime_c &+& \theta_c \mathcal{H} - k^2 \Psi \nonumber \\ &=& \begin{array}{ll} { \displaystyle (1-b)\alpha\mathcal{H} \frac{\rho_x}{\rho_c} \left[\theta_x - \theta_c \right]}. \\ \end{array}.\end{aligned}$$ The perturbed Einstein equations are well known, and we do not reproduce them here. They can be found in [@1995Ma], whose notation for the scalar metric perturbations we share. Initial conditions in the early radiation era {#early} ============================================= In [@Valiviita2008] it was shown that models with $\beta \neq 0$ suffered from an early time large-scale instability no matter how small the value of $\beta$. This was driven by a term proportional to $\beta$ on the right-hand side of equation (\[theta x\]). A term proportional to $\alpha$ also exists, which can be large if $w$ is close to $-1$ or $\alpha$ is made very large. In this section we examine how large this term needs to be to cause the non-adiabatic mode to be a growing one. We consider super-horizon scales ($k/\mathcal{H} \ll 1$) and assume adiabatic initial conditions. The gravitational potentials are dominated by fluctuations in the dominant fluid (radiation or matter). The well known result is that $\Phi \propto \Psi = \text{constant}$. The constant of proportionality in the radiation era is determined by the anisotropic stress generated by the neutrinos. In the absence of neutrinos or in the matter dominated era, the potentials are equal. These assumptions will be invalid only if perturbations in the dark energy are large enough to influence the gravitational potentials. As the dark energy has a very low background density in the radiation era, this can only happen if $\delta_x$ grows extremely large. Neglecting time derivatives of the gravitational potential, and keeping only leading order terms in $k/\mathcal{H}$, the dark energy equations (\[delta x\]) – (\[theta x\]) can be combined into a second order equation: $$\begin{aligned} \label{second order de} \delta_x^{\prime \prime} &+& \mathcal{H} \left(1-3w-\frac{(1+b)\alpha}{1+w} - 2\frac{\mathcal{H}^\prime}{\mathcal{H}^2} \right)\delta_x^{\prime} \nonumber \\ &+& 3\left(\mathcal{H}^2 \left(1-b\frac{\alpha}{1+w} \right) - \mathcal{H}^\prime \right)(1-w) \, \delta_x \nonumber \\ &=& \left(A\mathcal{H}^2 + B\mathcal{H}^\prime \right) \Psi,\end{aligned}$$ The constants $A$ and $B$ have values unimportant for our analysis. In the radiation era, $\mathcal{H} = \tau^{-1}$. The adiabatic mode is therefore an obvious solution: $\delta_x \propto \Psi = \text{constant}$. To find the remaining solutions, we define a new variable $\hat{\delta}_x = \delta_x + C\Psi$, with the constant $C$ chosen such that the right-hand side of (\[second order de\]) is equal to zero. In the radiation dominated era, we can then write: $$\begin{aligned} \label{final early time second order} \tau^2 \hat{\delta}_x^{\prime \prime} &+& \left(3-3w-\frac{(1+b)\alpha}{1+w} \right)\tau \hat{\delta}_x^{\prime} \nonumber \\ &+& 3(1-w)\left(2-b\frac{\alpha}{1+w}\right)\hat{\delta}_x = 0\end{aligned}$$ When $b=1$, equation (\[final early time second order\]) becomes formally the same equation found by He et al.[@He:2008], despite the differing assumptions made about the physics involved. In their investigation of perturbations given a background coupling of the form $Q = \alpha\mathcal{H}\rho_x$, they choose to set the net momentum exchange to zero ($Q^i_{(A)} = 0$), in contrast to our adoption of the form of momentum exchange used in [@Valiviita2008]. The differences between the $b=1$ choice of momentum exchange and zero net momentum exchange arise in the equations for the dark matter perturbations, which are not used in the above analysis, nor in the analysis by He et al. This leads to the same behaviour of dark energy perturbations. This is not true when $b=0$, and can result in different behaviour (oscillatory or non-oscillatory) for the same choice of parameters (see the remainder of this section). Note also that the simplifying assumptions, and their justifications, made in [@He:2008] differ to those made here: we have neglected terms that will be small due to choice of intial conditions, and simplified the result by extracting the adiabatic mode. In [@He:2008], terms are instead neglected that are found to be small from a numerical analysis. Solutions of equation (\[final early time second order\]) are power laws, $\hat{\delta}_x \propto \tau^{n_{\pm}}$. The index is given by: $$n_{\pm} = \frac{\Gamma}{2(1+w)} \pm \frac{\sqrt{\Delta}}{2(1+w)},$$ where we follow the notation of [@He:2008] and have defined the quantities $$\Gamma = 3w^2 + w + (1+b)\alpha - 2,$$ and $$\begin{aligned} \Delta &=& 9w^4 + 30w^3 + (13-6(b-1)\alpha)w^2 \nonumber \\ &+& 2w\left[(1+b)\alpha -14\right] + 4(2b-1)\alpha \nonumber \\ &+& (1+b)^2 \alpha^2 - 20.\end{aligned}$$ In the limit of $w$ very close to -1 (and assuming $\alpha$ is reasonably small), we can expand as a series in $(1+w)$, $$\begin{aligned} \frac{\Gamma}{2(1+w)} \approx -5/2 + \frac{(1+b)\alpha}{2(1+w)} &+& \frac{3(1+w)}{2} \nonumber \\ &+& \mathcal{O}(1+w)^2,\end{aligned}$$ $$\begin{aligned} \Delta &\approx& (1+3b)\alpha^2 + 2(7b-5)(1+w) \alpha \nonumber \\ &+& (6(1-b)\alpha -23)(1+w)^2 + \mathcal{O}(1+w)^3.\end{aligned}$$ When $\alpha = 0$, the non-adiabatic mode is decaying. But when the coupling is switched on, the second term in $\Gamma$ can become very large, resulting in $n_+ \gg 1$. For a range of $\alpha$ and $w$, which is much larger in the $b=0$ case, oscillatory behaviour can also result (due to $\Delta$ becoming negative). The instability means these coupled models suffer from all the problems outlined in [@Valiviita2008] for $\beta \neq 0$ models, unless the value of $\alpha$ is made small enough. The closer $w$ is to $-1$, the smaller $\alpha$ must be made to avoid the instability. This is in contrast to $\beta \neq 0$ models, which are unstable no matter how small the parameter $\beta$ is made. In the matter dominated era, we can carry out the same procedure, this time with $\mathcal{H} = 2\tau^{-1}$. We find, $$\frac{\Gamma}{2(1+w)} \approx -9/2 + \frac{(1+b)\alpha}{1+w} + 3(1+w) + \mathcal{O}(1+w)^2,$$ and $$\begin{aligned} \Delta &\approx& 4(1+3b)\alpha^2 + 12(5b-3)\alpha(1+w) \nonumber \\ &+& (24(1-b)\alpha-62)(1+w)^2 + \mathcal{O}(1+w)^3.\end{aligned}$$ Once again, the second term in $\Gamma$ can result in a rapidly growing dark energy fluctuation. We have solved equations (\[delta x\]) – (\[theta c\]) numerically in the matter dominated regime (Figure \[graph1\]), where we need not worry about the radiation fluid and its perturbations. The analytical agreement is excellent until the mode leaves the horizon ($k\tau \sim1$). Numerically we see that when this happens the mode begins to oscillate with a growing amplitude. ![\[graph1\]The evolution of the dark energy outside the horizon in a matter dominated universe, for modes of three different scales. We take $\alpha = 0.08$ and $w=-0.98$, and $b=1$. The agreement with the analytical approximation is excellent until the mode begins to leave the horizon ($k\tau \sim 1$). Vertical lines indicate when this occurs for each mode.](graph1) Sub-horizon evolution in the matter and radiation dominated eras {#sub} ================================================================ In the sub-horizon limit, $\mathcal{H}^2/k^2 \ll 1$, equation (\[delta x\]) yields, $$\label{small delta x} \delta_x^\prime + 3\mathcal{H}(1-w_x)\delta_x + (1+w_x) \theta_x -3(1+w_x)\Phi^\prime = 0.$$ Note the two terms on the right-hand side of equation (\[delta x\]) scale as $\mathcal{H}^2/k^2$. As these are the only two terms containing the coupling parameter $\alpha$, the simplified equation above does not contain the coupling parameter. One of the perturbed Einstein equations simplifies to Poisson’s equation in comoving coordinates, $$-k^2\Psi = 4\pi G a^2 \left(\rho_x \delta_x + \rho_c \delta_c \right).$$ Without the coupling, the dark energy perturbations are significantly suppressed on small-scales in comparison to dark matter perturbations, primarily due to its large speed of sound [@Bean]. The coupling does nothing to alter this fact unless the right-hand side of equation (\[theta x\]) makes a significant contribution. If the early time instability has been avoided this cannot be the case, as $\displaystyle \alpha/(1+w)$ will be small. Thus it is reasonable to expect the dark energy to remain suppressed on sub-horizon scales. We therefore neglect dark energy perturbations for the remainder of this section. By combining equations (\[delta c\]) and (\[theta c\]), we eliminate $\theta_c$ and find a second-order equation for the growth of the matter density perturbation. From the above argument, we have neglected dark energy perturbations. Keeping only the dominant gravitational terms, $$\begin{aligned} \label{second order w} \delta_c^{\prime \prime} &+& \mathcal{H}\left(1+2\alpha \frac{\rho_x}{\rho_c} \right) \delta_c^\prime \nonumber \\ &+& \alpha \frac{\rho_x}{\rho_c}\left(\mathcal{H}^\prime - \mathcal{H}^2(\alpha + 3w - 1) \right)\delta_c = -k^2 \Psi.\end{aligned}$$ We note that in the limit of $\alpha \to 0$, this reduces to the standard growth equation, with the well known growing mode $\delta_c \propto \tau^2$ in both matter and radiation eras. The additional terms are proportional to $\alpha r$ (recall $r$ is the ratio of dark energy to dark matter). In the matter dominated regime, then $\alpha r \ll 1$, and these terms will be negligible. Even when $r \sim 1$, the terms will be suppressed by the size of $\alpha$, which will be small itself. The dominant effect causing a deviation from standard linear growth of structure in the matter dominated regime will therefore be, as in an uncoupled cosmology, the influence of the dark energy upon the expansion rate. The growth of structure in a coupled model can therefore be treated in the matter dominated regime simply as an uncoupled model with an effective dark energy equation of state parameter $\displaystyle w_{\rm{eff}} = w + \alpha/3$. This will cease to be true only when the background energy density of matter is no longer well approximated by its usual $\rho_c \propto a^{-3}$ dependence, and the late time scaling behaviour becomes apparent. Sub-horizon evolution in the dark energy dominated era {#late} ====================================================== The coupling between dark energy and dark matter eventually leads to a constant ratio between the two dark components. With a small value of $\alpha$, the dark energy still dominates. We consider the evolution of structure once this equilibrium has been reached. Friedmann’s equation solves to yield $$\mathcal{H}=2(\alpha+3w+1)^{-1}\eta^{-1},$$ with the new time variable $\eta = \tau - \tau_\infty$. Note that as $\eta$ increases ($\tau$ decreases and approaches $\tau_\infty$), the scale-factor increases. The constant of integration, $\tau_\infty$, is the radius of the de Sitter event-horizon in the uncoupled case with a cosmological constant. The growth equation can then be written as, $$\begin{aligned} \eta^2 \delta_c^{\prime \prime} &+& \eta \frac{2-12w-4\alpha}{\alpha+3w+1} \delta_c^{\prime} \nonumber \\ &+& 2\frac{(3\alpha +9w - 1)(3w+\alpha) + \alpha w}{(\alpha+3w+1)^2}\delta_c = 0.\end{aligned}$$ This admits power law solutions, $\delta_c \propto \eta^m$, where $$m = \frac{5}{2} - \frac{3}{1+3w+\alpha} \pm \frac{1}{2} \sqrt{1-\frac{8\alpha}{w(1+3w+\alpha)^2}}.$$ In the range of $\alpha$ and $w$ relevant to the problem, then $m>0$. Recalling that $\eta$ decreases with increasing scale-factor, we see that the universe becomes steadily more homogeneous as it expands. We interpret this to be a combination of two effects. The first is the accelerating expansion, which slows and (without the coupling) eventually stops structure formation. This occurs, for example, in $\Lambda$CDM cosmology when the cosmological constant becomes dominant. The second effect is that dark energy is constantly being transformed into dark matter, via the coupling. As the rate is proportional to the density of the dark energy, and the dark energy density is essentially uniform, new dark matter is also created uniformally. This rising ‘background’ of dark matter reduces the relative value of the fluctuations, reducing $\delta_c$. We have also investigated both numerically and analytically the extreme late time behaviour, where $k\eta \ll 1$ and the modes can be thought of as leaving the horizon. We find the tend toward homogeneity continues, but with a much milder rate of decay. Conclusions =========== We have shown that constant $w$ models with the same form of energy-momentum exchange considered by [@Valiviita2008] suffer from an instability with $\alpha \neq 0$, even if $\beta = 0$. However the instabilities in these models are not as severe as those facing models with $\beta \neq 0$. There is at least some non-trivial region of parameter space where the instability can be avoided, although the value of $\alpha$ is now constrained both from background observables [@2008Q] and from stability requirements to be extremely small. Despite this, any non-zero value of $\alpha$ will lead to a late-time scaling regime, alleviating (even if not solving) the coincidence problem. It is unfortunate that with $\alpha$ constrained to such small values, we find any observable trace upon the growth of CDM structure will be negligible. Detecting a coupling of this form from measurements of large-scale structure is extremely doubtful, even with the precision promised by future experiments. We have said nothing up to this point of models of dark energy with a variable equation of state parameter, such as scalar-field (quintessence) models. The same caveats in [@Valiviita2008] apply here. Much of the above analysis will not apply in variable $w$ models, although some parameterisations such as the often used $w = w_0 + (1-a) w_a$ lead to fixed $w$ over large periods of time. Our analysis will apply during those epochs of constant $w$. We refer interested readers to recent work on quintessence with couplings of this or similar form (such as recent work [@Chong2008; @Corasaniti2008]). The future decay of dark matter fluctuations is an interesting result. It implies observers today find themselves close to the time of maximum inhomogeneity. The more the coincidence problem is alleviated, the closer to the late time scaling regime today becomes, and thus the closer to the peak of inhomogeneity. Without the coupling, observers find themselves at the time of the end of structure growth. The root cause in both cases is the acceleration of the universe only beginning today. Our position as apparently privileged observers in this fashion remains difficult to explain in any satisfactory way. B.M.J. acknowledges the support from theSTFC. We also thank Fergus Simpson for bringin interaction models to our attention. *Note added in proof:* after submission we became aware of the work by [@Gavela2009] in which an independant analysis of general instabilities in coupled models was carried out, as well as more specifically the model considered in this work. They provide cosmological parameter constraints for negative values of $\alpha$ and find a non-trivial region consistent with observations.

ArXiv

--- abstract: 'Two basic features of assemblages of unicellular plankton: (1) their high biodiversity and (2) the power-law structure of their abundance, can be explained by an allometric scaling of cell growth and mortality with respect to cell size. To show this, we describe a numerical study of a size-structured, multispecies, population-dynamic model; the model has a single resource, supporting an arbitrary number of phytoplankton and zooplankton species. If the number of plankton species is large enough, the death rate of prey and cell growth rate of predators have approximate allometric scalings with cell size. Together, these scalings give rise to an equilibrium distribution of abundance near the power law, on which many species can coexist. Scalings of this kind cannot be achieved if the number of species is small. This suggests that the conjunction of species-richness and power-law structures in plankton communities is more than a coincidence. Although the exact allometric scalings used here should not be expected in practice, exclusion of species should be relatively slow if they lie close to the power law. Thus the forces needed to achieve coexistence could be effective, even if they are relatively weak.' author: - 'Richard Law$^{1,\ast}$, Jos[é]{} A. Cuesta$^{2,3}$, Gustav W. Delius$^1$' bibliography: - 'plankton5.bib' title: '**Plankton: the paradox and the power law**' --- $^1$ Department of Mathematics and York Centre for Complex Systems Analysis, University of York, York, United Kingdom; $^2$ Grupo Interdisciplinar de Sistemas Complejos (GISC) and UC3M-BS Institute of Financial Big Data (IFiBiD), Departamento de Matemáticas, Universidad Carlos III de Madrid, Madrid, Spain; $^3$ Instituto de Biocomputación y Física de Sistemas Complejos (BIFI), Universidad de Zaragoza, Zaragoza, Spain. $\ast$ Corresponding author; e-mail: [email protected] Introduction ============ This paper is motivated by two widely observed features of aquatic ecosystems. The first is a great diversity of plankton taxa, in the case of phytoplankton seemingly unconstrained by the small number of resources for which they compete (the paradox of the plankton: @hutchinson:61). The second feature is a tendency for abundance of aquatic organisms to lie near a special power-law function of body mass. This function corresponds to a biomass density of aquatic assemblages that changes rather little, as body mass (logarithmically scaled) is increased; it is found in the size range of the plankton [@gaedke:92; @quinones:03; @sanmartin:06], and is thought to apply much more broadly from bacteria to whales [@sheldon:72]. It appears difficult to account for both these features simultaneously: they call both for a mechanism for species coexistence, and also for an emergent assemblage close to the power-law. Although both observations have separately been the subject of much research, and the need to link them is well recognized [@armstrong:99], a unified dynamical system that generates both features has not been described. In a phenomenological sense, coexistence of species is a prerequisite for the power law in assemblages of unicellular eukaryotic plankton. These organisms only double in size before cell division, and consequently a power law spanning the size range of unicellular plankton depends on the coexistence of multiple species. The purpose of the work here, and a mathematical paper that underpins it [@cuesta:16], is to examine the [*mechanistic*]{} link between them. For this we introduce a third ingredient: allometric scalings between ecological rates and cell size. These scalings are important because without them a power-law solution would not be possible. Some of the scalings, such those of metabolism and cell doubling time, are themselves well established empirically in unicellular plankton [@maranon:13; @lopez:14]. Other scalings are not, namely those of cell growth and death rates that stem from predation: these have to emerge directly from the predator-prey interactions. The core of this paper is to show that predation can generate the allometric scalings needed, leading to results consistent with a triangle of linked observations shown in Fig. \[fig:triangle\]. ![Three basic observed properties of assemblages of unicellular plankton; these motivate the multispecies, population-dynamic model in this paper. []{data-label="fig:triangle"}](images/fig_triangle.pdf){width="8cm"} Ecologists do not have a generally accepted mechanism for coexistence of a large number of plankton species [@roy:07]. Hutchinson’s own preferred solution of environmental fluctuations favouring different species (the intermediate disturbance hypothesis) is insufficient in general [@fox:13]. It slows down the rate of loss of species, but this is not equivalent to placing bounds on abundance of species favouring them when they are rare and restricting them when common [@chesson:00; @fox:13]. One promising idea for bounding population increase in microbial ecology is “killing the winner”, in which top-down control is strongest on the fastest-increasing prey [@thingstad:97; @winter:10]. This was developed with phage and bacteria in mind, and parallels the idea of predator-mediated coexistence used elsewhere in ecology [@leibold:96; @vage:14]. Here we extend the idea to predation in plankton assemblages. Neither do ecologists have a clear-cut model mechanism to generate the observed power-law equilibrium in assemblages of unicellular plankton. Ideas are better-developed in multicellular organisms from the study of so-called size-spectrum models [@silvert:78; @silvert:80; @benoit:04; @andersen:06; @capitan:10; @datta:10; @datta:11; @hartvig:11; @guiet:16], where organisms grow, often over orders of magnitude, by feeding on and killing smaller organisms [@datta:10; @plank:11]. But the power-law structure extends down to unicellular plankton [@gaedke:92; @quinones:03; @sanmartin:06] where feedbacks are different because organisms do no more than double in body mass. One suggestion is that, although plankton species may individually fluctuate in complicated ways, patterns become more regular after aggregation to the assemblage as a whole [@huisman:99; @scheffer:03]; it remains to be seen whether a power-law pattern after aggregation is likely to emerge from this. The work here builds on a mathematical theory proposed by @cuesta:16 for scale-invariant dynamics of a continuum of species with a continuous trait (a characteristic body size) that spans an unlimited range from zero upwards, in addition to a continuous body size within species conventionally used in size-spectrum models. These so-called trait-based models are tractable enough to allow some formal mathematical analysis of multispecies assemblages [@andersen:06; @hartvig:11]. However, they depend on assumptions: (1) that the trait is continuous (the number of species is infinite), and (2) that the trait is unbounded (i.e. the lower bound is zero, and the upper bound tends to infinity). Clearly, real-world assemblages only support a finite number of species, and must have lower and upper bounds to cell size. This calls for a numerical analysis as given in this paper. To keep a close link to the mathematical theory [@cuesta:16], we retain its assumptions about allometric scalings of ecological rates with cell mass; only the scalings of cell death and growth are allowed to emerge from the predator-prey interaction. However, it is not our intention to suggest that natural aquatic ecosystems satisfy the scalings exactly—there are many sources of variation that stand in the way of this. Our argument is that approximate scalings of processes with body mass in unicellular organisms bring them closer to ecological balance than has previously been thought. This slows down the rate of exclusion of species, with the consequence that other mechanisms for coexistence [@roy:07] can be more effective, even if their effects are not big. Note also that the plankton communities here are built on a single limiting resource to show that coexistence near the power law is possible even under the simplest and most exacting conditions; we are not, however, suggesting that there is only one limiting resource in natural plankton assemblages. Theory ====== We envisage an assemblage of unicellular plankton species, each species having a fixed characteristic cell size, defined as 1/2 of its maximum cell mass. For simplicity, mass is measured relative to that of a notional cell with equivalent spherical diameter (ESD) of 2 $\mu$m and mass $w_0$. Throughout, we write $w$ as the mass expressed on this relative scale, and denote the characteristic cell mass of species $i$ as $w_i$ on this relative scale. Lower and upper bounds on cell size are set to span much of the range of unicellular eukaryotic plankton, with ESDs from approximately 2 to 200 $\mu$m. The variables used are as follows. $N$ is the concentration of nutrients (mass volume$^{-1}$), and $P_i, Z_i$ are the total densities of a species of phytoplankton or zooplankton respectively, with species denoted by index $i$ (dimensions: volume$^{-1}$). However, to allow for allometric scalings in growth of cells, we disaggregate density down to cell mass within species. The state variables are therefore $p_i(w), z_i(w)$ for the density of phytoplankton and zooplankton respectively at cell mass $w$ (dimensions also volume$^{-1}$, because of the relative mass scale chosen above). All these variables are time dependent, but for brevity the time argument is omitted. The model depends on allometric scalings of cell growth and death with cell mass. We therefore start with these details of cell life histories, and use this as the input into the population dynamics of interacting species. In the population dynamics, phytoplankton species are coupled by competition for a single limiting resource acting on the growth of cells. Zooplankton and phytoplankton species are coupled by predation leading to death of phytoplankton cells and growth of zooplankton cells. Zooplankton species are coupled through predation by larger species on smaller ones. A scaling for growth of phytoplankton cells ------------------------------------------- The von Bertalanffy growth model is a good place to begin, as it is based on allometric scalings with body mass [@vonbertalanffy:57]. It is more often thought of in the context of multicellular organisms, and is used here on the basis that, in unicellular organisms, growth still stems from the balance between (a) resource uptake, and (b) loss through metabolism [@kempes:12]. For cell growth of a phytoplankton species $i$, we write $G^{(p)}_i(w)$: $$\frac{dw}{dt} = G^{(p)}_i(w) = \underbrace{A_iw^{\alpha}}_{(a)} - \underbrace{B_iw^{\beta}}_{(b)} . \label{eq:growth}$$ Here $t$ is time, $\alpha$ and $\beta$ ($\alpha <\beta$) are parameters that scale respectively the gain of mass through resource uptake and the loss of mass through metabolism with body mass $w$, and $A_i$ and $B_i$ are further species-dependent parameters for gain and loss respectively. It might seem from eq. that there are two separate allometric scalings with cell mass. However, it is known that over most of the size range of eukaryotic phytoplankton species there is an approximate scaling of the intrinsic rate of increase [@maranon:13]. This implies a single scaling of doubling time $T$ with respect to characteristic cell mass that has an exponent say $\xi$, i.e. $T \propto (w_i)^{\xi}$ (see definition of $w_i$ at start of this section). The scaling of doubling time also implies a scaling on how fast cells from species with different $w_i$s grow: cells that take longer to double, must also have a lower mass-specific growth rate. Terms $A_i$ and $B_i$ in eq. carry this dependence on $w_i$ and, with some algebra [@cuesta:16], eq. can be rewritten as: $$G^{(p)}_i(w,N) = w^{1-\xi}\left[\tilde a(N)\left(\frac{w}{w_i}\right)^{\alpha+\xi-1} -\tilde b\left(\frac{w}{w_i}\right)^{\beta+\xi-1}\right]. \label{eq:gp}$$ where $A_i(N) = \tilde a(N)\cdot(w_i)^{-(\alpha+\xi-1)}$, $B_i = \tilde b\cdot(w_i)^{-(\beta+\xi-1)}$, $\tilde a(N)$ and $\tilde b$ are independent of $w_i$, and parameters $\alpha$ and $\beta$ are such that $2w_i$ is always less than the asymptotic value of cell size achievable from eq. . The reason for writing the growth rate in this form is simply to make it clear that all phytoplankton species have the same basic growth function. This takes the form of a power function $w^{1-\xi}$ of cell mass, multiplied by a function of the ratio $w/w_i$. In other words, the growth function scales with cell mass. We have made use of a technical point that allometric scaling can be done with respect to $w$ or $w_i$; see Appendix \[ref:w\_wi\_scaling\]. Note that, in eq. , we have introduced a dependence of $\tilde a$ on the resource concentration $N$; $\tilde a$ is a strictly increasing function of $N$, since more resource leads to more gain in mass. This function is important because it has the effect of coupling the dynamics of all phytoplankton species through depletion of the resource. A scaling for growth of zooplankton cells {#ref:z_scalings} ----------------------------------------- We assume that zooplankton cells, like phytoplankton, grow through a balance between resource uptake and metabolism, as in eq. . However, scaling of resource-uptake needs more attention because these cells grow by consuming other cells, and therefore depend on the abundance of these prey. Thus, if this growth is to have an allometric scaling with cell size, it has to emerge from the predator-prey interactions. Here we introduce three assumptions, justifiable on empirical grounds, that make an allometric scaling possible, although by no means inevitable. The assumptions are as follows. (a) Zooplankton feeding depends on a volume sensed per unit time that scales with cell mass. This is based on observations of encounters of protists with phytoplankton [@delong:12]. (b) Feeding occurs on cells around a fixed ratio of the mass of the consumer. Size-based feeding is well documented in plankton, although the predator-prey size ratios are smaller in larger multicellular consumers [@wirtz:12]. Within the size range of unicellular organisms, the change in ratio is relatively small, and a fixed ratio is a reasonable approximation [@wirtz:12]. (c) Predation increases as prey species become abundant more than a simple law of mass action would allow; this acts as a stabilizing force on the population dynamics [@chesson:00], and is closely related to the notion of killing-the-winner [@thingstad:97; @winter:10]. These assumptions retain some elements of the complex food web that operates within plankton assemblages, avoiding the reduction to a simple bulk phytoplankton–zooplankton trophic connection [@boyce:15]. With these assumptions, the rate $S^{(p)}_{ij}(w,w')$ (respectively $S^{(z)}_{ij}(w,w')$) at which a zooplankton cell of species $i$ and size $w$ consumes a cell of species $j$ phytoplankton (respectively zooplankton) of size $w'$ ($w>w'$) is given by $$S^{(p)}_{ij}(w,w') = A w^{\nu} \times s(w/w') \times P_j^{\chi} \label{eq:sp}$$ $$S^{(z)}_{ij}(w,w') = \underbrace{A w^{\nu}}_{(a)} \times \underbrace{s(w/w')}_{(b)} \times \underbrace{ Z_j^{\chi}}_{(c)} , \label{eq:sz}$$ where (a), (b) and (c) formalize the three assumptions above. In part (a), $A$ sets the overall level of encounters, and $\nu$ is the scaling exponent. The function (b) is a feeding kernel that distributes predation around a fixed ratio of the predator cell mass, assumed for simplicity to be independent of predator species $i$ and prey species $j$. A positive value of $\chi$ in part (c) takes the mortality of prey species $j$ above the level of the law of mass action and would be expected to help stabilize the assemblage. We used the total density of prey $j$ ($P_j$, $Z_j$) on the grounds that fine sorting of predation at the level of the size-disaggregated distribution would be unrealistic in organisms that only double in size. This assumption could be changed with little effect on the results. The growth rate $G^{(z)}_i(w)$ of a single zooplankton cell of species $i$ at mass $w$ now replaces the simple input term (a) in eq. with the information on predation in eqs. , ; the metabolic loss is unchanged. This gives $$\begin{aligned} G^{(z)}_i(w) &=& \epsilon \int \sum_j w' S^{(p)}_{ij}(w,w') p_j(w') dw' \;\;\;\textrm{(gain from eating phytoplankton)}\nonumber\\ &+& \epsilon \int \sum_j w' S^{(z)}_{ij}(w,w') z_j(w') dw' \;\;\;\textrm{(gain from eating zooplankton)}\nonumber \\ &-& w^{1-\xi} \;\tilde b\left(\frac{w}{w_i}\right)^{\beta+\xi-1} \;\;\;\textrm{(loss from metabolism)}, \label{eq:gz}\end{aligned}$$ where $\epsilon$ is the efficiency with which prey mass is turned into predator mass. The gain terms have the form of a feeding rate at size $w$ on species $j$ at size $w'$, multiplied by the density of these prey and their cell mass $w'$, summed over prey species and integrated over prey cell masses, to get the total rate at which mass is consumed. A consistent allometric scaling of zooplankton growth requires that it should match the scaling of metabolism, $w^{1-\xi}$, and hence the scaling of growth in phytoplankton. There is nothing in eq. to ensure this happens. However, the scaling needed does emerge if the population densities lie on a power law, and we will show in Section \[ref:results\] below that the predator-prey dynamics can generate this structure. Such a power-law solution takes the form $$p_j(w)= w_j^{-\gamma} h(w/w_j) , \label{eq:phat}$$ where $\gamma$ is the power-law exponent. Showing the statement is true is a matter of algebra, and we provide a proof in Appendix \[ref:z\_growth\_scaling\]; see also @cuesta:16. The emergent scaling is important: it means that there is a set of densities on a power-law of the form given in eq. on which the growth rates of zooplankton species in eq. have the same allometric scaling of cell mass, $w^{1-\xi}$, as phytoplankton species. This is necessary for the triangle of observations in Fig. \[fig:triangle\] to emerge, although this does not preclude other outcomes that lack both the allometric scalings and also the power law (we will show an example in Fig. \[fig:loss\_of\_power-law\]). Note that the power-law exponent $\gamma$, expressed in terms of the other exponents $\nu$, $\xi$, $\chi$, is $$\gamma = 1+\frac{\nu+\xi}{1+\chi} . \label{eq:power-law}$$ That is all we know so far as growth is concerned; it is quite conceivable for a plankton assemblage neither to be on the power law, nor to relax onto it. A scaling for death rate ------------------------ We assume that most cell death comes from predation. With one caveat, this applies as much to zooplankton as to phytoplankton, because predators feed by cell size rather than by cell type. Cell death of prey is tied precisely to cell growth of predators because each consumption event in eqs. , is matched by a corresponding prey death. Reversing the indices so that $i$ is now a prey species of cell mass $w$, we write the rate $S^{(p)}_{ji}(w',w)$ (respectively $S^{(z)}_{ji}(w',w)$) as the rate at which a zooplankton cell of species $j$ and size $w'$ consumes a cell of species $i$ phytoplankton (respectively zooplankton) of size $w$ ($w'>w$). Then the per-capita death rate from predation is part (a) of the following equations: $$D_i^{(p)}(w) = \int \sum_j S^{(p)}_{ji}(w',w) z_j(w') dw' + d_0 w^{-\xi} \label{eq:dp}$$ $$D_i^{(z)}(w) = \underbrace{\int \sum_j S^{(z)}_{ji}(w',w) z_j(w') dw'}_{(a)} + \underbrace{d_0 w^{-\xi}}_{(b)} . \label{eq:dz}$$ In (a), the terms $S_{ji}^{(.)}$ for consumption by $j$ are multiplied by the density of these predators. The product is then summed over predator species, and integrated over predator cell masses, to get the total per-capita death rate from predation. The caveat about predation is that, if this was the only cause of death, the largest zooplankton would be free of all mortality; this is because we confine the study to unicellular organisms. Mortality-free dynamics would be unrealistic, and we therefore introduce some background intrinsic mortality with the same scaling in all species, given as part (b) of eqs. , . Predation by multicellular organisms would remove the need for this, although some background mortality is to be expected in any event. Thus the total per-capita death rate $D_i^{(p)}(w)$ of a phytoplankton cell (respectively $D_i^{(z)}(w)$ of a zooplankton cell) is given as the sum of (a) and (b) in eqs. , . As in cell growth, there is an allometric scaling of predation with respect to $w$ when the population densities lie on the power law eq. . This can be shown by the method used for cell growth (Appendix \[ref:z\_growth\_scaling\]), and in this case the scaling is $w^{-\xi}$. As noted for growth, the triangle of observations in Fig. \[fig:triangle\] would not be possible without the scaling, although outcomes without this and without the power law would be entirely feasible. Cell division rate ------------------ In addition to cell growth and death, a cell-division process is needed for the dynamics that follow. This is essentially a matter of larger cells splitting into two, but for realism, some variation in the size of daughter cells should be present immediately after cell division. In general, there are two ways of generating this. First, the rate of cell division in species $i$ can be a function $K_i^{(.)}(w)$ that becomes large as cell mass $w$ approaches $2w_i$. Secondly, there can be variation in size of the two daughter cells themselves. For a cell that divides at size $w'$, we define a function $Q(w|w')$ that describes the probability density that the daughters are of size $w$ and $w'-w$, concentrated around $w = w'/2$. The two functions $K_i^{(.)}(w)$ and $Q(w|w')$ ensure that arbitrary features of initial size distributions decay over time in the population dynamics. A dynamical system for the NPZ assemblage ----------------------------------------- So far, we have shown that an allometric scaling of cell growth and death with $w$ is present, if population densities lie on a power law of the form in eq. . However, whether this special set of population densities could actually emerge from the predator-prey interactions is a separate matter, as yet unanswered. This requires a model of multispecies population dynamics that can track the densities of cells over time. The model is built on the cell growth, death and division rates defined above, so that species are coupled through their predator-prey interactions and through their feeding on a common resource. Fig. \[fig:flows\] sketches the links involved. ![Mass flows in an assemblage of unicellular plankton living on a single resource. Heavy lines are log-scaled size distributions of cells within species $\textrm{log} p_i(\textrm{log}w)$, $\textrm{log}z_i(\textrm{log}w)$ for phytoplankton and zooplankton species respectively. Continuous arrows denote growth of cells. Dotted arrows show the source of food for cell growth; where growth comes from consuming smaller cells, this is accompanied by death of these cells. Dash-dot arrows denote cell division. []{data-label="fig:flows"}](images/fig_flows.pdf){width="12cm"} The state variables for the model are the resource concentration $N$, the size distributions of an arbitrary number ($m$) of phytoplankton species $p_1(w), \dots, p_m(w)$, and the size distributions of an arbitrary number ($n$) of zooplankton species $z_1(w), \dots, z_n(w)$. The exact form of the resource dynamics is not important, and we use a chemostat function $$\frac{dN}{dt}=c\bigg(1 - \frac{N}{\hat N}\bigg) - \underbrace{d \sum_{i=1}^m\, \int_0^{\infty} A_i(N)\, w^{\alpha}\, p_i(w)\, dw} _{\textrm{loss to phytoplankton}}. \label{eq:resource}$$ In the absence of phytoplankton, the resource settles to an equilibrium concentration $\hat N$. The term for loss of resource to phytoplankton starts with the rate of resource uptake by cells of mass $w$ in species $i$ in eq. , integrates over all cell masses of species $i$ allowing for their densities, and then sums over phytoplankton species $1, \dots, m$. The dimensionless parameter $d$ takes into account that growth is proportional to resource uptake, but not necessarily equal to it. In this way, loss rate from the nutrient pool in eq. matches the total uptake rate of resource, and couples together the dynamics of all phytoplankton species. For the phytoplankton dynamics we use a form of the size-based McKendrick$-$von Foerster equation, which is essentially a cell growth-division equation. This has a long history in cell biology [@fredrickson:67; @bell:67; @sinko:71; @diekmann:83; @Heijmans:84; @henson:03; @giometto:13], and is an appropriate model for the dynamics of unicellular phytoplankton size distributions, shown here for species $i$: $$\begin{aligned} \frac{\partial}{\partial t}p_i(w) = &-&\frac{\partial}{\partial w}\big[G^{(p)}_i(w,N)p_i(w)\big] \nonumber \\ &+&2\int_0^\infty Q(w|w')K_i^{(p)}(w')p_i(w') dw' \nonumber \\ &-&K_i^{(p)}(w)p_i(w) \nonumber \\ &-&D_i^{(p)}(w)p_i(w) , \label{eq:MvFP}\end{aligned}$$ with $G_i^{(p)}$ from eq. , and $D_i^{(p)}$ from eq. . The first term on the right-hand side is the rate at which cell growth leads to change in density at size $w$. The second term is the rate at which new cells are generated at size $w$ from cell division; the probability density $Q$ concentrates the new cells around $w'/2$. The third term is the rate at which cells disappear at size $w$ through cell division, $K_i^{(p)}$ making this term large near $2w_i$. The fourth term is the rate of cell loss from mortality. Similarly the zooplankton dynamics for species $i$ are given by $$\begin{aligned} \frac{\partial}{\partial t}z_i(w) = &-&\frac{\partial}{\partial w}\big[G^{(z)}_i(w,N)z_i(w)\big] \nonumber \\ &+&2\int_0^\infty Q(w|w')K_i^{(z)}(w')z_i(w') dw' \nonumber \\ &-&K_i^{(z)}(w)z_i(w) \nonumber \\ &-&D_i^{(z)}(w)z_i(w). \label{eq:MvFZ}\end{aligned}$$ with $G_i^{(z)}$ as in eq. , and $D_i^{(z)}$ as in eq. . Results on the NPZ equilibrium {#ref:results} ============================== Continuum model --------------- The continuum model allows some formal mathematical results to be given about the steady-state behaviour [@cuesta:16]. This model replaces the finite number of species in a bounded range of body size in eqs. , , with a continuum of species that have characteristic cell masses $w_*$ spanning the range from zero to infinity. Remarkably, it can be proved that this model has a power-law equilibrium. This is an equilibrium on which an infinite number of species coexist, notwithstanding the competition for resource among the phytoplankton and the predator-prey interactions between zooplankton and phytoplankton. The equilibrium has the form given in eq. , and the abundance of the characteristic masses $w_*$ scales with $w_*$ with the exponent given in eq. . In this case, the equations and the steady-state solution are scale invariant. The unique power-law equilibrium emerges from the dynamics as a direct consequence of the predator-prey interactions. The power-law structure of the zooplankton holds death rates from predation with a scaling exponent $-\xi$, consistent with the power law of prey abundance. At the same time, the power-law structure of the prey holds the growth rates of the zooplankton with a scaling exponent $1-\xi$, consistent with the power law of predator abundance. Whether this equilibrium could be stable without additional stabilising mechanisms is not known. Discretized model ----------------- For two reasons, it helps to go from the continuum model to a numerical analysis of the discretized, bounded system. First, the number of species is always finite in reality, and secondly there must always be lower and upper bounds on cell size in the plankton. In these circumstances, the perfect power-law equilibrium in the idealised continuum model [@cuesta:16] can never be achieved exactly, and a numerical analysis can show whether it is a useful guide to more realistic systems. In addition, a numerical stability analysis can show whether the near-power-law equilibrium is an attractor, in other words, whether plankton assemblages should actually be expected to move towards it; this has not been possible in the continuum model. The numerical analysis requires some functions and parameter values to be specified; these are given in Table \[table:functions\] and Table \[table:parameters\] in Appendix \[ref:parameter\_values\]. ![Absence of a power-law relationship between abundance and cell mass in an assemblage of 20 phytoplankton and a small number of zooplankton species. (a), (b), (c) One zooplankton species with diet breadth $\sigma = 1.0$. (d), (e), (f), Three zooplankton species with diet breadths $\sigma = 0.25$. (a), (d) Feeding kernels for zooplankton cells at characteristic cell sizes $w_i$ of each species. (b), (e) Size spectra for $w_i$ at time 600 of phytoplankton (filled circles), and zooplankton (open circles) species; the initial densities are shown as dots. (c), (f) Time series of the total densities of phytoplankton (continuous) and zooplankton (dashed) species. The time series were obtained from numerical integration of eqs. , , , with $a(N)$, $Q(w|w')$, $K_i^{(.)}(w)$ and $s(w/w')$ as in Table \[table:functions\], and parameter values as in Table \[table:parameters\] in Appendix \[ref:parameter\_values\]. []{data-label="fig:loss_of_power-law"}](images/fig_loss_of_power-law.pdf){width="14cm"} Note first that the NPZ dynamical system eqs. , , can reach an equilibrium far away from any power law (Fig. \[fig:loss\_of\_power-law\]). This is clear if, for instance, there is a just a single zooplankton species, restricting the size range over which predation takes place; Fig. \[fig:loss\_of\_power-law\]a shows the feeding kernel at the characteristic cell mass of this species. The densities of vulnerable phytoplankton species are driven down to low densities by this localised feeding of the predator, creating a corresponding ‘hole’ in the phytoplankton assemblage (Fig. \[fig:loss\_of\_power-law\]b). This leaves the zooplankton species without enough food, and it tends to zero density (Fig. \[fig:loss\_of\_power-law\]c). As the predator goes to extinction, the phytoplankton assemblage comes to rest at a point where the most vulnerable species have densities close to zero. Put another way, with only one zooplankton species, the predator-prey interaction does not allow the predation-death to settle near an allometric scaling near $-\xi$. More specialised feeding exacerbates the problem, as can be seen in Fig. \[fig:loss\_of\_power-law\]d,e,f, where there were initially three zooplankton species with narrow feeding kernels. Because the abundances are nowhere near a power law, this behaviour is not consistent with the full triangle of observations in Fig. \[fig:triangle\]. This is despite the fact that the phytoplankton species do coexist without the predator, and do have an allometric scaling of growth and death rates with respect to cell mass. The allometric scaling of death comes from the background mortality that remains in eq. in the absence of predation. The key numerical result of this paper is that a larger number of zooplankton species brings death (and cell growth) from predation approximately to the allometric scaling with cell mass needed for the power law. In effect, more zooplankton species distribute predation mortality better across the range of cell sizes. The outcome is then consistent with the full triangle of observations (Fig. \[fig:triangle\]), namely: coexistence of species, allometric scalings of ecological rates, and the power-law structure of the assemblage. Moreover, the numerical results show that the equilibrium can be locally asymptotically stable, given a sufficiently positive value of $\chi$. In other words, there is a neighbourhood of the near-power-law equilibrium within which all other initial size distributions return to the equilibrium. The equilibrium is at least a local attractor. ![Convergence to a near-power-law equilibrium of a plankton community of 20 phytoplankton species in presence of predation by 9 zooplankton species. (a), (b) Two contrasting initial conditions lead to the same equilibrium densities of $w_i$ (characteristic cell masses) of phytoplankton (filled circles), and zooplankton (open circles) species; the initial densities are shown as dots. (c) Complete size distributions within phytoplankton (continuous) and zooplankton (dashed) species at equilibrium. (d) Community spectra aggregated over species at equilibrium: size spectrum $p(\textrm{log}w)=\sum_i p_i(\textrm{log}w)$ (continuous line), biomass spectrum $b(\textrm{log}w)= \sum_i wp_i(\textrm{log}w)$ (dotted line). The dashed reference lines have a slope -1, and the dotted lines have slopes 2/7 and -5/7, corresponding to eq. after transforming from $w$ to $\textrm{log}w$. The set of phytoplankton species and the first three zooplankton species from Fig. \[fig:loss\_of\_power-law\] were used, together with seven further zooplankton species. Numerical integrations were carried out as in Fig. \[fig:loss\_of\_power-law\] with parameter values as in Table \[table:parameters\]. Integration brought the densities close to equiibrium, after which the equilibrium was obtained by Newton’s method. The Jacobian at equilibrium had a leading eigenvalue with a real part = -0.0018, implying local asymptotic stability of the equilibrium. []{data-label="fig:power-laws_PZ"}](images/fig_power-law.pdf){width="15cm"} We illustrate this result in Fig. \[fig:power-laws\_PZ\], in which we took an assemblage of 20 phytoplankton and 10 zooplankton species with random characteristic cell masses. The second largest zooplankton species tended to zero density, and the 29-species equilibrium (with the second-largest species at zero density) was locally asymptotically stable. Importantly, the equilibrium was quite close to a power law (i.e. close to linear in the double logarithmic plot); this is shown at the characteristic cell masses (Fig. \[fig:power-laws\_PZ\]a,b), and for the full size distributions of the species (Fig. \[fig:power-laws\_PZ\]c). It is notable that the two illustrative initial conditions in Fig. \[fig:power-laws\_PZ\]a,b are far from the equilibrium point, suggesting that the equilibrium has a basin of attraction substantially greater than the range of initial conditions corresponding to a local stability analysis. This ordered structure is generated by the predator-prey interaction. A slope of $2/7$ for the biomass spectrum is predicted from eq. , and the spectrum that emerged from the numerics is close to this, especially when away from the lower and upper boundaries of cell size (Fig. \[fig:power-laws\_PZ\]d). The lack of predation on species near the upper boundary would be expected to leave a relatively large footprint near the right-hand boundary, and this is indeed evident in the size distributions and overall densities of these species (Fig. \[fig:power-laws\_PZ\]c). Discussion ========== This study shows, for the first time, how predator-prey dynamics can drive unicellular plankton assemblages towards a power-law relationship between cell mass and abundance. With enough species, the near-power-law equilibrium can be an attractor, and the predator-prey interaction can bring the whole assemblage to this state. In this way we obtain a mechanism for coexistence of multiple plankton species near a power-law equilibrium, consistent with the triangle of observations in Fig. \[fig:triangle\]. The study gives some support to the conclusions of a mathematical analysis of the continuum model, in which an unlimited number of unicellular plankton species can coexist at a power-law equilibrium if cell growth predation and death all scale appropriately with cell mass [@cuesta:16]. The size-based modelling of predation adopted here improves the resolution of trophic structure in plankton assemblages, as called for by @boyce:15. It would readily extend to cover mixotrophic feeding [@flynn:13]. Also, a straightforward extension would be to move the upper bound on body mass to encompass metazoan plankton and fish, and we would expect this to remove some of the roughness at the right-hand side of the plankton spectrum in our computations. Taking this size-based approach through to larger organisms would help to provide a common formal framework for the flow of biomass more broadly through aquatic ecosystems, so-called end-to-end models [@fulton:10], eliminating the need for closure assumptions at the upper end of the plankton [e.g. @edwards:01] and at the lower end of fish assemblages [e.g. @datta:10; @hartvig:11]. The extension to more than one limiting resource would not change the qualitative picture we are presenting here. The caveat about the continuum (mathematical) model [@cuesta:16] is that, although it shows the power-law equilibrium exists, as it stands this equilibrium is not known to be an attractor. However, one further assumption can make the equilibrium stable to small displacements; this is that the dependence of predation on the density of prey is stronger than that given by the standard rule of mass-action. The effect is to make over-abundant species experience disproportionately large mortality. This is closely related to the idea of killing-the-winner in which prey (hosts) are controlled according to how fast they are increasing [@thingstad:97; @winter:10]. With this assumption in place, the exponents in the model are related by $\gamma = 1 + (\nu+\xi)/(1+\chi)$, and $-\gamma$ emerges as the exponent of the power-law relationship between abundance and characteristic cell mass. Such predation mortality could come about in various ways, for instance through the ideal-free distribution [@kacelnic:92], redistribution of predation according to prey species abundance [@armstrong:99], patchy spatial distributions of prey coupled to predator aggregation [@nachman:06], and vertical profiles of phytoplankton and zooplankton in the water column [@morosov:10]. A Holling type III functional response [@holling:59] has been suggested as a surrogate for such spatial effects [@nachman:06; @morosov:10], and was needed to generate plankton blooms in excitable-media models [@truscott:94]. The predation mortality has the effect that predators feed disproportionately on prey species with an excess of biomass, thereby reducing the size of peaks and troughs in the community biomass spectrum. This is not to discount the possible existence of other mechanisms that could stabilize the equilibrium. An important feature of the near-power-law equilibrium abundance is that it needs enough species to be present to generate allometric scalings for prey death and predator growth. The continuum (mathematical) model shows that the power-law equilibrium becomes exact as the number of species tends to infinity [@cuesta:16], and this suggests that the allometric scalings become easier to achieve as the number of prey and predator species increases. Broader diets, i.e. feeding over a greater range of prey sizes (coincidentally increasing connectance), would also help in generating the scalings. The coexistence of many species at equilibrium near the power law is not a consequence of more feeding niches with increasing species richness; there is, for instance, no lack of niche space for the few predator species in Fig. \[fig:loss\_of\_power-law\] at time 0. Neither is there a limiting similarity caused by niche overlap: species can be arbitrarily close together at the power-law equilibrium of the underlying continuum theory [@cuesta:16]. Coexistence of such similar species is consistent with recent molecular evidence of coexisting cryptic species [e.g. @amato:07; @mcmanus:09]. These features of the power-law equilibrium are unanticipated in the context of standard ecological theory. This theory does not lead to an expectation that an unlimited number of species could coexist, unconstrained by food-web connectance, and without the need for separate niches. In particular, a great deal of research on random matrices from @may:72 onwards leads to the view that ecological communities are less likely to be stable as species richness and connectance increase. We think it likely that there is structure in assemblages of aquatic plankton that random matrices have yet to take into account. Random matrices would need to generate equilibria that come close to a power-law relationship with cell size, and this probably requires appropriate allometric scalings of ecological rates. Note that the study of species-poor plankton communities gives no hint that coexistence could be readily achieved in species-rich communities. This invites an obvious and important question as to how a species-rich community close to the power-law abundance could be assembled from its component parts in the first place. That such assemblages exist is not in doubt [@gaedke:92; @quinones:03; @sanmartin:06]. However, it is notable that the biomass spectra of some lake communities have marked peaks and troughs, unlike oligotrophic ocean systems [@sprules:91; @quinones:03; @yurista:14], suggestive of an intermediate state in assembling plankton communities. Assembly is an interesting matter that needs further research. Clearly, many real-world processes prevent the existence of perfect power-law size spectra and exact allometric scalings of ecological rates. External processes include seasonal fluctuations that leave a strong footprint in community size spectra [@heath:95]. Exploitation of fish stocks causes major changes in size spectra further up the food web [@blanchard:05], which may be felt lower down through trophic coupling in the ecosystem. Internal processes include species-specific features of plankton life histories that a simple generic model cannot incorporate. For instance, the characteristic cell mass of diatoms decreases as they go through cycles of asexual reproduction; some taxonomic specificity in feeding has been observed [@jezbera:06] and there is evidence that unicellular taxonomic groups have different preferred predator-prey size ratios [@hansen:94]. The allometric scaling relationships may also change; for instance @maranon:13 found a positive relationship between intrinsic rate of increase and cell size below cell diameters $\simeq 5$ $\mu$m, and there is some change in the scaling of optimal prey size from small to large planktonic predators [@wirtz:12]. Bearing in mind these caveats, the strongest statement this paper can support is that unicellular plankton species are more balanced in their interactions when close to the power law, than when further away from it. This means that the forces driving exclusion of one species by another are likely to be weaker than has previously been thought. Correspondingly, the forces needed to counteract exclusion and to maintain species-rich plankton assemblages do not have to be so strong to work effectively. A corollary is that this balanced state of coexistence can be disrupted by processes, human or otherwise, that drive aquatic ecosystems further from the power law. Acknowledgements ================ RL and GWD were supported by EU Grant 634495 — MINOUW — H2020-SFS-2014-2015. JAC was supported by the Spanish mobility grant PRX12/00124 and project FIS2015-64349-P (MINECO/FEDER, UE). We thank A. D. Dean, J. W. Fox, J. Kolding, E. Mara[ñó]{}n, E. J. A. Minter, M. J. Plank and S. Vae for helpful discussions on this work. Appendices {#appendices .unnumbered} ========== Allometric scaling and power-law abundances {#ref:w_wi_scaling} =========================================== Allometric scaling of physiological rates with cell size plays an essential role in this paper. There is however a potential for confusion as to whether the allometric scaling depends on the cell size $w$ of an individual or on the characteristic cell size $w_i$ of a species. We clarify this by acknowledging that rates can depend on both $w$ and $w_i$, and that allometric scaling refers to the behaviour of the rate under a rescaling of both $w$ and $w_i$. If $R_i(w)=R(w,w_i)$ denotes some physiological rate for species $i$, then allometric scaling with an exponent $\rho$ means that the rate $R(w,w_i)$ satisfies $$R(\lambda w,\lambda w_i) = \lambda^\rho R(w,w_i)$$ for any positive scale factor $\lambda$. In other words, $R(w,w_i)$ is a homogeneous function of both arguments. This in turn means that the rate can be written in two alternative scaling forms $$R_i(w)=w_i^\rho \,r(w/w_i) =w^\rho\, \tilde{r}(w/w_i)$$ where $\tilde{r}(x)=x^{-\rho} r(x)$. Thus it does not matter whether one thinks of allometric scaling in terms of cell size $w$ or in terms of characteristic size $w_i$. The exponent will be the same in either case and only the scaling function of $w/w_i$ will differ. A similar issue also arises when abundance power laws are discussed: are these power laws in cell size $w$, or in the characteristic cell size of a species $w_i$? Again it helps to view the abundance as a function of both $w$ and $w_i$, as we do in this paper. Saying that the abundance has a power-law form then means that it is a homogeneous function of both arguments and thus can be written equally well in terms of a power of $w$ or a power of $w_i$. For example in the case of the steady state phytoplankton abundance it is equally valid to write $p_i(w) = w_i^{-\gamma-1}f_p(w/w_i)$ as it is to write $p_i(w) = w^{-\gamma-1}\tilde{f}_p(w/w_i)$. Allometric scaling of zooplankton growth {#ref:z_growth_scaling} ======================================== Here we show that there is an allometric scaling of zooplankton growth $G_i^{(z)}(w)$ of the form $w^{1-\xi}$ in eq. , if a multispecies plankton assemblage is at a power-law equilibrium. The argument begins by assuming a power-law equilibrium with densities $p_j(w_j)$ (respectively $z_j(w_j)$) at $w_j$ that scale with characteristic cell mass as $w_j^{-\gamma}$, where $\gamma$ is an unknown exponent. Also assume that the equilibrium densities at sizes other than $w_j$ can be written as $p_j(w)= w_j^{-\gamma} h(w/w_j).$ Make a change in variable $v=w'/w_j$ in the predation terms of eq. , giving $w'=w_jv$ and $dw'=w_jdv$; this allows $w_j^2$ to be factored out of the predation integrals. Substituting eq. into , allows a further $w_j^{-\gamma}$ to be factored out. The same factorisations apply to $z_j(w')$. Using eq. gives $P_j^\chi = w_j^{(1-\gamma)\chi}(\int h(v) dv)^\chi$; when substituted into eq. , this allows a further $w_j^{(1-\gamma)\chi}$ to be factored out. The same factorisation applies to $Z_j^\chi$. Lastly, $w^\nu$ is factored out of the predation integrals. Noting that $w_i^\zeta$ can always be written in the form $w^\zeta (w/w_i)^{-\zeta}$, means that the exponents can be collected together as powers of $w$, namely $2-\gamma+(1-\gamma)\chi+\nu$. This function of exponents can now be equated with the allometric scaling of metabolism $1-\xi$ so that the $1-\xi$ scaling applies throughout $G_i^{(z)}$. Parameter values {#ref:parameter_values} ================ The characteristic cell mass $w_i$ of species $i$ is defined as half its maximum cell mass. The discrete set of $w_i$s replace the continuum of species with characteristic cell mass $w_*$ in @cuesta:16. In phytoplankton, values for $w_i$ were chosen in the range $2 \rightarrow 30000$ pg, corresponding to upper limits on cell ESDs of $\sim 2 \rightarrow 40 \mu$m. Those of zooplankton were chosen in the range $200 \rightarrow 3 \times 10^6$ pg, corresponding to upper ESD limits of $\sim 10 \rightarrow 200 \mu$m. Staggering the cell size ranges in this way ensured that all phytoplankton were vulnerable to predation, and all zooplankton had food to eat. From a compilation of earlier studies, @tang:95 [Fig. 2] estimated algal cell division rate to scale with an exponent of about $-0.15$ with respect to cell size. Recently, it has been noted that cell division rates reach a maximum at a cell diameter $\simeq 5$ $\mu$m [@chen:10; @maranon:13]; the change is particularly clear in the transition from bacteria to eukaryotes [@kempes:12]. @maranon:13 gave an exponent for the rate of increase of cells on the right-hand side of the peak of approximately $ -0.15$, corresponding to a cell doubling time scaling as $\xi = 0.15$, the value used here. Metabolism was assumed to scale isometrically with cell mass $\beta = 1$, in keeping with observations on protists in @delong:10 and phytoplankton [@lopez:14]. The rate of gain in cell mass was assumed to scale with an exponent $\alpha = 0.85$. With the values chosen, the structure of eq. is simple, as $\alpha+\xi-1 = 0$. Note that $\alpha$ must be less than $\beta$ to ensure that metabolic loss eventually becomes greater than the gain in mass, as cells grow. [l l l p[6.5cm]{}]{} Name &Eq. &Function &Comments\ \ $ \tilde a(N) $ & & $\frac{\tilde a_\infty N}{r+N}$ & functional response for resource uptake\ \ $ K_i^{(.)}(w)$ & , & $G_i^{(.)}(w)\delta(w-2w_i)$ & cell division rate set to zero, except at $2w_i$ ($\delta$ is the Dirac $\delta$-function)\ \ $ Q(w|2w_i) $ & , & $\frac{1}{\sqrt{2\pi}\sigma_bw_i} \exp\big(-\frac{(w/w_i-1)^2}{2\sigma_b^2}\big)$ & distribution of cell masses following division at $2w_i$\ \ $ s(w/w')$ & , &$\frac{1}{\sigma \sqrt{2\pi}} \exp\big(-\frac{(\ln (w/w')-\Delta)^2}{2\sigma^2}\big)$ & feeding kernel\ \[table:functions\] [l l l l l l p[6.0cm]{}]{} Symbol & Value & Dimensions & Comments\ \ $ \xi $ & 0.15 &– & exponent scaling doubling time with cell mass\ \ &&&*Metabolic loss*\ $ \beta $ & 1 &– & exponent scaling metabolism with cell mass\ $ \tilde b $ & 0.5 &T$^{-1}$ & metabolism term\ \ &&&*Resource uptake*\ $ \alpha$ & 0.85 &– & exponent scaling resource uptake with cell mass\ $ \tilde a_\infty$ & 2 &T$^{-1}$ & resource uptake parameter\ \ &&&*Intrinsic death rate*\ $ d_0 $ & 0.1 &T$^{-1}$ & intrinsic death rate parameter\ \ &&&*Cell division*\ $\sigma_b$ & 0.05 & – & size range of daughter cells\ \ &&&*Resource dynamics*\ $ c $ & 100 & M V$^{-1}$ T$^{-1}$ & resource growth rate\ $ \hat N $ & 100 & V$^{-1}$ & resource equilibrium without plankton\ $ d $ & 1 & – & proportionality constant for resource uptake to cell growth\ $ r $ & 1 & V$^{-1}$ & type II functional response term\ \ &&&*Predation*\ $ A $ & 0.02 & V$^{-1}$ T$^{-1}$ & encounter parameter\ $ \nu $ & 0.85 & – & exponent volume sensed\ $ \Delta $ & 5 & – & log preferred predator prey mass ratio\ $ \sigma $ & 1, 0.25& – & diet breadth\ $ \epsilon$ & 0.6 & – & food conversion efficiency\ $ \chi $ & 0.4 & – & exponent for density dependence\ \ &&&*Numerics*\ $ \delta x $ & 0.025 & – & size step (logarithmic binning)\ $ \delta t $ & 0.0005 & T & time step\ $ t_{max} $ & 600 & T & time period for integration\ \[table:parameters\] We set $\tilde a_{\infty}$ in the function $\tilde a(N)$ (Table \[table:functions\]) to be substantially greater than $\tilde b$ (2, 0.5 respectively) so that phytoplankton cells would not shrink when the resource level was below its equilibrium level. It was still possible for there to be insufficient resource for phytoplankton cells to grow, in numerical integrations starting at high cell densities. The same could apply to zooplankton cells if the food available was too low. To deal with such cases, negative growth rates were replaced by zero, causing cells to stay at the same size, experiencing the level of mortality corresponding to that size. However, at the equilibrium point itself, the growth rate is positive in every species with a positive density. This is intuitive. If it was not so, cells would not get as far as cell division and the species concerned would be decreasing in density, not at equilibrium. The intrinsic death parameter $d_0$ was set at 0.1, so that most mortality would be caused by predation. Some residual mortality is needed for the largest zooplankton species [*in lieu*]{} of larger predators; for compararability, all species were given the same background value. The parameter $\sigma_b$ spreads out the size of daughter cells. We investigated the effects of this and chose $\sigma_b = 0.05$ as ensuring enough variation for effects of initial size structure within species to attenuate quickly. Parameter $A$ describes the overall level of encounters of zooplankton with their prey, and can be used to set the zooplankton abundance relative to the phytoplankton. We chose a value $A=0.02$ as bringing their abundance to within about one order of magnitude. Encounters scale with body mass with an exponent $\nu$; @delong:12 recorded 95 % confidence intervals $\approx 0.7 - 1.5$ for the exponent in a meta-analysis of protist-phytoplankton microcosms. A predator-prey mass ratio was obtained from the relationship between ESD of predators and prey taken from @wirtz:12. For unicellular plankton this is ESD$_P$ = 0.16 ESD$_Z$. Scaling this from diameter to volume and assuming neutral buoyancy gives a rounded value $\Delta = 5$. @wirtz:12 showed non-linearities at large zooplankton size (his Fig 4). Since our work was concerned with unicellular organisms, we worked with the left-hand end where the nonlinearity is small [see eq. 8 @wirtz:12]. Here it is $\bar r$ that matters, estimated as $\bar r = 0.16$ [@wirtz:12 page 5]. We varied the width of the feeding kernel to tune the connectance of the community. The value $\sigma = 1$ used here allows a high degree of connectedness, at the same time as ensuring prey cannot be large enough for cannibalism to take place. A food conversion efficiency $K = 0.6$ was assumed, following @hartvig:11. The parameter $\chi = 0.4$ describing the form of density-dependent predation was chosen by trial and error, as a small value that was often able to stabilise plankton assemblages. All computations were carried with log transformed cell mass ($x=\textrm{ln}w$); this is because of large range of cell masses involved ($\sim$ 1 to 10$^6$). Computations were started by choosing species $w_i$s from uniformly-distributed random numbers over a range of approximately $\textrm{ln}2$ to $\textrm{ln}30000$ for phytoplankton, and $\textrm{ln}300$ to $\textrm{ln}(4\times10^6)$ for zooplankton. The random $w_i$s were rounded to match the discretization of cell size used in the numerical integration. Eqs. , were discretized into steps of width $\delta x$ as given in Table \[table:parameters\]. Integration of the full system was by the standard Euler method, with step sizes as small as computationally feasible. To obtain the equilibrium point, integration was carried out up to time $t_{max}$ to get close to equilibrium. The equilibrium was found by the Newton-Raphson method; this uses the Jacobian matrix from which the leading eigenvalue at the equilibrium point was taken as the measure of local asymptotic stability.

ArXiv

--- abstract: | In the presence of the primordial magnetic field, initial vector (vorticity) perturbations produce cosmological Alfvén waves and leave imprints on cosmic microwave background (CMB) temperature and polarization anisotropy. We have investigated imprints of cosmological Alfvén waves in CMB anisotropy. For data constraints, we have used the power spectrum of the recent CMB observations, and correlations estimated from WMAP Internal Linear Combination (ILC) maps. Our analysis shows 3$\sigma$ evidence of cosmological Alfvén waves. Using the 3$\sigma$ limit from our analysis and the Alfvén velocity limit from the total energy density constraint, we impose a lower bound on the amplitude of primordial vector perturbation: $4\times 10^{-12}$ at $k_0=0.002/\mathrm{Mpc}$. address: '$^1$ Niels Bohr Institute, Blegdamsvej 17, DK-2100 Copenhagen, Denmark' author: - Jaiseung Kim$^1$ and Pavel Naselsky$^1$ bibliography: - '/home/tac/jkim/Documents/bibliography.bib' title: 'Cosmological Alfvén waves in the recent CMB data, and the observational bound on the primordial vector perturbation' --- Introduction ============ There are strong observational evidences on the existence of large-scale magnetic fields, whose strength is on the order of micro Gauss [@Magnetic_Fields_Lyman; @Magnetic_Fields_Galaxy; @Magnetic_Fields_Origin; @Magnetic_Fields_Faraday]. These magnetic fields are believed to be seeded by a small Primordial Magnetic Field (PMF) [@PMF_Cosmic_transition; @PMF_Electroweak; @PMF_QCD; @PMF_quark]. Primordial Magnetic Field may be considered to consist of a homogeneous and inhomogeneous (stochastic) magnetic fields. In this paper, we focus on the effect of Primordial Magnetic Field on CMB anisotropy associated with vector perturbations. Primordial vector perturbation, whose source includes topological defects and inhomogeneous primordial magnetic fields [@cosmic_defect_Vector; @cosmic_defect_CMB; @SPMF_CMB], decays rapidly with the expansion of the Universe, and therefore, does not leave observable signature on CMB anisotropy [@Foundations_Cosmology; @Cosmology]. However, primordial vector perturbation in the presence of a homogeneous primordial magnetic field induces Alfvén waves in the cosmic plasma, which may leave observable imprints on CMB anisotropy via a Doppler effect and an integrated Sachs-Wolfe effect [@PMF_acoustic_CMB; @inhomogeneous_Alfven; @pmf_small; @pmf_pol; @DKY; @Alfven_nl]. To be specific, cosmological Alfvén waves create correlations between $a_{lm}$, $a_{l\pm\Delta l,m\pm\Delta m}$, where $\Delta l=0, 1, 2$ and $\Delta m=0,1,2$ [@KLR]. Over the past several years, there have been great success in measurement of Cosmic Microwave Background (CMB) anisotropy by ground and satellite observations [@WMAP5:basic_result; @WMAP5:powerspectra; @WMAP5:parameter; @ACBAR; @ACBAR2008; @QUaD1; @QUaD2; @QUaD:instrument], and observational imprints of cosmological Alfvén waves in CMB anisotropy have been studied by several authors [@DKY; @KLR; @Chen; @Naselsky:PMF; @PMF_anomaly]. By investigating the effect on power spectrum and signature correlations in the recent CMB data, we attempt to constrain cosmological Alfvén waves, and also impose a lower bound on the primordial vector perturbation. The outline of this paper is as follows. In Section \[cmb\_pmf\], we briefly review the effect of cosmological Alfvén waves on CMB anisotropy. In Section \[correlation\], we discuss statistical properties of CMB in the presence of the Alfvén waves. In Section \[estimators\], we build the statistics sensitive to the imprints of Alfvén waves. In Section \[analysis\], we analyze the recent CMB data and present the result. In Section \[Discussion\], we make summary and conclusion. In \[leakage\], we discuss the effect of incomplete sky coverage on the analysis of Alfvén wave imprints. The effect of Alfvén waves on CMB anisotropy {#cmb_pmf} ============================================ According to linearized Einstein equations [@Spacetime], vector perturbation decays rapidly with the expansion of the Universe, and therefore, does not leave observable signature on CMB anisotropy, unless the initial values of vector perturbation are unusually high [@Foundations_Cosmology; @Cosmology]. In the presence of a homogeneous magnetic field, vector perturbation induces Alfvén waves in cosmic plasma, and leaves observable imprints on CMB anisotropy via a Doppler effect and an integrated Sachs-Wolfe effect [@DKY]. Durrer, Kahniashvili and Yates [@DKY] (hereafter, DKY) showed that cosmological Alfvén waves generate the fractional CMB anisotropy for a Fourier mode $\mathbf{k}$: $$\begin{aligned} \frac{\delta T}{T_0}(\hat {\mathbf{n}},\mathbf{k}) \approx \hat {\mathbf{n}}\cdot \mathbf{\Omega}(\mathbf k,\eta_{dec}) = \hat {\mathbf{n}}\cdot \mathbf{\Omega_0}\; v_A k\,\eta_{dec}\;\hat {\mathbf{B}}\cdot \hat {\mathbf{k}}, \end{aligned}$$ where $\hat {\mathbf{n}}$ and $\hat {\mathbf{B}}$ denote sky direction and a homogeneous magnetic field direction respectively, $\mathbf{\Omega}(\mathbf k,\eta_{dec})$ is the Gauge invariant linear combination associated with vector perturbations, $\eta_{dec}$ denotes the conformal time at the moment of baryon-photon decoupling, $v_A$ is Alfvén wave velocity and $T_0$ is the CMB monopole temperature $2.725 [\mathrm{K}]$ [@Fixen:dipole]. DKY assumed that vector perturbations are initially created by some random process and have the following statistical properties over an ensemble of universes: $$\begin{aligned} \langle \Omega^i_0(\mathbf k)\,\Omega^j_0(\mathbf k)\rangle =(\delta_{ij}-\hat k_i\,\hat k_j) P(k),\end{aligned}$$ $P(k)$, which is the power spectrum, is assumed to follow a simple power law: $$\begin{aligned} P(k)=A_v \frac{k^n}{k^{n+3}_0},\label{P} \end{aligned}$$ where $k_0$ is a pivot wavenumber and we set it to $0.002/\mathrm{Mpc}$, which is equal to the WMAP team’s pivot wavenumber for scalar perturbation [@WMAP3:parameter]. Possible sources of primordial vector perturbation include inhomogeneous stochastic primordial magnetic fields and topological defects [@cosmic_defect_Vector; @cosmic_defect_CMB; @SPMF_CMB]. CMB statistical properties {#correlation} ========================== The temperature anisotropy $\delta T(\theta,\phi)$ is conveniently decomposed in terms of spherical harmonics $Y_{lm}(\theta,\phi)$ : $$\begin{aligned} \delta T(\theta,\phi)=\sum_{lm} a_{lm}\,Y_{lm}(\theta,\phi),\end{aligned}$$ where $a_{lm}$ are the coefficients of decomposition, and $\theta$ and $\phi$ are a polar and an azimuthal angle. Kahniashvili, Lavrelashvili and Ratra [@KLR] (hereafter KLR) showed that the CMB anisotropy in the presence of cosmological Alfvén waves has the following statistical properties: $$\begin{aligned} \lefteqn{\langle a^*_{lm} a_{lm}\rangle = C^{0}_l+(3\cos^2\theta_B-1)\frac{l(l+1)}{(2l-1)(2l+3)}}\nonumber\\ &\times&\;\left\{\frac{l(l+1)+(l^2+l-3)\cos^2\theta_B}{3\cos^2\theta_B-1} - m^2\left[1-\frac{3}{l(l+1)}\right]\right\} I^{l,l}_d\,\label{correlation0} \end{aligned}$$ $$\begin{aligned} \langle a^*_{lm} a_{l,m+1} \rangle &=& -\sin2\theta_B \exp[-i\phi_B] I^{l,l}_d\,d^{m,m+1}_{ll},\label{correlation1}\end{aligned}$$ $$\begin{aligned} \langle a^*_{lm} a_{l,m+2} \rangle &=& -\frac{1}{2}\sin^2\theta_B \exp[-2i\phi_B] I^{l,l}_d\,d^{m,m+2}_{ll}, \label{correlation2}\end{aligned}$$ $$\begin{aligned} \langle a^*_{l,m} a_{l+2,m} \rangle&=&-(3\cos^2\theta_B-1) I^{l,l+2}_d\,d^{m,m}_{l,l+2}, \label{correlation3}\end{aligned}$$ $$\begin{aligned} \langle a^*_{l,m} a_{l+2,m+1}\rangle&=&\sin2\theta_B\exp[-i\phi_B] I^{l,l+2}_d\,d^{m,m+1}_{l,l+2}, \label{correlation4}\end{aligned}$$ $$\begin{aligned} \langle a^*_{l,m} a_{l+2,m-1}\rangle&=&\sin2\theta_B\exp[i\phi_B] I^{l,l+2}_d\,d^{m,m-1}_{l,l+2}, \label{correlation5}\end{aligned}$$ $$\begin{aligned} \langle a^*_{l,m} a_{l+2,m+2}\rangle&=&-\frac{1}{2}\sin^2\theta_B\exp[-2i\phi_B] I^{l,l+2}_d\,d^{m,m+2}_{l,l+2}, \label{correlation6}\end{aligned}$$ $$\begin{aligned} \langle a^*_{l,m} a_{l+2,m-2}\rangle&=&-\frac{1}{2}\sin^2\theta_B\exp[2i\phi_B] I^{l,l+2}_d\,d^{m,m-2}_{l,l+2},\label{correlation7}\end{aligned}$$ where $C^{0}_l$ is the power spectrum in the absence of Alfvén waves, $\theta_B$ and $\phi_B$ are the spherical coordinate of a PMF direction $\hat {\mathbf B}$, and $d^{m,m'}_{ll'}$ $(|m|\le l,\;|m'|\le l')$ are $$\begin{aligned} d^{m,m+1}_{l,l} &=&\frac{l^2+l-3}{(2l-1)(2l+3)} (m+\frac{1}{2})\sqrt{(l-m)(l+m+1)},\\ d^{m,m+2}_{l,l} &=& \frac{l^2+l-3}{(2l-1)(2l+3)}\sqrt{(l-m)(l-m-1)(l+m+1)(l+m+2)},\\ d^{m,m}_{l,l+2} &=&\frac{(l+3)l}{2(2l+3)\sqrt{(2l+1)(2l+5)}}\sqrt{((l+1)^2-m^2)(l-m+2)(l+m+2)},\\ d^{m,m+1}_{l,l+2} &=&\frac{(l+3)l}{2(2l+3)\sqrt{(2l+1)(2l+5)}} \sqrt{T((l+1)^2-m^2)(l+m+2)(l+m+3)},\\ d^{m,m-1}_{l,l+2}&=&\frac{(l+3)l}{2(2l+3)\sqrt{(2l+1)(2l+5)}}\sqrt{((l+1)^2-m^2)(l-m+2)(l-m+3)},\\ d^{m,m+2}_{l,l+2} &=& \frac{(l+3)l}{2(2l+3)\sqrt{(2l+1)(2l+5)}}\,\sqrt{\frac{(l+m+4)!}{(l+m)!}},\\ d^{m,m-2}_{l,l+2} &=&\frac{(l+3)l}{2(2l+3)\sqrt{(2l+1)(2l+5)}}\,\sqrt{\frac{(l-m+4)!}{(l-m)!}},\end{aligned}$$ and $$\begin{aligned} I^{ll'}_d&=&\frac{2\,T^2_0}{\pi}\int \,k^2 \left(P_0 \frac{k^{n_v}}{k^{n_v+3}_0}\right)\exp\left(-2\frac{k^2}{k^2_D}\right) v^2_A\,\left(\frac{\eta_{dec}}{\eta_0}\right)^2 j_l(k\eta_0)\,j_l'(k\eta_0)\,dk,\nonumber\\ &=&\frac{2\,T^2_0}{\pi}\frac{P_0\,v^2_A}{(k_0\eta_0)^{n_v+3}} \left(\frac{\eta_{dec}}{\eta_0}\right)^2 \int \,x^{n_v+2}\exp\left(-\frac{2 x^2}{(k_D\eta_0)^2}\right) j_l(x)j_l'(x)\,dx.\nonumber\\\label{Ill}\end{aligned}$$ $\eta_0$ is the present conformal time, and $v_A$ is the Alfvén velocity, which is given by $$\begin{aligned} v^2_A=\frac{B^2_0}{4\pi(\rho_r+p_r)},\;\;\;\;v_A\sim 4\times10^{-4}\frac{B_0}{10^{-9}\mathrm{Gauss}},\label{v_A}\end{aligned}$$ where $B_0$ is the magnitude of a homogeneous PMF, and $\rho_r$ and $p_r$ are the density and the pressure of photons [@DKY]. $k_D$ denotes the comoving wavenumber of the dissipation scale, due to photon viscosity and given by $\sim 10/\eta_{dec}$ [@DKY]. While DKY and KLR have obtained analytic results by neglecting the damping factor $\exp\left(-2k^2/k^2_D\right)$, we have computed Eq. \[Ill\] numerically. The damping effect is getting significant on multipoles $l\gtrsim 500$ [@DKY]. Primordial magnetic fields affect the expansion dynamics of the Universe, because of anisotropic pressures associated with primordial magnetic fields [@PMF_Constraint]. Considering the shear anisotropy due to primordial magnetic fields and the observed magnitude of CMB anisotropy ($\sim 10^{-5}$), Barrow et al. (1997) have derived an upper limit $B_0<2.27\times10^{-9}\,h/75$ Gauss on a homogeneous primordial magnetic field [@PMF_Constraint]. Using this upper limit and Eq. \[v\_A\], DKY showed $B_0\lesssim 10^{-8}\,\mathrm{Gauss}$ and $v_A\lesssim 10^{-3}$ [@DKY]. It is worth to note that correlations are invariant under the parity inversion of a PMF direction $\hat {\mathbf B}$. In other words, $(\theta_B,\phi_B)\rightarrow (\pi-\theta_B,\phi_B+\pi)$) does not affect Eq. \[correlation1\], \[correlation2\], \[correlation3\], \[correlation4\], \[correlation5\], \[correlation6\] and \[correlation7\]. note that it is the degeneracy of correlation, not that there should be two physical directions. Estimators ========== In order to constrain cosmological Alfvén waves, we construct the following statistics: $$\begin{aligned} C_l&=&(2l+1)^{-1}\sum_{m} a^*_{lm} a_{lm},\label{D0}\\ D^1_l&=&l^{-1}\sum_{m\ge 0} a^*_{lm} a_{l,m+1},\label{D1}\\ D^2_l&=&(2l-1)^{-1}\sum_{m} a^*_{lm} a_{l,m+2},\label{D2}\\ D^3_l&=&(2l+1)^{-1}\sum_{m} a^*_{l,m}\,a_{l+2,m},\label{D3}\\ D^4_l&=&(2l)^{-1}\sum_{m} a^*_{l,m}\,a_{l+2,m+1},\label{D4}\\ D^5_l&=&(2l-1)^{-1}\sum_{m} a^*_{l,m}\,a_{l+2,m+2}.\label{D5}\end{aligned}$$ Because of the reality condition $a_{l\,-m}=(-1)^m\,{a_{lm}}^*$, $\sum\limits_{m=-2l} a^*_{lm} a_{l,m+1}$ is always zero. Therefore, we defined $D^1_l$ to sum only terms of $m\ge 0$. There also exist the following indentities because of the reality condition: $$\begin{aligned} \sum_{m} a^*_{l,m}\,a_{l+2,m+1}=-\left[\sum_{m} a^*_{l,m}\,a_{l+2,m-1}\right]^*,\end{aligned}$$ $$\begin{aligned} \sum_{m} a^*_{l,m}\,a_{l+2,m+2}=\left[\sum_{m} a^*_{l,m}\,a_{l+2,m-2}\right]^*.\end{aligned}$$ Therefore, $\sum_{m} a^*_{l,m}\,a_{l+2,m-1}$ and $\sum_{m} a^*_{l,m}\,a_{l+2,m-2}$ are redundant and excluded from consideration. Note that the statistic used in [@DKY; @Chen] is equivalent to Eq. \[D3\]. Using Eq. \[correlation0\], \[correlation1\], \[correlation2\], \[correlation3\], \[correlation4\] and \[correlation6\], we may easily show that the expectation values of $C_l$ and $D^i_l$ are $$\begin{aligned} \bar C_l &=& C^0_l+\frac{l(l+1)}{3} I^{l,l}_d, \label{bar_D0}\\ \bar D^1_l &=& -\sin 2\theta_B \exp[-i\phi_B]\,\frac{I^{l,l}_d}{l}\sum_{m\ge 0} d^{m,m+1}_{ll},\label{bar_D1}\\ \bar D^2_l &=& -\frac{1}{2}\sin^2\theta_B \exp[-2i\phi_B]\,\frac{I^{l,l}_d}{2l-1}\sum_{m} d^{m,m+2}_{ll}, \label{bar_D2}\\ \bar D^3_l &=&-(3\cos^2\theta_B-1)\,\frac{I^{l,l+2}_d}{2l+1}\sum_{m} d^{m,m}_{l,l+2},\label{bar_D3}\\ \bar D^4_l &=&\sin 2\theta_B\exp[-i\phi_B]\,\frac{I^{l,l+2}_d}{2l}\sum_{m} d^{m,m+1}_{l,l+2}, \label{bar_D4}\\ \bar D^5_l &=&-\frac{1}{2}\sin^2\theta_B\exp[-2i\phi_B]\,\frac{I^{l,l+2}_d}{2l-1}\sum_{m}d^{m,m+2}_{l,l+2}. \label{bar_D5}\end{aligned}$$ We may also show that the variance of $D^i_l$ is $$\begin{aligned} \mathrm{Var}(C_l)&\approx& 2/(2l+1)(C^0_l+N_l)^2,\label{Var0}\\ \mathrm{Var}(D^1_l)&\approx& l^{-1}(C^0_l+N_l)^2,\label{Var1}\\ \mathrm{Var}(D^2_l)&\approx& (2l-1)^{-1}(C^0_l+N_l)^2,\label{Var2}\\ \mathrm{Var}(D^3_l)&\approx& (2l+1)^{-1}(C^0_l+N_l)(C^0_{l+2}+N_{l+2}),\label{Var3}\\ \mathrm{Var}(D^4_l)&\approx& (2l)^{-1}(C^0_l+N_l)(C^0_{l+2}+N_{l+2}),\label{Var4}\\ \mathrm{Var}(D^5_l)&\approx& (2l-1)^{-1}(C^0_l+N_l)(C^0_{l+2}+N_{l+2}),\label{Var5}\end{aligned}$$ where $N_l$ is noise power spectrum, and we have neglected correlations between distinct spherical harmonic modes (i.e. $\langle a^*_{lm} a_{l'm'}\rangle \approx C^0_l\,\delta_{ll'}\delta_{mm'}$). Though the underlying distribution for primordial vector perturbations is not necessarily Gaussian, the distribution function of $D^i_l-\bar D^i_l$ tends to Gaussian by the central limit theorem [@Math_methods]. Hence, the likelihood function of $\bar D^i_l$, given the data $D^i_l$, is $$\begin{aligned} \mathcal{L}(\bar D^i_l(\lambda_{\alpha})|D^i_l)=\frac{1}{(2\pi)^{\frac{N}{2}}{\left|\mathbf M\right|}^{\frac{1}{2}}}\exp[-\frac{1}{2}(\mathbf D-\mathbf {\bar D}(\lambda_{\alpha}))^{\dagger}\,{\mathbf M}^{-1}\,(\mathbf D-\mathbf {\bar D}(\lambda_{\alpha}))].\nonumber\\\label{likelihood}\end{aligned}$$ $\mathbf M$ is a $5\,l_\mathrm{max}\times 5\,l_\mathrm{max}$ nearly diagonal matrix, whose diagonal elements correspond to the variances of $D^1_l,\ldots,D^5_l$: $\mathbf M \approx \mathrm{diag} \left(\mathrm{Var}(D^1_{l_\mathrm{min}}),\,\ldots,\,\mathrm{Var}(D^5_{l_\mathrm{max}})\right)$. $N$ is the data vector size (i.e. $5\,l_\mathrm{max}$), and $\lambda_\alpha$ are the parameters associated with Alfvén waves : $\lambda_\alpha \in \{A_v v^2_A, n_v, \theta_B,\phi_B\}$. For low $l$, the deviation of $D^i_l-\bar D^i_l$ from Gaussian distribution is getting non-negligible. However, it does not affect the best-fit values, but only the error bars, since $\mathbf M$ is nearly diagonal, and thereofore the best-fit parameters are, in fact, determined by minimizing $\chi^2$. Since $\langle (C_l-\bar C_l) \,(D^i_l-\bar D^i_l)\rangle$ is negligible in comparison to the variance of $C_l$ or $D^i_l$, we may construct a full likelihood function as follows: $$\begin{aligned} \mathcal{L}(\bar C_l|C_l) \times \mathcal{L}(\bar D^i_l(\lambda_{\alpha})|D^i_l) \label{full_likelihood},\end{aligned}$$ where $\mathcal{L}(\bar C_l|C_l)$ is the likelihood function associated with CMB power spectra, whose approximate and optimal expression is found in [@WMAP1:parameter_method]. Note that cosmological Alfvén waves affect CMB power spectrum as well, and hence Eq. \[bar\_D0\] should be used for $\bar C_l$ in estimating $\mathcal{L}(\bar C_l|C_l)$. Since CMB power spectra and correlations (i.e. $\bar C_l$ and $\bar D^i_l$) possess non-trivial dependence on cosmological parameters (i.e. standard parameters + Alfvén wave parameters), we will resort to Markov Chains Montel Carlo (MCMC) likelihood analysis [@WMAP1:parameter_method; @CosmoMC]. The discussion on MCMC analysis and dataset used will be give in the next section. Analysis of the recent CMB observation data {#analysis} =========================================== We have investigated the effect of the WMAP team’s KQ75 and KQ85 mask on correlations (i.e. $D^i_l$ ), and found that significant amount of fictitious correlations are produced by the masks (refer to \[leakage\] for details). Therefore, we find sky maps, which require a Galactic mask, are not suitable for our analysis. Therefore, we have restricted our correlation estimation to whole sky maps obtained by variants of Internal Linear Combination (ILC) method: The WMAP team’s ILC map (WILC), Harmonic ILC (HILC) and Needlet ILC (NILC) [@WMAP3:temperature; @WMAP5:foreground; @HILCT; @HILCP; @NILC]. ![the power spectra of ILC maps and the WMAP concordance model[]{data-label="Cl"}](./figure/fig1.eps) In Fig. \[Cl\], we show the temperature power spectra of the ILC maps with that of the WMAP concordance model [@WMAP3:parameter; @WMAP5:parameter]. Noting that the power spectra of ILC maps are in good agreement with the model up to $l\le 200$, we have computed $D^i_{l}$ up to $l\le 200$. We have also made the analysis with $D^i_{l\le 250}$ for HILC, and $D^i_{l\le 250}$ and $D^i_{l\le 300}$ respectively for NILC. We have obtained the consistent results with the $D^i_{l\le 200}$ analysis. The confidence intervals are smaller, but the best-fit values turn out to be similar. From now on, all the results are from the $D^i_{l\le 200}$ analysis. parameter WILC HILC NILC --------------------- --------------------------------- ----------------------------------- --------------------------------- $\Omega_{b}\,h^2$ $0.023^{+0.002}_{-0.002}$ $0.024^{+0.002}_{-0.003}$ $0.024^{+0.002}_{-0.003}$ $\Omega_{c}\,h^2$ $0.109^{+0.023}_{-0.024}$ $0.104^{+0.03}_{-0.018}$ $0.108^{+0.025}_{-0.024}$ $\tau$ $0.089^{+0.085}_{-0.04}$ $0.09^{+0.082}_{-0.043}$ $0.089^{+0.08}_{-0.04}$ $n_s$ $1.015^{+0.262}_{-0.139}$ $1.028^{+0.242}_{-0.123}$ $1.018^{+0.251}_{-0.13}$ ${d\,n_s}/{d\ln k}$ $-0.008^{+0.068}_{-0.095}$ $-0.008^{+0.052}_{-0.095}$ $-0.009^{+0.066}_{-0.108}$ $\log[10^{10}A_s]$ $3.031^{+0.17}_{-0.27}$ $2.975^{+0.228}_{-0.234}$ $3.026^{+0.188}_{-0.321}$ $A_{\mathrm{sz}}$ $1.63^{+0.37}_{-1.63}$ $1.87^{+0.133}_{-1.867}$ $1.57^{+0.43}_{-1.57}$ $H_0$ \[km/s/Mpc\] $74.37^{+13.65}_{-10.58}$ $76.62^{+10.19}_{-13.02}$ $74.46^{+13.55}_{-10.98}$ $10^9 A_v\,v^2_A$ $3.13^{+5.66}_{-2.42}$ $4.16^{+4.99}_{-3.27}$ $3.87^{+5.28}_{-3.11}$ $n_v$ $-4.35^{+0.58}_{-0.7}$ $-4.45^{+0.6}_{-0.71}$ $-4.49^{+0.73}_{-0.7}$ $\theta^+_B$ ${15.7^\circ}^{+25}_{-15.7}$ ${1.6^\circ}^{+31.7}_{-1.6}$ ${11.7^\circ}^{+27.7}_{-11.7}$ $\phi^+_B$ ${112^\circ}^{+247.9}_{-111.8}$ ${113.2^\circ}^{+246.6}_{-113.1}$ ${119^\circ}^{+240.9}_{-119}$ $\theta^-_B$ ${164.3^\circ}^{+15.7}_{-25}$ ${178.4^\circ}^{+1.6}_{-31.7}$ ${168.3^\circ}^{+11.7}_{-27.7}$ $\phi^-_B$ ${292^\circ}^{+247.9}_{-111.8}$ ${293.2^\circ}^{+246.6}_{-113.1}$ ${299^\circ}^{+240.9}_{-119}$ : cosmological parameters of $\Lambda$CDM + ($A_v v^2_A$, $n_v$, $\theta_B$, $\phi_B$) constrained by power spectra (WMAP5YR + ACBAR + QUaD) + correlation (ILC) \[parameters\] By making small modifications to the `CosmoMC` package [@CosmoMC], we have included the parameter $\lambda_\alpha \in \{A_v v^2_A, n_v, \theta_B, \phi_B\}$, and computed the likelihood function given by Eq. \[full\_likelihood\]. Note that the power spectrum as well as correlations depend on $A_v v^2_A$ and $n_v$ (see Eq. \[bar\_D0\]). For data constraint, we have used the CMB power spectra of the recent CMB observations (WMAP5YR + ACBAR + QUaD) [@WMAP5:basic_result; @WMAP5:powerspectra; @ACBAR; @ACBAR2008; @QUaD1; @QUaD2; @QUaD:instrument], and the correlations estimated from ILC maps. In Table \[parameters\], we summarize 1$\sigma$ constraint and the best-fit values. As discussed in Sec. \[estimators\], our estimators are insensitive to the parity of a PMF direction. Hence, we quote two best-fit PMF directions, which are equally likely. In Fig. \[likelihood1\], we show the marginalized likelihood (solid lines) and mean likelihood (dotted lines) of $\{A_v\,v^2_A, n_v, \theta_B, \phi_B\}$. ![Likelihood of Alfvén wave parameters: WILC, HILC and NILC (from the left to the right), normalized to its peak.[]{data-label="likelihood1"}](./figure/fig2a.eps "fig:") ![Likelihood of Alfvén wave parameters: WILC, HILC and NILC (from the left to the right), normalized to its peak.[]{data-label="likelihood1"}](./figure/fig2b.eps "fig:") ![Likelihood of Alfvén wave parameters: WILC, HILC and NILC (from the left to the right), normalized to its peak.[]{data-label="likelihood1"}](./figure/fig2c.eps "fig:") In Fig. \[likelihood2\], we show likelihood in the plane of $A_v\,v^2_A$ versus $n_v$, which are highly correlated Alfvén wave parameters. We find there exists little correlation between Alfvén wave parameters and $\Lambda$CDM parameters, except for the scalar perturbation amplitude $A_s$. In Fig. \[likelihood2\], we show marginalized likelihood in the plane of $A_v\,v^2_A$ versus $\log[10^{10} A_s]$. ![Marginalized likelihood in the plane of $A_v\,v^2_A$ vs $\{n_v,\;\log[10^{10}A_s]\,\}$: WILC, HILC and NILC (from left to right), solid curves denote 1$\sigma$ and 2$\sigma$ contours.[]{data-label="likelihood2"}](./figure/fig3a.eps "fig:") ![Marginalized likelihood in the plane of $A_v\,v^2_A$ vs $\{n_v,\;\log[10^{10}A_s]\,\}$: WILC, HILC and NILC (from left to right), solid curves denote 1$\sigma$ and 2$\sigma$ contours.[]{data-label="likelihood2"}](./figure/fig3b.eps "fig:") ![Marginalized likelihood in the plane of $A_v\,v^2_A$ vs $\{n_v,\;\log[10^{10}A_s]\,\}$: WILC, HILC and NILC (from left to right), solid curves denote 1$\sigma$ and 2$\sigma$ contours.[]{data-label="likelihood2"}](./figure/fig3c.eps "fig:") Marginalized likelihood of PMF direction is shown in Fig. \[L\_direction\]. Note that a pair of directions associated by $(\theta_B,\phi_B)\leftrightarrow (\pi-\theta_B,\phi_B+\pi)$ possess equal likelihood. ![Marginalized likelihood of PMF direction: WILC, HILC, and NILC (from left to right), normalized to its peak[]{data-label="L_direction"}](./figure/fig4a.eps "fig:") ![Marginalized likelihood of PMF direction: WILC, HILC, and NILC (from left to right), normalized to its peak[]{data-label="L_direction"}](./figure/fig4b.eps "fig:") ![Marginalized likelihood of PMF direction: WILC, HILC, and NILC (from left to right), normalized to its peak[]{data-label="L_direction"}](./figure/fig4c.eps "fig:") As shown in Table \[parameters\] and Fig. \[likelihood1\], \[likelihood2\], \[L\_direction\], The result from various ILC maps are similar with each other. Since each ILC map contains distinct residual foregrounds and noise, consistency among the analysis with various ILC maps indicates that our results are affected insignificantly by residual foregrounds and noise in ILC maps. For good exploration on tails of the parameter $A_v v^2_A$, we have also set `CosmoMC` to check convergence for confidence intervals of $A_v v^2_A$. In Table \[3sigma\], we show 3$\sigma$ constraints on $A_v v^2_A$. Using our 3$\sigma$ limit on $A_v\,v^2_A$ and the upper limit on $v_A$ (refer to Sec. 3), we find $5\times 10^{-4}\lesssim A_v$ (refer to Eq. \[P\] for the definition of $A_v$). Taking into account Eq. \[P\], we find the amplitude of primordial vector perturbation is equal to $A_v k^3_0$. Hence, we impose the lower bound $4\times 10^{-12}$ on the amplitude of primordial vector perturbation. Power Spectra + ILC 3$\sigma$ confidence interval --------------------- ----------------------------------------------------- WILC $5.8\times 10^{-10}<A_v\,v^2_A<1.01\times 10^{-8}$ HILC $5.73\times 10^{-10}<A_v\,v^2_A<0.99\times 10^{-8}$ NILC $5.05\times 10^{-10}<A_v\,v^2_A<1.09\times 10^{-8}$ : $A_v v^2_A$ estimation at 3$\sigma$ confidence level[]{data-label="3sigma"} If there was a primordial magnetic field at the recombination epoch or after the Universe was re-ionized, it would induce Faraday rotation in CMB polarization [@PMF_B; @PMF_Faraday_obs]. Forecast on the PLANCK data constraint shows that we will be able to constrain a primordial magnetic field $\sim 2\times 10^{-9}\;\mathrm{Gauss}$ by investigating Faraday rotation effect in polarization data [@PMF_Faraday_obs]. However, we have not considered Faraday rotation effect in this work, since the signal-to-noise ratio in the currently available CMB polarization data is not high. Discussion {#Discussion} ========== We have constrained primordial magnetic field and primordial vector perturbation, by investigating the imprints of cosmological Alfvén waves on CMB anisotropy. We find there is little degeneracy between cosmological Alfvén wave parameters and $\Lambda$CDM parameters except for a scalar perturbation amplitude $A_s$. The results obtained with various ILC maps are consistent with each other. Therefore, we believe our results have not been affected significantly by residual foregrounds or anisotropic noise in ILC maps. Using our result and the upper bound on $v_A$ from the total energy density constraint, we impose a lower bound on the primordial vector perturbation $5\times 10^{-4}\lesssim A_v$ at the pivot scale $k_0=0.002/\mathrm{Mpc}$. Since Alfvén velocity $v_A$ is proportional to the magnitude of a primordial magnetic field, direct constraints on Alfvén velocity $v_A$ may be imposed by investigating Faraday rotation effect in CMB polarization data [@PMF_B; @PMF_Faraday_obs]. When the data from the PLANCK surveyor [@Planck_bluebook] are available, we are going to constrain Alfvén waves and Faraday rotation simultaneously, and hence impose a stronger constraint on the primordial vector perturbation. We thank anonymous referees for helpful comments, which lead to sharpening this work. We thank T. Kahniashvili and G. Lavrelashvili for useful discussions. We also thank J. Delabrouille and his colleagues for making their whole-sky CMB map (NILC) publically available. We acknowledge the use of the Legacy Archive for Background Data Analysis (LAMBDA) and ACBAR, and QUaD data. Our data analysis made the use of HEALPix [@HEALPix:Primer; @HEALPix:framework] and the `CosmoMC` package. This work is supported by FNU grant 272-06-0417, 272-07-0528 and 21-04-0355. Reference ========= incomplete sky coverage {#leakage} ======================= Incomplete sky coverage produces fictitious correlations by destroying the orthogonality of spherical harmonics. Spherical harmonic coefficients from partial sky coverage is related to the true ones as follows: $a_{lm} =\sum W_{ll'mm'} a^{\mathrm{true}}_{l'm'},$ where $ W_{ll'mm'}=\int d^2\hat{\mathbf{n}}\,W(\hat{\mathbf{n}}) Y^*_{lm}(\hat{\mathbf{n}})\,Y_{l'm'}(\hat{\mathbf{n}})$, and $W(\hat{\mathbf{n}})$ is zero in the masked region and one elsewhere. To estimate the amount of leakage, we assume CMB is purely Gaussian and has zero correlation: $$\begin{aligned} \langle (a^{\mathrm{true}}_{lm})^* a^{\mathrm{true}}_{l',m'}\rangle = C_l\,\delta_{ll'} \delta_{mm'}.\end{aligned}$$ We find the expectation value of $D^3_l$ of the pure Gaussian CMB is $$\begin{aligned} \lefteqn{\langle D^3_l\rangle=\frac{1}{2l+1}\sum_m\langle a^*_{l,m} a_{l+2,m}\rangle}\label{D_leakage}\\ &=&\frac{1}{2l+1}\sum_m\sum_{l'm'} C_{l'} W^*_{ll'mm'} W_{l+2,l'mm'}\nonumber\\ &\approx& \frac{C_{l+1} }{2l+1}\sum_{m}\sum_{l'm'}\int d^2\hat{\mathbf{n}}\,W^*(\hat{\mathbf{n}}) Y_{lm}(\hat{\mathbf{n}})\,Y^*_{l'm'}(\hat{\mathbf{n}})\int d^2\mathbf{\hat{n}'}\,W(\mathbf{\hat{n}'}) Y^*_{l+2\,m}(\mathbf{\hat{n}'})\,Y_{l'm'}({\mathbf{\hat{n}'}})\nonumber\\ &=&\frac{C_{l+1} }{2l+1} \sum_{m}\int d^2\hat{\mathbf{n}}\,W^*(\hat{\mathbf{n}})Y_{lm}(\hat{\mathbf{n}})\int d^2\mathbf{\hat{n}'} W(\mathbf{\hat{n}'}) \,Y^*_{l+2\,m}(\mathbf{\hat{n}'})\delta(\hat{\mathbf{n}}-\mathbf{\hat{n}'})\nonumber\\ &=&\frac{C_{l+1} }{2l+1} \sum_{m} \int d^2\hat{\mathbf{n}}\,|W(\hat{\mathbf{n}})|^2 Y_{lm}(\hat{\mathbf{n}}) Y^*_{l+2\,m}(\hat{\mathbf{n}}).\nonumber\end{aligned}$$ In the third equality, we have taken $C_{l'}$ out of the summation and equate it to $C_{l+1}$, since $W^*_{ll'mm'} W_{l+2,l'mm'}$ peaks sharply around $l'=l+1$, while $C_{l'}$ varies much slowly in comparison to $W^*_{ll'mm'} W_{l+2,l'mm'}$. In the fourth equality, we have used the identity $\sum_{l'm'} Y_{l'm'}(\hat{\mathbf{n}}) Y_{l'm'}(\mathbf{\hat{n}'}) =\delta(\hat{\mathbf{n}}-\mathbf{\hat{n}'})$. From Eq. \[D\_leakage\], we see the leakage from Gaussian CMB to $D^3_l$ can be estimated by: $$\begin{aligned} \frac{1}{2l+1}\sum_{m}\int d^2\hat{\mathbf{n}}\,|W(\hat{\mathbf{n}})|^2 Y_{lm}(\hat{\mathbf{n}}) Y^*_{l+2,m}(\hat{\mathbf{n}}).\label{Cl_leakage}\end{aligned}$$ The value as high as unity indicates that most of Gaussian CMB power is leaked into the fictitious correlation, while a zero value indicates no leakage. $l$ 10 20 100 200 ------ -------- -------- -------- -------- KQ75 $0.18$ $0.18$ $0.17$ $0.17$ KQ85 $0.12$ $0.12$ $0.12$ $0.12$ : D3 Leakage[]{data-label="D3_leakage"} We have estimated $D^i_l$ leakages for the WMAP team’s KQ75 and KQ85 mask and show them in Table \[D3\_leakage\]. We find $D^3_l$ leakage is as high as $\sim 20\%$ and other leakgages are $\sim 2\%$, which are all greater than the expected correlations, given Alfvén wave parameters in Table \[parameters\]. Therefore, we find CMB maps, which requires the KQ75 or KQ 85 mask, are not suitable for estimation of correlations induced by cosmological Alfvén waves.

ArXiv

Introduction. Macroscopic models for multilane traffic dynamics {#sec:intro} =============================================================== The study of traffic flow formed actually a novel branch of physics since the pioneering works by Lighthill and Whitham [@LW55], Richards [@R56], and, then, by Prigogine and Herman [@PH71]. It is singled out by the fact that in spite of [*motivated*]{}, i.e. a non-physical individual behavior of moving vehicles (they make up a so-called ensemble of “self-driven particles”, see, e.g., [@a1; @a2; @a3]), traffic flow exhibits a wide class of critical and self-organization phenomena met in physical systems (for a review see [@KL1; @KL2; @KL3]). Besides, the methods of statistical physics turn out to be a useful basis for the theoretical description of traffic dynamics [@H97]. The existence of a new basic phase in vehicle flow on multilane highways called the synchronized motion was recently discovered by Kerner and Rehborn [@KR2], impacting significantly the physics of traffics as a whole. In particular, it turns out that the spontaneous formation of moving jams on highways proceeds mainly through a sequence of two transitions: “free flow $\rightarrow$ synchronized motion $\rightarrow$ stop-and-go pattern” [@K98]. All the three traffic modes are phase states, meaning their ability to persist individually for a long time. Besides, the two transitions exhibit hysteresis [@K98; @KR1; @KR3], i.e., for example, the transition from the free flow to the synchronized mode occurs at a higher density and lower velocity than the inverse one. As follows from the experimental data [@KR2; @KR1; @KR3] the phase transition “free flow $\leftrightarrow$ synchronized mode” is essentially a multilane effect. Recently Kerner [@KL1; @KL2; @KL3] assumed it to be caused by “Z”-like form of the overtaking probability depending on the car density. The synchronized mode is characterized by substantial correlations in the car motion along different lanes because of the lane changing maneuvers. So, to describe such phenomena a multilane traffic theory is required. There have been proposed several macroscopic models dealing with multilane traffic flow and based on the gas-kinetic theory [@Ha97; @HG97; @HT98; @HT99; @HS99; @Siam], a compressible fluid model [@L97] generalizing the approach by Kerner and Konhäuser [@Km1; @Km2], and actually a model [@T99a; @T99b] dealing with the time dependent Ginzburg-Landau equation. All these models describe traffic flow in terms of the car density $\rho$, mean velocity $v$, and, may be, the velocity variance $\theta$ or ascribe these quantities to vehicle flow at each lane $\alpha$ individually. In other words, the quantities $\{\rho, v,\theta\}_{\alpha}$ are regarded as a complete system of the traffic flow state variables and if they are fixed then all the vehicle flow characteristics should be determined. The given models relate the self-organization phenomena actually to the vehicle flow instability caused by the delay in the driver response to changes in the motion of the nearest cars ahead. In fact, let us briefly consider their simplified version (cf. [@Wbook; @GTF5H]) which, nevertheless, catches the basic features taken into account: $$\begin{aligned} \frac{\partial \rho }{\partial t}+\frac{\partial (\rho v)}{\partial x} &=&0\,, \label{intro:1}\\ \frac{\partial v}{\partial t}+v\frac{\partial v}{\partial x} &=&-\frac{1}{ \rho }\frac{\partial {\mathcal{P}}}{\partial x}+\frac{1}{\tau ^{\prime }}({\mathcal{U}}-v)\,. \label{intro:2}\end{aligned}$$ Here the former term on the right-hand side of Eq. (\[intro:2\]), a so-called pressure term, reflects dispersion effects due to the finite velocity variance $\theta $ of the vehicles and the latter one describes the relaxation of the current velocity within the time $\tau ^{\prime }$ to a certain equilibrium value ${\mathcal{U}}\{\rho ,\theta \}$. In particular, for $${\mathcal{P}}\{\rho ,v,\theta \}=\rho \theta -\eta \frac{\partial v} {\partial x}\,, \label{nolabel}$$ where $\eta $ is a “viscosity” coefficient and the velocity variance $\theta $ is treated as a constant, we are confronted with the Kerner-Konhäuser model [@Km1; @Km2]. In the present form the relaxation time $\tau ^{\prime }$ characterizes the acceleration capability of the mean vehicle as well as the delay in the driver control over the headway (see, e.g., [@hw2; @HtoK; @hw1]). The value of the relaxation time is typically estimated as $\tau ^{\prime }\sim 30\,$s because it is mainly determined by the mean time of vehicle acceleration which is a slower process than the vehicle deceleration or the driver reaction to changes in the headway. The equilibrium velocity ${\mathcal{U}}\{\rho \}$ (here the fixed velocity variance $\theta $ is not directly shown in the argument list) is chosen by drivers keeping in mind the safety, the readiness for risk, and the legal traffic regulations. For homogeneous traffic flow the equilibrium velocity ${\mathcal{U}}\{\rho \}=\vartheta (\rho )$ is regarded as a certain phenomenological function meeting the conditions: $$\frac{d\vartheta (\rho )}{d\rho }<0\quad \text{and}\quad \rho \vartheta (\rho )\rightarrow 0\quad \text{as}\quad \rho \rightarrow \rho _{0}\,, \label{intro:3}$$ where $\rho _{0}$ is the upper limit of the vehicle density on the road. Since the drivers anticipate risky maneuvers of distant cars also, the dependence ${\mathcal{U}}\{\rho \}$ is nonlocal. In particular, it is reasonable to assume that the driver behavior is mainly governed by the car density $\rho $ at a certain distant “interaction point” $x_{a}=x+L^{\ast } $ rather than at the current one $x$, which gives rise to a new term in Eq. (\[intro:2\]) basing on the gas-kinetic theory [@HG97; @HT98]. Here, for the sake of simplicity following [@Wbook], we take this effect into account expanding $\rho (x+L^{\ast })$ and, then, ${\mathcal{U}}\{\rho \}$ into the Taylor series and write: $${\mathcal{U}}\{\rho \}=\vartheta (\rho )-v_{0}\frac{L^{\ast}}{\rho }\frac{\partial \rho }{\partial x}\,, \label{intro:4a}$$ where $v_{0}$ is a certain characteristic velocity of the vehicles. Then linearizing the obtained system of equations with respect to the small perturbations $\delta \rho ,\delta v\propto \exp (\gamma t+ikx)$ we obtain that the long-wave instability will occur if (cf. [@Km1; @Km2; @Wbook]) $$\label{intro:5} \tau ^{\prime }\left( \rho \vartheta _{\rho }^{\prime }\right) ^{2}>v_{0}L^{\ast }+\tau ^{\prime }\theta \,,$$ in the long-wave limit the instability increment $\mbox{Re}\,\gamma $ depends on $k$ as $$\mbox{Re}\,\gamma =k^{2}\left[ \tau ^{\prime }(\rho \vartheta _{\rho }^{\prime })^{2}-(v_{0}L^{\ast }+\tau ^{\prime }\theta )\right] \,, \label{intro:6}$$ and the upper boundary $k_{\text{max}}$ of the instability region in the $k$-space is given by the expression: $$k_{\text{max}}^{2}=\frac{\rho }{\tau \eta }\biggl\{ \biggl[ \frac{\tau ^{\prime }(\rho \vartheta_{\rho }^{\prime })^{2}}{(v_{0}L^{\ast }+\tau ^{\prime }\theta )}\biggr] ^{1/2}-1\biggr\}$$ (here $\vartheta _{\rho }^{\prime }=d\vartheta (\rho )/d\rho $). As follows from (\[intro:5\]) the instability can occur when the delay time $\tau ^{\prime }$ exceeds a certain critical value $\tau _{c}$ and for $\tau ^{\prime }<\tau _{c}$ the homogeneous traffic flow is stable at least with respect to small perturbations. Moreover, the instability increment attains its maximum at $k\sim k_{\text{max}}$, so, special conditions are required for a wide vehicle cluster to form [@Km1; @Km2; @Km3; @Km4]. In particular, in the formal limit $\tau ^{\prime }\rightarrow 0$ from Eq. (\[intro:2\]) we get $$v=\vartheta (\rho )-\frac{D}{\rho }\frac{\partial \rho }{\partial x}\,, \label{intro:7}$$ where $D=v_{0}L^{\ast }$ plays the role of the diffusion coefficient of vehicle perturbations. The substitution of (\[intro:7\]) into (\[intro:1\]) yields the Burgers equation $$\frac{\partial \rho }{\partial t}+\frac{\partial \left[ \rho \vartheta (\rho )\right] }{\partial x}=D\frac{\partial ^{2}\rho }{\partial x^{2}}\,,$$ which describes the vehicle flux stable, at least, with respect to small perturbations in the vehicle density $\rho $. However, the recent experimental data [@KR2; @K98; @KR1; @KR3; @Last] about traffic flow on German highways have demonstrated that the characteristics of the multilane vehicle motion are more complex (for a review see [@KL1; @KL2; @KL3]). In particular, there are actually three types of synchronized mode, the totally homogeneous state, homogeneous-in-speed and totally heterogeneous ones [@KR2]. Especially the homogeneous-in-speed state demonstrates the fact that, in contrast to the free flow, there is no direct relationship between the density and flux of vehicles in the synchronized mode, because their variations are totally uncorrelated [@KR2]. For example, an increase in the vehicle density can be accompanied by either an increase or decrease in the vehicle flux, with the car velocity being practically constant. As a result, the synchronized mode corresponds to a two-dimensional region in the flow-density plane ($j\rho$-plane) rather than to a certain line $j=\vartheta(\rho)\rho$ [@KR2]. Keeping in mind a hypothesis by Kerner [@KL2; @KL3; @KL2a] about the metastability of each particular state in this synchronized mode region it is natural to assume that there should be at least one additional state variable affecting the vehicle flux. The other important feature of the synchronized mode is the key role of some cars bunched together and traveling much faster than the typical ones, which enables us to regard them as a special car group [@KR2]. Therefore, in the synchronized mode the function of car distribution in the velocity space should have two maxima and we will call such fast car groups platoons in speed. Anomalous properties of the synchronized mode have been substantiated also in [@Last] using single-car-data. In particular, as the car density comes to the critical value $\rho_c$ of the transition “free flow $\leftrightarrow$ synchronized mode” the time-headway distribution exhibits a short time peak (at 0.8 sec.). This short time-headway corresponds to “...platoons of some vehicles traveling very fast–their drivers are taking the risk of driving “bumper-to-bumper” with a rather high speed. These platoons are the reason for occurrence of high-flow states in free traffic” [@Last]. The platoons are metastable and their destruction gives rise to the congested vehicle motion [@Kplat]. In the synchronized mode the weight of the short time-headways is less, however, almost every fourth driver falls below the 1-sec-threshold. In the vicinity of the transition “free flow $\leftrightarrow$ synchronized mode” the short time-headways have the greatest weight. In other words, at least near the given phase transition the traffic flow state is to be characterized by two different driver groups which separate from each other in the [*velocity*]{} space and, consequently, in multilane traffic flow there should be another relaxation process distinct from one taken into account by the model (\[intro:1\]), (\[intro:2\]). In order to move faster than the statistically averaged car a driver should permanently manoeuvring pass by the cars moving ahead. Meeting of several such “fast” drivers seems to cause the platoon formation. Obviously, to drive in such a manner requires additional efforts, so, each driver needs a certain time $\tau$ to get the decision whether or not to take part in these manoeuvres. Exactly the time $\tau$ characterizes the relaxation processes in the platoon evolution. It should be noted that the overtaking manoeuvres are not caused by the control over the headway distance and, thus, the corresponding transient processes may be much slower then the driver response to variations in the headway to prevent possible traffic accidents. The analysis of the obtained optimal-velocity function $V(\Delta x)$ demonstrates its dependence not only on the headway $\Delta x$ but also on the local car density. So, in congested flow the drivers supervises the vehicle arrangement or, at least, try to do this in a sufficiently large neighborhood covering several lanes. Another unexpected fact is that the synchronized mode is mainly distinctive not due to the car velocities at different lanes being equal. In the observed traffic flow various lanes did not exhibit a substantial difference in the car velocity even in the free flow. In agreement with the results obtained by Kerner [@KR2] the synchronized mode is singled out by small correlations between fluctuations in the car flow, velocity and density. There is only a strong correlation between the velocities at different lanes taken at the same time, however, it decreases sufficiently fast as the time difference increases. By contrast, there are strong long-time correlations between the flow and density for the free flow as well as the stop-and-go mode. In these phases the vehicle flow directly depends on the density. Thereby, the free flow, the synchronized mode and the jammed motion seem to be qualitatively distinct from one another at the microscopic level. So, it is likely that to describe macroscopically traffic phase transitions the set of the state variables $\{\rho, v,\theta\}_{\alpha}$ should be completed with an additional parameter (or parameters) reflecting the [*internal*]{} correlations in the car dynamics. In other words, this parameter has to be due to the “many-body” effects in the car interaction in contrast to such [*external*]{} variables as the mean car density and velocity being actually the zeroth and first moments of the “one-particle” distribution function. Thus, it can be regarded as an independent state variable of traffic flow. The derivation of macroscopic traffic equations based on a Boltzmann like kinetic approach [@W98] has also shown that there is an additional internal degrees of freedom in the vehicle dynamics. In any case a theory of unstable traffic flow has to answer, in particular, to a question of why its two phases, e.g., the free flow and the synchronized mode, can coexist and, thus, what is the difference between them as well as why the separating transition region (Fig. \[F1\]) does not widen but keeps a certain thickness. Besides, it should specify the velocity $u$ of this region depending on the traffic phase characteristics. There is a general expression relating the transition region velocity $u$ to the density and mean velocity of cars in the free flow and a developed car cluster: $\rho _{f}$, $v_{f}$, and $\rho _{cl}$, $v_{cl}$, respectively, that follows from the vehicle conservation [@LW55], namely, the Lighthill–Whitham formula: $$u=\frac{\rho _{cl}v_{cl}-\rho _{f}v_{f}}{\rho _{cl}-\rho _{f}}\,. \label{intro:4}$$ A specific model is to give additional relationships between the quantities $u$, $\rho _{f}$, $v_{f}$, and $\rho _{cl}$, $v_{cl}$ resulting from particular details of the car interaction. We note that a description similar to Eqs. (\[intro:1\]), (\[intro:2\]) dealing solely with the external parameters $\{\rho,v\}$ do not actually make a distinction between the free-flow and congested phases and their coexistence is due to the particular details of the car interaction. The transition “free flow $\leftrightarrow$ synchronized motion” is rather similar to aggregation phenomena in physical systems such as undercooled liquid when in a metastable phase (undercooled liquid) the transition to a new ordered (crystalline) phase goes through the formation of small clusters. Keeping in mind this analogy Mahnke [*et al.*]{} [@MP97; @MK99; @KM00] have proposed a kinetic approach based on stochastic master equation describing the synchronized mode formation that deals with individual free cars and their clusters. The cluster evolution is governed by the attachment and evaporation of the individual cars and the synchronized mode is regarded as the motion of a large cluster. To describe such phenomena in physical systems it was developed an effective macroscopic approach, called the Landau phase transition theory [@Landau], that introduces a certain order parameter $h$ characterizing the correlations, e.g., in the atom arrangement. In the present paper following practically the spirit of the Landau theory we develop a phenomenological approach to the description of the traffic flow instability that ascribes to the vehicle flux an additional [*internal*]{} parameter which will be also called the order parameter $h$ and allows for the effect of lane changing on the vehicle motion. In this way the free flow and the congested phases become in fact distinctive and solely the conditions of their coexistence as well the dynamics of the transition layer are the subject of specific models. Order parameter and the individual driver behavior {#sec:2} ================================================== We describe the vehicle flow on a multilane highway in terms of its characteristics averaged over the road cross-section, namely, by the car density $\rho $, the mean velocity $v$, and the order parameter $h$. The latter is the measure of the correlations in the car motion or, what is equivalent, of the car arrangement regularity forming due to the lane change by the “fast” drivers. Let us discuss the physical meaning of the order parameter $h$ in detail considering the free flow, synchronized mode and jammed traffic individually (Fig. \[FHDef\]). Physical meaning of the order parameter $h$ and its governing equation ---------------------------------------------------------------------- When vehicles move on a multilane highway without changing the lanes they interact practically with the nearest neighbors ahead only and, so, there should be no internal correlations in the vehicle flow at different lanes. Therefore, although under this condition the traffic flow can exhibit complex behaviour, for example, the “stop-and-go” waves can develop, it is actually of one-dimensional nature. In particular, the drivers that would prefer to move faster than then the statistically mean driver will bunch up forming the platoons headed by a relatively slower vehicle. When the cars begin to change lanes for overtaking slow vehicles the car ensembles at different lanes will affect one another. The case of this interaction is due to that a car during a lane change manoeuvre occupies, in a certain sense, two lanes simultaneously, affecting the cars moving behind it at both the lanes. Figure \[FHDef\]$b$ illustrates this interaction for cars 1 and 2 through car 4 changing the lanes. The drivers of both cars 1 and 2 have to regard car 4 as the nearest neighbor and, so, their motion will be correlated during the given manoeuvre and after it during the relaxation time $\tau'$. In the same way car 1 is affected by car 3 because the motion of car 4 directly depends on the behavior of car 3. The more frequently lane changing is performed, the more correlated traffic flow on a multilane highway. Therefore, it is reasonable to introduce the order parameter $h$ being the mean density of such car triplets normalized to its maximum possible for the given highway and to regard it as a measure of the mulitlane correlations in the vehicle flow. On the other hand the order parameter $h$ introduced in this way can be regarded as a measure of the vehicle arrangement regularity. Let us consider this question in detail for the free flow, synchronized mode, and jammed traffic individually. In the free flow the feasibility of overtaking makes the vehicle arrangement more regular because of platoon dissipation. So as the order parameter $h$ grows the free traffic becomes more regular. Nevertheless, in this case the density of the car mulitlane triplets remains relatively low, $h\ll 1$, and the vehicle ensembles should exhibit weak correlations. Whence it follows also that the mean car velocity $\vartheta $ is an increasing function of the order parameter $h$ in the free flow. In the jammed motion (Fig. \[FHDef\]$c$) leaving current lanes is hampered because of lack of room for the manoeuvres. So the car ensembles at different lanes can be mutually independent in spite of individual complex behavior. In the given case the order parameter must be small too, $h\ll 1$, but, in contrast, the car mean velocity should be a decreasing function of $h $. In fact, for highly dense traffic any lane change of a car requires practically that the neighbor drivers decelerating give place for this manoeuvres. Figure \[FHPhT\] illustrates the transition “free flow $\leftrightarrow $ synchronized mode”. As the car density grows in free flow, the “fast” drivers that at first overtake slow vehicles individually begin to gather into platoons headed by more “slow” cars among them but, nevertheless, moving faster than the statistically mean vehicle (Fig. \[FHPhT\]$a$). The platoons are formed by drivers preferring to move as fast as possible keeping short headways without lane changing. Such a state of the traffic flow should be sufficiently inhomogeneous and the vehicle headway distribution has to contain a short headway spike as observed experimentally in [@Last]. Therefore, even at a sufficiently high car density the free flow should be characterized by weak multilane correlations and not too great values of the order parameter $h_{f}$. The structure of these platoons is also inhomogeneous, they comprise cars whose drivers would prefer to move at different headways (for a fixed velocity) under comfortable conditions, i.e., when the cars moving behind a given car do not jam it or none of the vehicles moving on the neighboring lanes hinders its motion at the given velocity provided it changes the current lane. So, when the density of vehicles attains sufficiently high values and their mean velocity decreases remarkably with respect to the velocity on the empty highway some of the “fast” drivers can decide that there is no reason to move so slowly at such short headways requiring strain. Then they can either overtake the car heading the current platoon by changing lanes individually or leave the platoon and take vacant places (Fig. \[FHPhT\]$a$). The former has to increase the multilane correlations and, in part, to decrease the mean vehicle velocity because the other drivers should give place for this manoeuvres in a sufficiently dense traffic flow. The latter also will decrease the mean vehicle velocity because these places were vacant from the standpoint of sufficiently “fast” drivers only but not from the point of view of the statistically mean ones preferring to keep longer headways in comparison with the platoon headways. Therefore, the statistically mean drivers have to decelerate, decreasing the mean vehicle velocity. The two manoeuvre types make the traffic flow more homogeneous dissipating the platoons and smoothing the headway distribution (Fig. \[FHPhT\]$b$ and the low fragment). Besides, the single-vehicle experimental data [@Last] show that the synchronized mode is singled out by long-distant correlations in the vehicle velocities, whereas the headway fluctuations are correlated only on small scales, which justifies the assumptions of the synchronized mode being a more homogeneous state than the free flow. We think that the given scenario describes the synchronized mode formation which must be characterized by a great value of the order parameter, $h_{s}>h_{f}$, and a lower velocity in comparison with the free flow at the same vehicle density. In addition, whence it follows that, first, the left boundary of the headway distribution should be approximately the same for both the free flow and the synchronized mode near the phase transition, which corresponds the experimental data [@Last]. Second, since in this case the transition from the free flow to the synchronized mode leads to the decrease in the mean velocity, the “fast” driver will see no reason to alter their behaviour and to move forming platoons again until the vehicle density decreases and the mean velocity grows enough. It is reasonable to relate this characteristics to the experimentally observed hysteresis in the transition “free flow $\leftrightarrow $ synchronized mode” [@K98; @KR1; @KR3]. Third, for a car to be able to leave a given platoon the local vehicle arrangement at the neighboring lane should be of special form and when an event of the vehicle rearrangement occurs its following evolution depends also on the particular details of the neighboring car configuration exhibiting substantial fluctuations. Therefore, the synchronized mode can comprise a great amount of local metastable states and correspond to a certain two-dimensional region on the flow-density plane ($j\rho $-plane) rather than a line $j=\vartheta (\rho )\rho$, which matches the experimental data [@KR2] and the modern notion of the synchronized mode nature [@KL1; @KL2; @KL3]. This feature seems to be similar to that met in physical media with local order, for example, in glasses where phase transitions are characterized by wide range of controlling parameters (temperature, pressure, [*etc.*]{}) rather than their fixed values (see, e.g., [@Ziman]). This uncertainty of the synchronized mode, at least qualitatively, may be regarded as an effect of the internal fluctuations of the order parameter $h$ and at the first step we will ignore them assuming the order parameter $h$ to be determined in a unique fashion for fixed values of the vehicle density $\ \rho $ and the mean velocity $v$. Thus for a uniform vehicle flow we write: $$\tau \frac{dh}{dt}=-\Phi (h,\rho ,v)\,, \label{2.1}$$ where $\tau $ is the time required of drivers coming to the decision to begin or stop overtaking manoeuvres and the function $\Phi(h,\rho ,v)$ possesses a single stationary point $h=h(\rho,v)$ being stable and, thus, $$\frac{\partial \Phi }{\partial h}>0\,. \label{2.2}$$ The latter inequality is assumed to hold for all the values of the order parameter for simplicity. We note that equation (\[2.1\]) also allows for the delay in the driver response to changes on the road. However, in contrast with models similar to (\[intro:1\]) and (\[intro:2\]), here this effect is not the origin of the traffic flow instability and, thus, its particular description is not so crucial. Moreover, as discussed in the Introduction, the time $\tau$ characterizes the delay in the driver decision concerning to the lane changing but not the control over the headway, enabling us to assume $\tau\gg\tau'$. The particular value $h(v,\rho )$ of the order parameter results from the compromise between the danger of accident during changing lanes and the will to move as fast as possible. Obviously, the lower is the mean vehicle velocity $v$ for a fixed value of $\rho $, the weaker is the lane changing danger and the stronger is the will to move faster. Besides, the higher is the vehicle density $\rho $ for a fixed value of $v$, the stronger is this danger (here the will has no effect at all). These statements enable us to regard the dependence $h(v,\rho )$ as a decreasing function of both the variables $v$, $\rho $ (Fig. \[F4\]) and taking into account inequality (\[2.2\]) to write: $$\frac{\partial \Phi }{\partial v}>0\,,\quad \frac{\partial \Phi }{\partial \rho }>0\,, \label{2.3}$$ with the latter inequality stemming from the danger effect only. Equation (\[2.1\]) describes actually the behavior of the drivers who prefer to move faster than the statistically mean vehicle and whose readiness for risk is greatest. Exactly this group of drivers govern the value of $h$. There is, however, another characteristics of the driver behavior, it is the mean velocity $v=\vartheta (h,\rho )$ chosen by the [*statistically averaged*]{} driver taking into account also the danger resulting from the frequent lane changing by the “fast” drivers. This characteristics is actually the same as one discussed in the Introduction but depends also on the order parameter $h$. So, as a function of $\rho $ it meets conditions (\[intro:3\]). Concerning the dependence of $\vartheta(h,\rho)$ on $h$ we can state that generally this function should be increasing for small values of the car density, $\rho \ll \rho _{0}$, because in the given case the lane changing practically makes no danger to traffic and all the drives can overtake vehicle moving at lower speed without risk. By contrast, when the vehicle density is sufficiently high, $\rho \lesssim \rho _{0}$, only the most “impatient” drivers permanently change the lanes for overtaking, making an additional danger to the most part of other drivers. Therefore, in this case the velocity $\vartheta (h,\rho )$ has to decrease as the order parameter $h$ increases. For certain intermediate values of the vehicle density, $\rho\approx\rho_{c}$, this dependence is to be weak. Fig. \[F5\] shows the velocity $\vartheta(h,\rho)$ as a function of $h$ for different values of $\rho$, where, in addition, we assume the effect of order parameter $h\in (0,1)$ near the boundary points weak and set $$\frac{\partial \vartheta }{\partial h}=0\quad \text{at}\quad h=0\ \text{and}\ h=1\,. \label{2.3a}$$ We will ignore the delay in the relaxation of the mean velocity to the equilibrium value $v=\vartheta (h,\rho)$ because the corresponding delay time characterizes the driver control over the headway and should be short, as already discussed above. Then the governing equation (\[2.1\]) for the order parameter $h$ can be rewritten in the form: $$\tau \frac{dh}{dt}=-\phi (h,\rho )\,;\quad \phi(h,\rho )\stackrel{\text{def}}{=}\Phi [h,\rho ,\vartheta (h,\rho )]\,. \label{2.4}$$ For the steady state uniform vehicle flow the solution of the equation $\phi (h,\rho )=0$ specifies the dependence $h(\rho )$ of the order parameter on the car density. Let us, now, study its properties and stability. Nonmonotony of the $h(\rho )$ dependence and the traffic flow instability ------------------------------------------------------------------------- To study the local characteristics of the right-hand side of Eq. (\[2.4\]) we analyze its partial derivatives $$\begin{aligned} \frac{\partial \phi }{\partial h} &=&\frac{\partial \Phi }{\partial h}+ \frac{\partial \Phi }{\partial v}\frac{\partial \vartheta }{\partial h}\,, \label{2.6} \\ \frac{\partial \phi }{\partial \rho } &=&\frac{\partial \Phi }{\partial \rho} +\frac{\partial \Phi }{\partial v}\frac{\partial \vartheta }{\partial \rho } \,. \label{2.7}\end{aligned}$$ As mentioned above, the value of $\partial \Phi /\partial \rho $ is solely due to the danger during changing lanes, so this term can be ignored until the vehicle density $\rho$ becomes sufficiently high. In other words, in a certain region $\rho <\rho_h\lesssim \rho _{0}$ the derivative $\partial \phi /\partial \rho \sim (\partial \Phi /\partial v)(\partial \vartheta /\partial \rho )<0$ by virtue of (\[intro:3\]) and (\[2.3\]). So, the local behavior of the function $h(\rho)$ (meeting the equality $d\phi=0$ and, thus, $dh/d\rho = -(\partial\phi/\partial\rho)(\partial\phi/\partial h)^{-1}$) depends directly on the sign of the derivative $\partial \phi /\partial h$, it is increasing or decreasing for $\partial \phi /\partial h>0$ or $\partial \phi /\partial h<0$, respectively. For long-wave perturbations $\propto \exp \{ikx\}$ of the car distribution on a highway the density $\rho $ can be treated as a constant at the lower order in $k$. Therefore, according to Eq. (\[2.4\]) the steady state traffic flow is unstable if $\partial \phi /\partial h<0$. Due to (\[2.2\]) and (\[2.3a\]) the first term on the right-hand side of (\[2.6\]) is dominant in the vicinity of the lines $h=0$ and $h=1$, thus, in this region the curve $h(\rho )$ is increasing and the stationary state of the traffic flow is stable. For $\rho <\rho _{c}$ the value $\partial \vartheta /\partial h >0$ (Fig. \[F5\]), therefore, the whole region $\left\{ 0<h<1,\;0<\rho <\rho _{c}\right\} $ corresponds to the stable car motion. However, for $\rho >\rho _{c}$ there can be a region of the order parameter $h$ where the derivative $\partial \phi /\partial h$ changes the sign and the vehicle motion becomes unstable. Indeed, the solution $v=\eta (h,\rho )$ of the equation $\Phi (h,\rho ,v)=0$ can be regarded as the mean vehicle velocity controlled by the “fast” drivers and is decreasing function of $h$ because of $\partial \eta /\partial h=-(\partial \Phi /\partial h)/(\partial \Phi /\partial v)^{-1}$. So, once such “active” drivers become to change the lanes to move faster, they will do this as frequently as possible especially if the mean velocity decreases, which corresponds to a considerable increase in $h$ for a small decrease in $v$. So, it is quite natural to assume that the value of $\partial \eta /\partial h$ for $\rho >\rho _{c}$ is sufficiently small and $$\frac{\partial \phi }{\partial h} = \frac{\partial\Phi}{\partial v}\left( \frac{\partial \vartheta }{\partial h}-\frac{\partial \eta }{\partial h} \right)<0\,.$$ Under these conditions the instability region does exist, the curve $h(\rho)$ can look like “S” (Fig. \[F6\]) and its decreasing branch corresponds to the unstable vehicle flow. The lower increasing branch matches the free flow state of the car motion, whereas the upper one should be related to the synchronized mode because it is characterized by the order parameter coming to unity. Hysteresis and the fundamental diagram -------------------------------------- The obtained dependence $h(\rho)$ actually describes the first order phase transition in the vehicle motion. Indeed, when increasing the car density exceeds the value $\rho_1$ the free flow becomes absolutely unstable and the synchronized mode forms through a sharp jump of the order parameter. If, however, after that the car density decreases the synchronized mode will persist until the car density attains the value $\rho_2 < \rho_1$. It is a typical hysteresis and the region $(\rho_2, \rho_1)$ corresponds to the metastable phases of traffic flow. Let us, now, discuss a possible form of the fundamental diagram $j=j(\rho)$ showing the vehicle flux $j=\rho \vartheta [\rho ]$ as a function of the car density $\rho $, where, by definition, $\vartheta [\rho ]=\vartheta [h(\rho ),\rho ]$. It should be pointed out that here we confine our consideration to the region of not too large values of the car density, $\rho<\rho_h$, where the transition “free flow $\leftrightarrow$ synchronized mode” takes place. The transition “synchronized mode $\leftrightarrow$ jammed traffic” will be discussed below. Fig. \[F7\][$a$]{} displays the dependence $\vartheta (h,\rho )$ of the mean vehicle velocity on the density $\rho $ for the fixed limit values of the order parameter $h=0$ and 1. For a small values of $\rho $ these curves practically coincide with each other. As the vehicle density $\rho $ grows and until it comes close to the critical value $\rho _{c}$ when the lane change danger becomes substantial, the velocity $\vartheta (1,\rho )$ practically does not depend on $\rho $. So at the point $\rho _{c}$ at which the curves $\vartheta (1,\rho )$ and $\vartheta (0,\rho )$ meet each other, $\vartheta (1,\rho )$ is to exhibit sufficiently sharp decrease in comparison with the latter one. Therefore, on one hand, the function $j_{1}(\rho )=\rho \vartheta (1,\rho )$ has to be decreasing for $\rho > \rho _{c}$. On the other hand, at the point $\rho _{c}$ for $h\ll 1$ the effect of the lane change danger is not extremely strong, it only makes the lane change ineffective, $\partial \vartheta /\partial h\approx 0$ (Fig. \[F5\]). So it is reasonable to assume the function $j_{0}(\rho )=\rho \vartheta (0,\rho )$ increasing near the point $\rho _{c}$. Under the adopted assumptions the relative arrangement of the curves $j_{0}(\rho )$, $j_{1}(\rho )$ is demonstrated in Fig. \[F7\][$b$]{}, and Fig. \[F7\][$c$]{} shows the fundamental diagram of traffic flow resulting from Fig. \[F6\] and Fig. \[F7\][$b$]{}. Concluding the present section we note that in the given description of the driver behavior governing the order parameter $h$ the vehicle flux $j(h,\rho )=\rho \vartheta (h,\rho )$ is an external characteristics of traffic flow. So, the obtained form of the fundamental diagram does not follows directly from the developed model, but can be interpreted sufficiently reasonable. It can be rigorously justified if the critical point $\rho_{c}$ corresponds to the maximum of the flux $j(h^{*},\rho )$ for a certain fixed value $h^{*}$ of the order parameter. In other words, when the road capacity is exhausted and the following increase in the vehicle density leads to a decrease in the vehicle flux the drivers divide into two groups, the majority prefer to move at their own lanes whereas the most “impatient” drivers change the lanes as frequently as possible, giving rise to the traffic instability. This problem, however, deserves an individual investigation. Phase coexistence. Diffusion limited cluster motion {#sec:3} =================================================== The previous section has considered uniform traffic flow, so, analyzed actually the individual characteristics of the free flow and the synchronized mode. In the present section we study their coexistence, i.e., the conditions under which a car cluster of finite size forms. This problem, however, requires that the traffic flow model be defined concretely. Therefore, in what follows we will consider a certain simple model which illustrates the characteristic features of the car cluster self-organization without complex mathematical manipulations. As before, the model under consideration assumes the mean velocity relaxation to be immediate and modifies the governing equation (\[2.4\]) in such a way as to ascribe the order parameter $h$ to a local car group. In other words, we describe the vehicle flow by the Lighthill–Whitham equation with dissipation (see, e.g., [@N96] and also Introduction), replace the time derivative in Eq. (\[2.4\]) by the particle derivative, and take into account that the order parameter cannot exhibit substantial variations over scales $l\sim \theta^{1/2} \tau \lesssim v_{0}\tau $ ($\theta $ is the velocity variance, $v_0$ is the typical car velocity in the free flow). Namely, we write: $$\begin{aligned} \frac{\partial \rho }{\partial t}+\frac{\partial \left[ \rho \vartheta (h,\rho )\right] }{\partial x} &=&D\frac{\partial ^{2}\rho } {\partial x^{2}} \,, \label{3.1} \\ \tau \left[ \frac{\partial h}{\partial t}+\vartheta (h,\rho )\frac{\partial h}{\partial x} \right] &=&{\hat{{\cal L}}}\{h\}-\phi (h,\rho )+\xi (x,t)\,. \label{3.2}\end{aligned}$$ Let us discuss the meaning of the particular terms of the given model. The Burgers equation (\[3.1\]), as already discussed in Introduction, allows for the fact that drivers govern their motion taking into account not only the behavior of the nearest cars, but the state of traffic flow inside the whole field of their front view of length. The effective diffusivity $D$ can be estimated as $D\sim L^{*}v_{0}$, where $L^*\gg l$ is a front distance looked through by drivers assumed to be much greater that the scale $l$, so $$D\tau \sim lL^{*}\gg l^{2}\,. \label{3.3}$$ The function $\phi (h,\rho )$ is of the form $$\phi (h,\rho )\stackrel{\text{def}}{=}h(1-h)[a(\rho )-h]\,, \label{3.202}$$ where $$a(\rho )=\left\{ \begin{array}{lll} 1 & \;\text{for} & \rho <\rho _{c}\,, \\ (\rho _{c}+\Delta -\rho )/\Delta & \;\text{for} & \rho _{c}<\rho <\rho _{c}+\Delta \,, \\ 0 & \;\text{for} & \rho >\rho _{c}+\Delta \,. \end{array} \right.$$ It describes such a driver behavior that $h=0$ and $h=1$ are the unique stable values of the order parameter for $\rho <\rho _{c}$ and $\rho >\rho _{c}+\Delta $, respectively, whereas, for $\rho _{c}<\rho <\rho _{c}+\Delta $ the points $h=0$, $h=1$ are both locally stable and there is an additional unstable stationary point, namely, $h=a(\rho )$. The term $$\hat{{\cal L}}\{h\}\stackrel{\text{def}}{=} l^{2}\frac{\partial ^{2}h}{\partial x^{2}}+\frac{l} {\sqrt{2}}\frac{\partial h}{\partial x} \label{3.201}$$ governs spatial variations in the field $h(x,t)$ and takes into account that drivers mainly follow the behavior of cars in front of them and cars moving at the rear cannot essentially affect them. The mean car velocity depends on $h$ and $\rho $ as $$\rho \vartheta (h,\rho )\stackrel{\text{def}}{=} \rho \vartheta _{0}(1-h)+[\rho _{c}\vartheta _{0}-\nu (\rho -\rho _{c})]h\,. \label{3.2b}$$ The last term on the right-hand side of Eq. (\[3.2\]) characterizes the random fluctuations in the order parameter dynamics: $$\begin{aligned} \label{3.203a} \left\langle \xi (x,t)\right\rangle &=&0\,,\\ \label{3.203} \left\langle \xi (x,t)\xi (x^{\prime },t^{\prime })\right\rangle &=&\sigma ^{2}l\tau \delta (x-x^{\prime})\delta (t-t^{\prime })\,,\end{aligned}$$ where $\sigma $ is their dimensionless amplitude. Expressions (\[3.202\]) and (\[3.2b\]) gives the $h(\rho )$-dependence and the fundamental diagram shown in Fig. \[F8\] simplifying the one presented in Fig \[F7\]. If we ignore the random fluctuations of the order parameter $h$, i.e., set $\sigma =0$, then the model (\[3.1\]), (\[3.2\]) will give us an artificially long delay (much greater than $\tau $) in the order parameter variations from, for example, the unstable point $h=0$ to the stable point $h=1$. Such a delay can lead to a meaningless great increase of the vehicle density in the free flow without phase transition to congestion. In order to avoid this artifact and to allow for the effect of real fluctuations in the driver behavior we also will assume the amplitude $\sigma $ to obey the condition [@LM00]: $$\left( \frac{l}{L^{*}}\right) ^{5/4}\lesssim \sigma \ll 1 \label{3.5}$$ ($\sigma \ll 1$, because, otherwise, the traffic flow dynamics would be totally random). It should be noted that small random variations of the order parameter $h$ near the points $h=0$, $h=1$ going into the regions $h<0$ and $h>1$, respectively, do not come into conflict with its physical meaning as the measure of the car motion correlations. Indeed, the chosen values $h=0$ and $h=1$ can describe a renormalization of real correlation coefficients $\tilde{h}=\tilde{h}_{1} > 0$, $\tilde{h}_{2} < 1$. According to Eq. (\[3.2\]), for the order parameter $h$ the characteristic scale of its spatial variations is $l$, so, the layer $\Im _{h}$ separating the regions where $h\approx 0$ and 1 is of thickness about $l$. Due to inequality (\[3.3\]) the car density on such scales can be treated as constant. Therefore, the transition region ${\cal L}_{\rho }$ between practically the uniform free flow and the congested phase is of thickness determined mainly by spatial variations of the vehicle density and on such scales the layer $\Im _{h}$ can be treated as an infinitely thin interface. In addition, the characteristic time scale of the layer $\Im _{h}$ formation is about $\tau $, whereas it takes about the time $\tau _{\rho }\sim D/v_{0}^{2}\sim \tau (L^{*}/l)\gg \tau $ for the layer ${\cal L}_{\rho }$ to form. Thereby, when analyzing the motion of a wide car clusters we may regard the order parameter distribution $h(x,t)$ as quasi-stationary for a fixed value of the car density $\rho $. Let us, now, consider two possible limits of the layer $\Im _{h}$ motion under such conditions. Regular dynamics ---------------- In the region $\rho _{c}<\rho <\rho _{c}+\Delta$ until the value of $a(\rho)$ comes close to the boundaries $h=0$ and $h=1$ the effect of the random fluctuations is ignorable. In this case by virtue of the adopted assumptions the solution of Eq. (\[3.2\]) that describes the layer $\Im _{h}$ moving at the speed $u$ is of the form: $$h=\frac{1}{2}\left[ 1+\tanh \left( \frac{x-ut}{\lambda }\right) \right]\,. \label{3.4}$$ Here for the layer $\Im _{01}$ of the transition “free-flow $\rightarrow$ synchronized mode” and for the layer $\Im _{10}$ of the opposite transition (Fig. \[F9\]) $$\lambda _{01}=\frac{2\sqrt{2}}{\eta _{v}}l\,,\quad \lambda _{10}= -2\sqrt{2} \eta _{v}l \label{3.6}$$ $$\begin{aligned} u_{01} &=&\vartheta_0 - \frac{\Delta_v}{2} -\frac{l}{\sqrt{2}\eta _{v}\tau } \left[ 1+\eta _{v}-2a(\rho_i ) \right] \,, \label{3.7a} \\ u_{10} &=&\vartheta_0 - \frac{\Delta_v}{2} - \frac{l}{\sqrt{2}\tau }\left[ 2\eta _{v}a(\rho_i )- (\eta _{v}-1)\right] \,. \label{3.7b}\end{aligned}$$ where we introduced the quantities: $$\begin{aligned} \Delta _{v} &=&\vartheta (0,\rho_i )-\vartheta (1,\rho_i )\,, \\ \eta _{v} &=&\left[ 1+\left( \frac{\tau \Delta _{v}}{2\sqrt{2}l}\right) ^{2}\right] ^{1/2}+\frac{\tau \Delta _{v}}{2\sqrt{2}l}\,.\end{aligned}$$ and $\rho_i$ is the corresponding value of the car density inside the layers $\Im_{01}$ and $\Im_{10}$. Expressions (\[3.7a\]), (\[3.7b\]) describe the regular dynamics of the car cluster formation because the transition, for example, from the free flow to the synchronized phase at a certain point $x$ is induced by this transition at the nearest points. The dependence of the velocities $u_{01}$ and $u_{10}$ on the local car density $\rho _{i}$ is illustrated in Fig. \[F9\]. The characteristic velocities attained in this type motion can be estimated as $$\vartheta _{0}-u\sim \max \left\{ \vartheta _{0}\Delta /\rho _{c},\,l/\tau \right\} \,,$$ so, under the adopted assumptions the regular dynamics does not allow for the sufficiently fast motion of the layers $\Im _{h}$ upstream. Noise-induced dynamics ---------------------- As the car density $\rho $ tends to the critical values $\rho _{c}$ or $ \rho _{c}+\Delta $ the value of $a(\rho )$ comes close to the boundaries $ a(\rho _{c})=1$ and $a(\rho _{c}+\Delta )=0$, and the point $h=1$ or $h=0$ becomes unstable, respectively. In this case the effect of the random fluctuations $\xi (x,t)$ plays a substantial role. Namely, the phase transition, for example, from the free flow to the synchronized motion (for $\rho \approx \rho _{c}+\Delta $) is caused by the noise $\xi (x,t)$ and equiprobably takes place at every point of the region wherein $\rho \approx \rho _{c}+\Delta$ rather than is localized near the current position of the layer $\Im _{01}$. Under these conditions the motion of the layers $\Im _{h}$ can be qualitatively characterized by an extremely high velocity in both the directions, which is illustrated in Fig. \[F9\] by dashed lines. We note that the noise-induced motion, in contrast to the regular dynamics, is to exhibit significant fluctuations in the displacement of the layer $\Im _{h}$ as well as in its forms. This question is, however, a subject for an individual study. Diffusion limited motion of vehicle clusters -------------------------------------------- Let us, now, analyze the motion of a sufficiently large cluster that can form on a highway when the initial car density or, what is the same, the average car density $\bar{\rho}$ belongs to the metastable region, $\bar{ \rho}\in (\rho _{c},\,\rho _{c}+\Delta )$. The term “sufficiently large” means that the cluster dimension $L$ is assumed to be much greater than the front distance $L^{*}$ looked through by drivers, so, they cannot look round the congestion as a whole. Exactly, in this case a quasi-local description of traffic flow similar to the differential equations (\[3.1\]), (\[3.2\]) is justified. Converting to the frame $y=x-ut$ moving at the cluster velocity $u$, solving Eq. (\[3.1\]) individually for the free flow and the synchronized phase, and treating the layers $\Im _{h}$ as infinitely thin interfaces we get the following conclusion. Within the frameworks of the given model the car cluster moves upstream sufficiently fast, so, the motion of the layers $\Im _{01}$ and $\Im _{10}$ is governed by the noise $\xi (x,t)$. In this case the values of the car density at the layers $\Im _{01}$ and $\Im _{10}$ have to be $\rho _{j}\approx \rho _{c}+\Delta $ and $\rho _{f}\approx \rho _{c}$, respectively. Thereby, the cluster velocity $u$ is mainly determined by the car redistribution governed by the diffusion type processes. The latter feature is the reason why we refer to the cluster dynamics under such conditions as to the diffusion limited motion. The transition region ${\cal L}_{01}$ between practically the uniform free flow state and the cluster contains the exponential increase of the vehicle density inside the free flow phase from the value $\rho _{f}$ far from the “interface” $\Im _{01}$ up to $\rho _{j}\approx \rho _{c}+\Delta $ at $\Im _{01}$, $$\rho =\rho _{f}+(\rho _{j}-\rho _{f})\exp \{q_{f}y\}\,,$$ where $q_{f}=(\vartheta _{0}+\left| u\right| )/D\sim 1/L^{*}$ and the frame $ \{y\}$ is attached to the “interface” $\Im _{01}$. The transition region $ {\cal L}_{10}$ from the synchronized phase to the uniform free flow is to be localized inside the car cluster. So, it is characterized by the decrease in the vehicle density $\delta \rho \propto \exp \{q_{j}y\}$, where $q_{j}=(\left| u\right| -\nu )/D$, and the vehicle free flow leaving the cluster is uniform at all its points (Fig. \[F10\]$a$). The cluster velocity is directly determined by the motion of the interface $\Im _{01}$. Therefore, assuming also the cluster dimension $L$ large in comparison with $L^{*}$, from Eq. (\[3.1\]) we get the expression of the same form as the Lighthill–Whitham formula (\[intro:4\]) relating the cluster velocity $u$ and the vehicle flux characteristics on the both sides of the layer ${\cal L}_{01}$. Whence it follows that at the first approximation: $$u\approx -\nu \,, \label{3.8}$$ the value $q_{j}=0$, and the vehicle cluster is of the form shown in Fig. \[F10\]$a$ under the name “mesocluster”. Assuming the total number of cars on the highway of length $L_{\text{rd}}$ fixed we get the expression for the mesocluster dimension $L$: $$L=2L_{\text{rd}}\frac{\bar{\rho}-\rho _{c}}{\Delta }\,. \label{3.9a}$$ However, this result is justified only for sufficiently small values of $( \bar{\rho}-\rho _{c})/\Delta \ll 1\,$, when the cluster dimension is not too large, $Lq_{j}\ll 1$ (nevertheless, $L\gg L^{*}$). Exactly for this reason we refer to such clusters as mesoscopic ones. In order to study the opposite limit, $Lq_{j}\gg 1$, we have to take into account that the value $\rho _{f}$ is not rigorously equal to $\rho _{c}$ but practically is the root $\rho _{f}^{*}>\rho _{c}$ of the equation $u_{10}(\rho _{f}^{*})=-\nu $. In this case the Lighthill–Whitham formula (\[intro:4\]) gives the expression: $$u\simeq -\left[ \nu +(\vartheta _{0}+\nu )\frac{\rho _{f}^{*}-\rho _{c}} {\Delta }\right]$$ leading to the following estimates of the thickness $1/q_{j}$ of the transition region ${\cal L}_{10}$: $$1/q_{j}\sim \frac{D\Delta }{(\vartheta _{0}+\nu )(\rho _{f}^{*}-\rho _{c})} \sim L^{*}\frac{\Delta }{(\rho _{f}^{*}-\rho _{c})}\,.$$ The form of such a wide cluster is shown in Fig. \[F10\]$a$, its dimension is: $$L=L_{\text{rd}}\frac{\bar{\rho}-\rho _{c}}{\Delta }\,. \label{3.9b}$$ and the region of the mean car density corresponding to this limit is specified by the inequality: $$\frac{\rho ^{*}-\rho _{c}}{\Delta }\gg \frac{L^{*}}{L_{\text{rd}}} \frac{% \Delta }{(\rho _{f}^{*}-\rho _{c})}\,. \label{3.10}$$ The resulting dependence of the cluster dimension on the mean car density $\bar{\rho}$ is illustrated in Fig. \[F10\][$b$]{}. Phase transition “synchronized mode $\leftrightarrow$ jam”. Brief discussion {#sec:4} ============================================================================ In Sec. \[sec:2\] we have considered the phase transition between the free flow and the synchronized mode. However, according to the experimental data [@K98] there is an additional phase transition in traffic flow regarded as the transition between the synchronized motion and the jammed “stop-and-go” traffic. This transition occurs at extremely high vehicle densities $\rho$ coming close to the limit value $\rho _{0}$. The present section briefly demonstrates that the developed model for the driver behavior also predicts a similar phase transition at high car densities. To avoid possible misunderstandings we, beforehand, point out that the model in its present form cannot describe details of the transition “synchronized mode $\leftrightarrow$ jam” because we have not taken into account the delay in the driver response to variations in headway. The latter is responsible for the formation of the “stop-and-go” pattern, so, to describe the jammed traffic on multilane highways we, at least, should combine a governing equation for the order parameter $h$ and a continuity equation similar to (\[3.2\]), (\[3.1\]) with an equation for the car velocity relaxation similar to (\[intro:2\]). This question, however, is worthy of individual study. Besides, the approximations used in Sec. \[sec:3\] to characterize the synchronized mode at the car densities near $\rho_c$ do not hold here. In Sec. \[sec:2\] we have studied the dependence of the order parameter $h$ on the car density ignoring the first term on the right-hand side of Exp. (\[2.7\]) caused by the dangerous of lane changing. This assumption is justified when the car density is not to high. In extremely dense traffic flow, when the car density exceeds a certain value, $\rho > \rho_h\lesssim \rho_0$, changing lanes becomes sufficiently dangerous and the function $\Phi (h,v,\rho )$ describing the driver behavior is to depend strongly on the vehicle density in this region. In addition, the vehicle motion becomes sufficiently slow. Under such conditions the former term on the right-hand side of expression (\[2.7\]) should be dominant and, thus, $\partial\phi/\partial\rho >0$. Therefore, the stable vehicle motion corresponding to $\partial \phi /\partial h>0$ matches the decreasing dependence of the order parameter $h(\rho )$ on the vehicle density $\rho $ for $\rho >\rho_h$. So, as the vehicle density $\rho$ increases the curve $h(\rho )$ can again go into the instability region (in the $h\rho $-plane), which has to give rise to a jump from the synchronized mode with greater values of the order parameter to a new traffic state with its less values (Fig. \[F11\]). Obviously, this transition between the two congested phases also exhibit the same hysteresis as one describe in Sec. \[sec:2\]. We identify the latter traffic state with the jammed vehicle motion. Indeed, in extremely dense traffic lane change is practically depressed, making the car ensembles at different lanes independent of one another. So, in this case vehicle flow has to exhibit weak multilane correlations and we should ascribe to it small values of the order parameter $h$. It should be noted that the experimental single-vehicle data [@Last] demonstrates strong correlations of variations in the traffic flux and the car density for both the free flow and the “stop-and-go” motion. By contrast, the synchronized mode is characterized by small values of the cross-covariance between flow, speed, and density. In other words, for the free flow and the “stop-and-go” motion the traffic flux $j=\vartheta\rho$ should depends directly on the car density $\rho$, as it must in the present model if we set $h=0$. Finalizing the present section we point out that the given model treats the jammed phase as a “faster” vehicle motion then the synchronized mode at the [*same*]{} values of the order parameter. There is no contradiction with the usual view on the synchronized mode as a high flux traffic state. The latter corresponds to the traffic flow at the vehicle densities near the phase transition “free flow $\leftrightarrow$ synchronized mode” rather than close to the limit value $\rho_0$. Besides, ordinary driver’s experience prompts that a highly dense traffic flow can be blocked at all if one of the cars begin to change the lanes. Nevertheless, in order to describe, at least qualitatively, the real features of the phase transition “synchronized mode $\leftrightarrow$ stop-and-go waves” a more sophisticated model is required. The present description only relates it to the instability of the order parameter at high values of the vehicle density. Besides, the present analysis demonstrates also the nonmonotonic behavior of the order parameter as the car density increases even if we ignore the hysteresis regions and focus our attention the stabel vehicle flow regions only. It should be noted that a similar nonmonotonic dependence of the lane change frequency on the car density as well as the platoon formation has been found in the cellular automaton model for two-lane traffic [@CA]. Closing remarks\[sec:cr\] ========================= Concluding the paper we recall the key points of the developed model. We have proposed an original macroscopic approach to the description of multilane traffic flow based on an extended collection of the traffic flow state variables. Namely, in addition to such characteristics as the car density $\rho$ and mean velocity $v$ being actually the zeroth and first moments of the “one particle” distribution function we introduce a new variable $h$ called the “order parameter”. It stands for the [*internal*]{} correlations in the car motion along different lanes that are due to the lane changing manoeuvres. The order parameter, in fact, allows for the essentially “many-body” effects in the car interaction so it is treated as an independent state variable. Taking into account the general properties of the driver behavior we have stated a governing equation for the order parameter. Based on current experimental data [@KR2; @K98; @KR1; @KR3; @Last] we have assume the correlations in the car motion on multilane highways to be due to a small group of “fast” drivers, i.e. the drivers who move substantially faster than the statistically mean vehicle continuously overtaking other cars. These “fast” cars, on one hand, increase individually the total rate of vehicle flow but, on the other hand, make the accident danger greater and, thus, cause the statistically mean driver to decrease the velocity. The competition of the two effects depends on the car density and the mean velocity and, as shown, can give rise to the traffic flow instability. It turns out that the resulting dependence of the order parameter on the car density describes in the same way the experimentally observed sequence of phase transitions “free flow $\leftrightarrow$ synchronized motion $\leftrightarrow$ jam” typical for traffic flow on highways [@K98]. Besides, we have shown that both these transitions should be of the first order type and exhibit hysteresis, matching the experimental data [@K98; @KR1; @KR3]. The synchronized mode is characterized by a large value of the order parameter, whereas the free flow and the jam match its small values. The latter feature enables us to treat the jam as a phase comprising the vehicle flows at different lane with weak mutual interaction because of the lane changing being depressed. In order to illustrate the characteristic features of the car clusters that self-organizing under these conditions we have considered a simple model dealing only with the evolution of the car density and the order parameter. In particular, it is shown that in the steady state the car density inside the cluster and the free flow being in equilibrium with the cluster, as well as the velocity at which the cluster moves upstream are fixed and determined by the basic properties of the traffic flow. On the contrary, the size of the car cluster depends on the initial conditions. Finally, we would like to underline that the developed model takes into account only one effect causing the traffic flow instability. The other, the delay in the driver control over the headway, seems to be responsible for the “stop-and-go” waves in the jammed phase (for a review of the continuum description of this phenomena see, e.g., [@T99a; @T99b]). So, combining the two approaches into one model it enables detail description of a wide class of phenomena occurring in the transitions from free flow to the heavy congested phase on highways. In this way the order parameter model could describe also the formation of a local jam on a highway whose boundaries comprise both of the phase transitions. In the present form it fails to do this because the free flow and the jammed traffic are characterized by small values of the order parameter. Concerning a possible derivation of the order parameter model from the gas-kinetic theory we note that the appearance of the “fast” driver platoons demonstrates a substantial deviation of the car distribution function from the monotonic quasi-equilibrium form. So, to construct an adequate system of equations dealing with the moments of the distribution function a more sophisticated approximation is required. Acknowledgments {#acknowledgments .unnumbered} =============== One of us, I.A. Lubashevsky, would like to acknowledge the hospitality of Physics Department of Rostock University during the stay at which this work was carried out. M.J. Lighthill and G.B. Whitham, Proc. R. Soc. London Ser. A, [**229**]{}, 281, 317 (1955). P.I. Richards, Oper. Res., [**4**]{}, 42 (1956). I. Prigogine and R. Herman, [*Kinetic Theory of Vehicular Traffic*]{} (American Elsevier, New York, 1971). E. Ben-Jacob, I. Cohen, O. Shochet, A. Tenenbaum, A. Czirók, and T. Vicsek, Nature (London) [**368**]{}, 46 (1994). T. Vicsek, A. Czirók, E. Ben-Jacob, I. Cohen, and O. Shochet, Phys. Rev. Lett. [**75**]{}, 1226 (1995). H.J. Bussemaker, A. Deutsch, and E. Geigant, Phys. Rev. Lett. [**78**]{}, 5018 (1997). B.S. Kerner, Physics World [**12**]{}, 25 (August 1999). B.S. Kerner, in: *Transportation and Traffic Theory*, ed. A. Ceder (Pergamon, Amsterdam, 1999), p. 147. B.S. Kerner, Transp. Res. Rec., [**1678**]{}, 160 (1999). D. Helbing, [*Verkehrsdynamik*]{} (Springer-Verlag, Berlin, 1997). B.S. Kerner and H. Rehborn, Phys. Rev. E [**53**]{}, R4275 (1996). B.S. Kerner, Phys. Rev. Lett. [**81**]{}, 3797 (1998). B.S. Kerner and H. Rehborn, Phys. Rev. [ E]{} [**53**]{}, R1297 (1996). B.S. Kerner and H. Rehborn, Phys. Rev. Lett. [**79**]{}, 4030 (1997). D. Helbing, Physica A [**242**]{}, 175 (1997). D. Helbing and A. Greiner, Phys. Rev. E [**55**]{}, 5498 (1997). D. Helbing and M. Treiber, Phys. Rev. Lett. [**81**]{}, 3042 (1998). M. Treiber, A. Hennecke, and D. Helbing, Phys. Rev. E [**59**]{}, 239 (1999). V. Shvetsov and D. Helbing, Phys. Rev. E [**59**]{}, 6328 (1999). A. Klar and R. Wegener, SIAM J. Appl. Math. [**59**]{}, 983, 1002 (1999). H.Y. Lee, D. Kim, and M.Y. Choi, in: [*Traffic and Granular Flow ’97*]{}, ed. M. Schreckenberg and D.E. Wolf (Springer-Verlag, Singapore, 1998), p. 433. B.S. Kerner and P. Konhäuser, Phys. Rev. [E]{}, [**48**]{}, R2335 (1993). B.S. Kerner and P. Konhäuser, Phys. Rev. [E]{} [**50**]{}, 54 (1994). T. Nagatani, Physica A [**264**]{}, 581 (1999). T. Nagatani, Phys. Rev. E [**60**]{}, 1535 (1999). G.B. Whitham, [*Linear and Nonlinear Waves*]{} (Wiley & Sons, New York, 1974). D. Helbing, in: [*Traffic and Granular Flow*]{}, ed. D.E. Wolf, M. Schreckenberg, and A. Bachem (World Scientific Publishers, Singapore, 1996), p. 87. D. Helbing and B. Tilch, Phys. Rev. E [**58**]{}, 133 (1998). D. Helbing, in: [*Transportation Systems*]{}, Vol. 2, ed. M. Papageorgiou and A. Pouliezos (International Federation of Automatic Control, Chania, Greece, 1998), p. 809. D. Helbing, in: [*A Perspective Look at Nonlinear Media. From Physics to Biology and Social Sciences*]{}, edited by J. Parisi, S.C. Muller, and W. Zimmermann (Springer-Verlag, Berlin, 1998), p. 122. B.S. Kerner, P. Konhäuser, and M. Schilke, Phys. Rev. E [**51**]{}, 6243 (1995). B.S. Kerner, S.L. Klenov, and P. Konhäuser, Phys. Rev. E [**56**]{}, 4200 (1997). L. Neubert, L. Santen, A. Schadschneider, and M. Schreckenberg, [*Single-Vehicle Data of Highway Traffic – a Statistical Analysis*]{}, cond-mat/9905216, (to be published in Phys. Rev. E). B.S. Kerner, in: [*Proceedings of 3rd International Symposium on Highway Capacity*]{}, edited by R. Rysgaard, Vol. 2 (Road Directories, Ministry of Transport – Denmark, 1998), p. 621. B.S. Kerner, in: [*Traffic and Granular Flow’97*]{}, ed. M. Schreckenberg and D.E. Wolf (Springer-Verlag, Singapore, 1998), p. 239. C. Wagner, J. Stat. Phys. [**90**]{}, 1251 (1998). R. Mahnke and N. Pieret, Phys. Rev. E [**56**]{}, 2666(1997). R. Mahnke and J. Kaupužs, Phys. Rev. E [**59**]{}, 117 (1999). J. Kaupužs and R. Mahnke, Eur. Phys. J. B [**14**]{}, 793 (2000). for a review of applications to different physical systems see, e.g., J.-C. Tolédano and P. Tolédano, [*The Landau Theory of Phase Transitions*]{} (World Scientific Publishers, Singapore, 1987). J.M.Ziman, [*Models of Disorder*]{} (Cambridge University Press, Cambridge, 1979). K. Nagel, Phys. Rev. E [**53**]{}, 4655 (1996). To obtain estimate (\[3.5\]) we need a special analysis and its presentation would substantially overload the given paper. Moreover, the amplitude of the random fluctuations does not enter the obtained formulae, so, we refer this question to our forthcoming paper. W. Knospe, L. Santen, A. Schadschneider, and M. Schreckenberg, Physica A [**265**]{}, 614 (1999).

ArXiv

--- abstract: 'We propose a scheme for the observation of micro-macro entanglement in photon number based on amplifying and de-amplifying a single-photon entangled state in combination with homodyne quantum state tomography. The created micro-macro entangled state, which exists between the amplification and de-amplification steps, is a superposition of two components with mean photon numbers that differ by approximately a factor of three. We show that for reasonable values of photon loss it should be possible to detect micro-macro photon-number entanglement where the macro system has a mean number of one hundred photons or more.' author: - 'Roohollah Ghobadi$^{1,2}$, Alexander Lvovsky$^1$, and Christoph Simon$^1$' title: 'Creating and detecting micro-macro photon-number entanglement by amplifying and de-amplifying a single-photon entangled state' --- The goal of pushing the observation of quantum effects such as superpositions and entanglement towards the macroscopic level is currently being pursued in a number of different fields, including trapped ions [@Monroe96], superconducting circuits [@Friedman00], nano-mechanics [@Connell10], microwave cavities interacting with atoms in Rydberg states [@Deleglise08], atomic ensembles [@polzik], and non-linear optics [@yurke; @entlaser; @Ourjoumtsev07; @Demartini08]. Within non-linear optics, one can distinguish proposals based on Kerr non-linearities [@yurke], and proposals and experiments based on parametric down-conversion [@entlaser; @Ourjoumtsev07; @Demartini08]. The latter area has recently seen significant activity, a lot of which was stimulated by Ref. [@Demartini08], which claimed the creation and detection of entanglement in polarization between a single photon on one side and thousands of photons on the other. The state was created starting from a single polarization entangled photon pair, by greatly amplifying one of the photons using stimulated type-II parametric down-conversion (i.e. a two-mode squeezing interaction involving both polarization modes). Ref. [@Sekatski] subsequently showed that the evidence for micro-macro entanglement given in Ref. [@Demartini08] was not conclusive, and Ref. [@Raeisi11] showed that in order to rigorously prove the presence of entanglement for the state of Ref. [@Demartini08], one would need to be able to count the photons on the macro side with single-photon resolution, which is a significant technological challenge. Several other results also suggest that the observation of entanglement by direct measurement on macroscopic systems generally requires very high resolution, which also implies very low photon loss in the case of multi-photon states [@mermin; @peres; @entlaser; @sciarrino]. These results inspired the work of Ref. [@Raeisi12], which proposed to prove the existence of micro-macro polarization entanglement by de-amplifying the macro part of the state of Ref. [@Demartini08] back to the single-photon level. As entanglement cannot be created locally, if entanglement is detected at the single-photon level, this proves that micro-macro entanglement had to exist after the amplification stage. This approach has two advantages. On the one hand, the final measurement can be done by single-photon detection, which is much simpler than counting large photon numbers with single-photon resolution. On the other hand, the entanglement is primarily sensitive to loss between the amplification and de-amplification stages, which is easier to minimize in practice than the overall loss, which also includes detection inefficiency. Most of the previous work in this area was concerned with polarization (or spin) entanglement [@entlaser; @Demartini08; @mermin; @peres; @sciarrino; @Raeisi12]. Here we propose to apply the amplification-deamplification approach to create and detect a different - quite striking - type of entanglement, namely micro-macro entanglement in photon number. Instead of starting from a polarization entangled photon pair, we propose to start from a single-photon entangled state, which can be created by sending a single photon onto a beam splitter [@banaszek; @Lvovsky04; @VanEnk; @morin]. The presence of single-photon entanglement can be proven experimentally by homodyne tomography [@Lvovsky04], and also by a combination of interference and single-photon detection [@Cho05]. Using single-photon entanglement as a starting point, one can create a micro-macro photon-number entangled state by amplifying one side via stimulated type-I parametric down-conversion (i.e. a single-mode squeezer). The resulting state is a superposition of two components with largely different mean photon numbers. This distinguishes our proposal from another recent proposal, where micro-macro entanglement is created by displacing (rather than squeezing) one half of a single-photon entangled state [@Sekatski12]. In that case the mean photon numbers of the two superposed components are very similar. Micro-macro photon number entanglement as suggested here could in principle be used to test proposals for fundamental decoherence in energy [@energy-decoherence]. In the following we describe our proposal in more detail, taking into account the effects of photon loss, see also Figure 1. We start by sending a single photon onto a balanced beam splitter. The output state is a single-photon entangled state $$|\psi_{in}\rangle=\frac{|1\rangle_{A}|0\rangle_{B}+|0\rangle_{A}|1\rangle_{B}}{\sqrt{2}}. \label{in}$$ The photon in arm B is then subjected to the unitary evolution $S=e^{-iHt}$ where $$H=i\chi(a^{2}-a^{\dagger^{2}}) \label{sq}$$ is the single-mode squeezing Hamiltonian, which can be implemented by type-I parametric down conversion. The total state after the application of $S$ becomes $$|\psi_{s}\rangle=\frac{|1\rangle_{A}|S_{0}\rangle_{B}+|0\rangle_{A}|S_{1}\rangle_{B}}{\sqrt{2}},$$ where $$|S_{0}\rangle=S|0\rangle=\frac{1}{\sqrt{\cosh r}}\sum_{n=0}^{\infty}\frac{\sqrt{(2n)!}}{2^{n}n!}(-\tanh r)^{n}|2n\rangle$$ and $$|S_{1}\rangle=S|1\rangle=\frac{1}{\sqrt{(\cosh r){}^{3}}}\sum_{n=0}^{\infty}\frac{\sqrt{(2n+1)!}}{2^{n}n!}(-\textrm{tanh}r)^{n}|2n+1\rangle,$$ with $r=\chi t$. For large enough squeezing parameter $r$ the state $|\psi_{s}\rangle$ is a superposition of two components with largely different mean photon numbers. To see this, we calculate $$n_{0}=\langle S_{0}|a^{\dagger}a|S_{0}\rangle= \sinh^{2}(r) \label{n0}$$ and $$n_{1}=\langle S_{1}|a^{\dagger}a|S_{1}\rangle=1+3\sinh^{2}(r). \label{n1}$$ For large enough values of $r$ one has $n_{1}/n_{0}\sim3$. In Figure 2 we show the photon number distributions for the states $|S_0\rangle$ and $|S_1\rangle$, as well as their Wigner functions. Note that values of $r$ much greater than those used in the figure were achieved in the experiment of Ref. [@Demartini08], where thousands of photons were created on the macro side. So far we have discussed the amplification process. In order to study the effects of photon loss and subsequent de-amplification as shown in Figure 1, it is most convenient to work with the Wigner function. The Wigner function corresponding to the state of Eq. (\[in\]) is given by $$W_{in}(X_{A},P_{A},X_{B},P_{B})=\frac{1}{2}\sum_{m,n=0}^{m,n=1}W_{m,n}(X_{A},P_{A})W_{1-m,1-n}(X_{B},P_{B}),$$ where $W_{0,0}(X_{i},P_{i})=\frac{e^{-(X_{i}^{2}+P_{i}^{2})}}{\pi}$, $W_{1,0}(X_{i},P_{i})=\frac{\sqrt{2}(X_{i}+iP_{i})e^{-(P_{i}^{2}+X_{i}^{2})}}{\pi}=W^{*}_{0,1}(X_{i},P_{i})$ and $W_{1,1}(X_{i},P_{i})=\frac{(-1+2X_{i}^{2}+2P_{i}^{2})e^{-(P_{i}^{2}+X_{i}^{2})}}{\pi}$ with $i=A,B$. $X_{i}$,$P_{i}$ are the position and momentum quadratures respectively. The effect of squeezing in the phase space is simply given by the following transformation: $X_{B}\rightarrow e^{r}X_{B}$, $P_{B}\rightarrow e^{-r}P_{B}$. This implies that the Wigner function after squeezing is given by $W_{s}(X_{A},P_{A},X_{B},P_{B})=W_{in}(X_{A},P_{A},e^{r}X_{B},e^{-r}P_{B})$. In the absence of photon loss, the prepared macroscopic state in arm B would now undergo the de-amplification operation $S^{-1}$, which can be realized by changing the sign of $\chi$. Experimentally this can be done either by inverting the phase of the pump laser or by inverting the sign of the non-linear coefficient of the second non-linear crystal, in analogy to what is done in periodic poling [@Armstrong62]. In practice the de-amplification will always be preceded by a certain amount of photon loss. In this case the final state is no longer exactly equal to the initial state, in particular there will be higher order excitations in the number basis (beyond one). We will first discuss only loss between $S$ and $S^{-1}$, which is the most critical imperfection; loss before $S$ and after $S^{-1}$ will be discussed below. The effect of loss in phase space is described by a convolution [@Leonhardt] $$\begin{aligned} &&W_{s,\eta}(X_{A},P_{A},X_{B},P_{B})= \nonumber \\ &&\intop_{-\infty}^{\infty}dX_{B}'dP_{B}'W_{s}(X_{A},P_{A},X_{B}',P_{B}')F_{\eta}(X_{B},P_{B},X_{B}',P_{B}')\end{aligned}$$ with the attenuation kernel $$F_{\eta}(X_{B},P_{B},X_{B}',P_{B}')=\frac{\exp[-\frac{\eta}{1-\eta}((X_{B}'-\frac{X_{B}}{\sqrt{\eta}})^{2}+(P_{B}'-\frac{P_{B}}{\sqrt{\eta}})^{2}))]}{\pi(1-\eta)}.$$ The Wigner function of the final state after de-squeezing is given by $W_{s,\eta,s^{-1}}(X_{A},P_{A},X_{B},P_{B})=W_{s,\eta}(X_{A},P_{A},e^{-r}X_{B},e^{r}P_{B})$. Using Eqs.\[9,10\] one can then obtain the density matrix of the final state[@Leonhardt]. As mentioned above, on the $B$ side the final state will in general have higher-order components in the Fock basis. As it is difficult to quantify entanglement in high-dimensional systems, we will here focus on the projection of the final state onto the zero and first excitation subspace for mode $B$, i.e $$\rho_{p}=(I_{A}\otimes P_{B})\rho(I_{A}\otimes P_{B}),$$ where $I_{A}$ is the identity operator in mode $A$ and $P_{B}$ is the projection in subspace $\{|0\rangle_{B},|1\rangle_{B}\}$ in arm $B$. Since the local projection $P_B$ cannot create entanglement, any entanglement present in $\rho_p$ also had to be present in $\rho$. Similarly, any entanglement present in $\rho$ had to be present in the micro-macro state created by the amplification stage because the loss and de-amplification are also local processes. The projected density matrix $\rho_p$ has the following form in the Fock state basis, $$\rho_{p}=\left(\begin{array}{cccc} p_{00} & 0 & 0 & d'\\ 0 & p_{01} & d & 0\\ 0 & d^{\ast} & p_{10} & 0\\ d'^{\ast} & 0 & 0 & p_{11} \end{array}\right), \label{ro10}$$ where $p_{ij}$ is the probability to find $i$ photons in arm $A$ and $j$ photons in arm $B$; $d$ is the coherence term between $|1\rangle_{A}|0\rangle_{B}$ and $|0\rangle_{A}|1\rangle_{B}$ , $d'$ is the coherence between $|0\rangle_{A}|0\rangle_{B}$ and $|1\rangle_{A}|1\rangle_{B}$. One should note that the projected density matrix is not normalized. In fact the success probability of projection is given by $Tr(\rho_{p})=p_{00}+p_{01}+p_{10}+p_{11}$. In the initial state $p_{01}=p_{10}=d=\frac{1}{2}$, with all other coefficients equal to zero. One sees that the combination of amplification, loss in between, and de-amplification can create new population terms as well as a new coherence. However, certain coherences are still exactly zero (under otherwise ideal conditions). This can be understood by noting that neither the Hamiltonian nor the loss can create coherence between neighboring photon number states (in a given mode). The zero elements in Eq. (\[ro10\]) can also be understood in term of the reflection symmetry of the initial state and attenuation kernel Eq. (10). To characterize entanglement we use the concurrence[@Wootters98] which is defined as $$C(\rho_{p})=\max(0,\lambda_{1}-\lambda_{2}-\lambda_{3}-\lambda_{4})$$ where $\lambda_{i}$ are the eigenvalues in decreasing order of the Hermitian matrix $\sqrt{\sqrt{\rho_{p}}\tilde{\rho}_{p}\sqrt{\rho_{p}}}$ with $\tilde{\rho}_{p}=(\sigma_{y}\otimes\sigma_{y})\rho_{p}^{*}(\sigma_{y}\otimes\sigma_{y})$. For the density matrix in Eq.(\[ro10\]) the concurrence is given by $$C(\rho_{p})=\max\{0,2(|d|-\sqrt{p_{00}p_{11}}),2(|d'|-\sqrt{p_{10}p_{01}})\}.$$ Fig. 3(a) shows the concurrence as a function of mean photon number $n=\frac{n_0+n_1}{2}$ after squeezing for different values of attenuation $\eta$. One notes the high sensitivity of the concurrence to the attenuation. However, in practice it should be possible to keep losses between the two non-linear crystals very low, values of $\eta$ as high as 0.99 should be realistic. One promising approach would be to realize amplification and de-amplification in a single solid-state system, where the two active non-linear sections could be separated by a non-active spacer layer. Note that the experiment would typically be performed with femtosecond pulses, so it would not be difficult to make non-active layer thick enough to contain the entire pulse. The macro component of the micro-macro entangled state would then exist for a short span of time in that spacer layer. Fig. 3(b) shows that the success probability for projecting the system into the zero-or-one photon subspace decreases as the amount of loss and the mean photon number are increased, but that it is still quite significant in the regime under consideration. Fig. 3 suggests that micro-macro entanglement involving hundreds of photons might be observable with the proposed scheme. To confirm this suggestion, it is still important to study the effect of losses before the amplification and after the de-amplification, $\eta_1$ and $\eta_2$. These losses are harder to minimize in practice. In particular, $\eta_2$ also includes detector inefficiency. However, values of order $\eta_1=\eta_2=0.9$ should be achievable [@schnabel]. Figure 4 shows that a substantial amount of entanglement is still present in the system under these conditions for a mean photon number $n=100$. One can see that the micro-macro photon-number entanglement is much less sensitive to $\eta_{1}$ and $\eta_{2}$ than to $\eta$, similarly to the results of Ref. [@Raeisi12] for micro-macro polarization entanglement. The entanglement can be demonstrated experimentally by using full homodyne tomography, which was already used to demonstrate the non-locality of single-photon entanglement in Ref. [@Lvovsky04]. In this method the full density matrix is reconstructed from the joint quadrature statistics $p_{\theta_{A},\theta_{B}}(X_{A},X_{B})$ for different values of the local oscillator phases $\theta_{A}$ and $\theta_{B}$. In particular this allows one to reconstruct the density matrix in the zero and one-photon subspace which is relevant for us here. We have proposed to create and detect micro-macro photon-number entanglement by amplifying and then de-amplifying a single-photon entangled state. In particular, the present approach should allow the creation and detection of entangled states that are superpositions of two components with very different mean photon numbers (for example, 50 and 150). [*Acknowledgments.*]{} We thank S.Raeisi, S.Rahimi Keshari and B. Sanders for useful discussions. This work was supported by AITF and NSERC. [99]{} C. Monroe, D.M.Meekhof, B.E.King, and D.J.Wineland, Science [**272**]{}, 1131 (1996). J.R. Friedman, V. Patel, W.Chen, S.K. Tolpygo, and J.E. Lukens, Nature [**406**]{}, 43 (2000). A.D. O’Connell [*et al.*]{}, Nature [**464**]{}, 697 (2010); W. Marshall, C. Simon, R. Penrose, and D. Bouwmeester, Phys. Rev. Lett. [**91**]{}, 130401 (2003). S. Deleglise [*et al.*]{}, Nature [**455**]{}, 510 (2008). B. Julsgaard, A. Kozhekin, and E.S. Polzik, Nature [**413**]{}, 400 (2001). B. Yurke and D. Stoler, Phys. Rev. Lett. [**57**]{}, 13 (1986); H. Jeong, M. Paternostro, and T.C. Ralph, Phys. Rev. Lett. [**102**]{}, 060403 (2009). C. Simon and D. Bouwmeester, Phys. Rev. Lett. [**91**]{}, 053601 (2003). A. Ourjoumtsev, H. Jeong, R. Tualle-Brouri, and P. Grangier, Nature [**448**]{}, 784-786 (2007). F. De Martini, F. Sciarrino, and C. Vitelli, Phys. Rev. Lett. [**100**]{}, 253601 (2008). P. Sekatski, N. Brunner, C. Branciard, N. Gisin, and C. Simon, Phys. Rev. Lett. [**103**]{}, 113601 (2009); P. Sekatski [*et al.*]{}, Phys. Rev. A [**82**]{}, 053814 (2010). S. Raeisi, P. Sekatski, and C. Simon, Phys. Rev. Lett. [**107**]{}, 250401 (2011). N.D. Mermin, Phys. Rev. D [**22**]{}, 356 (1980). A. Peres, [*Quantum Theory: Concepts and Methods*]{} (Klouwer, Dordrecht, 2002); J. Kofler and C. Brukner, Phys. Rev. Lett. [**99**]{}, 180403 (2007). N. Spagnolo, C. Vitelli, M. Paternostro, F. De Martini, and F. Sciarrino, Phys. Rev. A [**84**]{}, 032102 (2011). S. Raeisi, W. Tittel, C. Simon, Phys. Rev. Lett. [**108**]{}, 120404 (2012). K. Banaszek and K. Wódkiewicz, Phys. Rev. Lett. [**82**]{}, 2009 (1999). S. A. Babichev, J. Appel, and A. I. Lvovsky, Phys. Rev. Lett. [**92**]{}, 193601 (2004). S.J. van Enk, Phys. Rev. A [**72**]{}, 064306 (2005). O. Morin [*et al.*]{}, arXiv:1206.5734. C.W. Chou [*et al.*]{}, Nature [**438**]{}, 828 (2005); J. Laurat, K. S. Choi, H. Deng, C. W. Chou, and H. J. Kimble, Phys. Rev. Lett. [**99**]{}, 180504 (2007). P. Sekatski [*et al.*]{}, arXiv:1206.1870. G.J. Milburn, Phys. Rev. A [**44**]{}, 5401 (1991); R. Gambini, R.A. Porto, and J. Pullin, Class. Quantum Grav. [**21**]{}, L51 (2004); C. Simon and D. Jaksch, Phys. Rev. A [**70**]{}, 052104 (2004). U. Leonhardt, [*Measuring the Quantum State of Light*]{} (Cambridge University Press, Cambridge, 1997). J. A. Armstrong, N. Bloembergen, J. Ducuing, and P. S. Pershan Phys. Rev. [**127**]{}, 1918 (1962) W.K. Wootters, Phys. Rev. Lett. [**80**]{}, 2245 (1998). H. Vahlbruch [*et al.*]{}, Phys. Rev. Lett. [**100**]{}, 033602 (2008).

ArXiv

--- abstract: 'The magneto-optical imaging technique is used to visualize the penetration of the magnetic induction in YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ thin films during surface resistance measurements. The in-situ surface resistance measurements were performed at 7 GHz using the dielectric resonator method. When only the microwave magnetic field $H_{rf}$ is applied to the superconductor, no $H_{rf}$-induced vortex penetration is observed, even at high rf power. In contrast, in the presence of a constant magnetic field superimposed on $H_{rf}$ we observe a progression of the flux front as $H_{rf}$ is increased. A local thermometry method based on the measurement of the resonant frequency of the dielectric resonator placed on the YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ thin film shows that the $H_{rf}$–induced flux penetration is due to the increase of the film temperature.' author: - Julien Kermorvant - 'Jean-Claude Mage' - Bruno Marcilhac - Yves Lemaître - 'Jean-François Bobo' - Cornelis Jacominus van der Beek date: - - title: 'Microwave heating-induced DC magnetic flux penetration in YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ superconducting thin films' --- Introduction ============= High Temperature Superconductor (HTS) thin films are now recognized as particularly suitable for high frequency signal processing. Due to their very low surface resistance $R_{s}$, as compared to normal metals, they allow for very efficient microwave signal filtering and detection. However, in the region of high power $P_{rf}$ of the incident microwave field, their application is limited by the strong dependence of $R_{s}$ on the $P_{rf}$–magnitude. A nonlinear increase of the surface resistance with the input rf power is commonly observed,[@Oates92; @Oates95; @Samoilova95; @Hampel1996; @Wosik97; @Anlage99; @Lahl2005; @Kermorvant2009] leading to a detrimental decrease of the $Q$-factor of the devices.[@Hein97] The origin of the nonlinear microwave losses in high $T_{c}$ superconductors has been studied by many groups. Among the cited causes, there are intrinsic phenomena such as the excitation of quasiparticles when the current density induced by the microwave magnetic field (of magnitude $H_{rf}$) becomes of the order of the pair-breaking current density,[@Dahm99] but also the limitation of the supercurrent in grain boundaries, and vortex motion [@Lahl2005; @Coffey1992; @Coffey1993; @Coffey1993ii] in the superconductor induced by the rf field. The interplay of the superposed ac and dc magnetic fields in superconducting thin films is extensively described in Refs. . Recent work has, however, clearly demonstrated that it is simply local Joule heating of the superconducting film by the microwave field that leads to the nonlinear behavior,[@Hampel1996; @Wosik97; @Kermorvant2009] and that the dissipation at the origin of the heating is due to the linear electromagnetic response of the films. In Ref. , we have introduced a valuable tool for local thermometry of superconducting films studied using a dielectric resonator. The variation of the resonator frequency as a function of the temperature depends essentially on the dielectric constant $\epsilon$ of the resonator; its calibration turns the latter into a precise local thermometer. This method has allowed us to measure the temperature of YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ films as a function of the rf input power (at 10 GHz) under nominally isothermal conditions. The observed temperature increase of the resonator and the YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ thin films as function of rf power unambiguously showed that the usually observed increase in the surface resistance is due to Joule heating, with a linear response dissipation mechanism. Candidate mechanisms at the origin of the heating can therefore be limited to quasi-particle dissipation [@Dahm99; @Mattis1958] and flux-flow losses.[@Coffey1992; @Coffey1993; @Coffey1993ii; @Bardeen1965; @Brandt1991; @Coffey1991] ![(Color online) Principle of MOI of superconductors. Thin drawn lines depict the magnetic flux as this traverses the superconductor and the MOL, thick black lines show the optical path of the impinging and reflected light, and the circled arrows illustrate the linear polarization direction of the light.[]{data-label="fig2"}](Figure1.pdf){width="1.01\linewidth"} ![Complex permeability of the (Lu,Bi)$_{3}$(Fe,Ga)$_{5}$O$_{12}$ magneto-optically active layer, measured at room temperature.[]{data-label="fig1"}](Figure2.pdf){width="1.01\linewidth"} Here, we study the influence of an rf magnetic field generated by a rutile dielectric resonator on the DC flux distribution in YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ thin films. For this, we have developed a dedicated set-up, that allows for the simultaneous measurement of the film surface resistance and the visualization of the magnetic flux distribution using the Magneto-Optical Imaging (MOI) Technique.[@Dorosinskii1992; @Jooss2002] MOI is based on the Faraday effect in which the rotation of the polarization plane of incident linearly polarized light is proportional to the magnetic induction component parallel to the wavevector of the incoming light. Since superconductors do not present a significant Faraday effect, one has to use a magneto-optical layer (MOL, with a strong Faraday effect) placed on top of a superconductor, as depicted in Fig. \[fig2\]. In the experiments described below, we use, as a MOL, Lu- and Bi-doped Yttrium-Iron Garnet thick films[@Uehara2009] with a ferromagnetic resonance at 0.9 GHz (see Fig. \[fig1\]). The application of a magnetic field of a few dozen mT will increase the ferromagnetic resonance frequency somewhat. However, at the experimental microwave field frequency of 7.0 GHz, the magnetization rotation leading to the Faraday effect will always be strongly overdamped, inhibiting the direct visualization of the microwave magnetic field. The MOI technique does permit the visualization of the modification of static flux structures in YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ thin films following microwave field application. Note that the development of a similar imaging system was reported in Ref. , however, the influence of the microwave magnetic field on flux penetration was not reported there. ![(Color online) Schematic view of the experimental assembly. Panel (a) shows a side view, while panel (b) depicts a top view for different lid apertures. The left-hand side cover allows the imaging of the central region of the sample, while the right-hand cover allows for the observation of the edge region. []{data-label="fig3"}](Figure3.pdf){width="0.9\linewidth"} Experimental details ==================== YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ films -------------------------------------- All YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ films under study in this work were cut from the same wafer, designated SY211 in Ref. . The YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ was deposited on a 500 $\mu$m-thick MgO substrate using cylindrical hollow cathode DC sputtering, and had a thickness $d =400$ nm. Its critical temperature is $T_{c} \simeq 86$ K and its critical current density $j_{c} = 5\times 10^{10}$ Am$^{-2}$ at $T = 77$ K. Magneto-Optical Imaging ----------------------- The MOL is a 5 $\mu$m-thick (Lu,Bi)$_{3}$(Fe,Ga)$_{5}$O$_{12}$ ferrimagnetic garnet film with in-plane magnetic anisotropy, grown on a 500 $\mu$m-thick Gd$_{3}$Gd$_{5}$O$_{12}$ (GGG) substrate.[@Uehara2009] The MOL is covered by a 100 nm-thick Al mirror layer and a 10 nm-thick TiO$_{2}$ protective layer. It is placed face-down on the superconducting film (Fig. \[fig2\]), whence it is observed through the transparent substrate using a polarized microscope with nearly crossed polarizers. In this configuration, the reflected light intensity increases as function of the local magnetic induction perpendicular to the garnet. Bright regions in the MO image correspond to regions of high magnetic flux density, while dark areas correspond to small or zero induction. This allows for the direct observation of magnetic flux penetration into the YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ films. The calibration of the luminous intensity in the absence of the superconductor, or measured at a point that is sufficiently far removed from the superconductor, allows one to convert the spatially resolved intensity maps to maps of the absolute value of the magnetic induction. Simultaneous MOI and Surface Resistance measurements ---------------------------------------------------- In order to measure the YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ film’s surface resistance during MOI, a number of specific modifications are required with respect to standard MOI and standard $R_{s}$ measurements. These are sketched in Fig. \[fig3\]. First of all, we have chosen to use the dielectric resonator technique with a rutile-phase TiO$_{2}$ resonator (of diameter 7 mm and height 3 mm) operating at 10 GHz for the $R_{s}$ measurements.[@Kermorvant2009] However, the presence of the MOL, in close contact with the superconducting film, prohibits one from placing the dielectric resonator directly on the film. The resonator was therefore installed on the side of the MgO substrate. The resonator is pressed to the substrate by the bottom lid of the square Cu cavity (of width 30 mm and height 7 mm) in which the whole assembly is placed. The presence of the substrate between the resonator and the film shifts the resonant frequency from 10 GHz to 7 GHz. The MgO substrate exhibits a low microwave loss tangent, $\tan \delta = 9 \times 10^{-6}$, and high thermal conductivity, $\kappa(90\,\,{\mathrm K}) = 290$ Wm$^{-1}$K$^{-1}$.[@Cahill] Hence, it does not introduce additional losses, and negligible measurement error on both the temperature and the resonant frequency when the temperature of the superconductor rises. The microwave field is excited using the sweeper of a HP 8510C vector network analyzer through a coupling loop in the Cu cavity, the position of which can be adjusted at low temperature using an XYZ stage. ![(Color online) (a) MOI of the flux distribution in the square YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ film at $H_{a} = 350$ Oe and $T = 50$ K. (b) Profiles of the magnetic induction $B$ along the horizontal line in (a), for successive applied fields between 0 and 250 Oe. (c) MOI of the flux distribution in the film at $T = 70$ K and $H_{a} =150$ Oe, applied after zero field cooling to 30 K and successive warming. (d) $B$–profiles taken along the horizontal line in (c), for different temperatures during warming to 70 K. []{data-label="fig4"}](Figure4.pdf){width="0.99\linewidth"} In usual MOI, the object to be imaged is cooled via direct thermal contact with a Cu sample holder attached to the cold head of the cryostat. For simultaneous $R_{s}$/MOI measurements, the intercalation of the dielectric resonator, with its low thermal conductivity,[@Thurber1965] prohibits this configuration. Thus, thermal contact is made via the top of the imaged specimen, through the top lid of the Cu cavity. This imposes a reduction of the magneto-optically imaged area of the sample under study. Imaging is performed through a circular aperture of diameter 8 mm in the cavity top lid. The presence of the Al mirror layer on the MOL does not pose additional problems, since it is shielded from the TiO$_{2}$ resonator by the YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ film. Results ======= MOI in dc magnetic field ------------------------- Figures \[fig4\](a,c) present standard MO images obtained on a square-shaped $5\times5$ mm$^{2}$ YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ film. The panels (a,c) illustrate the distribution of the magnetic induction in the superconducting film during the application of a constant magnetic field $H_{a} = 350$ Oe perpendicular to the film plane at the cavity temperature of 50 K, while panels (b,d) show the effect of increasing the cavity temperature after zero field cooling to 30 K, the application of $H_{a} = 150$ Oe and subsequent warming to 70 K. In all cases, the flux distribution accurately corresponds to the predictions of the Bean model,[@Bean1962; @Brandt1993; @Zeldov1994] which has that in the flux-penetrated areas of the film, the screening current density can only take on the value $\pm j_{c}$. The magnetic flux penetrates the superconductor from the edge of the sample, and is distributed according to the characteristic pillow-like shape expected for the divergence-free flow of the critical current in the thin film.[@Brandt1995] The flux distribution only depends on the parameter $H_{a}/j_{c}$, so that increasing either $H_{a}$ or temperature (with the concomitant decrease of $j_{c}$) both lead to a progression of the flux front to the film center. The position of the flux front, $x_{f}$, verifies the relation $$x_{f} = w \left[ 1 - \frac{1}{\cosh \left( \pi H_{a} /j_{c} d \right)} \right]$$ in which $w$ is the half-width of the YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ film. ![ (a) Surface resistance at 7 GHz as function of the input microwave power at 60 K, for zero applied field, and $H_{a} = 130$ Oe. (b) Resonant frequency as a function of the input microwave power at 60 K. []{data-label="fig5"}](Figure5.pdf){width="0.9\linewidth"} ![(Color online) (a) Schematic representation of the position of the obtained image relatively to the sample edge. The open green contour denotes the outline of the SC film. (b) MO image obtained after the sample was zero field cooled to 60 K, and a microwave field with input power of 22 dBm was applied. []{data-label="fig6"}](Figure6.pdf){width="1.0\linewidth"} ![(Color online) (a) Schematic representation of the position of the obtained image relative to the YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ thin film outline (open green contour). (b) MO image obtained after the sample was zero-field cooled to 60 K, and a dc field of 130 Oe applied.[]{data-label="fig7"}](Figure7.pdf){width="1.0\linewidth"} MOI with applied microwave magnetic field ----------------------------------------- In a first experiment, we have applied a microwave field to the superconductor, in the absence of a dc field. Since any magnetic flux should penetrate the superconducting film from the edge, we have used the right-hand Cu cavity cover of Fig. \[fig3\]. The film was cooled in Earth’s magnetic field, down to 60 K. After stabilization of the target temperature, the $TE_{011}$–mode of the dielectric resonator was excited in the low-power regime (-10 dBm applied to the coupling loop). In this mode, the rf electric field is parallel to the plane of the superconducting film, while the rf magnetic field is perpendicular to it. Next, the microwave power was increased from -10 dBm to +24 dBm in steps of 3 dBm. At each input power level, the resonator frequency and $Q$-factor were measured and the corresponding MO image recorded. Figure \[fig5\](a) renders the microwave power dependence of the surface resistance $R_{s}$, while Fig. \[fig5\](b) shows the power dependence of the resonant frequency $f_{0}$. No qualitative difference was observed with respect to the previously performed measurements at 10 GHz, with the resonator placed in direct contact with the superconducting film.[@Kermorvant2009] Both $R_{s}$ and the resonant frequency increase with rf power, indicating the nonlinear increase of microwave losses in the high microwave power regime. Figure \[fig6\](b) shows the MO image obtained at an input power level of 24 dBm. According to the $R_{s}$($P_{rf}$) characteristics in Fig. \[fig5\](a), the superconductor is clearly in the lossy regime. However, MOI does not reveal any vortex penetration into the superconductor, even in the high power regime. ![(a) Increase of the resonance frequency of the TiO$_{2}$ resonator as function of temperature, measured in the absence of the superconducting film. (b) Temperature increase of the superconducting film during the swept-power experiments of Fig. \[fig5\], as deduced from (a). []{data-label="fig8"}](Figure8.pdf){width="0.9\linewidth"} : 1. No microwave field; 2. 10 dBm; 3. 14 dBm; 4. 17 dBm; 5. 20 dBm; 6. 24 dBm.[]{data-label="fig9"}](Figure9.pdf){width="1.0\linewidth"} ![(Color online) Identical flux magnetic flux penetration into the YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ thin film exposed to a dc field of 130 Oe (a) at the sample holder temperature $T = 60$ K, with applied microwave power $P_{rf}$ = 17 dBm; (b) at the higher temperature of the sample holder $T = 62.586 $ K, chosen such that $\Delta f = 33.245$ MHz as in (a), but with $P_{rf}$ = 10 dBm. []{data-label="fig9"}](Figure10.pdf){width="1.0\linewidth"} MOI with microwave and dc magnetic field ---------------------------------------- A second series of experiments was performed by adding a constant magnetic field to the rf field. The superconducting sample was again zero-field cooled to 60 K, following which a dc magnetic field of 130 Oe was applied perpendicularly to the film surface. This leads to the magnetic flux penetration illustrated in Fig. \[fig7\]. Next, the microwave input was switched on in order to excite the dielectric resonator. The nominal temperature of the cold head and the dc magnetic field remain constant during the experiment. The microwave power was gradually increased in steps of approximately 3 dBm. At each input microwave power level the $Q$-factor and the resonant frequency of the resonator was measured and the corresponding MO image recorded. The obtained results concerning the power dependence of the surface resistance and the resonant frequency are again plotted in Figure \[fig5\]. We observe the usual power dependence of the surface resistance of YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ thin films, [*i.e.*]{}, a nonlinear increase of $R_{s}$ as function of $P_{rf}$. As shown by temperature dependence of $f_{0}$ depicted in Fig. \[fig8\], the simultaneous increase of the resonant frequency of the dielectric, here, by up to $\delta f = 33.245$ MHz for the highest injected power, can be understood as the result of local heating of the YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ film.[@Kermorvant2009] The simultaneously acquired MO images show no progression of the flux front in the linear regime of $R_{s}( P_{rf})$, but pronounced enhanced dc magnetic flux penetration in the nonlinear regime. A mechanism for ac magnetic field–induced dc flux penetration into thin superconducting films, the so-called vortex lattice shaking, was presented in Refs. . The conditions for vortex shaking to be effective are that the ac field strength be sufficient to drive the film into the critical state, which is the case here, and that the rf screening currents be of opposite polarity on the top and bottom film surfaces. In the present experimental configuration, the latter condition is not satisfied: the use of the $TE_{011}$ cavity mode means that the rf magnetic field to which the film is subjected induces screening currents of the same polarity on the top and bottom film surface, opposite in direction to the electric field in the rutile cavity. We therefore surmise that the progression of the flux front is due to the decrease of the critical current density $j_{c}$ associated with the increase in temperature of the YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ film, a conjecture that is checked by increasing the nominal temperature of the cold head and the sample holder using the incorporated heater, all the while maintaining a low microwave input power of 10 dBm. The temperature is adjusted so as to precisely yield a frequency change of the resonator $\delta f = 33.245$ MHz. We find that this corresponds to a temperature increase of $\Delta T = 2.586$ K. The MOI (Fig.  \[fig9\]) shows that the flux penetration precisely corresponds to that previously obtained by increasing the microwave power, showing that local heating is indeed at the origin of the enhanced flux penetration and the nonlinear dependence $R_{s}$ ( $P_{rf}$ ). Summary and Conclusions ======================= We have developed a experimental set-up that allows for the simultaneous imaging of magnetic flux penetration into superconducting samples using the magneto-optical technique, and the measurement of their surface resistance in the range 1 – 10 GHz. The method was applied to YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ thin films. No signature of the microwave magnetic field could be observed in magneto-optics. However, the application of a high-power microwave magnetic field significantly enhances dc flux penetration due to local heating of the superconducting film. This work was partially funded by the French National Research Agency ANR, under contract number ANR-07-BLAN-0242 ÔÔSURFÕÕ. [99]{} D. Oates, P.P. Nguyen, G. Dresselhaus, M.S. Dresselhaus, C.W. Lam, S.M. Ali, J. Superconductivity [**5**]{}, 361 (1992). Daniel E. Oates, Paul P. Nguyen, Gene Dresselhaus, Mildred S. Dresselhaus, Gad Koren and Emil Polturak, J. Superconductivity [**8**]{}, 725 (1995). T. B. Samoilova, Supercon. Sci. Technol. [**8**]{}, 259 (1995). G. Hampel, P. Kolodner, P.L. Gammel, P.A. Polakos, E. de Obaldia, P.M. Mankiewich, A. Anderson, R. Slattery, D. Zhang, G.C. Liang, C.F. Shi, Applied Physics Letters [**69**]{}, 571 (1996). Jaroslaw Wosik, Lei-ming Xie, Krzysztof Nesteruk, Dawei Li, John H Miller, Stuart A Long, J. Superconductivity [**10**]{}, 97 (1997). Steven M. Anlage, Wensheng Hu, C. P. Vlahacos, David Steinhauer, B. J. Feenstra, Sudeep K. Dutta, Ashfaq Thanawalla, and F. C. Wellstood, J. Supercond. [**12**]{}, 3 53-362 (1999) . J. Kermorvant, C.J. van der Beek, J.-C Mage, B. Marcilhac, Y. Lemaître, J. Briatico, R. Bernard, J. Villegas, J. Appl. Phys. [**106**]{}, 023912 (2009). P. Lahl and R. Wördenweber, J. Appl. Phys. [**97**]{}, 113911 (2005). T. Dahm and D.J. Scalapino, Phys. Rev. B [**60**]{}, 13125 (1999). M. Hein, W. Diete, M. Getta, S. Hensen, T. Kaiser, G. Müller, H. Piel, and H. Schilk, IEEE Trans. Appl. Supercond. [**7**]{}, 1264 (1997). D. C. Mattis and J. Bardeen, Phys. Rev. [**111**]{}, 412Ð417 (1958). John Bardeen and M. J. Stephen, Phys. Rev. [**140**]{}, A1197 (1965). Ernst Helmut Brandt, Phys. Rev. Lett. [**67**]{}, 2219 (1991) Mark W. Coffey and John R. Clem, Phys. Rev. Lett. [**67**]{}, 386 (1991). Mark W. Coffey and John R. Clem, Phys. Rev. B [**46**]{}, 11757 (1992). Mark W. Coffey and John R. Clem, Phys. Rev. B [**48**]{}, 342 (1993). Mark W. Coffey, Phys. Rev. B [**47**]{}, 15298 (1993). Ernst Helmut Brandt and Grigorii P. Mikitik, Phys. Rev. Lett. [**89**]{}, 027002-1(2002). Grigorii P. Mikitik and Ernst Helmut Brandt, Phys. Rev. B [**67**]{}, 134521 (2003). Grigorii P. Mikitik and Ernst Helmut Brandt, Phys. Rev. B [**69**]{}, 104511 (2004). G. P. Mikitik and E. H. Brandt, Journal of Low Temperature Physics [**139**]{}, 221 (2005). J. Wosik, C. Wang, L.-M. Xie, T.H. Johansen, Q.Y. Chen, and W.-K. Chu, Applied Superconductivity Conf., Aug. 4-9, 2002, Houston, TX; 15 L.A. Dorosinskii, M.V. Indenbom, V.I. Nikitenko, Yu.A. OssipÕyan, A.A. Polyanskii, and V.K. Vlasko-Vlasov, Physica C [**203**]{}, 149 (1992). Ch. Jooss, J Albrecht, H Kuhn, S Leonhardt and H Kronmüller, Rep. Prog. Phys. [**65**]{}, 651Ð788 (2002). M. Uehara, C.J. van der Beek, J. Gattacecca, V. Skidanov, Y. Quesnel, Geochemistry, Geophysics, Geosystems [**11**]{}, Q05Y09 (2009). See http://users.mrl.uiuc.edu/cahill/tcdata/tcdata.html W.R. Thurber, A.J. Mante, Physical Review [**139**]{}, 1655 (1965). C.P. Bean, Phys. Rev. Lett. [**8**]{}, 6 (1962). E.H. Brandt, M.V. Indenbom and A. Forkl, Europhys. Lett. [**22**]{}, 735 (1993). E. Zeldov, J.R. Clem, M. McElfresh and M. Darwin, Phys. Rev. B [**49**]{}, 9802 (1994). Ernst Helmut Brandt, Phys. Rev. Lett. [**74**]{}, 3025 (1995).

ArXiv

--- author: - | Maatoug Hassine$^1$ and Imen Kallel$^2$\ $^{1}$ FSM, Monastir University, Tunisie\ $^{2}$ Monastir University, Tunisie title: '**A fast reconstruction algorithm for geometric inverse problems using topological sensitivity analysis and Dirichlet-Neumann cost functional approach**' --- [**Abstract.**]{} This paper is concerned with the detection of objects immersed in anisotropic media from boundary measurements. We propose an accurate approach based on the Kohn-Vogelius formulation and the topological sensitivity analysis method. The inverse problem is formulated as a topology optimization one minimizing an energy like functional. A topological asymptotic expansion is derived for the anisotropic Laplace operator. The unknown object is reconstructed using a level-set curve of the topological gradient. The efficiency and accuracy of the proposed algorithm are illustrated by some numerical results. [**Keywords.**]{} geometric inverse problem, anisotropic Laplace, Kohn-Vogelius formulation, sensitivity analysis, topological optimization. [**2010 AMS subject classification.**]{} 49Q12, 65N21, 35N10 ¶ Ø [IR]{} Introduction ============ In this work we are interested on the detection of objects immersed in an anisotropic media from overdetermined boundary data. More precisely, let $\Omega\subset {I\!\!R}^d,\,\, d=2,3$ be a bounded domain with smooth boundary $\Gamma$. We assume that $\Gamma$ is partitioned into two parts $\Gamma_a$ (accessible) and $\Gamma_i$ (inaccessible) having both non-vanishing measure. Let $\O$ be an unknown object immersed inside the background domain $\Omega$ and having a smooth boundary $\Sigma = \partial \O$. The geometric inverse problem that we consider here can be formulated as follows: $-$ Giving two boundary data on $\Gamma_a$; an imposed flux $\Phi\in H^{-1/2}(\Gamma_a)$ and a measured datum $\psi_m\in H^{1/2}(\Gamma_a)$. $-$ Find the unknown location of the object $\O$ inside the domain $\Omega$ such that the solution $\psi$ of the anisotropic Laplace equation satisfies the following overdetermined boundary value problem $$\left \lbrace \begin{array}{rll} -\mbox{ div } (\gamma(x) \nabla \psi)&=F & \mbox{ in } \Omega\backslash \overline{\O},\\ \gamma(x)\nabla \psi \cdot{\bf n }&=\Phi & \mbox{ on } \Gamma_a,\\ \psi &=\psi_m & \mbox{ on } \Gamma_a,\\ \psi &=0 & \mbox{ on } \Gamma_i,\\ \gamma(x) \nabla\psi \cdot{\bf n } &= 0 & \mbox{ on } \Sigma, \end{array} \right.$$ where $\gamma$ is a scalar smooth function (of class $\mathcal C^1$) describing the physical properties of the medium $\Omega$, ${\bf n }$ is the exterior unit normal vector and $F\in L^2(\Omega)$ is a given source term. We assume that there exist two constants $c_0>0$ and $c_1>0$ such that $c_0\leq \gamma(x)\leq c_1,\,\forall x\in \Omega$. In this formulation the domain $\Omega\backslash \overline{\O}$ is unknown since the free boundary $\Sigma$ is unknown. This problem is ill-posed in the sense of Hadamard. The majority of works dealing with this kind of problems fall into the category of shape optimization and based on the shape differentiation technics. It is proved in [@BCD] that the studied inverse problems, treated as a shape optimization problems, are severely ill-posed (i.e. unstable), for both Dirichlet and Neumann conditions on the boundary $\Sigma$. Thus they have to use some regularization methods to solve them numerically. To solve this inverse problem, we extend the topological sensitivity analysis notion to the anisotropic case and we suggest an alternative approach based on the Kohn-Vogelius formulation [@ABB] and the topological gradient method [@AHM; @AK; @BHJM; @GH; @GB; @HJM]. We combine here the advantages of the Kohn-Vogelius formulation as a self regularization technique and the topological gradient approach as an accurate and fast method. In the first part of this paper, we derive a topological sensitivity analysis for a Kohn-Vogelius type functional valid for an arbitrary shaped geometric perturbation. The obtained results are based on a preliminary estimate describing the perturbation caused by the presence of a small geometry modification of the background domain $\Omega$. The proposed mathematical analysis is general and can be adapted for various partial differential equations (PDE). The second part of this paper is concerned with some numerical investigations. The obtained topological sensitivity function is used to built a simple, fast and accurate geometry reconstruction algorithm. The efficiency of the proposed algorithm is illustrated by some numerical results. This work is motivated by many practical problems for which it is necessary to detect the electrical properties of a media from boundary measurements. This kind of studies was realyzed for the clinical applications such as electrical impedance tomography [@CIN], the geophysical applications such as detection of the mineral deposits location in the earth [@parker], industrial applications such as non-destructive testing [@bowler], ... etc. The rest of the paper is organized as follows. In Section 2, we formulate the considered inverse problem as a topological optimization one. In Section 3, we discuss the perturbation caused by the presence of a small object inside the background domain $\Omega$. The Section 4 is devoted to the Kohn-Vogelius type function variation. In Section 5, we derive a topological sensitivity analysis for the anisotropic Laplace operator. In Section 6, we propose a one-iteration shape reconstruction algorithm. Formulation of the inverse problem ================================== In this section, we give the main steps of our analysis. Firstly, we introduce the Kohn-Vogelius formulation and we define the cost function to be minimized. Secondly, we present the perturbed problems and we describe the quantity to be estimated. The Kohn-Vogelius formulation ----------------------------- The Kohn-Vogelius formulation rephrases the considered inverse problem into a topological optimization one. In fact, the Kohn-Vogelius formulation leads to define for any given domain $\O\subset \Omega $ two forward problems. The first one is associated to the Neumann datum $\Phi$, it will be named the “Neumann problem”: $$\label{neum-pb} (\PB_N)\left\{ \begin{array}{l} \mbox{Find } \psi_N \in H^1(\Omega\backslash \overline{\O}) \mbox{ solving }\\ \begin{array}{rll} \displaystyle - \mbox{ div } (\gamma(x) \nabla \psi_N) =& F &\hbox{ in } \Omega \backslash \overline{\O}\\ \gamma(x) \nabla \psi_N \cdot{\bf n } =& \Phi &\hbox{ on } \Gamma_a \\ \psi_N =& 0 &\hbox{ on } \Gamma_i \\ \gamma(x) \nabla \psi_N \cdot{\bf n } =& 0 &\hbox{ on } \Sigma. \end{array} \end{array} \right.$$ The second one is associated to the Dirichlet datum (measured) $\psi_m$ $$\label{dir-pb} (\PB_D) \left\{ \begin{array}{l} \mbox{Find } \psi_D \in H^1(\Omega\backslash \overline{\O}) \mbox{ solving }\\ \begin{array}{rll} \displaystyle - \mbox{ div } (\gamma(x) \nabla \psi_D) =& F &\hbox{ in } \Omega \backslash \overline{\O}\\ \psi_D =& \psi_m &\hbox{ on } \Gamma_a\\ \psi_D =& 0 &\hbox{ on } \Gamma_i \\ \gamma(x) \nabla \psi_D \cdot{\bf n }=& 0 &\hbox{ on } \Sigma. \end{array} \end{array} \right.$$ One can see that if $\Sigma$ coincides with the actual boundary $\Sigma^*$ then the misfit between the solutions vanishes, ${\psi_D} = {\psi_N}$. According to this observation, we propose an identification process based on the minimization of the following energy like functional [@ABB] $$\begin{aligned} \label{misfit-func} \K (\Omega \backslash \overline{\O}) = \int_{\Omega \backslash \overline{\O}}\gamma(x)|\nabla \psi_D - \nabla \psi_N|^2 dx.\end{aligned}$$ The inverse problem can be formulated as a topological optimization one $$\min_{\O \subset \Omega} \K (\Omega \backslash \overline{\O}).$$ To solve this problem we will use the topological sensitivity analysis method. The sensitivity analysis method ------------------------------- This method consists in studying the variation of the function $\K$ with respect to the presence of a small object inside the background domain $\Omega$. To present the main idea of this method, we consider the case in which $\Omega$ contains a small object $\O_{z,\eps}$ that is centred at $z\in \Omega$ and has the shape $\O_{z,\eps}=z+\eps \omega\subset \Omega$, where $\eps>0$ and $\omega\subset {I\!\!R}^d$ is a given, fixed and bounded domain containing the origin, whose boundary $\partial \omega$ is of $\mathcal{C}^1$. The topological sensitivity analysis leads to an asymptotic expansion of the variation $\K (\Omega \backslash \overline{\O_{z,\eps}})-\K (\Omega)$ with respect to $\eps$. Using the Kohn-Vogelius formulation, one can define for each arbitrary location of $\O_{z,\eps}$ in the domain $\Omega$, two forward problems. The first one is associated to the Neuman datum $\Phi$ and called “the perturbed Neumann problem” $$\label{npb} (\PB^\eps_N)\left\{ \begin{array}{rll} \mbox{find } \psi^\eps_N \in H^1(\Omega \backslash \overline{\O_{z,\eps}})& \mbox{such that}\\ \displaystyle - \mbox{ div } (\gamma(x) \nabla \psi^\eps_N) =& F &\hbox{ in } \Omega \backslash \overline{\O_{z,\eps}} \\ \gamma(x) \nabla \psi^\eps_N \cdot{\bf n } =& \Phi &\hbox{ on } \Gamma_a \\ \psi^\eps_N =& 0 &\hbox{ on } \Gamma_i \\ \gamma(x) \nabla \psi^\eps_N \cdot{\bf n }=& 0 &\hbox{ on }\partial\O_{z,\eps}, \end{array} \right.$$ The second one is associated to the Dirichlet datum $\psi_m$ and called “the perturbed Dirichlet problem” $$\label{dpb} (\PB^\eps_D) \left\{ \begin{array}{rll} \mbox{find } \psi^\eps_D \in H^1(\Omega \backslash \overline{\O_{z,\eps}})& \mbox{such that}\\ \displaystyle - \mbox{ div } (\gamma(x) \nabla \psi^\eps_D) =& F &\hbox{ in } \Omega \backslash \overline{\O_{z,\eps}}\\ \psi^\eps_D =& \psi_m &\hbox{ on } \Gamma_a\\ \psi^\eps_D =& 0 &\hbox{ on } \Gamma_i \\ \gamma(x) \nabla \psi^\eps_D \cdot{\bf n }=& 0 &\hbox{ on } \partial\O_{z,\eps}. \end{array} \right.$$ In order to describe the presence of the object $\O_{z,\eps}$ inside the domain $\Omega$, we will use the shape function $$\begin{aligned} \K (\Omega \backslash \overline{\O_{z,\eps}}) =\int_{\Omega \backslash \overline{\O_{z,\eps}}}\gamma(x)|\nabla \psi^\eps_D - \nabla \psi^\eps_N|^2\, dx,\end{aligned}$$ Next, we will derive a topological sensitivity analysis for the function $\K$ with respect to the insertion of a small object $\O_{z,\eps}$ in $\Omega$. It leads to an asymptotic expansion of the form $$\begin{aligned} \K (\Omega \backslash \overline{\O_{z,\eps}})=\K (\Omega)+ \rho(\eps) \delta \K(z) + o(\rho(\eps)),\quad \forall z\in \Omega,\end{aligned}$$ where $\eps \longmapsto\rho(\eps)$ is a scalar positive function going to zero with $\eps$. The function $z\longmapsto \delta \K(z)$ is called the topological gradient and play the role of leading term of the variation $\K (\Omega \backslash \overline{\O_{z,\eps}})-\K (\Omega)$. In order to minimize the shape function $\K$, the best location $z$ of the object $\O_{z,\eps}$ in $\Omega$ is where $\delta\K$ is most negative. We start our analysis by estimating the perturbation caused by the presence of the small object $\O_{z,\eps}$ in the background domain $\Omega$. We will establish in the next section two estimates describing the behavior of the perturbed solutions with respect to $\eps$. In Section \[cost-funct-variation\], we will discuss the cost function variation. Based on the obtained estimates, we will derive in Section \[asymptotic\] a topological asymptotic expansion for the anisotropic laplace operator. Estimate of the perturbed solutions {#sec-estimate} =================================== In this paragraph, we establish two estimates describing the perturbation caused by the presence of the geometry modification $\O_{z,\eps}$ on the solutions of the Dirichlet and Neumann problems. To this end, we introduce two auxiliaries problems.\ The first one is related to the Neumann problem: $$\label{ext-neum} \left\{ \begin{array}{l} \mbox{Find } \varphi_N \in W^1(\mathbb{R}^d\backslash \overline{\omega}) \mbox{ such that }\\ \begin{array}{rll} \displaystyle - \Delta \varphi_N =& 0 &\hbox{ in } \mathbb{R}^d \backslash \overline{\omega}\\ \varphi_N =& 0 &\hbox{ at } \infty \\ \nabla \varphi_N \cdot{\bf n } =& - \nabla \psi_N (z)\cdot{\bf n }&\hbox{ on } \partial \omega . \end{array} \end{array} \right.$$ The second one is related to the Dirichlet problem: $$\label{ext-neum} \left\{ \begin{array}{l} \mbox{Find } \varphi_D \in W^1(\mathbb{R}^d\backslash \overline{\omega}) \mbox{ such that }\\ \begin{array}{rll} \displaystyle - \Delta \varphi_D =& 0 &\hbox{ in } \mathbb{R}^d \backslash \overline{\omega}\\ \varphi_D =& 0 &\hbox{ at } \infty \\ \nabla \varphi_D \cdot{\bf n } =& - \nabla \psi_D (z) \cdot{\bf n }&\hbox{ on } \partial \omega . \end{array} \end{array} \right.$$ The functions $\varphi_N$ and $\varphi_N$ can be expressed by a single layer potential on $\partial \omega$ (see [@DL]) on the following way $$\begin{aligned} \varphi_N(y)&=& \displaystyle \int_{\partial \omega} E(y-x)\eta_N(x)ds(x),\forall y \in \mathbb{R}^d \backslash \overline{\omega}\\ \varphi_D(y)&=& \displaystyle \int_{\partial \omega} E(y-x)\eta_D(x)ds(x),\, \forall y \in \mathbb{R}^d \backslash \overline{\omega},\end{aligned}$$ where $E$ is the fundamental solution of the Laplace problem in $\mathbb{R}^d$; $$E(y)= \left\{ \begin{array}{ll} -\displaystyle \frac{1}{2\pi} \log(|y|) &\mbox{ if }\,\,d=2,\\ \displaystyle \frac{1}{4\pi}\frac{1}{|y|}&\mbox{ if }\,\,d=3. \end{array} \right.$$ Here $\eta_N$ and $\eta_D$ belong to $H^{-1/2}(\partial \omega)$ and solve the following integral equations [@DL] $$\begin{aligned} \label{integ-N} && -\frac{\eta_N(y)}{2}+\int_{\partial\omega}\nabla E(y-x) \cdot {\bf n }\,\eta_N(x)ds(x)=-\nabla \psi^0_N(z) \cdot {\bf n },\quad y\in \partial \omega \\\label{integ-D} &&-\frac{\eta_D(y)}{2}+\int_{\partial\omega}\nabla E(y-x) \cdot {\bf n }\,\eta_D(x)ds(x)=-\nabla \psi^0_D(z) \cdot {\bf n },\quad y\in \partial \omega.\end{aligned}$$ The Neumann and Dirichlet perturbed solutions satisfy the following estimates. \[est-pert-sol\] There exist positive constants $c>0$, independent of $\eps$, such that $$\begin{aligned} \left\| \psi^\eps_N-\psi^0_N-\eps\varphi_N((x-z)/\eps) \right\|_{1,\Omega_{z,\eps}}\leq c\eps^{d/2},\\ \\ \left\| \psi^\eps_D-\psi^0_D-\eps\varphi_D((x-z)/\eps) \right\|_{1,\Omega_{z,\eps}}\leq c\eps^{d/2}.\end{aligned}$$ [Proof:]{} In order to prove the estimates established for the Neumann and Dirichlet perturbed solutions, we consider the following generic problem $$\label{auxil-eps} \left\{ \begin{array}{l} \mbox{Find }\psi^\eps \in H^1(\Omega_{z,\eps}) \mbox{ such that }\\ \begin{array}{rll} \displaystyle - \mbox{ div } (\gamma(x) \nabla \psi^\eps) =& F &\hbox{ in } \Omega_{z,\eps} \\ \gamma(x) \nabla \psi^\eps{\cdot n } =& \Phi &\hbox{ on } \Gamma_n \\ \psi^\eps =& \psi_m &\hbox{ on } \Gamma_d \\ \psi^\eps =& 0 &\hbox{ on } \Gamma_i \\ \gamma(x) \nabla \psi^\eps {\cdot n }=& 0 &\hbox{ on }\partial\O_{z,\eps}, \end{array} \end{array} \right.$$ where $\Gamma_n$ and $\Gamma_d$ are two parts of the boundary $\Gamma_a$ such that $\overline{\Gamma_a}=\overline{\Gamma_n}\cup\overline{\Gamma_d}$ and $\Gamma_n\cap\Gamma_d=\emptyset$. One can remark here that the considered problem (\[auxil-eps\]) has a general forme valid for the Neumann and Dirichlet cases. In fact, if $\Gamma_n=\emptyset$, we have $\Gamma_d=\Gamma_a$ and $\psi^\eps$ solves the Dirichlet problem $ (\PB^\eps_D)$. If $\Gamma_d=\emptyset$, we have $\Gamma_n=\Gamma_a$ and $\psi^\eps$ solves the Neumann problem $ (\PB^\eps_N)$. In the absence of any geometry perturbation (i.e. $\eps=0$), we have $\Omega_{z,\eps} =\Omega$ and $\psi^0$ solves $$\label{auxil-0} \left\{ \begin{array}{rll} \displaystyle - \mbox{ div } (\gamma(x) \nabla \psi^0) =& F &\hbox{ in } \Omega \\ \gamma(x) \nabla \psi^0 \cdot {\bf n } =& \Phi &\hbox{ on } \Gamma_n \\ \psi^0 =& \psi_m &\hbox{ on } \Gamma_d \\ \psi^0 =& 0 &\hbox{ on } \Gamma_i. \end{array} \right.$$ We denote by $\varphi$ the solution to the associated exterior problem $$\label{auxil-ext} \left\{ \begin{array}{rll} \displaystyle - \Delta \varphi =& 0 &\hbox{ in } \mathbb{R}^d \backslash \overline{\omega}\\ \varphi =& 0 &\hbox{ at } \infty \\ \nabla \varphi \cdot{\bf n } =& - \nabla \psi^0 (z)\cdot{\bf n }&\hbox{ on } \partial \omega . \end{array} \right.$$ Combining (\[auxil-eps\]), (\[auxil-0\]) and (\[auxil-ext\]), one can deduce that $\phi^\eps=\psi^\eps-\psi^0-\eps \varphi((x-z)/\eps)$ is solution to $$\left\{ \begin{array}{rll} \displaystyle - \mbox{ div } (\gamma(x) \nabla \phi^\eps) =& \nabla\gamma(x).\nabla_y \varphi((x-z)/\eps) &\hbox{ in } \Omega_{z,\eps} \\ \gamma(x) \nabla \phi^\eps \cdot{\bf n } =& -\gamma(x) \nabla_y \varphi((x-z)/\eps)\cdot{\bf n } &\hbox{ on } \Gamma_n \\ \phi^\eps =& -\eps \varphi((x-z)/\eps) &\hbox{ on } \Gamma_d\cup\Gamma_i \\ \gamma(x) \nabla \phi^\eps \cdot{\bf n }=& -\gamma(x) [\nabla \psi^0- \nabla \psi^0(z)]\cdot{\bf n } &\hbox{ on }\partial\O_{z,\eps}. \end{array} \right.$$ Due to the smoothness of $\gamma$ in $\Omega$, there exists $c>0$ such that $$\begin{aligned} \left\| \nabla\gamma(x).\nabla_y \varphi ((x-z)/\eps) \right\|_{0,\Omega_{z,\eps}} &\leq c & \left\|\nabla_y \varphi ((x-z)/\eps) \right\|_{0,\Omega_{z,\eps}},\,\\ \left\| \gamma(x) \nabla_y \varphi((x-z)/\eps)\cdot{\bf n }\right\|_{-1/2,\Gamma_n} &\leq c & \left\| \nabla_y \varphi((x-z)/\eps) \cdot{\bf n }\right\|_{-1/2,\Gamma_n}.\end{aligned}$$ By trace Theorem, it follows $$\begin{aligned} \left\| \varphi((x-z)/\eps)\right\|_{1/2,\Gamma_d\cup\Gamma_i}+\left\| \nabla_y \varphi((x-z)/\eps)\cdot{\bf n }\right\|_{-1/2,\Gamma_n} \leq c\,\left\| \varphi((x-z)/\eps)\right\|_{1,\Omega_R},\end{aligned}$$ where $\Omega_R=\Omega\backslash\overline{B(z,R)}$, with $R>0$ is a given radius such that $\overline{\O_{z,\eps}} \subset B(z,R)$ and $\overline{B(z,R)}\subset \Omega$. It is easy to check that the function $\varphi^\eps(x)=\varphi((x-z)/\eps)$ solves the following problem $$\label{auxil-ext-eps} \left\{ \begin{array}{rll} \displaystyle - \Delta \varphi^\eps =& 0 &\hbox{ in } \mathbb{R}^d \backslash \overline{\O_{z,\eps}}\\ \varphi^\eps =& 0 &\hbox{ at } \infty \\ \nabla \varphi^\eps \cdot{\bf n } =& - \nabla \psi^0 (z)\cdot{\bf n }&\hbox{ on } \partial \O_{z,\eps}. \end{array} \right.$$ Since $\displaystyle \int_{\partial \O_{z,\eps}} \nabla \psi^0 (z)\cdot{\bf n }ds=0$, using the change of variable $x=z+\eps y$, one can prove that there exists $c>0$ (independent of $\eps$) such that $$\begin{aligned} \label{est-1} \left\|\nabla_y \varphi^\eps \right\|_{0,\Omega_{z,\eps}}&&\leq c \eps^{d/2} \left\| \nabla \psi^0 (z)\cdot{\bf n }\right\|_{-1/2,\partial \omega},\\\label{est-2} \left\| \varphi^\eps\right\|_{1,\Omega_R} &&\leq c \eps^{d/2} \left\| \nabla \psi^0 (z)\cdot{\bf n }\right\|_{-1/2,\partial \omega}.\end{aligned}$$ The last estimates follow from a change of variable and the integral representation of the function $\varphi$. For more details and simular proof, one can consult [@GGM] for the elasticity problem or [@BHJM] for the Stokes problem. Now, we examine the boundary condition satisfied by $\phi^\eps$ on $\partial \O_{z,\eps}$. Using the smoothness of $\gamma$ and $\nabla\psi^0$ near the point $z$, one can derive $$\begin{aligned} \label{est-3} \left\| \gamma(x) [\nabla \psi^0(x)- \nabla \psi^0(z)]\cdot{\bf n }\right\|_{-1/2,\partial \O_{z,\eps}} \leq c\, \eps^{d/2}.\end{aligned}$$ Finally, combining (\[est-1\]), (\[est-2\]) and (\[est-3\]), one can deduce that the function $\phi^\eps=\psi^\eps-\psi^0-\eps \varphi((x-z)/\eps)$ satisfied the desired estimate $$\begin{aligned} \left\| \phi^\eps \right\|_{1,\Omega_{z,\eps}}\leq c\,\eps^{d/2}.\end{aligned}$$ Variation of the function $\K$ {#cost-funct-variation} ============================== This section is focused on the variation of the Kohn-Vogelius function $\K$ with respect to the presence of the small object $\O_{z,\eps}$ inside the domain $\Omega$. We will derive a simplified expression of the variation $\K (\Omega_{z,\eps})-\K (\Omega)$. The obtained result is presented in the following theorem. \[K-variation\] Let $\O_{z,\eps}$ be an arbitrary shaped object inserted inside the background domain $\Omega$. The variation $\K (\Omega_{z,\eps})-\K (\Omega)$ admits the expression $$\begin{aligned} \label{decomp-k} \nonumber \K (\Omega_{z,\eps})-\K (\Omega) &=& \int_{\partial\O_{z,\eps}}\gamma(x) \nabla \psi^0_D \cdot{\bf n } \, \psi^\eps_D\,ds -\int_{\partial\O_{z,\eps}}\gamma(x) \nabla \psi^0_N \cdot{\bf n } \, \psi^\eps_N\,ds \\ && + \int_{\O_{z,\eps}} F\,[\psi^0_D -\psi^0_N ]\,dx.\end{aligned}$$ [*Proof:*]{} From the definition of $\K$, we have $$\begin{aligned} \K (\Omega_{z,\eps})-\K (\Omega) &=& \int_{\Omega_{z,\eps}}\gamma(x)|\nabla \psi^\eps_D - \nabla \psi^\eps_N|^2\,dx- \int_{\Omega}\gamma(x)|\nabla \psi^0_D - \nabla \psi^0_N|^2\,dx\\ &=& \int_{\Omega_{z,\eps}}\gamma(x)| \nabla \psi^\eps_N|^2\,dx + \int_{\Omega_{z,\eps}}\gamma(x)|\nabla \psi^\eps_D|^2\,dx -2 \int_{\Omega_{z,\eps}}\gamma(x)\nabla \psi^\eps_D . \nabla \psi^\eps_N\,dx\\ && -\int_{\Omega}\gamma(x)| \nabla \psi^0_N|^2\,dx - \int_{\Omega}\gamma(x)|\nabla \psi^0_D|^2\,dx -2 \int_{\Omega}\gamma(x)\nabla \psi^0_D . \nabla \psi^0_N\,dx.\end{aligned}$$ Then, the variation $\K (\Omega_{z,\eps})-\K (\Omega)$ can be decomposed as $$\K (\Omega_{z,\eps})-\K (\Omega)= T_N(\eps)+T_D(\eps)-2 T_M(\eps),$$ where $T_N$ is the Neumann term $$T_N(\eps)= \int_{\Omega_{z,\eps}}\gamma(x)| \nabla \psi^\eps_N|^2\,dx - \int_{\Omega}\gamma(x)| \nabla \psi^0_N|^2\,dx,$$ $T_D$ is the Dirichlet term $$T_D(\eps) = \int_{\Omega_{z,\eps}}\gamma(x)|\nabla \psi^\eps_D|^2\,dx - \int_{\Omega}\gamma(x)|\nabla \psi^0_D|^2\,dx,$$ and $T_M$ is the mixed term $$T_M (\eps)= \int_{\Omega_{z,\eps}}\gamma(x)\nabla \psi^\eps_D . \nabla \psi^\eps_N\,dx - \int_{\Omega}\gamma(x)\nabla \psi^0_D . \nabla \psi^0_N\,dx.$$ Next, we shall examine each term separately.\ $-$ [*Calculate of the Neumann term.*]{} From the weak formulation of the problems $ (\PB^\eps_N)$ and $ (\PB^0_N)$ one can obtain $$\begin{aligned} T_N(\eps)&=& \int_{\Omega_{z,\eps}}F\,\psi^\eps_N\,dx + \int_{\Gamma_a}\Phi\,\psi^\eps_N \hbox{ds} - \int_{\Omega}F\,\psi^0_N\,dx - \int_{\Gamma_a}\Phi\,\psi^0_N \hbox{ds}\\ &=& \int_{\Omega_{z,\eps}}F (\psi^\eps_N - \psi^0_N)\,dx + \int_{\Gamma_a}\Phi (\psi^\eps_N - \psi^0_N) \hbox{ds} - \int_{\O_{z,\eps}}F\,\psi^0_N\,dx.\end{aligned}$$ Then, it follows $$\begin{aligned} T_N(\eps)&=& \int_{\Omega_{z,\eps}}\nabla \psi^\eps_N .\nabla (\psi^\eps_N - \psi^0_N)\,dx - \int_{\O_{z,\eps}}F\,\psi^0_N\,dx.\end{aligned}$$ Using the Green formula and the fact that $$\mbox{ div } (\gamma(x) \nabla (\psi^\eps_N - \psi^0_N)) =0 \hbox{ in } \Omega_{z,\eps},\,\, \psi^\eps_N - \psi^0_N=0 \hbox{ on }\Gamma_i\, \hbox{ and }\, \gamma(x) \nabla (\psi^\eps_N - \psi^0_N) \cdot{\bf n } = 0 \hbox{ on }\Gamma_a,$$ we deduce $$\begin{aligned} \nonumber T_N(\eps)&=& \int_{\partial\O_{z,\eps}} \gamma(x) \nabla (\psi^\eps_N - \psi^0_N) \cdot{\bf n } \psi^\eps_N \,ds - \int_{\O_{z,\eps}}F\,\psi^0_N\,dx,\\ \label{term-neumann} &=& - \int_{\partial\O_{z,\eps}} \gamma(x) \nabla \psi^0_N \cdot{\bf n } \psi^\eps_N \,ds - \int_{\O_{z,\eps}}F\,\psi^0_N\,dx.\end{aligned}$$ $-$ [*Calculate of the Dirichlet term.*]{} We have $$\begin{aligned} T_D(\eps) &=& \int_{\Omega_{z,\eps}}\gamma(x)|\nabla \psi^\eps_D|^2\,dx - \int_{\Omega}\gamma(x)|\nabla \psi^0_D|^2\,dx,\\ &=& \int_{\Omega_{z,\eps}}\gamma(x) \nabla (\psi^\eps_D - \psi^0_D)\nabla (\psi^\eps_D + \psi^0_D)\,dx - \int_{\O_{z,\eps}}\gamma(x)|\nabla \psi^0_D|^2\,dx.\end{aligned}$$ Using the Green formula and the fact that $$\mbox{ div } (\gamma(x) \nabla (\psi^\eps_D + \psi^0_D)) =2F \hbox{ in } \Omega_{z,\eps} \quad \hbox{ and }\quad \psi^\eps_D - \psi^0_D = 0 \hbox{ on }\Gamma,$$ we derive $$\begin{aligned} T_D(\eps)&=& \int_{\partial\O_{z,\eps}} \gamma(x) \nabla (\psi^\eps_D + \psi^0_D) \cdot{\bf n } (\psi^\eps_D - \psi^0_D)\,ds + 2 \int_{\Omega_{z,\eps}}F (\psi^\eps_D - \psi^0_D)\,dx \\ &&- \int_{\O_{z,\eps}}\gamma(x)|\nabla \psi^0_D|^2\,dx.\end{aligned}$$ Recalling that $\mbox{ div } (\gamma(x) \nabla \psi^0_D) =F \hbox{ in } \O_{z,\eps}$ and taking into account of the normal orientation one can write $$\int_{\O_{z,\eps}}\gamma(x)|\nabla \psi^0_D|^2\,dx= \int_{\O_{z,\eps}} F \, \psi^0_D\,dx - \int_{\partial\O_{z,\eps}} \gamma(x) \nabla \psi^0_D \cdot{\bf n } \, \psi^0_D \,ds.$$ Then, we obtain $$\begin{aligned} \label{term-dirichlet} T_D(\eps)&=& \int_{\partial\O_{z,\eps}} \gamma(x) \nabla \psi^0_D \cdot{\bf n } \psi^\eps_D \,ds + 2 \int_{\Omega_{z,\eps}}F (\psi^\eps_D - \psi^0_D)\,dx - \int_{\O_{z,\eps}} F \, \psi^0_D\,dx.\end{aligned}$$ $-$ [*Calculate of the mixed term.*]{} From the weak formulation of the problems $ (\PB^\eps_N)$ and $ (\PB^0_N)$ one can derive $$\begin{aligned} T_M (\eps)&=& \int_{\Omega_{z,\eps}}\gamma(x)\nabla \psi^\eps_D . \nabla \psi^\eps_N\,dx - \int_{\Omega}\gamma(x)\nabla \psi^0_D . \nabla \psi^0_N\,dx\\ &=&\int_{\Omega_{z,\eps}}F\,\psi^\eps_D\,dx + \int_{\Gamma_a}\Phi\,\psi^\eps_D \hbox{ds} -\int_{\Omega}F\,\psi^0_D\,dx - \int_{\Gamma_a}\Phi\,\psi^0_D \hbox{ds} \end{aligned}$$ Using the fact that $\psi^\eps_D - \psi^0_D = 0 \hbox{ on }\Gamma$, it follows $$\begin{aligned} \label{term-mixed} T_M (\eps)&=& \int_{\Omega_{z,\eps}}F \,(\psi^\eps_D - \psi^0_D)\,dx -\int_{\O_{z,\eps}}F\,\psi^0_D\,dx.\end{aligned}$$ Exploiting the obtained expressions (\[term-neumann\]), (\[term-dirichlet\]) and (\[term-mixed\]), one can easily deduce the desired result of Theorem \[K-variation\]. Asymptotic expansion {#asymptotic} ==================== In this section, we derive a topological asymptotic expansion for the Kohn-Vogelius function $\K$. The mathematical analysis is general and can be adapted for various partial differential equations. To this end, we introduce the polarization matrix $\mathcal M_\omega$. Thanks to the linearity of the integral equations (\[integ-N\]) and (\[integ-D\]), there exists a $d\times d$ matrix $\mathcal M_\omega$ such that $${\displaystyle\int}_{\partial \omega} \eta_N (y) y^T ds(y)=\mathcal M_\omega \nabla \psi^0_N(z) \mbox{ and } {\displaystyle\int}_{\partial \omega} \eta_D (y) y^T ds(y)=\mathcal M_\omega \nabla \psi^0_D(z).$$ The matrix $\mathcal M_\omega$ can be defined as $$(\mathcal M_\omega)_{ij}=\int_{\partial \omega} \eta_i y_j ds(y),\quad 1\leq i,j\leq d,$$ where $y_j$ is the $j$th component of $y\in \mathbb{R}^d$ and $\eta_i $ is the solution to $$-\frac{\eta_i(y)}{2}+\int_{\partial\omega}\nabla E(y-x) \cdot{\bf n } \,\eta_i(x)ds(x)=-e_i \cdot{\bf n } ,\quad y\in \partial \omega$$ with $\{e_i)_{1\leq i\leq d}$ is the canonical basis in $\mathbb{R}^d$. The topological sensitivity analysis with respect to the presence of an arbitrary shaped object is described by the following Theorem. \[th-kv\] Let $\O_{z,\eps}$ be an arbitrary shaped object inserted inside the background domain $\Omega$. The function $\K$ admits the asymptotic expansion $$\K(\Omega\backslash\overline{\O_{z,\eps}})=\K (\Omega)+ \eps^d \delta \K(z) + o(\eps^d),$$ with $\delta \K$ is the topological gradient $$\begin{aligned} \delta \K(x) &=& \gamma(x) \left\{\nabla \psi^0_N(x) \mathcal M_\omega \nabla \psi^0_N(z)- \nabla \psi^0_D(x) \mathcal M_\omega \nabla \psi^0_D(z) \right\}\\ && -2 |\omega|\, F(x) (\psi^0_N(x)-\psi^0_D(x)),\,\forall x\in \Omega.\end{aligned}$$ The polarization matrix $\mathcal M_\omega$ can be determined analytically in some cases. Otherwise, it can be approximated numerically.\ Particularly, in the case of circular or spherical object (i.e. $\omega = B(0,1)$), the matrix $\mathcal M_\omega$ is given by $$\mathcal M_\omega = 2\pi \mathcal I \mbox{ if } d=2 \mbox{ or } d=3,$$ where $\mathcal I $ is the $d\times d$ identity matrix. \[cor-bal\] If $\omega = B(0,1)$, the function $\K$ satisfies the following asymptotic expansion $$\begin{aligned} \K(\Omega_{z,\eps})-\K (\Omega)=2 \pi \,\varepsilon^d \delta \K(z)+o(\varepsilon^d),\end{aligned}$$ and the topological gradient $\delta \K$ admits the expression $$\begin{aligned} \delta \K(x) = \left\{ \begin{array}{ll} \gamma(x)\left(\left|\nabla \psi^0_N(x)\right|^2-\left|\nabla \psi^0_D(x)\right|^2\right)-F(x)\left( \psi^0_N(x)-\psi^0_D(x)\right),&\mbox{ if } d=2,\\ \gamma(x)\left(\left|\nabla \psi^0_N(x)\right|^2-\left|\nabla \psi^0_D(x)\right|^2\right)-\displaystyle \frac{4}{3}F(x)\left( \psi^0_N(x)-\psi^0_D(x)\right),&\mbox{ if } d=3. \end{array} \right.\end{aligned}$$ [*Proof of Theorem \[th-kv\]*]{} : It is established in Theorem \[K-variation\] that the variation of the function $\K$ can be rewritten as $$\begin{aligned} \label{decomp-k} \nonumber \K (\Omega_{z,\eps})-\K (\Omega) &=& \int_{\partial\O_{z,\eps}}\gamma(x) \nabla \psi^0_D\cdot{\bf n } \, \psi^\eps_D\,ds -\int_{\partial\O_{z,\eps}}\gamma(x) \nabla \psi^0_N \cdot{\bf n } \, \psi^\eps_N\,ds \\ && + \int_{\O_{z,\eps}} F\,(\psi^0_D -\psi^0_N )\,dx.\end{aligned}$$ Next, we will derive an asymptotic expansion with respect to $\eps$ for each term separately. $-$ [*Asymptotic expansion for the first integral term*]{}. The first term in (\[decomp-k\]) can be decomposed as $$\begin{aligned} \int_{\partial\O_{z,\eps}}\gamma(x) \nabla \psi^0_D\cdot{\bf n } \, \psi^\eps_D\,ds = \int_{\partial\O_{z,\eps}}\gamma(x) \nabla \psi^0_D\cdot{\bf n } \, (\psi^\eps_D(x)-\psi^0_D(x)-\eps \varphi_D((x-z)/\eps))\,ds\\ + \int_{\partial\O_{z,\eps}}\gamma(x) \nabla \psi^0_D\cdot{\bf n } \, \psi^0_D\,ds+ \eps \int_{\partial\O_{z,\eps}}\gamma(x) \nabla \psi^0_D\cdot{\bf n } \, \varphi_D((x-z)/\eps)\,ds.\end{aligned}$$ Using Theorem \[est-pert-sol\] and the fact that $\gamma(x)\nabla \psi^0_D(x)$ is uniformly bounded in $\overline{\O}_{z,\eps}$, one can obtain $$\begin{aligned} \label{estim-1} \int_{\partial\O_{z,\eps}}\gamma(x) \nabla \psi^0_D\cdot{\bf n } \, [\psi^\eps_D(x)-\psi^0_D(x)-\eps \varphi_D((x-z)/\eps)]\,ds = o(\eps^d).\end{aligned}$$ Due to the Green’s formula and taking into account of the normal orientation, $$\begin{aligned} \label{estim-2} \int_{\partial\O_{z,\eps}}\gamma(x) \nabla \psi^0_D\cdot{\bf n } (x) \, \psi^0_D(x)ds=-\int_{\O_{z,\eps}}\gamma(x) \nabla \psi^0_D(x) .\nabla \psi^0_D(x) dx + \int_{\O_{z,\eps}} F(x)\psi^0_D(x) dx.\end{aligned}$$ By the change of variable $x=z+\eps y$, one can write $$\begin{aligned} \eps \int_{\partial\O_{z,\eps}}\gamma(x) \nabla \psi^0_D\cdot{\bf n } \, \varphi_D((x-z)/\eps)ds= \eps^d \int_{\partial\omega}\gamma(z) \nabla \psi^0_D(z)\cdot{\bf n } \, \varphi_D(y)ds(y)\\ + \eps^d \int_{\partial\omega}[\gamma(z+\eps y)\nabla \psi^0_D(z+\eps y)-\gamma(z) \nabla \psi^0_D(z)]\cdot{\bf n } \, \varphi_D(y)ds(y).\end{aligned}$$ Using again the smoothness of the function $x\mapsto \gamma(x)\nabla \psi^0_D(x)$ near $z$, one can obtain $$\begin{aligned} \eps \int_{\partial\O_{z,\eps}}\gamma(x) \nabla \psi^0_D\cdot{\bf n } \, \varphi_D((x-z)/\eps)ds= \eps^d \gamma(z) \int_{\partial\omega} \nabla \psi^0_D(z)\cdot{\bf n } \, \varphi_D(y)ds(y)+o(\eps^d).\end{aligned}$$ Let $\xi^{\varphi_D}$ the unique solution to $$\left\{ \begin{array}{rll} \displaystyle - \Delta \xi^{\varphi_D} =& 0 &\hbox{ in } \omega \\ \xi^{\varphi_D} =& \varphi_D & \hbox{ on } \partial \omega . \end{array} \right.$$ Exploiting the Green’s formula, the fact that $\Delta (\nabla \psi^0_D(z)y)=0$ in $\omega$ and taking into account of the normal orientation, one can derive $$\begin{aligned} \int_{\partial\omega} \nabla \psi^0_D(z)\cdot{\bf n } \, \varphi_D(y)ds(y)&=& - \int_{\omega} \nabla \psi^0_D(z). \nabla \xi^{\varphi_D} (y)dy\\ &=& \int_{\partial\omega} \nabla \xi^{\varphi_D} (y)\cdot{\bf n } [\nabla \psi^0_D(z)y]ds(y).\end{aligned}$$ Recalling that the density $\eta_D$ is defined as the jump of the flux through the boundary $\partial \omega$ $$\begin{aligned} \eta_D(y)= -\nabla \psi^0_D(z)\cdot{\bf n } - \nabla \xi^{\varphi_D} (y)\cdot{\bf n } ,\quad y\in \partial \omega.\end{aligned}$$ It follows $$\begin{aligned} \int_{\partial\omega} \nabla \psi^0_D(z)\cdot{\bf n } \, \varphi_D(y)ds(y)=- \int_{\omega} [\nabla \psi^0_D(z)y]\eta_D(y)dy - \int_{\partial\omega}\nabla \psi^0_D(z)\cdot{\bf n } [\nabla \psi^0_D(z)y]ds(y).\end{aligned}$$ Then, we deduce $$\begin{aligned} \label{estim-3} \nonumber \eps \int_{\partial\O_{z,\eps}}\gamma(x) \nabla \psi^0_D\cdot{\bf n } \, \varphi_D((x-z)/\eps)ds= -\eps^d \gamma(z) \nabla \psi^0_D(x) .{\displaystyle\int}_{\partial \omega} \eta_D (y) y ds(y)\\ +\int_{\O_{z,\eps}}\gamma(z)\nabla \psi^0_D(z).\nabla \psi^0_D(z)dx+o(\eps^d).\end{aligned}$$ Combining (\[estim-1\]), (\[estim-2\]) and (\[estim-3\]), we have $$\begin{aligned} \int_{\partial\O_{z,\eps}}\gamma(x) \nabla \psi^0_D\cdot{\bf n } \, \psi^\eps_D\,ds = -\eps^d \gamma(z) \nabla \psi^0_D(x) .{\displaystyle\int}_{\partial \omega} \eta_D (y) y ds(y) + \int_{\O_{z,\eps}} F(x)\psi^0_D(x) dx\\ +\int_{\O_{z,\eps}}\gamma(z)\nabla \psi^0_D(z).\nabla \psi^0_D(z)dy-\int_{\O_{z,\eps}}\gamma(x) \nabla \psi^0_D(x) .\nabla \psi^0_D(x) dx +o(\eps^d).\end{aligned}$$ Finally, using a change of variable and the Taylor’s formula one can easily deduce that the first term in (\[decomp-k\]) admits the following expansion $$\begin{aligned} \label{est-term1} \int_{\partial\O_{z,\eps}}\gamma(x) \nabla \psi^0_D\cdot{\bf n } \, \psi^\eps_D\,ds =-\eps^d \gamma(z) \nabla \psi^0_D(x) .{\displaystyle\int}_{\partial \omega} \eta_D (y) y ds(y) + \eps^d |\omega|\, F(z)\psi^0_D(z)+o(\eps^d).\end{aligned}$$ $-$ [*Asymptotic expansion for the second integral term*]{}. The second term in (\[decomp-k\]) can be decomposed as $$\begin{aligned} \int_{\partial\O_{z,\eps}}\gamma(x) \nabla \psi^0_N\cdot{\bf n } \, \psi^\eps_N\,ds = \int_{\partial\O_{z,\eps}}\gamma(x) \nabla \psi^0_N\cdot{\bf n } (x) (\psi^\eps_N(x)-\psi^0_N(x)-\eps \varphi_N((x-z)/\eps))\,ds\\ + \int_{\partial\O_{z,\eps}}\gamma(x) \nabla \psi^0_N(x)\cdot{\bf n } \, \psi^0_N(x)\,ds+ \eps \int_{\partial\O_{z,\eps}}\gamma(x) \nabla \psi^0_N\cdot{\bf n } \, \varphi_N((x-z)/\eps)\,ds.\end{aligned}$$ The established estimates for the first term can be easily adapted for this term using $\psi^0_N$ instead of $\psi^0_D$ and $\varphi_N$ instead of $\varphi_D$. Then, one can prove that the second integral term in (\[decomp-k\]) satisfies the following expansion $$\begin{aligned} \label{est-term2} \int_{\partial\O_{z,\eps}}\gamma(x) \nabla \psi^0_N\cdot{\bf n } \, \psi^\eps_N \,ds =-\eps^d \gamma(z) \nabla \psi^0_N(x) .{\displaystyle\int}_{\partial \omega} \eta_N (y) y ds(y) + \eps^d |\omega|\, F(z)\psi^0_N(z)+o(\eps^d).\end{aligned}$$ $-$ [*Asymptotic expansion for the third integral term*]{}. The estimate of the third term is based on the change of variable $y=z+ \eps y$ and the smoothness of $\gamma$ in $\O_{z,\eps}$. Using Taylor’s Theorem, one can derive $$\begin{aligned} \label{est-term3} \int_{\O_{z,\eps}} F\,(\psi^0_D -\psi^0_N )dx = \eps^d |\omega| F(z) (\psi^0_D(z) -\psi^0_N (z)) + o(\eps^d).\end{aligned}$$ The desired asymptotic expansion of the function $\K$ follows from the estimates (\[est-term1\]), (\[est-term2\]) and (\[est-term3\]). Algorithm and numerical results =============================== In this section we consider the bidimentional case and we present a fast and simple one-iteration identification algorithm. Our numerical procedure is based on the formula described by Corollary \[cor-bal\]. The unknow object $\O$ is identified using a level set curve of the topological gradient $\delta \K$. More precisely, the unknow object $\O$ is likely to be located at zone where the topological gradient $\delta \K$ is negative. 0.4cm [**One-iteration algorithm:**]{} - Solve the two problems ($\PB_N^0$) and ($\PB_D^0$), - Compute the topological gradient $\delta \K(x),\, x\in \Omega$, - Determine the unknow object $$\O =\left\{x\in \Omega;\hbox{ such that } \delta \K(x)<c<0\right\},$$ where $c$ is a constant chosen in such a way that the cost function $\K$ decreases as most as possible. 0.2cm This one-iteration procedure has already been illustrated in [@AHM2] for the identification of cracks from overdetermined boundary data and in [@BHJM] for the detection of small gas bubbles in Stokes flow. Next, we presente some numerical results showing the efficiency and accuracy of our proposed one-iteration algorithm. In Figure 1, we test our algorithm on circular shape. In Figure 2, we consider the case of an elliptical shape. As one can observe, the domain to be detected is located at zone where the topological gradient is negative and it is approximated by a level set curve of the topological gradient $\delta \K$. The result is quite efficient. In Figure 3, we obtain an intersting reconstruction result for a non trivial shape. {width="4.5cm" height="3.1cm"} {width="4.5cm" height="3.1cm"} {width="4.5cm" height="3.1cm"} -0.6cm -0.6cm -0.6cm In Figure 4 we illustrate the case of a complex geometry. As one can see, we detect efficiency the location of the unknown domain but not its shape. The obtained result can serve as a good initial guess for an iterative optimization process based on the shape derivative. -0.6cm 0.3cm Conclusion ========== The presented work concerns the detection of objects immersed in anisotropic media from boundary measurements. The present approach is based on the Kohn-Vogelius formulation and the topological gradient method. A topological sensitivity analysis is derived for an energy like functional. An accurate and fast reconstruction algorithm is proposed. The efficiency and accuracy of the suggested algorithm are illustrated by some numerical results. The considered model can be viewed as a prototype of a geometric inverse problem arising in many applications. The presented approach is general and can be adapted for various partial differential equations. [99]{} M. Abdelwahed, M.Hassine, M. Masmoudi [*Optimal shape design for fluid flow using topological perturbation technique*]{}, Journal of Mathematical Analysis and Applications, 356, 2009, 548-563. Ammari H., Kang H., [*Reconstruction of small inhomogeneities from boundary measurements*]{}, Lecture Notes in Mathematics, 1846, Springer, 2004 Amstutz S., Horchani I., Masmoudi M. [*Crack detection by the topological gradient method*]{}, Control and Cybernetics, 34 (1), 2005, 81-101. Andrieux S., Baranger T. N., Ben Abda A. (2006), [*Solving Cauchy problems by minimizing an energy-like functional*]{}, Inverse Problems, [**22**]{}, 115-134. M. Badra F.Caubet and M. Dambrine, [*Detecting an obstacle immersed in a fluid by shape optimization methods*]{}, M3AS, 21 (10), 2069-2101, 2011. Ben Abda A., Hassine M., Jaoua M., Masmoudi M., [*Topological sensitivity analysis for the location of small cavities in Stokes flow*]{}, SIAM J. Contr. Optim., 48 (5), 2871–2900, 2009. Bowler J., [*Thin-skin eddy-current inversion for the determination of cracks shapes*]{}, Inverse Problems, 18, 1891-1905, 2002. M.Cheney, D.Isaacson J.C.Newell, [*Electrical impedance tomography*]{}, SIAM Review.,40, 85-101, 1999. R. Dautray and J.-L. Lions, [*Analyse mathématique et calcul numérique pour les sciences et les techniques*]{}, Masson, collection CEA 6 (1987). S. Garreau, Ph. Guillaume, M. Masmoudi, [*The topological asymptotic for PDE systems: The elastics case*]{}, SIAM J. contr. Optim. 39(6), 1756-1778, (2001). Ph.Guillaume, M.Hassine [*Removing holes in topological shape optimization*]{}, ESAIM, Control, Optimisation and Calculus of Variations, 14 (1), 2008, 160-191. Guzina B.B., Bonnet M. [*Small-inclusion asymptotic of misfit functionals for inverse problems in acoustics*]{}, Inverse Problems, 22, 2006, 1761-1785. M. Hassine, S. Jan, M. Masmoudi, [*From differential calculus to $0-1$ topological optimization*]{}, SIAM J. Cont. Optim., 45 (6), 1965-1987, 2007. Parker R.L., [*The inverse problem of resistivity sounding*]{}, Geophysics, 42, 2143-2158, 1984.

ArXiv

--- abstract: 'We report our study of the real-time charge counting statistics measured by a quantum point contact (QPC) coupled to a single quantum dot (QD) under different back-action strength. By tuning the QD-QPC coupling or QPC bias, we controlled the QPC back-action which drives the QD electrons out of thermal equilibrium. The random telegraph signal (RTS) statistics showed strong and tunable non-thermal-equilibrium saturation effect, which can be quantitatively characterized as a back-action induced tunneling out rate. We found that the QD-QPC coupling and QPC bias voltage played different roles on the back-action strength and cut-off energy.' author: - HaiOu Li - Ming Xiao - Gang Cao - Cheng Zhou - RuNan Shang - Tao Tu - GuangCan Guo - HongWen Jiang - GuoPing Guo title: 'Back-action Induced Non-equilibrium Effect in Electron Charge Counting Statistics' --- A quantum point contact (QPC) nearby a quantum dot (QD) is widely used to perform electron charge counting, which is important in many aspects such as the read-out of electron charge- or spin-based qubits [@Petta-Charge-Qubit-DQD; @Delft-Single-Spin-Readout]. However, the measurement with a QPC has inevitable side effects, known as the back-action [@Kouwenhoven-Backaction-Theory]. For example, the back-action can drive the QD electrons out of thermal equilibrium, facilitating ground to excited-state transitions [@Ensslin-Backaction-eVqpc; @Ludwig-Backaction-QPC-Current], and possibly causing qubit relaxation and dephasing [@Weizman-Backaction-WhichPath]. Thus, cautions have to be exercised to minimize the back-action, in order to use the QPC as a non-invasive read-out sensor. However, back-action is usually indirect and hard to quantitatively analyze. In QDs the excited tunneling through higher-energy levels has been utilized to study back-action[@Kouwenhoven-Backaction-Theory; @Ensslin-Backaction-eVqpc; @Ludwig-Backaction-QPC-Current; @Gustavsson-Backaction-Terahertz]. Here we found that we can relate the non-thermal-equilibrium effect in the charge counting statistics to back-action. This effect arises from the back-action induced tunneling directly out of the QD, without the help of excited levels. We developed a phenomenological model to quantitatively characterize the back-action strength. We also found the different contributions of the QD-QPC Coulomb coupling and QPC bias voltage. Especially we revealed that the back-action cut-off energy is proportional to the QD-QPC coupling, whereas independent of the QPC bias. We fabricated a QD with a QPC on side in a GaAs/AlGaAs heterostructure. The two-dimensional electron gas (2DEG) is 95 nm below the surface. The 2DEG has a density of $3.2\times10^{11} cm^{−2}$ and a mobility of $1.5\times10^{5} cm^{2}V^{-1}s^{-1}$. Fig. \[Figure1\] (a) shows the scanning electron microscopy (SEM) picture of the surface gates. Five gates LT, RT, LB, RB and P shaped the QD. Gate Q, along with LT and RT, formed a QPC channel to count the QD electron number via capacitive coupling. A small gap between LT and RT was created to maximize this coupling. The experiment was done in a Helium-3 refrigerator with base temperature of $240 mK$. We operate the QD in such a way that the left barrier is closed and the electrons only tunnel through the right barrier (tunneling rate conveniently controlled by gate RB). The voltages on LT and RT are set below pinch-off, so that no leakage tunneling to the QPC was found. Both the source and drain of the QD are grounded. A small dc bias is applied through the QPC channel. Fig. \[Figure1\] (b) shows the QPC response while gate P is used to control the number of QD electrons. The insert is a trace of random telegraph signal (RTS) for the $0e \leftrightarrow 1e$ transition. When studying the RTS statistics carefully, in some conditions we found large deviation from a thermal equilibrium picture. We studied all the last six electrons and found similar phenomena, except the complication of additional tunneling channels through spin excited levels for and only for the even electron numbers [@Ming-Backaction-SpinST]. Here we focus on the $0e \leftrightarrow 1e$ transition which contains all the basic features without this complication . Fig. \[Figure2\] (a) - (c) showed the RTS statistics with different QD-QPC gap opening (controlled by voltage $V_{T} \equiv V_{LT}=V_{RT}$), including the $0e$ and $1e$ occupancy ratio $R_{0/1}$, total tunneling rate $\Gamma^{total}$, tunneling out rate $\Gamma^{out}$, and tunneling in rate $\Gamma^{in}$. In comparison we also showed the simulation of the RTS statistics for a thermally activated two-level switching in dotted lines [@Uren-RTS-Review]: $\left\{ \begin{array}{llll} \Gamma^{out}=g_{n-1} \Gamma^{*} (1-f(\mu_{n})) \\ \Gamma^{in}= g_{n} \Gamma^{*} f(\mu_{n}) \\ \Gamma^{total}=1/(1/\Gamma^{out}+1/ \Gamma^{in}) \\ R_{n-1/n}=\Gamma^{out}/\Gamma^{in}=g_{n-1}/g_{n}e^{(\mu_{n}-E_{F}) / k_{B}T} \\ \end{array} \right. $ Here $g_{n}$ is the spin degeneracy; $\mu_{n}$ denotes the addition energy for the $n^{th}$ electron; Maximum tunneling rate $\Gamma^{*} \equiv (2\pi / \hbar) D \Delta^{2}$ where $D$ the electron density of energy and $\Delta$ is the tunneling matrix element; $f(\mu_{n})$ is the Fermi distribution function. We used the energy-voltage conversion factor $0.086 meV/mV$ obtained in transport experiment and temperature 240mK read by a sensor. Our data can be well described by the thermal equilibrium equations when $\mu_{n}-E_{F}>0$. However, a deviation appears when $\mu_{n}-E_{F}<0$, where the first electron is mostly trapped in the QD. $R_{0/1}$ should exponentially decay when $\mu_{1}$ drops below $E_{F}$ since the electron loses energy to tunnel out of the QD. Not as expected, we observed that $R_{0/1}$ saturates in this region. The same phenomena happens for the tunneling rates $\Gamma^{total}$ and $\Gamma^{out}$, but not for $\Gamma^{in}$. This saturation effect is found to increase with either the QD-QPC coupling or the QPC dc bias voltage. For example, in Fig. \[Figure2\] (a) we see that when $V_{T}=-1.40V$, $R_{0/1}$ starts to saturate when its value drops to $10^{-2}$, two magnitude larger than the saturation point $10^{-4}$ under the $V_{T}=-1.50V$ condition. We also observed stronger saturation effect by increasing the QPC dc bias voltage $V_{dc}$, as shown in Fig. \[Figure2\] (d). Although not open enough to allow direct tunneling, the QD-QPC gap increases their coupling strength drastically. The RTS amplitude reaches as high as $20 \%$. Such huge coupling must mean strong back-action as well. Principally the back-action strength is determined by the QPC condition and QD-QPC interaction together. The former can be controlled by the QPC dc bias voltage $V_{dc}$ [@Ensslin-Backaction-eVqpc; @Ludwig-Backaction-QPC-Current], and the latter can be controlled by voltage $V_{T}$. So by increasing either $V_{dc}$ or $V_{T}$ we introduce stronger back-action. The effect of back-action is illustrated in Fig. \[Figure2\] (e) and (f). When $\mu_{1}$ is below $E_{F}$, the QD should have been mostly filled with one electron. However, after receiving phonons emitted by the QPC, the electron gains energy and tunnels out. The outcome is that the QD empty occupancy increases and so does the ratio $R_{0/1}$. This phonon-assisted tunneling remains there when $E_{F} - \mu_{1}$ is less than a certain cut-off energy [@Ludwig-Backaction-DQD]. On the other hand, when $\mu_{1}$ is way above $E_{F}$, the phonon doesn’t apparently assist the electrons to tunnel inside since the 2DEG is supposed to be a huge reservoir in thermal equilibrium. This explains why $\Gamma^{in}$ doesn’t show saturation effect. We developed a phenomenological model about the effect of back-action: $\left\{ \begin{array}{llll} \Gamma^{out}=g_{n-1} [\Lambda^{out} + \Gamma^{*} (1-f(\mu_{n}))] \\ \Gamma^{in}= g_{n} [\Lambda^{in} + \Gamma^{*} f(\mu_{n})] \\ \Gamma^{total}=1/(1/\Gamma^{out}+1/ \Gamma^{in}) \\ R_{n-1/n}=\Gamma^{out}/\Gamma^{in} \\ \end{array} \right. $ Here we introduced two extra tunneling rates $\Lambda^{out}$ and $\Lambda^{in}$. They refer to the back-action driven tunneling out and in rates. Their values can be easily estimated from $\Gamma^{out}$ and $\Gamma^{in}$ in the extreme conditions: $\Gamma^{out} \approx g_{n-1}\Lambda^{out}$, $\Gamma^{in} \approx g_{n} \Gamma^{*}$ when $\mu_{n}<<E_{F}$; and $\Gamma^{out} \approx g_{n-1} \Gamma^{*}$, $\Gamma^{in} \approx g_{n}\Lambda^{in}$ when $\mu_{n}>>E_{F}$. We can immediately tell two basic features: $\Lambda^{out}$ is invariant and $\Lambda^{in}$ is negligible. $\Lambda^{out}$ is constant since the saturation tails are flat, implying a constant phonon spectrum before a certain cut-off energy. $\Lambda^{in}$ is found to be at least two-magnitude less than $\Lambda^{out}$ and two more magnitudes smaller than the other tunneling rates so we can safely ignore it. We showed our simulation as solid lines in Fig. \[Figure2\] (a) -(d). Good agreement with the experiment was found. We summarized all the simulation results in table I. As $V_{T}$ or $V_{dc}$ is increased, the back-action induced tunneling out rate $\Lambda^{out}$ shows observable increase. We use the percentage of $\Lambda^{out}$ on $\Gamma^{*}$ as a more objective measure since $\Gamma^{*}$ or $\Lambda^{out}$ alone could be affected in different experimental conditions. Table I shows that $\Lambda^{out}/ \Gamma^{*}$ steadily increases from $0.026\%$ to $1\%$ as we increase $V_{T}$ alone, and from $0.026\%$ to $0.8\%$ as we apply larger bias $V_{dc}$. Hence, we think that $\Lambda^{out}/ \Gamma^{*}$ serves as a sensitive quantitative measure of the back-action strength. $V_{T}$ $V_{dc}$ $\Lambda^{out}$ $\Gamma^{*}$ $\Lambda^{out}/\Gamma^{*}$ $T$ $E_{cut}$ ---------- ---------- ----------------- -------------- ---------------------------- --------- ----------- $-1.40V$ $0.5mV$ $15.0Hz$ $1.5kHz$ $1.0\%$ $0.32K$ $3.75meV$ $-1.46V$ $0.5mV$ $5.0Hz$ $1.6kHz$ $0.31\%$ $0.32K$ $3.40meV$ $-1.50V$ $0.5mV$ $0.5Hz$ $2.5kHz$ $0.026\%$ $0.24K$ $2.82meV$ $-1.50V$ $1.5mV$ $4.0Hz$ $2.4kHz$ $0.16\%$ $0.38K$ $2.85meV$ $-1.50V$ $2.5mV$ $20.0Hz$ $2.5kHz$ $0.8\%$ $0.45K$ $2.90meV$ : Simulation results at various back-action condition[]{data-label="Table1"} Also shown in Table I is that the electron temperature keeps warming up with $V_{dc}$, while only slightly increases with $V_{T}$. When $V_{T}$ goes from $-1.46V$ to $-1.40V$, the electron temperature nearly has no change although $\Lambda^{out}/ \Gamma^{*}$ increases by three times. This difference may indicate different back-action mechanisms. As we know, the inelastic back-action could be either direct or indirect. The QPC current heats up the electron bath and emits high frequency quasi-particles such as acoustic phonons [@Ludwig-Backaction-QPC-Current]. The partial re-absorption of these particles by the QD electrons causes indirect back-action. At large $V_{dc}$, the QPC current heats up the electron bath substantially and thus the indirect back-action could dominate. On the contrary, increasing the gap opening doesn’t cause severe heating. Instead it increases the direct QD-QPC Coulomb coupling strength efficiently. In this case the direct back-action should prevail. In general, both the indirect and direct back-action should exist. Another difference is in the cut-off energy $E_{cut}$, beyond where the saturation effect suddenly relieves. Fig. \[Figure3\] (a) - (b) show $R_{0/1}$ in a large energy scale for varying $V_{dc}$ or $V_{T}$. We determined $E_{cut}$ as where $R_{0/1}$ drops to $10^{-6}$ since we only take $10^{6}$ data points. The results are listed in the last column of Table I. $E_{cut}$ shows very small change (between $2.82meV$ and $2.90meV$) with $V_{dc}$, and increases dramatically (from $2.82meV$ to $3.75meV$) for increasing $V_{T}$. In both cases, the observed $E_{cut}$ is much larger than $|eV_{dc}|$. Since $|eV_{dc}|$ is the maximum single phonon energy the QPC emits [@Kouwenhoven-Backaction-Theory; @Ensslin-Backaction-eVqpc], in our system there must involve multi-phonon absorption process [@Ludwig-Backaction-DQD]. And the independence on $|V_{dc}|$ indicates that $E_{cut}$ is not limited by the number of phonons to absorb. Instead, from its increase with $V_{T}$, we conclude that $E_{cut}$ has to do with the QD-QPC coupling strength. We found that $V_{T}$ effectively and exclusively controls the QD-QPC coupling. Fig. \[Figure3\] (c) shows that with less negative $V_{T}$ the RTS amplitude, an indicator of the coupling strength, increases quickly. On the contrary, the other gate voltages such as $V_{LB}$ changes neither the coupling strength nor the cut-off energy, as shown in Fig. \[Figure3\] (d). Thus we suspect that $E_{cut}$ depends on $V_{T}$ through the QD-QPC coupling strength. However, at this moment we can not analytically determine this dependence. It may need further theoretical models, especially which incorporate multi-phonon process, to explicitly determine the cut-off energy. In conclusion, we found that the non-equilibrium effect in QPC charge counting statistics is a benchmark of its back-action. We gave a quantitative measure of the back-action strength. This paved a way for further study, such as the back-action driven spin excitations [@Ming-Backaction-SpinST]. This work was supported by the NFRP 2011CBA00200 and 2011CB921200, NNSF 10934006, 11074243, 10874163, 10804104, 60921091. [99]{} K. D. Petersson, et. al., Phys. Rev. Lett. 105, 246804 (2010). J. M. Elzerman, et. al., Nature 430, 432 (2004). R. Aguado and L. P. Kouwenhoven, Phys. Rev. Lett. 84, 1986 (2000). S. Gustavsson, et. al., Phys. Rev. Lett. 99, 206804 (2007). V. S. Khrapai, et. al., Phys. Rev. Lett. 97, 176803 (2006). E. Buks, et. al., Nature 391, 871 (1998). S. Gustavsson, et. al., Phys. Rev. B 78, 035324 (2008). G. Cao, et. al., arXiv:1109.5734v1 (2011). M. J. Kirton and M. J. Uren, Adv. Phys. 38, 367 (1989). D. Harbusch, et. al., Phys. Rev. Lett. 104, 196801 (2010).

ArXiv

--- abstract: 'We investigate type I multiple orthogonal polynomials on $r$ intervals which have a common point at the origin and endpoints at the $r$ roots of unity $\omega^j$, $j=0,1,\ldots,r-1$, with $\omega = \exp(2\pi i/r)$. We use the weight function $|x|^\beta (1-x^r)^\alpha$, with $\alpha,\beta >-1$ for the multiple orthogonality relations. We give explicit formulas for the type I multiple orthogonal polynomials, the coefficients in the recurrence relation, the differential equation, and we obtain the asymptotic distribution of the zeros.' author: - title: 'Jacobi-Angelesco multiple orthogonal polynomials on an $r$-star[^1]' --- Multiple orthogonal polynomials; Jacobi-Angelesco polynomials; recurrence relation; differential equation; asymptotic zero distribution. 33C45; 42C05. Introduction ============ Various families of multiple orthogonal polynomials have been worked out during the past few decennia, even though the notion of multiple orthogonality goes back at least to Hermite in the framework of Hermite-Padé approximation. There are two types of multiple orthogonal polynomials. Let $\vec{n}=(n_1,n_2,\ldots,n_r)$ be a multi-index of size $|\vec{n}| = n_1+n_2+\cdots+n_r$ and let $\mu_1,\ldots,\mu_r$ be positive measures for which all the moments exist. Type I multiple orthogonal polynomials for $(\mu_1,\ldots,\mu_r)$ are given by a vector $(A_{\vec{n},1}, \ldots, A_{\vec{n},r})$ of $r$ polynomials, with $\deg A_{\vec{n},j} = n_j-1$, such that the following orthogonality conditions hold: $$\sum_{j=1}^r \int x^k A_{\vec{n},j}(x) \, d\mu_j(x) = 0, \qquad 0 \leq k \leq |\vec{n}|-2,$$ with normalization $$\sum_{j=1}^r \int x^{|\vec{n}|-1} A_{\vec{n},j}(x) \, d\mu_j(x) = 1.$$ The type II multiple orthogonal polynomial for the multi-index $(n_1,\ldots,n_r)$ is the *monic* polynomial $P_{\vec{n}}$ of degree $|\vec{n}|$ for which the following orthogonality conditions hold: $$\int x^k P_{\vec{n}}(x)\, d\mu_j(x) = 0, \qquad 0 \leq k \leq n_j-1,$$ for $1 \leq j \leq r$. The orthogonality conditions for type I and type II multiple orthogonal polynomials give a linear system of $|\vec{n}|$ equations for the $|\vec{n}|$ unknown coefficients of the polynomials. If the solution exists and if it is unique, then we call the multi-index $\vec{n}$ a normal index, and if all multi-indices are normal, then the measures $(\mu_1,\ldots,\mu_r)$ are a perfect system. See [@Aptekarev] [@Ismail Ch. 23] [@NikiSor Ch. 4.3] for more information on multiple orthogonal polynomials (polyorthogonal polynomials). An important perfect system of measures was introduced by Angelesco[^2] in 1919 [@Angelesco] and later independently suggested by Nikishin [@Nikishin]. An Angelesco system has $r$ measures $\mu_1,\ldots,\mu_r$ where $\mu_j$ has support in an interval $\Delta_j$ and the intervals $\Delta_1,\ldots,\Delta_r$ are pairwise disjoint. Actually the intervals may be touching. Kalyagin [@Kalyagin] gave an explicit example of an Angelesco system which is basically a generalization of Jacobi polynomials. He considered the two intervals $[-1,0]$ and $[0,1]$ and investigated the type II multiple orthogonal polynomials $P_{n,m}$ satisfying $$\int_{-1}^0 P_{n,m}(x) x^k (1-x)^\alpha(1+x)^\beta |x|^\gamma\, dx = 0, \qquad 0 \leq k \leq n-1,$$ $$\int_{0}^1 P_{n,m}(x) x^k (1-x)^\alpha(1+x)^\beta |x|^\gamma\, dx = 0, \qquad 0 \leq k \leq m-1,$$ and investigated the asymptotic behavior of $P_{n,m}$ and later, with Ronveaux [@KalRon], found a third order differential equation, a four term recurrence relation and the asymptotic behavior of the ratio of two neighboring polynomials. We call these multiple orthogonal polynomials Jacobi-Angelesco polynomials. Type I Jacobi-Angelesco polynomials were only recently investigated for the case $\alpha=\beta=\gamma=0$ because they turn up in the analysis of Alpert multiwavelets [@GIWVA]. In this paper we will extend these type I Legendre-Angelesco and Jacobi-Angelesco polynomials to $r$ intervals. We take a special configuration for the $r$ intervals by having one common point $0$ and placing them on an $r$-star in the complex plane, with endpoints at the $r$ roots of unity $\omega^j$, $j=0,1,\ldots,r-1$, with $\omega = e^{2\pi i/r}$, see Figure \[fig:rstar\]. =0.6mm (100,100)(0,0) (50,50) (100,50) (65.45,97.55) (9.55,79.39) (9.55,20.61) (65.45,2.45) (54,53)[$0$]{} (104,53)[$1$]{} (70,100)[$\omega$]{} (5,84)[$\omega^2$]{} (5,12)[$\omega^3$]{} (70,0)[$\omega^4$]{} (50,50)[(0,0)(25,0)(50,0)]{} (50,50)[(0,0)(7.5,23.775)(15,47.55)]{} (50,50)[(0,0)(7.6,-23.775)(15,-47.55)]{} (50,50)[(0,0)(-20.225,14.695)(-40.45,29.39)]{} (50,50)[(0,0)(-20.225,-14.695)(-40.45,-29.39)]{} To preserve the symmetry, we take a weight function $w(x) = |x|^\beta (1-x^r)^\alpha$ and the measure $\mu_j$ is supported on the interval $\Delta_j = [0,\omega^{j-1}]$, $j=1,\ldots,r$ with this weight function as its Radon-Nikodym derivative. The orthogonality properties for the type I multiple orthogonal polynomials are then given by $$\sum_{j=1}^r \int_0^{\omega^{j-1}} x^k A_{\vec{n},j}(x) |x|^\beta (1-x^r)^\alpha\, dx = 0, \qquad 0 \leq k \leq |\vec{n}|-2,$$ and normalization $$\sum_{j=1}^r \int_0^{\omega^{j-1}} x^{|\vec{n}|-1} A_{\vec{n},j}(x) |x|^\beta (1-x^r)^\alpha\, dx = 1.$$ Observe that we are not using complex conjugation, hence the corresponding bilinear form is not an inner product. Nevertheless the type I multiple orthogonal polynomials will exist and they are unique, at least for multi-indices on the diagonal $\vec{n}=(n,n,\ldots,n)$ or near the diagonal $\vec{n} \pm \vec{e}_k$, where $\vec{e}_k$ is the $k$th unit vector in $\mathbb{Z}^r$. We will investigate these type I multiple orthogonal polynomials in Section \[secI\] where we give an explicit formula for the polynomials, prove their multiple orthogonality, give the recurrence coefficients in the nearest neighbor recurrence relations and obtain a differential equation, which we use to get the asymptotic distribution of the zeros. We give the results and the proofs for $r=2$ in full detail. In Section \[secII\] we consider the general case $r >1$ and again give an explicit expression for the type I multiple orthogonal polynomials, prove their multiple orthogonality, give the recurrence coefficients of the nearest neighbor recurrence relation near the diagonal, give a differential equation of order $r+1$, and work out the asymptotic distribution of the zeros. The results and the proofs are more complicated and technical, and we only outline the necessary modifications of the proofs for the case $r=2$ to general $r$. The type II Jacobi-Angelesco polynomials are somewhat easier to analyze, because for $\vec{n} = (n,n,\ldots,n)$ they are given by a Rodrigues type formula $$x^\beta (1-x^r)^\alpha P_{\vec{n}}(x) = C_n(\alpha,\beta) \frac{d^n}{dx^n} x^{\beta+n} (1-x^r)^{\alpha+n} ,$$ where $C_n(\alpha,\beta)$ is a constant that makes $P_{\vec{n}}$ a monic polynomial. These polynomials will not be considered in the present paper. Type I Jacobi-Angelesco polynomials for $r=2$ {#secI} ============================================= Type I Legendre-Angelesco polynomials appeared in [@GIWVA], where they were used to expand Alpert multiwavelets. In this case one has $r=2$ and the polynomials $(A_{n,m},B_{n,m})$ are such that $\deg A_{n,m} = n-1$, $\deg B_{n,m} = m-1$, and the orthogonality conditions are $$\int_{-1}^1 \bigl( A_{n,m}(x) \chi_{[-1,0]}(x) + B_{n,m}(x) \chi_{[0,1]}(x) \bigr) x^k \, dx = 0, \qquad 0 \leq k \leq n+m-2,$$ with the normalization given by $$\int_{-1}^1 \bigl( A_{n,m}(x) \chi_{[-1,0]}(x) + B_{n,m}(x) \chi_{[0,1]}(x) \bigr) x^{n+m-1} \, dx = 1.$$ An explicit expression for these polynomials was given in terms of two families of polynomials $p_n$ and $q_n$ given by $$\label{pn} p_n(x) = \sum_{k=0}^n \binom{n}{k} \binom{n+\frac{k}{2}}{n} (-1)^{n-k} x^k,$$ and $$\label{qn} q_n(x) = \sum_{k=0}^n \binom{n}{k} \binom{n+\frac{k-1}{2}}{n} (-1)^{n-k} x^k.$$ One has (see [@GIWVA Prop. 5 in §4.1]) The type I Legendre-Angelesco polynomials for multi-indices on the diagonal are given by $$B_{n+1,n+1}(x) = \frac12 \frac{(3n+2)!}{n!(2n+1)!} p_n(x), \quad A_{n+1,n+1}(x) = -B_{n+1,n+1}(-x),$$ and for $|n-m|=1$ one has $$\gamma_n B_{n+1,n}(x) = \binom{n+\frac{n}{2}}{n} q_n(x) - \binom{n+\frac{n-1}{2}}{n} p_n(x),$$ $$\gamma_n B_{n,n+1}(x) = \binom{n+\frac{n}{2}}{n} q_n(x) + \binom{n+\frac{n-1}{2}}{n} p_n(x),$$ and $$A_{n+1,n}(x) = B_{n,n+1}(-x), \quad A_{n,n+1}(x) = B_{n+1,n}(-x),$$ where the normalizing constant is given by $\gamma_n = 2(\frac{n}{2}+1)_n (2n)!/(3n+1)!$. We will extend this result by taking a more general Jacobi-type weight function and by using integration on an $r$-star with $r >2$. Explicit expression ------------------- Let us first consider the case $r=2$ and the weight function $w(x) = |x|^\beta (1-x^2)^\alpha$. The type I Jacobi-Angelesco polynomials $(A_{n,m},B_{n,m})$ then satisfy $$\int_{-1}^1 \bigl( A_{n,m}(x) \chi_{[-1,0]}(x) + B_{n,m}(x) \chi_{[0,1]}(x) \bigr) x^k |x|^\beta (1-x^2)^\alpha\, dx = 0, \qquad 0 \leq k \leq n+m-2,$$ with the normalization given by $$\int_{-1}^1 \bigl( A_{n,m}(x) \chi_{[-1,0]}(x) + B_{n,m}(x) \chi_{[0,1]}(x) \bigr) x^{n+m-1} |x|^\beta (1-x^2)^\alpha \, dx = 1.$$ The polynomials $A_{n,m}$ and $B_{n,m}$ on the diagonal can be expressed in term of the polynomials $$\label{pnab} p_n(x;\alpha,\beta) = \sum_{k=0}^n \binom{n}{k} \frac{\Gamma(n+\alpha+\frac{\beta+k}{2}+1)}{\Gamma(n+\alpha+1)\Gamma(\frac{\beta+k}{2}+1)} (-1)^{n-k} x^k.$$ \[thm22\] The type I Jacobi-Angelesco polynomials on the diagonal are given by $$B_{n+1,n+1}(x) = \frac12 \frac{(2\alpha+\beta+2n+2)_{n+1}}{n!} p_n(x;\alpha,\beta), \quad A_{n+1,n+1}(x) = -B_{n+1,n+1}(-x).$$ If we take $A_{n+1,n+1}(x)=-B_{n+1,n+1}(-x)$, then the integral for the orthogonality conditions is $$\begin{gathered} \int_{-1}^1 \bigl( A_{n+1,n+1}(x) \chi_{[-1,0]}(x) + B_{n+1,n+1}(x) \chi_{[0,1]}(x) \bigr) x^k |x|^\beta (1-x^r)^\alpha\, dx \\ = \bigl( 1 - (-1)^k \bigr) \int_0^1 B_{n+1,n+1}(x) x^k x^\beta (1-x^2)^\alpha\, dx . \end{gathered}$$ This is $0$ whenever $k$ is an even integer, so we only need to prove $$\label{odd} \int_0^1 p_n(x;\alpha,\beta) x^{2j+1} x^\beta (1-x^2)^\alpha \, dx = 0, \qquad 0 \leq j \leq n-1.$$ Take the polynomial $(1-x^2)^\ell -1 = \sum_{k=1}^\ell \binom{\ell}{k} (-1)^k x^{2k}$, then clearly $\bigl((1-x^2)^\ell-1)\bigr)/x$ is an odd polynomial of degree $2\ell-1$ and thus it is sufficient to prove $$\int_0^1 p_n(x;\alpha,\beta) \frac{(1-x^2)^\ell-1}{x} x^\beta (1-x^2)^\alpha \, dx = S_\ell-S_0 = 0, \qquad 1 \leq \ell \leq n,$$ where $$S_\ell = \int_0^1 p_n(x;\alpha,\beta) x^{\beta-1}(1-x^2)^{\ell+\alpha} \, dx .$$ Using the expression we find $$\begin{aligned} S_\ell &= \sum_{k=0}^n \binom{n}{k} \frac{\Gamma(n+\alpha+\frac{\beta+k}2+1)}{\Gamma(n+\alpha+1)\Gamma(\frac{\beta+k}{2}+1)} (-1)^{n-k} \int_0^1 x^{k+\beta-1}(1-x^2)^{\ell+\alpha}\, dx \\ &= \frac{\Gamma(\ell+\alpha+1)}{\Gamma(n+\alpha+1)} \sum_{k=0}^n \binom{n}{k} \frac{(-1)^{n-k}}{\beta+k} \frac{\Gamma(n+\alpha+\frac{\beta+k}2+1)}{\Gamma(\ell+\alpha+\frac{\beta+k}{2}+1)} ,\end{aligned}$$ where we used the beta integral $$\int_0^1 x^{k+\beta-1}(1-x^2)^{\ell+\alpha}\, dx = \frac12 \textup{B}\Bigl(\frac{k+\beta}{2},\ell+\alpha+1\Bigr) = \frac{\Gamma(\frac{k+\beta}2)\Gamma(\ell+\alpha+1)}{2 \Gamma(\frac{k+\beta}2+\ell+\alpha+1)} .$$ From this we see that $$S_\ell-S_0 = \frac{1}{(\alpha+1)_n} \sum_{k=0}^n \binom{n}{k} (-1)^{n-k} \bigl(\ell+\alpha+\frac{\beta+k}2+1\bigr)_{n-\ell} \frac{(\alpha+1)_\ell-\bigl(\alpha+\frac{\beta+k}2+1\bigr)_\ell}{\beta+k},$$ where we used the Pochhammer symbol $$(a)_n = \frac{\Gamma(a+n)}{\Gamma(a)}.$$ We see that for $1 \leq \ell \leq n$ this is of the form $$S_\ell-S_0 = \frac{1}{(\alpha+1)_n} \sum_{k=0}^n \binom{n}{k} (-1)^{n-k} \pi_{n-1}(k) ,$$ where $\pi_{n-1}(k)$ is a polynomial of degree $n-1$ in $k$. Hence by in Lemma \[lem23\], which we prove right after this, we see that $S_\ell-S_0=0$ for $1 \leq \ell \leq n$, proving the relations . For the normalization we need to show that $$\frac{2 n!}{(2n+2\alpha+\beta+2)_{n+1}} = 2 \int_0^1 p_n(x;\alpha,\beta) x^{2n+1} x^\beta (1-x^2)^\alpha\, dx = 2 (-1)^{n+1} (S_{n+1}-S_0).$$ Recall that $S_n-S_0=0$, so that $S_{n+1}-S_0 = S_{n+1}-S_n$, and $$S_{n+1}-S_n = - \sum_{k=0}^n \binom{n}{k} (-1)^{n-k} \frac{1}{2n+2\alpha+\beta+k+2},$$ and then the result follows from in Lemma \[lem23\]. In the proof of the previous theorem we used the following result. \[lem23\] For all integers $n \geq 1$ one has $$\label{lem23a} \sum_{k=0}^n \binom{n}{k} (-1)^{n-k} k^m = 0, \qquad 0 \leq m \leq n-1,$$ and $$\label{lem23b} \sum_{k=0}^n \binom{n}{k} (-1)^k \frac{1}{t+k} = \frac{n!}{(t)_{n+1}}, \qquad t \in \mathbb{R} \setminus \{0,-1,-2,\ldots,-n\}.$$ From Newton’s binomial formula $$(x+y)^n = \sum_{k=0}^n \binom{n}{k} x^k y^{n-k},$$ we find, after differentiating $m$ times with respect to $x$ $$\frac{n!}{(n-m)!} (x+y)^{n-m} = \sum_{k=m}^n \binom{n}{k} \frac{k!}{(k-m)!} x^{k-m} y^{n-k} .$$ If we take $x=1$ and $y=-1$, then for $0 \leq m \leq n-1$ $$\sum_{k=m}^n \binom{n}{k}(-1)^{n-k} (k-m+1)_m = 0.$$ Since $(k-m+1)_m$ is a monic polynomial of degree $m$ in $k$, this is equivalent with . For the second identity we use the beta integral $$\int_0^1 x^{t-1}(1-x)^n\, dx = \textup{B}(t,n+1) = \frac{n!}{(t)_{n+1}},$$ and if we expand $(1-x)^n$ then we also find $$\int_0^1 x^{t-1}(1-x)^n\, dx = \sum_{k=0}^n \binom{n}{k} (-1)^{k} \int_0^1 x^{k+t-1} \, dx = \sum_{k=0}^n \binom{n}{k} \frac{(-1)^k}{t+k}.$$ Comparison of both integrals gives . For the type I Jacobi-Angelesco polynomials above and below the diagonal we need a second family of polynomials $$\label{qnab} q_n(x;\alpha,\beta) = \sum_{k=0}^n \binom{n}{k} \frac{\Gamma(n+\alpha+\frac{\beta+k-1}{2}+1)}{\Gamma(n+\alpha+1)\Gamma(\frac{\beta+k-1}{2}+1)} (-1)^{n-k} x^k.$$ Observe that $q_n(x;\alpha,\beta) = p_n(x;\alpha,\beta-1)$. We then have \[thm24\] The type I Jacobi-Angelesco polynomials near the diagonal are given by $$\begin{aligned} \gamma_n(\alpha,\beta) B_{n+1,n}(x) &=& \nu_{n,1}(\alpha,\beta) q_n(x;\alpha,\beta) - \nu_{n,2}(\alpha,\beta) p_n(x;\alpha,\beta), \label{B2-down} \\ \gamma_n(\alpha,\beta) B_{n,n+1}(x) &=& \nu_{n,1}(\alpha,\beta) q_n(x;\alpha,\beta) + \nu_{n,2}(\alpha,\beta) p_n(x;\alpha,\beta), \label{B2-up} \end{aligned}$$ and $$\label{A2-updown} A_{n+1,n}(x) = B_{n,n+1}(-x), \quad A_{n,n+1}(x) = B_{n+1,n}(-x),$$ where $\nu_{n,1}(\alpha,\beta)$ and $\nu_{n,2}(\alpha,\beta)$ are the leading coefficients of $p_n(x;\alpha,\beta)$ and $q_n(x;\alpha,\beta)$ respectively $$\nu_{n,1}(\alpha,\beta) = \frac{\Gamma\bigl(n+\alpha+\frac{\beta+n}2 +1 \bigr)}{\Gamma(n+\alpha+1) \Gamma \bigl( \frac{\beta+n}2 +1 \bigr)}, \quad \nu_{n,2}(\alpha,\beta) = \frac{\Gamma\bigl(n+\alpha+\frac{\beta+n+1}2 \bigr)}{\Gamma(n+\alpha+1) \Gamma \bigl( \frac{\beta+n+1}2 \bigr)},$$ and $$\gamma_n(\alpha,\beta) = \frac{2 n! \nu_{n,1}(\alpha,\beta)}{(2n+2\alpha+\beta+1)_{n+1}} .$$ The degree of the polynomial $B_{n+1,n}$ is $n-1$ since the leading coefficients of $p_n$ and $q_n$ are cancelled by subtracting the polynomials, but the degree of $B_{n,n+1}$ is $n$ since we add the polynomials now. So it remains to prove the orthogonality. By we see that $$\begin{gathered} \int_{-1}^0 x^k A_{n+1,n}(x) |x|^\beta (1-x^2)^\alpha \, dx + \int_0^1 x^k B_{n+1,n}(x) x^\beta (1-x^2)^\alpha \, dx \\ = (-1)^k \int_0^1 x^k B_{n,n+1}(x) x^\beta (1-x^2)^\alpha\, dx + \int_0^1 x^k B_{n+1,n}(x) x^\beta (1-x^2)^\alpha\, dx . \end{gathered}$$ By using – this is equal to $$\begin{gathered} \left( 1+(-1)^k \right) \frac{\nu_{n,1}}{\gamma_n} \int_0^1 x^k q_n(x;\alpha,\beta) x^\beta (1-x^2)^\alpha\, dx \\ - \left( 1-(-1)^k \right) \frac{\nu_{n,2}}{\gamma_n} \int_0^1 x^k p_n(x;\alpha,\beta) x^\beta (1-x^2)^\alpha\, dx . \end{gathered}$$ When $k=2j+1$ is odd $(0 \leq j \leq n-1$) this is $$-2 \frac{\nu_{n,2}}{\gamma_n} \int_0^1 x^{2j+1} p_n(x;\alpha,\beta) x^\beta (1-x^2)^\alpha\, dx ,$$ which vanishes because of . When $k=2j$ is even $(0 \leq j \leq n-1)$ the integral reduces to $$2 \frac{\nu_{n,1}}{\gamma_n} \int_0^1 x^{2j} q_n(x;\alpha,\beta) x^\beta (1-x^2)^\alpha\, dx ,$$ and this vanishes because $q_n(x;\alpha,\beta) = p_n(x;\alpha,\beta-1)$ and again by . This proves the orthogonality. For the normalization we need $$\begin{gathered} 1 = \int_{-1}^0 x^{2n} A_{n+1,n}(x) |x|^\beta (1-x^2)^\alpha \, dx + \int_0^1 x^{2n} B_{n+1,n}(x) x^\beta (1-x^2)^\alpha \, dx \\ = 2 \frac{\nu_{n,1}}{\gamma_n} \int_0^1 x^{2n} q_n(x;\alpha,\beta) x^\beta (1-x^2)^\alpha\, dx . \end{gathered}$$ The latter integral is $$\begin{aligned} \int_0^1 x^{2n} q_n(x;\alpha,\beta) x^\beta (1-x^2)^\alpha\, dx &= \int_0^1 x^{2n+1} p_n(x;\alpha,\beta-1) x^{\beta-1}(1-x^2)^\alpha\, dx \\ &= \frac{n!}{(2n+2\alpha+\beta+1)_{n+1}}, \end{aligned}$$ which gives the normalizing constant $\gamma_n(\alpha,\beta)$. Recurrence relation ------------------- Multiple orthogonal polynomials satisfy a system of linear recurrence relations connecting the nearest neighbors [@wva]. For the type II multiple orthogonal polynomials they are $$\begin{aligned} xP_{n,m}(x) = P_{n+1,m}(x) + c_{n,m} P_{n,m}(x) + a_{n,m}P_{n-1,m}(x) + b_{n,m} P_{n,m-1}(x), \\ xP_{n,m}(x) = P_{n,m+1}(x) + d_{n,m} P_{n,m}(x) + a_{n,m}P_{n-1,m}(x) + b_{n,m} P_{n,m-1}(x).\end{aligned}$$ The recurrence relations are very similar for the type I multiple orthogonal polynomials: $$\begin{aligned} xQ_{n,m}(x) = Q_{n-1,m}(x) + c_{n-1,m} Q_{n,m}(x) + a_{n,m}Q_{n+1,m}(x) + b_{n,m} Q_{n,m+1}(x), \\ xQ_{n,m}(x) = Q_{n,m-1}(x) + d_{n,m-1} Q_{n,m}(x) + a_{n,m}Q_{n+1,m}(x) + b_{n,m} Q_{n,m+1}(x).\end{aligned}$$ The same recurrence relation holds with $Q_{n,m}$ replaced by $A_{n,m}$ or $B_{n,m}$. The coefficients $a_{n,m},b_{n,m}$ can be computed as $$a_{n,m} = \frac{\kappa_{n,m}}{\kappa_{n+1,m}}, \quad b_{n,m} = \frac{\lambda_{n,m}}{\lambda_{n,m+1}},$$ where $\kappa_{n,m}$ and $\lambda_{n,m}$ are the leading coefficients of $A_{n,m}$ and $B_{n,m}$ respectively. This can easily be seen by checking the leading coefficients in the recurrence relation. If we write $$A_{n,m}(x) = \kappa_{n,m} x^{n-1} + \delta_{n,m} x^{n-2} + \cdots, \quad B_{n,m}(x) = \lambda_{n,m} x^{m-1} + \epsilon_{n,m} x^{m-2} + \cdots,$$ then one also has $$c_{n-1,m} = \frac{\delta_{n,m}}{\kappa_{n,m}} - \frac{\delta_{n+1,m}}{\kappa_{n+1,m}} - \frac{\lambda_{n,m}}{\lambda_{n,m+1}} \frac{\kappa_{n,m+1}}{\kappa_{n,m}},$$ $$d_{n,m-1} = \frac{\epsilon_{n,m}}{\lambda_{n,m}} - \frac{\epsilon_{n,m+1}}{\lambda_{n,m+1}} - \frac{\lambda_{n+1,m}}{\lambda_{n,m}} \frac{\kappa_{n,m}}{\kappa_{n+1,m}}.$$ For the Jacobi-Angelesco polynomials near the diagonal we then have The recurrence coefficients for the Jacobi-Angelesco polynomials on the diagonal are $$a_{n,n} = \frac{n(n+\alpha)(2n+2\alpha+\beta)}{(3n+2\alpha+\beta+1)(3n+2\alpha+\beta)(3n+2\alpha+\beta-1)}, \quad b_{n,n} = a_{n,n},$$ and $$c_{n-1,n} = \frac{(2n+2\alpha+\beta-1) \Gamma(n+\alpha+\frac{n+\beta}{2}-1) \Gamma(\frac{n+\beta+1}{2})} {(3n+2\alpha+\beta-1) \Gamma(n+\alpha+\frac{n+\beta-1}{2}) \Gamma(\frac{n+\beta}{2})} , \quad d_{n,n-1} = -c_{n-1,n} .$$ The two leading coefficients can easily be obtained from Theorem \[thm22\] and Theorem \[thm24\] and give $$\lambda_{n+1,n+1} = \frac12 \frac{(2n+2\alpha+\beta+2)_{n+1}}{n!} \nu_{n,1}(\alpha,\beta), \quad \kappa_{n+1,n+1} = (-1)^{n+1} \lambda_{n+1,n+1},$$ and $$\lambda_{n,n+1} = \frac{\nu_{n,2}(\alpha,\beta)(2n+2\alpha+\beta+1)_{n+1}}{n!}, \quad \kappa_{n+1,n} = (-1)^n \lambda_{n,n+1}.$$ From this the coefficients $a_{n,n}$ and $b_{n,n}$ follow easily using the formulas above. For $c_{n-1,n}$ and $d_{n,n-1}$ one also needs the last but one leading coefficient and the calculus is a bit longer. Differential equation --------------------- \[thm:diff2\] For $\alpha,\beta >-1$ the polynomial $p_n(x;\alpha,\beta)$ given in satisfies the third order differential equation $$\begin{gathered} \label{diff2} x(1-x^2)y''' + \bigl(\beta+2 -(2\alpha+\beta+6)x^2\bigr) y'' + (n-1)(3n+4\alpha+2\beta+6) xy' \\ = n(n-1)(2n+2\alpha+\beta+2) y. \end{gathered}$$ From it is easy to see that $$\label{pndif} p_n'(x;\alpha,\beta) = np_{n-1}(x;\alpha+1,\beta+1),$$ and then of course also $$\label{pndif2} p_n''(x;\alpha,\beta) = n(n-1)p_{n-2}(\alpha+2,\beta+2).$$ Hence differentiation lowers the degree $n$ but increases the parameters $\alpha$ and $\beta$. On the other hand, one has for $\alpha,\beta>1$ $$\begin{gathered} \label{wpndif} \bigl( x^\beta (1-x^2)^\alpha p_n(x;\alpha,\beta) \bigr)' = x^{\beta-1}(1-x^2)^{\alpha-1} \bigl( (2n+2\alpha+\beta)p_{n+2}(x;\alpha-2,\beta-2) \\ -(3n+4\alpha+2\beta) x p_{n+1}(x;\alpha-1,\beta-1) \bigr). \end{gathered}$$ Indeed, if we work out the left hand side, then $$\bigl( x^\beta (1-x^2)^\alpha p_n(x;\alpha,\beta) \bigr)' = x^{\beta-1}(1-x^2)^{\alpha-1} \pi_{n+2}(x),$$ where $\pi_{n+2}$ is a polynomial of degree $n+2$ given by $$\label{pi1} \pi_{n+2}(x) = \bigl( \beta(1-x^2)-2\alpha x^2 \bigr) p_n(x;\alpha,\beta) +x(1-x^2)p_n'(x;\alpha,\beta).$$ One can check, by comparing coefficients and using , that $$\label{pi2} \pi_{n+2}(x) = (2n+2\alpha+\beta)p_{n+2}(x;\alpha-2,\beta-2) -(3n+4\alpha+2\beta) x p_{n+1}(x;\alpha-1,\beta-1),$$ but alternatively one can also observe that for $\alpha,\beta >0$ $$\begin{aligned} \int_0^1 x^{\beta-2} (1-x^2)^{\alpha-1} \pi_{n+2}(x) x^{2k+1} \,dx &=& \int_0^1 \bigl( x^\beta (1-x^2)^\alpha p_n(x;\alpha,\beta) \bigr)' x^{2k}\, dx \\ & = & - 2k \int_0^1 x^\beta (1-x^2)^\alpha p_n(x;\alpha,\beta) x^{2k-1}\, dx \\ & = & 0, \qquad 0 \leq k \leq n,\end{aligned}$$ where we used integration by parts and the orthogonality to odd powers . Therefore $\pi_{n+2}$ is a polynomial of degree $n+2$ which satisfies $n+1$ orthogonality conditions for odd powers with the weight $x^{\beta-2}(1-x^2)^{\alpha-1}$ on $[0,1]$, so it belongs to a linear space of polynomials of dimension $2$ and can be written as a linear combination of two linearly independent polynomials from that space. For $\alpha,\beta>1$ the polynomials $p_{n+2}(x;\alpha-2,\beta-2)$ and $xp_{n+1}(x;\alpha-1,\beta-1)$ are two such polynomials and by they are orthogonal to odd powers $x^{2k+1}$ for $0 \leq k \leq n$ with weight $x^{\beta-2}(1-x^2)^{\alpha-1}$ on $[0,1]$, hence $\pi_{n+2}(x) = a_n p_{n+2}(x;\alpha-2,\beta-2) +b_n xp_{n+1}(x;\alpha-1,\beta-1)$. The coefficients $a_n$ and $b_n$ can be found by comparing the leading coefficient and the constant coefficient. To find the differential equation we multiply by $x^{\beta+2}(1-x^2)^{\alpha+2}$ and differentiate to find $$\begin{gathered} x^{\beta+1}(1-x^2)^{\alpha+1} \Bigl[(\beta+2)(1-x^2)-2x^2(\alpha+2)\Bigr] p_n''(x;\alpha,\beta) \\ + x^{\beta+2}(1-x^2)^{\alpha+2} p_n'''(x;\alpha,\beta) \\ = n(n-1)x^{\beta+1}(1-x^2)^{\alpha+1} \Bigl( (2n+2\alpha+\beta+2)p_n(x;\alpha,\beta) \\ - (3n+4\alpha+2\beta+6)xp_{n-1}(x,\alpha+1,\beta+1) \Bigr),\end{gathered}$$ where we used the property for the right hand side. Remove the common factor $x^{\beta+1}(1-x^2)^{\alpha+1}$ and use , then the differential equation follows. Asymptotic zero behavior ------------------------ \[thm:zero2\] The asymptotic zero distribution of the polynomials $p_n(x;\alpha,\beta)$ given in is independent of $\alpha$ and $\beta$ and is given by a measure on $[0,1]$ with density $$\label{u2} u_2(x) = \frac{\sqrt{3}}{2\pi} \frac{(1+\sqrt{1-x^2})^{1/3} + (1-\sqrt{1-x^2})^{1/3}}{x^{1/3} \sqrt{1-x^2}}, \qquad 0 < x < 1.$$ Let $x_{1,n},x_{2,n},\ldots,x_{n,n}$ be the zeros of $p_n(x;\alpha,\beta)$. Since this is an Angelesco system, it is known that these zeros (which are the zeros of $B_{n+1,n+1}$) are simple and on the interval $(0,1)$ (see, e.g, [@NikiSor Prop. 3.4 in Ch. 4.3]). The normalized zero counting measure is $$\mu_n = \frac{1}{n} \sum_{j=1}^n \delta_{x_{j,n}},$$ and its Stieltjes transform is $$S_n(z) = \int_0^1 \frac{d\mu_n(x)}{z-x} = \frac{1}{n} \frac{p_n'(z;\alpha,\beta)}{p_n(z;\alpha,\beta)}.$$ The sequence $(\mu_n)_n$ is a sequence of probability measures on the compact interval $[0,1]$, and hence by Helley’s selection principle [@billingsley §25], it contains a subsequence $(\mu_{n_k})_k$ that converges weakly to a probability measure $\mu$ on $[0,1]$, i.e., $$\lim_{k \to \infty} \int_0^1 f(x)\, d\mu_{n_k}(x) = \lim_{k \to \infty} \frac{1}{n_k} \sum_{j=1}^{n_k} f(x_{j,n_k}) = \int_0^1 f(x)\, d\mu(x),$$ for every continuous function $f$ on $[0,1]$. The weak limit $\mu$ can depend on the subsequence, but we will show it is independent of the choice of converging subsequence. Observe that $p_n'(z) = n p_n(z) S_n(z)$ so that $$p_n''(z) = np_n'(z) S_n(z) + np_n(z) S_n'(z) = n^2 p_n(z) \left( S_n^2(z) + \frac{1}{n} S_n'(z) \right),$$ and $$p_n'''(z) = n^3p_n(z) \left( S_n^3(z) + \frac{3}{n} S_n(z)S_n'(z) + \frac{1}{n^2} S_n''(z) \right).$$ Insert this in the differential equation then $$\begin{gathered} \label{diffS} z(1-z^2)n^3 p_n(z) \left( S_n^3 + \frac{3}{n} S_n S_n' + \frac{1}{n^2} S_n'' \right) + [\beta+2-(2\alpha+\beta+6)z^2] n^2 p_n(z) \left( S_n^2 + \frac{1}{n} S_n' \right) \\ + (n-1)( 3n+4\alpha+2\beta+6) znp_n(z) S_n - n(n-1)(2n+2\alpha+\beta+2)p_n(z) = 0. \end{gathered}$$ The weak convergence of the sequence $(\mu_{n_k})_k$ to $\mu$ implies that $S_{n_k}$ converges uniformly on compact subsets of $\mathbb{C} \setminus [0,1]$ to the Stieltjes transform $S$ of $\mu$, $$S(z) = \int_0^1 \frac{d\mu(x)}{z-x}.$$ But then also $S_{n_k}'$ and $S_{n_k}''$ converge uniformly on compact subsets of $\mathbb{C} \setminus [0,1]$ to $S'$ and $S''$, respectively. Then taking the limit for $n=n_k \to \infty$ in , after dividing by $n^3p_n(z)$, gives the algebraic equation $$z(1-z^2) S^3(z) + 3zS(z) - 2 = 0.$$ Observe that this equation does not contain $\alpha$ and $\beta$ anymore. This algebraic equation has three solutions, and we need the solution that gives a Stieltjes transform, in particular we need the solution which is analytic on $\mathbb{C} \setminus [0,1]$ and $\lim_{z \to \infty} zS(z) = 1$. Solving the cubic equation gives the following three solutions $$S_1(z) = \frac{z^4-z^2+\bigl( (-1+(1-z^2)^{-1/2}) (z^2-1)z^2 \bigr)^{2/3}}{(z^3-z^2)\bigl( (-1+(1-z^2)^{-1/2})(z^2-1)z^2 \bigr)^{1/3}},$$ $$S_2(z) = - \frac{(1-i\sqrt{3}) \bigl( (-1+(1-z^2)^{-1/2})(z^2-1)^2z^2 \bigr)^{2/3} + (1+i\sqrt{3})z^2(z^2-1)} {(2z^3-2z) \bigl( (-1+(1-z^2)^{-1/2})(z^2-1)^2z^2 \bigr)^{1/3}},$$ $$S_3(z) = - \frac{(1+i\sqrt{3}) \bigl( (-1+(1-z^2)^{-1/2})(z^2-1)^2z^2 \bigr)^{2/3} + (1-i\sqrt{3})z^2(z^2-1)} {(2z^3-2z) \bigl( (-1+(1-z^2)^{-1/2})(z^2-1)^2z^2 \bigr)^{1/3}}.$$ One can check that $$\lim_{z \to \infty} zS_1(z) = -2, \quad \lim_{z\to \infty} zS_2(z) = \lim_{z \to \infty} zS_3(z) = 1,$$ hence either $S_2$ or $S_3$ is the desired solution. From the Stieltjes-Perron inversion formula (or Sokhotsky-Plemelj formula) $$u(x) = - \frac{1}{\pi} \lim_{\epsilon \to 0+} \Im S(x+i\epsilon) , \qquad 0 < x < 1,$$ we find that $S_2$ is the correct solution and it is the Stieltjes transform of the density . Hence every convergent subsequence has the same limit, and from the Grommer-Hamburger theorem [@Geronimo] it follows that $\mu_n$ converges weakly to the measure with density $u_2$. Type I Jacobi-Angelesco polynomials for general $r$ {#secII} =================================================== We now consider the type I Jacobi-Angelesco polynomials on the $r$-star for general $r \geq 1$. The results and the proofs are similar to the case $r=2$ but they are more complicated and technical. This is why we decided to explain the case $r=2$ in detail and only give the results and the modifications in the proof for the general case in this section. The type I Jacobi-Angelesco polynomials for the multi-index $(n_1,n_2,\ldots,n_r)$ on the $r$-star (see Figure \[fig:rstar\] for $r=5$) and parameters $\alpha,\beta >-1$ are given by the vector of polynomials $(A_{\vec{n},1}^{(\alpha,\beta)},\ldots,A_{\vec{n},r}^{(\alpha,\beta)})$ which is uniquely defined by 1. degree conditions: the degree of $A_{\vec{n},j}^{(\alpha,\beta)}$ is $n_j-1$, 2. orthogonality conditions $$\label{orthor} \sum_{j=1}^r \int_0^{\omega^{j-1}} x^k A_{\vec{n},j}^{(\alpha,\beta)}(x) |x|^\beta (1-x^r)^\alpha\, dx = 0, \qquad 0 \leq k \leq |\vec{n}|-2,$$ 3. normalization condition: $$\label{normr} \sum_{j=1}^r \int_0^{\omega^{j-1}} x^{|\vec{n}|-1} A_{\vec{n},j}^{(\alpha,\beta)}(x) |x|^\beta (1-x^r)^\alpha\, dx = 1.$$ We have used the weight $|x|^\beta(1-x^r)^\alpha$ on the $r$-star so that we can use the rotational symmetry and $\omega=e^{2\pi i/r}$ is the primitive $r$th root of unity. Explicit expression ------------------- The polynomials $A_{\vec{n},j}^{(\alpha,\beta)}$ on the diagonal can be expressed in terms of the polynomials $$\label{pnabr} p_n(x;\alpha,\beta) = \sum_{k=0}^n \binom{n}{k} \frac{\Gamma(n+\alpha+\frac{\beta+k}{r}+1)}{\Gamma(n+\alpha+1)\Gamma(\frac{\beta+k}{r}+1)} (-1)^{n-k} x^k.$$ Observe that for $r=2$ this coincides with . \[thm31\] The type I Jacobi-Angelesco polynomials on the diagonal $\vec{n}=(n+1,$ $n+1,\ldots,n+1)$ are given by $$A_{\vec{n},j}^{(\alpha,\beta)}(x) = \lambda_{n+1,r}^{(\alpha,\beta)} p_n(\omega^{-j+1}x;\alpha,\beta),$$ with normalizing constant $$\lambda_{n+1,r}^{(\alpha,\beta)} = \frac{1}{r} \frac{(rn+r\alpha+\beta+r)_{n+1}}{n!} .$$ The degree conditions are clearly satisfied. The orthogonality conditions become $$\lambda_{n+1,r}^{(\alpha,\beta)} \sum_{j=1}^r (\omega^{j-1})^{k+1} \int_0^1 x^k p_n(x;\alpha,\beta) x^\beta (1-x^r)^\alpha\, dx = 0 ,$$ for $0 \leq k \leq rn+r-2$. Since $\omega$ is the primitive $r$th root of unity, we have $$\sum_{j=1}^r (\omega^{j-1})^{k+1} = \begin{cases} 0, & \textup{if } k+1 \not\equiv 0 \bmod r, \\ r, & \textup{if } k+1 \equiv 0 \bmod r, \end{cases}$$ so we only need to prove $$\label{modr} \int_0^1 x^{rj-1} p_n(x;\alpha,\beta) x^\beta (1-x^r)^\alpha\, dx = 0, \qquad 1 \leq j \leq n.$$ We will consider the polynomials $\bigl((1-x^r)^\ell-1 \bigr)/x$ of order $r\ell-1$ and show that $S_\ell-S_0 = 0$ for $1 \leq \ell\leq n$, where $$S_\ell = \int_0^1 p_n(x;\alpha,\beta) x^{\beta-1} (1-x^r)^{\alpha+\ell}\, dx,$$ as we did in the proof of Theorem \[thm22\]. It turns out that for $1 \leq \ell \leq n$ $$S_\ell - S_0 = \frac{1}{r \Gamma(n+\alpha+1)} \sum_{k=0}^n \binom{n}{k} \pi_{n-1,\ell}\left(\frac{\beta+k}{r} \right),$$ where $\pi_{n-1,\ell}$ is a polynomial of degree $n-1$, and this vanishes because of in Lemma \[lem23\]. For the normalization we need to prove $$\label{norm} S_{n+1}-S_0 = S_{n+1}-S_n = (-1)^{n+1} \frac{n!}{(rn+r\alpha+\beta+r)_{n+1}},$$ and this follows from in Lemma \[lem23\]. Next, we will show that the type I Jacobi-Angelesco polynomials above the diagonal can be written as a linear combination of $r$ polynomials $p_n(x;\alpha,\beta-j)$ with $0 \leq j \leq r-1$. \[thm32\] Let $\vec{n}=(n,n,\ldots,n)$ and $\vec{e}_k$ be the unit vector in $\mathbb{N}^r$ with 1 on the $k$th position. The type I Jacobi-Angelesco polynomials $A_{\vec{n}+\vec{e}_k,j}^{(\alpha,\beta)}$ are given by $$\label{Aplus} A_{\vec{n}+\vec{e}_k,j}^{(\alpha,\beta)}(x) = A_{j-k \bmod r}(\omega^{-j+1}x) \omega^{-k+1} ,$$ where the polynomials $A_{\ell}$, $0 \leq \ell \leq r-1$ are given by $$\label{Aell} \tau_{n,r}^{(\alpha,\beta)} A_\ell(x) = \sum_{j=0}^{r-1} \frac{\omega^{\ell j}}{\nu_n^{(\alpha,\beta-j)}} p_n(x;\alpha,\beta-j),$$ with normalizing constant $$\label{tau} \tau_{n,r}^{(\alpha,\beta)} = \frac{r n! \Gamma(n+\alpha+1) \Gamma(\frac{\beta+n+1}{r}) \Gamma(rn+r\alpha+\beta+1)} {\Gamma(n+\alpha+ \frac{\beta+n+1}{r}) \Gamma(rn+n+r\alpha+\beta+2)},$$ and $\nu_n^{(\alpha,\beta)}$ is the leading coefficient of $p_n(x;\alpha,\beta)$ $$\label{nu} \nu_n^{(\alpha,\beta)} = \frac{\Gamma(n+\alpha+\frac{\beta+n}{r}+1)}{\Gamma(n+\alpha+1) \Gamma(\frac{\beta+n}{r}+1)} .$$ We will first determine the degree of the polynomials $A_\ell$. For $\ell=0$ we see that $\deg A_0 = \deg p_n(x;\alpha,\beta-j) =n$, which implies that $\deg A_{\vec{n}+\vec{e}_j,j}^{(\alpha,\beta)} = n$ for all $j=1,2,\ldots,r$. For $\ell=1,2,\ldots,r-1$ the coefficient of $x^n$ on the right hand side of is given by $$\sum_{j=0}^{r-1} \omega^{\ell j} = \frac{1-\omega^{r \ell}}{1-\omega^\ell} = 0.$$ Therefore for all $k \neq j$ we have $\deg A_{\vec{n}+\vec{e}_k,j}^{(\alpha,\beta)} < n$ and one can check that it is in fact $n-1$. For the orthogonality and the normalization we need the following integral to vanish for all $\ell = 0,1,2,\ldots,rn-1$ and to be equal to one for $\ell=rn$: $$\sum_{j=1}^r \int_0^{\omega^{j-1}} x^\ell A_{\vec{n}+\vec{e}_k,j}^{(\alpha,\beta)}(x) |x|^\beta (1-x^r)^\alpha\, dx ,$$ and this expression is equal to $$\frac{\omega^{-k+1}}{\tau_{n,r}^{(\alpha,\beta)}} \sum_{m=0}^{r-1} \frac{\omega^{m(-k+1)}}{\nu_{n}^{(\alpha,\beta-m)}} \sum_{j=1}^r (\omega^{\ell+m+1})^{j-1} \int_0^1 p_n(x;\alpha,\beta-m) x^{\ell+\beta} (1-x^r)^{\alpha}\, dx .$$ The second sum in this expression is $$\sum_{j=1}^r (\omega^{\ell+m+1})^{j-1} = \begin{cases} r, & \textrm{if } \ell+m+1 \equiv 0 \bmod r, \\ 0, & \textrm{if } \ell+m+1 \not\equiv 0 \bmod r. \end{cases}$$ Therefore we need to show that $$\int_0^1 p_n(x;\alpha,\beta-m) x^{rj-m-1+\beta} (1-x^r)^{\alpha}\, dx = 0, \qquad 1 \leq j \leq n,$$ and the latter follows from . For the normalization we need to show that $$\begin{aligned} 1 &=& \sum_{j=1}^r \int_0^{\omega^{j-1}} x^{rn} A_{\vec{n}+\vec{e}_k,j}^{(\alpha,\beta)}(x) |x|^\beta (1-x^r)^\alpha \, dx \\ &=& r \frac{\omega^{r(k-1)}}{\tau_{n,r}^{(\alpha,\beta)} \nu_{n}^{(\alpha,\beta-r+1)}} \int_0^1 p_n(x;\alpha,\beta-r+1) x^{rn+\beta} (1-x^r)^\alpha\, dx, \end{aligned}$$ and this follows from the explicit expressions and for $\tau_{n,r}^{(\alpha,\beta)}$ and $\nu_n^{(\alpha,\beta)}$ and the expression for the integral. We also give an explicit expression for the type I Jacobi-Angelesco polynomials below the diagonal, i.e., for $A_{\vec{n}-\vec{e}_k,j}^{(\alpha,\beta)}$, where $\vec{n}=(n,n,\ldots,n)$. \[thm33\] For every $r > 1$, $\alpha,\beta>-1$ and $\vec{n}=(n,n,\ldots,n)$ with $n > 0$ we have $$\begin{gathered} \label{Amin} \gamma_{n,r}^{(\alpha,\beta)} A_{\vec{n}-\vec{e}_k,j}^{(\alpha,\beta)}(x) = \omega^{j-1} \nu_{n-1}^{(\alpha,\beta)} p_{n-1}(\omega^{-j+1}x;\alpha,\beta-1) \\ - \omega^{k-1} \nu_{n-1}^{(\alpha,\beta-1)} p_{n-1}(\omega^{-j+1}x;\alpha,\beta) , \end{gathered}$$ where the normalizing constant $\gamma_{n,r}^{(\alpha,\beta)}$ is given by $$\label{gamma} \gamma_{n,r}^{(\alpha,\beta)} = \frac{r(n-1)! \Gamma(n+\alpha+\frac{n+\beta-1}{r})} {\Gamma(n+\alpha) \Gamma(\frac{n+\beta-1}{r}+1) (rn+r\alpha+\beta-1)_n} .$$ Observe that for $j\neq k$ the degree of $A_{\vec{n}-\vec{e}_k,j}^{(\alpha,\beta)}$ is $n-1$ but that for $j=k$ the leading term $x^{n-1}$ vanishes and the degree is $n-2$. For the orthogonality relations we need to verify that the following integral vanishes for $0 \leq \ell \leq rn-3$: $$\sum_{j=1}^r \int_0^{\omega^{j-1}} x^\ell A_{\vec{n}-\vec{e}_k,j}^{(\alpha,\beta)}(x) |x|^\beta (1-x^r)^\alpha\, dx ,$$ and this expression is equal to $$\begin{gathered} \frac{1}{\gamma_{n,r}^{(\alpha,\beta)}} \int_0^1 x^{\ell+\beta} (1-x^r)^\alpha \Bigr( \nu_{n-1}^{(\alpha,\beta)} \sum_{j=1}^r (\omega^{j-1})^{\ell+2} p_{n-1}(x;\alpha,\beta-1) \\ -\ \omega^{k-1} \nu_{n-1}^{(\alpha,\beta-1)} \sum_{j=1}^r (\omega^{j-1})^{\ell+1} p_{n-1}(x;\alpha,\beta) \Bigr)\, dx . \end{gathered}$$ The two sums involving the roots of unity $\omega$ are $$\sum_{j=1}^r (\omega^{\ell+2})^{j-1} = \begin{cases} r, & \textrm{if } \ell+2 \equiv 0 \bmod r, \\ 0, & \textrm{if } \ell+2 \not\equiv 0 \bmod r, \end{cases}$$ and $$\sum_{j=1}^r (\omega^{\ell+1})^{j-1} = \begin{cases} r, & \textrm{if } \ell+1 \equiv 0 \bmod r, \\ 0, & \textrm{if } \ell+1 \not\equiv 0 \bmod r, \end{cases}$$ so the integral vanishes whenever $\ell+2 \not\equiv 0 \bmod r$ and $\ell+1 \not\equiv 0 \bmod r$. In case $\ell=rj-2$ for $1 \leq j \leq n-1$ we need to verify $$\int_0^1 x^{rj-2+\beta} (1-x^r)^\alpha p_{n-1}(x;\alpha,\beta-1)\,dx = 0,$$ and this holds because of . In case $\ell=rj-1$ for $1 \leq j \leq n-1$ we need $$\int_0^1 x^{rj-1+\beta} (1-x^r)^\alpha p_{n-1}(x;\alpha,\beta)\,dx = 0,$$ which again follows from . For the normalization we need to show that $$\begin{aligned} 1 &=& \sum_{j=1}^r \int_0^{\omega^{j-1}} x^{rn-2} A_{\vec{n}-\vec{e}_k,j}^{(\alpha,\beta)}(x) |x|^\beta (1-x^r)^\alpha\, dx \\ &=& \frac{r}{\gamma_{n,r}^{(\alpha,\beta)}} \nu_{n-1}^{(\alpha,\beta)} \int_0^1 x^{rn+\beta-2} (1-x^r)^\alpha p_{n-1}(x;\alpha,\beta-1)\, dx ,\end{aligned}$$ and this follows by using the normalization given in Theorem \[thm31\]. Recurrence relation ------------------- For general $r$ the nearest neighbor recurrence relations for the type I multiple orthogonal polynomials are $$xQ_{\vec{n}}(x) = Q_{\vec{n}-\vec{e}_k}(x) + b_{\vec{n}-\vec{e}_k,k} Q_{\vec{n}}(x) + \sum_{\ell=1}^r a_{\vec{n},\ell} Q_{\vec{n}+\vec{e}_\ell}(x),$$ for $1 \leq k \leq r$, where $Q_{\vec{n}}(x) = \sum_{j=1}^r A_{\vec{n},j}(x) \chi_{[0,\omega^{j-1}]}(x)$. This relation can also be stated in terms of the individual polynomials on each part of the $r$-star and one has for every $j=1,2,\ldots,r$ $$xA_{\vec{n},j}(x) = A_{\vec{n}-\vec{e}_k,j}(x) + b_{\vec{n}-\vec{e}_k,k} A_{\vec{n},j}(x) + \sum_{\ell=1}^r a_{\vec{n},\ell} A_{\vec{n}+\vec{e}_\ell,j}(x),$$ for $1 \leq k \leq r$. For the type II multiple orthogonal polynomials the recurrence relations are $$xP_{\vec{n}}(x) = P_{\vec{n}+\vec{e}_k}(x) + b_{\vec{n},k} P_{\vec{n}}(x) + \sum_{\ell=1}^r a_{\vec{n},\ell} P_{\vec{n}-\vec{e}_\ell}(x),$$ for $1 \leq k \leq r$. An explicit expression for the recurrence coefficients for the Jacobi-Angelesco polynomials near the diagonal is given by \[prop34\] Let $\vec{n}=(n,n,\ldots,n)$ be a diagonal multi-index for $r \geq 1$. Then $$a_{\vec{n},k} = a_{n,r}^{(\alpha,\beta)} \omega^{2(k-1)}, \qquad 1 \leq k \leq r,$$ where $$a_{n,r}^{(\alpha,\beta)} = \frac{n(n+\alpha)(rn+r\alpha+\beta)}{r(rn+n+r\alpha+\beta)(rn+n+r\alpha +\beta + 1)} \frac{\Gamma(\frac{\beta+n+1}{r}) \Gamma(n+\alpha+\frac{\beta+n-1}{r})}{\Gamma(\frac{\beta+n-1}{r}+1) \Gamma(n+\alpha+\frac{\beta+n+1}{r})} .$$ Furthermore for $r >1$ $$b_{\vec{n}-\vec{e}_k,k} = b_{n,r}^{(\alpha,\beta)} \omega^{k-1}, \qquad 1 \leq k \leq r,$$ where $$b_{n,r}^{(\alpha,\beta)} = \frac{(n+\alpha+\frac{\beta-1}{r}) \Gamma(n+\alpha+\frac{n+\beta-2}{r}) \Gamma(\frac{n+\beta-1}{r}+1)} {(n+\alpha+\frac{n+\beta-1}{r}) \Gamma(n+\alpha+\frac{n+\beta-1}{r}) \Gamma(\frac{n+\beta-2}{r}+1)}.$$ If we write $$A_{\vec{n},k}(x) = \kappa_{\vec{n},k} x^{n-1} + \delta_{\vec{n},k} x^{n-2} + \cdots,$$ then $$a_{\vec{n},k} = \frac{\kappa_{\vec{n},k}}{\kappa_{\vec{n}+\vec{e}_k,k}},$$ and this ratio can be evaluated by using Theorem \[thm31\] which gives $$\kappa_{\vec{n},k} = \frac{1}{r} \frac{(rn+r\alpha+\beta)_n}{(n-1)!} \frac{\Gamma(n+\alpha+\frac{\beta+n-1}{r})}{\Gamma(n+\alpha)\Gamma(\frac{\beta+n-1}{r}+1)} \omega^{(-k+1)(n-1)},$$ and Theorem \[thm32\] which gives $$\kappa_{\vec{n}+\vec{e}_k,k} = \frac{(rn+r\alpha+\beta+1)_{n+1}}{n!} \frac{\Gamma(n+\alpha+\frac{\beta+n+1}{r})}{\Gamma(n+\alpha+1)\Gamma(\frac{\beta+n+1}{r})} \omega^{(-k+1)(n+1)}.$$ The coefficients $b_{\vec{n}-\vec{e}_k,k}$ can be computed in a similar way, but the computations are a bit longer and only the case $r>1$ is covered. For $r=1$ the computations are slightly different but in that case the result is known because this corresponds to Jacobi polynomials on $[0,1]$. The expression for $b_{\vec{n}-\vec{e}_k,k}$ is $$b_{\vec{n}-\vec{e}_k,k} = \frac{\delta_{\vec{n},k}}{\kappa_{\vec{n},k}} - \frac{\delta_{\vec{n}+\vec{e}_k,k}}{\kappa_{\vec{n}+\vec{e}_k,k}} - \sum_{\ell=1,\ell\neq k}^r \frac{\kappa_{\vec{n},\ell}}{\kappa_{\vec{n},k}} \frac{\kappa_{\vec{n}+\vec{e}_\ell,k}}{\kappa_{\vec{n}+\vec{e}_\ell,\ell}}.$$ From Theorem \[thm31\] one finds $$\delta_{\vec{n},k} = - \frac{(rn+r\alpha+\beta)_n\Gamma(n+\alpha+\frac{n+\beta-2}{r})} {r(n-2)!\Gamma(n+\alpha)\Gamma(\frac{n+\beta-2}{r}+1)} \omega^{(-k+1)(n-2)}, \qquad n\geq 2 ,$$ and from Theorem \[thm32\] $$\delta_{\vec{n}+\vec{e}_k,k} = - \frac{n\omega^{(-k+1)n}}{\tau_{n,r}^{(\alpha,\beta)}} \sum_{j=0}^{r-1} \frac{\Gamma(\frac{n+\beta-j}{r}+1)\Gamma(n+\alpha+\frac{n+\beta-1-j}{r}+1)} {\Gamma(\frac{n+\beta-1-j}{r}+1)\Gamma(n+\alpha+\frac{n+\beta-j}{r}+1)},$$ and for $\ell\neq k$ $$\kappa_{\vec{n}+\vec{e}_\ell,k} = - \frac{n \omega^{-\ell+1-(k-1)(n-1)}}{\tau_{n,r}^{(\alpha,\beta)}} \sum_{j=0}^{r-1} \omega^{(k-\ell)j} \frac{\Gamma(\frac{n+\beta-j}{r}+1)\Gamma(n+\alpha+\frac{n+\beta-1-j}{r}+1)} {\Gamma(\frac{n+\beta-1-j}{r}+1)\Gamma(n+\alpha+\frac{n+\beta-j}{r}+1)}.$$ Combining all these results gives the desired expression for $b_{\vec{n}-\vec{e}_k,k}$. Observe that one can easily find the asymptotic behavior of these recurrence coefficients as $n\to \infty$ by using [@NIST 5.11.12] $$\frac{\Gamma(n+a)}{\Gamma(n+b)} \sim n^{a-b}, \qquad n \to \infty,$$ which gives $$\lim_{n \to \infty} a_{n,r}^{(\alpha,\beta)} = \frac{r}{(r+1)^{2+2/r}}, \quad \lim_{n \to \infty} b_{n,r}^{(\alpha,\beta)} = \frac{r}{(r+1)^{1+1/r}}.$$ The limit for $b_{n,r}^{(\alpha,\beta)}$ is valid for $r>1$. Differential equation --------------------- In this section we will give a linear differential equation of order $r+1$ for the polynomial $p_n(x;\alpha,\beta)$ given in . The differential equation is a combination of lowering and raising operators for these polynomials, i.e., differential operators that lower or raise the degree of the polynomial and raise/lower the parameters $\alpha$ and $\beta$. \[lem35\] For the polynomial $p_n(x;\alpha,\beta)$ given in one has for $\alpha,\beta>-1$ the lowering property $$\label{pnabrlow} p_n'(x;\alpha,\beta) = np_{n-1}(x;\alpha+1,\beta+1) ,$$ and for $\alpha,\beta > r-1$ the raising property $$\label{pnabrrais} \bigl( x^\beta(1-x^r)^\alpha p_n(x;\alpha,\beta) \bigr)' = x^{\beta-1}(1-x^r)^{\alpha-1} \sum_{k=1}^r a_{k,n}^{(\alpha,\beta)} x^{r-k} p_{n+k}(x;\alpha-k,\beta-k),$$ where $$a_{k,n}^{(\alpha,\beta)} = (-1)^k \left[ \binom{r}{k} (r\alpha+\beta) + \binom{r+1}{k+1} kn \right] .$$ The lowering property can easily be proved by differentiating the expression . To prove we first observe that $$\label{raisPi} \bigl[ x^\beta(1-x^r)^\alpha p_n(x;\alpha,\beta) \bigr]' = x^{\beta-1}(1-x^r)^{\alpha-1} \pi_{n+r}(x),$$ where $\pi_{n+r}$ is a polynomial of degree $n+r$ given by $$\label{Pinr} \pi_{n+r}(x) = \bigl( \beta(1-x^r)-\alpha r x^r \bigr) p_n(x;\alpha,\beta) + x(1-x^r) p_n'(x;\alpha,\beta).$$ Integrating by parts shows that for $\alpha,\beta >0$ $$\begin{aligned} \int_0^1 x^{\beta-r}(1-x^r)^{\alpha-1} \pi_{n+r}(x) x^{r(j+1)-1}\, dx &=& \int_0^1 \bigl[ x^\beta(1-x^r)^\alpha p_n(x;\alpha,\beta) \bigr]' x^{rj}\, dx \\ &=& - rj \int_0^1 x^\beta (1-x^r)^\alpha p_n(x;\alpha,\beta) x^{rj-1}\, dx ,\end{aligned}$$ and by this is zero for $1 \leq j \leq n$ but also for $j=0$. So $\pi_{n+r}$ is a polynomial of degree $n+r$ which is orthogonal to all $x^{r\ell-1}$ for $1\leq \ell \leq n+1$ with weight $x^{\beta-r}(1-x^r)^{\alpha-1}$ on the interval $[0,1]$. These are $n+1$ orthogonality conditions. For $\alpha,\beta > r-1$ the $r$ polynomials $x^{r-k}p_{n+k}(x;\alpha-k,\beta-k)$, $1 \leq k \leq r$, have the same orthogonality conditions, they are all of degree $n+r$ and they are linearly independent, hence they span the linear space of polynomials of degree $n+r$ with the $n+1$ orthogonality conditions. Therefore $$\label{pinrsum} \pi_{n+r}(x) = \sum_{k=1}^r a_{k,n}^{(\alpha,\beta)} x^{r-k} p_{n+k}(x;\alpha-k,\beta-k),$$ for some $a_{k,n}^{(\alpha,\beta)}$, $k=1,2,\ldots,r$. To find these coefficients, one compares the coefficients of $x^{r-k}$ in the latter expansion and . With these operators one can find the differential equation. \[thm36\] For any $n \in \mathbb{N}$, $r\geq 1$ and $\alpha,\beta >-1$ the polynomial $y=p_n(x;\alpha,\beta)$ satisfies the differential equation $$\label{diffr} x(1-x^r)y^{(r+1)} + (r+\beta) y^{(r)} +\sum_{k=0}^{r} c_{k,n}^{(\alpha,\beta)} x^k y^{(k)} = 0,$$ where the coefficients $c_{k,n}^{(\alpha,\beta)}$ for $0 \leq k \leq r$ are given by $$\label{ckn} c_{k,n}^{(\alpha,\beta)} = (-1)^{r+k+1} (n-r+1)_{r-k} \left[ \binom{r}{k}(r\alpha+\beta)+ \binom{r+1}{k} (rn+r-kn) \right].$$ From the lowering operation one has $$p_n^{(r)}(x;\alpha,\beta) = \frac{n!}{(n-r)!}p_{n-r}(x;\alpha+r,\beta+r).$$ Multiplying both sides by $x^{\beta+r}(1-x^r)^{\alpha+r}$ and differentiating gives $$\begin{aligned} &x^{\beta+r}(1-x^r)^{\alpha+r} p_n^{(r+1)}(x;\alpha,\beta) \\ &+\ x^{\beta+r-1}(1-x^r)^{\alpha+r-1} \bigl[\beta+r)(1-x^r)- r(\alpha+r)x^r \bigr] p_{n}^{(r)}(x;\alpha,\beta) \\ = &\frac{n!}{(n-r)!}x^{\beta+r-1}(1-x^r)^{\alpha+r-1} \sum_{k=1}^r a_{k,n-r}^{(\alpha+r,\beta+r)}x^{r-k}p_{n-r+k}(x;\alpha+r-k,\beta+r-k),\end{aligned}$$ where we used the raising operation for the right hand side. Using the lowering operation one has $$p_{n-r+k}(x;\alpha+r-k;\beta+r-k) = \frac{(n-r+k)!}{n!} p_{n}^{(r-k)}(x;\alpha,\beta),$$ hence we have $$\begin{gathered} x(1-x^r)p_n^{(r+1)}(x;\alpha,\beta) + [(\beta+r)(1-x^r)-r(\alpha+r)x^r]p_n^{(r)}(x;\alpha,\beta) \\ - \sum_{k=1}^ra_{k,n-r}^{(\alpha+r,\beta+r)} \frac{(n-r+k)!}{(n-r)!} x^{r-k} p_n^{(r-k)}(x;\alpha,\beta)=0,\end{gathered}$$ or $$\begin{aligned} x(1-x^r)p_n^{(r+1)}(x;\alpha,\beta) + (\beta+r)p_n^{(r)}(x;\alpha,\beta) + \sum_{k=0}^rc_{k,n}^{(\alpha,\beta)} x^{k} p_n^{(k)}(x;\alpha,\beta)=0,\end{aligned}$$ where $$c_{k,n}^{(\alpha,\beta)} = - a_{r-k,n-r}^{(\alpha+r,\beta+r)} \frac{(n-k)!}{(n-r)!} ,$$ which gives with . Asymptotic zero behavior ------------------------ We now investigate the asymptotic distribution of the zeros of the type I Jacobi-Angelesco polynomials $A_{\vec{n},j}^{(\alpha,\beta)}$, $1 \leq j \leq r$, for the multi-index $\vec{n}=(n,n,\ldots,n)$ and $n\to \infty$. From Theorem \[thm31\] it is clear that the zeros of $A_{\vec{n},j}^{(\alpha,\beta)}$ are copies of the zeros of $p_n(x;\alpha,\beta)$ (which are on $[0,1]$, see Lemma \[lem37\]), but rotated to the interval $[0,\omega^{j-1}]$. Hence we only need to investigate the asymptotic distribution of the zeros of $p_n(x;\alpha,\beta)$ given in . First we prove that the zeros of $p_n(x;\alpha,\beta)$ are all on $(0,1)$ whenever $\alpha,\beta >-1$. For this we modify the standard proof of the location of zeros of orthogonal polynomials (see, e.g., [@Ismail Thm. 2.2.5]). \[lem37\] Let $\alpha,\beta >-1$, then all the zeros of $p_n(x;\alpha,\beta)$ given in are simple and lie in the open interval $(0,1)$. Suppose $x_1,\ldots,x_m$ are the zeros of odd multiplicity of $p_n(x;\alpha,\beta)$ that lie in $(0,1)$ and that $m < n$. Then consider the polynomial $q_m(x) = (x^r-x_1^r)(x^r-x_2^r)\cdots(x^r-x_m^r)$. This is a polynomial of degree $rm$ with $m$ real zeros on $(0,1)$ at the points $x_1,\ldots,x_m$ and no other sign changes on $(0,1)$. Hence $p_n(x;\alpha,\beta)q_m(x)$ has constant sign on $(0,1)$ so that $$\int_0^1 p_n(x;\alpha,\beta) q_m(x) x^{\beta+r-1}(1-x^r)^\alpha\, dx \neq 0.$$ But $q_m(x)x^{r-1}$ contains only powers $x^{rj-1}$ with $1 \leq j \leq m+1 \leq n$, hence by this integral is zero. This contradiction implies that $m \geq n$ and since $p_n(x;\alpha,\beta)$ can have at most $n$ zeros on $(0,1)$, we see that $m=n$. Denote the zeros of $p_n(x;\alpha,\beta)$ by $0 < x_{1,n} < x_{2,n} < \cdots < x_{n,n} < 1$. As before we will use the normalized zero counting measure $$\mu_n = \frac{1}{n} \sum_{j=1}^n \delta_{x_{j,n}}.$$ Then $(\mu_n)_n$ is a sequence of probability measures on $[0,1]$, and by Helley’s selection principle it will contain a subsequence $(\mu_{n_k})_k$ that converges weakly to a probability measure $\mu$ on $[0,1]$. If this limit is independent of the subsequence, then we call it the asymptotic zero distribution of the zeros of $p_n(x;\alpha,\beta)$. We will investigate this by means of the Stieltjes transform $$S_n(z) = \int_0^1 \frac{d\mu_n(x)}{z-x} = \frac{1}{n} \frac{p_n'(z;\alpha,\beta)}{p_n(z;\alpha,\beta)}, \quad S(z) = \int_0^1 \frac{d\mu(x)}{z-x} , \qquad z \in \mathbb{C} \setminus [0,1],$$ and use the Grommer-Hamburger theorem which says that $\mu_n$ converges weakly to $\mu$ if and only if $S_n$ converges uniformly on compact subsets of $\mathbb{C} \setminus [0,1]$ to $S$ and $zS(z) \to 1$ as $z \to \infty$ (see, e.g., [@Geronimo]). First we show that the weak limit of $(\mu_{n_k})_k$ has a Stieltjes transform $S$ which satisfies an algebraic equation of order $r+1$. \[prop32\] Suppose $\mu_{n_k}$ converges weakly to $\mu$, then the Stieltjes transform $S$ of $\mu$ satisfies $$\label{Salg} z(1-z^r) S^{r+1} + \sum_{\ell=0}^{r-1} (-1)^{r+\ell+1} \binom{r+1}{\ell} (r-\ell)z^\ell S^\ell = 0.$$ We first show that one can express the derivatives $p_n^{(j)}$ of $p_n(z;\alpha,\beta)$ in terms of $S_n$ and its derivatives: $$\label{jderpn} p_n^{(j)}(z;\alpha,\beta) = n^j p_n(z;\alpha,\beta) \bigr[ S_n^j(z) + \frac{1}{n} G_{n,j}(S_n,S_n',\ldots,S_n^{(j-1)}) \bigr], \qquad j \geq 0,$$ where $G_{n,j}$ is a polynomial in $j$ variables with coefficients of order $\mathcal{O}(1)$ in $n$. This polynomial is given recursively by $$\begin{gathered} G_{n,j}(x_1,\ldots,x_j) = x_1 G_{n,j-1}(x_1,x_2,\ldots,x_{j-1}) + (j-1)x_1^{j-2}x_2 \\ + \frac{1}{n} \sum_{k=1}^{j-1} \frac{\partial}{\partial x_k} G_{n,j-1}(x_1,\ldots,x_{j-1}) x_{k+1}, \end{gathered}$$ with $G_{n,0}=0$. This can be proved by induction on $j$. It is obvious for $j=0$ and for $j=1$ it follows from $G_{n,1}(x)=0$ and $$S_n(z) = \frac{1}{n} \frac{p_n'(z;\alpha,\beta)}{p_n(z;\alpha,\beta)}.$$ Suppose now it holds for $j$, then for $j+1$ we have $$\begin{gathered} p_n^{(j+1)}(z;\alpha,\beta) = \bigl( p_n^{(j)}(z;\alpha,\beta) \bigr)' = n^j p_n'(z;\alpha,\beta) \bigr[ S_n^j(z) + \frac{1}{n} G_{n,j}(S_n,S_n',\ldots,S_n^{(j-1)}) \bigr] \\ + n^j p_n(z;\alpha,\beta) \bigr[ j S_n' S_n^{j-1} + \frac{1}{n} \frac{d}{dz} G_{n,j}(S_n,S_n',\ldots,S_n^{(j-1)}) \bigr]. \end{gathered}$$ Here we can use $p_n'(z;\alpha,\beta) = n p_n(z;\alpha,\beta) S_n(z)$ and the chain rule $$\frac{d}{dz} G_{n,j}(S_n,S_n',\ldots,S_n^{(j-1)}) = \sum_{k=1}^j S_n^{(k)} \frac{\partial}{\partial x_k} G_{n,j}(S_n,S_n',\ldots,S_n^{(j-1)}),$$ which after collecting terms gives the desired formula for $j+1$. Now insert the expressions into the differential equation to find $$\begin{aligned} 0 &=& z(1-z^r) n^{r+1} p_n(z) \bigl[ S_n^{r+1} + \frac{1}{n} G_{n,r+1}(S_n,\ldots,S_n^{(r)}) \bigr] \\ & & +\ (r + \beta)n^r p_n(z) \bigl[ S_n^r + \frac{1}{n} G_{n,r} (S_n,\ldots,S_n^{(r-1)}) \bigr] \\ & & + \sum_{\ell=0}^r c_\ell(n) z^\ell n^\ell p_n(z) \bigl[ S_n^\ell + \frac{1}{n} G_{n,\ell}(S_n,\ldots,S_n^{(\ell-1)}) \bigr]. \end{aligned}$$ Divide by $n^{r+1} p_n(z)$ and let $n=n_k \to \infty$. From the convergence of $S_{n_k}$ to $S$ uniformly on compact subsets of $\mathbb{C} \setminus [0,1]$, it also follows that all the derivatives $S_{n_k}^{(j)}$ converge to $S^{(j)}$ uniformly on compact subsets of $\mathbb{C} \setminus [0,1]$. Furthermore, from $$\lim_{n \to \infty} \frac{c_\ell(n)}{n^{r-\ell+1}} = (-1)^{r+\ell+1} \binom{r+1}{\ell} (r-\ell),$$ so that in the limit we find the equation . Now observe that the equation does not depend on the subsequence $(n_k)_k$ anymore, so that every convergent subsequence has a limit $S$ satisfying equation . By using the binomial theorem, one can find $$\sum_{\ell=0}^{r+1} (-1)^{r+\ell+1} \binom{r+1}{\ell} (r-\ell)z^\ell S^\ell = -(zS+r) (zS-1)^r,$$ so that the algebraic equation simplifies to $$\label{Salg1} zS^{r+1} - (zS+r)(zS-1)^r = 0.$$ This equation has $r+1$ solutions but we are interested in the solution which is a Stieltjes transform of a probability measure on $[0,1]$. By using the Stieltjes-Perron inversion formula one can then find the asymptotic zero distribution measure $\mu$. \[thm39\] The asymptotic zero distribution of the polynomial $p_n(x;\alpha,\beta)$ as $n \to \infty$ is given by a measure which is absolutely continuous on $[0,1]$ with a density $u_r$ given by $u_r(x) = r x^{r-1} w_r(x^r)$, where $w_r$ is given by $$w_r(\hat{x}) = \frac{r+1}{\pi} \frac{1}{|\hat{x}'(\theta)|} =\frac{r+1}{\pi \hat{x}} \frac{\sin \theta \sin r\theta \sin (r+1)\theta}{|(r+1)\sin r\theta - re^{i\theta} \sin(r+1)\theta|^2},$$ where we used the change of variables $$\label{hatx} \hat{x} = x^r = \frac{1}{c_r} \frac{\bigl( \sin(r+1)\theta \bigr)^{r+1}}{\sin \theta (\sin r\theta)^r}, \qquad 0 < \theta < \frac{\pi}{r+1},$$ and $c_r=(r+1)^{r+1}/r^r$. The proof is along the same lines as in [@Neuschel] where the asymptotic zero distribution was obtained for Jacobi-Piñeiro polynomials. The algebraic equation can be transformed by taking $$W = \frac{zS}{zS-1}, \quad zS = \frac{W}{W-1},$$ which gives $$\label{Walg} W^{r+1} - (r+1) z^r W + rz^r = 0.$$ We look for a solution $W$ of the form $\rho e^{i\theta}$, where $\theta$ is real and $\rho >0$. Take $z=x \in [0,1]$ and insert $W=\rho e^{i\theta}$ into to find $$\rho^{r+1} e^{i(r+1)\theta} - (r+1) x^r \rho e^{i\theta} + rx^r = 0.$$ Hence the real and imaginary parts satisfy $$\begin{aligned} \rho^{r+1} \cos(r+1)\theta - (r+1) x^r \rho \cos \theta + rx^r & = & 0, \label{Wreal} \\ \rho^{r+1} \sin(r+1)\theta - (r+1) x^r \rho \sin \theta & = & 0. \label{Wimag}\end{aligned}$$ From we find $$\hat{x} := x^r = \frac{\rho^r \sin(r+1)\theta}{(r+1)\sin \theta},$$ and using this in , we find $$\rho(\hat{x}) = \frac{r}{r+1} \frac{\sin(r+1)\theta}{\sin r\theta} .$$ Combining both gives . Note that when $\theta \in (0,\pi/(r+1))$ one has $\rho > 0$. So for $\hat{x} \in [0,1]$ the equation has a solution of the form $\rho e^{i\theta}$. Observe that also $\rho e^{-i\theta}$ is a solution and in fact $$W_+(\hat{x}) = \lim_{\epsilon\to 0+} W(\hat{x}+i\epsilon) = \rho(\hat{x})e^{i\theta}, \quad W_-(\hat{x}) = \lim_{\epsilon\to 0+} W(\hat{x}-i\epsilon) = \rho(\hat{x})e^{-i\theta},$$ are the boundary values of the function $W$ which is analytic on $\mathbb{C} \setminus [0,1]$. From the Stieltjes-Perron inversion formula (or the Sokhotsky-Plemelj formula) we can compute the density $u_r$ as $$u_r(x) = \frac{1}{2\pi i} \bigl( S_-(x)-S_+(x) \bigr) = \frac{\rho}{\pi x} \frac{\sin \theta}{|\rho e^{i\theta}-1|^2} .$$ Writing everything in terms of $\hat{x}$ gives the required result. ![The density $u_r$ of the asymptotic zero distribution on $[0,1]$ for $r=1$ (solid), $r=2$ (dash), $r=3$ (dash-dot), $r=4$ (long dash) and $r=5$ (dots).[]{data-label="fig:asympzero"}](Asymp_zero.eps){width="4in"} We have plotted the densities $u_r$ on $[0,1]$ for $r=1,2,3,4,5$ in Figure \[fig:asympzero\]. The case $r=1$ corresponds to the density $$u_1(x) = \frac{1}{\pi} \frac{1}{\sqrt{x(1-x)}}, \qquad 0 < x < 1 ,$$ which is the well known arcsine distribution which is symmetric around $x=1/2$. For $r>1$ the symmetry is gone. The case $r=2$ corresponds to the density given in . When $x$ tends to the endpoints one has the behavior $$\begin{aligned} u_r(x) &\sim x^{-\frac{1}{r+1}}, \qquad x \to 0, \\ u_r(x) &\sim (1-x^r)^{-\frac12} , \qquad x \to 1, \end{aligned}$$ so that for fixed $r$ the density $u_r$ has a singulartity of order $\frac{1}{r+1}$ at the endpoint $0$ and a singularity of order $\frac12$ at the endpoint $1$. So for $r>1$ the zeros are more dense near the endpoint $1$ than near the point $0$. This is typical for an Angelesco systems where the zeros on all the other intervals $[0,\omega^{j-1}]$, $j=2,\ldots,r$ push the zeros on $[0,1]$ to the right. The asymptotic behavior of the zeros is similar to the asymptotic behavior of the zeros of Jacobi-Pineiro polynomials, which was studied by Neuschel and Van Assche [@Neuschel] and differs only by the change of variables $c_ry = x^r$. [99]{} A. Angelesco, *Sur deux extensions des fractions continues algébriques*, Comptes Rendus Acad. Sci. Paris **168** (1919), 262–265. A.I. Aptekarev, *Multiple orthogonal polynomials*, J. Comput. Appl. Math. **99** (1998), no. 1–2, 423–447. P. Billingsley, *Probability and Measure*, John Wiley & Sons, New York, 1979. E. Coussement, J. Coussement, W. Van Assche, *Asymptotic zero distribution for a class of multiple orthogonal polynomials*, Trans. Amer. Math. Soc. **360** (2008), 5571–5588. E.J.C. Dos Santos, *Monotonicity of zeros of Jacobi-Angelesco polynomials* Proc. Amer. Math. Soc. **145** (2017), no. 11, 4741–4750. J.S. Geronimo, T.P. Hill, *Necessary and sufficient condition that the limit of Stieltjes transforms is a Stieltjes transform*, J. Approx. Theory **121** (2003), no. 1, 54–60. J.S. Geronimo, P. Iliev, W. Van Assche, *Alpert multiwavelets and Legendre-Angelesco multiple orthogonal polynomials*, SIAM J. Math. Anal. **49** (2017), no. 1, 626–645. M.E.H. Ismail, *Classical and Quantum Orthogonal Polynomials in One Variable*, Encyclopedia of Mathematics and its Applications **98**, Cambridge University Press, 2005 (paperback edition 2009). V.A. Kalyagin, *A class of polynomials determined by two orthogonality relations*, Mat. Sb. (N.S.) **110** (152) (1979), no. 4, 609–627 (in Russian); translated in Math. USSR Sbornik **38** (1981), no. 4, 563–580. V. Kaliaguine, A. Ronveaux, *On a system of “classical” polynomials of simultaneous orthogonality*, J. Comput. Appl. Math. **67** (1996), no. 2, 207–217. T. Neuschel, W. Van Assche, *Asymptotic zero distribution of Jacobi-Piñeiro and multiple Laguerre polynomials*, J. Approx. Theory **250** (2016), 114–132. E.M. Nikishin, *A system of Markov functions*, Vestnik Moskov. Univ. Ser. I Mat. Mekh. **979**, no. 4, 60–63; translated in Moscow Univ. Math. Bull. **34** (1979). E.M. Nikishin, V.N. Sorokin, *Rational Approximations and Orthogonality*, Translations of Mathematical Monographs, vol. 92, Amer. Math. Soc., Providence, RI, 1991. F.W.J. Olver, D.W. Lozier, R.F. Boisvert, C.W. Clark (eds.), *NIST Handbook of Mathematical Functions*, NIST and Cambridge University Press, 2010. W. Van Assche, *Nearest neighbor recurrence relations for multiple orthogonal polynomials*, J. Approx. Theory **163** (2011), 1427–1448. W. Van Assche, E. Coussement, *Some classical multiple orthogonal polynomials*, Numerical Analysis 2000, vol. V, Quadrature and Orthogonal Polynomials, J. Compute. Appl. Math. **127** (2001), 317–347. [^1]: Supported by FWO research project G.0864.16N and EOS project PRIMA 30889451. [^2]: This is in fact Aurel Angelescu, a Romanian mathematician who wrote a PhD thesis in 1916 under supervision of Paul Appell at the Sorbonne in Paris.

ArXiv

--- abstract: 'Given an ensemble of mixed qubit states, it is possible to increase the purity of the constituent states using a procedure known as state purification. The reverse operation, which we refer to as dilution, produces a larger ensemble, while reducing the purity level of the systems. In this paper we find asymptotically optimal procedures for purification and dilution of an ensemble of independently and identically distributed mixed qubit states, for some given input and output purities and an asymptotic output rate. Our solution involves using the statistical tool of local asymptotic normality, which recasts the qubit problem in terms of attenuation and amplification of a single-mode displaced Gaussian state. Therefore, to obtain the qubit solutions, we must first solve the analogous problems in the Gaussian setup. We provide full solutions to all of the above, for the (global) trace-norm figure of merit.' author: - Peter Bowles - Mădălin Guţă - Gerardo Adesso date: 'March 21, 2011' title: Asymptotically optimal purification and dilution of mixed qubit and Gaussian states --- [^1] Introduction ============ When implementing any quantum information protocol, the states we wish to employ and manipulate are inevitably affected by decoherence effects, which diminish their purity and consequently their resource power. There exist several well-established methods to protect against such undesirable factors: strengthening the entanglement resource using distillation methods [@BBP] or employing a quantum error correction scheme [@S] to encode our ‘fragile’ states into some larger, more unyielding system. The method we study in this paper is that of state [*purification*]{} [@CEM; @KW], a procedure which takes as input an ensemble of identical copies of an arbitrary (unknown) state and produces as output a smaller ensemble of identical states with higher purity. This can be seen as a special case of the more general problem of inverting the effect of a noisy channel on ensembles of states, the channel being the depolarising one in the present study. There already exists several theoretical results for purification of $n$ i.i.d. (independently and identically distributed) mixed qubits, notably Refs. [@CEM; @KW], where optimal purification algorithms for various formulations of the purification problem are provided. Purification of an ensemble of mixed qubit states has also been found to occur in the context of ‘superbroadcasting’ [@DMP], an $n \rightarrow m$ cloning procedure which can actually result in purified clones for $n\geq4$ and sufficiently mixed input states (the noise present is merely shifted from local states into correlations between output states). For $n\geq m$, superbroadcasting is actually equivalent to the optimal purification procedure of [@CEM]. Experimentally, purification has been achieved in [@RDC], which implemented the methodology of [@CEM] and demonstrated optimal purification for the case $n=2$. Beyond the [*entanglementology*]{} (phenomenology of entanglement), judging the performance of a purification protocol requires a figure of merit (FoM) which measures the departure from the ideal transformation. Two types of FoM have been considered in the literature, with very different results. The [*local*]{} FoM is built upon the comparison of the [*reduced*]{} states of individual output systems with the target state. In this case, a complete reversal of the depolarising channel may be obtained asymptotically with the size of the input ensemble, and with arbitrarily high output rate $m/n$ [@KW]. The [*global*]{} FoM compares the [*joint*]{} state of the output with that of a product of independent target states. This is a more demanding criterion. For example if the output systems are independent and identically prepared then the global fidelity scales as $F(n,m)= F_n^m$ where $F_n<1$ is the fidelity of an individual output state with respect to the target state. Indeed, it has been shown [@KW] that no protocol can achieve asymptotic purification $\big(F(n,m) \to 1 \big)$ to [*pure*]{} target states at a finite rate $m/n$. The global figure or merit is relevant whenever we deal with the collective state of the output rather than the individual constituents, as in the case of state transfer between atomic ensembles and light. Additionally, it can serve as a “measure of correlations” when the individual constituents of the output states are known to be exactly in the target state, as in superbroadcasting. This hypothesis will however not be pursued in this paper. The above no-go theorem motivates us to consider the question whether the depolarising channel can be reversed with a positive asymptotic output rate, when the target states (i.e. the states prior to applying the depolarising channel) are [*mixed*]{}. We show that this is indeed possible, and compute the maximal purification rate for given input and target purity, and the optimal FoM for approximate purification at a fixed rate which is higher than the maximal one. We also consider the opposite process of [*dilution*]{} in which, starting from an ensemble of $n$ identically prepared states, we produce a [*larger*]{} ensemble consisting of $m$ independent, but more mixed states. Dilution shares similarities with the process of optimal $n \rightarrow m$ quantum cloning [@OQC], but while in cloning the rate $m/n$ is fixed, and one aims at generating clones as close as possible to the input states (with respect to a local or a global FoM), in a dilution procedure we set a target level of output purity and look for the optimal rate for generating such target states. qubit problem Gaussian problem ------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- state model $\begin{array}{c}\mbox{ensemble of $n$ i.i.d mixed qubits } \\ \rho_{\textbf{r}}^{\otimes n} \mbox{ [Eq.~(\ref{QubIN})]} \\ \mbox{$n \gg 1$: number of copies} \\ \mbox{$\textbf{r}$ with $\|\textbf{r}\| \le 1$: Bloch vector;} \end{array}$ $\begin{array}{c}\mbox{single-mode displaced Gaussian state} \\ \Phi_\alpha^s \mbox{ [Eq.~(\ref{GaussIN})]} \\ \mbox{$\alpha \in \mathbb{C}$: displacement;} \\ \mbox{$s \in (0,1)$: purity parameter} \end{array}$ input $\rho_{\textbf{r}_0+\textbf{u}/\sqrt{n}}^{\otimes n}$ $\Phi_\alpha^{s_1}$ target $\rho_{\lambda \textbf{r}_0+k\textbf{u}/\sqrt{m}}^{\otimes m}$ $\Phi_{k \alpha}^{s_2}$ procedure $\begin{array}{c}\mbox{purification} \\ \lambda>1, \ m<n\end{array}$ $\begin{array}{c}\mbox{attenuation} \\ s_2 < s_1, \ k<1\end{array}$ procedure $\begin{array}{c}\mbox{dilution} \\ \lambda < 1, \ m>n \end{array}$ $\begin{array}{c}\mbox{amplification} \\ s_2 > s_1, \ k>1\end{array}$ In deriving the asymptotic results, the key mathematical tool is that of local asymptotic normality (LAN), a fundamental ‘classical’ statistics technique [@LC] which was recently extended to the context of quantum statistical models [@GK; @GJK; @GK2; @GJ]. In the quantum case, LAN dictates that the collective state of $n$ i.i.d. quantum systems, can be approximated by a joint Gaussian state of a classical and a quantum continuous variable (CV) systems. This has been used to derive asymptotically optimal state estimation strategies for mixed states of arbitrary finite dimension [@GK2], and also in finding quantum teleportation benchmarks [@GBA] and optimal quantum learning procedures [@GKot] for multiple qubit states. The general strategy is to recast statistical problems involving $n$ i.i.d. quantum systems into the simpler setting of Gaussian states. The optimal solution for the corresponding Gaussian problems can then be used to construct asymptotically optimal procedures for the original one. In section \[sec3\] we sketch how this could be physically implemented, and more details can be found in [@GJK]. Following this methodology, we transform the qubit purification and dilation problems into those of optimal [*attenuation*]{} and [*amplification*]{} for a one-mode CV system in a Gaussian state, together with a classical real-valued Gaussian variable, both with known variance but unknown means. In attenuation we reduce the variance of a displaced Gaussian state, at the price of simultaneously reducing its amplitude, while in amplification we increase the amplitude, as well as the variance. For both problems we use a FoM based on maximum trace-norm distance, and show that the optimal attenuation channel is obtained by applying a beamsplitter, while the optimal amplification is implemented by a non-degenerate parametric amplifier. A similar scheme for the attenuation of Gaussian CV states has been proposed and experimentally implemented in [@AFF]. Parametric amplification has been investigated in [@HM; @CC; @CDG], and demonstrated experimentally in [@OUP]. In particular, the same amplifier is optimal for a FoM based on the minimum amount of added noise [@HM; @CC]. However, whilst these transformations are well known candidates for our protocols, to the best of our knowledge a proof of their optimality with respect to the FoM chosen in this paper had not been obtained in the literature. Our proof relies on a covariant channels optimisation technique developed in [@GM; @GBA]. We find that for given input and output purity parameters, there exists a range of values for the ratio $k$ between output and input displacement, such that attenuation or amplification can be realised perfectly, and we compute the maximal (optimal) value $k_0$, as a function the two purities. In the parameter range where the procedures cannot be accomplished perfectly, we give the exact expression for the optimal FoM. A schematic summary of the problems addressed in this paper is provided in Table \[table\]. The paper is organised as follows. In Section  \[sec2\] we formulate and solve the two quantum Gaussian problems, and the corresponding classical one. In Section \[sec3\] we use this result in conjunction with LAN to find asymptotically optimal purification and amplification channels for states of $n$ i.i.d. mixed qubits. We draw our concluding remarks in Section \[sec4\]. The proofs are collected in Appendix A. Optimal attenuation and amplification of Gaussian states {#sec2} ======================================================== Classical Case {#sec2a} -------------- Before we move onto the quantum case, it is instructive and relevant to consider the corresponding problems for classical random variables. In the classical scenario, the analogue of ‘attenuation’ (‘amplification’) is a procedure which reduces (increases) the mean and variance of a given random variable. The analogue to our quantum problem would then be to find a transformation $K$ which maps a real-valued normally distributed random variable $X\sim N(u,V_1)$ of arbitrary mean $u$ and fixed variance $V_1$, into a variable $Y\sim N(ku, V_2)$ such that the risk $$R_{\rm max}(K; V_{1}, V_{2}, k) =2 \sup_u\|K\big(N(u, V_1)\big)-N(ku,V_2)\|_{\text{tv}} \label{cfom}$$ is minimised. Here $k$ represents a fixed constant, where $0<k<1$ means attenuation and $k>1$ means amplification of the Gaussian variable $X$, and we choose the interesting case where $V_1>V_2$ in the case of attenuation, and $V_1<V_2$ for amplification. The notation $\|\mathbb{P}-\mathbb{Q}\|_{\text{tv}} = \sup\{|\mathbb{P}(A)-\mathbb{Q}(A)|:A\in\mathcal{F}\}$, for the $\sigma$-algebra $\mathcal{F}$, represents the total variation distance between the probability distributions $\mathbb{P}$ and $\mathbb{Q}$ which reduces to one-half of the $L_{1}$-distance between their probability densities in the case of mutually absolutely continuous distributions [@Torgersen]. The solutions of both classical and quantum versions of this problem rely on the notion of ‘covariance’. Consider the transformation $$X\mapsto K(X)=kX+Z \label{ccov}$$ where $X$ and $Z$ are independent random variables, $Z$ having a fixed variance and vanishing mean. Such a (classical) channel is covariant, in the sense that $$K(X+C) = K(X) +kC$$ for any constant $C$. Such transformations can be shown to not only minimise (\[cfom\]), but also to render it independent of expectation so that the FoM becomes $${R}_{\rm max}(K; V_{1},V_{2},k)=2 \|K\big(N(0, V_1)\big)-N(0,V_2)\|_{\text{tv}}. $$ It is easy to see that if $$k\leq k^{(c)}_{0} (V_{1}, V_{2}):= \sqrt{\frac{V_2}{V_1}} \label{cr}$$ then the target distribution can be achieved exactly, with the appropriate amount of Gaussian noise in the variable $Z$. As we shall see in the next section, there exists an analogous range (\[koeqn\]) for the quantum Gaussian transformation. As for the case $k> k_{0}^{(c)} (V_{1}, V_{2})$, it can be shown [@Torgersen] that the optimal choice for $Z$ in is $Z=0$, as one would expect, so that the optimal figure of merit is $$\begin{aligned} &&R_{\rm minmax}(V_{1},V_{2},k):= \inf_{K}R_{\rm max}(K;V_{1},V_{2},k) \nonumber\\ &&=\int{\bigg|\frac{1}{\sqrt{2\pi k^2 V_1}}e^{-\frac{x^2}{2k^2V_1}}-\frac{1}{\sqrt{2\pi V_2}}e^{-\frac{x^2}{2V_2}}\bigg|dx}. \label{clrisk}\end{aligned}$$ Henceforth, we will denote by $K^*$ the optimal transformation for the two cases discussed above. ![(Color online) Schematic phase-space diagram for (a) attenuation and (b) amplification of a displaced Gaussian state $\Phi^{s_1}_\alpha$. []{data-label="figps"}](phasespace2.pdf){width="8.5cm"} Quantum Case {#sec2b} ------------ In this Section we consider the following: given a Gaussian state $\Phi_\alpha$ of a one-mode CV quantum system, with known covariance and unknown displacement $\alpha$, we would like to optimally attenuate (amplify) it, that is transform it into a state with smaller (greater) covariance and displacement $k\alpha$, with the largest possible proportionality constant $k$. Let $$W_\alpha := \exp(\alpha a^\dagger-\bar{\alpha}a)$$ denote the Weyl operators where $\alpha\in\mathbb{C}$ and $a$, $a^\dagger$ the creation and annihilation operators satisfying $[a,a^\dagger]=\mathbf{1}$ and $$a|n\rangle=\sqrt{n}|n-1\rangle, \qquad n\geq 0,$$ where $\{|n\rangle\}_{n\geq0}$ is the Fock basis of the Hilbert space $\mathcal{H}$. For $0<s<1$ we denote by $\Phi^s$ the centred, phase invariant Gaussian state $$\label{eq.thermal} \Phi^s = (1-s)\sum^\infty_{n=0} s^n|n\rangle\langle n|,$$ and by displacing it we obtain the family of Gaussian states $$\Phi^s_\alpha := W_\alpha \Phi^s W^\dagger_\alpha. \label{GaussIN}$$ Given two different mixing parameters $s_1>s_2$ $(s_1<s_2)$ and a positive parameter $k<1$ $(k>1)$ we would like to find the optimal attenuation (amplification) channel which maps the state $\Phi^{s_1}_\alpha$ close to the state $\Phi^{s_2}_{\alpha}$ for an arbitrary displacement $\alpha$ (see Fig. \[figps\]). For any channel $P:\mathcal{T}_1(\mathcal{H}) \rightarrow \mathcal{T}_1(\mathcal{H})$ we define the FoM called the maximum risk $$R_{\rm max}(P; s_1,s_2,k) = \sup_{\alpha\in\mathbb{C}}\|P(\Phi^{s_1}_\alpha)-\Phi^{s_2}_{k\alpha}\|_1$$ and the minimax risk $$R_{\rm minmax}(s_1,s_2,k) = \inf_P R_{\rm max}(P; s_1,s_2,k). \label{mmx}$$ A channel is called ‘minimax’ if its maximum risk is equal to the minimax risk. We will show that (up to a trivial adjustment for a certain range of $k$’s) the optimal solutions to the attenuation and amplification problems are, respectively, the beamsplitter and parametric amplifier. \ We start by defining a specific channel denoted in both cases $P^{\star}$, then show that it is optimal and compute the minimax risk. For $s_1>s_2$ and $k<1$, the attenuation channel is implemented by the action of a beamsplitter with reflectivity $k$ acting on an input mode $a$ prepared in a state $\Phi^{s_1}_\alpha$, and a second ancillary mode $b$ in the vacuum state $\omega=|0\rangle\langle 0|$. The output mode $c$ of the channel is $$c = k^2a + \sqrt{1-k^2}b. \label{bs}$$ For $s_2>s_1$ and $k>1$, the channel is a parametric amplifier, whose action is represented by the following transformation on the input mode $a$ and an ancillary mode $b$ prepared in the vacuum state: $$c=ka + \sqrt{k^2-1}b^\dagger . \label{am}$$ We note that for each pair $(s_1,s_2)$ there exists a range of parameters $k$ for which $R_{\rm minmax}(s_1,s_2,k)=0$, i.e., the procedures can be accomplished perfectly. Indeed it can be easily verified that, for $k$ given by $$k^{\rm att}_0(s_1, s_2) = \sqrt{\frac{s_2(1-s_1)}{s_1(1-s_2)}}, \quad k^{\rm amp}_0(s_1, s_2)= \sqrt{\frac{1-s_1}{1-s_2}}, \label{koeqn}$$ the channels and respectively produce exactly the target state $\Phi^{s_{2}}_{k\alpha}$. Moreover, if $k<k_{0}$ then the output of $P^{\star}$ is the state $\Phi^{s}_{k\alpha}$ with $s<s_{2}$, and the target can be still perfectly achieved by adding an appropriate amount of Gaussian noise. For later use, when $k<k_{0}$ we will denote by the same symbol $P^\star$ this modified channel. From now on we consider the less trivial situation $k\geq k_{0}$, corresponding to the regime where perfect amplification or attenuation are impossible. We then state the following theorem and lemma, whose proofs are given in Appendix A: \[theorem21\] If $k<k_0$ then the minimax risk for attenuation (amplification) is zero. If $k\geq k_0$, the minimax procedure is $P^\star$, i.e. the beamsplitter (\[bs\]) in the case of attenuation, or the parametric amplifier (\[am\]) in the case of amplification: $$R_{\rm max}(P^\star;s_{1},s_{2},k)= R_{\rm minmax}(s_{1},s_{2},k). \label{minmax}$$ \[lemma22\] If $k<k_0$, then the minimax risk for attenuation (amplification) is given by $$R_{\rm minmax}(s_{1},s_{2},k)= 2(\tilde{s}^{m_{0}+1} - s_{2}^{m_{0}+1}) , \label{Rmaxpur}$$ where $m_0$ is the integer part of $$\ln [ (1-\tilde{s})/(1-s_{2})] /\ln(s_{2}/\tilde{s}),$$ and $\tilde{s}$ takes the values $$\tilde{s}_{\rm att} = \frac{s_1 k^2}{1-s_1 + s_1 k^2}, ~{\rm and}~ \tilde{s}_{\rm amp}= 1-\frac{1-s_1}{k^2}. \label{tildes}$$ in the case of attenuation and respectively amplification. The risk for both processes is plotted in Fig. \[figpuramp\] \[(a)-(d)\]. In Figure \[figk0\] we plot $k_0$ for both processes as a function of the input and output purity parameters $s_1$ and $s_2$. Asymptotically optimal purification and dilution for ensembles of qubits {#sec3} ======================================================================== We turn now to the problem of finding optimal purification and dilution schemes for ensembles of identical qubits. We denote by $\rho_{\textbf{r}}$ the qubit state with Bloch vector $\textbf{r}=(r_{x},r_{y},r_{z})$ $$\label{QubIN} \rho_{\textbf{r}}=\frac{1}{2}(\mathbf{1}+\textbf{r}\boldsymbol{\sigma})\,,$$ where $\textbf{r}\boldsymbol{\sigma}=r_x\sigma_{x} +r_y\sigma_{y} +r_z\sigma_{z}$ and $\sigma_{i}$ are the Pauli matrices. We are given $n$ identical qubits prepared in the state $\rho_{\textbf{r}}$ and we would like to produce $m$ identical qubits in the state $\rho_{\lambda \textbf{r}}$ with $m$ as large as possible, for a fixed positive parameter $\lambda$. When $\lambda >1$, the aim is to “purify” the state, and when $0<\lambda <1$ we want to “dilute” the state with the benefit of obtaining more copies. Clearly, for purification the output state is physical only if $\lambda$ satisfies $\lambda\| \textbf{r}\|\leq 1$. This can be achieved by letting $\lambda$ depend on $\textbf{r}$, or by restricting to those input states which satisfy the property. To illustrate the latter, suppose we would like to reverse the action of the depolarising channel $$C :\rho_{\textbf{r}}\mapsto \rho_{\textbf{r}/\lambda},$$ then the input states of the purification channel automatically satisfy the requirement. As to the former, our asymptotic analysis will produce a [*local*]{} FoM which only depends on the value of $\lambda$ at a particular state, so for simplicity we will assume it to be constant. A purification (dilution) procedure is a quantum channel $$Q_{n}:M( \mathbb{C}^{2^{n}})\to M( \mathbb{C}^{2^{m}})$$ mapping $n$-qubit states to $m$-qubit states, and its performance is measured by the FoM (risk) $$R(Q_n;\textbf{r},\lambda):=\|Q_{n}(\rho_{\textbf{r}}^{\otimes n})- \rho^{\otimes m}_{\lambda\textbf{r}}\|_1.$$ Note that this is a global rather than a local risk, in the sense that it measures the distance between the output and the [*joint*]{} product state, instead of comparing their restrictions to each single system. Note also that the risk at a fixed point $\textbf{r}$ can always be made equal to zero by simply preparing the target state for that point. To take into account the [*overall*]{} performance of a procedure, one can either integrate the risk with respect to a prior distribution over states (Bayesian statistics) or take the maximum over all states (frequentist statistics). We adopt the latter viewpoint, and in addition we will consider a more refined version of the maximum risk called [*local maximum risk*]{} around $\textbf{r}_0$ $$R_{\rm max}(Q_n;\textbf{r}_0,\lambda):=\sup_{\|\textbf{r}-\textbf{r}_0\|\leq n^{-\frac{1}{2}+\epsilon}} R(Q_n;\textbf{r},\lambda).$$ In asymptotic statistics the local maximum risk is more informative that the ‘global’ one since it captures the behaviour of the procedure around any point in the parameter space, rather than that of the worst case. The radius of the ball over which we maximise is slightly larger than the precision of $n^{-1/2}$ with which we can estimate the state parameters, so that the definition of the local risk does not amount to assuming any prior information about the parameter. Indeed one can use a small sample $n^{1-\epsilon}\ll n$ of the input systems to obtain a rough estimate of the Bloch vector $\textbf{r}$ such that the obtained estimator $\textbf{r}_{0}$ will be in a ball of size $n^{-1/2+\epsilon}$ around $\textbf{r}$, with probability converging to one as $n\to \infty$. With this additional information, one can then apply the purification (dilution) channel to the remaining systems, with no loss in the asymptotic optimal risk (see below). The local maximum risk is a standard FoM in asymptotic statistics and it is has been used in quantum statistics in [@GK; @GJK; @GKot] to which we refer for more details, and for its relation to Bayesian methods. Up to this point the number of input and output systems $n$ and $m$ were fixed, with $n$ considered to be large. However, for a non-trivial asymptotic analysis, $m$ should be an increasing function of $n$, more precisely we consider the optimal purification (dilution) procedure for a fixed rate $$\Lambda= \lim_{n \rightarrow \infty} \frac{m(n)}{n} >0.$$ Indeed from our fixed rate analysis it can easily be deduced that in the case of a sub-linear dependence $m(n)=o(n)$, one can produce $m$ output copies of arbitrary purity with vanishing local maximum risk. On the other hand, by similar reasonings, one may expect that if $m(n)/n$ is unbounded, then the best strategy should be to estimate the state and reprepare $m$ independent copies of the estimator (‘measure and prepare’ strategy [@Bae]). We leave this statement as a conjecture, and from now on we will assume that the rate $\Lambda$ is given and fixed. For any sequence $\mathbf{Q}:= \{Q_{n}\}$ of procedures we define the [*asymptotic local maximum risk*]{} at $\textbf{r}_0$ by $$R(\mathbf{Q};\textbf{r}_0,\lambda,\Lambda):= \limsup_{n\rightarrow\infty}R_{\rm max}(Q_n;\textbf{r}_0,\lambda),$$ and we would like to find an optimal (minimax) strategy whose asymptotic risk is equal to the [*minimax risk*]{} $$R_{\rm minmax}(\textbf{r}_0,\lambda,\Lambda) := \limsup_{n\rightarrow\infty}\inf_{Q_n}R_{\rm max}(Q_n;\textbf{r}_0,\lambda). \label{qmmx}$$ In other words, we will answer the following question: for given purification (dilution) constant scale factor $\lambda$ and input-output rate $\Lambda$, what is the minimax risk $R_{\rm minmax}(\textbf{r}_0,\lambda,\Lambda)$ and which is the procedure that achieves it? In particular, we will find that the minimax risk is zero for a range of parameters $(\textbf{r}_0,\lambda,\Lambda)$, and we will identify the maximum value $\Lambda_0^{\rm pur}$ ($\Lambda_0^{\rm dil}$) for which the purification (dilution) can be performed with asymptotically vanishing risk. These rates are the qubit analogues of the constants $k_0$ defined in . The main technical tool is the theory of [*local asymptotic normality*]{} (LAN) developed in [@GK; @GJK; @GJ; @GK2] as an extension of a key concept from (classical) asymptotic statistics [@LC; @vanderVaart]. LAN means that the joint quantum state of identically prepared (finite-dimensional) systems can be approximated in a strong sense by a quantum-classical Gaussian state of fixed variance, whose mean encodes the information about the parameters of the original state. In this way, a number of asymptotic problems can be reformulated in terms of Gaussian states, for which the explicit solution can be found, e.g. state estimation [@GK3], teleportation benchmarks [@GBA], quantum learning [@GKot], system identification [@G]. For the purposes of this paper we give a brief description of LAN for mixed qubit states. Let $$\rho_{\textbf{r}_0+ \textbf{u}/ \sqrt{n}}= \frac{1}{2}\big(\textbf{1}+(\textbf{r}_0+\textbf{u}/\sqrt{n})\boldsymbol{\sigma}\big)$$ denote a qubit state in a the neighbourhood of a fixed and known state $\rho_{\textbf{r}_0}$, which is uniquely characterised by an [*unknown*]{} local parameter $\textbf{u}$. The family of $n$-qubit states $$\label{eq.local.qubit.model.} \mathcal{P}_{n}:= \left\{ \rho_{\textbf{u}}^{n}:= \rho_{\textbf{r}_0+ \textbf{u}/ \sqrt{n}}^{\otimes n} : \| \textbf{u}\| \leq n^{\epsilon}\right\}$$ will be called the local statistical model at $\textbf{r}_0$. Additionally, we define a classical-quantum Gaussian model $$\label{eq.gaussian.model} \mathcal{N}:=\left\{ N_{\textbf{u}}\otimes \Phi_{\textbf{u}} :\textbf{u}\in \mathbb{R}^{3}\right\}$$ where $N_{\textbf{u}} := N(u_{z}, 1-\|\textbf{r}_{0}\|^{2})$ is a normal (Gaussian) distribution on $\mathbb{R}$ with mean $u_{z}$ and variance $1-\|\textbf{r}_{0}\|^{2}$, and $$\Phi_{\textbf{u}} = W_{\alpha} \Phi^{s} W_{\alpha}^{\dagger}, \qquad \alpha= \frac{u_{x}+iu_{y}}{2\|\textbf{r}_{0}\|},\,\, s= \frac{1-\|\textbf{r}_{0}\|}{1+\|\textbf{r}_{0}\|},$$ is a displaced thermal Gaussian state of a one-mode CV system (cf. Section \[sec2b\]) with known covariance matrix characterised by the purity parameter $s$ (with zero squeezing) and unknown means proportional to $(u_{x},u_{y})$. Now, the mathematical statement of LAN [@GK] is that there exist two sequences of channels $\mathbf{T}= \{T_{n}\}$ and $\mathbf{S}=\{S_{n}\}$ with $$\begin{aligned} T_{n}&:& M(\mathbb{C}^{2^{n}})\to L^{1}(\mathbb{R})\otimes \mathcal{T}_{1}\\ S_{n}&:& L^{1}(\mathbb{R})\otimes \mathcal{T}_{1}\to M(\mathbb{C}^{2^{n}})\end{aligned}$$ such that $$\begin{aligned} && \lim_{n\to \infty }\sup_{\|{\bf u} \|\leq n^\epsilon }\| T_{n} (\rho_{\textbf{u}}^{n})- N_{\textbf{u}}\otimes \Phi_{\textbf{u}}\|_{1} =0,\\ && \lim_{n\to \infty} \sup_{\|{\bf u} \|\leq n^\epsilon }\| \rho_{\textbf{u}}^{n}- S_{n}(N_{\textbf{u}}\otimes \Phi_{\textbf{u}} ) \|_{1} =0.\end{aligned}$$ In the above formulas, $\mathcal{T}_{1}$ is associated to the trace-class operators of the CV system, and the norm-one $\| \cdot \|_1$ denotes respectively the trace-norm for the quantum part and the $L_{1}$-norm for the classical part. The physical implementation for the channels $T_n$ and $S_n$, detailed in [@GKJ], is realised via a spontaneous emission coupling of the $n$ qubits to a Bosonic field, and subsequently letting the qubits ’leak’ into this environment. Since there is no correlation between atoms and field, the statistical model decouples into a Gaussian state $\Phi_{\textbf{u}}$ associated to the field, and a classical statistical mixture of atoms, distributed according to $N_{\textbf{u}}$. The corresponding operations of attenuation and amplification may then be carried out in the field in the optimal way. In our case we need to consider mixed qubit states, which means that the collective state has non-zero components in all irreducible representations of SU(2) (all values of the total spin). In fact the traces of the different blocks of given total spin form a probability distribution which (after centring and scaling) converges to the classical Gaussian component of the limit model in LAN. A typical block state of definite total spin can be mapped isometrically into the Fock space of a one mode CV system, and converges to the quantum Gaussian component of the limit model. This transfer can be implemented in principle by a creation-annihilation coupling with a Bosonic field in which the state ‘leaks’’ after a short time. The classical part (total spin) can be measured by coupling subsequently with another Bosonic field, and performing an indirect measurement of $L_{z}$ in the field. Since after the first step, all blocks are in the $|j,j\rangle$ state, a measurement of $L_{z}$ is implicitly a measurement of the total spin. The above convergence can be interpreted as follows: the quantum statistical models $\mathcal{P}_{n}$ can be mapped into the Gaussian model $\mathcal{N}$ and vice-versa, by means of physical operations (quantum channels) with vanishing norm-one error. From the statistical point of view, in many situations this convergence is strong enough to allow us to map a statistical problem concerning the model $\mathcal{P}_{n}$ to a similar one concerning the simpler model $\mathcal{N}$. In the case of purification or dilution of qubits, the mapping into a Gaussian problem is illustrated in the diagram below. We first give a detailed description of the steps involved, and then prove that our procedure is optimal (asymptotically minimax). $$\begin{CD} \rho^{\otimes n}_{\textbf{r}_0+\frac{\textbf{u}}{\sqrt{n}}} @>Q^{\star}_n>> \rho_{\lambda\textbf{r}_0+\frac{\textbf{u}^{\prime}}{\sqrt{m}}}^{\otimes m} \\ @V{T_n}VV @AA{S_m}A \\ N_{\textbf{u}}\otimes \Phi^{s_1}_{\textbf{u}} @>K^{\star}\otimes P^{\star}>> N_{\textbf{u}'}\otimes\Phi^{s_2}_{\textbf{u}'} \\ \end{CD}\label{cd}$$ ![(Color online) Schematic Bloch-sphere geometry for (a) purification and (b) dilution of qubits. A change in Bloch vector magnitude by factor $\lambda$ is reflected by a change in the corresponding Gaussian state displacement. The two are then related by $\lambda=k\sqrt{n/m}$.[]{data-label="figbloch"}](blochgeom3d.pdf){width="8.5cm"} [*Step 1: Localisation*]{}. We are given $n$ identical qubits in an arbitrary mixed state $\rho_{\textbf{r}}$. We measure a small proportion $n^{1-\epsilon}\ll n$ of the qubits, to obtain a rough estimator $\rho_{\textbf{r}_{0}}$ of the state. By standard concentration results, with asymptotically vanishing probability of error, the actual state is in a local neighbourhood of $\rho_{\textbf{r}_{0}}$ of size $n^{-1/2+\epsilon}$, so that the remaining qubits can be parametrised as in the local model . In the same time, the target single-system output state $\rho_{\lambda \textbf{r}}$ belongs to the local model $$\mathcal{Q}_{m}:= \left\{ \rho^{\otimes m}_{\lambda\textbf{r}_0 + \frac{\textbf{u}^{\prime}}{\sqrt{m} }} : \|\textbf{u}^{\prime}\| \leq \lambda \Lambda^{1/2-\epsilon} m^{\epsilon} \right\}$$ with local parameter (see Figure \[figbloch\]) $$\textbf{u}^{\prime}= \lambda \sqrt{\frac{m}{n}} \textbf{u} = \lambda \Lambda^{1/2} \textbf{u} := k\textbf{u}.$$ [*Step 2: Transfer to the Gaussian state*]{}. We apply the map $T_{n}$ to the qubits and obtain a classical random variable and a single-mode CV system whose states are approximately Gaussian \[see \] $$\label{eq.gaussian.input} N_{\textbf{u}} \otimes \Phi_{\textbf{u}} = N(u_{z} ,1- \|\textbf{r}_{0} \|^{2}) \otimes W_{\alpha} \Phi^{s} W_{\alpha}^{\dagger}.$$ [*Step 3: Optimal Gaussian purification (amplification)*]{}. Since $\mathcal{Q}_{m}$ is a local model around $\lambda\textbf{r}_{0}$, the corresponding parameter of the associated Gaussian model is $$s_{2}:= \frac{1- \lambda\|\textbf{r}_{0}\|}{1+\lambda\|\textbf{r}_{0}\|}.$$ and the family of Gaussian states is $$\label{eq.gaussian.output} N(ku_{z}, 1- \|k\textbf{r}_{0} \|^{2}) \otimes W_{k\alpha } \Phi^{s_{2}}W_{k\alpha}^{\dagger}.$$ In this step we attenuate (amplify) the Gaussian state in order to map it into, or close to , as described in Section \[sec2\]. This means that we apply the optimal channel $K^\star$ defined in section \[sec2a\] to the classical component $N_{\bf u}$, and the optimal quantum attenuation (amplification) channel $P^{\star}$ defined in Theorem \[theorem21\], to the quantum Gaussian component $\Phi_{\bf u}$. [*Step 4: Mapping back to the qubits*]{}. We apply the channel $S_{m}$ to the classical variable and the output of the attenuation (amplification) channel to obtain a state of $m$ qubits in the neighbourhood of the state $\rho_{\lambda\textbf{r}_{0}}$. ![(Color online) Parameter range defining the different regimes of qubit dilution. In the shaded region below the boundary curve $\lambda = \tilde\lambda(\|\textbf{r}_0\|)$ \[[Eq. (\[polybart\])]{}\], the threshold $k_0^{(c)}$ for a nonzero classical Gaussian risk is smaller than the corresponding threshold $k^{{\rm amp}}_0$ for a nonzero quantum Gaussian risk, and Cases 3 and 4 of Theorem \[theorem32\] apply for determining the optimal FoM for the qubit problem. Above the boundary curve, the situation is reversed, and Cases 2 and 4 of Theorem \[theorem32\] apply instead.[]{data-label="figpoly"}](figpoly.pdf){width="6cm"} By composing the channels employed in steps 2–4 we obtain the overall channel $$Q_{n}: = S_{n} \circ (K^{\star}\otimes P^{\star})\circ T_n,$$ which will be shown to be optimal. Recall that for $k\leq k_0$ \[see [Eq. (\[koeqn\])]{}\], the quantum component $\Phi^{s_2}_{k\alpha}$ of the Gaussian target state can be prepared exactly. The same is true for the classical component when $k \leq k_0^{(c)}$, where $$\label{k0c} k^{(c)}_0 = \sqrt{\frac{1-\lambda^2\|\textbf{r}_0\|^2}{1-\|\textbf{r}_0\|^2}}$$ is obtained by substituting $V_1=1-\|\textbf{r}_0\|^2$, $V_2=1-\lambda^2\|\textbf{r}_0\|^2$ for the variances in (\[cr\]). This means that the total risk has different expressions over the following three intervals: it is zero when $0<k\leq {\rm min}(k_0 ,k_0^{(c)}) $, it has one classical or quantum contribution for ${\rm min}(k_0 ,k_0^{(c)})< k \leq {\rm max}(k_0 ,k_0^{(c)} )$ and has both quantum and classical contributions for $k> {\rm max}(k_0 ,k_0^{(c)} )$. For purification (corresponding to Gaussian attenuation), the ordering $0<k_0^{\rm att}<k^{(c)}<1$ always holds, so the middle interval has a quantum contribution. However, for dilution (corresponding to Gaussian amplification), the ordering of $k_0$ and $k^{(c)}_0$ depends on the parameters $\|\textbf{r}_0\|$ and $ \lambda$. In particular, we see the appearance of a boundary which demarcates the two separate regimes of dilution, each defined by whether classical or quantum contributions to the risk take place first (see Fig. \[figpoly\]). Namely, $k_0^{(c)}<k^{{\rm amp}}_0$ for $$\label{polybart} \lambda < \tilde\lambda(\|\textbf{r}_0\|) \equiv \min\left\{1,\,\frac{1-\|\textbf{r}_0\|}{\|\textbf{r}_0\|}\right\}\,,$$ and $k_0^{(c)} \ge k^{{\rm amp}}_0$ otherwise. Notice that inequality (\[polybart\]) is always satisfied for $\|\textbf{r}_0\|\leq1/2$ for all values of $\lambda$. ![(Color online) Optimal input ($n$) vs output ($m$) rates $\Lambda_0$ of qubit production for the processes of perfect purification and dilution, \[Eq. (\[eqrate\])\], plotted versus the Bloch vector lengths before ($\|\textbf{r}_0\|$) and after ($\lambda \|\textbf{r}_0\|$) the protocols. The left (right) side of the three-dimensional surface, corresponding to the region $\lambda > 1$ ($\lambda <1$), represents the optimal rate $m/n$ ($n/m$) for mixed qubit purification (dilution).[]{data-label="figrates"}](qubrate.pdf){width="8.5cm"} The relation between the output qubit rate $\Lambda=m/n$ and the constants $\lambda,k$ can be inferred from the geometry of the Bloch sphere (see Fig. \[figbloch\]) $$\Lambda(\|\textbf{r}_0\|,\lambda,k)=\frac{k^2}{\lambda^2}. \label{bsg}$$ In particular, the maximum output rates for which the asymptotic risk is zero are obtained by setting: $k=k_0^{\rm att}$ for purification, and $$k=\left\{ \begin{array}{ll} k_0^{(c)}, & \lambda<\tilde\lambda(\|\textbf{r}_0\|)\,, \\ k_0^{\rm amp}, & \hbox{otherwise,} \end{array} \right.$$ for dilution. Explicitly, $$\begin{aligned} \label{eqrate} \Lambda_0^{\rm pur}(\|\textbf{r}_0\|,\lambda) &=& \frac{ \lambda^{-1} - \|\textbf{r}_0\|}{\lambda^2 (1- \|\textbf{r}_0\|)}\,; \nonumber \\ \\ \nonumber \Lambda_0^{\rm dil}(\|\textbf{r}_0\|,\lambda) &=& \left\{ \begin{array}{ll} \frac{\lambda^{-2}- \|\textbf{r}_0\|^2}{1- \|\textbf{r}_0\|^2}, & \ \ \lambda<\tilde\lambda(\|\textbf{r}_0\|)\,, \\ & \\ \frac{\|\textbf{r}_0\| + 1/\lambda}{\lambda^2 (\|\textbf{r}_0\| + 1)}, & \ \ \hbox{otherwise.} \end{array} \right.\end{aligned}$$ The optimal rates are plotted in Fig. \[figrates\]. We can now state main result of this section whose proof is given in Appendix A. \[theorem32\] The sequence of purification (dilution) maps $$Q^{\star}_n := S_m\circ (K^{\star}\otimes P^{\star}) \circ T_n$$ is locally asymptotically minimax. We distinguish four cases for the minimax risk. **Case** $\mathbf{1}$: Zero risk. If $k< {\rm min}(k_0, k_0^{(c)})$ then the minimax risk is zero. The optimal output rate $\Lambda_0$ is given in . **Case** $\mathbf{2}$: Quantum contribution. If $k_0\leq k \leq k^{(c)}_0$, then the purification (dilution) minimax risk at $\textbf{r}_0$ is equal to the risk of the optimal Gaussian attenuation (amplification) scheme (\[Rmaxpur\]) $$\label{riskqubeq} R_{\rm minmax}(\textbf{r}_0,\lambda,\Lambda)=R_{\rm minmax}(k,s_1,s_2).$$ **Case** $\mathbf{3}$: Classical contribution. If $k^{(c)}_0\leq k\leq k_0$, then the dilution minimax risk at $\textbf{r}_0$ is equal to the risk of the optimal classical Gaussian amplification scheme (\[clrisk\]): $$R_{\rm minmax}(\textbf{r}_0,\lambda,\Lambda)=R_{\rm minmax}(V_{1},V_{2},k).$$ **Case** $\mathbf{4}$: Classical and quantum contributions. If $k> {\rm max}(k_0,k^{(c)}_0)$, then the purification (dilution) minimax risk is $$\begin{aligned} \label{riskqubeqcl} &R_{\rm minmax}(\textbf{r}_0,\lambda,\Lambda)\nonumber \\ &=\int{dx}\sum_n\Bigg|\frac{\exp\bigg(-\frac{x^2}{2k^2(1-\|\textbf{r}_0\|^2)}\bigg)(1-\tilde{s})\tilde{s}^n}{\sqrt{2\pi k^2 \big(1-\|\textbf{r}_0\|^2\big)}}\nonumber \\ &-\frac{\exp\bigg(-\frac{x^2}{2(1-\lambda^2\|\textbf{r}_0\|^2)}\bigg)(1-s_2)s^n_2}{\sqrt{2\pi \big(1-\lambda^2\|\textbf{r}_0\|^2\big)}}\Bigg|\end{aligned}$$ where $\tilde{s}$ takes the values given in (\[tildes\]) in the case of attenuation (for qubit purification) and amplification (for qubit dilution) respectively. The optimal minimax risk for purification and dilution of qubits is plotted in Fig. \[figclass\] as a function of $k$. Conclusions {#sec4} =========== We have solved the practically relevant problem of optimal attenuation and amplification of displaced Gaussian states, with respect to the maximum norm-one distance FoM. As expected, the optimal channels are implemented by the beamsplitter and parametric amplifier respectively, where the ancillary state is provided by the vacuum in both cases. This solution was then used in conjunction with LAN, to construct optimal purification and dilution channels for ensembles of i.i.d. qubits as formulated in Theorem \[theorem32\]. In the Gaussian case, we give an explicit expression of the FoM as a function of the variance parameters $s_{1}$ and $s_{2}$ of input and output states and the attenuation (amplification) factor $k$. In particular we identify the optimal value $k_{0}(s_{1},s_{2})$ for which the protocol achieves the target state exactly. Similarly, in the multiple qubits case, we derive the FoM as a function of the input and output purity and the asymptotic input/output rate $\Lambda$, and identify the optimal rate $\Lambda_{0}$ for which the the protocol achieves the target collective state exactly. Both classical and quantum Gaussian contributions to the risk have to be taken into account to calculate the maximum rates, and to provide the optimal FoM for purification and dilution of multiple qubits, in the parameter range where the procedures cannot be accomplished perfectly. An interesting future project is to extend the techniques used in this paper to tackle the general problem of asymptotically optimal channel inversion for arbitrary channels on finite dimensional systems. Such a study may provide efficient strategies to counteract the effect of various types of noise and decoherence processes, beyond the depolarising channel considered in the present work. Proofs ====== Proof of Theorem \[theorem21\] {#proof-of-theorem-theorem21 .unnumbered} ------------------------------ As the proof follows the lines of similar results in [@GM; @GBA] we will briefly sketch the main ideas. A covariance argument [@C; @O] shows that one may restrict the attention to channels which are displacement-covariant, in the sense that $P(W_\xi \Phi W^\dagger_\xi) = W_\xi P(\Phi)W^\dagger_\xi$ for any input state $\rho$. For such channels the risk is independent of $\alpha$ and $${R}_{\max}(k,s_1,s_2,P) = \|P(\Phi^{s_1}) - \Phi^{s_2}\|_1$$ In the case of attenuation $(k<1)$ such channels are described by the linear transformation $$c_{\rm att} = ka + \sqrt{1-k^2}b$$ where $a$ is the input mode, $c_{\rm att}$ is the output and $b$ is an ancillary mode prepared in a state $\tau$. Since the channel is completely characterised by the state $\tau$, we will denote it by $P_\tau$. Similarly, for amplification $(k>1)$ the output of the channel $P_\tau$ is the mode $$c_{\rm amp} = ka + \sqrt{k^2-1}b^\dagger,$$ with $b$ prepared in the state $\tau$. By a second covariance argument with respect to phase rotations, and taking into account that $\Phi$ is invariant under phase rotations, we obtain that $\tau$ can be taken to be phase-invariant, i.e. it is a mixture of Fock states $\tau=\sum_i\tau_i|i\rangle\langle i|$. In this case the output state will be diagonal in the Fock basis and we write $P_\tau(\Phi^{s_1}) = \sum_ip^\tau_i|i\rangle\langle i|$, and in particular $p^\omega$ corresponds to the output state when the ancilla is the vacuum. Similarly, we denote the coefficients of the Gaussian state $\Phi^{s_1}$ and $\Phi^{s_2}$ by $p_i=(1-s_1)s^i_1$ and $q_i=(1-s_2)s^i_2$. The proof reduces now to showing that, for any $\tau$, $$\label{eq.norm.ineq} \|p^\tau - q\|_1 \geq \|p^\omega - q\|_1.$$ The key to proving this statement is the concept of stochastic ordering, whose definition we recall: Let $p=\{p_l:l\in\mathbb{N}\}$ and $q = \{q_l: l\in\mathbb{N}\}$ be two probability distributions over $\mathbb{N}$. We say that $p$ is stochastically smaller than $q$ $(p\preceq q)$ if $$\sum^m_{l=0} p_l \geq \sum^m_{l=0} q_l, \quad \forall m \geq 0$$ The following lemma holds for both the purification and the amplification scenarios: \[lemma42\] For any state $\tau$ the following stochastic ordering holds $$p^{\omega} \preceq p^{\tau}. \label{vacstoch}$$ We treat the attenuation and amplification separately but the idea is the same in both cases: we reduce the statement about stochastic ordering to a simpler one where the input mode is in the vacuum. [*Attenuation.*]{} We write the input mode as $a = \cosh(t)a_1 + \sinh(t)a^\dagger_2$ with $a_{1,2}$ two fictitious modes in the vacuum state, and $\tanh^2(t) := s_{1}$, which ensures that the state of $a$ is $\Phi^{s_{1}}$. Let $\tilde{t}$ be such that $\sinh(\tilde{t})=k\sinh(t)$ and denote $$T = \sqrt{1 - \frac{(1 - k^2)}{\cosh^2(\tilde{t})}}, \qquad R = \frac{\sqrt{1 - k^2}}{\cosh(\tilde{t})}.$$ Then $c_{pur}=ka+\sqrt{1-k^{2}}b $ can be written as $$\begin{aligned} c_{pur}& =& \cosh(\tilde{t})(Ta_1 + Rb) + \sinh(\tilde{t})a^\dagger_2 \\ &=& \cosh(\tilde{t})\tilde{b} + \sinh(\tilde{t})\tilde{a}^\dagger\end{aligned}$$ where $\tilde{b} := Ta_1 + Rb$ and $a_{2} $ was relabelled $\tilde{a}$. The state of the mode $\tilde{b}$ is given by $$\begin{aligned} \tilde{\tau} &=& \sum^\infty_{k=0} \tau_k \sum^k_{p=0} \binom{k}{p} T^{2(p-k)} R^{2k} |p\rangle\langle p|\\ & =& \sum^\infty_{p=0} \tilde{\tau}_p |p\rangle\langle p|\end{aligned}$$ so $\tilde{b}$ is in the vacuum state if and only if $b$ is in the vacuum. Thus it suffices to prove the stochastic ordering statement for the mode $c_{\rm att}$ written as a combination of $\tilde{b}$ and $\tilde{a}$ for an arbitrary diagonal state $\tilde{\tau}$ of $\tilde{b}$ and $\tilde{a}$ in the vacuum. Furthermore, since stochastic ordering is preserved under convex combinations, it suffices to prove the statement for any [*pure*]{} diagonal state $\tilde{\tau} = |k\rangle\langle k|$, $k\neq 0$. In this case the state of $c_{\rm att}$ is given by $$\begin{aligned} \rho^{\rm out}_{\rm att} &=& e^{-2g(k+1)}\sum^\infty_{l=0} \Gamma^{2l} \binom{l+k}{k} |l+k\rangle\langle l+k| \\ &:=& \sum^m_{l=0} d^{(k)}_l |l+k\rangle\langle l+k|\end{aligned}$$ where $\Gamma = \tanh({\tilde{t}})$ and $e^g = \cosh(\tilde{t})$. The stochastic ordering now reduces to showing that $\sum^m_{l=0} d^{(0)}_l \geq \sum^m_{l=0} d^{(k)}_l$ for all $m$. With the notation $\gamma = \Gamma^2$, we get $$\begin{aligned} \sum^{p+k}_{l=0}d^{(k)}_l &=& (1-\gamma)^{k+1} \sum^p_{l=0}\gamma^l\binom{l+k}{k} \\ &\leq& 1 - \gamma^{p+1} \sum^k_{r=0} (1-\gamma)^r\gamma^{k-r}\binom{k}{r} \\ &=& 1-\gamma^{p+1} = \sum^p_{l=0} d^{(0)}_l.\end{aligned}$$ [*Amplification.*]{} As before we write $a = \cosh(t)a_1 + \sinh(t)a^\dagger_2$ and define $\tilde{t}$ by $\cosh({\tilde{t}}) = k\cosh(t)$ and the beamsplitter coeficients $$T = \sqrt{1 - \frac{(1 - k^2)}{\sinh^2(\tilde{t})}} \qquad R = \frac{\sqrt{1 - k^2}}{\sinh(\tilde{t})}.$$ The output mode is now $$\begin{aligned} c_{\rm amp} &=& \sinh(\tilde{t})(Rb^\dagger + Ta^\dagger_2) + \cosh(\tilde{t})a_1 \\ &=& \sinh(\tilde{t})\tilde{b}^\dagger + \cosh(\tilde{t})\tilde{a}\end{aligned}$$ where we have relabelled $a_1$ by $\tilde{a}$ and introduced the mode $\tilde{b}=Rb^\dagger + Ta^\dagger_2$. As before, the state of $\tilde{b}$ is the vacuum if and only if $b$ is in the vacuum state, so it suffices to verify the statement for the state $\tilde{\tau}=|k\rangle \langle k|$ in which case the output state is $$\rho^{\rm out}_{\rm amp} = e^{-2g(k+1)}\sum^\infty_{l=0} \Gamma^{2l} \binom{l+k}{k} |l\rangle\langle l| = \sum^m_{l=0} d^{(k)}_l|l\rangle\langle l|$$ The relation $p^\omega \preceq p^\tau$ now follows from $$\begin{aligned} \sum^{p}_{l=0}d^{(k)}_l &=& (1-\gamma)^{k+1} \sum^p_{l=0}\gamma^l\binom{l+k}{k} \\ &\leq& 1-\gamma^{p+1}= \sum^p_{l=0} d^{(0)}_l.\end{aligned}$$ This ends the proof of Lemma \[lemma42\] for both cases. The following lemma completes the proof of Theorem \[theorem21\] by transforming the stochastic ordering into the desired norm inequality . Its proof [@GM; @GBA] uses the fact that $q \preceq p^{\omega} $ which is equivalent to the fact that $P^{\star}(\Phi^{s_{1}})$ is more noisy that $\Phi^{s_{2}}$. The latter is satisfied for $k\geq k_{0}$ as assumed in the theorem. Let $ p^{\prime}$ be a discrete probability distribution such that $p^{\omega}\preceq p^{\prime}$. Then $$\|p^\prime - q\|_1 \geq \|p^\omega - q\|_1.$$ Proof of Lemma \[lemma22\] {#proof-of-lemma-lemma22 .unnumbered} -------------------------- We use the notations introduced in the proof of Theorem \[theorem21\]. By expressing the quadrature variance of the input mode $a$ in terms of $t$ and $s_1$ we obtain $\sinh^2t = \frac{1}{e^{s_1} - 1}$. According to Theorem \[theorem21\] the output state of the optimal channel $P^{\star}$ is the Gaussian state $$P^{\star}( \Phi^{s_{1}})= \Phi^{\tilde{s}} = e^{-2g}\sum^\infty_{l=0}(1-e^{-2g})^l|l\rangle\langle l|$$ with $g$ taking different values in the attenuation and amplification cases. For the geometric distributions $p^{\omega}$ and $q$ we have $$\| \Phi^{\tilde{s}}- \Phi^{s_{2}} \|_{1}= \|p^{\omega}-q\|_{1} = 2(\tilde{s}^{m_{0}+1} - s_{2}^{m_{0}+1})$$ where $m_{0}$ is the largest integer such that $p^{\omega}_{m_{0}} \leq q_{m_{0}}$, more precisely $$m_{0}= \lfloor \ln [ (1-\tilde{s})/(1-s_{2})] /\ln(s_{2}/\tilde{s}) \rfloor$$ It remains to compute the concrete expressions of $\tilde{s}$ and implicitly of $m_{0}$ for the attenuation and amplification cases. For attenuation, making use of $\sinh\tilde{t} = k\sinh t$, we find $$\tilde{s}_{pur} = \frac{s_1 k^2}{1-s_1 + s_1 k^2}.$$ For amplification, we use $\cosh\tilde{t} = k\cosh t$ and find $$\tilde{s}_{amp}= 1-\frac{1-s_1}{k^2}.$$ Proof of Theorem \[theorem32\] {#proof-of-theorem-theorem32 .unnumbered} ------------------------------ We want to show that $Q^\star:=S_n\circ P^\star \circ T_n$ is the optimal purification or dilution procedure for $n$ i.i.d. qubits. The idea is that, by using LAN, we can show the qubit and Gaussian statistical problems to be equivalent, the Gaussian one (respectively for attenuation and amplification) being solved in Section \[sec2b\], which then allows us to recast the qubit problem in the Gaussian setup with a vanishing difference in the risks. We will consider the four separate cases: zero risk, solely quantum contribution, solely classical contribution, and both classical and quantum contributions. We will then use the Gaussian solution to show that $R_{\max}(\textbf{r}_0,Q^\star,\lambda)$ is less than or equal to the corresponding optimal Gaussian risk, then show that a strict inequality violates the optimality of this optimal solution. We begin by restricting $\textbf{r}$ to the local neighbourhood $\|\textbf{r}-\textbf{r}_0\|\leq n^{-\frac{1}{2}+\epsilon}$. This probability that the state fails to be in this region is $o(1)$ and has no influence on the asymptotic risk (see Lemma 2.1. in [@GJK]). We are now able to apply LAN, which maps input states $\rho^{\otimes n}_{\textbf{r}}$ close to some Gaussian state, say $\tilde{\Phi}_{\textbf{u}}$, via the channel $T_n$. We now consider the individual cases, which are each slight variations on the same proof: **Case**$\mathbf{1}$: $k<{\rm min}(k_0,k_0^{(c)})$. In this case both classical and quantum Gaussian channels have zero risk, so the asymptotic qubit risk is zero. **Case** $\mathbf{2}$: $k_0\leq k\leq k^{(c)}_0$. In this instance, the risk receives only a quantum contribution. Using contractivity of the CP maps $S_m$ and $P^\star$, the LAN convergence, and the fact that in this regime $K^\star N_{\bf u} = N_{{\bf u}^\prime}$ we obtain $$\begin{aligned} &R(Q^\star_n, \textbf{r},\lambda)\nonumber\\ &\qquad =\|\rho^m_{\textbf{u}'}-Q^\star_n(\rho^n_{\textbf{u}})\|_1\nonumber \\ &\qquad\leq\|\rho^m_{\textbf{u}^\prime}-S_m(\Phi^{s_2}_{\textbf{u}^\prime} \otimes N_{{\bf u}^\prime})\|_1\nonumber \\ &\qquad ~+\|S_m(\Phi^{s_2}_{\textbf{u}'} \otimes N_{{\bf u}^\prime}) )-S_m\big(P^\star \otimes K^\star \big(T_n(\rho^n_{\textbf{u}})\big)\big)\|_1\nonumber \\ &\qquad\leq\|\Phi^{s_2}_{\textbf{u}^\prime} \otimes N_{{\bf u}^\prime} -P^\star \otimes K^\star \big(T_n(\rho^n_{\textbf{u}})\big)\|_1+o(1)\nonumber \\ &\qquad = \|\Phi^{s_2}_{\textbf{u}^\prime} \otimes N_{{\bf u}^\prime} -P^\star \otimes K^\star \big( \Phi^{s_1}_{\textbf{u}} \otimes N_{\bf u}\big)\|_1 +o(1) \nonumber\\ &\qquad = \|\Phi^{s_2}_{\textbf{u}^\prime}-P^\star\big( \Phi^{s_1}_{\textbf{u}})\|_1 +o(1)\nonumber \\ &\qquad = R_{\rm minmax}(s_1,s_2,k) + o(1)\label{eq.upperbound}\end{aligned}$$ where $R_{\rm minmax}(s_1,s_2,k) $ is the minimax risk for the quantum Gaussian problem, obtained in Theorem \[theorem21\]. By taking supremum over $\|{\bf u}\|<n^\epsilon$ we get $$R_{\rm max}(Q^\star_n, \textbf{r}_0,\lambda) \leq R_{\rm minmax}(s_1,s_2,k).$$ which implies that $$R_{\rm minmax}(\textbf{r}_0,\lambda) \leq R_{\rm minmax}(s_1,s_2,k).$$ Next, we show by contradiction that this inequality cannot be strict. Suppose that there exists a sequence of purification or dilution procedures $\tilde{Q}_n$, which act on qubits and satisfies $R_{max}(\tilde{Q}_n\textbf{r}_0,\lambda)\leq R_{\rm minmax}(s_1,s_2,k)- \eta$ for some $\eta>0$ and $n>n_0$. We will use LAN to show that there exists a Gaussian dilution (amplification) channel whose risk is strictly smaller than the minimax risk, which is a contradiction. The general setup can be seen in (\[ocd\]) $$\begin{CD} \rho^{\otimes n}_{\textbf{u}} @>\tilde{Q}_n>> \rho^{\otimes m}_{\lambda\textbf{r}} \\ @AA{S_n}A @V{T_m}VV \\ \Phi^{s_1}_{\textbf{u}} @>\tilde{P}>> \Phi^{s_2}_{k\alpha} \\ \end{CD}\label{ocd}$$ Here LAN is restricted to a [*two dimensional*]{} family of rotated qubit states, for which the limit model is quantum Gaussian, with no classical component. Assuming $\Phi^{s_1}_{\textbf{u}}$ is in the domain of applicability of LAN (which can be effected by an adaptive measurement [@GBA]), we get the inequalities $$\begin{aligned} &\|T_m\circ\tilde{Q}_n\circ S_n(\Phi^{s_1}_{\textbf{u}})-\Phi^{s_2}_{\textbf{u}'}\|\nonumber \\ &\qquad\leq\|\tilde{Q}_n(\rho^n_{\textbf{u}})-\rho^m_{\textbf{u}'}\|_1+\|T_m(\rho^m_{\textbf{u}'})-\Phi^{s_2}_{\textbf{u}'}\|_1\nonumber \\ &\qquad\leq R_{\rm minmax}(s_1,s_2,k) -\eta + o(1).\end{aligned}$$ By taking the limit $n\to \infty$ we get the desired contradiction. **Case** $\mathbf{3}$: $k^{(c)}_0\leq k\leq k_0$. This case applies only to dilution and the risk receives only a classical contribution. The proof follows the same steps as the previous case, with the quantum Gaussian replaced by the classical one. The inequality becomes $$\begin{aligned} &R(Q^\star_n, \textbf{r} ,\lambda) \nonumber\\ & \qquad =\|\rho^m_{\textbf{u}'}-Q^\star_n(\rho^n_{\textbf{u}})\|_1\nonumber \\ &\qquad\leq \|N(0,1-\lambda^2\|\textbf{r}_0\|^2)-K^\star\big( N(0, 1-\|\textbf{r}_0\|^2)\big)\|_1+o(1)\nonumber \\ &\qquad = R_{\rm minmax}(V_{1},V_{2},k) + o(1).\end{aligned}$$ where $R_{\rm minmax}(V_{1},V_{2},k)$ is the optimal risk of Eq. (\[clrisk\]) and we have identified $V_1=1-\|\textbf{r}_0\|^2$, $V_2=1-\lambda^2\|\textbf{r}_0\|^2$. This implies $$R_{\rm minmax}(\textbf{r}_0,\lambda) \leq R_{\rm minmax}(V_{1},V_{2},k).$$ The equality is obtained by showing that strict inequality would lead to a classical amplification procedure whose risk is smaller than the minimax risk. **Case** $\mathbf{4}$: $k>{\rm max}(k^{(c)}_0, k_0)$. This case applies to both dilution and amplification, and both the quantum and classical channels contribute to the risk. $$\begin{aligned} &R_{\max}(Q_n^\star,\textbf{r},\lambda)\nonumber\\ &\qquad = \|\rho^m_{\textbf{u}'}-Q_n^\star(\rho^n_{\textbf{u}})\|_1\nonumber \\ &\qquad \leq\|\Phi^{s_2}_{\textbf{u}'}\otimes N_{{\bf u}^\prime}- K^\star\otimes P^\star\big( \Phi^{s_1}_{\textbf{u}}\otimes N_{{\bf u}} \big)\|_1+o(1)\nonumber \\ &\leq R_{\rm minmax}(V_{1},V_{2},s_1,s_2,k)+o(1).\label{a9}\end{aligned}$$ Here $R_{\rm minmax}(V_{1},V_{2},s_1,s_2,k)$ is the minimax norm-one risk for the problem of transforming the Gaussian state $\Phi^{s_1}_{\textbf{u}}\otimes N_{{\bf u}}$ into $\Phi^{s_2}_{\textbf{u}'}\otimes N_{{\bf u}^\prime}$. Since the quantum and classical components are independent and have different local parameters, the optimal channel is the product $K^\star\otimes P^\star$. The explicit expression of the minimax risk is $$\begin{aligned} &R_{\rm minmax}(V_{1},V_{2},s_1,s_2,k)\nonumber \\ &\quad = \int{dx}\sum_n\bigg|\frac{\exp\bigg(-\frac{x^2}{2k^2(1-\|\textbf{r}_0\|^2)}\bigg)}{\sqrt{2\pi k^2 \big(1-\|\textbf{r}_0\|^2\big)}}(1-\tilde{s})\tilde{s}^n\nonumber \\ &\qquad-\frac{\exp\bigg(-\frac{x^2}{2(1-\lambda^2\|\textbf{r}_0\|^2)}\bigg)}{\sqrt{2\pi \big(1-\lambda^2\|\textbf{r}_0\|^2\big)}}(1-s_2)s^n_2\bigg| \label{qacopt}\end{aligned}$$ Finally, the equality in (\[a9\]) can be proven by contradiction as in case 2. [99]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{} , , , , , , ****, (). , ****, (). , , , ****, (). M. Keyl and R. F. Werner, Annales Henri Poincare [**2**]{}, 1 (2001). , , , ****, (). , , , , ****, (). N. Gisin and S. Massar, Phys. Rev. Lett. [**79**]{}, 2153 (1997) . , ** (, ). , ****, (). , , , ****, (). , ****, (). , ****, (). , ****, (). M. Guţă, P. Bowles, and G. Adesso, Phys. Rev. A [**82**]{}, 042310 (2010). M. Guţă and W. Kotlowski, New J. Phys. [**12**]{}, 12303 (2010). , , , ****, (). , ****, (). ****, (). , , , , , ****, (). Z. Y. Ou, S. F. Pereira, and H. J. Kimble, Phys. Rev. Lett. [**70**]{}, 3239 (1993). , ****, (). , ** (, ). J. Bae and A. Acin, Phys. Rev. Lett. [**97**]{}, 030402 (2006). A. W. Van Der Waart, [*Asymptotic Statistics*]{} (Cambridge University Press, Cambridge, UK, 1998) , . M. Guţă, e–print arXiv:1007.0434 (2010). , , , , , ****, (). , , , , ****, (). [^1]: Corresponding author.\ Electronic address: [[email protected]]{}

ArXiv

--- abstract: | Recently, we have proposed *coordinated choices*, which are nondeterministic choices equipped with names. The main characteristic of coordinated choices is that they synchronize nondeterministic decision among choices of the same name. The motivation of the synchronization mechanism is to solve a theoretical problem. So, as a practical programming language, we still want to use coordinated choices like standard ones. In other words, we want to avoid synchronization. Now, there are two problems: (i) practically, it is a bit complicated work to write a program using coordinated choices in which execution synchronization never happens; and (ii) theoretically, it is unknown whether any programs using standard choices can be written by using only coordinated ones. In this paper, we define two simply typed lambda calculi called [$\lambda^\parallel$]{} equipped with standard choices and [$\lambda^{\parallel\omega}$]{} equipped with coordinated choices, and give compilation rules from the former into the latter. The challenge is to show the correctness of the compilation because behavioral correspondence between expressions before and after compiling cannot be defined directly by the compilation rules. For the challenge, we give an effect system for [$\lambda^{\parallel\omega}$]{} that characterizes expressions in which execution synchronization never happens. Then, we show that all compiled expressions can be typed by the effect system. As a result, we can easily show the correctness because the main concern of the correctness is whether synchronization happens or not. author: - Yuki Nishida - Atsushi Igarashi bibliography: - 'reference.bib' title: Compilation of Coordinated Choice --- Conclusion ========== We give a compilation algorithm from [$\lambda^\parallel$]{}, a simply typed lambda calculus with nondeterministic choices, into [$\lambda^{\parallel\omega}$]{}, a simply typed lambda calculus with coordinated choices; and show the compilation is sound and correct.

ArXiv

--- abstract: | We study regularity properties of solutions to reaction-diffusion equations ruled by the infinity laplacian operator. We focus our analysis in models presenting plateaus, i.e. regions where a non-negative solution vanishes identically. We obtain sharp geometric regularity estimates for solutions along the boundary of plateaus sets. In particular we show that the $(n-\epsilon)$-Hausdorff measure of the plateaus boundary is finite, for a universal number $\epsilon>0$. **Keywords:** Reaction-diffusion equations, infinity laplacian, regularity **AMS Subject Classifications:** 35J60, 35B65 author: - 'Damião J. Araújo Raimundo Leitão Eduardo V. Teixeira' title: Infinity Laplacian equation with strong absorptions --- Introduction ============ The mathematical analysis of problems involving the infinity Laplacian operator, $$\label{infty_lap eq} \Delta_\infty u := \sum\limits_{i,j} \partial_i u \partial_{ij} u \partial_{j} u = (D u)^TD^2u\; Du,$$ constitutes a beautiful chapter of the modern theory of partial differential equations, yet far from its denouement. The systematic study of problems involving the infinity laplacian operator has been originated by the pioneering works of G. Aronsson [@A1; @A2]. The initial purpose of this line of research is to answer the following natural question: given a bounded domain $O \subset \mathbb{R}^n$ and a Lipschitz function $g \colon \partial O \to \mathbb{R}$, find its best Lipschitz extension, $f$, in the sense that it agrees with $g$ on the boundary and for any $O' \Subset O$, if $f = h$ on $\partial O'$, then $\|f\|_{\text{Lip}(O')} \le \|h\|_{\text{Lip}(O')}$. Such a function $f$ is said to be an absolutely minimizing Lipschitz extension of $g$ in $O$. Jensen in [@J] has proven that a function in an absolutely minimizing Lipschitz extension if, and only if, it is a viscosity solution to the homogeneous equation $\Delta_\infty u = 0.$ That is, the infinity Laplacian rules the Euler-Lagrange equation associated to this $L^\infty$ minimization problem. Through the years, several different applications of the infinity Laplacian theory emerged in the literature, [@CMS; @PSS; @BCF], just to cite few. We refer to [@ACJ] for an elegant discussion on the theory of absolutely minimizing Lipschitz extensions. While, existence and uniqueness of viscosity solution to the homogeneous Dirichlet problem $\Delta_\infty h = 0$, in $O$, $u = g$, on $\partial O$ is nowadays fairly well established, obtaining improved regularity estimates for infinity harmonic functions remains a major open issue in the theory of nonlinear partial differential equations. The example of Aronsson $$h(x,y) = x^{{4}/{3}} - y^{{4}/{3}}$$ hints out to one of the most famous conjecture in this field: the first derivatives of infinity harmonic functions should be Hölder continuous with optimal exponent $\frac{1}{3}$. The best results known up to date are due to Evans and Savin, [@Evans_Savin], who proved that infinity harmonic functions in the plane are of class $C^{1,\alpha}$, for some $0<\alpha \ll 1$, see also [@Savin], and to Evans and Smart, [@ES], who obtained everywhere differentiability for infinity harmonic functions in any dimension. The theory of inhomogeneous infinity laplacian equations $\Delta_\infty u = f(X)$ is more recent and subtle. Lu and Wang in [@LW] has proven existence and uniqueness of continuous viscosity solutions to the Dirichlet problem $$\label{NH eq} \left \{ \begin{array}{rlll} \Delta_\infty u &=& f(X) &\text{ in } O \\ u &=& g &\text{ on } \partial O, \end{array} \right.$$ provided the source function $f$ does not change sign, i.e. either $\inf f > 0$ or else $\sup f < 0$. Uniqueness may fail if such a condition is violated, [@LW Appendix A]. While Lipschitz estimates and everywhere differentiability also hold for a function whose infinity laplacian is bounded in the viscosity sense, see [@E], no further regularity is so far known for inhomogeneous equations. This current work is devoted to the study of reaction-diffusion models ruled by the infinity Laplacian operator. Namely, for $\lambda > 0$ and $0 \leq \gamma < 3$, let $$\label{Lgamma} \mathcal{L}^{\,\gamma}_{\, \infty} \, v := \Delta_{\,\infty} v - \lambda (v^+)^\gamma$$ denote the $\infty$-diffusion operator with $\gamma$-strong absorption. The case $\gamma =0$ is related to the infinity-obstacle problem, [@RTU]. The constant $\lambda >0$ is called the Thiele modulus, which adjusts the ratio of reaction rate to diffusion–convection rate. Given a bounded domain $\Omega \subset \mathbb{R}^n$, $n\ge 2$, and a continuous, nonnegative boundary value datum $g \in C(\partial \Omega)$, we study existence, uniqueness and regularity issues to the Dirichlet problem $$\label{eq} \left\{ \begin{array}{rcc} \mathcal{L}^{\,\gamma}_{\, \infty} \, u = 0 & \mbox{in} & \Omega \\ u = \phi & \mbox{on} & \partial\Omega. \\ \end{array} \right.$$ An important feature in the mathematical formulation of equation is the possible existence of plateaus, i.e., a priori unknown regions where the function vanishes identically. Upon establishing existence of a viscosity solution, equation can be regarded as a inhomogeneous infinity laplacian equation; however the corresponding source function is not bounded away from zero. Notwithstanding, as a preliminar result, we show uniqueness, up-to-the-boundary continuity, and non-negativeness of viscosity solution to Equation , Theorem \[existence theorem\]. The proof is based on comparison principle methods, proven to hold for the operator $\mathcal{L}_\infty^\gamma$. The heart of the matter, though, lies on geometric regularity estimates for the solution to Equation . While it follows by classical considerations that bounded viscosity solutions are locally Lipschitz continuous, no further smoothness property can be inferred by the existing theory. The main result we show in this work assures that a viscosity solution to Equation is [*pointwisely*]{} of class ${C}^{\frac{4}{3-\gamma}}$ along the boundary of the non-coincidence set, $\partial \{u > 0 \}$, Theorem \[thmreg\]. One should notice that for each $0 < \gamma < 3$, the regularity estimate established in Theorem \[thmreg\] is superior than the optimal $C^{1,\frac{1}{3}}$-estimate, yet to be confirmed (or not), for infinity harmonic functions. Hence, it is clear that such a geometric, improved estimate cannot be extended inwards the non-coincidence set $\{u>0\}$. Nonetheless, such an estimate does enforce rather specific geometric information on the behavior of $u$ near the boundary of the coincidence set. By means of barriers, we show that such an estimate is optimal, Theorem \[nondegthm\], in the sense that $u$ detaches from its coincidence set precisely as $\text{dist}^{\frac{4}{3-\gamma}}$. This fact allows us to derive Hausdorff measure estimates for $\partial \{u>0\}$, Corollary \[H est\]. We conclude this introduction by pointing out that similar results can be derived to problems with more general absorption terms: $\Delta_\infty u = f(u)$. We have chosen to present this current article for $f(u) = \lambda (u^{+})^\gamma$ as to highlight the main novelties introduced in our analysis. Notations ========= In this article we shall use classical notations and terminologies, which, for the sake of the readers, we list below. The dimension of Euclidean space in which the equations and problems treated in this article are modeled into will be denoted by $n$. Given $\mathscr{O}$ a subset of the $\mathbb{R}^n$, we denote by $\partial\mathscr{O}$ its boundary. For $B_r(X) \subset \mathbb{R}^n$ we denote the open ball of radius $r>0$ centered at $X\in \mathbb{R}^n$. For the vectors $\vec{p}=(p_1,\cdots,p_n)$ and $\vec{q}=(q_1,\cdots,q_n)$, we consider $\langle \vec{p},\vec{q} \rangle$ the standard scalar product in $\mathbb{R}^n$ and $|\vec{p}|:=\sqrt{\langle \vec{p},\vec{p} \rangle}$ its Euclidean norm. The tensor product $\vec{p}\otimes \vec{q}$ denotes the matrix $(p_i \cdot q_j)_{1 \leq i,j \leq n}$. For a real function $\omega$ defined in a open subset of the $\mathbb{R}^n$, we denote by $$D\omega(X):=(\partial_j \omega(X))_{1 \leq j \leq n} \quad \mbox{and} \quad D^2\omega(X):=(\partial_{ij} \omega(X))_{1 \leq i,j \leq n}$$ its gradient and its hessian at the point $X\in \mathbb{R}^n$, where $\partial_i \omega$ is a $i$-th directional derivative of $\omega$ and $\partial_{ij}\omega$ the $j$-th directional derivative of $\partial_i\omega$. Fixed a domain $\Omega \subset \mathbb{R}^n$, we will call universal any positive constant that depends only on dimension, $\gamma$ and $\Omega$. For an operator $G \colon \mathscr{O} \times \mathbb{R}^n \times \mbox{Sym}(n) \to \mathbb{R}$ and a domain $\mathscr{O}\subset \mathbb{R}^n$, a continuous function $\omega \colon \mathscr{O} \to \mathbb{R}$ is called a *viscosity subsolution* of the equation $$\label{eq not} G(X,\omega,D\omega,D^2\omega)=0 \text{ in } \mathscr{O},$$ if whenever $ \varphi \in C^2$ is such that $\omega-\varphi$ has a local maximum at some point $Y \in \mathscr{O}$, then there holds $$G (Y,\omega(Y),D\varphi(Y),D^2\varphi(Y))\geq 0.$$ Similarly, a continuous function $\omega \colon \mathscr{O} \to \mathbb{R}$ is called a *viscosity supersolution* of equation , if $ \varphi \in C^2$ is such that $\varphi-\omega$ has a local maximum at some point $Y \in \mathscr{O}$, then there holds $$G (Y,\omega(Y),D\varphi(Y),D^2\varphi(Y))\leq 0.$$ We say $\omega$ a *viscosity solution* of the $G(X,\omega,D\omega,D^2\omega)=0$ when $\omega$ is both a subsolution and a supersolution. Preliminaries ============= In this Section we make a preliminar analysis on equation . Initially, we point that, for the purposes of this article, the Thiele modulus plays no important role, and hence, hereafter, we shall take $\lambda = 1$. We start off by verifying that any existing viscosity supersolution to , $\mathcal{L}_\infty^\gamma u \le 0$, is nonnegative. Indeed suppose the open set $\mathscr{O}(u):=\{u<0\}$ were nonempty. Then $u$ would satisfy in $\mathscr{O}(u)$ $$\begin{aligned} \label{Nonneg of u 1} \left \{ \begin{array}{rcr} \Delta_{\,\infty} u \leq 0, & \text{ in } & \mathscr{O}(u)\; \\ u= 0, & \text{ on } & \partial \mathscr{O}(u) . \end{array} \right.\end{aligned}$$ By the classical comparison principle for infinity-harmonic functions, see for instance [@J], $u \geq 0$ in $\mathscr{O}(u)$, which drives us to a contradiction. We now briefly comment on existence of a viscosity solution to the Dirichet problem . As usual it follows by an application of Perron’s method once comparison principle is established. Indeed, let us consider the functions $\overline{u}$ and $\underline{u}$, solutions to the following boundary value problems: $$\nonumber \begin{array}{ccc} \left\{ \begin{array}{rcc} \Delta_\infty\, \overline{u}= 0 & \mbox{in} & \Omega, \\ \overline{u} = \phi & \mbox{on} & \partial\Omega.\\ \end{array} \right. & \mbox{and} & \left\{ \begin{array}{rllcc} \Delta_\infty\, \underline{u} &=& \|\phi\|_{L^\infty(\partial\Omega)}^\gamma & \mbox{in} & \Omega, \\ \underline{u} &=& \phi & \mbox{on} & \partial\Omega.\\ \end{array} \right. \\ \end{array}$$ Existence of such solutions follows of standard arguments. We note that $\overline{u}$ and $\underline{u}$ are respectively, supersolution and subsolution to . Therefore by Comparison principle, Lemma \[comparison principle\] below, it is possible, under a direct application of Perron’s method, to obtain the existence of a viscosity solution in $C(\overline{\Omega})$ of , given by $$u(X):=\inf\{\omega(X)\; \vert\; \omega \;\mbox{is a supersolution of}\; \eqref{eq} \; \mbox{and} \; \underline{u}\, \leq \omega \leq \overline{u} \; \mbox{in} \;\overline{\Omega}\}.$$ Uniqueness also follows readily from comparison principle. We state these observations as a Theorem for future references. \[existence theorem\] Let $\Omega\subset \mathbb{R}^{n}$ be a bounded domain and $\varphi \in C\left( \partial \Omega \right)$ be a given nonnegative function. Then there exists a nonnegative function $u \in C\left( \overline{\Omega}\right)$ satisfying in the viscosity sense. Moreover, such a solution is unique. We now deliver a proof for comparison principle for the operator $\mathcal{L}_\infty^\gamma$. The reasoning is somewhat standard in the theory of viscosity solutions; we carry out the details for the reader’s convenience. \[comparison principle\] Let $u_1$ and $u_2$ be continuous functions in $\overline{\Omega}$ satisfying $$\mathcal{L}^\gamma_\infty \, u_1 \leq 0 \quad \mbox{and} \quad\mathcal{L}^\gamma_\infty \, u_2 \geq 0 \text{ in } \Omega.$$ If $u_1 \geq u_2$ on $\partial \Omega$, then $u_1 \geq u_2$ inside $\Omega$. Let us suppose, for the purpose of contradiction, that there exists $M_0>0$ such that $M_0= \sup\limits_{\overline{\Omega}}\left(u_2-u_1 \right)$. For each $\varepsilon>0$ small, define $$M_\varepsilon:= \sup_{\overline{\Omega}\times \overline{\Omega}} \left(u_2(X)-u_1(Y)-\frac{1}{2\varepsilon}|X-Y|^2 \right) < \infty.$$ Let $(X_\varepsilon,Y_\varepsilon) \in \overline{\Omega}\times \overline{\Omega}$ be a point where the maximum is attained. It follows as in [@UG lemma 3.1] that $$\label{e0} \lim\limits_{\varepsilon \to 0} \frac{1}{\varepsilon}|X_\varepsilon-Y_\varepsilon|^2=0, \quad \mbox{and} \quad \lim\limits_{\varepsilon \to 0} M_\varepsilon = M_0.$$ In particular we must have $$\label{e3} \lim\limits_{\varepsilon \to 0} X_\varepsilon = \lim\limits_{\varepsilon \to 0} Y_\varepsilon =: Z_0$$ where $u_2(Z_0)-u_1(Z_0)=M_0$. Moreover, one observes that $$M_0>0\geq \sup_{\partial\Omega} (u_2-u_1),$$ hence $X_\varepsilon \in \Omega'$ for some interior domain $\Omega' \Subset \Omega$ and $\varepsilon>0$ sufficiently small. Therefore, by [@UG Theorem 3.2] there exist $\mathscr{M},\mathscr{N} \in \mathscr{S}_n$ with $$\label{e1} \left(\frac{X_\varepsilon-Y_\varepsilon}{\varepsilon}, \mathscr{M} \right) \in \overline{J}^{2,+}_{\Omega} u_2(X_\varepsilon) \quad \mbox{and} \quad \left(\frac{Y_\varepsilon-X_\varepsilon}{\varepsilon}, \mathscr{N} \right) \in \overline{J}^{2,-}_{\Omega} u_1(Y_\varepsilon)$$ such that, $$\label{e2} -\frac{3}{\varepsilon} \left( \begin{array}{cc} I & 0 \\ 0 & I \\ \end{array} \right) \leq \left( \begin{array}{cc} \mathscr{M} & 0 \\ 0 & \mathscr{N} \\ \end{array} \right) \leq \frac{3}{\varepsilon} \left( \begin{array}{cc} I & -I \\ -I & I \\ \end{array} \right).$$ In particular, $\mathscr{M} \leq \mathscr{N}$. By and , we obtain $$\begin{array}{lll} (u_2(X_\varepsilon)^{+})^\gamma &\leq& \displaystyle \mathscr{M} \left(\frac{X_\varepsilon-Y_\varepsilon}{\varepsilon} \right) \cdot \left(\frac{X_\varepsilon-Y_\varepsilon}{\varepsilon} \right) \\ &\leq& \displaystyle \mathscr{N} \left(\frac{Y_\varepsilon-X_\varepsilon}{\varepsilon} \right) \cdot \left(\frac{Y_\varepsilon-X_\varepsilon}{\varepsilon} \right) \\ &\leq& \displaystyle (u_1(Y_\varepsilon)^{+})^\gamma. \end{array}$$ Therefore, $$\left(M_\varepsilon+u_1(Y_\varepsilon)+(2\varepsilon)^{-1} |X_\varepsilon-Y_\varepsilon|^2 \right)^{+} \leq u_1(Y_\varepsilon)^{+}.$$ By and and letting $\varepsilon \to 0$ in the estimate above gives $$(M_0+u_1(Z_0))^+ \leq u_1(Z_0)^+$$ which drives us to a contradiction since $u_1 \geq 0$ and $M_0>0$, by assumption. Geometric regularity estimates ============================== As previously mentioned, viscosity solutions to $$\label{localeq} \mathcal{L}^{\,\gamma}_\infty \, u = 0 \quad \mbox{in} \quad \Omega,$$ for $0 \leq \gamma < 3$, are locally Lipschitz continuous. This is the optimal regularity estimated available in the literature – there is hope to show $C^{1,\alpha}$ estimates for some $0<\alpha\le 1/3$, but certainly not beyond that. Surprisingly, in this Section we show a sharp, improved regularity estimate for $u$ along its free plateaus boundary $\partial \{u>0\} \cap \Omega$. The proof is based on a flatness improvement argument inspired by [@T; @T1]; see also [@T2] for improved estimates that hold solely along [*nonphysical*]{} free interfaces. Next Lemma provides a universal way to flatten a solution near a plateaus boundary point. In the sequel we shall apply such a Lemma in dyadic balls as to obtain the aimed regularity estimate at free plateaus boundary points. \[comp\] Given $\mu>0$, there exists a number $\kappa_\mu>0$, depending only on $\mu$ and dimension such that if $v \in C(B_1)$ satisfies $$v(0)=0, \quad 0\leq v \leq 1 \quad \mbox{in} \quad B_1$$ and $$\Delta_{\,\infty} v - \kappa^4 (v^+)^\gamma = 0 \quad \mbox{in} \quad B_1,$$ for $0 < \kappa \leq \kappa_\mu$, then $$\sup\limits_{B_{1/2}} v \leq \mu.$$ Let us suppose, for the sake of contradiction, that there exists $\mu_0>0$ and sequences $\{v_\iota\}_{\iota \in \mathbb{N}}$, $\{\kappa_\iota\}_{\iota \in \mathbb{N}}$ satisfying $$0\leq v_\iota \leq 1, \quad v_\iota(0)=0$$ and $$\Delta_{\,\infty} v_\iota - \kappa_\iota^4 (v_\iota^+)^\gamma=0 \quad \mbox{for} \quad \kappa_\iota = \mbox{o}(1),$$ such that, $$\label{supcomp} \sup\limits_{B_{1/2}} v_\iota > \mu_0.$$ By Lipschitz estimates, the sequence $\{v_\iota\}_{\iota \in \mathbb{N}}$ is pre-compact in the $C^{0,1}(B_{1/2})$ topology. Up to a subsequence, $v_\iota \to v_\infty$ locally uniform in $B_{2/3}$. Moreover, we have $v_\infty(0)=0$, $0\leq v_\infty \leq 1$ and $$\Delta_{\,\infty} v_\infty = 0 \quad \mbox{in} \quad B_1.$$ Therefore, by the maximum principle for infinity harmonic functions, we obtain $v_\infty \equiv 0$. This give us a contradiction to , if we choose $\iota \gg 1$. \[thmreg\] Let $u$ be a viscosity solution to equation and $X_0 \in \partial\{u>0\} \cap \Omega$. There exists a positive constant $C>0$ depending on, $\|u\|_{L^{\infty}(\Omega)}$, $(3-\gamma)$ and $\dist (X_0, \partial \Omega)$, such that $$\label{conclusion thmreg} u(X) \leq C \cdot |X-X_0|^{\frac{4}{3-\gamma}}$$ for $X \in \{u>0\}$ near $X_0$. We assume, with no loss of generality, that $X_0=0$ and $B_{1} \Subset\Omega$. Let us define $$\omega_1(X):= \tau \,u\left( \rho X \right) \quad \mbox{in} \quad B_{1},$$ for $\tau>0$ and $\rho>0$, constants to be determined universally. From the equation satisfied by $u$, we easily verify that $\omega_1$ satisfies $$\label{eqcomp} \Delta_\infty\, \omega_1 - \tau^{3-\gamma}\rho^4(\omega_1^+)^{\gamma}=0,$$ in the viscosity sense. If $\kappa_\star>0$ is the universal constant granted by previous Lemma \[comp\] when one takes $\mu=2^{-\frac{4}{3-\gamma}}$, we make the following choices in the definition of $\omega_1$: $$\tau := \|u\|_{L^\infty(\Omega)}^{-1} \quad \mbox{and} \quad \rho := \kappa_\mu \cdot \tau^{-\frac{3-\gamma}{4}}.$$ With such a (lucky) selection, $\omega_1$ fits into the framework of Lemma $\ref{comp}$, which ensures that $$\sup_{B_{1/2}} \omega_1 \leq 2^{-\frac{4}{3-\gamma}}.$$ In the sequel, we set $$\omega_{\,2}(X):= 2^{\frac{4}{3-\gamma}} \,\omega_1 \left( 2^{-1}X \right) \quad \mbox{in} \quad B_1.$$ We note that $\omega_2$ satisfies $\omega_2(0) = 0$, $0 \le \omega_2 \le 1$ and $$\Delta_\infty\, \omega_2 - \kappa_\star^4(\omega_2^+)^{\gamma}=0.$$ That is, we can apply Lemma \[comp\] to $\omega_2$ as well, yielding, after rescaling, $$\sup_{B_{1/4}} \omega_{1} \leq 2^{-2 \cdot \frac{4}{3-\gamma}}.$$ Now, we argue by finite induction. For each $k \in \mathbb{N}$, we define $$\omega_{\,k}(X):= 2^{\frac{4}{3-\gamma}} \,\omega_{\,k-1} \left( 2^{-1}X \right).$$ By the same reasoning employed above, we verify that $\omega_{\,k}(X)$ fits into the hypotheses of Lemma \[comp\], which gives after rescaling $$\label{sup} \sup_{B_{2^{-k}}} \omega_1 \leq 2^{-k \,\frac{4}{3-\gamma}}.$$ Finally, fixed a radius $0 < r \leq \dfrac{\rho}{2}$, we choose $k \in \mathbb{N}$ such that, $$2^{-(k+1)} < \frac{r}{\rho} \leq 2^{-k}.$$ Therefore, we estimate $$\sup\limits_{B_{r}} u \leq \sup\limits_{B_{\rho\, 2^{-k}}} u = \tau^{-1}\sup\limits_{B_{2^{-k}}} \omega_1,$$ yielding, by , $$\begin{array}{ccl} \sup\limits_{B_{r}} u & \leq & \tau^{-1} \cdot 2^{-k \, \frac{4}{3-\gamma}} \\ & \leq & \left(2^{\frac{4}{3-\gamma}}\tau^{-1} \right) \cdot \, 2^{-(k+1) \, \frac{4}{3-\gamma}} \\ & \leq & \left((\rho\tau)^{-1}2^{\frac{4}{3-\gamma}}\right) \cdot r^{\;\frac{4}{3-\gamma}}. \\ \end{array}$$ This concludes the proof of Theorem \[thmreg\]. (0,-2.7625)(15.106281,2.7625) (0.6969061,-2.1235113)[.]{} (2.9569058,1.0764886)[.]{} (10.783281,2.7625)(0.003281,-2.7625) (0.8141117,-1.6875)[$\{u > 0\}$]{} (5.5941114,2.0125)[$u$]{} (10.765,-1.3575)(0.0,-1.3425) (9.734114,-1.7075)[$\{u = 0\}$]{} (2.5317678,2.7225)(2.5117679,2.7025)(1.1117679,0.2225)(3.5117679,2.6025)(1.891768,-0.0375)(4.331768,2.6225)(2.7747452,-0.16855443)(4.951768,2.1225)(3.611768,-0.3175)(5.431768,1.6025)(4.311768,-0.5375)(5.951768,0.9625)(5.191768,-0.6975)(6.391768,0.4225) (13.168781,2.0875)[Zoom +]{} (7.741406,-1.3325) (13.583,2.2275)(11.123,0.3675) (11.123,0.7275)(13.563,0.7275) (11.143,2.2275)(11.443,1.3875)(12.703,0.7275)(13.203,0.7675) (12.685813,1.6325)[$\sim d^{\frac{4}{3-\gamma}}$]{} (6.371768,0.4225)(5.851768,-0.8175)(6.691768,-0.0175)(6.391768,-0.9775)(7.031768,-0.4375)(6.791768,-1.0575)(7.211768,-0.6975)(7.091768,-1.1175)(7.371768,-0.8575)(7.331768,-1.1575)(7.531768,-1.0175)(7.531768,-1.2375)(7.6517677,-1.1775)(7.711768,-1.3175) (7.971768,-1.1725)(7.503,-1.4975) (7.531768,-1.1975)(11.111768,2.2025) (7.531768,-1.4775)(11.131768,0.3825) (7.931768,-1.2175)(11.111768,0.7625) (7.931768,-1.4775)(13.571768,0.3825) (13.211768,0.7625) \[rmk1\] A careful scrutiny of the proof of Theorem \[regthm\] revels that the same regularity estimate holds for equations with non constant, bounded Thiele modulus: $$\Delta_\infty u = \lambda(X) \cdot u^\mu.$$ In this case, the constant $C>0$ appearing in , which bounds the $C^{\frac{4}{3-\mu}}$-growth estimate of $u$ away from the touching ground, depends only on $\|u\|_{L^{\infty}(\Omega)}$, $(3-\gamma)$, $\dist (X_0, \partial \Omega)$ and $\|\lambda\|_{L^{\infty}(\Omega)}$. This remark will be used in the future. We conclude this Section with an asymptotic Liouville type classification result. A stronger, quantitative version of this Theorem will be delivered later. \[bernstein\] Let $u$ be a viscosity solution to $$\mathcal{L}^{\,\gamma}_{\, \infty} \, u = 0 \quad \mbox{in} \quad \mathbb{R}^n$$ with $u(0)=0\,$. If $u(X)=\mathrm{o}(|X|^{\frac{4}{3-\gamma}})$ as $|X|\to \infty$, then $u\equiv 0$. For each positive number $\kappa \gg 1$, let us define $$u_{\kappa}(X):= u(\kappa\, X)\,\kappa^{-\frac{4}{3-\gamma}}.$$ It is easy to check that $$\mathcal{L}^{\,\gamma}_{\, \infty} \, u_\kappa = 0 \mbox{ in } B_1,$$ and $u_\kappa(0)=0$. Moreover, we note that $$\|u_\kappa\|_{L^\infty(B_1)} = \mbox{o}\,(1).$$ In fact, for each $\kappa \in \mathbb{N}$, let $X_\kappa \in \mathbb{R}^n$ be such that $u_\kappa(X_\kappa)= \sup\limits_{B_1}u_\kappa$. If $\lim\limits_{\kappa \to \infty} \kappa X_\kappa = \infty$, by the above assumption, we obtain $$u_\kappa(X_\kappa) \leq |\kappa X_\kappa|^{-\frac{4}{3-\gamma}}u(\kappa X_\kappa) \to 0, \quad \mbox{ as }\kappa \to \infty.$$ If the sequence $\{\kappa X_\kappa\}$ remains bounded, we easily obtain the limit above for $u_\kappa(X_\kappa)$. Applying Theorem \[thmreg\] we obtain $$\label{odeum} u_\kappa(X) \leq \mbox{o}\,(1)\cdot |X|^{\frac{4}{3-\gamma}} \quad \mbox{in} \quad B_{1/2}.$$ Now, if we assume that there is a $Z_0 \in \mathbb{R}^n$ such that $u(Z_0)>0$, we obtain from , $$\label{Beq05.5} \sup\limits_{B_{1/2}} \dfrac{u_\kappa(X)}{|X|^{\frac{4}{3-\gamma}}} \le \dfrac{{u(Z_0)}}{100 |Z_0|^{\frac{4}{3-\gamma}}},$$ provided $\kappa \gg 1$. We now estimate, for $\kappa \gg 2|Z_0|$, $$\label{Beq06} \dfrac{u(Z_0)}{|Z_0|^{\frac{4}{3-\gamma}}} \le \sup\limits_{B_{\kappa/2}} \dfrac{u(X)}{|X|^{\frac{4}{3-\gamma}}} \\ \le \sup\limits_{B_{1/2}} \dfrac{u_\kappa(X)}{|X|^{\frac{4}{3-\gamma}}} \\ \le \dfrac{{u(Z_0)}}{100 |Z_0|^{\frac{4}{3-\gamma}}},$$ which finally drives us to a contradiction, completing the proof of Theorem \[bernstein\]. Radial Analysis =============== In this intermediary section, we make a short pause as to analyze the radial boundary value problem $$\label{rad eq} \left \{ \begin{array}{rlll} \Delta_\infty u &=& \lambda (u^{+})^\gamma &\text{ in } B_R(X_0) \\ u &=& c &\text{ on } \partial B_R(X_0), \end{array} \right.$$ where $0 < c, \lambda < \infty$ are constants and $X_0 \in \mathbb{R}^n$. Herein we consider an arbitrary Thiele modulus $\lambda>0$, as to amplify the range of our analysis. Initially we observe that, by uniqueness and $\mathcal{O}(n)$ invariance of the infinity laplacian, it is plain that the solution of such a boundary value problem is radially symmetric. Indeed, for any $O \in \mathcal{O}(n)$, the function $v(X - X_0) := u(O(X-X_0))$ solves the same boundary value problem, hence, by uniqueness, $v(X) = u(X)$. Since $O \in \mathcal{O}(n)$ was taken arbitrary, it does follow that $u$ is radially symmetric. We then consider the following ODE related to , $$\label{rad edo} h''(h')^2 = \lambda (h^{+})^\gamma \quad \text{ in }\, (0,T)$$ satisfying the initial conditions: $h(0)=0$ and $h(T)=c$. Solving we obtain the solution $\,h(s)=\tau(\lambda,\gamma) \cdot s^{\,\frac{4}{3-\gamma}}$, where $$\label{tau} \tau(\lambda, \gamma)= \sqrt[3-\gamma]{\lambda \cdot \frac{(3-\gamma)^4}{64(1+\gamma)}} \quad \mbox{and} \quad \left( \frac{c}{\tau(\lambda,\gamma)} \right)^{\frac{3-\gamma}{4}}=:T.$$ Fixed $X_0 \in \mathbb{R}^n$ and $0<r<R$, let us assume the [*dead-core compatibility condition*]{} $$\label{DC cond1} R > T.$$ Define the following radially symmetric function $u : B_R(X_0) \setminus B_r(X_0) \to \mathbb{R}$ given by $$u(X):=h\left(|X-X_0|-r\, \right),$$ where $r=R-T$. One easily verifies that $u$ solves pointwise the equation $$\nonumber \Delta_\infty u = \lambda (u^{+})^\gamma \quad \text{ in } \quad B_R(X_0) \setminus B_r(X_0).$$ The boundary conditions: $u \equiv 0$ on $\partial B_r$ and $u \equiv c$ on $\partial B_R$ are also satisfied. Moreover, by the construction, for each $Z\in \partial B_r(X_0)$, we obtain $$\lim_{X \to Z} \nabla u(X) = h'(0^+).\frac{Z}{|Z|}=0.$$ Thus, extending $u\equiv 0$ in $B_r(X_0)$, we obtain a function in $B_R(X_0)$ satisfying $$\nonumber \Delta_\infty u = \lambda (u^{+})^\gamma \quad \text{ in } B_R(X_0).$$ We concluded that the function $$u(X):= \tau(\lambda,\gamma) \left(|X-X_0|-R + \left( \frac{c}{\tau(\lambda,\gamma)} \right)^{\frac{3-\gamma}{4}} \right)_+^{\frac{4}{3-\gamma}}$$ is the solution to . Its plateaus is precisely $B_r(X_0)$, where $$\label{setup} 0<r:=R-\left( \frac{c}{\tau(\lambda,\gamma)} \right)^{\frac{3-\gamma}{4}}.$$ Let us now deliver few elementary conclusions. Given a positive boundary data $c$, a radius $R>0$, a Thiele modulus $\lambda$, and an exponent $0\le \gamma < 3$, then 1. If the Thiele modulus $\lambda$ is sufficiently large (with bounds easily computable), then the radial boundary problem presents plateaus irrespective of $0\le \gamma <3$. 2. As one should expect, solution converges locally uniform in to zero as $\lambda$ goes to infinity. 3. On the other hand, fixed any small Thiele modulus $\lambda_0>0$, the boundary value problem has plateaus provided $\gamma$ is sufficiently close to 3; and indeed, solutions to go to zero as $\gamma \nearrow 3$. (0,-4.079871)(11.164836,3.9580135) (1.7024888,0.10496419)(0.0,-3.935036)(9.74,-3.995036)(11.08,0.08496411) (6.573795,-3.0703874)[.]{} (4.3137946,0.12961239)[.]{} (-0.045747217,-0.14301053)[(5.801751,-1.9347274)(3.2207081,0.9394911)]{} (5.744052,-1.9043761) (6.289036,-2.464376)[$R$]{} (-0.04506483,-0.14205708)[(5.76326,-1.906458)(1.3386139,0.359646)]{} (7.092489,-1.9043763)(8.12,-1.4750359)(8.76,1.6649641)(9.14,3.4649642) (0.08498112,-0.1404915)[(5.7645197,3.3909202)(3.3611968,0.5217588)]{} (5.762489,-1.935036)(7.18,-2.735036) (8.958896,-0.77437603)[$\sim (|X|-r)_+^{\frac{4}{3-\gamma}}$]{} (5.82,-1.8750359)(6.86,-1.6950359) (5.900911,-1.709376)[*$r$*]{} (1.94,3.344964)(2.1,-2.155036) (1.6690358,0.615624)[$c$]{} (4.44,-1.9643761)(3.2724888,-1.3150358)(2.8124888,1.5249641)(2.42,3.344964) (1.8434268,-3.594376)[Plateaus]{} (1.84,-3.375036)(1.84,-3.1150358)(4.76,-1.3750359)(5.08,-1.8750359) Now, if $v$ is an arbitrary solution to $$\Delta_\infty v = \lambda v_{+}^{\gamma}, \quad \text{ in } \Omega \subset \mathbb{R}^n,$$ and $X_0 \in \Omega$ is an interior point, define $\mathfrak{s} \colon (0, \dist(X_0, \partial \Omega)) \to \mathbb{R}_{+}$ by $$\mathfrak{s}(R) := \sup\limits_{B_R(X_0)} v.$$ If for some $0<R < \dist(X_0, \partial \Omega)$ , we have $$\mathfrak{s}(R) < \tau(\lambda, \gamma)R^{\frac{4}{3-\gamma}},$$ then $X_0$ is a plateaus point. In particular, we can improve Theorem \[bernstein\] to the following quantitative version: \[bernstein1\] Let $u$ be a viscosity solution to $$\label{eq thm B1} \Delta_\infty u = \lambda u_{+}^\gamma \quad \mbox{in} \quad \mathbb{R}^n.$$ If $$\label{cond thm B1} \limsup\limits_{|X| \to \infty} \dfrac{u(X)}{|X|^{\frac{4}{3-\gamma}}} < \sqrt[3-\gamma]{\lambda \cdot \frac{(3-\gamma)^4}{64(1+\gamma)}},$$ then $u\equiv 0$. Fixed $R>0$, let us consider $v \colon \overline{B_R} \to \mathbb{R}$, the solution to the boundary value problem $$\left \{ \begin{array}{rlll} \Delta_\infty v &=& \lambda (v^{+})^\gamma &\text{ in } B_R \\ v &=& \sup\limits_{\partial B_R}u &\text{ on } \partial B_R. \end{array} \right.$$ By comparison principle, Lemma , $u \leq v$ in $B_R$. It follows by hypothesis that, taking $R\gg 1$ sufficiently large, $$\label{lim rad} \sup\limits_{\partial B_R}\frac{u(X)}{R^{\frac{4}{3-\gamma}}} \leq \theta \cdot\tau(\lambda,\gamma)$$ for some $\theta<1$. For $R\gg 1$, the solution $v = v_R$ is given by $$\label{rad eq3} v(X)=\tau(\lambda,\gamma) \left(|X|-R + \left[\frac{\sup\limits_{\partial B_R}u}{\tau(\lambda,\gamma)}\right]^{\frac{3-\gamma}{4}} \right)_+^{\frac{4}{3-\gamma}}.$$ Finally, combining and , we get $$u(X) \leq \tau(\lambda,\gamma) \left(|X|- (1-\theta^{\frac{3-\gamma}{4}})R \right)_+^{\frac{4}{3-\gamma}},$$ Letting $R \to \infty$ we conclude the proof of the Theorem. We conclude by pointing out that Theorem \[bernstein1\], as stated, is sharp in the sense that one cannot remove the strict inequality in . Indeed, the function $$h(X) := \sqrt[3-\gamma]{\lambda \cdot \frac{(3-\gamma)^4}{64(1+\gamma)}} |X|^{\frac{4}{3-\gamma}},$$ solves in $\mathbb{R}^n$ and it clearly attains equality in . Minimal growth rate and measure estimates ========================================= In this section we show that the regularity estimate established in Theorem \[thmreg\] is indeed sharp. This is done by establishing a competing inequality which controls the minimal growth rate of the solution away from its free boundary. \[nondegthm\] Let $u \in C(\Omega)$ be a nonnegative viscosity solution to $$\label{lgamma} \mathcal{L}^{\,\gamma}_{\, \infty} \, u = 0 \quad \mbox{in} \quad \Omega$$ and $X_{0} \in \overline{\left\lbrace u> 0 \right\rbrace} \cap \Omega$. There exists a universal constant $c_0>0$, such that $$\begin{aligned} \label{nondeg est} \sup_{B_{r}\left( X_{0} \right)}u \geq c_0 \cdot r^{\frac{4}{3-\gamma}},\end{aligned}$$ for all $0<r <\dist (X_0, \partial \Omega)$. By continuity, it suffices to prove for points within the set $ \left\lbrace u> 0 \right\rbrace \cap \Omega'$. Initially define $$\begin{aligned} \psi \left( X \right) := c \cdot \vert X - X_{0}\vert^{\alpha},\end{aligned}$$ for $\alpha:= \frac{4}{3-\gamma}$ and $c>0$ a constant that will be fixed [*a posteriori*]{}. By direct computation, $$D \psi (X) = c \alpha |X-X_0|^{\alpha-1}\cdot \frac{X-X_0}{|X-X_0|}.$$ Continuing, direct computations further yield $$\begin{aligned} D^2 \psi \left( X \right) = c \alpha \left[(\alpha-1)|X-X_0|^{\alpha-2} \cdot \frac{(X-X_0)\otimes (X-X_0)}{|X-X_0|^2}\right. \\ \left. + |X-X_0|^{\alpha-2}\cdot \left( \text{Id}_{n \times n}-\frac{(X-X_0)\otimes(X-X_0)}{|X-X_0|^2}\right)\right].\end{aligned}$$ Therefore, we conclude $$\begin{aligned} \langle D^2\psi \cdot D\psi,D\psi \rangle \left( X \right) = \left( c \alpha \right)^3(\alpha-1) |X-X_0|^{2(\alpha-1)+(\alpha-2)}\end{aligned}$$ and hence, by selecting (and fixing) the constant $c$ within the range $$0 < c < \sqrt[3 - \gamma]{\dfrac{\left( 3 - \gamma \right)^{4} }{64\left( 1+ \gamma \right)}},$$ we reach $$\mathcal{L}^{\,\gamma}_{\, \infty} \psi < 0 = \mathcal{L}^{\,\gamma}_{\, \infty}\; u .$$ Now, for any ball $B_r(X_0) \subset \Omega$, there must exist a point $Z \in \partial B_{r}\left( X_{0}\right)$ such that $\psi \left( Z \right) < u\left( Z \right)$; otherwise, by comparison principle, Lemma \[comparison principle\], $\psi \ge u$ in the whole ball $ B_{r}\left( X_{0} \right)$. However, $0 = \psi \left( X_{0} \right) < u\left( X_{0} \right)$. In conclusion, we can estimate $$\begin{aligned} \sup_{B_{r}\left( X_{0} \right)}u \geq u\left( Y_{r} \right) \geq \psi \left( Y_{r} \right) = c \cdot r^{\frac{4}{3-\gamma}}\end{aligned}$$ and the Theorem is proven. \[H est\] Given a subdomain $\Omega' \Subset \Omega$, there exists a constant $\iota >0$ depending on $\|u\|_{L^\infty(\Omega)}, \gamma$ and $\Omega'$ such that for $u \in C(\Omega)$ a nonnegative, bounded viscosity solution to in $\Omega$, there holds $$\begin{aligned} \frac{\mathscr{L}^{n}\left( B_{r}\left( X_{0}\right) \cap \left\lbrace u> 0 \right\rbrace \right)}{r^{\,n}} \geq \iota,\end{aligned}$$ for any $X_0 \in \partial\{u>0\} \cap \Omega'$ and $0<r \ll 1$. In addition, for a universal constant $0< \sigma_{0} \le 1$, depending only on dimension and $\gamma$, the $(n-\sigma_0)$-Hausdorff measure of $\partial \{u>0\}$ is locally finite. In view of Theorem \[nondegthm\], for some $r>0$ fixed, it is possible to select a point $Y_0$ such that, $$\label{dens} u(Y_0)= \sup\limits_{B_r(X_0)} u \geq c_0 \cdot r^{\frac{4}{3-\gamma}}.$$ To conclude, we claim that for some $\delta>0$, chosen universally small, the following inclusion $$\label{inclusion} B_{\delta \cdot r}(Y_0) \subset \{u>0\}$$ holds. Indeed, by Theorem \[thmreg\], for $Z \in \partial\{u>0\}$, we reach $$u(Y_0) \leq C \cdot |Y_0 - Z|^{\frac{4}{3-\gamma}}.$$ Therefore, by and the inequality above, we find $$c_0 \cdot r^{\frac{4}{3-\gamma}} \leq C \cdot |Y_0-Z|^{\frac{4}{3-\gamma}}$$ and so, $$\left(\dfrac{c_0}{C}\right)^{\frac{3-\gamma}{4}}\cdot r \leq |\,Y_0-Z|.$$ Hence, taking $\delta>0$ sufficiently small, the inclusion claimed in is verified. We conclude with the analysis of the Hausdorff dimension of the free boundary. Let $X_{0} \in \partial \left\lbrace u > 0 \right\rbrace$. From the above reasoning, we can always select $$\begin{aligned} X'_{0} = \sigma Y_{r} + \left( 1 - \sigma \right)X_{0}.\end{aligned}$$ with $0<1-\sigma \ll 1$, such that $$\begin{aligned} B_{\sigma\frac{ r}{2}}\left( X'_{0} \right) \subset B_{\sigma} \left( Y_{r} \right) \cap B_{r}\left( X_{0}\right) \subset B_{r}\left( X_{0}\right) \setminus \partial \left\lbrace u > 0\right\rbrace.\end{aligned}$$ Hence the set $\partial \left\lbrace u> 0 \right\rbrace \cap \Omega'$ is $\left( \sigma / 2 \right)$-porous and therefore, by a classical result, see for instance [@KR Theorem 2.1], the Hausdorff dimension of $\partial \left\lbrace u> 0 \right\rbrace \cap \Omega'$ is at most $n - C\sigma^{n}$ for some dimensional constant $C >0$. The Hausdorff dimension estimate provided by Corollary \[H est\] assures in particular that the $\Leb$-Lesbegue measure of the plateaus boundary is zero, but no quantitative information is given on its precise Hausdorff dimension. We believe $\sigma_0 = 1$, and leave this is an open problem. The critical equation $\mathcal{L}_\infty^3$ ============================================ In this Section we turn our attention to the critical equation obtained as $\gamma \nearrow 3$, that is, $$\label{localeq3} \mathcal{L}^{\,3}_{\, \infty} \, u :=\Delta_{\, \infty}\, u - u^{3} = 0 \quad \mbox{in} \quad \Omega.$$ Such an operator is regarded to be critical, as all the estimates established so far deteriores when one let $\gamma$ converge to $3$. Certainly, one can treat equation as $$\Delta_\infty = (u^\delta) \cdot u^{3-\delta},$$ for any $\delta > 0$. In particular, it follows from Theorem \[thmreg\] that if $u$ vanishes at an interior point $\xi \in \Omega$, then $D^n u (\xi) = 0$, for all $n\in \mathbb{N}$. That is, any zero is an infinite order zero. Under the (very strong) assumption that $u$ is a real analytic function, one could conclude that $u \equiv 0$. As mentioned before, Lipschitz regularity is the best local estimate available in the literature for such a solution. Even in the best scenario possible, one could not expect estimates beyond $C^{1,\alpha}$. Thus assuming $u$ is real analytic would simply be artificial. Nonetheless, by means of geometric arguments, which explores the scalar invariance of the operator $\mathcal{L}_\infty^3$, we shall prove that indeed a positive solution to is prevented to vanish at an interior point. \[strongmp\] Let $u \in C(\Omega)$ be a nonnegative viscosity solution to . If there exists a point $X \in \Omega$ such that $u(X)=0$, then $u \equiv 0$ in $\Omega$. Let us suppose, for the purpose of contradiction, that the thesis of the theorem fails to hold. With no loss generality we assume $u(0) > 0$ and $$d:=\dist(0,\{u=0\}) < \dfrac{1}{10}\dist(0,\partial \Omega).$$ By comparison principle $u$ is locally bounded. We now build up the following auxiliary barrier function $$\nonumber \Phi_{\lambda}(|X|)= \left\{ \begin{array}{ccc} e^{-\lambda(d/2)^2}-\kappa_0 & \mbox{in} & B_{d/2}; \\ e^{-\lambda|X|^2}-\kappa_0 & \mbox{in} & B_d \setminus B_{d/2}; \\ 0 & \mbox{in} & \mathbb{R}^n \setminus B_d, \end{array} \right.$$ for $\kappa_0$ such that $\Phi_\lambda(d^+)=0$. By construction, one easily verifies that $$\label{gradphi} \inf\limits_{B_d \setminus B_{d/2}}|\nabla \Phi_\lambda| \geq \beta_0$$ for some $\beta_0>0$, easily computable if one desires. Moreover, direct computation yields $$\mathcal{L}^{\,3}_{\, \infty} \, \Phi_\lambda \geq 0 \quad \mbox{in} \quad B_d \setminus B_{d/2},$$ provided $\lambda \gg 1$. The important observation is that the operator $\mathcal{L}_{\infty}^3$ is invariant under scalar multiplication, that is, for any number $\theta>0$ $$\mathcal{L}^{\,3}_{\, \infty}(\theta \cdot \Phi_\lambda) = \mathcal{L}^{\,3}_{\, \infty}\, \Phi_\lambda \geq 0 = \mathcal{L}^{\,3}_{\, \infty}\, u \quad \mbox{in} \quad B_d \setminus B_{d/2}.$$ In addition, taking $0<\theta \ll 1$ we get $$\theta \cdot \Phi_\lambda \leq u \quad \mbox{in} \quad \partial B_d \cup \partial B_{d/2}.$$ Therefore, by comparison principle, Lemma \[comparison principle\], $$\label{cprin} \theta \cdot \Phi_\lambda \leq u \quad \mbox{in} \quad B_d \setminus B_{d/2}.$$ On the other hand, equation can be written as $$\Delta_\infty u = \left [u(X) \right ]\cdot u^{2} = \lambda(X) u^2,$$ for a bounded Thiele modulus $\lambda(X) = u(X)$. Hence, in view of Remark \[rmk1\], we obtain $$\sup\limits_{B_r(Y_0)} u \leq C \cdot r^{4},$$ for $Y_0 \in \partial B_d \cap \partial\{u>0\}$. Now, we choose $0<r_0 \ll 1$ such that $$C \cdot r_0^{4} \le \dfrac{1}{4} \theta \beta_0 \cdot r^{}_0.$$ Finally, by and , we reach $$\begin{aligned} 0< \theta\beta_0\cdot r_0 &\leq& \sup\limits_{B_{r_0}(Y_0)} \theta \cdot |\Phi_\lambda(X)-\Phi_\lambda(Y_0)| \\ &\leq& \sup\limits_{B_{r_0}(Y_0)} \theta\cdot \Phi_\lambda \\ &\leq& \sup\limits_{B_{r_0}(Y_0)} u \\ &\le & C \cdot r_0^{4} \\ &\le &\dfrac{1}{4} \theta \beta_0 \cdot r^{}_0,\end{aligned}$$ which gives us a contradiction. The proof of Theorem \[strongmp\] is complete. (0,-3.2182996)(11.421562,3.155876) (1.77,1.0558761)(0.0,-3.113464)(9.660778,-3.133464)(11.1,1.1265359) (8.164119,-2.2394755)[.]{} (5.9041185,0.9605243)[.]{} (-0.02513735,-0.17827968)[(7.2485213,-1.1129501)(2.5623963,0.8324325)]{} (7.251563,-1.0734642) (8.216547,-1.3934642)[$d$]{} (9.79,-1.024124)(9.05,-0.8641239)(8.67,-0.5841239)(8.535682,-0.10412391) (-0.46670577,-0.82827026)[(3.841094,-2.7034643)[$\{u > 0 \}$]{}]{} (-0.42838863,-0.25659156)[(1.1610937,-2.4434643)[$\{u =0\}$]{}]{} (4.3214064,-1.1384641)[$Y_0$]{} (10.481093,0.7765358)[$\Omega$]{} (7.0741405,-1.2484641)[$0$]{} (0.06,-3.1134644)(2.6,-2.7334645)(4.44,-2.073464)(4.66,-1.2534641)(4.88,-0.4334642)(3.4,-0.093464166)(3.52,0.22653583)(3.64,0.54653585)(5.14,1.0465358)(5.9,1.0265357) (-0.024212413,-0.17881462)[(7.2707715,-1.0755461)(1.3386142,0.359646)]{} (4.7,-1.1334641) (-0.026648972,-0.11256185)[(4.5711675,-1.1416575)(1.1793989,0.34184355)]{} (4.72,-1.0934643)(5.87,-0.50412387)(6.93,1.135876)(7.73,3.135876) (-0.0026213082,-0.17756745)[(7.2307715,-0.19554606)(1.3386142,0.359646)]{} (7.27,-1.104124)(8.53,-1.8041239) (4.6556816,-1.144124)(5.395682,-0.9841239)(5.775682,-0.7041239)(5.91,-0.2241239) (5.1757817,-0.3434641)[$\sim r_0^{4/3}$]{} (6.9782815,0.7865359)[$u$]{} (9.493594,-0.55946407)[$\theta\cdot \Phi_\lambda$]{} We note that in fact the proof of Theorem \[strongmp\] yields a Hopf-type lemma for the critical equation . [99]{} G. Aronsson *Extension of functions satisfying Lipschitz conditions.* Ark. Mat. 6, (1967) 551–561. G. Aronsson *On the partial differential equation $u_{x^2} u_{xx}+2u_xu_yu_{xy}+u_{y^2} u_yy=0$*. Ark. Mat. 7, (1968) 395–425. G. Aronsson, M. Crandall, P. Juutinen *A tour of the theory of absolute minimizing functions* (2004) Bull. Amer. Math. Soc., 41 (4), pp. 439-505. C. Bjorland; L. Caffarelli and A. Figalli [*Nonlocal tug-of-war and the infinity fractional Laplacian.*]{} Comm. Pure Appl. Math. 65 (2012), no. 3, 337–380. V. Caselles, J. Morel, and C. Sbert, [*An axiomatic approach to image interpolation*]{} , IEEE Trans. Image Process. 7 (1998), no. 3, 376–386. M.G. Crandall, H. Ishii, P.L. Lions *User’s guide to viscosity solutions of second-order partial differential equations* (1992) Bull. Amer. Math. Soc., 27 , pp. 1–67. D. dos Prazeres and E. Teixeira *Cavity problems in discontinuous media.* to appear in Calc. Var. Partial Differential Equations. L.C. Evans and O. Savin, [*$C^{1, \alpha}$ regularity for infinity harmonic functions in two dimensions*]{}, Calc. Var. Partial Differential Equations [**32**]{} (2008), 325–347. L.C. Evans and C.K. Smart, [*Everywhere differentiability of infinity harmonic functions*]{}, Calc. Var. Partial Differential Equations [**42**]{} (2011), 289–299. P. Koskela and S. Rohde *Hausdorff dimension and mean porosity* (1997) Math. Ann., 309 (4), pp. 593-609. E. Lindgren [*On the regularity of solutions of the inhomogeneous infinity Laplace equation.*]{} Proc. Amer. Math. Soc. [**142**]{} (2014), no. 1, 277–288. G. Lu and P. Wang *Inhomogeneous infinity Laplace equation* (2008) Adv. Math., 217 (4), pp. 1838–1868. R. Jensen [*Uniqueness of Lipschitz extensions minimizing the sup-norm of the gradient*]{}, Arch. Rat. Mech. Anal., 123 (1993), 51–74. O. Savin, [*$C^1$ regularity for infinity harmonic functions in two dimensions*]{}, Arch. Ration. Mech. Anal. [**176**]{} (2005), 351–361. Y. Peres, O. Schramm, S. Sheffield, and D. Wilson, [*Tug-of-war and the infinity Laplacian*]{}, J.Amer.Math.Soc. 22 (2009), no. 1, 167–210. J. Rossi and E. Teixeira [*A limiting free boundary problem ruled by Aronsson’s equation.*]{} Trans. Amer. Math. Soc. [**364**]{} (2012), 703–719. J. Rossi, E. Teixeira and J.M. Urbano [*Optimal regularity at the free boundary for the infinity obstacle problem.*]{} Interfaces Free Bound. 7 (2015), 381–398. E. Teixeira [*Sharp regularity for general Poisson equations with borderline sources.*]{} J. Math. Pures Appl. (9) 99 (2013), no. 2, 150–164. E. Teixeira *Regularity for quasilinear equations on degenerate singular sets.* Math. Ann. 358 (2014), no 1, 241–256. E. Teixeira [*Regularity for the fully nonlinear dead-core problem.*]{} To appear in Math. Ann. DOI 10.1007/s00208-015-1247-3 <span style="font-variant:small-caps;">Damião J. Araújo</span> <span style="font-variant:small-caps;">Raimundo Leitão</span>\ University of Florida/UNILABUniversidade Federal do Ceará\ Department of Mathematics Department of Mathematics\ Gainvesville, FL-USA 32611-8105 Fortaleza, CE-Brazil 60455-760\ `[email protected]` `[email protected]` <span style="font-variant:small-caps;">Eduardo V. Teixeira</span>\ Universidade Federal do Ceará\ Department of Mathematics\ Fortaleza, CE-Brazil 60455-760\ `[email protected]`

ArXiv

--- abstract: 'We propose to use semiclassical methods to treat laser control problems of chemical reaction dynamics. Our basic strategy is as follows: Laser-driven chemical reactions are considered to consist of two processes. One is the wavepacket propagation on an adiabatic potential energy surface (PES), and the other is the electronic transition between PES’s. Because the latter process is mathematically equivalent to nonadiabatic transitions between Floquet (dressed) states, we can control such a process using the semiclassical Zhu-Nakamura theory for nonadiabatic transitions. For the former process, we incorporate semiclassical propagation methods such as the Herman-Kluk propagator into optimization procedures like optimal control theory. We show some numerical examples for our strategies. We also develop a semiclassical direct algorithm to treat the adiabatic propagation and nonadiabatic transitions as a whole.' author: - | Hiroshi FUJISAKI, Yoshiaki TERANISHI$^{\spadesuit}$, Alexey KONDORSKIY,\ and Hiroki NAKAMURA[^1]\ [*Department of Theoretical Studies, Institute for Molecular Science*]{}\ [*$^{\spadesuit}$Japan Atomic Energy Research Institute*]{} title: ' [**Semiclassical approaches to controlling chemical reaction dynamics**]{}' --- Introduction ============ Controlling molecular dynamics is one of the intriguing targets in chemical physics. Since the works of Brumer and Shapiro, Tannor and Rice, and others, laser control of molecular dynamics has been developed relatively well for simple molecular processes [@RZ00]. Theoretically, in order to obtain an optimal field producing a desired product, optimal control theory (OCT) is one of the most natural and flexible vehicles [@RZ00]. It assumes a certain functional which should be maximized, from which one can derive equations of motion which in turn determines the optimal field. The equations including Schrödinger equation must be solved iteratively in general, hence its numerical cost becomes huge for large molecules. Although there exist some calculations for reduced dimensionality models of large molecules, it can be very difficult to control even a triatomic molecule quantum mechanically within the present state of art. Some strategies to circumvent this difficulty are strongly desired. Our basic strategy is as follows: Laser-driven chemical reactions are considered to consist of two processes. One is the wavepacket propagation on an adiabatic potential energy surface (PES), and the other is the electronic transition between PES’s. Because the latter process is mathematically equivalent to nonadiabatic transitions between Floquet (dressed) states, we can control such a process using the semiclassical Zhu-Nakamura theory for nonadiabatic transitions [@ZTN01]. (It has been shown in Refs. [@ZTN01; @NTN00; @NTN02] that this strategy is very successful.) In Sec. \[sec:HI\], we shall show another example of our strategy. A new thing here is that we treat a real diatomic molecule, HI, with its real PES information. We shall use the idea of [*complete reflection*]{} to control HI photodissociation to two different channels. A complication arises due to existence of three electronically excited states [@FTN02]. For controlling the wavepacket dynamics on a PES, on the other hand, we need to employ a general optimization procedure like OCT. There have been many applications of OCT in the literature; however, such works are often limited to rather small systems with less than three degrees of freedom. This is because the conventional methods of wavepacket propagation using [*grids*]{} are limited to rather small systems. [*Semiclassical wavepacket propagation methods*]{} are believed to overcome this kind of difficulty with use of classical trajectories. Hence it is very natural to examine the possibility of combining the semiclassical wavepacket propagation methods with OCT. In Sec. \[sec:OCT\], we employ one of semiclassical wavepacket propagation methods, the initial value representation (IVR) method using the Herman-Kluk propagator [@MB00], and incorporate it into OCT. It is known that the IVR method can approximately reproduce the full quantum results for many molecular systems, and its computational scalability is much less than that of the corresponding quantum calculations for large molecules. In addition, we can understand the “physical aspects” of the optimal field because the IVR method represents the result as a sum of the quantities calculated from classical trajectories. The effectiveness of the IVR method has been mainly tested for single PES systems, i.e., systems without nonadiabatic transitions. (For exceptions, see Ref. [@SM97].) On the other hand, we have the semiclassical Zhu-Nakamura theory of nonadiabatic transition [@ZTN01], and it is shown that it can be used to control even wavepacket excitation between PES’s under some conditions [@NTN02]. In Sec. \[sec:semi\], we examine the possibility to combine this kind of semiclassical theory of nonadiabatic transitions with the semiclassical wavepacket propagation method. Controlling a molecule by complete reflection: Application to HI photodissociation {#sec:HI} ================================================================================== We shall apply the strategy using complete reflection [@NTN00] to HI photodissociation. The [*ab initio*]{} potential energy curve (PEC) information is reproduced in Fig. 2 of [@FTN02]. There are three electronically excited states, and their transition dipole moments depend on the internuclear distance significantly. When applying a stationary laser field, we can consider this system as a curve-crossing system: one curve from the electronically ground state crosses with the other three curves from the electronically excited states. At crossings nonadiabatic transitions occur and such a phenomenon can be accurately treated by the semiclassical ZN formula. If the energy of a vibrational state satisfies a resonant-like condition, then the phenomenon, complete reflection, occurs: such a wavepacket cannot escape from the upper adiabatic PEC to the lower adiabatic PEC. (The lifetime is infinite.) This strategy can be used to control photodissociation and some model calculations were done in [@NTN00]. To control HI photodissociation, we need to control three crossings [*at the same time*]{}. Of course, this is [*mathematically*]{} impossible because the complete reflection occurs at a certain energy point, but it might work [*in practice*]{} at some time interval. As shown in Fig. \[fig:flux\_time\](a), at a certain energy ($\hbar \omega \simeq 3.58$ eV) and up to a certain time interval (14 ps), the two fluxes ($J_2(t), J_4(t)$) corresponding to two electronically excited PEC’s connected to H+I channel nearly equal zero because this energy satisfies the complete reflection condition for [*both*]{} states to a good extent. In this case, we have to prepare the vibrationally excited state with $v=4$ ($v=5$ is better from experimental points of view [@Kawasaki], but the qualitative result is the same). In addition, the contribution from $v=0$ component, which might remain as a major component even after infrared excitation of the vibrational state, is very small as shown in Fig. \[fig:flux\_time\](b). This means that the H+I$^*$ photodissociation can be nearly completely achieved by using the vibrational state with $v=4$ and the stationary laser field with $\hbar \omega \simeq 3.58$ eV. (In contrast, it is easy to achieve H+I photodissociation because only single excited PEC is asymptotically connected to H+I$^*$ which should be blocked.) We have not included the effect of rotation and temperature, which degrade the quality of control, but the actual experimental condition is expected to be much similar to our calculation [@Kawasaki]. ![ [ Time variation of the time-integrated fluxes $J_i(t)$ ($i=2,3,4$ corresponds to three electronically excited states) for the initial vibrational state with (a) $v=4$ and (b) $v=0$. The laser frequency is $\hbar \omega \simeq 3.58$ eV. ]{} []{data-label="fig:flux_time"}](flux_v_4_E0_17.7_time.eps){width="1.0\linewidth"} ![ [ Time variation of the time-integrated fluxes $J_i(t)$ ($i=2,3,4$ corresponds to three electronically excited states) for the initial vibrational state with (a) $v=4$ and (b) $v=0$. The laser frequency is $\hbar \omega \simeq 3.58$ eV. ]{} []{data-label="fig:flux_time"}](flux_v_0_E0_17.7_time.eps){width="1.0\linewidth"} Semiclassical implementation of optimal control theory {#sec:OCT} ====================================================== In the previous section, we have shown that the electronic transition can be controlled with the help of the semiclassical ZN formula. As mentioned in the introduction, however, laser-driven chemical reactions consist of two processes, and we need to control not only electronic transitions but also wavepacket propagation. There seems to be no simple semiclassical methods to deal with the latter, so we examine the possibility to combine the optimal control theory (OCT) with a semiclassical wavepacket propagation method with the Herman-Kluk propagator [@MB00]. As an OCT procedure, we take the Zhu-Botina-Rabitz method [@ZBR98], and substitute the semiclassical propagator into the control scheme. The numerical result for HD$^+$ molecule is shown in Fig. \[fig:OCT\]: The control here is to shift the vibrational ground state to 1 a.u. away from the equilibrium position, the target time $T$ is 100 fs, and the so called penalty factor $\alpha$ is 0.01. ![ [ (a) Semiclassical (solid line) and quantum (dashed line) calculations of the overlap integral $|\langle \varphi_T| \phi(t) \rangle|^2$ where $|\varphi_T \rangle$ is the target state. (b) Semiclassical (solid line) and quantum (dashed line) optimal fields obtained after one iteration. ]{} []{data-label="fig:OCT"}](correlation.eps){width="1.0\linewidth"} ![ [ (a) Semiclassical (solid line) and quantum (dashed line) calculations of the overlap integral $|\langle \varphi_T| \phi(t) \rangle|^2$ where $|\varphi_T \rangle$ is the target state. (b) Semiclassical (solid line) and quantum (dashed line) optimal fields obtained after one iteration. ]{} []{data-label="fig:OCT"}](field.eps){width="1.0\linewidth"} As can be seen, the semiclassical result quantitatively agrees with the quantum result. Unfortunately, however, this straightforward semiclassical calculation needs more computational time compared with the corresponding quantum calculation! This is because the optimal field expressed as $$%\nonumber \varepsilon(t)=-\frac{1}{\alpha} {\rm Im} \left[ \langle \phi(t)|\chi(t) \rangle \langle \chi(t)| \mu |\phi(t) \rangle \right]$$ requires ${\cal O}(N^2)$ calculations where $N$ is the number of the trajectories. Here $|\phi(t) \rangle$ and $|\chi(t) \rangle$ are semiclassically calculated wavefunctions with computational cost ${\cal O}(N)$; so $\langle \phi(t)|\chi(t) \rangle$ and $\langle \chi(t)| \mu |\phi(t) \rangle$ amount to such an order of calculation in general. ($\mu$ is the transition dipole moment.) One might expect that we can reduce the cost by decreasing the width of the Gaussian wavepacket in the HK propagator; but it turns out that the overall efficiency cannot be improved. The same difficulty appears when one semiclassically calculates general correlation functions as analyzed by Miller [@Miller01]. We are now planning to use a sort of [*genetic algorithm*]{} [@JR92] to circumvent this difficulty since the optimal fields in the genetic algorithm procedure are represented as “genes”, and there is no such a scaling problem. (The initially guessed field can be obtained from the concept of nonadiabatic transition between vibrational states.) Semiclassical IVR method for nonadiabatic transitions {#sec:semi} ===================================================== The semiclassical methods mentioned above [*separately*]{} treat the processes of adiabatic propagation on a single adiabatic PES and those of electronic transition. This means that the whole process should be divided into three steps: (i) propagation on an initial PES, (ii) nonadiabatic transition, and (iii) propagation on a final PES. The optimal field should be designed so that no unnecessary interference occurs between the steps. In order to overcome such a restriction and to use the unique pulse to control the process, semiclassical IVR methods for nonadiabatic transitions should be constructed. One of the purposes of this work is to develop such a method which generalizes the ordinary Herman-Kluk propagator for the case of several electronic states. For other studies in this direction, see [@SM97]. The research is still under way and preliminary results for a 1D case are presented. The model system employed is H$_2^+$ molecule composed of 1s$\sigma_g$ and 2p$\sigma_u$ states. An initial wavefunction is assumed to be a Gaussian of width 0.5 a.u. with center at 5 a.u. on the 1s$\sigma_g$ state. The external field is assumed to be a linearly polarized femtosecond pulse with the Gaussian profile of FWHM = 10 fs. The total duration of wavepacket propagation is 20 fs. The wavelength of the femtosecond pulse is 515 nm and the crossing point of the dressed electronic states is located at 4 a.u. The results for the wavepacket propagation are shown in Fig. \[fig:semi\]. ![ [ (a) Final wavepacket on the 2p$\sigma_u$ state calculated quantum mechanically (thick line) and semiclassically (thin line). The total population: Quantum =0.30; Semiclassical =0.28. (b) The same on the 1s$\sigma_g$ state. The total population: Quantum =0.70; Semiclassical =0.72. ]{} []{data-label="fig:semi"}](fig3a.eps){width="1.0\linewidth"} ![ [ (a) Final wavepacket on the 2p$\sigma_u$ state calculated quantum mechanically (thick line) and semiclassically (thin line). The total population: Quantum =0.30; Semiclassical =0.28. (b) The same on the 1s$\sigma_g$ state. The total population: Quantum =0.70; Semiclassical =0.72. ]{} []{data-label="fig:semi"}](fig3b.eps){width="1.0\linewidth"} Outlook ======= We have advocated the use of semiclassical methods for control problems in view of its general capability to treat multidimensional systems with less computational cost. If the dimensionality of the system is high, the effects of “chaos” might also affect the quality of control. We have shown that a chaotic system called a kicked rotor can be controlled by using the Zhu-Botina-Rabitz method even when the system becomes strongly chaotic in the classical limit [@TFM03]. It should be confirmed whether this is generic or not for multidimensional systems. For much larger systems, even semiclassical calculations themselves are formidable, and in such a situation we need to use some kind of mixed quantum-classical methods. They are very practical ways to treat multidimensional phenomena like vibrational relaxation in solvents [@Okazaki], so it is interesting to control such phenomena with use of mixed quantum-classical methods. [99]{} S. A. Rice and M. Zhao, [*Optical Control of Molecular Dynamics*]{}, John-Wiley & Sons (2000). C. Zhu, Y. Teranishi, and H. Nakamura, Adv. Chem. Phys. [**117**]{}, 127 (2001); H. Nakamura, [*Nonadiabatic transition: Concepts, Basic Theories and Applications*]{}, World Scientific, Singapore (2002). K. Nagaya, Y. Teranishi, and H. Nakamura, J. Chem. Phys. [**113**]{}, 6197 (2000). K. Nagaya, Y. Teranishi, and H. Nakamura, J. Chem. Phys. [**117**]{}, 9588 (2002). H. Fujisaki, Y. Teranishi, and H. Nakamura, J. Theor. Comp. Chem. [**1**]{}, 245 (2002); e-print quant-ph/0211142. See, B. R. McQuarrie and P. Brumer, Chem. Phys. Lett. [**319**]{}, 27 (2000), and references therein. X. Sun and W. H. Miller, J. Chem. Phys. [**106**]{}, 6346 (1997); G. Stock and M. Thoss, Phys. Rev. Lett. [**78**]{}, 578 (1997); M. Thoss and G. Stock, Phys. Rev. A [**59**]{}, 64 (1999); M. Thoss, W. H. Miller, and G. Stock, J. Chem. Phys. [**112**]{}, 10282 (2000). A. B. Alekseyev, H-P. Liebermann, D. B. Kokh, and R. J. Buenker, J. Chem. Phys. [**113**]{}, 6174 (2000); N. Balakrishnan, A. B. Alekseyev, and R. J. Buenker, Chem. Phys. Lett. [**341**]{}, 594 (2001); N. Balakrishnan, private communication. M. Kawasaki, private communication. W. Zhu, J. Botina, and H. Rabitz, J. Chem. Phys. [**108**]{}, 1953 (1998). W. H. Miller, J. Phys. Chem. A [**105**]{}, 2942 (2001). R. S. Judson and H. Rabitz, Phys. Rev. Lett. [**68**]{}, 1500 (1992). T. Takami, H. Fujisaki, and T. Miyadera, unpublished. T. Terashima, M. Shiga, and S. Okazaki, J. Chem. Phys. [**114**]{}, 5663 (2001). [^1]: e-mail: [email protected], [email protected], [email protected], [email protected]

ArXiv

--- abstract: 'An important measure of bipartite entanglement is the entanglement of formation, which is defined as the minimum average pure state entanglement of all decompositions realizing a given state. A decomposition which achieves this minimum is called an *optimal* decomposition. However, as for the entanglement of formation, there is not much known about the structure of such optimal decompositions, except for some special cases, like states of two qubits or isotropic states. Here we present a necessary and sufficient condition for a set of pure states of a finite dimensional bipartite systems to form an optimal decomposition. This condition is well suited to treat the question, whether the entanglement of formation is additive or not.' author: - Tobias Prager title: A necessary and sufficient condition for optimal decompositions --- One of the main ingredients in quantum information theory is entanglement. Therefore, there have been many attempts to understand this property better. While in the case of bi-partite pure states, entanglement is well understood and uniquely quantified under some general assumptions [@BBPS] this is not the case for mixed states. There have been many different entanglement measures proposed for such states, one of which is the *entanglement of formation* [@Bennet], defined as $$\begin{aligned} \label{eof} E_f(\rho):=\min\{\sum_i p_i E({|\psi_i\rangle}) \Big| \sum_i p_i {{|\psi_i\rangle}\!{\langle\psi_i|}}=\rho\},\end{aligned}$$ where $E$ denotes the entanglement measure on pure states, $$\begin{aligned} E({|\psi\rangle})=-{\operatorname{tr}}\sigma \log_2 \sigma, \quad \sigma={\operatorname{tr}}_B {{|\psi\rangle}\!{\langle\psi|}}.\end{aligned}$$ Closely related to the entanglement of formation are so called *optimal decompositions* of a state, which are those decompositions $\{(p_i,{|\psi_i\rangle}\}_{i}$, that achieve the minimum in eq. (\[eof\]). However, for all cases except ${{\mathbb C}}^2\otimes{{\mathbb C}}^2$ [@Wootters] and some highly symmetric states in higher dimensions [@tv], it is not known, how to compute an optimal decomposition for a given state nor the entanglement of formation itself. One property of the entanglement of formation is being affine on the convex set of states, generated by convex linear combinations of pure states from an optimal decomposition [@Uhl]. In other words, if the decomposition $\{(p_i,{|\psi_i\rangle}\}_{i=1}^n$ with $p_i>0$ is optimal, then the decomposition $\{(q_i,{|\psi_i\rangle}\}_{i=1}^n$ where the $q_i$ form an arbitrary probability distribution is also optimal. The optimality of a decomposition thus only depends on the states ${|\psi_i\rangle}$ and not on the corresponding probabilities. This is the justification to talk about the optimality of a set of pure states $\{{|\psi_i\rangle}\}_i$, by which we mean the optimality of the decompositions $\{(p_i,{|\psi_i\rangle})\}_i$. A question concerning the entanglement of formation, which is widely believed to be true but not yet answered, is the question of additivity, i.e. whether $E_f(\rho\otimes \sigma)=E_f(\rho)+E_f(\sigma)$ holds true for all $\rho$ and $\sigma$. This question can be directly reformulated in terms of decompositions: Given the optimal decompositions $\{(p_i,{|\psi_i\rangle})\}_{i=1}^n$ of $\rho$ and $\{(q_j,{|\phi_j\rangle})\}_{j=1}^m$ of $\sigma$, is the tensorproduct decomposition $\{(p_i q_j ,{|\psi_i\rangle}\otimes{|\phi_j\rangle})\}_{i,j=1}^{n,m}$, which is clearly a decomposition of $\rho \otimes \sigma$, also an *optimal* one? In terms of optimal sets, this is equivalent to the question, whether the set of pure states $\{{|\psi_i\rangle}\otimes{|\phi_j\rangle}\}_{i,j=1}^{n,m}$ is optimal given that $\{{|\psi_i\rangle}\}_{i=1}^n$ and $\{{|\phi_j\rangle}\}_{j=1}^m$ are optimal. To address this question we present a necessary and sufficient condition, to decide whether a given set of pure states is optimal. Before we continue, we need to fix some notations. A decomposition of a state $\rho$ is normally defined as a set of normalized vectors ${|\psi_i\rangle}$ with appropriate probabilities $p_i$ such that $\rho=\sum_{i} p_i {{|\psi_i\rangle}\!{\langle\psi_i|}}$. However, for computational and notational purposes it is often convenient to define a decomposition to be a set of non normalized vectors, whose squared norms correspond to the probabilities, i.e. we make the transition $\{(p_i,{|\psi_i\rangle})\}_{i}\rightarrow \{{|\tilde \phi_i\rangle}\}_{i}$ with ${|\tilde \phi_i\rangle}=\sqrt{p_i} {|\psi_i\rangle}$. To distinguish non normalized from normalized vectors we mark them with a tilde. In this notation we have $\rho=\sum_i {{|\tilde \phi_i\rangle}\!{\langle\tilde \phi_i|}}$ and $$\begin{aligned} E_f(\rho)=\min\big\{ \sum_i {{\cal E}}({|\tilde \phi_i\rangle}) \big| \sum_i {{|\tilde \phi_i\rangle}\!{\langle\tilde \phi_i|}}=\rho\big\},\end{aligned}$$ where we have introduced the homogeneously extended pure state entanglement $$\begin{aligned} {{\cal E}}({|\tilde\phi\rangle}):=\|\tilde\phi\|^2 E(\frac{{|\tilde\phi\rangle}}{\|\tilde\phi\|}) =-{\operatorname{tr}}\sigma \log_2 \frac{\sigma}{{\operatorname{tr}}\sigma}\end{aligned}$$ with $\sigma={\operatorname{tr}}_B {{|\tilde\phi\rangle}\!{\langle\tilde\phi|}}$. To find all possible decompositions $\{(q_j,{|\phi_j\rangle})\}_{j=1}^m$ of a state $\rho$ from a given decomposition $\{(p_i,{|\psi_i\rangle})\}_{i=1}^n$ ( e.g. the spectral decomposition) we will make repeated use of a theorem by Wootters et al. [@JW]. This theorem states, that two decompositions $\{(q_j,{|\phi_j\rangle})\}_{j=1}^m$ and $\{(p_i,{|\psi_i\rangle})\}_{i=1}^n$ generate the same state if and only if they are related by a right unitary matrix $U$, i.e. a matrix obeying $U^\dagger U=\openone$, in the following way: $$\begin{aligned} \sqrt{q_j}{|\phi_j\rangle}=\sum_{i=1}^n U_{ji} \sqrt{p_i}{|\psi_i\rangle}\end{aligned}$$ or in terms of non normalized decompositions $$\begin{aligned} \label{aes} {|\tilde \phi_j\rangle}=\sum_{i=1}^n U_{ji} {|\tilde \psi_i\rangle}.\end{aligned}$$ We now state our main result, a necessary and sufficient condition for the optimality of a set of pure states on a bipartite finite dimensional Hilbert space: **Theorem:** Let ${|\psi_i\rangle}\in {{\cal H}}_A\otimes {{\cal H}}_B$, $i=1,\ldots,n$ and $\sigma_{ij}:={\operatorname{tr}}_B {{|\psi_i\rangle}\!{\langle\psi_j|}}$. The set of pure states $\{{|\psi_i\rangle}\}_{i=1}^n$ is optimal if and only if $$\begin{aligned} \label{newcondeq} {{\cal E}}(\sum_{i=1}^n c_i {|\psi_i\rangle})\ge -{\operatorname{Re}}\big(\sum_{i,j=1}^n c_i \bar c_j {\operatorname{tr}}[\sigma_{ij} \log_2 \sigma_{ii}]\big) \end{aligned}$$ holds true for arbitrary $c_i \in {{\mathbb C}}$. First we prove sufficiency: Let $\{{|\psi_i\rangle}\}_{i=1}^n$ be a set of pure states for which eq. (\[newcondeq\]) is satisfied. Choosing an arbitrary probability distribution $\{p_i\}_{i=1}^n$ we can form a decomposition $\{(p_i,{|\psi_i\rangle})\}_{i=1}^n$ representing the state $\rho=\sum_{i=1}^n p_i {{|\psi_i\rangle}\!{\langle\psi_i|}}$. Every other decomposition $\{(q_j,{|\phi_j\rangle})\}_{j=1}^m$ of $\rho$ is obtained via a right unitary matrix $U$ by $$\begin{aligned} \sqrt{q_j} {|\phi_j\rangle}=\sum_{i=1}^n U_{ji} \sqrt{p_i} {|\psi_i\rangle}.\end{aligned}$$ From eq. (\[newcondeq\]) we get $$\begin{aligned} {{\cal E}}(\sqrt{q_j}{|\phi_j\rangle})&=& {{\cal E}}(\sum_{i=1}^n U_{ji} \sqrt{p_i} {|\psi_i\rangle})\\ &\ge&-{\operatorname{Re}}\Big(\sum_{i,k=1}^n U_{ji} \bar U_{jk} \sqrt{p_i p_k} {\operatorname{tr}}[\sigma_{ik} \log_2 \sigma_{ii}]\Big)\end{aligned}$$ with $\sigma_{ik}={\operatorname{tr}}{{|\psi_i\rangle}\!{\langle\psi_k|}}$. Summing over $j$ and using the right unitarity of $U$ we arrive at $$\begin{aligned} \lefteqn{\sum_{j=1}^m q_j E({|\phi_j\rangle})=\sum_{j=1}^m {{\cal E}}(\sqrt{q_j}{|\phi_j\rangle})}\\ &\ge& -{\operatorname{Re}}\sum_{i=1}^n p_i {\operatorname{tr}}[\sigma_{ii} \log_2 \sigma_{ii}]= \sum_{i=1}^n p_i E({|\psi_i\rangle}) \end{aligned}$$ which proves the optimality of the decomposition $\{(p_i,{|\psi_i\rangle})\}_{i=1}^n$ and therefore the optimality of the set $\{{|\psi_i\rangle}\}_{i=1}^n$. Next we prove necessity: Let $\{{|\psi_i\rangle}\}_{i=1}^n$ be an optimal set of pure states. Then the decomposition $\{(\frac{1}{n},{|\psi_i\rangle})\}_{i=1}^{n} =:\{{|\tilde \psi_i\rangle}\}_{i=1}^{n} $ representing the state $\rho=\sum_{i=1}^n \frac{1}{n} {{|\psi_i\rangle}\!{\langle\psi_i|}}$ is optimal. We now define a family of unitary $(n+1)\times (n+1)$-matrizes by $\tilde U(t):=\exp(tT)$, where t is a real parameter and $T$ is the skew hermitian matrix defined by $T_{n+1,i}=c_i$, $T_{i,n+1}=-\bar c_i$ and $T_{i,j}=0$ otherwise, i.e. $$\begin{aligned} T=\left( \begin{array}{ccccc} 0 & 0 &\cdots & 0 &-\bar c_1\\ 0 & 0 &\cdots & 0 &-\bar c_2\\ \vdots &\vdots& \ddots & \vdots &\vdots\\ 0 & 0 &\cdots & 0 &-\bar c_n\\ c_1 & c_2& \cdots & c_n& 0 \end{array} \right).\end{aligned}$$ Selecting the first $n$ columns of $\tilde U(t)$ we get a right unitary $(n+1)\times n$-matrix $U(t)$. Applying $U(t)$ to the optimal decomposition $\{{|\tilde \psi_i\rangle}\}_{i=1}^{n}$ according to eq. (\[aes\]) we get a new decomposition $\{{|\tilde \phi_j(t)\rangle}\}_{j=1}^{n+1}$ of $\rho$ which is defined by $$\begin{aligned} {|\tilde \phi_i(t)\rangle}&=&{|\tilde \psi_i\rangle}- \frac{t^2}{2} \sum_{j=1}^{n} c_j \bar c_i {|\tilde \psi_j\rangle} +{{\cal O}}(t^3),\; i=1,\ldots n \nonumber\\ {|\tilde \phi_{n+1}(t)\rangle}&=&t \sum_{i=1}^n c_i {|\tilde \psi_i\rangle}+{{\cal O}}(t^3)\label{phii}. \end{aligned}$$ This decomposition can not have less entanglement than the decomposition $\{{|\tilde \psi_i\rangle}\}_{i=1}^{n}$, which was assumed to be optimal, thus $$\begin{aligned} \label{oes} \sum_{i=1}^{n} {{\cal E}}({|\tilde\psi_i\rangle})\le \sum_{j=1}^{n+1} {{\cal E}}\big({|\tilde\phi_j(t)\rangle}\big) \quad \forall t\in {{\mathbb R}}\end{aligned}$$ holds true. At $t$=0, both sides of this inequality are equal. In the following, we will show that the first derivatives with respect to $t$ of both sides are also equal at $t=0$. Hence, the inequality must be fullfilled for the second derivatives at $t=0$ which will lead to eq (\[newcondeq\]). To proceed, we need the following lemma: **Lemma:** Let $A(t)$ be a family of positive matrices, which depend differentiably on $t\in(a,b)$. Then we have $$\begin{aligned} \label{fdhe} -\frac{d}{dt}{\operatorname{tr}}[A(t) \log_2 \frac{A(t)}{{\operatorname{tr}}A(t)}]= -{\operatorname{tr}}[A(t)' \log_2 \frac{A(t)}{{\operatorname{tr}}A(t)}] \end{aligned}$$ Additionaly, if $\frac{d}{dt} A(t)$ is zero at some point $t_0$ and if $A(t)$ is twice differentiable at $t_0$ we have $$\label{sdhe} \begin{split} -\frac{d^2}{dt^2}{\operatorname{tr}}[A(t) \log_2 \frac{A(t)}{{\operatorname{tr}}A(t)}]\Big|_{t=t_0}\!\!\!\!= -{\operatorname{tr}}[A(t_0){''} \log_2 \frac{A(t_0)}{{\operatorname{tr}}A(t_0)}]. \end{split}$$ Note that eq. (\[fdhe\]) is always well defined due to the positivity and differentiability of $A(t)$ whereas eq. (\[sdhe\]) may not be well defined. Let $A(t)=U(t) {{\text{diag}}}(\vec{\lambda}(t)) U^\dagger(t)$ with $U(t)$ unitary. Then we have $$\begin{aligned} \lefteqn{-\frac{d}{dt}{\operatorname{tr}}[A(t)\log_2 \frac{A(t)}{{\operatorname{tr}}A(t)}]}\\ &=&-\frac{d}{dt}\big(\sum_i \lambda_i(t)( \log_2 \lambda_i(t) -\log_2 \sum_j \lambda_j(t))\big)\\ &=&-\sum_i \lambda_i'(t)( \log_2 \lambda_i(t) -\log_2 \sum_j \lambda_j(t))\\ &=&-{\operatorname{tr}}[A'(t)( \log_2 A(t) -\log_2 {\operatorname{tr}}A(t))]\\ &=&-{\operatorname{tr}}[A(t)' \log_2 \frac{A(t)}{{\operatorname{tr}}A(t)}].\end{aligned}$$ In the last but one step, we have used that $U'(t)U^\dagger(t)+U(t) U^{\dagger \prime}(t)=0$, that the trace is invariant under cyclic permutations of the factors and that ${{\text{diag}}}(\vec\lambda(t))$ and $\log_2[{{\text{diag}}}(\vec\lambda(t))]$ commute. To prove the second part of the lemma, we use the fact that if $A'(t_0)$ is zero $\lambda_i'(t_0)$ is zero and we can always choose $U(t)$ such that $U'(t_0)=0$. This choice leads to $$\begin{aligned} \lefteqn{-\frac{d^2}{dt^2}{\operatorname{tr}}[A(t) \log_2 \frac{A(t)}{{\operatorname{tr}}A(t)}] \Big|_{t=t_0}=}\\ &=&-\sum_i \lambda_i{''}(t_0)( \log_2 \lambda_i(t_0) -\log_2 \sum_j \lambda_j(t_0))\\ &=&-{\operatorname{tr}}[A{''}(t_0) \log_2 \frac{A(t_0)}{{\operatorname{tr}}A(t_0)}].\end{aligned}$$ having used $U''(t_0)U^\dagger(t_0)+U(t_0) U^{\dagger\prime\prime}(t_0)=0$. which holds true, because $U'(t_0)=0$. Applying this lemma to $\tilde \pi_i(t):={\operatorname{tr}}_B {{|\tilde \phi_i(t)\rangle}\!{\langle\tilde \phi_i(t)|}}$, $i=1,\ldots,n$, we get $$\begin{aligned} \frac{d}{d t} {{\cal E}}({|\tilde \phi_i(t)\rangle})&=& -\frac{d}{dt}\big( {\operatorname{tr}}\tilde \pi_i(t) \log_2 \frac{\tilde \pi_i(t)}{{\operatorname{tr}}\tilde \pi_i(t)}\big)\\ &=&-{\operatorname{tr}}\tilde \pi'_i(t) \log_2 \frac{\tilde \pi_i(t)}{{\operatorname{tr}}\tilde \pi_i(t)}.\end{aligned}$$ According to eq. (\[phii\]), $\tilde \pi_i'(t)$ is zero at $t=0$. Therefore we can calculate the second derivative of ${{\cal E}}({|\tilde \phi_i(t)\rangle})$ at $t=0$ using the second part of the lemma, which leads to $$\begin{aligned} \frac{d^2}{d t^2} {{\cal E}}({|\tilde \phi_i(t)\rangle})\Big|_{t=0}&=&-{\operatorname{tr}}\tilde \pi{''}_i(0) \log_2 \frac{\tilde \pi_i(0)}{{\operatorname{tr}}\tilde \pi_i(0)}\end{aligned}$$ Using eq. (\[phii\]) we can further evaluate the righthand side. Defining $\tilde\sigma_{ij}:={\operatorname{tr}}_B {{|\tilde \psi_i\rangle}\!{\langle\tilde \psi_j|}}$, we arrive at $$\begin{aligned} \label{sd1} \frac{d^2}{d t^2} {{\cal E}}({|\tilde \phi_i(t)\rangle})\big|_{t=0} &=&\sum_{j=1}^n {\operatorname{tr}}[(c_i \bar c_j \tilde\sigma_{ij} +c_j \bar c_i \tilde\sigma_{ji})\log_2 \frac{\tilde\sigma_{ii}}{{\operatorname{tr}}\tilde\sigma_{ii}}]\nonumber\\ &=&2{\operatorname{Re}}\Big(\sum_{j=1}^n c_i \bar c_j {\operatorname{tr}}[\tilde\sigma_{ij}\log_2 \frac{\tilde\sigma_{ii}}{{\operatorname{tr}}\tilde\sigma_{ii}}] \Big)\nonumber\\ &=&\frac{2}{n}{\operatorname{Re}}\Big(\sum_{j=1}^n c_i \bar c_j {\operatorname{tr}}[\sigma_{ij}\log_2 \sigma_{ii}] \Big).\end{aligned}$$ This expression is always well defined because ${\operatorname{tr}}\sigma_{ij} \log_2 \sigma_{ii}$ is well defined, even if $\sigma_{ii}$ has a non vanishing kernel. This can be easily seen by expressing ${|\psi_i\rangle}$ and ${|\psi_j\rangle}$ in the Schmidt-basis of ${|\psi_i\rangle}$ and then evaluating ${\operatorname{tr}}\sigma_{ij} \log_2 \sigma_{ii}$ in this basis. To calculate the derivatives of ${{\cal E}}({|\tilde \phi_{n+1}(t)\rangle})$ we use the fact, that ${{\cal E}}(c {|\tilde \eta\rangle})=|c|^2 {{\cal E}}({|\tilde\eta\rangle})$ which leads to $$\begin{aligned} \frac{d}{dt} {{\cal E}}({|\tilde\phi_{n+1}(t)\rangle})\big|_{t=0}&=&\frac{d}{dt}[ t^2 {{\cal E}}(\sum_{i=1}^n c_i {|\tilde\psi_i\rangle}+{{\cal O}}(t^2))]\big|_{t=0}\nonumber\\&=&0\end{aligned}$$ and $$\begin{aligned} \label{sd2} \frac{d^2}{d t^2} {{\cal E}}({|\tilde\phi_{n+1}(t)\rangle})\big|_{t=0}&=&\frac{d^2}{d t^2}[ t^2 {{\cal E}}(\sum_{i=1}^n c_i {|\tilde\psi_i\rangle}+{{\cal O}}(t^2))]\big|_{t=0}\nonumber \\ &=&\frac{2}{n} {{\cal E}}(\sum_{i=1}^n c_i {|\psi_i\rangle}).\end{aligned}$$ Note that the first and second derivative of ${{\cal E}}(\sum_{i=1}^n c_i {|\tilde\psi_i\rangle}+{{\cal O}}(t^2))]\big|_{t=0}$ are well defined. As we have shown, the first derivatives of the right and left hand side of eq. (\[oes\]) are equal to 0. Therefore $$\begin{aligned} \label{sd} \frac{d^2}{d t^2}\Big( \sum_{i=1}^{n} {{\cal E}}({|\tilde\psi_i\rangle})\Big)\Big|_{t=0}\le \frac{d^2}{d t^2}\Big( \sum_{j=1}^{n+1} {{\cal E}}\big({|\tilde\phi_j(t)\rangle}\big)\Big)\Big|_{t=0}\end{aligned}$$ necessarily holds true, which in turn can be seen to be equivalent to eq. (\[newcondeq\]) by substituting eqs. (\[sd1\]) and (\[sd2\]) into the right hand side as well as taking into account, that the left hand side of eq. (\[sd\]) does not depend on $t$ and therefore is zero. Our condition for the optimality of a set of pure states is a generalization of a result by Benatti and Narnhofer [@BN], who proved a similar condition for the special case of a decompositions consisting only of *two* vectors. It is easily seen, that it is sufficient to demand eq. (\[newcondeq\]) to be true for all $c_i\in {{\mathbb C}}$ with $\|\vec c\|=1$ because we can cancel the square of the norm of the coefficient vector $\vec c$ on both sides of the inequality. Additionally, *if* the set $\{{|\psi_i\rangle}\}_{i=1}^n$ is optimal, the sum on the right side of eq. (\[newcondeq\]) is already real. This is implied by the fact, that for two states ${|\psi_k\rangle}$ and ${|\psi_l\rangle}$ from an optimal set, we have $$\begin{aligned} \label{nc} {\operatorname{tr}}\sigma_{kl} \log_2 \sigma_ {kk}= \overline{{\operatorname{tr}}\sigma_{lk} \log_2 \sigma_{ll}},\; \sigma_{kl}:={\operatorname{tr}}_B{{|\psi_k\rangle}\!{\langle\psi_l|}}. \end{aligned}$$ To prove this equation we define the decompositions $\{{|\tilde \phi_i(\theta)\rangle}\}_{i=1}^n$ with $$\begin{aligned} {|\tilde\phi^\pm_k(\theta)\rangle}&=&\frac{1}{\sqrt{n}}( \cos \theta {|\psi_k\rangle}\pm\sin \theta {|\psi_l\rangle})\\ {|\tilde\phi^\pm_l(\theta)\rangle}&=&\frac{1}{\sqrt{n}}(\mp\sin\theta {|\psi_k\rangle}+ \cos \theta {|\psi_l\rangle})\text{ and }\\ {|\tilde\phi^\pm_i(\theta)\rangle}&=&\frac{1}{\sqrt{n}} {|\psi_i\rangle} \quad i\neq k,l \end{aligned}$$ and $\{{|\tilde \eta_i(\theta)\rangle}\}_{i=1}^n$ with $$\begin{aligned} {|\tilde\eta^\pm_k(\theta)\rangle}&=&\frac{1}{\sqrt{n}}( \cos \theta {|\psi_k\rangle}\pm I\sin \theta {|\psi_l\rangle})\\ {|\tilde\eta^\pm_l(\theta)\rangle}&=&\frac{1}{\sqrt{n}}(\pm I\sin\theta {|\psi_k\rangle}+ \cos \theta {|\psi_l\rangle})\text{ and }\\ {|\tilde\eta^\pm_i(\theta)\rangle}&=&\frac{1}{\sqrt{n}} {|\psi_i\rangle} \quad i\neq k,l. \end{aligned}$$ For $\theta=0$ those decompositions are identical to $\{\frac{1}{\sqrt{n}}{|\psi_i\rangle}\}_{i=1}^n$ which is optimal by assumption. Hence $$\begin{aligned} \frac{d}{d\theta} \sum_{i=1}^n {{\cal E}}({|\tilde\phi^\pm_i(\theta)\rangle})\Big|_{t=0}\ge 0\text{ and } \frac{d}{d\theta} \sum_{i=1}^n {{\cal E}}({|\tilde\eta^\pm_i(\theta)\rangle})\Big|_{t=0}\ge0\end{aligned}$$ have to hold true, implying eq. (\[nc\]) by using the lemma and some basic algebra. This theorem is usefull to verify the optimality of a given decomposition or a given set of pure states, because we only have to optimize a function over the unit sphere in ${{\mathbb C}}^n$. In contrast, by using the straight forward condition $$\begin{aligned} \label{oldcond} \sum_{i=1}^n {{\cal E}}({|\tilde \psi_i\rangle})\le \sum_{j=1}^m {{\cal E}}(\sum_{i=1}^n u_{ji}{|\tilde\psi_i\rangle}) \; \forall \; U \text{ with } U^\dagger U=\mathbf{1}\end{aligned}$$ for the optimality of the decomposition $\{{|\tilde\psi_i\rangle}\}_{i=1}^n$ which is easily derived from the definition of an optimal decomposition and eq. (\[aes\]), we had to optimize a function over the set of all right unitary $m\times n$-matrices. In this case $m$ must be at least $[\operatorname{rank}\rho]^2$ with $\rho=\sum_{i=1}^n {{|\tilde \psi_i\rangle}\!{\langle\tilde \psi_i|}}$, which is the lowest known bound on the number of pure states in an optimal decomposition [@Uhlmann]. Clearly this task is harder to perform. However, as there is nothing for free, the information we get in both cases is different: In doing the optimization over the right unitary matrizes we arrive at an optimal decomposition from which the entanglement of formation can be easily calculated. In doing the optimization over the unit sphere in ${{\mathbb C}}^n$ , we only get the information whether the decomposition was optimal or not. Therefore our condition is of no direct use in finding optimal decompositions and thus calculating the entanglement of formation, but there is another important application, namely testing the additivity of the entanglement of formation. To perform this task, we have to check, whether the set of tensorproduct states $\{{|\psi^1_i\rangle}\otimes {|\psi^2_j\rangle}\}_{i,j=1}^{n,m}$ is optimal, given two optimal sets of states $\{{|\psi^1_i\rangle}\}_{i=1}^n$ and $\{{|\psi^2_j\rangle}\}_{j=1}^{m}$. Numerically this task is easier to perform using our theorem, than using eq.(\[oldcond\]). It might even be possible, that the new condition is helpful in proving additivity of the entanglement of formation analytically. Another information we get from our condition is an estimation of the entanglement of superpositions of states from an optimal set in terms of the entanglement of the states themselves and some overlap-terms. This can be seen by writing the diagonal terms in eq. (\[newcondeq\]) separately, $$\begin{aligned} \lefteqn{ {{\cal E}}(\sum_{i=1}^n c_i {|\psi_i\rangle})\ge \sum_{i=1}^n |c_i|^2 E({|\psi_i\rangle})-}\\ &&\qquad-{\operatorname{Re}}\big(\sum_{\substack{i,j=1\\i\neq j}}^n c_i \bar c_j {\operatorname{tr}}[\sigma_{ij} \log_2 \sigma_{ii}]\big) \quad \forall \; c_i \in {{\mathbb C}}. \end{aligned}$$ Finally we conclude that in some sense locally optimal decompositions are already optimal: In deriving eq. (\[newcondeq\]) we have only used derivatives and the value itself at the point $t=0$ in eq. (\[oes\]). Therefore it is sufficient to demand eq. (\[oes\]) to be true only in a small neighbourhood of $t=0$. This in turn is equivalent to demanding, that the decompositions $\{{|\tilde \psi_i\rangle}\}_{i=1}^{n}$ has lower average entanglement than every decomposition, which we get from it by a right unitary $(n+1)\times n$-matrix, which consists of the first $n$ columns of a unitary $(n+1)\times (n+1)$-matrix in an arbitrary small neighbourhood of the identity, i.e. demanding local optimality. But since eq. (\[newcondeq\]) is also a sufficient condition for optimality, we conclude that local optimality implies optimality. The author would like to thank A. Szkola, C. Witte and K.E. Hellwig for usefull discussions. The author gratefully acknowledges financial support from the Deutsche Forschungsgemeinschaft (DFG). [99]{} W. Wootters, Phys. Rev. Lett. **80**, 2245-2248 (1998) B.M. Terhal and K.G.H. Vollbrecht Phys. Rev. Lett. **85**, 2625 (2000) C.H. Bennet, H. Bernstein, S. Popescu and B. Schumacher, Phys.Rev.A **53**, 2986 (1996); C.H. Bennet, D.P. DiVicenzo, J.A. Smolin and W.K. Wootters, Phys. Rev. A **54**, 3824 (1996) F. Benatti, H. Narnhofer, A.Uhlmann, Rep. Math. Phys. **38**, 123 (1996) L. Hughston, R. Josza, W. Wootters, Physics Lett. A **183**, 14 (1993) F. Benatti, H. Narnhofer, quant-ph/0005126 K.Audenaert, F. Verstraete, B. De Moor, quant-ph/0006128 Armin Uhlmann, quant-ph/9704017

ArXiv

--- address: 'Dept. of Physics & MIT Kavli Institute, Massachusetts Institute of Technology, Cambridge, MA 02139' author: - Max Tegmark date: 'February 20 2005, published in [*“Universe or Multiverse?"*]{}, B. Carr ed., Cambridge University Press, 2007) ' title: The Multiverse Hierarchy --- \#1 \#2[\#1, \#2.]{} \#1 \#2 \#3[\#1, \#2. \#3.]{} \#1 \#2 \#3 \#4[\#1, \#2. \#3. \#4.]{} \#1 \#2 \#3 \#4 \#5[\#1, \#2. \#3. \#4. \#5.]{} \#1;\#2;\#3;\#4;\#5;\#6 \#1 \#2, [*\#3*]{}, [**\#4**]{}, \#5 (“\#6”) \#1;\#2;\#3;\#4;\#5 \#1 \#2, [*\#3*]{}, [**\#4**]{}, \#5 \#1;\#2;\#3;\#4;\#5 \#1 \#2, [*\#3*]{} (\#4: \#5) \#1;\#2;\#3;\#4;\#5;\#6 \#1 \#2, in [*\#3*]{}, ed. \#4 (\#5: \#6) \#1;\#2;\#3;\#4;\#5;\#6;\#7 \#1 \#2, in [*\#3*]{}, ed. \#4 (\#5: \#6), p\#7 \#1;\#2;\#3 \#1 \#2, \#3 \#1;\#2;\#3 \#1 \#2, \#3 \#1[Section \[\#1\]]{} \#1[\#1]{} \#1[$$\label{#1}} \def\eeq{$$]{} \#1[$$\begin{aligned} \label{#1}} \def\eeqa{\end{aligned}$$]{} \#1[equation (\[\#1\])]{} \#1[Equation (\[\#1\])]{} \#1[ (\[\#1\])]{} \#1[Section \[\#1\]]{} \#1[Section \[\#1\]]{} \#1[to 0pt[\#1]{}]{} \#1[$$]{} \#1[Figure \[\#1\]]{} \#1[Figure \[\#1\]]{} \#1[Section \[\#1\]]{} \#1[Section \[\#1\]]{} Parallel universes are now all the rage, cropping up in books, movies and even jokes: “You passed your exam in many parallel universes — but not this one." However, they are as controversial as they are popular, and it is important to ask whether they are within the purview of science, or merely silly speculation. They are also a source of confusion, since many forget to distinguish between different types of parallel universes that have been proposed. The farthest you can observe is the distance that light has been able to travel during the 14 billion years since the big-bang expansion began. The most distant visible objects are now about $4\times 10^{26}$ meters away[^1], and a sphere of this radius defines our observable universe, also called our [*Hubble volume*]{}, our [*horizon volume*]{} or simply our universe. In this article, I survey physics theories involving parallel universes, which form a natural four-level hierarchy of multiverses (Figure 1) allowing progressively greater diversity. - [**Level I:**]{} A generic prediction of cosmological inflation is an infinite “ergodic” space, which contains Hubble volumes realizing all initial conditions — including an identical copy of you about $10^{10^{29}}$m away. - [**Level II:**]{} Given the [*fundamental*]{} laws of physics that physicists one day hope to capture with equations on a T-shirt, different regions of space can exhibit different [*effective*]{} laws of physics (physical constants, dimensionality, particle content, ) corresponding to different local minima in a landscape of possibilities. - [**Level III:**]{} In unitary quantum mechanics, other branches of the wavefunction add nothing qualitatively new, which is ironic given that this level has historically been the most controversial. - [**Level IV:**]{} Other mathematical structures give different fundamental equations of physics for that T-shirt. The key question is therefore not whether there is a multiverse (since Level I is the rather uncontroversial cosmological concordance model), but rather how many levels it has. Below we will discuss at length the issue of evidence and whether this is science or philosophy. For now, the key point to remember is that [*parallel universes are not a theory, but a prediction of certain theories*]{}. For a theory to be falsifiable, we need not be able to observe and test all its predictions, merely at least one of them. Consider the following analogy: General Relativity Black hole interiors --------------------------- ------------------------------ Inflation Level I parallel universes Unitary quantum mechanics Level III parallel universes Because Einstein’s theory of General Relativity has successfully predicted many things that we [*can*]{} observe, we also take seriously its predictions for things we cannot observe, , that space continues inside black hole event horizons and that (contrary to early misconceptions) nothing funny happens right at the horizon. Likewise, successful predictions of the theories of cosmological inflation and unitary[^2] quantum mechanics have made some scientists take more seriously their other predictions, including various types of parallel universes. Let us conclude with two cautionary remarks before delving into the details. Hübris and lack of imagination have repeatedly caused us humans to underestimate the vastness of the physical world, and dismissing things merely because we cannot observe them from our vantage point is reminiscent of the ostrich with its head in the sand. Moreover, recent theoretical insights have indicated that Nature may be tricking us. Einstein taught us that space is not merely a boring static void, but a dynamic entity that can stretch (the expanding universe), vibrate (gravitational waves) and curve (gravity). Searches for a unified theory also suggest that space can “freeze”, transitioning between different phases in a landscape of possibilities just like water can be solid, liquid or gas. In different phases, effective laws of physics (particles, symmetries, .) could differ. A fish never leaving the ocean might mistakenly conclude that the properties of water are universal, not realizing that there is also ice and steam. We may be smarter than fish, but could be similarly fooled: cosmological inflation has the deceptive property of stretching a small patch of space in a particular phase so that it fills our entire observable universe, potentially tricking us into misinterpreting our local conditions for the universal laws that should go on that T-shirt. -1.5cm \[ZoomFig\] Level I: Regions beyond our cosmic horizon ========================================== Let us return to your distant twin. If space is infinite and the distribution of matter is sufficiently uniform on large scales, then even the most unlikely events must take place somewhere. In particular, there are infinitely many other inhabited planets, including not just one but infinitely many with people with the same appearance, name and memories as you. Indeed, there are infinitely many other regions the size of our observable universe, where every possible cosmic history is played out. This is the Level I multiverse. Evidence for Level I parallel universes --------------------------------------- Although the implications may seem crazy and counter-intuitive, this spatially infinite cosmological model is in fact the simplest and most popular one on the market today. It is part of the cosmological concordance model, which agrees with all current observational evidence and is used as the basis for most calculations and simulations presented at cosmology conferences. In contrast, alternatives such as a fractal universe, a closed universe and a multiply connected universe have been seriously challenged by observations. Yet the Level I multiverse idea has been controversial (indeed, an assertion along these lines was one of the heresies for which the Vatican had Giordano Bruno burned at the stake in 1600[^3]), so let us review the status of the two assumptions (infinite space and “sufficiently uniform” distribution). How large is space? Observationally, the lower bound has grown dramatically () with no indication of an upper bound. We all accept the existence of things that we cannot see but could see if we moved or waited, like ships beyond the horizon. Objects beyond cosmic horizon have similar status, since the observable universe grows by a light-year every year as light from further away has time to reach us[^4]. If anything, the Level I multiverse sounds trivially obvious. How could space not be infinite? Is there a sign somewhere saying “Space Ends Here–Mind the Gap”? If so, what lies beyond it? In fact, Einstein’s theory of gravity calls this intuition into question. Space could be finite if it has a convex curvature or an unusual topology (that is, interconnectedness). A spherical, doughnut-shaped or pretzel-shaped universe would have a limited volume and no edges. The cosmic microwave background radiation allows sensitive tests of such scenarios. So far, however, the evidence is against them. Infinite models fit the data, and strong limits have been placed on the alternatives (de Oliveira-Costa 2003; Cornish 2003). In addition, a spatially infinite universe is a generic prediction of the cosmological theory of inflation (Garriga & Vilenkin 2001b), so the striking successes of inflation listed below therefore lend further support to the idea that space is after all simple and infinite just as we learned in school. Another loophole is that space is infinite but matter is confined to a finite region around us–the historically popular “island universe” model. In a variant on this model, matter thins out on large scales in a fractal pattern. In both cases, almost all universes in the Level I multiverse would be empty and dead. But recent observations of the three-dimensional galaxy distribution and the microwave background have shown that the arrangement of matter gives way to dull uniformity on large scales, with no coherent structures larger than about 1024 meters. Assuming that this pattern continues, space beyond our observable universe teems with galaxies, stars and planets. -2.5cm What are Level I parallel universes like? {#ErgodicitySec} ----------------------------------------- The physics description of the world is traditionally split into two parts: initial conditions and laws of physics specifying how the initial conditions evolve. Observers living in parallel universes at Level I observe the exact same laws of physics as we do, but with different initial conditions than those in our Hubble volume. The currently favored theory is that the initial conditions (the densities and motions of different types of matter early on) were created by quantum fluctuations during the inflation epoch (see section 3). This quantum mechanism generates initial conditions that are for all practical purposes random, producing density fluctuations described by what mathematicians call an ergodic random field. [*Ergodic*]{} means that if you imagine generating an ensemble of universes, each with its own random initial conditions, then the probability distribution of outcomes in a given volume is identical to the distribution that you get by sampling different volumes in a single universe. In other words, it means that everything that could in principle have happened here did in fact happen somewhere else. Inflation in fact generates all possible initial conditions with non-zero probability, the most likely ones being almost uniform with fluctuations at the $10^{-5}$ level that are amplified by gravitational clustering to form galaxies, stars, planets and other structures. This means both that pretty much all imaginable matter configurations occur in some Hubble volume far away, and also that we should expect our own Hubble volume to be a fairly typical one — at least typical among those that contain observers. A crude estimate suggests that the closest identical copy of you is about $\sim10^{10^{29}}$m away. About $\sim 10^{10^{91}}$m away, there should be a sphere of radius 100 light-years identical to the one centered here, so all perceptions that we have during the next century will be identical to those of our counterparts over there. About $\sim 10^{10^{115}}$m away, there should be an entire Hubble volume identical to ours.[^5] This raises an interesting philosophical point that will come back and haunt us in : if there are indeed many copies of “you” with identical past lives and memories, you would not be able to compute your own future even if you had complete knowledge of the entire state of the cosmos! The reason is that there is no way for you to determine which of these copies is “you” (they all feel that they are). Yet their lives will typically begin to differ eventually, so the best you can do is predict probabilities for what you will experience from now on. This kills the traditional notion of determinism. How a multiverse theory can be tested and falsified {#TestingSec} --------------------------------------------------- Is a multiverse theory one of metaphysics rather than physics? As emphasized by Karl Popper, the distinction between the two is whether the theory is empirically testable and falsifiable. Containing unobservable entities does clearly [*not*]{} per se make a theory non-testable. For instance, a theory stating that there are 666 parallel universes, all of which are devoid of oxygen makes the testable prediction that we should observe no oxygen here, and is therefore ruled out by observation. As a more serious example, the Level I multiverse framework is routinely used to rule out theories in modern cosmology, although this is rarely spelled out explicitly. For instance, cosmic microwave background (CMB) observations have recently shown that space has almost no curvature. Hot and cold spots in CMB maps have a characteristic size that depends on the curvature of space, and the observed spots appear too large to be consistent with the previously popular “open universe” model. However, the average spot size randomly varies slightly from one Hubble volume to another, so it is important to be statistically rigorous. When cosmologists say that the open universe model is ruled out at 99.9% confidence, they really mean that if the open universe model were true, then fewer than one out of every thousand Hubble volumes would show CMB spots as large as those we observe — therefore the entire model with all its infinitely many Hubble volumes is ruled out, even though we have of course only mapped the CMB in our own particular Hubble volume. The lesson to learn from this example is that multiverse theories [*can*]{} be tested and falsified, but only if they predict what the ensemble of parallel universes is and specify a probability distribution (or more generally what mathematicians call a [*measure*]{}) over it. As we will see in , this measure problem can be quite serious and is still unsolved for some multiverse theories. Level II: Other post-inflation bubbles ====================================== If you felt that the Level I multiverse was large and hard to stomach, try imagining an infinite set of distinct ones (each symbolized by a bubble in Figure 1), some perhaps with different dimensionality and different physical constants. This is what is predicted by most currently popular models of inflation, and we will refer to it as the Level II multiverse. These other domains are more than infinitely far away in the sense that you would never get there even if you traveled at the speed of light forever. The reason is that the space between our Level I multiverse and its neighbors is still undergoing inflation, which keeps stretching it out and creating more volume faster than you can travel through it. In contrast, you could travel to an arbitrarily distant Level I universe if you were patient and the cosmic expansion decelerates.[^6] Evidence for Level II parallel universes ---------------------------------------- Inflation is an extension of the big bang theory and ties up many of the loose ends of that theory, such as why the universe is so big, so uniform and so flat. A rapid stretching of space long ago can explain all these and other attributes in one fell swoop (see reviews by Linde 1994 and Guth & Kaiser 2005). Such stretching is predicted by a wide class of theories of elementary particles, and all available evidence bears it out. Much of space is stretching and will continue doing so forever, but some regions of space stop stretching and form distinct bubbles, like gas pockets in a loaf of rising bread. Infinitely many such bubbles emerge (Figure 1, lower left, with time increasing upwards). Each is an embryonic Level I multiverse: infinite in size[^7] and filled with matter deposited by the energy field that drove inflation. Recent cosmological measurements have confirmed two key predictions of inflation: that space has negligible curvature and that the clumpiness in the cosmic matter distribution use to be approximately scale invariant. What are Level II parallel universes like? ------------------------------------------ The prevailing view is that the physics we observe today is merely a low-energy limit of a much more general theory that manifests itself at extremely high temperatures. For example, this underlying fundamental theory may be 10-dimensional, supersymmetric and involving a grand unification of the four fundamental forces of nature. A common feature in such theories is that the potential energy of the field(s) relevant to inflation has many different minima (sometimes called “metastable vacuum states”), and ending up in different minima corresponds to different effective laws of physics for our low-energy world. For instance, all but three spatial dimensions could be curled up (“compactified”) on a tiny scale, resulting in an effectively three-dimensional space like ours, or fewer could curl up leaving a 5-dimensional space. Quantum fluctuations during inflation can therefore cause different post-inflation bubbles in the Level II multiverse to end up with different effective laws of physics in different bubbles — say different dimensionality or different types of elementary particles, like two rather than three generations of quarks. In addition to such discrete properties as dimensionality and particle content, our universe is characterized by a set of dimensionless numbers known as [*physical constants*]{}. Examples include the electron/proton mass ratio $m_p/m_e\approx 1836$ and the cosmological constant, which appears to be about $10^{-123}$ in so-called Planck units. There are models where also such non-integer parameters can vary from one one post-inflationary bubble to another.[^8] In summary, the Level II multiverse is likely to be more diverse than the Level I multiverse, containing domains where not only the initial conditions differ, but perhaps the dimensionality, the elementary particles and the physical constants differ as well. This is currently a very active research area. The possibility of a string theory “landscape” (Bousso & Polchinski 2000; Susskind 2003), where the above-mentioned potential has perhaps $10^{500}$ different minima, may offer a specific realization of the Level II multiverse which would in turn have four sub-levels of increasing diversity: [**IId:**]{} different ways in which space can be compactified, which can allow both different effective dimensionality and different symmetries/elementary articles (corresponding to different topology of the curled up extra dimensions). [**IIc:**]{} different “fluxes” (generalized magnetic fields) that stabilize the extra dimensions (this sublevel is where the largest number of choices enter, perhaps $10^{500}$). [**IIb:**]{} once these two choices have been made, there may be a handful of different minima in the effective supergravity potential. [**IIa:**]{} the same minimum and effective laws of physics can be realized in a many different post-inflationary bubbles, each constituting a Level I multiverse. Before moving on, let us briefly comment on a few closely related multiverse notions. First of all, if one Level II multiverse can exist, eternally self-reproducing in a fractal pattern, then there may well be infinitely many other Level II multiverses that are completely disconnected. However, this variant appears to be untestable, since it would neither add any qualitatively different worlds nor alter the probability distribution for their properties. All possible initial initial conditions and symmetry breakings are already realized within each one. An idea proposed by Tolman and Wheeler and recently elaborated by Steinhardt & Turok (2002) is that the (Level I) multiverse is cyclic, going through an infinite series of Big Bangs. If it exists, the ensemble of such incarnations would also form a multiverse, arguably with a diversity similar to that of Level II. An idea proposed by Smolin (1997) involves an ensemble similar in diversity to that of Level II, but mutating and sprouting new universes through black holes rather than during inflation. This predicts a form of a natural selection favoring universes with maximal black hole production. In braneworld scenarios, another 3-dimensional world could be quite literally parallel to ours, merely offset in a higher dimension. However, it is unclear whether such a world (“brane”) deserves be be called a parallel universe separate from our own, since we may be able to interact with it gravitationally much as we do with dark matter. -0.4cm -0.1cm Fine-tuning and selection effects {#TuningSec} --------------------------------- Although we cannot interact with other Level II parallel universes, cosmologists can infer their presence indirectly, because their existence can account for unexplained coincidences in our universe. To give an analogy, suppose you check into a hotel, are assigned room 1967 and note that this is the year you were born. What a coincidence, you say. After a moment of reflection, however, you conclude that this is not so surprising after all. The hotel has hundreds of rooms, and you would not have been having these thoughts in the first place if you had been assigned one with a number that meant nothing to you. The lesson is that even if you knew nothing about hotels, you could infer the existence of other hotel rooms to explain the coincidence. As a more pertinent example, consider the mass of the sun. The mass of a star determines its luminosity, and using basic physics, one can compute that life as we know it on Earth is possible only if the sun’s mass falls into the narrow range between $1.6\times 10^{30}\kg$ and $2.4\times 10^{30}\kg$. Otherwise Earth’s climate would be colder than that of present-day Mars or hotter than that of present-day Venus. The measured solar mass is $M\sim 2.0\times 10^{30}\kg$. At first glance, this apparent coincidence of the habitable and observed mass values appears to be a wild stroke of luck. Stellar masses run from $10^{29}$ to $10^{32}\kg$, so if the sun acquired its mass at random, it had only a small chance of falling into the habitable range. But just as in the hotel example, one can explain this apparent coincidence by postulating an ensemble (in this case, a number of planetary systems) and a selection effect (the fact that we must find ourselves living on a habitable planet). Such observer-related selection effects are referred to as “anthropic” (Carted 1973), and although the “A-word” is notorious for triggering controversy, physicists broadly agree that these selection effects cannot be neglected when testing fundamental theories. In this weak sense, the anthropic principle is not optional. What applies to hotel rooms and planetary systems applies to parallel universes. Most, if not all, of the attributes set by symmetry breaking appear to be fine-tuned. Changing their values by modest amounts would have resulted in a qualitatively different universe–one in which we probably would not exist. If protons were $0.2\%$ heavier, they could decay into neutrons, destabilizing atoms. If the electromagnetic force were $4\%$ weaker, there would be no hydrogen and no normal stars. If the weak interaction were much weaker, hydrogen would not exist; if it were much stronger, supernovae would fail to seed interstellar space with heavy elements. If the cosmological constant were much larger, the universe would have blown itself apart before galaxies could form. Indeed, most if not all the parameters affecting low-energy physics appear fine-tuned at some level, in the sense that changing them by modest amounts results in a qualitatively different universe. Although the degree of fine-tuning is still debated (as exemplified in the rest of this book; see Barrow & Tipler 1986, Tegmark 1998 & Hogan (2000) for more technical reviews), these examples suggest the existence of parallel universes with other values of some physical constants. The existence of a Level II multiverse implies that physicists will never be able to determine the values of all physical constants from first principles. Rather, they will merely compute probability distributions for what they should expect to find, taking selection effects into account. The result should be as generic as is consistent with our existence. -0.3cm -0.6cm Level III: The many worlds of quantum physics ============================================= There may be a third type of parallel worlds that are not far away but in a sense right here. If the fundamental equations of physics are what mathematicians call [*unitary*]{}, as they so far appear to be, then the universe keeps branching into parallel universes as in the cartoon (, bottom): whenever a quantum event appears to have a random outcome, all outcomes in fact occur, one in each branch. This is the Level III multiverse. Although more debated and controversial than Level I and Level II, we will see that, surprisingly, this level adds no new types of universes. The quantum conundrum --------------------- In the early 20th century the theory of quantum mechanics revolutionized physics by explaining the atomic realm, which does not abide by the classical rules of Newtonian mechanics. Despite the obvious successes of the theory, a heated debate rages about what it really means. The theory specifies the state of the universe not in classical terms, such as the positions and velocities of all particles, but in terms of a mathematical object called a wave function. According to the Schrödinger equation, this state evolves over time in a fashion that mathematicians term “unitary”, meaning that the wave function rotates in an abstract infinite-dimensional space called Hilbert space. Although quantum mechanics is often described as inherently random and uncertain, the wave function evolves in a deterministic way. There is nothing random or uncertain about it. The sticky part is how to connect this wave function with what we observe. Many legitimate wave functions correspond to counterintuitive situations, such as a cat being dead and alive at the same time in a so-called superposition. In the 1920s physicists explained away this weirdness by postulating that the wave function “collapse” into some definite classical outcome whenever someone made an observation. This add-on had the virtue of explaining observations, but it turned an elegant, unitary theory into a kludgy, nonunitary one, since there was no equation specifying when or how this collapse occurred. The intrinsic randomness commonly ascribed to quantum mechanics is the result of this postulate, triggering Einstein’s objection that “God doesn’t play dice”. Over the years many physicists have abandoned this view in favor of one developed in 1957 by Princeton graduate student Hugh Everett III. He showed that the collapse postulate is unnecessary. Unadulterated quantum theory does not, in fact, pose any contradictions. Although it predicts that one classical reality gradually splits into superpositions of many such realities, observers subjectively experience this splitting merely as a slight randomness (), with probabilities in exact agreement with those from the old collapse postulate (de Witt 2003). This superposition of classical worlds is the Level III multiverse. What are Level III parallel universes like? ------------------------------------------- Everett’s many-worlds interpretation has been boggling minds inside and outside physics for more than four decades. But the theory becomes easier to grasp when one distinguishes between two ways of viewing a physical theory: the outside view of a physicist studying its mathematical equations, like a bird surveying a landscape from high above it, and the inside view of an observer living in the world described by the equations, like a frog living in the landscape surveyed by the bird.[^9]. From the bird perspective, the Level III multiverse is simple. There is only one wave function. It evolves smoothly and deterministically over time without any kind of splitting or parallelism. The abstract quantum world described by this evolving wave function contains within it a vast number of parallel classical story lines, continuously splitting and merging, as well as a number of quantum phenomena that lack a classical description. From their frog perspective, observers perceive only a tiny fraction of this full reality. They can view their own Level I universe, but a process called decoherence (Zeh 1970; Giulini 1996) — which mimics wave function collapse while preserving unitarity–prevents them from seeing Level III parallel copies of themselves. Whenever observers are asked a question, make a snap decision and give an answer, quantum effects in their brains lead to a superposition of outcomes, such as “Continue reading the article” and “Put down the article”. From the bird perspective, the act of making a decision causes a person to split into multiple copies: one who keeps on reading and one who doesn’t. From their frog perspective, however, each of these alter egos is unaware of the others and notices the branching merely as a slight randomness: a certain probability of continuing to read or not. As strange as this may sound, the exact same situation occurs even in the Level I multiverse. You have evidently decided to keep on reading the article, but one of your alter egos in a distant galaxy put down the magazine after the first paragraph. The only difference between Level I and Level III is where your doppelga[ä]{}ngers reside. In Level I they live elsewhere in good old three-dimensional space. In Level III they live on another quantum branch in infinite-dimensional Hilbert space (). Level III parallel universes: evidence & implications ----------------------------------------------------- The existence of Level III depends on one crucial assumption: that the time evolution of the wave function is unitary. So far experimenters have encountered no departures from unitarity. In the past few decades they have confirmed unitarity for ever larger systems, including carbon 60 buckyball molecules and kilometer-long optical fibers. On the theoretical side, the case for unitarity has been bolstered by the discovery of decoherence (see Tegmark & Wheeler 2001 for a popular review). Some theorists who work on quantum gravity have questioned unitarity; one concern is that evaporating black holes might destroy information, which would be a nonunitary process. But a recent breakthrough in string theory known as AdS/CFT correspondence suggests that even quantum gravity is unitary. If so, black holes do not destroy information but merely transmit it elsewhere. If physics is unitary, then the standard picture of how quantum fluctuations operated early in the big bang must change. These fluctuations did not generate initial conditions at random. Rather they generated a quantum superposition of all possible initial conditions, which coexisted simultaneously. Decoherence then caused these initial conditions to behave classically in separate quantum branches. Here is the crucial point: the distribution of outcomes on different quantum branches in a given Hubble volume (Level III) is identical to the distribution of outcomes in different Hubble volumes within a single quantum branch (Level I). This property of the quantum fluctuations is known in statistical mechanics as ergodicity. The same reasoning applies to Level II. The process of symmetry breaking did not produce a unique outcome but rather a superposition of all outcomes, which rapidly went their separate ways. So if physical constants, spacetime dimensionality and so on can vary among parallel quantum branches at Level III, then they will also vary among parallel universes at Level II. In other words, the Level III multiverse adds nothing new beyond Level I and Level II, just more indistinguishable copies of the same universes–the same old story lines playing out again and again in other quantum branches. The passionate debate about Everett’s theory therefore seems to be ending in a grand anticlimax, with the discovery of less controversial multiverses (Levels I and II) that are equally large. Needless to say, the implications are profound, and physicists are only beginning to explore them. For instance, consider the ramifications of the answer to a long-standing question: Does the number of universes exponentially increase over time? The surprising answer is no. From the bird perspective, there is of course only one quantum universe. From the frog perspective, what matters is the number of universes that are distinguishable at a given instant–that is, the number of noticeably different Hubble volumes. Imagine moving planets to random new locations, imagine having married someone else, and so on. At the quantum level, there are 10 to the $10^{118}$ universes with temperatures below $10^8$ kelvins. That is a vast number, but a finite one. From the frog perspective, the evolution of the wave function corresponds to a never-ending sliding from one of these $10$ to the $10^{118}$ states to another. Now you are in universe A, the one in which you are reading this sentence. Now you are in universe B, the one in which you are reading this other sentence. Put differently, universe B has an observer identical to one in universe A, except with an extra instant of memories. All possible states exist at every instant, so the passage of time may be in the eye of the beholder — an idea explored in Greg Egan’s 1994 science-fiction novel Permutation City and developed by physicist David Deutsch of the University of Oxford, independent physicist Julian Barbour, and others. The multiverse framework may thus prove essential to understanding the nature of time. -1.7cm Two world views --------------- The debate over how classical mechanics emerges from quantum mechanics continues, and the decoherence discovery has shown that there is a lot more to it than just letting Planck’s constant $\hbar$ shrink to zero. Yet as illustrates, this is just a small piece of a larger puzzle. Indeed, the endless debate over the interpretation of quantum mechanics — and even the broader issue of parallel universes — is in a sense the tip of an iceberg. In the Sci-Fi spoof “Hitchhiker’s Guide to the Galaxy”, the answer is discovered to be “42”, and the hard part is finding the real question. Questions about parallel universes may seem to be just about as deep as queries about reality can get. Yet there is a still deeper underlying question: there are two tenable but diametrically opposed paradigms regarding physical reality and the status of mathematics, a dichotomy that arguably goes as far back as Plato and Aristotle, and the question is which one is correct. - [**ARISTOTELIAN PARADIGM:**]{} The subjectively perceived frog perspective is physically real, and the bird perspective and all its mathematical language is merely a useful approximation. - [**PLATONIC PARADIGM:**]{} The bird perspective (the mathematical structure) is physically real, and the frog perspective and all the human language we use to describe it is merely a useful approximation for describing our subjective perceptions. What is more basic — the frog perspective or the bird perspective? What is more basic — human language or mathematical language? Your answer will determine how you feel about parallel universes. If you prefer the Platonic paradigm, you should find multiverses natural, since our feeling that say the Level III multiverse is “weird” merely reflects that the frog and bird perspectives are extremely different. We break the symmetry by calling the latter weird because we were all indoctrinated with the Aristotelian paradigm as children, long before we even heard of mathematics - the Platonic view is an acquired taste! In the second (Platonic) case, all of physics is ultimately a mathematics problem, since an infinitely intelligent mathematician given the fundamental equations of the cosmos could in principle [*compute*]{} the frog perspective, , compute what self-aware observers the universe would contain, what they would perceive, and what language they would invent to describe their perceptions to one another. In other words, there is a “Theory of Everything" (TOE) at the top of the tree in whose axioms are purely mathematical, since postulates in English regarding interpretation would be derivable and thus redundant. In the Aristotelian paradigm, on the other hand, there can never be a TOE, since one is ultimately just explaining certain verbal statements by other verbal statements — this is known as the infinite regress problem (Nozick 1981). Level IV: Other mathematical structures ======================================= Suppose you buy the Platonist paradigm and believe that there really is a TOE at the top of — and that we simply have not found the correct equations yet. Then an embarrassing question remains, as emphasized by John Archibald Wheeler: [*Why these particular equations, not others?*]{} Let us now explore the idea of mathematical democracy, whereby universes governed by other equations are equally real. This is the Level IV multiverse. First we need to digest two other ideas, however: the concept of a mathematical structure, and the notion that the physical world may be one. What is a mathematical structure? --------------------------------- Many of us think of mathematics as a bag of tricks that we learned in school for manipulating numbers. Yet most mathematicians have a very different view of their field. They study more abstract objects such as functions, sets, spaces and operators and try to prove theorems about the relations between them. Indeed, some modern mathematics papers are so abstract that the only numbers you will find in them are the page numbers! What does a dodecahedron have in common with a set of complex numbers? Despite the plethora of mathematical structures with intimidating names like orbifolds and Killing fields, a striking underlying unity has emerged in the last century: [*all*]{} mathematical structures are just special cases of one and the same thing: so-called formal systems. A formal system consists of abstract symbols and rules for manipulating them, specifying how new strings of symbols referred to as theorems can be derived from given ones referred to as axioms. This historical development represented a form of deconstructionism, since it stripped away all meaning and interpretation that had traditionally been given to mathematical structures and distilled out only the abstract relations capturing their very essence. As a result, computers can now prove theorems about geometry without having any physical intuition whatsoever about what space is like. shows some of the most basic mathematical structures and their interrelations. Although this family tree probably extends indefinitely, it illustrates that there is nothing fuzzy about mathematical structures. They are “out there” in the sense that mathematicians discover them rather than create them, and that contemplative alien civilizations would find the same structures (a theorem is true regardless of whether it is proven by a human, a computer or an alien). The possibility that the physical world is a mathematical structure ------------------------------------------------------------------- Let us now digest the idea that physical world (specifically, the Level III multiverse) [*is*]{} a mathematical structure. Although traditionally taken for granted by many theoretical physicists, this is a deep and far-reaching notion. It means that mathematical equations describe not merely some limited aspects of the physical world, but [*all*]{} aspects of it. It means that there is some mathematical structure that is what mathematicians call [*isomorphic*]{} (and hence equivalent) to our physical world, with each physical entity having a unique counterpart in the mathematical structure and vice versa. Let us consider some examples. A century ago, when classical physics still reigned supreme, many scientists believed that physical space was isomorphic to the mathematical structure known as $\R^3$: three-dimensional Euclidean space. Moreover, some thought that all forms of matter in the universe corresponded to various classical [*fields*]{}: the electric field, the magnetic field and perhaps a few undiscovered ones, mathematically corresponding to functions on $\R^3$ (a handful of numbers at each point in space). In this view (later proven incorrect), dense clumps of matter like atoms were simply regions in space where some fields were strong (where some numbers were large). These fields evolved deterministically over time according to some partial differential equations, and observers perceived this as things moving around and events taking place. Could, then, fields in three-dimensional space be the mathematical structure corresponding to the universe? No, since a mathematical structure cannot change — it is an abstract, immutable entity existing outside of space and time. Our familiar frog perspective of a three-dimensional space where events unfold is equivalent, from the bird perspective, to a four-dimensional spacetime where all of history is contained, so the mathematical structure would be fields in four-dimensional space. In other words, if history were a movie, the mathematical structure would not correspond to a single frame of it, but to the entire videotape. Given a mathematical structure, we will say that it has [*physical existence*]{} if any self-aware substructure (SAS) within it subjectively, from its frog perspective, perceives itself as living in a physically real world. What would, mathematically, such an SAS be like? In the classical physics example above, an SAS such as you would be a tube through spacetime, a thick version of what Einstein referred to as a world-line. The location of the tube would specify your position in space at different times. Within the tube, the fields would exhibit certain complex behavior, corresponding to storing and processing information about the field-values in the surroundings, and at each position along the tube, these processes would give rise to the familiar but mysterious sensation of self-awareness.From its frog perspective, the SAS would perceive this one-dimensional string of perceptions along the tube as passage of time. Although our example illustrates the idea of how our physical world can [*be*]{} a mathematical structure, this particular mathematical structure (fields in four-dimensional space) is now known to be the wrong one. After realizing that spacetime could be curved, Einstein doggedly searched for a so-called unified field theory where the universe was what mathematicians call a 3+1-dimensional pseudo-Riemannian manifold with tensor fields (top center in ), but this failed to account for the observed behavior of atoms. According to quantum field theory, the modern synthesis of special relativity theory and quantum theory, the universe (in this case the Level III multiverse) is a mathematical structure known as an algebra of operator-valued fields (top right in ). Here the question of what constitutes an SAS is more subtle (Tegmark 2000). However, this fails to describe black hole evaporation, the first instance of the Big Bang and other quantum gravity phenomena, so the true mathematical structure isomorphic to our universe, if it exists, has not yet been found. Mathematical democracy ---------------------- Now suppose that our physical world really is a mathematical structure, and that you are an SAS within it. This means that in the Mathematics tree of , one of the boxes is our universe. (The full tree is probably infinite in extent, so our particular box is not one of the few boxes from the bottom of the tree that are shown.) In other words, this particular mathematical structure enjoys not only mathematical existence, but physical existence as well. What about all the other boxes in the tree? Do they too enjoy physical existence? If not, there would be a fundamental, unexplained ontological asymmetry built into the very heart of reality, splitting mathematical structures into two classes: those with and without physical existence. As a way out of this philosophical conundrum, I have suggested (Tegmark 1998) that complete mathematical democracy holds: that mathematical existence and physical existence are equivalent, so that [*all*]{} mathematical structures exist physically as well. This is the Level IV multiverse. It can be viewed as a form of radical Platonism, asserting that the mathematical structures in Plato’s [*realm of ideas*]{}, the [*Mindscape*]{} of Rucker (1982), exist “out there” in a physical sense (Davies 1993), casting the so-called modal realism theory of David Lewis (1986) in mathematical terms akin to what Barrow (1991; 1992) refers to as “$\pi$ in the sky”. If this theory is correct, then since it has no free parameters, all properties of all parallel universes (including the subjective perceptions of SASs in them) could in principle be derived by an infinitely intelligent mathematician. Evidence for a Level IV multiverse ---------------------------------- We have described the four levels of parallel universes in order of increasing speculativeness, so why should we believe in Level IV? Logically, it rests on two separate assumptions: - [**Assumption 1:**]{} That the physical world (specifically our level III multiverse) is a mathematical structure - [**Assumption 2:**]{} Mathematical democracy: that all mathematical structures exist “out there” in the same sense In a famous essay, Wigner (1967) argued that “the enormous usefulness of mathematics in the natural sciences is something bordering on the mysterious", and that “there is no rational explanation for it". This argument can be taken as support for assumption 1: here the utility of mathematics for describing the physical world is a natural consequence of the fact that the latter [*is*]{} a mathematical structure, and we are simply uncovering this bit by bit. The various approximations that constitute our current physics theories are successful because simple mathematical structures can provide good approximations of how a SAS will perceive more complex mathematical structures. In other words, our successful theories are not mathematics approximating physics, but mathematics approximating mathematics. Wigner’s observation is unlikely to be based on fluke coincidences, since far more mathematical regularity in nature has been discovered in the decades since he made it, including the standard model of particle physics. A second argument supporting assumption 1 is that abstract mathematics is so general that [*any*]{} TOE that is definable in purely formal terms (independent of vague human terminology) is also a mathematical structure. For instance, a TOE involving a set of different types of entities (denoted by words, say) and relations between them (denoted by additional words) is nothing but what mathematicians call a set-theoretical model, and one can generally find a formal system that it is a model of. This argument also makes assumption 2 more appealing, since it implies that [*any*]{} conceivable parallel universe theory can be described at Level IV. The Level IV multiverse, termed the “ultimate Ensemble theory” in Tegmark (1997) since it subsumes all other ensembles, therefore brings closure to the hierarchy of multiverses, and there cannot be say a Level V. Considering an ensemble of mathematical structures does not add anything new, since this is still just another mathematical structure. What about the frequently discussed notion that the universe is a computer simulation? This idea occurs frequently in science fiction and has been substantially elaborated (, Schmidthuber 1997; Wolfram 2002). The information content (memory state) of a digital computer is a string of bits, say “$1001011100111001...$” of great but finite length, equivalent to some large but finite integer $n$ written in binary. The information processing of a computer is a deterministic rule for changing each memory state into another (applied over and over again), so mathematically, it is simply a function $f$ mapping the integers onto themselves that gets iterated: $n\mapsto f(n)\mapsto f(f(n))\mapsto...$. In other words, even the most sophisticated computer simulation is just yet another special case of a mathematical structure, and is already included in the Level IV multiverse. (Incidentally, iterating continuous functions rather than integer-valued ones can give rise to fractals.) A second argument for assumption 2 is that if two entities are isomorphic, then there is no meaningful sense in which they are not one and the same (Cohen 2003). This implies assumption 2 when the entities in question are a physical universe and a mathematical structure describing it, respectively. To avoid this conclusion that mathematical and physical existence are equivalent, one would need to argue that our universe is somehow made of stuff perfectly described by a mathematical structure, but which also has other properties that are not described by it. However, this violates assumption 1 and implies either that it is isomorphic to a more complicated mathematical structure or that it is not mathematical at all. The latter would be make Karl Popper turn in his grave, since those additional bells and whistles that make the universe non-mathematical by definition have no observable effects whatsoever. Another appealing feature of assumption 2 is that it provides the only answer so far to Wheeler’s question: [*Why these particular equations, not others?*]{} Having universes dance to the tune of all possible equations also resolves the fine-tuning problem of once and for all, even at the fundamental equation level: although many if not most mathematical structures are likely to be dead and devoid of SASs, failing to provide the complexity, stability and predictability that SASs require, we of course expect to find with 100% probability that we inhabit a mathematical structure capable of supporting life. Because of this selection effect, the answer to the question “what is it that breathes fire into the equations and makes a universe for them to describe?” (Hawking 1993) would then be “you, the SAS”. What are Level IV parallel universes like? ------------------------------------------ The way we use, test and potentially rule out any theory is to compute probability distributions for our future perceptions given our past perceptions and to compare these predictions with our observed outcome. In a multiverse theory, there is typically more than one SAS that has experienced a past life identical to yours, so there is no way to determine which one is you. To make predictions, you therefore have to compute what fractions of them will perceive what in the future, which leads to the following predictions: - [**Prediction 1:**]{} The mathematical structure describing our world is the most generic one that is consistent with our observations. - [**Prediction 2:**]{} Our future observations are the most generic ones that are consistent with our past observations. - [**Prediction 3:**]{} Our past observations are the most generic ones that are consistent with our existence. We will return to the problem of what “generic” means in (the measure problem). However, one striking feature of mathematical structures, discussed in detail in Tegmark (1997), is that the sort of symmetry and invariance properties that are responsible for the simplicity and orderliness of our universe tend to be generic, more the rule than the exception — mathematical structures tend to have them by default, and complicated additional axioms must be added to make them go away. In other words, because of both this and selection effects, we should not necessarily expect life in the Level IV multiverse to be a disordered mess. Discussion ========== We have seen that scientific theories of parallel universes form a four-level hierarchy, in which universes become progressively more different from ours. They might have different initial conditions (Level I), different effective physical laws, constants and particles (Level II), or different fundamental physical laws (Level IV). It is ironic that Level III is the one that has drawn the most fire in the past decades, because it is the only one that adds no qualitatively new types of universes. Whereas the Level I universes join seemlessly, there are clear demarcations between those within levels II and III caused by inflating space and decoherence, respectively. The level IV universes are completely disconnected and need to be considered together only for predicting your future, since “you” may exist in more than one of them. Future prospects ---------------- There are ample future prospects for testing and perhaps ruling out these multiverse theories. In the coming decade, dramatically improved cosmological measurements of the microwave background radiation, the large-scale matter distribution, , will test Level I by further constraining the curvature and topology of space and will test level II by providing stringent tests of inflation. Progress in both astrophysics and high-energy physics should also clarify the extent to which various physical constants are fine-tuned, thereby weakening or strengthening the case for Level II. If the current world-wide effort to build quantum computers succeeds, it will provide further evidence for Level III, since they would, in essence, be exploiting the parallelism of the Level III multiverse for parallel computation (Deutsch 1997). Conversely, experimental evidence of unitarity violation would rule out Level III. Finally, success or failure in the grand challenge of modern physics, unifying general relativity and quantum field theory, will shed more light on Level IV. Either we will eventually find a mathematical structure matching our universe, or we will bump up against a limit to the unreasonable effectiveness of mathematics and have to abandon Level IV. The measure problem {#MeasureSec} ------------------- There are also interesting theoretical issues to resolve within the multiverse theories, first and foremost the [*measure problem*]{}. As multiverse theories gain credence, the sticky issue of how to compute probabilities in physics is growing from a minor nuisance into a major embarrassment. If there are indeed many identical copies of you, the traditional notion of determinism evaporates. You could not compute your own future even if you had complete knowledge of the entire state of the multiverse, because there is no way for you to determine which of these copies is you (they all feel they are). All you can predict, therefore, are probabilities for what you would observe. If an outcome has a probability of, say, 50 percent, it means that half the observers observe that outcome. Unfortunately, it is not an easy task to compute what fraction of the infinitely many observers perceive what. The answer depends on the order in which you count them. By analogy, the fraction of the integers that are even is 50 percent if you order them numerically (1, 2, 3, 4, ...) but approaches 100 percent if you sort them digit by digit, the way your word processor would (1, 10, 100, 1,000, ...). When observers reside in disconnected universes, there is no obviously natural way in which to order them. Instead one must sample from the different universes with some statistical weights referred to by mathematicians as a “measure”. This problem crops up in a mild and treatable manner at Level I, becomes severe at Level II (see Tegmark 2004 for a detailed review), has caused much debate at Level III (de Witt 2003, Mukhanov 2005), and is horrendous at Level IV. At Level II, for instance, Linde, Vilenkin and others have published predictions for the probability distributions of various cosmological parameters. They have argued that different parallel universes that have inflated by different amounts should be given statistical weights proportional to their volume (, Garriga & Vilenkin 2001a). On the other hand, any mathematician will tell you that $2\times\infty=\infty$, so there is no objective sense in which an infinite universe that has expanded by a factor of two has gotten larger. Moreover, a finite universe with the topology of a torus is equivalent to a perfectly periodic universe with infinite volume, both from the mathematical bird perspective and from the frog perspective of an observer within it. So why should its infinitely smaller volume give it zero statistical weight? After all, even in the Level I multiverse, Hubble volumes start repeating (albeit in a random order, not periodically) after about 10 to the $10^{118}$ meters. If you think that is bad, consider the problem of assigning statistical weights to different mathematical structures at Level IV. The fact that our universe seems relatively simple has led many people to suggest that the correct measure somehow involves complexity. The pros and cons of parallel universes --------------------------------------- So should you believe in parallel universes? We have seen that this is not a yes/no question — rather, the most interesting issue is whether there are 0, 1, 2, 3 or 4 levels of multiverses. Figure 1 summarizes the evidence that we have discussed for the different levels. The principal arguments against them are that they are wasteful and that they are weird. The wastefulness argument is that multiverse theories are vulnerable to Occam’s razor because they postulate the existence of other worlds that we can never observe. Why should nature be so wasteful and indulge in such opulence as an infinity of different worlds? Yet this argument can be turned around to argue for a multiverse. What precisely would nature be wasting? Certainly not space, mass or atoms–the uncontroversial Level I multiverse already contains an infinite amount of all three, so who cares if nature wastes some more? The real issue here is the apparent reduction in simplicity. A skeptic worries about all the information necessary to specify all those unseen worlds. But an entire ensemble is often much simpler than one of its members. This principle can be stated more formally using the notion of algorithmic information content. The algorithmic information content in a number is, roughly speaking, the length of the shortest computer program that will produce that number as output. For example, consider the set of all integers. Which is simpler, the whole set or just one number? Naively, you might think that a single number is simpler, but the entire set can be generated by quite a trivial computer program, whereas a single number can be hugely long. Therefore, the whole set is actually simpler. Similarly, the set of all solutions to Einstein’s field equations is simpler than a specific solution. The former is described by a few equations, whereas the latter requires the specification of vast amounts of initial data on some hypersurface. The lesson is that complexity increases when we restrict our attention to one particular element in an ensemble, thereby losing the symmetry and simplicity that were inherent in the totality of all the elements taken together. In this sense, the higher-level multiverses are simpler. Going from our universe to the Level I multiverse eliminates the need to specify initial conditions, upgrading to Level II eliminates the need to specify physical constants, and the Level IV multiverse eliminates the need to specify anything at all. The opulence of complexity is all in the subjective perceptions of observers (Tegmark 1996) — the frog perspective. From the bird perspective, the multiverse could hardly be any simpler. The complaint about weirdness is aesthetic rather than scientific, and it really makes sense only in the Aristotelian worldview. Yet what did we expect? When we ask a profound question about the nature of reality, do we not expect an answer that sounds strange? Evolution provided us with intuition for the everyday physics that had survival value for our distant ancestors, so whenever we venture beyond the everyday world, we should expect it to seem bizarre. Thanks to clever inventions, we have glimpsed slightly beyond the frog perspective of our normal inside view, and sure enough, we have encountered bizarre phenomena whenever departing from human scales in any way: at high speeds (time slows down), on small scales (quantum particles can be at several places at once), on large scales (black holes), at low temperatures (liquid Helium can flow upward), at high temperatures (colliding particles can change identity), . A common feature of all four multiverse levels is that the simplest and arguably most elegant theory involves parallel universes by default. To deny the existence of those universes, one needs to complicate the theory by adding experimentally unsupported processes and ad hoc postulates: finite space, wave function collapse, ontological asymmetry, Our judgment therefore comes down to which we find more wasteful and inelegant: many worlds or many words. Perhaps we will gradually get more used to the weird ways of our cosmos, and even find its strangeness to be part of its charm. The author wishes to thank Anthony Aguirre, Aaron Classens, Marius Cohen, Angelica de Oliveira-Costa, Alan Guth, Shamit Kachru, Andrei Linde, George Musser, David Raub, Martin Rees, Harold Shapiro, Alex Vilenkin and Frank Wilczek for stimulating discussions and George Musser for extensive editing and improvement of the text of a shorter predecessor of this article that was published in [*Scientific American*]{}. This work was supported by NSF grants AST-0071213 & AST-0134999, NASA grants NAG5-9194 & NAG5-11099, a fellowship from the David and Lucile Packard Foundation and a Cottrell Scholarship from Research Corporation. References ========== Anglin J RZurek W H;1996;Phys. Rev. D;53;7327 ;Barbour J B;2001;The End of Time;Oxford Univ. Press;Oxford Barrow J D;1991;Theories of Everything;Ballantine;[New York]{} Barrow J D;1992;Pi in the Sky;Clarendon;Oxford Barrow J DTipler F J;1986;The Anthropic Cosmological Principle;Clarendon;Oxford Bousso RPolchinski J;2000;JHEP;6;6 Brundrit G B;1979;Q. J. Royal Astr. Soc.;20;37 Bucher M ASpergel D N 1999, [*Sci. Am.*]{} [**1/1999**]{} Carter B;1974;IAU Symposium 63;Longair S;Reidel;Dordrecht Chaitin G J;1987;Algorithmic Information Theory;Cambridge U. P;Cambridge Cohen M 2003, Master’s thesis, Dept. of Philosophy, Ben Gurion University of the Negev, Israel Davies P C W;1982;The Accidental Universe;Cambridge U. P;Cambridge Davies P;1993;The Mind of God;Touchstone;[New York]{} Davies P;1996;Are We Alone?;Basic Books;[New York]{} Deutsch D;1997;The Fabric of Reality;Allen Lane;[New York]{} Deutsch D;2003;Science and Ultimate Reality: From Quantum to Cosmos; Barrow J D, Davies P C WHarper C L;Cambridge Univ. Press;Cambridge B;2003;Science and Ultimate Reality: From Quantum to Cosmos; Barrow J D, Davies P C WHarper C L;Cambridge Univ. Press;Cambridge Egan G;1995;Permutation City;Harper;[New York]{} Garriga JVilenkin A;2001a;Phys. Rev. D;64;023507 Garriga JVilenkin A;2001b;Phys. Rev. D;64;043511. Giulini D, Joos E, Kiefer C, Kupsch J, Stamatescu I OZeh H D;1996;Decoherence and the Appearance of a Classical World in Quantum Theory;Berlin;Springer Guth ASteinhardt P J;1984;Sci. Am.;250;116 Hawking S;1993;A Brief History of Time;Touchstone;[New York]{} Lewis D;1986;On the Plurality of Worlds;Blackwell;Oxford Linde A;1994;Sci. Am.;271;32 Maldacena J;2003;Science and Ultimate Reality: From Quantum to Cosmos; Barrow J D, Davies P C WHarper C L;Cambridge Univ. Press;Cambridge Nozick R;1981;Philosophical Explanations;Harvard Univ. Press;Cambridge Rees M J;2002;Our Cosmic Habitat;Princeton Univ. Press;Princeton Rucker R;1982;Infinity and the Mind;Birkhauser;Boston Schmidthuber J 1997, [*A Computer Scientist’s View of Life, the Universe, and Everything*]{}, in [*Foundations of Computer Science: Potential-Theory-Cognition, Lecture Notes in Computer Science*]{}, C. Freksa, ed., (Springer: Berlin), http://www.idsia.ch/ juergen/everything/html.html Smolin L;1997;The Life of the Cosmos;Oxford Univ. Press;Oxford Steinhardt P JTurok N;2002;Science;296;1436 Susskind L;2003;hep-th/0302219 Tegmark M;1996;Found. Phys. Lett.;9;25 Tegmark M;1997;Class. Quant. Grav.;14;L69 Tegmark M;1998;Ann. Phys.;270;1 Tegmark M;2000;Phys. Rev. E;61;4194 Tegmark M;2002;Science;296;1427 Tegmark MWheeler J A;2001;Sci.Am.;2/2001;[68-75]{} Tegmark M;2004;astro-ph/0410281 Wigner E P;1967;Symmetries and Reflections;MIT Press;Cambridge Wolfram S;2002;A New Kind of Science;Wolfram Media;[New York]{} Zeh H D;1970;Found. Phys.;1;69 Zurek W;2003;Science and Ultimate Reality: From Quantum to Cosmos; Barrow J D, Davies P C WHarper C L;Cambridge Univ. Press;Cambridge [^1]: After emitting the light that is now reaching us, the most distant things we can see have receded because of the cosmic expansion, and are now about about 40 billion light years away. [^2]: As described below, the mathematically simplest version of quantum mechanics is “unitary”, lacking the controversial process known as wavefunction collapse. [^3]: Bruno’s ideas have since been elaborated by, , Brundrit (1979), Garriga & Vilenkin (2001b) and Ellis (2002), all of whom have thus far avoided the stake. [^4]: If the cosmic expansion continues to accelerate (currently an open question), the observable universe will eventually stop growing. [^5]: This is an extremely conservative estimate, simply counting all possible quantum states that a Hubble volume can have that are no hotter than $10^8$K. $10^{115}$ is roughly the number of protons that the Pauli exclusion principle would allow you to pack into a Hubble volume at this temperature (our own Hubble volume contains only about $10^{80}$ protons). Each of these $10^{115}$ slots can be either occupied or unoccupied, giving $N=2^{10^{115}}\sim 10^{10^{115}}$ possibilities, so the expected distance to the nearest identical Hubble volume is $N^{1/3}\sim 10^{10^{115}}$ Hubble radii $\sim 10^{10^{115}}$ meters. Your nearest copy is likely to be much closer than $10^{10^{29}}$ meters, since the planet formation and evolutionary processes that have tipped the odds in your favor are at work everywhere. There are probably at least $10^{20}$ habitable planets in our own Hubble volume alone. [^6]: Astronomical evidence suggests that the cosmic expansion is currently accelerating. If this acceleration continues, then even the level I parallel universes will remain forever separate, with the intervening space stretching faster than light can travel through it. The jury is still out, however, with popular models predicting that the universe will eventually stop accelerating and perhaps even recollapse. [^7]: Surprisingly, it has been shown that inflation can produce an infinite Level I multiverse even in a bubble of finite spatial volume, thanks to an effect whereby the spatial directions of spacetime curve towards the (infinite) time direction (Bucher & Spergel 1999). [^8]: Although the fundamental equations of physics are the same throughout the Level II multiverse, the approximate effective equations governing the low-energy world that we observe will differ. For instance, moving from a three-dimensional to a four-dimensional (non-compactified) space changes the observed gravitational force equation from an inverse square law to an inverse cube law. Likewise, breaking the underlying symmetries of particle physics differently will change the lineup of elementary particles and the effective equations that describe them. However, we will reserve the terms “different equations” and “different laws of physics” for the Level IV multiverse, where it is the fundamental rather than effective equations that change. [^9]: Indeed, the standard mental picture of what the physical world is corresponds to a third intermediate viewpoint that could be termed the [*consensus view*]{}. From your subjectively perceived frog perspective, the world turns upside down when you stand on your head and disappears when you close your eyes, yet you subconsciously interpret your sensory inputs as though there is an external reality that is independent of your orientation, your location and your state of mind. It is striking that although this third view involves both censorship (like rejecting dreams), interpolation (as between eye-blinks) and extrapolation (say attributing existence to unseen cities) of your inside view, independent observers nonetheless appear to share this consensus view. Although the inside view looks black-and-white to a cat, iridescent to a bird seeing four primary colors, and still more different to bee a seeing polarized light, a bat using sonar, a blind person with keener touch and hearing, or the latest overpriced robotic vacuum cleaner, all agree on whether the door is open. The key current challenge in physics is deriving this semiclassical consensus view from the fundamental equations specifying the bird perspective. In my opinion, this means that although understanding the detailed nature of human consciousness is an important challenge in its own right, it is [*not*]{} necessary for a fundamental theory of physics.

ArXiv

--- author: - 'B. Plez' - 'V.Hill' - 'R. Cayrel' - 'M. Spite' - 'B. Barbuy' - 'T.C. Beers' - 'P. Bonifacio' - 'F. Primas' - 'B. Nordström' date: 'Received/Accepted' title: 'Lead abundance in the uranium star CS31082-001 [^1]' --- Introduction ============ The detection of uranium in an old, very metal-poor star of the galactic halo, BPS , was first reported in Cayrel et al. ([@CHB01]). A greatly improved analysis, Hill et al. ([@HPC02]), (quoted as paper I) was made possible by a redetermination of crucial atomic data by Nilsson et al. ([@NIJL02; @NZL02]). Hill et al. have determined the abundance of U ($\log(\mathrm{U/H})+12. =-1.92 \pm 0.11$) and of Th ($\log(\mathrm{Th/H}) + 12 = -0.98 \pm 0.05$), in the usual scale $\log(n_H)=12.0$, in . These abundances have been used as cosmo-chronometers, comparing them to theoretical estimates of the initial production ratio. The time $ \Delta t$ in Gyr elapsed from the formation of the two actinides to now, is linked to the production ratio $\mathrm{(U/Th)_0}$ and the ratio measured in the star $\mathrm{(U/Th)_{now}}$ by the simple relation : $ \Delta t = 21.76[ \log(\mathrm{(U/Th)_0}) - \log(\mathrm{(U/Th)_{now}})]$ where the coefficient 21.76 is derived from the half-lives of $^{232}$Th and $^{238}$U. The superiority of the pair U/Th over the pair Th/Eu has been amply demonstrated, for example in Goriely & Clerbaux ([@GC99]), Goriely & Arnould ([@GA01]), or Wanajo et al. ([@WII02], see their fig.7). As both U and Th decay to the stable element lead, it is of great interest to know the abundance of lead in the star. In Hill et al. ([@HPC02]), only an upper limit to the lead abundance was given, and we report here the result of a new observation obtained at ESO Paranal to get this abundance. The time requested was 17 hours, enough to detect the minimum amount of lead coming from the decay of the observed elements $^{238}$U and $^{232}$Th into $^{206}$Pb, and $^{208}$Pb, respectively. In addition, lead may come from other channels, in particular from the decay of $^{235}$U into $^{207}$Pb, and from more unstable nuclides decaying very quickly to lead, such as $^{234}$U. Observations and reduction procedure ==================================== The observations were carried out with the ESO VLT using the UVES spectrograph with image slicer \#2, leading to a spectral resolution of $\simeq$80000. A total of 13 exposures were collected in service mode, reaching a total exposure time of 17 hours. The signal-to-noise ratio of the combined spectrum is around 600 per pixel. The data were reduced using the standard UVES pipeline (Ballester et al. [@bal00]). The signal we were looking for is very weak: a depression of only $\simeq$0.5 per cent expected at 4057.807 Å, in the red wing of a weak CH line located at 4057.718 Å. After correcting each spectrum for radial-velocity shifts, several methods for combining the 13 spectra were tested, including (i) a straight average of the best 10 spectra (those with no cosmic hits in that wavelength region), (ii) averaging the spectra after clipping points further away than 2.5$\sigma$ from the median of the distribution for each pixel, and (iii) averaging the 9 spectra closest to the median of the distribution for each pixel. All methods yielded a very similar result in the Pb region, and we display only one of them (average of the nine spectra closest to the median) in Fig.  \[fig1\], where the error bars represent the photon noise for each pixel. [{height="8cm"} {height="8cm"}]{} Spectral synthesis and comparison with observations =================================================== We used the same model atmosphere and spectrum synthesis code ([*turbospectrum*]{}: Plez et al. [@PSL93], Alvarez & Plez [@AP98]) as in Paper I, achieving complete self-consistency between the model and the spectra computations. Our spectrum synthesis of the Pb 4057Å region is displayed in Fig. \[fig1\], where it is compared to the observations. In addition to the photon noise itself (error bars in the observed points in Fig.\[fig1\]), various sources of uncertainties on the Pb abundance determination were examined[^2]: [*(i)*]{} Continuum placement: there are two clean continuum windows close to the Pb line, in the 4057.90 – 4058.0Å and 4058.4 – 4058.5Å intervals, which are used to achieve the best normalization of the observations to the synthetic spectra. None of the two windows are perfectly clean, the first containing a very faint $\ion{Mn}{I}$ line at 4057.949Å which seems slightly underestimated in the synthesis, while the second has a $\ion{Co}{I}$ line at 4058.599Å, slightly overestimated in the synthesis. The two extreme normalizations differ by 0.17%, which leads to a maximum uncertainty on the Pb abundance of 0.15dex. [*(ii)*]{} The wavelengths precision of the Pb and the blending CH line at 4057.718Å also affects the Pb abundance determination, but to a much smaller extent. We have tested that a reasonable maximum shift of 0.005Å (0.35km/s) affects the Pb abundance by at most 0.05dex. [*(iii)*]{} The isotopic ratio of $^{206}$Pb, $^{207}$Pb and $^{208}$Pb can impact the Pb line shape, slightly changing its effective central wavelength and hence affecting the abundance determination. Following Van Eck et al. ([@vaneck2003]), we considered five Pb components, one for each of the even isotopes 208 and 206 and the three hyperfine components for $^{207}$Pb. All wavelengths and oscillator strengths were adopted from Van Eck et al. ([@vaneck2003]). However, the two extreme cases of isotope ratios that we have considered (see next section) did not produce any noticeable difference in the spectrum synthesis and hence the derived Pb abundances. Considering the best fit spectrum displayed in Fig. \[fig1\], and the various sources of uncertainty outlined above, the Pb abundance in is constrained to be $\log \epsilon\mathrm{(Pb)} =-0.55\pm 0.15$dex, 0.35 dex below our upper limit in paper I. In the next section we discuss this result with respect to former attempts to measure Pb abundance in other very old stars, and with respect to the amount of lead expected from the decay of the actinides Th and U. We do not attempt to explain our result by theoretical arguments, considering that it is the observational result that justify this Letter. A more complete discussion of the impact of this new measurement on nuclear astrophysics will be included in a forthcoming paper also dealing with the analysis of our newly completed HST/STIS observations of . Discussion ========== Comparison with former observations of similar stars ---------------------------------------------------- Other r-process enhanced stars have been searched for Pb, leading only to upper limits or very uncertain detections, whether using the very weak $\lambda$4058Å line or the intrinsically stronger (but observable only from space) UV line $\lambda$2833Å. We report in Table \[otherstars\] the $\log$(Pb/Th) ratios for and the two other stars with secure upper limits: (Th from Sneden et al. [@SCL03], Pb from Hill et al. [@HPC02]) and (Cowan et al. [@CS02]). For completeness, we note that Sneden et al. ([@SCB98]) detected Pb from the UV line in , but the line was affected by a spike and the authors themselves regarded the Pb abundance in this star as very uncertain. Th is taken as reference element for the Pb abundance, because of the direct connection between these elements. In the 3 stars, which are as old or older than globular clusters, about half of the initial Th content has decayed into $^{208}$Pb, and half has survived. The Pb/Th ratio detected in clearly stands out as an [*extremely low value*]{} compared to any upper limit so far placed on an r-process enriched star. Comparison with the amount expected from the decay of $^{232}$Th,$^{235}$U and $^{238}$U ---------------------------------------------------------------------------------------- Clear channels of production of lead are the decay of $^{232}$Th into $^{208}$Pb , of $^{238}$U into $^{206}$Pb, and of $^{235}$U into $^{207}$Pb. The amount of $^{208}$Pb is fixed by the observation of $^{232}$Th now, and the knowledge of the decayed fraction after the matter has been isolated in the atmosphere of the star, at a known rate. Star CS31082-001 CS22892-052 BD+$17\degr$3248 ------------- --------------- ------------- ------------------ $-$2.9 $-$3.1 $-$2.0 $\log$Pb/Th $0.43\pm0.16$ $\leq 1.57$ $\leq 1.4$ : Pb in and two other very metal-poor stars.[]{data-label="otherstars"} The epoch of the nucleosynthesis of the photospheric matter of CS31082-001, cannot be more than 13.7 $\pm$ 0.2 Gyr ago (Big Bang epoch according to WMAP results, Spergel et al. [@SV03]), and should be at least as much as the age of globular clusters (13.2 $\pm$ 1.5 Gyr according to Chaboyer [@CHA01]), taking into consideration the very low metallicity, $\mathrm{[Fe/H]}= -2.9$ of the star. The median is 13.5, also the age of first stellar formation according to Kogut et al. ([@KS03]). Adopting $t=13.5 \pm 1.5$ Gyr for the age of the actinides in CS 31082-001, we easily derive both the original content in $^{232}$Th and $^{238}$U, and the fraction of them transformed into $^{208}$Pb and $^{206}$Pb. For example: $$\epsilon(^{206}\mathrm{Pb}) = \epsilon\mathrm{(^{238}U)_{now}}\times (2^{(t/\tau)}-1)$$ with $\tau=4.47$ Gyr, the half-life of $^{238}$U. A similar formula holds for $^{232}$Th and $^{208}$Pb with $\tau =14.05$. But $^{235}$U cannot be treated the same way, as there is not enough $^{235}$U left to have an observed value. We must then rely on theoretical works, usually done for reproducing the solar system isotopic abundances, but not necessarily adequate for which has a clear overabundance of the actinides with respect to the lighter [*r*]{}-elements, compared to the solar system. However we can hope that in the restricted mass range under consideration , $^{232}$Th to $^{238}$U, neutron exposures producing the right ratio $^{238}$U/$^{232}$Th may also produce the right ratio $^{235}$U/$^{238}$U. In Tables 1 and 2 of Goriely & Arnould ([@GA01]), several neutron exposures are considered with a wide set of mass models. Forgetting the solar system, we keep the exposures giving the right $^{238}$U/$^{232}$Th production ratio for CS 31082-001, compatible with an age 13.5 $\pm$ 1.5 Gyr. With this constraint, the production ratio $R$ of $^{235}$U/$^{238}$U lies between 0.67 and 0.87. Taking the median 0.77 seems a reasonable estimate. The full amount of produced $^{235}$U is converted into $^{207}$Pb in , because of the fast decay of this isotope. Table \[isotopes\] summarizes our findings. Interestingly, the case $R=1.0$ gives an amount of total lead equal to the observed one, leaving no other channel for the production of [*r*]{}-lead. We have not included the $^{234}$U decay to $^{206}$Pb channel here, because of the lack of published estimates of the $^{234}$U/$^{238}$U production ratio. This channel (which could be as high as the $^{235}$U contribution) can only increase the contribution of the actinide-path to the total production of Pb, thereby reducing even further any other production channel. A warning is appropriate here: our analysis is based on the LTE approximation, which must be questioned, especially in the blue and the UV in very metal-poor giants, where the continuum is in a large part due to Rayleigh scattering. We examine this in the next subsection. ----------------------------- ---------------- ---------------- ---------------- $^{238}$U$_{\mathrm{now}}$ $^{206}$Pb $^{232}$Th$_{\mathrm{now}}$ $^{208}$Pb R 0.77 0.59 1.00 $^{207}$Pb 0.075 $\pm$.03 0.058$\pm$.02 0.098$\pm$.04 tot. Pb 0.26$\pm$0.10 0.243$\pm$0.09 0.283$\pm$.11 $\log(\epsilon$(Pb)) $-$0.59$\pm$.2 $-$0.61$\pm$.2 $-$0.55$\pm$.2 ----------------------------- ---------------- ---------------- ---------------- : Three abundance patterns examined for the lead feature. The abundances of $^{207}$Pb are computed for 3 assumed values of the production ratio R=$^{235}\mathrm{U}/^{238}\mathrm{U}$, the first considered as the most probable, and the other two chosen 30 per cent below and above. All abundances are with respect to $10^{12}$ hydrogen atoms. []{data-label="isotopes"} NLTE versus LTE --------------- It would be very useful to have a NLTE analysis of Pb, as the lower level of the measured line is very deep, part of the ground-level term, and is of Pb I when most of lead is in the Pb II stage. In metal-poor stars these deep levels tend to be over-ionized, but not always. If the deep levels are indeed over-ionized, the LTE assumption predicts too large a population of the lower level, and an underestimated abundance. In an attempt to estimate the size of possible NLTE effects, we checked the J$_\nu$/B$_\nu$ ratio at the position of the Pb I line, as well as at 4076Å and 2035Å, corresponding to the ionization limit of the upper and lower levels of the transition, respectively. In these layers, the ratio J$_\nu$/B$_\nu$ of the mean intensity to the local Planck function is larger than one, but remains on the order of two. It is unlikely that the NLTE correction is larger than this factor of two, so the production of Pb, outside the decay of $^{232}$Th, $^{235}$U, and $^{238}$U is bound to be less or equal to the actinide production. Digression about the rapid neutron capture lead in the solar system ------------------------------------------------------------------- Our result clearly concerns a particular class of objects: very metal-poor stars born in the early days of the Galaxy, and strongly enhanced in [*r*]{}-process elements. Is the observed low ratio Pb/Th particular to this class of objects, or is it a more general property of the [*r*]{}-process in the mass range 206-238? There is now a general agreement that the fraction of Pb produced by the [*r*]{}-process in the solar system it practically unknown, after the discovery that zero-metal stars can produce a lot of [*s*]{}-lead (Goriely & Siess [@GS01], Van Eck et al. [@vaneck2003]). This has modified the estimates of the amount of [*s*]{}-lead produced in the Galaxy before the birth of the Sun (Gallino et al [@GAB98]), and of the [*r*]{}-lead, obtained by subtracting the [*s*]{}-lead from the total lead. If the [*r*]{}-lead of the solar system were mainly produced by the decay of the actinides, as for CS 31082-001, it is easy to verify that the [*r*]{}-lead in the solar system would be of the order of only 1 to 3 % of the total lead. Conclusions =========== In one of the very metal-poor stars showing a large enhancement of [*r*]{}-process elements we have now a true determination of the lead abundance, instead of upper limits, only. This abundance is very low,$-0.55 \pm 0.15$ dex in , about one dex below the former upper limits in and . Also, our result shows that, in the purely $r$-process enriched photosphere of , most of lead results from the decay of $^{232}$Th , $^{235}$U, and $^{238}$U. This places a limit on the amount of $^{235}$U which has contributed to the production of $^{207}$Pb, as well as on that of $^{234}$U. A non-LTE analysis of the spectrum is highly desirable, but hampered so far by the lack of photoionization cross-sections for Pb I. We are indebted to Prof. R. Gallino for informations on the production of lead by the [*s*]{}-process. T.C.B. acknowledges partial funding from NSF grants AST 00-98508 and 00-98549, and PHY 02-16783.BN acknowledges support from the Carlsberg Foundation and the Nordic Academy for Advanced Studies. Alvarez, R., & Plez, B. 1998,  330, 1109 Ballester, P., Modigliani, A., Boitquin, O., et al. 2000, The Messenger, 101, 31 Chaboyer, B. 2001, in [*Astrophysical ages and time scales*]{} Ed. T. von Hippel et al., ASP conf.series vol. 245, p.162 Cayrel, R., Hill, V., Beers, T.C. et al. 2001, Nature, 409, 691 Cowan, J.J., Sneden C., Burles, S. et al. 2002,  572, 861 Gallino, R., Arlandini, C., Busso, M., et al. 1998,  497, 388 Goriely, S., & Arnould, M. 2001,  379, 1113 Goriely, S., & Clerbaux B. 1999,  346, 798 Goriely, S., Siess, L. 2001,  378, L25 Hill, V., Plez,B., Cayrel, R., et al. 2002,  387, 560 (paper I) Kogut, A., Spergel, D.N., Barnes,C. 2003,  148, 161 Nilsson, H., Ivarsson, S., Johansson, S., & Lundberg, H., 2002,  381, 1090 Nilsson, H., Zhang, Z.G., Lundberg, H., et al. 2002,  382, 368 Plez, B., Smith, V.V., Lambert, D.L. 1993,  418, 812 Sneden, C., Cowan, J.J., Burris, D.L. et al. 1998,  496, 235 Sneden, C., Cowan, J.J., Lawler, J.E. et al. 2003,  591, 936 Spergel, D.N., Verde, L., Peiris, H.V. et al. 2003,  148, 175 Van Eck, S., Goriely, S., Jorissen, A., Plez, B., 2003,  404, 291 Wanajo, S., Itoh, N., Ishimaru, Y., Nozawa, S., Beers, T. C., 2002,   577, 853 [^1]: Based on observations obtained with the Very Large Telescope of the European Southern Observatory at Paranal, Chile [^2]: In addition, we noted a small, but clearly missing absorption component, in the synthesis around 4057.63Å, bluewards of the CH line in Fig \[fig1\]. This feature is too far from the Pb line to affect its abundance, but to check the nature of this unidentified feature, we compared our spectrum with the spectra of C-rich stars of similar temperature, gravity, and metallicity, and found that these stars also exhibited absorption missing in the synthesis around 4057.63Å, with an amplitude clearly linked to the C and N abundances. We therefore conclude that it is a CH or CN molecular line.

ArXiv

--- author: - bibliography: - 'bib.bib' title: 'WURBench: Toward Benchmarking Wake-up Radio-based Systems' --- Wake-up radios, Benchmarking methodology, Cyber-physical systems, Internet of Things, WURBench

ArXiv

--- abstract: 'We introduce generalized Kazdan-Warner equations on Riemannian manifolds associated with a linear action of a torus on a complex vector space. We show the existence and the uniqueness of the solution of the equation on any compact Riemannian manifold. As an application, we give a direct proof of the Hitchin-Kobayashi correspondence for the solutions of the Abelian vortex equations on a compact Kähler manifold which are associated with a linear action of a torus on a complex vector space. Our generalized Kazdan-Warner equations for a special action of a torus give rise to the $tt^\ast$ equations [@CV1; @GL1].' author: - Natsuo Miyatake title: 'Generalized Kazdan-Warner equations associated with a linear action of a torus on a complex vector space' --- Introduction ============ Generalized Kazdan-Warner equations ----------------------------------- Let $K$ be a connected subtorus of a real torus $T^d:=\U(1)^d$ with the Lie algebra $k\subseteq t^d$. We denote by $\iota^\ast: (t^d)^\ast \rightarrow k^\ast$ the dual map of the inclusion map $\iota: k\rightarrow t^d$. Let $u_1,\dots, u_d$ be a basis of $t^d$ defined by $$\begin{aligned} u_1\coloneqq&(\inum, 0, \dots, 0), \\ u_2\coloneqq&(0, \inum, 0,\dots, 0), \\ &\cdots \\ u_d\coloneqq&(0,\dots, 0,\inum).\end{aligned}$$ We denote by $u^1, \dots, u^d\in (t^d)^\ast$ the dual basis of $u_1, \dots, u_d$. Let $(\cdot, \cdot)$ be the metric on $t^d$ and $(t^d)^\ast$ satisfying $$\begin{aligned} (u_i, u_j)=(u^i, u^j)=\delta_{ij} \ \text{for all $i, j$},\end{aligned}$$ where $\delta_{ij}$ denotes the Kronecker delta. Let $(M, \gM)$ be a Riemannian manifold. We denote by $\Delta_{g_M}$ the geometric Laplacian $d^\ast d$. In this paper, we introduce the following equation on $M$. $$\begin{aligned} \label{generalized KW equation} \Delta_{g_M}\xi+\sum_{j=1}^d a_je^{(\iota^\ast u^j, \xi)}\iota^\ast u^j=w,\end{aligned}$$ where $\xi$ is a $k^\ast$-valued function on $M$ which is the solution of (\[generalized KW equation\]) for given nonnegative functions $a_1, \dots, a_d$ and a $k^\ast$-valued function $w$. We give some examples of equation (\[generalized KW equation\]). Let $d=1$, $K=\U(1)$. Then equation (\[generalized KW equation\]) is the [*Kazdan-Warner equation*]{} [@KW1]: $$\begin{aligned} \Delta_{g_M} f+he^f=c.\end{aligned}$$ It should be noted that in [@KW1] the sign of a given function $h$ is not assumed to be nonnegative. \[toda lattice\] Let $K$ be $\{(g_1, \dots, g_d)\in T^d\mid g_1\cdots g_d=1\}$. We consider equation (\[generalized KW equation\]) on an open subset $U$ of the complex plane $\C\simeq\R^2$ with the standard metric $dx\otimes dx+dy\otimes dy$. We define a map $\pi: T^d \longrightarrow K $ by $$\begin{aligned} \pi(g_1,\dots, g_d)=(g_1^{-1}g_2, g_2^{-1}g_3, \dots, g_d^{-1}g_1).\end{aligned}$$ Then $k^\ast$ is identified with $\{\theta_1u^1+\cdots+\theta_du^d\in (t^d)^\ast\mid \theta_1+\cdots +\theta_d=0\}$ by taking the dual of the derivative $\pi_\ast:k\rightarrow t^d$. We set $a_1=\cdots =a_d=4, \ w=0$. Then equation (\[generalized KW equation\]) is the following: $$\begin{aligned} \sum_{j=1}^d \left\{\frac{\partial^2}{\partial z \partial \bar{z}}\xi_j+e^{\xi_{j+1}-\xi_j}-e^{\xi_j-\xi_{j-1}}\right\}u^j=0, \label{toda equation}\end{aligned}$$ where $\xi_1, \dots, \xi_d$ are real valued functions on $U$ satisfying $\xi_1+\cdots+\xi_d=0$. Equation (\[toda equation\]) is known as the [*two-dimensional periodic Toda lattice with opposite sign*]{} [@GL1]. Let $\epsilon_1, \epsilon_2\in\{1,2\}.$ Define $K\coloneqq\{(g_1^{-1}g_2, g_2^{-\epsilon_2}, g_1^{-1}g_2, g_1^{\epsilon_1})\in T^4\mid g_1, g_2\in \U(1)\}$. Let $(M, g_M)=(\C\backslash \{0\}, dx\otimes dx+dy\otimes dy)$. We set $a_1=2$, $a_2=\frac{4}{\epsilon_2}$, $ a_3=2$, $a_4=\frac{4}{\epsilon_1}, w=0$. Let $\gamma\in k^\ast$. We set the boundary conditions as follows: $$\begin{aligned} &\xi(z)=(\gamma+o(1))\log |z| \ \text{as} \ |z|\rightarrow 0, \\ &\xi(z)\rightarrow 0 \ \text{as} \ z\rightarrow \infty. \end{aligned}$$ An element $\theta\in k^\ast$ can be written as $\theta=\theta_1u^1+\theta_2u^2$ by identifying $k^\ast$ with $(t^2)^\ast$. Then equation (\[generalized KW equation\]) is the following. $$\begin{aligned} &\frac{\partial^2}{\partial z \partial \bar{z}}\xi_1=e^{\epsilon_1\xi_1}-e^{\xi_2-\xi_1}, \notag \\ &\frac{\partial^2}{\partial z \partial \bar{z}}\xi_2=e^{\xi_2-\xi_1}-e^{-\epsilon_2\xi_2} \label{ttstar}\end{aligned}$$ with boundary conditions $\xi_i(z)=(\gamma_i+o(1))\log |z|$ as $|z|\rightarrow 0$ and $\xi_i(z)\rightarrow 0$ as $z\rightarrow \infty$ for $i=1,2$. Equation (\[ttstar\]) is known as one of the [*$tt^\ast$ equations*]{} [@CV1; @GL1]. For all $\gamma$ which satisfy $(\iota^\ast u^j,\gamma)+2\geq 0$ for any $j\in \{1,\dots, 4\}$, the equation has a unique solution [@GIL1; @GL1]. Therefore equation (\[generalized KW equation\]) can be considered as a generalization of the above examples. We call equation (\[generalized KW equation\]) [*generalized Kazdan-Warner equation*]{}. We solve equation (\[generalized KW equation\]) on any compact Riemannian manifold under the following assumption on $a_1, \dots, a_d$: 1. \[coefficients\] For each $j\in J_a$, $a_j^{-1}(0)$ is a set of measure $0$ and $\log a_j$ is integrable, where $J_a$ denotes $\{j\in\{1, \dots, d\}\mid \text{$a_j$ is not identically 0}\}$. Note that if $M$ is a complex manifold with a holomorphic hermitian bundle $(E, h) \rightarrow M$, then $a_1=|\Phi_1|^2, \dots, a_d=|\Phi_d|^2$ satisfy the condition $(\ast)$ for any holomorphic sections $\Phi_1, \dots, \Phi_d$ of $E$. Our main theorem is the following. \[main theorem\] *Let $(M, g_M)$ be an $m$-dimensional compact connected Riemannian manifold. We take non-negative $C^\infty$ functions $a_1, \dots, a_d $ and a $k^\ast$-valued $C^\infty$ function $w$. Assume $a_1, \dots, a_d$ satisfy the above condition $(\ast)$. Then the following (1) and (2) are equivalent:* (1) \[GKW equation has a solution\]The generalized Kazdan-Warner equation has a $C^\infty$ solution $\xi:M\rightarrow k^\ast$: $$\begin{aligned} \label{KWequation} \Delta_{g_M}\xi+\sum_{j=1}^d a_je^{(\iota^\ast u^j, \xi)}\iota^\ast u^j=w;\end{aligned}$$ (2) \[wbar is in the interior of the cone\] The given functions $a_1, \dots, a_d$ and $w$ satisfy $$\begin{aligned} \int_M w \ \vol\in\sum_{j\in J_a}\R_{>0}\iota^\ast u^j,\end{aligned}$$ where $\mu_{g_M}$ denotes the measure induced by $g_M$. Moreover if $\xi$ and $\xi^\prime$ are $C^\infty$ solutions of equation (\[KWequation\]), then $\xi-\xi^\prime$ is a constant which is in the orthogonal complement of $\sum_{j\in J_a}\R\iota^\ast u^j$. \[trivial\] The statement (\[GKW equation has a solution\]) immediately implies the statement (\[wbar is in the interior of the cone\]): if (\[GKW equation has a solution\]) holds, by integrating both sides of equation (\[KWequation\]), we have $$\begin{aligned} \sum_{j=1}^d \left(\int_M a_je^{(\iota^\ast u^j, \xi)}\ \vol\right)\iota^\ast u^j=\int_M w \ \vol.\end{aligned}$$ Hence it suffices to solve equation (\[KWequation\]) under the assumption of (\[wbar is in the interior of the cone\]) and to prove the uniqueness of the solution up to a constant which is in the orthogonal complement of $\sum_{j\in J_a}\R\iota^\ast u^j$. Our generalized Kazdan-Warner equation is defined for a vector subspace $k\subseteq t^d$ which is generated by rational vectors. Clearly our equation can be generalized to any real vector subspace of $t^d$. For such a generalized equation we have the same theorem by the same proof as for Theorem \[main theorem\]. Abelian vortex equations ------------------------ We shall discuss a relationship between the Hitchin-Kobayashi correspondence for the solutions of the Abelian vortex equations and the generalized Kazdan-Warner equation. Here we mean by the Abelian vortex equations gauge theoretic equations on a Kähler manifold associated with the diagonal action of the torus $K$ which are defined as follows: Let $\mu_K:\C^d\rightarrow k^\ast$ be a moment map for the diagonal action of $K$ which is defined by $$\begin{aligned} \langle \mu_K(z), v \rangle=\frac{1}{2}g_{\R^{2d}}(\inum vz, z) \ \text{for $v\in k$},\end{aligned}$$ where we denote by $g_{\R^{2d}}(\cdot, \cdot)$ the standard metric of $\C^d\simeq \R^{2d}$, and by $\langle \cdot, \cdot \rangle$ the natural coupling. The moment map $\mu_K$ is also denoted as $$\begin{aligned} \mu_K(z)=-\frac{1}{2}\sum_{j=1}^d\iota^\ast u^j |z_j|^2 \ \ \text{for $z=(z_1,\dots, z_d)\in\C^d$}.\end{aligned}$$ Let $P_K\rightarrow(X, \omega_X)$ be a principal $K$-bundle over a compact connected Kähler manifold $(X, \omega_X)$ of complex dimension $n$ and $\A_K$ the space of $C^\infty$ connections on $P_K$. We denote by $S$ the vector bundle $P_K\times_K \C^d$, and by $\Gamma(S)$ the set of all $C^\infty$ sections on $X$. We fix a $k^\ast$ valued $C^\infty$ function on $X$ which is denoted by $\tau$. The Abelian vortex equations associated with $\tau$ are the following: $$\begin{aligned} \label{abelian vortex} \begin{cases} {} F_A^{0, 2}=0, &\\ \bar{\partial}_A\Phi=0, &\\ \Lambda_{\omega_X}F_A+\mu_K(\Phi)+\tau=0, & \end{cases}\end{aligned}$$ for $A\in \A_K$ and $\Phi\in \Gamma(S)$, where we denote by $\Lambda_{\omega_X}$ the adjoint of $\omega_X\wedge$. The first equation says that a connection $A$ defines a holomorphic structure on $S$ and the second equation says that $\Phi$ is a holomorphic section. In the third equation, we identify $k$ with $k^\ast$ by the metric. Suppose $X$ is a Riemann surface. Let $K_X\rightarrow X$ be the canonical bundle of $X$. We can slightly modify the above Abelian vortex equation as follows: $$\begin{aligned} \label{hitchin equation} \begin{cases} \bar{\partial}_A\Phi=0, &\\ F_A+\mu_K(\Phi)=0, \end{cases}\end{aligned}$$ for $A\in \A_K$ and $\Phi\in \Gamma(S\otimes K_X)$. Note that in this setting it is not necessary to take a Kähler form. Let $K$ be $\{(g_1, \dots, g_d)\in T^d\mid g_1\cdots g_d=1\}$ and $\pi:T^d\rightarrow K$ a group homomorphism defined in Example \[toda lattice\]. We take a principal $T^d$-bundle $P_{T^d}\rightarrow X$. We define a principal $K$-bundle $P_K$ by $P_K\coloneqq P_{T^d}/\ker \pi$. A section $\Phi=(\Phi_1, \dots, \Phi_d)$ of $(P_K\times_{K} \C^d)\otimes K_X$ can be regarded as a section of $(P_{T^d}\times_{\rm Ad} \End(\C^d))\otimes K_X$ by $$\begin{aligned} (\Phi_1,\dots, \Phi_d)\simeq \frac{1}{\sqrt{2}}\left( \begin{array}{cccc} 0 & & &\Phi_d\\ \Phi_1 & 0 & &\\ & \ddots& \ddots& \\ &&\Phi_{d-1}&0 \end{array} \right).\end{aligned}$$ Then a solution $(A, \Phi)$ of equation (\[hitchin equation\]) solves the following [*Hitchin’s self-duality equations*]{} [@Hit1] (see also [@AF1; @Bar1; @DL1]): $$\begin{aligned} \begin{cases} \bar{\partial}_A\Phi=0, &\\ F_A+[\Phi\wedge\Phi^\ast]=0.& \end{cases}\end{aligned}$$ Hitchin-Kobayashi correspondence -------------------------------- Let $K_\C$ be the complexification of $K$. We define an action of the complex gauge group $C^\infty(X, K_\C)$ on $\A_K\times\Gamma(S)$ by $$\begin{aligned} \sigma\cdot(A,\Phi)\coloneqq(A+\sigma^{-1}\partial \sigma+\bar{\sigma}\bar{\partial}\bar{\sigma}^{-1}, \bar{\sigma}\Phi) \ \text{for $(A, \Phi)\in\A_K\times\Gamma(S)$}.\end{aligned}$$ Then the following equations are preserved by the above action. $$\begin{aligned} \begin{cases} F_A^{0, 2}=0,& \\ \bar{\partial}_A\Phi=0.& \end{cases}\end{aligned}$$ Further in addition to the above two equations the real gauge group $C^\infty(X, K)$ also preserves the following equation: $$\begin{aligned} \Lambda_{\omega_X}F_A+\mu_K(\Phi)+\tau=0.\end{aligned}$$ Then we have the following two quotient spaces: $$\begin{aligned} &\M\coloneqq\{\text{All solutions of the Abelian vortex equations (\ref{abelian vortex})}\}/C^\infty(X, K), \\ &\M_\C\coloneqq\{(A,\Phi)\in\A_K\times \Gamma(S)\mid\text{$F_A^{0,2}=0, \ \bar{\partial}_A\Phi=0$}\}/C^\infty(X, K_\C),\end{aligned}$$ and we have a map $\M\rightarrow \M_\C$ which is induced by the inclusion map. The following correspondence is known as a variant of the [*Hitchin-Kobayashi correspondence*]{}. \[known theorem\]([@Ban1; @Ba1; @Mu1]). [*Let $\A_K^{1,1}$ be $\{A\in\A_K\mid F_A^{0,2}=0\}$. We define $\lambda_\tau\in k^\ast$ by $$\begin{aligned} \lambda_\tau\coloneqq\int_X\left(\Lambda_{\omega_X}F_A+\tau\right) \ \frac{\omega_X^n}{n!} \ \ \text{for $A\in\A_K^{1,1}$.}\end{aligned}$$ Note $\lambda_\tau$ does not depend on the choice of the connection. We define $\M_\C^{\tau-st}$ by $$\begin{aligned} \M_\C^{\tau-st}\coloneqq\{(A,\Phi)\in\A_K^{1,1}\times \Gamma(S)\mid\bar{\partial}_A\Phi=0, \ \lambda_\tau\in\sum_{j\in J_\Phi}\R_{>0}\iota^\ast u^j \}/C^\infty(X, K_\C),\end{aligned}$$ where $J_\Phi$ denotes $\{j\in\{1,\dots, d\}\mid \text{The $j$-th component of $\Phi$ is not identically 0}\}$. Then the map $\M\rightarrow \M_\C$ induces the following bijection. $$\begin{aligned} \M\longrightarrow\M_\C^{\tau-st}.\end{aligned}$$* ]{} By integrating both sides of the equation $$\begin{aligned} \Lambda_{\omega_X}F_A+\mu_K(\Phi)+\tau=0\end{aligned}$$ we can immediately check $\M_\C^{\tau-st}$ includes the image of the map $\M\rightarrow \M_\C$. It should be noted in [@Ban1; @Mu1] the correspondence was established for much more general settings than ours. Baptista [@Ba1 Theorem 3.3] showed that for our setting the stability condition, i.e., a holomorphic characterization of the image of the map $\M\rightarrow\M_\C$ reduces to the following condition: $$\begin{aligned} \lambda_\tau\in\sum_{j\in J_\Phi}\R_{>0}\iota^\ast u^j.\end{aligned}$$ We also note that the Abelian vortex equations were firstly described in a language of mathematics by Jaffe and Taubes [@JT1; @Tau1]. They classified all finite energy solutions of the Abelian vortex equations over $\C$ which are associated with the standard action of $\U(1)$ on $\C$. Then Bradlow [@Br1] proved Theorem \[known theorem\] for the case that $d=1$ and $K=\U(1)$ by using the result of Kazdan and Warner [@KW1]. We remark that a special case of our generalized Kazdan-Warner equations appears in e.g., [@BW1; @Doa1] in relation to the Hitchin-Kobayashi correspondence for the solutions of the Abelian vortex equations. In [@BW1] for the case that the torus $K$ is given as $\{(g, \dots, g, g^{-1},\dots, g^{-1})\in T^{2d}\mid g\in \U(1)\}$, our generalized Kazdan-Warner equation is solved on any compact Riemannian manifold under a different assumption for given functions and from their result Theorem \[known theorem\] follows for the case that the torus $K$ is given as above. We shall see that Theorem \[main theorem\] gives a direct proof of the correspondence for the general case: Let $\Exp: C^\infty(X, k\oplus\inum k)\rightarrow C^\infty(X, K_\C)$ be the exponential map defined by $$\begin{aligned} \Exp(\xi+\inum \eta)\coloneqq(e^{\inum\langle \xi+\inum\eta, u^1\rangle}, \dots, e^{\inum\langle \xi+\inum\eta, u^d\rangle}).\end{aligned}$$ Then the following group isomorphism is induced by the exponential map: $$\begin{aligned} C^\infty(X, \inum k)\simeq C^\infty(X, K_\C)/C^\infty(X, K).\end{aligned}$$ Moreover we have the following: \[iff\] *For each $(A, \Phi)\in\A_K\times\Gamma(S)$ and $\xi\in C^\infty(X, k)$, the following are equivalent.* (1) Define $(A^\prime, \Phi^\prime)\coloneqq\Exp(-\inum\xi)\cdot (A, \Phi)$. Then $(A^\prime, \Phi^\prime)$ solves the following equation: $$\begin{aligned} \Lambda_{\omega_X}F_{A^\prime}+\mu_K(\Phi^\prime)+\tau=0;\end{aligned}$$ (2) The function $\xi$ solves the following equation: $$\begin{aligned} \label{equation} \Delta_{\omega_X}\xi+\frac{1}{2}\sum_{j=1}^d e^{2\langle u^j, \xi\rangle}|\Phi_j|^2\iota^\ast u^j=\Lambda_{\omega_X}F_A+\tau,\end{aligned}$$ where $\Phi_j$ denotes the $j$-th component of $\Phi$. Therefore Theorem \[main theorem\] gives a direct proof of Theorem \[known theorem\]. [**Acknowledgement.**]{} The author is grateful to his supervisor Professor Ryushi Goto for fruitful discussions and encouragements. Proof of Theorem \[main theorem\] ================================= Outline of the proof -------------------- We outline the proof of Theorem \[main theorem\]. As we mentioned in Remark \[trivial\], it suffices to solve equation (\[KWequation\]) under the assumption $$\begin{aligned} \int_M w \ \vol\in\sum_{j\in J_a}\R_{>0}\iota^\ast u^j\end{aligned}$$ and to prove the uniqueness of the solution up to a constant which is in the orthogonal complement of $\sum_{j\in J_a}\R\iota^\ast u^j$. We prove this using the variational method: We first define a functional $E$ whose critical point is a solution of equation (\[KWequation\]). Then we show that the functional $E$ is convex. The uniqueness of the solution of the equation follows from the convexity of $E$. Secondly, we show that the functional $E$ is bounded below and moreover the following estimate holds: $$\begin{aligned} |\xi|_{L^2}\leq(E(\xi)+C)^2+C^\prime E(\xi)+C^{\prime\prime}\end{aligned}$$ with some constants $C$, $C^\prime$ and $C^{\prime\prime}$ (see Lemma \[Key Lemma\]). Lemma \[Key Lemma\] essentially follows from a property of the moment map $\mu_K$ (see Proposition \[equivalent\] and Proposition \[the last proposition\]) and the Poincaré inequality. Finally by a method developed in [@Br2] we see that the functional $E$ has a critical point. Proof of Theorem \[main theorem\] --------------------------------- Hereafter we normalize the measure $\mu_{g_M}$ so that the total volume is 1: $$\begin{aligned} \Vol(M, g_M)\coloneqq\int_M1 \ d\mu_{g_M}=1.\end{aligned}$$ We define a functional $E: L^{2m}_3(M, k^\ast)\rightarrow \R$ by $$\begin{aligned} E(\xi)\coloneqq\int_M \Bigl\{\frac{1}{2}(d\xi, d\xi)+\sum_{j=1}^da_je^{(\iota^\ast u^j, \xi)}-(w, \xi)\Bigr\}\ \vol \ \text{for $\xi\in L^{2m}_3(M, k^\ast)$.}\end{aligned}$$ \[critical point\] *For each $\xi\in L^{2m}_3(M, k^\ast)$, the following are equivalent.* (1) \[critical point 1\]$\xi$ is a critical point of $E$; (2) \[critical point 2\]$\xi$ solves equation (\[KWequation\]). Moreover if $\xi$ solves equation (\[KWequation\]), then $\xi$ is a $C^\infty$ function. We have the following for each $\eta\in L^{2m}_3(M, k^\ast)$. $$\begin{aligned} \left.\frac{d}{dt}\right|_{t=0}E(\xi+t\eta)=\int_M (\Delta_{g_M}\xi+\sum_{j=1}^d a_j e^{(\iota^\ast u^j, \xi)}\iota^\ast u^j-w, \eta) \ \vol.\end{aligned}$$ Therefore (\[critical point 1\]) and (\[critical point 2\]) are equivalent. The rest of the proof follows from the elliptic regularity theorem. \[convexity\] *For each $\xi, \eta\in L^{2m}_3(M, k^\ast)$ and $t\in\R$, the following holds. $$\begin{aligned} \frac{d^2}{dt^2}E(\xi+t\eta)\geq0.\end{aligned}$$ Moreover the following are equivalent.* (1) There exists a $t_0\in\R$ such that $\left. \frac{d^2}{dt^2}\right|_{t=t_0}E(\xi+t\eta)= 0$; (2) $\frac{d^2}{dt^2}E(\xi+t\eta)= 0 $ for all $t\in \R$; (3) $\eta$ is a constant which is in the orthogonal complement of $\sum_{j\in J_a}\R\iota^\ast u^j$. A direct computation shows that $$\begin{aligned} \frac{d^2}{dt^2}E(\xi+t\eta)=\int_M\{(d\eta, d\eta)+\sum_{j=1}^da_je^{(\iota^\ast u^j, \xi+t\eta)}(\iota^\ast u^j, \eta)^2 \}\ \vol.\end{aligned}$$ This implies the claim. [*Let $\xi$ and $\xi^\prime$ be $C^\infty$ solutions of equation (\[KWequation\]). Then $\xi-\xi^\prime$ is a constant which is in the orthogonal complement of $\sum_{j\in J_a}\R\iota^\ast u^j$.* ]{} By Lemma \[critical point\], we have the following. $$\begin{aligned} \left. \frac{d}{dt}\right|_{t=0}E(t\xi+(1-t)\xi^\prime)=\left. \frac{d}{dt}\right|_{t=1}E(t\xi+(1-t)\xi^\prime)=0.\end{aligned}$$ Then Lemma \[convexity\] gives the result. Hereafter we assume $(\iota^\ast u^j)_{j\in J_a}$ generates $k^\ast$ for simplicity. \[Key Lemma\] [*The functional $E$ is bounded below. Further there exist non-negative constants $C$, $C^{\prime}$ and $C^{\prime \prime}$ such that $$\begin{aligned} |\xi|_{L^2}\leq(E(\xi)+C)^2+C^\prime E(\xi)+C^{\prime\prime}\end{aligned}$$ for all $\xi\in L^{2m}_3(M, k^\ast)$.* ]{} The following holds for each $\xi\in L^{2m}_3(M, k^\ast)$. $$\begin{aligned} &E(\xi) \\ =&\int_M \Bigl\{\frac{1}{2}(d\xi, d\xi)+\sum_{j=1}^da_je^{(\iota^\ast u^j, \xi)}-(w, \xi)\Bigr\} \ \vol \\ =&\int_M \Bigl\{\frac{1}{2}(d\xi, d\xi)-(w, \xi-\bar{\xi})\Bigr\} \ \vol +\sum_{j=1}^d\int_Ma_je^{(\iota^\ast u^j, \xi)}\ \vol-(\bar{w},\bar{\xi}) \\ =&\int_M \Bigl\{\frac{1}{2}(d\xi, d\xi)-(w, \xi-\bar{\xi}) \Bigr\}\ \vol+\sum_{j\in J_a}\int_Me^{\log a_j}e^{(\iota^\ast u^j, \xi)}\ \vol-(\bar{w},\bar{\xi}) \\ \geq& \int_M \Bigl\{\frac{1}{2}(d\xi, d\xi)-(w, \xi-\bar{\xi}) \Bigr\}\ \vol+\sum_{j\in J_a}(e^{\int_M\log a_j\ \vol})e^{(\iota^\ast u^j, \bar{\xi})}-(\bar{w},\bar{\xi}), \end{aligned}$$ where the final inequality follows from the Jensen’s inequality. Then by the Poincaré inequality and Proposition \[equivalent\] we see that the functional $E$ is bounded below. The rest of the proof follows from the above estimate of the functional $E$ and Proposition \[the last proposition\]. The remaining part is essentially same as the argument given in [@Br2]. Let $B>0$ a positive real number. We define $L^{2m}_3(M, k^\ast)_B$ by $$\begin{aligned} L^{2m}_3(M, k^\ast)_B\coloneqq\{\xi\in L^{2m}_3(M, k^\ast)\mid |\Delta_{g_M}\xi+\sum_{j=1}^da_je^{(\iota^\ast u^j, \xi)}\iota^\ast u^j-w|_{L^{2m}_1}^{2m}\leq B\}.\end{aligned}$$ By the same way as in [@Br2 Lemma 3.4.2] (we also refer the reader to [@AG1 Proposition 3.1]), we have the following. \[local minimum\] [ *If $E|_{L^{2m}_3(M, k^\ast)_B}$ attains a minimum at $\xi\in L^{2m}_3(M, k^\ast)_B$, then $E(\xi)$ is a global minimum.* ]{} Therefore the problem reduces to show $E|_{L^{2m}_3(M, k^\ast)_B}$ attains a minimum. To see this, we prove the following Lemma \[final lemma\]. \[final lemma\] [ *Let $(\xi_i)_{i\in\N}$ be a sequence of $L^{2m}_3(M, k^\ast)_B$ such that $$\begin{aligned} \lim_{i\to\infty}E(\xi_i)=\inf_{\eta\in L^{2m}_3(M, k^\ast)_B}E(\eta).\end{aligned}$$ Then we have $\sup_{i\in\N}|\xi_i|_{L^{2m}_3}<\infty$.* ]{} Before the proof of Lemma \[final lemma\], we recall the following. \[known lemma\]([@LT1 pp.72-73]) [ *Let $f\in C^2(M, \R)$ be a non-negative function. If $$\begin{aligned} \Delta_{g_M} f \leq C_0 f+C_1\end{aligned}$$ holds for some $C_0\in \R_{\geq0}$ and $C_1 \in \R$, then there is a positive constant $C_2$, depending only on $g_M$ and $C_0$, such that $$\begin{aligned} \max_{x\in M}f(x)\leq C_2(|f|_{L^1}+C_1).\end{aligned}$$* ]{} We have the following for each $i\in\N$. $$\begin{aligned} \frac{1}{2}\Delta_{g_M}|\xi_i|^2=&(\Delta_{g_M} \xi_i, \xi_i)-|d\xi_i|^2 \notag \\ \leq &(\Delta_{g_M}\xi_i+\sum_{j=1}^da_j\iota^\ast u^j-w, \xi_i)-(\sum_{j=1}^da_j\iota^\ast u^j-w, \xi_i) \notag \\ \leq &(\Delta_{g_M}\xi_i+\sum_{j=1}^da_je^{(\iota^\ast u^j, \xi_i)}\iota^\ast u^j-w, \xi_i)-(\sum_{j=1}^da_j\iota^\ast u^j-w, \xi_i), \label{final inequality}\end{aligned}$$ where the final inequality (\[final inequality\]) follows from the following argument: A direct computation shows that $$\begin{aligned} \frac{d}{dt}a_je^{t(\iota^\ast u^j, \xi_i)}(\iota^\ast u^j, \xi_i)=a_je^{t(\iota^\ast u^j, \xi_i)}(\iota^\ast u^j, \xi_i)^2\geq0.\end{aligned}$$ Hence $a_je^{t(\iota^\ast u^j, \xi_i)}(\iota^\ast u^j, \xi_i)$ monotonically increases with increasing $t\in\R$. Then we have the following. $$\begin{aligned} a_j(\iota^\ast u^j, \xi_i)\leq a_je^{(\iota^\ast u^j, \xi_i)}(\iota^\ast u^j, \xi_i),\end{aligned}$$ and thus we have the inequality (\[final inequality\]). Then Lemma \[known lemma\] and the $L^p$-estimate give the claim. [*The functional $E$ has a critical point.* ]{} By Lemma \[convexity\] and Lemma \[local minimum\], it is enough to show that $E|_{L^{2m}_3(M, k^\ast)_B}$ attains a minimum. We take a sequence $(\xi_i)_{i\in\N}$ so that $$\begin{aligned} \lim_{i\to\infty}E(\xi_i)=\inf_{\eta\in L^{2m}_3(M, k^\ast)_B}E(\eta).\end{aligned}$$ Then by Lemma \[final lemma\] there exits a subsequence $(\xi_{i_j})_{j\in\N}$ such that $(\xi_{i_j})_{j\in\N}$ weakly converges a $\xi_\infty\in L^{2m}_3(M, k^\ast)_B$. Since $E$ is continuous with respect to the weak topology, $E$ attains a minimum at $\xi=\xi_\infty$. Geometric invariant theory and the moment maps for linear torus actions ======================================================================= We give a brief review of the relationship between the geometric invariant theory and the moment maps for linear torus actions. General references for this section are [@Dol1; @Kin1; @Kir1; @KN1; @MFK1; @Nak1; @New1]. Notation -------- We first fix our notation. Let $K$ be a closed connected subgroup of a real torus $T^d:=\U(1)^d$ with the Lie algebra $k\subseteq t^d$. We denote by $\iota^\ast: (t^d)^\ast \rightarrow k^\ast$ the dual map of the inclusion map $\iota: k\rightarrow t^d$. Let $u_1,\dots, u_d$ be a basis of $t^d$ defined by $$\begin{aligned} u_1\coloneqq&(\inum, 0, \dots, 0), \\ u_2\coloneqq&(0, \inum, 0,\dots, 0), \\ &\cdots \\ u_d\coloneqq&(0,\dots, 0,\inum).\end{aligned}$$ We denote by $u^1, \dots, u^d\in (t^d)^\ast$ the dual basis of $u_1, \dots, u_d$. Let $(\cdot, \cdot)$ be the metric on $t^d$ and $(t^d)^\ast$ satisfying $$\begin{aligned} (u_i, u_j)=(u^i, u^j)=\delta_{ij} \ \ \text{for all $i, j$,}\end{aligned}$$ where $\delta_{ij}$ denotes the Kronecker delta. The diagonal action of $T^d$ on $\C^d$ induces an action of $K$ which preserves the Kähler structure of $\C^d$. Let $\mu_K:\C^d\rightarrow k^\ast$ be a moment map for the action of $K$ which is defined by $$\begin{aligned} \langle \mu_K(z), v \rangle=\frac{1}{2}g_{\R^{2d}}(\inum vz, z) \ \ \text{for $v\in k$},\end{aligned}$$ where we denote by $g_{\R^{2d}}(\cdot, \cdot)$ the standard metric of $\C^d\simeq \R^{2d}$, and by $\langle \cdot, \cdot \rangle$ the natural coupling. The moment map $\mu_K$ is also denoted as $$\begin{aligned} \mu_K(z)=-\frac{1}{2}\sum_{j=1}^d\iota^\ast u^j |z_j|^2 \ \ \text{for $z=(z_1,\dots, z_d)\in\C^d$}.\end{aligned}$$ Let $(\C^\times)^d$ be the complexification of $T^d$. We define the exponential map $\Exp: t^d\oplus \inum t^d\rightarrow (\C^\times)^d$ by $$\begin{aligned} \Exp(v+\inum v^\prime)=(e^{\inum\langle v+\inum v^\prime, u^1\rangle}, \dots, e^{\inum\langle v+\inum v^\prime, u^d\rangle}).\end{aligned}$$ We denote by $K_\C$ the complexification of $K$. Let $k_\Z\subseteq k$ be $\ker \Exp |_k$ and $(k_\Z)^\ast$ the dual. Note $(k_\Z)^\ast$ is naturally identified with $\sum_{j=1}^d\Z \ (\iota^\ast u^j/2\pi)$. For each $\alpha\in (k_\Z)^\ast$, we define a character $\chi_\alpha : K_\C\rightarrow \C^\times$ by $$\begin{aligned} \chi_\alpha(\Exp(v+\inum v^\prime))=e^{2\pi\inum\langle v+\inum v^\prime, \alpha\rangle}.\end{aligned}$$ Symplectic and GIT quotients ----------------------------- Let $\alpha\in(k_\Z)^\ast$. We define an action of $K_\C$ on $\C^d\times \C$ by $$\begin{aligned} g\cdot (z, v)\coloneqq(gz, \chi_\alpha(g)^{-1}v) \ \ \text{for $(z, v)\in\C^d\times\C$.}\end{aligned}$$ Let $R_\alpha$ be the invariant ring of the above action: $$\begin{aligned} R_\alpha\coloneqq\{\hat{f}(x, y)\in\C[x_1,\dots, x_d, y]\mid\text{$\hat{f}(g\cdot(x,y))=\hat{f}(x,y)$ for all $g\in K_\C$} \}.\end{aligned}$$ By the theorem of Nagata, $R_\alpha$ is finitely generated. For each $n\in\Z_{\geq0}$ let $R_{\alpha, n}$ be a space of polynomials defined by $$\begin{aligned} R_{\alpha, n}\coloneqq\{f(x)\in\C[x_1,\dots, x_d]\mid \text{$f(gx)=\chi_\alpha(g)^nf(x)$ for all $g\in K_\C$}\}\end{aligned}$$ Then $R_\alpha$ is naturally identified with $\bigoplus_{n\geq0}R_{\alpha, n}$. Define $\C^d{{/\mkern-6mu/}}_\alpha K_\C\coloneqq{\rm Proj}(\bigoplus_{n\geq0}R_{\alpha, n})$. This is called the [*geometric invariant theory (GIT) quotient*]{}. We say $z\in\C^d$ is $\alpha$-semistable if there exists a $f(x)\in R_{\alpha, n}$ with $n\in\Z_{>0}$ such that $f(z)\neq0$. We denote by $(\C^d)^{\alpha-ss}$ the set of all $\alpha$-semistable points. We remark the following fact. [ *Let $V$ be a complex vector space and $G\subseteq \GL(V)$ an algebraic subgroup. Then we have the following. $$\begin{aligned} \overline{\overline{G\cdot p}}=\overline{G\cdot p} \ \ \text{for all $p\in V$},\end{aligned}$$ where we denote by $\overline{\overline{G\cdot p}}$ the Euclidean closure, and by $\overline{G\cdot p}$ the Zariski closure. In particular, $G\cdot p$ is closed with respect to the Euclidean topology if and only if it is closed with respect to the Zariski topology.* ]{} See [@Mum1]. The GIT quotient can be described as follows. \[morphism\] [*There exists a categorical quotient $\phi:(\C^d)^{\alpha-ss}\rightarrow \C^d{{/\mkern-6mu/}}_\alpha K_\C$ which satisfies the following properties: for each $z, z^\prime\in(\C^d)^{\alpha-ss}$, $\phi(z)=\phi(z^\prime)$ holds if and only if $\overline{K_\C\cdot z}\cap\overline{K_\C\cdot z^\prime}\cap(\C^d)^{\alpha-ss}\neq\emptyset$ and further for each $q\in \C^d {{/\mkern-6mu/}}_\alpha K_\C$, $\phi^{-1}(q)$ contains a unique $K_\C$-orbit which is closed in $(\C^d)^{\alpha-ss}$.* ]{} See [@Dol1; @MFK1; @New1]. We define an equivalence relation $\sim$ on $(\C^d)^{\alpha-ss}$ as follows: $$\begin{aligned} z\sim z^\prime \Longleftrightarrow\overline{K_\C\cdot z}\cap\overline{K_\C\cdot z^\prime}\cap(\C^d)^{\alpha-ss}\neq\emptyset \ \ \text{for $z, z^\prime\in(\C^d)^{\alpha-ss}$}.\end{aligned}$$ Then by Proposition \[morphism\], $\C^d{{/\mkern-6mu/}}_\alpha K_\C$ is identified with $(\C^d)^{\alpha-ss}/\sim$. Moreover for each equivalent class there exists a $z\in(\C^d)^{\alpha-ss}$ such that $K_\C\cdot z=(\C^d)^{\alpha-ss}\cap \overline{K_\C\cdot z}$ and such a $z$ is unique up to a transformation of $K_\C$. $\alpha$-semistable points are characterized as follows. \[semistable\] *The following are equivalent for each $z\in\C^d$.* (1) \[semi1\] $z$ is $\alpha$-semistable; (2) \[semi2\] $\alpha$ satisfies the following: $$\begin{aligned} \alpha\in\sum_{j\in J_z}\Q_{\geq0}(\iota^\ast u^j/2\pi),\end{aligned}$$ where $J_z$ denotes $\{j\in\{1, \dots, d\}\mid z_j\neq 0\}$; (3) \[semi3\] $\alpha$ is in the cone generated by $(\iota^\ast u^j/2\pi)_{j\in J_z}$: $$\begin{aligned} \alpha\in\sum_{j\in J_z}\R_{\geq0}(\iota^\ast u^j/2\pi);\end{aligned}$$ (4) \[semi4\]For each $v\in \C\backslash\{0\}$, $\overline{K_\C\cdot(z, v)}$ does not intersect with $\C^d\times \{0\}$. $(\ref{semi1})\Leftrightarrow(\ref{semi2})$ This can be proved by essentially the same argument as in the proof of [@Kon1 Lemma 3.4]. $(\ref{semi2})\Leftrightarrow (\ref{semi3})$ This follows from the general theory of polyhedral convex cones. See [@Ful1]. $(\ref{semi1})\Rightarrow (\ref{semi4})$ Suppose $z$ is $\alpha$-semistable. We take a $f\in R_{n,\alpha}$ such that $n\in \Z_{>0}$ and $f(z)\neq 0$. We define a polynomial $\hat{f}(x, y)$ by $\hat{f}(x, y)\coloneqq y^nf(x)$. Then we have the following. $$\begin{aligned} \hat{f}(x, y)|_{\overline{K_\C\cdot(z, v)}}&\equiv v^nf(z), \\ \hat{f}(x, y)|_{\C^d\times\{0\}}&\equiv 0, \end{aligned}$$ and thus (\[semi4\]) holds. $(\ref{semi4})\Rightarrow (\ref{semi1})$ Suppose (\[semi4\]) holds. Then there exists a polynomial $\hat{f}(x, y)$ such that $$\begin{aligned} \hat{f}(x, y)|_{\overline{K_\C\cdot(z, v)}}&\equiv 1, \\ \hat{f}(x, y)|_{\C^d\times\{0\}}&\equiv 0. \end{aligned}$$ The polynomial $\hat{f}(x, y)$ can be written as $\hat{f}(x, y)=yf_1(x)+\cdots+y^mf_m(x)$. Take an $n\in\{1, \dots, m\}$ such that $f_n(x)\neq 0$. Then $f_n\in R_{n,\alpha}$ and $f_n(z)\neq 0$. The closed orbits are characterized as follows. \[closed orbits\] *The following are equivalent for each $z\in\C^d$.* (1) \[1\]$z$ is $\alpha$-semistable and the $K_\C$-orbit is closed in $(\C^d)^{\alpha-ss}$: $$\begin{aligned} K_\C\cdot z=\overline{K_\C\cdot z}\cap (\C^d)^{\alpha-ss};\end{aligned}$$ (2) \[2\] $\alpha$ satisfies the following: $$\begin{aligned} \alpha\in\sum_{j\in J_z}\Q_{>0}(\iota^\ast u^j/2\pi);\end{aligned}$$ (3) \[3\] $\alpha$ is in the interior of the cone generated by $(\iota^\ast u^j/2\pi)_{j\in J_z}$: $$\begin{aligned} \alpha\in\sum_{j\in J_z}\R_{>0}(\iota^\ast u^j/2\pi);\end{aligned}$$ (4) \[4\] The following holds: $$\begin{aligned} \sum_{j\in J_z}\R(\iota^\ast u^j/2\pi)=\sum_{j\in J_z}\R_{\geq0}(\iota^\ast u^j/2\pi)+\R_{\geq0}(-\alpha);\end{aligned}$$ (5) \[5\]For each $v\in \C\backslash\{0\}$, $K_\C\cdot (z, v)$ is closed; (6) \[6\] The following holds: $$\begin{aligned} \mu_K^{-1}(-\alpha)\cap K_\C\cdot z\neq \emptyset.\end{aligned}$$ $(\ref{1})\Rightarrow (\ref{5})$ Suppose (1) holds. By the general theory of algebraic groups, there uniquely exists a closed orbit which is contained in $\overline{K_\C\cdot(z,v)}$. Let $K_\C\cdot(z^\prime, v)$ be such a closed orbit. Then by Proposition \[semistable\], $z^\prime\in(\C^d)^{\alpha-ss}$. We take a sequence $(g_i)_{i\in\N}$ such that $$\begin{aligned} (z^\prime, v)=\lim_{i\to\infty}g_i\cdot(z,v).\end{aligned}$$ Therefore $z^\prime=\lim_{i\to\infty}g_i\cdot z$, and thus $z^\prime\in\overline{K_\C\cdot z}\cap (\C^d)^{\alpha-ss}$. Then (\[5\]) holds. $(\ref{5})\Rightarrow (\ref{1})$ Suppose (\[5\]) holds. Let $z^\prime\in\overline{K_\C\cdot z}\backslash K_\C\cdot z$. We take a sequence $(g_i)_{i\in\N}$ so that $z^\prime=\lim_{i\to\infty}g_i\cdot z$. Since $K_\C\cdot (z,1)$ is closed, $\lim_{i\to\infty}|\chi_\alpha(g_i)^{-1}|=\infty$. Then $\lim_{i\to\infty}(g_i^{-1}z^\prime, \chi_{\alpha}(g_i))\in\C^d\times\{0\}$. Therefore $z^\prime\notin(\C^d)^{\alpha-ss}$. $(\ref{2})\Leftrightarrow (\ref{3})\Leftrightarrow (\ref{4})$ This follows from the general theory of polyhedral convex cones. See [@Ful1]. $(\ref{3})\Leftrightarrow(\ref{6})$ We shall prove this in Proposition \[equivalent\]. $(\ref{4})\Leftrightarrow (\ref{5})$ See [@Nak1 pp.30-31]. The equivalence of (\[2\]) and (\[3\]) holds for any $\lambda\in k^\ast$: \[equivalent\] *Let $\lambda\in k^\ast$ and $z\in\C^d$. We define $l_{\lambda, z}:k\rightarrow \R$ by $$\begin{aligned} l_{\lambda, z}(v)\coloneqq\frac{1}{4}\sum_{j=1}^d|z_j|^2e^{2\langle u^j, v\rangle}-\langle\lambda, v\rangle.\end{aligned}$$ Then the following are equivalent.* (1) \[equivalent 1\] $\lambda$ is in the interior of the cone generated by $(\iota^\ast u^j/2\pi)_{j\in J_z}$: $$\begin{aligned} \lambda\in\sum_{j\in J_z}\R_{>0}\iota^\ast u^j;\end{aligned}$$ (2) \[equivalent 2\] The following holds: $$\begin{aligned} \mu_K^{-1}(-\lambda)\cap K_\C\cdot z\neq \emptyset;\end{aligned}$$ (3) \[equivalent 3\]$l_{\lambda, z}$ attains a minimum. Moreover if $v$ and $v^\prime$ be minimizers of $l_{\lambda, z}$, then $v-v^\prime$ is in the orthogonal complement of $\sum_{j\in J_z}\R\iota^\ast u^j$. We assume that $(\iota^\ast u^j)_{j\in J_z}$ generates $k^\ast$ for simplicity. Then a direct computation shows that $l_{\lambda, z}$ is strictly convex. Also we see that $v\in k$ satisfies $$\begin{aligned} \lambda=\frac{1}{2}\sum_{j=1}^de^{2\langle u^j, v\rangle}|z_j|^2\iota^\ast u^j\end{aligned}$$ if and only if $v$ is a critical point of $l_{\lambda, z}$. Therefore (\[equivalent 2\]) and (\[equivalent 3\]) are equivalent. Clearly, (\[equivalent 2\]) implies (\[equivalent 1\]). We shall prove that (\[equivalent 1\]) implies (\[equivalent 3\]). Suppose (\[equivalent 1\]) holds. It is enough to show that the following holds for each $v\neq 0$: $$\begin{aligned} \lim_{t\to\infty}l_{\lambda, z}(tv)=\infty.\end{aligned}$$ Clearly this holds for a $v$ such that $\langle \lambda, v\rangle<0$. We consider the case that $v$ satisfies $\langle\lambda, v\rangle\geq 0$. Then $\lambda$ is denoted as $\lambda=\sum_{j\in J_z}c_j\iota^\ast u^j$ for a family of positive numbers $(c_j)_{j\in J_z}$. It suffices to show that there exists a $j\in J_z$ such that $\langle u^j, v\rangle>0$. We assume the converse, i.e., we suppose that for all $j\in J_z$ we have $\langle u^j, v\rangle\leq0$. Since $v$ satisfies $\langle \lambda, v\rangle\geq 0$, we have $\sum_{j\in J_z} c_j\langle u^j, v\rangle=0$. This implies there exists a $j\in J_z$ such that $\langle u^j, v\rangle<0$. Also this suggests there exists a $j^\prime\in J_z$ which satisfies $\langle u^{j^\prime}, v\rangle>0$ and this contradicts the assumption. Therefore (\[equivalent 1\]) implies (\[equivalent 3\]). Then we see that (\[equivalent 1\]), (\[equivalent 2\]) and (\[equivalent 3\]) are equivalent. The rest of the proof follows from the convexity of $l_{\lambda, z}$. From Proposition \[morphism\], Proposition \[closed orbits\] and Proposition \[equivalent\], we have the following. [*The following map is bijective: $$\begin{aligned} \mu_K^{-1}(-\alpha)/K\longrightarrow(\C^d)^{\alpha-ss}/\sim.\end{aligned}$$* ]{} Finally, for the proof of Lemma \[Key Lemma\] we prove the following. \[the last proposition\] [*If the conditions of Proposition \[equivalent\] are satisfied, then we have the following. Let $V_z$ be $\sum_{j\in J_z}\R\iota^\ast u^j$. Define $\hat{l}_{\lambda, z}:V_z\rightarrow \R$ by $$\begin{aligned} \hat{l}_{\lambda, z}(v)\coloneqq l_{\lambda, z}(v^\ast)-|v|^{\frac{1}{2}},\end{aligned}$$ where $v^\ast\in k$ denotes the adjoint of $v$. Then $\hat{l}_{\lambda, z}$ attains a minimum.* ]{} A direct computation shows that the following map $$\begin{aligned} \R_{>0}&\longrightarrow \R \\ t&\longmapsto \hat{l}_{\lambda, z}(tv)\end{aligned}$$ is strictly convex for each $v\neq 0$. Further by the same way as in the proof of Proposition \[equivalent\], we have the following: $$\begin{aligned} \lim_{t\to \infty}\hat{l}_{\lambda, z}(tv)=\infty.\end{aligned}$$ Therefore $\hat{l}_{\lambda, z}(tv) \ (t\in\R_{\geq0})$ attains a minimum and this implies the claim. [99]{} E. Aldrovandi and G. Falqui, [*Geometry of Higgs and Toda fields on Riemann surfaces*]{}, J. Geom. Phys. 17 (1995), no. 1, 25–48. L. Álvarez-Cónsul and O. García-Prada, [*Hitchin-Kobayashi correspondence, quivers, and vortices*]{}, Comm. Math. Phys. 238 (2003), no. 1-2, 1–33. D. Banfield, [*The geometry of coupled equations in gauge theory*]{}, D. Phil. Thesis, University of Oxford, 1996. J. M. Baptista, [*Moduli spaces of Abelian vortices on Kähler manifolds*]{}, arXiv:1211.0012. D. Baraglia,[*Cyclic Higgs bundles and the affine Toda equations*]{}, Geom. Dedicata 174 (2015), 25–42. S.B. Bradlow, [*Vortices in holomorphic line bundles over closed Kähler manifolds*]{}, Comm. Math. Phys. 135 (1990), no. 1, 1-17. S.B. Bradlow, [*Special metrics and stability for holomorphic bundles with global sections*]{}, J. Diff. Geom. 33 (1991), no. 1, 169-213. J. A. Bryan and R. Wentworth, [*The multi-monopole equations for Kähler surfaces*]{}, Turkish J. Math. 20 (1996), no. 1, 119–128. S. Cecotti and C. Vafa, [*Topological–anti-topological fusion*]{}, Nuclear Phys. B 367 (1991), no. 2, 359–461. S. Dai and Q. Li, [*On cyclic Higgs bundles*]{}, Math. Ann., Nov 2018. A. Doan, [*Adiabatic limits and Kazdan-Warner equations*]{}, Calc. Var. Partial Differential Equations 57 (2018), no. 5, Art. 124, 25 pp. I. Dolgachev, [*Lectures on invariant theory*]{}, London Mathematical Society Lecture Note Series, 296. Cambridge University Press, Cambridge, 2003. xvi+220 pp. ISBN: 0-521-52548-9 W. Fulton, [*Introduction to toric varieties*]{}, Annals of Mathematics Studies, 131. The William H. Roever Lectures in Geometry. Princeton University Press, Princeton, NJ, 1993. xii+157 pp. ISBN: 0-691-00049-2. M. A. Guest, A. R. Its and C.-S. Lin, [*Isomonodromy aspects of the $tt^\ast$ equations of Cecotti and Vafa I. Stokes data*]{}, Int. Math. Res. Not. IMRN 2015, no. 22, 11745–11784. M. A. Guest and C.-S. Lin, [*Nonlinear PDE aspects of the $tt^\ast$ equations of Cecotti and Vafa*]{}, J. Reine Angew. Math. 689 (2014), 1–32. N. J. Hitchin, [*The self-duality equations on a Riemann surface*]{}, Proc. London Math. Soc. (3) 55 (1987), no. 1, 59–126. A. Jaffe and C. H. Taubes, [*Vortices and monopoles*]{}, Progress in Physics, 2. Birkhäuser, Boston, Mass., 1980. v+287 pp. ISBN: 3-7643-3025-2. A. D. King, [*Moduli of representations of finite-dimensional algebras*]{}, Quart. J. Math. Oxford Ser. (2) 45 (1994), no. 180, 515-530. F. C. Kirwan, [*Cohomology of quotients in symplectic and algebraic geometry*]{}, Mathematical Notes, 31. Princeton University Press, Princeton, NJ, 1984. i+211 pp. ISBN: 0-691-08370-3. G. Kempf and L. Ness, [*The length of vectors in representation spaces*]{}, Algebraic geometry (Proc. Summer Meeting, Univ. Copenhagen, Copenhagen, 1978), pp. 233–243, Lecture Notes in Math., 732, Springer, Berlin, 1979. H. Konno, [*The geometry of toric hyperkähler varieties*]{}, Toric topology, 241–260, Contemp. Math., 460, Amer. Math. Soc., Providence, RI, 2008. J. L. Kazdan and F. W. Warner, [*Curvature functions for compact 2-manifolds*]{}, Ann. of Math. (2) 99 (1974), 14-47. M. Lübke and A. Teleman, [*The Kobayashi-Hitchin correspondence*]{}, World Scientific Publishing Co., Inc., River Edge, NJ, 1995. x+254 pp. ISBN: 981-02-2168-1. D. Mumford, [*The red book of varieties and schemes*]{}, Second, expanded edition. Includes the Michigan lectures (1974) on curves and their Jacobians. With contributions by Enrico Arbarello. Lecture Notes in Mathematics, 1358. Springer-Verlag, Berlin, 1999. x+306 pp. ISBN: 3-540-63293-X. I. Mundet i Riera, [*A Hitchin-Kobayashi correspondence for Kähler fibrations*]{}, J. Reine Angew. Math. 528 (2000), 41-80. D. Mumford, J. Fogarty and F. Kirwan, [*Geometric invariant theory*]{}, Third edition. Ergebnisse der Mathematik und ihrer Grenzgebiete (2) \[Results in Mathematics and Related Areas (2)\], 34. Springer-Verlag, Berlin, 1994. xiv+292 pp. ISBN: 3-540-56963-4. H. Nakajima, [*Lectures on Hilbert schemes of points on surfaces*]{}, University Lecture Series, 18. American Mathematical Society, Providence, RI, 1999. xii+132 pp. ISBN: 0-8218-1956-9. P. E. Newstead, [*Introduction to moduli problems and orbit spaces*]{}, Tata Institute of Fundamental Research Lectures on Mathematics and Physics, 51. Tata Institute of Fundamental Research, Bombay; by the Narosa Publishing House, New Delhi, 1978. vi+183 pp. ISBN: 0-387-08851-2. C. H. Taubes, [*Arbitrary N-vortex solutions to the first order Ginzburg-Landau equations*]{}, Comm. Math. Phys. 72 (1980), no. 3, 277–292.

ArXiv

--- abstract: 'For a system composed of nine qubits, we show that natural interactions among the qubits induce the time evolution that can be regarded, at discrete times, as the Rabi oscillation of a logical qubit. Neither fine tuning of the parameters nor switching of the interactions is necessary. Although straightforward application of quantum error correction fails, we propose a protocol by which the logical Rabi oscillation is protected against all single-qubit errors. The present method thus opens a simple and realistic way of protecting the unitary time evolution against noise.' author: - Naoaki Kokubun - Akira Shimizu title: Protected Rabi oscillation induced by natural interactions among physical qubits --- Introduction ============ Decoherence of quantum states has been attracting much attention for long years [@SM02]. Many methods have been proposed for defeating the decoherence. As compared with other methods [@ZR97; @KBLW01; @VKL99; @Facchi05], quantum error correction (QEC) [@Sh95; @St96; @NC00; @Go97; @Pre98] has a great advantage of protecting against arbitrary errors if they only affect a single qubit (two-level system) in each [*logical*]{} qubit [@NC00]. Although QEC has been developed in the context of quantum computation, it is interesting and useful to apply QEC to protection of the unitary time evolution (Hamiltonian evolution) against noise. When trying to realize this, however, one encounters many physical problems, which are usually disregarded in discussions on the computational complexity [@NC00]. For example, some physical process may be much more difficult to realize than another, even if the number of the necessary steps for them differs ‘only by polynomial steps’ [@NC00]. Furthermore, fabrication of a controlled-NOT gate, which is one of the elementary quantum gates, is very difficult because it requires fine tuning of the coupling constants of the interactions and high-precision switching of them, even if one employs the excellent ideas of Refs. [@DiV00; @ML05]. Assembling a quantum circuit from the elementary gates is even more difficult, particularly when the circuit is large and complicated. Unfortunately, the circuit indeed becomes large and complicated when one tries to apply QEC to the Hamiltonian evolution, even for the simplest case such as the Rabi oscillation [@Pre98]. The largest and most complicated part of the circuit is the one that induces the encoded Hamiltonian evolution (such as the Rabi oscillation of a logical qubit) [*in a fault-tolerant manner*]{} [@NC00; @Go97]. Although a non-fault-tolerant circuit can be much simpler, such a circuit is too fragile to errors. It is therefore important to explore new methods, which are physically more feasible and natural, for inducing the encoded Hamiltonian evolution and thereby making QEC applicable. In this paper, we propose such a new method, choosing the Rabi oscillation as the Hamiltonian evolution to be protected. The method utilizes effective interactions that arise naturally among physical qubits. We show that the values of the parameters in the interactions are to a great extent arbitrary. Furthermore, switching of the interactions is unnecessary. Therefore, a system of a logical qubit with such interactions can be prepared easily by placing several two-level systems close to each other. Once such a system is prepared, it is driven spontaneously and flawlessly by the Schrödinger equation. This is much easier than to drive the system by a fault-tolerant quantum circuit. On the other hand, we argue physically that it is highly probable that unwanted interactions should also exist in such a system. While some of them are shown to be irrelevant, the others invalidate straightforward application of QEC. As a resolution we present a protocol, which we call the error-correction sequence. One can realize the protected Rabi oscillation by using the natural interactions (to induce the logical Rabi oscillation) and a quantum circuit for the error-correction sequence. This is much easier than realizing it wholly with a quantum circuit, because a fault-tolerant quantum circuit for inducing the logical Rabi oscillation, which is the largest and most complicated part of the full circuit, is unnecessary. Natural Hamiltonian for logical Rabi oscillation {#sec:Hamiltonian} ================================================ We employ a two-level system as a basic element, which we call a qubit or [*physical*]{} qubit. We represent operators acting on a qubit in terms of the Pauli operators $X,Y,Z$ (i.e., $\sigma_1, \sigma_2, \sigma_3$), which are not necessarily those for a physical spin. To apply QEC to the Rabi oscillation, $$e^{i \omega X t} |0 \rangle = \cos(\omega t) |0 \rangle + i \sin(\omega t) |1 \rangle, \label{eq:Rabi}$$ we replace a single qubit with a [*logical*]{} qubit which is composed of several qubits. The basis states $|0 \rangle, |1 \rangle$ ($+1$ and $-1$ eigenstate of $Z$, respectively) of a qubit correspond to $|0_L \rangle, |1_L \rangle$ of a logical qubit. The subspace (of the logical qubit) that is spanned by the latter is called the code space. For the reasons that will be described in Sec.\[sec:DandC\], we here take the Shor code [@Sh95], in which a logical qubit is composed of nine qubits and $$\begin{aligned} |0_L\rangle &=& {\bigl(|000\rangle+|111\rangle\bigr) \bigl(|000\rangle+|111\rangle\bigr) \bigl(|000\rangle+|111\rangle\bigr) \over 2^{3/2}}, %2\sqrt{2} \\ |1_L\rangle &=& {\bigl(|000\rangle-|111\rangle\bigr) \bigl(|000\rangle-|111\rangle\bigr) \bigl(|000\rangle-|111\rangle\bigr) \over 2^{3/2}}. %2\sqrt{2}.\end{aligned}$$ We have to induce the [*logical*]{} Rabi oscillation; $$e^{i \omega X_L t} |0_L \rangle = \cos(\omega t) |0_L \rangle + i \sin(\omega t) |1_L \rangle,$$ where $X_L$ is a logical Pauli operator; $X_L |0_L \rangle = |1_L \rangle$ and $X_L |1_L \rangle = |0_L \rangle$. Obviously, it can be induced if the Hamiltonian is $- \omega X_L$. \[Here and after, we take $\hbar=1$.\] This is an interaction among three or more qubits, for any code that can correct all single-qubit errors (Appendix \[app:3 or more\]). For the Shor code, $X_L$ can be represented in various ways, e.g., as $X_L = Z_3 Z_6 Z_9$ or $\Pi_{i=1}^9 Z_i$, where $Z_i$ acts on qubit $i$. In the following, we take $$X_L = Z_1 Z_4 Z_7.$$ Suppose that nine qubits (such as atoms, quantum dots, and so on) composing a logical qubit are placed close to each other as shown in Fig. \[qubits configuration\]. Then, as will be discussed in Sec. \[sec:effectiveH\], a three-qubit interaction proportional to $X_L$ ($=Z_1 Z_4 Z_7$) would be generated as an effective interaction. \[Similar three-qubit interactions were also discussed in Refs. [@3body1; @3body2].\] Unfortunately, however, if this interaction is strong enough unwanted two-qubit interactions proportional to $Z_1Z_4, Z_4Z_7, Z_7Z_1$ should also be strong, because otherwise the following unphysical conclusion would be drawn; if one of qubits $1,4,7$ is removed the other two qubits would have no interactions. Furthermore, interactions between other pairs of qubits, such as $Z_1 Z_2, Z_2 Z_3, \cdots$, would also exist in general. Therefore, a natural and simple Hamiltonian for the system of Fig. \[qubits configuration\] is $$H = H_{D}+H_{S}, \label{total hamiltonian}$$ where $$\begin{aligned} && \hspace{-3mm} H_{D} = -\omega Z_1 Z_4 Z_7 -J(k_1 Z_1Z_4+k_4 Z_4Z_7+ k_7 Z_7Z_1), \label{Rabi hamiltonian}\\ && \hspace{-3mm} H_{S} = \sum_{s=2,3,5,6,8,9} g_s Z_{s-1} Z_s. \label{Stabilizer hamiltonian}\end{aligned}$$ Here, $\omega$, $J$, $k_r$’s, $g_s$’s are real parameters. Since the signs of these parameters are irrelevant to the following discussions, we assume without loss of generality that they are positive. Furthermore, since three-qubit interactions are generally weaker than two-qubit interactions (see Sec. \[sec:effectiveH\]), we assume naturally that $$0< \omega \ll J. \label{omega<<J}$$ Although single-qubit terms may also exist, we can forget them because, as discussed in Appendix \[Single terms\], they are irrelevant to the following discussions. ![An example of the configuration of qubits that have the proposed Hamiltonian. The distances between the qubits are to a large extent arbitrary. We label the qubits inside and outside the central triangle by $r$ ($=1,4,7$) and $s$ ($=2,3,5,6,8,9$), respectively. []{data-label="qubits configuration"}](Figure_1.eps){width="0.4\linewidth"} Note that the operators $Z_{s-1} Z_s$ in $H_{S}$ do not change $|0_L \rangle$ or $|1_L \rangle$, i.e., they are elements of the stabilizer [@Go97] of the Shor code. Using this fact, we will show later by explicit calculations that the values of $g_s$’s are irrelevant. On the other hand, the two-qubit interactions in $H_D$ are not elements of the stabilizer, and hence drive the state out of the code space. Nevertheless, we will show in Sec. \[sec:kk’\] that the values of $J$ and $k_r$’s are fairly arbitrary as long as $\omega \ll J$. The value of $\omega$ is also unimportant because changing $\omega$ is just equivalent to changing the time scale. Therefore, the values (including signs) of all the parameters in $H$ (hence the distances between the qubits) are to a great extent arbitrary. This makes our scheme robust to fabrication errors. Once the system is thus fabricated, the law of the Nature drives it flawlessly if noise is absent. Difficulties and resolutions {#sec:concepts} ============================ We now discuss effects of noise. There are two difficulties in applying QEC straightforwardly to the system driven by $H$. We now explain them and resolutions. For simplicity, we explain the case where $k_1=k_4=k_7=1$. More general cases will be discussed in Sec. \[sec:kk’\]. We study the first difficulty by investigating the time evolution in the absence of noise, i.e., we calculate $ |\psi(t) \rangle \equiv e^{-i H t}|\psi_L^0 \rangle $, where $|\psi_L^0\rangle$ is a vector in the code space. We note that all terms in $H$ commute with each other, and that $H_S$ does not change $|\psi_L^0\rangle$ because all terms in $H_{S}$ are elements of the stabilizer. Using these facts and the relations $Z_1Z_4=Z_7X_L,Z_4Z_7=Z_1X_L,Z_7Z_1=Z_4 X_L$, we find $$\begin{aligned} |\psi(t) \rangle = & \left\{\cos^3(Jt)-i\sin^3(Jt)\right\} e^{i\omega tX_L}|\psi_L^0 \rangle \nonumber\\ & +\frac{i}{2}\sum_{r=1,4,7}e^{iJt}\sin(2Jt) Z_r %\left(Z_1+Z_4+Z_7\right) X_L e^{i\omega tX_L}|\psi_L^0 \rangle. \label{general period dynamics}\end{aligned}$$ When $\sin(2J t) \neq 0$, this state is out of the code space because of the last term. Therefore, we cannot perform QEC for phase errors at an arbitrary time, because the syndrome measurement [@NC00] to identify the errors misidentifies the last term as a wrong term generated by a phase-flip noise; if QEC for phase errors were performed with some intervals $\mu$ the time evolution would be affected as shown in Fig. \[Phase synd meas mono\], even when noise is absent. ![Probability of finding $|0_L\rangle$ plotted against time, for the logical Rabi oscillation (chain line), the Hamiltonian evolution by $H$ (dashed line), that affected by QEC for phase errors (solid line), which is performed repeatedly with some intervals $\mu$. []{data-label="Phase synd meas mono"}](Figure_2.eps){width="0.8\linewidth"} However, if we focus on the discrete times $$t_m \equiv m \tau \quad (m=0,1,2,\cdots),$$ then $|\psi(t_m)\rangle$ is in the code space, where $$\tau \equiv \pi/2J. \label{def:tau}$$ Therefore, we can perform QEC at $t=t_m$, for both phase and bit-flip errors. Furthermore, since $$|\psi(t_m)\rangle=e^{i\omega t_m X_L}|\psi_L^0\rangle = \left[ \cos(\omega t_m) + i \sin(\omega t_m)X_L \right] |\psi_L^0\rangle$$ apart from an irrelevant phase factor, the logical Rabi oscillation is realized at these discrete times, which we call the [*discrete logical Rabi oscillation*]{}. Since $\omega/J\ll 1$, the intervals $\tau$ of the discrete times are much shorter than the period $2 \pi/\omega$ of the Rabi oscillation. Hence, the discrete logical Rabi oscillation $\{|\psi(t_m)\rangle \}_{m=0,1,2\cdots}$ is quasi continuous as shown by the dots in Fig. \[Discrete Rabi oscillation\]. ![Probability of finding $|0_L\rangle$ plotted against time, for the logical Rabi oscillation (chain line) and the Hamiltonian evolution by $H$ (solid line). For clarity, we take $\omega /J$ ($\ll 1$) not so small; $\omega /J=0.1$. The dots represent the discrete logical Rabi oscillation. []{data-label="Discrete Rabi oscillation"}](Figure_3.eps){width="0.8\linewidth"} To discuss the second difficulty, let us study the time evolution in the presence of noise. Suppose, e.g., that the system has evolved freely from noise for $t<t'$, where $t_{m-1}<t'<t_m$, until a bit-flip noise $X_1$ acts on qubit $1$ at $t'$. Then the state at $t_m$ is evaluated as $$\begin{aligned} && e^{-iH(t_m- t' )}X_1e^{-iH t' }|\psi_L^0\rangle \nonumber\\ && \quad = e^{-i\left[g_2(2t'-t_m)+ \sum_{s \neq 2} g_st_m\right]} X_1e^{iJ(2 t' -t_m)(Z_1Z_4+Z_7Z_1)} \nonumber\\ && \quad \quad \times (iZ_4Z_7)^m e^{i \omega (2 t' -t_m)X_L} |\psi_L^0\rangle. \label{state.X1}\end{aligned}$$ The terms proportional to $g_s$’s are irrelevant because they contribute only to an overall phase factor. Therefore, $g_s$’s may take arbitrary values. The problem is that the above state is different from the correctable state $X_1|\psi(t_m)\rangle$, not only in the term generated by $Z_1Z_4+Z_7Z_1$ and $(iZ_4Z_7)^m$ but also in the wrong phase of the oscillation $\omega(2t'-t_m)$. That is, extra errors occur because the bit-flip error in qubit $1$ (or $4$ or $7$) is ‘propagated’ by $H$ to other qubits [^1]. As a result, QEC at $t_m$ cannot recover the correct state. To overcome this difficulty, we note that the syndrome measurement for bit-flip errors (unlike that for phase errors) does not misidentify the state of Eq. (\[general period dynamics\]) as a wrong state. Hence, one can successfully perform QEC for bit-flip errors frequently (i.e., with intervals $\nu$ which are much shorter than $\tau$) in the interval between $t_{m-1}$ and $t_{m}$ for all $m$. As will be confirmed in the next section, this reduces the probability of errors small enough. Our prescription is summarized as follows: Perform QEC for both phase and bit-flip errors at all $t_m$’s (i.e., with intervals $\tau$), and perform QEC for bit-flip errors repeatedly with intervals $\nu$ ($\ll \tau$). The latter intervals $\nu$ are not required to be regular. We call this protocol the [*error-correction sequence*]{}. Effects of the error-correction sequence {#sec:calculations} ======================================== To see how well the error-correction sequence protects the discrete logical Rabi oscillation against noise, let us calculate the time evolution for $t_0 < t \leq t_1$, i.e., for $0 < t \leq \tau$, quantitatively. We divide the interval $(0,t]$ into $N$ subintervals; $(0,\Delta t], (\Delta t, 2\Delta t], (2\Delta t, 3\Delta t], \cdots$, where $\Delta t \equiv t/N$. We model noise by the product of depolarizing channels [@NC00] $\Pi_{i=1}^9 {\cal E}_{\Delta t}^{(i)}$, where ${\cal E}_{\Delta t}^{(i)}$ acts on qubit $i$ at the end of every subinterval as $${\cal E}_{\Delta t}^{(i)}[ \rho ]\equiv (1-\epsilon \Delta t )\rho+ {\epsilon \Delta t \over 3} \sum_{\alpha=1}^3\sigma_{\alpha}^{(i)} \rho \sigma_{\alpha}^{(i)}.$$ Here, $\rho$ denotes an input state, and $\epsilon$ is a small positive parameter representing the strength of the interaction with the environment. The initial state at $t=0$ is denoted by $\rho_L^0$, which is assumed to be in the code space. We study its time evolution up to the first orders in $\epsilon \tau$ and $\omega \tau$, assuming that $$\epsilon \tau \ll 1 \mbox{ and } \omega \tau \ll 1, \label{small_parameters}$$ where the latter comes from condition (\[omega&lt;&lt;J\]). If noise and QEC were absent, $\rho_L^0$ would evolve into $$\rho_H(t) \equiv e^{-iH t} \rho_L^0 e^{iH t} =e^{-iH_D t} \rho_L^0 e^{iH_D t}.$$ When noise is present but QEC is not performed, on the other hand, $\Pi_{i=1}^9 {\cal E}_{\Delta t}^{(i)}$ acts at the end of every subinterval. When $N=2$, for example, $\rho_L^0$ evolves into $$\begin{aligned} & \prod_{i=1}^9{\cal E}_{\Delta t}^{(i)} \left[e^{-iH\Delta t} \prod_{i=1}^9{\cal E}_{\Delta t}^{(i)} \left[\rho_H(\Delta t)\right] e^{iH\Delta t} \right] \\ & =(1-9\epsilon t)\rho_H(t) +\frac{\epsilon}{3}\sum_{j=1}^2\Delta te^{-iH(t-j\Delta t)} \sum_{i=1}^9\sum_{\alpha=1}^3 \sigma_{\alpha}^{(i)}\rho_H(j\Delta t)\sigma_{\alpha}^{(i)} e^{iH(t-j\Delta t)}. \end{aligned}$$ By taking $N\rightarrow \infty$, we obtain the state at $t$ without QEC as $$\begin{aligned} \rho(t,\rho_L^0) \simeq & (1-9\epsilon t)\rho_H(t) \nonumber\\ & +\frac{\epsilon}{3} \int_0^{t}dt'\, e^{-iH(t-t')} \sum_{i=1}^9\sum_{\alpha=1}^3 \sigma_{\alpha}^{(i)}\rho_H(t') \sigma_{\alpha}^{(i)} e^{iH(t-t')}.\end{aligned}$$ We calculate how this state is corrected by the error-correction sequence, in which bit-flip errors are corrected with intervals $\nu$ and both bit-flip and phase errors are corrected at $t=\tau$. Although the intervals $\nu$ are not required to be regular, and $$n \equiv \tau/\nu \label{def:n}$$ is not required to take an integral value, we here assume for simplicity that $\nu$ is regular and $n$ is an integer. We label qubits in and outside the central triangle of Fig. \[qubits configuration\] by $r, r'$ ($=1,4,7$) and $s$ ($=2,3,5,6,8,9$), respectively. At $t=\nu$, QEC for bit-flip errors is performed. The pre-measurement state of the syndrome measurement is $\rho(\nu,\rho_L^0)$. The post-measurement state $\rho'(\nu)$ depends on the outcome of the syndrome measurement. For example, when the bit-flip error in qubit $s$ is detected (which happens with probability $2\epsilon \nu/3$), $$\rho'(\nu) = \frac{1}{2} X_s \rho_H(\nu) X_s +\frac{1}{2} Y_s\rho_H(\nu) Y_s.$$ By the recovery operation, $\rho'(\nu)$ is changed into $$\rho''(\nu) \equiv X_s \rho'(\nu) X_s = \frac{1}{2} \rho_H(\nu) +\frac{1}{2} Z_s\rho_H(\nu) Z_s,$$ which is a mixture of the correct state $\rho_H(\nu)$ and $Z_s\rho_H(\nu) Z_s$, the state with a phase error in qubit $s$. At this stage, QEC for [*phase*]{} error is [*not*]{} performed because $\rho_H(\nu)$ is out of the code space. At $t=2 \nu$, QEC for bit-flip errors is performed again. The pre-measurement state is $\rho(\nu, \rho''(\nu))$, where $\rho''(\nu)$ corresponds to one of possible outcomes of the previous syndrome measurement at $t=\nu$. We can calculate $\rho'(2\nu)$ and $\rho''(2\nu)$ in the same way as we have calculated $\rho'(\nu)$ and $\rho''(\nu)$. By repeating the arguments $n$ times, we obtain the probabilities of bit-flip errors during $0 < t <\tau$ and the corresponding states $\rho''(\tau)$ that are obtained at $t=n \nu = \tau$ by correcting the bit-flip errors. To the first orders in $\epsilon \tau$ and $\omega \tau$, they are given by $$\begin{array}{ccc} \hline \mbox{error} & \mbox{probability} & \mbox{corrected state } \rho''(\tau)\\ \hline \mbox{none} & 1-6\epsilon \tau & (1-3\epsilon \tau)\rho_H(\tau)+\frac{\epsilon \tau}{3}\sum_i Z_i\rho_H(\tau)Z_i, \\ X_s & 2\epsilon \tau/3 & \frac{1}{2}\rho_H(\tau)+\frac{1}{2}Z_s\rho_H(\tau)Z_s, \\ X_r & 2\epsilon \tau/3 & \frac{1}{2}\rho_e^{(r)}(\tau)+\frac{1}{2}Z_r\rho_e^{(r)}(\tau)Z_r, \vspace{1mm} \\ \hline \end{array} \label{eq:t=tau-bf}$$ where $$\rho_e^{(r)}(\tau) \equiv \int_0^{\nu} e^{2iJ Z_r X_L t'} \rho_H(\tau-2t') e^{-2iJ Z_r X_L t'} \frac{dt'}{\nu}.$$ Finally at $t=\tau$, phase errors in $\rho''(\tau)$ are detected and corrected. We denote the state after this QEC by $\rho'''(\tau)$. Since $\rho''(\tau)$ depends on which qubit has suffered from a bit-flip error for $0 < t < \tau$, so does $\rho'''(\tau)$. If a bit-flip error has occurred in no qubit or in qubit $s$, $\rho'''(\tau)$ agrees with the correct state $\rho_H(\tau)$. If, on the other hand, a bit-flip error has occurred in qubit $r$ (with probability $2\epsilon \tau/3$, see above), the conditional probability of each outcome of the syndrome measurement for phase errors and the corresponding $\rho'''(\tau)$ are given by [^2] $$\begin{array}{ccc} \hline \mbox{error} & \mbox{probability} & \mbox{corrected state } \rho'''(\tau)\\ \hline \mbox{none or } Z_r & \frac{3}{8}+\frac{1}{8}{\rm sinc}\frac{\displaystyle 4\pi}{\displaystyle n} & {\displaystyle \frac{a_n^{+}\rho_H(\tau)+a_n^{-}X_L\rho_H(\tau)X_L} {a_n^{+}+a_n^{-}} }, \\ Z_{r'} \ (r' \neq r) & \frac{1}{8}-\frac{1}{8}{\rm sinc}\frac{\displaystyle 4\pi}{\displaystyle n} & \frac{1}{2} \left[\rho_H(\tau)+X_L\rho_H(\tau)X_L \right]. \vspace{1mm} \\ \hline \end{array}$$ Here, ${\rm sinc}\, x \equiv (\sin x) /x$, $ a_n^{\pm} \equiv \frac{3}{16} +\frac{1}{16}{\rm sinc}{4 \pi \over n} \pm \frac{1}{4}{\rm sinc}{2 \pi \over n} $, and terms of $O(\epsilon \tau)$ and $O(\omega \tau)$ have been dropped because the probability that a bit-flip error has occurred is already of $O(\epsilon \tau)$. By averaging $\rho'''(\tau)$ over all possible branches, we obtain the average state $\rho_{c}(\tau)$ under the error-correction sequence as $$\rho_{c}(\tau)=\rho_H(\tau) -\epsilon \tau \left[ 1-{\rm sinc}\frac{2 \pi}{n} \right] \left[ \rho_H(\tau)-X_L\rho_H(\tau)X_L \right].$$ Therefore, $\rho_{c}(\tau)$ approaches the correct state $\rho_H(\tau)$ with increasing $n$. This can be seen more clearly from their trace distance [@NC00], which is calculated for $n \gg 1$ as $${1 \over 2} {\rm tr} \left| \rho_c(\tau) -\rho_H(\tau) \right| \simeq 2 \pi^2 \epsilon \tau L_{yz}(\tau)/3 n^2.$$ Here, $L_{yz}(\tau)$ denotes the length of the projection onto the $y$-$z$ plane of the Bloch vector of $\rho_H(\tau)$ in the code space. Hence, by taking $$n \gtrsim (1/\sqrt{\epsilon \tau}) \min \{ 1, \epsilon/\omega \}, \label{eq:n}$$ we can reduce the distance to about $ 6 L_{yz}(\tau) \max\{ (\epsilon \tau)^2, (\omega \tau)^2 \} $. Since $L_{yz}(\tau) = O(1)$, this is of the same order as the largest term that has been dropped in the above calculations. That is, we have successfully recovered the correct state at $t=\tau$ ($=t_1$), i.e., $\rho_{c}(t_1)=\rho_H(t_1)+ O\left(\max\{ (\epsilon \tau)^2, (\omega \tau)^2 \} \right)$. In a similar manner, we can evaluate $\rho_{c}(t_m)$ by taking $\rho_{c}(t_{m-1})$ as the initial state, and find that $$\rho_{c}(t_m)=\rho_H(t_m)+ O\left(\max\{ (\epsilon \tau)^2, (\omega \tau)^2 \} \right)$$ for all $m$. Therefore, the discrete logical Rabi oscillation is protected, with only $O\left(\max\{ (\epsilon \tau)^2, (\omega \tau)^2 \} \right)$ probability of failure, if we take $n$ as Eq. (\[eq:n\]). For example, we should take $n \gtrsim 10^2$ when $\epsilon \tau = \omega \tau =10^{-4}$. Figure \[Error-correction sequence\] demonstrates how the error-correction sequence corrects errors, i.e., how the solid line approaches the dashed line. ![ Probability of finding $|0_L\rangle$ plotted against time when $\rho_L^0=|0_L \rangle \langle 0_L|$, for the logical Rabi oscillation (chain line), the Hamiltonian evolution by $H$ (dashed line), that affected by noise (dotted line), and that corrected by the error-correction sequence (solid line). The dots represent the discrete logical Rabi oscillation. Here, $\omega/J=0.1$, $\epsilon \tau=\pi/100$, $\tau/\nu=6$. Inset: magnification around $t=\nu$. []{data-label="Error-correction sequence"}](Figure_4.eps){width="0.93\linewidth"} Arbitrariness of the parameters in $H_D$ {#sec:kk'} ======================================== It is clear from the results of Secs. \[sec:concepts\] and \[sec:calculations\] that the value of $J$ is arbitrary as long as $\omega \ll J$. On the other hand, we have assumed in those sections that $k_1=k_4=k_7=1$. In this section, we show that the error-correction sequence is successful also when $k_r$’s take other values. Recall that the error-correction sequence consists of two parts; QEC for both phase and bit-flip errors at all $t_m$’s, and QEC only for bit-flip errors with intervals $\nu$. The latter part is successful even when $k_r$’s are arbitrary real numbers, because in general a Hamiltonian which does not contain $X_i$’s and $Y_i$’s, such as the proposed $H$, cannot flip the bit of any physical qubit. Hence, the syndrome measurement for bit-flip errors does not misidentify the state evolved by such a Hamiltonian as a wrong state. Regarding the former part, we start with showing that $k_r$’s can be arbitrary integers. Note that QEC at $t_m$’s works well provided that the state of the qubits at $t_m$ would be in the code space if noise were absent. As discussed in Sec. \[sec:concepts\], this condition is satisfied when $k_1=k_4=k_7=1$, because $ |\psi(t_m)\rangle =e^{i\omega t_mX_L}|\psi_L^0\rangle $, which is certainly in the code space. When $k_r$’s are odd integers, we obtain the same result; $$\begin{aligned} |\psi(t_m)\rangle &=& e^{-iH_Dt_m}|\psi_L^0\rangle \nonumber\\ &=& e^{i\omega t_mX_L} e^{i\frac{m\pi}{2}(k_1 Z_1Z_4+k_4 Z_4Z_7+k_7 Z_7Z_1)} |\psi_L^0\rangle \nonumber\\ &=& e^{i\omega t_mX_L} \prod_{r=1,4,7} \left[\cos\frac{m k_r \pi}{2}+i\sin\frac{m k_r \pi}{2} Z_r Z_{r+3} \right] |\psi_L^0\rangle \nonumber\\ &=& e^{i\omega t_mX_L}|\psi_L^0\rangle, \end{aligned}$$ apart from irrelevant phase factors. Here, $Z_{10} \equiv Z_1$, and we have used $(Z_1Z_4)(Z_4Z_7)(Z_7Z_1)=1$. When $k_r$’s are general integers (not necessarily odd), on the other hand, we have to add a certain procedure to the error correction sequence. We explain this for the case where $k_1=1$ and either one of $k_4, k_7$ is even. In this case, we find that $$|\psi(t_m)\rangle = \begin{cases} e^{i\omega t_mX_L}|\psi_L^0\rangle & \mbox{for even $m$}, \\ Z_rZ_{r'}e^{i\omega t_mX_L}|\psi_L^0\rangle & \mbox{for odd $m$}. \end{cases}$$ Here, $r$ and $r'$ $(\neq r)$ each is $1,4$ or $7$ depending on $k_4,k_7$. For example, when $k_4$ is even and $k_7$ is odd, $ |\psi(t_m)\rangle =Z_7Z_1e^{i\omega t_mX_L}|\psi_L^0\rangle $ for odd $m$. Although this state is out of the code space, we note that the evolution into this state is not a stochastic process (such as evolution by noise) but a deterministic process induced by the known Hamiltonian $H$ [^3]. Hence, we can surely change this state to $e^{i\omega t_mX_L}|\psi_L^0\rangle$ by applying $Z_7Z_1$ just before QEC at $t_m$. By adding this procedure to the error correction sequence, we can successfully perform QEC at $t_m$’s. Thus, the error-correction sequence, supplemented with this additional procedure, works well when $k_r$’s are arbitrary integers. Note that if $k_r$’s have a common factor $K$, one can redefine $k_r$’s and $J$ as $$J' = KJ, \quad k'_r = k_r/K.$$ The corresponding terms in $H_D$ are then rewritten as $$J \sum_{r=1,4,7} k_r Z_r Z_{r+3} = J' \sum_{r=1,4,7} k'_r Z_r Z_{r+3}.$$ Hence, one can use $J'$ instead of $J$, which means, e.g., that $\tau' \equiv \pi / 2 J'$ is used instead of $\tau$. The error correction sequence has such flexibility. We next consider a more general case where $k_r$’s are rational numbers. Suppose, for example, that $k_1=1, k_4=3/2, k_7=5/3$. Then, one can redefine $k_r$’s and $J$ as $J' = J/6, k'_r = 6k_r$, and the corresponding terms in $H_D$ are rewritten as $$J \sum_{r=1,4,7} k_r Z_r Z_{r+3} = J'(6 Z_1Z_4+9 Z_4Z_7+ 10 Z_7Z_1).$$ Therefore, if one uses $\tau' \equiv \pi / 2 J'$ instead of $\tau$, the error correction sequence is successful. In general, if there exists a real number $\kappa $ such that $\kappa k_r$’s are integers and $$J' \equiv J/\kappa \gg \omega, \label{eq:J/kappa}$$ then the error correction sequence is successful if one uses $\tau' \equiv \pi / 2 J'$ instead of $\tau$. Finally, we consider the case where $k_r$’s are irrational numbers. We note that an irrational number can be well approximated by rational numbers. When $k_1 = \pi$ ($=3.14159\cdots$), for example, it can be approximated by $22/7$ ($=3.14285\cdots$), $333/106$ ($=3.14150\cdots$), and so on. Let $k_{1*}$ be such a rational number. The difference $k_1 - k_{1*}$ is negligible if $J \left| k_1 - k_{1*} \right| t \ll 1$. Therefore, for the time interval $t$ that satisfies $$t \ll 1/J \left| k_1 - k_{1*} \right|, \label{upperlimit-t}$$ this case reduces to the one where $k_r$’s are rational numbers. If one takes $k_{1*}$ such that $\left| k_1 - k_{1*} \right|$ is smaller, the upper limit of $t$ given by Eq. (\[upperlimit-t\]) becomes longer, whereas condition (\[eq:J/kappa\]) becomes harder to satisfy because the denominator of $k_{1*}$ becomes greater. To summarize this section, the error-correction sequence works well for fairly arbitrary values of $k_r$’s. Although it is better that one can successfully fabricate the system in such a way that $k_r$’s are integers, one can also accept most systems which have non-integral values of $k_r$’s (because of fabrication errors). This fact makes the preparation of the system easier. Derivation of the effective interactions {#sec:effectiveH} ======================================== The proposed Hamiltonian $H$ consists of Ising-type interactions and three-qubit interactions among physical qubits. We here discuss how they are generated as effective interactions from more fundamental interactions. Many physical systems can be candidates for physical qubits that have the proposed $H$. As an example, we here consider quantum dots in a semiconductor [@BDEJ95; @BW98]. To be more concrete, we assume that the spin of an electron in a dot is polarized by a high external magnetic field, so that we can forget about the spin degrees of freedom. We also assume that the potential barrier between the dots is high and thick so that electron tunneling between the dots is negligible. This and (possibly) the Coulomb interaction, by which states with two electrons in a single dot have much higher energies than states with a single electron, exclude double occupancy of a dot. For single-electron states of a dot, we assume that only the ground and the first excited states, denoted by $|0\rangle$ and $|1\rangle$, are relevant because higher states have much higher energies and/or the transition matrix elements to them are small. As a result, we can treat each dot as a system with two quantum levels, $|0\rangle$ and $|1\rangle$, i.e., as a qubit. For the reasons that will be explained below, we also assume that all dots in a logical qubit are asymmetric and different (in size and/or shape) so that accidental degeneracy is lifted. The effective Hamiltonian $H_{\rm eff}$ for a set of such qubits (dots) is the sum of single-qubit terms and the effective interactions. The effective interactions are derived from more elementary interactions $V, W, \cdots$, which are [*effective*]{} interactions among [*conduction*]{} electrons in [*homogeneous*]{} bulk semiconductors. On the other hand, $V, W, \cdots$ are derived from even more elementary interactions, such as the Coulomb interactions between electrons in vacuum. Since two- and three-body interactions have been studies in many physical systems (see, e.g., Refs. [@SCH03; @BMZ07]), we here consider a two-body interaction $V$ and a three-body interaction $W$. Generally, the latter is much weaker than the former [^4]. Since four- or more-body interactions are even weaker, we neglect them. We can represent $H_{\rm eff}$ as a polynomial of the Pauli operators. In general, it would have terms that include $X_i\equiv |0\rangle_i \, \null_i \langle 1 |+| 1 \rangle_i \, \null_i \langle 0 |$ and $Y_i\equiv -i| 0 \rangle_i \, \null_i \langle 1 |+i| 1 \rangle_i \, \null_i \langle 0 |$, where the subscript $i$ ($=1, 2, \cdots$) labels the qubits. Such terms are non-diagonal terms that are proportional to $|n\rangle \langle m|$ $(m\neq n)$, where $|n\rangle$ and $|m\rangle$ are product states of $| 0 \rangle_i $’s and $| 1 \rangle_i $’s (such as $\prod_i | 0 \rangle_i $). As discussed in Refs. [@BB04] and [@VC04] and in Appendix \[sec:XY\], contributions from the non-diagonal terms to the time evolution are negligible if $$\left| \frac{\langle n |(V+W)|m\rangle}{\Delta E_{nm}} \right| \ll 1 \quad \mbox{for every $n,m$ ($\neq n$)}, \label{eq:xi=small}$$ where $\Delta E_{nm}$ is the difference in energy of single qubit terms between $|n\rangle$ and $|m\rangle$. \[A more precise expression of this condition is given in Appendix \[sec:XY\], where $\langle n |H'|m\rangle$ corresponds to $\langle n |(V+W)|m\rangle$.\] In typical situations, $V$ and $W$ are significant only between [*adjacent*]{} dots (such as dots $1, 4, 7$, dots $1, 2$, dots $2, 3$, and so on, of Fig. \[qubits configuration\]) because $V$ and $W$ generally decrease as the distance is increased. In such a case, one can make $|\Delta E_{nm}|$ larger than $\left| \langle n |(V+W)|m\rangle \right|$ by making the sizes and/or shapes of adjacent dots different. One can also make $|\Delta E_{nm}|$ larger by modulating spatially the magnitude of the external magnetic field. If condition (\[eq:xi=small\]) is satisfied by these methods, one can drop non-diagonal terms, and hence $H_{\rm eff}$ reduces to $H$, which consists only of $Z_i=| 0 \rangle_i \, \null_i \langle 0 |-| 1 \rangle_i \, \null_i \langle 1 |$’s, when considering the time evolution. On the conditions and assumptions mentioned above, $H$ can be derived simply by taking the diagonal matrix elements, between $| n \rangle$’s, of the effective Hamiltonian for conduction electrons, $$H_0^{\rm el}+V+W, \label{H3:org}$$ where $H_0^{\rm el}$ denotes the non-interacting part, which includes the confining potential of the dots. We here present explicit results for the three qubits in the central triangle of Fig. \[qubits configuration\]. Interactions between the other qubits can be derived more easily in a similar manner. Since the potential barrier is high, the wavefunctions $\psi_r^0({\bm r})$ and $\psi_r^1({\bm r})$ of $|0\rangle_r$ and $|1\rangle_r$, respectively, are well localized within each dot. As a result, overlap of the wavefunctions of different dots is negligibly small, i.e., $\psi_r^a({\bm r})\psi_{r'}^b({\bm r})\simeq 0$ for $r\neq r'$ and for all $a,b$ ($=0,1$). Using this fact, we find that the effective Hamiltonian is given by $$-\frac{1}{2}\sum_{r=1,4,7}\zeta_r Z_r-\sum_{r>r'}J_{rr'}Z_rZ_{r'}-\omega Z_1Z_4Z_7, \label{H3:eff}$$ where, for $a,b,c=0,1$, $$\begin{gathered} \begin{split} \zeta_1=\zeta^0_1 -\frac{1}{2} \sum_{a,b} (-1)^{a}(V_{a \bullet b}+V_{a b \bullet}) -\frac{1}{4}\sum_{a,b,c}(-1)^aW_{abc} \label{eq:zeta} \end{split} \\ \begin{split} J_{14}=-\frac{1}{4}\sum_{a,b}(-1)^{a+b}V_{ab\bullet} -\frac{1}{8}\sum_{a,b,c}(-1)^{a+b}W_{abc}, \end{split} \\ \omega = -\frac{1}{8}\sum_{a,b,c}(-1)^{a+b+c}W_{abc},\end{gathered}$$ and similarly for the other $\zeta_r$’s and $J_{rr'}$’s. Here, $\zeta_r^0$ is the energy difference between $|1 \rangle_r$ and $|0 \rangle_r$, and $$\begin{aligned} V_{ab\bullet} &\equiv& \iint \left|\psi_{1}^{a}({\bm r}) \right|^2 V({\bm r},{\bm r}') \left| \psi_{4}^{b}({\bm r}') \right|^2 d{\bm r}d{\bm r}', \\ W_{abc} &\equiv& \iiint W({\bm r},{\bm r}',{\bm r}'') \left| \psi_{1}^{a}({\bm r}) \right|^2 \left| \psi_4^{b}({\bm r}') \right|^2 \left| \psi_7^{c}({\bm r}'') \right|^2 d{\bm r}d{\bm r}'d{\bm r}'',\end{aligned}$$ and similarly for $V_{a \bullet b}, V_{\bullet ab}$. In fact, one can easily verify that all the diagonal matrix elements of Eq.(\[H3:org\]), between $| n \rangle$’s, agree with those of Eq.(\[H3:eff\]). It is seen that the single-dot energy $\zeta_r$ is renormalized by the interactions $V$ and $W$, and the two-qubit effective interactions are generated from $V$ and $W$, whereas the three-qubit effective interaction is generated from $W$. Regarding the magnitudes of the effective coupling constants, $\omega$ is much smaller than $J_{rr'}$’s because the former is derived only from the weaker interaction $W$. Note that $\omega$ does not vanish by accidental degeneracy because we have assumed that all dots in a logical qubit are asymmetric and different. Since we can forget about the single-qubit terms (i.e., the first term of Eq. (\[H3:eff\])) as discussed in Appendix \[Single terms\], Eq. (\[H3:eff\]) agrees with the proposed $H_D$, Eq. (\[Rabi hamiltonian\]), where $J_{rr'}$ correspond to $k_r J$. Discussions and Conclusions {#sec:DandC} =========================== We have shown in Secs. \[sec:concepts\] and \[sec:calculations\] that two-qubit interactions in $H_D$ cause errors which are correctable not by the straightforward application of QEC but by the error-correction sequence. One might expect that such errors could be corrected more easily by using more elaborate codes such as the one in Ref. [@Ruskai]. If such codes are used, however, $X_L$ in $H_D$ becomes an interaction among three or more qubits. Generally, if $l$ qubits are crowded to induce an $l$-qubit interaction corresponding to $X_L$, unwanted interactions among $l'$ ($<l$) qubits are also induced, as we have discussed on $H$. For any code that can correct all single-qubit errors, some of such unwanted interactions are [*not*]{} elements of the stabilizer [^5]. If $l' \geq 3$ like the code of Ref. [@Ruskai], they cause errors which cannot be corrected even by the error-correction sequence. If $l'<3$ like the Shor code and the Steane code [@St96], they can be dealt with the error-correction sequence. We have also shown that the values of $g_s$’s in $H_S$ are arbitrary. Such great flexibility would not be obtained if we employed a non-degenerate code [@NC00], because its stabilizer does not include two-fold tensor products of the Pauli operators. For example, the Steane code is a non-degenerate code and hence it has less flexibility. For these reasons, we have employed in this paper the Shor code, which is a degenerate code with $l=3$ (because we can take $X_L=Z_1 Z_4 Z_7$) and $l'=2$. Possibility of use of other codes is worth exploring. It is also worth exploring the possibility of replacing a circuit for the syndrome measurements with another natural interactions. Our preliminary study indicates that this is basically possible, and more detailed studies are in progress. Furthermore, it is interesting to apply the present idea to general time evolutions (such as general SU(2) rotations) and/or to general systems (such as systems composed of many logical qubits). A possible way of realizing this may be mixed use of an Hamiltonian (such as the one of this paper) and simple quantum circuits. This might also be applicable to quantum simulations [@Feynman; @Lloyd]. Since these subjects are beyond the scope of the present paper, we leave them as subjects of future studies. In conclusion, we have shown that the Rabi oscillation of a logical qubit encoded by the Shor code can be induced by a Hamiltonian that is composed of natural short-range interactions among physical qubits (Sec. \[sec:Hamiltonian\]). The Hamiltonian replaces the most complicated part of a quantum circuit that would be necessary for inducing and protecting the logical Rabi oscillation. More specifically, the state driven by the proposed Hamiltonian agrees with the logical Rabi oscillation at discrete times $t_m=m\tau$ ($m=0,1,2,\cdots$), which is quasi continuous as shown in Fig. \[Discrete Rabi oscillation\]. We call it the discrete logical Rabi oscillation (Sec. \[sec:concepts\]). To prepare a physical system that has the proposed Hamiltonian, one has simply to place two-level systems (which are used as physical qubits), such as asymmetric quantum dots (Sec. \[sec:effectiveH\]), as shown in Fig. \[qubits configuration\]. The parameters of this system, such as the positions and the sizes of the dots, are to a great extent arbitrary because the proposed Hamiltonian has great flexibility (Secs. \[sec:Hamiltonian\] and \[sec:kk’\]). This makes the fabrication of the system easier. Once the fabrication is finished, one can measure the coupling constants of the effective interactions, and the important parameters such as $\tau$ can be calculated from them. To protect the discrete logical Rabi oscillation against noise, the ordinary QEC cannot be applied straightforwardly. However, we have shown that it can be protected by a new protocol, which we call the error-correction sequence (Secs. \[sec:concepts\] and \[sec:calculations\]). In this protocol, QEC for both phase and bit-flip errors is performed at $t_m$’s, whereas QEC only for bit-flip errors is performed frequently in the interval between $t_{m-1}$ and $t_m$ for all $m$. The frequency of the latter is determined by the strength of noise and the parameters of the effective interactions (Sec. \[sec:calculations\]). One can realize the protected Rabi oscillation by using the natural Hamiltonian (to induce the logical Rabi oscillation) and a quantum circuit for the error-correction sequence. This is much easier than realizing it wholly with a fault-tolerant quantum circuit. The authors thank Y. Matsuzaki for discussions. This work is partly supported by KAKENHI. $X_L$ is an interaction among three or more qubits {#app:3 or more} ================================================== Let $P_c$ be the projection operator onto the code space; $$P_c =|0_L\rangle \langle 0_L|+|1_L\rangle \langle 1_L|.$$ An $n$-qubit code which can correct all single-qubit errors satisfies the following condition [@NC00]; $$P_c \sigma_{\alpha}^{(i)} \sigma_{\beta}^{(j)} P_c =\chi_{i \alpha, j \beta} P_c % =\chi_{\alpha \beta}^{ij} P_c \quad (i,j=1,\cdots , n; \ \alpha,\beta=0,1,2,3). \label{eq:Pc}$$ Here, $\sigma_{\alpha}^{(i)}$ denotes the identity ($\alpha=0$) and Pauli ($\alpha=1,2,3$) operators acting on qubit $i$, and $\chi_{i \alpha, j \beta}$ is an element of some Hermitian matrix. If $X_L$ were a Pauli operator or a two-fold tensor product of Pauli operators, Eq. (\[eq:Pc\]) could not be satisfied. For example, if $X_L=X_1 X_2$ for some code the left-hand side of Eq. (\[eq:Pc\]) with $\sigma_{\alpha}^{(i)}=X_1, \sigma_{\beta}^{(j)}=X_2$ (i.e., $i=1, j=2, \alpha=\beta=1$) reduces to $$P_cX_1X_2P_c=P_cX_LP_c =|0_L\rangle \langle 1_L|+|1_L\rangle \langle 0_L|.$$ Since this is neither vanishing nor proportional to $P_c$, Eq. (\[eq:Pc\]) is not satisfied for any value of $\chi_{11}^{12}$. This means that such a code cannot correct all single-qubit errors. Therefore, $X_L$ is a three- or more-fold tensor product of the Pauli operators (which corresponds to an interaction among three or more qubits) for any code that can correct all single-qubit errors. Irrelevance of single-qubit terms {#Single terms} ================================= When two levels of a qubit have different energies, a single-qubit term, which represents the energy difference, arises in its effective Hamiltonian as discussed in Sec. \[sec:effectiveH\]. All effects of such single-qubit terms can be canceled if we do everything in the rotating frame [@VC04]. Although this fact seems to be known widely, we here explain it in order to clarify its meaning in the context of QEC. Let us investigate the time evolution of a state $\rho_+$ by the following Hamiltonian $$H_+ =-{1 \over 2}\sum_{i=1}^9\zeta_i Z_i + H =-{1 \over 2}\sum_{i=1}^9\zeta_i Z_i + H_D+H_S,$$ where $\zeta_i$’s are real numbers. We can go to the rotating frame (an interaction picture) by $U_0\equiv \exp({i \over 2}\sum_{i=1}^9 \zeta_iZ_i t)$, as $ \rho^{\text{rot}}=U_0^{\dagger} \rho_+U_0 $. It evolves according to $$i\frac{d}{dt}\rho^{\text{rot}}=[U_0^{\dagger} H U_0, \rho^{\text{rot}}]=[H,\rho^{\text{rot}}],$$ where we have used $[H,U_0]=0$. Therefore, $\rho^{\text{rot}}$ undergoes the same unitary evolution as that of $\rho$ of Sec. \[sec:calculations\]. Furthermore, it is easy to show that the depolarizing channel in the rotating frame is also the same as the one in Sec. \[sec:calculations\]. Thus, in the presence of noise, $\rho^{\text{rot}}$ evolves in the same manner as $\rho$ of Sec. \[sec:calculations\]. Therefore, the error correction sequence will be successful if we set the initial state $\rho^{\text{rot}}(0)$ in the code space and perform QEC in the rotating frame. For example, the observables for the syndrome measurement in the rotating frame are $ M_{b_1}^{\text{rot}} \equiv Z_1Z_2 $, $ M_{p_1}^{\text{rot}} \equiv X_1X_2X_3X_4X_5X_6 $, and so on. In the laboratory frame (Schrödinger picture), they are given by $ M_{b_1}=U_0 Z_1Z_2 U_0^{\dagger}=Z_1Z_2 $ and $ M_{p_1}=U_0 X_1X_2X_3X_4X_5X_6 U_0^{\dagger} =\prod_{i=1}^6 \exp({i \over 2}\zeta_i Z_i t)X_i\exp(-{i \over 2}\zeta_i Z_i t) $, respectively. Irrelevance of terms including $X, Y$ {#sec:XY} ===================================== It seems widely accepted by researchers of NMR that the non-diagonal terms, which include $X_i$’s and/or $Y_i$’s, in $H_{\rm eff}$ are irrelevant to the time evolution if condition (\[eq:xi=small\]) is satisfied (see, e.g., Refs. [@BB04] and [@VC04]). For completeness, we here show that this is indeed true under reasonable assumptions. Let us decompose $H_{\rm eff}$ as $$H_{\rm eff} = H_0 + H + H', \quad H_0 \equiv - {1 \over 2}\sum_i\zeta_i Z_i,$$ where $\zeta_i$ is the energy difference (that is renormalized, like Eq. (\[eq:zeta\]), by interections among dots) between $| 1 \rangle_i $ and $| 0 \rangle_i$, $H$ is a polynomial of $Z_i$’s only, and $H'$ consists of the other terms (such as $X_1 Y_1$, $X_1 X_2 Z_3$, and so on) which include $X_i$’s and/or $Y_i$’s. We denote a product state of $| 1 \rangle_i $’s and $| 0 \rangle_i $’s, such as $\prod_i | 1 \rangle_i $, by $|n \rangle$. In terms of such product states, $H_0$ and $H$ are diagonal, whereas $H'$ gives the off-diagonal elements. To characterize the magnitude of the latter, we define the parameter $\xi_{nm}$ by $$\xi_{nm} \equiv \begin{cases} \displaystyle % \left| {\langle n | H' | m \rangle \over \Delta E_{nm} } % \right| & \mbox{if } \langle n | H' | m \rangle \neq 0, \\ 0 & \mbox{if } \langle n | H'| m \rangle = 0, \end{cases}$$ where $\Delta E_{nm}$ denotes the difference of the eigenvalues of $H_0$ between $|n\rangle$ and $|m\rangle$. We also define $$\bar{\xi} \equiv \sqrt{\sum_{n,m} \left| \xi_{nm} \right|^2}. % \max_{n,m} \left| \xi_{n,m} \right|.$$ Consider the time evolution operator $U_{\rm eff}(t)$ generated by $H_{\rm eff}$. We can write it as $$U_{\rm eff}(t) \equiv \exp \left( -i H_{\rm eff} t \right) = U(t) e^{-i Q(t)}, % Q^\dagger(t),$$ where $$U(t) \equiv e^{- i (H_0+H) t} = e^{- i H_0 t} e^{- i H t}, % U(t) \equiv \exp(- i H t) % = \exp(- i H_0 t) \exp(- i H_{\rm int} t),$$ and $Q(t)$ is the Hermitian operator defined by $\displaystyle e^{-i Q(t)} \equiv U^\dagger(t) U_{\rm eff}(t)$. It is clear that $$Q(t) = 0 \mbox{ when } \bar{\xi}=0. \label{eq:Q=0}$$ If $\bar{\xi}$ were large then $Q(t)$ would be significant, particularly when $\Delta E_{nm} = 0$ for all $n,m$, for which $\bar{\xi}=\infty$. On the other hand, if $\bar{\xi}$ is small enough $\| Q(t) \|$ is expected to be small, where $\| \cdot \|$ denotes the operator norm. It is natural to assume that $$\mbox{Assumption 1: } \mbox{$Q(t)$ is continuous in $\xi_{nm}$'s in the neighborhood of $\bar{\xi}=0$}. \label{eq:Q=cont}$$ This assumption seems reasonable from the perturbation expansion of the time evolution operator in the interaction picture, which corresponds to $e^{i H_0 t } U_{\rm eff}(t) = e^{- i H t} e^{-i Q(t)}$; $$\begin{aligned} && 1 -i t \sum_{n} \langle n| H |n\rangle |n\rangle \langle n| -i\sum_{n} \sum_{m \, (\neq n)} \int_0^{t}e^{i\Delta E_{nm}t'}\,dt' \langle n | H' |m\rangle |n\rangle \langle m| +\cdots \nonumber\\ &&= 1 -i t \sum_{n} \langle n| H |n\rangle |n\rangle \langle n| -\sum_{n} \sum_{m \, (\neq n)} \left( e^{i\Delta E_{nm}t} - 1 \right) \xi_{nm} |n\rangle \langle m| +\cdots, \label{eq:expansion}\end{aligned}$$ each term of which is continuous with respect to $\xi_{nm}$. Assumption 1, together with Eq. (\[eq:Q=0\]), means that for any small positive number $\varepsilon$ there exists a positive number $\bar{\xi}_{\varepsilon, t}$ such that $$\| Q(t) \| < \varepsilon \mbox{ for all } \bar{\xi} < \bar{\xi}_{\varepsilon, t}.$$ In other words, for a given time period $[0, t)$ we can neglect $Q(t)$, i.e., we can regard $U_{\rm eff}(t)=U(t)$, if $\bar{\xi}$ is small enough. This means that the time evolution by $H_{\rm eff}$ takes place as if $|n\rangle$’s (which are eigenstates of $H_0+H$) were its eigenstates. That is, if we expand an initial state in terms of $|n\rangle$’s as $\sum_{n} c_n |n\rangle$, $$e^{-i H_{\rm eff} t} % \exp \left( -i H_{\rm eff} t \right) \sum_{n} c_n |n\rangle \simeq \sum_{n} c_n e^{-i \langle n|(H_0+H)|n\rangle t} % \exp \left( -i \langle n| H|n\rangle t \right) |n\rangle.$$ Note that the above argument is general in the sense that we have not assumed any specific forms for $H$ and $H'$. For example, the argument in Appendix A of Ref. [@BB04], where specific forms have been assumed, is essentially a special case of the present general argument. In the above argument, we have not excluded the possibility that $\bar{\xi}_{\varepsilon, t}$ increases with increasing $t$. This will not cause difficulty when one sets an upper limit of $t$. To be more complete, however, we here discuss dependence of $\bar{\xi}_{\varepsilon, t}$ on $t$. We note that the coefficients of the third term of Eq. (\[eq:expansion\]) are upper bounded as $$\left| \left( e^{i\Delta E_{nm}t} - 1 \right)\xi_{nm} \right| \leq 2 \left| \xi_{nm} \right| \leq 2 \bar{\xi}$$ for all $t$. This is due to the fact that $t$ appears only through the oscillatory factor $e^{i\Delta E_{nm}t}$. Since this is the case also for higher-order terms, we expect that $$\mbox{Assumption 2: } \mbox{$\bar{\xi}_{\varepsilon, t}$ has an upper bound $\bar{\xi}_\varepsilon$, which is independent of $t$}. % for all $t$}.$$ If this is true, then for any small positive number $\varepsilon$ and for all $t$ $$\| Q(t) \| < \varepsilon \mbox{ for all } \bar{\xi} < % \bar{\xi}_{\varepsilon, t} \leq \bar{\xi}_\varepsilon. \label{Q<xie}$$ In other words, we can regard $U_{\rm eff}(t)=U(t)$ even for long $t$ if $\bar{\xi}$ is small enough. [99]{} See, e.g., A. Shimizu and T. Miyadera, Phys. Rev. Lett. [**89**]{}, 270403 (2002), and references cited therein. P. Zanardi, M. Rasetti, Phys. Rev. Lett. [**79**]{}, 3306 (1997). J. Kempe, D. Bacon, D. A. Lidar, and K. B. Whaley, Phys. Rev. A. [**63**]{}, 042307 (2001). L. Viola, E. Knill, and S. Lloyd, Phys. Rev. Lett. [**82**]{}, 2417 (1999). P. Facchi [*et al*]{}., Phys. Rev. A. [**71**]{}, 022302 (2005). P. W. Shor, Phys. Rev. A [**52**]{}, 2493 (1995). A. M. Steane, Phys. Rev. Lett. [**77**]{}, 793 (1996). M. A. Nielsen and I. L. Chuang, [*Quantum Computation and Quantum Information*]{} (Cambridge University Press, Cambridge, 2000). D. Gottesman, arXiv:quant-ph/9705052v1. J. Preskill, Proc. R. Soc. Lond. A(1998) [**454**]{}, 385. D. P. DiVincenzo [*et al*]{}., Nature [**408**]{}, 339 (2000). M. Mohseni and D. A. Lidar, Phys. Rev. Lett. [**94**]{}, 040507 (2005). J. K. Pachos and E. Rico, Phys. Rev. A [**70**]{}, 053620 (2004). M. S. Tame, M. Paternostro, M. S. Kim, and V. Vedral, Phys. Rev. A [**73**]{}, 022309 (2006). A. Barenco, D. Deutsch, A. Ekert and R. Jozsa, Phys. Rev. Lett. [**74**]{}, 4083 (1995). A. Balandin and K. L. Wang, Superlattices Microstruct. [**25**]{}, 509 (1999). P. Soldán, M. T. Cvitaš, and J. M. Hutson, Phys. Rev. A [**67**]{}, 054702 (2003). H. P. Büchler, A. Micheli and P. Zoller, Nature Physics [**3**]{}, 726 (2007). S. C. Benjamin and S. Bose, Phys. Rev. A. [**70**]{}, 032314 (2004). L. M. K. Vandersypen and I. L. Chuang, Rev. Mod. Phys. [**76**]{}, 1037 (2004). M. B. Ruskai, Phys. Rev. Lett. [**85**]{}, 194 (2000). R. Feynman, Int. J. Phys. [**21**]{}, 467 (1982). S. Lloyd, Science [**273**]{}, 1073 (1996). [^1]: Unlike the bit-flip errors, phase errors are not propagated by $H$. For example, if a phase error $Z_1$ acts on qubit $1$ at $t'$, the state at $t_m$ is $ e^{-iH(t_m- t' )}Z_1e^{-iH t' }|\psi_L^0\rangle =Z_1e^{i\omega t_mX_L}|\psi_L^0\rangle $ because $Z_1$ commutes with $H$. This state will be corrected by QEC for phase errors, which is performed at $t_m$. [^2]: The phase flip in $r=1$ ($4, 7$) is equivalent to the phase flip in $s=2$ or $3$ ($5$ or $6$, $8$ or $9$) for the Shor code. [^3]: The values of all the parameters in $H$ can be measured experimentally after the system is fabricated. [^4]: For example, when the dielectric constant $\epsilon$ for the screened Coulomb interaction $U=e^2/\epsilon|{\bm r}-{\bm r}'|$ between electrons, which are located at ${\bm r}$ and ${\bm r}'$, weakly depends on the location ${\bm r}''$ of another electron as $\epsilon=\epsilon_0+\delta \epsilon({\bm r}'')$ ($\epsilon_0 \gg |\delta \epsilon({\bm r}'')|$), then $U \simeq V+W$, where $V \equiv e^2/\epsilon_0|{\bm r}-{\bm r}'|$ and $W \equiv -e^2 \delta \epsilon({\bm r}'') /(\epsilon_0)^2|{\bm r}-{\bm r}'|$. This $W$ is much weaker than $V$ because $\left| W/V \right| = \left| \delta \epsilon/\epsilon_0 \right| \ll 1$. [^5]: For example, suppose that $X_L=X_1X_4X_7$ and unwanted interactions are $X_1X_4$, $X_4X_7$, and $X_7X_1$. If these unwanted interactions were elements of the stabilizer, we would have $|1_L\rangle = X_L|0_L\rangle =X_1 X_4 X_7 |0_L\rangle =X_1 |0_L\rangle$, which shows that $X_1$ would be another expression of $X_L$. However, this is impossible for any code that can correct all single-qubit errors because, as discussed in Appendix \[app:3 or more\], $X_L$ should be a three- or more-fold tensor product of the Pauli operators.

ArXiv

--- author: - | [**Haim Judah**]{}\ Bar-Ilan University\ Abraham Fraenkel Center\ for\ Mathematical Logic\ Department of Mathematics and Computer Science\ 52-900 Ramat-Gan, Israel\ - | [**Saharon Shelah**]{}\ Hebrew University\ Institute of Mathematics\ Jerusalem, Israel title: Baire Property and Axiom of Choice --- \[section\] \[theorem\][Lemma]{} \[theorem\][Corollary]{} \[theorem\][Proposition]{} \[theorem\][Definition]{} Introduction ============ In 1979 Shelah proved that in order to obtain a model in which every set of reals has Baire property, a large cardinal assumption is not necessary. The model he constructed satisfied $\omega_{1}^{L} = \omega_{1}$. Therefore Woodin asked if we can get a model for “ZF + DC($\omega_{1}$) + each set of reals has Baire property”.\ Recall here that DC($\omega_{1}$) is the following sentence: > if ${\cal R}$ is a relation such that $(\forall X)(\exists Y)({\cal R}(X,Y))$ then there is a sequence $<Z_{\alpha}:\alpha<\omega_{1}>$ such that $$(\forall \alpha<\omega_{1})({\cal R}(<Z_{\beta}:\beta<\alpha>\ ,\ > Z_{\alpha})).$$ Note that DC($\omega_{1}$) implies the following version of choice: > if ${\cal R}\subseteq\omega_{1}\times{{\bf R}}$\ > then there exists a choice function $f:\omega_{1}\longrightarrow{{\bf R}}$ such that ${\cal R}(\alpha, f(\alpha))$ for each $\alpha<\omega_{1}$. In \[JS1\] we studied the consistency strength of “ZFC + variants of MA + suitable sets of reals have Baire property”. We showed that Baire property for $\Sigma^{1}_{3}$-sets of reals plus MA($\sigma$-centered) implied that $\omega_{1}$ is a Mahlo cardinal in L. The natural question that arises at this point is: > Do we need large cardinals to construct a model in which all projective sets of reals have Baire property and the union of any $\omega_{1}$ meager sets is meager? Note that if unions of $\omega_{1}$ many null sets are null then every $\Sigma^{1}_{2}$-set of reals is Lebesgue measurable. Consequently if each projective sets of reals has Baire property and any union of $\omega_{1}$ null sets is null then $\omega_{1}$ is inaccessible in L. The aim of the present paper is to prove the following two theorems: \[second\] If ZF is consistent then the following theory is consistent: > ZF + DC($\omega_{1}$) + “Every set of reals has Baire property” \[first\] If ZF is consistent then the following theory is consistent: > ZFC + “Every projective set of reals has Baire property” + “Any union of $\omega_{1}$ meager sets is meager” Our notation is standard and derived from \[Jec\]. There is one exception, however. We write $p\leq q$ to say that $q$ is a stronger condition then $p$. $\emptyset$ denotes the smallest element of a forcing notion. Basic definitions and facts =========================== In this section we recall some definitions and results from \[She\]. They will be applied in the next section. The basic tool in the construction of models in which definable sets have Baire property is the amalgamation. To define this operation we need the following definition. Recall that ${{\bf P}}{\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}{{\bf P}}'$ means ${{\bf P}}\subseteq{{\bf P}}'$ and each maximal antichain in ${{\bf P}}$ is a maximal antichain in ${{\bf P}}'$. For a forcing notion ${{\bf P}}$ let $\Gamma_{{{\bf P}}}$ be a ${{\bf P}}$-name for the generic subset of ${{\bf P}}$. Suppose that ${{\bf P}}{\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}\mbox{BA}({{\bf Q}})$. Then $({{\bf Q}}\!:\!{{\bf P}})$ is the ${{\bf P}}$-name of a forcing notion which is a subset of ${{\bf Q}}$, $$({{\bf Q}}\!:\!{{\bf P}})=\{q\in{{\bf Q}}: q\mbox{ is compatible with every } p\in\Gamma_{{{\bf P}}} \}.$$ Thus $p{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}q\in ({{\bf Q}}\!:\!{{\bf P}})$ if and only if every $p'\in{{\bf P}}$, $p'\geq p$ is compatible with $q$. Recall that if ${{\bf P}}{\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}\mbox{BA}({{\bf Q}})$ then forcing notions ${{\bf Q}}$ and ${{\bf P}}*({{\bf Q}}\!:\!{{\bf P}})$ are equivalent. Let ${{\bf P}}^{0}, {{\bf P}}^{1}$ and ${{\bf P}}^{2}$ be forcing notions. Suppose that $f_{1}:{{\bf P}}^{0}\stackrel{1-1}{\longrightarrow}\mbox{BA}({{\bf P}}^{1})$, $f_{2}:{{\bf P}}^{0}\stackrel{1-1}{\longrightarrow}\mbox{BA}({{\bf P}}^{2})$ are complete embeddings (i.e. they preserve order and $f_{i}[{{\bf P}}^{0}]{\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}\mbox{BA}({{\bf P}}^{i})\:$). We define [*the amalgamation of ${{\bf P}}^{1}$ and ${{\bf P}}^{2}$ over $f_{1}$, $f_{2}$* ]{} by ${{\bf P}}^{1}\times_{f_{1},f_{2}}{{\bf P}}^{2}=$ $$\{(p_{1},p_{2})\!\in\!{{\bf P}}^{1}\!\times\!{{\bf P}}^{2}: (\exists p\!\in\!{{\bf P}}^{0}) (p{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}\!\!\mbox{``}p_{1}\!\in\!({{\bf P}}^{1}\!: \!f_{1}[{{\bf P}}^{0}]) \:\&\: p_{2}\!\in\!({{\bf P}}^{2}\!:\!f_{2}[{{\bf P}}^{0}]) \mbox{''})\}$$ ${{\bf P}}^{1}\times_{f_{1},f_{2}}{{\bf P}}^{2}$ is ordered in the natural way: $(p_{1},p_{2})\leq(p_{1}',p_{2}')$ if and only if $p_{1}\leq p_{1}'$, $p_{2}\leq p_{2}'$. Note that ${{\bf P}}^{1},{{\bf P}}^{2}$ can be completely embedded into the amalgamation ${{\bf P}}^{1}\times_{f_{1},f_{2}}{{\bf P}}^{2}$ by $p_{1}\in{{\bf P}}^{1}\mapsto (p_{1},\emptyset)$ and $p_{2}\in{{\bf P}}^{2}\mapsto (\emptyset,p_{2})$. Thus we think of ${{\bf P}}^{1}\times_{f_{1},f_{2}} {{\bf P}}^{2}$ as an forcing notion extending both ${{\bf P}}^{1}$ and ${{\bf P}}^{2}$. The amalgamation is applied in constructing of Boolean algebras admitting a lot of automorphisms. The mapping $$f_{2}^{-1}{\circ}f_{1}^{-1}:f_{1}[{{\bf P}}^{0}]\longrightarrow{{\bf P}}^{2}$$ can be naturally extended to an embedding $$\phi:{{\bf P}}^{1}\longrightarrow{{\bf P}}^{1}\times_{f_{1},f_{2}}{{\bf P}}^{2}.$$ Now. suppose that ${{\cal B}}$ is a complete Boolean algebra such that for sufficiently many pairs $({{\bf P}}^{1},{{\bf P}}^{2})$ of complete suborders of ${{\cal B}}$ and for complete embeddings $f_{i}:{{\bf P}}^{0}\longrightarrow{{\bf P}}^{i}$, ($i=1,2$) the algebra ${{\cal B}}$ contains the amalgamation ${{\bf P}}^{1}\times_{f_{1},f_{2}}{{\bf P}}^{2}$. Then ${{\cal B}}$ is [*strongly Cohen-homogeneous*]{}: Suppose $\tau$ is a ${{\cal B}}$-name for an $\omega_{1}$-sequence of ordinals. Then there exists a complete subalgebra ${{\cal B}}'$ of the algebra ${{\cal B}}$ such that - $\tau$ is a ${{\cal B}}'$-name, - if ${{\cal B}}'{\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}{{\cal B}}''{\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}{{\cal B}}$, ${{\cal B}}'{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}$“$({{\cal B}}'':{{\cal B}}')$ is the Cohen algebra” and $f:{{\cal B}}''\longrightarrow{{\cal B}}$ is a complete embedding such that $f{{\mathord{{|}}}}{{\cal B}}'=\mbox{id}_{{{\cal B}}'}$\ then there exists an automorphism $\phi:{{\cal B}}\stackrel{\mbox{onto}}{\longrightarrow} {{\cal B}}$ extending $f$. For more details on extending homomorphisms see \[JuR\]. Solovay showed the connection between the strong homogeneity of the algebra ${{\cal B}}$ and the fact that in generic extensions via ${{\cal B}}$ all projective sets of reals have Baire property. Let ${\bf S}_{1}$ be the class of all $\omega_{1}$-sequences of ordinal numbers. \[Solovay\]  Let ${{\cal B}}$ be a strongly Cohen homogeneous complete Boolean algebra satisfying ccc. Suppose that for any ${{\cal B}}$-name $\tau$ for an $\omega_{1}$-sequence of ordinals ${{\cal B}}{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}$“the union of all meager Borel sets coded in ${{\bf V}}[\tau]$ is meager”.\ Then ${{\cal B}}{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}$“any set of reals definable over ${\bf S}_{1}$ has Baire property”. [[Proof]{} ]{}See theorem 2.3 of \[JuR\]. [width 6pt height 6pt depth 0pt ]{} The class ${\bf HOD}({\bf S}_{1})$ consists of all sets hereditarily ordinal definable over ${\bf S}_{1}$. \[Solovaybis\] Assume that every set of reals ordinal definable over ${\bf S}_{1}$ has Baire property. Then\ ${\bf HOD}({\bf S}_{1})\!\models$“ [**ZF**]{} + ${\bf DC}(\omega_{1})$ + every set of reals has Baire property”. [[Proof]{} ]{}See \[Sol\]. [width 6pt height 6pt depth 0pt ]{} In the next section we will built a model in which there exists an algebra ${{\cal B}}$ satisfying the assumptions of theorem \[Solovay\] and such that > ${{\cal B}}{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}$“the union of $\omega_{1}$ meager sets is meager”. To be sure that the algebra ${{\cal B}}$ satisfies ccc we will use the following notion. A triple $({{\bf P}},{{\cal D}}, \{E_{n}\}_{n\in\omega})$ is [*a model of sweetness* ]{} if 1. ${{\bf P}}$ is a notion of forcing and ${{\cal D}}$ is a dense subset of ${{\bf P}}$, 2. $E_{n}$ are equivalence relations on ${{\cal D}}$ such that - each $E_{n}$ has countably many equivalence classes (the equivalence class of the element $p\in{{\cal D}}$ in the relation $E_{n}$ will be denoted by $[p]_{n}$), - equivalence classes of all relations $E_{n}$ are upward directed, - if $\{p_{i}:i\leq\omega\}\subseteq{{\cal D}}$, $p_{i}\in [p_{\omega}]_{i}$ for all $i$ then for every $n<\omega$ there exists $q\in [p_{\omega}]_{n}$ which is stronger than all $p_{i}$ for $i\geq n$, - if $p,q\in{{\cal D}}$, $p\leq q$ and $n\in\omega$ then there exists $k\in\omega$ such that $$(\forall p'\in [p]_{k})(\exists q'\in [q]_{n})(p'\leq q').$$ Note that if $({{\bf P}},{{\cal D}},\{E_{n}\}_{n\in\omega})$ is a model of sweetness then ${{\bf P}}$ is $\sigma$-centered. We say that a model of sweetness $({{\bf P}}^{2},{{\cal D}}^{2},\{E_{n}^{2}\}_{n\in\omega})$ [*extends* ]{} a model $({{\bf P}}^{1},{{\cal D}}^{1},\{E_{n}^{1}\}_{n\in\omega})$ (we write\ $({{\bf P}}^{1},{{\cal D}}^{1},\{E_{n}^{1}\}_{n\in\omega})\: < \:({{\bf P}}^{2},{{\cal D}}^{2},\{E_{n}^{2}\}_{n\in\omega})$) whenever 1. ${{\bf P}}^{1}{\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}{{\bf P}}^{2}$, ${{\cal D}}^{1}\subseteq{{\cal D}}^{2}$ and $E_{n}^{1}=E_{n}^{2}{{\mathord{{|}}}}{{\cal D}}^{1}$ for each $n\in\omega$, 2. if $p\in{{\cal D}}^{1}$, $n\in\omega$ then $[p]_{n}^{2}\subseteq{{\cal D}}^{1}$, 3. if $p\leq q$, $p\in{{\cal D}}^{2},q\in{{\cal D}}^{1}$ then $p\in{{\cal D}}^{1}$. \[order\] [**a)**]{} : The relation $<$ is transitive on models of sweetness. [**b)**]{} : Suppose that $({{\bf P}}^{i},{{\cal D}}^{i},\{E_{n}^{i}\}_{n\in\omega})$ are models of sweetness such that $$({{\bf P}}^{i},{{\cal D}}^{i},\{E_{n}^{i}\}_{n\in\omega})\:<\: ({{\bf P}}^{j},{{\cal D}}^{j},\{E_{n}^{j}\}_{n\in\omega})$$ for all $i<j<\xi$ ($\xi<\omega_{1}$). Then $$\lim_{i<\xi}({{\bf P}}^{i},{{\cal D}}^{i},\{E^{i}_{n}\}_{n\in\omega}) = (\bigcup_{i<\xi}{{\bf P}}^{i},\bigcup_{i<\xi}{{\cal D}}^{i}, \{\bigcup_{i<\xi}E_{n}^{i}\}_{n\in\omega})$$ is a model of sweetness extending all models $({{\bf P}}^{i},{{\cal D}}^{i},\{E_{n}^{i}\}_{n\in\omega})$.[width 6pt height 6pt depth 0pt ]{} The sweetness may be preserved by the amalgamation. \[amalgamation\] Suppose that $({{\bf P}}^{i},{{\cal D}}^{i},\{E_{n}^{i}\}_{n\in\omega})$ for $i=1,2$ are models of sweetness and $f_{i}:{{\bf P}}^{0}\longrightarrow\mbox{BA}({{\bf P}}^{i})$ are complete embeddings. Then there exists a model of sweetness $({{\bf P}}^{1}\times_{f_{1},f_{2}}{{\bf P}}^{2},{{\cal D}}^{*},\{E_{n}^{*}\}_{n\in\omega})$ based on the amalgamation ${{\bf P}}^{1}\times_{f_{1},f_{2}}{{\bf P}}^{2}$ and extending both $({{\bf P}}^{1},{{\cal D}}^{1},\{E_{n}^{1}\}_{n\in\omega})$ and $({{\bf P}}^{2},{{\cal D}}^{2},\{E_{n}^{2}\}_{n\in\omega})$. [[Proof]{} ]{}see lemmas 7.5, 7.12 of \[She\]. [width 6pt height 6pt depth 0pt ]{} To ensure that our algebra satisfies > ${{\cal B}}{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}$“the union of $\omega_{1}$ meager sets is meager” we will use the Hechler order ${{\bf D}}$. Recall that ${{\bf D}}$ consists of all pairs $(n,f)$ such that $n\in\omega$, $f\in{\omega^{\textstyle \omega}}$. It is ordered by > $(n,f)\leq(n',f')$ if and only if > > $n\leq n'$, $f{{\mathord{{|}}}}n = f'{{\mathord{{|}}}}n$ and $(\forall k\!\in\!\omega) > (f(k)\leq f'(k))$. The forcing with ${{\bf D}}$ adds both a dominating real and a Cohen real. Consequently > ${{\bf D}}*\dot{{{\bf D}}}{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}$“the union of all Borel meager sets coded in the ground model is meager”. The iteration with ${{\bf D}}$ preserves sweetness. \[Hechler\] Let $({{\bf P}},{{\cal D}},\{E_{n}\}_{n\in\omega})$ be a model of sweetness and let $\dot{{{\bf D}}}$ be a ${{\bf P}}$-name for the Hechler forcing. Then there exists a model of sweetness $({{\bf P}}*\dot{{{\bf D}}},{{\cal D}}^{*},\{E_{n}^{*}\}_{n\in\omega})$ based on ${{\bf P}}*\dot{{{\bf D}}}$ and extending the model $({{\bf P}},{{\cal D}},\{E_{n}\}_{n\in\omega})$. [[Proof]{} ]{}Similar to the proof of lemmas 7.6, 7.11 of \[She\]. [width 6pt height 6pt depth 0pt ]{} The proof of the main result ============================ In this section we present proofs of theorems \[first\] and \[second\]. \[K\] Let ${{\cal K}}$ be the class consisting of all sequences\ $\bar{{{\bf P}}} =\ <(P^{i},M^{i}):\:i<\omega_{1}>$ such that 1. $M^{i}$ is a model of sweetness based on $P^{i}$, 2. if $i<j<\omega_{1}$ then $P^{i}{\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}P^{j}$. If $\:\bar{{{\bf P}}}\in{{\cal K}}$ is as above then we put $P^{\omega_{1}}=\bigcup_{i<\omega_{1}} P^{i}$. Note that if $\bar{{{\bf P}}}\in{{\cal K}}$ then each $P^{i}$ is $\sigma$-centered. Consequently $P^{\omega_{1}}$ satisfies ccc. We define the relation $\leq$ on ${{\cal K}}$. \[leq\] Let $\bar{{{\bf P}}}_{1},\bar{{{\bf P}}}_{2}\in{{\cal K}}$. We say $\bar{{{\bf P}}}_{1} \leq \bar{{{\bf P}}}_{2}$ if $P^{\omega_{1}}_{1}{\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}P^{\omega_{1}}_{2}$ and there exists a closed unbounded subset $C$ of $\omega_{1}$ such that [**(!)** ]{} : if $i\in C$ then $M^{i}_{1} < M^{i}_{2}$ [**(!!)**]{} : if $i\in C$, $q\in P^{\omega_{1}}_{1}$, $p\in P^{i}_{1}$ and $p{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}_{P^{i}_{1}}q\in(P^{\omega_{1}}_{1}\!:\! P^{i}_{1})$\ then $p{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}_{P^{i}_{2}}q\in(P^{\omega_{1}}_{2}\!:\! P^{i}_{2})$. Clearly the relation $\leq$ is transitive and reflexive. \[countable\] Suppose that $\bar{{{\bf P}}}_{m}\in{{\cal K}}$ for $m<\omega$ are such that $m_{1}<m_{2}<\omega$ implies $\bar{{{\bf P}}}_{m_{1}}\leq\bar{{{\bf P}}}_{m_{2}}$ (and let $C_{m_{1},m_{2}}$ witness it). Let $C=\bigcap_{m_{1}<m_{2}<\omega}C_{m_{1},m_{2}}$. Put $$P^{i}_{\omega}=\bigcup_{m<\omega}P^{\cap(C\backslash i)}_{m}, \ M^{i}_{\omega}=\lim_{m<\omega}M^{\cap(C\backslash i)}_{m}.$$ Then $\bar{{{\bf P}}}_{\omega} = \ <(P^{i}_{\omega},M^{i}_{\omega}):\: i<\omega_{1}\!>\ \in {{\cal K}}$ and $\bar{{{\bf P}}}_{m}\leq\bar{{{\bf P}}}_{\omega}$ for each $m<\omega$. [[Proof]{} ]{}First note that $C$ is a closed unbounded subset of $\omega_{1}$. Since $C\subseteq\bigcap_{m<\omega}C_{m,m+1}$ we may apply lemma \[order\] b) to conclude that each $M^{i}_{\omega}$ is a model of sweetness based on $P^{i}_{\omega}$. [Claim]{}:If $i<j<\omega_{1}$ then $P^{i}_{\omega}{\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}P^{j}_{\omega}$. Indeed, let $i<j$. We may assume that $i,j\in C$ (recall that $P^{i}_{\omega} = P^{\cap(C\backslash i)}_{\omega}$). Note that $P^{i}_{m}{\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}P^{i}_{\omega}$ and $P^{i}_{m}{\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}P^{j}_{m}$ for each $m\in\omega$. Let ${{\cal A}}\subseteq P^{i}_{\omega}$ be a maximal antichain. Clearly it is an antichain in $P^{j}_{\omega}$ but we have to prove that it is maximal. Let $q\in P^{j}_{\omega}$. Then $q\in P^{j}_{m}$ for some $m<\omega$. Let $$Z=\{r\in P^{i}_{m}: (\exists p_{r}\in{{\cal A}})(r{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}_{P^{i}_{m}}p_{r} \in (P^{i}_{\omega}:P^{i}_{m}))\}$$ Clearly $Z$ is dense in $P^{i}_{m}$. Hence we find $r\in Z$ such that $r{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}_{P^{i}_{m}} q\in (P^{j}_{m}\!:\!P^{i}_{m})$. Let $p_{r}\in{{\cal A}}$ witness $r\in Z$. Take $k$ such that $p_{r}\in P^{i}_{k}$, $m<k<\omega$. Consider $\bar{{{\bf P}}}_{m}$ and $\bar{{{\bf P}}}_{k}$. Since $i,j\in C\subseteq C_{m,k}$ we may apply condition [**(!!)**]{} to conclude that $$r{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}_{P^{i}_{k}} q\in(P^{j}_{k}:P^{i}_{k}).$$ By the choice of $p_{r}$ we have $$r{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}_{P^{i}_{m}} p_{r}\in(P^{i}_{k}:P^{i}_{m}).$$ Thus $p_{r}$ and $r$ are compatible and any $p'\in P^{i}_{k}$, $p'\geq r,p_{r}$ is compatible with $q$. Consequently $q$ and $p_{r}$ are compatible. The claim is proved. It follows from the above claim that $\bar{{{\bf P}}}_{\omega}\in{{\cal K}}$. [Claim]{}: The club $C$ witness that $\bar{{{\bf P}}}_{m}\leq\bar{{{\bf P}}}_{\omega}$ for each $m<\omega$. Indeed, first note that $$P^{\omega_{1}}_{\omega}=\bigcup_{i<\omega_{1}}P^{i}_{\omega}= \bigcup_{i<\omega_{1}}\bigcup_{m<\omega}P^{\cap(C\backslash i)}_{m}= \bigcup_{m<\omega}P^{\omega_{1}}_{m}.$$ Since $P^{\omega_{1}}_{m_{1}}{\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}P^{\omega_{1}}_{m_{2}}$ for each $m_{1}<m_{2}$ we see that $P^{\omega_{1}}_{m}{\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}P^{\omega_{1}}_{\omega}$. It follows from the definition of $M^{i}_{\omega}$ and lemma \[order\] that if $i\in C$ then $M^{i}_{m}<M^{i}_{\omega}$. Thus we have to check condition [**(!!)**]{} only. Suppose $i\in C$, $q\in P^{\omega_{1}}_{m}$, $p\in P^{i}_{m}$ and $p{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}_{P^{i}_{m}}q\in (P^{\omega_{1}}_{m}\!:\!P^{i}_{m})$. Assume $p\not{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}_{P^{i}_{\omega}}q\in (P^{\omega_{1}}_{\omega}\! : \!P^{i}_{\omega})$. Then we find $r\in P^{i}_{\omega}$ such that $r\geq p$ and $r$ is incompatible with $q$. Let $k>m$ be such that $r\in P^{i}_{k}$. Since $i\in C_{m.k}$ we have $p{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}_{P^{i}_{k}}q\in (P^{\omega_{1}}_{k}\!:\!P^{i}_{k})$ (by condition [**(!!)**]{} for $\bar{{{\bf P}}}_{m},\bar{{{\bf P}}}_{k}$). But $r{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}_{P^{i}_{k}}q\not\in (P^{\omega_{1}}_{k}\!:\!P^{i}_{k})$ - a contradiction. [width 6pt height 6pt depth 0pt ]{} \[omegaone\] Assume that - $\bar{{{\bf P}}}_{\xi}\in{{\cal K}}$ for $\xi<\omega_{1}$, - if $\xi<\zeta<\omega_{1}$ then $\bar{{{\bf P}}}_{\xi}\leq\bar{{{\bf P}}}_{\zeta}$ is witnessed by the club $C_{\xi,\zeta}\subseteq \omega_{1}$, - if $\delta< \omega_{1}$ is a limit ordinal and $i\in\bigcap_{\xi<\zeta<\delta}C_{\xi,\zeta}$ then $M^{i}_{\delta}=\lim_{\xi<\delta} M^{i}_{\xi}$. Let $$C=\{\delta<\omega_{1}:\delta\mbox{ is limit }\&\ (\forall\xi\!<\!\zeta\!<\!\delta)(\delta\in C_{\xi,\zeta})\}$$ and let $C(i)=\cap(C\backslash i)$ for $i<\omega_{1}$. Put $P^{i}_{\omega_{1}}=P^{C(i)}_{C(i)}$, $M^{i}_{\omega_{1}}=M^{C(i)}_{C(i)}$.\ Then $\bar{{{\bf P}}}_{\omega_{1}} \in{{\cal K}}$ and $(\forall\xi<\omega_{1})(\bar{{{\bf P}}}_{\xi}\leq\bar{{{\bf P}}}_{\omega_{1}})$. [[Proof]{} ]{}First note that the set $\{\delta\!<\!\omega_{1}:(\forall\xi\!<\!\zeta\!<\!\delta)(\delta \!\in\! C_{\xi,\zeta})\}$ is the diagonal intersection of clubs $\bigcap_{\xi<\zeta} C_{\xi,\zeta}$ (for $\zeta<\omega_{1}$). Hence $C$ is closed and unbounded and $\bar{{{\bf P}}}_{\omega_{1}}$ is well defined. [Claim]{}: If $i<j<\omega_{1}$ then $P^{i}_{\omega_{1}} {\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}P^{j}_{\omega_{1}}$. Indeed, suppose $i<j<\omega_{1}$. Then $P^{i}_{\omega_{1}} = P^{C(i)}_{C(i)}$, $P^{j}_{\omega_{1}} = P^{C(j)}_{C(j)}$ and we may assume that $C(i)<C(j)$. By \[K\] 2) we have that $P^{C(i)}_{C(i)} {\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}P^{C(j)}_{C(i)}$. Since $C$ consists of limit ordinals only and $C(j)\in\bigcap_{\xi<\zeta<C(j)}C_{\xi,\zeta}$ we get $P^{C(j)}_{C(j)} = \bigcup_{\xi<C(j)} P^{C(j)}_{\xi}$ (and it is a direct limit). Since $C(i)<C(j)$ we conclude $P^{C(j)}_{C(i)} {\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}P^{C(j)}_{C(j)}$ and consequently $P^{C(i)}_{C(i)}{\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}P^{C(j)}_{C(j)}$. The claim is proved. Since each $M^{i}_{\omega_{1}}$ is a model of sweetness based on $P^{i}_{\omega_{1}}$ we have proved that $\bar{{{\bf P}}}_{\omega_{1}}\in{{\cal K}}$. Let $\xi<\omega_{1}$. [Claim]{}: $P^{\omega_{1}}_{\xi}{\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}P^{\omega_{1}}_{\omega_{1}}$ First note that $$P^{\omega_{1}}_{\omega_{1}}=\bigcup_{i<\omega_{1}} P^{i}_{\omega_{1}}= \bigcup_{i<\omega_{1}} P^{C(i)}_{C(i)} = \bigcup_{\zeta,i<\omega_{1}} P^{i}_{\zeta} = \bigcup_{\zeta<\omega_{1}} P^{\omega_{1}}_{\zeta}.$$ Since $\zeta_{1}<\zeta_{2}<\omega_{1}$ implies $\bar{{{\bf P}}}_{\zeta_{1}} \leq \bar{{{\bf P}}}_{\zeta_{2}}$ we have $P^{\omega_{1}}_{\zeta_{1}} {\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}P^{\omega_{1}}_{\zeta_{2}}$ for $\zeta_{1}<\zeta_{2}<\omega_{1}$. Consequently $P^{\omega_{1}}_{\xi}{\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}P^{\omega_{1}}_{\omega_{1}}$. [Claim]{}: If $i\in C\backslash(\xi+1)$ then $M^{i}_{\xi}<M^{i}_{\omega_{1}}$. If $i\in C\backslash(\xi+1)$ then $C(i)=i>\xi$. Moreover it follows from our assumptions that $M^{i}_{i}=\lim_{\zeta<i}M^{i}_{\zeta}$. By lemma \[order\] we get $M^{i}_{\xi}<M^{i}_{i}=M^{C(i)}_{C(i)}=M^{i}_{\omega_{1}}$. [Claim]{}:Suppose $i\in C\backslash(\xi+1)$, $q\in P^{\omega_{1}}_{\xi}$, $p\in P^{i}_{\xi}$ and $p{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}_{P^{i}_{\xi}} q\in (P^{\omega_{1}}_{\xi}\!:\!P^{i}_{\xi})$. Then $p{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}_{P^{i}_{\omega_{1}}} q\in (P^{\omega_{1}}_{\omega_{1}}\!:\!P^{i}_{\omega_{1}})$. Assume not. Then we have $r\in P^{i}_{\omega_{1}}=P^{i}_{i}$, $r\geq p$ such that $r$ and $q$ are incompatible. There is $\zeta\in (\xi,i)$ such that $r\in P^{i}_{\zeta}$. Thus $p\not{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}_{P^{i}_{\zeta}} q\in (P^{\omega_{1}}_{\zeta}\!:\!P^{i}_{\zeta})$. Since $i\in C_{\xi,\zeta}$ we get a contradiction with condition [**(!!)**]{} for $\bar{{{\bf P}}}_{\xi} \leq \bar{{{\bf P}}}_{\zeta}$. We have proved that the club $C\backslash(\xi+1)$ witness $\bar{{{\bf P}}}_{\xi} \leq \bar{{{\bf P}}}_{\omega_{1}}$.[width 6pt height 6pt depth 0pt ]{} Suppose $\bar{{{\bf P}}}=\ <(P^{i},M^{i}):i<\omega_{1}>\ \in {{\cal K}}$. Let $$P^{i}_{D}=\{(p,\tau)\in P^{\omega_{1}}*\dot{{{\bf D}}}: p\in P^{i} \ \& \ \tau\mbox{ is a $P^{i}$-name }\}.$$ Note that $P^{i}_{D}$ is isomorphic to $P^{i}*\dot{{{\bf D}}}$. Let $M^{i}_{D}$ be the canonical model of sweetness based on $P^{i}_{D}$ and extending the model $M^{i}$ (see lemma \[Hechler\]). Let $$\bar{{{\bf P}}}_{D}= \ <(P^{i}_{D},M^{i}_{D}):i<\omega_{1}>.$$ \[Hechler2\] $\bar{{{\bf P}}}_{D}\in{{\cal K}}$, $\bar{{{\bf P}}}\leq\bar{{{\bf P}}}_{D}$ and $P^{\omega_{1}}_{D}=P^{\omega_{1}}*\dot{{{\bf D}}}$. [[Proof]{} ]{}The last assertion is a consequence of the fact that $P^{\omega_{1}}$ is a ccc notion of forcing. It follows from properties of Souslin forcing (cf \[JS2\]) that $P^{i}_{D}{\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}P^{j}_{D}$ provided $i<j$. Consequently $\bar{{{\bf P}}}_{D}\in{{\cal K}}$. To show $\bar{{{\bf P}}}\leq\bar{{{\bf P}}}_{D}$ note that $M^{i}<M^{i}_{D}$ for all $i<\omega_{1}$ and $P^{\omega_{1}}{\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}P^{\omega_{1}}_{D}$. Suppose now that $i<\omega_{1}$, $p\in P^{i}$, $q\in P^{\omega_{1}}$ and $p{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}_{P^{i}}q\in(P^{\omega_{1}}\!:\!P^{i})$. Assume that $p\not{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}_{P^{i}_{D}}q\in(P^{\omega_{1}}_{D}\!:\!P^{i}_{D})$. Then we find a condition $r=(r_{0},\tau)\in P^{i}_{D}$ above $p$ which is inconsistent with $q$. Note that $q$ may be a member of $P^{i}*\dot{{{\bf D}}}{\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}P^{j}$ (for some $j>i$) but we consider it as an element of $P^{\omega_{1}}$, while $r$ is an element of $P^{\omega_{1}}*\dot{{{\bf D}}}$. Consequently incompatibility of $q$ and $r$ means that $q$ and $r_{0}$ are not compatible. But $r_{0}\in P^{i}$ lies above $p$ - a contradiction. [width 6pt height 6pt depth 0pt ]{} \[BCD\] Suppose that ${{\cal B}},{{\cal C}},{{\cal D}},{{\cal C}}_{0}$ are complete Boolean algebras such that (1) : ${{\cal B}}{\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}{{\cal D}}{\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}{{\cal C}}$, ${{\cal C}}_{0}{\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}{{\cal C}}$ Let ${{\cal B}}_{0}={{\cal B}}\cap{{\cal C}}_{0}$, ${{\cal D}}_{0}={{\cal D}}\cap{{\cal C}}_{0}$ (note that ${{\cal B}}_{0}{\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}{{\cal D}}_{0}{\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}{{\cal C}}_{0}$). We assume that (2) : ${{\cal B}}{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}$“$({{\cal D}}\!:\!{{\cal B}})$ is a subset of $({{\cal C}}_{0}\!:\!{{\cal B}})$” (3) : if $b\in{{\cal B}}, b_{0}\in{{\cal B}}_{0}$ and $b_{0}{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}_{{{\cal B}}_{0}}b\in({{\cal B}}\!:\!{{\cal B}}_{0})$ then $b_{0}{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}_{{{\cal C}}_{0}}b\in({{\cal C}}\!:\!{{\cal C}}_{0})$. Then ($3^{*}$) : if $d\in{{\cal D}}, d_{0}\in{{\cal D}}_{0}$ and $d_{0}{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}_{{{\cal D}}_{0}}d\in({{\cal D}}\!:\!{{\cal D}}_{0})$ then $d_{0}{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}_{{{\cal C}}_{0}}d\in({{\cal C}}\!:\!{{\cal C}}_{0})$. [[Proof]{} ]{} [Claim]{}:Suppose $c\in{{\cal C}}_{0}, d_{0}\in{{\cal D}}_{0}$ and $d_{0}{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}_{{{\cal D}}_{0}}c\in({{\cal C}}_{0}\!:\!{{\cal D}}_{0})$. Then $d_{0}{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}_{{{\cal D}}}c\in({{\cal C}}\!:\!{{\cal D}})$. We have to prove that each $d\geq d_{0}, d\in{{\cal D}}$ is compatible with $c$. Let $d\geq d_{0}, d\in{{\cal D}}$. By (2) we find $b\in{{\cal B}}$ and $d_{1}\in{{\cal D}}_{0}$ such that $$b{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}_{{{\cal B}}}\mbox{``}d\in({{\cal D}}\!:\!{{\cal B}}) \ \& \ d\equiv_{({{\cal D}}:{{\cal B}})} d_{1} \mbox{''}$$ (the last means that $b{\!\cdot\!}d=b{\!\cdot\!}d_{1}$). Thus $b{\!\cdot\!}d_{1}{\!\cdot\!}d_{0} = b{\!\cdot\!}d{\!\cdot\!}d_{0} = b{\!\cdot\!}d_{0} \neq {\bf 0}$. We find $b_{0}\in {{\cal B}}_{0}$ such that $b_{0}{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}_{{{\cal B}}_{0}}b\in({{\cal B}}\!:\!{{\cal B}}_{0})$ and $b_{0}{\!\cdot\!}d_{1}{\!\cdot\!}d_{0}\neq {\bf 0}$ (it is enough to take $b_{0}$ such that $b_{0}{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}_{{{\cal B}}_{0}}b{\!\cdot\!}d_{1}{\!\cdot\!}d_{0}\in({{\cal D}}\!:\!{{\cal B}}_{0})$). Note that then $b_{0}{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}_{{{\cal C}}_{0}}b\in ({{\cal C}}\!:\!{{\cal C}}_{0})$ (by (3)). Since $b_{0}{\!\cdot\!}d_{1}{\!\cdot\!}d_{0} \in {{\cal D}}_{0}$ and it is stronger than $d_{0}$ we get $b_{0}{\!\cdot\!}d_{1}{\!\cdot\!}d_{0}{\!\cdot\!}c\neq {\bf 0}$. The last condition is stronger than $b_{0}$ and belongs to ${{\cal C}}_{0}$. Hence $b{\!\cdot\!}b_{0}{\!\cdot\!}d_{1}{\!\cdot\!}d_{0}{\!\cdot\!}c\neq {\bf 0}$. Finally note that $b{\!\cdot\!}b_{0}{\!\cdot\!}d_{1}{\!\cdot\!}d_{0}{\!\cdot\!}c\geq b{\!\cdot\!}d_{1} = b{\!\cdot\!}d\geq d$ so $d$ and $c$ are compatible. The claim is proved. Now suppose that $d\in{{\cal D}}, d_{0}\in{{\cal D}}_{0}$ and $d_{0}{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}_{{{\cal D}}_{0}} d\in ({{\cal D}}\!:\!{{\cal D}}_{0})$. Let $c\in{{\cal C}}_{0}, c\geq d_{0}$. Take $d^{*}\in{{\cal D}}_{0}$ such that $d^{*}\geq d_{0}$ and $d^{*}{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}_{{{\cal D}}_{0}} c\in ({{\cal C}}_{0}\!:\!{{\cal D}}_{0})$. By the claim we have $d^{*}{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}_{{{\cal D}}} c\in ({{\cal C}}\!:\!{{\cal D}})$. Since $d^{*}\geq d_{0}$ we have $d^{*}{\!\cdot\!}d\neq {\bf 0}, d^{*}{\!\cdot\!}d\in{{\cal D}}$ and consequently $d^{*}{\!\cdot\!}d{\!\cdot\!}c\neq {\bf 0}$. Hence $d$ and $c$ are compatible and we are done. [width 6pt height 6pt depth 0pt ]{} Suppose that $\bar{{{\bf P}}}_{0},\bar{{{\bf P}}}_{1},\bar{{{\bf P}}}_{2},\bar{{{\bf P}}}_{3} \in {{\cal K}}$ and the club $C\subseteq \omega_{1}$ witness that both $\bar{{{\bf P}}}_{0} \leq \bar{{{\bf P}}}_{1}$ and $\bar{{{\bf P}}}_{2}\leq \bar{{{\bf P}}}_{3}$. Assume that ${{\bf Q}}_{0},{{\bf Q}}_{2}$ are complete Boolean algebras such that for some $i_{0}< \omega_{1}$ - $\mbox{BA}(P^{\omega_{1}}_{0}){\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}{{\bf Q}}_{0}{\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}\mbox{BA}(P^{\omega_{1}}_{1})$, $\mbox{BA}(P^{\omega_{1}}_{2}){\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}{{\bf Q}}_{2} {\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}\mbox{BA}(P^{\omega_{1}}_{3})$ - $\mbox{BA}(P^{\omega_{1}}_{0}){{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}({{\bf Q}}_{0}\!:\! \mbox{BA}(P^{\omega_{1}}_{0}))\subseteq (\mbox{BA}(P^{i_{0}}_{1})\!:\! \mbox{BA}(P^{\omega_{1}}_{0}))$\ $\mbox{BA}(P^{\omega_{1}}_{2}){{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}({{\bf Q}}_{2}\!:\! \mbox{BA}(P^{\omega_{1}}_{2}))\subseteq (\mbox{BA}(P^{i_{0}}_{3})\!:\! \mbox{BA}(P^{\omega_{1}}_{2}))$ Let $f:{{\bf Q}}_{0}\longrightarrow{{\bf Q}}_{2}$ be an isomorphism such that $f[{{\bf Q}}_{0}\cap\mbox{BA}(P^{i}_{1})]={{\bf Q}}_{2}\cap\mbox{BA} (P^{i}_{3})$ for all $i\in C\backslash i_{0}$. For $i\in C\backslash i_{0}$ put $$P^{i}=\{(p_{1},p_{2})\in P^{\omega_{1}}_{1}\times_{\mbox{id},f} P^{\omega_{1}}_{3}: p_{1}\in P^{i}_{1}\ \& \ p_{2}\in P^{i}_{3}\},$$ where id stands for the identity on ${{\bf Q}}_{0}$. It follows from lemma \[BCD\] that $P^{i}$ is isomorphic to $P^{i}_{1} \times_{f_{1},f_{3}} P^{i}_{3}$, where $f_{3}=f{{\mathord{{|}}}}{{\bf Q}}_{0}\cap \mbox{BA}(P^{i}_{1})$ and $f_{1}$ is the identity on ${{\bf Q}}_{0} \cap \mbox{BA}(P^{i}_{1})$. Therefore we have the canonical model of sweetness $M^{i}$ based on $P^{i}$ and extending both models $M^{i}_{1}$ and $M^{i}_{2}$ (compare lemma \[amalgamation\]). Let $$\bar{{{\bf P}}}_{1}\times_{f}\bar{{{\bf P}}}_{3}=\ <(P^{i},M^{i}): i<\omega_{1}>.$$ Note that $\bigcup_{i<\omega_{1}}P^{i} = P^{\omega_{1}}_{1}\times_{\mbox{id},f} P^{\omega_{1}}_{3}$. \[amalgamation2\] $\bar{{{\bf P}}}_{1}\times_{f}\bar{{{\bf P}}}_{3}\in{{\cal K}}$ and $\bar{{{\bf P}}}_{1},\bar{{{\bf P}}}_{3}\leq \bar{{{\bf P}}}_{1}\times_{f}\bar{{{\bf P}}}_{3}$. [[Proof]{} ]{}To prove $\bar{{{\bf P}}}_{1}\times_{f}\bar{{{\bf P}}}_{3}\in{{\cal K}}$ we have to show the following [Claim]{}: $P^{i}{\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}P^{j}$ for each $i<j<\omega_{1}$, $i,j\in C\backslash i_{0}$. Let ${{\cal A}}\subseteq P^{i}$ be a maximal antichain and let $(p_{1},p_{2})\ \in P^{j}$. Let $q\in {{\bf Q}}_{0}$ be such that $$q{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}\mbox{``}p_{1}\in (P^{\omega_{1}}_{1}\!:\!{{\bf Q}}_{0})\ \& \ p_{2}\in (P^{\omega_{1}}_{3}\!:\!f[{{\bf Q}}_{0}])\mbox{''}.$$ Take $r_{1}\in P^{i}_{1}$ such that $r_{1}{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}_{P^{i}_{1}}\mbox{``} p_{1},q\in (P^{\omega_{1}}_{1}\!:\!P^{i}_{1})\mbox{''}$ (note that $q$ and $p_{1}$ are compatible). Next find $q'\in{{\bf Q}}_{0}$ such that $q'\geq q$ and $q'{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}r_{1}\in (P^{\omega_{1}}_{1}\!:\!{{\bf Q}}_{0})$ (recall that $r_{1}$ and $q$ are compatible). Since $p_{2}$ and $f(q')$ are compatible we find $r_{2}\in P^{i}_{3}$ such that $r_{2}{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}_{P^{i}_{3}}\mbox{``} p_{2},f(q')\in (P^{\omega_{1}}_{3}\!:\!P^{i}_{3})\mbox{''}$. Consider the pair $(r_{1},r_{2})$. There is $q''\in{{\bf Q}}_{0}, q''\geq q'$ such that $q''{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}r_{2}\in (P^{\omega_{1}}_{3}\!:\!f[{{\bf Q}}_{0}])$. Then $$q''{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}\mbox{``}r_{1}\in (P^{\omega_{1}}_{1}\!:\!{{\bf Q}}_{0})\ \& \ r_{2}\in (P^{\omega_{1}}_{3}\!:\!f[{{\bf Q}}_{0}])\mbox{''}$$ and consequently $(r_{1},r_{2})\in P^{i}$. Since $(r_{1},r_{2})$ has to be compatible with some element of ${{\cal A}}$ we are done. [Claim]{}:Suppose $q\in P^{\omega_{1}}_{1}$, $i\in C\backslash i_{0}$, $p\in P^{i}_{1}$ are such that $p{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}_{P^{i}_{1}} q\in(P^{\omega_{1}}_{1}\!:\!P^{i}_{1})$. Then $p{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}_{P^{i}} q\in (P^{\omega_{1}}\!:\!P^{i})$. Suppose $r\in P^{i}$ is stronger than $p$. Let $r=(r_{1},r_{2})$ and let $r_{0}\in {{\bf Q}}_{0}$ witness $r\in P^{\omega_{1}}_{1} \times_{\mbox{id},f} P^{\omega_{1}}_{3}$. We may get $r_{0}\in {{\bf Q}}_{0}\cap \mbox{BA}(P^{i}_{1})$. Remember that really we have $p\simeq\ (p,\emptyset)$, $q\simeq\ (q,\emptyset)$. Since $r_{0},r_{1}\in \mbox{BA}(P^{i}_{1})$ are compatible and $r_{1}\geq p$ we find $r_{1}^{*}\in P^{\omega_{1}}_{1}$ above $r_{0}, r_{1}$ and $q$. Then $(r_{1}^{*},r_{2})\in P^{\omega_{1}}$ and it is a condition stronger than both $(r_{1},r_{2})$ and $(q,\emptyset)$. The claim is proved. Since $M^{i}_{1}<M^{i}$ for each $i\in C\backslash i_{0}$ it follows from the above claim that $\bar{{{\bf P}}}_{1}\leq \bar{{{\bf P}}}_{1}\times_{f} \bar{{{\bf P}}}_{3}$ (and $C\backslash i_{0}$ is a witness for it). Similarly one can prove $\bar{{{\bf P}}}_{3}\leq \bar{{{\bf P}}}_{1}\times_{f} \bar{{{\bf P}}}_{3}$. [width 6pt height 6pt depth 0pt ]{} \[automorphism\] Suppose $\bar{{{\bf P}}}_{0},\bar{{{\bf P}}}_{1}\in{{\cal K}}$, $\bar{{{\bf P}}}_{0} \leq \bar{{{\bf P}}}_{1}$. Let ${{\bf Q}}_{0},{{\bf Q}}_{1}$ be complete Boolean algebras such that (for $k=0,1$): - $\mbox{BA}(P^{\omega_{1}}_{0}){\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}{{\bf Q}}_{k}{\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}\mbox{BA}(P^{\omega_{1}}_{1})$ - $\mbox{BA}(P^{\omega_{1}}_{0}){{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}$“$({{\bf Q}}_{k}\!:\! \mbox{BA}(P^{\omega_{1}}_{0}))$ is the Cohen algebra” Let $f:{{\bf Q}}_{0}\longrightarrow{{\bf Q}}_{1}$ be an isomorphism such that $f{{\mathord{{|}}}}\mbox{BA}(P^{\omega_{1}}_{0})=\mbox{id}$.\ Then there exist $\bar{p}\in {{\cal K}}$ and an automorphism $\phi:P^{\omega_{1}} \stackrel{\mbox{onto}}{\longrightarrow}P^{\omega_{1}}$ such that $\bar{{{\bf P}}}_{1} \leq \bar{{{\bf P}}}$ and $f\subseteq\phi$. [[Proof]{} ]{}We may apply lemma \[amalgamation2\] to get that $\bar{{{\bf P}}}_{2} = \bar{{{\bf P}}}_{1}\times_{f}\bar{{{\bf P}}}_{1}\in{{\cal K}}$. The amalgamation over $f$ produces an extension of $f$ — there is $f_{1}: P^{\omega_{1}}_{1}\longrightarrow P^{\omega_{1}}_{2}$ such that $f\subseteq f_{1}$ (we identify $p\in P^{\omega_{1}}_{1}$ with $(\emptyset, p)\in P^{\omega_{1}}_{2}$). Moreover $\bar{{{\bf P}}}_{1}, \bar{{{\bf P}}}_{2}, f_{1}$ satisfy assumptions of lemma \[amalgamation2\] and thus $\bar{{{\bf P}}}_{3}=\bar{{{\bf P}}}_{2}\times_{f_{1}}\bar{{{\bf P}}}_{2}\in {{\cal K}}$. If we identify $p\in P^{\omega_{1}}_{2}$ with $(p,\emptyset) \in \bar{{{\bf P}}}_{3}$ we get a partial isomorphism $f_{2}$ such that $f_{1} \subseteq f_{2}$ and $\mbox{rng}(f_{2})=P^{\omega_{1}}_{2}$. And so on, we build $\bar{{{\bf P}}}_{m}\in{{\cal K}}$ and partial isomorphisms $f_{m}$ such that $\bar{{{\bf P}}}_{m}\leq\bar{{{\bf P}}}_{m+1}$, $f_{m}\subseteq f_{m+1}$ and either $P^{\omega_{1}}_{m}\subseteq \mbox{dom}(f_{m})$ or $P^{\omega_{1}}_{m}\subseteq \mbox{rng}(f_{m})$. Next we apply lemma \[countable\] to conclude that $\bar{{{\bf P}}}_{\omega}\in{{\cal K}}$ and $f_{\omega}=\bigcup_{m\in \omega}f_{m}:P^{\omega_{1}}_{\omega} \stackrel{\mbox{onto}}{\longrightarrow}P^{\omega_{1}}_{\omega}$ is the desired automorphism. [width 6pt height 6pt depth 0pt ]{} We define the following notion of forcing - ${{\bf R}}=\{\bar{{{\bf P}}}\in{{\cal K}}: \bar{p}\in {{\cal H}}(\omega_{2})\}$ - $\leq_{{{\bf R}}}$ is the relation $\leq$ of \[leq\]. A notion of forcing ${{\bf P}}$ is $(\omega_{1}+1)$-strategically closed if the second player has a winning strategy in the following game of the length $\omega_{1}+1$. > For $i=0$ Player I gives $p_{0}\in{{\bf P}}$; > > Player I gives in the $i$-th move a dense subset $D_{i}$ of ${{\bf P}}$; > > Player II gives $p_{i+1}\geq p_{i}$, $p_{i+1}\in D_{i}$, for a limit $i$ Player II gives $p_{i}$ above all $p_{j}$ (for $j<i$). Player II looses if he is not able to give the respective element of ${{\bf P}}$ for some $i\leq \omega_{1}$. Note that $(\omega_{1}+1)$-strategically closed notions of forcings do not add new $\omega_{1}$-sequences of elements of the ground model. \[closed\] The forcing notion ${{\bf R}}$ is $\omega_{1}$-closed and $(\omega_{1}+1)$-strategically closed. Consequently forcing with ${{\bf R}}$ does not collapse $\omega_{1}$ and $\omega_{2}$. [[Proof]{} ]{}For the first assertion use lemma \[countable\]. The second follows from \[countable\] and \[omegaone\]. [width 6pt height 6pt depth 0pt ]{} Note that ${{\mathord{{|}}}}{{\bf R}}{{\mathord{{|}}}}=2^{\omega_{1}}$. Thus if we assume that $2^{\omega_{1}}=\omega_{2}$ then forcing with ${{\bf R}}$ does not collapse cardinals.\ Suppose ${{\bf V}}\models$GCH.\ Let $G\subseteq{{\bf R}}$ be a generic over ${{\bf V}}$. Let ${{\bf P}}=\bigcup\{P^{\omega_{1}}: \bar{{{\bf P}}}\in G\}$. 1. ${{\bf P}}$ is a ccc notion of forcing. 2. If $\tau$ is a ${{\bf P}}$-name for an $\omega_{1}$-sequence of ordinals then\ ${{\bf P}}{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}$“the union of all Borel meager sets coded in ${{\bf V}}[\tau]$ is meager”. 3. The Boolean algebra BA(${{\bf P}}$) is strongly Cohen-homogeneous. 4. ${{\bf P}}{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}$“any union of $\omega_{1}$ meager sets is meager”. [[Proof]{} ]{}1. Work in ${{\bf V}}$. Suppose that $\dot{{{\cal A}}}$ is a ${{\bf R}}$-name for an $\omega_{1}$-sequence of pairwise incompatible elements of ${{\bf P}}$. Let $\bar{{{\bf P}}}\in{{\bf R}}$. By proposition \[closed\] there is $\bar{{{\bf P}}}_{1}\geq\bar{{{\bf P}}}$ which decides all values of $\dot{{{\cal A}}}$. We may assume that all these elements belong to $P^{\omega_{1}}_{1}$. A contradiction. 2\. Let $\tau$ be a ${{\bf P}}$-name for an $\omega_{1}$-sequence of ordinals. Then $\tau$ is actually an $\omega_{1}$-sequence of (countable) antichains in ${{\bf P}}$. Therefore $\tau\in{{\bf V}}$ and it is a $P^{\omega_{1}}_{0}$-name for some $\bar{{{\bf P}}}_{0}\in G$. By density arguments we have that $(\bar{{{\bf P}}}_{D})_{D} \in G$ for some $\bar{{{\bf P}}}\geq\bar{{{\bf P}}}_{0}$ (compare lemma \[Hechler\]). Hence > ${{\bf P}}{{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}$“the union of all Borel meager sets coded in ${{\bf V}}[G][\tau]$ is meager” 3\. Work in ${{\bf V}}[G]$. Let $\tau$ be a ${{\bf P}}$-name for an $\omega_{1}$-sequence of ordinals. As in 2. we find $\bar{{{\bf P}}}_{0}\in G$ such that $\tau$ is a $P^{\omega_{1}}_{0}$-name. Suppose now that - $\mbox{BA}(P^{\omega_{1}}_{0}){\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}{{\cal B}}{\mathrel{\mathord{<}\!\!\raise 0.8 pt\hbox{$\scriptstyle\circ$}}}\mbox{BA}({{\bf P}})$, - $\mbox{BA}(P^{\omega_{1}}_{0}){{\mathrel{{\vrule height 6.9pt depth -0.1pt}\!\vdash}}}\mbox{``}({{\cal B}}\!:\! \mbox{BA}(P^{\omega_{1}}_{0}))$ is the Cohen algebra”, - $f:{{\cal B}}\longrightarrow \mbox{BA}({{\bf P}})$ is a complete embedding such that $f{{\mathord{{|}}}}\mbox{BA}(P^{\omega_{1}}_{0})=\mbox{id}$. Note that ${{\cal B}}$ and $f$ are determined by countably many elements. Each element of $\mbox{BA}({{\bf P}})$ is a countable union of elements of ${{\bf P}}$. Consequently ${{\cal B}}, f\in {{\bf V}}$ and there is $\bar{{{\bf P}}}_{1}\in G$ such that ${{\cal B}}, \mbox{rng}(f)\subseteq \mbox{BA}(P^{\omega_{1}}_{1})$, $\bar{{{\bf P}}}_{0} \leq \bar{{{\bf P}}}_{1}$. By density argument and lemma \[automorphism\] we find $\bar{{{\bf P}}}_{2}\in G$ and $f_{2}$ such that $\bar{{{\bf P}}}_{1}\leq\bar{{{\bf P}}}_{2}$ and $f_{2}$ is an automorphism of $\mbox{BA}(P^{\omega_{1}}_{2})$ extending $f$. Similarly, if $\bar{{{\bf P}}}_{4}\in G$, $\bar{{{\bf P}}}_{3}\leq\bar{{{\bf P}}}_{4}$ and $f_{3}$ is an automorphism of $\mbox{BA}(P^{\omega_{1}}_{3})$ then there are $\bar{{{\bf P}}}_{5}\in G$, $f_{5}$ such that $f_{5}$ is an automorphism of $\mbox{BA}(P^{\omega_{1}}_{5})$ extending $f_{3}$. It follows from the above that, in $V[G]$, we can extend $f$ to an automorphism of BA(${{\bf P}}$). 4\. Similar arguments as in 1. and 2. [width 6pt height 6pt depth 0pt ]{} Theorems \[first\] and \[second\] follow directly from the above proposition and theorems \[Solovay\] and \[Solovaybis\]. [References]{} \[Jec\] T.Jech, [*Set Theory*]{}, Academic Press, New York 1978. \[JS1\] H.Judah, S.Shelah, [*Martin’s axioms, measurability and equiconsistency results*]{}, Journal of Symbolic Logic, 54(1989). \[JS2\] H.Judah, S.Shelah, [*Souslin forcing*]{}, Journal of Symbolic Logic, 53(1988). \[JuR\] H.Judah, A.Roslanowski, [*On Shelah’s amalgamation*]{}, submitted to [**Set Theory of Reals**]{}, Proceedings of Bar-Ilan Conference in honour of Prof. Abraham Fraenkel, Bar-Ilan University 1991. \[She\] S.Shelah, [*Can you take Solovay’s inaccessible away?*]{}, Israel Journal of Mathematics, 48(1984), pp 1:47. \[Sol\] R.Solovay, [*A model of set theory in which every set of reals is Lebesgue measurable*]{}, Annals of Mathematics 92(1970), pp 1:56.

ArXiv

--- abstract: 'The challenge of unsupervised person re-identification (ReID) lies in learning discriminative features without true labels. This paper formulates unsupervised person ReID as a multi-label classification task to progressively seek true labels. Our method starts by assigning each person image with a single-class label, then evolves to multi-label classification by leveraging the updated ReID model for label prediction. The label prediction comprises similarity computation and cycle consistency to ensure the quality of predicted labels. To boost the ReID model training efficiency in multi-label classification, we further propose the memory-based multi-label classification loss (MMCL). MMCL works with memory-based non-parametric classifier and integrates multi-label classification and single-label classification in a unified framework. Our label prediction and MMCL work iteratively and substantially boost the ReID performance. Experiments on several large-scale person ReID datasets demonstrate the superiority of our method in unsupervised person ReID. Our method also allows to use labeled person images in other domains. Under this transfer learning setting, our method also achieves state-of-the-art performance.' author: - | Dongkai Wang  Shiliang Zhang\ Peking University\ [{dongkai.wang, slzhang.jdl}@pku.edu.cn]{} bibliography: - 'egbib.bib' title: 'Unsupervised Person Re-identification via Multi-label Classification' --- Introduction ============ ![Illustrations of multi-label classification for unsupervised person ReID. We target to assign each unlabeled person image with a multi-class label reflecting the person identity. This is achieved by iteratively running MPLP for prediction and MMCL for multi-label classification loss computation. This procedure guides CNN to produce discriminative features for ReID.[]{data-label="fig:idea"}](figures/illustration-class.pdf){width="1\linewidth"} Recent years have witnessed the great success of person re-identification (ReID), which learns discriminative features from labeled person images with deep Convolutional Neural Network (CNN) [@zheng2016person; @krizhevsky2012imagenet; @he2016deep; @su2015multi; @li2019pose; @su2017pose; @su2016deep; @su2017attributes]. Because it is expensive to annotate person images across multiple cameras, recent research efforts start to focus on unsupervised person ReID. Unsupervised person ReID aims to learn discriminative features from unlabeled person images. Compared with supervised learning, unsupervised learning relieves the requirement for expensive data annotation, hence shows better potential to push person ReID towards real applications. The challenge of unsupervised person ReID lies in learning discriminative features without true labels. To conquer this challenge, most of recent works [@yu2017cross; @zhong2019invariance; @Lv_2018_CVPR; @lin2018multi; @wang2018transferable] define unsupervised person ReID as a transfer learning task, which leverages labeled data on other domains for model initialization or label transfer. Among them, some works assign each image with a single-class label [@zhong2019invariance]. Some others leverage spatio-temporal cues or additional attribute annotations [@Lv_2018_CVPR; @lin2018multi; @wang2018transferable]. Detailed review of existing methods will be presented in Sec. \[sec:relatedwork\]. Thanks to the above efforts, the performance of unsupervised person ReID has been significantly boosted. However, there is still a considerable gap between supervised and unsupervised person ReID. Meanwhile, the setting of transfer learning leads to limited flexibility. For example, as discussed in many works [@long2015learning; @Yan_2017_CVPR; @wei2018person], the performance of transfer learning is closely related to the domain gap, *e.g.*, large domain gap degrades the performance. It is non-trivial to estimate the domain gap and select suitable source datasets for transfer learning in unsupervised person ReID. This paper targets to boost unsupervised person ReID without leveraging any labeled data. As illustrated in Fig. \[fig:idea\], we treat each unlabeled person image as a class and train the ReID model to assign each image with a multi-class label. In other words, the ReID model is trained to classify each image to multiple classes belonging to the same identity. Because each person usually has multiple images, multi-label classification effectively identifies images of the same identity and differentiates images from different identities. This in-turn facilitates the ReID model to optimize inter and intra class distances. Compared with previous methods [@lin2019bottom; @wu2018unsupervised], which classify each image into a single class, the multi-label classification has potential to exhibit better efficiency and accuracy. Our method iteratively predicts multi-class labels and updates the network with multi-label classification loss. As shown in Fig. \[fig:idea\], to ensure the quality of predicted labels, we propose the Memory-based Positive Label Prediction (MPLP), which considers both visual similarity and cycle consistency for label prediction. Namely, two images are assigned with the same label if they a) share large similarity and b) share similar neighbors. To further ensure the accuracy of label prediction, MPLP utilizes image features stored in the memory bank, which is updated with augmented features after each training iteration to improve feature robustness. Predicted labels allow for CNN training with a multi-label classification loss. Since each image is treated as a class, the huge number of classes makes it hard to train classifiers like Fully Connected (FC) layers. As shown in Fig. \[fig:idea\], we adopt the feature of each image stored in the memory bank as a classifier. Specifically, a Memory-based Multi-label Classification Loss (MMCL) is introduced. MMCL accelerates the loss computation and addresses the vanishing gradient issue in traditional multi-label classification loss [@zhang2013review; @Durand_2019_CVPR] by abandoning the sigmoid function and enforcing the classification score to 1 or -1. MMCL also involves hard negative class mining to deal with the imbalance between positive and negative classes. We test our approach on several large-scale person ReID datasets including Market-1501 [@zheng2015scalable], DukeMTMC-reID [@ristani2016performance] and MSMT17 [@wei2018person] without leveraging other labeled data. Comparison with recent works shows our method achieves competitive performance. For instance, we achieve rank-1 accuracy of 80.3% on Market-1501, significantly outperforming the recent BUC [@lin2019bottom] and DBC [@ding12dispersion] by 14.1% and 11.1%, respectively. Our performance is also better than the HHL [@zhong2018generalizing] and ECN [@zhong2019invariance], which use extra DukeMTMC-reID [@ristani2016performance] for transfer learning. Our method is also compatible with transfer learning. Leveraging DukeMTMC-reID for training, we further achieve rank-1 accuracy of 84.4% on Market-1501. In summary, our method iteratively runs MPLP and MMCL to seek true labels for multi-label classification and CNN training. As shown in our experiments, this strategy, although does not leverage any labeled data, achieves promising performance. The maintained memory bank reinforces both label prediction and classification. Our work also shows that, unsupervised training has potential to achieve better flexibility and accuracy than existing transfer learning strategies. Related Work {#sec:relatedwork} ============ This section briefly reviews related works on unsupervised person ReID, unsupervised feature learning, and multi-label classification. *Unsupervised person ReID* works can be summarized into three categories. The first category utilizes hand-craft features [@liao2015person; @zheng2015scalable]. However, it is difficult to design robust and discriminative features by hand. The second category [@fan2018unsupervised; @Fu_2019_ICCV] adopts clustering to estimate pseudo labels to train the CNN. However, these methods require good pre-trained model. Apart from this, Lin *et al.* [@lin2019bottom] treat each image as a cluster, through training and merging clusters, this method achieves good performance. The third category [@lin2018multi; @wang2018transferable; @yu2019unsupervised; @wei2018person; @deng2018image; @zhong2018generalizing; @zhong2019invariance; @Chen_2019_ICCV; @Wu_2019_ICCV; @Li_2019_ICCV; @Qi_2019_ICCV; @Zhang_2019_ICCV] utilizes transfer learning to improve unsupervised person ReID. Some works [@lin2018multi; @wang2018transferable] use transfer learning and minimize the attribute-level discrepancy by utilizing extra attribute annotations. MAR [@yu2019unsupervised] uses the source dataset as a reference to learn soft labels, which hence supervise the ReID model training. Generative Adversarial Network (GAN) is also utilized for transfer learning. PTGAN [@wei2018person] and SPGAN [@deng2018image] first generate transferred images from source datasets, then uses transferred image for training. HHL [@zhong2018generalizing] generates images under different cameras and trains network using triplet loss. ECN [@zhong2019invariance] utilizes transfer learning and minimizes the target invariance. Transfer learning requires a labeled source dataset for training. Our method differs with them that, it does not require any labeled data. It also achieves better performance than many transfer learning methods. *Unsupervised feature learning* aims to relieve the requirement on labeled data for feature learning. It can be applied in different tasks. Some works adopt unsupervised feature learning for neural network initialization. For example, RotNet [@komodakis2018unsupervised] predicts the rotation of image to learn a good representation. Li *et al.* [@li2018unsupervised] use motion and view as supervision to learn an initialization for action recognition. Some other works utilize unsupervised feature learning to acquire features for image classification and retrieval. [@iscen2018mining] utilizes manifold learning to seek positive and negative samples to compute the triplet loss. Wu *et al.* [@wu2018unsupervised] regard each image as single class, and propose a non-parametric softmax classifier to train CNN. Our work shares certain similarity with [@wu2018unsupervised], in that we also use non-parametric classifiers. However, we consider multi-label classification, which is important in identifying images of the same identity as well as differentiating different identities. *Multi-label classification* is designed for classification tasks with multi-class labels [@zhang2013review; @Durand_2019_CVPR; @wang2018transferable; @lin2018multi]. Durand *et al.* [@Durand_2019_CVPR] deal with multi-label learning based on partial labels and utilize GNN to predict missing labels. Wang *et al.* [@wang2018transferable; @lin2018multi] use multi-label classification to learn attribute feature. This paper utilizes multi-label classification to predict multi-class labels and focuses on learning identity feature for person ReID. To the best of our knowledge, this is an early work utilizing multi-label classification for unsupervised person ReID. Methodology =========== Formulation -----------  \[sec:formulate\] Given an unlabeled person image dataset $\mathcal X=\{x_1,x_2,...,x_n\}$, our goal is to train a person ReID model on $\mathcal X$. For any query person image $q$, the person ReID model is expected to produce a feature vector to retrieve image $g$ containing the same person from a gallery set $G$. In other words, the ReID model should guarantee $q$ share more similar feature with $g$ than with other images in $G$. We could conceptually denote the goal of person ReID as, $$g^* = \arg\min_{g\in G} \operatorname {dist}(f_g,f_q),$$ where $f \in \mathbb{R}^d$ is a $d$-dimensional L2-normalized feature vector extracted by the person ReID model. $\operatorname {dist}(\cdot)$ is the distance metric, *e.g.*, the L2 distance. To make training on $\mathcal X$ possible, we start by treating each image as an individual class and assign $x_i$ with a label $y_i$. This pseudo label turns $\mathcal X$ into a labeled dataset, and allows for the ReID model training. $y_i$ is initialized to a two-valued vector, where only the value at index $i$ is set to $1$ and the others are set to $-1$, *i.e.*, $$y_i[j]=\left\{ \begin{array}{rcl} 1 & & {j = i}\\ -1 & & {j \neq i} \end{array} \right.$$ Since each person may have multiple images in $\mathcal X$, the initial label vector is not valid in representing person identity cues. Label prediction is required to assign multi-class labels to each image, which can be used for ReID model training with a multi-label classification loss. Labels of $x_i$ can be predicted by referring its feature $f_i$ to features of other images, and find consistent feature groups. On the other hand, due to the huge number of image classes in $\mathcal X$, it is hard to train a multi-label classifier. One efficient solution is to use the $f_i$ as the classifier for the $i$-th class. This computes the classification score for any image $x_j$ as, $$c_j[i]= f_i^\top \times f_j,$$ where $c_j$ denotes the multi-label classification score for $x_j$. It is easy to infer that, both label prediction and multi-label classification require features of images in $\mathcal X$. We hence introduce a $n \times d$ sized memory bank $\mathcal{M}$ to store those features, where $\mathcal{M}[i] =f_i$. With $\mathcal{M}$, we propose the Memory-based Positive Label Prediction (MPLP) for label prediction and Memory-based Multi-label Classification Loss (MMCL) for ReID model training, respectively. As shown in Fig. \[fig:idea\], MPLP takes a single-class label as input and outputs the multi-label prediction $\bar y_i$ based on memory bank $\mathcal{M}$, *i.e.*, $$~\label{eq:mplp} \bar y_i = \operatorname {MPLP}(y_i, \mathcal{M}),$$ where $\operatorname {MPLP}(\cdot)$ denotes the MPLP module and $\bar y$ is the multi-class label. MMCL computes the multi-label classification loss by taking the image feature $f$, label $\bar y$, and the memory bank $\mathcal{M}$ as inputs. The computed loss $\mathcal{L}_{MMCL}$ can be represented as, $$\label{eq:mmcl} \mathcal{L}_{MMCL} = \sum_{i=1}^{n} \mathcal D(\mathcal{M}^\top \times f_i, \bar y_i),$$ where $\mathcal{M}^\top \times f_i$ computes the classification score, and $\mathcal D(\cdot)$ computes the loss by comparing classification scores and predicted labels. $\mathcal{M}$ is updated after each training iteration as, $$\mathcal{M}[i]^t = \alpha \cdot f_i + (1-\alpha) \cdot \mathcal{M}[i]^{t-1},$$ where the superscript $t$ denotes the $t$-th training epoch, $\alpha$ is the updating rate. $\mathcal{M}[i]^t$ is then L2-normalized by $\mathcal{M}[i]^t \leftarrow||\mathcal{M}[i]^t||_2$. It is easy to infer that, both MPLP and MMCL require robust features in $\mathcal{M}$ to seek reliable labels and classification scores, respectively. We use many data argumentation techniques to reinforce $\mathcal{M}$. In other words, each $\mathcal{M}[i]$ combines features of different augmented samples form $x_i$, hence it presents better robustness. More details are given in Sec. \[sec:implement\]. MPLP considers both similarity and cycle consistency to predict $\bar y_i$, making it more accurate than the classification score. This makes the loss computed with Eq.  valid in boosting the ReID model, which in-turn produces positive feedbacks to $\mathcal{M}[i]$ and label prediction. This loop makes it possible to train discriminative ReID models on unlabeled dataset. Implementations to MPLP and MMCL can be found in the following parts. Memory-based Positive Label Prediction {#section:mplp} -------------------------------------- As shown in Eq. , given an initial two-valued label $y_i$ of image $x_i$, MPLP aims to find other classes that $x_i$ may belong to. For $x_i$, MPLP first computes a rank list ${R}_{i}$ according to the similarity between $x_i$ and other features, *i.e.*, $$\label{eq:labelrank} {R}_{i} = \mathop{\arg \operatorname {sort}}_j (s_{i,j}), j \in [1,n],$$ $$\label{eq:mem_sim} s_{i,j} = \mathcal{M}[i]^\top \times \mathcal{M}[j],$$ where $s_{i,j}$ denotes the similarity score of $x_i$ and $x_j$. ${R}_{i}$ finds candidates for reliable labels for $x_i$, *e.g.*, labels at the top of rank list. However, variances of illumination, viewpoint, backgrounds, *etc*., would degrade the robustness of the rank list. *E.g.*, noisy labels may appear at the top of rank list. To ensure the quality of predicted labels, MPLP refers to the similarity score and cycle consistency for label prediction. **Label filtering by similarity score:** We first select positive label candidates for $x_i$ based on its rank list. Inspired by [@zhang2013review], that uses a threshold to select relevant labels for query, we select candidate labels with a predefined similarity threshold. Given a similarity score threshold $t$, $k_i$ label candidates can be generated by removing labels with similarity smaller than $t$, *i.e.*, $$\begin{aligned}\label{eq:t_select} {P}_{i} = {R}_{i}[1:k_i], \end{aligned}$$ where ${R}_{i}[k_i]$ is the last label with similarity score higher than $t$, ${P}_{i}$ is the collection of label candidates for $x_i$. $t$ largely decides the quantity of label candidates. It will be tested in Sec. \[section:pa\]. Eq.  adaptively finds different numbers of candidates labels for different images, which is better than selecting fixed number of labels, *i.e.*, the KNN in Fig. \[fig:visualization\]. We proceed to introduce cycle consistency to find positive labels from ${P}_{i}$.  \[fig:visualization\] **Label filtering by cycle consistency:** Inspired by k-reciprocal nearest neighbor[@jegou2007contextual; @zhong2017re], we assume that, if two images belong to the same class, their neighbor image sets should also be similar. In other words, two images should be mutual neighbor for each other if they can be assigned with similar labels. With this intuition, we propose a cycle consistency scheme to filter hard negative labels in $P_i$. MPLP traverses labels in $P_i$ from head to tail. For a label $j$ in ${P}_{i}$, MPLP computes its top-$k_i$ nearest labels with Eq. . If label $i$ is also one of the top-$k_i$ nearest labels of $j$, $j$ is considered as a positive label for $x_i$. Otherwise, it is treated as a hard negative label. The traverse is stopped when the first hard negative label is found. This leads to a positive label set $P^*_i$ as well as a hard negative label for image $x_i$. We denote the positive label set as, $${P}^*_{i} = {P}_{i}[1:l],$$ where $l$ satisfies $i \in {R}_{{P_i[l]}}[1:k_i] ~\&~ i \notin {R}_{{P_i[l+1]}}[1:k_i]$. As ${P}^*_{i}$ contains $l$ labels, $x_i$ would be assigned with a multi-class label $\bar y_i$ with $l$ positive classes, $$\bar y_i[j]=\left\{ \begin{array}{rcl} 1 & & {j\in {P}^*_{i}}\\ -1 & & {j \notin {P}^*_{i}} \end{array} \right.$$ As Fig. \[fig:visualization\] shows that, MPLP predicts accurate positive labels. Experimental evaluations will be presented in Sec. \[section:ablation\]. Memory-based Multi-label Classification Loss {#section:mmcl} -------------------------------------------- **Traditional multi-label classification loss:** The predicted multi-class labels are used for training the ReID model with a multi-label classification loss. In traditional multi-label classification methods, sigmoid and logistic regression loss is a common option [@zhang2013review; @Durand_2019_CVPR; @wang2018transferable; @lin2018multi]. For a task with $n$ classes, it adopts $n$ independent binary classifiers for classification. The loss of classifying image $x_i$ to class $j$ can be computed as, $$~\label{eq:mcl} \ell(j|x_i) = \log (1 + \exp (-\bar y_i[j]\times \mathcal{M}[j]^\top \times f_i)),$$ where $\mathcal{M}[j]^\top \times f_i$ computes the classification score of image $x_i$ for the class $j$. $\bar y_i[j]$ is the label of image $x_i$ for class $j$. With the loss at a single class, we can obtain the Multi-Label Classification (MCL) loss, *i.e.*, $\mathcal{L}_{MCL}$, $$\mathcal{L}_{MCL} = \sum_{i=1}^{n}\sum_{j=1}^{n}\ell(j|x_i),$$ where $n$ is the number of images in the dataset $\mathcal X$, which equals to the class number in our setting.  \[fig:gradientanalysis\] Because the $\mathcal{M}[j]^\top$ and $f_i$ are L2 normalized, the classification score is restricted between $[-1,1]$. This limits the range of sigmoid function in Eq. , making the loss non-zero even for correct classifications. This issue can be addressed by introducing a scalar $\tau$ on the classification score. This updates Eq.  as, $$~\label{eq:mcl-t} \ell_\tau(j|x_i) = \log (1 + \exp (-\bar y_i[j]\times \mathcal{M}[j]^\top \times f_i/\tau)).$$ We denote the corresponding MCL loss as $\mathcal{L}_{MCL-\tau}$. The gradient of $\mathcal{L}_{MCL-\tau}$ can be computed as, $$~\label{eq:grad_mcl} \begin{aligned} \frac{\partial{\mathcal{L}_{MCL-\tau}}}{\partial{f_i}} = -\frac{\exp(-\bar y_i[j] \mathcal{M}[j]^\top f_i /\tau)}{1+\exp(-\bar y_i[j]\mathcal{M}[j]^\top f_i/\tau)}\frac{\bar y_i[j]\mathcal{M}[j]}{\tau}. \end{aligned}$$ With Eq. , we illustrate the gradient of $\mathcal{L}_{MCL-\tau}$ with different values of $\tau$ when $\bar y_i[j]=1$ in Fig. \[fig:gradientanalysis\]. It is clear that, the updated MCL loss still suffers from substantial vanishing gradient issue as the classification score larger than 0.25 or smaller than -0.25. Another issue with MCL loss is that, our task involves a large number of classes, making the positive and negative classes unbalanced. Treating those negative classes equally in Eq.  may cause a model collapse. We hence proceed to propose MMCL to address those issues. **Memory-based Multi-label Classification Loss:** MMCL is proposed to address two issues in traditional MCL. For the first issue, since the score is bounded by $[-1,1]$, we can abandon the sigmoid function and directly compute the loss by regressing the classification score to 1 and -1. This simplifies the loss computation and improves the training efficiency. The loss of classifying image $x_i$ to class $j$ can be updated as, $$\ell^*(j|x_i) = ||\mathcal{M}[j]^\top \times f_i -\bar y_i[j]||^2,$$ where $f_i$ is the feature of image $x_i$. The second problem is the imbalance between positive and negative classes. MMCL introduces hard negative class mining to solve it. This is inspired by the sample mining in deep metric learning [@wu2017sampling], where hard negative samples are more informative for training. Similarly in our multi-label classification, the training should focus more on hard negative classes than easy negative classes. For $x_i$, its negative classes can be denoted as $R_{i} \backslash P^*_{i}$. We rank them by their classification scores and select the top $r\%$ classes as the hard negative classes. The collection of hard negative classes for $x_i$ can be denoted as ${N}_{i}, |{N}_{i}|=(n-|P^*_{i}|)\cdot r\%$. The MMCL is computed on positive classes and sampled hard negative classes as follows, $$\label{mmcl} \begin{aligned} \mathcal{L}_{MMCL} = \sum_{i=1}^{n} \frac{\delta}{|{P}^*_{i}|}\sum_{p \in \mathcal{P}^*_{i}}\ell^*(p|x_i) + \\ \frac{1}{|{N}_{i}|}\sum_{s \in {N}_{i}}\ell^*(s|x_i) \end{aligned}$$ where $\delta$ is a coefficient measuring the importance of positive class loss and negative class loss, which will be tested in experiments. We also illustrate the gradients of $\mathcal{L}_{MMCL}$ similarly when $\bar y_i[j]=1$, in Fig. \[fig:gradientanalysis\], where the gradient of $\mathcal{L}_{MMCL}$ can be computed as, $$\begin{aligned} \partial{\mathcal{L}_{MMCL}}/{\partial{f_i}}= 2\delta (\mathcal{M}[j]^\top \times f_i -\bar y_i[j])\mathcal{M}[j]. \end{aligned}$$ **Discussions:** Comparison between MCL and MMCL in Fig. \[fig:gradientanalysis\] clearly shows that, the vanishing gradient issue is effectively addressed by MMCL. Because of vanishing gradient, $\mathcal{L}_{MCL-\tau}$ won’t enforce the classifier to classify positive labels with large scores. This is harmful for decreasing the intra-class variance. Therefore, MMCL is more effective than MCL in optimizing the ReID model. Fig. \[fig:gradientanalysis\] also shows that, $\delta$ controls the magnitude of the gradient of MMCL. As discussed in [@NIPS2018_8094], mean square loss is inferior to log-based loss (*e.g.* cross entropy) when classification score is near the decision boundary. $\delta$ effectively solves this issue by scaling the gradient magnitude of MMCL. By adopting the hard negative class mining strategy, MMCL not only works for multi-label classification, it also could be applied in single-label classification, where the unbalanced class issue still exists. Compared with cross entropy loss and MCL, MMCL abandons activation functions like softmax and sigmoid, leading to more efficient computation. As discussed in many works [@morin2005hierarchical; @gutmann2010noise], a huge number of classes degrades the speed of softmax computation. Existing solutions include hierarchical softmax [@morin2005hierarchical] and noise-contrastive estimation [@gutmann2010noise]. As MMCL does not involve softmax computation, it does not suffer from such issues. Transfer Learning with Labeled Dataset {#section:transfer} -------------------------------------- Our method is also compatible with transfer learning setting. Given a dataset containing labeled person images, we can adopt the commonly used cross entropy loss and triplet loss on labeled data to train the model. The training loss can be denoted as $\mathcal{L}_{labeled}$. The overall training loss for transfer learning can be represented as the sum of MMCL and loss on labeled dataset, *i.e.*, $$\mathcal {L}_{transfer} = \mathcal{L}_{labeled} + \mathcal{L}_{MMCL}.$$ The performance of our methods on transfer learning will be tested in the next section. Experiments =========== Dataset and Evaluation Metrics ------------------------------ *Market-1501* [@zheng2015scalable] contains 32,668 labeled person images of 1,501 identities collected from 6 non-overlapping camera views. *DukeMTMC-reID* [@ristani2016performance] has 8 cameras and 36,411 labeled images of 1,404 identities. *MSMT17* [@wei2018person] is a newly released person ReID dataset. It is composed of 126,411 person images from 4,101 identities collected by 15 cameras. The dataset suffers from substantial variations of scene and lighting, and is more challenging than the other two datasets. All three datasets are collected under similar scenario, *i.e.*, campus, which makes the transfer learning possible. We follow the standard settings [@zheng2015scalable; @ristani2016performance; @wei2018person] on them to conduct experiments. Performance is evaluated by the Cumulative Matching Characteristic (CMC) and mean Average Precision (mAP).  \[figure:pa-sst\] Implementation Details {#sec:implement} ---------------------- All experiments are implemented on PyTorch. We use ResNet-50 [@he2016deep] as backbone to extract the feature and initialize it with parameters pre-trained on ImageNet [@5206848]. After pooling-5 layer, we remove subsequent layers and add a batch normalization layer [@ioffe2015batch], which produces a 2048-dim feature. During testing, we also extract the pooling-5 feature to calculate the distance. For multi-label classification, we allocate a memory bank to store L2 normalized image features. The memory bank is initialized to all zeros, and we start using MPLP for label prediction when the memory is fully updated 5 times (after 5 epochs). As mentioned in section \[section:mplp\], we leverage CamStyle [@zhong2018camera] as a data augmentation strategy for unlabeled images. Strategies like random crop, random rotation, color jitter, and random erasing are also introduced to improve the feature robustness. The input image is resized to 256\*128. We use SGD to optimize the model, the learning rate for ResNet-50 base layers are 0.01, and others are 0.1. The memory updating rate $\alpha$ starts from $0$ and grows linearly to $0.5$. We train the model for 60 epochs, and the learning rate is divided by 10 after every 40 epochs. The batch size for model training is 128. We fix the similarity threshold $t$ in MPLP as 0.6. In MMCL, the weight $\delta$ is fixed to 5 and we select the 1% top-ranked negative classes to compute the loss through the parameter analysis in Sec. \[section:pa\]. For transfer learning with labeled dataset, we apply the same batch size on the labeled dataset. A fully connected layer is added after batch normalization layer for classification. We optimize the $\mathcal {L}_{transfer}$ in section \[section:transfer\] following the same baseline training strategy as described in [@Fu_2019_ICCV]. Parameter Analysis {#section:pa} ------------------ This section aims to investigate some important hyper-parameters in our method, including the similarity score threshold $t$ in MPLP, coefficient $\delta$, and hard negative mining ratio $r$% in MMCL. Each experiment varies the value of one hyper-parameter while keeping others fixed. All experiments are conducted with unsupervised ReID setting on both Market-1501 and DukeMTMC-reID. [p[1cm]{}&lt;|cc|cc]{} & &\ & Rank-1 & mAP & Rank-1 & mAP\ 1 & 59.3 & 19.4 & 52.6 & 22.8\ 2 & 71.3 & 31.1 & 58.8 & 31.0\ 3 & 76.6 & 40.0 & 62.6 & 35.6\ 4 & 79.9 & 44.9 & 64.9 & 39.1\ 5 & **80.3** & **45.5** & **65.2** & **40.2**\ 6 & 78.1 & 45.0 & 65.0 & 39.8\ 7 & 73.2 & 41.3 & 64.1 & 39.4\ \[table:pa-delta\] **Similarity threshold $t$:** Fig.\[figure:pa-sst\] investigates the effect of similarity threshold $t$ in MPLP. We vary $t$ from $0.3$ to $0.7$ and test the model performance. A low similarity score $t$ will harm the model performance. For example, when $t$ is in range $[0.3,0.5]$, a substantial performance drop can be observed compared with larger $t$. This is because that, low similarity threshold introduces many negative labels. More accurate labels can be selected as $t$ becomes larger. However, too large $t$ decreases the number of selected labels. The best $t$ is 0.6 for both Market-1501 and DukeMTMC-reID. We hence set $t=0.6$. **Coefficient $\delta$:** Table \[table:pa-delta\] reports the analysis on coefficient $\delta$ of MMCL. As discussed in Sec. \[section:mmcl\], $\delta$ plays a role to scale the gradient of MMCL. $\delta=1$ means that we do not scale the gradient. In this case, the MMCL cannot produce large gradients to pull positive samples together, leading to bad performance. For example, the rank-1 accuracy is dropped to 59.3% on Market-1501 and 52.6% on DukeMTMC-reID. As $\delta$ becomes larger, MMCL effectively improves the similarity of positive samples, leading to better performance. However, too large $\delta$ may make the training unstable. According to Table \[table:pa-delta\], we set $\delta=5$. **Hard negative mining ratio $r$%:** Fig. \[figure:pa-hnmr\] shows effects of hard negative mining ratio $r$% in network training. $r=100$ means using all negative classes for loss computation. It is clear that, $r=100$ is harmful for the performance. This implies that, not all of the negative classes are helpful for unsupervised ReID training. As $r$ becomes smaller, hard negative mining would be activated and it boosts the performance. Too small $r$ selects too few negative classes, hence is also harmful for the performance. According to Fig. \[figure:pa-hnmr\], $r=1$ is used in our experiments. ![Evaluation of hard negative mining ratio $r$ in MMCL.[]{data-label="figure:pa-hnmr"}](figures/pa-hnmr.pdf) ---------------------- ---------- ---------- ---------- ---------- Rank-1 mAP Rank-1 mAP Supervised 87.1 68.3 75.9 57.9 ImageNet Pre-trained 7.8 2.1 5.1 1.4 MMCL + Single Label 49.0 17.8 42.0 16.6 MMCL + MPLP$\dagger$ [66.6]{} [35.3]{} [58.0]{} [36.3]{} MMCL + MPLP **80.3** **45.5** **65.2** **40.2** ---------------------- ---------- ---------- ---------- ---------- : Test of validity of MMCL and MPLP. $\dagger$ denotes not using CamStyle [@zhong2018camera] for data augmentation. \[table:ablation\] Ablation Study {#section:ablation} -------------- This part evaluates the effectiveness of MPLP and MMCL by making comparison with supervised learning, MMCL+single-class label, and MMCL+MPLP. Experimental results are reported in Table \[table:ablation\], where the performance of ImageNet pre-trained model is also reported as the baseline. As shown in the table, the supervised learning achieves high accuracy, *e.g.*, 87.1% in rank-1 accuracy and 68.3% in mAP on Market-1501. ImageNet pre-trained model performs badly on both Market-1501 and DukeMTMC-reID. MMCL with single-class pseudo labels boosts the baseline performance, indicating the validity of leveraging unlabeled dataset in training. Table \[table:ablation\] also shows that, combining MMCL with MPLP significantly boosts the performance, *e.g.*, from baseline 7.8% to 80.3% in rank-1 accuracy on Market-1501. Table \[table:ablation\] also shows that, CamStyle [@zhong2018camera] boosts the performance, indicating the importance of data augmentation as discussed in Sec. \[sec:formulate\]. **Effectiveness of MPLP:** To verify that MPLP is a reasonably good solution for label prediction, we compare MPLP against several other label prediction methods, *e.g.*, the KNN search and selection by Similarity Score (SS). Table \[table:labelandloss\] (a) summarizes the results. From Table \[table:labelandloss\] (a), we can observe that KNN (K=8, which achieves the best peformance) achieves 73.3% rank-1 accuracy and 35.4% mAP on Market-1501. Selecting positive labels by Similarity Score (SS) gains improvements over KNN. This indicates that, adaptively select positive labels with similarity threshold is more reasonable than fixing the positive label number for different images. Table \[table:labelandloss\] (a) also shows that, MPLP achieves the best performance. This indicates that combining cycle consistency and similarity score effectively ensure the quality of predicted labels. Visualization of predicted labels by MPLP can be found in Fig. \[fig:visualization\]. [l||c|c||c|c]{} & &\ & Rank-1 & mAP & Rank-1 & mAP\ MMCL + KNN &73.3 & 35.4 & 59.5 & 33.5\ MMCL + SS & 77.6 & 43.0 & 62.7 & 38.5\ MMCL + MPLP & **80.3** & **45.5** & **65.2** & **40.2**\ \ CE + Single Label & 31.4 & 9.9 & 31.5 & 11.8\ MMCL + Single Label & 49.0 & 17.8 & 42.0 & 16.6\ CE + Ground Truth & 85.5 & 65.5 & 70.4 & 50.3\ MMCL + Ground Truth & 86.5 & 67.2 & 74.5 & 56.9\ CE + MPLP & 65.0& 34.2 & 60.2 & 35.1\ MMCL + MPLP & 80.3 & 45.5 & 65.2 & 40.2\ \ \[table:labelandloss\] **Effectiveness of MMCL:** To test the validity of MMCL, this part proceeds to compare it against Cross Entropy (CE) loss with different training settings. Experimental results are summarized in Table \[table:labelandloss\] (b). We first test MPLP and CE using single-class labels for model learning. MMCL gets 49.0% rank-1 accuracy and 17.8% mAP on Market-1501, substantially better than CE. We further test MMCL and CE using ground truth labels for learning. Note that, we modify CE according to [@zhong2019invariance] to make it applicable in multi-class learning. With ground truth labels, MMCL still performs better than CE, especially on DukeMTMC-reID. We finally test MMCL and CE with labels predicted by MPLP. MMCL still outperforms CE by large margins. Note that, MMCL uses non-parameters classifiers for training. It still achieves comparable performance with the supervised learning in Table \[table:ablation\]. Table \[table:labelandloss\] (b) thus demonstrates the effectiveness of MMCL and the training paradigm of our method. \[table:sota-market-duke\] ---------------------------- -------- ---------- ---------- ---------- ---------- Rank-1 Rank-5 Rank-10 mAP Ours None **35.4** **44.8** **49.8** **11.2** PTGAN [@wei2018person] Market 10.2 - 24.4 2.9 ECN [@zhong2019invariance] Market 25.3 36.3 42.1 8.5 SSG [@Fu_2019_ICCV] Market 31.6 - 49.6 13.2 Ours (transfer) Market **40.8** **51.8** **56.7** **15.1** PTGAN [@wei2018person] Duke 11.8 - 27.4 3.3 ECN [@zhong2019invariance] Duke 30.2 41.5 46.8 10.2 SSG [@Fu_2019_ICCV] Duke 32.2 - 51.2 13.3 Ours (transfer) Duke **43.6** **54.3** **58.9** **16.2** ---------------------------- -------- ---------- ---------- ---------- ---------- : Comparison with state-of-the-art methods on MSMT17. \[table:sota-msmt17\] Comparison with the State of the Art ------------------------------------ We compare our method against state-of-the-art unsupervised learning and transfer learning approaches on Market-1501 [@zheng2015scalable], DukeMTMC-reID [@ristani2016performance] and MSMT17 [@wei2018person]. Table \[table:sota-market-duke\] and Table \[table:sota-msmt17\] summarize the comparison. Table \[table:sota-market-duke\] reports comparisons on Market-1501 and DukeMTMC-reID. We compare two types of methods, including unsupervised learning methods: LOMO [@liao2015person], BOW [@zheng2015scalable], BUC [@lin2019bottom] and DBC [@ding12dispersion], and transfer learning based approaches: PUL [@fan2018unsupervised], PTGAN [@wei2018person], SPGAN [@deng2018image], CAMEL [@yu2017cross], MMFA [@lin2018multi], TJ-AIDL [@wang2018transferable], HHL [@zhong2018generalizing], ECN [@zhong2019invariance], MAR [@yu2019unsupervised], PAUL [@Yang_2019_CVPR], SSG [@Fu_2019_ICCV], CR-GAN [@Chen_2019_ICCV], CASCL [@Wu_2019_ICCV], PDA-Net [@Li_2019_ICCV], UCDA [@Qi_2019_ICCV] and PAST [@Zhang_2019_ICCV]. We first compare with unsupervised learning methods. LOMO and BOW utilize hand-crafted features, which show lower performance. BUC and DBC treat each image as a single cluster then merges clusters, thus share certain similarity to our work. However, our method outperforms them by large margins. The reasons could be because: 1) BUC tries to keep different clusters with similar size, hence could suffer from the issue of imbalanced number of positive labels. MPLP could alleviate this issue by adaptively selecting positive labels for different images. 2) As discussed in ablation studies, MMCL performs better than the cross entropy loss when using memory bank as classifier. Therefore, MPLP and MMCL effectively boost our performance. Under the transfer learning setting, our method achieves the best performance on Market-1501 in Table \[table:sota-market-duke\]. For example, our rank-1 accuracy of on Market-1501 achieves 84.4%, when using DukeMTMC-reID as the source dataset. Similarly, we get 72.4% rank-1 accuracy on DukeMTMC-reID using Market-1501 as the source dataset. Although SSG and PAST achieve slightly better performance, our method is more flexible and can be used without labeled dataset. It is also interesting to observe that, with unsupervised learning setting, our method still outperforms several recent transfer learning methods, *e.g.*, our rank-1 accuracy of 80.3% on Market-1501 vs. 78.38% and 64.3% of PAST [@Zhang_2019_ICCV] and UCDA [@Qi_2019_ICCV]. We also conduct experiments on MSMT17, a larger and more challenging dataset. A limited number of works report performance on MSMT17, *i.e.*, PTGAN [@wei2018person], ECN [@zhong2019invariance], and SSG [@Fu_2019_ICCV]. As table \[table:sota-msmt17\] shows, our approach outperforms existing methods by large margins under both unsupervised and transfer learning settings. For example, our method achieves 35.4% and 43.6%/40.8% rank-1 accuracy respectively. This outperforms SSG by 11.4% in rank-1 accuracy. The above experiments on three datasets demonstrate the promising performance of our MPLP and MMCL. Conclusion ========== This paper proposes a multi-label classification method to address unsupervised person ReID. Different from previous works, our method works without requiring any labeled data or a good pre-trained model. This is achieved by iteratively predicting multi-class labels and updating the network with a multi-label classification loss. MPLP is proposed for multi-class label prediction by considering both visual similarity and cycle consistency. MMCL is introduced to compute the multi-label classification loss and address the vanishing gradient issue. Experiments on several large-scale datasets demonstrate the effectiveness of the proposed methods in unsupervised person ReID. **Acknowledgement** This work is supported in part by The National Key Research and Development Program of China under Grant No. 2018YFE0118400, in part by Beijing Natural Science Foundation under Grant No. JQ18012, in part by Natural Science Foundation of China under Grant No. 61936011, 61425025, 61620106009, 61572050, 91538111.

ArXiv

--- abstract: | In this note, we investigate the behaviour of suprema for band-limited spherical random fields. We prove upper and lower bound for the expected values of these suprema, by means of metric entropy arguments and discrete approximations; we then exploit the Borell-TIS inequality to establish almost sure upper and lower bounds for their fluctuations. Band limited functions can be viewed as restrictions on the sphere of random polynomials with increasing degrees, and our results show that fluctuations scale as the square root of the logarithm of these degrees. - Keywords and Phrases: Spherical Random Fields, Suprema, Metric Entropy, Almost Sure Convergence - AMS Classification: 60G60; 62M15, 53C65, 42C15 author: - | Domenico Marinucci[^1] and Sreekar Vadlamani\ Department of Mathematics, University of Rome Tor Vergata and\ Tata Institute for Fundamental Research, Bangalore title: 'A Note on Global Suprema of Band-Limited Spherical Random Functions ' --- Introduction ============ The analysis of the behaviour of suprema of Gaussian processes is one of the classical topics in probability theory ([@RFG],[@azaisbook]); in this note, we shall be concerned with suprema of band-limited random fields defined on the unit sphere $S^{2}$. More precisely, let $T:S^{2}\times \Omega \rightarrow \mathbb{R}$ be a measurable zero mean, finite variance Gaussian field defined on for some probability space $\left\{ \Omega ,\Im ,P\right\} ;$ we assume $T(.)$ is isotropic, e.g. the vectors$$\left\{ T(x_{1}),...,T(x_{k})\right\} \text{ and }\left\{ T(gx_{1}),...,T(gx_{k})\right\}$$have the same law, for all $k\in \mathbb{N}$, $x_{1},...,x_{k}\in S^{2}$ and $g\in SO(3),$ the group of rotations in $\mathbb{R}^{3}.$ It is then known that the field $\left\{ T(.)\right\} $ is necessarily mean square continuous ([@mp2012]) and the following spectral representation holds:$$T(x)=\sum_{\ell =0}\sum_{m=-\ell }^{\ell }a_{\ell m}Y_{\ell m}(x),$$where the *spherical harmonics* $\left\{ Y_{\ell m}\right\} $ form an orthonormal systems of eigenfunctions of the spherical Laplacian, $\Delta _{S^{2}}Y_{\ell m}=-\ell (\ell +1)Y_{\ell m}$ (see [@steinweiss],[marpecbook]{})$,$ while the random coefficients $\left\{ a_{\ell m}\right\} $ form a triangular array of complex-valued, zero-mean, uncorrelated Gaussian variables with variance $E\left\vert a_{\ell m}\right\vert ^{2}=C_{\ell },$ the *angular power spectrum of the field.* In the sequel, we shall adopt the following general model for the behavious of $\left\{ C_{\ell }\right\} ;$ as $\ell \rightarrow \infty ,$ there exist $\alpha >2$ and a positive rational function $G(\ell )$ such that $$C_{\ell }=G(\ell )\ell ^{-\alpha },\,\,\,\,0<c_{1}<G(\ell )<c_{2}<\infty \label{usucon}$$Spherical random fields have recently drawn a lot of applied interest, especially in an astrophysical environment (see [@bennett2012], [marpecbook]{}); closed form expressions for the density of their maxima and for excursion probabilities have been given in ([@ChengSchwar],[chengxiao]{},[@MarVad]). In particular, the latter references exploit the Gaussian Kinematic Fundamental formula by Adler and Taylor (see [@RFG]) to approximate excursion probabilities by means of the expected value of the Euler-Poincarè characteristic for excursion sets. It is then easy to show that$${\mathbb{E}}\mathcal{L}_{0}(A_{u}(T))=2\left\{ 1-\Phi (u)\right\} +4\pi \left\{ \sum_{\ell }\frac{2\ell +1}{4\pi }C_{\ell }\frac{\ell (\ell +1)}{2}\right\} \frac{u\phi (u)}{\sqrt{(2\pi )^{3}}},$$where $\phi ,\Phi $ denote density and distribution function of a standard Gaussian variable, while $A_{u}(T):=\left\{ x\in S^{2}:T(x)\geq u\right\} .$ It is also an easy consequence of results in Ch.14 of [@RFG] that there exist $\alpha >1$ and $\mu ^{+}>0$ such that, for all $u>\mu ^{+}$$$\left\vert {\mathbb{P}}\left\{ \sup_{x\in S^{2}}T(x)>u\right\} -2\left\{ (1-\Phi (u)+u\phi (u)\lambda \right\} \right\vert \leq \left\{ 4\pi \lambda \right\} \exp (-\frac{\alpha u^{2}}{2}), \label{before}$$ where$$\lambda :=\sum_{\ell }\frac{2\ell +1}{4\pi }C_{\ell }\frac{\ell (\ell +1)}{2},$$denotes the derivative of the covariance function at the origin, see again ([@ChengSchwar],[@chengxiao],[@MarVad]). When working on compact domains as the sphere, it is often of great interest to focus on sequences of band-limited random fields; for instance, a very powerful tool for data analysis is provided by fields which can be viewed as a sequence of wavelet transforms (at increasing frequencies) of a given isotropic spherical field $T.$ More precisely, take $b(.)$ to be a $C^{\infty }$ function, compactly supported in $[\frac{1}{2},2];$ having in mind the wavelets interpretation, it would be natural to impose the partition of unity property $\sum_{\ell }b^{2}(\frac{\ell }{2^{j}})\equiv 1,$ but this condition however plays no role in our results to follow. Let us now focus on the sequence of band-limited spherical random fields$$\beta _{j}(x):=\sum_{\ell =2^{j-1}}^{2^{j+1}}b(\frac{\ell }{2^{j}})a_{\ell m}Y_{\ell m}(x),$$which have a clear interpretation as wavelet components of the original field, and as such lend themselves to a number of statistical applications, see for instance [@bkmpAoS], [@cammar],[@npw1],[@pietrobon1]. Band-limited spherical fields have also been widely studied in other context of mathematical physics, although in such cases $b(.)$ is not necessarily assumed to be smooth, see for instance [@zelditch] and the references therein. In the sequel, it will be convenient to normalize the variance of $\left\{ \beta _{j}(x)\right\} $ to unity, and thus focus on $$\widetilde{\beta }_{j}(x):=\frac{\beta _{j}(x)}{\sqrt{\sum_{\ell }b^{2}(\frac{\ell }{2^{j}})\frac{2\ell +1}{4\pi }C_{\ell }}}\text{ .}$$The sequence of fields $\left\{ \widetilde{\beta }_{j}(x)\right\} $ has covariance functions$$\rho _{j}(x,y)=\frac{\sum_{\ell }b^{2}(\frac{\ell }{2^{j}})\frac{2\ell +1}{4\pi }C_{\ell }P_{\ell }(\left\langle x,y\right\rangle )}{\sum_{\ell }b^{2}(\frac{\ell }{2^{j}})\frac{2\ell +1}{4\pi }C_{\ell }}$$and second spectral moments$$\lambda _{j}:=\frac{\sum_{\ell }b^{2}(\frac{\ell }{2^{j}})\frac{2\ell +1}{4\pi }C_{\ell }P_{\ell }^{\prime }(1)}{\sum_{\ell }b^{2}(\frac{\ell }{2^{j}})\frac{2\ell +1}{4\pi }C_{\ell }}=\frac{\sum_{\ell }b^{2}(\frac{\ell }{2^{j}})\frac{2\ell +1}{4\pi }C_{\ell }\frac{\ell (\ell +1)}{2}}{\sum_{\ell }b^{2}(\frac{\ell }{2^{j}})\frac{2\ell +1}{4\pi }C_{\ell }},$$see [@MarVad]. For fixed $j,$ as in (\[before\]) it follows from results in [@RFG] that there exist $\alpha >1$ and $\mu ^{+}>0$ such that, for all $u>\mu ^{+}$$$\left\vert {\mathbb{P}}\left\{ \sup_{x\in S^{2}}\widetilde{\beta }_{j}(x)>u\right\} -2\left\{ (1-\Phi (u)+u\phi (u)\lambda _{j}\right\} \right\vert \leq \left\{ 4\pi \lambda _{j}\right\} \exp (-\frac{\alpha u^{2}}{2}). \label{sabato}$$However, here for $j\rightarrow \infty $ we also have $\lambda _{j}\rightarrow \infty ,$ whence the previous result clearly becomes meaningless. Intuitively, sample paths become rougher and rougher as $j$ grows, hence any fixed threshold is crossed with probability tending to one. In [@MarVad], uniform bounds for band-limited fields have indeed been established, covering even nonGaussian circumstances; however these bounds require a further averaging in the space domain for the fields considered, and this averaging ensures the uniform boundedness of $\lambda _{j};$ in these circumstances, the multiplicative constant on the right-hand side of (\[sabato\]) can be simply incorporated into the exponential choosing a different constant $1<\alpha ^{\prime }<\alpha .$ There is, however, a question that naturally arises for the cases where $\lambda _{j}$ diverges - e.g., whether it is possible to provide bounds on global suprema, allowing the thresholds to grow with frequency. This is a natural question for a number of statistical applications, for instance when considering thresholding estimates or multiple testing. Loosely speaking, the issue we shall be concerned with is then related to the existence of a growing sequence $\tau _{j}$ and positive constants $c_{1},c_{2},c_{1}^{\prime },c_{2}^{\prime }$ such that $$c_{1}\leq {\mathbb{E}}\left\{ \frac{\sup_{x\in S^{2}}\widetilde{\beta }_{j}(x)}{\tau _{j}}\right\} \leq c_{2},\text{ and /or }{\mathbb{P}}\left\{ c_{1}^{\prime }\leq \frac{\sup_{x\in S^{2}}\widetilde{\beta }_{j}(x)}{\tau _{j}}\leq c_{2}^{\prime }\right\} =1$$In fact, we shall be able to be more precise with our lower bounds. To make this statement more precise, it will be convenient to write $\ell _{j}:=2^{j};$ we shall then establish the following There exist positive constants $\gamma _{1},\gamma _{2}\geq 1,$ such that$$1\leq \liminf_{j}\frac{{\mathbb{E}}\left\{ \sup_{x\in S^{2}}\widetilde{\beta }_{j}(x)\right\} }{\sqrt{4\log \ell _{j}}}\leq \limsup_{j}\frac{{\mathbb{E}}\left\{ \sup_{x\in S^{2}}\widetilde{\beta }_{j}(x)\right\} }{\sqrt{4\log \ell _{j}}}\leq \gamma _{1},$$and$${\mathbb{P}}\left( 1\leq \liminf_{j}\frac{\left\{ \sup_{x\in S^{2}}\widetilde{\beta }_{j}(x)\right\} }{\sqrt{4\log \ell _{j}}}\leq \limsup_{j}\frac{\left\{ \sup_{x\in S^{2}}\widetilde{\beta }_{j}(x)\right\} }{\sqrt{4\log \ell _{j}}}\leq \gamma _{2}\right) =1.$$ The corresponding upper bounds are proved in Section 2, while the proofs for the lower bounds are collected in Section 3. The random functions $\left\{ \widetilde{\beta }_{j}(.)\right\} $ can be viewed as restrictions to the sphere of linear combinations of polynomials with increasing degree $p_{j}=2\ell _{j}$ ([@marpecbook])$.$ Our results can then be summarized by simply stating that as $j\rightarrow \infty ,$ the supremum of $\left\{ \widetilde{\beta }_{j}(.)\right\} $ grows as twice the square root of the logarithm of $p_{j}$. Metric Entropy and Upper Bounds =============================== The result we shall give in this Section is the following. There exist a positive constant $c$ such that, for all $j\in \mathbb{N}$$${\mathbb{E}}\left\{ \sup_{x\in S^{2}}\widetilde{\beta }_{j}(x)\right\} \leq c\sqrt{4\log \ell _{j}}. \label{1.1}$$Moreover there exist another positive constant $C$ such that$${\mathbb{P}}\left( \limsup_{j}\frac{\left\{ \sup_{x}\widetilde{\beta }_{j}(x)\right\} }{\sqrt{4\log \ell _{j}}}\leq C+\frac{1}{\sqrt{2}}\right) =1. \label{1.3}$$ (\[1.1\]) Define the canonical (Dudley) metric on $S^{2}$ as follows: $$d_{j}(x,y)=\sqrt{{\mathbb{E}}\left( \widetilde{\beta }_{j}(x)-\widetilde{\beta }_{j}(y)\right) ^{2}}=\sqrt{2-2\rho _{j}(x,y)},$$see [@RFG]. Note that since $\widetilde{\beta }_{j}(x)$ is isotropic, all the distances can be measured from one fixed point (say the north pole). Therefore, $$d_{j}^{2}(x,y)=2(1-\rho _{j}(\langle \cos \theta \rangle ))$$where $\theta :=\arccos \left\langle x,y\right\rangle $ is the usual geodesic distance on the sphere, and$$1-\rho _{j}(\cos \theta )=\frac{\sum_{\ell =2^{j-1}}^{2^{j+1}}b^{2}\left( \frac{\ell }{2^{j}}\right) \frac{(2l+1)}{4\pi }C_{\ell }\left( 1-P_{\ell }(\cos \theta )\right) }{\sum_{\ell =2^{j-1}}^{2^{j+1}}b^{2}\left( \frac{\ell }{2^{j}}\right) \frac{(2\ell +1)}{4\pi }C_{\ell }}.$$Now fix $\theta <1/(K\ell )$ and use Hilb’s asymptotics ([@szego]) to obtain$$P_{\ell }(\cos \theta )=\sqrt{\frac{\theta }{\sin \theta }}\,\,J_{0}((\ell +1/2)\theta )+\delta (\theta ),\,\,\,\,\,\,\,\,\,\,\delta (\theta )=O(\theta ^{2}),$$where $$J_{0}(z):=\sum_{k=1}^{\infty }(-1)^{k}\frac{x^{2k}}{2^{2k}(k!)^{2}}\text{ ,}$$is the standard Bessel function of zeroth order; note also that $$\lim_{K\rightarrow \infty }\sup_{\theta \leq (K\ell )^{-1}}\left\vert \frac{1-J_{0}((\ell +1/2)\theta )}{\ell ^{2}\theta ^{2}}-\frac{1}{4}\right\vert =\lim_{K\rightarrow \infty }\sup_{x\leq K^{-1}}\left\vert \frac{1-J_{0}(x)}{x^{2}}-\frac{1}{4}\right\vert =0,$$which means that for all $\delta >0,$ there exist $K_{\delta }$ small enough so that$$(\frac{1}{4}-\delta )\ell ^{2}\theta ^{2}\leq 1-J_{0}((\ell +1/2)\theta )\leq (\frac{1}{4}-\delta )\ell ^{2}\theta ^{2},\,\,\text{ for all }\theta <\frac{K_{\delta }}{\ell }.$$Combining these bounds with Hilb’s asymptotics, we get for $\theta <\frac{K_{\delta }}{\ell }$ $$(\frac{1}{4}-\delta )\ell ^{2}\theta ^{2}+O(\theta ^{2})\leq 1-P_{\ell }(\cos \theta )\leq (\frac{1}{4}+\delta )\ell ^{2}\theta ^{2}+O(\theta ^{2}).$$It follows that$$\begin{aligned} 1-\rho _{j}(\cos \theta ) &=&\frac{\sum_{\ell =2^{j-1}}^{2^{j+1}}b^{2}\left( \frac{\ell }{2^{j}}\right) \frac{(2l+1)}{4\pi }C_{\ell }\left( 1-P_{\ell }(\cos \theta )\right) }{\sum_{\ell =2^{j-1}}^{2^{j+1}}b^{2}\left( \frac{\ell }{2^{j}}\right) \frac{(2\ell +1)}{4\pi }C_{\ell }} \\ &\leq &\frac{1}{4}\theta ^{2}(2^{j+1}+1/2)^{2}+O(\theta ^{2})\end{aligned}$$and likewise$$1-\rho _{j}(\cos \theta )\geq \frac{1}{4}\theta ^{2}(2^{j-1}+1/2)^{2}+O(\theta ^{2}),$$thus implying that, for some constants $c_{1},c_{2}>0$ $$c_{1}\theta ^{2}2^{2j}\leq 1-\rho _{j}(\cos \theta )\leq c_{2}\theta ^{2}2^{2j}.$$We hence get$$c_{1}^{\prime }\theta ^{2}\leq \frac{d_{j}^{2}(0,(\theta ,\phi ))}{\ell _{j}^{2}}\leq c_{2}^{\prime }\theta ^{2},$$and more generally for $\xi _{1},\xi _{2}\in S^{2}$$$c_{1}^{\prime }d_{S}^{2}(\xi _{1},\xi _{2})\leq \frac{d_{j}^{2}(\xi _{1},\xi _{2})}{\ell _{j}^{2}}\leq c_{2}^{\prime }d_{S}^{2}(\xi _{1},\xi _{2}),$$where $d_{S}(\xi _{1},\xi _{2}):=\arccos (\left\langle \xi _{1},\xi _{2}\right\rangle )$ is the standard spherical distance. Now for $\varepsilon <C$ and $\theta <\frac{C}{\ell }$, define the sequence of $d_{j}$-balls $B_{d_{j}}(\xi _{jk},\varepsilon )=\{u\in S^{2}:d_{j}(\xi _{jk},u)\leq \varepsilon \},$ which can be rewritten as $$B_{d_{j}}(\xi _{jk},\varepsilon )=\{\xi \in S^{2}:d_{j}(\xi _{jk},\xi )=\ell _{j}d_{S}(\xi _{jk},\xi )\leq \varepsilon \}.$$Hence $B_{d_{j}}(\xi _{jk},\varepsilon )$ is a spherical cap of radius $\sim \frac{\varepsilon }{\ell },$ with Euclidean volume $$B_{d}(\xi _{jk},\varepsilon )\sim \frac{\varepsilon ^{2}}{\ell _{j}^{2}}.$$It follows that the number of $d_{j}$-balls needed to cover the sphere is asymptotic to $N_{j}(\varepsilon )\sim \frac{\ell _{j}^{2}}{\varepsilon ^{2}}.$ Consequently, by Theorem 1.3.3. of [@RFG], for any $\delta \in (0,\pi ]$ there exists a universal constant $K^{\ast }$ such that $$\begin{aligned} {\mathbb{E}}\left( \sup_{x}\widetilde{\beta }_{j}(x)\right) &\leq &K^{\ast }\int_{0}^{\delta }\sqrt{\log N_{j}(\varepsilon )}\,d\varepsilon \\ &\leq &K^{\ast }\left\{ \int_{0}^{C/\ell _{j}}\sqrt{\log N_{j}(\varepsilon )}\,d\varepsilon +\int_{C/\ell _{j}}^{\delta }\sqrt{\log N_{j}(\varepsilon )}\,d\varepsilon \right\} .\end{aligned}$$Clearly for $\varepsilon >C/\ell _{j}$ one has $N_{j}(\varepsilon )\leq c\ell _{j}^{4},$ whence$$\int_{C/\ell _{j}}^{\delta }\sqrt{\log N_{j}(\varepsilon )}\,d\varepsilon \leq c^{\prime }\sqrt{4\log \ell _{j}}.$$On the other hand $$\int_{0}^{\delta }\sqrt{\log N_{j}(\varepsilon )}\,d\varepsilon =\int_{0}^{\delta }\sqrt{2\log \left( \frac{\ell _{j}}{\varepsilon }\right) }\,d\varepsilon =\ell _{j}\int_{\sqrt{2\log \frac{\ell _{j}}{\delta }}}^{\infty }v^{2}\exp \left( -\frac{v^{2}}{2}\right) \,dv,$$with the change of variables $\frac{\varepsilon }{\ell _{j}}=\exp \left( -v^{2}/2\right) ,$ whence $$\begin{aligned} &&\int_{0}^{\delta }\sqrt{\log N_{j}(\varepsilon )}\,d\varepsilon \\ &=&\ell _{j}\left( \left. (-1)v\exp \left( -\frac{v^{2}}{2}\right) \right\vert _{\sqrt{2\log \frac{\ell _{j}}{\delta }}}^{\infty }+\int_{\sqrt{2\log \frac{l}{\delta }}}^{\infty }\exp \left( -\frac{v^{2}}{2}\right) \,dv\right) \\ &\leq &\ell _{j}\left( \frac{\delta }{\ell _{j}}\sqrt{2\log \left( \frac{\ell _{j}}{\delta }\right) }+\left( 2\log \left( \frac{\ell _{j}}{\delta }\right) \right) ^{-1/2}\frac{\delta }{\ell _{j}}\right) \\ &=&\delta \left( \sqrt{2\log \left( \frac{\ell _{j}}{\delta }\right) }+\left( 2\log \left( \frac{\ell _{j}}{\delta }\right) \right) ^{-1/2}\right) \leq c\sqrt{4\log \ell _{j}}.\end{aligned}$$Taking the same $C$ as in the entropy upper bound, and using the Borell-TIS inequality (cf. [@RFG]) we have $$\begin{aligned} {\mathbb{P}}\left( \left\{ \sup_{x}\widetilde{\beta }_{j}(x)\right\} >(C+\varepsilon )\sqrt{4\log \ell _{j}}\right) &\leq &{\mathbb{P}}\left( \Vert \widetilde{\beta }_{j}\Vert >E\Vert \widetilde{\beta }_{j}\Vert +\varepsilon \sqrt{4\log \ell _{j}}\right) \\ &\leq &\exp \left( -\frac{4\varepsilon ^{2}\log \ell _{j}}{2}\right) \\ &=&\exp \left( -\log \ell _{j}^{2\varepsilon ^{2}}\right) \\ &=&\frac{1}{\ell _{j}^{2\varepsilon ^{2}}}\rightarrow 0,\,\,\,\,\forall \varepsilon >0\end{aligned}$$ Now for (\[1.3\]), taking $\varepsilon >\frac{1}{\sqrt{2}}$ in the above expression, we obtain summable probabilities, and then by a simple application of the Borel-Cantelli Lemma we have that $${\mathbb{P}} \left( \limsup_{j}\frac{\Vert \widetilde{\beta }_{j}\Vert }{\sqrt{4\log \ell _{j}}}\geq C+\frac{1}{\sqrt{2}}\right) \leq \lim_{j\rightarrow \infty }\sum_{j^{\prime }=j}^{\infty } {\mathbb{P}} \left( \frac{\Vert \widetilde{\beta }_{j^{\prime }}\Vert }{\sqrt{4\log \ell _{j^{\prime }}}}\geq C+\frac{1}{\sqrt{2}}\right) =0.$$ Discretization and Lower Bounds =============================== As explained in the Introduction, this Section is devoted to the proofs for the lower bounds that follow. We have$$\liminf_{j}\frac{{\mathbb{E}}\left\{ \sup_{x\in S^{2}}\widetilde{\beta }_{j}(x)\right\} }{\sqrt{4\log \ell _{j}}}\geq 1. \label{3h30b}$$and$${\mathbb{P}}\left( \liminf_{j}\frac{\left\{ \sup_{x\in S^{2}}\widetilde{\beta }_{j}(x)\right\} }{\sqrt{4\log \ell _{j}}}\geq 1\right) =1. \label{3h30c}$$ We start showing that, for all $\delta >0$, $$\lim_{j}{\mathbb{P}} \left( \frac{\left\{ \sup_{x\in S^{2}}\widetilde{\beta }_{j}\right\} }{\sqrt{4\log \ell _{j}}}>1-\delta \right) =1. \label{3h30a}$$Note first that $\sup \widetilde{\beta }_{j}\geq \sup_{k}\widetilde{\beta }_{j,k}$, where $\{\widetilde{\beta }_{j,k}\}$ is any discrete sample taken from $\widetilde{\beta }_{j}$. Now, let us choose a grid of points such that the distance between them is at least $2^{-j(1-\delta )},$ for some $\delta >0$ - e.g., a $2^{-j(1-\delta )}$-net, see [@bkmpBer]. Note that the vectors $\beta _{j,\cdot }$ and $\widetilde{\beta }_{j,\cdot }$ both have cardinality of order $2^{2j(1-\delta )}$. By using the correlation inequality given in Lemma 10.8 of [@marpecbook], we have $${\mathbb{E}}\widetilde{\beta }_{j,k}\widetilde{\beta }_{j,k^{\prime }}\leq \frac{C}{(1+2^{j\delta })^{M}},$$where $M\in \mathbb{N}$ can be chosen arbitrarily large. The idea of the proof is to approximate these subsampled coefficients by means of a triangular array of Gaussian i.i.d. random variables, say $\widehat{\beta }_{j,k}$. More precisely, let $\Sigma _{j}$ be the covariance matrix of the Gaussian vector $\widetilde{\beta }_{j,\cdot }$; then define $\widehat{\beta }_{j,\cdot }=\Sigma _{j}^{-1/2}\widetilde{\beta }_{j,\cdot }$, which is clearly a vector of i.i.d. Gaussian variables, and let $\lambda _{j,\mathrm{max}}$ and $\lambda _{j,\mathrm{min}}$ be the largest and the smallest eigenvalues of the matrix $\Sigma _{j}$. Then $$\lambda _{j,\mathrm{max}},\lambda _{j,\mathrm{min}}=1+O(\varepsilon _{j}),$$ for a deterministic sequence $\left\{ \varepsilon _{j}\right\} $ which goes to zero faster than any polynomial (nearly exponentially). Indeed $$\begin{aligned} \lambda _{j,\mathrm{max}} &=&\sup_{x}x^{\prime }\Sigma _{j}x=\sup_{x}x^{\prime }\left( \Sigma _{j}-I+I\right) x=\sup_{x}x^{\prime }\left( \Sigma _{j}-I\right) x+1 \\ &\leq &\ell _{j}^{4(1-\delta )}\frac{C_{M}}{1+2^{j\delta M}}+1,\end{aligned}$$where the bound follows crudely from the cardinality of the off-diagonal terms in the matrix. Similarly, $$\begin{aligned} \lambda _{j,\mathrm{min}} &=&\inf_{x}x^{\prime }\Sigma _{j}x=\inf_{x}x^{\prime }\left( \Sigma _{j}-I+I\right) x \\ &=&\inf_{x}x^{\prime }\left( \Sigma _{j}-I\right) x+1\geq 1-\sup |x^{\prime }\left( \Sigma _{j}-I\right) x| \\ &\geq &1-\ell _{j}^{4}\frac{C_{M}}{1+\ell _{j}^{\delta M}}.\end{aligned}$$As a consequence, writing $\Vert \cdot \Vert _{2}$ for the Euclidean inner product in the appropriate dimension we have $$\begin{aligned} {\mathbb{E}}\left( \sup |\widetilde{\beta }_{j,k}-\widehat{\beta }_{j,k}|\right) &\leq &\sqrt{{\mathbb{E}}\Vert \widetilde{\beta }_{j,\cdot }-\widehat{\beta }_{j,\cdot }\Vert _{2}^{2}} \\ &=&\sqrt{{\mathbb{E}}\Vert (I-\Sigma ^{-1/2})\widetilde{\beta }_{j,\cdot }\Vert _{2}^{2}}\leq |1-\lambda _{\mathrm{max}}|\sqrt{E\Vert \widetilde{\beta }_{j,\cdot }\Vert _{2}^{2}} \\ &\leq &\ell _{j}^{4(1-\delta )}\frac{C_{M}}{1+\ell _{j}^{\delta M}}\cdot \ell _{j}^{2}=\ell _{j}^{2+4(1-\delta )}\frac{C_{M}}{1+\ell _{j}^{\delta M}}=O(\ell _{j}^{6-\delta M}).\end{aligned}$$We can now exploit a classical result by Berman ([@berman]) to conclude that $${\mathbb{P}} \left( \left|\frac{\sup_{k}\widetilde{\beta }_{j,k}}{\sqrt{4\log \ell _{j}}}-1\right|>\varepsilon \right) \rightarrow 0,$$as $j\rightarrow \infty $, for all $\varepsilon >0$. Thus (\[3h30a\]) is established; (\[3h30b\]) follows immediately, given that $\delta $ is arbitrary. To establish (\[3h30c\]), we use again the Borel-Cantelli Lemma, so that we need to prove that, for all $\varepsilon >0$ $$\sum_{j} {\mathbb{P}} \left( \frac{\left\{ \sup_{x\in S^{2}}\widetilde{\beta }_{j}\right\} }{\sqrt{4\log \ell _{j}}}<1-\varepsilon \right) <\infty .$$Clearly $${\mathbb{P}} \left( \frac{\left\{ \sup_{x\in S^{2}}\widetilde{\beta }_{j}\right\} }{\sqrt{4\log \ell _{j}}}<1-\varepsilon \right) \leq {\mathbb{P}} \left( \frac{\sup_{k}\widetilde{\beta }_{j,k}}{\sqrt{4\log \ell _{j}}}<1-\varepsilon \right) , \text{ for all }j,$$whence it suffices to prove that $$\sum_{j}{\mathbb{P}} \left( \frac{\sup_{k}\widetilde{\beta }_{j,k}}{\sqrt{4\log \ell _{j}}}<1-\varepsilon \right) <\infty .$$Now$${\mathbb{P}} \left( \frac{\sup_{k}\widetilde{\beta }_{j,k}}{\sqrt{4\log \ell _{j}}}<1-\varepsilon \right) ={\mathbb{P}} \left( \frac{\sup_{k}\left( \widehat{\beta }_{j,k}-\widehat{\beta }_{j,k}+\widetilde{\beta }_{j,k}\right) }{\sqrt{4\log \ell _{j}}}<1-\varepsilon \right)$$$$\begin{aligned} &\leq &{\mathbb{P}} \left( \frac{\sup_{k}\widehat{\beta }_{j,k}-\sup_{k}\left( \widehat{\beta }_{j,k}-\widetilde{\beta }_{j,k}\right) }{\sqrt{4\log \ell _{j}}}<1-\varepsilon \right) \\ &\leq &{\mathbb{P}} \left( \frac{\sup_{k}\widehat{\beta }_{j,k}-\sup_{k}\left\vert \widehat{\beta }_{j,k}-\widetilde{\beta }_{j,k}\right\vert }{\sqrt{4\log \ell _{j}}}<1-\varepsilon \right) \\ &=&{\mathbb{P}} \left( \frac{\sup_{k}\widehat{\beta }_{j,k}}{\sqrt{4\log \ell _{j}}}<1-\varepsilon +\frac{\sup_{k}\left\vert \widehat{\beta }_{j,k}-\widetilde{\beta }_{j,k}\right\vert }{\sqrt{4\log \ell _{j}}}\right) \\ &\leq &{\mathbb{P}} \left( \frac{\sup_{k}\widehat{\beta }_{j,k}}{\sqrt{4\log \ell _{j}}}<1-\frac{\varepsilon }{2}\right) +{\mathbb{P}} \left( \frac{\sup_{k}\left\vert \widehat{\beta }_{j,k}-\widetilde{\beta }_{j,k}\right\vert }{\sqrt{4\log \ell _{j}}}>\frac{\varepsilon }{2}\right) \\ &\leq &{\mathbb{P}} \left( \frac{\sup_{k}\widehat{\beta }_{j,k}}{\sqrt{4\log \ell _{j}}}<1-\frac{\varepsilon }{2}\right) +O(\ell _{j}^{6-\varepsilon M}).\end{aligned}$$The second term above is clearly summable, for all fixed $\varepsilon >0,$ by simply taking $M$ large enough. To check summability of the first term we write$${\mathbb{P}} \left( \frac{\sup_{k}\widehat{\beta }_{j,k}}{\sqrt{4\log \ell _{j}}}<1-\frac{\varepsilon }{2}\right) =\prod\limits_{k}{\mathbb{P}} \left( \frac{\widehat{\beta }_{j,k}}{\sqrt{4\log \ell _{j}}}<1-\frac{\varepsilon }{2}\right)$$ $$\begin{aligned} &=&\left( {\mathbb{P}} \left( \widehat{\beta }_{j,1}<(1-\frac{\varepsilon }{2})\sqrt{4\log \ell _{j}}\right) \right) ^{\ell _{j}^{2}} \\ &=&\left( 1-{\mathbb{P}} \left( \widehat{\beta }_{j,1}>(1-\frac{\varepsilon }{2})\sqrt{4\log \ell _{j}}\right) \right) ^{\ell _{j}^{2}} \\ &\leq &\left( 1-\frac{1}{(1-\frac{\varepsilon }{2})\sqrt{4\log \ell _{j}}}\left( 1-\frac{1}{(1-\frac{\varepsilon }{2})\sqrt{4\log \ell _{j}}}\right) \cdot \frac{1}{\ell _{j}^{2(1-\frac{\varepsilon }{2})^{2}}}\right) ^{\ell _{j}^{2}} \\ &\leq &\left( 1-\frac{1}{2(1-\frac{\varepsilon }{2})\sqrt{4\log \ell _{j}}}\cdot \frac{1}{\ell _{j}^{2(1-\frac{\varepsilon }{2})^{2}}}\right) ^{\ell _{j}^{2}},\end{aligned}$$where we have used Mill’s inequality for standard Gaussian variables, ${\mathbb{P}}\left\{ Z>z\right\} \geq \frac{z}{1+z^{2}}\phi (z).$ Since $(1-\frac{\varepsilon }{2})^{2}<1$, this term decays exponentially, and it is hence summable. The proof of (\[3h30c\]) is hence concluded. [99]{} **Adler, R. J. and Taylor, J. E.**, **(2007)** *Random Fields and Geometry*, Springer. **Azaïs, J.-M., Wschebor, M. (2005)** On the Distribution of the Maximum of a Gaussian Field with d Parameters. *Ann. Appl. Probab.* 15 no. 1A, 254–278. **Azaïs, J.-M., Wschebor, M. (2009)** Level Sets and Extrema of Random Processes and Fields. John Wiley & Sons, Inc., Hoboken, NJ. **Baldi, P., Kerkyacharian, G., Marinucci, D. and Picard, D. (2009)** Asymptotics for Spherical Needlets, *Annals of Statistics,* Vol. 37, No. 3, 1150-1171 **Baldi, P., Kerkyacharian, G., Marinucci, D. and Picard, D. (2009)** Subsampling Needlet Coefficients on the Sphere, *Bernoulli*,* *Vol. 15, 438-463 **Bennett, C.L. et al. (2012)** Nine-Year WMAP Observations: Final Maps and Results, arXiv:1212.5225 **Berman, S. (1962)** A Law of Large Numbers for the Maximum in a Stationary Gaussian Sequence, *Ann.Math.Stat.*, 33, 1, 93-97 **Cammarota, V., Marinucci, D. (2014)** On the Limiting Behaviour of Needlets Polyspectra, *Ann.Inst.H.Poinc.,* in press, arXiv:1307.4691 **Cheng, D. and Xiao, Y. (2012)** The Mean Euler Characteristic and Excursion Probability of Gaussian Random Fields with Stationary Increments, arXiv:1211.6693 **Cheng, D. and Schwartzman, A. (2013)** Distribution of the Height of Local Maxima of Gaussian Random Fields, arXiv:1307.5863 **Dodelson, S. (2003)** *Modern Cosmology*, Academic Press **Durrer, R. (2008)** *The Cosmic Microwave Background,* Cambridge University Press. **Malyarenko, A. (2012)**, *Invariant Random Fields on Spaces with a Group Action*, Probability and its Applications, Springer. **Marinucci, D. and Peccati, G. (2011)** *Random Fields on the Sphere. Representation, Limit Theorem and Cosmological Applications*, Cambridge University Press **Marinucci, D. and Peccati, G. (2012)** Mean Square Continuity on Homogeneous Spaces of Compact Groups, arXiv:1210.7676. **Marinucci, D. and Vadlamani, S. (2013)** High-Frequency Asymptotics for Lipscitz-Killing Curvatures of Excursion Sets on the Sphere, arXiv:1303.2456 **Narcowich, F.J., Petrushev, P. and Ward, J.D. (2006a)** Localized Tight Frames on Spheres, *SIAM Journal of Mathematical Analysis* Vol. 38, pp. 574–594 **Pietrobon, D., Amblard, A., Balbi, A., Cabella, P., Cooray, A., Marinucci, D. (2008)** Needlet Detection of Features in WMAP CMB Sky and the Impact on Anisotropies and Hemispherical Asymmetries, *Physical Review D*, D78 103504 **Stein, E.M. and Weiss, G. (1971)** *Introduction to Fourier Analysis on Euclidean Spaces.* Princeton University Press **Szego, G. (1975)** Orthogonal Polynomials, 4th edition, *American Mathematical Society, Colloquium Publications,* Vol. XXIII **Taylor, J.E. and Adler, R.J. (2009)** Gaussian Processes, Kinematic Formulae and Poincaré’s Limit. *Ann. Probab.* 37, no. 4, 1459–1482. **Taylor, J.E. and Vadlamani, S. (2013)** Random Fields and the Geometry of Wiener Space, *Ann. Probab.,* 41, 4, 2724-2754, arXiv: 1105.3839 **Zelditch, S. (2009)** Real and complex zeros of Riemannian random waves, *Contemp. Math.,* 484, 321–342 [^1]: Corresponding author; email address [email protected]. Research supported by the ERC Grants n. 277742 *Pascal,* “Probabilistic and Statistical Techniques for Cosmological Applications”.

ArXiv

[**On characterizations of $\mathcal{MT}(\lambda)$-functions**]{}\ \[0.2in\] **Wei-Shih Du**[^1] [Department of Mathematics, National Kaohsiung Normal University, Kaohsiung 82444, Taiwan]{} ------------------------------------------------------------------------ **Abstract:** In this paper, we introduce and share the new concept of $\mathcal{MT}(\lambda )$-functions and its some characterizations. **2010 Mathematics Subject Classification:** 26D07, 54C30. **Key words and phrases:** $\mathcal{MT}$-function (or $\mathcal{R}$-function), eventually nonincreasing sequence, eventually strictly decreasing sequence, $\mathcal{MT}(\lambda )$-function.\ ------------------------------------------------------------------------ [**1. Introduction and preliminaries**]{} Let $f$ be a real-valued function defined on $\mathbb{R}$. For $c\in \mathbb{R}$, we recall $$\limsup_{x\rightarrow c^{+}}f(x)=\inf_{\varepsilon >0}\sup_{c<x<c+\varepsilon }f(x)\text{.}$$ **Definition 1.1 (see \[1-6\]).**A function $\varphi :$ $[0,\infty )\rightarrow $ $[0,1)$ is said to be an $\mathcal{MT}$-$function$ (or $\mathcal{R}$-$function$) if $\limsup\limits_{s\rightarrow t^{+}}\varphi (s)<1$ for all $t\in \lbrack 0,\infty )$. It is obvious that if $\varphi :$ $[0,\infty )\rightarrow $ $[0,1)$ is a nondecreasing function or a nonincreasing function, then $\varphi $ is an $\mathcal{MT}$-function. So the set of $\mathcal{MT}$-functions is a rich class. However, it is worth to note that there exist functions which are not $\mathcal{MT}$-functions. **Example 1.2 (see \[2\]).**Let $\varphi :$ $[0,\infty )\rightarrow $ $[0,1)$ be defined by $$\quad \varphi (t):=\left\{ \begin{array}{cc} \frac{\sin t}{t} & \text{, if }t\in (0,\frac{\pi }{2}]\ \\ 0 & \text{, otherwise}.\end{array}\right.$$Since $\limsup\limits_{s\rightarrow 0^{^{+}}}\varphi (s)=1,$ $\varphi $ is not an $\mathcal{MT}$-function. Some characterizations of $\mathcal{MT}$-functions were established by Du \[2, Theorem 2.1\] and were applied in fixed point theory; for more detail, one can refer to \[1-6\] and references therein. In this paper, we introduce and share Du’s recent work on the new concept of $\mathcal{MT}(\lambda )$-functions and its some characterizations. [**2. New concept of $\mathcal{MT}(\lambda )$-functions and its characterizations**]{} Recall that a real sequence $\{a_{n}\}_{n\in \mathbb{N} }$ is called 1. *eventually strictly decreasing* if there exists $\ell \in \mathbb{N} $ such that $a_{n+1}<a_{n}$ for all $n\in \mathbb{N} $ with $n\geq \ell $; 2. *eventually strictly increasing* if there exists $\ell \in \mathbb{N} $ such that $a_{n+1}>a_{n}$ for all $n\in \mathbb{N} $ with $n\geq \ell $; 3. *eventually nonincreasing* if there exists $\ell \in \mathbb{N} $ such that $a_{n+1}\leq a_{n}$ for all $n\in \mathbb{N} $ with $n\geq \ell $; 4. *eventually nondecreasing* if there exists $\ell \in \mathbb{N} $ such that $a_{n+1}\geq a_{n}$ for all $n\in \mathbb{N} $ with $n\geq \ell $. Very recently, Du \[7\] proved some new characterizations of $\mathcal{MT}$-functions as follows. We give the proof for the sake of completeness and for the readers’ convenience. **Theorem 2.1 (see \[7\]).***Let* $\varphi :$* *$[0,\infty )\rightarrow $* *$[0,1)$* be a function. Then the following statements are equivalent.* 1. $\varphi $* is an* $\mathcal{MT}$*-function.* 2. *For each* $t\in \lbrack 0,\infty )$*, there exist* $r_{t}^{(1)}\in \lbrack 0,1)$* and* $\varepsilon _{t}^{(1)}>0$* such that* $\varphi (s)\leq r_{t}^{(1)}$* for all* $s\in (t,t+\varepsilon _{t}^{(1)})$*.* 3. *For each* $t\in \lbrack 0,\infty )$*, there exist* $r_{t}^{(2)}\in \lbrack 0,1)$* and* $\varepsilon _{t}^{(2)}>0$* such that* $\varphi (s)\leq r_{t}^{(2)}$* for all* $s\in \lbrack t,t+\varepsilon _{t}^{(2)}]$*.* 4. *For each* $t\in \lbrack 0,\infty )$*, there exist* $r_{t}^{(3)}\in \lbrack 0,1)$* and* $\varepsilon _{t}^{(3)}>0$* such that* $\varphi (s)\leq r_{t}^{(3)}$* for all* $s\in (t,t+\varepsilon _{t}^{(3)}]$*.* 5. *For each* $t\in \lbrack 0,\infty )$*, there exist* $r_{t}^{(4)}\in \lbrack 0,1)$* and* $\varepsilon _{t}^{(4)}>0$* such that* $\varphi (s)\leq r_{t}^{(4)}$* for all* $s\in \lbrack t,t+\varepsilon _{t}^{(4)})$*.* 6. *For any nonincreasing sequence* $\{x_{n}\}_{n\in \mathbb{N} }$* in* $[0,\infty )$*, we have* $0\leq \sup\limits_{n\in \mathbb{N} }\varphi (x_{n})<1$*.* 7. $\varphi $* is a function of contractive factor; that is, for any strictly decreasing sequence* $\{x_{n}\}_{n\in \mathbb{N} }$* in* $[0,\infty )$*, we have* $0\leq \sup\limits_{n\in \mathbb{N} }\varphi (x_{n})<1$*.* 8. *For any eventually nonincreasing sequence* $\{x_{n}\}_{n\in \mathbb{N} }$* in* $[0,\infty )$*, we have* $0\leq \sup\limits_{n\in \mathbb{N} }\varphi (x_{n})<1$*.* 9. *For any eventually strictly decreasing sequence* $\{x_{n}\}_{n\in \mathbb{N} }$* in* $[0,\infty )$*, we have* $0\leq \sup\limits_{n\in \mathbb{N} }\varphi (x_{n})<1$*.* **Proof.**The equivalence of statements (a)-(g) was indeed proved in \[2, Theorem 2.1\]. The implications “(h) $\Rightarrow $ (f)” and “(i) $\Rightarrow $ (g)” are obvious. Let us prove “(f) $\Rightarrow $ (h)”. Suppose that (f) holds. Let $\{x_{n}\}_{n\in \mathbb{N} }$ be an eventually nonincreasing sequence in $[0,\infty )$. Then there exists $\ell \in \mathbb{N} $ such that $x_{n+1}\leq x_{n}$ for all $n\in \mathbb{N} $ with $n\geq \ell $. Put $y_{n}=x_{n+\ell -1}$ for $n\in \mathbb{N} $. So $\{y_{n}\}_{n\in \mathbb{N} }$ is a nonincreasing sequence in $[0,\infty )$. By (f), we obtain $$0\leq \sup\limits_{n\in \mathbb{N} }\varphi (y_{n})<1\text{.}$$Let $\gamma :=\sup\limits_{n\in \mathbb{N} }\varphi (y_{n})$. Then $$0\leq \varphi (x_{n+\ell -1})=\varphi (y_{n})\leq \gamma <1\text{ \ \ for all }n\in \mathbb{N} \text{.}$$Hence we get $$0\leq \varphi (x_{n})\leq \gamma <1\text{ \ \ for all }n\in \mathbb{N} \text{ with }n\geq \ell \text{.}$$Let $$\eta :=\max \{\varphi (x_{1}),\varphi (x_{2}),\cdots ,\varphi (x_{\ell -1}),\gamma \}<1\text{.}$$Then $\varphi (x_{n})\leq \eta $ for all $n\in \mathbb{N} $. Hence $0\leq \sup\limits_{n\in \mathbb{N} }\varphi (x_{n})\leq \eta <1$ and (h) holds. Similarly, we can varify “(g) $\Rightarrow $ (i)”. Therefore, from above, we prove that the statements* *(a)-(i) are all logically equivalent. The proof is completed.$\Box $ In \[7\], Du first introduced the concept of $\mathcal{MT}(\lambda )$-functions. **Definition 2.2 (see \[7\]).**Let $\lambda \in $ $(0,1]$. A function $\mu :$ $[0,\infty )\rightarrow $ $[0,\lambda )$ is said to be an $\mathcal{MT}(\lambda )$-$function$ if $\limsup\limits_{s\rightarrow t^{+}}\mu (s)<\lambda $ for all $t\in \lbrack 0,\infty )$. Clearly, an $\mathcal{MT}$-function is an $\mathcal{MT}(1)$-function. It is quite obvious that $\mu $ is an $\mathcal{MT}(\lambda )$-function if and only if $\lambda ^{-1}\mu $ is an $\mathcal{MT}$-function$.$ **Remark 2.3 (see \[8\]).**Recall that a function $\varphi :$ $[0,\infty )\rightarrow \left[ 0,\frac{1}{2}\right) $ is said to be a $\mathcal{P}$-$function$ \[8\] if $\limsup\limits_{s\rightarrow t^{+}}\mu (s)<\frac{1}{2}$ for all $t\in \lbrack 0,\infty )$. So a $\mathcal{P}$-function is obviously an $\mathcal{MT}\left( \frac{1}{2}\right) $-function. The following characterizations of $\mathcal{MT}(\lambda)$-functions is an immediate consequence of Theorem 2.1. **Theorem 2.4 (see \[7\]).***Let* $\lambda \in $* *$(0,1]$* and let* $\mu :$* *$[0,\infty )\rightarrow $* *$[0,\lambda )$* be a function. Then the following statements are equivalent.* 1. $\mu $* is an* $\mathcal{MT}(\lambda )$*-function.* 2. $\lambda ^{-1}\mu $* is an* $\mathcal{MT}$*-function.* 3. *For each* $t\in \lbrack 0,\infty )$*, there exist* $\xi _{t}^{(1)}\in \lbrack 0,\lambda )$* and* $\epsilon _{t}^{(1)}>0$* such that* $\mu (s)\leq \xi _{t}^{(1)}$* for all* $s\in (t,t+\epsilon _{t}^{(1)})$*.* 4. *For each* $t\in \lbrack 0,\infty )$*, there exist* $\xi _{t}^{(2)}\in \lbrack 0,\lambda )$* and* $\epsilon _{t}^{(2)}>0$* such that* $\mu (s)\leq \xi _{t}^{(2)}$* for all* $s\in \lbrack t,t+\epsilon _{t}^{(2)}]$*.* 5. *For each* $t\in \lbrack 0,\infty )$*, there exist* $\xi _{t}^{(3)}\in \lbrack 0,\lambda )$* and* $\epsilon _{t}^{(3)}>0$* such that* $\mu (s)\leq \xi _{t}^{(3)}$* for all* $s\in (t,t+\epsilon _{t}^{(3)}]$*.* 6. *For each* $t\in \lbrack 0,\infty )$*, there exist* $\xi _{t}^{(4)}\in \lbrack 0,\lambda )$* and* $\epsilon _{t}^{(4)}>0$* such that* $\mu (s)\leq \xi _{t}^{(4)}$* for all* $s\in \lbrack t,t+\epsilon _{t}^{(4)})$*.* 7. *For any nonincreasing sequence* $\{x_{n}\}_{n\in \mathbb{N} }$* in* $[0,\infty )$*, we have* $0\leq \sup\limits_{n\in \mathbb{N} }\mu (x_{n})<\lambda $*.* 8. *For any strictly decreasing sequence* $\{x_{n}\}_{n\in \mathbb{N} }$* in* $[0,\infty )$*, we have* $0\leq \sup\limits_{n\in \mathbb{N} }\mu (x_{n})<\lambda $*.* 9. *For any eventually nonincreasing sequence* $\{x_{n}\}_{n\in \mathbb{N} }$* in* $[0,\infty )$*, we have* $0\leq \sup\limits_{n\in \mathbb{N} }\mu (x_{n})<\lambda $*.* 10. *For any eventually strictly decreasing sequence* $\{x_{n}\}_{n\in \mathbb{N} }$* in* $[0,\infty )$*, we have* $0\leq \sup\limits_{n\in \mathbb{N} }\mu (x_{n})<\lambda $*.* **Remark 2.5.**8, Lemma 3.1\] is a special case of Theorem 2.4 for $\lambda =\frac{1}{2}$. [**Acknowledgments**]{} This research was supported by Grant No. MOST 103-2115-M-017-001 of the Ministry of Science and Technology of the Republic of China. [**References**]{} 1. W.-S. Du, Some new results and generalizations in metric fixed point theory, Nonlinear Anal. 73 (2010) 1439-1446. 2. W.-S. Du, On coincidence point and fixed point theorems for nonlinear multivalued maps, Topology and its Applications 159 (2012) 49-56. 3. W.-S. Du, H. Lakzian, Nonlinear conditions for the existence of best proximity points, Journal of Inequalities and Applications, 2012, 2012:206. 4. W.-S. Du, On Caristi type maps and generalized distances with applications, Abstract and Applied Analysis, 2013, Volume 2013, Article ID 407219, 8 pages, http://dx.doi.org/10.1155/2013/407219. 5. W.-S. Du, E. Karapinar, A note on Caristi-type cyclic maps: related results and applications, Fixed Point Theory and Applications, 2013, 2013:344. 6. W.-S. Du, F. Khojasteh, Y.-N. Chiu, Some generalizations of Mizoguchi-Takahashi’s fixed point theorem with new local constraints, Fixed Point Theory and Applications, 2014, 2014:31. 7. W.-S. Du, New existence results of best proximity points and fixed points for $\mathcal{MT}(\lambda )$-functions, submitted. 8. H.K. Pathak, R.P. Agarwal, Y.J. Cho, Coincidence and fixed points for multi-valued mappings and its application to nonconvex integral inclusions, Journal of Computational and Applied Mathematics 283 (2015) 201-217. [^1]: [E-mail address: [email protected], [email protected]; Tel: +886-7-7172930 ext 6809; Fax: +886-7-6051061.]{}

ArXiv

--- author: - | Paulo Pires Pacheco\ Department of Physics, Imperial College London\ London, SW7 2AZ, UK - | Daniel Waldram\ Department of Physics, Imperial College London\ London, SW7 2AZ, UK\ Institute for Mathematical Sciences, Imperial College London\ London, SW7 2AZ, UK title: 'M-theory, exceptional generalised geometry and superpotentials' --- Introduction {#sec:intro} ============ Type II string backgrounds in $d$ dimensions which include non-trivial fluxes have a natural description in terms of Hitchin’s generalised geometry [@GCY; @Gualtieri; @H-GK; @H-brack], where the metric and NS–NS $B$-field are combined into a single geometrical object, transforming under $O(d,d)$. This description has proved very useful in, among other things: characterising supersymmetric backgrounds, finding new examples with non-zero fluxes and writing supersymmetric low-energy effective theories [@GLMW]–[@BFMMPZ]; describing topological string theories and generic $N=(2,2)$ $\sigma$-models [@Kap]–[@LSSW]; as well as motivating the structure of non-geometrical backgrounds [@Tfold]–[@Hull:2007jy]. The aim of this paper is to understand some details of how similar constructions based on the exceptional groups $E_{d(d)}$ can be used to describe M-theory, or more precisely, eleven-dimensional supergravity backgrounds. The general form of such constructions, as well as those arising from type II theories, has been described recently by Hull [@chris]. That work was partially motivated by the existence of so-called “non-geometrical” backgrounds, which appear consistent as string theory vacua and are typically dual to supergravity backgrounds, but do not themselves have a consistent global description in supergravity. This suggested considering extensions of generalised geometry based on generic $O(d,d)$ or $E_{d(d)}$ bundles, whereas only a subclass of such bundles arise in supergravity. The corresponding geometry was generically dubbed “extended” (or more specifically “M-geometry” for the generalisation of eleven-dimensional supergravity). Here, however, we will consider only supergravity backgrounds and concentrate on the physically important example of seven-dimensional backgrounds and hence the group $\operatorname{\mathit{E}_{7(7)}}$. There are two parts to the analysis. First to build the analog of the generalised geometry and then to describe the geometrical objects that characterise supersymmetric $N=1$ backgrounds in four dimensions. In analogy to the type II analysis in [@GLW1; @GLW2], we consider the particular application of writing the $N=1$ superpotential in a generic $\operatorname{\mathit{E}_{7(7)}}$-invariant form. The essential idea of the construction is that $O(d,d)$ symmetries of generalised $d$-dimensional geometry, which in string theory are related to T-duality symmetries, should be replaced by the U-duality exceptional symmetry groups $E_{d(d)}$ [@CJ; @Udual; @HT]. Since the U-duality connects all the bosonic degrees of freedom of eleven-dimensional supergravity, or for type II theories, both NS–NS and R–R degrees of freedom, this extension should provide a geometrisation of generic flux backgrounds. The fact that the full eleven-dimensional supergravity could be reformulated in terms of $E_{d(d)}$ objects was first pointed out by de Wit and Nicolai [@deWN] (for $\operatorname{\mathit{E}_{7(7)}}$) and then by Nicolai [@Nic] and Koepsell, Nicolai and Samtleben [@KNS] (for $E_{8(8)}$), the latter calling the construction “exceptional geometry”. Motivated by these authors’ and Hull’s nomenclature [@chris], we will refer to the variant of Hitchin’s generalised geometry relevant to eleven-dimensional (and type II) supergravity as EGG for “exceptional generalised geometry”. This reserves “extended” and “M” for generic M-theory backgrounds, potentially including non-geometrical examples. We should note here that recently there have also been more ambitious related proposals connecting infinite-dimensional exceptional algebras to supergravity. The original proposal of [@West1; @West2] had the goal of giving an eleven-dimensional covariant formulation of M-theory with $E_{11}$ invariance, as well as of lower-dimensional gauged supergravities [@RW], while the work of [@Nicolai] describes a gauged-fixed version of the supergravity dynamics in terms of an explicit $E_{10}$ coset construction. The paper is arranged as follows. In section \[sec:EGG\], after briefly reviewing generalised geometry, we describe how the analogous $\operatorname{\mathit{E}_{7(7)}}$-invariant EGG can be defined for a seven-dimensional manifold $M$. Much of this analysis appears in [@chris]. The corresponding exceptional generalised tangent space (EGT) encodes all the topological information of the conventional tangent space $TM$ as well as the topology of the “gerbe”, the analogue of a $U(1)$-bundle, on which the supergravity three-form $A$ is a connection. We give new details on how the tangent space bundle embeds into the EGG, as well as the precise form of the gerbes. We also define the analogue of the Courant bracket on the EGT which encodes the differential geometry. We then introduce the notion of an ${{\operatorname{\mathit{SU}}(8)/{\mathbb{Z}}_2}}$ structure on the EGT and show that this encodes supergravity metric $g$ and three-form $A$. (This is equivalent to the result [@chris], familiar from toroidal reductions [@CJ], that they parameterize an element of the coset space $\operatorname{\mathit{E}_{7(7)}}/({{\operatorname{\mathit{SU}}(8)/{\mathbb{Z}}_2}})$.) We show in particular how the non-linear Chern–Simons-like terms for $A$ appear naturally in this formalism. In section \[sec:SEGG\], we again start with review. We discuss the pure $O(6,6)$ spinors $\Phi^\pm$ which characterise generalised geometrical six-dimensional type II backgrounds preserving $N=2$ supersymmetry in four dimensions, and define an ${{\operatorname{\mathit{SU}}(3)\times\operatorname{\mathit{SU}}(3)}}$ structure on the generalised tangent space. We then make the analogous analysis for EGG. (A short related discussion of supersymmetric backgrounds appeared in section 8 of [@chris].) Without flux, for the background to have $N=1$ supersymmetry in four dimensions, the compactification manifold must have a $G_2$-structure. We show that the generalisation to EGG is that the EGT must admit an $\operatorname{\mathit{SU}}(7)$ structure. This is equivalent to the existence of a nowhere vanishing tensor $\phi$ transforming in the ${\mathbf{912}}$ of $\operatorname{\mathit{E}_{7(7)}}$. It is not a generic element but, to be stabilized by $\operatorname{\mathit{SU}}(7)$ must lie in a particular orbit under $\operatorname{\mathit{E}_{7(7)}}$. In section \[sec:W\] we give an application of the these results. We first derive the form of the superpotential for a generic $N=1$ compactification of eleven-dimensional supergravity slightly extending the results of [@HM]. We then use the structure $\phi$, which plays the role of the chiral scalar fields in the four-dimensional theory, to show that the superpotential can be rewritten in a $\operatorname{\mathit{E}_{7(7)}}$-invariant form. We conclude in section \[sec:concl\] with a brief discussion of our results and of possible further work. EGG in seven dimensions {#sec:EGG} ======================= In this section we discuss the structure of exceptional generalised geometry, focusing on the case of seven dimensions. Much of this analysis also appears in [@chris]. Here we approach the construction from the point of view of geometrical structures and include some new details on the gerbe topology and tangent space embedding, as well the definition of the analogue of the Courant bracket. Our particular motivation will be compactification of eleven-dimensional supergravity to a four-dimensional effective theory. However here we will first simply introduce the EGG formalism and leave the more detailed connection to supergravity to the following section. We start by reviewing the structure of generalised geometry [@GCY; @Gualtieri; @H-GK; @H-brack] since the EGG structure arises in a very analogous way. Review of generalised geometry {#sec:revGG} ------------------------------ In defining generalised geometry [@GCY; @Gualtieri; @H-GK; @H-brack] on a $d$-dimensional manifold, one starts with the (untwisted) generalised tangent space $E_0=TM\oplus TM^*$. Let us denote elements of $E_0$ by $X=x+\xi$. There is then a natural $O(d,d)$-invariant metric ${\eta}$ on $E$ given by $${\eta}(X,X) := i_x\xi .$$ In particular, ${\eta}$ is invariant under the $\operatorname{\mathit{GL}}(d)\subset O(d,d)$ action on the fibres of $TM$ and $T^*M$. The metric is also invariant under so called “$B$-shifts” where for any $B\in\Lambda^2T^*M$ we define $$\label{eq:B-shift} {\mathrm{e}}^B X := x + \left(\xi-i_xB\right) .$$ The differential geometry of the generalised tangent space is encoded in the Courant bracket which generalises the Lie bracket between two vectors. It is defined by $$\label{eq:Courant} [ x+\xi, y+\zeta] = [x,y] + \mathcal{L}_x\zeta - \mathcal{L}_y\xi - \tfrac{1}{2}{\mathrm{d}}\left(i_x\zeta - i_y\xi\right) ,$$ where $[x,y]$ is the usual Lie bracket of vector fields and $\mathcal{L}_x$ is the Lie derivative, so $\mathcal{L}_x\alpha=i_x{\mathrm{d}}\alpha+{\mathrm{d}}i_x\alpha$ for any form $\alpha$. Note that its automorphism group is not the full group of local $O(d,d)$ transformations but only the subgroup generated by diffeomorphisms and the $B$-shifts  (with ${\mathrm{d}}B=0$). Formally we can write this as the semidirect product $\operatorname{Diff}(M)\ltimes\Omega^2_{\text{closed}}(M)$. The usefulness of generalised geometry in string theory and supergravity is that the ordinary metric $g$ and NS–NS two-form field $B$ combine naturally into a single object, the so-called generalised metric $G$ [@Gualtieri]. This is an $O(2d)$-invariant metric which is compatible with ${\eta}$, that is ${\eta}^{-1}G{\eta}^{-1}=G^{-1}$. If we split into $TM$ and $T^*M$, we can write the $O(d,d)$ metric as the matrix $${\eta}= \frac{1}{2}\begin{pmatrix} 0 & {\mathbb{1}}\\ {\mathbb{1}}& 0 \end{pmatrix} , \qquad \text{where} \qquad X = \begin{pmatrix} x \\ \xi \end{pmatrix} .$$ The generic generalised metric can then be written as $$\label{eq:gen-G} G = \frac{1}{2}\begin{pmatrix} g - Bg^{-1}B & Bg^{-1} \\ - g^{-1}B & g^{-1} \end{pmatrix} ,$$ where $g$ is an ordinary Riemannian metric, and $B\in\Lambda^2T^*M$ is the NS-NS two-form. Note that one can also define $G$ as a product structure $\Pi$. Writing $\Pi={\eta}^{-1}G$ one has $\Pi^2={\mathbb{1}}$, and in addition ${\eta}(X,Y)={\eta}(\Pi X,\Pi Y)$. Hence $\Pi$ is a product structure, compatible with ${\eta}$, and the projections $\frac{1}{2}({\mathbb{1}}\pm\Pi)$ project onto two $d$-dimensional subspaces $C^\pm$, such that $E=C^+\oplus C^-$. From this perspective, given ${\eta}$, one sees that $G$ defines an $O(d)\times O(d)$ structure, and ${\eta}$ and $G$ decompose into the separate metrics on $C^+$ and $C^-$. One can view $g$ and $B$ as parametrising the coset space $O(d,d)/O(d)\times O(d)$. One can also write $$\label{eq:B-shiftG} G(X,Y) = G_0({\mathrm{e}}^B X, {\mathrm{e}}^B Y)$$ where $$G_0 = \frac{1}{2} \begin{pmatrix} g & 0 \\ 0 & g^{-1} \end{pmatrix} .$$ If $B$ leads to a non-trivial flux $H$ it is only locally defined as a two-form. Globally one must patch by gauge transformations, so on the overlap $U_{(\alpha)}\cap U_{(\beta)}$ one has $$B_{(\alpha)} - B_{(\beta)} = {\mathrm{d}}\Lambda_{(\alpha\beta)} ,$$ (so $\Lambda_{(\alpha\beta)}=-\Lambda_{(\beta\alpha)}$) while on the triple overlap $U_{(\alpha)}\cap U_{(\beta)}\cap U_{(\gamma)}$ $$\Lambda_{(\alpha\beta)} + \Lambda_{(\beta\gamma)} + \Lambda_{(\gamma\alpha)} = {\mathrm{d}}\Lambda_{(\alpha\beta\gamma)} ,$$ (implying $\Lambda_{(\alpha\beta\gamma)}=-\Lambda_{(\alpha\gamma\beta)}$ etc.). Mathematically this means $B$ is a connection on a gerbe (see for instance [@gerbes]). If the flux is quantised $H\in H^3(M,{\mathbb{Z}})$ then one has $g_{(\alpha\beta\gamma)}={\mathrm{e}}^{{\mathrm{i}}\Lambda_{(\alpha\beta\gamma)}}\in U(1)$ and these elements satisfy a cocycle condition on $U_{(\alpha)}\cap U_{(\beta)}\cap U_{(\gamma)}\cap U_{(\delta)}$ $$g_{(\beta\gamma\delta)}g^{-1}_{(\alpha\gamma\delta)} g_{(\alpha\beta\delta)}g^{-1}_{(\alpha\beta\gamma)} = 1 .$$ Formally the $g_{(\alpha\beta\gamma)}$ define the gerbe, while the $\Lambda_{(\alpha\beta)}$ define a “connective structure” on the gerbe. Together they encode the analogue of the topological data of a $U(1)$ gauge bundle. If $B$ is non-trivial, the form of the generalised metric  means that $G$ cannot really be an inner product on sections of $E_0=TM\oplus T^*M$. Instead, the generalised vectors must be sections of an extension $E$ $$\label{eq:gen-T} 0 \longrightarrow T^*M \longrightarrow E \longrightarrow TM \longrightarrow 0 ,$$ where on the intersection of two patches $U_{\alpha}\cap U_{\beta}$ one identifies $X_{(\alpha)}={\mathrm{e}}^{-{\mathrm{d}}{\Lambda_{(\alpha\beta)}}}X_{(\beta)}$, or in components $$\label{eq:patch} x_{(\alpha)} + \xi_{(\alpha)} = x_{(\beta)} + ( \xi_{(\beta)} + i_{x_{(\beta)}}{\mathrm{d}}\Lambda_{(\alpha\beta)} ) .$$ Here $\Lambda_{(\alpha\beta)}$ is the same one-form that appears in defining the connection $B$ (though of course it is independent of the particular choice of $B$). Thus we see that $E$ encodes both the topological structure of the tangent space $TM$ and the connective structure of the gerbe. Note that the form of the twisting  is such that, since the Courant bracket is $B$-shift invariant (when $B$ is closed), it can still be defined on sections of the twisted $E$. The exceptional generalised tangent space and $\operatorname{\mathit{E}_{7(7)}}$ {#sec:EGT} -------------------------------------------------------------------------------- We would now like to describe an exceptional generalised geometry (EGG), analogous to the generalised geometry of Hitchin, but relevant to the description of eleven-dimensional supergravity rather than simply the NS–NS sector of type II. We will concentrate on the case of a seven-dimensional manifold $M$. The basic construction has been described, in general dimension $d$, in [@chris] and is closely related to the work of [@deWN; @Nic; @KNS]. Introducing the generalised tangent space allowed one to construct objects transforming under $O(d,d)$. The $g$ and $B$ degrees of freedom, parametrising the generalised metric $G$, then define an element in the $O(d,d)/O(d)\times O(d)$ coset. This coset structure is familiar from the moduli space of toroidal compactifications of NS-NS sector ten-dimensional supergravity to $10-d$ dimensions (see for instance [@Tduality]). In addition, there is, of course, a stringy $O(d,d;{\mathbb{Z}})$ T-duality symmetry relating equivalent compactifications. The string winding and momentum charges transform in the $2d$-dimensional vector representation of $O(d,d)$, namely $TM\oplus T^*M$. If one includes Ramond–Ramond fields, or equivalently considers toroidal compactifications of eleven-dimensional supergravity [@CJ; @Udual], the moduli spaces are cosets $E_{d(d)}/H$ (where $E_{d(d)}$ is the maximally non-compact real form of the exceptional group and $H$ is the corresponding maximal compact subgroup). Elements in the coset are parametrised by the components of the eleven-dimensional supergravity fields on the compact space, namely the metric $g$ and three-form $A$ potential (and potentially its dual six-form ${\tilde{A}}$). The discrete subgroup $E_{(d(d)}({\mathbb{Z}})$ is the U-duality group relating different equivalent M-theory backgrounds [@HT]. The momentum and brane charges fill out a particular representation of $E_{d(d)}$. In particular, in $d=7$, the momentum, membrane, fivebrane and Kaluza–Klein monopole charges [@KKmono] fill out the ${\mathbf{56}}$ representation of $\operatorname{\mathit{E}_{7(7)}}$. (Note that the definition of $\operatorname{\mathit{E}_{7(7)}}$, together with some details of its various representations, is summarised in appendix \[app:E7\].) Given this extension from the T-duality group $O(d,d)$ to U-duality group $E_{d(d)}$, it is natural to introduce a corresponding extension of generalised geometry. For $d=7$, the analogue of the generalised tangent space is an “exceptional generalised tangent space” (EGT) which transforms in the ${\mathbf{56}}$ representation of $\operatorname{\mathit{E}_{7(7)}}$. As in generalised geometry, the $\operatorname{\mathit{GL}}(7,{\mathbb{R}})$ structure group of the tangent and cotangent spaces should be a subgroup of $\operatorname{\mathit{E}_{7(7)}}$, and we also expect the gauge transformations of $A$, like the $B$-shifts in generalised geometry, to be somehow embedded into $\operatorname{\mathit{E}_{7(7)}}$. The construction of the EGT is as follows. As described in appendix \[app:E7def\], there is an $\operatorname{\mathit{SL}}(8,{\mathbb{R}})$ subgroup of $\operatorname{\mathit{E}_{7(7)}}$ under which the ${\mathbf{56}}$ representation is given by $$E_0 = \Lambda^2V \oplus \Lambda^2V^*$$ where $V$ is the eight-dimensional fundamental representation of $\operatorname{\mathit{SL}}(8,{\mathbb{R}})$. The $\operatorname{\mathit{GL}}(7,{\mathbb{R}})$ structure group of the tangent space $TM$ embeds as (see appendix \[app:GL7\]) $$\label{eq:Vdecomp} V = \big[(\Lambda^7T^*M)^{1/4}\otimes TM \big] \oplus (\Lambda^7T^*M)^{-3/4} .$$ One then finds $$\label{eq:E0def} E_0 = (\Lambda^7T^*M)^{-1/2} \otimes \left[ TM \oplus \Lambda^2T^*M \oplus \Lambda^5T^*M \oplus (T^*M \otimes\Lambda^7T^*M) \right] .$$ (Note that the final term in brackets can also be written as $(\Lambda^7T^*M)^2\otimes\Lambda^6TM$). The bundle $(\Lambda^7T^*M)^{-1/2}$ is isomorphic to the trivial bundle, thus there is always a (non-canonical) isomorphism $$\label{eq:Eiso} E_0 \simeq TM \oplus \Lambda^2T^*M \oplus \Lambda^5T^*M \oplus (T^*M \otimes\Lambda^7T^*M) .$$ It is $E_0$ which is the (untwisted) exceptional generalised tangent space[^1]. Except for the overall tensor density factor of $(\Lambda^7T^*M)^{-1/2}$, we see that we can identify it as a sum of vectors, two-forms, five-forms and one-forms tensor seven-forms. Given the isomorphism  we can write $$\label{eq:Xdef} X = x + \omega + \sigma + \tau \in E_0 ,$$ where, writing the $\operatorname{\mathit{GL}}(7,{\mathbb{R}})$ indices explicitly, we have $x^m$, $\omega_{mn}$, $\sigma_{m_1\dots m_5}$ and $\tau_{m,n_1\dots n_7}$. Physically in M-theory we expect these to correspond to momentum, membrane, fivebrane and Kaluza–Klein monopole charge respectively. Recall that the T-duality symmetry group acting on the generalised tangent space was defined in terms of a natural $O(d,d)$-invariant metric. As discussed in appendix \[app:E7def\], the group $\operatorname{\mathit{E}_{7(7)}}$ is defined by, not a metric, but a symplectic structure $\Omega$ and symmetric quartic invariant $q$ on the 56-dimensional representation space. These are given explicitly in terms of $\operatorname{\mathit{SL}}(8,{\mathbb{R}})$ representations in the appendix , and are by definition $\operatorname{\mathit{GL}}(7,{\mathbb{R}})$, hence diffeomorphism, invariant. Having identified the $\operatorname{\mathit{GL}}(7,{\mathbb{R}})$ tangent space symmetry in $\operatorname{\mathit{E}_{7(7)}}$, we would next like to identify the analogues of the $B$-shifts. This is essentially contained in the original dimensional reduction of eleven-dimensional supergravity on $T^7$ [@CJ]. We note that the 133-dimensional adjoint representation of $\operatorname{\mathit{E}_{7(7)}}$ decomposes under $\operatorname{\mathit{GL}}(7,{\mathbb{R}})$ as $$\label{eq:adbundle} A = (TM\otimes T^*M) \oplus \Lambda^6TM \oplus \Lambda^6T^*M \oplus \Lambda^3TM \oplus \Lambda^3T^*M .$$ Given there is a three-form potential $A$ in eleven-dimensional supergravity, the analogue of $B$-shifts should be $A$-shifts generated by $A\in\Lambda^3T^*M$. In fact, we will also consider ${\tilde{A}}$-shifts with ${\tilde{A}}\in\Lambda^6T^*M$ corresponding to the dual six-form potential. This will be described in more detail in the next section. For now we simply note that $A$ and ${\tilde{A}}$ are both elements of the adjoint bundle . Their action on $X\in E$ is given in . It exponentiates to $$\begin{aligned} & {\mathrm{e}}^{A+{\tilde{A}}}X = x + \left[\omega + i_xA\right] + \big[ \sigma + A\wedge\omega + \tfrac{1}{2}A\wedge i_xA + i_x{\tilde{A}}\big] \\ &\quad + \big[ \tau + jA\wedge\sigma - j{\tilde{A}}\wedge\omega + jA\wedge i_x{\tilde{A}}+ \tfrac{1}{2}jA\wedge A \wedge \omega + \tfrac{1}{6}jA\wedge A \wedge i_xA \big], \end{aligned}$$ where we are using a notion for elements of $T^*M\otimes(\Lambda^7T^*M)$ defined in . Note that the action truncates at cubic order. The corresponding Lie algebra, unlike the case of $B$-shifts is not Abelian. We have the commutator $$\label{eq:A-algebra} [ A + {\tilde{A}}, A' + {\tilde{A}}'] = - A \wedge A' .$$ That is, two $A$-shifts commute to give a ${\tilde{A}}$ shift [@Cremmer:1998px]. When $A$ and ${\tilde{A}}$ are non-trivial one is led to defining a twisted EGT which encodes the patching of the potentials. This is again completely analogous to the generalised geometrical case. One starts with $$X_0 \in TM \oplus \Lambda^2T^*M \oplus \Lambda^5T^*M \oplus (T^*M \otimes\Lambda^7T^*M) .$$ On a given patch $U_{(\alpha)}$ we define the shifted element $$X_{(\alpha)} = {\mathrm{e}}^{A_{(\alpha)} + {\tilde{A}}_{(\alpha)}}X_0 .$$ In passing from one patch to another we have, on $U_{(\alpha)}\cap U_{(\beta)}$, $$\label{eq:Epatch} X_{(\alpha)} = {\mathrm{e}}^{{\mathrm{d}}{\Lambda_{(\alpha\beta)}}+{\mathrm{d}}{{\tilde{\Lambda}}_{(\alpha\beta)}}}X_{(\beta)} ,$$ provided the connections $A$ and ${\tilde{A}}$ patch as $$\label{eq:AtApatch} \begin{aligned} A_{(\alpha)} - A_{(\beta)} &= {\mathrm{d}}{\Lambda_{(\alpha\beta)}}, \\ {\tilde{A}}_{(\alpha)} - {\tilde{A}}_{(\beta)} &= {\mathrm{d}}{{\tilde{\Lambda}}_{(\alpha\beta)}}- \tfrac{1}{2}{\mathrm{d}}{\Lambda_{(\alpha\beta)}}\wedge A_{(\beta)} . \end{aligned}$$ As we will see in the next section this corresponds exactly to the patching of the three- and six-form potentials arising from eleven-dimensional supergravity. The patching  imply that $X_{(\alpha)}$ are sections of a twisted EGT which we will denote as $E$. Explicitly in components we have $$\begin{aligned} x_{(\alpha)} &= x_{(\beta)} , \\ \omega_{(\alpha)} &= \omega_{(\beta)} + i_{x_{(\beta)}}{\mathrm{d}}{\Lambda_{(\alpha\beta)}}, \\ \sigma_{(\alpha)} &= \sigma_{(\beta)} + {\mathrm{d}}{\Lambda_{(\alpha\beta)}}\wedge\omega_{(\beta)} + \tfrac{1}{2}{\mathrm{d}}{\Lambda_{(\alpha\beta)}}\wedge i_{x_{(\beta)}}{\mathrm{d}}{\Lambda_{(\alpha\beta)}}+ i_{x_{(\beta)}}{\mathrm{d}}{{\tilde{\Lambda}}_{(\alpha\beta)}}, \\ \tau_{(\alpha)} &= \tau_{(\beta)} + j{\mathrm{d}}{\Lambda_{(\alpha\beta)}}\wedge\sigma_{(\beta)} - j{\mathrm{d}}{{\tilde{\Lambda}}_{(\alpha\beta)}}\wedge\omega_{(\beta)} + j{\mathrm{d}}{\Lambda_{(\alpha\beta)}}\wedge i_{x_{(\beta)}}{\mathrm{d}}{{\tilde{\Lambda}}_{(\alpha\beta)}}\\ &\qquad {} + \tfrac{1}{2}j{\mathrm{d}}{\Lambda_{(\alpha\beta)}}\wedge{\mathrm{d}}{\Lambda_{(\alpha\beta)}}\wedge\omega_{(\beta)} + \tfrac{1}{6}j{\mathrm{d}}{\Lambda_{(\alpha\beta)}}\wedge{\mathrm{d}}{\Lambda_{(\alpha\beta)}}\wedge i_{x_{(\beta)}}{\mathrm{d}}{\Lambda_{(\alpha\beta)}}\end{aligned}$$ One can define $E$ formally via a series of extensions $$\label{eq:twistE} \begin{gathered} 0 \longrightarrow \Lambda^2T^*M \longrightarrow E'' \longrightarrow TM \longrightarrow 0 , \\ 0 \longrightarrow \Lambda^5T^*M \longrightarrow E' \longrightarrow E'' \longrightarrow 0 , \\ 0 \longrightarrow T^*M\otimes \Lambda^7T^*M \longrightarrow E \longrightarrow E' \longrightarrow 0 , \end{gathered}$$ in analogy with . As for the $B$-field, the potentials $A$ and ${\tilde{A}}$ are formally connections on gerbes. To define the connective structure of the gerbe we must define the patchings on successively higher-order intersections. For $A$, on the corresponding multiple intersections of patches we have $$\begin{aligned} \Lambda_{(\alpha\beta)} + \Lambda_{(\beta\gamma)} + \Lambda_{(\gamma\alpha)} &= {\mathrm{d}}\Lambda_{(\alpha\beta\gamma)} && \quad \text{on $U_{(\alpha)}\cap U_{(\beta)} \cap U_{(\gamma)}$}, \\ \Lambda_{(\beta\gamma\delta)} - \Lambda_{(\alpha\gamma\delta)} + \Lambda_{(\alpha\beta\delta)} - \Lambda_{(\alpha\beta\gamma)} &= {\mathrm{d}}\Lambda_{(\alpha\beta\gamma\delta)} && \quad \text{on $U_{(\alpha)}\cap U_{(\beta)} \cap U_{(\gamma)} \cap U_{(\delta)}$} . \end{aligned}$$ For a quantised flux ${\mathcal{F}}={\mathrm{d}}A_{(\alpha)}$ we have $g_{(\alpha\beta\gamma\delta)}={\mathrm{e}}^{{\mathrm{i}}\Lambda_{(\alpha\beta\gamma\delta)}}\in U(1)$ with the cocycle condition $$g_{(\beta\gamma\delta\epsilon)}g^{-1}_{(\alpha\gamma\delta\epsilon)} g_{(\alpha\beta\delta\epsilon)}g^{-1}_{(\alpha\beta\gamma\epsilon)} g_{(\alpha\beta\gamma\delta)} = 1 ,$$ on $U_{(\alpha)}\cap \dots \cap U_{(\epsilon)}$. For ${{\tilde{\Lambda}}_{(\alpha\beta)}}$ there is a similar set of structures, with the final cocycle condition defined on a octuple intersection $U_{(\alpha_1)}\cap\dots\cap U_{(\alpha_8)}$. The bundle $E$ encodes all the topological information of the supergravity background: the twisting of the tangent space $TM$ as well as the patching of the form potentials, but, as for  is independent of the particular choice of $A$ and ${\tilde{A}}$. Finally we would like to identify the analogue of the Courant bracket for the EGT. We look for a pairing with $A$- and ${\tilde{A}}$-shifts as automorphisms when ${\mathrm{d}}A={\mathrm{d}}{\tilde{A}}=0$. One finds the unique “exceptional Courant bracket” (ECB): $$\label{eq:ECB} \begin{split} \big[ x+&\omega+\sigma+\tau, x'+\omega'+\sigma'+\tau' \big] = \\ & [x,x'] + \mathcal{L}_x\omega' - \mathcal{L}_{x'}\omega - \tfrac{1}{2}{\mathrm{d}}\left(i_x\omega'-i_{x'}\omega\right) \\ & \quad + \mathcal{L}_x\sigma' - \mathcal{L}_{x'}\sigma - \tfrac{1}{2}{\mathrm{d}}\left(i_x\sigma'-i_{x'}\sigma\right) + \tfrac{1}{2}\omega\wedge{\mathrm{d}}\omega' - \tfrac{1}{2}\omega'\wedge{\mathrm{d}}\omega \\ & \quad + \tfrac{1}{2}\mathcal{L}_x\tau' - \tfrac{1}{2}\mathcal{L}_{x'}\tau + \tfrac{1}{2}\big( j\omega\wedge{\mathrm{d}}\sigma' - j\sigma'\wedge{\mathrm{d}}\omega \big) - \tfrac{1}{2}\big( j\omega'\wedge{\mathrm{d}}\sigma - j\sigma\wedge{\mathrm{d}}\omega' \big) \end{split}$$ If ${G_{\text{closed}}(M)}$ is the group generated by closed $A$ and ${\tilde{A}}$ shifts, the ECB is invariant under the $\operatorname{Diff}(M)\ltimes{G_{\text{closed}}(M)}$. The exceptional generalised metric and ${{\operatorname{\mathit{SU}}(8)/{\mathbb{Z}}_2}}$ structures {#sec:EGM} ---------------------------------------------------------------------------------------------------- Having defined the EGT, its topology and the corresponding bracket, one can then introduce, as in [@chris], the analog of the generalised metric which encodes the fields of eleven-dimensional supergravity. The motivation is that, when compactified on $T^7$, the moduli arising from the eleven-dimensional supergravity fields $g$ and $A$ parametrise a $\operatorname{\mathit{E}_{7(7)}}/({{\operatorname{\mathit{SU}}(8)/{\mathbb{Z}}_2}})$ coset space [@CJ]. Rather than consider a fixed element of the coset space, one takes one that is a function of position in the manifold $M$ [@deWN]. In the following, instead of starting with the coset, we show how the exceptional generalised metric can be defined as a generic ${{\operatorname{\mathit{SU}}(8)/{\mathbb{Z}}_2}}$ structure on the EGT. Fixing an element of the coset space $\operatorname{\mathit{E}_{7(7)}}/({{\operatorname{\mathit{SU}}(8)/{\mathbb{Z}}_2}})$ is equivalent to choosing a particular ${{\operatorname{\mathit{SU}}(8)/{\mathbb{Z}}_2}}$ subgroup of $\operatorname{\mathit{E}_{7(7)}}$. Making such a choice at each point in $M$ corresponds geometrically to a ${{\operatorname{\mathit{SU}}(8)/{\mathbb{Z}}_2}}$ structure on $E$. Such a structure can be defined as follows. The ${\mathbf{56}}$ representation decomposes into ${\mathbf{28}}+\bar{{\mathbf{28}}}$ under ${{\operatorname{\mathit{SU}}(8)/{\mathbb{Z}}_2}}$. Thus an ${{\operatorname{\mathit{SU}}(8)/{\mathbb{Z}}_2}}$ structure is equivalent to the existence of a decomposition of the (complexified) EGT $E\otimes{\mathbb{C}}=C\oplus\bar{C}$ where the fibres of $C$ transform in the ${\mathbf{28}}$ of ${{\operatorname{\mathit{SU}}(8)/{\mathbb{Z}}_2}}$. However this is the same as an almost complex structure $J$ with $J^2=-{\mathbb{1}}$ on $E$. For $J$ to define an ${{\operatorname{\mathit{SU}}(8)/{\mathbb{Z}}_2}}$ subgroup it must also be compatible with the $\operatorname{\mathit{E}_{7(7)}}$ structure. Recall that the latter was defined by a symplectic form $\Omega$ and a quartic invariant $q$. Compatibility requires $$\Omega(JX,JY) = \Omega(X,Y) , \qquad q(JX) = q(X) ,$$ or in other words $J\in\operatorname{\mathit{E}_{7(7)}}$. (Note the that first condition is just the usual condition between a symplectic and an almost complex structure required to define an Hermitian metric.) Such an almost complex structure $J$ defines an ${{\operatorname{\mathit{SU}}(8)/{\mathbb{Z}}_2}}$ structure on $E$. In contrast to the generalised geometry case where the $O(d)\times O(d)$ structure was equivalent to a compatible almost product structure satisfying $\Pi^2={\mathbb{1}}$, for ${{\operatorname{\mathit{SU}}(8)/{\mathbb{Z}}_2}}\subset\operatorname{\mathit{E}_{7(7)}}$ the structure is defined by a compatible almost complex structure satisfying $J^2=-{\mathbb{1}}$. Given $J$ and $\Omega$ one can then define the corresponding exceptional generalised metric (EGM) $G$ by $$G(X,Y) = \Omega(X,JY) ,$$ which gives a positive definite metric on $E$. We now turn to how one constructs the generic form of $J$ and hence $G$. Given a metric ${\hat{g}}_{ab}$ on the $\operatorname{\mathit{SL}}(8,{\mathbb{R}})$ representation space $V$, a natural way to define a particular almost complex structure $J_0$ (using the conventions of appendix \[app:E7def\], so that, in particular, pairs of indices $aa'$ and $bb'$ are antisymmetrised) is as $$J_0X = \begin{pmatrix} 0 & - {\hat{g}}^{ab}{\hat{g}}^{a'b'} \\ {\hat{g}}_{ab}{\hat{g}}_{a'b'} & 0 \end{pmatrix} \begin{pmatrix} x^{bb'} \\ x'_{bb'} \end{pmatrix} = \begin{pmatrix} - {\hat{g}}^{ab}{\hat{g}}^{a'b'}x'_{bb'} \\ {\hat{g}}_{ab}{\hat{g}}_{a'b'}x^{bb'} \end{pmatrix} .$$ By construction $J_0^2=-{\mathbb{1}}$. The corresponding EGM is $$G_0(X,Y) = {\hat{g}}_{ab}{\hat{g}}_{a'b'}x^{aa'}y^{bb'} + {\hat{g}}^{ab}{\hat{g}}^{a'b'}x'_{aa'}y'_{bb'} .$$ From the definitions  and  of $\Omega$ and $q$ it is clear that $J_0\in\operatorname{\mathit{E}_{7(7)}}$ provided $\det{\hat{g}}=1$. Under an infinitesimal $\operatorname{\mathit{E}_{7(7)}}$ transformation $\mu\in{\mathbf{133}}$ we have $$\delta J_0 = [ \mu , J_0 ] = \begin{pmatrix} \mu^{+aa'}{}_{bb'} & - 2 \mu^{+ab}{\hat{g}}^{a'b'} \\ - 2 \mu^+_{ab}{\hat{g}}_{a'b'} & - \mu^+_{aa'}{}^{bb'} \end{pmatrix}$$ where $\mu^\pm_{abcd}=\tfrac{1}{2}(\mu_{abcd}\pm *\mu_{abcd})$ and $\mu^\pm_{ab}=\tfrac{1}{2}(\mu_{ab}\pm\mu_{ba})$ where indices are raised and lowered using ${\hat{g}}$. Thus $J_0$ is invariant under the subgroup generated by $\mu^-_{ab}$ and $\mu^-_{abcd}$. As discussed in appendix \[app:SU8\] this is precisely ${{\operatorname{\mathit{SU}}(8)/{\mathbb{Z}}_2}}$ (see ). Given the embedding  of $\operatorname{\mathit{GL}}(7,{\mathbb{R}})\subset\operatorname{\mathit{SL}}(8,{\mathbb{R}})$ discussed in detail in appendix \[app:GL7\], we can define ${\hat{g}}$ in terms of a seven-dimensional metric $g$ as $$\label{eq:tgdef} {\hat{g}}_{ab} = (\det g)^{-1/4} \begin{pmatrix} g_{mn} & 0 \\ 0 & \det g \end{pmatrix} .$$ Acting on elements of $X=x+\omega+\sigma+\tau$ we have $$G_0(X,X) = 2\left(|x|^2 + |\omega|^2 + |\sigma|^2 + |\tau|^2 \right) ,$$ where $|\tau|^2=\frac{1}{7!}\tau_{m.n_1\dots n_7} \tau^{m,n_1\dots n_7}$, $|\sigma|^2=\frac{1}{5!}\sigma_{n_1\dots n_5}\sigma^{n_1\dots n_5}$ etc. and, so that the result is a scalar, we have dropped on overall factor of $(\det g)^{1/2}$, which is natural, since in writing $X=x+\omega+\sigma+\tau$ we are using the isomorphism . Given a seven-dimensional metric $g$ we have been able to write a particular ${{\operatorname{\mathit{SU}}(8)/{\mathbb{Z}}_2}}$ structure $J_0$. A generic structure, given all such structures lie in the same orbit, will be of the form $J=hJ_0h^{-1}$ where $h\in\operatorname{\mathit{E}_{7(7)}}$, or equivalently $G(X,Y)=G_0(h^{-1}X,h^{-1}Y)$. We write $h={\mathrm{e}}^\mu$ with the Lie algebra element $\mu=(\mu^a{}_b,\mu_{abcd})\in{\mathbf{133}}$. The elements $\mu^m{}_n$ generate the $\operatorname{\mathit{GL}}(7,{\mathbb{R}})$ subgroup and acting on $J_0$ simply change the form of the metric $g$. The additional components $\mu^8{}_m$ and $\mu^m{}_8$ modify the form of ${\hat{g}}$ . Since only $\mu^+_{ab}$ acts non-trivially on $J_0$, we need only consider transformations with, say, $\mu^m{}_8$. Similarly since only $\mu^+_{abcd}$ acts non-trivially we can generate a generic $J$ using only, say $\mu_{mnp8}$. However, $\mu^m{}_8$ and $\mu_{mnp8}$ transformations precisely correspond to the subgroup of $A$- and ${\tilde{A}}$-shifts. Thus, given a generic $g$ defining $G_0$, the generic EGM can be written as $$G(X,Y) = G_0({\mathrm{e}}^{-A-{\tilde{A}}}X,{\mathrm{e}}^{-A-{\tilde{A}}}Y) .$$ This is analogous to the form  of the generalised metric in generalised geometry. Note also that for non-trivial $A$ and ${\tilde{A}}$, $G_0$ is an EGM on the untwisted EGT given by $TM\oplus\Lambda^2T^*M\oplus\Lambda^5T^*M \oplus(T^*M\otimes\Lambda^7T^*M$), while $G$ is an EGM on the twisted bundle $E$ given by . Supersymmetric backgrounds and EGG {#sec:SEGG} ================================== We will now relate the EGG defined in the previous section to eleven-dimensional supergravity and in particular seven-dimensional supersymmetric backgrounds. After identifying the standard decomposition of the supergravity degrees of freedom on such backgrounds, we first briefly review the corresponding connection between type II backgrounds and generalised geometry before turning to EGG. The physical context we are interested in is where the eleven-dimensional spacetime is topologically a product of a four-dimensional “external” and a seven-dimensional “internal” space $$\label{eq:M11} M_{10,1} = M_{3,1} \times M_7$$ If $M_7$ is compact we can consider compactifying eleven-dimensional supergravity to give an effective four-dimensional theory. In particular, the effective theory could be supersymmetric. We could also look for particular examples of compactifications which are solutions of the supergravity field equations and preserve some number of supersymmetries. In either case, the geometry of $M_7$ is restricted, and some discussion of the latter case appears in [@chris]. The goal here is to understand how this restricted geometry can be naturally described in terms of EGG structures on $M_7$. We will also focus on the low-energy effective theory rather than the on-shell supersymmetric backgrounds. Effective theories and field decompositions {#sec:fdecomp} ------------------------------------------- Given the product , the tangent bundle decomposes as $TM_{10,1}=TM_{3,1}\oplus TM_7$ and all the supergravity fields can be decomposed under a local $\operatorname{\mathit{Spin}}(3,1)\times\operatorname{\mathit{Spin}}(7)\subset\operatorname{\mathit{Spin}}(10,1)$ symmetry. Normally one would derive a four-dimensional effective description by truncating the Kaluza–Klein spectrum of modes on $M_7$ to give a four-dimensional theory with a finite number of degrees of freedom. For instance, compactifying on a torus and keeping massless modes, one finds that the degrees of freedom actually arrange themselves into multiplets transforming under $\operatorname{\mathit{E}_{7(7)}}$ for the bosons and $\operatorname{\mathit{SU}}(8)$ for the fermions. However, one can also keep the full dependence of all eleven-dimensional fields on both the position on $M_{3,1}$ and $M_7$. One can then simply rewrite the eleven-dimensional theory, breaking the local $\operatorname{\mathit{Spin}}(10,1)$ symmetry to $\operatorname{\mathit{Spin}}(3,1)\times\operatorname{\mathit{Spin}}(7)$, so that it is analogous to a four-dimensional theory. This was done explicitly by de Wit and Nicolai [@deWN], retaining all 32 supersymmetries, where it was shown that in general the degrees of freedom fall into $\operatorname{\mathit{E}_{7(7)}}$ and ${{\operatorname{\mathit{SU}}(8)/{\mathbb{Z}}_2}}$ representations. In this paper we will ultimately be interested in such reformulations focusing on only four of the supercharges so that the theory has a structure analogous to $N=1$ four-dimensional supergravity. Note that formally the only requirement for making such rewritings is not that $M_{10,1}$ is topologically a product, but rather that the tangent space $TM_{10,1}$ decomposes into a four- and seven-dimensional part $$TM_{10,1} = T \oplus F .$$ For simplicity, here we will concentrate on the case of a product manifold, though all of our analysis actually goes through in the more general case, with the EGT defined in terms of $F$ rather than $TM_7$. The analogous analysis in terms of generalised geometry for type II compactifications was given in [@GLW1; @GLW2]. Let us briefly note how the fields decompose under $\operatorname{\mathit{Spin}}(3,1)\times\operatorname{\mathit{Spin}}(7)$. Our conventions for eleven-dimensional supergravity are summarised in appendix \[app:sugra\]. The degrees of freedom are the metric $g_{MN}$, three-form $A_{MNP}$ and gravitino $\Psi_M$. Consider first the $\operatorname{\mathit{Spin}}(3,1)$ scalars. The eleven-dimensional metric decomposes as a warped product $${\mathrm{d}}s^2(M_{11}) = {\mathrm{e}}^{2E}g^{(4)}_{\mu\nu}{\mathrm{d}}x^\mu{\mathrm{d}}x^\nu + g_{mn} {\mathrm{d}}x^m {\mathrm{d}}x^n ,$$ where $\mu=0,1,2,3$ denote coordinates on the external space. To get a conventionally normalised Einstein term in the four-dimensional effective theory we must take $${\mathrm{e}}^{-2E} = \sqrt{\det g} ,$$ where $\det g$ is the determinant of $g_{mn}$. In a conventional compactification, deformations of the internal metric $g_{mn}$ lead to scalar moduli fields in the effective theory. Moduli fields can also arise from the flux $F$. Keeping only $\operatorname{\mathit{Spin}}(3,1)$ scalar parts, one can decompose $$F = *_7{\tilde{{\mathcal{F}}}}\wedge {\mathrm{e}}^{4E}\epsilon_{(4)} + {\mathcal{F}}$$ where ${\mathcal{F}}\in\Lambda^4T^*M_7$ and ${\tilde{{\mathcal{F}}}}\in\Lambda^7T^*M_7$ and $\epsilon_{(4)}=\sqrt{-g^{(4)}}{\mathrm{d}}x^0\wedge\dots\wedge{\mathrm{d}}x^3$. The eleven-dimensional equation of motion and Bianchi identify  then decompose as $$\begin{aligned} {\mathrm{d}}{\tilde{{\mathcal{F}}}}+ \tfrac{1}{2}{\mathcal{F}}\wedge {\mathcal{F}}&= 0 , & \qquad \qquad {\mathrm{d}}\big({\mathrm{e}}^{4E}*_7 {\tilde{{\mathcal{F}}}}\big) &= 0 , \\ {\mathrm{d}}{\mathcal{F}}&= 0 , & \qquad \qquad {\mathrm{d}}\big({\mathrm{e}}^{4E}*_7 {\mathcal{F}}\big) + {\mathrm{e}}^{4E}*_7{\tilde{{\mathcal{F}}}}\wedge{\mathcal{F}}&= 0 \end{aligned}$$ so one can introduce, locally, $$\begin{aligned} {\mathcal{F}}& = {\mathrm{d}}A , \\ {\tilde{{\mathcal{F}}}}&= {\mathrm{d}}{\tilde{A}}- \tfrac{1}{2}A\wedge {\mathcal{F}}, \end{aligned}$$ where $A\in\Lambda^3T^*M_7$ and ${\tilde{A}}\in\Lambda^6T^*M_7$. By definition, ${\mathcal{F}}$ and ${\tilde{{\mathcal{F}}}}$ are globally elements of $\Lambda^4T^*M$ and $\Lambda^7T^*M$ respectively. Thus on $U_{(\alpha)}\cap U_{(\beta)}$ we have $$\begin{aligned} {\mathrm{d}}A_{(\alpha)} - {\mathrm{d}}A_{(\beta)} &= 0 \\ {\mathrm{d}}{\tilde{A}}_{(\alpha)} - {\mathrm{d}}{\tilde{A}}_{(\beta)} &= \tfrac{1}{2}\left(A_{(\alpha)}-A_{(\beta)}\right) \wedge{\mathrm{d}}A_{(\beta)} \end{aligned}$$ This implies that the potentials $A$ and ${\tilde{A}}$ must patch precisely as given by . We see that the twisting of the EGT  is precisely that corresponding to the supergravity potentials. Furthermore, given the discussion of the previous section \[sec:EGM\] the scalar degrees of freedom $g_{mn}$ and $A_{mnp}$ and ${\tilde{A}}_{m_1\dots m_6}$ scalars can be combined together as an EGM or equivalently an almost complex structure $J$ on $E$. Turning briefly to the remaining fields, there are 28 bosonic $\operatorname{\mathit{Spin}}(3,1)$ vector degrees of freedom coming from off-diagonal components of the metric $g_{\mu m}$ and from $A_{\mu mn}$. One usually also introduces the corresponding dual potentials giving a total of $56$. Finally for the fermionic degrees of freedom we decompose the eleven-dimensional gamma matrices as $$\label{gamma-decomp} \Gamma^\mu = {\mathrm{e}}^{-E} \gamma^\mu \otimes {\mathbb{1}}\qquad \Gamma^m = {{\mathrm{i}}\gamma_{(4)}}\otimes \gamma^m .$$ The seven-dimensional gamma-matrix conventions are defined in \[app:7d\], while the four-dimensional gamma matrices are chosen to satisfy $\{\gamma_\mu,\gamma_\nu\}=2g^{(4)}_{\mu\nu}{\mathbb{1}}$ and $\gamma_{\mu_1\dots\mu_4}=\gamma_{(4)}\epsilon^{(4)}_{\mu_1\dots\mu_4}$. The real eleven-dimensional spinors correspondingly decompose as $$\label{eq:spinor-decomp} \begin{aligned} \epsilon &= {\mathrm{e}}^{E/2}\theta_+\otimes{{\zeta}^c} + {\mathrm{e}}^{E/2}\theta_-\otimes{\zeta}\\ {\mathbf{32}} &= ({\mathbf{2}},{\mathbf{8}}) + (\bar{{\mathbf{2}}},\bar{{\mathbf{8}}}) \end{aligned}$$ where ${{\mathrm{i}}\gamma_{(4)}}\theta_\pm=\pm\theta_\pm$ (with $\theta_+^c=D\theta^*_+=\theta_-$ and $-\gamma_\mu^*=D^{-1}\gamma_\mu D$) are chiral four-dimensional spinors and ${{\zeta}^c}$ is a complex $\operatorname{\mathit{Spin}}(7)$ spinor. The factor of ${\mathrm{e}}^{E/2}$ and the choice of labelling ${\zeta}$ versus ${\zeta}^c$ are conventional. Thus $\Psi_\mu$ decomposes as eight spin-$\frac{3}{2}$ fermions, while $\Psi_m$ gives 56 spin-$\frac{1}{2}$ fermions. As discussed in appendix \[app:7d\] there is a natural embedding of $\operatorname{\mathit{SU}}(8)$ in the Clifford algebra $\operatorname{Cliff}(7,0;{\mathbb{R}})$ with the complex $\operatorname{\mathit{Spin}}(7)$ spinors transforming in the fundamental representation. In reformulating the eleven-dimensional theory, all the degrees of freedom, fermionic and bosonic arrange as $\operatorname{\mathit{SU}}(8)$ representations [@deWN]. Thus we can actually promote the $\operatorname{\mathit{Spin}}(3,1)\times\operatorname{\mathit{Spin}}(7)$ symmetry to $\operatorname{\mathit{Spin}}(3,1)\times\operatorname{\mathit{SU}}(8)$. This decomposition corresponds to the $N=8$ four-dimensional supergravity multiplet. It is summarized in table \[tab:N=8\] where ${\mathbf{r_s}}$ transforms as the ${\mathbf{r}}$ representation of $\operatorname{\mathit{SU}}(8)$ with $\operatorname{\mathit{Spin}}(3,1)$ spin ${\mathbf{s}}$. To summarise, from a EGG perspective, the scalar degrees of freedom define an ${{\operatorname{\mathit{SU}}(8)/{\mathbb{Z}}_2}}$ structure on the EGT. Given this structure (or rather the existence of a double cover $\operatorname{\mathit{SU}}(8)$ structure, which is not always guaranteed) one can then define $\operatorname{\mathit{SU}}(8)$ spinors and hence the fermionic degrees of freedom. This is the EGG analogue of requiring a metric, or $O(d)$ structure, and hence a set of vielbeins, before one can define ordinary spinors on a curved manifold. Review of generalized geometry of $N=2$ type II backgrounds ----------------------------------------------------------- The generic effective four-dimensional $N=2$ supersymmetric theories arising from type II supergravity compactified on $M_{3,1}\times M_6$ were analysed in terms of generalised geometry in [@GLW1; @GLW2]. The structure was as follows. The metric and $B$-field on $M_6$ combine into a generalised metric . This defines a $O(6)\times O(6)$ structure on the generalised tangent space , that is the decomposition $E=C^+\oplus C^-$. Assuming the double cover $\operatorname{\mathit{Spin}}(6)\times\operatorname{\mathit{Spin}}(6)$ exists, one can define $\operatorname{\mathit{Spin}}(6)$ spinors on $C^+$ and $C^-$ separately. In terms of the original ten-dimensional spinors one has the decomposition $$\begin{aligned} \epsilon^1 &= \theta^1_+ \otimes {\zeta}^1_- + \theta^1_- \otimes {\zeta}^1_+ \\ \epsilon^2 &= \theta^2_+ \otimes {\zeta}^2_\pm + \theta^2_- \otimes {\zeta}^2_\mp \end{aligned}$$ where in the second line one takes the upper sign for type IIA and the lower for type IIB. (Here ${\zeta}_\pm$ are complex, chiral $\operatorname{\mathit{Spin}}(6)$ spinors, with ${\zeta}_-={\zeta}_+^c$.) The two spinors ${\zeta}^1_+$ and ${\zeta}^2_+$ naturally transform under the two spin groups $\operatorname{\mathit{Spin}}(6)\times\operatorname{\mathit{Spin}}(6)$. Next, we want to concentrate on effective theories with $N=2$ supersymmetry in four dimensions. This means, we want to identify a fixed pair of spinors $({\zeta}^1_+,{\zeta}^2_+)$. The eight four-dimensional supersymmetric parameters are then parametrised by $(\theta^1_+,\theta^2_+)$. In order to be able to use these supersymmetries to decompose all modes of the ten-dimensional fields into $N=2$ multiplets, we must require that $({\zeta}^1_+,{\zeta}^2_+)$ are non-vanishing and globally defined. (We can then project the supersymmetry from an action on $M_{9,1}$ to an action on $M_{3,1}$.) But this condition is the same as requiring each spinor to define an $\operatorname{\mathit{SU}}(3)$ structure. Thus we have $$\text{$N=2$ effective theory} \Leftrightarrow \text{${{\operatorname{\mathit{SU}}(3)\times\operatorname{\mathit{SU}}(3)}}$ structure on $E$} .$$ The structure can be defined by the existence of the generalised metric $G$ on $E$ together with a pair of $\operatorname{\mathit{Spin}}(6)\times\operatorname{\mathit{Spin}}(6)$ spinors $({\zeta}^1_+,{\zeta}^2_+)$. A conventional $\operatorname{\mathit{SU}}(3)$ structure can similarly be defined by a ordinary metric $g$ together with a globally defined, nowhere vanishing spinor ${\zeta}_+$. However, one can also define the structure by a pair of real forms $J\in\Lambda^2T^*M_6$ and $\rho\in\Lambda^3T^*M_6$. In the case where the $\operatorname{\mathit{SU}}(3)$ structure is integrable, that is when $M_6$ is a Calabi–Yau manifold, $J$ is the Kähler form and $\rho$ is the real part of the holomorphic three-form[^2] $\Omega=\rho+{\mathrm{i}}\hat{\rho}$. Generically $(J,\rho)$ only define a $\operatorname{\mathit{SU}}(3)$ structure if $J\wedge\rho=0$ and $J\wedge J\wedge J=\frac{3}{2}\rho\wedge\hat{\rho}$. It is natural to ask if the ${{\operatorname{\mathit{SU}}(3)\times\operatorname{\mathit{SU}}(3)}}$ structure can be similarly defined in terms of $O(6,6)$ objects. It can, and the representations in question are the spinors of $\operatorname{\mathit{Spin}}(6,6)$. These are defined as follows. (For more details see appendix A of [@GLW2].) For $E_0=TM_6\oplus T^*M_6$ the spinor bundle $S(E_0)$ is isomorphic to the bundle of forms $$\label{spinbundles} S(E) = (\Lambda^7T^*M_6)^{-1/2} \otimes \Lambda^* T^* M_6 .$$ More generally, for $E$, an extension of the form , on any patch $U_{(\alpha)}$, a spinor $\Upsilon_{(\alpha)}\in S(E)$ is a sum of forms, with the patching $$\Upsilon_{(\alpha)} = {\mathrm{e}}^{-{\mathrm{d}}\Lambda_{(\alpha\beta)}}\Upsilon_{(\beta)} ,$$ where the action is by wedge product. Spinors of $O(6,6)$ are Majorana–Weyl. The positive and negative helicity spin bundles $S^\pm(E)$ are locally isomorphic to the bundles of even and odd forms ${\Lambda^\text{even/odd}}T^*M_6$. The Clifford action on $\Upsilon\in S(E)$, viewed as a sum of forms, is given by $$\label{cliff} (x+\xi)\cdot\Upsilon = i_x\Upsilon + \xi\wedge\Upsilon .$$ The usual spinor bilinear form on $S(E)$ is given by the Mukai pairing ${\big<{\cdot},{\cdot}\big>}$ on forms. Explicitly $$\label{eq:mukai} {\big<{\Upsilon},{\Upsilon'}\big>} = \sum_p (-)^{[(p+1)/2]} \Upsilon_{(p)} \wedge \Upsilon'_{(6-p)}\ ,$$ where the subscripts denote the degree of the component forms in $\Lambda^*T^*M_6$ and $[(p+1)/2]$ takes the integer part of $(p+1)/2$. Note that, given the isomorphism , the ordinary exterior derivative defines a natural generalised Dirac operator ${\mathrm{d}}:S^\pm(E)\to S^\mp(E)$. Given a generalised metric $G$, on can decompose $\operatorname{\mathit{Spin}}(6,6)$ spinors under $\operatorname{\mathit{Spin}}(6)\times\operatorname{\mathit{Spin}}(6)$. Projecting each subspace $C^\pm$ onto $TM$ defines a common $\operatorname{\mathit{Spin}}(6)$ subgroup. Under this group $\Upsilon\in S(E)$ transform as a bispinor, that is, as an element of $\operatorname{Cliff}(6,0;{\mathbb{R}})$. Explicitly one can write real $\Upsilon^\pm\in S^\pm$ as $$\label{66decomp} \Upsilon^\pm = \zeta_+\bar{\zeta}'_\pm \pm \zeta_-\bar{\zeta}'_\mp\ ,$$ where $\zeta_+$, $\zeta'_+$ are ordinary $\operatorname{\mathit{Spin}}(6)$ spinors and elements of the $\operatorname{\mathit{Spin}}(d)$ bundles $S^+(C^+)$ and $S^+(C^-)$ respectively. From this perspective $\Upsilon^\pm$ is a matrix. It can be expanded as $$\label{chichiepsilon} \Upsilon^\pm = \sum_p \frac{1}{8p!}\Upsilon^\pm_{m_1\dots m_p} \gamma^{m_1\dots m_p} ,$$ with $$\Upsilon^\pm_{m_1\dots m_p} = \operatorname{tr}(\Upsilon^\pm \gamma_{m_p\dots m_1}) \in \Lambda^pT^*M_6 ,$$ and where $\gamma^m$ are $\operatorname{\mathit{Spin}}(6)$ gamma-matrices and the trace is over the $\operatorname{\mathit{Spin}}(6)$ indices. For $\Upsilon^+$ only the even forms are non-zero, while for $\Upsilon^-$ the odd forms are non-zero. This gives an explicit realisation of the isomorphism between $S^\pm(E)$ and ${\Lambda^\text{even/odd}}T^*M_6$. One can introduce a pair of $\operatorname{\mathit{Spin}}(6)\times\operatorname{\mathit{Spin}}(6)$ spinors which define the ${{\operatorname{\mathit{SU}}(3)\times\operatorname{\mathit{SU}}(3)}}$ structure. Write the complex objects $$\label{purespinors} \Phi^+ := {\mathrm{e}}^{-B} {\zeta}^1_+\bar{{\zeta}}^2_+ , \qquad \Phi^- := {\mathrm{e}}^{-B} {\zeta}^1_+\bar{{\zeta}}^2_- ,$$ where again ${\mathrm{e}}^{-B}$ acts by wedge product. Note that in the special case where ${\zeta}^1_+={\zeta}^2_+$, the two $\operatorname{\mathit{SU}}(3)$ structures are the same, and we have $$\Phi^+ = \tfrac{1}{8} {\mathrm{e}}^{-B-{\mathrm{i}}J} , \qquad \Phi^- = - \tfrac{{\mathrm{i}}}{8}{\mathrm{e}}^{-B}\Omega .$$ Generically, each $\Phi^\pm$ individually defines an $\operatorname{\mathit{SU}}(3,3)$ structure on $E$. Provided these structures are compatible, together they define a common $\operatorname{\mathit{SU}}(3)\times\operatorname{\mathit{SU}}(3)$ structure. The requirements of compatibility, in terms of the Mukai pairing  is that [@GLW1] $$\label{eq:compat} \begin{aligned} {\big<{\Phi^+},{V\cdot\Phi^-}\big>} &= {\big<{\bar \Phi^+},{V\cdot\Phi^-}\big>}=0 \quad \forall \, V\in E , \\ {\big<{\Phi^+},{\bar{\Phi}^+}\big>} &= {\big<{\Phi^-},{\bar{\Phi}^-}\big>} . \end{aligned}$$ If $\Phi^\pm$ are given by  they are automatically compatible [@GMPT2]. However, one can also reverse the logic. The $\operatorname{\mathit{SU}}(3,3)$ structures can actually be defined using only the real parts $\Upsilon^\pm=\operatorname{Re}\Phi^\pm$. Furthermore any pair of $\operatorname{\mathit{Spin}}(6,6)$ spinors $(\Upsilon^+,\Upsilon^-)$ satisfying the conditions  define an ${{\operatorname{\mathit{SU}}(3)\times\operatorname{\mathit{SU}}(3)}}$ structure. So $$\text{${{\operatorname{\mathit{SU}}(3)\times\operatorname{\mathit{SU}}(3)}}$ structure on $E$} \Leftrightarrow (\Upsilon^+,\Upsilon^-) .$$ The spinor bundles $S^\pm(E)$ are 32-dimensional. The compatibility requirement  gives 13 conditions. Thus the space of $(\Upsilon^+\Upsilon^-)$ is 51-dimensional. Different structures $({\zeta}^+,{\zeta}^-)$ can be related by $O(6,6)$ transformations. Since a given structure is ${{\operatorname{\mathit{SU}}(3)\times\operatorname{\mathit{SU}}(3)}}$ invariant, the space of all structures should be related to the coset space $\Sigma=O(6,6)/{{\operatorname{\mathit{SU}}(3)\times\operatorname{\mathit{SU}}(3)}}$, which is 50-dimensional. From this perspective the compatible pair $(\Upsilon^+,\Upsilon^-)$ give an embedding, as a one-dimensional family of orbits, $$(\Upsilon^+, \Upsilon^-) : \frac{O(6,6)}{{{\operatorname{\mathit{SU}}(3)\times\operatorname{\mathit{SU}}(3)}}} \times {\mathbb{R}}^+ \hookrightarrow S^+(E)\oplus S^-(E) .$$ The additional ${\mathbb{R}}^+$ factor corresponds simply to an overall rescaling of the generalised spinors. The key point of introducing all these structures is that they provide a very simple way to characterise the effective theory. As shown in [@GLW1; @GLW2], $\Phi^\pm$ parametrise a special Kähler space underlying the vector or hypermultiplet scalar degrees of freedom, and there are simple expressions for the $N=2$ analogues of the superpotential in terms of $\Phi^\pm$. $N=1$ M-theory backgrounds, EGG and $\operatorname{\mathit{SU}}(7)$ structures ------------------------------------------------------------------------------ We would now like to identify the analogue of the $N=2$ ${{\operatorname{\mathit{SU}}(3)\times\operatorname{\mathit{SU}}(3)}}$ structure of type II theories for $N=1$ compactifications of eleven-dimensional supergravity. An ${{\operatorname{\mathit{SU}}(3)\times\operatorname{\mathit{SU}}(3)}}$ structure is the generalised geometrical extension of a conventional supersymmetric $\operatorname{\mathit{SU}}(3)$ structure and hence, when on-shell, generalises the notion of a Calabi–Yau three-fold. The structure introduced here is similarly the EGG generalisation of a $G_2$ structure on a seven-dimensional manifold. Identifying an $N=1$ background requires picking out four preferred supersymmetries out of 32, or equivalently a fixed seven-dimensional spinor ${{\zeta}}$ in the general decomposition . This decomposition is the most general compatible with four-dimensional Lorentz invariance [@LS], and generically defines a complex ${{\zeta}}$ on the internal space. Given an EGM $G$ we have an ${{\operatorname{\mathit{SU}}(8)/{\mathbb{Z}}_2}}$ structure on $E$, and ${{\zeta}}$ transforms as the fundamental representation ${\mathbf{8}}$ of double cover $\operatorname{\mathit{SU}}(8)$. To define a generic low-energy effective theory the spinor ${{\zeta}}$ must be globally defined and nowhere vanishing. The stabilizer group in $\operatorname{\mathit{SU}}(8)$ of a fixed element of the vector ${\mathbf{8}}$ representation is $\operatorname{\mathit{SU}}(7)$. Thus, given a fixed spinor ${\zeta}$ at each point of $M$, we see that $$\text{$N=1$ effective theory} \Leftrightarrow \text{$\operatorname{\mathit{SU}}(7)$ structure on $E$} .$$ The projection $E\to TM_7$ defines a $\operatorname{\mathit{GL}}(7,{\mathbb{R}})$ subgroup of $\operatorname{\mathit{E}_{7(7)}}$. Given a EGM $G$, this defines a $\operatorname{\mathit{Spin}}(7)\subset\operatorname{\mathit{SU}}(8)$ subgroup. We can then decompose ${\zeta}$ into real $\operatorname{\mathit{Spin}}(7)$ spinors, $${{\zeta}} = {\zeta}_1 + {\mathrm{i}}{\zeta}_2$$ Each ${\zeta}_i$ is stabilised by a $G_2\in\operatorname{\mathit{Spin}}(7)$ subgroup. Thus from the point of the view of the ordinary tangent space $TM_7$, if ${\zeta}_i$ are globally defined and non-vanishing we have a pair of $G_2$ structures. However, all we really require is a globally defined non-vanishing complex ${{\zeta}}$. Thus in general we may not have either $G_2$ structure. In analogy to the case where ${\zeta}$ is real and we have a single $G_2$ structure, we can define the complex bilinears $$\label{eq:bilins} \varphi_{mnp} = {\mathrm{i}}\bar{{\zeta}}^c\gamma_{mnp}{\zeta}\quad \text{and} \quad \psi_{mnpq} = (*_7\varphi)_{mnpq} = -\bar{{\zeta}}^c\gamma_{mnpq}{\zeta}.$$ Locally, the pair of ${\zeta}_i$ are preserved by a $\operatorname{\mathit{SU}}(3)$ group. We have seen that one way to define the structure is as the pair of EGM and $\operatorname{\mathit{SU}}(8)$ spinor $(G,{{\zeta}})$. However, as in the type II case, we can also find an element lying in a particular orbit in an $\operatorname{\mathit{E}_{7(7)}}$ representation which can also be used to define the structure. We expect that it can be defined as a spinor bilinear. As we discuss in a moment, this space should also correspond to the $N=1$ chiral multiplet space in the four-dimensional effective theory. Decomposing under $\operatorname{\mathit{SU}}(7)$, the ${\mathbf{56}}$ representation has no singlets so cannot have elements stabilized by $\operatorname{\mathit{SU}}(7)$. The adjoint ${\mathbf{133}}$ does have a singlet. In terms of the spinor ${{\zeta}}$, the singlet in $\mu\in{\mathbf{133}}$, using its decomposition ${\mathbf{133}}={\mathbf{63}}+{\mathbf{35}}+\bar{{\mathbf{35}}}$ under $\operatorname{\mathit{SU}}(8)$, can be written as $$\begin{aligned} \mu_0 &= \big( {\mu_0}^\alpha{}_\beta, \mu_{0\;\alpha\beta\gamma\delta}, \bar{\mu}_0{}^{\alpha\beta\gamma\delta} \big) , \\ &= \big( {{\zeta}}^\alpha\bar{{{\zeta}}}_\beta - \tfrac{1}{8}(\bar{{{\zeta}}}{{\zeta}})\delta^\alpha{}_\beta, 0, 0 \big) . \end{aligned}$$ However, it is easy to see that this is stabilized by $U(7)$ rather than $\operatorname{\mathit{SU}}(7)$. The next smallest $\operatorname{\mathit{E}_{7(7)}}$ representation is ${\mathbf{912}}$. (See appendix \[app:reps\] for our conventions for $\operatorname{\mathit{E}_{7(7)}}$ representations.) We can define the following $\operatorname{\mathit{SU}}(7)$-singlet complex element in terms of its $\operatorname{\mathit{SU}}(8)$ decomposition, that is ${\mathbf{912}}={\mathbf{36}}+{\mathbf{420}}+\bar{{\mathbf{36}}}+\bar{{\mathbf{420}}}$, $$\label{eq:phi0} \begin{aligned} \phi_0 &= (\phi_0{}^{\alpha\beta},\phi_0{}^{\alpha\beta\gamma}{}_\delta, \bar{\phi}_{0\;\alpha\beta}, \bar{\phi}_{0\;\alpha\beta\gamma}{}^\delta ) \\ &= ( {\zeta}^\alpha{\zeta}^\beta, 0 , 0 , 0 ) , \end{aligned}$$ Finally we can form the structure together with the form-field potentials $$\phi = {\mathrm{e}}^{A+{\tilde{A}}}\phi_0 .$$ Such a $\phi$ does indeed define a generic $\operatorname{\mathit{SU}}(7)$ structure. To see this, we first note that under an infinitesimal $\operatorname{\mathit{E}_{7(7)}}$ transformation we have $$\begin{aligned} \delta\phi_0{}^{\alpha\beta} &= (\mu^\alpha{}_\gamma{\zeta}^\gamma){\zeta}^\beta + {\zeta}^\alpha(\mu^\beta{}_\gamma{\zeta}^\gamma) , \\ \delta\phi_0{}^{\alpha\beta\gamma}{}_\delta &= 0 , \\ \delta\bar{\phi}_{0\;\alpha\beta} &= 0 ,\\ \delta\bar{\phi}_{0\;\alpha\beta\gamma}{}^\delta &= \mu_{\alpha\beta\gamma\epsilon}{\zeta}^\epsilon{\zeta}^\delta . \end{aligned}$$ Thus $\phi_0$ is stabilized by elements of $\operatorname{\mathit{E}_{7(7)}}$ satisfying $$\mu^\alpha{}_\beta {\zeta}^\beta = 0 , \qquad \mu_{\alpha\beta\gamma\delta}{\zeta}^\delta = 0 .$$ Since $\bar{\mu}_{\alpha\beta\gamma\delta}= *\mu_{\alpha\beta\gamma\delta}$, the second condition can only be satisfied if $\mu_{\alpha\beta\gamma\delta}=0$. Since $\mu^\alpha{}_\beta$ is an element of the adjoint of $\operatorname{\mathit{SU}}(8)$ we see that $\phi_0$ is indeed stabilised by $\operatorname{\mathit{SU}}(7)$. Since ${\mathrm{e}}^{A+{\tilde{A}}}\in\operatorname{\mathit{E}_{7(7)}}$ the stabilizer of $\phi$ must also be $\operatorname{\mathit{SU}}(7)$. Finally, note that the $\operatorname{\mathit{SU}}(8)$ representations were defined using the gamma matrices ${\hat{\gamma}}^a$ defined using the seven dimensional metric $g$. Since action by ${\mathrm{e}}^{A+{\tilde{A}}}$ generates a generic EGM, we see that (when taken with the choice of generic $g$ and spinor ${{\zeta}}$, which are implicit when we write ) the action of ${\mathrm{e}}^{A+{\tilde{A}}}$ must generate a generic element of the orbit under $\operatorname{\mathit{E}_{7(7)}}$. Note that we could also define a real object $\lambda=\operatorname{Re}\phi$ $$\lambda = {\mathrm{e}}^{A+{\tilde{A}}} ({\zeta}^\alpha{\zeta}^\beta,0,\bar{{\zeta}}_\alpha\bar{{\zeta}}_\beta,0) ,$$ which also manifestly defines the same $\operatorname{\mathit{SU}}(7)$ structure. Let $N(E)$ be the ${\mathbf{912}}$ representation space based on the EGT $E$, at each point $x\in M_7$, we can view $\lambda$ as an embedding of the coset space $$\lambda : \frac{\operatorname{\mathit{E}_{7(7)}}}{\operatorname{\mathit{SU}}(7)}\times{\mathbb{R}}^+ \hookrightarrow N(E) ,$$ where the ${\mathbb{R}}^+$ factor simply corresponds to a rescaling of $\lambda$. We will call this orbit subspace $\Sigma$. There should then be a natural complex structure on $\Sigma$ which allows one to define the holomorphic $\phi$. This is in analogy to $\Upsilon^\pm$ and $\Phi^\pm$ in generalised geometry. Note that $\lambda$ far from fills out the whole of the ${\mathbf{912}}$ representation space. Rather we are considering a very particular orbit. One could always write down the particular non-linear conditions which define the orbit, that is, the analogues of . Finally let us also consider how the supergravity fields decompose under the $\operatorname{\mathit{SU}}(7)$ subgroup and how these correspond to different $N=1$ multiplets. We have for $\operatorname{\mathit{SU}}(7)\subset\operatorname{\mathit{SU}}(8)$ $$\begin{aligned} {\mathbf{8}} &= {\mathbf{7}} + {\mathbf{1}} , & {\mathbf{35}} &= {\mathbf{35}} , \\ {\mathbf{28}} &= {\mathbf{21}} + {\mathbf{7}} , &\qquad {\mathbf{56}} &= {\mathbf{35}} + {\mathbf{21}} . \end{aligned}$$ This means we can arrange the degrees of freedom as in table \[tab:SU7\]. Note that the coset space $\operatorname{\mathit{E}_{7(7)}}/\operatorname{\mathit{SU}}(7)$ actually decomposes into ${\mathbf{35}}+{\mathbf{7}}+\bar{{\mathbf{35}}}+\bar{{\mathbf{7}}}$. Thus there are more degrees of freedom in $\lambda$ than chiral degrees of freedom. The same phenomenon appears in the type II case and is associated to the gauge freedom of the extra spin-$\tfrac{3}{2}$ multiplets. One solution is to assume in a given truncation of the theory that there are no ${\mathbf{7}}$ degrees of freedom. Note that in this picture we expect there to be a natural Kähler metric on the coset space $\Sigma=\operatorname{\mathit{E}_{7(7)}}/\operatorname{\mathit{SU}}(7)\times{\mathbb{R}}^+$ corresponding to the Kähler metric on the chiral scalar field space of $N=1$ theories [@inprog]. Application: the effective superpotential {#sec:W} ========================================= In the previous section we found the objects defining the $\operatorname{\mathit{SU}}(7)$ structure relevant to $N=1$ reformulations of eleven-dimensional supergravity. The elements $\phi\in\Sigma$ should correspond to the chiral multiplet degrees of freedom. As such there should be an analogue of the four-dimensional superpotential $W$, as a holomorphic function of $\phi$. In this section, we will derive the generic structure of $W$ and show that it can be written in an $\operatorname{\mathit{E}_{7(7)}}$ covariant form. This is the analogue of the corresponding generalised geometry calculation in the case of type II given in [@GLW1; @GLW2]. Note that the structure of $W$, for the special case of a $G_2$ structure was previously derived in [@HM]. Generic form of the effective superpotential {#sec:genW} -------------------------------------------- We will read off $W$ from the variation of the four-dimensional gravitino. Recall that the $N=1$ gravitino variations are given by $$\label{eq:4dsusy} \delta \psi_{\mu +} = \nabla_{\mu} \theta_{+} + \tfrac{1}{2}{\mathrm{i}}{\mathrm{e}}^{K/2} W \gamma_{\mu} \theta_{-} + \dots ,$$ where $W$ is the superpotential and $K$ the Kähler potential. The expressions for ${\mathrm{e}}^{K/2}W$ can then be derived directly from the eleven-dimensional gravitino variation (see appendix \[app:sugra\] for our conventions) $$\label{susy11} \delta \Psi_M = \nabla_M \epsilon + \tfrac{1}{288}\left( \Gamma_M{}^{NPQR} - 8 \delta_M^N\Gamma^{PQR}\right)F_{MNPQ} + \dots ,$$ where the dots denote terms depending on $\Psi_M$. We must first identify the correctly normalised four-dimensional gravitino $\psi_\mu$. A naive decomposition $\Psi_{M} = (\Psi_{\mu},\Psi_m)$ and identifying $\psi_\mu$ as part of $\Psi_\mu$, leads to cross-terms in the kinetic energy, so instead we first need to diagonalise the four-dimensional gravitino kinetic energy term. This requires the following shift $$\tilde{\Psi}_\mu := \Psi_\mu + \tfrac{1}{2}\Gamma_\mu\Gamma^m\Psi_m .$$ One further has to rescale by a factor of ${\mathrm{e}}^E$, and hence identify the four-dimensional gravitino $\psi_\mu$ as the $\operatorname{\mathit{SU}}(7)$ singlet part $$\tilde{\Psi}_\mu = {\mathrm{e}}^{E/2} \psi_{\mu +}\otimes{{\zeta}^c} + {\mathrm{e}}^{E/2} \psi_{\mu +}^c \otimes{\zeta}+ \dots$$ where the dots denote non-singlet terms. This rescaling by ${\mathrm{e}}^{E/2}$ is the reason for adopting the conventions in the spinor decomposition given in . Given we can rescale ${{\zeta}^c}$ by including factors in $\theta_+$, we can always choose a normalisation $$\label{norm} \bar{\zeta}{\zeta}= 1 .$$ This allows us to introduce the projectors $$\begin{aligned} \Pi_+ &:= \tfrac{1}{2}(1+{{\mathrm{i}}\gamma_{(4)}}) \otimes {{\zeta}^c}\bar{{\zeta}}^c \\ \Pi_- &:= \tfrac{1}{2}(1-{{\mathrm{i}}\gamma_{(4)}}) \otimes {\zeta}\bar{\zeta}\end{aligned}$$ such that $${\mathrm{e}}^{-E/2}\Pi_+ \tilde{\Psi}_\mu = \psi_{\mu +} \otimes {{\zeta}^c} .$$ It is now straightforward to calculate $\delta\psi_\mu$ in terms of ${{\zeta}^c}$, ${\mathcal{F}}$ and ${\tilde{{\mathcal{F}}}}$. By definition $$\begin{aligned} \delta\psi_\mu\otimes{{\zeta}^c} &= {\mathrm{e}}^{-E/2}\Pi_+\delta\tilde{\Psi}_\mu = {\mathrm{e}}^{-E/2}\Pi_+\left(\delta\Psi_\mu +\tfrac{1}{2}\Gamma_\mu\Gamma^m\delta\Psi_m\right) \\ &= \nabla_{\mu} \theta_{+}\otimes{{\zeta}^c} + \tfrac{1}{2}{\mathrm{i}}{\mathrm{e}}^{K/2}W\gamma_\mu\theta_-\otimes{{\zeta}^c} , \end{aligned}$$ which gives $${\mathrm{e}}^{K/2}W = \tfrac{1}{4}{\mathrm{i}}{\mathrm{e}}^E\left( 4\bar{{\zeta}}^c\gamma^m\nabla_m{\zeta}+ \tfrac{1}{4!}{\mathcal{F}}_{mnpq}\bar{{\zeta}}^c\gamma^{mnpq}{\zeta}- {\mathrm{i}}*_7{\tilde{{\mathcal{F}}}}\bar{{\zeta}}^c{\zeta}\right) .$$ We have used the fact that $\bar{{\zeta}}^c\gamma^m{\zeta}=0$ identically to remove $\nabla_mE$ terms. This expression can be put in a more standard form by writing: $${\zeta}\bar{{\zeta}}^c = \tfrac{1}{8} (\bar{{\zeta}}^c{\zeta}){\mathbb{1}}- \tfrac{1}{8\cdot 3!}(\bar{{\zeta}}^c\gamma_{mnp}{{\zeta}^c})\gamma^{mnp} .$$ Thus, since $\bar{{\zeta}}^c\gamma^m{\zeta}=\bar{{\zeta}}^c\gamma^{mn}{\zeta}=0$, we have $$\begin{split} (\bar{{\zeta}}^c\gamma^m\nabla_m{\zeta})(\bar{{\zeta}}^c{\zeta}) &= \bar{{\zeta}}^c\gamma^m\nabla_m({\zeta}\bar{{\zeta}}^c){\zeta}\\ &= \tfrac{{\mathrm{i}}}{8\cdot 3!}\nabla_m\varphi_{npq} (\bar{{\zeta}}^c\gamma^m\gamma^{npq}{\zeta}) \\ &= - \tfrac{{\mathrm{i}}}{8\cdot 4!}({\mathrm{d}}\varphi)_{mnpq}(*_7\varphi)^{mnpq} \\ &= - \tfrac{{\mathrm{i}}}{8}*_7(\varphi\wedge{\mathrm{d}}\varphi) , \end{split}$$ where we used . Similarly, $$\tfrac{1}{4!}{\mathcal{F}}_{mnpq}\bar{{\zeta}}^c\gamma^{mnpq}{\zeta}= - *_7 ({\mathcal{F}}\wedge\varphi) .$$ Hence $$\label{eq:Wfinal} \begin{split} *_7\, {\mathrm{e}}^{K/2} W &= \frac{1}{8}{\mathrm{e}}^{E}\left[ \frac{1}{\bar{{\zeta}}^c{\zeta}}\varphi\wedge{\mathrm{d}}\varphi - 2{\mathrm{i}}{\mathcal{F}}\wedge\varphi + 2{\tilde{{\mathcal{F}}}}\bar{{\zeta}}^c{\zeta}\right] \\ &= \frac{1}{8}\bar{{\zeta}}^c{\zeta}{\mathrm{e}}^{E}\left[ {\mathrm{d}}({\tilde{\varphi}}-{\mathrm{i}}A)\wedge({\tilde{\varphi}}-{\mathrm{i}}A) + 2{\mathrm{d}}{\tilde{A}}- {\mathrm{i}}{\mathrm{d}}(A\wedge{\tilde{\varphi}})\right] \end{split}$$ where we have introduced the renormalised ${\tilde{\varphi}}=\varphi/\bar{{\zeta}}^c{\zeta}$. In the case where ${{\zeta}^c}={\zeta}$ we have a global $G_2$ structure, our normalisation convention  implies that $\bar{{\zeta}}^c{\zeta}=1$ and one finds that  agrees with that derived in [@HM]. The generic $\operatorname{\mathit{SU}}(7)$ case differs form the simple $G_2$ case as through the pre-factor $\bar{{\zeta}}^c{\zeta}$ and the fact that $\tilde\varphi$ is no longer real. An $\operatorname{\mathit{E}_{7(7)}}$ covariant expression {#sec:covW} ---------------------------------------------------------- In this section we show that one may rewrite the superpotential term  in a manifestly $\operatorname{\mathit{E}_{7(7)}}$ invariant form using the $\operatorname{\mathit{SU}}(7)$ structure $\phi\in{\mathbf{912}}$. This is the analogue of the $O(6,6)$ invariant expressions for the $N=2$ prepotentials given in [@GLW2] for type IIA and IIB theories compactified on $\operatorname{\mathit{SU}}(3)\times\operatorname{\mathit{SU}}(3)$ structure backgrounds. We first need to introduce an embedding of the derivative operator into an $\operatorname{\mathit{E}_{7(7)}}$ representation. Given the $\operatorname{\mathit{GL}}(7)$ decomposition of the EGT given in , we see that, assuming for the moment we have a metric $g$, we can introduce an operator $$D = (D^{ab}, D_{ab}) \in {\mathbf{56}}$$ with $$\label{eq:Ddef} D^{mn} = D^{m8} = D_{mn} = 0 , \qquad D_{m8} = (\det g)^{1/4}\nabla_m .$$ Transforming to spinor indices gives $$D = (D^{\alpha\beta},\bar{D}_{\alpha\beta}) = (-\tfrac{1}{2\sqrt{2}}\gamma^{m\,\alpha\beta}\nabla_m, \tfrac{1}{2\sqrt{2}}\gamma^m_{\alpha\beta}\nabla_m) .$$ Given the derivative operator we can define its action on $\phi\in{\mathbf{912}}$. In particular using the product between ${\mathbf{56}}$ and ${\mathbf{912}}$ (see  or ) we can define an object in the ${\mathbf{133}}$ representation which we denote as $(D\cdot\phi)^{AB}=D_C\phi^{C(AB)}$ (where indices are raised and lowered using the symplectic structure $\Omega_{AB}$). The claim is that the superpotential can be written as $$\label{eq:E7W} (D\cdot\phi)\,\phi = - \left(\frac{3}{4\sqrt{2}}\,{\mathrm{e}}^{-E}{\mathrm{e}}^{K/2}W\right) \phi ,$$ where $(D\cdot\phi)\,\phi\in{\mathbf{912}}$ denotes the adjoint action of $D\cdot\phi$ on $\phi$ itself. The statement is that $(D\cdot\phi)\,\phi$ is itself an $\operatorname{\mathit{SU}}(7)$ singlet, proportional to $\phi$ and $W$ is related to the constant of proportionality. Note that we expect $W$ to be a holomorphic function of $\phi$, since $\phi$ encodes the chiral multiplet scalar fields, whereas the Kähler potential should be a function of both $\phi$ and its complex conjugate. This suggests we should identify ${\mathrm{e}}^E$ and ${\mathrm{e}}^{K/2}$, or equivalently ${\mathrm{e}}^{-K}=\textrm{const.}\times\sqrt{\det g}$. We return to this briefly in section \[sec:concl\]. To show  requires two steps. First recall that we defined $\phi={\mathrm{e}}^{A+{\tilde{A}}}\phi_0$. Writing $h={\mathrm{e}}^{A+{\tilde{A}}}\in\operatorname{\mathit{E}_{7(7)}}$, one can then define a new operator $D_0$ with a connection taking values in the adjoint of $\operatorname{\mathit{E}_{7(7)}}$ by conjugating by $h$. Making the $\operatorname{\mathit{E}_{7(7)}}$ indices explicit, we define $$D_0^{AB}{}_C = D^A\delta^B{}_C + \kappa^{AB}{}_C ,$$ where $\kappa^{AB}{}_C=(h^{-1})^B{}_ED^Ah^E{}_C$. One then has $D\cdot\phi={\mathrm{e}}^{A+{\tilde{A}}}(D_0\cdot\phi_0)$ where $$\begin{aligned} (D_0\cdot\phi_0)^{AB} &= \Omega_{CD} \Big( D^{C}\phi_0^{D(AB)} \\ & + \kappa^{CD}{}_E\phi_0^{E(AB)} + \kappa^{C(A}{}_E\phi_0^{DE|B)} + \kappa^{C(B}{}_E\phi_0^{D|A)E} \Big) , \end{aligned}$$ where we have used the fact that $h^A{}_BD^B=D^A$. The expression  can thus be rewritten as $$(D_0\cdot\phi_0)\,\phi_0 = - \left(\frac{3}{4\sqrt{2}}\,{\mathrm{e}}^{-E}{\mathrm{e}}^{K/2}W\right) \phi_0 .$$ It is then straightforward to calculate $(D_0\cdot\phi_0)\,\phi_0$. Using the Hadamard formula $${\mathrm{e}}^P Q {\mathrm{e}}^{-P} = Q + [P,Q] + \tfrac{1}{2}[P,[P,Q]] + \dots ,$$ for operators $P$ and $Q$, one can calculate $D_0$. Given the commutator algebra , we have $$\begin{split} {\mathrm{e}}^{-A-{\tilde{A}}}\nabla_m{\mathrm{e}}^{A+{\tilde{A}}} &= \nabla_m + \nabla_m(A+{\tilde{A}}) + \tfrac{1}{2}[\nabla_m(A+{\tilde{A}}),A+{\tilde{A}}] + \dots \\ &= \nabla_m + \nabla_mA+\nabla_m{\tilde{A}}-\tfrac{1}{2}\nabla_mA\wedge A \\ &:= \nabla_m + \kappa_m . \end{split}$$ Note that this expression truncates at quadratic order. Again the connection $\kappa_m$ takes values in the $\operatorname{\mathit{E}_{7(7)}}$ Lie algebra. In the ${{\operatorname{\mathit{SU}}(8)/{\mathbb{Z}}_2}}$ basis $\phi_0$ takes the form . Hence $$\begin{aligned} (\nabla_m+\kappa_m)\phi_0{}^{\alpha\beta} &= \nabla_m({\zeta}^\alpha{\zeta}^\beta) + (\kappa_m{}^\alpha{}_\gamma{\zeta}^\gamma){\zeta}^\beta + {\zeta}^\alpha(\kappa_m{}^\beta{}_\gamma{\zeta}^\gamma) , \\ (\nabla_m+\kappa_m)\phi_0{}^{\alpha\beta\gamma}{}_\delta &= 0 , \\ (\nabla_m+\kappa_m)\bar{\phi}_{0\;\alpha\beta} &= 0 , \\ (\nabla_m+\kappa_m)\bar{\phi}_{0\;\alpha\beta\gamma}{}^\delta &= \kappa_{m\,\alpha\beta\gamma\epsilon}{\zeta}^\epsilon{\zeta}^\delta . \end{aligned}$$ so that $$\begin{aligned} (D_0\cdot\phi_0)^\alpha{}_\beta &= \frac{3{\mathrm{i}}}{8\sqrt{2}} \Big[ \nabla_m({\zeta}^\alpha{\zeta}^\gamma)\gamma^m{}_{\gamma\beta} \\ & \qquad \qquad + \kappa_m{}^\alpha{}_\gamma{\zeta}^\gamma {\zeta}^\delta\gamma^m{}_{\delta\beta} - {\zeta}^\alpha\gamma^m{}_{\beta\gamma} \kappa_m{}^\gamma{}_\delta{\zeta}^\delta - {\zeta}^\alpha\gamma^{m\,\gamma\delta} \kappa_{m\,\gamma\delta\beta\epsilon}{\zeta}^\epsilon \Big] \\ (D_0\cdot\phi_0)_{\alpha\beta\gamma\delta} &= -\frac{{\mathrm{i}}}{2\sqrt{2}} \kappa_{m\,[\alpha\beta\gamma|\epsilon} {\zeta}^\epsilon\gamma^m{}_{\delta]\theta}{\zeta}^\theta . \end{aligned}$$ Finally, again using the identity ${\zeta}^\alpha\gamma^m_{\alpha\beta}{\zeta}^\beta=0$ we find $$\label{eq:nearly} \begin{aligned} \left[(D_0\cdot\phi_0)\,\phi_0\right]^{\alpha\beta} &= -\frac{3{\mathrm{i}}}{4\sqrt{2}} \left( {\zeta}^\gamma\gamma^m{}_{\gamma\delta}\nabla_m{\zeta}^\delta + {\zeta}^\gamma\gamma^m_{\gamma\delta} \kappa_m{}^\delta{}_\epsilon{\zeta}^\epsilon \right){\zeta}^\alpha{\zeta}^\beta \\ &= -\frac{3{\mathrm{i}}}{4\sqrt{2}}\Big[ \bar{{\zeta}}^c\gamma^m(\nabla_m+\kappa_m){\zeta}\Big]\phi_0^{\alpha\beta} , \end{aligned}$$ with all other components vanishing. We see that, as claimed, $(D_0\cdot\phi_0)\,\phi_0$ is proportional to $\phi_0$. Finally we recall that $\kappa_m$ corresponded to an “$A$-shift” of $\nabla_mA$ and a “${\tilde{A}}$-shift” of $\nabla_m{\tilde{A}}-\frac{1}{2}\nabla_mA\wedge A$. Only the ${{\operatorname{\mathit{SU}}(8)/{\mathbb{Z}}_2}}$ adjoint component $\kappa_m{}^\alpha{}_\beta$ survives in . Using the decomposition  together with the definitions  we find this component is given by $$\kappa_m = \tfrac{1}{4\cdot 4!}(\nabla_m A)_{npq} \gamma^{npq} - \tfrac{1}{4}{\mathrm{i}}\big[ *_7\big( \nabla_m{\tilde{A}}-\tfrac{1}{2}\nabla_mA\wedge A \big)\big]_n\gamma^n$$ and hence $$\begin{aligned} \bar{{\zeta}}^c\gamma^m(\nabla_m + \kappa_m){\zeta}&= \bar{{\zeta}}^c\gamma^m\nabla_m{\zeta}+ \tfrac{1}{4\cdot 4!}{\mathcal{F}}_{mnpq}\bar{{\zeta}}^c \gamma^{mnpq}{\zeta}- {\mathrm{i}}(*_7{\tilde{{\mathcal{F}}}}) \bar{{\zeta}}^c{\zeta}\\ &= -{\mathrm{i}}{\mathrm{e}}^{-E}{\mathrm{e}}^{K/2}W , \end{aligned}$$ as required. We now return briefly to a subtlety in the definition of $D$. As written , $D$ is only defined given a metric $g$. However $\phi$ defines an $\operatorname{\mathit{SU}}(7)$ structure on $E$ and hence a metric $g$ (and form fields $A$ and ${\tilde{A}}$). Thus, as written, we can think of $D$ as defined in terms of $\phi$. This is in contrast with the type II ${{\operatorname{\mathit{SU}}(3)\times\operatorname{\mathit{SU}}(3)}}$ case. There the exterior derivative defined a natural generalised Dirac operator ${\mathrm{d}}:S^\pm(E)\to S^\mp(E)$ independent of the structure (or specifically any metric). It is also in contrast with the final result: the final superpotential can be written  using only the exterior derivative. The indication is that one can actually define $D$ independently of the metric, such that it has a sensible action on $\phi$. In this sense the differential EGG is set up before introducing any structure, as is that case in generalised geometry. One would also expect that such a $D$ is dual to the ECB in the same way that the exterior derivative on $S^\pm(E)$ is dual to the ordinary Courant bracket [@H-brack]. A subtlety in the generalised geometrical case is that the isomorphism between $S^\pm(E)$ and $\Lambda^*T^*M$ is not unique: the natural isomorphism is to $(\Lambda^dT^*M)^{-1/2}\otimes\Lambda^*T^*M$ [@Gualtieri]. Thus to define the exterior derivative one must rescale by something in $(\Lambda^dT^*M)^{-1/2}$. We expect something similar in the definitions of $\phi$ and $D$, removing the need for $(\det g)^{1/4}$ in . Conclusions {#sec:concl} =========== In this paper we have discussed an extension of generalised geometry applicable to eleven-dimensional supergravity and for which the symmetry group is the continuous U-duality group $E_{d(d)}$. The general form of such constructions was recently discussed by Hull [@chris]. Here we specifically focused on generic $N=1$ flux compactifications to four dimensions, for which the relevant symmetry group is $\operatorname{\mathit{E}_{7(7)}}$. We showed that $N=1$ supersymmetry implies that there is $\operatorname{\mathit{SU}}(7)$ structure on this “exceptional generalised geometry” (EGG) defined by an element $\phi$ in a particular orbit in the ${\mathbf{912}}$ representation of $\operatorname{\mathit{E}_{7(7)}}$. This is the analogue of the pair of generalised spinors $\Phi^\pm$, each defining a generalised complex structure, which characterise $N=2$ type II backgrounds. In the four-dimensional theory it encodes the chiral multiplet scalar degrees of freedom. As an application we showed that the superpotential for generic $N=1$ flux compactifications could be written as an $\operatorname{\mathit{E}_{7(7)}}$-invariant homogeneous, holomorphic function of $\phi$. In fact, almost all the objects appearing in generalised geometry have analogues in EGG. There is an exceptional generalised tangent space $E$, now combining vectors, two-forms, five-forms and an eight-index tensor. This is twisted by gerbes which capture the topological information encoded in the patching of the three-form supergravity potential $A$ and its dual ${\tilde{A}}$ – the analogues of the NS–NS $B$-field. There is also a natural exceptional Courant bracket, encoding the differential structure on $E$. The seven-dimensional metric $g$ and potentials $A$ and ${\tilde{A}}$ then combine to define an ${{\operatorname{\mathit{SU}}(8)/{\mathbb{Z}}_2}}$ structure on $E$. This is the analogue of the generalised metric, combining metric and $B$-field, that defines an $O(d)\times O(d)$ structure in generalised geometry. If the background has $N=1$ supersymmetry this structure is further refined to $\operatorname{\mathit{SU}}(7)$, while for type II six-dimensional $N=2$ backgrounds the generalised structure is ${{\operatorname{\mathit{SU}}(3)\times\operatorname{\mathit{SU}}(3)}}\subset O(6)\times O(6)$. These parallels are partially summarised in table \[tab:compare\]. There are a number of obvious extensions of this work one would like to consider. Directly related to the results described here are the questions, first, of the definition of the derivative operator $D$ and, secondly, of the form of the Kähler potential [@inprog]. As discussed in section \[sec:covW\], we expect that the definition  of $D$ can be replaced with one written in terms of the ordinary partial derivative without need for a metric. This should, in an appropriate sense following [@H-brack], be the dual of the exceptional Courant bracket defined in . As for the superpotential the Kähler potential (or rather the Kähler metric) can be calculated directly by identifying the four-dimensional kinetic terms in the decomposition of the eleven-dimensional theory. One would again expect the potential to be a $\operatorname{\mathit{E}_{7(7)}}$ invariant. We know from the work on type II theories [@GLW1; @GLW2] that ${\mathrm{e}}^{-K}$ is proportional to the metric density $\sqrt{\det g}$. A similarly relation here would be compatible with the expression  for $W$ being a holomorphic function of $\phi$. In the type II case we also find [@GLW1; @GLW2] that ${\mathrm{e}}^{-K}$ is proportional to the Hitchin functional [@GCY; @Hfunc]. This is true both for the ordinary geometrical case for backgrounds parametrised by an $\operatorname{\mathit{SU}}(3)$ structure and for the extension to the generalised geometrical Hitchin functional for $\operatorname{\mathit{SU}}(3,3)$ structures $\Phi^\pm$. Hitchin has already introduced a functional for a conventional $G_2$ structure. One expects that the $\operatorname{\mathit{E}_{7(7)}}$-invariant Kähler potential should be the EGG generalisation of this functional for $\operatorname{\mathit{SU}}(7)$ structures defined by $\phi$ which includes the potential $A$ and ${\tilde{A}}$ degrees of freedom. This also should be the natural generalisation of the action for topological M-theory [@topM]. As the generalised geometrical functionals necessarily arose at one-loop in the topological B-model [@PW], one conjecture is that the putative EGG $\operatorname{\mathit{SU}}(7)$ functional would appear at one-loop in topological M-theory. However, note that the extension could also be to the generalised geometrical $G_2\times G_2$ functional [@BMSS]. An obvious question to address is to consider not the four-dimensional effective theory but the on-shell supersymmetric backgrounds [@inprog]. For type II theories, satisfying the six-dimensional Killing spinor equations is equivalent to simple differential conditions on the ${{\operatorname{\mathit{SU}}(3)\times\operatorname{\mathit{SU}}(3)}}$ structures $(\Phi^+,\Phi^-)$ [@GMPT; @JW; @GMPT2; @JW2]. One expects a very similar relation in EGG for the $\operatorname{\mathit{SU}}(7)$ structure $\phi$ using the operator $D$. These “integrability” conditions on the generalised geometry and EGG structures are the generalisations of the $G$-structure and intrinsic torsion classification of supersymmetric backgrounds [@G-struc]. One can also repeat the analysis here for compactifications of eleven-dimensional supergravity to other dimensions, or for that matter for type II theories where the EGG “geometrises” the Ramond–Ramond degrees of freedom [@chris]. There will also be relations between backgrounds in different dimensions. For instance the four-dimensional type II $N=2$ backgrounds [@GLW1; @GLW2] should be encoded in the dimensional reduction of the $N=1$ backgrounds discussed here. In particular, one notes that in terms of the generalised structures $O(6,6)\subset\operatorname{\mathit{E}_{7(7)}}$ and for supersymmetry ${{\operatorname{\mathit{SU}}(3)\times\operatorname{\mathit{SU}}(3)}}\subset\operatorname{\mathit{SU}}(7)$. Let us end by noting two further connections. Hull’s work [@chris] was partly motivated by the existence of non-geometrical backgrounds [@Tfold]–[@Hull:2007jy]. The exceptional generalised tangent space  is patched not by generic elements of $\operatorname{\mathit{E}_{7(7)}}$ but only elements in the subgroup $\operatorname{\mathit{GL}}(7)\ltimes{G_{\text{closed}}(M)}$, that is the usual geometrical patching of the tangent spaces together with $A$- and ${\tilde{A}}$-shifts (by exact forms). It is very natural to extend the twisting to generic $\operatorname{\mathit{E}_{7(7)}}$ bundles as discussed in [@chris]. Such spaces cannot directly describe non-geometrical backgrounds since the underlying space $M$ is still a conventional manifold. Nonetheless they are closely connected to the doubled T-fold and U-fold geometries of [@Tfold]. It was argued recently [@DPST] that the generic four-dimensional $N=8$ gauged supergravity theories arise from compactification on a 56-dimensional “megatorus” U-fold. Interestingly such generic supergravities are encoded by an embedding tensor [@dWST] which, like $\phi$, lies in the ${\mathbf{912}}$ representation of $\operatorname{\mathit{E}_{7(7)}}$. Here, we have focused on $N=1$ theories rather than $N=8$ in four dimensions. The appearance of the ${\mathbf{912}}$ representation is nonetheless indirectly connected to the embedding tensor. For $N=8$, the embedding tensor appears (via the $T$-tensor) in the supersymmetry variations of the eight gravitinos and 56 spin-$\frac{1}{2}$ fields. Decomposing under $\operatorname{\mathit{SU}}(8)$ the ${\mathbf{36}}$ representation appears in the former and the ${\mathbf{420}}$ in the latter. In order to define an $N=1$ theory, we further decompose under $\operatorname{\mathit{SU}}(7)$. The ${\mathbf{36}}$ representation decomposes as ${\mathbf{36}}={\mathbf{1}}+{\mathbf{7}}+{\mathbf{28}}$. The first term goes with the $N=1$ gravitino and corresponds to the superpotential term . In addition the ${\mathbf{420}}$ representation decomposes as ${\mathbf{420}}={\mathbf{224}}+{\mathbf{140}}+{\mathbf{35}}+{\mathbf{21}}$. The chiral multiplets are the ${\mathbf{35}}$ representation (see table \[tab:SU7\]), and so the ${\mathbf{35}}$ term in the $T$-tensor corresponds the derivative of the superpotential with respect to the chiral scalars. Thus the structure $\phi$ and the $N=8$ embedding tensor share a common ${\mathbf{35}}$ representation under the $\operatorname{\mathit{SU}}(7)$ decomposition. We would like to thank Jerome Gauntlett, Jan Louis and, in particular, Mariana Graña and Chris Hull for helpful discussions. P. P. P. thanks the FCT (part of the Portuguese Ministry of Education) for financial support under scholarship SFRH/BD/10889/2002. D. W. is supported by a Royal Society University Research Fellowship and thanks the Aspen Center for Physics, CEA/Saclay and the Isaac Newton Institute for Mathematical Sciences for hospitality during the completion of this work. Conventions {#app:convs} =========== Eleven-dimensional supergravity {#app:sugra} ------------------------------- We adopt conventions where the eleven-dimensional supergravity action takes the form (see for instance [@HM]) $$\label{action11} S = \frac{1}{2\kappa^2}\int_{M_{11}} \sqrt{-g}\left(R - \bar{\Psi}_M\Gamma^{MNP}D_N\Psi_P\right) - \tfrac{1}{2} F \wedge *F - \tfrac{1}{6} A \wedge F \wedge F + \dots ,$$ while the variation of the gravitino $\Psi_M$ is given by $$\delta \Psi_M = \nabla_M \epsilon + \tfrac{1}{288}\left( \Gamma_M{}^{NPQR} - 8 \delta_M^N\Gamma^{PQR}\right)F_{MNPQ} + \dots .$$ (The dots represent four-fermi terms and terms coupling $\Psi_M$ and $F$.) Here $M,N=0,1,\dots,10$ are eleven-dimensional indices, the metric $g$ has signature $(-,+,\dots,+)$ and $\Gamma_M$ are the eleven-dimensional gamma matrices satisfying $$\left\{ \Gamma_M, \Gamma_N \right\} = 2 g_{MN} {\mathbb{1}},$$ with $\Gamma_{M_1\dots M_{11}}={\mathbb{1}}\epsilon_{M_1\dots M_{11}}$ where the volume form $\epsilon$ is given by $\epsilon=\sqrt{-g}\,{\mathrm{d}}x^0\wedge\dots\wedge{\mathrm{d}}x^{10}$. The spinors $\epsilon$ and $\Psi_M$ are Majorana. Given the intertwining relation $-\Gamma_M^T=C^{-1}\Gamma_MC$, we define the conjugate spinor $\bar{\Psi}_M=\Psi_M^TC^{-1}$. Note that the equation of motion and Bianchi identity for the flux $F$ read $$\label{eq:Feom} {\mathrm{d}}* F + \tfrac{1}{2}F\wedge F = 0 , \qquad {\mathrm{d}}F = 0 .$$ $\operatorname{Cliff}(7,0)$ and seven-dimensional spinors {#app:7d} --------------------------------------------------------- Let us also fix our conventions for $\operatorname{\mathit{Spin}}(7)$ (see also Appendix C of [@CJ]). The Clifford algebra $\operatorname{Cliff}(7,0;{\mathbb{R}})$ is generated by the gamma matrices $\gamma_m$ with $m=1,\dots,7$ satisfying $$\{ \gamma_m, \gamma_n \} = 2 g_{mn}{\mathbb{1}}.$$ One finds $\operatorname{Cliff}(7,0;{\mathbb{R}})\simeq\operatorname{\mathit{GL}}(8,{\mathbb{C}})$ and hence the spinor representation of the Clifford algebra is complex and eight-dimensional. We define the intertwiners $A$ and $C$ by $$\gamma_M^\dag = A\gamma_m A^{-1} , \qquad -\gamma_m^T = C^{-1}\gamma_m C ,$$ with $A^\dag=A$, $C^T=C$ and such that $-\gamma_m^*=D^{-1}\gamma_mD$ with $D=CA^T$. Given a spinor ${\zeta}$ we define the conjugate spinors $$\bar{{\zeta}}^c={{\zeta}^c}^\dag A , \qquad {\zeta}= D {{\zeta}^c}^* .$$ Furthermore, we write $$\gamma_{m_1\dots m_7}=\gamma_{(7)}\epsilon_{m_1\dots m_7}$$ with $\epsilon=\sqrt{g}{\mathrm{d}}x^1\wedge \dots \wedge{\mathrm{d}}x^7$. Note that one can choose the gamma matrices such that $\gamma_{(7)}={\mathrm{i}}$. The intertwiner $A$ provides an hermitian metric on the spinor space, which is invariant under the subgroup $\operatorname{\mathit{SU}}(8)\subset\operatorname{Cliff}(7,0;{\mathbb{R}})$, with a Lie algebra spanned by $\{\gamma_{m_1,m_2},\gamma_{m_1m_2m_3},\gamma_{m_1\dots m_6},\gamma_{m_1\dots m_7}\}$. The even part of the Clifford algebra generated by the $\gamma_{mn}$ has $\operatorname{Cliff}(7,0;{\mathbb{R}})_{\text{even}}\simeq\operatorname{\mathit{GL}}(8,{\mathbb{R}})$ and hence a real spinor representation with ${\zeta}={\zeta}^c$. Thus the spin group $\operatorname{\mathit{Spin}}(7)\subset\operatorname{Cliff}(7,0;{\mathbb{R}})_{\text{even}}$ similarly has a real spinor representation. For real spinors, $\bar{{\zeta}}={{\zeta}}^TC^{-1}$, and $C^{-1}$ provides metric on the spin space. This is invariant under a $\operatorname{\mathit{Spin}}(8)$ group with Lie algebra spanned by $\{\gamma_{m_1m_2},\gamma_{m_1\dots m_6}\}$. This can alternatively be described by, for $a,b=1,\dots 8$ $$\label{eq:gamma8def} {\hat{\gamma}}_{ab} = \begin{cases} \,\,\,\,(\det g)^{-1/4}\gamma_{mn} & \text{if $a=m$, $b=n$} \\ \,\,\,\,(\det g)^{1/4}\gamma_m\gamma_{(7)} & \text{if $a=m$, $b=8$} \\ -(\det g)^{1/4} \gamma_n\gamma_{(7)} & \text{if $a=8$, $b=n$} \end{cases}$$ which generate the $\operatorname{\mathit{Spin}}(8)$ Lie algebra with metric $${\hat{g}}_{ab} = (\det g)^{-1/4} \begin{pmatrix} g_{mn} & 0 \\ 0 & \det g \end{pmatrix} .$$ Here we have introduced some factors of $\det g$ to match the form of ${\hat{g}}$ used elsewhere in the paper (in particular the decomposition under $\operatorname{\mathit{GL}}(7,{\mathbb{R}})$ given in section \[app:GL7\]). With these conventions, the spinors ${\zeta}$ are of positive chirality with respect to $\operatorname{\mathit{Spin}}(8)$. If we make the spinor indices explicit writing ${\zeta}^\alpha$ with $\alpha=1,\dots,8$ we can raise and lower spinor indices using the metric $C^{-1}$ so, for instance, $$\label{eq:bigammas} {\hat{\gamma}}_{ab\,\alpha\beta} = C^{-1}_{\alpha\gamma}{\hat{\gamma}}_{ab}{}^\gamma{}_\beta , \qquad {\hat{\gamma}}_{ab}{}^{\alpha\beta} = {\hat{\gamma}}_{mn}{}^\alpha{}_\gamma C^{\gamma\beta} .$$ One also has the useful completeness relations, reflecting $\operatorname{\mathit{Spin}}(8)$ triality, $$\label{eq:complete} \begin{aligned} {\hat{\gamma}}_{ab}{}^{\alpha\beta}{\hat{\gamma}}^{ab}{}_{\gamma\delta} &= 16 \delta^{[\alpha}_{[\gamma}\delta^{\beta]}_{\delta]} , \\ {\hat{\gamma}}_{ab}{}^{\alpha\beta}{\hat{\gamma}}^{cd}{}_{\alpha\beta} &= 16 \delta^{[c}_{[a}\delta^{d]}_{b]} . \end{aligned}$$ The exceptional Lie group $\operatorname{\mathit{E}_{7(7)}}$ {#app:E7} ============================================================ In this appendix we review some of the properties of the group $\operatorname{\mathit{E}_{7(7)}}$ relevant to this paper. A detailed definition of $\operatorname{\mathit{E}_{7(7)}}$ and an fairly exhaustive description of its properties can be found in Appendix B of [@CJ] which itself refers to the original work by Cartan [@Cartan]. Definition and the ${\mathbf{56}}$ representation {#app:E7def} ------------------------------------------------- The group $\operatorname{\mathit{E}_{7(7)}}$ can be defined by its action on the basic 56-dimensional representation as follows. Let $W$ be a real 56-dimensional vector space with a symplectic product $\Omega$, then $\operatorname{\mathit{E}_{7(7)}}$ is a subgroup of $\operatorname{\mathit{Sp}}(56,{\mathbb{R}})$ leaving invariant a particular quartic invariant $q$. Explicitly one can define $\Omega$ and $q$ using the $\operatorname{\mathit{SL}}(8,{\mathbb{R}})\subset\operatorname{\mathit{E}_{7(7)}}$ subgroup. If $V$ is an eight-dimensional vector space, on which $\operatorname{\mathit{SL}}(8,{\mathbb{R}})$ acts in the fundamental representation, then the ${\mathbf{56}}$ representation decomposes as $${\mathbf{56}} = {\mathbf{28}} + {\mathbf{28'}}$$ the ${\mathbf{28}}$ representation corresponding to $\Lambda^2V$ and the ${\mathbf{28'}}$ to $\Lambda^2V^*$. Note that using $\epsilon\in\Lambda^8V$, the totally antisymmetric form preserved by the $\operatorname{\mathit{SL}}(8,{\mathbb{R}})$ action on $V$, one can identify $\Lambda^2V^*$ with $\Lambda^6V$. In summary, one identifies $$\label{eq:Wdef} W = \Lambda^2V \oplus \Lambda^2V^*$$ and writes $X\in W$ as the pair $(x^{ab},x'_{ab})$ where $a,b=1,\dots,8$. The symplectic product $\Omega$ is then given by, where $A=1,\dots,56$ $$\label{eq:Odef} \Omega(X,Y) = \Omega_{AB} X^A Y^B = x^{ab}y'_{ab} - x'_{ab}y^{ab} ,$$ and the quartic invariant $q$ is $$\label{eq:qdef} \begin{aligned} q(X) &= q_{ABCD} X^A X^B X^C X^D \\ &= x^{ab} x'_{bc} x^{cd} x'_{da} - \tfrac{1}{4} x^{ab}x'_{ab}x^{cd} x'_{cd} \\ & \qquad + \tfrac{1}{96} \left( \epsilon_{abcdefgh} x^{ab} x^{cd} x^{ef} x^{gh} + \epsilon^{abcdefgh} x'_{ab} x'_{cd}x'_{ef}x'_{gh} \right) \end{aligned}$$ In what follows it will be useful to use a matrix notation where we write $$X^A = \begin{pmatrix} x^{aa'} \\ x'_{aa'} \end{pmatrix}$$ such that the symplectic form $$\Omega_{AB} = \begin{pmatrix} 0 & \delta^b_a\delta^{b'}_{a'} \\ \delta^a_b\delta^{a'}_{b'} & 0 \end{pmatrix}$$ Throughout it is assumed that *all pairs of primed and unprimed indices $(a,a')$ etc are antisymmetrised*. The ${\mathbf{133}}$ and ${\mathbf{912}}$ representations {#app:reps} --------------------------------------------------------- There are two other representations of interest in this paper. First is the adjoint. By definition it is a 133-dimensional subspace $A$ of the Lie algebra $\operatorname{\mathit{sp}}(56,{\mathbb{R}})$. It decomposes under $\operatorname{\mathit{SL}}(8,{\mathbb{R}})$ as $$\begin{aligned} A &= (V\otimes V^*)_0 \oplus \Lambda^4V^* \\ \mu &= (\mu^a{}_b, \mu_{abcd}) \\ {\mathbf{133}} &= {\mathbf{63}} + {\mathbf{70}} , \end{aligned}$$ where $(V\otimes V^*)_0$ denotes traceless matrices, so $\mu^a{}_a=0$. The action on the ${\mathbf{56}}$ representation is given by $$\begin{aligned} \delta x^{ab} &= \mu^a{}_cx^{cb} + \mu^b{}_cx^{ac} + {{*\mu}}^{abcd} x'_{cd} , \\ \delta x'_{ab} &= - \mu^c{}_ax'_{cb} - \mu^c{}_bx'_{ac} + \mu_{abcd} x^{cd} , \end{aligned}$$ with ${{*\mu}}^{a_1\dots a_4}=\frac{1}{4!}\epsilon^{a_1\dots a_8}\mu_{a_5\dots a_8}$. In terms of the matrix notation we have $\delta X^A=\mu^A{}_BX^B$ with $$\mu^A{}_B = \begin{pmatrix} 2\mu^a{}_b\delta^{a'}_{b'} & {{*\mu}}^{aa'bb'} \\ \mu_{aa'bb'} & - 2\mu^b{}_a\delta^{b'}_{a'} \end{pmatrix} .$$ Note that $\mu^{AB}=\mu^A{}_C\Omega^{-1\, CB}$ is a symmetric matrix. Taking commutators of the adjoint action gives the Lie algebra $\mu^{\prime\prime}=[\mu,\mu']$ $$\begin{aligned} \mu^{\prime\prime\, a}{}_b &= ( \mu^a{}_c\mu^{\prime\, c}{}_b - \mu^{\prime\, a}{}_c\mu^c{}_b ) + \tfrac{1}{3} ( {{*\mu}}^{ac_1c_2c_3}\mu'_{bc_1c_2c_3} - {{*\mu}}^{\prime\, ac_1c_2c_3}\mu_{bc_1c_2c_3} ) , \\ \mu^{\prime\prime}_{abcd} &= 4 ( \mu^e{}_{[a}\mu'_{bcd]e} + \mu^{\prime\, e}{}_{[a}\mu_{bcd]e} ) . \end{aligned}$$ The other representation of interest in this paper is the ${\mathbf{912}}$. The representation space $N$ decomposes under $\operatorname{\mathit{SL}}(8,{\mathbb{R}})$ as $$\begin{aligned} N &= S^2V \oplus (\Lambda^3V\otimes V^*)_0 \oplus S^2V^* \oplus (\Lambda^3V^*\otimes V)_0 \\ \phi &= (\phi^{ab},\phi^{abc}{}_d,\phi'_{ab}, \phi'_{abc}{}^d) \\ {\mathbf{912}} &= {\mathbf{36}} + {\mathbf{420}} + {\mathbf{36'}} + {\mathbf{420'}} , \end{aligned}$$ where $S^nV$ denotes the symmetric product and $(\Lambda^3V\otimes V^*)_0$ denotes traceless tensors, so $\phi^{abc}{}_c=0$. The adjoint action of $\operatorname{\mathit{E}_{7(7)}}$ on $\phi$ is given by $$\label{eq:muphi} \begin{aligned} \delta\phi^{ab} &= \mu^a{}_c\phi^{cb} + \mu^b{}_c\phi^{ac} - \tfrac{1}{3}( {{*\mu}}^{acde}\phi'_{cde}{}^b + {{*\mu}}^{bcde}\phi'_{cde}{}^a ) , \\ \delta\phi^{abc}{}_d &= 3\mu^{[a}{}_e\phi^{bc]e}{}_d - \mu^e{}_d\phi^{abc}{}_e + {{*\mu}}^{abce}\phi'_{ed} + {{*\mu}}^{ef[ab}\phi'_{efd}{}^{c]} - {{*\mu}}^{efg[a}\phi'_{efg}{}^{b}\delta^{c]}_d , \\ \delta\phi'_{ab} &= - \mu^c{}_a\phi'_{cb} - \mu^c{}_b\phi'_{ac} - \tfrac{1}{3}(\mu_{acde}\phi^{cde}{}_b + \mu_{bcde}\phi^{cde}{}_a ) , \\ \delta\phi'_{abc}{}^d &= - 3\mu^e{}_{[a}\phi'_{bc]e}{}^d + \mu^d{}_e\phi'_{abc}{}^e + \mu_{abce}\phi^{ed} + \mu_{ef[ab}\phi^{efd}{}_{c]} - \mu_{efg[a}\phi^{efg}{}_{b}\delta_{c]}^d . \end{aligned}$$ In terms of $\operatorname{\mathit{Sp}}(56,{\mathbb{R}})$ indices, we have $\phi^{ABC}$, corresponding to the Young tableau [$\young(AC,B)$]{} with $\phi^{ABC}\Omega_{AB}=0$. The different components are given by $$\begin{aligned} \phi^{aa'bb'cc'} &= - \tfrac{1}{12}( \epsilon^{abb'cc'efg}\phi'_{efg}{}^{a'} - \epsilon^{baa'cc'efg}\phi'_{efg}{}^{b'} ) , \\ \phi^{aa'bb'}{}_{cc'} &= 2\phi^{ab}\delta_c^{a'}\delta_{c'}^{b'} - \phi^{aa'b}{}_c\delta^{b'}_{c'} + \phi^{bb'a}{}_c\delta^{a'}_{c'} , \\ \phi^{aa'}{}_{bb'}{}^{cc'} &= \phi^{ac}\delta^{a'}_b\delta^{c'}_{b'} - 2\phi^{aa'c}{}_b\delta_{b'}^{c'} - \phi^{cc'a}{}_b\delta^{a'}_{b'} , \end{aligned}$$ with $\phi^{aa'}{}_{bb'}{}^{cc'} = - \phi_{bb'}{}^{aa'cc'}$ and identical expressions for $\phi_{aa'bb'cc'}$ etc. but with raised and lowered indices reversed. Finally we will also need the tensor product $${\mathbf{56}} \times {\mathbf{912}} = {\mathbf{133}} + \dots .$$ In terms of $\operatorname{\mathit{Sp}}(56,{\mathbb{R}})$ indices we have $\mu^{AB}=X^C\Omega_{CD}\phi^{D(AB)}$, while in terms of $\operatorname{\mathit{SL}}(8,{\mathbb{R}})$ components one finds $$\label{eq:Xphi} \begin{aligned} \mu^a{}_b &= \tfrac{3}{4}\left( x^{ac}\phi'_{cb} - x'_{bc}\phi^{ca}\right) + \tfrac{3}{4} \left( x^{cd}\phi'_{cdb}{}^a - x'_{cd}\phi^{cda}{}_b \right) , \\ \mu_{abcd} &= - 3 \left(\phi'_{[abc}{}^ex'_{d]e} + \tfrac{1}{4!}\epsilon_{abcdm_1\dots m_4} \phi^{m_1m_2m_3}{}_ex^{m_4e} \right) . \end{aligned}$$ A $\operatorname{\mathit{GL}}(7)$ subgroup {#app:GL7} ------------------------------------------ As described in section \[sec:EGT\], the tangent space structure group embeds in the action of $\operatorname{\mathit{E}_{7(7)}}$ on the EGT. To make this embedding explicit we must identify a particular $\operatorname{\mathit{GL}}(7,{\mathbb{R}})\subset\operatorname{\mathit{SL}}(8,{\mathbb{R}})\subset\operatorname{\mathit{E}_{7(7)}}$ subgroup. In this appendix we identify this group and give explicit expressions for part of the $\operatorname{\mathit{E}_{7(7)}}$ action in terms of $\operatorname{\mathit{GL}}(7,{\mathbb{R}})$ (that is spacetime tensor) representations. We start with the embedding of $\operatorname{\mathit{GL}}(7,{\mathbb{R}})$ in $\operatorname{\mathit{SL}}(8,{\mathbb{R}})$ given by the matrix $$\begin{pmatrix} (\det M)^{-1/4} M^m{}_n & 0 \\ 0 & (\det M)^{3/4} \end{pmatrix} \in \operatorname{\mathit{SL}}(8,{\mathbb{R}}),$$ where $M\in\operatorname{\mathit{GL}}(7,{\mathbb{R}})$. If $\operatorname{\mathit{GL}}(7,{\mathbb{R}})$ acts linearly on the seven-dimensional vector space $F$ this corresponds to the decomposition of the eight-dimensional representation space $V$ as $$V = (\Lambda^7F)^{-1/4}F \oplus (\Lambda^7F)^{3/4} .$$ The ${\mathbf{56}}$ representation  of $\operatorname{\mathit{E}_{7(7)}}$ then decomposes as[^3] $$W = (\Lambda^7F^*)^{-1/2} \otimes \left[ F \oplus \Lambda^2F^* \oplus \Lambda^5F^* \oplus F^*\otimes(\Lambda^7F^*) \right] .$$ We can write an element of $W$ as $$X = x + \omega + \sigma + \tau ,$$ where $x\in (\Lambda^7F^*)^{-1/2}\otimes F$ etc. If we write the index $m=1,\dots,7$ for the fundamental $\operatorname{\mathit{GL}}(7,{\mathbb{R}})$ representation, note that, ignoring the tensor density factor $(\Lambda^7F^*)^{-1/2}$, $\tau$ has the index structure $\tau_{m,n_1\dots n_7}$, where $n$ labels the $F^*$ factor and $n_1\dots n_7$ the $\Lambda^7F^*$ factor. We can make the identification between $\operatorname{\mathit{GL}}(7,{\mathbb{R}})$ indices and $\operatorname{\mathit{SL}}(8,{\mathbb{R}})$ indices explicit by writing (again ignoring the $(\Lambda^7F^*)^{-1/2}$ factor) $$\begin{aligned} x^{m8} &= x^m & x^{mn} = \sigma^{mn}_{1\dots 7} \\ x'_{m8} &= \tau_{m,1\dots 7} & x'_{mn} = \omega_{mn} , \end{aligned}$$ where $\sigma^{mn}_{p_1\dots p_7}= (7!/5!)\delta^m_{[p_1}\delta^n_{p_2}\sigma_{p_3\dots p_7]}$. We can similarly decompose the ${\mathbf{133}}$ representation. We find $$\begin{aligned} A &= (V\otimes V^*)_0 \oplus \Lambda^4V^* \\ &= F\otimes F^* \oplus \Lambda^6F \oplus \Lambda^6F^* \oplus \Lambda^3F \oplus \Lambda^3F^* . \end{aligned}$$ We will be particularly interested in the action of the $\Lambda^3F^*$ and $\Lambda^6F^*$ parts of ${\mathbf{133}}$ on $X$. Identifying $$\label{eq:AA'def} \begin{aligned} \mu_{mnp8} &= \tfrac{1}{2}A_{mnp} \in \Lambda^3V^* \\ \mu^m{}_8 &= - A^{\prime m}_{1\dots 7} \in \Lambda^6V^* \end{aligned}$$ where $A^{\prime m}_{p_1\dots p_7} =(7!/6!)\delta^m_{[p_1}{\tilde{A}}_{p_2\dots p_7]}$ we have the action in the Lie algebra $$\label{eq:AA'action} (A+{\tilde{A}})\cdot X = i_xA + \big(i_x{\tilde{A}}+ A\wedge\omega\big) + \big(jA\wedge\sigma - j{\tilde{A}}\wedge\omega\big) .$$ Here we have introduced a new notation. The symbol $j$ denotes the first pure $F^*$ index of sections of $F^*\otimes (\Lambda^7F^*)$. Hence $$\label{eq:jdef} \left(j\alpha^{(p+1)}\wedge\beta^{(7-p)}\right)_{m,n_1\dots n_7} := \frac{7!}{p!(7-p)!} \alpha_{m[n_1\dots n_p}\beta_{n_{p+1}\dots n_7]}$$ The ${{\operatorname{\mathit{SU}}(8)/{\mathbb{Z}}_2}}$ subgroup and spinor indices {#app:SU8} ---------------------------------------------------------------------------------- The maximal compact subgroup of $\operatorname{\mathit{E}_{7(7)}}$ is ${{\operatorname{\mathit{SU}}(8)/{\mathbb{Z}}_2}}$. In the supersymmetry transformations, the spinors transform in the fundamental representation under (the double cover) $\operatorname{\mathit{SU}}(8)$. Thus it is often useful to have the decomposition of the various $\operatorname{\mathit{E}_{7(7)}}$ representations in terms of ${{\operatorname{\mathit{SU}}(8)/{\mathbb{Z}}_2}}$. In particular, one can use the common $\operatorname{\mathit{Spin}}(8)$ subgroup to relate the decompositions under $\operatorname{\mathit{SL}}(8,{\mathbb{R}})$ and ${{\operatorname{\mathit{SU}}(8)/{\mathbb{Z}}_2}}$. Let ${\hat{\gamma}}^{ab}$ be $\operatorname{\mathit{Spin}}(8)$ gamma matrices defined in . We can raise and lower $\operatorname{\mathit{SL}}(8,{\mathbb{R}})$ indices using the metric ${\hat{g}}$. Similarly, spinor indices can be raised a lowered using $C^{-1}$. Under ${{\operatorname{\mathit{SU}}(8)/{\mathbb{Z}}_2}}$ the ${\mathbf{56}}$ representation decomposes as $$\begin{aligned} X &= ( x^{\alpha\beta}, \bar{x}_{\alpha\beta} ) \\ {\mathbf{56}} &= {\mathbf{28}} + {\mathbf{\bar{28}}} . \end{aligned}$$ If the symplectic product takes the form $$\label{eq:OdefSU8} \Omega(X,Y) = {\mathrm{i}}\left( x^{\alpha\beta}\bar{y}_{\alpha\beta} - \bar{x}_{\alpha\beta}y^{\alpha\beta} \right)$$ then $(x^{ab},x'_{ab})$ and $(x^{\alpha\beta},\bar{x}_{\alpha\beta})$ are related by $$\begin{aligned} x^{\alpha\beta} &= \tfrac{1}{4\sqrt{2}}(x^{ab}+{\mathrm{i}}x^{\prime ab}) {\hat{\gamma}}_{ab}{}^{\alpha\beta} , \\ \bar{x}_{\alpha\beta} &= \tfrac{1}{4\sqrt{2}}(x^{ab}-{\mathrm{i}}x^{\prime ab}) {\hat{\gamma}}_{ab\,\alpha\beta} , \end{aligned}$$ with ${\hat{\gamma}}_{ab}{}^{\alpha\beta}$ and ${\hat{\gamma}}_{ab\,\alpha\beta}$ given by , or equivalently $$\label{eq:SLSU} \begin{pmatrix} x^{\alpha\beta} \\ \bar{x}_{\alpha\beta} \end{pmatrix} = \frac{1}{4\sqrt{2}} \begin{pmatrix} {\hat{\gamma}}_{ab}{}^{\alpha\beta} & {\mathrm{i}}{\hat{\gamma}}^{ab\,\alpha\beta} \\ {\hat{\gamma}}_{ab\,\alpha\beta} & - {\mathrm{i}}{\hat{\gamma}}^{ab}{}_{\alpha\beta} \end{pmatrix} \begin{pmatrix} x^{ab} \\ x'_{ab} \end{pmatrix}$$ Recall that the $\operatorname{\mathit{SU}}(8)$ subgroup of $\operatorname{Cliff}(7,0;{\mathbb{R}})$ leaves the norm $\bar{{\zeta}}{{\zeta}}$ invariant. Since the defining ${\mathbf{56}}$ representation decomposes as a spinor bilinear, both the ${\mathbb{1}}$ and $-{\mathbb{1}}$ elements in $\operatorname{\mathit{SU}}(8)$ leave $X$ invariant and hence we see explicitly that the subgroup of interest of $\operatorname{\mathit{E}_{7(7)}}$ is actually ${{\operatorname{\mathit{SU}}(8)/{\mathbb{Z}}_2}}$. Viewing a 56-dimensional index either as a pair of $\operatorname{\mathit{SL}}(8,{\mathbb{R}})$ indices or as a pair of spinor indices, the relation  can be used to convert between $\operatorname{\mathit{SL}}(8,{\mathbb{R}})$ and ${{\operatorname{\mathit{SU}}(8)/{\mathbb{Z}}_2}}$ decompositions of any other $\operatorname{\mathit{E}_{7(7)}}$ representations. In particular, decomposing the adjoint representation $\mu$ under ${{\operatorname{\mathit{SU}}(8)/{\mathbb{Z}}_2}}$ as ${\mathbf{133}}={\mathbf{63}}+{\mathbf{35}}+{\mathbf{\bar{35}}}$ and writing $\mu=(\mu^\alpha{}_\beta,\mu^{\alpha\beta\gamma\delta}, \bar{\mu}_{\alpha\beta\gamma\delta})$, with $\bar{\mu}_{\alpha\beta\gamma\delta}={{*\mu}}_{\alpha\beta\gamma\delta}$ and $\mu^\alpha{}_\alpha=0$, one finds $$\begin{aligned} \delta x^{\alpha\beta} &= \mu^\alpha{}_\gamma x^{\gamma\beta} + \mu^\beta{}_\gamma x^{\alpha\gamma} + \mu^{\alpha\beta\gamma\delta} \bar{x}_{\gamma\delta} , \\ \delta\bar{x}_{\alpha\beta} &= - \mu^\gamma{}_\alpha \bar{x}_{\gamma\beta} - \mu^\gamma{}_\beta \bar{x}_{\alpha\gamma} + \bar{\mu}_{\alpha\beta\gamma\delta} x^{\gamma\delta} , \end{aligned}$$ with $$\label{eq:SU8decomp} \begin{aligned} \mu^\alpha{}_\beta &= \tfrac{1}{4}\mu^-_{ab}{\hat{\gamma}}^{ab\,\alpha}{}_\beta - \tfrac{1}{48}{\mathrm{i}}\mu^-_{abcd}{\hat{\gamma}}^{abcd\,\alpha}{}_\beta , \\ \mu^{\alpha\beta\gamma\delta} &= \tfrac{1}{16}\big( 2\mu^+_{ac}g^{(8)}_{bd} + {\mathrm{i}}\mu^+_{abcd} \big) {\hat{\gamma}}^{ab\,[\alpha\beta}{\hat{\gamma}}^{cd\,\gamma\delta]} . \end{aligned}$$ where $\mu^\pm_{ab}=\tfrac{1}{2}(\mu_{ab}\pm\mu_{ba})$ and $\mu^\pm_{abcd}=\tfrac{1}{2}(\mu_{abcd}\pm *\mu_{abcd})$, and the anti-symmetrisation in the second line is only over $\alpha$, $\beta$, $\gamma$ and $\delta$. Note that in both lines the contraction with the relevant combination of gamma matrices automatically projects onto $\mu_{ab}^\pm$ and $\mu_{abcd}^\pm$, so one could in practice leave out the $\pm$ superscripts. Finally we can similarly introduce the decomposition of the ${\mathbf{912}}$ representation under ${{\operatorname{\mathit{SU}}(8)/{\mathbb{Z}}_2}}$, writing $\phi=(\phi^{\alpha\beta},\phi^{\alpha\beta\gamma}{}_\delta, \bar{\phi}_{\alpha\beta},\bar{\phi}_{\alpha\beta\gamma}{}^\delta)$. The adjoint action of $\operatorname{\mathit{E}_{7(7)}}$ then takes exactly the same form as in  but with spinor indices replacing $\operatorname{\mathit{SL}}(8)$ indices. We can also write the ${\mathbf{56}}\times{\mathbf{912}}\to{\mathbf{133}}$ product in terms of spinor indices. One find $$\label{eq:XphiSU8} \begin{aligned} \mu^\alpha{}_\beta &= \tfrac{3}{4}{\mathrm{i}}\left( x^{\alpha\gamma}\bar{\phi}_{\gamma\beta} - \bar{x}_{\beta\gamma}\phi^{\gamma\alpha}\right) + \tfrac{3}{4}{\mathrm{i}}\left( x^{\gamma\delta}\bar{\phi}_{\gamma\delta\beta}{}^\alpha - \bar{x}_{\gamma\delta}\phi^{\gamma\delta\alpha}{}_\beta \right) , \\ \mu_{\alpha\beta\gamma\delta} &= - 3{\mathrm{i}}\left( \bar{\phi}_{[\alpha\beta\gamma}{}^\epsilon \bar{x}_{\delta]\epsilon} + \tfrac{1}{4!} \epsilon_{\alpha\beta\gamma\delta \mu_1\dots \mu_4} \phi^{\mu_1\mu_2\mu_3}{}_\epsilon x^{\mu_4\epsilon} \right) . \end{aligned}$$ The additional factor of ${\mathrm{i}}$ as compared to the $\operatorname{\mathit{SL}}(8)$ expression  comes from the addition factor of ${\mathrm{i}}$ between the $\operatorname{\mathit{SL}}(8)$ and ${{\operatorname{\mathit{SU}}(8)/{\mathbb{Z}}_2}}$ expressions for the symplectic form  and  respectively. [99]{} N. Hitchin, “Generalized Calabi-Yau manifolds,” Quart. J. Math. Oxford Ser.  [**54**]{} (2003) 281 \[arXiv:math.dg/0209099\]. M. Gualtieri, “Generalized Complex Geometry,” Oxford University DPhil thesis (2004) \[arXiv:math.DG/0401221\]. N. Hitchin, “Instantons, Poisson structures and generalized Kähler geometry,” Commun. Math. Phys.  [**265**]{}, 131 (2006) \[arXiv:math.dg/0503432\]. N. Hitchin, “Brackets, forms and invariant functionals,” arXiv:math.dg/0508618. S. Gurrieri, J. Louis, A. Micu and D. Waldram, “Mirror symmetry in generalized Calabi-Yau compactifications,” Nucl. Phys.  B [**654**]{}, 61 (2003) \[arXiv:hep-th/0211102\]. M. Grana, R. Minasian, M. Petrini and A. Tomasiello, “Supersymmetric backgrounds from generalized Calabi-Yau manifolds,” JHEP [**0408**]{} (2004) 046 \[arXiv:hep-th/0406137\]. P. Grange and R. Minasian, “Modified pure spinors and mirror symmetry,” Nucl. Phys.  B [**732**]{}, 366 (2006) \[arXiv:hep-th/0412086\]. C. Jeschek and F. Witt, “Generalised $G_2$-structures and type IIB superstrings,” JHEP [**0503**]{} (2005) 053 \[arXiv:hep-th/0412280\]. P. Berglund and P. Mayr, “Non-perturbative superpotentials in F-theory and string duality,” arXiv:hep-th/0504058. M. Grana, R. Minasian, M. Petrini and A. Tomasiello, “Generalized structures of $N=1$ vacua,” JHEP [**0511**]{} (2005) 020 \[arXiv:hep-th/0505212\]. M. Graña, J. Louis and D. Waldram, “Hitchin functionals in $N = 2$ supergravity,” JHEP [**0601**]{} (2006) 008 \[arXiv:hep-th/0505264\]. P. Koerber, “Stable D-branes, calibrations and generalized Calabi-Yau geometry,” JHEP [**0508**]{}, 099 (2005) \[arXiv:hep-th/0506154\]. L. Martucci and P. Smyth, “Supersymmetric D-branes and calibrations on general $N=1$ backgrounds,” JHEP [**0511**]{}, 048 (2005) \[arXiv:hep-th/0507099\]. M. Graña, “Flux compactifications in string theory: A comprehensive review,” Phys. Rept.  [**423**]{}, 91 (2006) \[arXiv:hep-th/0509003\]. C. Jeschek and F. Witt, “Generalised geometries, constrained critical points and Ramond-Ramond fields,” arXiv:math/0510131. I. Benmachiche and T. W. Grimm, “Generalized $N=1$ orientifold compactifications and the Hitchin functionals,” Nucl. Phys.  B [**748**]{}, 200 (2006) \[arXiv:hep-th/0602241\]. J. Louis and S. Vaula, “$N=1$ domain wall solutions of massive type II supergravity as generalized geometries,” JHEP [**0608**]{}, 058 (2006) \[arXiv:hep-th/0605063\]. R. Minasian, M. Petrini and A. Zaffaroni, “Gravity duals to deformed SYM theories and generalized complex geometry,” JHEP [**0612**]{}, 055 (2006) \[arXiv:hep-th/0606257\]. M. Grana, R. Minasian, M. Petrini and A. Tomasiello, “A scan for new $N=1$ vacua on twisted tori,” JHEP [**0705**]{} (2007) 031 \[arXiv:hep-th/0609124\]. M. Grana, J. Louis and D. Waldram, “$\operatorname{\mathit{SU}}(3)\times\operatorname{\mathit{SU}}(3)$ compactification and mirror duals of magnetic fluxes,” arXiv:hep-th/0612237. A. Tomasiello, “Reformulating Supersymmetry with a Generalized Dolbeault Operator,” arXiv:0704.2613 \[hep-th\]. S. Morris, “Doubled geometry versus generalized geometry,” Class. Quant. Grav.  [**24**]{}, 2879 (2007). P. Koerber and D. Tsimpis, “Supersymmetric sources, integrability and generalized-structure compactifications,” JHEP [**0708**]{}, 082 (2007) \[arXiv:0706.1244 \[hep-th\]\]. D. Cassani and A. Bilal, “Effective actions and $N=1$ vacuum conditions from ${{\operatorname{\mathit{SU}}(3)\times\operatorname{\mathit{SU}}(3)}}$ compactifications,” JHEP [**0709**]{}, 076 (2007) \[arXiv:0707.3125 \[hep-th\]\]. A. Butti, D. Forcella, L. Martucci, R. Minasian, M. Petrini and A. Zaffaroni, “On the geometry and the moduli space of beta-deformed quiver gauge theories,” arXiv:0712.1215 \[hep-th\]. A. Kapustin, “Topological strings on noncommutative manifolds,” Int. J. Geom. Meth. Mod. Phys.  [**1**]{}, 49 (2004) \[arXiv:hep-th/0310057\]. A. Kapustin and Y. Li, “Topological sigma-models with $H$-flux and twisted generalized complex manifolds,” arXiv:hep-th/0407249. U. Lindstrom, M. Rocek, R. von Unge and M. Zabzine, “Generalized Kähler geometry and manifest $N = (2,2)$ supersymmetric nonlinear sigma-models,” JHEP [**0507**]{}, 067 (2005) \[arXiv:hep-th/0411186\]. R. Zucchini, “Generalized complex geometry, generalized branes and the Hitchin sigma model,” JHEP [**0503**]{} (2005) 022 \[arXiv:hep-th/0501062\]. V. Pestun and E. Witten, “The Hitchin functionals and the topological B-model at one loop,” Lett. Math. Phys.  [**74**]{}, 21 (2005) \[arXiv:hep-th/0503083\]. U. Lindstrom, M. Rocek, R. von Unge and M. Zabzine, “Generalized Kähler manifolds and off-shell supersymmetry,” Commun. Math. Phys.  [**269**]{}, 833 (2007) \[arXiv:hep-th/0512164\]. V. Pestun, “Topological strings in generalized complex space,” Adv. Theor. Math. Phys.  [**11**]{}, 399 (2007) \[arXiv:hep-th/0603145\]. A. Kapustin and E. Witten, “Electric-magnetic duality and the geometric Langlands program,” arXiv:hep-th/0604151. W. Merrell, L. A. P. Zayas and D. Vaman, “Gauged (2,2) Sigma Models and Generalized Kahler Geometry,” JHEP [**0712**]{}, 039 (2007) \[arXiv:hep-th/0610116\]. A. Kapustin and A. Tomasiello, “The general (2,2) gauged sigma model with three-form flux,” JHEP [**0711**]{}, 053 (2007) \[arXiv:hep-th/0610210\]. U. Lindstrom, M. Rocek, I. Ryb, R. von Unge and M. Zabzine, “New $N=(2,2)$ vector multiplets,” JHEP [**0708**]{}, 008 (2007) \[arXiv:0705.3201 \[hep-th\]\]. N. Halmagyi and A. Tomasiello, “Generalized Kähler Potentials from Supergravity,” arXiv:0708.1032 \[hep-th\]. P. Grange and S. Schafer-Nameki, “Towards mirror symmetry à la SYZ for generalized Calabi-Yau manifolds,” JHEP [**0710**]{}, 052 (2007) \[arXiv:0708.2392 \[hep-th\]\]. A. Lawrence, T. Sander, M. B. Schulz and B. Wecht, “Torsion and Supersymmetry Breaking,” arXiv:0711.4787 \[hep-th\]. C. M. Hull, “A geometry for non-geometric string backgrounds,” JHEP [**0510**]{}, 065 (2005) \[arXiv:hep-th/0406102\]. J. Shelton, W. Taylor and B. Wecht, “Nongeometric flux compactifications,” JHEP [**0510**]{} (2005) 085 \[arXiv:hep-th/0508133\]. A. Dabholkar and C. Hull, “Generalised T-duality and non-geometric backgrounds,” JHEP [**0605**]{} (2006) 009 \[arXiv:hep-th/0512005\]. C. M. Hull, “Global Aspects of T-Duality, Gauged Sigma Models and T-Folds,” JHEP [**0710**]{} (2007) 057 \[arXiv:hep-th/0604178\]. C. M. Hull, “Doubled geometry and T-folds,” JHEP [**0707**]{} (2007) 080 \[arXiv:hep-th/0605149\]. J. Shelton, W. Taylor and B. Wecht, “Generalized flux vacua,” JHEP [**0702**]{}, 095 (2007) \[arXiv:hep-th/0607015\]. P. Grange and S. Schafer-Nameki, “T-duality with $H$-flux: Non-commutativity, T-folds and $G\times G$ structure,” Nucl. Phys.  B [**770**]{}, 123 (2007) \[arXiv:hep-th/0609084\]. R. A. Reid-Edwards, “Geometric and non-geometric compactifications of IIB supergravity,” arXiv:hep-th/0610263. C. M. Hull and R. A. Reid-Edwards, “Gauge Symmetry, T-Duality and Doubled Geometry,” arXiv:0711.4818 \[hep-th\]. C. M. Hull, “Generalised geometry for M-theory,” arXiv:hep-th/0701203. E. Cremmer and B. Julia, “The $\operatorname{\mathit{SO}}(8)$ Supergravity,” Nucl. Phys.  B [**159**]{} (1979) 141. B. Julia in *Supergravity and Superspace*, eds. S. W. Hawking and M. Rocek, CUP, Cambridge (1981);\ B. Julia, “Infinite Lie Algebras In Physics,” in *Unified field theories and beyond : proceedings*, 5th Johns Hopkins Workshop on Current Problems in Particle Theory, Johns Hopkins, Baltimore (1981). C. M. Hull and P. K. Townsend, “Unity of superstring dualities,” Nucl. Phys.  B [**438**]{} (1995) 109 \[arXiv:hep-th/9410167\]. B. de Wit and H. Nicolai, “$D = 11$ Supergravity With Local $\operatorname{\mathit{SU}}(8)$ Invariance,” Nucl. Phys. B [**274**]{}, 363 (1986). H. Nicolai, “$D = 11$ Supergravity with Local $\operatorname{\mathit{SO}}(16)$ Invariance,” Phys. Lett.  B [**187**]{}, 316 (1987). K. Koepsell, H. Nicolai and H. Samtleben, “An exceptional geometry for $d=11$ supergravity?,” Class. Quant. Grav.  [**17**]{}, 3689 (2000) \[arXiv:hep-th/0006034\]. P. C. West, “$E_{11}$ and M theory,” Class. Quant. Grav.  [**18**]{}, 4443 (2001) \[arXiv:hep-th/0104081\]. P. C. West, “$E_{11}$, $\operatorname{\mathit{SL}}(32)$ and central charges,” Phys. Lett.  B [**575**]{}, 333 (2003) \[arXiv:hep-th/0307098\]. F. Riccioni and P. C. West, “E(11)-extended spacetime and gauged supergravities,” JHEP [**0802**]{}, 039 (2008) \[arXiv:0712.1795 \[hep-th\]\]. T. Damour, M. Henneaux and H. Nicolai, “E(10) and a ’small tension expansion’ of M theory,” Phys. Rev. Lett.  [**89**]{}, 221601 (2002) \[arXiv:hep-th/0207267\];\ A. Kleinschmidt and H. Nicolai, “E(10) and SO(9,9) invariant supergravity,” JHEP [**0407**]{}, 041 (2004) \[arXiv:hep-th/0407101\]. T. House and A. Micu, “M-theory compactifications on manifolds with $G_2$ structure,” Class. Quant. Grav.  [**22**]{} (2005) 1709 \[arXiv:hep-th/0412006\]. N. J. Hitchin, “Lectures on special Lagrangian submanifolds,” arXiv:math/9907034. A. Giveon, M. Porrati and E. Rabinovici, “Target space duality in string theory,” Phys. Rept.  [**244**]{}, 77 (1994) \[arXiv:hep-th/9401139\]. C. M. Hull, “Gravitational duality, branes and charges,” Nucl. Phys.  B [**509**]{}, 216 (1998) \[arXiv:hep-th/9705162\]. E. Cremmer, B. Julia, H. Lu and C. N. Pope, “Dualisation of dualities. II: Twisted self-duality of doubled fields and superdualities,” Nucl. Phys.  B [**535**]{} (1998) 242 \[arXiv:hep-th/9806106\]. A. Lukas and P. M. Saffin, “M-theory compactification, fluxes and $AdS_4$,” Phys. Rev. D [**71**]{} (2005) 046005 \[arXiv:hep-th/0403235\]. P. Pires Pacheco and D. Waldram, work in progress. N. J. Hitchin, “The geometry of three-forms in six and seven dimensions,” arXiv:math/0010054. R. Dijkgraaf, S. Gukov, A. Neitzke and C. Vafa, “Topological M-theory as unification of form theories of gravity,” Adv. Theor. Math. Phys.  [**9**]{}, 603 (2005) \[arXiv:hep-th/0411073\],\ N. Nekrasov, “A la recherche de la M-theorie perdue. Z theory: Casing M/F theory,” arXiv:hep-th/0412021,\ A. A. Gerasimov and S. L. Shatashvili, “Towards integrability of topological strings. I: Three-forms on Calabi-Yau manifolds,” JHEP [**0411**]{}, 074 (2004) \[arXiv:hep-th/0409238\]. J. de Boer, P. de Medeiros, S. El-Showk and A. Sinkovics, “$G_2$ Hitchin functionals at one loop,” arXiv:0706.3119 \[hep-th\]. J. P. Gauntlett, D. Martelli, S. Pakis and D. Waldram, “$G$-structures and wrapped NS5-branes,” Commun. Math. Phys.  [**247**]{}, 421 (2004) \[arXiv:hep-th/0205050\] ;\ J. P. Gauntlett and S. Pakis, “The geometry of $D=11$ Killing spinors,” JHEP [**0304**]{}, 039 (2003) \[arXiv:hep-th/0212008\]. G. Dall’Agata, N. Prezas, H. Samtleben and M. Trigiante, “Gauged Supergravities from Twisted Doubled Tori and Non-Geometric String Backgrounds,” arXiv:0712.1026 \[hep-th\]. B. de Wit, H. Samtleben and M. Trigiante, “On Lagrangians and gaugings of maximal supergravities,” Nucl. Phys.  B [**655**]{}, 93 (2003) \[arXiv:hep-th/0212239\];\ B. de Wit, H. Samtleben and M. Trigiante, “The maximal $D = 4$ supergravities,” JHEP [**0706**]{}, 049 (2007) \[arXiv:0705.2101 \[hep-th\]\]. E. Cartan “Oeuvres completes T1 , Part 1 (These),” Gauthier-Villars Paris 1952 P. C. West, “$E_{11}$ origin of brane charges and U-duality multiplets,” JHEP [**0408**]{}, 052 (2004) \[arXiv:hep-th/0406150\];\ P. C. West, “Brane dynamics, central charges and $E_{11}$,” JHEP [**0503**]{}, 077 (2005) \[arXiv:hep-th/0412336\]. [^1]: Note that there is a second possible way to embed $\operatorname{\mathit{GL}}(7,{\mathbb{R}})$, and hence choice for $E_0$, analogous to the choice of spin-structures of $O(d,d)$ [@Gualtieri], where $E_0$ is defined as in  except with an overall factor of $\Lambda^7TM/|\Lambda^7TM|$. This bundle has a similar isomorphism to  but with $TM$ and $T^*M$ exchanged everywhere. [^2]: Note that $\hat{\rho}$ can be determined as a homogeneous function of $\rho$ of degree one. [^3]: Note that, up to the $(\Lambda^7F^*)^{-1/2}$ factor, the same embedding of $\operatorname{\mathit{GL}}(7,{\mathbb{R}})$ was identified in [@West-decomp]. This followed from an analysis of $E_{11}$ representations in [@West2].

ArXiv

--- abstract: 'We investigate the a matrix-type coring associated to a complete covering of an algebra, its Amitsur complex and propose a definition for the related Čech cohomology relative to the covering.' author: - | Andrzej Sitarz [^1] [^2]\ Institute of Physics, Jagiellonian University,\ Reymonta 4, 30-059 Kraków, Poland title: | On matrix type corings, algebra coverings\ and Čech cohomology --- Introduction ============ The definition and examples of algebra coverings, as given by Calow and Matthes [@CaMa] provided a nice set of examples of canonical corings [@BrWr]. In this note, we elaborate further on the construction and define a version of Čech cohomology respective to the covering using the cochain complexes based on Amitsur complex. Preliminaries ------------- We recall here the notation and definition. Let $A$ be a unital algebra and $I_i \subset A$, $i =1,2,\ldots N$, be two-sided ideals in $A$. We call $A_j = A/I_j$ and $A_{ij} = A/(I_i + I_j)$, further $B= \bigoplus_i A_i$ and $B' = \bigoplus_{i,j} A_{ij}$. The ideals $\{I_i\}_{i \in {\mathcal{I}}}$ are called a covering of $A$ if their intersection is $\{ 0 \}$. We call the covering [*complete*]{} if the following sequence of $A$-bimodules is exact: $$\xymatrix{0 \ar[r] & A \ar[r]^\pi &B \ar[r]^\tau & B'}$$ where $$\pi = \bigoplus_i \pi_i, \,\, \pi_i: A \to A_i,$$ and $$\tau: \bigoplus_i A_i \to \bigoplus_{i,j} A_{ij}: \tau(\bigoplus_i x_i) = \oplus_{i,j} \left( \pi^i_{ij} (x_i) - \pi^j_{ij} (x_j), \right),$$ where $\pi^i_{ij}: A/I_i \to A/(I_i+I_j)$. The coring ${\mathcal{C}}(A,\{I\}_{\mathcal{I}})$ associated to the covering $A, \{I_i\}_{i \in {\mathcal{I}}}$ is the Sweedler coring of the algebra extension $A \hookrightarrow B$. The coring ${\mathcal{C}}$ is a generalization of a matrix coring, where the $(i,j)$-matrix entry takes values in $A_i {\otimes}_A A_j$. The latter can be, for unital algebras identified (as $A$-bimodule) with $A_{ij}$. If we denote $e_{ij} = \pi^i(1_A) {\otimes}_A \pi^j(1_A)$ then for the coproduct and counit in ${\mathcal{C}}$ we have: $$\begin{aligned} {\Delta}e_{ij} &= \pi^i(1_A) {\otimes}_B \left( \sum_k \pi^k(1_A) \right) {\otimes}_A \pi_j(1_A) \\ &= \sum_k \left( \pi^i(1_A) {\otimes}_A \pi^k(1_A) \right) {\otimes}_B \left( \pi^k(1_A) {\otimes}_A \pi^j(1_A) \right) \\ &= \sum_k e_{ik} {\otimes}_B e_{kj}. \\ \varepsilon(e_{ij}) &= \delta_{ij} \pi_i(1_A). \end{aligned}$$ Note that although formally the coproduct formula is the same as in the case of a matrix coring, only in the case $A_i \equiv A, \, i \in I$, we recover the usual definition of the matrix coring. We have: The coring ${\mathcal{C}}$ is Galois if the covering is complete. The Čech cohomology of noncommutative coverings =============================================== In the classical situation of coverings of a topological space with open sets, with some additional data, we can easily introduce the notion of Čech cochain complex relative to the covering. We show that in the noncommutative situation, this is also possible and we relate the construction to the Amitsur complex of the coring related to the covering. However, for the purpose of the construction we need to adapt the notion of ringed space to the algebra case and define the corresponding complex. Čech complex of rings --------------------- Let ${\mathcal{I}}$ be a finite set of indices and $P({\mathcal{I}})$ be the set of its ordered subsets. We say that for any two ordered subsets $\zeta,\vartheta$ of ${\mathcal{I}}$, $\zeta \subset \vartheta$ if $\zeta$ is a subset of $\vartheta$ and the order in $\zeta$ is inherited from the order in $\vartheta$. We shall use $|\zeta|$ to denote cardinality of the subset $\zeta$. We assume that ${\mathcal{R}}$ is a functor from $P({\mathcal{I}})$ to the category of unital rings. That is, for any $\vartheta \in P({\mathcal{I}})$ there is a unital ring ${\mathcal{R}}(\vartheta)$, and for any $\zeta, \vartheta \in P({\mathcal{I}})$ such that $\zeta \subset \vartheta$ there is a ring homomorphism: $$r_\vartheta^\zeta: R(\zeta) \to R(\vartheta).$$ We assume that $r_\zeta^\zeta$ is always an identity morphism, and that for any three $\zeta \subset \eta \subset \vartheta \subset P({\mathcal{I}})$ we have: $$r^{\eta}_\vartheta \circ r_\zeta^\eta = r_{\zeta}^\eta. \label{r1}$$ We shall identify $R = {\mathcal{R}}(\emptyset)$. Since ${\mathcal{R}}(\emptyset) \hookrightarrow \bigoplus_i {\mathcal{R}}(\{i\})$, is an embedding, we have a similar situation of a ring extension as before and we can associate a Sweedler coring with the construction, which we shall call ${\mathcal{C}}({\mathcal{I}})$. The Amitsur complex of a coring ------------------------------- We recall here the definition of the Amitsur complex for the canonical Sweedler coring ${\mathcal{C}}= B {\otimes}_A B$ [@BrWi]. We use similar notation as before, here $A_{\mathcal{R}}= {\mathcal{R}}(\emptyset)$ and $B_{\mathcal{R}}= \bigoplus_i {\mathcal{R}}(\{i\})$. Let $\{C^n\}_{n=0,\ldots}$ be the following complex: $$C^0 = B_{\mathcal{I}}, \;\;\;\; C^n = B_{\mathcal{I}}^{\otimes_{A_{\mathcal{R}}} (n-1)} \cong {\mathcal{C}}^{\otimes_{B_{\mathcal{R}}} n},\; n \geq 1,$$ with the map: $$\begin{aligned} d (x) &= 1 {\otimes}_{A_{\mathcal{R}}} x - x {\otimes}_{A_{\mathcal{R}}} 1, \;\; x \in C^0, \\ d (\omega) &= 1 {\otimes}_{A_{\mathcal{R}}} \omega - b'(\omega) + (-1)^{n+1} \omega {\otimes}_{A_{\mathcal{R}}} 1, \;\; \omega \in C^n,\; n\geq 1, \end{aligned}$$ where $b'$ is the coboundary of the Hochschild bar complex constructed with the coring coproduct: $$b' (c_0 {\otimes}_{B_{\mathcal{R}}} c_1 {\otimes}_{B_{\mathcal{R}}} \cdots {\otimes}_{B_{\mathcal{R}}} c_n) = \sum_{i=0}^n (-1)^n c_0 {\otimes}_{B_{\mathcal{R}}} c_1 {\otimes}_{B_{\mathcal{R}}} \cdots \Delta c_i \cdots {\otimes}_{B_{\mathcal{R}}} c_n.$$ which, for the Sweedler corings is expressed using the equivalent presentation as: $$\begin{aligned} b'( x_0 {\otimes}_{A_{\mathcal{R}}} x_1 {\otimes}_{A_{\mathcal{R}}} \cdots {\otimes}_{A_{\mathcal{R}}} x_{n+1}) = \\ \sum_{i=0}^n (-1)^n x_0 {\otimes}_{A_{\mathcal{R}}} x_1 {\otimes}_{A_{\mathcal{R}}} \cdots x_i {\otimes}_{A_{\mathcal{R}}} 1 {\otimes}_{A_{\mathcal{R}}} x_{i+1} \cdots {\otimes}_{A_{\mathcal{R}}} x_{n+1}. \end{aligned}$$ We have (see [@BrWi] 29.5) The Amitsur complex of a Galois coring is acyclic. Hence the Amitsur complex of a coring related to a complete covering is acyclic. Therefore, to obtain more information, which arises from the covering we need to construct a different cochain complex, associated to the functor ${\mathcal{R}}$. Let $\phi$ be a map between the complex of the coring of covering ${\mathcal{C}}({\mathcal{R}})$ and the ${\mathbb{Z}}$-graded module $\bigoplus_{n \in {\mathbb{Z}}} S^n$, where $$S^n = \bigoplus_{\zeta \in P({\mathcal{I}}), |\zeta|=n} {\mathcal{R}}(\zeta).$$ defined as: $$\phi(y_{i_1} {\otimes}_{B_{\mathcal{R}}} y_{i_2} \cdots {\otimes}_{B_{\mathcal{R}}} y_{i_n}) = r^{i_1}_\zeta( \Phi_{I_1}(y_{i_1})) r^{i_2}_\zeta(\Phi_{I_2}(y_{i_2})) \cdots r^{i_n}_\zeta(\Phi_{I_n}(y_{i_n})),$$ for any $\zeta = \{ i_1, i_2, \ldots, i_n \} \in P(I)$, $y_{i_k} \in A_{i_k}$, and: $$\phi(y_{i_1} {\otimes}_{B_{\mathcal{R}}} y_{i_2} \cdots {\otimes}_{B_{\mathcal{R}}} y_{i_n}) \equiv 0,$$ for any $\{ i_1, i_2, \ldots, i_n \}$ which are not in $P({\mathcal{I}})$[^3]. The maps $\Phi_{i_k}$ denote some chosen ring homomorphisms between $A_{i_k}$ and ${\mathcal{R}}(\{i_k\})$, in fact, we have $\phi(y_i) \equiv \Phi_{i}(y_i)$. $S^n$ is a cochain complex, with the coboundary: map $d': S^n \to S^{n+1}$: $$d' x_\zeta = \sum_{\eta, \zeta \subset \eta, |\eta| = |\zeta|+1} (-1)^{|\eta/\zeta|} r_\eta^\zeta(x_\zeta),$$ where $|\eta/\zeta|$ denotes position on which the difference between $\eta$ and $\zeta$ occurs. Then $\phi$ is a morphism of cochain complexes. We call the cochain complex $S({\mathcal{R}}) = \{(S^n, d')\}_n$ the Čech complex associated to functor ${\mathcal{R}}$. First of all, let us see that the map $\phi$ is well-defined (the definition of $\phi$ is set for tensor products and we need to check that it remains correct for tensor products over $B_{\mathcal{R}}$.) We have: $$\begin{aligned} \phi( &\cdots y^{i_k} \pi_{i_k}(y) {\otimes}_{B_{\mathcal{R}}} y^{i_{k+1}} \cdots ) = \cdots r^{i_k}_\zeta \left( \Phi_{i_k}( y_{i_k} \pi_{i_k}(y) ) \right) r^{i_{k+1}}_\zeta \left( \Phi_{i_{k+1}}( y_{i_{k+1}})\right) \cdots \\ &= \cdots r^{i_k}_\zeta \left( \Phi_{i_k}( y_{i_k}) \Phi_{i_k}(\pi_{i_k}(y)) \right) r^{i_{k+1}}_\zeta \left( \Phi_{i_{k+1}}( y_{i_{k+1}})\right) \cdots \\ &= \cdots r^{i_k}_\zeta \left( \Phi_{i_k}( y_{i_k}) \right) r^{i_k}_\zeta \left( \Phi_{i_k}(\pi_{i_k}(y)) \right) r^{i_{k+1}}_\zeta \left( \Phi_{i_{k+1}}( y_{i_{k+1}})\right) \cdots \\ &= \cdots r^{i_k}_\zeta \left( \Phi_{i_k}( y_{i_k}) \right) r^{i_{k+1}}_\zeta \left( \Phi_{i_{k+1}}(\pi_{i_{k+1}}(y)) \right) r^{i_{k+1}}_\zeta \left( \Phi_{i_{k+1}}( y_{i_{k+1}})\right) \cdots \\ &= \cdots r^{i_k}_\zeta \left( \Phi_{i_k}( y_{i_k}) \right) r^{i_{k+1}}_\zeta \left( \Phi_{i_{k+1}}(\pi_{i_{k+1}}(y) y_{i_{k+1}}) \right) \cdots \\ &= \phi(\cdots y^{i_k} {\otimes}_{B_{\mathcal{R}}} r_{i_{k+1}}(y) y^{i_{k+1}} \cdots ). \end{aligned}$$ As the next step, we check that the map $d'$ is a coboundary: $$\begin{aligned} (d')^2 x_\zeta &= d' \left( \sum_{\eta, \zeta \subset \eta, |\eta| = |\zeta|+1} (-1)^{|\eta/\zeta|} r_\eta^\zeta(x_\zeta) \right) \\ &= \sum_{\vartheta, \vartheta \subset \zeta, |\zeta| = |\vartheta|+1} \sum_{\eta, \zeta \subset \eta, |\eta| = |\zeta|+1} (-1)^{|\vartheta/\zeta|} (-1)^{|\eta/\zeta|} r^\eta_\vartheta \circ r^\zeta_\eta (x_\zeta) \\ &= \sum_{\vartheta, \vartheta \subset \zeta, |\zeta| = |\vartheta|+2} \left( (-1)^{|\eta_1/\zeta|+|\vartheta/\eta_1|} + (-1)^{|\eta_2/\zeta|+|\vartheta/\eta_2|} \right) r^\zeta_\vartheta (x_\zeta) \\ &=0, \end{aligned}$$ where we have used (\[r1\]) and the fact that for each $\vartheta \subset \zeta$ such that $|\vartheta|+2 = |\zeta|$ there are two possibilities for $\vartheta \subset \eta_{1,2} \subset \zeta$ with $|\eta|=|\vartheta|+1$, which we denoted $\eta_1, \eta_2$. It is easy to verify that: $$(-1)^{|\eta_1/\zeta|+|\vartheta/\eta_1|} + (-1)^{|\eta_2/\zeta|+|\vartheta/\eta_2|} = 0,$$ which ends the verification. Finally, to verify that $\phi$ is a cochain complex morphism we calculate now using $x_i = \phi(y_i)$. First: $$\begin{aligned} d'\circ &\phi (x_{i_0} {\otimes}_R x_{i_1} \cdots {\otimes}_R x_{i_n}) = d' \left( r^{i_0}_\zeta(x_{i_0}) r^{i_1}_\zeta(x_{i_1}) \cdots r^{i_n}_\zeta(x_{i_n}) \right) \\ &= \sum_{i \in I/\zeta} \sum_{k=0}^{n+1} (-1)^k r^{\eta(k)}_\zeta \left( r^{i_0}_\zeta(x_{i_1}) r^{i_1}_\zeta(x_{i_1}) \cdots r^{i_n}_\zeta(x_{i_n}) \right) \\ &= \sum_{i \in I/\zeta} \sum_{k=0}^{n+1} (-1)^k r^{i_0}_{\eta(k)}(x_{i_0}) r^{i_1}_{\eta(k)}(x_{i_1}) \cdots r^{i_n}_{\eta(k)}(x_{i_n}) \end{aligned}$$ where ${\eta(k)}$ denotes the ordered subset of $P(I)$, containing $\zeta$ and $i \notin \zeta$ on $k\!+\!1$-th place. On the other hand, calculating $ \phi \circ d$, $$\begin{aligned} \phi \circ & d (x_{i_0} {\otimes}_R x_{i_2} \cdots {\otimes}_R x_{i_n}) = \\ &= \phi \left( \sum_{k=0}^{n+1} (-1)^k \phi(x_{i_0} {\otimes}_R x_{i_1} \cdots x_{i_{k-1}} {\otimes}_R (\sum_{i=1}^N r_i(1)) {\otimes}_R x_{i_k} \cdots {\otimes}_R x_{i_n}) \right) \\ &= \sum_{k=0}^{n+1} r^{i_0}_{\eta(k)}(x_{i_0}) r^{i_1}_{\eta(k)}(x_{i_1}) \cdots \left( \sum_{i \in I/\zeta} r^i_{\eta(k)}(1) \right) \cdots r^{i_n}_{\eta(k)}(x_{i_n}) \\ &= \sum_{i \in I/\zeta} \sum_{k=0}^{n+1} r^{i_0}_{\eta(k)}(x_{i_0}) r^{i_1}_{\eta(k)}(x_{i_1}) \cdots r^{i_n}_{\eta(k)}(x_{i_n}) \end{aligned}$$ which ends the proof. The trivial example of the functor ${\mathcal{R}}$ is given by a constant functor associating ring $R$ to every $\zeta \in P({\mathcal{I}})$. The coring ${\mathcal{C}}({\mathcal{R}})$ is then a full-matrix coring. The homology of the associated Čech cochain complex is: $$H^0(S) = R, \;\; H^1(S)= 0.$$ If $X$ is a ringed space with the functor $\Psi$ from open subsets of $X$ to the category of unital rings, and $U_i$, $i \in {\mathcal{I}}$ is a finite covering of $X$, then we construct ${\mathcal{R}}$ by setting: $${\mathcal{R}}(\{i_1,i_2,\ldots, i_k \}) = \Psi(U_{i_1} \cap U_{i_2} \cdots \cap U_{i_k}).$$ In the next section we shall discuss the relation of the constructed complex with the standard Čech cohomology. Čech cohomology of ringed algebras ---------------------------------- One of the examples of the functor ${\mathcal{R}}$ for a finite set ${\mathcal{I}}$ was based on the structure of a ringed space and a finite covering by open sets. We adapt the definition to the algebraic case: We call the algebra $A$ [*ringed*]{} if there exists a functor $\Phi$ which associates to any ideal $J \subset A$ a ring $\Phi(J)$, and a ring morphism $\Phi_J: A/J \to R(J)$, such that if $J_1 \subset J_2$ then the following diagram is commutative: $$\xymatrix{ A/J_1 \ar[d]^\Phi_{J_1} \ar[r]^\pi & A/J_2 \ar[d]^\Phi_{J_2} \\ \phi(J_1) \ar[r]^{\Phi(\pi)} & \Phi(J_2) }$$ An example of the ringed algebra is given by taking $\Phi(J) = A/J$ and $\Phi_J$ the identity map. Assume now that $A$ is a ringed algebra with the ringed structure $\Phi$ and $\{ J_i \}_{i \in {\mathcal{I}}}$ is a finite complete covering of $A$. Then by setting for every ordered subset of ${\mathcal{I}}$, $\zeta=\{i_1,\ldots, i_k\} \in P({\mathcal{I}})$ $${\mathcal{R}}(\zeta) = \Phi(I_{i_1} + \cdots + I_{i_k}),$$ and for the respective homomorphisms $r^\eta_\zeta: R(\eta) \to R(\zeta)$ to be images under $\Phi$ of projection morphisms. we obtain a functor ${\mathcal{R}}({\mathcal{I}})$. If $(A, \Phi)$ is a ringed algebra, and $\{I_i\}_{i \in {\mathcal{I}}}$ is its finite covering, we define the Čech cohomology of $A$ relative to the covering as the homology of the Čech cochain complex of the covering coring. Examples and applications ========================= Let $X$ be a topological space, and ${\cal O}_X$ be a presheaf of rings on $X$. We denote the functor associating a ring to any open set $U$ by $\Phi(U)$, and the ring morphisms corresponding to set inclusion $i_{U,V}: U \hookrightarrow V$ by $\Phi_U^V : \Phi(V) \to \Phi(U)$. Let $U_i$, $i \in I$ be a finite covering of $X$. Then from the presheaf ${\cal O}_X$ we can construct the data of a ringed space as in the previous section, the coring, the map $\phi$ and the respective cochain complexes. We have: The homology of the cochain complex $S({\mathcal{R}})$ for the coring given by the data $I, {\cal O}_X$, is the Čech cohomology of $X$ relative to the covering $\{U_i\}_I$. By rewriting the definition of $S^n,d'$ using the sets $U_i$ and the intersections, we explicitly recover the definition of Čech homology. The construction of Čech cohomology is possible for any matrix-type coring, by using the canonical functor ringed structure $\Phi$ and the covering implicitly defined by the coring construction. Conclusions =========== The formulation of [*noncommutative coverings*]{} in the language of algebra extensions and associated corings has enabled us to extend the definition of Čech cohomology to this setting. The Čech cohomology is mostly of combinatorial nature and strongly depends on the properties of the underlying noncommutative covering. It would be interesting to know, whether for some types of noncommutative coverings one could establish relation between this cohomology and some other cohomology theory of the algebra. [9]{} D. Calow, R. Matthes, Covering and gluing of algebras and differential algebras. J. Geom. Phys. 32 (2000), no. 4, 364–396. T, Brzeziński, R.Wisbauer, Corings and comodules. London Mathematical Society Lecture Note Series, 309. Cambridge University Press, Cambridge, 2003 T. Brzeziński, A.P.Wrightson, Complete coverings and Galois corings. Int. J. Geom. Methods Mod. Phys. 2 (2005), no. 5, 751–757 Bott, R., Tu, L. W.: Differential forms in algebraic topology. Springer-Verlag, Berlin, 1982 [^1]: The author acknowledges the Alexander von Humboldt Fellowship at the Mathematisches Institut der Heinrich-Heine-Universität, Universitätsstrasse 1, 40225 Düsseldorf, Germany [^2]: Partially supported by Polish Government grants 115/E-343/SPB/6.PR UE/DIE 50/2005–2008 and 189/6.PRUE/2007/7 [^3]: This happens when at least two of the indices are identical.

ArXiv

--- abstract: 'We present a new derivation of the equations governing the oscillations of slowly rotating relativistic stars. Previous investigations have been mostly carried out in the Regge–Wheeler gauge. However, in this gauge the process of linearizing the Einstein field equations leads to perturbation equations which as such cannot be used to perform numerical time evolutions. It is only through the tedious process of combining and rearranging the perturbation variables in a clever way that the system can be cast into a set of hyperbolic first order equations, which is then well suited for the numerical integration. The equations remain quite lengthy, and we therefore rederive the perturbation equations in a different gauge, which has been first proposed by Battiston et al. (1970). Using the ADM formalism, one is immediately lead to a first order hyperbolic evolution system, which is remarkably simple and can be numerically integrated without many further manipulations. Moreover, the symmetry between the polar and the axial equations becomes directly apparent.' author: - | Johannes Ruoff, Adamantios Stavridis and Kostas D. Kokkotas\ Department of Physics, Aristotle University of Thessaloniki, Thessaloniki 54006, Greece date: 'Accepted ???? Month ??. Received ???? Month ??; in original form ???? Month ??' title: Evolution equations for the perturbations of slowly rotating relativistic stars --- == == == == \#1[[ \#1]{}]{} \#1[[ \#1]{}]{} @mathgroup@group @mathgroup@normal@group[eur]{}[m]{}[n]{} @mathgroup@bold@group[eur]{}[b]{}[n]{} @mathgroup@group @mathgroup@normal@group[msa]{}[m]{}[n]{} @mathgroup@bold@group[msa]{}[m]{}[n]{} =“019 =”016 =“040 =”336 ="33E == == == == \#1[[ \#1]{}]{} \#1[[ \#1]{}]{} == == == == \[firstpage\] relativity – methods: numerical – stars: neutron – stars: oscillations – stars: rotation Introduction ============ The theory of non-radial perturbations of relativistic stars has been a field of intensive study for more than three decades, beginning with the pioneering paper of Thorne & Campolattaro in 1967. Their work was in turn based on previous studies of black hole perturbations initiated by Regge & Wheeler in 1957. Because of the quite involved and tedious computations of the perturbed field equations, the main focus remained on non-rotating neutron stars, although the foundations for computing rotating relativistic stellar models have already been laid by Hartle in 1967. Only in the early 90s, work began to shift to the perturbations of rotating relativistic neutron stars. In most works, the slow-rotation approximation was used to tackle the problem. Chandrasekhar & Ferrari (1991) studied the axisymmetric perturbations, where they established the coupling between the polar and axial modes induced by the rotation (polar or even parity modes are characterised by a sign change under parity transformation according to $(-1)^l$ while the axial ones change as $(-1)^{l+1}$). Shortly later, Kojima (1992) presented the first complete derivation of the coupled polar and axial perturbation equations. These equations were the starting point for investigations of relativistic rotational effects on stellar oscillations and associated instabilities. Most of the work considered the simpler task of solving the perturbation equations in the frequency domain and, as a result, calculations involved the determination of the eigenfrequencies rather than the solution of the time dependent equations. A somewhat different approach based on a Lagrangian description was used by Lockitch et al. (2001) with the focus on the computation of rotationally induced inertial modes. Following the more than 40 years old tradition, it was quite common to work in the Regge–Wheeler gauge, although some groups were using different gauges or the gauge invariant formulation of Moncrief (1974). In particular, a gauge introduced by Battiston, Cazzola & Lucaroni in 1971 in a series of papers to study the stability properties of non-radial oscillations in relativistic non-rotating stars, has not received proper attention (Battiston, Cazzola & Lucaroni 1971, Cazzola & Lucaroni 1972, 1974, 1978; Cazzola, Lucaroni & Semezato 1978). Since the perturbation equations of non-rotating stars are fairly simple, there is no real advantage of one gauge over the other. Note, however, that there was a long standing puzzle why in the Regge–Wheeler gauge the equations could be reduced to a fourth order system, whereas in the diagonal gauge used by Chandrasekhar & Ferrari (1991), which was previously used by Chandrasekhar (1983) in the context of black hole perturbations, only a fifth order system could be obtained. The latter system gave rise to an additional divergent solution, which had to be rejected on physical grounds. This discrepancy was finally solved by Price & Ipser (1991), who showed that the diagonal gauge was not completely fixed and still possessed one degree of gauge freedom, giving rise to the additional spurious solution. As computer power has been enormously increasing within the last decade, the problem of evolving the fully non-linear 3D Einstein equations in the time domain has finally become into reach of feasibility. Various groups around the world are now building robust codes that perform the time evolution of both single neutron stars or binaries in full 3D (see for instance Font et al. 2001, Shibata & Uryu 2001, and the review of Stergioulas 1998). Nevertheless, there is still considerable work in progress within perturbative approaches, which can provide us with a deeper understanding of the physics of relativistic star perturbations. Moreover, any trustable non-linear code must be able to reproduce the results from perturbation theory. Time evolutions of the perturbative equations have been carried out first for the axial equations (Andersson & Kokkotas 1996) and then for the polar equations using the Regge–Wheeler gauge by Allen et al. (1998) and Ruoff (2001a). Allen et al. (1998) managed to write down the evolution equations as two relatively simple wave equations for the metric perturbations and one wave equation for the fluid enthalpy perturbation inside the star. Ruoff (2001a) rederived these equations using the ADM formalism (Arnowitt et al. 1962). They were used to evolve and study initial data, representing the late stage of a binary neutron star head-on collision (Allen et al. 1999). Using the Arnowitt–Deser–Misner (ADM) formalism, the evolution equations for the axial perturbations of rotating stars have been brought into a suitable form for the numerical evolution by Ruoff & Kokkotas (2001a,b). Here, the resulting system of equations came out immediately as a first order system both in space and time, and it could be directly used for the numerical evolution without many further manipulations. In the non-rotating case, it is an easy task to transform the first order system into a single wave equation for just one metric variable. In the rotating case, however, this is not possible any more because of the rotational correction terms. When looking at the set of polar equations derived by Kojima (1992) it appears clearly that the presence of mixed spatial and time derivatives makes them not suitable for the numerical time integration. Nevertheless, using a number of successive manipulations and the introduction of many additional variables, we were able to cast the equations into a hyperbolic first order form. A more natural way to automatically obtain a set of equations, which is first order in time, is to use the ADM formalism. However, as we shall explain, even in that case the polar equations in the Regge–Wheeler gauge need to be further manipulated before they are suitable for a numerical integration. In general the ADM formalism yields a set of partial differential equations which are first order in time, but second order in space. For the numerical evolution, this is not ideal and one would rather prefer to have a pure first order system, or if possible a pure second order system, thus representing generalized wave equations. As we mentioned above, in the non-rotating case, it is easily possible to convert the perturbation equations into a set of wave equations. However, in the rotating case, this is not possible any more, even in the simple case when only axial perturbations are considered. To illustrate the problems associated with the Regge-Wheeler gauge, let us recall Einstein’s (unperturbed) evolution equations written in the ADM formalism: $$\begin{aligned} \label{eq_g} {\left(}\d_t - \cal L_\beta{\right)}\gamma_{ij} &=& - 2\alpha K_{ij}\;,\\ \label{eq_K} {\left(}\d_t - \cal L_\beta{\right)}K_{ij} &=& - \alpha_{;ij} + \alpha \left[R_{ij} + K^k_{\phantom{k}k}K_{ij} - 2K_{ik}K^k_{\phantom{i}j} - 4\pi {\left(}2T_{ij} - T^\nu_{\phantom{\nu}\nu}\gamma_{ij}{\right)}\right]\;,\end{aligned}$$ with $\alpha$ denoting the lapse function, $\beta^i$ the shift vector, $\cal L_\beta$ the Lie-derivative with respect to $\beta^i$, $\gamma_{ij}$ the metric of a space-like three dimensional hypersurface with Ricci tensor $R_{ij}$, and $K_{ij}$ its extrinsic curvature. It is obvious that the only second order spatial derivatives are $\d_i\d_j\alpha$ and $\d_i\d_j\gamma_{kl}$ with the latter coming from the Ricci tensor $R_{ij}$. This is still true for the linearized version of Eqs. (\[eq\_g\]) and (\[eq\_K\]). In the Regge–Wheeler gauge, we have a non-vanishing perturbation of the lapse $\alpha$ and of the diagonal components of the spatial perturbations $h_{ij}$. Using the notation of Ruoff (2001a), the polar perturbations can be written as $$\begin{aligned} \alpha &\sim& \sum_{l,m}{S_1^{lm}(t,r)Y_{lm}(\theta,\phi)}\;,\\ h_{ij} &\sim& \sum_{l,m}{{\left(}\begin{array}{ccc} S_3^{lm}(t,r) & 0 & 0\\ 0 & T_2^{lm}(t,r) & 0\\ 0 & 0 & \sin^2\theta\,T_2^{lm}(t,r)\\ \end{array}{\right)}Y_{lm}(\theta,\phi)}\;.\end{aligned}$$ The perturbation equations coming from Eq. (\[eq\_K\]) contain second order $r$-derivatives of $S_1$ and $T_2$. Note that they do not contain second order derivatives of $S_3$, because only the angular components of the metric get differentiated twice with respect to $r$. In the axial case there is only one perturbation function for the angular metric components, but it is set to zero in the Regge–Wheeler gauge. Therefore the ADM formalism immediately yields a first order system. The polar equations, in contrast, can be cast into a fully first order system only through the introduction of some auxiliary variables. In the non-rotating case, this is a fairly easy task, but for the rotating case, it turns out to become considerably more complicated. This is because the simple proportionality of $S_1$ and $S_3$, which holds in the non-rotating case and which makes is easy to replace $S_1$ by $S_3$, does not do so any more in the rotating case. Instead this relation involves rotational correction terms, and the replacement of $S_1$ by $S_3$ would lead to a considerable inflation of the equations. Consequently, instead of manipulating the perturbation equations in the Regge–Wheeler gauge, we look for a gauge in which the perturbation equations, by construction, do not show any second order spatial derivative. We have seen that the second derivatives come from the angular terms in the spatial metric and the lapse function. It seems therefore natural to set these components to zero. For the axial case this is already realized in the Regge–Wheeler gauge. It is only for the polar perturbations that we need a different gauge. From the seven polar components of the metric, Regge & Wheeler (1957) chose to set the components $V_1$, $V_3$ and $T_1$, which represent in the notation of Ruoff (2001a) the polar angular vector perturbations and one of the angular tensor perturbations, to zero. We now choose a different set, namely we set the angular terms in the spatial metric $T_1$, $T_2$ together with the lapse $S_1$ to zero and retain $V_1$ and $V_3$. With this choice we expect the ADM formalism to provide us with an evolution system without any second $r$-derivatives. We should mention again that this gauge has actually been introduced thirty years ago by Battiston, Cazzola & Lucaroni (1971) to derive the perturbation equations for non-radial oscillations of non-rotating neutron stars and to investigate in a subsequent series of papers their stability properties (Cazzola & Lucaroni 1972, 1974, 1978; Cazzola, Lucaroni & Semezato 1978a, 1978b). Particularly relevant is the first paper of the series, in which they proved the uniqueness of this gauge, hereafter referred to as the BCL gauge, and where they also showed the relation with the Regge–Wheeler gauge. In Section 2, we will use the ADM formalism to derive the time dependent perturbation equations for slowly rotating relativistic stars in the BCL gauge. Section 3 contains a brief discussion of the non-rotating limit and conclusions will be given in Section 4. In the Appendix, we present the perturbation equations as they follow directly from Einstein’s equations in a form similar to the equations in the Regge–Wheeler gauge given by Kojima (1992). Throughout the paper, we adopt the metric signature $(- + + +)$, and we use geometrical units with $c=G=1$. Derivatives with respect to the radial coordinate $r$ are denoted by a prime, while derivatives with respect to time $t$ are denoted by an over-dot. Greek indices run from 0 to 3, Latin indices from 1 to 3. The perturbation equations in the ADM formalism =============================================== The metric describing a slowly rotating neutron star reads in spherical coordinates ($t$, $r$, $\theta$, $\phi$) $$\begin{aligned} \label{metric} g_{\mu\nu} &=& {\left(}\begin{array}{cccc} -e^{2\nu} & 0 & 0 & -\omega r^2 \sin^2\theta\\ 0 & e^{2\lam} & 0 & 0\\ 0 & 0 & r^2 & 0\\ -\omega r^2 \sin^2\theta & 0 & 0 & r^2\sin^2\theta\\ \end{array}{\right)}\;,\end{aligned}$$ where $\nu$, $\lam$ and the “frame dragging” $\omega$ are functions of the radial coordinate $r$ only. With the neutron star matter described by a perfect fluid with pressure $p$, energy density $\eps$, and 4-velocity $$U^\mu = {\left(}e^{-\nu}, 0, 0, \Omega e^{-\nu}{\right)}\;,$$ the Einstein equations together with an equation of state $p=p(\eps)$ yield the well known TOV equations plus an extra equation for the frame dragging. To linear order, it is given by $$\label{drag} \varpi'' - {\left(}4\pi re^{2\lam}(p + \eps) - \frac{4}{r}{\right)}\varpi' - 16\pi e^{2\lam}{\left(}p + \eps{\right)}\varpi = 0\;,$$ where $$\varpi := \Omega - \omega$$ represents the angular velocity of the fluid relative to the local inertial frame. In the language of the ADM formalism, we have to express the above background metric (\[metric\]) in terms of lapse , (covariant) shift and the 3-metric, which we denote by $A$, $B_i$ and $\gamma_{ij}$, respectively. Explicitly, we have $$\begin{aligned} A &=& \sqrt{B^iB_i - g_{00}} \;=\; e^{\nu} + O(\omega^2)\;,\\ B_i &=& {\left(}0, 0, -\omega r^2 \sin^2\theta{\right)}\;,\\ \gamma_{ij} &=& {\left(}\begin{array}{ccc} e^{2\lam} & 0 & 0\\ 0 & r^2 & 0\\ 0 & 0 & r^2\sin^2\theta\\ \end{array}{\right)}\;.\end{aligned}$$ The extrinsic curvature of the space-like hypersurface described by $\gamma_{ij}$ can be computed by $$\begin{aligned} K_{ij} &=& \frac{1}{2A}{\left(}B^k\d_k\gamma_{ij} + \gamma_{ki}\d_jB^k + \gamma_{kj}\d_iB^k{\right)}\;,\end{aligned}$$ yielding the only non-vanishing components $$\begin{aligned} K_{13} &=& K_{31} \;=\; -\half\omega' e^{-\nu} r^2\sin^2\theta\;.\end{aligned}$$ The perturbations of the background lapse $A$, shift $B_i$, 3-metric $\gamma_{ij}$, extrinsic curvature $K_{ij}$, 4-velocity $U_i$, energy density $\eps$ and pressure $p$ will be denoted by $\alpha$, $\beta_i$, $h_{ij}$, $k_{ij}$, $u_i$, $\delta\eps$ and $\delta p$, respectively. The twelve evolution equations for $h_{ij}$ and $k_{ij}$ are obtained by linearizing the non-linear ADM equations (\[eq\_g\]) and (\[eq\_K\]). Working in the slow-rotation approximation, we keep only terms up to order $\Omega$ (or $\omega$). The background quantities $B^k$ and $K_{ij}$ are first order in $\Omega$, hence we can neglect any products thereof. Furthermore, it is $K = \gamma^{ij}K_{ij} = 0$. These circumstances lead to cancellations of various terms and the perturbations equations reduce to: $$\begin{aligned} \label{dtgij} \d_t h_{ij} &=& \d_i\beta_j + \d_j\beta_i - 2{\left(}A k_{ij} + K_{ij}\alpha + \Gamma^k_{\phantom{k}ij}\beta_k + B_k \delta\Gamma^k_{\phantom{k}ij}{\right)}\;,\\ \label{dtKij} \d_t k_{ij} &=& \alpha \left[R_{ij} + 4\pi(p - \eps)\gamma_{ij}\right] - \d_i\d_j\alpha + \Gamma^k_{\phantom{i}ij}\d _k\alpha + \delta \Gamma^k_{\phantom{i}ij}\d_kA\non\\ &&+ A \left[\delta R_{ij} + K_{ij}k - 2{\left(}K_i^{\phantom{i}k}k_{jk} + K_j^{\phantom{i}k}k_{ik}{\right)}+ 4\pi{\left(}(p - \eps)h_{ij} + \gamma_{ij}{\left(}\delta p - \delta\eps{\right)}- 2(p + \eps){\left(}u_i\delta u_j + u_j\delta u_i{\right)}{\right)}\right] \non\\ && + B^k\d_kk_{ij} + {\left(}\d_kK_{ij} - K_i^{\phantom{i}l}\d_j\gamma_{kl} - K_j^{\phantom{i}l}\d_i\gamma_{kl}{\right)}\beta^k + k_{ik}\d_jB^k + k_{jk}\d_iB^k + K_i^{\phantom{i}k}\d_j\beta_k + K_j^{\phantom{i}k}\d_i\beta_k\;,\end{aligned}$$ where $$\begin{aligned} k &:=& \gamma^{ij}k_{ij}\;,\\ \delta\Gamma^k_{\phantom{k}ij} &:=& \half\gamma^{km}{\left(}\d_ih_{mj} + \d_jh_{mi} - \d_mh_{ij} - 2\Gamma^{l}_{\phantom{i}ij} h_{lm}{\right)}\;,\\ \delta R_{ij} &:=& \d_k\delta\Gamma^k_{\phantom{i}ij} - \d_j\delta\Gamma^k_{\phantom{i}ik} + \Gamma^l_{\phantom{i}ij}\delta\Gamma^k_{\phantom{i}lk} + \Gamma^k_{\phantom{i}lk}\delta\Gamma^l_{\phantom{i}ij} - \Gamma^l_{\phantom{i}ik}\delta\Gamma^k_{\phantom{i}lj} - \Gamma^k_{\phantom{i}lj}\delta\Gamma^l_{\phantom{i}ik}\;.\end{aligned}$$ To obtain a closed set of evolution equations, we will also use the four evolution equations the fluid perturbations following from the linearized conservation law $\delta T^{\mu\nu}_{\phantom{\mu\nu};\mu} = 0$. Last but not least we need the four linearized constraint equations, which serve to construct physically valid initial data and to monitor the accuracy of the numerical evolution: $$\begin{aligned} \label{hhc} \gamma^{ij}\delta R_{ij} - h^{ij}R_{ij} - 2K^{ij}k_{ij} &=& 16\pi{\left(}\delta\eps + 2e^{-\nu}(p + \eps)(\Omega - \omega)\delta u_3{\right)}\;,\\ \label{mmc} -8\pi\left[(p + \eps)\delta u_i + u_i{\left(}\delta p + \delta\eps{\right)}\right]&=& \gamma^{jk}{\left(}\d _i k_{jk} - \d _j k_{ik} - \Gamma^l_{\phantom{i}ik}k_{jl} + \Gamma^l_{\phantom{i}jk} k_{il} - \delta\Gamma^l_{\phantom{i}ik}K_{jl} + \delta\Gamma^l_{\phantom{i}jk}K_{il}{\right)}\non\\ &&- h^{jk}{\left(}\d _i K_{jk} - \d _j K_{ik} - \Gamma^l_{\phantom{i}ik}K_{jl} + \Gamma^l_{\phantom{i}jk}K_{il}{\right)}\;.\end{aligned}$$ We assume the oscillations to be adiabatic, thus the relation between the Eulerian pressure perturbation $\delta p$ and density perturbation $\delta\eps$ is given by $$\label{adcond} \delta p = \frac{\Gamma_1p}{p + \eps}\delta\eps + p'\xi^r{\left(}\frac{\Gamma_1}{\Gamma} - 1{\right)}\;,$$ where $\Gamma_1$ represents the adiabatic index of the perturbed configuration, $\Gamma$ is the background adiabatic index $$\Gamma = \frac{p + \eps}{p}\frac{dp}{d\eps}\;,$$ and $\xi^r$ is the radial component of the fluid displacement vector $\xi^\mu$. The latter is related to the (covariant) 4-velocity perturbations $\delta u_\mu$ as follows $$\delta u_\mu = u^\nu h_{\mu\nu} + g_{\mu\nu}u^\lam\frac{\d\xi^\nu}{\d x^\lam} - \half u_\mu u^\kappa u^\lam h_{\kappa\lam}\;.$$ For the $r$ component, this gives us $$\label{xir} {\left(}\d_t + \Omega\d_\phi{\right)}\xi^r = e^{-2\lam}{\left(}e^\nu\delta u_r - \beta_r - \Omega h_{r\phi}{\right)}\;.$$ To proceed further, we expand the complete set of perturbations variables into spherical harmonics $Y_{lm} = Y_{lm}(\theta,\phi)$. This will enable us to eliminate the angular dependence and obtain a set of equations for the coefficients, which now only depend on $t$ and $r$. It is only then that we can finally choose our gauge. In principle, choosing the gauge amounts to providing prescriptions for lapse $\alpha$ and shift $\beta_i$. Those so-called slicing conditions determine how the space-like 3-metric foliates the 4-dimensional spacetime. In perturbation theory, the gauge can be used to set some of the ten metric perturbations to zero. We could, for instance, set $\alpha = \beta_i = 0$, and we would be left with only the six components $h_{ij}$. Note, that setting $\alpha$ to zero is possible, since this is only the perturbation of the background lapse $A$, and the latter does not vanish. However, our actual goal is to set some of the spatial perturbation components $h_{ij}$ to zero, in particular the angular components $h_{ab}$ with $a,b = \{\theta,\phi\}$. In principle we can prescribe the values of $h_{ij}$ only once for the initial data, and not throughout the evolution. The only possible way to keep $h_{ab}$ zero throughout the evolution is to choose our gauge such that the evolution equations for $h_{ab}$ become trivial, i.e. we have to have $$\begin{aligned} \d_t h_{ab} &=& 0\;,\quad a,b = \{\theta,\phi\}\;.\end{aligned}$$ We will see that this requirement leads to a unique algebraic condition for the shift vector $\beta_i$. With the definitions $$\begin{aligned} X_{lm} &:=& 2{\left(}\d_\theta - \cot\theta{\right)}\d_\phi Y_{lm}\;,\\ W_{lm} &:=& {\left(}\d^2_{\theta\theta} - \cot\theta\d_\theta - \frac{\d^2_{\phi\phi}}{\sin^2\theta}{\right)}Y_{lm} \;=\; {\left(}2\d^2_{\theta\theta} + l(l+1){\right)}Y_{lm}\;,\end{aligned}$$ we now expand the metric as follows. For the polar part we choose (symmetric components are denoted by an asterisk) $$\begin{aligned} \alpha &=& 0\;,\\ \beta_i^{polar} &=& \sum_{l,m}{{\left(}e^{2\lam}S_2^{lm},\, V_1^{lm}\d_\theta,\,V_1^{lm}\d_\phi{\right)}Y_{lm}}\;,\\ h_{ij}^{polar} &=& \sum_{l,m}{{\left(}\begin{array}{ccc} e^{2\lam}S_3^{lm} & V_3^{lm}\d_\theta & V_3^{lm}\d_\phi\\ \star & 0 & 0\\ \star & 0 & 0 \end{array}{\right)}Y_{lm}}\;,\label{p_metric}\end{aligned}$$ and the axial part is $$\begin{aligned} \beta_i^{axial} &=& \sum_{l,m}{{\left(}0, -V_2^{lm}\frac{\d_\phi}{\sin\theta},\, V_2^{lm}\sin\theta\d_\theta{\right)}Y_{lm}}\;,\\ h_{ij}^{axial} &=& \sum_{l,m}{{\left(}\begin{array}{ccc} 0 & \displaystyle{-V_4^{lm}\frac{\d_\phi}{\sin\theta}} & V_4^{lm} \sin\theta\d_\theta\\ \star & 0 & 0\\ \star & 0 & 0 \end{array}{\right)}Y_{lm}}\;.\end{aligned}$$ For the extrinsic curvature we have all six components $$\begin{aligned} k_{ij}^{polar} &=& \half e^{-\nu}\times\end{aligned}$$ $$\quad\sum_{l,m}{{\left(}\begin{array}{ccc} e^{2\lam}K_1^{lm}Y_{lm} & e^{2\lam}K_2^{lm}\d_\theta Y_{lm}& e^{2\lam}K_2^{lm}\d_\phi Y_{lm}\\ \star & {\left(}rK_4^{lm} - \Lambda K_5^{lm}{\right)}Y_{lm} + K_5^{lm} W_{lm} & K_5^{lm} X_{lm}\\ \star & K_5^{lm} X_{lm} & \sin^2\theta\left[{\left(}rK_4^{lm} - \Lambda K_5^{lm}{\right)}Y_{lm} - K_5^{lm} W_{lm}\right] \end{array}{\right)}}\;,$$ $$\begin{aligned} k_{ij}^{axial} &=& \half e^{-\nu}\sum_{l,m}{ {\left(}\begin{array}{ccc} 0 & \displaystyle{ -e^{2\lam}K_3^{lm}\frac{\d_\phi Y_{lm}}{\sin\theta}} & e^{2\lam}K_3^{lm}\sin\theta\d_\theta Y_{lm}\\ \star & \displaystyle{-K_6^{lm}\frac{X_{lm}}{\sin\theta}} & K_6^{lm}\sin\theta\,W_{lm}\\ \star & K_6^{lm}\sin\theta\,W_{lm} & K_6^{lm}\sin\theta\,X_{lm} \end{array}{\right)}}\;.\end{aligned}$$ Herein and throughout the whole paper, we use the shorthand notation $$\Lambda := l(l+1)\;.$$ We should note that the somewhat peculiar looking expansions for the coefficient $K_5^{lm}$ can actually be written as $$\begin{aligned} W_{lm} - \Lambda Y_{lm} &=& 2\d^2_{\theta\theta}Y_{lm}\;,\\ -\sin^2\theta{\left(}W_{lm} + \Lambda Y_{lm}{\right)}&=& 2{\left(}\cos\theta\sin\theta\d_\theta + \d^2_{\phi\phi}{\right)}Y_{lm}\;,\end{aligned}$$ which are essentially the diagonal terms of the Regge–Wheeler tensor harmonic $\Psi_{\alpha\beta}^{lm}$ (c.f. Eq. (20) of Ruoff 2001a). However, we prefer to write them in terms of $W_{lm}$ and $Y_{lm}$ because it is only for these quantities that simple orthogonality relations apply. Furthermore, we have to mention that in the definition of the polar components of the extrinsic curvature, we differ from the notation of Ruoff (2001a), where the meaning of $K_4$ and $K_5$ is reversed (c.f. Eq. (24)). Also, the expansion for the axial perturbations is not exactly the same as in Ruoff & Kokkotas (2001a,b). In their original paper, Battiston et al. (1971) did not use the ADM formalism to fix the gauge, instead they defined their gauge by directly setting $h_{tt}$, $h_{\theta\theta}$, $h_{\theta\phi}$ and $h_{\phi\phi}$ to zero. Since the relation between $h_{tt}$ and the lapse $\alpha$ is given by $$\label{ah_rel} h_{tt} = 2A\alpha + 2B^i\beta_i = 2e^{\nu}\alpha - 2\omega h_{t\phi}\;,$$ it follows that in the rotating case, $h_{tt} \ne 0$ although the lapse $\alpha$ vanishes. In the non-rotating case $\beta_i = 0$ and both $\alpha$ and $h_{tt}$ vanish. If we insisted on keeping a vanishing $h_{tt}$ also in the rotating case, we would obtain a non-vanishing lapse, giving us undesired second order spatial derivatives in the perturbation equations. Finally the fluid perturbations are decomposed as $$\begin{aligned} \delta u_i^{polar} &=& -e^\nu \sum_{l,m}{ {\left(}u_1^{lm}, u_2^{lm}\d_\theta, u_2^{lm}\d_\phi{\right)}Y_{lm}}\;,\\ \delta u_i^{axial} &=& -e^\nu \sum_{l,m}{ {\left(}0, -u_3^{lm}\frac{\d_\phi}{\sin\theta}, u_3^{lm}\sin\theta\d_\theta{\right)}Y_{lm}}\;,\\ \delta \eps &=& \sum_{l,m}{\rho^{lm} Y_{lm}}\;,\\ \delta p &=& {\left(}p + \eps{\right)}\sum_{l,m}{H^{lm} Y_{lm}}\;,\\ \xi^r &=& \bigg[\nu'{\left(}1 - \frac{\Gamma_1}{\Gamma}{\right)}\bigg]^{-1} \sum_{l,m}{\xi^{lm} Y_{lm}}\;.\end{aligned}$$ From Eq. (\[adcond\]), we have the relation $$\rho^{lm} = \frac{{\left(}p + \eps{\right)}^2}{\Gamma_1p}{\left(}H^{lm} - \xi^{lm}{\right)}\;.$$ For the sake of notational simplicity, we will from now on omit the indices $l$ and $m$ for the perturbation variables. With the above expansion, the evolution equations for $h_{ij}$ read: $$\begin{aligned} {\left(}\d_t + {\mbox{\rm i}}m\omega{\right)}S_3Y_{lm} &=& {\left(}2S_2' + 2\lam'S_2 - K_1{\right)}Y_{lm}\non\\ &&{} + 2\omega e^{-2\lam}{\left(}V_3' - \lam'V_3{\right)}\d_\phi Y_{lm} + 2\omega e^{-2\lam}{\left(}V_4' - \lam'V_4{\right)}\sin\theta\d_\theta Y_{lm}\;,\\ \label{v34a} \d_t{\left(}V_3\d_\theta - V_4\frac{\d_\phi}{\sin\theta}{\right)}Y_{lm} &=& {\left(}V_1' - \frac{2}{r}V_1 + e^{2\lam}{\left(}S_2 - K_2{\right)}{\right)}\d_\theta Y_{lm} - {\left(}V_2' - \frac{2}{r}V_2 - e^{2\lam}K_3{\right)}\frac{\d_\phi Y_{lm}}{\sin\theta}\non\\ &&{}- \omega \Lambda V_4\sin\theta Y_{lm}\;,\\ \label{v34b} \d_t{\left(}V_3 \d_\phi + V_4\sin\theta\d_\theta{\right)}Y_{lm} &=& {\left(}V_1' - \frac{2}{r}V_1 + e^{2\lam}{\left(}S_2 - K_2{\right)}{\right)}\d_\phi Y_{lm} + {\left(}V_2' - \frac{2}{r}V_2 - e^{2\lam}K_3{\right)}\sin\theta\d_\theta Y_{lm}\;,\\ \label{t123a} 0 &=& {\left(}2S_2 - \frac{\Lambda}{r}V_1 - K_4 + \frac{\Lambda}{r}K_5{\right)}Y_{lm} + 2\omega e^{-2\lam}{\left(}V_3\d_\phi Y_{lm} + V_4\sin\theta\d_\theta Y_{lm}{\right)}\;,\\ \label{t123b} 0 &=& {\left(}V_1 - K_5{\right)}W_{lm} + {\left(}V_2 - K_6{\right)}\frac{X_{lm}}{\sin\theta}\;,\\ \label{t123c} 0 &=& {\left(}V_1 - K_5{\right)}X_{lm} - {\left(}V_2 - K_6{\right)}\sin\theta W_{lm}\;.\end{aligned}$$ Still, in every equation a sum over all $l$ and $m$ is implied. From Eqs. (\[t123b\]) and (\[t123c\]) we immediately obtain our condition for the shift components $$\begin{aligned} V_1 &=& K_5\;,\\ V_2 &=& K_6\;,\end{aligned}$$ and from Eq. (\[t123a\]) it follows after multiplication with $Y^*_{lm}$ and integration over the 2-sphere that $$\begin{aligned} \label{s2} S_2 &=& \half K_4 - \omega e^{-2\lam}{\left(}{\mbox{\rm i}}mV_3 + \cLa V_4{\right)}\;,\end{aligned}$$ where we have defined the operator $\cLa$, which couples the equations of order $l$ to the equations of order $l+1$ and $l-1$, as $$\begin{aligned} \cLa A_{lm} &:=& \sum_{l'm'}A_{l'm'} \int_{S_2}{Y^*_{lm}\sin\theta\d_\theta Y_{l'm'}d\Omega} \;=\; (l-1)Q_{lm}A_{l-1m} - (l+2)Q_{l+1m}A_{l+1m}\;,\end{aligned}$$ with $$\begin{aligned} Q_{lm} &:=& \sqrt{\frac{(l-m)(l+m)}{(2l-1)(2l+1)}}\;.\end{aligned}$$ Later, we will also need $$\begin{aligned} \cLb A_{lm} &:=& \sum_{l'm'}A_{l'm'} \int_{S_2}{\d_\theta Y^*_{lm}\sin\theta Y_{l'm'}d\Omega} \;=\; -(l+1)Q_{lm}A_{l-1m} + lQ_{l+1m}A_{l+1m}\end{aligned}$$ and $$\begin{aligned} \cLc A_{lm} &:=& \sum_{l'm'}A_{l'm'} {\left(}l'(l'+1)\int_{S_2}{Y^*_{lm}\cos\theta Y_{l'm'}d\Omega} + \int_{S_2}{Y^*_{lm}\sin\theta{ \frac{\partial {}}{\partial {\theta}} }Y_{l'm'}d\Omega}{\right)}\non\\ &=& (l-1)(l+1)Q_{lm}A_{l-1m} + l(l+2)Q_{l+1m}A_{l+1m}\;.\end{aligned}$$ The operator $\cLc$ can actually be expressed in terms of $\cLa$ and $\cLb$: $$\begin{aligned} \cLc &=& -\half{\left(}\cLa(\Lambda-2) + \cLb\Lambda{\right)}\;.\end{aligned}$$ By making use of these relations we can eliminate the spherical harmonics and obtain the following simple set of evolution equations for the metric perturbations: $$\begin{aligned} \label{dts3} {\left(}\d_t + {\mbox{\rm i}}m\omega{\right)}S_3 &=& K_4' - K_1 + \lam'K_4 - 2\omega'e^{-2\lam}{\left(}{\mbox{\rm i}}mV_3 + \cLa V_4{\right)}\;,\\ \label{dtv3} {\left(}\d_t + {\mbox{\rm i}}m\omega{\right)}V_3 &=& K_5' - e^{2\lam}K_2 - \frac{2}{r}K_5 + \half e^{2\lam}K_4\;,\\ \label{dtv4} {\left(}\d_t + {\mbox{\rm i}}m\omega{\right)}V_4 &=& K_6' - e^{2\lam}K_3 - \frac{2}{r}K_6\;.\end{aligned}$$ In a similar way, we obtain the evolution equations for the six extrinsic curvature components, which are a little more lengthy: $$\begin{aligned} \label{dtk1} {\left(}\d_t + {\mbox{\rm i}}m\omega{\right)}K_1 &=& e^{2\nu-2\lam}\bigg[{\left(}\nu' + \frac{2}{r}{\right)}S_3' - 2\frac{\Lambda}{r^2}V_3' + 2\lam'\frac{\Lambda}{r^2}V_3 + 2{\left(}\frac{\nu'}{r} - \frac{\lam'}{r} - \frac{e^{2\lam} - 1}{r^2} + e^{2\lam}\frac{\Lambda}{2r^2}{\right)}S_3\bigg]\non\\ &&{}+ 8\pi e^{2\nu}{\left(}p + \eps{\right)}C_s^{-2}\left[{\left(}C_s^2 - 1{\right)}H + \xi\right] - 2e^{-2\lam}\omega'\bigg[{\mbox{\rm i}}m{\left(}K_5' - \frac{2}{r}K_5{\right)}+ \cLa{\left(}K_6' - \frac{2}{r}K_6{\right)}\bigg]\;,\\ \label{dtk2} {\left(}\d_t + {\mbox{\rm i}}m\omega{\right)}K_2 &=& e^{2\nu-2\lam}{\left(}{\left(}\nu' + \frac{1}{r}{\right)}S_3 - \frac{2}{r^2}V_3{\right)}\non\\ &&{} + \frac{{\mbox{\rm i}}mr^2}{2\Lambda}e^{-2\lam}\bigg[\omega' {\left(}K_4' - K_1 + \lam'K_4 - 4\frac{\Lambda - 1}{r^2}K_5{\right)}- 16\pi \varpi(p + \eps) {\left(}e^{2\lam}K_4 + 2e^{2\nu}u_1{\right)}\bigg]\non\\ &&{} - \frac{\omega'e^{-2\lam}}{\Lambda}\cLa{\left(}{\left(}\Lambda - 2{\right)}K_6{\right)}\;,\\ \label{dtk3} {\left(}\d_t + {\mbox{\rm i}}m\omega{\right)}K_3 &=& e^{2\nu-2\lam}\frac{\Lambda - 2}{r^2}V_4 + e^{-2\lam}\frac{\omega'}{\Lambda}{\left(}2{\mbox{\rm i}}mK_6 + {\left(}\Lambda - 2{\right)}\cLb K_5{\right)}\non\\ &&{}- \frac{r^2}{2\Lambda}e^{-2\lam}\cLb\bigg[ \omega'{\left(}K_4' - K_1 + \lam'K_4{\right)}- 16\pi \varpi(p + \eps){\left(}e^{2\lam}K_4 + 2e^{2\nu}u_1{\right)}\bigg]\;,\\ \label{dtk4} {\left(}\d_t + {\mbox{\rm i}}m\omega{\right)}K_4 &=& e^{2\nu-2\lam}\bigg[S_3' + 2{\left(}\nu' - \lam' + \frac{1}{r}{\right)}S_3 - \frac{2\Lambda}{r^2}V_3\bigg] + 8\pi r e^{2\nu}{\left(}p + \eps{\right)}C_s^{-2} \left[{\left(}C_s^2 - 1{\right)}H + \xi\right]\non\\ &&{}+ r{\left(}\cLa - \cLb{\right)}{\left(}\omega'K_3 + 16\pi e^{2\nu}\varpi(p + \eps)u_3{\right)}\;,\\ \label{dtk5} {\left(}\d_t + {\mbox{\rm i}}m\omega{\right)}K_5 &=& e^{2\nu-2\lam}{\left(}V_3' + {\left(}\nu' - \lam'{\right)}V_3 - \half e^{2\lam}S_3{\right)}\non\\ &&{} + \frac{r^2}{\Lambda}\left\{{\mbox{\rm i}}m \left[\omega'{\left(}\half K_4 - K_2{\right)}- 16\pi e^{2\nu}\varpi(p + \eps)u_2\right] - \cLb{\left(}\omega'K_3 + 16\pi e^{2\nu}\varpi(p + \eps)u_3{\right)}\right\}\;,\\ \label{dtk6} {\left(}\d_t + {\mbox{\rm i}}m\omega{\right)}K_6 &=& e^{2\nu-2\lam}{\left(}V_4' + {\left(}\nu' - \lam'{\right)}V_4{\right)}\non\\ &&{} - \frac{r^2}{\Lambda}\left\{{\mbox{\rm i}}m {\left(}\omega'K_3 + 16\pi e^{2\nu}\varpi(p + \eps)u_3{\right)}+ \cLb\left[\omega'{\left(}\half K_4 - K_2{\right)}- 16\pi e^{2\nu}\varpi(p + \eps)u_2\right]\right\}\;.\end{aligned}$$ It is worthwhile to point out the symmetry between the polar and axial equations. Each pair $V_3$ and $V_4$, $K_2$ and $K_3$, and $K_5$ and $K_6$ represent the polar and axial counterparts of a metric or extrinsic curvature perturbation. Thus, each associated pair of equations (\[dtv3\]) and (\[dtv4\]), (\[dtk2\]) and (\[dtk3\]), and (\[dtk5\]) and (\[dtk6\]) has basically the same structure, with only the polar equations containing additional terms as there are more polar variables than axial ones. The last missing set of evolution equations is the one for the fluid quantities, coming from $\delta{\left(}T^{\mu\nu}_{\phantom{\mu\nu};\mu}{\right)}= 0$ and from Eq. (\[xir\]): $$\begin{aligned} \label{dth} {\left(}\d_t + {\mbox{\rm i}}m\Omega{\right)}H &=& C_s^2\bigg\{e^{2\nu-2\lam}\bigg[u_1' + {\left(}2\nu' - \lam' + \frac{2}{r}{\right)}u_1 - e^{2\lam}\frac{\Lambda}{r^2}u_2\bigg] + \half K_1 - \frac{\Lambda}{r^2}K_5 + \frac{1}{r}K_4\non\\ &&{}\qquad+ \varpi e^{-2\lam} \bigg[{\mbox{\rm i}}m{\left(}V_3' + {\left(}\frac{2}{r} - \lam'{\right)}V_3 + e^{2\lam}{\left(}H - \half S_3{\right)}{\right)}+ \cLa{\left(}V_4' + {\left(}\frac{2}{r} - \lam'{\right)}V_4{\right)}\bigg]\bigg\}\non\\ &&{}- \nu'\bigg[e^{2\nu-2\lam}u_1 + \half K_4 + \varpi e^{-2\lam}{\left(}{\mbox{\rm i}}mV_3 + \cLa V_4{\right)}\bigg]\;,\\ \label{dtu1} {\left(}\d_t + {\mbox{\rm i}}m\Omega{\right)}u_1 &=& H' + \frac{p'}{\Gamma_1 p}\bigg[{\left(}\frac{\Gamma_1}{\Gamma} - 1{\right)}H + \xi\bigg] -{\mbox{\rm i}}m\bigg[e^{-2\nu}\varpi{\left(}K_5' - \frac{2}{r}K_5{\right)}+ {\left(}\omega' + 2\varpi{\left(}\nu' - \frac{1}{r}{\right)}{\right)}u_2\bigg]\non\\ &&{}-\cLa\bigg[e^{-2\nu}\varpi{\left(}K_6' - \frac{2}{r}K_6{\right)}+ {\left(}\omega' + 2\varpi{\left(}\nu' - \frac{1}{r}{\right)}{\right)}u_3\bigg]\;,\\ \label{dtu2} {\left(}\d_t + {\mbox{\rm i}}m\Omega{\right)}u_2 &=& H + \frac{\varpi}{\Lambda}{\left(}{\mbox{\rm i}}m{\left(}2u_2 - e^{-2\nu}{\left(}\Lambda - 2{\right)}K_5{\right)}+ 2\cLc u_3 - e^{-2\nu}\cLa{\left(}{\left(}\Lambda - 2{\right)}K_6{\right)}{\right)}- \frac{{\mbox{\rm i}}mr^2}{\Lambda}A\;,\\ \label{dtu3} {\left(}\d_t + {\mbox{\rm i}}m\Omega{\right)}u_3 &=& 2\frac{\varpi}{\Lambda}\bigg[{\mbox{\rm i}}m {\left(}u_3 + e^{-2\nu}K_6{\right)}- \cLc {\left(}u_2 + e^{-2\nu} K_5{\right)}\bigg] + \frac{r^2}{\Lambda}\cLb A\;,\\ \label{dtxi} {\left(}\d_t + {\mbox{\rm i}}m\Omega{\right)}\xi &=& \nu'{\left(}\frac{\Gamma_1}{\Gamma} - 1{\right)}\bigg[e^{2\nu-2\lam}u_1 + \half K_4 + \varpi e^{-2\lam}{\left(}{\mbox{\rm i}}mV_3 + \cLa V_4{\right)}\bigg]\;,\end{aligned}$$ where $$\begin{aligned} A &=& \varpi C_s^2\bigg\{e^{-2\lam}\left[u_1' + {\left(}2\nu' - \lam' + \frac{2}{r}{\right)}u_1 - e^{2\lam}\frac{\Lambda}{r^2}u_2\right] + e^{-2\nu}\left[\half K_1 - \frac{\Lambda}{r^2}K_5 + \frac{1}{r}K_4\right]\bigg\}\non\\ &&{}+ \left[\varpi{\left(}\nu' - \frac{2}{r}{\right)}+ \omega'\right] {\left(}e^{-2\lam}u_1 + \half e^{-2\nu}K_4{\right)}\;.\end{aligned}$$ In Eq. (\[dth\]), we have defined the sound speed $C_s$ as $$\begin{aligned} C^2_s &=& \frac{\Gamma_1}{\Gamma}\frac{dp}{d\eps}\;.\end{aligned}$$ The evolution equations comprise fourteen equations in total: four axial and ten polar ones. In the non-rotating case, they are equivalent to four wave equations, one for the axial and two for the polar metric perturbations plus one wave equation for the fluid variable $H$. The fluid equation for the axial velocity perturbation $u_3$ vanishes in the non-rotating case, whereas equation (\[dtxi\]) for the displacement variable $\xi$ does so in the barotropic case. Finally we have the four constraint equations. The Hamiltonian constraint reads $$\begin{aligned} \label{HC} 8\pi r^2e^{2\lam}\rho &=& rS_3' - \Lambda V_3' + {\left(}1 - 2r\lam' + \half e^{2\lam}\Lambda{\right)}S_3 + \Lambda{\left(}\lam' - \frac{1}{r}{\right)}V_3\non\\ &&{} + r^2e^{2\lam} \bigg[{\mbox{\rm i}}m{\left(}\half\omega'e^{-2\nu}K_2 + 16\pi\varpi{\left(}p + \eps{\right)}u_2{\right)}+ \cLa{\left(}\half\omega'e^{-2\nu}K_3 + 16\pi\varpi{\left(}p + \eps{\right)}u_3{\right)}\bigg]\;,\end{aligned}$$ and the three momentum constraints are $$\begin{aligned} \label{MC1} 8\pi r e^{2\nu}(p + \eps)u_1 &=& K_4' - \frac{\Lambda}{r}K_5' - K_1 + e^{2\lam}\frac{\Lambda}{2r}K_2 + \frac{\Lambda}{r^2}{\left(}1 + r\nu'{\right)}K_5 - \nu'K_4 + \frac{{\mbox{\rm i}}m}{4}r\omega'S_3\non\\ &&{}-{\left(}8\pi r{\left(}p + \eps{\right)}\varpi + 2e^{-2\lam}\omega'{\right)}{\left(}{\mbox{\rm i}}mV_3 + \cLa V_4{\right)}\;,\\ \label{MC2} 16\pi r e^{2\nu}(p + \eps)u_2 &=& -rK_2' + rK_1 + {\left(}r\nu' - r\lam' - 2{\right)}K_2 - \frac{2}{r}K_5 + K_4 + e^{-2\lam}\frac{r\omega'}{\Lambda} {\left(}2{\mbox{\rm i}}m V_3- {\left(}\Lambda - 2{\right)}\cLb V_4{\right)}\non\\ &&{} + \frac{{\mbox{\rm i}}mr^3}{\Lambda}\bigg[\half e^{-2\lam}\omega'S_3' - 16\pi\varpi(p + \eps){\left(}S_3 + C_s^{-2}{\left(}{\left(}C_s^2 + 1{\right)}H - \xi{\right)}{\right)}\bigg]\;,\\ \label{MC3} 16\pi r e^{2\nu}(p + \eps)u_3 &=& -rK_3' + {\left(}r\nu' - r\lam' - 2{\right)}K_3 + \frac{\Lambda - 2}{r}K_6 + e^{-2\lam}\frac{r\omega'}{\Lambda} {\left(}2{\mbox{\rm i}}m V_4 + {\left(}\Lambda - 2{\right)}\cLb V_3{\right)}\non\\ &&{} - \frac{r^3}{\Lambda}\cLb\bigg[\half e^{-2\lam}\omega'S_3' - 16\pi\varpi(p + \eps){\left(}S_3 + C_s^{-2}{\left(}{\left(}C_s^2 + 1{\right)}H - \xi{\right)}{\right)}\bigg]\;.\end{aligned}$$ The axial equations (\[dtv4\]), (\[dtk3\]), (\[dtk6\]), (\[dtu3\]) and (\[MC3\]) without the coupling terms to the polar perturbations are equivalent to Eqs. (7)–(10) and (12) of Ruoff & Kokkotas (2001b). Note, however, that therein a slightly different definition of the perturbation variables has been chosen. The non-rotating limit ====================== Although the non-rotating limit is well described by the wave equations given by Allen et al. (1998), it is instructive to consider it in the BCL gauge. This is obtained by setting $\Omega$ and $\omega$ to zero in all the evolution equations (\[dts3\])–(\[dtxi\]) and the constraints (\[HC\])–(\[MC3\]). As is well known, in this case the polar and axial parts of the equations completely decouple. For barotropic perturbations ($\Gamma_1 = \Gamma$), the polar evolution equations can then be easily transformed into three wave equations for the rescaled metric variables $S = e^{\nu-\lam} S_3$ and $V = e^{\nu-\lam}V_3/r$ and the rescaled fluid variable $\tilde H = e^{-\nu-\lam}H/r$: $$\begin{aligned} \label{waveS} { \frac{\partial^2 {S}}{\partial {t}^2} } &=& { \frac{\partial^2 {S}}{\partial {r_*}^2} }+ e^{2\nu-2\lam}\bigg[ {\left(}\nu'{\left(}\nu' - \lam'{\right)}+ 3\frac{\nu'}{r} + \frac{\lam'}{r} - \frac{3}{r^2} - e^{2\lam}\frac{\Lambda - 1}{r^2} - \lam''{\right)}S + \frac{4\Lambda}{r^2}{\left(}1 - r\nu'{\right)}V\bigg]\non\\ &&{} + 8\pi e^{2\nu - 2\lam}\bigg[{\left(}C_s^2 - 1{\right)}{\left(}\tilde \rho' + {\left(}\nu' - \frac{1}{r}{\right)}\tilde \rho{\right)}+ {\left(}C_s^2{\right)}'\tilde \rho\bigg]\;,\\ \label{waveV} { \frac{\partial^2 {V}}{\partial {t}^2} } &=& { \frac{\partial^2 {V}}{\partial {r_*}^2} } + e^{2\nu-2\lam}\bigg[{\left(}\frac{\nu'}{r} - \frac{\lam'}{r} + 2\frac{e^{2\lam} - 1}{r^2} - e^{2\lam}\frac{\Lambda}{r^2}{\right)}V - e^{2\lam}{\left(}\frac{\nu'}{r} + \frac{\lam'}{r} - \frac{1}{r^2}{\right)}S\bigg]\non\\ &&{} + 4\pi e^{2\nu}{\left(}C_s^2 - 1{\right)}\tilde \rho\;,\\ \label{waveH} { \frac{\partial^2 {\tilde H}}{\partial {t}^2} } &=& e^{2\nu-2\lam}\bigg\{C_s^2{ \frac{\partial^2 {\tilde H}}{\partial {r}^2} } - {\left(}C_s^2\lam' + \nu'{\right)}{ \frac{\partial {\tilde H}}{\partial {r}} }\non\\ &&{}\qquad\quad + \bigg[C_s^2{\left(}\lam'{\left(}\frac{3}{r} + \lam'{\right)}+ \frac{e^{2\lam} - 1}{r^2} - \lam'' - e^{2\lam}\frac{\Lambda}{r^2}{\right)}+ \frac{\lam'}{r} + 2\frac{\nu'}{r} + \frac{\nu'}{C_s^2}{\left(}\nu' + \lam'{\right)}\bigg]\tilde H\bigg\}\non\\ &&{} + e^{2\nu}\bigg\{\frac{r\nu'}{2}{\left(}C_s^2 - 1{\right)}{ \frac{\partial {S}}{\partial {r}} } + \bigg[C_s^2{\left(}\frac{\nu'}{2}{\left(}r\lam' - r\nu' + 6{\right)}+ \lam' - \frac{e^{2\lam} - 1}{r}{\right)}+ \frac{\nu'}{2}{\left(}r\lam' - r\nu' - 2{\right)}\bigg]S\non\\ &&{}\qquad\quad- \nu'\Lambda{\left(}C_s^2 - 1{\right)}V\bigg\}\;.\end{aligned}$$ In Eqs. (\[waveS\]) and (\[waveV\]), $r_*$ is the well-known tortoise coordinate, which is related to $r$ through $$\frac{d}{dr_*} = e^{\nu-\lam}\frac{d}{dr}\;.$$ Furthermore, one can express the energy density $\tilde\rho$ in terms of $\tilde H$, which in the barotropic case reduces to $$\tilde \rho = \frac{p + \eps}{C_s^2}\tilde H\,.$$ Although the equations in the first order form are quite simple, the above set of wave equations is more complicated than the equivalent set in the Regge–Wheeler gauge (Eqs. (14), (15) and (16) of Allen et al. 1998). This is particular so for the way in which the fluid variable $\tilde\rho$ (or equivalently $\tilde H$) couples to the metric variable $S$, where the derivative of $\tilde\rho$ enters. If the stellar model is based on a polytropic equation of state $p = \kappa \eps^\Gamma$, then the behaviour of $\tilde\rho$ at the stellar surface strongly depends on the polytropic index $\Gamma$. As discussed in Ruoff (2001a), $\tilde\rho$ actually diverges for $\Gamma > 2$. In this case the metric quantity $S$ would not even be ${\cal C}^0$, which can be troublesome for the numerical convergence. Although this could be a drawback of the BCL gauge, there is a clear advantage if one is interested in computing the gauge invariant Zerilli function $Z$ in the exterior. Following Moncrief (1974), the definition of the Zerilli function is $$\begin{aligned} Z &=& \frac{r^2{\left(}\Lambda k_1 + 4e^{-4\lam}k_2{\right)}}{r{\left(}\Lambda - 2{\right)}+ 2M}\end{aligned}$$ with $$\begin{aligned} k_1 = -2e^{-\lam-\nu}V\end{aligned}$$ and $$\begin{aligned} k_2 = \half e^{3\lam-\nu}S\;.\end{aligned}$$ In terms of $S$ and $V$ this gives us $$\begin{aligned} \label{Zer} Z &=& \frac{2r^2e^{-\lam-\nu}}{r{\left(}\Lambda - 2{\right)}+ 2M}{\left(}S - \Lambda V{\right)}\;,\end{aligned}$$ which is a simple algebraic relation in contrast to the relation in the Regge–Wheeler gauge, which includes a spatial derivative of one of the metric perturbations (see Eq. (20) of Allen et al. 1998, or Eq. (60) of Ruoff 2001a). In the Regge–Wheeler gauge, the two metric variables ($S$ and $F$ in the notation of Allen et al. 1998, and $S$ and $T$ in the notation of Ruoff 2001a) have different asymptotic behaviour at infinity, in particular one ($F$ or $T$) is linearly growing with $r$. It is only through the delicate cancellation of the growing terms that the Zerilli function remains finite at infinity. However, this cancellation can only happen if both metric variables exactly satisfy the Hamiltonian constraint. Any (numerical) violation leads to an incomplete cancellation, and the Zerilli function starts to grow at large radii. This makes it very difficult in the numerical time evolution to extract the correct amount of gravitational radiation emitted from the neutron star. With the above relation (\[Zer\]), we do not expect such difficulties to occur. Conclusions =========== We have presented the derivation of the perturbation equations for slowly rotating relativistic stars using the BCL gauge, which has been first used by Battiston, Cazzola & Lucaroni in 1971. This gauge is defined by setting $\alpha$, $h_{\theta\theta}$, $h_{\theta\phi}$ and $h_{\phi\phi}$ to zero. In the non-rotating case, the vanishing shift condition leads to a complete vanishing of $h_{tt}$, however, in the rotating case $h_{tt}$ becomes non-zero (see also Appendix). The advantage of the BCL gauge over the Regge–Wheeler gauge is that in the ADM formalism, the evolution equations a priori do not contain any second order spatial derivatives. Instead, one is immediately lead to a hyperbolic set of first order evolution equations, which can be directly used for the numerical time evolution without many further manipulations. Although it is in principle also possible to derive a hyperbolic set in the Regge–Wheeler gauge, the procedure is rather tedious and requires the introduction of carefully chosen new variables in order to replace the second order or mixed derivatives. The perturbation equations for slowly rotating relativistic stars form a set of fourteen evolution equations plus four constraints. In the non-rotating barotropic case, it is possible to cast the equations into a system of four wave equations, three for the polar and axial metric perturbations and one for the polar fluid perturbation, as it is the case in the Regge–Wheeler gauge. Although these wave equations are not simpler than the corresponding ones in the Regge–Wheeler gauge, the first order system actually is. Maybe the main advantage is the simple algebraic relation of the metric variables to the Zerilli function. It was demonstrated by Ruoff (2001a) that the accurate numerical evaluation of the Zerilli function in the Regge–Wheeler gauge is somewhat difficult and requires high resolution because a small numerical violation of the Hamiltonian constraint can lead to very large errors in the Zerilli function. This should not be the case in the BCL gauge as the relation (\[Zer\]) does not involve any derivatives. A further advantage of these evolution equations is that the inclusion of the source terms describing a particle orbiting the star can be very easily accomplished. This is not the case for the Regge–Wheeler gauge, as even for the non-rotating case, one is forced to include second order derivatives of the source terms (Ruoff 2001b). Since the source terms contain $\delta$-functions, one has to deal with second order derivatives thereof. In the axial case, no derivatives appear, and the perturbation equations with the source terms, which are given in Ruoff (2001b), are very simple. We expect the same to be the case for the polar equations in the BCL gauge. Here, it should be possible to plug the source terms into the equations for the extrinsic curvature without getting any derivatives. In subsequent papers, we will present results from the numerical evolution of the perturbation equations of slowly rotating relativistic stars in the BCL gauge with the particular focus on oscillations modes which are unstable with respect to gravitational radiation. As a step further, we will include the contribution of a test particle acting as a source of excitation for the stellar oscillations. acknowledgments {#acknowledgments .unnumbered} =============== We thank Nils Andersson, Luciano Rezzolla and Nikolaos Stergioulas for many helpful comments. J.R. is supported by the Marie Curie Fellowship No. HPMF-CT-1999-00364. A.S. is supported by the Greek National Scholarship foundation (I.K.Y.). This work has been supported by the EU Programme ’Improving the Human Research Potential and the Socio-Economic Knowledge Base’ (Research Training Network Contract HPRN-CT-2000-00137). [99]{} Allen G., Andersson N., Kokkotas K.D., Schutz B.F., 1998, Phys. Rev. D, 58, 124012 Allen G., Andersson N., Kokkotas K.D., Laguna P., Pullin J., Ruoff J., 1999, Phys. Rev. D, 60, 104021 Andersson N., Kokkotas K.D., 1996, Phys. Rev. Lett., 77, 4134 Arnowitt R., Deser S., Misner C.W., 1962, in Witten L., ed., Gravitation: An Introduction to Current Research. Wiley, New York, p.227 Battiston L., Cazzola P., Lucaroni L., 1971, Nuovo Cimento B, 3, 275 Cazzola P., Lucaroni L., 1972, Phys. Rev. D, 6, 950 Cazzola P., Lucaroni L., 1974, Phys. Rev. D, 10, 2038 Cazzola P., Lucaroni L., Semezato R., 1978, J. Math. Phys, 19, 237 Cazzola P., Lucaroni L., Semezato R., 1978, J. Math. Phys, 19, 901 Cazzola P., Lucaroni L., 1978, J. Math. Phys, 19, 2628 Chandrasekhar S., 1983, The Mathematical Theory of Black Holes, Oxford: Clarendon Press Chandrasekhar S., Ferrari V., 1991, Proc. R. Soc. Lond., A433, 423 Font J.A., Goodale T., Iyer S., Miller M., Rezzolla L., Seidel E., Stergioulas N., Tobias M., Suen W., 2001, in preparation Hartle J.B., 1967, ApJ, 150, 1005 Kojima Y., 1992, Phys. Rev. D, 46, 4289 Lockitch K.H., Andersson N., Friedman J.L., 2001, Phys. Rev. D, 63, 024019 Moncrief V., 1974, Ann. Phys., 88, 323 and 343 Price R.H., Ipser J.R., 1991, Phys. Rev. D, 44, 307 Regge T., Wheeler J.A., 1957, Phys. Rev., 108, 1063 Ruoff J., Kokkotas K.D., 2001a, MNRAS, gr-qc/0101105 Ruoff J., Kokkotas K.D., 2001b, MNRAS, gr-qc/0106073 Ruoff J., 2001a, Phys. Rev. D, 63, 064018 Ruoff J., Laguna P., Pullin J., 2001b, Phys. Rev. D, 63, 064019 Shibata M., Uryu K., 2001, to appear in H. Martel and J.C. Wheeler: Proceedings of the 20th Texas Symposium on Relativistic Astrophysics and Cosmology; astro-ph/0104409 Stergioulas N., 1998, Living Reviews in Relativity 1998-8, (http://www.livingreviews.org/Articles/Volume1/1998-8stergio/) Thorne K.S., Campolattaro A., 1967, ApJ, 149, 591; Erratum, 1968, ApJ, 152, 673 Appendix: The perturbation equations following from Einstein’s equations {#appendix-the-perturbation-equations-following-from-einsteins-equations .unnumbered} ======================================================================== Kojima (1992) derived the perturbation equations in the Regge–Wheeler gauge directly from the linearized Einstein equations without resorting to the ADM formalism. In this section we repeat this calculation using the BCL gauge. In order to facilitate the comparison with Kojima’s equations, who uses the more familiar notation of Regge & Wheeler (1957), we switch to a similar notation. In the Regge–Wheeler gauge, the quantities $h_0$ and $h_1$ denote the axial perturbations of $h_{t\{\phi,\theta\}}$ and $h_{r\{\phi,\theta\}}$, respectively, whereas the corresponding polar perturbations are set to zero. Since in the BCL gauge, the latter do not vanish, we denote them by $h_{0,p}$ and $h_{1,p}$, respectively, and to avoid confusion we denote the axial ones by $h_{0,a}$ and $h_{1,a}$. The remaining non-zero polar perturbations are then $H_1$ and $H_2$. Thus, the expansion of the metric in BCL gauge reads in this notation: $$h_{\mu\nu} = \sum_{lm} {\left(}\begin{array}{cccc} -2 \omega {\left(}h^{lm}_{0,p} \d_\phi + h^{lm}_{0,a} \sin\theta \d_\theta{\right)}& H^{lm}_1 & h^{lm}_{0,p} \d_\theta - h^{lm}_{0,a}/\sin\theta\d_\phi & h^{lm}_{0,p} \d_\phi + h^{lm}_{0,a} \sin\theta \d_\theta\\ \star & e^{2\lambda}H^{lm}_2 & h^{lm}_{1,p} \d_\theta - h^{lm}_{1,a}/\sin\theta\d_\phi & h^{lm}_{1,p}\d_\phi + h^{lm}_{1,a}\sin\theta \d_\theta\\ \star & \star & 0 & 0 \\ \star & \star & 0 & 0 \\ \end{array}{\right)}Y_{lm}\;. \label{hlm}$$ Note that here $h_{tt}$ is not zero, which is a consequence of the relation between the perturbation of the lapse $\alpha$ and $h_{tt}$ given by Eq. (\[ah\_rel\]). The relation between the above variables and the ones used in the previous sections is the following (we again omit the indices $l$ and $m$): $$\begin{aligned} H_1 &=& e^{2\lambda}K_4 - \omega {\left(}{\mbox{\rm i}}m V_3 + \cLa V_4 {\right)}\;,\\ H_2 &=& S_3\;,\\ h_{0,p} &=& K_5\;,\\ h_{0,a} &=& K_6\;,\\ h_{1,p} &=& V_3\;,\\ h_{1,a} &=& V_4\;,\\ R &=& -u_1\;,\\ V &=& -u_2\;,\\ U &=& -u_3\;.\end{aligned}$$ The extrinsic curvature components can be expressed as $$\begin{aligned} K_1 &=& 2 e^{-2\lambda} \left\{H_1' -\lambda'H_1 + \omega \left[ {\mbox{\rm i}}m{\left(}h_{1,p}' - \lambda'h_{1,p}{\right)}+ \cLa {\left(}h_{1,a}' - \lambda' h_{1,a}{\right)}\right]\right\} - \dot{H}_2 - {\mbox{\rm i}}m \omega H_2\;,\\ K_2 &=& e^{-2\lambda} \left[ h_{0,p}' - {2 \over r} h_{0,p} + H_1 - \dot{h}_{1,p} + \omega \cLa h_{1,a}\right]\;,\\ K_3 &=& e^{-2\lambda} \left[ h_{0,a}' - {2 \over r} h_{0,a} + \dot{h}_{1,a} - {\mbox{\rm i}}m \omega h_{1,a} \right]\;.\end{aligned}$$ A very often occurring combination of variables in the perturbation equations is $h_0' - \dot{h}_1$ for both the axial and polar cases, which we abbreviate with the following functions $$\begin{aligned} Z_a &=& h_{0,a}' - \dot{h}_{1,a}\;,\\ \label{Zeta} Z_p &=& h_{0,p}' - \dot{h}_{1,p}\;. \label{Zetae}\end{aligned}$$ The equations coming from the $(tt)$, $(tr)$, $(rr)$ and the addition of the $(\theta\theta)$ and $(\phi\phi)$ components can be written as $$A^{(I)}_{lm} + {\mbox{\rm i}}m C^{(I)}_{lm} + \cLb B^{(I)}_{lm} + \cLd \tilde{A}^{(I)}_{lm} = 0\;,\\ \label{tt}$$ with $$\begin{aligned} \cLd A_{lm} &:=& -\half{\left(}\cLa + \cLb{\right)}A_{lm} \;=\; Q_{lm}A_{l-1m} + Q_{l+1m}A_{l+1m}\end{aligned}$$ and $$\begin{aligned} A^{(tt)} &=& {2 e^{2\nu} \over r^2} \left[ r H_2' - \Lambda h_{1,p}' -16 \pi r^2 e^{2\lambda} C_s^{-2} {\left(}H - \xi {\right)}+ \Lambda {\left(}\lambda' - {1 \over r} {\right)}h_{1,p} + {\left(}1 - 2r \lambda' + { \Lambda e^{2\lambda} \over 2 } {\right)}H_2 \right]\;,\\ \label{A1tt} \tilde{A}^{(tt)} &=& 0\;, \\ \label{tildeA1} B^{(tt)} &=& 2 \omega Z_a' + {\left(}\omega' - 2\omega {\left(}\lambda' + \nu' - {2 \over r} {\right)}{\right)}Z_a - {4 \omega \over r} h_{0,a}' - 32 \pi \Omega {\left(}p+\eps{\right)}e^{2\nu+2\lambda} U\non \\ &&{}+ {2 \over r} \left[ -\omega' + \omega {\left(}2\nu' + 2\lambda' - { 2 \over r } - e^{2\lambda} {\Lambda - 2 \over r} {\right)}\right] h_{0,a}\;, \\ \label{B1tt} C^{(tt)} &=& 2 \omega {\left(}Z_p' - H_1' + e^{2\lambda} \dot{H}_2 {\right)}+ {\left(}\omega' - 2\omega {\left(}\lambda' + \nu'{\right)}{\right)}h_{0,p}' - {2\over r}\left[\omega'- 2\omega{\left(}\nu' + \lambda' - {e^{2\lambda}-1 \over r}{\right)}\right] h_{0,p} \non \\ &&{}- \left[\omega' - 2\omega{\left(}\nu' + \lambda' -{2 \over r}{\right)}\right] \dot{h}_{1,p} + {\left(}\omega' + 2\omega( \lambda' - \nu') {\right)}H_1 - 32\pi \Omega e^{2\nu+2\lambda} {\left(}p+\eps{\right)}V\;,\\ \label{C1tt} A^{(tr)} &=& {2 \over r} \dot{H}_2 + {\Lambda \over r^2} {\left(}Z_p + H_1 - 2 h_{0,p}' + 2\nu' h_{0,p} {\right)}+ 16 \pi{\left(}p + \eps{\right)}{\left(}e^{2\nu}R - H_1{\right)}\;,\\ \label{A2tr} \tilde{A}^{(tr)} &=& {2 \Lambda \omega \over r^2} h_{1,a} \;,\\ \label{tildeA2} B^{(tr)} &=& \left[ {\Lambda \omega \over r^2} - 16 \pi \Omega {\left(}p + \eps{\right)}\right] h_{1,a} \;,\\ \label{B2tr} C^{(tr)} &=& {\left(}{2 \omega \over r} + {\omega' \over 2} {\right)}H_2 - 16 \pi \Omega {\left(}p +\eps{\right)}h_{1,p} \;,\\ \label{C2tr} A^{(rr)} &=& \dot{H}_{1} + e^{2\nu}{\Lambda\over 2r}{\left(}\lambda' + {1 \over r} {\right)}h_{1,p} - 4 \pi r e^{2\nu+2\lambda} {\left(}p+\eps{\right)}H - {e^{2\nu} \over 2r} \left[ {\left(}2r \nu' + 1 {\right)}- {\Lambda \over 2} e^{2\lambda} \right] H_2 - e^{2\nu}{\Lambda \over 2r} h_{1,p}' \;,\\ \label{A3rr} \tilde{A}^{(rr)} &=& 0 \;,\\ \label{tildeArr} {B}^{(rr)} &=& \omega h_{0,a}' + {\omega' \over 2} h_{0,a} - {\left(}\omega + {r \omega' \over 4} {\right)}Z_a \;,\\ \label{B3rr} C^{(rr)} &=& \omega h_{0,p}' - {\left(}\omega + {r \omega' \over 4} {\right)}Z_p + {\left(}\frac{\omega'}{2} - e^{2\lambda}{\Lambda \omega \over r } {\right)}h_{0,p} + {\left(}\omega + {r \omega' \over 4} {\right)}H_1 \;,\\ \label{C3rr} A^{(\theta\theta+\phi\phi)} &=& -\ddot{H}_2 + 2 e^{-2\lambda} \left[\dot{H}_1' + {\left(}{1 \over r} - \lambda'{\right)}\dot{H}_1\right] - e^{2\nu-2\lambda} {\left(}\nu' + {1 \over r} {\right)}H_2' - {\Lambda \over r^2 }{\left(}\dot{h}_{0,p} - e^{2\nu-2\lambda} h_{1,p}'{\right)}\non\\ &&{}- 16 \pi e^{2\nu} (p+\eps) H - e^{2\nu} {\left(}{ \Lambda \over 2r^2 } + 16 \pi p {\right)}H_2 + { \Lambda \over r^2 } e^{2\nu-2\lambda} {\left(}\nu' -\lambda' {\right)}h_{1,p}\;, \\ \label{A4thph} \tilde{A}^{(\theta\theta+\phi\phi)} &=& 0 \;,\\ \label{tildeA4thph} {B}^{(\theta\theta+\phi\phi)} &=& 2 \omega e^{-2\lambda} {\left(}{h}_{0,a}'' - Z_a' + {\left(}{1\over r} - \lambda'{\right)}{\left(}h_{0,a} - Z_a{\right)}{\right)}+ 2\omega'e^{-2\lambda}{\left(}h_{0,a}' - {2\over r}h_{0,a}{\right)}- 16 \pi e^{2\nu} \varpi (p+\eps) U \;,\\ \label{B4thph} {C}^{(\theta\theta+\phi\phi)} &=& 2 \omega e^{-2\lambda}\left[ h_{0,p}'' - Z_p' + H_1' - e^{2\lambda}{\left(}\dot{H}_2 + {\Lambda \over r^2} h_{0,p}{\right)}+ {\left(}{1\over r} - \lambda'{\right)}{\left(}\dot{h}_{1,p} + H_1{\right)}\right]\non \\ &&{}+ e^{-2\lambda} \omega'{\left(}H_1 + 2h_{0,p}' - {4 \over r}h_{0,p}{\right)}- 16\pi\varpi e^{2\nu}{\left(}p + \eps{\right)}V\;. \label{C4thph}\end{aligned}$$ The $(t\theta)$ and $(r\theta)$ components are $$\Lambda a^{(I)}_{lm} + {\mbox{\rm i}}m d^{(I)}_{lm} + \cLc \tilde{a}^{(I)}_{lm} + \cLb \eta^{(I)}_{lm} = 0\;, \label{even}$$ with $$\begin{aligned} {a}^{(t\theta)} &=& - \dot{H}_2 + e^{-2\lambda} \left[H_1' - Z_p' + {2 \over r} h_{0,p}' + {\left(}\lambda' + \nu' - {2\over r} {\right)}Z_p + {\left(}\nu' - \lambda'{\right)}H_1 - { 2 \over r^2}{\left(}r\lambda' - r\nu' + e^{2\lambda} - 1{\right)}h_{0,p} \right]\non\\ &&{} + 16 \pi e^{2\nu}{\left(}p+\eps{\right)}V\;,\\ {d}^{(t\theta)} &=& e^{-2\lambda}\left[ 2\Lambda \omega {\left(}h_{1,p}' + {\left(}{1 \over r} - \nu' {\right)}h_{1,p}{\right)}+ \omega' {\left(}{r^2\over 2} H_2' + 2h_{1,p}{\right)}\right] - 16\pi r^2 \varpi {\left(}p+\eps{\right)}{\left(}H_2 + {\left(}1 + C_s^{-2}{\right)}H + C_s^{-2}\xi{\right)}\;,\\ \label{dttheta} \tilde{a}^{(t\theta)} &=& 2 \omega' e^{-2\lambda} h_{1,a} \;,\\ \label{tildea1ttheta} \eta^{(t\theta)} &=& -\Lambda \omega e^{-2\lambda} \left[ h_{1,a}' + {\left(}\nu' - \lambda'{\right)}h_{1,a} \right]\;, \\ \label{etattheta} a^{(r\theta)} &=& - \dot{Z}_p + {1 \over r} {\left(}2 e^{2\nu-2\lambda} - {\Lambda \over 2} {\right)}h_{1,p}' +\left[8\pi e^{2\nu}{\left(}p+\eps{\right)}+ {\Lambda \over 2r^2} {\left(}1 + r \lambda' {\right)}-{2\over r^2} e^{2\nu} \right] h_{1,p}\non \\ &&{}+ \left[ \nu' {\left(}e^{2\nu} - 1 {\right)}+ {1 \over 2r} {\left(}{\Lambda\over 2}e^{2\lambda} - 1 {\right)}\right] H_2 - 4 \pi r e^{2\lambda} {\left(}p+\eps{\right)}H\;, \\ \label{artheta} d^{(r\theta)} &=& 16 \pi r^2 \varpi{\left(}p+\eps{\right)}{\left(}H_1 + e^{2\nu} R{\right)}- \omega \Lambda{\left(}H_1 + Z_p - h_{0,p}{\right)}+ \omega'{\left(}{r^2\over 2}\dot{H}_2 - 2{\left(}\Lambda+2{\right)}h_{0,p}{\right)}\;,\\ \label{dtth} \tilde{a}^{(r\theta)} &=& 2 \omega' h_{0,a}\;,\\ \label{tildeartheta} \eta^{(r\theta)} &=& \Lambda \left[ \omega {\left(}h_{0,a}' - Z_a {\right)}+ \omega' h_{0,a} \right]\;. \label{etartheta}\end{aligned}$$ From the $(t\phi)$ and $(r\phi)$ components we get $$\Lambda b^{(I)}_{lm} + {\mbox{\rm i}}m c^{(I)}_{lm} + \cLc \tilde{b}^{(I)}_{lm} + \cLb \zeta^{(I)}_{lm} = 0\;, \label{odd}$$ with $$\begin{aligned} b^{(t\phi)} &=& - Z_a' + {\left(}\nu' + \lambda' - {2 \over r} {\right)}Z_a + {2 \over r} h_{0,a}' + 16 \pi e^{2\nu+2\lambda} {\left(}p+\eps{\right)}U - \left[ {2 \over r}{\left(}\nu' + \lambda' - {1 \over r} {\right)}- e^{2\lambda} {\Lambda - 2 \over r^2} \right] h_{0,a} \;,\\ \label{btphi} c^{(t\phi)} &=& -3 \Lambda \omega h_{1,a}' +\left[ \Lambda \omega {\left(}3 \lambda' - \nu' -{2 \over r} {\right)}- {\left(}\Lambda - 2{\right)}\omega' \right] h_{1,a} \;,\\ \label{ctphi} \tilde{b}^{(t\phi)} &=& - 2 e^{-2\lambda} \omega' h_{1,p} \;,\\ \label{btildetphi} \zeta^{(t\phi)} &=& 2\omega\Lambda{\left(}e^{2\lambda}H_2 + {\left(}\nu' -\lambda'{\right)}h_{1,p}{\right)}+ \omega'{\left(}\Lambda h_{1,p} - {r^2\over 2}H_2'{\right)}+ 16 \pi r^2 e^{2\lambda} \varpi {\left(}p+\eps{\right)}\left[ H_2 + {\left(}1 + C_s^{-2} {\right)}H + C_s^{-2} \xi \right]\;,\\ \label{dtphi} b^{(r\phi)} &=& \dot{Z}_a - {2 \over r} e^{2\nu-2\lambda}{h}_{1,a}' - e^{2\nu} \left[ { \Lambda - 2 \over r^2 } + {2 \over r} e^{-2\lambda} {\left(}\nu' - \lambda' {\right)}\right] h_{1,a} \;,\\ \label{brtheta} c^{(r\phi)} &=& \Lambda \omega h_{0,a}' + 2\left[{\left(}\Lambda + 1{\right)}\omega' - {\Lambda\omega\over r}\right] h_{0,a}\;,\\ \label{crph} \tilde{b}^{(r\phi)} &=& 2 \omega' h_{0,p}\;, \\ \label{tildebrphi} \zeta^{(r\phi)} &=& \omega'{\left(}\Lambda h_{0,p} - {r^2\over 2}\dot{H}_2{\right)}- 16\pi r^2\varpi {\left(}p+\eps{\right)}{\left(}e^{2\nu}R + H_1{\right)}\;. \label{drphi}\end{aligned}$$ From the $(\theta\phi)$ and the subtraction of $(\theta\theta)$ and $(\phi\phi)$ components we get $$\begin{aligned} \Lambda s_{lm} - {\mbox{\rm i}}m f_{lm} + \cLb g_{lm} &=& 0\;,\\ \label{thph} \Lambda t_{lm} + {\mbox{\rm i}}m g_{lm} + \cLb f_{lm} &=& 0\;, \label{thminusph}\end{aligned}$$ with $$\begin{aligned} f &=& \omega' r^2 e^{-2\lambda} {\left(}Z_p - {2 \over r} h_{0,p}{\right)}- 16 \pi r^2 \varpi e^{2\nu} {\left(}p+\eps{\right)}V \;,\\ \label{fthhp} g &=& -\omega' r^2 e^{-2\lambda} {\left(}Z_a - { 2 \over r} h_{0,a} {\right)}+ 16 \pi r^2 \varpi e^{2\nu} {\left(}p+\eps{\right)}U \;,\\ \label{gthph} s &=& - \dot{h}_{0,p} + e^{2\nu-2\lambda} {\left(}h_{1,p}' + {\left(}\nu' - \lambda'{\right)}h_{1,p}{\right)}- {e^{2\nu} \over 2}H_2 - {\mbox{\rm i}}m \omega h_{0,p} \;,\\ \label{sthph} t &=& -\dot{h}_{0,a} + e^{2\nu-2\lambda} {\left(}h_{1,a}' + {\left(}\nu' - \lambda'{\right)}h_{1,a}{\right)}- {\mbox{\rm i}}m \omega h_{0,a}\;. \label{tthhp}\end{aligned}$$ These equations are fully equivalent to the ones derived within the ADM formalism. Although they still contain some second order derivatives, they can be easily brought by introduction of a few auxiliary variables into a first order hyperbolic form or even into characteristic form, which is very useful for the numerical evolution. \[lastpage\]

ArXiv

--- abstract: 'We consider discrete-time Markov bridges, chains whose initial and final states coincide. We derive exact finite-time formulae for the joint probability distributions of additive functionals of trajectories. We apply our theory to time-integrated currents and frenesy of enzymatic reactions, which may include absolutely irreversible transitions. We discuss the information that frenesy carries about the currents and show that bridges may violate known uncertainty relations in certain cases. Numerical simulations are in perfect agreement with our theory.' author: - Édgar Roldán$^1$ and Pierpaolo Vivo$^2$ title: Exact Distributions of Currents and Frenesy for Markov Bridges --- introduction ============ Cutting-edge advances in physics and nanotechnology have paved the way in recent years to the quantitative analysis of fluctuation phenomena at mesoscopic scales. Important examples are the direct observation of molecular motors’ stepping [@svoboda1993direct], quantized charge transport in single-electron devices [@pekola2013single], rotation of synthetic mesoscopic machines [@kay2007synthetic], etc. Most non-equilibrium phenomena at small scales can be faithfully represented by Markov processes [@van1992stochastic]. In these models, the state of a system (e.g. active/inactive molecule) at a given time $t$ is described at a coarse-grained level by a discrete random variable (e.g. $X_t=\{1,2,...\}$) that changes its value at random times corresponding to transitions between different configurations (e.g. activation/deactivation of a molecule). Understanding fluctuations of thermodynamic quantities associated with a given [*trajectory*]{} (sequence of recordings) of a Markov process is one of the central goals of the emerging field of stochastic thermodynamics [@sekimoto2010stochastic; @jarzynski2011equalities; @seifert2012stochastic; @van2015ensemble; @martinez2017colloidal; @ciliberto2017experiments]. Recently, two different frameworks have established new universal thermodynamic properties of systems obeying Markovian dynamics, based on large deviations [@touchette2009large; @ellis2007entropy] and martingale theory [@doob1971martingale; @liptser2001statistics]. The large deviations approach has been instrumental in providing universal [*inequalities*]{} for the uncertainty of time-integrated currents in both discrete and continuous processes. The so called “thermodynamic uncertainty relations” [@barato2015thermodynamic; @gingrich2016dissipation; @pietzonka2016universal; @pietzonka2017finite; @proesmans2017discrete; @garrahan2017simple; @di2018kinetic; @maes2017frenetic; @ray2017dispersion; @dechant2018current; @shreshtha2019thermodynamic; @chiuchiu2018mapping; @shiraishi2018speed; @proesmans2018case; @koyuk2018generalization; @li2018quantifying; @carollo2018unravelling; @van2019uncertainty; @proesmans2019hysteretic], which have lately become an important line of research in its own right, provide only bounds for finite-time momenta. The martingale approach has instead led to universal [*equalities*]{}, which are however so far limited to a reduced set of thermodynamic currents [@neri2017statistics; @chetrite2011two; @ventejou2018progressive; @ge2018anomalous; @chetrite2019martingale; @manzano2019quantum]. The approach we put forward — inspired by the full counting statistics method from mesoscopic physics [@bagrets2003full; @bagrets2006full] — is very general, provides exact equalities for any finite time for a broad class of functionals, and is computationally inexpensive. In this article, we derive exact formulae – Eqs. (\[eq:key\],\[eq:key2\],\[eq:key3\]) – for the marginal and joint distributions of additive functionals of Markov chains obeying a specific constraint, namely Markov bridges. These are chains of finite length whose initial and final states coincide. For Markov bridges that may include absolutely irreversible transitions, our formula provides the exact marginal and joint distributions of a broad class of functionals, for instance:-1 - **Time-integrated currents**, which we simply call “currents”. These are functionals that change sign under time reversal of the trajectory. Examples: (i) the current between states $x$ and $x'$, given by the number of jumps $x\to x'$ minus the number of jumps $x' \to x$; (ii) any linear combination of currents between pairs of states in a network.-1 - **Frenetic quantities**, functionals that are invariant under time reversal of the trajectory. A paradigmatic example is given by the total number of jumps between any two different states in a network, which we simply call “frenesy" [@maes2018non; @baiesi2009nonequilibrium; @basu2015nonequilibrium; @basu2015statistical]. This quantity is often referred to as “traffic" or “activity". The study of bridges and constrained stochastic processes has a long history and many applications [@doob1957conditional; @fitzsimmons1993markovian; @horne2007analyzing; @majumdar2007brownian; @giardina2006direct; @morters2010brownian; @schehr2010area; @chetrite2015nonequilibrium; @majumdar2015effective; @szavits2015inequivalence; @benichou2016joint; @tizon2017structure; @mazzolo2017constrained; @adorisio2018exact; @perez2019sampling]. However, the discrete-time and discrete-space Markov setting is much less developed, especially in the context of stochastic thermodynamics. The analytical progress that we achieve in this work is especially welcome in the study of systems with absolutely irreversible transitions [@murashita2014nonequilibrium; @murashita2017gibbs; @monsel2018autonomous], which are notoriously harder to attack analytically. These kinds of systems are ubiquitous in e.g. biological processes, where biochemical and physiological reactions can often spontaneously occur only in one specific direction [@steinberg1986time; @chemla2008exact; @morin2015mechano; @battle2016broken; @dangkulwanich2013complete; @depken2013intermittent] and also in diffusion processes with stochastic resetting [@evans2011diffusion; @bhat2016stochastic; @roldan2016stochastic; @fuchs2016stochastic; @pal2017first; @roldan2017path; @montero2017continuous; @garcia2018path; @chechkin2018random; @masoliver2019telegraphic; @ahmad2019first; @pal2019local; @gupta2019stochastic]. In the following, we present the derivation of our main formulae, a swift review on celebrated uncertainty relations related to our work, and results for two examples: (i) a unicyclic enzymatic reaction; and (ii) a chemically-driven molecular motor with one absolutely irreversible transition. ![Illustration of two Markov bridges starting at time $T=0$ and ending at time $T=6$ in state $1$. For the red curve, we add (top black) the frenetic counter, which increases by one every time the system jumps between any two different states. For the same curve we add (bottom grey) the clockwise current counter, which increases (decreases) by one every time the system jumps from $x'$ to $x>x'$ ($x<x'$). -2 []{data-label="fig:1"}](figure_EPIC.pdf){width=".34\textwidth"} Theory ====== In this section, we develop our theory for the distribution of additive functionals for Markov bridges. We consider stationary discrete-time Markov chains of $T \geq 1$ steps defined over $K>1$ states. The probability to observe a trajectory $\omega_T \equiv ( X_0,X_1,\ldots,X_T)$, where each $X_t$ ($0\leq t\leq T$) belongs to the finite alphabet $\bm\xi=\{x_1,\ldots,x_K\}$, is given by $$P(\omega_T)=p(X_0)\pi(X_1|X_0)\cdots\pi(X_{T}| X_{T-1})\ .$$ Here, $p(X_0)$ is the probability of the initial state $X_0$, and $\pi(x|x')$ — arranged in the transition matrix $(\bm{\pi})_{x,x'}$ — denotes the conditional probability of jumping from state $x'$ to state $x$. A *Markov bridge* (MB) is a Markov chain constrained to terminate in the initial state (i.e. $X_0=X_T$). The probability to observe a trajectory of a Markov bridge reads $$P_{\rm MB}(\omega_T)=N_T^{-1} p(X_0)\pi(X_1|X_0)\cdots\pi(X_{T}| X_{T-1})\delta_{X_T,X_0}\ ,$$ where $\delta_{\alpha,\beta}$ is the Kronecker delta. The constant $N_T$ ensures that the probability is correctly normalized, $\sum_{\omega_T} P_{\rm MB}(\omega_T)=1$, where $\sum_{\omega_T} \equiv \sum_{X_0}\sum_{X_1}\cdots\sum_{X_T}$ with all sums running over the finite alphabet $X_i\in \bm\xi$. The inclusion of the bridge condition will provide a decisive technical advantage towards exact finite-time calculations, as we show below. Our interest is to provide exact finite-time statistics for additive Markov functionals of the form $$\bm{\Gamma} (\omega_T)=\sum_{t=0}^{T-1}\gamma(X_t,X_{t+1})\, ,$$ with the [*counter*]{} $\gamma(x',x)\in \mathbb{Z}$ for all $x,x'$. We note that $ \bm{\Gamma} (\omega_T)\in \mathbb{Z}$ is a random variable that depends on the full trajectory $\omega_T$. Denoting by $\Theta \omega_T$ the time-reversed trajectory — if allowed — key interesting physical examples of additive functionals are - Time-integrated currents, which obey $\gamma(x',x)=-\gamma(x,x')$, and are thus odd under time reversal $ \bm{\Gamma} (\omega_T)=-\bm{\Gamma} (\Theta\omega_T)$. - Frenetic quantities, obeying $\gamma(x',x)=\gamma(x,x')$ that are invariant under time reversal $ \bm{\Gamma} (\omega_T)=\bm{\Gamma} (\Theta\omega_T)$. For instance, in Fig. \[fig:1\] we show the frenetic counter $\gamma(x',x)=1-\delta_{x',x}$ and the clockwise current counter $\gamma(x',x)=\theta(x-x')-\theta(x'-x)$, with $\theta(y)=1$ if $y>0$ and $\theta(y)=0$ if $y\leq 0$. Exact finite-time distribution of a single functional ----------------------------------------------------- The probability distribution of any $\bm{\Gamma} (\omega_T)$ is $$P(\Gamma_T)\equiv \mathrm{Pr}\Big[\bm{\Gamma}(\omega_T)=\Gamma_T\Big]=\sum_{\omega_T}P_{\rm MB}(\omega_T)\delta_{\bm{\Gamma}(\omega_T),\Gamma_T}\ .$$ As noted above, we consider only functionals that can take integer values, i.e. $\Gamma_T \in \mathbb{Z}$. Introducing the well-known integral representation for the Kronecker delta, $\delta_{a,b}=(1/2\pi)\int_0^{2\pi} \text{d}k e^{\mathrm{i}k(a-b)}$, with $\mathrm{i}$ the imaginary unit and $a,b$ integers, we obtain $$\begin{aligned} \label{eq:5} P(\Gamma_T)&=&N_T^{-1}\int_0^{2\pi}\frac{\text{d}k}{2\pi}e^{\mathrm{i}k\Gamma_T}\sum_{X_T}\sum_{X_{T-1}}\pi_k(X_T|X_{T-1})\cdots\nonumber\\ &&\sum_{X_1}\pi_k(X_2|X_1)\sum_{X_0}\hat\pi_k(X_1|X_0)\delta_{X_0,X_T}\ ,\end{aligned}$$ where $\pi_k(x|x')\equiv \pi(x|x')e^{-\mathrm{i}k\gamma(x',x)}$ and $\hat\pi_k(x|x')\equiv p(x')\pi(x|x')e^{-\mathrm{i}k\gamma(x',x)}$. Using the Kronecker delta $\delta_{X_0,X_T}$ to kill the sum over $X_T$ we can rewrite  as $$\begin{aligned} \label{eq:6} P(\Gamma_T)&=&N_T^{-1}\int_0^{2\pi}\frac{\text{d}k}{2\pi}e^{\mathrm{i}k\Gamma_T}\sum_{X_0}\sum_{X_{T-1}}\cdots\sum_{X_1}\pi_k(X_0|X_{T-1})\nonumber\\ && \pi_k(X_{T-1}|X_{T-2})\cdots\pi_k(X_2|X_1)\hat\pi_k(X_1|X_0)\ ,\end{aligned}$$ where — in a way vaguely reminiscent of the *transfer matrix* approach to partition functions in disordered systems [@huang] — the multiple sum is now readily recognized as a trace of the product of $T$ matrices, using $\text{Tr} (A_1 \cdots A_m) = \sum_\ell \sum_{j_1}\cdots\sum_{j_{m-1}} (A_1)_{\ell,j_1} (A_2)_{j_1,j_2}\cdots (A_m)_{j_{m-1},\ell}$. Spelling this out in detail, we arrive at an analytical expression for the finite-time distribution of additive functionals $$\boxed{P(\Gamma_T)=\int_0^{2\pi}\frac{\text{d}k }{2\pi}e^{\mathrm{i}k \Gamma_T}\frac{\mathrm{Tr}\left[\bm{\pi}_k ^{T-1}\hat{\bm{\pi}}_k \right]}{\mathrm{Tr}\left[\bm{\pi}_0^{T-1}\hat{\bm{\pi}}_0\right]}}\ .\label{eq:key}$$ Here, $\mathrm{Tr}$ is the matrix trace, and the $K\times K$ matrices $\bm{\pi}$ and $\hat{\bm{\pi}}$ have elements $(\bm{\pi}_k )_{x,x'}=\pi_k (x|x')$ and $(\hat{\bm{\pi}}_k )_{x,x'}=\hat{\pi}_k (x|x')$, respectively. The *tilted matrix* $\bm{\pi}_k $ is a key object in the $T\to\infty$ theory of large deviations for Markov functionals [@touchette2009large], but here we make the most of it to extract [*finite time*]{} statistics. Further simplifications can be achieved if the initial probability $p(X_0)$ is uniform over all states, $p(X_0)=1/K$. In this case, the matrices $\bm{\pi}_k$ and $\hat{\bm{\pi}}_k$ are proportional to each other, and Eq.  simply reads $$P(\Gamma_T)= \int_0^{2\pi}\frac{\text{d}k }{2\pi}e^{\mathrm{i}k \Gamma_T}\frac{\mathrm{Tr}\left[\bm{\pi}_k ^{T}\right]}{\mathrm{Tr}\left[\bm{\pi}_0^{T}\right]}=\!\int_0^{2\pi}\frac{\text{d}k}{2\pi}e^{\mathrm{i}k\Gamma_T}\frac{\sum_{i}\lambda_i^T(k )}{\sum_{i}\lambda_i^T(0)}\ ,\label{formula2}$$ where $\lambda_i(k )$ with $i=1,\dots,K$ are the $K$ eigenvalues of the tilted matrix $\bm{\pi}_k $. This formula is particularly useful in the case where $K$ is not too large, or if the tilted matrix $\bm{\pi}_k $ has additional symmetries that facilitate an easier extraction of its eigenvalues, e.g. circulant matrices. Notably, from Eq.  and Eq. , the calculation of moments and cumulants for finite time is straightforward. As we will show below, Eq.  provides useful insights for simple examples of Markovian systems with homogeneous transition probabilities, see Sec. \[sec:iii\]. Furthermore, we will discuss how our main result  yields exact statistics for bridges in an example of a more complex Markovian process, see Sec. \[sec:iv\]. Joint distributions of finite-time functionals ---------------------------------------------- We can further exploit the multiplicative structure of the Fourier transform of the distribution to provide exact expressions for [*joint*]{} distributions of additive functionals for Markov bridges. For example, the joint probability distribution of any pair of functionals $(\bm{\Gamma}^{(1)} (\omega_T),\bm{\Gamma}^{(2)} (\omega_T))$ is $$\begin{aligned} \nonumber P(\Gamma^{(1)}_T,\Gamma^{(2)}_T) &\equiv \mathrm{Pr}\Big[\bm{\Gamma}^{(1)}(\omega_T) =\Gamma^{(1)}_T,\bm{\Gamma}^{(2)}(\omega_T) =\Gamma^{(2)}_T\Big]\\ &=\sum_{\omega_T}P_{\rm MB}(\omega_T)\delta_{\bm{\Gamma}^{(1)}(\omega_T),\Gamma^{(1)}_T}\delta_{\bm{\Gamma}^{(2)}(\omega_T),\Gamma^{(2)}_T}\ .\nonumber\end{aligned}$$ As noted above, we consider only functionals that can take integer values, i.e. $\Gamma^{(1)}_T,\Gamma^{(2)}_T \in \mathbb{Z}$. Introducing twice $\delta_{a,b}=(1/2\pi)\int_0^{2\pi} \text{d}k e^{\mathrm{i}k(a-b)}$, we obtain \[eq:5\] & P(\^[(1)]{}\_T,\^[(2)]{}\_T)=N\_T\^[-1]{}\_0\^[2]{}\_0\^[2]{}e\^[\[k\_1\^[(1)]{}\_T+k\_2 \^[(2)]{}\_T\]]{}\ &\_[X\_T]{}\_[X\_[T-1]{}]{}\_[k\_1,k\_2]{}(X\_T|X\_[T-1]{})\_[X\_1]{}\_[k\_1,k\_2]{}(X\_2|X\_1)\ &\_[X\_0]{}\_[k\_1,k\_2]{}(X\_1|X\_0)\_[X\_0,X\_T]{} , where $\pi_{k_1,k_2}(x|x')\equiv \pi(x|x')e^{-\mathrm{i}[k_1\gamma_1(x',x)+k_2 \gamma_2 (x',x)]}$ and $\hat\pi_{k_1,k_2}(x|x')\equiv p(x')\pi_{k_1,k_2}(x|x')$. Using the Kronecker delta $\delta_{X_0,X_T}$ to kill the sum over $X_T$ we obtain $$P(\Gamma_T^{(1)},\Gamma^{(2)}_T)=\!\iint_0^{2\pi}\frac{\text{d}k_1\text{d}k_2}{(2\pi)^2}e^{\mathrm{i}[k_1 \Gamma^{(1)}_T+k_2 \Gamma^{(2)}_T]}\frac{\mathrm{Tr}[\bm{\pi}_{k_1,k_2} ^{T-1}\hat{\bm{\pi}}_{k_1,k_2} ]}{\mathrm{Tr}[\bm{\pi}_{0,0}^{T-1}\hat{\bm{\pi}}_{0,0}]}\ .\label{eq:key2}$$ We can further generalize Eq.  to the case of $n> 2$ additive functionals. The joint finite-time distribution of a set of functionals $\Gamma_T^{(1)},\Gamma_T^{(2)},\dots,\Gamma_T^{(n)}$ is given by $$\boxed{\!P(\Gamma_T^{(1)},\dots,\Gamma^{(n)}_T)=\int_{\Omega_n} \frac{\text{d}^nk}{(2\pi)^n}\,e^{\mathrm{i}\sum_{\alpha=1}^n k_\alpha \Gamma^{(\alpha)}_T}\, \frac{\mathrm{Tr}[\bm{\pi}_{\bm{k} } ^{T-1}\hat{\bm{\pi}}_{\bm{k}} ]}{\mathrm{Tr}[\bm{\pi}_{\bm{0}}^{T-1}\hat{\bm{\pi}}_{\bm{0}}]}.\!}\label{eq:key3}$$ Here, $\text{d}^nk\equiv \text{d}k_1\cdots \text{d}k_n$, the tilted matrices are given by $\pi_{\bm{k}}(x|x')\equiv \pi(x|x')e^{-\mathrm{i}[k_1\gamma_1(x',x)+\cdots + k_n\gamma_n(x',x)]}$ and $\pi_{\bm{k}}(x|x')\equiv p(x')\pi_{\bm{k}}(x|x')$, respectively with $\gamma_i(x',x)$ the counter of the $i-$th functional. The integral $\Omega_n$ is done over $[0,2\pi]^n$ and $\mathbf{0} \equiv (0, \dots, 0)$. We remark that, in the long time limit, the rate functions for constrained trajectories (bridges) and for unconstrained trajectories are identical. For the case of a single additive functional, this result follows from Eq. : for large $T$, the dominant contribution to the integrand will come from $\lambda_{\mathrm{max}}(k)$, the largest eigenvalue of the tilted matrix. Setting $\Gamma_T = T\gamma_T$, with $\gamma_T$ an intensive parameter, we obtain $ P(\Gamma_T=T\gamma_T)\propto \int_0^{2\pi}\mathrm{d}k\ e^{T [\mathrm{i}k\gamma_T+\ln \lambda_{\mathrm{max}}(k) ]}$, which can be evaluated via a saddle-point approximation. Deforming the integration contour, this leads to $P(\Gamma_T)\sim e^{ T \psi(\Gamma_T/T)}$, where the rate function $\psi(x) = \max_k [k x - \ln\lambda_{\mathrm{max}}(k)]$ is the Legendre transform of the very same scaled cumulant generating function $\phi(k)=\ln\lambda_{\mathrm{max}}(k)$ that would be obtained for the unconstrained problem, see e.g. Ref. [@touchette2009large]. Appetizer over thermodynamic uncertainty relations ================================================== We now swiftly review recent universal bounds for the uncertainty of current-like and frenetic functionals of discrete Markovian stationary processes, so called [*thermodynamic uncertainty relations*]{} [@barato2015thermodynamic; @gingrich2016dissipation]. We stress that, at present, none of these results are available for Markov bridges. First, we discuss results for the uncertainty of time-integrated currents over a time window $T$, which we generically denote by $\mathbf{\Gamma}(\omega_T)=J_T$. For continuous-time stationary Markov processes, the relative uncertainty of any current obeys for all values of $T$ [@pietzonka2017finite; @horowitz2017proof] $$\frac{\text{var}(J_T)}{\langle J_T\rangle^2} \geq \frac{2k_{\rm B}}{\langle S_T\rangle}\, , \label{eq:tur1}$$ where $\langle J_T\rangle$ and $\text{var}(J_T)$ denote respectively the mean and variance of the current, and $\langle S_T\rangle$ is the average entropy produced up to time $T$. The latter is given by $\langle S_T\rangle =T(\Delta S /\Delta t)$, with $\Delta S=k_{\rm B} \sum_{x', x}p(x')\pi(x|x')\ln [\pi(x|x')p(x')/\pi(x'|x)p(x)]$ the entropy production per step and $k_{\rm B}$ the Boltzmann’s constant. A popular way to describe  is that one needs to dissipate a minimal amount of heat, given by $\mathsf{T}\langle S_T\rangle $ with $\mathsf{T}$ the temperature of the bath, in order to achieve the desired transport efficiency across the network [@ray2017dispersion; @dechant2018current]. For discrete-time Markov chains with time step $\Delta t$, Eq.  does not always hold, but should be instead replaced by a generalized uncertainty relation valid in the large $T$ limit, for both stationary and time-symmetric periodic driving [@proesmans2017discrete]-1 $$\frac{\text{var}(J_T)}{\langle J_T\rangle^2} \geq \frac{2(\Delta t/T)}{e^{\Delta S/k_{\rm B}}-1}\label{eq:tur2}\,.$$ Note that in the limit of small $\Delta t$, the right-hand side of Eq.  retrieves the right-hand side of Eq. , as it should. A comprehensive discussion on the discrete and continuous-time uncertainty relations for currents can be found in [@chiuchiu2018mapping], and a refreshing extension to non-Markovian “run-and-tumble" processes in [@shreshtha2019thermodynamic]. Among the class of time-integrated currents, a special place is reserved for the entropy production $S_T$ mentioned above, an additive functional defined by $\gamma(x',x) =k_{\rm B} \ln \left[p(x')\pi(x|x')/p(x)\pi(x'|x)\right]$ [@lebowitz1999gallavotti; @seifert2005entropy]. Specializing Eqs.  and  to $S_T$, one gets respectively $$\begin{aligned} \frac{\text{var}(S_T)}{\langle S_T\rangle} &\geq& 2k_{\rm B}\,, \label{eq:11} \\ \frac{\text{var}(S_T)}{\langle S_T\rangle} &\geq& \frac{2\Delta S/k_{\rm B}}{e^{\Delta S/k_{\rm B}}-1}\,,\label{eq:12}\end{aligned}$$ where  is valid at all times and  in the limit of large $T$.-1 So far, little is known about frenetic aspects of thermodynamic uncertainty relations [@garrahan2017simple; @maes2017frenetic; @di2018kinetic]. Important results include inequalities for the relative uncertainty of frenetic quantities in terms of the average frenesy [@di2018kinetic; @garrahan2017simple]. For continuous-time stationary Markov processes, the relative uncertainty of any additive functional $\Gamma_T$ obeys at all times $T$ the so-called “kinetic” uncertainty relation [@di2018kinetic] $$\frac{\text{var}(\Gamma_T)}{\langle \Gamma_T\rangle^2}\geq \frac{1}{\langle \Phi_T \rangle}\,,$$ where $\langle \Phi_T\rangle$ is the average frenesy. Specializing this relation to the frenesy itself, we get $$\frac{\text{var}(\Phi_T)}{\langle \Phi_T\rangle}\geq 1\, .\label{eq:garr}$$ To the best of our knowledge, analogous results for discrete-time processes are not currently available. We remark that all the uncertainty relations served in this appetizer involve ratios of cumulants for different kinds of functionals, which are easily accessible for finite time within our framework. In addition, none of these results concern cumulants evaluated along bridges, which are in general more “accurate" and produce more entropy than their unconstrained counterparts. -1 Currents and frenesy in unicyclic discrete-time enzymatic reactions {#sec:iii} =================================================================== Enzymatic reactions are often described using continuous-time Markov processes in which chemical reactions occur at random times [@van1992stochastic]. One simple yet paradigmatic example of a nonequilibrium cyclic enzymatic reaction is given by a single enzyme E that can convert substrate molecules S into product molecules P: $${\rm E}\leftrightarrows {\rm ES}\leftrightarrows {\rm EP}\leftrightarrows {\rm E} \label{eq:enz}\,.$$ Here we will assume that the enzyme is embedded in a thermal bath at temperature $\mathsf{T}$ that contains an excess stationary concentration of substrate S and product P molecules. In , E denotes the free enzyme, ES the enzyme bound to a substrate molecule and EP the enzyme bound to a product molecule. In continuous time, the probability for the enzyme to be in states E, ES and EP is described using master equations [@chemla2008exact; @ge2012stochastic]. ![Sketch of a three-state Markov chain describing a cyclic enzymatic reaction with homogeneous transition rates. In this model, the enzyme can be in three different states, E (free enzyme), ES (enzyme bounded to substrate) and EP (enzyme bounded to product). The transition probabilities between each of the states represented by symbols are shown with small letters. The red arrow illustrates the direction of the current that we are interested in. []{data-label="fig:2"}](model1.pdf){width=".25\textwidth"} {width="90.00000%"} In what follows, we consider a simplified discrete-time Markov model for the unicyclic enzymatic reaction given by Eq. , see Fig. \[fig:2\] for an illustration. In this model, the transition rates are considered to be homogeneous, yielding a biased motion in the clockwise direction ${\rm E}\to {\rm ES}\to {\rm EP}$. Mathematically, we describe the model as a three-state Markov chain with states ${\rm E}=1$, ${\rm ES}=2$, and ${\rm EP}=3$ with probabilities $p$ and $q$ to jump clockwise and counterclockwise, respectively. Thus, the transition matrix is $$\bm\pi=\begin{pmatrix} r & q & p\\ p & r & q\\ q & p & r \end{pmatrix}\ ,$$ with $r=1-(p+q)$ the probability to remain in a given state. We assume that the transition probabilities obey local detailed balance $p/q=e^{-Q/k_{\rm B}\mathsf{T}}$ with $Q$ ($-Q$) the heat absorbed (dissipated) by the enzyme into the bath in one clockwise (counterclockwise) transition. The cycle affinity of this enzyme is $\mathcal{A}=k_{\rm B}\mathsf{T}\ln [(p/q)^3] = 3k_{\rm B}\mathsf{T} \ln(p/q)$. ![Statistics of frenesy $\Phi_T$ (total number of jumps between different states) in Markov bridges for the three-state model in Fig. \[fig:2\] obtained from Monte Carlo simulations (symbols) and using Eq.  (solid lines). (a) Distribution of frenesy for parameters $p=0.47$ and $q=0.02$. (b) Fano factor of the frenesy $\Phi_T$ as a function of time $T$ for parameter $p=0.47$. The horizontal dashed line in (b) is set to one, following the uncertainty relation . In (a) and (b), the number of simulated bridges is $10^7$.[]{data-label="fig:4"}](Fig3.pdf){width=".37\textwidth"} We investigate bridges where the enzyme’s state is E at both time $0$ and time $T$. Two natural observables to measure in such process are the net current in the clockwise direction and the total number of jumps. These are simply related to entropy production and frenesy along cycles. In fact, the entropy produced along a bridge equals $S_T = (\mathcal{A}/3\mathsf{T}) J_T$, with $J_T$ the net number of jumps in the clockwise direction. We will provide exact formulae for the statistics of the clockwise current $J_T$, the frenesy $\Phi_T$ defined as the total number of jumps between different states, and for the joint statistics of $J_T$ and $\Phi_T$. Exact current statistics {#sec:iiia} ------------------------ The tilted matrix associated with the clockwise current $J_T\in \{-T,\dots,T\}$ is $$\bm\pi_k=\begin{pmatrix} r & q e^{\mathrm{i}k} & pe^{-\mathrm{i}k}\\ pe^{-\mathrm{i}k} & r & qe^{\mathrm{i}k}\\ qe^{\mathrm{i}k} & pe^{-\mathrm{i}k} & r \end{pmatrix}\ .$$ Because of the homogeneous nature of the transition matrix, we can apply directly formula  to calculate the distribution of any additive functional. In particular for $J_T$ we obtain, using the residue theorem and Faà di Bruno formula (see Appendix \[app:1\]): $$P(J_T) = N_T^{-1}p^T\sum_{j=0}^2 (\omega_j)^{2T}\,,\label{eq:FDB1}$$ for $J_T=T$, and $$\begin{aligned} P(J_T)&= N_T^{-1}\sum_{j=0}^2\frac{1}{(T-J_T)!}\sum_{k=1}^{T-J_T}[T]_k (p\omega_j^2)^{T-k}\nonumber\label{eq:FDB2} \\ &\times B_{T-J_T,k}\left(r,2q\omega_j,0,\ldots,0\right)\,,\end{aligned}$$ for $-T\leq J_T <T$. Here, $$N_T=\sum_{j=0}^2 (r+q\omega_j+p\omega_j^2)^T\,$$ is a normalization factor and $$\omega_j = e^{2\pi \mathrm{i} j/3}\,$$ is the $j-$th root ($j=0,1,2$) of the equation $z^3=1$. We have also introduced the falling factorial $[x]_n=\prod_{k=0}^{n-1} (x-k)$ and the incomplete Bell polynomials $B_{m,n}(x_1,x_2,...,x_{m-n+1}) $. With some algebra, it is possible to show that the distribution given by Eqs.  and  is real-valued and normalized. We perform numerical simulations of $10^7$ three-state Markov bridges, obtained by discarding from a large number of simulations of unconstrained chains those that did not meet the bridge constraint. The results shown in Fig. \[fig:3\]a are described very accurately by the formulae  and , which predict that the clockwise current is quantized in multiples of three, as required by the bridge condition. Furthermore, the Fano factor $\text{var}(S_T)/\langle S_T\rangle$ along bridges can be also computed exactly as shown in Figs. \[fig:3\]b-c. Not surprisingly, the bound  — valid for continuous-time processes — is violated at both small and large times and for a broad range of parameters. Interestingly, also the discrete-time uncertainty relation  is occasionally violated for finite time and especially in the limit of $\mathcal{A}$ large (Fig. \[fig:3\]c). The Fano factor often displays strongly non-monotonic behaviour as a function of $T$ (Fig. \[fig:3\]b) or of the bias (Fig. \[fig:3\]c). It will be very interesting in the future to tackle the question of what is the optimal time window that achieves the maximum accuracy in measurements of currents. Exact distribution of the frenesy {#sec:iiib} --------------------------------- The tilted matrix associated with the frenesy $\Phi_T\in \{0,\dots,T\}$ — the total number of jumps between different states — is $$\bm\pi_k=\begin{pmatrix} r & q e^{-\mathrm{i}k} & pe^{-\mathrm{i}k}\\ pe^{-\mathrm{i}k} & r & qe^{-\mathrm{i}k}\\ qe^{-\mathrm{i}k} & pe^{-\mathrm{i}k} & r \end{pmatrix}\ .$$ Following the same procedure as before in Sec. \[sec:iiia\], we find the remarkable closed expression for the distribution of the frenesy $$P(\Phi_T) = \frac{{T\choose T-\Phi_T} r^{T-\Phi_T} \sum_{j=0}^2 \label{eq:dfren} \left( q \omega_j+ p \omega_j^2 \right)^{\Phi_T}}{\sum_{j=0}^2 (r+ q \omega_j + p\omega_j^2 )^{T}}\,,$$ which again can be shown to be real-valued and normalized. We show in Fig. \[fig:4\]a that Eq.  reproduces, with excellent agreement, the frenetic distribution at any time $T$ obtained from numerical simulations of $10^7$ bridges. Interestingly, our formula reveals the complicated structure of the distribution, with many maxima, minima, and forbidden values due to the bridge constraint, which become less and less important as the time horizon grows. The Fano factor of the frenesy displays a non-trivial dependence on time and generically violates the uncertainty relation for the frenesy , which is valid for continuous-time processes and in the large $T$ limit. Notably, as revealed by Fig. \[fig:4\]b, the quest for improved bounds valid in topologically constrained processes is still open. Frenetic information about the current {#sec:ivc} -------------------------------------- As interesting corollaries of our theory, we use Eq.  to derive an elegant expression for the joint distribution of the finite-time clockwise current and the frenesy (see Appendix \[app:2\]): $$P(J_T,\Phi_T)=\left(3\frac{T! }{P_T! N^-_T! N^+_T! }\right) \frac{r^{P_T} q^{N_T^-} p^{N_T^+} }{\sum_{j=0}^2 (r+q\omega_j+p\omega_j^2)^T}\ . \label{eq:PJF}$$ Here the two integers $N^+_T =(\Phi_T+J_T)/2$ ($N^-_T =(\Phi_T-J_T)/2$) count the total number of jumps in the clockwise (counterclockwise) direction, and $P_T = T-\Phi_T$ is the persistence time, i.e. the amount of time spent without making a jump. This formula has support for $\Phi_T \in [0,T] $, $|J_T|\leq \Phi_T $ which we denote as [*frenetic cone*]{}, and $J_T$ an integer multiple of three. ![Joint distribution of current and frenesy for the three-state model in Fig. \[fig:2\]: theory (a,c) and simulations (b,d). The parameter values are $p=0.4, q=0.295$ (a,b) and $p=0.4, q=0.02$ (c,d). In all cases we used $T=11$ and in (b,d) we simulated $10^7$ bridges. The affinity of the different conditions are given by $\mathcal{A}=3k_{\rm B}\mathsf{T}\ln (p/q)$. The white lines illustrate the “frenetic cone" that limits the current values $|J_T|\leq \Phi_T$.[]{data-label="fig2d"}](spaceinvaders.pdf){width=".49\textwidth"} Using numerical simulations, we test Eq.  in close to equilibrium (Fig. \[fig2d\]a-b) and far from equilibrium (Fig. \[fig2d\]c-d) conditions. Notably, our formula reproduces with high accuracy the empirical distributions obtained from simulations with an accuracy of the order of $10^{-8}$. We can explore two interesting consequences of formula : the quantification of correlation between frenesy and current in terms of mutual information, and fluctuation theorems for joint distributions. Little is known about how much information frenetic quantities carry about currents both close and far from thermal equilibrium. To bridge this gap, we evaluate the mutual information (in bits) between the finite-time clockwise current and the frenesy, which is defined as $$I_{p,q}(J_T;\Phi_T) =\!\! \sum_{J_T,\Phi_T} P(J_T,\Phi_T) \log_2\!\left( \frac{P(J_T,\Phi_T)}{P(J_T)P(\Phi_T)}\right), \label{eq:info}$$ where $P(J_T)$, $P(\Phi_T)$ are the marginals computed in Sec. \[sec:iiia\] and Sec. \[sec:iiib\], respectively, and the sum runs over the support of the joint distribution $P(J_T, \Phi_T)$ given by Eq. .-1 We now analyze the behaviour of the mutual information  in the three-state model of the enzyme. In Fig. \[fig:info\]a, we plot the information $I_{p,q}(J_T;\Phi_T)$ as a function of the affinity $\mathcal{A}=3k_{\rm B}\mathsf{T}\ln(p/q)$ for a fixed value of $p$, and different values of the length $T$. The frenesy contains no information about the current for all values of $\mathcal{A}$ for the case of the shortest possible bridges. All these bridges return to the origin after two steps $T=2$ and thus $J_T=0$, $\Phi_T=2$. For longer bridges $T\geq 3$, we observe a very rich phenomenology. Quite surprisingly, for bridges of size $T=3$ the information content in the frenesy about the current decreases transiently with the bias strength in close-to-equilibrium conditions ($\mathcal{A}\lesssim 2k_{\rm B}\mathsf{T}$), and a similar phenomenon occurs for longer bridges ($T\geq 4$) at larger values of $\mathcal{A}$. In the latter case, the maximum value of the information for fixed $p$ and $T$ is achieved at intermediate values of $q$. In Fig. \[fig:info\]b, we show that the maximum value of the information $I_{\text{max},p} \equiv \text{max}_q I_{p,q}(J_T;\Phi_T)$ is an increasing function of $p$ and $T$. For small values of $p$, $I_{\rm max}$ increases smoothly with $T$, whereas when $p$ is large the maximum information develops a steep jump at $T=3$ from $0$ to $1$ bit. This result implies that, for $T$ small and $p+q$ large, one can obtain a maximum information of binary type – e.g. whether the current is positive or negative – by simply counting the number of jumps that occur in any direction. We also find that $I_{\text{max},p} \leq 2\,\text{bits}$ for all tested parameter values. ![Mutual information between the current and the frenesy in the three-states bridges for the enzymatic reaction sketched in Fig. \[fig:2\]. (a) Theoretical value of the mutual information $I_{p,q}(J_T;\Phi_T)$ given by Eq. , for fixed $p=0.4$ as a function of the cycle affinity (in units of $k_{\rm B}\mathsf{T}$), for different values of the bridge duration $T$. The circles indicate the maximum value of the mutual information $I_{\text{max},p} = \text{max}_q I_{p,q}(J_T;\Phi_T)$ for each value of $p$. (b) Theoretical value of the maximum mutual information $I_{\text{max},p}$ as a function of the bridge length $T$ for different values of the probability $p$ to jump clockwise (see legend). The lines are a guide to the eye. []{data-label="fig:info"}](frenetic_information.pdf){width=".37\textwidth"} Additional insights that can be gained from Eq.  are fluctuation theorems for joint distributions [@garcia2010unifying]. From the symmetry properties of , a joint fluctuation relation for iso-frenetic bridges follows: $$\frac{P(J_T,\Phi_T )}{P(-J_T,\Phi_T)} = e^{(\Delta S/k_{\rm B})J_T}\,.$$ Thus, the distribution of the current along bridges with fixed frenesy $\Phi_T$ (iso-frenetic bridges) is asymmetric with respect to change of sign of the current and its asymmetry is independent on the level of frenesy. which implies the cycle fluctuation relation for currents $P(J_T)/P(-J_T)=e^{(\Delta S/k_{\rm B})J_T}$ [@jia2016cycle; @polettini2018effective]. Frenetic fluctuations of molecular motors with broken detailed balance {#sec:iv} ====================================================================== In this section, we put our approach to the test by considering bridges in a four-state Markov chain with absolutely irreversible jumps. We consider a minimal four-state Markov model describing the motion of a molecular motor in a periodic track, see Fig. \[fig:5\] for an illustration. It is given by a simplification of existing models of active polymerization of e.g. DNA/RNA by molecular motors (polymerases) [@dangkulwanich2013complete; @depken2013intermittent]. In our minimal model, the free motor M (state $1$) can bind to a nucleoside triphosphate MT (state $2$) which serves as fuel with a probability of attachment $a$, and can enter a passive “backtracking” state B (state $4$) with a backtracking probability $b$. From the bound state MT, the motor can hydrolyze the fuel into nucleoside diphosphate changing its conformation to state MD (state $3$), and the fuel can be detached from the motor with a detachment probability $d$. Next, the motor in state MD can synthesize fuel with a small synthesis probability $s$ and can translocate the polymer irreversibly with probability $t$. Finally, the motor can recover from the backtracking state B with probability $r$. The existence of the absolutely irreversible step (red arrow in Fig. \[fig:5\]) may originate because of chemical and structural constraints in the polymerization process. ![Sketch of a four-state Markov chain describing the motion of a molecular motor along a periodic track. The states of the motor are denoted with circles: M (free motor), MT (motor bounded to adenosine triphosphate ATP), MT (motor bounded to adenosine diphosphate ADP) and B (motor in the backtracked state). The transition probabilities are indicated with small letters. For bridges whose initial and final state is M, we investigate the statistics of the number of jumps from state MD to state M (number of cycles) and the fraction of time spent in state B. []{data-label="fig:5"}](model2.pdf){width=".35\textwidth"} All in all, the dynamics of the model is described in terms of the following transition matrix $$\bm\pi=\begin{pmatrix} (1-a-b) & d & t & r\\ a & (1-d-h) & s & 0\\ 0 & h & (1-t-s)& 0 \\ b & 0 & 0 & 1-r \end{pmatrix}\ .$$ Here, we are specifically interested in two frenetic quantities along bridges: (i) the total number of cycles completed, given by the number of irreversible translocation jumps, described by the tilted matrix $$\bm\pi_k=\begin{pmatrix} (1-a-b) & d & te^{-\mathrm{i}k} & r\\ a & (1-d-h) & s & 0\\ 0 & h & (1-t-s)& 0 \\ b & 0 & 0 & 1-r \label{eq:m1} \end{pmatrix}\ ,$$ and (ii) the fraction of time spent by the motor in the backtracking state, which can be evaluated using the tilted matrix $$\bm\pi_k=\begin{pmatrix} (1-a-b) & d & t& r\\ a & (1-d-h) & s & 0\\ 0 & h & (1-t-s)& 0 \\ b & 0 & 0 & (1-r)e^{-\mathrm{i}k} \label{eq:m2} \end{pmatrix}\ .$$ ![Statistics of two frenetic properties of the 4-state model of molecular motor with absolute irreversibility for bridges of different duration $T$ (see legend): number of cycles (top) and fraction of time spent on backtracking (bottom). Symbols correspond to $10^7$ bridges obtained with Monte Carlo simulations and the lines are our theoretical predictions evaluated using Eq.  for the matrices  and . Values of the parameters: $a=0.7$, $b=0.02$, $d=0.05$, $h=0.9$, $s= 0.05$, $t= 0.7$ and $r= 0.1$. []{data-label="fig:6"}](Fig4.pdf){width=".37\textwidth"} Motivated by usual analyses in enzyme kinetics, we are interested in bridges M$\to$M of duration $T$, i.e. trajectories for which the motor is in the free state both at time $0$ and at time $T$. We thus compute the full distribution  for the specific initial condition $p(x)=\delta_{x,1}$. We do not report the full expression but only two significant figures of merit. First, the distribution of the number of cycles is skewed and develops a cusp for small values of $T$ (Fig. \[fig:6\]a). The theory describes the numerical simulations perfectly. Secondly, the distributions of the fraction of time spent in the backtracking state display a double-exponential-like behaviour (Fig. \[fig:6\]b). Remarkably, the accuracy of our formula extends beyond the first four decimal digits and perfectly reproduces the empirical results including their bulk and wildest kinks. ![Illustrations of applications of Markov bridges in biophysics. (a) Left: Kinetic model of a $n-$state enzymatic reaction ($n\geq 3$) with one reversible step of step size $d$. Adapted from [@chemla2008exact]. Right: Illustration of a single-molecule trace of a molecular motor with step size $d$. (b) Two sample configurations of a pinned polymer loop containing $T=7$ monomers. Adapted from [@huang2018exactly]. (c) Examples of cyclic DNA templates containing $T=5$ basepairs. The black boxes highlight the boundaries of the cyclic templates. []{data-label="fig:11"}](apps2.pdf){width=".45\textwidth"} summary and outlook =================== In this paper we have developed a comprehensive finite-time theory for fluctuations of Markov bridges. We have exhaustively tested the theory for currents and frenetic quantities for two examples including absolutely irreversible transitions. Markov bridges are yet a largely unexplored universe of stochastic processes while constituting a rather natural framework to model biochemical cycles, as shown here. We believe that our theory could also be extended to describe fluctuations of continuous time Markov bridges and of periodically-driven processes as stochastic heat engines. Our results are of particular interest in biophysics. Figure \[fig:11\] illustrates three specific examples where our theory could have potential impact: (a) molecular motors; (b) polymer physics; and (c) DNA sequencing. Molecular motors move along periodic tracks executing cyclic enzymatic reactions. In single molecule experiments it is often hard to detect all the internal states of a motor but only the jumps between two adjacent sites in the motor’s track (Fig. \[fig:11\]a, right panel), which often corresponds to the completion of a cycle, i.e. a Markov bridge (Fig. \[fig:11\]a, left panel). Markov bridges may also be applied to describe the fluctuations of a pinned polymer loop of a fixed length $T$ (Fig. \[fig:11\]b). In DNA sequencing, finite-time bridges correspond to “cyclic” templates containing $T$ basepairs where the first and last nucleotides coincide (see Fig. \[fig:11\]c). Our theory provides insights on e.g. the distribution of how many basepair changes (frenesy) may occur along cyclic templates. We have provided evidence that the existing uncertainty relations are insufficient to describe the “cost of accuracy" required for constrained fluctuations and further work is needed in this direction. We hope that this paper will trigger further research on universal fluctuation relations for bridges. It would be particularly interesting to analyze whether our results about the information that frenesy carries about currents (Sec. \[sec:ivc\]) can be discussed within the context of time-symmetric probes of nonequilibria. We thank R. Belousov, K. Proesmans, P. Pietzonka, M. Polettini, D. S. Golubev and M. Baiesi for fruitful discussions. PV acknowledges the stimulating research environment provided by the EPSRC Centre for Doctoral Training in Cross-Disciplinary Approaches to Non-Equilibrium Systems (CANES, EP/L015854/1). Author contributions: ER and PV conceived and directed the project, performed simulations, developed the theory, discussed the results, and wrote the paper. natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{} , , , , ****, (). , , , , , , , ****, (). , , , ****, (). , **, vol.  (, ). , **, vol. (, ). , ****, (). , ****, (). , ****, (). , , , , ****, (). , ****, (). , ****, (). , ** (, ). , ****, (). , , , , , , , **, vol.  (, ). , ****, (). , , , , ****, (). , , , ****, (). , , , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , (). , ****, (). , , , ****, (). , , , (). , , , ****, (). , , , , (). , , , (). , (). , (). , , , ****, (). , ****, (). , ****, (). , (). , , , , ****, (). , , , (). , ****, (). , , , , ****, (). , ** (, ). , , , ****, (). , in ** (, ), vol. , p. . , , , ****, (). , ****, (). , , , in ** (, ), pp. . , , , , ****, (). , in ** (, ), pp. . , , , ****, (). , **, vol.  (, ). , ****, (). , in ** (, ), vol. , pp. . , ****, (). , ****, (). , , , , ****, (). , , , ****, (). , ****, (). , , , , , ****, (). , (). , , , ****, (). , ****, (). , , , ****, (). , ****, (). , , , ****, (). , , , , , , , , ****, (). , , , , , , , ****, (). , , , , , , , ****, (). , , , ****, (). , ****, (). , , , ****, (). , , , , ****, (). , , , ****, (). , ****, (). , ****, (). , , , ****, (). , ****, (). , ****, (). , ****, (). , , , , , ****, (). , , , , (). , ****, (). , ** (, ). , ****, (). , ****, (). , ****, (). , , , ****, (). , , , , ****, (). , , , , ****, (). , (). , , , , , , ****, (). APPENDIX {#appendix .unnumbered} ======== Proof of Eqs.  and  {#app:1} ==================== Consider the tilted matrix $$\bm\pi_k=\begin{pmatrix} r & q e^{\mathrm{i}k} & pe^{-\mathrm{i}k}\\ pe^{-\mathrm{i}k} & r & qe^{\mathrm{i}k}\\ qe^{\mathrm{i}k} & pe^{-\mathrm{i}k} & r \end{pmatrix}\ .$$Calling $\zeta=e^{-\mathrm{i}k}$, the eigenvalues of this circulant matrix can be written as $$\lambda_j=r+\frac{q}{\zeta}\omega_j + p\zeta\omega_j^2\qquad j=0,1,2\,,$$ with $\omega_j=e^{2\pi\mathrm{i} j/3}$. For $k=0$, $\zeta=1$, and the normalization factor in  equals $$\sum_{i}\lambda_i^T(0)=\sum_{j=0}^2 (r+q\omega_j+p\omega_j^2)^T\,,$$ whereas for the numerator of  we have to compute $$\sum_{j=0}^2\int_0^{2\pi}\frac{\text{d}k}{2\pi}e^{\mathrm{i}k J_T}(r+q e^{\mathrm{i}k}\omega_j+p e^{-\mathrm{i}k}\omega_j^2)^T\ .$$ Changing variables $z=e^{\mathrm{i}k}$, we obtain $$\sum_{j=0}^2\oint_{|z|=1}\frac{\text{d}z}{2\pi\mathrm{i}~z^{1+T-J_T}}\left(q\omega_j z^2 + rz+p\omega_j^2\right)^T\ .$$ This integral can be evaluated using residues, noting that the integrand has a pole of order $1+T-J_T$ at $z=0$. Therefore we need to compute $$\sum_{j=0}^2\frac{1}{(T-J_T )!}\lim_{z\to 0}\frac{\partial^{T-J_T }}{\partial z^{T-J_T }}\left(q\omega_j z^2 + rz+p\omega_j^2\right)^T$$ obviously valid for $T\geq J_T $. The derivative can be computed using the Faà di Bruno formula $$\frac{\text{d}^n}{\text{d}x^n}f(g(x))=\sum_{k=0}^n f^{(k)}(g(x)) B_{n,k}\left(g^\prime(x),g^{\prime\prime}(x),\ldots,g^{(n-k+1)}(x)\right)\ ,$$ in terms of Bell polynomials $B_{n,k}(x_1,\ldots,x_{n-k+1})$. The case $J_T=T$ can be computed separately without any difficulty. We use the identification $n=T-J_T\geq 1$, $f(y)=y^T$ and $g(x)=q \omega_j x^2+rx+p\omega_j^2$, whose derivatives are $$\begin{aligned} g^\prime(x) &=2q\omega_j x+r\nonumber\\ g^{\prime\prime}(x) &=2q\omega_j\ . \end{aligned}$$ Using $f^{(k)}(y)=[T]_k y^{T-k}$ (where $[T]_k$ is the falling factorial) we get eventually $$\frac{\partial^{T-J_T }}{\partial z^{T-J_T }}\left(q\omega_j z^2 + rz+p\omega_j^2\right)^T= \sum_{k=1}^{T-J_T }[T]_k (q\omega_j z^2 + rz+p\omega_j^2)^{T-k} B_{T-J_T ,k}\left(2q\omega_j z+r,2q\omega_j,0,\ldots,0\right)\ .$$ Taking the limit $z\to 0$, we recover Eqs.  and  in the Main Text. Proof of Eq.  {#app:2} ============= Consider now the tilted matrix associated with the clockwise current $J_T\in \{-T,\dots,T\}$ and the frenesy $\Phi_T\in \{0,\dots,T\}$ — the total number of jumps between different states — $$\bm\pi_{k,\phi}=\begin{pmatrix} r & q e^{\mathrm{i}k-\mathrm{i}\phi} & pe^{-\mathrm{i}k-\mathrm{i}\phi}\\ pe^{-\mathrm{i}k-\mathrm{i}\phi} & r & qe^{\mathrm{i}k-\mathrm{i}\phi}\\ qe^{\mathrm{i}k-\mathrm{i}\phi} & pe^{-\mathrm{i}k-\mathrm{i}\phi} & r \end{pmatrix}\ .$$ The eigenvalues of this circulant matrix can be written as $$\lambda_j(k,\phi)=r+q e^{\mathrm{i}k-\mathrm{i}\phi} \omega_j + pe^{-\mathrm{i}k-\mathrm{i}\phi}\omega_j^2\qquad j=0,1,2\,,$$ with $\omega_j=e^{2\pi\mathrm{i} j/3}$. For $k=\phi=0$, the normalization factor equals $$\sum_{i}\lambda_i^T(0,0)=\sum_{j=0}^2 (r+q\omega_j+p\omega_j^2)^T\ .$$ For the numerator of Eq.  in the Main Text (specialized to uniform initial condition) we have to compute $$\sum_{j=0}^2\int_0^{2\pi}\frac{\text{d}k}{2\pi}\int_0^{2\pi}\frac{\text{d}\phi}{2\pi}e^{\mathrm{i}k J_T+\mathrm{i}\phi\Phi_T}(r+q e^{\mathrm{i}k-\mathrm{i}\phi}\omega_j+p e^{-\mathrm{i}k-\mathrm{i}\phi}\omega_j^2)^T\ .$$ Changing variables $z=e^{\mathrm{i}k}$ and $w=e^{\mathrm{i}\phi}$, we obtain $$\begin{aligned} &\sum_{j=0}^2\oint_{|z|=1}\oint_{|w|=1}\frac{\text{d}z\text{d}w}{(2\pi\mathrm{i})^2}z^{J_T-1-T}w^{\Phi_T-1-T}\left(rwz+qz^2\omega_j+p\omega_j^2\right)^T\\ &=\sum_{j=0}^2\sum_{\ell=0}^T{T\choose\ell}r^{T-\ell}\oint_{|z|=1}\oint_{|w|=1}\frac{\text{d}z\text{d}w}{(2\pi\mathrm{i})^2}z^{J_T-1-\ell}w^{\Phi_T-1-\ell}(qz^2\omega_j+p\omega_j^2)^\ell\\ &=\sum_{j=0}^2\sum_{\ell=0}^T{T\choose\ell}r^{T-\ell}\sum_{m=0}^\ell {\ell\choose m}(q\omega_j)^m (p\omega_j^2)^{\ell-m}\mathcal{S}(-J_T+1+\ell-2m)\mathcal{S}(-\Phi_T+1+\ell)\ , \end{aligned}$$ where $$\mathcal{S}(n)=\oint_{|z|=1}\frac{\text{d}z}{2\pi\mathrm{i}}\frac{1}{z^n}=\delta_{n,1}\ .$$ Therefore, the only allowed values for $\ell,m$ are $\ell=\Phi_T$ and $m=(\Phi_T-J_T)/2$ an integer between $0$ and $\Phi_T$. In summary we get $$P(J_T,\Phi_T)=\frac{{T\choose\Phi_T}{\Phi_T\choose(\Phi_T-J_T)/2}r^{T-\Phi_T}\sum_{j=0}^2 (q\omega_j)^{(\Phi_T-J_T)/2}(p\omega_j^2)^{(\Phi_T+J_T)/2}}{\sum_{j=0}^2 (r+q\omega_j+p\omega_j^2)^T}\ , \label{eq:F4}$$ which holds if: (i) $\Phi_T-J_T\in 2\mathbb{Z}$; and (ii) $\Phi_T \geq (\Phi_T-J_T)/2\geq 0$ i.e. $|J_T|\leq \Phi_T$. Otherwise $P(J_T,\Phi_T)=0$. Further simplification can be achieved by noting that the product of binomials in the numerator can be simplified $${T\choose\Phi_T}{\Phi_T\choose(\Phi_T-J_T)/2} = \frac{T!}{\Phi_T! (T-\Phi_T)!} \frac{\Phi_T!}{(\Phi_T+J_T)/2! (\Phi_T-J_T)/2!} = \frac{T!}{P_T! N_T^+!N_T^-!}\,,\label{eq:F5}$$ where we have introduced the integers $$P_T\equiv T-\Phi_T ;\quad N_T^+ = (\Phi_T+J_T)/2;\quad N_T^- = (\Phi_T-J_T)/2\label{eq:F6}$$ denoted as the persistence time, the number of jumps in the clockwise direction and the number of jumps in the counterclockwise direction, respectively. Using Eqs.  and  in  we find $$P(J_T,\Phi_T)= \left(\frac{T!}{P_T! N_T^+!N_T^-!}\right)r^{P_T}q^{N_T^-} p^{N_T^+} \frac{ \sum_{j=0}^2 \omega_j^{N_T^- + 2N_T^+}}{\sum_{j=0}^2 (r+q\omega_j+p\omega_j^2)^T}\ .\label{eq:F7}$$ We now perform the sum in the numerator in , denoting $K\equiv N_T^- + 2N_T^+$, and using De Moivre’s formula: $$\sum_{j=0}^2 \omega_j ^K = \sum_{j=0}^2 \left(\cos\left(\frac{2\pi K j}{3}\right) + \mathrm{i} \sin\left(\frac{2\pi K j}{3}\right)\right)\, .$$ The imaginary part vanishes because $ \sum_{j=0}^2 \sin\left(\frac{2\pi K j}{3}\right) = \sin\left(\frac{2\pi K }{3}\right)+ \sin\left(\frac{4\pi K }{3}\right)=0$, which holds for all integer values of $K$. The real part yields $ \sum_{j=0}^2 \cos \left(\frac{2\pi K j}{3}\right) =3 \, \delta_{K, 3\mathbb{Z}}$. Therefore, $$\sum_{j=0}^2 \omega_j ^K = 3 \, \delta_{K, 3\mathbb{Z}}\,,$$ and consequently the sum in the numerator in  equals $$\sum_{j=0}^2 \omega_j^{N_T^- + 2N_T^+} = 3\, \delta_{N_T^- + 2N_T^+, 3\mathbb{Z}}\,.$$ This implies $J_T$ is quantized in units of $3$, i.e. $J_T\in 3\mathbb{Z}$. We thus arrive at the compact expression in Eq.  in the Main Text.

ArXiv

**The birth of strange stars:** **kinetics, hydrodynamics and phenomenology** **of supernovae and GRBs** J. E. Horvath *Instituto de Astronomia, Geofísica e Ciências Atmosféricas* *Rua do Matão 1226, 05508-900 São Paulo SP, Brazil* [email protected] Abstract We present a short review of strange quark matter in supernovae and related explosions, with particular attention to the issue of the propagation of the combustion in the dense stellar environment. We discuss the instabilities affecting the flame and present some new results of application to the turbulent regime. The transition to the distributed regime and further deflagration-to-detonation mechanism are addressed. Finally we show that magnetic fields may be important for this problem, because they modify the flame through the dispersion relations which characterize the instabilities. A tentative classification of explosive phenomena according to the value of the average local magnetic field affecting the burning and the type of stellar system in which this conversion is taking place is presented. As a general result, we conclude that “short” conversion timescales are always favored, since the burning falls in either the turbulent Rayleigh-Taylor (or even the distributed) regime, or perhaps in the detonation one. In both cases the velocity is several orders of magnitude larger than $v_{lam}$, and therefore the latter is irrelevant in practice for this problem. Interesting perspectives for the study of this problem are still open and important issues need to be addressed. Introduction ============ Intensive work in the 60´s and 70´s definitely established the concept of elementary constituents of nucleons (quarks and gluons). At increasing center-of-mass energy in experimental searches of the elementary components ([*partons*]{}) of protons and other hadrons revealed new physics in need of a theoretical framework to be developed. The theory of “new” strong interactions (as opposed to the “old” nuclear physics) was constructed in parallel, first focused on classification schemes (or, as is called today, flavor physics) and later on finding a theory to describe the dynamics. The development and success of gauge theories in the ’70s eventually leaded to a non-abelian version based on the $SU(3)_c$ symmetry group [@QCD] as a “natural” candidate for a theory of strong interactions. The fundamental quantum number carried by the elementary constituents (quarks) was named “color”, and consequently the dynamics involving quarks and gauge fields (gluons) become known as [*Quantum Chromodynamics*]{} (QCD for short). It was considered by many somewhat puzzling that repeated efforts to find these entities as free particles (asymptotic states) failed. Subsequent work elaborated on a striking feature of the theory: that the interactions themselves preclude the appearance of the quarks and gluons outside ordinary hadrons, they remain confined inside them at low energies. Another property was soon demonstrated to hold when momentum transfer scales $Q$ became large enough. This is the so-called [*asymptotic freedom*]{}, and states that the colored particles behave as if they were free in the limit $Q \, \rightarrow \, 0$. Actually, there is an energy (or momentum) scale above which color quantum number is not confined any more, but how large the momentum transfer should be (or in other words, which is energy, as measured by the temperature or density of the ensemble allowing the deconfinement) is still a matter of debate. These developments mean that the early universe passed through a deconfinement $\rightarrow$ confinement phase transition along its cooling, although less certainty holds for the densities of the “natural” laboratories (neutron stars) in which [*compression*]{} would deconfine hadronic matter. The earliest calculations [@Baymetal] using reasonable models for both the confined and deconfined phases imprinted on successive researchers the uncertain conclusion that quarks and gluons (forming a state known as the quark-gluon plasma, or QGP) should appear at densities above a large threshold, say, $10 \times \rho_{0}$; with $\rho_{0}$ the nuclear saturation density. From the starting of these calculations it has proved very difficult to reliably determine the transition points, and also the nature of the transition itself (at least when full numerical calculations were out of sight [@num]). Most of the times the conclusions had to be extracted from simultaneous extrapolations of both a quark model, expected to be valid for $\rho \rightarrow \infty$, and an hadronic model valid around $\rho_{0}$ but uncertain much above it. Since there is no certainty in either one, the final result is always subject to reasonable doubts. The “induction” of a definite [*order*]{} of the transition because of the adopted functional forms of the thermodynamical quantities of both sides. Nevertheless these serious and honest attempts have proliferated until today, given that the transition is still elusive (the extensively studied finite temperature case still has some small uncertainty in the value of $T_{c}$ and a quite consensual assessment of the order, see [@T] for details). Recent analysis [@RHIC] of hadronic flows have added a lot of excitement to these topics, since it appears that the QGP was indeed produced in heavy ion collisions, but the asymptotic form is [*not*]{} reached, rather behaving as a glass-like system. Needless to say, this kind of studies attract a lot of attention and offers a concrete form to glimpse the deconfined state of hadronic matter, yet to be characterized and understood. Stable strange quark matter? ============================= While the study of the quark-gluon plasma occupied many studies in connection with the early Universe and compact stars, a much radical proposal emerged in the 80´s about it, which may be described as follows: it is true that the asymptotic freedom property guarantees that quarks and gluons will be the ground state of QCD at high densities/temperatures, but it says nothing about the ground state at lower densities or temperatures. The everyday experience strongly suggests that ordinary hadrons confine the quarks/gluons and thus constitute the “true” (in the sense of $\rho \rightarrow 0$ and $T \rightarrow 0$) ground state of hadronic matter. The emerging [*strange matter hypothesis*]{} came precisely to challenge this “common sense” statement: it says that the true ground state of hadronic matter is a particular form of the QGP , differing from the ordinary matter by the presence of a key quantum number (strangeness). This is counterintuitive to many people, but a careful look at the physical arguments shows no inconsistency whatsoever, at least in principle. An argument for the SQM being the true ground state can be made as follows: as is well-known the quantity that determines which phase is preferred is the Gibbs free energy per particle $G/n$ as a function of the pressure (we impose $T=0$ hereafter as appropriate for highly degenerate hadronic matter, it is easy to see that the term $-TS$ in the free energy disfavors SQM at high temperatures). As $P$ is increased starting from the neighborhood of the nuclear matter saturation point $\rho_{0}$ the asymptotic freedom guarantees that there has to be a switch from nuclear matter (N) to elementary hadronic constituents, that is, the lighter quarks $u$ and $d$. The point at which this is supposed to happen will be labelled as $P_{c}$. Thus, the doubts stated above about the appearance of the QGP inside neutron stars may be now restated as whether the pressure at the center is larger or smaller than $P_{c}$. However, this is where the concept of strangeness plays an important role. Strangeness is the flavor quantum number carried by $\Lambda$s and other heavy hadrons. At the elementary level, it is carried by a different quark $s$, with current mass in the ballpark of $\sim 100 \, MeV$, that is, light enough to be present at a few times the nuclear saturation density. While creating strangeness in hadrons costs energy (because strange hadrons are heavier than non-strange ones; for instance, the $\Lambda$s are heavier than the neutrons and so on); this trend is reversed inside the QGP. The reason is simply the Pauli exclusion principle: a new Fermi sea in the liquid (the one of the $s$ quark) allows a rearrangement of the energy, and sharing it lowers the energy per particle. However, the gain is not precisely known, but it is not impossible to imagine lowering the free energy per particle to a value that would be lower than the mass of the neutron $m_{n}$ even when $P \rightarrow 0$. If realized, this would preclude the (strange) QGP to decay into ordinary hadrons because this would [*cost*]{} energy and the SQM would have been created. Put it simply, the compression would liberate the elementary components that quickly create their own way of surviving. We stress that all these are bulk (i.e. large number) concepts, and it is central to the SQM hypothesis to reach a strangeness per baryon of the order one (and exactly one if the strange quark had no mass to deplete its relative abundance). This is not possible in a few-body system like a nucleus, because each weak decay creating a strangeness unit contributes roughly with a factor $G_{Fermi}^{2}$ to the amplitude, and thus the simultaneous decays are strongly suppressed; this is why it has been very difficult to produce even doubly strange nuclei, let alone higher multiplicity ones. However, once quarks roam free in the QGP they can easily decay by $u + d \, \rightarrow \, u + s$ because there is plenty of phase space for the products until equilibrium is reached. These bulk estimates have been always one way or another behind the idea of SQM. As it stands, the SQM hypothesis is very bold. It conjectures that every hadron we see around us is in a metastable state, and if conditions for creation of a large net strangeness were met, the matter would not make back ordinary hadrons (technically it is said that SQM constitutes a non-topological soliton stabilized against decays by a conserved charge, the baryon number, see [@FOGA] for a thorough discussion of this case and related ones). The general idea of reaching extreme conditions and stabilizing the QGP is already apparent in the paper of Bodmer [@Bodmer], later reintroduced and refined in references [@CK; @Tera; @Bjo] and colorfully discussed in the paper of Witten [@Ed], which was fundamental to give a big boost to SQM research. SQM as a theoretical construction is interesting, but finding it in nature would be infinitely more. Key questions of SQM such as whether it does exist or not, and whether it has been ever produced in the Universe are still unsolved. On the other hand, we begin for the first time to have the possibility of falsifying these basic questions mainly thanks to the new generation of space telescopes (HST, Chandra, XMM) and neutrino observatories (SNO, Kamioka, Icecube), to name just a few. These instruments may be used to look for exotic states in compact stars and their birth events. Many applications of SQM in astrophysics were foreseen during the first decade after its official birth [@Ed] and early infancy [@FJ] Since astrophysical insight has shown to be essential in the determination of fundamental questions related to SQM, we shall focus briefly in a very definite (and important) astrophysical problem, trying to give an assessment and pointing on the uncertainties and possible directions that may be explored in the near future. We thus restrict our discussion to SQM in compact stars, and more specifically, on how a seed of SQM may grow and propagate throughout a just-born neutron star. This has been a popular choice for an energy source in GRBs and core-collapse supernovae, therefore it is important to establish its basic features with confidence to build on them. SQM in protoneutron stars: effects in core-collapse supernovae ============================================================== As a “natural” environment in which SQM might form, core collapse supernovae has received reasonable attention [@BHV; @Matt; @chinos; @Anand]. Despite of more than three decades of theoretical research and hard numerical modelling, the processes that cause the explosion of massive stars are still not understood ([@Bur]). If, as the more recent and detailed numerical simulations suggest, the neutrino-driven mechanism works on special conditions only, the current paradigm for explaining massive star explosions would have to be deeply revised. “Conventional” physics has now turned attention to the role of rotation inside the progenitor and magnetic fields [@SN1; @SN2], possibly relating this problem to the GRB one [@Massimo]. Although it is still too early for making definitive conclusions, investigations including the possible transition to deconfined QCD phases may be relevant to this problem. The first studies of SQM in supernovae ([@Mac; @BHV; @LBV; @DPL; @Anand]) showed that this hypothetic subnuclear energy source is more than adequate to contribute to the explosion, and that some observed characteristics in the neutrino emission of SN1987A may be naturally explained within this scenario ([@Hat; @SatSuz; @Mac]) (a second peak in the neutrino emission is naturally predicted in these models, and such signal has been tentatively associated to the late neutrinos from SN1987A detected by Kamiokande, which have to be otherwise interpreted as a statistical gap within the current paradigm). From a wide perspective, supernovae are perhaps the only astrophysical events in which we could have the possibility of making a “multiwavelength” detection (neutrinos, various electromagnetic wavelengths, gravitational waves) of the process of SQM formation. However, these calculations are still in the infancy, and just bold expectations have been formulated. Some specific simulations [@FryerWoos] have addressed (negatively) a few questions posed in GRB models. In addition, a firmer detailed observational background would be needed, which imperatively needs the occurrence of a number of supernova explosions in the neighborhood of our galaxy, and thus is out of any human control (in turn the instrumentation must be improved greatly). Second, although the general picture of SQM formation in supernovae has been qualitatively constructed, no systematic calculations have been made. There are also many unresolved questions related to strong interactions at high densities, which introduce an uncomfortable degree of uncertainty in all conclusions. We shall attempt below to describe the basics of the SQM propagation problem, a subject that has been addressed in the literature over the years from the kinetic/energetics point of view [@Olinto; @MadsenOlesen; @Heiselberg; @Drago; @hindues], but has a high degree of complexity from the coupling to hydrodynamics, much in the same way thermonuclear supernovae do. We will be guided by the work done in the latter problem, even though most of our discussion is new (i.e. unpublished) for the specific problem of SQM propagation. We shall later attempt to sketch the effects of the magnetic fields for the propagation, which leads to a tentative classification of the different phenomenological events. SQM combustion dynamics: early stages ===================================== As discussed and agreed in the literature, a seed of SQM must become active or form following the standard bounce onto the former iron core. We shall not address this problem of the seed here, just assuming that by some of the proposed mechanisms [@AFO; @Mac] the seed of SQM is present within $\sim \, seconds$ after the bounce (if the quick appearance is bypassed, a late conversion could ensue [@Bombaci] but without effects in supernovae). The neutron-SQM interface must then propagate outwards powered by the energy release of converted neutrons, much in the same way as a laboratory combustion. It seems reasonable to assume the combustion to begin as a [*laminar*]{} deflagration, in which the diffusion of $s$ quarks set the scale for the flame length $l_{th}$. This has been actually the subject of early calculations [@varii; @Olinto; @MadsenOlesen; @Heiselberg], in which a plane front approximation was used to obtain the laminar velocity ${\bf u}_{lam}$ as a function of temperature, density and other relevant quantities. The result ${\bf u}_{lam} \leq \, 10^{4} cm \, s^{-1}$ suggested that a just-born NS would convert to a SS in $\sim \, 100 \, s$ or so. From the combustion theory point of view this is equivalent to decouple completely the kinetics of the burning from the hydrodynamics of the flow in the star. Nevertheless, the reasonable convergence of several approaches to the calculation of ${\bf u}_{lam}$ gives some confidence that the result is reliable within the approximations. In a situation as such (a combustion starting around the center and propagating outwards), it has been known for many years [@Darrieus; @Landau] that small perturbations are unstable for all wavelengths at a linear level. In fact Horvath and Benvenuto [@HB] calculated the perturbation growth for this specific problem with the resulting condition j\^[4]{} &lt; 4 g \_[1]{}\^[2]{} \_[2]{}\^[2]{} [1]{} where $j$ is the mass flux onto the flame, $\sigma$ the surface tension, $g$ the local gravitational acceleration and $\rho_{1}, \rho_{2}$ the densities of the “fuel” (neutron matter) and “ashes” (SQM) respectively. As it stands, this is impossible to satisfy for [*any*]{} deflagration (in particular, the laminar), because by its very definition $j^{2}=(P_{2}-P_{1})\rho_{2}\rho_{1}/(\rho_{2} - \rho_{1})$, and thus a deflagration which must obey $P_{2} < P_{1}$ and $\rho_{2} < \rho_{1}$ making the r.h.s a negative number. This way, the flame wrinkles in a timescale $\leq$ the dynamical timescale $\tau_{dyn} \sim \, 10^{-3} \, s$ (as appropriate in a protoneutron star). Thus, the strong statement made by Landau and Darrieus is confirmed at the linear level. Numerical calculations of the Landau-Darrieus instability beyond the linear level [@RoepkeHill] show the formation of [*cusps*]{}, leading to quadratic and higher-order terms in the dispersion relation and stabilizing the flame [@Zeld]. The flame acquires a cellular shape and accelerates, since the contact area between the fuel and ashes increases. The stationary, scale-invariant amplitude of this cusps leads to an acceleration of the flame, with velocity described in this regime as \_[cell]{} = [**u**]{}\_[lam]{} [(1 + 0.4(1 - )\^[2]{})]{} with $\mu = \rho_{2}/\rho_{1}$. The flame velocity is higher than in the laminar regime by a modest amount for all reasonable compression ratios $\mu$. An alternative cellular flame model has been developed by Blinnikov and Sasorov [@BlinSas]. They observe that the wrinkled flame can be described with a fractal model, which in the 2-D case yields \_[cell]{} \_[lam]{} [( [l]{})\^[D\_[cell]{}-2]{}]{} with $D_{cell}$ the fractal dimension of the surface and $l_{crit}$ a suitable minimum length. A calculation of the latter quantity finally yields \_[cell]{} \_[lam]{} [( [l\_[max]{}]{})\^[0.6(1-)\^[2]{}]{}]{} where we have imposed the radial distance $l_{max}$ as the maximum scale for which this theory is valid. Arguments related to the propagation of L-D unstable flames suggest that $l_{crit}$ may be identified with the [*Markstein*]{} length [@Markstein], or at least $\sim \, 100 l_{th}$. Eq. (4) leads to the same conclusion as before: there is a modest acceleration of the flame and stabilization at the small scales. In Ref.[@HB], the extreme assumption that velocity of the flame can not become supersonic, we obtained a (small) maximum length for this regime to hold. This should be rather interpreted as the scale beyond which the above L-D description breaks down definitely, due to combined additional physical effects that we now address. Since the gravitational pull is always being exerted onto the flame, one could have anticipated that the cell structure can not be scale-invariant indefinitely, and in fact disruption of the bubbles does occur [@NieHill95]. A [*turbulent cascade*]{} dominates the burning above certain length, which can be estimated from the point when the velocity of turbulent fluctuations ${\bf u'}(l)$ becomes equal to ${\bf u}_{cell}$. This defines the so-called [*Gibson scale*]{} $l_{gib}$ [@Peters]. Imposing a Kolmogorov spectrum (it is now established that this is more accurate than the so-called Bolgiano-Obukhov spectrum for 3D, whereas the latter is relevant for 2D models) ${\bf u'}(l) \propto l^{1/3}$, it can be shown that the scaling of $l_{gib}$ is l\_[gib]{} \^[3]{} where the fluctuations have to be normalized to the largest scales $L$ encountered in the system. Given that the turbulence itself can not become supersonic (the speed of sound is already $\sim \, c$ in the problem), and using the former value ${\bf u}_{cell} \geq 10^{4} cm \, s^{-1}$, we obtain for $l_{gib}$ the value of $\sim 10^{-4} cm$ as an extreme upper limit, and decreasing with time. This is, however, initially much larger than $l_{th}$ in the diffusive regime, and allows a classification of the burning into the [*flamelet*]{} regime: while the flame propagation is still determined by diffusion, the [*total burning*]{} is in turn controlled by turbulence in a turbulent region called the [*flame brush*]{}. In the flamelet regime, for all scales $\gg l_{gib}$, the turbulent velocity ${\bf u}_{turb}$ and front width $l_{turb}$ are determined by the Kolmogorov spectrum at the larger scales. The important point to stress here is that the turbulent eddy turnover controls the transport and fuel consumption, quite unlikely a pure laminar regime [@Kerstein; @Clavin]. Diffusion processes do not play the dominant role once the flamelet regime is achieved, quickly after the start of the combustion. We note that if $l_{gib}$ decreases below the value of $l_{th}$ one can no longer talk of a laminar regime and the burning is likely described by the [*distributed*]{} regime, in which turbulent eddies disrupt the flame and dominate the burning on macroscopic and microscopic scales. We shall assume that the flamelet regime exists and proceed to describe the large-scale physics, keeping in mind the possibility of being bypassed in favor of the distributed regime. SQM combustion dynamics: turbulent large-scale regime ===================================================== While at the small scales L-D instability affects the flame eventually leading to the flame brush in the way described above, on still larger scales, buoyancy of hot burned fuel (SQM) dominates the dynamics of the process as a consequence of the Rayleigh-Taylor instability. In fact, one obtains essentially the R-T results by letting ${\bf u}_{lam} \rightarrow 0$ in the L-D analysis. The classical solution of this problem [@Chandra] is well-known and indicates that in the linear regime the perturbations grow exponentially with a time scale $\tau_{RT} = (4\pi l/g)^{1/2} (\rho/\Delta \rho)^{1/2}$, with $\Delta \rho = \rho_{1}-\rho_{2}$. After the modes attain amplitudes similar to the originally unperturbed, the merging/fragmentation of the bubbles and the shear [*Kelvin-Helmholtz*]{} instability between bubble surfaces give rise to a turbulent mixing layer. In models with a single bubble scale, the velocity is \_[RT]{} [1]{} [1]{} with $At$ the [*Atwood number*]{} and $l$ the radius of the tube in the experiment [@DaviesTaylor]. In astrophysical problems a single-scale expression is seldom enough to describe the intrinsic multi-scale system, and a 1-D model containing most of the relevant physics, the so-called [*Sharp-Wheeler*]{} model [@SW] is widely used to calculate evolution of the flames. In this picture the combustion front advances into the cold unburned fuel with a speed of \_[SW]{} \~ [1]{} (1 - ) g t It is clear that the bubble radius evolves linearly with the distance to the center. The Sharp-Wheeler speed eq.(7) can be identified with the effective speed of the burning provided the latter is completed inside the R-T mixing zone. Fractal models have also been employed as an alternative description for the R-T regime [@TimWoos; @Ghezzi], with a generic prediction that can be summarized as \_[R-T]{} = [**u**]{}\_[lam]{} [( [L]{} )]{}\^[n/2]{} where $n=2(D-2)$ relates the index to the fractal index $D$, $L$ is the scale at which the turbulent velocities equal the R-T instability velocity, and $l_{min}$ is the smallest scale that can still deform the flame front (bounded from below by the Gibson scale defined above). For both the Sharp-Wheeler model eq.(7) and the fractal model of eq.(8) the velocity increase is very large respect to the “kinetic” laminar models for the same problem. This is quite analogous to the carbon burning regime in type I supernova models, in which all the hydrodynamical aspects are being considered together with the reaction kinetics. It is important to remark that in all the cases the flames can be still defined quite properly, and that energetics determined by Hugoniot curves are still valid, as they should. As performed in Type I SN studies, we plot in Fig. 1 the relevant velocities for the burning flame as a function of the scale. From a simple inspection of this figure, it is clear that the $n \rightarrow SQM$ combustion should accelerate substantially when evolving at relatively low radii (certainly $\ll \, 1 km$, still deep in the stellar interior. Further evolution of the flame will be discussed below, ending with some of the expected consequences and phenomenological features. SQM combustion dynamics: distributed regime and the transition to the detonations branch (DDT) ============================================================================================== The evolution of the flames described below is now quite well established and substantiated by several numerical simulations. One may wonder about the final outcome of the burning process when the flame is well within the R-T stage. The possibility that turbulence disrupts the flame on microscopic scales, which would not be well-defined any more, can be adopted as a rough intuitive description of the [*distributed regime*]{}. In the latter mixed regions of fuel and ashes burn in regions that have a distribution of temperatures interact strongly with the turbulence. Alternatively, the combustion may reach the edge of the star without reaching the distributed regime. More rigourously, the distributed regime can be characterized by the inequality $l_{gib} \, < \, l_{th}$. It is not clear whether this condition is achieved in the $n \rightarrow \, SQM$ conversion. As suggested above, it may be achieved directly in the early stages. However, and in spite that $l_{gib}$ decreases along the propagation, [**u’**]{} is clearly bounded from above by $c$. Therefore, $l_{gib}$ may be short for the distributed regime to be reached if it is not reached in the early stages, and this is a point that needs a detailed investigation. One may nevertheless entertain the possibility of a distributed regime in the problem because it is one of the expected pathways to the [*detonation*]{} branch of the combustions. In these scenarios a detonation (self-propagated combustion mediated by a shock) can start, for example, by means of the Zel’dovich gradient mechanism [@Zel2]. For this to occur, a macroscopic region of the mixed fuel/ashes should be able to burn “at once” (i.e. allowing a supersonic phase velocity), which requires a very shallow temperature gradient $\nabla T$. It is not known how large the critical macroscopic region should actually be, detailed calculations show that its value for the WD carbon burning problem is $L_{c} \sim 10^{4} cm$, and it is likely much smaller in our problem. Estimations of the size of the distributed flames yield essentially $l_{dist} = \alpha l_{th} Ka$, where $\alpha$ is a pure number and $Ka$ is the [*Karlovitz number*]{}, used in turbulence studies as a measure of the quotient of diffusive to eddy turnover time. Physically, if $l_{dist}$ can stretch to reach the $L_{c}$ value, the system would satisfy at least a necessary condition for a transition to detonation (since the deflagrations come first, this is call in the literature as Deflagration-to-Detonation Transition, or DDT, [@Khlo]). This condition can be combined with the expression $l_{gib} = l_{th}/Ka^{2}$ to yield the relation \^[1/3]{} Ka [( [L\_[c]{}]{} )]{}\^[1/3]{} converted into a bound on $\alpha$ when we observe that the distributed regime starts at $Ka \, > \, 1$. $L_{c}$ values larger than $\sim 10 \, cm$ would not allow the burning to become a detonation (DDT) by the Zel’dovich gradient mechanism. Thus, a necessary, but not sufficient, condition for the DDT can be established whenever $L_{c} \, \leq \, 10 \, cm$. Another condition for DDT within the gradient mechanism is related to the hierarchy of time scales of mixing, burning and dynamical. Contrary to the WD explosion problem, we have already seen that $\tau_{dyn}$ is always much longer than $\tau_{burn}$ (identified with the weak interaction time scale $\tau_{W} \, \sim 10^{-8} \, s)$ to create the strangeness). Therefore we have \_[mix]{} \_[W]{} \_[dyn]{} which yields, after substituting \_[W]{} using for the turbulent velocity fluctuations the estimate ${\bf u'}(L) = (1/2) \sqrt {g_{eff}L}$ [@NiemWoos], we obtain an [*upper*]{} bound on $L_{c}$ L\_[c]{} 10\^[-5]{} (1 - ) cm This is a small length over which to mix fluids, and would make the former condition on the Karlovitz number (eq. 9) irrelevant, likely leading to a DDT phenomenon immediately. Other mechanisms for DDT do exist, but is too difficult to discuss them in connection with our problem at this stage. From all the above discussion we believe it is clear that the examination of the laminar diffusive regime is just a part of the whole very complex problem. The full evolution of the burning $n \, \rightarrow \, SQM$ can be accurately described by using the so-called [*reactive Euler equations*]{} + [. ()]{} = 0 + [. ()]{} + P = 0 + [. ((E+P) [**u**]{})]{} = 0 + [([**u**]{}.)X\_[i]{}]{} = R (T, , X\_[i]{}) with $X_{i}$ are the fraction of each quarks and the reaction rates $R (T, \rho, X_{i})$ have to be calculated at finite temperature for the dense environment (see, i.e. [@Anand]). Due to enormously disparate length scales, ranging from $\sim$ few $fm$ to perhaps $\sim km$, a model of the flame can facilitate the calculations, otherwise it is known that resolving the full structure demands a huge computational investment [@WoosEtAlnew]. Role of $n \rightarrow SQM$ conversion in supernovae ==================================================== In the original proposal [@Mac; @BHV] of a fast combustion mode in supernovae, a newtonian calculation was employed to estimate the dynamical quantities, in particular the energy that could be transferred to the outer layer of a stalled shock in a massive star. We have seen that a complex but quick sequence of phenomena affects the flame, even if initially starts as a slow laminar combustion. If energy can not be directly transferred to the outer layers, SQM formation may still be important because of the production of neutrinos by appropriate reactions in the deconfined phase. The binding energy of the strange star has to be released as well [@BomDat], much in the same way as the binding energy of the neutron star in the standard picture. Although new fresh neutrinos could in principle produce a late revival of the stalled shock wave, other features than the total released energy are essential such as spectral features of the neutrino emission, and more importantly (if the transition happens to be somewhat delayed) the exact time of its occurrence, since if it occurs too late there will be no way to explode the star by the shock reheating mechanism at all. While it is still not clear whether the detonation mode is feasible, since it requires fast transport of heat to sustain the front and a working DDT mechanism (if it is not initiated “directly”), assuming the latter case, and since the conversion is not expected to be exothermic all the way down to zero pressure it is unavoidable that a detonation will become a standard shock wave beyond some radius (assuming the MIT Bag model for SQM this radius is the one for which $E - 3P= 4B$). This shock wave will propagate outwards and the question is whether or not it will be able to transfer its energy and complete the work unfinished by the unsuccessful prompt shock wave. In turn, a more moderate turbulent combustion (subsonic but still very fast) may be the final outcome instead of a detonation, and its propagation would mix the material on macroscopic scales due to the action of Landau-Darrieus and Rayleigh-Taylor instabilities. Its role in the reenergization of the stalled shock, possibly by neutrino transfer, has not been calculated as yet. A better understanding of the previous sequence of combustion processes will also give information about the timescale of the conversion of the star, which is closely related to the different observational signals. These calculations also constitute an important task for the near future. Delayed conversions, compact star structure and gamma-ray bursts ================================================================ Up to now we have considered the hydrodynamics of the reactive flows with the assumption of its occurrence well inside the first seconds after the prompt shock bounce. If the just-born protoneutron stars do not collapse to black holes due to accretion in the early stages [@LatPra], and within the SQM hypothesis, then pure strange stars, made up entirely of strange quark matter from the center to the surface, may be the compact remnants of supernovae. But even in the case of absolute stability, if the transition is [*not*]{} triggered during the supernova explosion, all observed “normal” neutron stars would be in a metastable state, which is quite difficult to imagine because of ISM contamination arguments [@AFO; @Jes; @MTH] and the mismatch $\tau_{conv} \, \ll \,\tau_{star}$ between the timescale in which favorable conditions for conversion occurs $\tau_{conv}$ and the lifetime of the star $\tau_{star}$. According to recent calculations the deconfinement transition is more likely to occur by heating and compression during the Kelvin-Helmholtz phase of proto-neutron star (PSN) evolution (see, for instance, [@LuBe]). If it did not happen there, once the PNS has cooled to temperatures below $\sim \, 1 MeV$, only accretion from a companion star or strangelet contamination would allow the transition (and many barriers may preclude its occurrence), even in the case where it is energetically favored. Thus, the existence of strange stars is determined not only by fundamental questions concerning the true ground state of dense matter but also by the exact physical conditions in the specific astrophysical environments together with the plausibility of the conversion mechanisms in these situations. The SQM conversion has been repeteadly associated with gamma-ray bursts. Many works in the past have explored the idea that the conversion of NM into SQM in NSs may be an energy source for GRBs ([@AFO1; @MX; @Hae; @CD; @BomDat; @Chinos2; @Rach; @PacHaens; @Berezhiani]). These models mostly address spherically symmetric conversions of the whole NS rendering isotropic gamma emission. Accumulating observational evidence suggests that at least “long” GRBs are strongly asymmetric, jet-like outflows, a feature that needs some crucial ingredient in the SQM physics formation/propagation to proceed. The “short” burst subclass was not obviously asymmetric prior to HETE2 and SWIFT data, but now evidence has mounted for a substantial asymmetry (but not extreme) in them. The association of Type Ib/c with a few GRBs has reinforced the investigation of underlying explosion mechanisms, and the absence of a temporal break in most of the light curves (interpreted in terms of a collimated jet effect) is a puzzling feature that might be related to the total energy budget in a yet unclear resolution. A new potentially important feature recently recognized in this class of models is that if a conversion to SQM actually begins near the center of an NS, the presence of a moderate magnetic field B ($ \sim 10^{13}$ G) will originate a prompt [*asymmetric*]{} gamma emission, which may be observed as a short, beamed GRB after the recovery of a fraction of the neutrino energy via $\nu {\bar\nu} \rightarrow e^{+}e{-} \rightarrow \gamma \gamma$ [@Lug]. The basic physical effect is again related to the instabilities described in the former sections: the influence of the magnetic field expected to be present in NS interiors quenches the growth of the hydrodynamic instabilities in the equatorial direction of the star (parallel to the magnetic field) while it allows them to grow in the polar one. As a result, the flame will propagate much faster in the polar direction, and this will result in a strong (transitory) asymmetry in the geometry of the just formed core of hot SQM, which will resemble a cylinder orientated in the direction of the magnetic poles of the NS. While it lasts, this geometrical asymmetry gives rise to a bipolar emission of the thermal neutrino-antineutrino pairs produced in the process of SQM formation. This is because almost all the thermal neutrinos generated in the process of SQM formation will be emitted in a free streaming regime through the polar cap surface, and not in other directions due to the opacity of the matter surrounding the cylinder. The neutrino-antineutrino pairs annihilate into electron-positron pairs just above the polar caps of the NS, giving rise to a relativistic fireball, thus providing a suitable form of energy transport and conversion to gamma-emission that may be associated to short gamma-ray bursts. A unifying scheme in which SQM appearance produces spherical ejection phenomena to highly asymmetric gamma beaming, as a more or less continuous function of the magnetic field $B$ and the astrophysical system under examination may be possible, and is tentatively sketched in Table 1. [**Table 1.**]{} Tentative classification of explosive events due to SQM in several stellar systems Mag. field (G) Type II SN LMXB-HMXB$^{\ast}$ AIC(?)$^{\dagger}$ -------------------- ------------- -------------------------- --------------------- -- $0 < B < 10^{12}$ “normal” SN spherical,weak short GRB UV-X flash $B \sim 10^{13}$ bipolar SN bipolar,strong short GRB bipolar UV-X flash $ B \geq 10^{14}$ ? jet-like,weak short GRB jet-like UV-X flash $B \gg 10^{15-16}$ – -no SQM formation- – $\ast$ only if $NM \rightarrow SQM$ conversion is sometimes suppressed when a NS is formed. $\dagger$ upper limit to the rate $\sim 10^{-4} yr^{-1} galaxy^{-1}$ needs to be revised if SQM burning occurs modifying nucleosynthetic yields. We are still very far from a thorough understanding of magnetic field effects, and a reliable simulation is even more challenging than simulating the $B=0$ reactive Euler equations (13-16). However, we believe that it is fair to say that magnetic fields are relevant for the physics of the conversion, even at moderate values. In summary, we may be witnessing an ultimate subnuclear energy source in action, powering SN-GRBs if SQM exists. Conclusions =========== We have presented a discussion of the main features of hypothetic $n \rightarrow SQM$ conversions inside neutron stars. We have shown that even if the initial state of the process could be a laminar deflagration, the hitherto ignored hydrodynamic instabilities (Landau-Darrieus and Rayleigh-Taylor) quickly take over and determine the propagation through the vast majority of the star, in a regime of turbulent deflagration [@HB; @Lug]. Models which ignore hydrodynamics altogether or just concentrate on the energy conditions to determine the combustion mode miss completely this important features. In particular, the association of long timescales (up to $10^{3}-10^{4} \, s$) of GRBs based on the identification of a laminar deflagration as the relevant timescale in the process is not tenable [@HZu]. Other proposed models differ in their kinetic aspects, for example, models in which energy is obtained by pairing quarks [@LH2; @LH3; @Hsu; @Ouy] typically operate on strong interaction timescales, and thus may be thought as an isocoric burning, i.e. much faster than the described instability scenario. Still other energy transfer mechanisms are possible [@Xu], and certainly the issue of neutrino transport from the reaction zone ahead has never been addressed in detail [@BenLug], although there is plenty of energy carried by them. It is still possible that all these regimes could be bypassed in favor of a “prompt” detonation mode started at the very central region, for example, by the sudden conversion of a macroscopic small region, further sending a shock wave with $\sim$ half of the initial overpressure [@Mamaia]. Propagation of such a combustion mode is in principle possible [@Mac; @Tokareva], but more detailed studies have yet to be performed on this problem by coupling properly the energy transport to the structure of the flame front. Models treating the conversion much in the same way as a plain phase transition are even more remotely relevant to the actual physics. Acknowledgements ================ Along several years of SQM research many people contributed to clarify and explain features of the physics and astrophysics of SQM to the author, and provided guidance in many respects. Among them we wish to acknowledge O.G. Benvenuto, H. Vucetich, G. Lugones, J.A. de Freitas Pacheco and I. Bombaci, colleagues and friends. The authors wish to thank the São Paulo State Agency FAPESP for financial support through grants and fellowships, and the partial support of the CNPq (Brazil). [99]{} see for example D.J. Gross, hep-ph/9210207 (1992); F. Wilczek, Int.J.Mod.Phys.A8:1359-1382,1993 and references therein. G. Baym and S. A. Chin, [*Phys. Lett. B*]{}[**62**]{}, 241 (1976); J. C. Collins and M.J. Perry [*Phys.Rev.Lett.*]{}[**34**]{}, 1353 (1975). see, for example, C.R. Allton et al., hep-ph/0206200 (2002) and references therein. S. Ejiri et al., hep-lat/0209012 (2002) and references therein. see for example, D. Karzeev, talk given at the XX International Symposium on Lattice Field Theory “Lattice 2002”, Cambridge, MA, June 24 - 29, (2002). J. Frieman, A.V. Olinto, M. Gleiser and C.Alcock, [*Phys. Rev. D*]{}[**40**]{}, 3241 (1989). A. R. Bodmer, [*Phys. Rev.D*]{}[**4**]{}, 1601 (1971). S.A. Chin and A.K. Kerman, [*Phys. Rev. Lett.*]{}[**43**]{}, 1292 (1979). H. Terazawa, Tokyo U. Report INS-336 (1979). J. D. Björken and L. McLerran, [*Phys. Rev. D*]{}[**20**]{}, 2353 (1979). E. Witten, [*Phys. Rev. D*]{}[**30**]{}, 272 (1984). E.Farhi and R.L. Jaffe, [*Phys. Rev.D*]{}[**30**]{}, 2379 (1984). O.G.Benvenuto, J.E.Horvath and H.Vucetich, [*Int. Jour. Mod. Phys. A*]{} [**4**]{}, 257 (1989). N.A. Gentile; M.B. Aufderheide; G.J. Mathews; F.D. Swesty and G.M. Fuller, [*Astrophys. J.*]{}[**414**]{}, 701 (1993). M. Takahara and K. Sato, [*Phys. Lett. B*]{}[**156**]{}, 17 (1985). J.D. Anand, A.Goyal, V.K. Gupta and S. Singh, ,[*Astrophys. J.*]{}[**481**]{}, 954 (1997). R.Buras, M. Rampp, H-Th. Janka and K. Kifonidis, [*Phys. Rev. Lett.*]{} [**90**]{}, 241101 (2003). A. Burrows, L. Dessart, E. Livne, C.D. Ott and J. Murphy, astro/ph0702539 (2007) A. Mezzacappa, J.M. Blondin, O.E.B. Messer, S.W. Bruenn,in [*AIP Conference Proceedings*]{}[**847**]{}, 179 (2006). O.G.Benvenuto and J.E.Horvath, [*Phys. Rev. Lett.*]{} [**63**]{}, 716 (1989). M. Della Valle, [*Chin. J. Astron. Astrophys. Supp.*]{}[**6**]{}, 315 (2006). G. Lugones, O.G.Benvenuto and H. Vucetich, [*Phys. Rev. D*]{} [**50**]{}, 6100 (1994). Z. Dai, Q. Peng and T. Lu,[*Astrophys. J.*]{}[**440**]{}, 815 (1995). T. Hatsuda, [*Mod.Phys.Lett.A*]{}[**2**]{}, 805 (1987). K. Sato and H. Suzuki, [*Phys.Rev.Lett.*]{}[**58**]{}, 2722 (1987). C. Fryer and S.E. Woosley, [*Astrophys. J.*]{}[**501**]{}, 780 (1998). G. Baym, E.W. Kolb, L. McLerran, T.P. Walker and R.L. Jaffe, [*Phys. Lett. B*]{}[**160**]{}, 181 (1985). A.V. Olinto, [*Phys. Lett. B*]{}[**192**]{}, 71 (1987) M.L. Olesen and J. Madsen, in [*Proceedings of the Workshop on Strange Quark Matter in Physics and Astrophysics*]{}, eds. J.Madsen e P.Haensel. [*Nuc.Phys.B Proc. Supp.*]{}[**24**]{}, 170 (1991). H. Heiselberg, G. Baym and C.J. Pethick, in [*Proceedings of the Workshop on Strange Quark Matter in Physics and Astrophysics*]{}, eds. J.Madsen e P.Haensel. [*Nuc.Phys.B Proc. Supp.*]{}[**24**]{}, 144 (1991). A. Drago, G. Pagliara and I. Parenti, [*Nuc. Phys. A*]{}[**782**]{}, 418 (2007) A. Bhattacharyya, S. K. Ghosh, P. S. Joardar, R. Mallick and S. Raha, [*Phys. Rev. C*]{}[**74**]{}, 065804 (2006) C.Alcock, E.Farhi and A.V.Olinto, [*Astrophys. J.*]{}[**310**]{}, 261 (1986). I. Bombaci, I. Parenti and I. Vidaña, in [*Superdense QCD Matter and Compact Stars*]{}, 353. D. Blaschke and D. Sedrakian (eds)(Springer, Dordretch 2006) G. Darrieus, unpublished (1938). L.D. Landau, [*Acta Physicochim. URSS*]{}[**19**]{}, 77 (1944). J.E. Horvath and O.G. Benvenuto, [*Phys. Lett. B*]{} [**213**]{}, 156 (1988). F.K. Röpke and W. Hillebrandt, in [*Cosmic Explosions, on the 10th Anniversary of SN1993J*]{} Proceedings of IAU Colloquium 192.J.M. Marcaide and K W. Weiler (eds), 333. (Springer Proceedings in Physics 99. Berlin 2005) Ya. B. Zel’dovich, [*J. Appl. Mech. Tech. Phys.*]{}[**7**]{}, 68 (1966). S.I. Blinnikov and P.V. Sasorov, [*Phys. Rev. E*]{}[**53**]{}, 4827 (1996). G. H. Markstein, [*J. Aeronaut. Sci.*]{}[**18**]{}, 199 (1951). J. Niemeyer and W. Hillebrandt, [**Astrophys. J.**]{}[**452**]{}, 769 (1995). N. Peters, in [*Proc. of 21st Int. Sym. Comb.*]{}, 1232 (Combustion Institute, Pittsburg 1988). J.C. Niemeyer and A. Kerstein, [*New Astron.*]{}[**2**]{}, 239 (1997). P. Clavin, [*Annu. Rev. Fluid. Mech.*]{}[**26**]{}, 321 (1994). S. Chandrasekhar, [*Hydrodynamic and Hydromagnetic Stability*]{} (Oxford University Press, Oxford 1961). R.M. Davies and G. Taylor, [*Proc. Roy. Soc. London*]{}[**200**]{}, 375 (1949). D.H. Sharp, [*Physica D*]{}[**12**]{}, 3 (1984) F. Timmes and S. E. Woosley, [*Astrophys. J.*]{}[**396**]{}, 649 (1992). C.R. Ghezzi, E.M. de Gouveia Dal Pino and J.E. Horvath, [*Astrophys. J. Lett.*]{}[**548**]{}, L193 (2001). Ya. B. Zel’dovich, V.B. Librovich, G.M. Makhviladze and G.I. Sivashinsky, [*Astronaut. Acta*]{}[**15**]{}, 313 (1970). A.M. Khokhlov, E.S. Oran and J.C. Wheeler, [*Astrophys. J.*]{}[**478**]{}, 678 (1997). J. C. Niemeyer and S.E. Woosley, [*Astrophys. J.*]{}[**4575**]{}, 740 (1997). M. Zingale, A.S. Almgren, J.B. Bell, M.S. Day, C.A. Rendleman and S.E. Woosley, [*Journal of Physics: Conference Series*]{}[**46**]{}, 385 (2006). I. Bombaci and B. Datta, [*Astrophys. J. Lett.*]{}[**530**]{}, L69 (2000). J.M. Lattimer and M. Prakash, [*Phys. Repts.*]{}[**333**]{}, 121 (2000). J. Madsen, [*Phys.Rev.Lett.*]{}[**61**]{}, 2909 (1988). G.A.Medina-Tanco and J.E.Horvath, [*Astrophys.J.*]{} [**464**]{}, 354 (1996). G.Lugones and O.G. Benvenuto, [*Mon. Not. R.A.S.*]{}[**304**]{}, L25 (1999). C.Alcock, E.Farhi and A.V.Olinto, [*Phys. Rev. Lett.*]{}[**57**]{}, 2088 (1986). F. Ma and B. Xie, [*Astrophys.J.Lett.*]{} [**462**]{}, L63 (1996). P. Haensel, B. Paczynski and P. Amsterdamski, [*Astrophys.J.*]{} [**375**]{}, 209 (1991). K.S. Cheng and Z.G. Dai, [*Phys. Rev. Lett.*]{}[**77**]{}, 1210 (1996). X.Y. Wang, Z.G. Dai, T. Lu, D.M. Wei and Y.F. Huang, [*Astron. Astrophys.*]{}[**357**]{}, 543 (2000). R. Ouyed and F. Sannino, [*Astron. Astrophys.*]{}[**387**]{}, 725 (2002). B. Paczynski and P. Haensel, [*MNRAS*]{} [**362**]{}, L4 (2005). Z. Berezhiani, I. Bombaci, A. Drago, F. Frontera and A. Lavagno, in [*Astronomical Society of the Pacific Conferences*]{}[**312**]{}, 407 (2004). G. Lugones, C.R. Ghezzi, E.M. de Gouveia Dal Pino and J.E. Horvath, [*Astrophys. J. Lett.*]{}[**581**]{}, L101 (2002). J.E. Horvath, in [*Proceedings of the Carpathian Summer School of Physics 2005*]{}, 188 (eds. S. Stoica, L. Trasche and R. Tribble). (Singapore: World Scientific 2006) I. Tokareva, A. Nusser, V. Gurovich and V. Folomeev, [*Int. Jour. Mod. Phys. D*]{} [**14**]{}, 33 (2005) ; I. Tokareva and A. Nusser, [*Phys.Lett. B*]{}[**639**]{}, 232 (2006). P. Haensel and J.L. Zdunik, in [*SWIFT and GRBs: Unveiling the Relativistic Universe*]{}, Venice, June 5-9, 2006. astro-ph/0701258. G.Lugones e J.E. Horvath, [*Phys. Rev. D*]{}[**66**]{}, 074017 (2002). G.Lugones e J.E. Horvath, [*Astron. Astrophys.*]{}[**403**]{}, 173 (2003). D. K. Hong, S.D.H. Hsu and F. Sannino, [*Phys. Lett. B*]{}[**516**]{}, 362 (2001). R. Ouyed, R. Raff and C. Vogt, [*Astrophys. J.*]{}[**632**]{}, 1001 (2005). A. Chen and R.Xu, astro-ph/0605285 (2006). O.G. Benvenuto and G. Lugones, [*MNRAS*]{}[**304**]{}, L25 (1999). At a given instant the regimes dominating the burning are shown as a function of the lengthscale. Below $\sim \, 100 \, l_{th}$ the laminar flame ensues. Cells appear above that scale and produce a weakly-dependent velocity (as described by fractal models, for instance). Above $l_{gib}$ cellular stabilization fails and above a transition scale the buoyancy ultimate dominates the burning ${\bf u_{RT}} \propto \, l^{1/2}$. It should be kept in mind that the distributed regime may be directly reached, disrupting the flame that no longer follows the regimes of Fig.1

ArXiv

--- abstract: 'We consider one possible implementation of Hadamard gate for optical and ion trap holonomic quantum computers. The expression for its fidelity determining the gate stability with respect to the errors in the single-mode squeezing parameter control is analytically derived. We demonstrate by means of this expression the cancellation of the squeezing control errors up to the fourth order on their magnitude.' author: - | V.I. Kuvshinov$^1$, A.V. Kuzmin$^2$\ Joint Institute for Power and Nuclear Research “Sosny”\ 220109 Krasina str., 99, Minsk, Belarus\ $1$ - e-mail: [email protected]\ $2$ - e-mails: [email protected]; [email protected]. title: Robust Hadamard gate for optical and ion trap holonomic quantum computers --- Holonomic quantum computation exploiting non-abelian geometrical phases (holonomies) was primarily proposed in Ref. [@Zanardi1999] and was further worked out in Ref. [@Pachos2000]. Various implementations of holonomic quantum computer (HQC) have been proposed recently. Namely, it was suggested to realize the HQC within quantum optics (optical HQC) [@OHQC]. Laser beams in a non-linear Kerr medium were used for this purpose. Two different sets of control devices could be used in this case. The first one consists of one and two mode displacing and squeezing devices. The second one includes SU(2) interferometers. Squeezing and displacement of the vibrational modes of the trapped ions were suggested to use for realization of HQC in Ref. [@ion]. This implementation of HQC is mathematically similar to the first embodiment of the optical HQC offered in Ref. [@OHQC]. Particularly, expressions for the adiabatic connection and holonomies are the same. Another proposed implementation of HQC was HQC with neutral atoms in cavity QED [@NeutralHQC]. The coding space was spanned by the dark states of the atom trapped in a cavity. Dynamics of the system was governed by the generalized $\Lambda$-system Hamiltonian. Mathematically similar semiconductor-based implementation of HQC was proposed in Ref. [@Semicond]. One-qubit gates were realized in the framework of the same generalized $\Lambda$-system as in Ref. [@NeutralHQC]. However its physical implementation exploits semiconductor excitons driven by sequences of laser pulses [@Semicond]. For the two-qubit gate the bi-excitonic shift was used. The generalized $\Lambda$-system with different Rabi frequencies parametrization was exploited recently for HQC implemented by Rf-SQUIDs coupled through a microwave cavity [@SQUID]. One more solid state implementation of HQC based on Stark effect was proposed in Ref. [@Stark]. Quantum computers including HQC are analog-type devices. Thus unavoidable errors in the assignment of the classical control parameters lead to an errors in quantum gates and in the case when the tolerance of quantum computation is not large enough the computation fails. This obstacle (inaccuracy) is also related to the decoherence problem [@Shor]. The effect of the errors originated from the imperfect control of classical parameters was studied for ${\bf CP}^n$ model of HQC in Ref. [@CPNEr] where the control-not and Hadamard gates were particularly considered. Approach based on the non-abelian Stokes theorem [@Simonov] was proposed in our previous Letter [@WePLA]. Namely, the general expression for fidelity valid for arbitrary implementation of HQC in the case of the single control error having the arbitrary size was derived. Simple approximate formulae was found in the small error limit. Adiabatic dynamics of a quantum system coupled to a noisy classical control field was studied in Ref. [@Gaitan]. It was demonstrated that stochastic phase shift arising in the off-diagonal elements of the system’s density matrix can cause decoherence. The investigation of the robustness of non-abelian holonomic quantum gates with respect to parametric noise due to stochastic fluctuations of the control parameters was presented in Ref. [@Robust]. In this work three stability regimes were discriminated for HQC model with logical qubits given by polarized excitonic states controlled by laser pulses. Noise cancellation effect for simple quantum systems was considered in Ref. [@Cancel]. The decoherence of HQC was discussed, for instance, in Refs. [@Decoh; @Refocus]. Berry phase in classical fluctuating magnetic field was considered in Ref. [@Palma]. Fidelity decay rates for HQC interacting with the stochastic environment were obtained recently in Ref. [@Buiv]. In this Letter we consider one-qubit gates for optical HQC and HQC on trapped ions. The mathematical model based on squeezing and displacing transformations of the qubit state is the same for both these implementations (compare [@OHQC] and [@ion]). We consider one possible implementation of Hadamard gate and analytically derive the expression for its fidelity determining the gate stability with respect to the errors in the single-mode squeezing parameter control. We demonstrate the cancellation of the control errors up to the fourth order on their magnitude. Let us briefly remind some results concerning HQC. In holonomic quantum computer non-abelian geometric phases (holonomies) are used for implementation of unitary transformations (quantum gates) in the subspace $C^N$ spanned on eigenvectors corresponding the degenerate eigenvalue of parametric isospectral family of Hamiltonians $F = \{H(\lambda) = U(\lambda) H_0 U(\lambda)^+\}_{\lambda \in M}$, where $U(\lambda)$ is unitary [@Zanardi1999]. The $\lambda$’s are the control parameters and $M$ represents the space of the control parameters. The subspace $C^N$ is called quantum code ($N$ is the dimension of the degenerate computational subspace). Quantum gates are realized when the control parameters are adiabatically driven along the loops in the control manifold $M$. The unitary operator mapping the initial state vector into the final one has the form $\bigoplus_{l=1}^R e^{i \phi_l} \Gamma_{\gamma}(A_\mu^l)$, where $l$ enumerates the energy levels of the system, $\phi_l$ is the dynamical phase, $R$ is the number of different energy levels of the system under consideration and the holonomy associated with the loop $\gamma \in M$ is given by: $$\label{Hol} \Gamma_{\gamma}(A_\mu) = \hat{P} \exp{\int_{\gamma} A_{\mu} d \lambda_{\mu} }.$$ Here $\hat{P}$ denotes the path ordering operator, $A_\mu $ is the matrix valued adiabatic connection given by the expression [@Wilczek]: $$\label{A} (A_\mu )_{mn} = <\varphi_m|U^+ \frac{\partial}{\partial \lambda_\mu} U|\varphi_n>,$$ where $|\varphi_k>, \quad k=\overline{1,N}$ are the constant basis vectors of the corresponding eigenspace $C^N$, index $\mu$ enumerates the classical control parameters of the system. Dynamical phase will be omitted bellow due to the suitable choice of the zero energy level. We shall consider the single subspace (no energy level crossings are assumed). For optical holonomic quantum computer (as well as for ion trap HQC) one-qubit unitary transformations are given as a sequence of single-mode squeezing and displacing operations $U = D(\eta) S(\nu)$. Here: $$\label{SD} \begin{array}{c} \medskip S(\nu) = \exp{\left(\nu a^+ a^+ - \overline{\nu}a a\right)}, \bigskip \\ D(\eta) = \exp{\left( \eta a^+ - \overline{\eta} a \right)} \end{array}$$ denotes single-mode squeezing and displacing operators respectively, $\nu = r_1 e^{i\theta_1}$ and $\eta = x+iy$ are corresponding complex control parameters, $a$ and $a^+$ are annihilation and creation operators. The line over the parameters denotes complex conjugate quantities. The full set of the connection as well as the field strength matrix components can be found in Refs. [@OHQC; @ion]. In this Letter we consider loops belonging to the planes $(x,r_1)|_{\theta_1 = 0}$ and $(y,r_1)|_{\theta_1 = 0}$ only. Corresponding field strength components are [@OHQC]: $$\label{F} \begin{array}{c} \medskip F_{x r_1}|_{\theta_1 = 0} = -2i\sigma_y \exp{(-2r_1)}, \bigskip \\ F_{y r_1}|_{\theta_1 = 0} = -2i\sigma_x \exp{(2 r_1)}. \end{array}$$ Here $\sigma_x$ and $\sigma_y$ are Pauli matrices. The corresponding holonomies for the loops $C_I \in (x,r_1)_{\theta_1 = 0}$ and $C_{II} \in (y,r_1)_{\theta_1 = 0}$ are given by [@OHQC]: $$\label{Gam} \begin{array}{c} \medskip \Gamma (C_I) = \exp{(-i\sigma_y \Sigma_I)}, \quad \Sigma_I = \int_{S(C_I)} dx dr_1 2 e^{-2r_1}, \bigskip \\ \Gamma (C_{II}) = \exp{(-i\sigma_x \Sigma_{II})}, \quad \Sigma_{II} = \int_{S(C_{II})} dy dr_1 2 e^{2r_1}, \end{array}$$ where $S(C_{I,II})$ are the regions in the planes $(x,r_1)|_{\theta_1 = 0}$ and $(y,r_1)|_{\theta_1 = 0}$ enclosed by the loops $C_I$ and $C_{II}$. Hadamard gate is widely used in various quantum algorithms, for example in quantum Fourier transform, for more details see [@Shor2]. It is given as follows: $$\label{Hadamard} H_0 = \frac{1}{\sqrt{2}} \left(\begin{array}{cc} 1 & 1 \\ 1 & -1 \ \end{array}\right).$$ We can obtain it by the two successive $y$ and $x$ rotations on $\pi/4$ and $\pi/2$ respectively: $$\label{iH0} -i H_0 = \Gamma(C_{II})_{\Sigma_{II}=\pi/2} \Gamma(C_I)_{\Sigma_I = \pi/4}.$$ The overall phase factor $(-i)$ is not essential for our purposes. From the experimentalist point of view it is more convenient to hold three control parameters fixed and adiabatically vary the fourth parameter. Thus we consider rectangular loops $C_I$ and $C_{II}$ with its sides to be parallel to the coordinate axes. For $C_I$ these sides are given by the lines $r_1 = 0$, $x=b_x$, $r_1 = d_x$, $x=a_x$. Here $a_x$ and $b_x$ can be chosen arbitrary and the lengths of the rectangle’s sides parallel to the $x$ axis are $l_x = b_x - a_x$. In the case of the gate considered $d_x$ is given by the following expression: $$\label{dx} d_x = - \frac{1}{2} \ln{\left(1-\frac{\pi}{4l_x}\right)}.$$ It immediately follows that $l_x > \pi/4$. In the same way we set the loop $C_{II}$ as the rectangle composed by the lines $r_1 = 0$, $y=b_y$, $r_1 = d_y$, $y=a_y$ and find that: $$\label{dy} d_y = \frac{1}{2}\ln{\left( 1+\frac{\pi}{2l_y} \right)},$$ where $l_y = b_y - a_y$. To hold zero value of the squeezing control parameter is more easy from the experimental point of view than non-zero one. One encounters problems when trying to keep non-zero squeezing parameter and simultaneously adiabatically change the displacing parameter as well as vise versa. In this Letter we restrict ourselves by the errors in the single-mode squeezing parameter control. To take into account the errors in assignment of the single-mode squeezing parameter $r_1$ we have to replace $d_x$ by $d_x + \delta r_x (x)$ and $d_y$ by $d_y + \delta r_y (y)$. In this case we obtain that parameters $\Sigma_I$ and $\Sigma_{II}$ entering into the expressions (\[Gam\]) are replaced by: $$\label{Sigma} \begin{array}{c} {\displaystyle \medskip \Sigma_I^{\prime} = \Sigma_I + e^{-2d_x} \int_{a_x}^{b_x} dx \left( 1 - e^{-2\delta r_x} \right) = \Sigma_I + \delta \Sigma_I,} \bigskip \\ {\displaystyle \Sigma_{II}^{\prime} = \Sigma_{II} + e^{2d_y} \int_{a_y}^{b_y} dy \left( e^{2\delta r_y}-1 \right) = \Sigma_{II} + \delta \Sigma_{II}.} \end{array}$$ Therefore the perturbed Hadamard gate is given by the following expression: $$\label{PertH} -i H = \Gamma(C_{II})|_{\Sigma_{II}^{\prime}} \Gamma(C_I)|_{\Sigma_I^{\prime}}.$$ Using (\[Gam\]) and (\[Sigma\]) we obtain: $$\label{LongH} \begin{array}{c} {\displaystyle \medskip -i H = - \frac{1}{\sqrt{2}} \left( \cos{\delta \Sigma_I} - \sin{\delta \Sigma_I} \right) \left( I \sin{\delta \Sigma_{II}} + i \sigma_x \cos{\delta \Sigma_{II}} \right)- } \bigskip \\ {\displaystyle - \frac{i}{\sqrt{2}}\left( \cos{\delta \Sigma_I} + \sin{\delta \Sigma_I} \right) \left( \sigma_z \cos{\delta \Sigma_{II}} - \sigma_y \sin{\delta \Sigma_{II}} \right)}, \end{array}$$ where $\sigma_x$, $\sigma_y$ and $\sigma_z$ are Pauli matrixes and $I$ is $2\times 2$ identity matrix. Fidelity of the Hadamard gate determining the gate stability with respect to the errors in the control of the single-mode squeezing parameter $r_1$ is: $$\label{fid} f_j = \sqrt{|<j|iH_0^+(-iH)|j>|^2}, \quad j=0,1.$$ Here $|0>$ and $|1>$ are the basis vectors of the qubit. Substituting expressions (\[Hadamard\]) and (\[LongH\]) into formulae (\[fid\]) we obtain: $$\label{fidel} f = |\cos{\delta \Sigma_I}|.$$ Here we see that fidelity does not depend on $j$ and $f_j \equiv f$. As well it is evident that fidelity does not depend on errors made in the $(y,r_1)_{\theta_1 = 0}$ plane. The reason is that the corresponding $x$-rotation up to the overall phase factor is just the classical not-gate. Using expressions (\[Sigma\]), (\[fidel\]) and assuming that the control error $\delta r_x (x)$ much less than unity for all $x$ and have the zero average value at the line segment $[a_x,b_x]$ we find: $$\label{fidelity} f \simeq \left|\cos{\left[<\delta r_x^2> \left(2l_x - \frac{\pi}{2}\right)\right]}\right| \simeq 1 - \left(<\delta r_x^2>\right)^2 \left[ l_x \sqrt{2} - \frac{\pi}{2\sqrt{2}} \right]^2,$$ where $$\label{r2} <\delta r_x^2> = \frac{1}{l_x} \int_{a_x}^{b_x} \delta r_x^2 (x) dx.$$ Since $l_x > \pi/4$ fidelity $f=1$ at $l_x = \pi/4$ and equal or less than unity for all over values of the parameters as it should be. As well fidelity is equal to unity for $l_x^{(n)} = \pi/4 + \pi n /(2<\delta r_x^2>)$ with integer $n > 0$. We believe that this result stems from the fact that we have restricted ourselves by the errors in the squeezing parameter control only. If we take into account errors in assignment of the over control parameters, especially fluctuations of displacing parameters $x$ and $y$ while the squeezing parameter is being adiabatically changed, fidelity of the Hadamard gate will be less than unity for all $l_x^{(n)}, \quad n>0$. However, there are reasons to believe that local maxima at these points will still remain. To clarify it is the task for the further investigations. As well we would like to note the cancellation of the squeezing control errors up to the fourth order on their magnitude that is obviously follows from the expression (\[fidelity\]) where we took into account that linear terms were absent in the cosine Taylor expansion. Thus, in this Letter we have analytically derived the expression for the fidelity determining the Hadamard gate stability with respect to the errors in the single-mode squeezing parameter control. We have demonstrated the cancellation of the control errors up to the fourth order on their magnitude under the restrictions and conditions stated above. [xxxx]{} Zanardi P., Rasetti M. [*Phys. Lett. A*]{} [**264**]{} (1999) 94. Pachos J., Zanardi P., Rasetti M. [*Phys. Rev. A*]{} [**61**]{} (2000) 010305. Pachos J., Chountasis S. [*Phys. Rev. A*]{} [**62**]{} (2000) 052318. Pachos J. [*Phys. Rev. A*]{} [**66**]{} (2002) 042318. Recati A., Calarco T., Zanardi P., Cirac J.I., Zoller P. [*Phys.Rev. A*]{} [**66**]{} (2002) 0302309. Solinas P., Zanardi P., Zanardi N., Rossi F. [*e-Print:*]{} quant-ph/0301090. Zhang P., Wang Z.D., Sun J.D., Sun C.P. [*e-print:*]{} quant-ph/0407069. Bernevig B.A., Zhang S.-C. [*e-print:*]{} quant-ph/0402165. Shor P.W. [*e-print:*]{} quant-ph/9605011. Ellinas D., Pachos J. [*e-print:*]{} quant-ph/0009043. Simonov Yu.A. [*Yad. Fiz.*]{} [**50**]{} (1989) 213; Simonov Yu.A. [*Usp. Fiz. Nauk*]{} [**166**]{} (1996) 337; Kuz’menko D.S., Simonov Yu.A., Shevchenko V.I. [*Usp. Fiz. Nauk*]{} [**174**]{} (2004) 3. Kuvshinov V.I., Kuzmin A.V. [*Phys. Lett. A*]{} [**316**]{} (2003) 391. Gaitan F. [*E-print:*]{} quant-ph/0312008. Solinas P., Zanardi P., Zanghi N. [*e-print:*]{} quant-ph/0312109. Solinas P., Zanghi N. [*e-print:*]{} quant-ph/0409020. Fuentes-Guridi I., Girelli F., Livine E. [*e-print:*]{} quant-ph/0311164. Cen L.-X., Zanardi P. [*e-print:*]{} quant-ph/0403143. De Chiara G., Palma G.M. [*e-print:*]{} quant-ph/0303155. Kuvshinov V.I., Buividovich P.V. [*e-print:*]{} hep-th/0502175. Wilczek F., Zee A. [*Phys. Rev. Lett.*]{} [**52**]{} (1984) 2111. Shor P.W. [*e-print:*]{} quant-ph/9508027.

ArXiv

--- abstract: 'We provide an analytical investigation of the entanglement dynamics for a system composed of an arbitrary number of qubits dissipating into a common environment. Specifically we consider initial states whose evolution remains confined on low dimensional subspaces of the operators space. We then find for which pairs of qubits entanglement can be generated and can persist at steady state. Finally, we determine the stationary distribution of entanglement as well as its scaling versus the total number of qubits in the system.' author: - 'Laleh Memarzadeh[^1]' - 'Stefano Mancini[^2]' title: Entanglement dynamics for qubits dissipating into a common environment --- Introduction ============ Entanglement is synonymous of quantum correlations that cannot be explained by any local classical theory. Initially this notion was relegated to foundational issues, but in the last decades it has pervaded quantum information and other areas of physics [@entrev]. Being a purely quantum phenomenon it was considered fragile under contamination of environment noise in open quantum systems. Nevertheless, in recent years it has been shown that it can persist for long time up to stationary conditions in several contexts [@clark03] (see also [@polzik10] for a striking experiment on long living entanglement). More importantly it has been shown that the environment can play a constructive role in establishing entanglement [@eassist; @braun02]. Quite remarkably it happens that even without any interaction among subsystems a common dissipative environment is able to induce entanglement [@braun02]. However, this possibility has been proved true only for systems composed by two subsystems due to technical difficulties arising when accounting for more subsystems. Here we overcome these difficulties and achieve results for an arbitrary number of subsystems by considering a limited number of initially available excitations and by exploiting a power expansion for the dissipation superoperator. This allows us to restrict the analysis to a low dimensional subspace of the operators space. We find that pairwise entanglement is not created at any time between initially excited qubits. Instead, it is created and persists at steady state for pairs of initially excited and initially not excited qubits. The same holds true for pairs of qubits initially in the ground state, but in such a case the amount of entanglement is negligibly smaller than the previous case. System Dynamics =============== Let us consider a system of $n$ qubits with associated Hilbert space $\mathcal{H}\simeq\mathbb{C}^{2\otimes n}$. Let $\{|0\ra,|1\ra\}^{\otimes n}$ be the orthonormal basis, with $|0\rangle$ (resp. $|1\rangle$) the ground (resp. excited) single qubit state. We want to study the dynamics of this system of qubits when they dissipate together into an environment at zero temperature. It will be governed by the Lindblad master equation [@qnoise]: $$\label{me} \dot\rho(t)=2\sigma \rho(t) \sigma^{\dag} -\sigma^\dag \sigma\rho(t)- \rho(t)\sigma^\dag \sigma \equiv\mathcal{D}\rho(t),$$ where :=\_[i=1]{}\^n\_i, with $\sigma_i:=|0\ra\la 1|$ for the $i$th qubit. The dissipation rate has been set equal to $1$ for the sake of simplicity. Let us describe the strategy we will put into practice to solve . The fomal solution reads $\rho(t)=e^{t\mathcal{D}}\rho(0)$. To explicitly find $\rho(t)$ let us first write the Taylor expansion: $$\label{Taylor1} \rho(t)=\rho(0)+t\mathcal{D}\rho(0)+\frac{t^2}{2!}\mathcal{D}^2\rho(0)+\frac{t^3}{3!}\mathcal{D}^3\rho(0)+\cdots .$$ Now notice that once the initial state $\rho(0)$ has been chosen, the dynamics (being purely dissipative at zero temperature) can only decrease the number of initial excitation present in $\rho(0)$. That is, repeated applications of $\mathcal{D}$ to $\rho(0)$ will leave the state within a subspace $\mathbb{H}_{\rho(0)}\subset \mathbb{H}$ of the Hilbert space $\mathbb{H}=\mathcal{H}\otimes\mathcal{H}^*$ where $\mathcal{H}^*$ stands for the dual of $\mathcal{H}$. After having identified $\mathbb{H}_{\rho(0)}$, i.e. a set of operators on $\mathcal{H}$ spanning $\mathbb{H}_{\rho(0)}$, we will write down $\rho(0)$ as linear combination of such operators with unknown time dependent coefficients. Then, by inserting this expansion into Eq. we will derive a set of linear differential equations for the unknown coefficients. In this way if the initial state contains a small number of excitation we can hope to provide an analytical solution. In fact given a number of initial excitations $e$ ($e\ll n$) the following inequality holds true: dim \_[(0)]{} ( 2\^e )\^2 ( 2\^[n]{} )\^2 = dim . \[eqdim\] Finally, notice that does not satisfy the condition for the uniqueness of a stationary solution. This condition requires that the only operators commuting with the Lindblad operator $\sigma$ must be multiples of identity [@Spohn]. Hence we should expect different steady states depending on the choice of $\rho(0)$. One initial excitation ($e=1$) ============================== In this section we assume that initially only the $k$th qubit is in the excited state while all the others are in the ground state. Let us denote this state as $|k\rangle$, then $\rho(0)=|k\ra\la k|$. By applying $\mathcal{D}$ to $\rho(0)=|k\ra\la k|$ we can easily find the following closed relations: $$\begin{aligned} \mathcal{D}|k\ra\la k|&=&2|G\ra\la G|-(|E_{\not k}\ra\la k|+|k\ra\la E_{\not k}|),\notag\\ \mathcal{D}|G\ra\la G|&=&0,\notag\\ \mathcal{D}\left(|E_{\not k}\ra\la k|+|k\ra\la E_{\not k}|\right)&=&4(n-1)|G\ra\la G|-2|E_{\not k}\ra\la E_{\not k}|\notag\\ &&-2(n-1)|k\ra\la k|\notag\\ &&-n(|E_{\not k}\ra\la k|+|k\ra\la E_{\not k}|),\notag\\ \mathcal{D}|E_{\not k}\ra\la E_{\not k}|&=&2(n-1)^2|G\ra\la G|\notag\\ &&-2(n-1)|E_{\not k}\ra\la E_{\not k}|\notag\\ &&-(n-1)(|E_{\not k}\ra\la k|+|k\ra\la E_{\not k}|),\notag\\ \label{cr1}\end{aligned}$$ where we have defined $$\begin{aligned} \label{defGE} |G\ra:=|0\ra^{\otimes n},\qquad |E_{\not k}\ra:=\sum_{i\neq k}^n|i\ra,\end{aligned}$$ ($|i\ra$ stands for a state in which the $i$th qubit is in the excited state and all the others are in the ground state). With the help of Eq. we identify the subspace \_[(0)]{}&=&span{ |GG|, |kk|, |E\_[k]{}E\_[k]{}|, .\ &&. (|E\_[k]{}k|+|kE\_[k]{}|) }. Then we can write (t)&=&a\_0(t) |GG| +a\_1(t) |kk| +a\_2(t) |E\_[k]{}E\_[k]{}|\ &&+a\_3(t) (|E\_[k]{}k|+|kE\_[k]{}|), which, upon insertion into , leads to the set of differential equations \_0&=&2a\_1+2(n-1)\^2a\_2+4(n-1)a\_3 ,\ \_1&=&-2a\_1-2(n-1)a\_3 ,\ \_2&=&-2(n-1)a\_2-2a\_3 ,\ \_3&=&-a\_1-(n-1)a\_2-n a\_3 , with initial conditions $a_0(0)=a_2(0)=a_3(0)=0, a_1(0)=1$. Solving them we arrive at the density operator $$\begin{aligned} \label{rt} \rho(t)&=&(1-f(t))^2|k\ra\la k|+f(t)(2-nf(t))|G\ra\la G|\notag\\ &&+f(t)^2|E_{\not k}\ra\la E_{\not k}|\notag\\ &&-f(t)(1-f(t))(|E_{\not k}\ra\la k|+|k\ra\la E_{\not k}|),\end{aligned}$$ where $f(t):=\frac{1}{n}(1-e^{-nt})$. To analyze the pairwise entanglement between the qubit $k$ and a generic other qubit $j$ we compute the partial trace of overall the other qubits and obtain the following density operator $$\begin{aligned} \rho_{k,j}&=&(1-f(t))^2|10\ra\la 10|+f^2(t)|01\ra\la 01|\notag\\ &&-f(t)(1-f(t))\left(|10\ra\la 01|+|01\ra\la 10|-2|00\ra\la 00|\right).\notag\\ \label{rhokj1}\end{aligned}$$ Using the concurrence [@Wootters] as a measure of entanglement of $\rho_{k,j}$ in Eq.(\[rhokj1\]) we find its time evolution as $$\label{CkjT} C_{k,j}(t)=\frac{2}{n}\left(1-e^{-nt}\right)\left(1-\frac{1-e^{-nt}}{n}\right).$$ As it is shown in Fig.\[C1Excitation\], entanglement is generated between the $k$th qubit (initially in the excited state) and an arbitrary qubit $j$ (initially in the ground state), although there is no direct interaction between them. The maximum amount of entanglement decreases with the system size $n$, but the concurrence achieves its maximum value faster when the system size is larger. ![Concurrence $C_{k,j}$ versus time $t$. From top to bottom the system size is $n=2,6,10,14$.[]{data-label="C1Excitation"}](C1Excitation) Letting $t$ going to infinity in Eq.(\[CkjT\]) we can find the behavior of stationary entanglement versus the system size $n$ as $$C_{k,j}(\infty)=\frac{2(n-1)}{n^2}.$$ This is shown in Fig.\[C1ExcitationStationary\]. ![Stationary concurrence $C_{k,j}$ versus system size $n$.[]{data-label="C1ExcitationStationary"}](C1ExcitationStationary) To study the entanglement between any two qubits initially in the ground state, we compute the reduced density matrix of $j$th and $m$th qubits ($j,m\neq k$). By referring to Eq.(\[rt\]) it is easy to show that \_[j,m]{}&=&(1-2f\^2(t))|0000|\ &&+f\^2(t)(|10+|01)((10|+01|). \[rhojm1\] The concurrence [@Wootters] of $\rho_{j,m}$ in Eq.(\[rhojm1\]) results $$\label{CmnT} C_{_{j,m}}(t)=2\left(\frac{1-e^{-nt}}{n}\right)^2.$$ Its behavior is shown in Fig.\[ConcurrenceMN1excitation\]. From Eq.(\[CmnT\]) it is easy to see that at steady state it is $C_{j,m}(\infty)=\frac{2}{n^2}$, which is negligible compared to $C_{k,j}(\infty)$ for large $n$. Therefore for large systems, we have a star graph (see Fig.\[Star\]) as steady state. ![Concurrence $C_{j,m}$ versus time $t$. From top to bottom the system size is $n=2,6,10,14$.[]{data-label="ConcurrenceMN1excitation"}](ConcurrenceMN1excitation) ![Pictorial representation of the leading stationary correlations (edges) among qubits (circles). Red (reps. blue) circles represents qubits initially in the excited (resp. ground) state.[]{data-label="Star"}](Star) Two initial excitations ($e=2$) =============================== In this section we assume that initially two qubits, say the $k$th and the $l$th, are in the excited state, while all the others are in the ground state. Let us denote this state as $|k,l\rangle$, then $\rho(0)=|k,l\ra\la k,l|$. By applying $\mathcal{D}$ to $\rho(0)=|k,l\ra\la k,l|$ we can find the following closed relations: $$\begin{aligned} \label{l} \mathcal{D}|k,l\ra\la k,l|&=&2|k+l\ra\la k+l|-4|k,l\ra\la k,l|-\Lambda,\notag\\ \mathcal{D}|k+l\ra\la k+l|&=&8|G\ra\la G|-2\Omega-4|k+l\ra\la k+l|,\notag\\ \mathcal{D}|G\ra\la G|&=&0,\notag\\ \mathcal{D}\Omega&=&2(n-2)\left( 4|G\ra\la G|-|k+l\ra\la k+l| \right) -n\Omega-4|E_{\not k\not l}\ra\la E_{\not k\not l}|,\notag\\ \mathcal{D}|E_{\not k\not l}\ra\la E_{\not k\not l}|&=&(n-2)\left(2(n-2)|G\ra\la G|-\Omega-2| E_{\not k\not l}\ra\la E_{\not k\not l}|\right),\notag\\ \mathcal{D}\Lambda&=&4(n-2)\left(|k+l\ra\la k+l|-|k,l\ra\la k,l|\right)-2|B\ra\la B|+4(\Omega-\Pi)-(n+2)\Lambda,\notag\\ \mathcal{D}| B\ra\la B|&=&2(n-2)^2|k+l\ra\la k+l|+2(n-2)(2\Omega-\Lambda)+8| E_{\not k\not l}\ra\la E_{\not k\not l} |-2n |B\ra\la B|-4\Gamma,\notag\\ \mathcal{D}\Pi&=&(n-3)\left(2\Omega-\Lambda\right)-2(n-2)\Pi-\Gamma,\notag\\ \mathcal{D}\Gamma&=&2(n-3)\left(4| E_{\not k\not l} \ra\la E_{\not k\not l} |-|B\ra\la B|\right)+2(n-2)(n-3)\Omega-8|H\ra\la H|-2(n-2)\Pi-3(n-2)\Gamma,\notag\\ \mathcal{D}|H\ra\la H|&=&(n-3)\left(2(n-3) | E_{\not k\not l}\ra\la E_{\not k\not l} |-4 |H\ra\la H|-\Gamma\right), \label{cr2}\end{aligned}$$ where, in addition to Eq., we have defined $$\begin{aligned} |k+l\ra&:=&|k\ra+|l\ra, \notag\\ | E_{\not k\not l} \ra&:=&\sum_{i\neq k,l}^n|i\ra, \notag\\ |B\ra&:=&\sum_{i\neq k}|i,k\ra+\sum_{i\neq l}|i,l\ra, \notag\\ |H\ra&:=&\sum_{h>i}|h,i\ra, \hskip 2cm h,i\neq k,l, \notag\\\end{aligned}$$ and $$\begin{aligned} \Omega&:=&| E_{\not k\not l}\ra\la k+l|+|k+l\ra\la E_{\not k\not l}|, \notag\\ \Lambda&:=&|B\ra\la k,l|+|k,l\ra\la B|, \notag\\ \Pi&:=&|H\ra\la k,l|+|k,l\ra\la H|, \notag\\ \Gamma&:=&|B\ra\la H|+|H\ra\la B|.\end{aligned}$$ With the help of Eq. we can identify the subspace \_[(0)]{}&=&span{ |GG|, |E\_[kl]{}E\_[kl]{}|, |k+lk+l|,, .\ &&. |k,lk,l|, |B B|, |HH|, ,, }.\ Then we can write $$\begin{aligned} \rho(t)&=&b_0(t) |G\ra\la G|+b_1(t) | E_{\not k\not l}\ra\la E_{\not k\not l} |+b_2(t) |k+l\ra\la k+l|\notag\\ &&+b_3(t) \Omega+b_4(t) |k,l\ra\la k,l| +b_5(t) |B\ra\la B|\notag\\ &&+b_6(t) |H\ra\la H|+b_7(t) \Lambda+b_8(t) \Pi+b_9(t) \Gamma. \label{rhoexpb}\end{aligned}$$ By inserting this expansion of $\rho(t)$ into Eq. we arrive at the following set of differential equations: \_0(t)&=&2(n-2)\^2b\_1(t)+8b\_2(t)+8(n-2)b\_3(t),\[eqb0\]\ (t)&=&M [**v**]{}(t), \[eqv\] where ${\bf{v}}=(b_1,b_2,b_3,b_4,b_5,b_6,b_7,b_8,b_9)^{\sf T}$ and $$M=\left(\begin{array}{cccccccccc} -2(n-2)&0&-4&0&8&2(n-3)^2&0&0&8(n-3)\cr 0&-4&-2(n-2)&2&2(n-2)^2&0&4(n-2)&0&0\cr -(n-2)&-2&-n&0&4(n-2)&0&4&2(n-3)&2(n-2)(n-3)\cr 0&0&0&-4&0&0&-4(n-2)&0&0\cr 0&0&0&0&-2n&0&-2&0&-2(n-3)\cr 0&0&0&0&0&-4(n-3)&0&0&-8\cr 0&0&0&-1&-2(n-2)&0&-(n+2)&-(n-3)&0\cr 0&0&0&0&0&0&-4&-2(n-2)&-2(n-2)\cr 0&0&0&0&-4&-(n-3)&0&-1&-3(n-2) \end{array}\right),$$ with the initial conditions $b_0(0)=0$, ${\bf v}(0)=(0,0,0,1,0,0,0,0,0)^{\sf T}$. It is worth remarking that Eq. has been kept out of Eq. to make the matrix $M$ non-singular. Still $M$ is not normal, hence cannot be diagonalized. Nevertheless Eq. can be solved using the Laplace transform (s)&:=&\_0\^e\^[-st]{}[**v**]{}(t)dt\ &=&(s I\_[99]{}-M)\^[-1]{}[**v**]{}(0), where $I_{9\times 9}$ denotes the $9\times 9$ identity matrix. By inverting the Laplace transform we get the coefficients $b_i,\;i=1,\ldots 9$. Then, $b_0$ straightforwardly follows by integrating Eq.. Their explicit expressions are given in Appendix A. The reduced density operator of qubits $k$ and $l$ (initially in excited states) results from Eq. as $$\begin{aligned} &&\rho_{_{k,l}}(t)=b_{_4}(t)|11\ra\la 11|\notag\\ &&+\left(b_{_2}(t)+(n-2)b_{_5}(t)\right)(|10\ra+|01\ra)(\la 10|+\la 01|)\notag\\ &&+\left(b_{_0}(t)+(n-2)\left(b_{_1}(t)+\frac{(n-3)}{2}b_{_6}(t)\right)\right)|00\ra\la 00|.\notag\\\end{aligned}$$ By computing the concurrence [@Wootters] of $\rho_{k,l}(t)$ it is possible to see that the two qubit are not entangled at any time. The reduced density operator of qubit $k$ and a generic qubit $j$ initially in the ground state derived from Eq. is given by $$\begin{aligned} \rho_{_{k,j}}(t)&=&b _{_5}(t)|11\ra\la 11|\notag\\ &+&\left(b_{_2}(t)+(n-3)b_{_5}(t)+b_{_4}(t)\right)|10\ra\la 10|\notag\\ &+&(b_{_1}(t)+b_{_5}(t)+(n-3)b_{_6}(t))|01\ra\la 01|\notag\\ &+&(b_{_3}(t)+(n-3)b_{_9}(t)(|10\ra\la 01|+|01\ra\la 10|)\notag\\ &+&\Big(b_{_0}(t)+(n-3)b_{_1}(t)+b_{_2}(t) +(n-3)b_{_5}(t) \notag\\ &&+\frac{(n-3)(n-4)}{2}b_{_6}(t)\Big) |00\ra\la 00|. \label{rhokl2}\end{aligned}$$ Then, the concurrence [@Wootters] of the state (\[rhokl2\]) results $$\begin{aligned} C_{k,j}(t)&=&-2(b_{3}(t)+(n-3)b_{9}(t)+b_{7}(t))\notag\\ &&-2\Big[b_{5}(t)(b_{0}(t)+(n-3)b_{1}(t)+b_{2}(t)\notag\\ &&+(n-3)b_{5}(t)+\frac{(n-3)(n-4)}{2}b_{6}(t)\Big]^{1/2}.\notag\\ \label{Ckj}\end{aligned}$$ Blue lines in Fig.\[CkmCmn\] show the behavior of entanglement between qubits $k$ and $j$ quantified by $C_{k,j}$ vs time. It is generated as qubits start interacting with the environment and, as it can be seen, the stationary entanglement is achieved faster as the system size increases, though the maximum value decreases. In the limit of $t\to \infty$ Eq.(\[Ckj\]) becomes C\_[k,j]{}()&=&\ &-&\^[1/2]{} ,\ which for large $n$ behaves like $\frac{2}{n}$. \]t\] ![Blue lines: concurrence $C_{k,j}$ versus time $t$. Green Lines: concurrence $C_{j,m}$ versus time $t$. From top to bottom the system size is $n=4,7,10,13$.[]{data-label="CkmCmn"}](CkmCmn "fig:") The other possible case study is the entanglement between qubits initially in ground sate, say qubits $j$ and $m$ ($j,m\neq k,l$). Their reduced density operator, derived from Eq., reads $$\begin{aligned} &&\rho_{_{j,m}}=b_{_6}(t)|11\ra\la 11|\notag\\ &&+(b_{_1}(t)+2b_{_5}(t)+(n-4)b_{_6}(t))\notag\\ &&\times(|10\ra+|01\ra) (\la 10|+\la 01|)\notag\\ &&+\Big(b_{_0}(t)+(n-4)b_{_1}(t)+2b_{_2}(t)+b_{_4}(t) \notag\\ &&+2(n-4)b_{_5}(t)+\frac{(n-4)(n-5)}{2}b_{_6}(t)\Big)|00\ra\la 00|.\notag\\ \label{rhojm2}\end{aligned}$$ The concurrence [@Wootters] of the state (\[rhojm2\]) reads C\_[j,m]{}(t)&=&2(b\_[1]{}(t)+2b\_[5]{}(t)+(n-4)b\_[6]{}(t))\ &-&2\^[1/2]{}.\ In the stationary condition $C_{j,m}$ becomes C\_[j,m]{}()&=&\ &-&\^[1/2]{} , which vanishes for large system size. Green lines in Fig.\[CkmCmn\] show that the amount of entanglement generated between qubits initially in the ground state is negligible compared to that of qubits $k$ and $j$. Therefore in the stationary state we have a bipartite graph as depicted in Fig.\[TwoColorGraph\]. ![Pictorial representation of the leading stationary correlations (edges) among qubits (circles). Red (reps. blue) circles represents qubits initially in the excited (resp. ground) state.[]{data-label="TwoColorGraph"}](TwoColorGraph) Concluding Remarks ================== In conclusion, we have studied the entanglement dynamics of a system composed of an arbitrary number of qubits dissipating into a common environment. The consideration of models like this is becoming increasingly pressing with the continuing miniaturization of information processing devices. For instance, in solid state implementations qubits may be so closely spaced that they can experience the same environment [@Hu]. By considering initial states whose evolution is confined on low dimensional subspaces of the operators space we found that pairwise entanglement is not created at any time between initially excited qubits. Instead, it is created and persists at steady state for pairs of initially excited and initially not excited qubits. The same holds true for pairs of qubits initially in the ground state, but in such a case the amount of entanglement is negligibly smaller than the previous case. This lead to the emergence of structures (subgraphs) which makes this subject appealing also from the perspective of complex networks [@Barab]. Finally, it is worth noticing that the extension of the presented analysis to states with an arbitrary number $e$ of initial excitations can be done in a computationally efficient way due to the polynomial scaling of $dim\mathbb{H}_{\rho(0)}$ vs $e$. We can arrive at such a scaling (much tighter with respect to the bound given in Eq.(\[eqdim\])), by simply observing the results in Sec.III and IV. In fact, in Sec.III the density matrix can be written, in the vector basis $\{|G\rangle, |k\rangle, |E_{\not k}\rangle\}$, as $$\begin{aligned} \rho=\left(\begin{array}{ccc} a_0 & 0 & 0 \\ 0 & a_1 & a_3\\ 0 & a_3 & a_2 \end{array}\right). \label{rhoIII}\end{aligned}$$ Furthermore, in Sec.IV the density matrix can be written, in the vector basis $\{|G\rangle, |E_{\not k, \not l}\rangle, |k+l\rangle, |k,l\rangle, |B\rangle, |H\rangle\}$, as $$\begin{aligned} \rho=\left(\begin{array}{cccccc} b_0 & 0 & 0 & 0 & 0 & 0 \\ 0 & b_1 & b_3 & 0 & 0 & 0 \\ 0 & b_3 & b_2 & 0 & 0 & 0 \\ 0 & 0 & 0 & b_4 & b_7 & b_8 \\ 0 & 0 & 0 & b_7 & b_5 & b_9 \\ 0 & 0 & 0 & b_8 & b_9 & b_6 \end{array}\right). \label{rhoIV}\end{aligned}$$ Then, given the diagonal block structure of the above matrices and , we can argue that for a number of initially excitations $e$ the density matrix will have $e+1$ blocks of dimensions $1,2,\cdots, e+1$. Each block of dimension $r$ (with $1\le r \le e+1$) has $\frac{r(r+1)}{2}$ number of real parameters. Since the total number of parameters in the matrix determines the dimension of $\mathbb{H}_{\rho(0)}$, we will have $dim \mathbb{H}_{\rho(0)}=\sum_{r=1}^{e+1}\frac{r(r+1)}{2}=(1+e)(2+e)(3+e)/6$ which is polynomial of degree 3 of $e$. Solution of Eqs. (\[eqb0\]) and (\[eqv\]) ========================================= b\_0&=&(-) ,\ b\_1&=&(1-e\^[-nt]{})\^2(1-e\^[-2(n-2)t]{}) ,\ b\_2&=&-,\ b\_3&=& ,\ b\_4&=&()\^2,\ b\_5&=&+ +-\ &&-+()\^2,\ b\_6&=&4(-2+)\^2,\ b\_7&=&4+2 +2+2\ &&+- ,\ b\_8&=&4(+.\ &&.+-2 -+) ,\ b\_9&=&-2(-2-.\ &&.++2 -+) . [99]{} R. Horodecki, P. Horodecki, M. Horodecki, K. Horodecki, Rev. Mod. Phys. **81**, 865 (2009) M. B. Plenio and S. F. Huelga, Phys. Rev. Lett. **88**, 197901 (2002); S. Clark, A. Peng, M. Gu and S. Parkins , Phys. Rev. Lett. **91**, 177901 (2003); S. Mancini and J. Wang, Eur. Phys. J. D **32**, 257 (2005); L. Hartmann, W. Dur, H.J. Briegel, Phys. Rev. A **74**, 052304 (2006); D. Angelakis, S. Bose and S. Mancini, Europhys. Lett. **85**, 20007 (2007). H. Krauter, C. A. Muschik, K. Jensen, W. Wasilewski, J. M. Petersen, J. I. Cirac, E. S. Polzik, Phys. Rev. Lett. **107**, 080503 (2011). B. Ghosh, A. S. Majumdar and N. Nayak, Phys. Rev. A **74**, 052315 (2006); L. Memarzadeh and S. Mancini, Phys. Rev. A **83**, 042329 (2011); M. Rafiee, C. Lupo, H. Mokhtari and S. Mancini Phys. Rev. A **85**, 042320 (2012): S. F. Huelga, A. Rivas and M. B. Plenio, Phys. Rev. Lett. **108**, 160402 (2012). D. Braun, Phys. Rev. Lett. **89**, 277901 (2002); F. Benatti, R. Floreanini and M. Piani, Phys. Rev. Lett. **91**, 070402 (2003). H. P. Breuer and F. Petruccione, *The Theory of Open Quantum Systems*, Oxford University Press, Oxford (2002). H. Spohn, Rev. Mod. Phys. **52**, 569 (1980). W.K. Wootters, Phys. Rev. Lett., **80**, 2245 (1998). Y. Hu, Phys. Rev. A **75**, 012314 (2007). R. Albert and A. L. Barabàsi, Rev. Mod. Phys. **74**, 47 (2002); S. Perseguers, M. Lewenstein, A. Acín and J. I. Cirac, Nat. Phys. **6**, 539 (2010). [^1]: email: [email protected] [^2]: email: [email protected]

ArXiv

--- abstract: | In numerous diluted magnetic semiconductor (DMS) systems, the competition takes place between the short-range antiferromagnetic (AF) superexchange interactions and the long-range Ruderman-Kittel-Kasuya-Yosida (RKKY) coupling mediated by the charge carriers. Such a situation strongly influences the critical temperature, the maximization of which constitutes a challenging task in DMS physics and technology. The aim of the paper is to discuss theoretically the limiting effect of AF interactions between nearest-neighbour magnetic ions on the stability of inhomogeneous ferromagnetic state in a model diluted magnetic system reflecting some crucial features of DMS. The modified molecular field-based model is constructed to account for the magnetic inhomogeneity. The behavior of the system is studied as a function of the ratio of superexchange integral to effective ferromagnetic coupling integral, including the possibility of clustering/anticlustering tendency for the magnetic ions. The ground state of the system is analysed. The critical temperature is found to change non-monotonically with the concentration of magnetic ions and decrease severely for larger concentrations. The behavior of the system significantly differs from the predictions of the usual homogeneous mean-field model. Brief comparison with selected experimental results for (Zn,Mn)Te is provided. author: - 'K. Szałowski$\,$' title: | Nearest-neighbour antiferromagnetic interaction as a limiting factor\ for critical temperature in model DMS system --- Introduction ============ Maximization of the critical temperature of DMS is of paramount importance from applicational point of view. One of the known obstacles is the presence of short-range superexchange interactions competing with the charge carrier-mediated RKKY interaction supporting the ferromagnetic state in those systems [@Dietl]. This results in significant reduction of effective magnetic moment and concentration of magnetically active impurity ions, which is observed experimentally. From the theoretical point of view, such systems are challenging due to the presence of magnetic frustration, positional disorder, and thus appearance of inhomogeneous magnetic orderings. Our aim is to construct a simple model of a disordered system with competing interactions (capturing some essential features met in DMS systems) and study its mean-field solution, focusing on the critical temperature. Model DMS system and its thermodynamics ======================================= We consider a model site-diluted magnetic system consisting of spin $S$ impurity ions, distributed over the $N$ fcc lattice sites, with atomic concentration equal to $x$. This reflects the situation in typical DMS systems, where the substitutional ions of magnetic impurities occupy the sites of fcc (or slightly distorted fcc) lattice. The Hamiltonian of the model is the following: $$\label{EqHamiltonian} \mathcal{H}=-\frac{1}{2}\sum_{i,j}^{}{J\left(R_{ij}\right)\xi_i\,\mathbf{S}_i\,\xi_j\,\mathbf{S}_j}.$$ The site dilution is introduced by the occupation number operators $\xi_{i}$, which take the value of 1 for a lattice site occupied by a magnetic impurity ion and 0 otherwise. The interaction between the spins is characterized by the exchange integral $J\left(R_{ij}\right)$. This interaction includes a long-range carrier-mediated RKKY coupling (driving the possible ferromagnetic order in DMS) and as well the antiferromagnetic (AF) interaction originating from superexchange mechanism, limited to nearest-neighbour (NN) magnetic ions [@Dietl]. In a strongly diluted magnetic system, the dominant (or at least significant) number of impurity ions either lacks NN (single magnetic ions) or is involved in pairs without NN (isolated pairs) [@Behringer]. Therefore, the single magnetic ions experience only the long-range interaction, while the ions in pairs are additionally coupled antiferromagnetically. As a result, conditions for inhomogeneity in the magnetic ordering arise in the system. In order to take this feature into account in the thermodynamic description, we develop a modified molecular-field approximation (MFA) approach, proceeding along the lines sketched in our work [@Kilanski1]. The population of all magnetic impurities is divided into single ions and NN isolated pairs. The probability that a given impurity is an isolated ion equals to $p$. Therefore, the probability that it belongs to a NN pair reads $1-p$. All the clusters of larger size are neglected in the present approach. Let us note that the lattice site occupations by magnetic impurities may not be purely random. The presence of an impurity in a given lattice site modifies the probability of occupying the NN sites, which then equals to $x\left(1+\alpha\right)$, where $\alpha$ is a Warren-Cowley parameter for NN, fulfilling some necessary inequalities [@Szalowski1]. Therefore, the probability $p$ reads in general $p=\left[1-x\left(1+\alpha\right)\right]^{12}$ for fcc lattice with 12 NN. MFA is known to work well in presence of the long-range interactions. Therefore, the density matrices describing the quantum states of single spins are assumed to take the MFA form (see e.g. [@Szalowski1]). For the spin of a single ion we use the density matrix $$\pmb{\rho}^{(1)}=\frac{1}{\mathcal{Z}^{(1)}} \,\exp\left(\frac{\lambda^{(1)}}{k_{\rm B}T} \,\mathbf{S}^z\right),\,\,\mathcal{Z}^{(1)}=\mathrm{Tr} \exp\left(\frac{\lambda^{(1)}}{k_{\rm B}T}\right),$$ depending on the molecular field parameter $\lambda^{(1)}$ and yielding the magnetization $m^{(1)}=S\mathcal{B}_{S}\left(S\frac{\lambda^{(1)}}{k_{\rm B}T}\right)$. Here, $T$ is the absolute temperature, $k_{\rm B}$ is the Boltzmann constant and $S\mathcal{B}_{S}$ is a Brillouin function for spin $S$. For the ions belonging to NN pairs we accept the density matrices $$\pmb{\rho}^{(2)}_{\pm}=\frac{1}{\mathcal{Z}^{(2)}_{\pm}} \,\exp\left(\frac{\lambda^{(2)}_{\pm}}{k_{\rm B}T} \,\mathbf{S}^z\right),\,\,\mathcal{Z}^{(2)}_{\pm}=\mathrm{Tr} \exp\left(\frac{\lambda^{(2)}_{\pm}}{k_{\rm B}T}\right)$$ with the parameters $\lambda^{(2)}_{\pm}$ and $m^{(2)}_{\pm}=S\mathcal{B}_{S}\left(S\frac{\lambda^{(2)}_{\pm}}{k_{\rm B}T}\right)$, respectively. The signs $+$ and $-$ allow to distinguish between the non-equivalent ions in a pair. The total magnetization of a pair reads $m^{(2)}=m^{(2)}_{+}+m^{(2)}_{-}$. The total magnetization of the system per lattice site is a sum of contributions from the single ions and the pairs, namely $\overline{m}=\left\langle\left\langle \mathbf{S}^z\right\rangle\right\rangle_{r}=p\,m^{(1)}+\left(1-p\right)m^{(2)}/2$. Let us denote by $\left\langle \cdots \right\rangle$ the thermodynamic average over a canonical ensemble, while the configurational average over atomic disorder is $\left\langle \cdots \right\rangle_{r}$. Calculation of the thermodynamic and configurational average of the Hamiltonian (\[EqHamiltonian\]) is performed by using the postulated density matrices. The way of calculating the two-site configurational averages over the disorder reflects the inhomogeneity in the system. For the sites $i$ and $j$ being NN, we assume $\left\langle\left\langle \xi_{i}\,\mathbf{S}_{i}\,\xi_{j}\,\mathbf{S}_{j}\right\rangle\right\rangle_{r}=x\left(1-p\right)m^{(2)}_{+}m^{(2)}_{-}/2$. For a pair of non-NN sites $\left\langle\left\langle \xi_{i}\,\mathbf{S}_{i}\,\xi_{j}\,\mathbf{S}_{j}\right\rangle\right\rangle_{r} = x^2\left[p^2 \left(m^{(1)}\right)^2\right.\newline +\left.p\left(1-p\right)\,m^{(1)}m^{(2)} /2 +\left(1-p\right)^2\,\left(m^{(2)}\right)^2/4\right]$. The internal energy of the system $U=\left\langle\left\langle \mathcal{H}\right\rangle\right\rangle_{r}$ yields: $$\begin{aligned} \label{EqEnthalpy} &\frac{U}{N} = -x^2J_{F}\Biggl[\frac{p^2}{2} \left(m^{(1)}\right)^2 +\frac{p\left(1-p\right)}{2}\,m^{(1)}m^{(2)}\Biggr.\nonumber\\ &\Biggl.+\,\frac{\left(1-p\right)^2}{8}\left(m^{(2)}\right)^2\Biggr]-\frac{1-p}{2}\,xJ_{AF}\,m^{(2)}_{+}m^{(2)}_{-}.\end{aligned}$$ The parameter $J_{F}=\sum_{k=2}^{\infty}{z_{k}\,J^{\rm RKKY}\left(r_{k}\right)}$ characterizes in a convenient way the long-range RKKY interaction, which is assumed to yield the net ferromagnetic coupling $J_{F}>0$. Here, $z_{k}$ denotes the number of sites for $k$-th coordination zone in the fcc lattice, having the radius of $r_{k}$. The entropy can be expressed as: $$\begin{aligned} \label{EqEntropy} &\frac{\mathcal{S}}{N}=x\Biggl\{ k_{\rm B}\Biggl[p\,\ln \mathcal{Z}^{(1)}+\frac{1-p}{2}\left(\ln \mathcal{Z}^{(2)}_{+}+\ln \mathcal{Z}^{(2)}_{-}\right)\Biggr]\Biggr.\nonumber\\ &-\Biggl.\frac{1}{T}\left[p\lambda^{(1)}m^{(1)}+\frac{1-p}{2}\left(\lambda^{(2)}_{+}m^{(2)}_{+}+\lambda^{(2)}_{-}m^{(2)}_{-}\right)\right]\Biggr\}.\end{aligned}$$ Knowing the above, the Helmholtz free energy can be calculated as $F=U - \mathcal{S}T$. From the variational minimization of the Helmholtz energy with respect to the three molecular field parameters, the following set of equations is obtained: $$\begin{aligned} \label{EqEquationsforLambda} \lambda^{(1)}&=\left(pm^{(1)}+\frac{1-p}{2}\,m^{(2)}\right)xJ_{F}\nonumber\\ \lambda^{(2)}_{+}&=J_{AF}m^{(2)}_{-} +\left(pm^{(1)}+\frac{1-p}{2}\,m^{(2)}\right) xJ_{F}\nonumber\\ \lambda^{(2)}_{-}&=J_{AF}m^{(2)}_{+} +\left(pm^{(1)}+\frac{1-p}{2}\,m^{(2)}\right) xJ_{F}.\end{aligned}$$ Solving the equations self-consistently allows to find the magnetizations of all the subsystems. Ground state ------------ ![ The ground-state phase diagram of the model as dependent on the concentration of magnetic impurities $x$ and the relative strength of AF coupling $J_{AF}/J_{F}$ for various values of a Warren-Cowley parameter for NN, describing the non-randomness in the distribution of impurities.[]{data-label="FigGroundState"}](fig1){width="\linewidth"} {width=".80\textwidth"} The ground-state of the system (at the temperature $T=0$) can be determined basing on the minimization of the internal energy given by Eq. (\[EqEnthalpy\]). Such a procedure shows the existence of two possible magnetic orderings: a phase with the ferromagnetic order of the single ions and the antiferromagnetic order within the pairs of ions (F-AF) and a frustrated phase with all the spins ordered ferromagnetically (F-F), depending on the relative strength of antiferromagnetic and ferromagnetic couplings $J_{AF}/J_{F}$ and the concentration of magnetic component $x$. The phase boundary follows from the comparison of the internal energies of both phases and is described by the expression $$\left(J_{AF}/J_{F}\right)_{c}=-\frac{1}{2}\,x_{c}\left\{1+\left[1-x_{c}\left(1+\alpha\right)\right]^{12}\right\}.$$ In the Fig. \[FigGroundState\] the stability ranges of both phases are illustrated for purely random distribution of the magnetic ions as well as for the clustering ($\alpha>0$) and anticlustering ($\alpha<0$) tendency. Let us note that the stability range of F-F phase may be expected to shift toward larger concentrations $x$ when the larger clusters of AF-coupled ions are included in the model. Temperature magnetization curves -------------------------------- By solving numerically the self-consistent set of equations (\[EqEquationsforLambda\]) together with the relations between magnetizations and molecular field parameters following from the choice of the density matrices, the temperature dependence of the magnetization can be found. Since the system of equations has in general more than one solution, it is necessary to select the one consistent with the ground state (discussed above). All the further calculations assume $S=5/2$. Fig. 2 presents the example calculations performed for a rather weak AF interactions, $J_{AF}/J_{F}=-0.05$. For the lower concentration (corresponding to F-AF state), an interesting feature is that the increase of the temperature causes the appearance of net non-zero magnetization $m^{(2)}$ of the ion pairs. The magnetization of the single ion system $m^{(1)}$ vanishes at certain temperature $T_{c}$, together with the net magnetization of the pairs $m^{(2)}$ and hence the total magnetization of the system $\overline{m}$. However, $m^{(2)}_{\pm}$ do not vanish itself and the spins of ions in pairs remain antiferromagnetically polarized up to the Néel temperature $T_{N}>T_{c}$. The situation is quantitatively different for the larger concentration $x$, corresponding to F-F ground state. As visible in Fig. 2, both the single ions and the pairs are ferromagnetically polarized up to the critical temperature $T_{c}$ and the magnetizations of all the subsystems vanish simultaneously. Critical temperature -------------------- For the homogeneous ferromagnetic state (i.e. for $J_{AF}=0$) we get the reference critical temperature $T_{c}=S\left(S+1\right)xJ_{F}/3k_{\rm B}$. On the other hand, in the limit of a rather strong superexchange we deal with two completely decoupled systems and then only the single magnetic ions contribute to the ferromagnetic state, leading to the critical temperature of $T_{c}=S\left(S+1\right)x\,pJ_{F}/3k_{\rm B}$. ![ The normalized critical temperature (at which the total magnetization of the system vanishes) as a function of the magnetic impurities concentration $x$, for various relative magnitudes of AF coupling $J_{AF}/J_{F}$ and for randomly distributed impurities.[]{data-label="FigCriticalTemperature"}](fig3){width="\linewidth"} The usual procedure of finding the critical temperature within MFA approach relies on the linearization of the relations between magnetization and molecular field parameters resulting from the choice of the density matrices [@Szalowski1]. It is tempting to apply this procedure here, however it should be preceded by the careful analysis of the temperature magnetization curves, since the magnetizations of the subsystems in the F-AF phase do not vanish simultaneously. As a consequence, the critical temperature $T_{c}$ should be determined numerically on the basis of the temperature magnetization curves. As for $T>T_{c}$ the magnetization $m^{(1)}$ is zero, then for finding the Néel temperature the usual linearization procedure can be utilized, leading to the result $T_{N}=S(S+1)\left|J_{AF}\right|/3k_{\rm B}$, being independent of the existence of the ferromagnetic interactions in the system. On the other hand, for F-F phase, all the components of magnetization vanish together and the linearization procedure gives $$\begin{aligned} \label{EqTc} T_{c}=&\frac{S\left(S+1\right)}{6 k_{\rm B}} \bigg[ \sqrt{x^2J_{F}^2+2xJ_{F}J_{AF}\left(1-2p\right)+J_{AF}^2}\bigg.\nonumber\\ &+\bigg. xJ_{F}+J_{AF}\bigg].\end{aligned}$$ Fig. (\[FigCriticalTemperature\]) presents the dependence of the normalized critical temperature $k_BT_c/J_F$ (at which the total magnetizaton of the system vanishes) on the atomic concentration of the magnetic component $x$. A purely random, uncorrelated distribution of the magnetic impurity ions over the lattice is assumed ($\alpha=0$). The plot is prepared for varying relative strength of AF interaction, the figure of merit being the ratio of NN antiferromagnetic exchange constant $J_{AF}$ to the ferromagnetic long-range interaction parameter $J_F$. Both the asymptotes - the homogeneous MFA result neglecting the existence of AF interaction and the limiting behavior of $T_{c}$ for strong $J_{AF}$ - are shown with the bold lines. It is visible that the presence of $J_{AF}$ always lowers the critical temperature below its homogeneous MFA value. Within the range of F-AF phase, for weak $J_{AF}$ the critical temperature increases monotonically with $x$. Increasing the strength of AF interaction causes the critical temperature to exhibit a local maximum for certain value of $x$ and then significantly drop. On the other hand, for F-F phase, $T_{c}$ changes almost perfectly linearly with $x$, however its value is reduced over a constant depending on $J_{AF}$. It is noticeable that for $J_{AF}/J_{F}<-0.059$ the critical temperature exhibits a jump with increasing $x$, between F-AF and F-F phase (as shown by the dotted vertical lines). The envelope of the maximum critical temperatures $T_{c}\left(x_{c}\right)$ achieved in the F-AF phase within this range is also shown. ![ The normalized critical temperature as a function of the magnetic impurities concentration $x$, for $J_{AF}/J_{F}=-0.05$ and for various values of the Warren-Cowley parameter for NN.[]{data-label="FigWarrenCowley"}](fig4){width="\linewidth"} The influence of the clustering or anticlustering tendency of magnetic ions is analysed for the weak $J_{AF}$ in the Fig. \[FigWarrenCowley\]. It can be observed that the effect of clustering ($\alpha>0$) limits the $T_{c}$ value and is similar to the result of increasing the strength of AF coupling. On the contrary, the tendency of anticlustering ($\alpha<0$) reduces the influence of AF interaction, and in the limiting case of $\alpha=-1$ (without NN) restores the ordinary homogeneous MFA result for $T_{c}$. This indicates the significant importance of the possible non-randomness of the magnetic ions distribution in DMS in presence of the superexchange. {width=".80\textwidth"} Comparison with experimental data for (Z,M)T ============================================ It might be instructive to present briefly some results of application of our model to a representative DMS system. For this purpose we selected the p-doped Zn$_{1-x}$Mn$_{x}$Te, a II-VI semiconductor exhibiting charge carriers-driven ferromagnetism, which has been subject to experimental and theoretical studies of Ferrand *et al.* [@Ferrand1]. In order to calculate the parameter $J_{F}$, we make use of the RKKY exchange integral in the following form: $$\begin{aligned} J^{RKKY}\left(r\right)=&\frac{m_{h}a^6 \left(N_0\beta\right)^2}{8\pi h}k_{\rm F}^4\,e^{-r/\lambda}\nonumber\\ &\times \frac{\sin \left(2 k_{\rm F}r\right)-\left(2 k_{\rm F}r\right)\cos\left(2 k_{\rm F}r\right)}{\left(2 k_{\rm F}r\right)^4}.\end{aligned}$$ After the work [@Ferrand1], we assume the interaction to be mediated dominantly by the heavy holes of effective mass $m_{h}=0.60\,m_{e}$, therefore $k_{\rm F}=\left(3\pi^2 p_{h}\right)^{1/3}$ is a wavevector for the heavy holes in the valence band. The spin-hole exchange integral $N_0\beta=-1.1$ eV and the lattice constant $a=6.1$ $\mathrm{\AA}$ are accepted. The limited mean free path of the holes in a disordered system is reflected by an exponential damping of RKKY coupling with a characteristic length $\lambda$. The dominant AF superexchange integral for the NN in (Zn,Mn)Te equals to $J_{AF}=-9$ K (after the Ref. [@Bindilatti1]). However, the same experimental results indicate that the AF interaction between Mn impurities in this DMS is not limited to NN ions. Thus, the exchange integrals $J^{AF}\left(r_{k}\right)$, known up to the distance of the fifth neighbors, are included in the calculation of $J_{F}$, so that finally $J_{F}=\sum_{k=2}^{\infty}{z_{k}\,\left[J^{\rm RKKY}\left(r_{k}\right)+J^{AF}\left(r_{k}\right)\right]}$. In the Fig. \[FigExp\] we present the comparison of our model predictions with the experimental data for two representative samples studied in the Ref. [@Ferrand1]. For both samples the Curie-Weiss temperature (close to the critical temperature $T_{c}$) was measured and the temperature $T_{F}$ characterizing only the ferromagnetic charge carriers-driven interactions was evaluated. In the model of Ferrand *et al.* is was done using the empirical characteristic temperature $T_{AF}$, determined from the experiments on the samples with sufficiently low concentrations of charge carriers. In the Fig. \[FigExp\] the temperatures $T_{c}$ and $T_{F}$ taken from [@Ferrand1] are marked by empty and full circles, respectively. The experimentally determined concentrations of Mn impurities $x$ and holes $p_{h}$ are indicated in the plots. For both samples also the effective concentration of magnetic impurities was obtained from the low-temperature saturation magnetization, what allowed us to establish the value of $p$ and thus calculate the Warren-Cowley parameter $\alpha$ necessary to reproduce $p$ for the given concentration $x$. Using the above-mentioned parameters, we reproduce the experimentally observed critical temperatures using the characteristic damping lengths $\lambda$ slightly larger than the lattice constant, as given in the plots. Moreover, the same set of parameters allows us to calculate the characteristic ferromagnetic temperature $T_{F}=xS\left(S+1\right)\sum_{k=2}^{\infty}{z_{k}\,J^{\rm RKKY}\left(r_{k}\right)}/3$ (which does not include AF superexchange interactions, to conform with the meaning of experimentally determined $T_{F}$ in the work [@Ferrand1]). The values of $T_{c}$ and $T_{F}$ were calculated numerically as a function of magnetic impurity concentration (assuming constant $\alpha$). It can be observed that for the sample with lower Mn concentration $x=0.019$ (Fig. \[FigExp\](a)), we reproduce well both the value of the critical temperature and the characteristic ferromagnetic temperature within our model. On the other hand, for the sample with higher Mn concentration $x=0.038$ (Fig. \[FigExp\](b)), the value of $T_{F}$ is slightly overestimated. Such a difference might be attributed to the influence of larger clusters of the magnetic ions, which is more pronounced for larger $x$. Especially, the clusters containing odd number of the ions contribute to the ground-state magnetization (contrary to the AF-polarized pairs of ions) and thus have an impact on the value of $p$ extracted from the experimental data. Let us emphasize that in our model we do not make use of the empirically determined characteristic temperature $T_{AF}$, but we use only the values of superexchange and RKKY exchange integrals present in the Hamiltonian (\[EqHamiltonian\]). Conclusions =========== For our model DMS system, it is noticeable that the critical temperature is rather sensitive to the presence and the relative strength of the AF interactions between NN as well as to the clustering or anticlustering of the magnetic ions. Application of the model to (Zn,Mn)Te indicates that the experimental results can be satisfactorily reproduced, especially for low concentrations of magnetic impurities, by constructing the systematic thermodynamic treatment of the system Hamiltonian. The future developments of the model may especially include taking into account the existence of the larger clusters of magnetic ions (e.g. triples), since the relative abundance of such clusters rises rapidly with the concentration of impurities. It may also be worthwhile to include the AF interactions between further neighbors in the magnetic sublattice. Acknowledgement {#acknowledgement .unnumbered} =============== Prof. T. Balcerzak is gratefully acknowledged for fruitful discussions and critical reading of the manuscript. [\[1\]]{} T. Dietl, in: Handbook of Magnetism and Advanced Magnetic Materials vol. 5, edited by H. Kronmüller and S. Parkin (J. Wiley & Sons, 2007), pp. 2774-2789. R.E. Behringer, J. Chem. Phys. **29**, 537 (1958). L. Kilanski, R. Szymczak, W. Dobrowolski, K. Sza[ł]{}owski, V.E. Slynko, and E.I. Slynko, Phys. Rev. B **82**, 094427 (2010). K. Sza[ł]{}owski, and T. Balcerzak, Phys. Rev. B **77**, 115204 (2008). D. Ferrand, J. Cibert, A. Wasiela, C. Bourgognon, S. Tatarenko, G. Fishman, T. Andrearczyk, J. Jaroszyński, S. Koleśnik, T. Dietl, B. Barbara, and D. Dufeu, Phys. Rev. B **63**, 085201 (2001). V. Bindilatti, E. ter Haar, N. F. Oliveira, Jr., Y. Shapira, and M. T. Liu, Phys. Rev. Lett. **80**, 5425 (1998).

ArXiv

--- abstract: 'A limiting factor in the ability to interpret isotope effect measurements in cuprates is the absence of sufficiently accurate data for the whole phase diagram; there is precise data for $T_{c}$, but not for the antiferromagnetic transition temperature $T_{N}$. Extreme sensitivity of $T_{N}$ to small changes in the amount of oxygen in the sample is the major problem. This problem is solved here by using the novel compound (Ca$_{0.1}$La$_{0.9}$)(Ba$_{1.65}$La$_{0.35}$)Cu$_{3}$O$_{y}$ for which there is a region where $T_{N}$ is independent of oxygen doping. Meticulous measurements of $T_{N}$ for samples with $^{16}$O and $^{18}$O find the absence of an oxygen isotope effect on $T_{N}$ with unprecedented accuracy. A possible interpretation of our finding and existing data is that isotope substitution affects the normal state charge carrier density.' author: - 'E. Amit' - 'A. Keren' - 'J. S. Lord' - 'P. King' title: A Precise Measurement of the Oxygen Isotope Effect on the Néel temperature in Cuprates --- Isotope substitution is a powerful experimental tool used to investigate complex systems. Ideally, the isotope substitution affects only one parameter, for example, a phonon frequency which is directly related to the nuclear mass. However, the strong coupling of many parameters in cuprates highly limit the ability to interpret isotope effect (IE) experiments. Eventhough the oxygen isotope substitution is known to affect the superconductivity transition temperatures $T_{c}$, it is unclear whether it primarily impacts phonons, polarons, magnons, doping or other physical properties.[@Lee] The isotope effect is usually described using the isotope coefficient $\alpha$ via the relation $${T_{q}\propto M^{-\alpha_{q}}} \label{alpha}$$ where $T_{q}$ is a phase transition temperature, $M$ is the isotope mass and $q=C, N,$ and $g$ for the superconducting (SC), antiferromagnetic (AFM) and spin glass critical temperatures, respectively. In many conventional superconductors $\alpha$ was found to be very close to 0.5.[@Maxwell; @Reynolds] The explanation of this $\alpha$ in terms of Cooper pairs glued by phonons was one of the triumphs of the BCS theory for metallic superconductors.[@BCS] In cuprates the isotope effect is much more complicated and $\alpha$ is not single valued and varies across the phase diagram. The consensus today for YBCO like compounds is that in the SC phase, close to optimal doping, the oxygen IE is $$\alpha_{C}^{od}=0.018\pm0.005, \label{alphaOD}$$ which is very small but non zero.[@Khasanov] On the SC dome $\alpha_{C}$ increases as the doping decreases.[@Pringle; @Keller] In the glassy state less data is available, but it seems that the isotope effect reverses sign and $\alpha_{g}$ becomes negative. In extremely underdoped samples, where long range AFM order prevails at low temperatures, data is scarce, controversial, and has relatively large error bars.[@Khasanov; @Zhao] The most recent measurements with Y$_{y}$Pr$_{1-y}$Ba$_{2}$Cu$_{3}$O$_{7-\delta}$ show that $\alpha_{N}=0.02(3)$ in the parent compound.[@Khasanov] There are also several theories dealing with the variation of $\alpha$ along the phase diagram,[@Zhao2; @A.; @Bill; @J.; @F.; @Baugher; @D.Fisher; @Kresin; @Serbyn] but since $\alpha_{C}^{od}$ and $\alpha_{N}$ are within an error bar of each other one cannot contrast these theories with experiments. In particular, it is impossible to tell whether the same glue that holds the spins together holds the cooper pair together, or not. Increasing the accuracy of the IE measurements of the Néel temperature will shed light on the role of isotope substitution. ![The (Ca$_{0.1}$La$_{0.9}$)(Ba$_{1.65}$La$_{0.35}$)Cu$_{3}$O$_{y}$ phase diagram [@ROfer]. The antiferromagnetic, spin glass, and superconducting phases are represented in red, green and blue, respectively. For $y<6.6$ the Néel temperature does not depend on $y$. This enables accurate measurements of the oxygen isotope effect on the Néel temperature $T_{N}$.[]{data-label="x01"}](x01.EPS){width="8cm"} As mentioned above, a major limitation on the measurements of $\alpha_{N}$ is the strong dependence of $T_{N}$ on doping. For example, in Y$_{y}$Pr$_{1-y}$Ba$_{2}$Cu$_{3}$O$_{7-\delta}$, $T_{N}$ decreases with increasing doping at a rate $\Delta T_{N}=2.5$ K per $\Delta y=0.01$. This strong temperature dependence is, of course, common to many other cuprates. As a consequence, it is very difficult to prepare two samples with exactly the same $T_{N}$ even with the same isotope. The smallest fluctuation in either $y$ or $\delta$ may lead to a huge fluctuation in $T_{N}$ regardless of the isotope effect. This is not the case for the (Ca$_{0.1}$La$_{0.9}$)(Ba$_{1.65}$La$_{0.35}$)Cu$_{3}$O$_{y}$ where $T_{N}$ is constant for oxygen density $y<6.6$, as can be seen in Fig. \[x01\].[@ROfer] For our experiments, four sintered pellets were prepared using standard techniques.[@Goldschmidt] Two of the pellets were enriched with $^{18}$O isotope and two with $^{16}$O isotope in the same procedure: The samples were placed simultaneously in two closed tubes, each with different isotope gas, and then they were heated to allow the isotope to diffuse into the sample. In order to achieve a higher percentage of gas, the enrichment was repeated several times. The $^{18}$O isotope content in the samples was determined based on measurements of gas composition being in equilibrium with the sample during the exchange. Balzers Prisma mass spectrometer was used to analyze in situ isotopic composition of the atmosphere. After the exchange process was performed, the weight increase of the sample was also determined as the light $^{16}$O isotope was exchanged with the heavy $^{18}$O. The isotope enrichment in the samples measured by both methods looked to be higher than 80%. Finally a Thermal Analysis experiment (TA) was performed for the investigated samples after all experiments described in this work were fulfilled in order to verify the isotope enrichment. The samples were heated up to 1200$^{\circ}$C in the NETSCH STA 449C Jupiter analyzer in a stream of helium. During the TA experiments, ion current signals for the $^{18}$O$_{2}$, $^{18}$O$^{16}$O and $^{16}$O$_{2}$ molecules were measured using a mass spectrometer (ThermoStar Pfeiffer Vacuum). The results are shown on Fig. \[MS\]. The isotope content deduced from these measurements (comparing peak areas for the signals of particular oxygen molecules) was larger than 70%. ![Ion current signals for the $^{18}$O$_{2}$, $^{18}$O$^{16}$O and $^{16}$O$_{2}$ molecules (red, green and blue color, respectively) obtained during heating of the sample in a stream of helium. This graph shows that the $^{18}$O isotope fraction in the samples is bigger than 70% (see text).[]{data-label="MS"}](MS){width="8cm"} The oxygen IE was measured using zero-field muon spin rotation/relaxation ($\mu$SR). We particularly used the ISIS facility, which allows low muon relaxation rate and rotation frequency measurements. This is ideal for measurements near magnetic phase transitions where the muon signal varies on a long time scale. $\mu$SR data at temperatures close to the phase transition are shown in Fig. \[MuSR raw\]. As the temperature is lowered from 383.4 K the relaxation rate increases. At $T=$378.5 K oscillations appear in the data indicating the presence of long range magnetic order. The frequency of oscillations and the relaxation rate increase as the temperature is further lowered. The formula $$P_{z}(t)=P_{m}(ae^{-\lambda_{1}t}+(1-a)e^{-\lambda_{2}t}cos(\omega t))+P_{n}e^{-\Delta t} \label{muonfit}$$ was fitted to the muon polarization, where $P_{m}$, $\lambda_{1}$, $\lambda_{2}$ and $\omega$ are the polarization, relaxation rates, and frequency of muons spin in the fraction of the samples which is magnetic, and $a$ is the weighting factor between the muons experiencing transverse field and longitudinal field. This factor is close to 2/3 and temperature-independent. $P_{n}$ and $\Delta$ are the polarization and relaxation of the spin of muons that stopped in the non-magnetic volume of the sample. The solid lines in Fig. \[MuSR raw\](a) represent the fits. [raw.EPS]{} [compare.EPS]{} The AFM order parameter is determined by the frequency $\omega=\gamma B$, where $B$ is the average magnetic field at muon site and $\gamma$ is the muon gyromagnetic ratio. This frequency can be easily extracted from the $\mu$SR data well below the transition but is very difficult to determined near the transition. A second approach is to treat the magnetic volume fraction $P_{m}$ as the order parameter.[@ROfer2; @Khasanov] $P_{m}$ can be followed more closely to $T_{N}$, although this parameter also has large error bars when $\omega$ is not well defined. We used an alternative approach similar to Ref. [@Shay]; we define an order parameter which, does not require a fit, via the relation $$\mathcal{OP}(T)\equiv\frac{<P_{inf}>-<P(T)>}{<P_{inf}>-<P(0)>} \label{muonopuse}$$ where $<P(T)>$ is the average polarization at temperature $T$, and $<P_{inf}>$ is the average polarization above the transition. The denominator normalizes $\mathcal{OP}$ to 1 at zero temperature. All three order parameters for one sample are shown in Fig. \[MuSR raw\](b). The transition temperatures determined using the different order parameters are in good agreement. The advantages of $\mathcal{OP}$ are clear: it is a model free and has very small uncertainties. In Fig. \[MuSR OIE\] we present the $\mathcal{OP}$ for two samples with $^{16}$O and one with $^{18}$O around 380 K. A wide temperature range from 50 to 410 K is shown in the inset for an experiment done on separate occasion, in which two samples of $^{18}$O and one with $^{16}$O were examined. We determine $T_{N}$ by fitting a straight line to the data in the main panel of Fig. \[MuSR OIE\] in the temperature range 378 to 382 K, for each sample, and taking the crossing with the temperature axis. We find that $T_{N}^{18}$=382.49(0.34) K and $T_{N}^{16}$=382.64(0.29) K. For 100% isotope substitution the isotope exponent is determined by $$\alpha_{N}=-\frac{M_{o}^{16}}{T_{N}^{16}}\frac{T_{N}^{18}-T_{N}^{16}}{M_{o}^{18}-M_{o}^{16}}. \label{alphaDef}$$ When taking into account the isotopic fraction in the samples we obtain: $$\alpha_{N}=0.005\pm0.011. \label{alphaN}$$ This result indicates that $\alpha_{N}<\alpha_{C}^{od}$ (see Eq. \[alphaOD\]) beyond the error bars, and is consistent with no isotope effect on $T_{N}$. [Khasanov.eps]{} One possible interpretation of these results is that magnetic excitations are not relevant for superconductivity since the isotopes affect $T_{c}$ without affecting $T_{N}$. This approach was presented, for example, by Zhao *et al.*.[@Zhao2] They found that samples enriched with $^{18}$O have longer penetration depth $\lambda$ than samples enriched with $^{16}$O with the same amount of oxygen per unit cell. $\lambda$ is related to the SC carrier density $n_{s}$ and effective mass $m^{\ast}$ by $\lambda^{-2}\propto n_{s}/m^{\ast}$ so a priory both $n_{s}$ and $m^{\ast}$ can be affected by isotope substitution.[@Uemura] They ruled out the possibility that the number of carriers concentration varies by demonstrating that the thermal expansion coefficient of samples with different isotopes are the same. The authors therefore concluded that the IE changes the mass of the cooper pairs, which could be explained by polaronic supercarriers. An alternative interpretation is that the isotopes affect the efficiency of doping, as suggested in Ref. [@Kresin]. To demonstrate this interpretation we present in Fig. \[Khasanov phase diagram\](a) the critical temperatures in Y$_{y}$Pr$_{1-y}$Ba$_{2}$Cu$_{3}$O$_{7-\delta}$ for the two different isotopes taken from Ref. [@Khasanov]. $T_{N}^{18}$ seems to be a bit higher than $T_{N}^{16}$, but $T_{c}^{18}$ is a bit lower than $T_{c}^{16}$. However, if we define an efficiency parameter $K_{i}$ which relates the number of holes $p$ to the number of oxygens in the unit cell $y$ via $p=K_{i}y$, where $i$ stands for the isotope type, we can generate a unified phase diagram. This is demonstrated in Fig. \[Khasanov phase diagram\](b) and \[Khasanov phase diagram\](c) for $T_{N}$ and $T_{c}$, respectively. In these graphs three different values of $K_{18}$ are used to generate $p$ while keeping $K_{16}=1$. When using $K_{18}=0.98$, both curves of $T_{N}$ and $T_{c}$ versus $p$ for the two different isotopes collapse to the same curve. Similar scaling of the doping axis was applied to the (Ca$_{x}$La$_{1-x}$)(Ba$_{1.75-x}$La$_{0.25+x}$)Cu$_{3}$O$_{y}$ system [@ROfer] and was explained by NQR.[@Amit] Next we discuss the IE on the stiffness in the above scenario. We assume that the effective mass of the SC charge carrier $m^{\ast}$, the critical $p$ where superconductivity starts $p_{crit}$, and where $T_c$ is optimal $p_{opt}$, are not affected by the isotope substitution as suggested by Fig. \[Khasanov phase diagram\](c) and demonstrated in Ref. [@Amit]. The stiffness can be measured by the muon transverse relaxation rate $\sigma$, and it is expected that:[@Uemura] $$\sigma^i=C(p^i-p_{crit}) \label{sigma1}$$ where $C$ is a constant. Dividing the differential of $\sigma$ from Eq. \[sigma1\] by the relaxation at optimal doping yields: $$\frac{d\sigma}{\sigma_{opt}}=\frac{dp}{p_{opt}-p_{crit}}=\frac{y(K_{16}-K_{18})}{p_{opt}-p_{crit}} \label{sigma2}$$ The expected change in the stiffness due to isotope substitution can be calculated from Eq. \[sigma2\] using $p_{opt}=1$, $p_{crit}=0.42$ (which are extracted from Fig. \[Khasanov phase diagram\](b)) and $\sigma_{opt}=3.0$ $\mu s^{-1}$.[@Khasanov2] For $y=0.8$, $K_{16}=1$ and $K_{18}=0.98$ we get $d\sigma=0.083$ $\mu s^{-1}$. This value is consistent with the experimental value of $d\sigma=0.08(1)$ $\mu s^{-1}$ reported in Ref. [@Khasanov2]. In other words a 2% difference in the doping efficiency between the two isotopes can explain both the variations in phase diagram in Fig. \[Khasanov phase diagram\](a) and the variation in the stiffness. Our experiment shows that oxygen isotope substitution does not affect the Néel temperature and therefore does not play a role in magnetic excitations. However, the isotope effect of $T_{c}$ does not necessarily imply that phonons play a role in cuprate superconductivity. We show that an isotope dependent doping efficiency can explain the variation in $T_{c}$ and in the magnetic penetration depth between samples rich in $^{16}$O or $^{18}$O. We would like to thank the ISIS pulsed muon facility at Rutherford Appleton Laboratory, UK for excellent muon beam conditions. This work was funded by the Israeli Science Foundation and the Posnansky research fund in high temperature superconductivity. [99]{} P. A. Lee, Rep. Prog. Phys. **71**, 012501 (2008). E. Maxwell, Phys. Rev. **78**, 477 (1950). C. A. Reynolds, B. Serin, W. H. Wright, and L. B. Nesbitt, Phys. Rev. **78**, 487 (1950). J. Bardeen, L. N. Cooper, and J. R. Schrieffer, Phys. Rev. **108**, 1175 (1957). D. J. Pringle, G. V. M. Williams, and J. L. Tallon, Phys. Rev. B **62**, 12527 (2000). . R. Ofer *et al.*, Phys. Rev. B **74**, 220508(R) (2006). D. Goldschmidt *et al.*, Phys. Rev. B **48**, 532 (1993). R. Ofer, A. Keren, O. Chmaissem and A. Amato, Phys. Rev. B **78**, 140508(R) (2008).

ArXiv

\[section\] \[proposition\][Theorem]{} \[proposition\][Corollary]{} \[proposition\][Lemma]{} \[proposition\][Definition]{} \[proposition\][Remark]{} [**Core models in the presence of Woodin cardinals**]{} [ by Ralf Schindler]{} [ www.logic.univie.ac.at/${}^\sim$rds]{} Introduction. ============= In this paper we show that the core model might exist even if there are Woodin cardinals in $V$. This observation is not new. Woodin [@hugh], in his proof that ${\sf AD}_{\mathbb R}$ implies the ${\sf AD}_{\mathbb R}$ hypothesis, constructed models of ${\sf ZFC}$ in which there are fully iterable extender models with Woodin cardinals which satisfy (among other things) a weak covering property. Steel [@john], in his proof that $M_n$ satisfies $V = HOD$, gave an argument which appears to be a special case of what we shall do in this paper. However, the general method for constructing the core model in the theory ${\sf ZFC}$ + “there is a measurable cardinal above $n$ Woodin cardinals” + “$M_{n+1}^\#$ does not exist” which we shall present here does seem to be new. This method might in turn admit generalizations, but we do not know how to do it. We shall indicate that our method might have applications; we shall prove that if $M$ is an ultrapower of $V$ by an extender with countably closed support then $K^M$ is an iterate of $K$ (although $K$ might not be fully iterable). We let $\Sigma$ denote the following partial function. $\Sigma({\cal T}) = b$ if and only if ${\cal T}$ is an iteration tree of limit length and either \(a) ${\rm cf}({\rm lh}({\cal T})) > \omega$ and $b$ is the unique cofinal branch through ${\cal T}$, or else \(b) $b$ is the unique cofinal branch through ${\cal T}$ such that ${\cal Q}(b,{\cal T})$ exists and is weakly iterable. We say that an iteration tree ${\cal T}$ is [*according to*]{} $\Sigma$ if and only if $[0,\lambda)_T = \Sigma({\cal T} \upharpoonright \lambda)$ for every limit ordinal $\lambda < {\rm lh}({\cal T})$. Let ${\cal M}$ be a $k$-sound premouse, and let $\delta \in {\cal M}$. We say that ${\cal M}$ is [*weakly iterable above*]{} $\delta$ if for all weak $k$-embeddings $\pi \colon {\bar {\cal M}} \rightarrow {\cal M}$ with $\delta \in {\rm ran}(\pi)$, $\Sigma$ witnesses that ${\bar {\cal M}}$ is $(k,\omega_1+1)$ iterable above $\pi^{-1}(\delta)$ (i.e., with respect to iteration trees in which all extenders used have critical point $\geq \pi^{-1}(\delta)$). If $\delta = 0$ then we omit “above $0$.” By arguments of Woodin and Neeman, the $K$ of the next section can be shown to be not “fully iterable.” We have to introduce a new kind of iterability in order to be able to formulate to which extent $K$ will be iterable after all. \[delta-iterability\] Let ${\cal M}$ be a premouse, and let ${\vec \delta} = (\delta_1 < \delta_2 < ... < \delta_n)$ be a sequence of inaccessible cardinals. We say that ${\cal M}$ is ${\vec \delta}$ iterable via $\Sigma$ provided the following holds true. Suppose that ${\cal T}$ is a putative iteration tree on ${\cal M}$ which is according to $\Sigma$. Suppose also that we can write $${\cal T} = {\cal T}_0^{ \ \frown} {\cal T}_1^{ \ \frown} ... {}^\frown {\cal T}_n$$ where the following assumptions are met. \(a) For each $k \leq n$, ${\cal T}_k$ is an iteration tree of length $\theta_k$ (possibly $\theta_k = 1$, i.e., ${\cal T}_k$ is trivial), \(b) for each $k < n$, $\theta_k$ is a successor ordinal and ${\cal T}_{k+1}$ is an iteration tree on ${\cal M}^{{\cal T}_k}_{\theta_k-1}$, the last model of ${\cal T}_k$, \(c) for each $k \leq n$, if $k > 0$ then ${\rm crit}(E_\alpha^{{\cal T}_k}) > \delta_k$ whenever $\alpha+1 < {\rm lh}({\cal T}_k)$ and if $k < n$ then ${\rm lh}(E_\alpha^{{\cal T}_k}) < \delta_{k+1}$ whenever $\alpha+1 < {\rm lh}({\cal T}_k)$, and \(c) for each $k < n$, either $\theta_k < \delta_{k+1}$ or else $\theta_k \in \{ \delta_{k+1}, \delta_{k+1}+1 \}$ and there is an unbounded $A_k \subset \delta_{k+1}$ and a non-decreasing sequence $(\Omega_i \colon i \in A_k)$ of inaccessible cardinals below $\delta_{k+1}$ converging to $\delta_{k+1}$ such that if $j \leq i \in A_k$ then $j \leq_{T_k} i$, $\pi_{0 j}^{{\cal T}_k} \ {\rm " } \ \Omega_j \subset \Omega_j$, and $\pi_{ji}^{{\cal T}_k} \upharpoonright \Omega_j = {\rm id}$. Then either ${\cal T}$ has successor length and its last model is well-founded, or else ${\cal T}$ has limit length and $\Sigma({\cal T})$ is well-defined. Even if ${\cal M} \cap {\rm OR} < \delta_n$ then it makes sense to say that ${\cal M}$ is ${\vec \delta}$ iterable. If ${\cal M} \cap {\rm OR} \geq \delta_k$ and if ${\cal T}$ is as above then it is easy to see that $\delta_1$, ..., $\delta_k$ are not moved by the relevant embeddings in ${\cal T}$. If ${\cal T}_k \upharpoonright \delta_{k+1}$ has length $\delta_{k+1}$ then there is a unique cofinal branch through ${\cal T}_k \upharpoonright \delta_{k+1}$ by Definition \[delta-iterability\] (c), and hence $\Sigma({\cal T}_k \upharpoonright \delta_{k+1})$ is then well-defined. The construction. ================= Fix $n < \omega$. Throughout this section we shall assume that $\delta_1 < \delta_2 < ... < \delta_n$ are Woodin cardinals, that $\Omega > \delta_n$ is measurable, and that $M_{n+1}^\#$ does not exist. We aim to construct $K$, “the core model of height $\Omega$.” We sometimes write $\delta_0 = 0$ and $\delta_{n+1} = \Omega$. We shall recursively construct $K||\delta_k$, $k \leq n+1$, in such a way that the following assumptions will hold inductively. ${}_{1,k}. \ $ Suppose that $k>0$. If $\kappa$ is measurable with $\delta_{k-1} < \kappa \leq \delta_k$ then $K||\kappa$ is the core model over $K||\delta_{k-1}$ of height $\kappa$. ${}_{2,k}. \ $ Suppose that $k>1$. Then $\delta_1$, ..., $\delta_{k-1}$ are Woodin cardinals in $K||\delta_k$. ${}_{3,k}. \ $ $K||\delta_k$ is ${\vec \delta}$ iterable. Now suppose that $K||\delta_k$ has already been contructed, where $k<n+1$. Suppose also that [**A**]{}${}_{1,k}$, [**A**]{}${}_{2,k}$, and [**A**]{}${}_{3,k}$ hold. We aim to construct $K||\delta_{k+1}$ in such a way that [**A**]{}${}_{1,k+1}$, [**A**]{}${}_{2,k+1}$, and [**A**]{}${}_{3,k+1}$ hold. Let us first run the $K^c(K||\delta_k)$ construction inside $V_{\delta_{k+1}}$. Let ${\cal N}_\xi$ and ${\cal M}_\xi = {\frak C}_\omega({\cal N})$ be the models of this construction. We would let the construction break down just in case that we reached some $\xi < \delta_{k+1}$ with $\rho_\omega({\cal N}_\xi) < \delta_k$. We’ll show in a minute that this is not the case. We shall prove the following statements under the assumption that $\xi \leq \delta_{k+1}$ and ${\cal N}_\xi$ is defined. ${}_{1,k,\xi}. \ $ Suppose that $\xi>0$. Then $\delta_1$, ..., $\delta_k$ are Woodin cardinals in ${\cal N}_\xi$. of [**A**]{}${}_{1,k,\xi}$. By [**A**]{}${}_{2,k}$ it suffices to show that $\delta_k$ is a Woodin cardinal in ${\cal N}_\xi$, where we assume that $k>0$. Let $f \colon \delta_k \rightarrow \delta_k$, $f \in {\cal N}_\xi$, and let $g \colon \delta_k \rightarrow \delta_k$ be defined by $g(\xi) =$ the least $V$-inaccessible strictly above $f(\xi) \cup \delta_{k-1}$. As $\delta_k$ is Woodin in $V$, there is some $\kappa < \delta_k$ and some $V$-extender $E \in V_{\delta_k}$ such that if $\pi \colon V \rightarrow_E M$ then $M$ is transitive, $V_{\pi(g)(\kappa)} \subset M$, and $\pi(K||\delta_k) \cap V_{\pi(\kappa)} = K||\delta_k \cap V_{\pi(\kappa)}$. [**A**]{}${}_{1,k}$ then gives, by standard arguments, a total extender $F \in K||\delta_k$ such that if $\sigma \colon K||\delta_k \rightarrow_F K'$ then $K'$ is transitive and $K \models V_{\sigma(f)(\kappa)} \subset K'$. $\square$ ([**A**]{}${}_{1,k,\xi}$) ${}_{2,k,\xi}. \ $ ${\cal N}_\xi$ is weakly iterable above $\delta_k$. of [**A**]{}${}_{2,k,\xi}$. Cf. [@CMIP §§1, 2, and 9]. $\square$ ([**A**]{}${}_{2,k,\xi}$) ${}_{3,k,\xi}. \ $ Suppose that there is no $\delta > \delta_k$ which is definably Woodin in ${\cal N}_\xi$. Then ${\cal N}_\xi$ is ${\vec \delta}$ iterable via $\Sigma$. of [**A**]{}${}_{3,k,\xi}$. Suppose not. Let $${\cal T} = {\cal T}_0^{ \ \frown} {\cal T}_1^{ \ \frown} ... {}^\frown {\cal T}_n$$ be a putative iteration tree on ${\cal N}_\xi$ according to $\Sigma$ and as in Definition \[delta-iterability\] such that either ${\rm lh}({\cal T})$ is a successor ordinal and ${\cal T}$ has a last ill-founded model or else ${\rm lh}({\cal T})$ is a limit ordinal and $\Sigma({\cal T})$ is undefined. (Notice that if $k<n$ then we must have that ${\cal T}_{k+1}$, ..., ${\cal T}_n$ are trivial.) Suppose for a second that ${\rm lh}({\cal T})$ is a limit ordinal. Then ${\rm lh}({\cal T}) \notin \{ \delta_1, ..., \delta_n \}$, as otherwise $\Sigma({\cal T})$ would clearly be defined by Definition \[delta-iterability\] (c). Set $\delta = \delta({\cal T})$, and let $W = K^c({\cal M}({\cal T}))$ (built of height ${\rm OR}$). $W$ is well-defined and in fact $\delta \notin \{ \delta_1 , ..., \delta_n \}$ is a Woodin cardinal in $W$ due to the fact that $\Sigma({\cal T})$ is undefined (cf. [@CMMW]). By [**A**]{}${}_{1,k,\xi}$ and by the proof of [@FSIT Theorem 11.3], we’ll also have that $\delta_1$, ..., $\delta_n$ are Woodin cardinals in $W$, and $\Omega$ will be measurable in $W$. As $M_{n+1}^\#$ does not exist we know that $W||\Omega^{++}$ can’t be weakly iterable. If ${\rm lh}({\cal T})$ is a successor ordinal then we let $\delta$ and $W$ be undefined. Now let $\Theta$ be regular and large enough, and pick $${\bar H} \stackrel{\pi}{\rightarrow} H' \stackrel{\sigma}{\rightarrow} H_\Theta$$ such that ${\bar H}$ and $H'$ are transitive, ${\rm Card}({\bar H}) = \aleph_0$, and $\{ {\cal N}_\xi , {\cal T} \} \subset {\rm ran}(\sigma \circ \pi)$. If $k>0$ then we shall also assume that ${\rm Card}(H') < \delta_k$ and ${\rm ran}(\sigma) \cap \delta_k \in \delta_k$. (If $k=0$ then we allow $\sigma = {\rm id}$.) Set ${\bar {\cal N}} = (\sigma \circ \pi)^{-1}({\cal N}_\xi)$ and ${\bar {\cal T}} = (\sigma^{-1} \circ \pi)({\cal T})$. If ${\rm lh}({\cal T})$ is a limit ordinal then we shall also assume that $W||\Omega^{++} \in {\rm ran}(\sigma \circ \pi)$, and setting ${\bar W} = (\sigma \circ \pi)^{-1}(W||\lambda)$ we may and shall assume that ${\bar W}$ is not $\omega_1+1$ iterable. We also write ${\cal N}' = \pi({\bar {\cal N}}) = \sigma^{-1}({\cal N}_\xi)$ and $\delta' = \sigma^{-1}(\delta_k)$ ($= {\rm crit}(\sigma)$ if $k>0$). Suppose that $k>0$, and let $\kappa < \delta_k$ be measurable with ${\cal N}' \cap {\rm OR} < \kappa$. Then ${\cal N}'$ is $\Sigma$ coiterable with $K||\kappa$. of Claim 1. Suppose not. Let ${\cal W}$ and ${\cal W}'$ denote the putative iteration trees arising from the coiteration of ${\cal N}'$ with $K||\kappa$. ${\cal W}$ is above $\delta'$ and ${\cal W}'$ is above $\delta_{k-1}$. We assume that ${\cal W}$ has a last ill-founded model or else that $\Sigma({\cal W})$ is not defined. By [**A**]{}${}_{2,k,\xi}$, we must have that ${\cal W}$ has limit length and $\Sigma({\cal W})$ is not defined. By [**A**]{}${}_{3,k}$, $\Sigma({\cal W}')$ is well-defined, i.e., ${\cal Q}({\cal W}')$ does exists. Let us pick $$\tau \ \colon \ H'' \rightarrow H_\Theta$$ such that $H''$ is transitive, ${\rm Card}(H'') = \aleph_0$, and $\{ {\cal N}', {\cal W}, {\cal Q}({\cal W}') \} \subset {\rm ran}(\tau)$. Set ${\bar {\cal N}}' = \tau^{-1}({\cal N}')$ and ${\bar {\cal W}} = \tau^{-1}({\cal W})$. As ${\cal N}_\xi$ is weakly iterable above $\delta_k$, we certainly have that ${\bar {\cal N}}'$ is $\omega_1+1$ iterable above $\tau^{-1}(\delta')$. In particular, ${\cal Q}({\bar {\cal W}})$ exists, and ${\cal Q}({\bar {\cal W}}) \trianglelefteq {\cal M}_{\Sigma({\bar {\cal W}})}^{\bar {\cal W}}$. However, we must have that ${\cal Q}({\bar {\cal W}}) = \tau^{-1}({\cal Q}({\bar {\cal W}'}))$. In particular, ${\cal Q}({\bar {\cal W}}) \in H''$, and therefore by standard arguments $\Sigma({\cal W})$ is well-defined after all. $\square$ (Claim 1) ${\bar {\cal N}}$ is $\omega_1+1$ iterable. of Claim 2. Well, if $k=0$ then this readily follows from [**A**]{}${}_{2,k,\xi}$. Let us now assume that $k>0$. Let $\kappa < \delta_k$ be measurable with ${\cal N}' \cap {\rm OR} < \kappa$. By Claim 1, we may let ${\cal W}$ and ${\cal W}'$ denote the iteration trees arising from the (successful) $\Sigma$ coiteration of ${\cal N}'$ with $K||\kappa$. By [**A**]{}${}_{1,k}$, we shall have that $\pi_{0 \infty}^{{\cal W}} \colon {\cal N}' \rightarrow {\cal M}^{{\cal W}'}_\infty ||\gamma$ for some $\gamma$, where $\pi_{0 \infty}^{{\cal W}} \upharpoonright \delta' = {\rm id}$. We then have that $$\pi_{0 \infty}^{{\cal W}} \ \circ \ \pi \ \colon \ {\bar {\cal N}} \rightarrow {\cal M}_\infty^{{\cal W}'} || \gamma$$ witnesses, by [**A**]{}${}_{3,k}$, that ${\bar {\cal N}}$ is $\omega_1+1$ iterable via $\Sigma$. $\square$ (Claim 2) In particular, ${\cal T}$ must have limit length. By Claim 2, ${\cal Q}({\bar {\cal T}})$ exists (and is $\omega_1+1$ iterable). Set ${\cal Q} = {\cal Q}({\bar {\cal T}}) \trianglelefteq {\cal M}_{\Sigma({\bar {\cal T}})}^{{\bar {\cal T}}}$, and set ${\bar \delta} = (\sigma \circ \pi)^{-1}(\delta)$. We have that ${\bar \delta}$ is a cutpoint in ${\bar W}$, ${\bar \delta}$ is a Woodin cardinal in ${\bar W}$, and ${\bar W}$ is $\omega_1+1$ iterable above ${\bar \delta}$. Of course, ${\cal Q} || {\bar \delta} = {\bar W}||{\bar \delta}$. Let ${\cal V}'$ and ${\cal V}$ denote the iteration trees arising from the (successful) comparison of ${\bar {\cal Q}}$ with ${\bar W}$. We shall have that $\pi_{0 \infty}^{\cal V} \colon {\bar W} \rightarrow {\cal M}^{{\cal V}'}_\infty || \gamma'$ for some $\gamma'$. By Claim 2, this embedding witnesses that ${\bar W}$ is $\omega_1+1$ iterable. Contradiction! $\square$ ([**A**]{}${}_{3,k,\xi}$) ${}_{4,k,\xi}. \ $ $\rho_\omega({\cal N}_\xi) \geq \delta_k$. of [**A**]{}${}_{3,k,\xi}$. We may trivially assume that $k>0$. Suppose that $\rho = \rho_\omega({\cal N}_\xi) < \delta_k$. Let $\Theta$ be large enough, and pick $\pi \colon {\bar H} \rightarrow H_\Theta$ such that ${\bar H}$ is transitive, ${\rm Card}({\bar H}) < \delta_k$, ${\rm ran}(\pi) \cap \delta_k \in (\delta_k \setminus (\rho+1))$, and $\{ {\cal N}_\xi , \delta_k \} \subset {\rm ran}(\pi)$. Set ${\bar {\cal N}} = \pi^{-1}({\cal N}_\xi)$ and ${\bar \delta} = \pi^{-1}(\delta_k)$. Let $\kappa < \delta_k$ be measurable such that $\kappa > {\bar {\cal N}} \cap {\rm OR}$. By [**A**]{}${}_{3,k,\xi}$ (or rather by the proof of Claim 1 in the proof of [**A**]{}${}_{3,k,\xi}$), we may now successfully coiterate ${\bar {\cal N}}$ with $K||\kappa$. [**A**]{}${}_{1,k}$ implies that any witness to $\rho = \rho_\omega({\bar {\cal N}})$ is an element of $K||\kappa$. But any such witness is also a witness to $\rho = \rho_\omega({\cal N}_\xi)$. Contradiction! $\square$ ([**A**]{}${}_{3,k,\xi}$) We have shown that ${\cal N}_{\delta_{k+1}}$ exists, is ${\vec \delta}$ iterable, and that $\delta_1$, ..., $\delta_k$ are Woodin cardinals in ${\cal N}_{\delta_{k+1}}$. Let us write $K^c(K||\delta_k)$ for ${\cal N}_{\delta_{k+1}}$. Now let $\Omega$ be a measurable cardinals with $\delta_k < \Omega \delta_{k+1}$. Using $K^c(K||\delta_k)||\Omega$ we may define $K_\Omega$ as the core model over $K||\delta_k$. We shall have the canonical embedding $\pi_\Omega \colon K_\Omega \rightarrow K^c(K||\delta_k)||\Omega$ (cf. [@maximality]). If $\Omega$ and $\Omega'$ are measurable cardinals with $\delta_k < \Omega \leq \Omega' \leq \delta_{k+1}$ then $K_\Omega \trianglelefteq K_{\Omega'}$ and $\pi_\Omega = \pi_{\Omega'} \upharpoonright K_\Omega$. Let us define $K||\delta_{k+1}$ as the “union” of the $K_\Omega$, where $\Omega$ is a measurable cardinal with $\delta_k < \Omega \delta_{k+1}$. The “union” of the maps $\pi_\Omega$ then give a canonical embedding $\pi \colon K||\delta_{k+1} \rightarrow K^c(K||\delta_k)$. It is now easy to see that [**A**]{}${}_{1,k+1}$, [**A**]{}${}_{2,k+1}$, and [**A**]{}${}_{3,k+1}$ hold true. An application. =============== Suppose that $\delta_1 < ... < \delta_n$ are Woodin cardinals, that $\Omega > \delta_n$ is measurable, and that $M_{n+1}^\#$ does not exists. Let $E \in V_\Omega$ be an extender with countably closed support. Let $\pi \colon V \rightarrow_E M$ where $M$ is transitive. Let $K$ denote the core model of height $\Omega$ constructed in the previous section. There is then an iteration tree ${\cal T}$ on $K$ as in Definition \[delta-iterability\] such that ${\cal M}_\infty^{\cal T} = \pi(K)$ and $\pi_{0 \infty}^{\cal T} = \pi \upharpoonright K$. [Proof.]{} This follows by the methods of [@iterates-of-K]. $\square$ [99]{} B. Mitchell and J. Steel, [*Fine structure and iteration trees*]{}. B. Mitchell and R. Schindler, [*$K^c$ without large cardinals in $V$*]{}. E. Schimmerling and J. Steel, [*The maximality of the core model*]{}. R. Schindler, [*Iterates of $K$*]{}. J. Steel, [*The core model iterability problem*]{}. J. Steel, [*Core models with more Woodin cardinals*]{}. J. Steel, [*private communication*]{}. W.H. Woodin, [*private communication*]{}.

ArXiv

--- abstract: 'The concept of qudit (a d-level system) cluster state is proposed by generalizing the qubit cluster state (Phys. Rev. Lett. **86**, 910 (2001)) according to the finite dimensional representations of quantum plane algebra. We demonstrate their quantum correlations and prove a theorem which guarantees the availability of the qudit cluster states in quantum computation. We explicitly construct the network to show the universality of the one-way computer based on the defined qudit cluster states and single-qudit measurement. And the corresponding protocol of implementing one-way quantum computer can be suggested with the high dimensional “Ising" model which can be found in many magnetic systems.' author: - 'D.L. Zhou' - 'B. Zeng' - 'Z. Xu' - 'C.P. Sun' title: 'Quantum computation based on d-level cluster state ' --- Introduction ============ Quantum computers can process computational tasks that are intractable with classical computers. The reason lies in the fact that quantum computing systems composed by qubits (two-level quantum systems) possess mysterious quantum coherence, such as entanglement (or quantum correlation), which has no counterpart in the classical realm [@Ni]. Recently, an important kind of entangled states, cluster states [@Bri], was introduced with remarkable property—the maximal connectedness, *i.e.*, each pair of qubits can be projected onto maximally entanglement state with certainty by single-qubit measurements on all the other qubits. More surprisingly, it was shown that the cluster states can be used to build a one-way universal quantum computer, in which all the operations can be implemented by single-qubit measuments only [@Rau]. As one of the key elements to realize such a scalable quantum computer, a celebrated theorem proved by R. Raussendorf and H. J. Briegel provides a simple criterion for the functioning of gate simulations on such quantum computer [@Rau1]. It was pointed out that the protocol of cluster state computers can be easily realized in practical physical systems since the creation of the cluster states needs only Ising-type interactions [@Bri]. In fact, it has been demonstrated that the Hamiltonian with such interactions can be easily found in the solid state lattice systems with proper spin-spin interactions [@Bri], even in the system for cold atoms in optical lattices [@Du]. Theoretically, it is natural to ask whether the concept of cluster states would have its counterpart in higher dimensional Hilbert space, *i.e.*, the qudits cases, since most of the available physical systems can not be treated as a two-level systems even in an approximate way. The answer to this question is affirmative. Using a pair of non-commutative operators $X$ and $Z$, which will be identified with the $d$ dimensional irreducible representations of Manin’s quantum plane algebra (QPA) in Sec. II [@Sun], we find that the qudit cluster states $|\phi \rangle_{\mathcal{C}}$ can be defined as a common eigenstate of the tensor product operators $$X^{\dag}_a\bigotimes\limits_{b\in {\mathtt{nbgh}}(a)}Z_{b},$$ where the lower indexes $a$ and $b$ denote qudit $a$ and qudit $b$ in the cluster, and index $b$ is taken in the neighborhood of index $a$ depending on the cluster structure. Based on this definition of qudit cluster states, we further show that we can build an one way universal quantum computer by explicitly demonstrating how to construct all single qudit unitary gates and one imprimitive two-qudit gate. We organize our paper as follows: We first briefly review the finite dimensional representations of quantum plane algebra in Sec. II, which provides the main mathematical tools in this article. As a non-trivial generalization of the qubit cluster state, the qudit cluster state is defined according to the quantum plane algebra, and its essential properties of quantum correlations are analyzed in Sec. III. In Sec.IV, the one-way quantum computer based on qudit cluster states is proposed with crucial supports from the proof of a central theorem. Similar to that for qubit case [@Rau1], this theorem guarantees the functioning of gate simulations on those computers. Its proof depends on the subtle understanding about the quantum correlations of the qudit cluster states. In Sec. V, the universality of the qudit cluster state computers is proved through the explicit construction of all the single- and one two-qudit logic gates. Finally, we give our conclusion in the last section. Finite dimensional representaions of quantum plane algebra ========================================================== In this section, we will review the mathematics for finite dimensional representations of Manin’s quantum plane algebra, which will be the main mathematical tool to describe not only qudit cluster states but also unitary transformation on the Hilbert space. The Manin’s quantum plane is defined by $$XZ=qZX, \label{qpc}$$where $q$ is a complex number. Mathematically, it can be proved that the associated algebra generated by $Z$, $X$ possesses a $d-$dimensional irreducible representation only for $q^{d}=1$ [@Sun]. In this article, we take $q\equiv q_d\equiv e^{i \frac {2\pi}{d}}$. This special case is first introduced by Weyl [@Wey] and whose completeness is first proved by Schwinger [@Sch]. Obviously, when $d=1,\,q=1$, $X$ and $Z$ can be regarded as the ordinary coordinates of $R^{2}$ plane; When $d=2,\,q=-1$, $X$ and $Z$ can be identified with the Pauli matrices $\sigma _{x}$ and $\sigma _{z}$. In this sense $Z$ and $X$ can be regarded as the so-called generalized Pauli operators" [@Bar1; @San; @Dab; @Got; @Pat; @Kni]. In fact, when $q=q_d$, $Z^{d}$ and $X^{d}$ commutate with the algebra generators, so they belong to the center of QPA. The Shur’s lemma tells us $Z^{d}$ and $X^{d}$ are constants multiples of the $d-$dimensional identity matrix, *i.e.*, $Z^{d}=zI$and $X^{d}$ $=xI$. In general we can normalize them to the identity. Since the complex field $C$ is algebraically closed, there must exist an eigen-state $% |0\rangle $, which satisfies $$Z|0\rangle=|0\rangle.$$ According to Eq. (\[qpc\]), we obtain all the eigenvalue equations for operator $Z$ $$Z|k\rangle=q_d^k|k\rangle, \, (k\in Z_d),$$ where $|k\rangle =X^{\dagger k}|0\rangle$. This also implies $$X|k\rangle =|k-1\rangle. \label{xiz}$$In the $Z$-diagonal representation, the matrices of $X$ and $Z$ are: $$Z=\left[ \begin{array}{cccccc} 1 & 0 & 0 & \cdots & 0&0 \\ 0 & q_d & 0 & \cdots & 0&0 \\ \vdots & \vdots & \vdots& \ddots & \vdots & \vdots\\ 0 & 0 & 0 & \cdots & q_d^{d-2}&0 \\ 0 & 0 & 0 & \cdots & 0& q_d^{d-1}% \end{array} \right],$$ $$X=\left[ \begin{array}{cccccc} 0 & 1 & 0 & \cdots &0& 0 \\ 0 & 0 & 1 & \cdots & 0&0 \\ \vdots & \vdots & \vdots& \ddots & \vdots & \vdots\\ 0 & 0 & 0 & \cdots &1& 0 \\ 1 & 0 & 0 & \cdots &0& 0% \end{array} \right].$$ From Eq. (\[xiz\]), we have $$X|x(0)\rangle=|x(0)\rangle,$$ where $$|x(0)\rangle=\frac{1}{\sqrt{d}}\sum\limits_{k=0}^{d-1}|k\rangle.$$ Similar to the eigenvalue equation of $Z$, we have $$X|x(j)\rangle=q_d^j|x(j)\rangle,$$ where $$|x(j)\rangle=Z^j|x(0)\rangle=\frac{1}{\sqrt{d}}\sum\limits_{k=0}^{d-1}q_d^{jk}|k\rangle.$$ Corresponding to this representation, we can also define $Z(d)$ algebra as generated by $Z$, $X$. Its all basis elements $$B=\{Z^{j}X^{k},\, (j, k\in Z_{d})\},\label{oba}$$are called unitary operator bases in Ref. [@Sch]. The general commutation relations for any two basis elements are $$X^{j}Z^{k}=q_d^{jk} Z^{k}X^{j}. \label{unibase}$$ In addition, we can replace the generators $Z$ and $X$ with two other elements in the basis. First, Let $(m,n)$ be the greatest common factor of integers $m$ and $n$. Then if $(m_1,n_1)=1,\;(m_1,n_1\in Z_d)$, we can take $$\bar{Z}=q_d^{-\frac{d-1}{2}m_1n_1}Z^{m_1}X^{n_1},$$ where the factor before $Z^m X^n$ makes $\bar{Z}$ have the same eigenvalues with $Z$. To maintain Eq. (\[qpc\]), we take $$\bar{X}=q_d^{-\frac{d-1}{2}m_2n_2}Z^{m_2}X^{n_2},$$ where $(m_2,n_2)=1,\;(m_2,n_2\in Z_d)$, and $m_1n_2-m_2n_1=1$. From another viewpoint, $\bar{Z}$ and $\bar{X}$ defines a unitary transformation $U$ $$\bar{Z}=UZU^{\dagger}, \quad \bar{X}=UXU^{\dagger}.$$ By the above definition, it is easy to check that all this kind of unitary transformations form a group. In fact, it is the so-called Clifford group, which is a useful concept in universal quantum computation. Qudit cluster states in quantum plane ===================================== To generalize the concept of qubit cluster states to qudit cases, we first restrict ourselves to one dimensional lattices for the sake of the conceptual simplicity. First, let us recall the definition of one-dimensional cluster states for $N$ qubits. For a $N$-site lattice, each qubit is attached to a site. As a novel multi-qubit entanglement state, the cluster state is written as $$|\phi \rangle_{\mathcal{C}} =\frac{1}{2^{N/2}}\bigotimes\limits_{a=1}^{N}(|0\rangle _{a}+|1\rangle _{a}(\sigma_{z})_{a+1}),\label{cqb1}$$ where $(\sigma _{i})_{a}\,(i=x,y,z)$ are the Pauli matrices assigned for site $a$ in the lattice, and $$\sigma_z|s\rangle=(-1)^s |s\rangle, \, (s\in\{0,1\}).$$ Analogy to Eq. (\[cqb1\]), it is natural to conjuncture that the qudit cluster state in one dimension as $$|\phi \rangle_{\mathcal{C}} =\frac{1}{d^{N/2}}\bigotimes\limits_{a=1}^{N}\left( \sum\limits_{k=0}^{d-1}|k\rangle _{a}Z^{k}_{a+1}\right),\label{qdcs}$$ where $$Z_a |k\rangle_a=q_d^k |k\rangle_a, \,\forall a.$$ Now we present one of our main result: The qudit cluster state in one dimension defined by Eq. (\[qdcs\]) is a common eigen-state, with eigen-values being equal to $1$, of the operators $X^{\dag}_a\bigotimes\limits_{b\in {\mathtt{nbgh}}(a)}Z_{b}$ , *i.e.*, $$X^{\dag}_a\bigotimes\limits_{b\in {\mathtt{nbgh}}(a)}Z_{b}|\phi\rangle_{\mathcal{C}} =|\phi \rangle_{\mathcal{C}}, \label{eqclu}$$ where $${\mathtt{nbgh}}(a)= \left\{ \begin{array}{ll} \{2\},& a=1,\\ \{N-1\},& a=N,\\ \{a-1,a+1\},& a\notin \{1,N\}. \end{array} \right.$$ **Proof**. To prove this theorem, we notice that qudit cluster state (\[qdcs\]) can be constructed in the following procedure. We first prepare a product state$$|+\rangle =\bigotimes\limits_{a=1}^{N}|x(0)\rangle _{a},$$then apply a unitary transformation $$S=\prod\limits_{b-a=1}S_{ab} \label{clugen}$$to the state $|+\rangle $. Here $S_{ab}$ is defined by as an intertwining operator $$S_{ab}|j\rangle_a |k\rangle_b =q_d^{jk}|j\rangle_a |k\rangle_b.$$It is easy to prove $$|\phi \rangle_{\mathcal{C}} =S|+\rangle.$$ Since $X^{\dag}_a|+\rangle =|+\rangle $, it is easy to check $$SX^{\dag}_aS^{\dag }|\phi \rangle_{\mathcal{C}} =|\phi \rangle_{\mathcal{C}}.$$ The next step is to prove $$SX^{\dag}_aS^{\dag }=X^{\dag}_a\bigotimes\limits_{b\in {\mathtt{nbgh}}(a)}Z_{b}$$To this end, for $a,b,c\in\{1,...,N\}$, we observe that $$S_{ab}X^{\dag}_aS_{ab}^{\dag }=X^{\dag}_a\otimes Z_{b},$$ $$S_{ab}X^{\dag}_bS_{ab}^{\dag }=Z_{a}\otimes X^{\dag}_b,$$ $$S_{ab}X^{\dag}_c S_{ab}^{\dag }=X^{\dag}_c,\, \forall c\notin \{ a,b\},$$ and $$S_{ab}Z_{c}S_{ab}^{\dag}=Z_{c},\, \forall c.$$ This completes the proof. $\Box$ Although the above proof is restricted to one dimensional cluster, it is convenient to generalize from one-dimensional qudit cluster to more complex clusters whether in two or three dimensional space. In fact, for a general cluster $C$ with one qudit on each site, the cluster state $|\phi \rangle_{\mathcal{C}}$ is defined by the following eigen-equations: $$X^{\dag}_a\bigotimes\limits_{b\in {\mathtt{nbgh}}(a)}Z_{b}|\phi \rangle_{\mathcal{C}} =|\phi\rangle_{\mathcal{C}}. \label{eqgclu}$$Formally, the definition of a general cluster is the same as one dimensional case. Different clusters correspond to different relations of neighbours. Now we will discuss the properties of quantum correlations in the above cluster states under single qudit measurements. For simplicity, we still restrict ourselves to one dimensional case. First, let us discuss how to describe a von Neumann measurement for single qudit. Although the generator $Z$ (or $X$) is not Hermitian, *i.e.*, the eigenvalues of $Z$ (or $X$) are not real, the non-degenerate eigenstates of $Z$ (or $X$) can still form a complete orthogonal basis of $H_{d}$. Therefore they can be used to define a von Neumann measurement. For example, when we make a measurement marked by $Z$, we mean that we can obtain different results corresponding to different eigenstates of $Z$. Next, we discuss the minimal number of single qudit measurements needed to destroy all the quantum correlations in qudit cluster states. For the one dimensional cluster state, we find all quantum entanglement will be destroyed by measuring $Z_{2a}, \,(a=1,2,\cdots, [N/2])$. Finally, we will discuss the most remarkable property - maximally connected - of qudit cluster state, *i.e.*, each pair of qudits in the cluster can be projected onto maximally entanglement state with certainty by single-qudit measurement with all the other qudits. In fact, to project arbitrary two qudits in one dimensional cluster onto maximally entanglement state, we only need to measure $X$ for the qudits between them, and $Z$ for all the other qudits. It is easy to find that two or three dimensional qudit cluster states is also maximally connected. We only need to find a one-dimensional path connecting these two qudits and measure $Z$ for all the other qudits which are not on the path, which reduces the two or three dimensional problem to one dimension. We now discuss the problem of physical implementation of our one-way quantum computer. Physically the qudit cluster state (\[qdcs\]) can be created by Hamiltonian $$H=-\hbar g\sum\limits_{(a,b)}N^{(z)}_{a}N^{(z)}_{b}, \,(g>0) \label{ham}$$ where $(a,b)$ denotes sites $a$ and $b$ are nearest neighbors in the cluster; and $N^{(z)}$ is defined as $$N^{(z)}=\sum\limits_{k=0}^{d-1}|k\rangle k\langle k|$$Then we find the intertwining operator $S$ in Eq. (\[clugen\]) has the the explicit form $$S=exp\left( -\frac {i} {\hbar} H t_{\mathcal{C}}\right),$$where the evolution time $t_{\mathcal{C}}=\frac {2\pi} {dg}$. To associate with more familiar Hamitonian in physics, let us define the spin-$\frac{(d-1)}{2}$ operator of $z$ direction $$s_{z}=N^{(z)}-\frac {d-1} {2}.$$Then we can rewrite Eq. (\[ham\]) as $$H=\frac {d-1} {2}\hbar g \sum\limits_{a}\nu_a {(s_z)}_a-\hbar g\sum\limits_{(a,b)}{(s_z)}_{a}{(s_{z})}_{b},$$where $\nu_a$ is the number of nearest neighbors for qudit $a$ in the cluster. Obviously, the interaction Hamiltonian $$H_{I}=-\hbar g\sum\limits_{(a,b)}(s_{z})_{a}(s_{z})_{b},$$which is the ferromagnetic Ising type interaction with spin-$\frac{d-1}{2}$. Those kinds of Ising model, other than the usual spin-$1/2$ Ising model, has been one of the most actively studied systems in condenced matter and statistical physics due to their rich variety of critical and multicritical phenomena. For example, the spin - $1$ Ising model with nearest- neighbor interactions and a single-ion potential is know as the Blume-Emery-Griffiths (BEG) model [@Blu], the spin - $3/2$ Ising model was introduced to explain phase transitions in $DyVO4$ and its phase diagrams were obtained within the mean-field approximation [@Siv]. In addition, higher spin Ising models can be associated with the magnetic properties of artificially fabricated superlattices. Such lattices consist of two or more ferromagnetic materials have been widely studied over the years, because their physical properties differ dramatically from simple solids formed from the same materials. The development of film deposition techniques has aroused great interest in the synthesis and study of superlattices in other materials. A number of experimental[@Kwo; @Maj; @Kre; @Cam1] and theoretical works[@Izm; @Mat; @Hin; @Qu; @Fis; @Gri1; @Sy; @Sab1] have been devoted to those directions. Measurement based quantum computation with qudit ================================================ As discussed above, the qudit cluster states exhibit the same features in quantum entanglement as that for the qubit cluster states. Then a question arises naturally: Can these natures of quantum correlations be available for constructing the universal quantum computations? We will give an affirmative answer to this question in the following two sections. In this section, we further generalize the basic concept of “single qudit quantum measurement" and the corresponding measurement based quantum computation (MBQC) on qubit clusters to that on qudit clusters. Along the line to construct MBQC for the qubit case, we will formulate the corresponding theorem which relates unitary transformation to quantum entanglement exhibited by the qudit cluster states. Quantum computations with qudit clusters inherit all basic concepts of those with qubit clusters. They include the basic procedure of simulation of any unitary gate, the concatenation of gate simulation and the method to deal with the random measurement results. Here we will give a $d$-dimensional parallel theorem as generalization of the central theorem $1$ in Ref. [@Rau1]. Before formulating our central theorem, let us introduce the basic elements for quantum computing with qudit clusters. The main task of quantum computing with qudits is to simulate arbitrary quantum gate $g$ defined on $% n $ qudits Hilbert space. For this purpose, the first step for quantum computing on qudit clusters is to find out a proper cluster $\mathcal{C}(g)$. Then we divide it into three sub-clusters: the input cluster $\mathcal{C}% _{I}(g)$, the body cluster $\mathcal{C}_{M}(g)$, and the output cluster $% \mathcal{C}_{O}(g)$. As usual we require that the input and output clusters have the same rank (i.e. the same number of qudits), $|\mathcal{C}_{I}(g)|=|\mathcal{C}_{O}(g)|=n$. Then we prepare the initial state in $$|\Psi(\mathtt{in})\rangle_{\mathcal{C}(g)} =|\psi(\mathtt{in})\rangle_{\mathcal{C} _{I(g)}} |+\rangle _{\mathcal{C }_{M(g)}\cup \mathcal{C}_{O(g)}}. \label{Psi}$$ In this first step of quantum computing, we entangle the qudits on the qudit cluster by using the cluster state generator $S$, *i.e.*, $$|\Phi(\mathtt{in})\rangle_{\mathcal{C}(g)} =S|\Psi(\mathtt{in})\rangle_{\mathcal{C}(g)}. \label{Phi}$$ This step brings the structure information of the qudit cluster into our computing process, and thus relates it with the corresponding qudit cluster state. The second step is to measure all qudits on the cluster in special space-time dependent basis according to a given measurement pattern (MP). The definition of MP is given as follows. A measurement pattern ${\mathcal{M}}_{{\mathcal{C}}}$ on a cluster ${\mathcal{C}}$ is a set of unitary matrixes $${\mathcal{M}}_{{\mathcal{C}}}=\left\{ u_{a}Z_{a}u_{a}^{\dagger }\,|\,\,a\in {\mathcal{C}},\,u_a\in SU(d)\right\} ,$$which determines the one-qudit measured operators $N^{(u)}_{a}$ on ${\mathcal{C}}$, with the explicit form $$N^{(u)}_{a}=\sum_{s=0}^{d-1}u_{a}|s\rangle_{a}s\,{_{a}\;\!\!}\langle s|{u_{a}}^{\dagger}.$$ If this measurement pattern $\mathcal{M}_{\mathcal{C}}$ operates on the initial state $|\Phi(\mathtt{in})\rangle_{\mathcal{C}(g)}$, the set of measurement outcomes $$\{{s}\}_{\mathcal{C}}=\left\{ {s}_{a}\in Z_d\,|\,\,a\in \mathcal{C}\right\}$$is obtained. Then, modulo norm factor, the resulting state $|\Psi _{\mathcal{% M}}\rangle _{\mathcal{C}}$ is given by $$|{\Phi}_{\mathcal{M}_{\mathcal{C}}}^{\{s\}}\rangle =P^{\{s\}}_{\mathcal{M}_{\mathcal{C}}}\, |\Phi(\mathtt{in})\rangle_{\mathcal{C}(g)},$$ where the pure state projection $$P^{\{{s}\}}_{\mathcal{M}_\mathcal{C}}=\bigotimes_{k\in \mathcal{C}% }u_{k}|s\rangle_{k}\, {_{k}\;\!\!}\langle s|u^{\dagger }_{k}.$$It is worthy pointing out that we always measure $X$ for the input qudits and $Z$ for the output, which is independent of gate $g$, *i.e.*, $$\begin{aligned} \mathcal{M}_{\mathcal{C}_I(g)}&=&\{X_i, \,i\in \mathcal{C}_I(g)\},\\ \mathcal{M}_{\mathcal{C}_O(g)}&=&\{Z_i, \,i\in \mathcal{C}_O(g)\}.\end{aligned}$$ Then we reach the final step, to associate the measurement values with the result of gate $g$ acting on the initial state. From the above standard procedure of qudit clusters quantum computation, we learn that what is crucial for this scheme is to associate a given gate $g$ with a measurement pattern. Although by now we have no general optimal operational procedure to do this for practical problem, the following theorem provides a useful tool in realizing specific gates on the qudit clusters. \[coruni\] Suppose that the state $|\psi \rangle _{{\mathcal{C}}(g)}=P^{\{{s}\}}_{{\mathcal{M}}_{ \mathcal{C}_{M}}(g)}\,|\phi \rangle _{{\mathcal{C}}(g)}$ obeys the $2n$ eigenvalue equations $$\begin{aligned} X_{{\mathcal{C}}_{I}(g),i}\left( UX_{i}U^{\dagger }\right)_{{\mathcal{C }}_{O}(g)}|\psi \rangle _{{\mathcal{C}}(g)} & = & q_d^{-\lambda _{x,i}}|\psi \rangle _{{\mathcal{C}}(g)} \label{check1}\\ {Z^{\dagger }}_{{\mathcal{C}}_{I}(g),i}\left( UZ_{i}U^{\dagger }\right)_{{\mathcal{C }}_{O}(g)} |\psi \rangle _{{\mathcal{C}}(g)} & = & q_d^{-\lambda _{z,i}}|\psi \rangle _{{\mathcal{C}}(g)},\label{check2}%\end{aligned}$$with $\lambda _{x,i},\lambda _{z,i}\in Z_{d}$ and $1\leq i\leq n$. Then, according to the above standard quantum computing procedure, we have $$P^{{\{s\}}}_{{\mathcal{M}}_{\mathcal{C}_{I}(g)}}P^{{\{s\}}}_{\mathcal{M}_{{\mathcal{C}} _{M}(g)}}\,|\Phi(\mathtt{in})\rangle_{{\mathcal{C}}(g)} \propto(\prod |s_{i}\rangle )_{{\mathcal{C}}_{I}(g)\cup {\mathcal{C}} _{M}(g)}{|\psi ({\mathtt{out}})\rangle }_{{\mathcal{C}}_{O}(g)},\label{out}$$ where the input and output state in the simulation of $g$ are related via $$|\psi ({\mathtt{out}})\rangle =UU_{\Sigma }\,|\psi ({\mathtt{in}})\rangle ,$$ where $U_{\Sigma }$ is a byproduct operator given by $$U_{\Sigma }=\bigotimes_{({\mathcal{C}}_{O}(g)\ni i)=1}^{n}(Z_{i})^{-\lambda _{x,i}-{s}_{i}}(X_{i})^{\lambda _{z,i}}.$$ **Proof**. Let us begin with the case when $$|\psi(\mathtt{in})\rangle_{\mathcal{C}_I(g)}=|\{t\}\rangle_{\mathcal{C}_I(g)},$$ with $$\{t\}=\{t_1 t_2 \cdots t_n\}.$$ To associate with $|\psi\rangle_ {{\mathcal{C}}(g)}$, the initial input state is written as $$|\psi(\mathtt{in})\rangle_{\mathcal{C}_I(g)}= (\sqrt{d})^n P^{\{t\}}_{\mathcal{M}^{\prime}_{\mathcal{C}_I(g)}}|+\rangle_{\mathcal{C}_I(g)},\label{psi}$$ where $$\mathcal{M}^{\prime}_{\mathcal{C}_I(g)}=\{Z_i,\, i\in \mathcal{C}_I(g)\}.$$ In terms of Eqs. (\[Psi\],\[Phi\],\[out\],\[psi\]), we have $$(\prod |s_{i}\rangle )_{{\mathcal{C}}_{I}(g)\cup {\mathcal{C}} _{M}(g)}{|\psi ({\mathtt{out}})\rangle }_{{\mathcal{C}}_{O}(g)}\propto P^{\{s\}}_{\mathcal{M}_{\mathcal{C}_I(g)}} P^{\{t\}}_{\mathcal{M}^{\prime}_{\mathcal{C}_I(g)}}|\psi\rangle_ {{\mathcal{C}}(g)}.$$ To find out the equations for the final state ${|\psi ({\mathtt{out}})\rangle}_{{\mathcal{C}}_{O}(g)}$, we take $P^{\{s\}}_{\mathcal{M}_{\mathcal{C}_I(g)}} P^{\{t\}}_{\mathcal{M}^{\prime}_{\mathcal{C}_I(g)}}$ acting on both sides of Eq. (\[check1\]) and Eq. (\[check2\]): $$\left( UX_{i}U^{\dagger }\right) _{{\mathcal{C}}_{O}(g)}{|\bar{\psi} ({\mathtt{out}})\rangle}_{{\mathcal{C}}_{O}(g)} = q_d^{-s_i-\lambda _{x,i}}{|\psi ({\mathtt{out}})\rangle}_{{\mathcal{C}}_{O}(g)},\label{a1'}$$ $$\left( UZ_{i}U^{\dagger }\right) _{{\mathcal{C}}_{O}(g)}{|\psi ({\mathtt{out}})\rangle}_{{\mathcal{C}}_{O}(g)} = q_d^{t_i-\lambda _{z,i}}{|\psi ({\mathtt{out}})\rangle}_{{\mathcal{C}}_{O}(g)},\label{a1}$$ where the input state for $|\bar{\psi} ({\mathtt{out}})\rangle$ $$|\bar{\psi} ({\mathtt{in}})\rangle=X_i^{\dagger}|\{t\}\rangle.$$ Before drawing a conclusion, we need to check the final state is not a zero vector. In fact, from Eqs. (\[check1\]) and (\[check2\]), the state $U^{\dagger}|\psi \rangle _{{\mathcal{C}}(g)}$ is the simutaneous eigen-state of operators $X_{{\mathcal{C}}_{I}(g),i}X_{{\mathcal{C}}_{O}(g),i}$ and ${Z^\dagger}_{{\mathcal{C}}_{I}(g),i}Z_{{\mathcal{C}}_{O}(g),i}$. We can directly evaluate this state and find out that it has every components in $Z$-diagonal representation of the input part. Consequently the final state is indeed a nonzero vector. So from Eq. (\[a1’\]) and Eq. (\[a1\]), we obtain $$|\psi(\mathtt{out})\rangle_{{\mathcal{C}}_{O}(g)} =e^{i \eta(t)} UU_{\Sigma}\,|\{t\} \rangle_{{\mathcal{C}}_{O}(g)}. \label{az}$$ To further determine the relation between the output state and the input state, let us consider the other case when $$|\psi^\prime (\mathtt{in})\rangle_{{\mathcal C}_I(g)}=|+\rangle_{\mathcal{C}_I(g)}.$$ At this time, the final state $$|\{s_i\}\rangle_{{{\mathcal{C}}_{I}(g)}\cup {{\mathcal{C}}_{M}(g)}}\, |\psi^{\prime} (\mathtt{out})\rangle_{{\mathcal C}_O(g)}= P^{{\{s\}}}_{{\mathcal{C}}_{I}(g)}({X})|\psi\rangle_ {{\mathcal{C}}(g)}.$$ Let $P^{{\{s\}}}_{\mathcal{M}_{{\mathcal{C}}_{I}(g)}}$ apply on both sides of Eq. (\[check1\]) and Eq. (\[check2\]), we obtain $$\left( UX_iU^{\dagger }\right)_{{\mathcal{C }}_{O}(g)}|\psi^{\prime} (\mathtt{out})\rangle_{{\mathcal C}_O(g)}= q_d^{-s_i-\lambda _{x,i}}|\psi^{\prime} (\mathtt{out})\rangle_{{\mathcal C}_O(g)},$$ $$\left( UZ_iU^{\dagger }\right)_{{\mathcal{C }}_{O}(g)}|\bar{\psi^{\prime}} (\mathtt{out})\rangle_{{\mathcal C}_O(g)}= q_d^{-\lambda _{z,i}}|\psi^{\prime} (\mathtt{out})\rangle_{{\mathcal C}_O(g)},$$ where the input state for $|\bar{\psi^{\prime}} (\mathtt{out})\rangle$ $$|\bar{\psi^{\prime}} (\mathtt{in})\rangle=Z^{\dagger}_i|+\rangle.$$ From the above equations, we obtain $$|\psi^{\prime} (\mathtt{out})\rangle_{{\mathcal C}_O(g)} =e^{i \chi} UU_{\Sigma}\,|+ \rangle_{{\mathcal{C}}_{O}(g)}. \label{ax}$$ Substitute Eq. (\[az\]) into Eq. (\[ax\]), $$|\psi^{\prime} (\mathtt{out})\rangle_{{\mathcal C}_O(g)} =UU_{\Sigma} d^{-\frac n 2}\sum_{\{t\}}e^{i \eta(t)} \,|\{t\} \rangle_ {{\mathcal{C}}_{O}(g)}. \label{axz}$$ Comparing Eq. (\[ax\]) to Eq. (\[axz\]), we obtain $$e^{i \eta(t)}=e^{i \chi}.$$ This completes the proof.$\, \Box$ This theorem tells us that, since the cluster states have remarkable quantum correlations, which has been discussed in Sec. II, they play an essential role in the realization of arbitrary unitary gates. More precisely, as long as one cluster can be used to process a unitary gate for the cluster state, it will work for arbitrary input states. Therefore, it is sufficient to check the conditions for the cluster states, *i.e.*, Eq. (\[check1\]) and Eq. (\[check2\]). To be emphasized, the special features of the above theorem arising from qudits is that it is expressed not only in terms of unitary operators $ X$ and $Z$ , but also in terms of their conjugates. For $d=2$, it exactly reduces to the Theorem $1$ of Ref. [@Rau1]. Before using the theorem to construct a specific unitary gate, we need to explain how to deal with the byproduct part $U_{\Sigma }$. The basic idea is to move $U_{\Sigma }$ to the front of $U$ according to the commutation relations between $U_{\Sigma }$ and $U$. To complete this operation, the general strategy is to divide measurements into several steps, in which the subsequent measurements depend on the results of previous measurements. In the next section, we will use specific examples to demonstrate how to construct all basic elementary gates with the help of this theorem. Universality of qudit cluster quantum computation ================================================= It is well known that, for qubit quantum computing network, a finite collection of one qubit unitary operations and ${CNOT}$ gate is enough to construct any unitary transformation in the network. This conclusion has been shown to work well for qudit case [@Bry]. More precisely, the collection of all one-qudit gates and any one imprimitive two-qudit gate is exactly universal for arbitrary quantum computing, where a primitive two-qudit gate is such a gate which maps all separate states to separate states. Therefore, in order to prove universality of quantum computation with qudit clusters, we only need to construct those basic element gates, and then to integrate them to realize arbitrary unitary gate. In this section, based on Theorem [coruni]{} in Sec. III, we will build certain qudit clusters to simulate the basic elementary gates, *i.e.*, any one-qudit unitary operation and one imprimitive two-qudit unitary transformation. Realizations of single qudit unitary transformations ---------------------------------------------------- Let us start with single qudit gates. First, we introduce a proposition for any unitary transformation in $d$ dimensional Hilbert space. Let $\{N_i, \;(i\in Z_{d^2-1})\}$ be a Hermitian basis of the operator space for $d$ dimensional Hilbert space, then any unitary transformation $U$ has the form $$U={q_d}^{\sum_i \alpha_i N_i}=\prod_i{q_d}^{ \beta_i N_i},$$ where $\alpha_i$ and $\beta_i$ are real numbers. The first equality is obvious, and we refer to Ref. [@pur] for the proof of the second equality. By the above proposition, we can divide any qudit gate into a product of more simple basic ones, *i.e.*, single parameter unitary transformations. Now, we need to find $d\times d$ independent $N_i$ to simulate all unitary gates for one qudit. To use the definition of qudit cluster states, we expect that the single parameter unitary transformations must have deep relations with the basis elements of QPA. In fact, we find such a way to introduce $d\times d$ one parameter unitary transformations. This originates from that the fact that some of the basis elements of QPA can be used to define a state basis. We find that all the unitary transformations that do not change every basis state up to a phase are defined by the property of multi-values for complex functions. For example, for operator $Z$, we can define $$Z^{\beta}(\{m\})=q_d^{\beta N(Z,\{m\})}, \,( \beta\in \mathcal{R}),$$ where $$N(Z,\{m\})=\sum_{n=0}^{d-1} |n\rangle (n+m_n d)\langle n|,\,(\forall m_n\in \mathcal{Z}).$$ Although the above definitions include infinity unitary transformations, there are only $d$ independent ones, which can be used to describe the following type of unitary transformations $$U_Z (\{\alpha\})|n\rangle =q_d^{ \alpha_n} |n\rangle, \,(\forall n\in Z_d, \alpha\in\mathcal{R}).$$ Obviously, these $d$ independent unitary transformations, also the corresponding $N(Z,\{m\})$, can take the place of $\{Z^n, n \in d\}$ in the unitary basis. A similar argument can be generalized to $\bar{Z}$, which is defined in Sec. II. More precisely, for operator $\bar{Z}$, we define $$\bar{Z}^{\beta}(\{m\})=q_d^{\beta N(\bar{Z},\{m\})}, \,( \beta\in \mathcal{R}),$$ where $$N(\bar{Z},\{m\})=\sum_{n=0}^{d-1} |n(\bar{Z})\rangle (n+m_n d)\langle n(\bar{Z})|,\,(\forall m_n\in \mathcal{Z}),$$ with $$\bar{Z}|n(\bar{Z})\rangle=q_d^n|n(\bar{Z})\rangle.$$ In the following, we will show that we can select $d\otimes d$ independent Hermitian operators from $N(\bar{Z},\{m\})$. When $d$ is a prime number, a convenient choice is to take $\bar{Z}$ from the operator set $$\{Z, X, ZX, ZX^2,\cdots, ZX^{d-1}\}.$$ Because each $\bar{Z}$ defines $d-1$ independent $N(\bar{Z},\{m\})$ besides the identity, we obtain $d^2\,(=(d-1)(d+1)+1)$ independent Hermitian operators. When $d$ is not a prime number, we can choose the independent $N(\bar{Z},\{m\})$ by the following procedure. First, we take $\bar{Z}$ from $\{Z, \, X,\, ZX\}$, and we can obtain $3(d-1)$ independent $N(\bar{Z},\{m\})$ besides the identity, which can take the place of the set of basis elements $$S=\{ Z^n, X^n, Z^nX^n, \,(n\in Z_d)\}.$$ Then we take an element $\bar{Z}\notin S$, find out the elements in $\{\bar{Z}^n, \, n\in Z_d\}$ that is not in the set $S$, add these elements into $S$, and take the new independent $N(\bar{Z},\{m\})$, whose number is the number of new elements in set $S$. We repeat the above step until set $S=B$, then we obtain $d^2$ independent Hermitian operators. Therefore if we can do all the above basic unitary transformations $\bar{Z}^{\beta}(\{m\}$, we will declare that we can do all single qudit unitary transformations. Here we adopt the following strategy: First, we realize $Z^{\alpha}(\{m\}$ on a five qudit cluster, as a basic single qudit transformation. To associate with all the other single qudit unitary transforamtion $\bar{Z}^{\alpha}_{\{m\}}$, we observe an useful fact $$\bar{Z}^{\alpha}(\{m\})=U Z^{\alpha}(\{m\}){U}^{\dagger},\label{clif}$$ where $U$ (or $U^{\dagger}$) can be taken as an element in the Clifford group as defined in Sec. II. Based on this observation, we further discuss how to implement these elements in the Clifford group on clusters. Then we only need to connect these clusters in the given order to realized unitary gate $\bar{Z}^{\alpha}(\{m\})$. If we succeed to pass the above procedure, in principle we can make any single qudit unitary gate in $SU(d)$. Now, let us realize these basic unitary transformations with a specifically-designed qudit cluster, which will be given as follows. ### Five qudit cluster realization of $X^\alpha(\{m\})$ and $Z^\alpha(\{m\})$ In this subsection, we aim to realize the basic single qudit unitary transformation $Z^\alpha(\{m\})$ on a five qudit cluster designed as in Fig. \[fig5\], which is a linear array of five qudits. By the way, we also use the same cluster to implement $X^\alpha(\{m\})$, which demonstrates that the same cluster with different measurement patterns can realize different unitary transformations. ![Five qudit cluster used in realization of $X^\alpha(\{m\})$ and $Z^\alpha(\{m\})$. A circle represents one qudit, number $n$ in the circle means the $n$-th qudit, $in$ or $out$ denote the input or output part of the cluster, and two qudits which are connected by a line is neighbors. []{data-label="fig5"}](5qx.eps) The corresponding cluster state is defined by the following system of equations: $$\begin{aligned} X_{1}^{\dagger }Z_{2}|\phi \rangle_{\mathcal{C}} &=&|\phi \rangle_{\mathcal{C}} , \label{5qdit1} \\ Z_{1}X_{2}^{\dagger }Z_{3}|\phi \rangle_{\mathcal{C}} &=&|\phi \rangle_{\mathcal{C}}, \label{5qdit2} \\ Z_{2}X_{3}^{\dagger }Z_{4}|\phi\rangle_{\mathcal{C}} &=&|\phi \rangle_{\mathcal{C}} , \label{5qdit3} \\ Z_{3}X_{4}^{\dagger }Z_{5}|\phi \rangle_{\mathcal{C}} &=&|\phi \rangle_{\mathcal{C}} , \label{5qdit4} \\ Z_{4}X_{5}^{\dagger }|\phi \rangle_{\mathcal{C}} &=&|\phi \rangle_{\mathcal{C}} . \label{5qdit5}\end{aligned}$$It follows from Eqs. (\[5qdit1\]-\[5qdit5\]) that $$\begin{aligned} X_{1}X_{3}^{\dagger }X_{5}|\phi \rangle_{\mathcal{C}} &=&|\phi \rangle_{\mathcal{C}} , \\ Z_{1}^{\dagger }X_{2}X_{4}^{\dagger } Z_{5}|\phi\rangle_{\mathcal{C}} &=&|\phi \rangle_{\mathcal{C}} . \label{}\end{aligned}$$Form Eq. (\[5qdit5\]), we obtain $$Z_4^{\dagger \alpha}(\{m\}) X_5^{\alpha}(\{m\})|\phi \rangle_{\mathcal{C}}=|\phi \rangle_{\mathcal{C}}.$$ Here we emphasizes that when we obtain the above equation, we have used the following condition On $\{m\}$. If $n_4+n_5=0 \;(Mod(d))$, then $$n_4+m_{n_4}+n_5+m_{n_5}=0,\; (n_4, n_5\in Z_d).$$ From the above four equations, we have $$\begin{aligned} X_{1}X_{3}^{\dagger }\left(X_5^{\alpha}\left(\{m\}\right)X_{5}X_5^{\dagger \alpha}\left(\{m\}\right)\right) |\phi \rangle_{\mathcal{C}} &=&|\phi \rangle_{\mathcal{C}} , \\ Z_{1}^{\dagger }X_{2}\left(Z_4^{\dagger \alpha}(\{m\})X_{4}^{\dagger }Z_4^{\alpha}(\{m\})\right)\left(X_5^{\alpha}(\{m\})Z_{5}X_5^{\dagger \alpha}(\{m\})\right)|\phi \rangle_{\mathcal{C}} &=&|\phi \rangle_{\mathcal{C}} . \label{x20}\end{aligned}$$For a measurement pattern $\{X_{2}, \,X_{3}^{\dagger },\, Z_4^{\dagger \alpha}(\{m\})X_{4}^{\dagger }Z_4^{\alpha}(\{m\})\}$, Theorem \[coruni\] concludes that the simulated unitary transformation is $X_5^{\alpha}(\{m\})U_{\Sigma }$, where $$U_{\Sigma}=Z_5^{-s_1-s_3}X_5^{s_2+s_4}.$$ Because $U_{\Sigma }$ depends on the measurement results and can not be moved to the front of $X_5^{\alpha}(\{m\})$ trivially, different measurement results lead to different unitary transformations. In order to realize unique gate $X_5^{\alpha}(\{m\})$, we complete the measurement in two steps: In the first step, we measure $\{X_{1},X_{2},X_{3}^{\dagger }\}$. Then, when the obtained values are $s_{1}$ , $\lambda _{x}\,(=s_{3})$ and $s_{2}$, the byproduct operator reads $U_{\Sigma }=Z^{-s_{3}-s_{1}}X^{\lambda _{z}}$. At this time, $\lambda _{z}$ is still unknown since it depends on the measured result $s_{4}$. However, notice that $ Z_{5}^{-s_{3}-s_{1}}X_5^{\alpha}(\{m\})Z_{5}^{s_{3}+s_{3}}$ is diagonal in the $X$ representation, we have $$Z_{5}^{-s_{3}-s_{1}}X_5^{\alpha}(\{m\})Z_{5}^{s_{3}+s_{1}}=\prod_{\{m\}} X_5^{\alpha^{s_1 s_3}_{\{m\}}}(\{m\}). \label{sepx}$$ From Eq. (\[sepx\]), we can obtain $d$ equations of $\{\alpha^{s_1 s_3}_{\{m\}}\}$, which determine the values of $% \{\alpha^{s_1 s_3}_{\{m\}}\}$. Now we make a new choice depending on the known measurement results. Also from Eqs. (\[5qdit1\]-\[5qdit5\]), we obtain the following equation instead of Eq. (\[x20\]) $$\begin{aligned} &&Z_{1}^{\dagger }X_{2}\left(\prod_{\{m\}} Z_4^{\dagger \alpha^{s_1 s_3}_{\{m\}}}(\{m\})X_{4}^{\dagger }\prod_{\{m\}} Z_4^{\alpha^{s_1 s_3}_{\{m\}}}(\{m\})\right) \nonumber \\ &&\left(\prod_{\{m\}} X_5^{\alpha^{s_1 s_3}_{\{m\}}}(\{m\})Z_{5}\prod_{\{m\}} X_5^{\dagger \alpha^{s_1 s_3}_{\{m\}}}(\{m\})\right)|\phi \rangle_{\mathcal{C}} =|\phi \rangle_{\mathcal{C}} . \label{x201}\end{aligned}$$Measuring the fourth qudit in the base $$\prod_{\{m\}} Z_4^{\dagger \alpha^{s_1 s_3}_{\{m\}}}(\{m\})X_{4}^{\dagger }\prod_{\{m\}} Z_4^{\alpha^{s_1 s_3}_{\{m\}}}(\{m\}),$$we obtain the value $s_{4}$ and therefore $\lambda _{z}=s_{2}+s_{4}$. According to Theorem \[coruni\], we obtain the final operation $$UU_{\Sigma }=Z_{5}^{-s_{3}-s_{1}}X_5^{s_2+s_4} X_5^{\alpha}(\{m\}).$$Finally, by measuring $Z_{5}$ we obtain the correct result $$s=s_{5}+s_{2}+s_{4}. \label{corres}$$ The above equation does not mean that the final result depends only on the measurements to the second, fourth, and fifth qudit, because different values of $s_{1}$ and $s_{3}$ correspond to different measurement on the fourth qudit. Based on the same cluster, we can also implement the single qudit rotation $% Z^\alpha (\{m\})$. Similarly, we first measure $\{X_{1},X_{2},x^{\dagger}_4\}$; Then from Eqs. (\[5qdit1\]-\[5qdit5\]), we obtain $$\begin{aligned} X_{1}\left(\prod_{m}Z_3^{\alpha^{s_2 s_4}_{\{m\}}} (\{m\})X_{3}^{\dagger }\prod_{m}Z_3^{\dagger \alpha^{s_2 s_4}_{\{m\}}} (\{m\})\right) && \nonumber \\ \left(\prod_{m}Z_5^{\alpha^{s_2 s_4}_{\{m\}}} (\{m\})X_{5}\prod_{m}Z_5^{\dagger \alpha^{s_2 s_4}_{\{m\}}} (\{m\})\right)|\phi \rangle_{\mathcal{C}} &=&|\phi \rangle _{\mathcal{C}}, \\ Z_1^{\dagger }X_{2}X_{4}^{\dagger }Z_{5}|\phi \rangle_{\mathcal{C}} &=&|\phi \rangle_{\mathcal{C}} .\end{aligned}$$ where $\{\alpha^{s_2 s_4}_{\{m\}}\}$ is determined by $$X_{5}^{s_{2}+s_{4}}Z_5^{\alpha}({\{m\}})X_{5}^{-s_{2}-s_{4}}=\prod_{\{m\}}Z_5^{\alpha_{\{m\}}^{s_2 s_4}}({\{m\}}).$$ At the same time we make another measurement on $$\prod_{m}Z_3^{\alpha^{s_2 s_4}_{\{m\}}} (\{m\})X_{3}^{\dagger }\prod_{m}Z_3^{\dagger \alpha^{s_2 s_4}_{\{m\}}} (\{m\}).$$ According to theorem \[coruni\], we conclude that the simulated unitary transformation indeed is $$Z^{-s_{3}-s_{1}}X^{s_{2}+s_{4}}Z_5^{\alpha}({\{m\}}).$$ The correct result and the measurement values are also related by Eq. ([corres]{}). ### Realizations of single qudit elements in Clifford group As implied in Eq. (\[clif\]), we only need to realize the single qudit elements in the Clifford group. It is easy to show that not all elements in the Clifford group are required. In fact, we only need do the elements defined as $$U^{mn}Z{U^{mn}}^\dagger=\bar{Z},\label{dmn}$$ where $$\bar{Z}=q_d^{-\frac{d-1} {2} mn}Z^mX^n,$$ with $$(m,n)=1.$$ We will show that we can do all the above Clifford unitary transformations through a series of four basic types of unitary transformations. The first is defined as $$\begin{aligned} U^{1n}Z{{U^{1n}}^\dagger}&=&q_d^{-\frac{d-1} {2} n}Z X^{n},\\ U^{1n}X{{U^{1n}}^\dagger}&=&X;\end{aligned}$$ The second is defined as $$\begin{aligned} U^{n1}Z{{U^{n1}}^\dagger}&=&q_d^{-\frac{d-1} {2} n}Z^{n} X,\\ U^{n1}X{{U^{n1}}^\dagger}&=&Z^{\dagger};\end{aligned}$$ The third is defined as $$\begin{aligned} V Z{{V}^\dagger}&=&q_d^{-\frac{d-1} {2}}Z X,\\ V X{{V}^\dagger}&=&X;\end{aligned}$$ The last is defined as $$\begin{aligned} W Z{{W}^\dagger}&=&Z,\\ W X{{W}^\dagger}&=&q_d^{-\frac{d-1} {2}}Z X.\end{aligned}$$ For any unitary transformation $U^{mn}$ partially defined by Eq. (\[dmn\]), it can be factorized into a product of a series of the above four basic unitary transformations. **Proof**. Let us prove it by induction. We denote $N_d\equiv\{1,2,\cdots, d\}$. When $m=1$ or $n=1$, it is the first or the second types of unitary transformations. Let us suppose the above theorem is valid at $m\in N_d$ or $n\in N_d$, *i.e.*, we can do $$U^{mn}\qquad (m \;\mbox{or}\; n\in N_d,\, (m,n)=1 ).$$ For arbitrary positive integer $n$ that is relatively prime to $d+1$, there exists $$n=i(d+1)+n^{\prime}, \quad (i\in Z_{\infty},\, n^{\prime}\in N_d).$$ Obviously, $$n^{\prime}\in N_d,\, (d+1,n^{\prime})=1.$$ According to the assumption, we can do $U^{(d+1) n^{\prime}}$. Then let $V$ operate $i$ times, $$V^i U^{(d+1)n^{\prime}} {V^{\dagger}}^i=U^{(d+1)n}.$$ According to the assumption, we can also do $U^{n^{\prime}(d+1) }$. Then let $W$ operate $i$ times, $$W^i U^{n^{\prime}(d+1)} {W^{\dagger}}^i=U^{n(d+1)}.$$ Therefore the theorem is valid for $m \;\mbox{or}\; n \in N_{d+1}$. This completes the proof. $\Box$ Now we come to constructing the four basic unitary transformations. We will prove that the first (including the third) and the fourth can be realized on the five qudit cluster as Fig. [\[fig5\]]{}. According to Eqs. (\[5qdit1\]-\[5qdit5\]), we have $$\begin{aligned} X_{1}X_{3}^{\dagger }X_{5}|\phi \rangle_{\mathcal{C}} &=&|\phi \rangle_{\mathcal{C}} , \\ Z_{1}^{\dagger }X_{2}({Z_{4}}^n X_{4})^{\dagger } Z_{5}{X_{5}}^n|\phi\rangle_{\mathcal{C}} &=&|\phi \rangle_{\mathcal{C}} . \label{}\end{aligned}$$ According to theorem \[coruni\], when the measurement pattern is $\{X_{2},\,X_{3}^{\dagger },\,({q_d}^{-\frac {d-1} {2} n}{Z_{4}}^n X_{4})^{\dagger }\}$, the corresponding unitary transformation is $${q_d}^{\frac {d-1} {2} (s_1+s_3)n}Z^{-s_1-s_3}X^{s_2+s_4-n(s_1+s_3)} U^{1n}.$$ Also from Eqs. (\[5qdit1\]-\[5qdit5\]), we obtain $$\begin{aligned} X_{1}Z_{3}X_{3}^{\dagger }X_{4}^{\dagger}Z_{5}X_{5}|\phi \rangle_{\mathcal{C}} &=&|\phi \rangle_{\mathcal{C}} , \\ Z_{1}^{\dagger }X_{2}X_{4}^{\dagger } Z_{5}|\phi\rangle_{\mathcal{C}} &=&|\phi \rangle_{\mathcal{C}} . \label{}\end{aligned}$$ According to theorem \[coruni\], when the measurement pattern is $\{X_{2},\,{q_d}^{\frac {d-1} {2} n}Z_{3}X_{3}^{\dagger },\, X_{4}^{\dagger }\}$, the corresponding unitary transformation is $${q_d}^{\frac {d-1} {2} (s_2+s_4)}Z^{s_2-s_1-s_3}X^{s_2+s_4} W.$$ To realize the unitary transformation $U^{n1}$, we need a cluster composed of six qudits as Fig. \[fig6a\]. ![Six qudit cluster used in realization of $U^{n1}$. Meanings of the symbols in this Figure are the same as Fig. \[fig5\].[]{data-label="fig6a"}](6qta.eps) The cluster state is defined by the following system of equations $$\begin{aligned} X_{1}^{\dagger }Z_{2}|\phi _{\mathcal{C}}\rangle &=&|\phi \rangle_{\mathcal{C}} , \label{6qdita1} \\ Z_1 X_{2}^{\dagger }Z_{3}|\phi\rangle_{\mathcal{C}} &=&|\phi \rangle_{\mathcal{C}} , \label{6qdita2} \\ Z_{2}X_{3}^{\dagger }Z_{4}|\phi\rangle_{\mathcal{C}} &=&|\phi \rangle_{\mathcal{C}} , \label{6qdita3} \\ Z_{3}X_{4}^{\dagger }Z_{5}|\phi\rangle_{\mathcal{C}} &=&|\phi \rangle_{\mathcal{C}}, \label{6qdita4} \\ Z_{4}X_{5}^{\dagger }Z_{6}|\phi \rangle_{\mathcal{C}} &=&|\phi \rangle_{\mathcal{C}} , \label{6qdita5} \\ Z_{5}X_{6}^{\dagger }|\phi\rangle_{\mathcal{C}} &=&|\phi \rangle_{\mathcal{C}} . \label{6qdita6}\end{aligned}$$It follows from the above equations that $$\begin{aligned} X_{1}X_{3}^{\dagger}X_{5} Z_{6}^{\dagger}|\phi \rangle_{\mathcal{C}} &=&|\phi\rangle_{\mathcal{C}} , \\ Z_{1}^{\dagger}X_{2}{Z_{4}}^nX_{4}^{\dagger}{X_{5}^{\dagger}}^n{Z_{6}}^nX_{6}|\phi \rangle_{\mathcal{C}} &=&|\phi \rangle_{\mathcal{C}}.\end{aligned}$$ When the measurement pattern is $% \{X_{2},\,X_{3}^{\dagger},\,{q_d}^{\frac {d-1} {2} n}{Z_{4}}^nX_{4}^{\dagger},\,X_{5}\}$, the corresponding unitary transformation is $${q_d}^{(s_1+s_3)(s_2+s_4-ns_5+\frac {d-1} {2} n)}Z^{n(s_5-s_1-s_3)-s_2-s_4}X^{-s_1-s_3} U^{n1}.$$ Realization of an imprimitive two qudit gate -------------------------------------------- Now we come to the construction for simulating two qudit operations. ![Six qudit cluster used in realization of an imprimitive two qudit gate $T$. Meanings of the symbols in this Figure are the same as Fig. \[fig5\].[]{data-label="fig6"}](6qt.eps) The cluster composed of six qudits as Fig. \[fig6\] is considered with the following system of equations $$\begin{aligned} X_{1}^{\dagger }Z_{3}|\phi _{C}\rangle &=&|\phi \rangle_{\mathcal{C}} , \label{6qdit1} \\ X_{2}^{\dagger }Z_{4}|\phi\rangle_{\mathcal{C}} &=&|\phi \rangle_{\mathcal{C}} , \label{6qdit2} \\ Z_{1}X_{3}^{\dagger }Z_{4}Z_{5}|\phi\rangle_{\mathcal{C}} &=&|\phi \rangle_{\mathcal{C}} , \label{6qdit3} \\ Z_{2}Z_{3}X_{4}^{\dagger }Z_{6}|\phi\rangle_{\mathcal{C}} &=&|\phi \rangle_{\mathcal{C}}, \label{6qdit4} \\ Z_{3}X_{5}^{\dagger }|\phi \rangle_{\mathcal{C}} &=&|\phi \rangle_{\mathcal{C}} , \label{6qdit5} \\ Z_{4}X_{6}^{\dagger }|\phi\rangle_{\mathcal{C}} &=&|\phi \rangle_{\mathcal{C}} . \label{6qdit6}\end{aligned}$$It follows from the above equations that $$\begin{aligned} X_{1}X_{5}^{\dagger }|\phi \rangle_{\mathcal{C}} &=&|\phi\rangle_{\mathcal{C}} , \\ X_{2}X_{6}^{\dagger }|\phi\rangle_{\mathcal{C}} &=&|\phi \rangle_{\mathcal{C}} , \\ Z_{1}^{\dagger }X_{3}Z_{5}^{\dagger}X_{6}^{\dagger }|\phi \rangle_{\mathcal{C}} &=&|\phi \rangle_{\mathcal{C}} , \\ Z_{2}^{\dagger }X_{4}Z_{6}^{\dagger}X_{5}^{\dagger }|\phi \rangle_{\mathcal{C}} &=&|\phi\rangle_{\mathcal{C}}.\end{aligned}$$ By measuring the system according to the measurement pattern $% \{X_{1}X_{2}X_{3}X_{4}Z_{5}Z_{6}\}$, the simulated two qudit gate $T$ satisfies, and is also defined by $$\begin{aligned} TX_{5}T^{\dagger } &=&X_{5}^{\dagger }, \\ TX_{6}T^{\dagger } &=&X_{6}^{\dagger }, \\ TZ_{5}T^{\dagger } &=&Z_{5}^{\dagger} X_{6}^{\dagger }, \\ TZ_{6}T^{\dagger } &=&Z_{6}^{\dagger} X_{5}^{\dagger }.\end{aligned}$$According to Theorem $\ref{coruni}$, the above measurement pattern realize the following unitary gate $$TZ_{5}^{-s_{1}}X_{5}^{s_{3}}Z_{6}^{-s_{2}}X_{6}^{s_{4}}={q_d}^{ s_{1}s_{2}}Z_{5}^{s_{1}}X_{5}^{s_{2}-s_{3}}Z_{6}^{s_{2}}X_{6}^{s_{1}-s_{4}}T.$$ The next task is to prove that $T$ is an imprimitive two qudit operation. Ref. [@Bry] tells us that a two gate $V$ is primitive if and only if $V=S_{1}\otimes S_{2}$ or $V=(S_{1}\otimes S_{2})P$. Here, $S_{1}$ and $S_{2}$ are different single qudit operators, $P$ is the interchanging operator obeying $P|x\rangle \otimes |y\rangle =|y\rangle \otimes |x\rangle $. Based on this fact, we can easily conclude that a primitive operator always maps a single qudit operator to another single qudit operator. Obviously, the above two qudit operator $T$ is imprimitive. Another way to prove it is to evaluate the unitary transformation directly. Then we can find that it maps all $Z\otimes Z-$ bases to the maximum entangle states. As demonstrated in this section, any single qudit unitary gate and one imprimitive two qudit gate can be realized on qudit clusters. Therefore, the measurement based quantum computing on qudit clusters is universal. Conclusions =========== We have introduced the concept of qudit cluster state in terms of finite dimensional representations of QPA. Based on these qudit cluster states, we have built all the elements of qudit clusters needed for implementation of universal measurement-based quantum computations. With generalizations of cluster states and measurement patterns, most of the results in qubit cluster can work well for qudit clusters in parallel ways. We also show that there still exists the celebrated theorem guaranteeing the availability of qudit cluster states to quantum computations. To prove the universality of this quantum computation, we show that we can implement all single qudit unitary transformations and one imprimitive two qudit gate on specific qudit clusters. In addition, we propose to build a one-way universal quantum computer with qudit cluster states practically since the high dimensional Isingmodel can be used to generate such cluster state dynamically. The authors would like to thank Prof. X. F. Liu for useful discussions. The work of D. L. Z is partially supported by the National Science Foundation of China (CNSF) grant No. 10205022. The work of Z. X is supported by CNSF (Grant No. 90103004, 10247002). The work of C. P. S is supported by the CNSF( grant No. 10205021)and the knowledged Innovation Program (KIP) of the Chinese Academy of Science. It is also funded by the National Fundamental Research Program of China with No 001GB309310. [99]{} M. A. Nielsen and I. S. Chuang, *Quantm computation and quantum information*, Cambridge University Press (2000). H. J. Briegel and R. Raussendorf, Phys. Rev. Lett. **86**, 910 (2001). R. Raussendorf and H. J. Briegel, Phys. Rev. Lett. **86**, 5188 (2001). R. Raussendorf, D. E. Browne, and H. J. Briegel, quant-ph/0301052 (2003). L. M. Duan, E. Demler, and M. D. Lukin, cond-mat/0201564 (2002). C. P. Sun, in “Quantum Group and Quantum Integrable Systems", ed by M. L. Ge, World Scientific, 1992, p.133; M. L. Ge, X. F. Liu, C. P. Sun, J. Phys A-Math Gen **25** (10): 2907, (1992). S. D. Bartlett, H. de Guise, and B. C. Sanders, Phys. Rev. **A65**, 052316 (2002). B. C. Sanders, S. D. Barlett, and H. de Guise, quant-ph/0208008(2002). J. Jamil, X. G. Wang, and B. C. Sanders, quant-ph/0211185 (2002). D. Gottemann, A. Kitaev and J. Preskill, Phys. Rev. **A65**, 044303 (2002). J. Patera and H. Zassenhaus, J. Math. Phys. **29**, 665 (1988). E. Knill, quant-ph/9608048 (1996). H. Weyl, *Theory of groups and quantum mechanics*, New York: E. P. Dutton Co., (1932). J. Schwinger, Proceedings of the National Academy of Sciences, **46** (4), 570 (1960). R. R. Puri, *Mathematical methods of quantum optics*, Springer press (2001). J. L. Brylinski and R. Brylinski, quant-ph/0108062 (2001). M. Blume, V. J. Emery, and R. B. Griffiths, Phys. Rev. **A4**, 1071 (1971). J. Sivardiere and M. Blume, Phys. Rev. **B5**, 1126 (1972). J. Kwo et al., Phys. Rev. Lett. **55**, 1402 (1985). C. F.Majkrzak et al., Phys. Rev. Lett. **56**, 2700 (1986). J. J. Krebs, P. Lubirtz, A. Chaiken, and G. A. Prinz, Phys. Rev. Lett. **63**, 1645 (1989). R. E. Camiey, J. Kwo, M. Hong, and C. L. Chien, Phys. Rev. Lett. **64**, 2703 (1990). N. Sh. Izmailian, arXiv:hep-th/9603080. A. Mattoni, M. Bagagiolo, and A. Saber1, Chinese Journal of Physics **40**, 3, 2002. L. L. Hinchey and D. L. Mills, Phys. Rev. **B33**, 3329 (1986). B. D. Qu, W. L. Zhong, P. L. Zhang, Phys. Lett. **A189**, 419 (1994). F. Fisman, F. Schwabl, and D. Schwenk, Phys. Lett. **A121**, 192 (1987). J. Gritz-Saaverda, F. Aguilra-Grania, and J. L. Moran-Lopez, Solid State Commun. **82**, 5891 (1992). H. K. Sy and M. H. Ow, J. Phys. Condens. Matter, **4**, 5891 (1992). A. Saber, A. Ainane, M. Saber, I. Essaoudi, F. Dujardin, and B. St¡äeb¡äe, Phys. Rev. **B60**, 4149 (1999).

ArXiv

--- abstract: 'We propose a space-frequency (SF) block coding scheme for a multiple-input multiple-output (MIMO) orthogonal frequency-division multiplexing (OFDM) system using antennas with reconfigurable radiation patterns. In this system, each element of the antenna array at the transmitter side is assumed to be reconfigurable so that it can independently change the physical characteristics of its radiation pattern. The proposed block code is full rate and benefits from spatial, frequency, and reconfigurable radiation pattern state diversity over frequency-selective fading channels. We provide simulation results to demonstrate the performance of the proposed block coding technique and make comparisons with that of the previous SF coding schemes in MIMO-OFDM systems. The results indicate that the proposed code achieves higher diversity and coding gain compared to other available SF codes.' author: - bibliography: - 'IEEEabrv.bib' - 'Reference.bib' title: 'Space-Frequency Block Code for MIMO-OFDM Communication Systems with Reconfigurable Antennas' --- Frequency-selective fading channels, multiple-input multiple-output-orthogonal frequency-division multiplexing (MIMO-OFDM) systems, space-frequency (SF) coding, reconfigurable antennas. Introduction ============ Reconfigurable antennas can be used in multiple-input multiple-output (MIMO) communication systems to increase the capacity and reliability of wireless links [@cetiner2006mimo; @piazza2008design; @frigon2008dynamic; @li2009capacity; @grau2008reconfigurable]. In a reconfigurable MIMO system, the characteristics of each antenna radiation pattern can be changed by placing switching devices such as Microelectromechanical Systems (MEMS), varactor diodes, or field-effect transistor (FET) within the antenna structure [@weedon2001mems; @caloz2005electromagnetic; @won2006reconfigurable]. As a result, a system employing reconfigurable antennas is able to alter the propagation characteristics of the wireless channel into a form that leads to a better signal quality at the receiver. In fact, by using reconfigurable antennas and designing a proper code, we can achieve an additional diversity gain that can further improve the performance of wireless communication systems. There are several works in the literature on designing efficient codes for reconfigurable MIMO systems in order to take advantage of the antenna reconfigurability. In [@grau2008reconfigurable], authors have proposed a MIMO system equipped with reconfigurable antennas at the receiver that can achieve a diversity order that equals to the product of the number of transmit antennas, the number of receive antennas and the number of reconfigurable states of the receive antennas. They have shown that this diversity gain is achievable only under certain channel propagation conditions and using an appropriate coding technique. Later on, in [@fazel2009space] authors extended the concept by using reconfigurable elements at both transmitter and receiver sides. In their work, they have introduced a state-switching transmission scheme, called space-time-state block coding (STS-BC), to further utilize the available diversity in the system over flat fading wireless channels. However, their coding scheme does not exploit the frequency diversity offered by the multipath propagation channels between each transmit and receive antenna pair. To obtain frequency diversity in multipath environment, a space-frequency (SF) block code was first proposed by authors in [@agrawal1998space], where they used the existing space-time (ST) coding concept and constructed the code in frequency domain. Later works [@lee2000space; @bolcskei2000space; @lu2000space; @blum2001improved; @hong2002robust; @lee2000space1] also used similar strategies to develop SF codes for MIMO-OFDM systems. However, the resulting SF codes achieved only spatial diversity, and they were not able to obtain both spatial and frequency diversities. To address this problem, a subcarrier grouping method has been proposed in [@molisch2002space] to further enhance the diversity gain while reducing the receiver complexity. In [@su2003obtaining], a repetition mapping technique has been proposed that obtains full-diversity in frequency-selective fading channels. Although their proposed technique achieves full-diversity order, it does not guarantee full coding rate. Subsequently, a block coding technique that offers full-diversity and full coding rate was derived [@su2005full; @fazel2008quasi]. However, the SF codes proposed in the above studies and other similar works on the topic are not able to exploit the state diversity available in reconfigurable multiple antenna systems. In this paper, we propose a coding scheme for reconfigurable MIMO-OFDM systems that achieves multiple diversity gains, including, space, frequency, and state. Basically, the proposed scheme consists of a code that is sent over transmit antennas, OFDM tones, and radiation states. In order to obtain state diversity, we configure each transmit antenna element to independently switch its radiation pattern to a direction that can be selected according to different optimization criteria, e.g., to minimize the correlation among different radiation states. We construct our proposed code based on the fundamental concept of rotated quasi-orthogonal space-time block codes (QOSTBC) [@tirkkonen2001optimizing; @sharma2003improved; @su2004signal]. By using the rotated QOSTBC, the proposed coding structure provides rate-one transmission (i.e., one symbol per frequency subcarrier per radiation state) and leads to a simpler Maximum Likelihood (ML) decoder. As the simulation results indicate, our proposed code outperforms the existing space-frequency codes substantially.  The rest of this paper is organized as follows. In Section \[sec:ch\_sys\_model\], we introduce the channel and system model for a reconfigurable MIMO-OFDM system. In Section \[sec:code\_deg\], we briefly discuss the code design for reconfigurable multiple antenna systems. Simulation results are presented in Section \[sec:results\], and finally conclusions are drawn in Section \[sec:conc\]. *Notation:* Throughout this paper, we use capital boldface letters for matrices, and lowercase boldface letters for vectors. $(\cdot)^T$ denotes transpose of a vector. $\mathcal{C}$ stands for the set of complex valued numbers. Operator $\text{diag}(a_1, a_2, \cdots, a_n)$ represents a diagonal $n \times n$ matrix whose diagonal entries are $a_1, a_2, \cdots , a_n$. $\lfloor \cdot \rfloor$ stands for the floor operation and ${\bf I}_N$ represents the $N \times N$ identity matrix. Operator $\text{col}\{\cdot\}$ stacks up the matrices on top of each other. Channel and System Models for Reconfigurable MIMO-OFDM Systems {#sec:ch_sys_model} ============================================================== Consider a MIMO-OFDM system with $M_t$ reconfigurable elements at the transmitter where each of these elements is capable of electronically changing its radiation pattern and creating $P$ different radiation states as shown in Fig. \[fig:Sys\_Blk\]. In this system, we assume the receiver antenna array consist of $M_r$ omni-directional elements with fixed radiation patterns. Moreover, we consider an $N_c$-tone OFDM modulation and frequency-selective fading channels with $L$ independent propagation paths between each pair of transmit and receiver antenna in each radiation state. The channel gains are quasi-static over one OFDM symbol interval. The channel impulse response between transmit antenna $i$ and receive antenna $j$ in the $p$-th radiation state can be modelled as $$\begin{aligned} h_{p}^{i,j}(\tau) = \sum_{l=0}^{L-1} \alpha_{p}^{i,j}(l)\delta(\tau-\tau_{l,p}), \label{eq:channel_t}\end{aligned}$$ where $\tau_{l,p}$ is the $l$-th path delay in the $p$-th radiation state, and $\alpha_{p}^{i,j}(l)$ is the complex amplitude of the $l$-th path between the $i$-th reconfigurable transmit antenna and the $j$-th receive antenna in the $p$-th radiation state. The average total received power is normalized to one. The frequency response of the channel at the $n$-th subcarrier between transmit antenna $i$ and receive antenna $j$ in the $p$-th radiation state is given by $$\begin{aligned} H_{p}^{i,j}(n) = \sum_{l=0}^{L-1} \alpha_{p}^{i,j}(l) e^{-j2\pi n \Delta f \tau_{l,p}}, \label{eq:channel_f}\end{aligned}$$ where $\Delta f = 1/T_s$ is the subcarrier frequency spacing and $T_s$ is the OFDM symbol duration. The space-frequency codeword transmitted during the $p$-th radiation state, ${\bf C}_p \in \mathcal{C}^{M_t \times N_c}$, can be expressed as $${\bf C}_{p} = \left[ \begin{array}{cccc} c_p^1(0) & c_p^1(1) & \cdots & c_p^1(N_c-1) \\ c_p^2(0) & c_p^2(1) & \cdots & c_p^2(N_c-1) \\ \vdots & \vdots & \ddots & \vdots \\ c_p^{M_t}(0) & c_p^{M_t}(1) & \cdots & c_p^{M_t}(N_c-1) \end{array} \right], \label{eq:SF}$$ where $c_{p}^i(n)$ denotes the data symbol transmitted by transmit antenna $i$ on the $n$-th subcarrier during the $p$-th radiation state. At the receiver, after cyclic prefix removal and FFT, the received frequency domain signal of the $n$-th subcarrier and $p$-th radiation state at the $j$-th receive antenna can be written as $$\begin{aligned} y_{p}^j (n) = \sqrt{\frac{\gamma}{M_t}} \sum_{i=1}^{M_t} H_{p}^{i,j}(n) c_{p}^i(n) + z_{p}^j(n), \label{eq:Rx_Signal}\end{aligned}$$ where $H_{p}^{i,j}(n)$ is the frequency response of the channel at the $n$-th subcarrier between transmit antenna $i$ and receive antenna $j$ in the $p$-th radiation state as defined in (\[eq:channel\_f\]), $z_{p}^j(n)$ is the additive complex Gaussian noise with zero mean and unit variance at the $n$-th subcarrier, and $\gamma$ is the received signal-to noise ratio (SNR). The received signal during the $p$-th radiation state ${\bf y}_p = [{\bf y}_p^T(0) \; {\bf y}_p^T(1) \; \cdots \; {\bf y}_p^T(N_c-1)]^T$ with ${\bf y}_p(n) = [y_{p}^1 (n) \; y_{p}^2 (n) \; \cdots \; y_{p}^{M_r} (n)]^T$, can be written as $$\begin{aligned} {\bf y}_p = \sqrt{\frac{\gamma}{M_t}} {\bf H}_p {\bf c}_p + {\bf z}_p, \label{eq:Rx_Signal_state}\end{aligned}$$ where $${\bf H}_p = \text{diag}\{{\bf H}_p(0),\; {\bf H}_p(1),\; \cdots,\; {\bf H}_p(N_c-1)\} \label{eq:channel_matrix_state}$$ is the channel matrix, ${\bf c}_p = \text{vec} ({\bf C}_p)$ is the transmitted codeword, and ${\bf z}_p \in \mathcal{C}^{N_cM_r \times 1}$ is the noise vector during the $p$-th radiation state. In (\[eq:channel\_matrix\_state\]), ${\bf H}_p(n)$ is an $M_r \times M_t$ channel matrix with entries defined in (\[eq:channel\_f\]). The SF codeword over all $P$ radiation states can be represented as $${\bf C} = \big[{\bf C}_1, \; {\bf C}_2,\;\cdots, \; {\bf C}_P\big], \label{eq:Code}$$ where ${\bf C}_p$ is given in (\[eq:SF\]). The received signals over all radiation states is defined by ${\bf y} = [{\bf y}_1^T \; {\bf y}_2^T \; \cdots \; {\bf y}_P^T]^T \in \mathcal{C}^{PN_c M_r \times 1}$ and can be represented by $$\begin{aligned} {\bf y} = \sqrt{\frac{\gamma}{M_t}} {\bf H} {\bf c} + {\bf z}, \label{eq:Rx_Signal_state}\end{aligned}$$ where ${\bf c} = \text{vec} ({\bf C})$, ${\bf H} = \text{diag} \{{\bf H}_1,\; {\bf H}_2,\; \cdots,\; {\bf H}_P\} \in \mathcal{C}^{PN_cM_r \times PN_cM_t} $ is the overall channel matrix, and ${\bf z} = [{\bf z}_1^T \; {\bf z}_2^T \; \cdots \; {\bf z}_P^T]^T \in \mathcal{C}^{PN_c M_r \times 1}$ is the noise vector. $$\setcounter{equation}{13} {\bf C}_1 = \frac{1}{4} \left[ \begin{array}{ccccccccc} \mathcal{S}_1^1 & -\mathcal{S}^{1^*}_2 & \mathcal{S}^1_3 & -\mathcal{S}^{1^*}_4 & \cdots & \mathcal{S}_1^M & -\mathcal{S}^{M^*}_2 & \mathcal{S}^M_3 & -\mathcal{S}^{M^*}_4 \\ \mathcal{S}_2^1 & \mathcal{S}^{1*}_1 & \mathcal{S}^1_4 & \mathcal{S}^{1*}_3 & \cdots & \mathcal{S}_2^M & \mathcal{S}^{M*}_1 & \mathcal{S}^M_4 & \mathcal{S}^{M*}_3 \\ \end{array} \right]$$ $${\bf C}_2 = \frac{1}{4} \left[ \begin{array}{ccccccccc} \mathcal{S}_5^1 & -\mathcal{S}^{1^*}_6 & \mathcal{S}^1_7 & -\mathcal{S}^{1^*}_8 & \cdots & \mathcal{S}_5^M & -\mathcal{S}^{M^*}_6 & \mathcal{S}^M_7 & -\mathcal{S}^{M^*}_8 \\ \mathcal{S}_6^1 & \mathcal{S}^{1*}_5 & \mathcal{S}^1_8 & \mathcal{S}^{1*}_7 & \cdots & \mathcal{S}_6^M & \mathcal{S}^{M*}_5 & \mathcal{S}^M_8 & \mathcal{S}^{M*}_7 \\ \end{array} \right] \label{eq:codewordP4}$$ Space-Frequency Code Design for Reconfigurable MIMO-OFDM Systems {#sec:code_deg} ================================================================ In this section, we present our proposed coding scheme for a reconfigurable antenna system, where each antenna elements can independently change its radiation pattern direction. In particular, we construct the code based on the principle of a quasi-orthogonal coding structure for an arbitrary number of transmit antennas and radiation pattern states. In each radiation state, we consider a coding strategy where the SF codeword is a concatenation of ${\bf G}_p^{m^T}$ as follows: $${\bf C}_p = [ {\bf G}_p^{1^T} {\bf G}_p^{2^T} \cdots {\bf G}_p^{M^T} {\bf 0}^T_{N_c-MLM_t} ], \label{eq:SFS_code}$$ where $M=\lfloor \frac{N_c}{LM_t}\rfloor$ and ${\bf 0}_N$ is the all-zeros $N \times N$ matrix. In this expression, ${\bf 0}_N$ will disappear if $N_c$ is an integer multiple of $LM_t$. In this work, for simplicity, we assume $N_c=LM_tq$, for some integer $q$. Each ${\bf G}_p^{m}$ matrix, $m \in \{1,2,\cdots,M\}$, takes the following form: $${\bf G}_p^{m} = \text{col}\{ {\bf X}_1, \; {\bf X}_2, \; \cdots, \; {\bf X}_L\},$$ where ${\bf X}_l$ is the $M_t \times M_t$ block coding matrix which is equivalent to an Alamouti code structure for $M_t=2$. To maintain simplicity in our presentation, we design the code for $M_t=2$ transmit antennas, however, extension to $M_t>2$ is possible by following the similar procedure. In the case of having two transmit antennas, ${\bf X}_l = {\bf A}\big(x_1, x_2\big)$, where $${\bf A}\big(x_1, x_2\big) = \left[ {\begin{array}{cc} x_1 & x_2 \\ -x_2^* & x_1^* \\ \end{array} } \right],$$ and therefore ${\bf G}_p^{m}$ can be expressed as $${\bf G}_p^{m} = \left[ \begin{array}{c} {\bf A}(\mathcal{S}_{2(p-1)L+1}^m,\mathcal{S}_{2(p-1)L+2}^m)\\ {\bf A}(\mathcal{S}_{2(p-1)L+3}^m,\mathcal{S}_{2(p-1)L+4}^m)\\ \vdots \\ {\bf A}(\mathcal{S}_{2pL-1}^m,\mathcal{S}_{2pL}^m)\\ \end{array} \right]. \label{eq:G}$$ In (\[eq:G\]), $\mathcal{S}_i^m$ is a set of combined symbols, computed as $$\begin{aligned} &\left[ \begin{array}{cccc} \mathcal{S}_{\scriptscriptstyle 1}^m \;\mathcal{S}_{\scriptscriptstyle 3}^m \;\cdots \;\mathcal{S}_{\scriptscriptstyle 2PL-1}^m \end{array} \right]^T = \Theta \left[ \begin{array}{cccc} s_{\scriptscriptstyle 1}^m \;s_{\scriptscriptstyle 3}^m \;\cdots \; s_{\scriptscriptstyle 2PL-1}^m \end{array} \right]^T, \nonumber\\ &\left[ \begin{array}{cccc} \mathcal{S}_{\scriptscriptstyle 2}^m \;\mathcal{S}_{\scriptscriptstyle 4}^m \;\cdots \;\mathcal{S}_{\scriptscriptstyle 2PL}^m\;\;\; \end{array} \right]^T = \Theta \left[ \begin{array}{cccc} s_{\scriptscriptstyle 2}^m \;s_{\scriptscriptstyle 4}^m \;\cdots \;s_{\scriptscriptstyle 2PL}^m\;\;\; \end{array} \right]^T, \label{eq:codewordP3}\end{aligned}$$ where $\{s_{\scriptscriptstyle 1}^m,\cdots,s_{\scriptscriptstyle 2PL}^m\}$ is a block of symbols belonging to a constellation $\mathcal{A}$, $\Theta = {\bf U} \times \text{diag}\{1,e^{j\theta_1},\ldots,e^{j\theta_{PL-1}}\}$ and ${\bf U}$ is a $PL \times PL$ Hadamard matrix. The $\theta_i$’s are the rotation angles. Different optimization strategies can be used to find the optimal values of rotation angles $\theta_i$’s, such that they maximize the coding gain. The objective function in this optimization is defined as the minimum Euclidean distance between constellation points. As an example, consider a reconfigurable MIMO-OFDM system with $M_t=2$ transmit antennas, $P=2$ radiation states, and $L=2$ multipaths. In this scenario, the transmitted codewords ${\bf C}_1$ and ${\bf C}_2$ given in (\[eq:codewordP4\]) are constructed according to (\[eq:SFS\_code\]). The entries of ${\bf C}_p$ are computed using (\[eq:codewordP3\]). As a result, we obtain ${\bf C}_1^T$ as $$\begin{aligned} \left[ \begin{array}{cc} s_1^1+\tilde{s}_3^1+\hat{s}_5^1+\check{s}_7^1 & s_2^1+\tilde{s}_4^1+\hat{s}_6^1+\check{s}_8^1 \\ -s_2^{1^*}-\tilde{s}_4^{1^*}-\hat{s}_6^{1^*}-\check{s}_8^{1^*} & s_1^{1^*}+\tilde{s}_3^{1^*}+\hat{s}_5^{1^*}+\check{s}_7^{1^*} \\ s_1^1-\tilde{s}_3^1+\hat{s}_5^1-\check{s}_7^1 & s_2^1-\tilde{s}_4^1+\hat{s}_6^1-\check{s}_8^1 \\ -s_2^{1^*}+\tilde{s}_4^{1^*}-\hat{s}_6^{1^*}+\check{s}_8^{1^*} & s_1^{1^*}-\tilde{s}_3^{1^*}+\hat{s}_5^{1^*}-\check{s}_7^{1^*} \\ \vdots & \vdots\\ s_1^M+\tilde{s}_3^M+\hat{s}_5^M+\check{s}_7^M & s_2^M+\tilde{s}_4^M+\hat{s}_6^M+\check{s}_8^M \\ -s_2^{M^*}-\tilde{s}_4^{M^*}-\hat{s}_6^{M^*}-\check{s}_8^{M^*} & s_1^{M^*}+\tilde{s}_3^{M^*}+\hat{s}_5^{M^*}+\check{s}_7^{M^*} \\ s_1^M-\tilde{s}_3^M+\hat{s}_5^M-\check{s}_7^M & s_2^M-\tilde{s}_4^M+\hat{s}_6^M-\check{s}_8^M \\ -s_2^{M^*}+\tilde{s}_4^{M^*}-\hat{s}_6^{M^*}+\check{s}_8^{M^*} & s_1^{M^*}-\tilde{s}_3^{M^*}+\hat{s}_5^{M^*}-\check{s}_7^{M^*} \\ \end{array} \right], \label{eq:codewordP5}\end{aligned}$$ and ${\bf C}_2^T$ as $$\left[ \begin{array}{cc} s_1^1+\tilde{s}_3^1-\hat{s}_5^1-\check{s}_7^1 & s_2^1+\tilde{s}_4^1-\hat{s}_6^1-\check{s}_8^1 \\ -s_2^{1^*}-\tilde{s}_4^{1^*}+\hat{s}_6^{1^*}+\check{s}_8^{1^*} & s_1^{1^*}+\tilde{s}_3^{1^*}-\hat{s}_5^{1^*}-\check{s}_7^{1^*} \\ s_1^1-\tilde{s}_3^1-\hat{s}_5^1+\check{s}_7^1 & s_2^1-\tilde{s}_4^1-\hat{s}_6^1+\check{s}_8^1 \\ -s_2^{1^*}+\tilde{s}_4^{1^*}+\hat{s}_6^{1^*}-\check{s}_8^{1^*} & s_1^{1^*}-\tilde{s}_3^{1^*}-\hat{s}_5^{1^*}+\check{s}_7^{1^*} \\ \vdots & \vdots\\ s_1^M+\tilde{s}_3^M-\hat{s}_5^M-\check{s}_7^M & s_2^M+\tilde{s}_4^M-\hat{s}_6^M-\check{s}_8^M \\ -s_2^{M^*}-\tilde{s}_4^{M^*}+\hat{s}_6^{M^*}+\check{s}_8^{M^*} & s_1^{M^*}+\tilde{s}_3^{M^*}-\hat{s}_5^{M^*}-\check{s}_7^{M^*} \\ s_1^M-\tilde{s}_3^M-\hat{s}_5^M+\check{s}_7^M & s_2^M-\tilde{s}_4^M-\hat{s}_6^M+\check{s}_8^M \\ -s_2^{M^*}+\tilde{s}_4^{M^*}+\hat{s}_6^{M^*}-\check{s}_8^{M^*} & s_1^{M^*}-\tilde{s}_3^{M^*}-\hat{s}_5^{M^*}+\check{s}_7^{M^*} \\ \end{array} \right], \label{eq:codewordP6}$$ where $\tilde{s}_i = e^{j\theta_1}s_i$, $\hat{s}_i = e^{j\theta_2}s_i$, and $\check{s}_i = e^{j\theta_3}s_i$. Note that the above codeword provides rate-one transmission (i.e., one symbol per OFDM tone per radiation state). Simulation Results {#sec:results} ================== In this section, we present simulation results for both conventional and reconfigurable MIMO-OFDM systems. The reconfigurable multiple antenna system employs antenna elements capable of dynamically changing their radiation pattern directions at the transmitter, where in this work, we consider that each element has $P = 2$ radiation states. The conventional MIMO-OFDM system uses omni-directional antenna elements with fixed radiation pattern at both transmitter and receiver ends. For both systems, we consider $M_t = 2$ antennas at the transmitter and $M_r = 1$ antenna at the receiver and an OFDM modulation technique with $N_c = 128$ subcarriers as well as a cyclic prefix equal to or longer than the maximum channel delay spread. In our simulations, we consider that the receiver has perfect channel state information. We also assume that the symbols are chosen from a BPSK constellation, leading to a spectral efficiency of 1 bit/sec/Hz if the cyclic prefix overhead is ignored. The average symbol power per transmit antenna is set to be $E_s = 1/M_t$ and the noise variance is $\sigma_n^2 = 1/\gamma$. We carry out the simulations for a $2$-ray equal power channel model for two different delay spreads. Furthermore, for a reconfigurable antenna system, we assume the same delay spread for both radiation states (i.e., $\tau_{l,1} = \tau_{l,2}$). The channel coefficients $\alpha_{p}^{i,j}(l)$ are zero-mean identically-distributed Gaussian random variables with a variance of $\sigma_{l,p}^2$. We assumed that they are independent for each multipath, transmit antenna and radiation pattern state. The powers of all paths in each radiation state are normalized such that $\sum_{l=0}^{L-1} \sigma_{l,p}^2 = 1$. For our proposed QOSF scheme, the rotation angles are chosen as $\theta_1 = \frac{\pi}{4}$, $\theta_2 = \frac{\pi}{2}$, and $\theta_3 = \frac{3\pi}{4}$. ![BER vs. SNR for a reconfigurable multi-antenna system with $M_t=2$, $P = 2$, $M_r=1$ in a $2$-ray channel with a delay spread of $5 \mu s$[]{data-label="fig:BER_tau5"}](./Figures/BERvsSNR0,20_SF_QOSF_QOSFS_Nsc128_Ncp21_2RayCh_Tau5){width="3.8in"} Fig. \[fig:BER\_tau5\] shows the bit error rate (BER) performance of the proposed code in multipath propagation channels with a delay spread of $\tau = 5 \mu s$. As shown in this figure, the proposed code outperforms those of [@su2005full] and [@fazel2008quasi]. In particular, at a bit error rate of $10^{-5}$, the performance improvement compared to [@su2005full] and [@fazel2008quasi] is nearly $7$ and $6$ dB, respectively. This performance improvement demonstrates the superiority of our proposed scheme which is due to the extra diversity gain offered by the use of reconfigurable antenna elements. Fig. \[fig:BER\_tau20\] depicts the BER performance of the proposed code for a delay spread of $\tau = 20 \mu s$. It is evident from the figure that at a BER of $10^{-5}$, our proposed coding scheme outperforms the codes presented in [@su2005full] and [@fazel2008quasi] by about $6$ and $4$ dB, respectively. Compared to the results in Fig. \[fig:BER\_tau5\], it can be seen that as delay spread increases, the BER performance improves. This is due to benefiting from lower correlation between subcarriers, and therefore higher frequency diversity in multipath propagation channels. ![BER vs. SNR for a reconfigurable multi-antenna system with $M_t=2$, $P = 2$, $M_r=1$ in a $2$-ray channel with a delay spread of $20 \mu s$ []{data-label="fig:BER_tau20"}](./Figures/BERvsSNR0,20_SF_QOSF_QOSFS_Nsc128_Ncp21_2RayCh_Tau20){width="3.8in"} Conclusions {#sec:conc} =========== We proposed a space-frequency coding technique for MIMO-OFDM systems using antennas with reconfigurable radiation patterns. The proposed code is constructed based on the principle of quasi-orthogonal coding scheme and consists of a block of transmitted symbols expanding over space, frequency, and radiation state dimensions. We provided simulation results to demonstrate the performance of the proposed coding scheme and make comparisons with that of the previous SF coding schemes. In these experiments, it has been shown that the proposed code provides additional diversity and coding gains compared to the previously designed SF codes in MIMO-OFDM systems.

ArXiv

--- abstract: | Detectability of isolated black holes (IBHs) without a companion star but emitting X-rays by accretion from dense interstellar medium (ISM) or molecular cloud gas is investigated. We calculate orbits of IBHs in the Galaxy to derive a realistic spatial distribution of IBHs, for various mean values of kick velocity at their birth $\upsilon_{\rm avg}$. X-ray luminosities of these IBHs are then calculated considering various phases of ISM and molecular clouds, for a wide range of the accretion efficiency $\lambda$ (a ratio of the actual accretion rate to the Bondi rate) that is rather uncertain. It is found that detectable IBHs mostly reside near the Galactic Centre (GC), and hence taking the Galactic structure into account is essential. In the hard X-ray band, where identification of IBHs from other contaminating X-ray sources may be easier, the expected number of IBHs detectable by the past survey by [*NuSTAR*]{} towards GC is at most order unity. However, 30–100 IBHs may be detected by the future survey by [*FORCE*]{} with an optimistic parameter set of $\upsilon_{\rm avg} = 50 \ \mathrm{km\ s^{-1}}$ and $\lambda = 0.1$, implying that it may be possible to detect IBHs or constrain the model parameters. author: - | Daichi Tsuna,$^{1,2}$[^1] Norita Kawanaka,$^{3,4}$ and Tomonori Totani$^{5,1}$\ $^{1}$Research Center for the Early Universe (RESCEU), the University of Tokyo, Hongo, Tokyo 113-0033, Japan\ $^{2}$Department of Physics, School of Science, the University of Tokyo, Hongo, Tokyo 113-0033, Japan\ $^{3}$Department of Astronomy, Graduate School of Science, Kyoto University, Kitashirakawa Oiwake-cho, Sakyo-ku Kyoto, 606-8502, Japan\ $^{4}$Hakubi Center, Yoshida Honmachi, Sakyo-ku, Kyoto 606-8501, Japan\ $^{5}$Department of Astronomy, School of Science, the University of Tokyo, Hongo, Tokyo 113-0033, Japan bibliography: - 'IBH.bib' date: 'Accepted XXX. Received YYY; in original form ZZZ' title: 'X-ray Detectability of Accreting Isolated Black Holes in Our Galaxy' --- \[firstpage\] accretion, accretion discs – black hole physics – Galaxy: general – X-rays: ISM – X-rays: stars Introduction ============ A black hole is thought to form in the last stage of stellar evolution, when a massive star gravitationally collapses at the end of its life. Owing to the development of X-ray observation technology, more than 20 strong black hole candidates have been detected in our Galaxy as X-ray binaries (see, e.g. @RM06 for a review). An even stronger proof of the existence of stellar mass black holes has been obtained by the recent detections of gravitational waves from binary mergers of two black holes [@GW150914; @GW151226; @GW170104; @GW170814; @GW170608]. However it is expected that there are many more black holes without a companion star, which are often called isolated black holes (IBHs). The number of black holes that formed in the Milky Way in the past is estimated to be $\sim 10^8$, based on stellar evolution theory, total stellar mass of the Galaxy, observed metallicity and chemical evolution modeling [e.g., @Shapiro83; @vandenHeuvel92; @Samland98; @Caputo17]. IBHs, which are probably occupying more than half of the total black hole population [@Fender13], are expected to shine by accreting surrounding gas. If mass accretion onto IBHs can be described by the Bondi spherical accretion formula [@HL39; @BH44; @Bondi52], the accretion rate is proportional to $\rho \upsilon^{-3}$, where $\rho$ and $\upsilon$ are density of surrounding gas and IBH velocity, respectively. Therefore IBHs that plunge into a dense gas cloud (which mainly exists in the Galactic disc) with a sufficiently low velocity can obtain a large mass accretion. Such IBHs may be brighter than isolated neutron stars (INSs), because of heavier masses and slower velocities, though INSs are expected to be more abundant than IBHs by an order of magnitude [@Arnett89; @ST10]. There have been many studies on the detectability of such accreting INSs [@Ostriker70; @Treves91; @BM93; @Treves00; @Perna03; @Ikhsanov07] and IBHs [@Shvartsman71; @Grindlay78; @Carr79; @McDowell85; @CP93; @PP98; @Fujita98; @Armitage99; @Grindlay01; @Agol02; @Maccarone05; @MT05; @Sartore10; @Motch12; @Barkov12; @Fender13; @Ioka17; @Matsumoto17]. A number of observational searches have also been performed in the past for such accreting INSs and IBHs [@Stocke95; @Walter96; @Wang97; @Schwope99; @Chisholm03; @Muno06]. Although several INSs candidates have been identified by their thermal emission (e.g. “The Magnificent Seven” identified by the [*ROSAT*]{} satellite; see e.g. [@Haberl07; @Kaplan08] for reviews), no accretion-powered INSs or IBHs have been detected so far. In estimates of accretion-powered IBH detectability, there are some sources of uncertainty. One is the natal kick velocity $\upsilon_{\rm kick}$ of black holes. In spite of intense theoretical and observational studies (@WvP [@JN04; @Gualandris05; @Miller-Jones09; @Fryer12; @Repetto12; @Reid14; @Wong14; @Repetto15; @Mandel16; @Wysocki17]; for a recent extensive review see @Belczynski16), the distribution and mean value of $\upsilon_{\rm kick}$ (hereafter denoted as $\upsilon_{\rm avg}$) are still highly uncertain. BH kick velocities may be smaller or comparable to that of neutron star kicks (typically a few hundred km s$^{-1}$; @Hobbs05) that can be inferred from proper motion of pulsars. Another source of uncertainty is the efficiency of accretion, $\lambda$, which is the ratio of actual accretion rate onto IBHs to the Bondi rate. The spherical Bondi accretion formula can be applied only when angular momentum of accretion flow is negligible. Once angular momentum becomes important, the accretion flow should be described by other modes, such as radiatively inefficient accretion flow (RIAF) [@Ichimaru77; @NY95; @KFM08]. It is generally expected that accreting matter is lost by outflow on various scales, and hence the actual accretion reaching to BH horizons can be significantly reduced. In other words, observational searches for IBHs will give constraints on these parameters of $\upsilon_{\rm avg}$ and $\lambda$. The aim of this paper is to make the best estimate currently possible for the number and luminosities of IBHs that are formed by normal stellar evolution and shining by accretion from ISM. We then evaluate detectability by the past and present X-ray observations such as [*ROSAT*]{} and [*NuSTAR*]{}, and especially we focus on the prospect of a future project [*FORCE*]{} [@FORCE] in hard X-ray band. Since luminous IBHs are expected to be found in dense gas regions, hard X-ray observation would be particularly powerful because of less absorption. Discrimination of IBHs from other populations of X-ray sources is also important, and the hard X-ray band is useful to examine the spectral difference between IBHs and cataclysmic variables [@RM06; @Nobukawa16]. In order to obtain a more reliable estimate than previous studies, we calculate the spatial distribution of IBHs by solving their orbits in the gravitational potential of the Galaxy with various values of kick velocities. Then accretion rate is estimated by using a realistic gas distribution in the Galaxy based on latest observations. The paper is organized as follows. In Section \[sec:formulations\] we give formulations of our calculations, with discussions about the plausible ranges of $\upsilon_{\rm avg}$ and $\lambda$. Then main results, including the expected number of detectable IBHs for the wide ranges of $\upsilon_{\rm avg}$ and $\lambda$, are presented in Section \[sec:results\] with prospects of future observations. After some discussion on the caveats in our work in Section \[sec:discussion\] we conclude in Section \[sec:conclusions\]. Formulations {#sec:formulations} ============ The Galactic Structure ---------------------- In a realistic picture within the framework of cosmological galaxy and structure formation driven by cold dark matter, the dynamical structure of our Galaxy should have evolved in time (e.g. @Amores17). However, it is difficult to construct a reliable model of accurate time evolution for our Galaxy, and here we use a simplified model in which the Galactic structure does not evolve, and is composed of three components: the central bulge, the Galactic disc, and a spherical dark halo surrounding them. First we introduce the model of the Galactic gravitational potential that we use, and then we present our model of IBH birth location and history. Here we use a cylindrical coordinate system of radius, azimuth, and height ($r,\theta,z$), with the $z$-direction perpendicular to the Galactic plane. We follow the gravitational potential model of @Irrgang13 (the Model II). For the gravitational potential of the spherical bulge and disc, the model proposed by @MN75 is assumed, which is described as $$\begin{aligned} \phi_i(r,z)=-\frac{GM_i}{\sqrt{r^2+\left(a_i+\sqrt{z^2+b_i^2}\right)^2}}. \label{eq:potential_bd}\end{aligned}$$ where $i=1,2$ represent the bulge and disc respectively. The spherical model of @Wilkinson99 is assumed for the dark halo potential: $$\begin{aligned} \phi_{\mathrm{halo}}=-\frac{GM_h}{R_h}\ln\left(\frac{\sqrt{R^2+R_h^2}+R_h}{R}\right), \label{eq:potential_h}\end{aligned}$$ where $R\equiv \sqrt{r^2+z^2}$. @Irrgang13 obtained the values of the constants $M_i,a_i,b_i,M_h,$ and $R_h$ by comparison with observational data, e.g. the Galactic rotation curve and the mass densities in the solar vicinity. These are: $$\begin{aligned} M_1&=4.07\times 10^{9}\ \mathrm{M_{\sun}},\ a_1=0,\ b_1=0.184\ \mathrm{kpc} \\ M_2&=6.58\times 10^{10}\ \mathrm{M_{\sun}},\ a_2=4.85\ \mathrm{kpc},\ b_2=0.305\ \mathrm{kpc}\\ M_h&=1.62\times 10^{12}\ \mathrm{M_{\sun}},\ R_h=200\ \mathrm{kpc}. \label{eq:value_Galaxy}\end{aligned}$$ We assume that the IBH birth rate per unit volume has an exponential radial profile, $\rho_{\rm IBHb} \propto \exp( - r / r_d )$ with the scale length $r_d = $ 2.15 kpc of the stellar disc [@Licquia15]. Along the disc height $\rho_{\rm IBHb}$ is assumed to be uniform in the region of $|z| < h$, where we adopt $h = 75$ pc which is the scale height of molecular clouds in the Galaxy (see Table \[tab:ISMtable\] in Section \[ssec:ISM\]). A spherical exponential profile is assumed for $\rho_{\rm IBHb}$ in the bulge, as $\rho_{\rm IBHb} \propto \exp(-R/R_b)$ and $R_b = $ 120 pc, following [@Sofue13], who found that the spherical exponential profile fits better to the observed data than the de Vaucouleurs law that is conventional as the profile of spheroidal galaxies. The total number of IBHs born in the past in the Galaxy is set to $N_{\rm IBH} = 1 \times 10^8$, and of course the final number of the observed IBHs found in this work scales with this parameter. Recent observations show that the Galactic bulge contains $\sim 15$ per cent of the total stellar mass in the Galaxy, while the remaining $\sim 85$ per cent are in the disc [@Licquia15]. Hence we determine the IBH birth rate in the disc and bulge so that the disc-to-bulge number ratio of the total IBH numbers is the same as that of stellar mass. This is a reasonable assumption provided that the initial mass function (IMF) does not depend on time or location throughout the Milky Way (for a discussion on IMF variability over cosmic history and within the Milky Way, see @Bastian10 and @Wegg17 respectively). The IBH birth rate in the disc is assumed to be constant with time from 10 Gyrs ago to now. On the other hand, the bulge is composed mostly by old stellar populations, which were presumably formed in the early stage of the history of the Galaxy. Quantitative star formation history of the bulge is still under debate (see @Nataf16 for a review), and here we assumed a simple history that bulge stars formed uniformly in a time period of $2$ Gyrs spanning from $10$ to $8$ Gyrs ago, which is within the range of uncertainty about the bulge star formation history. IBH Initial Conditions ---------------------- Here we describe the initial conditions of IBHs, namely the BH mass distribution and initial velocities. We adopt the IBH mass distribution of @Ozel10 obtained from observation of X-ray binaries: a normal distribution of average $7.8\ \mathrm{M_{\sun}}$ and standard deviation $1.2\ \mathrm{M_{\sun}}$. This distribution is consistent with a parallel Bayesian estimation by @Farr11. It is assumed that the mass change due to accretion is negligible. The initial velocity of an IBH is calculated as the sum of the velocity of the progenitor star and the kick velocity given at the time of the IBH formation. Velocities of IBH progenitors formed in the disc are assumed to follow the rotation velocity of the Milky Way consistent with the potential model of @Irrgang13, which is approximated as $$\begin{aligned} \upsilon_\phi= \begin{cases} 265-1875(r-0.2)^2&\mathrm{km\ s^{-1}}\ \ \ ($for$\ r<0.2)\\ 225+15.625(r-1.8)^2&\mathrm{km\ s^{-1}}\ \ \ ($for$\ 0.2<r<1.8)\\ 225+3.75(r-1.8)&\mathrm{km\ s^{-1}}\ \ \ ($for$\ 1.8<r<5.8)\\ 240&\mathrm{km\ s^{-1}} \ \ \ ($for$\ r>5.8) \end{cases} \label{eq:rotation_curve}\end{aligned}$$ where $r$ is measured in kpc. The motion of stars in the bulge is instead dominated by random motion rather than rotation. Thus we assume that the progenitors of IBHs in the bulge have a Maxwell-Boltzmann velocity distribution with the mean of $130\ \mathrm{km\ s^{-1}}$, which is consistent with velocity measurements of stars located near the Galactic Centre (hereafter GC) [@Kunder12]. The kick velocity distribution of BHs is hardly known. It is often supposed that a natal kick speed decreases with BH mass, as expected in the case of a fixed momentum. This implies a reduced BH kick speed compared to neutron stars [e.g., @Fryer12]. A study of 233 pulsars by @Hobbs05 concluded that neutron star kicks obey a Maxwell-Boltzmann distribution with 1D standard deviation $\sigma = 265\ \mathrm{km\ s^{-1}}$, which corresponds to an average 3D kick velocity of about $420\ \mathrm{km\ s^{-1}}$. If we simply extrapolate this result to black holes with conserved momentum, we obtain an average kick speed as low as $50\ \mathrm{km\ s^{-1}}$. However some studies [@Repetto12; @Repetto15] claim that this may not be the case, and the present locations of some X-ray binary systems require a natal kick broadly in the range of 100–500 km s$^{-1}$, which is comparable to neutron star kicks. @Janka13 proposed a possible theoretical explanation about this result. This is still a matter of debate, and here we assume a Maxwell-Boltzmann kick velocity distribution with a 3D average velocity $\upsilon_{\mathrm{avg}}$ in the range of 50–400 km/s, and we assume that the kick speed is not correlated with BH mass for simplicity. If the initial kick velocity is generated by the Monte-Carlo method obeying the Maxwell-Boltzmann distribution, the probability of getting a velocity much lower than the average is small, but such IBHs have a high chance of detection by higher Bondi accretion rate. Because of the limitation of computing time, the number of orbital calculations ($N_{\rm MC} \sim 10^6$) is much smaller than the actual IBH number $N_{\rm IBH} \sim 10^8$, and the small velocity IBHs are not well sampled by the Monte-Carlo generation. Therefore we set the grids of kick velocity for the orbital calculation that is more uniform than the Maxwell-Boltzmann distribution, and multiply the probability distribution of the kick velocity in the final output of the detectable number of IBHs (e.g., X-ray source counts). Equation of Motion {#ssec:EQM} ------------------ IBHs formed and kicked in our Galaxy will move following the Galactic gravitational potential. By simple calculation using eq. \[eq:potential\_bd\] and eq. \[eq:potential\_h\] we obtain the equations of motion as follows: $$\begin{aligned} \frac{dr}{dt}=&\upsilon_r\\ \frac{dz}{dt}=&\upsilon_z\\ \frac{d\upsilon_r}{dt}=&-\frac{\partial \Phi}{\partial r}+\frac{j_z^2}{r^3} \nonumber\\ =&\frac{j_z^2}{r^3}-\sum_{i=1,2}\frac{GM_ir}{\left\{r^2+[a_i+(z^2+b_i^2)^{1/2}]^2\right\}^{3/2}}\nonumber \\ &-\frac{GM_h}{R_h}\cdot\frac{r}{\sqrt{r^2+z^2}}\left[\frac{R_h}{\sqrt{r^2+z^2}\sqrt{r^2+z^2+R_h^2}} \right]\\ \frac{d\upsilon_z}{dt}=&-\frac{\partial\Phi}{\partial z} \nonumber \\ =&-\sum_{i=1,2}\frac{GM_iz[a_i+(z^2+b_i^2)^{1/2}]}{\left\{r^2+[a_i+(z^2+b_i^2)^{1/2}]^2\right\}^{3/2}\sqrt{z^2+b_i^2}}\nonumber \\ &-\frac{GM_h}{R_h}\cdot\frac{z}{\sqrt{r^2+z^2}}\left[\frac{R_h}{\sqrt{r^2+z^2}\sqrt{r^2+z^2+R_h^2}}\right]. \label{eq:eq_of_motion}\end{aligned}$$ We have used $\Phi\equiv \phi_1+\phi_2+\phi_{\mathrm{halo}}$ to denote the total Milky Way potential. Here $j_z\equiv r\upsilon_\theta$ is the z-axis specific angular momentum, and since the potential $\Phi$ is independent of $\theta$, $j_z$ will be conserved. Thus the rotational velocity $\upsilon_\theta$ can be obtained from the conservation of $j_z$, which greatly simplifies our calculation. The four equations are integrated using the 4th-order Runge-Kutta method, and as a result the present location and velocity of each IBH are obtained. Dynamical friction by stars or gas in molecular clouds [@Ostriker99; @MT05; @Inoue17] is not considered in our calculation of IBH orbits. These effects are larger for more massive black holes, but negligible for stellar mass black holes. Interstellar Gas {#ssec:ISM} ---------------- Once the present location and velocity of the IBHs are calculated, the next information needed for estimating the accretion rate is the profile of interstellar gas clouds in the Milky Way. We consider five ISM phases that differ by temperature and density [@BH00]. The densest are the molecular clouds composed mostly of $\mathrm{H_2}$, followed by the cold neutral medium mostly made up of cold H gas. These two types of gases are expected to be the regions where accretion becomes large enough to make IBHs observable. However the other three phases, the warm neutral medium (warm H[I]{}), warm ionized medium (warm H[II]{}), and hot ionized medium (hot H[ II]{}) occupy the majority of the volume. Table \[tab:ISMtable\] lists the ISM parameters around the solar neighbourhood adopted in this work. For the densest two phases (molecular clouds and cold H[I]{}) we assume that the probability distribution of gas particle density at a given point is described by a power law with an index $\beta$ in the range $n_1 < n < n_2$ [@Agol02]. Thus the volume filling fraction of gas with density $n$ to $n + dn$ is expressed as $(d\xi / dn) dn$, and $$\begin{aligned} \frac{d\xi (n)}{dn}=\frac{\beta-1}{n_1^{1-\beta}-n_2^{1-\beta}} \, \tilde{\xi}(r, z) n\, ^{-\beta} \ \ (n_1<n<n_2) \ , \label{eq:fillingfactor_vs_density}\end{aligned}$$ where $\tilde{\xi}(r, z)$ is the volume filling fraction integrated over $n$. In calculation of final IBH numbers, we set many bins in $n_1 < n < n_2$, and when an IBH is found in these ISM phases after the orbital calculation, X-ray luminosity is calculated for each $n$ bin. Then IBHs of the luminosity corresponding to varoius bins are summed up with the probability distribution $d\xi/dn$ multiplied. On the other hand, the other three phases (warm H[I]{}, warm and hot H[II]{}) are represented by a single gas density from @BH00. The mid-plane volume filling fraction around the solar neighbourhood [@BH00] and the disc scale heights [@Agol02] of these five phases are shown as $\tilde{\xi}_0^{\rm BR}$ and $H_d$, respectively, where the subscript 0 indicates the value at the solar neighbourhood. Effective sound velocity, which includes turbulent velocity that becomes dominant in cold phases, is also necessary to estimate the Bondi accretion rate. This is set to $c_s=3.7(n/100\ {\rm cm}^{-3})^{-0.35}$ km s$^{-1}$ for molecular clouds [@MT05] from the observational results of turbulent velocity by @Larson81, $c_s=150$ km s$^{-1}$ for the hot H[ II]{} phase, and $10$ km s$^{-1}$ for the other three phases [@Ioka17]. It is assumed that the gas densities of each ISM phase are constant, but volume filling fractions depend on the location in the Galaxy. The filling fractions of molecular clouds, cold and warm H[I]{} phases are determined to match the surface density profile $\Sigma(r)$ of $\mathrm{H_2}$ and H[I]{} recently obtained by @NS16. We assume that the gas distribution along the height from the disc plane is uniform in the region of $|z| < H_d$, and $H_d$ is constant in the Galaxy. Hence the filling fraction can be calculated from the surface density as $$\begin{aligned} \tilde{\xi}(r, z) = \frac{\Sigma(r)}{2H_d\mu} \frac{1}{n} \ \label{eq:fillingfactor}\end{aligned}$$ when $|z| < H_d$ but zero otherwise, where $\mu=2.72m_p$ and $1.36m_p$ for molecular and atomic gas clouds respectively, and $m_p$ is the proton mass. The gas density $n$ of warm H[I]{} is simply given in Table \[tab:ISMtable\], but that for molecular clouds and cold H[I]{} should be replaced by the mean density of the power-law distribution, $$\begin{aligned} \langle n \rangle = \frac{1}{\tilde{\xi}} \int_{n_1}^{n_2} n \frac{\partial \xi}{\partial n} dn = \frac{\beta-1}{\beta-2} \frac{n_1^{2-\beta}-n_2^{2-\beta}}{n_1^{1-\beta}-n_2^{1-\beta}} \ . \label{eq:fillingfactor}\end{aligned}$$ Observed $\Sigma(r)$ of H[I]{} includes both cold and warm H[I]{}, and we assume that the relative proportion of these two phases is 3.1:3.5 and constant throughout the Galaxy, which is calculated by $\Sigma_0 = 2 H_d \mu \tilde{\xi}_0^{\rm BR} n$ at the solar neighbourhood with the parameters given in Table \[tab:ISMtable\]. Then $\tilde{\xi}(r, z)$ for the three phases of $\mathrm{H_2}$ and H[I]{} has been determined throughout the Galaxy. It should be noted that the filling fraction may become larger than the unity depending on $\Sigma(r)$ in this formulation, but we confirmed that the total of $\tilde{\xi}$ for these three phases is less than one everywhere in the Galaxy with the observed values of $\Sigma(r)$. Then the filling fractions $\tilde{\xi}(r, z)$ of the other two phases (warm and hot H[II]{}) are determined as follows. At a given location, the total filling fraction $\tilde{\xi}_{\rm H2+HI}$ of the three phases ($\mathrm{H_2}$ and cold/warm H[I]{}) is calculated. (Note that some of the three phases do not exist depending on the height $z$.) Then the remaining filling fraction, $1 - \tilde{\xi}_{\rm H2+HI}$, is distributed into two H[II]{} phases when $|z| \le 1$ kpc, assuming that the ratio of $\tilde{\xi}_0^{\rm BR}$ for the solar neighbourhood is constant throughout the Galaxy. In regions of $1 < |z| \le 3$ kpc the fraction $1 - \tilde{\xi}_{\rm H2+HI}$ is filled only by hot H[II]{}. The filling fractions thus determined in this work are different from $\tilde{\xi}_0^{\rm BR}$ even at a mid-plane point of the Sun’s Galactocentric distance ($r = R_0 = $ 8.3 kpc, @Gillessen09 [@Russeil17]; this is consistent with @Irrgang13 model II as well), which are also shown in Table \[tab:ISMtable\]. Phase $n_1\mathrm{[cm^{-3}]}$ $n_2\mathrm{[cm^{-3}]}$ $\beta$ $\tilde{\xi}_0^{\rm BR}$ $H_d$ $c_s[\mathrm{km\ s^{-1}}]$ $\Sigma_0 [\mathrm{M_{\sun}\ pc^{-2}}]$ $\tilde{\xi}(r=8.3{\rm \ kpc})$ ------------------ ------------------------- ------------------------- --------- -------------------------- ---------- ------------------------------------- ----------------------------------------- --------------------------------- Molecular clouds $10^2$ $10^5$ $2.8$ 0.001 $75$ pc $3.7(n/100\ {\rm cm}^{-3})^{-0.35}$ $2.3$ $0.0004$ Cold H[I]{} $10^1$ $10^2$ $3.8$ 0.02 $150$ pc $10$ $3.1$ $0.026$ Warm H[I]{} – 0.35 $500$ pc $10$ $3.5$ $0.46$ Warm H[II]{} – 0.20 $1$ kpc $10$ $2.0$ $0.16$ Hot H[II]{} – 0.43 $3$ kpc $150$ $0.17$ $0.37 $ Mass Accretion from ISM ----------------------- There are studies that claim the accretion onto compact objects could be much less than the Bondi accretion rate, due to material outflow in the process of accretion. These claims are supported by theoretical modelings of accretion flows including outflow [@BB99] and hydrodynamical and MHD simulations (see @Perna03). These studies suggest that the ratio of the actual accretion to the Bondi accretion, $\lambda$, scales as $(R_{\mathrm{in}}/R_\mathrm{out})^p$, where $R_{\mathrm{in}}$ and $R_\mathrm{out}$ are the inner and outer radius of the accretion flow respectively, and index $p$ being an uncertain number around $0.5-1$ [@Yuan14]. The inner radius $R_{\mathrm{in}}$ is generally expected to be about a few to few tens times the Schwartzschild radius, but $R_{\mathrm{out}}$ should be dependent on the angular momentum at the Bondi radius. If angular momentum is sufficiently large to make the accretion flow rotationally supported at the Bondi radius, $R_{\mathrm{out}}$ will be comparable to the Bondi radius. This is for the case of Sgr A\*, as assumed by several authors [@Yuan03; @Totani06], and in this case $R_{\mathrm{in}}/R_\mathrm{out}$ will be extremely small, down to $10^{-8}-10^{-9}$. If we simply adopt this and use $p=0.5-1$, we get $\lambda$ no higher than $10^{-4}$. However for Sgr A\* a smaller index $p=0.27$ is preferred from fit to observations [@Yuan03], which gives $\lambda \sim 0.01$. This agrees with the observation of nearby active galaxies by @Pellegrini05, who estimated $\lambda$ to be around $0.01$. We do not know, however, whether this can be applicable to stellar-mass black holes, since their Bondi radii are very different in scale from supermassive black holes. The study by @Perna03 discusses the case for accretion onto INSs, which concluded that $\lambda \lesssim 10^{-3}$ is consistent with the null detection of accreting INSs by [ *ROSAT*]{}. However, neutron stars have magnetic fields and hard surfaces which significantly affect the accretion rate [e.g., @Toropina12]. Thus we cannot simply assume that IBHs would follow this constraint. Some studies [e.g., @Fujita98; @Agol02; @Ioka17; @Matsumoto17; @Inoue17] used observations of the density [@Armstrong95] or velocity fluctuations [@Larson81] of the interstellar medium, deriving that $R_{\mathrm{out}}$ is much smaller than the Bondi radius. They obtain $R_{\mathrm{out}}\sim 10^5 R_{s}$, which gives a range $\lambda=10^{-4}$–$10^{-2}$ for $p=0.5$–$1$. However the observational results for ISM density fluctuations include significant uncertainties, and, more importantly, the observation by @Armstrong95 targets ionized hot gas. Thus we cannot apply this relation in the case when IBHs accrete neutral molecular gas, which is the most observable case. Estimates based only on ISM velocity fluctuations would be a lower limit of the initial angular momentum, because it would increase by density fluctuations and the velocity of the black hole. The IBH velocity is typically much larger than the turbulent velocity of the interstellar medium. To summarize, there is a large uncertainty about the accretion efficiency $\lambda$ both theoretically and observationally. Therefore here we simply test $\lambda$ in the range of $10^{-3}$–$10^{-1}$. Although this may be rather optimistic, we adopt this because later we will find in our calculations that IBHs may be detectable by future surveys only when $\lambda\gtrsim 0.01$. IBH Luminosity and Flux {#ssec:lum_flux} ----------------------- We estimate the bolometric luminosity $L$ from the Bondi-Hoyle accretion rate with the $\lambda$-factor, as $$\begin{aligned} \dot{M}=&\lambda \cdot 4\pi\frac{(GM)^2\rho}{(\upsilon^2+c_s^2)^{3/2}}\nonumber \\ \approx& 3.7\times 10^{15}\mathrm{g\ s^{-1}}\nonumber \\ &\cdot \left(\frac{\lambda}{0.1}\right)\left(\frac{M}{10\ \mathrm{M_{\sun}}}\right)^2\left(\frac{\rho}{10^3\ \mathrm{cm^{-3}}\ m_p}\right)\left[\frac{\upsilon^2+c_s^2}{(10\ \mathrm{km\ s^{-1}})^2}\right]^{-3/2} \ , \label{eq:accretion_rate}\end{aligned}$$ where $G$ is the gravitational constant, $M$ the BH mass, $\rho$ the gas mass density, $m_p$ the proton mass, $\upsilon$ the speed of the BH relative to the interstellar gas, and $c_s$ the effective sound speed taken from Table \[tab:ISMtable\]. To obtain the luminosity we apply the treatment of @MT05, which takes into account the transition from the standard disc to the RIAF (radiatively-inefficient accretion flow) mode in low accretion rate regime. At high accretion rates the BH accretion is described with the standard disc model, where the radiation efficiency $\eta \equiv L / (\dot M c^2)$ is constant and the luminosity is proportional to the accretion rate. However when the accretion rate drops below a threshold, the disc would switch to the RIAF phase and $\eta$ becomes proportional to the accretion rate, making the luminosity proportional to the square of the accretion rate [@NY95; @KFM08]. The threshold is expected to be around 1/10 of the Eddington accretion rate, and hence $$\begin{aligned} \dot{M}_{\rm th} = \epsilon_{\rm th} \dot{M}_{\mathrm{Edd}} =1.4\times 10^{18}\mathrm{g\ s^{-1}} \left(\frac{M}{10\ \mathrm{M_{\sun}}}\right) \left( \frac{\epsilon_{\rm th}}{0.1} \right) \left(\frac{\eta_{\mathrm{std}}}{0.1}\right)^{-1}, \label{eq:accretion_threshold}\end{aligned}$$ where $\dot M_{\rm Edd} \equiv L_{\rm Edd} / (\eta_{\rm std} c^2)$ is the Eddington accretion rate corresponding to the Eddington luminosity, and $\eta_{\rm std}$ is the radiation efficiency in the standard disc regime. Then requiring that $\eta$ changes continuously around the threshold, we model $$\begin{aligned} \eta= \begin{cases} \eta_{\rm std} (\dot{M}/\dot{M}_{\rm th}) &(\mathrm{when}\ \dot{M}< \dot{M}_{\rm th}) \\ \eta_{\rm std} &(\mathrm{when}\ \dot{M}_{\mathrm{th}} < \dot{M} < 2 \dot{M}_{\rm Edd}). \end{cases} \label{eq:eta}\end{aligned}$$ We adopt the standard values of $\epsilon_{\rm th} = 0.1$ and $\eta_{\rm std} = 0.1$ in all of our calculations in this work, and uncertainties about these parameters are discussed in in Section \[sec:discussion\]. The bolometric luminosity is then calculated as $$\begin{aligned} L = & \eta \dot{M}c^2\nonumber \\ = & 3.4 \times 10^{37} \mathrm{erg\ s^{-1}}\nonumber \\ \cdot & \eta \lambda \left(\frac{M}{10\ \mathrm{M_{\sun}}}\right)^2\left(\frac{\rho}{10^3\ \mathrm{cm^{-3}}\ m_p}\right)\left[\frac{\upsilon^2+c_s^2}{(10\ \mathrm{km\ s^{-1}})^2}\right]^{-3/2} \ , \end{aligned}$$ which becomes $$\begin{aligned} L = & 9.0 \times 10^{32} \mathrm{erg\ s^{-1}}\nonumber\\ \cdot & \left(\frac{\lambda}{0.1}\right)^2 \left(\frac{M}{10\ \mathrm{M_{\sun}}}\right)^3\left(\frac{\rho}{10^3\ \mathrm{cm^{-3}}\ m_p}\right)^2\left[\frac{\upsilon^2+c_s^2}{(10\ \mathrm{km\ s^{-1}})^2}\right]^{-3} \label{eq:Xray_luminosity_RIAF}\end{aligned}$$ in the RIAF regime and $$\begin{aligned} L = & 3.4 \times 10^{35} \mathrm{erg\ s^{-1}}\nonumber\\ \cdot & \left(\frac{\lambda}{0.1}\right) \left(\frac{M}{10\ \mathrm{M_{\sun}}}\right)^2\left(\frac{\rho}{10^3\ \mathrm{cm^{-3}}\ m_p}\right)\left[\frac{\upsilon^2+c_s^2}{(10\ \mathrm{km\ s^{-1}})^2}\right]^{-3/2} \label{eq:Xray_luminosity_standard_disk}\end{aligned}$$ in the standard disc regime. When the accretion rate largely exceeds the Eddington limit, the accretion flow would be described by the slim disc rather than the standard disc. In this regime we adopt the formula by @Watarai00: $$\begin{aligned} L = 2 L_{\rm Edd} \left[1+\ln\left(\frac{\dot{M}/\dot{M}_{\rm Edd}}{2}\right)\right] \label{eq:Xray_luminosity_Watarai}\end{aligned}$$ for $\dot{M} > 2 \dot{M}_{\rm Edd}$, which is smoothly connected to the standard disc regime when $\eta_{\rm std} = 0.1$ is assumed. IBH Spectrum {#ssec:IBH_spectrum} ------------ Although the spectrum of IBHs is essentially unknown, past studies [e.g., @Agol02; @Fender13] assumed that their characteristics are similar to those observed from BH binaries. The spectrum of BH binaries is divided into two categories depending on the accretion rate: the soft state and the hard state [@KFM08]. BH binaries are considered to show the soft state spectrum in the standard disk regime (i.e. when the accretion rate is high), and the radiation spectrum is dominated in X-rays by a multi-temperature black body radiation from an optically-thin accretion disk. The hard state is typical for low-accreting BHs in the RIAF regime, and the radiation is described as a power law with a photon index of $\zeta = 1.4$–$2.1$ [@RM06], where the differential photon spectrum is $dF_{\rm ph}/d\epsilon_{\rm ph} \propto \epsilon_{\rm ph}^{-\zeta}$. Due to the low density of the interstellar matter, we expect that a majority of IBHs are in the RIAF regime. Therefore, following @Fender13, we assume that the IBH spectrum is a power-law with $\zeta = 1.6$. We will see in Section \[ssec:Flux\_and\_Properties\] that most of the detectable IBHs are indeed in the RIAF regime. We assess quantitatively in Section \[sec:discussion\] how our results would change by varying the photon index. We also assume that the bolometric luminosity $L$ is dominantly radiated in the X-ray band of 0.1–100 keV, and then the fraction $f_{\rm band}$ of the luminosity in the observed band is $f_{\rm band} = 0.31$ for the [*NuSTAR*]{} and [*FORCE*]{} bands ($10$ – $40$ keV), and $f_{\rm band} = 0.17$ for the [*ROSAT*]{} band ($0.1$ – $2.4$ keV). Absorptions ----------- Furthermore we introduce two parameters, $f_{\rm MC}$ and $f_{\rm MW}$, to take into account the photoelectric absorption of X-rays. When an IBH is accreting in a molecular cloud, the flux is reduced by a factor of $f_{\rm MC}$ by absorption within the cloud. We calculate $f_{\rm MC}$ assuming a hydrogen column density of $N_H=5\times 10^{21} {\rm \ cm^{-2}}$, which is calculated from the density-size relation of molecular clouds [@Larson81]. According to this relation, $N_H$ is not sensitive to the size of molecular clouds, though there exists a scatter from the mean relation by up to an order of magnitude. We calculate using the model of @Wilms00 that there is a significant photoelectric absorption of $f_{\rm MC} \sim 0.3$ in the [*ROSAT*]{} band, but $f_{\rm MC} \sim 1$ for other satellites. X-rays are absorbed also by ISM in the Galaxy along the line of sight to the observer by a factor of $f_{\rm MW}$. The value of $f_{\rm MW}$ depends on the location of an IBH, and if it is located in GC, X-rays in the [*ROSAT*]{} band are seriously absorbed by $f_{\rm MW} \sim 0.01$ with a large column density of $N_H \sim 6\times 10^{22}{\rm cm^{-2}}$ [@Baganoff03; @Muno09]. On the other hand, the absorption is negligible in the hard X-ray [*NuSTAR*]{} and [ *FORCE*]{} band. These are taken into account when our results are compared with observational constraints. Finally, we assume that the emission is isotropic, and flux measured on the Earth is calculated assuming that the Sun is located in the mid-plane (i.e., $z = 0$) at the Galactocentric distance of $R_0 = 8.3$ kpc. Results {#sec:results} ======= Distribution of IBHs in the Milky Way {#ssec:distribution_result} ------------------------------------- For a given set of model parameters, we generate typically $N_{\rm MC} = 10^6$–$5\times 10^6$ IBHs in our Galaxy with physical quantities obeying the distributions described in previous sections by the Monte Carlo method. The number of generated IBHs is larger for larger $\upsilon_{\mathrm{avg}}$ because of its higher probability of escaping from the Galaxy potential. Then the final estimate of detectable IBHs will be scaled to match the real number of IBHs in the Galaxy, $N_{\rm IBH}$. The present spatial distribution of IBHs in the Galaxy, after orbital calculations described in Section \[ssec:EQM\], is shown in Fig. \[fig:IBHdistribution\] for two values of average kick velocity, $\upsilon_{\mathrm{avg}}=50$ and $400$ $\mathrm{km\ s^{-1}}$. We find an obvious trend that IBH distribution becomes more extended from the GC with increasing $\upsilon_{\rm avg}$. Note that a portion of IBHs have positive total (kinetic plus potential) energy and eventually escape the Galaxy. The fraction is negligibly small for $\upsilon_{\mathrm{avg}}=50\ \mathrm{km\ s^{-1}}$, whereas it increases to 0.01, 3, 17, and 37 per cent for $\upsilon_{\mathrm{avg}}=$ $100, 200, 300$ and $400\ \mathrm{km\ s^{-1}}$, respectively. [cc]{} {width="1.0\linewidth"} {width="1.0\linewidth"} Fig. \[fig:Comparison\_of\_IBHdistribution\] shows the surface number density of IBHs on the Galactic plane, in comparison with the uniform (i.e. constant surface density) distribution and an exponential distribution with the scale length of 2.15 kpc, which was adopted for the initial IBH distribution at their birth. It can be seen that the distribution after orbital evolution becomes more extended than the initial exponential shape, and this effect becomes stronger with larger kick velocity. It should also be noted that there is an excess of IBHs near the GC within 1 kpc, which is due to the contribution from the Galactic bulge component. ![The surface number density of IBHs on the Galactic plane. Histograms are for present-day IBHs after orbital evolution from their birth, with two different values of average kick velocity. The total number of IBHs is normalized to $1\times 10^8$. For comparison, uniform and exponential (with a scale 2.15 kpc) distributions are also shown.[]{data-label="fig:Comparison_of_IBHdistribution"}](distribution_comparison.eps){width="1.0\linewidth"} X-ray Source Counts {#ssec:Flux_and_Properties} ------------------- Here we present X-ray source counts, i.e., number of IBHs as a function of X-ray flux. According to the modelling presented in the previous section, we can assign an X-ray flux for each of the IBHs whose location and velocity have been calculated by time integration of their orbits. Figure \[fig:histogram\_curve\] shows the cumulative X-ray source counts into the direction of the GC, for a few different values of the average kick velocity $\upsilon_{\mathrm{avg}}$ and accretion efficiency $\lambda$. Here calculation is for the hard X-ray band where absorptions are negligible, i.e., $f_{\rm MC} = f_{\rm MW} = 1$. Figure \[fig:histogram\_each\_phase\] is the same but shows the contribution from each ISM phase in the case of the GC direction. Figures \[fig:histogram\_curve\_allsky\_minabs\] and \[fig:histogram\_curve\_allsky\_maxabs\] are the same as Figure \[fig:histogram\_curve\] but for all sky. To compare with the [*ROSAT*]{} result, we assumed $f_{\rm MC} = 0.3$ for the absorption in molecular clouds in the [*ROSAT*]{} band. The absorption in ISM depends on IBH locations, and here we show two extreme cases: no absorption ($f_{\rm MW}=1$) and absorption towards GC ($f_{\rm MW}=10^{-2}$) in Figures \[fig:histogram\_curve\_allsky\_minabs\] and \[fig:histogram\_curve\_allsky\_maxabs\], respectively. It should be noted that the number of IBHs generated by Monte Carlo ($N_{\rm MC}$) is 1–2 orders of magnitude smaller than the actual number of IBHs in the Galaxy, $N_{\rm IBH}$, because of the limited computing time. The results shown in these figures are scaled up to match the actual number of $N_{\rm IBH}$. One may consider that in this case our calculation cannot resolve a population of IBHs whose number is smaller than $N_{\rm IBH}/N_{\rm MC}$ in the Galaxy. However the results shown in Figures \[fig:histogram\_curve\_allsky\_minabs\] and \[fig:histogram\_curve\_allsky\_maxabs\] well extend to the region of small number ($\ll 1$) in all sky. This is because we consider the weight of probability distribution about kick velocity and gas density in molecular clouds and cold H[I]{} ISM. The difference of $N_{\rm MC}$ and $N_{\rm IBH}$ may also change the distance to the nearest IBH from the Sun, but it is not important because the nearest IBHs are not the major component in the detectable IBHs (see Section \[ssec:comparison\], Fig. \[fig:distance\_from\_sun\]). ![Cumulative X-ray source counts of IBHs with X-ray flux greater than $F_X$, inside the 0.5 deg$^2$ region towards the GC direction. A hard X-ray energy band of $10$ – $40$ keV is assumed, and in this energy band the absorptions in molecular clouds or ISM are negligible (i.e. $f_{\rm MC}=f_{\rm MW}=1$). Six curves are shown for the combination of $\upsilon_{\mathrm{avg}}=50$ or $400 \ \mathrm{km\ s^{-1}}$, and $\lambda$ = $0.1, 0.01, 0.001$, as indicated in the Figure. Conservative upper limits from the observations by [*Chandra*]{} ($0.5$–$8$ keV), [*XMM-Newton*]{} ($2$–$12$ keV), and [*NuSTAR*]{} ($10$–$40$ keV) are shown, by requiring that the number of IBHs should not exceed that of all X-ray sources detected toward the GC. The vertical line shows the expected sensitivity of a proposed mission [*FORCE*]{} in $10$–$40$ keV energy band. []{data-label="fig:histogram_curve"}](flux_vs_number_GC.eps){width="1.0\linewidth"} ![Same as Fig. \[fig:histogram\_curve\], but broken into contributions from each ISM phase. An average kick velocity of $\upsilon_{\rm avg} = 50\ \mathrm{km\ s^{-1}}$ and $\lambda=0.1$ are assumed for this figure. It should be noted that the contribution from IBHs in the hot H[II]{} ISM phase is too faint to appear in this plot.[]{data-label="fig:histogram_each_phase"}](histogram_each_phase.eps){width="1.0\linewidth"} ![Same as Fig. \[fig:histogram\_curve\], but for IBHs in all sky observed in the [*ROSAT*]{} 0.1 – 2.4 keV band. Here we assume the absorption correction as $f_{\rm MC}=0.3$ and $f_{\rm MW}=1$.[]{data-label="fig:histogram_curve_allsky_minabs"}](flux_vs_number_allsky_minabs.eps){width="1.0\linewidth"} ![ Same as Fig. \[fig:histogram\_curve\_allsky\_minabs\], but $f_{\rm MW}=0.01$ is assumed.[]{data-label="fig:histogram_curve_allsky_maxabs"}](flux_vs_number_allsky_maxabs.eps){width="1.0\linewidth"} Comparison with Observations and Future Detectability {#ssec:comparison} ----------------------------------------------------- As a comparison with observations, we first consider the [*ROSAT*]{} all-sky survey [@ROSAT]. Fig. \[fig:histogram\_curve\_allsky\_minabs\] implies that many IBHs may be detected by the [*ROSAT*]{} sensitivity of $\sim 1\times 10^{-12} {\rm erg/s/cm^2}$ in all sky with optimistic model parameters, but it should be noted that the flux plotted in this figure is not corrected for absorption by intervening ISM, and evidently it is more difficult to detect if maximal absorption is assumed (see Fig. \[fig:histogram\_curve\_allsky\_maxabs\]). Fig. \[fig:distance\_from\_sun\] shows the distribution of distance from the Sun to IBHs detectable by the [*ROSAT*]{} all sky survey. It can be seen that IBHs towards GC are dominated by those located around GC. This is not only because IBHs are concentrated around GC, but also because the abundance of molecular clouds around GC is high. This result tells us that considering only the Solar vicinity is not sufficient for estimating the detectability of IBHs, but it is necessary to take the entire Galaxy, especially the GC region, into consideration. Soft X-ray flux in the [*ROSAT*]{} band is seriously reduced by a factor of $f_{\rm MW} \sim 0.01$ due to absorption by ISM along the sightline to GC, and hence it is difficult to derive a strong constraint on IBH parameters even if no IBH is included in the [*ROSAT*]{} catalog. For the GC direction, past surveys by [*Chandra*]{} (flux limit $5 \times 10^{-14} \ \rm erg \ s \ cm^{-2}$ in $0.5$–$8$ keV, survey area 1.6 deg$^2$), [*XMM-Newton*]{} ($2.5 \times 10^{-14} \ \rm erg \ s \ cm^{-2}$ in $2$–$10$ keV, 22.5 deg$^2$), and [*NuSTAR*]{} ($6 \times 10^{-13} \ \rm erg \ s \ cm^{-2}$ in $10$–$40$ keV, 0.6 deg$^2$) have detected 9017, 2204, and 70 sources, respectively [@Muno09; @Warwick11; @Hong16]. Here, the flux limits for [ *Chandra*]{} and [*XMM-Newton*]{} are those corrected for absorption by ISM, as given in the references. The correction factor is $\sim 3$ for [*Chandra*]{}, but it is negligible in the [*NuSTAR*]{} band. The expected number of IBHs with a parameter set of $\lambda = 0.1$ and $\upsilon_{\rm avg}$ = 50 km/s are 36, $2.4\times 10^{2}$, and 4, respectively (when $f_{\rm band}=0.3$ for [*Chandra*]{} and [ *XMM-Newton*]{} bands are assumed). Though the number expected for [*XMM-Newto*]{}n largely exceeds unity, it is difficult to discriminate IBHs for other populations of X-ray sources. Therefore we conservatively set upper limits that the number of IBHs cannot exceed those of all detected sources in these surveys. These upper limits are shown in Fig. \[fig:histogram\_curve\], which do not give a strong constraint on IBH parameters. As mentioned in the Introduction, hard X-ray band may be useful to discriminate IBHs from other X-ray populations. However, the expected number of IBHs detectable by the past [*NuSTAR*]{} survey is at most of order unity. Therefore we consider a survey towards GC by the proposed mission [*FORCE*]{}. A sensitivity limit of about $1\times 10^{-14}\ \mathrm{erg\ s^{-1} cm^{-2}}$ in the $10$–$40$ keV can be achieved by 100 ksec observation for each field-of-view, by the improved angular resolution compared with [*NuSTAR*]{}. The sensitivity flux limit is indicated in Fig. \[fig:histogram\_curve\]. A total survey area of about 1 deg$^2$ is possible with a realistic telescope time. We show in Fig. \[fig:detection\_number\] the expected number of IBH detections by a survey by [*FORCE*]{} as a function of $\upsilon_{\rm avg}$ with two values of $\lambda$, assuming a total survey area of 0.5 and 1.5 deg$^2$. The expected number becomes much larger than unity at an optimistic parameter region of $\lambda = 0.1$ and $\upsilon_{\rm avg} \lesssim 100$ km s$^{-1}$, and hence no detection of an IBH by this survey can exclude this parameter region, provided that discrimination of IBHs from other X-ray source populations is successful. ![Distribution of distance from the Sun to IBHs detectable by the [*ROSAT*]{} all-sky survey or by the future [*FORCE*]{} survey towards GC of survey area $0.5$ deg$^2$. Here $\upsilon_{\rm avg}$ and $\lambda$ are set to $50\ \mathrm{km\ s^{-1}}$ and $0.1$, respectively. For [*ROSAT*]{} the absorption factors of $f_{\rm MC}=0.3$ and $f_{\rm MW}=1$ are used like Fig. \[fig:histogram\_curve\_allsky\_minabs\], but the histogram in reality can be significantly modified by adopting location-dependent $f_{\rm MW}$.[]{data-label="fig:distance_from_sun"}](hist_distance.eps){width="1.0\linewidth"} ![The expected number of IBH detections by a survey by [ *FORCE*]{} towards GC, as a function of mean kick velocity $\upsilon_{\rm avg}$ for two values of $\lambda$. The sensitivity is assumed to be $1\times 10^{-14}\ \mathrm{erg\ s^{-1} cm^{-2}}$ in the $10$–$40$ keV, and two different survey area (0.5 and 1.5 deg$^2$) are considered, as shown by the solid and dashed curves. For comparison, the expected number by the past [*NuSTAR*]{} survey [@Hong16] is also shown.[]{data-label="fig:detection_number"}](detection_count_final.eps){width="1.0\linewidth"} Interpretation of Source Counts {#ssec:interpretation} ------------------------------- The shape of the X-ray source counts in Fig. \[fig:histogram\_curve\] can be understood as follows. The shape of $\lambda = 0.1$ curves is a broken power-law with the break flux of $\sim 3 \times 10^{-14}\ \mathrm{erg\ s^{-1} cm^{-2}}$. The power index is $\alpha\sim 0.5$ and $0.7$ at the lower and higher flux regimes, respectively, where $N(>F_X)\propto F_X^{-\alpha}$. As seen in Fig. \[fig:histogram\_each\_phase\], source counts are dominated by IBHs in molecular clouds. When velocity of an IBH is less than $c_s$, accretion rate and luminosity are determined by gas density $n$ with a weak dependence on $\upsilon$ (see eq. \[eq:Xray\_luminosity\_RIAF\]). Since the probability distribution of gas density is a power-law, $d\xi/dn\propto n^{-2.8}$, the probability of finding a density larger than $n$ is $P(>n) \propto n^{-1.8}$. In addition, the number of black holes that are lower than $c_s (\ll v_{\rm avg})$ scales as $N(<c_s)\propto c_s^3 \propto n^{-1.05}$ due to the assumed Maxwell-Boltzmann distribution of kick velocity. In this regime the IBH luminosity is approximately proportional to $n^2 c_s^{-6} \propto n^{4.1}$ by eq. (\[eq:Xray\_luminosity\_RIAF\]), and we obtain $\alpha\sim (1.8+1.05)/4.1 \sim 0.7$, which is consistent with the slope observed at fluxes above the break. The break flux should correspond to that for the minimum density of molecular clouds, $n_1 = 10^2\ \mathrm{cm^{-3}}$, with $\upsilon\sim c_s= 3.7\ \mathrm{km \ s^{-1}}$. Using these parameters, we find $\sim 3 \times 10^{-14}(\lambda/0.1)^2\ \mathrm{erg\ s^{-1} cm^{-2}}$ at GC for the mean BH mass of $M=7.8\ \mathrm{M_{\sun}}$, which is consistent with the break flux found in the figure. On the other hand, the sharp drop of the counts at the bright end of $F_X \sim 10^{-8}\ \mathrm{erg\ s^{-1}\ cm^{-2}}$ common for all curves corresponds to the transition luminosity, where the RIAF regime switches into the standard disc regime. The luminosity here is equivalent to $0.1$ times the Eddington luminosity, which gives $F_X \sim 10^{-8}\ \mathrm{erg\ s^{-1}\ cm^{-2}}$ when an IBH at GC is assumed. The source counts below the break flux is dominated by IBHs having velocities larger than $c_s$ and hence fainter flux. When $\upsilon > c_s$ but still $\upsilon < \upsilon_\mathrm{avg}$, The number of IBHs with a velocity lower than $\upsilon$ roughly scales as $N(<\upsilon)\propto \upsilon^3$. Therefore $\alpha\sim 0.5$ should be found in this range, which is also consistent with the curves in the figure. The source counts in all sky (Figures \[fig:histogram\_curve\_allsky\_minabs\] and \[fig:histogram\_curve\_allsky\_maxabs\]) show similar breaks and drops to those found in the counts towards GC. However they appear more smoothed out because distances of IBHs from the Sun are more widely distributed with less concentration to GC, as seen in Fig. \[fig:distance\_from\_sun\]. Properties of Detectable IBHs ----------------------------- Fig. \[fig:physical\_properties\] shows the distribution of IBH velocities $\upsilon$ (with respect to the frame of the Galactic rotation) and IBH masses, for IBHs detectable by the future [ *FORCE*]{} survey towards GC. As expected from eq. \[eq:Xray\_luminosity\_RIAF\], detectable IBHs are dominated by those with low velocities. The velocity distribution becomes flatter when $\upsilon$ becomes close to the effective sound speed of the molecular gas ($c_s \lesssim 10\ \mathrm{km\ s^{-1}}$). On the other hand, the observable IBH mass distribution is almost the same as that of the entire IBH population, though the peak of the former is slightly shifted to larger mass. This is because IBH masses are narrowly distributed, although X-ray luminosity is proportional to the cube of BH mass in the RIAF regime (eq. \[eq:Xray\_luminosity\_RIAF\]). It is interesting to see the relative contributions to detectable IBHs from those formed in the Galactic disc or bulge. This is shown as the expected number of IBHs detectable by the future [*FORCE*]{} survey, for several values of $\upsilon_{\rm avg}$ and assuming $\lambda = 0.1$, in Table \[tab:bulge\_or\_disk\_0.1\]. We see that the disc fraction becomes smaller as $\upsilon_{\rm avg}$ increases, because IBHs in the disc region are more efficiently expelled by a large kick velocity from dense gas regions than those in the bulge. Since the expected number of bulge IBHs is always much less than unity, IBHs detectable by [*FORCE*]{} would be mostly of the disc origin, though they are located around GC. [cc]{} {width="1.0\linewidth"} {width="1.0\linewidth"} ------- --------- --------------------- --------- --------------------- --------------------- $50$ $100$ $200$ $300$ $400$ bulge $0.018$ $1.4\times 10^{-3}$ $0.067$ $3.7\times 10^{-3}$ $9.1\times 10^{-3}$ disc $32$ $4.5$ $0.79$ $0.19$ $0.086$ ------- --------- --------------------- --------- --------------------- --------------------- : Expected number of IBHs detectable by the future [*FORCE*]{} survey, showing separately those born in the Galatctic disc and bulge. For all calculations $\lambda=0.1$ is assumed. []{data-label="tab:bulge_or_disk_0.1"} Discussion {#sec:discussion} ========== In this section we mention some caveats and important notes of our work. Parameters for Radiative Efficiency of Accretion Flow ----------------------------------------------------- The parameter $\epsilon_{\rm th}$ in Section \[ssec:lum\_flux\] is sensitive to the viscosity parameter $\alpha$. [@NY95] have shown that $\epsilon_{\rm th}$ can become approximately $10^{-3}$ – $10^{-1}$ with $\alpha=0.03$–$0.3$. When $\epsilon_{\rm th}$ smaller than our value $10^{-1}$ is assumed, the radiation efficiencies of IBHs whose accretion rates are smaller than the threshold would be larger. This increases the luminosities of IBHs in the RIAF regime, thus significantly increasing the number of detections. For example, by changing $\epsilon_{\rm th}$ to $10^{-3}$ we find that the number of IBHs detectable by [*FORCE*]{} is increased to $2.8\times 10^{2}$ by a factor of about ten, assuming $\upsilon_{\rm avg} = 50$ km s$^{-1}$, $\lambda = 0.1$, and the survey area $0.5$ deg$^2$. The parameter $\eta_{\rm std}$ is likely less uncertain, which is known to be 0.06 for a Schwarzschild black hole and 0.4 for a maximally rotating Kerr black holes [e.g., @Thorne74]. IBH Spectrum {#ibh-spectrum} ------------ Considering a different X-ray spectrum of IBHs will change $f_{\rm band}$ of each telescope and consequently the flux observed in each band. We have assumed a power-law spectrum because detectable IBHs are almost always in the RIAF regime, and if we consider the photon index to be $\zeta = $1.4 (2.1) and maintain all the other assumptions, $f_{\rm band}$ of [*ROSAT*]{} and [*FORCE*]{} changes to 0.092 (0.55) and 0.33 (0.16), respectively, from the values of $f_{\rm band} = 0.17$ and $0.31$ for $\zeta = 1.6$. From the discussions in Section \[ssec:interpretation\], the number of detections by [*ROSAT*]{} and [*FORCE*]{} changes by the flux limit to the power of at most 0.7. Then the change of detectable IBH number is at most a factor of two. Identification of IBHs ---------------------- If an IBH candidate is found in hard X-ray surveys like [*FORCE*]{}, it is then necessary to distinguish from other sources emitting X-rays, such as cataclysmic variables (CVs) or X-ray binaries. Soft and/or thermal sources, such as CVs, may be discriminated by selecting hard power-law spectrum sources as expected for RIAF. Discrimination from X-ray binaries may be possible by looking for a binary companion using infrared telescopes [e.g., @Matsumoto17]. Our calculation predicts that IBHs will be found preferentially in dense molecular clouds, and IBH candidates would be particularly strong if they are embedded in molecular clouds around GC found by radio observations. Background AGNs may be removed by checking variability time scales and long-term proper motions [@Fender13]. The purpose of this work is to predict the number of IBHs that are bright enough to be detected, and detailed observational strategies to identify them are beyond the scope of this paper. Conclusions {#sec:conclusions} =========== In this work we estimated the number and X-ray luminosity of IBHs accreting from ISM or molecular cloud gas, and investigate detectability by past and future surveys, by taking into account the realistic structure of our Galaxy. The orbit of each IBH is calculated by integrating the equation of motion in the Galactic potential, and luminosity is calculated considering various phases of ISM and molecular clouds. An important result is that most of the detectable IBHs reside near the GC, not only for surveys targeted to GC but also for the [*ROSAT*]{} survey in all sky. This demonstrates the importance of considering the entire Galactic profile in a search for IBHs. The detectable number of IBHs was calculated with two main model parameters: the average of kick velocity $\upsilon_{\mathrm{avg}}$ and the ratio of actual accretion to the Bondi accretion rate $\lambda$. We found that a few tens of IBHs would be detected by [*ROSAT*]{} with an optimistic parameter set of $\upsilon_{\mathrm{avg}}$ = 50 km s$^{-1}$ and $\lambda = 0.1$, if we ignore absorption. However, most of such IBHs are in GC and soft X-rays should be severely absorbed in ISM, and hence non-detection of IBHs by [*ROSAT*]{} does not give a strong constraint. The expected number for the survey by [*XMM-Newton*]{} towards GC is a few hundred for the same parameter set, and IBHs may have been detected in the [*XMM-Newton*]{} sources, though discrimination from other X-ray source populations may be difficult. The hard X-ray band may have an advantage about discrimination, but the expected number for the survey performed by [*NuSTAR*]{} is at most order unity ($\sim$4). The future [*FORCE*]{} survey towards GC in hard X-ray may detect 30–100 with the same parameter set, although this depends on the actual survey parameters. It should be noted that $\lambda$ can be smaller than 0.1 depending on the physics of accretion, and in such a case IBH detection may be difficult even in the foreseeable future. Acknowledgements {#acknowledgements .unnumbered} ================ The authors express their deepest thanks to Masayoshi Nobukawa and Koji Mori for information on the [*FORCE*]{} satellite and many other valuable comments. We thank the anonoymous referee for suggestions which greatly improved the manuscript of this work. We also thank the faculty members at the University of Tokyo, notably Noriyuki Matsunaga, Toshikazu Shigeyama, Hideyuki Umeda, Kazuhiro Shimasaku, and Kazumi Kashiyama, for fruitful comments and discussions. NK is supported by the Hakubi project at Kyoto University. TT was supported by JSPS KAKENHI Grant Numbers JP15K05018 and JP17H06362. \[lastpage\] [^1]: [email protected]

ArXiv

--- author: - | A.J.E.M. Janssen\ Eindhoven University of Technology,\ Department of Mathematics and Computer Science,\ P.O. Box 513, 5600 MB Eindhoven, The Netherlands.\ E-mail address: {[email protected]}\ Tel.: +31-402474541 title: 'Sharp bound on the radial derivatives of the Zernike circle polynomials (disk polynomials)' --- \ \ \ \ \ [**Abstract.**]{}\ We sharpen the bound $n^{2k}$ on the maximum modulus of the $k^{{\rm th}}$ radial derivative of the Zernike circle polynomials (disk polynomials) of degree $n$ to $n^2(n^2-1^2)\cdot ... \cdot(n^2-(k-1)^2)/2^k(1/2)_k$. This bound is obtained from a result of Koornwinder on the non-negativity of connection coefficients of the radial parts of the circle polynomials when expanded into a series of Chebyshev polynomials of the first kind. The new bound is shown to be sharp for, for instance, Zernike circle polynomials of degree $n$ and azimuthal order $m$ when $m=O(\sqrt{n})$ by using an explicit expression for the connection coefficients in terms of squares of Jacobi polynomials evaluated at 0.\ \ [**Keywords**]{}: Zernike circle polynomial, disk polynomial, radial derivative, Chebyshev polynomial, connection coefficient, Gegenbauer polynomial. Introduction {#sec1} ============ \ The Zernike circle polynomials (ZCPs) $Z_n^m$ are defined for integer $n$ and $m$ such that $n-|m|$ is even and non-negative by \[e1\] Z\_n\^m(re\^[i]{})=e\^[im]{}R\_n\^[|m|]{}(r)=e\^[im]{}r\^mP\^[(0,|m|)]{}\_(2r\^2-1) , where $0\leq r\leq 1$, $\varp\in\dR$, and where $R_n^{|m|}(r)$ is its radial part with $P_p^{(\alpha,\beta)}(x)$ the Jacobi polynomial of degree $p=0,1,...$ corresponding to the weight function $(1-x)^{\alpha}(1+x)^{\beta}$, $-1\leq x\leq1$. The Zernike circle polynomials are directly related to the disk polynomials $R_{k,l}^{(\alpha)}$, see [@ref8], 18.37(i) on pp. 477–478, by $Z_n^m=R_{\frac12(n+m),\frac12(n-m)}^{(0)}$. The ZCPs were introduced by Zernike [@ref10] in connection with his phase contrast method and further elaborated by Nijboer in his thesis [@ref7] on the diffraction theory of aberrations for optical systems with circular pupils. The ZCPs were also used by Cormack [@ref2] in the context of computerized tomography with functions on the disk to be reconstructed from their line integrals. The usefulness of the ZCPs in these fields arises from the fact that they are a complete orthogonal set of functions on the unit disk that have a closed-form expression for both their 2-D Fourier transform and their Radon transform. For the basic properties of ZCPs and their application in optical diffraction theory, see [@ref1], Appendix VII and Chapter 9, Section 2. For further mathematical properties of the ZCPs, among which an addition-type formula, see [@ref3]. In a recent statistical study of the Radon transform in a medical imaging context, there is interest in upper bounds for the modulus of the radial derivatives of the ZCPs, see [@ref6], Appendix B, where it is shown that for $k=0,1,...$ \[e2\] \_[0r1]{}|(R\_n\^[|m|]{})\^[(k)]{}(r)|n\^[2k]{} ,      0r1 , with $f^{(k)}(r)$ denoting the $k^{{\rm th}}$ derivative of $f$ at $r$. We shall sharpen this bound to \[e3\] \_[0r1]{}|(R\_n\^[|m|]{})\^[(k)]{}(r)| ,      0r1 , where $(\alpha)_0=1$, $(\alpha)_k=\alpha(\alpha+1)\cdot...\cdot(\alpha+k-1)$, $k=1,2,...$ (Pochhammer symbol), and where the right-hand side of (\[e3\]) is to be interpreted as 1 for the case that $k=0$. Observe that the right-hand side of (\[e3\]) vanishes when $k>n$. The inequality $|R_n^{|m|}(r)|\leq1$, $0\leq r\leq1$, i.e., the case $k=0$ in (\[e3\]), follows from non-negativity of the connection coefficients that occur when $R_n^{|m|}(r)$ is expanded in a series of Chebyshev polynomials of the first kind, as was shown by Koornwinder [@ref5], and the fact that $R_n^{|m|}(1)=1$. The general case $k=1,2,...$ follows by explicit results for derivatives of Chebyshev polynomials in terms of Gegenbauer polynomials, having an explicit expression for the maximum modulus of these. The inequality in (\[e3\]) is sharp, in the sense that for large $n$ and $m=O(\sqrt{n})$, the inequality is obtained within a factor of the order $1/\sqrt{n}$. This is shown by using a result of Janssen [@ref3], Section 5, that gives the connection coefficients when expanding $R_n^{|m|}$ into a series of Chebyshev polynomials in an explicit form. Preliminary result {#sec2} ================== \ In [@ref3], Section 5, it is shown that \[e4\] R\_n\^[|m|]{}()=\_[j=0]{}\^[n/2]{}a\_[n-2j]{}(n-2j) ,      0 , where $\lfloor n/2\rfloor$ is the largest integer not exceeding $n/2$, and, for non-negative integer $i$ with $n-i$ non-negative and even, \[e5\] a\_i=a\_i(m)=\_i(12)\^l(P\_p\^[(,)]{}(0))\^2 , in which $\eps_0=1$, $\eps_1=\eps_2=\cdots =2$, and \[e6\] p=,  q=,  s=,  t=,  =,   = , with \[e7\] l=(|m|,i) ,      r=(|m|,i) . Note that $p$, $q$, $s$, $t$, $\gamma$, $\delta$ in (\[e6\]) are non-negative and integer since $n$, $m$, $i$ have same parity. The $a_i$ are non-negative, and this confirms Koornwinder’s result. Since $R_n^{|m|}(1)=P_{\frac12(n-|m|)}^{(0,|m|)}(1)=1$, we have \[e8\] |R\_n\^[|m|]{}()|\_[j=0]{}\^[n/2]{}a\_[n-2j]{}=R\_n\^[|m|]{}(1)=1 , showing (\[e3\]) for $k=0$. The result (\[e4\]–\[e7\]) gives specific information as to when $a_i=0$. For instance, when $i=0$, we have $l=|m|$, $r=0$, and so $\gamma=\delta$, $p=\tfrac12(n-|m|)$. Therefore $a_0=0\Leftrightarrow \tfrac12(n-|m|)$ is odd, since $P_p^{(\gamma,\gamma)}(0)=0\Leftrightarrow p$ is odd, see [@ref8], second line of Table 18.61 on p. 444. Similarly, when $m=0$ we have $a_i=0$ if and only if $\tfrac12(n-i)$ is odd. An important case of a non-vanishing $a_{n-2j}$ is $a_n=a_n(m)$ (so $j=0$) that is used in Section \[sec3\] where we discuss sharpness of the bound in (\[e3\]). Then $l=i=n$, $r=|m|$, $p=0$, and since $P_0^{(\gamma,\delta)}(0)=1$ , we thus get \[e9\] a\_n=2(12)\^n( n\ ) ,      n&gt;0 . In a similar manner, see [@ref9], bottom of p. 161 for $P_1^{(\gamma,\delta)}(0)$ and $P_2^{(\gamma,\delta)}(0)$, $$\begin{aligned} \label{e10} & \mbox{} & a_{n-2}=a_n\,\dfrac{|m|^2}{n}~,~~~~~~n-2>0\,,~~n-2\geq|m|~; \nonumber \\[2mm] & & a_{n-4}=a_n\,\dfrac{(n-|m|^2)^2}{2n(n-1)}~,~~~~~~n-4>0\,,~~n-4\geq|m|~.\end{aligned}$$ Proof of the new bound {#sec3} ====================== \ Let $k=1,2,...\,$. Setting $r=\cos\vart$ in (\[e4\]) and using that $\cos(n-2j)\,\vart=T_{n-2j}(\cos\vart)$, with $T_l$ the Chebyshev polynomial of the first kind and of degree $l=0,1,..\,$, we have \[e11\] (R\_n\^[|m|]{})\^[(k)]{}(r)=\_[j=0]{}\^[n/2]{}a\_[n-2j]{}T\_[n-2j]{}\^[(k)]{}(r) . From [@ref8], 18.9.21 on p. 447, 18.7.4 on p. 444 and 18.9.19 on p. 446 (used $k-1$ times), we have \[e12\] T\_[n-2j]{}\^[(k)]{}(r)=2\^[k-1]{}(k-1)!(n-2j)C\_[n-2j-k]{}\^k(r) , where $C_i^k$ is the Gegenbauer polynomial of degree $i$ corresponding to the weight function $(1-x^2)^{k-1/2}$, $-1\leq x\leq1$ with $C_i^k=0$ for $i={-}1,{-}2,...\,$. The result (\[e12\]) can be found in this form in [@ref9], first item in line 6 of p. 188, from which we borrow the notation $C_i^k$, rather than the notation $C_i^{(k)}$ as used in [@ref8], Chapter 18, to avoid confusion with the operation of taking the $k^{{\rm th}}$ derivative. From [@ref8], 18.14.4 on p. 450, we have that $|C_i^k(r)|$ is maximal (strictly when $i>0$) at $r={\pm}1$, with maximum value \[e13\] C\_i\^k(1)= . Since $a_{n-2j}\geq0$, we have from (\[e11\]) and (\[e12\]) that $|(R_n^{|m|})^{(k)}(r)|$ is maximal at $r=1$, with maximal value $$\begin{aligned} \label{e14} (R_n^{|m|})^{(k)}(1) & = & \sum_{j=0}^{\lfloor n/2\rfloor}\,a_{n-2j}\,T_{n-2j}^{(k)}(1) \nonumber \\[3.5mm] & = & 2^{k-1}(k-1)!\,\sum_{j=0}^{\lfloor n/2\rfloor}\,a_{n-2j}(n-2j)\,\frac{(2k)_{n-2j-k}}{(n-2j-k)!}~.\end{aligned}$$ The maximum at $r=1$ is strict when $n>0$ since, see (\[e9\]), $a_n>0$. The quantity $i(2k)_{i-k}/(i-k)!$ is positive and increasing in $i=k,k+1,...\,$, while $\sum_{j=0}^{\lfloor n/2\rfloor}\,a_{n-2j}=1$ and $a_{n-2j}\geq0$ for $j=0,1,...,\lfloor n/2\rfloor$. Therefore \[e15\] 2\^[k-1]{}(k-1)!na\_n(R\_n\^[|m|]{})\^[(k)]{}(1)2\^[k-1]{}(k-1)!n . The third member in (\[e15\]) can be rewritten as $$\begin{aligned} \label{e16} 2^{k-1}(k-1)!\,n\,\frac{(2k)_{n-k}}{(n-k)!} & = & \frac{2^{k-1}(k-1)!}{(2k-1)!}\,n\,\frac{(n+k-1)!}{(n-k)!} \nonumber \\[3mm] & = & \frac{n^2(n^2-1^2)\cdot ...\cdot(n^2-(k-1)^2)}{2^k(1/2)_k}~,\end{aligned}$$ and this gives the upper bound (\[e3\]) as well as a lower bound for\ $\max_{0\leq r\leq1}\, |(R_n^{|m|})^{(k)}(r)|$. Sharpness of the bounds in (\[e15\]) {#sec4} ==================================== \ We have \[e17\] (R\_n\^[|m|]{})\^[(k)]{}(1)=\_[j=0]{}\^[n/2]{}a\_[n-2j]{}(m)T\_[n-2j]{}\^[(k)]{}(1) , where $T_1^{(k)}(1)$ are given for $i=k,k+1,...$ by \[e18\] T\_i\^[(k)]{}(1)== 2\^[k-1]{}(k-1)!  , and vanish for $i<k$. When (for a given $n$ and $m$) $k$ increases, the terms $a_{n-2j}(m)$ in the series in (\[e7\]) with small $j$ get relatively more weight than those with large $j$. It is therefore expected that the inquality in (\[e3\]), i.e., the upper bound in (\[e15\]), tends to be sharp for small $k$ while the lower bound in (\[e15\]) tends to be sharp for large $k$. Indeed, there is equality in the second inquality in (\[e15\]) when $k=0$ and in the first inequality in (\[e15\]) when $k=n,n-1$ (for then the only non-vanishing term in the series in (\[e17\]) is the one with $j=0$). For the cases $k=1,2\,$, we have \[e19\] (R\_n\^[|m|]{})’(1)=|m|+2w ,      (R\_n\^[|m|]{})”(1)=|m|(|m|-1)+2w(w+|m|-1) , where $w=\frac14(n-|m|)(n+|m|+2)$. With $n^2$, $\frac13\,n^2(n^2-1)$ being the right-hand side of (\[e3\]) for $k=1,2\,$, we have from (\[e19\]) \[e20\] 12 ,      38 when $n\pr\infty$ and $m=o(n)$. The identities in (\[e19\]) follow from $$\begin{aligned} \label{e21} (R_n^{|m|})'(r) & = & \sum_{j=0}^{\frac12(n-1-|m-1|)}\,(n-2j)\,R_{n-1-2j}^{|m-1|}(r) \nonumber \\[3.5mm] & & +~\sum_{j=0}^{\frac12(n-1-|m+1|)}\,(n-2j)\,R_{n-1-2j}^{|m+1|}(r)~,\end{aligned}$$ see [@ref4], (10) and add the $\pm$-cases that occur there, where we also recall that $R_{n-1-2j}^{|m\pm 1|}(1)=1$ for $j$ in the summation ranges in the two series in (\[e21\]). Actually, (\[e21\]) was used in [@ref6], Appendix B, to show the bound in (\[e2\]) by induction. In general, it can be shown by induction from (\[e21\]) that the ratio of $(R_n^{|m|})^{(k)}(1)$ and the right-hand side of (\[e3\]) converges to $(1/2)_k/(1)_k$ as $n\pr\infty$ and $m$, $k$ are fixed. In the case that $k=n-2$, we have $$\begin{aligned} \label{e22} (R_n^{|m|})^{(n-2)}(1) & = & a_n(m)\,T_n^{(n-2)}(1)+a_{n-2}(m)\,T_{n-2}^{(n-2)}(1) \nonumber \\[3mm] & = & a_n(m)\,T_n^{(n-2)}(1)\Bigl(1+\frac{|m|^2}{n}~\frac{1}{n(2n-3)}\Bigr)~,\end{aligned}$$ where the first item in (\[e10\]) and the second form in (\[e18\]) for $T_{n-2}^{(n-2)}(1)$ and $T_n^{(n-2)}(1)$ has been used. Hence, the lower bound in (\[e15\]) is also sharp for $k=n-2$ when $n\pr\infty$, and a similar thing can be concluded for $k=n-4$ from the second item in (\[e10\]). We next consider $a_n(m)$ as given by (\[e9\]) for $n>0$. We have for $0\leq m\leq\frac12\,n$ by Stirling’s formula \[e23\] a\_n(m)=()\^ ()\^ (1+O(1n)) . In the case that $|m|=O(\sqrt{n})$, this gives \[e24\] a\_n(m)=([-]{})(1+O()) . Therefore, by (\[e15\]), the upper bound (\[e3\]) is attained within a factor $O(1/\sqrt{n})$ when $|m|=O(\sqrt{n})$. There are also cases that neither bound in (\[e15\]) is sharp. For instance, when $n=m=2k$ with large $k$, the ratio of $(R_n^{|m|})^{(k)}(1)$ and the lower bound and upper bound in (\[e15\]) are of the order $(32/37)^k$ and $(8/27)^k$, respectively, and thus tend to $\infty$ and 0 exponentially fast as $k\pr\infty$.\ \ [**Acknowledgement**]{}.\ The author wishes to thank B.M.M. de Weger for demonstrating non-triviality of the problem of deciding whether $P_p^{(\gamma,\delta)}(0)=0$ (non-negative integer $\gamma$, $\delta$ and $p$), already for degrees $p$ as low as 4, 5, 6.\ \ \ \ [99]{} M. Born, E. Wolf, Principles of Optics, seventh ed., Cambridge University Press, Cambridge, 1999. A.M. Cormack, Representation of a function by its line integrals, with some radiological applications, J. Appl. Phys. 34 (1963) 2722–2727 and J. Appl. Phys. 35 (1964) 2908–2913. A.J.E.M. Janssen, New analytic results for the Zernike circle polynomials from a basic result in the Nijboer-Zernike diffraction theory, J. Europ. Opt. Soc. Rap. Public. 6 (2011) 11028. A.J.E.M. Janssen, Zernike expansion of derivatives and Laplacians of the Zernike circle polynomials, J. Opt. Soc. Am. A 31 (2014) 1604–1613. T. Koornwinder, Positivity proofs for linearization and connection coefficients of orthogonal polynomials satisfying an addition formula, J. London Math. Soc. (2) 18 (1978) 101–114. T. Kutta, N. Bissantz, J. Chown, H. Dette, The empitical process of residuals from an inverse regression, Math. Methods Statist. 28 (2019) 104–126. B.R.A. Nijboer, The diffraction theory of aberrations (Ph.D. thesis, Rijksuniversiteit Groningen, 1942), Wolters, Groningen, 1942. F.W.J. Olver, D.W. Lozier, R.F. Boisvert, C.W. Clark, NIST Handbook of Mathematical Functions, Cambridge University Press, Cambridge, 2010. F.G. Tricomi, Vorlesungen über Orthogonalreihen, Springer-Verlag, Berlin, 1955. F. Zernike, Diffraction theory of the knife-edge test and its improved version, the phase-contrast method, Physica 1 (1934) 689–704.

ArXiv

--- abstract: 'The concept of Universal Shower Profile is used to characterize the average behavior of high energy cosmic rays. The shape variables contain important information about composition. They are independent of the primary cross-section by construction, but affected by other hadronic parameters, like multiplicity. The two variables give access to the average nuclear mass of the sample and their compatibility serves as a test of hadronic models.' address: - 'LIP, Av. Elias Garcia, 14-1, 1000-149 Lisboa, Portugal' - 'Departamento de Física, IST, Av. Rovisco Pais, 1049-001 Lisboa, Portugal' author: - 'S. Andringa' - 'R. Conceição' - 'F. Diogo' - 'M. Pimenta' bibliography: - 'Bib-hX.bib' title: Sensitivity to primary composition and hadronic models from average shape of high energy cosmic ray shower profiles --- Extensive Air Shower ,Longitudinal Profile ,Electromagnetic Component ,Shape Variables ,Mass Composition ,Hadronic Interaction Models Introduction {#sec:Intro} ============ The development of high energy cosmic ray showers in the atmosphere gives indirect information on the primary particle. It is, however, governed by very high energy hadronic interactions for which there is no unique description, only different phenomenological models that try to extrapolate the accelerator data obtained at lower energies and different phase space regions. In general the maximum number of particles and the depth at which that maximum occurs are used to characterize the showers. The first is a measurement of energy, while the the second statistically distinguishes primary particles and depends on the cross-section of the first interaction. More information is however hidden in the details of the shower longitudinal profile shape [@USPV; @UnivM; @UnivG]. Electromagnetic (and muonic [@USPmu]) longitudinal shower profiles are well described by the Gaisser-Hillas function, which can be parametrized in $N'=N/N_{max}$ and $X'=X-X_{max}$ as: $$N'=\left(1+\frac{RX'}{L}\right)^{R^{-2}} \exp{\left(-\frac{X'}{LR}\right)} %N'=\exp\Bigl(-\frac{1}{2}\Bigl(\frac{X'}{L}\Bigr)^2\Bigr)\prod_{n=3}^{\infty}{\exp\Bigl(-\frac{R^{n-2}}{n}\Bigl(-\frac{X'}{L}\Bigr)^n\Bigr)}$$ With the above parametrization, introduced in [@USPV], the profile can be recognized as a Gaussian, with width [**L**]{}, with an asymmetry introduced by non-zero values of [**R**]{}. Notice that after translation to the maximum, the variations due to the point of first interaction disappear[^1]. We will start by introducing the Universal Shower Profile (USP), showing that the shape of the energy deposit longitudinal profile is almost universal. Both shape variables, [**L**]{} and [**R**]{}, tend to different values according to energy, primary particle and hadronic interaction model. We will then study the construction of the USP for a given energy, regardless of the possible mixed compositions. In fact, both parameters are a function of the average nuclear mass, with different energy evolutions. We will find that the different hadronic interaction models define specific possible regions for the two parameters, which can then be used to exclude or constrain present models or help in the construction of new ones. Shower Profile Universality =========================== The universality of the shower longitudinal profiles, when expressed in $N'\equiv N/N_{max}$ and $X'\equiv X-X_{max}$ was discussed in [@USPV], where it was shown that [**L**]{} and [**R**]{} could give extra information on the shower properties. Here we concentrate on what can be learned from the average showers. In fact, the stability of the above variables within showers initiated by similar primaries means that by collecting only a limited number of events, important information can already be obtained, with a significant reduction of statistical uncertainties. The average shower profile is not exactly described by a Gaisser-Hillas and so the values of [**L**]{} and [**R**]{} can vary slightly according to the $X'$ interval considered in their determination. In order to increase the available data in a real experiment, the total $X'$ interval should be minimized. However, for a good determination of the width, [**L**]{}, a minimum range around the maximum is needed, while for [**R**]{} the rising tail needs to be included[@USPV]. From now on, we will always consider the range of $X' \in [-350,200]$ g/cm$^2$. ![Shape parameters [**L**]{} (left) and [**R**]{} (right) as a function of the number of events used to get the average shower. The showers were generated using QGSJet-II as the high energy hadronic interaction model. Color codes show different primaries at $\log(E/{\rm eV})=19.0$.[]{data-label="fig:RL"}](./fig_Levts.eps "fig:"){width="45.00000%"} ![Shape parameters [**L**]{} (left) and [**R**]{} (right) as a function of the number of events used to get the average shower. The showers were generated using QGSJet-II as the high energy hadronic interaction model. Color codes show different primaries at $\log(E/{\rm eV})=19.0$.[]{data-label="fig:RL"}](./fig_Revts.eps "fig:"){width="45.00000%"} ![Shape parameters [**L**]{} (left) and [**R**]{} (right) as a function of the number of events used to get the average shower. The showers were generated using QGSJet-II as the high energy hadronic interaction model. Darker colors show fits to the USP of proton and iron showers at both $\log(E/{\rm eV})=18.5$ and $\log(E/{\rm eV})=19.0$; light colors correspond to the USP obtained by mixing events generated at two energies evenly.[]{data-label="fig:RLene"}](./fig_Lene.eps "fig:"){width="45.00000%"} ![Shape parameters [**L**]{} (left) and [**R**]{} (right) as a function of the number of events used to get the average shower. The showers were generated using QGSJet-II as the high energy hadronic interaction model. Darker colors show fits to the USP of proton and iron showers at both $\log(E/{\rm eV})=18.5$ and $\log(E/{\rm eV})=19.0$; light colors correspond to the USP obtained by mixing events generated at two energies evenly.[]{data-label="fig:RLene"}](./fig_Rene.eps "fig:"){width="45.00000%"} Figure \[fig:RL\] shows the build up of the USP variables, with the collection of events. The energy deposit shower profiles were generated with CONEX[@Conex1; @Conex2], using the QGSJet-II.03[@QGSII1; @QGSII2] model, for different primaries at fixed energy of $\log(E/{\rm eV})=19$. After around one hundred events, the average shape becomes stable, with a good separation for the different primaries, in terms of both variables. The USP variables evolve with energy, and figure \[fig:RLene\] shows the results for proton and iron primaries of different energies. Average values of [**L**]{} and [**R**]{} are reached when different energies are mixed. In general, data analysis will mix a range of energies, and care must be taken in the interpretation of the results. However, it can be seen that even for a wide interval of $\log(E/{\rm eV}) \in [18.5,19.0]$, light and heavy primaries can still be distinguished. USP and Mass Composition ======================== ![Shape parameters [**L**]{} (left) and [**R**]{} (right) as a function of $\left< ln(A) \right>$. The showers were generated using QGSJet-II as the high energy hadronic interaction model. Color codes show different combinations of primaries, generated at $\log(E/{\rm eV})=19$.[]{data-label="fig:RL1D"}](./fig_L_lnA.eps "fig:"){width="45.00000%"} ![Shape parameters [**L**]{} (left) and [**R**]{} (right) as a function of $\left< ln(A) \right>$. The showers were generated using QGSJet-II as the high energy hadronic interaction model. Color codes show different combinations of primaries, generated at $\log(E/{\rm eV})=19$.[]{data-label="fig:RL1D"}](./fig_R_lnA.eps "fig:"){width="45.00000%"} ![Relation of the shape parameters of the universal shower profile with the $\left< ln(A) \right>$. The showers were generated using QGSJet-II as the high energy hadronic interaction model. Color codes show the $\left< ln(A) \right>$ for different combinations of primaries, generated at $\log(E/{\rm eV})=19$, and crosses mark the results for pure composition samples.[]{data-label="fig:RL2D"}](./fig_umbrella_lnA_detail.eps){width="85.00000%"} The values of [**L**]{} and [**R**]{} have an almost linear dependence with the logarithm of the nuclear mass, $ln(A)$, as can be seen in fig. \[fig:RL1D\]. There, in addition to the pure samples, also different mixtures of primaries are presented. Even in these cases, the stability around an average shape is easily achieved. The values of both parameters give two measurements of $\left< ln(A) \right>$, with small variation according to the mixture of the sample. A pure sample of helium primaries will have an only slightly lower value of [**R**]{} and higher value of [**L**]{} than a mixture of proton and iron primaries with the same average logarithmic mass. Since the values of both parameters correlate with $\left< ln(A) \right>$, a well defined region can be expected in the [**L, R**]{} plane, as shown in fig. \[fig:RL2D\]. Notice also that the same analysis can be extended to include $X_{max}$, bearing in mind that in that case there is an influence of the primary cross-section: a big change of the cross-section (as, for example, suggested in ref. [@xsec]) would then be apparent as a consistent reading of $\left< ln(A) \right>$ between [**L**]{} and [**R**]{}, with an inconsistency for $\left< X_{max} \right>$. Primary composition and hadronic parameters =========================================== The different values of the shape variables [**L**]{} and [**R**]{} according to primary composition, energy and hadronic interaction models can be used to understand the hadronic parameters governing them. While the new observables are insensitive to the point of first interaction, they can be deeply changed by the most usual processes in cascade development. ![Shape parameters [**(L, R)**]{} of the universal shower profile , generated at $\log(E/{\rm eV})=19$, for the different mass composition combinations in several models QGSJet-II.03, SIBYLL2.1, EPOS1.99, QGSJet01c.[]{data-label="fig:models"}](./fig_RL_models.eps){width="85.00000%"} The correlated evolution of [**L**]{} and [**R**]{} is characteristic of the hadronic interaction model. Figure \[fig:models\] shows the comparison of four different high energy hadronic interaction models: QGSJet01c[@QGS01], SIBYLL2.1[@SYBILL21] and EPOS1.99[@EPOS], in addition to QGSJet-II.03, for showers at $\log(E/{\rm eV})=19$. The measurement of one point in this plane can be used to check the compatibility of each model with experimental data. It can be seen, for example, that EPOS has larger [**L**]{} values for proton showers than the others; it is even more interesting to note that QGSJet-II spans a range of the [**L, R**]{} plane which is quite different from the other three models, namely in what concerns the larger [**R**]{} values for heavy nuclei. This might reflect the fact that multiplicity is much higher in this model. High values of [**R**]{} characterize heavy primaries, which also have larger multiplicities in the first interactions, and the same kind of physical characteristics that enable the distinction of primaries is expected to be present between hadronic interaction models. Figure \[fig:scaling\] shows the evolution of the parameters as a function of energy, for the different models. Notice that the electromagnetic energy is given by $E_{em}=\sqrt{2\pi} L \left.\frac{dE}{dX}\right|_{max}$ and [**L**]{} should be mainly determined by the primary energy. There is a relative change of the order of 5% between primaries, and of the order of 1% between models. On the other hand, [**R**]{} migth be related to multiplicity, with large differences between primaries. We highlight these trends by using the scaled energy, $\log(E/A)$, inspired in the superposition model for showers initiated by nuclei. The scaling is reflected within each model, with all primaries and energies lying approximately in two lines: one for QGSJet-II and another one grouping all other models. ![Shape parameters [**L**]{} (left) and [**R**]{} (right) as a function of energy for four different primaries (p, He, N and Fe) at $\log(E/{\rm eV})=19.0$ in all models, shown in colored dots, and also $\log(E/{\rm eV})=18.5$ in QGSJet-II and QGSJet01c, shown as squares and $\log(E/{\rm eV})=18.5$ in QGSJet-II, as triangles. In the case of [**R**]{}, the energy is scaled by mass number, to highlight the hidden dependency (see text).[]{data-label="fig:scaling"}](./fig_L_E.eps "fig:"){width="45.00000%"} ![Shape parameters [**L**]{} (left) and [**R**]{} (right) as a function of energy for four different primaries (p, He, N and Fe) at $\log(E/{\rm eV})=19.0$ in all models, shown in colored dots, and also $\log(E/{\rm eV})=18.5$ in QGSJet-II and QGSJet01c, shown as squares and $\log(E/{\rm eV})=18.5$ in QGSJet-II, as triangles. In the case of [**R**]{}, the energy is scaled by mass number, to highlight the hidden dependency (see text).[]{data-label="fig:scaling"}](./fig_R_EA.eps "fig:"){width="45.00000%"} Summary and Conclusions ======================= A Universal Shower Profile can be constructed with a limited number of events to characterize the average behavior of high energy cosmic rays. By constructing an average longitudinal shower profile shape, the statistical errors associated to the detection of single events are significantly reduced. The average shape can be described with the variables [**L**]{} and [**R**]{}, which contain important information on composition. While being independent of the first interaction point by construction, the Universal Shower Profile shape keeps information about the other parameters governing shower development which are different among primaries and models, namely multiplicity. Within each model, [**L**]{} and [**R**]{} give two measurements of $\left< A \right>$, with a reduced dependence on the primary cross-section. The consistency between both is a powerful test of the modeling of the hadronic interactions. Acknowledgments {#acknowledgments .unnumbered} =============== We would like to thank all our colleagues in the Auger-LIP group for reading the manuscript. This work is partially funded by Fundação para a Ciência e Tecnologia (CERN/FP123611/2011 and SFRH/BPD/73270/2010), and funding of MCTES through POPH-QREN-Tipologia 4.2, Portugal, and European Social Fund. [^1]: Although changes in the first interaction point affect the atmospheric profile crossed by the air shower, the corresponding effect on the profile shape in gcm$^{-2}$ is negligible when compared to the differences between the several hadronic interaction models used for the shower development, for example.

ArXiv

--- abstract: 'The discovery of a gravitational lens candidate is reported. The quasars SDSS J111611.73+411821.5 and SDSS J111610.68+411814.4 are recognized as two images of the same object, being strongly lensed by the closer galaxy SDSS J111611.03+411820.9. The source is located at a redshift of $z \sim 3$, while the redshift of the lens galaxy is $z \sim 0.25$. The separation of the images is large, $\sim 13$ arcsec. Commonly used models of the mass distribution for the lens galaxy with values of the parameters in the expected range describe the positions and fluxes of the images.' author: - | [*Eusebio Sánchez Álvaro,*]{}\ [*Francisco Javier Rodríguez Calonge*]{}\  \ [*CIEMAT*]{}\ [*Avda. Complutense 22*]{}\ [*E-28040 Madrid (Spain)*]{}\ date:   title: A New Gravitational Lens Candidate with Large Image Separation in the SDSS DR5 Data --- Introduction {#introduction .unnumbered} ============ Gravitational lenses are widely used for astrophysical and cosmological studies. Individually analyzed, they can provide a better understanding of both the lens and the source. The lensing effect depends only on the distribution of mass within the lens, and not on the type of matter, and can be used to study the distribution of dark matter in galaxy or galaxy cluster halos [@ref:darkmatter]. The statistical properties of the lens systems are sensitive to the cosmological constant [@ref:cosmocons], the global Hubble constant [@ref:hubble], or the formation and evolution of galaxies [@ref:evogal]. Currently, about 100 lensed quasars have been discovered[^1], but the recent development of massive sky surveys is increasing this number quickly. The discovery of a new lens candidate in the Sloan Digital Sky Survey (SDSS)[^2] data is reported here. Several gravitational lenses have been discovered previously in the SDSS data [@ref:sdss_lenses]. Collecting a large sample of gravitational lenses with well-defined statistical properties is important to obtain consistent cosmological and astrophysical conclusions, and the SDSS gives a good oportunity to obtain such a sample. In the analysis presented here, the quasars SDSS J111610.8+411814.4 and SDSS J111611.73+411821.5 are identified as two images of the same quasar, being lensed by the closer galaxy SDSSJ111611.03+411820.9. Both quasars are in the spectroscopic catalog of the survey, while the galaxy is a photometric object. The system is studied and modelled to check the consistency of the gravitational lens hypothesis. Observation of the Lens Candidate {#observation-of-the-lens-candidate .unnumbered} ================================= The Sloan Digital Sky Survey {#the-sloan-digital-sky-survey .unnumbered} ---------------------------- The SDSS [@ref:sdss] is a large-area imaging survey of the north Galactic cap. It has covered nearly a quarter of the Celestial Sphere (8000 square degrees). The survey has recorded around $1.3\times10^8$ objects with well-calibrated photometry [@ref:sdss_photometry]. The photometric errors are typically less than 0.03 magnitude [@ref:sdss_magnitudes]. In addition to that, there is a spectroscopic survey, that has collected spectra for roughly $10^6$ galaxies and $10^5$ quasars, and has covered a total area of 5713 square degrees using an optimized tiling algorithm [@ref:sdss_tiling]. The survey uses a dedicated telescope [@ref:sdss_telescope] of 2.5m, located at the Apache Point Observatory, New Mexico. The imaging survey uses the drift-scan mode of the 142 Mpixel camera [@ref:sdss_camera] to obtain data in five broad bands [@ref:sdss_filter] , $u$ $g$ $r$ $i$ $z$, centered at 3561, 4676, 6176, 7494 and 8873 Å respectively. The images are processed automatically by a photometric pipeline [@ref:sdss_pipeline], and are calibrated astrometrically [@ref:sdss_astrometry] and photometrically [@ref:sdss_photometry]. Some objects are selected from the imaging data for spectroscopy using a variety of algorithms, that include a deeper sample of galaxies [@ref:sdss_galaxies] and quasars [@ref:sdss_qso]. The spectra are observed by means of a pair of double spectrographs able to take 640 spectra simultaneously. They are fed by optical fibers each 3” in diameter. The spectra cover the wavelength range from 3800 to 9200 Å , with a resolution of $\lambda/\Delta \lambda \sim 2000$. The results presented here correspond to the Data Release 5 (DR5) of the SDSS [@ref:sdss_dr5]. Previous data releases are described elsewhere [@ref:sdss_datareleases]. Observation of the Lens Candidate {#observation-of-the-lens-candidate-1 .unnumbered} --------------------------------- The identification of the lens candidate presented here was performed during a study of close quasar pairs. The SDSS image of the system can be seen in Figure \[fig:lensimage\], where the two images of the quasar are labelled QSO 1 and QSO 2, and are compatible with point-like objects, and the lens galaxy is clearly visible between them as an extended object, labelled GAL. The large separation between the images makes the system a very clear candidate, since the objects are perfectly resolved even in the SDSS spectroscopic sample, and no additional follow-up is necessary. The main astrometric and photometric features of the objects are described in Table \[tab:features\]. The two images of the quasar are included in the spectroscopic sample of the SDSS. Their spectra can be seen in Figure \[fig:qsospectra\]-top. The angular separation between the two quasar images is large, $13.7''$, and the separation of each image and the lens galaxy is $7.6''$ and $7.9''$. The observed redshifts of the two images using the CIV line are $3.01\pm0.04$ and $3.01 \pm 0.03$. The optimal redshifts from the SDSS spectroscopic calculations are $2.971\pm0.003$ and $2.980 \pm 0.005$. The redshifts are compatible within errors. Moreover, the shapes of both spectra are very similar. The ratio of spectra is shown in Figure \[fig:qsospectra\]-bottom, and it is constant along the full range of considered wavelengths. Considering both quasars as lensed images of a single object is then supported both by the similarity of the spectra and by the consistency of the redshifts. There is no spectrum of the lens galaxy, which is located in the middle of the two images of the quasar, but the photometric redshift computed for it in the SDSS data is $z=0.18 \pm 0.03$ using the templates algorithm [@ref:photoz] or $z=0.25 \pm 0.03$ using the modified neural network method [@ref:photoz2]. The two values are only marginally compatible, but the lens system is very weakly dependent on this redshift. The most sensitive quantity is the time delay between images, that has not been explored in this analysis. For the other parameters, the full analysis has been performed using both values of the redshift and the conclusions are the same, with compatible values for all the parameters. The numbers quoted in the text are for a redshift of 0.25. Lens Model {#lens-model .unnumbered} ========== The gravitational lens hypothesis is analysed using two mass models, a Singular Isothermal Ellipsoid (SIE) and a Singular Isothermal Sphere with shear (SISS). The [gravlens/lensmodel]{} software [@ref:lensmodel] is used. The assumed errors are $\pm 0.1''$ in the positions and 20% in the fluxes. These errors are larger than the actual errors on the measurements, and are adopted to account for possible external perturbations. Both models have 8 parameters (the lens position, the Einstein radius, the ellipticity or shear, the position angle, the source position and the source flux) and the system offers 8 constraints (the image positions, the lens position and the fluxes of the images), what means that the data should be described perfectly in any fit to these models, since there are no degrees of freedom. However, sensible values for the parameters should be obtained from the fits. The measured positions of the images relative to the central galaxy in the lens plane are shown in Figure \[fig:lensangles\] as dots, while the fitted positions are shown as open circles. Both models, SIE and SISS are able to describe the observed positions, as expected. Also the fitted source position and the critical curves for the lens system are presented. The parameters for the lens obtained from each model are shown in Table \[tab:siesiss\]. The mass of the lens galaxy, defined as all the mass contained inside the Einstein radius, is $M_L = (5.3 \pm 0.1) \times 10^{12} M_{\odot}$ in both models. The fitted ellipticity for the lens galaxy in the SIE model is smaller than the measured ellipticity for the light, which is around 0.3 in all the bands, as presented in Table \[tab:lensgalaxy\]. The fitted shear in the SISS model is in the typical range for lens systems. The fitted position angles are not compatible with the measured angle for the light, shown in Table \[tab:lensgalaxy\], what means that some slight external perturbation may be present. In fact, there seems to be a group or cluster of galaxies between the source and the lens galaxy, at $z \sim 0.5$, that may also contribute to the lensing. Since the redshifts for the source quasar and the lens galaxy are known, it is possible to compute the angular diameter distances, for a flat cosmology with $\Omega_{\Lambda}=0.7$, $\Omega_{M}=0.3$ and $h=0.71$. Then, the expected magnifications of the quasar images can be obtained from the lens models. The expected ratio of fluxes of the two images of the quasar is found to be $0.38 \pm 0.02$ for both models, in agreement with the observed value, $0.38 \pm 0.01$, as expected. The observed value has been obtained from a fit to the ratio of spectra of Figure \[fig:qsospectra\]-bottom and confirmed from the differences of magnitudes from Table \[tab:features\]. The velocity dispersion of the mass producing the lens in the SIE model can be determined from the Einstein radius, and it is found to be $\sigma = 560$ km/s, higher than expected for a typical elliptical galaxy. This fact is not strange, since the mass distribution producing the lens extends for $5.80 \pm 0.11$ effective radii from the center of the galaxy. The effective radius of the galaxy, taken as the radius that contains half of the petrosian flux of the galaxy in the r-band is read from the SDSS database. Then, using differents mass distributions (described in [gravlens/lensmodel]{}), the mass inside the effective radius can be inferred from the mass contained inside the Einstein radius. When the SIE model is used to evaluate the mass of the galaxy inside the effective radius, the obtained value is $M_{R_{eff}}^{SIE}=(0.91 \pm 0.02) \times 10^{12} M_{\odot}$. The corresponding velocity dispersion for this radius is $\sigma_{R_{eff}}=232$ km/s. These are typical values for an elliptical galaxy. If a NFW profile is used, the mass contained in the effective radius is $M_{R_{eff}}^{NFW}=(2.33 \pm 0.05) \times 10^{12} M_{\odot}$. An extended part of the dark matter halo is contributing to the lens effect. The contribution of the halo is confirmed by the fact that the mass-to-luminosity ratio of the lens galaxy is $M_{L}/L_{L} \sim 170 M_{\odot}/L_{\odot}$. This value is in the expected range if the dark matter halo contributes to the lens. The absolute luminosity of the galaxy has been obtained from the SDSS database. The mass-to-luminosity ratios inside the effective radius, are $\sim 30 M_{\odot}/L_{\odot}$ for the isothermal models and $\sim 100 M_{\odot}/L_{\odot}$ for the NFW model. The value for the NFW is much higher than expected for a typical elliptical galaxy. The value for the isothermal profile is high, but can be still in the expected range for elliptical galaxies. Conclusions {#conclusions .unnumbered} =========== A candidate to strong lensed quasar by a closer galaxy has been found in the SDSS DR5 data. The system fulfills the conditions to be considered a gravitational lens. There are two point-like images with very similar redshifts ($\sim 3$) and spectra and there is a galaxy between the images with a redshift smaller than the images ($\sim 0.25$). The angular separation between the two images is large($\sim 13 ''$) The fitted Einstein radius for the lens is large, $\theta_{E} = 7.68 \pm 0.09 ''$. The total mass inside the Einstein radius is $M_{L}=(5.3 \pm 0.1) \times 10^{12} M_{\odot}$. The mass-to-luminosity ratio of the lens galaxy at the Einstein radius, or 5.8 effective radii, is high, $M_L/L_{L}>100 M_{\odot}/L_{\odot}$. This is compatible with a substantial contribution of the dark matter halo to the lensing mass. The extrapolation of the mass value from the Einstein radius to the effective radius with different mass distributions shows that the isothermal models are a good description for an elliptical galaxy, with mass and velocity dispersion in the expected range. The orientation angle resulting from the fit (in both models) is misaligned with respect to the orientation angle of light, which may indicate some small contribution of an external perturbation. In fact, there seems to be a group or cluster of galaxies at $z \sim 0.5$ that may have some contribution to the lensing of the quasar. Acknowledgments {#acknowledgments .unnumbered} =============== Funding for the Sloan Digital Sky Survey (SDSS) and SDSS-II has been provided by the Alfred P. Sloan Foundation, the Participating Institutions, the National Science Foundation, the U.S. Department of Energy, the National Aeronautics and Space Administration, the Japanese Monbukagakusho, and 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 (ARC) for the Participating Institutions. The Participating Institutions are the American Museum of Natural History, Astrophysical Institute Potsdam, University of Basel, University of Cambridge, Case Western Reserve University, The University of Chicago, Drexel University, Fermilab, the Institute for Advanced Study, the Japan Participation Group, The 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]{} C. S. Kochanek, Astrophys. J. [**373**]{} (1991) 354,\ R. B. Metcalf & P. Madau, Astrophys. J. [**563**]{} (2001) 9,\ M. Chiba, Astrophys. J. [**565**]{} (2002) 17,\ N. Dalal & C. S. Kochanek, Astrophys. J. [**572**]{} (2002) 25,\ P. L. Schechter & J. Wamsbsganss, Astrophys. J. [**580**]{} (2002) 685,\ L. V. E. Koopmans & T. Treu, Astrophys. J. [**583**]{} (2003) 606,\ D. Rusin, C. S. Kochanek & C. R. Keeton, Astrophys. J. [**595**]{} (2003) 29. E. L. Turner, J. P. Ostriker & J. R. Gott, Astrophys. J. [**284**]{} (1984) 1,\ E. L. Turner, Astrophys. J. Lett. [**365**]{} (1990) L43,\ M. Fukugita, T. Futamase & M. Kasai, Mon. Not. Roy. Astron. Soc. [**246**]{} (1990) 24,\ K.-H. Chae [*et al.*]{}, Phys. Rev. Lett. [**89**]{} (2002) 151301. S. Refsdal, Mon. Not. Roy. Astron. Soc. [**128**]{} (1964) 307,\ S. Refsdal, Mon. Not. Roy. Astron. Soc. [**132**]{} (1966) 101. C. S. Kochanek [*et al.*]{}, Astrophys. J. [**543**]{} (2000) 131,\ C. S. Kochanek & M. White, Astrophys. J. [**559**]{} (2001) 559 531,\ C. R. Keeton, Astrophys. J. [**561**]{} (2001) 46,\ M. Oguri, Astrophys. J. [**580**]{} (2002) 2,\ O. Ofek, H.-W. Rix & D. Maoz, Mon. Not. Roy. Astron. Soc. [**343**]{} (2003) 639. D. Johnston, Astron. J. [**126**]{} (2003) 2281,\ N. Inada, Astron. J. [**126**]{} (2003) 666,\ N. Inada, Nature [**426**]{} (2003) 810,\ M. Oguri, Proc. Astron. Soc. Jap. [**56**]{} (2004) 399,\ B. Pindor, Astron. J. [**127**]{} (2004) 1318,\ N. Morgan, Mon. Not. Roy. Astron. Soc. [**349**]{} (2004) 882,\ N. Inada, Astron. J. [**130**]{} (2005) 1967,\ M. Oguri, Astrophys. J. [**622**]{} (2005) 106,\ A. Bolton, Astrophys. J.L [**624**]{} (2005) 21,\ N. Inada, Astron. J. [**131**]{} (2006) 1934,\ B. Pindor, Astron. J. [**131**]{} (2006) 41.\ D. G. York [*et al.*]{}, Astron. J. [**120**]{} (2000) 1579. D. W. Hogg, D. P. Finkbeiner, D. J. Schlegel & J. E. Gunn, Astron. J. [**122**]{} (2001) 2129,\ J. A. Smith [*et al.*]{}, Astron. J. [**123**]{} (2002) 2121,\ Z. Ivezic [*et al.*]{}, Astron. Nachr. [**325**]{} (2004) 583,\ D. L. Tucker [*et al.*]{}, Astron. Nachr. [**327**]{} (2006) 821. R. Lupton, J. E. Gunn & A. Szalay, Astron. J. [**118**]{} (1999) 1406. M. R. Blanton [*et al.*]{}, Astron. J. [**125**]{} (2003) 2276. J. E. Gunn [*et al.*]{}, Astron. J. [**131**]{} (2006) 2332. J. E. Gunn [*et al.*]{}, Astron. J. [**116**]{} (1998) 3040. M. Fukugita, T. Ichikawa, J. E. Gunn, M. Doi, K. Shimasaku & D. P. Schneider, Astron. J. [**111**]{} (1996) 1748. R. Lupton [*et al.*]{} 2001, ASP Conf. Ser. 238,  Astronomical Data Analysis Software and Systems X,  ed. F. R. Harndern Jr., F. A. Primini and H. E. Payne (San Francisco: Astr. Soc. Pac.) (astro-ph/0101420). J. R. Pier [*et al.*]{}, Astron. J. [**125**]{} (2003) 1559. D. J. Eisenstein [*et al.*]{}, Astron. J. [**122**]{} (2001) 2267,\ M. A. Strauss [*et al.*]{}, Astron. J. [**124**]{} (2002) 1810.\ G. T. Richards [*et al.*]{}, Astron. J. [**123**]{} (2002) 2945. Adelman-McCarthy [*et al.*]{}, Astrophys. J. Suppl. (2007) Submitted. C. Stoughton [*et al.*]{}, Astron. J. [**123**]{} (2002) 485 (Early Data Release),\ K. Abazajian [*et al.*]{}, Astron. J. [**126**]{} (2003) 2081 (DR1),\ K. Abazajian [*et al.*]{}, Astron. J. [**128**]{} (2004) 502 (DR2),\ K. Abazajian [*et al.*]{}, Astron. J. [**129**]{} (2005) 1755 (DR3),\ J. Adelman-McCarthy [*et al.*]{} Astrophys. J. Suppl. [**162**]{} (2006) 38 (DR4).\ I. Csabai [*et al.*]{}, Astron. J. [**125**]{} (2003) 580. A. A. Collister & O. Lahav, Publ. Astron. Soc. Pac. [**16**]{} (2004) 345. C. R. Keeton, astro-ph/0102340 (2001), \ C. R. Keeton, astro-ph/0102341 (2001). Object GAL QSO 1 QSO 2 -------------------- ------------------ ------------------ ------------------ $\Delta$ra ($''$) 0.0 10.541 0.590 $\Delta$dec ($''$) 0.0 -5.227 -6.473 $u$ $21.75 \pm 0.28$ $20.40 \pm 0.05$ $21.36 \pm 0.11$ $g$ $19.79 \pm 0.02$ $18.54 \pm 0.01$ $19.45 \pm 0.01$ $r$ $18.62 \pm 0.01$ $18.15 \pm 0.01$ $19.15 \pm 0.01$ $i$ $18.03 \pm 0.01$ $17.93 \pm 0.01$ $19.00 \pm 0.01$ $z$ $17.64 \pm 0.03$ $17.96 \pm 0.02$ $19.03 \pm 0.04$ : \[tab:features\] Astrometry and photometry of the lens objects. The difference in right ascension and declination with respect to the lens galaxy is given in arcsec. The coordinates of the galaxy are ra$=169.045962^{\circ}$, dec$=41.305817^{\circ}$. The given magnitudes are the SDSS model magnitudes in the five bands. Model $\theta_{Einstein}$ ($''$) Mass ($10^{12} M_{\odot}$) Shear Ellipticity Angle ($^{\circ}$) ------- ---------------------------- ---------------------------- ----------------- ----------------- -------------------- SIE $7.68 \pm 0.07$ $5.3 \pm 0.1$ – $0.06 \pm 0.02$ $263 \pm 6$ SISS $7.68 \pm 0.07$ $5.3 \pm 0.1$ $0.02 \pm 0.01$ – $263 \pm 6$ : \[tab:siesiss\] Lens parameters obtained from fits to two models of the mass distribution: SIE and SISS. Ellipticity or shear plus orientation angle may indicate some influence of external perturbations. Band Ellipticity Orientation Angle ($^{\circ}$) ------ ----------------- -------------------------------- -- $u$ $0.41 \pm 0.73$ 142 $g$ $0.34 \pm 0.06$ 156 $r$ $0.28 \pm 0.03$ 157 $i$ $0.30 \pm 0.03$ 153 $z$ $0.26 \pm 0.06$ 157 : \[tab:lensgalaxy\] Main physical characteristics of the lens galaxy in the five SDSS bands. The ellipticity and the orientation angle are the De Vacouleurs values read from the SDSS database. The orientation angle is defined in the SDSS standard way, east of north. ![SDSS image of the lens candidate. The two images of the quasar are compatible with point-like objects, while the galaxy is an extended object located in the middle of the two quasar images. \[fig:lensimage\]](lens_candidate.eps){width="15.0truecm"} ![Top: Spectra of the two QSO images taken in the SDSS spectroscopic survey. Bottom: Ratio of both spectra. The value is constant along the considered wavelenghts range. \[fig:qsospectra\]](qso_spectra.eps){width="16.0truecm"} ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![Measured (dots) and fitted (open circles) positions of the QSO images and the lens galaxy for the two mass models considered, SIE (top) and SISS (bottom). The size of the dots and the open circles does not correspond to the errors, and is chosen to make them visible. The critical curves for the models and the fitted positions of the source are also plotted. \[fig:lensangles\] ](sie_model_positions.eps "fig:"){width="8.0truecm"} ![Measured (dots) and fitted (open circles) positions of the QSO images and the lens galaxy for the two mass models considered, SIE (top) and SISS (bottom). The size of the dots and the open circles does not correspond to the errors, and is chosen to make them visible. The critical curves for the models and the fitted positions of the source are also plotted. \[fig:lensangles\] ](siss_model_positions.eps "fig:"){width="8.0truecm"} ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- [^1]: http://cfa-www.harvard.edu/castles [^2]: http://www.sdss.org

ArXiv

--- abstract: 'Galaxy mergers are expected to have a significant role in the mass assembly of galaxies in the early Universe, but there are very few observational constraints on the merger history of galaxies at $z > 2$. We present the first study of galaxy major mergers (mass ratios $>$ 1:4) in mass-selected samples out to $z \approx 6$. Using all five fields of the HST/CANDELS survey and a probabilistic pair count methodology that incorporates the full photometric redshift posteriors and corrections for stellar mass completeness, we measure galaxy pair-counts for projected separations between 5 and 30 kpc in stellar mass selected samples at $9.7 < \log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) < 10.3$ and $\log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) > 10.3$. We find that the major merger pair fraction rises with redshift to $z\approx6$ proportional to $(1+z)^{m}$, with $m = 0.8\pm0.2$ ($m = 1.8\pm0.2$) for $\log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) > 10.3$ ($9.7 < \log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) < 10.3$). Investigating the pair fraction as a function of mass ratio between 1:20 and 1:1, we find no evidence for a strong evolution in the relative numbers of minor to major mergers out to $z < 3$. Using evolving merger timescales we find that the merger rate per galaxy ($\mathcal{R}$) rises rapidly from $0.07\pm 0.01$ Gyr$^{-1}$ at $z < 1$ to $7.6\pm 2.7$ Gyr$^{-1}$ at $z = 6$ for galaxies at $\log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) > 10.3$. The corresponding co-moving major merger rate density remains roughly constant during this time, with rates of $\Gamma \approx 10^{-4}$ Gyr$^{-1}$ Mpc$^{-3}$. Based on the observed merger rates per galaxy, we infer specific mass accretion rates from major mergers that are comparable to the specific star-formation rates for the same mass galaxies at $z > 3$ - observational evidence that mergers are as important a mechanism for building up mass at high redshift as in-situ star-formation.' author: - Kenneth Duncan - 'Christopher J. Conselice' - Carl Mundy - Eric Bell - Jennifer Donley - Audrey Galametz - Yicheng Guo - 'Norman A. Grogin' - Nimish Hathi - Jeyhan Kartaltepe - Dale Kocevski - 'Anton M. Koekemoer' - 'Pablo G. Pérez-González' - 'Kameswara B. Mantha' - 'Gregory F. Snyder' - Mauro Stefanon title: | Observational constraints on the merger history of galaxies since $z\approx6$:\ Probabilistic galaxy pair counts in the CANDELS fields --- Introduction ============ Galaxies grow their stellar mass in one of two distinct ways. They can grow by forming new stars from cold gas that is either accreted from their surroundings or already within the galaxy. Alternatively, they can also grow by merging with other galaxies in their local environment. Although observations suggest that both channels of growth play have played equal roles in the build-up of massive galaxies over the last eleven billion years [e.g., @Bell:2006ey; @Bundy:2009jw; @Bridge:2010ft; @2010ApJ...719..844R; @Ownsworth:2014gt; @mundy2017], there are few observational constraints on their relative roles in the early Universe. On-going star-formation within a galaxy is to date by far the easiest, and most popular, of the two growth mechanisms to measure and track through cosmic time. The numerous ways of observing star-formation: UV emission, optical emission lines, radio and far-infrared emissions, have allowed star-formation rates of individual galaxies to be estimated deep into the earliest epochs of galaxy formation [see e.g. @Hopkins:2006bq; @Behroozi:2013fg; @Madau:2014gt for compilations of these measurements]. However, in contrast to measuring galaxy star-formation rates, measuring the merger rates of galaxies is a significantly more tricky task, yet at least as equally important for many reasons. Despite the difficulty in measuring merger rates, studying the merger history of galaxies is vital for understanding more than just the mass build-up of galaxies. Mergers are thought to play a crucial role in structure evolution [@1972ApJ...178..623T; @2002MNRAS.333..481B; @Dekel:2009bn], as well as the triggering of star-bursts and active galactic nuclei activity [@Silk:1998up; @Hopkins:2008gr; @2011MNRAS.418.2043E; @Chiaberge:2015ip]. Mergers are also correlated with super-massive black hole mergers, which may be the origin of a fraction of gravitational wave events that future missions such as LISA [@2017arXiv170200786A] will detect. Two main avenues exist for studying the fraction of galaxies undergoing mergers at a given epoch (and hence the merger rate). The first method relies on counting the number of galaxies that exist in close pairs, for example @Zepf:1989gv, @Burkey:1994hn, @Carlberg:1994ke, @Woods:1995il, @Yee:1995bc, @Neuschaefer:1997vr, @Patton:2000kt, and @LeFevre:2000iq . This method assumes that galaxies in close proximity, a galaxy pair, are either in the process of merging or will do so within some characteristic timescale. The second method relies on observing the morphological disturbance that results from either ongoing or very recent merger activity [e.g., @Reshetnikov:2000un; @Conselice:2003jz; @Conselice:2008de; @Lavery:2004dd; @Lotz:2006is; @Lotz:2008kr; @Jogee:2009iz]. These two methods are complementary, in that they probe different aspects and timescales within the process of a galaxy merger. However, it is precisely these different merger timescales which represent one of the largest uncertainties in measuring the galaxy merger rate [e.g., @Kitzbichler:2008gi; @Conselice:2009bi; @Lotz:2010ie; @Lotz:2010hf; @Hopkins:2010ip]. The major merger rates of galaxies have been well studied out to redshifts of $z \leq 2.5$ [@Conselice:2003jz; @Bluck:2009in; @LopezSanjuan:2010cz; @Lotz:2011bu; @Bluck:2012dh], but fewer studies have extended the analysis beyond this. Taking into account systematic differences due to sample selection and methodology, there is a strong agreement that between $z = 0$ and $z\approx 2 - 3$ the merger fraction increases significantly [@Conselice:2003jz; @Bluck:2009in; @LopezSanjuan:2010cz; @Bluck:2012dh; @Ownsworth:2014gt]. @2009MNRAS.397..208C presented the first tentative measurements of the merger fractions at redshifts as high as $4 \leq z \leq 6$, making use of both pair-count and morphological estimates of the merger rate. For both estimates, the fraction of galaxies in mergers declines past $z\gtrsim4$, supporting the potential peak in the galaxy merger fraction at $1 \lesssim z \lesssim 2$ reported by @Conselice:2008de ([-@Conselice:2008de]; morphology) and @RyanJr:2008ka ([-@RyanJr:2008ka]; close pairs). However, as the analysis of @2009MNRAS.397..208C was limited to only optical photometry in the very small but deep Ultra Deep Field [@2006AJ....132.1729B], the results were subject to uncertainties due to small sample sizes and limited photometric redshift and stellar mass estimates. When studying galaxy close pair statistics, to satisfy the close pair criterion two galaxies must firstly be within some chosen radius (typically 20 to 50 kpc) in the plane of the sky and, in many studies, within some small velocity offset along the redshift axis [other studies, e.g. @2010ApJ...719..844R deproject into 3D close pairs]. The typical velocity offset required is $\Delta 500~\rm{km~s}^{-1}$, corresponding to a redshift offset of $\delta z / (1+z) = 0.0017$. However, this clearly leads to difficulties when studying the close pair statistics within deep photometric surveys, as the scatter on even the best photometric redshift estimates is $\delta z / (1+z) \approx 0.01$ to $0.04$ [e.g. @Molino:2014iz]. To estimate the merger fractions of galaxies in wide-area photometric redshift surveys or at high-redshift, a methodology that allows us to overcomes the limitations of redshift accuracy in these surveys is required. The method used must correct or account for the pairs observed in the plane of the sky that are due to chance alignments along the line-of-sight. Various approaches have been used to overcome this limitation, including the use of de-projected two-point correlation functions [@Bell:2006ey; @2010ApJ...719..844R], correcting for chance pairs by searching over random positions in the sky [@Kartaltepe:2007dv], and integrating the mass or luminosity function around the target galaxy to estimate the number of expected random companions [@LeFevre:2000iq; @Bluck:2009in; @Bundy:2009jw]. The drawback of these methods is that they are unable to take into account the effects of the redshift uncertainty on the derived properties, such as rest-frame magnitude or stellar mass, potentially affecting their selection by mass or luminosity. @LopezSanjuan:2014uj [LS15 hereafter] present a new method for estimating reliable merger fractions through the photometric redshift probability distribution functions (posteriors) of galaxies. By making use of all available redshift information in a probabilistic manner, this method has been shown to produce accurate merger fractions in the absence of spectroscopic redshift measurements. In this paper we apply this PDF close pair technique presented in , and further developed by us in @mundy2017 using deep ground based NIR surveys. In this paper we apply this methodology, with some new changes, to all five of the fields in the CANDELS [@2011ApJS..197...35G; @Koekemoer:2011br] photometric survey in order to extend measurements of the major merger fraction of mass-selected galaxies out to the highest redshifts currently possible, $z \sim 6$. This allows us to determine how mergers are driving the formation of galaxies through 12.8 Gyr of its history when the bulk of mass in galaxies was put into place [e.g. @Madau:2014gt]. By doing this we are also able to test the role of minor mergers at lower redshifts, and how major mergers compare with star-formation for the build up of stellar mass in galaxies over the bulk of cosmic time. Crucially, thanks to the availability of extensive narrow- and medium-band surveys in a subset of these fields, we are also able to directly explore the effects of redshift precision on our method and resulting merger constraints. The structure of this paper is as follows: In Section \[merger-sec:data\] we briefly outline the photometric data and the derived key galaxy properties used in this analysis. In Section \[merger-sec:method\] we describe the probabilistic pair-count method of and @mundy2017 as implemented in this work. In Section \[merger-sec:results\] we present our results, including comparison of our observations with the predictions of numerical models of galaxy evolution and comparable studies in the literature. In Section \[merger-sec:discussion\], we discuss our results and their implications. Finally, Section \[merger-sec:summary\] presents our summary and conclusions for the results in this paper. Throughout this paper all quoted magnitudes are in the AB system [@1983ApJ...266..713O] and we assume a $\Lambda$-CDM cosmology ($H_{0} = 70$ kms$^{-1}$Mpc$^{-1}$, $\Omega_{m}=0.3$ and $\Omega_{\Lambda}=0.7$) throughout. Quoted observables are expressed as actual values assuming this cosmology unless explicitly stated otherwise. Note that luminosities and luminosity-based properties such as observed stellar masses scale as $h^{-2}$ while distances such as pair separation scale as $h^{-1}$. Data {#merger-sec:data} ==== The photometry used throughout this work is taken from the matched UV to mid-infrared multi-wavelength catalogs in the CANDELS field based on the CANDELS WFC3/IR observations combined with the existing public photometric data in each field. The published catalogs and the data reduction involved are each described in full in their respective catalog release papers: GOODS South [@Guo:2013ig], GOODS North (Barro et al. *in prep*), COSMOS [@2017ApJS..228....7N], UDS [@Galametz:2013dd] and EGS [@2017ApJS..229...32S]. Imaging Data ------------ ### HST Near-infrared and Optical Imaging The near-infrared WFC3/IR data observations of the CANDELS survey [@2011ApJS..197...35G; @Koekemoer:2011br] comprise of two tiers, a DEEP and a WIDE tier. In the CANDELS DEEP survey, the central portions of the GOODS North and South fields were observed in the WFC3 F105W ($Y_{105}$), F125W ($J_{125}$) and F160W ($H_{160}$) filters in five separate epochs. In fields flanking the DEEP region, GOODS North and South were also observed to shallower depth (two epochs) in the same filters as part of the CANDELS WIDE tier. Additionally, the northern-most third of GOODS South comprises WFC3 Early Release Science [ERS, @2011ApJS..193...27W] region and was observed in F098M ($Y_{98}$), $J_{125}$ and $H_{160}$. Within the GOODS South DEEP region also lies the Hubble Ultra Deep Field [WFC3/IR HUDF: @Ellis:2012iw; @Koekemoer:2013db see also @Bouwens:2010dk [-@Bouwens:2010dk] and @2013ApJS..209....6I [[email protected]]] with extremely deep observations also in $Y_{105}$, $J_{125}$ and $H_{160}$. As part of the CANDELS WIDE survey, the COSMOS, UDS and EGS fields were observed in the WFC3 $J_{125}$ and $H_{160}$ filters to two epochs. Finally, in addition to the CANDELS observations, all five CANDELS fields have also been observed in the alternative J band filter, F140W ($JH_{140}$), as part of the 3D-HST survey [@Brammer:2012bu; @Skelton:do]. The 3D-HST observations, processed in the same manner as the CANDELS observations, are included in the photometry catalogs used in this work. For the GOODS North and South fields, the optical HST images from the Advanced Camera for Surveys (ACS) images are version v3.0 of the mosaiced images from the GOODS HST/ACS Treasury Program, combining the data of @2004ApJ...600L..93G with the subsequent observations obtained by @2006AJ....132.1729B where available and the parallel F606W and F814W CANDELS observations [@2011ApJS..193...27W; @Koekemoer:2011br]. Altogether, each GOODS field was observed in the F435W ($B_{435}$), F606W ($V_{606}$), F775W ($i_{775}$), F814W ($I_{814}$) and F850LP ($z_{850}$) bands. For COSMOS, UDS and EGS, optical ACS imaging in $V_{606}$ and $I_{814}$ is provided by the CANDELS parallel observations in combination with available archival observations [EGS: @Davis:2007ek]. All WFC3 and ACS data were reduced and processed following the method outlined in @Koekemoer:2011br. ### Spitzer Observations Being extremely well-studied extragalactic fields, all of the five fields have deep *Spitzer*/IRAC [@Fazio:2004eb] observations at 3.6, 4.5, 5.8 and 8.0$\mu m$ taken during *Spitzer’s* cryogenic mission. For the GOODS North and South fields the cryogenic mission observations GOODS Spitzer Legacy project (PI: M. Dickinson). The wider COSMOS field was observed as part of the S-COSMOS survey [@Sanders:ed]. The UDS was surveyed as part of the *Spitzer* UKIDSS Ultra Deep Survey (SpUDS; PI: Dunlop). And finally, part of the EGS was observed by @Barmby:hi, with subsequent observations extending the coverage (PID 41023, PI: Nandra). In addition to the legacy cryogenic data, subsequent observations in both the 3.6 and 4.5$\mu$m have since been made during the *Spitzer* Warm Mission as part of both the SEDS [@Ashby:2013cc] and S-CANDELS [@Ashby:fh] surveys, significantly increasing the depth of 3.6 and 4.5$\mu$m over the wider CANDELS area. All of the IRAC data available within the CANDELS footprints were combined and reprocessed, first as part of the SEDS survey [@Ashby:2013cc] and later as part of S-CANDELS [@Ashby:fh]. Due to their earlier publication date, the IRAC data in the published GOODS South and UDS catalogs make use of the SEDS data, while the remaining fields (GOODS North, COSMOS and EGS) use the latest S-CANDELS mosaics. Full details of the IRAC data and its reduction can therefore be found in the respective SEDS or S-CANDELS survey papers. ### Ground-based observations Complementary to the space based imaging of HST and Spitzer, each CANDELS field has also been surveyed by a large number of ground-based telescope and surveys. As these extensive ancillary ground-based observations vary from field to field, we do not present the full details for each field, instead we again refer the interested reader to the corresponding individual release papers for each field: GOODS South [@Guo:2013ig], GOODS North (Barro et al. *in prep.*), COSMOS [@2017ApJS..228....7N], UDS [@Galametz:2013dd] and EGS [@2017ApJS..229...32S]. In addition to the ground-based photometry outlined in the primary CANDELS release papers, in the GOODS North field we also include the medium-band imaging from the Survey for High-z Absorption Red and Dead Sources [SHARDS; @2013ApJ...762...46P]. SHARDS uses 25 medium-band filters between wavelengths of 500-900 nm over an area of 130 arcmin$^{2}$ in the GOODS-N region. This imaging was taken with the 10.4 m Gran Telescopio Canarias (GTC), and by itself gives effectively a spectral resolution of about R=50 down to limits of AB$\approx 26.5$ mag. One of the major goals of the SHARDS survey is to find emission and absorption line galaxies at redshifts up to $z \sim 5$. However, the fine wavelength sampling also makes it a powerful dataset for producing precise photo-$z$ estimates for all source types. Similarly, in the GOODS South field we also include the *Subaru* medium band imaging presented in @Cardamone:2010jf. Source photometry and deconfusion --------------------------------- All of the CANDELS survey catalogs have been produced using the same photometry method, full details which can be found in the respective catalog papers [e.g. @Guo:2013ig; @Galametz:2013dd]. In summary, photometry for the HST bands was done using <span style="font-variant:small-caps;">SExtractor</span>’s [@Bertin:1996hf] dual image mode, using the WFC3 H band mosaic as the detection image in each field and the respective ACS/WFC3 mosaics as the measurement image after matching of the point-spread function (PSF, individual to each field). For all ground-based and *Spitzer* IRAC bands, deconvolution and photometry was done using template fitting photometry (TFIT). We refer the reader to @2007PASP..119.1325L, @2012ApJ...752...66L, and the citations within for further details of the TFIT process and the improvements gained on multi-wavelength photometry. As with the broad-band imaging, photometry for the medium-band imaging was performed using the same TFIT forced photometry procedure employed during the main catalog production [@Guo:2013ig] - with positions based on the corresponding WFC3 $H_{160}$ imaging [@2013ApJ...762...46P and J. Donley, priv. communication for GOODS North and South respectively]. Image depths and detection completeness estimates {#merger-sec:completeness} ------------------------------------------------- Due to the tiered observing strategy employed for the CANDELS survey and the limitations imposed on the tiling of individual exposures, the final $H_{160}$ science images used for the catalog source detections are somewhat in-homogeneous. Not only is there significant variation in image depth across the five CANDELS fields, but each field itself is inhomogeneous. To overcome these limitations whilst still making full use of the deepest available areas, we divide each of the CANDELS fields into sub-fields based on the local limiting magnitude (as determined from the RMS maps of the $H_{160}$ science images). [![Distribution of area with a given limiting magnitude (1$\sigma$ within an area of 1 arcsec$^2$) for each of the five CANDELS fields. The vertical dashed lines show the limiting magnitudes used to define the a) ‘Wide 1’, b) ‘Wide 2’, c) ‘Deep’ and d) ‘Ultra-deep’ sub-fields within each field. The corresponding total area covered (in arcmin$^2$) is also shown for each sub-field. Note that the range of limiting magnitudes shown excludes that reached by the HUDF, hence the area of GOODS South corresponding to the HUDF is not plotted. We refer the interested reader to the individual catalog release papers for an illustration of the spatial distribution of these depths (see Section. \[merger-sec:data\] for references).[]{data-label="merger-fig:lim_depths"}](plots/LimitingMags.pdf "fig:"){width="0.99\columnwidth"}]{} Fig. \[merger-fig:lim\_depths\] illustrates the distribution of area with a given limiting magnitudes (within an area of 1 arcsec$^2$ at 1$\sigma$; $H_{160}^{\text{lim}}$) for each of the five CANDELS fields. While the difference in depth between the WIDE and DEEP tiers of the survey are very clear, there is also noticeable variation in limiting magnitude between fields of with same number of HST observation epochs (COSMOS, UDS and EGS). The observed difference in field depth is primarily due to the different locations on the sky in which the CANDELS fields are located, the ability to schedule HST time to observe these fields, and how the orbits are divided into exposure times. Together these constraints determined the differences in the CANDELS tiling strategies and the resulting exposure times for each pointing [@Koekemoer:2011br; @2011ApJS..197...35G]. As a result of this tiling and scheduling constraints, the EGS pointings are 10-15% longer than in COSMOS and are as a result slightly deeper, with the UDS field in between these two. Additionally, the fields also have different background levels as they are in different portions of the sky, and these different background levels result in different effective depths being reached. This creates the variety of depths for the WIDE and DEEP epochs highlighted by Fig. \[merger-fig:lim\_depths\]. Based on the distributions observed in Fig. \[merger-fig:lim\_depths\], we define four sets of sub-fields based on the following limiting magnitude ranges: $H_{160}^{\text{lim}} < 27.87 ~\textup{mag}$ (Wide 1), $27.87 \leq H_{160}^{\text{lim}} < 28.3 ~\textup{mag}$ (Wide 2), $28.3 \leq H_{160}^{\text{lim}} < 29.1 ~\textup{mag}$ (Deep) and $H_{160}^{\text{lim}} \geq 29.1 ~\textup{mag}$ (Ultra-deep). The sub-sets of observed galaxies are then simply defined based on the measured $H_{160}^{\text{lim}}$ at the position of the galaxy. To ensure consistent estimates of the respective source detection limits, we performed new completeness simulations across all five CANDELS fields. These simulations include a realistic range of input (magnitude-dependent) morphologies based on the observed structural properties of galaxies in the CANDELS fields [@vanderWel:2012eu]. Full details of how the completeness simulations were performed are outlined in Appendix \[app:completeness\]. In Fig. \[merger-fig:gs\_completeness\], we present an example plot illustrating the measured source recovery fraction as a function of magnitude for each of the sub-fields within the GOODS South field. Due to the effects of source confusion and chance alignment with brighter sources in the field, it can be seen that the catalogs are 100% complete at only the very brightest magnitudes. For this field, the $80\%$ completeness limits range from $H_{160} = 25.29 \textup{mag}$ for the shallowest observations down to $H_{160} = 27.26 \textup{mag}$ mag for the Ultra-deep field. In Table \[tab:completeness\], we present the measured completeness limits for image regions of different limiting magnitude for each CANDELS field. Figures illustrating the detection completeness for all fields are included for reference in Appendix \[app:completeness\]. [![Example detection completeness estimates, showing the fraction of recovered sources as a function of $H_{160}$ magnitude for the GOODS South field. The vertical dashed lines show the magnitude at which the recovery fraction equals 80% for each sub-field.[]{data-label="merger-fig:gs_completeness"}](plots/GS_completeness.pdf "fig:"){width="0.95\columnwidth"}]{} [lcccccccc]{} \[tab:completeness\] GOODS South & 33.88 & 25.29 & 33.75 & 25.91 & 96.74 & 26.44 & 5.15 & 27.26\ GOODS North & 39.23 & 25.28 & 57.11 & 25.77 & 76.6 & 26.56 & 0.0 & -\ COSMOS & 48.21 & 25.35 & 147.89 & 25.74 & 5.88 & 26.23 & 0.0 & -\ UDS & 56.82 & 25.46 & 141.38 & 25.95 & 3.95 & 26.28 & 0.0 & -\ EGS & 13.36 & 25.43 & 164.66 & 26.06 & 26.65 & 26.29 & 0.0 & -\ Total Area (Average) & 191.5 & 25.36 & 544.7 & 25.9 & 209.8 & 26.46 & 5.15 & 27.26\ Photometric redshifts {#merger-sec:photoz} --------------------- Photometric redshift (photo-$z$) estimates for all five fields are calculated following a variation of the method presented in @Duncan:2017ul and @Duncan:2017wu. In summary, template-fitting estimates are calculated using the <span style="font-variant:small-caps;">eazy</span> photometric redshift code [@Brammer:2008gn] for three different template sets and incorporates zero-point offsets to the input fluxes and additional wavelength dependent errors [we refer the reader to @Duncan:2017ul for details]. Templates are fit to all available photometric bands in each field as outlined in Section \[merger-sec:data\]. Additional empirical estimates using a Gaussian process redshift code [<span style="font-variant:small-caps;">GPz</span>; @2016MNRAS.462..726A] are then estimated using a subset of the available photometric bands (further details discussed below). Finally, after calibration of the individual redshift posteriors (Section \[sec:pdf\_calibration\]), the four estimates are then combined in a robust statistical framework through a hierarchical Bayesian (HB) combination to produce a consensus redshift estimate. For the GOODS North field, we also calculate an additional second set of photo-$z$ estimates incorporating the SHARDS medium-band photometry based using only template fitting. The template fits for the GOODS North + SHARDS photometry are calculated using the default <span style="font-variant:small-caps;">eazy</span> template library. To account for the spatial variation in filter wavelength intrinsic to the SHARDS photometry [see @2013ApJ...762...46P], the fitting for each source is done using its own unique set of filter response functions specific to the expected SHARDS filter central wavelengths at the source position. ### Luminosity priors in template fitting and HB combination {#merger-sec:priors} When calculating the redshift posteriors for each template fit, we do not make use of a luminosity-dependent redshift prior as is commonly done to improve photometric redshift accuracy [@Brammer:2008gn; @Dahlen:2013eu], i.e. we assume a luminosity prior which is flat with redshift. Luminosity dependent priors such as the one implemented in <span style="font-variant:small-caps;">eazy</span> rely on mock galaxy lightcones which accurately reproduce the observed (apparent) luminosity function. Current semi-analytic models do agree well with observations at $z < 2$ [@Henriques:2012gsa], but increasingly diverge at higher redshift [@Lu:2014kl] and may not represent an ideal prior. Even in the case of an empirically calculated prior [e.g. @Duncan:2017wu] that may not suffer from these limitations, the use of a prior which is dependent only on a galaxy’s luminosity and not its color or wider SED properties could significantly bias the estimation of close pairs using redshift posteriors. As an example, we can imaging a hypothetical pair of galaxies at identical redshifts and with identical stellar population properties such that the only difference is the stellar mass of the galaxy (i.e. the star-formation histories differ only in normalization). If a luminosity-dependent prior is then applied, the posterior probability distribution for each galaxy will be modified differently for each galaxy and could erroneously decrease the integrated pair probability. ### Gaussian process redshift estimates In addition to the primary template-based estimates outlined in the previous section, our consensus photo-$z$s also incorporate empirical photo-$z$ estimates based on the Gaussian process redshift code <span style="font-variant:small-caps;">GPz</span> [@2016MNRAS.455.2387A; @2016MNRAS.462..726A]. Our implementation of the <span style="font-variant:small-caps;">GPz</span> code in this work, includes magnitude and color-dependent weighting of the spectroscopic training sample, and follows the procedure outlined in @Duncan:2017wu, to which we refer the reader for additional details. The spectroscopic training sample for each field was taken from a compilation of those available in the literature (Hathi, priv. communication), with additional spectroscopic quality cuts applied based on the quality flags provided by each survey. To maximise the training sample available, we train <span style="font-variant:small-caps;">GPz</span> using only a subset of the available filters that are common to multiple fields: $V_{606}$, $I_{814}$, $J_{125}$, $H_{160}$ from HST (additionally $B_{435}$ for GOODS North and South) as well as the 3.6 and 4.5$\mu$m IRAC bands of *Spitzer*. In practice, the resulting <span style="font-variant:small-caps;">GPz</span> estimates have significantly higher scatter ($\sigma_{\textup{NMAD}} \approx 10\%$)[^1] and out-lier fraction ($\gtrsim 15\%$) than their corresponding template estimates. Nevertheless, we include the <span style="font-variant:small-caps;">GPz</span> estimates within the Hierarchical Bayesian combination procedure as they can serve to break color degeneracies inherent within the template estimates in a more sophisticated manner than a simple luminosity prior (see Sec. \[merger-sec:priors\]). ### Calibrating redshift posteriors {#sec:pdf_calibration} In @Hildebrandt:2008jh, @Dahlen:2013eu and more recently @2016MNRAS.457.4005W and @Duncan:2017ul, it is shown that the redshift probability density functions output by photometric redshift codes can often be an inaccurate representation of the true photometric redshift error. This inaccuracy can be due to under- or over-estimates of photometric errors, or a result of systematic effects such as the template choices. Whatever the cause, the effect can result in significantly over- or underestimated confidence intervals whilst still producing good agreement between the best-fit $z_{\rm{phot}}$ and the corresponding $z_{\rm{spec}}$. Although this systematic effect may be negated when measuring the bulk properties of larger galaxy samples, the method central to this paper relies on the direct comparison of individual redshift posteriors. It is therefore essential that the posterior distributions used in the analysis accurately represent the true uncertainties. Given this known systematic effect, we therefore endeavor to ensure the accuracy of our redshift posteriors before undertaking any analysis based on their posteriors. A key feature of the photo-$z$ method employed in this work is the calibration of the redshift posteriors for all estimates included in the Bayesian combination [@Duncan:2017ul; @Duncan:2017wu]. Crucially, this calibration is done as a function of apparent magnitude, rather than as a global correction, minimizing any systematic effects that could result from biases in the spectroscopic training sample. An additional step in the calibration procedure introduced in this work is the correction of bias in the posteriors by shifting the posteriors until the Euclidean distance between the measured and optimum $\hat{F}(c)$ is minimized [@Gomes:2017ut]. This additional correction is necessary due to the very high precision offered by the excellent photometry available in these fields (and the correspondingly low scatter in the resulting estimates) and prevents unnecessary inflation of the uncertainties to account for this bias during the subsequent calibration of the posterior widths. {width="80.00000%"} In Fig. \[merger-fig:pdf\_calibration\] we present cumulative distribution, $\hat{F}(c)$, of threshold credible intervals, $c$, for our final consensus photo-$z$ estimate. For a set of redshift posterior predictions which perfectly represent the redshift uncertainty, the expected distribution of threshold credible intervals should be constant between 0 and 1, and the cumulative distribution should therefore follow a straight 1:1 relation, i.e. a quantile-quantile plot. If there is over-confidence in the photometric redshift errors, i.e the $P(z)$s are too sharp, the $\hat{F}(c)$ curves will fall below the ideal 1:1 relation. Likewise, under-confidence results in curves above this line. Remaining bias in the estimates can manifest as steeper or shallow gradients and offsets in the intercepts at $c = 0$ and $c = 1$. From Fig. \[merger-fig:pdf\_calibration\], we can see that overall the accuracy of the photo-$z$ uncertainties is very high across a very broad range in apparent magnitude. For the GOODS North + SHARDS estimates, there remains a small amount of over-confidence in the photo-$z$ uncertainties. Additionally, for the EGS field there remains a magnitude dependent trend in the photo-$z$ posterior accuracy. Uncertainties for bright sources are slightly under-estimated while those for faint sources are slightly over-estimated. [![Robust scatter ($\sigma_{\textup{NMAD}}$: black circles, left-hand scale) and outlier fraction ($\frac{\left | \Delta z \right |}{1+z_{\rm{spec}}} > 0.15$, blue triangles, right-hand scale) for the galaxies in our samples with available high quality spectroscopic redshifts and with photometric redshift fits which pass our selection criteria. The position of the filled circle/triangle within each bin shows the average spectroscopic redshifts within that bin. Error-bars for the outlier fractions indicate the 1-$\sigma$ binomial uncertainties and lighter blue downward triangles indicate upper limits. For the GOODS North field, gray points (and dotted line) illustrate the scatter for the GOODS North + SHARDS redshift estimates.[]{data-label="merger-fig:specz_comp1"}](plots/photoz_z_stats.pdf "fig:"){width="1.02\columnwidth"}]{} ### Photo-$z$ quality statistics In Fig. \[merger-fig:specz\_comp1\] we illustrate the photometric redshift quality for each CANDELS field as a function of redshift. Following the same metrics as in @Molino:2014iz and , we find that the quality of our photometric redshifts is excellent given the high-redshifts being studied and the broadband nature of the photometry catalog. We find a normalized median absolute deviation of between $\sigma_{\text{NMAD}} \lesssim 1\%$ and $\sigma_{\text{NMAD}} \lesssim 5\%$, depending on redshift. As with most spectroscopic redshift comparison samples, the typically bright nature of the galaxies with high quality spectroscopic redshift may present a biased representation of the quality of the photometric redshifts. We can see this effect in the comparison in Fig. \[merger-fig:specz\_comp1\], by comparing the different $\sigma_{\textup{NMAD}}$ values for the different fields. It may be initially surprising that we find poorer agreement between the photometric and spectroscopic redshifts (w.r.t outlier fraction) at $z>3$ for the GOODS North and South fields compared to EGS and UDS, given that these fields significantly deeper HST data available. In fact, it is the increased level of spectroscopic completeness at fainter magnitudes and higher redshifts that is the reason for the apparently poorer performance in GOODS fields, with spectroscopic redshifts for a greater number of sources for which photo-$z$ are more difficult to measure. However, overall we are still getting good photometric redshifts for the fainter systems. The basis of our analysis is the full redshift posteriors for which we have high confidence in the accuracy and precision. Stellar mass estimates {#merger-sec:stellarmass} ---------------------- The stellar mass as a function of redshift, $\textup{M}_{*} (z)$, for each galaxy is measured using a modified version of the SED code introduced in @Duncan:2014gh. Rather than estimating the best-fit mass (or mass likelihood distribution) for a fixed input photometric or spectroscopic redshift, we instead estimate the stellar mass at all redshifts in the photo-$z$ fitting range. Specifically, we calculate the least-squares weighted mean: $$\textup{M}_{*} (z) = \frac{\sum_{t}^{} w_{t}(z) \text{M}_{*,t}(z)} {\sum_{t} w_{t}(z)}$$ where the sum is over all galaxy template types, $t$, with ages less than the age of the Universe at the redshift $z$, and $\textup{M}_{\star,t}(z)$ is the optimum stellar mass for each galaxy template (Equation \[eq:temp\_norm\]). The weight, $w_{t}(z)$, is determined by $$w_{t}(z) = \exp(-\chi_{t}^{2}(z)/2),$$ where $\chi^{2}_{t}(z)$ is given by: $$\label{eqn:chi} \chi^{2}_{t}(z) = \sum_{j}^{N_{filters}} \frac{(\textup{M}_{\star,t}(z) F_{j,t}(z) - F_{j}^{obs})^2} {\sigma_{j}^{2}}.$$ The sum is over $j$ broadband filters available for each galaxy, its observed photometric fluxes, $F_{j}^{\textup{obs}}$ and corresponding error, $\sigma_{j}$. We note that due to computing limitations, we do not include the available medium-band photometry when estimating stellar masses. The optimum scaling for each galaxy template type (normalized to 1 M$_{\odot}$), $\textup{M}_{\star,t}$, is calculated analytically by setting the differential of Equation \[eqn:chi\] equal to 0 and rearranging to give: $$\label{eq:temp_norm} \textup{M}_{\star,t}(z) = \frac{\sum_{j} \frac{F_{j,t}(z)F_{j}^{obs}}{\sigma^{2}_{j}} } { \sum_{j} \frac{F_{j,t}(z)^{2}}{\sigma^{2}_{j}} }.$$ In this work we also incorporate a so-called “template error function" to account for uncertainties caused by the limited template set and any potential systematic offsets as a function of wavelength. The template error function and method applied to our stellar mass fits is identical to that outlined in @Brammer:2008gn and included in the initial photometric redshift analysis outlined in Section \[merger-sec:photoz\]. Specifically, this means that the total error for any individual filter, $j$, is given by: $$\sigma_{j} = \sqrt{\sigma_{j,obs}^{2} + \left ( F_{j,obs}\sigma_{\textup{temp}}(\lambda_{j}) \right )^{2} }$$ where $\sigma_{j,obs}$ is observed photometric flux error, $F_{j,obs}$ its corresponding flux and $\sigma_{\textup{temp}}(\lambda_{j})$ the template error function interpolated at the pivot wavelength for that filter, $\lambda_{j}$. We note that in addition to estimating the stellar mass, this method also provides a secondary measurement of the photometric redshift, whereby $P(z) \propto \sum_{t} w_{t}(z) $. We use an independently estimated redshift posterior in the pair analysis in place of those generated by the marginalised redshift likelihoods from the stellar mass fits due to the higher precision and reliability offered by our hierarchical Bayesian consensus photo-$z$ estimates. For the @Bruzual:2003ckb templates used in our stellar mass fitting we allow a wide range of plausible stellar population parameters and assume a @Chabrier:2003ki IMF. Model ages are allowed to vary from 10 Myr to the age of the Universe at a given redshift, metallicities of 0.02, 0.2 and 1 Z$_{\odot}$, and dust attenuation strength in the range $0 \le A_{V} \le 3$ assuming a @2000ApJ...533..682C attenuation curve. The assumed star-formation histories follow exponential $\tau$-models ($SFR \propto e^{-t/\tau}$), both decreasing and increasing (negative $\tau$), for characteristic timescales of $\left | \tau \right | = $ 0.25, 0.5, 1, 2.5, 5, 10, plus an additional short burst ($\tau = 0.05$) and continuous star-formation models ($\tau \gg1/H_{0}$). Nebular emission is included in the model SEDs assuming a relatively high escape fraction $f_{\text{esc}} = 0.2$ [@Yajima:2010fb; @Fernandez:2011cw; @Finkelstein:2012hr; @Robertson:2013ji] and hence a relatively conservative estimate on the contribution of nebular emission. As in @Duncan:2014gh, we assume for the nebular emission that the gas-phase stellar metallicities are equivalent and that stellar and nebular emission are attenuated by dust equally. To ensure that our stellar mass estimates do not suffer from significant systematic biases we compare our best-fitting stellar masses (assuming $z = z_{\text{peak}}$) with those obtained by averaging the results of several teams within the CANDELS collaboration [@Santini:2015hh]. Although there is some scatter between the two sets of mass estimates, we find that our best-fitting masses suffer from no significant bias relative to the median of the CANDELS estimates (see Fig. \[merger-fig:mass\_candels\] in the Appendix). Some of the observed scatter can be attributed to the fact that the photometric redshift assumed for the two sets of mass estimates is not necessarily the same. Overall, we are therefore confident that the stellar population modelling employed here is consistent with that of the wider literature. We find no systematic error relative to other mass estimates which make use of stellar models and assume the same IMF. However, standard caveats with regards to stellar masses estimated using stellar population models still apply (see discussion in @Santini:2015hh). Close pair methodology {#merger-sec:method} ====================== The primary goal of analysing the statistics of close pairs of galaxies is to estimate the fraction of galaxies which are in the process of merging. From numerical simulations such as @Kitzbichler:2008gi, it is well understood that the vast majority of galaxy dark matter halos within some given physical separation will eventually merge. For spectroscopic studies in the nearby Universe, a close pair is often defined by a projected separation, $r_{\textup{p}} $, in the plane of the sky of $r_{\textup{p}} < 20~\rm{to}~50~ h^{-1}$ kpc, and a separation in redshift or velocity space of $ \Delta v \leq 500 \textup{ km s}^{-1}$. Armed with a measure of the statistics of galaxies that satisfy these criteria within a sample, we can then estimate the corresponding pair fraction, $f_{\textup{P}}$, defined as $$f_{\textup{P}} = \frac{N_{\textup{pairs}}} {N_{\textup{T}}},$$ where $N_{\textup{pairs}}$ and $N_{\textup{T}}$ are the number of galaxy pairs and the total number of galaxies respectively within some target sample, e.g. a volume limited sample of mass selected galaxies. Note that $N_{\text{pairs}}$ is the number of galaxy pairs rather then number of galaxies *in* pairs which is up to factor of two higher [@Patton:2000kt], depending on the precise multiplicity of pairs and groups. In this work, we analyse the galaxy close pairs through the use of their photo-$z$ posteriors. The use of photo-$z$ posterior takes into account the uncertainty in galaxy redshifts in the pair selection, and the effect of the redshift uncertainty on the projected distance and derived galaxy properties. As presented in this method is able to directly account for random line-of-sight projections that are typically subtracted from pair-counts through Monte Carlo simulations. In the following section we outline the method as applied in this work and how it differs to that presented in in the use of stellar mass instead of luminosity when defining the close pair selection criteria, as well as our use of flux-limited samples and the corresponding corrections. Sample cleaning --------------- Before defining a target-sample, we first clean the photometric catalogs for sources that have a high likelihood of being stars or image artefacts. A common method for identifying stars in imaging is though optical morphology of the sources in the high-resolution HST imaging. The exclusion of objects with high <span style="font-variant:small-caps;">SExtractor</span> stellarity parameters (i.e. more point-like sources) could potentially bias the selection by erroneously excluding very compact neighbouring galaxies and AGN instead of stars. Therefore, when cleaning the full photometric catalog to produce a robust sample of galaxies, we define stars as sources that have a high <span style="font-variant:small-caps;">SExtractor</span> stellarity parameter ($> 0.9$) in the $H_{160}$ imaging *and* have an SED that is consistent with being a star. Using <span style="font-variant:small-caps;">eazy</span>, we fit the available optical to near-infrared photometry (with rest-frame wavelength $< 2.5\mu$m) for each field with the stellar library of @Pickles:1998er while fixing the redshift to zero. We then classify as a star any object which has $\chi^{2}_{\text{Star}}/N_{\textup{filt,S}} < \chi^{2}_{\text{Galaxy}}/N_{\textup{filt,G}}$, where $\chi^{2}_{\text{Galaxy}}$ and $\chi^{2}_{\text{Star}}$ are the best-fit $\chi^{2}$ obtained when fitting the galaxy templates used in Section \[merger-sec:photoz\] and stellar templates respectively, normalized by the corresponding number of filters used in the fitting ($N_{\textup{filt,G}}$, $N_{\textup{filt,S}}$). Based on the combined classification criteria, we exclude $\lesssim 0.4\%$ of objects per field. Thus, the fraction of sources excluded by this criterion is very small so should not present a significant bias in the following analysis. Additionally, to prevent erroneous SED fits (either photo-$z$ or stellar mass estimates) in sources with photometry contaminated by artefacts due to bright stars in the field (and their diffraction spikes) or edge effects, we also exclude sources which have flags in the photometry flag map [see e.g. @Guo:2013ig; @Galametz:2013dd]. Based on inspection of the photo-$z$ quality for all of the sources identified in this initial cut we find the published catalog flags to be overly conservative, with the overall quality of the photo-$z$ for flagged sources comparable to those of un-flagged objects. To exclude only objects for which the photometric artefacts will adversely affect the results in this work, we apply an additional selection criteria: excluding sources which are flagged and have $\chi^{2}_{\text{Galaxy}}/N_{\textup{filt,G}} > 4$, indicative of bad SED fits. Given these criteria, we exclude between 0.71% and 3.3% of sources in each field. Selecting initial potential close pairs {#merger-sec:initial} --------------------------------------- Once an initial sample has been selected based on redshift (see Section \[merger-sec:photoz\]), we then search for projected close pairs between the target and full galaxy samples. The initial search is for close pairs which have a projected separation less than the maximum angular separation across the full redshift range of interest (corresponding to the desired physical separation). Duplicates are then removed from the initial list of close pairs (with the primary galaxy determined as the galaxy with the highest stellar mass at its corresponding best-fit photo-$z$) to create the list of galaxy pairs for the posterior analysis. Because the posterior analysis makes use of all available information to determine the pair fractions, it is applied to all galaxies within the initial sample simultaneously, with the redshift and mass ranges of interest determined by the selection functions and integration limits outlined in the following sections. The pair probability function {#merger-sec:ppf} ----------------------------- For a given projected close pair of galaxies within the full galaxy sample, the combined redshift probability function, $\mathcal{Z}(z)$, is defined as $$\label{eq:Zz} \mathcal{Z}(z) = \frac{2 \times P_{1}(z) \times P_{2}(z)}{P_{1}(z) + P_{2}(z)} = \frac{P_{1}(z) \times P_{2}(z)}{N(z)}$$ where $P_{1}(z)$ and $P_{2}(z)$ are the photo-$z$ posteriors for the primary and secondary galaxies in the projected pair. The normalization, $N(z) = (P_{1}(z) + P_{2}(z))/2$, is implicitly constructed such that $\int_{0}^{\infty} N(z) dz = 1 $ and $\mathcal{Z}(z)$ therefore represents the number of fractional close pairs at redshift $z$ for the projected close pairs being studied. Following Equation \[eq:Zz\], when either $P_{1}(z)$ or $P_{2}(z)$ is equal to zero, the combined probability $\mathcal{Z}(z)$ also goes to zero. This can be seen visually for the example galaxy pairs in Fig. \[merger-fig:pairs\_pz\] (black line). The total number of fractional pairs for a given system is then given by $$\mathcal{N}_{z} = \int_{0}^{\infty} \mathcal{Z}(z) dz.$$ and can range between 0 and 1. As each initial target galaxy can have more than one close companion, each potential galaxy pair is analysed separately and included in the total pair count. Note that because the initial list of projected pairs is cleaned for duplicates before analysing the redshift posteriors, if the two galaxies in a system (with redshift posteriors of $P_{1}(z)$ and $P_{2}(z)$) both satisfy the primary galaxy selection function, the number of pairs is not doubly counted. [![Example redshift posteriors and integrated $\mathcal{Z}(z)$ for three projected pairs within the DEEP region of the GOODS South fields. In all panels, the blue dashed line corresponds to the redshift PDF for the primary galaxy, while the red dotted line is that of the projected companion. The solid black line shows the cumulative integrated $\mathcal{Z}(z)$ for the galaxy pair. Inset cutouts show the $H_{160}$ image centered on the primary galaxy (with $\operatorname{arcsinh}$ scaling), with the primary and secondary galaxies to match their corresponding $P(z)$. The black circle illustrates the maximum pair search radius at the peak of the primary galaxy $P(z)$.[]{data-label="merger-fig:pairs_pz"}](plots/stamp_4_8321_8089.pdf "fig:"){width="\columnwidth"} ![Example redshift posteriors and integrated $\mathcal{Z}(z)$ for three projected pairs within the DEEP region of the GOODS South fields. In all panels, the blue dashed line corresponds to the redshift PDF for the primary galaxy, while the red dotted line is that of the projected companion. The solid black line shows the cumulative integrated $\mathcal{Z}(z)$ for the galaxy pair. Inset cutouts show the $H_{160}$ image centered on the primary galaxy (with $\operatorname{arcsinh}$ scaling), with the primary and secondary galaxies to match their corresponding $P(z)$. The black circle illustrates the maximum pair search radius at the peak of the primary galaxy $P(z)$.[]{data-label="merger-fig:pairs_pz"}](plots/stamp_7_10336_10261.pdf "fig:"){width="\columnwidth"} ![Example redshift posteriors and integrated $\mathcal{Z}(z)$ for three projected pairs within the DEEP region of the GOODS South fields. In all panels, the blue dashed line corresponds to the redshift PDF for the primary galaxy, while the red dotted line is that of the projected companion. The solid black line shows the cumulative integrated $\mathcal{Z}(z)$ for the galaxy pair. Inset cutouts show the $H_{160}$ image centered on the primary galaxy (with $\operatorname{arcsinh}$ scaling), with the primary and secondary galaxies to match their corresponding $P(z)$. The black circle illustrates the maximum pair search radius at the peak of the primary galaxy $P(z)$.[]{data-label="merger-fig:pairs_pz"}](plots/stamp_27_22371_22279.pdf "fig:"){width="\columnwidth"} ]{} In Fig. \[merger-fig:pairs\_pz\] we show three examples of projected pairs within the DEEP region of CANDELS GOODS South that satisfy the selection criteria applied in this work (Section \[merger-sec:results\]). Two of the the pairs have a high probability of being a real pair within the redshift range of interest ($\mathcal{N}>0.8$) while the third pair (middle panel) has only a partial chance of being at the same redshift. ### Validating photometric line-of-sight probabilities with spectroscopic pairs {#merger-sec:specz_pairs} Due to the relatively high spectroscopic completeness within the CANDELS GOODS-S field thanks to deep surveys such as the MUSE UDF and WIDE surveys [@Bacon:2015eh; @2018arXiv181106549U respectively], precise spectroscopic redshifts are available for a number of close projected pairs within the field. Calculating a mass-selected pair-fraction based on spectroscopic pairs is beyond the scope of this work due to the corrections required for the complicated spectroscopic selection functions. However, the sample of available spectroscopic pairs does allow us to test the reliability of the photo-$z$ based line-of-sight pair probabilities ($\mathcal{N}_{z}$). ![Ratio of total integrated photo-$z$ pairs ($\sum_{i}\mathcal{N}_{z,i}$) to total number of spectroscopic pairs as a function of velocity separation ($\Delta v$) and redshift for projected close pairs within the CANDELS GOODS-S spectroscopic sample. Datapoints for different velocity cuts are offset in redshift for clarity.[]{data-label="merger-fig:spec_pairs"}](plots/specz_photz_pair_ratio.pdf){width="0.95\columnwidth"} After applying a magnitude cut based on the GOODS South completeness limits and a stellar-mass cut on the primary galaxy of $9.7 < \log_{10}({\text{M}_{\star}}/ \text{M}_{\odot})$, we find all potential pairs by searching for other galaxies with spectroscopic redshifts within 30 kpc of each primary galaxy. For each of these potential pairs, we then calculate the integrated number of photo-$z$ pairs, $\mathcal{N}_{z} = \int_{z_{\textup{min}}}^{z_{\textup{max}}} \mathcal{Z}(z) dz$, in four redshift bins from $z=0.5$ to $z=6$. Figure \[merger-fig:spec\_pairs\] shows how the number of integrated photo-$z$ pairs compares to the number of spectroscopic pairs after applying different cuts on velocity separation. We find that the integrated number of photo-$z$ pairs is comparable to the spectroscopic pair counts with velocity separations of up to $<2000~\textup{km~s}^{-1}$ at all redshifts. At low redshift the photo-$z$ pair probabilities over-estimate the number of pairs at separations of $<500~\textup{km~s}^{-1}$, the typical definition used in spectroscopic pair fraction studies, by $\approx 50\%$. However, above $z > 1.5$ we find that the photo-$z$ pairs are fully consistent with the spectroscopic definition within the uncertainties. In Section \[merger-sec:results\] and  \[merger-sec:discussion\] we will discuss how the redshift dependence observed in Figure \[merger-fig:spec\_pairs\] on our final results and the conclusions drawn. The cause of the redshift dependency observed in Figure \[merger-fig:spec\_pairs\] is not immediately clear. Naively, we would expect the increased photo-$z$ scatter/outlier fraction at high redshift to result in the photo-$z$ measurements probing broader velocity offsets. For now, we note that the photo-$z$ pair probabilities are able to effectively probe velocity separations that are a factor of $\approx 3-12\times$ smaller than the scatter within photo-$z$s themselves ($\Delta v = 500~\textup{km~s}^{-1} \approx 0.0017\times(1+z)$) - illustrating the power of the statistical pair count approach. ### Incorporating physical separation and stellar mass criteria The combined redshift probability function defined in Equation \[eq:Zz\] ($\mathcal{Z}(z)$) takes into account only the line-of-sight information for the potential galaxy pair, therefore two additional redshift dependent masks are required to enforce the remaining desired pair selection criteria. These masks are binary masks, equal to one at a given redshift if the selection criteria are satisfied and zero otherwise. As above, we follow the notation outlined in and define the angular separation mask, $\mathcal{M}^{\theta}(z)$, as $$\mathcal{M}^{\theta}(z) = \begin{cases} 1, & \text{ if } \theta_{\textup{min}}(z) \leq \theta \leq \theta_{\textup{max}}(z)\\ 0, & \text{ otherwise, } \end{cases},$$ where the angular separation between the galaxies in a pair as a function of redshift is denoted $\theta (z)$. The angular separation is a function of the projected distance $r_{p}$ and the angular diameter distance, $d_{A}(z)$, for a given redshift and cosmology, i.e. $\theta_{\textup{max}}(z) = r^{\textup{max}}_{p} / d_{A}(z)$ and $\theta_{\textup{min}}(z) = r^{\textup{min}}_{p} / d_{A}(z)$. The pair selection mask, denoted as $\mathcal{M}^{\rm{pair}}(z)$, is where our method differs to that outlined by . Rather than selecting galaxy pairs based on the luminosity ratio, we instead select based on the estimated stellar mass ratio. We define our pair-selection mask as $$\label{eq:sel_mask} \mathcal{M}^{\rm{pair}}(z) = \begin{cases} 1, & \text{ if } {\text{M}_{\star}}^{\rm{lim},1}(z) \leq \rm{M}_{\star,1}(z) \leq \rm{M}_{\star,max} \\ & \text{ and }~ {\text{M}_{\star}}^{\rm{lim},2}(z) \leq \rm{M}_{\star,2}(z) \\ 0, & \text{ otherwise. } \end{cases}$$ where $\rm{M}_{\star,1}(z)$ and $\rm{M}_{\star,2}(z)$ are the stellar mass as a function of redshift, details of how $\rm{M}_{\star}(z)$ is calculated for each galaxies are discussed in Section \[merger-sec:stellarmass\]. The flux-limited mass cuts, ${\text{M}_{\star}}^{\rm{lim},1}(z)$ and ${\text{M}_{\star}}^{\rm{lim},2}(z)$, are given by $$\label{eq:fluxlim_pri} {\text{M}_{\star}}^{\rm{lim},1}(z)= max \{ {\text{M}_{\star}}^{\rm{min}}, {\text{M}_{\star}}^{\rm{flux}}(z) \}$$ and $$\label{eq:fluxlim_sec} {\text{M}_{\star}}^{\rm{lim},2}(z) = max \{ \mu {\text{M}_{\star}}^{1}(z), {\text{M}_{\star}}^{\rm{flux}}(z) \}$$ respectively, where ${\text{M}_{\star}}^{\rm{flux}}(z)$ is the redshift-dependent mass completeness limit outlined in Section \[merger-sec:weights\_flux\] and ${\text{M}_{\star}}^{min}$ and ${\text{M}_{\star}}^{max}$ are the lower and upper ranges of our target sample of interest. The mass ratio $\mu$ is typically defined as $\mu > 1/4$ for major mergers and $1/10 < \mu < 1/4$ for minor mergers. Throughout this work we set $\mu = 1/4$ by default, unless otherwise stated. The pair selection mask ensures the following criteria are met at each redshift: firstly, it ensures the primary galaxy is within the mass range of interest. Secondly, that the mass ratio between the primary and secondary galaxy is within the desired range (e.g. for selecting major or minor mergers). Finally, that both the primary and secondary galaxy are above the mass completeness limit at the corresponding redshift. We note that the first criteria of Equation \[eq:sel\_mask\] also constitutes the selection function for the primary sample, given by $$\label{eq:pri_sel} S(z) = \begin{cases} 1, & \text{ if } {\text{M}_{\star}}^{\rm{lim},1}(z) \leq \rm{M}_{\star,1}(z) \leq \rm{M}_{\star,max} \\ 0, & \text{ otherwise. } \end{cases}$$ [![Redshift-dependent stellar mass estimations for the example close pairs shown in Fig. \[merger-fig:pairs\_pz\]. In all panels the blue dashed line corresponds to the stellar mass for the primary galaxy, while the red dotted line is that of the projected companion. The blue shaded regions illustrate the range of secondary galaxy masses that satisfy the selected merger ratio criteria - $\mu > 1/4$. Dashed gray lines indicate the stellar mass selections applied in this study.[]{data-label="merger-fig:pairs_mass"}](plots/stamp_4_8321_8089_mass.pdf "fig:"){width="\columnwidth"} ![Redshift-dependent stellar mass estimations for the example close pairs shown in Fig. \[merger-fig:pairs\_pz\]. In all panels the blue dashed line corresponds to the stellar mass for the primary galaxy, while the red dotted line is that of the projected companion. The blue shaded regions illustrate the range of secondary galaxy masses that satisfy the selected merger ratio criteria - $\mu > 1/4$. Dashed gray lines indicate the stellar mass selections applied in this study.[]{data-label="merger-fig:pairs_mass"}](plots/stamp_7_10336_10261_mass.pdf "fig:"){width="\columnwidth"} ![Redshift-dependent stellar mass estimations for the example close pairs shown in Fig. \[merger-fig:pairs\_pz\]. In all panels the blue dashed line corresponds to the stellar mass for the primary galaxy, while the red dotted line is that of the projected companion. The blue shaded regions illustrate the range of secondary galaxy masses that satisfy the selected merger ratio criteria - $\mu > 1/4$. Dashed gray lines indicate the stellar mass selections applied in this study.[]{data-label="merger-fig:pairs_mass"}](plots/stamp_27_22371_22279_mass.pdf "fig:"){width="\columnwidth"}]{} With these three properties in hand for each potential companion galaxy around our primary target, the pair-probability function, $\textup{PPF}(z)$, is then given by $$\label{eq:PPF} \textup{PPF}(z) = \mathcal{Z}(z) \times \mathcal{M}^{\theta}(z) \times \mathcal{M}^{\text{pair}}(z).$$ In Fig. \[merger-fig:pairs\_mass\], we show the estimated stellar mass as a function of redshift for the three example projected pairs shown in Fig. \[merger-fig:pairs\_pz\]. Additionally, the redshift ranges where all three additional pair selection criteria are shown by the gray shaded region. For the first and third galaxy pairs with high probability of being a pair along the line-of-sight, the separation criteria and mass selection criteria are also satisfied at the relevant redshift. In contrast, the second potential pair (with $\mathcal{N}_{z} = 0.477$) does not satisfy the stellar mass criteria at all redshifts of interest and therefore has a significantly reduced final pair-probability of $\int_{0}^{\infty} \textup{PPF}(z) dz = 0.238$. In Section \[merger-sec:pair\_frac\] we outline how these individual pair-probability functions are combined to determine the overall pair-fraction, but first we outline the steps taken to correct for selection effects within the data. Correction for selection effects -------------------------------- As defined by , the pair-probability function in Equation \[eq:PPF\] is affected by two selection effects. Firstly, the incompleteness in search area around galaxies that are near the image boundaries or near areas affected by bright stars (Section \[merger-sec:weights\_area\]). And secondly, the selection in photometric redshift quality (Section \[merger-sec:weights\_osr\]). In addition, because in this work we use a flux-limited sample rather than one that is volume limited (as used by ), we must also include a further correction to account for this fact. ### The redshift-dependent mass completeness limit {#merger-sec:weights_flux} Since the photometric survey we are using includes regions of different depth and high-redshift galaxies are by their very nature quite faint, restricting our analysis to a volume-limited sample would necessitate excluding the vast majority of the available data. As such, we choose to use a redshift-dependent mass completeness limit determined by the flux limit determined by the survey. Due to the limited number of galaxy sources available, determining the strict mass completeness continuously as a function of redshift entirely empirically [@Pozzetti:2010gw] is not possible. Instead, we make use of a method based on that of @Pozzetti:2010gw, using the available observed stellar mass estimates to fit a functional form for the evolving 95% stellar mass-to-light limit. Following @Pozzetti:2010gw, the binned empirical mass limit is determined by selecting galaxies which are within a given redshift bin, then scaling the masses of the faintest 20% such that their apparent magnitude is equal to the flux limit. The mass completeness limit for a given redshift bin is then defined as the mass corresponding to the 95th percentile of the scaled mass range. To accurately cover the full redshift range of interest, we apply this method to two separate sets of stellar mass measurements. Firstly at $z\leq 4$ we use the best-fitting stellar masses estimated for each of the CANDELS photometry catalogs used in this work. Secondly, at $z \geq 3.5$ we make use of the full set of high-redshift Monte Carlo samples of @Duncan:2014gh to provide improved statistics and incorporate the significant effects of redshift uncertainty on the mass estimates in this regime. The resulting mass completeness at $z > 1$ in bins with width $\Delta z = 0.5$ are shown in Fig. \[merger-fig:mass\_comp\] assuming a flux-limit equal to the appropriate corresponding ‘WIDE 2’-depth $80\%$ detection completeness limit. Based on the binned empirical completeness limits, we then fit a simple polynomial function to the observed ${\text{M}_{\star}}/\textit{L}$ redshift evolution. By doing so we can estimate the mass completeness as a continuous function of redshift. [ ![*Top:* Mass completeness limit corresponding to the flux limits of the WIDE 2 depth sub-fields in the CANDELS survey. Dark red circles correspond to the 95% completeness limits at $z \geq 3.5$ derived from the stellar mass estimates of @Duncan:2014gh, lighter red circles show the equivalent estimates for the stellar mass estimates of this work for all five fields (smaller circles show estimates for individual fields). The continuous blue line shows the completeness limits corresponding to a maximally old (at a given redshift) single burst stellar population. The functional form (3rd order polynomial) fitted to the empirical mass completeness estimates is shown by the dashed red line. *Bottom:* Estimated mass completeness limits for each of the sub-field depths: the functional form for the $95\%$ stellar mass-to-light limit has been scaled to the $80\%$ detection completeness limit for each sub-field (as determined in Section \[merger-sec:completeness\]). The shaded regions show the range of detection completeness limits covered by the CANDELS fields (Table \[tab:completeness\]) with the area-weighted average for each sub-field depth shown by the solid, dashed and dotted blue lines respectively. Relevant mass selection limits are shown as horizontal red dashed and dotted lines for illustrative purposes. []{data-label="merger-fig:mass_comp"}](plots/MassCompleteness.pdf "fig:"){width="0.95\columnwidth"} ![*Top:* Mass completeness limit corresponding to the flux limits of the WIDE 2 depth sub-fields in the CANDELS survey. Dark red circles correspond to the 95% completeness limits at $z \geq 3.5$ derived from the stellar mass estimates of @Duncan:2014gh, lighter red circles show the equivalent estimates for the stellar mass estimates of this work for all five fields (smaller circles show estimates for individual fields). The continuous blue line shows the completeness limits corresponding to a maximally old (at a given redshift) single burst stellar population. The functional form (3rd order polynomial) fitted to the empirical mass completeness estimates is shown by the dashed red line. *Bottom:* Estimated mass completeness limits for each of the sub-field depths: the functional form for the $95\%$ stellar mass-to-light limit has been scaled to the $80\%$ detection completeness limit for each sub-field (as determined in Section \[merger-sec:completeness\]). The shaded regions show the range of detection completeness limits covered by the CANDELS fields (Table \[tab:completeness\]) with the area-weighted average for each sub-field depth shown by the solid, dashed and dotted blue lines respectively. Relevant mass selection limits are shown as horizontal red dashed and dotted lines for illustrative purposes. []{data-label="merger-fig:mass_comp"}](plots/MassCompleteness_subfields.pdf "fig:"){width="0.95\columnwidth"}]{} A common choice of template for estimating the strict ${\text{M}_{\star}}/\textit{L}$ completeness is a maximally old single stellar population (continuous blue line in the top panel of Fig. \[merger-fig:mass\_comp\], assuming a formation redshift of $z = 12$ and sub-solar metallicity of $Z = 0.2 Z_{\odot}$). However, since the vast majority of galaxies above $z\sim3$ are expected to be actively star-forming, this assumption significantly overestimates the actual completeness mass at high-redshift (hence under-estimating the completeness). The redshift-dependent mass limit, ${\text{M}_{\star}}^{\rm{flux}}(z)$, is defined as $$\label{eq:fluxlim} \log_{10}({\text{M}_{\star}}^{\rm{flux}}(z)) = 0.4\times( H_{{\text{M}_{\star}}/\textit{L}}(z) - H^{\rm{lim}})$$ where $H^{\rm{lim}}$ is the $H_{160}$ magnitude at the flux-completeness limit in the field or region of interest and $H_{{\text{M}_{\star}}/\textit{L}}(z)$ is the $H_{160}$ magnitude at a given redshift of the fitted functional form normalized to 1 $\rm{M}_{\odot}$. In the bottom panel of Fig. \[merger-fig:mass\_comp\] we show the redshift-dependent mass limit corresponding to each of the sub-field depths outlined in Section \[merger-sec:completeness\]. Also shown in this plot are lines corresponding to the stellar mass ranges we wish to probe for major mergers ($\mu > 1/4$) around galaxies with stellar mass of $9.7 < \log_{10} {\text{M}_{\star}}\leq 10.3$ and $\log_{10} {\text{M}_{\star}}\geq 10.3$ (hatched region). For a primary galaxy with a mass close to the redshift-dependent mass-limit imposed by the selection criteria $S(z)$, the mass range within which secondary pairs can be included may be reduced, i.e. $\mu{\text{M}_{\star}}^{1}(z) < {\text{M}_{\star}}^{\rm{lim}}(z) < {\text{M}_{\star}}^{1}(z)$. In Fig. \[merger-fig:mass\_comp\_eg\] we illustrate this for a galaxy with $\log_{10} {\text{M}_{\star}}\approx 10.3$ in the redshift range $2.5 < z \leq 3$ (red) and a $\log_{10} {\text{M}_{\star}}\approx 9.7$ at $1.5 < z < 2$ (green). The darker shaded regions shows the area in the parameter space of $z$ vs ${\text{M}_{\star}}$ where potential secondary galaxies with merger ratios $>1/4$ are excluded by the redshift-dependent mass-completeness cut. To correct for the potential galaxy pairs that may be lost by the applied completeness limit, we make a statistical correction based on the stellar mass function at the redshift of interest - analogous to the luminosity function-based corrections first presented in @Patton:2000kt. The flux-limit weight, $w^{\text{flux}}_{2}(z)$, applied to every secondary galaxy found around each primary galaxy, is defined as $$w^{\text{flux}}_{2}(z) = \frac{1} {W_{2}(z)},$$ where $$W_{2}(z) = \frac{ \int_{{\text{M}_{\star}}^{\text{lim}}(z)}^{{\text{M}_{\star}}^{\text{1}}(z)} \phi({\text{M}_{\star}}|z)d{\text{M}_{\star}}} { \int_{\mu {\text{M}_{\star}}^{1}(z)}^{{\text{M}_{\star}}^{\text{1}}(z)} \phi({\text{M}_{\star}}|z)d{\text{M}_{\star}}}$$ and $ \phi({\text{M}_{\star}}|z) $ is the stellar mass function at the corresponding redshift. The redshift-dependent mass limit is ${\text{M}_{\star}}^{\text{lim}}(z) = max \{ \mu {\text{M}_{\star}}^{1}(z), {\text{M}_{\star}}^{\rm{flux}}(z) \} $, where ${\text{M}_{\star}}^{\rm{flux}}(z)$ is defined in Equation \[eq:fluxlim\] (dashed blue line in Fig. \[merger-fig:mass\_comp\_eg\]). By applying this weight to all pairs associated with a primary galaxy, we get the pair statistics corresponding to $\mu {\text{M}_{\star}}^{1}(z) \leq {\text{M}_{\star}}^{\text{2}}(z) \leq {\text{M}_{\star}}^{\text{1}}(z)$ (the volume limited scenario, e.g. the total red or green shaded areas in Fig. \[merger-fig:mass\_comp\_eg\]). Note that because this correction is based on the statistically expected number density of galaxies as a function of mass, representative numbers of detected secondary galaxies above the completeness limit are still required. [ ![Illustration of the parameter space where the statistical stellar-mass completeness correction is in effect. The example illustrates the relevant mass limits and selection ranges for a redshift bin of two different bins within the ‘WIDE 2’ sub-fields: a primary mass selection of $9.7 < \log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) < 10.3$ at $1.5 < z < 2$ (green) and a $\log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) > 10.3$ selection at $2.5 < z < 3$ (red).[]{data-label="merger-fig:mass_comp_eg"}](plots/MassCompleteness_corr_eg.pdf "fig:"){width="0.95\columnwidth"}]{} As in @Patton:2000kt, we also assign additional weights to the primary sample in order to minimize the error from primary galaxies that are closer to the flux limit (i.e. with redshift posteriors weighted to higher redshifts) as these galaxies will have fewer numbers of *observed* pairs. The primary flux-weight, $w_{\text{flux}}^{1}(z)$ is defined as $$w^{\text{flux}}_{1}(z) = W_{1}(z) = \frac{ \int_{{\text{M}_{\star}}^{\text{lim}}(z)}^{{\text{M}_{\star}}^{\rm{max}}} \phi({\text{M}_{\star}}|z)d{\text{M}_{\star}}} { \int_{{\text{M}_{\star}}^{\rm{min}}}^{{\text{M}_{\star}}^{\rm{max}}} \phi({\text{M}_{\star}}|z)d{\text{M}_{\star}}}$$ where ${\text{M}_{\star}}^{\rm{min}}$ and ${\text{M}_{\star}}^{\rm{max}}$ are the lower and upper limits of the mass range of interest for the primary galaxy sample, the redshift-dependent lower limit is defined as $ {\text{M}_{\star}}^{\text{lim}}(z) = max \{ {\text{M}_{\star}}^{\rm{min}}, {\text{M}_{\star}}^{\rm{flux}}(z) \} $, and the remaining parameters are as outlined above. For volume-limited samples (where ${\text{M}_{\star}}^{\rm{flux}}(z) < \mu{\text{M}_{\star}}^{1}(z) $ at all redshifts) both of the flux-limit weights are equal to unity. The stellar mass functions (SMF) parameterizations as a function of redshift, $ \phi({\text{M}_{\star}}|z) $, are taken from @Mortlock:2014et at $z \leq 3$, @Santini:2012jq at $3 < z < 3.5$ and @Duncan:2014gh at $z \geq 3.5$. When selecting redshift bins in which to estimate the merger fraction, we ensure that the bins are chosen to match the bins in which the SMF are constrained (i.e. the SMF used to weight the merger fraction is the same across the bin). Tests performed when applying the same methodology to wide-area datasets in @mundy2017 indicate that results are robust to the choice of specific SMF and that results presented later in the paper would not be significantly affected if alternative SMF are assumed. Furthermore, we note that this correction assumes that the shape of the SMFs for satellite galaxies does not differ from those measured for the full population. Observational constraints at low redshift indicate that such an assumption is valid [@2016MNRAS.459.2150W], but direct constraints at higher redshift are not currently available. ### Image boundaries and excluded regions {#merger-sec:weights_area} A second correction which must be taken into account is to the search area around primary galaxies that lie close to the boundaries of the survey region. Because of the fixed physical search distance, this correction is also a function of redshift, so it must be calculated for all redshifts within the range of interest. In addition to the area lost at the survey boundaries, it is also necessary to correct for the potential search area lost due to the presence of large stars and other artefacts, around which no sources are included in the catalog (see Section \[merger-sec:initial\]). We have taken both of these effects into account when correcting for the search areas by creating a mask image based on the underlying photometry mosaics. Firstly, we define the image boundary based on the exposure map corresponding to the $H_{160}$ photometry used for object detection. Next, for every source excluded from the sample catalog based on its classification as a star or image artefact by our photometric or visual classification, the area corresponding to that object (from the photometry segmentation map) is set to zero in our mask image. Finally, areas of photometry which are flagged in the flag map (and excluded based on their corresponding catalog flags) are also set to zero. To calculate the area around a primary galaxy that is excluded by these effects, we perform aperture ‘photometry’ on the generated mask images. Photometry is performed in annuli around each primary galaxy target, with inner and outer radii of $\theta_{\rm{min}}(z)$ and $\theta_{\rm{max}}(z)$ respectively. The area weight is then defined as $$w_{\rm{area}}(z) = \frac{1} {f_{ \rm{area}}(z)}$$ where $f_{ \rm{area}}(z)$ is the sum of the normalized mask image within the annulus at a given redshift divided by the sum over the same area in an image with all values equal to unity. By measuring the area in this way we are able to automatically take into account the irregular survey shape and any small calculation errors from quantization of areas due to finite pixel size. Despite the relatively small survey area explored in this study (and hence a higher proportion of galaxies likely to lie near the image edge), the effect of the area weight on the estimated pair fractions is very small. To quantify this, we calculate the pair averaged area weights, $\left \langle w_{\rm{area}} \right \rangle$, such that $$\left \langle w_{\rm{area}}^{i,j} \right \rangle = \frac{\int \text{PPF}^{i,j}(z) w_{\rm{area}}^{i}(z) \rm{d}z}{\int \text{PPF}^{i,j}(z) \rm{d}z},$$ where $w_{\rm{area}}^{i}(z)$ is the redshift dependent area weight for a primary galaxy $i$, and $\text{PPF}^{i,j}(z)$ the corresponding pair-probability function for primary galaxy and a secondary galaxy $j$. Of the full sample of primary galaxies, less than $10\%$ have average area weights greater than 1.01 (where a primary galaxy has multiple pairs, we take the average of $\left \langle w_{\rm{area}}^{i,j} \right \rangle$ over all secondary galaxies). Furthermore, only $\approx 2\%$ of primary galaxies have average weights $\left \langle w_{\rm{area}}^{i,j} \right \rangle > 1.1$ and only $0.15\%$ have weights $>1.5$ (e.g. sources which lie very close to the edge of the survey field). The effects of area weights on the final estimated merger fractions will therefore be minimal. Nevertheless, we include these corrections in all subsequent analysis. ### The Odds sampling rate {#merger-sec:weights_osr} In the original method outlined in , and also applied in @mundy2017, an additional selection based on the photometric redshift quality, or odds $\mathcal{O}$ parameter. The original motivation for this additional selection criteria (and subsequent correction), as outlined partially in @Molino:2014iz, is that by enforcing the odds cut they are able to select a sample for which the posterior uncertainties are accurate. Due to the extensive magnitude dependent photo-$z$ posterior calibration applied in this work and the fact that our resulting redshift posteriors are well calibrated at all magnitudes, we do not include this additional criteria. Therefore, we do not apply the additional odds sampling rate weighting terms outlined in @mundy2017. ### The combined weights Taking both of the above effects into account, the pair weights for each secondary galaxy found around a galaxy primary are given by $$\label{eq:flux_weights_2} w_{\text{2}}(z) = w_{1,\text{area}}(z) \times w_{1,\text{flux}}(z) \times w_{2,\text{flux}}(z) $$ The weights applied to every primary galaxy in the sample are then given by $$\label{eq:flux_weights_1} w_{\text{1}} = w_{1,\text{flux}}(z) $$ These weights are then applied to the integrated pair-probability functions for each set of potential pairs to calculate the merger fraction. The greatest contribution to the total weights primarily comes from the secondary galaxy completeness weights, $w_{2,\text{flux}}(z)$, with additional non-negligible contributions from the primary completeness. Furthermore, the largest additional uncertainty in the total weights results from the mass completeness weights. Final integrated pair fractions {#merger-sec:pair_frac} ------------------------------- With the pair probability function and weights calculated for all potential galaxy pairs, the total integrated pairs fractions can then be calculated as follows. For each galaxy, $i$, in the primary sample, the number of associated pairs, $N_{\rm{pair}}^{i}$, within the redshift range $z_{\rm{min}} < z < z_{\rm{max}}$ is given by $$N_{\text{pair}}^{i} = \sum_{j} \int_{z_{\rm{min}}}^{z_{\rm{max}}} w_{\text{2}}^{j}(z)\times \text{PPF}_{j}(z) dz$$ where $j$ indexes the number of potential close pairs found around the primary galaxy, $\text{PPF}_{j}(z)$ the corresponding pair-probability function (Equation \[eq:PPF\]) and $w_{\text{2},j}(z)$ its pair weight (Equation \[eq:flux\_weights\_2\]). The corresponding weighted primary galaxy contribution, $N_{1,i}$, within the redshift bin is $$N_{1,i} = \sum_{i}\int_{z_{\rm{min}}}^{z_{\rm{max}}} w_{1,i}(z) \times P_{i}(z) \times S_{1,i}(z){} dz$$ where $S_{1,i}(z)$ is the selection function for the primary galaxies given in Equation \[eq:pri\_sel\], $P_{i}(z)$ its normalized redshift probability distribution and $w_{1,i}$ its weighting. In the case of a primary galaxy with stellar mass in the desired range with its redshift PDF contained entirely within the redshift range of interest, $N_{1,i} = w_{1,i}$, and hence always equal or greater than unity. The estimated pair fraction $f_{\rm{P}}$ is defined as the number of pairs found for the target sample divided by the total number of galaxies in that sample. In the redshift range $z_{\rm{min}} < z < z_{\rm{max}}$, $f_{\rm{P}}$ is then given by $$f_{\rm{P}} = \frac{\sum_{i} N_{\text{pair},i}}{\sum_{i} N_{1,i}}$$ where $i$ is summed over all galaxies in the primary sample. For a field consisting of different sub-fields, this sum becomes $$f_{\rm{P}} = \frac{\sum_{k}\sum_{i} N_{\text{pair},k,i}}{\sum_{k}\sum_{i} N_{1,k,i}}$$ where $k$ is indexed over the number sub-fields (e.g. 4: ‘Wide 1’, ‘Wide 2’, ‘Deep’ and ‘Ultra Deep’). The mass completeness limit used throughout the calculations is set by the corresponding $H_{160}$ depth within each field. Results {#merger-sec:results} ======= In this section we investigate the role of mergers in forming massive galaxies up to $z \approx 6$. We first investigate and describe a purely observationally quantity, the pair fraction, using the full posterior pair-count analysis described in the previous section, within eight redshift bins from $z = 0.5$ to $z = 6.5$. We carry this out within stellar mass cuts of $9.7 < \log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) < 10.3$ and $\log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) > 10.3$. We also perform the pair searches in annuli with projected separations of $5 \leq r_{\text{p}} \leq 30$. The minimum radius of 5 kpc is typically used in pair counting studies to prevent confusion of close sources due to the photometric or spectroscopic fibre resolution. Although the high-resolution HST photometry allows for reliable deblending at radii smaller than this [@2007PASP..119.1325L; @Galametz:2013dd], we adopt this radius for consistency with previous results. Later in this section, we then calculate observational constraints placed on merger *rates* for these galaxies, using physically motivated merger-time scales to explore both the merger rate per galaxy and the merger rate density over time since $z = 6$. Evolution of the major pair fraction {#merger-sec:mergerfraction} ------------------------------------ ### Observed pair fractions in CANDELS In this section we present measurements of the observed pair fraction, $f_{\text P}$ of massive galaxies from $z=0.5$ to $z \sim 6$ in the combined CANDELS multi-wavelength datasets. Our results are shown in Fig. \[merger-fig:merger\_frac\], where we plot our derived pair fractions for each of the five fields as well the overall constraints provided by the combined measurements. The measured values and their corresponding statistical errors are presented in Table \[tab:fmerger\]. The errors on our $f_\text{P}$ values are estimated using the common bootstrap technique of @Efron:1979uf [@EFRON:HJ2mD4hg]. The standard error, $\sigma_{f_{\text{P}}}$, is defined as $$\label{eq:bootstrap_err} \sigma_{f_{\rm{P}}} = \sqrt{\frac{\sum_{i,N} \left (f_{\text{m},i} - \left \langle f_{\text{P}} \right \rangle \right)^{2}} {(N-1)}, }$$ where $f_{\text{P}}^{i}$ is the estimated merger fraction for a randomly drawn sample of galaxies (with replacement) from the initial sample (for $N$ independent realisations) and $\left \langle f_{\text{P}} \right \rangle = \left( \sum_{i} f_{\text{P},i} \right) / N $. {width="1.4\columnwidth"} {width="1.4\columnwidth"} Only regions (i.e. ‘Wide 1’, ‘Wide 2’, ‘Deep’ and ‘Ultra Deep’) that are complete in stellar mass to the primary galaxy selection mass at the bin upper redshift limit are included in the estimate for a given field. The same completeness cuts are applied when calculating the combined ‘All CANDELS’ estimates, with only the contributing datapoints plotted in Fig. \[merger-fig:merger\_frac\]. When calculating the combined pair fraction estimates, we include only one measurement from GOODS North, specifically the estimates incorporating the SHARDS medium-band photometry. As can be seen in Fig. \[merger-fig:merger\_frac\], there is a variance in the derived pair fraction across the five CANDELS fields. However, given the statistical uncertainties within each field, we find that the individual measurements are consistent across the wide range in redshifts. In all fields, we find a systematic trend with redshift, such that the pair fraction increases towards higher redshifts for primary galaxies in both the $9.7 < \log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) < 10.3$ and $\log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) > 10.3$ mass selected samples. In the lower stellar mass bin explored in this work, the fall in completeness for the shallower CANDELS fields is evident at higher redshifts, with constraints provided primarily by the Hubble Ultra-Deep Field region within the GOODS South field. However, overall we find that the pair counts for the lower mass range show a similar increase in the pair fraction up to until $z \sim 3$. Above this redshift the constraints are limited to measurements of the upper limit, i.e. finding no significant probability of pairs around the small number of galaxies that lie in the mass-complete sample . [cccccccc]{} $0.5 \leq z < 1.0$ & $0.063 \pm 0.020$ & $0.073 \pm 0.018$ & $0.061 \pm 0.017$ & $0.069 \pm 0.015$ & $0.046 \pm 0.016$ & $0.084 \pm 0.022$ & $0.065 \pm 0.007$\ $1.0 \leq z < 1.5$ & $0.074 \pm 0.021$ & $0.091 \pm 0.023$ & $0.078 \pm 0.021$ & $0.125 \pm 0.023$ & $0.088 \pm 0.021$ & $0.140 \pm 0.024$ & $0.104 \pm 0.010$\ $1.5 \leq z < 2.0$ & $0.133 \pm 0.032$ & $0.132 \pm 0.026$ & $0.146 \pm 0.033$ & $0.188 \pm 0.030$ & $0.142 \pm 0.025$ & $0.200 \pm 0.027$ & $0.167 \pm 0.012$\ $2.0 \leq z < 2.5$ & $0.201 \pm 0.056$ & $0.123 \pm 0.051$ & $0.117 \pm 0.047$ & $< 0.913$ & $0.241 \pm 0.049$ & $0.250 \pm 0.039$ & $0.222 \pm 0.023$\ $2.5 \leq z < 3.0$ & $0.224 \pm 0.064$ & $0.198 \pm 0.069$ & $0.206 \pm 0.057$ & & & & $0.216 \pm 0.041$\ $3.0 \leq z < 3.5$ & $< 0.384$ & & & & & & $< 0.387$\ \ $z$ & GS & GN & GN (SHARDS) & COSMOS & UDS & EGS & All\ $0.5 \leq z < 1.0$ & $0.041 \pm 0.025$ & $0.068 \pm 0.023$ & $0.044 \pm 0.019$ & $0.061 \pm 0.016$ & $0.065 \pm 0.022$ & $0.053 \pm 0.019$ & $0.054 \pm 0.008$\ $1.0 \leq z < 1.5$ & $0.047 \pm 0.022$ & $0.067 \pm 0.025$ & $0.051 \pm 0.020$ & $0.049 \pm 0.021$ & $0.055 \pm 0.019$ & $0.075 \pm 0.019$ & $0.057 \pm 0.008$\ $1.5 \leq z < 2.0$ & $0.049 \pm 0.025$ & $0.084 \pm 0.027$ & $0.075 \pm 0.027$ & $0.092 \pm 0.024$ & $0.069 \pm 0.021$ & $0.147 \pm 0.028$ & $0.093 \pm 0.011$\ $2.0 \leq z < 2.5$ & $0.087 \pm 0.050$ & $0.053 \pm 0.029$ & $0.050 \pm 0.026$ & $0.112 \pm 0.040$ & $0.054 \pm 0.026$ & $0.150 \pm 0.045$ & $0.090 \pm 0.015$\ $2.5 \leq z < 3.0$ & $0.085 \pm 0.047$ & $0.150 \pm 0.060$ & $0.119 \pm 0.045$ & $0.120 \pm 0.058$ & $0.083 \pm 0.040$ & $0.099 \pm 0.050$ & $0.100 \pm 0.019$\ $3.0 \leq z < 3.5$ & $0.119 \pm 0.102$ & $0.270 \pm 0.140$ & $0.156 \pm 0.099$ & $0.220 \pm 0.126$ & $0.109 \pm 0.070$ & $0.097 \pm 0.078$ & $0.141 \pm 0.038$\ $3.5 \leq z < 4.5$ & $< 0.182$ & $0.249 \pm 0.153$ & $< 0.214$ & $0.197 \pm 0.126$ & $< 0.168$ & $0.101 \pm 0.078$ & $0.118 \pm 0.038$\ $4.5 \leq z < 5.5$ & $< 0.307$ & $0.437 \pm 0.190$ & $0.229 \pm 0.186$ & $< 0.334$ & $0.410 \pm 0.337$ & $0.184 \pm 0.152$ & $0.221 \pm 0.081$\ $5.5 \leq z < 6.5$ & $< 0.549$ & $< 0.301$ & $< 0.275$ & & $0.783 \pm 0.552$ & $0.386 \pm 0.238$ & $0.374 \pm 0.146$\ \[tab:fmerger\] ### Comparison to literature A large number of previous studies have explored the redshift evolution of galaxy pair counts in mass or (absolute) magnitude selected samples [@LeFevre:2000iq; @Conselice:2003jz; @Kartaltepe:2007dv; @Bluck:2009in; @Bluck:2012dh; @Bundy:2009jw; @LopezSanjuan:2010cz; @LopezSanjuan:2014uj; @Man:2011jo; @2016ApJ...830...89M; @2014MNRAS.444..906F]. However, these past studies employ a wide range of criteria in selecting close pairs (mass ranges, separation radius, line-of-sight selection/correction methods), making it difficult to direct comparisons with the observations presented in this work. The majority of merger rate studies typically focus on the most massive galaxies, i.e. $\log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) > 11$. For studies at $z > 1$, such massive galaxies are above our typical flux and mass completeness limits and are bright enough for obtaining accurate spectroscopic redshift, they therefore represent the most robust samples studied to date [@Bluck:2009in; @Man:2011jo]. However, given that these massive galaxies are increasingly rare at higher redshifts [@Ilbert:2013dq; @Muzzin:2013bl; @Mortlock:2014et; @Duncan:2014gh], the small field of view of the CANDELS fields does not probe a large enough volume to detect statistically significant samples of these galaxies. We are therefore unable to compare our results with these previous works at the same mass limit $\log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) > 11$, irrespective of any difference in pair selection radii. Nevertheless, a range of literature results that select galaxy pairs with comparable mass and pair separation criteria exist. In Fig. \[merger-fig:merger\_frac2\] we plot the combined CANDELS major merger pair count observations presented in this work alongside other published measurements that employ the same mass limits and projected separation cuts. From @mundy2017 we plot the pair fractions for the three wide area optical surveys used in that work for a primary galaxy mass cut of $\log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) > 10.3$ following the same method employed in this paper (priv. communication). Additionally, we plot the recent results of @2018MNRAS.475.1549M who employ a different pair count methodology to the same underlying CANDELS photometric datasets. To illustrate the latest results on spectroscopic pair counts at high redshift, in the upper panel of Fig. \[merger-fig:merger\_frac2\] we plot the major pair fractions presented by for spectroscopically selected pairs with separation $<25$ kpc and primary galaxy stellar mass $\log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) > 9.5$ (median masses from $9.9 \lesssim \log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) \lesssim 10.3$). In the lower panel of Fig. \[merger-fig:merger\_frac2\] we plot the pair fraction over the redshift $1.9 < z < 4$ as presented by @Tasca:2014gz, with pairs also defined by $<25$ kpc separation and a median primary galaxy mass of $\log_{10}({\text{M}_{\star}}/ \text{M}_{\odot})= 10.3$. Both sets of spectroscopic measurements are in good agreement with the higher pair fractions measured in this work. [^2] Finally, we also plot the parameterized pair fraction evolution calculated for the EAGLE [@Schaye:2014gk] hydrodynamical simulation presented by @Qu:2016hd. Although the mass limits and merger ratio selections presented in @Qu:2016hd match closely the ranges explored in this work, we note that the pair separation criteria employed are dependent on the half-stellar mass of each primary galaxy and are therefore mass and redshift dependent (typically between 10 and 30 kpc for the redshift and mass range presented here). We therefore caution against over-interpretation of any comparison between the simulation results and those presented in this work. {width="1.4\columnwidth"} {width="1.4\columnwidth"} In addition to the literature comparison, in Fig. \[merger-fig:merger\_frac2\] we also plot our best-fit parameterization of the observations presented in this work. The redshift evolution of the galaxy pair fraction has been previously parametrized in a number of ways, but primarily as a power-law with respect to $(1+z)$ such that the observed pair fraction goes as $$\label{eq:power} f_{\rm P}(z) = f_{0} \times (1+z)^{\rm m}.$$ However, other studies have found that the pair, and thus inferred merger, fraction shows evidence of a decline at redshifts higher than around $z \sim 1.5$ to $z\sim2.5$ [e.g. @Conselice:2008de; @2016ApJ...830...89M; @2018MNRAS.475.1549M]. To test whether there is any statistical evidence for a turn-over in the pair fraction at high redshift we therefore fit both the power-law form and a two-component model of a power-law form and an exponential: $$\label{eq:power+exp} f_{\rm P}(z) = f_{0} \times (1+z)^{\rm m} \times \exp(\tau (1+z)).$$ We fit these two models to the observational results in both mass ranges using a likelihood-based regression optimised through Markov chain Monte Carlo fitting [@ForemanMackey:2013io] and incorporating an additional intrinsic scatter term, $s$, within the uncertainties such that $\sigma^{2}_{\textup{tot}} = \sigma_{f_{P}}^{2} + s^{2}f_{\rm P}(z)^{2}$. In all fits we use a permissive prior that is flat in linear space with very broad boundary conditions for the shape parameters and a flat log prior for the intrinsic scatter, $s$. The resulting median values and marginalized 1-$\sigma$ uncertainties for both sets of parameterizations are presented in Table \[tab:fitvalues\] alongside the Bayesian Information Criterion (BIC) for each fit. [cccccc]{}\ $9.7 < \log_{10} \textup{M}_{\star} < 10.3$ & $0.024^{+0.005}_{-0.004}$ & $1.775^{+0.205}_{-0.196}$ & - & $0.009^{+0.100}_{-0.009}$ & -121.5\ $\log_{10} \textup{M}_{\star} > 10.3$ & $0.032^{+0.009}_{-0.007}$ & $0.844^{+0.216}_{-0.235}$ & - & $0.002^{+0.036}_{-0.002}$ & -218.1\ & & & &\ \ $9.7 < \log_{10} \textup{M}_{\star} < 10.3$ & $0.030^{+0.009}_{-0.007}$ & $4.431^{+1.721}_{-1.590}$ & $-1.028^{+0.621}_{-0.672}$ & $0.010^{+0.094}_{-0.010}$ & -120.2\ $\log_{10} \textup{M}_{\star} > 10.3$ & $0.033^{+0.008}_{-0.007}$ & $0.439^{+1.085}_{-0.939}$ & $0.131^{+0.291}_{-0.363}$ & $0.001^{+0.024}_{-0.001}$ & -214.7\ \[tab:fitvalues\] Based on the BIC, we find that there is no strong statistical evidence ($\Delta\textup{BIC} > 10$) for a power-law plus exponential form for the evolution of the pair fraction in either mass bin. Rather, we find that the two models are formally indistinguishable ($0 < \Delta\textup{BIC} < 4$) given our statistical uncertainties. This result is in contrast to the conclusions drawn by @2018MNRAS.475.1549M from pair count measurements based on the same underlying datasets. We attribute this difference primarily to the incorporation of flux-limit corrections that account for pairs which are un-observed due to selection effects [as is also done in @mundy2017]. We note that in choosing to fit the power-law distribution to binned data, we are potentially subject to biases in the best-fitting power-law slope [@2004EPJB...41..255G; @2007EPJB...58..167B]. Quantitative comparison of the best-fitting slopes should therefore be made with this caveat in mind. However, our key conclusions regarding the statistical evidence for or against a redshift turnover are robust to this problem. ### The effects of photometric redshift precision on measured pair-counts {#merger-sec:pz_precision} In Fig. \[merger-fig:merger\_frac\] we present pair fraction measurements for the CANDELS GOODS North field using two separate photo-$z$ estimates, both with and without the inclusion of the SHARDS medium-band photometry [@2013ApJ...762...46P]. As illustrated in Fig. \[merger-fig:specz\_comp1\], the photo-$z$ estimates incorporating SHARDS are $\sim5\times$ more precise at $z \lesssim 1.5$ than those without. We are therefore able to explore the effect of redshift precision on the results obtained by our pair-count methodology given the same galaxy sample. Across all redshift bins, we find that the observed pair fractions between both GOODS North estimates are in agreement within the statistical uncertainties. However, the GOODS North + SHARDS pair fractions are systematically lower by $\approx 30\%$ on average at these redshifts - comparable to the scatter observed between different CANDELS fields. To further investigate the effect of redshift uncertainty and the reliability of our pair-count method, we perform an additional test to investigate the potential for residual contamination of the observed pair-counts by chance line-of-sight projections. Previous attempts to estimate pair-counts using photo-$z$s have estimated the number of true galaxy pairs by subtracting a statistical estimate of the number of random line-of-sight pairs from the observed pair counts. This correction is typically done using Monte Carlo simulations where the source positions have been randomized across the field, [e.g. @Kartaltepe:2007dv; @2018MNRAS.475.1549M]. In @2018MNRAS.475.1549M, the *chance* pairs at separations of $<30$ kpc were found to contribute between $\sim75$ to 85$\%$ of the observed pairs for a stellar mass cut of $\log_{10} \textup{M}_{\star} > 10.3$. A key advantage of our method is that it does not treat the projected pairs as binary, i.e. contributing either 0 or 1 to the pair count. Rather, the probabilistic pair-count accounts for the fact that even though the 1-$\sigma$ photo-$z$ uncertainties of two galaxies may overlap, the integrated possibility of the two galaxies being at the same redshift will be less than one.[^3] If the method is performing as designed, chance projected pairs that are unassociated should therefore not contribute significantly to the pair count. However, as illustrated by the comparison with spectroscopic pairs in Section \[merger-sec:specz\_pairs\], there may still remain some contamination at low-redshift from chance projections due to imperfect or outlier photo-$z$s. Due to the inhomogeneity in depth across many of the CANDELS fields, creation of fully releastic random catalogs that account for the variation in depth (and hence relative source-counts) are non trivial. We therefore perform our test on EGS as it is the most homogeneous field with more than 80% of its area having almost identical $H_{160}$ limiting magnitudes and the remaining area having very similar depths. These results can be generalized accross all of our CANDELS fields. To estimate the residual contamination from un-associated projected pairs, we produce 10 catalogs where the source positions have been re-drawn randomly from within the $H_{160}$ observation footprint and run the full pair-count analysis for the $\log_{10} \textup{M}_{\star} > 10.3$ stellar mass cut. The background contamination is then estimated based on the median pair-count over the 10 random catalogs. Averaged over all redshift bins, we find that the random pairs can account for 29% of the observed pairs in this field – directly comparable to the difference we see for the high-precision SHARDS sample compared to the broadband only measurements. This fraction also represents a conservative upper limit due to increased signal from the larger scale clustering at a given redshift over the field (while positions were randomized, the redshift distributions still represent those of the small survey area). Regardless, the maximum size of this effect is not large. When fitting the power-law and power-law plus exponential models to the EGS field data points alone, we find that our conclusions on the redshift evolution of the pair fraction are unchanged. The best-fitting power law for the EGS pair-fractions before subtracting the contamination is $$f(z) = 0.045^{+0.019}_{-0.014} \times (1+z)^{0.762^{+0.328}_{-0.359}}.$$ While after subtracting the contamination for chance pairs we find $$f(z) =0.043^{+0.026}_{-0.017} \times (1+z)^{0.488^{+0.474}_{-0.545}}.$$ The power-law only parametrisation remains the best fit after subtraction of the random pairs, but formally the two models are still statistically indistinguishable ($\Delta\textup{BIC} = 2.7$). As this effect is not large enough to affect any of the conclusions presented in the following section and has not been applied to previous [@mundy2017], we do not apply the correction to the full pair fraction results. In Section \[merger-sec:discussion\], we discuss further how this systematic might effect the conclusions on the merger history of massive galaxies. Minor merger pair fractions {#merger-sec:minor} --------------------------- Minor mergers, with mass ratios between 10:1 and 4:1, are predicted in some galaxy formation models as one of the dominant ways in which mass is added to massive galaxies. However, almost no direct observational information is available to determine the role of minor mergers [some studies such as @Ownsworth:2014gt observationally infer their importance]. This quantity was previously examined in more massive galaxies by @Bluck:2012dh for the GOODS NICMOS Survey, and more recently by @2016ApJ...830...89M and @mundy2017. The depth of CANDELS data used in this work means we can investigate the pair fraction for galaxies in our sample down to mass ratios as low as 20:1 or lower. While we are not able to measure these ratios out to our highest redshifts of $z \sim 6$ due to the mass completeness limits, we can investigate the evolution of these minor pairs over the epoch of peak galaxy formation ($z \lesssim 3$). [ {width="100.00000%"}]{} In Fig. \[merger-fig:merger\_frac\_mu\] we show the measured *cumulative* pair fraction for five different redshift ranges between $0.5 \leq z < 3$. We plot these pair fractions $f_{\rm P}$ as a function of the mass ratio $\mu =$ M$_{\star,\rm pri}$/M$_{\star,\rm sec}$ where ‘pri’ and ‘sec’ denote the stellar mass of the more and less massive galaxy involved in the merger, respectively. As expected, we find that the cumulative merger fraction smoothly increases with mass ratio. To parametrize the pair fractions as a function of mass ratio $\mu$, we fit the following functional form for each redshift bin: $$\label{eq:mu_func} f_{\rm P}(>\mu) = A \times \left(\frac{1}{\mu} -1 \right)^{B}.$$ Table \[tab:minor\_func\] shows the corresponding parameter fits for each of the redshift bins. As can be seen through these fits, there is no significant change in the slope of this relation between merger mass ratio and the resulting pair fraction. The shallow slope we find for the *cumulative* pair fractions indicates that at larger mass ratio differences (smaller $\mu$), the observed pair fraction decreases for greater mass ratios (more minor mergers). This result qualitatively confirms the findings of @2016ApJ...830...89M and @mundy2017 for more massive samples of galaxies. [ccc]{} $0.5 \leq z < 1.0$ & $-1.472^{+0.037}_{-0.040}$ & $0.413^{+0.042}_{-0.041}$\ $1.0 \leq z < 1.5$ & $-1.522^{+0.037}_{-0.039}$ & $0.540^{+0.040}_{-0.039}$\ $1.5 \leq z < 2.0$ & $-1.291^{+0.032}_{-0.033}$ & $0.515^{+0.041}_{-0.040}$\ $2.0 \leq z < 2.5$ & $-1.299^{+0.046}_{-0.051}$ & $0.491^{+0.078}_{-0.076}$\ $2.5 \leq z < 3.0$ & $-1.346^{+0.065}_{-0.078}$ & $0.582^{+0.160}_{-0.147}$\ \[tab:minor\_func\] Similarly, within this range we also do not see a significant decline in the values for the normalization ($A$), such that the observed history of galaxy pairs over this redshift range from $0.5 < z < 3$ is fairly constant, as seen previously in the redshift evolution of the pair fraction for major mergers (Fig. \[merger-fig:merger\_frac2\]). This suggests that minor mergers are following the major mergers in terms of their commonality at these redshifts. Evolution of galaxy merger rates {#merger-sec:mergerrate} -------------------------------- The major or minor merger pair fraction is a purely observational quantity and not a fundamental parameter for deriving evolution (such as the star-formation rate). Furthermore, comparisons of pair fractions between different redshift bins and methodologies can be difficult, as different methods of finding mergers having different time-scale sensitivities. This is analogous to measuring the star-formation rate using e.g., UV fluxes, H$\alpha$ fluxes or FIR fluxes. Each flux is a representation of some aspect of the star-formation rate, but each one is sensitive only to certain types of stars and over certain time-scales. Thus the conversion between flux and star-formation rate for these different fluxes has to be done differently for each method. Likewise, a similar situation exists when examining pair and merger fractions measured using different mass/luminosity criteria, different separations and when using pairs or structure/morphology. A more fundamental property of interest is therefore the merger rate, either the average time between mergers per galaxy ($\mathcal{R}$) or the overall merger rate, specifically the merger rate density measured in units of co-moving Mpc$^{3}$ (denoted $\Gamma$ in this work). ### Major merger rates Conversion of the observed pair fraction to a merger rate per galaxy is typically defined as $$\label{eq:old_rate_conv} \mathcal{R} (>{\text{M}_{\star}}, z) = \frac{f_{\text{P}}(>{\text{M}_{\star}}, z) \times C_{\rm merg}}{\tau_{\text{m}}(z)}$$ where $f_{\text{P}}(>{\text{M}_{\star}}, z)$ is the pair fraction at redshift $z$ and masses greater than ${\text{M}_{\star}}$ (Section \[merger-sec:mergerfraction\]), $C_{\rm merg}$ is the average fraction of those pairs that will eventually merge into a single galaxy and $\tau_{\text{m}}(z)$ the corresponding merger timescale at a given redshift. The merger timescale can be derived either empirically [@Conselice:2009bi] or through simulations [@Kitzbichler:2008gi; @Lotz:2010ie; @Lotz:2010hf; @2017MNRAS.468..207S], with different morphology or pair criteria having different timescales within the merger process. Simulations using N-body models of this merger process have measured the time-scales for mergers of galaxies with different masses, mass ratios, and other merger properties @Lotz:2010ie. Typically these have been found by e.g., @Conselice:2009bi [@Lotz:2010ie] to be around $\tau_{\rm{m}} = 0.3-0.7$ Gyr for pairs with projected separation of $\leq 20$ and $\leq 30$ kpc, respectively. These values are based on the average timescales for those separations and similar (baryonic) mass ratios of 1:3. The additional factor, $C_{\rm merg}$, is necessary because two galaxies that appear as a pair only have some probability to merge over a given time-scale. The orbital parameters of some galaxy pairs can result in a very long dynamical friction time-scale, resulting in merger timescale longer than the Hubble time. From simulations, this value computed over all possible merging scenarios is typically $C_{\rm merg} = 0.6$ [@Conselice:2014ct] but this value will also depend on the specific mass and redshift dependent. In this work we estimate the merger rates using the redshift dependent merger observability timescale of @2017MNRAS.468..207S, such that $$\mathcal{R} (>{\text{M}_{\star}}, z) = \frac{f_{\text{P}}(>{\text{M}_{\star}}, z)}{\tau_{\text{P}}(z)}.$$ The redshift dependent merger observability timescale, $\tau_{\text{P}}(z)$, is calculated by modelling the timescale required to reconcile the intrinsic merger rates of galaxies in the Illustris simulation [@Vogelsberger:2014gw; @Genel:2014dh] with the estimated pair counts of galaxies from the simulation. This evolving time-scale incorporates the effects accounted for by $C_{\rm merg}$ in Eq. \[eq:old\_rate\_conv\], and is defined as: $$\label{eq:merger_timescale} \tau_{\text{P}}(z) = 2.4 \times (1+z)^{-2} \textup{Gyr}.$$ We note that the pair criteria employed by @2017MNRAS.468..207S differ from those in this work, with a primary galaxy mass range of $10.5 < \log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) < 11$ and pair separation radii of 10 to 50 kpc. The overall normalization of the timescales therefore represents a significant systematic uncertainty, particularly in the case of the $9.7 < \log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) < 10.3$ sample. Despite these systematic uncertainties, our assumed observability timescales presented by @2017MNRAS.468..207S represent the best currently available and the most plausible avenue for inferring merger rates from observed pair counts. In addition to these systematic uncertainties, we also highlight that there is likely significant scatter in the merging timescales on a pair-by-pair basis [see Fig. 6 of @2017MNRAS.468..207S]. [ ![Estimated major merger rate per galaxy as a function of redshift for galaxies with stellar mass $9.7 < \log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) < 10.3$ (top) and $\log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) > 10.3$ (bottom) assuming the redshift dependent merger timescales of @Snyder:2016ta. Also shown are the merger rates based on the close-pair statistics of , @mundy2017 and @2018MNRAS.475.1549M, assuming the same redshift dependent timescale. The gold line and shaded region in each figure show the best-fitting power-law model from Figure \[merger-fig:merger\_frac\] converted to merger rates using our assumed merger timescales (Equation \[eq:merger\_timescale\]). The right-hand scale illustrates the inferred specific mass accretion rate through major mergers based on the observed merger rate (see text). For reference, we also show the observed specific star-formation rates for similar mass galaxies as a function of redshift [green shaded region; @2014ApJS..214...15S].[]{data-label="merger-fig:merger_rate_per_galaxy"}](plots/merger_rate_9_7.pdf "fig:"){width="\columnwidth"} ![Estimated major merger rate per galaxy as a function of redshift for galaxies with stellar mass $9.7 < \log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) < 10.3$ (top) and $\log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) > 10.3$ (bottom) assuming the redshift dependent merger timescales of @Snyder:2016ta. Also shown are the merger rates based on the close-pair statistics of , @mundy2017 and @2018MNRAS.475.1549M, assuming the same redshift dependent timescale. The gold line and shaded region in each figure show the best-fitting power-law model from Figure \[merger-fig:merger\_frac\] converted to merger rates using our assumed merger timescales (Equation \[eq:merger\_timescale\]). The right-hand scale illustrates the inferred specific mass accretion rate through major mergers based on the observed merger rate (see text). For reference, we also show the observed specific star-formation rates for similar mass galaxies as a function of redshift [green shaded region; @2014ApJS..214...15S].[]{data-label="merger-fig:merger_rate_per_galaxy"}](plots/merger_rate_10_3.pdf "fig:"){width="\columnwidth"}]{} With these caveats in mind, in Fig. \[merger-fig:merger\_rate\_per\_galaxy\] we present the merger rate per galaxy as a function of redshift implied by the observed pair counts in this work. We find an increase in the merger rate over all redshifts such that the highest merger rates are found for galaxies at the highest redshifts where we can probe. In Fig. \[merger-fig:merger\_rate\_per\_galaxy\] we also plot the best-fit power law and power-law + exponential parameterizations from Table \[tab:fitvalues\] convolved with the observability timescale. The pair count results of @mundy2017, @2018MNRAS.475.1549M and converted using the same merger timescale are also shown. In the higher mass bin we find that there is excellent agreement with the merger rates measured in the Illustris hydrodynamical simulation by @RodriguezGomez:2015hw. However, at $9.7 < \log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) < 10.3$, the pair counts measured for the CANDELS fields imply merger rates that are significantly larger than those presented in Illustris. @Snyder:2016ta. [cc]{}\ $0.5 \leq z < 1.0$ & $0.08 \pm 0.01$\ $1.0 \leq z < 1.5$ & $0.22 \pm 0.02$\ $1.5 \leq z < 2.0$ & $0.53 \pm 0.04$\ $2.0 \leq z < 2.5$ & $0.97 \pm 0.09$\ $2.5 \leq z < 3.0$ & $1.26 \pm 0.21$\ $3.0 \leq z < 3.5$ & $< 2.91$\ $3.5 \leq z < 4.5$ & $< 14.15$\ &\ \ $0.5 \leq z < 1.0$ & $0.07 \pm 0.01$\ $1.0 \leq z < 1.5$ & $0.12 \pm 0.02$\ $1.5 \leq z < 2.0$ & $0.29 \pm 0.03$\ $2.0 \leq z < 2.5$ & $0.39 \pm 0.06$\ $2.5 \leq z < 3.0$ & $0.58 \pm 0.10$\ $3.0 \leq z < 3.5$ & $1.07 \pm 0.25$\ $3.5 \leq z < 4.5$ & $1.21 \pm 0.37$\ $4.5 \leq z < 5.5$ & $3.29 \pm 1.07$\ $5.5 \leq z < 6.5$ & $7.59 \pm 2.69$\ \[tab:mergerpergal\] Although more informative than the merger fraction alone, the merger rate per galaxy is an average over all galaxies at a given mass and redshifts. We are also interested in knowing what the true merger rate is - that is how many merger are occurring per unit time per unit volume as a function of redshift. Similarly to previous studies, we define the comoving merger rate density, $\Gamma$, as: $$\label{eq:merger_rate} \Gamma(>{\text{M}_{\star}}, z) = f_{\text{p}}(>{\text{M}_{\star}}, z)n_{\text{c}}({\text{M}_{\star}}, z)\tau_{\text{P}}(z)^{-1},$$ where $f_{\text{p}}(>{\text{M}_{\star}}, z)$ is, as before, the mass and redshift-dependent galaxy pair fraction, $n_{\text{c}}(>{\text{M}_{\star}}, z)$ the comoving number density for galaxies with stellar mass $>{\text{M}_{\star}}$ and $\tau_{\text{P}}(z)$ the redshift dependent merger observability timescale. The comoving number densities for galaxies with $9.7 < \log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) < 10.3$ and $\log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) \geq 10.3$ are estimated from the same stellar mass function parameterizations used for the mass completeness weights: @Mortlock:2014et at $z \leq 3$, @Santini:2012jq at $3 < z < 3.5$, and @Duncan:2014gh at $z \geq 3.5$. Errors on the number densities are estimated by perturbing the Schechter function parameters based on their quoted errors and recalculating the integrated number density. This step is then repeated $10^{4}$ times and the lower and upper 1-$\sigma$ errors are taken as the and percentiles. In Fig. \[merger-fig:merger\_rate\] we show the resulting merger rates calculated following Equation \[eq:merger\_rate\]. We also compare in Fig. \[merger-fig:merger\_rate\] our results with those from Mundy et al. (2017) and @2018MNRAS.475.1549M. Here we see that the volume merger rates for the mass-selected samples are relatively constant with redshift, albeit with significant uncertainties in the highest redshift bins (the statistical uncertainties are dominated by the poor constraints on the high mass end of the stellar mass function). Given the additional statistical uncertainties in the cumulative number densities, the results for $\Gamma$ are in significantly less tension than for $\mathcal{R}$. [ ![Estimated comoving major merger rate as a function of redshift for galaxies with stellar mass $9.7 < \log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) < 10.3$ (top) and $\log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) > 10.3$ (bottom) assuming the redshift dependent merger timescales of @2017MNRAS.468..207S. Also shown are the merger rates based on the close-pair statistics of @mundy2017 and @2018MNRAS.475.1549M, assuming the same redshift dependent timescale. The right-hand scale illustrates the inferred mass accretion rate density from major mergers based on the observed merger rate (see text).[]{data-label="merger-fig:merger_rate"}](plots/merger_rate_dens_9_7.pdf "fig:"){width="\columnwidth"} ![Estimated comoving major merger rate as a function of redshift for galaxies with stellar mass $9.7 < \log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) < 10.3$ (top) and $\log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) > 10.3$ (bottom) assuming the redshift dependent merger timescales of @2017MNRAS.468..207S. Also shown are the merger rates based on the close-pair statistics of @mundy2017 and @2018MNRAS.475.1549M, assuming the same redshift dependent timescale. The right-hand scale illustrates the inferred mass accretion rate density from major mergers based on the observed merger rate (see text).[]{data-label="merger-fig:merger_rate"}](plots/merger_rate_dens_10_3.pdf "fig:"){width="\columnwidth"}]{} [cc]{}\ $0.5 \leq z < 1.0$ & $4.72^{+2.06}_{-1.69}$\ $1.0 \leq z < 1.5$ & $3.61^{+0.93}_{-0.84}$\ $1.5 \leq z < 2.0$ & $4.85^{+1.64}_{-1.31}$\ $2.0 \leq z < 2.5$ & $7.76^{+4.35}_{-3.17}$\ $2.5 \leq z < 3.0$ & $7.99^{+5.67}_{-4.12}$\ $3.0 \leq z < 3.5$ & $< 94.83$\ $3.5 \leq z < 4.5$ & $< 345.28$\ \ $0.5 \leq z < 1.0$ & $1.80^{+0.99}_{-0.79}$\ $1.0 \leq z < 1.5$ & $1.24^{+0.37}_{-0.34}$\ $1.5 \leq z < 2.0$ & $1.35^{+0.52}_{-0.44}$\ $2.0 \leq z < 2.5$ & $1.29^{+0.82}_{-0.62}$\ $2.5 \leq z < 3.0$ & $1.28^{+1.00}_{-0.70}$\ $3.0 \leq z < 3.5$ & $3.65^{+4.70}_{-2.69}$\ $3.5 \leq z < 4.5$ & $0.92^{+4.60}_{-1.08}$\ $4.5 \leq z < 5.5$ & $2.28^{+64.87}_{-2.99}$\ $5.5 \leq z < 6.5$ & $1.01^{+171.86}_{-1.37}$\ \[tab:merger\_rate\_dens\] ### Minor merger rates Despite the strong observational constraints on the evolution of the pair fraction as a function of merger ratio, it is not currently possible to derive strong conclusions on the actual minor merger *rates* and their corresponding mass growth. As illustrated in the previous section, assumptions on the merger timescales used to convert pair fractions to merger rates have significant effects on the estimated merger rates. Detailed simulations of the merger timescale as a function of mass, mass-ratio and redshift (whether physical or observability) are not currently available. Simulations of isolated mergers at low-redshift indicate that the timescales of minor mergers could be longer than those of major mergers [@Lotz:2010ie; @Lotz:2010hf], for mass ratios of 1:9 the average timescale increases only by $\approx 50\%$. The effect of longer timescales would be to decrease the predicted merger rate for minor mergers compared to that for major mergers, reducing their importance as a channel for galaxy growth. However, these simulations do not take into account the broader effects of projection effects and redshift uncertainties explored by @2017MNRAS.468..207S that may dominate the merger timescales for pair counts at high redshift. The observational results presented here illustrate that when such simulations are available, it will be possible to place detailed constraints on the complete merger histories of galaxies out to these high redshifts. Merger Mass Accretion Rates --------------------------- The rapid rise in merger rates per galaxy observed in Fig. \[merger-fig:merger\_rate\_per\_galaxy\] mirrors that observed in the specific star-formation rate evolution of galaxies over this period, [e.g. @Stark:2013ix; @Schenker:2013ep; @Duncan:2014gh; @2014ApJS..214...15S and references therein]. Growth through major mergers may therefore still represent a significant role in the formation of the earliest galaxies at $z > 3$, a fundamental prediction of hierarchical structure formation. Based on simple assumptions for the average mass accreted per major merger, in this section we present estimates of the stellar mass growth corresponding to the merger rates presented in Section \[merger-sec:mergerrate\]. Due to the large systematic uncertainty inherent in the conversion from pair fractions to merger rates, interpretation of results from more complex modelling approaches would still be dominated by the same systematic limitations. ### Specific mass accretion rates Analogous to the specific star-formation rate, the specific mass accretion rate can simply defined as $\dot{M}/M = \mathcal{R}(z)\bar{\mu}$, where $\bar{\mu}$ is the median mass ratio. By integrating the distribution of pair fraction as a function of merger ratio presented at $2.5 < z < 3$ in Fig. \[merger-fig:merger\_frac\_mu\], we calculate an average ‘major merger’ mass ratio of $\bar{\mu} = 0.53$. Based on the lack of observed evolution in pair fraction as a function of mass ratio (Fig. \[merger-fig:merger\_frac\_mu\]) we make the assumption of $\bar{\mu} = 0.53$ at all redshifts. In Fig. \[merger-fig:merger\_rate\_per\_galaxy\] we plot the resulting specific mass accretion rate for our sample in the right-hand twin axis. We find values for the specific mass accretion rate which vary between $0.07$ and $\sim7$ Gyr$^{-1}$ for the major mergers in our sample. Also plotted in Fig. \[merger-fig:merger\_rate\_per\_galaxy\] for reference is the median specific star-formation rate (sSFR, plus intrinsic scatter) for star-forming galaxies out to $z\sim6$, as described by the functional form presented in @2014ApJS..214...15S. During the period of peak galaxy formation ($1 < z < 3$), star-formation in massive galaxies is clearly the dominant form of mass growth. However, modulo the large systematic uncertainties in both estimates, the sSFR and specific merger mass accretion rate begin to converge at $z > 3$. This implies that at the highest redshifts, the amount of mass added to galaxies through major mergers may be directly comparable to that added by in-situ star-formation. However, it is also the case that some of the star formation we see is being produced in the merging events associated with these galaxies. We cannot separate at this point the merger contributed to the non-merger triggered star formation, but suffice it to say, a significant fraction of the mass in these galaxies is being added in some form by the merger process. ### Mass accretion rate density A second important observational property is the integrated mass accretion rate density from major mergers, $\rho_{\textup{M}}$. As above, we make a simple assumption that the average mass added per merger event is equal to $\bar{\mu}\times \bar{M_{\star}}$, where $\bar{\mu} = 0.53$ based on the $2.5 < z < 3$ bin and $\bar{M_{\star}}$ is calculated from the stellar mass function in this same bin. The resulting mass accretion rate density estimates are illustrated by the twin axis in Fig. \[merger-fig:merger\_rate\]. We note that while there is variation in $\bar{M_{\star}}$ between redshift bins, typically $\pm 0.05~\textup{dex}$, it is smaller than the large systematic uncertainties in the merger timescales used to derive $\Gamma$. We therefore present only this fiducial conversion in Fig. \[merger-fig:merger\_rate\] for ease of interpretation. Interpreting the estimated $\rho_{\textup{M}}$ presented in Fig. \[merger-fig:merger\_rate\], we find that the merger rate density is fairly constant, and this extends down to the lowest redshifts when we include results from @mundy2017. We find no clear peak in the integrated merger rate, at least for galaxies with masses M$_{*} > 9.7$. This is in stark contrast to the cosmic star-formation rate density which peaks at $z \sim 2$ [@Madau:2014gt]. The difference in merger rate and merger rate density redshift evolution can be reconciled by the fact that, while the number of mergers per galaxy is going down at lower redshifts, the number of galaxies above that mass limit is increasing and these two effects average each other out – at least for the mass range probed in this work. Discussion - The evolution of galaxy mergers at $0.5 < z < 6$ {#merger-sec:discussion} ============================================================= It is fairly well established that there appears to be a disagreement between the observed merger history and models, particularly at high redshifts [e.g. @Bertone:2009jc; @Jogee:2009iz]. Recent studies have attempted to alleviate this discrepancy with the idea that it results from observational studies selecting galaxies (and their merger ratios) by stellar mass, while model predictions have often been based on *baryonic* mass [@2016ApJ...830...89M]. The significant rise in the gas-fraction of galaxies at higher redshift would mean that pairs of merging galaxies with stellar mass ratio of $\mu << 1/4$ could have a baryonic mass ratio that would classify it as a major merger $\mu > 1/4$ - therefore increasing the observed number of major mergers. Two recent observational studies [@2016ApJ...830...89M; @2018MNRAS.475.1549M] have supported this picture, finding significantly greater numbers of major merger pairs at $z > 1$ based on flux ratios when compared with stellar mass ratios. However, simulations that explore mergers as a function of *stellar* mass [@RodriguezGomez:2015hw] can also significantly over-predict major merger rates at high redshift with respect to those presented in observational studies [@2016ApJ...830...89M; @mundy2017]. It is important to remember that the conversion of observed pair fractions to a merger rate requires the assumption of a corresponding merger timescale. This merger time-scale is critical but difficult to measure, and in the past has been taken to be a constant through cosmic time. @Snyder:2016ta revealed through forward modelling of galaxy pair-counts in simulations that the merger time-scale for galaxy pairs declines as $\sim (1+z)^{2}$. When we use these new evolving time-scales to estimate the merger rate from pair fractions, the ‘observed’ merger rate is found to increase with redshift at a rate that is more comparable to those predicted by hydrodynamical simulations that previous work would suggest [@RodriguezGomez:2015hw]. Because the merger time-scale is shorter at higher redshifts, this means that although we see a gentle rise in pair fraction with redshift, there are significantly more mergers actually occurring due to the fact that the time-scale for these mergers to occur is much faster at higher redshifts. This suggests that mergers are a more common process, by a factor of $>10$ at $z = 6$ compared with $z = 1$. The reason we do not see as many mergers ongoing is clearly because the time-scales for them to occur is much quicker than it is at lower redshifts. Based on the results presented in Section \[merger-sec:mergerrate\], we conclude that the assumed timescale is the origin of the discrepancy between the observations and simulation results for the merger history (and not necessarily the use of stellar mass selections). What we generally find is that, while we agree well with the predicted merger rates at higher masses, our observations now imply a higher merger rate than predicted for galaxies with $9.7 < \log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) < 10.3$. Some of this discrepancy may be accounted for by the expected mass-dependency of merger timescales. @Kitzbichler:2008gi find a merger timescale in $N$-body simulations that varies as $\propto M_{\star}^{-0.3}$, yielding expected timescales for $9.7 < \log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) < 10.3$ that are $\approx40\%$ longer than for $\log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) > 10.3$, and hence merger rates that are lower by the same amount. Further investigation is required to establish whether any remaining offset is physical or a result of additional mass-dependence in the merger observability timescales. For minor mergers ($\mu < \frac{1}{4}$) we find that difference between the observations and theory gets larger at the lowest mass ratio of mergers. At face value, our observations suggest that minor mergers may not be as common or as important in the galaxy formation process than what is predicted in the Illustris simulation [@RodriguezGomez:2015hw]. However, given the simplistic prescription used in this work convert from pair count to merger rates (and vice-versa), the source of this discrepancy may also lie in this critical assumption. Only a small mass-ratio dependence in the merger observability timescales would be required to alleviate the observed tension. Finally, given that the redshift evolution of pair fractions for massive galaxies ($\log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) \geq 10.3$) observed in this work is in good agreement with other recent studies [@2016ApJ...830...89M; @mundy2017], our key conclusion on the rapid rise in merger rates is not necessarily unique to our observed pair fractions. Furthermore, although the recent work by @2018MNRAS.475.1549M using the same CANDELS dataset find different evolution in the observed galaxy pair fraction, when incorporating the evolving merger timescales of @Snyder:2016ta the authors draw similar conclusions to those presented in this study. However, the higher observed pair fractions at $2 < z < 3$ in this work (see Fig.\[merger-fig:merger\_frac\]) and the extension to higher redshift mean that this is the first instance where the observed merger rates per galaxy are shown to rise at a rate that so closely matches those of simulations out to the very earliest epoch of galaxy formation. Despite these advances, there still remain key uncertainties in estimating galaxy merger rates that future studies can address. From the additional tests performed in this study (see Sections \[merger-sec:specz\_pairs\] and \[merger-sec:pz\_precision\]), we know there are still systematic uncertainties in the pair fractions obtained from photometric redshifts on the order of $\sim30\%$. However, in the final inferred merger rates, these uncertainties are dominated by the larger uncertainty in the merger timescales (or observability timescale). With larger simulation volumes and the improved number statistics these allow, extensions to the forward modelling of @Snyder:2016ta would enable estimates of merger observability timescales (and the scatter therein) as a function of mass, merger ratio and redshift to much greater precision. This increased understanding of merger timescales, more than any increase in redshift precision or reliability, is key to placing meaningful observational constraints on the assembly history of massive galaxies. Summary {#merger-sec:summary} ======= Using the full CANDELS data set we present a study of galaxy major mergers up to $z = 6$, and minor mergers up to $z = 3$, for massive galaxies with $\log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) \geq 10.3$ and $9.7 < \log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) < 10.3$. This is the first analysis at such early times in the universe, and uses the deepest data over a relatively large area where this type of analysis can be preformed. The results of this study have implications for a host of other areas of galaxy formation and astrophysics, including star-formation triggering, black hole growth and AGN activity, galaxy assembly as well as the number of super-massive black hole mergers in the early universe. This last issue is an important one for gravitational wave detections through future missions such as LISA (Conselice & Duncan 2019, in prep). As part of our analysis we have made new stellar mass and photometric redshift measurements for galaxies in all five CANDELS fields, including the full photometric redshift posteriors and stellar masses estimate at all likely redshift steps. The summary of our findings are: 1. For both $\log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) \geq 10.3$ and $9.7 < \log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) ,\leq 10.3$, the fraction of galaxies in major pairs (mass ratios of $0.25 \leq \mu \leq 1$) increases monotonically as a function of redshift out to $z\sim6$. 2. We furthermore find that the merger rate increases up to the highest redshifts explored ($z \sim 6$). This is due to the fact that we use new scaling laws from simulations which show that the merger observability time-scale declines at higher redshifts at $\sim (1+z)^{2}$ [@2017MNRAS.468..207S]. This differs significantly from previous work whereby the merger rate appeared to decline at higher redshifts. 3. Based on our observed merger rates, we infer that at $z > 3$, major mergers may play an increasingly important role in the mass growth of star-forming galaxies - significantly more so than at the peak of galaxy formation. 4. While the cumulative pair fraction increases for more minor mergers down to a mass ratio of 1:20 for galaxies at $1 < z < 3$, the relative number of minor mergers is lower than predicted by simulations. Between these redshifts we also do not find a significant change in the fraction of galaxies merging at any merger mass ratio we probe, suggesting that the merger history for both minor and major mergers mimic each other at these epochs. Overall, our conclusions are that observational constraints of mergers in massive galaxies are now consistent with hierarchical models of galaxy formation. At the highest redshifts, mass growth from major mergers may be comparable to or even higher than in-situ star-formation. To probe at even higher redshifts, or lower mass galaxies at $z < 6$, will require deeper surveys with the James Webb Space Telescope. Alternatively wide-area surveys at comparable depths to CANDELS will probe volumes sufficient to provide samples of even more massive galaxies that are large enough to perform similar analyses [e.g. the Euclid Deep fields; @2011arXiv1110.3193L]. In addition to providing vital new observational constraints on galaxy formation, our results can be used to predict the number of likely events gravitational wave detectors such as LISA will find due to merger of super-massive black holes that exist at the centres of these galaxies. Thus, overall our results lead to a suite of implications that we will explore in future papers. Acknowledgements {#acknowledgements .unnumbered} ================ The research leading to these results has received funding from the European Union Seventh Framework Programme FP7/2007-2013/ under grant agreement number 607254. This publication reflects only the author’s view and the European Union is not responsible for any use that may be made of the information contained therein. KJD acknowledges support from the ERC Advanced Investigator programme NewClusters 321271. We would also like to acknowledge funding from the Science and Technology Facilities Council (STFC) and the Leverhulme Trust. PGP-G wishes to acknowledge support from Spanish Government MINECO Grant AYA2015-63650-P. KBM acknowledges support from the HST archival research grant HST-AR-15040. This work is based on observations taken by the CANDELS Multi-Cycle Treasury Program with the NASA/ESA HST, which is operated by the Association of Universities for Research in Astronomy, Inc., under NASA contract NAS5-26555. The VUDS spectroscopic data is in this work is based on data obtained with the European Southern Observatory Very Large Telescope, Paranal, Chile, under Large Program 185.A-0791, and made available by the VUDS team at the CESAM data center, Laboratoire d’Astrophysique de Marseille, France. Consistent completeness simulations for all CANDELS fields {#app:completeness} ========================================================== Our methodology for determining the detection completeness follows a procedure similar to those presented in the CANDELS release papers [@Galametz:2013dd; @Guo:2013ig], adding a representative range of mock galaxies to the $H_{160}$ detection images and attempting to recover them using the same photometry procedures used to produce the science catalogs. For the morphological distribution of our input mock sources, we assume an empirical distribution that is dependent on apparent magnitude. Firstly, we divide the galaxies with parametric morphology measurement of @vanderWel:2012eu into bins of apparent $H_{160}$ magnitude. Next, for every mock galaxy with a given assigned magnitude, we assign a morphology (effective radius, Sersic index and ellipticity) by randomly sampling a morphology drawn from the corresponding magnitude bin of the real galaxy sample. To maximise the final number statistics of detected sources at faint magnitudes, we assume a power-law magnitude distribution which results in $\approx 10\times$ more input sources at the faint magnitude limit ($H_{160} = 30$) as at the bright limit ($H_{160} = 22$). One critical assumption to note is that we assume the morphological distribution of sources below the magnitude limit of the @vanderWel:2012eu sample morphologies is similar to those just brighter than the limit. While in true physical terms this assumption is not likely to be valid for the key properties such as size, for the image resolution of HST any further evolution in size would have minimal effect. Additionally, the observed distribution of morphologies in our faintest bin is very similar to the completeness corrected morphology distribution for Lyman break galaxies observed by @2004ApJ...600L.107F and therefore likely represent a valid assumption. ![Size and morphology distributions of the mock sources inserted into the CANDELS $H_{160}$ images for completeness simulations. The measured distributions are taken from the catalog presented in @vanderWel:2012eu. Top: Surface brightness distrubution parametrized by Sersic indices. Middle: Circularised effective radii (kpc). Bottom: Measured ellipticity.[]{data-label="merger-fig:MorphologyDists"}](plots/MorphologyDists.pdf){width="50.00000%"} Once the morphologies of the mock sources has been assigned, we then insert the mock galaxy images into the respective $H_{160}$ image for each field, 3000 sources at a time, and then process the images through the same <span style="font-variant:small-caps;">SExtractor</span> process as used to produce the original CANDELS photometry catalogs. This process was repeated 75 times for each field, yielding an average number of mock sources detections per field of 70,000 (typically 100 to 3000 *detected* galaxies in the magnitude bins corresponding to 50% completeness). {width="45.00000%"} {width="45.00000%"} {width="45.00000%"} {width="45.00000%"} {width="45.00000%"} We note that specific care was taken to ensure that the correct combination of image release, software version and extraction parameters were used for each CANDELS field. With no additional mock sources were added to the $H_{160}$ science image, we confirm that we obtain the exact number of galaxy detections as presented in the official CANDELS releases (we refer the reader to the respective release papers for precise numbers). Stellar mass consistency checks =============================== Although there is clearly visible scatter, the majority of mass estimates are in very good agreement with the team estimate with no significant bias and relatively small scatter. Furthermore, the scatter is most significant at masses of $\log_{10}({\text{M}_{\star}}/ \text{M}_{\odot}) < 9$, well below the range probed in this analysis. Note that the redshift assumed for the stellar mass fits differs between the results of this paper and the team redshift, much of the scatter is therefore a result of small differences in redshift and due to issues in the mass estimate. When assuming identical redshifts (i.e. the best available redshift from the CANDELS photo-z releases) the scatter and biases are reduced even further. ![Comparison of stellar mass estimates from this work (minimum-$\chi^{2}$) with those of the median CANDELS team mass estimates for each of the full CANDELS photometry catalog. The median offset in each field is indicated in each panel.[]{data-label="merger-fig:mass_candels"}](plots/mass_comparison_candels.pdf){width="55.00000%"} \[lastpage\] [^1]: The normalized median absolute deviation is defined as $\sigma_{\text{NMAD}} = 1.48 \times \rm{median} \left ( \frac{\left | \Delta z \right |}{1+z_{\rm{spec}}} \right )$, see @Dahlen:2013eu. [^2]: We note that while naming convention varies between studies [e.g. ‘companion fraction’; @2018MNRAS.475.1549M], all literature values plotted correspond to the same observational quantity: the number of galaxy pairs divided by the number of primary galaxies within the sample. [^3]: Conversely, two galaxies separated in redshift by more than 1-$\sigma$ will still have a non-zero possibility of being at the same redshift.

ArXiv

Supplementary Figures {#supplementary-figures .unnumbered} ===================== ![The self-energy $\Sigma(\omega)$ computed for the 3D TI slab as discussed in the Methods section of the main text. [**a,**]{} The real part and [**b,**]{} The imaginary part.](figs1.pdf){width="\textwidth"} Supplementary Note 1: Spin non-conservation and the role of spin relaxation in the spin Hall effect {#supplementary-note-1-spin-non-conservation-and-the-role-of-spin-relaxation-in-the-spin-hall-effect .unnumbered} =================================================================================================== As pointed out in the main text, spin is not a conserved quantity in this system. However, the non-conservation of spin does not completely break off the connection between spin density and spin current density. It only requires that a spin relaxation term be added to the continuity equation and taken with care. The bulk of our system is disorder-free where time reversal symmetry prohibits any spin accumulation at a steady state, thus spin accumulation and relaxation can only happen on the surface. An empirical equation for surface spin density $s$ and bulk spin current $j^s$ can be written down as $$\frac{{\mathrm{d}}s}{{\mathrm{d}}t}=j^s-R(s)$$ where $R(s)$ is the surface spin relaxation rate which can be roughly expressed as $s/\tau_s$. At a steady state, one has $$j^s=R(s)\approx\frac{s}{\tau_s}$$ from which the accumulated surface spin density $s$ is determined. If there were no such spin flip term, the system would never reach a steady state. The situation for the spin Hall effect is subtly different from the Hall effect. In the Hall effect, the accumulated surface charge causes a lateral voltage drop to resist further accumulation of charge, such that at a steady state the lateral current $j\equiv 0$. In the spin Hall effect, however, the spin accumulation does not lead to any counter force for the spin current, and there is a persistent bulk spin current as long as the longitudinal electric field exists. This term has to be canceled by a spin flip term to reach the steady state. The situation is also different from the 2D case. In the main text, we have argued that for the steady state of a 2D quantum spin Hall system there is actually no voltage drop along the conduction channel, but only across the contact. Therefore the lateral spin current $j^s\equiv 0$ similar to the Hall effect case. Thus it’s possible to define a conserved spin current for a 2D system, but not for a 3D system where the voltage drops across the system itself. Among the literature of spin Hall effect, some [@SHI] did attempt to define a conserved spin component. Yet from the above argument it seems that in the spin Hall effect, a properly defined, physically observable “spin” which manifests itself as a surface spin accumulation should be nonconserved. Supplementary Note 2: About the Dyakonov-Perel spin relaxation mechanism {#supplementary-note-2-about-the-dyakonov-perel-spin-relaxation-mechanism .unnumbered} ======================================================================== A very tricky question regarding the Dyakonov-Perel spin relaxation mechanism in our system is the lack of ${\mathbf{k}}$ space. Under high disorder $\epsilon_F\tau\lesssim 1$, the impurity potential cannot be regarded as a perturbation thus wave vector ${\mathbf{k}}$ becomes an ill-defined quantity. Talking about the spin precessional random walk in this situation seems an unjustified story. However, the lack of $k$-space is only true when we treat $H_0+U$ as a whole. The spin random walk picture of D-P spin relaxation mechanism is actually an interaction picture which splits the Hamiltonian into a free part $H_0$ and an interaction part $U$. $H_0$ provides the energy eigenstate bases while $U$ accounts for the time evolution of the wave function. The interaction Hamiltonian $U$ does not have to be much smaller than $H_0$. Wave vector ${\mathbf{k}}$ is perfectly defined for $H_0$, which justifies the language of spin random walk. To be more specific, we compare the situation of our system to a traditional 2DEG with Rashba spin splitting which is known to exhibit D-P mechanism. The Hamiltonian in this case is $$H=H_b({\mathbf{k}})+H_s+U$$ where $H_b({\mathbf{k}})$ is a spin-independent band energy, $H_s$ is the spin splitting energy which can be expressed as $\hbar v_F {\mathbf{\sigma}}\cdot{\mathbf{k}}$, $U$ is the scattering term. Here $v_F$ is just a parameter with no meaning of “Fermi velocity”. The criterion for D-P mechanism is $|H_b|\gg|U|\gg|H_s|$ or $H_b\gg \hbar/\tau \gg \hbar v_F k$. The latter part of this criterion simply means the spin splitting structure is completely blurred by scattering. While on a 3D TI surface the entire band Hamiltonian is just the spin splitting energy, it certainly means a complete destruction of the k-space. The only difference from the traditional case is the lack of $H_b$ term. However, we will argue in the following that $H_b$ is not essential to the D-P mechanism. We investigate the evolution of the system during a time $t$. We divide $t$ into a lot of infinitesimal intervals $$t=\Delta t_1+\Delta t_2+...+\Delta t_N$$ The time evolution operator correspondingly breaks into $${\mathrm{e}}^{-{\mathrm{i}}\frac{H}{\hbar}t}={\mathrm{e}}^{-{\mathrm{i}}\frac{H}{\hbar}\Delta t_1}{\mathrm{e}}^{-{\mathrm{i}}\frac{H}{\hbar}\Delta t_2}...{\mathrm{e}}^{-{\mathrm{i}}\frac{H}{\hbar}\Delta t_N}$$ Each interval can be separated in terms of the three terms of $H$ $${\mathrm{e}}^{-{\mathrm{i}}\frac{H}{\hbar}\Delta t}={\mathrm{e}}^{-{\mathrm{i}}\frac{H_b}{\hbar}\Delta t}{\mathrm{e}}^{-{\mathrm{i}}\frac{H_s}{\hbar}\Delta t}{\mathrm{e}}^{-{\mathrm{i}}\frac{U}{\hbar}\Delta t}$$ Now we consider an initial state $$|{\mathbf{k}}{\rangle}\otimes|{\mathbf{\sigma}}{\rangle}$$ with momentum ${\mathbf{k}}$ and spin in the ${\mathbf{\sigma}}$ direction ($|{\mathbf{\sigma}}{\rangle}$’s are actually the coherent states of spin which form an over complete set in the spin space). Acting ${\mathrm{e}}^{-{\mathrm{i}}\frac{U}{\hbar}\Delta t}$ on $|{\mathbf{k}}{\rangle}\otimes|{\mathbf{\sigma}}{\rangle}$ will scatter it to a different ${\mathbf{k}}'$ with the amplitude determined by $U_{{\mathbf{k}}'{\mathbf{k}}}$ $${\mathrm{e}}^{-{\mathrm{i}}\frac{U}{\hbar}\Delta t}|{\mathbf{k}}{\rangle}\otimes|{\mathbf{\sigma}}{\rangle}=\sum_{{\mathbf{k}}'} \left(\delta_{{\mathbf{k}}'{\mathbf{k}}}-{\mathrm{i}}\frac{U_{{\mathbf{k}}'{\mathbf{k}}}}{\hbar}\Delta t\right)|{\mathbf{k}}'{\rangle}\otimes|{\mathbf{\sigma}}{\rangle}$$ but leaving the spin vector ${\mathbf{\sigma}}$ unchanged. Acting ${\mathrm{e}}^{-{\mathrm{i}}\frac{H_s}{\hbar}\Delta t}$ on $|{\mathbf{k}}{\rangle}\otimes|{\mathbf{\sigma}}{\rangle}$ will precess the spin vector about the axis ${\mathbf{k}}$ by an angle $${\mathrm{e}}^{-{\mathrm{i}}\frac{H_s}{\hbar}\Delta t}|{\mathbf{k}}{\rangle}\otimes|{\mathbf{\sigma}}{\rangle}=|{\mathbf{k}}{\rangle}\otimes|{\mathbf{\sigma}}+2v_F{\mathbf{k}}\Delta t\times{\mathbf{\sigma}}{\rangle}$$ but leaving the momentum ${\mathbf{k}}$ unchanged. Acting ${\mathrm{e}}^{-{\mathrm{i}}\frac{H_b}{\hbar}\Delta t}$ on $|{\mathbf{k}}{\rangle}\otimes|{\mathbf{\sigma}}{\rangle}$ does not change anything but simply induces a phase factor $${\mathrm{e}}^{-{\mathrm{i}}\frac{H_b}{\hbar}\Delta t}|{\mathbf{k}}{\rangle}\otimes|{\mathbf{\sigma}}{\rangle}={\mathrm{e}}^{-{\mathrm{i}}\frac{H_b({\mathbf{k}})}{\hbar}\Delta t}|{\mathbf{k}}{\rangle}\otimes|{\mathbf{\sigma}}{\rangle}$$ Now we may assign each time interval $\Delta t_i$ an available ${\mathbf{k}}_i$ to form an integral path $$\label{eq:2} {\mathbf{k}}_1\times\Delta t_1 \rightarrow {\mathbf{k}}_2\times\Delta t_2 \rightarrow {\mathbf{k}}_3\times\Delta t_3 \rightarrow ... \rightarrow {\mathbf{k}}_N\times\Delta t_N$$ The final state is just a sum over all paths. Consider two extreme cases: (1) $|H_s|\gg|U|$. In this case the spin vector precesses by an appreciable angle far before the momentum has an appreciable probability to get scattered to a different value. The scattering is essentially an adiabatic rotation of the spin vector to the new energy eigenstate. Hence spin relaxation and momentum relaxation are bound together and we have $\tau_s=\tau$. (2) $|H_s|\ll|U|$. In this case the momentum gains an appreciable probability to be scattered to a different value far before the spin vector precesses by an appreciable angle. This will result in the precessional random walk picture of D-P mechanism. We thus expect $\tau_s \sim 1/\tau$. The $H_b$ term, however, will cause some restriction to the above picture through the additional phase factor ${\mathrm{e}}^{-{\mathrm{i}}\frac{H_b({\mathbf{k}})}{\hbar}\Delta t}$. Consider a virtual variation to the path (\[eq:2\]): we slightly change the lengths of $\Delta t_i$ and $\Delta t_{i+1}$ to $\Delta t_i+\delta$ and $\Delta t_{i+1}-\delta$. Since we have assumed $|H_b|\gg|H_s|$ and $|H_b|\gg|U|$, if $H_b({\mathbf{k}}_i)\neq H_b({\mathbf{k}}_{i+1})$, we can choose the value of $\delta$ such that the amplitude contributions by $H_s$ and $U$ remain almost unchanged but the phase factor by $H_b$ changes drastically. Consequently, summing over these paths will result in cancelation. The only exception is paths with $$H_b({\mathbf{k}}_1)=H_b({\mathbf{k}}_2)=...=H_b({\mathbf{k}}_N)$$ where the contribution of $H_b$ becomes a trivial global phase factor. Therefore, we see that the presence of the $H_b$ term simply restricts available paths to those on the constant energy contour of $H_b({\mathbf{k}})$. Now for the surface of a 3D TI without the $H_b$ term, we simply remove the restriction that ${\mathbf{k}}$ must stay on a constant energy contour. The precessional random walk picture still holds even though there is no semi-classical orbital motion. Supplementary Note 3: Difficulties in existing models for the minimum conductivity in 3D TIs {#supplementary-note-3-difficulties-in-existing-models-for-the-minimum-conductivity-in-3d-tis .unnumbered} ============================================================================================ In this section we address in details why the two currently existing models do not explain the minimum conductivity satisfactorily. In [@tran_5], the authors showed a resistance peak of $70\ \Omega$ while tuning the gate voltage applied to a $10\ \mathrm{nm}$-thick $\mathrm{Bi}_2\mathrm{Se}_3$ thin film. Considering the $1:8$ aspect ratio of the conduction channel, this resistance converts to a $560\ \Omega$ square resistivity ($\sim50e^2/h$). The authors attributed this conductance to electrons hopping in a bulk impurity band. Charged impurities in $\mathrm{Bi}_2\mathrm{Se}_3$ are believed to have a relatively large Bohr radius $a_B\approx 4\ \mathrm{nm}$, which is comparable with the average spacing between impurities at a typical impurity concentration ($\sim 10^{19}\ \mathrm{cm}^{-3}$). This may result in a considerable hopping amplitude between impurity orbitals and contribute to some conduction if this impurity band is partially occupied. Based on this model, the 2D carrier density contributed by impurities is $n_{2D}<10^{19}\ \mathrm{cm}^{-3}\times10\ \mathrm{nm}=10^{13}\ \mathrm{cm}^{-2}$. To account for the residue conduction, the mobility of such hopping is then greater than $1000\ \mathrm{cm}^2\mathrm{V}^{-1}\mathrm{s}^{-1}$, which is unreasonably high. Based on a similar consideration, Ref. [@tran_9] extracted a slightly lower impurity band mobility of $380\ \mathrm{cm}^2\mathrm{V}^{-1}\mathrm{s}^{-1}$ but still seems too high for a hopping mechanism, which should typically be below $1\ \mathrm{cm}^2\mathrm{V}^{-1}\mathrm{s}^{-1}$ [@hopping_mobility]. Moreover, hopping electrons should also contribute to Hall coefficient depending on the occupancy of the impurity band. If impurity conduction dominates in the region near the charge neutral point, tuning the surface states shouldn’t cause a significant change in Hall coefficient. Although Ref.[@tran_5] did not report a Hall coefficient polarity switching, a similar experiment by [@tran_4] did report such switching in Ca-doped $\mathrm{Bi}_2\mathrm{Se}_3$ thin films. Ca-doping is expected to induce even higher impurity levels compared to exfoliated single crystals. Therefore impurity band conduction does not seem to be a good explanation of the residual conductivity. Ref. [@tran_6] adopts another explanation which attributes the conductance residue to electron/hole puddles formed when the Fermi level is close to the Dirac point. This model inherits from a similar study in graphene which concludes that the main source of scattering in graphene is unscreened long range Coulomb scattering [@trans_gra_thy]. This long range interaction results in a surface potential fluctuation in a relatively large length scale, where electrons can be semi-classically thought to form “puddles”. However, it has been demonstrated that in the most common 3D TIs such as $\mathrm{Bi}_2\mathrm{Se}_3$, the dominant impurity source is Se vacancies, which is short range and cannot be thought in terms of a semi-classical potential fluctuation. On the other hand, if the minimum conductivity $\sim 5e^2/h$ observed in [@tran_6] indeed comes from long range potential fluctuation, a brief estimation reveals that the residue carrier density is $n^\ast\approx10^{12}\ \mathrm{cm}^{-2}$, which corresponds to a potential fluctuation of $120\ \mathrm{meV}$. Those puddle-like residue carriers actually form a lot of mini-pn-junctions and should not be as mobile as uniform carriers. Therefore, the actually required fluctuation is even larger to account for the large $\sigma$. The potential fluctuation on the surface of $\mathrm{Bi}_2\mathrm{Se}_3$ can be directly measured through scanning tunneling spectroscopy, which has already been carried out by several groups. Ref. [@pfluctuation_2] did report a typical fluctuation of about $120\ \mathrm{mV}$, but suggested this fluctuation is structural rather than disorder-induced. Moreover, the morphology does not really look like “puddles” but rather some “spikes”. On the other hand, Ref. [@pfluctuation_1] reported a much smaller value around $10\ \mathrm{mV}$, suggesting such potential fluctuation is quite sample-dependent and cannot be universally adopted to explain the residue conductivity. [10]{} Murakami, S., Nagaosa, N. & Zhang., S. C. Spin-Hall insulator. , [**93,**]{} 156804 (2004). Sacepe, B. [*et al.*]{} Gate-tuned normal and superconducting transport at the surface of a topological insulator. , [**2,**]{} 575 (2011). He, L. [*et al.*]{} Surface-dominated conduction in a 6 nm thick $\mathrm{Bi}_2\mathrm{Se}_3$ thin film. , [**12,**]{} 1486–1490 (2012). Lee, J.-S., Kovalenko, M. V., Huang, J., Chung, D. S., & Talapin, D. V. Band-like transport, high electron mobility and high photoconductivity in all-inorganic nanocrystal arrays. , [**6,**]{} 348–352 (2011). Checkelsky, J. G., Hor, Y. S., Cava, R. J., & Ong., N. P. Bulk band gap and surface state conduction observed in voltage-tuned crystals of the topological insulator $\mathrm{Bi}_2\mathrm{Se}_3$. , [**106,**]{} 196801 (2011). Kim, D. [*et al.*]{} Surface conduction of topological dirac electrons in bulk insulating $\mathrm{Bi}_2\mathrm{Se}_3$. , [**8,**]{} 459–463 (2012). Adam, S., Hwang, E. H., Galitski, V. M., & Das Sarma., S. A self-consistent theory for graphene transport. , [**104,**]{} 18392–18397 (2007). Liu, Y. [*et al.*]{} Charging dirac states at antiphase domain boundaries in the three-dimensional topological insulator $\mathrm{Bi}_2\mathrm{Se}_3$. , [**110,**]{} 186804 (2013). Beidenkopf, H. [*et al.*]{} Spatial fluctuations of helical dirac fermions on the surface of topological insulators. , [**7,**]{} 939–943 (2011).

ArXiv

--- abstract: 'We present the analysis of the radial distributions and kinematic properties of the multiple stellar populations (mPOPs) hosted in the globular cluster (GC) NGC 6352 as part of the *Hubble Space Telescope* ([*HST*]{}) “UV Legacy Survey of Galactic Globular Clusters” program. NGC 6352 is one of the few GCs for which the mPOP tagging in appropriate color-magnitude diagrams is clear in all evolutionary sequences. We computed high-precision stellar proper motions for the stars from the cluster’s core out to 75 arcsec ($\sim$1.5 core radii, or $\sim$0.6 half-light radii). We find that, in the region explored, first- and second-generation stars share the same radial distribution and kinematic properties. Velocity dispersions, anisotropy radial profiles, differential rotation, and level of energy equipartition, all suggest that NGC 6352 is probably in an advanced evolutionary stage, and any possible difference in the structural and kinematic properties of its mPOPs have been erased by dynamical processes in the core of the cluster. We also provide an estimate of the mass of blue stragglers and of main-sequence binaries through kinematics alone. In general, in order to build a complete dynamical picture of this and other GCs, it will be essential to extend the analyses presented in this paper to the GCs’ outer regions where some memories of the initial differences in the mPOP properties, and those imprinted by dynamical processes, might still be present.' author: - Mattia Libralato - Andrea Bellini - Giampaolo Piotto - Domenico Nardiello - 'Roeland P. van der Marel' - Jay Anderson - 'Luigi R. Bedin' - Enrico Vesperini title: 'The *Hubble Space Telescope* UV Legacy Survey of Galactic Globular Clusters. XVIII. Proper-motion kinematics of multiple stellar populations in the core regions of NGC 6352.' --- Introduction ============ \[tab:log\] GO PI Instrument/Camera Filter $N$ $\times$ Exp. Time Epoch ------- ------------ ------------------- -------- ------------------------------------- --------------- 10775 Sarajedini ACS/WFC F606W $4 \times 140$ s , $1 \times 7$ s 2006 April F814W $4 \times 150$ s , $1 \times 7$ s 12746 Kong ACS/WFC F625W $2 \times 150$ s 2012 February F658N $1 \times 643$ s , $1 \times 650$ s WFC3/UVIS F336W $5 \times 400$ s , $1 \times 410$ s 13297 Piotto WFC3/UVIS F275W $2 \times 706$ s 2013 August $2 \times 800$ s 2014 May F336W $2 \times 311$ s 2013 August $2 \times 311$ s 2014 May F438W $1 \times 58$ s 2013 August $1 \times 72$ s 2014 May The *Hubble Space Telescope* ([*HST*]{}) “UV Legacy Survey of Galactic Globular Clusters” program [GO-13297, PI: Piotto; see @2015AJ....149...91P] has started a systematic, photometric analysis of the multiple stellar populations (mPOPs) hosted in globular clusters (GCs). We now know that essentially all GCs studied with the proper tools show the presence of mPOPs, and that the mPOPs found in different GCs are characterized by a wide variety of properties. After the initial studies focused on the identification of mPOPs, we worked on the full characterization of the mPOP properties. We have analyzed most of the photometric pieces of information contained in this Treasury survey, but the wealth of information contained in this data set is still far from being completely explored and revealed. We have begun to study the kinematic properties of the mPOPs thanks to state-of-the-art proper motions (PMs) in light of the recent work on this research field [@2010ApJ...710.1032A; @2013ApJ...771L..15R; @2015ApJ...810L..13B; @2018ApJ...853...86B; @2018ApJ...861...99L; @2018MNRAS.479.5005M]. In this work, we continue the investigation of the GC NGC 6352 started by @2015MNRAS.451..312N. NGC 6352 is a metal-rich ($\rm [Fe/H] = -0.67$, ) GC in the Bulge direction with a mass of $6.1 \times 10^4$ $M_\odot$ [@2019MNRAS.482.5138B]. @2015MNRAS.451..312N demonstrated the presence of two populations clearly distinguishable in the red-giant branch (RGB), sub-giant branch (SGB), main sequence (MS), asymptotic giant branch (AGB), and horizontal branch (HB) with UV-optical color-magnitude diagrams (CMDs). The first-generation (1G) population (hereafter, POPa) has a primordial chemical composition, while the second-generation (2G) population (POPb) is almost coeval ($\Delta$Age = $10\pm120$ Myr), slightly enhanced in He ($\Delta Y = 0.029\pm 0.006$), and in Na . In this paper, we focus our attention on the structural and kinematic properties of the mPOPs in this cluster. We first compute [*HST*]{}-based, high-precision PMs to select members of NGC 6352 among the multitude of (Bulge and Disk) field stars. We then separate POPa and POPb, and study their radial distributions and internal kinematics. Data sets and reduction {#obs} ======================= We made use of `_flc` exposures[^1] taken with the Wide-Field Channel (WFC) of the Advanced Camera for Survey (ACS) and with the Ultraviolet-VISible (UVIS) channel of the Wide-Field Camera 3 (WFC3). The complete list of observations[^2] is shown in Table \[tab:log\]. {width="1.\textwidth"} The data reduction is a combination of a first- and a second-pass photometric stages, and was performed as in @2018ApJ...853...86B and @2018ApJ...861...99L. We refer to these papers for an extensive description of the work flow. The first-pass photometry allows us to detect the brightest, most isolated sources in each exposure in a single finding wave, and measure position and flux of these objects via point-spread-function (PSF) fit. The publicly-available, spatially-variable [*HST*]{}library PSFs[^3] are tailored to each exposure, and stellar positions are corrected for geometric distortion using the state-of-the-art corrections available for [*HST*]{}detectors [@2006acs..rept....1A; @2009PASP..121.1419B; @2011PASP..123..622B]. Positions and fluxes are then used to build a common reference-frame system. The second-pass photometry employs all images at once to enhance the contribution of faint sources. All close-by neighbors are subtracted from each image prior to estimate position and flux of a given source, thus improving the measurements in crowded regions. The main differences with @2018ApJ...861...99L are that (i) we used the Gaia Data Release 2 to set up orientation (X and Y axes toward West and North, respectively) and pixel scale (40 mas yr$^{-1}$) of our reference frame system, and (ii) we run the second-pass photometry tool separately for the three data sets (GO-10775, GO-12746, and GO-13297) as done in @2018ApJ...853...86B, so as to measure stars that might have moved by more than 1 pixel from one epoch to another. We calibrated our photometry on to the Vega-mag system following the prescriptions given in, e.g., @2017ApJ...842....6B and @2018MNRAS.481.3382N. We measured bright, isolated stars on the `_drc` exposures using aperture photometry with aperture of 5 pixels, and corrected for the finite aperture. We then computed the 2.5$\sigma$-clipped median value of the magnitude difference between our photometry and the aperture-corrected `_drc`-based magnitudes. Finally, we calibrated our photometry by adding to our instrumental magnitudes this median difference and the Vega-mag zero-point given in the STScI website[^4]. PMs were computed by combining the multi-epoch [*HST*]{}data as in @2014ApJ...797..115B, to which we refer for the detailed description of the method. We also corrected for spatially-variable, high- and low-frequency systematic effects following the descriptions of @2014ApJ...797..115B [@2018ApJ...853...86B] and @2018ApJ...861...99L. An overview of the final PM catalog of NGC 6352 is presented in Fig. \[fig:overview1\]. The median 1D PM error of stars at the level of the MS turn-off ($18 < {m_{\rm F606W}}< 19$) is of $\sim$32 $\mu$as yr$^{-1}$. For comparison, stars in this field at the same magnitude level have a median 1D PM error of $\sim$360 $\mu$as yr$^{-1}$ in the Gaia DR2. Figure \[fig:overview2\] shows the PM along $\alpha \cos\delta$ and $\delta$ directions as a function of the [$m_{\rm F606W}$]{}magnitude, $({m_{\rm F606W}}-{m_{\rm F814W}})$ color, X and Y master-frame positions. No clear systematic trends at levels comparable to the random errors for the best-measured stars arise from these plots. {width="1.\textwidth"} Only objects measured in all the available filters of GO-10775 and GO-13297 were considered in the analysis, thus the field of view (FoV) at disposal is that covered by the GO-13297 WFC3/UVIS data. Stars are considered as well measured if: (i) their quality of PSF (`QFIT`) parameter is larger than the 95-th percentile at any given magnitude[^5], (ii) their magnitude rms is lower than the 95-th percentile at any given magnitude[^6], (iii) they are measured in at least 40% of the images, (iv) their fraction of neighbor flux within the fitting radius before neighbor subtraction is less than 1, (v) their shape parameter `RADXS` [excess/deficiency of flux outside of the fitting radius with respect to the PSF prediction, see @2008ApJ...678.1279B] is lower than the 85-th percentile at any given magnitude, (vi) their flux is at least 3$\sigma$ above the local sky, (vii) their reduced $\chi^2$ of the PM fit is lower than 2, and (viii) their rejection rate in the PM fit is lower than 30%. Finally, we also excluded stars with a PM error larger than half the local velocity dispersion $\sigma_\mu$ of the closest 100 cluster stars, and stars outside a radius of 75 arcsec from the cluster center (i.e., the four corners of the FoV) to avoid edge effects. These last two conditions are applied only in the kinematic analysis. Dissecting NGC 6352 {#tag} =================== We identified the two populations of NGC 6352 in all evolutionary sequences by combining UV and optical CMDs and color-color diagrams, similarly to what is described in @2017MNRAS.464.3636M and @2018ApJ...861...99L. We initially corrected for the differential reddening affecting our photometry following the prescription of, e.g., and @2017ApJ...842....7B. NGC 6352 is in the direction of the Bulge and the extinction is high [$E(B-V)=0.22$, @1996AJ....112.1487H 2010 edition]. We assumed that cluster members are affected by the same amount of intra-cluster differential reddening. For each star, we selected a sample of close-by cluster stars and measured their median shift along the reddening direction with respect to a fiducial line in the CMD. Since NGC 6352 hosts two distinct populations, we selected the most populous group (POPb) to define the fiducial line. We tailored the correction for each CMD we present in this paper. {width="1.\textwidth"} We first defined the RGB, SGB, and MS regions, and then separated POPa and POPb by means of the pseudo two-color diagrams “chromosome maps” [@2017MNRAS.464.3636M]. We drew by hand two fiducial lines along RGB, SGB and MS in the [$m_{\rm F814W}$]{}versus $({m_{\rm F275W}}-{m_{\rm F814W}})$ and [$m_{\rm F814W}$]{}versus C$_{\rm F275W,F336W,F438W}$$=$$({m_{\rm F275W}}-{m_{\rm F336W}})-({m_{\rm F336W}}-{m_{\rm F438W}})$ planes, and rectified these sequences by defining $\Delta$color$=$$(\rm color-fiducial_{\rm red})/(fiducial_{\rm blue}-fiducial_{\rm red})$, where “color” is either $({m_{\rm F275W}}-{m_{\rm F814W}})$ or C$_{\rm F275W,F336W,F438W}$. The chromosome maps of NGC 6352 are presented in Fig. \[fig:overviewphot\]. The primordial POPa and the 2G stars of POPb are shown in red and azure, respectively. The two populations are quite obvious to separate in the RGB, SGB, AGB, and HB. The identification of the two populations in the MS is based on the Hess diagram of the chromosome map (inset in the MS panel of Fig. \[fig:overviewphot\]). We also separated equal-mass binaries that seem to belong to POPa and POPb as follows. First, we defined equal-mass MS binaries objects $\sim$0.75 mag brighter than the MS fiducial in the [$m_{\rm F814W}$]{}versus $({m_{\rm F606W}}-{m_{\rm F814W}})$ CMD (left panel of Fig. \[fig:overviewbin\]). Then, in the [$m_{\rm F814W}$]{}versus C$_{\rm F275W,F336W,F438W}$ CMD, we tagged as POPa binaries the equal-mass binaries brighter than the reddest edge of the POPa MS. The remaining binaries belong to the POPb (right panel of Fig. \[fig:overviewbin\]). Note that this classification for the equal-mass binaries does not consider the case of equal-mass mixed 1G$+$2G binaries, which are expected theoretically [e.g., @2016MNRAS.457.4507H] and might have different properties from 1G$+$1G and 2G$+$2G systems. ![Overview of the mPOP tagging along the equal-mass MS binaries. In the [$m_{\rm F814W}$]{}versus $({m_{\rm F606W}}-{m_{\rm F814W}})$ CMD (left panel) we initially selected as equal-mass MS binaries objects 0.75 mag brighter than the MS fiducial line (yellow, solid line). The MS fiducial line shifted by 0.75 mag is shown as a yellow, dashed line in the CMD as reference. Black points are stars belonging to either POPa or POPb, while gray dots are all other objects. We then defined POPa/ POPb equal-mass MS binaries stars brighter/fainter than the reddest edge (green, dashed line) of the POPa MS in the [$m_{\rm F814W}$]{}versus C$_{\rm F275W,F336W,F438W}$ CMD (right panel).[]{data-label="fig:overviewbin"}](fig4.pdf){width="1.\columnwidth"} mPOPs spatial distribution {#distr} ========================== Our [*HST*]{}observations of NGC 6352 cover a FoV out to $\sim$2$r_{\rm c}$ [$r_{\rm c} = 49.8$ arcsec, @1996AJ....112.1487H 2010 edition] from the cluster center. We analyzed the radial distributions of the populations along the RGB, SGB, and MS as follows. We chose as workbench the [$m_{\rm F814W}$]{}versus C$_{\rm F275W,F336W,F438W}$ CMD. We rectified the sequences by means of fiducial lines as described in Sect. \[tag\]. The fiducial lines were made by linearly interpolating the median color and magnitude in different (0.5, 0.1, and 0.5 mag for RGB, SGB, and MS, respectively) magnitude bins. We computed the histogram of the $\Delta$C$_{\rm F275W,F336W,F438W}$ color adopting a bin width of 0.05 mag. This methodology is however sensitive to the bin width and the starting point of the histogram. To ensure a bias-free estimate of the fraction of POPa and POPb stars, we computed the histograms 10000 times, each time adding a random noise to the points, and averaged them. The noise added to each star was randomly picked from a Gaussian distribution with $\sigma$ equal to the C$_{\rm F275W,F336W,F438W}$ photometric error of the star. This way we removed the dependencies on the bin width (by blurring or sharpening the distributions) and on the starting point of the histogram (by shifting the stars in the rectified CMD). The histograms of RGB, SGB, and MS stars present two distinct peaks. We fitted the final average histograms with a pair of Gaussian functions and estimated the fraction of stars belonging to POPa and POPb in a statistical fashion as described in @2013ApJ...765...32B. Figures \[fig:rgb\_rd\], \[fig:sgb\_rd\] and \[fig:ms\_rd\] present the radial distributions of the mPOPs along the RGB, SGB, and MS, respectively. The left panels show the [$m_{\rm F814W}$]{}versus C$_{\rm F275W,F336W,F438W}$ CMD of the stars in the entire FoV. The rectified CMDs are displayed in the middle-left panels. The fiducial lines used to rectify the CMD are shown as azure and red lines. The average histogram and the dual Gaussian functions for all stars in the FoV are shown in the middle-right panels. The ratios POPa,b/(POPa+POPb) over the entire FoV are shown as red and azure solid horizontal lines, respectively, in the rightmost panels. The dashed horizontal lines indicate the $\pm 1\sigma$ thresholds. We repeated the same procedure shown in these three panels for stars in equally-populated radial bins (two bins for the RGB, two for the SGB and five for the MS). The population ratio computed in each radial bin is shown with a filled circle in the rightmost panels of Figs. \[fig:rgb\_rd\], \[fig:sgb\_rd\] and \[fig:ms\_rd\]. The number of bins in the RGB and SGB was chosen to have about 100 stars in each bin, while for the MS five bins are a good compromise between the need of a high statistics in each bin (about 360 stars per bin) and map possible radial features in the distributions. We also changed the number of radial bins for the MS and found consistent results within the error bars. The ratios POPa/(POPa+POPb) are listed in Table \[tab:raddist\]. We find that the population ratio does not vary with the distance from the cluster center, and POPb is slightly more numerous than POPa. Our analysis is focused on the innermost regions of the cluster, which are the first parts where any initial radial gradient in the population ratio is erased by the effects of dynamical evolution, as shown in @2013MNRAS.429.1913V. Our result is therefore consistent with the theoretical expectations. Considering that NGC 6352 has a relatively short half-mass relaxation time [t$_{\rm hm}\simeq 0.8$ Gyr, @1996AJ....112.1487H 2010 edition], it is quite possible that the two populations are completely mixed over the entire cluster extension, but data covering a larger radial interval are necessary to further explore this issue. Finally, it is worth noticing that @2017MNRAS.464.3636M estimated the global fraction of 1G RGB stars in NGC 6352. They found a ratio of POPa RGB stars equal to $0.474 \pm 0.035$, in agreement with our independent estimate for the RGBa stars ($0.452 \pm 0.035$). #### HB, AGB, and equal-mass MS binaries The number of stars in the AGB and HB is too small to perform an analysis of their radial distributions. However, the relative number of POPa stars is $\sim$38% and 40% of the total for AGB and HB, respectively, in agreement with the results for RGB, SGB, and MS. We also find the relative number of POPa stars for the equal-mass MS binaries is $\sim$33%. In Fig. \[fig:bin\_rd\], we present the cumulative radial distribution of equal-mass binaries of POPa and POPb. The binary stars of the two populations are mixed in the cluster innermost regions. Recently, @2018MNRAS.tmp.3147H have shown that binary-star spatial mixing can be delayed by a number of dynamical processes affecting binaries (ionization, recoil, ejection). A more extended radial coverage than the one available in our [*HST*]{}data is required to test whether binaries are not mixed yet or NGC 6352 is sufficiently dynamically old to have reached complete spatial mixing also for its binary stars. \[tab:raddist\] [cccc]{} $r_{\rm min}$ & $r_{\rm max}$ & $\overline{r}$ & POPa/(POPa+POPb)\ $[$arcsec$]$ & $[$arcsec$]$ & $[$arcsec$]$ &\ \ 0.0 & 93.80 & 47.14 & $0.452 \pm 0.035$\ 0.0 & 46.09 & 28.31 & $0.475 \pm 0.050$\ 46.09 & 93.80 & 65.97 & $0.485 \pm 0.050$\ \ 0.0 & 101.66 & 48.37 & $0.487 \pm 0.030$\ 0.0 & 47.87 & 28.95 & $0.492 \pm 0.044$\ 47.87 & 101.66 & 67.65 & $0.485 \pm 0.044$\ \ 0.0 & 103.29 & 50.77 & $0.445 \pm 0.011$\ 0.0 & 30.34 & 20.34 & $0.457 \pm 0.025$\ 30.34 & 44.11 & 37.49 & $0.428 \pm 0.024$\ 44.11 & 57.89 & 51.04 & $0.435 \pm 0.025$\ 57.89 & 70.55 & 63.98 & $0.419 \pm 0.024$\ 70.55 & 103.29 & 80.93 & $0.464 \pm 0.025$\ {width="75.00000%"} {width="75.00000%"} {width="75.00000%"} ![Cumulative distributions of equal-mass 1G (red) and 2G (azure) MS binaries.[]{data-label="fig:bin_rd"}](fig8.pdf){width="\columnwidth"} Internal kinematics of NGC 6352 {#kin} =============================== We analyzed the internal kinematics of the members of NGC 6352. In the following analysis, the velocity dispersions were obtained by correcting the observed scatter of the PMs for the uncertainties of the individual PMs as in @2010ApJ...710.1063V. Velocity dispersions in mas yr$^{-1}$ were converted in km s$^{-1}$ by assuming a cluster distance of 5.6 kpc [@1996AJ....112.1487H 2010 edition]. mPOP kinematics {#mpopkin} --------------- We analyzed the combined ($\sigma_\mu$), radial ($\sigma_{\rm Rad}$), and tangential ($\sigma_{\rm Tan}$) velocity dispersions as a function of distance from the cluster center for each mPOP separately and in equally-populated radial bins. Figure \[fig:mpopall\] presents the velocity-dispersion (left) and anisotropy radial profiles (right) for the mPOPs along the RGB (top), SGB (middle), and MS (bottom), respectively. The value of each point is computed using 32 (43) stars for RGBa (RGBb), 55 (65) stars for SGBa (SGBb), 145 (150) stars for MSa (MSb), respectively. As a reference, we also draw the average velocity dispersion and anisotropy of the entire sample (horizontal lines) and the corresponding $\pm$1$\sigma$ uncertainties (shaded regions). Figure \[fig:mpopall\] shows that POPa (1G) and POPb (2G) have the same kinematics and are kinematically isotropic within $\sim$2$\sigma$. Considering that the data probe the cluster’s innermost regions within $\sim$1.5$r_{\rm c}$, or $\sim$0.6$r_{\rm h}$ [$r_{\rm c} = 49.8$ and $r_{\rm h} = 123.6$ arcsec, @1996AJ....112.1487H 2010 edition], this result is expected. The innermost regions are the first to relax and no significant anisotropy is, in general, expected close to the cluster’s center [see, e.g., @2016MNRAS.455.3693T]. For example, @2013ApJ...771L..15R and @2015ApJ...810L..13B found the presence of some velocity anisotropy in the 2G stars at distances larger than 1.5-2 $r_{\rm h}$. A study of the kinematics in the cluster’s outermost regions would be necessary to further explore the possible presence of anisotropy in the velocity distribution due to internal dynamical processes and the cluster’s interaction with the external tidal field [see, e.g., @2016MNRAS.455.3693T]. {width="\textwidth"} #### HB and equal-mass MS binaries We analyzed the internal kinematics of HB stars as a whole. Figure \[fig:mpophb\] shows $\sigma_\mu$ and the anisotropy for the HB stars (red and azure points for POPa and POPb, respectively) as a function of radius. Again, the two populations hosted along the HB of NGC 6352 are isotropic and share the same kinematics within the errors. We also computed the velocity dispersions of the mPOPs along the MS-binary sequence. Due to the low-number statistics, we made only one radial bin containing all stars in the FoV. The results are presented in Fig. \[fig:mpopbin\]. The azure and red points show the velocity-dispersion of MS binaries. MS binaries of POPa and POPb share the same kinematics and are isotropic. We were not able to analyze AGB stars because of the very few stars at disposal. ![Combined velocity dispersion (left panel) and anisotropy (right panel) as a function of radius for the two populations in the HB of NGC 6352. Each bin is made up by 10 HB stars.[]{data-label="fig:mpophb"}](fig10.pdf){width="\columnwidth"} ![$\sigma_\mu$ as a function of distance from the cluster center for equal-mass MS binaries is shown in the left panel. Each bin represents the average value of all binaries (11 and 16 for POPa and POPb, respectively). The right panel shows the tangential-to-radial anisotropy profile for the same stars.[]{data-label="fig:mpopbin"}](fig11.pdf){width="\columnwidth"} mPOP differential rotation -------------------------- We investigated the possible presence of differential rotation. We cannot directly measure any signature of cluster’s rotation in the plane of the sky because the systemic rotation signal is absorbed by the six-parameter linear transformations. However, there are different methods to infer the presence of rotation that rely on the skewness of the PMs in the tangential direction [see @2017ApJ...850..186H; @2018ApJ...853...86B; @2018ApJ...861...99L]. Figure \[fig:mpoprot\] shows the PMs of NGC 6352 in the $\mu_{\rm Tan}$ versus $\mu_{\rm Rad}$ plane. We measured the amount of skew in the $\mu_{\rm Tan}$ with (1) the value $G_1$ and the significance test $Z_{G1}$ [@Cramer1997], and (2) the third-order Gauss-Hermite moment $h_3$ [@1993ApJ...407..525V]. We find $$\textrm{POPa:}\left\{ \begin{array}{l} G_1 = 0.09 \textrm{ , } Z_{G_1} = 1.05 \\ h_3 = 0.021 \pm 0.026 \\ \end{array} \right. \, ,$$ $$\textrm{POPb:}\left\{ \begin{array}{l} G_1 = 0.09 \textrm{ , } Z_{G_1} = 1.27 \\ h_3 = 0.021 \pm 0.024 \\ \end{array} \right. \, . \\$$ These values are consistent with an absence of differential rotation in NGC 6352. Recently, @2018MNRAS.481.2125B estimated the amount of rotation in GCs by means of the Gaia DR2 PMs. The authors measured for NGC 6352 $\mu_{\rm Tan}(r<3r_{\rm h}) = 0.001^{+0.014}_{-0.013}$ mas yr$^{-1}$ and a 1$\sigma$ upper limit of $V/\sigma = 0.09$ ($V$ is the value of the rotation peak, $\sigma$ is the central velocity dispersion): values consistent with no rotation in the plane of the sky. Furthermore, they found a correlation between $V/\sigma$ and the relaxation time. As GCs advance in their evolution, they gradually lose their initial rotation [see, e.g., @2007MNRAS.377..465E; @2017MNRAS.469..683T and references therein]. The absence of rotation in NGC 6352 could be another indication of its advanced dynamical age. ![POPa (left panel) and POPb (right panel) PMs in the $\mu_{\rm Tan}$ versus $\mu_{\rm Rad}$ plane.[]{data-label="fig:mpoprot"}](fig12.pdf){width="\columnwidth"} Level of Energy Equipartition {#equip} ----------------------------- Globular clusters evolve toward a state of increasing energy equipartition ($\sigma_\mu \propto m^{-\eta}$ with $m$ the stellar mass and $\eta=0.5$ for a complete energy equipartition), but they are not expected to reach a complete equipartition [@2013MNRAS.435.3272T; @2016MNRAS.458.3644B; @2017MNRAS.464.1977W]. We divided the MS in 7 equally-populated magnitude bins, and computed $\sigma_\mu$ and median magnitude of the $\sim$230 stars in each bin. We then transformed these magnitudes in stellar masses by using a set of Dartmouth isochrones [@2008ApJS..178...89D]. NGC 6352 has an age of $\sim$13.0 Gyr [e.g., @2017MNRAS.468.1038W], $\rm [Fe/H] = -0.67$ , $\rm [\alpha/Fe] = 0.4$, $E(B-V) = 0.22$, and it is at a distance of 5.6 kpc [@1996AJ....112.1487H 2010 edition]. @2015MNRAS.451..312N analyzed the He content and relative age of the two populations hosted in NGC 6352, finding $\Delta Y = 0.029 \pm 0.006$ and $\rm \Delta Age = 10 \pm 120$ Myr. The best-fit isochrones are derived using a technique similar to that described in @2015MNRAS.451..312N to estimate the He difference between POPa and POPb in NGC 6352. We compared the colors of the observed fiducial lines of the two populations in the [$m_{\rm F606W}$]{}versus $({m_{\rm F606W}}-{m_{\rm F814W}})$ CMD with the colors of the fiducial lines of synthetic CMDs built from a set of isochrones with $\rm [Fe/H] = -0.67$ and $\rm [\alpha/Fe] = 0.4$. We treated both mPOPs at the same time, using the same isochrones except for the He content ($Y=0.256$ for POPa and $Y=0.285$ for POPb). We let distance, reddening and age vary, and defined the best-fit isochrones as those that provided the lowest $\chi^2$. We find $$\left\{ \begin{array}{rl} E(B-V) & = 0.25 \\ \textrm{Distance} & = 5.3 \textrm{ kpc} \\ \textrm{Age} & = 13.0 \textrm{ Gyr} \\ \end{array} \right.$$ The best isochrones for POPa and POPb are shown in the left panel of Fig. \[fig:equip\] as red and azure lines, respectively. The MS of NGC 6352 is made up by 45% of POPa stars and 55% of POPb stars. Therefore, we weighted the median mass in each magnitude bin by these ratios. Finally, we fitted the $\sigma_\mu$ versus mass values in a log-log plane with a weighted least-square straight line and find $$\eta = 0.334 \pm 0.182 \phantom{,}$$ The result is presented in the right panel of Fig. \[fig:equip\]. The black points picture the velocity dispersions of the seven MS bins; the black line is the best fit to these points \[We postpone the explanation of the azure and green points to the next section.\]. We computed the value of $\eta$ by using isochrones with slightly different ages, $\rm [Fe/H]$, and by neglecting the presence of the mPOPs, and obtained consistent results within the errors. This value of $\eta$ provides further evidence that NGC 6352 is in an advanced stage of its dynamical evolution, and it is in general consistent with the theoretical predictions of @2013MNRAS.435.3272T and @2017MNRAS.464.1977W, although our uncertainty on the value of $\eta$, mainly due to the low statistics and the small mass range considered, is large. {width="\textwidth"} The mass of blue stragglers and MS binaries {#mass} ------------------------------------------- Blue stragglers (BSs) and MS-binary systems are more massive than a typical (single) star in a GC. If the GC has some degree of energy equipartition, we expect these massive objects to have a lower velocity dispersion than the other stars. @2016ApJ...827...12B computed the mass of the BSs in several GCs by assuming that the velocity dispersions and masses of these stars scale as $$\alpha = \frac{\sigma_{\rm BSS}}{\sigma_{\rm MSTO}} = \left(\frac{M_{\rm BSS}}{M_{\rm MSTO}}\right)^{-\eta} \phantom{,}.$$ We calculated the mass of the BSs in NGC 6352 in a similar fashion, and extended the same methodology to the equal-mass MS binaries. For the BSs (blue dots in the left panel of Fig. \[fig:bss\]), we computed the combined velocity dispersion of these stars in 3 equally-populated radial bins of 5 stars each. We chose as reference population all RGB, SGB, and MS stars brighter than 0.5 mag below the MS turn-off (black points in the CMD in Fig. \[fig:bss\]), and computed the velocity dispersions of these reference objects in 11 radial bins[^7]. The velocity-dispersion radial profiles of BSs and reference stars are shown in the right panel of Fig. \[fig:bss\] in blue and black, respectively. For each radial bin of the reference population, we drew 10000 realizations of the average velocity dispersion by adding a random Gaussian noise with $\sigma$ equal to the error in the velocity dispersion. We then interpolated the median values of these distributions by means of a 3rd-order polynomial function with a flat center. The best fit is shown in Fig. \[fig:bss\]. The gray region represents the 1$\sigma$ error of the fitted profiles. We performed a similar computation for the BS, but this time we interpolated the velocity-dispersion profile with the same 3rd-order polynomial function of the reference population rescaled by a factor $\alpha$. The blue line in Fig. \[fig:bss\] depicts the best fit for the BSs, the pale-blue region is the 1$\sigma$-error region. We find $\alpha = 0.77 \pm 0.12$. Given a MS turn-off mass of 0.83 $M_\odot$ and assuming that the SGB and RGB stars have the same dynamical mass of those at the MS turn-off, we derive a mass for the BSs of $1.82 \pm 0.37$ $M_\odot$. Our estimate is slightly larger, although in agreement at 1$\sigma$ level, than the typical mass of a BS [between 1.0 and 1.6 $M_\odot$; see, e.g., @2018ApJ...860...36F]. We also estimated the average mass of the equal-mass MS binary systems (green dots in Fig. \[fig:msbin\]) as done for the BSs. This time we chose as reference the MS stars 0.75 mag fainter than the equal-mass binary region (black points in Fig. \[fig:msbin\]). We find $\alpha = 0.80 \pm 0.09$. The average mass of the MS stars (computed as described in Sect. \[equip\]) in the magnitude interval considered in the analysis is $\sim$0.72 $M_\odot$, which gives us an average mass of the binary systems of $1.41 \pm 0.30$ $M_\odot$. This is consistent with a system made by two stars of 0.72 $M_\odot$ each. Returning back to Fig. \[fig:equip\], its right panel shows the average velocity dispersion of all BSs and MS binaries as a function of mass. The filled dots refer to the mass computed in this Section and, by construction, they are aligned with the best fit of the MS stars from which we computed the value of $\eta$. If we assume an average mass of the BSs and MS binaries of 1.3 $M_\odot$ [the mid-range value of typical BS masses from @2018ApJ...860...36F] and 1.44 $M_\odot$ (twice the average MS stars at the same color level of the binaries), respectively, we can have an independent check of the goodness of our estimate of $\eta$. The empty dots plotted by assuming these theoretical values of the mass for BSs and MS-binary systems seem to be in agreement with the prediction of the MS stars (black line), and confirm the level of energy equipartition we computed with this [*HST*]{}data set. {width="\textwidth"} Comparison with the literature ------------------------------ As an external check, we made a comparison with the work of @2019MNRAS.482.5138B. The authors fitted $N$-body models to ground-based radial velocities, [*HST*]{}-based mass functions, and the Gaia DR2 PMs, and derived structural parameters for 144 Galactic GCs[^8]. First, we compared our [*HST*]{}-based velocity-dispersion radial profiles with those obtained with the Gaia DR2. We limited our analysis to the RGB stars to be consistent with @2019MNRAS.482.5138B. We computed the average $\sigma_\mu$ of all stars within 10 arcsec from the cluster center, and in 4 equally-populated (35 stars per bin) radial bins in the remaining part of the FoV. The values of $\sigma_\mu$ computed in this work are shown in black in Fig. \[fig:rgbprof\]. The red points are the velocity dispersions computed with the Gaia DR2 PMs by @2019MNRAS.482.5138B, rescaled to the cluster distance adopted in our work. The [*HST*]{}- and Gaia-based velocity dispersions are in excellent agreement with each other. @2019MNRAS.482.5138B provided an estimate of the central velocity dispersion $\sigma_0 = 3.5$ km s$^{-1}$, or $\sigma_0 = 3.33$ km s$^{-1}$ upon rescaling for the distance adopted in our work. We fitted our velocity-dispersion radial profiles with a 3rd-order polynomial (forced to be flat at the center) and extrapolated a value of $\sigma_0 = 3.47$ km s$^{-1}$, consistent with the value of @2019MNRAS.482.5138B. Finally, @2019MNRAS.482.5138B also computed the value of $\eta$ for stars in the mass range 0.5-0.8 $M_\odot$. They find $\eta = 0.36$, which is consistent with the value we derived from our [*HST*]{}-based PMs in Sect. \[equip\]. {width="\textwidth"} ![Combined velocity dispersion as a function of distance from the cluster center for RGB stars. Black points represent the kinematics of the RGB stars measured in this work, red points are those obtained by @2019MNRAS.482.5138B with the Gaia DR2 PMs. The black line is the best fit to our data. The red, dashed line is set at the central value of the velocity dispersion computed by @2019MNRAS.482.5138B.[]{data-label="fig:rgbprof"}](fig16.pdf){width="\columnwidth"} Conclusions =========== In this paper we presented a comprehensive characterization of the structural and kinematic properties of the mPOPs in the innermost regions of the Galactic GC NGC 6352. We used data from the [*HST*]{}UV survey GO-13297 to identify 1G (POPa) and 2G (POPb) stars along the RGB, SGB, and MS, finding a similar fraction of POPa stars ($\sim$45%) in each evolutionary group. We also find no evidence of a significant variation in the fraction of POPa stars with distance from the cluster center. As shown in previous theoretical studies, the innermost regions are the first where any initial radial gradients is erased, and the lack of a radial variation in the fraction of stars belonging to the two populations is therefore consistent with the theoretical expectations. We studied the cluster’s internal kinematics by means of [*HST*]{}-based, high-precision PMs. There is no sign of internal rotation, anisotropy in the velocity distributions, and differences between the kinematic properties of POPa and POPb stars. We also measured the dependence of the velocity dispersion on the stellar mass, and provided a quantitative estimate of the level of energy equipartition reached in the cluster’s inner regions. NGC 6352 is probably in its advanced evolutionary stages, and any difference in the structural and kinematic properties of mPOPs might have been erased by dynamical processes over the entire cluster’s extension. It will be essential to extend the analysis presented in this paper to the cluster’ outer regions where some memories of the initial or dynamically-induced differences might still be retained. Acknowledgments {#acknowledgments .unnumbered} =============== M.L. and A.B. acknowledge support from STScI grant GO-13297. The authors thank the anonymous Referee for the thoughtful suggestions that improved the quality of the paper. Based on observations with the NASA/ESA *HST*, obtained at the Space Telescope Science Institute, which is operated by AURA, Inc., under NASA contract NAS 5-26555. This work has made use of data from the European Space Agency (ESA) mission [*Gaia*]{} (<https://www.cosmos.esa.int/gaia>), processed by the [*Gaia*]{} Data Processing and Analysis Consortium (DPAC, <https://www.cosmos.esa.int/web/gaia/dpac/consortium>). Funding for the DPAC has been provided by national institutions, in particular the institutions participating in the [*Gaia*]{} Multilateral Agreement. Anderson, J., & King, I. R. 2006, Instrument Science Report ACS 2006-01, 34 pages, Anderson, J., & Bedin, L. R. 2010, , 122, 1035 Anderson, J., & van der Marel, R. P. 2010, , 710, 1032 Baldwin, A. T., Watkins, L. L., van der Marel, R. P., et al. 2016, , 827, 12 Baumgardt, H., Hilker, M., Sollima, A., & Bellini, A. 2019, , 482, 5138 Bedin, L. R., King, I. R., Anderson, J., et al. 2008, , 678, 1279-1291 Bellini, A., & Bedin, L. R. 2009, , 121, 141 Bellini, A., Anderson, J., & Bedin, L. R. 2011, , 123, 622 Bellini, A., Piotto, G., Milone, A. P., et al. 2013, , 765, 32 Bellini, A., Anderson, J., van der Marel, R. P., et al. 2014, , 797, 115 Bellini, A., Vesperini, E., Piotto, G., et al. 2015, , 810, L13 Bellini, A., Anderson, J., Bedin, L. R., et al. 2017a, , 842, 6 Bellini, A., Anderson, J., van der Marel, R. P., et al. 2017b, , 842, 7 Bellini, A., Libralato, M., Bedin, L. R., et al. 2018, , 853, 86 Bianchini, P., van de Ven, G., Norris, M. A., Schinnerer, E., & Varri, A. L. 2016, , 458, 3644 Bianchini, P., van der Marel, R. P., del Pino, A., et al. 2018, , 481, 2125 Carretta, E., & Gratton, R. G. 1997, , 121, 95 Cramer, D., 1997, “Basic Statistics for Social Research”. Routledge. Dotter, A., Chaboyer, B., Jevremovi[ć]{}, D., et al. 2008, , 178, 89-101 Ernst, A., Glaschke, P., Fiestas, J., Just, A., & Spurzem, R. 2007, , 377, 465 Feltzing, S., Primas, F., & Johnson, R. A. 2009, , 493, 913 Ferraro, F. R., Lanzoni, B., Raso, S., et al. 2018, , 860, 36 Gaia Collaboration, Prusti, T., de Bruijne, J. H. J., et al. 2016a, , 595, A1 Gaia Collaboration, Brown, A. G. A., Vallenari, A., et al. 2018, , 616, A1 Goldsbury, R., Richer, H. B., Anderson, J., et al. 2010, , 140, 1830-1837 Harris, W. E. 1996, , 112, 1487 Heyl, J., Caiazzo, I., Richer, H., et al. 2017, , 850, 186 Hong, J., Vesperini, E., Sollima, A., et al. 2016, , 457, 4507 Hong, J., Patel, S., Vesperini, E., Webb, J. J., & Dalessandro, E. 2018, , arXiv:1812.01229 Libralato, M., Bellini, A., van der Marel, R. P., et al. 2018, , 861, 99 Milone, A. P., Piotto, G., Bedin, L. R., et al. 2012, , 540, A16 Milone, A. P., Piotto, G., Renzini, A., et al. 2017, , 464, 3636 Milone, A. P., Marino, A. F., Mastrobuono-Battisti, A., & Lagioia, E. P. 2018, , 479, 5005 Nardiello, D., Piotto, G., Milone, A. P., et al. 2015, , 451, 312 Nardiello, D., Libralato, M., Piotto, G., et al. 2018, , 481, 3382 Piotto, G., Milone, A. P., Bedin, L. R., et al. 2015, , 149, 91 Richer, H. B., Heyl, J., Anderson, J., et al. 2013, , 771, L15 Tiongco, M. A., Vesperini, E., & Varri, A. L. 2016, , 455, 3693 Tiongco, M. A., Vesperini, E., & Varri, A. L. 2017, , 469, 683 Trenti, M., & van der Marel, R. 2013, , 435, 3272 van der Marel, R. P., & Franx, M. 1993, , 407, 525 van der Marel, R. P., & Anderson, J. 2010, , 710, 1063 Vesperini, E., McMillan, S. L. W., D’Antona, F., & D’Ercole, A. 2013, , 429, 1913 Wagner-Kaiser, R., Sarajedini, A., von Hippel, T., et al. 2017, , 468, 1038 Webb, J. J., & Vesperini, E. 2017, , 464, 1977 [^1]: `_flt` images pipeline corrected for the charge-transfer-efficiency defects as described in @2010PASP..122.1035A. [^2]: DOI reference: [^3]: [http://www.stsci.edu/$\sim$jayander/STDPSFs/](http://www.stsci.edu/~jayander/STDPSFs/). [^4]: <http://www.stsci.edu/hst/acs/analysis/zeropoints> for ACS/WFC and <http://www.stsci.edu/hst/wfc3/phot_zp_lbn> for WFC3/UVIS. [^5]: The closer the `QFIT` is to 1, the better is the PSF fit. In addition to the 95-th percentile cut, we consider as well measured all objects with a `QFIT` value higher than 0.99. We also discard all sources with `QFIT` lower than 0.6. [^6]: Similarly as for the `QFIT`, we also keep all sources with a magnitude rms lower than 0.05 mag, and discard those with a rms higher than 0.4 mag. [^7]: One bin of 37 stars was defined by all stars within 10 arcsec from the center of the GC. Five equally-populated bins of 118 stars each were defined from 10 arcsec to the core radius, and five bins of 81 stars each from the core radius to the limit of the analyzed FoV. [^8]: <https://people.smp.uq.edu.au/HolgerBaumgardt/globular/>.

ArXiv

--- author: - | Elizabeth M. Sweeney^1,2^, Russell T. Shinohara^3^, Blake E. Dewey^2^, Matthew K. Schindler^2^,\ John Muschelli^1^, Daniel S. Reich^1,2,4,5^, Ciprian M. Crainiceanu^1^, Ani Eloyan^1^\ bibliography: - 'mybib.bib' title: 'Relating multi-sequence longitudinal intensity profiles and clinical covariates in new multiple sclerosis lesions' --- ^1^Department of Biostatistics, The Johns Hopkins University, Baltimore, MD 21205 ^2^Translational Neuroradiology Unit, Division of Neuroimmunology and Neurovirology, National Institute of Neurological Disease and Stroke, National Institute of Health, Bethesda, MD 20892 ^3^Department of Biostatistics and Epidemiology, Center for Clinical Epidemiology and Biostatistics, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA 19104 ^4^Department of Radiology, The Johns Hopkins University School of Medicine, Baltimore, MD 21287 ^5^Department of Neurology, The Johns Hopkins University School of Medicine, Baltimore, MD 21287 Corresponding Author: Elizabeth Sweeney (317)698-5700 [email protected] Abstract {#abstract .unnumbered} ======== Structural magnetic resonance imaging (MRI) can be used to detect lesions in the brains of multiple sclerosis (MS) patients. The formation of these lesions is a complex process involving inflammation, tissue damage, and tissue repair, all of which are visible on MRI and potentially modifiable by pharmacological therapy, including neuroprotective and reparative agents now in development. Here we characterize the lesion formation process on longitudinal, multi-sequence structural MRI from 34 MS patients, scanned on average once every 37 days (sd 52.3, range \[13, 889\]), with an average of 21 scans each (sd 8.0, range \[10, 37\]). We then relate the longitudinal changes we observe within lesions to disease subtype and therapeutic interventions. In this article, we first outline a pipeline to extract voxel level, multi-sequence longitudinal profiles from four MRI sequences within lesion tissue. We then propose two models to relate clinical covariates to the longitudinal profiles. The first model is a principal component analysis (PCA) regression model, which collapses the information from all four profiles into a single scalar value. We find that the score on the first PC identifies areas of slow, long-term intensity changes within the lesion at a voxel level, as validated by two experienced clinicians (a neuroradiologist and a neurologist). On a quality scale of 1 to 4, with 4 being the highest, the neuroradiologist gave the score on the first PC a median rating of 4 (95% CI: \[4,4\]), and the neurologist gave it a median rating of 3 (95% CI: \[3,3\]). In the PCA regression model, we find that treatment with disease modifying therapies (p-value $<$ 0.01), steroids (p-value $<$ 0.01), and being closer to the boundary of abnormal signal intensity (p-value $<$ 0.01) are associated with a return of a voxel to intensity values closer to that of normal-appearing tissue. The second model is a function-on-scalar regression, which allows for assessment of the individual time points at which the covariates are associated with the profiles. In the function-on-scalar regression both age and distance to the boundary were found to have a statistically significant association with the profiles at some time point. The methodology presented in this article shows promise for both understanding the mechanisms of tissue damage in the disease MS and may prove to be useful for evaluating the impact of treatments for the disease in clinical trials. Keywords {#keywords .unnumbered} ======== Structural magnetic resonance imaging; Multi-sequence imaging; Longitudinal study; Multiple sclerosis; Longitudinal lesion behavior; Principal component analysis regression; Function-on-scalar regression; Expert rater trial Abbreviations {#abbreviations .unnumbered} ============= - Confidence Interval (CI) - Fluid attenuated inversion recovery (FLAIR) - National Institute of Neurological Disease and Stroke (NINDS) - Normal-appearing white matter (NAWM) - Magnetic resonance imaging (MRI) - Multiple sclerosis (MS) - Principal component (PC) - Principal component analysis (PCA) - Proton density-weighted (PD) - Relapsing remitting MS (RRMS) - Secondary progressive MS (SPMS) - Standard deviation (sd) - T1-weighted (T1) - T2 -weighted (T2) - Tesla (T) Introduction ============ Structural magnetic resonance imaging (MRI) can be used to detect lesions in the brains of multiple sclerosis (MS) patients. The formation of these lesions is a complex process involving inflammation, tissue damage, and repair, all of which MRI has been shown to be sensitive to. The McDonald criteria for diagnosis of MS have emphasized the key role of dissemination of lesions in the central nervous system on MRI not only in space, but also in time [@polman2011diagnostic]. Characterizing the longitudinal behavior of lesions on structural MRI is therefore likely to be important for monitoring disease progression and response to therapy, and for understanding the etiology of the disease. The clinico-radiological paradox refers to the poor association between clinical findings and radiological extent of involvement on MRI using traditional volumetric measures [@barkhof2002clinico]. Here we address this paradox by studying the association between the longitudinal behavior of lesions on structural MRI and clinical covariates. Previous work to characterize the longitudinal behavior of lesions on structural MRI has only involved a single structural MRI sequence, proton density-weighted (PD) imaging [@meier2003time; @meier2006mri; @meier2007time]. These approaches used a parametric model at the voxel level. Voxel-specific, biologically relevant features were then produced: hyperintensity above baseline, residual hyperintensity, and length of time at a given intensity level. It was found that the maximal insult within a lesion occurred at the center of the lesion, that lower initial intensity within a lesion was predictive of repair, and that most lesion activity did not last beyond 10 weeks. Some limitations of these studies were: (1) the MRI for these analysis consisted of frequent, bi-weekly scans; and (2) only one MRI sequence (PD) was used to characterize the behavior of the lesions, which ignores information known to be available in the other sequences [@mcfarland2002role]. Previous work designed to relate longitudinal changes in lesions on MRI and clinical covariates was recently published by [@ghassemi2014quantitative]. The change over a 2-year period in normalized T1-weighted (T1) intensity within new T2-weighted (T2) lesions was compared in pediatric and adult-onset MS patients to investigate differences in the behaviors of lesions between the two populations. A generalized linear mixed-effects model was used to relate clinical covariates, such as disease duration and treatments, to changes in intensity in the MRI. No statistically significant relationships were found, but a difference in the change in intensity between adults and pediatric MS patients was found, with pediatric patients showing more recovery than adults. Work has also been done to relate longitudinal changes in lesion intensity to sample size calculations for clinical trials. [@reich2015sample] used the change in the 25th percentile of intensity-normalized PD signal within a lesion over time to calculate sample sizes for clinical trials data for different lengths of trials. The 25th, 50th, and 75th percentiles of multiple MRI sequences were assessed, and it was found that the 25th percentile of the normalized PD yielded the smallest sample size requirements. In this study, we use MRI acquired from the National Institute of Neurological Disease and Stroke (NINDS), with subjects being scanned on average once every 37 days (sd 52.3, range \[13, 889\]) and an average of 21 scans per subject (sd 8.0, range \[10, 37\]). We first describe a pipeline for extracting the longitudinal profiles in lesion intensities from the T1, T2, T2-weighted fluid attenuated inversion recovery (FLAIR), and PD sequences. Next, we describe two models for relating the longitudinal information from these profiles to clinical covariates. The first model, a PCA regression model, collapses the information from the profiles into a single scalar. This scalar value is found to identify areas of slow and persistent, long-term intensity changes. This finding is validated by an expert rater trial with two raters, a neuroradiologist and a neurologist. The scalar value is then related to clinical information in a mixed-model regression. The second model, a function-on-scalar regression, allows for assessment of the time points at which the covariates are impacting the profiles. {width="4in"} \[fig:scan\_time\] Material and Methods ==================== In this section, we first describe the patient demographics, followed by the experimental methods, including the image acquisition and the image preprocessing. Next, in the subsection *Longitudinal Profile Pipeline*, we describe a pipeline for extracting the multi-sequence longitudinal profiles of the voxel intensities from the four sequences within lesion tissue. In the subsection *Modeling the Association with Clinical Covariates*, we describe two models that relate the clinical information to the four intensity profiles. Last, in the subsection *Expert Validation*, we describe a study to validate the lesion segmentation and the outcome from the first model. All analysis, except for image preprocessing, was performed in the R environment [@R] using the R package oro.nifti [@oro]. Patient Demographics -------------------- For this analysis, we use 60 subjects scanned at the NINDS, with the earliest scan performed in 2000 and the most recent scan performed in 2008. Three subjects were excluded during the expert validation because it was found that the longitudinal registration had failed, causing overall poor segmentation of areas of abnormal signal intensity. After exclusion of these subjects and subjects that did not have voxels that met a prespecified inclusion criteria, described in detail later in this section, there were 34 subjects left in the analysis. The 34 subjects included in the analysis had an average of 21 scans each (sd 8.0, range \[10, 37\]). Figure \[fig:scan\_time\] shows the time points at which each of the 34 subjects were scanned. Each row of the plot corresponds to a subject, and each point in the plot represents an MRI study, with time from the subject’s baseline visit in years along the horizontal axis. The total follow-up time per subject was on average 2.2 years (sd 1.2, range \[0.9, 5.5\]). The mean age of the subjects at baseline was 37 years (sd 10.1, range \[18,60\]). At baseline, there were 30 subjects with relapsing remitting MS (RRMS) and 4 subjects with secondary progressive MS (SPMS) . There were 20 females and 14 males,14 subjects on disease-modifying treatment, and 2 subjects received steroids at the baseline visit. The disease-modifying treatments and use of steroids for many of these subjects changed at subsequent follow-up visits. Experimental Methods -------------------- ### Image Acquisition and Preprocessing Whole-brain 2D FLAIR, PD, T2, and 3D T1 volumes were acquired in a 1.5 tesla (T) MRI scanner (Signa Excite HDxt; GE Healthcare, Milwaukee, Wisconsin) using the body coil for transmission. The 2D FLAIR, PD, and T2 volumes were acquired using fast-spin-echo sequences and the 3D T1 volume using a gradient-echo sequence. The PD and T2 volumes were acquired as short and long echoes from the same sequence. The scanning parameters were clinically optimized for each acquired image. For image preprocessing, we use Medical Image Processing Analysis and Visualization (http:// mipav.cit.nih.gov) and the Java Image Science Toolkit (http://www.nitrc.org/projects/jist) [@lucas2010java]. We interpolate all images for each subject at each visit to a voxel size of $1$ $mm^3$ and rigidly align all of the T1 volumes to the Montreal Neurological Institute standard space [@fonov2009unbiased]. We rigidly align the T2, FLAIR, and PD image to the T1 for each subject at each visit. We then find the average of the T1 volume for all the studies, for each subject, and register the longitudinal collection of the T1, T2, FLAIR, and PD images for the subject to this average image. We remove extracerebral voxels using a skull-stripping procedure [@carass2007joint]. We automatically segment the entire brain using the T1 and FLAIR images [@shiee2010topology]. After preprocessing, studies were manually quality controlled by a researcher with over three years experience with structural MRI (EMS). Studies with motion or other artifacts were removed. Longitudinal Profile Pipeline ----------------------------- The procedure for extracting the longitudinal voxel level lesion profiles is divided into four steps: (1) identifying voxels with new lesion formation, (2) intensity normalization, (3) temporal alignment, and (4) temporal interpolation. All voxels in this analysis are part of incident or enlarging lesions detected during the subject’s follow-up period. All voxels that are part of lesions that existed at baseline are excluded from the analysis. Figure \[fig:longit\] shows an example of a lesion that is detected during a subject’s follow-up period. The first row of Figure \[fig:longit\] shows the multiple MRI sequences at one time point (from left to right, the FLAIR, T2, PD, and T1 sequences). In each sequence, a red box shows an area with a lesion that develops during the follow-up period. The subsequent 4 rows of the figure show the longitudinal behavior within this red box. Each column of the figure shows a different MRI study, starting at 98 days after baseline in the far left column and going until 343 days after baseline. The lesion in the red box is first seen 175 days after baseline. {width="4.5in"} \[fig:longit\] ### Identifying Voxels with New Lesion Formation When identifying voxels with new lesion formation, we distinguish between areas that contain vasogenic edema (which we will refer to simply as “edema") and actual lesion, which both manifest as areas of abnormal signal intensity, especially on the T2-weighted sequences. For this analysis, we are interested in areas with tissue damage, as opposed to edema. To identify areas with new lesion formation, we first find areas in the MRI with new abnormal signal intensity, which includes both edema and lesion. We then segment lesions by analyzing subsequent visit data (details provided below). SuBLIME segmentation of voxel level lesion incidence and enlargement is a method for detecting voxels that are part of an area of new abnormal signal intensity between two MRI studies [@sweeney2013automatic]. For each subject, we produce SuBLIME maps between the respective sets of consecutive MRI studies. We exclude all abnormal signal intensity areas that contained fewer than 27 voxels, as these areas could be artifact or noise. We then produce cross-sectional lesion segmentations using OASIS segmentation of abnormal signal presence [@sweeney2013oasis]. As the signal from edema disappears rapidly from the MRI after the lesion formation, we locate the incident abnormal signal voxels using SuBLIME, but only include the voxels that are detected by OASIS at the following study visit, as these voxels should not contain edema. Therefore, only voxels that have an MRI study within 40 days after SuBLIME detects the area of abnormal signal intensity, where the intensity remains in the OASIS maps, are considered as lesion tissue and used in this analysis, as by this time edema would subside. We use expert validation by a neuroradiologist and a neurologist, both with experience in MS imaging, to confirm that this method is identifying lesion tissue, which we describe in detail in the subsection *Expert Validation*. Figure \[fig:longit\] shows the SuBLIME segmentation for each study, the OASIS segmentation for each study, and a map of the distance to the boundary of the SuBLIME segmentation. The row corresponding to the SuBLIME segmentation is further divided into edema and lesion voxels using the method described above. Only voxels that are part of lesion tissue are used in the analysis. ### Intensity Normalization Structural MRI is acquired in arbitrary units. Therefore, in addition to pulse sequence similarity, intensity normalization is paramount for comparing intensities in a voxel over time within subject and for comparing voxel intensities between subjects. We normalize each sequence separately on each scan by calculating the mean and standard deviation over a mask of the normal-appearing white matter (NAWM) from the brain segmentation described in the subsection *Image Acquisition and Preprocessing* [@shiee2010topology]. We then subtract the mean from the intensity in each voxel and divide by the standard deviation [@shinohara2011population; @shinohara2014statistical]. Let $S_{ilv} (t)$ be the observed intensity from imaging sequence $S$ in voxel $v$ for subject $i$ in lesion $l$ at study time $t$, with $S = $ FLAIR, T1, T2, and PD. Let $\mu_{Si} (t) $ and $\sigma_{Si} (t) $ be the mean and standard deviation, respectively, over the NAWM mask for sequence $S$ at scan time $t$ for subject $i$ . Then the normalized intensity in voxel $v$ in lesion $l$ for subject $i$ at scan time $t$ is: $$S^N_{ilv} (t) = \frac{S_{ilv}(t) - \mu_{Si}(t) }{ \sigma_{Si} (t) } \nonumber$$ Thus, all image intensities are expressed as a departure, in multiples of standard deviation of white matter intensities, from the subject’s mean normal-appearing white matter (NAWM) in each imaging sequence. The normalized intensity longitudinal profiles from the lesion in Figure \[fig:longit\] for all four sequences can be seen in the first column of Figure \[fig:trajectories\]. From top to bottom in the first column of Figure \[fig:trajectories\] we have the profiles from 150 randomly sampled voxels from the lesion in Figure \[fig:longit\] for the FLAIR, T2, PD, and T1 sequences. Each line in the plot represents the longitudinal profile from a single voxel. The x-axis shows the time in days from the baseline visit, with the point of lesion incidence denoted by a vertical dashed line, and the y-axis shows the normalized sequence intensities. {width="6in"} \[fig:trajectories\] ### Temporal Alignment The date of the study visit at which SuBLIME detects the lesion voxels is considered the time of incidence for this voxel. If a voxel is determined to be a new or enlarging lesion by SuBLIME more than once over the follow-up time, the first occurrence is considered to be the time of lesion incidence for that voxel. Voxel profiles from incident lesions during the follow-up of each subject are aligned in time, using the time of incidence as time 0, therefore any observations before incidence have a negative time and after lesion incidence have a positive time. Let $t'$ denote this aligned time scale. Then we have $S^N_{ilv}(t')$, where $S^N_{ilv}(0)$ indicates the intensity in sequence $S$ at the time of lesion incidence. ### Temporal Interpolation Next we perform a temporal linear interpolation so that all voxels are observed on the same time grid. In this work, we are interested in the lesion dynamics only after lesion incidence, therefore we perform the linear interpolation within the window after lesion incidence and up to 200 days post-incidence. The end point of 200 days is selected as it has been previously found that new T2 lesions show the most dramatic changes in intensity for three to four months [@meier2007time], and we opt to be conservative and include data beyond this reported stabilization point. Voxels are selected for the analysis if the subject has at least one visit 200 days or more after lesion incidence. Of the 60 subjects in this analysis, 34 have voxel profiles meeting this inclusion criteria, after removing the three subjects for poor longitudinal registration. We linearly interpolate over a grid of 0 to 200 days by increments of 5 days so that all profiles are observed on the same time grid. We denote the vector of observations from a voxel over this time grid for sequence $S$ as $S^N_{ilv}$, where $S^N_{ilv}$ is a $1 \times 41$ vector. The second column of Figure \[fig:trajectories\] shows the temporally registered and linearly interpolated profiles over the period of 0 to 200 days for the lesion in Figure \[fig:longit\] for the same 150 randomly sampled voxels as shown in the first column. Modeling the Association with Clinical Covariates ------------------------------------------------- We propose two models for relating the information in the lesion profiles to the clinical covariates. The clinical information for each subject at each study visit is MS disease subtype, age, sex, an indicator of treatment with steroids, an indicator of disease-modifying treatment, and distance to the boundary of an area of abnormal signal intensity. An example of distance to the boundary of an area of abnormal signal intensity for the lesion in Figure \[fig:longit\] can be seen in the eighth row of the Figure \[fig:longit\]. We center age at the mean age over all of the voxel level observations of 36 years. During the observation period, many of the subjects were enrolled in clinical trials at NINDS to test various experimental therapies. Our indicator of disease-modifying treatment indicates treatment with any of the Food and Drug Administration approved treatments interferon beta 1-a (intramuscular or subcutaneous), interferon beta 1-b, and glatiramer acetate, as well as experimental therapy. As many of the covariates change over time, we model the relationship between the lesion profiles and the value of the covariate at the time of lesion incidence for the particular profiles. For the following analysis, we have a total of 57,908 voxels from 315 lesions in 34 subjects. ### Principal Component Analysis and Regression To perform PCA, we first concatenate the voxel profiles for each voxel from the four sequences together. For each sequence and at each voxel there is a $1 \times 41$ vector of longitudinal intensities, $S^N_{i l v}$. Let $I_{ilv}$ denote the $1 \times 164$ dimensional vector of the four concatenated profiles, $S^N_{i l v}$, from subject $i$ lesion $l$ and voxel $v$. More precisely, $$I_{ilv} = \{ FLAIR^N_{ilv}, T1^N_{ilv}, T2^N_{ilv}, PD^N_{ilv} \}$$ and we index the entries of $I_{ilv}(j)$, where $j=1,\ldots,164$ is the $j^{th}$ entry of the concatenated vector. Note that we first remove the mean from the concatenated profiles and then perform a PCA on these concatenated profiles. Let $\phi_k$ denote the $k^{th}$ PC, where $\phi_k$ is also indexed by $j$. The relationship between the score on the $k^{th}$ PC, $\xi_{ilv}(k)$, a scalar, and the observed trajectory for $I_{ilv} \left( j \right)$ is: $$I_{ilv} \left( j \right) =\sum_{k=1}^K \xi_{ilv}(k) \phi_k \left( j \right) ,$$ with $\xi_{ilv}(k)$ being the scalar score on the $k^{th}$ eigenfunction. We will focus on the first PC, $\phi_1$, and the score on this component, $\xi_{ilv}(1) $. The first PC is found to identify slow, long-termintensity changes at the voxel level within lesions, with positive PC scores corresponding to a return of the voxel to intensity values closer to that of normal-appearing tissue and negative scores corresponding to the voxels maintaining intensity values closer to those at lesion incidence. We use expert validation from a neuroradiologist and a neurologist to confirm this finding, with details provided in the subsection *Expert Validation*. The score on the first PC, $\xi_{ilv}(1)$, collapses the full profiles at each voxel from the four sequences into a single scalar. We use only the score on the first PC in this analysis, as the other PCs explain only $25\%$ of the variation in the data and were not found to identify any biological processes. To assess variability in both the mean and the first PC, we bootstrap this procedure by resampling subjects with replacement 1000 times [@efron1994introduction]. We now introduce a linear mixed-effects model to relate the score on the first PC to the clinical covariates [@mcculloch2001generalized]. We use the value of the covariate at the time of lesion incidence for the particular profiles, which can vary within subject. Thus, for added precision, the covariates that change over time are indexed by the subject index $i$, lesion index $l$ and voxel index $v$, as voxels from the same lesion may have different times of incidence. For example, the sex of the subject does not change by time of lesion incidence, so it is only indexed by $i$. In contrast, age of the subject changes with lesion incidence and is indexed by $i$, $l$ and $v$. We also add random effects for subject and lesion, which we denote by $b_i$ and $b_l$, respectively, with both following a normal distribution: $b_i \sim N \left(0, \sigma_i^2 \right)$ and $b_l \sim N \left(0, \sigma_l^2 \right)$, where $\sigma^2$ denotes the variance of the random effects. We consider the following basic model for the association between $\xi_{ilv}(1)$ and the covariates: $$\begin{aligned} \xi_{ilv}(1) & = \beta_0 + \beta_1 \text{SPMS}_{ilv} + \beta_{2} \text{ Distance} _{ilv} + \beta_3 \text{Age}_{ilv} + \beta_4 \text{ (Age $-$ 4)}_{+ ilv} \nonumber \\ & + \beta_{5} \text{Steroids}_{ilv} + \beta_{6} \text{Male}_{i} + \beta_{7} \text{Treatment}_{ilv} + b_i + b_l + \epsilon_{ilv} \nonumber \end{aligned}$$ We assume that the error terms are independent identically distributed, with each $\epsilon_{ilv}$ following a normal distribution, $\epsilon_{ilv} \sim N \left(0, \sigma_{\epsilon}^2\right) $. In the model, the term SPMS is an indicator of having SPMS where the comparison group is RRMS. Note that the age term has been centered at the mean age of 36 years. The term $\text{ (Age $-$ 4)}_{+ ilv} = \text{Age}_{ilv} \cdot 1( \text{Age}_{ilv}>4)$ is a spline term for centered age over 4 years (or age over 40 years), which was included in the model after visualizing the relationship between the scores on the first PC and age. This visualization can be found in the Appendix. We also investigated simpler models with the same mixed-effects structure, but where we considered each covariate separately. To test for associations, we use two procedures. First, we perform a parametric bootstrapping procedure [@efron1994introduction], and second we calculate p-values using a normal approximation for the distribution of the fixed-effects in the mixed-effects model [@barr2013random]. We use 1000 bootstrap samples for the bootstrap procedure. We perform the parametric bootstrap because steroid use and disease subtype of SPMS did not always appear in the nonparametric bootstrap sample. A complete description of the parametric bootstrap procedure is found in the Appendix. We also use the normal approximation, as this approximation has been found to be a reasonable approximation for the distribution of the fixed-effects in most settings [@barr2013random]. Expert Validation ----------------- We use expert validation to determine the quality of the lesion segmentation (excluding edema tissue) and the score on the first PC for identifying areas of slow, long-term intensity change. For the validation we use two raters, a neuroradiologist with 11 years of experience with research in MS (DSR) and a neurologist with 4 years of experience with research in MS (MKS). For each lesion, we first determine the axial slice of the image that contains the largest number of voxels with abnormal signal intensity. Then for each lesion the two raters are presented the following: (1) the full axial slice for the FLAIR, T2, PD, and T1 volumes that contains the largest number of voxels with abnormal signal intensity; (2) the entire collection of longitudinal scans for a box containing the abnormal signal intensity in the FLAIR, T2, PD, and T1 volumes for this axial slice; (3) the segmentation of the lesion and edema tissue within this box; (4) the score on the first PC for the voxels segmented as lesion tissue within this box and a scale for the intensities within this image; (5) the entire collection of longitudinal scans for the FLAIR, T2, PD, and T1 weighted volumes within this box with the score for the first PC overlaid on the images for each scan after lesion incidence. The raters are then asked to rate the quality of the lesion tissue segmentation and the score on the first PC for identifying areas of slow, long-termintensity changes on an integer scale from 1 to 4, with each rating corresponding to the following: (1) failed miserably; (2) some redeeming features; (3) passed with minor errors; and (4) passed. Examples of the images presented to the raters for each lesion that received a rating of 1,2,3, and 4 for the score on the first PC by both raters are provided in the Appendix. Forty-seven lesions are selected at random to be repeated in the analysis to assess intra-rater reliability. We report the median of the ratings of the lesion segmentation and the score on the first PC for each rater over all lesions. To assess between-rater reliability, we report the Coehn’s kappa coefficients for the two raters separately, for both the ratings, over all of the lesions. We also assess the within-rater agreement by reporting the kappa coefficient for the set of repeated lesions for each rater, for both the ratings. We also report the kappa coefficient for the rating of the lesion segmentation and the score on the first PC for all lesions, for each rater, to determine if the quality of the segmentation and the quality of the score on the first PC are related. We nonparametrically bootstrap with replacement the subjects 1000 times to produce the confidence intervals for the median of the ratings for each rater and the kappa coefficients for within- and between-rater agreement as well as the relationship between the quality of the segmentation and the score on the first PC. ### Function-on-Scalar Regression The previous model is an attempt to collapse the information from the four profiles (across sequences) for a voxel into a single scalar at each voxel. As an alternative, we also fit a function-on-scalar regression model [@reiss2010fast], where we can investigate the relationship between the covariates of interest and the profile at each time point. We fit a function-on-scalar regression model for each sequence separately. For simplicity of notation, we now use $t$ for the registered time, as opposed to $t'$. The outcome in the model is the full lesion intensity profile: $$\begin{aligned} S^N_{i l v}(t) & = \beta'_0(t) + \beta_1'(t) \text{SPMS}_{ilv} + \beta_{2}'(t) \text{ Distance} _{ilv} + \beta_3'(t) \text{Age}_{ilv} + \beta_4'(t) \text{ (Age $-$ 4 )}_{+ ilv} \nonumber \\ & + \beta_{5}'(t) \text{Steroids}_{ilv} + \beta_{6}'(t) \text{Male}_{i} + \beta_{7}'(t) \text{Treatment}_{ilv} + \epsilon_{ilv} \left( t \right) \nonumber \end{aligned}$$ for $S = $ FLAIR, T1, T2, and PD. To fit the model, we use a two-step function-on-scalar regression implemented in the R package refund [@refund]. The procedure first fits a scalar-on-scalar regression at each individual time point. Then the resulting coefficient functions are smoothed over time using a cubic spline basis with a penalty on the second derivative. To assess the variability in the coefficient functions and provide bootstrapped ,point wise $95\%$ confidence intervals, we non-parametrically bootstrap subjects 1000 times. When samples do not contain subjects with a covariate, for example the indicator of steroids, we remove this sample from the bootstrap and replace it with another sample. The difference between the function-on-scalar regression and the PCA regression model is that PCA collapses the entire temporal intensity profile of the voxel into a scalar. By contrast the function-on-scalar regression investigates the association at every time point. While function-on-scalar regression is more comprehensive and interpretable, it is more appropriate when there are strong functional effects that are not captured by one or two principal components. Results ======= Principal Component Analysis and Regression ------------------------------------------- In Figure \[fig:PCload\] A we show the mean profiles for each sequence over the registered 200 day period, and in Figure \[fig:PCload\] B we show the first PC for each sequence over the registered 200 day period. The first PC explains 75% (95% CI: \[72%, 76%\]) of the variation in the concatenated longitudinal profiles. The first PC, $\phi_1$, is split into the different sequences for purposes of presentation. The subfigures for both the mean and the first PC show the bootstrapped $95\%$ confidence intervals. Recall that the normalization procedure puts the volumes into units of standard deviations above the mean of the NAWM. Therefore a value of 0 on the image corresponds to the average value of NAWM from the particular MRI scan. The mean profiles for the FLAIR, T2 and PD are all above 0 throughout the time course, as lesions are hyperintense on these sequences. In contrast, the mean profile for the T1 sequence is below 0, as lesions are hypointense on this sequence. The first PC for the FLAIR, T2 and PD is negative throughout the time course, with values closer to 0 at lesion incidence (time 0). Positive scores on this PC indicate a decrease in the signal in these sequences, which corresponds to a return of the voxel to intensity values closer to that of normal-appearing tissue. In contrast, negative scores indicate the voxel maintaining intensity values closer to those at lesion incidence, with more hypointensity than the average profile. Similarly, for T1 the first PC is positive throughout the time course, with values closer to 0 at lesion incidence. Positive scores on this PC indicate increased signal on the T1. As lesions are hypointense on the T1, this also indicates a return of the voxel to intensity values closer to that of normal-appearing tissue. Negative scores again correspond to the voxels maintaining intensity values closer to those at lesion incidence. Therefore, we consider the score on the first PC to capture slow, long-termintensity changes within the lesion at a voxel level. In the last row of Figure \[fig:longit\] we see the PC scores from the lesion that is shown in the figure. Here we see that the positive scores indicate areas of the lesion that return to values of normal-appearing tissue, while the negative scores show areas that remain at the intensity values at lesion incidence. {width="6in"} \[fig:PCload\] {width="6in"} \[fig:coefficient\] In the second-stage analysis, we fit both univariate and multivariate mixed-effects models to investigate the relationship between the covariates and the score on the first PC. The estimates of the coefficients from both models are shown in the bar plots in Figure \[fig:coefficient\], with asterisks indicating statistical significance at the $5\%$ level using the bootstrapped 95% confidence intervals. Tables containing the coefficient estimates, standard errors, t-statistics, p-values using the normal approximation, and 95% bootstrapped confidence intervals can be found in the Appendix for both the univariate and the multivariate models. There are no differences in the conclusions determined by the normal approximation and the bootstrapped 95% confidence intervals. For continuous covariates, such as age, the coefficient is interpreted as the expected change in the score on the first PC for a one unit increase in the covariate. For binary variables, such as disease subtype, the coefficient is interpreted as the difference in the expected change in the first PC in the specified group. Therefore, positive coefficients are indicative of the voxel returning to intensity values closer to normal-appearing tissue with an increase in the covariate, while negative coefficients are indicative of the voxel maintaining the intensities at lesion incidence with an increase in the covariate (or in some rare cases having intensities that have an increasing departure from those of normal-appearing tissue over time with an increase in the covariate). The results indicate that voxels that are farther away from the boundary have increased risk for maintaining abnormal signal intensity. In this model, the coefficient for distance to the boundary has a value of -9.39 (95% CI: \[-9.56, -9.25\] ), indicating that for a one voxel (or $1$ $mm$) increase in distance away from the boundary (toward the center of the lesion) the average value of the first PC decreases by 9.39, adjusting for the other coefficients and the random effects. In the last row of Figure \[fig:longit\], we see this spatial relationship between the score on the first PC and the distance to the lesion boundary; with positive scores near the boundary and negative scores near the center of the lesion. In both models, we found the use of disease-modifying treatment and steroids to be associated with return of a voxel to the value of normal-appearing tissue. The coefficient for treatment has a value of 5.39 (95% CI: \[4.67, 6.08\]), indicating that when subjects are on treatment the average value of the first PC increases by 5.39, adjusting for the other coefficients and the random effects. The use of steroids has a similar interpretation, with a coefficient value of 4.26 (95% CI: \[2.67, 5.85\]). In the Appendix, density plots for the score on the first PC by disease subtype, treatment with steroids, disease-modifying treatment, and sex are provided. Expert Validation ----------------- Expert validation is used to determine the quality of the lesion segmentation (excluding edema tissue) and the score on the first PC for identifying areas of slow, long-termintensity change. The distributions of the ratings for the two raters for both the lesion segmentation and the score on the first PC are shown in Figure \[fig:ratings\]. The first row of plots in Figure \[fig:ratings\] shows the distribution of the ratings for the lesion segmentation and the second row shows the ratings for the score on the first PC. Plots in the left column are ratings by the neuroradiologist, and plots on the right column are ratings by the neurologist. The median rating for both the lesion segmentation and the first PC by the neuroradioloist are 4 (95% CI: \[4,4\]), which is a rating of passed, the highest possible rating. The median rating for both the lesion segmentation and the first PC by the neurologist are 3 (95% CI: \[3,3\]), which is a rating of passed with minor errors. Note that criteria for assigning scores were not discussed between the two raters prior to their respective analyses. {width="6in"} \[fig:ratings\] The Cohen’s kappa coefficients for the within- and between-rater agreement for both the lesion segmentation and the score on the first PC are shown in Table \[table:pearson\]. The values for the kappa coefficient range between $0$ and $1$, with a value of $1$ indicating total agreement and $0$ indicating no agreement. The within-rater agreement for the lesion segmentation and the score on the first PC are higher for the neuroradiologist than the neurologist. There is only modest agreement between the neuroradiologist and neurologist on both ratings, with a kappa coefficient of 0.29 (95% CI: \[0.18, 0.41\]) for the lesion segmentation and 0.24 (95% CI: 0.11, 0.39) for the score on the first PC. This is due, in part, to the fact that the neurologist spread ratings of the studies between 3 and 4, while the neuroradiologist gave more ratings of 4. The kappa coefficient for the agreement between the rating of the lesion segmentation and the score on the first PC is 0.97 (95% CI: 0.93, 1.00) for the neuroradiologist and 0.68 (95% CI: 0.58, 0.78) for the neurologist. The high correlation between these ratings, especially for the neuroradiologist, indicates that the quality of the segmentation impacts the quality of the score on the first PC. Comments from the raters mirrored this finding, as many of the low scores for both the lesion segmentation and the first PC were due to: (1) missing the first time point of lesion incidence and segmenting it as new lesion at a later time point; (2) not segmenting the entire lesion; and (3) parts of the same lesion being segmented (unnecessarily) at different time points. As both the ratings for the lesion segmentation and the score on the first PC were high, the quality of the lesion segmentation does not appear to be negatively impacting the method. [ |l|l|l| ]{}\ & Neuroradiologist & Neurologist\ Neuroradiologist & 0.92; (0.76,0.99) & 0.29; (0.18, 0.41)\ Neurologist & & 0.75; (0.62, 0.86)\ [ |l|l|l| ]{}\ Neuroradiologist & Neurologist\ 0.92; (0.76, 0.99) & 0.24; (0.11, 0.39)\ & 0.72; (0.51, 0.86)\ Function-on-Scalar Regression ----------------------------- The resulting coefficient functions from the function-on-scalar regression with bootstrapped, point wise $95\%$ confidence intervals with the FLAIR profile as the outcome are shown in Figure \[fig:fosrflair\]. Similar figures for the T2, PD, and T1 profiles as outcome are provided in the Appendix. The coefficient functions for continuous variables in the function-on-scalar regression model are interpreted as the change in the expected profile at each time point for a one unit increase in the covariate. Similarly, for binary variables, the coefficient function is interpreted as the change in the expected profile for the specified group. For the FLAIR profiles, the coefficient functions corresponding to distance to the boundary and age have bootstrapped $95\%$ confidence intervals that do not overlap with 0 across any of the time points, and are therefore statistically significant at the $.05$ level. The coefficient function for distance to the boundary is greater than 0 throughout the entire trajectory, indicating that the further away from the boundary the voxel is, the more the FLAIR hyperintensity is maintained within the voxel. For a one voxel (or $1$ $mm$) increase in distance away from the boundary (toward the center of the lesion) the average normalized intensity of the trajectory increases by around $0.5$ at all time points, adjusting for the other coefficients and the random effects. The result for distance from the boundary agrees with the results from the PCA regression model. {width="6in"} \[fig:fosrflair\] Discussion {#discussion .unnumbered} ========== We describe a pipeline to extract voxel level, multi-sequence longitudinal profiles from the FLAIR, T2, PD, and T1 sequences within new lesions, as well as two models to relate the clinical information to the four profiles. The methodology presented here shows promise for both understanding the time course of tissue damage in the disease MS and may prove to be useful for evaluating the impact of neuroprotective or reparative treatments for the disease, for example in clinical trials. Indeed, reliable methods to evaluate such treatments, which are currently in development, are lacking at present. In contrast to prior studies of change in lesion intensity in clinical trials, our work is focused on voxel level analysis, and therefore it can provide spatial information about intensity recovery and does not artificially reduce the size of the data set. This may have implications on the sample size calculations for clinical trials. These methods are also broadly applicable to other imaging modalities and disease areas, in which longitudinal intensity profiles may lead to more sensitive and specific biomarkers. The inference from both of the models in regards to disease-modifying treatment should only be taken as a proof-of-concept for the relationship between the imaging and the clinical covariates. The models may suffer from confounding by indication, which arises when individuals who are on a treatment are different from those who do not receive treatment, due to unobserved considerations. In the multivariate model, we adjust for age, sex, and disease subtype, but unobservable differences related to treatment choice may cause biased results. However, bias in terms of treatment effect would most plausibly result in underestimation of improvements, as more aggressive therapies are usually given to subjects with more aggressive or refractory disease. Thus, our findings might underestimate what would be observed in a randomized trial of disease-modifying therapy. While many of the coefficient functions from the function-on-scalar regression are not found to be statistically different from 0, this model may have more power with more subjects. For the bootstrap procedure we only have 34 subjects, resulting in wide confidence intervals for the estimated coefficient functions. In contrast, the regression using score outcomes identifies strong associations between specific covariates and longitudinal patterns of longitudinal intensities. The two models presented in this work are fit voxel-wise and therefore may be sensitive to major misregistration within a study and between longitudinal studies for the same subject. The models are also sensitive to local displacement of tissue due to transient swelling in and around or resorption of lesion tissue. We therefore do not call the slow changes in intensity within the voxels that are observed “tissue repair," as we cannot be certain that the change is not due to misregistration or displacement of tissue from the lesions themselves. We do observe a relationship between the return of voxels to the intensity of normal-appearing tissue and both disease-modifying treatment and treatment with steroids, and therefore find this measure useful and deserving of further study. We also see an association with the distance to the boundary of the lesion and slow, long-termintensity changes – with voxels near the boundary of the lesion returning to baseline intensity and voxels near the center of the lesion maintaining abnormal signal intensity. Future work to assess tissue repair may involve investigating a nonlinear registration within individual lesions. The methods described here use only conventional clinical imaging for patients with MS, namely FLAIR, T2, PD, and T1. While this is beneficial for using the methodology in a clinical trial setting or for analysis of retrospective imaging studies, one could also incorporate advanced imaging into the method. For example, magnetization transfer ratio imaging [@van1999axonal], quantitative T1-weighted imaging [@filippi2000quantitative], and diffusion tensor imaging [@filippi2001diffusion] have been studied in MS lesions. The longitudinal dynamics of lesions on these images could be incorporated into our framework to better understand the behavior of lesions over time and the impact of disease-modifying therapies on this behavior. For this analysis, all MRI studies are acquired on a single 1.5 T MRI scanner at one imaging center. Similar analysis could be performed at higher field strength, but for this analysis we use a 1.5 T dataset for the availability of the large retrospective cohort study over a long period of time. Although different scanning parameters were used for the acquisitions, further investigation is warranted into the robustness of the methods to changes in scanner, changes in magnetic field strength, as well changes in the imaging center. Acknowledgments {#acknowledgments .unnumbered} =============== The project described is supported in part by the NIH grants RO1 EB012547 from the National Institute of Biomedical Imaging and Bioengineering, RO1 NS060910 and RO1 NS085211 from the National Institute of Neurological Disorders and Stroke (NINDS), T32 AG021334 from the National Institute of Aging, and RO1 MH095836 from the National Institute of Mental Health. The research is also supported by the Intramural Research Program of NINDS. The content is solely the responsibility of the authors and does not necessarily represent the official views of the funding agencies Appendix ======== Parametric Bootstrapping Procedure {#parametric-bootstrapping-procedure .unnumbered} ---------------------------------- Let $B$ be the number of bootstrap samples to be performed and let $b$ index these $B$ samples. Let $Y_{ilv}$ be the outcome for an observation indexed by $i$, $l$, and $v$. Let $\boldsymbol{X}$ be the design matrix and $\boldsymbol{\beta}$ be the vector of the coefficients. For this analysis we have a model of the form: $$\begin{aligned} Y_{ilv} & = \boldsymbol{X \beta} + b_i + b_l + \epsilon_{ilv} \nonumber \end{aligned}$$ where $b_i \sim N \left(0, \sigma_{i}^2 \right)$ and $b_l \sim N \left(0, \sigma_{l}^2 \right)$ are random intercepts, and $\epsilon_{ilv} \sim N \left(0, \sigma_{\epsilon}^2 \right)$ is an error term. For the parametric model, we fit the above mixed-effect model to get an estimate of $\boldsymbol{\beta}$, which we denote as $\boldsymbol{\hat \beta}$. We then fix this estimate, and keep $\boldsymbol{X \hat \beta}$. Using the fitted variances, $\hat \sigma_i^2$, $\hat \sigma_l^2$ and $\hat \sigma_{\epsilon}^2$, we generate a random intercept for each lesion from a $N \left(0, \hat \sigma_{i}^2 \right)$ distribution, a random intercept for each subject from a $N \left(0, \hat \sigma_{l}^2 \right)$, and random noise for each voxel from a $N \left(0, \hat \sigma_{\epsilon}^2 \right)$. We then add the random intercepts and noise to $\boldsymbol{X \hat \beta}$ for the corresponding observation and use this as our outcome to refit the model and get out bootstrapped coefficient vector $\boldsymbol{\beta_{b} ^{*}}$. To obtain the bootstrap sample, we repeat this procedure $B$ times. Expert Validation {#expert-validation-2 .unnumbered} ----------------- Examples of the set of evaluation images presented to the experts for each lesion are shown in Figures \[fig:rate4\], \[fig:rate3\], \[fig:rate2\], and \[fig:rate1\]. The first row of the figures shows the full axial slice for the FLAIR, T2, PD, and T1 volumes that contains the largest number of voxels with abnormal signal intensity. The second through fourth rows show the entire collection of longitudinal scans for a box containing the abnormal signal intensity in the FLAIR, T2, PD, and T1 weighted volumes for this axial slice. The scans are displayed in chronological order, from first time point to last time point, from left to right. The fifth row shows the segmentation of the lesion and edema tissue within this box at each time point. The sixth row shows the score on the first PC for the voxels segmented as lesion tissue, displayed at the time of lesion incidence for each voxel. The seventh throughout tenth row shows the entire collection of longitudinal scans for the FLAIR, T2, PD, and T1 weighted volumes within this box with the score for the first PC overlaid on the images for each scan after lesion incidence. The last row shows the scale for the score on the first PC. The figures show examples of the four different ratings for the score on the first PC. Both raters rate the scans as either (1) failed miserably, (2) some redeeming features, (3) passed with minor errors or (4) passed. {width="6in"} \[fig:rate4\] {width="6in"} \[fig:rate3\] {width="6in"} \[fig:rate2\] {width="6in"} \[fig:rate1\] Principal Component Analysis and Regression {#principal-component-analysis-and-regression-2 .unnumbered} ------------------------------------------- Table \[table:multivariate\] shows the coefficient estimates, standard errors, t-statistics, the p-values using the normal approximation, and the 95% bootstrapped confidence intervals for the multivariate PCA regression model. Table \[table:univariate\] shows the same for the individual univariate PCA regression models. Estimate Standard Error t-value p-value 95% Bootstrapped CI ---------------------- ---------- ---------------- --------- --------- --------------------- SPMS 2.15 4.41 0.49 0.63 (-6.19, 10.93) Distance to Boundary -9.39 0.08 -123.74 0.00 (-9.56, -9.25) Age -0.21 0.18 -1.16 0.25 (-0.57, 0.13) (Age $-$ 4)$_{+}$ -0.10 0.23 -0.42 0.68 (-0.54, 0.35) Steroids 4.26 0.79 5.42 0.00 (2.67, 5.85) Male 1.16 2.55 0.45 0.65 (-3.94, 6.61) Treatment 5.39 0.36 15.03 0.00 (4.67, 6.08) Intercept 8.89 1.92 4.64 0.00 (5.17, 12.85) : **Coefficient estimates, standard errors, t-statistics, p-values, and bootstrapped 95% confidence intervals for the multivariate PCA regression model.**[]{data-label="table:multivariate"} Estimate Standard Error t-value p-value 95% Bootstrapped CI ---------------------- ---------- ---------------- --------- --------- --------------------- SPMS 0.65 4.11 0.16 0.88 ( -7.71, 9.18) Distance to Boundary -9.37 0.08 -123.18 0.00 (-9.52, -9.22) Age 0.89 0.19 4.58 0.00 (0.51, 1.23) (Age $-$ 4)$_{+}$ -1.55 0.24 -6.40 0.00 (-1.95, -1.14) Steroids 6.03 0.78 7.77 0.00 (4.55, 7.59) Male 0.43 2.43 0.18 0.86 (-4.32, 4.97) Treatment 4.48 0.38 11.76 0.00 (3.67, 5.25) : **Coefficient estimates, standard errors, t-statistics, p-values, and bootstrapped 95% confidence intervals for the univariate PCA regression model.**[]{data-label="table:univariate"} Density Plots {#density-plots .unnumbered} ------------- Density plots for the score on the first PC by disease subtype, treatment with steroids and disease modifying therapies, and sex are shown below. A plot of the score on the first PC versus age is also shown below. Disease subtype and sex were not found to be statistically significant in the linear mixed-effects model, as there were few subjects in our analysis with SPMS or that were male. {width="5in"} \[fig:den\] {width="5in"} \[fig:fosrt2\] Function-on-Scalar Regression {#function-on-scalar-regression-2 .unnumbered} ----------------------------- The coefficient functions from the function-on-scalar regression with bootstrapped $95\%$ confidence intervals with the T2, PD, and T1 profile as the outcome are shown below. Similar to using the FLAIR profile as the outcome, only the distance to the boundary and age were found to be different from 0 at any point along the profile. {width="5in"} \[fig:fosrt2\] {width="5in"} \[fig:fosrpd\] {width="5in"} \[fig:fosrt1\]

ArXiv

--- abstract: 'Human activity recognition has drawn considerable attention recently in the field of computer vision due to the development of commodity depth cameras, by which the human activity is represented as a sequence of 3D skeleton postures. Assuming human body 3D joint locations of an activity lie on a manifold, the problem of recognizing human activity is formulated as the computation of activity manifold-manifold distance (AMMD). In this paper, we first design an efficient division method to decompose a manifold into ordered continuous maximal linear patches (CMLPs) that denote meaningful action snippets of the action sequence. Then the CMLP is represented by its position (average value of points) and the first principal component, which specify the major posture and main evolving direction of an action snippet, respectively. Finally, we compute the distance between CMLPs by taking both the posture and direction into consideration. Based on these preparations, an intuitive distance measure that preserves the local order of action snippets is proposed to compute AMMD. The performance on two benchmark datasets demonstrates the effectiveness of the proposed approach.' author: - | Yaqiang Yao, Yan Liu, Huanhuan Chen,\ School of Computer Science and Technology,\ University of Science and Technology of China, Hefei, China\ [email protected], [email protected], [email protected] bibliography: - 'reference.bib' title: | Skeleton-based Activity Recognition\ with Local Order Preserving Match of Linear Patches --- Introduction {#sec:introduction} ============ In computer vision and machine learning communities, human activity recognition has become one of the most appealing studies [@vrigkas2015review; @xu2017learning] for its wide applications. Previous RGB-based work focused on extracting local space-time features from 2D images. Recently, with the introduction of real-time depth cameras and corresponding human skeleton extraction methods [@shotton2013real], the studies of activity recognition have been greatly promoted in terms of depth maps-based methods [@rahmani2016histogram] and skeleton-based methods [@wang2014learning]. In particular, [@yao2011does] verified that skeleton data alone can outperform other low-level image features for human activity recognition. The main reason is that the 3D skeleton poses are invariant to the viewpoint and appearance, such that activities vary less from actor to actor. Several specially designed descriptors such as HOJ3D [@xia2012view], Cov3DJ [@hussein2013human], and HOD [@gowayyed2013histogram] employed the conclusion and achieve decent performance. Related Work ------------ Different from human posture recognition, the temporal relation between adjacent frames poses a challenge to the activity recognition task. Human body represented with 3D skeleton can be viewed as an articulated system in which rigid segments are connected with several joints. In this way, we can treat a human activity as an evolution of the spatial configuration of these segments. Based on this perspective, [@gong2014structured] addressed the human activity recognition as the problem of structured time series classification. From the view of temporal dynamics modeling, existing methods for human activity recognition fall into two categories: state space models and recurrent neural networks (RNNs). State space models, including linear dynamic system [@chaudhry2013bio], hidden Markov model [@lv2006recognition; @wu2014leveraging] and conditional restricted Boltzmann machine [@taylor2007modeling], treated the action sequence as an observed output produced by a Markov process, whose hidden states are used to model the dynamic patterns. In contrast, RNNs utilize their internal state (memory) instead of dynamic patterns to process an action sequence of inputs [@du2015hierarchical; @zhu2016co; @song2017end]. However, [@elgammal2004inferring] showed that the geometric structure of activity would not be preserved in temporal relation and proposed to learn the representation of activity with manifold embedding. In particular, the authors nonlinearly embedded activity manifolds into a low dimensional space with LLE and found that the temporal relation of the input sequence was preserved in the obtained embedding to some extent. Manifold based representation and related algorithms have attracted much attention in image and video analysis. In consideration of temporal dimension, [@wang2007learning] exploited locality preserving projections to project a given sequence of moving silhouettes associated with an action video into a low-dimensional space. Modeling each image set with a manifold, [@wang2012manifold] formulated the image sets classification for face recognition as a problem of calculating the manifold-manifold distance (MMD). The authors extracted maximal linear patches (MLPs) to form nonlinear manifold and integrated the distances between pairs of MLPs to compute MMD. Similar to image sets that each set is composed of images from the same person but covering variations, human body 3D joint locations of an activity can be viewed as a non-linear manifold embedded in a higher-dimensional space. However, in this case, MLP is not a proper decomposition for activity manifold since it may disorder the geometric structure of action sequence. Our Contributions ----------------- In this paper, we propose a new human activity recognition approach based on the manifold representation of 3D joint locations by integrating the advantages of the temporal relation modeling with the manifold embedding. Rather than modeling the dynamical patterns of the sequence explicitly, manifold learning methods preserve the local geometric properties of activity sequence by embedding it into a low-dimensional space. In this way, human activity is denoted as a series of ordered postures residing on a manifold embedded in a high dimensional space. To construct the sequence of meaningful low-dimensional structures on an activity manifold, we design an efficient division method to decompose an action sequence into the ordered CMLPs based on the nonlinearity degree. Different from the division method proposed in [@yao2018human] which divides the action sequence into two sub-sequences each time, our division algorithm is more flexible in that an action sequence can be divided into more than two sub-sequences according to a predefined threshold. The CMLP corresponding to an action snippet is regarded as a local maximal linear subspace. Motivated by the Cov3DJ proposed in [@hussein2013human], we combine the major posture of action snippet with the main direction of evolution to represent the local maximal linear subspace. In particular, the major posture and main direction are computed with the mean of joints locations and the first principal component of the corresponding covariance matrix, respectively. Based on the intuition that a reasonable distance measure between actions snippets should take both the major posture distance (MPD) and the main direction distance (MDD) between action snippets into consideration, we define the activity manifold-manifold distance (AMMD) as the pairwise matching of adjacent action snippets in the reference and the test activity manifolds to preserve the local order of action snippets. Our approach is evaluated on two popular benchmarks datasets, KARD dataset [@gaglio2015human] and Cornell Activity Dataset (CAD-60) [@sung2012unstructured]. Experimental results show the effectiveness and competitiveness of the proposed approach in comparison with the state-of-the-art methods. In summary, the main contributions of this paper include three aspects: - We design an efficient division method to decompose an activity manifold into ordered continuous maximal linear patches (CMLPs) with $k$ sequential neighbors graph. - A reasonable distance measure between CMLPs that takes into account both the major posture and the main direction of an action snippet is defined. - Based on the distance between CMLPs, an activity manifold-manifold Distance (AMMD) that incorporates the sequential property of action snippets is proposed to discriminate different activities. The Proposed Approach {#sec:recognition} ===================== This section presents the proposed approach for human activity recognition. We first describe the algorithm for the construction of continuous maximal linear patch (CMLP), which decomposes an activity manifold into a sequence of CMLPs viewed as action snippets. Next, we represent CMLP with major posture and main direction, and propose the definition of the distance measure between CMLPs based on this representation. Finally, the activity manifold-manifold distance (AMMD) is computed to discriminate the different activities. {width="\textwidth"} Continuous Maximal Linear Patch ------------------------------- Local linear models on a manifold are linear patches, whose linear perturbation is characterized by the deviation of the geodesic distances from the Euclidean distances between points. Here the Euclidean distance and the geodesic distance are computed with $l_{2}$-norm and the Dijkstra’s algorithm, respectively. The Dijkstra’s algorithm is based on the $k$ nearest neighbors graph in which each vertex is connected to its nearest $k$ vertices with the Euclidean metric. We extend the previous MLP [@wang2012manifold] to a new concept termed continuous maximal linear patch (CMLP). The aim of the construction algorithm is to guarantee that each CMLP only contains meaningful successive postures so that it can be regarded as an action snippet. In view of a rational hypothesis that adjacent postures would close to each other in Euclidean metric, we define a $k$ sequential neighbors graph to compute the geodesic distance between postures in action sequence as follows: **$k$ sequential neighbors graph:** A graph in which each vertex are connected to its previous and next $k$ vertices in temporal order. Formally, a human activity is a sequence of human postures $\mathbf{P}=\{\mathbf{p}_1,\mathbf{p}_2,\cdots,\mathbf{p}_F\}$, where $\mathbf{p}_f \in \mathbb{R}^{D}$ is a $D$-dimensional column vector ($D=3\times J$ is 3 coordinates of $J$ human joints), and $F$ is the number of postures. Assume these postures lie on a low-dimensional manifold $\mathcal{M}$ composed of several subspaces, we aim to construct a sequence of CMLPs $\textmd{C}_{i}$ from $\textmd{P}$, $$\begin{split} &\textmd{P}=\{\textmd{C}_1,\textmd{C}_2,\cdots,\textmd{C}_m\}, \\ &\textmd{C}_{i}|_{i=1}^{m}=[\mathbf{p}^i_1,\mathbf{p}^i_2,\cdots,\mathbf{p}^i_{F_i}], \quad \sum_{i=1}^{m} F_i = F, \end{split}$$ where $m$ is the total number of CMLPs and each action snippet $\textmd{C}_{i}$ contains $F_{i}$ postures. An efficient division method based on nonlinear score is proposed to construct CMLP. In particular, the current action snippet only contains the first posture, and we include the next posture to current action snippet until the nonlinear score of current action snippet exceeds a defined threshold $\delta$. The next action snippet initialized with empty set continues this process. An illustration of constructed CMLPs is presented in Figure \[fig:cmlp\]. The nonlinearity score $\beta_i$ to measure the CMLP nonlinearity degree is defined as in [@wang2012manifold], $$\label{eq:nonlinearity} \beta_i = \frac{1}{F_i^{2}} \sum_{t \in \textmd{C}_{i}} \sum_{s \in \textmd{C}_{i}} r(\mathbf{p}_t,\mathbf{p}_s),$$ where $r(\mathbf{p}_t,\mathbf{p}_s)=d_{G}(\mathbf{p}_t,\mathbf{p}_s) / d_{E}(\mathbf{p}_t,\mathbf{p}_s)$ is the ratio of the geodesic distance $d_{G}$ and the Euclidean distance $d_{E}$ computed by $k$ sequential neighbors graph. We average the ratios between each pair of postures $\mathbf{p}_t$ and $\mathbf{p}_s$ in $\textmd{C}_{i}$ to obtain a robust measurement of nonlinearity degree, and the computation of $\beta_i$ can be efficiently carried out. \    A activity sequence $\mathbf{P} = \{\mathbf{p}_1,\mathbf{p}_2,\cdots,\mathbf{p}_F\}$;\    The nonlinearity degree threshold $\delta$;\    The number of sequential neighbors $k$.\    Local linear model sequences $\textmd{C}_{i}|_{i=1}^{m}$. Initialization:\    $\textmd{C} = \varnothing$, $i = 1$, $\textmd{C}_{i} = \varnothing$, $\beta^{(i)} = 0$, $f=1$;\    Euclidean distance matrix $D_{E} = \varnothing$;\    Geodesic distance matrix $D_{G} = \varnothing$;\    Distance ratio matrix $R = \varnothing$; Update $\textmd{C}_{i} = \textmd{C}_{i} \cup \mathbf{p}_f$, $\mathbf{P} = \mathbf{P} - \{\mathbf{p}_f\}$; Expand $D_E$, $D_G$, $R$ to include $\mathbf{p}_f$; Compute the nonlinearity score $\beta^{(i)}$ with Eq. ; Update $\textmd{C} = \textmd{C} \cup \textmd{C}_{i}$, $i = i + 1$; Reset $\textmd{C}^{(i)}=\varnothing$, $\beta^{(i)} = 0$, $D_E = \varnothing$, $D_G = \varnothing$, $R = \varnothing$; Update $f = f + 1$; Update $\textmd{C} = \textmd{C} \cup \textmd{C}_{i}$; $\textmd{C}$; The improved CMLP not only inherits the ability of MLP to span a maximal linear patch, but also holds the intrinsic structure of successive postures which imply the evolution of corresponding human action snippet. A nonlinearity degree threshold $\delta$ is utilized to control the trade-off between the accuracy of representation and the range of a CMLP. Specifically, a smaller $\delta$ leads to a better accurate representation but a shorter range, and vice versa. Obviously, to make the algorithm applicable, $\delta$ is supposed to be specified to a value larger than $1$ to construct meaningful CMLP sequence. The algorithm of construction of CMLP is summarized in Algorithm \[alg:Framwork\]. The index $f$ and $i$ indicate the current posture and the current CMLP, respectively. After the initialization of the distance matrix and the distance ratio matrix, we include current posture $\mathbf{p}_f$ into current CMLP and compute the nonlinear score of current CMLP. If the nonlinear score is greater than threshold $\delta$, we obtain the first CMLP and reset distance matrix and distance ratio matrix to the initial values. Otherwise, the index of the current posture is assigned to the next posture. These procedures continue until the entire sequence is divided into several CMLPs. Distance Measure between CMLPs ------------------------------ An activity manifold is decomposed into ordered CMLPs, and each CMLP can be regarded as a linear patch spanned by the continuous postures. We represent a linear patch with its center and the first principal component of the covariance matrix, which specify the major posture and main direction of the evolution of an action snippet, respectively. For a CMLP $\textmd{C}_{i}$ denoted by a sequence of postures $\left[\mathbf{p}_1^i,\mathbf{p}_2^{i},\cdots,\mathbf{p}^i_{F_i}\right]$, the major posture is averaged on all postures in this CMLP, $$\mathbf{u}_i=\frac{1}{F_i} \sum_{f=1}^{F_i}{\mathbf{p}^i_f},$$ where $\mathbf{p}^i_f$ is the $f$-th posture of the CMLP $\textmd{C}_{i}$. The sample covariance matrix can be obtained with the formula, $$\mathbf{\Sigma}_i = \frac{1}{F_i} \sum_{f=1}^{F_i}{(\mathbf{p}^i_f-\mathbf{u}_i)(\mathbf{p}^i_f-\mathbf{u}_i)^\top}.$$ By performing eigen-decomposition on the symmetric matrix $\mathbf{\Sigma}_i$, the covariance matrix can be factorized as $$\mathbf{Q}^{-1}\mathbf{\Lambda} \mathbf{Q}=\mathbf{\Sigma}_i,$$ where the diagonal matrix $\mathbf{\Lambda}$ contains the real eigenvalues of $\mathbf{\Sigma}_i$ on its diagonal elements, and $\mathbf{Q}$ is the orthogonal matrix whose columns are the eigenvectors of $\mathbf{\Sigma}_i$ and corresponds to the eigenvalues in $\mathbf{\Lambda}$. The eigenvector that is associated with the largest eigenvalue of $\mathbf{\Sigma}_i$ is denoted by $\mathbf{v}_i$. For distance measure between two subspaces, the commonly used method is principal angles [@bjorck1973numerical], which is defined as the minimal angles between any two vectors of the subspaces. In particular, let $\mathcal{S}_1$ and $\mathcal{S}_2$ be subspaces of $\mathbb{R}^D$ with dimensions $d_1$ and $d_2$ respectively, and $d=\min(d_1,d_2)$. The $j$-th principal angle $0\leq\theta_j\leq\pi/2\ (j\in\{1,\cdots,d\})$ between $\mathcal{S}_1$ and $\mathcal{S}_2$ are defined recursively as follows, $$\begin{aligned} &\cos(\theta_j)=\max\limits_{\mathbf{x}_j\in\mathcal{S}_1}\max\limits_{\mathbf{y}_j\in\mathcal{S}_2} \mathbf{x}_j^\top\mathbf{y}_j, \\ \text{s.t.} \ & \Vert\mathbf{x}_j\Vert=\Vert\mathbf{y}_j\Vert=1, \ \mathbf{x}_j^\top\mathbf{x}_i=\mathbf{y}_j^\top\mathbf{y}_i=0, \\ & \text{where}\ i=1,\cdots,j-1. \end{aligned}$$ The vector pairs $(\mathbf{x}_j,\mathbf{y}_j)$ are called the $j$-th principal vectors. Denote the orthonormal bases of $\mathcal{S}_1$ and $\mathcal{S}_2$ with $\mathbf{S}_1$ and $\mathbf{S}_2$, respectively, the principal angles can be computed straightforward based on the singular value decomposition of $\mathbf{S}_1^\top\mathbf{S}_2$. Concretely, the cosine of the $j$-th principle angle is the $j$-th singular value of $\mathbf{S}_1^\top\mathbf{S}_2$. Various subspace distance definitions have been proposed based on principal angles. For example, max correlation and min correlation are defined using the smallest and largest principal angles, respectively, while [@edelman1999geometry] employed all principal angles in their subspace distance. However, these definitions fail to reflect the difference of subspace positions since principal angles only characterize the difference in direction variation. To derive a better distance measure between CMLPs, we take both the subspace position and direction variation into consideration to measure the main posture distance (MPD) and main direction distance (MDD) between the corresponding action snippets, respectively. The MPD between two CMLPs $\textmd{C}_{i}$ and $\textmd{C}_{j}$ is related to the cosine similarity of $\mathbf{u}_i$ and $\mathbf{u}_j$, $$\begin{aligned} \label{eq:MPD} d_{P}(\textmd{C}_{i},\textmd{C}_{j}) &= (1 - \cos^{2}\alpha)^{1/2} = \sin\alpha, \\ \cos\alpha &=\frac{\mathbf{u}_{i}^\top\mathbf{u}_{j}}{\left(||\mathbf{u}_{i}||\cdot||\mathbf{u}_{j}||\right)}. \end{aligned}$$ In contrast to previous work that assigns weights to each eigenvector, in our case, the MDD is simply defined as the sine distance between first eigenvectors $\mathbf{v}_{i}$ and $\mathbf{v}_{j}$ of two CMLPs $\textmd{C}_{i}$ and $\textmd{C}_{j}$, $$\begin{aligned} \label{eq:MDD} d_{D}({\textmd{C}_{i},\textmd{C}_{j}}) &= (1 - \cos^{2}\beta)^{1/2} = \sin\beta, \\ \cos\beta &=\frac{\mathbf{v}_{i}^\top\mathbf{v}_{j}}{\left(||\mathbf{v}_{i}||\cdot||\mathbf{v}_{j}||\right)}. \end{aligned}$$ The employment of sine distance on both MPD and MDD leads to our distance definition between CMLPs, $$\begin{aligned} d_c(\textmd{C}_{i},\textmd{C}_{j}) &= \bigg(d_{P}^2(\textmd{C}_{i},\textmd{C}_{j}) + d_{D}^2({\textmd{C}_{i},\textmd{C}_{j}})\bigg)^{1/2} \\ &= (\sin^2\alpha + \sin^2\beta)^{1/2} \\ &= (2 - \cos^2\alpha - \cos^2\beta)^{1/2}. \end{aligned}$$ This distance is then used as the basis for the following distance measure between action manifolds. {width="80.00000%"} Activity Manifold-Manifold Distance ----------------------------------- Given the reference and test activity manifold denoted as $\mathcal{R} = \{\textmd{C}_{r1}, \textmd{C}_{r2},\cdots,\textmd{C}_{rm}\}$ and $\mathcal{T} = \{\textmd{C}_{t1}, \textmd{C}_{t2},\cdots,\textmd{C}_{tn}\}$, respectively, where $\textmd{C}_{ri}$ and $\textmd{C}_{tj}$ are CMLPs, we aim to measure Activity Manifold-Manifold Distance (AMMD) based on the distance between CMLPs. An intuitive definition of the manifold to manifold distance is proposed in [@wang2012manifold], $$\begin{aligned} d(\mathcal{R},\mathcal{T})&=\sum_{i=1}^m\sum_{j=1}^n w_{ij}d_c(\textmd{C}_{ri},\textmd{C}_{tj}), \\ \text{s.t.}&\ \sum_{i=1}^m\sum_{j=1}^n w_{ij}=1,\ w_{ij}\geq 0. \end{aligned}$$ This definition integrates all pairwise subspace to subspace distance and is a many-to-many matching problem. The difficulty is how to determine the weight $w_{ij}$ between subspaces $\textmd{C}_{ri}$ and $\textmd{C}_{tj}$. Although earth mover’s distance ($1$-st Wasserstein distance) [@rubner2000earth] can be employed compute $w_{ij}$, its computational complexity is too high. In practice, all weights are set as an equal constant value $\frac{1}{m+n}$. In the scenario of face recognition with image set (FRIS) [@wang2012manifold], the authors believed that the closet subspace pair deserves the most emphasis and defined the manifold to manifold distance as the distance of closest subspace pair from these two manifolds as follows, $$\begin{aligned} d(\mathcal{R},\mathcal{T})=&\min\limits_{\textmd{C}_{ri}\in\mathcal{R}}d(\textmd{C}_{ri},\mathcal{T}) \\ =&\min\limits_{\textmd{C}_{ri}\in\mathcal{R}}\min\limits_{\textmd{C}_{tj}\in\mathcal{T}}d_c(\textmd{C}_{ri},\textmd{C}_{tj}). \end{aligned}$$ It is easy to find out that the weight of the closet pair is set to $1$ and all the other weights are set to $0$ in this case. The best-suited subspaces distance is one of the most appropriate manifold-manifold distances for FRIS problem. However, it cannot be applied to our activity recognition problem since this distance ignores the temporal relationship between actions snippets. To preserve the local order of action snippets in distance definition, we propose to match the pairwise adjacent two CMLPs from the test manifold to the reference manifold and obtain the following distance, $$\begin{aligned} &d(\mathcal{R},\mathcal{T}) \\ =&\sum_{j=1}^{m-1} \min_{i \in [1,n)}\left[d_c(\textmd{C}_{ri},\textmd{C}_{tj})+d_c(\textmd{C}_{r,i+1},\textmd{C}_{t,j+1})\right]. \end{aligned}$$ As illustrated in Figure \[fig:PPDistance\], for each CMLP pair extracted in test manifold $\mathcal{T}$, we find the most similar pair from reference action manifold $\mathcal{R}$, and the sum of all pairwise distances amounts to the AMMD. Afterward, the unknown activity is assigned to the class that has the closest AMMD over all reference action classes, $$\text{label}\quad l=\operatorname*{arg\,min}\limits_{c}\{d(\mathcal{R}_{c}, \mathcal{T})\},$$ where $d(\mathcal{R}_c,\mathcal{T})$ is the distance between the $c$-th class reference action manifold and the test action manifold. Experiments {#sec:experiment} =========== We study the performance of our approach on two popular benchmarks, KARD dataset [@gaglio2015human] and Cornell Activity Dataset (CAD-60) [@sung2012unstructured]. Both of them records 15 joint locations for the participated subjects. In all experiments, the hyperparameters, the number of linked sequential neighbor and the nonlinearity degree threshold, are selected with cross-validation. KARD Dataset ------------ The KARD dataset contains 18 activities collected by Gaglio et al. [@gaglio2015human]. These activities include ten gestures and eight actions, and are grouped into three subsets as listed in Table \[tab:KARDGroup\]. The obtained sequences are collected on 10 different subjects that perform each activity 3 times, thus, there are 540 skeleton sequences in this dataset. According to the previous work [@gaglio2015human], KARD dataset is split under three different setups and two modalities in the experiment. Specifically, the three experiments setups A, B, and C utilize one-third, two-thirds, and half of the samples for training, respectively, and the rest for testing. The activities constituting the dataset are split into the five groups: Gestures, Actions, Activity Set 1, 2, and 3 (three subsets). From subset 1 to 3, the activities become increasingly difficult to recognize due to the increase of similarity between activities. Note that Actions are more complex than Gestures. Methods Accuracy % ------------------------ ------------ [@gaglio2015human] 84.8 [@cippitelli2016human] 95.1 The Proposed Approach **99.3** : Accuracies on the KARD dataset under the “new-person" setting.[]{data-label="tab:KARDResultsNewPerson"} ![Confusion matrix on the KARD dataset under the “new-person” setting.[]{data-label="fig:KARDConfusionMatrix"}](KARD_ConfusionMatrix){width="\linewidth"} All results on this dataset are obtained with the parameter setting: $k = 5$, $\delta = 1.04$. In consideration of the randomness existing in the dataset splitting procedure, we run each experimental setup 10 times and present the mean performance in Table \[tab:KARDResultsAll\]. The proposed approach outperforms all other methods on four out of five subsets under all experimental setups but narrowly lost to the method in [@cippitelli2016human] on the Actions subset. The reason is that the CMLP representation is a linear descriptor, which might fail to capture some nonlinear features of complex activities and is unable to discriminate the subtle difference between the similar activities as a result. In addition, we perform the experiment in the “new-person" scenario, i.e., a leave-one-subject-out setting. The experimental setting is in line with that in [@cippitelli2016human]. Table \[tab:KARDResultsNewPerson\] presents the results of the proposed approach compared with the state-of-the-arts. It can be observed that our approach achieves the best result with an accuracy of $99.3\%$, which exceeds the second best result by $4.2\%$. Figure \[fig:KARDConfusionMatrix\] illustrates the confusion matrix, which shows that the proposed approach classifies all activities correctly with only slight confusion between activities $toss$ $paper$ and $take$ $umbrella$. The reason is that the representations with 3D joint locations are almost the same in these two activities, and the proposed approach is prone to confuse the activities based on the limited information obtained from the linear descriptor. This confusion directly degrades the performance of the experimental setup “Actions" in Table \[tab:KARDResultsAll\]. We believe that it would be sensible to explore the addition of RGB or depth image information in our future work. In summary, this newly proposed approach achieved impressive performance on the above human activity recognition tasks in our current experimental setting. Cornell Activity Dataset ------------------------ Methods Accuracy % ------------------------ ------------ [@wang2014learning] 74.7 [@koppula2013learning] 80.8 [@hu2015jointly] 84.1 [@cippitelli2016human] 93.9 The Proposed Approach **99.6** : Accuracies on the CAD-60 dataset under the “cross-person” setting.[]{data-label="tab:CAD-60ResultsNewPerson"} ![Confusion matrix on the CAD-60 dataset under the “cross-person” setting.[]{data-label="fig:CAD60ConfusionMatrix"}](CAD60_ConfusionMatrix){width="\linewidth"} Cornell Activity Dataset 60 (CAD-60) [@sung2012unstructured] is a human activity dataset comprising of twelve unique activities. Four different human subjects (one is left-handed and others are right-handed, two males and two females) are asked to perform three or four common activities in five different environments, including bathroom, bedroom, kitchen, living room and office. The experimental setting of leave-one-person-out cross-validation is adopted as in [@wang2014learning] that the person in the training would not appear in the testing for each environment. To eliminate the influence from the left-handed subject, if the $y$-coordinate of the right hand is smaller than the left hand, we interchange the $y$-coordinate of left and right hands, ankles and shoulders, to transform skeleton positions of the left-handed persons to those of the right-handed ones. Here, the number of sequential neighbors and nonlinearity degree threshold are set as $k=1$ and $\delta = 1.2$, respectively. The recognition performance is shown in Table \[tab:CAD-60ResultsNewPerson\] by averaging the accuracies on all possible splits (totally 20). The proposed approach achieves an accuracy of $99.6\%$, which outperforms the results of the comparative methods. Figure \[fig:CAD60ConfusionMatrix\] shows the confusion matrix of the performance obtained by our proposed approach. It can be observed that the proposed approach classifies all the actions correctly except minor confusion on two $cooking$ actions: $stirring$ and $chopping$, which is probably caused by the inaccurate human skeletons information. In conclusion, the appealing recognition result demonstrates that our approach can effectively capture the evolution of human activities only based on human 3D joints locations. ![The recognition accuracies of different distance measurements on the KARD dataset with respect to the threshold $\delta$ under the “new person" setting.[]{data-label="fig:KARDAnalysis"}](KARD_Threshold){width="\linewidth"} Parameter Analysis ------------------ There are two key parameters in our approach: the nonlinearity degree threshold $\delta$ and the number of linked sequential neighbors $k$. The nonlinearity degree threshold determines the granularity of continuous linear maximal patches, while the number of linked sequential neighbors quantifies the topology preservation in the computation of geodesic distance. To evaluate the sensitivity of the proposed approach with respect to these two parameters, we conduct experiments with different parameter values under the leave-one-subject-out setting on the KARD dataset. We first fix the number of neighbors $k$ and adjust the nonlinearity degree threshold $\delta$ from $1$ to $2.1$ with step size $0.01$. Figure \[fig:KARDAnalysis\] illustrates the corresponding experimental performance. The relatively small gap between the worst and the best results under each distance measurement validates that the proposed approach is quite robust with respect to the value of $\delta$. Generally speaking, lower $\delta$ could lead to a better performance since a smaller CMLP yields more representative action snippet. Then, we adjust the number of neighbors $k$ from $1$ to $7$ with step size $1$ and fix the nonlinearity degree threshold $\delta = 1.01,1.04,1.09,1.14$, respectively. The obtained result is illustrated in Figure \[fig:KARDAnalysisK\], which shows that with the increase of $k$, the recognition accuracy gets increasingly higher when $\delta$ is small. However, while $\delta$ is large, the recognition accuracy shows a decline with the increase of $k$. Overall, the best recognition accuracy is usually obtained with a small $k$ matched with a large $\theta$ or vice verse. In some sense, two parameters are not totally independent on determining the final performance of our approach, both parameters cooperate with each other to construct the most representative CMLPs. Distance Measure Methods Comparison ----------------------------------- ![The recognition accuracy of the proposed approach on KARD dataset with respect to the number of sequential neighbors $k$ under the “new-person" setting.[]{data-label="fig:KARDAnalysisK"}](KARD_K){width="\linewidth"} The proposed distance measure includes two parts: main posture distance (MPD) and main direction distance (MDD) between action snippets. Intuitively, MDD is more discriminative than MPD since the evolution of the main direction is more important than the position of subspace in the activity recognition problem. However, the MPD is complementary to MDD to some extent. To demonstrate the strength of the proposed distance measure, we compare the different combination of distance measure between the CMLPs and sequence matching algorithm. The results are shown in Figure \[fig:KARDAnalysis\], in which dynamic time warping (DTW) is a template matching algorithm that calculates an optimal match between two given sequences under some certain restrictions. The curve of MPD is almost above the curve of MDD, which confirms our intuition that the major posture features more important than main tendency feature for recognition. As expected, MPD holds major posture representation and MDD keeps the ability to describe the evolution of action snippet. Thus the combination of these two distance measurements performs the best. In general, the combination MPD+MDD+AMMD obtains the best results in most cases. Conclusion {#sec:conclusion} ========== In this paper, we present a novel human activity recognition approach that utilizes a manifold representation of 3D joint locations. Considering that an activity is composed of several compact sub-sequences corresponding to meaningful action snippets, the 3D skeleton sequence is decomposed into ordered continuous maximal linear patches (CMLPs) on the activity manifold. The computation of activity manifold-manifold distance (AMMD) preserves the local order of action snippets and is based on the pairwise distance between CMLPs, which takes into account the major posture and the main direction of action snippets. Experimental results show better performance of our approach in comparison with the state-of-the-art approaches. In practice, there often exists local temporal distortion and periodic patterns in the action sequence. By viewing action snippets as samples from a probability distribution, we attempt to introduce the Wasserstein metric to measure the distance between the action snippets for activity recognition in the future work.

ArXiv

--- abstract: 'We study the core hole-electron correlation in coherently coupled molecules by energy dispersive near edge x-ray absorption fine-structure spectroscopy. In a transient phase, which exists during the transition between two bulk arrangements, 1,4,5,8-naphthalene-tetracarboxylicacid-dianhydride multilayer films exhibit peculiar changes of the line shape and energy position of the x-ray absorption signal at the C *K*-edge with respect to the bulk and gas phase spectra. By a comparison to a theoretical model based on a coupling of transition dipoles, which is established for optical absorption, we demonstrate that the observed spectroscopic differences can be explained by an intermolecular delocalized core hole-electron pair. By applying this model we can furthermore quantify the coherence length of the delocalized core exciton.' author: - 'M. Scholz' - 'F. Holch' - 'C. Sauer' - 'M. Wiessner' - 'A. Schöll' - 'F. Reinert' title: 'Core Hole-Electron Correlation in Coherently Coupled Molecules' --- X-ray absorption measurements at high-resolution beam lines of third-generation synchrotron sources are a powerful tool for the investigation of the electronic and vibronic structure of molecules [@hofer_photoemission_1990; @fohlisch_beyond_1998; @scholl_electron-vibron_2004]. From an analysis of high-resolution data one can gain insight into the complex response of the electronic system and of the molecular frame on the electronic excitation. Upon excitation of a core electron into an unoccupied molecular orbital, a particularly strong reaction of the electronic system in the vicinity of the core excitation occurs. This core hole-electron correlation, which is often referred to as core exciton, has to be included for a correct description of the experimental data. However, the nature of core excitons is not clear in molecular materials, since experimental information, e.g., on the spatial expansion, is usually complicated to extract from x-ray absorption data due to various other interfering effects . Thus, it is not clear whether the excitation is mainly localized on one molecule or on a subgroup of a molecule, or if it extends over several molecules. In this work we demonstrate the importance of the core hole-electron correlation in near edge x-ray absorption fine-structure (NEXAFS) data of organic molecules. On the example of the organic dye 1,4,5,8-naphthalene-tetracarboxylicacid-dianhydride (NTCDA) we show that under certain circumstances the molecules exhibit an extraordinarily strong intermolecular coupling, which leads to a redshift of electronic transitions and a substantial narrowing of the vibronic progressions in the NEXAFS spectra with respect to the gas phase. The respective experimental data can be described by a coupling of the transition dipoles to the neighboring molecules, thus resulting in a significant intermolecular delocalization of the core hole-electron pair. While this effect is well established for optical excitations [@muller_exciton_2011; @wurthner_j-aggregates:_2011], the calculated transition dipoles in core excitations are substantially smaller and thus a more local character of the excitation may generally be expected. The delocalization of the core excition upon x-ray absorption implies not only the transfer of the excited electron but also of the core hole. While for diatomic molecules, such as $N_2$ and $O_2$, experimental evidence for the nonlocal character of the core hole exists [@bjorneholm_doppler_2000; @rolles_isotope-induced_2005; @schoffler_ultrafast_2008], an intermolecular delocalization of a core hole is not easily understood. For the model system NTCDA \[see molecular structure in Fig. \[fig:fig1\](b)\] on a Ag(111) surface, different structural phases exist for multilayer films [@gador_manipulation_1998; @scholl_electron-vibron_2004; @braatz_vibrational_2012]. Growth at substrate temperatures below 100 K results in morphologically smooth films with preferentially flat lying molecules and without long-range order [@gador_manipulation_1998]. Heating the samples to temperatures above 180 K leads to an increasing crystallinity in the films and changes the molecular orientation to a more upright standing configuration. During this transition a transient phase occurs, which exists for only about 10 min and which shows the characteristic changes of the signature of the NEXAFS spectra of particular interest for this work. The fast acquisition of x-ray absorption data is thus crucial and the experiments were therefore performed at the beam line UE52-PGM of BESSY II in Berlin Ref. [@note2]. The beam line was operated in the energy dispersive mode, where the beam is focused on the sample without an exit slit, thus leading to a spatial dispersion of photon energies on the sample in vertical direction [@batchelor_energy-dispersive_2007]. By use of an electron spectrometer with adapted spatial resolution (VG-SCIENTA R4000), NEXAFS spectra can be recorded in a multichanneling mode with a possible acquisition time of less than 1 sec and an energy resolution better than 50 meV at the C *K*-edge.\ ![(a) C K-NEXAFS spectra of a NTCDA multilayer film as prepared at 95 K, recorded with *s* (black line) and *p* polarization (gray line) of the incident beam. (b) Close-up of the $\pi^*$-resonances (denoted as A–C) recorded with *p* polarization, compared to the result of a Hartree-Fock calculation with the GSCF3 code [@kosugi_strategies_1987]. The theoretical energy scale was shifted and scaled to fit the experimental data. The calculated electronic transitions are marked according to the respective carbon 1s sites C1–C4 (see inlay) and final state orbitals $\pi^*$, 2$\pi^*$, 3$\pi^*$. The shaded region indicates the energy window recorded in the energy dispersive mode, which is displayed in (c).[]{data-label="fig:fig1"}](fig1.pdf){width="50.00000%"} Figure \[fig:fig1\](a) shows the C *K*-NEXAFS spectra of a four monolayer (4 ML) thick NTCDA film which was prepared and measured at a substrate temperature of 95 K [@note2]. The absorption signal of the $\pi^*$-resonances between 282 eV and 290 eV shows a characteristic dependence on the polarization of the incident light, which allows determining the average molecular orientation angle with respect to the substrate surface [@stohr_nexafs_1992]. The strong dichroism in Fig. \[fig:fig1\](a) indicates a preferential flat lying orientation of the NTCDA molecules in the film, consistent with previous investigations [@gador_manipulation_1998; @scholl_electron-vibron_2004]. Figure \[fig:fig1\](b) shows the $\pi^*$ resonances of the *p* polarized spectrum on an expanded energy scale. According to Hartree-Fock calculations [@kosugi_strategies_1987] plotted on the bottom of Figure \[fig:fig1\]b), the spectrum can be explained by the superposition of several electronic transitions from the four symmetry-inequivalent carbon 1s sites C1–C4 \[see inset in Fig. \[fig:fig1\](b)\] to the final state orbitals 1$\pi^*$, 2$\pi^*$, 3$\pi^*$. In addition, the electronic transitions show a characteristic coupling to vibrations, leading to the rich fine structure of the three main signals, denoted by A, B, and C in Fig. \[fig:fig1\](b) [@scholl_electron-vibron_2004]. While the spectra in Fig. \[fig:fig1\](a) and 1(b) were recorded by scanning the photon energy in the normal beam line mode with an acquisition time of about 30 min per spectrum, Fig. \[fig:fig1\](c) shows an energy dispersive NEXAFS experiment of the energy regime of feature B \[shaded region in Fig. \[fig:fig1\](b)\]. The integration time was 30 sec and the data are in very good agreement with feature B of Fig. \[fig:fig1\](b). The fine structure is well resolved, and according to the calculations in Fig. \[fig:fig1\](b) the signal in the photon energy range between 285.1 and 286.0 eV corresponds to the two electronic transitions C4 1s-2$\pi^*$ and C3 1s-2$\pi^*$, which are originating from carbon atoms located on the naphthalene core \[see Fig. \[fig:fig1\](b)\]. As demonstrated in Ref. [@scholl_electron-vibron_2004], these electronic transitions lead to the excitation of vibrations of the aromatic core with vibrational energies of around 90 meV. The color plot in Fig. \[fig:fig2\](a) shows the evolution of feature B with time during an energy dispersive NEXAFS experiment. The sample (thickness about 4 ML) was prepared at 95 K and then quickly (in about 6 min) heated to 220 K in order to induce the transition from flat lying to upright standing molecules. On the y axis *t*=0 refers to the time when a constant sample temperature of 220 K was reached. On the bottom of Fig. \[fig:fig2\](a), i.e., directly after preparation at 95 K, the spectra resemble the typical absorption signal for films prepared at low temperature \[see Fig. \[fig:fig1\](c)\]. This is demonstrated by Fig. \[fig:fig2\](b), which displays spectra derived after integration of the intensity map in Fig. \[fig:fig2\](a) over intervals of 3 min at the times indicated by the white horizontal lines. Shortly after a substrate temperature of 220 K is reached, the line shape of feature B is significantly altered. Starting at *t*=2 min the intensity of the absorption peaks at 285.4 eV and 285.7 eV increases, while it decreases in between. In addition, the line width of the most prominent signals decreases, which can be observed best in the spectrum in Fig. \[fig:fig2\](b) extracted after 8 min at 220 K. Moreover, an energy shift of the main spectral components towards lower energy occurs, which is indicated by the white dashed lines in Fig. \[fig:fig2\](a). The energy shift sums up to about $\sim$ 80 meV after 35 min. While the overall intensity increases until about *t*=12 min, the absorption signal is constantly decreasing after the sample was kept at 220 K for more than 12 min.\ ![(a): 2D color map of the energy dispersive NEXAFS spectra of feature *B* recorded with *s* polarized light plotted against time. Yellow refers to high, blue to low intensity. The 4 ML thick film was prepared at 95 K and heated to 220 K within 1 min. *t*=0 refers to the time when a constant sample temperature of 220 K was reached. The NEXAFS data was collected continuously during annealing for about 45 min. The vertical white dashed lines are guidelines for the eye to illustrate the energy shift of the main components. (b) NEXAFS spectra of feature B derived by integration of the intensity map of (a) over 3 min at the times indicated by the horizontal white dashed lines (black dots: original data; solid line: 3-point average).[]{data-label="fig:fig2"}](fig2.pdf){width="50.00000%"} ![(a) Illustration of the coupling of NTCDA molecules in the transient phase. The transition dipole moments are oriented perpendicular to the molecular plane. (b) Schematic energy level diagram of adjacent NTCDA molecules after resonant core excitation. See text for details. (c) Result of a coherent exciton scattering calculation: The adiabatic component of the electronic excitation (0-0) and the corresponding vibronic progression (0–1, 0–2, 0–3) are plotted for isolated (gray) and coupled (black) molecules. (d) NEXAFS spectra of NTCDA feature B in the gas phase [@holch_new_2011], in the disordered condensed phase [@scholl_electron-vibron_2004], in the ordered condensed phase after annealing [@scholl_electron-vibron_2004], and in the transient phase (from bottom). The spectra were normalized arbitrarily for better comparison of the line shape.[]{data-label="fig:fig3"}](fig3.pdf){width="50.00000%"} This variation of general intensity can be explained by the change of orientation angle of the NTCDA molecules from flat lying to a more upright configuration, which occurs after the sample temperature is increased to 220 K [@gador_manipulation_1998]. This leads to an immediate intensity increase in the C *K*-NEAXFS data recorded with *s* polarization [@stohr_nexafs_1992]. However, this intensity increase is counteracted by a change of film morphology, since the increased sample temperature also leads to a roughening of the originally closed film and to formation of 3D islands. Because of the limited probing depth of about 1 nm [@graber_experimental_2011] of the Auger electrons mainly contributing to the NEXAFS signal in electron yield mode, this second process leads to an overall reduction of the absorption signal. Moreover, the morphology change includes material transport and occurs on a slower time scale than the change of orientation angle, thus accounting for the observed delayed intensity decrease [@Scholz_2012]. We will in the following concentrate on the particular changes of the spectral signature of the NTCDA C K-NEXAFS which occur during the transition. Figure \[fig:fig3\](d) compares the NEXAFS spectra of feature B for four different NTCDA samples. On the bottom gas phase data are displayed, which is representative of the isolated molecule [@holch_new_2011]. Film growth at low sample temperature (*T*=95 K) results in the disordered phase with preferentially flat lying molecules (second spectrum from bottom) [@scholl_electron-vibron_2004], while annealing to temperatures above 180 K results in the ordered phase with more upright molecular orientation (third spectrum from bottom) [@scholl_electron-vibron_2004]. The data of the transient phase during the transition are plotted on top. The line shape in the disordered phase resembles the gas phase, while the entire spectrum is redshifted by about 100 meV upon condensation. The ordering of the NTCDA molecules upon annealing leads to a slight decrease of the line width and to a more pronounced shoulder on the low energy side of feature B. This trend is followed if the transient phase is examined. The most intense peak at 285.7 eV is now very sharp and the low energy shoulder has developed into a well distinguishable peak at 285.4 eV. Moreover, additional peaks can be resolved at 285.5 and 285.6 eV as well as at 285.8 eV. The energetic separation of these signals from the prominent peaks at 285.4 and 285.7 eV corresponds well to the vibronic energy of the NTCDA modes coupled to the C4 1s-2$\pi^*$ and C3 1s-2$\pi^*$- transitions, respectively [@scholl_electron-vibron_2004]. We thus straightforwardly interpret the spectrum of feature B in the transient phase as due to the prominent adiabatic transitions C4 1s-2$\pi^*$ at 285.4 eV and C3 1s-2$\pi^*$ at 285.7 eV with adjacent vibronic states. The main difference from the other condensed and gas phase spectra lies in a reduced line width of the involved signals and in a change of the vibronic envelope, which obviously shifts intensity towards the adiabatic transition in the transient phase. An explanation for this observation can be given on the basis of a coupling of the transition dipole moments of adjacent molecules. The coupling strength strongly depends on the arrangement of the molecules and on the symmetry of the respective transition moments. In the dipole approximation the coupling strength $V_{nm}$ between molecules n and m can be described by $$V_{nm}=\left|\vec{\mu}_n\right|\left|\vec{\mu}_m\right|\cdot\left(\frac{\vec{n}_{\vec{\mu}_n}\vec{n}_{\vec{\mu}_m}}{R^3_{nm}}-\frac{3(\vec{n}_{\vec{\mu}_n}\vec{r}_{nm})(\vec{n}_{\vec{\mu}_m}\vec{r} _{nm})}{R^5_{nm}}\right),$$ were $\vec{\mu}_n$ and $\vec{\mu}_m$ are the transition dipole moments on molecules n and m, respectively, which are separated by $\vec{r}$. For a C 1s-2$\pi^*$ transition with a dipole moment perpendicular to the molecular plane, the coupling is thus largest for a head-to-tail arrangement of the dipoles (commonly referred to as J aggregates); i.e., a configuration as sketched in Fig. \[fig:fig3\](a).\ The proposed dipole-dipole coupling mechanism is illustrated in Fig. \[fig:fig3\](b). The C 1s-2$\pi^*$ excitation of molecule A couples to the same transition on the neighboring molecule B.\ The changes of the vibrational envelope observed in our NEXAFS data can be understood in the framework of the coherent exciton scattering (CES) theory [@briggs_sum_1970; @eisfeld_j-band_2002; @eisfeld_j-_2006; @walczak_exchange_2008]. This one-particle Green’s function approach obeys Dyson’s equation and describes the excitation of coherently delocalized quasiparticles. In other words, this means that the hole-electron pair propagates to the adjacent molecule fast compared to the molecular vibration frequencies. Therefore, the originally excited molecule remains preferentially in its vibrational ground state. As a result, intensity is shifted towards the adiabatic transition in the vibrational envelope observed in experiment. This is illustrated by Fig. \[fig:fig3\](c), which shows the result of a CES calculation of the vibronic progression of a single molecule and of coupled molecules in a configuration as sketched in Fig. \[fig:fig3\](a) [@note]. The intensity ratio of the adiabatic transition (0-0) to the higher vibrational states (0–1, 0–2, and 0–3) changes strongly, and in the case of sufficiently strong coupling the oscillator strength accumulates in the 0-0 transition. Moreover, a marked energy shift of the entire vibronic progression towards lower energy can be observed in the case of the coupled molecules, which also matches our experimental observation qualitatively \[see Fig. \[fig:fig2\](a)\]. In the transient phase we also obverse a reduced line width of the C 1s-2$\pi^*$ transitions, as evidently seen in Fig. \[fig:fig3\](d). For a quantitative analysis the FWHM of the gas phase and of the transient phase spectra was estimated to 211 ($\pm$10) meV and 74 ($\pm$10) meV respectively, for the most prominent signals in the data of Fig. \[fig:fig3\](d). According to Knapp et al. [@knapp_lineshapes_1984] the decrease of the FWHM $\Delta$ of the absorption signal with the number of coherently coupled molecules $N$ can be described by the equation: $$\Delta\propto\frac{1}{\sqrt{N}}.$$ With the FWHM derived from our NEXAFS data of isolated and coherently coupled molecules, i.e. $\Delta_{isol.}$ and $\Delta_{coup.}$, respectively, a number of about $N$=8$\pm$3 can be estimated in the transient phase.\ In conclusion, we have reported on an x-ray absorption experiment which demonstrates the importance of a correct consideration of the core hole-electron correlation for condensates of organic molecules. During the transition from a disordered to an ordered film arrangement, a transient phase occurs for NTCDA/Ag(111) which is characterized by a distinctly different spectral signature. The spectra of this transient phase exhibit a reduced line width, a redshift, as well as a substantial change of the vibronic profile. All these experimental observations can be explained by the coupling of the transition dipole moments of adjacent molecules, a model, which is well established for optical spectroscopies. This result has two immediate implications: Firstly, it shows that the excitation delocalizes over several molecules (a number of about 10 can be determined from our data). This comprises the existence of a delocalized core hole, which is involved in the excitation. This implication is not understood and needs further consideration. Note that other models have been considered but could not explain our experimental observation. Secondly, the very accurate explanation of our experimental data by this model based on the coupling of transition dipoles strongly suggests that the oscillator strength of the transition dipole moments involved in the core-to-valence excitations is substantially larger \[a value of about 0.37 a.u. can be estimated from the experimental data and eq. (1)\] than derived from calculations on isolated molecules (about 0.04 a.u. for the respective NTCDA transitions [@scholl_electron-vibron_2004]). This discrepancy may be explained by the additional contribution of the wave function overlap between neighboring molecules [@fuckel_theoretical_2008; @curutchet_does_2008; @yamagata_nature_2011; @yamagata_designing_2012]. For close-packed $\pi$-conjugated molecules in a face-to-face arrangement, neighboring dyes can have a considerable overlap of the frontier orbital wave functions [@kera_very_2002; @koch_evidence_2006; @ueno_electron_2008; @machida_highest-occupied-molecular-orbital_2010]. This can lead to increased intermolecular charge transfer transition dipole moments [@bredas_charge-transfer_2004; @bredas_molecular_2009]. This so-called short-range component of the coupling becomes significant, e.g., for naphthalene dimers already at a separation of about 6 Å [@scholes_electronic_1994] and can thus be expected to be substantial in the present case, where the intermolecular distance between the NTCDA molecules is only around 3 Å. Moreover, the relatively small coupling values for core-to-valence excitations predicted by theory are usually derived from a point-dipole approach, which may find its restrictions in the present case of a very dense molecular packing, where the intermolecular distance of around 3 Å is much smaller than the size of the NTCDA molecules ($\sim$9 Å along the long axis). Further efforts to develop a comprehensive theoretical description of the core excitation in molecular condensates are thus highly desirable. We thank Patrick Hoffmann (BESSY), David Batchelor (KIT) and the BESSY staff for assistance during beam time. We acknowledge stimulating discussions with Alexander Eisfeld and Peter Jakob. This work was financially supported by the Deutsche Forschungsgemeinschaft (Grants No. FOR1162, No. GRK1221, and No. RE1469/9-1) and the Bundesministerium für Bildung und Forschung (Grant No. 03SF0356B).\ [36]{}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 [****, ()](http://www.ncbi.nlm.nih.gov/pubmed/10041882),  [****, ()](\doibase 10.1103/PhysRevLett.81.1730) [****, ()](\doibase 10.1103/PhysRevLett.93.146406) [****,  ()](\doibase 10.1103/PhysRevLett.54.1960) [****, ()](\doibase 10.1103/PhysRevLett.74.614) [****,  ()](\doibase 10.1016/j.cplett.2006.08.110) [****,  ()](\doibase 10.1103/PhysRevB.83.241203) [****,  ()](\doibase 10.1002/anie.201002307) [****,  ()](\doibase 10.1103/PhysRevLett.84.2826) [****,  ()](\doibase 10.1038/nature04040) [****,  ()](\doibase 10.1126/science.1154989) [****,  ()](\doibase 10.1209/epl/i1998-00135-4) [****, ()](\doibase doi:10.1063/1.3699030) @noop [ ]{} [****,  ()](\doibase 10.1016/j.nima.2007.02.108) @noop [[[en]{}**]{}]{} (, ) [****,  ()](\doibase 10.1007/BF00528137) [****,  ()](\doibase 10.1016/j.susc.2011.01.033) @noop [ ]{} [****,  ()](\doibase 10.1016/j.elspec.2011.05.006) [****,  ()](\doibase 10.1088/0022-3700/3/12/011) [****,  ()](\doibase 10.1016/S0301-0104(02)00594-3) [****,  ()](\doibase 10.1016/j.chemphys.2005.11.015) [****,  ()](\doibase doi:10.1063/1.2823730) @noop [ ]{} [****,  ()](\doibase 10.1016/S0301-0104(84)85174-5) [****,  ()](\doibase 10.1063/1.2829531) [****, ()](\doibase 10.1021/jp7106507) [****,  ()](\doibase doi:10.1063/1.3590871) [****,  ()](\doibase 10.1021/jp309407r) [****,  ()](\doibase 10.1016/S0009-2614(02)01302-7) [****,  ()](\doibase 10.1103/PhysRevLett.96.156803) [****,  ()](\doibase 10.1016/j.progsurf.2008.10.002) [****,  ()](\doibase 10.1103/PhysRevLett.104.156401) [****, ()](\doibase 10.1021/cr040084k) [****,  ()](\doibase 10.1021/ar900099h) [****,  ()](\doibase 10.1021/j100068a017)

ArXiv

--- abstract: 'We present Monte Carlo simulations of the dynamical evolution of the Oort cloud over the age of the Solar System, using an initial sample of one million test comets without any cloning. Our model includes perturbations due to the Galactic tide (radial and vertical) and passing stars. We present the first detailed analysis of the injection mechanism into observable orbits by comparing the complete model with separate models for tidal and stellar perturbations alone. We find that a fundamental role for injecting comets from the region outside the loss cone (perihelion distance $q > 15$ AU) into observable orbits ($q < 5$ AU) is played by stellar perturbations. These act in synergy with the tide such that the total injection rate is significantly larger than the sum of the two separate rates. This synergy is as important during comet showers as during quiescent periods and concerns comets with both small and large semi-major axes. We propose different dynamical mechanisms to explain the synergies in the inner and outer parts of the Oort Cloud. We find that the filling of the observable part of the loss cone under normal conditions in the present-day Solar System rises from $<1$% for $a < 20\,000$ AU to about 100% for $a \gta 100\,000$ AU.' author: - Hans Rickman - Marc Fouchard - Christiane Froeschlé - 'Giovanni B. Valsecchi' title: 'Injection of Oort Cloud Comets: The Fundamental Role of Stellar Perturbations' --- Introduction ============ When analyzing the distribution of original inverse semi-major axes of long-period comets, concluded that the near-parabolic spike of this distribution reveals a distant reservoir of comets (the ‘Oort Cloud’). His favoured mechanism of injection of comets from this reservoir into observable orbits (, with small perihelion distances) was the passage of stars in the vicinity of the reservoir, whereby the long-term reshuffling of angular momenta would ensure a continuous infeed into the innermost part of the Solar System. Until the 1980’s stellar perturbation was the only mechanism considered, when issues concerning the injection of comets from the Oort Cloud were discussed (, ; ; ; ; ). However, by that time it became clear that the tidal action of the Galaxy as a whole must also have an important influence – especially the part corresponding to the $z$-dependent disk potential [@HEI.TRE:86]. This was verified by noting that the Galactic latitudes of perihelia of new Oort Cloud comets have a double-peaked distribution that is characteristic of the disk tide [@DEL:87]. An important paper by treated the formation of the Oort Cloud and showed that the characteristic time scale for changing the perihelion distances, independent of the semi-major axis, is shorter for the Galactic disk tide than for the stellar perturbations. This has been verified by later analytic work, , by , and further numerical simulations of Oort Cloud evolution (, ) have given support to the dominance of Galactic tides for comet injection. Consequently, stellar perturbations have come to be practically neglected as a source of comet injection – except when discussing “comet showers” [@HIL:81] arising from rare stellar passages through the dense, inner parts of the Oort Cloud. The importance of stellar perturbations for randomizing the orbital distribution of the Oort Cloud and thus keeping the relevant infeed trajectories of the disk tide populated over long time scales has been realized (see and references therein), but the actual injection is often seen as due only to the tide. Hence it should be limited to semi-major axes large enough for the tidal perturbation to bring the cometary perihelion at once from outside the “Jupiter-Saturn barrier” (, perihelion distance $q \gta 15$ AU) into the observable region ($q < 5$ AU). The result is that one expects new comets to have $a_{ori} \gta 3\times 10^4$ AU [@BAI.STA:88; @LEVetal:01; @FER:05]. On the other hand, some recent papers indicate that this picture may have problems. The fractional population of the observable region – if fed only by Galactic tides – is small enough, and the orbital periods long enough, that the estimated total population of the Oort Cloud may be uncomfortably large [@CHA.MOR:07]. And in addition, when non-gravitational effects are included into orbit determinations for new Oort Cloud comets [@KRO:06], the resulting original orbits tend to be of shorter periods, having smaller semi-major axes – often much smaller than $3\times 10^4$ AU. Meanwhile, we have developed fast and accurate methods to treat both the Galactic tides [@BREetal:07; @FOUetal:07b] and stellar perturbations [@RICetal:05] in Monte Carlo simulations of Oort Cloud dynamics. This has allowed us to perform calculations, to be presented here, where the cloud is represented by as many as $10^6$ sample comets and integrated over a time exceeding the age of the Solar System. This amounts to $5\times 10^{15}$ comet-years of individual evolutions (or only slightly smaller due to the loss of comets during the simulation), which is much more than in all previous long-term Oort cloud simulations – , $3\times 10^{13}$ comet-years for , $4\times 10^{14}$ comet-years for , and $\sim 4\times 10^{13}$ comet-years for who used cloning. simulated $\simeq 7\times 10^{15}$ comet-years but only thanks to extensive cloning during the course of the simulation. In fact her long-term simulations (4.5 Gyr) concerned only $\simeq10^4$ “tokens”, , comets actually followed, while these were meant to represent $\simeq150$ times as many comets by means of cloning. Our work is the first to study the mechanism of injection of comets from the Oort Cloud over the age of the Solar System by simulating and comparing different dynamical models. The reason why models involving both the Galactic tide and stellar perturbations gave a much higher flux of injected comets than those involving only stars [@HEIetal:87; @HEI:90] was never clarified, since comparisons with models involving only the tide were not made. In the present paper we concentrate on a comprehensive comparison of combined and separate models, thus describing and analyzing for the first time the synergy effect of Galactic tide and stars. We also take special care to define correctly the encounter velocities in our sample of passing stars, thereby arriving at somewhat larger values than those used previously. Finally, we study the relative filling of the observable part of the loss cone and the distribution of inverse semi-major axes of the injected comets. These studies are, however, only preliminary, since our current simulations do not include planetary perturbations, and thus we cannot account for those comets that arrive into the observable region after having “diffused” across the Jupiter-Saturn barrier in several revolutions. Our calculations are presented in Sect. 2, and in Sects. 3–5 we describe our results in terms of the distribution of injection times into the inner planetary system, the flux of new, observable comets as a function of time, and the distributions of inverse semi-major axis and Galactic latitude of perihelion as well as loss cone filling at representative epochs. In Sect. 6 we discuss the results and summarize our conclusions. Calculations ============ As a simplifying assumption, we consider the Oort Cloud to have been formed instantaneously at a given epoch, and that its orbital distribution was isotropic to begin with. Thus the initial conditions are chosen with flat distributions of $\cos i_o$, $\omega_o$, $\Omega_o$ and $M_o$ (we use common notations for the orbital elements, and the angles may be defined with respect to an arbitrary frame of reference). We consider a thermalized initial state of the cloud, where the semi-major axes ($a_o$) are chosen in the interval $3\times 10^3 < a_o < 1 \times 10^5$ AU with a probability density $\propto a_o^{-1.5}$ [@DUNetal:87]. The initial eccentricities ($e_o$) are chosen with a density function $\propto e_o$ in such a way that the perihelia are outside the planetary system ($q>32$ AU). We thus initialise $1\times 10^6$ comets. The Galactic parameters used for calculating the tidal effects are the same as described in earlier papers [@FOUetal:07b]. The most important one for comparison with other investigations is the mid-plane disk density, which we take as 0.1 $M_\odot$pc$^{-3}$ [@HOL.FLY:00]. This is in agreement with , while used 0.18 $M_\odot$pc$^{-3}$ [@BAH:84]. The simulation runs with a predefined set of $197\,906$ stellar encounters, occurring at random times during a lapse of $t_{max}=5\times 10^9$ yr, with random solar impact parameters up to $d_{max}=4\times 10^5$ AU, and with random stellar masses and velocities. Our procedure for creating each of these encounters is as follows. Let $\xi$ denote a stochastic, random number with a uniform probability distribution on the interval $[0,1]$. The solar impact parameter is chosen as $d=\xi_d^{1/2}\times d_{max}$, and the time of the encounter (specifically, the time of the star’s perihelion passage) is $t=\xi_t\times t_{max}$. The direction of stellar motion with respect to the Sun is defined in terms of Galactic latitude and longitude ($b,\ell$) such that $\sin b=2\xi_b-1$ and $\ell=\xi_\ell\times2\pi$, , it has an isotropic distribution. The point at which the initial stellar velocity cuts the orthogonal impact plane is situated on a circle of radius $d$ around the Sun, and its location is defined by an azimuthal angle ($a$) such that $a=\xi_a\times2\pi$. Next we choose the type of the star. We use 13 categories as in with parameters listed in Table \[tab:enc\]. To each category we associate one value of the stellar mass. These masses are generally taken as those of the archetypal spectral classes along the main sequence according to . However, in contrast to our earlier investigations, we replace the archetypal mass of 18 $M_\odot$ for B0 stars by a weighted average of 9 $M_\odot$, considering that the less massive, later types (B2, B5) are much more common than the earlier ones. The relative encounter frequencies $f_i$ of Table \[tab:enc\] are taken from , where they were derived from the respective products of number density and mean velocity, $n_i\langle v_i\rangle$. A random number $\xi_i$ is used to pick a stellar category $i$ with the probability $f_i/\sum f_i$. ------ ------------- ------------ ----------- --------------- -------------------- ------------ Type Mass Enc. freq. $v_\odot$ $\sigma_\ast$ $\langle V\rangle$ $\sigma_V$   ($M_\odot$)   (km/s) (km/s) (km/s) (km/s) B0 9 0.005 18.6 14.7 24.6 6.7 A0 3.2 0.03 17.1 19.7 27.5 9.3 A5 2.1 0.04 13.7 23.7 29.3 10.4 F0 1.7 0.15 17.1 29.1 36.5 12.6 F5 1.3 0.08 17.1 36.2 43.6 15.6 G0 1.1 0.22 26.4 37.4 49.8 17.1 G5 0.93 0.35 23.9 39.2 49.6 17.9 K0 0.78 0.34 19.8 34.1 42.6 15.0 K5 0.69 0.85 25.0 43.4 54.3 19.2 M0 0.47 1.29 17.3 42.7 50.0 18.0 M5 0.21 6.39 23.3 41.8 51.8 18.3 wd 0.9 0.72 38.3 63.4 80.2 28.2 gi 4 0.06 21.0 41.0 49.7 17.5 ------ ------------- ------------ ----------- --------------- -------------------- ------------ : Stellar parameters. The types are mostly MK types for main sequence stars; ‘wd’ indicates white dwarfs, and ‘gi’ indicates giant stars. The encounter frequencies are given in number per Myr within 1 pc. The following two columns list the solar apex velocity with respect to the corresponding type, and the spherical Maxwellian velocity dispersion. The last two columns give the mean heliocentric encounter velocity and its standard deviation according to our results. []{data-label="tab:enc"} Finally, we choose the speed of the stellar motion in the following way. The velocity dispersions ($\sigma_{\ast i}$) listed in Table \[tab:enc\] are taken from , and they correspond to the semi-axes of the velocity ellipsoids ($\sigma_{ui}$, $\sigma_{vi}$, $\sigma_{wi}$) listed by using: $\sigma_{\ast i}^2=\sigma_{ui}^2+\sigma_{vi}^2+\sigma_{wi}^2$. For the peculiar velocity ($v_\ast$) of a star with respect to its LSR, we use a spherical Maxwellian by taking $\eta_u$, $\eta_v$ and $\eta_w$ as three random numbers, each with a Gaussian probability distribution with expectance 0 and variance 1, and computing $v_\ast=\sigma_{\ast i}\left\{(\eta_u^2+\eta_v^2+\eta_w^2)/3\right\}^{1/2}$. The star’s heliocentric velocity is found by combining the vector ${\mathbf v}_\ast$ with the Sun’s peculiar velocity with respect to the star’s LSR (“apex velocity”) ${\mathbf v}_\odot$, whose absolute value is listed in Table \[tab:enc\] for each stellar category. We assume a random relative orientation of the two vectors and thus compute: $$\label{eq:velocity} V=\bigl\{v_{\odot i}^2+v_\ast^2-2v_{\odot i}v_\ast\cdot C\bigr\}^{1/2}$$ where $C=\cos\theta$ is taken as $C=2\xi_C-1$, and $\theta$ is the angle between the two vectors. Within each stellar category we have to account for the fact that the contribution to the encounter flux is proportional to $V$. Thus we define a constant, large velocity $V_{0i} = v_{\odot i} + 3\sigma_{\ast i}$ for each category, such that $V$ is always smaller than $V_{0i}$, and we take a new random number $\xi_V$ and keep the value just found for $V$, if $\xi_V<V/V_{0i}$. Otherwise we repeat the computation of $V$ until the $\xi_V$ condition is fulfilled. This procedure was not followed in our previous investigations, leading to underestimates of the average stellar velocities. Further underestimates were caused by programming errors, and we caution the reader that the stellar velocities in @RICetal:04  were systematically too small. This is clearly seen by comparing Fig. 1 of with the data in Table \[tab:enc\], which yield a mean velocity of 53 km/s with a dispersion of $\simeq20$ km/s. A few comments on the mean stellar encounter velocity are in order. The average peculiar velocity of stars in the solar neighbourhood is $\simeq40$ km/s. This value was given by , and combining it in quadrature with a typical solar apex velocity of 23 km/s for the most common stellar categories (Table \[tab:enc\]), one gets a mean heliocentric velocity of $\simeq46$ km/s. were the first to introduce the flux-weighting into the selection of random velocities, but they neglected the solar apex velocity. In fact, their flux-weighting was somewhat different from ours, because they considered only one of the three velocity components, namely, the radial heliocentric velocity. But the encounter flux is sensitive to the velocity with respect to the impact plane, , the full speed of the star, instead of just the radial component. We have found that this difference has only a small effect on the resulting mean velocity, but we mention it for the sake of completeness. In both cases, we find that the weighting raises the mean velocity by $6-7$ km/s. This explains our mean velocity of 53 km/s as resulting from including both the solar motion and the flux-weighting procedure. Finally, let us compare with the mean encounter velocity of $\simeq46$ km/s in . This resulted from a list of 92 stellar encounters within 5 pc and 1 Myr of the present, compiled with the aid of Hipparcos data, but the authors showed that there was a serious bias against faint absolute magnitudes in this sample, affecting all stars with $M_V>4$. Thus, the stars with the highest velocities were essentially lacking, and the resulting mean velocity is that of the inherently brighter, slower moving stars. Comparing with other investigators, we note that both , and based their stellar encounter frequencies on the analysis by , who ignored the solar motion – thereby underestimating the relative frequency of encounters with massive stars that have small velocity dispersions – and neglected the contribution of the massive giants. Our encounter sample contains as much as 3.5% of massive stars, , the B0, A0, A5, F0, F5 and ‘gi’ categories in Table \[tab:enc\] with an average mass of 2.3 $M_\odot$, while counting the stars in the absolute magnitude range that corresponds to this average mass in , one arrives at $<1$% of the total encounter frequency. Since the massive stars have an average $M/V$ ratio $\sim 10$ times larger than the red dwarf stars that dominate the sample, each such star will affect $\sim 100$ times as many comets. Hence one easily realizes that in our case a large fraction of the total stellar effect comes from the massive star category that is downplayed by the other investigators. This, to some extent, compensates for two other effects that make the stellar perturbations less efficient in our simulation. One is the larger encounter velocities, as already described, and the other is the total encounter frequency within 1 pc, which in our case is 10.5 per Myr, while for the others it is 13.1 per Myr. Our calculations of the heliocentric impulse imparted to the comet (at time $t$) are done using the Sequential Impulse Approximation, which guarantees a good accuracy at a low cost of computing time [@RICetal:05]. During the simulation we keep track of all the perihelion passages with their $q$ values. At each perihelion a decision is taken about which method to use for the Galactic tide perturbation during the coming orbital period. The fastest method is a mapping [@BREetal:07] that analytically computes the orbital elements at the subsequent perihelion, but this is used only for elliptic orbits with semi-major axis less than a critical value that depends on the eccentricity [@FOUetal:07b]. Otherwise we use numerical integration with a symplectic integrator for KS-regularized equations of motion [@LAS.ROB:01] in case $1/a>10^{-5}$ AU$^{-1}$ and the 15$^{th}$ order RADAU integrator [@EVE:85] for $1/a<10^{-5}$ AU$^{-1}$. During the orbital period in question, normally, several stellar passages occur. On those occasions the osculating cometary orbit is subject to an instantaneous impulse. In the numerical integration regime for the Galactic tide, one always comes back to perihelion. But in the mapping regime, when the starting orbital period has elapsed, the comet may not be at perihelion because of the intervening stellar perturbations, and we then resort to numerical integration until the next perihelion passage takes place. The simulation proceeds for a maximum of 5 Gyr, unless an end state is reached. There are two such end states: either the comet reaches perihelion with $q<q_c=15$ AU (it is lost due to planetary perturbations), or the comet reaches $r=4\times 10^5$ AU in outbound motion (it escapes directly into interstellar space). What we have just described is the full simulation of the “combined” model including both the Galactic tide and stellar perturbations. In addition, we have run two simulations that include only one or the other of the two dynamical effects. Injection time ============== We will first consider the time needed to shift any cometary starting perihelion distance $q_o\,>\,32$ AU into a perihelion distance $q\,<\,15$ AU. This is the time $t_{\rm inj}$ required to inject a comet into the target zone, and we call it the injection time. We have thus scrutinized all three simulations, and for each injected comet in every simulation we noted its value of $t_{\rm inj}$. Let us now compare the statistics of injection times between all three dynamical models. The range of initial semi-major axes $3\,000\,<\,a_o\,<\, 100\,000$ AU is divided, according to $\log a_o$, into ten equal intervals. For each interval the following statistical parameters concerning the injection time are computed: its median value, its lower and upper quartiles (surpassed by 75%, respectively 25%, of the values), its lower and upper deciles (surpassed by 90%, respectively 10%, of the values), and finally its lower and upper percentiles (surpassed by 99%, respectively 1%, of the values). Figure \[fig:dqt\] presents the comparison of $t_{\rm inj}$ statistics by means of two plots. The left one (Fig. \[fig:dqt\]a) compares the model with only the Galactic tides to the one with only stellar perturbations, while in the right one (Fig. \[fig:dqt\]b) the tides-only model is compared to the combined model. In each case we plot the statistical quantities versus $a_o$. The tides-only model is represented by filled squares, and for the other models we use half-filled circles. At any particular value of $a_o$, the symbols for each model are joined by vertical bars. A slight horizontal shift between the symbols has been introduced for easy distinction of the models, but the real $a_o$ intervals are identical. The median values have been joined by curves (dotted for the tides-only model, and solid for the other ones). The grey dots show individual injections for the stars-only model (Fig. \[fig:dqt\]a) and the combined model (Fig. \[fig:dqt\]b). (10,12) (0,6)[![Injection time versus initial semi-major axis. The semi-major axis range is split into ten equal intervals of $\log a$. For each interval we plot different statistical parameters characterizing the distribution of injection times, as explained in the text. The dotted line labelled $T_g$ corresponds to the tidal torquing time shown in Fig. 2 of Duncan [*et al.*]{} (1987), $T_s$ to the corresponding stellar torquing time, and $T_r$ to the period of ($q,i_G$) oscillation imposed by the vertical tide. The left panel (Fig. \[fig:dqt\]a) compares the model with only tides to the one with only stars, and the grey dots are individual stellar injections. In the right panel (Fig. \[fig:dqt\]b) the stars-only model is replaced by the combined model.[]{data-label="fig:dqt"}](dqt-G-C-rev1.eps "fig:"){height="6cm"}]{} (0,0)[![Injection time versus initial semi-major axis. The semi-major axis range is split into ten equal intervals of $\log a$. For each interval we plot different statistical parameters characterizing the distribution of injection times, as explained in the text. The dotted line labelled $T_g$ corresponds to the tidal torquing time shown in Fig. 2 of Duncan [*et al.*]{} (1987), $T_s$ to the corresponding stellar torquing time, and $T_r$ to the period of ($q,i_G$) oscillation imposed by the vertical tide. The left panel (Fig. \[fig:dqt\]a) compares the model with only tides to the one with only stars, and the grey dots are individual stellar injections. In the right panel (Fig. \[fig:dqt\]b) the stars-only model is replaced by the combined model.[]{data-label="fig:dqt"}](dqt-rev1.eps "fig:"){height="6cm"}]{} For comparison with , who plotted a similar diagram (their Fig. 2), we include three lines. The one labelled $T_g$ shows their “tidal torquing time” and the one labelled $T_s$ shows the corresponding “stellar torquing time”, both as functions of $a$. These are meant to represent the typical time required to decrease the perihelion distance from 25 AU to 15 AU in the two cases. The third line labelled $T_r$ refers to the period of ($q,i_G$) oscillation due to the tidal component normal to the Galactic plane [@DUNetal:87]. By inspecting Fig. \[fig:dqt\], we can make the following observations. First, compare the tides-only median curve with the $T_r$ line. The two agree fairly well in the range from $a\simeq6000$ to $25\,000$ AU. This is natural, because $T_r$ is twice the average time it takes for the vertical tide to bring any Oort Cloud comet into the target zone, as long as it is on a relevant trajectory with $q_{\rm min}<15$ AU. For $a\lta6000$ AU the median flattens out at about 2.5 Gyr, and this is due to the limit of our simulated interval at 5 Gyr. Had we let the simulation run for a much longer time, we would have seen the median curve follow the $T_r$ line to even smaller $a$ values. For $a\gta25\,000$ AU we see how the median curve turns upwards, while $T_r$ continues to decrease. This can be explained as a result of a quick stripping of comets from all trajectories with $q_{\rm min}<15$ AU, after which these have to be repopulated through the action of the non-integrable part of the tides. Since this works on a much longer time scale, it is obvious that the median of $t_{\rm inj}$ has to increase. Already at this point we see evidence that the mean injection time – even in the tides-only model – does not follow the prediction of the $T_r(a)$ dependence at all semi-major axes. Studying the Oort Cloud over a long enough time allows other parts of the Galactic tide than the simple, vertical component to take control of comet injections, at least in the outer parts of the cloud. But consider now the median curves of the two models that involve stellar perturbations. They are mutually quite similar, but they differ strongly from the tides-only curve except at $a\lta 6000$ AU. The mutual similarity – in spite of a much larger number of grey points (injections) in Fig. \[fig:dqt\]b – means that the same basic mechanism is at work. We identify this as the angular momentum reshuffling by stellar perturbations. In Fig. \[fig:dqt\]a (stars-only model) this in itself makes comets diffuse all over angular momentum space so that some reach the target zone. In Fig. \[fig:dqt\]b (combined model) the same angular momentum diffusion repopulates the “infeed trajectories” (with $q_{\rm min} <15$ AU) of the vertical Galactic tide, whereupon the comets are injected at a rate given by $T_r(a)$. We interpret the flatness of the median curves at a level of roughly half the duration of our simulation as evidence that the time scale of angular momentum reshuffling is short enough to guarantee an injection rate that is at least as large during the first half ($0-2.5$ Gyr) as during the second half ($2.5-5$ Gyr). This is in agreement with the thermalization time scale reported by . The tendency for a slight decrease of the median $t_{\rm inj}$ at the largest semi-major axes is likely due to a progressive depletion of the outer parts of the cloud during the course of the simulation, while the reshuffling time scale is relatively short. We thus realize that the behaviour of the median injection time, generally speaking, has very little to do with any of the theoretical time scales. Let us now instead consider the lowest percentiles, since these give information on the quickest injections that – in principle – might be governed by the $T_g$ or $T_s$ dependences. The lowest percentile of the tides-only model indeed decreases with $a_o$ roughly parallel to the $T_g$ line in the inner core of the cloud ($a<10\,000$ AU), but we nonetheless see a somewhat smaller slope. This tendency gets stronger with increasing $a_o$ and finally turns into an increase outside $30\,000$ AU. We interpret this as due to the same repopulation of infeed trajectories by the non-integrable part of the tide as we discussed in relation to the median curve. The lowest percentile of the stars-only model shows a fall-off with $a_o$ that is interestingly slow in comparison with the $T_s$ line. This appears to be related to the horizontal bands of grey points, which are cometary showers. For each $a_o$ interval in the inner core, the timing of the lowest percentile is that of the first shower reaching into that interval. The larger $a_o$, the sooner such a shower appears. But the showers also get weaker, being caused by more and more distant stellar encounters. Thus, in the outer parts of the cloud they are no longer of significance for defining the lowest percentile. Since this is instead controlled by a growing number of usual, inefficient stars passing through the outer regions, one has to wait longer. When we look at the lowest percentile of the combined model, we see that it follows the same decrease as the tides-only model in the inner core. Indeed, with a much larger number of injections, the comet showers have lost their importance, and as we shall see in Sect. 4, during the first Gyr the injections are largely controlled by the Galactic tides. But outside the inner core the lowest percentile now behaves with respect to that of the tides-only model in a similar way as the median does, and the reason is the same. Going to larger semi-major axes, in both models we see an increasing number of late injections, although for different reasons, and these determine the behaviour of most statistical parameters, causing them to decrease less rapidly than the $T_g$ line. In summary we can state that we have found the theoretical time scales of to give a rather poor representation of our statistics of injection times. On the other hand, we find evidence in the combined model for an important role being played by the repopulation of tidal infeed trajectories via stellar encounters – something that may be described as a synergy effect. This being said, one nonetheless realizes that $T_r$ is one of the basic time scales that govern this synergetic injection, the other one being the angular momentum reshuffling time scale of stellar perturbations. Let us now move to a discussion of the rate of injections and how this depends on time. Time dependent injection flux ============================= The upper part of Fig. \[fig:flux\_all\] shows a histogram plot of the number of comets injected into the observable region as a function of time from the beginning till the end of the simulation. Three histograms are shown together: the one in black corresponds to a model with only Galactic tides, and the grey one to a model including only stellar perturbations. Finally, the top, white histogram is for the combined model that includes both tides and stars. ![The upper diagram shows the number of comets entering the observable zone per 50 Myr versus time. The white histogram corresponds to the combined model, the black histogram to the Galactic tide alone, and the grey histogram to the passing stars alone. The asterisks indicate the number of comets remaining in our simulation for the combined model at every 500 Myr with scale bars to the right. The middle diagram shows the excess number of injections into the observable region per 50 Myr in the combined model with respect to the sum of the stars-only and tides-only models. The lower diagram shows this excess expressed in percent of the mentioned sum. []{data-label="fig:flux_all"}](flux-all-5d7.eps){width="12.cm"} The first thing to note is the gradual decline of the injection flux of the tides-only model over a few Gyr, after which it stays at a very low level. The reason is clear. In the beginning, the phase space trajectories that in the regular dynamics imposed by the vertical tide will periodically lead into the “loss cone” ($q<15$ AU) are populated just as densely as any phase space domain and thus furnish an important flux of injections during the first period of ($e,i_G$) oscillations. This amounts to a typical time scale of $< 1$ Gyr for much of the initial cloud, but the population on these trajectories is depleted by each entry into the loss cone, and there is no efficient way to replenish them without including stellar passages. On the longer time scale, we see only the feeble flux coming from (1) the infeed into the tidal injection trajectories by the non-integrable part of the tide; (2) the inner parts of the cloud, where the period of oscillation is very long [@FOUetal:06]. The other two histograms include the effects of stellar passages, and the stars are the same in both simulations. Therefore, we see the same comet showers appearing and the same quasi-quiescent periods in between. The white area at the top of each bin corresponds to the extra contribution of the combined model as compared with that of the stars only. If the numbers plotted in the white, grey and black histograms are called $N_C$, $N_S$ and $N_G$, respectively, we can define $\Delta N_C = N_C - N_S - N_G$ as an absolute measure of this extra contribution.[^1] Already at first glance, looking at the later part of the simulation, we see that this is very significant. In the two lower panels of the Figure, we plot histograms of $\Delta N_C$ and $\tau = \Delta N_C/(N_S+N_G)$, , the extra contribution expressed as a fraction of $N_S+N_G$. The basic observations are as follows. The spiky nature of the grey histogram is due to comet showers caused by close stellar encounters (we will briefly treat these below). While during the first Gyr the level of $N_G$ is generally higher than that of $N_S$, this situation gets reversed after more than two Gyr. Even outside the main showers, $N_S$ is then at a somewhat higher level than $N_G$. The white histogram, showing $N_C$, shares the spikes of the strongest showers, but the contrast between the spikes and the background is less than in the grey histogram. Indeed, the $\Delta N_C$ histogram shows no spikes at all, and the general level does not seem to correlate with the stellar injection rate, as illustrated by Fig. \[fig:cor\_syn1\]. Therefore, during the later part of the simulation, the $\tau$ parameter shows fluctuations anticorrelated with those of $N_S$. It reaches a few hundred percent, when $N_S$ drops to its lowest levels, but sometimes decreases to nearly zero during the peaks of $N_S$. ![$\Delta N_C$ versus $N_S$ (plotted on a log scale), where the numbers refer to injections of comets into the observable region during intervals of 10 Myr.[]{data-label="fig:cor_syn1"}](synergie-DC-log.eps){width="8.cm"} In order to smooth out those fluctuations we present in Table \[tab:tau\] time averages of $\tau$ over 1 Gyr periods along with the corresponding integrals of $N_C$, $N_S$ and $N_G$. During the first Gyr the flux of the combined model is not much larger than the sum of the two fluxes with separate effects, and the difference is just a small fraction of the total flux. But toward the end the synergy effect of the combined model, as measured by $\Delta N_C$, has grown – on the average – to nearly the same level as $N_S+N_G$. During the last Gyr we find that $\langle N_C\rangle$ is about 2.5 times larger than $\langle N_S\rangle$ in fair agreement with earlier estimates by and . After an initial, relatively fast decrease due to the emptying of the tidal infeed trajectories, $\langle N_C\rangle$ continues to decrease approximately in proportion to the total number of Oort Cloud comets ($N_{OC}$), and $\langle N_S\rangle$ and $\langle N_G\rangle$ show similar behaviours. Model $ [0 - 1]$ Gyr $ [1 - 2]$ Gyr $ [2 - 3]$ Gyr $ [3 - 4]$ Gyr $ [4 - 5]$ Gyr ---------------------- ---------------- ---------------- ---------------- ---------------- ---------------- G 2128 797 481 307 248 S 1425 1555 1030 717 511 C 3618 3141 2412 1733 1274 $\langle\tau\rangle$ 1.8% 33.6% 59.6% 69.2% 67.9% : Number of comets entering the observable region during periods of 1 Gyr. Model G corresponds to the Galactic tide alone, S to passing stars alone, and C to Galactic tide and passing stars together. $\langle\tau\rangle$ is the increment from the sum of the two first rows (Galactic tide plus passing stars separately) to the third row (Galactic tide and passing stars together). []{data-label="tab:tau"} Looking in detail at the $\Delta N_C$ and $\tau$ histograms in Fig. \[fig:flux\_all\] for the beginning of our simulation, we see that they start from negative values and turn into positive ones after $\sim0.5$ Gyr. Thus, in the very beginning, the sum of the separate fluxes is larger than the combined flux. This phenomenon was found by , whose calculations were limited to only 5 Myr, and as they explained, it is typical of a situation where both tides and stars individually are able to fill the loss cone to a high degree. We will discuss this point again in Sect. 5. The large amount of synergy ($\tau\sim70$%) seen in the later part of our simulation is remarkable and indicates that both the tides and the stars on their own are seriously inefficient in filling the loss cone. It is only by means of the synergy of both effects that we are able to explain an important degree of loss cone filling at the current epoch. We will look at this closer in Sect. 5 by separating the injection flux into different ranges of semi-major axis. For the moment we emphasize that [*treating comet injections from the Oort Cloud in the contemporary Solar System simply as a result of the Galactic tide is not a viable idea*]{}. Already in Sect. 3 we identified a mechanism that offers a likely explanation of the synergy effect, , the repopulation of tidal infeed trajectories via stellar encounters. But note in Fig. \[fig:flux\_all\] that the initial flux of the model with tides only is not matched by the white areas in the later part of the simulation. Thus, even though there is an ongoing replenishment of the tidal infeed trajectories due to the randomizing effect of stellar encounters, this replenishment is not complete. [*The critical trajectories remain largely depleted, and models that do not take this fact into account will overestimate the tidal contribution to the injection flux, as well as the efficiency of tides in filling the loss cone.*]{} The most important synergy mechanism of the Galactic tide and stellar perturbations is that the latter are able to repopulate the critical phase space trajectories that in the quasi-regular dynamics imposed by the tide lead into the loss cone [@DYB:02; @FER:05]. Our results appear to verify and quantify this picture. But in addition we see hints that a different effect is also at work. In Sect. 5 we will show that the distribution of $1/a$ of newly injected, observable comets – even during the typical, quiescent periods – has a significant extension inside the limit (at $a\simeq3\times10^4$ AU), where the tide becomes able to feed comets from outside the loss cone into observable orbits. Our explanation for this effect is as follows. In qualitative terms, when the Galactic tide is in the process of injecting a comet into the observable region from the region outside the loss cone, and stellar perturbations are added, the latter will sometimes aid in decreasing the perihelion distance of the comet ($\Delta q_\ast < 0$), and sometimes they will counteract the tide ($\Delta q_\ast > 0$). To first order, the two effects will cancel. But if we consider how much the critical value ($1/a_c$) of the inverse semi-major axis for tidal injection into the observable region ($a_c\simeq 3\times10^4$ AU) is decreased by a typical positive $\Delta q_\ast$ or increased by the same typical value of $\Delta q_\ast$ in the negative direction, we find that the latter effect dominates. Thus, a negative $\Delta q_\ast$ causes a larger gain of comets with $a<a_c$ than the loss of comets with $a>a_c$ caused by a positive $\Delta q_\ast$ of the same size. This holds for any nearly uniform distribution of $1/a$ in the Oort Cloud. In mathematical terms, consider the expression for the maximum possible decrease of $q$ in one revolution due to the Galactic tide: $$\label{eq:max_pert} \Delta q = - S q^{1/2} z^{-7/2}$$ where $z=1/a$ and $S=2.8\times10^{-15}$ [@BYL:86], counting $q$ and $a$ in AU. This would hold for a Galactic latitude of perihelion of $\pm45^\circ$. We take this favourable orbital orientation as an example, but the following arguments apply for any other orientation as well. Next, consider a particular value ($q_p$) of the perihelion distance preceding the injection into an observable orbit. Using Eq. (\[eq:max\_pert\]), one can write down an approximate condition for the critical value $z=z_0$ in order to bring the comet into the observable region ($q<q_\ell=5$ AU): $$\label{eq:ener_lim} q_p = q_\ell + S q_p^{1/2} z_0^{-7/2}$$ Eq. (\[eq:ener\_lim\]) defines a unique relation between $q_p$ and $z_0$, and by differentiating one easily finds that $q_p$ decreases monotonously with $z_0$, while the second derivative is always positive. Considering thus an arbitrary point ($z_0,q_p$) satisfying Eq. (\[eq:ener\_lim\]), we may introduce stellar perturbations by adding a term $-\Delta q_\ast$ to the right-hand member of Eq. (\[eq:ener\_lim\]), and we can write: $$\label{eq:el_stars} q_p = q_\ell + S q_p^{1/2} z_1^{-7/2} - \Delta q_\ast$$ where $z_1$ is the new critical value of $z$. Hence, ($z_1,q_p+\Delta q_\ast$) also satisfies Eq. (\[eq:ener\_lim\]). Without the stellar perturbation all comets with $q=q_p$ and $z<z_0$ are injected into $q<q_\ell$, and including the stellar perturbation, the condition changes into $z<z_1$. From the negative slope of the $q_p(z_0)$ relation it is obvious that a negative value of $\Delta q_\ast$ yields $z_{1+}>z_0$, and the same positive value yields $z_{1-}<z_0$. It is also obvious from the positive curvature of the graph that $z_{1+}-z_0>z_0-z_{1-}$. If $\vert\Delta q_\ast\vert$ is very small, this difference is negligible, but if it is large enough to compete with $\vert\Delta q_G\vert$, the effect will be important. The latter is indeed often the case, when we discuss injections from just outside the loss cone ($q_p\simeq15$ AU), as has been shown, , by and . Therefore, the gain of comets with $z>z_0$ occurs over a larger interval than the loss with $z<z_0$ for a symmetric distribution of stellar perturbations. Another issue is the distribution function of $z$ for the Oort Cloud. Our simulations start with a probability density $f(z)\propto z^{-1/2}$ as appropriate for an Oort Cloud formed according to the model of . In such a situation there would be more comets per unit interval of $z$ at $z<z_0$ than at $z>z_0$, and the gain effect would be counteracted and possibly turned into a net loss of injected comets. However, an interesting result of our simulations is that the gradual loss of comets from the Oort Cloud changes the distribution of $1/a$. In agreement with a recent study by , we find that the loss of comets from the outer parts of the cloud is not compensated by diffusion from the inner parts, so that after 4.5 Gyr, when the number of comets has decreased from $1\times10^6$ to $7.6\times10^5$, $f(z)$ has become roughly flat over the range from $a=20\,000$ to $100\,000$ AU. This shows that we have to expect a net gain of newly injected, observable comets resulting from a synergy of $\Delta q_G$ and $\Delta q_\ast$. Moreover, there should be a shift of comets from outside to inside the tidal injection limit – probably explaining why we see a significant flux of new comets all the way down to $a\simeq20\,000$ AU. Although we cannot provide exact numbers, it appears that the secondary synergy mechanism due to what we may call “constructive interference” of the two effects – even though it certainly exists – is not the dominating one. The tentative evidence comes from the relative smoothness of the $\Delta N_C$ histogram (Fig. \[fig:flux\_all\]) and the lack of correlation between $\Delta N_C$ and $N_S$ (Fig. \[fig:cor\_syn1\]). These properties are expected of the repopulation of tidal infeed trajectories because of the long response time ($\sim$ several $10^8$ yr) for tidal infeed on the $T_r$ time scale (Fig. \[fig:dqt\]). But if the constructive interference had been very important, we would have expected $\Delta N_C$ to increase immediately upon an increase of $N_S$ – with the caveat that visible peaks of our $N_S$ histogram might arise primarily from a temporary infeed of inner Oort Cloud comets with $z>>z_o$, which do not contribute to the interference. Further detailed studies are needed to clarify this issue. ![Same as Fig. \[fig:flux\_all\] (upper panel), but versus time per 10 Myr for the combined model only.[]{data-label="fig:flux_10d6"}](flux-C-1d7.eps){width="12.cm"} The cometary showers, displayed in Fig. \[fig:flux\_10d6\] by means of a histogram of the injection flux of the combined model with a bin width of only 10 Myr, are seen to be quite important for the injection of comets from the Oort Cloud, as expected and as found by other authors previously (, ). We are saving a detailed analysis of those for a later paper. At present, we can only remark that the results presented here are hard to compare with the treatments of cometary showers by or . The first of these papers treated only comets with $a=10\,000$ or $20\,000$ AU with a full dynamical model and then only for a time interval of less than 200 Myr. The second gave only a brief account of a simulation for 4.5 Gyr and then only for injections from $q>10$ to $q<10$ AU instead of our requirement that the perihelion has to fall substantially from $q>15$ to $q<5$ AU. Orbital element distributions of observable comets ================================================== Figure \[fig:IN\_4.55d9-4.72d9\] shows the distributions of the opposite of the inverse semi-major axis $(-1/a)$ and the sine of the Galactic latitude of perihelion (for clarity we use the absolute value $|\sin b|$) of the comets entering the observable region, , heliocentric distance smaller than 5 AU, during a typical 170 Myr interval near the end of our simulation, where no strong comet showers are registered. We show an average of three such periods, , $4.38-4.55$ Gyr, $4.55-4.72$ Gyr and $4.80-4.97$ Gyr. In fact, comparing the three data sets, we find a rather good agreement, so that tentatively, the expected error of the mean is not very large. Three histograms are shown for each quantity: the one in black is for the model with Galactic tides only, the grey one is for the model with only stellar perturbations, and the white one shows the combined model. \[cc\][$-1/a\times 10^5~{\rm (AU^{-1})}$]{} (11.8,6.) (7.54,0.)[![Distributions of $-1/a$, where $a$ is the semi-major axis (top panels) and $|\sin b|$, where $b$ is the Galactic latitude of perihelion (bottom panels), for the comets entering the observable region during 170 Myr near the end of the simulation. When present, numbers in the top-left corners of $-1/a$ distribution panels correspond to comets with $-1/a<-1 \times 10^{-4}~{\rm AU}^{-1}$. The left column corresponds to the model with Galactic tide alone, the middle column to passing stars alone, and the right column to the model with both effects.[]{data-label="fig:IN_4.55d9-4.72d9"}](IN-C-quiet_end_moy.eps "fig:"){width="4.cm" height="6.cm."}]{} (3.85,0.)[![Distributions of $-1/a$, where $a$ is the semi-major axis (top panels) and $|\sin b|$, where $b$ is the Galactic latitude of perihelion (bottom panels), for the comets entering the observable region during 170 Myr near the end of the simulation. When present, numbers in the top-left corners of $-1/a$ distribution panels correspond to comets with $-1/a<-1 \times 10^{-4}~{\rm AU}^{-1}$. The left column corresponds to the model with Galactic tide alone, the middle column to passing stars alone, and the right column to the model with both effects.[]{data-label="fig:IN_4.55d9-4.72d9"}](IN-S-quiet_end_moy.eps "fig:"){width="4.cm" height="6.cm"}]{} (0,0.)[![Distributions of $-1/a$, where $a$ is the semi-major axis (top panels) and $|\sin b|$, where $b$ is the Galactic latitude of perihelion (bottom panels), for the comets entering the observable region during 170 Myr near the end of the simulation. When present, numbers in the top-left corners of $-1/a$ distribution panels correspond to comets with $-1/a<-1 \times 10^{-4}~{\rm AU}^{-1}$. The left column corresponds to the model with Galactic tide alone, the middle column to passing stars alone, and the right column to the model with both effects.[]{data-label="fig:IN_4.55d9-4.72d9"}](IN-G-quiet_end_moy.eps "fig:"){width="4.cm" height="6.cm"}]{} After more than 4 Gyr the Galactic tides alone are practically only able to inject comets into the observable region if $a>50\,000$ AU, so that the non-integrable part of the tides may provide new comets into the emptied infeed trajectories of the vertical component. Thus the feeble flux of new observable comets is strictly confined to the outermost parts of the Oort cloud. If only the stellar perturbations are at work, the injected comets are almost as few as in the case of the Galactic tides. However, the distribution of $- 1/a$ shows that the stellar perturbations are relatively efficient injectors of comets with semi-major axes in the whole range from $25\,000$ to more than $100\,000$ AU, and there is some marginal infeed all the way into the inner core. Note that this concerns a time interval without any strong comet showers. When both the processes are at work, the number of comets entering the observable zone is 206, about 86% more than the sum of the two separate contributions ($39 + 72$). This estimate of $\tau$ is a bit higher than for the entire 1 Gyr interval, listed in Table \[tab:tau\], because the three intervals have been selected as particularly calm, avoiding even the smaller peaks of $N_S$ that can be seen in Fig. \[fig:flux\_all\]. We have shown above that larger values of $N_S$ lead to smaller values of $\tau$. The distribution of $ -1/a$ is as wide as for the stellar perturbations alone. However, the picture has changed, since the additional 86% of the comets are strongly concentrated to the interval from $-4\times 10^{-5}$ to $-2\times 10^{-5}~{\rm AU}^{-1}$ ($25\,000\,<\,a\,<\,50\,000$ AU). The local values of $\Delta N_C$ for the five $1/a$ intervals $(0-1)$, $(1-2)$, $(2-3)$, $(3-4)$ and $(4-5)\times10^{-5}$ AU$^{-1}$ are $-2$, $-22$, $+63$, $+36$ and $+10$, respectively. We will comment on the negative values of the first entries in connection with Fig. \[fig:IN\_6.3d8-8d8\]. We see that the mechanism of synergy that increases the flux of injections in the combined model prefers the range of semi-major axis ($a>30\,000$ AU) where the vertical Galactic tide is able to provide the injections, once the relevant trajectories are populated. But there is an important extension of the synergy to smaller semi-major axes as well, extending at least to $a\simeq20\,000$ AU. We conclude that both the above-described synergy mechanisms must be at work. The repopulation mechanism is obviously important, but the shift to smaller semi-major axes can only be explained by the ‘constructive interference’ mechanism. Looking at the distributions of $|\sin b|$, indeed the signature of the Galactic tide is clearly present in the left diagram and absent in the middle one. However, it appears again to some extent in the right-hand diagram, where the combined model is presented. Thus we have evidence that the synergetic injection of comets in the combined model carries an imprint in the latitudes of perihelia similar to that of the Galactic tide, though the feature is strongly subdued. In fact, while the subdued tidal imprint is consistent with an important role being played by the ‘constructive interference’ synergy mechanism, our combined model does not appear to reproduce the observed $|\sin b|$ distribution of new Oort Cloud comets. An in-depth study of this problem and a consideration of ways out of this possible dilemma are deferred to future papers. The shaping of the $b$ distribution by the Galactic tide was first simulated numerically for a realistic Oort Cloud model by . However, the left panels of Fig. \[fig:IN\_4.55d9-4.72d9\] show a behaviour that is in stark contrast to their results. Practically all our tidal injections occur for $a>50\,000$ AU, where found no tidal imprint in the $b$ distribution because of complete loss cone filling independent of $b$. In the light of our results this can be seen as an artifact of their assumption of complete randomization of the Oort Cloud orbit distribution. Indeed, as we shall find below (Table \[tab:flc\_gsc\]), the tidal loss cone filling for $50\,000<a<100\,000$ AU towards the end of our simulation is far from complete, and therefore we see the imprint of the tide in our $b$ distribution. In Fig. \[fig:IN\_3.85d9-3.86d9\] we show the corresponding distribution of $-1/a$ and $|\sin b|$ for the 10 Myr interval from $3.85$ to $3.86$ Gyr, where Fig. \[fig:flux\_10d6\] shows that the number of observable comets has a high peak due to a moderately strong shower. Occurring near the middle of the period, this shower dominates the time-integrated injection rate. The trigger is a M5 star with an impact parameter $d\simeq2\,000$ AU and a velocity $v_\star\simeq18$ km/s. \[cc\][$-1/a\times 10^5~{\rm (AU^{-1})}$]{} (11.8,6) (7.56,0.)[![Same as Fig. \[fig:IN\_4.55d9-4.72d9\] but considering the comets that enter the observable region during a shower between $3.85$ and $3.86$ Gyr. The shower is due to a M5 star with impact parameter $2\,055$ AU, velocity 17.7 km/s and mass $0.21~M_\odot$.[]{data-label="fig:IN_3.85d9-3.86d9"}](IN-C-shower.eps "fig:"){width="4.cm" height="6.cm"}]{} (3.80,0.)[![Same as Fig. \[fig:IN\_4.55d9-4.72d9\] but considering the comets that enter the observable region during a shower between $3.85$ and $3.86$ Gyr. The shower is due to a M5 star with impact parameter $2\,055$ AU, velocity 17.7 km/s and mass $0.21~M_\odot$.[]{data-label="fig:IN_3.85d9-3.86d9"}](IN-S-shower.eps "fig:"){width="4.cm" height="6.cm"}]{} (0,0.)[![Same as Fig. \[fig:IN\_4.55d9-4.72d9\] but considering the comets that enter the observable region during a shower between $3.85$ and $3.86$ Gyr. The shower is due to a M5 star with impact parameter $2\,055$ AU, velocity 17.7 km/s and mass $0.21~M_\odot$.[]{data-label="fig:IN_3.85d9-3.86d9"}](IN-G-shower.eps "fig:"){width="4.cm" height="6.cm"}]{} The mid and right-hand $-1/a$ distributions show that, as soon as stars are involved, the injection of comets now extends over the whole cloud, including an important fraction from the inner core with $a\,<\,10\,000$ AU. In fact, the synergy effect is now very strong in the range from $10\,000$ to $20\,000$ AU, amounting to $\tau>150$%. This is unexpected on the basis of both the above-mentioned mechanisms, since we are discussing orbits too far inside the tidal injection limit. We are instead led to hypothesize a different mechanism. In the present case we are comparing the number of comets injected by a particular, deeply penetrating star from the mentioned range of semi-major axes in the stars-only vs the combined model. In the absence of the Galactic tides it is likely that orbits with perihelia close to but outside the loss cone have been depleted by the preceding cometary showers, while in the combined model the disk tide provides a regular transfer of comets into this zone on a Gyr time scale, thus compensating for the losses. This means that the synergy works in the opposite sense compared to the normal situation outside the showers. [*During a shower the tides are providing the material for injections by the stars, while in the normal situation the stars are providing the material for tidal injections.*]{} The absence of a synergy in the inner core may be explained by the very long time scale for tidal torquing, or by a smaller degree of depletion of the source region for stellar injections. Naturally, in the stars-only model the shower does not carry any signature in the distribution of $|\sin b|$. The combined model does not exhibit any significant signature either, but there may nonetheless be a slight tendency. In case this is real, it might possibly reveal a somewhat more efficient synergy in the $10\,000-20\,000$ AU range for the orbits experiencing a faster tidal decrease of $q$. \[cc\][$-1/a\times 10^5~{\rm (AU^{-1})}$]{} (11.8,6) (7.56,0.)[![Same as Fig. \[fig:IN\_4.55d9-4.72d9\] but considering the comets entering the observable region during 170 Myr near the beginning of the simulation.[]{data-label="fig:IN_6.3d8-8d8"}](IN-C-quiet_deb.eps "fig:"){width="4.cm" height="6cm"}]{} (3.80,0.)[![Same as Fig. \[fig:IN\_4.55d9-4.72d9\] but considering the comets entering the observable region during 170 Myr near the beginning of the simulation.[]{data-label="fig:IN_6.3d8-8d8"}](IN-S-quiet_deb.eps "fig:"){width="4.cm" height="6cm"}]{} (0.,0.)[![Same as Fig. \[fig:IN\_4.55d9-4.72d9\] but considering the comets entering the observable region during 170 Myr near the beginning of the simulation.[]{data-label="fig:IN_6.3d8-8d8"}](IN-G-quiet_deb.eps "fig:"){width="4.cm" height="6cm"}]{} Let us now consider the situation at the beginning of the simulation, before the tides have had the time to completely empty the tidal infeed trajectories in the outer part of the cloud. The results are shown in Fig. \[fig:IN\_6.3d8-8d8\] for a period between $0.63$ and $0.80$ Gyr, when no strong showers are noted. The number of comets entering into the observable region is 282, 128 and 463 for the tides, the stars, and the combined model, respectively. The action of the tides is still quite strong, since the infeed trajectories in the interval $20\,000<a<50\,000$ AU are not yet severely depleted (cf. Fig. \[fig:dqt\]). Therefore the net synergy effect amounts to only $13$%. The local $\Delta N_C$ values for the same five $1/a$ intervals as we discussed in connection with Fig. \[fig:IN\_4.55d9-4.72d9\] are in this case $-5$, $-66$, $+35$, $+61$ and $+19$. We do not see any significant synergy effect for more tightly bound orbits. The distribution of positive synergy over the $2-5\times10^{-5}$ AU$^{-1}$ range is similar to that of Fig. \[fig:IN\_4.55d9-4.72d9\], and our conclusions about the relevance of the two mechanisms are the same. Note that in both cases we see negative $\Delta N_C$ values in the two outermost $1/a$ ranges ($0-2\times10^{-5}$ AU$^{-1}$). The fundamental reason is the one discussed above in order to explain the negative $\tau$ values in the very beginning of the simulation, , a saturation effect of the loss cone when both injection effects individually are able to cause a large degree of filling. The distributions of $|\sin b|$ exhibit the same features as in Fig. \[fig:IN\_4.55d9-4.72d9\] and lead to the same conclusion: when both tides and stars act in synergy, the signature of the Galactic tide may be seen but appears quite marginal. Since we noted in Sect. 4 that the replenishment of the tidal infeed trajectories by stellar perturbations is not complete during the later part of our simulation, an obvious consequence is that the filling of the loss cone for the relevant semi-major axes cannot be complete either. To quantify this statement, we consider the rate of perihelion passages $\dot{n}$ with $q<q_0$ as a function of the semi-major axis, assuming complete loss cone filling and a completely thermalized Oort cloud [@HIL:81; @BAI.STA:88]: $$\label{eq:loss_cone} \dot{n}(a)=N_{\rm OC} \cdot f(a) \cdot \frac{2\,q_0}{a}\cdot a^{-3/2}.$$ Here, $N_{\rm OC}$ is the number of comets in the entire Oort cloud (initially $10^6$ in our model) and $f(a)$ the frequency function describing the distribution of semi-major axes: $f(a) \propto a^{-1.5}$ initially in our model. We have computed injection rates in the combined model for the three time intervals of Figs. \[fig:IN\_4.55d9-4.72d9\]–\[fig:IN\_6.3d8-8d8\] using Eq. (\[eq:loss\_cone\]) and finding the integrals $\int N_{\rm OC} f(a) a^{-5/2} da$ over different ranges of $-1/a$ directly from the simulation output at neighbouring moments. The calculation of these integrals is done by simply adding the values of $a^{-5/2}$ of all the comets found in the relevant range. Multiplying by the length of each interval and putting $q_0\,=\,5$ AU, we find the numbers $N_{\rm comp}$ listed in Table \[tab:flc\]. --------------------------- ---------------- --------------- -------------- ---------------- --------------- -------------- ---------------- --------------- -------------- $\Delta (1/a)$ $(10^{-5}~{\rm AU}^{-1})$ $N_{\rm comp}$ $N_{\rm sim}$ $f_{\rm lc}$ $N_{\rm comp}$ $N_{\rm sim}$ $f_{\rm lc}$ $N_{\rm comp}$ $N_{\rm sim}$ $f_{\rm lc}$ $(0-2) $ 106 111 $\sim100$% 3.2 2 $\sim60$% 48 40 83% $(2-3) $ 303 189 62% 10 6 $\sim60$% 160 83 52% $(3-4) $ 616 111 18% 24 2 $\sim8$% 367 46 13% $(4-5) $ 1044 23 2.2% 44 6 14% 692 14 2.0% $(5-10)$ 15600 23 0.15% 740 29 3.9% 12100 13 0.11% $ >10 $ 672000 6 0.0009% 37200 31 0.08% 626000 10 0.0016% --------------------------- ---------------- --------------- -------------- ---------------- --------------- -------------- ---------------- --------------- -------------- : Numbers of comet injections during the time intervals of Figs. \[fig:IN\_4.55d9-4.72d9\]–\[fig:IN\_6.3d8-8d8\] for different ranges of inverse semi-major axis, as computed from Eq. (\[eq:loss\_cone\]) and found from our simulation of the combined model. The ratio of simulated to computed number, expressed in percent, is also listed in each case.[]{data-label="tab:flc"} The numbers of comet injections for each $1/a$ range and each time interval can be read off from the figures, and they are listed as $N_{\rm sim}$ in the Table along with the ratios $N_{\rm sim}/N_{\rm comp}$, which give the filling factor of the observable part of the loss cone ($f_{\rm lc}$). We find that this factor is close to 100% in the beginning of the simulation for $a>50\,000$ AU and remains $>80$% for such semi-major axes even towards the end during quiescent periods. But the factor drops rapidly with decreasing $a$ to values near 2% at $a\simeq20\,000$ AU. These results may be compared with those of , who used a similar procedure for deriving $f_{\rm lc}$. She did not consider semi-major axes $a>40\,000$ AU, and inside this limit we find somewhat smaller filling factors than she did, consistent with the fact that we use a lower value for the Galactic mid-plane density and somewhat higher stellar velocities. Note that the filling factors have decreased somewhat, when we compare the final quiescent period with the initial one. Except in the outermost parts of the Oort Cloud, there is always a depletion of comets in the regions of phase space near the tidal infeed trajectories and in the vicinity of the loss cone, and this depletion grows slowly with time. The numbers $N_{\rm sim}$ found for the shower period are too small to be statistically useful for the outer parts of the cloud, and the filling factors listed are very uncertain. However, we see an obvious effect in the inner parts, when comparing $f_{\rm lc}$ with the corresponding values of quiescent periods. The shower increases $f_{\rm lc}$ by factors $\sim20-100$, and thus the overall flux exhibits the peak seen in Fig. \[fig:flux\_10d6\] due to comets with $a<20\,000$ AU. --------------------------- ------------ ------------ ------------ ------------ ------------ ------------ $\Delta (1/a)$ $(10^{-5}~{\rm AU}^{-1})$ Tidal Stellar Combined Tidal Stellar Combined $(0-1) $ $\sim400$% $\sim100$% $\sim200$% $\sim100$% $\sim100$% $\sim100$% $(1-2) $ 86% 60% $\sim100$% 30% 45% 78% $(2-3) $ 36% 10% 62% 0.6% 10% 52% $(3-4) $ 6.5% 1.1% 18% – 2.3% 13% $(4-5) $ 0.09% 0.3% 2.2% – 0.5% 2.0% $(5-10)$ – 0.1% 0.15% – 0.06% 0.11% $ >10 $ – 0.0006% 0.0009% – 0.0008% 0.0016% --------------------------- ------------ ------------ ------------ ------------ ------------ ------------ : Filling factors for the observable part of the loss cone, computed for different ranges of semi-major axis and separately for the three dynamical models (tides-only, stars-only, and combined). []{data-label="tab:flc_gsc"} We have already made the remark that neither $\Delta N_C$ nor $\tau$ provides a fully satisfactory measure of the synergy effect, because they do not account for the difference of the number of Oort Cloud comets between different dynamical models – especially towards the end of the simulation. After 5 Gyr the total number of comets in the combined model is only $\sim80$% of that in the tides-only model, and if we concentrate on comets with $50\,000<a<100\,000$ AU where the losses are the largest, the ratio of the two models is only 35%. In order to compensate for such effects we have computed the $f_{\rm lc}$ parameter separately for the three models and for all the ranges of $1/a$, and we present the results in Table \[tab:flc\_gsc\]. The time periods referred to are the quiescent periods of Figs. \[fig:IN\_4.55d9-4.72d9\] and \[fig:IN\_6.3d8-8d8\]. The outermost energy range is empty in all models, when the simulation starts, but it gets populated quickly – at least when stars are involved. We interpret the very large value of $f_{\rm lc}$ in the tides-only model at the beginning as evidence that the radial tide has not extracted comets into this energy range in a uniform manner, so that our assumption of thermalization when deriving $N_{\rm comp}$ is not justified. To a lesser extent this appears to be true also in the combined model, where stars have extracted many more comets. It is likely that this extraction too – at the early time in question – has not populated all the angular momenta in a thermalized fashion. However, the statistics is too poor to be confident about such conclusions. In any case, the loss cone filling is extremely efficient for all models, thus explaining the negative values of $\Delta N_C$. For the next energy range we see the saturation effect again, especially in the beginning. At the end, the value of $f_{\rm lc}$ in the combined model is close to the sum of those in the other two models. Hence there is no apparent synergy in this case, but probably a real synergy has been concealed by the saturation effect. In any case the large negative value of $\Delta N_C$ results entirely from the small number of comets in the combined model, as discussed above. In the next three energy ranges ($20\,000<a<50\,000$ AU) we see that $f_{\rm lc}$ in the combined model is much larger than the sum of the two other entries, and for $a<20\,000$ AU the effect continues: adding the tides to the stars increases the loss cone filling by a factor $1.5-2$. Discussion and conclusions {#sec:conclu} ========================== We have simulated the evolution of the Oort cloud over 5 Gyr, using for initial conditions a relaxed model with a distribution of semi-major axis $f(a) \propto a^{-1.5}$ within the interval $3\,000 - 100\,000$ AU. This model is based on the results of simulation of Oort Cloud formation and evolution by . However, we do not find this to be a steady distribution. More comets are lost from the outer parts of the cloud than can be replaced from inside, so that our model cloud evolves into a distribution close to $f(a) \propto a^{-2}$ – , flat in $1/a$. Our dynamical model has two main limitations. We do not treat encounters with very massive Galactic perturbers, such as star clusters or Giant Molecular Cloud complexes, the justification being that they occur so rarely that the current Solar System is unlikely to feel the direct reverberations of any such encounter, and that even if they modify the structure of the Oort cloud, our interest is not primarily in its dynamical history but rather in the way stars and Galactic tides currently interact when injecting observable comets. Moreover, we do not treat planetary perturbations in any direct manner. Like most previous investigators ([*e.g.*]{}, ) we use a loss cone defined by a limiting perihelion distance (in our case, 15 AU) outside which no planetary effects are included and inside which all comets are considered lost from the cloud through perturbations by Jupiter and Saturn. In terms of “transparency” of the planetary system [@DYB.PRE:97; @DYB:05], our model is completely opaque ($P=1$). This means that we are limiting our attention to a subset of the observed population of “new” Oort cloud comets, , those that have jumped directly from $q\,>\,15$ AU into their observed orbits with $q\,<\,5$ AU. We are neglecting the rest of the population, which consists of comets that passed perihelia in the outer part of the loss cone without being perturbed away before arriving into observable orbits. We are also neglecting a possible contribution to the observed new comets by a “leakage” from the scattered disk [@LEVetal:06]. Therefore we prefer not to draw any conclusions in this paper regarding the total number of Oort Cloud comets or the exact values of the filling factors. Nor do we claim to make any prediction on the detailed shape of the $1/a$ distribution of new Oort Cloud comets, until we have included the planetary perturbations in a realistic manner. We have shown that the concept of tidal and stellar torquing time scales [@DUNetal:87] gives a very incomplete picture of the speed of comet injection, whether it may concern Galactic tides or stellar encounters. The distribution of injection times is largely shaped by other effects – like comet showers or the repopulation of the emptied infeed trajectories of the disk tide due to the non-integrable part of the tides or stellar perturbations. We have also shown how – for semi-major axes large enough for the tide to populate observable orbits – the regions of the phase space occupied by trajectories leading into the loss cone get depleted during the first Gyr of Oort cloud evolution. This would leave little chance for the tide to produce a significant number of observable comets at the current time, were it not for the capability of stellar perturbations to replenish the tidal infeed trajectories. We have indeed demonstrated that, during the later parts of our simulation, there is a very important synergy effect of the Galactic tide and stellar perturbations such that the combined injection rate is on the average $\sim70$% larger than that of the stars alone plus that of the tide alone. This synergy is strongest for semi-major axes between $\sim20\,000-50\,000$ AU but continues all the way into the inner core. During comet showers the synergy effect in the outer parts of the cloud practically disappears, but the one affecting the inner parts becomes very important. We have identified two mechanisms for the synergy during quiescent periods in the outer parts of the Oort Cloud. One is that the stellar perturbations provide a supply of new comets that replenishes the depleted tidal infeed trajectories, and the other is that the gain of comet injections, when stellar perturbations decrease the perihelion distance, dominates over the loss caused by opposing perturbations. For the synergy of the inner cloud we hypothesize that the Galactic tides provide the material for stellar injections by slowly feeding the region of phase space in the vicinity of the loss cone. Thus, the general picture spawned by our results is that injection of comets from the Oort Cloud is essentially to be seen as a team work involving both tides and stars. It appears meaningless to rank the two effects in terms of strength or efficiency. Indeed, for the smaller semi-major axes the Galactic tide does not dominate the injection of comets, contrary to the conclusions of and .[^2] It only contributes to a synergy with stellar perturbations, and without the stars one would not have any injections of comets with $a\,\lta\,20\,000$ AU. The distribution of Galactic latitudes of perihelia of the observable comets exhibits a maximum for $|\sin b|\simeq0.5$ as expected in the tides-only model, but in the combined model this feature can hardly be seen at all. The tides form part of the synergetic injection, but their imprint is largely washed out by the stellar contribution. But, likely due to the role of the tides in helping the stars to create comet showers, the pattern can be seen at least as clearly during a shower as during the quiescent periods. Therefore, it tentatively appears that the shape of the observed $b$ distribution can not be used to indicate whether we are experiencing any shower at present. However, since none of our model distributions appears to agree with the observed one, we have to defer any conclusions to future papers. It may be interesting to see, for instance, if the leakage from the scattered disk into the Oort Cloud with an ensuing direct transfer into observable new comets may alleviate the problem. We have measured the filling of the observable part of the loss cone by comparing our simulated injection rates for different intervals of semi-major axis with the rates of observable perihelion passages ($q\,<\,5$ AU) computed for a completely thermalized distribution of cometary orbits involving a filled loss cone. The deficiency of our simulated rate likely reflects not only a lack of comets in the loss cone but a general depletion in a wider phase space region in its vicinity, as remarked by . Our results can be compared with hers, and in contrast to her inference that $f_{\rm lc}$ may level out at $\sim60$% for $a\gta30\,000$ AU, we find an average filling factor during quiescent periods in the current Solar System, which drops steadily from $\sim100$% at $a>100\,000$ AU to 1% or less at $a<20\,000$ AU in the combined model. However, there are important differences between the two investigations, one being that she simulated a much shorter time period than we do, and in addition our parameters for the Galactic tides and stellar encounters also differ from hers. In agreement with , we find that cometary showers do not significantly increase the loss cone filling at large semi-major axes. However, near $25\,000$ AU there is an abrupt change into the regime of the inner cloud, where the filling factor increases by orders of magnitude during moderate to strong events. The showers of course involve direct injections by the passing stars, but the synergy with the Galactic tide is as important as during quiescent periods. Acknowledgments {#acknowledgments .unnumbered} =============== We are greatly indebted to the referees of this paper, Ramon Brasser and Julio A. Fernández, for having pointed out various weaknesses in a previous version, which inspired us to exert more care in all parts of our analysis. M.F. is grateful to GDRE 224 CNRS INdAM, GREFI-MEFI for financial support. For H.R., this work was supported by Grants nr. 78/06 and 119/07:1 of the Swedish National Space Board. The work of G.B.V. was supported by the contract ASI/INAF I/015/07/0. [99]{} Allen, C.: 1985, [*Astrophysical Quantities, 3rd ed.*]{} London: Athlone Press. , J. N.: 1984, ‘[Self-consistent determinations of the total amount of matter near the sun]{}’. , 169–181. , M. E. and C. R. [Stagg]{}: 1988, ‘[Cratering constraints on the inner Oort cloud - Steady-state models]{}’. , 1–32. Brasser R., M.J. Duncan, and H.F. Levison: 2008, ‘[Embedded star clusters and the formation of the Oort Cloud III: The Galactic phase]{}’. , in press. Breiter, S., M. Fouchard, R. Ratajczak, and W. Borczyk: 2007, ‘[Two fast integrators for the Galactic tide effects in the Oort Cloud]{}’. , 1151–1162. , J.: 1986, ‘[The effect of the Galaxy on cometary orbits]{}’. , 263–273. , S. and A. [Morbidelli]{}: 2007, ‘[Coupling dynamical and collisional evolution of small bodies II. Forming the Kuiper belt, the Scattered Disk and the Oort cloud]{}’. , 468–480. , A. H.: 1987, ‘[Galactic Tides Affect the Oort Cloud - an Observational Confirmation]{}’. , 913–918. , M., T. [Quinn]{}, and S. [Tremaine]{}: 1987, ‘[The formation and extent of the solar system comet cloud]{}’. , 1330–1338. , P. A.: 2002, ‘[Simulating observable comets. I. The effects of a single stellar passage through or near the Oort cometary cloud]{}’. , 283–292. , P. A.: 2005, ‘[Simulating observable comets. II. Simultaneous stellar and galactic action]{}’. , 783–790. , P. A. and H. [Pr[ȩ]{}tka]{}: 1997, ‘[The Galactic Disk Tidal Force: Simulating the Observed Oort Cloud Comets]{}’. In: I. M. [Wytrzyszczak]{}, J. H. [Lieske]{}, and R. A. [Feldman]{} (eds.): [*IAU Colloq. 165: Dynamics and Astrometry of Natural and Artificial Celestial Bodies*]{}. p. 149. , V. V., D. J. [Asher]{}, and M. E. [Bailey]{}: 2007, ‘[The fundamental role of the Oort cloud in determining the flux of comets through the planetary system]{}’. , 779–789. , E.: 1985, ‘[An efficient integrator that uses Gauss-Radau spacings]{}’. In: A. [Carusi]{} and G. B. [Valsecchi]{} (eds.): [*ASSL Vol. 115: IAU Colloq. 83: Dynamics of Comets: Their Origin and Evolution*]{}. p. 185. , J. A.: 1980, ‘[Evolution of comet orbits under the perturbing influence of the giant planets and nearby stars]{}’. , 406–421. , J. A. (ed.): 2005, ‘[Comets - Nature, Dynamics, Origin and their Cosmological Relevance]{}’, Vol. 328 of [*Astrophysics and Space Science Library*]{}. , M., Ch. [Froeschl[é]{}]{}, S. [Breiter]{}, R. [Ratajczak]{}, , and H. [Valsecchi]{}, G. [Rickman]{}: 2007, ‘[Methods to Study the Dynamics of the Oort Cloud Comets II: Modelling the Galactic Tide]{}’. In: D. [Benest]{}, C. [Froeschlé]{}, and E. [Lega]{} (eds.): [*Topics in Gravitational dynamics*]{}, Vol. 729 of [*Lecture Notes in Physics, Berlin Springer Verlag, in press*]{}. pp. 271–293. , M., Ch. [Froeschl[é]{}]{}, G. [Valsecchi]{}, and H. [Rickman]{}: 2006, ‘[Long-term effects of the Galactic tide on cometary dynamics]{}’. , 299–326. , J., P. R. [Weissman]{}, R. A. [Preston]{}, D. L. [Jones]{}, J.-F. [Lestrade]{}, D. W. [Latham]{}, R. P. [Stefanik]{}, and J. M. [Paredes]{}: 2001, ‘[Stellar encounters with the solar system]{}’. , 634–659. , J.: 1990, ‘[Monte Carlo simulations of the Oort comet cloud]{}’. , 104–121. , J. and S. [Tremaine]{}: 1986, ‘[The influence of the galactic tidal field on the Oort comet cloud]{}’. , 13–26. , J., S. [Tremaine]{}, and C. [Alcock]{}: 1987, ‘[The Frequency and Intensity of Comet Showers from the Oort Cloud]{}’. , 269–288. , J., S. [Tremaine]{}, P. [Weissman]{}, and R. [Greenberg]{}: 1991, ‘[Sky Distributions of Oort Cloud Comets During and Outside of Showers]{}’. In: [*Lunar and Planetary Institute Conference Abstracts*]{}, Vol. 22 of [*Lunar and Planetary Institute Conference Abstracts*]{}. p. 553. , J. G.: 1981, ‘[Comet showers and the steady-state infall of comets from the Oort cloud]{}’. , 1730–1740. , J. and C. [Flynn]{}: 2000, ‘[The local density of matter mapped by Hipparcos]{}’. , 209–216. , P. and S. [Tremaine]{}: 1985, ‘[Have interstellar clouds disrupted the Oort comet cloud?]{}’. , 1548–1557. , M.: 2006, ‘[Non-Gravitational Effects in Long-Period Comets and the Size of the Oort Cloud]{}’. , 385–412. , J. and P. [Robutel]{}: 2001, ‘[High order symplectic integrators for perturbed Hamiltonian systems]{}’. , 39–62. , H. F., L. [Dones]{}, and M. J. [Duncan]{}: 2001, ‘[The Origin of Halley-Type Comets: Probing the Inner Oort Cloud]{}’. , 2253–2267. , H. F., M. J. [Duncan]{}, L. [Dones]{}, and B. J. [Gladman]{}: 2006, ‘[The scattered disk as a source of Halley-type comets]{}’. , 619–633. , J. J. and J. J. [Lissauer]{}: 2002, ‘[Characteristics and Frequency of Weak Stellar Impulses of the Oort Cloud]{}’. , 228–240. , J. J. and P. G. [Whitman]{}: 1989, ‘[The Galactic disk tidal field and the nonrandom distribution of observed Oort cloud comets]{}’. , 389–401. , O. A.: 2006, ‘[Dynamical evolution of Oort cloud comets to near-Earth space]{}’. In: A. [Milani]{}, G. Valsecchi, and Vokrouhlický (eds.): [*Near Earth Objects, our Celestial Neighbors: Opportunity and Risk*]{}, Vol. 236 of [*IAU Symposium*]{}. pp. 43–54. Mihalas, D. and J. Binney: 1981, [*Galactic Astronomy: Structure and Kinematics*]{}. San Francisco: Freeman. , J. H.: 1950, ‘[The structure of the cloud of comets surrounding the Solar System and a hypothesis concerning its origin]{}’. , 91–110. , F. and F. [Mignard]{}: 1985, ‘[Dynamical evolution of the Oort cloud. I - A Monte Carlo simulation. II - A theoretical approach]{}’. , 1–30. , H.: 1976, ‘[Stellar perturbations of orbits of long-period comets and their significance for cometary capture]{}’. , 92–105. , H., M. [Fouchard]{}, G. B. [Valsecchi]{}, and Ch. [Froeschl[é]{}]{}: 2005, ‘[Algorithms for Stellar Perturbation Computations on Oort Cloud Comets]{}’. , 411–434. , H., Ch. [Froeschl[é]{}]{}, Cl. [Froeschl[é]{}]{}, and G. B. [Valsecchi]{}: 2004, ‘[Stellar perturbations on the scattered disk]{}’. , 673–681. , P. and P. [Hut]{}: 1986, ‘[Dynamics of Cometary Showers]{}’. In: [*Lunar and Planetary Institute Conference Abstracts*]{}, Vol. 17 of [*Lunar and Planetary Institute Conference Abstracts*]{}. pp. 935–936. , P. R.: 1979, ‘[Physical and dynamical evolution of long-period comets]{}’. In: R. L. [Duncombe]{} (ed.): [*Dynamics of the Solar System*]{}, Vol. 81 of [*IAU Symposium*]{}. pp. 277–282. \[lastpage\] [^1]: Towards the end of our simulation the number of Oort Cloud comets has decreased in all three models but most in the combined one. We then have about $930\,000$, $840\,000$ and $760\,000$ comets in the tides-only, stars-only and combined models, respectively. This means that $\Delta N_C$ actually underestimates the extra contribution of the combined model. [^2]: The main reason for this discrepancy is that the Heisler papers considered injections into the loss cone – mainly by slight perturbations of $q$ across the limiting value $q_c\,=\,10$ AU – while we consider large jumps from $q\,>\,15$ AU into the observable region with $q\,<\,5$ AU. Interestingly, commented that the injection into orbits with $a \, \lta \, 20\,000$ AU and $q\,<\,2$ AU is indeed dominated by stellar perturbations.

ArXiv

--- abstract: 'We show higher interior regularity for the Westervelt equation with strong nonlinear damping term of the $q$-Laplace type. Secondly, we investigate an interface coupling problem for these models, which arise, e.g., in the context of medical applications of high intensity focused ultrasound in the treatment of kidney stones. We show that the solution to the coupled problem exhibits piecewise $H^2$ regularity in space, provided that the gradient of the acoustic pressure is essentially bounded in space and time on the whole domain. This result is of importance in numerical approximations of the present problem, as well as in gradient based algorithms for finding the optimal shape of the focusing acoustic lens in lithotripsy.' address: - | Insitut für Mathematik, Alpen-Adria-Universität Klagenfurt\ Universitätsstra[ß]{}e 65-57, 9020 Klagenfurt am Wörthersee, Austria - | Insitut für Mathematik, Alpen-Adria-Universität Klagenfurt\ Universitätsstra[ß]{}e 65-57, 9020 Klagenfurt am Wörthersee, Austria author: - 'Vanja Nikoli'' c' - Barbara Kaltenbacher title: On higher regularity for the Westervelt equation with strong nonlinear damping --- nonlinear acoustics, interface coupling, Westervelt’s equation, $q$-Laplace 35L05, 35L20 Introduction ============ High intensity focused ultrasound has numerous applications starting from the treatment of kidney and bladder stones, via thermo therapy and ultrasound cleaning to sonochemistry. Due to the nonlinear effects observed in the propagation of ultrasound in these cases, such as the appearance of sawtooth solutions, models of nonlinear acoustics and their rigourous mathematical treatment have become of great interest in recent years.\ One of the most popular models for the nonlinear propagation of ultrasound is the Westervelt equation $$\begin{aligned} \label{westervelt} (1-2ku)\ddot{u} -c^2 \Delta u - b\Delta \dot{u} = 2k(\dot{u})^2,\end{aligned}$$ expressed here in terms of the acoustic pressure $u$, where a dot denotes time differentiation, $b$ the diffusivity and $c$ the speed of sound, $k = \beta_a / \lambda$, $\lambda=\varrho c^2$ is the bulk modulus, $\varrho$ is the mass density, $\beta_a = 1 + B/(2A)$, and $B/A$ represents the parameter of nonlinearity. A detailed derivation of can be found in [@HamiltonBlackstock], [@manfred] and [@Westervelt].\ Westervelt’s equation is a quasilinear wave equation which can degenerate due to the factor $1-2ku$. This means that any analysis of this equation has to include bounding away from zero this term, i.e. finding an essential bound for $u$. That has been so far achieved by means of employing the Sobolev embedding $H^2(\Omega) \hookrightarrow L^{\infty}(\Omega)$ (cf. [@KL08], [@KLV10]), which implies that the solution of the Westervelt equation has to exhibit $H^2$ regularity in space. However, achieving $H^2$-regularity is too high of a demand in the case of coupling acoustic regions with different material parameters.\ To remedy this issue, Westervelt’s equation is considered with an added nonlinear damping term $$\begin{aligned} \label{westervelt_damp} (1-2ku)\ddot{u}-c^2\Delta u -\div(b((1-\delta) +\delta|\nabla \dot{u}|^{q-1})\nabla \dot{u}) =2k(\dot{u})^2,\end{aligned}$$ cf., [@BKR13], [@VN], where $\delta \in (0,1)$, $q \geq 1$, $q >d-1$, and $d \in \{1,2,3\}$ is the dimension of the spatial domain $\Omega\subseteq{{\mathbb{R}}}^d$ on which is considered. Since degeneracy is avoided with the help of the embedding $W^{1,q+1}(\Omega) \hookrightarrow L^\infty(\Omega)$, using this model allows to show existence of weak solutions with $W^{1,q+1}$ regularity in space, and in turn well-posedness of the acoustic-acoustic coupling problem.\ Considerations of the acoustic-acoustic coupling are motivated by lithotripsy where a silicone acoustic lens focuses the ultrasound traveling through a nonlinearly acoustic fluid to a kidney stone (see fig. 1). The interface coupling is modeled by the presence of spatially varying coefficients in the weak form of the equation (see [@BGT97] for the linear and [@BKR13] and [@VN] for the nonlinear case) as follows: $$\label{ModWest_coupled} \begin{cases} \text{Find} \ u \ \text{such that} \vspace{1.5mm}\\ \int_0^T \int_{\Omega} \{\frac{1}{\lambda(x)}(1-2k(x)u)\ddot{u} \phi+\frac{1}{\varrho(x)}\nabla u \cdot \nabla \phi + b(x)(1-\delta(x))\nabla \dot{u} \cdot \nabla \phi \vspace{1.5mm}\\ \quad \quad \quad +b (x)\delta(x)|\nabla \dot{u}|^{q-1} \nabla \dot{u} \cdot \nabla \phi-\frac{2k(x)}{\lambda(x)}(\dot{u})^2 \phi\} \, dx \, ds =0 \vspace{1.5mm}\\ \text{holds for all test functions} \ \phi \in \tilde{X}=L^2(0,T;W_0^{1,q+1}(\Omega)), \end{cases}$$ with $(u,\dot{u})\vert_{t=0}=(u_0,u_1)$. In this model $b$ stands for the quotient between the diffusivity and the bulk modulus, while the other coefficients maintain their meaning. The coefficients are allowed to jump only over the interface, i.e. the boundary of the lens. For notational brevity, we emphasized the space dependence of coefficients in , while omitting space and time dependence of $u$ in the notation.\ fig. 1: Schematic of a power source in lithotripsy\ based on the electromagnetic principle The first goal of the present paper is to show higher interior regularity results for solutions of . We will show that $u \in H^1(0,T;H_{loc}^2(\Omega))$ and $|\nabla \dot{u}|^{\frac{q-1}{2}}\nabla \dot{u} \in L^2(0,T;H^1_{loc}(\Omega))$. Although $q$-Laplace and parabolic $q$-Laplace equation have been extensively studied in the past (see [@lindqvist], [@U], [@Dibendetto], [@Friedman] and references given therein), regularity results in literature on hyperbolic equations with damping of the $q$-Laplace type are sparse and have so far been concerned with local and global well-posedness (see [@W], [@Gao], [@BKR13], [@VN]).\ Secondly, we will consider the coupled problem and show that the solution to is piecewise $H^2$ regular in space under the assumption that the gradient of the acoustic pressure remains essentially bounded in space and time. This result is crucial in future numerical approximations of the present problem, as well as in gradient based algorithms for finding the optimal shape of the focusing acoustic lens, where $H^2$ regularity of $u$ is needed in order to express the shape derivative in terms of integrals over the boundary of the lens, see [@NK15]. Overview -------- The paper is organized as follows. We begin in Section \[preliminaries\] by recalling certain basic results from the theory of finite differences, as well as some helpful inequalities which are needed later. In Section \[ModWest\_interior\_reg\] we prove higher interior regularity for the Westervelt equation . Section \[interior\_ModWestcoupled\] extends these results to the coupled problem. In Section \[boundary\], we will show that we can obtain higher regularity up to the boundary of the subdomains if the gradient of the acoustic pressure is essentially bounded in space and time and the subdomains are sufficiently regular. Preliminaries ============= In what is to follow, we will need to employ difference quotient approximations to weak derivatives. Assume that $\Omega \subset \mathbb{R}^d$, $d \in \{1,2,3\}$ is an open, connected set with Lipschitz boundary. Let $V \subset \subset \Omega$. $D_r^l$ will stand for the $r$-th difference quotient of size $l$ $$D_r^l u(x,t)=\frac{u(x+l e_r,t)-u(x,t)}{l}, \quad r \in [1,d],$$ for $x \in V$, $l \in \mathbb{R}$, $ 0<|l|< \frac{1}{2}\text{dist}(V,\partial \Omega)$. Then $D^lu:=(D_1^lu, \ldots,D_d^lu)$. We recall the integration by parts formula for difference quotients $$\int_{V} u \, D_r^l \varphi \, dx =-\int_{V } D^{-l}_r u \, \varphi \, dx,$$ where $\varphi \in C_c^\infty(V)$, $0<|l|<\frac{1}{2}\text{dist}(V,\partial \Omega)$, as well as the product rule $$D_r^l(\varphi u)=\varphi^l D_r^l u+uD_r^l \varphi,$$ with $\varphi^l(x,t):=\varphi(x+l e_r,t)$. We will also need the following result (cf. Theorem 3, Section 5, [@evans]): \[reg\_evans\] Assume $0 \leq q < \infty$ and $u \in W^{1,q+1}(\Omega)$ Then for each $V \subset \subset \Omega$ $$\|D^lu\|_{L^{q+1}(V)} \leq C\|\nabla u\|_{L^{q+1}(\Omega)},$$ for some constant $C$ and all $0<|l|<\frac{1}{2}\text{dist}(V,\partial \Omega)$.\ Assume $0 < q < \infty$, $u \in L^{q+1}(\Omega)$, and there exists a constant $C$ such that $\|D^lu\|_{L^{q+1}(V)} \leq C$ for all $0<|l|<\frac{1}{2}\text{dist}(V,\partial \Omega)$. Then $$u \in W^{1,q+1}(V), \ \text{with} \ \|\nabla u\|_{L^{q+1}(V)} \leq C.$$ Essential inequalities {#essential-inequalities .unnumbered} ---------------------- Before proceeding further, let us also recall several useful inequalities that we will need when handling the $q$-Laplace damping term. They can be found in Chapter 10, [@lindqvist] and Appendix, [@LiuYan]. From now on, $C_q$ will be used to denote a generic constant depending only on $q$. For any $x,y \in \mathbb{R}^d$ it holds $$\begin{aligned} \label{ineq1} ||x|^{q-1}x-|y|^{q-1}y|\leq C_q |x-y|(|x|+|y|)^{q-1}, \ q>0,\end{aligned}$$ $$\begin{aligned} \label{ineq3} ||x|^{q-1}x-|y|^{q-1}y|\geq 2^{-1}|x-y|^{2}(|x|+|y|)^{q-1} \geq 2^{1-q} |x-y|^{q+1} \geq 0, \ q \geq 1.\end{aligned}$$ $$\begin{aligned} \label{ineq4} \frac{4}{(q+1)^2}||x|^{\frac{q-1}{2}}x-|y|^{\frac{q-1}{2}}y|^2\leq (|x|^{q-1}x-|y|^{q-1}y)\cdot(x-y), \ q\geq 1,\end{aligned}$$ $$\begin{aligned} \label{ineq5} ||x|^{q-1}x-|y|^{q-1}y| \leq q(|x|^{\frac{q-1}{2}}+|y|^{\frac{q-1}{2}}) \left| |x|^{\frac{q-1}{2}}-|y|^{\frac{q-1}{2}}\right|, \ q \geq 1.\end{aligned}$$ We will also need Young’s inequality (see for instance Appendix B, [@evans]) in the form $$\begin{aligned} \label{Young} \quad \quad \quad |xy| \leq \varepsilon |x|^r+C(\varepsilon, r) |y|^{\frac{r}{r-1}} \quad (\varepsilon >0, \ 1<r< \infty),\end{aligned}$$ with $C(\varepsilon, r)=(r-1)r^{\frac{r}{r-1}}\varepsilon ^{-\frac{1}{1-r}}$. Notation {#notation .unnumbered} -------- By $C_{X,Y}^\Omega$ we will denote the norm of the embedding operator $X(\Omega)\to Y(\Omega)$ between two function spaces over the domain $\Omega$. Interior regularity for the Westervelt equation with strong nonlinear damping {#ModWest_interior_reg} ============================================================================= In this section, we will establish higher interior regularity for the equation with constant coeffcients. Let us first consider the following Dirichlet problem: $$\label{ModWest_Dirichlet} \begin{cases} (1-2ku)\ddot{u}-c^2\Delta u-b\,\text{div}\,(((1-\delta) +\delta|\nabla \dot{u}|^{q-1})\nabla \dot{u}) =2k(\dot{u})^2 \quad \text{ in } \Omega \times (0,T] , \\ u|_{\partial \Omega} =0 \quad \text{ for } t \in (0,T], \\ (u,\dot{u})|_{t=0}=(u_0, u_1) \,\, \text{ in } \Omega, \\ \end{cases}$$ with the following assumptions on the coefficients and the exponent $q$: $$\begin{aligned} \label{coeff_ModWest_Dirichlet} c^2, \ b>0,\ \delta \in (0,1), \ k\in \mathbb{R}, \ q>d-1, \ q \geq 1.\end{aligned}$$ The weak formulation reads as $$\label{ModWest_Dirichlet_weak} \begin{cases} \int_0^T \int_{\Omega} \{(1-2ku)\ddot{u} \phi+c^2\nabla u \cdot \nabla \phi + b(1-\delta)\nabla \dot{u} \cdot \nabla \phi \vspace{1.5mm}\\ \quad \quad \quad +b\delta|\nabla \dot{u}|^{q-1} \nabla \dot{u} \cdot \nabla \phi-2k(\dot{u})^2 \phi\} \, dx \, ds =0 \vspace{1.5mm}\\ \text{holds for all test functions} \ \phi \in \tilde{X}=L^2(0,T;W_0^{1,q+1}(\Omega)), \end{cases}$$ with $(u,\dot{u})=(u_0,u_1)$. We recall the following well-posedness result (cf. Theorem 2.3, [@BKR13]): \[ModWest\_Dirichlet\_localw\] (Local well-posedness) Let assumptions hold. For any $T>0$ there is a $\kappa_T>0$ such that for all $u_0,u_1\in W_0^{1,q+1}(\Omega)$ with $$\begin{aligned} |u_1|_{L^2(\Omega)}^2 +|\nabla u_0|_{L^2(\Omega)}^2 + |\nabla u_1|_{L^2(\Omega)}^2 + |\nabla u_1|_{L^{q+1}(\Omega)}^{q+1} +|\nabla u_0|_{L^{q+1}(\Omega)}^2\leq \kappa_T^2 \end{aligned}$$ there exists a weak solution $u\in \cW \subset X$ of , where $X=H^2(0,T;L^2(\Omega))\cap C^{0,1}(0,T;W_0^{1,q+1}(\Omega))$, and $$\begin{aligned} \label{defcW} \cW =\{v\in X &:& \|\ddot{v}\|_{L^2(0,T;L_2(\Omega))}\leq \bar{m} \wedge \|\nabla \dot{v}\|_{L^\infty(0,T;L^2(\Omega))}\leq \bar{m}\nonumber\\ && \wedge \|\nabla \dot{v}\|_{L^{q+1}(0,T;L^{q+1}(\Omega))}\leq \bar{M} \wedge (v,\dot{v})=(u_0,u_1)\}\end{aligned}$$ with $$\label{smallnessMbar} 2|k| C_{W_0^{1,q+1},L^\infty}^\Omega (\kappa_T+ T^{\frac{q}{q+1}} \bar{M}) <1$$ and $\bar{m}$ sufficiently small, and $u$ is unique in $\cW $. In [@BKR13], the issue of possible degeneracy of the Westervelt equation due to the factor $1-2ku$ is resolved by means of the embedding $W_0^{1,q+1}(\Omega) \hookrightarrow L^{\infty}(\Omega)$, valid for $q>d-1$, and the following estimate $$\begin{aligned} |u(x,t)|\leq&\, C^{\Omega}_{W_0^{1,q+1},L^{\infty}}|\nabla u(t)|_{L^{q+1}(\Omega)}\\ \leq& \, C^{\Omega}_{W_0^{1,q+1},L^{\infty}}|\nabla u_0+\int_0^t \nabla \dot{u} \, ds \,|_{L^{q+1}(\Omega)}\\ \leq&\, C^{\Omega}_{W_0^{1,q+1},L^{\infty}}\Bigl(|\nabla u_0|_{L^{q+1}(\Omega)}+\Bigl(t^q \int_0^t \int_\Omega |\nabla \dot{u}(y,s)|^{q+1}\, dy \, ds \Bigr)^{\frac{1}{q+1}}\Bigr),\end{aligned}$$ which leads to the bound $$\label{degeneracy} \begin{aligned} &1-a_0< 1-2ku < 1+a_0, \\ & a_0:= 2|k|C^{\Omega}_{W_0^{1,q+1},L^{\infty}}(|\nabla u_0|_{L^{q+1}(\Omega)}+T^{\frac{q}{q+1}}\|\nabla \dot{u}\|_{L^{q+1}(0,T;L^{q+1}(\Omega))}). \end{aligned}$$ Due to the embedding $W^{1,q+1}(\Omega) \hookrightarrow C^{0,1-\frac{d}{q+1}}(\overline{\Omega})$, we also know that $u$ is H" older continuous in space, i.e. $u \in C^{0,1}(0,T;C^{0,1-\frac{d}{q+1}}(\overline{\Omega}))$. Higher interior regularity -------------------------- We will establish higher interior regularity by following the difference-quotient approach (see, for instance, Theorem 4, Section 6.3.2, [@evans]). Let us denote $$\begin{aligned} F=|\nabla \dot{u}|^{\frac{q-1}{2}}\nabla \dot{u}, \quad F^l=|\nabla \dot{u}^l|^{\frac{q-1}{2}}\nabla \dot{u}^l.\end{aligned}$$ As a by-product of the following proof we will also obtain $F \in H^{1}_{loc}(\Omega)$ by adapting the idea of Bojarski and Iwaniec for $q$-harmonic functions (see, for instance, Section 4, [@lindqvist]) to our model. \[thm:interior\_reg\] (Higher interior regularity) Let assumptions hold true, $u_0 \in H^2(\Omega) \cap W_0^{1,q+1}(\Omega)$, $u_1 \in W_0^{1,q+1}(\Omega)$, and let $u$ be the weak solution of . Then $u \in H^1(0,T;H^2_{loc}(\Omega))$ and $|\nabla \dot{u}|^{\frac{q-1}{2}}\nabla \dot{u} \in L^2(0,T;H^1_{loc}(\Omega))$. Choose any open set $V \subset \subset \Omega$ and an open set $W$ such that $V \subset \subset W \subset \subset \Omega$. We then introduce a smooth cut-off function $\zeta$ such that $$\begin{aligned} \begin{cases} \zeta=1 \ \text{on} \ V, \ \zeta=0 \ \text{on} \ \Omega \setminus W, \\ 0 \leq \zeta \leq 1. \end{cases}\end{aligned}$$ Let $|l|>0$ be small and choose $r \in \{1,\ldots, d\}$. We are then allowed to use $$\phi:=-D_r^{-l}(\zeta^2 D_r^l \dot{u}){\chi}_{[0,t)}, \ \ t \in[0,T]$$ as a test function in , which results in $$\begin{aligned} \label{interior:est1} &\frac{1}{2}\Bigl[\int_{\Omega_i}(1-2ku^l)(\zeta D_r^l \dot{u})^2\, dx\Bigr]_0^t+\frac{1}{2}c^2\Bigl[\int_{\Omega}|\zeta D_r^l \nabla u|^2\, dx\Bigr]_0^t \nonumber\\ &+b(1-\delta)\int_0^T \int_{\Omega_i}|\zeta D_r^l \nabla \dot{u}|^2\, dx \, ds +\frac{4}{(q+1)^2}b \delta\int_0^t \int_{\Omega} |\zeta D_r^l F|^{2} \, dx \, ds\nonumber\\ \leq& \,2k\int_0^t \int_{\Omega}D_r^l u \, \ddot{u} \, \zeta^2 D_r^l \dot{u} \, dx \, ds -c^2\int_0^t \int_{\Omega} \zeta \nabla \zeta \cdot D_r^l \nabla u\, D_r^l \dot{u} \, dx \, ds \nonumber\\ &-2b(1-\delta)\int_0^t \int_{\Omega} \zeta \nabla \zeta \cdot D_r^l \nabla \dot{u}\, D_r^l \dot{u}_i \, dx \, ds \\ &-2b\delta_i \int_0^t \int_{\Omega} D_r^l(|\nabla \dot{u}|^{q-1}\nabla \dot{u}) \cdot \zeta \nabla \zeta\, D_r^l \dot{u} \, dx \, ds +k\int_0^t \int_{\Omega_i}(\dot{u}^l+2\dot{u})(\zeta D_r^l \dot{u})^2 \, dx \, ds. \nonumber\end{aligned}$$ Here we have made use of the estimate $$\begin{aligned} \int_0^t \int_{\Omega}\zeta^2 D_r^l(|\nabla \dot{u}|^{q-1}\nabla \dot{u}) D_r^l\nabla \dot{u} \, dx \, ds \geq \frac{4}{(q+1)^2}\int_0^t \int_{\Omega} \frac{1}{l^2}\zeta^2||\nabla \dot{u}^l|^{\frac{q-1}{2}}\nabla \dot{u}^l-|\nabla \dot{u}|^{\frac{q-1}{2}}\nabla \dot{u}|^{2}\, dx \, ds,\end{aligned}$$ which follows from . Next, we estimate the terms on the right hand side containing $\zeta \nabla \zeta$. We have $$\begin{aligned} & - c^2\int_0^t \int_{\Omega} \zeta \nabla \zeta \cdot D_r^l \nabla u D_r^l \dot{u} \, dx \, ds-2b(1-\delta)\int_0^t \int_{\Omega} \zeta \nabla \zeta \cdot D_r^l \nabla \dot{u} D_r^l \dot{u} \, dx \, ds \\ \leq& \, C \int_0^T \int_{\Omega} \zeta (|D_r^l \nabla u| + |D_r^l \nabla \dot{u}| )|\nabla \zeta D_r^l \dot{u}|\, dx \, ds \\ \leq& \, \varepsilon T\|\zeta D_r^l \nabla u\|^2_{L^\infty(0,T;L^2(\Omega))}+ \varepsilon\|\zeta D_r^l \nabla \dot{u}\|^2_{L^2(0,T;L^2(\Omega))}+ \frac{C}{\varepsilon}\|\nabla \dot{u}\|^2_{L^2(0,T;L^2(\Omega))},\end{aligned}$$ where $C$ depends on $c, b, \delta$ and $|\nabla \zeta|_{L^\infty(W)}$ and we have used Lemma \[reg\_evans\], (a). By employing estimate and H" older’s inequality we obtain $$\begin{aligned} &-2b\delta \int_0^t \int_{\Omega} D_r^l(|\nabla \dot{u}|^{q-1}\nabla \dot{u})\cdot \zeta \nabla \zeta \, D_r^l \dot{u} \, dx \, ds \\ \leq& \, 2 b \delta \int_0^T \int_{\Omega} \frac{1}{l}|D_r^l \dot{u}| \left| |\nabla \dot{u}^l|^{q-1}\nabla \dot{u}^l-|\nabla \dot{u}|^{q-1}\nabla \dot{u} \right| \zeta|\nabla \zeta|\, dx \, ds \\ \leq&\, 2 q b \delta \int_0^T \int_{\Omega}|\nabla \zeta D_r^l \dot{u}|(|\nabla \dot{u}^l|^{\frac{q-1}{2}}+|\nabla \dot{u}|^{\frac{q-1}{2}})|\zeta D_r^l F| \, dx \, ds \\ \leq&\, 2 q b \delta \Bigl\{ \int_0^T \int_{\Omega}|\nabla \zeta D_r^l \dot{u}|^{q+1}\, dx \, ds \Bigr \}^{\frac{1}{q+1}}\Bigl\{ \int_0^T \int_{\text{supp} \zeta} (|\nabla \dot{u}^l|^{\frac{q-1}{2}}+|\nabla \dot{u}|^{\frac{q-1}{2}})^{\frac{2(q+1)}{q-1}}\, dx\, ds\Bigr\}^{\frac{q-1}{2(q+1)}}\\ & \quad \quad \quad \times \Bigl\{\int_0^T \int_{\Omega}|\zeta D_r^l F|^2 \, dx \, ds \Bigr\}^{1/2}.\end{aligned}$$ The second integral can be majorized with the help of Minkowski’s inequality by $$\begin{aligned} &\Bigl\{ \int_0^T \int_{\text{supp} \zeta} (|\nabla \dot{u}^l|^{\frac{q-1}{2}}+|\nabla \dot{u}|^{\frac{q-1}{2}})^{\frac{2(q+1)}{q-1}}\, dx\, ds\Bigr\}^{\frac{q-1}{2(q+1)}} \\ \leq&\, \Bigl\{ \int_0^T \int_{\text{supp} \zeta} (|\nabla \dot{u}^l|^{q+1} \, dx \, ds\Bigr\}^{\frac{q-1}{2(q+1)}} + \Bigl\{ \int_0^T \int_{\text{supp} \zeta} (|\nabla \dot{u}|^{q+1} \, dx \, ds\Bigr\}^{\frac{q-1}{2(q+1)}} \\ \leq&\, 2 \Bigl\{ \int_0^T \int_{\Omega} (|\nabla \dot{u}|^{q+1} \, dx \, ds\Bigr\}^{\frac{q-1}{2(q+1)}}=2\|\nabla \dot{u}\|^{\frac{q-1}{2}}_{L^{q+1}(0,T;L^{q+1}(\Omega))},\end{aligned}$$ for small $|l|$. Utilizing Young’s inequality then yields $$\begin{aligned} &-2b\delta \int_0^t \int_{\Omega_i} D_r^l(|\nabla \dot{u}|^{q-1}\nabla \dot{u})\cdot \zeta \nabla \zeta \, D_r^l \dot{u} \, dx \, ds \\ \leq&\, b\delta \varepsilon \int_0^T \int_{\Omega}|\zeta D_r^l F|^2 \, dx \, ds +C(\|\nabla \dot{u}\|^{q+1}_{L^{q+1}(0,T;L^{q+1}(W))}+\|D_r^l \dot{u}\|^{q+1}_{L^{q+1}(0,T;L^{q+1}(W))}),\end{aligned}$$ where $C>0$ depends on $b_i$, $\delta_i$, $\varepsilon$, $q$ and $|\nabla \zeta|_{L^{\infty}(W)}$. Note that the first term in the last line can be absorbed by the $b_i\delta_i$- term on the left hand side in for sufficiently small $\varepsilon>0$. The two remaining terms on the right hand side in can be estimated as follows $$\begin{aligned} &k\int_0^t \int_{\Omega}(\dot{u}^l+2\dot{u})\zeta^2(D_r^l \dot{u})^2 \, dx \, ds+2k\int_0^t \int_{\Omega}\zeta^2 \ddot{u}\, D_r^l u \, D_r^l \dot{u} \, dx \, ds\\ \leq& \,C\Bigl(\|\dot{u}\|_{L^\infty(0,T;L^\infty(\Omega))}\|\zeta D_r^l \dot{u}\|^2_{L^2(0,T;L^2(W))}\\ &+(C_{H^1,L^4}^{\Omega})^2\|\ddot{u}\|_{L^2(0,T;L^2(\Omega))}\|\zeta D_r^l u\|_{L^\infty(0,T;H^1(\Omega))}\|\zeta D_r^l \dot{u}\|_{L^2(0,T;H^1(\Omega))}\Bigr) \\ \leq& C\Bigl(\|\dot{u} \|_{L^\infty(0,T;L^\infty(\Omega))}\|D_r^l \dot{u}\|^2_{L^2(0,T;L^2(\Omega))}\\ &+\bar{m}^2\frac{1}{4\varepsilon}(\|D_r^l u\|^2_{L^\infty(0,T;L^2(W))}+\|\zeta D_r^l \nabla u\|^2_{L^\infty(0,T;L^2(\Omega))}) \\ &+\varepsilon\|D_r^l \dot{u}\|^2_{L^2(0,T;L^2(W))}+\varepsilon\|\zeta D_r^l \nabla \dot{u}\|^2_{L^2(0,T;L^2(\Omega))}\Bigr)\, .\end{aligned}$$ Altogether, for sufficiently small $\varepsilon>0$ and $\bar{m}$, we can achieve that $$\begin{aligned} &\|\zeta D_r^l \dot{u}\|^2_{L^\infty(0,T;L^2(\Omega))}+\|\zeta D_r^l \nabla u_i\|^2_{L^\infty(0,T;L^2(\Omega_i))}+\|\zeta D_r^l \nabla \dot{u}\|^2_{L^2(0,T;L^2(\Omega))}\\ &+\|\zeta D_r^l F\|^{q+1}_{L^{2}(0,T;L^{2}(\Omega))}\\ \leq& \, C (\|\nabla \dot{u}\|^{q+1}_{L^{\infty}(0,T;L^\infty(\Omega))}+\|D_r^l \dot{u}\|^{q+1}_{L^{q+1}(0,T;L^{q+1}(W))}+\|\dot{u}\|_{L^\infty(0,T;L^\infty(\Omega))}\|D_r^l \dot{u}\|^2_{L^2(0,T;L^2(W))}\\ &+\|D_r^l u\|^2_{L^\infty(0,T;L^2(W))}+\|D_r^l \dot{u}\|^2_{L^2(0,T;L^2(W))}+|D_r^l u_1|^2_{L^2(\Omega)}+|D_r^l \nabla u_0|^2_{L^2(\Omega)}).\end{aligned}$$ By remembering the definition of $\zeta$ and Lemma \[reg\_evans\], we finally arrive at $$\begin{aligned} &\| D_r^l \dot{u}_i\|^2_{L^\infty(0,T;L^2(V))}+\| D_r^l \nabla u\|^2_{L^\infty(0,T;L^2(V))} +\|D_r^l \nabla \dot{u}\|^2_{L^2(0,T;L^2(V))}+\|D_r^l F\|^{2}_{L^{2}(0,T;L^{2}(V))}\\ \leq& \, C (\|\nabla \dot{u}\|^{q+1}_{L^{\infty}(0,T;L^\infty(\Omega))}+(1+\|\dot{u} \|_{L^\infty(0,T;L^\infty(\Omega))})\|\nabla \dot{u}\|^2_{L^2(0,T;L^2(\Omega))}+\|\nabla u\|^2_{L^\infty(0,T;L^2(\Omega))}\\ &+|\nabla u_1|^2_{L^2(\Omega)}+|u_0|^2_{H^2(\Omega)}),\end{aligned}$$ for $r \in [1,d]$, sufficiently small $|l|>0 $ and sufficiently large $C>0$ which does not depend on $l$. By employing Lemma \[reg\_evans\], we can conclude that $u \in H^{1}(0,T;H^{2}_{loc}(\Omega))$ and $|\nabla \dot{u}|^{\frac{q-1}{2}}\nabla \dot{u} \in L^2(0,T;H^1_{loc}(\Omega))$. As a simple consequence of the previous proposition, we can obtain H" older continuity of $u$ (see Section 4, [@lindqvist]). Indeed, since $F \in L^2(0,T;H^1_{loc}(\Omega))$ and $d \in \{1,2,3\}$, due to Sobolev’s embedding theorem we have that $F \in L^2(0,T;L^{6}_{loc}(\Omega))$. This implies that $u \in W^{1,q+1}(0,T;W^{1,3(q+1)}_{loc}(\Omega))$. We can than conclude that $u \in W^{1,q+1}(0,T;C^{0,\alpha}_{loc}(\Omega))$, where $\alpha=1-\frac{d}{3(q+1)}$.\ When $d \in \{1,2\}$ we can do even better. According to Sobolev’s embedding theorem, $u \in W^{1,q+1}(0,T;C_{loc}^{1,\frac{1}{2}}(\Omega))$ if $d=1$, and $u \in W^{1,q+1}(0,T;C_{loc}^{0,\gamma}(\Omega))$ if $d=2$, where $\gamma \in (0,1)$. Altogether, we have Let the assumptions of Theorem \[thm:interior\_reg\] hold true. Then $$\begin{aligned} \label{H_reg} u \in \begin{cases} W^{1,q+1}(0,T;C^{0,1-\frac{1}{q+1}}_{\text{loc}}(\Omega)) \ \text{if} \ d=3, \vspace{1mm} \\ W^{1,q+1}(0,T;C^{0,\gamma}_{\text{loc}}(\Omega)) \ \text{with} \ \gamma \in (0,1), \ \text{if} \ d=2, \vspace{1mm} \\ W^{1,q+1}(0,T;C^{1,\frac{1}{2}}_{\text{loc}}(\Omega)) \ \text{if} \ d=1. \end{cases}\end{aligned}$$ Neumann problem for the Westervelt equation ------------------------------------------- Let us also consider the Neumann problem for the Westervelt equation with strong nonlinear damping: $$\begin{aligned} \label{ModWest_Neumann} \begin{cases} (1-2ku)\ddot{u}-c^2\Delta u-b\,\text{div}(((1-\delta) +\delta|\nabla \dot{u}|^{q-1})\nabla \dot{u}) =2k(\dot{u})^2 \, \text{ in } \Omega \times (0,T], \vspace{2mm} \\ c^2 \frac{\partial u}{\partial n}+b((1-\delta)+\delta|\nabla \dot{u}|^{q-1})\frac{\partial \dot{u}}{\partial n}=g \ \ \text{on} \ \partial\Omega \times (0,T], \vspace{2mm} \\ (u,\dot{u})\vert_{t=0}=(u_0, u_1) \ \ \text{on} \ \overline{\Omega}, ,\\ \end{cases}\end{aligned}$$ with the same assumptions on coefficients. Problem is locally well-posed thanks to the following result (cf. Theorem 2.5, [@VN]): \[thm:W1\] Let $g \in L^{\infty }(0,T;W^{-\frac{q}{q+1},\frac{q+1}{q}}(\partial \Omega))$, $\dot{g} \in L^{\frac{q+1}{q}}(0,T;W^{-\frac{q}{q+1},\frac{q+1}{q}}(\partial \Omega))$, and $u_0, u_1 \in W^{1,q+1}(\Omega)$. For sufficiently small initial and boundary data, final time $T$ and $\bar{m}$ and $\bar{M}$ there exists a unique weak solution $u \in \cW \subset X$ of , where $X=H^2(0,T;L^2(\Omega))\cap C^{0,1}(0,T;W^{1,q+1}(\Omega))$, and $$\label{defcW1} \begin{split} \cW =\{v\in X :& \ \|\ddot{v}\|_{L^2(0,T;L^2(\Omega))}\leq \bar{m} \wedge \| \dot{v}\|_{L^{\infty}(0,T;H^1(\Omega))}\leq \bar{m}\\ & \wedge \| \nabla \dot{v}\|_{L^{q+1}(0,T;L^{q+1}(\Omega))}\leq \bar{M} \}. \end{split}$$ By inspecting the proof of Proposition \[thm:interior\_reg\], we immediately obtain higher interior regularity result for the present model, since the cut-off function used in the proof vanishes near the boundary: \[interior\_ModWest\_Neumann\] Let the assumptions hold true, $u_0 \in H^2(\Omega) \cap W^{1,q+1}(\Omega)$, $u_1 \in W^{1,q+1}(\Omega)$, and let $u$ be the weak solution of . Then $u \in H^1(0,T;H^2_{loc}(\Omega))$ and $|\nabla \dot{u}|^{\frac{q-1}{2}}\nabla \dot{u} \in L^2(0,T;H^1_{loc}(\Omega))$. Moreover, holds. Interior regularity for the coupled problem {#interior_ModWestcoupled} =========================================== Let us now assume that $\Omega \subset \mathbb{R}^d$, $d \in \{1,2,3\}$, is a bounded domain with Lipschitz boundary $\partial \Omega$, and $\Omega_{+}$ a subdomain, representing the lens, such that $\bar{\Omega}_{+} \subset \Omega$ and $\Omega_{+}$ has Lipschitz boundary $\partial \Omega_+=\Gamma$. \ Lens $\Omega_+$ and fluid $\Omega_-$ regions\ We denote by $\Omega_{-}=\Omega \setminus \bar{\Omega}_{+}$ the part of the domain representing the fluid region. We then have $\partial \Omega_-=\Gamma \cup \partial \Omega$.\ $n_{+}$, $n_{-}$ will stand for the unit outer normals to lens $\Omega_{+}$ and fluid region $\Omega_{-}$. Restrictions of a function $v$ to $\Omega_{+,-}$ will be denoted by $v_{+}$, $v _{-}$ and $\llbracket v \rrbracket:=v_+-v_-$ will denote the jump over $\Gamma$.\ Note that the assumption on the regularity of the subdomains will be strengthened to $C^{1,1}$ when showing higher regularity up to the boundary of the subdomains.\ \ The coefficients in will only be allowed to jump over the interface $\Gamma$: $$\label{coeff} \begin{cases} b,\varrho, \lambda, \delta, k \in L^{\infty}(\Omega), \\ b_i:=b\vert_{\Omega_i}, \varrho_i:=\varrho\vert_{\Omega_i}, \lambda_i:=\lambda\vert_{\Omega_i} >0, \ \delta_i:=\delta\vert_{\Omega_i} \in (0,1), \ k_i:=k\vert_{\Omega_i} \in \mathbb{R} \quad \text{for} \ i \in \{+,-\}. \end{cases}$$ We assume that $q \geq 1$, $q>d-1$. The strong formulation of reads as follows: $$\begin{aligned} \label{ModWest_coupled_Dirichlet} \begin{cases} \frac{1}{\lambda(x)}(1-2k(x)u)\ddot{u}-\div(\frac{1}{\varrho(x)}\nabla u) -\text{div}\Bigl(b(x)((1-\delta(x))+\delta(x)|\nabla \dot{u}|^{q-1})\nabla \dot{u}\Bigr) \vspace{1.5mm} \\ =\frac{2k(x)}{\lambda(x)}(\dot{u})^2 \quad \text{ in } \Omega_{+} \cup \Omega_{-}, \vspace{1.5mm} \\ \llbracket u \rrbracket =0 \quad \text{on} \ \Gamma=\partial \Omega_+, \vspace{1.5mm} \\ \Bigl \llbracket \frac{1}{\varrho} \frac{\partial u}{\partial n_{+}}+b(1-\delta)\frac{\partial \dot{u}}{\partial n_{+}}+b\delta|\nabla \dot{u}|^{q-1}\frac{\partial \dot{u}}{\partial n_{+}}\Bigr \rrbracket =0 \quad \text{on} \ \Gamma=\partial \Omega_+, \vspace{1mm} \\ u=0 \quad \text{on} \ \partial \Omega, \vspace{1mm}\\ (u,\dot{u})\vert_{t=0}=(u_0,u_1). \end{cases}\end{aligned}$$ This model was studied (in an equivalent one domain formulation) in [@BKR13]. We will utilize the following well-posedness result (cf. Theorem 2.3 and Corollary 4.1, [@BKR13]): (Local well-posedness)\[aa\_corollary\] Let $q>d-1$, $q \geq 1$ and the assumptions hold. For any $T>0$ there is a $\kappa_T>0$ such that for all $u_0, u_1 \in W_0^{1,q+1}(\Omega)$ with $$\begin{aligned} |u_1|^2_{L^2(\Omega)}+|\nabla u_0|^2_{L^2(\Omega)}+|\nabla u_1|^2_{L^2(\Omega)}+|\nabla u_0|^2_{L^{q+1}(\Omega)}+|\nabla u_1|^{q+1}_{L^{q+1}(\Omega)} \leq \kappa_T^2\end{aligned}$$ there exists a unique solution $u\in \cW \subset X$ of , where $X=H^2(0,T;L^2(\Omega))\cap C^{0,1}(0,T;W_0^{1,q+1}(\Omega))$, and $$\begin{aligned} \cW =\{v\in X &:& \|\ddot{v}\|_{L^2(0,T;L^2(\Omega))}\leq \bar{m} \wedge \|\nabla \dot{v}\|_{L^\infty(0,T;L^2(\Omega))}\leq \bar{m} \\ && \wedge \|\nabla \dot{v}\|_{L^{q+1}(0,T;L^{q+1}(\Omega))}\leq \bar{M} \wedge (v,\dot{v})=(u_0,u_1) \},\nonumber \end{aligned}$$ with $$2\overline{k}C_{W_0^{1,q+1},L^\infty}^\Omega(\kappa_T+ T^{\frac{q}{q+1}} \bar{M}) <1,$$ and $\bar{m}$ sufficiently small. For simplicity of exposition, higher interior and later boundary regularity will be obtained under the assumption that the coefficients in are piecewise costant functions, i.e. $$\label{coeff_const} \begin{cases} b_i:=b\vert_{\Omega_i},\varrho_i:=\varrho\vert_{\Omega_i}, \lambda_i:=\lambda\vert_{\Omega_i}, \delta_i:=\delta\vert_{\Omega_i}, k_i:=k\vert_{\Omega_i} \ \text{are constants}, \\ b_i, \varrho_i, \lambda_i >0, \quad \delta_i \in (0,1), \quad k_i \in \mathbb{R} \quad \text{for} \ i \in \{+,-\}. \end{cases}$$ We denote $\underline{\omega}=\min\{|\omega_+|,|\omega_-|\}$, $\overline{\omega}=\max\{|\omega_+|,|\omega_-|\}$, where $\omega \in \{b, \varrho, \lambda, \delta, k\}$. The proof of Theorem \[thm:interior\_reg\] can be carried over in a straightforward manner to the coupled problem to show higher interior regularity within each of the subdomains: \[interior\_reg\_coupled\] Assume that $q\geq 1$, $q>d-1$, $u_0\vert_{\Omega_i} \in H^2(\Omega_i)$, $i\in\{+,-\}$, $u_0,u_1 \in W_0^{1,q+1}(\Omega)$ and assumptions on the coefficients hold true. Let $u$ be the weak solution of . Then $u_i \in H^1(0,T;H^2_{loc}(\Omega_i))$ and $|\nabla \dot{u}_i|^{\frac{q-1}{2}}\nabla \dot{u}_i \in L^2(0,T;H^1_{loc}(\Omega_i))$, $i \in \{+,-\}$. Moreover, holds with $u$ replaced by $u_i$ and $\Omega$ by $\Omega_i$, $i \in \{+,-\}$. Neumann problem for the coupled system -------------------------------------- Let us also consider the coupled problem with Neumann boundary conditions on the outer boundary of the fluid subdomain: $$\begin{aligned} \label{coupled_Neumann} \begin{cases} \frac{1}{\lambda(x)}(1-2k(x)u)\ddot{u}-\div(\frac{1}{\varrho(x)}\nabla u) -\text{div}(b(x)((1-\delta(x))+\delta(x)|\nabla \dot{u}|^{q-1})\nabla \dot{u}) \vspace{1.5mm} \\ =\frac{2k(x)}{\lambda(x)}(\dot{u})^2 \quad \text{ in } \Omega_{+} \cup \Omega_{-}, \vspace{1.5mm} \\ [u]=0 \quad \text{on} \ \Gamma=\partial \Omega_+, \vspace{1.5mm} \\ \Bigl[\frac{1}{\varrho} \frac{\partial u}{\partial n_{+}}+b(1-\delta)\frac{\partial \dot{u}}{\partial n_{+}}+b\delta|\nabla \dot{u}|^{q-1}\frac{\partial \dot{u}}{\partial n_{+}}\Bigr]=0 \quad \text{on} \ \Gamma=\partial \Omega_+, \vspace{1.5mm} \\ \frac{1}{\varrho} \frac{\partial u}{\partial n_{+}}+b(1-\delta)\frac{\partial \dot{u}}{\partial n_{+}}+b\delta|\nabla \dot{u}|^{q-1}\frac{\partial \dot{u}}{\partial n_{+}}=g \quad \text{on} \ \partial \Omega, \vspace{1.5mm}\\ (u,\dot{u})\vert_{t=0}=(u_0,u_1). \end{cases}\end{aligned}$$ We will show that the higher regularity result is valid for this model as well. The weak form of the problem is given as follows: $$\begin{aligned} &\int_0^T \int_\Omega \Bigl\{\frac{1}{\lambda}(1-2ku) \ddot{u}\phi + \frac{1}{\varrho} \nabla u \cdot \nabla \phi + b((1-\delta) +\delta |\nabla \dot{u}|^{q-1})\nabla \dot{u} \cdot \nabla \phi - \frac{2k}{\lambda}(\dot{u})^2 \phi \Bigr\} \, dx \, ds \\ =& \, \int_0^T \int_{\partial \Omega}g \phi \, dx \, ds, \end{aligned}$$ for all $\phi\in L^2(0,T;W^{1,q+1}(\Omega))$, with initial conditions $(u_{0},u_{1})$. We recall the following well-posedness result (cf. [@VN]): \[prop:Neumann\] Let assumptions hold. Let $q>d-1$, $q \geq 1$, $g \in L^{\infty }(0,T;W^{-\frac{q}{q+1},\frac{q+1}{q}}(\partial \Omega))$, $\dot{g} \in L^{\frac{q+1}{q}}(0,T;W^{-\frac{q}{q+1},\frac{q+1}{q}}(\partial \Omega))$, and $u_0, u_1 \in W^{1,q+1}(\Omega)$. For sufficiently small initial and boundary data and final time $T$, there exists a unique weak solution $u \in \cW \subset X$ of , where $X= H^2(0,T;L^2(\Omega)) \cap C^{0,1}(0,T;W^{1,q+1}(\Omega))$ and $$\begin{split} \cW =\{v\in X :& \ \|\ddot{v}\|_{L^2(0,T;L^2(\Omega))}\leq \bar{m} \wedge \|\dot{v}\|_{L^{\infty}(0,T;H^1(\Omega))}\leq \bar{m}\\ & \wedge \| \nabla \dot{v}\|_{L^{q+1}(0,T;L^{q+1}(\Omega))}\leq \bar{M} \}. \end{split}$$ The interior regularity result can again be transferred from Corollary \[interior\_reg\_coupled\] to the present model: \[interior\_reg\_Neumann\] Let the assumptions of Proposition \[prop:Neumann\] and assumptions on the coefficients hold true, let $u_0\vert_{\Omega_i} \in H^2(\Omega_i)$, $i\in\{+,-\}$, $u_0,u_1 \in W^{1,q+1}(\Omega)$, and let $u$ be the weak solution of . Then $u_i \in H^1(0,T;H^2_{loc}(\Omega_i))$ and $|\nabla \dot{u}_i|^{\frac{q-1}{2}}\nabla \dot{u}_i \in L^2(0,T;H^1_{loc}(\Omega_i))$, $i \in \{+,-\}$. Furthermore, holds with $u$ replaced by $u_i$ and $\Omega$ by $\Omega_i$, $i \in \{+,-\}$. Boundary regularity for the coupled problem {#boundary} =========================================== We will show next that the $H^2$-regularity result can be extended up to the boundary of each of the subdomains under the assumption that the gradient of $u$ is essentially bounded in time and space on the whole domain. For this property to hold, we will need to smoothen out the subdomains, i.e. assume that they are $C^{1,1}$ regular. The proof will expand on the approach taken in Lemma 3.6, [@Ammari]. \[thm:boundary\_reg\_coupled\_Dirichlet\](Boundary $H^2$-regularity) Let the assumptions of Corollary \[interior\_reg\_coupled\] hold and let $\partial \Omega$ and $\Gamma=\partial \Omega_+$ be $C^{1,1}$ regular. If $u \in W^{1,\infty}(0,T;W^{1,\infty}(\Omega))$ and $\|\nabla \dot{u}\|_{L^{\infty}(\Omega)(0,T;L^\infty(\Omega))}$ is sufficiently small, then $u_{+,-} \in H^{1}(0,T;H^{2}(\Omega_{+,-}))$. We will only show that $u_+ \in H^{1}(0,T;H^{2}(\Omega_+))$, since $u_- \in H^{1}(0,T;H^{2}(\Omega_-))$ follows analogously.\ *Step 1: Straightening the boundary.* We begin by straightening the boundary through the change of coordinates near a boundary point (cf. Theorem 4, Section 6.3.2, [@evans]). Choose any point $x_0 \in \partial \Omega_+$. There exists a ball $B=B_r(x_0)$ for some $r>0$ and a $C^{1,1}$-diffeomorphism $\Psi:B \rightarrow \Psi(B) \subset \mathbb{R}^d$ such that $\text{det}|\nabla \Psi|=1$, $U^{\prime}=\Psi(B)$ is an open set, $\Psi(B \cap \Omega_+) \subset \mathbb{R}^d_+$ and $\Psi(B\cap \Gamma) \subset \partial \mathbb{R}^d_+$, where $\mathbb{R}^d_+$ is the half-space in the new coordinates.\ \ [straightening out the boundary]{} We change the variables and write $$\begin{aligned} &y=\Psi(x), \ x \in B, \\ &x=\Phi(y), \ y \in U^{\prime}.\end{aligned}$$ Then we have $\Psi(B \cap \Omega_+)=\{y \in U^{\prime}: y_n>0\}$. We denote $$\begin{aligned} B^+=B_{\frac{r}{2}}\cap \Omega_+, \ G=\Psi(B_{\frac{r}{2}}(x_0)), \ G^+=\Psi(B^+).\end{aligned}$$ Then $G \subset \subset U^{\prime}$ and $G^{+} \subset G$. We define $$\begin{aligned} w(y,t):= u(\Phi(y),t), \ (y,t) \in U^{\prime} \times [0,T].\end{aligned}$$ It immediately follows that $w(t):=w(\cdot,t) \in W^{1,q+1}(U^\prime)$. We now transform the original equation on $B \times [0,T]$ into an equation on $U^{\prime} \times [0,T]$: $$\label{eq:boundary_reg} \begin{aligned} & \int_{U^{\prime}}\Bigl\{\frac{1}{\hat{\lambda}}(1-2\hat{k}w(t))\ddot{w}(t)\phi+\displaystyle \sum_{i,j=1}^d\Bigl(\hat{\sigma}_{ij}D_iw(t)D_j\phi+\hat{\xi}_{ij}D_i \dot{w}_i(t) D_j \phi \\ &+|J_\Phi^T\nabla \dot{w}(t)|^{q-1}\hat{\eta}_{ij}D_i\dot{w}(t)D_j \phi\Bigr)-\frac{2\hat{k}}{\hat{\lambda}}(\dot{w}(t))^2\phi\Bigr\} \, dy=0, \end{aligned}$$ for a.e. $t \in [0,T]$, and all $\phi \in W^{1,q+1}_0(U^{\prime})$, where $D_iw=\frac{\partial{w}}{\partial y_i}$, and $$\begin{aligned} \label{def_coeff} & \hat{\lambda}(y)=\lambda(\Phi(y)), \ \hat{k}(y)=k(\Phi(y)), \nonumber\\ & \hat{\sigma}_{ij}=\displaystyle \sum_{r=1}^{d} \frac{1}{\varrho(\Phi(y))}\frac{\partial \Phi_i}{\partial x_r}(\Phi(y))\frac{\partial \Phi_j}{\partial x_r}(\Phi(y)),\nonumber\\ &\hat{\xi}_{ij}(y)=\displaystyle \sum_{r=1}^{d}b(\Phi(y))(1-\delta(\Phi(y)))\frac{\partial \Phi_i}{\partial x_r}(\Phi(y))\frac{\partial \Phi_j}{\partial x_r}(\Phi(y)),\\ &\hat{\eta}_{ij}(y)=\displaystyle \sum_{r=1}^{d}b(\Phi(y))\delta(\Phi(y))\frac{\partial \Phi_i}{\partial x_r}(\Phi(y))\frac{\partial \Phi_j}{\partial x_r}(\Phi(y)). \nonumber\end{aligned}$$ Note that $D_r\hat{\sigma}_{ij}, D_r\hat{\xi}_{ij}, D_r\hat{\eta}_{ij} \in L^\infty(U^{\prime})$ for $r\in\{1,\ldots, d-1\}$ since $\Psi$ and $\Phi$ are $C^{1,1}$ mappings and $C^{1,1}=W^{2,\infty}$ (cf. Chapter 2, Section 2.6.4, [@DZ]). It can be shown (cf. Section 6.3.2, [@evans]) that $$\label{coercivity} \begin{aligned} &\displaystyle \sum_{i,j=1}^d \hat{\sigma}_{ij} \phi_{i} \phi_j \geq K_1|\phi|^2, \ \displaystyle \sum_{i,j=1}^d \hat{\xi}_{ij} \phi_{i} \phi_j \geq K_1|\phi|^2, \\ &\displaystyle \sum_{i,j=1}^d \hat{\eta}_{ij} \phi_{i} \phi_j \geq K_1|\phi|^2, \ \ \forall (y,\phi) \in U^{\prime} \times \mathbb{R}^d. \end{aligned}$$ Next, we choose a domain $W^{\prime}$ such that $G \subset \subset W^{\prime} \subset \subset U^{\prime}$ and select a cut-off function such that $$\begin{aligned} \begin{cases} \zeta=1 \ \text{on} \ G, \ \ \zeta=0 \ \text{on} \ \mathbb{R}^d \setminus W^{\prime},\\ 0 \leq \zeta \leq 1. \end{cases}\end{aligned}$$ Let $|l|>0$ be small and choose $r \in \{1,\ldots, d-1\}$. Note that since we consider directions parallel to the interface now, we have $D_r^l \hat{\lambda}=0$, $D_r^l \hat{k}=0$ and that there exists a constant $K_2>0$ such that $$\begin{aligned} \label{est_coeff} |D_r^l \hat{\sigma}_{ij}(y)|<K_2, \ |D_r^l \hat{\xi}_{ij}(y)|<K_2, \ |D_r^l \hat{\eta}_{ij}(y)|<K_2\end{aligned}$$ for a.e. $y \in W^{\prime}$, $1 \leq i,j \leq d$ and sufficiently small $|l|$.\ *Step 2: Existence of second order derivatives $ D_j D_i w \in H^1(0,T;L^2(G^+))$, $j \neq d$.* We then use $\phi=-D_r^{-l}(\zeta^2D_r^l\dot{w}(t))$ as a test function in , which, after integration with respect to time, results in $$\begin{aligned} &\frac{1}{2}\Bigl[\int_{W^{\prime}}\frac{1}{\hat{\lambda^l}}(1-2\hat{k}^lw^l)(\zeta D_r^l \dot{w})^2\, dy\Bigr]_0^t+\frac{1}{2}\Bigl[\int_{W^{\prime}}\displaystyle \sum_{i,j=1}^d \hat{\sigma}_{ij}^l \zeta^2 D_r^l D_i w D_r^l D_j w\, dy\Bigr]_0^t\\ &+\int_0^t \int_{W^{\prime}}\displaystyle \sum_{i,j=1}^d \hat{\xi}_{ij}^l \zeta^2 D_r^l D_i \dot{w} D_r^l D_j \dot{w}\, dy \, ds \\ =& \,\int_0^t \int_{W^{\prime}}\frac{2\hat{k}^l}{\hat{\lambda}^l} D_r^l w \ddot{w} \zeta^2 D_r^l \dot{w} \, dy \, ds+\int_0^t \int_{W^{\prime}}\frac{\hat{k}^l}{\hat{\lambda}^l}(\dot{w}^l+2\dot{w})(\zeta D_r^l \dot{w})^2 \, dy \, ds \\ &-\int_0^t \int_{W^{\prime}}\displaystyle \sum_{i,j=1}^d D_r^l (\hat{\sigma}_{ij})D_i w \,( \zeta^2D_r^lD_j\dot{w}+2\zeta D_j \zeta D_r^l \dot{w})\, dy \, ds\\ &-2\int_0^t \int_{W^{\prime}}\displaystyle \sum_{i,j=1}^d\hat{\sigma}^l_{ij} D_r^l D_i w \, \zeta D_j\zeta D_r^l \dot{w}\, dy \, ds \\ &-\int_0^t \int_{W^{\prime}}\displaystyle \sum_{i,j=1}^d D_r^l (\hat{\xi}_{ij})D_i \dot{w} \,( \zeta^2D_r^lD_j\dot{w}+2\zeta D_j \zeta D_r^l \dot{w})\, dy \, ds\\ &- 2\int_0^t \int_{W^{\prime}}\displaystyle \sum_{i,j=1}^d \hat{\xi}_{ij} D_i \dot{w} \, \zeta D_j \zeta D_r^l \dot{w}\, dy \, ds\\ &-\int_0^t \int_{W^{\prime}}\displaystyle \sum_{i,j=1}^d D_r^l(|J_\Phi^T\nabla \dot{w}|^{q-1}\hat{\eta}_{ij}D_i\dot{w}) (\zeta^2 D_r^l D_j \dot{w}+2\zeta D_j \zeta D_r^l \dot{w})\, dy \, ds.\end{aligned}$$ We can estimate the last term on the right hand side as follows $$\begin{aligned} &-\int_0^t \int_{W^{\prime}}\displaystyle \sum_{i,j=1}^d D_r^l(|J_\Phi^T\nabla \dot{w}|^{q-1}\hat{\eta}_{ij}D_i\dot{w}) (\zeta^2 D_r^l D_j \dot{w}+2\zeta (D_j \zeta) D_r^l \dot{w})\, dy \, ds \\ =& \, -\int_0^t \int_{W^{\prime}}\displaystyle \sum_{i,j=1}^d |(J_\Phi^T\nabla \dot{w})^l|^{q-1}\hat{\eta}^l_{ij} D_r^lD_i\dot{w} (\zeta^2 D_r^l D_j \dot{w}+2\zeta (D_j \zeta) D_r^l \dot{w})\, dy \, ds \\ &-\int_0^t \int_{W^{\prime}}\displaystyle \sum_{i,j=1}^dD_r^l(|J_\Phi^T\nabla \dot{w}|^{q-1}\hat{\eta}_{ij}) D_i\dot{w} (\zeta^2 D_r^l D_j \dot{w}+2\zeta( D_j \zeta) D_r^l \dot{w})\, dy \, ds \\ \leq& \, C\|\nabla \dot{w}\|^{q-1}_{L^\infty(0,T;L^\infty(W^\prime))}(\displaystyle \sum_{i=1}^d \|\zeta D_r^lD_i\dot{w}\|^2_{L^2(0,T;L^2(W^\prime))}+\|\nabla \zeta\|_{L^\infty(0,T;L^\infty(W^\prime))}\|\nabla \dot{w}\|^2_{L^2(0,T;L^2(W^\prime))} )\end{aligned}$$ with $C$ independent of $\nabla\zeta$. Due to the rest of the terms on the right hand side can be estimated analogously to the estimates in the proof of Theorem \[thm:interior\_reg\], which for sufficiently small $\|\nabla \dot{w}\|_{L^\infty(0,T;L^\infty(W^\prime))}$ leads to $$\begin{aligned} &\|\zeta D_r^l \dot{w}\|^2_{L^\infty(0,T;L^2(W^\prime))}+\displaystyle \sum_{i=1}^d \|\zeta D_r^l D_i w \|^2_{L^\infty(0,T;L^2(W^\prime))} +\displaystyle \sum_{i=1}^d \|\zeta D_r^l D_i \dot{w}\|^2_{L^2(0,T;L^2(W^\prime))} \\ \leq& \,C((1+\|\dot{w}\|_{L^\infty(0,T;L^\infty(W^\prime))})\|\nabla w\|^2_{L^\infty(0,T;L^2(W^\prime))} \\ &+(1+\|\nabla \dot{w}\|^{q-1}_{L^\infty(0,T;L^\infty(W^\prime))})\|\nabla \dot{w}\|^2_{L^2(0,T;L^2(W^\prime))} +|\nabla \dot{w}(0)|^2_{L^2(W^\prime)}+|w(0)|^2_{H^2(W^\prime)}).\end{aligned}$$ Recalling the definition of $\zeta$ and employing Lemma \[reg\_evans\] yields $D_j D_i w \in H^1(0,T;L^2(G^+))$ for $1\leq i \leq d$, $1\leq j \leq d-1$.\ *Step 3: Existence of second order derivative $ D_d D_d w \in H^1(0,T;L^2(G^+))$.* It remains to show that $D_{dd}w:=D_d D_d w \in H^1(0,T;L^2(G^+))$. From , after integration by parts, we obtain $$\begin{aligned} \label{boundary_eq} &\int_{G^+} \{\hat{\sigma}_{dd}D_dw(t)+\hat{\xi}_{dd}D_d\dot{w}(t)+|J_\Phi^T\nabla \dot{w}(t)|^{q-1}\hat{\eta}_{dd}D_d \dot{w}(t)\} D_d \phi \, dy \nonumber\\ =&\,\int_{G^+}\Bigl\{-\frac{1}{\hat{\lambda}}(1-2\hat{k}w(t))\ddot{w}(t)+\frac{2\hat{k}}{\hat{\lambda}}(\dot{w}(t))^2 \\ &+\displaystyle \sum_{j=1}^{d-1}\displaystyle \sum_{i=1}^d \Bigl(D_j(\hat{\sigma}_{ij}D_iw(t))+D_j(\hat{\xi}_{ij}D_i \dot{w}(t)) +D_j(|J_\Phi^T\nabla \dot{w}(t)|^{q-1}\hat{\eta}_{ij}D_i \dot{w}(t))\Bigr)\Bigr\}\phi \, dy,\\ &=: \,\int_{G^+}\hat{f}(w)(t)\phi \, dy, \nonumber\end{aligned}$$ for $\phi \in C_0^{\infty}(G^+)$, a.e. in $[0,T]$. Since the right hand side of the equation is well-defined, we conclude that for a.e. $t \in [0,T]$ the weak derivative of $\hat{\sigma}_{dd}D_dw(t)+\hat{\xi}_{dd}D_d\dot{w}(t)+|J_\Phi^T\nabla \dot{w}|^{q-1}\hat{\eta}_{dd}D_d \dot{w}(t)$ with respect to $y_d$ exists on $G^+$. Furthermore, for a.e. $t \in [0,T]$ the weak derivative satisfies $$\label{weak_derivative} \begin{aligned} &-D_d(\hat{\sigma}_{dd}D_dw(t)+\hat{\xi}_{dd}D_d\dot{w}(t)+|J_\Phi^T\nabla \dot{w}|^{q-1}\hat{\eta}_{dd}D_d \dot{w}(t)) =\hat{f}(w)(t) \end{aligned}$$ on $G^+$. From what we have shown, it follows that $\hat{f}(w) \in L^2(0,T;L^2(G^+))$. We set $$z(t):=\hat{\sigma}_{dd}D_dw(t)+\hat{\xi}_{dd}D_d\dot{w}(t)+|J_\Phi^T\nabla \dot{w}|^{q-1}\hat{\eta}_{dd}D_d \dot{w}(t),$$ and $$\tilde{\xi}_{dd}(t,D_d\dot{w}(t)):= \hat{\xi}_{dd}+|J_\Phi^T\nabla \dot{w}(t)|^{q-1}\hat{\eta}_{dd},$$ (suppressing in the notation dependence on $D_1\dot{w}(t),\ldots,D_{d-1}\dot{w}(t)$ which we already know to be smooth anyway) so that relation reads $$-D_d z(t)= \hat{f}(w)(t),$$ where $$\label{ODE} z(t)=\hat{\sigma}_{dd}D_dw(t)+\tilde{\xi}_{dd}(t,D_d\dot{w}(t))D_d\dot{w}(t).$$ Since $\hat{f}(w)\in L^2(0,T;L^2(G^+))$, and using the fact that $D_j D_i w \in H^1(0,T;L^2(G^+))$ for $1\leq i \leq d$, $1\leq j \leq d-1$, we have $z\in L^2(0,T;H^1(G^+))$. On the other hand, due to assuming that $u \in W^{1,\infty}(0,T;W^{1,\infty}(\Omega))$, we know also that $z \in L^\infty(0,T;L^\infty(G^+))$. Therefore we conclude that $$z \in L^2(0,T;H^1(G^+)) \cap L^\infty(0,T;L^\infty(G^+)).$$ Since represents an ODE (pointwise a.e. in space) for $D_d w(t)$, it can be resolved as follows: $$\label{Dddow} \begin{aligned} D_dw(t)=& \, \exp \Bigl(-\int_0^t\frac{\hat{\sigma}_{dd}}{\tilde{\xi}_{dd}(\tau,D_d\dot{w}(\tau))}\, d\tau\Bigr) \Bigl(\int_0^t\frac{z(\tau)}{\tilde{\xi}_{dd}(\tau,D_d\dot{w}(\tau))} \exp\Bigl(\int_0^\tau\frac{\hat{\sigma}_{dd}}{\tilde{\xi}_{dd}(\rho,D_d\dot{w}(\rho))}\, d\rho \Bigr)\, d\tau \\ &+ D_dw(0)\Bigr), \end{aligned}$$ and then also $D_d\dot{w}(t)$ can be expressed in terms of $z, \hat{\sigma}_{dd}, \tilde{\xi}_{dd}$: $$\label{Dddotw} \begin{aligned} D_d\dot{w}(t)=& \,\frac{z(t)}{\tilde{\xi}_{dd}(t,D_d\dot{w}(t))} -\frac{\hat{\sigma}_{dd}}{\tilde{\xi}_{dd}(t,D_d\dot{w}(t))} \exp\left(-\int_0^t\frac{\hat{\sigma}_{dd}}{\tilde{\xi}_{dd}(\tau,D_d\dot{w}(\tau))}\, d\tau\right)\\ &\left(\int_0^t\frac{z(\tau)}{\tilde{\xi}_{dd}(\tau,D_d\dot{w}(\tau))} \exp\left(\int_0^\tau\frac{\hat{\sigma}_{dd}}{\tilde{\xi}_{dd}(\rho,D_d\dot{w}(\rho))}\, d\rho\right)\, d\tau + D_dw(0)\right) \end{aligned}$$ Since the right hand side depends on $D_d\dot{w}(t)$, this is rather a fixed point equation than an explicit expression for $D_d\dot{w}(t)$. We therefore consider the fixed point operator $\mathcal{T}:M\to M$, defined by the right hand side of , i.e., $$\mathcal{T}=\mathcal{S}\circ\mathcal{R},$$ where $\mathcal{R}:M\to \hat{M}$, $\hat{M}=\{\phi \in M: \phi \geq K_1\}$, is the superposition operator associated with $\tilde{\xi}_{dd}$, i.e., $\mathcal{R}(v)(t)=\tilde{\xi}_{dd}(t,v(t))$, and $\mathcal{S}:\hat{M} \to M$ $$\begin{aligned} \mathcal{S}(r)(t)=& \,\frac{z(t)}{r(t))} -\frac{\hat{\sigma}_{dd}}{r(t)} \exp\left(-\int_0^t\frac{\hat{\sigma}_{dd}}{r(\tau)}\, d\tau\right) \left(\int_0^t\frac{z(\tau)}{r(\tau)} \exp\left(\int_0^\tau\frac{\hat{\sigma}_{dd}}{r(\rho)}\, d\rho\right)\, d\tau + D_dw(0)\right).\end{aligned}$$ Note that both $\hat{\sigma}_{dd}$ and $\tilde{\xi}_{dd}$ are bounded from below by $K_1$ due to . We can then conclude that $\mathcal{T}$ is a self-mapping on $$M=M_0=L^\infty(0,T;L^\infty(G^+)),$$ and on $$M=M_1=L^2(0,T;H^1(G^+))\cap L^\infty(0,T;L^\infty(G^+))\,,$$ since $z \in L^2(0,T;H^1(G^+)) \cap L^\infty(0,T;L^\infty(G^+))$. Moreover, $$\mathcal{R}'(D_d\dot{w}) =\frac{\partial}{\partial v} \Bigl(\hat{\xi}_{dd}+|J_\Phi^T(D_1\dot{w},\ldots,D_{d-1}\dot{w},v)|^{q-1}\hat{\eta}_{dd}\Bigr)\vert_{v=D_d\dot{w}}$$ is small if $\|\nabla \dot{u}_i\|_{L^{\infty}(0,T;L^\infty(\Omega_i))}$ and therefore $\|\nabla \dot{w}\|_{L^{\infty}(0,T;L^\infty(G^+))}$ is small. This implies that $\mathcal{T}$ is a contraction on $$\tilde{M}_0=\{v\in M_0 \, : \ \|v-D_d\dot{w}\|_{L^\infty(0,T;L^\infty(G^+))}\leq\gamma\}$$ and on $$\tilde{M}_1=\{v\in M_1 \, : \ \|v-D_d\dot{w}\|_{L^\infty(0,T;L^\infty(G^+))}\leq\gamma\}\,.$$ for $\gamma$ sufficiently small. Thus the fixed point equation $v=\mathcal{T}(v)$ has a unique solution $v_0$ in $\tilde{M}_0$ and it also has a uniqe solution $v_1$ in $\tilde{M}_1$, and both have to coincide $v_1=v_0$ by uniqueness on $\tilde{M}_0\supseteq \tilde{M}_1$. On the other hand, obviously $D_d\dot{w}$ lies in $\tilde{M}_0$, solves this fixed point equation and thus has to coincide with $v_0$, hence also with $v_1$. This proves that $D_d\dot{w}\in M_1\subseteq L^2(0,T;H^1(G^+))$.\ By transforming $w$ back to $u$, we can conclude that $u \in H^1(0,T;H^2(B^+))$. The assertion then follows from the fact that the boundary is compact and can be covered by a finite set of balls $\{B_{r_i/2}(x_i)\}_{i=1}^N$. Higher boundary regularity was obtained under the assumption that $u$ belongs to $W^{1,\infty}(0,T;W^{1,\infty}(\Omega))$; this was necessitated by the presence of the $q$-Laplace damping term in the equation. The assumption is equivalent to assuming Lipschitz continuity of $u$ in time and space, i.e. $u \in C^{0,1}(0,T;C^{0,1}(\overline{\Omega}))$ (see Theorem 4, Chapter 5, [@evans]). Neumann problem for the coupled system -------------------------------------- It remains to show $H^2$-regularity up to the boundary for the Neumann problem . Let the assumptions of Corollary \[interior\_reg\_Neumann\] hold, let $\partial \Omega$ and $\Gamma=\partial \Omega_+$ be $C^{1,1}$ regular and let $u$ be the weak solution of . Furthermore, assume that $g \in L^2(0,T;H^{1/2}(\partial \Omega))$. If $u \in W^{1,\infty}(0,T;W^{1,\infty}(\Omega))$ and $\|\nabla \dot{u}\|_{L^{\infty}(\Omega)(0,T;L^\infty(\Omega))}$ is sufficiently small, then $u_i \in H^1(0,T;H^2(\Omega_i))$, $i \in \{+,-\}$. We will show that $u_- \in H^1(0,T;H^2(\Omega_-))$, since the regularity on $\Omega_+$ follows as in the proof of Theorem \[thm:boundary\_reg\_coupled\_Dirichlet\]. We begin again as before, by straightening the boundary around $x_0 \in \partial \Omega_- \setminus \partial \Omega_+$. There exists a ball $B=B_r(x_0)$ for some $r>0$ and a $C^{1,1}$-diffeomorphism $\Psi:B \rightarrow \Psi(B) \subset \mathbb{R}^d$ such that $\text{det}|\nabla \Psi|=1$, $U^{\prime}=\Psi(B \cap \Omega_-) \subset \mathbb{R}^d_+$ is an open set, and $\Psi(B\cap \partial \Omega_-) \subset \partial \mathbb{R}^d_+$. We change the variables and write $$\begin{aligned} &y=\Psi(x), \ x \in B, \\ &x=\Phi(y), \ y \in U^{\prime}.\end{aligned}$$ We denote $ G:=\Psi(B_{\frac{r}{2}}\cap \Omega_-) \subset \subset U^{\prime}$. We define $$\begin{aligned} w(y,t):= u(\Phi(y),t), \quad (y,t) \in U^{\prime} \times [0,T],\end{aligned}$$ and transform the orginial equation from $(B \cap \Omega) \times [0,T]$ to $U^\prime \times [0,T]$: $$\label{eq:boundary_reg_Neumann} \begin{aligned} & \int_{U^{\prime}}\Bigl\{\frac{1}{\hat{\lambda}}(1-2\hat{k}w(t))\ddot{w}(t)\phi+\displaystyle \sum_{i,j=1}^d\Bigl(\hat{\sigma}_{ij}D_iw(t)D_j\phi+\hat{\xi}_{ij}D_i \dot{w}_i(t) D_j \phi \\ &+|J_\Phi^T\nabla \dot{w}(t)|^{q-1}\hat{\eta}_{ij}D_i\dot{w}(t)D_j \phi\Bigr)-\frac{2\hat{k}}{\hat{\lambda}}(\dot{w}(t))^2\phi\Bigr\} \, dy=\int_{\partial U^\prime} g \phi |J^{-T}_{\Phi}n|_{\mathbb{R}^d}\, dx, \end{aligned}$$ for a.e. $t \in [0,T]$, and all $\phi \in W^{1,q+1}(U^{\prime})$, and $\hat{\lambda}$, $\hat{k}$, $\hat{\sigma}_{ij}$, $\hat{\xi}_{ij}$, and $\hat{\eta}_{ij}$ are defined as in . We then again use $\phi=-D_r^{-l}(\zeta^2D_r^l\dot{w}(t))$, $r \in \{1,\ldots, d-1\}$ as a test function and proceed with the estimates like in the proof of Theorem \[thm:boundary\_reg\_coupled\_Dirichlet\]. The only difference here is the need to estimate the boundary integral over $\partial U^\prime$ appearing in the weak form, therefore we focus our attention solely on estimating this term: $$\begin{aligned} &-\int_0^t\int_{\partial U^\prime} g D_r^{-l}(\zeta^2D_r^l\dot{w}) |J^{-T}_{\Phi}n|\, dx \, ds \\ =& \, \int_0^t\int_{B\cap\partial \Omega_-} g D_r^{-l}(\zeta^2D_r^l\dot{u}) \, dx \, ds \\ \leq& \, C \|g\|_{L^2(0,T;H^{1/2}(\partial \Omega_-))} \|\zeta^2D_r^l\dot{u}\|_{L^2(0,T;H^{1/2}(\partial \Omega_-))} \\ \leq& \, C \|g\|_{L^2(0,T;H^{1/2}(\partial \Omega_-))} \|\zeta D_r^l\dot{u}\|_{L^2(0,T;H^{1}(\Omega_-))}\\ \leq& \, \varepsilon \|\zeta D_r^l\dot{u}\|^2_{L^2(0,T;H^{1}(\Omega))}+\frac{1}{4\varepsilon}C^2 \|g\|^2_{L^2(0,T;H^{1/2}(\partial \Omega_-))}.\end{aligned}$$ Showing that $D_{dd}w \in H^1(0,T;L^2(G))$ follows as in the proof of Theorem \[thm:boundary\_reg\_coupled\_Dirichlet\], since we use $\phi \in C^{\infty}_0(G)$ in . Acknowledgments {#acknowledgments .unnumbered} =============== The financial support by the FWF (Austrian Science Fund) under grant P24970 is gratefully acknowledged as well as the support of the Karl Popper Kolleg “Modeling-Simulation-Optimization”, which is funded by the Alpen-Adria-Universit" at Klagenfurt and by the Carinthian Economic Promotion Fund (KWF). References {#references .unnumbered} ========== [00]{} N. D. Alikakos and L. C. Evans, Continuity of the gradient of the solutions of certain dgenerate parabolic equations, J. Math. pures et appl., 62, 253–268, (1983). H. Ammari, D. Chen, J. Zou, Well-posedness of a Pulsed Electric Field Model in Biological Media and its Finite Element Approximation, arXiv:1502.06803. A. Bamberger, R. Glowinski and Q.H. Tran, A domain decomposition method for the acoustic wave equation with discontinuous coefficients and grid change. SIAM J. Numer. Anal. 34, 603–639, (1997). R. Brunnhuber, B. Kaltenbacher and P. Radu, Relaxation of regularity for the Westervelt equation by nonlinear damping with application in acoustic-acoustic and elastic-acoustic coupling, Evolution Equations and Control Theory, 3(4):595 - 626, (2014). M. C. Delfour and J.P. Zolesio, Shapes and Geometries, 2nd edn, SIAM, (2001). E. DiBenedetto, Degenerate parabolic equations, Universitext, New York, NY: Springer-Verlag, xv 387, (1993.) E. DiBendetto and A. Friedman, Regularity of solutions of nonlinear degenerate parabolic systems, J. Reine Angew. Math., 349, 83–128, (1984). L. C. Evans, Partial Differential Equations, 2nd edn, American Mathematical Society, Providence, (1998). H. Gao and T. F. Ma, Global solutions for a nonlinear wave equation with the $p-$Laplacian operator, Electronic J. Qualitative Theory Differ. Equ., 11, 1–-13, (1999). M. F. Hamilton and D. T. Blackstock, Nonlinear Acoustics, Academic Press, New York, (1997). B. Kaltenbacher and I. Lasiecka, Global existence and exponential decay rates for the Westervelt equation, Discrete and Continuous Dynamical Systems Series S, 2, 503–525, (2009). B. Kaltenbacher, I. Lasiecka and S. Veljović, Well-posedness and exponential decay for the Westervelt equation with inhomogeneous Dirichlet boundary data, J. Escher et al (Eds): Progress in Nonlinear Differential Equations and Their Applications, 60, 357–387, (2011). M. Kaltenbacher, Numerical Simulations of Mechatronic Sensors and Actuators, Springer, Berlin, (2004). G. M. Lieberman, Boundary regularity for solutions of degenerate elliptic equations, Nonlinear Anal., 12, 1203–1219, (1988). P. Lindqvist, Notes on the p-Laplace equation, Lecture notes, University of Jyväskylä, (2006). W. Liu and N. Yan, Quasi-norm local error estimators for $p$-Laplacian, SIAM Journal on Numerical Analysis, 39, 100–127, (2002). V. Nikoli' c, Existence results for the Westervelt equation with nonlinear damping and Neumann as well as absorbing boundary conditions, J. Math. Anal. Appl., doi:10.1016/j.jmaa.2015.02.076, (2015). V. Nikoli' c and B. Kaltenbacher, Sensitivity analysis for shape optimization of a focusing acoustic lens in lithotripsy, in preparation. N. N. Ural’tseva, Degenerate quasilinear elliptic systems, Zap. Na. Sem. Leningrad. Otdel. Mat. Inst. Steklov. (LOMI), 7, 184–-222, (1968). P.J. Westervelt, Parametric acoustic array, The Journal of the Acoustic Society of America, 35: 535–537, (1963). Z. Wilstein, Global well-posedness for a nonlinear wave equation with $p$-Laplacian damping, PhD thesis, University of Nebraska-Lincoln, (2011).

ArXiv

--- abstract: 'We have studied the problem of the dissipative motion of Bloch walls considering a totally anisotropic one dimensional spin chain in the presence of a magnetic field. Using the so-called “collective coordinate method” we construct an effective Hamiltonian for the Bloch wall coupled to the magnetic excitations of the system. It allows us to analyze the Brownian motion of the wall in terms of the reflection coefficient of the effective potential felt by the excitations due to the existence of the wall. We find that for finite values of the external field the wall mobility is also finite. The spectrum of the potential at large fields is investigated and the dependence of the damping constant on temperature is evaluated. As a result we find the temperature and magnetic field dependence of the wall mobility.' address: - | $^{1}$Departamento de Física, Facultad de Ciencias Exactas y Naturales,\ Universidad de Buenos Aires, RA-1428, Buenos Aires, Argentina. - | $^{2}$Instituto de Física “Gleb Wataghin”,\ Departamento de Física do Estado Sólido e Ciência dos Materiais,\ Universidade Estadual de Campinas, 13083-970, Campinas, SP, Brasil. - '$^{3}$Dept. of Physics, University of California, Riverside CA, 92521, USA.' author: - | M.A. Despósito[^1]$^{1}$, A. Villares Ferrer$^{2}$, A.O. Caldeira$^{2}$\ and A.H. Castro Neto$^{3}$ title: Mobility of Bloch Walls via the Collective Coordinate Method --- Introduction ============ It is a well-known fact that ultimately due to magnetic dipole interaction, different domains are formed in magnetic systems [@Lan]. In many situations, the physical region separating two different magnetic domains—the domain wall—must be treated as a physical entity because it has a characteristic behaviour when acted by external agents. For instance, it is known that the response of a magnetic system to a frequency dependent external magnetic field depends on whether domain walls are present [@Dillon]. Domain walls can also move throughout the system and this motion happens to be dissipative [@Land]. A particularly interesting kind of domain wall is commonly found in low dimensional ferromagnetic systems. These are the so-called Bloch walls[@Malozemoff]. It is known that these walls perform dissipative motion [@Land] due to the presence of the elementary excitations which can be scattered by the wall as it moves and the momentum transferred to them reduces the speed of the wall. The primary aim of this work is to study the influence of finite temperatures in the mobility of these Bloch walls. For this purpose we start by considering a microscopic model for a one dimensional ferromagnet containing hard and easy-axis anisotropies and subject to an external magnetic field. A semiclassical picture provides us with the localized solutions for the spin configurations which are the solitons corresponding to the walls. Making use of a recently developed method for the analysis of the dissipative dynamics of solitons [@CN; @Caldeira], in which the “collective coordinate method” [@R.R.] is used to transform the original Hamiltonian into one of a particle coupled to an infinite set of modes, we show that the Bloch wall behaves like a Brownian particle. The advantage of using this method is that we keep closer contact with the microscopic details of the system and the mobility is naturally calculated as a function of the temperature. The information from the microscopic scattering processes between the Bloch wall and the residual modes can be obtained from the knowledge of the phase shifts of the associated spectral problem. In the case of reflectionless potentials, as it happens for vanishing anisotropies or external field, the motion of the wall is undamped. If this is not the case, the reflection coefficient does not vanish and the mobility is finite. The outline of this paper is as follows. In Sec. II we present the model. The dynamics of its static solution is investigated in Sec. III and there we also show how to obtain an effective Hamiltonian for the Bloch wall coupled to the residual magnetic excitations. In Sec. IV the mobility of the Bloch wall is studied in terms of the scattering phase shifts of the second variation problem. The case of large external fields is investigated in Sec.V where the phase shifts and the damping constant are explicitly evaluated. Finally, we present our conclusions in Sec. VI. The Model and its static solutions ================================== In this work we consider a one dimensional magnetic system composed by an array of spins lying along the $\hat{z}$-direction. Furthermore, let us assume that there is an easy-plane anisotropy which tries to keep the spins on the $x$-$y$ plane and, on top of this, an in-plane anisotropy tending to align them along the $\hat{x}$-direction. This is a totally anisotropic model which is described by a XYZ model of magnetic systems defined by the Hamiltonian $$H=-\sum_{\langle ij\rangle }\left( J_xS_i^{(x)}S_j^{(x)}+J_yS_i^{(y)}S_j^{(y)}+J_zS_i^{(z)}S_j^{(z)}\right) -\frac \mu \hbar B\sum_iS_i^{(x)} \label{Hspin}$$ where ${J_x>J_y>J_z>0}$, ${S_j^{(\alpha )}}$ is the $\alpha $ component (${% \alpha =x,y,z}$) of the $i^{th}$ spin of the system, ${\mu }$ is the modulus of the magnetic moment on each site and $B$ is the external magnetic field. The ferromagnetic $XYZ$ model is actually defined for ${B=0}$ and this is the starting point of our analysis. As we can see from ($\ref{Hspin}$), the ground state of this system is the configuration where all the spins are aligned in the $\hat{x}$-direction. However, there is another possible configuration which is a local minimum of the energy functional and cannot be obtained from the previous uniform configuration by any finite energy operation. Let us imagine that we describe our spins classically by vectors $${\bf S}_i=S(\sin \theta _i\cos \varphi _i,\sin \theta _i\sin \varphi _i,\cos \theta _i) \label{semispin}$$ where $\theta _i$ and $\varphi _i$ are the polar angles of the $i^{th}$ spin. In this representation the above-mentioned configuration consists of all $\theta _i$’s equal to ${\pi /2}$  and  $\varphi _i$’s equal to zero or $\pi $. However, there are other configurations in which $\theta _i=\pi /2,\varphi _i=0$ if ${i\rightarrow -\infty }$ and ${\varphi _i=\pi }$ if ${% i\rightarrow \infty }$ which are approximately (only because $\theta _i$’s may slightly vary [@Stamp]) local minima of the energy functional of the system. So, ${\bf S}_i$ winds around the $\hat{z}$ direction starting at ${% (\theta ,\varphi )}$ = ${(\pi /2,0)}$ and ending at ${(\theta ,\varphi )}$ = ${(\pi /2,\pi )}$. The so-called $\pi $-Bloch wall [@Winter] is one example of these configurations where $\varphi _i$ varies from $0$ to $\pi $ without making a complete turn around the $\hat{z}$ axis. Later on we will see the specific form of this configuration when we consider the system in the continuum limit. It will then be shown that Bloch walls are related to soliton-like solutions of the non-linear equations which control the spin dynamics in the semi-classical approximation. If we now turn the external field $B$ on what happens is that the degeneracy between ${\varphi =0}$ and ${\varphi =\pi }$ is broken. For ${B>0}$ it is clear from ($\ref{Hspin}$) that ${\varphi =0}$ has lower energy than ${% \varphi =\pi }$ which is now a metastable configuration of the system. In this circumstance the system still presents a local minimum of the energy functional. The only difference is that whereas $\theta_i$ is still approximately ${\pi /2}$, $\varphi _i$ starts and ends at zero as -${\infty <i<\infty }$. The 2$\pi $-Bloch wall is now the configuration where $\varphi _i$ winds only once around $\hat{z}$. In any of the two cases mentioned above, there is no way we could spend a finite energy to transform the Bloch wall into the uniform configuration. We would need to turn an infinite number of spins over an anisotropy energy barrier. We say that these two configurations are topologically distinct. Another important point is that ($\ref{Hspin}$) is translation invariant and this is reflected by the translation invariance of the Bloch wall. This means that the region about which the spins wind up can be centered anywhere on the $\hat{z}$ axis. In reality they can even move with constant speed along that direction. These structures can be obtained by mapping the original Hamiltonian ($\ref {Hspin}$) into a $1+1$ field theoretical model such as the $\varphi ^4$, sine-Gordon or any other appropriate model. This can be done by simply proceeding a bit further with the semi-classical description for the spins we have seen in (\[semispin\]) and writing the Hamilton equations of motion for $\theta _i(t)$ and $\varphi _i(t)$. After having done that we take the continuum limit $\theta _i\rightarrow \theta (z,t)$ and $\varphi _i(t)\rightarrow \varphi (z,t)$ and write $$\theta (z,t)\approx \frac \pi 2+\alpha (z,t),$$ where ${\alpha (z,t)\ll \ 1}$. Assuming that the variations of $\varphi $ and $\alpha $ from site to site of the spin chain are small and linearizing the equations of motion with respect to $\alpha $ one obtains $$\begin{aligned} \dot{\varphi} &=&\alpha 2S(J_x\cos ^2\varphi +J_y\sin ^2\varphi -J_z) \\ \dot{\alpha} &=&a^2S(J_x\sin ^2\varphi +J_y\cos ^2\varphi )\frac{\partial ^2\varphi }{\partial z^2}-\sin \varphi \left[\frac{\mu B}\hbar +2S(J_x-J_y)\cos \varphi \right]\end{aligned}$$ where $a$ is the lattice spacing. Then, eliminating $\alpha $ from these equations, we get an effective equation of motion for ${\varphi (z,t)}$ of the form $$\frac 1{c^2}\frac{\partial ^2\varphi }{\partial t^2}-\frac{\partial ^2\varphi }{\partial z^2}=-A_1\sin \varphi -A_2\sin 2\varphi \label{mofi}$$ where $$c^2\cong 2a^2S^2J_xJ_y\left( 1-\frac{J_z}{J_y}\right) , \label{c}$$ $$A_1=\frac {\mu B}{a^2SJ_x \hbar} , \label{A1}$$ and $$A_2=\frac 1{a^2}\left( 1-\frac{J_y}{J_x}\right) . \label{A2}$$ Notice that if $J_x=J_y$ one has $A_2=0$ whereas if $B=0$ it turns out that $% A_1=0$. So, as we can derive the r.h.s. of (\[mofi\]) from a potential energy density $U(\varphi )$ given by $$U(\varphi )=A_1(1-\cos \varphi )+\frac{A_2}2(1-\cos 2\varphi ) \label{pot}$$ we see that $A_1$ controls the potential energy barrier due to the presence of $B\neq 0$ and $A_2$ controls the anisotropy energy barrier. The static solutions (${\partial \varphi /\partial t=0}$) of Eq. (\[mofi\]) are obtained using that [@R.R.] $$z-z_0=\int_{\varphi (z_0)}^{\varphi (z)}\frac{d\varphi ^{\prime }}{\sqrt{% 2U(\varphi ^{\prime })}}$$ are the solitons of the system. In particular, the examples of Bloch walls we gave above are the solitons $$\begin{aligned} \varphi (z) &=&2\tan ^{-1}[\exp \sqrt{2A_2}(z-z_0)]\quad \text{ if }A_1=0 \label{fia10} \\ \varphi (z) &=&4\tan ^{-1}[\exp \sqrt{A_1}(z-z_0)]\quad \;\text{ if }A_2=0 \label{fia20}\end{aligned}$$ while for the general case of finite anisotropy and magnetic field the solution is the 2$\pi $-Bloch wall $$\varphi (z)=2\tan ^{-1}[\frac{\cosh \rho }{\sinh (z/\lambda )}] \label{figen}$$ where we define $$\begin{aligned} \lambda &=&1/\sqrt{A_1+2A_2} \\ \cosh \rho &=&\sqrt{1+\frac{2A_2}{A_1}}\end{aligned}$$ The soliton (\[figen\]) can be expressed as a superposition of two twisted $\pi $-Bloch walls [@Braun] with arguments $\lambda ^{-1}(z-z_0)\pm \rho $. We mention here there is another static solution of (\[mofi\]), the so- called [*nucleus*]{} [@Long] which correspond to a superposition of two untwisted $\pi $-Bloch walls [@Braun]. This solution is topologically distinct from the previous one. Dynamics of Bloch walls ======================= The quantum dynamics of our spin system can be analyzed by studying the quantum mechanics of the field theory described by the action $$S[\varphi ]=JS^{2}a\int_{-\infty }^{+\infty }\int_0^tdz\,dt\,\left\{ \frac 1{2c^2}% \left( \frac{\partial \varphi }{\partial t}\right) ^2-\ \frac 12\left( \frac{% \partial \varphi }{\partial z}\right) ^2-U(\varphi )\right\} . \label{action}$$ The next step is to quantize the system described by (\[action\]). The standard way to carry this program forward is to evaluate [@R.R.] $$G(t)=\text{tr}\int {\cal D}\varphi \ \exp \ \frac i\hbar S[\varphi ] \label{gt}$$ where the functional integral has the same initial and final configurations and tr means to evaluate it over all such configurations. As the functional integral in (\[gt\]) is impossible to be evaluated for a potential energy density as in (\[pot\]) we must choose an approximation to do it. Since we are already considering large spins ${(S\gg \ \hbar /2)}$, and consequently in the semi-classical limit, let us take this approximation as the appropriate one for our case. The semi-classical limit ($\hbar \rightarrow 0$) turns out to be very easily tractable within the functional integral formulation of quantum mechanics [@R.R.]. It is simply the stationary phase method applied to (\[gt\]). Moreover, since we are only interested in static solutions, the functional derivative of $S$ happens to be the equation of motion (\[mofi\]) when ${% \partial \varphi /\partial t=0}$. Its solutions can be either constant (uniform magnetization) or the solitons (Bloch walls) we mentioned in (\[fia10\]-\[figen\]). Since we are interested in studying the magnetic system in the presence of walls it is obvious that we must pick up one of those localized solutions as the stationary “point” in the configuration space and the second functional derivative of (\[pot\]) should be evaluated at this configuration. When this is done we are left with an eigenvalue problem that reads $$\left\{ -\frac{d^2}{dz^2}+U^{\prime \prime }(\varphi _s)\right\} \psi _n(z-z_0)=\kappa _n^2\psi _n(z-z_0) \label{Sle}$$ where $\varphi _s$ is denoting the soliton-like solution about which we are expanding $\varphi (z,t)$. Now one can easily show that $d\varphi _s/dz$ is a solution of (\[Sle\]) with $% \kappa _n=0$. The existence of this mode is related to the translation invariance of the Lagrangian in (\[action\]) and this makes the functional integral in (\[gt\]) blow up in the semi-classical limit (Gaussian approximation). The way out of this problem is the so-called collective coordinate method which was developed by field-theorists in the seventies (see [@R.R.] and references therein). It consists of keeping the expansion of the field configurations about $\varphi _s(z)$ as $$\varphi (z,t)=\varphi _s(z-z_0)+\sum_{n=1}^\infty c_n\psi _n\left( z-z_0\right) \label{expfi}$$ but regarding the $c$-number $z_0$ as a position operator. Eq. (11) is then substituted in the Hamiltonian $$H=JS^{2}a\int dx\left\{ \frac{c^{2}\Pi ^2}2+\ \frac 12 \left( \frac {d\varphi }{dz}\right) ^{2}+U(\varphi )\right\} ,$$ where ${\ \Pi =\frac {1}{c} (\frac{\partial \varphi }{\partial t})}$, which can be transformed into [@Castro] $$H=\frac 1{2M_s}\ (P-\sum_{mn}\hbar g_{mn}b{_n^{+}}b_m)^2+\sum \hbar \Omega _nb{_n^{+}}b_n. \label{Heff}$$ where $\Omega _{n}\equiv c\kappa _{n}$. In the Hamiltonian (\[Heff\]), $P$ stands for the momentum canonically conjugated to $z_0$, $$M_s=\frac {2JS^{2}a}{c^{2}}\int_{-\infty }^{+\infty }dzU(\varphi _s(z)) \label{sm}$$ is the soliton mass [@R.R.] and the coupling constants $g_{mn}$ are given by $$g_{mn}=\frac 1{2ia}\left[ \left( \frac{\Omega _m}{\Omega _n}\right) ^{1/2}+\left( \frac{\Omega _n}{\Omega _m}\right) ^{1/2}\right] \int dz\psi _m(z)\frac{d\psi _n(z)}{dz}. \label{gmn}$$ The operators $b^{+}$ and $b$ are respectively creation and annihilation operators for the excitations of the magnetic systems (magnons) in the presence of the wall. It should also be stressed that Eq. (\[Heff\]) is not an exact result. It is only valid in the limit $\hbar \rightarrow 0$ or, to be more precise, when $g^2\hbar \rightarrow 0$ where $g^2 \equiv \frac{1}{JS^{2}a}$ is the coupling constant that originally appears in $U(g,\varphi )$. It must also be emphasized that we have neglected inelastic terms such as $b^{+}b^{+}$ or $bb$ because these are only important if the wall moves at high speed $(v>c)$ originating Cherenkov-like radiation of the elementary excitations of the medium. This approximation also means that the number of excitations in the medium is conserved. Mobility of the Bloch wall ========================== At this point we are ready to start to study properties such as the mobility of the wall because we have been able to map that problem into the Hamiltonian (\[Heff\]) which on its turn has been recently used to study the mobility of polarons, heavy particles and solitons in general. We shall not discuss this specific problem in this paper and urge those interested in the details of this calculation to follow them in references [@CN; @Castro; @Neto; @CyN; @Des]. The result that can be obtained reads [@CyN] $$\gamma (t)=\frac \hbar {2M}\int_0^\infty \!\int_0^\infty d\omega d\omega ^{\prime }S(\omega ,\omega ^{\prime })(\omega -\omega ^{\prime })[n(\omega )-n(\omega ^{\prime })]\cos (\omega -\omega ^{\prime })t \label{g}$$ where $\gamma (t)$ is the damping function (the inverse of the mobility), $$n(\omega )=\frac 1{e^{\beta \hbar \omega }-1}$$ is the Bose function and $$S(\omega ,\omega ^{\prime })=\sum_{mn}|g_{mn}|^2\delta (\omega -\Omega _n)\delta (\omega ^{\prime }-\Omega _m) \label{sww}$$ is the so-called scattering function. In the long time limit $\gamma (t)$ can, to a good approximation, be written as $$\gamma (t)\cong \bar{\gamma}(T)\delta (t) \label{gtT}$$ and $\bar{\gamma}(T)$ is given by [@CyN] $$\bar{\gamma}(T)=\frac 1{2\pi M_s}\int_0^\infty dE\ {\cal R}(E)\frac{\beta E\ e^{\beta E}}{(e^{\beta E}-1)^2} \label{damp}$$ where ${\cal R}(E)$ is the reflection coefficient of the “potential” $% U^{\prime \prime }(\varphi _s)$ in the Schrödinger-like equation (\[Sle\]). Notice that (\[damp\]) is only valid if the states involved in (\[sww\]) are scattering states (see section [**V.A**]{} below for details). One important point that should be emphasized here is that there are parameters of the non-linear field equations for which the localized solutions render $U^{\prime \prime }(\varphi _s)$ a reflectionless potential. These are genuine solitons and for these the mobility is infinite. One may realize this is what happens for the Bloch walls (\[fia10\]) and (\[fia20\]). In these cases, the “potential” appearing in (\[Sle\]) can be written as $$U^{\prime \prime }(z)=\eta ^2(1-2% %TCIMACRO{\limfunc{sech} } %BeginExpansion \mathop{\rm sech} %EndExpansion \nolimits^2\eta z), \label{u0}$$ where $\eta ^2=A_1$ for vanishing anisotropy and $\eta ^2=2A_2$ for vanishing external field. The spectrum of (\[u0\]) contains a bound state with zero energy $$\psi _0=\sqrt{\frac \eta 2}% %TCIMACRO{\limfunc{sech} } %BeginExpansion \mathop{\rm sech} %EndExpansion (\eta z),\qquad \kappa _0^2=0 \label{gm}$$ which constitutes the translation mode of the domain wall (Goldstone mode), and a continuum of quasiparticles modes (magnons) given [@MF] by $$\psi_{n}(x)=\frac{1}{\sqrt{L}}\left[ \frac{k_{n}+i\eta \tanh(\eta z)}{k_{n}+i\eta} \right] e^{ik_{n}z}, \label{qp}$$ where $$k_{n}=\frac{2n\pi}{L}-\frac{\delta(k_{n})}{L}, \qquad \delta(k)=\arctan \left[ \frac{2\eta k}{k^{2}-\eta^{2}} \right].$$ It is known that the reflection coefficient ${\cal R}$ for a general symmetric potential can be expressed in terms of the corresponding phase shifts as [@CyN] $${\cal R}(k)=\sin ^2\left( \delta ^e(k)-\delta ^o(k)\right), \label{Rsin}$$ where $\delta ^e$ and $\delta ^o$ are the even and odd scattering phase shifts, respectively. Then, re-expressing (\[qp\]) in terms of even and odd defined parity states, it is easy to prove that this potential belongs to the class of reflectionless because its phase shifts are given by $$\delta ^{e,o}(k)=\arctan (\eta /k), \label{ps0}$$ that do not distinguish between odd and even parities. Nevertheless, when both the anisotropy and external field are finite, the reflection coefficient is nonvanishing and consequently the 2$\pi $-Bloch wall (\[figen\]) has a finite mobility. In this case, the spectral problem (\[Sle\]) can be rewritten as $$\left\{ -\frac{d^2}{dz^2}+V(z) \right\} \psi_n(z)=\kappa^2_n \psi_n(z), \label{Sle2}$$ where the potential $V(z)$ is expressed as $$\begin{aligned} V(z)&=&-\frac{ %TCIMACRO{\limfunc{sech} } %BeginExpansion \mathop{\rm sech} %EndExpansion \nolimits^2(\rho)}{\lambda^2} \left[\tanh(\frac{z}{\lambda}+\rho)\tanh(\frac{z}{\lambda}-\rho)- %TCIMACRO{\limfunc{sech} } %BeginExpansion \mathop{\rm sech} %EndExpansion \nolimits (\frac{z}{\lambda}+\rho) %TCIMACRO{\limfunc{sech} } %BeginExpansion \mathop{\rm sech} %EndExpansion \nolimits (\frac{z}{\lambda}-\rho)\right] \nonumber \\ &&-\frac{\tanh^2(\rho)}{\lambda^2} \left[\left( \tanh(\frac{z}{\lambda}+\rho)\tanh(\frac{z}{\lambda}-\rho)- %TCIMACRO{\limfunc{sech} } %BeginExpansion \mathop{\rm sech} %EndExpansion \nolimits (\frac{z}{\lambda}+\rho) %TCIMACRO{\limfunc{sech} } %BeginExpansion \mathop{\rm sech} %EndExpansion \nolimits (\frac{z}{\lambda}-\rho)\right)^{2} \right. \nonumber \\ &&-\left. %TCIMACRO{\limfunc{sech} } %BeginExpansion \mathop{\rm sech} %EndExpansion \nolimits ^2(\frac{z}{\lambda}+\rho) %TCIMACRO{\limfunc{sech} } %BeginExpansion \mathop{\rm sech} %EndExpansion \nolimits ^2(\frac{z}{\lambda}-\rho) \left(\sinh(\frac{z}{\lambda}+\rho)\sinh(\frac{z}{\lambda}-\rho) \right)^{2} \right]. \label{PotV}\end{aligned}$$ Now, for all finite values of $\lambda $ and $\rho$, the translational invariance of the system persists and as a consequence the potential (\[PotV\]) has a zero energy state that is given by $$\psi _0\propto %TCIMACRO{\limfunc{sech} } %BeginExpansion \mathop{\rm sech} %EndExpansion (\frac{z}{\lambda}+\rho )+% %TCIMACRO{\limfunc{sech} } %BeginExpansion \mathop{\rm sech} %EndExpansion (\frac{z}{\lambda}-\rho )\ , \label{fi0g}$$ which is nothing but the Goldstone mode of the Bloch walls for finite anisotropy and external field. In order to obtain an expression for the damping constant (\[damp\]) we need an expression for the odd and even phase shifts of (\[PotV\]). Unfortunately, their analytical evaluation is very complicated for all finite values of $\lambda $ and $\rho $, and in what follows we study the situation of large fields. 2$\pi $-Bloch walls for large fields ==================================== In this section we evaluate the scattering phase shifts in the situation of large external fields and provide an explicit expression for the damping constant. Scattering phase shifts ----------------------- In the case of large fields ($\rho \ll 1$) the Shrödinger-like equation (\[Sle2\]) can be written as $$\left\{ -\frac{d^2}{dz^2}+V(z) \right\} \psi_{n}(z)=(\kappa^{2}_{n}-\eta^{2} -\frac{\rho^{2}}{\lambda^{2}})\psi_{n}(z),$$ where the potential (\[PotV\]) is now reduced to the sum of the reflectionless contribution and a perturbation coming from the presence of the large field , explicitly $$V(z)=V_0(z)+\rho ^2V_1(z), \label{Vap}$$ where $$V_0(z)=-2% %TCIMACRO{\limfunc{sech} } %BeginExpansion \mathop{\rm sech} %EndExpansion \nolimits^2 \left( \frac{z}{\lambda} \right) \label{Vap0}$$ and $$V_1(z)=1-8\tanh ^2\left(\frac{z}{\lambda}\right) %TCIMACRO{\limfunc{sech} } %BeginExpansion \mathop{\rm sech} %EndExpansion \nolimits^2 \left( \frac{z}{\lambda} \right) . \label{Vap1}$$ In order to obtain the even and odd scattering phase shifts $\Delta ^{e,o}$ corresponding to a particle in a one dimensional symmetric potential like (\[Vap\]), we will use of a 1D version of the Fredholm theory [@Gott], which states that $$\pi A^{e,o}(E)\cot (\Delta ^{e,o})=1+{\cal P}\int_0^\infty dE\frac{% A^{e,o}(E^{\prime })}{E-E^{\prime }}, \label{piacot}$$ where $\Delta ^{e,o}$ are the phase shifts originated by both contributions, the first coming from the reflectionless potential $V_{0}$, and the other associated to the high field perturbation $V_{1}$. On the other hand the even and odd spectral functions, $A^{e,o}(E)$, can be calculated from the series expansion (see [@Gott] for details) $$A(E)=-\langle E | V(z) | E \rangle + {\cal P}\int_{0}^{\infty} \frac{dE_{1}}{E-E_{1}} \left| \begin{array}{cc} \langle E |V(z) |E \rangle & \langle E|V(z)|E_{1} \rangle \\ \langle E_{1}|V(z)|E\rangle & \langle E_{1}|V(z)|E_{1} \rangle \\ \end{array} \right|+\cdots \label{Aap}$$ where ${\cal P}$ stands for the Cauchy principal value. Clearly the expression (\[Aap\]) cannot be analytically evaluated to all orders. On the other hand, making use of (\[ps0\]) and considering that $\rho$ is small enough, the expression for the the phase shifts (\[piacot\]) can be written up to first order in $\rho^{2}$ as $$\tan\Delta^{e,o}=\frac{\eta}{k}+ \rho^2 \frac{\pi A^{e,o}_{1}} {1+2B^{e,o}_{0}}, \label{nova1}$$ where $$A_{1}=-\langle E |V_{1} (z)| E \rangle, \label{yyy}$$ and $$B^{e,o}_{0}={\cal P} \int_{-\infty}^{\infty} \frac{A^{e,o}_{0}(k^{\prime}) k^{\prime} dk^{\prime}}{k^2-k^{\prime 2}}, \qquad A_{0}=-\langle E|V_{0}(z)|E\rangle. \label{rmi}$$ Using a convenient basis set, the three expressions given by (\[yyy\]) and (\[rmi\]) can be analytically evaluated (see the apendix) yielding $$A^{e,o}_{0}(k)=\frac{2M}{\hbar^2}\left[ \frac {1}{\pi k \lambda } \pm \frac{1}{\sinh (\pi k \lambda)} \right], \label{aeok2}$$ $$A^{e,o}_{1}(k)=\frac{8 \rho^2 M}{\hbar^2}\left[ \frac {1}{\pi k \lambda } \mp (2k^2-\lambda^{-2})\frac{\lambda^2}{\sinh (\pi k \lambda)} \right], \label{aeok}$$ $$B^{e,o}= \pm \frac{4M}{\hbar^2} \sum _{n=1}^\infty (-1)^{n+1} \frac {n}{((k/\lambda)^{2} + n^2)}. \label{beo}$$ Now we can finally write down an expression for the phase shifts $\Delta^{e,o}$ by substituting (\[aeok\]) and (\[beo\]) in (\[nova1\]). In so doing one gets $$\tan \Delta ^{e,o}(k) = \frac{{\textstyle \eta}}{{\textstyle k}}+ \frac{ {\textstyle 8 \pi \hbar^{-2} M \rho^2 }}{ {\textstyle 1} \pm \frac{{\textstyle 8M}}{{\textstyle \hbar^2}} {\displaystyle \sum _{n=1}^\infty} {\textstyle (-1)^{n+1}} \frac {{\textstyle n}}{{\textstyle (k/\lambda)^{2} + n^2}}} \left[ \frac{{\textstyle 1}}{{\textstyle k \pi \lambda}} \mp \frac{{\textstyle 2k^2 \lambda^{2}-1}}{{\textstyle 3 \sinh(k \pi \lambda)}} \right] \label{deltaeo}$$ Looking at (51) we realize that, whereas $\Delta ^{e}(k)$ remains almost unchanged as a function of $k$ for $\rho \neq 0$ (see Fig.1), $\Delta ^{o}(k)$ presents a completely different structure. As it is shown in Fig.2, the odd phase shift for $\rho \neq 0$ display a $\pi$ amplitude discontinuity characteristic of a resonance. This discontinuity becomes sharper as $\rho \rightarrow 0$, giving no contribution to the reflection coefficient, because as expected from (\[ps0\]), the even and the odd phase shifts approach each other for small $\rho$ (see Fig.3). In this situation the $\pi$ amplitude discontinuity is still present, but as an isolated one, does not contribute to the reflection coefficient. We can now turn our atention to the behaviour of the phase shifts near $k=0$. In order to do this we will use the 1D version of the Levinson’s theorem [@Bart] for one-dimensional symmetric potentials. Levinson’s theorem establishes that: $$\begin{aligned} \Delta ^e(k &=&0)=\pi (n^e-\frac 12), \nonumber \\ \Delta ^o(k &=&0)=\pi n^o, \label{levi}\end{aligned}$$ where $n^e$ and $n^o$ are the number of even and odd parity bound states. As it can be seen in Fig. 1, the even parity state phase shift is always $\pi/2$ for $k=0$ for any finite value of $\rho $, then from (\[levi\]) it is clear that our system has an even parity bound state for any value of $\rho$. The existence of this state is in complete agreement with the translational invariance of the system and corresponds to the Goldstone mode. The analysis of the existence of an odd parity bound state is slightly different. As can be seen for (\[ps0\]) the correct definition of $\delta^o$ according to (\[levi\]) is $$\delta^{o}= \left\{ \begin{array}{ll} 0, & \mbox{ $ k=0 $ } \\ \arctan (\eta / k), & \mbox{$k \neq 0$} \end{array} \right. \label{pinky}$$ because there is no odd parity bound state for the $\rho=0$ case. Following the same idea and realizing that (\[deltaeo\]) behaves in the same way as (\[ps0\]) for small enough $k$, we conclude that $\Delta^o(0)=0$ and therefore there is no odd bound state in the $\rho \neq 0$ case. Actually, $\Delta^e(0) \approx \Delta^o(0)$ for large values of k whereas they start to deviate from one another as k decreases. Equation (\[pinky\]) represents an extreme situation when $\delta^o(k)$ discontinously jumps from $\delta^e(k)$ to $\delta^o(k)=0$ when $k \rightarrow 0$. This is the only way we can reconcile the absence of an odd bound state and the reflectiolessness of the potential. Therefore, the spectrum of (\[Vap\]) is composed by: i) the $\psi _0$ solution (\[fi0g\]) corresponding to the translation mode of the wall (Goldstone mode) and ii) the $\psi _k$ solutions which constitute the continuum modes and correspond to magnons. The damping coefficient ----------------------- In order to find the damping coefficient we must compute ${\cal R}(k)$. This can be done by inserting (\[deltaeo\]) into the general expression $${\cal R}(k)=\sin^{2}(\Delta ^{e}(k)-\Delta ^{o}(k)).$$ Although we have its analytical form in the present approximation, we had better plot it for the whole range of the momentum $ k$ for various ratios of $ A_{2}/A_{1}$ (anisotropy/external field) (see Fig. 4). Having done that, one can immediately integrate this function in expression (\[damp\]) which finally allows us to describe the damping as a function of the temperature as shown in Fig.5. As it can be seen, the damping constant is linear for the whole range of temperatures. This result can be obtained directly from (\[damp\]). For $T$ high enough the damping constant can be approximated by $$\bar{\gamma}(T)\simeq \frac 1{2\pi M_s\beta }\int_0^\infty dE\ \frac{{\cal R}% (E)}E\propto T \label{g1}$$ which is linear on $T$, independently of the explicit form of ${\cal R}(E)$. In the low temperature regime we can write $$\bar{\gamma}(T)\simeq \frac 1{2\pi M_s}\int_0^\infty dE\ {\cal R}(E)\beta Ee^{-\beta E} \label{aoc}$$ where $E$ always presents a gap given by $E^2_{g}=[(\eta\lambda)^2+\rho^2]\hbar^2c^2/\lambda^2$ and therefore and exponentially small damping. However, this expression must be studied in the limit in which we are interested; namely, $\rho \rightarrow 0$. Unlike $\bar{\gamma}(T)$ given by (\[g1\]) the expression (\[aoc\]) cannot be analitically estimated in a trivial way because ${\cal R}(E)$ has only been computed numerically. Here it must be stressed that this result is only reliable for not too low temperatures because we have employed the odd phase shift $\Delta^o(k)$ from Fig.2 and its computation clearly do not account for its correct values for $k \rightarrow 0$. This behaviour is due to the fact that for very low energies one can not approximate $A(E)$ in (\[Aap\]) by its first term although this approximation works well for the even phase shift $\Delta^e(k)$. This turns out to be a good description of the mobility of the wall for extremely high fields if we keep the above explanation in mind. Conclusions =========== In this paper we have analyzed the dissipative dynamics of Bloch walls in a one dimensional anisotropic ferromagnet in the presence of an external magnetic field. In particular, we have considered the limit of high magnetic fields although there is no reason why one should not apply the same methods to the low field case. The only difference is that the scattering problem with which one has to deal is more straightforward in the high field case. Our predictions are that the damping coefficient $\bar{\gamma}(T)$ presents a linear behaviour as is plotted in Fig.5, for a vast range of temperatures. Acknowledgments =============== One of us (M.A.D.) would like to acknowledge the financial support from FAEP (Fundo de Apoio ao Ensino e Pesquisa) during his visit to UNICAMP. A.V.F. wishes to thank FAPESP (Fundação de Amparo à Pesquisa no Estado de São Paulo) for a scholarship, A.O.C. kindly acknowledges the partial support from the CNPq (Conselho Nacional de Desenvolvimento Cientifíco e Tecnológico) whereas A.H.C.N. acknowledges support from the A.P. Sloan foundation and support provided by the DOE for research at Los Alamos National Laboratory. Appendix ======== In this appendix we show how to obtain the expressions of the even and odd spectral functions (\[rmi\]) and (\[yyy\]). Suppose we have a particle in a one dimensional symmetric potential of the form $V=V_0+gV_1$ confined to a region ($-L,+L)$ with $L$ much larger than the range of the potential $V$. The asymptotic form of the wave functions for $V \neq 0$ are given by $$\begin{aligned} |z\rangle ^e &=&\sqrt{\frac 1L}\cos (k|z|+\Delta ^e(k)), \nonumber \\ |z\rangle ^o &=&\sqrt{\frac 1L}% %TCIMACRO{\limfunc{sgn} } %BeginExpansion \mathop{\rm sgn} %EndExpansion (z)\sin (k|z|+\Delta ^o(k)), \label{figa}\end{aligned}$$ for $|z|\rightarrow \infty$. If $V=0$ the wave functions have the same structure as in (\[figa\]) with $\Delta ^{e,o}=0$. Because the wave functions must vanish at $z=\pm L$ one realizes that $$\frac{\delta E_n}{\Delta E_n}=-\frac 1\pi \Delta ^{e,o}$$ where $\Delta E_n=E_{n+1}^0-E_n^0$ and $\delta E_n=E_n-E_n^0.$ Following closely the prescriptions given in [@Gott] for the 1-D case, the spectral functions (\[rmi\]) and (\[yyy\]) are given by $$A^{e,o}_{i}(E)=-\int_{-\infty }^{+\infty }dzV_i(z)\langle z|E\rangle _{e,o}\langle E|z\rangle _{e,o}, \qquad i=0,1 \label{Aap2}$$ where the states $\langle E|z\rangle_{e,o}$ can be obtained by taking $$\langle z|E\rangle_{e,o}=\lim_{L \rightarrow \infty}\frac{|z \rangle}{\sqrt{\Delta E_{n}}}.$$ Explicitly, $$\langle z|E\rangle =\frac 1{\sqrt{2\pi k}} \left\{ \begin{array}{ll} \cos (kx), & \mbox{ for even parity} \\ \sin (kx), & \mbox{for odd parity} \end{array} \right. \label{zeo}$$ Inserting (\[zeo\]) and (\[Vap1\]) in (\[Aap2\]) we have $$A_{1}^{e}+A_{1}^{o}=\frac{8M\rho ^2}{\pi \hbar^2 k} \int_{-\infty }^{+\infty} %TCIMACRO{\limfunc{sech} } %BeginExpansion \mathop{\rm sech} %EndExpansion \nolimits ^2(\frac{z}{\lambda}) \tanh ^2 (\frac{z}{\lambda})dz,$$ which can be easily evaluated with the substitution $y=\tanh z/\lambda$, yielding $$A_{1}^{e}+A_{1}^{o}=\frac {16 \rho ^2 M}{ \pi \hbar^2 \lambda k}. \label{amas}$$ On the other hand we have $$A_{1}^{e}-A_{1}^{o}=\frac{8M\rho ^2}{\pi \hbar^2 k} \int_{-\infty }^{+\infty } %TCIMACRO{\limfunc{sech} } %BeginExpansion \mathop{\rm sech} %EndExpansion \nolimits ^2(\frac{z}{\lambda}) \tanh ^2 (\frac{z}{\lambda}) \cos(2k \lambda)dz,$$ that can be writen as $$A_{1}^{e}-A_{1}^{o}=\frac{8M\rho ^2}{\pi \hbar^2 k} \int_{-\infty }^{+\infty } \left( %TCIMACRO{\limfunc{sech} } %BeginExpansion \mathop{\rm sech} %EndExpansion \nolimits ^2(\frac{z}{\lambda})- %TCIMACRO{\limfunc{sech} } %BeginExpansion \mathop{\rm sech} %EndExpansion \nolimits ^4(\frac{z}{\lambda}) \right) \cos(2k \lambda)dz$$ which can be analytically evaluated [@Grad] yielding $$A_{1}^{e}-A_{1}^{o}=\frac {16 \rho ^2 M}{ \hbar^2 \sinh (\pi k \lambda)} \left[\frac{1}{3}-\frac{2 k^2 \lambda^2}{3} \right]. \label{amen}$$ Therefore combining (\[amas\]) and (\[amen\]) we have (\[aeok\]). In the same fashion it can be shown that $$A_{0}^{e}+A_{0}^{o}=\frac {4 M}{ \pi \hbar^2 k \lambda} \qquad $and$ \qquad A_{0}^{e}-A_{0}^{o}=\frac {{\textstyle 4 M}}{ {\textstyle \hbar^2 \sinh ( \pi k \lambda)}}$$ which immediately gives (\[aeok2\]). Now we can evaluate the Cauchy principal value in (\[rmi\]) which reads $$B_{0}^{e,o}=\frac{2M}{\hbar^2}{\cal P}\int_{-\infty}^{+\infty} \left[\frac{1}{\pi k^{\prime} \lambda} \pm \frac{1}{\sinh (\pi k^{\prime} \lambda)} \right] \frac{k^{\prime} dk^{\prime}}{k^2-k^{\prime 2}}. \label{bras}$$ The first term on the right hand side of (\[bras\]) is clearly zero. Therefore, using the product expansion of $\sinh (\pi z)$ fuction [@Grad] its second term becomes $$B_{0}^{e,o}=\pm \frac{2M}{\pi \hbar^2}{\cal P}\int_{-\infty}^{+\infty}\frac{dq}{k^2-q^2} \prod_{n=1}^{\infty}\frac{n^2}{n^2+q^2},$$ where $q=k/\lambda$. Going to the complex plane, the previous expression can be analitically evaluated as $$B_{0}^{e,o}=\pm \frac{4M}{\hbar^2}\sum_{n=1}^{\infty}\frac{(-1)^n n}{(k/\lambda)^2+n^2}.$$ L. D. Landau and E. M. Lifshitz in “Electrodynamics of the Continuous Media,” [*Pergamon Press Ltd.*]{}, (1960). J. F. Dillon Jr. in “Magnetism,” ed. by G. T. Rado and H. Suhl, [*Academic Press Inc. Vol. III*]{}  (1963). L. D. Landau and E. M. Lifshitz, “Physik A,” [*Soviet Union  8*]{}, 153, (1935). A. P. Malozemoff and J. C. Slonczewski in “Magnetic Domain Walls in Bubble Materials,” [*Academic Press Inc.*]{}, (1979) A. H. Castro Neto and A. O. Caldeira [*Phys. Rev. B, [**46**]{}*]{} 8858 (1992). A. H. Castro Neto and A. O. Caldeira [*Phys. Rev.B, [**50**]{}*]{} 4863 (1994). R. Rajaraman in “Solitons and Instantons,” (North Holland, Amsterdam, 1992). P.C.E. Stamp, E. M. Chudnovsky and B. Barbara, [*Int. J. Modern Physics  B6.*]{}, 1355, (1992). J.M.Winter, [*Phys. Rev*]{} [**124**]{}, 452 (1962) H. Braun, [*Phys. Rev. B*]{} [**50**]{}, 16485 (1994) K.A.Long and A.R.Bishop,[* J.Phys.A* ]{}[**12**]{}, 1325 (179) A. H. Castro Neto and A. O. Caldeira [*Phys. Rev. E, [**48**]{}*]{} 4037 (1993). A. H. Castro Neto and A. O. Caldeira [*Phys. Rev. Lett, [**67**]{}*]{} 1960 (1991). A. O. Caldeira and A. H. Castro Neto [*Phys. Rev. B, [**52**]{}*]{} 4198 (1995). M.A.Despósito, [*J. Phys. A*]{} [**29**]{}, 1147 (1996). P.M. Morse and H. Feshbach, Methods of Theoretical Physics (McGraw-Hill, New York, 1953) K. Gottfried “Quantum Mechanics: Vol I” (W.A.Benjamin Inc, New York, 1966) G. Barton, [*J.Phys.A*]{} [**18**]{} 479 (1985) I.S.Gradsteyn and I.M.Ryshik, “Table of Integrals, Series and Products” (Academic Press, New York, 1965). to 9.0cm [to 9cm [ ]{} ]{} to 9.0cm [to 9cm [ ]{} ]{} to 9.0cm [to 9cm [ ]{} ]{} to 9.0cm [to 9cm [ ]{} ]{} to 9.0cm [to 9cm [ ]{} ]{} [^1]: E-mail: [email protected]

ArXiv

--- author: - 'Victor Kalvin, Alexey Kokotov' title: Determinant of Laplacian on tori of constant positive curvature with one conical point --- We find an explicit expression for the zeta-regularized determinant of (the Friedrichs extensions) of the Laplacians on a compact Riemann surface of genus one with conformal metric of curvature $1$ having a single conical singularity of angle $4\pi$. Introduction ============ Let $X$ be a compact Riemann surface of genus one and let $P\in X$. According to [@CLW], Cor. 3. 5. 1, there exists [*at most*]{} one conformal metric on $X$ of constant curvature $1$ with a (single) conical point of angle $4\pi$ at $P$. The following simple construction shows that such a metric, $m(X, P)$, in fact always exists (and due to [@CLW] is unique). Consider the spherical triangle $T=\{(x_1, x_2, x_3)\in S^2\subset {\mathbb R}^3: x_1\geq 0, x_2\geq 0, x_3\geq 0\}$ with all three angles equal to $\pi/2$. Gluing two copies of $T$ along their boundaries, we get the Riemann sphere ${\mathbb P}$ with metric $m$ of curvature $1$ and three conical points $P_1, P_2, P_3$ of conical angle $\pi$. Consider the two-fold covering $$\mu: X(Q) \to {\mathbb P}$$ ramified over $P_1$, $P_2$, $P_3$ and some point $Q\in {\mathbb P}\setminus \{P_1, P_2, P_3\}$. Lifting the metric $m$ from ${\mathbb P}$ to the compact Riemann surface $X(Q)$ of genus one via $\mu$, one gets the metric $\mu^*m$ on $X(Q)$ which has curvature $1$ and the unique conical point of angle $4\pi$ at the preimage $\mu^{-1}(Q)$ of $Q$. Clearly, any compact surface of genus one is (biholomorphically equivalent to) $X(Q)$ for some $Q\in {\mathbb P}\setminus \{P_1, P_2, P_3\}$. Now let $X$ be an arbitrary compact Riemann surface of genus one and let $P$ be any point of $X$. Take $Q\in {\mathbb P}$ such that $X=X(Q)$ and consider the automorphism $\alpha:X\to X$ (the translation) of $X$ sending $P$ to $\mu^{-1}(Q)$. Then $$m(X, P)=\alpha^*(\mu^*(m))=(\mu\circ \alpha)^*(m)\,.$$ Now introduce the scalar (Friedrichs) self-adjoint Laplacian $\Delta(X, P):=\Delta^{m(X, P)}$ on $X$ corresponding to the metric $m(X, P)$. For any $P$ and $Q$ from $X$ the operators $\Delta(X, P)$ and $\Delta(X, Q)$ are isospectral and, therefore, the $\zeta$-regularized (modified, i. e. with zero modes excluded) determinant ${\rm det}\Delta(X, P)$ is independent of $P\in X$ and, therefore, is a function on moduli space ${\cal M}_1$ of Riemann surfaces of genus one. The main result of the present work is the following explicit formula for this function: $$\label{result} {\rm det}\Delta(X, P) = C_1\,|\Im \sigma||\eta(\sigma)|^4F(t)=C_2\,{\rm det}\Delta^{(0)}(X)F(t),$$ where $\sigma$ is the $b$-period of the Riemann surface $X$, $C_1$ and $C_2$ are absolute constants, $\eta$ is the Dedekind eta-function, $\Delta^{(0)}$ is the Lapalacian on $X$ corresponding to the flat conformal metric of unit volume, the surface $X$ is represented as the two-fold covering of the Riemann sphere ${\mathbb C}P^1$ ramified over the poits $0, 1, \infty$ and $t\in {\mathbb C}\setminus \{0, 1\}$, and $$F(t)=\frac{|t|^{\frac{1}{24}}|t-1|^{\frac{1}{24}}}{(|\sqrt{t}-1|+|\sqrt{t}+1|)^{\frac{1}{4}}}\,.$$ As it is well-known, the moduli space ${\cal M}_1$ coincides with the quotient space $$\left( {\mathbb C}\setminus \{0, 1\}\right)/G\,,$$ where $G$ is a finite group of order $6$, generated by transformations $t\to \frac{1}{t}$ and $t\to 1-t$. A direct check shows that $F(t)=F(\frac{1}{t})$ and $F(t)=F(1-t)$ and, therefore, the right hand side of (\[result\]) is in fact a function on ${\cal M}_1$. Using the classical relation (see, e. g. [@Clemens] (3.35)) $$t=-\left(\frac{\Theta[^1_0](0\,|\,\sigma)}{\Theta[^0_1](0\,|\, \sigma)}\right)^4\,,$$ one can rewrite the right hand side as a function $\sigma$ only. The classical (see [@Polch]) relation ${\rm det}\Delta^{(0)}=C\,|\Im \sigma||\eta(\sigma)|^4$ used in (\[result\]), implies that (\[result\]) can be considered as a version of Polyakov’s formula (relating determinants of the Laplacians corresponding to two [*smooth*]{} metrics in the same conformal class) for the case of two conformally equivalent metrics on a torus: one of them is smooth and flat, another is of curvature one and has one (very special) singular point. Metrics on the base and on the covering ======================================= Here we find an explicit expression for the metric $m$ on the Riemann sphere ${\mathbb P}={\mathbb C}P^1$ of curvature $1$ and with three conical singularities at $P_1=0$, $P_2=1$ and $P_3=\infty$. The stereographic projection (from the south pole) maps the spherical triangle $T$ onto quarter of the unit disk $\{z\in {\mathbb C}; |z|\leq 1,\ 0\leq {\rm Arg}\, z\leq \pi/2\}$. The conformal map $$\label{map} z\mapsto w=\left(\frac{1+z^2}{1-z^2}\right)^2$$ sends this quarter of the disk to the upper half-plane $H$; the corner points $i, 0, 1$ go to the points $0, 1$ and $\infty$ on the real line. The push forward of the standard round metric $$\frac{4|dz|^2}{(1+|z|^2)^2}$$ on the sphere by this map gives rise to the metric $$\label{basemetric}m=\frac{|dw|^2}{|w||w-1|(|\sqrt{w}+1|+|\sqrt{w}-1|)^2}$$ on $H$; clearly, the latter metric can be extended (via the same formula) to ${\mathbb C}P^1$. The resulting curvature one metric on ${\mathbb C}P^1$ (also denoted by $m$) has three conical singularities of angle $\pi$: at $w=0$, $w=1$ and $ w=\infty$. Consider a two-fold covering of the Riemann sphere by a compact Riemann surface $X(t)$ of genus $1$ $$\label{cov} \mu: X(t)\to {\mathbb C}P^1$$ ramified over four points: $0, 1, \infty$ and $t\in {\mathbb C}\setminus \{0, 1\}$. Clearly, the pull back metric $\mu^*m$ on $X(t)$ is a curvature one metric with exactly one conical singularity. The singularity is a conical point of angle $4\pi$ located at the point $\mu^{-1}(t)$. Determinant of Laplacian as function of critical value $t$ ========================================================== The analysis from [@KK] in particular implies that one can introduce the standard Ray-Singer $\zeta$-regularized determinant of the (Friedrichs) self-adjoint Laplacian $\Delta^{\mu^*m}$ in $L_2(X(t), \mu^*m)$ $${\rm det}\,\Delta^{\mu^*m}:=\exp\{-\zeta'_{\Delta^{\mu^*m}}(0)\}\,,$$ where $\zeta'_{\Delta^{\mu^*m}}$ is the operator zeta-function. In this section we establish a formula for the variation of $\zeta'_{\Delta^{\mu^*m}}(0)$ with respect to the parameter $t$ (the fourth ramification point of the covering . The derivation of this formula coincides almost verbatim with the proof of [@KK Proposition 6.1], therefore, we will give only few details. For the sake of brevity we identify the point $t$ of the base ${\mathbb C}P^1$ with its (unique) preimage $\mu^{-1}(t)$ on $X(t)$. Let $Y(\lambda; \,\cdot\, )$ be the (unique) special solution of the Helmholz equation (here $\lambda$ is the spectral parameter) $(\Delta^m -\lambda)Y=0$ on $X\setminus \{t\}$ with asymptotics $Y(\lambda) (x)=\frac{1}{x}+O(x)$ as $x\to 0$, where $x(P)=\sqrt{\mu(P)-t}$ is [ it the distinguished holomorphic local parameter]{} in a vicinity of the ramifiction point $t\in X(t)$ of the covering (\[cov\]). Introduce the complex-valued function $\lambda\mapsto b(\lambda)$ as the coefficient near $x$ in the asymptotic expansion $$Y(x, \bar x; \lambda)=\frac{1}{x}+c(\lambda)+a(\lambda)\bar x+b(\lambda)x +O(|x|^{2-\epsilon}) \text {\ as } x\to 0.$$ The following variational formula is proved in [@KK Propositon 6.1]: $$\label{var1} \partial_t (-\zeta'_{\Delta^{\mu^*m}}(0))=\frac{1}{2}\left(b(0)-b(-\infty)\right).$$ The value $b(0)$ is found in [@KK Lemma 4.2]: one has the relation $$\label{me}b(0)=-\frac{1}{6}S_{Sch}(x)\Big|_{x=0},$$ where $S_{Sch}$ is the Schiffer projective connection on the Riemann surface $X(t)$. Since $\lambda=-\infty$ is a local regime, in order to find $b(-\infty)$ the solution $Y$ can be replaced by a local solution with the same asymptotic as $x\to 0$. A local solution $\widehat Y$ with asymptotic $$\widehat Y(u, \bar u; \lambda)=\frac{1}{u}+\hat c(\lambda)+\hat a(\lambda)\bar u+\hat b(\lambda)u +O(|u|^{2-\epsilon}) \text {\ as } u\to 0$$ in the local parameter $u^2=z-s$ was constructed in [@KK Lemma 4.1] by separation of variables; here $z$ and $w=\mu(P)$ (resp. $s$ and $t$) are related by  (resp. by  with $z=s$ and $w=t$) and $\hat b(-\infty)=\frac{1}{2}\frac{\bar s}{1+|s|^2}$. One can easily find the coefficients $A(t)$ and $B(t)$ of the Taylor series $u=A(t)x+B(t)x^3+O(x^5)$. As a local solution replacing $Y$ we can take $A(t)\widehat Y$. This immediately implies $b(-\infty)=A^2(t)\hat b(-\infty)-B(t)/A(t)$. A straightforward calculation verifies that $$\label{Victor} b(-\infty)=\partial_t \log \left ({|t||t-1|(|\sqrt{t}+1|+|\sqrt{t}-1|)^2}\right )^{1/4}.$$ Observe that the right hand side in  is actually the value of $\partial_w \log \rho(w, \bar w)^{-1/4}$ at $w=t$, where $\rho(w, \bar w)$ is the conformal factor of the metric ; this is also a direct consequence of [@VK Lemma 4]. Using (\[var1\]) together with (\[me\]) and (\[Victor\]), we are now able to derive an explicit formula for ${\rm det}\Delta^{\mu^*m}$. Explicit formula for the determinant ==================================== Equations (\[var1\]), (\[me\]) and (\[Victor\]) imply that the determinant of the Laplacian ${\rm det}\,\Delta^{\mu^*m}=\exp\{-\zeta'_{\Delta^{\mu^*m}}(0)\}$ can be represented as a product $$\label{prelim} {\rm det}\,\Delta^{\mu^*m}=C\,|\Im \sigma||\tau(t)|^2\left| \frac{1}{|t||t-1|(|\sqrt{t}+1|+|\sqrt{t}-1|)^2}\right|^{1/8}\,$$ where $\tau(t)$ is the value of the Bergman tau-function (see [@KokKor1], [@KokKor2], [@KokStra]) on the Hurwitz space $H_{1, 2}(2)$ of two-fold genus one coverings of the Riemann sphere, having $\infty$ as a ramification point at the covering, ramified over $1, 0, \infty$ and $t$. More specifically, $\tau$ is a solution of the equation $$\frac{\partial \log \tau}{\partial t}=-\frac{1}{12}S_B(x)|_{x=0}\,$$ where $S_B$ is the Bergman projective connection on the covering Riemann surface $X(t)$ of genus one and $x$ is the distinguished holomorphic parameter in a vicinity of the ramification point $t$ of $X(t)$. We remind the reader that the Bergman and the Schiffer projective connections are related via the equation $$S_{Sch}(x)=S_B(x)-6\pi(\Im\sigma)^{-1} v^2(x)\,$$ where $v$ is the normalized holomorphic differential on $X(t)$ and that the Rauch variational formula (see, e. g., [@KokKor1]) implies the relation $$\frac {\partial\log \Im \sigma}{\partial t}=\frac{\pi}{2}(\Im \sigma)^{-1} v^2(x)|_{x=0}\,.$$ The needed explicit expression for $\tau$ can be found e. g. in [@KokStra f-la (18)] (it is a very special case of the explicit formula for the Bergman tau-function on general coverings of arbitrary genus and degree found in [@KokKor2] as well as of a much earlier formula of Kitaev and Korotkin for hyperelliptic coverings [@KitKor]). Namely, [@KokStra f-la (18)] implies that $$\label{tau} \tau=\eta^2(\sigma)\left[\frac{v(\infty)^3}{v(P_1)v(P_2)v(Q)}\right]^{\frac{1}{12}},$$ where $P_1$ and $P_2$ are the points of the $X(t)$ lying over $0$ and $1$, $Q$ is a point of $X(t)$ lying over $t$ and $\infty$ denotes the point of the covering curve $X(t)$ lying over the point at infinity of the base ${\mathbb C}P^1$; $v$ is an arbitrary nonzero holomorphic differential on $X(t)$; and, say, $v(P_1)$ is the value of this differential in the distinguished holomorphic parameter at $P_1$. (One has to take into account that $\tau=\tau_I^{-2}$, where $\tau_I$ is from [@KokStra].) Taking $$v=\frac{dw}{\sqrt{(w(w-1)(w-t)}}\,,$$ and using the following expressions for the distinguished local parameters at $P_1$, $P_2$, $Q$ and $\infty$ $$x=\sqrt{w}; \ \ x=\sqrt{w-1};\ \ x=\sqrt{w-t}; \ \ x=\frac{1}{\sqrt w}$$ one arrives at the relations (where $\sim$ means $=$ up to insignificant constants like $\pm 2$, etc.) $$v(P_1)\sim\frac{1}{\sqrt{t}}; \ \ v(P_2)\sim\frac{1}{\sqrt{t-1}}; \ \ v(Q)\sim\frac{1}{\sqrt{t(t-1)}}; \ \ v(\infty)\sim1.$$ These relations together with (\[tau\]) and (\[prelim\]) imply (\[result\]). [100]{} Ching-Li Chai, Chang-Shou Lin, Chin-Lung Wang, Mean field equation, hyperelliptic curves and modular forms: I, Cambridge Journal of Mathematics, Vol. 3, N 1-2, 2015 C. Clemens, A scrapbook of complex curve theory, Grad. Studies in Math., Vol 55 V. Kalvin, A. Kokotov, Metrics of constant positive curvature, Hurwitz spaces and ${\rm det \Delta}$, IMRN, 2018; in press; arXiv:1612.08660 A. Kokotov, D. Korotkin, Tau-functions on Hurwitz spaces, Mathematical Physics, Analysis and Geometry, 7 (2004), no. 1, 47–96. A. Kokotov, D. Korotkin, Isomonodromic tau-function of Hurwitz Frobenius manofolds, Int. Math. Res. Not. IMRN (2006), pp. 1-34 A.Kokotov, I. Strachan, On the isomonodromic tau-function for the Hurwitz spaces of branched coverings of genus zero and one, Mathematical Research Letters, 12, 2005, no. 5-6, 857–-875. V. Kitaev, D. Korotkin, On solutions of the Schlesinger equations in terms of theta-functions, International Mathematics Research Notices, 1998, no. 17,877–-905. V. Kalvin, On Determinants of Laplacians on Compact Riemann Surfaces Equipped with Pullbacks of Conical Metrics by Meromorphic Functions, in preparation J. Polchinski, Evaluation of the one loop string path integral. Comm. Math. Phys. 104 (1986), no. 1, 37–47

ArXiv

--- abstract: 'A Fermi’s Golden Rule for population transfer between instantaneous eigenstates of elliptical quantum billiards with oscillating boundaries is derived. Thereby, both the occurrence of the recently observed resonant population transfer between instantaneous eigenstates \[F. Lenz et al. New J. Phys., [**13**]{}, 103019, 2011\] and the empirical criterion stating that these transitions occur when the driving frequency matches the mean difference of the latter are explained. As a second main result a criterion judging which resonances are resolvable in a corresponding experiment of certain duration is provided. Our analysis is complemented by numerical simulations for three different driving laws. The corresponding resonance spectra are in agreement with the predictions of both criteria.' author: - Jakob Liss - Benno Liebchen - Peter Schmelcher title: 'Analysis of resonant population transfer in time-dependent elliptical quantum billiards' --- Introduction ============ Classical driven billiards of varying geometry have been subject to intensive research during the last years [@Loskutov1; @Loskutov2; @Leonel1; @Leonel2; @Leonel3; @Leonel4; @Leonel5; @Matrasulov1; @ClEllBill1; @ClEllBill2; @ExpFermiOrig; @BennoExpFermi; @LRA]. While billiards are, in general, important models to study aspects of nonlinear dynamics, semi-classics or (quantum) chaos [@stockmann; @reichl], driven billiards additionally facilitate the study of non-equilibrium dynamics. As one of the key topics concerning driven billiards, Fermi acceleration (FA) and the related conditions for its occurrence have gained much attention [@Loskutov1; @ExpFermiOrig; @BennoExpFermi; @Leonel3; @Leonel4; @Leonel5; @ClEllBill1; @ClEllBill2; @LRA]. FA describes the unbounded growth of energy of particles that repeatedly interact with a time-dependent potential that is usually modeled by a moving billiard boundary and was originally proposed by Enrico Fermi as a possible mechanism to explain high-energetic cosmic radiation [@Fermi]. The infamous Fermi-Ulam model (FUM) is basically a one-dimensional billiard with a moving boundary and it was found that FA is present in the FUM only for non-smooth driving laws [@liebermann]. The general conditions for the occurrence of FA are still under debate. Originally, it was assumed that a sufficient condition for the occurrence of FA in a driven two-dimensional billiard is the presence of chaotic regions in the phase space of the corresponding static billiard [@LRA]. However, it turned out, that driving an ovally shaped billiard which has a mixed phase space in a certain mode does not lead to FA [@PhdLenz]. On the other hand, it was shown that the classical driven elliptical billiard does show FA although its static counterpart is completely integrable [@ClEllBill1; @ClEllBill2]. Furthermore, while correlated motion suppresses FA for smooth driving in the FUM, it was found that correlations can even cause exponential FA for smooth driving laws in a related two-dimensional model [@ExpFermiOrig; @BennoExpFermi]. Although it is known that periodically driven quantum billiards with a discrete Floquet spectrum can not exhibit FA [@BoundedEnergy], it is natural to complement the study of the classical dynamics of a system by analyzing its quantum behavior. While one finds many studies to the quantum version of the one-dimensional FUM (see [@QFUM1; @QFUM2; @QFUM3] and references therein), literature is very sparse on driven quantum billiards of higher dimensions [@Flrespop; @DQB1; @DQB2]. In particular, [@Flrespop] presents a method to solve the time-dependent elliptical quantum billiard. The main result was the numerical observation of resonances in the population transfer probability between instantaneous energy eigenstates. These transitions could be reproduced in an effective Rabi-model and captured by a criterion stating that resonances occur whenever the difference of corresponding time-averaged energy eigenvalues matches an integer multiple of the driving frequency. However, an explanation for this criterion was not given in [@Flrespop]. Here, we develop a systematic perturbative analysis of population transfer for the system analyzed in [@Flrespop] and a generalized driving law. In this framework a Fermi’s Golden Rule [@goldenRule] is derived for elliptical quantum billiards with oscillating boundaries which explains the key observations in [@Flrespop], i.e. the occurrence of resonant population transfer between instantaneous eigenstates and the empirical criterion relating these resonances with the spectrum of instantaneous eigenstates and the driving frequency. As a second major result, we provide a criterion to decide whether a predicted resonance can be resolved in a possible ’experiment’ of a certain duration. Finally, the numerical studies in [@Flrespop] are complemented by a corresponding analysis of further driving laws. The predictions derived within our perturbative analysis will be shown to provide a perfect agreement with the numerical results in all cases. This work is structured as follows: Chapter \[ch:setup\] provides a short summary of the solution of the time-dependent elliptical quantum billiard as developed in [@Flrespop], followed by transformations that bring the Schrödinger equation into a form being convenient for the application of time-dependent perturbation theory. In chapter \[ch:perturbative\_analysis\], we will calculate the transition rate between two instantaneous eigenstates per unit time is calculated in first order perturbation theory and an approximate population dynamics in the near-resonant case is derived. We find a criterion for the resolvability of predicted resonances in a possible experiment of certain duration. Finally, in chapter \[ch:numerical\_results\] we present and analyze numerical results for three different periodic driving laws. Time-dependent elliptical billiard and its analytic treatment ============================================================= In the following, we first summarize our approach to a numerical solution of the time-dependent Schrödinger equation of the elliptical billiard as presented in [@Flrespop]. We will transform the time-dependent Schrödinger equation (TDSE) into a convenient form and finally develop a perturbative approach for the periodically driven billiard in the second part of this chapter. Setup {#ch:setup} ----- A wave function $\Psi(\vec{x},t)$ in a driven elliptical billiard obeying the TDSE $$\dot \imath \hbar \partial_t \Psi(\vec{x},t) = -\frac{\hbar^2}{2 \mu} \Delta \Psi(\vec{x},t) \textrm{,} \label{Eq:SE1}$$ is subject to Dirichlet boundary conditions, $\Psi(\vec{x},t)|_{\partial B} = 0$, on a boundary $\partial B$ of elliptical shape: $$\partial B = \left\{\vec{x} = (x,y)^{\intercal} \in \mathbb{R}^2 \left| \frac{x^2}{a^2} + \frac{y^2}{b^2} = 1\right.\right\} \textrm{.}$$ Here, the semi-axes of the elliptical boundary, $a$ and $b$, are assumed to be arbitrary smooth functions of time, i.e. $a=a(t)$ and $b=b(t)$. The time-dependent boundary conditions can be handled by a coordinate transformation [@Flrespop], $$\rho_t: \left(\begin{array}{c} x\\ y \end{array}\right) \mapsto \left(\begin{array}{c} \eta\\ \xi \end{array}\right) = \left(\begin{array}{cc} \frac{1}{a(t)} & 0\\ 0 & \frac{1}{b(t)} \end{array}\right) \left(\begin{array}{c} x\\ y \end{array}\right) \label{Eq:coordtrafo}$$ that maps the time-dependent elliptical boundary onto a static boundary of the shape of a unit circle. Applying (\[Eq:coordtrafo\]) to (\[Eq:SE1\]) together with a unitary transformation, $$U(x,y,t) = \exp\left(-\frac{\dot \imath \mu}{2 \hbar}\left(\frac{\dot a(t) x^2}{a(t)} + \frac{\dot b(t) y^2}{b(t)}\right)\right) \textrm{,} \label{Eq:U}$$ and extracting a volume-dependent prefactor $\sqrt{a(t) b(t)}$ from the wave function $\Psi(\vec{x},t)$, we are led to an effective SE $$\dot \imath \hbar \partial_t \Lambda(\eta,\xi,t) = H^{e}(\eta,\xi,t) \Lambda(\eta,\xi,t) \textrm{,} \label{Eq:SE2}$$ where the effective Hamiltonian $H^{e}$ contains time derivatives of the prefactor $\sqrt{a(t) b(t)}$ and of the unitary transformation $U$ of the left-hand side of the TDSE: $$H^{e}(\eta,\xi,t) = \frac{-\hbar^2}{2 \mu} \left(\frac{1}{a^2(t)}\frac{\partial^2}{\partial \eta^2} + \frac{1}{b^2(t)}\frac{\partial^2}{\partial \xi^2} \right) + \frac{1}{2} \mu \left(a(t)\ddot a(t) \eta^2 + b(t)\ddot b(t) \xi^2\right) \textrm{.} \label{Eq:effective_Hamiltonian}$$ The introduction of the unitary transformation (\[Eq:U\]) ensures that $H^{e}$ is Hermitian. Due to the extracted prefactor $\sqrt{a(t) b(t)}$, the effective wave function $$\Lambda(\eta,\xi,t) := \sqrt{a(t)b(t)} \ U(\rho_t^{-1}(\eta,\xi),t) \Psi(\rho_t^{-1}(\eta,\xi),t)$$ is normalized to $1$ on the domain boundary of the unit circle $C := \left\{\vec{x} = (x,y)^{\intercal} \in \mathbb{R}^2 \left| x^2 + y^2 \leq 1\right.\right\}$ and the coordinate transformation (\[Eq:coordtrafo\]) makes $\Lambda$ subject to the Dirichlet boundary condition $\Psi(\vec{x},t)|_{\partial C} = 0$. The reader is referred to [@Flrespop] for a similar, more detailed derivation of the effective Hamiltonian and equations of motion. A complete set of orthonormal functions on $C$ is given by the eigenfunctions of the static circular billiard [@circular_billiard1; @circular_billiard2]: $$\Phi_{n,m}(\rho,\phi) = \frac{1}{\sqrt{\pi} J_{m+1}(k_{m,n})} J_{m}(k_{m,n} \rho) e^{\dot \imath m \phi} \textrm{.} \label{Eq:eigenfunctions_circular_billiard}$$ $\rho$ and $\phi$ can be calculated from $\eta$ and $\xi$ by $\eta = \rho \cos \phi$ and $\xi = \rho \sin \phi$. $J_m$ is the cylindrical Bessel function of order $m$ and $k_{m,n}$ is its $n$-th root. $n$ and $m$ are called radial, resp. angular, quantum number for obvious reasons. If we expand the effective wave function $\Lambda$ in terms of the eigenfunctions of the static circular billiard, the effective SE (\[Eq:SE2\]) becomes a linear homogeneous ordinary differential equation of first order in time and can, thus, be solved numerically by standard methods [@Flrespop]. A main result of [@Flrespop] was the observation of resonant population transfer between so-called instantaneous eigenstates of $$H_M = \frac{-\hbar^2}{2 \mu} \left(\frac{1}{a^2(t)}\frac{\partial^2}{\partial \eta^2} + \frac{1}{b^2(t)}\frac{\partial^2}{\partial \xi^2} \right) \textrm{.} \label{Eq:Mathieu_Hamiltonian}$$ We understand instantaneous eigenstates as follows: The semi-axes $a$ and $b$ are parameters of $H_M$ that change in time. If we evolve our system solely by $H_M$ in the SE, start the system in an initial state that corresponds to an eigenstate of $H_M$ at $t=0$ and change $a$ and $b$ sufficiently slowly, then we define the instantaneous eigenstate of $H_M$ at time $t$ as the time-evolved wave function of the system at time $t$ in accordance to the adiabatic theorem of quantum mechanics [@adiabatensatz]. The Hamiltonian $H_M$ (\[Eq:Mathieu\_Hamiltonian\]) is part of the effective Hamiltonian $H^e$ (\[Eq:effective\_Hamiltonian\]). Its complementary part is $$H_F = H^e - H_M = \frac{1}{2} \mu \left(a(t)\ddot a(t) \eta^2 + b(t)\ddot b(t) \xi^2\right) \textrm{.} \label{Eq:coupling_hamiltonian}$$ Population transfer between instantaneous eigenstates of $H_M$ takes place by two different mechanisms in the billiard. First, as $a$ and $b$ are of course not changed sufficiently slowly, diabatic population transfer between the instantaneous eigenstates of $H_M$ will take place. Additionally, the Hamiltonian $H_F$ triggers population transfer as it is non-diagonal in the basis set of instantaneous eigenstates of $H_M$. Introducing the volume of the elliptical billiard, $V(t) = a(t)b(t) \textrm{,}$ and the ratio of the semi-axes, $r(t) = b(t)/a(t)$, $H_M$ can be rewritten in the much more convenient form $$H_M = \frac{\hbar^2}{\mu V(t)} M(r(t)) \textrm{,}$$ where we call $$M(r) := -\frac{1}{2}\left(r \frac{\partial^2}{\partial \eta^2} + \frac{1}{r}\frac{\partial^2}{\partial \xi^2}\right) \label{Eq:Mathieu_op}$$ the Mathieu operator as its eigenfunctions are just ordinary and modified Mathieu functions as they appear in the solutions of the static elliptical billiard. If we label the eigenstates of $M(r)$ by $\Ket{n;r}$ with eigenvalue $q_n(r)$, $\Ket{n;r(t)}$ are, of course, the instantaneous eigenstates of $H_M$ and $E_n(t) = \frac{\hbar^2}{\mu V(t)} q_n(r(t))$ the corresponding instantaneous eigenvalues of $H_M$, i.e. we have $$H_M = \sum_{n = 1}^{\infty} \Ket{n;r(t)}\frac{\hbar^2 q_n(r(t))}{\mu V(t)}\Bra{n;r(t)} \textrm{.} \label{Eq:Mathieu_Ham_diagonal_form}$$ Note that $M(r)$ is invariant upon sign-change of $\eta$ and $\xi$. One can therefore choose its eigenstates $\Ket{n;r}$ such that they are also eigenstates of the parity operators that change the sign of $\eta$ or $\xi$. In this context, we will refer to $\Ket{n;r}$ having even or odd $\eta$-, resp. $\xi$-, parity. Note that the effective Hamiltonian $H^e$ (\[Eq:effective\_Hamiltonian\]) is also invariant upon sign-change of $\eta$ and $\xi$ and, consequently, only couples instantaneous eigenstates that have the same $\eta$- and $\xi$- parity. The Hilbert space therefore splits into four uncoupled Hilbert subspaces. We choose the following ansatz for the effective wave function $\Lambda$, $$\Ket{\Lambda(t)} = \sum_n c_n(t) e^{-\frac{\dot \imath}{\hbar} \phi_n(t)} \Ket{n;r(t)} \textrm{,}$$ with time-dependent expansion coefficients $c_n(t)$ and $$\phi_n(t) := \int_0^t dt' E_n(t') = \int_0^t dt' \frac{\hbar^2}{\mu V(t')} q_n(r(t')) \label{Eq:phase_dim}$$ being the time-integrated instantaneous eigenvalues of $H_M$. If we put this ansatz into the the SE (\[Eq:SE2\]) and note that $ \partial_t \Ket{n;r(t)} = \dot r (\partial_r \Ket{n;r})|_{r = r(t)} \equiv \dot r \partial_r \Ket{n;r(t)} \textrm{,}$ we get a SE for the coefficients $c_n(t)$: $$\begin{aligned} \dot \imath \hbar \partial_t c_n(t) &=& \sum_m c_m(t) e^{-\frac{\dot \imath}{\hbar} (\phi_m(t)-\phi_n(t))} \left(\Braket{n;r(t)|H_F(t)|m;r(t)} - \dot \imath \hbar \dot r(t) \Braket{n;r(t)|\partial_r|m;r(t)} \right) \nonumber\\ \label{Eq:SE_coeff_1}\end{aligned}$$ For non-degenerate eigenstates $\Ket{n;r}$ and $\Ket{m;r}$, one can express the second matrix element on the right-hand side of (\[Eq:SE\_coeff\_1\]) as $$\Braket{n;r|\partial_r|m;r} = \begin{cases} \frac{\Braket{n;r|(\partial_r M(r))|m;r}}{q_m(r) - q_n(r)} & \textrm{ for } n \neq m\\ \Braket{n;r|\partial_r|n;r} & \textrm{ for } n = m \textrm{.} \end{cases} \label{Eq:diabatic_coupling_1}$$ It is now interesting to notice that the representation of $M(r)$ in the eigenbasis of the static circular billiard (\[Eq:eigenfunctions\_circular\_billiard\]), $\Braket{\Phi_{n',m'}|M(r)|\Phi_{n,m}}$, is not only a Hermitian, but a real symmetric matrix (cf. Eq. (\[Eq:mathieu\_elements\]) in appendix \[App:matrix\_elements\]). We can therefore choose the expansion coefficients of the eigenstates of $M(r)$ in the eigenbasis of the static circular billiard, $\Braket{\Phi_{n',m'}|n;r}$, to be real. It follows that also the expansion coefficients of $\partial_r\Ket{n;r}$, $\Braket{\Phi_{n',m'}|\partial_r|n;r}$, are real. Thus, $$\Braket{n;r|\partial_r|n;r} = \sum_{n',m'} \Braket{n;r|\Phi_{n',m'}}\Braket{\Phi_{n',m'}|\partial_r|n;r} \label{Eq:Berry_expansion_circular}$$ is also real. On the other hand, due to normalization of the eigenstates $\Ket{n;r}$, expression (\[Eq:Berry\_expansion\_circular\]) has to be purely imaginary and is therefore identical zero. Noting that $$\begin{aligned} \partial_r M(r) &=& \frac{\partial^2}{\partial \eta^2} - \frac{1}{r^2}\frac{\partial^2}{\partial \xi^2} \nonumber\\ &=& \frac{1}{r} \left(r \frac{\partial^2}{\partial \eta^2} - \frac{1}{r}\frac{\partial^2}{\partial \xi^2} \right) \nonumber\\ &=& \frac{M(\dot \imath r)}{\dot \imath r} \label{Eq:parameter_derivative_M}\end{aligned}$$ further simplifies Eq. (\[Eq:diabatic\_coupling\_1\]). From now on, we will restrict ourselves to periodic driving laws, i.e. $a(t+\frac{2 \pi}{\omega}) = a(t)$ and $b(t+\frac{2 \pi}{\omega}) = b(t)$, and cases where all populated instantaneous eigenstates are non-degenerate. These conditions also include all cases that are discussed in [@Flrespop]. We further introduce a rescaled dimensionless time $\tau := \frac{\omega}{2 \pi} t$ and finally put Eqs. (\[Eq:diabatic\_coupling\_1\]) and (\[Eq:parameter\_derivative\_M\]) back into (\[Eq:SE\_coeff\_1\]). $$\begin{aligned} \dot \imath \partial_{\tau} c_n(\tau) &=& \sum_m c_m(\tau) e^{-\frac{2 \pi \dot \imath}{\hbar \omega} (\phi_m(\tau)-\phi_n(\tau))} \frac{\omega}{2 \pi \hbar} \Braket{n;r(\tau)|H_F(\tau)|m;r(\tau)} \nonumber\\ &+& \sum_{m \neq n} c_m(\tau) e^{-\frac{2 \pi \dot \imath}{\hbar \omega} (\phi_m(\tau)-\phi_n(\tau))} \frac{\dot r(\tau)}{r(\tau)} \frac{\Braket{n;r(\tau)|M(\dot \imath r(\tau))|m;r(\tau)}}{q_n(r(\tau)) - q_m(r(\tau))} \textrm{.} \label{Eq:SE_coeff_2}\end{aligned}$$ We point out that the modulus of the first term in (\[Eq:SE\_coeff\_2\]) depends linearly on the driving frequency $\omega$ while the modulus of the second term is independent of $\omega$. We therefore expect the first term to be dominating for large driving frequencies, while the second one should be dominant for small driving frequencies and should especially couple neighboring instantaneous eigenstates due to the denominator $q_n(r(\tau)) - q_m(r(\tau))$. Obviously, due to periodic driving, all terms on the right-hand side of (\[Eq:SE\_coeff\_2\]) but the coefficients $c_m(\tau)$ and the phase factors $\exp(-2 \pi \dot \imath (\phi_m(\tau)-\phi_n(\tau))/\hbar \omega)$ are one-periodic functions in $\tau$. It is therefore possible to represent them by discrete Fourier transforms. Before we do so, we split $\phi_m(\tau)-\phi_n(\tau)$ into a non-periodic part $\hbar \nu_{mn} \tau := (\phi_m(1)-\phi_n(1)) \tau $ and a one-periodic part $\hbar \Delta\nu_{mn}(\tau)$: $$\phi_m(\tau)-\phi_n(\tau) = \hbar \nu_{mn} \tau + \hbar \Delta\nu_{mn}(\tau) \textrm{.} \label{Eq:split_phi}$$ We then combine the one-periodic phase factor $\exp(-2 \pi \dot \imath \Delta\nu_{mn}(\tau)/\omega)$ with the other one-periodic terms on the right-hand side of (\[Eq:SE\_coeff\_2\]) and Fourier transform the results: $$\begin{aligned} \omega \sum_{l = -\infty}^{l = \infty} F_l^{nm} e^{-2 \pi \dot \imath l \tau} &=& e^{-\frac{2 \pi \dot \imath}{\omega} \Delta\nu_{mn}(\tau)} \frac{\omega}{2 \pi \hbar} \Braket{n;r(\tau)|H_F(\tau)|m;r(\tau)} \label{Eq:FT_pert_coupling} \\ \sum_{l = -\infty}^{l = \infty} D_l^{nm} e^{-2 \pi \dot \imath l \tau} &=& \begin{cases} e^{-\frac{2 \pi \dot \imath}{\omega} \Delta\nu_{mn}(\tau)} \frac{\dot r(\tau)}{r(\tau)} \frac{\Braket{n;r(\tau)|M(\dot \imath r(\tau))|m;r(\tau)}}{q_n(r(\tau)) - q_m(r(\tau))} & \textrm{ for } n \neq m\\ 0 & \textrm{ for } n = m \textrm{.} \end{cases} \label{Eq:FT_diab_coupling}\end{aligned}$$ Before we put (\[Eq:FT\_pert\_coupling\]) and (\[Eq:FT\_diab\_coupling\]) back into the SE (\[Eq:SE\_coeff\_2\]), it is useful to perform a unitary transformation, $c_n(\tau) = \exp(-\dot \imath \omega F_{0}^{nn} \tau) b_n(\tau)$. Note that we do not include $D_0^{nn}$ in the unitary transformation as it is zero by definition (\[Eq:FT\_diab\_coupling\]) and that $F_{0}^{nn}$ is completely independent of $\omega$. This unitary transformation, together with Eqs. (\[Eq:FT\_pert\_coupling\]) and (\[Eq:FT\_diab\_coupling\]), leads via (\[Eq:SE\_coeff\_2\]) to a SE for the coefficients $b_n$: $$\dot \imath \dot b_n(\tau) = \sum_{\substack{m,l \\ m \neq n \textrm{ for } l = 0}} e^{2 \pi \dot \imath \theta_{l}^{nm} \tau} (\omega F_l^{nm} + D_l^{nm}) b_m(\tau) \textrm{,} \label{Eq:FT_SE_1}$$ where we have defined the abbreviation $$\theta_{l}^{nm} := \frac{\nu_{nm}}{\omega} + \frac{\omega}{2 \pi} (F_{0}^{nn} - F_{0}^{mm}) - l \textrm{.} \label{Eq:theta_def}$$ Note that the solution of Eq. (\[Eq:FT\_SE\_1\]) determines the complete physics of periodically driven elliptical quantum billiards. Perturbative analysis {#ch:perturbative_analysis} --------------------- We will now use time-dependent perturbation theory (TDPT) to find an approximate solution of Eq. (\[Eq:FT\_SE\_1\]) in first order. To do so, we formally affix a parameter $\lambda$ to $F_l^{nm}$ and $D_l^{nm}$ to keep track of the order of perturbation and will set $\lambda$ to 1 at the end of our calculation: $F_l^{nm} = \lambda \cdot F_l^{nm}$, $D_l^{nm} = \lambda \cdot D_l^{nm}$. An expansion of $b_n(\tau)$ in $\lambda$ gives: $b_n(\tau) = \sum_{p = 0}^{\infty} \lambda^{p} b_n^{(p)}(\tau)$. As $\lambda$ should track the order of perturbation, it is natural to choose the initial values of $b_n^{(p)}$ as $b_n^{(p)}(0) = \delta_{p,0} \cdot b_n(0)$. Inserting this ansatz into (\[Eq:FT\_SE\_1\]) and equating equal powers of $\lambda$ yields up to first order: $$\begin{aligned} \dot \imath \dot b_n^{(0)} &=& 0 \Rightarrow b_n^{(0)} = \textrm{const.} = b_n(0) \\ \nonumber\\ \dot \imath \dot b_n^{(1)} &=& \sum_{\substack{m,l \\ m \neq n \textrm{ for } l = 0}} e^{2 \pi \dot \imath \theta_{l}^{nm} \tau} (\omega F_l^{nm} + D_l^{nm}) b_m(0) \\ \nonumber\\ \Rightarrow b_n^{(1)}(\tau) &=& \sum_{\substack{m,l \\ m \neq n \textrm{ for } l = 0}} \frac{e^{2 \pi \dot \imath \theta_{l}^{nm} \tau} - 1}{2 \pi \dot \imath \theta_{l}^{nm}} (\omega F_l^{nm} + D_l^{nm}) b_m(0) \label{Eq:first_order_coeff} \ \textrm{.}\end{aligned}$$ ### Population transfer probability We are now able to calculate the population transfer probability between two instantaneous eigenstates which will lead to a systematical understanding of resonant population transfer as it was observed in [@Flrespop]. For this purpose we assume that the wave function $\Ket{\Lambda}$ was initially in the (undriven) eigenstate $\Ket{k;r}$ and then calculate the evolution of the population of the eigenstate $\Ket{n;r}$ ($n \neq k$). Population transfer in first order gives: $$p_{nk}^{1}(\tau) := |b_n^{(1)}(\tau)|^2 = \sum_{l,l'} \frac{e^{2 \pi \dot \imath \theta_{l}^{nk} \tau} - 1}{2 \pi \theta_{l}^{nk}} \cdot \frac{e^{- 2 \pi \dot \imath \theta_{l'}^{nk} \tau} - 1}{2 \pi \theta_{l'}^{nk}} (\omega F_l^{nk} + D_l^{nk})(\omega F_{l'}^{*nk} + D_{l'}^{*nk}) \textrm{.} \label{Eq:pnk_1}$$ We would like to calculate a population transition rate per unit time from (\[Eq:pnk\_1\]) which is defined as $\Gamma_{nk}^{1} := \lim_{\tau \rightarrow \infty} p_{nk}^{1}(\tau)/\tau$. Note that $(e^{2 \pi \dot \imath \theta_{l}^{nk} \tau} - 1)/2 \pi \theta_{l}^{nk} = \dot \imath e^{\pi \dot \imath \theta_{l}^{nk} \tau} \sin(\pi \theta_{l}^{nk} \tau)/(\pi \theta_{l}^{nk})$ grows linearly with $\tau$ for $\theta_{l}^{nk} = 0$ while it oscillates periodically with an amplitude $1/\pi \theta_{l}^{nk}$ (which is independent of $\tau$) for $\theta_{l}^{nk} \neq 0$. Due to $\theta_{l}^{nk} - \theta_{l'}^{nk} = l' -l$, we can therefore neglect all terms in (\[Eq:pnk\_1\]) with $l \neq l'$ for $\tau$ being sufficiently large. $$\begin{aligned} \Gamma^{nk}_{1} &:=& \lim_{\tau \rightarrow \infty} \frac{p_{nk}^{1}(\tau)}{\tau} \nonumber\\ &=& \lim_{\tau \rightarrow \infty} \sum_l \frac{\sin^2 \pi \theta_{l}^{nk} \tau}{\tau (\pi \theta_{l}^{nk})^2} \left|\omega F_l^{nk} + D_l^{nk}\right|^2 \nonumber\\ &=& \sum_l \delta(\theta_{l}^{nk}) \left|\omega F_l^{nk} + D_l^{nk}\right|^2 \textrm{.} \label{Eq:transition_rate_calc}\end{aligned}$$ By applying appropriate transformations, we have handled the time-dependent boundary conditions of the billiard by introduction of a time-dependent external potential. This enabled us to derive Eq. (\[Eq:transition\_rate\_calc\]) which is a Fermi’s Golden Rule [@goldenRule] for driven elliptical quantum billiards. It states that efficient population transfer in first order between the instantaneous eigenstates $\Ket{k;r}$ and $\Ket{n;r}$ is only possible for $\theta_{l}^{nk}=0$. We can now use (\[Eq:theta\_def\]) to calculate corresponding resonance frequencies, $$\omega_{res}^{nk,l} = \frac{l \pm \sqrt{l^2 - 4 \nu_{nk} \delta F_0^{nk}}}{2 \delta F_0^{nk}} \textrm{,} \label{Eq:exact_resonance cond}$$ where $2 \pi \cdot \delta F_0^{nk} := F_0^{nn} - F_0^{kk}$ has been defined. Numerical experience shows that $\delta F_0^{nk}$ is usually a very small quantity. The “$+$”-term in (\[Eq:exact\_resonance cond\]) thus corresponds to a very large resonance frequency. Restricting ourselves to not too strongly driven billiards, we will neglect this term from now on. If we develop the “$-$”-term in (\[Eq:exact\_resonance cond\]) about $\delta F_0^{nk} \approx 0$ and use the definition of $\nu_{nk}$ in Eq. (\[Eq:split\_phi\]) above, we find: $$l \cdot \omega_{res}^{nk,l} = \nu_{nk} = \int_0^1 E_n(\tau') - E_k(\tau') d\tau' \label{Eq:appr_resonance_cond}$$ Thus, only when the one-period average difference of two instantaneous energy eigenvalues matches an integer multiple of the driving frequency, resonant population transfer between the corresponding instantaneous eigenstates can occur. This is precisely the empirically found criterion in [@Flrespop] and has herewith a theoretical basis. The result justifies to call the Fourier summation index $l$ “photon process order” of a population transfer in analogy to the interaction of light and matter. ### Applicability of first order TDPT Not all predicted resonance frequencies (\[Eq:exact\_resonance cond\]) are of equal importance with respect to their experimental observation and we will now derive a criterion to discriminate them. In the resonant case $\theta_{l}^{nk}=\theta_{l'}^{nk}=0$, Eq. (\[Eq:pnk\_1\]) reduces to: $$p_{nk}^{1} = \tau^2 \sum_l \left|\omega F_l^{nk} + D_l^{nk}\right|^2 \label{Eq:p_nk_in_res}$$ The reader is reminded that (\[Eq:p\_nk\_in\_res\]) only holds for $n \neq k$, while for $n = k$, $p_{kk}^{1}=0$ holds as $\theta_{l}^{kk}=0$ implies $l=0$ and this case just had been excluded from the summation in Eq. (\[Eq:first\_order\_coeff\]). Consequently, the instantaneous eigenstate $\Ket{k;r}$ gets exclusively depopulated in first order TDPT. We can therefore calculate the time $\tau_{\textrm{int}}$ at which the population $p_k$ of the instantaneous eigenstate $\Ket{k;r}$ gets negative and therefore unphysical: $$\begin{aligned} p_k(\tau_{\textrm{int}}) = 1 - \sum_{\theta_{l}^{nk} \equiv 0} p_{nk}^{1}(\tau_{\textrm{int}}) \stackrel{!}{=} 0 \label{Eq:p_k} \\ \Rightarrow \tau_{\textrm{int}} = \frac{1}{\sqrt{\sum_{\theta_{l}^{nk} \equiv 0} \left|\omega F_l^{nk} + D_l^{nk}\right|^2}} \textrm{.} \label{Eq:tau_int}\end{aligned}$$ The summation index $\theta_{l}^{nk} \equiv 0$ in (\[Eq:p\_k\]) and (\[Eq:tau\_int\]) means that it should only be summed over states $n$ and photon process orders $l$ that satisfy the resonance condition $\theta_{l}^{nk} \equiv 0$. This means in all practical examples that the sum only consists of a single term. $\tau_{\textrm{int}}$ is a measure for how fast a population probability transfer takes place. It is, thus, reasonable that we will not be able to fully resolve resonances that correspond to an interaction time $\tau_{\textrm{int}}$ that is much larger than the actual runtime $\tau_{\textrm{run}}$ of a possible experiment. In this case, population transfer will have been stopped before the maximal theoretically possible amount of population probability will have been transferred from one instantaneous eigenstate to the other and our ability to resolve a resonance in corresponding observations is diminished. On the other hand, we understand that the transition rate (\[Eq:transition\_rate\_calc\]) has been calculated in the limit $\tau \to \infty$ and the included $\delta$-function is the result of a convergence process. In order to have the system meet the predictions of first order TDPT, $\tau_{\textrm{int}}$ should be large enough such that a delta function $\delta(\theta)$ is a good approximation of $\sin^2 \pi \theta \tau/\tau (\pi \theta)^2$ as it appears in the derivation of (\[Eq:transition\_rate\_calc\]). Obviously, such criterion depends on the density of the $\theta_{l}^{nk}$ about $\theta = 0$. We therefore define a lower threshold $$\tau_{\textrm{low}} := \max_{|\theta_{l}^{nk}| \neq 0} \frac{1}{|\theta_{l}^{nk}|} \label{Eq:def_tau_low}$$ where only $\theta_{l}^{nk}$ should be considered in (\[Eq:def\_tau\_low\]) whose corresponding resonant probability transitions (i.e. for the case $\theta_{l}^{nk}=0$) have interaction times of the order of magnitude of $\tau_{\textrm{run}}$ such that they are relevant for the experiment. In summary, we expect predicted resonances to be fully resolved if $$\tau_{\textrm{low}} \ll \tau_{\textrm{int}} < \tau_{\textrm{run}} \label{Eq:res_criterion}$$ holds. The discussion of concrete driving laws in chapter \[ch:numerical\_results\] shows that this criterion is in excellent agreement with our numerical simulations. ### Rotating wave approximation Interestingly, (\[Eq:res\_criterion\]) justifies a rotating wave approximation in (\[Eq:FT\_SE\_1\]) [@vogel2006quantum]. This allows us to calculate approximate population dynamics of the system that, in contrast to (\[Eq:pnk\_1\]), conserve the total population probability. For simplicity, we will assume that there is only one $\theta_{l}^{nk}$ close to zero. A rotating wave approximation simply sets all other terms in (\[Eq:FT\_SE\_1\]) that do not contain this $\theta_{l}^{nk}$ to zero as they are comparatively fast oscillating, such that we are left with with an effective two-level system: $$\begin{aligned} \dot \imath \dot b_n(\tau) &=& e^{2 \pi \dot \imath \theta_{l}^{nk} \tau}(\omega F_l^{nk} + D_l^{nk}) b_k(\tau) \nonumber\\ \dot \imath \dot b_k(\tau) &=& e^{-2 \pi \dot \imath \theta_{l}^{nk} \tau} (\omega F_l^{kn} + D_l^{kn}) b_n(\tau) \ \textrm{.} \label{Eq:rabi_eq}\end{aligned}$$ The behavior of such a system is very well understood. A discussion in terms of Bloch equations is, for instance, given in [@mandel1995optical]. Eqs. (\[Eq:rabi\_eq\]) especially explain why the population dynamics in [@Flrespop] are reminiscent of Rabi oscillations. The effective Rabi-frequency $\Omega_{\textrm{eff}}$ can be calculated (cf. [@mandel1995optical]) to be $$\Omega_{\textrm{eff}} = \sqrt{\left(2 \pi \theta_{l}^{nk}\right)^2 + 4 |\omega F_l^{nk} + D_l^{nk}|^2} = \sqrt{\left(2 \pi \theta_{l}^{nk}\right)^2 + 4 \frac{1}{\tau_{\textrm{int}}^2}}$$ which yields a beating period $T_B$ of the population dynamics $$T_B := \frac{2 \pi}{\Omega} = \frac{\pi \tau_{\textrm{int}}}{\sqrt{1 + \left(\pi \theta_{l}^{nk} \tau_{\textrm{int}}\right)^2}} \textrm{.} \label{Eq:Beating_period}$$ In summary, if we assume the system to have initially been in state $k$, the population dynamics of state $n$ is given by: $$p_n(\tau) = \frac{\sin^2\left(\frac{\pi \tau}{T_B}\right)}{1 + (\pi \theta_{l}^{nk} \tau_{\textrm{int}})^2} \textrm{.} \label{Eq:rabi_population}$$ Numerical results and discussion {#ch:numerical_results} ================================ In this chapter we will present full numerical simulations of driven elliptical billiards and analyze the results with the developed perturbation theory of chapter \[ch:perturbative\_analysis\]. Details to the numerical calculation of the predicted quantities can be found in appendix \[App:matrix\_elements\]. All numerical calculations have been run for $\tau_{\textrm{run}} = 100$ periods of driving and $\hbar$ and $\mu$ have, w.l.o.g., been set to 1. We will always drive the semi-axis $a(t)$ harmonically, i.e. $$a(t) = a_0 + A \sin(\omega t) \textrm{,} \label{Eq:a(t)}$$ and adjust $b(t)$ such that the billiard is driven in different ways as will be specified later. To be able to compare the different driving laws, we have chosen to keep the following parameters fixed: $$a_0 := a(t=0) = 1\textrm{, } A = 0.1 \textrm{ and } b_0 := b(t=0) = \sqrt{0.51} \textrm{.} \label{Eq:parameter_choices}$$ These parameters are the same as in [@Flrespop]. The energy $E(\tau)$ will be a key observable for the analysis of the billiard dynamics. It is calculated from the expectation value of the Hamiltonian $H = -\frac{\hbar^2}{2 \mu} \Delta$ as it appears in Eq. (\[Eq:SE1\]): $$E(\tau) = \Braket{\Psi(\tau)|-\frac{\hbar^2}{2 \mu} \Delta|\Psi(\tau)} \textrm{.}$$ If we apply again the coordinate transformation (\[Eq:coordtrafo\]) and the unitary transformation $U$ (\[Eq:U\]), the energy reads $$E(\tau) = \Braket{\Lambda(\tau)|U^{\dagger}(\tau)H_M(\tau)U(\tau)|\Lambda(\tau)} \textrm{,}$$ where $H_M(\tau)$ is given by (\[Eq:Mathieu\_Hamiltonian\]). We can therefore calculate $E(\tau)$ by determining the population $p_n(\tau)$ of the eigenstates of $U^{\dagger}(\tau)H_M(\tau)U(\tau)$ in $\Ket{\Lambda(\tau)}$, weighting these populations with the respective eigenvalues $E_n(\tau)$ of $U^{\dagger}(\tau)H_M(\tau)U(\tau)$ and sum the results up: $$E(\tau) = \sum_n E_n(\tau) p_n(\tau) \textrm{.} \label{Eq:energy_by_pop}$$ Note that, due to $U$ being a unitary transformation, the energy eigenvalues $E_n(\tau)$ of $U^{\dagger}(\tau)H_M(\tau)U(\tau)$ are actually identical to the instantaneous eigenvalues $E_n(\tau)$ of $H_M(\tau)$. From now on, we refer to the instantaneous eigenstates of $U^{\dagger}(\tau)H_M(\tau)U(\tau)$ as energy eigenstates. We understand in particular that the instantaneous eigenstates $\Ket{n;r(\tau)}$ of $H_M(\tau)$ are in general not identical to the energy eigenstates, but unitarily transformed energy eigenstates, given by $U^{\dagger}\Ket{n;r(\tau)}$. Note that $U$ is also invariant upon sign-change of $\eta$ and $\xi$, such that an instantaneous eigenstate $\Ket{n;r(\tau)}$ and its corresponding energy eigenstate $U^{\dagger}\Ket{n;r(\tau)}$ have the same $\eta$- and $\xi$- parity. We will initialize the system in the fourth energy eigenstate (at $\tau=0$) and calculate the populations $p_n(\tau)$ upon driving. Note that in [@Flrespop] instantaneous eigenstates $\Ket{n;r(\tau)}$ were used as initial states and for population analyses. Interestingly, we find that the overlap $|\Braket{n;r(\tau)|U(\tau)|n;r(\tau)}|^2$ of all relevant instantaneous eigenstates $\Ket{n;r(\tau)}$ with their respective energy eigenstates is greater than 94.5% for the later analyzed parameter regimes in chapter \[ch:numerical\_results\]. We thus expect that the energy eigenstates are similar to the instantaneous eigenstates and also show similar population dynamics. Consequently, we will from now on disregard the differences between $\Ket{n;r(\tau)}$ and the energy eigenstates $U^{\dagger}\Ket{n;r(\tau)}$ when analyzing the billiard dynamics perturbatively and will subsequently refer to $\Ket{n;r(\tau)}$ simply as energy eigenstates. This approximation enables us to predict the seemingly complicated population dynamics of the energy eigenstates by our perturbation theory for instantaneous eigenstates. Although the population dynamics of the energy eigenstates are not expected to be qualitatively different, some predictions may be compromised quantitatively. For instance, due to the different actions of $U^{\dagger}(\tau)$ on different $\Ket{n;r(\tau)}$, shifts of the resonance frequencies (\[Eq:exact\_resonance cond\]) are to be expected. However, as $U$ (\[Eq:U\]) approaches unity for vanishing $\omega$, these shifts will rather be observed for larger resonance frequencies. We will also find that the resonance shifts become more negligible for higher order photon processes. In the approximation of instantaneous eigenstates $\Ket{n;r(\tau)}$ being energy eigenstates, a transition to a higher excited state increases the energy $E(\tau)$ (\[Eq:energy\_by\_pop\]) while a transition to a lower excited state decreases it. We can therefore determine if a population transition occurs at a certain driving frequency $\omega$ upon simulation time $\tau_{\textrm{run}}$ by recording the maximal and minimal energy of the billiard in dependence of $\omega$. Axes-ratio-preserving driving law {#ch:axes_ratio} --------------------------------- In the following, we provide numerical solutions for various driving laws and analyze them with the developed perturbation theory of chapter \[ch:perturbative\_analysis\]. A simple but illustrative driving law is the so-called axes-ratio-preserving driving law which merely rescales the billiard by varying its volume $V(\tau)$ while keeping the ratio of the semi-axes $r(\tau)$ constant for all times upon driving. As $a(t)$ is given by Eq. (\[Eq:a(t)\]), we find $ b(t) = r_0 a(t)$ and choose $r_0 = \sqrt{0.51}$ to satisfy Eq. (\[Eq:parameter\_choices\]). The axes-ratio-preserving driving law has the nice property that the instantaneous eigenstates $\Ket{n;r(\tau)}$ become time-independent due to fixed $r(\tau) = r_0$, while the eigenvalues $E_n(\tau)$ stay time-dependent as can be seen from Eq. (\[Eq:Mathieu\_Ham\_diagonal\_form\]). Their variation is solely given by the global prefactor $1/V(\tau)$ which particularly prevents crossings of energy eigenvalues. Fig. \[Fig:ratio\_eigenvalues\] shows a sample of eigenvalue curves for one period of driving. The fourth energy eigenstate has even $\eta$- and $\xi$- parity. Thus, only energy eigenstates in the corresponding sub-Hilbert space couple to the chosen initial state. ![Energy eigenvalues $E_n(\tau)$ of eigenstates $\Ket{n;r(\tau)}$ with even $\eta$- and $\xi$- parity eigenvalues for the axes-ratio-preserving driving law. The instantaneous shape of the ellipse at five different values of $\tau$ is drawn below the energy eigenvalue curves. Parameters: $a_0 = 1$, $b_0 = \sqrt{0.51}$, $A = 0.1$.[]{data-label="Fig:ratio_eigenvalues"}](inst_eigen_legend.eps){width="49.00000%"} In Fig. \[Fig:ratio\_resonances\] the dependence of the maximal and minimal energy that has been reached upon driving as a solution of the TDSE in Eq. (\[Eq:SE1\]) is plotted depending of the driving frequency $\omega$. We clearly see sharp peaks and dips at certain driving frequencies. The vertical lines represent our predictions of resonance frequencies according to Eq. (\[Eq:exact\_resonance cond\]). Note that the observed resonances deviate slightly from the predicted ones, especially for larger driving frequencies. This is due to the unitary transformation $U$ (\[Eq:U\]) that has been neglected in our considerations, i.e. we apply perturbation theory only to the instantaneous eigenstate $\Ket{n;r(\tau)}$ that is most populated in the energy eigenstate $U^{\dagger}(\tau)\Ket{n;r(\tau)}$. Beside this anticipated small deviation, we find a very good agreement of the numerical calculations with our predictions. All resonances with a comparatively small interaction time $\tau_\textrm{int}$ have been resolved, while resonances with very large interaction times could not be observed. Naturally, for interaction times that are longer than (half) the runtime of an experiment $\tau_{\textrm{run}}$, a full population transition from the initial state to some other energy eigenstate can not happen according to Eq. (\[Eq:rabi\_population\]). This is the reason why some peaks in Fig. \[Fig:ratio\_resonances\] that correspond to transitions to the same energy eigenstate possess different heights. It is interesting to note that, although the runtime $\tau_{\textrm{run}}$ of our numerical simulations was only 100 periods of driving, resonances that correspond to an interaction time of up to 2000 periods of driving could still be partly resolved in Fig. \[Fig:ratio\_resonances\]. Tab. \[Tab:ratio\_resonances\] in appendix \[App:theoretical\_predictions\] provides numerical values for all predicted resonance frequencies between 0 and 16 that have an interaction time of less than 2000. It also shows that the lower threshold $\tau_{\textrm{low}}$ is always much smaller than the corresponding interaction time $\tau_\textrm{int}$, such that the first part of Eq. (\[Eq:res\_criterion\]) is fulfilled and TDPT of first order is applicable. ![Dependence of the maximal and minimal energy on the driving frequency $\omega$ for the axes-ratio-preserving driving law. The vertical lines show all predicted resonance frequencies (\[Eq:exact\_resonance cond\]) with an interaction time $\tau_{\textrm{int}}$ of less than 2000. The darker the lines are, the longer is the corresponding interaction time. Numerical values can be taken from Tab. \[Tab:ratio\_resonances\] in appendix \[App:theoretical\_predictions\]. Parameters like in Fig. \[Fig:ratio\_eigenvalues\].[]{data-label="Fig:ratio_resonances"}](resonances_tau_thre_2000.eps){width="49.00000%"} One might wonder about a structure of several small peaks, especially for frequencies $\omega \gtrsim 10.5$. We assume that these smaller, not predicted peaks correspond to transitions of second order where population is first transferred to one excited state and then from this state again transferred to yet another energy eigenstate. This is, for instance, supported by a population analysis in Fig. \[Fig:ratio\_pop\_1\] for the small peak at $\omega = 10.81$. We see that the mean (or envelope behaviour) population of the energetical ground state decreases while the population amplitude of the tenth energy eigenstate increases. This may be interpreted as an interaction of these two states that consecutively leads to a transfer of population that was initially transferred to the ground state and is then pushed to the tenth energy eigenstate. Such a process is not included in the time-dependent perturbation theory of first-order in chapter \[ch:perturbative\_analysis\] and the dynamics visualized in Fig. \[Fig:ratio\_pop\_1\] are a precursor to the breakdown of this simple theory for higher driving frequencies where indirect transitions become more and more important. ![Population dynamics $p_n(\tau)$ for the axes-ratio-preserving driving law at $\omega = 10.81$. Parameters like in Fig. \[Fig:ratio\_eigenvalues\].[]{data-label="Fig:ratio_pop_1"}](population_w_1081_1.eps){width="49.00000%"} It is also interesting that the resonance at $\omega \approx 15.17$ that corresponds to a 4-photon transition from the initial state to the 22nd energy eigenstate is so well resolved although the interaction time of this resonance is much larger than the interaction time of several resonances that are much worse resolved. The reason for this is that the 22nd energy eigenstate has a much higher energy eigenvalue than, for instance, the seventh energy eigenstate. Thus, the energy is significantly increased for already a small amount of transferred population probability from the initial state to the 22nd energy eigenstate. ![Population dynamics $p_n(\tau)$ for the axes-ratio-preserving driving law at $\omega = 15.17$. Parameters like in Fig. \[Fig:ratio\_eigenvalues\].[]{data-label="Fig:ratio_pop_2"}](population_w_1517_1.eps){width="49.00000%"} Other driving laws {#ch:other} ------------------ We have seen in the last chapter that the predictions of TDPT work very well for the axes-ratio-preserving driving law. To demonstrate the general applicability of the perturbation theory of chapter \[ch:perturbative\_analysis\], we will analyze two further driving laws. The so-called breathing driving law $b(t) = a(t) - a_0 + b_0$, where $a(t)$ is again given by Eq. (\[Eq:a(t)\]), was already discussed in [@Flrespop]. Fig. \[Fig:breathing\_eigenvalues\] shows the eigenvalues of energy eigenstates with even $\eta$- and $\xi$- parity for one period of driving. We see that crossings of energy eigenvalues are, in contrast to Fig. \[Fig:ratio\_eigenvalues\], now possible as $r(t)$ is no longer kept constant. The eigenvalues of the lowest excited states are, though, very similar to the ones in Fig. \[Fig:ratio\_eigenvalues\] and, consequently, the resonances in Fig. \[Fig:breathing\_resonances\] resemble the ones in Fig. \[Fig:ratio\_resonances\]. In Fig. \[Fig:breathing\_resonances\] more resonances can be resolved due to a sufficiently small interaction time $\tau_{\textrm{int}}$. This observation can be understood as follows: While for the axes-ratio-preserving driving law the transition matrix $D_l^{nm}$ (\[Eq:FT\_diab\_coupling\]) is identical to zero due to $\dot r(t) = 0$, this is not the case for the breathing driving law. The additionally resolved resonances for the breathing law correspond, thus, to Landau-Zener transitions with $\dot r(t)$ being the Landau-Zener velocity [@zener]. This role of $D_l^{nm}$ triggering Landau-Zener transitions will be even more pronounced for the next driving law that is presented. ![Energy eigenvalues $E_n(\tau)$ of eigenstates $\Ket{n;r(\tau)}$ with even $\eta$- and $\xi$- parity for the breathing driving law. Parameters like in Fig. \[Fig:ratio\_eigenvalues\].[]{data-label="Fig:breathing_eigenvalues"}](inst_eigen_legend_br.eps){width="49.00000%"} ![Analogue of Fig. \[Fig:ratio\_resonances\] for the breathing driving law. Numerical values can be taken from Tab. \[Tab:breathing\_resonances\] in appendix \[App:theoretical\_predictions\]. Parameters like in Fig. \[Fig:ratio\_eigenvalues\].[]{data-label="Fig:breathing_resonances"}](resonances_tau_thre_2000_br.eps){width="49.00000%"} The so-called volume-preserving driving law is just the opposite of the axes-ratio-preserving driving law. It keeps the volume $V(t)$ of the billiard fixed, while varying the ratio of the semi-axes $r(t)$. Thus, $b(t)$ depends on $a(t)$ (\[Eq:a(t)\]) as $b(t) = a_0 b_0/a(t)$. Fig. \[Fig:vol\_eigenvalues\] shows the corresponding energy eigenvalues. We see that the fourth and seventh energy eigenvalue get close upon driving such that we expect that the transition matrix $D_l^{nm}$ (\[Eq:FT\_diab\_coupling\]) couples these states strongly. We point out that we can arbitrarily control how close these eigenvalues get upon driving by choosing $r(t)$ appropriately. As $D_l^{nm}$ (\[Eq:FT\_diab\_coupling\]) depends on $\omega$ only through the phase factor $e^{-\frac{2 \pi \dot \imath}{\omega} \Delta\nu_{mn}(\tau)}$, we expect it to be slowly varying with $\omega$, thus setting an upper bound on the interaction time (\[Eq:tau\_int\]) of resonant population transitions between the fourth and seventh energy eigenstate even for small $\omega$ and corresponding large photon process orders $l$. ![Energy eigenvalues $E_n(\tau)$ of eigenstates $\Ket{n;r(\tau)}$ with even $\eta$- and $\xi$- parity eigenvalues for the volume-preserving driving law. Parameters like in Fig. \[Fig:ratio\_eigenvalues\].[]{data-label="Fig:vol_eigenvalues"}](inst_eigen_legend_vol.eps){width="49.00000%"} This expectation is fully confirmed by Fig. \[Fig:vol\_resonances\]. All resolved resonances with $\omega < 5.5$ correspond exclusively to transitions between the fourth and seventh energy eigenstate. As $U$ (\[Eq:U\]) gets close to identity for small driving frequencies $\omega$, the induced resonance shift is negligible. One might wonder, why most of the resonances at small driving frequencies have numerically not been fully resolved although their interaction time is short enough. In review of Eq. (\[Eq:appr\_resonance\_cond\]), we understand that a detuning in the driving frequency is also multiplied by the photon process order $l$. Thus, one has to adjust the driving frequency very carefully to resolve a multiple photon resonance. We assume that the frequency grid in Fig. \[Fig:vol\_resonances\] is not fine enough to resolve all predicted resonances although it should in principle be possible. ![Analogue of Fig. \[Fig:ratio\_resonances\] for the volume-preserving driving law. Numerical values can be taken from Tab. \[Tab:vol\_resonances\] in appendix \[App:theoretical\_predictions\]. Parameters like in Fig. \[Fig:ratio\_eigenvalues\].[]{data-label="Fig:vol_resonances"}](resonances_tau_thre_2000_vol.eps){width="49.00000%"} Finally, population analyses close to the resonance frequencies confirm the Rabi-like behavior of the population dynamics as predicted by Eq. (\[Eq:rabi\_population\]). This can be especially well illustrated for resonances with small interaction times. As an example, Fig. \[Fig:vol\_pop\_2\] shows almost perfect Rabi-like population dynamics of the fourth and seventh energy eigenstate in the nearly resonant case of a three photon process at $\omega = 3.32$. Comparison of the observed beating periods with the corresponding interaction times $\tau_{\textrm{int}}$ in Tab. \[Tab:vol\_resonances\] in appendix \[App:theoretical\_predictions\] gives, even quantitatively, a very good agreement as predicted by Eq. (\[Eq:Beating\_period\]). We point out that as we can tune the strength of the transition matrix $D_l^{nm}$ by choosing how close the energy eigenvalues get upon driving, we can also tune the interaction time $\tau_{\textrm{int}}$ in the regime of weak driving where it is mainly determined by $D_l^{nm}$. Hence, we can, in principle, also control the beating period of the present effective two-level Rabi-system. ![Population dynamics $p_n(\tau)$ for the volume-preserving driving law at $\omega = 3.32$. Parameters like in Fig. \[Fig:ratio\_eigenvalues\].[]{data-label="Fig:vol_pop_2"}](population_w_332_1_vol.eps){width="49.00000%"} Brief Summary ============= A time-dependent perturbative approach for elliptical quantum billiards with oscillating boundaries has been developed. As our major results we have obtained a Fermi Golden Rule, predicting the driving frequencies yielding resonant population transfer between instantaneous eigenstates as observed in [@Flrespop] and a criterion allowing to decide which of these resonances are observable in a corresponding experiment of certain duration. Extensive numerical simulations have been performed for three different driving laws which are in excellent agreement with our predictions. Particularly for the volume preserving driving law, due to the change of the billiard geometry upon driving, Landau-Zener transitions have been analyzed to take place. Depending only weakly on the driving frequency, these transitions allow for resonant population transfer also for very weak driving. We have shown, that the billiard dynamics can be reduced in this regime to an effective two-level system with in principle arbitrarily tunable oscillation period.\ Further interesting phenomena beyond the scope of our perturbative description can be expected in the numerically challanging regime of strong driving. Acknowledgements ================ B.L. thanks the Landesexzellenzinitiative Hamburg “Frontiers in Quantum Photon Science”, which is funded by the Joachim Herz Stiftung, for financial support. Matrix elements {#App:matrix_elements} =============== Introducing the matrices $$\begin{aligned} \hat{f}^1 &=& \sum_{n,m,n',m'} \Ket{\Phi_{n,m}}\delta_{m,m'} f^1_{nmn'}\Bra{\Phi_{n',m'}} \\ \hat{f}^2 &=& \sum_{n,m,n',m'} \Ket{\Phi_{n,m}}\delta_{m,m'} f^2_{nmn'}\Bra{\Phi_{n',m'}} \\ \hat{f}^3 &=& \sum_{n,m,n',m'} \Ket{\Phi_{n,m}}\delta_{(m-2),m'} f^3_{nmn'}\Bra{\Phi_{n',m'}} \\ \hat{f}^4 &=& \sum_{n,m,n',m'} \Ket{\Phi_{n,m}}\delta_{(m-2),m'} f^4_{nmn'}\Bra{\Phi_{n',m'}} \\ \hat{f}^5 &=& \sum_{n,m,n',m'} \Ket{\Phi_{n,m}}\delta_{(m+2),m'} f^5_{nmn'}\Bra{\Phi_{n',m'}} \\ \hat{f}^6 &=& \sum_{n,m,n',m'} \Ket{\Phi_{n,m}}\delta_{(m+2),m'} f^6_{nmn'}\Bra{\Phi_{n',m'}} \textrm{,}\end{aligned}$$ with matrix elements $$\begin{aligned} f^1_{nmn'} &=& \frac{-k^2_{m,n}}{4} \delta_{n,n'} \label{f1}\\ f^2_{nmn'} &=& \frac{1}{2 J_{m+1}(k_{m,n}) J_{m+1}(k_{m,n'})} \int_0^1 J_m(k_{m,n}r) J_m(k_{m,n'}r) r^3 dr \label{f2}\\ \nonumber\\ f^3_{nmn'} &=& \frac{k^2_{m-2,n'}}{4 J_{m+1}(k_{m,n}) J_{m-1}(k_{m-2,n'})} \int_0^1 J_m(k_{m,n}r) J_m(k_{m-2,n'}r) r dr \label{f3}\\ f^4_{nmn'} &=& \frac{1}{4 J_{m+1}(k_{m,n}) J_{m-1}(k_{m-2,n'})} \int_0^1 J_m(k_{m,n}r) J_{m-2}(k_{m-2,n'}r) r^3 dr \label{f4}\\ \nonumber\\ f^5_{nmn'} &=& \frac{k^2_{m+2,n'}}{4 J_{m+1}(k_{m,n}) J_{m+3}(k_{m+2,n'})} \int_0^1 J_m(k_{m,n}r) J_m(k_{m+2,n'}r) r dr \label{f5}\\ f^6_{nmn'} &=& \frac{1}{4 J_{m+1}(k_{m,n}) J_{m+3}(k_{m+2,n'})} \int_0^1 J_m(k_{m,n}r) J_{m+2}(k_{m+2,n'}r) r^3 dr \label{f6} \textrm{,} \end{aligned}$$ where $J_m$ is again the cylindrical Bessel function of order $m$ and $k_{m,n}$ is its $n$-th root. We have a convenient form of representing $H_M$ (\[Eq:Mathieu\_Hamiltonian\]), $M(r)$ (\[Eq:Mathieu\_op\]) and $H_F(\tau)$ (\[Eq:coupling\_hamiltonian\]) in the eigenbasis $\{\Ket{\Phi_{n,m}}\}_{n,m}$ (\[Eq:eigenfunctions\_circular\_billiard\]) of the static circular billiard: $$H_M = g_1(\tau) \hat{f}^{1} + g_3(\tau) \left( \hat{f}^{3} + \hat{f}^{5} \right) \textrm{,}$$ $$M(r) \ = -\left(r + \frac{1}{r}\right) \hat{f}^{1} - \left(r - \frac{1}{r}\right) \left( \hat{f}^{3} + \hat{f}^{5} \right) \textrm{,} \label{Eq:mathieu_elements}$$ $$H_F(\tau) \ = g_2(\tau) \hat{f}^{2} + g_4(\tau) \left(\hat{f}^{4} + \hat{f}^{6}\right) \textrm{.} \label{H_pert_elements}$$ Diagonalizing $M(r)$ yields the instantaneous eigenstates $\Ket{n;r}$ and their eigenvalues $q_n(r)$. One could in principle calculate the energy eigenvalues $E_n(\tau) = \frac{\hbar^2}{\mu V(\tau)} q_n(r(\tau))$ from the $q_n(r)$, but it turns out that diagonalizing $H_M$ directly increases the numerical precision of the energy eigenvalues. Note that the time-dependent factors $g_i(\tau)$ as well as the matrix elements $f^i_{nmn'}$ are the same as in [@Flrespop]. However, the matrix elements $f^i_{nmn'}$ have been reduced to a much simpler form, using orthonormality relations of the Bessel functions. $$\begin{aligned} g_1(\tau) &=& -\frac{\hbar^2}{\mu}\left(\frac{1}{a(\tau)^2} + \frac{1}{b(\tau)^2}\right) \label{g1_t}\\ g_2(\tau) &=& \mu \left(a(\tau)\ddot a(\tau) + b(\tau)\ddot b(\tau)\right) \label{g2_t}\\ g_3(\tau) &=& -\frac{\hbar^2}{\mu}\left(\frac{1}{a(\tau)^2} - \frac{1}{b(\tau)^2}\right) \label{g3_t}\\ g_4(\tau) &=& \mu \left(a(\tau)\ddot a(\tau) - b(\tau)\ddot b(\tau)\right) \label{g4_t}\end{aligned}$$ Further note that the sign of $g_3(\tau)$ is inverted in comparison with [@Flrespop]. We can now calculate the transition matrix elements $D_l^{nm}$ (\[Eq:FT\_diab\_coupling\]) and $F_l^{nm}$ (\[Eq:FT\_pert\_coupling\]): $$\begin{aligned} D_l^{nm} &=& v_{1,l}^{nm} + v_{2,l}^{nm} \\ F_l^{nm} &=& v_{3,l}^{nm} + v_{4,l}^{nm}\end{aligned}$$ $$\begin{aligned} v_{1,l}^{nm} &=& - \dot \imath \int_0^1 d\tau e^{2 \pi \dot \imath l \tau} e^{-\frac{2 \pi \dot \imath}{\omega} \Delta\nu_{mn}(\tau)} \left(r - \frac{1}{r}\right) \frac{\dot r}{r} \frac{\Braket{n;r|\hat{f}^1|m;r}}{q_n(r) - q_m(r)} \nonumber\\ \label{pert_zener_1}\\ \nonumber\\ v_{2,l}^{nm} &=& - \dot \imath \int_0^1 d\tau e^{2 \pi \dot \imath l \tau} e^{-\frac{2 \pi \dot \imath}{\omega} \Delta\nu_{mn}(\tau)} \left(r + \frac{1}{r}\right) \frac{\dot r}{r} \frac{\Braket{n;r|\hat{f}^3 + \hat{f}^5|m;r}}{q_n(r) - q_m(r)} \nonumber\\ \label{pert_zener_2}\\ \nonumber\\ v_{3,l}^{nm} &=& \frac{1}{2 \pi \hbar} \int_0^1 d\tau e^{2 \pi \dot \imath l \tau} e^{-\frac{2 \pi \dot \imath}{\omega} \Delta\nu_{mn}(\tau)} g_2(\tau) \Braket{n;r|\hat{f}^2|m;r} \nonumber\\ \label{pert_acc_1} \\ \nonumber\\ v_{4,l}^{nm} &=& \frac{1}{2 \pi \hbar} \int_0^1 d\tau e^{2 \pi \dot \imath l \tau} e^{-\frac{2 \pi \dot \imath}{\omega} \Delta\nu_{mn}(\tau)} g_4(\tau) \Braket{n;r|\hat{f}^4 + \hat{f}^6|m;r} \ \textrm{.} \nonumber \\ \label{pert_acc_2}\end{aligned}$$ After calculating these quantities and diagonalizing $M(r)$ and $H_M$, it is straightforward to reproduce the theoretical predictions contained in this work. Theoretical predictions {#App:theoretical_predictions} ======================= $\omega_{res}^{n4,l}$ $\tau_{\textrm{int}}$ $\tau_{\textrm{low}}$ state $n$ order $l$ ----------------------- ----------------------- ----------------------- ----------- ----------- 3.966 304 0.181 1 3 5.030 328 0.482 7 2 5.122 1014 0.493 10 4 5.944 39.4 0.271 1 2 6.829 133 0.657 10 3 7.720 1098 0.743 13 4 10.09 40.5 0.970 7 1 10.24 17.3 0.985 10 2 10.29 144 0.990 13 3 11.84 4.91 0.540 1 1 15.11 1647 1.77 22 4 15.44 18.7 1.48 13 2 : Numerical values for all predicted resonance frequencies between the values 0 and 16 with an interaction time $\tau_{\textrm{int}}$ less than 2000. Information is provided on the corresponding lower threshold $\tau_{\textrm{low}}$, the quantum number of the coupling instantaneous eigenstate $n$ and the photon process order $l$ of the resonance.[]{data-label="Tab:ratio_resonances"} $\omega_{res}^{n4,l}$ $\tau_{\textrm{int}}$ $\tau_{\textrm{low}}$ state $n$ order $l$ ----------------------- ----------------------- ----------------------- ----------- ----------- 2.584 589 0.254 7 4 3.446 150 0.339 7 3 3.972 309 0.178 1 3 5.128 1207 0.504 10 4 5.170 40.0 0.508 7 2 5.954 39.9 0.268 1 2 6.836 151 0.672 10 3 7.807 361 0.728 13 4 10.25 18.8 1.01 10 2 10.35 11.0 1.02 7 1 10.41 63.7 0.970 13 3 11.86 4.90 0.534 1 1 13.18 1284 3.29 20 4 15.21 1019 1.87 22 4 15.62 11.2 1.44 13 2 : Analogue to Tab. \[Tab:ratio\_resonances\] but for the breathing driving law.[]{data-label="Tab:breathing_resonances"} $\omega_{res}^{n4,l}$ $\tau_{\textrm{int}}$ $\tau_{\textrm{low}}$ state $n$ order $l$ ----------------------- ----------------------- ----------------------- ----------- ----------- 0.3682 1985 0.0367 7 3 0.3823 1889 0.0381 7 2 0.3976 1975 0.0397 7 4 0.4733 1176 0.0472 7 2 0.4970 842 0.0496 7 3 0.5232 741 0.0522 7 4 0.5522 712 0.0551 7 1 0.5847 525 0.0583 7 2 0.6213 303 0.0620 7 3 0.6627 185 0.0661 7 1 0.7100 124 0.0708 7 4 0.7647 88.6 0.0763 7 2 0.8284 63.7 0.0826 7 3 0.9037 45.0 0.0902 7 2 0.9941 31.3 0.0992 7 4 1.105 22.0 0.110 7 2 1.243 15.7 0.124 7 3 1.420 11.5 0.142 7 4 1.657 8.51 0.165 7 1 1.989 6.42 0.199 7 2 2.487 4.95 0.248 7 3 3.317 3.94 0.332 7 1 4.982 3.31 0.500 7 4 5.905 411 0.271 1 2 6.228 1134 0.555 13 3 7.795 691 0.691 13 2 8.559 875 1.98 20 4 9.926 28.9 1.01 10 2 10.04 2.97 1.02 7 3 10.28 347 2.33 20 4 10.42 218 0.912 13 1 11.74 5.06 0.538 1 2 12.86 136 2.83 20 3 : Analogue to Tab. \[Tab:ratio\_resonances\] but for the volume-preserving driving law.[]{data-label="Tab:vol_resonances"} [9]{} A. Loskutov, A. Ryabov and E. D. Leonel, Physica A [**389**]{}, 5408 (2010). D. R. da Costa, C. P. Dettmann and E. D. Leonel, Phys. Rev. E [**83**]{}, 066211 (2011). E. D. Leonel and C. P. Dettmann, Phys. Lett. A [**376**]{}, 1669 (2012). D. U. Matrasulov, U. R. Salomov, G. M. Milibaeva and N. E. Iskandarov, Physica D [**240**]{}, 470 (2011). F. Lenz, F. K. Diakonos and P. Schmelcher, Phys. Rev. Lett. [**100**]{}, 014103 (2008). A. Loskutov, A. B. Ryabov and L. G. Akinshin, J. Phys. A [**33**]{}, 7973 (2000). D. F. M. Oliveira, J. Vollmer and E. D. Leonel, Physica D [**240**]{}, 389 (2011). E. D. Leonel and L. A. Bunimovich, Phys. Rev. E [**82**]{}, 016202 (2010). E. D. Leonel and L. A. Bunimovich, Phys. Rev. Lett. [**104**]{}, 224101 (2010). F. Lenz, C. Petri, F. K. Diakonos and P. Schmelcher, Phys. Rev. E [**82**]{}, 016206 (2010). K. Shah, D. Turaev and V. Rom-Kedar, Phys. Rev. E [**81**]{}, 056205 (2010). B. Liebchen, R. Büchner, C. Petri, F. K. Diakonos, F. Lenz and P. Schmelcher, New J. Phys. [**13**]{}, 093039 (2011). A. Y. Loskutov, A. B. Ryabov and L. G. Akinshin, J. Exp. Theor. Phys. [**89**]{}, 966 (1999). L. E. Reichl *The transition to chaos*, (Springer, New York, , 1992). H.-J. St[ö]{}ckmann, *Quantum Chaos: An Introduction*, (Cambridge University Press, 1999). E. Fermi, Phys. Rev. [**75**]{}, 1169 (1949). A. J. Lichtenberg and M. A. Liebermann, *Regular and Chaotic Dynamics*, (Springer, New York, 1992). F. Lenz, *Classical and quantum dynamics of driven elliptical billiards*, PhD thesis (2009). T. Hogg and B. A. Huberman, Phys. Rev. Lett. [**48**]{}, 711 (1982). P. Seba, Phys. Rev. A [**41**]{}, 2306 (1990). J. F. Willemsen, Phys. Rev. E [**50**]{}, 3116 (1994). K. Nakamura, S. K. Avazbaev, Z. A. Sobirov, D. U. Matrasulov and T. Monnai, Phys. Rev. E [**83**]{}, 041133 (2011). F. Lenz, B. Liebchen, F. K. Diakonos and P. Schmelcher, New J. Phys. [**13**]{}, 103019 (2011). D. Cohen and D. A. Wisniacki, Phys. Rev. E [**67**]{}, 026206, (2003). R. L. Liboff and M. A. Porter, Chaos [**10**]{}, 366 (2000). J. Orear and E. Fermi, *Nuclear physics: a course given by Enrico Fermi at the University of Chicago*, (University of Chicago Press, 1950). W. Vogel and D. G. Welsch, *Quantum Optics*, (Wiley-VCH, 2006). L. Mandel and E. Wolf, *Optical Coherence and Quantum Optics*, (Cambridge University Press, 1995). M. Born, and V. Fock, Zeitschr. Phys. A [**51**]{}, 165 (1928). R. W. Robinett, Am. J. Phys. [**64**]{}, 440 (1996). R. W. Robinett, Europ. J. Phys. [**24**]{}, 231 (2003). C. Zener, Proc. R. Soc. Lond. A [**137**]{}, 696 (1932).

ArXiv

--- abstract: 'To predict whether a coronal mass ejection (CME) will impact Earth, the effects of the background on the CME’s trajectory must be taken into account. We develop a model, ForeCAT (Forecasting a CME’s Altered Trajectory), of CME deflection due to magnetic forces. ForeCAT includes CME expansion, a three-part propagation model, and the effects of drag on the CME’s deflection. Given the background solar wind conditions, the launch site of the CME, and the properties of the CME (mass, final propagation speed, initial radius, and initial magnetic strength), ForeCAT predicts the deflection of the CME. Two different magnetic backgrounds are considered: a scaled background based on type II radio burst profiles and a Potential Field Source Surface (PFSS) background. For a scaled background where the CME is launched from an active region located between a CH and streamer region the strong magnetic gradients cause a deflection of 8.1${^{\circ}}$ in latitude and 26.4${^{\circ}}$ in longitude for a 10$^{15}$ g CME propagating out to 1 AU. Using the PFSS background, which captures the variation of the streamer belt position with height, leads to a deflection of 1.6${^{\circ}}$ in latitude and 4.1${^{\circ}}$ in longitude for the control case. Varying the CME’s input parameters within observed ranges leads to the majority of CMEs reaching the streamer belt within the first few solar radii. For these specific backgrounds, the streamer belt acts like a potential well that forces the CME into an equilibrium angular position.' author: - 'C. Kay and M. Opher' - 'R. M. Evans' title: 'Forecasting a Coronal Mass Ejection’s Altered Trajectory: ForeCAT' --- Introduction ============ The Sun explosively releases magnetized plasma known as coronal mass ejections (CMEs). @Gop09a cataloged CMEs observed with the Solar and Heliospheric Observatory (SOHO) mission’s Large Angle and Spectrometric Coronagraph (LASCO). This catalog includes CMEs with a wide range of speeds (200-2500 km s$^{-1}$), masses (10$^{13}$-10$^{16}$ g) and kinetic energies (10$^{27}$-10$^{33}$ ergs). Earth-directed CMEs can drive space weather phenomena, producing aurora but also potentially damaging power grids. Shocks driven by CMEs can accelerate particles. At Earth these energetic particles can damage satellites and harm astronauts. The better we understand the trajectory of a CME through the heliosphere, the better we can predict the effects at Earth and throughout the rest of the heliosphere. Since the beginning of CME observations in the 1970s, CME deflections have been observed [@Mac86]. @CaB04 and @Kil09 discuss the trend of high latitude CMEs deflecting toward the equator during solar minimum conditions. Both attribute the deflection to polar coronal holes (CHs). @CaB04 emphasize the role of the fast wind affecting the CME’s expansion. @Kil09 suggest that CMEs cannot penetrate the polar CH magnetic fields which then guide the CME to the equator. @Kil09 also note the correlation between the direction of CME deflections and the decreased tilt of the heliospheric current sheet (HCS) at solar minimum. At other times of the the solar cycle the increased complexity of the HCS configuration may lead to more variation in the direction of deflection. @Xie09 find that slow CMEs ($\le$ 400 km s$^{-1}$) follow a pattern of deflection toward the streamer belt (SB) but some fast CMEs move away from the SB. @Xie09 observe a correlation between the deflection of the CME and the distance between the CME source and the SB for the slow CMEs, but find no such correlation for the fast CMEs. They also find that fast CMEs statistically tend to deflect less than slow ones. Recent observational studies show that CMEs can undergo strong deflections close to the Sun, however, below 5 ${R_{\Sun}}$ deflection cannot be distinguished from nonuniform expansion. Longitudinal deflections are observable using multiple coronagraph viewpoints after the launch of the Solar TErrestrial RElations Observatory (STEREO) with the Sun Earth Connection Coronal and Heliospheric Investigation (SECCHI). These observations confirmed that deflections can occur in both longitude and latitude [@Isa13; @Liu10a; @Liu10b; @Lug10; @Rod11]. @Byr10 reconstruct the 2008 December 12 CME in three dimensions (3D) using an elliptical tie-pointing method. By matching the positions of edges in STEREO Ahead and Behind images they fit a 3D ellipsoid to the CME. They estimate a latitudinal change of 30${^{\circ}}$ in the midpoint of their CME front during propagation up to 7 ${R_{\Sun}}$, but beyond this distance the latitude remains relatively constant. @Liu10b reconstruct the 3D behavior of several events using geometric triangulation: the forward modeling of a flux rope-like structure with self-similar expansion [@The06; @The09]. @Liu10b find a 13${^{\circ}}$ westward deflection within 15 ${R_{\Sun}}$ for the 2007 November 14 CME and about 10${^{\circ}}$ westward within 20 ${R_{\Sun}}$ for the 2008 December 12 CME, but do not address the latitudinal deflection calculated by @Byr10. @Liu10b suggest that this systematic westward deflection may be a universal feature due to the magnetic field connecting the Sun and the CME. If the solar magnetic field is frozen into the plasma of the CME, then corotation with the Sun would explain this motion. The east-west asymmetry driven by a systematic westward deflection of CMEs was first observed in @Wan04. @Wan04 attribute the deflection to the Parker spiral and the speed of CMEs: CMEs traveling faster than the solar wind will deflect to the east and CMEs traveling slower than the solar wind will deflect to the west. @Isa13 use a combination of forward modeling of STEREO-SECCHI and SOHO-LASCO coronagraph images and Grad-Shafranov reconstruction to reconstruct the full three dimensional trajectory of a CME out to 1 AU. @Isa13 reconstruct 15 CMEs from between the minimum of Solar Cycle 23 and the maximum of Solar Cycle 24. The latitudinal deflection of these CME far exceeds the longitudinal deflections. Latitudinal deflections up to 35${^{\circ}}$ are observed and the maximum longitudinal deflection is only 5.4${^{\circ}}$. CME deflection is also studied through the use of magnetohydrodynamic (MHD) simulations. @Lug11 present a MHD simulation of the 2005 August 22 CME using the Space Weather Modeling Framework [@Tot11]. This CME was launched from an anemone active region (AR) within a CH. @Lug11 find that magnetic forces drive a deflection of 10-15${^{\circ}}$ within 8 ${R_{\Sun}}$ which is smaller than the 40-50${^{\circ}}$ expected from observations. The simulated CME is initiated with an out-of-equilibrium flux rope so that the CME does not match the observations within three solar radii of the Sun. The simulated CME reaches its maximum speed of 1500 km s$^{-1}$, only 1.5 minutes after initiation, but beyond 3 ${R_{\Sun}}$ the propagation speed matches the observed value of 1250 km s$^{-1}$. @Lug11 note that the difference in propagation at low heights, where the magnetic forces should be the strongest, could explain some of the discrepancy between the observed and simulated deflections. @Zuc12 compare a MHD simulation to the 2009 September 21 CME which was observed to deflect 15${^{\circ}}$ toward the HCS. In their MHD simulation, reconnection creates an imbalance in the magnetic pressure and tension forces causing the CME to deflect toward the SB. A CH’s influence on a CME has been quantified by defining force vectors based on the CH parameters. To study correlations between a CME’s deflection and the distance, $r$, from its source location to a CH with area, $A$, @Cre06 introduce a force, $\textbf{F}=A/r \; \hat{r}$. This force points in $\hat{r}$, defined as the direction pointing away from the CH toward the CME. They find a correlation between the direction of **F** and the direction of the CME deflection suggesting that CHs do influence the CME motion. @Gop09b define an “coronal hole influence parameter” (CHIP) similar to the force of @Cre06, but also incorporate the magnetic field strength (B) of the CH, $\textbf{F}=B^2A / r \; \hat{r}$. @Gop09b find good agreement between the direction a CME propagates after it deflects and the direction given by vector sum of all the individual F-vectors from nearby CHs. @Gop10 update the r-dependence to $r^{-2}$ giving a final form $\textbf{F}=B^2 A /r^2 \; \hat{r}$. @Moh12 compare Solar Cycle 23 CMEs originating from disk center with their CHIP parameter as a function of the solar cycle and CME type. Driverless shocks tend to have the largest CHIP value, magnetic clouds (MC) the smallest, with non-MCs falling in between. The CHIP values are smallest during the rising phase. @Moh12 suggest CHs deflect CMEs away from the Sun-Earth line which provides support for the idea that all CMEs may be flux ropes, the distinction between MCs, non-MCs and driverless shocks being a matter of viewing perspective. @She11 and @Gui11 consider gradients in the magnetic energy density of the background solar corona as an explanation for the observed deflection. At the distances of their observations ($\ge 1.5 {R_{\Sun}}$) these gradients point toward the streamer region. During solar minimum conditions the streamer region is generally centered near the equator so mainly latitudinal deflections will occur. At other times, the coronal magnetic field becomes more complex so a wider variety of gradient directions exists. @She11 present a theoretical approach that compares favorably with observations. @Gui11 extend the work with additional observations and find that the direction of deflection tends to agree with the direction of the background gradients. Similar to the deflection of CMEs, @Pan11 investigate the rolling motion of prominences/filaments. They find that the prominences tend to roll away from CHs before they form flux ropes. @Pan11 suggest that the filament motion could be explained by local magnetic force imbalances within the filament arcade, whereas the non-radial motion of CMEs would result from similar imbalances on global scales. This paper presents a model for CME deflection near the Sun by considering the effects of magnetic pressure gradients as well as magnetic tension. Magnetic energy dominates the free energy budget of the ambient plasma in the lower corona, so magnetic forces play an important role in the deflection of CMEs near the Sun. The closer to the Sun a CME is, the stronger the surrounding coronal magnetic fields and therefore the stronger the forces that act upon a CME. The magnetic field strength falls off quickly with distance so the magnetic forces should as well. Other effects, not included in ForeCAT, can cause deflection such as interactions with other CMEs propagating through the interplanetary medium [@Lug12; @Tem12] or variations in the speed of the background solar wind. Spatial velocity variations can distort the shape of a CME, seen in observations [@Sav10] and in MHD models [@Wan03]. If unbalanced, these effects on opposite sides of the CME could cause deflection. We focus only on the magnetic forces close to the Sun, ignoring magnetic reconnection. The model, called ForeCAT (Forecasting a CME’s Altered Trajectory), calculates the deflection of a CME within a plane defined by global magnetic pressure gradients. The deflection motion of the CME not only depends on the magnetic forces but requires models for the CME expansion and propagation as well. ForeCAT uses the expansion model from the melon-seed-overpressure-expansion model (MSOE) of @Sis06. A three-part propagation model, similar to @Zha06, determines the CME’s radial motion. The CME starts with a slow rise phase which transitions to an acceleration phase, then finally to a constant speed propagation phase. ForeCAT also includes the effects of drag hindering the CME’s nonradial motion, so that the CME cannot propagate freely in a direction quasi-perpendicular to the solar wind flow. ForeCAT’s radial propagation model results from fitting observations of CMEs affected by drag so ForeCAT does not explicitly include drag in the radial direction. The paper is organized as follows: Section 2 contains analytic descriptions of the magnetic forces driving the deflection. Section 3 describes the expansion, propagation and drag models; section 4 presents two models for the background solar magnetic field: a scaled model and a Potential Field Source Surface (PFSS) model. Section 5 shows the results of a test case using the scaled magnetic background; section 6 investigates ForeCAT’s sensitivity to input parameters, both CME parameters and values assumed in the analytic propagation model. Section 7 includes results from using the PFSS magnetic field model. Section 8 looks at deflection from local gradients related to the AR. Analytic Model of CME Deflection ================================ Deflection Plane ---------------- In order to simplify the treatment of the CME deflection in the lower corona, we restrict the calculations within ForeCAT to a plane called the “deflection plane”. In ForeCAT, magnetic forces drive the deflection so the background coronal magnetic field gradients at the location from which the CME launches determine the direction of the deflection plane. The normal to this plane is defined as the cross product of the direction of initial radial CME motion and the direction of the dominant background magnetic pressure gradients. The calculation of the deflection plane normal vector uses the direction of the gradients in the magnetic pressure at a single location, which requires picking a specific height. It is expected that the direction of the magnetic pressure gradient will change with distance from the Sun. At smaller distances the local effects of the AR from which the CME is launched dominate the gradient, and at further distances effects from global features such as CHs and SBs dominate. For the magnetic background used in this study, the effects from global features dominate at distances of 2 ${R_{\Sun}}$ or larger. ForeCAT uses the direction of the gradient at this 2 ${R_{\Sun}}$ to define the deflection plane to capture the effects of the CH and SB. Figure 1 shows four panels illustrating how different features determine the gradients at different heights. The figures show a constant height from a MHD simulation using the Space Weather Modeling Framework (@Tot11, @van10, @Eva12, see section 4 for details) centered around the AR from which the CME is launched. The figure contains color contours corresponding to the logarithm of the magnetic pressure and the arrows show the direction of the nonradial magnetic pressure gradient unit vectors. The white dot indicates the latitude and longitude from which the CME launches. All panels use the same color contour scale. The strongest magnetic pressure occurs close to the AR, visible in Figure 1a (a distance of 1.05 ${R_{\Sun}}$). Figure 1b shows a distance of 1.5 ${R_{\Sun}}$ where both the local effects of the AR and the global effects of the CH and SB influence the gradients. In Figure 1c and 1d (distances of 2 ${R_{\Sun}}$ and 3 ${R_{\Sun}}$) the streamer region that becomes the HCS can be seen as a minimum in the magnetic pressure. In these panels the magnetic pressure is weaker than magnetic pressure in Figure 1a by 2-3 orders of magnitude. At larger distances, the gradients transition from being dominated by local features, such as ARs, to a more uniform configuration, determined by the global structure of CHs and SBs. For this background, 2 ${R_{\Sun}}$ is the smallest radius at which the gradients are dominated by the global effects. These global gradients are present closer to the Sun, but can only be easily separated from the local gradients at larger distances. Between 2 ${R_{\Sun}}$ and 3 ${R_{\Sun}}$ the direction of the gradient at the CME launch position changes by less than six degrees. ForeCAT uses the value at the smaller radius where the magnetic field is stronger as more deflection will occur near that height. While the AR’s magnetic field does affect the CME’s propagation, and it is included in the calculations, the current focus of ForeCAT is deflection due to global gradients resulting from the orientation of CHs and SBs and the differences in these magnetic fields. The effects of the AR are explored in section 8. The normal for the deflection plane is given by $$\vec{n}=\vec{R}_0 \times -\vec{\nabla}\left(\frac{B^2}{8\pi}\right)$$ where $\vec{R}_0$ is the initial radial vector for the CME and $\vec{\nabla}B^2/8\pi$ is the gradient in the magnetic field pressure at a distance of 2 ${R_{\Sun}}$. Figure 2 illustrates how the deflection plane is selected. Figure 2a shows color contours of the magnetic field strength at distances of 1.05 and 2 ${R_{\Sun}}$, analogous to Figure 1a and 1c, in three dimensions using a color scale appropriate for the range at each distance. At 1.05 ${R_{\Sun}}$ the red lines indicate the approximate position of the CHs. The black circle marks the latitude and longitude of the CME’s initial position. The radial vector $\vec{R_0}$ extends from the center of the Sun through this point. The black line shows the orientation of the deflection plane, defined by the direction of $-\vec{\nabla}B^2 / 8\pi$ at the black circle at 2 ${R_{\Sun}}$. The radial direction and the gradient vector from Figure 2a define the deflection plane in Figure 2b. The schematic in Figure 2b includes an example deflection plane and the Sun’s surface. As shown in Figure 2b, the deflection plane can be tilted, it need not be an equatorial or meridional plane. Figure 3 contains a schematic showing the Sun-CME configuration: a) shows the CME, b) shows within the deflection plane. Figure 3a shows the Sun in white and the flux-rope-like CME in yellow, as well as the intersection of the CME and deflection plane. Figure 3b represents features within the deflection plane. The background solar magnetic field is defined in polar coordinates, $R$ and $\phi$, with the origin at the center of the Sun. A second set of polar coordinates, $r$ and $\theta$, with origin at the point on Sun ($ =1{R_{\Sun}}$) from which the CME is launched, is used to calculate deflection forces on the CME. The set of Cartesian coordinates, with origin also at the center of the Sun, allow conversion between the two sets of polar coordinates: $$R=\sqrt{x^2+y^2} \qquad r=\sqrt{(x-1)^2+y^2}$$ $$\phi=\tan^{-1}\left(\frac{y}{x}\right) \qquad \theta=\tan^{-1}\left(\frac{y}{x-1}\right)$$ where $x$, $y$, $R$, and $r$ have units of ${R_{\Sun}}$ and $x$ and $y$ correspond respectively to the $\vec{R_0}$ and $\vec{\nabla}B^2/8\pi$ directions in Figure 2. In this geometry the CME launches along the x-axis which corresponds to $\theta=\phi=0$. A circle, initially of radius $L_0$, represents the cross-section of the CME in the deflection plane. For the cross-section to be a perfect circle requires the CME to be perpendicular to the deflection plane. Deviations from this orientation will introduce small errors into the calculation of the CME density as the cross-section will take an elliptical shape within the deflection plane. ForeCAT uses the deflection forces on two “edges”, marked with X’s in Figure 3, to calculate the total deflection. These edges correspond to the points on the circle that lie on a line running through the center of the circle and perpendicular to the $\hat{r}$ direction at that point. Averaging the $\phi$ values of the two edges gives the central position angle (CPA) of the CME, which equals the $\phi$ position of the center of the circle: $$CPA=\frac{\phi_{1}+\phi_2}{2}$$ where 1 and 2 refer to the two edges. The CPA is calculated using the change in the Sun-centered angle, comparable to latitudinal or longitudinal changes of CMEs in coronagraph observations. The net deflection force on the two edges determines the change in the $\theta$ position of the CME ($\theta\rightarrow\theta'$). Before a CME detaches from the solar surface the deflection motion will occur with respect to the position where the footpoints are anchored. Accordingly, ForeCAT calculates deflection forces in the $\hat{\theta}$ direction. Different analytic models describe the change in distance ($R\rightarrow R$’), change in CME radius ($L \rightarrow L$’), and effects of nonradial drag, separate from the deflection (see section 3). No change occurs in the CPA for a CME propagating without deflection and with uniform expansion. Deviations from the original CPA correspond to deflection or non-uniform expansion, however ForeCAT only includes CMEs with uniform expansion (section 3.2). Deflection Forces ----------------- ForeCAT calculates CME deflection due to the magnetic tension and the magnetic pressure gradient. Imbalance of these forces between the two edges causes a net force in the $\hat{\theta}$ direction, driving deflection. All forces within this model are volumetric so that the acceleration equals the force divided by the density. ### Magnetic Tension In general the force due to magnetic tension can be expressed as $$F_{\kappa}=\vec{\kappa}\frac{B^2}{4\pi}$$ where $\kappa=1/R_C$ is the curvature and $R_C$ is the radius of curvature. The tension force points toward the center of curvature. As the CME expands into the surrounding medium the external magnetic field will drape around it. The curvature of the draped magnetic field can be approximated then as equal to the CME curvature with $R_C$ as the radius of the CME cross-section within the deflection plane. The draping of the coronal magnetic field is not restricted to the deflection plane so ForeCAT includes a $\cos \alpha$ factor to account for this, assuming the radius of curvature does not change. The angle $\alpha$ is the angle between the deflection plane and the direction of the draping of the background solar magnetic field lines around the CME. In principle $\alpha$ will vary in time. The final tension force on each edge is $$\vec{F}_{\kappa}=\mp\frac{1}{L}\frac{B^2}{4\pi}\cos\alpha\;\hat{\theta}$$ where the top edge (defined as the edge with the largest $y$ value in the Cartesian coordinate system in Figure 3b) has the negative sign and the bottom the positive. Only for a background magnetic field that is symmetric about the CME will the two forces balance. ### Magnetic Pressure Gradient The component of the magnetic pressure gradient perpendicular to the radial direction also leads to deflection: $$F_{\nabla P}=-\nabla_{\perp}\frac{B^2}{8\pi}$$ where the $\perp$ corresponds to gradients perpendicular to the direction of the magnetic field according to the definition of the Lorentz force. The magnetic pressure gradient expression used in ForeCAT includes the $\cos\alpha$ factor to account for draping out of the deflection plane. Since the background magnetic field lines drape around the CME, at the edges the direction of the perpendicular gradient within the deflection plane is the $\hat{\theta}$ direction. Equation 5 can be recast as $$\vec{F}_{\nabla P}=-\frac{B}{4\pi R}\frac{\partial B}{\partial \phi}\cos(\theta-\phi)\cos\alpha\;\hat{\theta}$$ with the force directed in the $\hat{\theta}$ direction due to the $\cos(\theta-\phi)$ term which results from taking the $\hat{\theta}$ component of the gradient in $\phi$. As the CME propagates away from the Sun the orientation of the background magnetic pressure gradients may change. When this occurs deflection occurs out of the original deflection plane. The net out of plane deflection is minimal as the magnetic forces decrease with distance. ### Total Deflection Force The net volumetric deflection force is given by the sum of equations 4 and 6. $$\vec{F} = \left(\mp\frac{1}{L}\frac{B^2}{4\pi}\cos\alpha-\frac{B}{4\pi R}\frac{\partial B}{\partial \theta}\cos(\theta-\phi) \cos\alpha\right) \hat{\theta}$$ where $\hat{\theta}=-\sin\theta\hat{x}+\cos\theta\hat{y}$, which changes with time as $\theta$ changes. Dividing the deflection force by the CME density gives the acceleration of each CME edge. The density is defined as: $$\rho=\frac{M}{\pi^2 r L^2}$$ which $M_{CME}$ is the CME mass and the volume is approximated using a uniform curved cylinder of length $\pi r$ and cross-section $\pi L^2$. The mass of the CME is assumed to be constant. CMEs tend to accrete mass as they travel. @Vou10 analyze the mass evolution of CMEs in the low corona using the coronagraph brightness and find that CMEs tend to increase in mass in the corona below 10 ${R_{\Sun}}$. ForeCAT’s assumption of a constant mass will cause underestimations of the density, leading to an overestimate of the acceleration cause by deflection. The density evolves in time due to the expansion of the CME (see section 3.1). The acceleration of edges in the $\theta$ direction is a linear acceleration with $x$ and $y$ components. The equations of motion for an edge are: $$\begin{aligned} x(t+\Delta t) &=& x(t)+v_{def,x}(t)\Delta t-0.5\frac{F(t)}{\rho(t)}\sin\theta (t) \Delta t^2\\ y(t+\Delta t) &=& y(t)+v_{def,y}(t)\Delta t+0.5\frac{F(t)}{\rho(t)}\cos\theta (t) \Delta t^2 \nonumber\end{aligned}$$ where $v_{def,x}$ and $v_{def,y}$ are the velocities of the edge in the $\hat{x}$ and $\hat{y}$ direction resulting from deflection $$\begin{aligned} v_{def,x}(t)&=&-\int^t_0 \frac{F(t)}{\rho (t)} \sin\theta (t) dt\\ v_{def,y}(t)&=&\quad\int^t_0 \frac{F(t)}{\rho (t)} \cos\theta (t) dt \nonumber\end{aligned}$$ The deflection equals the change in the CME’s CPA $$CPA(t)=\frac{1}{2}\left[\tan^{-1}\left(\frac{y_1(t)}{x_1(t)}\right)+\tan^{-1}\left(\frac{y_2(t)}{x_2(t)}\right)\right]$$ where the 1 and 2 refer to the two CME edges. Since initially the CPA equals zero, the total deflection at any time equals the CME’s current CPA. The total deflection within the deflection plane can be converted into a change in latitude and longitude using the orientation of the plane. Description of CME Motion ========================= To calculate the total deflection, the radial propagation, expansion, and nonradial drag of the CME must be incorporated as they affect the position of the CME edges over time. CME Expansion ------------- Several analytic models describing CME evolution exist. These models focus on the radial propagation of CMEs and do not account for deflections. @Pne84 introduces the melon-seed model, which @Sis06 later develop into the melon-seed-overpressure-expansion (MSOE) model. More complex models exist such as that of @Che96, which treats the CME as a flux rope containing two different plasmas, representing the cavity and embedded prominence, and triggers the eruption by increasing the poloidal magnetic field. ForeCAT uses the MSOE model’s description of CME expansion. The MSOE model modifies a classical hydrodynamic solution for an overpressure of a spherical cavity. The hydrodynamic solution (see @Mil68) is driven by an adiabatic gas overpressure that can be treated as a fluid “source.” @Sis06 change the adiabatic overpressure to a magnetic overpressure. This results in the following expression for the change in the CME radius L (see Appendix for more details). $$\frac{\partial^2L}{\partial t^2}=\frac{1}{L}\left[-\frac{3}{2}\left(\frac{dL}{dt}\right)^2 + \left(\frac{L_0}{L}\right)^4\frac{(A_{h0}^2-A_{SW0}^2)}{2}\frac{\rho_{SW0}}{\rho_{SW}}\right]$$ The subscript $0$ indicates initial values evaluated at t=0. $A_{h0}$ and $A_{SW0}$ refer to Alfvén speeds calculated using the initial background solar wind density and either the magnetic field of the CME ($A_{h0}$) or the background solar magnetic field strength ($A_{SW0}$). The expansion equation depends on the background solar wind density $\rho_{SW}$ which requires assuming some solar wind density profile. ForeCAT uses the expression for the density from @Che96, also used by @Sis06: $$\rho_{SW}(R)=6.68\times10^{-16}\left[3\left(\frac{{R_{\Sun}}}{R}\right)^{12}+\left(\frac{{R_{\Sun}}}{R}\right)^4\right]+3.84\times10^{-19}\left(\frac{{R_{\Sun}}}{R}\right)^{2} \mathrm{g} \; \mathrm{cm}^{-3}$$ Using the profile from @Che96, the CH regions are scaled down by a constant value as $\rho_{CH}(R)=0.25\rho_{SW}(R)$, which produces a CH density profile closer to that of observations [@Guh98; @Doy99]. The value 0.25 results from assuming constant mass flux and a solar wind speed roughly double the slow wind for the fast wind [@Mcc00]. Figure 4 shows the analytic density model and several radial profiles from a MHD simulation, the same simulation shown in Fig. 1 and 2. The details of the simulation are discussed in section 4.1. The MHD profiles come from different locations: above an AR (red), a CH (blue), and the SB (green). The standard analytic model (solid black) is shown in addition to the scaled CH analytic model (dashed black). Close to the Sun the analytic model overestimates the MHD solution by nearly an order of magnitude. Near 20 ${R_{\Sun}}$, the outer boundary of the MHD simulation domain, the analytic and MHD profiles for the SB and CH have better agreement. However, some discrepancy still exists for the AR. The effects of the chosen analytic density profile will be explore in a future work. CME Propagation --------------- ForeCAT adopts an analytic expression for the radial propagation speed based on empirical fits. The Lorentz force that drives the CME deflection also causes radial motion of the CME, however, ForeCAT makes use of an empirical model. @Zha01 [@Zha04] and @Zha06 present a three phase description of CME propagation (initiation or gradual rise, impulsive acceleration, and propagation) and the connection to X-ray flare observations. The initiation phase occurs first as the CME slowly rises, then due to an instability or reconnection the CME lifts off and begins rapidly accelerating away from the Sun. The transition to the acceleration phase often correlates with the onset of flare activity. Both the initiation and propagation phases occur within a few solar radii of the Sun [@Zha06]. The final phase is the propagation phase where comparatively little CME acceleration occurs. In a statistical study of 50 CMEs, @Zha06 found average main accelerations of 330.9 m s$^{-2}$ whereas the average residual acceleration during the propagation phase was only 0.9 m s$^{-2}$. @Vrs07 and @Bei11 identify the Lorentz force as the driving mechanism behind the impulsive acceleration in the radial direction, explaining why this phase occurs so close to the Sun’s surface. The observational studies of both @Bei11 and @Jos11 show that the acceleration phase tends to end by a height of about 2 ${R_{\Sun}}$. Although the acceleration phase contains the most rapid acceleration, a CME continues to accelerate during the propagation phase, @Che10 refer to this as the post-impulsive-phase acceleration. The @Che10 study of several hundred CMEs results in a mean post-impulsive-phase acceleration value equal to -11.9 m s$^{-1}$ with individual values ranging between -150 and 180 m s$^{-1}$. The data from @Zha06 cover a similar range of post-impulsive-phase accelerations as the data from @Che10 but the two means differ as a result of using different subsets of CMEs from the LASCO catalog. The positive and negative post-impulsive-phase acceleration values imply that CMEs can either accelerate or decelerate in the third phase due to the drag force from the CME’s interaction with the solar wind. ForeCAT adopts the three part propagation model for the radial dynamics. ForeCAT uses a constant velocity for the initiation and propagation phase and a constant acceleration in the acceleration phase. We define the radial distance at which the CME transitions from the gradual to acceleration phase as $R_{ga}$ and that from the acceleration to propagation phase as $R_{ap}$. We assume that the initiation phase lasts until the center of the CME cross-section reaches a distance $R_{ga}=1.5{R_{\Sun}}$ and then the acceleration occurs until $R_{ap}=3.0{R_{\Sun}}$. ForeCAT uses a single representative value for each transition, as well as a constant value for the gradual velocity of the CME in the initiation phase, $v_g$=80 km s$^{-1}$. @Zha06 observe $v_g$ between tens of km s$^{-1}$ up to 100 km s$^{-1}$. Section 6 contains analysis of ForeCAT’s sensitivity to the parameters $R_{ga}$, $R_{ap}$, and $v_g$. Given the above assumptions, the CME’s radial propagation is described by its final velocity, $v_f$ at the propagation phase. From the kinematic evolution of the CME during acceleration phase $$v_f^2=v_g^2+2a(1.5{R_{\Sun}})$$ which corresponds to an acceleration, a, $$a=\frac{v_f^2-v_g^2}{3 {R_{\Sun}}}$$ which allows us to describe the CME’s radial velocity over time as $$\begin{aligned} v_{r}&=&v_g \qquad 1.0 {R_{\Sun}}\leq R \leq R_{ga} \\ \nonumber v_{r}&=&v_g+0.5a(t-t_{ga}) \qquad R_{ga} \le R \leq R_{ap} \\ \nonumber v_r&=&v_f \qquad R \ge R_{ap} \end{aligned}$$ where t$_{ga}$ is the time at which the CME reaches R$_{ga}$. Equation 16 produces CME velocity profiles similar to those present in Figure 1 of @Zha06, with the exception of flat initial and propagation phases. Nonradial Drag -------------- We include the nonradial drag as the component of the drag force in the $\hat{\phi}$ direction which results from the interaction of CME with the solar wind. ForeCAT does not explicitly calculate drag in the radial direction since the propagation model describes a CME’s radial motion. To calculate the nonradial drag, ForeCAT uses the expression for the acceleration due to drag from @Car04: $$\vec{a}_D=-\frac{C_dA\rho_{SW}}{M_{CME}}(\vec{v}_{CME}-\vec{v}_{SW})|\vec{v}_{CME}-\vec{v}_{SW}|$$ where $C_d$ is the drag-coefficient, $A$ the cross-sectional area in the direction of the drag, and $v_{SW}$ the solar wind speed. @Car04 use this equation to describe the radial drag on a CME, but the same physical process governs drag in all directions. The solar wind is approximated as entirely radial so that the solar wind velocity term equal zero in the expression for the drag in the $\hat{\phi}$ direction. Close to the Sun this approximation is the least accurate but it allows ForeCAT to include nonradial drag without invoking a complete solar wind velocity model. The cross-sectional area in the direction of the drag can be expressed as: $$A=\frac{\pi}{2}\left((r+L)^2-(r-L)^2\right)=2\pi rL$$ based on the definition of the CME structure in section 2. Setting $C_d$=1 and taking the component of the CME velocity in the radial direction, equation 17 becomes: $$\vec{a}_d=-\frac{2\pi rL\rho_{SW}}{M_{CME}}\left(-v_x\sin(\phi)+v_y\cos(\phi))\right|-v_x\sin(\phi)+v_y\cos(\phi)|\hat{\phi}$$ where $v_x$ and $v_y$ are the x and y components of the CME’s velocity. In section 6 we explore ForeCAT’s sensitivity to the drag coefficient and discuss models for $C_d$. Background Solar Magnetic Field =============================== To calculate a CME’s deflection, ForeCAT includes the magnetic structure of various features such as CHs, ARs, and SBs. The background solar magnetic field therefore is crucial. Two different background magnetic field models were explored with ForeCAT: a “scaled” background and a Potential Field Source Surface (PFSS) background. Scaled Background ----------------- The scaled background uses the background magnetic field from the output of an MHD steady state solar wind from the Space Weather Modeling Framework (SWMF, @Tot12, @van10) using a magnetogram as input. Alfvén waves drive the background solar wind and surface Alfvén wave damping adds heating [@Eva12]. The magnetic field values from a ring at $R=1.15{R_{\Sun}}$ within the deflection plane (defined using the magnetic pressure gradients at a distance of 2 ${R_{\Sun}}$) within $\pm 90{^{\circ}}$ of the CME launch location yield discrete points for the magnetic field strength as a function of angle. By extracting values at low heights, $B(\phi )$ includes the signatures of the solar features (CHs, SBs and ARs). ForeCAT uses the MHD background only at $R=1.15{R_{\Sun}}$ and uses extrapolations for larger radii based on observational studies of the solar magnetic field. ForeCAT uses these extrapolations because, as described below, the MHD solution falls unrealistically quickly. The extrapolations differ between ARs and non-AR locations, also described below. The extrapolations for ForeCAT’s magnetic field model result from observations of the solar magnetic field versus distance. Observational studies of the magnetic field of ARs fit the profile of the magnetic field versus distance with the form $B=B_0R^{-\alpha}$. @Dul78 present a compilation of observational data(including data from Helios, Mariner 10, and various ground-based solar telescopes) of the magnetic field above an active region. The study finds that $B=0.5[(R/{R_{\Sun}})-1]^{-1.5}$ agrees within a factor of three for all the observations. @Pat87 use Helios measurements of Faraday rotation and find a best fit between 3 and 10 ${R_{\Sun}}$ using a combination of $\alpha=2$ and $\alpha=3$. More recent Faraday rotation measurements have been acquired for $R$ between 6.2 and 7.1 ${R_{\Sun}}$ using the Very Large Array which agree with a coefficient of $\alpha$=1.3 [@Spa05] . In order to study shock development in the corona, @Man03 use a background magnetic field combining a $R^{-2}$ term for the quiet sun and a dipole term ($\propto \; R^{-3}$) to represent the ARs. @Gop11 use the standoff distance of CME-driven shocks to determine the magnetic field profile between 6 and 23 ${R_{\Sun}}$ and find good agreement with @Dul78 and @Spa05. Fitting polynomials to these points allows generalization of the discrete magnetic field points to a function that can be used for all $\phi$ angles. In addition, it allows for calculation of analytic derivatives. Separate polynomials describe the AR and the quiet sun (QS, defined as the region outside of the AR). The ranges of the polynomials are determined by the location of local maxima and minima in $B(\phi )$, having the polynomials break at inflection points provides the best fit. First polynomials are fit to the QS yielding a function $B_{QS}(R,\phi)$. The QS magnetic field is subtracted from the MHD result and then the AR polynomials, $B_{AR}(R,\phi)$ , are fit to the residual magnetic field. Figure 5a shows the simulation data (solid red line) as well as the sum of the QS and AR best-fit polynomials (dashed black line) for $B(R=1.15{R_{\Sun}},\phi)$ for $-90{^{\circ}}\le\phi\le 90{^{\circ}}$. Within the deflection plane, the AR corresponds to two local maxima in $B(\phi)$ whereas both the SB and CH represent local minima. The center of the AR is a local minimum corresponding to the polarity inversion line (PIL) located in between the two maxima corresponding to the opposite polarity flux systems. Weaker magnetic field exists at the SB minimum than the CH minimum and this asymmetry produces the gradients that drive deflection. This example contains strong gradients due to the proximity of the CH and SB. Any coronal configuration will have gradients leading to deflection, the magnitude of the deflection depending on the magnitude of the gradients. Scaling the values from $R=1.15{R_{\Sun}}$ determines the background magnetic field strength at other radii. ForeCAT’s model treats the AR like a dipole so that the magnetic field falls as $R^{-3}$. Outside the AR, for the QS the model uses the $R^{-2}$ dependence commonly used for open field lines. This combination of scaling is the same as that of @Man03. Equation 20 gives the ForeCAT scaled magnetic field. $$B (R, \phi) = B_{QS}(1.15 {R_{\Sun}}, \phi) \left(\frac{1.15 {R_{\Sun}}}{R}\right)^2 +B_{AR}(1.15 {R_{\Sun}}, \phi) \left(\frac{1.15 {R_{\Sun}}}{R}\right)^3$$ @Man03 use type II radio bursts to infer the behavior of the background solar magnetic field. The type II radio bursts are believed to result from shock waves propagating outward in the corona [@Nel85]. The speed of the disturbance driving the shock can be used to infer the background Alfvén speed and because the type II emissions occur at the local background plasma frequency the Alfvén speed can be used to determine the background magnetic field strength. @Man03 compare the Alfvén profile from their magnetic field model with the general behavior of type II radio bursts. The combination of a scaling of $R^{-2}$ and $R^{-3}$ yields favorable comparisons to the type II radio observations. In particular, the model produces an Alfvén profile with a local minimum and maximum in the low corona which allows for the formation, decay, and reformation of shocks within 6 ${R_{\Sun}}$, reproducing a two shock wave behavior seen in some type II radio observations [@Gop02]. Figure 5a shows the scaled model magnetic field strength for several different radii ($R$=1.5, 2.0, and 2.5 ${R_{\Sun}}$). As the radius increases the signatures of the individual solar features weaken but are still present at 2.5 ${R_{\Sun}}$. The MHD results for $R$=1.5 are also included. ForeCAT uses analytic fits to observations rather than the results of MHD simulations because for $R<2{R_{\Sun}}$ , as seen in Figure 5a, the MHD magnetic field strength decreases quicker with distance than observations, closer to $r^{-6}$ or $r^{-8}$ depending on the region (CH or SB versus AR). Recent advancements in the MHD model have included a two-temperature (electron and proton) formalism, including the effects of field-aligned heat conduction, radiative cooling, collisional coupling, and wave heating [@Dow10; @van10; @Sok13]. In addition, a Finite Difference Iterative Potential Solver (FDIPS; @Tot11) can be used to initialize the magnetic field in place of the spherical harmonic expansion approach. However, the wave-driven model output (with either FDIPS or harmonic coefficients) does not show a difference in the rapid decrease of the magnetic field magnitude with radial distance. Beyond 2 ${R_{\Sun}}$ the MHD magnetic field falls as the expected r$^{-2}$, but, as discussed in Section 5, the deflection of the CME depends crucially on the magnetic strength at these distances. The rapid decrease of the magnetic field in MHD simulations will lead to an underestimate of the magnetic deflection, which could explain the discrepancy between the observed and simulated CME in @Lug11. @Eva08 show that the steepness of the MHD profiles would allow slow CMEs to drive shocks low in the corona and that the Alfvén speed profiles do not have the characteristic “valley” and “hump” shape seen in analytic models. Using a scaled model, we capture a slower decrease of B with distance, consistent with some type II radio observations. However, this model does not allow a change in the angular position with distance of coronal structures such as the SB. PFSS Background --------------- PFSS models were first used to describe the solar magnetic field in the late 1960’s [@Alt69; @Sch69]. If the magnetic field is assumed to be potential, it can be described using a sum of Legendre polynomials. The harmonic coefficients can be determined from a magnetogram and by assuming the magnetic field becomes entirely radial at the source surface height. The magnetic field at any location can be calculated using the harmonic coefficients. The literature contains extensive discussion of the details of PFSS calculations and the model’s ability to reproduce observed conditions [@Hoe82; @Luh02; @Neu98; @Ril06; @Wan92; @Wan93]. ForeCAT uses a PFSS magnetic field strength calculated using radial harmonic coefficients [@Wan92] from the Michelson Doppler Imager on SOHO [@Sch95], calculated with a source surface height of 2.5 ${R_{\Sun}}$. The PFSS magnetic field is calculated using coefficients for Legendre polynomials up to order 90. Higher order polynomials represent spatially smaller features and decay faster with distance. Since the magnetic field strength (which drives ForeCAT deflection) is strongest close to the Sun, not including the higher orders could make a difference in the CME deflection. Figure 5b contains the PFSS magnetic field (in blue), as well as the MHD results (in red) for $R$ = 1.05, 1.15, 2.0 and 2.5 ${R_{\Sun}}$ within $\pm 50 {^{\circ}}$ of the location from which the CME launches. In Fig. 5b the PFSS model and the MHD model agree. @Ril06 find similar agreement between the MHD and PFSS solutions when strong currents are not present. Both models also show a clear change in the angular magnetic field profile with distance, an effect that the scaled model cannot capture. The “rigid” magnetic minima of the scaled model exists at 1.15 ${R_{\Sun}}$ but at 2.5 ${R_{\Sun}}$ the formation of the HCS near -5${^{\circ}}$ causes a different magnetic minima. This change of the magnetic minima will affect the CME’s magnetic deflection. The PFSS model and the MHD model fall similarly with distance. The PFSS background will underestimate magnetic forces compared to the scaled background. Numerical Implementation and Test Case ====================================== Equations 9, 12 and 16 form a set that describes the evolution of the CME as it propagates away from the Sun for the model, ForeCAT. Initializing the equations requires a radius of the cross-section of the CME within the deflection plane, $L_0$, height, $r_0$, the CME mass, $M_{CME}$, the final propagation velocity, $v_g$, and the magnetic field strength of the CME that causes the initial overpressure, $B_0$. We assume that the angle $\alpha$ equals zero throughout the simulation (no draping outside of the deflection plane) and therefore find a maximum deflection angle. The model also needs the background magnetic field configuration. ForeCAT integrates these equations numerically using a second-order Taylor expansion for the position so that the error is of order $\Delta t^3$. ForeCAT yields a deflection of the CPA over time, as well as the trajectory of the CME as it deflects. In the control case, the following values are chosen for the free parameters of ForeCAT: $M_{CME}$=10$^{15}$ g, $v_{g}$=475 km s$^{-1}$, $L_{0}$=0.15 ${R_{\Sun}}$ and $B_0$= 15 G. These input parameters represent an average CME mass and the velocity corresponds to the mean value from the @Gop09a analysis of the SOHO/LASCO survey of CMEs before the end of 2006. @Sis06 use a similar value of CME magnetic strength. The CME begins at a height of 0.25 ${R_{\Sun}}$ so the model captures some of the gradual phase of radial propagation. The CME launches from AR 0758 of Carrington Rotation (CR) 2029. This corresponds to the magnetic field background shown in Figure 5, this case using a scaled magnetic background. The deflection plane was defined using the magnetic pressure gradient at $R$=2${R_{\Sun}}$ and a latitude of -8${^{\circ}}$ and Carrington longitude of 130.6${^{\circ}}$. Strong gradients that exist between the SB and CH should cause a large deflection. Figure 6a shows the CME’s propagation out to a distance of about 10 ${R_{\Sun}}$ in a Cartesian coordinate system with the Sun at the origin. The figure shows the diameter of the CME parallel to the y-axis (shown with a red line in Figure 3) in 1 minute time-steps. Figure 6b shows the CPA (Eq. 11) of the CME versus distance out to 1 AU. The CME deflects -27.0${^{\circ}}$ in the deflection plane during propagation out to 1 AU. This is equivalent to a change of -8.1${^{\circ}}$ in latitude and -26.4${^{\circ}}$ in longitude. The majority of the deflection occurs while the CME is in the gradual rise and acceleration phases ($R\le$3${R_{\Sun}}$). By 5 ${R_{\Sun}}$ the CME comes close to a constant angular position: the CPA changes less than 1${^{\circ}}$ between 5 ${R_{\Sun}}$ and 1 AU. Fig. 6b shows that beyond about 10 ${R_{\Sun}}$ the CME’s angular motion reverses and it slowly moves in the opposite direction. This motion causes a change in the CPA of less than a degree and can be explained by a change in the direction of the forces acting upon the CME. The net deflection force comes from summing over the two CME edges. Figure 7 shows the magnetic tension and magnetic pressure gradient force, respectively in red and blue, versus distance from the center of the Sun. The figure also shows the total force (tension plus pressure gradient) in black. Fig. 7 highlights the strongest forces, which occur close to the Sun. Beyond 1.7 ${R_{\Sun}}$ the forces have decreased by several orders of magnitude from the values during the first few time-steps and are not included in the figure. The force continues to decrease as the magnetic field decreases with distance. Initially both the magnetic tension and magnetic pressure gradients force the CME toward the SB, the tension force being about twice as strong as the magnetic pressure gradient force. The CME motion can be explained by considering the angular magnetic profiles in Fig. 5 as a series of potential barriers and wells since the deflection forces all depend on the magnetic field strength. Initially, the magnetic pressure gradient force at the edge near the CH (hereafter CH edge) points toward the CH because of the strong magnetic field of the AR. The magnetic pressure gradient force on the edge near the SB (hereafter SB edge) points towards the SB initially. Because of the strong gradients near the SB, the magnetic pressure gradient force on the SB edge has a larger magnitude than the force on the CH edge so the net magnetic pressure gradient force points toward the SB. The tension force always points toward the CME center for each edge so the direction of the total tension force will always be toward the edge in the weakest background magnetic field. Initially, the net magnetic tension points towards the SB. Both of these forces cause the CME to start moving toward the SB. As the CME moves toward the SB the CH edge will interact with the potential barrier of the AR. The CH edge in the control case starts close to the AR maximum so it quickly reaches the peak in $B(\phi )$ when the CME is at a distance of 1.26 ${R_{\Sun}}$. After crossing the peak the magnetic pressure gradient force on the CH edge changes sign as the edge moves toward the PIL. This force again changes direction as the CH edge crosses the PIL, then one final change at 1.36 ${R_{\Sun}}$ as it crosses the second maxima of the AR and continues the motion toward the SB. Until 1.34 ${R_{\Sun}}$ the magnetic pressure gradient force on SB edge continues to point toward the SB. The SB edge then crosses the minimum in $B(\phi )$ at the SB so the magnetic pressure gradient force switches direction. The CME continues to move toward the SB until the SB edge pushes far enough into the SB potential well for the forces on the SB edge to overcome the forces on the CH edge. After decelerating the SB-directed motion the magnetic pressure gradient forces cause the CME to begin move away from the SB. By the time this occurs, the CME is several radii from the Sun so the force is minimal compared to the forces that initiated the deflection process. However, this process does cause the CPA to change by a little less than a degree between 5 ${R_{\Sun}}$ and 1 AU. The edge positions also affect the contribution of the magnetic tension force. The tension force does not vary substantially as a result of the CH edge’s motion through the PIL. Until the CPA reaches the SB, the CH edge remains in higher background magnetic field strength than the SB edge so the tension force always pushes the CME toward the SB. The tension force decreases quickly with both time and distance as the CME expands and moves away from the Sun toward regions of lower magnetic field strength. CMEs deflected only as a result of magnetic forces will always head toward the minima in the magnetic field. Observations have shown that CMEs do tend to head toward the HCS [@Kil09; @Gui11; @She11]. For the control case, the magnetic background possesses strong global magnetic gradients. These gradients cause the CME to reach the SB. For other Carrington Rotations with weaker global magnetic gradients this might not be the case. The model is also limited by the inclusion of only magnetic deflection forces. Other factors, not included in ForeCAT, such as interactions with other CMEs, effects of spatial variations in solar wind speed, or reconnection may still affect observed CMEs. ForeCAT does not include different plasma properties (density and temperature) for the SB compared to the rest of the solar wind background, which may also affect the SB’s interaction with a CME. Streamer blowouts should occur with a different background with weaker magnetic gradients. Parameter Sensitivity: Potential Deflection Angles ================================================== We explore ForeCAT’s sensitivity to the free parameters of the model, such as $M_{CME}$. The mass is increased to 1x10$^{16}$ g while all other free parameters are unchanged. Figure 8 shows the CME trajectory, analogous to Figure 6. The more massive CME deflects -25.3${^{\circ}}$, 1.7${^{\circ}}$ less than the original case which deflected -27.0${^{\circ}}$. At each distance, the deflection forces have comparable magnitudes to those on the control case. Again these forces initially deflect the more massive CME toward the SB. Figure 9 (analogous to Fig. 7 for the control case) shows that as with the control case, the forces change direction as the CME interacts with the AR and potential well of the SB. The more massive CME crosses the SB minimum in B($\phi$) closer to the Sun so when the forces change sign they have a larger magnitude than when this occurs for the control case. This causes the nonradial motion of the more massive CME to slow down faster than for the control case. The more massive CME also penetrates further into the SB potential well. Around 1.45 ${R_{\Sun}}$ the magnetic tension changes directions when the background magnetic field strength near the SB edge becomes stronger than that near the CH edge. Beyond 1.6 ${R_{\Sun}}$ the magnetic pressure gradient force is stronger than the control cases values. The strength and direction of these forces cause the CME to move further away from the SB, ultimately yielding a decrease in total deflection. Next, deflection angles were calculated by varying both $M_{CME}$ and $v_f$, as shown in Figure 10a, and $B_0$ and $L_0$, as shown in Figure 10b. The color indicates the change in the CPA between 1 ${R_{\Sun}}$ and 1 AU. Figure 10a shows the total deflection for varying CME mass between 10$^{14}$ and 10$^{16}$ g and varying the final propagation speed between 300 and 1000 km s$^{-1}$ while keeping all other parameters fixed. The contour plots come from a sample of 625 CMEs (25$\times$25). As seen in the individual test cases, more massive CMEs tend to deflect less. The variation for masses ranging over two orders of magnitude is only 2${^{\circ}}$ showing that the strong gradients in this magnetic background force everything to the SB. Faster CMEs tend to deflect less, but for any single mass the variance with velocity is half a degree. Slower radial velocity causes a decrease in deflection similar to the behavior observed in the test cases. The slower CME spends more time in a region with high forces causing it to reach the SB closer to the Sun. As with the high mass CME the strong forces move the CME further back toward the direction it came from, decreasing the total deflection. Figure 10b shows contours of the deflection versus the initial CME magnetic field, $B_0$, and initial CME cross-section radius, $L_0$. These parameters determine the evolution of the CME’s volume. $B_{0}$ is varied from 10 to 25 G and $L_{0}$ between 0.05 and 0.30 ${R_{\Sun}}$. Magnetic field strengths less than 10 G did not provide sufficient expansion to stop the CME from collapsing in on itself due to the large magnetic tension forces from the strong magnetic field background. These parameters lead to three distinct populations with different deflection behavior. The majority of the CMEs follow a pattern similar to that of the test cases. Stronger expansion and larger initial size tend to lead to less deflection. The second population is located in the lower left side of Fig. 10b below the band between (0.05 ${R_{\Sun}}$, 21 G) and (0.13 ${R_{\Sun}}$, 10G) corresponding to a deflection of -28${^{\circ}}$. The smallest CMEs with weakest expansion had the strongest deflection and show slightly different behavior. During propagation to 2 ${R_{\Sun}}$ these CMEs quickly deflect around 10-15${^{\circ}}$. The CH edge of the CME does not cross the angular position of the AR in this time. Between 2 ${R_{\Sun}}$ and about 60 ${R_{\Sun}}$ the CH edge of the CME moves toward the SB and crosses the first magnetic maximum (corresponding to the angular position of one polarity flux system of the AR) causing an additional 5${^{\circ}}$ of deflection. Near 60 ${R_{\Sun}}$ the CH edge passes over the final magnetic maximum (corresponding to the angular position of the other polarity flux system of the AR) leading to an additional 10-15${^{\circ}}$ of deflection by 1 AU. Of the 625 CMEs, 41 CMEs display this sort of behavior. The third population falls in the solid red region in the upper right corner of Fig. 10b. The CMEs with the strongest expansion and initially large size tend to deflect substantially less than the other two sets. 90 of the CMEs deflect between -13 and -16${^{\circ}}$. These CMEs deflect less because the CME reaches an equilibrium position with the SB edge near the SB minimum and the CH edge near the CH minimum. Due to initial size, strong expansion, or some combination thereof, the CH edge of the CME never crosses over the magnetic maximum at the angular position of the AR. The same method used to explore the influence of the CME properties can be used to analyze some of the parameters in the propagation model. The assumed CME radial velocity profile may affect the net deflection. For the control case the propagation parameters $R_{ga}$, $R_{ap}$, and $v_g$ were varied while all other parameters were unchanged. The ranges for each parameter are 1.25 ${R_{\Sun}}\le R_{ga}\le$2.25 ${R_{\Sun}}$, 2.5 ${R_{\Sun}}\le R_{ap}\le$4 ${R_{\Sun}}$, and 25 km s$^{-1} \le v_g \le$100 km s$^{-1}$. For these ranges, the final deflection angle varied by less than 0.2${^{\circ}}$. Therefore we conclude that the parameters chosen for the propagation model do not greatly influence the deflection of the CME. We explore the sensitivity of the deflection to different values of $C_d$. Using MHD simulations @Car96 show that values of $C_d$ between 1 and 3 are appropriate for the acceleration phase of a CME. @For06 use $C_d=\tanh(\beta)$ where $\beta$ equals the ratio of the thermal and magnetic pressure. Close to the Sun, $\beta <<1$ so $C_d$ will be small. We use larger values than $\tanh(\beta)$, using constant values of $C_d$ between 0.25 and 10, similar to the range of $C_d$ in @Car04. We find that these values yield deflections varying by 2${^{\circ}}$ for the control case. Stronger drag causes less deflection but ultimately the CME still deflects to the SB because of the strong magnetic gradients specific to this background. The drag changes the distance at which the CME begins interacting with the SB. With a weaker background the chosen drag coefficient may have a more significant effect. We explore as well other expressions of $C_d$ contained in the literature. @Sis06 contains two models of the drag coefficient versus distance: a linear model: $$C_d=1+\frac{5R}{1 \; AU}$$ and a quadratic model: $$C_d=\left(1+\frac{1.45R}{1 \; AU}\right)^2$$ These models produce deflections less than 0.01${^{\circ}}$ smaller than the control case with $C_d$=1. Deflection with PFSS Background =============================== We run the control case using a PFSS background to see the effects of the position of the SB varying with height. The PFSS model uses the same set of coefficients as used to initialize the MHD solution. As seen in Fig. 5b at 2.5 ${R_{\Sun}}$ the HCS forms around -5${^{\circ}}$, over 20${^{\circ}}$ away from the minimum in the scaled background. Figure 11 shows the CPA versus distance for the control case with the PFSS background (in black) and the scaled background (in blue). The PFSS control case deflects -4.4${^{\circ}}$, a deflection of -1.6${^{\circ}}$ in latitude and -4.1${^{\circ}}$ in longitude. As seen in the scaled case, the CME deflects to the minimum in the magnetic field strength but that position has changed because of the nature of the PFSS model. The forces of the PFSS model decrease much quicker with distance than the scaled model, however, with this case the magnetic minimum moves closer to the initial CME position. For cases where the magnetic minimum is further from the initial CME position the rapid decrease in forces could cause the CME to only deflect in the direction of the minimum, not fully to it. In future work we will continue to explore the sensitivity of the deflection to the magnetic background. To better understand the difference between the models requires comparisons within a background with weaker gradients or where the magnetic minimum moves away from the initial CME location. Future work will also include the effects of including the nonradial drag force. As mentioned in section 6, the deflection is sensitive to the chosen drag coefficient but complete exclusion of nonradial drag produces even more variation. Fig. 11 includes both PFSS and scaled runs without drag as dashed lines. The effect for the PFSS case is smaller since the total deflection is smaller but for the scaled model we see a difference of nearly 30${^{\circ}}$ in the cases with and without nonradial drag. Effects of Active Regions ========================= We explore here the effects that an AR can have on a CME’s deflection. We define a new deflection plane based on the orientation of the AR. Close to the Sun the magnetic pressure gradients exhibit complex behavior (Fig. 1a) and cannot be used to define the deflection plane. Deflection from an AR will result from imbalances between the different polarity flux systems of the AR. We define the overall gradient of the AR using the positions of the point within each polarity containing the strongest magnetic field. This gradient replaces the gradient vector in the deflection plane calculation. Figure 12 contains the scaled magnetic profile within the deflection plane calculated using this AR vector, analogous to Figure 5. The MHD model does not capture the full complexity of the magnetic field in an AR but it does include some variation between the opposite polarity flux systems, which is more pronounced in this plane than the original deflection plane. The system near the CH has a stronger magnetic field than the system near the SB, and the magnetic field decreases near the polarity inversion line (PIL) between the two systems. The results presented here use the control case parameters (see Section 5), but launched from $\phi$=0${^{\circ}}$ within the new deflection plane, close to the local minimum corresponding to the PIL. This CME is deflected -24.6${^{\circ}}$ during propagation out to 1 AU. The global magnetic gradients still contribute within this plane and the heightened asymmetry between the opposite polarity flux systems drives additional deflection. As a result, we determine that close to the Sun it may be necessary to redefine the deflection plane along the CMEs trajectory if we wish to accurately predict an actual CME’s deflection. Discussion ========== ForeCAT shows that magnetic forces alone can be responsible for significant CME deflection. The model excludes several other factors known to deflect CMEs, such as interaction with other CMEs, spatially varying background solar wind, or reconnection, yet still results in deflections of comparable magnitude to observed deflections. Calculating the magnetic forces along the CME trajectory relies on many fundamental assumptions about the CMEs behavior. These assumptions may prevent the current version of ForeCAT from predicting the precise behavior of actual CMEs, but allow us to demonstrate the importance of magnetic forces for CME deflection. Future works will continue to improve the model and refine some of the underlying assumptions. The description of the CME flux rope used assumes that the toroidal axis of the CME lies within the deflection plane. Any deviation from the CME being perfectly perpendicular to the deflection plane will result in an elliptical CME cross-section. The CME cross-section contributes to the tension calculation. To first order, a tilt between the toroidal CME axis and the normal to the deflection plane could be accounted for by calculating the radius of curvature as $\kappa = 1 / (L \sin(\beta))$ where $\beta$ is the angle between the deflection plane and the toroidal CME axis. The present version of ForeCAT assumes that $\beta$=90${^{\circ}}$ which will lead to an underestimate of the curvature and the magnetic tension force for an elliptical CME cross-section. CME-driven shocks would also distort the draping of the background magnetic field around the CME. Shocks are known to change the orientation of the magnetic field: fast-mode shocks rotate the magnetic field away from the shock normal, slow-mode shocks rotate the magnetic field toward the shock normal. A shock would cause the background magnetic field to rotate which would affect how the field then drapes around the CME. The draping out of the deflection plane would affect not only the direction of the magnetic tension force, but the magnitude may change as well if the magnetic field drapes around a region of the CME with different curvature. The calculation of the forces is restricted to the two points where the deflection forces should be the strongest. In reality the force should be integrated over the full surface of the CME. By assuming a solid CME body we assume as well that the motion of the CME cross-section in the deflection plane applies to the entire CME. Different cross-sections will feel different forces which should be accounted for when trying to compare to specific observations. This will be addressed in a future work. The assumption of a solid CME body also affects the ForeCAT results. Close to the Sun deflection, rotation, and nonuniform expansion cannot be distinguished. @Nie12 reconstruct the 3D trajectory of the 16 June 2010 CME, a CME with significant rotation. They find that not accounting for the CME rotation can cause substantial errors in the calculation of the CME size leading to an overestimate of the latitudinal CME expansion. Using a global MHD model, including deflection, rotation, and expansion, @Eva11 present three simulations of flux ropes with different orientations embedded in the same background solar wind. They show that the resulting shape and dynamical evolution of the CME are highly dependent on the initial CME orientation. In general, external forces likely will cause a combination of deflection, rotation and deformation. What is interpreted in coronagraph observations as rotation or nonuniform expansion could be differential deflection along the CME. We assume that the external forces cause the CME to deflect rather than to deform. The rotation and non-uniform expansion will be addressed in a future study. The magnetic background is assumed to remain in a static configuration. However reconnection can occur during the eruption and evolution of the CME. The reconnection will alter the background magnetic field, transforming magnetic energy into kinetic and thermal energy, potentially affecting the background magnetic pressure gradients. In addition, when magnetic field lines draping around the CME reconnect the tension force will become unbalanced leading to deflection similar to that of @Lug11 and @Zuc12. The CME’s radial motion is predetermined using the analytic model. A CME’s motion results, at least in part, from the same Lorentz force that drives the magnetic deflection. As the CME deflects to regions of weaker magnetic background the radial acceleration of the CME will be affected as well. Feedback between the radial and non-radial motion could lead to deflections different from those determined with an analytic radial propagation model. Conclusions =========== This manuscript presents ForeCAT, a model of CME deflection using deflection forces from both magnetic pressure gradients and magnetic tension. ForeCAT relies on many simplifying assumptions, but several test cases show that magnetic forces alone can cause deflections of similar magnitude as observed deflections. Future work will refine these assumptions and allow for comparisons between ForeCAT results and specific observed CMEs. The current ForeCAT model has already yielded several insights into CME deflection. The magnetic forces cause the CME to deflect towards the SB, the minimum in the magnetic field strength. For most CMEs the magnetic forces are sufficiently strong that the majority of the deflection occurs within the first several solar radii. The chosen magnetic background contains strong magnetic gradients so that the deflected CME reaches the SB. In a weaker background, deflection will move CMEs toward the SB, but may not be capable of fully deflecting the CME to the SB. Deflection will also change due to interactions with other CMEs, spatially varying background solar wind velocities, or reconnection, the effects of which are not included in ForeCAT. The inclusion of variations in temperature and density in the background plasma could also affect how the CME reaches its equilibrium angular position. An exploration in parameter space shows variation in the final deflection for a wide range of input parameters, for this specific background. The majority of CMEs deflect fully to the SB within a couple solar radii. Two different subsets of CMEs exhibit different behavior. Initially small CMEs with little expansion deflect the most. These CMEs do not reach the SB close to the Sun but instead partially deflect and continue along a nearly constant angular position until a secondary period of strong deflection occurs around 60 ${R_{\Sun}}$. This secondary deflection deflects the CMEs further since they interact with the SB much further from the Sun where the gradients are weaker. Initially large CMEs with strong expansion deflect the least. These CMEs remain above the AR, an equilibrium angular position is found with each edge in a potential minimum on either side of the AR. The relative strength of the magnetic minima at the SB and the CH cause a slight deflection toward the SB. The PFSS background yields different deflections than the scaled background. The scaled background contains a more realistic radial dependence assumes that the angular dependence is fixed. The PFSS background decreases too quickly with radial distance, but the angular location of the streamer belt is not fixed. These two differences between the models lead to differences in deflection, 4.4${^{\circ}}$ compared to 27${^{\circ}}$ in the PFSS and scaled backgrounds, respectively. CME deflections depend strongly on the magnetic background which will be a focus of future work. ForeCAT can be extended to uses other than just solar CMEs. Using the AR deflection plane and a more complex model of the AR magnetic field, ForeCAT should be able to capture the rolling motion of prominences. Given some approximation of the background magnetic field, ForeCAT can probe the space weather conditions of planetary systems around other stars. The magnetic fields of low mass stars can reach several kG [@Rei07], far exceeding solar values, so significant CME deflections could occur depending on the properties of the ejecta. The authors would like to thank the anonymous referee for his/her comments. C. K. is supported by NSF CAREER ATM-0747654. R. M. E. is supported through an appointment to the NASA Postdoctoral Program at GSFC, administered by Oak Ridge Associated Universities through a contract with NASA. Overpressure Expansion ====================== The equation for expansion results from a series of modifications to a classic hydrodynamics problem of a spherical over pressured region expanding in a fluid [@Mil68]. This appendix presents the derivation of the final equation, starting from the hydrodynamics. The momentum equation can be written as $$\frac{d\vec{v}}{dt} = \vec{F} - \frac{1}{\rho}\nabla P$$ where $\vec{v}$ is the velocity, $\vec{F}$ represents external forces, $\rho$ is density and P is pressure. The total velocity derivative has two contributions: the local and convective components. For incompressible fluids, the convective term becomes $\frac{1}{2}\nabla v ^2$. Rearranging gives $$\frac{\partial \vec{v}}{\partial t} = - \nabla \left( \int\frac{\partial P}{\rho}+\frac{1}{2}v^2\right)$$ where no external forces is assumed. For an irrotational fluid the velocity can be written as the negative gradient of a scalar field ($v=-\nabla\phi$) so that A2 becomes $$\nabla\left( \int\frac{\partial P}{\rho}+\frac{1}{2}v^2 - \frac{\partial \phi}{\partial t}\right)=0$$ or integrating $$\frac{P}{\rho}+\frac{1}{2}v^2 - \frac{\partial \phi}{\partial t} = C(t)$$ The formalism of a fluid source can be used to simplify equation A4. A source emits $4\pi m$ of volume per time, where $m$ is the strength of the source. Applying conservation of mass in 3D and assuming only radial velocities ($v=v_r$) gives $m=r^2v$. Plugging this into $v=-\nabla\phi$ and integrating both sides with respect to $r$ yields $m=\phi r$ or $v=\frac{\phi}{r}$. $$\frac{P}{\rho}+\frac{1}{2}\left(\frac{\phi}{r}\right)^2 - \frac{\partial \phi}{\partial t} = C(t)$$ The spherical overexpanding cavity is considered using the fluid source description. Assuming at some time the cavity has radius $R$, at the edge of the cavity ($r=R$) the change in radius is defined to be $R'$ (the same as $v$ since the velocity is only radial) which corresponds to $\phi=RR'$ giving a source strength $m=R^2R'$. The scalar field then has an $r$-dependence $$\phi = \frac{R^2R'}{r}$$ where $r$ is not just limited to the radius of the cavity. Taking the partial time derivative of A6 and rewriting A5 gives $$\frac{P}{\rho}+\frac{1}{2}\left(\frac{R^2R'}{r^2}\right)^2-\frac{R^2R''+2RR'^2}{r}=0$$ where $C$ is set equal to zero because as $r$ goes to infinity the pressure should be negligible. Looking at $r=R$ $$\frac{P}{\rho}+\frac{1}{2}(R')^2-RR''+2R'^2=0$$ or $$R''=\frac{1}{R}\left(\frac{P}{\rho}-\frac{3}{2}(R')^2\right)$$ For a cavity dominated by the magnetic pressure $P \propto B^2$. Assuming mainly poloidal magnetic field then $B$ must fall as $R^{-2}$ to conserve magnetic flux. The pressure then changes as $$\frac{P}{P_0}=\left(\frac{R_0}{R}\right)^{4}$$ and assuming a magnetic overpressure $$P=\frac{B^2_{CME}}{8 \pi}-\frac{B^2_{SW}}{8 \pi}$$ Dividing by the initial solar wind density $\rho_{SW0}$ gives $$P_0=\rho_{SW0}\frac{A^2_{h0}-A^2_{SW0}}{2}$$ where $A_{h0}$ is a hybrid Alfvén speed using the CME initial overpressure strength and the initial background solar wind density whereas $A_{SW0}$ uses the initial solar wind magnetic field. Using A9, A10, and A12 leads to a final expression $$R''=\frac{1}{R}\left(-\frac{3}{2}(R')^2 +\frac{\rho_{SW0}}{\rho}\left(\frac{R_0}{R}\right)^4\left(\frac{A^2_{h0}-A^2_{SW0}}{2}\right)\right)$$ which is the same as equation 12 with $L$ replacing $R$. Altschuler, M. D., Newkirk, G. 1969, Sol. Phys., 9, 131 Bein, B. M., Berkebile-Stoiser, S., Veronig, A. M., et al. 2011, ApJ, 738, 191 Byrne, J. P., Maloney, S. A., Refojo, J. M., Gallagher, P. T. 2010, Nature Communications, 1, 74 Cargill, P. J., Chen, J., Spicer, D. S., Zalesak, S. T. 1996, J. Geophys. Res., 101, A3 Cargill, P. J. 2004, Solar Phys., 221, 135 Chen, J. 1996, J. Geophys. Res., 101. 27499 Cheng, X., Zhang, J., Ding, M. D., Poomvises, W. 2010, ApJ, 712, 752 Cremades, H., Bothmer, V. 2004, A&A, 422, 307 Cremades, H., Bothmer, V., Tripathi, D. 2006, Adv. in Space Res., 38, 4611 Downs, C., Roussev, I. I., van der Holst, B., et al. 2010, ApJ, 712, 1219 Doyle, J. G, Teriaca, L., Banerjee, D. 1999, A&A, 349, 956 Dulk, G. A., McLean, D. J. 1978, Solar Phys., 57, 279 Evans, R. M., Opher, M., Manchester, W. B., Gombosi, T. I. 2008, ApJ, 687, 1355 Evans, R. M., Opher, M., Gombosi, T. I 2011, ApJ, 728, 41 Evans, R. M., Opher, M., Oran, R., et al. 2012, ApJ, 756, 155 Forbes, T. G., Linker, J. A., Chen, J. 2006, Space Sci. Rev., 123, 251 Gleeson, L. J., Axford, W. I. 1976, J. Geophys. Res., 81, 19 Gopalswamy, N., Kaiser, M. L. 2002, Adv. Space Res., 29, 307 Gopalswamy, N., Yashiro, S., Michalek, G., et al. 2009a, Earth Moon Planets, 104, 295 Gopalswamy, N., Mäkelä, P., Xie, H., Akiyama, S., Yashiro, S. 2009b, JGR, 114, A3 Gopalswamy, N., Mäkelä, P., Xie, H., Akiyama, S., Yashiro, S. 2010, Twelfth Int. Solar Wind Conf., 1216, 452 Gopalswamy, N., Yashiro S. 2011, ApJ Let., 736, L17 Gui, B., Shen, C., Wang, Y., et al. 2011, 271, 11 Guhathakurta, M., Fisher, R. 1998, ApJ, 499, L215 Hoeksema, J. T., Wilcox, J. M., Scherrer. P. H. 1982, J. Geophys. Res., 87, A12, 331 Isavnin, A., Vourlidas, A., Kilpua, E. K. J. 2013, Solar Phys. 284, 203 Joshi, A. D., Srivastava, N. 2011 ApJ, 739, 8 Kilpua, E. K. J., Pomoell, J., Vourlidas, A., Vainio, R., Luhmann, J., et al. 2009, Ann. Geophys., 27, 4491 Kumar, P., Cho, K.-S, Bong, S.-C., Park, S.-H., Kim, Y. H. 2012, ApJ, 746, 67 Liu, R., Liu, C., Török, T., Wang, Y., Wang, H. 2012, ApJ, 757, 150 Liu, Y., Davies, J. A., Luhmann, J. G., et al. 2010a, ApJ Lett., 710, 82 Liu, Y., Thernisien, A., Luhmann, J. G., et al. 2010b, ApJ, 722, 1762 Lugaz, N., Hernandez-Charpak, J. N., Roussev, I. I., et al. 2010, ApJ, 715, 493 Lugaz, N., Downs, C., Shibata, K., et al. 2011, ApJ, 738, I27 Lugaz, N., Farrugia, C. J., Davies, J. A., et al. 2012, ApJ, 759, 68 Luhmann, J. G., Li, Y., Arge, N., Gazis, P. R., Ulrich, R. 2002, J. Geophys. Res, 107, A8, 1154 MacQueen, R. M., Hundhausen, A. J., Conover, C. W. 1986, J. Geophys. Res., 91, 31 Mann, G., Klassen, A., Aurass, H., Classen, H. T. 2003, Astron. & Astrophys., 400, 329 McComas, D. J., Barraclough, B. L., Funsten, H. O. et al. 2000, J. Geo. Res., 105, A5, 10419 Milne-Thomson, L. M. 1968, Theoretical Hydrodynamics, The Macmillan Company, New York Mohamed, A. A., Gopalswamy, N., Yashiro, et al. 2012, J. Geophys. Res., 117, A01103 Nelson, G. J., Melrose, D. B. 1985, in Solar Radiophysics: Studies of Emission from the Sun at Metre Wavelengths, ed. D. J. McLean & N. R. Labrum (Cambridge: Cambridge Univ. Press), 333 Neugebauer, M., Forsyth, R. J., Galvin, A. B., et al. 1998, J. Geophys. Res., 103, A7, 14587 Nieves-Chinchilla, T., Colannino, R., Vourlidas, A., et al. 2012, J. Geophys. Res., 117, A06106 Panasenco, O., Martin, S., Joshi, A. D., Srivastava, N. 2011, J. Atmos. Solar Terr. Phys., 73, 1129 Pätzold, M., Bird, M. K., Volland, H., et al. 1987, Sol. Phys., 109, 91 Pneuman, G. W. 1984, Solar Phys., 94, 387 Reiners, A., Basri, G. 2006, ApJ, 656, 1121 Riley, P., Linker, J. A., Mikić, Z., et al. 2006, ApJ, 653, 1510 Rodriguez, L., Mierla, M., Zhukov, A. N., West, M., Kilpua, E. 2011, Solar Phys., 270, 561 Savani, N. P., Owens, M. J., Rouillard, A. P., Forsyth, R. J., Davies, J. A. (2010), ApJ, 714, L128 Shatten, K. H., Wilcox, J. M., Ness, N. F. 1969, 6, 442 Scherrer, P. H., Bogart, R. S., Bush, R. I., et al. 1995, Sol. Phys., 162, 129 Shen, C., Wang, Y., Gui, B., Ye, P., Wang, S. 2011, Solar Phys., 269, 389 Siscoe, G. L., Crooker, N. U., Elliot, H. A. 2006, Solar Phys., 239, 293 Sokolov, I. V., van der Holst, B., Oran, R., et al. 2013, ApJ, 764, 23 Spangler, S. R. 2005, Space Sci. Rev, 121, 189 Su, Y., Surges, V., van Ballegooijen, A., Deluca, E., Golub, L. 2011, ApJ, 734, 53 Temmer, M., Vrsnak, B., Rollet, T., et al. 2012, ApJ, 747, 57 Thernisien, A. F. R., Howard, R. A., Vourlidas 2006, ApJ, 652, 763 Thernisien, A., Vourlidas, A., Howard, R. A. 2009, Sol. Phys. 256, 111 Tóth, G., van der Holst, B., Huang, Z. 2011, ApJ, 732, 102 Tóth, G., van der Holst, B., Sokolov, I. V., et al. 2012, J. Comput. Phys., 231, 870 van der Holst, B., Manchester, W. B., IV, Franzin, R. A., et al. 2010, ApJ 725, 1373 Vourlidas, A. H., Esfandiari, E., Patsourakos, S., Yashiro, S., Michalek, G. 2010, ApJ, 722, 1522 Vrsnak, B., Maricić, D., Stanger, A., et al. 2007, Solar Phys., 241, 85 Wang, Y.-M., Sheeley, N. R. 1992, ApJ, 392, 310 Wang, Y.-M. 1993, J. Geophys. Res. Wang, A. H., Wu, S. T., Tan, A. 2003, AIP Conf. Proc. 679, 457; doi: 10.1063/1.1618633 Wang, Y., Shen, C., Wang, S., Pinzhong, Y. 2004, Solar Phys., 222, 329 Xie, H., St. Cyr, O. C., Gopalswamy, N., et al. 2009, Sol. Phys., 259, 143 Zhang, J., Dere, K. P., Howard, R. A., Kundu, M. R., White, S. M. 2001, ApJ, 559, 452 Zhang, J., Dere, K. P., Howard, R. A., Vourlidas, A. 2004, ApJ, 604, 420 Zhang, J., Dere, K. P. 2006, ApJ, 649, 1100 Zuccarello, F. P., Bemporad, A., Jacobs, C., et al. 2012, ApJ, 744, 66

ArXiv

--- abstract: 'We present a combined ion imaging and density functional theory study of the dynamics of the desorption process of rubidium and cesium atoms off the surface of helium nanodroplets upon excitation of the perturbed $6s$ and $7s$ states, respectively. Both experimental and theoretical results are well represented by the pseudodiatomic model for effective masses of the helium droplet in the desorption reaction of $m_{\mathrm{eff}}/m_{\mathrm{He}}\approx$10 (Rb) and 13 (Cs). Deviations from this model are found for Rb excited to the $6p$ state. Photoelectron spectra indicate that the dopant-droplet interaction induces relaxation into low-lying electronic states of the desorbed atoms in the course of the ejection process.' author: - 'J. von Vangerow' - 'A. Sieg' - 'F. Stienkemeier' - 'M. Mudrich' - 'A. Leal' - 'D. Mateo' - 'A. Hernando' - 'M. Barranco' - 'M. Pi' title: 'Desorption Dynamics of Heavy Alkali Metal Atoms (Rb, Cs) off the Surface of Helium Nanodroplets' --- \[sec:Intro\]Introduction ========================= Helium nanodroplets are fascinating many-body quantum systems which feature unique properties such as an extremely low internal temperature (0.38K), nanoscopic superfluidity, and the ability to efficiently cool and aggregate embedded species (dopants). Therefore, He nanodroplets are widely used as nearly ideal spectroscopic matrices for high resolution spectroscopy of isolated atoms, molecules, and clusters [@Stienkemeier:2001; @Toennies:2004; @Stienkemeier:2006; @Barranco:2006]. While most studies so far pertain to the structure and time-independent spectroscopy of doped He nanodroplets, the dynamics initiated by laser-excitation or ionization of either the dopants or the droplets themselves moves into the focus of current research. A limited number of time-resolved experiments has been carried out with pure [@Kornilov:2011; @BuenermannJCP:2012] and doped He droplets [@Droppelmann:2004; @Doeppner:2005; @Claas:2006; @Przystawik:2008; @Mudrich:2009; @PentlehnerPRL:2013; @PentlehnerPRA:2013; @Goede:2013] using femtosecond pump-probe techniques. Likewise, theoretical models of pure and doped He nanodroplets have mostly been restricted to static structure and to excitation spectrum calculations [@Barranco:2006; @Whaley:1994; @Kwon:2000; @Chin:1995; @Krotscheck:2001]. Only recently, the development of time-dependent density functional theory (TDDFT) methods applicable to microscopic superfluids [@Giacomazzi:2003; @Lehtovaara:2004] has opened the way to a time-dependent description of doped He droplets in a range of sizes comparable to those used in the experiment [@Mateo:2011; @Hernando:2012; @Mateo:2013; @Mateo:2014]. Dopants consisting of alkali (Ak) metal atoms or molecules are particularly interesting due to their weak attractive interaction with He droplets which results in their location in shallow dimple states at the droplet surface [@Dalfovo:1994; @Ancilotto:1995; @Stienkemeier:1995]. Upon electronic excitation, Ak atoms tend to desorb off the He droplet as a consequence of the repulsive interaction caused by the overlap of their extended electronic orbitals with the surrounding He [@Reho:2000; @Schulz:2001; @Callegari:2011]. The only known exceptions are Rb and Cs atoms excited to their lowest excited states [@Auboeck:2008; @Theisen:2011]. The dynamics of the desorption process of excited Ak atoms off the surface of He droplets has been recently studied in detail experimentally using the velocity-map imaging technique applied to Li, Na and Rb atoms, and theoretically using TDDFT for Li and Na [@Hernando:2012; @Fechner:2012]. The calculated He droplet response following the dopant excitation process from $ns$ to $(n+1)s$ states was found to be quite complex involving different types of density waves propagating through the droplet while the Ak dopant is ejected within a few picoseconds [@Hernando:2012]. In spite of this, the experiments show that the kinetic energy of the desorbed atom depends linearly on the excitation energy of the dopant. This conspicuous result, also reproduced by the TDDFT simulations, gives further support to the pseudodiatomic model which has already been successfully applied to interpreting the absorption spectra as well as the ion velocity distributions [@Stienkemeier:1996; @Buenermann:2007; @Loginov:2011; @Lackner:2011; @Fechner:2012]. According to this model, the dynamics of the excited AkHe$_N$ complex follows that of a dissociating diatomic molecule [@Busch:1972] where He$_N$ plays the role of one single atom in this pseudo-diatom. The part of the He droplet that effectively interacts with the Ak atom was found to have an effective mass $m_{\mathrm{eff}}\approx 15$ and $m_{\mathrm{eff}}\approx 25$ amu for Li and Na, respectively [@Hernando:2012]. In the present work, we extend previous ion imaging and TDDFT studies to the heaviest stable Ak metal atoms Rb and Cs. We again find linear dependences of the ion kinetic energies upon laser photon energy in both experiment and theory. From these we infer the effective mass of the interacting He droplet for the desorption of Rb and Cs excited to the perturbed $6s$ and $7s$ states, respectively. While most excited Ak atoms interact repulsively with a He nanodroplet as a whole, some excited states experience local attraction with one or a few He atoms. Therefore, as the excited Ak atom is expelled from the droplet surface, a bound AkHe molecule or in some cases small AkHe$_n$, $n=2,3$ complexes can form [@Reho:1997; @Droppelmann:2004; @Bruehl:2001; @Mudrich:2008; @Schulz:2001; @Giese:2012]. These so called ‘exciplexes’ are characterized by having bound vibronic states as long as the complex is electronically excited. Upon spontaneous decay into the electronic ground state the exciplex decomposes. For such excited states of the Ak atom, the desorption dynamics may be expected to deviate from that described by the simple dissociating pseudo-diatom model. In our previous experiment on Rb-doped He droplets excited into the $6p\Pi$ state, the ion kinetic energy distributions indicated that desorption of excited Rb atoms and RbHe exciplexes proceeds along the repulsive pseudodiatomic potential which correlates to the closest-lying excited $6p$ state of the free Rb atom. However, the photoelectron spectra clearly revealed that a large fraction of the desorbed Rb atoms have electronically relaxed into lower-lying levels. The photoelectron spectra contained components of the $6p$ state and of lower-lying levels ($4d$ and 5$p_{3/2}$) [@Fechner:2012]. Previously, droplet-induced relaxation of excited Rb atoms was only observed within the 5$p_{3/2,\,1/2}$ fine-structure doublet [@Bruehl:2001]. For Rb and Cs injected into bulk superfluid He fast relaxation of the lowest excited $p_{3/2}$ state into the $p_{1/2}$ and probably to the $s_{1/2}$ ground state was found to proceed within $\sim 30$ ps [@Takahashi:1993]. For Na-doped He nanodroplets, droplet-induced electronic relaxation was first observed only for higher-lying excitations with principal quantum numbers $n>6$, where the dopant-droplet interaction induces significant mixing of electronic configurations [@Loginov:2011]. In a more recent study, even for the $3d$, $5s$ and $4d$-states the authors found indications for droplet-induced decay into lower-lying levels [@Loginov:2014]. Interestingly, the presence of the relaxation channels was also visible in the speed distributions of the desorbed atoms, which contained multiple components. High-lying Rydberg states were found to completely relax into levels $n\leq 7$. Based on these observations, the authors suggested that droplet-induced relaxation proceeds via level-crossings of the pseudodiatomic potential curves which occur while the local He droplet environment of the excited Na dopant dynamically rearranges. Efficient He droplet-induced electronic relaxation was also observed for barium [@Loginov:2012] and for the transition metal atoms silver [@Loginov:2007], chromium [@Kautsch:2013] and copper [@Lindebner:2014], which are submerged in the droplet interior. Note, however, that the light Ak metals Li and Na were not found to electronically relax by droplet interactions when excited into the lowest excited $s$-states (orbital angular momentum $\ell =0$) [@Hernando:2012]. In contrast, in the present study on Rb and Cs atoms in their lowest excited $s$-states we detect *exclusively* relaxed electronic levels in the photoelectron spectra. We discuss the apparent discrepancy between the ion and electron measurements in terms of the desorption dynamics and electron energetics. Experimental ============ The experiments presented here are performed using the same setup as described previously [@Fechner:2012]. In short, a continuous beam of He nanodroplets with a mean size ranging from 200 to 17000 He atoms per droplet is generated by varying the temperature $T_0$ of a cryogenic nozzle with a diameter of $5\,\mu$m [@Toennies:2004; @Stienkemeier:2006]. An adjacent vacuum chamber contains a vapor cell filled with bulk metallic Rb or Cs heated to 85$^\circ$C and 70$^\circ$C, respectively. In the detector chamber further downstream, the He droplet beam intersects a dye laser beam (Sirah Cobra, pulse length $~10$ns, pulse energy $~10\,\mu$J, repetition rate $1\,$kHz) at right angles in the center of a velocity map imaging (VMI) spectrometer. The laser is linearly polarized along the direction of the He droplet beam, which is perpendicular to the symmetry axis of the VMI spectrometer. We record single events per image frame for which the coordinates are determined using the centroid method. Velocity-map photoelectron and photoion images are transformed into kinetic energy distributions using standard Abel inversion programs [@Vrakking:2001; @Garcia:2004]. Theoretical approach ==================== In order to model the absorption spectra as well as the dynamic response of the excited doped He droplets we describe the doped He droplets within the Density Functional Theory (DFT) framework [@Barranco:2006]. The basic ingredients of our approach are described in detail in Refs. [@Hernando:2012; @Mateo:2013]. Let us just recall that we have used the Born-Oppenheimer approximation to factorize the electronic and nuclear wavefunctions, the Franck-Condon approximation which assumes that the atomic nuclei do not change their positions or momenta during the electronic transition, and the diatomics-in-molecules approximation (pseudodiatomic model) [@Ellison:1963]. We have first obtained the structure of the Rb-droplet and Cs-droplet complexes in the ground state. Throughout this work we have used the Orsay-Trento (OT) density functional [@Dalfovo:1995] neglecting the backflow term. The Rb-He and Cs-He ground state pair potentials $V_X$ have been taken from Ref. [@Patil:1991]. Due to the large mass of Rb and Cs compared to that of He, we describe them as classical particles in the dynamics while their effect in the statics is incorporated as an external field acting upon the droplet [@Mateo:2013]. Accordingly, the energy of the system is written as $$\begin{aligned} E[\rho] &=& \int d \mathbf{r} \, \frac{\hbar^2}{2m_{\mathrm{He}}}\big|\nabla \sqrt{\rho(\mathbf{r})}\big|^2 + {\cal E}_{\mathrm{He}}[\rho(\mathbf{r})] \nonumber \\ &+& \int d \mathbf{r} \rho(\mathbf{r})V_X(|\mathbf{r}_\mathrm{Ak} -\mathbf{r}|) \; , \label{eq1}\end{aligned}$$ where ${\cal E}_{\mathrm{He}}$ is the OT potential energy density per unit volume, Ak represents either the Rb or Cs atom, and $\rho$ is the He particle density. Upon variation, one obtains the Euler-Lagrange equation that has to be solved to determine the equilibrium density $\rho_0(\mathbf{r})$ of the droplet and the location of the dopant Rb or Cs atom $\mathbf{r}_{{\rm Ak}_0}$ [@Buenermann:2007]. Schematically, $$\frac{\delta}{\delta\rho} \left( \frac{\hbar^2}{2m_{\mathrm{He}}}\big|\nabla \sqrt{\rho}\big|^2 + {\cal E}_{\mathrm{He}} \right) + V_X = \mu \; , \label{eq2}$$ where $\mu$ is the chemical potential of the He droplet that throughout this paper is made of $N=1000$ atoms. To explore other locations of the Ak atom around its equilibrium position in the surface dimple, we have minimized the energy submitted to a constraint [@Mateo:2013]. This will be useful for determining the mean kinetic energy of the ejected Ak atom as a function of the excess excitation energy. Equation (\[eq2\]) has been solved in cartesian coordinates using a spatial grid of 0.4 Å and a $200 \times 200 \times 250$ points mesh. The derivatives have been calculated with 13-point formulas. Extensive use of fast-Fourier techniques has been made to efficiently calculate the energy density and dopant-droplet interaction potentials [@Hernando:2012; @Mateo:2013]. The dynamics is triggered by the sudden substitution of the Ak-He ground state pair potential by the excited one. Within TDDFT, we represent the He droplet by a complex effective wavefunction $\Psi_{\mathrm{He}}(\mathbf{r},t)$ such that $\rho(\mathbf{r},t) = |\Psi_{\mathrm{He}}(\mathbf{r},t)|^2$. The position of the Ak atom $\mathbf{r}_{\mathrm{Ak}}(t)$ obeys Newton’s equation. For excitations involving two $s$ states, the evolution equations derived in Ref. [@Mateo:2013] adopt a simple form, namely $$\begin{aligned} i\hbar\frac{\partial}{\partial t} \Psi_{\mathrm{He}} &=& \left[ -\frac{\hbar^2}{2m_{\mathrm{He}}}\nabla^2 + \frac{\delta {\cal E}_{\mathrm{He}}}{\delta \rho(\mathbf{r})} + V_{ns}(\mathbf{r}- \mathbf{r}_{\mathrm{Ak}}) \right] \Psi_{\mathrm{He}} \nonumber \\ m_{\mathrm{Ak}} \ddot{\mathbf{r}}_{\mathrm{Ak}} &=& - \nabla_{\mathbf{r}_{\mathrm{Ak}}} \left[ \int d \mathbf{r} \rho(\mathbf{r}) V_{ns}(\mathbf{r}- \mathbf{r}_{\mathrm{Ak}}) \right] \; . \label{eq3}\end{aligned}$$ In the above equations, $V_{ns}$ with $n=6(7)$ is the $6s(7s)$ excited Rb(Cs)-He pair potential [@Pascale:1983]. The initial configuration to solve Eqs. (\[eq3\]) is the static dopant-droplet configuration, either at equilibrium or with the dopant sitting in another position around the surface dimple, $\Psi(\mathbf{r}, t=0)$= $\sqrt{\rho_0(\mathbf{r})}$, $\mathbf{r}_{\rm Ak}(t=0)= \mathbf{r}_{{\rm Ak}_0}$. The initial velocity of the Ak dopant is set to zero. Equations (\[eq3\]) have been solved using the same grid as for the static problem and a time step of 0.5 fs. We have used a predictor-corrector method [@Ralston:1960] fed by a few time steps obtained by a fourth-order Runge-Kutta algorithm. ![Sketch of the excitation and ionization scheme of Rb attached to He nanodroplets. Upon excitation of the RbHe$_N$ complex to a repulsive pseudodiatomic potential [@Callegari:2011], the Rb atom departs from the droplet surface and is ionized by a second photon from the same laser pulse.[]{data-label="fig:scheme"}](Fig1_scheme){width="80.00000%"} Photoions ========= In this work we focus on the $ns\Sigma \rightarrow (n+1)s\Sigma$ transitions of the RbHe$_N$ and CsHe$_N$ pseudo-diatoms, where $n=5,\, 6$ denotes the principal quantum number of the atomic ground states of Rb and Cs, respectively. The excitation scheme is represented in Fig. \[fig:scheme\] for Rb, where we use the pseudodiatomic potential energy curves computed by Callegari and Ancilotto [@Callegari:2011]. The ionic potential is obtained by integration of the Rb$^+$-He pair potential [@Koutselos:1990] over the He density distribution corresponding to the Rb ground state configuration, which we assume to be frozen [@Buenermann:2007]. Since the Ak-He interaction in the excited $(n+1)s\Sigma$ states is purely repulsive, the excited Ak atoms detach from the He droplets as neat atoms. Subsequent ionization by the absorption of a second photon from the same nanosecond laser pulse yields atomic ions which we detect with the VMI spectrometer. ![Simulated (a, b) and measured (c, d) photoionization spectra of He nanodroplets doped with Rb and Cs. The filled curves in (a, b) show Franck-Condon calculations based on Rb-He$_N$ and Cs-He$_N$ pseudodiatomic potentials [@Callegari:2011]; the blue lines show the Rb $6s\Sigma$ (a) and Cs $7s\Sigma$ (b) absorption profiles obtained from the present atomic-like DFT sampling method. []{data-label="fig:PIspectrum"}](Fig2_PIspectrum_rev){width="80.00000%"} Using these potentials [@Callegari:2011], we have obtained the Rb and Cs absorption spectra by calculating wave functions and Franck-Condon factors for the pseudodiatomic transitions using R. LeRoy’s program BCONT 2.2  [@bcont]. The results are depicted in Fig. \[fig:PIspectrum\]. We have also calculated the Rb $5s\Sigma \rightarrow 6s\Sigma$ and Cs $6s\Sigma \rightarrow 7s\Sigma$ absorption band contours by employing the atomic-like DFT sampling method described in Ref. [@Mateo:2011; @MateoPRB:2011], shown in that figure as blue lines. The vertical dashed lines indicate the atomic transitions. Both experimental and theoretical absorption spectra are characterized by broad bands which are blue-shifted with respect to the free atomic transitions. The blue-shift of the transitions of Rb attached to He droplets results from the fact that all excited pseudodiatomic potentials are repulsive whereas the ground state is slightly attractive (Fig. \[fig:scheme\]). The widths of the absorption contours reflect the width of the ground state wave function which is mapped onto the excited repulsive potential upon excitation. While the calculated Franck-Condon profile of the Rb $6p\Pi$ transition and the experimental spectrum is satisfactory, the Franck-Condon profile of the $6s\Sigma$ transition is slightly red-shifted with respect to the experimental contour, whereas the DFT result is blue-shifted (Fig. \[fig:PIspectrum\] a and c). The photoionization spectrum of Cs (Fig. \[fig:PIspectrum\] d) features a maximum in the range $\sim 18\,300$-$19\,000$ cm$^{-1}$ associated with the $7s\Sigma$ transition. The corresponding DFT calculation (Fig. \[fig:PIspectrum\] b) yields a peak centered at $19\,635$ cm$^{-1}$, which is again significantly blue-shifted. The DFT calculation thus overestimates the atomic shift, being unclear which part of the disagreement has to be attributed to deficiencies of the model and which to inaccuracies of the excited Ak-He pair potentials we are using [@Pascale:1983]. Note that a broad Rb$^+$ ion signal level which features a step around 20700 cm$^{-1}$ is measured around the $6s\Sigma$ feature, as previously observed in photoion [@Fechner:2012] and laser-induced fluorescence spectra [@Pifrader:2010] around the $6p\Pi$ transition. This contribution may be due to photoionization of Rb$_2$ dimers which fragment into Rb$^+$. In particular, at wave numbers below 20700 and above 21000 cm$^{-1}$ we observe significant Rb$^+_2$ signals in the tof mass spectrum which could be due to resonance-enhanced ionization via the $2^3\Sigma_u$ and via the $2^1\Sigma_u$, $2^1\Pi_u$, or $3^3\Pi_g$ states of Rb$_2$, respectively [@Lozeille:2006]. ![Top: Raw (a) and inverse Abel transformed (b) velocity-map Rb$^+$ ion images recorded when exciting Rb atoms attached to He nanodroplets on the transition $5s\Sigma\rightarrow 6s\Sigma$ at the laser wave number $\bar{\nu}=20\,800\,$cm$^{-1}$. The laser polarization direction is indicated by the vertical arrow. Bottom: Rb$^+$ ion kinetic energy distributions inferred from ion images recorded at the indicated laser wave numbers. []{data-label="fig:ionimages"}](Fig3_IonImages){width="70.00000%"} The dynamics of the laser-induced desorption process of Rb and Cs atoms is studied by recording velocity-map ion images. Fig. \[fig:ionimages\] (a) and (b) displays the raw and inverted Rb$^+$ ion images taken upon excitation to the $6s\Sigma$ state of the RbHe$_N$ complex at the laser wave number $\bar{\nu}=20\,800\,$cm$^{-1}$. The image features a circular intensity distribution with a pronounced anisotropy of the angular dependence. The intensity maxima are directed along the polarization axis of the laser (yellow arrow), as expected for the parallel $5s\Sigma\rightarrow 6s\Sigma$ transition [@Hernando:2012; @Fechner:2012]. The velocity-map ion images of Cs recorded at the $6s\Sigma\rightarrow 7s\Sigma$ transition around $\bar{\nu}=18\,700$ cm$^{-1}$ closely resemble those for Rb. In the measurements of Rb excited to the $5p\Pi$ state around $\bar{\nu}=24\,100$ cm$^{-1}$, the opposite anisotropy is observed as expected for a perpendicular $\Sigma\rightarrow\Pi$ transition in the frame of the pseudodiatomic model [@Fechner:2012]. From these images we infer the ion kinetic energy distributions (KED) by applying inverse Abel transformation and angular integration. Typical examples of such KED for excitation around the maximum of the $6s\Sigma$ band are depicted in Fig. \[fig:ionimages\]. Similarly to the previous measurements with Li and Na [@Hernando:2012], the KED consist of well-resolved maxima of widths $\sim 70$ meV which shift toward higher kinetic energies as the photon energy increases. ![(a) Total Rb$^+$ ion counts and tof peak intensity (red), (b) Rb$^+$ mean kinetic energies, (c) anisotropy parameters $\beta_2$ inferred from ion images recorded at various laser wavelengths around the maximum of the Rb $6s\Sigma$ absorption band.[]{data-label="fig:AnalysisIons"}](Fig4_AnalysisIons_rev){width="80.00000%"} Figure \[fig:AnalysisIons\] presents a compilation of the results of the analysis of the ion images recorded around the Rb $6s\Sigma$ band as a function of the laser wave number. The total Rb$^+$ ion counts \[black squares in a)\] reproduce the photoionization spectrum (red diamonds) with some systematic deviation at wave numbers $>$20850 cm$^{-1}$ of unknown origin. The mean values of the KED inferred from the images, shown in Fig. \[fig:AnalysisIons\] (b), nearly linearly increases with laser wave number. In addition, Fig. \[fig:AnalysisIons\] (c) shows the variation of the anisotropy parameter $\beta_2$ within a 4 sigma range around the KED intensity maximum inferred from the angular distributions $I(\theta )$ by fitting to the general expression for the probability distribution of one-photon transitions $ I(\theta )\propto 1+\beta_2P_2(\cos\theta )$ [@Zare:1972]. For laser wave numbers close to the maximum of the $6s\Sigma$ absorption band (20700-21100 cm$^{-1}$) we obtain beta=1.9(3). Within the experimental error this is consistent with the value $\beta_2=2$ expected for excitation of an ideal diatomic molecule at a parallel $\Sigma -\Sigma$ transition. In this case the angular distribution of dissociation products takes the form $I(\theta )\propto\cos^2\theta$. This result nicely confirms the validity of the pseudodiatomic model and the assignment of the spectral band to the parallel $5s\Sigma \rightarrow 6s\Sigma$ transition. However, in the wings of the absorption peak we find the anisotropy of the angular ion distribution to be significantly reduced. This may be due to the contribution of fragment ions from Rb$_2$ dimers which are present to a small extent in the droplet beam. Besides, it is conceivable that dynamic deformations of the local He droplet environment during the departure of the Rb atom induce perturbations of the electronic configuration of the excited Rb atom which are not accounted for in the pseudodiatomic picture. ![(Color online) Velocity (solid line, left scale) and displacement from its equilibrium location at the surface dimple (dashed line, right scale) of the desorbing Rb (a) and Cs (b) atoms as a function of time after excitation of the $6s\Sigma$ state.[]{data-label="fig:VelPosCs"}](Fig5_VelPos){width="80.00000%"} In order to obtain more detailed insight into this process, we have simulated the ejection of a Rb atom from the nominal $6s$ state and of a Cs atom from the nominal $7s$ state using TDDFT calculations. The velocities and positions as a function of time for Rb and Cs ejected from the equilibrium position at the surface dimple are shown in Fig. \[fig:VelPosCs\]. It can be seen that the Cs atom reaches an asymptotic velocity of $\sim 230$ m/s after a time evolution of $\sim 1.25$ ps. By this time, the Cs atom is $\sim 2$ Å away from its original equilibrium position at the dimple. The corresponding values for the Rb atom are $\sim 350$ m/s, $\sim 1$ ps and $\sim 2.7$ Å, respectively. In both cases, the recoil velocity of the He droplet is small, of the order of 7.5 m/s. The different evolution of positions and velocities for Rb and Cs is mainly due to the different masses of the two species. The fact that both curves are smooth and monotonously increasing with time implies that the desorption proceeds impulsively. Thus, although in the first stages of the dynamics surface vibrations and highly non-linear density waves are excited in the droplet which take a large part of the energy deposited in the system upon photo excitation, the desorption dynamics is rather insensitive to them. ![Mean kinetic energies of desorbed Rb (a) and Cs (b) atoms upon excitation to the $6s\Sigma$ and $7s\Sigma$ states, respectively. Straight lines: linear fits to the theoretical and experimental data.[]{data-label="fig:ExcessEnergies"}](Fig6_ExcessEnergies){width="80.00000%"} Detailed information about the kinematics of the desorption process can be gained from the kinetic energies of the desorbing dopants as a function of the excess excitation energy (difference between photon energy and internal energy of the free Rb or Cs atom in the $5s$ and $6s$ states) [@Hernando:2012]. The results are shown in Fig. \[fig:ExcessEnergies\] and compared to the experimental mean kinetic energies. The calculated points have been obtained by starting the dynamic simulation from different positions of the Ak obtained by a constrained minimization of the total energy of the Ak-He$_N$ complex as indicated in Sec. III. For both Rb and Cs dopants, the kinetic energy displays a linear dependence on the excess excitation energy. This dependence indicates that in spite of its apparent complexity, the ejection process is well represented by the pseudodiatomic model [@Busch:1972; @Hernando:2012]. Indeed, imposing the energy and linear momentum conservation in the instantaneous ejection of the Ak atom from the droplet one obtains $$E_{kin} =\eta (\hbar\omega-\hbar\omega_0) \; . %\equiv \eta \Delta \hbar \omega \; . \label{eq4}$$ Here, $\omega$ denotes the laser frequency and $\omega_0$ is the atomic transition frequency. Within this model, the value of the slope $\eta$ is related to the mass $m_{\mathrm{eff}}$ of the part of the He droplet that effectively interacts with the Ak atom [@Hernando:2012] by $$\eta= \frac{m_{\mathrm{eff}}}{m_{\mathrm{eff}}+m_{\mathrm{Ak}}} \Longrightarrow m_{\mathrm{eff}}=\frac{\eta}{1-\eta} \, m_{\mathrm{Ak}} \; . %\equiv \xi m_{\mathrm{Ak}}. \label{eq5}$$ By fitting the experimental and simulation data to the expression Eq. (\[eq4\]) one obtains a theoretical value $m_{\mathrm{eff}} \sim 40.7$ amu (10.2 He atoms) for Rb as compared with the experimental value of $39.6$ amu (9.9 He atoms). The corresponding values for Cs are $m_{\mathrm{eff}} \sim 52.0$ amu (13.0 He atoms) from theory and $51.6$ amu (12.9 He atoms) from experiment. The results for Rb and Cs, together with those obtained for Li and Na in Ref. [@Hernando:2012], are collected in Table \[table1\] and plotted in Fig. \[fig:meff\]. One observes an increase of $m_{\mathrm{eff}}$ with the mass of the Ak atom, as indicated by Eq. (\[eq5\]), although the prefactor $\eta/(1-\eta)$ has the opposite behavior. ![Experimental, theoretical and estimated values of the effective mass of the He droplet in the desorption process of various alkali species excited to their first excited $s$-states.[]{data-label="fig:meff"}](Fig8b_meff){width="80.00000%"} The conspicuous dependence of the effective mass of the helium droplet $m_{\mathrm{eff}}$ on the Ak dopant mass $m_{\mathrm{Ak}}$ is expected to be mainly determined by two effects. On the one hand, the geometric structure of the excited Ak-droplet system is different for each species due to slight variations of the ground state equilibrium configuration [@AncilottoZPD:1995] (radius of the surface dimple, distance of the Ak atom from the surface) as well as due to a varying mean radius of the excited Ak atom orbital $r_e$. On the other hand, the kinematics of the dissociation process induces an Ak mass-dependence, irrespective of the differing geometric initial conditions. The geometric effect is estimated by computing the geometrical overlap of the electron orbit of the excited Ak atom with the adjacent He atoms of the dimple. Based on the He dimple parameters specified in Ref. [@AncilottoZPD:1995] and on values for the mean orbital radius $\langle r_e\rangle$ we calculate the number of He atoms in the overlap volume $V_{\mathrm{eff}}$ of the excited Ak orbit and He dimple surface, $N_{\mathrm{He, eff}}=V_{\mathrm{eff}}\rho_{\mathrm{eff}}$. Here, $\rho_{\mathrm{eff}}$ is taken as half the bulk value $\rho_{\mathrm{He}}=0.0218$ Å$^{-3}$ which roughly matches the average He density within the overlap volume due to its location dimple surface where the density smoothly falls off [@AncilottoZPD:1995]. The mean orbital radius is approximated by [@Gallagher:1994; @Loginov:2014] $$\langle r_e\rangle=\frac 3 2 a_0 (n-\delta_l)^2,$$ where $a_0$ is the Bohr radius and $\delta_l$ is the quantum defect of the Ak excited state. The corresponding values of $\langle r_e\rangle$ and $N_{\mathrm{He, eff}}$ are added to Table \[table1\] and to Fig. \[fig:meff\]. The kinematic effect of the varying mass of the desorbing Ak atom is probed by solving the classical equations of motion of the Ak atom being repelled off a linear chain of effective, mutually non-interacting He atoms, each containing the mass of 7 He atoms which roughly equals the number of He atoms in the first surface layer next to the Ak dopant [@AncilottoZPD:1995]. The initial spacing between the He “layers´´ is taken as the average distance between He atoms in the droplets, $3.6$ Å [@Peterka:2007]. The distance between the Ak atom and the fist He “layer´´ is held fixed at $5.5$ Å and the same Ak-He interaction potential $V_{\mathrm{Ak-He}}(d)=0.2\exp\left(-d/3-1\right)$ (in atomic units) is used for all Ak species. The trajectories of the Ak atoms closely follow those shown in Fig. \[fig:VelPosCs\] and the trajectories of the He layers show that mostly the first He layer participates in the desorption dynamics. Accordingly, the effective mass of the He droplet (approximated by the linear chain of atoms with the mass of the He layers) only slightly exceeds the mass of the first He layer, 28 amu, see the solid line in Fig. \[fig:meff\]. While the He effective mass in this simple kinematic model matches the experimental and DFT values for Rb, the variation as a function of Ak dopant mass is not sufficiently well reproduced (Fig. \[fig:meff\]). The simple estimate based on the geometric Ak-He orbital overlap, however, shows a strong variation of the effective mass in surprisingly good agreement with the experimental values. We therefore conclude that the difference in the number of interacting He atoms for the different Ak species is likely related to the difference in the dimple structure and excited electron orbit rather than to the kinematics of the desorption process. ![Evolution of the He density distributions of the CsHe$_{1000}$ (left column) and RbHe$_{1000}$ (right column) systems after excitation to their $(n+1)s\Sigma$ states. (Multimedia view)[]{data-label="fig:DensityPlots"}](Fig7_RbCsDensities){width="60.00000%"} ![Evolution of the He density profile of the AkHe$_{1000}$ system along the symmetry axis. Three supersonic wave fronts are identified and labeled by 1 to 3. Equidensity lines corresponding to 0.5 and 0.1 times the He saturation density, 0.0218 Å$^{-3}$, representing the surface region of the droplet, are shown in white. []{data-label="fig:waves"}](Fig8_Waves){width="70.00000%"} The detailed picture of the dynamics of the He droplet upon excitation of the Ak atom is obtained from the DFT calculations. Fig. \[fig:DensityPlots\] shows the evolution of the CsHe$_{1000}$ and RbHe$_{1000}$ complexes after the Ak atom has been excited. It can be seen the dramatic changes in the droplet density caused by the excitation and subsequent ejection of the dopant. Figure \[fig:waves\] shows the evolution of the He cross-sectional density profiles of a He$_{1000}$ droplet doped with a Rb and a Cs atom for the first 5 ps. Initially, the droplet extends along the $z$ symmetry axis from about 0 to 44 Å, and the Ak atom is located in a dimple at the droplet surface (near $z=0$). Excitation of the Ak atom to the $(n+1)s$ state causes the dimple first to deepen due to the highly repulsive Ak-He interaction in the $(n+1)s\Sigma$ state. The associated compression of the He droplet lasts up to $\sim~1$ ps, as shown in the figure. Following this compression, the He surface bounces back and the dimple starts being filled. The more distant part of the droplet (near $z=42$ Å) is unperturbed and at rest, indicating that during these first ps the energy deposited in the droplet goes to its internal excitation and not to its center-of-mass motion. Figures \[fig:DensityPlots\] and \[fig:waves\] reveal that the excitation of the Cs and Rb atoms launches highly non-linear density waves into the droplet. In the case of Rb, the first perturbation front, labeled as 1, moves at $\sim 900$ m/s. This perturbation generates carrier waves with a phase velocity of $\sim 430$ m/s, modulated by supersonic envelope fronts with growing intensity. The ones with highest intensity, labeled as 2, have a group velocity of $\sim 700$ m/s. Next, a high intensity wave appears traveling at $\sim 410$ m/s (labeled as 3), which generates secondary waves propagating backwards. In the case of Cs, the velocities of the fronts are 880 m/s, 675 m/s, and 410 m/s, respectively. A similar behavior was found in Ref. [@Hernando:2012] for Na and Li atoms. ![Experimental mean kinetic energies of Rb atoms and of RbHe exciplexes ejected out of He droplets upon excitation of the $6p\Pi$ state of the RbHe$_N$ complex. The lines are linear fits to the data.[]{data-label="fig:ExcessEnergiesRb6pPi"}](Fig9_Rb6pPiFitKER){width="80.00000%"} As an extension of our previous ion imaging measurements at the Rb $6p\Pi$ band [@Fechner:2012] we analyze here the mean ion kinetic energy as a function of the excess energy. Since in the $6p\Pi$ configuration the Rb-He pair potential along the internuclear axis is attractive [@Pascale:1983], RbHe exciplexes are formed with roughly 40% relative abundance [@Fechner:2012]. Therefore, we record ion images for Rb$^+$ and RbHe$^+$ ions separately and extract the mean ion kinetic energies for each of the two species. Figure \[fig:ExcessEnergiesRb6pPi\] shows that the data points lie on a straight line that surprisingly intercepts the abscissa at a finite value of the excess energy of about $-22$ meV. Using Eq. (\[eq5\]), from the slope of the line $\eta$ one obtains $m_{\mathrm{eff}}=46.0$ (11.5 He atoms) for Rb, and $m_{\mathrm{eff}}=53.2$ amu (13.3 He atoms) for RbHe, slightly larger than the corresponding value for the $6s\Sigma$ state. The fact that the extrapolation of the $6p\Pi$ experimental data to zero kinetic energy yields a finite energy shift at zero kinetic energy, at variance with the extrapolation of the $6s\Sigma$ data, discloses an intrinsic limitation of the method used to analyze the results. The pseudodiatomic approximation, even if appropriate for the description of a direct dissociation via a purely repulsive state, does not account for other effects which are present in the dissociation kinematics of the $(n+1)p$ excitation. In the case of the $6p\Pi$ state of Rb, the dopant-He interaction contains both repulsive and attractive contributions, the latter inducing the formation of exciplexes. It is conceivable that the binding energy of the RbHe exciplex may be converted into additional translational energy upon desorption of RbHe. This interpretation has recently been invoked to rationalize the negative excess energy offset measured for NaHe exciplexes formed upon excitation of Na into the droplet-perturbed $3d$ state [@Loginov:2014]. The binding energy of RbHe in the $6p\Pi$ state amounts to about 8 meV, which does not account for the observed energy shift alone. Additional internal energy may be released into translational motion of the desorbing Rb by droplet-induced relaxation of population from the upper $6p_{3/2}$ into the lower $6p_{1/2}$ spin-orbit state of Rb. In that case, the excess energy axis would be down-shifted as represented by the horizontal top scale of Fig. \[fig:ExcessEnergiesRb6pPi\] provided the droplet effective mass is the same ($m_{\mathrm{eff}}=46.0$ amu) for this additional acceleration of the Rb atom due to spin-orbit relaxation. However, the atomic spin-orbit splitting (9.6 meV) does not fully account for the observed shift. Only the assumption that both spin-orbit and binding energy of the RbHe exciplex are fully converted into translational energy would explain the energy offset for RbHe. The nearly coinciding kinetic energies of Rb and RbHe may indicate that Rb$^+$ ions are actually produced by dissociative ionization of RbHe, the latter being the dominant product of the desorption reaction. Thus, it seems that the pseudodiatomic model no longer strictly applies when the internal degrees of freedom of the constituent atom are involved in the dynamics. Note that for the case of the desorption of sodium (Na) atoms excited to the 3$p$ state deviations from the pseudodiatomic model were also observed [@LoginovPhD:2008]. However, in contrast to the Rb case discussed here, a positive value for the abscissa intercept was found. TDDFT studies of Ak atoms ejected from the $(n+1)p$ excited states could help elucidate this open issue, but improved Ak-He pair potentials have to be previously obtained. Photoelectrons ============== Complementary information about the dynamics following laser excitation of Ak atoms attached to He nanodroplets is obtained from imaging photoelectrons. In the experiment, velocity-map photoelectron images are obtained by simply reversing the polarity of the voltages applied to the repeller and extractor electrodes [@Fechner:2012]. A typical raw and inverse Abel transformed image recorded at the laser wave number $\bar{\nu}=21\,400$ cm$^{-1}$ is depicted in the upper and lower half of Fig. \[fig:PES\] (a), respectively. The image clearly contains three separated ring structures, indicating that ionization occurs out of three Rb atomic orbitals. The faint ring structure between the two inner rings in the low half of Fig. \[fig:PES\] (a) is an artifact of the inverse Abel tranformation caused by the limited statistics. As for the ion images, we again convert the electron images into angular distributions and electron kinetic energy spectra. The latter are shown in Fig. \[fig:PES\] (b) for Rb and in (c) for Cs. ![(a) Raw (upper half) and Abel inverted photoelectron image (lower half) of Rb$^+$ ions recorded with Rb-doped He nanodroplets excited into the $6s\Sigma$ state at the laser wave number $\bar{\nu}=21\,400$ cm$^{-1}$. (b) and (c) Photoelectron spectra of Rb and Cs inferred from images recorded at the laser wave numbers 21400 cm$^{-1}$ and 18700 cm$^{-1}$, respectively. The vertical bars represent the relative populations in the respective atomic states.[]{data-label="fig:PES"}](Fig10_PES_rev){width="70.00000%"} Surprisingly, all the photoelectron spectra recorded within the Rb $6s\Sigma$ band reveal contributions of the Rb $5p_{1/2}$, $5p_{3/2}$ and $4d$ atomic levels. No electron signal associated with the $6s$ state is detected within the noise level, although the $6s$ state is the dominant atomic component of the originally excited $6s\Sigma$ state of the RbHe$_N$ complex [@Callegari:2011]. The same holds for Cs excited to the $7s\Sigma$ state. Only one peak is present in the spectrum due to ionization out of the Cs $5d$ state. The $6p$ states, which are probably populated as in the Rb case, are not detected because of insufficient photon energy for one-photon ionization of the $6p$ states. This result is at odds with the previous measurements of Li and Na excited to $(n+1)s$ states and to our previous photoelectron spectra recorded at the Rb $6p\Pi$ band where on resonance the dominant photoelectron signals came from the correlating atomic $6p$ state. The lower lying $4d$ and $5p_{3/2,\,1/2}$ states became particularly apparent for off-resonant excitation. In the present case, however, the absence of the Rb $6s$ and Cs $7s$ photoelectron signals is probably due to the particularly small photoionization cross sections of about $0.01$ Mb which result from Cooper minima close to the laser wave numbers used in the experiment [@Moskvin:1963; @Lahiri:1986]. For comparison, the detected states have photoionization cross sections $>10$ Mb [@Aymar:1984; @Lahiri:1986]. In the case of Na attached to He nanodroplets, the appearance of lower-lying atomic states was attributed to the short radiative life time of the excited level as compared to the laser pulse length [@Loginov:2011]. In the Rb and Cs cases, however, as for the Rb $6p\Pi$ state previously studied [@Fechner:2012], the lifetimes of the free Rb and Cs atoms in the $6s$ and $7s$ states ($\sim 50$ ns [@Heavens:1961; @Lahiri:1986]) by far exceed the laser pulse length (9 ns). Moreover, the appearance of the $4d$ state of Rb and of the $5d$ state of Cs cannot be explained by spontaneous radiative decay due to selection rules. Merely the Rb $5p$ photoelectron signal may contain a contribution from radiative decay. Therefore, we attribute the population of lower-lying electronic states to He droplet-induced relaxation. Whether this relaxation mechanism is predominantly non-radiative or whether the dopant-droplet interaction induces fast radiative decay even at nominally forbidden transitions cannot be determined from theses measurements. The vertical bars in Fig. \[fig:PES\] (b) and (c) depict the relative populations of the detected states as inferred from the peak integrals weighted by the photoionization cross sections. The corresponding values of the undetected Rb $6s$ and Cs $7s$-states reflect the noise level and can only be considered as upper bounds. Thus, while the populations of the Rb $6s$ and Cs $7s$ states and of the lower lying states (Rb $5s$ and Cs $6s$, $6p$) are undetermined, the Rb $5p_{3/2}$ and $4d$ states are nearly equally populated. When assuming that the Rb $5p$ level is populated purely by radiative decay, this population corresponds to a fraction of about 13% of the original $6s$ population whereas the $5p_{1/2}$ state is populated only by 7% [@Heavens:1961]. However, the fact that the measured population of the $5p_{1/2}$ state only amounts to about 8% of the $5p_{3/2}$ population indicates that an additional droplet-induced decay process is active. In the case of Cs only the $5d$ state is detected so no quantitative comparison with other states can be made. The detection of photoelectrons exclusively out of relaxed states seems to contradict the results from ion imaging which clearly demonstrate that desorption proceeds according to the pseudodiatomic model for a fixed $(n+1)s\Sigma$ electronic configuration. For Na excited into the droplet-perturbed states $5s$ and $4d$, the presence of relaxation channels was also observed in the speed distributions of the desorbed atoms [@Loginov:2014]. A broad, nearly laser wavelength independent component extending out to velocities $\sim$1500 m/s (kinetic energy $\sim$270 meV) was assigned to atoms having undergone relaxation to the lower $3d$-level. However, in the present experiments on Rb and Cs in the $(n+1)s$-state, no such broad component of the ion distributions is observed (see Fig. \[fig:ionimages\]). The range of kinetic energies observed (Fig. \[fig:ExcessEnergies\]) matches well the values expected for dissociation to proceed along the Rb and Cs $(n+1)s\Sigma$ potentials, see Fig. \[fig:scheme\]. Furthermore, we have considered the possibility that the photoelectron peaks from relaxed states could be associated with Rb$_2^+$ and Cs$_2^+$ dimer ions. However, the relative yield of dimers falls far below the proportion of photoelectrons in relaxed states. Besides, the dependence of the signal intensity of the relaxed electrons on the Rb and Cs vapor pressure in the doping cells clearly indicates that these electrons correlate to Rb$^+$ and Cs$^+$ atomic ions. In addition, the possible correlation of these electrons with large ion masses, resulting from unfragmented ion-doped He droplets, was probed by performing dedicated time-of-flight measurements using a different detection unit. The measured proportion of large cluster ions to Rb$^+$ again stayed well behind that of relaxed electrons to (undetected) electrons out of the Rb $6s$-state. However, due to the uncertainty in determining the relative detection efficiency for large ions, this possibility cannot strictly be ruled out. Thus, our observations seem to imply that electronic relaxation occurs with some time delay with respect to the strong repulsive interaction which accelerates the dopant atom away from the droplet surface. He induced electronic couplings may be facilitated by the formation of a compressed shell of He atoms around the dopant in the course of desorption ($t=$0.5-1.5 ps, see Fig. \[fig:DensityPlots\]). Theoretical modeling of the coupled electron dynamics of excited dopant-droplet complexes as well as time and mass-resolved ion and electron imaging experiments are needed to resolve this puzzling issue. ![Total photoion signal and photoelectron counts (a), relative abundances of electrons out of different atomic states (b), and anisotropy parameters $\beta_2$ (c) and $\beta_4$ (d) inferred from electron images recorded a various laser wavelengths around the maximum of the Rb $6s\Sigma$ absorption band.[]{data-label="fig:AnalysisElectrons"}](Fig11_AnalysisElectrons_rev){width="80.00000%"} For the sake of completeness we present in Fig. \[fig:AnalysisElectrons\] the results of analyzing a series of photoelectron images taken within the Rb $6s\Sigma$ band. The slight shift to higher wave numbers of the spectral feature measured by detecting electrons with respect to ions \[Fig. \[fig:AnalysisElectrons\] (a)\] likely results from contributions of ionized Rb$_2$ in the electron measurement. The relative yields of photoelectrons out of the relaxed states $4d$, $5p_{3/2}$, and $5p_{1/2}$ are depicted in Fig. \[fig:AnalysisElectrons\] (b). Similarly to our previous measurements around the Rb $6p\Pi$ state, the relative populations of the lowest detected levels increase as the laser is detuned below the droplet resonance. This change in relative populations likely reflects the variation of the monomer to dimer ratio. Changing relaxation rates into the various target electronic states of Rb due to droplet interactions may also contribute. The anisotropy parameters $\beta_2$ and $\beta_4$, which characterize the angular distribution of emitted electrons by two-photon ionization [@Reid:2003], are depicted in Fig. \[fig:AnalysisElectrons\] (c) and (d). The values of $\beta_2$ remain nearly constant within the accuracy of the measurement over the excitation spectrum. The $\beta_4$ values for the $5p$ states are roughly consistent with zero for all laser wave numbers. This indicates vanishing alignment of the electron orbitals as previously found for Rb $6p\Pi$ excitation [@Fechner:2012]. However, the $4d$ orbital appears to retain a certain degree of orbital alignment when exciting on the blue side of the Rb $6s\Sigma$ band. Likely, this is due to faster desorption when exciting further up on the repulsive branch of the Rb-He$_N$ potential. \[sec:Summary\]Summary ====================== We have studied the desorption dynamics of the heavy alkali metal atoms Rb and Cs off the surface of He nanodroplets, initiated by excitation to the perturbed $6s$ and $7s$ states, respectively. As for Li and Na adatoms [@Hernando:2012], the calculations reveal a complex response of the helium droplet to the impulsive perturbation induced by the excitation of the Rb and Cs adatoms. We find significant local deformations of the droplets and three distinct types of non-linear density waves which propagate through the droplets at different speeds. Nevertheless, both the measured and theoretically calculated mean kinetic energies of the desorbed atoms, which are in excellent agreement, can be modeled as a simple pseudodiatomic direct photodissociation reaction driven by a highly repulsive interaction. We find values of the effective mass of the He droplet interacting with Rb and Cs of about 10 and 13 He atoms, respectively. Deviations from this simple model are found experimentally for the desorption dynamics of Rb on helium droplets excited to the $6p$ state. The photoelectron spectra measured upon excitation to the perturbed $6s$ and $7s$ states evidence significant electronic relaxation of the desorbed Rb and Cs atoms into lower-lying states, at variance with analogous measurements using the light alkali species Li and Na attached to He droplets. While the ion and electron measurements appear to be contradictory, possible correlations of the observed electrons with other ion signals can largely be ruled out. This puzzling issue will be further studied by measuring photoelectron spectra with femtosecond time-resolution in pump-probe experiments. Further theoretical work in this direction is also planned. The authors gratefully acknowledge support by DGI, Spain (FEDER) under Grants No. FIS2011-28617-C02-01, by Generalitat de Catalunya (2009SGR1289), and by the Deutsche Forschungsgemeinschaft. AL has been supported by the ME (Spain) FPI program, Grant No. BES-2012-057439. Animated views (mpeg-files) of the evolution of the helium density distributions upon excitation of rubidium and cesium adatoms are available as Supporting Information. This material is available free of charge via the Internet at http://pubs.acs.org. \[1\][\#1]{} @ifundefined [71]{} Stienkemeier, F.; Vilesov, A. F. Electronic spectroscopy in [H]{}e droplets. *J. Chem. Phys.* **2001**, *115*, 10119–10137 Toennies, J. P.; Vilesov, A. F. Superfluid helium droplets: A uniquely cold nanomatrix for molecules and molecular complexes. *Angew. Chem. Int. Ed.* **2004**, *43*, 2622–2648 Stienkemeier, F.; Lehmann, K. K. Spectroscopy and dynamics in helium nanodroplets. *J. Phys. B: At. Mol. Opt. Phys.* **2006**, *39*, R127 – R166 Barranco, M.; Guardiola, R.; Hern[á]{}ndez, S.; Mayol, R.; Navarro, J.; Pi, M. Helium Nanodroplets: an Overview. *J. Low Temp. Phys.* **2006**, *142*, 1–81 Kornilov, O.; B[ü]{}nermann, O.; Haxton, D. J.; Leone, S. R.; Neumark, D. M.; Gessner, O. Femtosecond Photoelectron Imaging of Transient Electronic States and Rydberg Atom Emission from Electronically Excited He Droplets. *J. Phys. Chem. A* **2011**, *115*, 7891–7900 Bünermann, O.; Kornilov, O.; Haxton, D. J.; Leone, S. R.; Neumark, D. M.; Gessner, O. Ultrafast probing of ejection dynamics of Rydberg atoms and molecular fragments from electronically excited helium nanodroplets. *J. Chem. Phys.* **2012**, *137*, 214302 Droppelmann, G.; B[ü]{}nermann, O.; Schulz, C. P.; Stienkemeier, F. Formation Times of [R]{}b[H]{}e Exciplexes on the Surface of Superfluid versus Normal Fluid Helium Nanodroplets. *Phys. Rev. Lett.* **2004**, *93*, 023402 Döppner, T.; Fennel, T.; Diederich, T.; Tiggesbäumker, J.; Meiwes-Broer, K. H. Controlling the Coulomb Explosion of Silver Clusters by Femtosecond Dual-Pulse Laser Excitation. *Phys. Rev. Lett.* **2005**, *94*, 013401 Claas, P.; Droppelmann, G.; Schulz, C. P.; Mudrich, M.; Stienkemeier, F. [Wave packet dynamics of potassium dimers attached to helium nanodroplets]{}. *J. Phys. B* **2006**, *39*, S1151 Przystawik, A.; G[ö]{}de, S.; D[ö]{}ppner, T.; Tiggesb[ä]{}umker, J.; Meiwes-Broer, K.-H. Light induced collapse of metastable magnesium complexes formed in helium nanodroplets. *Phys. Rev. A* **2008**, *78*, 021202 Mudrich, M.; Heister, P.; Hippler, T.; Giese, C.; Dulieu, O.; Stienkemeier, F. Spectroscopy of triplet states of Rb$_2$ by femtosecond pump-probe photoionization of doped helium nanodroplets. *Phys. Rev. A* **2009**, *80*, 042512 Pentlehner, D.; Nielsen, J. H.; Slenczka, A.; Mølmer, K.; Stapelfeldt, H. Impulsive Laser Induced Alignment of Molecules Dissolved in Helium Nanodroplets. *Phys. Rev. Lett.* **2013**, *110*, 093002 Pentlehner, D.; Nielsen, J. H.; Christiansen, L.; Slenczka, A.; Stapelfeldt, H. Laser-induced adiabatic alignment of molecules dissolved in helium nanodroplets. *Phys. Rev. A* **2013**, *87*, 063401 G[ö]{}de, S.; Irsig, R.; Tiggesb[ä]{}umker, J.; Meiwes-Broer, K.-H. Time-resolved studies on the collapse of magnesium atom foam in helium nanodroplets. *New J. Phys.* **2013**, *15*, 015026 Whaley, K. B. Structure and dynamics of quantum clusters. *Int. Rev. Phys. Chem.* **1994**, *13*, 41–84 Kwon, Y.; Huang, P.; Patel, M. V.; Blume, D.; Whaley, K. B. Quantum solvation and molecular rotations in superfluid helium clusters. *J. Chem. Phys.* **2000**, *113*, 6469–6501 Chin, S. A.; Krotscheck, E. Systematics of pure and doped $^{4}\mathrm{He}$ clusters. *Phys. Rev. B* **1995**, *52*, 10405–10428 Krotscheck, E.; Zillich, R. Dynamics of $^4$He droplets. *J. Chem. Phys.* **2001**, *115*, 10161–10174 Giacomazzi, L.; Toigo, F.; Ancilotto, F. Dynamics of liquid $^4$He in confined geometries from time-dependent density functional calculations. *Phys. Rev. B* **2003**, *67*, 104501 Lehtovaara, L.; Kiljunen, T.; Eloranta, J. Efficient numerical method for simulating static and dynamic properties of superfluid helium. *J. Comput. Phys.* **2004**, *78*, 194 Mateo, D.; Jin, D.; Barranco, M.; Pi, M. Excited electron-bubble states in superfluid $^4$He: A time-dependent density functional approach. *J. Chem. Phys.* **2011**, *134*, 044507 Hernando, A.; Barranco, M.; Pi, M.; Loginov, E.; Langlet, M.; Drabbels, M. Desorption of alkali atoms from $^4$He nanodroplets. *Phys. Chem. Chem. Phys.* **2012**, *14*, 3996–4010 Mateo, D.; Hernando, A.; Barranco, M.; Loginov, E.; Drabbels, M.; Pi, M. Translational dynamics of photoexcited atoms in $^4$He nanodroplets: the case of silver. *Phys. Chem. Chem. Phys.* **2013**, *15*, 18388–18400 Mateo, D.; Leal, A.; Hernando, A.; Barranco, M.; Pi, M.; Cargnoni, F.; Mella, M.; Zhang, X.; Drabbels, M. Communication: Nucleation of quantized vortex rings in 4He nanodroplets. *J. Chem. Phys.* **2014**, *140*, 131101 Dalfovo, F. Atomic and molecular impurities in $^4$[H]{}e clusters. *Z. Phys. D* **1994**, *29*, 61–66 Ancilotto, F.; DeToffol, G.; Toigo, F. Sodium dimers on the surface of liquid $^4$[H]{}e. *Phys. Rev. B* **1995**, *52*, 16125–16129 Stienkemeier, F.; Ernst, W. E.; Higgins, J.; Scoles, G. On the use of liquid helium cluster beams for the preparation and spectroscopy of the triplet states of alkli dimers and other weakly bound complexes. *J. Chem. Phys.* **1995**, *102*, 615–617 Reho, J.; Higgins, J.; Callegari, C.; Lehmann, K. K.; Scoles, G. Alkali-helium exciplex formation on the surface of helium nanodroplets. [I]{}. [D]{}ispersed emission spectroscopy. *J. Chem. Phys.* **2000**, *113*, 9686–9693 Schulz, C. P.; Claas, P.; Stienkemeier, F. Formation of [K]{}$^*$[He]{} exciplexes on the surface of helium nanodroplets studied in real time. *Phys. Rev. Lett.* **2001**, *87*, 153401 Callegari, C.; Ancilotto, F. Perturbation Method to Calculate the Interaction Potentials and Electronic Excitation Spectra of Atoms in He Nanodroplets. *J. Phys. Chem. A* **2011**, *115*, 6789–6796 Aub[ö]{}ck, G.; Nagl, J.; Callegari, C.; Ernst, W. E. Electron Spin Pumping of Rb Atoms on He Nanodroplets via Nondestructive Optical Excitation. *Phys. Rev. Lett.* **2008**, *101*, 035301 Theisen, M.; Lackner, F.; Ernst, W. E. Rb and Cs Oligomers in Different Spin Configurations on Helium Nanodroplets. **2011**, *115*, 7005–7009 Fechner, L.; Gr[ü]{}ner, B.; Sieg, A.; Callegari, C.; Ancilotto, F.; Stienkemeier, F.; Mudrich, M. Photoionization and imaging spectroscopy of rubidium atoms attached to helium nanodroplets. *Phys. Chem. Chem. Phys.* **2012**, *14*, 3843 –– 3851 Stienkemeier, F.; Higgins, J.; Callegari, C.; Kanorsky, S. I.; Ernst, W. E.; Scoles, G. Spectroscopy of alkali atoms ([Li, Na, K]{}) attached to large helium clusters. *Z. Phys. D* **1996**, *38*, 253–263 B[ü]{}nermann, O.; Droppelmann, G.; Hernando, A.; Mayol, R.; Stienkemeier, F. Unraveling the Absorption Spectra of Alkali Metal Atoms Attached to Helium Nanodroplets. *J. Phys. Chem. A* **2007**, *111*, 12684 – 12694 Loginov, E.; Callegari, C.; Ancilotto, F.; Drabbels, M. Spectroscopy on Rydberg States of Sodium Atoms on the Surface of Helium Nanodroplets. *J. Phys. Chem. A* **2011**, *115*, 6779–6788 Lackner, F.; Krois, G.; Theisen, M.; Koch, M.; Ernst, W. E. Spectroscopy of nS[,]{} nP[,]{} and nD Rydberg series of Cs atoms on helium nanodroplets. *Phys. Chem. Chem. Phys.* **2011**, *13*, 18781–18788 Busch, G. E.; Wilson, K. R. Triatomic Photofragment Spectra. II. Angular Distributions from NO$_2$ Photodissociation. *J. Chem. Phys.* **1972**, *56*, 3638–3654 Reho, J.; Callegari, C.; Higgins, J.; Ernst, W. E.; Lehmann, K. K.; Scoles, G. Spin-orbit effects in the formation of the [Na-He]{} excimer on the surface of [H]{}e clusters. *Faraday Discussion* **1997**, *108*, 161–174 Br[ü]{}hl, F. R.; Trasca, R. A.; Ernst, W. E. Rb–He exciplex formation on helium nanodroplets. *J. Chem. Phys.* **2001**, *115*, 10220–10224 Mudrich, M.; Droppelmann, G.; Claas, P.; Schulz, C.; Stienkemeier, F. Quantum interference spectroscopy of RbHe exciplexes formed on helium nanodroplets. *Phys. Rev. Lett.* **2008**, *100*, 023401 Giese, C.; Mullins, T.; Grüner, B.; Weidemüller, M.; Stienkemeier, F.; Mudrich, M. Formation and relaxation of RbHe exciplexes on He nanodroplets studied by femtosecond pump and picosecond probe spectroscopy. *J. Chem. Phys.* **2012**, *137*, 244307 Takahashi, Y.; Sano, K.; Kinoshita, T.; Yabuzaki, T. Spectroscopy of alkali atoms and molecules in superfluid helium. *Phys. Rev. Lett.* **1993**, *71*, 1035–1038 Loginov, E.; Drabbels, M. Dynamics of Excited Sodium Atoms Attached to Helium Nanodroplets. *J. Phys. Chem. A* **2014**, *118*, 2738–2748 Loginov, E.; Drabbels, M. Spectroscopy and dynamics of barium-doped helium nanodroplets. *J. Chem. Phys.* **2012**, *136*, 154302 Loginov, E.; Drabbels, M. Excited State Dynamics of Ag Atoms in Helium Nanodroplets†. *J. Phys. Chem. A* **2007**, *111*, 7504–7515 Kautsch, A.; Koch, M.; Ernst, W. E. Electronic Relaxation after Resonant Laser Excitation of Cr in Superfluid Helium Nanodroplets. *J. Phys. Chem. A* **2013**, *117*, 9621–9625 Lindebner, F.; Kautsch, A.; Koch, M.; Ernst, W. E. Laser ionization and spectroscopy of Cu in superfluid helium nanodroplets. *Int. J. Mass Spectrom.* **2014**, *365 - 366*, 255 – 259 Vrakking, M. J. J. An iterative procedure for the inversion of two-dimensional ion/photoelectron imaging experiments. *Rev. Sci. Instr.* **2001**, *72*, 4084 Garcia, G. A.; Nahon, L.; Powis, I. Two-dimensional charged particle image inversion using a polar basis function expansion. *Rev. Sci. Instrum.* **2004**, *75*, 4989–4996 Ellison, F. O. A Method of Diatomics in Molecules. I. General Theory and Application to H$_2$O. *J. Am. Chem. Soc.* **1963**, *85*, 3540 Dalfovo, F.; Lastri, A.; Pricaupenko, L.; Stringari, S.; Treiner, J. Structural and dynamical properties of superfluid helium. *Phys. Rev. B* **1995**, *52*, 1193 Patil, S. H. Adiabatic potentials for alkali-inert gas systems in the ground state. *J. Chem. Phys.* **1991**, *94*, 8089–8095 Pascale, J. Use of *l*-dependent preudopotemtials in the study of alkali-metal-atom-[H]{}e systems. [T]{}he adiabatic molecular potentials. *Phys. Rev. A* **1983**, *28*, 632–644 Ralston, A.; Wilf, H. S. *Mathematical methods for digital computers*; John Wiley and Sons, New York, 1960 Koutselos, A. D.; Mason, E. A.; Viehland, L. A. Interaction universality and scaling laws for interaction potentials between closed‐shell atoms and ions. *J. Chem. Phys.* **1990**, *93*, 7125–7136 LeRoy, R. J.; Kraemer, G. T. [BCONT]{} 2.2. [C]{}omputer Program for Calculating Absorption Coefficients, Emission Intensities or (Golden Rule) Predissociation Rates. [T]{}he source code and manual for this program may be obtained from “Computer Programs´´ link at http://leroy.uwaterloo.ca. University of Waterloo Chemical Physics Research Report CP-650R$^2$, 2004 Mateo, D.; Hernando, A.; Barranco, M.; Mayol, R.; Pi, M. Absorption spectrum of atomic impurities in isotopic mixtures of liquid helium. *Phys. Rev. B* **2011**, *83*, 174505 Pifrader, A.; Allard, O.; Auböck, G.; Callegari, C.; Ernst, W. E.; Huber, R.; Ancilotto, F. One- and two-photon spectroscopy of highly excited states of alkali-metal atoms on helium nanodroplets. *J. Chem. Phys.* **2010**, *133*, 164502 Lozeille, J.; Fioretti, A.; Gabbanini, C.; Huang, Y.; Pechkis, H.; Wang, D.; Gould, P.; Eyler, E.; Stwalley, W.; Aymar, M.; Dulieu, O. Detection by two-photon ionization and magnetic trapping of cold Rb$_2$ triplet state molecules. *Eur. Phys. J. D* **2006**, *39*, 261 –– 269 Zare, R. N. Photoejection Dynamics. *Mol. Photochem.* **1972**, *44*, 1 Ancilotto, F.; Cheng, E.; Cole, M. W.; Toigo, F. The binding of alkali atoms to the surfaces of liquid helium and hydrogen. *Z. Phys. D* **1995**, *98*, 323–329 Gallagher, T. *Rydberg Atoms*; Cambridge University Press: Cambridge, U.K., 1994 Peterka, D. S.; Kim, J. H.; Wang, C. C.; Poisson, L.; Neumark, D. M. Photoionization Dynamics of Pure Helium Droplets. *J. Phys. Chem. A* **2007**, *111*, 7449 – 7459 Loginov, E. [P]{}hotoexcitation and [P]{}hotoionization [D]{}ynamics of [D]{}oped [L]{}iquid [H]{}elium-4 [N]{}anodroplets. Ph.D. thesis, [É]{}cole [P]{}olytechnique [F]{}[é]{}d[é]{}rale de [L]{}ausanne, 2008 Moskvin, Y. V. Photoionization of atoms and recombination of ions in the vapors of alkali metals. *Opt. Spectrosc.* **1963**, *15*, 316 – 318 Lahiri, J.; Manson, S. T. Oscillator-strength distributions for discrete and continuum transitions of excited states of cesium. *Phys. Rev. A* **1986**, *33*, 3151–3165 Aymar, M.; Robaux, O.; Wane, S. Central-field calculations of photoionisation cross sections of excited states of Rb and Sr$^+$ and analysis of photoionisation cross sections of excited alkali atoms using quantum defect theory. *J. Phys. B* **1984**, *17*, 993 – 1007 Heavens, O. Radiative Transition Probabilities of the Lower Excited States of the Alkai Metals. *J. Opt. Soc. Am.* **1961**, *51*, 1058–1061 Reid, K. L. Photoelectron angular distributions. *Annu. Rev. Phys. Chem.* **2003**, *54*, 397 – 424 Ak m$_{Ak}$ (exp) $\eta$ (exp) $\eta$ (th) m$_{\mathrm{eff}}$ (exp) m$_{\mathrm{eff}}$ (th) $\langle r_e\rangle$ (Å) $m_{\mathrm{eff}}$ (overlap) ----------- -- ---------------- -- -------------- -- ------------- -- -------------------------- -- ------------------------- -- -------------------------- -- ------------------------------ Li 6.94 0.687 0.756 15.2 21.5 5.35 17.0 Na 23.0 0.516 0.583 24.6 32.2 5.55 22.1 Rb 85.5 0.327 0.324 41.9 41.0 6.54 41.5 Cs 132.9 0.281 0.273 51.8 50.5 6.78 53.8 \[0.5ex\] : \[table1\] Characteristics of the experimental and theoretical kinetic energy distributions of the desorbed alkali atoms, see text for details. All masses are given in amu.

ArXiv

--- abstract: | We present the results of a search for standard model Higgs boson production with decay to $WW^{*}$, identified through the leptonic final states $e^{+}e^{-}\bar{\nu} \nu ,~e^{\pm}\mu^{\mp} \bar{\nu} \nu$ and $\mu^{+}\mu^{-} \bar{\nu} \nu$. This search uses 360 $\rm{pb}^{-1}$ of data collected from $p\bar{p}$ collisions at $\sqrt{s}$ = 1.96 TeV by the upgraded Collider Detector at Fermilab (CDF II). We observe no signal excess and set 95% confidence level upper limits on the production cross section times branching ratio for the Higgs boson to $WW^{*}$ or any new scalar particle with similar decay products. These upper limits range from 5.5 to 3.2 $\rm{pb}$ for Higgs boson masses between 120 and 200 GeV/$c^2$. author: - The CDF Collaboration title: ' Search for a Neutral Higgs Boson Decaying to a $W$ Boson Pair in $p\bar{p}$ Collisions at $\sqrt{s} = 1.96$ TeV ' --- The Higgs mechanism is a leading candidate for electroweak symmetry breaking and consequently for mass generation of the $W$ and $Z$ bosons without violation of local gauge invariance. A manifestation of this mechanism is the existence of a neutral scalar particle, the Higgs boson [@higgs], which has not been observed to date. Its mass is a free parameter in the standard model (SM), but its couplings to other particles of known mass are fully specified at tree level. Direct searches at the CERN $e^{+}e^{-}$ collider (LEP) yielded a lower limit for the Higgs boson mass of $m_{H}>$ 114.4 GeV/$c^{2}$ at 95$\%$ confidence level (C.L.) [@lep2lim]. Precision electroweak measurements indirectly predict a Higgs boson mass of $91^{+45}_{-32}$ GeV/$c^{2}$ [@ewprec]. At the Tevatron, the dominant production mechanism for the SM Higgs boson is gluon-gluon fusion through heavy quark loops. Branching fractions for the various decay channels of the Higgs boson depend on its mass. For masses below about $135\,{\rm GeV}/c^2$ the dominant decay is $H \rightarrow b\bar{b}$, while heavier Higgs bosons decay predominantly to $WW^\ast$ [@xsecnlo], where $W^{*}$ indicates a $W$ boson that can be off mass-shell. For the $b \bar{b}$ decay mode, the requirement of associated production of the Higgs with vector bosons ($p\bar{p}\rightarrow WH/ZH$) can greatly improve the signal purity [@run1_VH]. For the $WW^{*}$ decay mode, the leptonic decays of $W$ bosons give a clean enough signature that the inclusive single Higgs production process gives the best search sensitivity. The next-to-leading order (NLO) production cross section [@xsecnlo] times branching ratio for a SM Higgs boson, $\sigma(p\bar{p}\rightarrow H)\times$BR($H\rightarrow WW^{*})$, ranges from 0.036 to 0.25 $\rm{pb}$ for Higgs masses of 110-200 GeV/$c^2$. This Letter presents the results of a direct search for a Higgs boson in the channel $gg \rightarrow H \rightarrow WW^{*}\rightarrow \ell^{+}\nu\ell^{-}\bar{\nu}$ ($\ell=e, \mu, \tau$), identified by the “dilepton” final states $e^{+}e^{-}, e^{\pm}\mu^{\mp}$ or $\mu^{+}\mu^{-}$. We also include the efficiency for leptonically decaying taus to e or $\mu$. This is the first search in this channel by the CDF Collaboration. A similar search in this channel was recently performed by the DØ Collaboration [@hwwd0]. The data sample used for this analysis was collected with the CDF II detector at the Fermilab Tevatron between 2002 and 2004, and corresponds to an integrated luminosity of approximately 360 $\rm{pb}^{-1}$ [@lumdetail]. For this integrated luminosity, the cross section limits we are able to place on Higgs production are a factor of approximately 10-50 larger than the SM expectation, based on the NLO calculation. However, the production cross-section can be enhanced in extensions to the SM due to new particles e.g., a fourth generation fermion family [@4gen], contributing at higher order to the gluon-gluon fusion Higgs production process. CDF II is a detector with approximate azimuthal and forward-backward symmetry and it is fully described elsewhere [@cdfdet]. It consists of a charged-particle tracking system in a 1.4 T magnetic field and segmented electromagnetic and hadronic calorimeters surrounded by muon detectors. The electromagnetic and hadronic sampling calorimeters surrounding the solenoid are used to measure the energy of interacting particles in the pseudo-rapidity range $|\eta|<3.6$ [@eta]. The calorimeters are divided into projective geometry towers. This analysis uses both central ($|\eta|<1.1$) and end-plug detectors ($1.2<|\eta|<2.0$) to identify electron candidates. A set of drift chambers located outside the central hadron calorimeters and another set behind a 60 cm iron shield help detect muons in the region $|\eta|<$ 0.6. Additional drift chambers and scintillation counters detect muons in the region 0.6 $\leq|\eta|\leq 1.0$. Events used for this analysis are collected using the following triggers [@trigele; @trigplug]: an inclusive central electron ($|\eta|<1.1$) trigger requiring an electron with $E_{T}>$ 18 GeV, an inclusive central muon ($|\eta|<1.0$) trigger requiring a muon with $p_{T}>$ 18 GeV/$c$, or a trigger for events with a forward electron (1.2 $\leq|\eta|\leq 2.0$) with $E_{T}>$ 20 GeV and missing transverse energy, $~\met>$ 15 GeV [@met]. After the event reconstruction, event selection criteria which retain high $H\rightarrow WW^{*}$ signal efficiency while minimizing the effect of background contamination are applied. Some selection requirements are mass dependent, as the event kinematics and topology change as functions of $m_{H}$. The selection requires two oppositely charged lepton candidates consistent with originating from the same vertex, with $p_{T}>$ 20 GeV/$c$ for the trigger lepton and $p_{T}>$ 10 GeV/$c$ for the second one. The leptons are also required to be isolated in both the calorimeter and the tracking chamber [@caliso], and the dilepton invariant mass $m_{\ell\ell}$ is required to be greater than $16~{\rm{GeV}}/c^2$, in order to remove events from the $c\bar{c}$/$b\bar{b}$ resonances. After removal of events identified as cosmic rays or electrons from photon conversions [@trigele], we count the jets [@jetdef] with $E_{T}>$ 15 GeV and $|\eta|<2.5$. Signal events do not typically have high-$E_T$ jets in the final state, but can occasionally have lower-$E_T$ jets from initial state gluon radiation. On the other hand, $t\bar{t}$ pairs decay primarily to $W^{+}W^{-}b\bar{b}$ and thus tend to have at least two jets in the final state. This background is reduced by selecting only events satisfying one of the following criteria: no jets with $E_{T}>15$ GeV, or only one jet with $15<E_{T}<55$ GeV, or 2 jets, each with $15<E_{T}< 40$ GeV. Events with more than two jets with $E_{T}>15$ GeV are also rejected. After the selection criteria described above, the dominant surviving background is Drell-Yan production of $\ell^{+}\ell^{-}$ pairs, which is suppressed by requiring that $\met>m_{H}/4$. The events with missing energy due to a mis-measurement of the jet energy, or $Z\rightarrow\tau\tau$ events with missing energy arising from a leptonic tau decay, are removed by requiring the azimuthal angle between the $\met$ and the closest jet or lepton to be at least $20^\circ$, if $\met<$ 50 GeV. To further reduce the large $Z/\gamma^{*}$ background, the dilepton invariant mass is required to be $m_{\ell\ell} < m_{H}/2 - 5$ GeV/$c^2$. Finally, the scalar sum of the $p_{T}$ of the two leptons and the $\met$ is required to be below the Higgs mass. The kinematic cuts described above exploit the correlations in the $W$ pairs produced by the decay of a Higgs boson and suppress SM $WW$ production. These correlations are due to angular momentum conservation in the decay of a spin-zero Higgs boson. Since $W$ bosons decay into left-handed leptons and right-handed anti-leptons, and since the W bosons in the decay $H\rightarrow WW^{*}$ have opposite helicities, the final state lepton pairs and also the neutrino pairs tend to be azimuthally aligned in Higgs decay. This implies that the signal events tend to have smaller $m_{\ell\ell}$ and azimuthal angle between leptons ($\Delta \phi$) and larger $\met$, as compared with production of SM $WW$ pairs. These differences are further exploited in the final stages of the analysis, when the $\Delta \phi$ distribution of the data is compared with the background and signal predictions. The acceptance for identifying $H\rightarrow WW^{*}\rightarrow \ell\nu\ell\nu$ events with the above selection criteria is calculated as a function of the Higgs boson mass using [PYTHIA]{} [@pythia] Monte Carlo, after a [GEANT]{}-based [@geant] simulation of the CDF detector response. The total acceptance is a product of the geometric and kinematic acceptance, the lepton identification efficiencies, the trigger efficiencies, and the topological cut efficiencies. It does not include the branching fraction of $W$ leptonic decays. The total acceptance ranges from 3.0$\%$ to 6.5$\%$, depending on the Higgs mass, and is summarized in Table \[table:hwwacc\]. Approximately 25% of the expected signal are $ee$ events, 25% $\mu\mu$, and 50% $e\mu$. $m_{H}$ (GeV/$c^2$) 120 140 160 180 200 --------------------------------- ------ ------ ------ ------ ------ BR($H\rightarrow$ $WW^{*})(\%)$ 13 48 90 94 74 Total acceptance($\%$) 3.15 4.56 6.47 6.41 5.54 : \[table:hwwacc\] The branching ratio BR($H\rightarrow$ $WW^{*}$) and the total acceptance of the signal after all the selection criteria. The total acceptance is calculated with respect to the number of $p\bar{p} \rightarrow H \rightarrow WW^{*}\rightarrow \ell^{+}\nu\ell^{-}\bar{\nu}$ events. The systematic uncertainty on the acceptance is 6$\%$ resulting from uncertainties in the modeling of the initial state radiation by [PYTHIA]{} (3$\%$), and uncertainties on the gluon parton distribution functions (4$\%$) [@pdferror], jet energy scale (1$\%$), track isolation ($<$2$\%$), electron and muon trigger efficiencies($<$1$\%$), and electron and muon identification efficiencies (2$\%$). In addition, a 6% uncertainty on the integrated luminosity is applied to the expected number of events for all processes [@lum]. $m_{H}$ (GeV/$c^{2}$) 120 140 160 180 200 ------------------------- ------------------- ------------------ ------------------ ------------------ ------------------ $WW$ 5.49 $\pm$ 0.66 7.98 $\pm$ 0.96 9.79 $\pm$ 1.18 9.89 $\pm$ 1.19 9.19 $\pm$ 1.11 $Z/\gamma^{*}$ 1.63 $\pm$ 0.42 1.01 $\pm$ 0.26 0.76 $\pm$ 0.20 0.83 $\pm$ 0.21 0.96 $\pm$ 0.25 $W+\rm{jets}$/$\gamma$ 4.57 $\pm$ 0.90 3.49 $\pm$ 0.81 2.48 $\pm$ 0.69 1.70 $\pm$ 0.46 1.20 $\pm$ 0.37 $WZ+ZZ$ 0.25 $\pm$ 0.03 0.37 $\pm$ 0.05 0.40 $\pm$ 0.05 0.49 $\pm$ 0.07 1.16 $\pm$ 0.15 $t\bar{t}$ 0.12 $\pm$ 0.01 0.21 $\pm$ 0.02 0.35 $\pm$ 0.04 0.46 $\pm$ 0.05 0.58 $\pm$ 0.06 Total Background 12.06 $\pm$ 1.19 13.08 $\pm$ 1.28 13.78 $\pm$ 1.38 13.37 $\pm$ 1.30 13.09 $\pm$ 1.21 $H\rightarrow$ $WW^{*}$ 0.090 $\pm$ 0.008 0.32 $\pm$ 0.03 0.58 $\pm$ 0.05 0.41 $\pm$ 0.03 0.20 $\pm$ 0.02 Data 7 14 16 19 17 After all selection requirements, the background events come predominantly from $WW$ pair production (about 70% of the total for $m_H = 160$ GeV/$c^2$) [@WW], $Z/\gamma^{*}$, $W + {\rm jets}$, and $W+\gamma$. Smaller backgrounds include $WZ$, $ZZ$, and $t\bar{t}$ production. A summary of these contributions as a function of Higgs mass is given in Table \[table:grdtable\]. The diboson ($WW, WZ, ZZ$), $Z/\gamma^{*}$ and $t\bar{t}$ backgrounds are determined using [PYTHIA]{} Monte Carlo, followed by the CDF II detector simulation. We normalize the total number of events for these processes to recent theoretical cross-sections [@wwxsec; @ttbarxsec]. To estimate the $W+\gamma$ background we use a matrix element generator [@baur1] and use [PYTHIA]{} for the initial state QCD radiation and hadronization. The background from $W + {\rm jets}$, where a jet or track is misidentified as a lepton (electron or muon), is determined from the data and called the “fake background.” We first determine the probability that a jet with a large fraction of its energy deposited in the electromagnetic calorimeter is misidentified as an electron, and the probability that a minimum ionizing track is misidentified as a muon. These probabilities are termed fake rates. The fake rate for each lepton type is calculated using an average of four inclusive jet samples (triggered with at least one jet with $E_{T} > $ 20, 50, 70 or 100 GeV). We subtract the contribution from sources of real leptons ($W$ and $Z$ decays) and parametrize the fake rates as a function of jet transverse energy (for electrons) or track transverse momentum (for muons). The background is determined by weighting the jets from a data sample of $(W \rightarrow \ell\nu) + \rm{jets}$ events by the fake rates. For data events passing the previously described selection criteria, we search for an excess of events with small azimuthal angle between the leptons, $\Delta \phi$. A binned likelihood is used to compare the azimuthal angle distribution in the data with a combination of expected distributions from the SM background processes. Figure \[fig:hww\_fig1\] shows the $\Delta \phi$ distributions for SM backgrounds, for Higgs masses of 140 and 160 GeV/$c^2$, and for the data. We observe no evidence for a signal over the SM expectations. We calculate upper limits on the production cross section times branching ratio, $\sigma_{H} \times {\rm{BR}}(H \rightarrow WW^\ast$), using a Bayesian procedure. We consider three components in the data: $H \rightarrow WW^\ast$, SM $WW$, and other SM processes ($WZ$, $ZZ$, Drell Yan, $W+\rm{jets}$/$\gamma$) labeled as “other”. The expected number of events in each $\Delta \phi$ bin is $$\begin{aligned} \mu = f_{WW}\cdot n_{WW} + f_{other}\cdot n_{other} + \nonumber \\ f_{HWW} \cdot (\epsilon \cdot {\cal{L}} \cdot \sigma_{H} \cdot BR(H\rightarrow WW^{*})),\nonumber \label{eq1}\end{aligned}$$ where $f_{WW}$, $f_{other}$ and $f_{HWW}$ represent the expected fraction of the specified categories of events falling in each $\Delta \phi$ bin, $n_{WW}$ and $ n_{other}$ are the expected numbers of $WW$ and non-$WW$ background events, and $\epsilon$, $\cal L$ and $\sigma_{H}$ correspond to efficiency, integrated luminosity, and $H$ production cross section. A posterior density is obtained by multiplying the Poisson likelihood function with Gaussian prior densities for the integrated luminosity, background normalizations, and the signal efficiency: {height="7.6cm" width="0.9\columnwidth"} {height="7.6cm" width="0.9\columnwidth"} $$\begin{aligned} L = \prod_{i=1}^{N_{bins}}{\frac{\mu_{i}^{n_{i}} \cdot e^{-\mu_{i}}}{n_{i}!} \times G(n_{WW},\sigma_{WW}) \times G(n_{other},\sigma_{other})} \nonumber \\ \times G(\epsilon,\sigma_{\epsilon}) \times G(\cal L,\sigma_{\cal L})\nonumber \label{eq2}\end{aligned}$$ where $n_{i}$ is the number of events observed in the data, and $G(n,\sigma_n)$ are Gaussian constraints for parameter $n$ with uncertainty $\sigma_{n}$. The prior density for $\sigma \times$ BR($H \rightarrow WW^\ast$) is assumed uniform. The posterior density is then integrated over all parameters except for $\sigma \times {\rm{BR}}(H\rightarrow WW^\ast$), for which a 95% confidence level upper limit is obtained by calculating the $95^{\rm th}$ percentile of the resulting distribution. ![\[fig:vh\_fig3\] Summary of the Run II CDF $95\%$ confidence level upper limits on $\sigma(p\bar{p} \rightarrow H)\times$BR($H\rightarrow WW^{*}$). Shown for comparison are the standard model prediction, the 4th generation model prediction [@4gen] and the region excluded by the LEP experiments. The prediction for the 4th generation model assumes that 4th family fermions have a mass $m_{4}$ = 200 GeV/$c^{2}$.](plot_hwwlimits_final.eps){height="7.5cm" width="0.95\columnwidth"} The expected and observed upper limits on the cross section times branching ratio, $\sigma \times {\rm{BR}}(H\rightarrow WW^\ast$), for different Higgs masses are shown in Table \[table:limits\]. The expected limits are calculated using 1000 simulated experiments, assuming no signal, for each Higgs mass. The median value of the limits obtained from these experiments is chosen as the [*a priori*]{} upper limit. $m_{H}$ (GeV/$c^2$) 120 140 160 180 200 ----------------------------- ----- ----- ----- ----- ----- Expected Limits ($\rm{pb}$) 7.1 4.8 3.5 3.4 4.0 Observed Limits ($\rm{pb}$) 4.5 4.6 3.2 4.3 5.2 : \[table:limits\] The expected and observed 95$\%$ C.L. limits on $\sigma(p\bar{p}\rightarrow H)\times {\rm{BR}}(H\rightarrow WW^{*})$. In conclusion, observing no signal in the direct search for $H \rightarrow WW^\ast$, with the subsequent decay of the $W$ bosons to leptons, we have set mass dependent limits at 95% C.L. on $\sigma(p\bar{p}\rightarrow H)\times{\rm{BR}}(H\rightarrow WW^{*})$. This search is potentially sensitive to other new physics models such as the example in Figure \[fig:vh\_fig3\]. We thank the Fermilab staff and the technical staffs of the participating institutions for their vital contributions. This work was supported by the U.S. Department of Energy and National Science Foundation; the Italian Istituto Nazionale di Fisica Nucleare; the Ministry of Education, Culture, Sports, Science and Technology of Japan; the Natural Sciences and Engineering Research Council of Canada; the National Science Council of the Republic of China; the Swiss National Science Foundation; the A.P. Sloan Foundation; the Bundesministerium für Bildung und Forschung, Germany; the Korean Science and Engineering Foundation and the Korean Research Foundation; the Particle Physics and Astronomy Research Council and the Royal Society, UK; the Russian Foundation for Basic Research; the Comisión Interministerial de Ciencia y Tecnología, Spain; in part by the European Community’s Human Potential Programme under contract HPRN-CT-2002-00292; and the Academy of Finland. [99]{} P. W. Higgs, Phys. Lett. [**12**]{}, 132 (1964). R. Barate [*et al.*]{} (ALEPH Collaboration), Phys. Lett. B [**565**]{}, 61 (2003). J. Alcaraz [*et al.*]{}, hep-ex/0511027 (2005). A. Djouadi, J. Kalinowski and M. Spira, Comput. Phys. Commun. [**108**]{}, 56 (1998); S. Catani, D. de Florian, M. Grazzini, P. Nason, hep-ph/0306211. D. Acosta [*et al.*]{} (CDF Collaboration), Phys. Rev. Lett. 95, 051801 (2005); A. Abulencia [*et al.*]{}, Phys. Rev. Lett. 96, 081803 (2006). V. M. Abrazov [*et al.*]{} (DØ Collaboration), Phys. Rev. Lett. 96, 011801 (2006). The integrated luminosities for $ee$, $e\mu$ and $\mu\mu$ are $363 \pm 22$, $356 \pm 21$, and $364 \pm 22~{\rm{pb}^{-1}}$ respectively. The differences for the various channels are due to different detector quality requirements. E. Arik [*et al.*]{}, Phys. Rev. D[**66**]{}, 033003 (2002); E. Arik [*et al.*]{}, hep-ph/0502050. D. Acosta [*et al.*]{} (CDF Collaboration), Phys. Rev. D[**71**]{}, 032001 (2005). In the CDF coordinate system, $\theta$ and $\phi$ are the polar and azimuthal angles, respectively, with respect to the proton beam direction ($z$ axis). The pseudorapidity $\eta$ is defined as $-\ln \tan(\theta/2)$. The transverse momentum of a particle is $p_{T} = p \sin\theta$. The analogous quantity using calorimeter energies, defined as $E_{T} = E \sin\theta$, is called transverse energy. D. Acosta [*et al.*]{} (CDF Collaboration), hep-ex/0508029, submitted to Phys. Rev. D. D. Acosta [*et al.*]{} (CDF Collaboration), PRD-RC 71, 051104 (2005). The missing transverse energy, $\met$, is defined as $|-\sum \Et^i \,\hat{n}_i|$, where $\hat{n}_i$ is the unit vector in the transverse plane pointing from the interaction point to the energy deposition in calorimeter cell $i$. The calorimeter isolation $I_{cal}$ is defined as the extra energy deposited in the calorimeter cone of radius $\Delta R=\sqrt{{\Delta \phi}^{2}+ {\Delta \eta}^{2}}= 0.4$ around the lepton cluster. The $I_{cal}$ is required to be less than 10 $\%$ of the lepton $E_{T}$. The track isolation $I_{trk}$ is defined as the sum $p_T$ of all the tracks in a cone of radius $\Delta R = 0.4$ around the lepton candidate track, but excluding it. The $I_{trk}$ is required to be less than 10$\%$ of the lepton track $p_{T}$. A jet is defined as a cluster of calorimeter towers above an energy threshold of 3 GeV, within fixed radius $\Delta R= 0.4$. The jet $E_{T}$ is corrected for the calorimeter response and multiple interactions. See: G. C. Blazer and B. L. Flaugher, Ann. Rev. of Nucl. and Part. Sci. Vol 49 (1999). T. Sjöstrand [*et al.*]{}, Comput. Phys. Commun. [**135**]{}, 238 (2001). We use [PYTHIA]{} V6.2. R. Brun and F. Carminati, CERN Program Library Long Writeup W5013, 1993 (unpublished). We compared the acceptances obtained using MRST 72 and 75 PDF sets, corresponding to $\alpha_{s}$ = 0.1175 and 0.1125. Also we include the effect of the eigenvector variations around the fit minima, using CTEQ6M PDF sets. S. Klimenko, J. Konigsberg, and T. Liss, FERMILAB-FN-0741 (2003); D. Acosta [*et al.*]{}, Nucl.Instrum. Methods [**A 494**]{}, 57 (2002). D. Acosta [*et al.*]{} (CDF Collaboration), Phys. Rev. Lett. 94, 211801 (2005). J. M. Campbell, R. K. Ellis, Phys. Rev. D [**60**]{}, 113006 (1999). M. Cacciari [*et al.*]{}, hep-ph/0303085; N. Kidonakis and R. Vogt, Phys. Rev. D [**68**]{}, 114014 (2003). U. Baur and E. L. Berger, Phys. Rev. D[**47**]{}, 4889 (1993).

ArXiv

--- abstract: 'Radiative transfer effects due to overlapping X-ray lines in a high-temperature, optically-thick, highly-ionized medium are investigated. One particular example, where the [Ly$\alpha$]{} doublet ($2~^2P_{1/2;3/2}$ – $1~^2S_{1/2}$) coincide in frequency with the [Ly$\zeta$]{} lines ($7~^2P_{1/2;3/2}$ – $1~^2S_{1/2}$) is studied in detail to illustrate the effects on the properties of the emergent line spectrum. We solve the radiative transfer equation to study the energy transport of resonance line radiation in a static, infinite, plane-parallel geometry, which is used to compute the destruction/escape probabilities for each of the lines for various total optical thicknesses of the medium, as well as destruction probabilities by sources of underlying photoelectric opacity. It is found that a large fraction of the  [Ly$\alpha$]{} line radiation can be destroyed by , which can result in an reversal of the  [Ly$\alpha$]{}/ [Ly$\alpha$]{} line intensity ratio similar to what may be seen under non-solar abundances. Photoelectric absorption by ionized carbon and nitrogen can also subsequently increase the emission line intensities of these ions. We show that line ratios, which are directly proportional to the abundance ratios in optically thin plasmas, are not good indicators of the true [[CNO]{}]{} abundances. Conversely, global spectral modeling that assumes optically thin conditions may yield incorrect abundance estimates when compared to observations, especially if the optical depth is large. Other potentially important overlapping lines and continua in the X-ray band are also identified and their possible relevance to recent high resolution spectroscopic observations with [[*Chandra*]{}]{} and [[*XMM-Newton*]{}]{} are briefly discussed.' author: - Masao Sako title: 'CNO Line Radiation and the X-ray Bowen Fluorescence Mechanism in Optically-Thick, Highly-Ionized Media' --- Introduction ============ Radiative transfer effects are capable of producing line spectra that differ substantially from those emitted under optically thin conditions. As resonance line photons travel through a medium that is optically thick to its own radiation, multiple non-coherent scattering substantially alters the line profiles of the emergent spectrum. Line transfer also affects the emergent line intensities through line or continuum absorption, in which case the energy is either (1) re-radiated through other discrete transitions or (2) is used in heating the ambient gas. In hydrogen-like ions, for example, the upper levels of the higher-series Lyman transitions ($\beta$, $\gamma$, $\delta$, etc.) have finite probabilities of decaying to excited states. This implies that the Lyman line photons will be destroyed and re-radiated in the Balmer, Paschen, etc. series if the line optical depths are sufficiently large. Ratios of Lyman to Balmer line intensities, for example, can then be used to estimate the optical depth, as well as other local physical parameters, such as the density and temperature (see, e.g., @drake80, and references therein). Another well-known example of such an effect is the Bowen fluorescence mechanism [@bowen34; @bowen35], where it was realized that the coincidence in the wavelengths of (He$^{+}$) [Ly$\alpha$]{} and (O$^{2+}$) $2p$ – $3d$ transitions allows the conversion of [Ly$\alpha$]{} line photons into the upper levels of , which then decay to produce lines in the optical/[[UV]{}]{} band known as the Bowen lines at $\lambda \approx 3000 \sim 4000$ Å. These line were commonly found to be extraordinarily strong in the optical spectra of planetary nebulae and Seyfert galaxies. @unno55 presented the first complete analysis of the [Ly$\alpha$]{} transfer and the conversion efficiencies of the Bowen lines, and demonstrated that the observed spectral properties can be reasonably well-described by this mechanism. Subsequently, @weymann69 presented a more detailed calculation using an exact form of the redistribution function and a realistic ionization balance calculation (see, also, @harrington72 [@kallman80; @eastman85; @netzer85]). In the simplest approximation generally adopted for spectral modeling, one assumes that line photons created in the gas escape the medium without any further interaction with the constituent atoms in the medium. This is generally referred to as the optically-thin limit or the coronal approximation in the context of collisionally ionized plasma. In this limit, the ionization balance is determined solely by the local temperature and is assumed to be entirely decoupled from the radiation field. The emission spectra computed under these assumptions seem to apply fairly well to the observed X-ray spectra of a wide variety of sources including stellar coronae, supernova remnants, hot diffuse gas in starburst galaxies, and the intergalactic medium in clusters of galaxies. For photoionized plasmas, on the other hand, codes such as Cloudy [@cloudy] and XSTAR [@xstar; @bautista01] treat the transfer of X-ray lines using either the escape probability method or the assumption of complete redistribution. While these assumptions yield reasonable results for [*isolated*]{} lines with moderate line optical depths ($\tau \la 10$), it does not properly describe line scattering when absorption in the damping wings becomes important (i.e., for $\tau \ga 100 \sim 1000$ depending on the Voigt parameter; see, e.g., @shine75). Such conditions might be relevant for a wide variety of astrophysical environment including ionized surface layers of accretion disks, extended circumsource regions in [[AGN]{}]{}, and possibly shocked regions in the stellar wind of hot young stars. The purpose of this paper is to demonstrate that radiative transfer effects can alter intensities of some of the brightest emission lines in the X-ray band, in particular, the hydrogen- and helium-like lines from carbon, nitrogen, and oxygen. Non-solar [[CNO]{}]{} abundance ratios have been inferred in a number of sources observed with the grating spectrometers on [[*Chandra*]{}]{} and [[*XMM-Newton*]{}]{}. The relativistic emission line interpretation of the [[*XMM-Newton*]{}]{} [[RGS]{}]{} spectra of the Seyfert 1 galaxies [MCG]{}$-$6-30-15 and Mrk 766 presented by @branduardi01, for example, requires the nitrogen emission lines, which are presumably formed in an optically thick accretion disk, to be much stronger than that of the oxygen line assuming solar abundance ratios (see also, @mason03). Similarly, @kinkhabwala02 conclude that nitrogen in the extended circumnuclear regions in the Seyfert 2 galaxy [NGC]{} 1068 is overabundant by a factor of $\sim 3$ based on measurements of the hydrogen- and helium-like line intensities with the [[RGS]{}]{} (see also, @brinkman02 [@ogle03]). @jimenez02 also report abnormally high nitrogen line intensities relative to those of oxygen in the [[RGS]{}]{} spectrum of the low-mass X-ray binary Her X-1, and interpreted as an overabundance of nitrogen due to H-burning by the [[CNO]{}]{}-cycle. All of the modeling, however, are based on rather simple models that do not include detailed line transfer. We show that the effects of line overlap alone can explain some of the observed anomalies, in particular the strength of the nitrogen lines, if the optical depth is large, and that detailed transfer models are required to infer accurate values for the [[CNO]{}]{} abundances with X-ray spectroscopic data. We note, however, that the theoretical model presented here is highly idealized in order to simply illustrate the general effects and describe the relevant spectroscopic details, and is certainly not adequate for quantitative comparisons with observations of complicated systems. The paper is organized as follows. In §2, we describe the spectroscopic properties of the frequency range near the  ([O$^{7+}$]{}) [Ly$\alpha$]{} doublet transitions, which contains a pair of overlapping  ([N$^{6+}$]{}) [Ly$\zeta$]{} lines and underlying sources of continuum opacity, which is usually dominated by K-shell absorption of H- and He-like carbon ([C$^{4+}$]{} and [C$^{5+}$]{}) and He-like nitrogen ([N$^{6+}$]{}). We compute the destruction probabilities of  [Ly$\alpha$]{} due to line absorption by the  [Ly$\zeta$]{} transition and continuum absorption by carbon and nitrogen for temperatures ranging from cool photoionized media ($kT \sim 10~\rm{eV}$) to hot collisionally ionized plasmas ($kT \sim 1~\rm{keV}$), which is described in §3. We demonstrate that the wavelength coincidence allows efficient conversion of  [Ly$\alpha$]{} line radiation into the  lines, and may be misidentified as a [[CNO]{}]{} abundance anomaly (§4). Finally, in §5, a few other potentially important overlaps are identified and their effects on the global X-ray spectrum under optically thick conditions are briefly discussed. Spectroscopy ============ In the non-relativistic, hydrogenic approximation, the transition energy of the  [Ly$\alpha$]{} line ($n = 2 \rightarrow 1$) coincides exactly with that of the  [Ly$\zeta$]{} ($n = 7 \rightarrow 1$) line; i.e., $Z_a^2(1-1/n^2) = Z_b^2(1-1/m^2)$ holds exactly for $Z_a = 8$, $n = 2$ and $Z_b = 7$, $m = 7$. More precisely, the  [Ly$\alpha$]{} transition is a doublet at wavelengths $\lambda_{\rm{Ly}\alpha_1} = 18.96711$ Å ($653.680~\rm{eV}$) and $\lambda_{\rm{Ly}\alpha_2} = 18.97251$ Å ($653.494~\rm{eV}$) that correspond to decays to the ground state from the $^2P_{3/2}$ and $^2P_{1/2}$ levels, respectively. The oscillator strength of the [Ly$\alpha_1$]{} line is twice that of the [Ly$\alpha_2$]{} line. The [Ly$\zeta$]{} transition is also a doublet with $\lambda_{\rm{Ly}\zeta_1} = 18.97411$ Å ($653.439~\rm{eV}$) and $\lambda_{\rm{Ly}\zeta_2} = 18.97418$ Å ($653.436~\rm{eV}$). Since the wavelength difference of $0.07$ mÅ is much too small to be resolved at any reasonable temperature where [N$^{6+}$]{} can exist, we simply treat them as a single line with an oscillator-strength-averaged wavelength of $\lambda = 18.97413$ Å ($653.438~\rm{eV}$; see, Figure \[fig:grot\]). The  [Ly$\alpha$]{} and  [Ly$\zeta$]{} transition wavelengths are adopted from @johnson85 and @garcia65, respectively, which include reduced-mass and quantum electrodynamical corrections to the eigenvalues of the Dirac equation[^1]. Oscillator strengths and radiative decay rates are calculated using the Flexible Atomic Code ([[FAC]{}]{}[^2]; @fac). All of the relevant atomic parameters are listed in Table \[tbl1\]. The relative cross sections assuming solar abundances ($A_{\rm{N}}/A_{\rm{O}}$ = 0.13; @anders89) and a temperature of $kT = 10 ~\rm{eV}$, typical for a photoionization-dominated medium at this level of ionization, are shown in Figures \[fig:cross10\]. At this temperature, the  [Ly$\zeta$]{} lies approximately 2 Doppler widths away from the  [Ly$\alpha_2$]{} line and 10 Doppler widths from the [Ly$\alpha_1$]{} line. The cross section is almost three orders of magnitude lower than that of the  lines, which implies that significant scattering between the lines occur when the total optical depth is $\ga 10^3$. Similarly, at a temperature of $kT = 50~\rm{eV}$, the  [Ly$\zeta$]{} lies $\sim 1$ Doppler width away from the  [Ly$\alpha_2$]{} line and 4 Doppler widths from the [Ly$\alpha_1$]{} line, as shown in Figure \[fig:cross50\]. The emission coefficient in this spectral range is dominated almost entirely by the  [Ly$\alpha$]{} lines. The intrinsic  [Ly$\zeta$]{} emission coefficient is approximately two to three orders of magnitude lower in both collision-dominated and recombination-dominated plasmas with solar $A_{\rm{N}}/A_{\rm{O}}$ abundance ratio. If the medium is photoionized by some external source of continuum radiation, photoexcitation may enhance the line emissivities of the  lines, but this is important only in the surface layer within $\tau \la 10$, as external radiation will not be able to penetrate deeper into the medium. Since we are primarily interested in the properties of the Bowen mechanism at large optical depths, we ignore the  [Ly$\zeta$]{} intrinsic source term entirely. Absorption of an  line photon by  is followed by one of the following two processes: (1) re-emission in the [Ly$\zeta$]{} line, which results in a pure scattering event or (2) destruction of the line through cascades via the upper levels ($n = 2 - 6$), followed eventually by decay to the ground level. Collisional excitation and de-excitation from these levels are important at only extremely high electron densities ($n_{\rm e,crit} \sim 10^{21}~\rm{cm}^{-3}$), and so we do not consider these processes here. We note, however, that thermalization of the line photons via collisions may occur at sufficiently high density and high optical depth, such that $n_e \langle{N}\rangle_{\rm line} \ga n_{\rm e,crit}$, where $\langle{N}\rangle_{\rm line}$ is the mean number of scatterings experienced by the line photons as they propagate through the medium. The branching ratios for the two possible processes can be determined from the radiative decay rates from the $7p$ level to all possible levels. We use radiative rates calculated with [[FAC]{}]{} except for the two-photon decay rates, which are adopted from @drake86. The branching ratios for processes (1) and (2) are found to be 0.797 and 0.203, respectively. The final fate of the 20.3% of the line photons that initially decay to an upper level can be understood by computing the cascade matrices; quantities that represent the probabilities of one state decaying to another state via all possible channels [@seaton59]. We find, for example, that 58% of them produce the Balmer, Paschen, etc. lines and eventually result in two-photon emission from the $2~^2S_{1/2}$ level, while the remaining 42% produce the Lyman lines in the soft X-ray band. Continuum opacity from ions that coexist with  is dominated primarily by  (ionization potential $\chi = 31.622$ Å),  ($\chi = 25.303$ Å), and  ($\chi = 22.458$ Å) for both collisionally and photoionized plasmas in ionization and thermal equilibrium. The threshold wavelength of  ($\chi = 18.587$ Å) lies shortward of the [Ly$\alpha$]{} line. Lower charge states of carbon and nitrogen may contibute very small amounts of continuum opacity as well. For solar abundances, the typical value for the combined continuum opacity relative to the [Ly$\alpha$]{} line opacity is $\sim 10^{-4}$, but can range anywhere from a $\rm{few} \times 10^{-6}$ to as high as $\sim 10^{-3}$ depending on the level of ionization and the temperature. In ionization equilibrium, a carbon or nitrogen ion that absorbs the  [Ly$\alpha$]{} line will be photoionized, and will eventually recombine to form emission lines and continua. In a recombination-dominated plasma at a temperature of $kT \sim 100~\rm{eV}$, for example, calculations show that approximately 50% of the total number of recombinations produce one of the Lyman lines. Under most circumstances, the Compton cross section is roughly two orders of magnitude below the continuum absorption cross section. This implies that lines photons are preferentially absorbed before they are Compton scattered and, therefore, continuum absorption is the dominant mechanism that competes with the line processes. Compton scattering effects will, however, be important in extremely highly-ionized regions, where only trace abundances of bound ions can survive or in low-metalicity plasmas with abundances $\la 1$% solar. These will be ignored entirely in the present work. Computational Method ==================== To illustrate the importance of line overlap, we assume an infinite, symmetric slab with a pre-specified source distribution of line photons. In reality, dynamical effects, such as velocity gradients and turbulence, as well as inhomogenieties that exist, for example, in accretion disks complicate the problem, but we ignore them in the present study. Although resonance line scattering, in general, produces polarized light, we assume that the radiation field is unpolarized (cf., @lee94). To simpify the computational procedure even further, the medium is assumed to be isothermal with a constant level of ionization. We solve the radiative transfer equation in the frequency range near the  [Ly$\alpha$]{} lines by properly accounting for scattering and absorption by the various overlapping lines, as well as the presence of a background continuum opacity source. Using the emergent intensity, we then compute the escape probabilities of the original  [Ly$\alpha$]{} line photons and conversion efficiencies into the  lines and photoelectric absorption as functions of the total optical depth of the slab. The radiative transfer equations for rays propagating in the $+\mu$ and $-\mu$ directions ($\mu~\equiv~cos~\theta$, where $\theta$ is the angle from the normal $z$-axis) in a plane-parallel medium can be written as, $$\label{eq:rte1} \pm \mu \frac{\partial I(z,\nu,\pm\mu)}{\partial z} = \sum_i k_i(\nu)[S_i(z,\nu) - I(z,\nu,\pm\mu)]$$ where $\theta$ is the angle measured from the $z$-direction, which we define to be normal to the slab. The right-hand-side is summed is over all possible sources of opacity and emissivity in the slab. We express the frequency in terms of the frequency shift from the reference line center at frequency $\nu_0$ in units of the Doppler width $\Delta \nu_D$, i.e., $x = (\nu - \nu_0)/\Delta \nu_D$, and write the line absorption coefficient (cm$^{-1}$) as $k_i(x) = k_{iL} \phi_a(x)$. Since most X-ray resonance lines have large radiative decay rates ($A_r \ga 10^{12} ~\rm{s}^{-1}$), the population of the upper level is usually negligible compared to that of the ground state. Denoting the transition oscillator strength as $f_{ul}$ and the lower-level density as $n_i$ ($\rm{cm}^{-3}$), we can then write, $$\label{eq:kL} k_{iL} = \frac{\pi e^2}{m_e c} \frac{f_{ul} n_i}{\Delta \nu_D},$$ and $\phi_a(x)$ is the normalized Voigt profile, $$\label{eq:voigt} \phi_a(x) = \frac{H(a,x)}{\sqrt{\pi}} = \frac{a}{\pi^{3/2}} \int^{\infty}_{-\infty} \frac{e^{-t^2} dt}{(x-t)^2 + a^2}.$$ Adopting the notation of @hummer80, we define $\tau$ to be the mean optical scale, which increases in the negative $z$-direction through the slab. Therefore, $d\tau = - k_L dz$, which is related to the line center optical depth $\tau_0$ by, $\tau_0 = \phi_a(0) \tau$. We assume that the continuum is purely absorbing with zero local emissivity. The source function then contains only the line terms, and can be written as [@hummer69], $$S_{iL}(\tau,x) = \frac{1 - \epsilon_i}{\phi_{i,a}(x)} \int_{-\infty}^{+\infty} R(x,x') J(\tau,x') dx' + G_i(\tau),$$ where the first term represents the scattering term and $G_i(\tau)$ is the intrinsic source term. The quantity $\epsilon_i$ is the destruction probability of the line per scattering (i.e., $\epsilon = 0.203$ for the  [Ly$\zeta$]{} line) and $J(\tau,x)$ is the angle-averaged intensity, $$J(\tau,x) = \frac{1}{2} \int_{-1}^{+1} I(\tau,x,\mu)d\mu.$$ The scattering term is computed assuming partial redistribution in a Voigt profile, in which photons scatter coherently in the atoms’ frame. Compared to the assumption of complete redistribution, this provides a more accurate representation of line scattering in an optically thick medium where scattering in the damping wings is important. The frequency redistribution function $R(x,x')$ is defined as the probability that a photon of frequency $x$ absorbed by an ion is re-emitted at frequency $x'$, and can be written as, $$R(x,x') = \frac{1}{\pi^{3/2}} \int^{+\infty}_{-\infty} e^{-u^2} \left\{\tan^{-1}\left[\frac{{\rm{max}}(x,x') + x}{a}\right] - \tan^{-1}\left[\frac{{\rm{min}}(x,x') - x}{a}\right] \right\} du,$$ which was first derived by @henyey41 (see, also @hummer62, and references therein). Accounting for the angular redistribution makes negligible difference [@milkey75] and is, therefore, ignored. The intrinsic source term is frequency independent and can be produced, for example, through collisional excitation and or radiative cascades following recombination. The exact nature of the excitation mechanism is not important for our present purposes. The transfer equation for multiple overlapping lines can then be written as, $$\label{eq:rte2} \pm \mu \frac{dI_\nu(x,\mu,\tau)}{d\tau} = \sum_{i} \left[ I(x,\mu,\tau) \alpha_{i}(x) + (1 - \epsilon_{i}) \frac{\alpha_{i}(x)}{\phi_a(x)} \int^{\infty}_{-\infty} R_{i}(x', x) J(\tau,x') dx' + \alpha_{i}(x) G_{i}(\tau) \right],$$ where $\alpha_i(x)$ represent the absorption coefficients normalized to that of the reference line at $x = 0$, and are assumed to be independent of depth. Therefore, the first, second, and third terms on the right hand side of Eq. \[eq:rte2\] represent absorption, scattering, and intrinsic emission, respectively. The transfer equation is solved using the Feautrier method [@feautrier64], which requires a discretized version of Eq. \[eq:rte2\] in angular, frequency, and optical depth variables. Defining the total opacity $\phi_{\rm{tot}}(x) = \sum\phi_{i,a}(x) + \beta$, where $\beta$ is the continuum opacity relative to the reference line opacity ($\beta = k_C/k_{L}$) assumed to be independent of frequency across the line, we define a mean-intensity-like quantity, $$j_{\mu\nu} = \frac{1}{2}[I(x,+\mu,\nu) + I(x,-\mu,\nu)],$$ and rewrite Eq. \[eq:rte2\] as a second order differential equation, $$\label{eq:rte3} \mu^2 \frac{d^2j_{\mu\nu}}{d\tau^2} = \phi^2_{\rm{tot}}(x) j_{\mu\nu} + \phi_{\rm{tot}}(x) \sum_{i=0} \alpha_i(x) S_i(\tau, x).$$ The presence of temperature and ion abundance gradients will introduce an additional term on the right hand side, which is ignored for simplicity. We assume a symmetric slab and specify boundary conditions on the surface and at the midplane of the slab. Assuming no incident radiation on either boundaries, they can be written as, $$\label{eq:bc_top} \mu \left. \frac{dj_{\mu\nu}}{d\tau} \right \vert_{\tau = 0} = j_{\mu\nu}$$ on the surface and, $$\label{eq:bc_mid} \mu \left. \frac{dj_{\mu\nu}}{d\tau} \right \vert_{\tau = \tau_{\rm{tot}}/2} = 0$$ at the midplane. These boundary conditions are also discretized using a second order method described by @auer67. The frequency integral is represented as a quadrature sum using Simpson’s formula with $N_x = 100 - 200$ points depending on the temperature and the total optical depth of the slab. We use a 2-point Gaussian quadrature for the angular integral in the interval $0 \leq \mu \leq 1$ ($N_\mu = 2$). The optical depth is discretized into 12 points per dex. Extensive testing shows that the above discretization scheme yields reliable results. Energy conservation (see, e.g., @hummer80) is satisfied to within $\sim 2$% for most cases studied in this paper, and is no larger than $\sim 5$% even in the worst cases. Conversion Efficiencies ======================= We solve the transfer equation for a source function distributed uniformly within a slab of ionized material for a range of total slab thicknesses. We define the  [Ly$\alpha_1$]{} line to be our reference line and measure frequency shifts in units of its Doppler width. The emissitivities of the  [Ly$\alpha_1$]{} and [Ly$\alpha_2$]{} lines are assumed to be proportional to their respective statistical weights of the upper levels, i.e., $G_{\rm{Ly}\alpha_1}(\tau) = 2 G_{\rm{Ly}\alpha_2}(\tau)$, which is approximately correct for both collision-dominated and recombination-dominated line emission. The absolute numerical values for the source functions are not important for the present purposes, and without any loss of generality, we can write, $$\label{eq:src_norm} \sum_i \int_0^{\tau_{\rm{tot}}} d\tau_i G_i(\tau) \equiv 1.$$ The solution to Eq. \[eq:rte3\] – \[eq:bc\_mid\] yields the frequency-dependent intensity at each optical depth and angular points, which can then be used to compute the mean number of scatterings for each of the lines, their mean optical scales, and the escape probabilities. Given the normalization condition (Eq. \[eq:src\_norm\]), the fraction of the original photons that are absorbed by a transition with destruction probability $\epsilon_i > 0$ can be expressed in terms of the mean number of scatterings [@hummer69; @hummer80]. The fraction absorbed and destroyed by line $i$ can be written as, $$f_{iL} = \frac{\epsilon_i}{1 - \epsilon_i}\langle N \rangle_i = \frac{\epsilon_i}{1 - \epsilon_i} \int_0^{\tau_{\rm{tot}}} d\tau_i \int_{-\infty}^{\infty} \phi_a(x) S_{iL}(\tau,x) dx,$$ where $\langle N \rangle_i$ is the mean number of scatterings. Similarly, the fraction of photons absorbed in the continuum can be expressed in terms of the mean optical scale $\langle l \rangle$ [@ivanov73], i.e., $$f_C = \beta \langle l \rangle = \beta \int_0^{\tau_{\rm{tot}}} d\tau \int_{-\infty}^{\infty} J(\tau,x) dx.$$ Finally, the fraction of photons that escape through both faces of the slab is simply, $$f_{\rm{esc}} = \int_{-\infty}^{\infty} dx \int_{0}^{1} I(\tau = 0, x,\mu)~\mu~d\mu,$$ where $I(\tau = 0,x,\mu)$ is the emergent intensity on the surface. Fractional conversion efficiencies as a function of the total threshold optical depth at the  edge ($\tau_{\rm{O~VIII}}$) at temperatures of $kT = 10 ~\rm{eV}$ and $50 ~\rm{eV}$ are shown in Figures \[fig:yield10\] and \[fig:yield50\], respectively. The underlying continuum opacity is assumed to be $0.1~\tau_{\rm{O~VIII}}$. At $kT = 10 ~\rm{eV}$, the line-center cross section of the  [Ly$\alpha_1$]{} line is $6.9 \times 10^{-16}~\rm{cm}^2$ and is a factor of $7.0 \times 10^3$ times larger than the corresponding continuum absorption cross section at threshold, which is assumed to be $9.9 \times 10^{-20}~\rm{cm}^2$ for  [@verner96]. For $\tau_{\rm{O~VIII}} = 1$ and $kT = 10 ~\rm{eV}$ (Figure \[fig:yield10\]), only $\sim 40$% of the  [Ly$\alpha$]{} photons escape the medium, while $\sim 40$% are absorbed by  and $\sim 20$% are absorbed by the background continuum. The total line-center opacity in the  [Ly$\alpha_1$]{} line is $6.9 \times 10^3$. We find the mean number of scatterings for the  [Ly$\alpha_1$]{} line to be $2.3 \times 10^3$, which implies that thermalization through particle collisions is relevant if the gas density is higher than $\sim 10^{18}~\rm{cm}^{-3}$ (see §2). At a temperature of $kT = 50 ~\rm{eV}$, the overlap between the  and  lines is larger, and the conversion to  is enhanced relative to the amount absorbed in the continuum. The temperature dependence of the conversion efficiencies can be understood from the curves plotted in Figure \[fig:yield\_kt\]. At low temperatures, the  [Ly$\zeta$]{} line lies in the wings of the  [Ly$\alpha$]{} lines and so the conversion process is relatively inefficient. As the temperature is increased, the amount of overlap between the lines increases rapidly up to a temperature of $kT \sim 40~\rm{eV}$, which is where the separation between the  [Ly$\zeta$]{} and  [Ly$\alpha_1$]{} lines is approximately equal to one Doppler width. This temperature is representative of a photoionized plasma at this level of ionization. Above this temperature, the amount of line overlap increases more slowly than the temperature dependence of the line cross section, which decreases approximately as $\propto T^{-1/2}$, and so the efficiency drops again with temperature. Finally, we show in Figure \[fig:yield2\] the effect of increasing the overlapping continuum opacity on the conversion efficiencies. From the derived conversion efficiencies, one can estimate the emergent intensities of each of the lines for a given total optical depth. At $kT = 10~\rm{eV}$ and a total optical depth of $\tau_{\rm{O~VIII}} = 1$, approximately 40% of the original  [Ly$\alpha$]{} photons escape through the surfaces of the slab, 40% of then are absorbed by , and the remaining 20% are absorbed in the continuum (see Figure \[fig:yield10\]). Assuming that the medium is very highly ionized consisting only of H-like carbon, nitrogen, and oxygen, and further assuming that the emission coefficient ratios of the [Ly$\alpha$]{} lines are proportional to their abundance ratios, (i.e., [[CNO]{}]{} [Ly$\alpha$]{} relative line intensities of $1.0:0.13:0.45$), the emergent radiation field will have an  [Ly$\alpha$]{} line intensity of $0.40$, a  [Ly$\alpha$]{} intensity of $0.26$, and a carbon line intensity of $0.55$. Here, we have simply assumed that 33% of the photons absorbed by result in the  [Ly$\alpha$]{} line and 50% of the photons absorbed by produce the  [Ly$\alpha$]{} line (see §2). Therefore, the oxygen to nitrogen line ratio decreased from 7.7 to 1.5 (a factor of $\sim 5$) and the oxygen to carbon line ratio decreased from 2.2 to 0.73 (a factor of $\sim 3$) simply due to finite opacity effects. We do not attempt to compute the global emergent spectrum, which requires a self-consistent calculation of the  [Ly$\alpha$]{} line source functions including a complete set of ions and a realistic ionization balance calculation. To accurately compute the emergent [[CNO]{}]{} line intensities, one must properly account for all of the processes that increase the oxygen line source function. The forest of iron L-shell line emission shortward of the  edge at $\lambda = 14.228$ Å is particularly important and is likely to be the dominant contributor to the source function. A more detailed investigation will be presented in a future article. Conclusions and Discussion ========================== We have identified a pair of overlapping resonance lines, which alters the intensities of some of the most prominent lines in X-ray spectra of cosmic sources under optically thick conditions, analogous to the – Bowen fluorescence mechanism. A non-negligible fraction of the  [Ly$\alpha$]{} photons can be converted into  emission lines, which can be misinterpreted as an anomalous [[[CNO]{}]{}]{} abundance pattern. In general, oxygen emission lines will be suppressed, while both nitrogen and carbon lines will be enhanced compared to those emitted under optically thin conditions. As briefly mentioned earlier, detections of non-solar [[CNO]{}]{} abundance ratios have been made in a variety of sources using soft X-ray data obtained with the Reflection Grating Spectrometer onboard [[*XMM-Newton*]{}]{} and the [[*Chandra*]{}]{} Low Energy Transmission Grating Spectrometer (see, e.g., @branduardi01 [@kinkhabwala02; @jimenez02; @brinkman02; @mason03; @leutenegger03]). In all cases, the nitrogen lines are observed to be stronger than expected relative to the oxygen lines. The carbon lines, on the other hand, are stronger in some sources and weaker in others. Although it is highly likely that non-solar overabundances play a major role in producing the anamolous line fluxes, it is also possible that a substantial part of it is due to the Bowen mechanism described in this paper. Detailed modeling and careful inspections of other lines in the observed spectrum are required to break the degeneracy between overabundance and radiative transfer effects. Dynamical effects, such as turbulent motion and velocity gradients, which are not included in the present calculations, may reduce or enhance the conversion efficiency of  [Ly$\alpha$]{} into  depending on the geometrical configuration of the ionized medium. Turbulence, for example, can suppress the efficiency if the turbulent length scale is much smaller than the mean free path of the  line photons. In the presence of velocity gradients, the efficiency can be enhanced in the direction where the  [Ly$\zeta$]{} line is shifted into the  [Ly$\alpha$]{} line core. These effects coupled with density inhomogeneities will complicate the problem further and must be studied carefully on a case-by-case basis. There are a few other potentially important line overlaps in the X-ray band that are worth mentioning. For example, one of the three brightest lines at wavelengths of 17.096 Å, 17.051 Å, and 16.780 Å typically referred to as the M2, 3G, 3F lines [@brown98], respectively, coincide in wavelength with the high-$n$ series lines of  at 17.10068 Å ($n = 7$), 17.02584 Å ($n = 8$), etc., and 16.786 Å ($n = \infty$). In this case,  dominates the line opacity, while contributes most of the intrinsic source function. Another example is the overlap between the lines at 16.071 and 16.004 Å [@brown02] with the [Ly$\beta$]{} lines at 16.00666 Å and 16.00552 Å. These examples are certainly not exhaustive and it is highly likely that they must all be accounted for collectively for accurately predicting the global X-ray spectrum of optically thick sources. The author thanks Roger Blandford, Steven Kahn, and Duane Liedahl for valuable discussions, and Mingfeng Gu for his help with the Flexible Atomic Code. This work was supported by [[NASA]{}]{} through [[*Chandra*]{}]{} Postdoctoral Fellowship Award Number [PF]{}1-20016 issued by the [[*Chandra*]{}]{} X-ray Observatory Center, which is operated by the Smithsonian Astrophysical Observatory for and behalf of [[NASA]{}]{} under contract [ NAS]{}8-39073. Anders, E. & Grevesse, N. 1989, , 53, 197 Auer, L. 1967, , 150, L53 Bautista, M. A. & Kallman, T. R. 2001, , 134, 139 Bowen, I. S. 1934, , 46, 146 Bowen, I. S. 1935, , 81, 1 Branduardi-Raymont, G., Sako, M., Kahn, S. M., Brinkman, A. C., Kaastra, J. S., & Page, M. J. 2001, , 365, L140 Brinkman, A. C., Kaastra, J. S., van der Meer, R. L. J., Kinkhabwala, A., Behar, E., Kahn, S. M., Paerels, F. B. S., & Sako, M. 2002, , 396, 761 Brown, G. V., Beiersdorfer, P., Liedahl, D. A., Widmann, K., & Kahn, S. M. 1998, , 502, 1015 Brown, G. V., Beiersdorfer, P., Liedahl, D. A., Widmann, K., Kahn, S. M., & Clothiaux, E. J. 2002, , 140, 589 Drake, G. W. F. 1986, , 34, 2871 Drake, S. A. & Ulrich, R. K. 1980, , 42, 351 Eastman, R. G. & MacAlpine, G. M. 1985, , 299, 785 Feautrier, P. 1964, C. R. Acad. Aci. Paris, 258, 3189 Ferland, G. J., Korista, K. T., Verner, D. A., Ferguson, J. W., Kingdon, J. B., & Verner, E. M. 1998, , 110, 761 Garcia, J. D. & Mack, J. E. 1965, Journal of the Optical Society of America, 55, 654 Gu, M. F. 2003, , 582, 1241 Harrington, J. P. 1972, , 176, 127 Henyey, L. G. 1941, Proc. Nat. Acad. Sci., 26, 50 Hummer, D. G. 1962, , 125, 21 Hummer, D. G. 1969, , 145, 95 Hummer, D. G. & Kunasz, P. B. 1980, , 236, 609 Ivanov, V. V. 1973, Transfer of radiation in spectral lines (NBS Special Publication, Washington: US Department of Commerce, National Bureau of Standards, 1973, English language edition) Jimenez-Garate, M. A., Hailey, C. J., Herder, J. W. d., Zane, S., & Ramsay, G. 2002, , 578, 391 Johnson, W. R. & Soff, G. 1985, Atomic Data and Nuclear Data Tables, Vol. 33, p.405, 33, 405 Kallman, T. & McCray, R. 1980, , 242, 615 Kallman, T. R. & McCray, R. 1982, , 50, 263 Kinkhabwala, A., Sako, M., Behar, E., Kahn, S. M., Paerels, F., Brinkman, A. C., Kaastra, J. S., Gu, M. F., et al. 2002, , 575, 732 Lee, H.-W., Blandford, R. D., & Western, L. 1994, , 267, 303 Leutenegger, M. A., Kahn, S. M., & Ramsay, G. 2002, , 585, 1015 Mason, K. O., Branduardi-Raymont, G., Ogle, P. M., Page, M. J., Puchnarewicz, E. M., Behar, E., C' ordova, F. A., Davis, S., et al. 2003, , 582, 95 Milkey, R. W., Shine, R. A., & Mihalas, D. 1975, , 202, 250 Netzer, H., Elitzur, M., & Ferland, G. J. 1985, , 299, 752 Ogle, P. M., Brookings, T., Canizares, C. R., Lee, J. C., & Marshall, H. L. 2003, , 2003, 402, 849 Seaton, M. J. 1959, , 119, 90 Shine, R. A., Milkey, R. W., & Mihalas, D. 1975, , 201, 222 Unno, W. 1955, , 7, 81 Verner, D. A., Ferland, G. J., Korista, K. T., & Yakovlev, D. G. 1996, , 465, 487 Weymann, R. J. & Williams, R. E. 1969, , 157, 1201 [ccccc]{} [Ly$\alpha_1$]{}& 18.96711 & $2.77 \times 10^{-1}$ & $2.58 \times 10^{12}$ & 1.000\ [Ly$\alpha_2$]{}& 18.97251 & $1.39 \times 10^{-1}$ & $2.58 \times 10^{12}$ & 1.000\ [Ly$\zeta$]{}& 18.97413 & $4.81 \times 10^{-3}$ & $2.97 \times 10^{10}$ & 0.797\ ![Partial Grotrian diagrams of the  and  transitions relevant for the X-ray Bowen fluorescence mechanism. The probabilites for each channel $P \equiv 1 - \epsilon$ are indicated.[]{data-label="fig:grot"}](f1.eps) [^1]: see, also the [[NIST]{}]{} Atomic Spectra Database – http://physics.nist.gov/cgi-bin/AtData/main\_asd [^2]: ftp://space.mit.edu/pub/mfgu/

ArXiv

22.5cm 15.0cm 1.0in 1.0in Tomo Munehisa and Yasuko Munehisa\ 0.5in Faculty of Engineering, Yamanashi University\ Kofu, Yamanashi, 400 Japan\ 1.5in We propose new approach to numerical study of quantum spin systems. Our method is based on a fact that one can use any set of states for the path integral as long as it is complete. We apply our method to one-dimensional quantum spin system with next-to-nearest neighbor interactions. We found remarkable improvement in negative sign problem. [**Section 1 Introduction**]{} Recently quantum spin systems have obtained much interests among people in various fields. One reason for it is that Haldane found characteristic property of the quantum spin system which is difficult to imagine in the classical case [@fdmh]. Another reason comes from the possible relationship between the antiferromagnetic system on a 2-dimensional square lattice and the high $T_c$ materials [@hightc]. One powerful tool to numerically investigate quantum spin systems is Monte Carlo approach using the Suzuki-Trotter formula [@st]. Study through this method has brought us very intriguing results on the ferromagnetic system. If one applies this method to the antiferromagnetic system, however, one often encounters the so-called negative sign (N.S.) problem. N.S. problem, which becomes more serious on larger lattices, makes it very difficult to get statistically meaningful results in numerical calculations. Let us explain what the N.S. problem is. To this purpose we show how the Monte Carlo method is applied to the quantum spin $1/2$ system whose Hamiltonian is $$\hat{H} = \sum \vec{\sigma}_i \vec{\sigma}_j,$$ where $\vec{\sigma}_i=(\sigma_i^x,\sigma_i^y,\sigma_i^z)$ represents Pauli matrix on site $i$ of a lattice and sum runs over all nearest neighbors. The partition function $Z$ is calculated by $$Z = tr( e^{-\beta \hat{H} } )$$ Using the Suzuki-Trotter formula it can be written by the classical partition function as follows. $$Z = \lim_{n \rightarrow \infty} tr \lbrace (e^{-\beta \hat{H}_1/n } e^{-\beta \hat{H}_2/n } ) ^n \rbrace$$ $$= \lim_{n \rightarrow \infty} \sum_{ \lbrace \alpha_i, \alpha'_i \rbrace } \prod_{i=1}^n < \alpha_i \mid e^{-\beta \hat{H}_1/n} \mid \alpha'_i > < \alpha'_i \mid e^{-\beta \hat{H}_2/n} \mid \alpha_{i+1} >,$$ $$\hat{H} = \hat{H}_1 + \hat{H}_2.$$ Here $\mid \alpha_i>$ represents a state of the system, $\lbrace \mid \alpha_i> \rbrace $ a complete set of states $$\hat{1} = \sum_{\alpha_i} \mid \alpha_i > < \alpha_i \mid,$$ and $\hat{H}$ is divided into two parts so that every term in $\hat{H}_1$ ($\hat{H}_2$) commutes with other terms within that partial Hamiltonian. Since the expectation value of $e^{-\beta \hat{H}/n}$ is a real number we can apply Monte Carlo methods to this partition function if $n$, which is called trotter number, is kept finite. There would have been no N.S. problem if positivity of the expectation value were always guaranteed. However in several cases of much physical interest, antiferromagnetic quantum spin system for example, we do have some negative expectation values. If total product of expectation values over one configuration can become negative, then the N.S. problem may occur. In order to numerically calculate some physical quantity $\langle A \rangle $ one should subtract contributions of negatively signed configurations, $A_{-}$, from those of positively signed ones, $A_{+}$, namely, $$\langle A \rangle ={{A_{+}-A_{-}} \over {Z_{+}-Z_{-}}},$$ where $Z_{+}(Z_{-})$ is number of configurations with positive (negative) weight. The result would suffer from serious cancellation when $Z_{-} \simeq Z_{+}$. Note that negativity of expectation values does not always bring the N.S. problem. On square lattices one does not need to worry about it even in the antiferromagnetic case because the total product of expectation values is always positive. On the triangular lattice, on the contrary, the N.S. problem is quite serious. In this paper we describe our prescription to obtain meaningful numerical results for systems to which the conventional formulation is useless because of the N.S. problem. Our strategy is to find rearranged basis of the system with which no serious cancellations take place. We do not mean that the N.S. problem is completely solved by our approach, but we show our method is successful to improve the numerical results when applied to a quantum spin system. In following sections we concentrate ourselves to one dimensional quantum spin $1/2$ system with the next-to-nearest neighbor interactions, which is the simplest one among systems suffering from serious N.S. problem. We show what choice of complete set of states is more suitable in this case to carry out Monte Carlo simulations. In section 2 we present our method. Numerical results will be given in section 3 and final section is devoted to discussions on technical problems. Physical properties of the system will be discussed in other paper. 0.5in [**Section 2 Method**]{} Hamiltonian of the quantum spin $1/2$ system with next-to-nearest neighbor interactions on a one-dimensional chain is $$\hat{H}= -{J \over 2} \sum_{i=1}^N (\vec{\sigma}_i\vec{\sigma}_{i+1} + \vec{\sigma}_i\vec{\sigma}_{i+2}),$$ where $N$ is number of sites on the chain and $\vec{\sigma}_{N+i} \equiv \vec{\sigma}_i$ (periodic boundary condition). The system is ferromagnetic (antiferromagnetic) for positive (negative) $J$. We imply the coupling constant between next-to-nearest neighbors is equal to that between nearest neighbors. It is partly because for simplicity and partly because this case brings most serious N.S. problem for negative $J$. The partition function $Z$ is given by $Z=tr(e^{-\beta \hat{H}})$. First let us describe the conventional approach. State on each site is represented by $z$-component of the spin, namely up and down, or $+$ and $-$. In this representation states of the system are given by $$\mid \alpha > = \mid s_1,s_2,s_3,...,s_N > ,$$ where $s_i=+$ or $-$. The identity operator is then $$\hat{1} = \sum_{\lbrace s_i \rbrace} \mid s_1,s_2,...,s_N > < s_1,s_2,...,s_N \mid .$$ To use this identity operator in the Suzuki-Trotter formula, we divide the Hamiltonian into four parts as schematically shown in Fig. 1(a). $$Z = \lim_{n \rightarrow \infty} tr \lbrace (e^{-\beta \hat{H}_1/n } e^{-\beta \hat{H}_2/n } e^{-\beta \hat{H}_3/n } e^{-\beta \hat{H}_4/n } ) ^n \rbrace ,$$ $$\hat{H}_1 = -{J \over 2} \sum_{i=1}^{N/2} \vec{\sigma}_{2i-1}\vec{\sigma}_{2i},$$ $$\hat{H}_2 = -{J \over 2} \sum_{i=1}^{N/2} \vec{\sigma}_{2i}\vec{\sigma}_{2i+1},$$ $$\hat{H}_3 = -{J \over 2} \sum_{i=1}^{N/4} (\vec{\sigma}_{4i-3}\vec{\sigma}_{4i-1} + \vec{\sigma}_{4i-2}\vec{\sigma}_{4i}),$$ $$\hat{H}_4 = -{J \over 2} \sum_{i=1}^{N/4} (\vec{\sigma}_{4i-1}\vec{\sigma}_{4i+1} + \vec{\sigma}_{4i }\vec{\sigma}_{4i+2}).$$ With these complete sets and partial Hamiltonians we obtain the partition function $Z_C^{(n)}$ which we use in Monte Carlo calculations, $$Z_C^{(n)} = \sum_{\lbrace \alpha_j, \alpha'_j, \alpha''_j, \alpha'''_j \rbrace } \prod_{j=1}^n <\alpha_j \mid e^{-\beta\hat{H}_1/n} \mid \alpha'_j> <\alpha'_j \mid e^{-\beta\hat{H}_2/n} \mid \alpha''_j>$$ $$\ \ \ \ \ \ \times <\alpha''_j \mid e^{-\beta\hat{H}_3/n} \mid \alpha'''_j> <\alpha'''_j \mid e^{-\beta \hat{H}_4/n} \mid \alpha_{j+1}> ,$$ where suffix $j$ numbers sites along the trotter axis and $\alpha_{n+1} \equiv \alpha_1$. Now we turn to our method. First we will show our idea using spin system of two sites which we denote site $a$ and site $b$. The Hamiltonian of this system is simply $\hat{H}_{ab}= -J \vec{\sigma_a} \vec{\sigma_b} $. In the conventional approach to this two-spin system the complete set is given by $ \lbrace \mid +_a,+_b >, \mid +_a,-_b >, \mid -_a,+_b >, \mid -_a,-_b > \rbrace .$ Instead of this set we can adopt another complete set composed of diagonalized states of $\hat H_{ab}$, $$\lbrace \mid 1 >, \mid \oplus >, \mid \ominus >, \mid -1 > \rbrace ,$$ where $$\mid 1 > =\mid +_a , +_b > ,$$ $$\mid \oplus > = {1\over \sqrt{2} } (\mid +_a , -_b >+ (\mid -_a , +_b >),$$ $$\mid \ominus > ={1\over\sqrt{2} } (\mid +_a , -_b >- (\mid -_a , +_b >),$$ $$\mid -1 > = \mid -_a , -_b > .$$ The partition function is then easily calculated using eigenvalues $E_1$, $E_\oplus$, $E_\ominus$ and $E_{-1}$, $$Z_{ab}=tr( e^{-\beta \hat{H}_{ab}})$$ $$= e^{-\beta E_1}+e^{-\beta E_{\oplus}}+ e^{-\beta E_{\ominus}}+ e^{-\beta E_{-1}}.$$ We apply this diagonalization to the spin system of $N$ sites. For this purpose we rewrite the Hamiltonian in the following form $$\hat{H}= -{J \over 2} \sum_{i=1}^{N/2} (\vec{\sigma}_{a,i}\vec{\sigma}_{a,i+1} + \vec{\sigma}_{b,i}\vec{\sigma}_{b,i+1} + \vec{\sigma}_{a,i}\vec{\sigma}_{b,i} + \vec{\sigma}_{b,i}\vec{\sigma}_{a,i+1}),$$ where we denote odd and even sites with suffix $a$ and $b$, respectively, $$\vec{\sigma}_{a,i} \equiv \vec{\sigma}_{2i-1},\ \ \vec{\sigma}_{b,i} \equiv \vec{\sigma}_{2i}.$$ We employ the complete set where the operators $\vec{\sigma}_{a,i} \vec{\sigma}_{b,i} $ are diagonalized. Then each state of the system is represented by combinations of $N/2$ two-site diagonalized states, $$\mid \alpha > = \mid S_1,S_2,...,S_{N/2} >,$$ where $S_i$ stands for diagonalized state on the $i$-th spin pair, namely $S_i=1_i$, $\oplus_i$, $\ominus_i$ or $-1_i$. Then it becomes necessary to divide the Hamiltonian into “odd” and “even” parts $$\hat{H}=\hat{H_o}+\hat{H_e},$$ where, as shown in Fig. 1(b), $$\hat{H_o} = -{J \over 2} \sum_{i=1}^{N/4} (\vec{\sigma}_{a,2i-1} \vec\sigma_{a,2i} + \vec\sigma_{b,2i-1} \vec\sigma_{b,2i} + \vec\sigma_{b,2i-1} \vec\sigma_{a,2i}$$ $$\ \ \ \ \ + {1 \over 2} \vec\sigma_{a,2i-1} \vec\sigma_{b,2i-1} + {1 \over 2} \vec\sigma_{a,2i} \vec\sigma_{b,2i} ),$$ $$\hat{H_e} = -{J \over 2} \sum_{i=1}^{N/4} (\vec{\sigma}_{a,2i} \vec\sigma_{a,2i+1} + \vec\sigma_{b,2i} \vec\sigma_{b,2i+1} + \vec\sigma_{b,2i} \vec\sigma_{a,2i+1}$$ $$\ \ \ \ \ + {1 \over 2} \vec\sigma_{a,2i-1} \vec\sigma_{b,2i-1} + {1 \over 2} \vec\sigma_{a,2i} \vec\sigma_{b,2i} ).$$ The partition function with these partial hamiltonians is $$Z=\lim_{n \rightarrow \infty} tr \lbrace (e^{-\beta \hat{H}_o/n} e^{-\beta \hat{H}_e/n} )^n \rbrace .$$ Inserting identity operators made of $\alpha$’s between the exponents we obtain partition function $Z_M^{(n)}$ to use in our numerical study, $$Z_M^{(n)} = \sum_{ \lbrace \alpha_j,\alpha'_j \rbrace } \prod_{j=1}^{n} <\alpha_j \mid e^{-\beta \hat{H}_o/n} \mid \alpha'_{j} > <\alpha'_j \mid e^{-\beta \hat{H}_e/n} \mid \alpha_{j+1}>.$$ [**Section 3 Numerical Results**]{} In this section we present Monte Carlo results we obtained using the new partition function stated in section 2. In order to figure how much quantitative improvement we can make, we also show some results from the conventional approach. Monte Carlo simulations are carried out on a two dimensional lattice, one of which is space direction of size $N$ and another is the trotter direction having $2n$ ($4n$) sites in our (conventional) method with the trotter number $n$. In each simulation with new partition function $Z_{M}^{(n)}$ typically one thousand configurations are generated in order to reach thermal equilibrium and ten thousand configurations after thermalization are used to measure system’s energy, $E$, and ratio of negatively weighted configurations to total configurations, $P$, $$E=-{{\partial} \over {\partial \beta}} ln Z,$$ $$P={{Z_-} \over {Z_+ + Z_-}}.$$ Here one new configuration is obtained by updating spins on all sites using the heat-bath method. In each update we make simultaneous spin flips on sites along the smallest closed path (local flips) and along the trotter axis ($n$-direction global flips) according to their probabilities. From technical reasons we did not include local and global flips which change $z$-component of $S_i$ of $i$-th spin pair by 2 [@fn1]. We also made global flips along the space axis ($x$-direction global flips) available, but we excluded them in simulations presented in this paper because we have found their effects on system’s energy are negligible. In simulations with the conventional partition function $Z_{C}^{(n)}$ we employ Metropolis algorithm. Here, in addition to the local flips and usual $n$-direction global flips, we found it necessary to make a kind of twisted $n$-direction global flips shown in Fig. 1(c) so that enough phase space is guaranteed. We carry out twenty thousand updates for the thermalization, one hundred thousand for the measurement. Let us first present results for the ferromagnetic system. In this case the conventional method should be effective since no N.S. problem exists. Fig. 2 shows results on an $N=8$ chain by the conventional method as a function of temperature $T$, which is defined by $1/J \beta$, together with exact values obtained by diagonalization methods. We see conventional method works very well. Our method is also successful, but we have no special reason to prefer it [@fn2]. Next we will show results for the antiferromagnetic system. In Fig. 3(a) we present ratio $P$ versus $T$ with $N$=8 chain by both methods. As we see in the figure, even when $n=2$ the ratio $P$ with the conventional partition function increases quite rapidly as $T$ decreases. It is therefore difficult to obtain statistically meaningful results by this method. By our new method, on the other hand, $P$ remains much smaller for larger $n$ and for lower $T$ so that the N.S. problem is less serious. Fig. 3(b) plots system’s energy on $N=8$ and $n=2$, 3 and 4 lattices in comparison with exact values. The results indicate we can get reliable values by new method. How far is our method useful when the size of the chain is enlarged? Fig. 3(c), which shows how ratio $P$ in our method increases for larger $N$’s, answers this question. We see that with our present choice of states reliable calculation on $N \geq 32$ and $n \geq 4$ lattices is difficult when $T < 1.0$. 0.5in [**Section 4 Discussions**]{} In this paper we suggested a new approach for Monte Carlo study of quantum spin systems suffering from the negative sign problem. Essential point of our method is to choose a set of states for the path integral which is appropriate to numerical calculation. Conventional choice of the set, whose states consist of eigenstates of $z$-component of the Pauli matrix on each lattice site, is the simplest one but the N.S. problem turns out to be serious with this set. Since any other choice is possible as long as the set is complete, we employ a set made of eigenstates of diagonalized Hamiltonian for every two neighboring sites. We believe this choice is better approximation to true ground state and excited states. We applied our method to one-dimensional quantum spin $1/2$ system with next-to-nearest neighbor interactions, which suffers from the N.S. problem in the antiferromagnetic case. Then remarkable improvements were found as shown in Section 3. Monte Carlo results also showed, however, that even in our method one would encounter difficulties for low $T$ on large lattices because the ratio of negatively signed configurations increases up to $\sim$ 0.5 when temperature goes down or lattice size is enlarged. So we should emphasize that what we present here is not a way to completely solve the N.S. problem but a prescription to improve numerical calculations. Comments on ergodicity are in order. In quantum Monte Carlo simulations it is known that local flips are not sufficient to guarantee ergodic update of the system [@erg]. One should add some kinds of global spin flips — $n$-direction global flips to change $z$-component of spins, $x$-direction global flips and transverse global flips to change the winding number, and so on — in updating procedure in order to make phase space big enough to observe satisfactorily precise results. What global flips are necessary and sufficient is not a trivial question, but from technical point of view it is preferred to include minimum kinds of global flips which are as simple as possible. From our experiences $n$-direction global flips seem inevitable for this purpose. As results in Section 3 indicate, we can measure system’s energy without any other flips in wide range of temperature. Transverse global flips and $x$-direction global flips, however, might be important for low temperatures. It should be investigated in future study. Although a complete solution for N.S. problem could not be obtained, our results show much improvement in numerical study, which inspires us with confidence that quantum Monte Carlo methods are effective if a set of states for the path integral is appropriately chosen. More improvements would be possible by selecting larger cluster of spins to construct states which diagonalize the Hamiltonian. [99]{} F. D. M. Haldane, Phys. Lett. [**A93**]{}, 464 (1983); Phys. Rev. Lett. [**50**]{}, 1153 (1983). P. W. Anderson, Science [**235**]{}, 1196 (1987).\ E. Manousakis, Rev. Mod. Phys. [**63**]{}, 1 (1991). M. Suzuki, Prog. Theor. Phys. [**56**]{}, 1454 (1976); J. Stat. Phys. [**43**]{}, 883 (1986).\ M. ${\rm Makivi\acute{c}}$, [*Computational Approaches in condensed-matter physics*]{}, edited by S. Miyashita, M. Imada and H. Takayama (Springer-Verlag, Berlin, 1992), p.129. Contribution from $n$-direction global flips with $\Delta S_i^z = 2$ should be estimated in future work because it is not clear whether two successive $\Delta S_i^z = 1$ updates can realize all $\Delta S_i^z=2 $ configurations. It should be noted that in our method configurations with negative weights do appear even in the ferromagnetic system, though they are small in number. M. Marcu, [*Quantum Monte Carlo Methods*]{}, edited by M. Suzuki (Springer-Verlag, Berlin, 1987), p.64.\ S. Miyashita, [*Quantum Simulation*]{}, edited by J. D. Doll, J. E. Gubernatis (World Scientific, Singapore, 1990), p.238. [**Figure captions**]{} [**Fig. 1**]{} \(a) Way to divide Hamiltonian on $N$=8 chain in the conventional approach. Filled circles, squares, diamonds and triangles indicate terms in $\hat{H}_1, \hat{H}_2, \hat{H}_3$ and $\hat{H}_4$, respectively. \(b) Way to divide Hamiltonian on $N$=8 chain in our approach, where odd and even sites are renumbered as shown. Open triangles and diamonds represent terms in $\hat{H}_1$ and $\hat{H}_2$, respectively. Open circles denote terms halved by $\hat{H}_1$ and $\hat{H}_2$. \(c) Examples for twisted $n$-direction global flips when $N=8$. Open circles and plaquettes with vertical trotter direction denote site locations and elementary four spin interactions, respectively. All spins on each path shown by bold solid line should be simultaneously flipped. [**Fig. 2**]{} Energy in the ferromagnetic case versus temperature $T$. Data are calculated on $N$=8 chain by diagonalization method (exact) and by the conventional approach for trotter number $n=2$, 3 and 4. For each $T$ and $n$ twenty thousand sweeps are done for the thermalization, one hundred thousand for measurement. [**Fig. 3**]{} Data obtained by new method in the antiferromagnetic case. Each measurement is done using ten thousand configurations generated after one thousand sweeps for the thermalization. \(a) Ratio of negatively signed configurations on $N$=8 chain as a function of temperature. Results from conventional approach are also plotted for comparison. In simulations by conventional method twenty thousand sweeps are done for the thermalization, one hundred thousand for measurement. \(b) Energy of the system observed on $N$=8 and $n$=2, 3 and 4 lattices. \(c) Ratio of negatively signed configurations as a function of the number of lattice sites. Trotter number $n$ is kept 4 here.

ArXiv

--- author: - 'Yoske Sumitomo, Markus Rummel' bibliography: - 'myrefs.bib' title: '**De Sitter Vacua from a D-term Generated Racetrack Uplift**' --- KEK-TH-1755 [**De Sitter Vacua from\ a D-term Generated Racetrack Uplift**]{} [Markus Rummel${}^{1}$ and Yoske Sumitomo${}^{2}$, ]{} 0.6cm ${^1}$Rudolph Peierls Centre for Theoretical Physics, University of Oxford,\ 1 Keble Road, Oxford, OX1 3NP, United Kingdom\ ${}^2$ High Energy Accelerator Research Organization, KEK,\ 1-1 Oho, Tsukuba, Ibaraki 305-0801 Japan\ 0.4cm Email: [markus.rummel at physics.ox.ac.uk, sumitomo at post.kek.jp](mailto:[email protected], [email protected]) 1.0cm Introduction ============ Dark Energy is the dominant source causing the current accelerated expansion of the universe, as has been confirmed by observations [@Riess:1998cb; @Schmidt:1998ys; @Bennett:2012zja; @Ade:2013zuv]. Although there exist some possibilities explaining Dark Energy, a tiny positive cosmological constant would be the prime candidate, in perfect agreement with recent observations [@Bennett:2012zja; @Ade:2013zuv]. If one wants to understand the purely theoretically origin of this cosmological constant, we should promote Einstein gravity to be consistent with its quantum formulation. String theory is quite motivated for this purpose as it is expected to provide the quantum nature of gravity as well as particle physics. A cosmological constant could be realized in the context of flux compactifications [@Giddings:2001yu; @Dasgupta:1999ss] of 10D string theories, where a vacuum expectation value of the moduli potential at minima contributes to the vacuum energy in a four-dimensional space-time universe. Since there exist many possible choices of quantized fluxes and also a number of types of compactifications, the resultant moduli potential including a variety of minima forms the string theory landscape (see reviews [@Douglas:2006es; @Grana:2005jc; @Blumenhagen:2006ci; @Silverstein:2013wua; @Quevedo:2014xia; @Baumann:2014nda]). Although there exist many vacua in the string theory landscape, when we naively stabilize the moduli and obtain the minima, negative cosmological constants seem likely to come by. Hence an ’uplift’ mechanism from the negative vacuum energy keeping stability should be important to realize an accelerated expanding universe. Some possible ways of the uplift mechanism have been proposed in string compactifications. - Explicit SUSY breaking achieved by brane anti-brane pairs contributes positively in the potential, and thus can be used for the uplift [@Kachru:2002gs; @Kachru:2003aw; @Kachru:2003sx]. When the D3 brane anti-brane pairs are localized at the tip of a warped throat, the potential energy may be controllable due to a warping factor. As the uplift term contributes to the potential at ${\cal O}({\cal V}^{-4/3})$, which appears larger than the F-term potential for stabilization which is in general $\mathcal{O}(\ll {\mathcal{V}}^{-2})$, de Sitter ($dS$) vacua with tiny positive cosmological constant may be achieved as a result of tuning of warping. A caveat of this proposal is that the SUSY breaking term needs to compensate the entire Anti-de Sitter ($AdS$) energy, so it is an open question if the SUSY breaking term, originally treated in a probe approximation, can be included as a backreaction in supergravity appropriately. - As an alternative uplift mechanism, one may use the complex structure sector [@Saltman:2004sn]. In the type IIB setup, the complex structure moduli as well as the dilaton are often stabilized at a supersymmetric point. Owing to the no-scale structure, the potential for the complex structure sector is positive definite $V_{\rm c.s.} = e^{K} |DW|_{\rm c.s.}^2 \sim {\cal O} ({\mathcal{V}}^{-2})$. So when we stay at the SUSY loci, the potential is given convex downward in general and thus tractable. However, if one stabilizes the complex structure sector at non-supersymmetric points, then there appears a chance to have a positive contribution in the potential without tachyons, that may be applied for the uplift with a tuning. See also recent applications of this mechanism [@Danielsson:2013rza; @Blaback:2013qza; @Kallosh:2014oja]. - When we include the leading order $\alpha'$-correction coming of ${\cal O}(\alpha'^3)$ in the Kähler potential [@Becker:2002nn] which breaks the no-scale structure, this generates a positive contribution in the effective potential if the Euler number $\chi$ of the Calabi-Yau is given by a negative value [@Balasubramanian:2004uy]. This positive term can balance the non-perturbative terms in the superpotential such that stable $dS$ vacua can be achieved in this Kähler Uplift model [@Westphal:2006tn; @Rummel:2011cd; @deAlwis:2011dp] (see also [@Westphal:2005yz]). In the simplest version of the Kähler uplifting scenario, there is an upper bound on the overall volume of the Calabi-Yau such that one may worry about higher order $\alpha'$-corrections. However, this bound can be significantly relaxed when embedded in a racetrack model [@Sumitomo:2013vla]. - It has been proposed that the negative curvature of the internal manifold may be used for $dS$ constructions as it contributes positive in the scalar potential [@Silverstein:2007ac]. Motivated by this setup, there were many attempts constructing $dS$ vacua [@Haque:2008jz; @Flauger:2008ad; @Danielsson:2009ff; @Caviezel:2009tu; @deCarlos:2009fq; @deCarlos:2009qm; @Dong:2010pm; @Andriot:2010ju; @Danielsson:2010bc; @Danielsson:2011au; @Danielsson:2012et; @Danielsson:2012by; @Blaback:2013ht]. Using the necessary constraint for the extrema [@Hertzberg:2007wc; @Wrase:2010ew] and for the stability [@Shiu:2011zt], we see that the existence of minima requires not only negative curvature, but also the presence of orientifold planes. - When the stabilization mechanism does not respect SUSY, D-terms can provide a positive contribution to the potential if the corresponding D7-brane is magnetically charged under an anomalous $U(1)$ [@Burgess:2003ic; @Cremades:2007ig; @Krippendorf:2009zza]. The potential of D-terms arises of order ${\cal O} ( {\cal V}^{-n})$ with $n\geq 2$, depending on the cycle that the D7-brane wraps. If we take into account the stabilization of matter fields having a non-trivial VEV, originating from fluxed D7-branes wrapping the large four-cycle, then the uplift contribution becomes ${\cal O} ({\cal V}^{-8/3})$ [@Cremades:2007ig]. So a relatively mild suppression, for instance by warping, is required for this volume dependence. See also recent applications to explicit scenarios in [@Cicoli:2012vw; @Cicoli:2013mpa] and also together with a string-loop correction in fibred Calabi-Yaus [@Cicoli:2011yh]. - Recently, it has been proposed that a dilaton-dependent non-perturbative term can also work for the uplift mechanism toward $dS$ vacua [@Cicoli:2012fh]. The non-perturbative term depends on both the dilaton and a vanishing blow-up mode which is stabilized by a D-term. Since the D-term turns out to be trivial at the minima due to a vanishing cycle, the non-trivial dilaton as well as the vanishing cycle dependence generate the uplift term within the F-term potential. In this setup, the given uplift term is proportional to $e^{-2b \langle s \rangle}/{\cal V}$. Although the volume does not suppress the uplift term so much, we may expect an exponential fine-tuning by the dilaton dependence to balance with the moduli stabilizing F-term potential. In this paper, we introduce an uplifting term of the form $e^{-a_s \tau_s}/{\cal V}^2$, where $\tau_s$ is the volume of a small 4-cycle, which naturally balances with the stabilizing F-term potential in the Large Volume Scenario (LVS). The following ingredients are necessary for this mechanism: - one non-perturbative effect on a 4-cycle $D_2$ to realize the standard LVS moduli stabilization potential, - another non-perturbative effect on a different cycle $D_3$, - a D-term constraint that enforces the volumes $\tau$ of the two 4-cycles to be proportional $\tau_s \sim \tau_2 \propto \tau_3$ via a vanishing D-term potential. Hence, the minimal number of Kähler moduli for this uplifting scenario is $h^{1,1}_+=3$. At the level of the F-term potential the effective scalar potential reduces to the standard LVS moduli stabilization potential plus the mentioned uplifting term yielding metastable $dS$ vacua. The Kähler moduli are stabilized at large values avoiding dangerous string- and $\alpha'$-corrections. Compared to [@Cicoli:2012fh], the dilaton can take rather arbitrary values determined by fluxes as there is no tuning required to keep the uplifting term suppressed. Note that for $h^{1,1}_+=2$, a racetrack setup with two non-perturbative effects on one small cycle does not allow stable $dS$ vacua in the LVS. Hence, we have to consider at least two small cycles and a relating D-term constraint to construct $dS$ vacua. We also have to consider a necessary condition for coexistence of the vanishing D-term constraint with the non-perturbative terms in superpotential. If the rigid divisors for the two non-perturbative effects intersect with the divisor on which the D-term constraint is generated via magnetic flux, we have to worry whether the VEV of matter fields that are generated by this magnetic flux are given accordingly such that the coefficients of non-perturbative terms remain non-zero. On the other hand, we may avoid additional zero mode contributions in a setup with minimal intersections. A general constraint is that the non-perturbative effects and D-term potential have to fulfill all known consistency condition, for instance requiring rigid divisors, avoiding Freed-Witten anomalies [@Minasian:1997mm; @Freed:1999vc] and saturating D3, D5, and D7 tadpole constraints. We expect these constraints to become less severe as the number of Kähler moduli increases beyond $h^{1,1}_+=3$ as in principle the degrees of freedom such as flux choices and rigid divisors increases. This paper is organized as follows. We illustrate the uplift proposal generated through the multi-Kähler moduli dependence in the F-term potential and the required general geometric configuration in Section \[genmech\_sec\] and give some computational details in Appendix \[app\_fluxconstraint\]. We further discuss the applicability of the uplift mechanism in more general Swiss-Cheese type Calabi-Yau manifolds in Section \[sec:real-more-moduli\]. D-term generated racetrack uplift - general mechanism {#genmech_sec} ===================================================== We illustrate the uplift mechanism by a D-term generated racetrack in Calabi-Yaus with the following properties: there are two small 4-cycles and two linear combinations of these small cycles that are rigid such that the existence of two non-perturbative terms is guaranteed in the superpotential avoiding additional fermionic zero modes from cycle deformations or Wilson lines. We show that this setup in general allows to stabilize the moduli in a $dS$ vacuum at large volume. Geometric setup and superpotential ---------------------------------- We consider an orientifolded Calabi-Yau $X_3$ with $h^{1,1}_+\geq3$ with the following general volume form of the divisors $D_i$ $${\mathcal{V}}= \frac16 \left(\sum_{i,j,k=1}^{h^{1,1}_+}\kappa_{i,j,k} t_i t_j t_k \right)\,,\label{gen2cycles}$$ in terms of 2-cycle volumes $t_i$ and intersection numbers $$\kappa_{ijk} = \int_{X_3} D_i \wedge D_j \wedge D_k\,.$$ The 4-cycle volumes are given as $$\tau_i = \frac{\partial {\mathcal{V}}}{\partial t_i} = \frac12 \kappa_{ijk} t_j t_k\,.\label{tauandt}$$ We assume that $X_3$ has a Swiss-Cheese structure with a big cycle named $D_a$ and at least two small cycles $D_b$ and $D_c$, i.e., its volume form can be brought to the form $${\mathcal{V}}= \gamma_a \tau_a^{3/2} - \gamma_b \tau_b^{3/2} - \gamma_c \tau_c^{3/2} - {\mathcal{V}}_{\text{rest}}\,,$$ with ${\mathcal{V}}_{\text{rest}}$ parametrizing the dependence of the volume on the remaining $h^{1,1}_+-3$ moduli. Now let us assume there are two rigid divisors $D_2$ and $D_3$ of which a linear combination forms the small cycles $D_b$ and $D_c$. $$\begin{aligned} \begin{aligned} D_2 &= d_{2b} D_b + d_{2c} D_c\,,\\ D_3 &= d_{3b} D_b + d_{3c} D_c\,. \end{aligned}\label{tau2taua}\end{aligned}$$ Even if there do not exist two divisors $D_2$ and $D_3$ that are rigid, one might still be able to effectively ‘rigidify’ one or more divisors by fixing all the deformation moduli of the corresponding D7-brane stacks via a gauge flux choice [@Bianchi:2011qh; @Cicoli:2012vw; @Louis:2012nb]. Under these assumptions, the superpotential in terms of the Kähler moduli $T_i = \tau_i + i\, \zeta_i$ is of the form $$W = W_0 + A_2 e^{-a_2 T_2} + A_3 e^{-a_3 T_3} = W_0 + A_2 e^{-a_2 \left(d_{2b} T_b + d_{2c} T_c \right)} + A_3 e^{-a_3 \left(d_{3b} T_b + d_{3c} T_c \right)}\,,\label{Wnonpert}$$ with non-zero $A_2$, $A_3$ and $W_0$ being the Gukov-Vafa-Witten flux superpotential [@Gukov:1999ya]. D7-brane and gauge flux configuration {#d7config3moduli_sec} ------------------------------------- The orientifold plane O7 induces a negative D7 charge of $-8[O7]$ that has to be compensated by the positive charge of D7-branes. In general the O7 charge can be cancelled by introducing a Whitney brane with charge $8[O7]$ [@Collinucci:2008pf]. The non-perturbative effects of  can be either generated by ED3-instantons or gaugino condensation. For the latter, we choose a configuration with $N_2$ D7-branes on $D_2$ and $N_3$ D7-branes on $D_3$. In this case, the exponential coefficients of the non-perturbative terms in  are $a_2 = 2 \pi / N_2$ and $a_3 = 2 \pi / N_3$. The corresponding gauge group is either $SO(N)$ or $Sp(N)$ (which becomes $SU(N)$ if gauge flux is introduced), depending on if the divisor lies on the orientifold plane or not. Furthermore we introduce a third stack of $N_D$ branes on a general linear combination $D_D$ of basis divisors that is not either $D_2$ or $D_3$. This stack will introduce a D-term constraint that reduces the F-term effective scalar potential by one degree of freedom/ Kähler modulus. In the case of $h^{1,1}_+=3$ this corresponds to a two Kähler moduli LVS potential plus an uplift term that allows $dS$ vacua as we will show in Section \[effectiveVgen\_sec\].[^1] Note that in general all required D7-brane stacks have to be consistent with possible factorizations of the Whitney brane that cancels the O7 charge [@Collinucci:2008pf; @Cicoli:2011qg]. The D-term constraint is enforced via a Fayet-Illiopoulos (FI) term $$\xi_D = \frac{1}{{\mathcal{V}}} \int_{X_3} D_D \wedge J \wedge \mathcal{F}_D = \frac{1}{{\mathcal{V}}} q_{Dj} t_j\,,\label{FID}$$ where $J=t_i D_i$ is the Kähler form on $X_3$ and $q_{Dj} = \tilde f^k_D \kappa_{Djk}$ is the anomalous $U(1)$-charge of the Kähler modulus $T_j$ induced by the magnetic flux $\mathcal{F}_D = \tilde f^k_D D_k$ on $D_D$. We choose flux-quanta $\tilde f^k_D$ such that $\xi_D = 0$ in  implies $$\tau_c = c\, \tau_b\,,\label{tauctaubprop}$$ with a constant $c$ depending on flux quanta and triple intersection numbers. In a concrete example it is important to check that a constant $c$ in  is realized which is consistent with stabilizing the moduli inside the Kähler cone of the manifold. An important constraint arises from the requirement of two non-vanishing non-perturbative effects $A_2, A_3\neq 0$ on generally intersecting cycles $D_2$ and $D_3$. The cancellation of Freed-Witten anomalies requires the presence of fluxes $\mathcal{F}$ on the D7-branes wrapping these divisors that can potentially forbid the contribution from gaugino condensation in the superpotential. This gauge invariant magnetic flux $\mathcal{F}$ is determined by the gauge flux $F$ on the corresponding D7-brane and pull-back of the bulk $B$-field on the wrapped four-cycle via $$\mathcal{F} = F - B\,.$$ If $D_2$ and $D_3$ intersect each other, the $B$-field can in general not be used to cancel both of theses fluxes to zero. However, it is still possible that both fluxes $\mathcal{F}_2$ and $\mathcal{F}_3$ can be chosen to be effectively trivial, such that no additional zero modes and FI-terms are introduced. These zero-modes would be generated via charged matter fields arising at the intersection of D7-brane stacks or from the bulk D7 spectrum. The constraint has to be checked on a case-by-case basis. We work out a sufficient condition on the intersections $\kappa_{ijk}$ for $\mathcal{F}_2$ and $\mathcal{F}_3$ to be trivial for the case of $D_2$ and $D_3$ not intersecting any other divisors $\kappa_{2,j,k} = \kappa_{3,j,k} =0$ for $j,k\neq 2,3$ in Appendix \[app\_fluxconstraint\]. Furthermore, it has to be checked that $\mathcal{F}_D$ does not generate any additional zero-modes at the intersections of $D_D$ with $D_2$ and $D_3$. Finally, the chosen D7-brane and gauge flux setup has to be consistent with D3, D5 and D7 tadpole cancellation. As for every explicit construction this has to be checked on a case-by-case basis for the particular manifold under consideration. We do not expect tadpole cancellation to be in general more restrictive than in e.g., the $AdS$ LVS. In particular, we do not require a large number of D7-branes [@Westphal:2006tn] and/or racetrack effect [@Sumitomo:2013vla] on a particular single divisor to achieve a large volume as in the Kähler Uplifting scenario. Effective potential of the Kähler moduli {#effectiveVgen_sec} ---------------------------------------- We start with a slightly simplified model where the F-term potential $$V_F = e^{K} \left( K^{\alpha\bar{\beta}} D_\alpha W \overline{D_\beta W} - 3 |W|^2 \right)\,,\label{VF}$$ is given by $$\begin{split} K=& - 2 \ln \left({\cal V} + {\xi \over 2}\right), \quad {\cal V} = (T_a + \bar{T}_a)^{3/2} - (T_b + \bar{T_b})^{3/2} - (T_c + \bar{T_c})^{3/2},\\ W =& W_0 + A_2 e^{-a_2 T_b} + A_3 e^{-a_3(T_b + T_c)}, \end{split} \label{F-term potential-simplified}$$ where we have used equal intersection numbers $\gamma$ and assumed stabilization of the dilaton and complex structure moduli via fluxes [@Giddings:2001yu]. The values of these parameters are not essential for the uplift dynamics we illustrate in this paper. The superpotential in  corresponds to a particular choice of the general linear combination in . The model  is known to include the solutions of the LVS [@Balasubramanian:2005zx] that stabilizes the moduli in a non-supersymmetric way in the presence of the leading $\alpha'$-correction [@Becker:2002nn] and one non-perturbative term. The $\alpha'$-correction is given by $\xi \propto - \chi g_s^{-3/2}$ where $\chi$ is the Euler number of the Calabi-Yau manifold.[^2] The D-term potential is given through the magnetized D7-branes wrapping the Calabi-Yau divisor $D_i$ [@Haack:2006cy]: $$V_D = {1 \over \operatorname*{Re \,}(f_D)} \left(\sum_j c_{Dj} \hat{K}_j \varphi_j - \xi_D\right)^2\,, \label{D-term potential}$$ where the gauge kinetic function $$\text{Re}\,(f_D) = \frac12\, \int_{D_D} J \wedge J - \frac{1}{2g_s}\int_{D_D} \mathcal{F}_D \wedge \mathcal{F}_D \,,\label{fD}$$ and $\varphi_j$ are matter fields associated with the diagonal $U(1)$ charges $c_{Dj}$ of a stack of D7-branes and the FI-term $\xi_D$ is defined in . Now we redefine the coordinates: $$T_s \equiv {1\over 2 }\left(T_b + T_c \right), \quad Z \equiv {1\over 2} \left(T_b - T_c \right).$$ When the D7-branes wrapping the divisor $D_D$ are magnetized and the matter fields are stabilized either at $\langle \varphi_i \rangle = 0$ or satisfying $ \langle \sum c_{ij} \hat{K}_j \varphi_j \rangle =0$, the D-term potential may become $$V_D \propto {1\over \operatorname*{Re \,}(f_D)}{1 \over {\cal V}^2} \left(\sqrt{\tau_b} - \sqrt{\tau_c} \right)^2\,, \label{D-term potential in simple model}$$ using $\xi_D \propto \sqrt{\tau_b} - \sqrt{\tau_c}$ implied by the flux $\mathcal{F}_D$, see  where we use $c=1$ for simplicity. In the large volume limit, the F-term potential generically scales as ${\cal O} ({\cal V}^{-3})$ in the minima given in the LVS model. Stabilizing the Kähler moduli at ${\cal O} ({\cal V}^{-3})$ then requires a vanishing D-term potential, i.e., $\tau_b = \tau_c$ corresponding to $z \equiv \operatorname*{Re \,}Z=0$. Thanks to the topological coupling to the two-cycle supporting magnetic flux, the imaginary mode of the $Z$ modulus is eaten by a massive $U(1)$ gauge boson through the Stückelberg mechanism. Since the gauge boson has a mass of order of the string scale ${\cal O} ({\cal V}^{-1/2})$, the degree of freedom of $\operatorname*{Im \,}Z$ charged under the anomalous $U(1)$ as well as the gauge boson is integrated out at the high scale. Hence, we are left with the stabilization of the remaining moduli fields by the F-term potential. F-term uplift\[sec:f-term-uplift\] ---------------------------------- Next we will consider the stabilization by the F-term potential given in (\[F-term potential-simplified\]). We are interested in LVS like minima ${\cal V} \sim e^{\hat{a}_i \tau_i}$ realizing an exponentially large volume. Then the leading potential of order ${\cal V}^{-3}$ is given by $$\begin{split} V & \sim {3 W_0^2 \xi \over 4 {\cal V}^3} + {2 W_0 \over {\cal V}^2} \left( a_2 A_2 {\tau_b} e^{-a_2 \tau_b/2} + a_3 A_3 (\tau_b + \tau_c) e^{-a_3 (\tau_b + \tau_c)/2}\right) \\ & + {2 \over 3 {\cal V}} \left(a_2^2 A_2^2 \sqrt{\tau_b} e^{- a_2 \tau_b} + a_3^2 A_3^2 (\sqrt{\tau_b}+\sqrt{\tau_c}) e^{-a_3 (\tau_b + \tau_c)} + 2 a_2 a_3 A_2 A_3 \sqrt{\tau_b} e^{-a_2 \tau_b/2 - a_3 (\tau_b + \tau_c)/2} \right)\,, \end{split} \label{effective potential at larger volume 0}$$ where the imaginary directions are stabilized at $\operatorname*{Im \,}T_i= 0$, and $\operatorname*{Im \,}T_a$ is stabilized by non-perturbative effects that are omitted in (\[F-term potential-simplified\]), inducing a very small mass for $\operatorname*{Im \,}T_a$ and with negligible influence on the stabilization of the other moduli. Although the general minima of $\operatorname*{Im \,}T_i$ are given by $a_i \operatorname*{Im \,}T_i = m_i \pi$ with $m_i \in {\mathbb Z}$, the different solutions just change the sign of $A_i$ and thus we can simply have the potential of the above form. As the D-term stabilizes $\tau_c = \tau_b$, the resultant potential becomes $$\begin{split} V &\sim {3 W_0^2 \xi \over 4 {\cal V}^3} + {4 W_0 \over {\cal V}^2} \left(a_2 A_2 \tau_s e^{-a_2 \tau_s} + 2 a_3 A_3 \tau_s e^{-2 a_3 \tau_s}\right)\\ &+ {2 \sqrt{2} \over 3 {\cal V}} \left({a_2^2 A_2^2 \sqrt{\tau_s}} e^{-2 a_2 \tau_s} + {4 a_3^2 A_3^2 \sqrt{\tau_s}} e^{-4 a_3 \tau_s} + 2 a_2 a_3 A_2 A_3 \sqrt{\tau_s} e^{-(a_2 + 2 a_3)\tau_s} \right)\,, \end{split} \label{effective potential at larger volume with ts}$$ where we have defined $\tau_s = \operatorname*{Re \,}T_s$. One may consider that this form of the potential looks similar to the racetrack type. Although cross terms of $A_2, A_3$ appear due to the $T_b$ dependence, the important point for the uplift mechanism demonstrated in this paper is that the cross terms between $T_b$ dependence of the $A_2$ term and $T_c$ dependence of the $A_3$ term appear at ${\cal O} ({\cal V}^{-4})$.[^3] If the cross term appears at ${\cal O}({\cal V}^{-3})$, it disturbs uplifting to $dS$. We further redefine the fields and parameters such that there are no redundant parameters affecting the stabilization: $$\begin{split} &x_s = a_2 \tau_s, \quad {\cal V}_x = {\cal V} a_2^{3/2},\\ &c_i = {A_i \over W_0},\quad \xi_x = \xi a_2^{3/2}, \quad \beta = {2 a_3 \over a_2}. \end{split} \label{redefined parameters}$$ Then the effective potential at order ${\cal O}({\cal V}^{-3})$ becomes $$\begin{split} \hat{V} \equiv \left({a_2^{-3} W_0^{-2}}\right) V \sim {3\xi_x \over 4 {\cal V}_x^3} + {4 c_2 x_s \over {\cal V}_x^2}e^{-x_s} + {2\sqrt{2} c_2^2 x_s^{1/2}\over 3 {\cal V}_x} e^{-2 x_s } + {4 \beta c_3 x_s \over {\cal V}_x^2} e^{-\beta x_s}. \end{split} \label{redefined potential}$$ We have neglected the term proportional to $c_3^2$ and the cross term between $c_2, c_3$ in the expression above. In fact, these terms are not important for the uplift mechanism of our interest, and we will justify this assumption a posteriori later. Since the uplift term comes together with $e^{-\beta x_s}$, this term contributes of ${\cal O}({\cal V}_x^{-3})$ when $\beta \sim {\cal O} (1)$. Hence, it contributes at the same order as the stabilizing F-term potential and no suppression factor provided by warping or dilaton dependence is required. Before performing the uplift, we consider the LVS solution by setting $c_3 =0$. We use a set of parameters: $$c_2 = -0.01, \quad \xi_x = 5. \label{test input parameter}$$ The extremal equations $\partial_I \hat{V} = 0$ at $c_3=0$ can be simplified as $$\xi_x = {64 \sqrt{2} (x_s-1) x_s^{5/2} \over (4 x_s -1)^2}, \quad c_2 = - {6\sqrt{2} (x_s-1) x_s^{1/2} \over (4 x_s -1) {\cal V}_x} e^{-x_s}.$$ Solving the equations above, we obtain $${\cal V}_x \sim 467, \quad x_s \sim 1.50.\label{VxxsAds}$$ We can easily check that this solution gives an $AdS$ vacuum. Note that when we have just two moduli fields ${\cal V}_x, x_s$ in the LVS, the positivity of $\xi_x$ automatically guarantees the stability of the minima since the required condition $x_i > 1$ is satisfied (see e.g. [@Rummel:2013yta]). Now we consider non-zero $c_3$ for the uplift. As $c_3$ increases, the vacuum energy of the potential minimum increases and eventually crosses the Minkowski point. In Figure \[fig:uplift-illustration\], we illustrate the behavior of the minimum point by changing the value of $c_3$. Interestingly, the volume increases as the vacuum energy increases, suggesting that the effective description of the theory will be more justified toward $dS$ vacua. On the other hand, the minimum value of the Hessian decreases. Destabilization occurs when the uplift term dominates the entire potential. As this happens at higher positive values of the cosmological constant, there certainly exist a range of parameters yielding stable $dS$ vacua within this setup. As a reference, we show numerical values of parameters close to crossing the Minkowski point. When we use $$\beta = {5\over 6}, \label{test beta value}$$ the minimum reaches Minkowski at $$c_3 \sim 4.28 \times 10^{-3}, \quad {\cal V}_x \sim 3240, \quad x_s \sim 3.07.$$ So we see that the volume increases quite drastically from the $AdS$ vacuum . Since $c_3$ remains small compared to the input value of $c_2$, we see that our approximation neglecting the term proportional to $c_3^2$ is justified. ![Illustration of the D-term generated racetrack uplift mechanism. We plot the cosmological constant $\hat \Lambda$ vs $c_3$, $\text{min}(\partial^2 \hat V)$ and ${\mathcal{V}}_x$ at the minima of the potential, especially near the Minkowski point.[]{data-label="fig:uplift-illustration"}](lambda-c3.pdf "fig:"){width="20.5em"} ![Illustration of the D-term generated racetrack uplift mechanism. We plot the cosmological constant $\hat \Lambda$ vs $c_3$, $\text{min}(\partial^2 \hat V)$ and ${\mathcal{V}}_x$ at the minima of the potential, especially near the Minkowski point.[]{data-label="fig:uplift-illustration"}](lambda-ddv.pdf "fig:"){width="20em"} ![Illustration of the D-term generated racetrack uplift mechanism. We plot the cosmological constant $\hat \Lambda$ vs $c_3$, $\text{min}(\partial^2 \hat V)$ and ${\mathcal{V}}_x$ at the minima of the potential, especially near the Minkowski point.[]{data-label="fig:uplift-illustration"}](lambda-vol.pdf "fig:"){width="20.5em"} In fact, it is not difficult to see how these values change when the presence of $c_3^2$ terms and cross terms between $c_2,c_3$ in the potential (\[effective potential at larger volume with ts\]) are taken into account. With the input parameters we used in (\[test input parameter\]), the Minkowski vacuum is obtained when $$c_3 \sim 5.11 \times 10^{-3}, \quad {\cal V}_x \sim 2860, \quad x_s \sim 2.61. \label{test solution with c_3^2}$$ Since the obtained values are not significantly different from the case where $c_3^2$ terms and cross terms between $c_2, c_3$ are neglected, we conclude a posteriori that the uplift term is dominated by the term linear in $c_3$. Let us comment on the stabilization of the axionic partner of each modulus field. As stated, the imaginary mode of $Z$ is eaten up by a massive gauge boson and hence integrated out at the high scale. The axionic partner of the big divisor $T_a$ is stabilized by non-perturbative effects yielding a tiny mass. The remaining modulus $\operatorname*{Im \,}T_s$ is stabilized by the F-term potential as the D-term potential does not depend on the latter. In the approximated potential up to ${\cal O}({\mathcal{V}}^{-3})$, the Hessian of $ y_s = a_2 \operatorname*{Im \,}T_s$ is $$\partial_{y_s}^2 \hat{V}|_{\rm ext} \sim 5.14 \times 10^{-10}\,,$$ where we have included $c_3^2$ and $c_2, c_3$ cross terms, and used the solution (\[test solution with c\_3\^2\]) and $\operatorname*{Im \,}T_{a} = \operatorname*{Im \,}T_i =0$. Thus all Kähler moduli are stabilized. Analytical estimate ------------------- It is difficult to analytically derive a generic condition for the D-term generated racetrack uplift since the formulas are still complicated enough even after using several approximations. However, some of the expressions can be simplified under an additional reasonable approximation. In this subsection, we illustrate some analytical analyses for a better understanding. Since we checked that the uplift mechanism works even at linear approximation of the uplift parameter $c_3$, we only keep terms up to linear order in $c_3$ and neglect the higher order terms including cross terms. The extremal condition $\partial_i \hat{V} = 0$ of the potential (\[redefined potential\]) is now simplified by $$\begin{split} &c_2 \sim -{6\sqrt{2x_s} (x_s-1) \over 4 x_s -1} {e^{x_s} \over {\cal V}_x} + c_3 {\beta (\beta x_s -1) \over x_s -1} e^{(1-\beta)x_s} ,\\ &\xi_x \sim {64\sqrt{2} x_s^{5/2} (x_s-1) \over (4x_s-1)^2} - c_3 {32 \beta x_s^2 \left(2(\beta+2)x_s + \beta-7 \right) \over 9 (x_s-1) (4 x_s-1)} e^{-\beta x_s} {\cal V}_x . \end{split} \label{extremal condition of effective potential}$$ Although our interest is the uplift toward $dS$ vacua, we have to cross the Minkowski point along the way. Thus, the condition that the minimum structure holds when uplifted to Minkowski vacua is a necessary condition for the $dS$ uplift mechanism. The condition for Minkowski at the extrema $\hat{V}|_{\rm ext} = 0$ reads $$\begin{split} c_3 \sim {18 \sqrt{2} (\beta-1) x_s^{3/2} \over \beta (4 (\beta-1) x_s - 3)^2} {e^{\beta x_s} \over {\cal V}_x}. \end{split} \label{minkowski condition}$$ Next, we proceed to check the stability at the Minkowski point. Although we know the conditions to check the stability, the formula of the Hessian is yet too complicated to perform an analytical analysis. So we further focus on the region satisfying $x_s \gg 1$. The region with $x_s \gg 1$ is motivated since the $AdS$ minimum points, before adding an uplift term, are guaranteed to have a positive Hessian since all eigenvalues are positive definite when satisfying $x_s > 1$ in LVS type stabilizations (see e.g. [@Rummel:2013yta]). Furthermore, higher instanton corrections can be safely neglected. As shown in Figure \[fig:uplift-illustration\], the minima can be uplifted keeping the positivity of the Hessian until reaching the destabilization point with a relatively high positive vacuum energy. Hence, having $x_s \gg 1$ is motivated to see the basic feature of the D-term generated racetrack uplift mechanism. Since there is no reason to take $\beta$ to be small/large, we consider $\beta \sim {\cal O} (1)$. When we use the approximation $x_s \gg 1$, a component of the Hessian and the determinant at the extrema become (\[extremal condition of effective potential\]) $$\begin{split} \partial_{{\cal V}_x}^2 \hat{V}|_{\rm ext} \sim& {6\sqrt{2} x_s^{3/2} \over {\cal V}_x^5} - c_3 {8 \beta x_s e^{-\beta x_s} \over {\cal V}_x^4} \sim {6\sqrt{2} x_s^{3/2} \over {\cal V}_x^5},\\ \det \left(\partial_i \partial_j \hat{V} \right)|_{\rm ext} \sim& {162 x_s^2 \over {\cal V}_x^8} + c_3 {24\sqrt{2} \beta(\beta^2 + \beta -2) x_s^{5/2} e^{-\beta x_s} \over {\cal V}_x^7} \sim {54 (1- \beta) x_s^2 \over {\cal V}_x^8}\,, \end{split} \label{analytical hessian at minkowski}$$ where in the last step of both equations, we have used the Minkowski condition (\[minkowski condition\]). According to Sylvester’s criterion, the positivity of a matrix can be checked by the positivity of the determinant of all sub-matrices. Thus it is enough to check the positivity of the quantities in (\[analytical hessian at minkowski\]). Therefore we conclude that the stability at the Minkowski point requires $\beta < 1$. This condition is clearly satisfied in the previous numerical example following (\[test beta value\]), which may justify the crude approximations we took in this subsection. Note that the Hessian of the imaginary mode is guaranteed to be positive under the above used approximations: $$\partial_{y_s}^2 \hat{V}|_{\rm ext} \sim {6\sqrt{2} x_s^{3/2} \over {\cal V}_x^3} - c_3 {4 \beta^2 (\beta +1) x_s e^{-\beta x_s} \over {\cal V}_x^2} \sim {6\sqrt{2} x_s^{3/2} \over {\cal V}_x^3}.$$ Finally, let us check the extremal and Minkowski conditions in the limit $x_s \gg 1$. Now all conditions are simplified to be $$\begin{split} \xi_x \sim 4 \sqrt{2} x_s^{3/2}, \quad c_2 \sim -{3 \sqrt{x_s} \over \sqrt{2}} {e^{x_s} \over {\cal V}_x}, \quad c_3 \sim {9 \over 4 \sqrt{2x_s} \beta (1-\beta)} {e^{\beta x_s} \over {\cal V}_x}. \end{split}$$ We see that the minimum point needs $\xi_x >0$ and $c_2 < 0$ in agreement with the minimum requirement of the two-moduli LVS at $AdS$. The stability condition $\beta < 1$ suggests $c_3 > 0$. In fact, the extremal condition for $\xi_x, c_2$ is simply the leading order approximation of each first term in as the $c_3$ contribution appears sub-dominant. This justifies that the linear approximation for $c_3$ is compatible with $x_s \gg 1$. Hence, we can regard the last term in the potential (\[redefined potential\]) as the uplift term. On realization in models with more moduli\[sec:real-more-moduli\] ================================================================= In this section, we show that the uplift mechanism works well in the presence of additional Kähler moduli in Swiss-Cheese type Calabi-Yau compactifications. We consider a simple toy model with $h^{1,1}_+=4$, which captures the essential features of the D-term generated racetrack uplift mechanism defined by $$\begin{split} K=& - 2 \ln \left({\cal V} + {\xi \over 2}\right), \quad {\cal V} = (T_a + \bar{T}_a)^{3/2} - (T_b + \bar{T_b})^{3/2} - (T_c + \bar{T_c})^{3/2} - (T_e + \bar{T_e})^{3/2},\\ W =& W_0 + A_2 e^{-a_2 T_b} + A_3 e^{-a_3 (T_b+T_c)} + A_4 e^{- a_4 T_e}. \end{split}$$ Again we are interested in the case of a Swiss-Cheese volume for moduli stabilization of the LVS type. Note that we used the name $T_e$ to avoid confusion with $T_D$. Taking into account the D-term potential generated by the magnetized D7-branes wrapping the divisor $Z$, we assume again that the $Z =\frac12 (T_b - T_c)$ modulus is stabilized at $Z=0$. Setting $a_4 = a_2$ for simplicity, the effective potential at ${\cal V}^{-3}$ from the F-terms is given by $$\begin{split} {\hat{V}} \equiv \left({a_2^{-3} W_0^{-2}}\right) V \sim& {3\xi_x \over 4 {\cal V}_x^3} + {4 c_2 x_s \over {\cal V}_x^2}e^{-x_s} + {2\sqrt{2} c_2^2 x_s^{1/2}\over 3 {\cal V}_x} e^{-2 x_s} + {4 c_4 x_4 \over {\cal V}_x^2}e^{-x_4} + {2\sqrt{2} c_4^2 x_4^{1/2}\over 3 {\cal V}_x} e^{-2 x_4}\\ &+ {4 \beta c_3 x_s \over {\cal V}_x^2} e^{-\beta x_s} + {\sqrt{2} c_3^2 x_s^{1/2}\over 3 {\cal V}_x} e^{-2 \beta x_s} + {2\sqrt{2}\beta c_2 c_3 x_s^{1/2} \over 3 {\cal V}_x} e^{-(1+\beta)x_s}, \end{split}$$ where we have further defined $x_e = a_4 \tau_e, \beta=2a_3/a_2$ and $c_4 = A_4/ W_0$ in addition to . Here we included the term proportional to $c_3^2$ as well as the cross term $c_2 c_3$ even though they are potentially subleading. When we use a set of parameters: $$c_2 = -0.0114, \quad c_4 = -3.38 \times 10^{-4}, \quad \xi_x = 19, \label{input of 4 moduli model}$$ then the $AdS$ LVS minimum at $c_3=0$ is located at $$\begin{split} {\cal V}_x \sim 2740, \quad x_s \sim 2.60, \quad x_e \sim 1.12. \end{split} \label{4modminval}$$ The stability of multi-Kähler moduli models of the LVS type is ensured if the constraint $x_i>1$ is satisfied [@Rummel:2013yta]. Hence, the extremal point  is stable. Now we add the uplift terms $c_3\neq 0$ and $\beta=5/6$. The minimum with the input parameters (\[input of 4 moduli model\]) reaches Minkowski at $$\begin{split} c_3 \sim 4.55 \times 10^{-3}, \quad {\cal V}_x \sim 5.64 \times 10^4, \quad x_s \sim 5.45, \quad x_e \sim 2.26. \end{split}$$ Although the volume is drastically changed during the uplift toward $dS$ vacua, we can check the stability of the minimum by plugging the values into the Hessian, similarly to the simple three moduli model. The cosmological constant can further increase in the positive region keeping the stability until the minima exceeds the potential barrier where decompactification happens. Discussion ========== We have proposed an uplift mechanism using the structure of at least two small Kähler moduli $T_b$ and $T_c$ in Swiss-Cheese type compactifications. The uplift contribution arises as an F-term potential when using a D-term condition which fixes $\operatorname*{Re \,}T_b = c \operatorname*{Re \,}T_c$ at a higher scale, where $c$ is determined by magnetized fluxes on D7-branes. The uplift term becomes of the form $e^{-a_s \tau_s}/{\cal V}^2$ at large volumes, and hence it can naturally balance with the stabilizing potential in the Large Volume Scenario (LVS), without requiring suppressions in the coefficient, for instance, by warping or a dilaton dependent non-perturbative effect. In addition, we have shown that the D-term generated racetrack uplift works in the presence of additional Kähler moduli. Together with the fact that constraints on the uplift parameters are rather relaxed, i.e., $\beta< 1$ and $c_3>0$, this makes us optimistic that there should be many manifolds admitting the proposed uplift mechanism. Since the proposed uplift mechanism requires certain conditions for a D-term constraint and two non-vanishing non-perturbative effects, it should be interesting if we can construct an explicit realization of this model in a particular compactification. Such an explicit construction requires to match all known consistency conditions such as cancellation of Freed-Witten anomalies and cancellation of the D3, D5, and D7 tadpole [@Cicoli:2011qg; @Cicoli:2012vw; @Louis:2012nb; @Cicoli:2013mpa; @Cicoli:2013zha]. We hope to report on an explicit example in another paper. Furthermore, the phenomenological aspect of the proposed uplift mechanism should be interesting. Even though the moduli are essentially stabilized as in the LVS, the resultant behavior of the mass spectrum and/or soft SUSY breaking terms may be different depending on which uplift mechanism we employ to realize the $dS$ vacuum. Finally, in this paper, we concentrated on analyzing the structure of $dS$ minima. However, the structure of the potential is also changed by the uplift term in regions that might be important for including inflationary dynamics. We relegate the analysis of possible inflation scenarios as well as phenomenological consequences compared to other uplift proposals to future work. Acknowledgments {#acknowledgments .unnumbered} =============== We would like to thank Joseph P. Conlon and Roberto Valandro for valuable discussions and important comments, and also the organizers of the workshop “String Phenomenology 2014” held at ICTP Trieste, Italy, where some of the results of this paper were presented. YS is grateful to the Rudolph Peierls Centre for Theoretical Physics, University of Oxford where part of this work was done for their hospitality and support. MR is supported by the ERC grant ‘Supersymmetry Breaking in String Theory’. This work is partially supported by the Grant-in-Aid for Scientific Research (No. 23244057) from the Japan Society for the Promotion of Science. Conditions for avoiding additional zero-modes {#app_fluxconstraint} ============================================= In this appendix, we give a sufficient condition on the intersections $\kappa_{ijk}$ for $\mathcal{F}_2$ and $\mathcal{F}_3$ to be trivial for the case of $D_2$ and $D_3$ not intersecting any other divisors $\kappa_{2,j,k} = \kappa_{3,j,k} =0$ for $j,k \neq 2,3$. This is a necessary condition for the non-perturbative effects on $D_2$ and $D_3$ to contribute to the superpotential, which is crucial for the uplift mechanism considered in this work. In order to avoid Freed-Witten anomalies the gauge flux on the D7-branes has to satisfy $$F + \frac{c_1(D)}{2} \in H^2(X_2,\mathbb{Z})\,.$$ In particular, if $D$ is non-spin, i.e., $c_1(D)$ is odd, $F$ is always non-zero. Using $c_1(D)=-D$, the magnetic fluxes on $D_2$ and $D_3$ become $$F_i = f_i^2 D_2 + f_i^3 D_3 + \frac{D_i}{2} \quad \text{with} \quad f_i^k \in \mathbb{Z} \quad \text{for} \quad i,k=2,3\,.$$ Since $D_2$, $D_3$ and $D_D$ are all intersecting we have only one choice for the $B$-field to cancel one $\mathcal{F}$. We choose the $B$-field without loss of generality such that $\mathcal{F}_2 = F_2 - B = 0$. In this case, we get $$\mathcal{F}_3 = F_3 - F_2 = \left(f_3^2 - f_2^2 -\frac12\right) D_2 + \left(f_3^3 - f_2^3 +\frac12 \right) D_3\,.$$ In order to avoid additional FI-terms and/or zero-modes via chiral matter at brane intersections or in the bulk spectrum of the D7-brane stacks on $D_2$ and $D_3$, we have to demand the magnetic fluxes $\mathcal{F}_2$ and $\mathcal{F}_3$ to be effectively trivial which is the case for $$0 = \int_{D_i} \mathcal{F}_i \wedge J = \int_{X_3} J \wedge D_i \wedge \mathcal{F}_i \qquad \text{for} \quad i=2,3\,,\label{FIvanish23}$$ for the Kähler form $J = t_1 D_1 + t_2 D_2 + t_3 D_3$. This condition is trivially fulfilled for the zero flux $\mathcal{F}_2$. For $\mathcal{F}_3$, becomes $$\begin{aligned} \begin{aligned} 0 &= \int_{X_3} J \wedge D_3 \wedge \mathcal{F}_3\,,\\ &= \int_{X_3} (t_1 D_1 + t_2 D_2 + t_3 D_3) \wedge D_3 \wedge \left[ \left(f_3^2 - f_2^2 -\frac12\right) D_2 + \left(f_3^2 - f_2^2 +\frac12\right) D_3\right]\,,\\ &= t_2 \left[ \left(f_3^2 - f_2^2 -\frac12\right)\kappa_{223} + \left(f_3^2 - f_2^2 +\frac12\right)\kappa_{233} \right] + t_3 \left[ \left(f_3^2 - f_2^2 -\frac12\right)\kappa_{233} + \left(f_3^2 - f_2^2 +\frac12\right)\kappa_{333} \right]\,,\label{F3det} \end{aligned}\end{aligned}$$ using the intersection form . For general $t_2,t_3\neq 0$ is fulfilled if $$\frac{2 f^2_3 - 2 f_2^2 - 1}{2 f^3_3 - 2 f_2^3 + 1} = -\frac{\kappa_{333}}{\kappa_{233}} = -\frac{\kappa_{233}}{\kappa_{223}}\,,\label{fluxcond}$$ where the last condition can be rewritten as $\kappa_{233}^2 = \kappa_{223} \kappa_{333}$. Clearly, can not be fulfilled for general intersection numbers. The intersections that can accommodate the condition of trivial gauge flux $\mathcal{F}_2$ and $\mathcal{F}_3$ are the following: - $\kappa_{223}=\kappa_{233}=\kappa_{333}\neq0$ or - $\kappa_{222}=\kappa_{223}=\kappa_{233}\neq0$ or - $\kappa_{233} = Z^n$, $\kappa_{223} = Z^m$, $\kappa_{333} = Z^k$ with $Z$ being an odd integer and integers $k+m=2n$ or - $\kappa_{223} = Z^n$, $\kappa_{222} = Z^m$, $\kappa_{233} = Z^k$ with $Z$ being an odd integer and integers $k+m=2n$, i.e., for either of these conditions there exist flux quanta $f_i^k$ such that $\mathcal{F}_2$ and $\mathcal{F}_3$ are trivial. The second and fourth condition stem from choosing the $B$-field such that $\mathcal{F}_3=0$. [^1]: In the case of $D_D$ being a linear combination of only $D_2$ and $D_3$ this divisor is only meaningful if $D_2$ and $D_3$ intersect as a linear combination of non-intersecting and rigid, i.e., local, four-cycles would not make sense. This is the reason we consider non-zero intersections between $D_2$ and $D_3$ in the first place as opposed to the more simple setup ${\mathcal{V}}\sim \tau_1^{3/2} - \tau_2^{3/2} - \tau_3^{3/2} - {\mathcal{V}}_{\text{rest}}$. [^2]: Recently, it has been argued that the leading correction in both $\alpha'$ and string coupling constants on $SU(3)$ structure manifold comes with the Euler characteristic of the six-dimensional manifold as well as Calabi-Yau compactifications [@Grana:2014vva]. [^3]: Note that this would be more obvious if we start from a toy setup with $W = W_0 + A_2 e^{-a_2 T_b} + A_3 e^{-a_3 T_c}$, although one might not obtain the D-term constraint like $\tau_c = c \tau_b$.

ArXiv

--- abstract: 'The author introduces the concept of harmonically $s$-convex functions and establishes some Ostrowski type inequalities and Hermite-Hadamard type inequality of these classes of functions.' address: | Department of Mathematics, Faculty of Arts and Sciences,\ Giresun University, 28100, Giresun, Turkey. author: - İmdat İşcan title: 'Ostrowski type inequalities for harmonically $s$-convex functions' --- Introduction ============ Let $f:I\mathbb{\rightarrow R}$, where $I\subseteq \mathbb{R}$ is an interval, be a mapping differentiable in $I^{\circ }$ (the interior of $I$) and let $a,b\in I^{\circ }$ with $a<b.$ If $\left\vert f^{\prime }(x)\right\vert \leq M,$ for all $x\in \left[ a,b\right] ,$ then the following inequality holds$$\left\vert f(x)-\frac{1}{b-a}\int_{a}^{b}f(t)dt\right\vert \leq M(b-a)\left[ \frac{1}{4}+\frac{\left( x-\frac{a+b}{2}\right) ^{2}}{\left( b-a\right) ^{2}}% \right] \label{1-1}$$for all $x\in \left[ a,b\right] .$ This inequality is known in the literature as the Ostrowski inequality (see [@O38]), which gives an upper bound for the approximation of the integral average $\frac{1}{b-a}% \int_{a}^{b}f(t)dt$ by the value $f(x)$ at point $x\in \left[ a,b\right] $. For some results which generalize, improve and extend the inequalities([1-1]{}) we refer the reader to the recent papers (see [@ADDC10; @L12] ). In [@HM94], Hudzik and Maligranda considered the following class of functions: A function $f:I\subseteq %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{+}\rightarrow %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion $ where $% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{+}=\left[ 0,\infty \right) $, is said to be $s$-convex in the second sense if$$f\left( \alpha x+\beta y\right) \leq \alpha ^{s}f(x)+\beta ^{s}f(y)$$for all $x,y\in I$ and $\alpha ,\beta \geq 0$ with $\alpha +\beta =1$ and $s$ fixed in $\left( 0,1\right] $. They denoted this by $K_{s}^{2}.$ It can be easily seen that for $s=1$, $s$-convexity reduces to ordinary convexity of functions defined on $[0,\infty )$. In [@DF99], Dragomir and Fitzpatrick proved a variant of Hermite-Hadamard inequality which holds for the $s$-convex functions. Suppose that $f:% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{+}\mathbb{\rightarrow }% %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion _{+}$ is an $s$-convex function in the second sense, where $s\in \lbrack 0,1) $ and let $a,b\in \lbrack 0,\infty )$, $a<b$. If $f\in L\left[ a,b% \right] $, then the following inequalities hold $$2^{s-1}f\left( \frac{a+b}{2}\right) \leq \frac{1}{b-a}\dint% \limits_{a}^{b}f(x)dx\leq \frac{f(a)+f(b)}{s+1}\text{.} \label{1-2}$$the constant $k=\frac{1}{s+1}$ is the best possible in the second inequality in (\[1-2\]). The above inequalities are sharp. For recent results and generalizations concerning $s$-convex functions see [@ADK11; @DF99; @HBI09; @I13b; @KBO07]. In [@I13c], the author gave harmonically convex and established Hermite-Hadamard’s inequality for harmonically convex functions as follows: Let $I\subset %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion \backslash \left\{ 0\right\} $ be a real interval. A function $% f:I\rightarrow %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion $ is said to be harmonically convex, if   $$f\left( \frac{xy}{tx+(1-t)y}\right) \leq tf(y)+(1-t)f(x) \label{1-3}$$for all $x,y\in I$ and $t\in \lbrack 0,1]$. If the inequality in (\[1-3\]) is reversed, then $f$ is said to be harmonically concave. Let $f:I\subset %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion \backslash \left\{ 0\right\} \rightarrow %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion $ be a harmonically convex function and $a,b\in I$ with $a<b.$ If $f\in L[a,b]$ then the following inequalities hold $$f\left( \frac{2ab}{a+b}\right) \leq \frac{ab}{b-a}\dint\limits_{a}^{b}\frac{% f(x)}{x^{2}}dx\leq \frac{f(a)+f(b)}{2}. \label{1-4}$$The  above inequalities are sharp. The goal of this paper is to introduce the concept of the harmonically $s$-convex functions, obtain the similar the inequalities (\[1-4\]) for harmonically $s$-convex functions and establish some new inequalities of Ostrowski type for harmonically $s$-convex functions. Main Results ============ Let $I\subset \left( 0,\infty \right) $ be an real interval. A function $% f:I\rightarrow %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion $ is said to be harmonically $s-$convex (concave), if   $$f\left( \frac{xy}{tx+(1-t)y}\right) \leq \left( \geq \right) t^{s}f(y)+(1-t)^{s}f(x) \label{2-1}$$for all $x,y\in I$ , $t\in \lbrack 0,1]$ and for some fixed $s\in \left( 0,1% \right] $. Let $I\subset \left( 0,\infty \right) $ be an real interval and $% f:I\rightarrow %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion $ is a function, then ; 1. if f is $s$-convex and nondecreasing function then f is harmonically $s $-convex. 2. if f is harmonically $s$-convex and nonincreasing function then f is $s $-convex. Since $f:\left( 0,\infty \right) \rightarrow %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion ,\ f(x)=x,$ harmonically convex function, we have $$\frac{xy}{tx+(1-t)y}\leq ty+(1-t)x \label{2-1a}$$for all $x,y\in \left( 0,\infty \right) $ , $t\in \lbrack 0,1]$ (see also [@CD11 page 4]). The proposition (1) and (2) is easily obtained from the inequality (\[2-1a\]). Let $s\in \left( 0,1\right] $ and $f:\left( 0,1\right] \rightarrow \left( 0,1% \right] ,\ f(x)=x^{s}$. Since $f$ is $s$-convex (see  [@HM94]) and nondecreasing function, f is harmonically $s-$convex. The following result of the Hermite-Hadamard type holds. \[2.2\]Let $f:I\subset \left( 0,\infty \right) \rightarrow %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion $ be an harmonically $s$-convex function and $a,b\in I$ with $a<b.$ If $f\in L[a,b]$ then the following inequalities hold:$$2^{s-1}f\left( \frac{2ab}{a+b}\right) \leq \frac{ab}{b-a}\dint\limits_{a}^{b}% \frac{f(x)}{x^{2}}dx\leq \frac{f(a)+f(b)}{s+1}. \label{2-2}$$ Since $f:I\rightarrow %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion $ is an harmonically $s$-convex function, we have, for all $x,y\in I$ (with $% t=\frac{1}{2}$ in the inequality (\[2-1\]) )$$f\left( \frac{2xy}{x+y}\right) \leq \frac{f(y)+f(x)}{2^{s}}$$Choosing $x=\frac{ab}{ta+(1-t)b},\ y=\frac{ab}{tb+(1-t)a}$, we get$$f\left( \frac{2ab}{a+b}\right) \leq \frac{f\left( \frac{ab}{tb+(1-t)a}% \right) +f\left( \frac{ab}{ta+(1-t)b}\right) }{2^{s}}$$Further, integrating for $t\in \lbrack 0,1]$, we have$$f\left( \frac{2ab}{a+b}\right) \leq \frac{1}{2^{s}}\left[ \dint% \limits_{0}^{1}f\left( \frac{ab}{tb+(1-t)a}\right) dt+\dint\limits_{0}^{1}f\left( \frac{ab}{ta+(1-t)b}\right) dt\right] \label{2-2a}$$Since each of the integrals is equal to $\frac{ab}{b-a}\dint\limits_{a}^{b}% \frac{f(x)}{x^{2}}dx$, we obtain the left-hand side of the inequality ([2-2]{}) from (\[2-2a\]). The proof of the second inequality follows by using (\[2-1\]) with $x=a$ and $y=b$ and integrating with respect to $t$ over $[0,1]$. In order to prove our main theorems, we need the following lemma: \[2.1\]Let $f:I\subset %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion \backslash \left\{ 0\right\} \rightarrow %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion $ be a differentiable function on $I^{\circ }$ and $a,b\in I$ with $a<b$. If $f^{\prime }\in L[a,b]$ then $$\begin{aligned} &&f(x)-\frac{ab}{b-a}\dint\limits_{a}^{b}\frac{f(u)}{u^{2}}du \\ &=&\frac{ab}{b-a}\left\{ \left( x-a\right) ^{2}\dint\limits_{0}^{1}\frac{t}{% \left( ta+(1-t)x\right) ^{2}}f^{\prime }\left( \frac{ax}{ta+(1-t)x}\right) dt\right. \\ &&-\left. \left( b-x\right) ^{2}\dint\limits_{0}^{1}\frac{t}{\left( tb+(1-t)x\right) ^{2}}f^{\prime }\left( \frac{bx}{tb+(1-t)x}\right) dt\right\}\end{aligned}$$ Integrating by part and changing variables of integration yields$$\begin{aligned} &&\frac{ab}{b-a}\left\{ \left( x-a\right) ^{2}\dint\limits_{0}^{1}\frac{t}{% \left( ta+(1-t)x\right) ^{2}}f^{\prime }\left( \frac{ax}{ta+(1-t)x}\right) dt\right. \\ &&-\left. \left( b-x\right) ^{2}\dint\limits_{0}^{1}\frac{t}{\left( tb+(1-t)x\right) ^{2}}f^{\prime }\left( \frac{bx}{tb+(1-t)x}\right) dt\right\} \end{aligned}$$$$\begin{aligned} &=&\frac{1}{x(b-a)}\left[ b\left( x-a\right) \dint\limits_{0}^{1}tdf\left( \frac{ax}{ta+(1-t)x}\right) +a\left( b-x\right) \dint\limits_{0}^{1}tdf\left( \frac{bx}{tb+(1-t)x}\right) \right] \\ &=&\frac{1}{x(b-a)}\left[ b\left( x-a\right) \left\{ \left. tf\left( \frac{ax% }{ta+(1-t)x}\right) \right\vert _{0}^{1}-\dint\limits_{0}^{1}f\left( \frac{ax% }{ta+(1-t)x}\right) dt\right\} \right] \\ &&+\frac{1}{x(b-a)}\left[ a\left( b-x\right) \left\{ \left. tf\left( \frac{bx% }{tb+(1-t)x}\right) \right\vert _{0}^{1}-\dint\limits_{0}^{1}f\left( \frac{bx% }{tb+(1-t)x}\right) dt\right\} \right] \\ &=&f(x)-\frac{ab}{b-a}\dint\limits_{a}^{b}\frac{f(u)}{u^{2}}du.\end{aligned}$$ \[2.3\]Let $f:I\subset \left( 0,\infty \right) \rightarrow %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion $ be a differentiable function on $I^{\circ }$, $a,b\in I$ with $a<b,$ and $% f^{\prime }\in L[a,b].$ If $\left\vert f^{\prime }\right\vert ^{q}$ is harmonically $s$- convex on $[a,b]$ for $q\geq 1,$ then for all $x\in \left[ a,b\right] $, we have$$\left\vert f(x)-\frac{ab}{b-a}\dint\limits_{a}^{b}\frac{f(u)}{u^{2}}% du\right\vert \label{2-3}$$$$\begin{aligned} &\leq &\frac{ab}{b-a}\left\{ \left( x-a\right) ^{2}\left( \lambda _{1}(a,x,s,q,q)\left\vert f^{\prime }\left( x\right) \right\vert ^{q}+\lambda _{2}((a,x,s,q,q)\left\vert f^{\prime }\left( a\right) \right\vert ^{q}\right) ^{\frac{1}{q}}\right. \\ &&+\left. \left( b-x\right) ^{2}\left( \lambda _{3}(b,x,s,q,q)\left\vert f^{\prime }\left( x\right) \right\vert ^{q}+\lambda _{4}(b,x,s,q,q)\left\vert f^{\prime }\left( b\right) \right\vert ^{q}\right) ^{\frac{1}{q}}\right\} ,\end{aligned}$$where $$\lambda _{1}(a,x,s,\vartheta ,\rho )=\frac{\beta \left( \rho +s+1,1\right) }{% x^{2\vartheta }}._{2}F_{1}\left( 2\vartheta ,\rho +s+1;\rho +s+2;1-\frac{a}{x% }\right) ,$$$$\lambda _{2}(a,x,s,\vartheta ,\rho )=\frac{\beta \left( \rho +1,1\right) }{% x^{2q}}._{2}F_{1}\left( 2q,\rho +1;\rho +s+2;1-\frac{a}{x}\right) ,$$$$\lambda _{3}(b,x,s,\vartheta ,\rho )=\frac{\beta \left( 1,\rho +s+1\right) }{% b^{2\vartheta }}._{2}F_{1}\left( 2\vartheta ,1;\rho +s+2;1-\frac{x}{b}% \right) ,$$$$\lambda _{4}(b,x,s,\vartheta ,\rho )=\frac{\beta \left( s+1,\rho +1\right) }{% b^{2\vartheta }}._{2}F_{1}\left( 2\vartheta ,s+1;\rho +s+2;1-\frac{x}{b}% \right) ,$$$\beta $ is Euler Beta function defined by$$\beta \left( x,y\right) =\frac{\Gamma (x)\Gamma (y)}{\Gamma (x+y)}% =\dint\limits_{0}^{1}t^{x-1}\left( 1-t\right) ^{y-1}dt,\ \ x,y>0,$$and $_{2}F_{1}$ is hypergeometric function defined by $$_{2}F_{1}\left( a,b;c;z\right) =\frac{1}{\beta \left( b,c-b\right) }% \dint\limits_{0}^{1}t^{b-1}\left( 1-t\right) ^{c-b-1}\left( 1-zt\right) ^{-a}dt,\ c>b>0,\ \left\vert z\right\vert <1\text{ (see \cite{AS65}).}$$ From Lemma \[2.1\], Power mean inequality and the harmonically $s$-convexity of $\left\vert f^{\prime }\right\vert ^{q}$ on $[a,b],$we have$$\begin{aligned} &&\left\vert f(x)-\frac{ab}{b-a}\dint\limits_{a}^{b}\frac{f(u)}{u^{2}}% du\right\vert \\ &\leq &\frac{ab}{b-a}\left\{ \left( x-a\right) ^{2}\dint\limits_{0}^{1}\frac{% t}{\left( ta+(1-t)x\right) ^{2}}\left\vert f^{\prime }\left( \frac{ax}{% ta+(1-t)x}\right) \right\vert dt\right. \\ &&+\left. \left( b-x\right) ^{2}\dint\limits_{0}^{1}\frac{t}{\left( tb+(1-t)x\right) ^{2}}\left\vert f^{\prime }\left( \frac{bx}{tb+(1-t)x}% \right) \right\vert dt\right\} \end{aligned}$$$$\begin{aligned} &\leq &\frac{ab\left( x-a\right) ^{2}}{b-a}\left( \dint\limits_{0}^{1}1dt\right) ^{1-\frac{1}{q}} \label{2-3a} \\ &&\times \left( \dint\limits_{0}^{1}\frac{t^{q}}{\left( ta+(1-t)x\right) ^{2q}}\left[ t^{s}\left\vert f^{\prime }\left( x\right) \right\vert ^{q}+(1-t)^{s}\left\vert f^{\prime }\left( a\right) \right\vert ^{q}\right] dt\right) ^{\frac{1}{q}} \notag\end{aligned}$$$$\begin{aligned} &&+\frac{ab\left( b-x\right) ^{2}}{b-a}\left( \dint\limits_{0}^{1}1dt\right) ^{1-\frac{1}{q}} \\ &&\times \left( \dint\limits_{0}^{1}\frac{t^{q}}{\left( tb+(1-t)x\right) ^{2q}}\left[ t^{s}\left\vert f^{\prime }\left( x\right) \right\vert ^{q}+(1-t)^{s}\left\vert f^{\prime }\left( b\right) \right\vert ^{q}\right] dt\right) ^{\frac{1}{q}},\end{aligned}$$where an easy calculation gives$$\dint\limits_{0}^{1}\frac{t^{q+s}}{\left( ta+(1-t)x\right) ^{2q}}dt=\frac{% \beta \left( q+s+1,1\right) }{x^{2q}}._{2}F_{1}\left( 2q,q+s+1;q+s+2;1-\frac{% a}{x}\right) , \label{2-3b}$$$$\dint\limits_{0}^{1}\frac{t^{q+s}}{\left( tb+(1-t)x\right) ^{2q}}dt=\frac{% \beta \left( 1,q+s+1\right) }{b^{2q}}._{2}F_{1}\left( 2q,1;q+s+2;1-\frac{x}{b% }\right) ,$$$$\dint\limits_{0}^{1}\frac{t^{q}(1-t)^{s}}{\left( ta+(1-t)x\right) ^{2q}}dt=% \frac{\beta \left( q+1,s+1\right) }{x^{2q}}._{2}F_{1}\left( 2q,q+1;s+q+2;1-% \frac{a}{x}\right) ,$$$$\dint\limits_{0}^{1}\frac{t^{q}(1-t)^{s}}{\left( tb+(1-t)x\right) ^{2q}}dt=% \frac{\beta \left( s+1,q+1\right) }{b^{2q}}._{2}F_{1}\left( 2q,s+1;s+q+2;1-% \frac{x}{b}\right) . \label{2-3c}$$Hence, If we use (\[2-3b\])-(\[2-3c\]) in (\[2-3a\]), we obtain the desired result. This completes the proof. In Theorem \[2.3\], additionally, if $|f^{\prime }(x)|\leq M$, $x\in \left[ a,b\right] ,$ then inequality $$\begin{aligned} &&\left\vert f(x)-\frac{ab}{b-a}\dint\limits_{a}^{b}\frac{f(u)}{u^{2}}% du\right\vert \\ &\leq &\frac{ab}{b-a}M\left\{ \left( x-a\right) ^{2}\left( \lambda _{1}(a,x,s,q,q)+\lambda _{2}((a,x,s,q,q)\right) ^{\frac{1}{q}}\right. \\ &&+\left. \left( b-x\right) ^{2}\left( \lambda _{3}(b,x,s,q,q)+\lambda _{4}(b,x,s,q,q)\right) ^{\frac{1}{q}}\right\} \end{aligned}$$holds. \[2.4\]Let $f:I\subset \left( 0,\infty \right) \rightarrow %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion $ be a differentiable function on $I^{\circ }$, $a,b\in I$ with $a<b,$ and $% f^{\prime }\in L[a,b].$ If $\left\vert f^{\prime }\right\vert ^{q}$ is harmonically $s$- convex on $[a,b]$ for $q\geq 1,$ then for all $x\in \left[ a,b\right] $, we have$$\left\vert f(x)-\frac{ab}{b-a}\dint\limits_{a}^{b}\frac{f(u)}{u^{2}}% du\right\vert \label{2-4}$$$$\begin{aligned} &\leq &\frac{ab}{b-a}\left( \frac{1}{2}\right) ^{1-\frac{1}{q}}\left\{ \left( x-a\right) ^{2}\left( \lambda _{1}(a,x,s,q,1)\left\vert f^{\prime }\left( x\right) \right\vert ^{q}+\lambda _{2}(a,x,s,q,1)\left\vert f^{\prime }\left( a\right) \right\vert ^{q}\right) ^{\frac{1}{q}}\right. \\ &&+\left. \left( b-x\right) ^{2}\left( \lambda _{3}(b,x,s,q,1)\left\vert f^{\prime }\left( x\right) \right\vert ^{q}+\lambda _{4}(b,x,s,q,1)\left\vert f^{\prime }\left( b\right) \right\vert ^{q}\right) ^{\frac{1}{q}}\right\} \end{aligned}$$where $\lambda _{1}$, $\lambda _{2}$, $\lambda _{3}$ and $\lambda _{4}$ are defined as in Theorem \[2.3\]. From Lemma \[2.1\], Power mean inequality and the harmonically $s$-convexity of $\left\vert f^{\prime }\right\vert ^{q}$ on $[a,b],$we have$$\begin{aligned} &&\left\vert f(x)-\frac{ab}{b-a}\dint\limits_{a}^{b}\frac{f(u)}{u^{2}}% du\right\vert \label{2-4a} \\ &\leq &\frac{ab\left( x-a\right) ^{2}}{b-a}\left( \dint\limits_{0}^{1}tdt\right) ^{1-\frac{1}{q}} \notag \\ &&\times \left( \dint\limits_{0}^{1}\frac{t}{\left( ta+(1-t)x\right) ^{2q}}% \left[ t^{s}\left\vert f^{\prime }\left( x\right) \right\vert ^{q}+(1-t)^{s}\left\vert f^{\prime }\left( a\right) \right\vert ^{q}\right] dt\right) ^{\frac{1}{q}} \notag\end{aligned}$$$$\begin{aligned} &&+\frac{ab\left( b-x\right) ^{2}}{b-a}\left( \dint\limits_{0}^{1}tdt\right) ^{1-\frac{1}{q}} \\ &&\times \left( \dint\limits_{0}^{1}\frac{t}{\left( tb+(1-t)x\right) ^{2q}}% \left[ t^{s}\left\vert f^{\prime }\left( x\right) \right\vert ^{q}+(1-t)^{s}\left\vert f^{\prime }\left( b\right) \right\vert ^{q}\right] dt\right) ^{\frac{1}{q}}\end{aligned}$$$$\begin{aligned} &\leq &\frac{ab}{b-a}\left( \frac{1}{2}\right) ^{1-\frac{1}{q}}\left\{ \left( x-a\right) ^{2}\left( \lambda _{1}(a,x,s,q,1)\left\vert f^{\prime }\left( x\right) \right\vert ^{q}+\lambda _{2}(a,x,s,q,1)\left\vert f^{\prime }\left( a\right) \right\vert ^{q}\right) ^{\frac{1}{q}}\right. \\ &&+\left. \left( b-x\right) ^{2}\left( \lambda _{3}(b,x,s,q,1)\left\vert f^{\prime }\left( x\right) \right\vert ^{q}+\lambda _{4}(b,x,s,q,1)\left\vert f^{\prime }\left( b\right) \right\vert ^{q}\right) ^{\frac{1}{q}}\right\} \end{aligned}$$ This completes the proof. In Theorem \[2.4\], additionally, if $|f^{\prime }(x)|\leq M$, $x\in \left[ a,b\right] ,$ then inequality $$\begin{aligned} &&\left\vert f(x)-\frac{ab}{b-a}\dint\limits_{a}^{b}\frac{f(u)}{u^{2}}% du\right\vert \\ &\leq &\frac{ab}{b-a}M\left( \frac{1}{2}\right) ^{1-\frac{1}{q}}\left\{ \left( x-a\right) ^{2}\left( \lambda _{1}(a,x,s,q,1)+\lambda _{2}((a,x,s,q,1)\right) ^{\frac{1}{q}}\right. \\ &&+\left. \left( b-x\right) ^{2}\left( \lambda _{3}(b,x,s,q,1)+\lambda _{4}(b,x,s,q,1)\right) ^{\frac{1}{q}}\right\} \end{aligned}$$holds. \[2.5\]Let $f:I\subset \left( 0,\infty \right) \rightarrow %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion $ be a differentiable function on $I^{\circ }$, $a,b\in I$ with $a<b,$ and $% f^{\prime }\in L[a,b].$ If $\left\vert f^{\prime }\right\vert ^{q}$ is harmonically $s$- convex on $[a,b]$ for $q\geq 1,$ then for all $x\in \left[ a,b\right] $, we have$$\left\vert f(x)-\frac{ab}{b-a}\dint\limits_{a}^{b}\frac{f(u)}{u^{2}}% du\right\vert \label{2-5}$$$$\begin{aligned} &\leq &\frac{ab}{b-a}\left\{ \lambda _{5}^{1-\frac{1}{q}}(a,x)\left( x-a\right) ^{2}\left( \lambda _{1}(a,x,s,1,1)\left\vert f^{\prime }\left( x\right) \right\vert ^{q}+\lambda _{2}(a,x,s,1,1)\left\vert f^{\prime }\left( a\right) \right\vert ^{q}\right) ^{\frac{1}{q}}\right. \\ &&+\left. \lambda _{5}^{1-\frac{1}{q}}(b,x)\left( b-x\right) ^{2}\left( \lambda _{3}(b,x,s,1,1)\left\vert f^{\prime }\left( x\right) \right\vert ^{q}+\lambda _{4}(b,x,s,1,1)\left\vert f^{\prime }\left( b\right) \right\vert ^{q}\right) ^{\frac{1}{q}}\right\} \end{aligned}$$where $$\lambda _{5}(\theta ,x)=\frac{1}{x-\theta }\left\{ \frac{1}{\theta }-\frac{% \ln x-\ln \theta }{x-\theta }\right\} ,$$and $\lambda _{1}$, $\lambda _{2}$, $\lambda _{3}$ and $\lambda _{4}$ are defined as in Theorem \[2.3\]. From Lemma \[2.1\], Power mean inequality and the harmonically $s$-convexity of $\left\vert f^{\prime }\right\vert ^{q}$ on $[a,b],$we have$$\begin{aligned} &&\left\vert f(x)-\frac{ab}{b-a}\dint\limits_{a}^{b}\frac{f(u)}{u^{2}}% du\right\vert \label{2-5a} \\ &\leq &\frac{ab\left( x-a\right) ^{2}}{b-a}\left( \dint\limits_{0}^{1}\frac{t% }{\left( ta+(1-t)x\right) ^{2}}dt\right) ^{1-\frac{1}{q}} \notag \\ &&\times \left( \dint\limits_{0}^{1}\frac{t}{\left( ta+(1-t)x\right) ^{2}}% \left[ t^{s}\left\vert f^{\prime }\left( x\right) \right\vert ^{q}+(1-t)^{s}\left\vert f^{\prime }\left( a\right) \right\vert ^{q}\right] dt\right) ^{\frac{1}{q}} \notag\end{aligned}$$$$\begin{aligned} &&+\frac{ab\left( b-x\right) ^{2}}{b-a}\left( \dint\limits_{0}^{1}\frac{t}{% \left( tb+(1-t)x\right) ^{2}}dt\right) ^{1-\frac{1}{q}} \\ &&\times \left( \dint\limits_{0}^{1}\frac{t}{\left( tb+(1-t)x\right) ^{2}}% \left[ t^{s}\left\vert f^{\prime }\left( x\right) \right\vert ^{q}+(1-t)^{s}\left\vert f^{\prime }\left( b\right) \right\vert ^{q}\right] dt\right) ^{\frac{1}{q}}.\end{aligned}$$It is easily check that$$\dint\limits_{0}^{1}\frac{t}{\left( ta+(1-t)x\right) ^{2}}dt=\frac{1}{x-a}% \left\{ \frac{1}{a}-\frac{\ln x-\ln a}{x-a}\right\} , \label{2-5b}$$$$\dint\limits_{0}^{1}\frac{t}{\left( tb+(1-t)x\right) ^{2}}dt=\frac{1}{b-x}% \left\{ \frac{\ln b-\ln x}{b-x}-\frac{1}{b}\right\} ,$$Hence, If we use (\[2-3b\])-(\[2-3c\])  for $q=1$ and (\[2-5b\]) in (\[2-5a\]), we obtain the desired result. This completes the proof. In Theorem \[2.5\], additionally, if $|f^{\prime }(x)|\leq M$, $x\in \left[ a,b\right] ,$ then inequality $$\begin{aligned} &&\left\vert f(x)-\frac{ab}{b-a}\dint\limits_{a}^{b}\frac{f(u)}{u^{2}}% du\right\vert \\ &\leq &\frac{ab}{b-a}M\left\{ \lambda _{5}^{1-\frac{1}{q}}(a,x)\left( x-a\right) ^{2}\left( \lambda _{1}(a,x,s,1,1)+\lambda _{2}(a,x,s,1,1)\right) ^{\frac{1}{q}}\right. \\ &&+\left. \lambda _{5}^{1-\frac{1}{q}}(b,x)\left( b-x\right) ^{2}\left( \lambda _{3}(b,x,s,1,1)+\lambda _{4}(b,x,s,1,1)\right) ^{\frac{1}{q}% }\right\} \end{aligned}$$holds. \[2.6\]Let $f:I\subset \left( 0,\infty \right) \rightarrow %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion $ be a differentiable function on $I^{\circ }$, $a,b\in I$ with $a<b,$ and $% f^{\prime }\in L[a,b].$ If $\left\vert f^{\prime }\right\vert ^{q}$ is harmonically $s$-convex on $[a,b]$ for $q>1,\;\frac{1}{p}+\frac{1}{q}=1,$ then$$\left\vert f(x)-\frac{ab}{b-a}\dint\limits_{a}^{b}\frac{f(u)}{u^{2}}% du\right\vert \label{2-6}$$$$\begin{aligned} &\leq &\frac{ab}{b-a}\left( \frac{1}{p+1}\right) ^{\frac{1}{p}}\left\{ \left( x-a\right) ^{2}\left( \lambda _{1}(a,x,s,q,0)\left\vert f^{\prime }\left( x\right) \right\vert ^{q}+\lambda _{2}(a,x,s,q,0)\left\vert f^{\prime }\left( a\right) \right\vert ^{q}\right) ^{\frac{1}{q}}\right. \\ &&+\left. \left( b-x\right) ^{2}\left( \lambda _{3}(b,x,s,q,0)\left\vert f^{\prime }\left( x\right) \right\vert ^{q}+\lambda _{4}(b,x,s,q,0)\left\vert f^{\prime }\left( b\right) \right\vert ^{q}\right) ^{\frac{1}{q}}\right\} .\end{aligned}$$where $\lambda _{1}$, $\lambda _{2}$, $\lambda _{3}$ and $\lambda _{4}$ are defined as in Theorem \[2.3\]. From Lemma \[2.1\], Hölder’s inequality and the harmonically convexity of $\left\vert f^{\prime }\right\vert ^{q}$ on $[a,b],$we have$$\begin{aligned} &&\left\vert f(x)-\frac{ab}{b-a}\dint\limits_{a}^{b}\frac{f(u)}{u^{2}}% du\right\vert \\ &\leq &\frac{ab\left( x-a\right) ^{2}}{b-a}\left( \dint\limits_{0}^{1}t^{p}dt\right) ^{\frac{1}{p}} \\ &&\times \left( \dint\limits_{0}^{1}\frac{1}{\left( ta+(1-t)x\right) ^{2q}}% \left[ t^{s}\left\vert f^{\prime }\left( x\right) \right\vert ^{q}+(1-t)^{s}\left\vert f^{\prime }\left( a\right) \right\vert ^{q}\right] dt\right) ^{\frac{1}{q}}\end{aligned}$$$$\begin{aligned} &&+\frac{ab\left( b-x\right) ^{2}}{b-a}\left( \dint\limits_{0}^{1}t^{p}dt\right) ^{\frac{1}{p}} \\ &&\times \left( \dint\limits_{0}^{1}\frac{1}{\left( tb+(1-t)x\right) ^{2q}}% \left[ t^{s}\left\vert f^{\prime }\left( x\right) \right\vert ^{q}+(1-t)^{s}\left\vert f^{\prime }\left( b\right) \right\vert ^{q}\right] dt\right) ^{\frac{1}{q}}\end{aligned}$$$$\begin{aligned} &\leq &\frac{ab}{b-a}\left( \frac{1}{p+1}\right) ^{\frac{1}{p}}\left\{ \left( x-a\right) ^{2}\left( \lambda _{1}(a,x,s,q,0)\left\vert f^{\prime }\left( x\right) \right\vert ^{q}+\lambda _{2}(a,x,s,q,0)\left\vert f^{\prime }\left( a\right) \right\vert ^{q}\right) ^{\frac{1}{q}}\right. \\ &&+\left. \left( b-x\right) ^{2}\left( \lambda _{3}(b,x,s,q,0)\left\vert f^{\prime }\left( x\right) \right\vert ^{q}+\lambda _{4}(b,x,s,q,0)\left\vert f^{\prime }\left( b\right) \right\vert ^{q}\right) ^{\frac{1}{q}}\right\} .\end{aligned}$$This completes the proof. In Theorem \[2.6\], additionally, if $|f^{\prime }(x)|\leq M$, $x\in \left[ a,b\right] ,$ then inequality $$\begin{aligned} &&\left\vert f(x)-\frac{ab}{b-a}\dint\limits_{a}^{b}\frac{f(u)}{u^{2}}% du\right\vert \\ &\leq &\frac{ab}{b-a}M\left( \frac{1}{p+1}\right) ^{\frac{1}{p}}\left\{ \left( x-a\right) ^{2}\left( \lambda _{1}(a,x,s,q,0)+\lambda _{2}(a,x,s,q,0)\right) ^{\frac{1}{q}}\right. \\ &&+\left. \left( b-x\right) ^{2}\left( \lambda _{3}(b,x,s,q,0)+\lambda _{4}(b,x,s,q,0)\right) ^{\frac{1}{q}}\right\} \end{aligned}$$holds. \[2.7\]Let $f:I\subset \left( 0,\infty \right) \rightarrow %TCIMACRO{\U{211d} }% %BeginExpansion \mathbb{R} %EndExpansion $ be a differentiable function on $I^{\circ }$, $a,b\in I$ with $a<b,$ and $% f^{\prime }\in L[a,b].$ If $\left\vert f^{\prime }\right\vert ^{q}$ is harmonically $s$-convex on $[a,b]$ for $q>1,\;\frac{1}{p}+\frac{1}{q}=1,$ then$$\left\vert f(x)-\frac{ab}{b-a}\dint\limits_{a}^{b}\frac{f(u)}{u^{2}}% du\right\vert$$$$\begin{aligned} &\leq &\frac{ab}{b-a}\left\{ \left( \lambda _{1}(a,x,0,p,p)\right) ^{\frac{1% }{p}}\left( x-a\right) ^{2}\left( \frac{\left\vert f^{\prime }\left( x\right) \right\vert ^{q}+\left\vert f^{\prime }\left( a\right) \right\vert ^{q}}{s+1}\right) ^{\frac{1}{q}}\right. \\ &&+\left. \left( \lambda _{3}(b,x,0,p,p)\right) ^{\frac{1}{p}}\left( b-x\right) ^{2}\left( \frac{\left\vert f^{\prime }\left( x\right) \right\vert ^{q}+\left\vert f^{\prime }\left( b\right) \right\vert ^{q}}{s+1}% \right) ^{\frac{1}{q}}\right\} .\end{aligned}$$where $\lambda _{1}$, $\lambda _{2}$, $\lambda _{3}$ and $\lambda _{4}$ are defined as in Theorem \[2.3\]. From Lemma \[2.1\], Hölder’s inequality and the harmonically convexity of $\left\vert f^{\prime }\right\vert ^{q}$ on $[a,b],$we have$$\begin{aligned} &&\left\vert f(x)-\frac{ab}{b-a}\dint\limits_{a}^{b}\frac{f(u)}{u^{2}}% du\right\vert \\ &\leq &\frac{ab\left( x-a\right) ^{2}}{b-a}\left( \dint\limits_{0}^{1}\frac{% t^{p}}{\left( ta+(1-t)x\right) ^{2p}}dt\right) ^{\frac{1}{p}} \\ &&\times \left( \dint\limits_{0}^{1}\left[ t^{s}\left\vert f^{\prime }\left( x\right) \right\vert ^{q}+(1-t)^{s}\left\vert f^{\prime }\left( a\right) \right\vert ^{q}\right] dt\right) ^{\frac{1}{q}}\end{aligned}$$$$\begin{aligned} &&+\frac{ab\left( b-x\right) ^{2}}{b-a}\left( \dint\limits_{0}^{1}\frac{t^{p}% }{\left( tb+(1-t)x\right) ^{2p}}dt\right) ^{\frac{1}{p}} \\ &&\times \left( \dint\limits_{0}^{1}\left[ t^{s}\left\vert f^{\prime }\left( x\right) \right\vert ^{q}+(1-t)^{s}\left\vert f^{\prime }\left( b\right) \right\vert ^{q}\right] dt\right) ^{\frac{1}{q}}\end{aligned}$$$$\begin{aligned} &\leq &\frac{ab}{b-a}\left\{ \left( \lambda _{1}(a,x,0,p,p)\right) ^{\frac{1% }{p}}\left( x-a\right) ^{2}\left( \frac{\left\vert f^{\prime }\left( x\right) \right\vert ^{q}+\left\vert f^{\prime }\left( a\right) \right\vert ^{q}}{s+1}\right) ^{\frac{1}{q}}\right. \\ &&+\left. \left( \lambda _{3}(b,x,0,p,p)\right) ^{\frac{1}{p}}\left( b-x\right) ^{2}\left( \frac{\left\vert f^{\prime }\left( x\right) \right\vert ^{q}+\left\vert f^{\prime }\left( b\right) \right\vert ^{q}}{s+1}% \right) ^{\frac{1}{q}}\right\} .\end{aligned}$$This completes the proof. In Theorem \[2.7\], additionally, if $|f^{\prime }(x)|\leq M$, $x\in \left[ a,b\right] ,$ then inequality $$\begin{aligned} &&\left\vert f(x)-\frac{ab}{b-a}\dint\limits_{a}^{b}\frac{f(u)}{u^{2}}% du\right\vert \\ &\leq &\frac{ab}{b-a}M\left( \frac{2}{s+1}\right) ^{\frac{1}{q}}\left\{ \left( \lambda _{1}(a,x,0,p,p)\right) ^{\frac{1}{p}}\left( x-a\right) ^{2}\right. \\ &&+\left. \left( \lambda _{3}(b,x,0,p,p)\right) ^{\frac{1}{p}}\left( b-x\right) ^{2}\right\} \end{aligned}$$holds. [99]{} M. Alomari, M. Darus, S. S. Dragomir, and P. Cerone, Ostrowski type inequalities for functions whose derivatives are $s$-convex in the second sense, Appl. Math. Lett. 23 (2010), no. 1, 1071-1076. M. W. Alomari, M. Darus, and U. S. Kirmaci, Some inequalities of Hermite-Hadamard type for $s$-convex functions, Acta Mathematica Scientia B31, No.4 (2011), 1643–1652. M. Abramowitz and I.A. Stegun (Eds.), *Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical Tables*, Dover, New York, 1965. P. Cerone, S.S. Dragomir, Mathematical Inequalities, CRC Press, Taylor and Francis Group, Boca Raton, London, New York, 2011. S.S. Dragomir, S. Fitzpatrick, The Hadamard’s inequality for $% s$-convex functions in the second sense, Demonstratio Math. 32 (4) (1999), 687–696. S. Hussain, M. I. Bhatti, and M. Iqbal, Hadamard-type inequalities for $s$-convex functions. I, Punjab University Journal of Mathematics, 41 (2009), 51–60. H. Hudzik , L. Maligranda, Some remarks on $s$-convex functions, Aequationes Math., 48 (1994), 100–111. I. Iscan, New estimates on generalization of some integral inequalities for $s$-convex functions and their applications, International Journal of Pure and Applied Mathematics, 86**,** No.4 (2013), in press. I. Iscan, On Some New Hermite-Hadamard type inequalities for s-geometrically convex functions, arXiv:1304.0431. Avilable online at http://arxiv.org/abs/1304.0431. U. S. Kirmaci, M. Klaričić Bakula, M.E. Özdemir, and J. Pečarić, Hadamard-type inequalities for $s$-convex functions, Applied Mathematics and Computation, 193**, **No.1 (2007), 26–35. Z. Liu, A note on Ostrowski type inequalities related to some $% s$-convex functions in the second sense, Bull. Korean Math. Soc. 49 (2012), No. 4, pp. 775-785. Available online at http://dx.doi.org/10.4134/BKMS.2012.49.4.775. A. Ostrowski, Über die Absolutabweichung einer differentiebaren funktion von ihren integralmittelwert, Comment. Math. Helv. 10 (1938) 226–227.

ArXiv

--- abstract: 'This paper considers broadcast channels with $L$ antennas at the base station and $m$ single-antenna users, where each user has perfect channel knowledge and the base station obtains channel information through a finite rate feedback. The key observation of this paper is that the optimal number of on-users (users turned on), say $s$, is a function of signal-to-noise ratio (SNR) and other system parameters. Towards this observation, we use asymptotic analysis to guide the design of feedback and transmission strategies. As $L$, $m$ and the feedback rates approach infinity linearly, we derive the asymptotic optimal feedback strategy and a realistic criterion to decide which users should be turned on. Define the corresponding asymptotic throughput per antenna as the *spatial efficiency*. It is a function of the number of on-users $s$, and therefore, $s$ should be appropriately chosen. Based on the above asymptotic results, we also develop a scheme for a system with finite many antennas and users. Compared with other works where $s$ is presumed constant, our scheme achieves a significant gain by choosing the appropriate $s$. Furthermore, our analysis and scheme is valid for heterogeneous systems where different users may have different path loss coefficients and feedback rates.' author: - | Wei Dai, Youjian (Eugene) Liu and Brian Rider\ University of Colorado at Boulder\ 425 UCB, Boulder, CO, 80309, USA\ [email protected], [email protected], [email protected] bibliography: - 'Bib/\_BC\_Feedback.bib' - 'Bib/\_Jindal.bib' - 'Bib/\_Liu\_Dai.bib' - 'Bib/\_Tse.bib' title: | How Many Users should be Turned On\ in a Multi-Antenna Broadcast Channel?$^{*}$ --- broadcast channel, finite rate feedback, spatial efficiency \[sec:Introduction\]Introduction ================================ It is well known that multiple antennas can improve the spectral efficiency. This paper considers broadcast channels with $L$ antennas at the base station and $m$ single-antenna users. To achieve the full benefit, perfect channel state information (CSI) is required at both receiver and transmitter. Perfect CSI at the receiver can be obtained by estimation from the received signal. However, if CSI at the transmitter (CSIT) is obtained from feedback, perfect CSIT requires an infinite feedback rate. As this is not feasible in practice, it is important to analyze the effect of finite rate feedback and design efficient strategy accordingly. The feedback models for broadcast channels are described as follows. To save feedback rate on power control, we assume a power on/off strategy where each user is either turned on with a constant power or turned off, and that the number of on-users (the users turned on) is a constant, say $s$, independent of the channel realization. For any given channel realization, the users quantize their channel states into finite bits and feedback the corresponding indices to the base station. After receiving the feedback from users, the base station decides which users should be turned on and then forms beamforming vectors for transmission. Broadcast channels with feedback have been studied in [@Sharif_IT05_MIMO_BC_Feedback; @Jindal_IT06sub_BC_Feedback]. Ideally, if the base station has the perfect CSI, zero-forcing transmission avoids interference among users. However, with only finite rate feedback on CSI, the base station does not know the perfect channel state information and therefore interference from other users is inevitable. The interference gets so strong at high signal-to-noise ratio (SNR) regions that the system throughput is upper bounded by a constant even when SNR approach infinity. This phenomenon is called interference domination and was reported on in [@Sharif_IT05_MIMO_BC_Feedback; @Jindal_IT06sub_BC_Feedback]. The analysis is based on the assumption that the number of on-users $s$ always equals to the number of antennas at the base station $L$ ($L\le m$ is typically assumed). To limit the interference to a desired level, Sharif and Hassibi let $m$ grow exponentially with $L$ such that there are $L$ near orthogonal users with high probability [@Sharif_IT05_MIMO_BC_Feedback]. In both scenarios, a homogeneous system is assumed where all the users share the same path loss coefficient and feedback resource. Different from the above approaches, this paper studies heterogeneous broadcast systems, where different users may have different path loss coefficients and feedback rates. Furthermore, different from [@Sharif_IT05_MIMO_BC_Feedback], we focus on systems with a relatively small number of users. Note that a cooperative communication network can often be viewed as a composition of multi-access and broadcast sub-systems with a small number of users. Research on broadcast systems of small size provides insights into cooperative communications. For such systems, we solve the interference domination problem by choosing the appropriate number of on-users $s$. The reason that random beams construction in [@Sharif_IT05_MIMO_BC_Feedback] fails in our small size systems is elaborated in Theorem \[thm:Random-Beams\]. Our solution is based on the asymptotic analysis where $L,m,s$ and the feedback rates approach infinity linearly with constant ratios among them. This type of asymptotics is applied to systems of small size. The main asymptotic results are: - It is asymptotically optimal to quantize the channel directions only and ignore the channel magnitude information. The asymptotically optimal feedback function and codebook are derived accordingly. - A realistic on/off criterion is proposed to decide which users should be turned on. - The corresponding throughput per antenna converges to a constant, defined as the *spatial efficiency*. It is a function of the normalized number of on-users $\bar{s}=\frac{s}{L}$. Further, there exists a unique $\bar{s}\in\left(0,1\right)$ to maximize the the spatial efficiency. We develop a scheme to choose the appropriate $s$ for systems with finite $L$ and $m$. Simulations show that the gain achieved by choosing $s$ is significant compared with the strategies where $s\equiv L$. In addition, our scheme has the following advantages. - It is valid for heterogeneous systems. - The set of on-users is independent of the channel realization. As a result, computation complexity is low since we do not have to perform a user selection computation every fading block. - Only on-users need to feedback CSI, which saves a large amount of feedback resource. This paper is organized as follows. The system model is introduced in Section \[sec:System-Model\]. Then Section \[sec:Asymptotic-Analysis\] performs the asymptotic analysis obtaining insights into system design, and quantifies the spatial efficiency. Based on the asymptotic results, a practical scheme is developed in Section \[sec:Finite-System-Design\] for systems with finite many antennas and users. Finally, conclusions are summarized in Section \[sec:Conclusion\]. \[sec:System-Model\]System Model ================================ Consider a broadcast channel with $L$ antennas at the base station and $m$ single-antenna users. Assume that the base station employs zero forcing transmitter. Let $\gamma_{i}\in\mathbb{R}\backslash\mathbb{R}^{-}$ ($1\le i\le m$) be the path loss coefficient for user $i$. Then the signal model for user $i$ is$$Y_{i}=\sqrt{\gamma_{i}}\mathbf{h}_{i}^{\dagger}\left(\sum_{j=1}^{m}\mathbf{q}_{j}X_{j}\right)+W_{i},$$ where $Y_{i}\in\mathbb{C}$ is the received signal for user $i$, $\mathbf{h}_{i}\in\mathbb{C}^{L\times1}$ is the channel state vector for user, $\mathbf{q}_{j}\in\mathbb{C}^{L\times1}$ is the zero-forcing beamforming vector for user $j$, $X_{j}\in\mathbb{C}$ is the source signal for the user $j$ and $W_{i}\in\mathbb{C}$ is the complex Gaussian noise with zero mean and unit variance $\mathcal{CN}\left(0,1\right)$. Here, we assume that $\mathbf{q}_{j}^{\dagger}\mathbf{q}_{j}=1$ and the Rayleigh block fading channel model: the entries of $\mathbf{h}_{i}$ are independent and identically distributed (i.i.d.) $\mathcal{CN}\left(0,1\right)$. Without loss of generality, we assume that $L\le m$; if $L>m$, adding $L-m$ users with $\gamma_{i}=0$ yields an equivalent system with $L^{\prime}=m$. For the above broadcast system, it is natural to assume a total power constraint $\sum_{i=1}^{m}\mathrm{E}\left[\left|X_{i}\right|^{2}\right]\leq\rho$. Further, for implementation simplicity, we assume a power on/off strategy with a constant number of on-users as follows. [A1)]{} : Power on/off strategy: a source $X_{i}$ is either turned on with a constant power $P_{\mathrm{on}}$ or turned off. It is motivated by the fact that this strategy is near optimal for single user MIMO system [@Dai_05_Power_onoff_strategy_design_finite_rate_feedback]. [A2)]{} : A constant number of on-users: we assume that the number of on-users $s$ ($1\le s\le m$) is a constant independent of the specific channel realizations. With this assumption, $P_{\mathrm{on}}=\frac{\rho}{s}$. Here, $s$ is a function of $\rho$, $\gamma_{i}$ and feedback rate. This assumption is different from the one in [@Sharif_IT05_MIMO_BC_Feedback; @Jindal_IT06sub_BC_Feedback], where $s=L$ always ($L\leq m$ is assumed there). The finite rate feedback model is then described as follows. Assume that both base station and user $i$ knows $\gamma_{i}$[^1] but only user $i$ knows the channel state realization $\mathbf{h}_{i}$ perfectly. For given channel realizations $\mathbf{h}_{1}\cdots\mathbf{h}_{m}$, user $i$ quantizes his channel $\mathbf{h}_{i}$ into $R_{i}$ bits and then feeds the corresponding index to the base station. Formally, let $\mathcal{B}_{i}=\left\{ \hat{\mathbf{h}}\in\mathbb{C}^{L\times1}\right\} $ with $\left|\mathcal{B}_{i}\right|=2^{R_{i}}$ be a channel state codebook for user $i$. Then the quantization function is given by$$\begin{aligned} \mathfrak{q}:\mathbb{C}^{L\times1} & \rightarrow\mathcal{B}_{i}\\ \mathbf{h}_{i} & \mapsto\hat{\mathbf{h}}_{i}.\end{aligned}$$ In Section \[sub:quantization-function\] and \[sub:codebooks\], we will show how to design $\mathfrak{q}$ and $\mathcal{B}$ respectively. After receiving feedback information from users, the base station decides which $s$ users should be turned on and forms zero-forcing beamforming vectors for them. Let $A_{\mathrm{on}}$ be the set of $s$ on-users. The zero-forcing beamforming vectors $\mathbf{q}_{i}$’s $i\in A_{\mathrm{on}}$ is calculated as follows. Let $\mathcal{P}_{i}^{\perp}$ be the plane generated by $\left\{ \hat{\mathbf{h}}_{j}:\; j\in A_{\mathrm{on}}\backslash\left\{ i\right\} \right\} $. Let $\mathcal{P}_{i}$ be the orthogonal complement of $\mathcal{P}_{i}^{\perp}$ and $t$ be the dimensions of $\mathcal{P}_{i}$. Define $\mathbf{T}_{i}\in\mathbb{C}^{L\times t}$ the matrix whose columns are orthonormal and span the plane $\mathcal{P}$. Then $\mathbf{q}_{i}$ is the *unitary projection* of $\hat{\mathbf{h}}_{i}$on $\mathbf{T}_{i}$$$\mathbf{q}_{i}:=\frac{\mathbf{T}_{i}\mathbf{T}_{i}^{\dagger}\hat{\mathbf{h}}_{i}}{\left\Vert \mathbf{T}_{i}\mathbf{T}_{i}^{\dagger}\hat{\mathbf{h}}_{i}\right\Vert }.\label{eq:zero-forcing-rule}$$ Here, if $s=1$ and $A_{\mathrm{on}}=\left\{ i\right\} $, $\mathbf{T}_{i}$ is a $L\times L$ unitary matrix and $\mathbf{q}_{i}=\hat{\mathbf{h}}_{i}/\left\Vert \hat{\mathbf{h}}_{i}\right\Vert $. \[sec:Asymptotic-Analysis\]Asymptotic Analysis =============================================== In order to obtain insights into system design, this section performs asymptotic analysis by letting $L,m,R_{i}\mathrm{'s}\rightarrow\infty$ linearly. The quantization function $\mathfrak{q}$ and asymptotically optimal codebook $\mathcal{B}_{i}$ are derived in Section \[sub:quantization-function\] and \[sub:codebooks\] respectively. Then Section \[sub:On/off-Criterion\] develops a realistic on/off criterion to decide which users should be turned on. Finally Section \[sub:The-Spatial-Efficiency\] computes the corresponding spatial efficiency. \[sub:quantization-function\]Design of Quantization Function ------------------------------------------------------------ Generally speaking, full information of $\mathbf{h}_{i}$ contains the direction information $\mathbf{v}_{i}:=\mathbf{h}_{i}/\left\Vert \mathbf{h}_{i}\right\Vert $ and the magnitude information $\left\Vert \mathbf{h}_{i}\right\Vert $. In our Rayleigh fading channel model, it is well known that $\mathbf{v}_{i}$ and $\left\Vert \mathbf{h}_{i}\right\Vert $ are independent. Intuitively, joint quantization of $\mathbf{v}_{i}$ and $\left\Vert \mathbf{h}_{i}\right\Vert $ is preferred. Interestingly, Theorem \[thm:magnitude-concentration\] implies that there is no need to quantize the channel magnitudes. Indeed, as $L,m\rightarrow\infty$ linearly, all users’ channel magnitudes concentrate on a single value with probability one. \[thm:magnitude-concentration\]For $\forall\epsilon>0$, as $L,m\rightarrow\infty$ with $\frac{m}{L}\rightarrow\bar{m}\in\mathbb{R}^{+}$,$$\Pr\left(\underset{1\le i\le m}{\max}\;\frac{1}{L}\left\Vert \mathbf{h}_{i}\right\Vert ^{2}\ge1+\epsilon\right)\rightarrow0,$$ and $$\Pr\left(\underset{1\le i\le m}{\min}\;\frac{1}{L}\left\Vert \mathbf{h}_{i}\right\Vert ^{2}\le1-\epsilon\right)\rightarrow0.$$ The proof of Theorem \[thm:magnitude-concentration\] is omitted due to the space limitation. An important fact behind the proof is that whether the users’ channel magnitudes concentrate or not depends on the relationship between $L$ and $m$: this concentration happens in our asymptotic region where $L$ and $m$ are of the same order. To fully understand Theorem \[thm:magnitude-concentration\], it is important to realize that the Law of Large Numbers does not imply that all users’ channel magnitudes will concentrate. According to the Law of Large Numbers, $\frac{1}{L}\left\Vert \mathbf{h}_{i}\right\Vert \rightarrow1$ almost surely for any *given* $i$. However, if $m$ approaches infinity exponentially with $L$, there are certain number of users whose channel magnitudes are larger than others’, and therefore it may be still beneficial to quantize and feedback channel magnitude information. This phenomenon is illustrated by the following example. (A case where magnitude information is beneficial) Consider a broadcast channel with $\gamma_{1}=\cdots=\gamma_{m}=1$. As $L,m\rightarrow\infty$ with $\log\left(m\right)/L\rightarrow\bar{m}^{\prime}\in\mathbb{R}^{+}$, there exists an $\epsilon>0,$ $\delta_{1}>0$ and $\delta_{2}>0$ such that $$\frac{1}{L}\log\left|\left\{ i:\;\frac{1}{L}\left\Vert \mathbf{h}_{i}\right\Vert ^{2}>1+\frac{\epsilon}{2}\right\} \right|\rightarrow\delta_{1},$$ and $$\frac{1}{L}\log\left|\left\{ i:\;\frac{1}{L}\left\Vert \mathbf{h}_{i}\right\Vert ^{2}<1-\frac{\epsilon}{2}\right\} \right|\rightarrow\delta_{2}$$ with probability one. Note that there are a set of users whose channel magnitudes are $\epsilon$-larger than another set of users. It may be worth to let the base station know which users have stronger channels. Theorem \[thm:magnitude-concentration\] implies that it is sufficient to quantize the channel direction information only and omit the channel magnitude information. For this quantization, the codebook is given by $\mathcal{B}_{i}=\left\{ \mathbf{p}\in\mathbb{C}^{L\times1}:\;\left\Vert \mathbf{p}\right\Vert =1\right\} $ with $\left|\mathcal{B}_{i}\right|=2^{R_{i}}$. We use the following quantization function $$\begin{aligned} \mathfrak{q}:\mathbb{C}^{L\times1} & \rightarrow\mathcal{B}_{i}\nonumber \\ \mathbf{h}_{i} & \mapsto\mathbf{p}_{i}=\underset{\mathbf{p}\in\mathcal{B}_{i}}{\arg\;\max}\;\left|\mathbf{v}_{i}^{\dagger}\mathbf{p}\right|,\label{eq:quantization-fn}\end{aligned}$$ where $\mathbf{v}_{i}$ is the channel direction vector. \[sub:codebooks\]Asymptotically Optimal Codebooks ------------------------------------------------- Given the quantization function (\[eq:quantization-fn\]), the distortion of a given codebook $\mathcal{B}_{i}$ is the average chordal distance between the actual and quantized channel directions corresponding to the codebook $\mathcal{B}_{i}$ and defined as $$D\left(\mathcal{B}_{i}\right):=1-\mathrm{E}_{\mathbf{h}_{i}}\left[\underset{\mathbf{p}\in\mathcal{B}_{i}}{\max}\left|\mathbf{v}_{i}^{\dagger}\mathbf{p}\right|^{2}\right].$$ The following lemma bounds the minimum achievable distortion for a given codebook rate (usually called the distortion rate function). \[lem:dist-rate-fn\]Define $D^{*}\left(R\right)\triangleq\underset{\mathcal{B}:\;\left|\mathcal{B}\right|\leq2^{R}}{\inf}D\left(\mathcal{B}\right)$. Then $$\begin{aligned} & & \frac{L-1}{L}2^{-\frac{R}{L-1}}\left(1+o\left(1\right)\right)\leq D^{*}\left(R\right)\nonumber \\ & & \quad\quad\quad\leq\frac{\Gamma\left(\frac{1}{L-1}\right)}{L-1}2^{-\frac{R}{L-1}}\left(1+o\left(1\right)\right),\label{eq:quantization-bds}\end{aligned}$$ and as $L$ and $R$ approach infinity with $\frac{R}{L}\rightarrow\bar{r}\in\mathbb{R}^{+}$, $$\underset{\left(L,R\right)\rightarrow\infty}{\lim}D^{*}\left(R\right)=2^{-\bar{r}}.$$ The following Lemma shows that a random codebook is asymptotically optimal with probability one. \[lem:random-codes-asymptotically-optimal\]Let $\mathcal{B}_{\mathrm{rand}}$ be a random codebook where the vectors $\mathbf{p}\in\mathcal{B}_{\mathrm{rand}}$’s are independently generated from the isotropic distribution. Let $R=\log\left|\mathcal{B}_{\mathrm{rand}}\right|$. As $L,R\rightarrow\infty$ with $\frac{R}{L}\rightarrow\bar{r}\in\mathbb{R}^{+}$, for $\forall\epsilon>0$, $$\underset{\left(L,R\right)\rightarrow\infty}{\lim}\;\mathrm{Pr}\left\{ \mathcal{B}_{\mathrm{rand}}:\; D\left(\mathcal{B}_{\mathrm{rand}}\right)>2^{-\bar{r}}+\epsilon\right\} =0.$$ The proofs of Lemma \[lem:dist-rate-fn\] and \[lem:random-codes-asymptotically-optimal\] are given in our paper [@Dai_05_Quantization_Grassmannian_manifold]. Due to the asymptotic optimality of random codebooks, we assume that the codebooks $\mathcal{B}_{i}$’s $i=1,\cdots,m$ are independent and randomly constructed throughout this paper. \[sub:On/off-Criterion\]On/off Criterion ---------------------------------------- After receiving feedback from users, the base station should decide which $s$ users should be turned on. Ideally, for given channel realizations $\mathbf{h}_{1},\cdots,\mathbf{h}_{m}$, the optimal set of on users $A_{\mathrm{on}}^{*}$ should be chosen to maximize the instantaneous mutual information. Note that the base station only knows the quantized version of channel states $\mathbf{p}_{1},\cdots,\mathbf{p}_{m}$. It can only estimate the instantaneous mutual information through $\mathbf{p}_{i}$’s. The set $A_{\mathrm{on}}^{*}$ is given by$$\begin{aligned} A_{\mathrm{on}}^{*}= & \underset{A_{\mathrm{on}}:\;\left|A_{\mathrm{on}}\right|=s}{\arg\;\max}\;\sum_{i\in A_{\mathrm{on}}}\log\left(1+\right.\nonumber \\ & \quad\quad\quad\left.\frac{\gamma_{i}\frac{\rho}{s}\left|\mathbf{p}_{i}^{\dagger}\mathbf{q}_{i}\right|^{2}}{1+\gamma_{i}\frac{\rho}{s}\sum_{j\in A_{\mathrm{on}}\backslash\left\{ i\right\} }\left|\mathbf{p}_{i}^{\dagger}\mathbf{q}_{j}\right|^{2}}\right).\label{eq:optimal-Aon}\end{aligned}$$ However, finding $A_{\mathrm{on}}^{*}$ requires exhaustive search, whose complexity exponentially increases with $m$. The random orthonormal beams construction method in [@Sharif_IT05_MIMO_BC_Feedback] does not work for our asymptotically large system either. In [@Sharif_IT05_MIMO_BC_Feedback], the base station randomly constructs $L$ orthonormal beams $\mathbf{b}_{1},\cdots,\mathbf{b}_{L}$, finds the users with highest signal-to-noise-plus-interference ratios (SINRs) through feedback from users, and then transmits to these selected users. There, the SINR calculation for user $i$ is related to the quantity $\underset{1\le k\le L}{\max}\left|\mathbf{h}_{i}^{\dagger}\mathbf{b}_{k}\right|$. However, Theorem \[thm:Random-Beams\] below shows that in the asymptotic region where $L$ and $m$ are of the same order, all users’ channels are near orthogonal to all of the $L$ orthonormal beams $\mathbf{b}_{i}$’s. Therefore, all users’ maximum SINRs (maximum over $L$ given orthonormal beams) approach zero uniformly with probability one. The method in [@Sharif_IT05_MIMO_BC_Feedback] fails in our asymptotically large system. \[thm:Random-Beams\]Given $\forall\epsilon>0$ and any $L$ orthonormal beams $\mathbf{b}_{k}\in\mathbb{C}^{L\times1}$ $1\le k\le L$, as $L,m\rightarrow\infty$ linearly with $\frac{m}{L}\rightarrow\bar{m}\in\mathbb{R}^{+}$, $$\underset{\left(L,m\right)\rightarrow\infty}{\lim}\;\Pr\left(\underset{1\le i\le m,\;1\le k\le L}{\max}\;\frac{1}{L}\left|\mathbf{h}_{i}^{\dagger}\mathbf{b}_{k}\right|>\epsilon\right)=0.$$ The proof is omitted due to the space limitation. In this paper, we take another approach where the on/off decision is independent of channel directions. We start with the throughput analysis for a specific on-user $i\in A_{\mathrm{on}}$. Note that $$Y_{i}=\sqrt{\gamma_{i}}\mathbf{h}_{i}^{\dagger}\mathbf{q}_{i}X_{i}+\left(\sqrt{\gamma_{i}}\mathbf{h}_{i}^{\dagger}\sum_{j\in A_{\mathrm{on}}\backslash\left\{ i\right\} }\mathbf{q}_{j}X_{j}+W\right).$$ The signal power and interference power for user $i$ are given by $$P_{\mathrm{sig},i}=\frac{\rho}{s}\gamma_{i}\left|\mathbf{h}_{i}^{\dagger}\mathbf{q}_{i}\right|^{2}\label{eq:signal-power}$$ and $$P_{\mathrm{int},i}=\frac{\rho}{s}\gamma_{i}\sum_{j\in A_{\mathrm{on}}\backslash\left\{ i\right\} }\left|\mathbf{h}_{i}^{\dagger}\mathbf{q}_{j}\right|^{2}\label{eq:interference-power}$$ respectively. Note that the influence of the users in $A_{\mathrm{on}}\backslash\left\{ i\right\} $ on user $i$ only occurs through their directions $\mathbf{q}_{j}$’s $j\in A_{\mathrm{on}}\backslash\left\{ i\right\} $. If the choice of on-users $A_{\mathrm{on}}$ is independent of their channel directions, then $\mathbf{h}_{i}$ and $\mathbf{q}_{j}$’s are independent. In this case, $P_{\mathrm{sig},i}$ and $P_{\mathrm{int},i}$ can be quantified as $L,m,s,R_{i}\mathrm{'s}\rightarrow\infty$. The result is given in the following proposition. \[pro:on-user-i-throughput\]Let $\left|A_{\mathrm{on}}\right|=s$ and $L,m,s,R_{i}\mathrm{'s}\rightarrow\infty$ with $\frac{m}{L}\rightarrow\bar{m}$, $\frac{s}{L}\rightarrow\bar{s}$ and $\frac{R_{i}}{L}\rightarrow\bar{r}_{i}$. Assume that $\mathbf{v}_{i}$’s $i\in A_{\mathrm{on}}$ are independent. Then for $\forall i\in A_{\mathrm{on}}$, $$P_{\mathrm{sig},i}\rightarrow\frac{\rho}{\bar{s}}\gamma_{i}\left(1-2^{-\bar{r}_{i}}\right)\left(1-\bar{s}\right),$$ $$P_{\mathrm{int},i}\rightarrow\rho\gamma_{i}2^{-\bar{r}_{i}},$$ and therefore$$\begin{aligned} \mathcal{I}_{i} & :=\log\left(1+\frac{P_{\mathrm{sig},i}}{1+P_{\mathrm{int},i}}\right)\\ & \rightarrow\log\left(1+\eta_{i}\frac{1-\bar{s}}{\bar{s}}\right),\end{aligned}$$ with probability one, where$$\eta_{i}:=\frac{\rho\gamma_{i}\left(1-2^{-\bar{r}_{i}}\right)}{1+\rho\gamma_{i}2^{-\bar{r}_{i}}}.\label{eq:eta-zf}$$ This proposition may not be true if $\mathbf{v}_{j}$’s ($j\in A_{\mathrm{on}}\backslash\left\{ i\right\} $) are not independent of $\mathbf{v}_{i}$. Indeed, for example, if other users are chosen such that their channel directions are as orthogonal to user $i$ as possible, the interference to user $i$ is less than that achieved by our choice where channel directions are not taken into consideration. This claim is verified by the fact that $\exists\epsilon>0$ such that $$\underset{A_{\mathrm{on}}\backslash\left\{ i\right\} }{\min}\;\sum_{j\in A_{\mathrm{on}}\backslash\left\{ i\right\} }\left|\mathbf{h}_{i}^{\dagger}\mathbf{q}_{j}\right|^{2}<\sum_{j\in A_{\mathrm{on},\mathrm{rand}}\backslash\left\{ i\right\} }\left|\mathbf{h}_{i}^{\dagger}\mathbf{q}_{j}\right|^{2}-\epsilon$$ with probability one as $L,m,s\rightarrow\infty$ linearly, where $A_{\mathrm{on,}\mathrm{rand}}\backslash\left\{ i\right\} $ denotes a random choice of $A_{\mathrm{on}}\backslash\left\{ i\right\} $. Proposition \[pro:on-user-i-throughput\] shows that the user $i$’s asymptotic throughput is a constant independent of the specific channel realization $\mathbf{h}_{i}$ with probability one. Based on Proposition \[pro:on-user-i-throughput\], we select the set of $s$ on-users $A_{\mathrm{on}}$ such that $\left|A_{\mathrm{on}}\right|=s$ and $$A_{\mathrm{on}}=\left\{ i:\;\eta_{i}\ge\eta_{j}\;\mathrm{for}\;\forall j\notin A_{\mathrm{on}}\right\} ;\label{eq:on-off-criterion}$$ if there are multiple candidates, we randomly choose one of them. It is the asymptotically optimal on/off selection if the on/off decision is independent of the channel direction information. The difference between the throughput achieved by optimal on/off criterion in (\[eq:on-off-criterion\]) and the proposed one in (\[eq:optimal-Aon\]) remains unknown. \[sub:The-Spatial-Efficiency\]The Spatial Efficiency ---------------------------------------------------- We define the spatial efficiency (bits/sec/Hz/antenna) as$$\bar{\mathcal{I}}\left(\bar{s}\right):=\underset{\left(L,m,s,R_{i}\mathrm{'s}\right)\rightarrow\infty}{\lim}\bar{\mathcal{I}}^{\left(L\right)},$$ where $L,m,s,R_{i}\mathrm{'s}\rightarrow\infty$ in the same way as before, $\bar{\mathcal{I}}^{\left(L\right)}$ is the average throughput per antenna given by $$\bar{\mathcal{I}}^{\left(L\right)}:=\mathrm{E}_{\mathcal{B}_{i}\mathrm{'s},\mathbf{h}_{i}\mathrm{'s}}\left[\frac{1}{L}\sum_{i\in A_{\mathrm{on}}}\log\left(1+\frac{P_{\mathrm{sig},i}}{1+P_{\mathrm{int},i}}\right)\right],$$ and $A_{\mathrm{on}}$, $P_{\mathrm{sig},i}$ and $P_{\mathrm{int},i}$ are defined in (\[eq:on-off-criterion\]), (\[eq:signal-power\]) and (\[eq:interference-power\]) respectively. We shall quantify $\bar{\mathcal{I}}\left(\bar{s}\right)$ for a given $\bar{s}$. Define the empirical distribution of $\eta_{i}$ as $$\mu_{\eta}^{\left(m\right)}\left(\eta\le x\right):=\frac{1}{m}\left|\left\{ \eta_{i}:\;\eta_{i}\le x\right\} \right|,$$ and assume that $\mu_{\eta}:=\lim\mu_{\eta}^{\left(m\right)}$ exists weakly as $L,m,R_{i}\mathrm{'s}\rightarrow\infty$. In order to cope with $\mu_{\eta}$’s with mass points, define $$\int_{x^{+}}^{\infty}f\left(\eta\right)d\mu_{\eta}:=\underset{\Delta x\downarrow0}{\lim}\int_{x+\Delta x}^{\infty}f\left(\eta\right)d\mu_{\eta}$$ for $\forall x\in\mathbb{R}$, where $f$ is a integrable function with respect to $\mu_{\eta}$. Then $\bar{\mathcal{I}}\left(\bar{s}\right)$ is computed in the following theorem. \[thm:spatial-efficiency\]Let $L,m,s,R_{i}\mathrm{'s}\rightarrow\infty$ with $\frac{m}{L}\rightarrow\bar{m}$, $\frac{s}{L}\rightarrow\bar{s}$ and $\frac{R_{i}}{L}\rightarrow\bar{r}_{i}$. Define $$\eta_{\bar{s}}:=\sup\left\{ \eta:\;\bar{m}\int_{\eta}^{\infty}d\mu_{\eta}>\bar{s}\right\} .$$ Then as $\bar{s}\notin\left(0,1\right)$, $\bar{\mathcal{I}}\left(\bar{s}\right)=0$. If $\bar{s}\in\left(0,1\right)$, $$\begin{aligned} \bar{\mathcal{I}}\left(\bar{s}\right) & =\bar{m}\int_{\eta_{\bar{s}}^{+}}^{\infty}\log\left(1+\eta\frac{1-\bar{s}}{\bar{s}}\right)d\mu_{\eta}\nonumber \\ & \quad+\left(\bar{s}-\bar{m}\int_{\eta_{\bar{s}}^{+}}^{\infty}d\mu_{\eta}\right)\log\left(1+\eta_{\bar{s}}\frac{1-\bar{s}}{\bar{s}}\right).\label{eq:zf-spatial-efficiency}\end{aligned}$$ We are also interested in finding the optimal $\bar{s}$ to maximize $\bar{\mathcal{I}}\left(\bar{s}\right)$. Unfortunately, $\bar{\mathcal{I}}\left(\bar{s}\right)$ is not a concave function of $\bar{s}$ in general. Furthermore, the measure $\mu_{\eta}$ may contain mass points. The optimization of $\bar{\mathcal{I}}\left(\bar{s}\right)$ is therefore a non-convex and non-smooth optimization problem. The following theorem provides a criterion to find the optimal $\bar{\mathcal{I}}\left(\bar{s}\right)$. \[thm:optimal-s\]$\bar{\mathcal{I}}\left(\bar{s}\right)$ is maximized at a unique $\bar{s}^{*}\in\left(0,1\right)$ such that$$\begin{aligned} & 0\in\left[\underset{\Delta\bar{s}\rightarrow0}{\lim\;\inf}\frac{\bar{\mathcal{I}}\left(\bar{s}^{*}\right)-\bar{\mathcal{I}}\left(\bar{s}^{*}-\Delta\bar{s}\right)}{\Delta\bar{s}},\right.\nonumber \\ & \quad\quad\left.\underset{\Delta\bar{s}\rightarrow0}{\lim\;\sup}\frac{\bar{\mathcal{I}}\left(\bar{s}^{*}\right)-\bar{\mathcal{I}}\left(\bar{s}^{*}-\Delta\bar{s}\right)}{\Delta\bar{s}}\right].\label{eq:optimal-s}\end{aligned}$$ The proof is omitted due to the space limitation. The $\bar{\mathcal{I}}\left(\bar{s}^{*}\right)$ is the maximum achievable spatial efficiency for the proposed power on/off strategy. \[sec:Finite-System-Design\]Finite Dimensional System Design ============================================================ Based on the asymptotic results in Theorem \[thm:spatial-efficiency\]-\[thm:optimal-s\], we now propose a scheme for systems with finite $L$ and $m$. Throughput Estimation for Finite Dimensional Systems ---------------------------------------------------- While asymptotic analysis provide many insights, we do not apply asymptotic results directly for a finite dimensional system. The reason is that in asymptotic analysis $\frac{1}{L}\rightarrow0$ while $\frac{1}{L}$ cannot be ignored for a system with small $L$. In the following, we first calculate the main order term of the throughput for user $i\in A_{\mathrm{on}}$ and then explain the difference between asymptotic analysis and finite dimensional system analysis explicitly. To obtain the main order term, proceed as follows. Note that the throughput for user $i\in A_{\mathrm{on}}$ ($\left|A_{\mathrm{on}}\right|=s$) is $$\begin{aligned} \mathcal{I}_{i} & =\mathrm{E}\left[\log\left(1+\frac{P_{\mathrm{sig},i}}{1+P_{\mathrm{int},i}}\right)\right]\\ & =\log\left(1+\frac{\mathrm{E}\left[P_{\mathrm{sig},i}\right]}{1+\mathrm{E}\left[P_{\mathrm{int},i}\right]}\right)\\ & \quad+\mathrm{E}\left[\log\left(\frac{1+P_{\mathrm{sig},i}+P_{\mathrm{int},i}}{1+\mathrm{E}\left[P_{\mathrm{sig},i}\right]+\mathrm{E}\left[P_{\mathrm{int},i}\right]}\right)\right]\\ & \quad-\mathrm{E}\left[\log\left(\frac{1+P_{\mathrm{int},i}}{1+\mathrm{E}\left[P_{\mathrm{int},i}\right]}\right)\right],\end{aligned}$$ where $P_{\mathrm{sig},i}$ and $P_{\mathrm{int},i}$ are defined in (\[eq:signal-power\]) and (\[eq:interference-power\]). The following theorem calculates $\mathrm{E}\left[P_{\mathrm{sig},i}\right]$ and $\mathrm{E}\left[P_{\mathrm{int},i}\right]$ for finite dimensional systems. \[thm:Average-sig-int-power\]Let $\mathcal{B}_{i}$’s be randomly constructed and $D_{i}=\mathrm{E}_{\mathcal{B}_{i}}\left[D\left(\mathcal{B}_{i}\right)\right]$ for all $1\le i\le m$. For randomly chosen $A_{\mathrm{on}}$ and $i\in A_{\mathrm{on}}$, if $1\le s\le L$$$\mathrm{E}\left[P_{\mathrm{sig},i}\right]=\gamma_{i}\rho\frac{L}{s}\left[\left(1-D_{i}\right)\left(1-\frac{s-1}{L}\right)+D_{i}\frac{s-1}{L\left(L-1\right)}\right],\label{eq:zf-average-signal}$$ and $$\mathrm{E}\left[P_{\mathrm{int},i}\right]=\gamma_{i}\rho\frac{L}{s}\frac{s-1}{L-1}D_{i};\label{eq:zf-average-interference}$$ if $s>L$, $\mathrm{E}\left[P_{\mathrm{sig},i}\right]=0$. The calculation of $\mathrm{E}\left[P_{\mathrm{sig},i}\right]$ and $\mathrm{E}\left[P_{\mathrm{int},i}\right]$ relies on quantification of $D_{i}$. In general, it is difficult to compute $D_{i}$ precisely. Note that the upper bound in (\[eq:quantization-bds\]) is derived by evaluating the average performance of random codebooks (see [@Dai_05_Quantization_Grassmannian_manifold] for details). We use its main order term to estimate $D_{i}$: $$D_{i}\approx\frac{\Gamma\left(\frac{1}{L-1}\right)}{L-1}2^{-\frac{R_{i}}{L-1}}.$$ Define $$\mathcal{I}_{\mathrm{main},i}:=\log\left(1+\frac{\mathrm{E}\left[P_{\mathrm{sig},i}\right]}{1+\mathrm{E}\left[P_{\mathrm{int},i}\right]}\right).\label{eq:main-order-term}$$ It can be verified from Proposition \[pro:on-user-i-throughput\] that $\mathcal{I}_{i}=\mathcal{I}_{\mathrm{main},i}+o\left(1\right)$ and therefore $\mathcal{I}_{\mathrm{main},i}$ is the main order term of $\mathcal{I}_{i}$. Then the difference between asymptotic analysis and finite dimensional systems analysis is clear. In the limit, $\frac{s-1}{L}\rightarrow\bar{s}$ and $\frac{R_{i}}{L-1}\rightarrow\bar{r}_{i}$. However, for finite dimensional systems, simply substituting these asymptotic values into (\[eq:zf-average-signal\]-\[eq:main-order-term\]) directly introduces unpleasant error, especially when $L$ is small. Therefore, to estimate $\mathcal{I}_{i}$ ($\forall i\in A_{\mathrm{on}}$) for finite dimensional systems, we have to rely on (\[eq:zf-average-signal\]-\[eq:main-order-term\]). A Scheme for Finite Dimensional Systems --------------------------------------- Given system parameters $L$, $m$, $\gamma_{i}$’s and $R_{i}$’s, a practical scheme needs to calculate the appropriate $s$ and $A_{\mathrm{on}}$. This process is described in the following. For a given $s$, the set of $A_{\mathrm{on}}$ is decided as follows: we first calculate $\mathcal{I}_{\mathrm{main},1},\cdots,\mathcal{I}_{\mathrm{main},m}$ according to (\[eq:main-order-term\]) and then choose the $s$ users with the largest $\mathcal{I}_{\mathrm{main},i}$’s to turn on; if there exists any ambiguity, random selection is employed to resolve it. For example, if $\mathcal{I}_{\mathrm{main},1}>\mathcal{I}_{\mathrm{main},2}>\cdots>\mathcal{I}_{\mathrm{main},m}$, the user $1,2,\cdots,s$ are turned on. If $\mathcal{I}_{\mathrm{main},1}=\mathcal{I}_{\mathrm{main},2}=\cdots=\mathcal{I}_{\mathrm{main},m}$, the $s$ on-users are randomly selected from all the $m$ users. Note again, $A_{\mathrm{on}}$ is independent of the channel realization. The appropriate $s$ is chosen as follows. Let $$\mathcal{I}_{\mathrm{main}}\left(s\right)=\underset{A_{\mathrm{on}}:\;\left|A_{\mathrm{on}}\right|=s}{\max}\;\sum_{i\in A_{\mathrm{on}}}\mathcal{I}_{\mathrm{main},i}.$$ Here, note that $\mathcal{I}_{\mathrm{main},i}$ is a function of $s$. For a given broadcast system, we choose the number of on-users to be $$s_{\mathrm{main}}^{*}=\underset{1\le s\le L}{\arg\;\max}\;\mathcal{I}_{\mathrm{main}}\left(s\right).$$ Although the above procedure involves exhaustive search, the corresponding complexity is actually low. First, the calculations are independent of instantaneous channel realizations. Only system parameters $L$, $m$, $\gamma_{i}$’s and $R_{i}$’s are needed. Provided that $\gamma_{i}$’s change slowly, the base station does not need to recalculate $s_{\mathrm{main}}^{*}$ and $A_{\mathrm{on}}$ frequently. Second, $R_{i}=R_{j}$ in most systems. For such systems and a given $s$, the $s$ on-users are just simply the users with the largest $\gamma_{i}$’s. After calculating $s_{\mathrm{main}}^{*}$ and $A_{\mathrm{on}}$, the base station broadcast $A_{\mathrm{on}}$ to all the users. For each fading block, the system works as follows. - At the beginning of each fading block, the base station broadcasts a single channel training sequence to help all the users estimate their channel states $\mathbf{h}_{i}$’s. - After estimating their $\mathbf{h}_{i}$’s, the on-users quantize $\mathbf{h}_{i}$’s into $\mathbf{p}_{i}$’s according to (\[eq:quantization-fn\]) and feed the corresponding indices to the base station. - The base station then calculates the transmit beamforming vectors $\mathbf{q}_{i}$’s according to (\[eq:zero-forcing-rule\]), and then transmits $\mathbf{q}_{i}X_{i}$’s. \[Fairness Scheduling\]For systems with $\gamma_{i}\neq\gamma_{j}$ or $R_{i}\neq R_{j}$, there may be some users always turned off according to the above scheme. Fairness scheduling is therefore needed to ensure fairness of the system. There are many ways to perform fairness scheduling. Since fairness is not the primary concern of this paper, we only give an example as follows. Given $m$ users, the base station calculates the corresponding $s_{\mathrm{main}}^{*}$ and $A_{\mathrm{on}}$, and then turns on the users in $A_{\mathrm{on}}$ for the first fading block. At the second fading block, the base station considers the users who have not been turned on $\left\{ 1,\cdots,m\right\} \backslash A_{\mathrm{on}}$. It calculates the corresponding $s_{\mathrm{main}}^{*}$ and $A_{\mathrm{on}}$, and then turns on the users in the new $A_{\mathrm{on}}$. Proceed this process until all users have been turned on once. Then start a new scheduling cycle. Simulation Results ------------------ Fig. \[cap:Rate-ZF\] gives the simulation results for the proposed scheme using zero-forcing. In the simulations, $L=m=4$. For simplicity, we assume that $\gamma_{1}=\gamma_{2}=\cdots=\gamma_{m}=1$ and $R_{1}=R_{2}=\cdots=R_{m}=R_{\mathrm{fb}}$. With these assumptions, the $s$ on-users can be randomly chosen from all the $m$ users. Without loss of generality, we assume that $A_{\mathrm{on}}\equiv\left\{ 1,\cdots,s\right\} $. Let $\mathcal{I}\left(s\right)=\sum_{i\in A_{\mathrm{on}}}\mathcal{I}_{i}.$ In Fig. \[cap:Rate-ZF\], the solid lines are the simulations of $\mathcal{I}\left(s_{\mathrm{main}}^{*}\right)$ while the dashed lines are the theoretical calculation of $\mathcal{I}_{\mathrm{main}}\left(s_{\mathrm{main}}^{*}\right)$. The simulation results show that the optimal $s$ is a function of $\rho$ and $R_{\mathrm{fb}}$. For example, $s=1$ is optimal when $\rho\in\left[15,20\right]$dB and $R_{\mathrm{fb}}=6$ bits, while $s=3$ is optimal for the same SNR region as $R_{\mathrm{fb}}$ increases to 12 bits. The reason behind it is that the interference introduced by finite rate quantization is larger when $R_{\mathrm{fb}}$ is smaller: when $R_{\mathrm{fb}}$ is small, the base station needs to turn off some users to avoid strong interference as SNR gets very large. We also compare our scheme with the schemes where the number of on-users is a presumed constant (independent of $\rho$ and $R_{\mathrm{fb}}$). The throughput of schemes with presumed $s$ is presented in dotted lines. From the simulation results, the throughput achieved by choosing appropriate $s$ is always better than or equals to that with presumed $s$. Specifically, compared to the scheme in [@Jindal_IT06sub_BC_Feedback] where $s=L=4$ always, our scheme achieves a significant gain at high SNR by turning off some users. \[sec:Conclusion\]Conclusion ============================ This paper considers heterogeneous broadcast systems with a relatively small number of users. Asymptotic analysis where $L,m,s,R_{i}\rightarrow\infty$ linearly is employed to get insight into system design. Based on the asymptotic analysis, we derive the asymptotically optimal feedback strategy, propose a realistic on/off criterion, and quantify the spacial efficiency. The key observation is that the number of on-users should be appropriately chosen as a function of system parameters. Finally, a practical scheme is developed for finite dimensional systems. Simulations show that this scheme achieves a significant gain compared with previously studied schemes with presumed number of on-users. [^1]: There are many ways in which the base station obtains $\gamma_{i}$. A simple example could be that the base station measures the feedback signal strength.

ArXiv

--- author: - 'E. Valenti, M. Zoccali, A. Mucciarelli, O. A. Gonzalez, F. Surot Madrid, D. Minniti, M. Rejkuba' - 'L. Pasquini' - 'G. Fiorentino' - 'G. Bono R. M. Rich M. Soto' bibliography: - 'mybiblio.bib' title: 'The central velocity dispersion of the Milky Way bulge [^1]' --- [Recent spectroscopic and photometric surveys are providing a comprehensive view of the Milky Way bulge stellar population properties with unprecedented accuracy. This in turn allows us to explore the correlation between kinematics and stellar density distribution, crucial to constraint the models of Galactic bulge formation.]{} [The Giraffe Inner Bulge Survey (GIBS) revealed the presence of a velocity dispersion peak in the central few degrees of the Galaxy by consistently measuring high velocity dispersion in three central most fields. Due to suboptimal distribution of these fields, all being at negative latitudes and close to each other, the shape and extension of the sigma peak is poorly constrained. In this study we address this by adding new observations distributed more uniformly and in particular including fields at positive latitudes that were missing in GIBS.]{} [MUSE observations were collected in four fields at $(l, b)= (0^\circ, +2^\circ), (0^\circ, -2^\circ)$, $(+1^\circ, -1^\circ)$, and $(-1^\circ, +2^\circ)$. Individual stellar spectra were extracted for a number of stars comprised between $\sim$500 and $\sim$1200, depending on the seeing and the exposure time. Velocity measurements are done by cross-correlating observed stellar spectra in the CaT region with a synthetic template, and velocity errors obtained through Monte Carlo simulations, cross-correlating synthetic spectra with a range of different metallicities and different noise characteristics.]{} We measure the central velocity dispersion peak within a projected distance from the Galactic center of $\sim$280pc, reaching $\sigma V_{GC}\sim$140km/s at b=-1$^\circ$. This is in agreement with the results obtained previously by GIBS at negative longitude. The central sigma peak is symmetric with respect to the Galactic plane, with a longitude extension at least as narrow as predicted by GIBS. As a result of the Monte Carlo simulations we present analytical equations for the radial velocity measurement error as a function of metallicity and signal-to-noise ratio for giant and dwarf stars. {width="18cm"} Introduction ============ Recent photometric and spectroscopic surveys of the Galactic bulge are providing a wealth of data to explore the spatial distribution, chemical content and kinematics of its stellar population. A special 2016 edition of the Publications of the Astronomical Society of Australia (vol. 33) provides several reviews on the bulgenobservational properties [e.g., @zoccali+16; @babusiaux16]. Stellar kinematics and spatial distribution, in particular, are thought to be strongly correlated with the bulge formation process. Two main scenarios have been proposed for bulge formation. The first one through evolution of the disk, when the latter had been mostly converted into stars. In this case the bulge is expected to have the shape of a bar, though vertically heated into a boxy/peanut, with corresponding kinematics. The second one is the hierarchical merging of gas rich sub-clumps coming either from the disk or from satellite structures. In this case both the spatial distribution and the kinematics are expected to be more isotropic. Obviously, a combination of these two scenarios could have also led to the formation of the Galactic bulge we observed today. Early kinematical surveys covering a large bulge area such as BRAVA [@rich+07; @howard+09] and ARGOS [@freeman+13; @ness+13a; @ness+13b] derived a rotation curve that looked cylindrical, supporting the conclusion that the bulge had been formed exclusively via disk dynamical instabilities [@shen+10]. These studies, however, were limited – by crowding and interstellar extinction – to latitudes $|b|>4^{\circ}$. By using data from the GIRAFFE Inner Bulge Survey (GIBS), [@zoccali+14] found that the radial velocity dispersion ($\sigma$) exhibits a strong increase resulting in a peak with $\sim$140 km/s, confined within a radius of $\sim$250 pc from the Galactic center. It was later demostrated that this peak is spatially associated to a peak in star counts [@valenti+16], hence in stellar mass, and it is slightly dominated by metal poor stars [@zoccali+17]. Indeed, there is now consensus that the inner Galactic bulge hosts two components that are best separated in metallicity (\[Fe/H\]) , but also show different spatial distribution [@ness+12; @dekany+13; @rojas-arriagada+14; @pietrukowicz+15; @gran+16; @zoccali+17], kinematics [@babusiaux+10; @zoccali+17] and \[Mg/Fe\] ratio [@hill+11]. The velocity dispersion peak found from GIBS data was constrained by three fields, at galactic coordinates ($l,b$) = ($-0.26^{\circ}$, $-1.40^{\circ}$) (0.27$^{\circ}$, $-2.13^{\circ}$) and ($-0.58^{\circ}$, $-1.98^{\circ}$), respectively. The velocity dispersion was derived from samples of 441, 435 and 111 stars, respectively. The need to obtain intermediate resolution optical spectra for many stars, with a signal-to-noise ratio (SNR) high enough to allow us to measure Calcium II Triplet (CaT) metallicity, restricted the position of the GIBS innermost fields to the bulge hemisphere at negative latitudes. In order to constrain the shape and spatial extension of the $\sigma$-peak, we analyse here new data obtained with the MUSE IFU spectrograph at the ESO VLT, in fields closer to the Galactic center at both positive and negative latitudes. Observations and data reduction =============================== Three fields, hereafter named p0m2, p0p2 and m1p2, located in the innermost bulge regions were observed with MUSE during the Science Verification campaign. Another one, consisting of two adjacent pointings, named p1m1-A and p1m1-B, was observed in Service Mode as part of a filler program 99.B-0311A (PI: Valenti) for which only 4 hours were executed of the 76 hours originally approved. Table\[tab:log\] lists the Galactic coordinates, exposure times, image quality and interstellar extinction [@gonzalez+12] for all the fields. The two pointings of the p1m1 field were observed under quite different seeing conditions, but they are so close to each other that they have the same velocity dispersion, and are thus treated as a single field hereafter. MUSE [@muse] is the integral field spectrograph at the Nasmyth B focus of the Yepun (VLT-UT4) telescope at ESO Paranal Observatory. It provides 1 square arcmin field of view, with a spatial pixel of 0.2", and a mean spectral resolution of $R\approx3000$. The observations were carried out in seeing limited mode (WFM-noAO) by using the so-called Nominal setup, which yields a continuous wavelength coverage between 4750$\AA$ and 9350$\AA$. For all the fields, we used similar observing strategy but different total integration time (see Table\[tab:log\]): a combination of on-target sub-exposures each $\sim$1000sec long, taken with a small offsets pattern (i.e. $\sim$1.5") and $90^\circ$ rotations in order to optimise the cosmics rejection and obtaining a uniform combined dataset in terms of noise properties. The processing of the raw data was performed with the MUSE pipeline [v.1.5, @muse-pipe]. The entire pipeline data reduction cascade consists of two main steps: [*i)*]{} creating all necessary calibrations to remove the instrument signature from each target exposures, such as bias, flats, bad pixels map, instrument geometry, illumination, astrometry correction, line spread function, response curve for flux calibration, and wavelength solution map; and [*ii)*]{} constructing, for each target field, the final datacube by combining the different science exposures processed during the previous step. In addition to the final datacube, the pipeline optionally produces the so-called Field-of-View (FoV) images by convolving the MUSE datacube with the transmission curve of various filters. For this work we produced FoV images in $V$-Johnson, $R$-Cousins and $I$-Cousins. Fig.\[fig:images\] shows the color image of each target field obtained combining the $V$, $R$ and $I$ FoV images, and the position of the four fields in the velocity dispersion map provided by the GIBS survey. Clearly, the number of stars detected in each field is affected both by the different seeing conditions and by the extinction of the field. ------------------------------------------------------------------------------------------------------------------------------------------------------ Field $l$     $b$     Exp. Time FWHM E($J-Ks$) -------- ----------------- ------------------- ------------------ ------------------------------------------------------------------------ ----------- p0m2 +0.26$^{\circ}$ $-$2.14$^{\circ}$ 6 $\times$ 1000s 0.6“ & 0.36\ 0.86 m1p2 & $-$1.00$^{\circ}$ & +2.00$^{\circ}$ & 2 $\times$ 1000s & 0.5” p0p2 0.00$^{\circ}$ +2.00$^{\circ}$ 3 $\times$ 1000s 1.1“ & 0.90\ 0.87 p1m1-A & +1.20$^{\circ}$ & $-$1.00$^{\circ}$ & 6 $\times$ 1066s & 1.2” p1m1-B +1.20$^{\circ}$ $-$1.00$^{\circ}$ 6 $\times$ 1066s 0.9" & 0.86\ ------------------------------------------------------------------------------------------------------------------------------------------------------ : Galactic coordinates, reddening, image quality and observations Log of the observed fields. \[tab:log\] Extraction of the spectra ------------------------- The procedure adopted to extract the spectra for all the stellar sources present in the target fields consists of two main steps: [*i)*]{} the creation of a master star list for each field; and [*ii)*]{} the reconstruction of the spectrum of each star in the list, by using the star flux as measured in the MUSE final data cubes. We first performed standard aperture photometry, with DAOPHOT [@daophot], on the FoV images to obtain a master list of all sources with significant counts ($>4\sigma$) above the background. Due to the relatively modest crowding, aperture or PSF-fitting photometry yield virtually identical results, therefore we used aperture photometry hereafter (see below). Fig.\[fig:cmd\] shows the derived color-magnitude diagrams (CMDs) for all the fields, either in the ($R$, $V-R$) or in the ($R$, $R-I$) instrumental plane. The latter was used for the p1m1 field because, due to its higher extinction, the $V$ image had the lowest SNR. Here the impact of the different seeing and total exposure time is also very clear, with the p0p2 field being the least populated, due to the combination of relatively poor seeing and short total exposure time. The p1m1 field is the closest one to the Galactic plane, therefore showing a prominent disk main sequence (MS) that is both more prominent and extends to brighter magnitudes in comparison to the other fields. This is due to the fact that at $b=-1^{\circ}$ the optical depth of the thin disk is larger. The presence of bright blue stars is very evident also in the FoV images of Fig.\[fig:images\] ![Instrumental CMD of the observed bulge fields as derived by running aperture photometry on the MUSE FoV images. The name of the fields and total number of detected stars are given.[]{data-label="fig:cmd"}](evalentiFig2.pdf){width="9cm"} The final MUSE data cubes were then sliced along the wavelength axis into 3681 monochromatic images (i.e. single planes) sampling the target stars from 4750$\AA$ to 9350$\AA$ with a step in wavelength of 1.25$\AA$. Aperture photometry was performed on each of these images (task [PHOT]{} of [IRAF]{}[^2]) with an aperture radius $\sim$ 1.5$\times \,<FWHM>$, where $<FWHM>$ is the average image quality measured over the wavelength range $\lambda >$6000$\AA$. Finally, for each star in the master list, the corresponding spectrum was obtained by assigning to each wavelength the flux measured on the corresponding monochromatic image. For a given field, a single value was used for the aperture radius, since the FWHM variation, as measured across the entire wavelength range, is about half a pixel, independent from the mean FWHM value. It is worth mentioning that this [*photometric*]{} approach to the extraction of IFU spectra has the advantage of successfully addressing the issue of sky subtraction residuals often present in the final data cubes. Indeed, it is well known that the sky subtraction may be not always optimal, leading to the presence of artefacts (e.g. weak emission line residuals and/or [*pcygni*]{}-like profiles) in the final spectra. By contrast, any such residuals present in the single plane images are fully taken into account by the photometric procedure, which estimates a local sky background for each source present in the master list. Several attempts at using PSF-fitting photometry on the monochromatic images were performed. They were finally discarded because the majority of the stars were lost in a few of the monochromatic images, corresponding to the bottom of their strongest absorption lines. When the star flux is close to the sky level, aperture photometry still assigns a meaningful flux value, while PSF-fitting photometry just discards the star from the list. This is not a negligible issue, given that the strong absorption lines are very important in the measurement of radial velocities. On the other hand, given the modest crowding of the images, the photometry from aperture and PSF-fitting yielded similar quality result, at least on the FoV images. Therefore we judged not necessary, in this case, to try and overcome the problem of non-convergence of the PSF in the low signal regime. The analysis of more crowded fields such as the inner regions of dense star clusters might require some different approach. Spectra for 1203, 861, 496 and 502 stars were reconstructed in the p0m2, m1p2, p0p2 and p1m1 fields, respectively. Examples of typical extracted spectra, zoomed in the CaT region, for stars of different magnitudes are given in Fig.\[fig:specs\]. We show spectra for the field with the best combination of seeing and exposure time (p0m2) and the one with the worst seeing (p1m1-A). The average SNR, as measured in the CaT wavelength range, of field p0m2, m1p2, p0p2 and p1m1 stars in the faintest 0.5 mag bin is $\sim$20, 15, 10, 10 respectively. ![Example of typical spectra in the CaT region, for stars in the brightest 0.5 mag bin, for red clump stars, and star at the very faint-end magnitude range. For each spectrum, the instrumental R magnitude of the star is given.[]{data-label="fig:specs"}](evalentiFig3.pdf){width="6.7cm"} Radial velocities and velocity errors ===================================== We measured the heliocentric radial velocity (RV) of all stars detected in the observed fields through cross-correlation with a synthetic template by using the [IRAF]{} task [*fxcor*]{}. Specifically, we adopted for all stars the same synthetic spectra of a relatively metal rich (\[Fe/H\]$=-0.4$dex) K giant and performed the cross-correlation between the model and the observed spectra in the wavelength range bracketing the CaT lines. To assess the effect that the use of a single metallicity template may have on the derived velocities, in the case of p0m2 stars field, we also used 2 additional synthetic templates with metallicities (\[Fe/H\]$=-1.3$dex and $+0.2$dex) that bracket the typical metallicity distribution function observed in the GIBS fields by @zoccali+17. We found that the RV derived with the metal-poor and metal-rich models always agree within $\leq\,1$km/s, thus confirming what already noticed by @zoccali+14 that the metallicity of the adopted synthetic template has a very minor effect on the derived RV. ![RV error ($\epsilon_{RV}$) profile as a function of the spectra SNR for giants (bottom panel) and dwarf (top panel) of different metallicity. The inset shows the $\epsilon_{RV}$ profile of giants in the medium-low SNR regime.[]{data-label="fig:err_snr"}](evalentiFig4.pdf){width="8cm"} The error in the derived RV could not be estimated from repeated measurements because, for each target field , we only have one single data-cube. Therefore, the uncertainty ($\epsilon_{RV}$) has been estimated by means of MonteCarlo simulations. The main sources of uncertainties are the SNR, the spectral resolution, and the sampling. In order to evaluate their impact on the derived RVs we generated different sets of artificial MUSE spectra, varying SNR and metallicity, reproducing the observed ones. We started from synthetic spectra calculated with the code [SYNTHE]{} [@Sbordone+04], assuming typical parameters of a giant ($T_{eff}$= 4500 K, log g=1.5, $v_{turb}$=2 km/s) and a dwarf star ($T_{eff}$= 6500 K, log g=4.5, $v_{turb}$=1 km/s), and considering a grid of metallicity between \[Fe/H\]=$-$3.0 and +0.5dex with a step of 0.5dex. These synthetic spectra have been convolved with a Gaussian profile to reproduce the spectral resolution of MUSE and then resampled at the same pixel size of the observed extracted spectra (1.25 Å/pixel). Poisson noise was added to the synthetic spectra in order to reproduce different noise conditions, from SNR$\sim$10 to SNR$\sim$100 with steps of 10. At the end, for each metallicity and SNR, a sample of 500 synthetic spectra with randomly added noise was generated and their RVs measured through cross-correlation technique ([*fxcor*]{}) with the original synthetic spectrum as template. The dispersion of the derived RVs of each sample has been assumed as 1$\sigma$ uncertainty in the RV measurement for a given SNR and metallicity. We derived the following relations that link the radial velocity error to SNR and metallicity for giant stars (1): $$$$\ln(\epsilon_{RV})=4.209 - 0.997\ln({\rm SNR})-0.029{\rm [Fe/H]}+0.058{\rm [Fe/H]}^{2}$$$$ and for dwarf stars (2): $$$$\ln(\epsilon_{RV})=4.624 - 1.023\ln({\rm SNR})-0.159{\rm [Fe/H]}+0.120{\rm [Fe/H]}^{2}$$$$ The behaviour of $\epsilon_{RV}$ as a function of SNR for different values of \[Fe/H\] is shown in Fig.\[fig:err\_snr\] for giant and dwarf stars. For giant stars $\epsilon_{RV}$ increases rapidly for SNR smaller than 30, reaching uncertainties at SNR= 10 of 6.8 and 12.5 km/s for \[Fe/H\]=+0.5 and –3.0 dex, respectively, while at high SNR $\epsilon_{RV}$ is almost constant and close $\sim$1 km/s. At a given SNR, $\epsilon_{RV}$ increases as decreasing \[Fe/H\], due to the weakening of the CaT lines, while for \[Fe/H\] larger than -1.0 dex the curves are almost indistinguishable. A similar general behaviour is found also for the dwarf stars, but with larger uncertainties because of the weakness of the CaT lines: in particular at SNR= 10, the relation provides $\epsilon_{RV}$= 9.2 km/s for \[Fe/H\]=+0.5 dex, while at lower metallicities the uncertainties increase dramatically (up to $\sim$45 km/s for \[Fe/H\]=–3.0 dex). Note that we limited this procedure to the CaT lines spectral region, in a window between 8450 Åand 8700 Å, hence these relations are specific to the case of MUSE in this spectral window. For each field, Table\[tab:errors\] lists the errors on the RV estimates for stars in the faintest 0.5mag bin as derived by using the above mentioned relations for two different metallicity values: \[Fe/H\]=–1dex and \[Fe/H\]=+0.5dex, which represent the metal-poor and metal-rich edge of the typical bulge metallicity distribution function. In particular, due to the differences in the magnitude depth among the different fields, the values quoted in Table\[tab:errors\] have been derived by using equation (2) for the p0m2, m1p2 and p0p2 fields, whereas for p1m1 we have adopted the relation for giants (i.e. equation (1)). As expected, we found that at fixed SNR (i.e. magnitude) metal-rich stars have typically smaller radial velocity error. However, the variation over the entire bulge metallicity range is $\leq\,2$km/s (see Table\[tab:errors\] and Fig.\[fig:err\_snr\]). [cccc]{} Field & SNR & $\epsilon_{RV}$ (MP) & $\epsilon_{RV}$ (MR)\ &&km/s&km/s\ p0m2 & 20 & 6.3 & 4.5\ m1p2 & 15 & 8.4 & 6.1\ p0p2 & 10 & 12.8 & 9.2\ p1m1$^a$& 10 & 7.4 & 6.8\ \ \ Velocity dispersion =================== In order to measure the velocity dispersion of bulge stars, it is important to take into account the contamination by foreground disk stars, which are known to have a smaller velocity dispersion [@Ness+16; @Robin+17]. An estimate of the actual disk velocity dispersion can be attempted by selecting foreground disk MS stars in the instrumental CMD of each observed field. This is shown in Figs. \[fig:p0m2Sigma\]-\[fig:p1m1Sigma\]. Thanks to the good seeing, longer exposure time and relatively low reddening, the p0m2 field has the best defined CMD, reaching fainter magnitudes (Fig. \[fig:p0m2Sigma\]). In this field, bona fide bulge-RGB/disk-MS stars (red/blue symbols, respectively) are selected as having both $R<16$ and $(V-R)$ larger/smaller than $0.35$, respectively. The cuts isolate 75 disk MS stars, shown in blue in the CMD, and 206 bulge RGB stars shown in red. Stars fainter than $R=16$, plotted in green, cannot be safely assigned to either population. The top-right panels of Figs. \[fig:p0m2Sigma\]-\[fig:p1m1Sigma\] shows the heliocentric radial velocity versus magnitude, for all the stars, with the same color coding as before. It is clear that disk MS stars have a lower velocity dispersion, but, as expected, their radial velocity distribution is contaminated by bulge stars, both blue stragglers and sub giant branch stars. In fact, the radial velocity histogram shown at the bottom of the right panel clearly shows the presence of outliers at $|RV|>150$ km/s. Indeed, if these stars are excluded, by a simple cut at $|RV|<120$ km/s, the radial velocity dispersion drops to a value of $\sigma_{\rm RV}=45$ km/s, consistent in all three fields at $b=\pm2^{\circ}$. This exercise allows us to conclude that in the region of the CMDs above the old MS turnoff, where we can safely separate disk foreground from bulge stars by means of a color cut, the velocity dispersion of the disk is significantly lower than that of the bulge. Therefore, in order to include bulge MS stars in our analysis, we need to allow for the presence of two components with different kinematics. In the field at $b=-1^{\circ}$ the data do not reach the bulge MS, and therefore the foreground disk MS and the bulge RGB can be separated by just a color cut, at $(R-I)_{\rm inst}=0.7$. In Fig.\[fig:p1m1Sigma\], the panels on the right show that disk stars, with the same selection imposed for the other fields ($|RV|<120$ km/s), have a velocity dispersion of 39 km/s. This value is lower than the 45 km/s found at $b=\pm2^{\circ}$, consistent with the fact that this field, at $b=-1^{\circ}$, samples more thin disk stars, having a smaller velocity dispersion. Bulge RGB stars, on the other hand, have a velocity dispersion $\sigma=119$ km/s. This value will not be further refined, because it already includes all the bulge stars measured in this field, with a negligible contamination from disk stars. ![[*left:*]{} CMD of the global targets sampled in p0m2 color-coded according to their evolutionary phase. Red and blue symbols refer to the bona-fide bulge-RGB and disk-MS stars, respectively. For the bona-fide sample, the total number of stars and their radial velocity dispersion are also given. Green circles mark either bulge- or disk-MS and MS-TO stars. [*Top right:*]{} heliocentric RV as a function of the star magnitude of all stars adopting the same color code as in the left panel. Blue solid symbols refer to disk-MS stars with $|RV|<120$ km/s. [*Bottom right:*]{} heliocentric radial velocity distribution of disk-MS stars. The velocity dispersion of the total disk-MS sample and of the sub-sample obtained after applying a cut at $|RV|<120$ km/s are given.[]{data-label="fig:p0m2Sigma"}](evalentiFig5.pdf){width="7cm"} ![Same as Fig.\[fig:p0m2Sigma\] for m1p2 field.[]{data-label="fig:m1p2Sigma"}](evalentiFig6.pdf){width="7cm"} ![Same as Fig.\[fig:p0m2Sigma\] for p0p2 field.[]{data-label="fig:p0p2Sigma"}](evalentiFig7.pdf){width="7cm"} ![Same as Fig.\[fig:p0m2Sigma\] for p1m1 field. In this case the bulge MS falls below the limit magnitude, and therefore bulge RGB stars can be separated from disk MS stars by means of a simple color cut.[]{data-label="fig:p1m1Sigma"}](evalentiFig8.pdf){width="6.7cm"} The RV distributions for all the sampled stars, in the 3 fields at $b=\pm2^{\circ}$, are shown in Fig.\[fig:Sigma\]. As demonstrated above, the best-fit to the observed velocity distribution in all fields is obtained with a combination of two Gaussian components with approximately the same mean ($(RV_b-RV_d)\sim$10km/s) but very different $\sigma$. The foreground disk stars contaminating the bulge sample have a velocity distribution with smaller dispersion (blue dashed line in Fig.\[fig:Sigma\]). Specifically, we found the velocity dispersion of the disk population along the line of sight towards p0m2, m1p2 and p0p2 fields to be $\sigma$=45, 40, 35km/s, respectively. These values agree with the APOGEE findings [@Ness+16] for disk stars in the foreground of the bulge, within a distance of 3 kpc, as well as those reported by @Robin+17 for thick disk stars in the solar neighbourhood. In Table\[tab:results\] we list the mean heliocentric RV and velocity dispersion measured for the bulge stars as obtained from the best-fit to the velocity distribution of the global sample. In addition, following the prescription of @ness+13b, we provide the mean galactocentric radial velocity ($V_{GC}$) by correcting the mean heliocentric value for the Sun motion with respect to the Galactic center. ![Normalised heliocentric radial velocity distribution function for all sampled stars observed in the p0m2 (top panel), m1p2 (middle panel) and p0p2 (bottom panel) fields. The best-fit to the velocity distribution (solid black line) is obtained by using a combination of two gaussian (red and blue solid lines) functions whose sigma are also reported in each panel. The residuals of the best-fit are shown for each field in the insets.[]{data-label="fig:Sigma"}](evalentiFig9.pdf){width="8.5cm"} ------- ----------------- ------------ --------------- ----------------------------- ------ Field $<V_{Helio}>$ $<V_{GC}>$ $\sigma$ N$_{b}$/N$_{tot}$ N$_{tot}$ (km/s) (km/s) (km/s) (%) p0m2 0 $\pm$ 4.4 +9.8 135 $\pm$ 3.1 79.5 1203 m1p2 $-$14 $\pm$ 4.6 $-$8.8 125 $\pm$ 3.3 87.5 861 p0p2 10 $\pm$ 7.5 19.2 137 $\pm$ 5.3 82.4 496 p1m1 1 $\pm$ 6.6 14.7 119 $\pm$ 4.7 67.3 502 ------- ----------------- ------------ --------------- ----------------------------- ------ : Mean heliocentric and galactocentric radial velocity, and velocity dispersion measured for the bulge stars in the observed fields\[tab:results\] ![Mean galactocentric radial velocity ([*top*]{}) and velocity dispersion ([*bottom*]{}) as a function of the Galactic longitude, for different latitude as listed in the labels. Big color symbols refer to fields observed with MUSE, whereas small crosses mark the innermost GIBS fields at $b=-2^{\circ}$ (black) and $b=-1^{\circ}$ (grey). The solid lines represent the expected trend of the radial velocity and velocity dispersion according to the maps derived in @zoccali+14 [equations 1 and 2]. Error bars are derived from the sampling.[]{data-label="fig:SigmaConfr"}](evalentiFig10.pdf){width="9cm"} ![Profile of the stellar projected density, as traced by RC stars, in the inner few degrees of the Galactic bulge, from @valenti+16. The figure is a zoom of their Fig. 4, with solid/dashed lines referring to negative/positive latitudes, respectively.[]{data-label="fig:Nstars"}](evalentiFig11.pdf){width="8cm"} Discussion and Conclusions ========================== We have measured radial velocities for several hundreds bulge stars in each of four fields located within $l=\pm1.5^{\circ}$ and $b=\pm2^{\circ}$, with the IFU spectrograph MUSE@VLT. All the fields are confined within a projected radius of 280 parsecs from the Galactic center, assuming the latter at 8 kpc from the Sun. The aim of this work is to assess the presence of a large peak in velocity dispersion in this inner region of the Galaxy, previously identified by @zoccali+14 based on GIBS survey, and to constrain its shape (see Fig. \[fig:images\] for a zoom of the inner region of the velocity dispersion map derived in that work). Figure \[fig:SigmaConfr\] (bottom) shows the central velocity dispersion peak as measured in the five innermost GIBS fields (black and gray small points, at $b=-2^{\circ}$ and $b=-1^{\circ}$, respectively) and as predicted in other Galactic positions according to the interpolated surface derived in that paper (black and grey curves). The curve was assumed to be symmetric above and below the Galactic plane, therefore the prediction for negative or positive latitude is identical by definition. The upper panel of Fig. \[fig:SigmaConfr\] shows the same but for the radial velocity. The new values derived in the present work are plotted in Fig.\[fig:SigmaConfr\] with large colored symbols. They confirm both the presence of the central velocity dispersion peak, and its absolute value, reaching $\sigma V_{\rm GC}$ $\sim$ 140 km/s at its center. We also confirm that the peak is symmetric above and below the plane, as the two measurements at ($l,b$)=(0$^{\circ}$,-2$^{\circ}$) and (0$^{\circ}$,+2$^{\circ}$) are mutually consistent. With the present data we cannot constrain the latitude extension of the peak better than what was done in [@zoccali+14], who found that the peak would disappear at the latitude of Baade’s window ($b=-4^{\circ}$). We can however constrain the longitude extension of the peak, which we show to be at least as narrow as predicted by GIBS in longitude. In fact, the new fields at $l=\pm1^{\circ}$ have a velocity dispersion that is lower than the prediction of the GIBS maps. In @valenti+16 we have derived maps of stellar projected density and stellar mass from star counts in the VVV PSF catalogues, using red clump stars as tracers of the total number of stars (and stellar mass). We found the presence of a peak in stellar density in the inner few degrees of the Galaxy, that is reproduced here in Fig. \[fig:Nstars\]. This demonstrates that the sudden increase in velocity dispersion is likely due to the presence of a large concentration of stars (/mass) in the inner Galaxy. The Galactic position and spatial extension of the peak roughly coincides with the sigma peak characterised here, even if its detailed shape is somehow different. In particular, while the sigma peak is still rather sharp at $b=\pm2^{\circ}$, the peak in star counts is already very shallow at these latitudes. On the other hand, while the conversion from observed velocity and mass requires dynamical modelling, as it requires hypothesis on the orbit distribution and their possible anisotropy, it is qualitatively expected that the effect of a mass concentration on the stellar velocity is felt down to some distance from the mass source. Hence it is not surprising that the sigma peak is more spatially extended than the stellar density peak. One thing that deserves further study is the fact that, according to @zoccali+17, metal-poor stars slightly dominate the stellar density at $b=1^{\circ}$ (their Fig. 7), but the velocity dispersion is higher for metal-rich stars, in the same field (their Fig. 12). This is a clear evidence that the conversion between velocity dispersion and mass involves at least another parameter (the anisotropy of the orbit distribution) and that this parameter is different for metal-poor and metal-rich stars. The data provided here should be included in the (chemo)-dynamical models of the Galaxy [e.g., @dimatteo+15; @debattista+17; @portail+17; @fragkoudi+18] in order to properly take into account the mass distribution of the inner few degrees of the Milky Way. EV, MZ, OAG, DM and MR gratefully acknowledge the Aspen Center for Physics where this work was partially completed. The Aspen Center for Physics is supported by the National Science Foundation grant PHY-1066293. During their stay in Aspen, MZ and DM were partially supported by a grant from the Simons Foundation. Support for MZ and DM is provided by the BASAL CATA Center for Astrophysics and Associated Technologies through grant PFB-06, and the Ministry for the Economy, Development, and Tourism’s Programa Iniciativa Científica Milenio through grant IC120009, awarded to Millenium Institute of Astrophysics (MAS). EV and MZ also acknowledge support from FONDECYT Regular 1150345. DM acknowledges support from FONDECYT Regular 117012. [^1]: Based on observations taken at the ESO Very Large Telescope with the MUSE instrument under programme IDs 060.A-9342 (Science Verification; PI: Valenti/Zoccali/Kuijken), and 99.B-0311A (SM; PI: Valenti). [^2]: 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.

ArXiv

--- abstract: 'Full-duplex (FD) technology is envisaged as a key component for future mobile broadband networks due to its ability to boost the spectral efficiency. FD systems can transmit and receive simultaneously on the same frequency at the expense of residual self-interference and additional interference to the network compared with half-duplex (HD) transmission. This paper analyzes the performance of wireless networks with FD multi-antenna base stations (BSs) and HD user equipments (UEs) using stochastic geometry. Our analytical results quantify the success probability and the achievable spectral efficiency and indicate the amount of cancellation needed for beneficial FD operation.' author: - | Italo Atzeni, *Member*, *IEEE*, and Marios Kountouris, *Senior Member*, *IEEE* [^1]\ [^2] bibliography: - 'IEEEabrv.bib' - 'ref\_Huawei.bib' title: | Full-Duplex MIMO Small-Cell Networks\ with Interference Cancellation --- Full duplex, interference cancellation, multiple antennas, performance analysis, small cells, stochastic geometry, ultra-dense networks. Introduction {#sec:Intro} ============ Full-duplex (FD) communication is an emerging technology that has been recognized as one of the promising solutions to cope with the demand for high data rates. Indeed, allowing the network nodes to transmit and receive over the same time/frequency resources can potentially double spectral efficiency with respect to the half-duplex (HD) counterparts (i.e., time- and frequency-division duplex) [@Sab14]. However, there are three major technical challenges hindering the implementation of FD cellular networks. Third, the concurrent, aggressive utilization of both forward and reverse links doubles the interference between neighboring cells [@Goy15]. Recently, there has also been an increasing interest in network densification as a means to fulfill the performance requirements of 5th generation (5G) wireless systems [@Bhu14]. In particular, ultra-dense networks (UDNs), i.e., the dense and massive deployment of small-cell base stations (BSs), is regarded as a key enabler for providing higher data rates and enhanced coverage by exploiting spatial reuse. Interestingly, small-cell BSs prove particularly suitable for the deployment of FD technology thanks to their reduced transmit power and the low mobility of their user equipments (UEs). This hybrid FD/HD architecture can be used either for serving an UL node and a DL node separately (with two independent data flows) or for relaying purposes to increase coverage between an UL node and a DL node (with the same data flow being received, amplified, and re-transmitted by the FD BS) [@Sab14]. Related Work {#sec:Intro_SOTA} ------------ Due to the extra interference terms introduced in FD mode (see Figure \[fig:topology\]), it is not clear how the network throughput and the aggregate interference will behave in dense multi-cell FD systems. Several recent works have examined the performance of large-scale FD networks using stochastic geometry, which is a powerful mathematical framework that provides models and tools for efficiently analyzing the performance of UDNs and heterogeneous cellular/ad hoc networks [@Hae12]. References [@Ton15; @Lee15] study the performance of bipolar networks and multi-tier heterogeneous networks, respectively, consisting of both HD and FD nodes, and quantify the impact of imperfect cancellation. Interestingly, the two papers reach the same conclusion that operating all nodes in either FD or HD mode maximizes the area spectral efficiency compared with a mixture of the two modes. On the other hand, [@Goy16] shows that raising the proportion of FD nodes increases the outage probability and thus highlights the inherent tradeoff between coverage and throughput. All these works assume single-antenna nodes, whereas multiple-input multiple-output (MIMO) nodes are considered by [@Moh15] in a single-cell setting with randomly located DL UEs. Furthermore, [@Pso16] analyzes the effect of directional antennas in reducing the overall interference in FD cellular networks and shows that directionality alone can passively suppress the . The asymptotic performance of massive MIMO-enabled backhaul nodes serving FD small-cell BSs is studied in [@Tab16], where zero-forcing beamforming allows to mitigate the interference among multiple backhaul data streams. The scenario of with massive antenna arrays is considered in [@Sho16], which also assumes mobile UEs with FD capabilities. ![System model with FD BSs relaying between HD UL and DL nodes, with corresponding desired signals and interferences.[]{data-label="fig:topology"}](./img/topology) Motivation and Contributions {#sec:Intro_contr} ---------------------------- [ Nevertheless, while massive antenna arrays are appropriate for macro-cell backhaul nodes, they are not suitable for small-cell BSs typical of UDN scenarios, which are usually equipped with a low-to-moderate number of transmit/receive antennas. Hence, it is meaningful to investigate the performance of dense FD small-cell networks with multi-antenna nodes, which is a promising and practically relevant solution for future mobile broadband networks. In particular, it is important to analyze the impact of array gain and interference cancellation techniques in mitigating the additional interference introduced by FD mode.]{} . However, these are very coarse approximations: the former neglects the strong line-of-sight signal component between transmit and receive antennas [@Dua12], whereas the latter is only meaningful when digital cancellation is applied [@Rii11]. As a matter of fact, the residual channel is known to be subject to Rician fading[^3] and, therefore, its modeling in a MIMO context represents a challenging problem when receive combining and transmit beamforming techniques are employed. - Using tools from stochastic geometry, we study the performance of wireless networks with randomly distributed FD MIMO relays and derive tight bounds for the probability of successful transmission. - We appropriately model the channel using Rician fading and we derive the distribution of the power for arbitrary receive combining and transmit beamforming strategies. The resulting expression approximately follows a gamma distribution and can be readily incorporated into existing frameworks for the performance analysis of UDNs. - We consider interference cancellation techniques at the receive side of both the and the (multi-antenna) DL nodes. The rest of the paper is structured as follows. The system model is introduced in Section \[sec:SM\]. Section \[sec:SP\] presents our main results on the performance analysis of FD MIMO small-cell networks. The proposed analytical framework is extended in Section \[sec:IC\] to accommodate interference cancellation techniques at both the FD BSs and the HD DL nodes. In Section \[sec:NUM\], numerical results are reported to corroborate our theoretical findings and to establish under which conditions FD mode outperforms HD mode. Finally, Section \[sec:END\] summarizes our contributions and draws some concluding remarks. System Model {#sec:SM} ============ Network Model {#sec:SM_NS} ------------- We consider a set of small-cell BSs operating in FD mode. Each FD BS acts as a relay between an UL node and a DL node,[^4] both operating in HD mode, during a given time slot; all communications occur in the same frequency band. This general scenario is depicted in Figure \[fig:topology\] and can be used to model, for instance, the two-hop communication between a backhaul node and a mobile UEs, or self-backhauled small cells. In our setting, the FD BSs are equipped with multiple receive/transmit antennas, . Alternatively, our model can be seen as an instance of multi-antenna UL/DL where nodes perform space division multiple access (SDMA) and send/receive one stream to/from the FD BS, hence being equivalently seen as single-antenna nodes by each FD BS . In short, our model consists of a single-input multiple-output transmission followed by a multiple-input single-output transmission.[^5] Let us thus introduce the stationary, independently marked Poisson point process (PPP) $\Phi_{\mathrm{m}} \triangleq \big\{ (x_{i}, \widetilde{m}(x_{i}), \widehat{m}(x_{i})) \big\}$ on $\Real^{2} \times \Real^{2} \times \Real^{2}$. with fixed distances of the desired links given by $\widetilde{R} \triangleq \| x - \widetilde{m}(x) \|$ and $\widehat{R} \triangleq \| x - \widehat{m}(x) \|$, $\forall x \in \Phi$. Therefore, we have $\widetilde{m}(x) = x + \widetilde{R} (\cos \widetilde{\varphi}_{x}, \sin \widetilde{\varphi}_{x})$ and $\widehat{m}(x) = x + \widehat{R} (\cos \widehat{\varphi}_{x}, \sin \widehat{\varphi}_{x})$, with $\{ \widetilde{\varphi}_{x}, \widehat{\varphi}_{x} \}_{x \in \Phi}$ independent and uniformly distributed in $[0,2 \pi]$. For convenience, in the rest of the paper we use the notation $\widetilde{m}_{x} \triangleq \widetilde{m}(x)$ and $\widehat{m}_{x} \triangleq \widehat{m}(x)$. Channel Model {#sec:SM_CM} ------------- We assume that the FD BSs and the HD UL nodes transmit with constant powers $\widehat{\rho}$ and $\widetilde{\rho}$, respectively. Furthermore, the FD BSs are equipped with $N_{{\mathrm{R}}}$ receive antennas and $N_{{\mathrm{T}}}$ transmit antennas. The propagation through the wireless channel is characterized as the combination of a pathloss attenuation and a small-scale fading. Given transmitting node $x$ and receiving node $z$, we use the following notation. The pathloss between nodes $x$ and $z$ is given by the function $\ell (x,z) \triangleq \|x - z\|^{-\alpha}$, with pathloss exponent $\alpha > 2$. The channels are denoted by $\H_{x z} \in \Compl^{N_{{\mathrm{R}}} \times N_{{\mathrm{T}}}}$ if $x, z \in \Phi$, as $\h_{x z} \in \Compl^{N_{{\mathrm{R}}} \times 1}$ if $x \in \widetilde{\Phi}$ and $z \in \Phi$, as $\h_{x z} \in \Compl^{N_{{\mathrm{T}}} \times 1}$ if $x \in \Phi$ and $z \in \widehat{\Phi}$, and as $h_{x z} \in \Compl$ if $x \in \widetilde{\Phi}$ and $z \in \widehat{\Phi}$; in particular, $\H_{x x}$ models the at $x \in \Phi$ resulting from its own transmission. We assume that all the channels, except the channel, are subject to Rayleigh fading with elements distributed independently as $\setC \setN (0, 1)$. On the other hand, the channel is subject to Rician fading [@Dua12] and, therefore, and standard deviation of $\H_{x x}$ as (cf. [@Tep03]) $$\begin{aligned} \label{eq:mu_nu} \mu \triangleq \sqrt{\frac{K \Omega}{K+1}}, \qquad \nu \triangleq \sqrt{\frac{\Omega}{K+1}}.\end{aligned}$$ In addition, let $s_{x}$ represent the data symbol transmitted by $x$ with $\Exp [ |s_{x}|^2 ] = 1$, whereas the additive noise at $x$ is denoted by $\n_{x} \in \Compl^{N_{{\mathrm{R}}}}$ if $x \in \Phi$ and by $n_{x} \in \Compl$ if $x \in \widehat{\Phi}$, with elements distributed independently as $\setC \setN (0, \sigma^{2})$. Lastly, $\v_{x} \in \Compl^{N_{{\mathrm{R}}}}$ and $\w_{x} \in \Compl^{N_{{\mathrm{T}}}}$ denote the receive combining and the transmit beamforming vectors applied by $x \in \Phi$, respectively, with $\| \v_{x} \|^{2} = \| \w_{x} \|^{2} = 1$. SINR Characterization {#sec:SM_SINR} --------------------- In this section, we characterize the signal-to-interference-plus-noise ratio (SINR) at the FD BSs and at the HD DL nodes, which is then used to analyze the probability of successful transmission, also termed as *success probability*, in the next section. Our analysis focuses on a *typical FD BS*, indexed by $0$, and on its corresponding HD UL/DL nodes, referred to as *typical HD UL/DL nodes* and indexed by $\widetilde{m}_{0}$ and $\widehat{m}_{0}$, respectively. The two-hop link between these nodes is representative of the whole network, as detailed next. **First Hop:** Consider the typical FD BS located at the origin of the Euclidean plane and indexed by $0$. Due to Slivnyak’s theorem [@Hae12 Ch. 8.5] and to the stationarity of $\Phi$, the statistics of the typical BS’s signal reception are representative of the statistics seen by any FD BS: we can thus write $\ell (x,0) = r_{x}^{-\alpha}$, with $r_{x} \triangleq \| x \|$ being the distance of $x$ from the typical FD BS. Hence, the received signal at the typical FD BS is given by $$\begin{aligned} \label{eq:y_1} \y_{0} \triangleq & \ \underbrace{\sqrt{\widetilde{\rho}} \widetilde{R}^{-\frac{\alpha}{2}} \h_{\widetilde{m}_{0} 0} s_{\widetilde{m}_{0}}}_{\mathrm{(a)}} + \sum_{x \in \Phi} \underbrace{\sqrt{\widehat{\rho}} r_{x}^{-\frac{\alpha}{2}} \H_{x 0} \w_{x} s_{x}}_{\mathrm{(b)}} + \sum_{x \in \Phi} \underbrace{\sqrt{\widetilde{\rho}} r_{\widetilde{m}_{x}}^{-\frac{\alpha}{2}} \h_{\widetilde{m}_{x} 0} s_{\widetilde{m}_{x}}}_{\mathrm{(c)}} + \underbrace{\sqrt{\widehat{\rho}} \H_{0 0} \w_{0} s_{0}}_{\mathrm{(d)}} + \n_{0}\end{aligned}$$ where (a) represents the desired signal, (b) and (c) indicate the interference coming from FD BS $x$ and its associated HD UL node $\widetilde{m}_{x}$, respectively, and (d) represents the . Given the receive combining vector $\v_{0}$, the resulting SINR reads as $$\begin{aligned} \label{eq:SINR_1} {\mathrm{SINR}}_{0} & \triangleq \frac{\widetilde{\rho} \widetilde{R}^{-\alpha} S_{\widetilde{m}_{0} 0}}{I_{0} + \sigma^{2}}\end{aligned}$$ where we have defined $$S_{x 0} \triangleq \left\{ \begin{array}{ll} |\v_{0}^{\herm} \H_{x 0} \w_{x}|^{2}, & x \in \Phi \\ |\v_{0}^{\herm} \h_{x 0}|^{2}, & x \in (\widetilde{\Phi} \cup \widetilde{m}_{0}) \end{array} \right.$$ and where $I_{0}$ is the overall interference power at the typical FD BS, i.e., $$\begin{aligned} \label{eq:I_1} I_{0} & \triangleq \sum_{x \in \Phi} \big( \widehat{\rho} r_{x}^{-\alpha} S_{x 0} + \widetilde{\rho} r_{\widetilde{m}_{x}}^{-\alpha} S_{\widetilde{m}_{x} 0} \big) + \widehat{\rho} S_{0 0}.\end{aligned}$$ The success probability of the first hop is derived in Section \[sec:SP\_1\]. **Second Hop:** Again, following Slivnyak’s theorem and due to the stationarity of $\widehat{\Phi}$, the statistics of the typical HD DL node’s signal reception are representative of the statistics seen by any HD DL node: we can thus write $\ell (x,\widehat{m}_{0}) = r_{x}^{-\alpha}$, with $r_{x} \triangleq \| x - \widehat{m}_{0} \|$ being the distance of $x$ from the typical HD DL node. Hence, the received signal at the typical HD DL node is given by $$\begin{aligned} \nonumber & \hspace{-15mm} y_{\widehat{m}_{0}} \triangleq \ \underbrace{\sqrt{\widehat{\rho}} \widehat{R}^{-\frac{\alpha}{2}} \h_{0 \widehat{m}_{0}}^{\herm} \w_{0} s_{0}}_{\mathrm{(a)}} + \sum_{x \in \Phi} \underbrace{\sqrt{\widehat{\rho}} r_{x}^{-\frac{\alpha}{2}} \h_{x \widehat{m}_{0}}^{\herm} \w_{x} s_{x}}_{\mathrm{(b)}} \\ \label{eq:y_2} & \hspace{30mm} + \sum_{x \in \Phi} \underbrace{\sqrt{\widetilde{\rho}} r_{\widetilde{m}_{x}}^{-\frac{\alpha}{2}} h_{\widetilde{m}_{x} \widehat{m}_{0}} s_{\widetilde{m}_{x}}}_{\mathrm{(c)}} + \underbrace{\sqrt{\widetilde{\rho}} r_{\widetilde{m}_{0}}^{-\frac{\alpha}{2}} h_{\widetilde{m}_{0} \widehat{m}_{0}} s_{\widetilde{m}_{0}}}_{\mathrm{(d)}} + n_{\widehat{m}_{0}}\end{aligned}$$ where (a) represents the desired signal, (b) and (c) indicate the interference coming from FD BS $x$ and its associated HD UL node $\widetilde{m}_{x}$, respectively, and (d) represents the inter-node interference . The resulting SINR reads as $$\begin{aligned} \label{eq:SINR_2} {\mathrm{SINR}}_{\widehat{m}_{0}} & \triangleq \frac{\widehat{\rho} \widehat{R}^{-\alpha} S_{0 \widehat{m}_{0}}}{I_{\widehat{m}_{0}} + \sigma^{2}}\end{aligned}$$ where we have defined $$S_{x \widehat{m}_{0}} \triangleq \left\{ \begin{array}{ll} |\h_{x \widehat{m}_{0}}^{\herm} \w_{x}|^{2}, & x \in (\Phi \cup 0) \\ |h_{x \widehat{m}_{0}}|^{2}, & x \in (\widetilde{\Phi} \cup \widetilde{m}_{0}) \end{array} \right.$$ and where $I_{\widehat{m}_{0}}$ is the overall interference power at $\widehat{m}_{0}$, i.e., $$\begin{aligned} \label{eq:I_2} I_{\widehat{m}_{0}} \triangleq \sum_{x \in \Phi} \big( \widehat{\rho} r_{x}^{-\alpha} S_{x \widehat{m}_{0}} + \widetilde{\rho} r_{\widetilde{m}_{x}}^{-\alpha} S_{\widetilde{m}_{x} \widehat{m}_{0}} \big) + \widetilde{\rho} r_{\widetilde{m}_{0}}^{-\alpha} S_{\widetilde{m}_{0} \widehat{m}_{0}}.\end{aligned}$$ The success probability of the second hop is derived in Section \[sec:SP\_2\]. For the sake of simplicity, we focus on the interference-limited case, where $I_{0} \gg \sigma^{2}$ and $I_{\widehat{m}_{0}} \gg \sigma^{2}$, and consider the signal-to-interference ratio (SIR). Our analysis can be extended with straightforward yet more involved calculations to the general case. Success Probability {#sec:SP} =================== $$\begin{aligned} \label{eq:P_suc} {\textcolor{black}{{\underline{\mathrm{P}}_{\mathrm{suc}}}(\theta) = {\mathrm{P}_{\mathrm{suc}}}^{(1)} (\theta) {\mathrm{P}_{\mathrm{suc}}}^{(2)} (\theta)}}.\end{aligned}$$ In the rest of the section, we assume that the FD BSs adopt maximum ratio combining (MRC) and maximum ratio transmission (MRT), i.e., the receive combining and transmit beamforming vectors are given by $$\begin{aligned} \v_{x} = \frac{\h_{\widetilde{m}_{x} x}}{\| \h_{\widetilde{m}_{x} x} \|}, \qquad \w_{x} = \frac{\h_{x \widehat{m}_{x}}}{\| \h_{x \widehat{m}_{x}} \|}\end{aligned}$$ respectively. Different combining configurations are considered in Section \[sec:IC\] to study the impact of interference cancellation at the receiver. Success Probability of the First Hop {#sec:SP_1} ------------------------------------ In this section, we analyze the success probability of the first hop ${\mathrm{P}_{\mathrm{suc}}}^{(1)} = \Pr [{\mathrm{SINR}}_{0} > \theta]$, i.e., the probability of successful transmission from the typical HD UL node to the typical FD BS. Considering ${\mathrm{SINR}}_{0}$ in , since MRC is adopted, we have $S_{\widetilde{m}_{0} 0} \sim \chi_{2 N_{{\mathrm{R}}}}^{2}$ (desired signal) and $S_{x 0} \sim \chi_{2}^{2}$, $\forall x \in \Phi \cup \widetilde{\Phi}$ (interferers).[^6] Regarding the power $S_{0 0}$, the following lemma provides a tight approximation of the distribution of the power under Rician fading. \[lem:SI\] [ Let $\v_{x}$, $\w_{x}$, and $\H_{x x}$ be independent; in addition, assume that the (non-normalized) elements of $\v_{x}$ and $\w_{x}$ are distributed independently as $\setC \setN (0,1)$. Then, the power $S_{x x} = |\v_{x}^{\herm} \H_{x x} \w_{x}|^{2}$ approximately follows a gamma distribution, i.e., $S_{x x} \sim \Gamma (a,b)$, with shape parameter $a$ and scale parameter $b$ given by[^7] $$\begin{aligned} \label{eq:ab} a \triangleq \frac{(\mu^{2} + \nu^{2})^{2}}{\eta \mu^{4} + 2 \mu^{2} \nu^{2} + \nu^{4}}, \qquad b \triangleq \frac{\eta \mu^{4} + 2 \mu^{2} \nu^{2} + \nu^{4}}{\mu^{2} + \nu^{2}}\end{aligned}$$ respectively, where $\mu$ and $\nu$ are the and standard deviation, respectively, of the channel $\H_{x x}$ (see ) and where we have defined $$\begin{aligned} \eta \triangleq \frac{4 N_{{\mathrm{R}}} N_{{\mathrm{T}}} - (N_{{\mathrm{R}}}+1) (N_{{\mathrm{T}}}+1)}{(N_{{\mathrm{R}}}+1) (N_{{\mathrm{T}}}+1)}.\end{aligned}$$]{} See Appendix \[sec:A\_SI\]. \[rem:SI\] [ The assumptions of Lemma \[lem:SI\] are very mild. First of all, $\v_{x}$ and $\w_{x}$ are generally chosen as the result of some linear processing of, respectively, $\h_{\widetilde{m}_{x} x}$ and $\h_{x \widehat{m}_{x}}$ (such channels are subject to Rayleigh fading by assumption) followed by power normalization. Note that the assumption of $\v_{x}$ and $\w_{x}$ being MRC and MRT vectors is not required. Besides, $\h_{\widetilde{m}_{x} x}$ and $\h_{x \widehat{m}_{x}}$ are independent on one another and on $\H_{x x}$, and the same holds for $\v_{x}$ and $\w_{x}$. Note that the only practically relevant case where the above assumptions are not satisfied is when $\v_{x}$ (resp. $\w_{x}$) zero-forces the equivalent channel $\H_{x x} \w_{x}$ (resp. $\v_{x}^{\herm} \H_{x x}$): however, this case trivially implies $S_{x x} = 0$ (this scenario is examined in Section \[sec:IC\_2\]).]{} ![PDF of the power for $N_{{\mathrm{R}}} = N_{{\mathrm{T}}} = 4$, $K=1$, and $\Omega = - 60$ dB: empirical histogram versus analytical approximation from Lemma \[lem:SI\].[]{data-label="fig:distr_SI"}](./img/distr_SI) Lemma \[lem:SI\] represents a key result of this paper since it provides a formal characterization of the power experienced by a FD MIMO node with arbitrary receive combining and transmit beamforming vectors. Such distribution of the power is based uniquely on the knowledge of the parameters $K$ and $\Omega$, whose values are available either by design or by measurements, and can be readily incorporated into existing frameworks for the performance analysis of UDNs. Figure \[fig:distr\_SI\] shows the accuracy of the approximated distribution derived in Lemma \[lem:SI\]. The next theorem provides the success probability of the first hop. \[th:P\_suc1\] [ Consider the interference term $I_{0}$ in . The success probability of the first hop is given by $$\begin{aligned} \label{eq:P_suc1} {\mathrm{P}_{\mathrm{suc}}}^{(1)} (\theta) & = \sum_{n=0}^{N_{{\mathrm{R}}} - 1} \bigg[ \frac{(- s)^{n}}{n!} \frac{\mathrm{d}^{n}}{\mathrm{d} s^{n}} \mathcal{L}_{I_{0}}(s) \bigg]_{s = \theta \widetilde{\rho}^{-1} \widetilde{R}^{\alpha}}\end{aligned}$$ where $$\begin{aligned} \label{eq:LI_1} \mathcal{L}_{I_{0}}(s) \triangleq \frac{1}{(1 + s b \widehat{\rho})^{a}} \exp \big( - \lambda \Upsilon(s) \big)\end{aligned}$$ is the Laplace transform of $I_{0}$, where we have defined $$\begin{aligned} \label{eq:Upsilon} \Upsilon (s) \triangleq 2 \pi \int_{0}^{\infty} \bigg( 1 - \frac{1}{1 + s \widehat{\rho} r^{-\alpha}} \Psi(s, r) \bigg) r \mathrm{d} r\end{aligned}$$ with $$\begin{aligned} \label{eq:Psi} \Psi (s, r) \triangleq \frac{1}{2 \pi} \int_{0}^{2 \pi} \frac{\mathrm{d} \varphi}{1 + s \widetilde{\rho} (r^{2} + \widetilde{R}^{2} + 2 r \widetilde{R} \cos \varphi)^{- \frac{\alpha}{2}}}.\end{aligned}$$ ]{} See Appendix \[sec:A\_P\_suc1\_th\]. The array gain resulting from the employment of multiple receive antennas appears evident from Theorem \[th:P\_suc1\]: the larger is $N_{{\mathrm{R}}}$, the more terms are included in the summation of ${\mathrm{P}_{\mathrm{suc}}}^{(1)}(\theta)$ in (the same applies for Theorem \[th:P\_suc2\]); note that all terms in the summation are positive since the $n$-th derivative of $\mathcal{L}_{I_{0}}(s)$ are negative for odd $n$. Expressions of the form of arise frequently, among other cases, when multiple antennas are involved, and are widely used throughout the paper. A useful upper bound for this type of expression is provided in the following proposition.[^8] \[pro:Alzer\] [ For any $\setL_{X}(s^{\prime}) \triangleq \Exp_{X} \big[ e^{-s^{\prime} X} \big]$ and $N > 1$, the following inequality holds: $$\begin{aligned} \label{eq:Alzer} \sum_{n=0}^{N - 1} \bigg[ \frac{(- s)^{n}}{n!} \frac{\mathrm{d}^{n}}{\mathrm{d} s^{n}} \mathcal{L}_{X}(s) \bigg]_{s = s^{\prime}} < \sum_{n=1}^{N} (-1)^{n - 1} {{N}\choose{n}} \setL_{X} \big( n \big( \Gamma(N + 1) \big)^{- \frac{1}{N}} s^{\prime} \big).\end{aligned}$$]{} See Appendix \[sec:A\_Alzer\]. Given the integral form of $\Upsilon (s)$ in , the success probability ${\mathrm{P}_{\mathrm{suc}}}^{(1)}(\theta)$ is not in closed form and needs to be evaluated numerically; nonetheless, we derive the following closed-form lower and upper bounds. \[cor:P\_suc1\] [ The Laplace transform of $I_{0}$ in is bounded as $\mathcal{L}_{I_{0}}(s) \in \big[ \mathcal{L}_{I_{0}}^{(\min)}(s), \mathcal{L}_{I_{0}}^{(\max)}(s) \big]$, with $$\begin{aligned} \label{eq:LI_1min} \mathcal{L}_{I_{0}}^{(\min)}(s) & \triangleq \frac{1}{(1 + s b \widehat{\rho})^{a}} \exp \big( - \lambda \Upsilon^{(\max)}(s) \big), \\ \label{eq:LI_1max} \mathcal{L}_{I_{0}}^{(\max)}(s) & \triangleq \frac{1}{(1 + s b \widehat{\rho})^{a}} \exp \big( - \lambda \Upsilon^{(\min)}(s) \big)\end{aligned}$$ where we have defined $$\begin{aligned} \label{eq:Upsilon_min} \Upsilon^{(\min)}(s) & \triangleq (1 + \tfrac{2}{\alpha}) (\widetilde{\rho}^{\frac{2}{\alpha}} + \widehat{\rho}^{\frac{2}{\alpha}}) \frac{\pi^{2} s^{\frac{2}{\alpha}}}{\alpha \sin \big( \frac{2 \pi}{\alpha} \big)}, \\ \label{eq:Upsilon_max} \Upsilon^{(\max)}(s) & \triangleq 2 (\widetilde{\rho}^{\frac{2}{\alpha}} + \widehat{\rho}^{\frac{2}{\alpha}}) \frac{\pi^{2} s^{\frac{2}{\alpha}}}{\alpha \sin \big( \frac{2 \pi}{\alpha} \big)}.\end{aligned}$$ Then, the lower and upper bounds on the success probability of the first hop ${\mathrm{P}_{\mathrm{suc}}}^{(1)} (\theta)$, are obtained by replacing $\mathcal{L}_{I_{0}}(s)$ in with $\mathcal{L}_{I_{0}}^{(\min)}(s)$ and $\mathcal{L}_{I_{0}}^{(\max)}(s)$, respectively. ]{} See Appendix \[sec:A\_P\_suc1\_cor\]. [ In order to efficiently compute the derivatives of the bounds –, one can resort to the well-known general Leibniz rule for the differentiation of the product of two functions $f(s) g(s)$ [@Abr72 Eq. 3.3.8]: for instance, for $\mathcal{L}_{I_{0}}^{(\min)}(s)$, we can write $f(s) = \frac{1}{(1 + s b \widehat{\rho})^{a}}$ and $g(s) = \exp \big( -\lambda \Upsilon^{(\max)}(s) \big)$. In turn, the derivatives of $g(s)$ can be computed using Faà di Bruno’s formula [@Joh02] for the differentiation of the composition of two functions $g(s) = (g_{1} \circ g_{2})(s)$, with $g_{1}(s) = \exp(s)$ and $g_{2}(s) = - \lambda \Upsilon^{(\max)}(s)$. These considerations apply equivalently to the bounds provided in Corollary \[cor:P\_suc2\].]{} The following corollary provides a sufficient condition under which FD mode outperforms HD mode in terms of spectral efficiency for the case of single receive antenna. \[cor:HD\] [ Consider the first hop assuming that $N_{{\mathrm{R}}}=1$. The achievable spectral efficiency when the BSs operate in FD mode is lower bounded by $$\begin{aligned} \label{eq:T_FD} \mathrm{SE}_{\mathrm{FD}}^{(\min)}(\theta) \triangleq 2 \mathcal{L}_{I_{0}}^{(\min)}(\theta \widetilde{\rho}^{-1} \widetilde{R}^{\alpha}) \log_{2}(1+\theta).\end{aligned}$$ When the BSs operate in HD mode (i.e., when $\widehat{\rho}=0$), the achievable spectral efficiency is given by $$\begin{aligned} \label{eq:T_HD} \mathrm{SE}_{\mathrm{HD}}(\theta) \triangleq \exp \bigg( - \lambda \frac{2 \pi^{2} (\theta \widetilde{R}^{\alpha})^{\frac{2}{\alpha}}}{\alpha \sin \big( \frac{2 \pi}{\alpha} \big)} \bigg) \log_{2}(1+\theta).\end{aligned}$$ Then, $\mathrm{SE}_{\mathrm{FD}}^{(\min)}(\theta) \geq \mathrm{SE}_{\mathrm{HD}}(\theta)$ whenever the density $\lambda$ satisfies $$\begin{aligned} \lambda \leq \frac{\alpha \sin \big( \frac{2 \pi}{\alpha} \big)}{2 \pi^{2} (\theta \widetilde{\rho}^{-1} \widehat{\rho} \widetilde{R}^{\alpha})^{\frac{2}{\alpha}}} \log \bigg( \frac{2}{(1 + b \theta \widetilde{\rho}^{-1} \widehat{\rho} \widetilde{R}^{\alpha})^{a}} \bigg).\end{aligned}$$]{} The proof is straightforward from Theorem \[th:P\_suc1\] and Corollary \[cor:P\_suc1\]. Evidently, if the density $\lambda$ exceeds a certain threshold, using twice the bandwidth in FD mode does not compensate for the additional interference due to the concurrent UL/DL transmissions and, therefore, HD mode becomes optimal. Success Probability of the Second Hop {#sec:SP_2} ------------------------------------- In this section, we analyze the success probability of the second hop ${\mathrm{P}_{\mathrm{suc}}}^{(2)}(\theta) = \Pr [{\mathrm{SINR}}_{\widehat{m}_{0}} > \theta]$, i.e., the probability of successful transmission from the typical FD BS to the typical HD DL node. Considering ${\mathrm{SINR}}_{\widehat{m}_{0}}$ in , since MRT is adopted, we have $S_{0 \widehat{m}_{0}} \sim \chi_{2 N_{{\mathrm{T}}}}^{2}$ (desired signal) and $S_{x \widehat{m}_{0}} \sim \chi_{2}^{2}$, $\forall x \in \Phi \cup \widetilde{\Phi}$ (interferers). The success probability of the second hop is given next in Theorem \[th:P\_suc2\], whereas its lower and upper bounds are provided in Corollary \[cor:P\_suc2\]. \[th:P\_suc2\] [ Consider the interference term $I_{\widehat{m}_{0}}$ in . The success probability of the second hop is given by $$\begin{aligned} \label{eq:P_suc2} {\mathrm{P}_{\mathrm{suc}}}^{(2)} (\theta) = \sum_{n=0}^{N_{{\mathrm{T}}} - 1} \bigg[ \frac{(- s)^{n}}{n!} \frac{\mathrm{d}^{n}}{\mathrm{d} s^{n}} \mathcal{L}_{I_{\widehat{m}_{0}}}(s) \bigg]_{s = \theta \widehat{\rho}^{-1} \widehat{R}^{\alpha}}\end{aligned}$$ where $$\begin{aligned} \label{eq:LI_2} \mathcal{L}_{I_{\widehat{m}_{0}}}(s) \triangleq \Psi(s, \widehat{R}) \exp \big( -\lambda \Upsilon(s) \big)\end{aligned}$$ is the Laplace transform of $I_{\widehat{m}_{0}}$, with $\Upsilon(s)$ and $\Psi (s, r)$ defined in and in , respectively. ]{} See Appendix \[sec:A\_P\_suc2\_th\]. \[cor:P\_suc2\] [ The Laplace transform of $I_{\widehat{m}_{0}}$ in is bounded as $\mathcal{L}_{I_{\widehat{m}_{0}}}(s) \in \big[ \mathcal{L}_{I_{\widehat{m}_{0}}}^{(\min)}(s), \mathcal{L}_{I_{\widehat{m}_{0}}}^{(\max)}(s) \big]$, with $$\begin{aligned} \label{eq:LI_2min} \hspace{-1mm} \mathcal{L}_{I_{\widehat{m}_{0}}}^{(\min)}(s) & \triangleq \frac{1}{1 + s \widetilde{\rho} | \widetilde{R} - \widehat{R} |^{-\alpha}} \exp \big( - \lambda \Upsilon^{(\max)}(s) \big), \\ \label{eq:LI_2max} \hspace{-1mm} \mathcal{L}_{I_{\widehat{m}_{0}}}^{(\max)}(s) & \triangleq \frac{1}{1 + s \widetilde{\rho} (\widetilde{R} + \widehat{R})^{-\alpha}} \exp \big( - \lambda \Upsilon^{(\min)}(s) \big)\end{aligned}$$ with $\Upsilon^{(\min)}(s)$ and $\Upsilon^{(\max)}(s)$ defined in and in , respectively. Then, the lower and upper bounds on the success probability of the second hop ${\mathrm{P}_{\mathrm{suc}}}^{(2)} (\theta)$, are obtained by replacing $\mathcal{L}_{I_{\widehat{m}_{0}}}(s)$ in with $\mathcal{L}_{I_{\widehat{m}_{0}}}^{(\min)}(s)$ and $\mathcal{L}_{I_{\widehat{m}_{0}}}^{(\max)}(s)$, respectively. ]{} See Appendix \[sec:A\_P\_suc2\_cor\]. Observe that the bounds and are more accurate when $\widetilde{R} \gg \widehat{R}$ or $\widetilde{R} \ll \widehat{R}$ due to the presence of the first multiplicative term. This condition is easily verified, for instance, when the HD UL node (resp. HD DL node) is a backhaul node and the HD DL node (resp. HD UL node) is a mobile UE, with the former being likely much farther away from the FD small-cell BS with respect to the latter. Interference Cancellation {#sec:IC} ========================= In the previous section, we have considered a MRC/MRT configuration at the FD BSs. This section analyzes interference cancellation at the receive side of both the FD BSs and the HD DL nodes; for the latter, we further extend our analysis to the case of HD DL nodes with multiple receive antennas. We consider PZF, which represents an efficient and low-complexity spatial interference cancellation technique for multi-antenna receivers [@Atz16b; @Jin08]. If a node is equipped with $N$ receive antennas, the PZF receiver allows to cancel $M \leq N - 1$ interference contributions while using the remaining degrees of freedom to boost the desired received signal.[^9] Observe that, when $M=0$, the PZF receiver reduces to the MRC case analyzed in Section \[sec:SP\_1\]. Focusing on the first hop, we consider two possible receive configurations: - Each FD BS cancels the interference coming from the $M$ nearest FD BSs (cf. Section \[sec:IC\_1\]); - Each FD BS cancels the (cf. Section \[sec:IC\_2\]). Observe that the above configurations can be also combined, e.g., by simultaneously cancelling the $M-1$ nearest nodes and the . Focusing on the second hop, we assume multiple receive antennas at the HD DL nodes and consider the following receive configuration: - Each HD DL node cancels the inter-node interference (cf. Section \[sec:IC\_3\]). First Hop: Cancelling the Nearest $M$ FD BSs {#sec:IC_1} -------------------------------------------- In interference-limited scenarios, it is often beneficial to cancel the interference coming from a certain number of surrounding nodes. Let us assume that the points of $\Phi$ are indexed such that their distances from the typical FD BS is in increasing order, i.e., $\{ r_{x_{i}} \leq r_{x_{i+1}} \}_{i=1}^{\infty}$, and let us suppose that each FD BS adopts PZF at the receiver to cancel its $M$ nearest FD BSs. The resulting overall interference power at the typical FD BS is given by (cf. ) $$\begin{aligned} \label{eq:I_1_M} I_{0}^{\textnormal{\tiny{PZF-$M$}}} & \triangleq \sum_{\substack{x_{i} \in \Phi \\ i > M}} \widehat{\rho} r_{x_{i}}^{-\alpha} S_{x_{i} 0} + \sum_{x_{i} \in \Phi} \widetilde{\rho} r_{\widetilde{m}_{x_{i}}}^{-\alpha} S_{\widetilde{m}_{x_{i}} 0} + \widehat{\rho} S_{0 0}.\end{aligned}$$ The success probability of the first hop with PZF is given in the next theorem. \[th:P\_suc1\_M\] [ Consider the interference term $I_{0}^{\textnormal{\tiny{PZF-$M$}}}$ in . The success probability of the first hop is given by $$\begin{aligned} \label{eq:P_suc1_M} {\mathrm{P}_{\mathrm{suc}}}^{(1)}(\theta) = \sum_{n=0}^{N_{{\mathrm{R}}} - M - 1} \bigg[ \frac{(- s)^{n}}{n!} \frac{\mathrm{d}^{n}}{\mathrm{d} s^{n}} \mathcal{L}_{I_{0}^{\textnormal{\tiny{PZF-$M$}}}}(s) \bigg]_{s = \theta \widetilde{\rho}^{-1} \widetilde{R}^{\alpha}}\end{aligned}$$ where $$\begin{aligned} \label{eq:L_I1_M} \mathcal{L}_{I_{0}^{\textnormal{\tiny{PZF-$M$}}}}(s) \triangleq \frac{1}{(1 + s b \widehat{\rho})^{a}} \Exp_{\Phi} \bigg[ \prod_{\substack{x_{i} \in \Phi \\ i \leq M}} \frac{1}{1 + s \widetilde{\rho} r_{\widetilde{m}_{x_{i}}}^{-\alpha}} \bigg] \Exp_{\Phi} \bigg[ \prod_{\substack{x_{i} \in \Phi \\ i > M}} \frac{1}{1 + s \widehat{\rho} r_{x_{i}}^{-\alpha}} \frac{1}{1 + s \widetilde{\rho} r_{\widetilde{m}_{x_{i}}}^{-\alpha}} \bigg]\end{aligned}$$ is the Laplace transform of $I_{0}^{\textnormal{\tiny{PZF-$M$}}}$.]{} See Appendix \[sec:A\_P\_suc1\_th\_1\]. The tradeoff between array gain and interference cancellation appears evident from Theorem \[th:P\_suc1\_M\]: the larger is $M$, the larger is $\mathcal{L}_{I_{0}^{\textnormal{\tiny{PZF-$M$}}}}(s)$ in , but also the less terms are included in the summation of ${\mathrm{P}_{\mathrm{suc}}}^{(1)}(\theta)$ in . Note that a similar expression of the success probability can be obtained if the $M$ nearest HD UL nodes are cancelled. Unfortunately, a closed-form expression of $\mathcal{L}_{I_{0}^{\textnormal{\tiny{PZF-$M$}}}}(s)$ in is not available and, in order to obtain a more tractable expression, one can resort to the approximation provided in the following corollary. \[cor:P\_suc1\_M\] [ The Laplace transform of $I_{0}^{\textnormal{\tiny{PZF-$M$}}}$ in can be tightly approximated by $$\begin{aligned} \label{eq:LI_1_M_a} \mathcal{L}_{I_{0}^{\textnormal{\tiny{PZF-$M$}}}}(s) \simeq \frac{1}{(1 + s b \widehat{\rho})^{a}} \exp \big( - \lambda (\Upsilon_{1}(s,M) + \Upsilon_{2}(s,M)) \big)\end{aligned}$$ where we have defined $$\begin{aligned} \label{eq:Upsilon1} \Upsilon_{1}(s, M) & \triangleq 2 \pi \int_{0}^{d_{M}} \big( 1 - \Psi (s, r) \big) r \diff r, \\ \label{eq:Upsilon2} \Upsilon_{2}(s, M) & \triangleq 2 \pi \int_{d_{M}}^{\infty} \bigg( 1 - \frac{1}{1 + s \widehat{\rho} r^{-\alpha}} \Psi (s, r) \bigg) r \diff r\end{aligned}$$ with $\Psi (s, r)$ defined in and $$\begin{aligned} \label{eq:d_M} d_{M} \triangleq (\lambda \pi)^{-\frac{1}{2}} \frac{\Gamma (M + \tfrac{1}{2})}{\Gamma(M)}.\end{aligned}$$]{} The approximation in is obtained using the framework [@Atz16b] where $d_{M}$ in is the average distance between the typical FD BS and its $M$-th nearest FD BS, i.e., $d_{M} = \Exp[r_{x_{M}}]$ [@Hae12 Ch. 2.9.1]. First Hop: Cancelling the {#sec:IC_2} -------------------------- As discussed in Section \[sec:Intro\], a strong greatly reduces the SINR of the received signals and implicitly sets an upper bound on the transmit power of the FD BSs. Hence, in presence of low attenuation, spatial cancellation at the receiver may be necessary to preserve the SINR of the received signal . Suppose that each FD BS adopts PZF at the receiver to suppress its .[^10] The resulting overall interference power at the typical FD BS is given by (cf. ) $$\begin{aligned} \label{eq:I_1_SI} I_{0}^{\textnormal{\tiny{PZF-SI}}} \triangleq \sum_{x \in \Phi} \big( \widehat{\rho} r_{x}^{-\alpha} S_{x 0} + \widetilde{\rho} r_{\widetilde{m}_{x}}^{-\alpha} S_{\widetilde{m}_{x} 0} \big).\end{aligned}$$ The success probability of the first hop with cancellation is given in the next theorem. \[th:P\_suc1\_SI\] [ Consider the interference term $I_{0}^{\textnormal{\tiny{PZF-SI}}}$ in . The success probability of the first hop is given by $$\begin{aligned} \label{eq:P_suc1_SI} {\mathrm{P}_{\mathrm{suc}}}^{(1)}(\theta) & = \sum_{n=0}^{N_{{\mathrm{R}}} - 2} \bigg[ \frac{(- s)^{n}}{n!} \frac{\mathrm{d}^{n}}{\mathrm{d} s^{n}} \exp \big( - \lambda \Upsilon(s) \big) \bigg]_{s = \theta \widetilde{\rho}^{-1} \widetilde{R}^{\alpha}}.\end{aligned}$$ ]{} The expression in can be readily obtained from the proof of Theorem \[th:P\_suc1\] (see Appendix \[sec:A\_P\_suc1\_th\_1\]). Note that removing the greatly simplifies the computation of the success probability. Now, we wish to answer to the following question: *is it better to use one degree of freedom to suppress the or to cancel the nearest interfering FD BS?* This issue is meaningful in cases where the FD BSs can devote no more than one antenna for interference cancellation (e.g., when the density $\lambda$ is very high). A comparative sufficient condition for this choice is provided in the following corollary. \[cor:canc\_UL\] [ Cancelling the is, on average, more beneficial than cancelling the nearest FD BS if $$\begin{aligned} \label{eq:canc_UL} \frac{4 a}{\pi} \log \bigg( 1 + b \theta \frac{\widehat{\rho}}{\widetilde{\rho}} \widetilde{R}^{\alpha} \bigg) \geq {}_{2}F_{1} \bigg( 1, \frac{2}{\alpha}, 1 + \frac{2}{\alpha}, - \frac{\widetilde{\rho}}{\theta \widehat{\rho} (2 \widetilde{R} \sqrt{\lambda})^{\alpha}} \bigg)\end{aligned}$$ where ${}_{2}F_{1}(a,b,c,x)$ denotes the Gauss hypergeometric function [@Gra07 Sec. 9.1].]{} See Appendix \[sec:A\_P\_suc1\_canc1\]. Since the right-hand side of is increasing with $\lambda$, Corollary \[cor:canc\_UL\] formalizes that, on average, the overcomes the interference produced by the nearest FD BS when the density $\lambda$ is below a certain threshold. In fact, the nearest interferer approaches the typical FD BS as the density $\lambda$ increases, and the corresponding average interference power becomes stronger. \[cor:canc\_imp\_CSI\] [ ]{} Second Hop: Cancelling the Inter-Node Interference {#sec:IC_3} -------------------------------------------------- So far we have assumed single-antenna HD DL nodes. In this section, we extend our analysis to the case of HD DL nodes with multiple receive antennas. First, we aim at answering the following question: *is it better to use one degree of freedom to cancel the inter-node interference or to cancel the nearest interfering FD BS?* A comparative sufficient condition for this choice is provided in the following corollary. \[cor:canc\_DL1\] [ Cancelling the inter-node interference is, on average, more beneficial than cancelling the nearest FD BS if $$\begin{aligned} \label{eq:canc_DL1} \frac{4}{\pi} \log \bigg( 1 + \frac{\widetilde{\rho}}{\widehat{\rho}} \bigg( \frac{\widetilde{R}}{\widehat{R}} + 1 \bigg)^{-\alpha} \bigg) \geq {}_{2}F_{1} \bigg( 1, \frac{2}{\alpha}, 1 + \frac{2}{\alpha}, - \frac{\widehat{\rho}}{\theta \widetilde{\rho} (2 \widehat{R} \sqrt{\lambda})^{\alpha}} \bigg).\end{aligned}$$]{} See Appendix \[sec:A\_P\_suc2\_canc1\]. The interpretation of Corollary \[cor:canc\_DL1\] is similar to that of Corollary \[cor:canc\_UL\] in the sense that, on average, the inter-node interference is stronger than the interference produced by the nearest FD BS when the density $\lambda$ is below a certain threshold. Of particular interest is the case where the HD UL/DL nodes are mobile UEs served simultaneously within the same small cell. In a small-cell scenario, the distance between UL and DL UEs is generally very short, which causes a severe inter-node interference at the latter [@Ale16]. Since the vast majority of commercial mobile UEs is currently equipped with two receive antennas, it is meaningful to examine the impact of inter-node interference cancellation at the HD DL nodes. Therefore, assume that each FD BS serves a pair of HD UL/DL mobile UEs, with the latter equipped with two receive antennas; for simplicity, we suppose that the FD BSs transmit with a single antenna (which encompasses the case of multi-antenna FD BSs performing SDMA). The following corollary provides a sufficient condition for cancelling the inter-node interference, where an evident tradeoff arises between the density $\lambda$ and the radius of the small cell. \[cor:canc\_DL2\] [ Suppose that the HD DL node is equipped with two receive antennas and let $\widetilde{R} = \widehat{R} = R$. Cancelling the inter-node interference improves the success probability of the second hop if $$\begin{aligned} \label{eq:canc_DL2} R \leq \frac{\Delta(\theta)}{\sqrt{\lambda}}\end{aligned}$$ where we have defined $$\begin{aligned} \Delta(\theta) \triangleq \theta^{1-\frac{1}{\alpha}} \widetilde{\rho} \widehat{\rho}^{\frac{1}{\alpha}-1} \sqrt{\frac{2^{-(2 \alpha + 1)} \alpha}{\Upsilon^{(\min)}(1) (2^{-\alpha} \theta \widetilde{\rho} \widehat{\rho}^{-1} + 1)}}\end{aligned}$$ with $\Upsilon^{(\min)}(s)$ defined in .]{} See Appendix \[sec:A\_P\_suc2\_canc2\]. [ Fixing $\widetilde{R} = \widehat{R} = R$ in Corollary \[cor:canc\_DL2\] models a scenario where the pair of HD UL/DL mobile UEs are located at opposite edges of a small cell with radius $R$. Hence, the condition in can be interpreted as follows: *i)* if the radius $R$ is lower than a certain threshold, then one receive antenna should be invested to cancel the inter-node interference; *ii)* if the density $\lambda$ exceeds a certain threshold, the interference from the other nodes becomes too strong and the two antennas should be used to enhance the signal reception.]{} Numerical Results and Discussion {#sec:NUM} ================================ In this section, we present numerical results to assess our theoretical findings. In particular, we aim at answering the following questions: *i)* *under which conditions does FD mode yield performance gains with respect to HD mode?* And *ii)* *what is the impact of interference cancellation at both the FD BSs and the (multi-antenna) DL nodes on the network performance?* These points are addressed next in Sections \[sec:NUM\_TG\] and \[sec:NUM\_IC\], respectively. Unless otherwise stated, we focus on the scenario where each acts as a relay between a HD macro-cell backhaul node and a HD mobile UE. The parameters $a$ and $b$ of the are computed according to in Lemma \[lem:SI\], where $\mu$ and $\nu$ are obtained from with Rician $K$-factor $K=1$ (see [@Dua12] for an experimental characterization of $K$) and attenuation $\Omega=-60$ dB. Lastly, the pathloss exponent is $\alpha = 4$ and the SINR threshold is $\theta=0$ dB. Furthermore, it is evident from Figure \[fig:P\_suc\] that employing multiple antennas produces substantial SINR gains and compensates for the additional interference generated in FD mode. ![Success probability: simulation and analytical bounds versus density $\lambda$, with $\Omega=-60$ dB, $\theta=0$ dB, and for different antenna configurations.[]{data-label="fig:P_suc"}](./img/P_suc) Throughput Gain {#sec:NUM_TG} --------------- We now focus our attention on the first hop in order to analyze the feasibility of FD mode. With this objective in mind, we introduce the *minimum throughput gain* as performance metric, which is defined as $$\begin{aligned} \mathrm{TG}^{(\min)}(\theta) \triangleq \frac{\mathrm{SE}_{\mathrm{FD}}^{(\min)}(\theta)}{\mathrm{SE}_{\mathrm{HD}}(\theta)}\end{aligned}$$ with $\mathrm{SE}_{\mathrm{FD}}^{(\min)}(\theta)$ and $\mathrm{SE}_{\mathrm{HD}}(\theta)$ defined in and in , respectively, for the single-antenna case. This metric represents the worst-case gain of FD mode over HD mode in terms of throughput, with $\mathrm{TG}^{(\min)}(\theta) > 1$ indicating that FD mode outperforms the equivalent HD setup. ![Minimum throughput gain of FD mode over HD mode in the first hop against the SINR threshold $\theta$, with $\lambda=10^{-4}$ BSs/m$^{2}$, $\Omega=-60$ dB, and for different numbers of receive antennas.[]{data-label="fig:TG_theta"}](./img/TG_Omega) ![Minimum throughput gain of FD mode over HD mode in the first hop against the SINR threshold $\theta$, with $\lambda=10^{-4}$ BSs/m$^{2}$, $\Omega=-60$ dB, and for different numbers of receive antennas.[]{data-label="fig:TG_theta"}](./img/TG_theta) Notably, the analytical tools presented in Section \[sec:SP\_1\] allow to evaluate the effect of multiple receive antennas in mitigating the , which represents a crucial issue in FD communications. Figure \[fig:TG\_Omega\] plots the minimum throughput gain against the attenuation $\Omega$ with $\lambda=10^{-4}$ and $\theta=0$ dB. Hence, we have $\mathrm{TG}^{(\min)}(\theta) > 1$ even for moderate values of the attenuation, namely: $\Omega \leq -47$ dB for $N_{{\mathrm{R}}}=N_{{\mathrm{T}}}=1$, $\Omega \leq -43$ dB for $N_{{\mathrm{R}}}=N_{{\mathrm{T}}}=2$, $\Omega \leq -35$ dB for $N_{{\mathrm{R}}}=N_{{\mathrm{T}}}=8$, and $\Omega \leq -29$ dB for $N_{{\mathrm{R}}}=N_{{\mathrm{T}}}=32$ dB. On the other hand, the minimum throughput gain is analyzed in Figure \[fig:TG\_theta\] as a function of the SINR threshold $\theta$ with $\lambda=10^{-4}$ and $\Omega=-60$ dB: in this respect, it is shown that FD mode improves the performance with respect to HD mode for any reasonable value of $\theta$. Interference Cancellation {#sec:NUM_IC} ------------------------- Lastly, we consider interference cancellation at both the FD BSs and at the (multi-antenna) HD DL nodes and analyze its impact on the network performance. In doing so, we make use of the tools developed for PZF receivers proposed in Section \[sec:IC\]. Focusing on the first hop, assume that the FD BSs are equipped with $N_{{\mathrm{R}}} = 2$ receive antennas and that they employ PZF to cancel either the or the nearest FD BS. Figure \[fig:canc\_M\] plots the success probability of the first hop against the density $\lambda$ and compares the above scenarios with the case of no interference cancellation. On the one hand, when the attenuation is low (i.e., $\Omega=-50$ dB), whereas MRC is the best option for higher densities; on the other hand, for a higher attenuation (i.e., $\Omega=-80$ dB), it is always better to use both antennas for array gain. Hence, since the array gain obtained with just two antennas is significant with respect to the single-antenna case (see Figure \[fig:P\_suc\]), one should always exploit both antennas for boosting the desired received signal unless $\Omega$ is very low. ![Success probability of the second hop considering DL nodes with two receive antennas against the density $\lambda$, with $\theta=0$ dB and for two different values of the radius of the small cell $R$.[]{data-label="fig:canc_DL"}](./img/canc_M) ![Success probability of the second hop considering DL nodes with two receive antennas against the density $\lambda$, with $\theta=0$ dB and for two different values of the radius of the small cell $R$.[]{data-label="fig:canc_DL"}](./img/canc_DL) Consider now the scenario described in Section \[sec:IC\_3\], where each single-antenna FD BS serves a pair of HD UL/DL mobile UEs located at opposite edges of a small cell with radius $R$, with the DL UE equipped with two receive antennas. Figure \[fig:canc\_DL\] plots the success probability of the second hop against the density $\lambda$ with inter-node interference cancellation using PZF and compares it with the MRC case. Note that the crossing point between the corresponding curves can be recovered exactly from Corollary \[cor:canc\_DL2\]. As expected, suppressing the inter-node interference becomes detrimental at high densities, where both antennas at the mobile UE should be used for boosting the desired received signal. Recall that this represents a best-case scenario and inter-node interference cancellation becomes even more desirable for random UE locations within the small cell (see Remark \[rem:SI\]). Conclusions {#sec:END} =========== In this paper, we investigate the success probability and spectral efficiency performance of full-duplex (FD) multiple-input multiple-output (MIMO) small-cell networks using tools from stochastic geometry. In particular, the use of the extra degrees of freedom – brought by multiple antennas – for either desired signal power increase or interference cancellation is studied. Simulation results show the beneficial effect of multiple antennas in mitigating the additional interference introduced by FD mode and demonstrate the feasibility of FD technology in practical scenarios even for moderate values of the attenuation. In this respect, partial zero forcing is shown to be a promising antenna processing technique for beneficial FD operation. Success Probability of the First Hop {#success-probability-of-the-first-hop} ==================================== Proof of Lemma \[lem:SI\] {#sec:A_SI} ------------------------- In this appendix, we derive the approximate distribution of the power $S_{x x}$; for notational simplicity, in the following we omit the sub-indices in the beamforming vectors and in the channel matrix and write $S_{x x} \triangleq |\v^{\herm} \H \w|^{2}$. Let $\v \triangleq (v_{i})_{i=1}^{N_{{\mathrm{R}}}}$, $\w \triangleq (w_{j})_{j=1}^{N_{{\mathrm{T}}}}$, and $\H \triangleq \big( (h_{i j})_{i=1}^{N_{{\mathrm{R}}}} \big)_{j=1}^{N_{{\mathrm{T}}}}$. Hence, assuming that $\v$, $\w$, and $\H$ are independent (see Remark \[rem:SI\]), we can write $$\begin{aligned} \label{eq:vHw} |\v^{\herm} \H \w|^{2} = \sum_{i,k=1}^{N_{{\mathrm{R}}}} \sum_{j,\ell=1}^{N_{{\mathrm{T}}}} v_{i}^{*} v_{k} h_{i j} h_{k \ell}^{*} w_{j} w_{\ell}^{*}.\end{aligned}$$ Then, building on the central limit theorem for causal functions [@Pap62], we can approximate a sum of positive random variables $X = \sum_{i} X_{i}$ using the gamma distribution with shape and scale parameters given by $$\begin{aligned} \label{eq:gamma_ab} a = \frac{(\Exp[X])^{2}}{\Var[X]}, \qquad b = \frac{\Var[X]}{\Exp[X]}\end{aligned}$$ respectively, with $\Var[X]$ denoting the variance of the random variable $X$. First of all, we outline the statistics of $\H$, $\v$, and $\w$. Recalling that , we have $\Exp \big[ |h_{i j}|^{2} \big] = \mu^{2} + \nu^{2}$ and $\Exp \big[ |h_{i j}|^{4} \big] = \mu^{4} + 4 \mu^{2} \nu^{2} + 2 \nu^{4}$. On the other hand, since $\| \v \|^{2} = \| \w \|^{2} = 1$, we have $\Exp \big[ |v_{i}|^{2} \big] = \frac{1}{N_{{\mathrm{R}}}}$ and $\Exp \big[ |w_{j}|^{2} \big] = \frac{1}{N_{{\mathrm{T}}}}$. Furthermore, we can write $v_{i} = \frac{\bar{v}_{i}}{\| \bar{\v} \|}$ and $w_{j} = \frac{\bar{w}_{j}}{\| \bar{\w} \|}$, where $\bar{v}_{i}, \bar{w}_{j} \sim \setC \setN (0,1)$ are the non-normalized coefficients of $\v$ and $\w$ (see Remark \[rem:SI\]), respectively, with $\bar{\v} \triangleq (\bar{v}_{i})_{i=1}^{N_{{\mathrm{R}}}}$ and $\bar{\w} \triangleq (\bar{w}_{j})_{j=1}^{N_{{\mathrm{T}}}}$. It follows that $|v_{i}|^{4} = \frac{|\bar{v}_{i}|^{4}}{\| \bar{\v} \|^{4}}$, where $$\begin{aligned} \| \bar{\v} \|^{4} & = (\bar{\v}^{\herm} \bar{\v})^{2} = \sum_{i=1}^{N_{{\mathrm{R}}}} |\bar{v}_{i}|^{4} + \sum_{\substack{i,j=1 \\ i \neq j}}^{N_{{\mathrm{R}}}} |\bar{v}_{i}|^{2} |\bar{v}_{j}|^{2}, \\ \Exp \big[ \| \bar{\v} \|^{4} \big] & = N_{{\mathrm{R}}} \Exp \big[ |\bar{v}_{i}|^{4} \big] + N_{{\mathrm{R}}} (N_{{\mathrm{R}}} - 1) \big( \Exp \big[ |\bar{v}_{i}|^{2} \big] \big)^{2} = N_{{\mathrm{R}}} (N_{{\mathrm{R}}} + 1), \\ \Exp \big[ |v_{i}|^{4} \big] & = \frac{\Exp \big[ |\bar{v}_{i}|^{4} \big]}{N_{{\mathrm{R}}} (N_{{\mathrm{R}}} + 1)} = \frac{2}{N_{{\mathrm{R}}} (N_{{\mathrm{R}}} + 1)}, \\ \Exp \big[ |v_{i}|^{2} |v_{j}|^{2} \big] & = \frac{\big( \Exp \big[ |\bar{v}_{i}|^{2} \big] \big)^{2}}{N_{{\mathrm{R}}} (N_{{\mathrm{R}}} + 1)} = \frac{1}{N_{{\mathrm{R}}} (N_{{\mathrm{R}}} + 1)}\end{aligned}$$ and, likewise, from $|w_{j}|^{4} = \frac{|\bar{w}_{j}|^{4}}{\| \bar{\w} \|^{4}}$, we have $\Exp \big[ |w_{j}|^{4} \big] = \frac{2}{N_{{\mathrm{T}}} (N_{{\mathrm{T}}} + 1)}$ and $\Exp \big[ |w_{i}|^{2} |w_{j}|^{2} \big] = \frac{1}{N_{{\mathrm{T}}} (N_{{\mathrm{T}}} + 1)}$. In order to obtain the parameters of the gamma function introduced in , we need to derive the second and fourth moments of $|\v^{\herm} \H \w|$. Let us define $\Sigmab \triangleq \diag (\sigma_{i})_{i=1}^{N_{\min}}$, where $\sigma_{i}$ denotes the $i$-th singular value of $\H$ and $N_{\min} \triangleq \min(N_{{\mathrm{R}}}, N_{{\mathrm{T}}})$: recalling and the above properties of the beamforming vectors, we obtain $$\begin{aligned} \label{eq:vHw1} \Exp \big[ |\v^{\herm} \H \w|^{2} \big] & = \Exp \bigg[ \sum_{i=1}^{N_{{\mathrm{R}}}} \sum_{j=1}^{N_{{\mathrm{T}}}} |v_{i}|^{2} |h_{i j}|^{2} |w_{j}|^{2} \bigg] = \frac{1}{N_{{\mathrm{R}}} N_{{\mathrm{T}}}} \sum_{i=1}^{N_{{\mathrm{R}}}} \sum_{j=1}^{N_{{\mathrm{T}}}} \Exp \big[ |h_{i j}|^{2} \big] = \mu^{2} + \nu^{2}\end{aligned}$$ and $$\begin{aligned} \nonumber \Exp \big[ |\v^{\herm} \H \w|^{4} \big] & = \Exp \bigg[ \sum_{i=1}^{N_{\min}} |v_{i}|^{4} |\sigma_{i}|^{4} |w_{j}|^{4} \bigg] + 2 \Exp \bigg[ \sum_{\substack{i,j=1 \\ i \neq j}}^{N_{\min}} |v_{i}|^{2} |v_{j}|^{2} |\sigma_{i}|^{2} |\sigma_{j}|^{2} |w_{i}|^{2} |w_{j}|^{2} \bigg] \\ \nonumber & = \frac{2}{N_{{\mathrm{R}}} {N_{{\mathrm{T}}} (N_{{\mathrm{R}}} + 1) (N_{{\mathrm{T}}} + 1)}} \bigg( 2 \Exp \bigg[ \sum_{i=1}^{N_{\min}} |\sigma_{i}|^{4} \bigg] + \Exp \bigg[ \sum_{\substack{i,j=1 \\ i \neq j}}^{N_{\min}} |\sigma_{i}|^{2} |\sigma_{j}|^{2} \bigg] \bigg) \\ \nonumber & = \frac{2}{N_{{\mathrm{R}}} {N_{{\mathrm{T}}} (N_{{\mathrm{R}}} + 1) (N_{{\mathrm{T}}} + 1)}} \bigg( \Exp \bigg[ \sum_{i=1}^{N_{\min}} |\sigma_{i}|^{4} \bigg] + \Exp \bigg[ \sum_{i,j=1}^{N_{\min}} |\sigma_{i}|^{2} |\sigma_{j}|^{2} \bigg] \bigg) \\ \nonumber & = \frac{2}{N_{{\mathrm{R}}} {N_{{\mathrm{T}}} (N_{{\mathrm{R}}} + 1) (N_{{\mathrm{T}}} + 1)}} \Big( \Exp \big[ \tr \big( (\H^{\herm} \H)^{2} \big) \big] + \Exp \big[ \big( \tr (\H^{\herm} \H) \big)^{2} \big] \Big) \\ \label{eq:vHw4} & = \frac{4 N_{{\mathrm{R}}} N_{{\mathrm{T}}}}{(N_{{\mathrm{R}}} + 1) (N_{{\mathrm{T}}} + 1)} \mu^{4} + 4 \mu^{2} \nu^{2} + 2 \nu^{4}\end{aligned}$$ the fact that $\Exp \big[ \sum_{i,j} |\sigma_{i}|^{2} |\sigma_{j}|^{2} \big] = \Exp \big[ \sum_{i} |\sigma_{i}|^{2} \big] \Exp \big[ \sum_{j} |\sigma_{j}|^{2} \big]$ and $$\begin{aligned} \sum_{i=1}^{N_{\min}} |\sigma_{i}|^{2} = \tr (\Sigmab^{\herm} \Sigmab) = \tr (\H^{\herm} \H), \qquad \sum_{i=1}^{N_{\min}} |\sigma_{i}|^{4} = \tr \big( (\Sigmab^{\herm} \Sigmab)^{2} \big) = \tr \big( (\H^{\herm} \H)^{2} \big)\end{aligned}$$ $$\begin{aligned} \vspace{1mm} \Exp \big[ \big( \tr (\H^{\herm} \H) \big)^2 \big] & \! = \! N_{{\mathrm{R}}} N_{{\mathrm{T}}} \Big( \Exp \big[ |h_{i j}|^{4} \big] \! + \! (N_{{\mathrm{R}}} N_{{\mathrm{T}}} - 1) \Exp \big[ |h_{i j}|^{2} \big] \Big), \\ \vspace{1mm} \Exp \big[ \tr \big( (\H^{\herm} \H)^{2} \big) \big] & \! = \! N_{{\mathrm{R}}} N_{{\mathrm{T}}} \Big( \Exp \big[ |h_{i j}|^{4} \big] \! + \! (N_{{\mathrm{R}}} + N_{{\mathrm{T}}} + 2) \big( \Exp \big[ |h_{i j}|^{2} \big] \big)^{2} \! \! + \! (N_{{\mathrm{R}}} - 1)(N_{{\mathrm{T}}} - 1) \big( \Exp [h_{i j}] \big)^{4} \Big).\end{aligned}$$ Since $\Var \big[ |\v^{\herm} \H \w|^{2} \big] = \Exp \big[ |\v^{\herm} \H \w|^{4} \big] - \big( \Exp \big[ |\v^{\herm} \H \w|^{2} \big] \big)^{2}$, we readily obtain $a$ and $b$ in by applying . This concludes the proof. Proof of Theorem [\[th:P\_suc1\]]{} {#sec:A_P_suc1_th} ----------------------------------- The success probability of the first hop is given by $$\begin{aligned} {\mathrm{P}_{\mathrm{suc}}}^{(1)} (\theta) = \Pr \bigg[ \frac{\widetilde{\rho} \widetilde{R}^{-\alpha} S_{\widetilde{m}_{0} 0}}{I_{0}} > \theta \bigg] = \Pr \big[ S_{\widetilde{m}_{0} 0} > \theta \widetilde{\rho}^{-1} \widetilde{R}^{\alpha} I_{0} \big] = \Exp_{I_{0}} \big[ \bar{F}_{S_{\widetilde{m}_{0} 0}} \big( \theta \widetilde{\rho}^{-1} \widetilde{R}^{\alpha} I_{0} \big) \big]\end{aligned}$$ where $I_{0}$ is defined in and denotes the overall interference at the typical FD BS. Since the latter is equipped with $N_{{\mathrm{R}}}$ receive antennas, the power of its desired signal is distributed as $S_{\widetilde{m}_{0} 0} \sim \chi_{2 N_{{\mathrm{R}}}}^{2}$: hence, our case falls into the general framework [@Hun08] and ${\mathrm{P}_{\mathrm{suc}}}^{(1)}(\theta)$ in results from applying (see footnote \[fn:chi2N\]) $$\begin{aligned} \Exp_{I_{0}} \big[ \bar{F}_{S_{\widetilde{m}_{0} 0}} \big( s I_{0} \big) \big] = \Exp_{I_{0}} \bigg[ e^{- s I_{0}} \sum_{n=0}^{N_{\mathrm{R}}-1} \frac{\big( s I_{0} \big)^{n}}{n!} \bigg] = \sum_{n=0}^{N_{\mathrm{R}}-1} \bigg[ \frac{(-s)^{n}}{n!} \frac{\diff^{n}}{\diff s^{n}} \setL_{I_{0}}(s) \bigg].\end{aligned}$$ On the other hand, building on [@Ton15 Th. 1], the Laplace transform of $I_{0}$ is obtained as $$\begin{aligned} \nonumber \mathcal{L}_{I_{0}}(s) & = \Exp \big[ e^{-s I_{0}} \big] \\ \nonumber & = \Exp \big[ \exp (- s \widehat{\rho} S_{0 0}) \big] \Exp \bigg[ \prod_{x \in \Phi} \exp \big( - s (\widehat{\rho} r_{x}^{-\alpha} S_{x 0} + \widetilde{\rho} r_{\widetilde{m}_{x}}^{-\alpha} S_{\widetilde{m}_{x} 0}) \big) \bigg] \\ \nonumber & = \frac{1}{(1 + s b \widehat{\rho})^{a}} \Exp_{\Phi} \bigg[ \prod_{x \in \Phi} \frac{1}{1 + s \widehat{\rho} r_{x}^{-\alpha}} \frac{1}{1 + s \widetilde{\rho} r_{\widetilde{m}_{x}}^{-\alpha}} \bigg] \\ \label{eq:LI_1a} & = \frac{1}{(1 + s b \widehat{\rho})^{a}} \exp \bigg( - \lambda \int_{\Real^{2}} \bigg( 1 - \frac{1}{1 + s \widehat{\rho} r_{x}^{-\alpha}} \frac{1}{1 + s \widetilde{\rho} r_{\widetilde{m}_{x}}^{-\alpha}} \bigg) \diff x \bigg)\end{aligned}$$ Finally, the expression in follows from $$\begin{aligned} \label{eq:m_tilde1} r_{\widetilde{m}_{x}} = \| x + \widetilde{R} (\cos \varphi, \sin \varphi) \| = \sqrt{r_{x}^{2} + \widetilde{R}^{2} + 2 r_{x} \widetilde{R} \cos \varphi}\end{aligned}$$ with $\varphi$ uniformly distributed in $[0, 2\pi]$. This completes the proof. Proof of Proposition [\[pro:Alzer\]]{} {#sec:A_Alzer} -------------------------------------- For any random variable $X$ and $N > 1$, we have that (see footnote \[fn:chi2N\]) $$\begin{aligned} \label{eq:Alzer1} \sum_{n=0}^{N - 1} \bigg[ \frac{(- s)^{n}}{n!} \frac{\mathrm{d}^{n}}{\mathrm{d} s^{n}} \mathcal{L}_{X}(s) \bigg]_{s = s^{\prime}} = 1 - \frac{\Exp_{X} \big[ \gamma(N,s^{\prime} X) \big]}{\Gamma(N)}\end{aligned}$$ Now, we use Alzer’s inequality [@Alz97], by which $$\begin{aligned} \frac{\gamma(N,x)}{\Gamma(N)} > (1- e^{-c x})^{N}\end{aligned}$$ with $c \triangleq \big( \Gamma(N+1) \big)^{-\frac{1}{N}}$ for $N > 1$. Now, expanding the expectation term in , we have $$\begin{aligned} \nonumber \frac{\Exp_{X} \big[ \gamma(N, s^{\prime} X) \big]}{\Gamma(N)} & > \Exp_{X} \big[ (1 - e^{- c s^{\prime} X})^{N} \big] \\ \nonumber & = \Exp_{X} \bigg[ \sum_{n=0}^{N} (-1)^{n} \binom{N}{n} e^{- n c s^{\prime} X} \bigg] \\ \label{eq:Alzer2} & = \sum_{n=0}^{N} (-1)^{n} \binom{N}{n} \setL_{X}(n c s^{\prime}).\end{aligned}$$ , where $-(-1)^{n} = (-1)^{n-1}$ and $\binom{N}{0} \setL_{X}(0)=1$. Proof of Corollary [\[cor:P\_suc1\]]{} {#sec:A_P_suc1_cor} -------------------------------------- Building on [@Ton15 Th. 3], we have that $\Upsilon(s)$ in is bounded as $\Upsilon(s) \in \big[ \Upsilon^{(\min)}(s), \Upsilon^{(\max)}(s) \big]$, with $\Upsilon^{(\min)}(s)$ and $\Upsilon^{(\max)}(s)$ defined in and in , respectively. Then, the lower and upper bounds on $\mathcal{L}_{I_{0}}(s)$ in – readily follow. Success Probability of the Second Hop {#success-probability-of-the-second-hop} ===================================== Proof of Theorem [\[th:P\_suc2\]]{} {#sec:A_P_suc2_th} ----------------------------------- The success probability of the second hop is given by $$\begin{aligned} {\mathrm{P}_{\mathrm{suc}}}^{(2)} (\theta) = \Pr \bigg[ \frac{\widehat{\rho} \widehat{R}^{-\alpha} S_{0 \widehat{m}_{0}}}{I_{\widehat{m}_{0}}} > \theta \bigg] = \Pr \big[ S_{0 \widehat{m}_{0}} > \theta \widehat{\rho}^{-1} \widehat{R}^{\alpha} I_{\widehat{m}_{0}} \big] = \Exp_{I_{\widehat{m}_{0}}} \big[ \bar{F}_{S_{0 \widehat{m}_{0}}} \big( \theta \widehat{\rho}^{-1} \widehat{R}^{\alpha} I_{\widehat{m}_{0}} \big) \big]\end{aligned}$$ where $I_{\widehat{m}_{0}}$ is defined in and denotes the overall interference at the typical HD DL node. Since the typical FD BS is equipped with $N_{{\mathrm{T}}}$ transmit antennas, the power of the desired signal is distributed as $S_{0 \widehat{m}_{o}} \sim \chi_{2 N_{{\mathrm{T}}}}^{2}$ and the expression in is obtained following similar steps as in Appendix \[sec:A\_P\_suc1\_th\]. On the other hand, building again on [@Ton15 Th. 1], the Laplace transform of $I_{\widehat{m}_{0}}$ is obtained as $$\begin{aligned} \nonumber \mathcal{L}_{I_{\widehat{m}_{0}}}(s) & = \Exp \big[ e^{-s I_{\widehat{m}_{0}}} \big] \\ \label{eq:LI_2a} & = \Exp \big[ \exp (- s \widetilde{\rho} r_{\widetilde{m}_{0}}^{-\alpha} S_{\widetilde{m}_{0} \widehat{m}_{0}}) \big] \Exp \bigg[ \prod_{x \in \Phi} \exp \big( - s (\widehat{\rho} r_{x}^{-\alpha} S_{x \widehat{m}_{0}} + \widetilde{\rho} r_{\widetilde{m}_{x}}^{-\alpha} S_{\widetilde{m}_{x} \widehat{m}_{0}}) \big) \bigg]\end{aligned}$$ where $\Exp \big[ \exp (- s \widetilde{\rho} r_{\widetilde{m}_{0}}^{-\alpha} S_{\widetilde{m}_{0} \widehat{m}_{0}}) \big] = \Psi(s,\widehat{R})$ from the moment-generating function of the exponential distribution and , . This concludes the proof. Proof of Corollary [\[cor:P\_suc2\]]{} {#sec:A_P_suc2_cor} -------------------------------------- Given the definition of $\Psi(s,r)$ in , it is not difficult to find the following bounds: $$\begin{aligned} \label{eq:Psi_bounds} \Psi(s,r) \in \bigg[ \frac{1}{1 + s \widetilde{\rho} |r - \widetilde{R}|^{-\alpha}} , \frac{1}{1 + s \widetilde{\rho} (r + \widetilde{R})^{-\alpha}} \bigg].\end{aligned}$$ Then, the lower and upper bounds on $\mathcal{L}_{I_{\widehat{m}_{0}}}(s)$ in – are a straightforward result of combining and Corollary \[cor:P\_suc1\]. Interference Cancellation {#interference-cancellation} ========================= Proof of Theorem \[th:P\_suc1\_M\] {#sec:A_P_suc1_th_1} ---------------------------------- The proof follows similar steps as in Appendix \[sec:A\_P\_suc1\_th\]. By applying PZF, the typical FD BS uses $N_{{\mathrm{R}}}-M$ receive antennas to match its desired received signal and, therefore, the power of the latter is distributed as $S_{\widetilde{m}_{0} 0} \sim \chi_{2 (N_{{\mathrm{R}}}-M)}^{2}$, yielding the expression in (see [@Atz16b] for details). On the other hand, the Laplace transform of $I_{0}^{\textnormal{\tiny{PZF-$M$}}}$ in is obtained by removing the interference contribution of the first $M$ FD BSs and $$\begin{aligned} \label{eq:L_I1_M1} \Exp_{\Phi} \bigg[ \prod_{\substack{x_{i} \in \Phi \\ x_{i} > M}} \frac{1}{1 + s \widehat{\rho} r_{x_{i}}^{-\alpha}} \prod_{x_{i} \in \Phi} \frac{1}{1 + s \widetilde{\rho} r_{\widetilde{m}_{x_{i}}}^{-\alpha}} \bigg] = \Exp_{\Phi} \bigg[ \prod_{\substack{x_{i} \in \Phi \\ x_{i} \leq M}} \frac{1}{1 + s \widehat{\rho} r_{\widetilde{m}_{x_{i}}}^{-\alpha}} \prod_{\substack{x_{i} \in \Phi \\ x_{i} > M}} \frac{1}{1 + s \widehat{\rho} r_{x_{i}}^{-\alpha}} \frac{1}{1 + s \widetilde{\rho} r_{\widetilde{m}_{x_{i}}}^{-\alpha}} \bigg].\end{aligned}$$ Finally, the two products in the expectation on the right-hand side of are independent and can be thus separated. Proof of Corollary \[cor:canc\_UL\] {#sec:A_P_suc1_canc1} ----------------------------------- The average power is larger than the average interference power from the nearest FD BS if $$\begin{aligned} \label{eq:canc_UL1} \frac{1}{(1 + s b \widehat{\rho})^{a}} \leq \Exp \bigg[ \frac{1}{1 + s \widehat{\rho} r_{x_{1}}^{-\alpha}} \bigg]\end{aligned}$$ where $$\begin{aligned} \nonumber \Exp \bigg[ \frac{1}{1 + s \widehat{\rho} r_{x_{1}}^{-\alpha}} \bigg] & = 2 \pi \lambda \int_{0}^{\infty} \frac{1}{1 + s \widehat{\rho} r^{-\alpha}} e^{- \pi \lambda r^{2}} r \diff r \\ \label{eq:canc_UL2} & \simeq \exp \bigg( - 2 \pi \lambda \int_{0}^{d_{1}} \bigg( 1 - \frac{1}{1 + s \widehat{\rho} r^{-\alpha}} \bigg) r \diff r \bigg).\end{aligned}$$ Proof of Corollary \[cor:canc\_imp\_CSI\] {#sec:A_P_suc1_canc2} ----------------------------------------- $$\begin{aligned} {\textcolor{black}{\v_{x} = \frac{ \big( \I_{N_{{\mathrm{R}}}} - \hat{\H}_{x x} \w_{x} (\hat{\H}_{x x} \w_{x})^{\sharp} \big) \h_{\widetilde{m}_{x} x}}{\| (\I_{N_{{\mathrm{R}}}} - \hat{\H}_{x x} \w_{x} (\hat{\H}_{x x} \w_{x})^{\sharp}) \h_{\widetilde{m}_{x} x} \|}}}\end{aligned}$$ $$\begin{aligned} {\textcolor{black}{\Exp \big[ |\v^{\herm} \E \w|^{2} \big] \leq \Exp \bigg[ \sum_{i=1}^{N_{{\mathrm{R}}}} \sum_{j=1}^{N_{{\mathrm{T}}}} |v_{i}|^{2} |h_{i j}|^{2} |w_{j}|^{2} \bigg] = \frac{1}{N_{{\mathrm{R}}} N_{{\mathrm{T}}}} \sum_{i=1}^{N_{{\mathrm{R}}}} \sum_{j=1}^{N_{{\mathrm{T}}}} \Exp \big[ |e_{i j}|^{2} \big] = \epsilon^{2}}}\end{aligned}$$ Proof of Corollary \[cor:canc\_DL1\] {#sec:A_P_suc2_canc1} ------------------------------------ The average inter-node interference power is larger than the average interference power from the nearest FD BS if $$\begin{aligned} \Psi(s,\widehat{R}) \leq \Exp \bigg[ \frac{1}{1 + s \widehat{\rho} r_{x_{1}}^{-\alpha}} \bigg].\end{aligned}$$ Then, we use the upper bound of $\Psi(s,r)$ in and $s = \theta \widehat{\rho}^{-1} \widehat{R}^{\alpha}$, and the rest of the proof follows similar steps as in Appendix \[sec:A\_P\_suc1\_canc1\]. Proof of Corollary \[cor:canc\_DL2\] {#sec:A_P_suc2_canc2} ------------------------------------ Recall the definition of $\setL_{I_{\widehat{m}_{0}}} (s)$ in . Considering the scenario where the FD BSs transmit with one antenna to HD DL nodes with two receive antennas, we can build on Theorem \[th:P\_suc2\] to write the success probabilities of the second hop with and without inter-node interference cancellation as $$\begin{aligned} \mathrm{P}_{\mathrm{suc},1}^{(2)} (\theta) & \triangleq \exp \big( - \lambda \Upsilon(s) \big), \\ \mathrm{P}_{\mathrm{suc},2}^{(2)} (\theta) & \triangleq \bigg[ \setL_{I_{\widehat{m}_{0}}} (s) - s \frac{\diff}{\diff s} \setL_{I_{\widehat{m}_{0}}} (s) \bigg]_{s = \theta \widehat{\rho}^{-1} \widehat{R}^{\alpha}}\end{aligned}$$ respectively. Now, let us fix $\widetilde{R} = \widehat{R} = R$ and let us consider the upper bound on $\Psi(s,r)$ in and $\Upsilon^{(\min)} (s)$ in . After some algebraic manipulations, we have that $\mathrm{P}_{\mathrm{suc},1}^{(2)} (\theta) \geq \mathrm{P}_{\mathrm{suc},2}^{(2)} (\theta)$ when the condition in is satisfied. [^1]: Part of this work was presented at GLOBECOM 2015 in San Diego, CA, USA [@Atz15a] and at Asilomar 2016 in Pacific Grove, CA, USA [@Atz16b]. [^2]: The authors are with the Mathematical and Algorithmic Sciences Lab, Paris Research Center, Huawei Technologies Co., Ltd., France (email: [email protected]; [email protected]). [^3]: Before applying active cancellation, the magnitude of the channel is modeled using a Rician distribution with large $K$-factor due to the strong line-of-sight component; after applying active cancellation, the line-of-sight component is reduced, resulting in smaller $K$-factor [@Dua12]. [^4]: We assume that the FD BSs adopt a perfect decode-and-forward relaying scheme; the study of imperfect schemes goes beyond the scope of this paper. [^5]: [^6]: \[fn:chi2N\]We define a random variable $X \sim \chi_{2 N}^{2}$ to have probability density function (PDF) $f_{X}(x) = \frac{x^{N-1} e^{-x}}{\Gamma(N)}$; its CCDF is given by $\bar{F}_{X}(x) = 1 - \frac{\gamma(N,x)}{\Gamma(N)} = e^{-x} \sum_{n=0}^{N-1} \frac{x^{n}}{n!}$. [^7]: We define a random variable $X \sim \Gamma(a,b)$ with shape parameter $a$ and scale parameter $b$ to have PDF $f_{X}(x) = \frac{x^{a-1} e^{-x/b}}{b^a \Gamma(a)}$. [^8]: A lower bound with a similar expression can be also obtained; however, such bound is usually not sufficiently tight and it is thus not considered. [^9]: [^10]: Instead of nulling the completely, one can adopt the partial cancellation proposed in [@Atz16] at the receiver to enhance the UL throughput.

ArXiv

--- abstract: 'The stationary state of stochastic processes such as reaction-diffusion systems can be related to the ground state of a suitably defined quantum Hamiltonian. Using this analogy, we investigate the applicability of a real space renormalisation group approach, originally developped for quantum spin systems, to interacting particle systems. We apply the technique to an exactly solvable reaction-diffusion system and to the contact process (both in $d=1$). In the former case, several exact results are recovered. For the contact process, surprisingly good estimates of critical parameters are obtained from a small-cell renormalisation.' author: - | Jef Hooyberghs[^1] , Carlo Vanderzande\ Departement WNI, Limburgs Universitair Centrum\ 3590 Diepenbeek, Belgium\ title: '**[Real space renormalisation for reaction-diffusion systems]{}**' ---  \ Introduction ============ Reaction-diffusion systems and other interacting particle systems are relevant for the description of several phenomena in physics, chemistry and biology [@VK]. In the past, they have been mainly modelled by (non-linear) partial differential equations [@MB], a description which implicitly contains a mean-field assumption. Such a description is however no longer appropriate in low dimensions where fluctuations are important. To take these into account one turns to a description of the reaction-diffusion system in terms of a stochastic process. This can for example be realised by adding a noise term to the partial differential equation. In recent years however, particular attention has been paid to models defined on a lattice. It has been found that these can be related to a number of interesting topics in modern statistical mechanics such as growing interfaces [@KPZ], phase transitions into an absorbing state [@DR], exactly solvable quantum spin chains [@SH], persistence exponents [@DerPer] and so on. The systems which we will study here are defined on a lattice but evolve in continuous time. At each site of the lattice, one can have (hard core) ‘particles’ which can perform a random walk and/or can undergo one or several ‘reactions’. In this paper we will limit ourselves to systems with one type of particle . Each lattice site can then be either empty (${\emptyset}$) or be occupied by a particle (${\mbox{A}}$). As an example, consider a system in which particles perform random walks and where two particles on neighbouring sites can ‘annihilate’ (i.e. undergo the reaction ${\mbox{A}}+{\mbox{A}}\to {\emptyset}+ {\emptyset}$). In a simple mean field approach the density of particles $c(t)$ in this system decays asymptotically as $1/t$. It is common to introduce a critical exponent $\theta$ which describes the decay of the density ($c(t) \sim t^{-\theta}$) and which therefore in mean field theory equals $1$. An exact solution of the diffusion-annihilation model in $d=1$ (where it is equivalent to the $T=0$ Glauber dynamics of an Ising model on the dual lattice) shows however that $\theta=1/2$ [@AF]. This latter value is also found experimentally in systems which are thought to be described by the diffusion-annihilation model. Moreover, $\theta$ shows a large universality across materials and initial conditions. The same value for $\theta$ is also found when the experimental situation is described more correctly by a diffusion-coagulation (${\mbox{A}}+{\mbox{A}}\to {\emptyset}+ {\mbox{A}}, {\mbox{A}}+{\mbox{A}}\to {\mbox{A}}+{\emptyset}$) model [@exp]. Hence, as in the theory of equilibrium critical phenomena one needs a scheme which at the same time explains the observed universality and gives precise values for critical exponents such as $\theta$. One is therefore naturally led to a search for renormalisation group (RG) approaches to stochastic systems. Mathematically, the stochastic process is a continuous time Markov process. Using the formal equivalence between the master equation and the Schrödinger equation in imaginary time, one can set up a field theoretic formulation for stochastic systems which in turn can be used in the construction of a renormalisation approach [@LC]. Critical exponents can then be calculated in an $\epsilon$-expansion around the upper critical dimension, which in this case equals $2$. This approach is by now well established and has been applied to many interesting systems. As an example, we mention the much studied branching and annihilating random walks (BARW) [@BAW], where besides diffusion and annihilation, particles can undergo branching processes (${\mbox{A}}\to (m+1){\mbox{A}}$). The competition between annihilation and branching leads to a non-equilibrium phase transition where the stationary state particle density ${c_{st}}(\equiv \lim_{t \to \infty} c(t))$ goes to zero in a continuous way as a function of the rates of the different processes. Near this transition, several critical exponents (static and dynamic) can be introduced. It has been shown, both numerically and using the field theoretic RG, that the universality class of the transition for BARW is completely determined by the parity of $m$ [@CT]. For $m$ odd, the transition falls in the universality class of directed percolation, whereas for $m$ even a new universality class appears. On the basis of very precise simulations, Jensen [@Jensen] conjectured that $\theta=2/7$ and $\beta=13/14$ when $m$ is even ($\beta$ describes the way ${c_{st}}$ goes to zero near the transition). So far, a precise analytical calculation of these exponents has however not been possible. Menyhárd and Ódor propose the value $\beta=1$ on the basis of a perturbation around mean field theory [@MO]. Using a loop expansion at fixed dimension, Cardy and Täuber find $\beta=4/7$ [@CT]. Thus, there is clearly room for the introduction of new, analytical approaches to the BARW and related models. In this paper, we investigate the possibilities of a real space RG approach to interacting particle systems. Our starting point is again the equivalence between a stochastic system and a quantum mechanical one. In the past, several real space RG approaches to quantum lattice systems have been introduced. We must mention here as an example the [*density matrix renormalisation group*]{} (DMRG) which has been very succesful [@White1]. Currently, several research teams in the world are investigating the applicability of the DMRG to reaction-diffusion systems [@CH],[@Pdmrg]. Unfortunately, the name DMRG is a bit a misnomer, since one rarely calculates RG flows in this approach and hence it is not easy to decide on questions of universality using this technique. Moreover, the approach is purely numerical. Instead in this work, we will use an appropriate perturbative extension of the so called standard (also called SLAC) approach [@SLAC1]. We have applied this RG method to the study of the stationary state properties of some interacting particle systems. We found that the method works surprisingly well in several cases which we studied. These include a solvable reaction diffusion model (with diffusion, coagulation and decoagulation) and the contact process, which is also thought to be in the directed percolation universality class. This paper is organised as follows. In section II we briefly introduce the description of stochastic processes in a quantum mechanical language. In section III, we discuss a real space approach to the ground state of quantum (spin or fermion) chains. Our method is perturbative and in first order coincides with the standard (or SLAC) approach. We discuss how the technique can be derived from more conventional (Niemeijer - Van Leeuwen) RG approaches to quantum lattice systems by taking the zero temperature limit. The method as such is not new but, in our opinion, not very well known. We also discuss how critical exponents for stochastic systems can be obtained from such an approach. In section IV we study properties of the stationary state of some solvable reaction-diffusion systems. In section V, we give our results for the contact process. Finally, in section VI we present some conclusions and an outlook on further applications of the real space RG technique. Quantum formalism of reaction-diffusion systems =============================================== In this section we discuss the relation between stochastic processes in continuous time on the one hand and quantum mechanics on the other hand [@Doi],[@Peliti]. This relation has been (independently) discovered by several authors, and is now well established. We only give a brief overview, with the main purpose of fixing our notation. Consider a one dimensional lattice of $L$ sites and let $\eta=\{\eta_1,\ldots,\eta_L\}$ ($\eta_i=0(1)$ when no (a) particle is present at site $i$) be the microscopic configuration of the particle system. Furthermore, we denote by $P(\eta,t)$ the probability that the system is in configuration $\eta$ at time $t$. The time evolution of $P(\eta,t)$ is determined by the transition rates $w_0,w_1,\ldots,w_s$ of the model. In general one of these is used to fix the time scale (say $w_0=1$). We will collect the remaining rates in a vector $\vec{w}=(w_1,\ldots,w_s)$. Next, we introduce a state vector $|P(t)\rangle$ $$\begin{aligned} |P(t)\rangle = \sum_\eta P(\eta,t) |\eta\rangle\end{aligned}$$ The time evolution of $|P(t)\rangle$ is described by the master equation $$\begin{aligned} \frac{d|P(t)\rangle}{dt}= - H |P(t)\rangle \label{1.1}\end{aligned}$$ where $H$ is a $(2^L \times 2^L)$ matrix which depends on all the transition rates of the system. In particular the matrix element $-H_{\zeta,\xi} \geq 0$ (for $\zeta \neq \xi$) equals the transition rate to go from configuration $\zeta$ to configuration $\xi$. Due to the conservation of probability the diagonal elements are given by $$\begin{aligned} H_{\zeta,\zeta}= - \sum_{\xi \neq \zeta} H_{\xi,\zeta} \label{1.2}\end{aligned}$$ This relation implies that the sum of the elements in a column of $H$ equals zero, a condition which we will refer to as the [*stochasticity condition*]{}. It is now common to consider (\[1.1\]) as a Schrödinger equation (in imaginary time) and call $H$ the Hamiltonian. In contrast to the quantum mechanical situation $H$ is not necessarily hermitian. Hence, left and right eigenvectors are in generally not related by simple transposition. In situations where there are only local interactions we can write $H$ in terms of local operators $H=\sum_i H_i$. For reaction-diffusion systems with one type of particle, it is common to interpret the model as a spin model (where a spin up (down) corresponds the absence (presence) of a particle) so that one writes $H$ in terms of Pauli-matrices or in terms of the matrices $E^{\alpha\beta}$ where $(E^{\alpha\beta})_{\sigma,\tau}=\delta_{\alpha\sigma}\delta_{\beta\tau}$. The stochasticity condition implies the existence of at least one eigenvalue which equals zero and an associated left eigenvector $\langle s|$ given by $$\begin{aligned} \langle s| =\sum_\eta \langle \eta|\end{aligned}$$ Moreover, it can be shown that the real part of all the eigenvalues is non-negative. The formal solution of the master equation (\[1.1\]) is $$\begin{aligned} |P(t)\rangle = e^{-Ht} |P(0)\rangle\end{aligned}$$ The expectation value of an operator, say $X$, corresponding with a physical quantity (such as the density of particles) can then be written as $$\begin{aligned} \langle X(t) \rangle &=& \sum_\eta X(\eta) P(\eta,t) \nonumber \\ &=& \langle s| X |P(t)\rangle \nonumber\\ &=& \langle s| X e^{-Ht} |P(0)\rangle \label{1.3}\end{aligned}$$ In this paper, we will be mainly interested in stationary state properties of the stochastic system. In that limit, (\[1.3\]) is determined by the groundstate(s) $|0\rangle$ of $H$. In particular, (\[1.3\]) becomes $$\begin{aligned} X_{st}=\lim_{t\to\infty}\langle X(t)\rangle = \langle s| X |0\rangle \label{1.3.1}\end{aligned}$$ in the case of a non-degenerate ground state $|0\rangle$. In summary, when we want to study properties in the stationary state of a stochastic system we are in need of a (real space) RG which is suitable for the ground state of quantum spin systems. In the next section, such an approach will be presented. Ground state renormalisation for quantum spin systems ===================================================== It is convenient to remind the reader briefly on the history of real space RG approaches to quantum systems. In a first development, approaches working at finite temperature were introduced [@RD],[@ST]. These were direct extensions of the first real space RG developped by Niemeijer and Van Leeuwen [@NVL]. The SLAC approach was later introduced as an independent approach to the ground state of quantum lattice systems. It was at first claimed to be non-perturbative. In a little known paper [@SVD], it was however shown that the SLAC approach can be obtained from the finite temperature technique by applying a suitable perturbation expansion (see below) to first order and by sending the temperature $T \to 0$. We now briefly describe the two methods and their relation. As will be explained below, the possibility to turn the SLAC approach into a perturbative technique is of importance for the study of some stochastic systems. The SLAC or standard approach was introduced in [@SLAC1]. It was used a lot in the early eighties to study ground state properties of several quantum spin and fermion chains (for a review, see [@PfJ]). As usual in real space RG methods, the lattice is divided into cells, each containing $b$ sites. The Hamiltonian $H$ is then divided into an intracell part $H_{0}$ and an intercell part $V$. If $\alpha$ labels the cells, one can in the particular case of one dimension write $$\begin{aligned} H = \sum_{\alpha} \left( H_{0,\alpha} + V_{\alpha,\alpha+1} \right)\end{aligned}$$ As a first step, the Hamiltonian within one cell $H_{0,\alpha}$ is diagonalised exactly. Let $H_{0,\alpha}$ have eigenvalues $E_{n,\alpha}$ with corresponding right (left) eigenvectors $|n\rangle_\alpha (_\alpha\langle n|)$. One then selects two low lying eigenstates (for example the ground state $|0\rangle_\alpha$ and the first excited state $|1\rangle_\alpha$) and considers them as effective spin states for the cell: $|+_\alpha\rangle'=|0\rangle_\alpha$ and $|-_\alpha\rangle'=|1 \rangle_\alpha$. Renormalised lattice states $|\sigma\rangle'$ can then be constructed by making tensor products over all cells: $|\sigma\rangle'=\otimes_{\alpha} |\sigma_\alpha\rangle'$. These states span a $2^{L/b}$ dimensional vector space ${\cal W}$. The renormalisation transformation, which always involves an elimination of degrees of freedom, is now performed by projecting the original Hamiltonian onto ${\cal W}$. Mathematically this is achieved by introducing a $2^L \times 2^{L/b}$-matrix $T_2$ whose columns contain the vectors $|\sigma\rangle'$ together with a $2^{L/b} \times 2^L$-matrix $T_1$ whose rows contain the vectors $'\langle \sigma|$ (which are constructed from the left eigenvectors of $H_{0,\alpha}$). Then the renormalised Hamiltonian $H'={\cal R}(H)$ is defined as $$\begin{aligned} H'={\cal R}(H) = T_1 H T_2 \label{2.1}\end{aligned}$$ The transformation (\[2.1\]) will define a mapping in the parameter space $\vec{w}'=f(\vec{w})$, from which fixed points and critical exponents can be determined as we will explain further below. The procedure which was defined above at first sight seems to be rather [*ad hoc*]{} and non-perturbative. Further insight in the method was obtained when its relation with more conventional real space RG approaches was discovered [@SVD]. These approaches work at finite temperatures and are a direct extension of the Niemeijer-Van Leeuwen real space approach to the case of quantum spin systems. In such an approach, the eigenstates of $H_{0,\alpha}$ are divided into two groups according to some criterion (for example, a majority rule). Each group is associated with an effective spin state for the cell. Within each group there are $2^{b-1}$ states. Hence, cell states can be denoted as $|\sigma_\alpha,q_\alpha\rangle$, where $\sigma_\alpha=\pm 1,\ q_\alpha=1,\ldots,2^{b-1}$. Tensor products of these states will be denoted by $|\sigma,q\rangle=\otimes_{\alpha} |\sigma_\alpha,q_\alpha\rangle$. Finally let ${\cal H}=-H/(kT)$ be the reduced Hamiltonian. The matrix elements of the renormalised (reduced) Hamiltonian $\cal{H}'$ are then obtained by performing a partial trace $$\begin{aligned} '\langle \sigma|e^{{\cal H}'}|\sigma'\rangle' = Tr_q\ \langle \sigma,q|e^{{\cal H}}|\sigma',q\rangle \label{2.1.b}\end{aligned}$$ Usually, it is impossible to work out this mapping exactly. In any explicit calculation, it is therefore necessary to perform an expansion in the intercell Hamiltonian $V$. This is achieved by using the Feynman identity $$\begin{aligned} e^{\cal H}= e^{{\cal H}_0} \ T_{\lambda}\left[ \exp{\left( \int_{0}^{1} e^{-\lambda {\cal H}_0} {\cal V} e^{\lambda{\cal H}_0}d\lambda\right)}\right] \label{2.2}\end{aligned}$$ where $T_{\lambda}$ is a time ordering operator and ${\cal H}_0$ and ${\cal V}$ are respectively the reduced intracell and intercell Hamiltonian. A similar expansion is made on the left hand side of (\[2.1.b\]). In reference [@SVD] it was shown that if one expands (\[2.2\]) [*to first order*]{} in $V$, and then takes the limit $T \to 0$, one recovers the SLAC approach. This relation is useful for several reasons. First and most importantly, it shows that the SLAC approach is perturbative and one obtains a procedure on how to calculate higher order corrections. Secondly, since the finite temperature RG is constructed to conserve the partition function, one is garanteed that the SLAC approach conserves the ground state energy. Finally, this relation allows a consistent study of finite and zero temperature properties of a quantum system in thermal equilibrium. Going back to reaction-diffusion systems, which are non-equilibrium systems, we are thus garanteed that our RG approach conserves the ground state energy. In principle, we also have a recipe to calculate perturbative corrections to the SLAC approach. In this paper, we will limit ourselves to calculations in first order. Higher order corrections usually lead to a proliferation of terms in the Hamiltonian. However, as we will discuss in our conclusions, the existence of these higher order terms will be necessary for a proper study of the BARW with $m=2$, or for any other model in which $H$ contains terms involving three or more sites. While in this paper we will only study stationary state properties (corresponding to ground state properties of $H$), it is our feeling that the finite temperature extension of the SLAC-approach will be useful to study finite $t$ properties of the reaction-diffusion systems. We now turn to a discussion of how stationary state properties can be determined from the RG mapping $\vec{w}'=f(\vec{w})$. To fix ideas, let us assume that this equation has a non-trivial fixed point at $\vec{w}=\vec{w^\star}$, with one relevant scaling field (which in linear approximation is proportional to $\Delta w_1=w_1-w_1^\star$) whose scaling dimension is $y_{w_1}$. From standard RG theory it then follows that near criticality, the correlation length $\xi$ will diverge as $\xi \sim |\Delta w_1|^{\nu_\perp}$ with $\nu_\perp=1/y_{w_1}$. In general it will be so that after renormalisation $w'_0 \neq 1$. Hence time needs to be rescaled again, which is achieved by dividing $H'$ by $w'_0$. The number $w'_0(\vec{w}^\star)$ therefore teaches us how time rescales under a rescaling of space. We can use this quantity to calculate the exponent $z$ as $$\begin{aligned} w'_0(\vec{w}^\star)=b^{-z} \label{2.2.1}\end{aligned}$$ From $\nu_\perp$ and $z$, scaling [@ScSP] gives us the exponent $\nu_\parallel=\nu_\perp/z$ which determines the divergence of the relaxation time near the critical point. Next, we turn to the calculation of the particle density ${c_{st}}$. For a translationally invariant system, we can, using (\[1.3.1\]), write ${c_{st}}$ as $$\begin{aligned} {c_{st}}= \langle s| E^{11} |0\rangle \end{aligned}$$ which under the RG transforms (\[2.1\]) transforms as $$\begin{aligned} {c_{st}}&=&\langle s|T_1^{-1} T_1 E^{11} T_2 T_2^{-1}|0\rangle\\ &=& '\langle s| (E^{11})'|0\rangle' \label{2.3}\end{aligned}$$ The simplest possible case is the one in which the renormalisation of the operator $E^{11}$ does not involve other operators so that $(E^{11})'=T_1 E^{11} T_2=a(\vec{w}) E^{11}$. In that case (\[2.3\]) becomes (where we now explicitly denote the dependence of ${c_{st}}$ on the transition rates) $$\begin{aligned} {c_{st}}(\vec{w})= a(\vec{w}) {c_{st}}(\vec{w}') \label{2.4}\end{aligned}$$ This relation can be iterated along the RG-flow, and hence the density of particles can be obtained as an infinite product if one knows the density at the fixed point $\vec{w}_t^{\star}$ which attracts $\vec{w}$ (where the $t$ reflects the fact that this attractive fixed point is trivial and not critical). $$\begin{aligned} {c_{st}}(\vec{w})= \left[\prod_{i=0}^{\infty} a(\vec{w}^{(i)})\right] {c_{st}}(\vec{w}_t^{\star}) \label{2.4.b}\end{aligned}$$ In principle correlation functions can be calculated in a similar way. Near the critical fixed point $\vec{w}^{\star}$, we get from (\[2.4\]) for the singular part of ${c_{st}}$ to leading order in $\Delta w_1$ $$\begin{aligned} {c_{st}}(\Delta w_1) = a(\vec{w}^{\star}) {c_{st}}(b^{y_{w_1}}\Delta w_1) \label{2.5}\end{aligned}$$ From this relation we see that $a(\vec{w}^{\star})$ determines the rescaling of the particle density at criticality. We write $a(\vec{w}^{\star})=b^{D-d}$ where $D$ can then be interpreted as the fractal dimension of the sites that are occupied by particles at criticality. Finally, from (\[2.5\]) we get the behaviour of the particle density close to the critical point as $$\begin{aligned} {c_{st}}(\Delta w_1) \sim \left(\Delta w_1\right)^{(d-D)/y_{w_1}} \label{2.6}\end{aligned}$$ so that we obtain the scaling relation $\beta=(d-D)\nu_\perp$. Finally, $\theta$ can be obtained from the scaling relation $\theta=\beta/\nu_\parallel$ [@ScSP]. It is straightforward to extend these reasonings to the case where the transformation of $E^{11}$ is more complicated. We thus see that a complete characterisation of the stationary state particle density and of all the critical properties of the system can be obtained from our RG approach. In the next two sections we test our method on simple stochastic systems. The first one is an exactly solvable reaction-diffusion system with a trivial transition. The second one is the well known contact process which has a non trivial transition thought to be in the DP universality class. Renormalisation for a simple reaction-diffusion system ====================================================== We consider a model with diffusion, decoagulation (which is the process $A+{\emptyset}\to A+A$, ${\emptyset}+A\to A+A$) and coagulation. We will use the diffusion rate to fix the time scale, so that our model has two independent rates which we will denote as $w_1$ (decoagulation) and $w_2$ (coagulation). The local Hamiltonian $H_i$ in this case can most conveniently be written as a $4\times 4$-matrix $$\begin{aligned} H_i = \left(\begin{array}{rrrr} 0 & 0 & 0 & 0\\ 0 & 1+w_1 & -1 &-w_2\\ 0 & -1 & 1+w_1 &-w_2\\ 0 & -w_1 & -w_1 &2w_2\end{array}\right) \label{3.4}\end{aligned}$$ or in terms of the matrices $E^{\alpha\beta}$ as $$\begin{aligned} H_i &=& E^{11}_i E^{00}_{i+1} + E^{00}_i E^{11}_{i+1} + 2w_2(E^{11}_i E^{11}_{i+1}) + w_1(E^{11}_i E^{00}_{i+1} + E^{00}_i E^{11}_{i+1}) \nonumber \\ &-& E^{01}_i E^{10}_{i+1} - E^{10}_i E^{01}_{i+1} - w_2(E^{01}_i E^{11}_{i+1} + E^{11}_i E^{01}_{i+1}) -w_1(E^{11}_i E^{10}_{i+1} + E^{10}_i E^{11}_{i+1}) \label{4.1}\end{aligned}$$ With a similarity transformation this Hamiltonian can be mapped onto that of a free fermion system [@MSimL], from which many exact results can be obtained. To renormalise this model, we take $b=2$, so that $H_{0,\alpha}$ is exactly given by (\[4.1\]). The spectrum of $H_i$ can be calculated trivially. The ground state is doubly degenerate and we identify the corresponding eigenstates as effective spin states. In particular, we have (we use spin language where $|+\rangle$ ($|-\rangle$) denotes a vacancy (an $A$-particle)) $$\begin{aligned} |+\rangle' &=& |++\rangle \nonumber\\ '\langle +|&=&\langle ++| \nonumber\\ |-\rangle'&=& \frac{1}{N} \left[ |+-\rangle + |-+\rangle + r|--\rangle \right] \nonumber \\ '\langle -|&=&\langle +-| + \langle -+| + \langle--|\end{aligned}$$ where $r=w_1/w_2$ and $N=2+r$. Notice that we normalise the states as probability vectors (and not quantum mechanically). Next, we calculate $H'$ using (\[2.1\]). We find that the renormalised Hamiltonian contains the same terms as (\[4.1\]). Moreover, the transformation conserves the stochasticity condition. The renormalised diffusion rate is $1/N$. We divide the Hamiltonian by this factor (rescaling of time). Then $H'$ has completely the same form as $H$ but with renormalised values for $w_1$ and $w_2$. The RG equations for $w_1$ and $w_2$ are $$\begin{aligned} w_1' = w_1 \frac{1+w_1+w_2}{w_2} \label{4.2.a}\\ w_2'= w_2 \frac{1+w_1+w_2}{2w_2 + w_1} \label{4.2.b}\end{aligned}$$ The flow generated by these equations is shown in figure 1. There is a fixed point at $w_2=1,w_1=0$ (pure coagulation fixed point). The line $w_2=1$ is an invariant line. The RG equations, linearised at the fixed point, have the relevant eigenvalue $y_{w_1}=1$, together with a marginal eigenvalue in the $w_2$ direction. (12,8) (0,0)[(1,0)[12]{}]{} (0,0)[(0,1)[8]{}]{} (-0.6,7.7)[$w_1$]{}(11.7,-0.4)[$w_2$]{} (3.91,-0.07)[$\bullet$]{}(4,-0.4)[1]{}(4,0)[(0,1)[8]{}]{}(4,4)[(0,1)[0.1]{}]{} (2,0)[(1,0)[0.1]{}]{}(6,0)[(-1,0)[0.1]{}]{}(0,4)[(0,1)[0.1]{}]{} (1,0)(3.8,0)(3.8,8)(3.56,4)[(1,4)[0.01]{}]{} (0,0)(2,0)(2,8)(1.842,4)[(0,1)[0.1]{}]{} (10,0.5)(4.2,0.5)(4.2,7)(10,0.49)[(1,0)[2]{}]{}(4.19,7)[(0,1)[1]{}]{}(4.62,4)[(-1,4)[0.03]{}]{} (12,1.4)(5,2)(5,8)(6.17,4)[(-1,2)[0.05]{}]{} Next, we turn to a calculation of particle density in the stationary state using the scheme outlined in the previous section. Projecting $E^{11}$ onto $\cal{W}$ we find the simple recursion $$\begin{aligned} (E^{11})' = \frac{1+r}{N} E^{11} \label{4.2.c}\end{aligned}$$ From this we have $$\begin{aligned} a(w_1,w_2) = \frac{1+w_1/w_2}{2+w_1/w_2}\end{aligned}$$ Since also $w_1'/w_2'$ depends only on the ratio $w_1/w_2$ we arrive at the conclusion that the particle density only depends on that ratio. Its precise value can then be calculated by making a product of $a(w_1,w_2)$ along the RG flow. From figure 1 we see that all points with $w_1>0$ flow to points where $w_1/w_2 \to \infty$. In that limit ${c_{st}}=1$. Then, using (\[2.4.b\]) we can obtain the particle density. In figure 2 we plot the result for ${c_{st}}$ as a function of $r$ together with the exact result [@MSimL] $$\begin{aligned} {c_{st}}(r)=\frac{r}{r+1} \label{2.exact}\end{aligned}$$ Within the numerical accuracy, both results coincide. Hence it seems that we recover the exact result! This might seem surprising since our calculation is only precise to first order in $V$. With hindsight the accuracy of our result can be understood by the fact that the ground state of the whole system is a product of one particle states. Nevertheless, this calculation illustrates nicely the use of the method. Moreover it gives an RG-flow for the problem which cannot be obtained in any other way. To conclude this section, we calculate the exponent $\beta$ which determines the behaviour of ${c_{st}}$ near $r=0$. From the exact result (\[2.exact\]), one obviously has $\beta=1$. On the other hand, from (\[4.2.c\]), we have $a(0,1)=1/2$, hence $D=0$ and using $y_{w_1}=1$ we recover the exact result $\beta=1$. At this place it is appropriate to mention that for other simple reaction diffusion systems we can also recover exactly known results. An example is a model with diffusion, decoagulation and death (which is the process $A+{\emptyset}\to{\emptyset}+{\emptyset},\ {\emptyset}+A\to{\emptyset}+{\emptyset}$). This model undergoes a first order transition [@MSimL]. When the decoagulation rate is greater then the death rate, ${c_{st}}=1$, whereas when the opposite inequality holds, one has ${c_{st}}=0$. The RG recovers this exact result. As usual the first order transition is controlled by a discontinuity fixed point. Also for a model with diffusion, annihilation and pair creation ${\emptyset}+{\emptyset}\to A+A$, we recover the exact result, first obtained by Z. Rácz [@Racz], that ${c_{st}}$ grows with the square root of the pair creation rate. The contact proces ================== The contact proces was originally introduced as a simple model for an epidemic [@Harris]. In that interpretation a particle corresponds to sick person, and a vacancy to a healthy individual. In the process particles can disappear ($A\to 0$) with a rate $w_0=1$. Empty sites can become occupied with a rate $\lambda z/2$, where $z$ is the number of occupied neighbours of the empty site (this represents contamination in epidemic terms). This model cannot be solved exactly, but its critical exponents are known to high accuracy [@DR]. On the basis of numerical data, and from symmetry arguments, it is generally believed that the model is in the DP-universality class [@DPJ],[@DPG]. The quantum Hamiltonian corresponding to this model contains one-site terms (for the process $A\to 0$) and two-site terms (for the contamination process). The local Hamiltonian for the model is $$\begin{aligned} H_i = \left(\begin{array}{rrrr} 0 & 0 & -1 & 0\\ 0 & \lambda/2 & 0 & -1\\ 0 & 0 & 1+\lambda/2 & 0\\ 0 & -\lambda/2 & -\lambda/2 &1\end{array}\right) \label{5.1}\end{aligned}$$ We have split the whole Hamiltonian in such a way that $H_i$ contains one two-site contribution and one one-site term. In this way, when performing the RG , both $H_{0,\alpha}$ en $V_{\alpha,\alpha+1}$ will contain an equal number of one-site and two-site terms. This garantees, at least for the contact process, that the ground state of the intracell Hamiltonian is a doublet. These states are then the natural candidates to be used as effective cell spins. We will perform a renormalisation for this model using a cell with $b=3$. The calculation is straightforward and can most effectively be done using $^\copyright$Mathematica. It came as a surprise to us that also in this case, there is no proliferation of interactions in the renormalised Hamiltonian. The RG-equation for $\lambda$ is $$\begin{aligned} \lambda'=\frac{\lambda^3 (2+\lambda) (8+10\lambda+4\lambda^2+\lambda^3)}{4(16+40 \lambda+37 \lambda^2 +18\lambda^3+4\lambda^4)} \label{5.2}\end{aligned}$$ This equation has one non-trivial repulsive fixed point at $\lambda=\lambda^\star=3.22319$. This value is surprisingly close to the best known numerical value for the contact process which is $\lambda_c=3.2978$ (all numerical results are taken from [@JDS]). From linearising the RG-equations near the fixed point, we obtain $y_{w_1}=.8886$, from which we obtain $\nu_\perp=1.1253$, to be compared with the numerically determined value of $1.0972$. As explained in section 3, one can obtain the exponent $z$ from the rescaling of the unit of time. In this case we have $$\begin{aligned} w_0'=\frac{4(16+40\lambda+37 \lambda^2+18\lambda^3+4\lambda^4)}{(8 +10\lambda+4\lambda^2+\lambda^3)^2} \label{5.2.b}\end{aligned}$$ from which we obtain $z=.6858$ (to be compared with $z=.636$). Finally we need to determine $D$. As explained in section 3, one therefore has to renormalise the operator $E^{11}$. In a cell with $b \geq 3$ this can be done in several ways. Either one projects the operator $E^{11}_m$ where $m$ is a site at or near the middle of the cell. Alternatively we can take an ‘average’ operator $E^{11}_a$ $$\begin{aligned} E^{11}_a=\frac{1}{b}\sum_{i=1}^b E^{11}_i\end{aligned}$$ We have performed both calculations. They give respectively $D=.6391$ and $D=.6940$, from which we obtain $\beta=.4061$, respectively $\beta=.3444$. These values should be compared with the precise numerical estimate $\beta=.2769$. Other exponents can be obtained using scaling relations. We get $\nu_\parallel=1.6409, \theta=.2099$ whereas the best known values are $\nu_\parallel=1.736, \theta=.1597$. We thus see that, taking into account the smallness of the cell considered, our estimates of $\lambda_c$ and all critical exponents are close to the known values. In figure 3 we finally plot ${c_{st}}(\lambda)$ as obtained from our RG approach. At the scale of the figure it almost coincides with the results from other approaches. We are currently extending our calculations for the contact process to larger cell sizes. We hope that our approach, combined with suitable extrapolation techniques, is able to give very precise estimates of critical exponents. Conclusions =========== In this paper we have investigated the applicability to reaction-diffusion processes of a quantum real space RG method. We first studied simple processes for which exact information is available. In three cases, these exact results are reproduced in calculations done on very small cells. So far, we encountered one system in which the RG predicted wrong results. This is a model with diffusion, coagulation and birth (${\emptyset}+{\emptyset}\to{\emptyset}+A,\ {\emptyset}+{\emptyset}\to A+{\emptyset}$). When the diffusion rate equals the coagulation rate, this model belongs to a class which is, at least partially, integrable [@PeshelI]. For small birth rates $w_3$, ${c_{st}}$ is known to grow as $w_3^{1/3}$. The application of our technique to that model raised several difficulties. The ground state of $H_{0,\alpha}$ is a singlet, and for the first excited state there is a crossing of energy levels (in finite systems). Hence, there is no obvious choice for the effective cell states. For a specific choice we made, the renormalised Hamiltonian $H'$ turns out to have a form different from the original Hamiltonian $H$, but there is a similarity mapping this $H'$ onto $H$ with renormalised couplings, and an extra decoagulation term. If we define the full RG as the projection (\[2.1\]), followed by the similarity we are able to obtain a flow in parameter space. Unfortunately, our results indicate that for cells with $b=2$ and $b=3$, ${c_{st}}$ grows as $w_3^{1/2}$. We hope to clarify the RG for this model in the future. For the contact process, which contains a non-trivial critical point, rather accurate estimates for the location of the critical point and for critical exponents are obtained from a calculation on a cell of $3$ sites. This justifies the hope that by going to larger cells and using good extrapolation techniques very precise exponent estimates can be obtained. We are currently performing such calculations. Another project we hope to carry through is a study of the BARW with $m=2$. As stated in the introduction, there are few reliable analytical results on the critical properties of this system. A particularly nice model that is known to be in this universality class was introduced by Menyhárd. It is a non-equilibrium Ising model (NEKIM) with Glauber dynamics at zero temperature and Kawasaki dynamics at infinite temperature [@NEKIM]. It was recently shown that this model is selfdual [@MuS]. The ground state energy of the quantum Hamiltonian corresponding with this model is for all finite systems again doubly degenerated, so that effective cell states can be defined unambiguously. The quantum Hamiltonian for the NEKIM contains a three-site interaction term. Under the RG, performed to first order, such a term will be mapped onto a two-cell interaction term. However, by extending the SLAC-approach to second order in $V$, as discussed in section 3, one could generate a renormalised three-cell interaction. In fact, from a mathematical point of view, the Hamiltonian of the NEKIM is rather similar to that of a transverse Ising model with three spin interactions [@TIM31],[@TIM32]. That model also has a selfduality and was renormalised successfully by using our approach to second order [@CVD]. An alternative model in the universality class of the even-$m$-BARWis a recently introduced variant of the contact process in which particles disappear and are contaminated in pairs [@NCP]. Since this model doesn’t involve any diffusion, it may be more simple to analyse. It has of course to be admitted that real space RG methods in general involve some ill understood approximations. Nevertheless, it is our opinion that the results presented here give considerable hope that our RG can succesfully be used to further understand the critical behaviour of non exactly solved systems such as the BARW. [**Acknowledgement**]{} We thank M. Henkel, P. Leoni and G. M. Schütz for some interesting discussions on the subject of this paper. We also thank the IUAP for financial support. One of us (JH) thanks the “Fonds voor Wetenschappelijk Onderzoek - Vlaanderen” for financial support. [Bau89]{} Van Kampen N.G.,[*Stochastic processes in physics and chemistry*]{}, North-Holland (1981) Murray J.D., [*Mathematical biology*]{}, Springer-Verlag (1989) Plischke M., Rácz Z., Liu D., Phys. Rev. B [**35**]{}, 3485 (1987) Dickman R. in [*Nonequilibrium statistical mechanics in one dimension*]{} edited by Privman V., Cambridge University Press (1997) Schütz G.M.,[*Integrable stochastic many-body systems*]{}, Habilitationsthesis, Universität Bonn (1998) Derrida B., Hakim V., Pasquier V., Phys. Rev. Lett. [**75**]{}, 751 (1995) Family F., Amar J.J., J. Stat. Phys [**65**]{} 1235 (1991) Kroon R., Sprik R. in [*Nonequilibrium statistical mechanics in one dimension*]{} edited by Privman V., Cambridge University Press (1997) Lee B.P., J. Phys.A [**27**]{},2633 (1994); Lee B.P., Cardy J.L. , J. Stat. Phys. [**80**]{}, 971 (1995) Takayasu H., Tretyakov A. Yu., Phys. Rev. Lett. [**68**]{}, 3060 (1992) Cardy J.L., Täuber U.C., Phys. Rev. Lett. [**77**]{}, 4780 (1996); Cardy J.L., Täuber U.C., J. Stat. Phys. [**90**]{} 1 (1998) Jensen I., Phys. Rev.E [**50**]{}, 3623 (1994) Menyhárd N., Ódor G., J. Phys.A [**28**]{}, 4505 (1995) White S.R., Phys. Rev.B [**48**]{}, 10345 (1993) Carlon E., Henkel M., Schollwöck U., cond-mat preprint/9902041 Kaulke M., Peshel I., Eur. Phys. J. B [**5**]{}, 727 (1998) Drell S.D., Weinstein M., Yankielowicz S., Phys. Rev.D [**14**]{}, 487 (1976) Doi M., J. Phys.A [**9**]{}, 1479 (1976) Peliti L., J. Phys. (Paris) [**46**]{}, 1469 (1984) Rogiers J., Dekeyser R., Phys. Rev.B [**13**]{}, 4886 (1976) Stella A.L., Toigo F., Phys. Rev.B [**17**]{}, 2343 (1978) Niemeijer Th., van Leeuwen J.M.J. in [*Phase Transitions and critical phenomena*]{} vol.6, edited by Domb C. and Green M.S., Academic Press (1976) Stella A.L., Vanderzande C., Dekeyser R., Phys. Rev.B [**27**]{}, 1812 (1983) Pfeuty P., Jullien R., Penson K.A. in [*Real space renormalisation*]{} edited by Burkhardt Th., van Leeuwen J.M.J., Springer-Verlag (1982) Grassberger P., de la Torre A., Ann. Phys. (NY) [**122**]{}, 373 (1979) Henkel M., Orlandini E., Santos J., Ann. Phys. (NY) [**259**]{}, 163 (1997) Rácz Z., Phys. Rev. Lett. [**55**]{}, 1707 (1985) Harris T.E., Ann. Prob. [**2**]{}, 969 (1974) Janssen H.K., Z. Phys.B [**42**]{}, 151 (1981) Grassberger P., Z. Phys.B [**47**]{}, 365 (1982) Jensen I., Dickman R., J. Stat. Phys. [**71**]{}, 89 (1993) Peshel I., Rittenberg V., Schultze U., Nucl. Phys.B [**430**]{}\[FS\], 633 (1994) Menyhárd N., J. Phys.A [**27**]{}, 6139 (1994) Mussawisade K., Santos J.E., Schütz G.M., J. Phys.A [**31**]{} 4381 (1998) Turban L., J. Phys.C [**15**]{}, L65 (1982); Penson K. A., Jullien R., Pfeuty P., Phys. Rev.B [**26**]{},6334 (1982) Vanderzande C., Iglói F., J. Phys.A [**20**]{}, 4539 (1987) Vanderzande C., [*Ph. D. thesis*]{}, Universiteit Leuven (1984) Inui N., Tretyakov A. Yu., Phys. Rev. Lett. [**80**]{}, 5148 (1998) [^1]: Aspirant Fonds voor Wetenschappelijk Onderzoek - Vlaanderen

ArXiv

--- abstract: 'Despite remarkable success in practice, modern machine learning models have been found to be susceptible to adversarial attacks that make human-imperceptible perturbations to the data, but result in serious and potentially dangerous prediction errors. To address this issue, practitioners often use adversarial training to learn models that are robust against such attacks at the cost of weaker generalization accuracy on unperturbed test sets. The conventional wisdom is that more training data should shrink the generalization gap between adversarially-trained models and standard models. However, we study the training of robust classifiers for both Gaussian and Bernoulli models under $\ell_\infty$ attacks, and we prove that more data may actually increase this gap. Furthermore, our theoretical results identify if and when additional data will finally begin to shrink the gap. Lastly, we experimentally demonstrate that our results also hold for linear regression models, which may indicate that this phenomenon occurs more broadly.' author: - Lin Chen - 'Yifei Min[^1]' - Mingrui Zhang - Amin Karbasi bibliography: - 'reference.bib' title: More Data Can Expand the Generalization Gap Between Adversarially Robust and Standard Models --- Introduction ============ As modern machine learning models continue to gain traction in the real world, a wide variety of novel problems have come to the forefront of the research community. One particularly important challenge has been that of adversarial attacks [@szegedy2013intriguing; @goodfellow2014explaining; @kos2018adversarial; @carlini2018audio]. To be specific, given a model with excellent performance on a standard data set, one can add small perturbations to the test data that can fool the model and cause it to make wrong predictions. What is more worrying is that these small perturbations can possibly be designed to be imperceptible to human beings, which raises concerns about potential safety issues and risks, especially when it comes to applications such as autonomous vehicles where human lives are at stake. The problem of adversarial robustness in machine learning models has been explored from several different perspectives since its discovery. One direction has been to propose attacks that challenge these models and their training procedures [@carlini2017adversarial; @gu2014towards; @athalye2018obfuscated; @papernot2016limitations; @moosavi2016deepfool]. In response, there have been works that propose more robust training techniques that can defend against these adversarial attacks [@he2017adversarial; @zhang2018efficient; @weng2018towards; @wong2018provable; @raghunathan2018certified; @raghunathan2018semidefinite; @shaham2018understanding; @cohen2019certified; @lecuyer2019certified]. For robust training, one promising approach is to treat the problem as a minimax optimization problem, where we try to select model parameters that minimize the loss function under the strongest feasible perturbations [@madry2017towards; @xu2012robustness]. Overall, adversarial training may be computationally expensive [@bubeck2019adversarial; @nakkiran2019adversarial], but it can lead to enhanced resistance towards adversarially modified inputs. Although adversarially robust models tend to outperform standard models when it comes to perturbed test sets, recent studies have found that such robust models are also likely to perform worse on standard (unperturbed) test sets [@tsipras2018robustness; @raghunathan2019adversarial]. We refer to the difference in accuracy on unperturbed test sets as the generalization gap. This paper focuses on the question of whether or not this generalization gap can be closed. Theoretical work by @schmidt2018adversarially has shown that adversarial models require far more data than their standard counterparts to reach a certain level of test accuracy. This supports the general understanding that adversarial training is harder than standard training, as well as the conventional wisdom that more data helps with generalization. However, when it comes to the generalization gap, things may not be so simple. In this paper, we identify two regimes during the adversarial training process. In one regime, more training data eventually helps to close the generalization gap, as expected. In the other regime, the generalization gap will surprisingly continue to grow as more data is used in training. The data distribution and the strength of the adversary determine the regime and the existence of the two regimes indicates a fundamental phase transition in adversarial training. Our Contributions ----------------- In our analysis of the generalization gap, we assume the robust model is trained under $\ell_\infty$ constrained perturbations. We study two classification models including a Gaussian model and a Bernoulli model, as well as a simple linear regression model. For the Gaussian model, we theoretically prove that during the training of a robust classifier there are two possible regimes that summarize the relation between the generalization gap and the training sample size (see \[thm:gaussian\]). More specifically, let $n$ denote the number of training data points. Suppose the perturbation that the adversary can add is constrained to the $\ell_\infty$ ball of radius ${\varepsilon}$. In the strong adversary regime (i.e. large ${\varepsilon}$ compared to the signal strength of the data), the gap always increases and has an infinite data limit. In contrast, in the weak adversary regime, there exists a critical point that marks the boundary between two stages. For all $n$ less than this threshold, we have the increasing stage where the gap monotonically increases. Beyond this threshold, we will eventually reach another stage where the gap strictly decreases. It is important to note that, even in the weak adversary regime, it is possible to make this threshold arbitrarily large, which means adding data points will always expand the generalization gap. For the Bernoulli model, we show similar results (see \[thm:bernoulli\]). Although the curve for the generalization gap will be oscillating (see \[fig:bernoulli\]), we prove that it manifests in a general increasing or decreasing trend. We further explore a simple one-dimensional linear regression and experimentally verify that the phase transition also exists. The primary implication of our work is that simply adding more data will not always be enough to close the generalization gap. Therefore, fundamentally new ideas may be required if we want to be able to train adversarially robust models that do not sacrifice accuracy on unperturbed test sets. Related Work ============ There is an existing body of work studying adversarially robust models and their generalization. In this section we briefly discuss some of the papers that are most relevant to our work. #### Trade-off between robustness and standard accuracy What initially motivated our work is the experimental finding that standard accuracy and adversarial robustness can sometimes be incompatible with each other [@papernot2016towards; @tsipras2018robustness]. These works empirically show that using more data for adversarial training might decrease the standard accuracy. Additionally, this decline becomes more obvious when the radius of perturbation ${\varepsilon}$ increases. This causes a generalization gap between robust and standard models. The side effect of a large perturbation has also been studied by @dohmatob2019generalized who shows that it is possible to adversarially fool a classifier with high standard accuracy if ${\varepsilon}$ is large. @ilyas2019adversarial explore the relation between the perturbation ${\varepsilon}$ and the features learned by the robust model. Their results suggest that a larger ${\varepsilon}$ tends to add more weight onto non-robust features and consequently the model may miss useful features which should be learned under standard setting. @diochnos2018adversarial consider both error region setting and study the classification problem where data is uniformly distributed over $\left\{ 0,1 \right\}^d$. They show that under this $\ell_0$ perturbation setting the adversary can fool the classifier into having arbitrarily low accuracy with at most ${\varepsilon}= O(\sqrt{d})$ perturbation. @zhang2019theoretically theoretically study the trade-off between robustness and standard accuracy from a perspective of decomposition. More specifically, they decompose the robust error into a standard error and a boundary error that would be affected by the perturbation. Their decomposition further leads to a new design of defense. #### Sample complexity for generalization The generalization of adversarially robust models has different properties from the standard ones, especially in sample complexity. @schmidt2018adversarially study Gaussian mixture models in $d$-dimensional space and show that for the standard model only a constant number of training data points is needed, while for the robust model under $\ell_\infty$ perturbation a training dataset of size $\Omega(d)$ is required. Their work is in a different direction to ours: their main result focuses on dimension-dependent bounds for sample complexity, while we quantify the effect of the amount of training data on adversarial generalization and we prove the existence of a phase transition under two binary classification models. @bubeck2019adversarial analyze the computational hardness in training a robust classifier in the statistical query model. They prove that for a binary classification problem in $d$ dimensions, one needs polynomially (in $d$) many queries to train a standard classifier while exponentially many queries to train a robust one. @garg2019adversarially consider a setting where the adversary has limited computational power and show that there exist learning tasks that can only be robustly solved when faced with such limited adversaries. @yin2019rademacher and @khim2018adversarial prove generalization bounds for linear classifiers and neural networks via Rademacher complexity. In addition, @yin2019rademacher show the adversarial Rademacher complexity is always no less than the standard one and is dimension-dependent. @montasser2019vc show the widely used uniform convergence of empirical risk minimization framework, or more generally, any proper learning rule, might not be enough for robust generalization. They prove the existence of a hypothesis class where any proper learning rule gives poor robust generalization accuracy under the PAC-learning setting, while improper learning can robustly learn any class. @cullina2018pac study generalization under the PAC-learning setting and prove a polynomial upper bound for sample complexity that depends on a certain adversarial VC-dimension. @diochnos2019lower study PAC-learning under the error region setting and prove a lower bound for sample complexity that is exponential in the input dimension. #### Other relevant work @bhagoji2019lower use optimal transport to derive lower bounds for the adversarial classification error. For a binary classification problem, they prove a relation between the best possible adversarial robustness and the optimal transport between the two distributions under a certain cost. Another line of work analyzes adversarial examples via concentration of measure and show that their existence is inevitable under certain conditions [@gilmer2018adversarial; @fawzi2018adversarial; @shafahi2018adversarial; @mahloujifar2019curse]. Preliminaries ============= Notation {#sub:notation} -------- We use the shorthand $ [d] $ to denote the set $ \{1,2,\dots,d\} $ for any positive integer $ d $. We use $ \cN(\mu,\Sigma) $ to denote the multivariate Gaussian distribution with mean vector $ \mu $ and covariance matrix $ \Sigma $. If $ u, v\in \bR^d $ are two $ d $-dimensional vectors, the $ j $-th component of $ u $ is denoted by $ u(j) $. The inner product of $ u $ and $ v $ is denoted by $ \langle u, v\rangle $. If $ A $ is a positive semi-definite matrix, let the semi-norm induced by $ A $ be $ \|u\|_A = \sqrt{u^\top A u} $. Let $ B_u^\infty({\varepsilon}) $ denote the $ \ell_\infty $ ball centered at $ u $ and with radius $ {\varepsilon}$, [[*i.e.*]{}]{}, $ B_u^\infty({\varepsilon}) = \{ v\in \bR^d: \| u-v \|_\infty \le {\varepsilon}\} $. In our problem setup in \[sub:problem\_setup\], the ball $ B_u^\infty $ is the set of allowed perturbed vectors for the adversary, where $ {\varepsilon}$ is the perturbation budget. We define the Heaviside step function $ H $ to be $$H(x) = \begin{cases} 1, & \text{for }x > 0\,;\\ \sfrac{1}{2}, & \text{for } x = 0\,;\\ 0, & \text{for } x < 0\,. \end{cases}$$ Problem Setup {#sub:problem_setup} ------------- Suppose that the data $(x,y)$ is drawn from an unknown distribution $\cD$, where $x$ is the input and $y$ is the label. For example, in a classification problem, we have $(x,y)\in \bR^d\times \{\pm 1\}$; in a regression problem, we have $ (x,y)\in \bR^d\times \bR $. Given a model parameter $ w \in \Theta\subseteq \bR^{p} $ and a data point $(x,y)$, the loss of the model parameterized by $ w $ on the data point $ (x,y) $ is denoted by $ \ell(x,y;w) $. The training dataset $ {D_\textnormal{train}}= \{(x_i,y_i)\}_{i=1}^n$ consists of $n$ data points sampled i.i.d. from the distribution $\cD$. Given the training dataset with size $n$, we respectively define the optimal standard and robust models trained on $ {D_\textnormal{train}}$ by $$\label{eq:wstd_wrob_general} \begin{split} {w^\textnormal{std}}_n ={}& \operatorname*{arg\,min}_{w\in \Theta} \frac{1}{n}\sum_{i=1}^n \ell(x_i,y_i;w)\,,\\ {w^\textnormal{rob}}_n = {}& \operatorname*{arg\,min}_{w\in \Theta} \frac{1}{n} \sum_{i=1}^n \max_{\tilde{x}_i\in B^\infty_{x_i}({\varepsilon})} \ell(\tilde{x}_i,y_i;w)\,,\\ \end{split}$$ The optimal standard model $ {w^\textnormal{std}}$ is the minimizer of the total training loss $ \frac{1}{n}\sum_{i=1}^n \ell(x_i,y_i;w) $. In the definition of the optimal robust model $ {w^\textnormal{rob}}$, we take into consideration the adversarial training for each data point, [[*i.e.*]{}]{}, the inner maximization $ \max_{\tilde{x}_i\in B^\infty_{x_i}({\varepsilon})} \ell(\tilde{x}_i,y_i;w) $. We assume that the adversary is able to perturb each data item $ x_i $ within an $ \ell_\infty $ ball centered at $ x_i $ and with radius $ {\varepsilon}$. The best robust model is the minimizer of the total training loss with adversarial training. Note that both $ {w^\textnormal{std}}$ and $ {w^\textnormal{rob}}$ are functions of the training dataset and thereby also random variables. If we have a model parametrized $ w $ and the test dataset $ {D_\textnormal{test}}= \{(x'_i,y'_i)\}_{i=1}^{n'} $ consists of $ n' $ data points sampled i.i.d.  from $ \cD $, the test loss of $ w $ is given by $${L_{\textnormal{test}}}(w) = \bE\left[ \frac{1}{n'}\sum_{i=1}^{n'} \ell(x'_i,y'_i;w) \right] =\bE_{(x,y)\sim \cD}\left[ \ell(x,y;w) \right]\,.$$ Additionally, we define the generalization gap $ g_n $ between the standard and robust classifiers by $$ \begin{split} g_n ={}& \bE_{\{(x_i,y_i)\}_{i=1}^n\stackrel{\textnormal{i.i.d.}}{\sim} \cD}\left[ {L_{\textnormal{test}}}({w^\textnormal{rob}}) - {L_{\textnormal{test}}}({w^\textnormal{std}}) \right]\\ ={} & \bE_{\{(x_i,y_i)\}_{i=1}^n\stackrel{\textnormal{i.i.d.}}{\sim} \cD} \left[ \bE_{(x,y)\sim \cD}[\ell(x,y;{w^\textnormal{rob}})]\right.\\ & \left.- \bE_{(x,y)\sim \cD}[\ell(x,y;{w^\textnormal{std}})]\right]\,. \end{split}$$ Classification {#sec:classification} ============== In this section, we study a binary classification problem, where we have each data point $ (x,y) \in \bR^d \times \{\pm 1\} $. For any model parameter $ w\in \bR^d $, we consider the loss function $\ell(x,y;w) = -y\langle w,x\rangle$ [@yin2019rademacher; @khim2018adversarial]. The parameter $ w $ is constrained on the $ \ell^\infty $ ball $ \Theta = \{ w\in \bR^d \mid \|w\|_\infty \le W \} $, where $W$ is some positive real number. Under this setup, the best standard and robust classifier are given as follows. $$\label{eq:wstd_wrob} \begin{split} {w^\textnormal{std}}_n ={}& \operatorname*{arg\,min}_{\|w\|_\infty\le W} \frac{1}{n}\sum_{i=1}^n -y_i\langle w,x_i\rangle\\ ={}& \operatorname*{arg\,max}_{\|w\|_\infty\le W} \sum_{i=1}^n y_i\langle w,x_i\rangle\,, \\ {w^\textnormal{rob}}_n = {}& \operatorname*{arg\,min}_{\|w\|_\infty\le W} \frac{1}{n} \sum_{i=1}^n \max_{\tilde{x}_i\in B^\infty_{x_i}({\varepsilon})} \left( -y_i\langle w,\tilde{x_i}\rangle \right)\\ = {}& \operatorname*{arg\,max}_{\|w\|_\infty\le W} \sum_{i=1}^n \min_{\tilde{x}_i\in B^\infty_{x_i}({\varepsilon})} y_i\langle w,\tilde{x_i}\rangle\,. \end{split}$$ The generalization gap $ g_n $ between the standard and robust classifiers is given by $$\label{eq:generalization_gap} \begin{split} g_n ={} & \bE_{\{(x_i,y_i)\}_{i=1}^n\stackrel{\textnormal{i.i.d.}}{\sim} \cD} \left[ \bE_{(x,y)\sim \cD}[y\langle {w^\textnormal{std}},x\rangle]\right.\\ & \left.- \bE_{(x,y)\sim \cD}[y\langle {w^\textnormal{rob}},x\rangle]\right]\,. \end{split}$$ In this paper, we investigate how the generalization gap $ g_n $ evolves with the amount of data. Intuitively, one might conjecture that the generalization gap should satisfy the following properties: 1. First, the gap should always be non-negative. This means that the robust classifier incurs a larger test (generalization) loss than the standard classifier, as there is no free lunch and robustness in adversarial training would compromise generalization performance. \[it:gn-positive\] 2. Second, more training data would close the gap gradually; in other words, the gap would be decreasing with respect to the size of the training dataset. \[it:gn-decrease\] 3. Third, in the infinite data limit ([[*i.e.*]{}]{}, when the size of the training dataset tends to infinity), the generalization gap would eventually tend to zero. \[it:gn-zero\] Our study corroborates \[it:gn-positive\] but denies \[it:gn-decrease\] and \[it:gn-zero\] in general. The implication of this is not only that current adversarial training techniques sacrifice standard accuracy in exchange for robustness, but that simply adding more data may not solve the problem. [0.5]{} {width="\linewidth"} [0.5]{} {width="\linewidth"} Gaussian Model -------------- The Gaussian model is specified as follows. Let $ (x,y)\in \bR^d \times \{\pm 1\} $ obey the distribution such that $ y\sim \operatorname{Unif}(\{\pm 1\}) $ and $ x\mid y\sim \cN(y\mu, \Sigma) $, where $ \mu(j)\ge 0 $ for $ \forall j\in [d] $ and $ \Sigma=\operatorname{diag}(\sigma(1)^2,\sigma(2)^2,\dots,\sigma(d)^2) $. We denote this distribution by $ (x,y)\sim {\cD_{\textnormal{Gau}}}$. \[thm:gaussian\] Given i.i.d. training data $ (x_i,y_i)\sim {\cD_{\textnormal{Gau}}}$ with $ n $ data points, if we define the standard and robust classifier as in (denoted by ${w^\textnormal{std}}$ and ${w^\textnormal{rob}}$, respectively) and define the generalization gap $g_n$ as in , we have 1. $ g_n \ge 0 \hspace{0.05in} \forall \ n\ge 1 $; \[it:gn\_ge0\] 2. The infinite data limit equals $$\lim_{n\to\infty} g_n = 2W \sum_{j\in [d]:\mu(j)>0} \mu(j) H\left( \frac{{\varepsilon}}{\mu(j)}-1 \right)\,;$$ \[it:gn\_lim\] 3. If $ {\varepsilon}< \min_{j\in [d]:\mu(j)>0} \mu(j) $, $ g_n $ is strictly increasing in $ n $ when $$n < \min_{j\in [d]:\mu(j)>0} \max\left\{ \frac{3}{2} ,2\log\frac{1}{1-{\varepsilon}/\mu(j)} \right\} \left( \frac{\sigma(j)}{\mu(j)} \right)^2 \,,$$ and it is strictly decreasing in $ n $ when $$n \ge \max_{j\in [d]: \mu(j)>0}\left( K_0 + 2\log \frac{1}{1-{\varepsilon}/\mu(j)} \right)\left( \frac{\sigma(j)}{\mu(j)} \right)^2\,,$$ where $ K_0 $ is a universal constant. \[it:gn\_increase\] 4. If $ {\varepsilon}> \| \mu \|_\infty $, $ g_n $ is strictly increasing for all $ n\ge 1 $.\[it:gn\_always\_increase\] Part \[it:gn\_ge0\] of \[thm:gaussian\] states that the generalization of the robust classifier is never better than the standard one. Part \[it:gn\_lim\] quantifies the size of the gap as the size of the training dataset $n$ goes to infinity. The main implication here is that the gap will always converge to some finite limit, which may be zero if the strength of the adversary $\epsilon$ is small enough. Parts \[it:gn\_increase\] and \[it:gn\_always\_increase\] describe the two different possible regimes. Part \[it:gn\_increase\] states that if the strength of the adversary is not too large, then there will be two stages: an initial stage where the generalization gap is strictly increasing in $n$, followed by a secondary stage where the generalization gap is strictly decreasing in $n$. On the other hand, part \[it:gn\_always\_increase\] states that a large $\epsilon$ will result in a generalization gap that is strictly increasing (but still tending towards some finite limit). In order to better describe and visualize the implications of \[thm:gaussian\], we consider a special case where $ \mu = (\mu_0,\dots,\mu_0) $ and $ \Sigma = \sigma_0^2 I $. \[cor:gaussian\] Assume that $ W=1 $, $ \mu(j) = \mu_0 \ge 0 $, and $ \sigma(j) = \sigma_0 > 0 $ for all $ j\in [d] $. The infinite data limit equals $$\lim_{n\to\infty} g_n = 2d\mu_0 H\left( \frac{{\varepsilon}}{\mu_0} - 1 \right) = \begin{cases} 2d\mu_0, & \text{for } \frac{{\varepsilon}}{\mu_0}>1\,;\\ d \mu_0, & \text{for } \frac{{\varepsilon}}{\mu_0}=1\,;\\ 0, & \text{for } \frac{{\varepsilon}}{\mu_0}<1\,. \end{cases}$$ If $ {\varepsilon}< \mu_0 $, we have $ g_n $ is strictly increasing when $$n < \max\left \{ \frac{3}{2}, 2\log \frac{1}{1-{\varepsilon}/\mu_0} \right \}\left( \frac{\sigma_0}{\mu_0} \right)^2\,,$$ and it is strictly decreasing when $$n \ge \left( K_0 + 2\log \frac{1}{1-{\varepsilon}/\mu_0} \right) \left( \frac{\sigma_0}{\mu_0} \right)^2\,,$$ where $ K_0 $ is a universal constant. If $ {\varepsilon}> \mu_0 $, we have $ g_n $ is strictly increasing for all $ n\ge 1 $. \[cor:gaussian\] is essentially a simplified version of parts \[it:gn\_increase\] and \[it:gn\_always\_increase\] of \[thm:gaussian\] where we cleanly divide between a weak adversary regime and a strong adversary regime at a threshold ${\varepsilon}= \mu_0$. We illustrate the generalization gap $ g_n $ vs. the size of the training dataset in \[fig:gaussian\], where we set $ W = d = \mu = 1 $ and $ \sigma = 2 $. The curve $ {\varepsilon}= 1 $ belongs to the strong adversary regime, while the remaining curves belong to the weak adversary regime. In the weak adversary regime, the evolution of $ g_n $ can be divided into two stages, namely the increasing and decreasing stages (part \[it:gn\_increase\] of \[thm:gaussian\]). The duration of the increasing stage is $$\Theta \left( \left( \frac{\sigma_0}{\mu_0} \right)^2 \log\frac{1}{1-{\varepsilon}/\mu_0} \right)\,.$$ This duration is controlled by the ratio $ {\varepsilon}/\mu_0 $, as well as the reciprocal of the signal-to-noise ratio (SNR), [[*i.e.*]{}]{}, $ \frac{\sigma_0}{\mu_0} $. A larger SNR and an $ {\varepsilon}$ closer to $ \mu_0 $ lead to a shorter increasing stage. It can be observed in \[fig:gaussian\] that for the curves with $ {\varepsilon}= 0.2,0.5,0.7,0.9,0.95 $, a larger $ {\varepsilon}$ results in a longer duration of the increasing stage. After the increasing stage, the generalization gap will eventually begin to decrease towards some finite limit (given by part \[it:gn\_lim\] of \[thm:gaussian\]) if sufficient training data is provided. In addition, we would like to remark that the duration relies on the data and the strength of the adversary and could be potentially arbitrarily large; in other words, without full information about the true data distribution and the power of the adversary, one cannot predict when the increasing stage will terminate. In the strong adversary regime, the generalization gap expands from the very beginning. In the infinite data limit, the gap approaches $ d \mu_0 $ if $ {\varepsilon}= \mu_0 $, and it approaches $ 2d \mu_0 $ if $ {\varepsilon}> \mu_0 $. Bernoulli Model --------------- In this subsection, we investigate the Bernoulli model defined as follows. Let $ (x,y)\in \bR^d \times \{\pm 1\} $ obey the distribution such that $ y\sim \operatorname{Unif}(\{\pm 1\}) $ and for $ \forall j \in [d] $ independently, $$x(j) = \begin{cases} y\cdot \theta(j) &\text{with probability } \frac{1+\tau}{2}\,,\\ -y\cdot \theta(j) & \text{with probability } \frac{1-\tau}{2}\,, \end{cases}$$ where $ \theta\in \bR^d_{\ge 0} $ and $ \tau\in (0,1) $. We denote this distribution by $ (x,y)\sim {\cD_{\textnormal{Ber}}}$. The parameter $ \tau $ controls the signal strength level. When $ \tau =0 $ (lowest signal strength), $ x(j) $ takes the value of $ +1 $ or $ -1 $ uniformly at random, irrespective of the label $ y $. When $ \tau=1 $ (highest signal strength), we have $ x(j) = y\cdot \theta(j) $ almost surely. We illustrate the generalization gap $ g_n $ vs. the size of the training dataset (denoted by $ n $) in \[fig:bernoulli\], where we set $ W = d = \theta = 1 $ and $ {\varepsilon}= 0.2 $. We observe that all curves $ g_n $ oscillate around the other curves labeled $ s_n $. Although the figure shows that the curves $ g_n $ are not monotone, they all exhibit a monotone trend, which is characterized by $ s_n $. As a result, we will not show that $ g_n $ is monotonically increasing or decreasing (as shown in \[fig:bernoulli\], it is not monotone). Alternatively, we will show that $ g_n $ resides in a strip centered around $ s_n $ and $ s_n $ displays (piecewise) monotonicity. Additionally, the height of the strip shrinks at a rate of $ O\left( \frac{1}{\sqrt{n}} \right) $; in other words, it can be shown that $$\left |g_n-s_n \right | \le O\left( \frac{1}{\sqrt{n}} \right), \quad \forall n \ge 1\,.$$ \[thm:bernoulli\] Given i.i.d. training data $ (x_i,y_i)\sim {\cD_{\textnormal{Ber}}}$ with $ n $ data points, if we define the standard and robust classifier (denoted by $ {w^\textnormal{std}}$ and $ {w^\textnormal{rob}}$, respectively) as in and define the generalization gap $ g_n $ as in , we have 1. $ g_n\ge 0 $ for $ \forall n\ge 1 $; 2. The infinite data limit equals $$\lim_{n\to\infty} g_n = 2W\tau \sum_{j\in [d]:\theta(j)>0} \theta(j) H\left(\frac{{\varepsilon}}{\theta(j)\tau}-1\right)\,,$$ where $ H $ is the Heaviside step function. Furthermore, there exists a positive constant $ C_0 \le \frac{\sqrt{10}+3}{6\sqrt{2\pi}}\approx 0.4097 $ and a sequence $ s_n $ such that $ |g_n-s_n|\le \frac{8C_0W\tau \| \theta \|_1(\tau ^2+1)}{\sqrt{n}\sqrt{1-\tau ^2}} $ and 1. If $ \frac{{\varepsilon}}{\tau} < \min_{j\in [d]:\theta(j)>0} \theta(j) $, $ s_n $ is strictly increasing in $ n $ when $$n < \left ( \frac{1}{\tau^2}-1 \right ) \max \left \{ \frac{3}{2}, 2\min_{j\in [d]:\theta(j)>0} \log \frac{1}{1-\frac{{\varepsilon}}{\theta(j)\tau}} \right \}$$ and strictly decreasing in $ n $ when $$n \ge \left( \frac{1}{\tau^2}-1 \right) \left( K_0 + 2 \max_{j\in [d]: \theta(j)>0} \log \frac{1}{1-\frac{{\varepsilon}}{\theta(j)\tau}} \right) \,,$$ where $ K_0 $ is a universal constant; 2. If $ \frac{{\varepsilon}}{\tau}\ge \| \theta \|_{\infty} $, $ s_n $ is strictly increasing for all $ n \ge 1 $. Again, to explain the implications of \[thm:bernoulli\], we explore the following special case where $ W=1 $ and $ \theta = (\theta_0,\dots,\theta_0) $. Assume $ W=1 $ and that $ \theta(j) = \theta_0 > 0 $ holds for all $ j\in [d] $. The infinite data limit equals $$\label{eq:ber-gn} \begin{split} \lim_{n\to\infty} g_n ={}& 2\tau d \theta_0 H\left( \frac{{\varepsilon}}{\theta_0\tau} -1 \right)\\ ={}& \begin{cases} 2\tau d \theta_0, & \text{for } {\varepsilon}> \theta_0 \tau\,;\\ \tau d \theta_0, & \text{for } {\varepsilon}= \theta_0 \tau\,;\\ 0, & \text{for } {\varepsilon}< \theta_0 \tau\,. \end{cases} \end{split}$$ If $ {\varepsilon}< \theta_0\tau $, $ s_n $ is strictly increasing in $ n $ when $$n < \left ( \frac{1}{\tau^2}-1 \right ) \max \left \{ \frac{3}{2}, 2 \log \frac{1}{1-{\varepsilon}/(\theta_0 \tau)} \right \} \,,$$ and it is strictly decreasing when $$n \ge \left( \frac{1}{\tau^2}-1 \right) \left( K_0 + 2 \log \frac{1}{1-{\varepsilon}/(\theta_0 \tau)} \right) \,,$$ where $ K_0 $ is a universal constant. If $ {\varepsilon}\ge \theta_0\tau $, $ s_n $ is strictly increasing for all $ n\ge 1 $. Similar to the Gaussian model, there also exist two regimes. One is the weak adversary regime where $ {\varepsilon}< \theta_0 \tau $, while the other is the strong adversary regime where $ {\varepsilon}\ge \theta_0\tau $. Recall that in \[fig:bernoulli\], we set $ W=d=\theta=1 $ and $ {\varepsilon}= 0.2 $. Therefore the values $ \tau = 0.1 $ and $ \tau = 0.2 $ lie in the strong adversary regime, while the values $ \tau = 0.5 $ and $ \tau = 0.7 $ belong to the weak adversary regime. In the weak adversary regime, the critical point is when $$\label{eq:ber-critical} n \approx \Theta\left( \left( \frac{1}{\tau^2}-1 \right) \log \frac{1}{1-{\varepsilon}/(\theta_0\tau)} \right)\,.$$ Before this critical point, the strip center $ s_n $ that the generalization gap $ g_n $ oscillates around is strictly increasing; it is strictly decreasing after the critical point and eventually vanished as $n\to \infty$. Note that when $ \tau\to 0 $, both terms ($ \left( \frac{1}{\tau^2}-1 \right) $ and $ \log \frac{1}{1-{\varepsilon}/(\theta_0\tau)} $) in blow up and thereby the increasing stage elongates infinitely. The increasing and decreasing stages of the weak adversary regime are confirmed by the two curves $ \tau = 0.5 $ and $ \tau = 0.7 $ in \[fig:bernoulli\]. In the strong adversary regime, the strip center $ s_n $ displays a similar trend as the generalization gap in the Gaussian model; i.e., it is strictly increasing from the very beginning (see the two curves $ \tau=0.1 $ and $ \tau=0.2 $ in \[fig:bernoulli\]). Recall that under the Bernoulli model, the strong/weak adversary regime is determined by the ratio $ \frac{{\varepsilon}}{\theta_0\tau} $, while under the Gaussian model, it is determined by the ratio $ \frac{{\varepsilon}}{\mu_0} $. Nevertheless, note that in the binary classification, $\theta_0 \tau$ is the mean (in one coordinate) of the positive class, just like $\mu_0$ in the Gaussian scenario. These two ratios are thus closely related. We would also like to remark that limits of $g_n$ in \[fig:bernoulli\] follow the theoretical results outlined in . In particular, if we are in the weak adversary regime, the limit of $g_n$ always tends to 0. On the other hand, in the strong adversary regime, the limit is non-zero and proportional to $\tau$. Discussion ---------- One common observation from \[thm:gaussian\] and \[thm:bernoulli\] is that the duration of the increasing stage heavily depends on the ratio between ${\varepsilon}$ and the coordinate-wise mean of the positive class (i.e. $\mu_0$ and $\theta_0 \tau$). Note that the mean can be interpreted as half the distance between the centers of positive and negative classes in the space of $x$. Thus, another way to view this result is that if the strength of the adversary is relatively large compared to the distance between classes, then we will have a long increasing stage. One interesting implication of this can be seen in regression vs. classification tasks. Intuitively, one might look at a regression task as a classification task with infinitely many classes. Therefore, depending on the distribution that $x$ is sampled from, we could end up with a very small distance between class centers and thus we would expect a very long increasing stage. Regression {#sec:regression} ========== [0.24]{} {width="\linewidth"} [0.24]{} {width="\linewidth"} [0.24]{} {width="\linewidth"} [0.24]{} {width="\linewidth"} In this section, we explore the problem of linear regression, where we have each data point $ (x,y)\in \bR^d\times \bR $ and the linear model is represented by a vector $ w\in \bR^d $. The loss function is defined by $ \ell(x,y;w) = (y-\langle w,x\rangle )^2 $. We assume the following data generation process. First, we sample $ x_i $ from some distribution $ P_X $. Given the fixed true model $ w^* $, we set $ y_i = \langle w^*,x_i\rangle + \delta $, where $ \delta\sim \cN(0,\sigma^2) $ is the Gaussian noise. The parameter space $ \Theta $ is the entire $ \bR^d $. Given the training dataset $ {D_\textnormal{train}}= \{(x_i,y_i)\}_{i=1}^n$, if we define $ X = [x_1,\dots,x_n]^\top $ and $ y = [y_1,\dots,y_n]^\top $, the best standard model has a closed form [@graybill1961introduction]. $${w^\textnormal{std}}= (X^\top X)^{-1} X^\top y\,.$$ \[ob:regression-wstd-wrob\] presents the form of the best robust model in the linear regression problem. \[ob:regression-wstd-wrob\] The best robust model in the linear regression problem is given by $${w^\textnormal{rob}}_n = \operatorname*{arg\,min}_{w\in \bR^d}\frac{1}{n} \sum_{i=1}^{n} \sum_{j=1}^{d}\left( |y_i - w(j)x_i(j)| + {\varepsilon}|w(j)| \right)^2\,.$$ The optimization problem in \[ob:regression-wstd-wrob\] is completely decoupled in each dimension. Optimizing each dimension separately, we have for all $ j \in [d] $, $${w^\textnormal{rob}}_n(j) = \operatorname*{arg\,min}_{w\in \bR} \sum_{i=1}^{n} \left( |y_i - w x_i(j)| + {\varepsilon}|w| \right)^2\,.$$ \[ob:regression-gn\] In the linear regression problem, The generalization gap equals $$g_n = \| {w^\textnormal{rob}}_n - w^* \|^2_{\bE_{x\sim P_X}[xx^\top]} - \| {w^\textnormal{std}}_n - w^* \|^2_{\bE_{x\sim P_X}[xx^\top]}\,.$$ \[ob:regression-gn\] shows that the generalization gap not only depends on the difference vectors $ ({w^\textnormal{rob}}_n-w^*) $ and $ ({w^\textnormal{std}}_n-w^*) $ but also the matrix $ \bE_{x\sim P_X}[xx^\top] $. This matrix weights each dimension of the difference vectors and thereby influences the generalization gap. To avoid the complication incurred by the different weightings of the matrix $ \bE_{x\sim P_X}[xx^\top] $ across the dimensions, we investigate two one-dimensional linear regression problems ($ d=1 $) with the data input $ x $ sampled from a standard normal distribution and a shifted Poisson distribution, respectively. To be specific, in the first study, we consider $ x $ sampled from the standard normal distribution $ \cN(0,1) $. In the second study, the data input $ x $ is drawn from $ \operatorname{Poisson}(5) + 1 $ (in order to avoid $ x = 0 $); in other words, $x-1$ obeys the $\operatorname{Poisson}(5)$ distribution. In both studies, we set the true model $ w^* = 1 $ and the noise obeys $ \delta\sim \cN(0,1) $ ([[*i.e.*]{}]{}, $ \sigma^2=1 $). In light of \[ob:regression-gn\], we obtain that if the linear regression problem is one-dimensional, the generalization gap equals $$g_n = \bE_{(x,y)\sim \mathcal{D}}\left( ({w^\textnormal{rob}}_n - w^* )^2 - ({w^\textnormal{std}}_n - w^*)^2 \right) \bE_{x\sim P_X}[x^2]\,.$$ Since $ g_n $ is proportional to $ \left( ({w^\textnormal{rob}}_n - w^* )^2 - ({w^\textnormal{std}}_n - w^*)^2 \right) $ with $ \bE_{x\sim P_X}[x^2] $ being a constant, we call $ g_n/\bE_{x\sim P_X}[x^2] $ the *scaled generalization gap* and plot it against the size of the training dataset (denoted by $ n $) in \[fig:regression\]. \[fig:reg-gaussian1\] shows the result for the first study with $ n $ ranging from $ 1 $ to $ 20 $. For a clear presentation, \[fig:reg-guassian5\] provides a magnified plot for $ 5\le n\le 20 $. Our first observation is that in the Gaussian case, the generalization gap $ g_n $ always expands with more data, even if $ {\varepsilon}$ is as small as $ 0.05 $. This may be because if we sort $ n $ i.i.d. standard normal random variables $ x_1,\dots,x_n $ in ascending order and obtain $ x_{\pi(1)} \le x_{\pi(2)} \le \cdots \le x_{\pi(n)} $, the difference between two consecutive numbers ([[*i.e.*]{}]{}, $ x_{\pi(i+1)} - x_{\pi(i)} $) becomes smaller as $ n $ becomes larger. As we discussed in \[sec:classification\], the monotone trend of $ g_n $ is determined by the ratio of $ {\varepsilon}$ to half the distance between the positive and negative classes. The ratio is $ \frac{{\varepsilon}}{\mu_0} $ in the Gaussian model and it is $ \frac{{\varepsilon}}{\theta_0\tau} $ in the Bernoulli model. The regression problem may be viewed as a classification problem with infinitely many classes. The difference between two consecutive numbers is the analog of the distance between the means of difference classes. Since the difference reduces as $ n $ becomes larger (points are more densely situated), the ratio increases and therefore we observe a wider generalization gap. Our second observation regarding the Gaussian data is that the generalization gap is (very) negative at the initial stage. In particular, when $ n=1 $, the gap $g_1$ is between $ -10^6 $ and $ -10^7 $. The reason is that when $ n=1 $, we have $$ {\mathbb{E}}\left[ ({w^\textnormal{std}}_1-w^*)^2 \right] = \infty\,. $$ Because of the robustness, $ {w^\textnormal{rob}}$ is more stablized and therefore $ \bE[\left( {w^\textnormal{rob}}_1-w^* \right)^2] $ is finite. Since the generalization gap $ g_1 $ is proportional to their difference $ \bE[\left( {w^\textnormal{rob}}_1-w^* \right)^2] - {\mathbb{E}}\left[ ({w^\textnormal{std}}_1-w^*)^2 \right] $, the gap $g_1$ is indeed $ -\infty $. We present a proof of $g_1 = -\infty$ in \[thm:g1-minus-infty\]. \[thm:g1-minus-infty\] In the one-dimensional linear regression problem, if $ x_1\sim \cN(0,1) $, $ \delta\sim \cN(0,1) $, and $ y_1 = w^* x+\delta $, the generalization gap $ g_1 $ with only one training data point is $ -\infty $. \[fig:poisson1\] presents the result for the Poisson input with $ {\varepsilon}$ varying from $ 1.0 $ to $ 6.0 $. \[fig:poisson5\] illustrates the result corresponding to large $ {\varepsilon}$ that ranges from $ 7.0 $ to $ 15.0 $. We see two different regimes in \[fig:poisson1\] and \[fig:poisson5\]. \[fig:poisson1\] represents the weak adversary regime where the generalization gap shrinks with more training data. \[fig:poisson5\] represents the strong adversary regime in which the gap expands with more training data. Furthermore, given the same size of the training dataset, the gap increases with $ {\varepsilon}$. The result for the Poisson input is in sharp contrast to the Gaussian input. It appears that for any small $ {\varepsilon}$, the generalization gap will increase with more data in the Gaussian setting, as the real line becomes increasingly crowed with data points. In the Poisson setting, whilst the Poisson distribution is infinitely supported as well, the minimum distance between two different data points is one (recall that the Poisson distribution is supported on natural numbers). A weak adversary with a small $ {\varepsilon}$ is unable to drive the generalization gap into an increasing trend. Additionally, recalling that the mean of $ \operatorname{Poisson}(5) + 1 $ is $ 6 $, the value $ {\varepsilon}= 6 $ exactly separates the weak and strong adversary regimes in these two figures. Note that all $ {\varepsilon}$ values in \[fig:poisson1\] are $ \le 6 $, while all those in \[fig:poisson5\] are $ >6 $. Unlike the Gaussian setting for linear regression, we never observe a negative generalization gap, even if $n = 1$. This observation supports our theoretical finding, which is summarized in \[thm:g1-minus-infty-poisson\]. \[thm:g1-minus-infty-poisson\] In the one-dimensional linear regression problem, if $ x_1\sim \operatorname{Poisson}(\lambda)+1 $, $ \delta\sim \cN(0,1) $, and $ y_1 = w^* x+\delta $ with $|w^*|\geq 1$, the generalization gap $ g_1 $ with only one training data point is non-negative, finite, and increases with ${\varepsilon}$. Conclusion ========== In this paper, we study the generalization gap between adversarially robust models and standard models. We analyze two classification models (the Gaussian model and the Bernoulli model), and we also explore the linear regression model. We theoretically find that a larger training dataset won’t necessarily close the generalization gap and may even expand it. In addition, for the two classification models, we prove that the generalization gap is always non-negative, which indicates that current adversarial training must sacrifice standard accuracy in exchange for robustness. For the Gaussian classification model, we identify two regimes: the strong adversary regime and the weak adversary regime. In the strong adversary regime, the generalization gap monotonically expands towards some non-negative finite limit as more training data is used. On the other hand, in the weak adversary regime, there are two stages: an increasing stage where the gap increases with the training sample size, followed by a decreasing stage where the gap decreases towards some finite non-negative limit. Broadly speaking, the ratio between the strength of the adversary and the distance between classes determines which regime we will fall under. In the Bernoulli model, we also prove the existence of the weak and strong adversary regimes. The primary difference is that generalization gap is oscillating instead of monotone. However, we also show that these oscillating curves have strip centers that display very similar behavior to the Gaussian curves. Our findings are further validated by a study of the linear regression model, which experimentally exhibits similar behavior and may indicate that our results hold for an even broader class of models. The ultimate goal of adversarial training is to learn models that are robust against adversarial attacks, but do not sacrifice any accuracy on unperturbed test sets. The primary implication of our work is that this trade-off is provably unavoidable for existing adversarial training frameworks. Acknowledgements {#acknowledgements .unnumbered} ================ We would like to thank Mohammad Mahmoody for helpful conversations and comments and thank Marko Mitrovic for his help in preparation of the paper. Proof of \[thm:gaussian\] {#sec:proof-gaussian} ========================= \[lem:phi\] Define the function $ \phi(x) = 2\Phi(x) - \Phi(x(1+\delta)) - \Phi(x(1-\delta)) $, where $ \Phi $ denotes the CDF of the standard normal distribution. We have 1. If $ \delta >0 $, $ \lim_{x\to\infty} \phi(x) = H(\delta-1) $, where $ H $ is the Heaviside step function. 2. If $ \delta \in (0,1) $, there exists $$\sqrt{ \max\{ \sfrac{3}{2},2\log\frac{1}{1-\delta} \}} < x_0 < \sqrt{K_0+2\log\frac{1}{1-\delta}}$$ such that the function $ \phi(x) $ is strictly increasing on $ (0, x_0) $ and strictly decreasing on $ (x_0,\infty) $, where $ K_0 >0 $ is a universal constant. 3. If $ \delta\ge 1 $, the function $ \phi(x) $ is strictly increasing on $ (0,\infty) $. First, we compute the derivative of $ \phi(x) $ and obtain $$\phi'(x) = \frac{1}{\sqrt{2 \pi }} e^{-\frac{1}{2} (\delta +1)^2 x^2} \left(-\delta +(\delta -1) e^{2 \delta x^2}+2 e^{\frac{1}{2} \delta (\delta +2) x^2}-1\right) \,.$$ If we define $ h(a) = -\delta +(\delta -1) a^{2 \delta }+2 a^{\frac{1}{2} \delta (\delta +2) }-1 $ for $ a\ge 1 $, we have $ \phi'(x) = \frac{1}{\sqrt{2 \pi }} e^{-\frac{1}{2} (\delta +1)^2 x^2} h(e^{x^2}) $. It can be observed that $ h(1) = 0 $. We first consider the case where $ \delta\ge 1 $. The derivative of $ h(a) $ with respect to $ \delta $ is given by $$\frac{\partial h(a)}{\partial \delta} = a^{2 \delta }+2 \left((\delta +1) a^{\frac{\delta ^2}{2}}+(\delta -1) a^{\delta }\right) a^{\delta } \log (a)-1\,,$$ which is non-negative when $ a\ge 1 $ and $ \delta \ge 1 $. Therefore, we deduce $ h(a)\ge h(a)\rvert_{\delta=1} = 4 a^{3/2} \log (a)+a^2-1 $. Since the right-hand side is increasing in $ a $, we get $ h(a)\ge h(a)\rvert_{\delta=1,a=1} = 0 $ and the equality is attained when $ a=0 $. In other words, $ h(a)>0 $ if $ a>1 $, which implies $ \phi'(x)>0 $ for $ x>0 $. Therefore, the function $ \phi(x) $ is strictly increasing on $ (0,\infty) $. Next, we compute the limit $ \lim_{x\to \infty} \phi(x) $. If $ \delta >1 $, when $ x $ goes to $ \infty $, we have $ x(1+\delta) $ goes to $ \infty $ as well, while $ x(1-\delta) $ goes to $ -\infty $. Recall that $ \Phi(x) $ is a CDF. Since $ \lim_{x\to\infty} \Phi(x)=1 $ and $ \lim_{x\to -\infty} \Phi(x)=0 $, we obtain that $ \lim_{x\to\infty} \phi(x) = 2-1-0 = 1 $. If $ \delta = 1 $, we have $ \phi(x) = 2\Phi(x) - \Phi(2x) - \Phi(0) =2\Phi(x) - \Phi(2x) -1/2 $. Therefore, we obtain $ \lim_{x\to\infty}\phi(x) = 2-1-1/2=1/2 $. In the sequel, we assume $ \delta\in (0,1) $. In this case, when $ x $ goes to $ \infty $, both $ x(1+\delta) $ and $ x(1-\delta) $ go to $ \infty $. Therefore, we get $ \lim_{x\to\infty}\phi(x) = 2-1-1 = 0 $. The derivative of $ h(a) $ is given by $$h'(a) = \delta a^{\delta -1} \left((\delta +2) a^{\frac{\delta ^2}{2}}+2 (\delta -1) a^{\delta }\right) \,.$$ If $ \delta\in (0,1) $, the function $ h'(a) $ is positive on $ (1, a_0) $ and negative on $ (a_0,\infty) $, where $ a_0=\left( \frac{2+\delta}{2(1-\delta)} \right)^{\frac{1}{\delta-\delta^2/2}} $. Therefore, if $ \delta>0 $, the function $ h(a) $ is strictly increasing on $ (1,a_0) $ and strictly decreasing on $ (a_0,\infty) $. Since $ h(1)=0 $ and $ \lim_{a\to\infty}h(a) = -\infty $, we deduce that $ h(a) $ has a unique root $ a_1 $ on $ (1,\infty) $ and $ a_0<a_1 $. We claim that $ a_0(\delta)= \left( \frac{2+\delta}{2(1-\delta)} \right)^{\frac{1}{\delta-\delta^2/2}} $ is increasing with respect to $ \delta\in (0,1) $. We define $ f(\delta) = \log a_0(\delta) = \frac{\log \left(\frac{\delta +2}{2 (1-\delta )}\right)}{\delta -\frac{\delta ^2}{2}} $ and $ f_1(\delta) = 4 \left(\delta ^3-3 \delta +2\right) \log \left(\frac{\delta +2}{2-2 \delta }\right)+6 (\delta -2) \delta $. The derivative of $ f(\delta) $ is given by $ f'(\delta) = \frac{4 \left(\delta ^3-3 \delta +2\right) \log \left(\frac{\delta +2}{2-2 \delta }\right)+6 (\delta -2) \delta }{(\delta -2)^2 \delta ^2 \left(\delta ^2+\delta -2\right)} = \frac{f_1(\delta)}{(\delta -2)^2 \delta ^2 \left(\delta ^2+\delta -2\right)} $. The derivative of $ f_1(\delta) $ is $ f'_1(\delta) = 12 \left(\delta ^2-1\right) \log \left(\frac{\delta +2}{2-2 \delta }\right) $. Since $ \delta>0 $, we have $ \delta+2 > 2-2\delta $ and thus $ \log \left(\frac{\delta +2}{2-2 \delta }\right) > 0 $. Therefore $ f'_1(\delta) < 0 $ on $ (0,1) $. As a result, $ f_1(\delta) $ is decreasing on $ (0,1) $ and thus for $ \delta\in (0,1) $, we have $ f_1(\delta) < f_1(0) = 0 $. Since $ (\delta -2)^2 \delta ^2 \left(\delta ^2+\delta -2\right) < 0 $ holds for $\delta\in (0,1) $, the derivative $ f'(\delta)>0 $ on $ (0,1) $. Therefore the function $ f(\delta) $ is increasing on $ (0,1) $ and for $ \forall \delta\in (0,1) $, we have $ f(\delta) \ge \lim_{\delta\to 0+} f(\delta) = \frac{3}{2} $. Thus $ a_0(\delta) $ is increasing on $ (0,1) $ and $ a_0(\delta)\ge e^{3/2} $. Since $ a_0<a_1 $, we have $ a_1>e^{3/2} $. Next, we show that $ a_1 > \frac{1}{(1-\delta)^2} $. Since $a_1 > e^{3/2} $, it suffices to show $a_1 > \frac{1}{(1 - \delta)^2 }$ for $\delta > 1-e^{-\sfrac{3}{4}}>\sfrac{1}{2}$. As a result, in what follows, we assume $ \sfrac{1}{2} < \delta < 1 $. First, since $ \sfrac{1}{2} < \delta < 1 $, the following inequality holds $$\label{eq:le_onehalf} (1-\delta)^{\delta(2+\delta)} < (1-\delta)^{\sfrac{5}{4}} < \frac{1}{2}\,.$$ Therefore, we deduce $$(1-\delta)^{(1-\delta)^2} + (1-\delta)^{\delta(2+\delta)} (1+\delta) < 1 + (1-\delta)^{\delta(2+\delta)} (1+\delta) < 1 + 2(1-\delta)^{\delta(2+\delta)} < 1+2\cdot \frac{1}{2} = 2\,,$$ where we use in the last inequality. Since $$(1-\delta)^{\delta(2+\delta)}\left( 1+\delta+(1-\delta)^{1-4\delta} \right) = (1-\delta)^{(1-\delta)^2} + (1-\delta)^{\delta(2+\delta)} (1+\delta) < 2\,,$$ we have $$2(1-\delta)^{-\delta(2+\delta)} > 1+\delta+(1-\delta)^{1-4\delta}\,.$$ We are in a position to evaluate $ h\left( \frac{1}{(1-\delta)^2} \right) $: $$h\left( \frac{1}{(1-\delta)^2} \right) = 2 (1-\delta )^{-\delta (\delta +2)}-(1-\delta )^{1-4 \delta }-(1 + \delta ) > 0\,.$$ Therefore we get $h(a_1) = h(\frac{1}{(1-\delta)^2}) > 0 $. Recall that $ a_1 $ is the unique root of $ h(a) $ and that $ h(a) >0 $ if $ a\in (0,a_1) $ while $ h(a)<0 $ if $ a > a_1 $. Therefore we have $ a_1 > \frac{1}{(1-\delta)^2} $ if $ \delta > \sfrac{1}{2} $. As a consequence, we conclude $a_1 > \max \{ e^{3/2} , \frac{1}{(1-\delta)^2} \}$. In the final part of the proof, we derive an upper bound for $ a_1 $. We consider the function $$h_1(\delta,K) = -(\delta +1) (1-\delta )^{\delta (\delta +2)}+2 K^{\frac{\delta ^2}{2}+\delta }-(1-\delta )^{(1-\delta )^2} K^{2 \delta }\,.$$ First, we claim that there exists $ K_1>0 $ such that for all $ K>K_1 $, $ h_1(\delta,K) $ is decreasing in $ K $ for every given $ \delta\in (0,1) $. To see this, we compute its derivative with respect to $ K $: $$\frac{\partial h_1}{\partial K} = \delta K^{\delta -1} \left((\delta +2) K^{\frac{\delta ^2}{2}}-2 (1-\delta )^{(\delta -1)^2} K^{\delta }\right)\,.$$ The derivative is negative if $ K > h_2(\delta)\triangleq (\delta -1)^2 \left(\frac{6}{\delta +2}-2\right)^{\frac{2}{(\delta -2) \delta }} $. The function $ h_2(\delta) $ is continuous on $ (0,1) $. Since $ \lim_{\delta\to 0+} h_2(\delta) = e^{3/2} $ and $ \lim_{\delta\to 1-} h_2(\delta) = \frac{9}{4} $, $ h_2(\delta) $ is bounded on $ (0,1) $. Therefore, if we set $ K_1 = \sup_{\delta\in (0,1)} h_2(\delta) $, we have $ \frac{\partial h_1}{\partial K} < 0 $ and thereby $ h_1(\delta,K) $ is decreasing for all $ K > K_1 $. In the second step, we expand $ h_1 $ with respect to $ \delta $ around $ 0 $ and with the Peano form of the remainder $$h_1(\delta,K) = \delta ^2 \left(3 \log (K)-\log ^2(K) + r_K(\delta)\right)\,,$$ where $ \lim_{\delta\to 0+} r_K(\delta) = 0 $. Since $ 3 \log (K)-\log ^2(K) < 0 $ for all $ K\ge 21 $, there exists $ \delta_0(K) > 0 $ such that for all $ \delta \in (0, \delta_0(K)) $, we have $ h_1(\delta, K) < 0 $; in particular, for $ K_2 = \max\{21, K_1\} $, there exists $ \delta_0(K_2) > 0 $ such that for all $ \delta \in (0, \delta_0(K_2)) $, we have $ h_1(\delta, K_2) < 0 $. We use the shorthand $ \delta_2 \in (0,1) $ to denote $ \delta_0(K_2) $. Since for every given $ \delta\in (0,1) $, $ h_1(\delta,K) $ is decreasing for all $ K> K_2 \ge K_1 $, we have $$\label{eq:small_delta} h_1(\delta,K)<0,\quad \forall \delta\in (0,\delta_2) , K>K_2 \,.$$ Let $ h_3(\delta) = (1-\delta )^2 \log (1-\delta ) $. Its derivative is $ h'_3(\delta) = -(1-\delta ) (2 \log (1-\delta )+1) $, which is positive if $ \delta > 1-e^{-\sfrac{1}{2}} $ and negative if $ \delta < 1-e^{-\sfrac{1}{2}} $. Therefore, for all $ \delta\in (0,1) $, we have $ h_3(\delta) \ge h_3(1-e^{-\sfrac{1}{2}}) = -\frac{1}{2 e} $. As a consequence, for all $ \delta\in (0,1) $, we have $$\label{eq:ge_onehalf} (1-\delta)^{(1-\delta)^2} \ge \exp(-\frac{1}{2 e}) > \sfrac{1}{2}\,.$$ If $ K > K_3 = 4^{\frac{1}{\delta_2-\delta_2^2/2}} $, we have for all $ \delta \in [\delta_2,1) $ it holds that $ K > 4^{\frac{1}{\delta_2-\delta_2^2/2}} \ge 4^{\frac{1}{\delta-\delta^2/2}} $ and therefore $$\label{eq:2k} 2K^{\delta^2/2+\delta} - \frac{1}{2} K^{2\delta}<0\,.$$ Hence, for $ \delta\in [\delta_2,1) $ and $ K>K_3 $, the follow inequalities hold $$\label{eq:large_delta} \begin{split} h_1(\delta,K) ={}& -(\delta +1) (1-\delta )^{\delta (\delta +2)}+2 K^{\frac{\delta ^2}{2}+\delta }-(1-\delta )^{(1-\delta )^2} K^{2 \delta }\\ <{} & 2 K^{\frac{\delta ^2}{2}+\delta }-(1-\delta )^{(1-\delta )^2} K^{2 \delta }\\ <{} & 2 K^{\frac{\delta ^2}{2}+\delta }-\frac{1}{2} K^{2 \delta } < 0\,, \end{split}$$ where the first inequality holds because $ (\delta +1) (1-\delta )^{\delta (\delta +2)} > 0 $, the second inequality holds due to , and the final inequality holds because of . Combining and , we deduce that for $ K_4 = \max\{K_2,K_3\} + 1 $ and all $ \delta\in (0,1) $, we have $ h_1(\delta,K_4) < 0 $. Notice that $$h\left( \frac{K_4}{(1-\delta)^2} \right) = \frac{h_1(\delta,K_4)}{(1-\delta )^{\delta (\delta +2)}} < 0\,.$$ Therefore, we have $ a_1 < \frac{K_4}{(1-\delta)^2} $. Since $ \phi'(x) = \frac{1}{\sqrt{2 \pi }} e^{-\frac{1}{2} (\delta +1)^2 x^2} h(e^{x^2}) $, we have $ e^{x_0^2} = a_1 $ and the function $ \phi(x) $ is strictly increasing on $ (0,x_0) $ and strictly decreasing on $ (x_0,\infty) $. Recalling $ \frac{K_4}{(1-\delta)^2} > a_1 > \max \{ e^{3/2} , \frac{1}{(1-\delta)^2} \}$ and setting $ K_0 = \log K_4 >0 $, we have $ \sqrt{K_0+2\log\frac{1}{1-\delta}} > x_0 = \sqrt{\log(a_1)} > \sqrt{ \max\{ \sfrac{3}{2},2\log\frac{1}{1-\delta} \}} $. \[lem:wstd\_wrob\_gap\] Given $ {\varepsilon}>0 $ and training data $ \{(x_i,y_i)\}_{i=1}^n \subseteq \bR^d\times \{\pm 1\} $ with $ n $ data points, if we define the standard and robust classifier by $$\begin{aligned} {w^\textnormal{std}}_n ={}& \operatorname*{arg\,max}_{\|w\|_\infty\le W} \sum_{i=1}^n y_i\langle w,x_i\rangle\,, \\ {w^\textnormal{rob}}_n = {}& \operatorname*{arg\,max}_{\|w\|_\infty\le W} \sum_{i=1}^n \min_{\tilde{x}_i\in B^\infty_{x_i}({\varepsilon})} y_i\langle w,\tilde{x_i}\rangle\,, \end{aligned}$$ we have $ {w^\textnormal{std}}- {w^\textnormal{rob}}= W[\operatorname{sign}(u) - \operatorname{sign}(u-{\varepsilon}\operatorname{sign}(u))] $, where $ u = \frac{1}{n}\sum_{i=1}^n y_i x_i $. The first step is to compute the inner minimization in the expression of the robust classifier. If $ y_i=1 $, the minimizer of $ \min_{\tilde{x}_i\in B^\infty_{x_i}({\varepsilon})} y_i\langle w,\tilde{x_i}\rangle $ is $ x_i-{\varepsilon}\operatorname{sign}(w) $. If $ y_i=-1 $, its minimizer is $ x_i + {\varepsilon}\operatorname{sign}(w) $. Therefore, in both cases, the minimizer is $ x_i - y_i{\varepsilon}\operatorname{sign}(w) $ and $$\min_{\tilde{x}_i\in B^\infty_{x_i}({\varepsilon})} y_i\langle w,\tilde{x_i}\rangle = y_i\langle w, x_i - y_i{\varepsilon}\operatorname{sign}(w)\rangle = y_i(\langle x_i,w\rangle -y_i{\varepsilon}\|w\|_1) = y_i\langle x_i,w\rangle - {\varepsilon}\|w\|_1\,.$$ Thus we have $$\label{eq:equiv-rob} \sum_{i=1}^n \min_{\tilde{x}_i\in B^\infty_{x_i}({\varepsilon})} y_i\langle w,\tilde{x_i}\rangle = \sum_{i=1}^n (y_i\langle x_i,w\rangle - {\varepsilon}\|w\|_1) = n(\langle u,w\rangle - {\varepsilon}\|w\|_1) = n \sum_{\substack{j\in [d]: \\w(j)\ne 0}} (u(j)w(j) - {\varepsilon}|w(j)|)\,,$$ where $ u = \frac{1}{n} \sum_{i=1}^n y_i x_i $. By the definition of the robust classifier, $ {w^\textnormal{rob}}_n $ is a maximizer of . We only consider $ j\in [d] $ such that $ w(j)\ne 0 $. If $ u(j)\ne 0 $, we have $ \operatorname{sign}({w^\textnormal{rob}}_n(j)) = \operatorname{sign}(u(j)) $; otherwise, we can always flip the sign of $ {w^\textnormal{rob}}(j) $ and make larger (note that the first term $ \langle u,w\rangle $ will increase and the second term $ - {\varepsilon}\|w\|_1 $ remains unchanged). If $ u(j)=0 $, to maximize the second term $ - {\varepsilon}\|w\|_1 $, $ {w^\textnormal{rob}}_n(j) $ has to be zero. Therefore, we conclude that $ \operatorname{sign}({w^\textnormal{rob}}_n) = \operatorname{sign}(u) $ and obtain $${w^\textnormal{rob}}_n = \operatorname*{arg\,max}_{\|w\|_\infty\le W} \langle u-{\varepsilon}\operatorname{sign}(u),w\rangle = W\operatorname{sign}(u-{\varepsilon}\operatorname{sign}(u))\,.$$ The standard classifier equals $${w^\textnormal{std}}_n = \operatorname*{arg\,max}_{\|w\|_\infty\le W} \langle w,\sum_{i=1}^n y_ix_i\rangle = W\operatorname{sign}(u)\,.$$ Therefore, we obtain that $ {w^\textnormal{std}}- {w^\textnormal{rob}}= W[\operatorname{sign}(u) - \operatorname{sign}(u-{\varepsilon}\operatorname{sign}(u))] $. Given the training data $ \{(x_i,y_i)\}_{i=1}^n $, \[lem:wstd\_wrob\_gap\] implies that the generalization gap is given by $$\begin{aligned} \bE_{(x,y)\sim \cD}[y\langle {w^\textnormal{std}}-{w^\textnormal{rob}}, x\rangle ] ={}& \frac{\langle {w^\textnormal{std}}-{w^\textnormal{rob}}, \mu\rangle - \langle {w^\textnormal{std}}-{w^\textnormal{rob}}, -\mu\rangle}{2} = \langle {w^\textnormal{std}}-{w^\textnormal{rob}}, \mu\rangle\\ ={}& W\sum_{j\in [d]} \mu(j)[ \operatorname{sign}(u(j)) - \operatorname{sign}(u(j)-{\varepsilon}\operatorname{sign}(u(j))) ] \,, \end{aligned}$$ where $ u = \frac{1}{n}\sum_{i=1}^n y_i x_i $. Note that $ u $ is distributed as $ \cN(\mu,\frac{1}{n}\Sigma) $. We have $ u(j)\sim \cN(\mu(j),\frac{\sigma(j)^2}{n} ) $. Therefore, we deduce $$\begin{aligned} & \bE_{u(j)\sim \cN(\mu(j), \frac{\sigma(j)^2}{n})} [\operatorname{sign}(u(j)) - \operatorname{sign}(u(j)-{\varepsilon}\operatorname{sign}(u(j)))]\\ ={}& 2[\Pr[0<u(j)<{\varepsilon}]-\Pr[-{\varepsilon}<u(j)<0]]\\ ={}& 2[\Phi(\frac{\sqrt{n}}{\sigma(j)}({\varepsilon}-\mu(j))) - \Phi(-\frac{\sqrt{n}}{\sigma(j)}\mu(j)) - \Phi(-\frac{\sqrt{n}}{\sigma(j)}\mu(j)) + \Phi(-\frac{\sqrt{n}}{\sigma(j)}({\varepsilon}+\mu(j))) ]\\ ={}& 2[2\Phi(\frac{\sqrt{n}}{\sigma(j)}\mu(j))- \Phi(\frac{\sqrt{n}}{\sigma(j)}(\mu(j)+{\varepsilon})) - \Phi(\frac{\sqrt{n}}{\sigma(j)}(\mu(j)-{\varepsilon}))]\,, \end{aligned}$$ where $ \Phi $ denotes the CDF of the standard normal distribution. We are in a position to compute $ g_n $: $$\label{eq:gn} \begin{split} g_n ={}& 2W\sum_{j\in [d]:\mu(j)\ne 0} \mu(j) [2\Phi(\frac{\sqrt{n}}{\sigma(j)}\mu(j))- \Phi(\frac{\sqrt{n}}{\sigma(j)}(\mu(j)+{\varepsilon})) - \Phi(\frac{\sqrt{n}}{\sigma(j)}(\mu(j)-{\varepsilon}))]\\ ={}& 2W\sum_{j\in [d]:\mu(j)\ne 0} \mu(j) [2\Phi(\frac{\sqrt{n}}{\sigma(j)}\mu(j))- \Phi(\frac{\sqrt{n}}{\sigma(j)}\mu(j)(1+{\varepsilon}')) - \Phi(\frac{\sqrt{n}}{\sigma(j)}\mu(j)(1-{\varepsilon}'))] \,, \end{split}$$ where $ {\varepsilon}' = \frac{{\varepsilon}}{\mu(j)} $. The second derivative of $ \Phi(x) $ is $ \Phi''(x) = -\frac{e^{-\frac{x^2}{2}} x}{\sqrt{2 \pi }} $ and it is non-positive if $ x\ge 0 $. This implies the concavity of $ \Phi $ on $ [0,\infty) $. By Jensen’s inequality, we have $$\Phi(\frac{\sqrt{n}}{\sigma(j)}\mu(j))- \frac{1}{2}(\Phi(\frac{\sqrt{n}}{\sigma(j)}\mu(j)(1+{\varepsilon}')) + \Phi(\frac{\sqrt{n}}{\sigma(j)}\mu(j)(1-{\varepsilon}'))) \ge 0\,.$$ Therefore $ g_n\ge 0 $. When $ n $ goes to $ \infty $, if $ mu(j)>0 $, $ \frac{\sqrt{n}}{\sigma(j)}\mu(j) $ goes to $ \infty $ as well. By \[lem:phi\], we get $$\lim_{n\to\infty} g_n = 2W \sum_{j\in [d]:\mu(j)>0} \mu(j) H\left( \frac{{\varepsilon}}{\mu(j)}-1 \right)\,.$$ If $ {\varepsilon}< \min_{j\in [d]:\mu(j)>0} \mu(j) $, we have for all $ j\in [d] $ such that $ \mu(j)>0 $, it holds that $ {\varepsilon}<\mu(j) $. Recalling and by \[lem:phi\], we deduce that $ g_n $ is strictly increasing if $$\frac{\sqrt{n}}{\sigma(j)}\mu(j) < \sqrt{ \max\left\{ \frac{3}{2},2\log\frac{1}{1-{\varepsilon}/\mu(j)} \right\}}$$ for $\forall j\in [d] $ such that $ \mu(j)>0 $. In other words, $ g_n $ is strictly increasing when $$n < \min_{j\in [d]:\mu(j)>0} \max\left\{ \frac{3}{2} ,2\log\frac{1}{1-{\varepsilon}/\mu(j)} \right\} \left( \frac{\sigma(j)}{\mu(j)} \right)^2 \,.$$ Since $ \phi(x) $ is strictly decreasing when $ x $ is sufficiently large ([[*i.e.*]{}]{}, $ x \ge \sqrt{K_0 + 2\log \frac{1}{1-\delta}} $, where $ K_0 $ is a universal constant), we have $ g_n $ is strictly decreasing if $$\frac{\sqrt{n}}{\sigma(j)} \mu(j) \ge \sqrt{K_0 + 2\log \frac{1}{1-{\varepsilon}/\mu(j)}}$$ for $ \forall j\in [d] $ such that $ \mu(j) > 0 $. In other words, $ g_n $ is strictly increasing when $$n \ge \max_{j\in [d]: \mu(j)>0}\left( K_0 + 2\log \frac{1}{1-{\varepsilon}/\mu(j)} \right)\left( \frac{\sigma(j)}{\mu(j)} \right)^2\,.$$ If $ {\varepsilon}> \| \mu \|_\infty $, we have for all $ j\in [d] $ such that $ \mu(j)>0 $, it holds that $ {\varepsilon}> \mu(j) $. \[lem:phi\] gives that $ g_n $ is strictly increasing for all $ n\ge 1 $. Proof of \[thm:bernoulli\] {#sec:proof-bernoulli} ========================== \[lem:bet\] There exists a positive constant $ C_0 $ such that if $ X_1,X_2,\dots $ are i.i.d. random variables with $ \bE[X_1]=0 $, $ \bE[X_1^2] = \sigma^2 > 0 $, and $ \bE[|X_1|^3]=\rho<\infty $, and if we define $ Y_n = \frac{1}{n}\sum_{i=1}^n X_i $ and denote the CDF of $ \frac{Y_n\sqrt{n}}{\sigma} $ by $ F_n $, then for all $ x $ and $ n $, $$|F_n(x)-\Phi(x)| \le \frac{C_0\rho}{\sigma^3 \sqrt{n}}\,,$$ where $ \Phi(x) $ is the CDF of the standard normal distribution. @shevtsova2014absolute established the upper bound $ C_0\le 0.4748 $. In the Bernoulli case where the cardinality of the support of $ X_1 $ is $ 2 $, @schulz2016 showed that $ C_0 \le \frac{\sqrt{10}+3}{6\sqrt{2\pi}}\approx 0.4097 $. \[lem:binom\_dist\] If $ X\sim \operatorname{Bin}(n,p) $, $p>\sfrac{1}{2} $, $ \delta >0 $ and $x = \sfrac{n}{2}$, we have $ \Pr[X\in (x,x+\delta)] \ge \Pr[X\in (x-\delta, x)] $. We first note that the two intervals $(x,x+\delta)$ and $(x-\delta, x)$ are symmetric about $x=\sfrac{n}{2}$. Thus if $\exists l \in (x-\delta, x)$, such that $\Pr(X=l) > 0$, [[*i.e.*]{}]{}, $l$ is an integer, then there exists a positive number $k$ such that $l=\sfrac{n}{2}-k$, and $\sfrac{n}{2}+k$ is an integer and falls on $(x,x+\delta)$. Actually, this is a bijection: there is a set $K$ of positive numbers such that $\{\sfrac{n}{2}-k: k \in K\} = (x-\delta,x)\cap \mathbb{Z}$, and $\{\sfrac{n}{2}+k: k \in K\} = (x,x+\delta)\cap \mathbb{Z}$. So we have $$\begin{aligned} \Pr[X\in (x-\delta, x)] &{}= \Pr[X \in \{\sfrac{n}{2}-k: k \in K\}] \\ &{}= \sum_{k \in K} \binom{n}{\sfrac{n}{2}-k}p^{\sfrac{n}{2}-k}(1-p)^{\sfrac{n}{2}+k} \\ &{}\leq \sum_{k \in K} \binom{n}{\sfrac{n}{2}-k}p^{\sfrac{n}{2}+k}(1-p)^{\sfrac{n}{2}-k} \\ &{} = \Pr[X \in \{\sfrac{n}{2}+k: k \in K\}] \\ &{} = \Pr[X\in (x,x+\delta)],\end{aligned}$$ where the inequality holds because $p > \sfrac{1}{2}$. For $ i\in [n] $ and $ j \in [d] $, let $ B_i(j) \stackrel{\textnormal{i.i.d.}}{\sim} \operatorname{Ber}(\frac{1+\tau}{2}) $. We have $ x_i(j) = (2B_i(j)-1)y_i\theta(j) $ and $ \bE[x_i\mid y_i] = y_i\theta\tau $. If we define $ u(j) = \frac{1}{n}\sum_{i=1}^{n} y_ix_i(j) $, we have $$u(j) = \frac{1}{n}\sum_{i=1}^{n} (2B_i(j)-1)\theta(j)\,.$$ Given the training data $ \{(x_i,y_i)\}_{i=1}^n $, \[lem:wstd\_wrob\_gap\] implies that the generalization gap is given by $$\begin{aligned} \bE_{(x,y)\sim \cD}[y\langle {w^\textnormal{std}}-{w^\textnormal{rob}}, x\rangle ] ={}& \frac{\langle {w^\textnormal{std}}-{w^\textnormal{rob}}, \theta\tau\rangle - \langle {w^\textnormal{std}}-{w^\textnormal{rob}}, -\theta\tau\rangle}{2} = \tau\langle {w^\textnormal{std}}-{w^\textnormal{rob}}, \theta\rangle\\ ={}& W\tau \sum_{j\in [d]} \theta(j)[ \operatorname{sign}(u(j)) - \operatorname{sign}(u(j)-{\varepsilon}\operatorname{sign}(u(j))) ] \,. \end{aligned}$$ For $ j\in [d] $ such that $ \theta(j)>0 $, taking the expectation over $ u(j) $, we deduce $$\begin{aligned} & \bE_{u(j)} [\operatorname{sign}(u(j)) - \operatorname{sign}(u(j)-{\varepsilon}\operatorname{sign}(u(j)))]\\ ={}& 2[\Pr[0<u(j)<{\varepsilon}]-\Pr[-{\varepsilon}<u(j)<0]]\\ ={}& 2\left( \Pr\left [\frac{n}{2}< \sum_{i=1}^{n} B_i(j) < \frac{n}{2}\left( \frac{{\varepsilon}}{n}+1 \right) \right ] - \Pr\left [ \frac{n}{2}\left( 1-\frac{{\varepsilon}}{n} \right) < \sum_{i=1}^{n} B_i(j) < \frac{n}{2}\right ] \right)\,.\end{aligned}$$ Since $ x_i(j) = (2B_i(j)-1)y_i\theta(j) $, the sum $ \sum_{i=1}^{n}B_i(j) $ obeys the distribution $ \operatorname{Bin}(n, \frac{1+\tau}{2}) $, where $ \frac{1+\tau}{2} > \frac{n}{2} $. \[lem:binom\_dist\] implies that $$\Pr\left [\frac{n}{2}< \sum_{i=1}^{n} B_i(j) < \frac{n}{2}\left( \frac{{\varepsilon}}{n}+1 \right) \right ] \ge \Pr\left [ \frac{n}{2}\left( 1-\frac{{\varepsilon}}{n} \right) < \sum_{i=1}^{n} B_i(j) < \frac{n}{2}\right ]\,.$$ Therefore, we have $ \bE_{u(j)} [\operatorname{sign}(u(j)) - \operatorname{sign}(u(j)-{\varepsilon}\operatorname{sign}(u(j)))]\ge 0 $. Since $$g_n = W\tau \sum_{j\in [d]} \theta(j) \bE_{u(j)} [\operatorname{sign}(u(j)) - \operatorname{sign}(u(j)-{\varepsilon}\operatorname{sign}(u(j)))]$$ and we assume $ \theta(j)\ge 0 $, we deduce $ g_n\ge 0 $. Next, we show that $ g_n $ is contained in a strip centered at $ s_n $ and with width $ O\left( \frac{1}{\sqrt{n}} \right) $. We have $$\begin{aligned} & \bE_{u(j)} [\operatorname{sign}(u(j)) - \operatorname{sign}(u(j)-{\varepsilon}\operatorname{sign}(u(j)))]\\ ={}& 2[\Pr[0<u(j)<{\varepsilon}]-\Pr[-{\varepsilon}<u(j)<0]]\\ ={}& 2[\Pr[ -\frac{\tau\sqrt{n}}{\sqrt{1-\tau^2}} \le \frac{1}{\sqrt{n}} \sum_{i=1}^{n} X_i \le \frac{(\frac{{\varepsilon}}{\theta(j)}-\tau)\sqrt{n}}{\sqrt{1-\tau^2}} ]- \Pr[-\frac{(\tau+\frac{{\varepsilon}}{\theta(j)})\sqrt{n}}{\sqrt{1-\tau^2}} \le \frac{1}{\sqrt{n}}\sum_{i=1}^{n} X_i \le -\frac{\tau\sqrt{n}}{\sqrt{1-\tau^2}}]]\\ ={}& 2[F_n(\frac{(\frac{{\varepsilon}}{\theta(j)}-\tau)\sqrt{n}}{\sqrt{1-\tau^2}}) + F_n(-\frac{(\tau+\frac{{\varepsilon}}{\theta(j)})\sqrt{n}}{\sqrt{1-\tau^2}}) - 2F_n(-\frac{\tau\sqrt{n}}{\sqrt{1-\tau^2}}) ] \end{aligned}$$ where $ X_i = \frac{2(B_i(j)-\frac{1+\tau}{2})}{\sqrt{(1-\tau^2)}} $ and $ F_n $ is the CDF of $ \frac{1}{\sqrt{n}}\sum_{i=1}^{n}X_n $. The third absolute moment of $ X_1 $ is $ \bE[|X_1|^3] = \frac{\tau ^2+1}{\sqrt{1-\tau ^2}} $. By \[lem:bet\], we get $$\begin{aligned} &\left| \bE_{u(j)} [\operatorname{sign}(u(j)) - \operatorname{sign}(u(j)-{\varepsilon}\operatorname{sign}(u(j)))]\right.\\ & \left. - 2\left( \Phi(\frac{(\frac{{\varepsilon}}{\theta(j)}-\tau)\sqrt{n}}{\sqrt{1-\tau^2}}) + \Phi(-\frac{(\tau+\frac{{\varepsilon}}{\theta(j)})\sqrt{n}}{\sqrt{1-\tau^2}}) - 2\Phi(-\frac{\tau\sqrt{n}}{\sqrt{1-\tau^2}}) \right) \right| \le \frac{8C_0 }{\sqrt{n}} \cdot \frac{\tau ^2+1}{\sqrt{1-\tau ^2}}\,, \end{aligned}$$ where $ \Phi $ denotes the CDF of the standard normal distribution and $ C_0 \le \frac{\sqrt{10}+3}{6\sqrt{2\pi}}\approx 0.4097 $. If we define $ \phi(x,\delta) = 2\Phi(x) - \Phi(x(1+\delta)) - \Phi(x(1-\delta)) $, using the relation $ \Phi(-x) = 1-\Phi(x) $, we have $$\Phi(\frac{(\frac{{\varepsilon}}{\theta(j)}-\tau)\sqrt{n}}{\sqrt{1-\tau^2}}) + \Phi(-\frac{(\tau+\frac{{\varepsilon}}{\theta(j)})\sqrt{n}}{\sqrt{1-\tau^2}}) - 2\Phi(-\frac{\tau\sqrt{n}}{\sqrt{1-\tau^2}}) = \phi\left (\frac{\tau\sqrt{n}}{\sqrt{1-\tau^2}}, \frac{{\varepsilon}}{\theta(j)\tau}\right )\,.$$ Therefore, we obtain $$\left| \bE_{u(j)} [\operatorname{sign}(u(j)) - \operatorname{sign}(u(j)-{\varepsilon}\operatorname{sign}(u(j)))] - 2\phi\left (\frac{\tau\sqrt{n}}{\sqrt{1-\tau^2}}, \frac{{\varepsilon}}{\theta(j)\tau}\right ) \right| \le \frac{8C_0 }{\sqrt{n}} \cdot \frac{\tau ^2+1}{\sqrt{1-\tau ^2}}\,.$$ If we set $$\label{eq:sn} s_n = 2W\tau \sum_{j\in [d]:\theta(j)>0}\theta(j) \phi\left (\frac{\tau\sqrt{n}}{\sqrt{1-\tau^2}}, \frac{{\varepsilon}}{\theta(j)\tau}\right )\,,$$ we have $$\label{eq:gn_sn_bound} \left | g_n - s_n \right | \le \frac{8C_0W\tau \| \theta \|_1(\tau ^2+1)}{\sqrt{n}\sqrt{1-\tau ^2}}\,.$$ \[lem:phi\] implies that $ \lim_{n\to\infty} s_n = 2W\tau \sum_{j\in [d]:\theta(j)>0} \theta(j) H\left( \frac{{\varepsilon}}{\theta(j)\tau} - 1 \right) $. Since we have shown in that $ |g_n-s_n|\le O\left( \frac{1}{\sqrt{n}} \right) $, we deduce $ \lim_{n\to\infty} g_n = 2W\tau \sum_{j\in [d]:\theta(j)>0} \theta(j) H\left( \frac{{\varepsilon}}{\theta(j)\tau} - 1 \right) $. If $ \frac{{\varepsilon}}{\tau} < \min_{j\in [d]:\theta(j)>0} \theta(j) $, we have for $ \forall j\in [d] $ such that $ \theta(j) > 0 $, it holds that $ \frac{{\varepsilon}}{\theta(j)\tau} < 1 $. \[lem:phi\] implies that $ \phi\left (\frac{\tau\sqrt{n}}{\sqrt{1-\tau^2}}, \frac{{\varepsilon}}{\theta(j)\tau}\right ) $ is strictly increasing in $ n $ when $$\frac{\tau\sqrt{n}}{\sqrt{1-\tau^2}} < \sqrt{\max\left \{ \sfrac{3}{2},2\log\frac{1}{1-{\varepsilon}/(\theta(j)\tau)} \right \}}\,,$$ or equivalently when $ n < \left( \frac{1}{\tau^2}-1 \right) \max\{ \sfrac{3}{2},2\log\frac{1}{1-{\varepsilon}/(\theta(j)\tau)} \} $. \[lem:phi\] also implies that it is strictly decreasing when $$\frac{\tau\sqrt{n}}{\sqrt{1-\tau^2}} \ge \sqrt{K_0 + 2\log \frac{1}{1-{\varepsilon}/(\theta(j)\tau)}}\,,$$ or equivalently when $ n \ge \left( \frac{1}{\tau^2}-1 \right) \left( K_0 + 2\log \frac{1}{1-{\varepsilon}/(\theta(j)\tau)} \right) $, where $ K_0 $ is a universal constant. In light of the relation between $ s_n $ and $ \phi\left (\frac{\tau\sqrt{n}}{\sqrt{1-\tau^2}}, \frac{{\varepsilon}}{\theta(j)\tau}\right ) $ shown in , we deduce that $ s_n $ is strictly increasing when $ n < \left( \frac{1}{\tau^2}-1 \right) \max\{ \sfrac{3}{2}, 2\min_{j\in [d]:\theta(j)>0} \log\frac{1}{1-{\varepsilon}/(\theta(j)\tau)} \} $ and that it is strictly decreasing when $ n \ge \left( \frac{1}{\tau^2}-1 \right) \left( K_0 + 2 \max_{j\in [d]: \theta(j)>0} \log \frac{1}{1-{\varepsilon}/(\theta(j)\tau)} \right) $. If $ \frac{{\varepsilon}}{\tau} \ge \| \theta \|_\infty $, we have for $ \forall j\in [d] $ such that $ \theta(j)>0 $, it holds that $ \frac{{\varepsilon}}{\theta(j)\tau} > 1 $. \[lem:phi\] implies that $ \phi\left (\frac{\tau\sqrt{n}}{\sqrt{1-\tau^2}}, \frac{{\varepsilon}}{\theta(j)\tau}\right ) $ is strictly increasing in $ n $ for all $ n\ge 1 $. As a consequence, the sequence $ s_n $ is strictly increasing for all $ n\ge 1 $. Proof of \[ob:regression-wstd-wrob\] {#sec:proof-wrob-regression} ==================================== By definition, we have $$\begin{aligned} {w^\textnormal{rob}}_n &{}= \operatorname*{arg\,min}_{w\in \bR^d}\frac{1}{n}\sum_{i=1}^n \max_{\tilde{x}_i\in B^\infty_{x_i}({\varepsilon})} (y-\langle w, \tilde{x_i} \rangle)^2 \\ &{}= \operatorname*{arg\,min}_{w\in \bR^d}\frac{1}{n}\sum_{i=1}^n \max_{\tilde{x}_i\in B^\infty_{x_i}({\varepsilon})} \sum_{j=1}^d[y_i(j) - w_i(j)\tilde{x_i}(j)]^2 \\ &{}= \operatorname*{arg\,min}_{w\in \bR^d}\frac{1}{n}\sum_{i=1}^n \sum_{j=1}^d \max_{\tilde{x}_i\in B^\infty_{x_i}({\varepsilon})}[y_i(j) - w_i(j)\tilde{x_i}(j)]^2. \end{aligned}$$ Note that the maximum of $[y_i(j) - w_i(j)\tilde{x_i}(j)]^2$ for $\tilde{x_i} \in [x_i-{\varepsilon},x_i+{\varepsilon}]$ can only be obtained at one of the two end points, we have $$\begin{aligned} &{}\max_{\tilde{x}_i\in B^\infty_{x_i}({\varepsilon})}[y_i(j) - w_i(j)\tilde{x_i}(j)]^2 \\ &{}= \max \{[y_i(j)-w_i(j)(x_i(j)\pm {\varepsilon})]^2 \} \\ &{}= \max\{[y_i(j)-w_i(j)x_i(j) \pm w_i(j){\varepsilon}]^2\} \\ &{}= \max \{(y_i - w(j)x_i(j))^2 + (w_i(j){\varepsilon})^2 \pm 2(y_i - w(j)x_i(j))(w_i(j){\varepsilon}) \} \\ &{}= (y_i - w(j)x_i(j))^2 + (w_i(j){\varepsilon})^2 + 2|(y_i - w(j)x_i(j))(w_i(j){\varepsilon})| \\ &{} = \left( |y_i - w(j)x_i(j)| + {\varepsilon}|w(j)| \right)^2.\end{aligned}$$ Combining the two equations above proves \[ob:regression-wstd-wrob\]. Proof of \[ob:regression-gn\] {#sec:proof-regression-gn} ============================= By definition, the generalization error of a linear regression estimator $w_n$ is $$\begin{aligned} L(w_n) &{}= \bE_{x\sim P_x, \delta \sim \cN(0,\sigma^2)}(\langle w^*-w_n, x\rangle + \delta)^2 \\ &{}= \bE_{x\sim P_X}[\langle w^*-w_n, x\rangle]^2 + \bE_{\delta \sim \cN(0,\sigma^2)}[\delta^2] \\ &{}= \bE_{x\sim P_X} (w^*-w_n)^\top xx^\top (w^*-w_n) + \sigma^2 \\ &{}= (w^*-w_n)^\top \bE_{x\sim P_X}[xx^\top] (w^*-w_n)+ \sigma^2\\ &{}= \| w_n - w^* \|^2_{\bE_{x\sim P_X}[xx^\top]}+ \sigma^2,\end{aligned}$$ where the second equation holds because of the independence between $x$ and $\delta$, and $\bE[\delta]=0$. Applying this equation to both ${w^\textnormal{rob}}$ and ${w^\textnormal{std}}$, we have $$\begin{aligned} g_n &{}= L({w^\textnormal{rob}}_n) - L({w^\textnormal{std}}_n) \\ &{}= \| {w^\textnormal{rob}}_n - w^* \|^2_{\bE_{x\sim P_X}[xx^\top]} - \| {w^\textnormal{std}}_n - w^* \|^2_{\bE_{x\sim P_X}[xx^\top]}\,.\end{aligned}$$ Proof of \[thm:g1-minus-infty\] {#sec:proof-g1-minus-infty} =============================== The standard estimator with one sample is given by ${w^\textnormal{std}}_1 = y_1/x_1$. By \[ob:regression-gn\] we can compute $$\begin{aligned} {\mathbb{E}}\left[ ({w^\textnormal{std}}_1-w^*)^2 \right] ={}&\bE \left[ \left( \frac{y_1}{x_1} - w^* \right)^2 \right] \\={}& \bE_{x_1, \delta\sim \cN(0,1)}\left[ \left( \frac{\delta}{x_1} \right)^2 \right] \\={}&\bE_{\delta\sim \cN(0,1)} \left[ \delta^{2}\right] \cdot \bE_{x_1\sim \cN(0,1)} \left[ x_1^{-2}\right] \\={}& 1 \cdot \int_{-\infty}^\infty \frac{1}{\sqrt{2 \pi}} \frac{1}{x^2} e^{-\left(x^2 \right)/2} \ dx \\={}& \infty\end{aligned}$$ where the second equality is by the assumption of the model, the third equality is by Tonelli’s Theorem [@rudin1964principles] and the independence of $x_1$ and the noise $\delta$, the fourth equality follows from the density function of standard Normal, and the last equality holds because the integration of $1/x^2$ is infinite in any neighborhood around zero. For the robust estimator we have $$\begin{aligned} {w^\textnormal{rob}}_1= \operatorname*{arg\,min}_{w \in \bR} \left( |y_1 - x_1 w |+ {\varepsilon}|w|\right)^2 = \operatorname*{arg\,min}_w |y_1 - x_1 w |+ {\varepsilon}|w|, \end{aligned}$$ and the minimizer is given by $$\begin{aligned} {w^\textnormal{rob}}_1 = \begin{cases} \frac{y_1}{x_1} &{} \textnormal{if} \ |x_1| \geq {\varepsilon}\\0 &{} \textnormal{if} \ |x_1| < {\varepsilon}\end{cases}.\end{aligned}$$Therefore we can compute $$\begin{aligned} {\mathbb{E}}\left[ ({w^\textnormal{rob}}_1-w^*)^2 \right] ={}& {\mathbb{E}}_{x_1 , \delta \sim \cN(0,1)} \left[ ({w^\textnormal{rob}}_1-w^*)^2 \right] \\ ={}& \int_{|x| < {\varepsilon}} {\mathbb{E}}_{\delta \sim \cN(0,1)} \left[ \left(0-w^*\right)^2 \right] \frac{1}{\sqrt{2 \pi}} e^{- x^2 /2} \ dx \\ {}& + \int_{|x| \geq {\varepsilon}} {\mathbb{E}}_{\delta \sim \cN(0,1)} \left[ \left(\frac{y_1}{x}-w^*\right)^2 \right] \frac{1}{\sqrt{2 \pi}} e^{- x^2 /2} \ dx \\ ={}& \int_{|x| < {\varepsilon}} \left(w^*\right)^2 \frac{1}{\sqrt{2 \pi}} e^{- x^2 /2} \ dx \\ {}& + \int_{|x| \geq {\varepsilon}} {\mathbb{E}}_{\delta \sim \cN(0,1)} \left[ \left(\frac{\delta}{x}\right)^2 \right] \frac{1}{\sqrt{2 \pi}} e^{- x^2 /2} \ dx \\ ={}& (w^*)^2 \int_{|x| < {\varepsilon}} \frac{1}{\sqrt{2 \pi}} e^{- x^2 /2} \ dx + \int_{|x| \geq {\varepsilon}} \frac{1}{x^2} \frac{1}{\sqrt{2 \pi}} e^{- x^2 /2} \ dx \\ \leq{}& (w^*)^2 \int_{|x| < {\varepsilon}} \frac{1}{\sqrt{2 \pi}} e^{- x^2 /2} \ dx + \frac{1}{{\varepsilon}^2}\int_{|x| \geq {\varepsilon}} \frac{1}{\sqrt{2 \pi}} e^{- x^2 /2} \ dx \\ \leq{}& (w^*)^2 + \frac{1}{{\varepsilon}^2} < \infty\end{aligned}$$where the second equality is again by Tonelli’s Theorem, the third equality is by the assumption of the model, the fourth equality is by distribution of $\delta$, the fifth inequality holds since $1/x^2 \leq 1/{\varepsilon}^2$ for $|x|\geq{\varepsilon}$, and the sixth inequality holds since the integral of the density function is less than or equal to 1. Altogether we have $g_1 = \left({\mathbb{E}}\left[ ({w^\textnormal{rob}}_1-w^*)^2 \right] - {\mathbb{E}}\left[ ({w^\textnormal{std}}_1-w^*)^2 \right]\right)\bE_{x\sim \cN(0,1)}[x^2] = - \infty$. Proof of \[thm:g1-minus-infty-poisson\] {#sec:proof-g1-minus-infty-poisson} ======================================= As shown in \[sec:proof-g1-minus-infty\], we have ${w^\textnormal{std}}_1 = y_1/x_1$, and $$\begin{aligned} {w^\textnormal{rob}}_1 = \begin{cases} \frac{y_1}{x_1} &{} \text{if } |x_1| \geq {\varepsilon}\,, \\0 &{} \text{if } |x_1| < {\varepsilon}\,. \end{cases}\end{aligned}$$ As a result, when $x_1 \geq {\varepsilon}$, we have ${w^\textnormal{std}}_1={w^\textnormal{rob}}_1$. In order to obtain the generalization gap, we only need to consider cases where $x_1 < {\varepsilon}$. Specifically, in this case, we have $$\begin{aligned} ({w^\textnormal{rob}}_1-w^*)^2 - ({w^\textnormal{std}}_1-w^*)^2 &{}= (0-w^*)^2 - (y_1/x_1-w^*)^2 \\ &{}= (w^*)^2 - (\frac{w^*x_1+\delta}{x_1}-w^*)^2 \\ &{}= (w^*)^2 - \frac{\delta^2}{x_1^2}.\end{aligned}$$ Therefore, $$\begin{aligned} g_1 &{}= \left({\mathbb{E}}_{(x,y)\sim \mathcal{D}} \left[ ({w^\textnormal{rob}}_1-w^*)^2 - ({w^\textnormal{std}}_1-w^*)^2 \right]\right)\bE_{x\sim \operatorname{Poisson}(\lambda)+1}[x^2] \\ &{}= \bE_{x\sim \operatorname{Poisson}(\lambda)+1}[x^2] \cdot \sum_{1\leq k < {\varepsilon}}\Pr[x_1=k] \bE[(w^*)^2-\frac{\delta^2}{k^2}] \\ &{}= \bE_{x\sim \operatorname{Poisson}(\lambda)+1}[x^2] \cdot \sum_{1\leq k < {\varepsilon}}\Pr[x_1=k] [(w^*)^2-\frac{1}{k^2}]\end{aligned}$$ Since by assumption $|w^*|\geq 1$, we have $[(w^*)^2-\frac{1}{k^2}] \geq 0,\ \forall k \geq 1$. Thus $g_1$ is non-negative, and also an increasing function with respect to ${\varepsilon}\geq 0$. We note that for $0 \leq {\varepsilon}\leq 1$, we have ${w^\textnormal{std}}_1={w^\textnormal{rob}}_1$, thus $g_1 = 0$. Also, since $[(w^*)^2-\frac{1}{k^2}] \leq (w^*)^2$, we have $$g_1 \leq \bE_{x\sim \operatorname{Poisson}(\lambda)+1}[x^2] \cdot \sum_{1\leq k < {\varepsilon}}\Pr[x_1=k] (w^*)^2 \leq \bE_{x\sim \operatorname{Poisson}(\lambda)+1}[x^2] \cdot (w^*)^2 < \infty.$$ Test Loss of Standard and Robust models ======================================= The focus of this paper is the generalization gap between the standard and adversarially robust models, which is defined as the difference between the test loss of two models. To further demonstrate this gap, we empirically study the test loss of both standard and adversarially robust models and illustrate the test error versus the size of the training dataset. Let $ n $ denote the size of the training dataset in the sequel. [0.5]{} ![Test loss vs. the size of the training dataset under the Gaussian and Bernoulli model in the classification problem.[]{data-label="fig:loss"}](pix/classreward-guas.pdf "fig:"){width="\linewidth"} [0.5]{} ![Test loss vs. the size of the training dataset under the Gaussian and Bernoulli model in the classification problem.[]{data-label="fig:loss"}](pix/classreward-bern.pdf "fig:"){width="\linewidth"} [0.5]{} ![Scaled test loss $ L_n/\bE_{x\sim P_X}[x^2] $ vs. the size of the training dataset (denoted by $ n $) in the linear regression problem. First two plots correspond to $x$ being sampled from the standard normal distribution $\cN(0,1)$ and last two plots correspond to $\operatorname{Poisson}(5) + 1$. Each curve in a plot represents a different choice of ${\varepsilon}$. []{data-label="fig:loss-regr"}](pix/regrloss-gauss.pdf "fig:"){width="\linewidth"} [0.5]{} ![Scaled test loss $ L_n/\bE_{x\sim P_X}[x^2] $ vs. the size of the training dataset (denoted by $ n $) in the linear regression problem. First two plots correspond to $x$ being sampled from the standard normal distribution $\cN(0,1)$ and last two plots correspond to $\operatorname{Poisson}(5) + 1$. Each curve in a plot represents a different choice of ${\varepsilon}$. []{data-label="fig:loss-regr"}](pix/regrloss-gauss-start5.pdf "fig:"){width="\linewidth"} [0.5]{} ![Scaled test loss $ L_n/\bE_{x\sim P_X}[x^2] $ vs. the size of the training dataset (denoted by $ n $) in the linear regression problem. First two plots correspond to $x$ being sampled from the standard normal distribution $\cN(0,1)$ and last two plots correspond to $\operatorname{Poisson}(5) + 1$. Each curve in a plot represents a different choice of ${\varepsilon}$. []{data-label="fig:loss-regr"}](pix/regrloss-pois-small.pdf "fig:"){width="\linewidth"} [0.5]{} ![Scaled test loss $ L_n/\bE_{x\sim P_X}[x^2] $ vs. the size of the training dataset (denoted by $ n $) in the linear regression problem. First two plots correspond to $x$ being sampled from the standard normal distribution $\cN(0,1)$ and last two plots correspond to $\operatorname{Poisson}(5) + 1$. Each curve in a plot represents a different choice of ${\varepsilon}$. []{data-label="fig:loss-regr"}](pix/regrloss-pois-large.pdf "fig:"){width="\linewidth"} In \[fig:loss\], we plot the test loss (also known as the generalization error) versus the size of the training dataset under the Gaussian and Bernoulli data generation model in the classification problem studied in \[sec:classification\]. The test loss is defined in \[sub:problem\_setup\]. All the model parameters are set to be identical to those in \[fig:classification\]. In \[fig:loss-gaussian,fig:loss-bernoulli\], dashed curves represent the test loss of the standard model. Each solid curve represents the test loss of an adversarially robust model with a different ${\varepsilon}$. The generalization gap illustrated in \[fig:gaussian,fig:bernoulli\] is given by the difference between the curves of the corresponding adversarially robust model and the standard model in \[fig:loss-gaussian,fig:loss-bernoulli\], respectively. \[fig:loss-gaussian\] shows the Gaussian data generation model. We observe that the test loss of the standard model converges to $ -1 $ quickly as the size of the training dataset increases. The threshold between the strong and the weak adversary regimes is marked by ${\varepsilon}= 1$. We can see that for ${\varepsilon}>1$, the test loss monotonically approaches $ 1 $ with the size of the training dataset varying from $ 1 $ to $ 20 $. In this regime, more training data hurts the generalization of the robust model. In the weak adversary regime, we have three observations. First, in general, the loss will eventually decrease and go towards $ -1 $. This indicates that the robust models in this regime will reach the standard model in terms of the test loss in the infinite data limit. Second, their convergence to $ -1 $ is slower with an ${\varepsilon}$ larger and more close to the threshold ($ {\varepsilon}=1 $). The curve that corresponds to ${\varepsilon}=1$ converges to $ 1 $ and therefore the generalization gap tends to $ 1 $ for $ {\varepsilon}=1 $. Third, the test loss is not necessarily monotonically decreasing in the training dataset size $n$. Particularly, for ${\varepsilon}=0.9$ and ${\varepsilon}=0.95$, the losses decrease at the initial and final stage and exhibit an increasing trend at the intermediate stage. During this intermediate increasing stage, more training data actually hurts the generalization of the adversarially robust model. The result of the Bernoulli data generation model is shown in \[fig:loss-bernoulli\]. Recall that the values $ \tau = 0.1 $ and $ \tau = 0.2 $ lie in the strong adversary regime, while the values $ \tau = 0.5 $ and $ \tau = 0.7 $ belong to the weak adversary regime. In the weak adversary regime ($\tau = 0.5$ and $\tau =0.7$), the test loss of the standard and robust models declines with more training data and tends to the same limit, resulting in a zero generalization gap. In the strong adversary regime ($\tau = 0.1$ and $\tau=0.2$), with more training data, the test loss of the standard models decreases, while the test loss of the robust models increases. As a result, it results in an expanding generalization gap as presented in \[fig:bernoulli\]. In \[fig:loss-regr\], we illustrate the scaled test loss versus the size of the training dataset for the linear regression model that we considered in \[sec:regression\]. The model parameters are identical to those in \[fig:regression\]. \[fig:regrloss-gauss\] shows the result of the Gaussian data model and \[fig:regrloss-gauss-start5\] is a magnified plot of the same result (for $ n\ge 5 $). We observe that the test loss of the standard model converges to zero quickly. Regarding the robust models, for $ {\varepsilon}$ values less than $ 0.75 $, the test loss decreases to zero with more training data, and it declines more quickly for smaller $ {\varepsilon}$ values. The test loss increases with more data if $ {\varepsilon}$ is large ($ {\varepsilon}=1.0 $); in this case, more training data again hurts the generalization of robust models in the linear regression problem. For the Poisson data (\[fig:regrloss-pois-small\] and \[fig:regrloss-pois-large\]), the test loss of the standard model declines with more training data and tends to zero. Regarding the robust models, recall that $ {\varepsilon}$ values less than or equal to $ 6.0 $ belong to the weak adversary regime, while the remaining $ {\varepsilon}$ values ([[*i.e.*]{}]{}, those greater than $ 6.0 $) belong to the strong adversary regime. In the weak adversary regime, the test loss of robust models decreases with more training data. In contrast, the test loss exhibits an increasing trend in the strong adversary regime and thus the generalization is hurt by more training data. [^1]: First two authors contributed equally.

ArXiv

--- abstract: 'We address the role of short range interactions for spinless fermions in the hyperhoneycomb lattice, a three dimensional (3D) structure where all sites have a planar trigonal connectivity. For weak interactions, the system is a line-node semimetal. In the presence of strong interactions, we show that the system can be unstable to a 3D quantum anomalous Hall phase with loop currents that break time reversal symmetry, as in the Haldane model. We find that the low energy excitations of this state are Weyl fermions connected by surface Fermi arcs. We show that the 3D anomalous Hall conductivity is $e^{2}/(\sqrt{3}ah)$, with $a$ the lattice constant.' author: - 'Sang Wook Kim, Kangjun Seo, and Bruno Uchoa' title: 3D Quantum Anomalous Hall Effect in Hyperhoneycomb Lattices --- *Introduction. $-$* The quantum Hall conductivity describes dissipationless transport of electrons in a system that breaks time reversal symmetry (TRS) due to an external applied magnetic field. In two dimensions (2D), the current is carried through the edges [@Thouless1982], and the Hall conductivity $\sigma_{xy}$ is quantized in units of $e^{2}/h$. In three dimensions (3D), the Hall conductivity is not universal and has an extra unit of inverse length. As shown by Halperin [@Halperin1987], the 3D conductivity tensor on a lattice has the form $\sigma_{ij}=e^{2}/(2\pi h)\epsilon_{ijk}G_{k}$, where $G$ is a reciprocal lattice vector (it could be zero). The realization of the 3D quantum Hall effect has been proposed in systems with very anisotropic Fermi surfaces [@Balicas; @McKerman; @Bernevig], or else in line-node semimetals [@Guinea; @Mullen2015; @Moessner; @Kim], where the Fermi surface has the form of a line of Dirac nodes [@Burkov; @Lu; @Yang; @Rappe; @Weng; @Yu; @Heykikila; @Chen; @Ezawa; @Wang; @Bian; @Bian2; @Chan; @Xie; @Li]. Equally interesting would be to realize the 3D quantum anomalous Hall (QAH) effect [@Xu; @Xu2], where the anomalous Hall conductivity emerges from the topology of the 3D band structure in the absence of Landau levels. The first proposal for a Chern insulator system was the Haldane model [@Haldane1988] on the honeycomb lattice, where loop currents break TRS and can produce a non-zero Chern number in the bulk states. Hyperhoneycomb lattices have the same planar trigonal connectivity of the honeycomb lattice (see Fig. 1a), and hence could provide a natural system for the emergence of a 3D QAH conductivity. This lattice has been experimentally realized in honeycomb iridates [@Modic] as a candidate for the Kitaev model [@Kitaev] and awaits to be realized as a line-node semimetal. In this Letter, we describe the 3D QAH state that emerges from interactions in a hyperhoneycomb lattice with spinless fermions. This state competes with a CDW state, and produces a very anisotropic gap around a line of Dirac nodes in the semimetallic state. Due to a broken inversion symmetry, the QAH gap changes sign along the nodal line, forming Weyl points connected by Fermi arcs [@Wan; @Armitage]. We show that the QAH conductivity of the surface states is $e^{2}/(\sqrt{3}ah)$, with $a$ the lattice constant. ![(color online) a) Hyperhoneycomb lattice with four sublattices, indicated by the different color sites. All sites are trigonally connected with planar links spaced by $120^{\circ}$. The two planes are rotated by $\pi/2$ along the $z$ direction, which has a screw symmetry. b) 3D Brillouin zone of the crystal. In the semi-metallic state, a closed zero energy line of Dirac nodes (Dirac loop) is shown in the red curve on the $k_{z}=0$ plane (gray area). The black arrows indicate the reciprocal lattice vectors. \[fig:unit-cell\]](Fig1){width="0.84\linewidth"} *Lattice model.$-$* We start from ** the tight binding model of the hyperhoneycomb lattice, which has four atoms per unit cell and planar links spaced by $120^{\circ}$, as shown in Fig. 1a. The lattice has three vector generators $\mathbf{a}_{1}=(\sqrt{3},0,0)$, $\mathbf{a}_{2}=(0,\sqrt{3},0)$ and $\mathbf{a}_{3}=(-\sqrt{3}/2,\sqrt{3}/2,3)$, and the corresponding reciprocal lattice vectors $\mathbf{b}_{1}=(2\pi/\sqrt{3},0,-\pi/3)$, $\mathbf{b}_{2}=(0,-2\pi/\sqrt{3},\pi/3)$ and $\mathbf{b}_{3}=(0,0,2\pi/3)$. For a model of spinless fermions, which could physically result from a strong Rashba spin orbit coupling, the kinetic energy is $\mathcal{H}_{0}=-t\sum_{\left\langle i,j\right\rangle }(a_{i}^{\dagger}a_{j}+h.c.)$, where $a_{i}$ destroys an electron on site $i$, $t$ is the hopping energy and $\langle ij\rangle$ denotes nearest neighbor (NN) sites. In the four-sublattice basis, the Hamiltonian is a 4$\times$4 matrix [@Mullen2015] $$\mathcal{H}_{0}=-t\left(\begin{array}{cccc} 0 & \Theta_{x} & 0 & \text{e}^{ik_{z}}\\ \Theta_{x}^{*} & 0 & \text{e}^{-ik_{z}} & 0\\ 0 & \text{e}^{ik_{z}} & 0 & \Theta_{y}\\ \text{e}^{-ik_{z}} & 0 & \Theta_{y}^{*} & 0 \end{array}\right),\label{eq:Ho}$$ where $\Theta_{\gamma}\equiv2\text{e}^{ik_{z}/2}\cos(\sqrt{3}k_{\gamma}/2)$, with $\gamma=x,\,y$, and **$\mathbf{k}$** is the momentum away from the center of the Brillouin zone (BZ). The electronic structure has a doubly degenerate zero energy line of nodes in the form of a Dirac loop at the $k_{z}=0$ plane, **$\mathbf{k}_{0}(s)\equiv(k_{x}(s),k_{y}(s),0)$** in some parametrization that satisfies the equation $4\cos(\sqrt{3}k_{x}(s)/2)\cos(\sqrt{3}k_{y}(s)/2)=1$, as schematically depicted in Fig. 1b. The projected low energy Hamiltonian has the form $$\mathcal{H}_{0,p}(\mathbf{q})=\left[v_{x}(s)q_{x}+v_{y}(s)q_{y}\right]\sigma_{x}+v_{z}(s)q_{z}\sigma_{y},\label{eq:2}$$ where $\mathbf{q}\equiv\mathbf{k}-\mathbf{k}_{0}(s)$ is the momentum away from the nodal line, $\sigma_{x},\,\sigma_{y}$ are Pauli matrices, with $v_{x}(s)=\frac{\sqrt{3}}{2}t\sin(\sqrt{3}k_{x}(s)/2)/(1+\alpha^{2})$, $v_{y}(s)=\frac{\sqrt{3}}{2}\alpha^{2}t\sin(\sqrt{3}k_{y}(s)/2)/(1+\alpha^{2})$ and $v_{z}=-3t\alpha/(1+\alpha^{2})$ the quasiparticle velocities, and $\alpha(s)=2\cos(\sqrt{3}k_{x}(s)/2)$. Hamiltonian (\[eq:2\]) corresponds to the low energy spectrum $$\epsilon_{0}(\mathbf{q})=\sqrt{(v_{x}q_{x}+v_{y}q_{y})^{2}+v_{z}^{2}q_{z}^{2}},\label{epsilon0}$$ that is gapless along the nodal line. ![(color online) $xz$ (a) and $yz$ (b) planes of the hyperhoneycomb lattice, with sublattices $A,\,B\,,C$ and $D$. Complex NNN hopping terms $\chi_{ij}$ give rise to current loops with flux $\Phi$. The lowest energy state has $\Phi_{1}=-\Phi_{2}$, which corresponds to a zero total flux in the unit cell, with purely imaginary $\chi_{ij}$. \[fig:Phase-diagram-1\]](Fig2){width="0.95\linewidth"} The total Hamiltonian is $\mathcal{H}=\mathcal{H}_{0}+\mathcal{H}_{I}$, where $$\mathcal{H}_{I}=V_{1}\!\sum_{\left\langle i,j\right\rangle }(\hat{n}_{i}-1)(\hat{n}_{j}-1)+V_{2}\negthickspace\!\sum_{\left\langle \left\langle i,j\right\rangle \right\rangle }(\hat{n}_{i}-1)(\hat{n}_{j}-1),\label{HI}$$ is the interaction term, with $\hat{n}_{i}=a_{i}^{\dagger}a_{i}$ the density operator on site $i$, and $V_{1}$ and $V_{2}$ are the repulsion between NN and next-nearest neighbors (NNN) sites, respectively. For spinless fermions, one possible instability is a charge density wave (CDW) state that corresponds to a charge imbalance among the different sublattices. The CDW state is defined by the four component order parameter $\rho_{\alpha}=\langle a_{i}^{\dagger}a_{i}\rangle-\rho_{0}$ with $i\in\alpha$ belonging to sublattice $\alpha=A\,,B,\,C,\,D$, as shown in Fig 2, and $\rho_{0}$ a uniform density. At the neutrality point, the local densities at the four sites of the unit cell add up to zero, $\sum_{\alpha}\rho_{\alpha}=0$. The nodal line is protected by a combination of TRS and mirror symmetry along the $z$ axis. The state where $\rho_{A}=-\rho_{B}=\rho_{C}=-\rho_{D}$, namely $(\rho,-\rho,\rho,-\rho)$, breaks the mirror symmetry and opens the largest gap among all possible charge neutral configurations of $\rho_{\alpha}$. The more symmetric state $(\rho,\rho,-\rho,-\rho)$ does not open a gap. Hence, the former state is the dominant CDW instability. We will not consider other possible states that enlarge the size of the unit cell [@Grushin], such as an $n$-site CDW state, with $n>4$. The other dominant instability is the QAH state, where complex hopping terms between NNN sites lead to loop currents in the $xz$ and $yz$ planes, as shown in Fig. 2. Each plane can have loop currents with opposite flux ($\Phi$), producing zero magnetic flux in the unit cell, in analogy with the 2D case in the honeycomb lattice [@Haldane1988]. The QAH order parameter is defined as $\chi_{ij}=\langle a_{i}^{\dagger}a_{j}\rangle$, where $i$ and $j$ sites are connected by NNN vectors [@Raghu2008]. We define the Ansatz $\chi_{ij}=\chi\text{e}^{i\phi_{ij}}$ for $i,\,j\in\{A,\,C\}$ sublattices and $\chi_{ij}=\chi\text{e}^{i\bar{\phi}_{ij}}$ for $i,\,j\in\{B,D\}$, where $\chi$ is real. Due to particle-hole symmetry, $\chi_{ij}$ is purely imaginary and hence $\phi,\bar{\phi}=\pm\frac{\pi}{2}$. The state that minimizes the free energy of the system has total zero flux in the unit cell, $\Phi_{1}=-\Phi_{2}$ (see Fig. 2), when the magnetic flux lines can more easily close. The QAH order parameter is $\chi_{ij}=\pm i\chi$ for NNN sites and zero otherwise, with the $+$ sign following the convention of the arrows in Fig. 2. We perform a mean-field decomposition of the NN interaction in the CDW state ($\rho$) and of the NNN repulsion in the QAH order parameter $\chi_{ij}$. For simplicity, we absorb the couplings $V_{1}$ and $V_{2}$ in the definition of the order parameters, $\rho V_{1}\to\rho$ and $\chi V_{2}\to\chi$, which have units of energy from now on. The effective interaction in ** the four-sublattice basis is $$\mathcal{H}_{I}^{{\rm MF}}=\left(\begin{array}{cccc} \chi g-3\rho & 0 & -\chi f & 0\\ 0 & -\chi g+3\rho & 0 & \chi f^{*}\\ -\chi f^{*} & 0 & \chi g-3\rho & 0\\ 0 & \chi f & 0 & -g\chi+3\rho \end{array}\right)\!,\label{eq:Ho-1}$$ where $$g(\mathbf{k})=2\left[\sin\left(\sqrt{3}k_{x}\right)+\sin\left(\sqrt{3}k_{y}\right)\right],\label{g}$$ and $$f(\mathbf{k})=2\left[\text{e}^{i3k_{z}/2}\sin\left(\sqrt{3}k_{x}/2\right)+\text{e}^{-i3k_{z}/2}\sin\left(\sqrt{3}k_{y}/2\right)\right].\label{f}$$ The mean-field Hamiltonian $\mathcal{H}^{\text{MF}}=\mathcal{H}_{0}+\mathcal{H}_{I}^{\text{MF}}$ has an additional constant energy term $E_{0}=6\rho^{2}/V_{1}+16\chi^{2}/V_{2}$ that is reminiscent of the decomposition of the interactions to quadratic form. The phase diagram follows from the numerical minimization of the free energy $F$ with respect to $\rho$ and $\chi$ at zero temperature, $\partial F/\partial\chi=\partial F/\rho=0$. The semimetal state is unstable to a CDW order at the critical coupling $V_{1,c}=0.41t$, and to a QAH phase at $V_{2,c}=1.51t$. The CDW and QAH states compete with each other, as shown in Fig. 3. Fluctuation effects are expected to be less dramatic in 3D compared to the more conventional 2D case [@Raghu2008; @Motruck; @Scherer]. Hence, the mean-field phase diagram is likely a reliable indication of the true instabilities of the fermionic lattice for the spinless case. In real crystals, screening and elastic effects lead to a distortion of the lattice in the CDW state, in order to minimize the Coulomb energy due to electron-ion coupling, which can be high [@Cowley]. While the CDW appears to be the leading instability over the QAH state, the elastic energy cost to displace the ions and equilibrate the charge in the electron-ion system may hinder the CDW order and favor the QAH phase when $V_{2}>V_{2,c}$. ![(color online) a) Mean field phase diagram for spinless fermions. The node-line semimetal phase (NLSM) turns into the CDW state at the critical value $V_{1}=0.41t$ and into the QAH phase in 3D at $V_{2}=1.51t$. The CDW is fully gapped, while the QAH phase has nodes around the Dirac loop. \[fig:Phase-diagram\]](Fig3){width="0.56\linewidth"} *Low energy Hamiltonian*$.-$Integrating out the two high energy bands using perturbation theory, the effective low energy Hamiltonian (\[eq:2\]) of the nodal line becomes massive, as expected. The leading correction to Hamiltonian (2) around the nodal line to lowest order in $\rho$ and $\chi$ has the form of a mass term $$\mathcal{H}_{I,p}\left(\mathbf{q}\right)=-\left[3\rho+m(\mathbf{k}_{0})+v_{x}^{\prime}q_{x}+v_{y}^{\prime}q_{y}\right]\sigma_{3},\label{eq:deltaH}$$ where $$m(\mathbf{k}_{0})=\chi\left(g(\mathbf{k}_{0})+\frac{2}{\alpha+\frac{1}{\alpha}}f(\mathbf{k}_{0})\right)\label{eq:m}$$ gives the QAH mass at the nodal line, with $v_{\gamma}^{\prime}(s)=2\chi\{\cos(\sqrt{3}k_{\gamma}(s))+\frac{1}{\alpha+1/\alpha}\cos(\sqrt{3}k_{\gamma}(s)/2)\}$ and $\alpha(s)$ defined below Eq. (\[eq:2\]). The low energy spectrum is $$\epsilon(\mathbf{q})=\pm\sqrt{\epsilon_{0}^{2}(\mathbf{q})+\left[3\rho+m(\mathbf{k}_{0})+v_{x}^{\prime}q_{x}+v_{y}^{\prime}q_{y}\right]^{2}},\label{eq:epsilon}$$ which describes either a uniformly gapped state in the CDW phase ($\rho\neq0$, $\chi=0$) or a non-uniform QAH gap $(\rho=0,\chi\neq0)$ with six nodes at the zeros of $m(\mathbf{k}_{0})$, as indicated in Fig. 4. The CDW state breaks mirror symmetry along the $z$ axis, but preserves the screw axis symmetry and hence creates a fully gapped state that is rotationally symmetric along the nodal line. The QAH state on the order hand breaks inversion symmetry. The mass term (\[eq:m\]) changes sign at six zeros along the nodal line, as shown in Fig. 4b. Two zeros are located along the diagonal direction of the nodal line, at the points $\pm\mathbf{Q}_{1}=\pm\left(-\frac{2\pi}{3\sqrt{3}},\frac{2\pi}{3\sqrt{3}},0\right)$. The other four zeros of $m(\mathbf{k}_{0})$ are symmetrically located around that direction, at $\pm\mathbf{Q}_{2}=\pm(Q_{+},Q_{-},0)$ and $\pm\mathbf{Q}_{3}=\mp(Q_{-},Q_{+},0)$, as shown in Fig. 4, with $Q_{\pm}=\frac{1}{\sqrt{3}}\text{arccos}(\frac{\sqrt{17}-1}{4})\pm\frac{1}{\sqrt{3}}\arccos(\frac{3-\sqrt{17}}{4})$. The position of the nodal points extracted from the low energy Hamiltonian (\[eq:deltaH\]) is in agreement with the values calculated numerically from Hamiltonians (\[eq:Ho\]) and (\[eq:Ho-1\]) in the regime where $\chi\ll t$. For larger values of $\chi$, the nodal points $\pm\mathbf{Q}_{2}$ and $\pm\mathbf{Q}_{3}$ can move in the $k_{z}=0$ plane, as the position of the nodal line is renormalized by the interactions. The two nodal points in the diagonal $\pm\mathbf{Q}_{1}$ remain fixed. ![(color online) a) Contour plot of the mass gap of the QAH state (\[eq:m\]) around the Dirac loop (red line). Momenta are in units of $\pi$. The gap vanishes at six points along the nodal line indicated by the blue dots, where the contours collapse. b) Schematic picture of the sign of the QAH gap around the nodal line: blue line ($m>0$); orange line ($m<0$). At the nodes, the low energy excitations are Weyl fermions, with helicities $\gamma=+1$ (black dots) and $\gamma=-1$ (white). The Weyl points are intercepted by four planes oriented in the $(1\bar{1}0)$ direction (diagonal lines). Those planes form domain walls separating slices of the BZ with distinct Chern numbers. Gray area: $\nu=1$. Dark gray: $\nu=-1$. Light blue: $\nu=0$. \[fig:mass gap\]](Fig4){width="0.97\linewidth"} {width="0.81\linewidth"} Expanding the the mass term around the zeros of $m(\mathbf{k}_{0})$, the low energy quasiparticles around the nodes are Weyl fermions. Performing a rotation of the quasiparticle momenta into a new basis $p_{x}=(q_{x}-q_{y})/\sqrt{2},$ $p_{y}=-q_{z}$ and $p_{z}=(q_{x}+q_{y})/\sqrt{2}$, the expansion around the the nodes at $\pm\mathbf{Q}_{1}$ gives the low energy Hamiltonian $$\mathcal{H}_{\pm Q_{1}}(\mathbf{p})=\mathbf{h}_{\pm Q_{1}}(\mathbf{p})\cdot\vec{\sigma}=\sum_{i=x,y,z}v_{0,i}^{(\pm)}p_{i}\sigma_{i},\label{eq:hWeyl}$$ with $\mathbf{p}$ the momentum away from the nodes and $v_{0,x}^{(\pm)}=\pm3\sqrt{2}t/4,\:v_{0,y}=3t/2$ and $v_{0,z}=\sqrt{3/2}\chi$ the corresponding velocities in the rotated basis. The equation above describes two Weyl points with opposite helicities $\gamma=(2\pi)^{-2}\int_{\Omega}\text{d}^{2}p\,\hat{h}\cdot(\partial_{p_{x}}\hat{h}\times\partial_{p_{y}}\hat{h})=\pm1$, and hence broken TRS, with $\hat{h}=\mathbf{h}/|\mathbf{h}|$ a unitary vector and $\Omega$ the surface of a small sphere enclosing each node. Similarly, the expansion around the nodes $\pm\mathbf{Q}_{2}$ and $\pm\mathbf{Q}_{3}$ give Hamiltonians of Weyl fermions with helicities $\pm1$, as indicated in Fig. 4b. *Anomalous Hall conductivity$.-$* The Weyl points delimit a topological domain wall between slices of the BZ parallel to the $(1\bar{1}0)$ plane. Each slice in the light gray region in Fig. 4b crosses the nodal line twice and has a well defined Chern number $\nu=+1$. The slices in the dark gray regions across the domain walls have opposite Chern number $\nu=-1$, as the QAH mass changes sign simultaneously at the two Weyl points (with the same helicity) where each domain wall intersects the nodal line. The BZ slices in the light blue region do not cross the nodal line and have zero Chern number. The 3D QAH conductivity is defined as $\sigma_{ij}=(e^{2}/h)(2\pi)^{-3}\int_{BZ}\text{d}^{3}k\sum_{n\in\textrm{filled}}(\frac{\partial}{\partial k_{i}}A_{j}-\frac{\partial}{\partial k_{j}}A_{i})$, where $A_{j}=-i\left\langle \psi_{n}\right|\frac{\partial}{\partial k_{j}}\left|\psi_{n}\right\rangle $ is the Berry connection of the $n$-th occupied Block band integrated over the entire BZ [@Haldane2]. For the hyperhoneycomb lattice in the QAH state, $$\sigma_{ij}=\frac{e^{2}}{2\pi h}\int_{C}\text{d}k_{k}\epsilon_{ijk}\nu_{(k)}(\mathbf{k}_{0})=\frac{e^{2}}{2\pi h}\epsilon_{ijk}(\mathbf{b}_{1}+\mathbf{b}_{2})_{k},\label{eq:sigmaij}$$ where $\mathbf{b}_{1}+\mathbf{b}_{2}=\left(2\pi/\sqrt{3},-2\pi/\sqrt{3},0\right)a^{-1}$ is a reciprocal lattice vector, restoring the lattice constant $a$. $\nu_{(j)}(\mathbf{k}_{0})=0,\,\pm1$ is the Chern number of a slice of the BZ oriented in the $j=x,y,z$ direction, intersecting the nodal line $\mathbf{k}_{0}(s)$ at two points, and $C\in[k_{j,\text{min}}(s),k_{j,\text{max}}(s)]$. Therefore, we find that $$\sigma_{yz}=\sigma_{xz}=e^{2}/(\sqrt{3}ha),\label{sigmayz}$$ while $\sigma_{xy}=0$. In the 3D QAH phase, the bulk of the system is a semimetal with topologically protected Weyl quasiparticles [@Xu], while charge currents spontaneously emerge on the $[100]$ and $[010]$ surfaces of the crystal. *Surface states$.-$* The presence of Weyl points in the QAH state implies in the existence of Fermi arcs on the surfaces of the lattice, connecting nodes with opposite helicities. In Fig. 5a, we numerically calculate the Fermi arcs in the (001) surface Brillouin zone, as shown in the solid blue lines. The nodes at $\pm\mathbf{Q}_{2}$ are connected by a Fermi arc crossing the center of the BZ, while the pair of nodes at $\mathbf{Q}_{1}$, $-\mathbf{Q}_{3}$ and $-\mathbf{Q}_{1}$, $\mathbf{Q}_{3}$ are connected by short Fermi arcs directed along the nodal line. In Fig. 5b, we scan the energy spectrum of the $k_{z}=0$ plane along the $k_{x}$ axis along three paths indicated by the dotted horizontal lines in panel 5a. Line 1 ($k_{y}=3\pi/(4\sqrt{3}$) intersects a Fermi arc close to the node at $\mathbf{Q}_{1}$, as indicated by the arrow in the left panel of Fig. 5b, which has a zero energy crossing in the vicinity of a node. The scan on line 2, at $k_{y}=\pi/(2\sqrt{3})$, does not intercept a Fermi arc, as shown in the center panel of Fig. 5b. The third path at $k_{y}=\pi/(10\sqrt{3})$ crosses the Fermi arc near the center of the zone, as indicated by the zero energy mode shown in the right panel of Fig. 5b. *Conclusions$.-$* We have shown that hyperhoneycomb lattices with spinless fermions may host a 3D QAH effect, which competes with a CDW state. The 3D anomalous Hall conductivity is $e^{2}/(\sqrt{3}ha)$. Due to the symmetry of the mass term, which spontaneously breaks inversion symmetry around the nodal line, the low energy excitations of the QAH state have a rich structure, with Weyl fermions in bulk and topologically protected surface states. *Acknowledgements.$-$* We acknowledge F. Assaad, S. Parameswaran and K. Mullen for helpful discussions. S. W. Kim and B.U. acknowledge NSF CAREER Grant No. DMR- 1352604 for support. K. S. acknowledges the University of Oklahoma for support. *Note.$-$* Recently, we became aware of a related work [@Murakami], which predicted the conditions for the emergence of Weyl points in nodal-line semimetals from symmetry arguments. [10]{} D. J. Thouless, M. Kohmoto, M. P. Nightingale, and M. Den Nijs, Phys. Rev. Lett. **49**, 405 (1982). B. I. Halperin, Jpn. J. Appl. Phys. **26** (1987). L. Balicas, G. Kriza, and F. I. B. Williams, Phys. Rev. Lett. **75**, 2000 (1995). S. K. McKernan, S. T. Hannahs, U. M. Scheven, G. M. Danner, and P.M. Chaikin, Phys. Rev. Lett. **75**, 1630 (1995). B. A. Bernevig, T. L. Hughes, S. Raghu, and D. P. Arovas, Phys. Rev. Lett. **99**, 146804 (2007). F. Guinea and D. Arovas, Phys. Rev. B **78**, 245416 (2008). K. Mullen, B. Uchoa, and D. T. Glatzhofer, Phys. Rev. Lett. **115**, 026403 (2015). L.-K. Lim and R. Moessner, Phys. Rev. Lett. **118**, 016401 (2017). J.-W. Rhim and Y. B. Kim, Phys. Rev. B **92**, 045126 (2015). A. A. Burkov, M. D. Hook and L. Balents, Phys. Rev. B **84**, 235126 (2011). L. Lu, L. Fu, J. D. Joannopoulos and M. Soljačić, Nat. Photonics **7**, 294 (2013). S. A. Yang, H. Pan, and F. Zhang, Phys. Rev. Lett. **113**, 046401 (2014). Y. Kim, B. J. C. Wieder, C. L. Kane, and A. Rappe, Phys. Rev. Lett. **115**, 036806 (2015). H. Weng, Y. Liang, Q. Xu, Y. Rui, Z. Fang, X. Dai and Y. Kawa, Phys. Rev. B **92**, 045108 (2015). R. Yu, H. Weng, Z. Fang, X. Dai, and X. Hu, Phys. Rev. Lett. **115**, 036807 (2015). T. T. Heikkila and G. E. Volovik, JETP Lett. **93**, 59 (2011). Y. Chen, Y. Xie, S. A. Yang, H. Pan, F. Zhang, M. L. Cohen, and S. Zhang, Nano Letters **15**, 6974 (2015). L. S. Xie, L. M. Schoop, E. M. Seibel, Q. D. Gibson, W. Xie, and R. J. Cava, APL Mater. **3**, 083602 (2015). M. Ezawa, Phys. Rev. Lett. **116**, 127202 (2016). J.-T. Wang, H. Weng, S. Nie, Z. Fang, Y. Kawazoe, and C. Chen, Phys. Rev. Lett. **116**, 195501 (2016). G. Bian *et.al.*, Nat. Commun. **7**, 10556 (2016). G. Bian *et al.*, Phys. Rev. B **93**, 121113(R) (2016). Y.-H. Chan, C.-K. Chiu, M.Y. Chou, and A. P. Schnyder, Phys. Rev. B **93**, 205132 (2016). R. Li, H. Ma, X. Cheng, S. Wang, D. Li, Z. Zhang, Y. Li, and X.-Q. Chen, Phys. Rev. Lett. **117**, 096401 (2016). G. Xu, H. Weng, Z. Wang, X. Dai, and Z. Fang, Phys. Rev. Lett. **107**, 186806 (2011). G. Xu, J. Wang, C. Felser, X.-L. Q, and S.-C. Zhang, Nano Lett. **15**, 2019 (2015). F. D. M. Haldane, Phys. Rev. Lett. **61**, 18 (1988). K. A Modic et al., Nat. Commun. **5**, 1 (2014). A. Kitaev, Ann. Phys. (Amsterdam) **321**, 2 (2006). X. Wan, A. M. Turner, A. Vishwanath, and S. Y. Savrasov, Phys. Rev. B **83**, 205101 (2011). N. P. Armitage, E. J. Mele, Ashvin Vishwanath, arXiv:1705.01111 (2017). A. G. Grushin, E. V. Castro, A. Cortijo, F. de Juan, M. A. H. Vozmediano, B. Valenzuela, Phys. Rev. B **87**, 085136 (2013). S. Raghu, X.-L. Qi, C. Honerkamp, and S.-C. Zhang, Phys. Rev. Lett. **100** 156401 (2008). J. Motruk, A.G. Grushin, F. de Juan, and F. Pollmann, Phys. Rev. B **92**, 085147 (2015). D.D. Scherer, M.M. Scherer, C. Honerkamp, Phys. Rev. B **92**, 155137 (2015). R. D. Cowley, Adv. Phys. **29**, 1 (1980). F. D. M. Haldane Phys. Rev. Lett. **93**, 206602 (2004). R. Okugawa and S. Murakami, arXiv:1706.08551 (2017).

ArXiv

--- abstract: 'We introduce a new method for reconstructing the primordial power spectrum, $P(k)$, directly from observations of the Cosmic Microwave Background (CMB). We employ Singular Value Decomposition (SVD) to invert the radiation perturbation transfer function. The degeneracy of the multipole $\ell$ to wavenumber $k$ linear mapping is thus reduced. This enables the inversion to be carried out at each point along a Monte Carlo Markov Chain (MCMC) exploration of the combined $P(k)$ and cosmological parameter space. We present best–fit $P(k)$ obtained with this method along with other cosmological parameters.' author: - Gavin Nicholson - 'Carlo R. Contaldi' - Paniez Paykari bibliography: - 'paper.bib' title: Reconstruction of the Primordial Power Spectrum by Direct Inversion --- Introduction ============ The primordial power spectrum of scalar, curvature perturbations $\Phi({\vec k})$ is defined as, $$\label{eq:pkdef} P(k) \equiv \frac{k^3}{2\pi^2}\delta^3(\vec{k}-\vec{k'})\langle\Phi(\vec{k})\Phi^*(\vec{k'})\rangle ,$$ where $k\equiv |\vec k|$ is the wavenumber. The spectrum encodes the initial conditions for the system of coupled Einstein–Boltzmann equations which describe the evolution of density and radiation perturbations about the FRW background. The spectrum itself is considered as a unique window into the era approaching the Planck time. In most models a period of scalar field driven inflation is used to solve the cosmological problems and set the near scale-invariant form of $P(k)$ prior to the radiation dominated epoch. In this case a simple power law parametrised by an amplitude $A_s$ and spectral index $n_s$ suffices to describe the initial conditions to sufficient accuracy for the current data. Higher order contributions such as a mild curvature $d n_s/d\ln k$ have also been explored although it is not strictly required by the best–fit models. A separate approach to studying the physics behind this early phase is to drop any model dependent assumptions of near scale invariance and allow more general functional forms of the initial spectrum. One drawback of this approach is the increase in parameter space which needs to be explored which increases the complexity of the data fitting step. A second drawback is the limited information content of the observations; the effect of sample variance, limited range of scales probed, and the degeneracy in the mapping of $k$ to angular multipoles on the sky means that one cannot expect to constrain arbitrarily complex functions with many degrees of freedom. However there is scope to go beyond the model dependent approach. This is particularly true if one is interested in constraining the presence of features on the spectrum. These could be in the form of ‘glitches’ or step–like features which would be otherwise unconstrained. In fact many models have been proposed which predict features on the spectrum due to, for example, features on the inflaton potential [@Starobinsky:1992ts; @Adams:2001vc; @Wang:2002hf; @Hunt:2004vt; @Joy:2007na; @Hunt:2007dn; @Pahud:2008ae; @Lerner:2008ad], a small number of $e$-folds [@Contaldi:2003zv; @Powell:2006yg; @Nicholson:2007by], or other more exotic sources of non-standard behaviour [@Lesgourgues:1999uc; @Feng:2003zua; @Mathews:2004vu; @Jain:2008dw; @Romano:2008rr]. There are two approaches to reconstructing $P(k)$; parameterisation and direct inversion. None of the various methods have shown conclusive evidence for a departure from near scale-invariance of $P(k)$. Despite this there have been tantalising hints of anomalous features in the CMB. One example of this is that the first year WMAP results gave an indication of a cut-off in $P(k)$ on large scales. With subsequent data releases the significance of this feature has been reduced, although future observations of the polarisation of the CMB may provide more conclusive evidence [@Nicholson:2007by; @Jain:2009pm]. In [@Nicholson:2009pi] we also showed evidence for a dip in power at $k\approx0.002$Mpc$^{-1}$. Numerous parametertic searches for features with a similar form to those in complex inflationary models have been performed along with simple binning of $P(k)$ [@Bridle:2003sa; @Contaldi:2003zv; @Parkinson:2004yx; @Sinha:2005mn; @Sealfon:2005em; @Mukherjee:2005dc; @Bridges:2005br; @Bridges:2006zm; @Covi:2006ci; @Joy:2008qd; @Verde:2008zz]. Methods of direct inversion which make no assumptions about the early universe model being probed [@Hannestad:2000pm; @Wang:2000js; @Matsumiya:2001xj; @Shafieloo:2003gf; @Bridle:2003sa; @Kogo:2003yb; @Mukherjee:2003cz; @Mukherjee:2003ag; @Hannestad:2003zs; @Kogo:2004vt; @TocchiniValentini:2004ht; @Leach:2005av; @Shafieloo:2006hs; @Shafieloo:2007tk; @Nagata:2008tk; @Nagata:2008zj; @Nicholson:2009pi] are hampered by the singular nature of the transfer function that takes $P(k)$ and transfers it onto the CMB or LSS modes. In general this causes the process of estimation to be prohibitively slow so as not to allow joint estimation of a free $P(k)$ with cosmological parameters. Instead one usually assumes a set of cosmological parameters, this allows the use of a fiducial CMB photon transfer function to integrate the primordial curvature perturbation into today’s photon distribution perturbation. However, this hides a significant degeneracy between features in the primordial power spectrum, and the physical parameters which determine the height and position of acoustic peaks in the CMB. It has been pointed out [@Hu:2003vp; @Nicholson:2009pi; @Mortonson:2009qv] that adding polarisation information or LSS data can help break some of the degeneracies. In this paper we propose a new method for direct inversion of $P(k)$ from CMB observation using Singular Value Decomposition (SVD). The method is fast enough to allow us to carry out a joint estimation of $P(k)$ and the cosmological parameters. The form of $P(k)$ is derived from the SVD inversion. We also show that there are regions of $k$ for which polarisation data has the potential to more accurately constrain $P(k)$. The paper is organised as follows; in section \[method\] we introduce the SVD method and test it against known input models in section \[results\]. We also show current constraints from the joint estimation of cosmological parameters and $P(k)$ reconstructions in section \[current\]. We discuss our results and conclude in section \[discussion\]. Direct Inversion of $P_k$ by Singular Value Decomposition {#method} ========================================================= Direct primordial power spectrum reconstruction requires the inversion of the following relations $$\label{eq:intro-aps} C_\ell^{XY} =\int\limits_{0}^{\infty} \frac{\rm{d}k}{k} \Delta_{\ell}^{X}(k) \Delta_{\ell}^{Y}(k) P(k),$$ where $X$ and $Y$ represent $T$, $E$, or $B$-type anisotropies, $C_\ell^{XY}$ are the angular power spectra for the $XY$ combination and the $\Delta^{X}_{\ell}(k)$ are the photon perturbation transfer functions. The transfer functions are obtained by integrating the full Einstein-Boltzmann system of differential equations [@cmbfast; @camb]. These describe the evolution of perturbations in the photon distribution functions in the presence of gravity and other sources of stress-energy. The functions determine all of the structure in the anisotropy spectra which arise after the initial conditions are set. Most notably the $C_\ell^{XY}$ contain distinct peaks due to the acoustic oscillation of the tightly coupled photon-baryon fluid in gravitational potential wells at the time of last scattering. The aim of any inversion method is to distinguish such features from any structure in the initial perturbation spectrum. For a finite sampling of the wavenumber space $k$ Eq. (\[eq:intro-aps\]) can be recast as an operator acting on the primordial spectrum $P_k$ $$\label{eq:rld-start} C_\ell = \sum\limits_k F_{\ell k} P_k,$$ with operator $$\label{eq:operator} F^{X Y}_{\ell k} = \Delta \ln k\, \Delta_{\ell k}^{X} \Delta_{\ell k}^{Y},$$ where $\Delta \ln k$ are the logarithmic $k$ intervals for the discrete sampling chosen in the integration of the system of equations. A solution for $P_k$ cannot be obtained from a direct inversion of the $F^{X Y}_{\ell k}$ as it is numerically singular. This is due to the high level of degeneracy in the transfer functions relating the power at any wavenumber $k$ to angular multipoles $\ell$. We instead approximate the inversion by using the SVD method, first reducing the degeneracy of the system and then inverting using the remaining orthogonal modes. The transfer functions can be factorised as $$\label{eq:SVDdecom} F_{\ell k}= \sum_{\ell'k'}U^{\,}_{\ell \ell'} \Lambda^{\,}_{\ell' k'} V^\dagger_{k' k}\,,$$ where the matrices $\bf U$ and $\bf V$ are unitary and of dimensions $n_\ell$ and $n_k$ respectively, and $\bf \Lambda$ is a non-negative, diagonal matrix with diagonal elements $\lambda_k$. For this application, the dimensions of the matrices are that $n_\ell < n_k$ i.e. there are more equations than modes of interest. This results in some of the diagonal elements of $\bf \Lambda$ being singular (numerically zero) which prevent the inversion of the transfer matrix. The SVD method allows one to invert such a system by nulling the singular modes. This is achieved by creating an inverse $\bf \Lambda^{-1}$ where the diagonal elements are $1/\lambda_k$ except where the value of $\lambda_k$ is singular in which case it is replaced by $0$. In practice we rank order the factorised modes in descending order of $\lambda_k$ and all modes with condition number less than a threshold $\epsilon \mbox{max}(\lambda_k)$ are nulled, $\epsilon \approx 0.038$ for this work. Thus the method is a ‘$k$-to-$\ell$’ compression of the system where we keep the least degenerate modes connecting the 3d Fourier space to the 2d angular multipole space. This is not to be confused with a signal-to-noise compression of the data which aims to select with respect to orthogonal modes of the covariance of the observations [@Bond:1998qg]. It is instructive to look at the first few orthogonal modes given by the columns of the $\bf U$ matrix. We plot the first six in Fig. \[fig:basisvec\]. These $\ell$-space modes are the least degenerate (or best determined) in the mapping provided by the CMB physics. In other words, in the absence of sample and noise variance, these modes pick out the $\ell$-range where observing the CMB will have the highest impact upon the reconstructed $P(k)$. Not surprisingly, the first few modes are peaked around the angular scales where the acoustic signal from each polarisation combination is maximised. As these are the best constrained vectors in $\ell$ space in the absence of all errors, they are not necessarily the basis vectors of $P_k$ which are most accurately constrained, [@Hu:2003vp] show what they are for WMAP. We assume that this ordering of singular values is the optimal method for sorting the columns in $U$ and $V$. {width="16.25cm"} Once $\bf \Lambda^{-1}$ has been computed the primordial power spectrum can be reconstructed by inverting a set of observed $C_\ell$ using $$\begin{aligned} \label{eq:pkrecon} P_k & = & \sum_{\ell' k' \ell}V^{\,}_{kk'} \Lambda_{k' \ell'}^{-1} U_{\ell' \ell}^\dagger C^{\,}_\ell\,, \nonumber \\ & \approx & \sum_\ell F_{k\ell}^{-1}C^{\,}_\ell\,. \end{aligned}$$ Our choice of $\epsilon \approx 0.038$ is a conservative one with approximately 200 non-singular modes for a typical transfer matrix. Reconstructing the primordial power spectrum in this way means that we are not using any degenerate modes which carry no information in $k$ space but can increase the scatter in the reconstructed spectrum. However we are still susceptible to the scatter in the observed $C_\ell$ since we have not used any noise weighting in this scheme. In practice we start our inversion process with a guess input spectrum, parameterised by the usual form $A_s k^{n_s-1}$. From this and our fiducial cosmological model we obtain a $C_\ell^\textrm{model}$ spectrum. We use this to calculate the residual spectrum, $$\label{eq:residual} C_\ell^\textrm{res}=C_\ell^\textrm{obs}-C_\ell^\textrm{model},$$ so as to minimise the error induced in $k$-space by the cut-offs in $\ell$, both on large and small scales. To remove high frequency oscillations in the data we apply a low-pass filter to the resultant $P_k$. The following algorithm was used, $$\label{eq:lowpass} P_k^\textrm{low-pass}=\alpha P_k + (1-\alpha)P_{k-1},$$ where $\alpha$ was taken to be $0.05$. This method of smoothing leaves the first few points in the series strongly influenced by $P_1$. Therefore one should take any effects seen at low $k$ with a pinch of salt as these points are highly correlated. The covariance matrix was altered appropriately by a lower-triangular matrix representing this filter. We then proceed to bin the reconstructed power spectrum using the optimal binning method of [@Paykari:2009ac]. This binning method estimates a series of ranges in $k$ over which the signal-to-noise in the measured primordial power spectrum is constant. Many of the data points are highly correlated with their nearest neighbours and optimal binning gives a clear indication of the scales on which we have independent information. The centre of the $k$ bins chosen by the cosmological tool [CAMB]{} [@camb] change when the input cosmological parameters are altered. This is a problem when it is run over many realisations as in the case of a Markov Chain. We choose the optimal binning method to find the standardised output of centres and sizes of $k$ bins for each call of the [CAMB]{} routine. To find the optimal binning of the reconstructed power spectrum we investigate how the uncertainty in the $C_{\ell}$ transfers into uncertainty in the primordial power spectrum. For this purpose, we need to define the primordial power spectrum as a series of top-hat bins: $$P(k)=\sum_{B}w_{B}(k)Q_{B\:,}$$ where $Q_{B}$ is the power in each bin $B$ and $w_{B}=1$ if $k\in B$ and $0$ otherwise. To obtain the errors for these bins we define the Fisher matrix for the $C_{\ell}$ by $$M_{\ell\ell^{\prime}}=(\delta C_{\ell\ell^{\prime}})^{-1}\:,$$ where $\delta C_{\ell\ell^{\prime}}$ is the diagonal matrix of the squares of the variances in each measurement of $C_{\ell}$. To transfer the given errors from the $C_{\ell}$ to other parameters we use the Jacobian, $$M_{\alpha\beta}=\sum_{\ell\ell^{\prime}}M_{\ell\ell^{\prime}}\frac{\partial C_{\ell}}{\partial\theta_{\alpha}}\frac{\partial C_{\ell^{\prime}}}{\partial\theta_{\beta}}\;,$$ where, $\theta_{\alpha}$ and $\theta_{\beta}$ represent the bins of the primordial power spectrum. The derivative of the $C_{\ell}$ with respect to the primordial power spectrum is the average radiation transfer function in each bin: $$\frac{\partial C_{\ell}}{\partial P(k)}=\int_{k_{min}^{B}}^{k_{max}^{B}}\frac{dk}{k}\Delta_{\ell}^{X}(k)\Delta_{\ell}^{Y}(k)\;.\label{eq:Cl_pPS_derivative}$$ To calculate the signal-to-noise ratio in each bin we take the inverse square root of the diagonal elements of $M_{\alpha\beta}$ to be the noise and the amplitude of the primordial power spectrum to be the signal. We then arrange the bins to have the same signal-to-noise over our $k$ range. Our algorithm will result in more bins where the signal-to-noise ratio is greater, sampling more finely where the signal is strongest. We construct a signal vector, $S$, which contains the amplitude of the primordial power spectrum for all the bins and weight our Fisher matrix by this vector $$\left(\frac{S}{N}\right)^2_{\alpha\beta}=S_{\alpha}M_{\alpha\beta}S_{\beta}\,,\label{eq:(S/N)^2 FM}$$ where there is no Einstein summation. The square root of the diagonal elements of this matrix are the $S/N$ of the bins. We start our algorithm with the maximum number of bins possible in our $k$ range. This is set by the usual properties of the Fourier transform. These imply that the scale of the survey not only determines $k_{min}$, but also gives a lower bound upon the resolution, $\Delta k_{min}$: narrower bins would become highly correlated. Therefore, we set up a series of bins with the properties $$k_{min}=\frac{\ell_{min}}{d_{A}}=\frac{2}{d_{A}}\:\:\:\textnormal{and}\:\:\: (\Delta k)_{min}=\frac{\Delta\ell}{d_{A}}=\frac{1}{d_{A}}\:,$$ where $d_{A}=14.12\textnormal{Gpc}$ (value given by WMAP$5$) is the angular diameter distance to the surface of last scaterring. We set $k_{max}=0.08\textnormal{Mpc}{}^{-1}$ as the reconstruction process is limited past this regime due to the cut-off in $\ell$ at $1000$. A Fisher matrix is then constructed for this set of bins and weighted by the signal vector. We choose a signal-to-noise value, that cannot be less than the maximum value seen in any bin, as our target ratio. The binwidths are increased in order to reach the target value at each bin. To obtain this ‘optimal’ binning we iterate until the bin with the smallest signal-to-noise ratio is within $5\%$ of the target ratio. Reconstruction with cosmological parameter fitting -------------------------------------------------- The reconstruction method described above is fast and can be carried out at each random sample of a MCMC exploration of the cosmological parameter space. Inserting the reconstruction as part of an MCMC exploration we can account for the variance induced in the primordial power spectrum due to the dependence of the radiation transfer function on the cosmology. We have modified the [cosmomc]{} [@cosmomc] package by introducing the reconstruction at each chain evaluation using the inverse of the transfer function computed for each combination of parameters acting on the ’observed’ CMB angular power spectrum. The reconstructed spectrum is then itself used to compute the final $C_\ell$ which are used to calculate the likelihood at the chain step. In principle the chains would probe the reduced set of parameters; the physical densities of baryons $\Omega_bh^2$, and of cold dark matter $\Omega_ch^2$, the angular diameter distance parameter $\theta$, and optical depth parameter $\tau$. The primordial power spectrum parameters $n_s$ and $A_s$ become irrelevant and need not be probed since the power spectrum is being reconstructed directly. However, in practice, we [*do*]{} include power law spectral parameters which determined the shape the template model (\[eq:residual\]) and we marginalise over the spectral parameters in order to account for any sensitivity of the reconstruction to the assumed $C_\ell^{\textrm{model}}$. The immediate advantage of combining the reconstruction with an MCMC method is that we can then calculate the variance in the resulting spectrum due to the random nature of the transfer function. We do this by including the binned amplitudes for the reconstructed spectrum as ‘derived’ parameters when analysing the chains. The covariance of the chains is then mapped into a covariance for the binned power spectrum. We also need to account for the variance due to the errors in the observed CMB data. This is not accounted for in the MCMC chains since we always use the same observed $C_\ell$ data to reconstruct the spectrum. In principle this contribution to the variance and that from the transfer function are correlated, however this is difficult to quantify without including MCMC steps over realisations of the observations. We therefore make a conservative estimate of the final error in the reconstructed spectrum by adding the covariance matrix obtained from the MCMC chain and that obtained by rotating the error matrix of the observed $C_\ell$ as $$\delta P_{BB'} = \sum_{\ell\ell'}F^{-1}_{B\ell}\delta C_{\ell\ell'}F^{-1}_{B'\ell'}\,$$ where $F^{-1}_{B\ell}$ is the bin-averaged contribution from $ F^{-1}_{k\ell}$. Application of the SVD Inversion {#results} ================================ Tests on simulated CMB data --------------------------- {width="14cm"} {width="14cm"} {width="14cm"} {width="14cm"} {width="14cm"} We start by testing the reconstruction algortihm on a set of input spectra with known features as in [@Nicholson:2009pi]. Fig. \[fig:testingSVD\] shows that the method accurately reconstructs test features on the input spectra. We used a fiducial cosmological model of $\Omega_bh^2=0.0226$, $\Omega_ch^2=0.108$, $\theta=1.041$ and $\tau=0.076$. The starting spectrum used to obtain $C_\ell$ model is shown by the black, long-dashed line. We assume there are no errors on the input $C_\ell^\textrm{obs}$ (between $\ell=2$ and $\ell=1000$) and observe that the reconstructed $P_k$ generally picks out the input features. We limit our method to fit between $k=0.0013$ and $k=0.08$ as this is the range with the highest signal-to-noise in the WMAP data. The features chosen are the same as in [@Nicholson:2009pi] to allow comparisons with this method. They are a standard power law with $n_s=0.96$ but with an amplitude $90$% of that of the best-fit WMAP model, a running $n_s$ model with $dn_s/d\ln k = -0.037$, a power law with a sharp, compensated feature at $k=0.02$ Mpc$^{-1}$ [@Starobinsky:1992ts; @Lesgourgues:1999uc; @Adams:2001vc; @Feng:2003zua; @Mathews:2004vu; @Joy:2007na; @Hunt:2007dn; @Jain:2008dw; @Lerner:2008ad] and a power law with superimposed sinusoidal oscillations [@Wang:2002hf; @Martin:2003sg; @Martin:2004yi; @Pahud:2008ae; @Romano:2008rr]. All four features are clearly recovered to varying degrees when using $TT$, $TE$ or $EE$, however we find a phase offset between the reconstructed ad input spectrum for the case where the input spectrum includes oscillations as in the lower panels of Fig. \[fig:testingSVD\]. The offset is stable with respect to the presence of the smoothing kernel, the number of singular values cut from the inversion, and with repsect to the number of $k$ bins and range. The reconstructed $P_k$ given by both $TE$ and $EE$ contains ‘glitches’ not present in the $TT$ reconstruction. These regions correspond to regions where there is little information int he $TE$ and $EE$ spectra and the reconstruction is still degenerate. The cosmological parameters obtained from a traditional Markov Chain Monte Carlo search are used to give us our fiducial operator, $F_{\ell k}$. The parameters however have errors present upon them, which are not usually incorporated into errors on the final reconstructed $P_k$. If inaccurate parameters have been used to calculate the fiducial operator very specific signatures would be expected to show up in the reconstructed from of $P_k$. We show in Fig. \[fig:incorrectparamest\] how these signatures appear in the reconstructed $P_k$ for $TT$ anisotropies for four cosmological parameters, $\Omega_bh^2$, $\Omega_ch^2$, $h$ and $\tau$. If any features with this form are observered one should attribute this to an incorrect estimation of the parameters and not some fundamental physics. A check for this would be if the corresponding features also show up in the reconstructed $TE$ and $EE$ spectra. We have also tested the inversion on simulated CMB data with similar experimental properties as the recently launched Planck satelite mission [@planck]. We assume a total of 12 detectors with NETs of 64$\mu K/\sqrt{s}$ observing 80% of the sky over 12 months with a resolution of 7 arcminutes FWHM. We calculate errors around our fiducial CMB best-fit models in both total intensity and polarisation spectra for this experimental setup and use these together with $C_\ell$ samples on the fiducial model to test the inversion method’s properties. We consider multipoles of $\ell < 1000$ for both total intensity spectra and polarisation. We have not taken into account any residual error from foreground subtraction in our forecasts. Thus our forecast are on the optimistic side of the accuracy achievable in the case of polarisation where foreground removal will certainly have a significant impact on errors at $\ell < 1000$. In the case of total intensity spectra we are significantly underestimating the accuracy achievable by Planck as we expect to obtain well measured $C_\ell$’s well above $\ell$ of $1000$. We do not consider these modes as it significantly increases the time required to perform the SVD. Consideration of total-intensity modes past $\ell$ of $1000$ will increase the accuracy obtainable for Planck on $P_k$, it also allows one to probe $k$ past $0.08$. It would be desirable to perform this process with a greater $\ell$ range when the Planck data becomes available. To compare the accuracy obtainable with each of the anisotropy types we show the degree to which they each reconstruct a simple power law $P_k$ in Fig. \[fig:testingSVD-errors\]. We use the same fiducial cosmological model as in Fig. \[fig:testingSVD\] with a standard power law input $P_k$ equal to the spectra being reconstructed. All the $C_\ell^\textrm{obs}$ were placed on the fiducial model. Over the whole range of $k$ total-intensity modes best constrain $P_k$. But it is also true that there are regions in $k$-space, for example between $k=0.022$ and $k=0.035$, where considering only the $TE$ measurements can give us a more accurate estimation of $P_k$. $EE$ measurements approach the accuracy of $TT$ at a number of points, there are however regions where the errors become so large that any reconstructed $P_k$ is meaningless (these correspond to the troughs of the $EE$ spectrum). The regions of $\ell$-space corresponding to the most accurately measured $k$ regions are the peaks of the spectra. Both $TE$ and $EE$ spectra are not significantly affected by changes to $P_k$ at very low $k$, so errors in this region are artificially small. We also tested the combined $P(k)$ and parameter estimation MCMC search as described in the previous section. The optimal binning method found $128$ bins for the Planck experiment, where our target signal-to-noise value in each bin is $10$. We choose this value so as to have approximately the same number of bins across the range $0.01<k<0.03$ where WMAP best probes $P_k$. In Fig. \[fig:cosmomc-params\] we show the results of a cosmological parameter estimation from this process for a simulated Planck experiment. We find that the input parameters were accurately recovered after this process. We show the final reconstructed $P_k$ in Fig. \[fig:cosmomc-pk\]. The red line is the reconstructed $P_k$ at its best-fit point. The error bars we show are obtained from combining the error from the marginalised distributions with the reconstruction errors given the observed CMB data. The errors are centred around the mean of the marginalised distribution for each bin. It is important to note that the errors are highly correlated. This explains the reduced scatter in the mean values compared to the size of the plotted errors. As was seen in Fig. \[fig:incorrectparamest\] changing a single parameter by a small amount (in the manner a MCMC search does) creates a very specific signature on $P_k$ for each parameter, where changing it slightly has a correlated effect upon the whole range of $k$. This explains the very high correlations observed accross the whole $k$ range. It is at odds with the errors on $P(k)$ associated with those of $C_\ell$ at any best-fit point which are not correlated across large ranges of $k$. Constraints from current CMB observations {#current} ========================================= {width="15cm"} {width="15cm"} {width="15cm"} We used two sets of currently available data to estimate the cosmological parameters in conjunction with a free unparameterised $P_k$. The first is the WMAP 5-year data alone [@wmap5], in the second we combine this with that of SNIa, HST and BBN [@Perlmutter:1998np; @Riess:2004nr; @Astier:2005qq; @Kowalski:2008ez; @Freedman:2000cf]. The second set combination chosen because the non-CMB sets do not depend upon the form of $P_k$ and can therefore give us independent and tighter constraints on most of the cosmological parameters. We run this data through the MCMC tool [COSMOMC]{} in the same manner as we did for the test Planck data. In Fig. \[fig:wmap-pk\] we show the current constraints on $\Omega_bh^2$, $\Omega_ch^2$, $\theta$ and $\tau$. The solid red vertical lines indicates the WMAP only best fit values when $P(k)$ is parameterised in the usual fashion, the black solid curves represent the marginalised probability distribution of the WMAP only data. There is no disagreement between this and the WMAP bestfit model. The blue dotted line shows the marginalised probability distribution of WMAP including the other datasets. Here we observe some tension at around the $1\sigma$ level in $\Omega_ch^2$. The inclusion of this data does not move the position of the likelihood peak significantly for each parameter. The optimal binning method gave us 86 bins in our k range for WMAP. We used the minimum $\Delta k$ as described previously, this gave a target signal-to-noise ratio of $6$. The binned reconstructed $P_k$ are shown in Figs. \[fig:wmap-cosmomc\] and \[fig:wmap-cosmomc-sn\] for each of the two data sets. As in the Planck case, the results of the MCMC are highly correlated across the whole $k$ range, in contrast to a single reconstruction which is not. We have similarly added the errors from the MCMC to a single reconstruction in each bin in quadrature. The red lines are the reconstructed $P_k$ at their best-fit points. They error bars are centred around the best-fit marginalised value of each bin. For the WMAP only run we find that the limiting errors on $P_k$ in the range $0.0075 < k < 0.05$ come from uncertainty in the cosmological parameters, whereas this limiting range is around $0.0075 < k < 0.04$ when we include the other datasets. No significant deviation from the standard power-law case is observed in either run. Discussion ========== The SVD based reconstruction method we have outlined provides is fast enough to be incorporated into full MCMC parameter fitting runs. We have tested the method and shown that it recovers the overall features of input spectra. We have applied the method to forecasted Planck data and current WMAP 5-year results. These results have allowed us to consistently combine the reconstruction with a full exploration of the parameter likelihoods for the first time. We have seen that the limiting factor in constraining the primordial spectrum over a large range of wavenumbers $k$ comes from the uncertainty in cosmological parameters. Any claims of a detection of a feature in $P_k$ must necessarily confront the degeneracy with the cosmological parameter space . This effect will be less important when the Planck data is released, however it must still be considered as the unprecedented accuracy offered by future data may lead to premature claims of a detection of an interesting feature. We observed some tension between the WMAP only best-fit model, when $P(k)$ is parameterised with the normal amplitude and tilt, and our method in the marginalised probability distribution of $\Omega_ch^2$. They disagree at around the $1\sigma$ level which is not overly significant, however it is an indication that there may be some departure from a standard power law in $P(k)$. Planck will certainly determine if this is the case. There are other currently available CMB datasets, which could expand the range of $k$ probed. However expanding the method to include multiple data sets is a non-trivial task due to the binning of the $C_\ell$ data for sub-orbital experiments and the increased running time required for the expanded $k$-range. We leave this analysis for future work. In the future, as CMB polarisation data becomes increasingly accurate, it will be desirable to perform a joint inversion of total intensity data along with polarisation data. It is not clear how to extend the SVD based method to include all polarisation modes simultaneously since a HOSVD (Higher-Order SVD) step would probably be required. On the other hand this would give the best estimate of $P_k$ given any dataset and would help to reduce the correlations found in the reconstructed $P_k$ by increasing the degrees of freedom that can be effectively constrained. This work was supported by a STFC studentship. We acknowledge the use the Imperial College high performance computing service[^1]. [^1]: http://www.imperial.ac.uk/ict

ArXiv

LPENSL-TH-11-18 **[Complete spectrum of quantum integrable lattice models associated to $\mathcal{U}_{q} (\widehat{gl_{n}})$ by separation of variables]{}** [**J. M. Maillet** **and**    **G. Niccoli**]{} - **Abstract.** In this paper we apply our new separation of variables approach to completely characterize the transfer matrix spectrum for quantum integrable lattice models associated to fundamental evaluation representations of $\mathcal{U}_{q} (\widehat{gl_{n}})$ with general quasi-periodic boundary conditions. We consider here the case of generic deformations associated to a parameter $q$ which is not a root of unity. The Separation of Variables (SoV) basis for the transfer matrix spectral problem is generated by using the action of the transfer matrix itself on a generic co-vector of the Hilbert space, following the general procedure described in our paper [@MaiN18]. Such a SoV construction allows to prove that for general values of the parameters defining the model the transfer matrix is diagonalizable and with simple spectrum for any twist matrix which is also diagonalizable with simple spectrum. Then, using together the knowledge of such a SoV basis and of the fusion relations satisfied by the hierarchy of transfer matrices, we derive a complete characterization of the transfer matrix eigenvalues and eigenvectors as solutions of a system of polynomial equations of order $n+1$. Moreover, we show that such a SoV discrete spectrum characterization is equivalently reformulated in terms of a finite difference functional equation, the quantum spectral curve equation, under a proper choice of the set of its solutions. A construction of the associated $Q$-operator induced by our SoV approach is also presented. Introduction ============ In this paper we make use of our new approach to generate the SoV basis [MaiN18,MaiN18a,MaiN18b]{} to characterize the complete spectrum of quantum integrable lattice models with general quasi-periodic boundary conditions associated to the fundamental evaluation representations of higher rank trigonometric Yang-Baxter algebras. More in detail, in the framework of the quantum inverse scattering method [FadS78,FadST79,FadT79,Skl79,Skl79a,FadT81,Skl82,Fad82,Fad96]{}, these models are constructed by using the rank $n$ principal gradation $R$-matrix [BabdeVV81,Jim86]{} solution of the Yang-Baxter equation associated to the quantum group $\mathcal{U}_{q}(\widehat{gl_{n+1}})$ [KulR83a,KulR82,Jim85,Dri87,ChaP94]{}. This $R$-matrix admits a nontrivial set of scalar solutions to the Yang-Baxter equation. Such symmetries of the $R$-matrix, here called twist $K$-matrices, allow for the definition of integrable quasi-periodic boundary conditions for the corresponding integrable quantum models. Our SoV basis is generated for these general twists, under the assumption that the twist matrix $K$ has simple spectrum. As previously shown in our work [MaiN18]{}, the existence of such a SoV basis implies the simplicity of the transfer matrix spectrum. Moreover if the twist matrix $K$ is diagonalizable with simple spectrum it implies that the transfer matrix is also diagonalizable with simple spectrum for almost any choice of the parameters of the representations. The transfer matrix spectrum of these integrable quantum models has been analyzed also by other approaches, in particular, for diagonal twists, in the framework of the fusion relations [@KulRS81; @KulR83] and analytic Bethe ansatz [@Res83; @Res83a; @KirR86; @Res87; @KunNS94], the nested Bethe ansatz [@KulR81; @KulR83; @BelR08; @PakRS18], with also first interesting results toward the computation of correlations functions [BelPRS12,BelPRS12a,BelPRS13,BelPRS13a,PakRS14,PakRS14a,PakRS15,PakRS15a,PakRS15b,LiaS18]{}. Let us also note that for anti-periodic boundary conditions an eigenvalue analysis by a functional approach has been developed in [@HaoCLYSW16]. The quantum version of the separation of variables method has been pioneered by Sklyanin in a series of works [@Skl85; @Skl90; @Skl92; @Skl92a; @Skl95; @Skl96] in particular to tackle models for which the standard algebraic Bethe ansatz cannot be applied. Since the Sklyanin’s original papers this approach has been successfully implemented and partially generalized to several classes of integrable quantum models mainly associated to different representations of quantum algebras related of rank one type, e.g. for the 6-vertex and 8-vertex Yang-Baxter algebras and reflection algebras as well as to their dynamical deformations [Skl85,Skl90,Skl92,Skl92a,Skl95,Skl96,BabBS96,Smi98a,Smi01,DerKM01,DerKM03,DerKM03b,BytT06,vonGIPS06,FraSW08,MukTV09,MukTV09a,MukTV09c,AmiFOW10,NicT10,Nic10a,Nic11,FraGSW11,GroMN12,GroN12,Nic12,Nic13,Nic13a,Nic13b,GroMN14,FalN14,FalKN14,KitMN14,NicT15,LevNT15,NicT16,KitMNT16,MarS16,KitMNT17,MaiNP17]{}. The interest in developing the separation of variables method is mainly due to some important built-in features as the ability to provide a direct proof of the completeness of the spectrum characterization as well as some first elements towards the dynamics like scalar products and form factors. Important analysis toward the SoV description of higher rank cases have been presented in [@Skl96; @Smi01; @MarS16; @GroLMS17; @DerV18]. Here, we solve the long-standing problem to systematically introduce a quantum separation of variable approach capable to completely characterize the transfer matrix spectrum associated to the higher rank representations of the trigonometric Yang-Baxter algebra. While our approach bypass the construction of the so-called Sklyanin’s commuting $B$-operator family [Skl85,Skl90,Skl92,Skl92a,Skl95,Skl96]{}, the results on the rank one representations as well as some evidence from the short lattices for the higher rank representations [@MaiN18], plus some recent analysis developed in [@RyaV18] for the rational higher rank situation, confirm that our SoV basis construction can nevertheless reproduce the Sklyanin’s SoV basis (i.e. the $B$-operator eigenbasis). This is the case under some special choice of the generating covector, i.e. the covector from which our SoV basis is constructed by the action of a chosen set of commuting conserved charges. This type of connection for the trigonometric representations, in particular, in relation to the SoV results obtained in [@Smi01] would be interesting to study further. However, as anticipated above, our strategy following [@MaiN18], is instead to make a direct use of our SoV basis construction to obtain the complete characterization of the spectrum (eigenvalues and eigenvectors) using in particular the hierarchy of fusion relations for the transfer matrices. Here we consider the transfer matrices associated to general twist matrices $K$ diagonalizable and with simple spectrum for the trigonometric $gl_n$ ($n\geq 2$ ) Yang-Baxter algebras in the fundamental evaluation representations. We will first obtain a complete characterization of the spectrum in terms of the set of solutions to a given system of $\mathsf{N}$ polynomial equations of degree $n+1$ in $\mathsf{N}$ unknowns, where $\mathsf{N}$ is the number of lattice sites. Second, we introduce the so-called quantum spectral curve functional equation and we provide the exact characterization of the set of its solutions which generates the complete transfer matrix spectrum associating to any eigenvalue solution exactly one nonzero eigenvector. These results allow also to point out, as already done for other quantum integrable models [@MaiN18; @MaiN18a; @MaiN18b], how the SoV basis in our construction can be equivalently obtained by the action of the commuting family of $Q$-operator. This connection is important as it allows to bring in our SoV approach results of the Baxter’s Q-operator method [@Bax73-tot; @Bax73; @Bax76; @Bax77; @Bax78; @PasG92; @BatBOY95; @YunB95; @BazLZ97; @AntF97; @BazLZ99; @Der99; @Pro00; @Kor06; @Der08; @BazLMS10; @Man14] ; of special interests are then the results presented in [@MT15] for the higher rank case. In our approach we show that this $Q$-operator satisfies with the transfer matrices the quantum spectral curve equation and that it can be reconstructed, making use of our Sov basis, in terms of the monodromy matrix entries. In order to make easier the reading of our results we have decided to present them first in the rank 2 case, namely for the fundamental representations generated by the principal gradation $R$-matrix associated to $\mathcal{U}_{q}(\widehat{gl_{3}})$. Then these results are extended to the generic higher rank cases associated to the fundamental evaluation representations of $\mathcal{U}_{q}(\widehat{gl_{n}})$. This article is organized as follows. In section 2, we introduce the fundamental evaluation representation of the rank two trigonometric Yang-Baxter algebra, the corresponding quantum spectral invariants of the model and then we list some general analytic properties they satisfy. Then in subsection 2.2 we construct our SoV covector basis and we state the first consequences on the transfer matrix spectrum. The section 3 is then dedicated to the presentation of our results on the transfer matrix spectrum characterization. In subsection 3.1, we derive the SoV discrete characterization of the transfer matrix spectrum in terms of solutions to a system of polynomial equations of degree 3. In subsection 3.2, we give an equivalent characterization in terms of the solutions to a functional equation of third order type, the so-called quantum spectral curve equation. In subsection 3.3 we also show that our SoV characterization of the transfer matrix eigenvectors allows for their rewriting in an algebraic Bethe ansatz form. In section 4, we define the framework of the general higher rank $n$ case by introducing the corresponding fundamental evaluation representations of the $\mathcal{U}_{q}(\widehat{gl_{n}})$ Yang-Baxter algebra, the associated quantum spectral invariants and some of their general analytic properties. In subsection 4.2 we construct our SoV covector basis for this general rank $n$ case. The section 5 presents the complete transfer matrix spectrum characterization. We first derive the SoV discrete spectrum characterization in subsection 5.1 while in subsection 5.2 we show its equivalence to the quantum spectral curve, a functional equation of difference type of order $n+1$ for the $Q$-operator that we construct using the knowledge of our SoV basis. Some important technical proofs are gathered in two appendices. In appendix A, for the rank two case, we provide a proof of the complete characterization of the spectrum which is based on the explicit calculation of the transfer matrix action on our SoV covector basis. While this proof can be generalized to the general rank $n$ along a similar path described in [@MaiN18a] for the rational case, we provide in appendix B a proof of the SoV discrete characterization of the transfer matrix spectrum which bypass the computation of the action of the transfer matrix on the SoV covector basis. Transfer matrices for fundamental evaluation representations of $\mathcal{U}_{q}(\widehat{gl_{3}})$ =================================================================================================== We consider here the trigonometric Yang-Baxter algebra generated by the principal gradation $R$-matrix [@BabdeVV81; @Jim86] associated to the quantum group $\mathcal{U}_{q}(\widehat{gl_{3}})$[^1]:$$R_{a,b}^{\left( P\right) }(\lambda )=\left( \begin{array}{ccc} a_{1}(\lambda ) & \lambda ^{1/3}b_{1} & \lambda ^{-1/3}b_{2} \\ \lambda ^{-1/3}c_{1} & a_{2}(\lambda ) & \lambda ^{1/3}b_{3} \\ \lambda ^{1/3}c_{2} & \lambda ^{-1/3}c_{3} & a_{3}(\lambda )\end{array}\right) \in \text{End}(V_{a}\otimes V_{b}),$$where $V_{a}\cong V_{b}\cong \mathbb{C}^{3}$ and we have defined:$$\begin{aligned} & a_{j}(\lambda )\left. =\right. \left( \begin{array}{ccc} \lambda q^{\delta _{j,1}}-1/(\lambda q^{\delta _{j,1}}) & 0 & 0 \\ 0 & \lambda q^{\delta _{j,2}}-1/(\lambda q^{\delta _{j,2}}) & 0 \\ 0 & 0 & \lambda q^{\delta _{j,3}}-1/(\lambda q^{\delta _{j,3}})\end{array}\right) ,\text{ \ \ }\forall j\in \{1,2,3\}, \notag \\ & b_{1}\left. =\right. \left( \begin{array}{ccc} 0 & 0 & 0 \\ q-1/q & 0 & 0 \\ 0 & 0 & 0\end{array}\right) ,\text{ \ }b_{2}\left. =\right. \left( \begin{array}{ccc} 0 & 0 & 0 \\ 0 & 0 & 0 \\ q-1/q & 0 & 0\end{array}\right) ,\text{ \ }b_{3}\left. =\right. \left( \begin{array}{ccc} 0 & 0 & 0 \\ 0 & 0 & 0 \\ 0 & q-1/q & 0\end{array}\right) , \notag \\ & c_{1}\left. =\right. \left( \begin{array}{ccc} 0 & q-1/q & 0 \\ 0 & 0 & 0 \\ 0 & 0 & 0\end{array}\right) ,\text{ \ }c_{2}\left. =\right. \left( \begin{array}{ccc} 0 & 0 & q-1/q \\ 0 & 0 & 0 \\ 0 & 0 & 0\end{array}\right) ,\text{ \ }c_{3}\left. =\right. \left( \begin{array}{ccc} 0 & 0 & 0 \\ 0 & 0 & q-1/q \\ 0 & 0 & 0\end{array}\right) .\end{aligned}$$Note that here we have chosen to present the $R$-matrix in a Laurent polynomial form but clearly it can be rewritten as well in a trigonometric form. This $R$-matrix is a solution of the Yang-Baxter equation:$$R_{12}(\lambda /\mu )R_{13}(\lambda )R_{23}(\mu )=R_{23}(\mu )R_{13}(\lambda )R_{12}(\lambda /\mu )\in \text{End}(V_{1}\otimes V_{2}\otimes V_{3}),$$and it is related by a similarity transformation to the so-called homogeneous gradation $R$-matrix for $U_{q}(\widehat{gl_{3}})$:$$R_{a,b}^{(H)}(\lambda )=\left( \begin{array}{ccc} a_{1}(\lambda ) & \lambda b_{1} & \lambda b_{2} \\ c_{1}/\lambda & a_{2}(\lambda ) & \lambda b_{3} \\ c_{2}/\lambda & c_{3}/\lambda & a_{3}(\lambda )\end{array}\right) \in \text{End}(V_{a}\otimes V_{b}),$$which is also a solution of the Yang-Baxter equation. More in detail, it holds:$$R_{a,b}^{\left( P\right) }(\lambda )=S_{a}(\lambda )R_{a,b}^{(H)}(\lambda )S_{a}^{-1}(\lambda ),$$where$$S(\lambda )=\left( \begin{array}{ccc} 1 & 0 & 0 \\ 0 & \lambda ^{2/3} & 0 \\ 0 & 0 & \lambda ^{4/3}\end{array}\right) ,$$such connection has been first remarked in [@Jim86]. Let us comment that this two solutions of the Yang-Baxter equation generate the same quantum integrable models with diagonal quasi-periodic boundary conditions. Indeed, any diagonal $3\times 3$ matrix $K\in $End$(V)$ is a symmetry for both these $R$-matrices:$$R_{12}^{\left( P/H\right) }(\lambda )K_{1}K_{2}=K_{2}K_{1}R_{12}^{\left( P/H\right) }(\lambda )\in \text{End}(V_{1}\otimes V_{2}),$$and defined the monodromy matrices:$$\begin{aligned} M_{a}^{(P/H|K)}(\lambda |\{\xi \}) &=&\left( \begin{array}{ccc} A_{1}^{(K)}(\lambda ) & B_{1}^{(K)}(\lambda ) & B_{2}^{(K)}(\lambda ) \\ C_{1}^{(K)}(\lambda ) & A_{2}^{(K)}(\lambda ) & B_{3}^{(K)}(\lambda ) \\ C_{2}^{(K)}(\lambda ) & C_{3}^{(K)}(\lambda ) & A_{3}^{(K)}(\lambda )\end{array}\right) _{a} \\ &\equiv &K_{a}R_{a,\mathsf{N}}^{\left( P/H\right) }(\lambda /\xi _{\mathsf{N}})\cdots R_{a,1}^{\left( P/H\right) }(\lambda /\xi _{1})\in \text{End}(V_{a}\otimes \mathcal{H}),\end{aligned}$$where $\mathcal{H}=\bigotimes_{n=1}^{\mathsf{N}}V_{n}$, and the transfer matrices$$T_{1}^{\left( P/H|K\right) }(\lambda |\{\xi \})\equiv \text{tr}_{a}M_{a}^{\left( P/H|K\right) }(\lambda |\{\xi \})\in \text{End}\mathcal{H}$$then the following identity holds for the homogeneous chains:$$\left. T_{1}^{\left( P|K\right) }(\lambda |\{\xi \})\right\vert _{\xi _{i}=0}=\left. T_{1}^{\left( H|K\right) }(\lambda |\{\xi \})\right\vert _{\xi _{i}=0},$$which implies our statement on the equality of the Hamiltonians with diagonal quasi-periodic boundary conditions generated by these two $R$-matrices. However, one has to remark that for the inhomogeneous models the above identity does not hold. One motivation to consider the fundamental models associated with the principal gradation $R$-matrix is the set of symmetry $K$-matrices enjoyed by this Yang-Baxter solution together with the simpler properties under co-product action, and hence under fusion. Indeed, the set of solutions of the scalar Yang-Baxter equation for the principal gradation $R$-matrix reads:$\allowbreak $$$K^{\left( a\right) }=\delta _{a,1}\left( \begin{array}{ccc} \alpha & 0 & 0 \\ 0 & \beta & 0 \\ 0 & 0 & \gamma\end{array}\right) +\delta _{a,2}\left( \begin{array}{ccc} 0 & 0 & \alpha \\ \beta & 0 & 0 \\ 0 & \gamma & 0\end{array}\right) +\delta _{a,3}\left( \begin{array}{ccc} 0 & \beta & 0 \\ 0 & 0 & \gamma \\ \alpha & 0 & 0\end{array}\right) ,$$for any complex value of $\alpha ,$ $\beta $ and $\gamma $, while for the homogeneous gradation $R$-matrix the matrices $K^{\left( 2\right) }$ and $K^{\left( 3\right) }$ are symmetries if and only if $\alpha =0$. Note that the matrices $K^{\left( 2\right) }$ and $K^{\left( 3\right) }$ are diagonalizable and their eigenvalues reads:$$\mathsf{k}_{0}=\sqrt[3]{\alpha \beta \gamma },\mathsf{k}_{1}=-(-1)^{1/3}\mathsf{k}_{0},\mathsf{k}_{2}=(-1)^{2/3}\mathsf{k}_{0}$$so that for $\alpha \beta \gamma \neq 0$ these matrices have simple spectrum and it holds:$$\text{det}K^{\left( a\right) }=\alpha \beta \gamma \text{ \ }\forall a\in \{1,2,3\}.$$It is also interesting to remark that the principal gradation $R$-matrix directly generates, the $q$-deformed antisymmetric projector $P_{a,b}^{-}\in $End$ (V_{a}\otimes V_{b})$ which is used in the fusion representations for the $\mathcal{U}_{q}(\widehat{gl_{3}})$ case. In particular, it holds:$$\begin{aligned} & \frac{R_{12}^{\left( P\right) }(1/q)}{2(1/q-q)}=P_{a,b}^{-} \notag \\ & =\left( \begin{array}{ccccccccc} 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\ 0 & 1/2 & 0 & -q^{1/3}/2 & 0 & 0 & 0 & 0 & 0 \\ 0 & 0 & 1/2 & 0 & 0 & 0 & -1/(2q^{1/3}) & 0 & 0 \\ 0 & -1/(2q^{1/3}) & 0 & 1/2 & 0 & 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 & 0 & 1/2 & 0 & -q^{1/3}/2 & 0 \\ 0 & 0 & -q^{1/3}/2 & 0 & 0 & 0 & 1/2 & 0 & 0 \\ 0 & 0 & 0 & 0 & 0 & -1/(2q^{1/3}) & 0 & 1/2 & 0 \\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\end{array}\right) .\end{aligned}$$ Fundamental properties of the hierarchy of transfer matrices ------------------------------------------------------------ Here, we collect some relevant known properties [@KulRS81; @KulR83; @KunNS94] of the fused transfer matrices for the representations under consideration. The transfer matrices:$$T_{1}^{\left( K\right) }(\lambda )\equiv \text{tr}_{a}M_{a}^{(K)}(\lambda ),\text{ \ \ \ }T_{2}^{\left( K\right) }(\lambda )\equiv \text{tr}_{a,b}P_{a,b}^{-}M_{b}^{(K)}(\lambda )M_{a}^{(K)}(\lambda /q),$$satisfy the following commutation relations:$$\left[ T_{1}^{\left( K\right) }(\lambda ),T_{1}^{\left( K\right) }(\mu )\right] =\left[ T_{1}^{\left( K\right) }(\lambda ),T_{2}^{\left( K\right) }(\mu )\right] =\left[ T_{2}^{\left( K\right) }(\lambda ),T_{2}^{\left( K\right) }(\mu )\right] =0.$$The quantum determinant:$$q\text{-det}M^{(K)}(\lambda )\equiv \text{tr}_{abc}P_{abc}^{-}M_{c}^{(K)}(\lambda )M_{b}^{(K)}(\lambda /q)M_{a}^{(K)}(\lambda /q^{2}) \label{Bound-q-detU_1-tri}$$is a central element of the algebra, i.e.$$\lbrack q\text{-det}M^{(K)}(\lambda ),M_{a}^{(K)}(\mu )]=0.$$ Furthermore, let us define the operators $\mathsf{N}_{i}\in \text{End}(\mathcal{H})$ by the following action:$$\mathsf{N}_{i}\otimes _{n=1}^{\mathsf{N}}|n,a_{n}\rangle =\otimes _{n=1}^{\mathsf{N}}|n,a_{n}\rangle \sum_{n=1}^{\mathsf{N}}\delta _{i,a_{n}} \text{, } \forall \ i\in \{1,2,3\},$$in the basis$$|n,a_{n}\rangle =\left( \begin{array}{c} \delta _{1,a_{n}} \\ \delta _{2,a_{n}} \\ \delta _{3,a_{n}}\end{array}\right) _{n}\in V_{n}\text{ \ }\forall n\in \{1,...,\mathsf{N}\}\text{ and }a_{n}\in \{1,2,3\}\text{,}$$for which we have:$$\mathsf{N}_{1}+\mathsf{N}_{2}+\mathsf{N}_{3}=\mathsf{N} \ .$$These operators generalize to higher rank the symmetry of the $R$-matrix given in the rank one case by the third component of spin. Then it holds: The transfer matrices $T_{1}^{\left( K\right) }(\lambda )$ and $T_{2}^{\left( K\right) }(\lambda )$ satisfy the following properties: i\) For any $K$-matrix defining a symmetry of the $R$-matrix, $T_{2}^{\left( K\right) }(\lambda )$ has the following $2 \mathsf{N}$ central zeroes: $$T_{2}^{\left( K\right) }(\pm q\xi _{b})=0\text{ \ }\forall b\in \{1,...,\mathsf{N}\},$$the quantum determinant reads:$$q\text{-det}M^{(K)}(\lambda )=\text{det}K\text{ }\prod_{b=1}^{\mathsf{N}}(\lambda q/\xi _{b}-\xi _{b}/(q\lambda ))(\lambda /(q\xi _{b})-(q\xi _{b})/\lambda )(\lambda /(q^{2}\xi _{b})-(q^{2}\xi _{b})/\lambda ),$$and the following fusion identities hold for any $b \in \{1,...,\mathsf{N}\}$:$$\begin{aligned} T_{1}^{\left( K\right) }(\xi _{b})T_{1}^{\left( K\right) }(\xi _{b}/q) &=&T_{2}^{\left( K\right) }(\xi _{b}), \\ T_{1}^{\left( K\right) }(\xi _{b})T_{2}^{\left( K\right) }(\xi _{b}/q) &=&q\text{-det}M^{(K)}(\xi _{b}).\end{aligned}$$ Moreover, in the case of a diagonal twist $K^{(1)}$: ii\) $\lambda ^{\mathsf{N}}T_{1}^{\left( K\right) }(\lambda )$ is a degree polynomial in $\lambda ^{2}$ with the following asymptotics:$$T_{1}^{(\pm \infty |K)}\equiv \lim_{\log \lambda \rightarrow \pm \infty }\lambda ^{\mp \mathsf{N}}T_{1}^{\left( K\right) }(\lambda )=\left( \pm 1\right) ^{\mathsf{N}}\frac{\alpha q^{\pm \mathsf{N}_{1}}+\beta q^{\pm \mathsf{N}_{2}}+\gamma q^{\pm \mathsf{N}_{3}}}{\prod_{n=1}^{\mathsf{N}}\xi _{n}^{\pm 1}}.$$ iii\) $\lambda ^{2\mathsf{N}}T_{2}^{\left( K\right) }(\lambda )$ is a degree $2\mathsf{N}$ polynomial in $\lambda ^{2}$ with the following asymptotics:$$T_{2}^{(\pm \infty |K)}\equiv \lim_{\log \lambda \rightarrow \pm \infty }\lambda ^{\mp 2\mathsf{N}}T_{2}^{\left( K\right) }(\lambda )=\frac{\alpha \beta q^{\pm (\mathsf{N}_{1}+\mathsf{N}_{2})}+\alpha \gamma q^{\pm (\mathsf{N}_{1}+\mathsf{N}_{3})}+\beta \gamma q^{\pm (\mathsf{N}_{2}+\mathsf{N}_{3})}}{q^{\pm \mathsf{N}}\prod_{n=1}^{\mathsf{N}}\xi _{n}^{\pm 2}}.$$ iv\) The operators $N_i$ commute with the transfer matrices.\ In the case of non-diagonal twists $K^{(2)}$ or $K^{(3)}$: v\) $\lambda ^{(\mathsf{N}-1/3)}T_{1}^{\left( K\right) }(\lambda )$ is a degree $\mathsf{N}-1$ polynomial in $\lambda ^{2}$. vi\) $\lambda ^{(2\mathsf{N}-2/3)}T_{2}^{\left( K\right) }(\lambda )$ is a degree $2\mathsf{N}-1$ polynomial in $\lambda ^{2}$. Let us introduce the functions$$\begin{aligned} f_{l,\mathbf{h}}^{(a,m)}(\lambda )& =\left( \prod_{b=1}^{\mathsf{N}}\frac{\lambda /\xi _{b}^{(-1)}-\xi _{b}^{(-1)}/\lambda }{\xi _{l}^{(h_{l})}/\xi _{b}^{(-1)}-\xi _{b}^{(-1)}/\xi _{l}^{(h_{l})}}\right) ^{\delta _{m,2}}\left( \frac{t_{h_{1},...,h_{\mathsf{N}}}\lambda /\xi _{n}+\xi _{n}/(t_{h_{1},...,h_{\mathsf{N}}}\lambda )}{t_{h_{1},...,h_{\mathsf{N}}}+1/t_{h_{1},...,h_{\mathsf{N}}}}\right) ^{\delta _{1,a}} \notag \\ & \times \prod_{b\neq n,b=1}^{\mathsf{N}}\frac{\lambda /\xi _{b}^{(h_{b})}-\xi _{b}^{(h_{b})}/\lambda }{\xi _{n}^{(h_{n})}/\xi _{b}^{(h_{b})}-\xi _{b}^{(h_{b})}/\xi _{n}^{(h_{n})}},\text{\ \ \ \ \ \ }\xi _{b}^{(h)}\left. =\right. \xi _{b}/q^{h},\text{ }t_{h_{1},...,h_{\mathsf{N}}}\left. =\right. q^{-\sum_{a=1}^{\mathsf{N}}h_{a}},\end{aligned}$$that are well defined under the assumption that $q$ is not a root of unity and $h_{n} \in \{0,...,n-1\}$ for any $n\in \{1,...,\mathsf{N}\}$. We also define the following functions of the operators $N_i$: $$T_{1,\mathbf{h}}^{(K^{\left( a\right) },\infty )}(\lambda )=\delta _{1,a}\frac{\alpha \cosh \eta \mathsf{N}_{1}+\beta \cosh \eta \mathsf{N}_{2}+\gamma \cosh \eta \mathsf{N}_{3}}{\cosh (\eta \sum_{b=1}^{\mathsf{N}}h_{b})}\prod_{b=1}^{\mathsf{N}}(\lambda /\xi _{b}^{(h_{b})}-\xi _{b}^{(h_{b})}/\lambda ), \label{T_1,h-asymp}$$and$$\begin{aligned} T_{2,\mathbf{h}}^{(K^{\left( a\right) },\infty )}(\lambda )& =\delta _{1,a}\frac{\alpha \beta \cosh \eta (\mathsf{N}_{1}+\mathsf{N}_{2})+\alpha \gamma \cosh \eta (\mathsf{N}_{1}+\mathsf{N}_{3})+\beta \gamma \sinh \eta (\mathsf{N}_{2}+\mathsf{N}_{3})}{\cosh (\eta \sum_{b=1}^{\mathsf{N}}h_{b})} \notag \\ & \times \prod_{b=1}^{\mathsf{N}}(\lambda /\xi _{b}^{(h_{b})}-\xi _{b}^{(h_{b})}/\lambda )(\lambda /\xi _{b}^{(-1)}-\xi _{b}^{(-1)}/\lambda ) \ .\end{aligned}$$Then the next corollary holds: \[Coro-U-q(gl\_3)\]The transfer matrix $T_{1}^{(K^{\left( a\right) })}(\lambda )$ and $T_{2}^{(K^{\left( a\right) })}(\lambda )$ admit the following interpolation formulae:$$T_{1}^{(K^{\left( a\right) })}(\lambda )=T_{1,\mathbf{h}}^{(K^{\left( a\right) },\infty )}(\lambda )+\sum_{n=1}^{\mathsf{N}}f_{n,\mathbf{h}}^{\left( a,1\right) }(\lambda )T_{1}^{(K^{\left( a\right) })}(\xi _{n}^{(h_{n})}),$$and$$T_{2}^{(K^{\left( a\right) })}(\lambda )=T_{2,\mathbf{h}}^{(K^{\left( a\right) },\infty )}(\lambda )+\sum_{n=1}^{\mathsf{N}}f_{n,\mathbf{h}}^{\left( a,2\right) }(\lambda )T_{2}^{(K^{\left( a\right) })}(\xi _{n}^{\left( h_{n}\right) }),$$under the assumption that $q$ is not a root of unity and $h_{n} \in \{0, 1, 2\}$ for any $n\in \{1,...,\mathsf{N}\}$. Moreover, the following sum rules are satisfied:$$\begin{aligned} & \delta _{1,a}\left( \alpha \sinh \eta (\mathsf{N}_{1}-\sum_{b=1}^{\mathsf{N}}h_{b})+\beta \sinh \eta (\mathsf{N}_{2}-\sum_{b=1}^{\mathsf{N}}h_{b})+\gamma \sinh \eta (\mathsf{N}_{3}-\sum_{b=1}^{\mathsf{N}}h_{b})\right) \notag \\ & =\sum_{n=1}^{\mathsf{N}}\frac{T_{1}^{(K^{(a)})}(\xi _{n}^{\left( h_{n}\right) })}{2\prod_{b\neq n,b=1}^{\mathsf{N}}\xi _{n}^{\left( h_{n}\right) }/\xi _{b}^{\left( h_{b}\right) }-\xi _{b}^{\left( h_{b}\right) }/\xi _{n}^{\left( h_{n}\right) }},\end{aligned}$$and$$\begin{aligned} &\delta _{1,a}\left( \alpha \beta \sinh \eta (\mathsf{N}_{1}+\mathsf{N}_{2}-\sum_{b=1}^{\mathsf{N}}h_{b})+\alpha \gamma \sinh \eta (\mathsf{N}_{1}+\mathsf{N}_{3}-\sum_{b=1}^{\mathsf{N}}h_{b})+\beta \gamma \sinh \eta (\mathsf{N}_{2}+\mathsf{N}_{3}-\sum_{b=1}^{\mathsf{N}}h_{b})\right) \notag \\ & =\sum_{n=1}^{\mathsf{N}}\frac{T_{2}^{(K^{(a)})}(\xi _{n}^{\left( h_{n}\right) })}{2\prod_{b\neq n,b=1}^{\mathsf{N}}(\xi _{n}^{\left( h_{n}\right) }/\xi _{b}^{\left( h_{b}\right) }-\xi _{b}^{\left( h_{b}\right) }/\xi _{n}^{\left( h_{n}\right) })\prod_{b=1}^{\mathsf{N}}(\xi _{n}^{\left( h_{n}\right) }/\xi _{b}^{\left( -1\right) }-\xi _{b}^{\left( -1\right) }/\xi _{n}^{\left( h_{n}\right) })}.\end{aligned}$$$T_{1}^{(K^{\left( a\right) })}(\lambda )$ then completely characterizes $T_{2}^{(K^{\left( a\right) })}(\lambda )$ in terms of the fusion equations by:$$T_{2}^{(K^{\left( a\right) })}(\lambda )=T_{2,\mathbf{h=0}}^{(K^{\left( a\right) },\infty )}(\lambda )+\sum_{n=1}^{\mathsf{N}}f_{n,\mathbf{h=0}}^{\left( a,2\right) }(\lambda )T_{1}^{(K^{\left( a\right) })}(\xi _{n}/q)T_{1}^{(K^{\left( a\right) })}(\xi _{n}).$$ We have to use just the known central zeroes and asymptotic behavior to prove the above interpolation formula once $T_{2}^{(K^{\left( a\right) })}(\xi _{n})$ is given by the fusion equations. The sum rules just follow from the fact that in this trigonometric case we know the asymptotic behavior of the transfer matrices in two points (at $\lambda$ going to zero and to infinity) while we still reconstruct these degree $\mathsf{N}$ polynomials in $\mathsf{N}$ points which leads to the sum rule. SoV covector basis generated by transfer matrix action ------------------------------------------------------ The general Proposition 2.6 of our article [@MaiN18] for the construction of the SoV covector basis applies in particular to the fundamental representations of the trigonometric Yang-Baxter algebra. The twist $K^{\left( a\right) }$ are diagonalizable and with simple spectrum $3\times 3$ matrices, as soon as $\alpha ,$ $\beta $ and $\gamma $ are all different in the case $a=1$ and $\alpha \beta \gamma \neq 0$ in the case $a=2$ and $a=3$. Let us denote by $K_{J}$ the diagonal form of the matrix $K$ and $W_K$ the invertible matrix defining the change of basis:$$K=W_{K}K_{J}W_{K}^{-1}\text{ \ with \ }K_{J}=\left( \begin{array}{ccc} \mathsf{k}_{1} & 0 & 0 \\ 0 & \mathsf{k}_{2} & 0 \\ 0 & 0 & \mathsf{k}_{3}\end{array}\right) .$$Then the following theorem holds: For any diagonalizable $3\times 3$ twist matrix $K$ having simple spectrum, the following set:$$\langle h_{1},...,h_{\mathsf{N}}|\equiv \langle S|\prod_{n=1}^{\mathsf{N}}(T_{1}^{(K)}(\xi _{n}))^{h_{n}}\text{ \ for any }\{h_{1},...,h_{\mathsf{N}}\}\in \{0,1,2\}^{\otimes \mathsf{N}},$$forms a covector basis of $\mathcal{H}$, for almost any choice of $\langle S| $ and of the inhomogeneities under the condition $$\xi _{a}\neq \xi _{b}q^h \,\,\,\,\,\forall a\neq b\in \{1,...,\mathsf{N}\} \text{ and } h\in\{-2,-1,0,1,2\}. \label{Inhomog-cond}$$Moreover, a proper choice of the state $\langle S|$ has the following tensor product form:$$\langle S|=\bigotimes_{a=1}^{\mathsf{N}}(x,y,z)_{a}\Gamma _{W}^{-1},\text{ \ \ }\Gamma _{W}=\bigotimes_{a=1}^{\mathsf{N}}W_{K,a}$$under the condition $x\,y\,z\neq 0$. As a corollary of the general Proposition 2.6 of [@MaiN18], this set of covectors is a covector basis of $\mathcal{H}$ as soon as we can show that the covectors:$$(x,y,z)_{a}W^{-1},(x,y,z)_{a}W^{-1}K,(x,y,z)_{a}W^{-1}K^{2},$$or equivalently:$$(x,y,z)_{a},(x,y,z)_{a}K_{J},(x,y,z)_{a}K_{J}^{2},$$form a basis in $V_{a}\cong \mathbb{C}^{3}$, that is that the following determinant is non-zero:$$\text{det}||\left( (x,y,z)K_{J}^{i-1}e_{j}(a)\right) _{i,j\in \{1,2,3\}}||=-xyzV(\mathsf{k}_{0},\mathsf{k}_{1},\mathsf{k}_{2}),$$which leads to the given requirements on the components $x,y,z\in \mathbb{C}$ of the three dimensional covector. Transfer matrix spectrum in our SoV approach: the $\mathcal{U}_{q}(\widehat{gl_{3}})$ case ========================================================================================== Discrete spectrum characterization by SoV ----------------------------------------- For any given twist matrix $K$ diagonalizable and with simple spectrum, the following characterization of the transfer matrix spectrum holds: \[Discrete-Ch-3\_q\]The spectrum of $T_{1}^{(K)}(\lambda )$ is characterized by:$$\Sigma _{T^{(K)}}=\bigcup_{0\leq l+m\leq \mathsf{N}}\Sigma _{T^{(K)}}^{\left( l,m\right) },$$where $l, m$ are positive integers and $$\Sigma _{T^{(K)}}^{\left( l,m\right) }=\left\{ t_{1}(\lambda )=t_{1}(l,m,\alpha,\beta,\gamma)\prod_{b=1}^{\mathsf{N}}(\lambda /\xi _{b}-\xi _{b}/\lambda )+\sum_{a=1}^{\mathsf{N}}f_{n,\mathbf{h=0}}^{\left( a,1\right) }(\lambda )x_{a} \right\}, \label{SET-T}$$for any set $\{x_1, \dots, x_{\mathsf{N}} \}$ belonging to $S_{T^{(K)}}^{\left( l,m\right)}$ where we have defined: $$t_{1}(l,m,\alpha,\beta,\gamma)\equiv\delta _{1,a}\left( \alpha \cosh \eta l+\beta \cosh \eta m+\gamma \cosh \eta (l+m-\mathsf{N})\right)$$ under the assumption that the $3\times 3$ twist matrix $K$ is simple and diagonalizable and the inhomogeneities satisfy $(\ref{Inhomog-cond})$. Here, $S_{T^{(K)}}^{\left( l,m\right) }$ stand for the set of solutions $\{x_1, \dots,x_{\mathsf{N}}\}$ to the following system of $\mathsf{N}$ cubic equations:$$\begin{aligned} & x_{n}[\delta _{1,a}\left( \alpha \beta \cosh \eta (l+m)+\alpha \gamma \cosh \eta (\mathsf{N}-m)+\beta \gamma \sinh \eta (\mathsf{N}-l)\right) \prod_{b=1}^{\mathsf{N}}(\xi _{n}^{(1)}/\xi _{b}-\xi _{b}/\xi _{n}^{(1)}) \notag \\ & \times (\xi _{n}^{(1)}/\xi _{b}^{(-1)}-\xi _{b}^{(-1)}/\xi _{n}^{(1)})+\sum_{m=1}^{\mathsf{N}}f_{m,\mathbf{h=0}}^{\left( a,2\right) }(\xi _{m}^{(1)})t_{1}(\xi _{m}^{(1)})x_{m}]\left. =\right. q\text{-det}M^{(K)}(\xi _{a}),\end{aligned}$$in $\mathsf{N}$ unknown $\{x_{1},...,x_{\mathsf{N}}\}$. Moreover, $T_{1}^{(K)}(\lambda )$ is diagonalizable with simple spectrum and for any $t_{1}(\lambda )\in \Sigma _{T^{(K)}}$ the associated unique eigenvector $|t\rangle $ has the following wave function in the covector SoV basis:$$\langle h_{1},...,h_{\mathsf{N}}|t\rangle =\prod_{n=1}^{\mathsf{N}}t_{1}^{h_{n}}(\xi _{n}), \label{SoV-Ch-T-eigenV}$$where the overall normalization has been fixed by imposing $\langle S|t\rangle =1$. The transfer matrix fusion equations:$$T_{1}^{\left( K\right) }(\xi _{b})T_{2}^{\left( K\right) }(\xi _{b}/q)=q\text{-det}M^{(K)}(\xi _{b}),\text{ }\forall b\in \{1,...,\mathsf{N}\}, \label{scalar-fusion-1}$$when rewritten for the eigenvalues take exactly the form of the given system of $\mathsf{N}$ cubic equations in $\mathsf{N}$ unknowns $\{x_{1},...,x_{\mathsf{N}}\}$, once we use the known analyticity and central zeroes. Consequently, this system has to be satisfied and the given characterization of any eigenvector $|t\rangle $ is implied. The reverse statement has to be shown now. In particular, we have to prove that given a polynomial of the above form satisfying this system then it is an eigenvalue of the transfer matrix $T_{1}^{(K)}(\lambda )$. This is done by proving that the vector $|t\rangle $ defined in $(\ref{SoV-Ch-T-eigenV})$ is a transfer matrix eigenvector using our SoV basis:$$\langle h_{1},...,h_{\mathsf{N}}|T_{1}^{(K)}(\lambda )|t\rangle =t_{1}(\lambda )\langle h_{1},...,h_{\mathsf{N}}|t\rangle ,\text{ }\forall \{h_{1},...,h_{\mathsf{N}}\}\in \{0,1,2\}^{\otimes \mathsf{N}}. \label{Eigen-cond-sl3}$$Let us point out that as a consequence of the Corollary \[Coro-U-q(gl\_3)\], in order to prove this identity it is enough to prove it in a generic $\mathsf{N}$-upla of points $\xi _{a}^{(k_{a})}$ for any $a\in \{1,...,\mathsf{N}\}$. Indeed, let us assume that it holds:$$\langle h_{1},...,h_{\mathsf{N}}|T_{1}^{(K)}(\xi _{a}^{(k_{a})})|t\rangle =t_{1}(\xi _{a}^{(k_{a})})\langle h_{1},...,h_{\mathsf{N}}|t\rangle ,\text{ }\forall \{h_{1},...,h_{\mathsf{N}}\}\in \{0,1,2\}^{\otimes \mathsf{N}},$$then we have that:$$\begin{aligned} & \delta _{1,a}\langle h_{1},...,h_{\mathsf{N}}|\left( \alpha \sinh \eta (\mathsf{N}_{1}-\sum_{b=1}^{\mathsf{N}}k_{b})+\beta \sinh \eta (\mathsf{N}_{2}-\sum_{b=1}^{\mathsf{N}}k_{b})+\gamma \sinh \eta (\mathsf{N}_{3}-\sum_{b=1}^{\mathsf{N}}k_{b})\right) |t\rangle \notag \\ & =\langle h_{1},...,h_{\mathsf{N}}|\left( \sum_{n=1}^{\mathsf{N}}\frac{T_{1}^{(K^{(a)})}(\xi _{n}^{\left( k_{n}\right) })}{2\prod_{b\neq n,b=1}^{\mathsf{N}}\xi _{n}^{\left( k_{n}\right) }/\xi _{b}^{\left( k_{b}\right) }-\xi _{b}^{\left( k_{b}\right) }/\xi _{n}^{\left( k_{n}\right) }}\right) |t\rangle \\ & =\left( \sum_{n=1}^{\mathsf{N}}\frac{t_{1}(\xi _{n}^{\left( k_{n}\right) })}{2\prod_{b\neq n,b=1}^{\mathsf{N}}\xi _{n}^{\left( k_{n}\right) }/\xi _{b}^{\left( k_{b}\right) }-\xi _{b}^{\left( k_{b}\right) }/\xi _{n}^{\left( k_{n}\right) }}\right) \langle h_{1},...,h_{\mathsf{N}}|t\rangle \\ & =\delta _{1,a}( \alpha \sinh \eta (l-\sum_{b=1}^{\mathsf{N}}k_{b})+\beta \sinh \eta (m-\sum_{b=1}^{\mathsf{N}}k_{b})+\gamma \sinh \eta (\mathsf{N}-(l+m)-\sum_{b=1}^{\mathsf{N}}k_{b})) \langle h_{1},...,h_{\mathsf{N}}|t\rangle ,\end{aligned}$$which implies that for $a=1$, $|t\rangle $ is an eigenvector of the charges $\mathsf{N}_{1}$, $\mathsf{N}_{2}$ and $\mathsf{N}_{3}$ with eigenvalues, respectively, $l,$ $m$ and $\mathsf{N}-(l+m)$ which in turn fix the asymptotics of the transfer matrices in that case. That is for $a=1$, any $t_{1}(\lambda )$ in $\left( \ref{SET-T}\right) $, $t_{1}(\lambda )$ and $|t\rangle $ can be transfer matrix eigenvalue and eigenvector only associated to the common eigenspace of $\mathsf{N}_{1}$ and $\mathsf{N}_{2}$ corresponding to the nonnegative integer eigenvalues $l$ and $m$, respectively. Notice that if $a \neq 1$ the asymptotic term is zero and the $\mathsf{N}_i$ are no longer symmetries of the transfer matrices and so we don’t need to distinguish those values in the discussion. Let $h_{a}=0,1$ and $h_{b}\in \{0,1,2\}$ for any $b\in \{1,...,\mathsf{N}\}\backslash a$, then we have the following identities:$$\begin{aligned} \langle h_{1},...,h_{\mathsf{N}}|T_{1}^{(K^{\left( a\right) })}(\xi _{a})|t\rangle & =\langle h_{1},...,h_{a}+1,...,h_{\mathsf{N}}|t\rangle \notag \\ & =t_{1}(\xi _{a})\langle h_{1},...,h_{a},...,h_{\mathsf{N}}|t\rangle , \label{Id-step1}\end{aligned}$$as a direct consequence of the definition of the covector SoV basis and of the state $|t\rangle $. So that we are left with the proof of the statement in the case $h_{a}=2$. In this case we want to prove that it holds:$$\langle h_{1},...,h_{\mathsf{N}}|T_{1}^{(K^{\left( a\right) })}(\xi _{a}/q)|t\rangle =t_{1}(\xi _{a}/q)\langle h_{1},...,h_{a},...,h_{\mathsf{N}}|t\rangle , \label{Id-step2}$$the proof is done by induction on the number of zeros contained in $\{h_{1},...,h_{\mathsf{N}}\}\in \{0,1,2\}^{\otimes \mathsf{N}}$. It is developed just following the same steps we have developed in the case of the fundamental representation of the $Y(gl_{3})$ rational Yang-Baxter algebra in our paper [@MaiN18]. In fact we have only to take into account the different functional form of the transfer matrix, i.e. they are Laurent polynomials and not simple polynomials, and the fact that the asymptotic behavior of the transfer matrices are not central for $a=1$ but take fixed values in any common eigenspace of $\mathsf{N}_{1}$ and $\mathsf{N}_{2}$ that is stable by the action of the transfer matrices since for $a=1$ they commute with each $\mathsf{N}_{i}$. For completeness we dedicate Appendix A to make explicit these steps of the proof. Spectrum characterization by quantum spectral curve --------------------------------------------------- The discrete characterization of the transfer matrix spectrum derived in the previous section in our SoV basis allows us to introduce the following quantum spectral curve functional reformulation. Let us first introduce the functions:$$\begin{aligned} \delta _{3}(\lambda ) &=&\delta _{1}(\lambda )\delta _{1}(\lambda /q)\delta _{1}(\lambda /q^{2}), \\ \delta _{2}(\lambda ) &=&\delta _{1}(\lambda )\delta _{1}(\lambda /q), \\ \delta _{1}(\lambda ) &=&\delta _{0}\prod_{a=1}^{\mathsf{N}}(\lambda q/\xi _{a}-\xi _{a}/(\lambda q)),\end{aligned}$$ \[Functio-Ch-3\_q\]Here we consider the case of a twist $K$ which is diagonal and with simple spectrum, then the entire functions $t_{1}(\lambda )$ is a $T_{1}^{\left( K\right) }(\lambda )$ transfer matrix eigenvalue belonging to $\Sigma _{T^{(K)}}^{\left( \nu_{1},\nu_{2}\right) }$, with $\nu_{1},$ $\nu_{2}$ two nonnegative integers satisfying:$$\nu_{1}+\nu_{2}\leq\mathsf{N},$$iff there exists an unique Laurent polynomial of the form:$$\varphi _{t}(\lambda )=\prod_{a=1}^{\mathsf{M}}(\lambda /\lambda _{a}-\lambda _{a}/\lambda )\text{\ \ with }\mathsf{M}\leq \mathsf{N}\text{ and }\lambda _{a}\neq \xi _{n}\text{ }\forall (a,n)\in \{1,...,\mathsf{M}\}\times \{1,...,\mathsf{N}\}, \label{Phi-form}$$solution of the following quantum spectral curve functional equation:$$\delta _{3}(\lambda )\varphi _{t}(\lambda /q^{3})-\delta _{2}(\lambda )t_{1}(\lambda /q^{2})\varphi _{t}(\lambda /q^{2})+\delta _{1}(\lambda )t_{2}(\lambda /q)\varphi _{t}(\lambda /q)-q\text{-det}M_{a}^{(K)}(\lambda )\varphi _{t}(\lambda )\left. =\right. 0$$where we have defined$$\begin{aligned} t_{2}(\lambda )& =\left( \alpha \beta \cosh \eta (\nu_{1}+\nu_{2})+\alpha \gamma \cosh \eta (\nu_{1}+\nu_{3})+\beta \gamma \sinh \eta (\nu_{2}+\nu_{3})\right) \notag \\ & \times \prod_{b=1}^{\mathsf{N}}(\lambda /\xi _{b}-\xi _{b}/\lambda )(\lambda /\xi _{b}^{(-1)}-\xi _{b}^{(-1)}/\lambda )+\sum_{n=1}^{\mathsf{N}}f_{n,\mathbf{h=0}}^{\left( a\right) }(\xi _{n}^{(1)})t_{1}(\xi _{n}^{(1)})t_{1}(\xi _{n}),\end{aligned}$$with $\nu_3 =\mathsf{N} -\nu_1 -\nu_2$ and fixed[^2]:$$\delta _{0}=\mathsf{k}_{i}\text{ for one fixed }i\in \{1,2,3\},$$with$$\mathsf{M}=\mathsf{M}-\nu_{i}.$$Moreover, up to an overall normalization the common transfer matrix eigenvector $|t\rangle $ admits the following separate representation:$$\langle h_{1},...,h_{\mathsf{N}}|t\rangle =\prod_{a=1}^{\mathsf{N}}\delta _{1}^{h_{a}}(\xi _{a})\varphi _{t}^{h_{a}}(\xi _{a}/q)\varphi _{t}^{2-h_{a}}(\xi _{a}).$$ This is a special case of the proof that will be given in the general $U_{q}(\widehat{gl_{n}})$ case. Let us comment that the Theorem \[Discrete-Ch-3\_q\] applies for any integrable boundary conditions while the previous Theorem [Functio-Ch-3\_q]{} applies only to the case of diagonal twists. The non-diagonal case is not presented explicitly but we can similarly derive a functional equation of third order type which however is of inhomogeneous type, if we ask that the $\varphi _{t}(\lambda )$ has the same Laurent polynomial form as indicated in $\left( \ref{Phi-form}\right) $. While at this stage this is a simple exercise we think that the main interesting question to investigate about the non-diagonal twists is if, with a different (periodicity) definition of the $\varphi _{t}(\lambda )$ function, we can reestablish an homogeneous equation as it has been proven for the $U_{q}(\widehat{gl_{2}})$ case in [@NicT15]. ABA rewriting of transfer matrix eigenvectors --------------------------------------------- An equivalent rewriting of algebraic Bethe ansatz type for the transfer matrix eigenvectors can be derived on the basis of their SoV representation. Let us first define one common eigenvector of the transfer matrix $T_{1}^{\left( K\right) }(\lambda )$ and $T_{2}^{\left( K\right) }(\lambda )$: Let $K$ be a diagonal $3\times 3$ matrix having three distinct eigenvalues $\mathsf{k}_i$:$$K=\left( \begin{array}{ccc} \mathsf{k}_{1} & 0 & 0 \\ 0 & \mathsf{k}_{2} & 0 \\ 0 & 0 & \mathsf{k}_{3}\end{array}\right) ,$$then:$$|t_{0}\rangle =\bigotimes_{a=1}^{\mathsf{N}}\left( \begin{array}{c} 1 \\ 0 \\ 0\end{array}\right) _{a},$$is a common eigenvector of the transfer matrices $T_{1}^{\left( K\right) }(\lambda )$ and $T_{2}^{\left( K\right) }(\lambda )$:$$\begin{aligned} T_{1}^{\left( K\right) }(\lambda )|t_{0}\rangle & =|t_{0}\rangle t_{1,0}(\lambda )\text{ \ with }t_{1,0}(\lambda )=\mathsf{k}_{1}\prod_{a=1}^{\mathsf{N}}(\lambda q/\xi _{a}-\xi _{a}/(\lambda q))+(\mathsf{k}_{2}+\mathsf{k}_{3})\prod_{a=1}^{\mathsf{N}}(\lambda /\xi _{a}-\xi _{a}/\lambda ), \\ T_{2}^{\left( K\right) }(\lambda )|t_{0}\rangle & =|t_{0}\rangle t_{2,0}(\lambda )\text{ \ with }t_{2,0}(\lambda )=\left\{ \begin{array}{l} \prod_{a=1}^{\mathsf{N}}(\lambda /(q\xi _{a})-(q\xi _{a})/\lambda )(\mathsf{k}_{3}\mathsf{k}_{2}\prod_{a=1}^{\mathsf{N}}(\lambda /\xi _{a}-\xi _{a}/\lambda ) \\ +(\mathsf{k}_{2}\mathsf{k}_{1}+\mathsf{k}_{3}\mathsf{k}_{1})\prod_{a=1}^{\mathsf{N}}(\lambda q/\xi _{a}-\xi _{a}/(\lambda q)),\end{array}\right.\end{aligned}$$and the quantum spectral curve$$\delta _{3}(\lambda )-\delta _{2}(\lambda )t_{1}(\lambda /q^{2})+\delta _{1}(\lambda )t_{2}(\lambda /q)-q\text{-det}M_{a}^{(K)}(\lambda )\left. =\right. 0$$ with constant $\varphi _{t}(\lambda )$ is satisfied by the couple of eigenvalues $t_{1,0}(\lambda )$ and $t_{2,0}(\lambda )$ for $\delta _{0}=\mathsf{k}_{1}.$ This is a standard result which follows by proving that it holds:$$A_{i}^{(I)}(\lambda )|t_{0}\rangle =|t_{0}\rangle \prod_{a=1}^{\mathsf{N}}(\lambda q^{\delta _{i,1}}/\xi _{a}-\xi _{a}/(\lambda q^{\delta _{i,1}})),\text{ \ }C_{i}^{(I)}(\lambda )|t_{0}\rangle =0,\text{ \ }i\in \{1,2,3\},$$where the upper index $I$ stands for the identity twist matrix, from which it is simple to verify by direct computation that the $t_{1,0}(\lambda )$ and $t_{2,0}(\lambda )$ satisfies the fusion equations $(\ref{scalar-fusion-1})$ and that it holds:$$\begin{aligned} t_{1,0} &\equiv &\lim_{\lambda \rightarrow \pm \infty }\lambda ^{\mp \mathsf{N}}t_{1,0}(\lambda )=\left( -1\right) ^{\frac{1\pm 1}{2}\mathsf{N}}\left( \mathsf{k}_{1}q^{\pm \mathsf{N}}+\mathsf{k}_{2}+\mathsf{k}_{3}\right) \left( \prod_{a=1}^{\mathsf{N}}\xi _{a}\right) ^{\mp 1}, \\ t_{2,0} &\equiv &\lim_{\lambda \rightarrow \pm \infty }\lambda ^{\mp 2\mathsf{N}}t_{2,0}(\lambda )=\left( \prod_{a=1}^{\mathsf{N}}\xi _{a}\right) ^{\mp 2}q^{\mp \mathsf{N}}(\mathsf{k}_{3}\mathsf{k}_{2}+(\mathsf{k}_{2}\mathsf{k}_{1}+\mathsf{k}_{3}\mathsf{k}_{1})q^{\pm \mathsf{N}}),\end{aligned}$$so that $t_{1,0}(\lambda )$ satisfies the SoV characterization of the eigenvalues of $T_{1}^{\left( K\right) }(\lambda )$. Observing now that it holds:$$t_{1,0}(\xi _{a})=\delta _{1}(\xi _{a})\text{ \ \ for any }a\in \{1,...,\mathsf{N}\}$$it follows that the associated $\varphi _{t}(\lambda )$ satisfies the equations:$$\varphi _{t}(\xi _{a})=\varphi _{t}(\xi _{a}/q)\text{ \ \ for any }a\in \{1,...,\mathsf{N}\}$$and so $\varphi _{t}(\lambda )$ is a constant. In our SoV basis we can now define the operator $\mathbb{B}^{(K)}\left( \lambda \right) $ as the one parameter family of commuting operators through the following characterization:$$\langle h_{1},...,h_{\mathsf{N}}|\mathbb{B}^{(K)}\left( \lambda \right) =\text{ }b_{h_{1},...,h_{\mathsf{N}}}(\lambda )\langle h_{1},...,h_{\mathsf{N}}|,$$where we have defined$$b_{h_{1},...,h_{\mathsf{N}}}(\lambda )=\prod_{a=1}^{\mathsf{N}}(\lambda /\xi _{a}-\xi _{a}/\lambda )^{2-h_{a}}(\lambda q/\xi _{a}-\xi _{a}/(q\lambda ))^{h_{a}} \ .$$Then the following corollary holds: The following algebraic Bethe ansatz type representation$$|t\rangle =\prod_{a=1}^{\mathsf{M}}\mathbb{B}^{(K)}(\lambda _{a})|t_{0}\rangle \text{\ \ with }\mathsf{M}\leq \mathsf{N}\text{ and }\lambda _{a}\neq \xi _{n}\text{ }\forall (a,n)\in \{1,...,\mathsf{M}\}\times \{1,...,\mathsf{N}\},$$holds for the unique (up to trivial scalar multiplication) eigenvector $|t\rangle $ associated to any given $t_{1}(\lambda )\in \Sigma _{T^{(K)}}\equiv \bigcup_{\forall \nu_{i}\geq 0\text{ : }\nu_{1}+\nu_{2}\leq \mathsf{N}}\Sigma _{T^{(K)}}^{\left( \nu_{1},\nu_{2}\right) }$. Here the $\lambda _{a}$ are the roots of the unique Laurent polynomial $\varphi _{t}(\lambda )$ associated to $t_{1}(\lambda )\in \Sigma _{T^{(K)}}$. The proof is standard, the following chain of identities holds$$\begin{aligned} \langle h_{1},...,h_{\mathsf{N}}|\prod_{a=1}^{\mathsf{M}}\mathbb{B}^{(K)}(\lambda _{a})|t_{0}\rangle & =\prod_{j=1}^{\mathsf{M}}b_{h_{1},...,h_{\mathsf{N}}}(\lambda _{j})\text{ }\langle h_{1},...,h_{\mathsf{N}}|t_{0}\rangle \notag \\ & =\prod_{j=1}^{\mathsf{M}}\prod_{a=1}^{\mathsf{N}}(\lambda _{j}/\xi _{a}-\xi _{a}/\lambda _{j})^{2-h_{a}}(\lambda _{j}q/\xi _{a}-\xi _{a}/(q\lambda _{j}))^{h_{a}}\prod_{a=1}^{\mathsf{N}}\delta _{1}^{h_{a}}(\xi _{a}) \notag \\ & =\prod_{a=1}^{\mathsf{N}}\delta _{1}^{h_{a}}(\xi _{a})\varphi _{t}(\xi _{a})^{2-h_{a}}\varphi _{t}(\xi _{a}/q)^{h_{a}},\end{aligned}$$which coincides with the SoV characterization of the transfer matrix eigenvector once we recall that:$$\langle h_{1},...,h_{\mathsf{N}}|t_{0}\rangle =\prod_{a=1}^{\mathsf{N}}\delta _{1}^{h_{a}}(\xi _{a}).$$ Transfer matrices for fundamental evaluation representations of $\mathcal{U}_{q}(\widehat{gl_{n}})$ ==================================================================================================== Let us consider now the general higher rank $n-1$, with $n\geq 3$ case. In particular, here we take the following $R$-matrix:$$\begin{aligned} R_{a,b}(\lambda )& =\left( \frac{\lambda }{q}-\frac{q}{\lambda }\right) \sum_{k=1}^{n}E_{kk}^{(a)}\otimes E_{kk}^{(b)}+\left( \lambda -\frac{1}{\lambda }\right) \sum_{p=1}^{n}\sum_{k=1,k\neq p}^{n}E_{kk}^{(a)}\otimes E_{pp}^{(b)} \notag \\ & +\left( q-\frac{1}{q}\right) \sum_{1\leq k<p\leq n}\left( \lambda ^{(n-2(p-k))/n}E_{kp}^{(a)}\otimes E_{pk}^{(b)}+\lambda ^{-(n-2(p-k))/n}E_{pk}^{(a)}\otimes E_{kp}^{(b)}\right) \in \text{End}(V_{a}\otimes V_{b})\end{aligned}$$which is the *trigonometric* principal gradation solution[^3] [@BabdeVV81; @Jim86] of the Yang-Baxter equation:$$R_{12}(\lambda /\mu )R_{13}(\lambda )R_{23}(\mu )=R_{23}(\mu )R_{13}(\lambda )R_{12}(\lambda /\mu )\in \text{End}(V_{1}\otimes V_{2}\otimes V_{3}),$$where $V_{i}\simeq \mathbb{C}^{n}$ for any $i\in \{1,2,3\}$, and it is associated to the fundamental evaluation representations of $\mathcal{U}_{q}(\widehat{gl_{n}})$ [KulR83a,KulR82,Jim85,Dri87,ChaP94]{}. Above, we have used the standard notation for the elementary matrices $E_{lm}\in $End$(V\simeq \mathbb{C}^{n})$, $(l,m)\in \{1,...,n\}\times \{1,...,n\}$:$$\left( E_{lm}\right) _{\alpha \beta }=\delta _{\alpha l}\delta _{m\beta }\text{ \ }\forall (\alpha ,\beta )\in \{1,...,n\}\times \{1,...,n\}.$$In this paper, we analyze the fundamental representations of these rank $n-1$ trigonometric Yang-Baxter algebras, associated to the following monodromy matrices:$$M_{a}^{(K)}(\lambda )\equiv K_{a}R_{a,\mathsf{N}} (\lambda /\xi _{\mathsf{N}})\cdots R_{a,1}(\lambda /\xi _{1})\in \text{End}(V_{a}\otimes \mathcal{H}),$$where $\mathcal{H}=\bigotimes_{n=1}^{\mathsf{N}}V_{n}$ and $K\in $End$(V)$ is a symmetry (twist matrix):$$R_{12}(\lambda )K_{1}K_{2}=K_{2}K_{1}R_{12}(\lambda )\in \text{End}(V_{1}\otimes V_{2}).$$Then the one parameter family of operators$$T_{1}^{\left( K\right) }(\lambda )\equiv \text{tr}_{a}M_{a}^{\left( K\right) }(\lambda )\in \text{End}\mathcal{H},$$are the associated commuting transfer matrices. Here, we focus our attention on the case of diagonal quasi-periodic boundary conditions which are associated to generic diagonal $n\times n$ twist matrices having pairwise distinct eigenvalues $\mathsf{k}_{i}$:$$K=\left( \begin{array}{cccc} \mathsf{k}_{1} & 0 & \cdots & 0 \\ 0 & \mathsf{k}_{2} & \ddots & \vdots \\ \vdots & \ddots & \ddots & 0 \\ 0 & \cdots & 0 & \mathsf{k}_{n}\end{array}\right) .$$For this class of representations we will prove the complete spectrum characterization of the transfer matrix in terms of a specific class of polynomial solutions to the quantum spectral curve equation, an homogeneous functional equation to the finite difference of order $n$. Let us comment that, as for the case $n=3$, the symmetry of the principal gradation $R$-matrix extends also to non-diagonal twist matrices[^4] and that our construction of the SoV basis applies for these cases, as well as the derivation of the complete SoV characterization of the transfer matrix spectrum. However, for the non-diagonal twist matrices a natural reformulation of the transfer matrix spectrum leads to an inhomogeneous functional equation. We have decided to develop the case of non-diagonal twist matrix in some future analysis where, as already discussed at the end of section 3.2, the main interesting question is if under an apropiate choice of the $Q$-functions one can derive an homogeneous quantum spectral curve characterization. Such a statement indeed holds for both the fundamental and higher spin representations of the rank one trigonometric Yang-Baxter algebra, as proven in [@NicT15]. Fundamental properties of the hierarchy of transfer matrices ------------------------------------------------------------ Let us introduce the following operators, $\mathsf{N}_{i}\in $End$(\mathcal{H})$$$\mathsf{N}_{i}\otimes _{l=1}^{\mathsf{N}}|l,a_{l}\rangle =\otimes _{l=1}^{\mathsf{N}}|l,a_{l}\rangle \sum_{l=1}^{\mathsf{N}}\delta _{i,a_{l}}\text{ \ }\forall i\in \{1,...,n\},$$in the basis$$|l,a_{l}\rangle =\left( \begin{array}{c} \delta _{1,a_{l}} \\ \delta _{2,a_{l}} \\ \vdots \\ \delta _{n,a_{l}}\end{array}\right) _{l}\in V_{l}\text{ \ }\forall l\in \{1,...,\mathsf{N}\}\text{ and }a_{l}\in \{1,...,n\}\text{,}$$so that it holds:$$\sum_{i=1}^{\mathsf{N}}\mathsf{N}_{i}=\mathsf{N}.$$We will denote by $\nu_i$ the eigenvalues of the operators $\mathsf{N}_{i}$. Then we can recall the following relevant and known properties of the fused transfer matrices: The higher transfer matrices:$$T_{m}^{\left( K\right) }(\lambda )\equiv \text{tr}_{a_{1},...,a_{m}}P_{a_{1},...,a_{m}}^{-}M_{a_{m}}^{(K)}(\lambda )\cdots M_{a_{1}}^{(K)}(\lambda /q^{m-1}),\text{ \ \ }m\in \{1,...,n\}$$defines one parameter families of mutually commuting operators: $$\left[ T_{i}^{\left( K\right) }(\lambda ),T_{j}^{\left( K\right) }(\mu )\right] =0,\text{ \ \ }i,j\in \{1,...,n\}.$$The last quantum spectral invariant$$q\text{-det}M^{(K)}(\lambda )\equiv T_{n}^{\left( K\right) }(\lambda ),$$the so-called quantum determinant, is a central element of the algebra with the following explicit form:$$q\text{-det}M^{(K)}(\lambda )=\text{det}K\text{ }\prod_{b=1}^{\mathsf{N}}(\lambda q/\xi _{b}-\xi _{b}/(q\lambda ))\prod_{k=1}^{n-1}(\lambda /(q^{k}\xi _{b})-(q^{k}\xi _{b})/\lambda ).$$Moreover, the quantum spectral invariants have the following analyticity properties: a\) The following fusion identities holds for any $a \in \{1, \dots, \mathsf{N}\}$:$$T_{1}^{\left( K\right) }(\xi _{a})T_{m-1}^{\left( K\right) }(\xi _{a}/q)=T_{m}^{\left( K\right) }(\xi _{a}),\text{ \ }\forall m\in \{2,...,n\},$$and the following $(m-1)\mathsf{N}$ central zero conditions: $$\text{ }T_{m}^{\left( K\right) }(\pm q^{r}\xi _{a})=0\text{ \ }\forall r\in \{1,...,m-1\},a\in \{1,...,\mathsf{N}\},$$holds for any $m\in \{1,...,n-1\}$ and for any above diagonal $K$-matrix. b\) $\lambda ^{m\mathsf{N}}T_{m}^{\left( K\right) }(\lambda )$ is a degree $m\mathsf{N}$ polynomial in $\lambda ^{2}$ with the following asymptotics:$$T_{m}^{(\pm \infty |K)}\equiv \lim_{\log \lambda \rightarrow \pm \infty }\lambda ^{\mp \mathsf{N}}T_{m}^{\left( K\right) }(\lambda )=\left( \pm 1\right) ^{m\mathsf{N}}\frac{\sigma _{m}^{(n)}(\mathsf{k}_{1}q^{\pm \mathsf{N}_{1}},...,\mathsf{k}_{n}q^{\pm \mathsf{N}_{n}})}{q^{\pm m(m-1)\mathsf{N}/2}\prod_{n=1}^{\mathsf{N}}\xi _{n}^{\pm m}},$$in the case of the diagonal twist $K$, where $\sigma _{m}^{(n)}$ is the standard symmetric polynomial of degree $m$ in $n$ variables. Let us introduce the functions$$\begin{aligned} f_{l,\mathbf{h}}^{(m)}(\lambda )&=\frac{t_{h_{1},...,h_{\mathsf{N}}}\lambda /\xi _{l}+\xi _{l}/(t_{h_{1},...,h_{\mathsf{N}}}\lambda )}{t_{h_{1},...,h_{\mathsf{N}}}+1/t_{h_{1},...,h_{\mathsf{N}}}}\prod_{b=1}^{\mathsf{N}}\prod_{r=1}^{m-1}\frac{\lambda /\xi _{b}^{(-r)}-\xi _{b}^{(-r)}/\lambda }{\xi _{l}^{(h_{l})}/\xi _{b}^{(-r)}-\xi _{b}^{(-r)}/\xi _{l}^{(h_{l})}} \notag \\ &\times\prod_{b\neq l,b=1}^{\mathsf{N}}\frac{\lambda /\xi _{b}^{(h_{b})}-\xi _{b}^{(h_{b})}/\lambda }{\xi _{n}^{(h_{n})}/\xi _{b}^{(h_{b})}-\xi _{b}^{(h_{b})}/\xi _{n}^{(h_{n})}}, \\\end{aligned}$$ well defined under the assumption that $q$ is not a root of unity and $h_{l} \in \{0, \dots, n-1\}$ for any $l\in \{1,...,\mathsf{N}\}$, and$$\begin{aligned} T_{m,\mathbf{h}}^{(K,\infty )}(\lambda |\mathsf{N}_{1},...,\mathsf{N}_{n})& =\frac{\sum_{1\leq i_{1}<i_{2}<\cdots <i_{m-1}<i_{m}\leq n}\prod_{k=1}^{m}\mathsf{k}_{i_{k}}\cosh (\eta \sum_{k=1}^{m}\mathsf{N}_{i_{k}})}{\cosh (\eta \sum_{a=1}^{\mathsf{N}}h_{a})} \notag \\ & \times \prod_{b=1}^{\mathsf{N}}(\lambda /\xi _{b}^{(h_{b})}-\xi _{b}^{(h_{b})}/\lambda )\prod_{r=1}^{m-1}(\lambda /\xi _{b}^{(-r)}-\xi _{b}^{(-r)}/\lambda ), \label{Asym-Tm}\end{aligned}$$then the following corollary holds: Under the assumption that $q$ is not a root of unity and $h_{l} \in \{0, \dots, n-1\}$ for any $l\in \{1,...,\mathsf{N}\}$, the following interpolation formulae:$$T_{m}^{(K)}(\lambda )=T_{m,\mathbf{h}}^{(K,\infty )}(\lambda |\mathsf{N}_{1},...,\mathsf{N}_{n})+\sum_{l=1}^{\mathsf{N}}f_{l,\mathbf{h}}^{\left( m\right) }(\lambda )T_{m}^{(K)}(\xi _{l}^{(h_{l})})$$holds for the transfer matrix $T_{m}^{(K)}(\lambda )$ with $m\in \{1,...,n-1\}$ together with the following sum rules:$$\begin{aligned} & \sum_{1\leq i_{1}<i_{2}<\cdots <i_{m-1}<i_{m}\leq n}\prod_{k=1}^{m}\mathsf{k}_{i_{k}}\cosh (\eta \sum_{k=1}^{m}\mathsf{N}_{i_{k}}-\sum_{l=1}^{\mathsf{N}}h_{l}) \notag \\ & =\sum_{l=1}^{\mathsf{N}}\frac{T_{m}^{(K)}(\xi _{l}^{\left( h_{l}\right) })}{2\prod_{b\neq l,b=1}^{\mathsf{N}}(\xi _{l}^{\left( h_{l}\right) }/\xi _{b}^{\left( h_{b}\right) }-\xi _{b}^{\left( h_{b}\right) }/\xi _{l}^{\left( h_{l}\right) })\prod_{b=1}^{\mathsf{N}}\prod_{r=1}^{m-1}(\xi _{l}^{\left( h_{l}\right) }/\xi _{b}^{\left( -r\right) }-\xi _{b}^{\left( -r\right) }/\xi _{l}^{\left( h_{l}\right) })}.\end{aligned}$$The fusion equations allow to completely characterize all the $T_{m}^{(K)}(\lambda )$ in terms of $T_{1}^{(K)}(\lambda )$ by the following interpolation formulae:$$T_{m}^{(K)}(\lambda )=T_{m,\mathbf{h=0}}^{(K,\infty )}(\lambda )+\sum_{l=1}^{\mathsf{N}}f_{n,\mathbf{h=0}}^{\left( m\right) }(\lambda )T_{m-1}^{(K)}(\xi _{l}/q)T_{1}^{(K)}(\xi _{l}).$$ SoV covector basis generated by transfer matrix action ------------------------------------------------------ The following theorem holds as a corollary of Proposition 2.6 of [MaiN18]{}: Let $K$ be a $n\times n$ simple and diagonalizable symmetry of the $R$-matrix, then the following set:$$\langle h_{1},...,h_{\mathsf{N}}|\equiv \langle S|\prod_{n=1}^{\mathsf{N}}(T_{1}^{(K)}(\xi _{n}))^{h_{n}}\text{ \ for any }\{h_{1},...,h_{\mathsf{N}}\}\in \{0,...,n-1\}^{\otimes \mathsf{N}},$$forms a covector basis of $\mathcal{H}$, for almost any choice of $\langle S| $ and of the inhomogeneities satisfying $(\ref{Inhomog-cond})$. A proper choice of the state $\langle S|$ has the following tensor product form:$$\langle S|=\bigotimes_{l=1}^{\mathsf{N}}(x_{1},...,x_{n})_{l}\Gamma _{W}^{-1},\text{ \ \ }\Gamma _{W}=\bigotimes_{a=1}^{\mathsf{N}}W_{K,a}$$under the condition $\prod_{l=1}^{n}x_{l}\neq 0$, where $W$ is the invertible matrix defining the similarity transformation to the diagonal matrix $K_{J}$ by $K=W_{K}K_{J}W_{K}^{-1}$. Transfer matrix spectrum in our SoV approach: the $\mathcal{U}_{q}(\widehat{gl_{n}})$ case ========================================================================================== Discrete spectrum characterization by SoV ----------------------------------------- In the following we need $n-1$ Laurent polynomials in $\lambda $:$$t_{1}^{\left( K,\{x\},\{\nu_{i}\}\right) }(\lambda )=T_{m+1,\mathbf{h=0}}^{(K,\infty )}(\lambda |\nu_{1},...,\nu_{n})+\sum_{l=1}^{\mathsf{N}}f_{l,\mathbf{h=0}}^{\left( 1\right) }(\lambda )x_{l}, \label{Func-form-1}$$where we have used the notation $T_{m+1,\mathbf{h=0}}^{(K,\infty )}(\lambda |\nu_{1},...,\nu_{n})$ for the eigenvalue of the already defined asymptotic operator $T_{m+1,\mathbf{h=0}}^{(K,\infty )}(\lambda |{\mathsf{N}}_{1},...,{\mathsf{N}}_{n})$ on the common eigenspaces of the operators ${\mathsf{N}}_i$ and $$t_{m+1}^{(K,\{x\},\{\nu_{i}\})}(\lambda )=T_{m+1,\mathbf{h=0}}^{(K,\infty )}(\lambda |\nu_{1},...,\nu_{n})+\sum_{l=1}^{\mathsf{N}}f_{l,\mathbf{h=0}}^{\left( m+1\right) }(\lambda )t_{m}^{(K^{\left( a\right) },\{x\},\{\nu_{i}\})}(\xi _{l}/q)x_{l},$$for any $m\in \{1,...,n-2\}$, which are as well functions of a $n\times n$ twist matrix $K$, of a point $\{x_{1},...,x_{\mathsf{N}}\}\in \mathbb{C}^{\mathsf{N}}$ and of an $n$-upla $\{\nu_{1},...,\nu_{n}\}$ of nonnegative integers (the eigenvalues of the operators $\mathsf{N}_i$) satisfying:$$\sum_{l=1}^{n}\nu_{l}=\mathsf{N}.$$Then, the following characterization of the transfer matrix spectrum holds: \[ch-discrete-U\_q-n\] We consider a twist $K$ matrix symmetry which is diagonal and with simple spectrum and inhomogeneity parameters in generic position. Then the spectrum of $T_{1}^{(K)}(\lambda )$ is characterized by:$$\Sigma _{T^{(K)}}=\bigcup_{\forall \nu_{i}\geq 0\text{ : }\sum_{l=1}^{n}\nu_{l}=\mathsf{N}}\Sigma _{T^{(K)}}^{\left( \{\nu_{i}\}\right) },$$where$$\Sigma _{T^{(K)}}^{\left( \{\nu_{i}\}\right) }=\left\{ t_{1}(\lambda ):t_{1}(\lambda )=t_{1}^{\left( K,\{x\},\{\nu_{i}\}\right) }(\lambda ),\text{ \ }\forall \{x_{1},...,x_{\mathsf{N}}\}\in S_{T^{(K)}}^{\left( \{\nu_{i}\}\right) }\right\} , \label{SET-T-n}$$and $S_{T^{(K)}}^{\left( \{\nu_{i}\}\right) }$ is defined as the set of solutions to the next system of $\mathsf{N}$ polynomial equations of order $n $:$$x_{a}t_{n-1}^{\left( K,\{x\},\{\nu_{i}\}\right) }(\xi _{a}/q)\left. =\right. q\text{-det}M^{(K)}(\xi _{a}), \label{Ch-System-SoV-n}$$in $\mathsf{N}$ unknown $\{x_{1},...,x_{\mathsf{N}}\}$. Moreover, $T_{1}^{(K)}(\lambda ,\{\xi \})$ is diagonalizable with simple spectrum and$$\langle h_{1},...,h_{\mathsf{N}}|t\rangle =\prod_{n=1}^{\mathsf{N}}t_{1}^{h_{n}}(\xi _{n}) \label{SoV-Ch-T-eigenV-n}$$uniquely characterizes the eigenvector $|t\rangle $ associated to any fixed $t_{1}(\lambda )\in \Sigma _{T^{(K)}}$ in our SoV basis. The proof works by some simple modifications of the case of the Yangian $Y(gl_{n})$ fundamental representations developed in our second paper [MaiN18a]{}. We have just to handle the fact that the asymptotic behavior of the transfer matrices is now not central in the full representation space but only in each common eigenspaces of all the operators $\mathsf{N}_{i}$. Since those commute with all transfer matrices, the proof can be achieved in each of these subspaces and hence in the full Hilbert space. In appendix we present an alternative proof of our statement which is a corollary of the diagonalizabilty and simplicity of the transfer matrix spectrum which follows from the Proposition 2.7 of our paper [MaiN18]{}. Spectrum characterization by quantum spectral curve --------------------------------------------------- Let us first introduce the functions:$$\delta _{1}(\lambda )=\delta _{0}\prod_{a=1}^{\mathsf{N}}(\lambda q/\xi _{a}-\xi _{a}/(\lambda q)),\text{ \ }\delta _{m}(\lambda )=\prod_{i=0}^{m-1}\delta _{1}(\lambda /q^{i}),$$ then the discrete SoV characterization of the transfer matrix spectrum implies: Under the same conditions of the previous theorem, the entire functions $t_{1}(\lambda )$ is a $T_{1}^{\left( K\right) }(\lambda )$ transfer matrix eigenvalue belonging to[^5] $\Sigma _{T^{(K)}}^{\left( \{\nu_{i}\}\right) }$, with $\nu_{i}$ nonnegative integers satisfying:$$\sum_{l=1}^{n}\nu_{l}=\mathsf{N},$$iff there exists the unique Laurent polynomial:$$\varphi _{t}(\lambda )=\prod_{a=1}^{\mathsf{M}}(\lambda /\lambda _{a}-\lambda _{a}/\lambda )\text{\ \ with }\mathsf{M}\leq \mathsf{N}\text{ and }\lambda _{a}\neq \xi _{m}\text{ }\forall (a,m)\in \{1,...,\mathsf{M}\}\times \{1,...,\mathsf{N}\}, \label{Phi-form-n}$$such that $t_{1}(\lambda )$, $t_{m}(\lambda )\equiv t_{m}^{(K^{\left( a\right) },\{t_{1}(\xi _{1}),...,t_{1}(\xi _{\mathsf{N}})\},\{\nu_{i}\})}(\lambda )$ and $\varphi _{t}(\lambda )$ are solutions of the following quantum spectral curve functional equation:$$\sum_{l=0}^{n}\left( -1\right) ^{l}\delta _{l}(\lambda )\varphi _{t}(\lambda /q^{l})t_{n-l}(\lambda /q^{l})\left. =\right. 0 \label{Q-Spectral-curve-U(gl_n)}$$where $t_{0}(\lambda )=1$ and we have to fix[^6]:$$\delta _{0}=\mathsf{k}_{i}\text{ \ for one fixed }i\in \{1,...,n\},$$and$$\mathsf{M}=\mathsf{N}-\nu_{i}.$$Moreover, up to a normalization, the corresponding common transfer matrix eigenvector $|t\rangle $ admits the following separate representation:$$\langle h_{1},...,h_{\mathsf{N}}|t\rangle =\prod_{a=1}^{\mathsf{N}}\delta _{1}^{h_{a}}(\xi _{a})\varphi _{t}^{h_{a}}(\xi _{a}/q)\varphi _{t}^{n-h_{a}}(\xi _{a}).$$ Let us start proving that the asymptotics of the functional equation are indeed compatibles with those of the transfer matrix eigenvalues. That is, if we assume that $t_{1}(\lambda )\in \Sigma _{T^{(K)}}^{\left( \{n_{i}\}\right) }$, then we have to show that the leading asymptotics associated to the degree $\mathsf{M}+n\mathsf{N}$ of the l.h.s. of the equation must be zero and vice versa. Let us remark that from the known asymptotics $T_{m}^{(\pm \infty |K)}$ of the transfer matrices, the following identities hold:$$\lim_{\log \lambda \rightarrow \pm \infty }\lambda ^{\mp (n-a)\mathsf{N}}T_{n-a}^{(K,\infty )}(\lambda /q^{a}|\mathsf{N}_{1},...,\mathsf{N}_{n})=\frac{\left( \pm 1\right) ^{(n-a)\mathsf{N}}\sigma _{n-a}^{(n)}(\mathsf{k}_{1}q^{\pm \mathsf{N}_{1}},...,\mathsf{k}_{n}q^{\pm \mathsf{N}_{n}})}{q^{\pm (n-a)(a+(n-a-1)/2)\mathsf{N}}\prod_{n=1}^{\mathsf{N}}\xi _{n}^{\pm (n-a)}},$$while it is easy to verify that it holds:$$\lim_{\log \lambda \rightarrow \pm \infty }\lambda ^{\mp a\mathsf{N}}\delta _{a}(\lambda )=\frac{\left( \pm 1\right) ^{a\mathsf{N}}\mathsf{k}_{i}^{a}}{q^{\pm ((a-1)(a-2)/2-1)\mathsf{N}}\prod_{n=1}^{\mathsf{N}}\xi _{n}^{\pm a}},$$where we have imposed the choice $\delta _{0}=$ $\mathsf{k}_{i}$ and$$\lim_{\log \lambda \rightarrow \pm \infty }\lambda ^{\mp \mathsf{M}}\varphi _{t}(\lambda /q^{a})=\frac{\left( \pm 1\right) ^{\mathsf{M}}}{q^{\pm a\mathsf{M}}},$$from which it follows:$$\begin{aligned} \lim_{\log \lambda \rightarrow \pm \infty }\lambda ^{\mp (\mathsf{M}+n\mathsf{N})}(l.h.s.)_{\left( \ref{Q-Spectral-curve-U(gl_n)}\right) }& =\frac{\left( \pm 1\right) ^{(\mathsf{M}+n\mathsf{N})}\sum_{l=0}^{n}\left( -\mathsf{k}_{i}\right) ^{l}q^{\pm l(\mathsf{N}-\mathsf{M})}\sigma _{n-l}^{(n)}(\mathsf{k}_{1}q^{\pm \nu_{1}},...,\mathsf{k}_{n}q^{\pm \nu_{n}})}{q^{\pm n(n-1)\mathsf{N}/2}\prod_{n=1}^{\mathsf{N}}\xi _{n}^{\pm n}} \\ & =\frac{\left( \pm 1\right) ^{(\mathsf{M}+n\mathsf{N})}\sum_{l=0}^{n}\left( -\mathsf{k}_{i}q^{\pm \nu_{i}}\right) ^{l}\sigma _{n-l}^{(n)}(\mathsf{k}_{1}q^{\pm \nu_{1}},...,\mathsf{k}_{n}q^{\pm \nu_{n}})}{q^{\pm n(n-1)\mathsf{N}/2}\prod_{n=1}^{\mathsf{N}}\xi _{n}^{\pm n}} \\ & =0,\end{aligned}$$where according to our choice $\delta _{0}=\mathsf{k}_{i}$ we also fix $\mathsf{N}-\mathsf{M}=\nu_{i}$. The last identity to zero, holding for any choice of $i\in \{1,..,n\}$, as a trivial consequence of the defining identity of the symmetric polynomials:$$\sum_{l=0}^{n}\left( -\lambda \right) ^{l}\sigma _{n-l}^{(n)}(x_{1},...,x_{n})=\left( -1\right) ^{n}\prod_{a=1}^{n}(\lambda -x_{a}),$$which is zero if and only if $\lambda =x_{i}$ for any fixed $i\in \{1,..,n\}$. Vice versa, if $t_{1}(\lambda )$ satisfies with the polynomial $t_{m}(\lambda )$ and $\varphi _{t}(\lambda )$ the functional equation then it is a degree $\mathsf{N}$ Laurent polynomial in $\lambda $ with leading coefficients forced to be:$$t_{1}^{(\pm )}\equiv \lim_{\log \lambda \rightarrow \pm \infty }\lambda ^{\mp \mathsf{N}}t_{1}(\lambda )=\left( \pm 1\right) ^{\mathsf{N}}\frac{\sigma _{1}^{(n)}(\mathsf{k}_{1}q^{\pm n_{1}},...,\mathsf{k}_{n}q^{\pm n_{n}})}{\prod_{n=1}^{\mathsf{N}}\xi _{n}^{\pm 1}},$$as a consequence of the asymptotic of the satisfied functional equation for any $\delta _{0}=\mathsf{k}_{i}$ we also fix $\mathsf{N}-\mathsf{M}=\nu_{i}$. Now that the asymptotic behavior is verified the proof of the theorem follows mainly the same steps used for the Yangian $Y(gl_{n})$ case. For completeness let us reproduce them here. We complete first the proof of the fact that given a $t_{1}(\lambda )$ entire function satisfying with the polynomials $t_{m}(\lambda )$ and $\varphi _{t}(\lambda )$ the functional equation implies that it is a transfer matrix eigenvalue. Let us observe now that, for $\lambda =\xi _{a}$ it holds:$$\delta _{1+j}(\xi _{a})=0,\text{ for }1\leq j\leq n-1,\text{ }\delta _{1}(\xi _{a})\neq 0,\text{ det}_{q}M_{a}^{(K)}(\xi _{a})\neq 0,$$and the quantum spectral curve in these points reads:$$\frac{\delta _{1}(\xi _{a})\varphi _{t}(\xi _{a}/q)}{\varphi _{t}(\xi _{a})}=\frac{\text{det}_{q}M_{a}^{(K)}(\xi _{a})}{t_{n-1}(\xi _{a}/q)}. \label{Ch-1-Q-n}$$Consider instead $1\leq s\leq n-1$, then for $\lambda =\xi _{a}q^{s}$ it holds:$$\begin{aligned} \delta _{r\geq s+2}(\xi _{a}q^{s}) &=&0,\text{ }t_{n-b}(\xi _{a}q^{s-b})=0,\text{ for any }0\leq b\leq s-1 \\ \text{\ }\delta _{r\leq s+1}(\xi _{a}q^{s}) &\neq &0,\end{aligned}$$and the quantum spectral curve in these points reads:$$\frac{\delta _{s+1}(\xi _{a}q^{s})\varphi _{t}(\xi _{a}/q)}{\delta _{s}(\xi _{a}q^{s})\varphi _{t}(\xi _{a})}=\frac{t_{n-s}(\xi _{a})}{t_{n-s-1}(\xi _{a}/q)}. \label{Ch-2-Q-n}$$Then the chain of identities:$$\frac{\delta _{s+1}(\xi _{a}q^{s})}{\delta _{s}(\xi _{a}q^{s})}=\delta _{1}(\xi _{a})\text{ \ for any }1\leq s\leq n-1,$$imply the following ones:$$t_{m+1}(\xi _{a})=t_{m}(\xi _{a}/q)t_{1}(\xi _{a}),\text{ }\forall m\in \{1,...,n-1\},a\in \{1,...,\mathsf{N}\}.$$So that $t_{m}(\lambda )$ are eigenvalues of the transfer matrices $T_{m}^{\left( K\right) }(\lambda )$, for the same eigenvector $|t\rangle $, thanks to the SoV characterization given in our previous theorem. Let us now prove the reverse statement. Let $t_{1}(\lambda )$ be eigenvalue of the transfer matrix $T_{1}^{\left( K\right) }(\lambda )$ then we have to prove the existence of $\varphi _{t}(\lambda )$ a Laurent polynomial which satisfies the quantum spectral curve with the $t_{m}(\lambda )$. By imposing the following set of conditions:$$\delta _{1}(\xi _{a})\frac{\varphi _{t}(\xi _{a}/q)}{\varphi _{t}(\xi _{a})}=t_{1}(\xi _{a}), \label{Discrete-Ch-Q-n}$$we characterize uniquely a Laurent polynomial $\varphi _{t}(\lambda )$ of the form $\left( \ref{Phi-form}\right) $. Indeed, following the same steps given in the proof of the Theorem 4.1 of our second paper [@MaiN18a] for the Yangian $Y(gl_{n})$ case, we have that there exists a unique Laurent polynomial $\varphi _{t}(\lambda )$ of the form $\left( \ref{Phi-form}\right) $ with some degree $\mathsf{M}\leq \mathsf{N}$ so that one is left with the proof of the identity $\mathsf{M}=\mathsf{N}-\nu_{i}$. This is done just generalizing to the present case the argument based on the sum rules as presented in the proof of the Theorem 4.3 of our first paper [MaiN18]{}, see equations (4.69) and (4.70) there. Here, we recall that this characterization of $\varphi _{t}(\lambda )$ indeed implies that the functional equation is satisfied. The l.h.s. of the quantum spectral curve is a Laurent polynomial in $\lambda $ of maximal degree $n\mathsf{N}+\mathsf{M}\leq (n+1)\mathsf{N}$ then to prove that it is identically zero it is enough to show that it is zero in $(n+1)\mathsf{N}$ distinct points. Indeed, when this is the case the above argument on the sum rules shows that the leading coefficients of the quantum spectral curve are indeed zero. The chosen distinct points are the following $(n+1)\mathsf{N}$ ones $\xi _{a}q^{k_{a}}$, for any $a\in \{1,...,\mathsf{N}\}$ and $k_{a}\in \left\{ -1,0,...,n-1\right\} $. For $\lambda =\xi _{a}/q$ we have:$$\delta _{r}(\xi _{a}/q)=0\text{ \ for any }1\leq r\leq n,\text{ as well as det}M_{a}^{(K)}(\xi _{a}/q)=0,$$so that for any $a\in \{1,...,\mathsf{N}\}$ the quantum spectral curve equation is satisfied while in the remaining $n\mathsf{N}$ points this equation coincides with the $n\mathsf{N}$ equations ($\ref{Ch-1-Q-n}$) and ($\ref{Ch-2-Q-n}$) which are satisfied by the transfer matrix eigenvalues as they are all equivalent to the discrete characterization ([Discrete-Ch-Q-n]{}) thanks to the fusion equations. Let us verify now the equivalence of the SoV characterization of the transfer matrix eigenvector with the one presented in this theorem. It is enough to multiply by the non-zero product of the $\varphi _{t}^{n-1}(\xi _{a})$ over all the $a\in \{1,...,\mathsf{N}\}$ the eigenvector $|t\rangle $ getting our result: $$\prod_{a=1}^{\mathsf{N}}\varphi _{t}^{n-1}(\xi _{a})\prod_{a=1}^{\mathsf{N}}t_{1}^{h_{a}}(\xi _{a})\overset{\left( \ref{Discrete-Ch-Q-n}\right) }{=}\prod_{a=1}^{\mathsf{N}}\delta _{1}^{h_{a}}(\xi _{a})\varphi _{t}^{h_{a}}(\xi _{a}/q)\varphi _{t}^{n-1-h_{a}}(\xi _{a}).$$ $Q$-operator reconstruction by SoV ---------------------------------- The $Q$-operator commuting family, satisfying the quantum spectral curve equation with the transfer matrices at the operator level, can be constructed in terms of the fundamental transfer matrix thanks to the above SoV characterization of the transfer matrix spectrum. Let us denote by $\delta _{b}^{(i)}(\lambda )$ the polynomials defined in the previous section just making explicit that we have fixed $\delta _{0}=\mathsf{k}_{i}$ for some fixed $i\in \{1,...,n\}$. Moreover, let us define the following $\mathsf{N}\times \mathsf{N}$ operator matrix of elements:$$\lbrack C_{i,\xi _{\mathsf{N}+1}}^{(T_{1}^{\left( K\right) })}]_{ab}=-\delta _{ab}\,\frac{T_{1}^{\left( K\right) }(\xi _{a})}{\delta _{1}^{(i)}(\xi _{a})}+\prod_{\substack{ c=1 \\ c\neq b}}^{\mathsf{N}+1}\frac{\xi _{a}/(q\xi _{c})-(q\xi _{c})/\xi _{a}}{\xi _{b}/\xi _{c}-\xi _{c}/\xi _{b}}\qquad \forall a,b\in \{1,\ldots ,\mathsf{N}\},$$and the rank one central matrix:$$\lbrack \Delta _{\xi _{\mathsf{N}+1}}(\lambda )]_{ab}=\frac{\lambda /\xi _{\mathsf{N}+1}-\xi _{\mathsf{N}+1}/\lambda }{\lambda /\xi _{b}-\xi _{b}/\lambda }\frac{\prod_{c=1}^{\mathsf{N}}(\xi _{a}/(q\xi _{c})-(q\xi _{c})/\xi _{a})}{\prod_{c=1,c\neq b}^{\mathsf{N}+1}(\xi _{b}/\xi _{c}-\xi _{c}/\xi _{b})}\qquad \forall a,b\in \{1,\ldots ,\mathsf{N}\}, \label{Rank-one}$$then it holds: Given[^7] $\xi _{\mathsf{N}+1}\neq \xi _{i\leq N}$, then for almost any values of the parameters $\{\xi _{i\leq N}\}$ and of the nonzero eigenvalues $\{\mathsf{k}_{j\leq n}\}$ of the diagonal and simple spectrum twist matrix $K$, the following Laurent polynomial family of $Q$-operators:$$Q_{_{i}}(\lambda )=\frac{\text{det}_{N}[C_{i,\xi _{\mathsf{N}+1}}^{(T_{1}^{\left( K\right) })}+\Delta _{\xi _{\mathsf{N}+1}}(\lambda )]}{\text{det}_{N}[C_{i,\xi _{\mathsf{N}+1}}^{(T_{1}^{\left( K\right) })}]}\prod_{c=1}^{\mathsf{N}}\frac{\lambda /\xi _{c}-\xi _{c}/\lambda }{\xi _{\mathsf{N}+1}/\xi _{c}-\xi _{\mathsf{N}+1}/\xi _{c}},$$satisfies the operator quantum spectral curve equation$$\sum_{b=0}^{n}\delta _{b}^{(i)}(\lambda )Q_{i}(\lambda -b\eta )T_{n-b}^{\left( K\right) }(\lambda -b\eta )=0,$$where we have defined $T_{0}^{\left( K\right) }(\lambda )\equiv 1$, and moreover $Q_{_{i}}(\xi _{a})$ are invertible operators for any $a\in \{1,\ldots ,\mathsf{N}\}$. The SoV characterization of the transfer matrix spectrum and the proof of its reformulation in terms of the quantum spectral curve functional equation imply this corollary. Indeed, following the same proof given in the case of the fundamental representations of the Yangian $Y(gl_{n})$, see appendix B of our second paper [@MaiN18a], one can prove that the Laurent polynomial $\varphi _{t}(\lambda )$ of the form $\left( \ref{Phi-form}\right) $ solution of the quantum spectral curve equation has the following determinant representation:$$\varphi _{t}^{(i)}(\lambda )=\frac{\text{det}_{N}[C_{i,\xi _{\mathsf{N}+1}}^{(t_{1})}+\Delta _{\xi _{\mathsf{N}+1}}(\lambda )]}{\text{det}_{N}[C_{i,\xi _{\mathsf{N}+1}}^{(t_{1})}]}\prod_{c=1}^{\mathsf{N}}\frac{\lambda /\xi _{c}-\xi _{c}/\lambda }{\xi _{\mathsf{N}+1}/\xi _{c}-\xi _{\mathsf{N}+1}/\xi _{c}},$$obtained by substituting to the transfer matrix $T_{1}^{\left( K\right) }(\xi _{a})$ the corresponding eigenvalue $t_{1}(\xi _{a})$. As a consequence of the Proposition 2.7 of [@MaiN18], the transfer matrix $T_{1}^{\left( K\right) }(\lambda )$ is diagonalizable and with simple spectrum in our current setting. The Laurent polynomial family $Q_{_{i}}$-operator is then completely characterized by its action on the eigenbasis of the transfer matrix:$$Q_{_{i}}(\lambda )|t\rangle =|t\rangle \varphi _{t}^{(i)}(\lambda ),$$for any eigenvalue $t_{1}(\lambda )$ and uniquely associated eigenvector $|t\rangle $ of the transfer matrix $T_{1}^{\left( K\right) }(\lambda )$, which is equivalent to the characterization given in the corollary. This also imply that this operator family satisfies the quantum spectral curve equation with the transfer matrices. Acknowledgements {#acknowledgements .unnumbered} ================ J. M. M. and G. N. are supported by CNRS and ENS de Lyon. Appendix A ========== In this appendix we complete the proof of the Theorem \[Discrete-Ch-3\_q\] by computing the direct computation of the action of transfer matrices on the SoV covector basis. Once again let us comment that these computations are obtained by adapting those of the fundamental representations of the $Y(gl_{3})$ rational Yang-Baxter algebra in our first paper [@MaiN18], taking into account the fact that the transfer matrices commute with the operators $\mathsf{N}_i$ that define their (non-central) asymptotic behavior. hence, all computations can be done in each common eigenspaces of these operators that give a complete decomposition of the full Hilbert space. In order to complete the proof of Theorem \[Discrete-Ch-3\_q\], we have to prove that in the case $h_{j}=2$ the following identities hold $$\langle h_{1},...,h_{\mathsf{N}}|T_{1}^{(K^{\left( a\right) })}(\xi _{j}/q)|t\rangle =t_{1}(\xi _{j}/q)\langle h_{1},...,h_{a},...,h_{\mathsf{N}}|t\rangle ,$$and this is done by making an induction on the number $R$ of zeros contained in $\{h_{1},...,h_{\mathsf{N}}\}\in \{0,1,2\}^{\otimes \mathsf{N}}$. Let us start proving this identity for $R$ $=0$, the fusion identities imply:$$\langle h_{1},...,h_{a}=2,...,h_{\mathsf{N}}|T_{1}^{(K^{\left( a\right) })}(\xi _{j}/q)|t\rangle =\langle h_{1},...,h_{j}=1,...,h_{\mathsf{N}}|T_{2}^{(K^{\left( a\right) })}(\xi _{j})|t\rangle ,$$so that:$$\begin{aligned} \langle h_{1},...,h_{j}^{\prime } &=&1,...,h_{\mathsf{N}}|T_{2}^{(K^{\left( a\right) })}(\xi _{j})|t\rangle =\left. T_{2,\mathbf{h=1}}^{(K^{\left( a\right) },\infty )}(\xi _{j})\right\vert _{\mathsf{N}_{1}=l,\mathsf{N}_{2}=m}\langle h_{1},...,h_{j}^{\prime },...,h_{\mathsf{N}}|t\rangle \\ &&+\sum_{n=1}^{\mathsf{N}}f_{n,\mathbf{h}=\mathbf{1}}^{(a,2)}(\xi _{j})\langle h_{1},...,h_{j}^{\prime },...,h_{\mathsf{N}}|T_{2}^{\left( K\right) }(\xi _{n}/q)|t\rangle ,\end{aligned}$$thanks to the following interpolation formula:$$T_{2}^{\left( K\right) }(\xi _{j})=\left. T_{2,\mathbf{h=1}}^{(K^{\left( a\right) },\infty )}(\xi _{j})\right\vert _{\mathsf{N}_{1}=l,\mathsf{N}_{2}=m}+\sum_{n=1}^{\mathsf{N}}f_{n,\mathbf{h}=\mathbf{1}}^{(a,2)}(\xi _{j})T_{2}^{\left( K\right) }(\xi _{n}/q).$$Then, being $R=0$, it follows:$$\begin{aligned} \langle h_{1},...,h_{a}^{\prime },...,h_{\mathsf{N}}|T_{2}^{(K^{\left( a\right) })}(\xi _{j})|t\rangle & =\left. T_{2,\mathbf{h=1}}^{(K^{\left( a\right) },\infty )}(\xi _{j})\right\vert _{\mathsf{N}_{1}=l,\mathsf{N}_{2}=m}\langle h_{1},...,h_{j}^{\prime },...,h_{\mathsf{N}}|t\rangle \notag \\ & +\sum_{n=1}^{\mathsf{N}}q\text{-det}M^{(K^{\left( a\right) })}(\xi _{n})f_{n,\mathbf{h}=\mathbf{1}}^{(a,2)}(\xi _{j})\langle h_{1},...,h_{n}^{\prime \prime },...,h_{\mathsf{N}}|t\rangle ,\end{aligned}$$where $h_{n}^{\prime \prime }=h_{n}-1$ for $n\neq j$ and $h_{j}^{\prime \prime }=h_{j}^{\prime }-1=0$. Now the function:$$t_{2}(\lambda )=\left. T_{2,\mathbf{h=1}}^{(K^{\left( a\right) },\infty )}(\lambda )\right\vert _{\mathsf{N}_{1}=l,\mathsf{N}_{2}=m}+\sum_{n=1}^{\mathsf{N}}f_{n,\mathbf{h}=\mathbf{0}}^{(a,2)}(\lambda )t_{1}(\xi _{n}/q)t_{1}(\xi _{n}), \label{function t2-def}$$satisfies by definition the equations:$$\begin{aligned} t_{2}(\xi _{n}) &=&t_{1}(\xi _{n}/q)t_{1}(\xi _{n}),\text{ }\forall n\in \{1,...,\mathsf{N}\}, \\ t_{1}(\xi _{n})t_{2}(\xi _{n}/q) &=&q\text{-det}M^{(K^{\left( a\right) })}(\xi _{n}),\text{ }\forall n\in \{1,...,\mathsf{N}\},\end{aligned}$$where the the quantum determinant equation is indeed a consequence of the definition of $t_{1}(\lambda )$. Then we get:$$\begin{aligned} & \langle h_{1},...,h_{j}^{\prime },...,h_{\mathsf{N}}|T_{2}^{(K^{\left( a\right) })}(\xi _{j})|t\rangle \left. =\right. \notag \\ & =\left( \left. T_{2,\mathbf{h=1}}^{(K^{\left( a\right) },\infty )}(\xi _{j})\right\vert _{\mathsf{N}_{1}=l,\mathsf{N}_{2}=m}+\sum_{n=1}^{\mathsf{N}}t_{2}(\xi _{n}/q)f_{n,\mathbf{h}=\mathbf{1}}^{(a,2)}(\xi _{j})\right) \langle h_{1},...,h_{j}^{\prime },...,h_{\mathsf{N}}|t\rangle , \\ & =t_{2}(\xi _{n})\langle h_{1},...,h_{j}^{\prime },...,h_{\mathsf{N}}|t\rangle \\ & =t_{1}(\xi _{n}/q)\langle h_{1},...,h_{j}=2,...,h_{\mathsf{N}}|t\rangle ,\end{aligned}$$where we have used the interpolation formula:$$t_{2}(\xi _{j})=\left. T_{2,\mathbf{h=1}}^{(K^{\left( a\right) },\infty )}(\xi _{j})\right\vert _{\mathsf{N}_{1}=l,\mathsf{N}_{2}=m}+\sum_{n=1}^{\mathsf{N}}t_{2}(\xi _{n}/q)f_{n,\mathbf{h}=\mathbf{1}}^{(a,2)}(\xi _{j}),$$i.e. we have shown our identity $(\ref{Id-step2})$ for $R=0$. Then we can do our proof by induction; we assume that it holds for the generic $\{h_{1},...,h_{\mathsf{N}}\}\in \{0,1,2\}^{\otimes \mathsf{N}}$ containing $R-1$ zeros and we prove it for the generic $\{h_{1},...,h_{\mathsf{N}}\}\in \{0,1,2\}^{\otimes \mathsf{N}}$ containing $R$ zeros. Let us fix the generic $\{h_{1},...,h_{\mathsf{N}}\}\in \{0,1,2\}^{\otimes \mathsf{N}}$ with $h_{j}=2$ and let us denote with $\pi $ a permutation of $\{1,...,\mathsf{N}\} $ such that:$$\left. \begin{array}{l} h_{\pi (i)}=0,\text{ }\forall i\in \{1,...,R\}, \\ h_{\pi (i)}=1,\text{ }\forall i\in \{R+1,...,R+S\}, \\ h_{\pi (i)}=2,\text{ }\forall i\in \{R+S+1,...,\mathsf{N}\},\end{array}\right.$$with $j=\pi (R+S+1)$. Let us use now the following interpolation formula:$$T_{2}^{\left( K\right) }(\xi _{j})=\left. T_{2,\mathbf{k}}^{(K^{\left( a\right) },\infty )}(\xi _{j})\right\vert _{\mathsf{N}_{1}=l,\mathsf{N}_{2}=m}+\sum_{n=1}^{\mathsf{N}}f_{n,\mathbf{k}}^{(a,2)}(\xi _{j})T_{2}^{\left( K\right) }(\xi _{n}^{\left( k_{n}\right) }),$$where we have defined $\mathbf{k}$ by:$$\left. \begin{array}{l} k_{\pi (i)}=1,\text{ }\forall i\in \{1,...,R\}, \\ h_{\pi (i)}=2,\text{ }\forall i\in \{R+1,...,\mathsf{N}\},\end{array}\right.$$then it holds: $$\begin{aligned} \langle h_{1},...,h_{j}^{\prime }& =1,...,h_{\mathsf{N}}|T_{2}^{(K^{\left( a\right) })}(\xi _{j})|t\rangle =\left. T_{2,\mathbf{k}}^{(K^{\left( a\right) },\infty )}(\xi _{j})\right\vert _{\mathsf{N}_{1}=l,\mathsf{N}_{2}=m}\langle h_{1},...,h_{j}^{\prime },...,h_{\mathsf{N}}|t\rangle \notag \\ & +\sum_{n=1}^{R}f_{\pi (n),\mathbf{k}}^{(a,2)}(\xi _{j})\langle h_{1},...,h_{j}^{\prime },...,h_{\mathsf{N}}|T_{2}^{\left( K\right) }(\xi _{\pi (n)})|t\rangle \notag \\ & +\sum_{n=R+1}^{\mathsf{N}}f_{\pi (n),\mathbf{k}}^{(a,2)}(\xi _{j})\langle h_{1},...,h_{j}^{\prime },...,h_{\mathsf{N}}|T_{2}^{\left( K\right) }(\xi _{\pi (n)}/q)|t\rangle .\end{aligned}$$and which by the fusion identity reads:$$\begin{aligned} \langle h_{1},...,h_{j}^{\prime },...,h_{\mathsf{N}}|T_{2}^{(K^{\left( a\right) })}(\xi _{j})|t\rangle & =\left. T_{2,\mathbf{k}}^{(K^{\left( a\right) },\infty )}(\xi _{j})\right\vert _{\mathsf{N}_{1}=l,\mathsf{N}_{2}=m}\langle h_{1},...,h_{j}^{\prime },...,h_{\mathsf{N}}|t\rangle \notag \\ & +\sum_{n=1}^{R}f_{\pi (n),\mathbf{k}}^{(a,2)}(\xi _{j})\langle h_{1}^{(n)},...,h_{\mathsf{N}}^{(n)}|T_{1}^{\left( K\right) }(\xi _{\pi (n)}/q)|t\rangle \notag \\ & +\sum_{n=R+1}^{\mathsf{N}}q\text{-det}M^{(K^{\left( a\right) })}(\xi _{\pi (n)})f_{\pi (n),\mathbf{k}}^{(a,2)}(\xi _{j})\langle h_{1}^{(n)},...,h_{\mathsf{N}}^{(n)}|t\rangle , \label{T2-id-2}\end{aligned}$$where we have defined:$$h_{\pi (m)}^{(n)}=\left\{ \begin{array}{l} h_{\pi (m)}+\theta (R-m)\delta _{m,n}\text{ \ for }n\leq R \\ h_{\pi (m)}-\theta (m-(R+1))\delta _{m,n}-\delta _{m,R+S+1}\text{ \ for }R+1\leq n\end{array}\right. .$$To compute $\langle h_{1}^{(n)},...,h_{\mathsf{N}}^{(n)}|T_{1}^{\left( K\right) }(\xi _{\pi (n)}/q)|t\rangle $ for $n\leq R$, we use the following interpolation formula:$$T_{1}^{\left( K\right) }(\xi _{\pi (n)}/q)=\left. T_{1,\mathbf{k}^{\prime }}^{(K^{\left( a\right) },\infty )}(\xi _{\pi (n)}/q)\right\vert _{\mathsf{N}_{1}=l,\mathsf{N}_{2}=m}+\sum_{r=1}^{\mathsf{N}}f_{r,\mathbf{k}^{\prime }}^{(a,1)}(\xi _{\pi (n)}/q)T_{1}^{\left( K\right) }(\xi _{r}^{(k_{r}^{\prime })}),$$where we have defined:$$k_{\pi (m)}^{\prime }=\left\{ \begin{array}{l} 0\text{ \ for }m\leq R+S+1 \\ 1\text{ \ for }R+S+2\leq m\end{array}\right. ,$$which gives:$$\begin{aligned} \langle h_{1}^{(n)},...,h_{\mathsf{N}}^{(n)}|T_{1}^{\left( K\right) }(\xi _{\pi (n)}/q)|t\rangle & =\left. T_{1,\mathbf{k}^{\prime }}^{(K^{\left( a\right) },\infty )}(\xi _{\pi (n)}/q)\right\vert _{\mathsf{N}_{1}=l,\mathsf{N}_{2}=m}\langle h_{1}^{(n)},...,h_{\mathsf{N}}^{(n)}|t\rangle \notag \\ & +\sum_{r=1}^{R+S+1}f_{\pi (r),\mathbf{k}^{\prime }}^{(a,1)}(\xi _{\pi (n)}/q)\langle h_{1}^{(n)},...,h_{\mathsf{N}}^{(n)}|T_{1}^{\left( K\right) }(\xi _{\pi (r)})|t\rangle \notag \\ & +\sum_{r=R+S+2}^{\mathsf{N}}f_{\pi (r),\mathbf{k}^{\prime }}^{(a,1)}(\xi _{\pi (n)}/q)\langle h_{1}^{(n)},...,h_{\mathsf{N}}^{(n)}|T_{1}^{\left( K\right) }(\xi _{\pi (r)}/q)|t\rangle ,\end{aligned}$$which becomes:$$\begin{aligned} \langle h_{1}^{(n)},...,h_{\mathsf{N}}^{(n)}|T_{1}^{\left( K\right) }(\xi _{\pi (n)}/q)|t\rangle & =\left. T_{1,\mathbf{k}^{\prime }}^{(K^{\left( a\right) },\infty )}(\xi _{\pi (n)}/q)\right\vert _{\mathsf{N}_{1}=l,\mathsf{N}_{2}=m}\langle h_{1}^{(n)},...,h_{\mathsf{N}}^{(n)}|t\rangle \notag \\ & +\sum_{r=1}^{R+S+1}f_{\pi (r),\mathbf{k}^{\prime }}^{(a,1)}(\xi _{\pi (n)}/q)t_{1}(\xi _{\pi (r)})\langle h_{1}^{(n)},...,h_{\mathsf{N}}^{(n)}|t\rangle \notag \\ & +\sum_{r=R+S+2}^{\mathsf{N}}f_{\pi (r),\mathbf{k}^{\prime }}^{(a,1)}(\xi _{\pi (n)}/q)t_{1}(\xi _{\pi (r)}/q)\langle h_{1}^{(n)},...,h_{\mathsf{N}}^{(n)}|t\rangle ,\end{aligned}$$where in the second line we have used the identity $(\ref{Id-step1})$ while in the third line the identity $(\ref{Id-step2})$, which holds by assumption being $R-1$ the number of zeros in $\{h_{1}^{(n)},...,h_{\mathsf{N}}^{(n)}\}$. So that we have shown for any $n\leq R$:$$\langle h_{1}^{(n)},...,h_{\mathsf{N}}^{(n)}|T_{1}^{\left( K\right) }(\xi _{\pi (n)}/q)|t\rangle =t_{1}(\xi _{\pi (n)}/q)\langle h_{1}^{(n)},...,h_{\mathsf{N}}^{(n)}|t\rangle ,$$and substituting it in the second line of (\[T2-id-2\]), we get:$$\begin{aligned} \langle h_{1},...,h_{j}^{\prime },...,h_{\mathsf{N}}|T_{2}^{(K^{\left( a\right) })}(\xi _{j})|t\rangle & =\left. T_{2,\mathbf{k}}^{(K^{\left( a\right) },\infty )}(\xi _{j})\right\vert _{\mathsf{N}_{1}=l,\mathsf{N}_{2}=m}\langle h_{1},...,h_{j}^{\prime },...,h_{\mathsf{N}}|t\rangle \notag \\ & +\sum_{n=1}^{R}t_{1}(\xi _{\pi (n)}/q)f_{\pi (n),\mathbf{k}}^{(a,2)}(\xi _{j})\langle h_{1}^{(n)},...,h_{\mathsf{N}}^{(n)}|t\rangle \notag \\ & +\sum_{n=R+1}^{\mathsf{N}}q\text{-det}M^{(K^{\left( a\right) })}(\xi _{\pi (n)})f_{\pi (n),\mathbf{k}}^{(a,2)}(\xi _{j})\langle h_{1}^{(n)},...,h_{\mathsf{N}}^{(n)}|t\rangle ,\end{aligned}$$and so $\langle h_{1},...,h_{j}^{\prime },...,h_{\mathsf{N}}|T_{2}^{(K^{\left( a\right) })}(\xi _{j})|t\rangle $ reads:$$\begin{aligned} & \left( \left. T_{2,\mathbf{k}}^{(K^{\left( a\right) },\infty )}(\xi _{j})\right\vert _{\mathsf{N}_{1}=l,\mathsf{N}_{2}=m}+\sum_{n=1}^{R}t_{1}(\xi _{\pi (n)})t_{1}(\xi _{\pi (n)}/q)f_{\pi (n),\mathbf{k}}^{(a,2)}(\xi _{j})+\sum_{n=R+1}^{\mathsf{N}}t_{2}(\xi _{\pi (n)}/q)f_{\pi (n),\mathbf{k}}^{(a,2)}(\xi _{j})\right) \notag \\ & \times \langle h_{1},...,h_{j}^{\prime },...,h_{\mathsf{N}}|t\rangle \notag \\ & =t_{2}(\xi _{j}/q)\langle h_{1},...,h_{j}^{\prime }=1,...,h_{\mathsf{N}}|t\rangle =t_{1}(\xi _{j}/q)\langle h_{1},...,h_{j}=2,...,h_{\mathsf{N}}|t\rangle ,\end{aligned}$$i.e. we have proven our formula $(\ref{Id-step2})$. Finally, taken the generic $\{h_{1},...,h_{\mathsf{N}}\}\in \{0,1,2\}^{\otimes \mathsf{N}}$ with:$$\left. \begin{array}{l} h_{\pi (i)}=0,\text{ }\forall i\in \{1,...,R\}, \\ h_{\pi (i)}=1,\text{ }\forall i\in \{R+1,...,R+S\}, \\ h_{\pi (i)}=2,\text{ }\forall i\in \{R+S+1,...,\mathsf{N}\},\end{array}\right.$$and by using the interpolation formula:$$T_{1}^{\left( K\right) }(\lambda )=\left. T_{1,\mathbf{k}}^{(K^{\left( a\right) },\infty )}(\lambda )\right\vert _{\mathsf{N}_{1}=l,\mathsf{N}_{2}=m}+\sum_{n=1}^{\mathsf{N}}f_{n,\mathbf{p}}^{(a,1)}(\lambda )T_{1}^{\left( K\right) }(\xi _{n}^{\left( p_{n}\right) }),$$where we have defined $\mathbf{p}$ by:$$\left. \begin{array}{l} p_{\pi (i)}=0,\text{ }\forall i\in \{1,...,R+S\}, \\ p_{\pi (i)}=1,\text{ }\forall i\in \{R+S+1,...,\mathsf{N}\},\end{array}\right.$$then it holds: $$\begin{aligned} \langle h_{1},...,h_{\mathsf{N}}|T_{1}^{\left( K\right) }(\lambda )|t\rangle & =\left. T_{1,\mathbf{k}}^{(K^{\left( a\right) },\infty )}(\lambda )\right\vert _{\mathsf{N}_{1}=l,\mathsf{N}_{2}=m}\langle h_{1},...,h_{\mathsf{N}}|t\rangle \notag \\ & +\sum_{n=1}^{R}f_{\pi (n),\mathbf{p}}^{(a,1)}(\lambda )\langle h_{1},...,h_{\mathsf{N}}|T_{1}^{\left( K\right) }(\xi _{\pi (n)})|t\rangle \notag \\ & +\sum_{n=R+1}^{\mathsf{N}}f_{\pi (n),\mathbf{p}}^{(a,1)}(\lambda )\langle h_{1},...,h_{\mathsf{N}}|T_{1}^{\left( K\right) }(\xi _{\pi (n)}/q)|t\rangle\end{aligned}$$then by using in the second line the identity $(\ref{Id-step1})$ and $(\ref{Id-step2})$ in the third line we get:$$\begin{aligned} \langle h_{1},...,h_{\mathsf{N}}|T_{1}^{\left( K\right) }(\lambda )|t\rangle & =\left( \left. T_{1,\mathbf{k}}^{(K^{\left( a\right) },\infty )}(\lambda )\right\vert _{\mathsf{N}_{1}=l,\mathsf{N}_{2}=m}+\sum_{n=1}^{\mathsf{N}}f_{\pi (n),\mathbf{p}}^{(a,1)}(\lambda )t_{1}(\xi _{\pi (n)}^{(p_{\pi (n)})})\right) \langle h_{1},...,h_{\mathsf{N}}|t\rangle \\ & =t_{1}(\lambda )\langle h_{1},...,h_{\mathsf{N}}|t\rangle,\end{aligned}$$which complete the proof of our theorem. Appendix B ========== In this appendix, we provide a proof of the discrete SoV characterization of the transfer matrix spectrum given in Theorem \[ch-discrete-U\_q-n\] bypassing the computation of the transfer matrix action in the SoV basis. The proof is presented bellow in the case of the rational fundamental representations of $Y(\widehat{gl_{n}})$. Then one can either use the argument that the fundamental evaluation representations of $\mathcal{U}_{q}(\widehat{gl_{n}})$ lead under the rational limit to the rational ones, so inferring that the same result has to hold for the trigonometric case too for almost any values of the parameters. Otherwise one can just repeat the same type of proof directly in the trigonometric case, only taking into account that the asymptotic behavior for the trigonometric case are not central in the full representation space but only in the common eigenspaces of the operators $\mathsf{N}_{i}$. Moreover, the case of non-fundamental representation can be handled similarly. In fact, the proof of the Theorem 2.3 of our third paper [MaiN18b]{} can be seen as the first step in the proof by induction for these non-fundamental representations. For the fundamental representations of $Y(gl_{n})$, the quantum separation of variable characterization of the first transfer matrix spectrum reads$$\Sigma _{T^{(K)}}=\left\{ t_{1}(\lambda ):t_{1}(\lambda )=t_{1}^{\left( K,\{x\}\right) }(\lambda ),\text{ \ }\forall \{x_{1},...,x_{\mathsf{N}}\}\in S_{T}\right\} ,$$in terms of the solutions to the following system $S_{T}$ of $\mathsf{N}$ polynomial equations of degree $n$:$$x_{a}t_{n-1}^{\left( K,\{x\}\right) }(\xi _{a}-\eta )\left. =\text{det}K\right. q\text{-det}M^{(I)}(\xi _{a}),$$in $\mathsf{N}$ unknown $\{x_{1},...,x_{\mathsf{N}}\}$, where we recall the definitions used in our second paper [@MaiN18]: $$t_{1}^{\left( K,\{x\}\right) }(\lambda )=\text{tr\thinspace }K\text{ }\prod_{a=1}^{\mathsf{N}}(\lambda -\xi _{a})+\sum_{a=1}^{\mathsf{N}}g_{a,\mathbf{h}=0}^{\left( 1\right) }(\lambda )x_{a}, \label{t1-form}$$and:$$t_{m+1}^{\left( K,\{x\}\right) }(\lambda )=\prod_{b=1}^{\mathsf{N}}\prod_{r=1}^{m}(\lambda -\xi _{b}-r\eta )\left[ T_{m+1,\mathbf{h}=\mathbf{0}}^{(K,\infty )}(\lambda )+\sum_{a=1}^{\mathsf{N}}g_{a,\mathbf{h}=\mathbf{0}}^{\left( m+1\right) }(\lambda )x_{a}t_{m}^{\left( K,\{x\}\right) }(\xi _{a}-\eta )\right] , \label{Rec-Func-form-m}$$for any $m\in \{1,...,n-2\}$, and$$\begin{aligned} T_{m,\mathbf{h}}^{(K,\infty )}(\lambda ) &=&\text{tr}_{1,...,m}\left[ P_{1,...,m}^{-}K_{1}K_{2}\cdots K_{m}\right] \prod_{b=1}^{\mathsf{N}}(\lambda -\xi _{b}^{(h_{n})}), \\ g_{a,\mathbf{h}}^{\left( m\right) }(\lambda ) &=&\prod_{b\neq a,b=1}^{\mathsf{N}}\frac{\lambda -\xi _{b}^{(h_{b})}}{\xi _{a}^{(h_{a})}-\xi _{b}^{(h_{b})}}\prod_{b=1}^{\mathsf{N}}\prod_{r=1}^{m-1}\frac{1}{\xi _{a}^{(h_{a})}-\xi _{b}^{(-r)}}.\end{aligned}$$Here we are interested in giving a proof of this characterization bypassing the computation of the action of the first transfer matrix in the SoV basis. The fact that any eigenvalue defines a solutions of this system follows from the fusion relations. So the only nontrivial thing to show is that indeed any solution of the above system defines one eigenvalue. The Theorem of Bezout[^8] states that the above system of polynomial equations admits $n^{\mathsf{N}}$ solutions if the $\mathsf{N}$ polynomials, defining the system, have no common components[^9]. The transfer matrix, being diagonalizable and with simple spectrum, has exactly $n^{\mathsf{N}}$ distinct eigenvalues and so, under the condition of no common components, there are indeed exactly $n^{\mathsf{N}}$ distinct solutions to the above system and each one is uniquely associated to a transfer matrix eigenvalue. We have to show now that the condition of no common components indeed holds for almost any values of the parameters. The proof of this statement can be done by induction on $n-1$ the rank of the Yang-Baxter algebra. Let us start with the rank 1 case, i.e. $n=2$ and fundamental representations of $Y(gl_{2})$. Here, we fix the eigenvalue of the twist matrix to be $\mathsf{k}_{1}\neq 0$ and $\mathsf{k}_{2}=0$, then the system of equations reads:$$t_{1}^{\left( K,\{x\}\right) }(\xi _{a})t_{1}^{\left( K,\{x\}\right) }(\xi _{a}-\eta )=x_{a}t_{1}^{\left( K,\{x\}\right) }(\xi _{a}-\eta )\left. =\text{det}K\right. q\text{-det}M^{(I)}(\xi _{a})=0, \label{Sys-n=2}$$now taking into account that by definition $t_{1}^{\left( K,\{x\}\right) }(\lambda )$ is a degree $\mathsf{N}$ polynomial in $\lambda $ and that it holds:$$\xi _{a}^{(h)}\neq \xi _{b}^{(k)}\text{ \ \ }\forall h,k\in \{0,1\},a\neq b\in \{1,...,\mathsf{N}\},$$then a solution to the system can be obtained iff for any $a\in \{1,...,\mathsf{N}\}$ there exists a unique $h_{a}\in \{0,1\}$ such that $t_{1}^{\left( K,\{x\}\right) }(\xi _{a}^{(h_{a})})=0$, or equivalently:$$t_{1,\mathbf{h}}^{\left( K,\{x\}\right) }(\lambda )=\mathsf{k}_{1}\prod_{a=1}^{\mathsf{N}}(\lambda -\xi _{a}^{(h_{a})}).$$So we have that the system has exactly $2^{\mathsf{N}}$ distinct solutions associated to the $2^{\mathsf{N}}$ distinct $\mathsf{N}$-upla $\mathbf{h}=\{h_{1\leq n\leq \mathsf{N}}\}$ in $\bigotimes_{n=1}^{_{\mathsf{N}}}\{0,1\}$. So there are no common components for $\mathsf{k}_{1}\neq 0$ and $\mathsf{k}_{2}=0$, and being the polynomials defining the system (\[Sys-n=2\]) also polynomial in twist matrix eigenvalues we infer that this statement is true for almost any choice of $\mathsf{k}_{1}$ and $\mathsf{k}_{2}$. So we have proven our statement for $n=2$. Let us now prove it for $n=3$, we fix here the twist matrix eigenvalues as it follows $\mathsf{k}_{1}\neq 0$, $\mathsf{k}_{2}\neq 0$, $\mathsf{k}_{2}\neq \mathsf{k}_{1}$ and $\mathsf{k}_{3}=0$, then the system of equations reads:$$x_{a}t_{2}^{\left( K,\{x\}\right) }(\xi _{a}-\eta )=\text{det}K\text{ }q\text{-det}M^{(I)}(\xi _{a})=0, \label{Sys-n=3}$$where by definition it holds$$x_{a}t_{1}^{\left( K,\{x\}\right) }(\xi _{a}-\eta )=t_{2}^{\left( K,\{x\}\right) }(\xi _{a}), \label{Fusion-1-1}$$so that it holds too$$t_{2}^{\left( K,\{x\}\right) }(\xi _{a})t_{2}^{\left( K,\{x\}\right) }(\xi _{a}-\eta )=0, \label{Compl-Sys-n=3}$$now taking into account that by definition $t_{2}^{\left( K,\{x\}\right) }(\lambda )$ is a degree $2\mathsf{N}$ polynomial in $\lambda $, zero in the points $\xi _{a}+\eta $ for any $a\in \{1,...,\mathsf{N}\}$, it follows that a solution to the system (\[Compl-Sys-n=3\]) can be obtained iff for any $a\in \{1,...,\mathsf{N}\}$ there exists a unique $h_{a}\in \{0,1\}$ such that $t_{2}^{\left( K,\{x\}\right) }(\xi _{a}^{(h_{a})})=0$, or equivalently:$$t_{2,\mathbf{h}}^{\left( K,\{x\}\right) }(\lambda )=\mathsf{k}_{1}\mathsf{k}_{2}\prod_{a=1}^{\mathsf{N}}(\lambda -\xi _{a}-\eta )(\lambda -\xi _{a}^{(h_{a})}).$$So that the system (\[Compl-Sys-n=3\]) has exactly $2^{\mathsf{N}}$ distinct solutions associated to the $2^{\mathsf{N}}$ distinct $\mathsf{N}$-upla $\mathbf{h}=\{h_{1\leq n\leq \mathsf{N}}\}$ in $\bigotimes_{n=1}^{_{\mathsf{N}}}\{0,1\}$. Now for any fixed $\mathbf{h}\in \bigotimes_{n=1}^{_{\mathsf{N}}}\{0,1\}$ we can define a permutation $\pi _{\mathbf{h}}\in S_{\mathsf{N}}$ and a nonnegative integer $m_{\mathbf{h}}\leq \mathsf{N}$ such that:$$h_{\pi _{\mathbf{h}}(a)}=0\text{ \ }\forall a\in \{1,...,m_{\mathbf{h}}\}\text{ \ and \ }h_{\pi _{\mathbf{h}}(a)}=1\text{ \ }\forall a\in \{m_{\mathbf{h}}+1,...,\mathsf{N}\}\text{.}$$It is easy to remark now that fixed $\mathbf{h}\in \bigotimes_{n=1}^{_{\mathsf{N}}}\{0,1\}$ then (\[Fusion-1-1\]), for $a\in \{1,...,m_{\mathbf{h}}\}$, and (\[Sys-n=3\]) are satisfied iff it holds:$$x_{\pi _{\mathbf{h}}(a)}=t_{1}^{\left( K,\{x\}\right) }(\xi _{\pi _{\mathbf{h}}(a)})=0\text{ \ }\forall a\in \{1,...,m_{\mathbf{h}}\}. \label{Induced-zeros}$$Indeed, if this is not the case for a given $b\in \{1,...,m_{\mathbf{h}}\}$, then the (\[Sys-n=3\]) implies $t_{2,\mathbf{h}}^{\left( K,\{x\}\right) }(\xi _{h_{\pi _{\mathbf{h}}(b)}}-\eta )=0$ which is not compatible with our choice of $t_{2,\mathbf{h}}^{\left( K,\{x\}\right) }(\lambda )$. So, for any fixed $\mathbf{h}\in \bigotimes_{n=1}^{_{\mathsf{N}}}\{0,1\}$, we are left with the requirement to satisfy the fusion equation (\[Fusion-1-1\]) for $a\in \{m_{\mathbf{h}}+1,...,\mathsf{N}\}$ which results in the following system of equation:$$t_{1}^{\left( K,\{x\},\mathbf{h}\right) }(\xi _{\pi _{\mathbf{h}}(a)})t_{1}^{\left( K,\{x\},\mathbf{h}\right) }(\xi _{\pi _{\mathbf{h}}(a)}-\eta )=t_{2,\mathbf{h}}^{\left( K,\{x\}\right) }(\xi _{\pi _{\mathbf{h}}(a)}),\text{ }\forall a\in \{m_{\mathbf{h}}+1,...,\mathsf{N}\},$$where $t_{1}^{\left( K,\{x\},\mathbf{h}\right) }(\lambda )$ is a degree $\mathsf{N}$ polynomial in $\lambda $ of the form (\[t1-form\]) with the $m_{\mathbf{h}}$ zeros given by (\[t1-form\]). Then let us define the following degree $\mathsf{N}-m_{\mathbf{h}}$ polynomial in $\lambda $:$$\bar{t}_{1}^{\left( K,\{x\},\mathbf{h}\right) }(\lambda )=t_{1}^{\left( K,\{x\},\mathbf{h}\right) }(\lambda )/\prod_{a=1}^{m_{\mathbf{h}}}(\lambda -\xi _{\pi _{\mathbf{h}}(a)}),$$and the degree $2(\mathsf{N}-m_{\mathbf{h}})$ polynomial in $\lambda $:$$\begin{aligned} \bar{t}_{2,\mathbf{h}}^{\left( K,\{x\}\right) }(\lambda ) &=&t_{2,,\mathbf{h}}^{\left( K,\{x\}\right) }(\lambda )/\prod_{a=1}^{m_{\mathbf{h}}}\left[ (\lambda -\xi _{\pi _{\mathbf{h}}(a)})(\lambda -\xi _{\pi _{\mathbf{h}}(a)}-\eta )\right] \\ &=&\mathsf{k}_{1}\mathsf{k}_{2}\prod_{a=1+m_{\mathbf{h}}}^{\mathsf{N}}(\lambda -\xi _{\pi _{\mathbf{h}}(a)}-\eta )(\lambda -\xi _{\pi _{\mathbf{h}}(a)}+\eta ),\end{aligned}$$ the previous system of equations simplifies to:$$\bar{t}_{1}^{\left( K,\{x\},\mathbf{h}\right) }(\xi _{\pi _{\mathbf{h}}(a)})\bar{t}_{1}^{\left( K,\{x\},\mathbf{h}\right) }(\xi _{\pi _{\mathbf{h}}(a)}-\eta )=\bar{t}_{2,\mathbf{h}}^{\left( K,\{x\}\right) }(\xi _{\pi _{\mathbf{h}}(a)}),\text{ }\forall a\in \{m_{\mathbf{h}}+1,...,\mathsf{N}\}.$$Such a system coincides with the system associated to the case $n=2$ for a lattice with $\mathsf{N}-m_{\mathbf{h}}$ sites and inhomogeneities $\xi _{\pi _{\mathbf{h}}(a)}$ with $a\in \{m_{\mathbf{h}}+1,...,\mathsf{N}\}$. Indeed, $\bar{t}_{2,\mathbf{h}}^{\left( K,\{x\}\right) }(\lambda )$ is just the quantum determinant for such a lattice associated to the $2\times 2$ twist matrix with distinct non-zero eigenvalues $\mathsf{k}_{1}\neq 0$ and $\mathsf{k}_{2}\neq 0$ and $\bar{t}_{1}^{\left( K,\{x\},\mathbf{h}\right) }(\lambda )$ has the functional form of a transfer matrix eigenvalue with asymptotic given by the trace of this twist matrix. Now, we can use our result for $n=2$ to state that this system has exactly 2$^{\mathsf{N}-m_{\mathbf{h}}}$ distinct solutions, which allows to count the full set of solutions to our original system:$$\sum_{m=0}^{\mathsf{N}}2^{\mathsf{N}-m}\left( \begin{array}{c} \mathsf{N} \\ m\end{array}\right) =3^{\mathsf{N}},$$ where we have used that for any fixed $m\in \{1,...,\mathsf{N}\}$ the number of $\mathbf{h}\in \bigotimes_{n=1}^{_{\mathsf{N}}}\{0,1\}$ such that $m_{\mathbf{h}}=m$ is exactly given by the binomial symbol:$$\label{binomial} \left( \begin{array}{c} \mathsf{N} \\ m\end{array}\right) =\frac{\mathsf{N}!}{(\mathsf{N}-m)!m!}.$$So we proved that the system has exactly $3^{\mathsf{N}}$ distinct solutions and no common components for $\mathsf{k}_{1}\neq 0$, $\mathsf{k}_{2}\neq 0$ and $\mathsf{k}_{3}=0$, and being the polynomials defining the system ([Sys-n=3]{}) also polynomial in twist matrix eigenvalues we can infer that this statement is true for almost any choice of three distinct eigenvalues $\mathsf{k}_{1}$, $\mathsf{k}_{2}$ and $\mathsf{k}_{3}$. So we have proven our statement for $n=3$. At this point it is easy to understand how to implement the proof by induction, i.e. we assume that the statement is proven for the rank $n-1$ case and we prove it for the rank $n$ case and this is done in the case of a diagonalizable and simple spectrum $(n+1)\times (n+1)$ twist matrix with pairwise distinct eigenvalues $\mathsf{k}_{a}\neq 0$, for any $a\in \{1,...,n\}$, and $\mathsf{k}_{n+1}=0$. Then following similar steps to those illustrated above, we see that the function $t_{n}^{\left( K,\{x\}\right) }(\lambda )$ is forced to take the form$$t_{n,\mathbf{h}}^{\left( K,\{x\}\right) }(\lambda )=\prod_{a=1}^{\mathsf{N}}\mathsf{k}_{a}(\lambda -\xi _{a}^{(h_{a})})\prod_{r=1}^{n-1}(\lambda -\xi _{a}-r\eta ),$$associated to the $2^{\mathsf{N}}$ distinct $\mathsf{N}$-upla $\mathbf{h}=\{h_{1\leq n\leq \mathsf{N}}\}$ in $\bigotimes_{n=1}^{_{\mathsf{N}}}\{0,1\}$ and that for any fixed $\mathbf{h}$ the system is reduced to that associated to the case of rank $n-1$ with general diagonalizable and simple spectrum $n\times n$ twist matrix with eigenvalues $\mathsf{k}_{a}\neq 0$, for any $a\in \{1,...,n\}$, on a number of site $\mathsf{N}-m_{\mathbf{h}}$. Then using the induction we know that this system admits $n^{\mathsf{N}-m_{\mathbf{h}}}$ distinct solutions for any such $\mathbf{h}$ and that for any fixed $m\in \{1,...,\mathsf{N}\}$ the number of $\mathbf{h}\in \bigotimes_{n=1}^{_{\mathsf{N}}}\{0,1\}$ such that $m_{\mathbf{h}}=m$ is exactly given by the binomial symbol (\[binomial\]). So that the total counting gives:$$\sum_{m=0}^{\mathsf{N}}n^{\mathsf{N}-m}\left( \begin{array}{c} \mathsf{N} \\ m\end{array}\right) =(n+1)^{\mathsf{N}},$$which proves the no common component statement for the rank $n$ case too when we repeat the polynomiality argument of the dependence w.r.t. the twist matrix eigenvalues. [99]{} J. M. Maillet, and G. Niccoli. On quantum separation of variables (2018) 091417. J. M. Maillet, and G. Niccoli. Complete spectrum of quantum integrable lattice models associated to $Y(gl_n)$ by separation of variables . J. M. Maillet, and G. Niccoli. On quantum separation of variables: beyond fundamental representations, to appear. L. D. Faddeev and E. K. Sklyanin. Quantum-mechanical approach to completely integrable field theory models. , 23:902–904, 1978. L. D. Faddeev, E. K. Sklyanin, and L. A. Takhtajan. Quantum inverse problem method [I]{}. , 40:688–706, 1979. Translated from Teor. Mat. Fiz. 40 (1979) 194-220. L. A. Takhtadzhan and L. D. Faddeev. The quantum method of the inverse problem and the [H]{}eisenberg [XYZ]{} model. *Russ. Math. Surveys*, 34(5):11–68, 1979. E. K. Sklyanin. Method of the inverse scattering problem and the non-linear quantum [S]{}chr[ö]{}dinger equation. *Sov. Phys. Dokl.*, 24:107–109, 1979. E. K. Sklyanin. On complete integrability of the [L]{}andau-[L]{}ifshitz equation. Preprint LOMI E-3-79, 1979. L. D. Faddeev and L. A. Takhtajan. Quantum inverse scattering method. , C 1:107, 1981. E. K. Sklyanin. Quantum version of the inverse scattering problem method. , 19:1546–1595, 1982. L. D. Faddeev. Integrable models in $(1 + 1)$-dimensional quantum field theory. In J. B. Zuber and R. Stora, editors, *Les Houches 1982, Recent advances in field theory and statistical mechanics*, pages 561–608. Elsevier Science Publ., 1984. L. D. Faddeev. How algebraic [B]{}ethe ansatz works for integrable model. Les Houches Lectures, 1996. O. Babelon, H.J. de Vega and C.M. Viallet, Solutions of the factorization equations from Toda field theory *Nucl. Phys. B*, 190 \[FS3\] (1981) 542-552. M. Jimbo, Quantization of solitons and the restricted sine-gordon model. , 102: 537–547 (1996). . . , [23]{}:[2435–41]{}, [1983]{}. P. P. Kulish and N. [Yu]{}. Reshetikhin. -invariant solutions of the [Y]{}ang-[B]{}axter equation and associated quantum systems. , 34:1948–1971, 1986. translated from Zap. Nauch. Sem. LOMI 120, 92-121 (1982). . . , [10]{}:[63-69]{}, [1985]{}. . . , [10]{}:[798-820]{}, [1986]{}. . . , (1994). P. P. Kulish, N. Yu. Reshetikhin, and E. K. Sklyanin. ang-[B]{}axter equation and representation theory [I]{}. *Lett. Math. Phys.*, 5:393–403, 1981. P.  P. Kulish, N. Yu. Reshetikhin, Diagonalization of GL(N) invariant transfer matrices and quantum N-wave system (Lee model), (1983) L591. N. Yu. Reshetikhin, A method of functional equations in the theory of exactly solvable quantum systems (1983) 205. N. Yu. Reshetikhin, The functional equation method in the theory of exactly soluble quantum systems (1983) 691. A.  N. Kirillov and N. Yu. Reshetikhin, Yangians, Bethe ansatz and combinatorics (1986) 199. N. Yu. Reshetikhin, The spectrum of the transfer matrices connected with Kac-Moody algebras **14** (1987) 235. A. Kuniba, T. Nakanishi, J. Suzuki, Functional relations in solvable lattice models I: Functional relations and representation theory, (1994) 5215. P.  P. Kulish, N. Yu. Reshetikhin, Generalized Heisenberg ferromagnet and the Gross-Neveu model, *Sov. Phys. JETP*, **53** (1981) 108. S. Belliard and E. Ragoucy The nested Bethe ansatz for ’all’ closed spin chains (2008) **41** 295202. S. Pakuliak, E. Ragoucy and N. Slavnov, Nested Algebraic Bethe Ansatz in integrable models: recent results (2018) **6**. S. Belliard, S. Pakuliak, E. Ragoucy and N. A. Slavnov Highest coefficient of scalar products in SU(3)-invariant integrable models (2012) P09003. S. Belliard, S. Pakuliak, E. Ragoucy and N. A. Slavnov The algebraic Bethe ansatz for scalar products in SU (3)-invariant integrable models (2012) P10017. S. Belliard, S. Pakuliak, E. Ragoucy and N. A. Slavnov Bethe vectors of GL(3)-invariant integrable models, (2013) P02020. S. Belliard, S. Pakuliak, E. Ragoucy and N. A. Slavnov Form factors in SU(3)-invariant integrable models, (2013) P04033. S. Pakuliak, E. Ragoucy and N. A. Slavnov Form factors in quantum integrable models with GL(3)-invariant R-matrix, (2014) 343. S. Pakuliak, E. Ragoucy and N. A. Slavnov Determinant representations for form factors in quantum integrable models with the GL(3)-invariant R-matrix, (2014) 1566. S. Pakuliak, E. Ragoucy and N. A. Slavnov GL(3)-Based Quantum Integrable Composite Models. I. Bethe Vectors, *SYGMA* **11** (2015) 063. S. Pakuliak, E. Ragoucy and N. A. Slavnov GL(3)-Based Quantum Integrable Composite Models. II. Form Factors of Local Operators (2015) 064. S. Pakuliak, E. Ragoucy and N. A. Slavnov Form factors of local operators in a one-dimensional two-component Bose gas, (2015) 435001. A. Liashyk and N. A. Slavnov. On [B]{}ethe vectors in gl3 -invariant integrable models. , 06:18, 2018. K. Hao, J. Cao, G.-L. Li, W.-L. Yang, K. Shi and Y. Wang, Exact solution of an su(n) spin torus (2016) 073104. E. K. Sklyanin. The quantum [T]{}oda chain. In N. Sanchez, editor, *Non-Linear Equations in Classical and Quantum Field Theory*, pages 196–233. Springer Berlin Heidelberg, 1985. E. K. Sklyanin. Functional [B]{}ethe [A]{}nsatz. In B.A. Kupershmidt, editor, *Integrable and Superintegrable Systems*, pages 8–33. World Scientific, Singapore, 1990. E. K. Sklyanin. Quantum inverse scattering method. [S]{}elected topics. In Mo-Lin Ge, editor, *Quantum Group and Quantum Integrable Systems*, pages 63–97. Nankai Lectures in Mathematical Physics, World Scientific, 1992. E. K. Sklyanin. Separation of variables in the classical integrable sl3 magnetic chain. , 150:181–191, 1992. E. K. Sklyanin. Separation of variables. [N]{}ew trends. , 118:35–60, 1995. E. K. Sklyanin. Separation of variables in the quantum integrable models related to the yangian \[sl(3)\]. , 80:1861–1871, 1996. O. Babelon, D. Bernard, and F. A. Smirnov. Quantization of solitons and the restricted sine-gordon model. *Comm. Math. Phys.*, 182(2):319–354, Dec 1996. F. A. Smirnov. Structure of matrix elements in the quantum [T]{}oda chain. , 31(44):8953, 1998. F. Smirnov. . , 2001. S. E. Derkachov, G.P. Korchemsky, and A. N. Manashov. . , 617:375–440, 2001. S. E. Derkachov, G. P. Korchemsky, , and A. N. Manashov. Separation of variables for the quantum [SL(2,$\mathbb{\ R}$)]{} spin chain. , 07:047, 2003. S. E. Derkachov, G. P. Korchemsky, and A. N. Manashov. Baxter [Q]{}-operator and separation of variables for the open [SL(2,$\mathbb{R}$)]{} spin chain. , 10:053, 2003. A. Bytsko and J. Teschner. . , 39:12927, 2006. G. von Gehlen, N. Iorgov, S. Pakuliak, and V. Shadura. . , 39:7257, 2006. H. Frahm, A. Seel, and T. Wirth. Separation of variables in the open [XXX]{} chain. , 802:351–367, 2008. E. Mukhin, V. Tarasov, and A. Varchenko. On separation of variables and completeness of the [B]{}ethe ansatz for quantum gln [G]{}audin model. , 51:137–145, 2009. E. Mukhin, V. Tarasov, and A. Varchenko. Schubert calculus and the representations of the general linear group. , 22:909–940, 2009. E. Mukhin, V. Tarasov, and A. Varchenko. The [B]{}. and [M]{}. [S]{}hapiro conjecture in real algebraic geometry and the [B]{}ethe ansatz. , 170:863–881, 2009. L. Amico, H. Frahm, A. Osterloh, and T. Wirth. Separation of variables for integrable spin-boson models. , 839(3):604 – 626, 2010. G. Niccoli and J. Teschner. . , 2010:P09014, 2010. G. Niccoli. . , 835:263–283, 2010. G. Niccoli. Completeness of [B]{}ethe [A]{}nsatz by [S]{}klyanin [SOV]{} for cyclic representations of integrable quantum models. , 03:123, 2011. H. Frahm, J. H. Grelik, A. Seel, and T. Wirth. Functional [B]{}ethe ansatz methods for the open [XXX]{} chain. , 44:015001, 2011. N. Grosjean, J. M. Maillet, and G. Niccoli. On the form factors of local operators in the lattice sine-[G]{}ordon model. , 2012:P10006, 2012. N. Grosjean and G. Niccoli. . *J. Stat. Mech.*, 2012:P11005, 2012. G. Niccoli. Non-diagonal open spin-1/2 [XXZ]{} quantum chains by separation of variables: Complete spectrum and matrix elements of some quasi-local operators. , 2012:P10025, 2012. G. Niccoli. Antiperiodic spin-1/2 [XXZ]{} quantum chains by separation of variables: Complete spectrum and form factors. , 870:397–420, 2013. G. Niccoli. An antiperiodic dynamical six-vertex model: [I]{}. [C]{}omplete spectrum by [SOV]{}, matrix elements of the identity on separate states and connections to the periodic eight-vertex model. , 46:075003, 2013. G. Niccoli. Form factors and complete spectrum of [XXX]{} antiperiodic higher spin chains by quantum separation of variables. , 54:053516, 2013. N. Grosjean, J. M. Maillet, and G. Niccoli. . , 16:1103, 2015. S. Faldella and G. Niccoli. approach for integrable quantum models associated with general representations on spin-1/2 chains of the 8-vertex reflection algebra. , 47:115202, 2014. S. Faldella, N. Kitanine, and G. Niccoli. Complete spectrum and scalar products for the open spin-1/2 [XXZ]{} quantum chains with non-diagonal boundary terms. , 2014:P01011, 2014. N. Kitanine, J. M. Maillet, and G. Niccoli. Open spin chains with generic integrable boundaries: Baxter equation and [B]{}ethe ansatz completeness from separation of variables. , 2015:P05015, 2014. G. Niccoli and V. Terras. Antiperiodic [XXZ]{} chains with arbitrary spins: Complete eigenstate construction by functional equations in separation of variables. , 105:989–1031, 2015. D. Levy-Bencheton, G. Niccoli, and V. Terras. Antiperiodic dynamical 6-vertex model by separation of variables [II]{}: Functional equation and form factors. 2016:033110. G. Niccoli and V. Terras. The 8-vertex model with quasi-periodic boundary conditions. , 49:044001, 2016. N. Kitanine, J. M. Maillet, G. Niccoli, and V. Terras. On determinant representations of scalar products and form factors in the [SoV]{} approach: the [XXX]{} case. , 49:104002, 2016. D. Martin and F. Smirnov. . , 106:469–484, 2016. N. Kitanine, J. M. Maillet, G. Niccoli, and V. Terras. The open [XXX]{} spin chain in the sov framework: scalar product of separate states. , 50(22):224001, 2017. J. M. Maillet, G. Niccoli, and B. Pezelier. . , 2:009, 2017. N. Gromov, F. Levkovich-Maslyuk, and G. Sizov. New construction of eigenstates and separation of variables for [SU]{}([N]{}) quantum spin chains. , 09:111, 2017. S. E. Derkachov and P. A. Valinevich. Separation of variables for the quantum [SL]{}(3,[C]{}) spin magnet: eigenfunctions of [S]{}klyanin [B]{}-operator. arXiv:1807.00302 \[math-ph\]. P. Ryan, D. Volin Separated variables and wave functions for rational gl(N) spin chains in the companion twist frame arXiv:1810.10996. . . , [76]{}:[1–24 ; 25–47 ; 48–71]{}, [1973]{}. . . , [9(2)]{}:[145–182]{}, [1973]{}. . . , [15]{}:[485–503]{}, [1976]{}. . . , [76]{}:[1–14]{}, [1977]{}. . . , [289]{}:[315–346]{}, [1978]{}. . . , [330]{}:[523]{}, [1990]{}. . . , [28]{}:[2759]{}, [1995]{}. C. M. Yung and M. T. Batchelor. Exact solution for the spin-s [XXZ]{} quantum chain with non-diagonal twists. , 446:461–484, 1995. V. V. Bazhanov, S. L. Lukyanov, and A. B. Zamolodchikov. Integrable structure of conformal field theory [II]{}. [Q]{}-operator and [DDV]{} equation. , 190:247–278, 1997. hep-th/9604044. A. Antonov and B. Feigin. Quantum group representation and [B]{}axter equation. , 392:115–122, 1997. hep-th/9603105. V. V. Bazhanov, S. L. Lukyanov, and A. B. Zamolodchikov. Integrable structure of conformal field theory. [III]{}: The [Y]{}ang-[B]{}axter relation. , 200:297–324, 1999. hep-th/9805008. S. E. Derkachov. Baxter’s [Q]{}-operator for the homogeneous [XXX]{} spin chain. , 32:5299–5316, 1999. solv-int/9902015. G. P. Pronko. On the [B]{}axter’s [Q]{} operator for the [XXX]{} spin chain. , 212:687–701, 2000. hep-th/9908179. C. Korff. A [Q]{}-operator for the twisted [XXX]{} model. , 39:3203–3219, 2006. math-ph/0511022. S. E. Derkachov. Factorization of [R]{}-matrix and [B]{}axter’s [Q]{}-operator. , 151:2848, 2008. math/0507252. V. V. Bazhanov, T. Lukowski, C. Meneghelli, and M. Staudacher. A shortcut to the [Q]{}-operator. , page P11002, 2010. V. V. Mangazeev. Q-operators in the six-vertex model. , 886:166–184, 2014. . . , [21]{}:[1189–1371]{}, [2017]{}. . . [^1]: As already stressed, in all this article we assume that the deformation parameter $q=e^{\eta}$ is not a root of unity [^2]: That is we have to fix $\delta _{0}$ to be one of the three distinct eigenvalues of the matrix$K$ and then the degree of the Laurent polynomial $\varphi _{t}(\lambda )$ is fixed by:$$\mathsf{M}=\mathsf{N}-\nu_{1},\text{ for }\delta _{0}=\alpha ,\text{ }\mathsf{M}=\mathsf{N}-\nu_{2},\text{ for }\delta _{0}=\beta ,\text{ }\mathsf{M}=\mathsf{N}-\nu_{3},\text{ for }\delta _{0}=\gamma \text{, with }\nu_{3}=\mathsf{N}-(\nu_{1}+\nu_{2}).\notag$$ [^3]: These $R$-matrices are associated to general values of $q$, the root of unity case has been also analyzed see for example [@Au-YP06] for a review. However in all the present article we assume that $q$ is not a root of unity. [^4]: In the special case associated to the so-called anti-periodic boundary conditions a first eigenvalue analysis has been developed in [HaoCLYSW16]{}. [^5]: i.e. the transfer matrix eigenvalues associated to the common eigenspace of the $\mathsf{N}_{1},...,\mathsf{N}_{n}$ with eigenvalues $\nu_{1},...,\nu_{n}.$ [^6]: That is we have to fix $\delta _{0}$ to be one of the $n$ distinct eigenvalue $\mathsf{k}_{i}$ of the twist matrix $K$ and then the degree of the Laurent polynomial $\varphi _{t}(\lambda )$ is fixed by $\mathsf{M}=\mathsf{N}-\nu_{i}$. [^7]: Note that we can fix for example $\xi _{N+1}=\xi _{h}-\eta $ for any fixed $h\in \{1,\ldots ,N\}$. [^8]: See for example William Fulton (1974). Algebraic Curves. Mathematics Lecture Note Series. W.A. Benjamin. [^9]: Indeed, if there are common components the system admits instead an infinite number of solutions.

ArXiv

--- abstract: 'Radio-frequency (RF) impairments, which intimately exist in wireless communication systems, can severely limit the performance of multiple-input multiple-output (MIMO) systems. Although we can resort to compensation schemes to mitigate part of these impairments, a certain amount of residual impairments always persists. In this paper, we consider a training-based point-to-point MIMO system with residual transmit RF impairments (RTRI) using spatial multiplexing transmission. Specifically, we derive a new linear channel estimator for the proposed model, and show that RTRI create an estimation error floor in the high signal-to-noise ratio (SNR) regime. Moreover, we derive closed-form expressions for the signal-to-noise-and-interference ratio (SINR) distributions, along with analytical expressions for the ergodic achievable rates of zero-forcing, maximum ratio combining, and minimum mean-squared error receivers, respectively. In addition, we optimize the ergodic achievable rates with respect to the training sequence length, and demonstrate that finite dimensional systems with RTRI generally require more training at high SNRs than those with ideal hardware. At last, we extend our analysis to large-scale MIMO configurations, and derive deterministic equivalents of the ergodic achievable rates. It is shown that, by deploying large receive antenna arrays, the extra training requirements due to RTRI can be eliminated. In fact, with sufficiently large number of receive antennas, systems with RTRI may even need less training than systems with ideal hardware.' author: - 'Xinlin Zhang,  Michail Matthaiou,  Mikael Coldrey,  and Emil Björnson, [^1][^2] [^3][^4] [^5]' bibliography: - 'IEEEabrv.bib' - 'Reference.bib' title: 'Impact of Residual Transmit RF Impairments on Training-Based MIMO Systems' --- Hardware impairments, large-scale MIMO, pilot optimization, random matrix theory, training-based channel estimation Introduction ============ Point-to-point multiple-input multiple-output (MIMO) systems offer wireless communication with high data rates, without requiring additional bandwidth or transmit power. The pioneering works of [@telatar1999capacity; @foschini1998a] demonstrated a linear growth in capacity in rich scattering environments by deploying multiple antennas at both the transmitter and receiver sides. However, to fully reap the advantages that MIMO systems can offer, perfect instantaneous channel state information (CSI) is essential, especially at the receiver. In many practical systems, a training-based (or pilot-based) transmission scheme is usually adopted to acquire CSI. In such systems, the transmitter sends a block of symbols which contain both pilot and data information. The receiver estimates the instantaneous channel realization and uses the acquired CSI to retrieve the intended data. Under ideal radio-frequency (RF) hardware assumptions, channel estimation is a well investigated topic (e.g., [@hassibi2003; @biguesh2006training; @mikael1; @emil2010est; @Love2014Training1_JSTSP; @Love2014Training2_JSTSP]); however, these assumptions are quite unrealistic in practice. RF impairments, such as in-phase/quadrature-phase (I/Q) imbalance, high power amplifier non-linearities, and oscillator phase noise, are known to have a detrimental impact on practical MIMO systems [@schenk2008rf; @studer2010residual; @Jingya2014IQI_TCOM; @Giuseppe2014PA_TCOM]. The influence of such individual hardware impairments is usually mitigated by using analog and digital signal processing algorithms [@schenk2008rf]. However, these techniques are not able to completely remove hardware impairments, such that a certain amount of residual distortions always remains. These residual impairments stem from, for example, time-varying hardware characteristics which cannot be accurately parameterized and estimated, as well as, the randomness induced by different types of noise and imperfect compensation schemes. The statistical properties of residual impairments have been investigated in a number of existing works. The authors in [@schenk2008rf; @studer2010residual] characterized and verified experimentally that the residual distortion caused by the impairments behaves as additive and independent Gaussian noise. This Gaussian behavior can be understood by the central limit theorem, when the distortions from many independent and different sources add up together. In particular, for Gaussian input, one can interpret the above observation using the Bussgang Theorem [@Bussgang], which states that the output process of a Gaussian input process through a non-ideal hardware is the sum of a scaled input process and a distortion process which is uncorrected with the input process [@Ochiai2002Clipped_TCOM; @Wenyi2012Framework_TCOM]. The very recent paper [@Ulf2014Impairments_GLOBECOM] has compared the above additive Gaussian model with a sophisticated hardware impairment model in massive MIMO systems, which comprised several main sources of hardware impairments. Their results verified that the above Gaussian approximation is sufficiently accurate despite of its simplicity. The impact of residual hardware impairments has been only scarcely investigated in few works recently. In [@Emil2013imp_COML; @xinlin2014capacity_ICC; @Michalis2013Dhop_TCOM; @Michalis2013TwoWay_COML; @emil2013est], the impact of residual RF impairments was studied under different system scenarios, with a few important conclusions being drawn. For example, [@Emil2013imp_COML; @xinlin2014capacity_ICC] showed that impairments fundamentally limit the channel capacity in the high signal-to-noise (SNR) regime, while for low transmit power, such impact is negligible. Moreover, [@xinlin2014capacity_ICC] has shown that by deploying large antenna arrays at both the transmitter and receiver, the influence of impairments can be alleviated. Nonetheless, these results hold only for optimal receivers which induce high computational complexity, and are not of particular interest for larger-scale MIMO topologies. In [@emil2013est], the authors claimed that the channel capacity is mainly limited by the residual impairments at the user terminals, and also proved that large-scale MIMO systems were able to tolerate larger hardware impairments than conventional small-scale MIMO systems. The authors therein also showed a few limiting effects induced by impairments in the large-antenna regime, such as estimation accuracy and capacity ceiling; however, they did not pursue any resource allocation analysis, which is of pivotal importance for the performance optimization of training-based MIMO systems. Motivated by the above mentioned limitations, in this paper, we consider a training-based MIMO system affected by residual transmit RF impairments (RTRI). Our main goal is to investigate the impact of RTRI on the rates that are achievable with different linear receivers, as well as to find the optimal training length. More specifically, our main contributions are: - We derive a linear minimum mean-squared error (LMMSE) channel estimator of the channel with optimized training matrix structure. We analytically show that RTRI introduce an irreducible estimation error floor in the high SNR regime. <!-- --> - We derive closed-form expressions which tightly approximate the SINR distributions (outage probabilities) and ergodic achievable rates for zero-forcing (ZF), maximum ratio combining (MRC), and minimum mean-squared error (MMSE) receivers. In contrast to [@emil2013est], these results hold *for any finite number of antennas*. We observe achievable rate ceilings in the high SNR regime, which are analytically quantified, while in the low SNR regime, we show that the impact of RTRI vanishes. Moreover, we demonstrate that MRC receivers are much more resilient to hardware impairments than ZF and MMSE receivers. - We find the optimal training length of MIMO systems with RTRI. We show that, if ZF and MMSE receivers are adopted, the optimal training length can be much larger than that of systems with ideal hardware. In general, higher levels of RTRI impose more training requirements. - Finally, we extend our analysis to a general large-scale MIMO scenario. With the help of random matrix theory, we derive *deterministic equivalents* of the ergodic achievable rates. The optimal training length is thereafter found by solving an asymptotically optimal convex optimization problem. An interesting observation is that with large number of receive antennas, systems with RTRI may need even less training than ideal hardware systems. The paper is structured as follows: The system model, including channel training and data transmission, is described in Section \[sec:SystemModel\]. The estimation phase is analyzed in Section \[sec:estimation\]. In Section \[sec:datatransmission\], we derive the achievable rates of different linear receivers, along with the optimal training length. We extend our analysis to large-scale MIMO systems in Section \[sec:MassiveMIMO\]. Finally, conclusions are drawn in Section \[sec:Conclusion\]. *Notation:* Upper and lower case boldface letters denote matrices and vectors, respectively. The trace of a matrix is expressed by $\mathrm{tr}\left(\cdot\right)$. The $n \times n$ identity matrix is represented by $\mathbf I_n$. The expectation operation is $\mathbb {E} [\cdot]$, while the matrix determinant is denoted by det$(\cdot)$. The superscripts $(\cdot)^ H$, $(\cdot)^{-1}$ and $(\cdot)^{\dagger}$ stand for Hermitian transposition, matrix inverse and pseudo-inverse, respectively. The Frobenius norm and spectral norm are denoted by $\left\|\cdot\right\|_F$ and $\left\|\cdot\right\|_2$, respectively. The symbol $\mathcal{CN}\left(\mathbf m, \boldsymbol\Sigma\right)$ denotes a multi-variate circularly-symmetric complex Gaussian (CSCG) distribution with mean $\mathbf m$ and covariance $\boldsymbol\Sigma$. For any matrix $\mathbf H\in\mathbb C^{m\times n}$, $\mathbf h_i$ is the $i$-th column of $\mathbf H$. Signal and System Models {#sec:SystemModel} ======================== We consider a block fading channel with coherence length ${T}$. During each block, the channel is constant, and is a realization of the uncorrelated Rayleigh fading model. Channel realizations between different blocks are assumed to be independent. System Model With RTRI ---------------------- RF impairments exist inherently in practical wireless communication systems. Due to these impairments, the transmitted signal will be distorted during the transmission processing, thereby causing a mismatch between the intended signal and what is actually transmitted. Even though compensation schemes are usually adopted to mitigate the effects of these impairments, there is always some amount of residual impairments. In [@schenk2008rf; @studer2010residual], the authors have shown that these residual impairments on the transmit side act as additive noise. Furthermore, experimental results in [@studer2010residual] revealed that such RTRI behave like zero-mean complex Gaussian noise, but with the important property that their average power is proportional to the average signal power. For sufficient decoupling between different RF chains, such impairments are statistically independent across the antennas. Moreover, impairments during different channel uses are also assumed to be independent.[^6] We now denote the RTRI noise as $\boldsymbol\Delta$. Then, the input-output relationship of a training-based MIMO system with $N_t$ transmit antennas and $N_r$ receive antennas, within a block of $T$ symbols, can be expressed as $$\mathbf Y = \sqrt{\frac{\rho}{N_t}}\mathbf H (\mathbf S + \boldsymbol \Delta) + \mathbf V, \label{eq:SysModelAll}$$ where ${\mathbf S}\in\mathbb{C}^{N_t\times T}$ is the transmitted signal containing both pilot symbols and data symbols,[^7] and $\mathbf H\in\mathbb C^{N_r\times N_t}$ is the channel matrix. The receiver noise and the received signal are denoted as $\mathbf V\in\mathbb{C}^{N_r\times T}$ and $\mathbf Y\in\mathbb{C}^{N_r\times T}$, respectively. Each element of $\mathbf H$ and $\mathbf V$ follows a $\mathcal{CN}(0,1)$ distribution. The average SNR at each receive antenna is denoted by $\rho$. At last, according to the above discussion, we can characterize the RTRI noise $\boldsymbol\Delta$ according to $$\begin{aligned} \boldsymbol\delta_{i} \sim \mathcal{CN}\left(\mathbf{0}, \delta^2 \mathbf I_{N_t} \right), \mathbb E\left[\boldsymbol\delta_{i}\boldsymbol\delta_{j}^H\right] = \mathbf 0\notag\ i, j= 1, 2, \dots, T, i\neq j,\end{aligned}$$ where ${\boldsymbol\delta}_{i}$ denotes the $i$th column of $\boldsymbol\Delta$, and the proportionality parameter $\delta$ characterizes the level of residual impairments in the transmitter. Note that $\delta$ appears in practical applications as the error vector magnitude (EVM) [@holma2011LTE], which is commonly used to measure the quality of RF transceivers. For instance, 3GPP LTE has EVM requirements in the range $\left[0.08,0.175 \right]$ [@holma2011LTE], where smaller EVMs are required to support high rates. The relationship between $\delta$ and EVM is defined as $$\mathrm{EVM} \triangleq \sqrt{\frac{\mathbb{E}_{\boldsymbol\Delta}\left[ \left\| \boldsymbol\Delta \right\|_F^2 \right]}{\mathbb{E}_{\mathbf S}\left[ \left\| \mathbf S\right\|_F^2 \right]}} = \delta.$$ Evidently, when $\delta = 0$, the system model simplifies to the ideal hardware case. The EVM can be measured in practice using well-established experiment setups, e.g. [@Agilent8Hints]. We can now decompose the system model in (\[eq:SysModelAll\]) into the training and data transmission phases as follows: ### Training Phase $$\begin{aligned} \label{eq:SysModelTraining} \mathbf Y_p = \sqrt{\frac{\rho}{N_t}}\mathbf H \left( \mathbf S_p\! +\! \boldsymbol\Delta_p \right) + \mathbf V_p, ~\mathrm{tr}\left(\mathbf S_p\mathbf S_p^H\right) = N_tT_p,\end{aligned}$$ where $\mathbf S_p\in\mathbb C^{N_t\times T_p}$ is the deterministic matrix of training sequences sent over $T_p$ channel uses, that is known by the receiver, $\mathbf Y_p$ is the $N_r\times T_p$ received matrix. The distortion noise caused by the RTRI is characterized as $$\begin{aligned} {\boldsymbol\delta_p}_{i}\sim \mathcal{CN}\left(\mathbf 0, \delta^2\mathbf I_{N_t}\right), \mathbb E\left[{\boldsymbol\delta_p}_{i}{\boldsymbol\delta_p}_{j}^H\right] = \mathbf 0, ~i,j=1,2,\dots,T_p,i\neq j,\end{aligned}$$ where ${\boldsymbol\delta_p}_{i}$ denotes the $i$th column of $\boldsymbol\Delta_p$. ### Data Transmission Phase $$\begin{aligned} \label{eq:SysModelData} \mathbf Y_d \!= \!\sqrt{\frac{\rho}{N_t}}\mathbf H \left( \mathbf S_d \!+\! \boldsymbol\Delta_d\right)\! +\! \mathbf V_d, ~\mathbb{E}\Big[\mathrm{tr}\left(\mathbf {S}_d^H\mathbf S_d\right)\Big]\! =\! N_t T_d,\end{aligned}$$ where $\mathbf S_d \in \mathbb{C}^{N_t \times T_d}$ is the matrix of data symbols sent over $T_d$ channel uses, $\mathbf Y_d$ is the $N_r \times T_d$ received signal matrix. The distortion noise caused by RTRI is characterized as $$\begin{aligned} {\boldsymbol\delta_d}_{i}\sim \mathcal{CN}\left(\mathbf 0, \delta^2\mathbf I_{N_t}\right), \mathbb E\left[{\boldsymbol\delta_d}_{i}{\boldsymbol\delta_d}_{j}^H\right] = \mathbf 0,~i,j=1,2,\dots,T_d,i\neq j,\end{aligned}$$ where ${\boldsymbol\delta_d}_{i}$ denotes the $i$th column of $\boldsymbol\Delta_d$. Note that $\mathbf S = \left[\mathbf S_p ~\mathbf S_d\right]$ and $\boldsymbol\Delta = \left[\boldsymbol\Delta_p ~\boldsymbol\Delta_d\right]$, and the conservation of coherence block length yields $T = T_p + T_d$. We only consider the RTRI at the transmitter side since it is more dominant than its counterpart at the receiver. In practical communication systems, the received signal will firstly pass through a low noise amplifier (LNA), such that the post-processed power for the rest of the receiver circuit (e.g. demodulator, mixer) is always within a certain range, no matter how much power is actually captured by the antenna. Moreover, even if we model the receiver residual impairments in the same way as the RTRI, i.e., its variance is proportional to the received power that is received by the antenna, it is still less dominant in some practically interesting cases; for example, low SNR applications and large-scale MIMO with many receive antennas [@xinlin2014capacity_ICC]. Channel Estimation {#sec:estimation} ================== In this section, we analyze the impact of RTRI on the training phase. Our objective is to characterize how estimation accuracy is affected by RTRI. Conditioned on the availability of RTRI distortion distribution at the receiver, the current channel realization within each block can be estimated through the received signal $\mathbf Y_p$. The received signal contains both Gaussian distributed terms $\mathbf H\mathbf S_p$ and $\mathbf V_p$ and double Gaussian distributed term $\mathbf H\boldsymbol\Delta_p$; therefore, closed-form MMSE estimator is difficult to determine. Thus, we resort to LMMSE estimator, whose closed-form expression can be derived through tractable manipulations, and which performs better than other non-Bayesian estimators, e.g., least square (LS) estimator. Given the signal model in , we can derive the LMMSE estimator by directly extending the classical estimation results [@biguesh2006training], which is given by $$\label{eq:estimator}\vspace{-0.3cm} \hat{\mathbf H} = \sqrt{\frac{\rho}{N_t}}\mathbf Y_p\bigg(\frac{\rho}{N_t} \mathbf S_p^H{\mathbf S_p} + \left( \delta^2\rho + 1 \right)\mathbf I_{T_p} \bigg)^{-1}\mathbf S_p^H.$$ For any given training symbol matrix $\mathbf S_p$, minimizes the mean-squared error (MSE) from the true channel, which is defined as $\mathrm{MSE} \triangleq \mathrm{tr}\left(\tilde{\mathbf C}\right)$, where $\tilde{\mathbf C}\!\triangleq\!\mathbb E\left[ \tilde{\mathbf H}^H\tilde{\mathbf H}\right]$ represents the estimation error covariance matrix, and $\tilde{\mathbf H} \triangleq \mathbf H - \hat{\mathbf H}$ is the estimation error. Furthermore, using a similar approach as in [@biguesh2006training], we find that the optimal training sequence which minimizes the MSE should satisfy $\mathbf S_p\mathbf S_p^H = T_p\mathbf I_{N_t}$, and the corresponding MSE reduces to $$\label{eq:MSE_expression}\vspace{-0.3cm} \mathrm{MSE} = \frac{N_rN_t}{1 + \epsilon}, ~\mathrm{with}~ \epsilon \triangleq \frac{\rho T_p}{N_t(\rho\delta^2+1)}.$$ The resulting estimation error covariance matrix therefore becomes $\tilde{\mathbf C} = \frac{N_r}{1+\epsilon}\mathbf I_{N_t}$. It is straightforward to show that $\tilde{\mathbf H}$ has i.i.d. entries with zero mean. The variance of each element in $\tilde{\mathbf H}$, which is also defined here as the normalized MSE (NMSE), can be expressed as $\sigma_{\tilde{\mathbf H}}^2 = \frac{1}{N_rN_t}\mathrm{tr}\left(\tilde{\mathbf C}\right) = \frac{1}{1+\epsilon}$. By the orthogonality principle of LMMSE estimates [@kaybook1], each element in $\hat{\mathbf H}$ has a variance of $\sigma_{\hat{\mathbf H}}^2 = 1 - \sigma_{\tilde{\mathbf H}}^2 = \frac{\epsilon}{1+\epsilon}$. If we examine at the expression of the NMSE, it is trivial to find that the estimator reaches an error floor at high SNRs, which is quantified by taking $\rho\rightarrow\infty$ as $$\vspace{-0.3cm} \mathrm{NMSE}_\mathrm{limit} = \frac{1}{1+\frac{T_p}{N_t\delta^2}}.\label{eq:MSE_limit}$$ Obviously, the value of this floor depends on the level of RTRI; in general, large RTRI will cause severe degradation of the channel estimates. We can also see from (\[eq:MSE\_limit\]) that, for a fixed level of RTRI, an increase in the training sequence length $T_p$ decreases the MSE monotonically. Figure \[fig:MSE\] shows the NMSE, $\sigma_{\tilde{\mathbf H}}^2$, of a MIMO system with $N_t = N_r = 4$ for different levels of impairments. In this case, we use $T_p\!=\!4$ channel uses to transmit pilot symbols, which is the minimum length required to estimate all channel dimensions. Without the existence of RTRI, increasing the transmit power decreases the NMSE monotonically towards zero. However, in the presence of RTRI, we observe a fundamentally different behavior. Specifically, when the transmit power becomes high, impairments will generate an irreducible error floor, which was explicitly quantified in . ![Normalized mean-squared error (MSE) for different levels of impairments ($N_t = N_r = 4, T = 100, T_p = 4$).[]{data-label="fig:MSE"}](Normalized_MSE.pdf){height="35.00000%"} Data Transmission {#sec:datatransmission} ================= In this section, we investigate the data transmission of a training-based MIMO system with spatial multiplexing scheme. We seek to analyze the impact of RTRI on the ergodic achievable rates of three most common linear receivers, namely ZF, MRC and MMSE receivers, and to find the optimal training length that maximizes the corresponding ergodic achievable rates. Specifically, for a given coherence time $T$, we can formulate the following optimization problem: $$\begin{aligned} \label{eq:Optimization1} &\textrm{maximize} ~~~R\notag\\\vspace{-0.4cm} &\textrm{subject to} ~~~N_t\leq T_p<T,\end{aligned}$$ where $R$ is the ergodic achievable rate parameterized by $T_p$. We derive SINR distributions for each type of receiver to evaluate $R$. The optimal training length $T_p^{*}$ is found through a line search over all possible $T_p$ values that satisfy the conditions in (\[eq:Optimization1\]). In the following, we pursue our analysis for exact SNRs, as well as, high and low SNRs. Exact SNR analysis {#sec:RateExact} ------------------ We start by writing the $N_r\times 1$ received vector through (\[eq:SysModelData\]) $$\begin{aligned} {{\mathbf y}_d} & = \sqrt{\frac{\rho}{N_t}}{\hat {\mathbf H}}{\mathbf s_d} + \sqrt{\frac{\rho}{N_t}}\left({\tilde {\mathbf H}}{\mathbf s_d} + \mathbf H{\boldsymbol\delta_d}\right) +\mathbf v_d\\ & = \sqrt{\frac{\rho}{N_t}}\hat {\mathbf h}_k{s_d}_k + \underbrace{ \sqrt{\frac{\rho}{N_t}}\left(\sum_{i = 1, i \neq k}^{N_t}\hat {\mathbf h}_i{s_d}_i + {\tilde {\mathbf H}}{\mathbf s_d} + \mathbf H{\boldsymbol\delta_d}\right) +\mathbf v_d}_{\triangleq{{\mathbf z}_d}_k}, ~k = 1, 2, \dots, N_t\end{aligned}$$ where ${s_d}_k$ is the transmitted data symbol on the $k$-th transmit antenna, and ${\mathbf z_d}_k$ is the total effective noise plus interference on the $k$-th spatial stream. To recover the signal ${s_d}_k$, ${{\mathbf y}_d}$ is weighted by a $1\times N_r$ vector $\hat {\mathbf a}_k^H$, which for ZF, MRC and MMSE linear receivers is respectively given by $$\hat {\mathbf a}_k^H = \begin{cases} \hat {\mathbf g}_k^H, & \textrm{for ZF}\\ \hat{\mathbf h}^H_k, & \textrm{for MRC}\\ \hat{\mathbf h}^H_k\left(\mathbf R_{\mathbf z_{dk}}\right)^{-1} & \textrm{for MMSE} \end{cases}$$ where $\hat {\mathbf g}_k$ is the $k$-th column of $\hat {\mathbf G}$, with $\hat {\mathbf G} \triangleq \hat{\mathbf H}^{\dagger} = \hat{\mathbf H}\left(\hat{\mathbf H}^H\hat{\mathbf H}\right)^{-1}$, and $\mathbf R_{{\mathbf z_d}_k} \triangleq \mathbb E\left[{\mathbf z_{dk}}{\mathbf z^H_{dk}} | \hat{\mathbf H}\right]$ is the covariance matrix of ${{{\mathbf z}_d}_k}$, which is explicitly given by $$\mathbf R_{{\mathbf z_d}_k} = \frac{\rho}{N_t}\sum_{i = 1, i \neq k}^{N_t}\hat {\mathbf h}_i{\hat {\mathbf h}_i}^H + \frac{\rho \delta^2}{N_t}\hat{\mathbf H}\hat{\mathbf H}^H + \left( \frac{\rho+\rho\delta^2 +1+\epsilon}{1+\epsilon} \right)\mathbf I_{N_r}.$$ Multiplying the received signal with the weighting vector, and conditioned on that $\hat{\mathbf H}^H$ is known at the receiver, the SINR of ZF, MRC and MMSE receivers on the $k$-th spatial stream are respectively given as follows $$\vspace{-0.5cm} \gamma_{k,\mathrm{ZF}} = \frac{1}{\delta^2 + c_0\left[\left(\bar {\mathbf H}^H\bar {\mathbf H}\right)^{-1}\right]_{k,k}},\label{eq:SINR_ZF}$$ $$\gamma_{k,\mathrm{MRC}} = \frac{ \lvert {\bar {\mathbf h}_k}^H{\bar {\mathbf h}_k} \rvert^2 }{\sum_{i=1,i\neq k}^{N_t} \lvert {\bar {\mathbf h}_k}^H{\bar {\mathbf h}_i} \rvert^2 +\delta^2\sum_{i=1}^{N_t} \lvert {\bar {\mathbf h}_k}^H{\bar {\mathbf h}_i} \rvert^2 + c_0\lvert {\bar {\mathbf h}_k}^H{\bar {\mathbf h}_k}\rvert},\label{eq:SINR_MRC}$$ $$\gamma_{k,\mathrm{MMSE}} = {\bar {\mathbf h}_k}^H\left( \sum_{i=1,i\neq k}^{N_t}{\bar {\mathbf h}_i}{\bar {\mathbf h}_i}^H + \delta^2\sum_{i=1}^{N_t}{\bar {\mathbf h}_i}{\bar {\mathbf h}_i}^H +c_0\mathbf I_{N_r} \right)^{-1}\bar {\mathbf h}_k,\label{eq:SINR_MMSE}$$ where we have defined $c_0 \triangleq \frac{N_t\left(\rho+\rho\delta^2+1+\epsilon\right)}{\rho \epsilon}$ for notational convenience, and $\bar {\mathbf h}_i$ denotes the $i$th column of $\bar {\mathbf H}$. We also define $\bar{\mathbf H} \triangleq \frac{1}{\sigma^2_{\hat{\mathbf H}}}\hat{\mathbf H}$, which has independent and approximately $\mathcal{CN}(0,1)$ entires.[^8] Note that for ZF receivers, we require $N_r\geq N_t$, whilst we consider arbitrary antenna configurations for MRC and MMSE receivers. We now give the SINR cumulative distribution functions (CDFs) for these receivers in the following proposition, which, to best of our knowledge, are new. \[prop:cdf\] For ZF, MRC and MMSE linear receivers, the CDF of the output SINR of an arbitrary spatial stream can be tightly approximated as[^9] $$\begin{aligned} \label{eq:cdf_ZF} F_{\Gamma_{\mathrm{ZF}}}(\gamma) &= \begin{cases} 1-e^{-\frac{c_0\gamma}{1-\delta^2\gamma}}\sum\limits_{k=0}^{N_r-N_t}\frac{\left({\frac{c_0\gamma}{1-\delta^2\gamma}}\right)^k}{k!}, &0\leq\gamma<\frac{1}{\delta^2},\\ 1, &\gamma \geq \frac{1}{\delta^2},\\ \end{cases}\\ \label{eq:cdf_MRC} F_{\Gamma_{\mathrm{MRC}}}(\gamma) &= \begin{cases} 1-\frac{e^{-\frac{c_0\gamma}{1-\delta^2\gamma}} }{\left(\frac{1+\gamma}{1-\delta^2\gamma}\right)^{N_t-1}}\sum\limits_{k=0}^{N_r-1}\sum\limits_{p=0}^{k}\frac{\alpha_{p,k}\left( \frac{c_0\gamma}{1-\delta^2\gamma} \right)^k}{\left( \frac{1+\gamma}{1-\delta^2\gamma} \right)^p}, &0\leq\gamma<\frac{1}{\delta^2},\\ 1, &\gamma \geq \frac{1}{\delta^2},\\ \end{cases}\\ \label{eq:cdf_MMSE} F_{\Gamma_{\mathrm{MMSE}}}(\gamma) &= \begin{cases} 1- \frac{e^{-\frac{c_0\gamma}{1-\delta^2\gamma}}}{\left(\frac{1+\gamma}{1-\delta^2\gamma}\right)^{N_t-1}}\sum\limits_{k=0}^{N_r-1}\beta_k\left(\frac{c_0\gamma}{1-\delta^2\gamma}\right)^{k}, &0\leq\gamma<\frac{1}{\delta^2},\\ 1, &\gamma \geq \frac{1}{\delta^2}, \end{cases}\end{aligned}$$ where $\alpha_{p,k} \triangleq \frac{{N_t+p-2\choose p} \left(\frac{1+\delta^2}{c_0}\right)^p }{(k-p)!}$, and $\beta_k \triangleq \sum\limits_{p = \max(0,k-N_t+1)}^{k} \frac{{N_t-1\choose k-p}\left(\frac{c_0}{1+\delta^2}\right)^{p-k}}{p!}$. See Appendix \[appendix:zf\_cdf\]. The above closed-form expressions for the SINR CDF distributions incorporate the statistical impact of RTRI, and most importantly, are valid for any finite number of transmit and receive antennas. Note that for $\delta = 0$, the above SINR expressions reduce to the ones in [@Mckay2009LinearReceiver] with ideal hardware assumptions. Mathematically speaking, (\[eq:cdf\_ZF\])-(\[eq:cdf\_MMSE\]) are equivalent to the outage probabilities,[^10] since $$P_\mathrm{out}(x) = \mathrm{Pr}\{\gamma\leq x\} = F_\Gamma(x).$$ Note that Proposition \[prop:cdf\] showcases the important fact that the outage probability is always 1 for $\gamma\geq \frac{1}{\delta^2}$. This implies that an SINR wall exists, which cannot be crossed by simply increasing the transmit power. The value of this wall is inversely proportional to the square of the level of impairments. Note that similar conclusions were also drawn in [@Michalis2013Dhop_TCOM] in the context of dual-hop relaying. We now turn our attention to the ergodic achievable rates, and invoke the following expression of the ergodic achievable rate [@Mckay2009LinearReceiver Lemma 1]: $$\begin{aligned} \label{eq:cdf_rate} R &= \frac{T_d}{T}\mathbb E\left[ N_t\log_2\left(1+\gamma\right)\right]\notag\\ &=\frac{T_dN_t}{\ln(2)T}\int_0^{\infty}\frac{1-F_{\Gamma}(\gamma)}{1+\gamma}d\gamma,\end{aligned}$$ where $\gamma$ is found in (\[eq:SINR\_ZF\]), (\[eq:SINR\_MRC\]) and (\[eq:SINR\_MMSE\]) for ZF, MRC and MMSE receivers, respectively, and the function $F_{\Gamma}(\gamma)$ represents the CDF of $\gamma$. Specifically, the corresponding ergodic achievable rates are given in the following theorem. \[theorem:anal\_rates\] For training-based MIMO systems with RTRI, the ergodic achievable rates of ZF, MRC and MMSE receivers can be tightly approximated as $$\!\!\!\!R_\mathrm{ZF} = \frac{T_dN_t}{\ln(2)T}\sum_{k=0}^{N_r-N_t}c_0^k\left(e^{\frac{c_0}{\delta^2+1}}E_{k+1}\left(\frac{c_0}{\delta^2+1}\right)-e^{\frac{c_0}{\delta^2}}E_{k+1}\left(\frac{c_0}{\delta^2}\right)\right),\label{eq:anal_ZF_rate}$$ $$\begin{aligned} \!\!\!\!R_\mathrm{MRC} &= \frac{T_dN_t}{\ln(2)T}\sum_{k=0}^{N_r-1}\sum_{p=0}^{k}(-1)^{p+N_t}\alpha_{p,k}\delta^{2(p+N_t-1)}\Gamma(k+1)\notag\\ &\>\>\>\>\times\!\left(\!\!\Psi\!\left(1,-k\!+\!1;\!\frac{c_0}{\delta^2}\right) \!-\! \sum_{i=1}^{p+N_t}\!\!\left(\!-\frac{c_0}{\delta^2(\delta^2\!\!+\!\!1)}\!\right)^{p+N_t-i}\!\!\!\!\!\!\Psi\!\left(p\!+\!N_t\!-\!i\!+\!1,p\!+\!N_t\!-\!k\!-\!i\!+\!1;\frac{c_0}{\delta^2\!\!+\!\!1}\right)\!\!\right)\!,\label{eq:anal_MRC_rate}\\\vspace{-0.2cm} \!\!\!\!R_\mathrm{MMSE} &= \frac{T_dN_t}{\ln(2)T}\sum_{k=0}^{N_r-1}(-1)^{N_t}\beta_k\delta^{2(N_t-1)}\Gamma(k+1)\notag\\ &\>\>\>\>\times\!\left(\!\Psi\!\left(1,-k\!+\!1;\frac{c_0}{\delta^2}\right)\! -\! \sum_{i=1}^{N_t}\left(-\frac{c_0}{\delta^2(\delta^2\!\!+\!\!1)}\right)^{N_t-i}\!\!\!\Psi\left(N_t\!-\!i\!+\!1,-k\!+\!N_t\!-\!i\!+\!1;\frac{c_0}{\delta^2\!+\!1}\right)\!\!\right), \label{eq:anal_MMSE_rate}\end{aligned}$$ where $E_n(z)$ and $\Gamma(n)$ are the exponential integral function [@gradshteyn2007a Eq. (8.211.1)] and the Gamma function [@gradshteyn2007a Eq. (8.310.1)], respectively, while $\Psi(a,b;z)$ is the regularized hypergeometric function [@gradshteyn2007a Eq. (9.210.2)]. See Appendix \[appendix:anal\_rates\]. These ergodic achievable rates are valid for arbitrary antenna configurations, and can be efficiently evaluated and easily implemented. Unfortunately, they offer limited physical insights, and we, therefore, now elaborate on the low and high SNR regimes. Low SNR analysis, $\rho\rightarrow 0$ {#sec:RateLow} -------------------------------------- For some practical applications (e.g. battery-limited users or sensors), low SNR is of particular interest. Specifically, for ZF, MRC, and MMSE receivers, we have the following results. ### ZF receivers As $\rho\rightarrow 0$, the ZF SINR becomes $$\gamma_{\mathrm{ZF}}^{\rho\rightarrow 0} = \frac{T_p\rho^2}{N_t^2\left[\left(\bar {\mathbf H}^H\bar {\mathbf H}\right)^{-1}\right]_{k,k}}.$$ Consequently, we can approximate the ergodic achievable rate as below $$\begin{aligned} R_{\mathrm{ZF}}^{\rho\rightarrow 0} &= \mathbb E\left[\frac{T_dN_t}{T}\log_2\left( 1+\frac{T_p\rho^2}{N_t^2\left[\left(\bar {\mathbf H}^H\bar {\mathbf H}\right)^{-1}\right]_{k,k}}\right) \right]\notag\\ &\overset{(a)}\approx\frac{T_p(T-T_p)\rho^2}{\ln(2)TN_t}\mathbb E\left[ \frac{1}{\left[\left(\bar {\mathbf H}^H\bar {\mathbf H}\right)^{-1}\right]_{k,k}}\right]\notag\\ &\overset{(b)}=\frac{T_p(T-T_p)(N_r-N_t+1)}{\ln(2)TN_t}\rho^2,\label{eq:ZFLowRate}\end{aligned}$$ where $(a)$ is obtained by using truncated Taylor’s expansion, and $(b)$ is obtained by noticing that $1\slash\left[\left(\bar {\mathbf H}^H\bar {\mathbf H}\right)^{-1}\right]_{k,k}$ follows a chi-squared distribution with $2(N_r-N_t+1)$ degrees of freedom and is scaled by $1\slash 2$ [@winters1994diversity]. ### MRC receivers For very low transmit power, we can remove negligible terms from (\[eq:SINR\_MRC\]), and arrive at the following MRC SINR expression $$\gamma_{\mathrm{MRC}}^{\rho\rightarrow 0} = \frac{T_p\rho^2{\lvert {\bar {\mathbf h}_k}^H{\bar {\mathbf h}_k} \rvert}}{N_t^2}.$$ Through similar mathematical manipulations, we get the achievable rate of MRC receivers at low SNRs as $$R_{\mathrm{MRC}}^{\rho\rightarrow 0} \approx\frac{T_p(T-T_p)\rho^2}{\ln(2)TN_t}\mathbb E\left[ \lvert {\bar {\mathbf h}_k}^H{\bar {\mathbf h}_k} \rvert\right] \overset{(a)}=\frac{T_p(T-T_p)N_r}{\ln(2)TN_t}\rho^2,\label{eq:MRCLowRate}$$ where $(a)$ is achieved by using the fact that $\lvert {\bar {\mathbf h}_k}^H{\bar {\mathbf h}_k} \rvert$ is a complex chi-square distributed random variable with $2N_r$ degrees of freedom. ### MMSE receivers For MMSE receivers, we first rewrite (\[eq:SINR\_MMSE\]) as $$\begin{aligned} \gamma_{\mathrm{MMSE}} &\overset{(a)}= \frac{{\bar {\mathbf h}_k}^H\left(\sum_{i=1,i\neq k}^{N_t}{\bar {\mathbf h}_i}{\bar {\mathbf h}_i}^H + \frac{c_0}{1+\delta^2} \mathbf I_{N_r} \right)^{-1}{\bar {\mathbf h}_k}}{1+\delta^2+\delta^2{\bar {\mathbf h}_k}^H\left(\sum_{i=1,i\neq k}^{N_t}{\bar {\mathbf h}_i}{\bar {\mathbf h}_i}^H + \frac{c_0}{1+\delta^2}\mathbf I_{N_r} \right)^{-1}{\bar {\mathbf h}_k} }\label{eq:MMSE_SINR_trasform1}\\ &\overset{(b)}=\frac{1-\left[ \left(\mathbf I_{N_t} + \frac{1+\delta^2}{c_0}\bar{\mathbf H}^H\bar{\mathbf H} \right)^{-1} \right]_{k,k}}{\delta^2 + \left[ \left(\mathbf I_{N_t} + \frac{1+\delta^2}{c_0}\bar{\mathbf H}^H\bar{\mathbf H} \right)^{-1} \right]_{k,k}}\label{eq:MMSE_SINR_trasform2},\end{aligned}$$ where $(a)$ follows by applying Lemma \[lemma:MatrixInv\] in Appendix A , while $(b)$ is obtained by noticing that the quadratic form on the numerator of (\[eq:MMSE\_SINR\_trasform1\]) is a classic MMSE SINR expression with a modified noise power, hence it equals [$\frac{1}{\left[ \left(\mathbf I_{N_t} + \frac{1+\delta^2}{c_0}\bar{\mathbf H}^H\bar{\mathbf H} \right)^{-1} \right]_{k,k}} - 1$ ]{}[@Ping2006MMSE]. As $\rho\rightarrow 0$, we can easily find that the SINR and the ergodic achievable rate approach the following limits, respectively $$\begin{aligned} \gamma_{\mathrm{MMSE}}^{\rho\rightarrow 0} &= \frac{T_p\rho^2{\lvert {\bar {\mathbf h}_k}^H{\bar {\mathbf h}_k} \rvert}}{N_t^2},\\ R_{\mathrm{MMSE}}^{\rho\rightarrow 0} &=\frac{T_p(T-T_p)N_r}{\ln(2)TN_t}\rho^2.\label{eq:MMSELowRate}\end{aligned}$$ Comparing the ergodic achievable rates of different linear receivers in (\[eq:ZFLowRate\]), (\[eq:MRCLowRate\]) and (\[eq:MMSELowRate\]), we find that the achievable rates have no dependence on $\delta^2$, i.e., the impact of RTRI has vanished in the low transmit power regime. Moreover, it can also be easily conjectured that the optimal training length always equals half of the coherence length, i.e., $T_p^* = T\slash 2$, such that as much as half of the coherence block should be devoted to channel training, which is in line with the result in [@hassibi2003]. More importantly, to maintain a certain level of performance, we can increase the number of receive antennas, while reducing the transmit power as fast as $\sqrt{N_r}$ and $\sqrt{N_r-N_t+1}$, for MRC (or MMSE) and ZF receivers, respectively. We can also achieve this power scaling by using less transmit antennas. Finally, as one would expect, MMSE and MRC receivers are equally good in the low SNR regime, while ZF receiver performs badly. High SNR analysis, $\rho\rightarrow\infty$ {#sec:RateHigh} ------------------------------------------ From the expressions in (\[eq:anal\_ZF\_rate\])-(\[eq:anal\_MMSE\_rate\]), we can see that for a certain level of RTRI and system setup, the ergodic achievable rates only depend on the transmit power via the term $c_0$. As $\rho\rightarrow\infty$, $c_0\rightarrow\bar{c}_0 \triangleq \frac{\delta^2(1+\delta^2)N_t^2}{T_p}$. Therefore, the ergodic achievable rates of these linear receivers saturate in the high SNR regime. We can easily quantify these rate limits by substituting $c_0 = \bar{c}_0$ back into the corresponding achievable rate expressions in (\[eq:anal\_ZF\_rate\])-(\[eq:anal\_MMSE\_rate\]). Numerical Illustrations ----------------------- In Fig. \[fig:CDFs\], we plot the outage probabilities of linear receivers for different values of the SINR threshold. We consider two antenna configurations, i.e., $N_t = N_r = 5$ and $N_t = 5, N_r = 30$, at $\rho = 30$dB. We observe that for a certain SINR threshold, the outage probability of a hardware-impaired system is systematically higher than that of an ideal-hardware system. More importantly, RTRI create a SINR wall, which cannot be crossed by increasing the transmit power. On the other hand, for the ideal hardware case, the outage probability converges smoothly to 1. However, increasing the number of receive antennas can make the outage probability approach the limit. When comparing these receivers, we also find that MRC is much more resilient to hardware impairments than ZF and MMSE receivers, while ZF receivers are the most sensitive to hardware impairments. For instance, for $N_r = N_t = 5$, the outage probability $10^{-2}$ of ZF receivers exhibits a power penalty of $11$dB, whilst for MMSE and MRC receivers this value is $4$dB and $0$dB, respectively. ![Outage probability of linear receivers for different SINR threshold $x$.[]{data-label="fig:CDFs"}](CDFs.pdf){height="50.00000%"} Figure \[fig:Rate\_LinRec\_All\] illustrates the ergodic achievable rates with ZF, MRC and MMSE linear receivers and optimized $T_p^*$. The information-theoretic capacity, which can be approached by optimal receivers (OR), are generated according to our previous work [@xinlin2014training_ICC]. These curves provide upper bounds on the achievable rates obtained by linear receivers. The OR structure can be a maximum-likelihood receiver with joint decoding, or a MMSE receiver with successive interference cancellation (SIC). These curves are used as performance benchmarks. In particular, we consider a MIMO system with $N_t = N_r = 4$ and coherence block $T=200$. The theoretical curves are plotted according to the ergodic achievable rate expressions in (\[eq:anal\_ZF\_rate\])-(\[eq:anal\_MMSE\_rate\]), while the numerical curves are plotted based on Monte-Carlo simulations. Clearly, there is an exact agreement between analytical and numerical results, thereby validating the correctness of the analytical expressions. At low SNRs, RTRI have negligible impact on the achievable rates for all receivers considered herein, while at high SNRs, RTRI decrease the achievable rates of ZF and MMSE receivers and generate finite achievable rate ceilings; however, RTRI do not have substantial impact on the performance of MRC receivers in this regime, due to the dominating inter-stream interference. As anticipated, MMSE receivers perform the best across the whole SNR regime. In contrast to ideal hardware systems, we can see that the performance gap between optimal receivers and MMSE/ZF receivers in hardware-impaired systems diminishes as SNR increases. In other words, for hardware-impaired systems, MMSE and ZF receivers become near optimal in the high SNR regime. With the same system setup, we demonstrate the optimal training length $T_p^*$ for systems with different linear receivers in Fig. \[fig:Tp\_LinRec\_All\], $T_p^*$ values for optimal receivers are drawn according to [@xinlin2014training_ICC Section IV-C]. For low SNR values, considerable amount of training is needed. As we have proved in Section \[sec:RateLow\], $T_p^*$ converges to $\frac{T}{2}$, as $\rho\rightarrow0$. As the SNR increases, ideal hardware systems require less channel training, until the pilot length reaches the minimum possible value $T_p^*=N_t$. However, for systems with RTRI, the conclusions are different. If ZF or MMSE receivers are considered, it is noteworthy that more pilot symbols are needed in the high SNR regime; this phenomenon indicates that, although the effective SINR saturates due to RTRI, increasing the training length can still provide performance gains by improving the estimation accuracy. If MRC receivers are considered, there is no such extra training benefit. This is due to the fact that MRC receivers are dominated by inter-stream interference for both ideal and hardware-impaired systems, such that, the RTRI distortions have only insignificant impact on the system performance. As a result, the optimal training length for hardware-impaired systems does not deviate much from that of ideal hardware systems. Moreover, MMSE and ZF receivers generally require more training, while MRC receivers require less training, in comparison to optimal receivers. ![Ergodic achievable rates with different linear receivers and optimized $T_p$ ($N_t = N_r = 4, T = 200$).[]{data-label="fig:Rate_LinRec_All"}](Rate_LinearAndOR.pdf){height="50.00000%"} ![Optimal training sequence length for linear receivers with various levels of RTRI ($N_t = N_r = 4, T = 200$).[]{data-label="fig:Tp_LinRec_All"}](Tp_LinearAndOR.pdf){height="50.00000%"} Extension to Large-Scale MIMO Systems {#sec:MassiveMIMO} ===================================== Large-scale MIMO systems are considered as a key enabler for next generation wireless systems, since it can offer huge improvements in throughput, as well as, energy efficiency. Moreover, they are very robust against small-scale fading and intentional jamming [@Marzetta2010Massive_TWCOM]. For large-scale MIMO systems with ideal hardware, the training overhead has been analyzed in, for example, [@Love2014Training1_JSTSP; @Love2014Training2_JSTSP], where it has been shown that the training requirement does not necessarily to be increased linearly with the number of antennas. However, large-scale MIMO systems are expected to deploy lower-cost radio frequency (RF) components, which are more prone to hardware impairments. Motivated by this important observation, in this section, we will investigate the impact of RTRI on large-scale systems. We derive deterministic equivalents of the ergodic achievable rates, and thereafter find the optimal training length. We assume that the coherence time $T$ is big enough to support our large dimensional analysis. As we will show later, our deterministic equivalents are accurate even for small number of antennas; therefore the dependence on large $T$ is not essential [@Hoydis2013Massive]. Deterministic Equivalents ------------------------- For each type of linear receiver, we now present the deterministic equivalent approximations of the SINR and achievable rates in the large system limit, i.e., for $N_r, N_t \rightarrow\infty$ with a finite ratio $\beta \triangleq \frac{N_r}{N_t} \in [1, \infty)$.[^11] \[theorem:bothlarge\] As $N_r, N_t \rightarrow\infty$, with a finite ratio $\beta = \frac{N_r}{N_t}\in[1, \infty)$,[^12] the SINRs of ZF, MRC, and MMSE receivers, almost surely converge to the following deterministic equivalents, respectively $$\begin{aligned} \gamma_{\mathrm{ZF}}^{N_r,N_t\rightarrow\infty} &\xrightarrow{a.s.} \bar{\gamma}_{\mathrm{ZF}} \triangleq \frac{\beta-1}{\delta^2(\beta-1)+c_1}\label{eq:deter_ZF},\\ \gamma_{\mathrm{MRC}}^{N_r,N_t\rightarrow\infty} &\xrightarrow{a.s.} \bar{\gamma}_{\mathrm{MRC}} \triangleq \frac{\beta}{1+\delta^2 + c_1+\delta^2\beta} \label{eq:deter_MRC},\\ \gamma_{\mathrm{MMSE}}^{N_r,N_t\rightarrow\infty} &\xrightarrow{a.s.} \bar{\gamma}_{\mathrm{MMSE}} \triangleq \frac{\frac{1}{2c_1}\left(-d + \sqrt{d^2 + \frac{4\beta c_1}{1+\delta^2}} \right)}{1+\delta^2 + \frac{\delta^2}{2c_1}\left(-d + \sqrt{d^2 + \frac{4\beta c_1}{1+\delta^2}} \right)}\label{eq:deter_MMSE},\end{aligned}$$ where $c_1 \triangleq \frac{\rho+\rho\delta^2+1+\bar{\epsilon}}{\rho\bar{\epsilon}}$, $\bar{\epsilon} \triangleq \frac{T_p\rho}{N_t(\rho\delta^2+1)}$, and $d\triangleq\frac{c_1}{1+\delta^2}+1-\beta$. See Appendix \[appendix:bothlarge\]. Theorem \[theorem:bothlarge\] provides us with deterministic approximations of the SINRs in the large-antenna regime, which do not rely on random channel realizations, as long as the ratio $\beta$ is fixed. The above results are quite general and tractable, and as we will show later, are very accurate even for small number of antennas. Note that some practical interesting system setups, for instance, $N_t \ll N_r$ (usually appearing in the multi-user MIMO uplink), can be treated as a special case of Theorem \[theorem:bothlarge\]. In particular, we will give this result in the following corollary. \[corollary:Nr\_Large\] For the case $N_t\ll N_r$, such that $\beta\rightarrow\infty$, the SINRs of ZF, MRC, and MMSE receivers, converge almost surely to the same deterministic equivalent$$\label{eq:Nr_Large} \gamma^{\beta\rightarrow\infty} \stackrel{a.s.}\longrightarrow \bar{\gamma} = \frac{1}{\delta^2}.$$ This corollary is consistent with the asymptotic results in [@Emil2013imp_COML]. From the above equation, we can see that the SINR reaches a limit when $N_t\ll N_r$. This limit, which is only quantified by the level of RTRI, is consistent with the outage probability analysis in (\[eq:cdf\_ZF\])-(\[eq:cdf\_MMSE\]). In this case, increasing the transmit power does not provide any SINR gain. After establishing the asymptotic SINRs, we proceed to derive the achievable rates. Using the continuous mapping theorem [@Van2000Asymptotic Theorem 2.3] and dominated convergence theorem [@Billingsley2008Probability Theorem 16.4], we can show that the achievable rates by different receivers converge to the following deterministic equivalent $$\label{eq:Rate_Asymp} R^{N_r, N_t\rightarrow\infty} \longrightarrow \bar{R}= \left(1-\frac{T_p}{T}\right)N_t\log_2\left(1+\bar{\gamma}\right),$$ where $\bar{\gamma}$ is found in (\[eq:deter\_ZF\]), (\[eq:deter\_MRC\]) and (\[eq:deter\_MMSE\]) for ZF, MRC and MMSE, respectively. Note that one can easily prove that $\bar{R}$ is a strictly concave function in $T_p$ [@Boyd2004Convex]. Then, as proved in [@Jakob2011Training_TSP Theorem 4], optimizing $R^{N_r, N_t\rightarrow\infty}$ is asymptotically identical to optimizing $\bar{R}$. Therefore, solving the following optimization problem will return the desired $T_p^*$: $$\begin{aligned} \label{eq:Optimization} &\textrm{maximize} ~~~\bar{R}^{N_r, N_t\rightarrow\infty}\notag\\ &\textrm{subject to} ~~~N_t\leq T_p<T.\end{aligned}$$ With the results in Theorem \[theorem:bothlarge\] and (\[eq:Rate\_Asymp\]), we can find $T_p^*$, for example, via a line search over different $T_p$ values (e.g., using the bisection method). As for the special case in Corollary \[corollary:Nr\_Large\], it follows readily that the achievable rates of different receivers all converge to $$R^{\beta\rightarrow\infty} \longrightarrow \bar{R} = \left(1-\frac{T_p}{T}\right)N_t\log_2\left(1+\frac{1}{\delta^2}\right).$$ Obviously, in this scenario, $T_p^* = N_t$ channel uses should be devoted for training. Furthermore, the number of transmit antennas should not be larger than half of the coherence time, i.e., $N_t\leq \frac{T}{2}$, otherwise, the channel training overhead will cause a decrease in the achievable rate. Numerical Illustrations ----------------------- Figure \[fig:DeviationAsymp\] compares the deterministic equivalents of the ergodic achievable rates with the ergodic achievable rates we derived in the previous section . For MMSE receivers, we see an excellent match in all cases, even for low dimensional systems. For small number of antennas, the achievable rate approximations of ZF receivers at low SNRs and of MRC receivers at high SNRs have moderate deviations from the true values, while these deviations vanish quickly as $N_r$ increases. Comparing these four plots, we can conclude that, for high SNR values, RTRI always make the convergence to the asymptotic expressions faster, whilst for low SNRs, RTRI slow down the convergence. We can also see that, increasing the ratio $\beta$ always guarantees better convergence. In Fig. \[fig:MassiveTp\], we illustrate the optimal training length $T_p^*$ in the large-antenna regime. We note that in the upper subplot, where the system is “not so massive”, RTRI still create extra training requirements in the high SNR regime. However, when the system is equipped with large receive antenna arrays, the optimal training length is reduced, and the optimal $T_p^*$ may be even less than those of ideal hardware systems. The reason behind this phenomenon is that the improvement provided by estimation saturates very fast, and we can still have certain gain by allocating more time for data transmission. ![Deviation of the achievable rate deterministic equivalents from the ergodic achievable rates ($T = 500$).[]{data-label="fig:DeviationAsymp"}](Convergence.pdf){height="75.00000%"} ![Optimal training length in the asymptotic regime ($T = 500$).[]{data-label="fig:MassiveTp"}](Massive_Tp.pdf){height="55.00000%"} Conclusions {#sec:Conclusion} =========== In this paper, we analyzed the performance of training-based MIMO systems with residual hardware impairments. We derived a LMMSE channel estimator for the proposed model, and deduced the irreducible estimation error floor in the high SNR regime. We derived SINR distributions for ZF, MRC and MMSE linear receivers, and found the corresponding closed-form ergodic achievable rates. We observed that the achievable rates saturated in the high SNR regime. The optimal training lengths were thereafter found through optimizing the ergodic achievable rates. Generally, RTRI imposed more training requirement at high SNRs for MMSE and ZF receivers. However, for MRC receivers, there was no difference in the optimal training length as compared to ideal hardware systems. Moreover, for large-scale systems, we derived deterministic equivalents of the ergodic achievable rates, which were shown to be accurate even for small number of antennas. In this case, we found that by deploying large receive antenna arrays, the extra training requirement imposed by RTRI was alleviated; with sufficiently large number of receive antennas, systems with RTRI needed even less training than ideal hardware systems. The reason behind this phenomenon is that, the improvement offered by channel estimation saturates very fast, while we can still have performance gain by allocating more channel uses to data transmission. Thus, it is important to find a good tradeoff between channel estimation and data transmission. Useful random matrix theory results =================================== \[lemma:MatrixInv\][*(Matrix inversion lemma [@Hoydis2013Massive])*]{} Let $\mathbf A\in\mathbb C^{N\times N}$ be a Hermitian invertible matrix, and let $\mathbf x\in\mathbb C^{N\times 1}$ and $\tau$ be arbitrary vector and scalar, respectively. Then, we have $$\label{eq:matrixinversion} \mathbf x(\mathbf A + \tau\mathbf x\mathbf x^H)^{-1} = \frac{\mathbf x^H\mathbf A^{-1}}{1 + \tau\mathbf x^H\mathbf A^{-1}\mathbf x},$$ \[lemma:TraceLemma\][*(Trace lemma [@Bai2009RMBook Lemma B.26])*]{} Let $\mathbf A\in\mathbb C^{N\times N}$ and $\mathbf x$, $\mathbf y \sim \mathcal{CN}(\mathbf 0, \frac{1}{N}\mathbf I_N)$. Assume that $\mathbf A$ has uniformly bounded spectral norm with respect to $N$, and that $\mathbf x$ and $\mathbf y$ are mutually independent, and independent of $\mathbf A$. Then, as $N\rightarrow\infty$, we have $$\begin{aligned} \mathbf x^H\mathbf A\mathbf x-\frac{1}{N}\mathrm{tr}\left(\mathbf A\right)\stackrel{a.s.}\longrightarrow 0\\ \mathbf x^H\mathbf A\mathbf y \stackrel{a.s.}\longrightarrow 0.\end{aligned}$$ \[lemma:Rank1Lemma\][*(Rank-1 perturbation lemma [@Bai1995Empirical])* ]{} Let $z < 0$, $\mathbf A, \mathbf B\in\mathbb C^{N\times N}$ with $\mathbf B$ Hermitian nonnegative definite, and $\mathbf v\in\mathbb{C}^N$, then $$\left| \mathrm{tr}\left((\mathbf B - z\mathbf I_N)^{-1} - (\mathbf B + \mathbf v\mathbf v^H - z\mathbf I_N)^{-1}\right)\mathbf A\right|\leq\frac{||\mathbf A||_2}{|z|}.$$ \[lemma:STLemma\] Let $\mathbf A\in\mathbb C^{N\times n}$, denote $F^{\mathbf A^H\mathbf A}$ as the eigenvalue CDF of $\mathbf A^H\mathbf A$, and $m_{F^{\mathbf A^H\mathbf A}}$ as the Stieltjes transform of $F^{\mathbf A^H\mathbf A}$ . Then, for $z\in\mathbb C\backslash \mathbb R^+$[^13], we have [@Couillet2011a Lemma 3.1] $$\frac{n}{N}m_{F^{\mathbf A^H\mathbf A}}(z) = m_{F^{\mathbf A\mathbf A^H}}(z) + \frac{N-n}{N}\frac{1}{z}.$$ CDF of the SINR with ZF receivers {#appendix:zf_cdf} ================================= Denoting $x \triangleq 1/{\left[\left(\mathbf H^H \mathbf H\right)^{-1}\right]_{k,k}}$, we know from [@winters1994diversity] that $x$ follows a chi-square distribution with $2(N_r-N_t+1)$ degrees of freedom and is scaled by $1\slash2$, with its CDF given by $$F_X(x) = 1-e^{-x}\sum\limits_{k=0}^{N_r-N_t}\frac{x^k}{k!}.$$ Rewrite the SINR of ZF receiver as $$\gamma_{\mathrm{ZF}} = \frac{X}{\delta^2X+c_0},$$ then we finally have $$F_{\Gamma_{\mathrm{ZF}}}(\gamma) = \Pr(\Gamma_{\mathrm{ZF}}\leq\gamma) = F_X\left(\frac{c_0\gamma}{1-\delta^2\gamma}\right).$$ CDF of the SINR with MRC receivers {#appendix:mrc_cdf .unnumbered} ================================== First, we rewrite $\gamma_{\mathrm{MRC}}$ as $$\begin{aligned} \gamma_{\mathrm{MRC}} &= \frac{\bar{\mathbf h}_k^H\bar{\mathbf h}_k}{(1+\delta^2)\sum\limits_{i=1,i\neq N_t}^{N_t}\frac{|\bar{\mathbf h}_k^H\bar{\mathbf h}_i|^2}{\bar{\mathbf h}_k^H\bar{\mathbf h}_k} + \delta^2\bar{\mathbf h}_k^H\bar{\mathbf h}_k +c_0}\notag\\ & = \frac{Z}{(1+\delta^2)Y+\delta^2Z+c_0},\end{aligned}$$ where $Z$ is chi-square distributed with $2N_r$ degrees of freedom, and $Y$ is chi-square distributed with $2(N_t-1)$ degrees of freedom with PDF [@Shah1998MRC] $$f_Y(y) = \frac{e^{-y}y^{N_t-2}}{(N_t-2)!}.$$ We can now express the CDF of the SINR with MRC as $$\begin{aligned} F_{\Gamma_{\mathrm{MRC}}}(\gamma) &= \Pr(\Gamma_{\mathrm{MRC}}\leq\gamma)\\ &=\int_0^{\infty}F_Z\left(\frac{(1+\delta^2)\gamma}{1-\delta^2\gamma}y+\frac{c_0\gamma}{1-\delta^2\gamma}\right)f_Y(y)dy\notag\\ & = 1 - \frac{e^{-\frac{c_0\gamma}{1-\delta^2\gamma}}}{(N_t-2)!}\sum\limits_{k=0}^{N_r-1}\frac{\left(\frac{c_0\gamma}{1-\delta^2\gamma}\right)^k}{k!}\sum\limits_{p=0}^{k}{k\choose p}\left(\frac{1+\delta^2}{c_0}\right)^p\int_{0}^{\infty}y^{N_t+p-2}e^{-\frac{1+\gamma}{1-\delta^2\gamma}y}dy\label{eq:MRC_CDF_int}\\ & = 1-\frac{e^{-\frac{c_0\gamma}{1-\delta^2\gamma}}}{(N_t-2)!}\sum\limits_{k=0}^{N_r-1}\frac{\left(\frac{c_0\gamma}{1-\delta^2\gamma}\right)^k}{k!}\sum\limits_{p=0}^{k}{k\choose p}\left(\frac{1+\delta^2}{c_0}\right)^p\frac{(N_t+p-2)!}{\left(\frac{1+\gamma}{1-\delta^2\gamma}\right)^{N_t+p-1}}\\ & = 1-\frac{e^{-\frac{c_0\gamma}{1-\delta^2\gamma}}}{\left(\frac{1+\gamma}{1-\delta^2\gamma}\right)^{N_t-1}}\sum\limits_{k=0}^{N_r-1}\sum\limits_{p=0}^{k}\underbrace{\frac{{{N_t+p-2}\choose p}}{(k-p)!}\left(\frac{1+\delta^2}{c_0}\right)^p}_{\triangleq{\alpha_{p,k}}}\frac{\left(\frac{c_0\gamma}{1-\delta^2\gamma}\right)^k}{\left(\frac{1+\gamma}{1-\delta^2\gamma}\right)^{p}}.\end{aligned}$$ The integral in \[eq:MRC\_CDF\_int\] can be evaluated via the identity [@gradshteyn2007a Eq. (3.381.4)]. CDF of the SINR with MMSE receivers {#appendix:mmse_cdf .unnumbered} =================================== We start from the equivalent form in (\[eq:MMSE\_SINR\_trasform1\]), and have $$\gamma_{\mathrm{MMSE}} = \frac{W}{1+\delta^2+\delta^2W},$$ where $W \triangleq {{\bar {\mathbf h}_k}^H\left(\sum_{i=1,i\neq k}^{N_t}{\bar {\mathbf h}_i}{\bar {\mathbf h}_i}^H + \frac{c_0}{1+\delta^2} \mathbf I_{N_r} \right)^{-1}{\bar {\mathbf h}_k}}$ is a classical MMSE SINR expression, with CDF given by [@Gao1998MMSE Eq. 11] and [@Mckay2009LinearReceiver Eq. 8] $$F_W(w) = 1- e^{-\frac{c_0x}{1+\delta^2}}\sum\limits_{k=0}^{N_r-1}\frac{A_k(w)}{k!}\left(\frac{c_0w}{1+\delta^2}\right)^k,$$ where $$A_k(w) \triangleq \left\{ \begin{array}{rl} &1, ~~~~~~~~~\mbox{ if $N_r\geq N_t+k$}, \\ &\frac{1+\sum\limits_{i=1}^{N_r-k-1}{N_t-1\choose i}x^i}{\left(1+x\right)^{N_t-1}}, \mbox{ otherwise.} \end{array} \right.$$ Using a similar approach as in the above proofs, followed by some simple manipulations, we obtain the final result. Proof of Theorem \[theorem:anal\_rates\] {#appendix:anal_rates} ======================================== Achievable rate of ZF receivers ------------------------------- Firstly, we can easily prove that the SINR should be in the range $(0,\frac{1}{\delta^2})$, which also holds for MRC and MMSE receivers. Therefore, substituting (\[eq:cdf\_ZF\]) into (\[eq:cdf\_rate\]), we have $$\begin{aligned} R_\mathrm{ZF} &= \frac{T_dN_t}{\ln(2)T}\int_0^{\frac{1}{\delta^2}} \frac{e^{-\frac{c_0\gamma}{1-\delta^2\gamma}}}{1+\gamma}\sum_{k=0}^{N_r-N_t}\frac{\left(\frac{c_0\gamma}{1-\delta^2\gamma}\right)^k}{k!}d\gamma\\ &\overset{(a)}=\frac{T_dN_t}{\ln(2)T}\sum_{k=0}^{N_r-N_t}\left(\frac{c_0}{\delta^2}\right)^k\int_0^{\infty}\frac{e^{-x}x^k}{(1+x)(\delta^2+(\delta^2+1)x)}dx\label{eq:anal_ZF_step2}\\ &\overset{(b)}= \frac{T_dN_t}{\ln(2)T}\sum_{k=0}^{N_r-N_t}\left(\frac{c_0}{\delta^2}\right)^k\left(\int_0^{\infty}\frac{e^{-x}x^k}{x+\frac{\delta^2}{\delta^2+1}}dx - \int_0^{\infty}\frac{e^{-x}x^k}{x+1}dx \right),\label{eq:anal_ZF_step3}\end{aligned}$$ where $(a)$ is obtained by a change of variables $\frac{c_0\gamma}{1-\delta^2\gamma}\rightarrow x$ and exchanging the order of integration and summation, and $(b)$ is obtained by expanding the integrand in (\[eq:anal\_ZF\_step2\]). Finally, applying [@gradshteyn2007a Eq. (3.383.10)] in (\[eq:anal\_ZF\_step3\]) and the fact that $E_n(z)=z^{n-1}\Gamma(1-n,z)$, where $\Gamma(n,z)$ is the upper incomplete gamma function [@gradshteyn2007a Eq. (8.350.2)], concludes the proof. The achievable rate of MMSE receivers can be deduced in a similar way. Achievable rate of MRC receivers -------------------------------- Substituting the SINR with MRC in (\[eq:cdf\_ZF\]) into (\[eq:cdf\_rate\]), and applying the change of variables $\frac{c_0\gamma}{1-\delta^2\gamma}\rightarrow x$, as well as changing the order of summation and integration, we have $$\label{eq:anal_MRC_step1} R_\mathrm{MRC} = \frac{T_dN_t}{\ln(2)T}\sum_{k=0}^{N_r-1}\sum_{p=0}^{k}\alpha_{p,k}\left(\frac{c_0}{\delta^2+1}\right)^{p+N_t}\frac{1}{\delta^2}\int_0^{\infty}\frac{e^{-x}x^k}{\left(x + \frac{c_0}{\delta^2}\right)\left(x+\frac{c_0}{\delta^2+1}\right)^{p+N_t}}dx.$$ Note that for any $a, b\in\mathbb C, a\neq b$, we can easily prove the following identity: $$\frac{1}{(x+a)(x+b)^n} = \frac{(b-a)^{-n}}{x+a} - \sum_{i=1}^{n}\frac{(b-a)^{-i}}{(x+b)^{n-i+1}}.$$ Applying this result to (\[eq:anal\_MRC\_step1\]), and using the integral identity in [@Prudnikov1986integrals Eq. (2.3.6.9)] followed by some algebraic manipulations, we arrive at the desired result. Proof of Theorem \[theorem:bothlarge\] {#appendix:bothlarge} ====================================== Deterministic equivalent of ZF SINR ----------------------------------- The singular value decomposition (SVD) of $\bar{\mathbf H}$ is denoted by $\bar{\mathbf H} = \mathbf U\boldsymbol\Sigma\mathbf V^H$, where $\mathbf U$ and $\mathbf V$ are unitary matrices, while $\boldsymbol \Sigma$ is a $N_r\times N_t$ diagonal matrix with the diagonal elements containing the singular values of $\bar{\mathbf H}$. Consequently, we have $$\begin{aligned} \left[\left(\bar {\mathbf H}^H\bar {\mathbf H}\right)^{-1}\right]_{k,k} \!= \left[\mathbf V\boldsymbol\Lambda\mathbf V^H\right]_{k,k} = \mathbf v_k^H\boldsymbol\Lambda\mathbf v_k \underset{(a)}{\stackrel{a.s.}\longrightarrow} \frac{1}{N_t}\textrm{tr}\left(\boldsymbol\Lambda\right) \stackrel{a.s.}\longrightarrow \frac{1}{N_r\!-\!N_t},\end{aligned}$$ where $\boldsymbol\Lambda$ is a diagonal matrix with its diagonal elements being the eigenvalues of $\left(\bar {\mathbf H}^H\bar {\mathbf H}\right)^{-1}$. Note that $\mathbf v_k$ is known to satisfy the conditions of $\mathbf x$ in Lemma \[lemma:TraceLemma\] [@Eldar2003ZFAsym_TIT], therefore $(a)$ follows readily. Substituting the above result into the ZF SINR expressions in (\[eq:SINR\_ZF\]) completes the proof. Deterministic equivalent of MRC SINR ------------------------------------ Rewriting the SINR of MRC, we have $$\begin{aligned} \gamma_{k,\mathrm{MRC}} &= \frac{\mid\bar{\mathbf h}_k^H\bar{\mathbf h}_k\mid^2}{\bar{\mathbf h}_k^H\left( (1+\delta^2)\bar{\mathbf H}\bar{\mathbf H}^H - \bar{\mathbf h}_k\bar{\mathbf h}_k^H + c_0\mathbf I_{N_r} \right)\bar{\mathbf h}_k}\\ &=\frac{\mid\bar{\mathbf h}_k^H\bar{\mathbf h}_k\mid^2}{(1+\delta^2)\bar{\mathbf h}_k^H\left( \sum_{i=1, i\neq k}^{N_t}\bar{\mathbf h}_i\bar{\mathbf h}_i^H+ \frac{c_0}{1+\delta^2}\mathbf I_{N_r} \right)\bar{\mathbf h}_k + \delta^2\mid\bar{\mathbf h}_k^H\bar{\mathbf h}_k\mid^2}\\ &\underset{(a)}{\stackrel{a.s.}\longrightarrow} \frac{1}{\frac{1+\delta^2}{N_r^2}\mathrm{tr}\left( \sum_{i=1, i\neq k}^{N_t}\bar{\mathbf h}_i\bar{\mathbf h}_i^H+ \frac{c_0}{1+\delta^2}\mathbf I_{N_r} \right) + \delta^2}\\ &\underset{(b)}{\stackrel{a.s.}\longrightarrow}\frac{\beta}{1+\delta^2 + c_1 + \delta^2\beta},\end{aligned}$$ where $(a)$ and $(b)$ are obtained by applying Lemma \[lemma:TraceLemma\]. Deterministic equivalent of MMSE SINR ------------------------------------- First, rewrite the MMSE SINR expression in (\[eq:MMSE\_SINR\_trasform1\]) as $$\begin{aligned} \gamma_{k,\mathrm{MMSE}} = \frac{\tilde\gamma_{k,\mathrm{MMSE}} }{1+\delta^2+\delta^2\tilde\gamma_{k,\mathrm{MMSE}} },\label{eq:MMSE_SINR2}\end{aligned}$$ where $\tilde\gamma_{k,\mathrm{MMSE}}\triangleq {\bar {\mathbf h}_k}^H\left(\sum_{i=1,i\neq k}^{N_t}{\bar {\mathbf h}_i}{\bar {\mathbf h}_i}^H + \frac{c_0}{1+\delta^2} \mathbf I_{N_r} \right)^{-1}{\bar {\mathbf h}_k}$. As $N_r, N_t \rightarrow \infty$ with $\beta = \frac{N_r}{N_t} \in [1, \infty)$, we have $$\begin{aligned} \tilde\gamma_{k,\mathrm{MMSE}} {\stackrel{a.s.}\longrightarrow} m_{F_{\bar{\mathbf H}^H\bar{\mathbf H}}}\left(\frac{c_0}{1+\delta^2}\right) &= \mathrm{tr}\left(\left(\bar{\mathbf H}\bar{\mathbf H}^H + \frac{c_0}{1+\delta^2}\mathbf I_{N_r}\right)^{-1}\right)\label{eq:STtrans1}\\ &\overset{(a)}{=} \frac{1+\delta^2}{c_0}(N_r-N_t) + \mathrm{tr}\left(\left(\bar{\mathbf H}^H\bar{\mathbf H} + \frac{c_0}{1+\delta^2}\mathbf I_{N_t}\right)^{-1}\right)\notag\\ &\overset{(b)}{=} \frac{1+\delta^2}{c_0}(N_r-N_t) + \sum_{i=1}^{N_t}\frac{1+\delta^2}{c_0(1 + \bar{\mathbf h}_i^H\left(\bar{\mathbf H}\bar{\mathbf H}^H + \frac{c_0}{1+\delta^2}\mathbf I_{N_r}\right)^{-1}\bar{\mathbf h}_i)}\notag\\ &\underset{(c)}{\stackrel{a.s.}\longrightarrow}\frac{1+\delta^2}{c_0}(N_r-N_t) + \frac{(1+\delta^2)N_t}{c_0(1+m_{F_{\bar{\mathbf H}^H\bar{\mathbf H}}}(\frac{c_0}{1+\delta^2}))},\label{eq:STtrans4}\end{aligned}$$ where $F_{\bar{\mathbf H}^H\bar{\mathbf H}}$ is the eigenvalue CDF of $\bar{\mathbf H}\bar{\mathbf H}^H$, and $m_{F_{\bar{\mathbf H}^H\bar{\mathbf H}}}\left(\frac{c_0}{1+\delta^2}\right)$ is the Stieltjes transform of $F_{\bar{\mathbf H}^H\bar{\mathbf H}}$ with argument $\frac{c_0}{1+\delta^2}$ [@Couillet2011a], $(a)$ is obtained by using Lemma \[lemma:STLemma\], $(b)$ is obtained by using the matrix inversion identity (\[eq:matrixinversion\]) and Lemma \[lemma:Rank1Lemma\], while $(c)$ is obtained by using Lemma \[lemma:TraceLemma\]. Combining (\[eq:STtrans1\]) and (\[eq:STtrans4\]), we can easily obtain $$\tilde\gamma_{k,\mathrm{MMSE}} {\stackrel{a.s.}\longrightarrow} m_{F_{\bar{\mathbf H}^H\bar{\mathbf H}}}\left(\frac{c_0}{1+\delta^2}\right) = \frac{1}{2c_1}\left(-d + \sqrt{d^2 + \frac{4\beta c_1}{1+\delta^2}} \right).$$ Substituting this result into (\[eq:MMSE\_SINR2\]) and applying the continuous mapping theorem [@Van2000Asymptotic Theorem 2.3] concludes the proof. [^1]: X. Zhang is with the Department of Signals and Systems, Chalmers University of Technology, 412 96, Gothenburg, Sweden (email: [email protected]). [^2]: M. Matthaiou is with the School of Electronics, Electrical Engineering and Computer Science, Queen’s University Belfast, Belfast, U.K. and with the Department of Signals and Systems, Chalmers University of Technology, 412 96, Gothenburg, Sweden (email: [email protected]). [^3]: M. Coldrey is with Ericsson Research Gothenburg, Sweden (email: [email protected]). [^4]: E. Björnson is with the Department of Electrical Engineering (ISY), Linköping University, Linköping, Sweden (email: [email protected]). [^5]: Parts of this paper were published at the IEEE International Conference on Communications (ICC), Sydney, Australia, June 2014 [@xinlin2014training_ICC]. [^6]: Residual impairments refer to the aggregate noise that is left after properly compensating the hardware-impaired transmit signals in each transmission step. Since it is conventionally assumed that the transmitted symbols across different channel uses are independent, it is also reasonable to assume that the residual impairments are also independent over each symbol. The paper \[11\] provides experimental justification of this assumption. More generally speaking (not restricted to RTRI), the effects of transceiver distortions, such as clipping in power amplifier nonlinearities, quantization errors, can be well-approximated as memoryless functions \[16\], thus they are also independent over symbols/time. [^7]: Note that we choose $\mathbf S$ to be full rank, i.e., all $N_t$ antennas are used to transmit signals. Although this is inefficient for the low SNR regime, it is a meaningful scheme when no CSI at the transmitter (CSIT) is available [@Jorswieck2004Transmission_TSP]. [^8]: Due to the multiplicative term $\mathbf H\boldsymbol\Delta_p$, which appears in the estimated channel $\hat{\mathbf H}$, $\bar{\mathbf H}$ is not exactly Gaussian distributed. However, we can verify easily through simulations that this it can be tightly approximated as Gaussian distributed even for high levels of RTRI. [^9]: Although our analytical results are based on the approximation mentioned in Footnote 3, we henceforth use equality signs without significant notational abuse. Note that our simulation results indicate that the rate approximations remain remarkably tight for a plethora of system parameters. [^10]: The outage probability per stream is always the lowest when transmitting only one stream, however, it renders a lower average throughput. Thus, in the following numerical analysis, we use all $N_t$ antennas for capacity-approaching transmission. Note that, for the case which only $N_t'<N_t$ antennas are used for transmission, we can evaluate the outage probability by substituting $N_t = N_t'$ into Proposition \[prop:cdf\]. [^11]: Since the SINRs of each stream are statistically independent, we drop the dependence on $k$ for the sake of clarity. [^12]: For ZF receivers, we restrict $\beta$ to be in the range $(1, \infty)$. For $\beta =1$, one can show that $\bar{\mathbf H}^H\bar{\mathbf H}$ has at least one eigenvalue converging to zero [@Bai1993Smallest], which makes the ZF analysis intractable. Due to page limit, we ignore the case of $\beta = 1$, one can find a detailed study of this case, for example, in [@Eldar2003ZFAsym_TIT]. [^13]: The notation $z\in\mathbb C\backslash \mathbb R^+$ denotes all $z\in\mathbb C, z \not\in\mathbb R^+$.

ArXiv

--- author: - 'D. Heath Jones, Ben W. Stappers' - 'Bryan M. Gaensler' date: 'Received ; accepted ' title: 'Discovery of an optical bow-shock around pulsar B0740$-$28[^1]' --- Introduction ============ Pulsar wind nebulae (PWN) around high-velocity pulsars provide a primary insight into the interaction between a pulsar and its environment. Specifically, optical observations of such nebulae provide important information on pulsar velocities, and on the density, temperature and composition of the ambient medium. However only three pulsars are known to power optical bow-shock PWNe: B2224$+$65 (the ‘Guitar Nebula’, Cordes, et al. [@cordes93]), and the two millisecond pulsars (MSPs) B1957$+$20 (Kulkarni & Hester [@kulkarni88]) and J0437$-$4715 (Bell et al. [@bell95]). All three of these pulsars have high spin-down luminosities and/or high space velocities. However, these pulsars differ markedly in their spin-periods, ages and magnetic field strengths, highlighting the variety of pulsar winds which can be probed by these sources. The nebula associated with the neutron star RX J1856.5$-$3754 further exemplifies the variety of neutron stars known to power such nebulae (van Kerkwijk & Kulkarni [@vankerkwijk01]). We have therefore initiated a search for optical bow-shocks around other pulsars in order to characterise the properties of the associated pulsars, pulsar winds and ambient environments. --------------------------------- --------------------------- ---------------------- Spin Period$^a$ P (s) 0.1667 Period Derivative$^a$ $\dot{P}$ 1.68$\times10^{-16}$ Spin-down Luminosity$^a$ $\dot{E}$ (ergs s$^{-1}$) 1.4$\times10^{35}$ Dispersion Measure Distance$^b$ d$_{\rm DM}$ (kpc) 1.9$^{+0.4}_{-0.3}$ H[i]{} Distance$^c$ d$_{\rm HI}$ (kpc) 1.4 - 7.7 Proper Motion$^d$ $\mu$ () $29 \pm 1$ Position Angle$^d$ $\theta$ ($^\circ$) $269 \pm 1$ --------------------------------- --------------------------- ---------------------- : Properties of PSR B0740$-$28[]{data-label="physParams"} $(a)$ Arzoumanian et al. ([@arzoumanian94]). $(b)$ Taylor & Cordes ([@taylor93]). $(c)$ Koribalski et al. ([@koribalski95]). $(d)$ Bailes et al. ([@bailes90]).\ During the first two nights of this programme (January 4 and 5, 2001), we discovered an optical bow-shock nebula around the radio pulsar B0740$-$28. Discovered at Bologna (Salter [@salter70]), this pulsar was subsequently shown to be rapidly spinning down (McCulloch et al.[@mcculloch73]). Its correspondingly high spin-down luminosity, $\dot{E}=4\pi I\dot{P}/P^3$, (where $I=10^{45}I_{45}$ g cm$^{2}$ is the pulsar moment of inertia; see Table \[physParams\]), but likely moderate distance and high transverse velocity make it a promising target for powering an observable  nebula. We note that the maximum distance derived from  measurements given in Table \[physParams\] is large, but in Section 5 we argue for a distance closer to that derived from the dispersion measure. See Chatterjee & Cordes ([@chatterjee02]) for a recent review. ------------------- ------------------- ----------------- Date 4, 5 January 2001 14 April 2001 Instrument SUSI-2/NTT TTF$^a$/AAT Pixel Scale () 0.161 0.373 Filter$^b$ (nm) 656/7 656/2 Exposure Time (s) 6480 2700 Transparency photometric photometric Mean Seeing () $\sim 0.9$ $\sim 1.4$ Continuum Frame None 600 s in $R1^c$ ------------------- ------------------- ----------------- : Log of Observations for the PSR B0740$-$28 field.\[table2\][]{data-label="observations"} $(a)$ using the blue Fabry-Perot and $R0$ (668/21) filter.\ $(b)$ central wavelength and FWHM in nanometers.\ $(c)$ TTF $R1$ filter (707/26).\ Observations and Reduction ========================== The nebula around PSR B0740$-$28 was discovered in narrowband frames taken through the 656/7 nm filter of SUSI-2 at the 3.5 m New Technology Telescope (NTT), La Silla (Fig. \[fig1\]$a$). A separate continuum-subtracted narrowband frame (Fig. \[fig1\]$b$) was derived from subsequent imaging with the Taurus Tunable Filter (TTF; Bland-Hawthorn & Jones 1998) at the 3.9 m Anglo-Australian Telescope (AAT). Figure 1($c$) shows the NTT image in Fig. \[fig1\]($a$) smoothed with a $5 \times 5$-pixel boxcar kernel. The observations are summarised in Table \[table2\]. Raw frames were treated using standard techniques for bias and flat-field correction, image registration and co-addition. CCD defects and cosmic rays were removed using cross-pixel interpolation on the final frame, in order to preserve as much of the faint nebula signal as possible. {width="6cm"} {width="6cm"} {width="6cm"} {width="6cm"} An astrometric solution was derived for the higher-resolution NTT image using the [iraf]{} task [*ccmap*]{}. A Legendre polynomial with 11 coefficients was found to give the best fit in $x$ and $y$ to 24 USNO Catalog reference stars across the field. The fitting option used in [ccmap]{} sees a linear term (with coefficients dependent on translation, scaling, rotation and skew), computed separately from a distortion term, which is a polynomial fit to the residuals of the linear term. No distortion residuals were evident in a test transformation of the USNO star coordinates to those of the image. The RMS scatter was $(\Delta \alpha, \Delta \delta)$ = ($\pm 0.028$,$\pm 0.038$) and uniform across the field. The location of PSR B0740$-$28 was determined by applying the proper motion determined by Bailes et al. (1990) to the position measured by Fomalont et al. ([@fomalont92]). This places the pulsar at $(\alpha, \delta) = (07$:$42$:$49.041, -28$:$22$:$44.02)$ (J2000) at the mean time of observation, 13 inside the leading edge of the nebula on its long axis (Fig \[fig1\]). It is not coincident with the prominent star near the head of the nebula. Bailes et al. (1990) determine a proper motion position angle (measured through east) of $\theta = 269\pm1^\circ$, [*i.e.*]{} due west. This is consistent with the direction of the motion implied by the symmetry axis of the nebula and confirms the association. Flux and Morphology =================== The optical emission from pulsar wind nebulae is purely Balmer. The  flux density in the NTT image was calibrated using observations of the planetary nebula flux standards 235.3$-$03.9 and 278.6$-$06.7 (Dopita & Hua [@dopita97]). Atmospheric extinction was estimated using the CTIO measurements of Stone & Baldwin ([@stone83]). For the planetary nebula standards, the 656/7 filter admits flux from the adjacent  and   lines, in addition to . The contribution for 235.3$-$03.9 (37%) is more significant than 278.6$-$06.7 (15%), although the 656/7 fluxes were corrected for it in both cases. The  flux over the entire PSR B0740$-$28 nebula is $5 \times 10^{-5}$ , around two orders of magnitude fainter than the nebulae powered by J0437$-$4715 (Bell et al. [@bell95]) and B1957$+$20 (Kulkarni & Hester [@kulkarni88]). The  images in Fig. \[fig1\] show a distinctive key-hole shape which can be divided into three main regions. The most western component, or “head”, has an almost circular shape, with the pulsar located at the leading edge. Approximately 20 from the front of the shock, the nebula broadens into a fan-shaped tail and a bridge of emission connecting the two sides of the nebula can be seen. At the eastern limit of emission, some 45 from the apex, the nebula appears to show a boundary with opposite concavity to that of the apex, closing the tail. Figure 1($b$) shows no emission within the interior of the head although deeper images are needed to confirm this. Variations in the surface brightness across the extent of the nebula are seen in Fig. 1$(a)$ and Fig. 1$(b)$; see Table \[fluxes\] for a summary. The knot at Region A, located at the juncture of the “fan” and “head” section, and its southern counterpart both have surface brightness greater than that of the overall nebula. Nebular Geometry ================ When a star moves at a supersonic velocity with respect to the interstellar medium (ISM), a bow-shock forms at the interface between the stellar wind and the ambient medium. For pulsars the wind is relativistic. Assuming an isotropic wind, the stand-off distance, $r_{\rm w}$, separating the pulsar from the contact discontinuity of the shock, can be derived from the balance between ram pressure and the outflowing wind pressure, $$r_{\rm w} = \left(\frac{\dot{E}}{4\pi c \rho_0v_{\rm p}^{2}}\right)^{1/2}~{\rm cm} \label{shock_radius}$$ where $\rho_0 = 2.2\times10^{-24}n_{\rm H}$gcm$^{-3}$ is the ISM density, $n_{\rm H}$ is the hydrogen number density and $v_{\rm p}$ (cm s$^{-1}$) is the space velocity of the pulsar. The shock apex was determined by finding the peak emission lying along the pulsar proper motion direction and is located 13 from the pulsar. However this does not correspond to $r_{\rm w}$, as the Balmer emission comes from a thin shell upstream from the contact discontinuity. Aldcroft et al. ([@aldcroft92]) find for B1957+20 that the measured pulsar-apex separation is 30% larger than $r_{\rm w}$. Hence we adopt $r_{\rm w} = $10 $= 0.005 d_{\rm kpc}$ pc, where $d_{\rm kpc}$ is the pulsar distance in kpc. The proper motion, when corrected for galactic rotation, corresponds to a projected space velocity of $v_{\rm p} = 1.3\times10^7d_{\rm kpc}/{\rm cos}i$ , where $i$ is the angle of this velocity vector to the plane of the sky. Substituting $v_{\rm p}$ into Eqn. (\[shock\_radius\]) and rearranging gives: $$n_{\rm H} = \frac{4{\rm cos}^4i~I_{45}}{d_{\rm kpc}^4}~{\rm cm}^{-3}. \label{nh}$$ Using the upper and lower limits to the distance given in Table \[physParams\] and assuming that $i$ is small then $0.002 \simlt n_{\rm H} \simlt 1$ cm$^{-3}$. If we assume the dispersion measure distance $d_{\rm kpc} = 2$, (which we argue for in Section 5), then the density is n$_{\rm H} \sim 0.25$ cm$^{-3}$. ----------------- ------------------------------------- Region Surface Brightness (see Fig. 1$a$) (ergs$^{-1}$cm$^{-2}$arcsec$^{-2}$) A 7$\times10^{-19}$ B 4$\times10^{-19}$ C 5$\times10^{-19}$ entire nebula 3$\times10^{-19}$ ----------------- ------------------------------------- : Mean Surface Brightness of Different Regions[]{data-label="fluxes"} Recent derivations of the geometry of thin momentum-balance bow-shocks give a generic form for the shock that scales with stand-off distance (Chen et al. [@chen96]; Wilkin [@wilkin96]). We have compared the shape of the PSR B0740$-$28 nebula with the expression of Chen et al. ([@chen96]), $$\frac{x}{r_{\rm w}} + 1 = \frac{3}{10}\biggl( \frac{y}{r_{\rm w}} \biggr)^2 + \biggl( \frac{3}{280}\frac{y}{r_{\rm w}} \biggr)^4. \label{chen}$$ Here, our axes lie in the plane of the sky, such that the $y$-axis lies along the proper motion direction, $x$ is perpendicular to it and the origin is located at the pulsar. The best fit to 14 locations around the “head” gives a stand-off distance of 12$\pm 0.1$ and proper motion position angle of $276 \pm 3^{\circ}$ which are consistent with the measured values. However some deviation from an exact fit occur on the southern edge of the shock front. Modelling by Wilkin ([@wilkin00]) shows that either wind asymmetry or density gradients in the ISM are able to reproduce such variations. Observations of PWNe around other pulsars have demonstrated marked anisotropies in the wind flow away from the pulsar, (Hester [@hester98]; Gaensler et al. [@gaensler02]), while large variations in the ISM density are seen in the region of the Guitar nebula (Chatterjee & Cordes [@chatterjee02]). Discussion\[discussion\] ======================== Cordes (1996) shows that the  flux density in the nebula can be expressed in terms of observables as, $$F_{\alpha} = \frac{10^{-2.47}Xn_{\rm H}v_7^3\theta_s^2}{{\rm cos}^5i}~{\rm photons\,cm}^{-2}\,{\rm s}^{-1}, \label{flux}$$ where $X$ is the neutral fraction, $v_7 = v_{\rm p}/10^7$ and $\theta_{\rm s}$ is the observed angular distance. The $2$ kpc distance derived from the dispersion measure (Taylor & Cordes [@taylor93]) is consistent with the amount of reddening and closed shape of the nebula, in the following ways. The observed reddening in the direction of PSR B0740$-$28 is $A_{\rm v} \approx 1.5$ magnitudes at $d \approx 2$ kpc and distances beyond, corresponding to a reduction in  flux by a factor of $\sim 3$ (Neckel & Klare [@neckel80]). Combining this with a neutral fraction $X = 0.01$, (reasonable for a warm ionised ISM with the n$_{\rm H} \sim 0.25$ cm$^{-3}$ calculated earlier from Eqn. \[nh\] with $d_{\rm kpc} = 2$), gives good agreement with the observed flux. At larger distances the density derived from Eqn. (\[nh\]) implies the pulsar is in the hot phase of the ISM where the neutral fraction is very small, thereby further reducing the predicted flux. The direction of the nebula with respect to the plane of the sky will also affect the measured flux. Using the shock morphology formula given in Eqn. (\[chen\]), we generated a three-dimensional model of the shock and then determined its shape projected on the plane of the sky for various values of $i$. We find that to best match the observed stand-off distance and opening angle $i$ must be small ($\simlt 25^{\circ}$) and thus its contribution to the low flux is small. The overall morphology of the PSR B0740$-$28 nebula is very similar to the Guitar nebula where the “neck” and “body” regions (Cordes et al. [@cordes93]) are analogous to our “head” and “fan” regions respectively. However, there are some differences; the head is more circular than the cylindrical neck of the Guitar nebula and the fan is more rectilinear than the bulbous body of the Guitar nebula. We note, however, that both nebulae show stronger  emission in the transition between the narrow and broad regions. These could indicate regions of increased density on sub-parsec scales, given that the full extent of the nebula is $\sim 0.6$ pc if 2 kpc away. Alternatively, the stronger emission may indicate the transition between different physical regimes of the nebula. The Guitar nebula emission shows two distinct bubble regions (Cordes et al. [@cordes93]). The fainter, inner bubble is analogous to the “bridge” of emission that we see in PSR B0740$-$28. It is possible to generate these apparently closed structures via projection of a truly open structure. This can occur if the nebula is slightly inclined, as is shown in Fig 1($d$) for a cone assumed to be inclined at 20$^\circ$. The apparently closed nature of the PSR B0740–28 nebula suggests that in the rear region of the nebula, the ram pressure due to the pulsar motion is no longer important and we could instead be seeing the point where the pressure in the wind is balanced by the ISM pressure — it is possible that the break between the “head” and “tail” components of the PWN may correspond to the transition between these two sources of confinement. However, this explanation is not consistent with the entire PWN being embedded in gas of uniform density, since the pulsar’s motion must be supersonic to form a bow shock, but in which case there can be no pressure balance at the rear of the PWN. In order to explain the observed morphology, we propose that the pulsar has recently passed through a strong density gradient in the ISM, such that the pulsar’s motion is supersonic in the comparatively dense medium ($n_{\rm H} \sim 0.25$ cm$^{-3}$) through which it is currently propagating, but was subsonic in low density material ($n_{\rm H} \la 0.003$ cm$^{-3}$) surrounding the closed end of the PWN. Similar arguments can be invoked to explain the morphology of the Guitar PWN powered by PSR B2224+65. Considering the latter region, the requirement that the pulsar be moving slower that the local sound speed implies an upper limit to the pulsar’s distance of 3–4 kpc, as no ISM component has a sound speed faster than $\sim500$ km s$^{-1}$. If PSR B0740–28 is at 2 kpc, then the pulsar wind pressure, $P_{\rm w} = \dot{E}/(4\pi r_b^2c)$ approximately balances the ISM pressure, $P_{ISM} \sim 5\times10^{-14}$ Pa, at the the location of the closed region (where $r_{\rm b}$ is the distance to the back of the nebula), providing further support for the lower distance argued for above. In the interpretation we propose, the closed region of the PWN corresponds to the static or “ghost” nebulae proposed by Blandford et al. ([@blandford73]). The faintness of this region is consistent with the expectation that such sources should generally never be observable (Gaensler et al. [@gaensler00]); in the case of PSR B0740–28, it was only through the presence of brighter emission near the pulsar that this closed region was identified. Bucciantini & Bandiera ([@bucciantini01]) have recently shown that classical thin-shock models are not appropriate for pulsar bow-shock nebulae. However, here we find reasonable agreement between the classical models and the shape of the head of the PSR B0740$-$28 nebula. It is likely though that the microphysical processes they discuss are responsible for at least some of the morphological characteristics of these PWNe. Indeed Bucciantini & Bandiera ([@bucciantini01]) point out that perhaps PSR B0740$-$28 is an intermediate case between the nebulae associated with the MSPs and the Guitar nebula. Interestingly, they also predict that, based on the ISM conditions, this nebula should be readily detectable. However, we find that it is considerably less-so than the others. Clearly, increasing the presently small sample of optical bow-shock nebulae remains a key step towards being able to generalise about the ISM conditions within the vicinity of pulsars. We are grateful to the NTT Team and J. Bland-Hawthorn for the quality of the data obtained at the NTT and AAT respectively. We are also grateful to the referee, R. Bandiera, whose comments helped to improve the final draft. BWS is supported by NWO Spinoza grant 08-0 to E. P. J. van den Heuvel. BMG acknowledges the support of a CfA Clay Fellowship. DHJ would like to acknowledge R. A. Faulkner (1945–2001) for encouraging an early interest in astronomy. Aldcroft, T. L., Romani, R. W. & Cordes, J. M. 1992, ApJ, 400, 638 Arzoumanian, Z., Nice, D. J., Taylor, J. H. & Thorsett, S. E. 1994, ApJ, 422, 671 Bailes, M., Manchester, R. N., Kesteven, M. J., Norris, R. P. & Reynolds, J. E. 1990, MNRAS, 247, 322 Bell, J. F., Bailes, M., Manchester, R. N., Weisberg, J. M. & Lyne, A. G. 1995, ApJ 440, L81 Blandford, R. D., Ostriker, J. P., Pacini, F. & Rees, M. J. 1973, A&A, 23, 145 Bland-Hawthorn, J. & Jones, D. H. 1998, in Optical Astronomical Instrumentation, Proc. SPIE 3355, ed. S. D’Odorico, 855 Bucciantini, N. & Bandiera, R. 2001, A&A 375, 1032 Chatterjee, S. & Cordes, J., 2002, (astro-ph/0201062) Chen, Y., Bandiera, R. & Wang, Z. 1996, ApJ, 469, 715 Cordes, J. M., Romani, R. W. & Lundgren, S. C. 1993, Nature, 362, 133 Cordes, J. M. 1996, in IAU Colloq 160, p393 Dopita, M. A. & Hua, C. T. 1997, ApJS 108, 515 Fomalont E. B., Goss, W. M., Lyne, A. G., Manchester, R. N. & Justtanont, K. 1992, MNRAS 258, 497 Gaensler, B. M., Stappers, B. W., Frail, D. A., Moffett, D. A., Johnston, S. & Chatterjee, S. 2000, MNRAS 318, 58 Gaensler, B. M., Arons, J., Kaspi, V. M., et al. 2002, ApJ 509, in press (astro-ph/0110454) Hester, J. J., 1998, Mem. Soc. Astron. It. 69, 883 Koribalski, B., Johnston, S., Weisberg, J. M. & Wilson, W. 1995, ApJ, 441, 756 Kulkarni, S. R. & Hester, J. J. 1988, Nature 335, 801 McCulloch, P. M., Komesaroff, M. M., Ables, J. G., Hamilton, P. A. & Rankin, J. M. 1973, Astrophys.  Lett. 14, 169 Neckel, Th. & Klare, G. 1980, A&ASS 42, 251 Salter, C. J. 1970, IAUC 2295 Stone R. P. S. & Baldwin J. A 1983, MNRAS 204, 347 Taylor, J. H. & Cordes, J. M. 1993, ApJ, 411, 674 Wilkin, F. P. 1996, ApJL 459, 31 Wilkin, F. P. 2000, ApJ 532, 400 van Kerkwijk, M. H. & Kulkarni, S. R. 2001, A&A submitted, (astro-ph/0110065) [^1]: Based on observations made with ESO Telescopes at the La Silla Observatory (Programme 66.D–0046) and under the AAT Service Mode Programme

ArXiv

--- abstract: 'Theory of scattering of a quantum-mechanical particle on a cosmic string is developed. S-matrix and scattering amplitude are determined as functions of the flux and the tension of the string. We reveal that, in the case of the nonvanishing tension, the high-frequency limit of the differential scattering cross section does not coincide with the differential cross section for scattering of a classical pointlike particle on a string.' --- =0.3cm 0.9cm **On effects of non-Euclidean geometry in quantum theory** 0.3cm Yu.A. Sitenko $ ^{a\dagger}$, N.D. Vlasii $^{ab\star}$ 0.1cm $^{a}$ *Bogolyubov Institute for Theoretical Physics, National Academy of Sciences of Ukraine,* *14 b Metrologichna str., Kyiv 03143 , Ukraine* $^{b}$ *Department of Physics, Taras Shevchenko National University of Kyiv,* *6 Academician Glushkov ave., Kyiv 03680, Ukraine* 0.2cm 0.1cm **Keywords**: cosmic string, Bohm-Aharonov effect 0.1cm (based on talks given at: the International Workshop “Frontiers of Particle Astrophysics”, June 21-24, 2004, Kyiv, Ukraine; the George Gamow Memorial International Conference “Astrophysics and Cosmology after Gamow - Theory and Observations”, August 8-14, 2004, Odessa, Ukraine; the IV International Conference “Non-Euclidean Geometry in Modern Physics and Mathematics”, September 7-11, 2004, Nizhny Novgorod, Russia) $ \begin{array}{ll} ^{\dagger}\mbox{{\it e-mail address:}} & \mbox{[email protected]} \\ ^{\star}\mbox{{\it e-mail address:}} & \mbox{[email protected]} \end{array} $ Introduction {#s1} ============ Usually, the effects of non-Euclidean geometry are identified with the effects which are due to the curvature of space. It can be immediately shown that this is not the case and there are spaces which are flat but non-Euclidean. A simplest example is given by a twodimensional space (surface) which is obtained from a plane by cutting a segment of a certain angular size and then sewing together the edges. The resulting surface is the conical surface which is flat but has a singular point corresponding to the apex of the cone. To be more precise, the intrinsic (Gauss) curvature of the conical surface is proportional to the twodimensional delta-function placed at the apex; the coefficient of proportionality is the deficit angle. Usual cones correspond to positive values of the deficit angle, i.e. to the situation when a segment is deleted from the plane. But one can imagine a situation when a segment is added to the plane; then the deficit angle is negative, and the resulting flat surface can be denoted as a saddle-like cone. The deleted segment is bounded by the value of $2\pi$, whereas the added segment is unbounded. Thus, deficit angles for possible conical surfaces range from $-\infty$ to $2\pi$. It is evident that an apex of the conical surface with the positive deficit angle can play a role of the convex lens, whereas an apex of the conical surface with the negative deficit angle can play a role of the concave lens. Really, two parallel trajectories coming from infinity towards the apex from different sides of it, after bypassing it, converge (and intersect) in the case of the positive deficit angle, and diverge in the case of the negative deficit angle. This demonstrates the non-Euclidean nature of conical surfaces. It is interesting that this item provides a basis for understanding such physical objects as cosmic strings. In the present paper we shall discuss peculiarities of quantum theory and its quasiclassical limit, which are due to non-Euclidean geometry of locally flat space-times. Space-time in the presence of a cosmic string {#concl} ============================================= Cosmic strings are topological defects which are formed as a result of phase transitions with spontaneous breakdown of symmetries at early stages of evolution of the universe, see, e.g., reviews in Refs.[@Hi; @Vi]. In general, a cosmic string is characterized by two quantities: flux $$\Phi=\int\limits_{\rm core}d^2x\sqrt{g}B^3,$$ and tension $$\mu=\frac{1}{16\pi G}\int\limits_{\rm core}d^2x\sqrt{g}R;$$ here the integration is over the transverse section of the core of the string, $B^3$ is the field strength which is directed along the string axis, $R$ is the scalar curvature, $G$ is the gravitational constant, and units $\hbar=c=1$ are used. The space-time metric outside the string core is $$ds^2=dt^2-(1-4G\mu)^{-1}d{\stackrel{\sim}{r}}^2- (1-4G\mu)\stackrel{\sim}{r}^2d\varphi^2-dz^2= dt^2-dr^2-r^2d\stackrel{\sim}{\varphi}^2-dz^2,$$ where $$\stackrel{\sim}{r}=r\sqrt{1-4G\mu}, \quad 0\leq\varphi<2\pi,\quad 0\leq\stackrel{\sim}{\varphi}<2\pi (1-4G\mu).$$ A surface which is transverse to the axis of the string is isometric to the surface of a cone with a deficit angle equal to $8\pi G\mu$. Such space-times were known a long time ago (M. Fierz, unpublished, see footnote in Ref.[@We]) and were studied in detail by Marder [@Ma]. In the present context, as cosmological objects and under the name of cosmic strings, they were introduced in seminal works of Kibble [@Ki] and Vilenkin [@Vil]. A cosmic string resulting from a phase transition at the scale of the grand unification of all interactions is characterized by the values of tension $$\mu\sim(10^{-7}\div10^{-6})G^{-1}.$$ The nonvanishing of the string tension leads to various cosmological consequences and, among them, to a very distinctive gravitational lensing effect. A possible observation of such an effect has been reported recently [@Sa], and this has revived an interest towards cosmic strings. The flux parameter (1) is nonvanishing for the so-called gauge cosmic strings, i.e. strings corresponding to spontaneous breakdown of local symmetries. If tension vanishes ($\mu=0$), then a gauge cosmic string becomes a magnetic string, i.e. a tube of the magnetic flux lines in Euclidean space. If the tube is impenetrable for quantum-mechanical charged particles, then scattering of the latter on the magnetic string depends on flux $\Phi$ periodically with period $2\pi e^{-1}$ ($e$ is the coupling constant – charge of the particle). This is known as the Bohm-Aharonov effect [@Aha], which has no analogue in classical physics, since the classical motion of charged particles cannot be affected by the magnetic flux from the impenetrable for the particles region. The natural question is, how the nonvanishing string tension ($\mu\neq0$) influences scattering of quantum-mechanical particles on the string. Thus, the subject of cosmic strings, in addition to tantalizing phenomenological applications, acquires a certain conceptual importance. Quantum scattering on a cosmic string {#concl} ===================================== Due to nonvanishing flux $\Phi$ and tension $\mu$, the quantum scattering of a test particle by a cosmic string is a highly nontrivial problem. It is impossible to choose a plane wave as the incident wave, because of the long-range nature of the interaction inherent in this problem. A general approach to quantum scattering in the case of long-range interactions was elaborated by Hormander [@Ho]. This approach covers the cases of scattering on a Coulomb center and on a magnetic string ($\mu=0$), but is not applicable to the case of scattering on a cosmic string ($\mu\neq 0$). Therefore the last case needs a special consideration and a thorough substantiation. When the effects of the core structure of a cosmic string are neglected and the transverse size of the core is negligible, the field strength and the scalar curvature are presented by twodimensional delta-functions. Scattering of a quantum-mechanical particle on an idealized (without structure) cosmic string was considered in Refs.[@Hoo; @De; @So; @Si2]. A general theory of quantum-mechanical scattering on a cosmic string, permitting to take into account the effects of the core structure, was elaborated in Ref.[@Si5]. According to this theory, the $S$-matrix in the momentum representation is $$\begin{aligned} S(k,\varphi;\,\,k',\varphi')=\frac{1}{2}\frac{\delta(k-k')}{\sqrt{kk'}}\left\{\Delta (\varphi-\varphi'+\frac{4G\mu\pi}{1-4G\mu})\exp\left[-\frac{ie\Phi}{2(1-4G\mu)} \right]+\right. \nonumber \\ +\left.\Delta\left(\varphi-\varphi'-\frac{4G\mu\pi}{1-4G\mu}\right) \exp\left[\frac{ie\Phi}{2(1-4G\mu)}\right]\right\}+\delta(k-k') \sqrt{\frac{i}{2\pi k}}f(k,\,\,\varphi-\varphi'),\end{aligned}$$ where the initial $({\bf k})$ and final $({\bf k}')$ twodimensional momenta of the particle are written in polar variables, $f(k,\,\,\varphi-\varphi')$ is the scattering amplitude, and $\Delta(\varphi)= \frac{1}{2\pi}\sum\limits_{n=-\infty}^{\infty}e^{in\varphi}$ is the angular part of the twodimensional delta-function. Note that in the case of short-range interaction one has $2\Delta(\varphi-\varphi')$ instead of the figure brackets in Eq.(6). Thus, one can see that, due to the long-range nature of interaction, even the conventional relation between $S$-matrix and scattering amplitude is changed, involving now a distorted unity matrix (first term in Eq.(6)) instead of the usual one, $\delta(k-k')\Delta(\varphi-\varphi')(kk')^{-1/2}$. In view of the comparison with the Bohm-Aharonov effect [@Aha], we shall be interested in the situation when the string core is impenetrable for the particle. The scattering amplitude in this case takes form: $$f(k,\,\,\varphi)=f_0(k,\,\,\varphi)-\sqrt{\frac{2}{\pi ik}}\sum\limits_{n=-\infty}^{\infty}\exp[in\varphi-i(\alpha_n-|n|)\pi] \frac{J_{\alpha_n}(kr_c)}{H_{\alpha_n}^{(1)}(kr_c)},$$ where $r_c$ is the radius of the string core, $J_\nu(u)$ and $H_\nu^{(1)}(u)$ are the Bessel and the first-kind Hankel functions of order $\nu$, $$\alpha_n=\left|n-\frac{e\Phi}{2\pi}\right|(1-4G\mu)^{-1},$$ and $$\begin{aligned} f_0(k,\varphi)&=&\frac{1}{\sqrt{2\pi i k}} \left\{ \frac{\exp\left[i[[\frac{e\Phi}{2\pi}]](\varphi+ \frac{4G\mu\pi}{1-4G_\mu})-\frac{ie\Phi}{2(1-4G\mu)}\right]}{1-\exp \left[-i\left(\varphi+\frac{4G\mu\pi}{1-4G\mu}\right)\right]}-\right. \nonumber \\ &-&\left.\frac{\exp\left[i[[\frac{e\Phi}{2\pi}]]\left(\varphi- \frac{4G\mu\pi}{1-4G\mu}\right)+\frac{ie\Phi}{2(1-4G\mu)}\right]}{1 -\exp\left[-i\left(\varphi-\frac{4G\mu\pi}{1-4G\mu}\right)\right]} \right\}\end{aligned}$$ is the amplitude of scattering on an idealized (without structure) cosmic string, $[[u]]$ is the integer part of $u$. Sum over $n$ in Eq.(7) describes the core structure effects. In the low-frequency limit $(k\rightarrow 0)$ these effects die out, and the differential cross section (i. e. the square of the absolute value of the amplitude) takes form $$\begin{aligned} \frac{d\sigma}{d\varphi}&=&\frac{1}{4\pi k} \left\{ \frac{1}{2\sin^2\left[\frac{1}{2}\left(\varphi+\frac{4G\mu\pi}{1- 4G\mu}\right)\right]}+\frac{1}{2\sin^2\left[\frac{1}{2}\left(\varphi- \frac{4G\mu\pi}{1-4G\mu}\right)\right]}-\right. \nonumber \\ &-&\left. \frac{\cos\left[\frac{e\Phi}{1-4G\mu}-\left(2[[\frac{e\Phi}{2\pi}]]+1\right) \frac{4G\mu\pi}{1-4G\mu}\right]}{\sin\left[\frac{1}{2}\left(\varphi+ \frac{4G\mu\pi}{1-4G\mu}\right)\right]\sin\left[\frac{1}{2}\left(\varphi- \frac{4G\mu\pi}{1-4G\mu}\right)\right]} \right\}.\end{aligned}$$ Differential cross section in the limit of high frequency of\ scattered particle {#concl} ============================================================= In the high-frequency limit $(k\rightarrow \infty)$ the first term in Eq.(7) dies out, and the differential cross section takes form $$\begin{aligned} \frac{d\sigma}{d\varphi}&=&\frac{1}{2}r_c(1-4G\mu)^2\left|\sum\limits_{l}\sqrt{\cos[\frac{1}{2} (1-4G\mu)(\varphi-\pi+2l\pi)]}\times \right.\nonumber \\ &\times& \left.\exp\{ie\Phi l-2ikr_c\cos[\frac{1}{2}(1-4G\mu)(\varphi-\pi+2l\pi)]\}\right|^2,\end{aligned}$$ where the summation is over integer $l$ satisfying condition $$-\frac{\varphi}{2\pi}-\frac{2G\mu}{1-4G\mu}<l<-\frac{\varphi}{2\pi}+1+\frac{2G\mu}{1-4G\mu}.$$ Note that results (10) and (11) are periodic in the value of flux $\Phi$ with period equal to $2\pi e^{-1}$. This feature is common with the scattering on a purely magnetic string ($\mu=0$). The difference is that the Bohm-Aharonov differential cross section in the low frequency limit ($k\rightarrow 0$) diverges in the forward direction, $\varphi=0$, while Eq.(10) diverges in two symmetric directions, $\varphi=\pm4G\mu(1-4G\mu)^{-1}$. The difference becomes much more crucial in the high-frequency limit $(k\rightarrow \infty)$. In the $\mu=0$ case one gets $$\frac{d\sigma}{d\varphi}=\frac{1}{2}r_c\sin\frac{\varphi}{2},$$ which is the cross section for scattering of a classical pointlike particle by an impenetrable cylindrical shell of radius $r_c$; evidently, the dependence on fractional part of $e\Phi(2\pi)^{-1}$ disappears in this limit. In the $\mu\neq 0$ case the dependence survives, see Eq.(11). In particular, if $0<\mu<(8G)^{-1}$, which is most interesting from the phenomenological point of view, then the cross section at $k\rightarrow\infty$ takes the following form in the region of the cosmic string shadow, $-\frac{4G\mu\pi}{1-4G\mu}<\varphi<\frac{4G\mu\pi}{1-4G\mu}$: $$\begin{aligned} \frac{d\sigma}{d\varphi}&=&r_c(1-4G\mu)^2\left(\cos[\frac{1}{2}(1-4G\mu)\varphi]\sin(2G\mu\pi) +\right. \nonumber \\ &+&\left.\sqrt{\sin^2(2G\mu\pi)-\sin^2[\frac{1}{2}(1-4G\mu)\varphi]}\right. \left.\cos\left\{e\Phi+4kr_c\sin[\frac{1}{2}(1-4G\mu)\varphi]\cos(2G\mu\pi)\right\}\right).\end{aligned}$$ Integrating Eq.(14) over the region of the shadow and the appropriate expression (which is independent of $\Phi$) over the region out of the shadow, we obtain the total cross section in the $k\rightarrow\infty$ limit: $$\sigma_{\rm tot}=2r_c(1-4G\mu).$$ The high-frequency limit is usually identified with the quasiclassical limit. Although this identification is valid for the total cross section, it is found to be invalid for the differential cross section, see Eqs.(11) and (14) revealing the periodic dependence on the flux, which is a purely quantum effect. These results are generalized to the case of scattering of a particle with spin. 0.2cm Acknowledgements ================= This work was supported by the State Foundation for Basic Research of Ukraine (project 2.7/00152). [99]{} A.Vilenkin, E.P.S.Shellard, [*Cosmic Strings and Other Topological Defects*]{}, Cambridge Univ. Press, Cambridge (1994). M.B.Hindmarsh, T.W.B.Kibble, Rep. Progr. Phys. [**58**]{}, 477 (1995). J.Weber, J.A.Wheeler, Rev. Mod. Phys. [**29**]{}, 509 (1957). L.Marder, Proc. Roy. Soc. London [**A252**]{}, 45 (1959). T.W.B.Kibble, J. Phys. [**A9**]{}, 1387 (1976); Phys. Rep. [**67**]{}, 183 (1980). A.Vilenkin, Phys. Rev. [**D23**]{}, 852 (1981); [**D24**]{}, 2082 (1981). M.Sazhin et al., Mon. Not. Roy. Astron. Soc. [**343**]{}, 353 (2003); astro-ph/0406516. Y.Aharonov, D.Bohm, Phys. Rev. [**115**]{}, 485 (1959). L.Hormander, [*Analysis of Linear Partial Differential Operators IV*]{}, Springer–Verlag, Berlin (1985). G.’t Hooft, Commun. Math. Phys. [**117**]{}, 685 (1988). S.Deser, R.Jackiw, Commun. Math. Phys. [**118**]{}, 495 (1988). P. de Sousa Gerbert, R.Jackiw, Commun. Math. Phys. [**124**]{}, 229 (1989). Yu.A.Sitenko, Nucl. Phys. [**B372**]{}, 622 (1992). Yu.A.Sitenko, A.V.Mishchenko, JETP [**81**]{}, 831 (1995).

ArXiv

--- abstract: 'The cosmic web is one of the most striking features of the distribution of galaxies and dark matter on the largest scales in the Universe. It is composed of dense regions packed full of galaxies, long filamentary bridges, flattened sheets and vast low density voids. The study of the cosmic web has focused primarily on the identification of such features, and on understanding the environmental effects on galaxy formation and halo assembly. As such, a variety of different methods have been devised to classify the cosmic web – depending on the data at hand, be it numerical simulations, large sky surveys or other. In this paper we bring twelve of these methods together and apply them to the same data set in order to understand how they compare. In general these cosmic web classifiers have been designed with different cosmological goals in mind, and to study different questions. Therefore one would not [*a priori*]{} expect agreement between different techniques however, many of these methods do converge on the identification of specific features. In this paper we study the agreements and disparities of the different methods. For example, each method finds that knots inhabit higher density regions than filaments, etc. and that voids have the lowest densities. For a given web environment, we find substantial overlap in the density range assigned by each web classification scheme. We also compare classifications on a halo-by-halo basis; for example, we find that 9 of 12 methods classify around a third of group-mass haloes (i.e. $M_{\rm halo}\sim10^{13.5}h^{-1}M_{\odot}$) as being in filaments. Lastly, so that any future cosmic web classification scheme can be compared to the 12 methods used here, we have made all the data used in this paper public.' author: - bibliography: - 'Allrefs.bib' title: '[Tracing the Cosmic Web]{}[^1]' --- \[1\] \[1\][[~~\#1~~]{}]{} \[1\] \[1\] [Cosmology: theory – dark matter – large-scale structure of the Universe. Methods: data analysis]{} Introduction {#section:intro} ============ On Megaparsec scales the matter and galaxy distribution is not uniform, but defines an intricate multi-scale inter-connected network which is known as the *cosmic web* [@bondweb1996]. It represents the fundamental spatial organization of matter on scales of a few up to a hundred Megaparsec. Galaxies, intergalactic gas and dark matter arrange themselves in a salient wispy pattern of dense compact clusters, long elongated filaments, and sheetlike tenuous walls surrounding near-empty void regions. Ubiquitous throughout the entire observable Universe, such patterns exist at nearly all epochs, albeit at smaller scales. It defines a complex spatial pattern of intricately connected structures, displaying a rich geometry with multiple morphologies and shapes. This complexity is considerably enhanced by its intrinsic multiscale nature, including objects over a considerable range of spatial scales and densities. For, a recent up-to-date report on a wide range of relevant aspects of the cosmic web, we refer to the volume by [@iau308]. The presence of the weblike pattern can be easily seen in the spatial distribution of galaxies. Its existence was suggested by early attempts to map the nearby cosmos in galaxy redshift surveys [@gregthomp1978; @joeveer1978b; @lapparent1986; @geller1989; @shectman1996] Particularly iconic was the publication of the *slice of the Universe* by [@lapparent1986]. Since then, the impression of a weblike arrangement of galaxies has been confirmed many times by large galaxy redshift surveys such as 2dFGRS [@colless2003; @weyschaap2009], the Sloan Digital Sky Survey SDSS [@tegmark2004] and the 2MASS redshift survey [@huchra2012], as well as by recently produced maps of the galaxy distribution at larger cosmic depths such as VIPERS [@vipers2014]. From cosmological N-body simulations [e.g. @springmillen2005; @illustris2014; @eagle2015] and recent Bayesian reconstructions of the underlying dark matter distribution in the Local Universe [@hess2013; @kitaura2013; @2014MNRAS.445..988N; @leclercq2015; @sorce2016], we have come to realize that the weblike pattern is even more pronounced and intricate in the distribution of dark matter. The Components of the Cosmic Web -------------------------------- The most prominent and defining features of the cosmic web are the filaments. The most outstanding specimen in the local Universe is the Pisces-Perseus chain [@giovanelli1985]. A recent systematic inventory of filaments in the SDSS galaxy redshift distribution has been catalogued by [@tempel2014] [also see @jones2010; @sousbie2011b]. Filaments appear to be the highways of the Universe, the transport channels along which mass and galaxies get channelled into the higher density cluster regions [@haarlemwey1993; @2004ApJ...603....7K] and which define the connecting structures between higher density complexes [@bondweb1996; @colberg2005; @weybond2008; @aragon2010]. On the largest scales, filaments on scales of 10 up to 100 Mpc, are found to connect complexes of superclusters - such as the great attractor [@1988ApJ...326...19L], the Shapley concentration or more recently the Vela supercluster [@2017MNRAS.466L..29K] - as was, for example, indicated by the work of [@2004ApJ...606...25B], [@romanodiaz2007] and [@2015MNRAS.452.1052L]. By contrast, the tenuous sheetlike membranes are considerably more difficult to find in the spatial mass distribution traced by galaxies. Their low surface density renders them far less conspicuous than the surrounding filaments, while they are populated by galaxies with a considerably lower luminosity [see e.g. @cautun2014]. When looking at the spatial structure outlined by clusters, we do recognise more prominent flattened supercluster configurations, often identified as *Great Walls*, which is a reflection of their dynamical youth. Particularly outstanding specimens are the CfA Great Wall [@geller1989], the Sloan Great Wall [@2005ApJ...624..463G], and most recently the BOSS Great Wall and the well established supergalactic plane [@1953AJ.....58...30D; @2000MNRAS.312..166L]. Along with filaments, the large void regions represent the most prominent aspect of the Megaparsec scale Universe. These are enormous regions with sizes in the range of $20-50h^{-1}$ Mpc that are practically devoid of any galaxy, usually roundish in shape and occupying the major share of space in the Universe [see @weyiau308 for a recent review]. Forming an essential and prominent aspect of the [*cosmic web*]{} [@bondweb1996], voids are instrumental in the spatial organisation of the cosmic web [@icke1984; @sahni1994; @shethwey2004; @einasto2011b; @aragon2013]. The first indications for their existence was found in early galaxy redshift samples [@chincarini1975; @gregthomp1978; @zeldovich1982], while the discovery of the 50 Mpc size Boötes void by [@kirshner1981], [@kirshner1987] and the CfA study by [@lapparent1986] established them as key aspects of the large scale galaxy distribution. Recent studies have been mapping and cataloguing the void population in the Local Universe [@fairall1998; @pan2012; @sutter2012], and even that in the implied dark matter distribution [@leclercq2015b]. In the immediate vicinity of our Milky Way, one of the most interesting features is in fact the Local Void whose diameter is around 30 Mpc [@tully_atlas1987]. Its effectively repulsive dynamical influence has been demonstrated in studies of cosmic flows in the local volume [@tully2008], while a recent study even indicated the dominant impact of a major depression at a distance of more than 100 Mpc [the so-called “dipole repeller”, @hoffman2017]. Physics and Dynamics of the Cosmic Web -------------------------------------- The cosmic web is a direct result of two physical drivers, which are at the heart of the current paradigm of structure formation. The first is that the initial density field is a Gaussian random field, described by a power spectrum of density fluctuations [@adler1981; @bbks]. The second is that these perturbations evolve entirely due to gravity [@peebles1980]. Gravitational instability is responsible for increasing the contrast in the universe, as rich over-dense regions grow in mass and density while shrinking in physical size, and as empty voids expand and come to dominate the volume of the universe. Once the gravitational clustering process begins to go beyond the linear growth phase, we see the emergence of complex patterns and structures in the density field. Within the gravitationally driven emergence and evolution of cosmic structure the weblike patterns in the overall cosmic matter distribution do represent a universal but possibly transient phase. As borne out by a large array of N-body computer experiments of cosmic structure formation [e.g. @springmillen2005; @illustris2014; @eagle2015], web-like patterns defined by prominent anisotropic filamentary and planar features — and with characteristic large underdense void regions — are the natural outcome of the gravitational cosmic structure formation process. They are the manifestation of the anisotropic nature of gravitational collapse, and mark the transition from the primordial (Gaussian) random field to highly nonlinear structures that have fully collapsed into halos and galaxies. Within this context, the formation and evolution of anisotropic structures are the product of anisotropic deformations accurately described by the Zeld’d’dovich formalism in the mildly nonlinear stage, driven by gravitational tidal forces induced by the inhomogeneous mass distribution. In other words, it is the anisotropy of the force field and the resulting deformation of the matter distribution which are at the heart of the emergence of the weblike structure of the mildly nonlinear mass distribution [also see @bondweb1996; @hahn2007a; @weybond2008; @forero2009]. This idea was first pointed out by @zeldovich1970 [also see @icke1973] who described, in the now seminal *“Zel’dovich approximation”* framework, how gravitational collapse amplifies any initial anisotropies and gives rise to highly anisotropic structures. Accordingly, the final morphology of a structure depends on the eigenvalues of the deformation tensor. Sheets, filaments and clusters correspond to domains with one, two and three positive eigenvalues, while voids correspond to regions with all negative eigenvalues. Based on this realization, [@doroshkevich1970] derived a range of analytical predictions for structure emerging from an initial field of Gaussian perturbations. In the emerging picture of structure formation, also known as Zel’dovich’s pancake picture, anisotropic collapse has a well defined sequence, with regions first contracting along one axis to form sheets, then along the second axis to produce filaments and only at the end to fully collapse along each direction [@shandzeld1989; @shandsuny2009]. Following up on this, the early evolution of the cosmic web can be understood in detail in terms of the singularities and caustics that are arising in the matter distribution as a result of the structure of the corresponding flow field [see @shandzeld1989; @hidding2014]. Indeed, one of the most interesting recent developments in our understanding of the dynamical evolution of the cosmic web has been the uncovering of the intimate link between the emerging anisotropic structures and the multistream migration flows involved in the buildup of cosmic structure [@shandarin2011; @shandarin2012; @falck2012; @neyrinck2012; @abel2012]. Also recent observational advances have enabled new profound insights into the dynamical processes that are shaping the cosmic web in our Local Universe. In particular the Cosmicflows-2 and Cosmicflows-3 surveys of galaxy peculiar velocities in our Local Universe have produced tantalizing results [@courtois2013; @tully2014], opening up a window on the flows of mass along and towards structures in the local cosmic web. Amongst others, these studies show the sizeable impact of low-density void regions on the dynamics in the vicinity of the Milky Way and have allowed the velocity shear based V-web identification of weblike components in the local Universe [@2015MNRAS.452.1052L; @2015ApJ...812...17P; @hoffman2017]. The extension of the Zel’dovich approximation, the *adhesion approximation*, allows further insights into the hierarchical buildup of the cosmic web [@gurbatov1989; @kofman1990; @kofman1992; @hidding2012]. By introducing an artificial viscosity term, the adhesion approximation mitigates some of the late-time limitations of the Zel’dovich approximation. It also leads to a profound understanding of the link between the evolving phase-space structure of the cosmic matter distribution and the tendency to continuously morph the emerging spatial structure into one marked by ever larger structures [see also @1995PhR...262....1S for a review of analytical extensions to the Zel’dovich approximation]. Interestingly, for a considerable amount of time the emphasis on anisotropic collapse as agent for forming and shaping structure in the Zel’dovich pancake picture was seen as the rival view to the purely hierarchical clustering picture. In fact, the successful synthesis of both elements culminated in the *cosmic web* theory [@bondweb1996], which stresses the dominance of filamentary shaped features and appears to provide a successful description of large scale structure formation in the $\Lambda$CDM cosmology. This theoretical framework pointed out the dynamical relationship between the filamentary patterns and the compact dense clusters that stand out as the nodes within the cosmic matter distribution: filaments as cluster-cluster bridges [also see @bondweb1996; @weyedb1996; @colberg2005; @weybond2008]. In the overall cosmic mass distribution, clusters — and the density peaks in the primordial density field that are their precursors — stand out as the dominant features for determining and outlining the anisotropic force field that generates the cosmic web. The cosmic web theory embeds the anisotropic evolution of structures in the cosmic web within the context of the hierarchically evolving mass distribution [@bondmyers1996]. Meanwhile, complementary analytical descriptions of a hierarchically evolving cosmic web within the context of excursion set theory form the basis for a statistical evaluation of its properties [@shethwey2004; @shen2006]. Significance and Impact of the Cosmic Web ----------------------------------------- Understanding the nature of the cosmic web is important for a variety of reasons. Quantitative measures of the cosmic web may provide information about the dynamics of gravitational structure formation, the background cosmological model, the nature of dark matter and ultimately the formation and evolution of galaxies. Since the cosmic web defines the fundamental spatial organization of matter and galaxies on scales of one to tens of Megaparsecs, its structure probes a wide variety of scales, form the linear to the nonlinear regime. This suggests that quantification of the cosmic web at these scales should provide a significant amount of information regarding the structure formation process. As yet, we are only at the beginning of systematically exploring the various structural aspects of the cosmic web and its components towards gaining deeper insights into the emergence of spatial complexity in the Universe [see e.g. @cautun2014]. The cosmic web is also a rich source of information regarding the underlying cosmological model. The evolution, structure and dynamics of the cosmic web are to a large extent dependent on the nature of dark matter and dark energy. As the evolution of the cosmic web is directly dependent on the rules of gravity, each of the relevant cosmological variables will leave its imprint on the structure, geometry and topology of the cosmic web and the relative importance of the structural elements of the web, i.e. of filaments, walls, cluster nodes and voids. A telling illustration of this is the fact that void regions of the cosmic web offer one of the cleanest probes and measures of dark energy as well as tests of gravity and General Relativity. Their structure and shape, as well as mutual alignment, are direct reflections of dark energy [@parklee2007; @platen2008; @leepark2009; @lavaux2010; @lavaux2012; @bos2012; @2015MNRAS.446L...1S; @pisani2015]. Given that the measurement of cosmological parameters depends on the observer’s web environment [e.g. @2014MNRAS.438.1805W], one of our main objectives is to develop means of exploiting our measures of filament structure and dynamics, and the connectivity characteristics of the weblike network, towards extracting such cosmological information. Method Web types Input Type Main References ---------------------------------------- ------------------ ----------- ---------------------- -----------------------------------   Adapted Minimal Spanning Tree (MST) filaments haloes Graph & Percolation @2014MNRAS.438..177A   Bisous filaments haloes Stochastic @tempel2014 [@2016AC....16...17T] FINE filaments haloes Stochastic @gonzalez2010   Tidal Shear Tensor (T-web) all particles Hessian @forero2009 Velocity Shear Tensor (V-web) all particles Hessian @2012MNRAS.425.2049H CLASSIC all particles Hessian @2012MNRAS.425.2443K   NEXUS+ all particles Scale-Space, Hessian @cautun2013 Multiscale Morphology Filter-2 (MMF-2) all except knots particles Scale-Space, Hessian @aragon2007 @aragon2014   Spineweb all except knots particles Topology @aragon2010a DisPerSE all except knots particles Topology @sousbie2011   ORIGAMI all particles Phase-Space @falck2012 [@falck2015] MultiStream Web Analysis (MSWA) all particles Phase-Space @Ramachandara_Shandarin:15   \[table:methods\] Perhaps the most prominent interest in developing more objective and quantitative measures of large-scale cosmic web environments concerns the environmental influence on the formation and evolution of galaxies, and the dark matter halos in which they form [see e.g. @hahn2007b; @hahnphd2009; @cautun2014]. The canonical example of such an influence is that of the origin of the rotation of galaxies: the same tidal forces responsible for the torquing of collapsing protogalactic halos [@hoyle1951; @peebles1969; @doroshkevich1970] are also directing the anisotropic contraction of matter in the surroundings. We may therefore expect to find an alignment between galaxy orientations and large scale filamentary structure, which indeed currently is an active subject of investigation [e.g. @aragon2007a; @leepen2000; @jones2010; @codis2012; @tempel2012; @2013MNRAS.428.2489L; @tempel2013; @trowland2013; @trowlandphd2013; @aragon2014; @2016MNRAS.457..695P; @hirv2017; @2017MNRAS.464.4666G]. Some studies even claim this implies an instrumental role of filamentary and other weblike environments in determining the morphology of galaxies [see e.g. @pichon2016 for a short review]. Indeed, the direct impact of the structure and connectivity of filamentary web on the star formation activity of forming galaxies has been convincingly demonstrated by @2009Natur.457..451D [see also @2009ApJ...703..785D [@2015MNRAS.449.2087D; @2015MNRAS.454..637G; @aragon2016]]. Such studies point out the instrumental importance of the filaments as transport conduits of cold gas on to the forming galaxies, and hence the implications of the topology of the network in determining the evolution and final nature. Such claims are supported by a range of observational findings, of which the morphology-density relation [@dressler1980] is best known as relating intrinsic galaxy properties with the cosmic environment in which the galaxies are embedded [see e.g. @kuutma2017]. A final example of a possible influence of the cosmic web on the nature of galaxies concerns a more recent finding that has lead to a vigorous activity in seeking to understand it. The satellite galaxy systems around the Galaxy and M31 have been found to be flattened. It might be that their orientation points at a direct influence of the surrounding large scale structures [see @ibata2013; @cautun2015; @2015MNRAS.452.1052L; @2015ApJ...800...34G; @2015ApJ...799...45F; @2016ApJ...829...58G], for example a reflection of local filament or local sheet. Detection and Classification of Cosmic Web Structure ---------------------------------------------------- To enable further advances in the astronomical issues addressed above, we need to establish a more objective description and quantification of the structure seen in the cosmic web. However, extracting such topological and morphological information from a discrete set of points, provided by either an $N$-body simulation or a galaxy survey, is very difficult. As such, many different methods have been developed to tackle this problem (reviewed in depth in Section 4). Some of the problems faced by observational surveys include sampling errors, projection effects, observational errors, incomplete sky coverage, magnitude limits, as well as various biases (e.g. Malmquist bias, selection bias). On the other hand, $N$-body simulations return the full 6-dimensional phase space and density field of the simulated universe at any desired epoch. A method that takes full advantage of this often unobservable information cannot be directly applied to observations, but can be applied to simulations constrained to match observations [e.g. @2016arXiv160100093L]. For this reason, methods that are developed specifically for the analysis of numerical simulations, may be completely inapplicable to current observational data sets and vice versa. Yet the numerous articles in the literature which attempt to study the cosmic web often refer to the same structural hierarchy: knots, filaments, sheets and voids. Here, we use a numerical simulation to compare classifiers, that, regardless of their position on the theoretical to observational spectrum, speak the same language of knots, filaments, sheets, and voids. In the spirit of previous structure finder comparison projects [@Colberg2008; @Knebe2011 etc.], we present a comparison of cosmic web identification codes and philosophies. However, our comparison differs significantly from e.g. the seminal Santa Barbara comparison project [@1999ApJ...525..554F] or other tests of codes which purport to model the same physical process [e.g. @2012MNRAS.423.1726S; @2013MNRAS.435.1618K]. Instead, the methods compared here were developed for very different purposes, to be applied to different kinds of data and with different goals in mind. Some of the methods are based on treating galaxies (haloes) as points; while others were developed to be applied to density or velocity fields. Furthermore, unlike halo finders seeking collapsed or bound objects, there is no robust analytical theory [such as the spherical top hat collapse model of @1999MNRAS.308..119S] which we may use as a guide for how we expect different cosmic web finders to behave. Therefore, we enter into this comparison fully expecting large disagreements between the methods examined. Outline ------- This paper is laid out as follows: in Section 2 we group the different methods into “families” that follow broadly similar approaches. In Section 3, we present the test dataset that has been used as the basis for our comparison. In Section 4, we review each method that has taken part in the comparison. In Section 5, we describe the results of the comparison. In Section 6 we summarise our results and draw conclusions. Web Identification Methods:    Classification {#section:methods} ============================================= It is a major challenge to characterise the structure, geometry and connectivity of the cosmic web. The complex spatial pattern – marked by a rich geometry with multiple morphologies and shapes, an intricate connectivity, a lack of symmetries, an intrinsic multiscale nature and a wide range of densities – eludes a sufficiently relevant and descriptive analysis by conventional statistics to quantify the arrangement of mass and galaxies. Many attempts to analyze the clustering of mass and galaxies at Megaparsec scales have been rather limited in their ability to describe and quantify, let alone identify, the features and components of the cosmic web. Measures like the two-point correlation function, which has been the mainstay of many cosmological studies over the past nearly forty years [@peebles1980], are not sensitive to the spatial complexity of patterns in the mass and galaxy distribution. The present paper seeks to compare the diverse range of more sophisticated techniques that have been developed over the past few years to address the spatially complex Megaparsec scale patterns delineated by mass and galaxies in the Universe. In the present study we compare the results and web evaluations and identifications of 12 different formalisms. They are diverse, involving different definitions for the physical identity of the structural features, as well as employing different means of turning these definitions into practical identification tools. The various different methods that have been developed can largely be grouped into five main classes: - [**Graph and Percolation techniques.**]{} The connectedness of elongated supercluster structures in the cosmic matter distribution was first probed by means of percolation analysis, introduced and emphasized by Zel’dovich and coworkers [@zeldovich1982; @shandzeld1989; @shandarin2004]. A related graph-theoretical construct, the minimum spanning tree (MST) of the galaxy distribution, was extensively analyzed by Bhavsar and collaborators [@barrow1985; @graham1995; @colberg2007] in an attempt to develop an objective measure of filamentarity. @colberg2007 set out to identify filaments and their adjoining clusters, using an elaborate set of criteria for the identification of features based on the branching of MSTs. In our study, we involve the MST based algorithm developed by [@2014MNRAS.440L.106A] for identification of filaments and void regions in the GAMA survey [@2014MNRAS.438..177A]. - [**Stochastic methods.**]{} This class of methods involves the statistical evaluation of stochastic geometric concepts. Examples are filament detection algorithms based on the Bayesian sampling of well-defined and parameterized stochastic spatial (marked) point processes that model particular geometric configurations. and @2016AC....16...17T use the Bisous model as an object point process of connected and aligned cylinders to locate and catalogue filaments in galaxy surveys. One of the advantages of this approach is that it can be applied directly on the original galaxy point field, given by the positions of the galaxies centres, without requiring the computation of a continuous density field. These methods are computationally very demanding. A thorough mathematical nonparametric formalism involving the medial axis of a point cloud, as yet for 2-D point distributions, was proposed by [@genovese2010]. It is based on a geometric representation of filaments as the medial axis of the data distribution. Also solidly rooted within a geometric and mathematical context is the more generic geometric inference formalism developed by [@chazal2009]. It allows the recovery of geometric and topological features of the supposedly underlying density field from a sampled point cloud on the basis of distance functions. In addition, we also see the proliferation of tessellation-based algorithms. Following specific physical criteria, [@gonzalez2010] put forward a promising combination of a tessellation-based density estimator and a dynamical binding energy criterion [also see @weyschaap2009]. We may also include another recent development in this broad class of methods. [@leclercq2015; @leclercq2015c] describe a highly interesting framework for the classification of geometric segments using information theory. [@leclercq2016] have previously compared a few cosmic-web classifiers to each other, judging them on the basis of their information content. <!-- --> - [**Geometric, Hessian-based methods.**]{} A large class of approaches exploits the morphological and (local) geometric information included in the Hessian of the density, tidal or velocity shear fields [e.g. @aragon2007; @hahn2007a; @forero2009; @bond2010a; @2012MNRAS.421L.137L; @cautun2013]. Based on the realization that the formation and dynamical evolution of the cosmic web is tied to the tidal force field [see @bondweb1996], [@hahn2007a] developed an elaborate classification scheme based on the signature of the tidal tensor [also see @hahn2007b]. A further extension and elaboration of this tidal field based scheme was developed by [@forero2009], while also the multiscale Nexus formalism incorporates versions that classify weblike features on the tidal tensor signature [@cautun2013 see below] Following a similar rationale and focusing on the link between emerging weblike structures and the nature of the velocity flow in and around these features, in a sense following up on the classic realization of such a connection by [@zeldovich1970], Libeskind, Hoffman and collaborators forwarded the V-web technique [@2012MNRAS.425.2049H; @2012MNRAS.421L.137L; @2013MNRAS.428.2489L; @2013ApJ...766L..15L; @2014MNRAS.441.1974L; @2014MNRAS.443.1274L; @2015MNRAS.452.1052L; @2015MNRAS.453L.108L; @2015MNRAS.446.1458M; @2016MNRAS.460..297M; @2016MNRAS.458..900C; @2016MNRAS.457..695P]. Its classification is explicitly based on the signature of the velocity shear field. Instead of using the tidal or velocity sheer field configuration, one may also try to link directly to the morphology of the density field itself [@aragon2007; @bond2010a; @cautun2013]. Though this allows a more detailed view of the multiscale matter distribution, it is usually more sensitive to noise and less directly coupled to the underlying dynamics of structure formation than the tidal field morphology. A single scale dissection of the density field into its various morphological components has been defined by [@bond2010a], and applied to N-body simulations and galaxy redshift samples [also see @bond2010a; @bond2010b; @choi2010]. - [**Scale-space Multiscale Hessian-based methods.**]{} While most of the Hessian-based formalisms are defined on one particular (smoothing) scale for the field involved, explicit multiscale versions have also been developed. The MMF/Nexus Multiscale Morphology Filter formalism of [@aragon2007] and [@cautun2013] look at structure from a [*Scale Space*]{} point of view, where the (usually Gaussian) smoothing scale of the field, defines an extra dimension. This formalism takes into account the multiscale character of the cosmic mass distribution by assessing at each spatial location the prominence of structural signatures, set by the signature of the Hessian of the field involved [@aragon2007; @cautun2013]. A somewhat similar multiscale approach was followed by the Metric Space Technique described by [@wu2009], who applied it to a morphological analysis of SDSS-DR5. While the original MMF method [@aragon2007] only involved the density field, the Nexus formalism extended this to a versatile algorithm that classifies the cosmic web on the basis of a multiscale filter bank applied to either the density, tidal, velocity divergence or velocity shear fields. Applying the technique to the logarithm of the density increases its sensitivity and dynamical range and allows the approach to attain its optimal form, the so called [<span style="font-variant:small-caps;">NEXUS+</span>]{} method, revealing both major filamentary arteries as well as tiny branching tendrils [@cautun2013]. - [**Topological methods.**]{} While the Hessian-based methods concentrate on criteria of the local geometric structure of density, velocity or tidal field, another family of techniques seeks to assess the cosmic web by studying the connectivity and topological properties of the field involved. A typical example involves the delineation of underdense void basins in the large scale mass distribution by means of the [*Watershed Transform*]{}, in the form of the Watershed Void Finder [@platen2007] and ZOBOV [@neyrinck2008]. The Spineweb procedure [@aragon2010] extends this to an elaborate scheme for tracing the various weblike features – filaments, sheets and voids – on purely topological grounds. Spineweb achieves this by identifying the central axis of filaments and the core plane of walls with the boundaries between the watershed basins of the density field. While the basic Spineweb procedure involves one single scale, the full multiscale spineweb procedure allows a multiscale topological characterization of the cosmic web [@aragon2010b; @aragon2013]. In essence, the Spineweb procedure is a practical implementation of the mathematics of Morse theory [@morse1934]. Morse theory describes the spatial connectivity of the density field on the basis of its singularity structure, i.e. on the location and identity of the singularities - maxima, minima and saddle points - and their spatial connectivity by means of the characteristic lines defined by the gradient field. [@colombi2000] first described the role of Morse theory in a cosmological context, which subsequently formed the basis of the [*skeleton analysis*]{} by [@novikov2006] (2-D) and [@sousbie2008] (3-D). This defined an elegant and mathematically rigorous tool for filament identification. In a considerably more versatile elaboration of this, invoking the power of topological persistence to identify topologically significant features, [@sousbie2011] has formulated the sophisticated [<span style="font-variant:small-caps;">DisPerSE</span>]{} formalism that facilitates the detection of the full array of structural features in the cosmic mass distribution [also see @sousbie2011b]. Nonetheless, most of its applications are directed towards outlining the filaments. A further development along these lines, invoking the information provided by persistence measures, is that advocated by [@shivashankar2016]. - [**Phase-space methods.**]{} Most closely connected to the dynamics of the cosmic web formation process are several recently proposed formalisms that look at the phase-space structure of the evolving mass distribution [@abel2012; @falck2012; @shandarin2012]. They are based on the realization that – in cosmologies in which the intrinsic velocity dispersion of particles in the primordial universe is small – the evolving spatial mass distribution has the appearance of a 3D sheet folding itself in 6D phase space, a [*phase space sheet*]{}. By assessing its structure in full phase space, these formalisms trace the mass streams in the flow field reflecting the emergence of nonlinear structures. Noting that the emergence of nonlinear structures occurs at locations where different streams of the corresponding flow field cross each other, these phase-space methods provide a dynamically based morphological identification of the emerging structures. This class of methods contains the [<span style="font-variant:small-caps;">ORIGAMI</span>]{} formalism [@falck2012; @falck2015], the phase-space sheet methods of [@shandarin2011] [also see @Ramachandara_Shandarin:15] and [@abel2012], and the Claxon formalism [@hidding2017]. The Claxon approach incorporates the modelling of the nonlinear evolution of the cosmic mass distribution by means of the adhesion formalism [@gurbatov1989; @hidding2012], in order to identify and classify the singularities – shocks – emerging in the evolving structure. Claxon states that these singularities trace the skeleton of the cosmic web. Test data: Simulation and Data set ================================== Web Identification Methods:   \       Description & Details =============================== The following section describes each method as well as the practical details in the analysis of this data set. See Table \[table:methods\] for a brief summary. {width="\textwidth"} {width="\textwidth"} -0.5truecm {width="\textwidth"} ---------------------------------------------------------- ------- ----------- -------- ------- ------- ----------- -------- ------- ------- ----------- -------- ------- Method Knots Filaments Sheets Voids Knots Filaments Sheets Voids Knots Filaments Sheets Voids MST – – – – – – – – – 0.941 – 0.023 Bisous – 0.051 – – – 0.286 – – – 0.377 – – FINE – – – – – – – – – 0.411 – – V-web 0.003 0.034 0.204 0.755 0.097 0.235 0.331 0.337 0.231 0.317 0.293 0.159 T-web 0.013 0.149 0.413 0.425 0.166 0.380 0.319 0.135 0.328 0.415 0.211 0.045 CLASSIC 0.006 0.053 0.238 0.703 0.121 0.239 0.324 0.315 0.271 0.276 0.290 0.163 [<span style="font-variant:small-caps;">NEXUS+</span>]{} 0.001 0.113 0.228 0.657 0.084 0.488 0.250 0.178 0.245 0.658 0.088 0.006 MMF-2 – 0.078 0.190 0.732 – 0.295 0.197 0.508 – 0.909 0.072 0.019 Spineweb – 0.361 0.307 0.332 – 0.600 0.235 0.165 – 0.971 0.027 0.001 DisPerSE – 0.239 0.373 0.388 – 0.621 0.254 0.125 – 0.797 0.158 0.044 ORIGAMI 0.074 0.064 0.123 0.738 0.489 0.131 0.137 0.243 0.898 0.067 0.024 0.010 MSWA 0.001 0.007 0.088 0.903 0.070 0.106 0.264 0.560 0.641 0.219 0.130 0.009 ---------------------------------------------------------- ------- ----------- -------- ------- ------- ----------- -------- ------- ------- ----------- -------- ------- 0.5cm Comparison and Results {#sec:results} ====================== Here we present a visual and quantitative comparison of the different methods. We focus on comparing general features of the cosmic web: mass and volume filling fractions, density distributions and halo mass functions in each environment. As already mentioned, all methods were applied to the same simulation and they all used, depending on the method, either the dark matter particle distribution or the FOF halo catalogue. Visual comparison {#sec:cube} ----------------- We begin our analysis by performing a visual comparison of the various web finders. Fig. \[fig:env\_slice\_a\] and Fig. \[fig:env\_slice\_b\] show the environments returned by the web identification methods that took part in the comparison. Each panel shows the same $2\hmpc$ thick slice through the simulation box. Broadly speaking, there are two types of methods: the ones that return multiple cosmic web environments (i.e. voids, sheets, filaments and possibly knots; these are shown in the Fig. \[fig:env\_slice\_a\]) and the ones that identify only filaments (shown in Fig. \[fig:env\_slice\_b\]). Among the first type, [<span style="font-variant:small-caps;">DisPerSE</span>]{}, MMF and Spineweb do not identify knots. For the second type of methods, we show either the grid cells identified as filaments (the [Bisous]{} method) or the positions of the haloes associated to filaments for the methods that did not return a web classification for each volume element (the [<span style="font-variant:small-caps;">FINE</span>]{} and [<span style="font-variant:small-caps;">MST</span>]{} methods). A number of general points are immediately visible from inspection of Figs. \[fig:env\_slice\_a\] and  \[fig:env\_slice\_b\] (in no particular order): - [<span style="font-variant:small-caps;">DisPerSE</span>]{} provides no knots, and its filaments are relatively thick compared to the other methods. - [<span style="font-variant:small-caps;">MMF-2</span>]{} and [Spineweb]{} fill much of the simulation’s volume with sheets and filaments. - [<span style="font-variant:small-caps;">ORIGAMI</span>]{} ascribes much of the over-dense volume as knots – owing primarily to the fact that these regions contains haloes which have undergone shell crossing along three orthogonal axes. - The Hessian methods ([<span style="font-variant:small-caps;">NEXUS+</span>]{}, [T-web]{}, [V-web]{} and [<span style="font-variant:small-caps;">CLASSIC</span>]{}) have a mix of knots, filaments and sheets, with voids dominating the under dense volume. - The [Bisous]{} model and [<span style="font-variant:small-caps;">MST</span>]{} seem to more or less agree with each other, whereas the [<span style="font-variant:small-caps;">FINE</span>]{} method ascribes far fewer haloes to filaments. {width="\textwidth"} It is important to note that some of these methods (specifically [<span style="font-variant:small-caps;">NEXUS+</span>]{}, [<span style="font-variant:small-caps;">MMF-2</span>]{}, [T-web]{} and [V-web]{}) have been designed, to various degrees, to reproduce the visual impression of the cosmic web. Furthermore, given that voids are by definition under-dense regions, it is ideologically unlikely that a given method would be designed to identify clusters deep inside voids[^2]. Density PDF ----------- The relationship between the cosmic web and the density field can be quantified by studying the probability distribution function (PDF) of the density field for each volume element (grid cell) as a function of web environment. This is shown in Fig. \[fig:den\_PDF\], where the total density PDF for this simulation (computed on a regular grid with cell spacing of 2$\hmpc$) is shown in black and is the same in all panels; we quantify the density by normalising to the mean density of the universe, $\delta=\rho/\bar{\rho}$. Note that only those methods that assign web classification to volume elements are included here – the [<span style="font-variant:small-caps;">FINE</span>]{} and [<span style="font-variant:small-caps;">MST</span>]{} methods assign a cosmic web environment only to haloes and are therefore excluded. The median of each PDF is denoted by the corresponding arrow. ### Knots In Fig. \[fig:den\_PDF\](a) we show that knots are characterised by a wide variety of environmental densities. Although the [T-web]{}, [V-web]{} and [<span style="font-variant:small-caps;">CLASSIC</span>]{} roughly agree, they differ substantially from the fourth Hessian method, [<span style="font-variant:small-caps;">NEXUS+</span>]{}, which has a much narrower and higher distribution of densities. Indeed [<span style="font-variant:small-caps;">NEXUS+</span>]{} is in closer agreement with [<span style="font-variant:small-caps;">MSWA</span>]{}. [<span style="font-variant:small-caps;">ORIGAMI</span>]{} peaks at roughly the same density as [V-web]{}, although is a little narrower. Perhaps the strongest conclusion we can draw from Fig. \[fig:den\_PDF\](a) is that the local density by itself is a poor proxy for being considered a knot by any given method. Or, conversely, where knots are found, their density may differ by an order of magnitude or more. ### Filaments In Fig. \[fig:den\_PDF\](b) we show the PDF of densities for cells identified as filaments. Qualitatively, the picture is similar to that for knots, but pushed to slightly lower densities. There also appears to be a weak convergence of the median density among methods. Namely, although the widths of the PDF are similar, their medians are more strongly in agreement (with the exception of [<span style="font-variant:small-caps;">MSWA</span>]{}), and span less than an order of magnitude. [<span style="font-variant:small-caps;">MSWA</span>]{} stands out here in labelling higher density cells as filamentary; the [Bisous]{} model (the only filament only model that can participate in this test) closely resembles [<span style="font-variant:small-caps;">ORIGAMI</span>]{}, while the PDFs of three of the Hessian methods ([T-web]{}, [V-web]{} and [<span style="font-variant:small-caps;">CLASSIC</span>]{}) have similar shapes but are offset with respect to each other. The PDF of [Spineweb]{} peaks at the lowest density. ### Sheets The density PDF for cells labelled as sheets, shown in Fig. \[fig:den\_PDF\](c), displays more coherence than those of knots or filaments. Despite the PDFs still varying widely among the web finders, the median densities of the PDFs are roughly similar and take values around $\delta=0$. The median of the set of density PDFs moves to lower values, although, like the PDFs for knots and filaments, there is still a wide variety of permissible environments. Three pairs of methods produce nearly identical PDFs: [<span style="font-variant:small-caps;">NEXUS+</span>]{} and [<span style="font-variant:small-caps;">ORIGAMI</span>]{}, [<span style="font-variant:small-caps;">DisPerSE</span>]{}, [T-web]{}, and [<span style="font-variant:small-caps;">MMF-2</span>]{}, and [V-web]{} and [<span style="font-variant:small-caps;">CLASSIC</span>]{}. Again, the PDF of [Spineweb]{} peaks at the lowest density. ### Voids The best agreement between methods is found in regions denoted as voids, as shown in Fig. \[fig:den\_PDF\](d). The void density PDFs show less diversity and generally have the same shape. The spread in medians is small: less than 0.2 dex. As voids purport to be the most under-dense regions in the universe, they also make up the greatest fraction of the simulation’s volume (as can be inferred by the overlap between the void density PDF and the total density PDF). It can thus be said that the methods studied here all agree that the majority of the simulation volume is indeed categorised as void. ### Trends in the density PDFs The cosmic web classification is layered: knots are embedded in filaments, which, in turn, reside in sheets, which, in turn constitute the boundaries between different void basins. As our analysis of the cosmic web moves from knots to voids, the median of the density distribution of each method and for each web type moves to lower values [*in tandem.*]{} Although for a given web type there may be a wide variety of permissible density environments across the analysed methods, each method follows a similar trend. The peak of the density PDF moves to lower and lower densities, with most methods converging in the lowest density and most abundant environment in the simulation: voids. Mass and volume fraction ------------------------ We continue the cosmic web finder’s comparison with a study of the volume and mass filling fractions that are ascribed to a specific cosmic web type. These quantities are shown in Fig. \[fig:vff\_mff\] for knots, filaments, sheets and voids. The mass fraction is found by summing up the particles in all the cells with the same cosmic web type and dividing by the total number of particles in the simulation. The volume fraction is found by counting all the cells with the same cosmic web type and dividing by the total number of cells. Note that for these tests we have a 100$^3$ grid with (2$\hmpc)^3$ cells. {width="\textwidth"} {width="\textwidth"} {width="\textwidth"} - [**Knots:**]{} as [<span style="font-variant:small-caps;">ORIGAMI</span>]{} makes no distinction between knots and haloes, it is perhaps unsurprising that this method finds that nearly half the simulation’s mass is confined in $\sim7\%$ of the volume. Most other methods tag far fewer cells as knots, claiming they constitute below $\sim1\%$ of the volume with between 10%-20% of the mass. Interestingly the mass-volume fraction relation for knots follows a fairly tight linear proportionality – the more mass found in knots, the more volume, regardless of method used. - [**Filaments:**]{} A similar, but slightly weaker proportionality between mass and volume fraction is found for filament regions. Here, [Spineweb]{} and [<span style="font-variant:small-caps;">DisPerSE</span>]{} place roughly 60% of the simulations mass in filaments which occupy some $\sim35\%$ and $\sim25\%$ of the simulation volume, respectively. Unlike knots, there is considerably more spread in the relationship between mass and volume fractions amongst the methods, although a linear relationship is still discernible to the eye. Similar to knots, [<span style="font-variant:small-caps;">MSWA</span>]{} continues to place virtually none of the volume and roughly $\sim10\%$ of the simulation’s mass in filaments. The [Bisous]{} model – the only one of the filament-only models that can participate in this comparison – finds very similar filament volume and mass filling fractions as [<span style="font-variant:small-caps;">CLASSIC</span>]{}, [V-web]{} and [<span style="font-variant:small-caps;">MMF-2</span>]{}, with some $\sim5\%$ of the simulations volume and $\sim30\%$ of the simulations mass labelled as filaments. To summarize, the filament volume filling fraction spans from virtually nothing ([<span style="font-variant:small-caps;">MSWA</span>]{}) to more than a third of the volume ([Spineweb]{}); while the filament mass filling fraction spans roughly double that range, from $\sim10\%$ ([<span style="font-variant:small-caps;">MSWA</span>]{}) to $\sim60\%$ ([<span style="font-variant:small-caps;">DisPerSE</span>]{}). - [**Sheets:**]{} The spread of the sheet mass filling fraction is quite tight, with most methods assigning $\sim30\%\pm5\%$ of the total mass to sheets (with the exception of [<span style="font-variant:small-caps;">ORIGAMI</span>]{} and [<span style="font-variant:small-caps;">MMF-2</span>]{}, which find lower values). However, the sheet volume filling fractions vary substantially between methods, ranging from less than $\sim10\%$ for MWSA to more than 40% for T-web. As in knots and filaments, MWSA continues to assign only a small volume fraction to sheets. - [**Voids:**]{} The volume fraction associated to voids shows three distinct groups: three methods with $\sim 40\%$ ([<span style="font-variant:small-caps;">DisPerSE</span>]{}, [Spineweb]{} and [T-web]{}), five with $\sim 70\%$ ([<span style="font-variant:small-caps;">NEXUS+</span>]{}, [<span style="font-variant:small-caps;">ORIGAMI</span>]{}, [V-web]{}, [<span style="font-variant:small-caps;">MMF-2</span>]{} and [<span style="font-variant:small-caps;">CLASSIC</span>]{}), and one with $\sim90\%$ ([<span style="font-variant:small-caps;">MSWA</span>]{}). For the first group of finders ([<span style="font-variant:small-caps;">DisPerSE</span>]{}, [Spineweb]{} and [T-web]{}), the mass fraction is more or less the same at around $10-15\%$. For the second group ([<span style="font-variant:small-caps;">NEXUS+</span>]{}, [<span style="font-variant:small-caps;">ORIGAMI</span>]{}, [V-web]{}, [<span style="font-variant:small-caps;">MMF-2</span>]{} and [<span style="font-variant:small-caps;">CLASSIC</span>]{}), the mass fraction in voids spans a large range from $\sim 15\%$ for [<span style="font-variant:small-caps;">NEXUS+</span>]{} to $\gsim50\%$ for [<span style="font-variant:small-caps;">MMF-2</span>]{}. In general it is apparent that the mass fraction assigned to void regions spans a large range. It is interesting to note how the void mass filling fractions of these methods have flipped compared to their estimate for the filament mass fraction. In summary, the various methods predict fairly large ranges for the volume and mass fractions assigned to a given web type. Given the substantial differences in how these methods identify the web components, it is not very surprising that there are large discrepancies in these fractions. That said, in each plot of Fig. \[fig:vff\_mff\], clusters of methods can be identified which have similar values of either the volume fraction, mass fraction or both. The values for the mass and volume fractions are shown in the first two columns of Table \[tab:mff\]. Halo assignment and mass functions ---------------------------------- We now compare how the web environment assigned to haloes varies across cosmic web finders. For most methods, each halo is assigned the cosmic web environment of the cell in which its centre is located in. For the filament only methods ([Bisous]{}, [<span style="font-variant:small-caps;">FINE</span>]{}, and [<span style="font-variant:small-caps;">MST</span>]{}), the methods themselves directly identify which haloes are part of a filamentary structure. In Fig. \[fig:HMF\] we show the halo mass function for the entire halo sample and for each web type. We find a mixed picture, with substantial variations in the halo mass function of web types. Despite this, there are also agreements. For example, all the methods place the most massive haloes (i.e. $M \gsim 10^{14}M_{\odot}$) into knots. Similar trends are visible in how the filament halo mass function behaves – the mass functions are similarly valued at low masses and show a “knee” that precipitates a quick decline in the mass function. The agreement of mass functions in filaments is strongest (except the phase-space methods, [<span style="font-variant:small-caps;">ORIGAMI</span>]{} and [<span style="font-variant:small-caps;">MSWA</span>]{}), and the shape of the halo mass function in filaments is the closest one to the total halo mass function. Indeed, [<span style="font-variant:small-caps;">MMF-2</span>]{} and [Spineweb]{} place nearly all haloes in filaments: the green dashed and dot-dashed curves are only visibly separate from the black line below $\sim10^{12.5}h^{-1}M_{\odot}$. The last column of Table \[tab:mff\] shows how much mass is locked up in haloes of a given web type for each web finder. The four Hessian methods ([<span style="font-variant:small-caps;">NEXUS+</span>]{}, [T-web]{}, [V-web]{} and [<span style="font-variant:small-caps;">CLASSIC</span>]{}) agree that around 20-30% of mass in haloes is found in knot haloes. This is dramatically different from [<span style="font-variant:small-caps;">DisPerSE</span>]{}, [<span style="font-variant:small-caps;">MMF-2</span>]{} and [Spineweb]{} which do not identify haloes as belonging to knots, and the phase-space methods ([<span style="font-variant:small-caps;">ORIGAMI</span>]{} and [<span style="font-variant:small-caps;">MSWA</span>]{}), which place the bulk of halo mass in knots. It is interesting to note that the methods that do not identify knots, but do identify filaments, sheets and voids ([Spineweb]{}, [<span style="font-variant:small-caps;">MMF-2</span>]{}, and [<span style="font-variant:small-caps;">DisPerSE</span>]{}) place the overwhelming bulk of halo mass in filaments (with $\gsim 80\%$ of all halo mass in filaments). All methods also agree that haloes in voids have the least amount of total halo mass, although they disagree on exactly how much this is, with methods predicting either $\sim 15\%$ ([V-web]{}, [<span style="font-variant:small-caps;">CLASSIC</span>]{}), $\sim 5\%$ ([T-web]{}, [<span style="font-variant:small-caps;">DisPerSE</span>]{}) or $\lsim 1\%$ ([<span style="font-variant:small-caps;">NEXUS+</span>]{}, [<span style="font-variant:small-caps;">ORIGAMI</span>]{}, [<span style="font-variant:small-caps;">MSWA</span>]{}, [Spineweb]{}, [<span style="font-variant:small-caps;">MMF-2</span>]{}) of halo mass in voids. It is important to compare the environment tag associated to haloes on a halo per halo basis too, not only globally as is the case when comparing halo mass functions. To accomplish this, we ask the following question: for haloes in a given mass range, how many methods agree that some fraction of these have the same cosmic web environment? The answer to this question is shown in Fig. \[fig:halo\_agree\]. To better understand our analysis, lets consider the panel of Fig. \[fig:halo\_agree\](a), which gives the agreement across methods for individual knot haloes. For high halo masses, $M\gsim10^{14}h^{-1}M_{\odot}$, the panel shows that most such haloes ($\sim90\%$) are assigned to knots by all the six methods (namely: [<span style="font-variant:small-caps;">NEXUS+</span>]{}, [T-web]{}, [V-web]{}, [<span style="font-variant:small-caps;">CLASSIC</span>]{}, [<span style="font-variant:small-caps;">ORIGAMI</span>]{}, and [<span style="font-variant:small-caps;">MSWA</span>]{}) that identify knot environments (dot-dashed red curve). Conversely, 60% of the smallest haloes (with $M\sim10^{11}h^{-1}M_{\odot}$) are not assigned by any method to the knot environment (solid black curve). In between the two extreme masses, we find two bell-like curves where haloes with $M\sim10^{12}h^{-1}M_{\odot}$ are assigned to knots by only one method (black dotted curve), and haloes with $M\sim10^{13}h^{-1}M_{\odot}$ (black dashed curve) are assigned to knots by the two phase-space methods, [<span style="font-variant:small-caps;">MSWA</span>]{} and [<span style="font-variant:small-caps;">ORIGAMI</span>]{}. In Fig. \[fig:halo\_agree\](b), we show the agreement among filament haloes. Note the two peaks in the blue dashed and blue dot-dashed lines at $M_{\rm halo}\approx 10^{13.5}$: nine methods agree that $\sim30\%$ of haloes of this mass are in filaments while 10 methods agree that at least 10% of halos of this mass are in filaments. Here, four methods ([<span style="font-variant:small-caps;">DisPerSE</span>]{}, [Spineweb]{}, [<span style="font-variant:small-caps;">MMF-2</span>]{} and [<span style="font-variant:small-caps;">MST</span>]{}) place the most massive haloes in filaments. Fig. \[fig:halo\_agree\](c) and Fig. \[fig:halo\_agree\](d) indicate that no method puts the most massive haloes in sheets or voids. Specifically this means that no haloes with $M\gsim10^{14}h^{-1}M_{\odot}$ are found in sheets, and no haloes with $M\gsim10^{13.5}h^{-1}M_{\odot}$ are found in voids, by any method. The degree of agreement of web classifiers on a halo per halo basis varies accordingly to the spatial distribution of haloes, as we illustrate in Fig. \[fig:halo\_agree\_dot\]. Here, each halo is coloured by how many methods agree on its given classification. Because the number of methods capable of assigning haloes to a given web type changes (e.g filament only finders can’t identify knot haloes, etc) the colour scheme is not identical in each panel (see caption for exact colour explanation). In general if many of the capable methods agree on a specific halo’s classification the halo is shown in black; if around half of the capable methods agree, the halo is plotted in blue. If a small number of capable methods agree, the halo is plotted in red. If no method assigns a halo a given classification, the halo is plotted in grey. Fig. \[fig:halo\_agree\_dot\] “Knot” and “Filament” halo panels shows quite clearly that the haloes where the most methods agree belong to a biased set and are not simply random. Knot haloes find the most agreement in the densest areas of the simulation – a reassuring result. Similarly, those haloes which by eye appear to define the filamentary network too have the most agreements. Accordingly none of the haloes in either the densest parts of the simulation or in the filaments are assigned as void haloes (appearing as grey points). Sheets appear, as often is the case, as tenuous structures. Fig. \[fig:halo\_agree\_dot\] indicates that most or many methods are likely to agree on a specific halo’s classification based on its location. Summary and Conclusions ======================= Large galaxy redshift surveys (e.g. 2dFGRS, SDSS, 2MRS) reveal that at Megaparsec scales the Universe has a salient weblike structure. On these scales, the galaxies and the matter distribution in the universe has arranged itself into a complex weblike network of dense, interconnected knots, elongated filaments, two-dimensional sheets, and large nearly-empty voids. These cosmic environments characterise the universe on the largest scales. One important aspect of the cosmic web is its multi-scale character, manifesting itself in the existence of weblike structures over a sizeable range of scale. High-resolution simulations have revealed that such structures can be found down to very small scales, even down to the virial radius of haloes, and that they play a prominent role in the accretion of cold gas onto young and assembling protogalaxies in the early Universe [@2012MNRAS.422.1732D]. It ties in with the results of a range of recent studies that have analysed the role of environment on the formation and the evolution of galaxies . Furthermore, theoretical studies have suggested that around half of the warm gas in the Universe is hiding in a “warm-hot- intergalactic medium”, presumably in the filaments of the cosmic web [e.g. @2015Natur.528..105E]. It has therefore become of key importance to gain more insight into the structure and dynamics of the weblike universe, and into the interaction of the cosmic web with galaxy scale processes. The cosmic web is one of the most intriguing and striking patterns found in nature, rendering its analysis and characterization far from trivial. This is evidenced by the many elaborate descriptions that have been developed. The absence of an objective and quantitative procedure for identifying and isolating knots, filaments, sheets and voids in the cosmic matter distribution has been a major obstacle in investigating the structure and dynamics of the cosmic web. The overwhelming complexity of the individual structures and their connectivity, the huge range of densities and the intrinsic multi-scale nature prevent the use of simple tools. Over the past years, we have seen the introduction and proliferation of many new approaches and techniques. These methods are very varied in how they identify the cosmic web environments; being designed with different cosmological data in mind and to answer different questions. These issues are compounded since the techniques available to theorists and simulators differ substantially from those employed by observers. This makes it even more important to understand how the various web identification methods compare with each other. The main driver of this paper is to quantify in a systematic way both the similarities and differences between cosmic web finders. There is no well motivated common framework to objectively define the constituents of the cosmic web, so there is no way of judging which methods are successful or which ones are - in some objective way - “better”. As such, the goal is to compare the output of the various methods to better relate studies that make use of different web identification methods. We proceeded by comparing several basic properties of the cosmic web: the mass and volume filling fraction of each component, the density distribution and the halo mass function in each environment, and a halo by halo comparison of their environment tag. For this, we asked the authors of each method to apply their technique to the same data, the output of an N-body simulation, and to return the resulting web classification in a common format. We find a substantial diversity in the properties of the cosmic web across the various methods. This is to be expected given the challenges inherent in identifying the cosmic web and the multitude of approaches undertaken in doing so. In spite of this, we also find many similarities across the methods. Some of the most important agreements are: - Voids correspond to the most underdense regions and are consistently identified as such by all the methods. The voids occupy the largest volume fraction, with the majority of methods finding a ${\sim}70\%$ volume filling fraction. - Most methods, except [<span style="font-variant:small-caps;">ORIGAMI</span>]{} and [T-web]{}, find that knots contain ${\sim}10\%$ of the total mass in less than $1\%$ of the volume of the universe. - All the methods find that the density PDF systematically shifts towards lower densities as we go from knots to filaments, than to sheets and voids. Despite this trend, there is still a substantial overlap between the density PDF of different environments, which suggests that a simple density is inadequate for cosmic web identification. - Most massive haloes, $M\gsim10^{14}h^{-1}M_{\odot}$, are classified as residing in knot environments by all the methods that identify knots. - The voids are only sparsely populated with haloes and they lack completely massive haloes with $M\gsim10^{13.5}h^{-1}M_{\odot}$. We have a very incomplete knowledge of what is the effect of environment on galaxy formation and evolution or of what is the cosmological information encoded in the cosmic web pattern. The lack of knowledge is a result of the limitations of analytical approaches in modelling these non-linear processes. Each web finder captures different aspects of this very complex pattern, i.e. the cosmic web, so it is a worthwhile pursuit to analyze the connection between the environments identified by each method and the effect on galaxies and cosmological constraints. Acknowledgments {#acknowledgments .unnumbered} =============== The authors thank the Lorentz Center in Leiden for the organizational and financial support of this meeting, which ultimately led to this paper. In addition, we are grateful to NOVA and NWO for the financial support to facilitate the workshop. The authors would like to thank Adi Nusser, Christoph Pichon and Dmitri Pogosyan for useful discussions. NIL acknowledges and thanks Jenny Sorce for useful conversations. RvdW and SS, and (with a separate grant) MN and MAC, acknowledge support from the New Frontiers of Astronomy and Cosmology program at the Sir John Templeton Foundation. MC is supported by ERC Advanced Investigator grant COSMIWAY (grant number GA 267291) and the UK STFC grant ST/L00075X/1. MN was supported by the UK Science and Technology Facilities Council (ST/L00075X/1). BF acknowledges financial support from the Research Council of Norway (Programme for Space Research). ET acknowledges the support by the ETAg grants IUT26-2, IUT40-2, and by the European Regional Development Fund (TK133). FK and SEN thank Arman Khalatyan for assistance. AK is supported by the [*Ministerio de Economía y Competitividad*]{} and the [*Fondo Europeo de Desarrollo Regional*]{} (MINECO/FEDER, UE) in Spain through grant AYA2015-63810-P as well as the Consolider-Ingenio 2010 Programme of the [*Spanish Ministerio de Ciencia e Innovación*]{} (MICINN) under grant MultiDark CSD2009-00064. He also acknowledges support from the [*Australian Research Council*]{} (ARC) grant DP140100198. He further thanks William Fitzsimmons for fortune. SEN acknowledges support by the Deutsche Forschungsgemeinschaft under grant NU 332/2-1. GY acknowledge financial support from MINECO/FEDER under research grant AYA2015-63810-P. Affiliations {#affiliations .unnumbered} ============ [*$^1$Leibniz-Institute für Astrophysik Potsdam (AIP), An der Sternwarte 16, D-14482 Potsdam, Germany\ $^2$Kapteyn Astronomical Institute, University of Groningen, P.O. Box 800, 9700 AV Groningen, The Netherlands\ $^3$Institute for Computational Cosmology, Durham Univerity, South Road, DH1 3LE, United Kingdom\ $^4$Institute of Theoretical Astrophysics, University of Oslo, PO Box 1029 Blindern, N-0315, Oslo, Norway\ $^5$Tartu Observatory, Observatooriumi 1, 61602 Tõravere, Estonia\ $^{6}$Kavli Institute for Particle Astrophysics and Cosmology, Stanford University, Menlo Park, CA 94025, USA\ $^{7}$Department of Physics, Stanford University, Stanford, CA 94305, USA\ $^{8}$NASA Ames Research Center N232, Moffett Field, Mountain View, CA 94035, U.S.A.\ $^{9}$Instituto Astronomico de Ensenada,UNAM, Mexico\ $^{10}$Departamento de Física, Universidad de los Andes, Cra. 1 No. 18A-10, Edificio Ip, Bogotá, Colombia\ $^{11}$Instituto de Astrof[í]{}sica, Pontificia Universidad Catolica de Chile, Santiago, Chile\ $^{12}$Centro de Astro-Ingenier[í]{}a, Pontificia Universidad Catolica de Chile, Santiago, Chile\ $^{13}$Observatoire de la Cote d’Azur, Laboratoire Lagrange, Boulevard de l’Observatoire, CS, 34229, 06304 NICE, France\ $^{14}$Institute of Cosmology and Gravitation, University of Portsmouth, Portsmouth PO1 3FX, UK\ $^{15}$Janusz Gil Institute of Astronomy, University of Zielona Góra, ul. Szafrana 2, 65-516 Zielona Góra, Poland\ $^{16}$Racah Institute of Physics, Hebrew University of Jerusalem, Givat Ram, Jerusalem, 91904 Israel\ $^{17}$Instituto de Astrofìsica de Canarias (IAC), C/Vía Láctea, s/n, E-38200, La Laguna, Tenerife, Spain\ $^{18}$Departamento Astrofísica, Universidad de La Laguna (ULL), E-38206 La Laguna, Tenerife, Spain\ $^{19}$Departamento de Física Teórica, Módulo 15, Facultad de Ciencias, Universidad Autónoma de Madrid, 28049 Madrid, Spain\ $^{20}$Astro-UAM, UAM, Unidad Asociada CSIC\ $^{21}$Scuola Normale Superiore, Piazza dei Cavalieri 7, I-56126 Pisa, Italy\ $^{22}$Universidad de Buenos Aires, Facultad de Ciencias Exactas y Naturales, Buenos Aires, Argentina\ $^{23}$CONICET-Universidad de Buenos Aires, Instituto de Astronomía y Física del Espacio (IAFE), Buenos Aires, Argentina\ $^{24}$Department of Physics and Astronomy, University of Kansas, Lawrence, Kansas 66045, USA\ $^{25}$ICRAR, M468, University of Western Australia, Crawley, WA 6009, Australia\ $^{26}$Université de Lorraine, Institut Elie Cartan de Lorraine, 54506 Vandoeuvre-lés-Nancy Cedex, France\ $^{27}$Institut de Mécanique Céleste et Calcul des Ephémérides (IMCCE), Observatoire de Paris, 75014 Paris, France\ $^{28}$Institut d’Astrophysique de Paris, CNRS UMR 7095 and UPMC, 98bis, bd Arago, F-75014 Paris, France* ]{} [^1]: This paper is the outcome of the “Tracing the Cosmic Web” Lorentz Center workshop, held in Leiden, 17-21 February of 2014. [^2]: The measure of a density depends on scale: large enough volumes that include relatively small over-densities can, on average, be well below the mean density and thus considered voids.

ArXiv

--- abstract: 'We give an overview of the work done with the LNE-SYRTE fountain ensemble during the last five years. After a description of the clock ensemble, comprising three fountains FO1, FO2 and FOM, and its newest developments, we review recent studies of several systematic frequency shifts. This includes the distributed cavity phase shift which we evaluate for the FO1 and FOM fountains, applying the techniques of our recent work on FO2. We also report calculations of the microwave lensing frequency shift for the three fountains, review the status of the blackbody radiation shift, and summarize recent experimental work to control microwave leakage and spurious phase perturbations. We give current accuracy budgets. We also describe several applications in time and frequency metrology: fountain comparisons, calibrations of the international atomic time, secondary representation of the SI second based on the $^{87}$Rb hyperfine frequency, absolute measurements of optical frequencies, tests of the T2L2 satellite laser link, and review fundamental physics applications of the LNE-SYRTE fountain ensemble. Finally, we give a summary of the tests of the PHARAO cold atom space clock performed using the FOM transportable fountain.' author: - 'J. Guéna, M. Abgrall, D. Rovera, P. Laurent, B. Chupin, M. Lours, G. Santarelli, P. Rosenbusch, M.E. Tobar, R. Li, K. Gibble, A. Clairon, and S. Bize[^1] [^2] [^3][^4][^5][^6] [^7]' title: 'Progress in atomic fountains at LNE-SYRTE' --- =1 Introduction ============ Atomic fountain clocks provide the most accurate realization of the SI second and define the accuracy of the widely used international atomic time (TAI). Ten years ago, a few formal TAI calibration reports were available each year. Now, there are normally several formal reports each month as a result of a great deal of effort in several metrology institutes to improve the reliability and long term operation. At LNE-SYRTE, the fountain ensemble typically provides more than 20 TAI calibrations each year, about half of all reports. These large improvements in reliability and corresponding capability to perform lengthy measurements yield an enhanced ability to investigate smaller and smaller frequency shifts, to compare fountain clocks with more stringent uncertainties and to test their long term behavior. In turn, applications using atomic fountain clocks have also developed and benefited from the improvements of the last few years. In this article, we give an overview of developments and of applications of the LNE-SYRTE atomic fountain ensemble during the last five years. LNE-SYRTE fountain ensemble =========================== Overview -------- LNE-SYRTE has operated three atomic fountains for more than a decade. The first, FO1, is a $^{133}$Cs fountain which has been in operation since 1994 [@Clairon1995]. The second, FOM, was originally a prototype for the PHARAO[^8] cold atom space clock [@Laurent1998] and was later modified to be a transportable fountain clock. The third, FO2, is a dual fountain which operates with $^{87}$Rb and $^{133}$Cs simultaneously [@Guena2010]. Development for more than a decade has continuously improved each of these. Our description here reflects the present configuration. The three fountains share several features. All of them are gathering atoms in a Lin $\perp$ Lin optical molasses with laser beams oriented to launch atoms vertically in the so-called (1,1,1) configuration[^9] [@Gibble1993]. The 6 laser beams for the optical molasses are obtained from 6 fiber-coupled collimators, which are designed, machined and tuned to ensure proper alignment at the few $100~\mu$rad level when placed against the corresponding reference surfaces of the vacuum chamber. Similar fiber-coupled collimators provide the other laser beams for state selection and detection. In all three fountains, including the $^{87}$Rb part of FO2, the laser system is on a separate optical bench and the light is coupled to the vacuum system with optical fibers. The lasers are home-built extended cavity diode lasers, most of which use a narrow band interference filter for wavelength selection [@Baillard2006]. This design is used in the PHARAO cold atom space clock (see \[subsec\_PHARAO\]). Also, it has recently become commercially available. The required power is obtained either by injection locking another laser diode or using a tapered semi-conductor laser amplifier. In all three fountains, atoms are launched with the moving molasses technique [@Clairon1991] at a velocity of $\sim 4$ m.s$^{-1}$ and a temperature of $\sim 1~\mu$K. Atoms exit the launch sequence in the upper hyperfine state. For normal clock operation, microwave and laser pulses further select the atoms in the $m_F=0$ Zeeman sub-state of the lower hyperfine state ($|F=3,m_F=0\rangle$ for $^{133}$Cs, $|F=1,m_F=0\rangle$ for $^{87}$Rb). All three fountains also have at least one layer of magnetic shield that surrounds the entire vacuum system. Notably, the molasses and the detection region are shielded, which is important to operate in the presence of the strong magnetic field fluctuations as of many urban environments, including the Observatoire de Paris. The largest, vertical magnetic field component in the lower part of the fountain (molasses and detection) is measured by a flux-gate magnetometer and actively stabilized with a set of horizontal coils distributed over the height of the fountain. Two or three additional cylindrical shields with end caps surround only the interrogation region to further attenuate magnetic field fluctuations. Inside the innermost shield, a solenoid and a set of compensation coils provide a static magnetic field in the range of $80$ to $200$ nT. The field is homogeneous to $10^{-3}$ and stable to $2\times 10^{-5}$ from 1 s to several days, as established by spectroscopy of the $|F,m_F=1\rangle \longrightarrow |F+1,m_F=1\rangle$ field sensitive transition. This corresponds to a stability of $\sim 4~$pT. All three fountains also use the same strategy to control temperature and thermal gradients in the interrogation region. The innermost layer surrounding the interrogation is made of a material with high thermal conductivity (copper or aluminium alloy) and is well isolated from the environment to ensure good temperature uniformity of this layer. The temperature of this innermost layer is monitored using platinum resistors and is allowed to drift slowly, following temperature fluctuations in the room with a typical time constant of 1 day. The measured temperature is used to infer the blackbody radiation temperature seen by the atoms. With this scheme, the temperature of the innermost layer is uniform to much better than $0.1$ K and no large gradients exist between the inner region and the environment. This essentially removes the need to care about the impact of the necessary openings (for the atomic trajectories, pumping,...) on the effective blackbody temperature for the atoms. Using heating devices to change or actively stabilize the temperature of the innermost layer induced a significant degradation of the temperature uniformity, which was difficult to compensate, even with segmented heaters. Consequently, this approach was abandoned. The microwave interrogation cavities in all three fountains are $TE_{011}$ copper resonators with two independent microwave feedthroughs. The microwave synthesizers used to feed the cavities are home built. Despite significant differences, these synthesizers share the capability to use the ultra-low phase noise reference signal (see below \[subsec\_ultrastablereference\]) from a cryogenic sapphire oscillator (CSO) without significant degradation at the level of one to a few parts in $10^{15}$ at $1$ s [@Chambon2007]. Consequently, the short-term instability of all three fountains is typically limited by the atomic quantum projection noise [@Santarelli1999] and technical noise in the atom detection. The best short-term stability to date for the operation of an atomic fountain with a high atom number is $1.6\times 10^{-14}\tau^{-1/2}$ as reported in [@Vian2005] and exemplified in Fig. \[fig\_beststability\]. The present version of the synthesizers also incorporate a phase-stable microwave switch [@Santarelli2009] which helps to assess and remove leakage fields that may affect the interrogation process, independent of other systematic effects. Some details are given in \[subsec\_microwaveleakage\]. For all three fountains, the computer system allows automated control of most of the important parameters, enabling routine implementation of multiple interleaved measurements with different configurations to study systematic shifts. A large number of parameters (laser powers, temperatures in the interrogation region, temperatures in the environment, at the Cs or Rb sources, humidity,...) are frequently monitored. The microwave power is regularly optimized and the magnetic field is verified via spectroscopy of a field sensitive transition, with automated computer controlled sequences. FO1 {#subsec_FO1} --- FO1 uses a 2 dimensional magneto-optical trap (2DMOT) [@Dieckmann1998] as a source of slow atoms to load the optical molasses. An advantage of this type of source is that it confines the high alkali vapor pressure to the 2DMOT chamber, yielding a large number of cold atoms in the molasses while keeping source consumption low. It also facilitates a better vacuum in the molasses region, thereby minimizing perturbations in the detection region and the contamination of the interrogation, including the microwave cavity, by background alkali vapor. With higher laser powers in the 2DMOT, it should be possible to reproduce the atom numbers previously obtained with a chirped-slowed thermal atomic beam, that gave the best reported fountain stability [@Vian2005], while consuming an order of magnitude less of alkali metal. So far, the laser power devoted to the 2DMOT yields $\sim 4$ times fewer atoms in the molasses than with our best chirped-cooled atomic beam. Note that 2DMOT loading of the molasses has some disadvantages. Besides the quite intricated mechanical and optical design and the need for extra laser sources, we observe a degradation of the 2DMOT laser windows that are in contact with the relatively high alkali vapor pressure (up to $10^{-4}$ Pa) after a few years of almost continuous operation. Also, loading the molasses asymmetrically from the slow atom beam from the 2DMOT, leads to a non-spherical, complicated atomic cloud distribution which makes the investigation of some systematic frequency shifts more difficult. Note however, that by using the adiabatic passage method to change the atom number [@Pereira2002], the extrapolation of the cold collision shift to zero density is essentially unaffected by distortions of the atomic cloud’s shape. The microwave cavity in FO1 has a diameter of $42.99$ mm and a length of $42.88$ mm. The below-cutoff guides that let the atoms pass through the cavity have a diameter of $12$ mm and a length of $60$ mm. At the two ends of the cutoff guides, an additional diaphragm with a diameter of $11$ mm blocks atoms that could travel close to the copper walls. Each of the 2 opposing microwave feedthroughs includes a resonant, polarization-filtering rectangular waveguide section. In the latest implementation, the cavity has a loaded quality factor of $14000$. The interrogation region in FO1 is currently equipped with Stark plates. These can be used to measure the sensitivity of the clock transition to static electric fields to determine the Stark coefficient for the blackbody radiation shift correction. Alternatively, FO1 can be equipped with a blackbody radiator to directly measure the blackbody radiation shift (see \[subsec\_BBR\] below). FO2 dual Rb/Cs fountain ----------------------- A schematic view of the FO2 dual Rb/Cs fountain is shown in Fig. \[fig\_FontaineRbCs\]. It uses 2 independent 2DMOT sources for Cs and for Rb. For Rb, only the $27.8\%$ abundant isotope is used in the fountain and the use of a 2DMOT eliminates the $^{87}$Rb background atoms, as for Cs, and also the larger background from the unwanted $^{85}$Rb isotope. The Rb and Cs molasses are superimposed with dichroic collimators for each molasses beam, where the 852 nm light (Cs) and the 780 nm light (Rb) come from 2 independent optical benches through polarizing fibers and are then combined with a dielectric dichroic beamsplitter and collimated with an achromatic lens [@Chapelet2008]. Similar dichroic collimators are used for all other laser beams, the pushing beam for the state selection, detection beams and repumper for detection. As shown in Fig. \[fig\_FontaineRbCs\], there are a total of 4 microwave cavities for state selection and interrogation of Rb and Cs atoms. Because both interrogation cavities are in the interrogation region, their temperature must be uniform and therefore the same as the environment temperature, which removes any possibility to independently tune each cavity. Practically, the Rb and Cs cavities are a single copper assembly where the Rb and Cs cavity resonances were mechanically tuned relative to one another, after a lengthy and tedious process. The resonance frequencies had to be verified under vacuum and tightly controlled thermal conditions, after each tuning step that was a controlled polishing of the two cavity spacers [@Sortais2001a][@Bize2001a]. The Rb and Cs microwave cavities have the same aspect ratio, with the radii scaled by the ratio of the hyperfine frequencies. The cavity for Cs is located at the top. It has a diameter of $50.00$ mm and a length of $26.87$ mm. The Rb cavity has a diameter of $67.25$ mm and a length of $36.01$ mm. The below-cutoff end cap tubes have a diameter of $12$ mm and a length of $\sim 50$ mm. The 2 opposing microwave feedthroughs for each cavity include a resonant, polarization filtering rectangular waveguide section. The loaded quality factor is $7000$ for Cs and $6000$ for Rb. The Rb and Cs clocks in FO2 routinely operate simultaneously since 2009 [@Guena2010]. Two independent but connected and synchronized computer systems allow the coordination of multiple interleaved measurements with Cs and Rb, as well as all other tasks for each clock, such as monitoring the temperature, the magnetic field, etc. Thus far, the nominal configuration is to launch the Rb and Cs clouds almost simultaneously, but at different velocities, so that the interrogation times are simultaneous, i.e. the Rb cloud is at the center of the Rb cavity when the Cs cloud is at the center of the Cs cavity, on both the way up and the way down. In this way, the two atomic clouds do not interact with each other during the interrogation process, avoiding interspecies collisions. The two clouds also remain well separated as they fall to the detection region, allowing the time-resolved detection of Rb and Cs [@Guena2010]. As a result, the Rb and Cs clocks run simultaneously without impacting the performance of the other. The flexible computer systems allow for a large number of other possibilities which will be explored in the future. For instance, one cloud can be purposely launched through the other with a well controlled and yet variable speed, in order to study interspecies collisions at varying energies [@Legere1998], based either on measurements of the cold collision frequency shift or on direct detection of the scattered wave [@Hart2007]. FOM transportable fountain {#subsec_FOM} -------------------------- Photographs of the transportable fountain FOM are shown in Fig. \[fig\_FOM\]. FOM has two major subsystems. The first is the vacuum system where atoms are manipulated. The second is the optical laser bench, its electronics, the microwave synthesizer and the computer system that operates the fountain, all of which are mounted on a single frame. Both of these subsystems include vibration damping devices necessary from transportation. When disconnected, the two subsystems fit into a small size truck. A battery powers the ion pump to maintain the vacuum of the system. After transportation, the two subsystems are reconnected (polarizing optical fibers for the laser light, microwave cables, wires for monitoring) and the system is allowed to thermally stabilize for a few hours. After a limited number of manual re-optimizations (laser power through the optical fibers), a computer controlled sequence is started which automatically finds the laser set points and activates the laser stabilization loops [@Allard2004]. At remote locations, close-to-nominal operation of the fountain is typically recovered in less than 2 days. One notable feature of the transportable fountain FOM is that the microwave synthesizer has two possible modes of operation. At LNE-SYRTE and at laboratories where a high quality metrological reference signal is available, the microwave synthesizer is synchronized to this external reference and the FOM fountain works as a frequency standard calibrating the frequency of the external reference. In the second configuration, FOM autonomously delivers a primary reference signal. Here, a BVA quartz oscillator synchronizes the microwave synthesizer and a digital loop locks the frequency of this BVA quartz oscillator to the atoms. The microwave interrogation cavity in FOM has a diameter of $41.06$ mm and a length of $65.66$ mm. The two opposing microwave feedthroughs have direct evanescent coupling: the coaxial input cable is terminated in a small non-resonant cavity, with an antenna located a few millimeters away from the coupling iris of the cavity. The cavity has a loaded quality factor of $21000$. Ultra-stable reference and frequency distribution scheme {#subsec_ultrastablereference} -------------------------------------------------------- Figure \[fig\_ClockEnsemble\] is a schematic of the LNE-SYRTE fountain ensemble. A Cryogenic Sapphire Oscillator (CSO) [@Luiten1995b][@Mann2001] is the local oscillator for all of the atomic fountains. Its ultra-low phase noise allows the fountains to operate at the quantum projection noise limit [@Santarelli1999], producing short-term instabilities as low as $1.6\times 10^{-14}$ at 1 s ([@Vian2005] and Fig. \[fig\_beststability\]). As shown in Fig. \[fig\_ClockEnsemble\], a finely tunable[^10] microwave $11.98$ GHz signal is first derived from the free-running CSO. This frequency is divided down to 100 MHz and compared to the output of one of the LNE-SYRTE hydrogen masers, H1 in Fig. \[fig\_ClockEnsemble\]. A computer driven digital phase-lock loop (time constant $\sim 1000$ s) corrects the frequency of the $11.98$ GHz microwave signal to phase lock the 100 MHz signals. This $11.98$ GHz reference has the exquisite phase noise properties of the CSO for time scales less than $\sim 1000$ s and the mid and long-term frequency stability of the hydrogen maser, while providing the connection to LNE-SYRTE timescales, to the International Atomic Time (TAI) and to other laboratories via GPS and TWSTFT satellite time transfer. This $11.98$ GHz signal is the high-performance reference from which all other signals are derived. A 1 GHz source is the starting point of the microwave synthesis for FO1, FO2-Rb and FOM. Since FO2 is located next door to the CSO, a more direct synthesis from 11.98 GHz to 9.192 GHz could be used for FO2-Cs [@Chambon2005]. To compare to optical clocks through optical frequency combs, a $8.985$ GHz is also generated. Finally, an additional synthesizer generates signals in the 6 to 9 GHz range for testing (see \[subsec\_microwaveleakage\]). The design and characterization of the low phase noise electronics that generates these reference signals are reported in [@Chambon2005][@Chambon2007]. Phase noise power spectral densities can be found in [@Mann2001] for the CSO alone, and in [@Chambon2005], [@Chambon2007] and [@Dawkins2010] for the derived reference signals. Figure \[fig\_ClockEnsemble\] shows 100 to 300 m long actively compensated optical fibers links that are used to distribute the reference signals. These links are simplified versions of similar long-distance links [@Daussy2005], due to the short distance and limited insertion losses. The reference signal modulates the current of a $1.55~\mu$m diode laser that is launched into a standard telecom fiber. At the remote end, a fast photodiode detects the intensity modulation and delivers the signal to the application. A fraction of this recovered signal is used to modulate a second diode laser that is launched back into the fiber. At the emission site, the returned signal is detected by another fast photodiode and then mixed with the reference signal to detect the phase variations imposed by the two passes propagation through the fiber. A temperature controlled fiber spool of 50 to 100 m is inserted to control the length of the link to cancel these phase variations with a computer operated servo-loop (time constant $\sim 10$ s). For the high performance links at $6-9$ GHz and $8.985$ GHz, an additional fast actuated piezo-driven fiber stretcher is used (bandwidth $\sim 400$ Hz). Characterization of these links has shown stabilities $\sim 10^{-14}$ at 1 s for the 1 GHz carrier and $\sim 2\times 10^{-15}$ at 1 s for the $6-9$ GHz and $8.985$ GHz carriers, and a residual frequency offset well below $10^{-16}$. A computer acquires and analyzes the doubled-passed phase signals to assess the performance in real-time and detect anomalies. New solution for the ultra-low noise reference oscillator --------------------------------------------------------- The CSO allows a large improvement in the fountain short-term stability but requires cumbersome and costly liquid helium refills every 26 days. Each refill lasts several hours and degrades the performance of the CSO for nearly half a day. An alternate solution for fully continuous operation of an ultra-low noise microwave reference would be to use a pulsed-tube cryocooler for the CSO [@Hartnett2010a][@Hartnett2010][@Grop2010][@Grop2009]. However, advances in laser frequency stabilization and optical frequency combs offer the possibility to replace the CSO with a fully continuous, cryogen-free and maintenance-free ultra-low noise microwave generation system. Figure \[fig\_FemtoComb\] shows a schematic of such a laser-based system. An ultra-stable laser is realized by stabilizing a reliable $1.55~\mu$m laser, such as an Erbium-doped fiber laser or an extended cavity laser, to an ultra-stable reference Fabry-Perot cavity. Several Fabry-Perot stabilized lasers have demonstrated short-term stabilities well below $10^{-15}$ [@Young1999][@Millo2009b][@Jiang2011], similar to stabilities reported for cryogenic sapphire oscillators [@Mann2001]. Note however that simplified versions of these cavities can be sufficient for application to atomic fountains. Microwave signals with short-term frequency stabilities in the mid $10^{-15}$ range at 1 s are indeed sufficient to achieve the quantum projection noise limit at $10^{-14}$ at 1 s. This instability is not easy to surpass for other reasons: it is difficulty to increase the number of useful atoms per cycle, to cope with the tremendous cold collision shift that would result from more atoms, and to suppress certain detection noise sources. A short ($\leq 10$ cm) reference Fabry-Perot cavity with mirrors made of Ultra Low Expansion glass with a comparatively simple temperature stabilization system can deliver instabilities less than $2\times 10^{-15}$ between 1 and 1000 s, which is well suited to atomic fountains. As shown in Fig. \[fig\_FemtoComb\], the ultra-stable laser is used to stabilize the repetition rate of the optical frequency comb generated by a fiber based femtosecond laser. A well-chosen harmonic, here 12 GHz, is detected and divided down to 100 MHz to allow a phase comparison with a hydrogen maser, similar to what is done for the CSO (see \[subsec\_ultrastablereference\]). Here, a digital phase lock loop shifts the frequency of the acousto-optic modulator (AOM) and thereby the ultra-stable light that controls the comb repetition rate, ensuring the long-term stability of the 12 GHz reference that ultimately provides the $9.192$ GHz interrogation signal for the fountains. Ultra-low noise microwave generation with fiber-based optical frequency combs was demonstrated in [@Lipphardt2008] and [@Millo2009c]. Atomic fountains using microwave signals derived from optical references reported fountain short-term stabilities of $3.5\times 10^{-14}\tau^{-1/2}$ [@Millo2009a] and $7.4\times 10^{-14}\tau^{-1/2}$ [@Weyers2009a]. For both, the fountain stability was quantum projection noise limited and the microwave signal alone was compatible with a fountain stability $\leq 10^{-14}\tau^{-1/2}$. Subsequent progress on fiber-based optical frequency combs will enable the implementation of a fully operational version of the system of Fig. \[fig\_FemtoComb\]. Note that an alternative approach for achieving quantum noise limited stability without an ultra-low noise microwave sources is the continuous atomic fountain [@Guena2007][@DiDomenico2011a]. Recent studies of systematic shifts =================================== Distributed cavity phase shift {#subsec_DCP} ------------------------------ The distributed cavity phase (DCP) shift is a residual first order Doppler shift that occurs when the moving atoms interact with the field inside the microwave cavity that has a spatial phase variation. Up until recently, measurements have been unable to reveal this effect and to reasonably agree with calculations. On the one hand, this was because the corresponding frequency shifts can be quite small, the measured frequency shifts give only indirect information about the phase variations, this shift and its changes with fountain parameters can be intricated with other systematic effects. On the other hand, accurately calculating the actual phase distribution and its corresponding frequency shift was a challenging numerical problem. In [@Li2004] and [@Li2010b], a new theoretical approach was proposed to solve the problem of the cavity field computation. The field, represented explicitly as the sum of a large standing wave $\mathbf{H}_0(\mathbf{r})$ and of a small field, $\mathbf{g}(\mathbf{r})=\sum\mathbf{g}_m(\rho,z)\cos(m\phi)$, that describes the effects of wall losses and cavity feeds, is expressed as an azimuthal Fourier series. Only the lowest azimuthal terms are relevant $m=0$, $m=1$ and $m=2$. We have recently published measurements of the DCP shift performed with $^{133}$Cs in FO2 [@Guena2011] where, for the first time, quantitative agreement between theory and experiment is found. This study led to a better characterization and a quantitative evaluation of the contributions to the uncertainty of the DCP shift. It highlighted the importance of controlling the fountain geometry, notably the effective tilt of the fountain and the detection inhomogeneity, to reduce the DCP shift uncertainty. Based on the experimentally validated model, new cavity designs were proposed [@Li2010b] to reduce the DCP shift and its uncertainty. These measurements lowered the DCP shift uncertainty for FO2-Cs clock from $3\times 10^{-16}$ to $8.4\times 10^{-17}$ and defined a set of measurements to establish the DCP uncertainty of a fountain, an approach adopted in [@Li2011][@Weyers2012]. Here, we apply this approach to evaluate the DCP uncertainty of FO1, FOM and FO2-Rb. The $m=1$ term corresponds to power flowing horizontally from one side of the cavity to the other. This contribution can be probed by measuring the fountain frequency as a function of the tilt of the launch direction. The uncertainty for this term is determined by the experimental uncertainty in frequency change as a function of tilt, after correctly balancing the feeds to null the tilt sensitivity, and by the uncertainty of the fountain tilt for nominal operation [@Guena2011]. When we purposely exaggerate the $m=1$ DCP shifts by asymmetrically feeding the cavities of FO1 and FOM, the measurements show tilt sensitivities that are much smaller than those of FO2-Cs. This is largely due to the higher quality factor of their cavities and to their smaller cavity apertures. Also, when the two feeds are adjusted to produce equal Rabi pulse areas, FO1 and FOM show no measured tilt sensitivities at the level of $(1.5\pm 1.7)\times 10^{-16}$ mrad$^{-1}$ and $(1.8 \pm 1.1)\times 10^{-16}$ mrad$^{-1}$ respectively, along the axis of the feeds. For FO2-Rb, this sensitivity is $(1.8\pm 1)\times 10^{-16}$ mrad$^{-1}$. Tilt sensitivities in the perpendicular direction remain to be measured for FO1 and FOM. Based on the findings of [@Guena2011], we can reasonably apply, for the time being, the same bound as in the parallel direction. For FO2-Rb, the measured perpendicular sensitivity is $\leq 0.9\times 10^{-16}$ mrad$^{-1}$. Using the same methods as in [@Guena2011] (detected atoms *versus* tilt, frequency difference between the 2 asymmetric feedings), we estimate the uncertainty on the effective launch direction to be $1~$mrad for FO1, $0.2~$mrad for FO2-Rb and $0.7$ mrad for FOM in the direction of the feeds and $0.7~$mrad, $0.7~$mrad and $0.7~$mrad in the direction perpendicular to the feeds. The $m=0$ and $m=2$ terms are calculated from the phase distributions [@Li2004][@Li2010b] and the corresponding response of the fountain. This second step takes as an input a number of parameters of the fountain geometry such as the cloud size, the cloud temperature and the geometry of the detection. Figure \[fig\_contrastFO1FOM\] shows the measured contrast of the Rabi oscillations at upward and downward passages in the microwave cavity, which is used to ascertain the fountain geometry to compute the DCP shift [@Guena2011]. The same curve, measured for FO2-Cs, is shown in [@Guena2011a]. The resulting DCP shift corrections and their uncertainties are presented in Table \[tab\_DCP\]. Several of these uncertainties are expected to decrease when improved measurements of the tilt sensitivities and tilt uncertainties will be available. $m$ ----- ----------------------- ---------------- ---------------- ------------------------ ---------- 0 -0.7 $\pm$ 0.4 -0.4 $\pm$ 0.2 -1.5 $\pm$ 0.7 -0.2 $\pm$ 0.1 0 $\pm$ 0.8 0 $\pm$ 0.7 0 $\pm$ 1.2 0 $\pm$ 0.5 -0.7 $\pm$ 0.9 -0.4 $\pm$ 0.8 -1.5 $\pm$ 1.4 -0.2 $\pm$ 0.5 1 -9 $\pm$ 18 ($\dag$) 0 $\pm$ 4.4 0 $\pm$ 13 9.0 $\pm$ 6.2 ($\dag$) 0 $\pm$ 22 0 $\pm$ 6.3 0 $\pm$ 13 0 $\pm$ 6.2 to feeds -9 $\pm$ 27 0 $\pm$ 7.1 0 $\pm$ 16 9.0 $\pm$ 9.5 2 0 $\pm$ 1.4 ($\ddag$) -8.8 $\pm$ 4.4 -5.6 $\pm$ 2.8 -5.1 $\pm$ 2.6 0 $\pm$ 3.2 0 $\pm$ 3.0 0 $\pm$ 2.2 0 $\pm$ 1.3 0 $\pm$ 3.5 -8.8 $\pm$ 5.3 -5.6 $\pm$ 3.5 -5.1 $\pm$ 2.9 -9.7 $\pm$ 27 -9.2 $\pm$ 8.9 -7.1 $\pm$ 16 3.7 $\pm$ 9.9 Microwave lensing ----------------- The external degrees of freedom of the atoms evolve dynamically during the Ramsey interrogation. The interaction of the atoms with the standing-wave field in the microwave cavity on the upward passage slightly modifies the atomic motion, leading to a predicted shift of the clock frequency. This effect was first called “the microwave recoil" [@Wolf2004], connecting to the simple situation where an atom interacts with a plane wave as in optical spectroscopy [@Hall1976]. In the optical case, an atom absorbs a photon from a plane wave. The shift, expressed in energy, is $(\hbar k )^2/ (2 m_{at})$, corresponding to the kinetic energy of the atom with the momentum $\hbar k$ of the absorbed photon. For microwave photons, the corresponding fractional frequency shift is $1.5\times 10^{-16}$ for $^{133}$Cs, comparable to the best reported fountain accuracy. In 2006 a new approach to this effect provided better physical insight and made the calculations tractable to treat both transverse dimensions [@Gibble2006]. The microwave standing wave and its spatially varying field amplitude produces a radial magnetic dipole force which acts as a lens to focus (or defocus) the atomic wave packets, hence the name “microwave lensing frequency shift". To calculate the shift with reasonable accuracy, it is crucial to properly treat the apertures in the fountain. They truncate the downwardly traveling atomic cloud and this introduces a difference between the two dressed states, which experience opposite lensing deflections on the upward cavity traversal. Here, we report calculations of the microwave lensing for the LNE-SYRTE fountains. We apply the approach that was recently used to evaluate the NPL-CsF2 and PTB-CSF2 fountains [@Li2011][@Weyers2012]. The results are summarized in Table \[tab\_MicrowaveLensing\_a\]. The simple “analytic" result [@Li2011] (1$^{\mathrm{st}}$ line in Table \[tab\_MicrowaveLensing\_a\]) neglects all apertures on the upward passage and only treats dipole energy that is quadratic ($\propto \rho^2$) in the radial coordinate $\rho$. Since the initial clouds in the LNE-SYRTE fountains are not point-like sources, this overestimates the microwave lensing shift because it includes atoms that would be cut by the cavity apertures, which experience much larger dipole forces away from the cavity axis. Including an aperture at the first cavity passage eliminates these atoms from the calculation, giving a smaller and more realistic shift (2$^{\mathrm{nd}}$ line in Table \[tab\_MicrowaveLensing\_a\]). Adding the lowest cutoff aperture for the upward passage, the variation of the Rabi tipping angles, and the all-order dipole forces (beyond $\rho^2$), gives a small correction (3$^{\mathrm{rd}}$ line), as does including the detection inhomogeneities and imperfect Ramsey fringe contrast, yielding the predicted shift. [|l|c|c|c|c|]{} &[FO1]{} &[FO2-Cs]{} & [FOM]{} & [FO2-Rb]{}\ [Analytic]{} & $7.8$ & $8.4$ & $11.9$ & $7.0$\ & $6.1$ & $7.0$ & $9.2$ & $5.8$\ & $5.7$ & $6.5$ & $8.6$ & $5.6$\ & $6.5$ & $7.3$ & $9.0$ & $6.8$\ Blackbody radiation shift {#subsec_BBR} ------------------------- The blackbody radiation (BBR) shift is one of the largest systematic corrections and a significant source of uncertainty in atomic fountains. The bulk of this shift is simply given by the scalar Stark shift of the clock transition [@Itano1982], its response to a static electric field. A high accuracy measurement of the Stark coefficient was made early on at LNE-SYRTE with FO1 [@Simon1998] and subsequently used by most groups for their BBR corrections. These measurements where made with applied electric fields ranging from $50$ to $150$ kV.m$^{-1}$, significantly higher than the $831.9$ V.m$^{-1}$ rms value of the BBR field at 300 K. The greatly improved measurement capability enabled by the CSO (see \[subsec\_ultrastablereference\]) provides an opportunity to measure the Stark coefficient at an electric field strength that is much closer to a typical BBR field. The early measurements [@Simon1998] were specifically tested for a possible deviation from a purely quadratic dependance on the electric field. This test did not show any such deviation, which was expected since there was no theoretical basis to expect significant higher order terms given the measurement accuracy and field strengths. Nevertheless, it was important to re-examine these measurements at much lower fields given the intervening conflicting measurements and calculations of [@Micalizio2004][@Levi2004] and [@Godone2005]. These new measurements [@Rosenbusch2007] at field strengths ranging from $1.5$ to $25$ kV.m$^{-1}$, are in excellent agreement with the earlier measurements of [@Simon1998]. To extract the scalar Stark coefficient for the BBR clock correction from these measurements, the small tensor contribution has to be subtracted [@Simon1998]. In [@Ulzega2006], a sign error in the earlier theory [@Sandars1967][@Angel1968] for this tensor part was found, which was then confirmed experimentally [@Ulzega2007]. Taking into account this revisited theory, we get a corrected value for the scalar Stark coefficient: $k_0=-2.282(4)\times 10^{-10}$ Hz.V$^{-2}$.m$^{2}$. The change in the BBR fractional frequency correction from the previous value of [@Simon1998] is $6.8\times 10^{-17}$, $80\%$ of which is due to the sign correction of the theory. The uncertainty of $k_0$ continues to be dominated by the uncertainty on the geometry of the copper plates used to apply the electric field in the fountain. The value agrees with the most recent theoretical calculations [@Angstmann2006a][@Beloy2006][@Palchikov2003]. It also agrees with several other calculations [@Feichtner1965][@Lee1975][@Itano1982][@Micalizio2004], after some [@Feichtner1965][@Micalizio2004] are corrected for omitted contributions pointed out in [@Angstmann2006a] and [@Beloy2006]. The only discrepancy that seems to remain, even after the omitted contributions are included, is the calculation of [@Ulzega2006] (see the table summary in [@Rosenbusch2007]). We also note that a thermal atomic beam experiment recently provided another Stark shift coefficient measurement [@Robyr2011]. The still preliminary value is in agreement with the LNE-SYRTE value and is $\sim 5$ times less accurate. We have also performed direct measurements of the BBR shift by inserting a blackbody tube in the FO1 fountain [@Zhang2004], exposing atoms to thermal radiation at temperatures ranging from $300$ to $440$ K. Measurements were in excellent agreement with the above mentioned scalar Stark shift coefficient, but with a $22$ times larger uncertainty [@Rosenbusch2007]. On going comparisons between a room temperature and a cryogenic fountain [@Heavner2011] should yield another measurement of the Cs BBR shift at room temperature. Regarding the BBR coefficient of the $^{87}$Rb hyperfine frequency, the experimental status is unchanged from [@Bize1999]. On the theoretical side, high accuracy *ab initio* calculations have recently been published in [@Angstmann2006][@Safronova2010a]. Microwave leakage {#subsec_microwaveleakage} ----------------- In the last few years, we have developed a new approach to assess frequency shifts due to microwave leakage. Ideally, after the state selection process, atoms interact with the interrogating microwaves only in the microwave cavity during the upwards and downwards traversals. In practice, unintended interactions can occur with a residual microwave field that is coherent with the probing field. Being unintended, such a residual field is by nature uncontrolled, with a potentially complex structure with both a large standing-wave and a large traveling-wave component. Such an extraneous field, which is difficult to exclude at the required level with purely electric means, leads to a frequency shift which generally has a complicated variation with the fountain parameters, especially the microwave power [@Weyers2006][@Jefferts2005a]. If present, the frequency shift is generally difficult to disentangle from other systematic errors such as the distributed cavity phase or the microwave lensing frequency shifts. Our latest experimental approach to evaluate microwave leakage shifts uses microwave synthesizers, which not only will have intrinsically low microwave leakage but also can be switched off when the atoms are outside the microwave cavity. The challenge is to develop a switch that does not cause a large phase transient of the transmitted field after it is switched on. Figure \[fig\_synthesizer\] depicts the architecture of our synthesizers. The interrogation signal at $9.192$ GHz is generated by mixing a high power $8.992$ GHz with a small amount of $200$ MHz signal to produce only the minimum needed power. This architecture provides an ability to switch the carrier at the output by switching the $200$ MHz signal. To minimize phase transients, the $200$ MHz is switched using a Mach-Zehnder interferometric switch. The attenuator and the phase shifter are adjusted so that, when the PIN diode switch is closed, letting the signal pass through, destructive interference gives a dark fringe fringe at the output of the interferometer. When the PIN diode switch is open, the on-state signal passes the direct channel of the interferometer with no component that is likely to produce a phase transient. Specifically, the phase transient occurring in the PIN diode as it is switched is circumvented. The development of a heterodyne phase transient analyzer [@Santarelli2009] was crucial to the successful implementation of switchable microwave synthesizers in Fig. \[fig\_synthesizer\] with low phase transients. This analyzer was used extensively to select components and identify the critical parameters to achieve the desired performance. Furthermore, this device can also be used to unveil spurious phase perturbations that are synchronous with the fountain cycle, with sensitivities as low as $1~\mu$rad.s$^{-1}$ at the $9.192$ GHz carrier frequency. Such perturbations can occur if some elements of the fountain sequence (switching of fairly high levels of RF power in acousto-optic modulators, switching of the state selection microwave synthesizer, etc) perturb the interrogation synthesizer through ground loops or other electromagnetic compatibility weaknesses. Required for the heterodyne phase transient analyzer, an independent reference synthesizer, the $6-9$ GHz synthesis “for other purposes" in Fig. \[fig\_ClockEnsemble\], is distributed through an actively compensated fiber link. With a suitably characterized switchable synthesizer, it becomes possible to test for a putative microwave leakage by measuring the clock frequency with and without switched microwaves [@Santarelli2009]. All the LNE-SYRTE fountains (including FO2-Rb) are now equipped with a switchable synthesizer. Furthermore, synchronous perturbations are periodically searched for with the heterodyne phase transient analyzer. Accuracy budget --------------- Our approach to deal with other systematic shifts have been reported elsewhere and have not changed significantly since our last reviews [@Bize2004][@Bize2005]. In FO1 and FO2-Cs, the large cold collision shift [@Gibble1993] is extrapolated to zero density in real time through interleaved measurements at a high density and a halved density. We use the interrupted adiabatic passage method [@Pereira2002] which gives accurate ratios of the atomic densities. Owing to the much lower atom number dependent effects in $^{87}$Rb [@Fertig2000][@Sortais2000][@Bize2001], we change the atom number and density with the more conventional ways, either loading time or Rabi flopping in the state selection cavity. It is also possible to adopt an approach where a constant atom number is used most of the time, the collision shift being corrected based on a previously measured coefficient. The coefficient is measured with dedicated comparisons of high and low density, performed a few times per year. This approach is also used for FOM, which has a comparatively low cold collision shift due to a smaller atom number. Note that recently, another method has been proposed and demonstrated in order to deal with the cold collision shift of Cs [@Szymaniec2007][@Szymaniec2011]. The accuracy budget for the LNE-SYRTE fountains is given in Table \[tab\_accuracy\]. We give the most recent status, taking into account the latest evaluations of the distributed cavity phase shift and microwave lensing frequency shifts as reported in Table \[tab\_DCP\] and \[tab\_MicrowaveLensing\_a\]. In the absence of an independent calculation and/or experimental investigation, the uncertainty for the microwave lensing is conservatively taken to be equal to the calculated value. The last row of Table \[tab\_accuracy\] gives the status in accuracy budgets prior to 2011. ------------------------------------------------- ------------------- ------------------ ----------------- ------------------ [FO1]{} [FO2-Cs]{} [FOM]{} [FO2-Rb]{} $-1274.5 \pm 0.4$ $-1915.9\pm0.3$ $-305.6\pm 1.2$ $-3465.5\pm0.7$ [Blackbody radiation]{} $172.6 \pm 0.6$ $168.0\pm0.6$ $165.6\pm0.6$ $122.8\pm1.3$ [Collisions and cavity pulling]{} $70.5\pm 1.4$ $112.0\pm 1.2$ $28.6 \pm 5.0$ $2.0\pm2.5$ [Distributed Cavity Phase shift]{} $-1.0\pm 2.7$ $-0.9\pm 0.9 $ $-0.7\pm 1.6$ $0.4\pm 1.0$ [Spectral purity & leakage]{} $<1.0$ $<0.5$ $<4.0$ $<0.5$ [Ramsey & Rabi pulling]{} $<1.0$ $<0.1$ $<0.1$ $<0.1$ [Microwave lensing]{} $-0.7\pm0.7$ $-0.7\pm0.7$ $-0.9\pm0.9$ $-0.7\pm0.7$ [Second order Doppler shift]{} $<0.1$ $<0.1$ $<0.1$ $<0.1$ [Background collisions]{} $<0.3$ $<1.0$ $<1.0$ $<1.0$ [Total]{} $-1033.1\pm 3.5$ $-1637.5\pm 2.1$ $-113.0\pm 6.9$ $-3341.0\pm 3.3$ [Prior to 2011 (see note [@CommentAccuracy])]{} $-1031.4\pm4.1$ $-1635.9\pm 3.8$ $-111.4\pm8.1$ $-3340.7\pm4.2$ ------------------------------------------------- ------------------- ------------------ ----------------- ------------------ Applications in time and frequency metrology ============================================ $^{133}$Cs Fountain comparisons ------------------------------- For more than a decade, the development of LNE-SYRTE fountain ensemble of Fig. \[fig\_ClockEnsemble\] has allowed high-accuracy comparisons between atomic fountain clocks. This has especially been the case in the last few years during which hundreds of days of clock comparisons have accumulated. To deal with the differences in clock cycles, mode of operation, schedule for verifying the magnetic field, etc, we apply the following data processing scheme. Each fountain produces measurements, corrected for all systematic biases, at a rate of 1 measurement per cycle. When an interleaved sequence of high and low density measurements is used, this first step requires to use these interleaved data to determine the cold collision shift coefficient for a particular time window, and then to correct the high and low density frequency measurements accordingly. The individual fountain measurements are averaged over time intervals of 864 s (1/100th of a day) starting at the beginning of the day. Intervals are discarded if the number of valid fountain cycles is below a chosen threshold, for instance, due to an automated verification of the magnetic field, a laser going out of lock, or a phase/frequency jump in the output of the cryogenic oscillator. Other filters can be applied based on fault of the CSO, one of the fiber links, or other subsystems. This leaves a set of validated data where the sampling sequence is matched between all fountains. These data represent measurements by each fountain of the frequency of the common ultra-stable reference, i.e. the 11.98 GHz weakly phase locked to one of the hydrogen masers (see Fig. \[fig\_ClockEnsemble\]), which are used for calibration of the international atomic time (see \[subsec\_TAI\] below). As a third step, the frequency difference between the synchronous validated data of two fountains eliminates the frequency of the common-mode ultra-stable reference and allows a direct comparison. An example of frequency instability between FOM and FO2 over 50 effective days of averaging is shown in Fig. \[fig\_StabilityFOMFO2\]. Here, the overall short-term instability is $1.3\times 10^{-13}\tau^{-1/2}$ and the measurement resolution is $\sim 2$ part in $10^{16}$ after averaging for $4.5$ days. The best reported instability for this type of comparison is $5.0\times 10^{-14}\tau^{-1/2}$ [@Bize2004][@Bize2005][@Vian2005]. ------------- --------------------- --------- --------- ------------- ---------------- ----------- ---------------- ----------- [Duration (days)]{} [$D$]{} [$U$]{} [$|D/U|$]{} [$\chi_D^2$]{} [$Q_D$]{} [$\chi_0^2$]{} [$Q_0$]{} $320$ $-3.0$ $9.1$ $0.33$ $0.28$ $0.996$ $0.37$ $0.983$ [FOM-FO1]{} $240$ $4.1$ $9.2$ $0.44$ $0.12$ $0.999$ $0.28$ $0.989$ [FO1-FO2]{} $500$ $-5.7$ $6.2$ $0.92$ $1.48$ $0.55$ $2.12$ $0.001$ ------------- --------------------- --------- --------- ------------- ---------------- ----------- ---------------- ----------- Fountain comparisons following the above data processing scheme are now performed regularly at LNE-SYRTE. It was recently incorporated into a fully automated procedure. In addition to daily updates of the fountain comparisons, an overview of the systems’ parameters is generated to help assess the status of the clock ensemble in quasi-realtime with hourly updates. To produce fully validated data for fountain comparisons, for example for timescale steering and TAI calibrations, the entire data processing is critically scrutinized and redone if necessary. Figure \[fig\_DiffFontaines\] shows the monthly average of fractional frequency differences $D$ between the three Cs fountain pairs from November 2006 to April 2011. These data correspond to a total of $\sim 1000$ effective days of fountain comparison. The error bars are the total uncertainties for a given clock pair, which are dominated by the systematic uncertainties of the two fountains. The overall fractional frequency differences $D$ (weighted averages) are reported in Table \[tab\_comparison\], together with the corresponding overall uncertainties $U$. $|D/U|$ represents the deviation from a null difference scaled to the $1\sigma$ total uncertainty. The reduced $\chi^{2}$, either with respect to the mean ($\chi_D^2$) or with respect to zero ($\chi_0^2$) gives a further consistency check of these data. Finally, the goodness-of-fit[^11] as defined in [@NR1992] can be computed, either with respect to the mean ($Q_D$) or with respect to zero ($Q_0$). The $Q$’s very close to 1 for FOM-FO2 and FOM-FO1 indicate that fluctuations about the means are smaller than the total uncertainties. $Q_0$ which tests deviations (and/or fluctuations) from the expected null difference shows that FOM-FO2 and FOM-FO1 agree within the total uncertainties. Instead the low $Q_D$ and $Q_0$ for the more stringent FO1-FO2 comparison would indicate a slightly underestimated systematic effect, of a few $10^{-16}$, which is under investigation. Note that for the data of Fig. \[fig\_DiffFontaines\], the modifications from our recent study of the DCP shift (see \[subsec\_DCP\] and [@Guena2011]) were not yet implemented. Consequently, the DCP shift is one potential candidate to explain these deviations. Future comparisons, after a full implementation to all fountains of the modified approach to the DCP shift, will clarify this point. Note that a somewhat similar analysis of worldwide comparisons of fountain primary frequency standards exploiting their calibrations of TAI is reported in [@Parker2010]. A secondary representation of the SI second ------------------------------------------- Following the early demonstration of the low cold collision shift in $^{87}$Rb [@Fertig2000][@Sortais2000], LNE-SYRTE has performed a series of high accuracy absolute measurements of the $^{87}$Rb hyperfine frequency against Cs fountain clocks [@Bize1999][@Marion2003][@Guena2010]. The 2003 comparison was the most accurate of any atomic frequency comparison at the time. Since then, it has been superseded by the rapidly improving optical frequency measurements. The $^{87}$Rb hyperfine frequency was the first secondary representation of the SI second recognized by BIPM [@CCTF2004]. Its recommended value is given by the 2002 LNE-SYRTE measurement: 6 834 682 610.904324 Hz with a recommended uncertainty of $21~\mu$Hz or $3$ parts in $10^{15}$. This recommended uncertainty was chosen to be $\sim 3$ times larger than the actual uncertainty of the measurement ($1.3\times 10^{-15}$). Our latest reported measurement ($4~\mu$Hz uncertainty or $5.9$ parts in $10^{16}$), together with an analysis of the consistency of all our measurements can be found in [@Guena2010]. An independent $^{87}$Rb *versus* $^{133}$Cs fountain comparison has been reported for the first time in 2010. Agreement with LNE-SYRTE measurements is claimed at the $10^{-15}$ level but the actual result of the measurement has not yet been reported [@Ovchinnikov2011][@Li2011b][@Ovchinnikov2011a]. Also notable is the construction of at least 8 Rb fountains at the US Naval Observatory [@Peil2006][@Peil2011]. Contributions to TAI {#subsec_TAI} -------------------- Since 2007, LNE-SYRTE fountains have measured the frequency of one of LNE-SYRTE H-masers almost continuously. These data are used to calibrate the International Atomic Time (TAI). Figure \[fig\_TAI\] shows all calibrations of TAI by LNE-SYRTE cesium fountains from November 2004. Each data point corresponds to a measurement, by one fountain, of the H-maser average frequency over a $20-30$ days period, reaching a typical statistical uncertainty of $3\times 10^{-16}$, which is augmented in quadrature by an uncertainty due to dead times of typically $1.5\times 10^{-16}$ [@CircularT]. Between the 2007 and 2010 period, LNE-SYRTE reported 78 formal evaluations to BIPM, out of a total of 161 fountain reports worldwide, i.e. LNE-SYRTE has contributed almost $50\%$ of all reports since 2007. For contributions to TAI, the gravitational redshift corrections are $-69.3\times 10^{-16}$, $-65.4\times 10^{-16}$ and $-68.7\times 10^{-16}$ for FO1, FO2 and FOM respectively, with an uncertainty of $1.0\times 10^{-16}$ corresponding to a $1$ m uncertainty of height above the geoid. Contributions to TAI also provide a vehicle to compare to other fountain clocks around the world. The latest update on these comparisons appears in [@Parker2010]. Since mid-2006, fountain data, including from FO2-Rb, are used to steer the French Atomic Time TA(F) on a monthly basis. TA(F) is a timescale based on an ensemble of commercial Cs clocks located in France and elaborated at LNE-SYRTE. This steering has improved the long-term stability of the timescale with respect to TAI [@Uhrich2008]. Absolute measurements of optical frequencies -------------------------------------------- The accuracy and stability of optical atomic clocks have now largely surpassed those of microwave clocks, and notably atomic fountain clocks [@Chou2010][@Lemke2009a][@Rosenband2008][@Ludlow2008]. This opens a path to redefining of the SI second with an optical transition [@CGPM2009]. One of the prerequisites to a redefinition is absolute frequency measurements of potential optical transitions with the highest accuracy. The LNE-SYRTE fountain clock ensemble (Fig. \[fig\_ClockEnsemble\]) offers some unique features to contribute to this task. This includes the availability of several fountains for redundant measurements, of an ultra-low noise reference signal that provides the best reported stability to date between a microwave and optical clock [@Baillard2007b], and of the transportable fountain FOM for measuring optical frequency at other laboratories. Absolute frequency measurements of optical frequencies can also probe the stability of fundamental constants [@Blatt2008][@Fischer2004][@Fortier2007][@Peik2004], determine the Rydberg constant [@Mohr2008] and test Quantum Electrodynamics (QED) [@Karshenboim2005]. Next, we give an overview of absolute optical frequency measurements made with LNE-SYRTE fountain ensemble. ### Hydrogen $\mathrm{1S-2S}$ transition measured with $\mathrm{FOM}$ at $\mathrm{MPQ}$ Garching, Germany {#subsec_H1S2S} At the Max Planck Institut für Quantenoptik (MPQ) in Garching (Germany), we have measured the absolute frequency of the 1S-2S optical transition via 2 photon excitation at 246 nm [@Niering2000][@Fischer2004][@Parthey2011a]. Three measurement campaigns were completed in 1999, 2003 and 2010. For these experiments a femto-comb laser is used to measure the frequency of the excitation laser. The repetition rate is locked to a hydrogen maser whose frequency is simultaneously measured directly by FOM. The noise of the maser signal is filtered by phase-locking a BVA quartz oscillator to improve the measurement stability. The last measurement in 2010 had the best fractional uncertainty: $4.2\times 10^{-15}$ [@Parthey2011a]. The measured 1S-2S transition frequency is 2 466 061 413 187 035(10) Hz. High resolution spectroscopy of hydrogen is interesting to test highly accurate atomic structure calculations for QED tests [@Pohl2010]. Also, the recent capture of antihydrogen at CERN [@AlphaCollaboration2011] is a first step toward cooling of antimatter to enable high resolution spectroscopy of antihydrogen for fundamental tests of charge, parity, time reversal (CPT) symmetry and the gravity of antimatter. ### $^{40}\mathrm{Ca}^{+}$ optical clock measured with $\mathrm{FOM}$ at the University of Innsbruck, Austria In 2007 at the Institut für Experimentalphysik and Quantenoptik, University of Innsbruck (Austria), we measured the absolute frequency of the $4s~^{2}S_{1/2}-3d~^{2}D_{5/2}$ optical quadrupole transition of a trapped single $^{40}$Ca$^{+}$ ion. The frequency of this transition is 411 042 129 776 393.2 (1.0) Hz [@Chwalla2009]. This corresponds to a fractional frequency uncertainty of 2.4 parts in $10^{15}$. During this experiment, FOM was used in the autonomous configuration mentioned in \[subsec\_FOM\], where a BVA quartz oscillator is frequency locked to the FOM spectroscopic signal. The output of the BVA quartz oscillator synchronizes the repetition rate of the femto-comb laser that measures the probe laser frequency for the $^{40}$Ca$^{+}$ transition. This quadrupole transition in $^{40}$Ca$^{+}$ ion is another candidate for an optical clock and for tests of the stability of fundamental constants. ### $^{87}\mathrm{Sr}$ and $^{88}\mathrm{Sr}$ optical lattice clocks measured at $\mathrm{SYRTE}$ {#subsec_Sr} LNE-SYRTE is developing two Sr optical lattice clocks. During their development, several absolute frequency measurements of the Sr clock frequency were made [@Baillard2007b][@Baillard2007][@LeTargat2006][@Courtillot2003]. The best reported uncertainty for these measurements is $2.6\times 10^{-15}$ [@Baillard2007b]. Both Titanium:Sapphire based and Er-doped fiber based optical frequency combs were used to measure the frequency of the 698 nm laser light stabilized to the Sr atoms. The 2006 measurements of the $^{87}$Sr used a fiber optical frequency comb transported from the Physikalisch Technische Bundesanstalt (PTB) in Braunschweig (Germany) [@LeTargat2006][@Baillard2007b]. As already mentioned, the ultra-stable reference based on a CSO (Fig. \[fig\_ClockEnsemble\]) allows the best reported short-term stability to date for comparing an optical clock to a microwave clock: $6\times 10^{-14}\tau^{-1/2}$. In the recent measurements, the optical frequency comb is stabilized to an ultra-stable laser as described in [@Dawkins2010], making the repetition rate of the femtosecond laser an ultra-stable microwave signal. This signal is then measured against the CSO based ultra-stable reference which is in turn measured by the fountain clocks. Simultaneously, the beat between the optical frequency comb and the 698 nm light is counted to extract its absolute frequency. Combining these with similar measurements by other institutes, stringently tests the stability of fundamental constants [@Blatt2008]. ### Measurement of $^{199}\mathrm{Hg}$ and $^{201}\mathrm{Hg}$ optical clock transition frequencies at $\mathrm{SYRTE}$ LNE-SYRTE is developing an optical lattice clock based on neutral mercury. The first absolute frequency measurements of the $^{199}$Hg and $^{201}$Hg optical clock transition were in 2008 [@Petersen2008a]. These initial measurements improved the previous knowledge of the transition frequency by more than 4 orders of magnitude, to a fractional frequency uncertainty of $5\times 10^{-12}$. More recently, measurements against LNE-SYRTE ultra-stable reference were used to perform the first experimental determination of the magic wavelength for Hg [@Yi2011]. Ongoing development of the Hg clock is expected to soon yield measurements at the $10^{-15}$ level and beyond. Other measurements with the transportable fountain $\mathrm{FOM}$ ----------------------------------------------------------------- We have performed frequency comparisons with the engineering model of the space clock PHARAO (see \[subsec\_PHARAO\] below and [@Laurent2006]) between 2007 to 2009, at the Centre National d’Etudes Spatiales (CNES) in Toulouse, France. We have also verified the frequency performance of the ACES architecture [@Cacciapuoti2007]. The test included a ground model of the onboard data handling unit (XPLC), the Frequency Comparison and Distribution Package (FCDP), the PHARAO engineering model, and a ground model of the Space Hydrogen Maser (SHM). The two clocks are combined to generate a timescale with the short-term stability of SHM and the long-term stability and accuracy of PHARAO. Remote comparisons via T2L2 satellite laser link ------------------------------------------------ In 2010, time transfer by the T2L2 laser link *via* the JASON2 satellite [@Guillemot2008] was tested by comparing FOM at the Observatoire de la Côte d’Azur (OCA) in Grasse (France) with the other LNE-SYRTE fountains at the Observatoire de Paris. The common time transfer techniques, carrier phase GPS and TWSTFT, were also used in parallel. Notable in this comparison is the large gravitational redshift, $-1.384\times 10^{-13}$, at the OCA altitude of 1268 m. Preliminary results [@Samain2011] show that all the time transfer methods are consistent to within 2 ns over 2 months. The frequency difference between the remote fountains is measured with a typical uncertainty of $1\times 10^{-15}$, which determines the gravitational redshift difference to $10^{-2}$. It now remains to refine the analysis and to evaluate the ultimate performances of the T2L2 time transfer link. Note that at MPQ, at CNES and at OCA, FOM also provided calibration of the TAI by using a GPS receiver and a carrier phase GPS analysis software developed by CNES or by NRCAN. This is the first time that a primary frequency standard contributes to steering the TAI from different sites with different gravitational redshift [@Rovera2011]. Applications in fundamental physics =================================== One of the most interesting applications of atomic clocks is to test fundamental physics. In this section, we review contributions of LNE-SYRTE fountain ensemble to such tests. Stability of constants ---------------------- Repeated high accuracy comparisons between atomic (or molecular) frequencies can test the stability of fundamental constants such as the fine structure constant $\alpha$ or the electron to proton mass ratio $m_e/m_p$ (see, for instance, [@Karshenboim2000b] [@Flambaum2008a] and references therein). Laboratory experiments usefully complement tests over cosmological timescales since the interpretation does not rely on any assumptions about a cosmological model [@Uzan2003][@Uzan2011]. A first test performed with LNE-SYRTE fountain ensemble comes from a series of $^{87}$Rb *versus* $^{133}$Cs hyperfine frequency comparisons made during the development of the FO2 dual Rb/Cs fountain [@Marion2003][@Bize2004][@Bize2005] and [@Guena2008]. In this last 2008 report, a putative time variation of the ratio of the two hyperfine frequencies is constrained to $d\ln(\nu_{\mathrm{Rb}}/\nu_{\mathrm{Cs}})/dt=(-3.2\pm 2.3)\times 10^{-16}$ yr$^{-1}$. In terms of fundamental constants, this result yields $d\ln(\alpha^{-0.49} [g_{\mathrm{Rb}}/g_{\mathrm{Cs}}])/dt=(-3.2\pm 2.3)\times 10^{-16}$ yr$^{-1}$ where $g_{\mathrm{Rb}}$ and $g_{\mathrm{Cs}}$ are the nuclear g-factors in $^{87}$Rb and $^{133}$Cs. Expressing these g-factors in terms of fundamental parameters of the Standard Model [@Flambaum2004][@Flambaum2006] gives: $d\ln(\alpha^{-0.49} [m_q/\Lambda_{\mathrm{QCD}}]^{-0.025})/dt=(-3.2\pm 2.3)\times 10^{-16}$ yr$^{-1}$, where $\Lambda_{\mathrm{QCD}}$ is the mass scale of Quantum Chromodynamics (QCD). An improved analysis including more recent measurements, with FO2 in the dual fountain configuration, will improve this value significantly. A second test comes from a series of absolute frequency measurements of the $^{87}$Sr optical lattice clock by LNE-SYRTE (see \[subsec\_Sr\]), the University of Tokyo, Japan and JILA, Boulder, Colorado, USA [@Blatt2008]. Together, these constrain the putative variation of $\nu_{\mathrm{Sr}}/\nu_{\mathrm{Cs}}$ to $d\ln(\nu_{\mathrm{\mathrm{Sr}}}/\nu_{\mathrm{\mathrm{Cs}}})/dt=(-7\pm 18)\times 10^{-16}$ yr$^{-1}$ corresponding to $d\ln(\alpha^{2.77}[m_q/\Lambda_{\mathrm{QCD}}]^{-0.039}[m_e/\Lambda_{\mathrm{QCD}}])/dt=(7\pm 18)\times 10^{-16}$ yr$^{-1}$. The same measurements constrain the putative variation of this same combination of constants with the gravitational potential to $c^2 d\ln(\alpha^{2.77}[m_q/\Lambda_{\mathrm{QCD}}]^{-0.039}[m_e/\Lambda_{\mathrm{QCD}}])/dU=(-5.8\pm 8.9)\times 10^{-6}$, where $c$ is the speed of light and $U$ the gravitational potential [^12]. This test relies on exploiting the yearly modulation of the gravitational potential of the Sun due the eccentricity of the Earth orbit. Measurements of the H(1S-2S) transition performed at MPQ Garching (see \[subsec\_H1S2S\]) using the transportable fountain FOM as a reference, offer a third test: $d\ln(\nu_{\mathrm{H(1S-2S)}}/\nu_{\mathrm{Cs}})/dt=(-3.2\pm 6.3)\times 10^{-15}$ yr$^{-1}$ [@Fischer2004]. This translates to $d\ln(\alpha^{2.83}[m_q/\Lambda_{\mathrm{QCD}}]^{-0.039}[m_e/\Lambda_{\mathrm{QCD}}])/dt=(3.2\pm 6.3)\times 10^{-15}$ yr$^{-1}$, a constraint which will soon improve as a result of a recent measurement campaign [@Parthey2011a]. A fourth test comes from a series of comparisons between the $^{133}$Cs and the hydrogen hyperfine frequencies. LNE-SYRTE fountain’s contributions to this test were the calibrations of TAI reported in BIPM *Circular T* (see \[subsec\_TAI\]) enabling the connection with 4 hydrogen masers participating to the elaboration of the NIST AT1 atomic timescale (see for instance [@Parker1997]). Other Cs fountain contributions to the experiment came from PTB in Germany and from the Istituto Nazionale di Ricerca Metrologica (INRIM) in Italy. The analysis was performed at the National Institute of Standards and Technology (NIST) Boulder, Colorado, USA [@Ashby2007] and gives $|c^2 d\ln(\nu_{\mathrm{H}}/\nu_{\mathrm{Cs}})/dU|=(0.1\pm 1.4)\times 10^{-6}$, corresponding to $|c^2 d\ln(\alpha^{0.83}[m_q/\Lambda_{\mathrm{QCD}}]^{0.11})/dU|=(0.1\pm 1.4)\times 10^{-6}$. Note that instead of using $\alpha$, $m_q/\Lambda_{\mathrm{QCD}}$ and $m_e/\Lambda_{\mathrm{QCD}}$ as independent variables, $\alpha$, $\mu=m_e/m_p$ and $m_q/m_p$ can be used. The link between the two approaches is $d\ln(m_p/\Lambda_{\mathrm{QCD}})\simeq 0.048\times d\ln(m_q/\Lambda_{\mathrm{QCD}})$ [@Flambaum2006]. Test of Lorentz Invariance in the matter sector ----------------------------------------------- The LNE-SYRTE FO2 fountain has tested the anisotropy of space as in Hughes-Drever experiments (see for instance [@Will2006]). For this experiment, the fountain sequence is tailored to probe opposing Zeeman transitions in $^{133}$Cs to test for a putative variation of their frequencies when the orientation of the quantization axis changes (here, due the Earth rotation) with respect to a supposedly preferred frame [@Wolf2006], such as the frame of the Cosmic Microwave Background. The experiment was interpreted within the framework of the Lorentz violating Standard Model Extension (SME), where it is sensitive to proton parameters corresponding to a largely unexplored region of the SME parameter space. The constraints for 4 parameters, already constrained by other measurements, were improved by as much as 13 orders of magnitude and 4 parameters were constrained for the first time. Operating FO2 in the dual Rb/Cs configuration [@Guena2010] opens new possibilities for improving and complementing these tests. Other tests ----------- The development of LNE-SYRTE fountain ensemble also offered a test of Lorentz Local Invariance in the photon sector, not using the fountains, but simply comparing the CSO to a hydrogen maser (see Fig. \[fig\_ClockEnsemble\]) for a duration that now exceeds 10 years. This experiment is to date the most stringent Kennedy-Thorndike test by a factor of $500$. The latest update is published in [@Tobar2010]. The experiment can also be interpreted in the SME framework [@Wolf2004a]. Development of the PHARAO cold atom space clock {#subsec_PHARAO} =============================================== Atomic Clocks Ensemble in Space (ACES) is a European Space Agency (ESA) fundamental physics mission originally proposed by Laboratoire Kastler Brossel and SYRTE. It is based on the operation of highly stable and accurate atomic clocks in the microgravity environment of the International Space Station (ISS) [@Cacciapuoti2007]. The time scale generated by the ACES clocks on-board the ISS is delivered to Earth through a high-performance two-way time and frequency transfer link. The clock signal is used to perform space-to-ground as well as ground-to-ground comparisons of atomic frequency standards. The ACES scientific objectives cover both fundamental physics and applications. Tests of Special and General Relativity will be performed with improved accuracy, as well as a search for temporal variations of fundamental constants. On the application side, frequency comparisons between distant clocks, both space-to-ground and ground-to-ground, will be performed worldwide with unprecedented resolution. ACES will demonstrate a new type of “relativistic geodesy", which, based on a precision measurement of the Einstein’s gravitational red-shift, will resolve differences in the Earth gravitational potential at the $10$ cm level. Finally, ACES will contribute to the improvement of global navigation satellite systems (GNSS) and to future evolutions of these systems. It will demonstrate new methods to monitor the ocean surface based on scatterometric measurements of the GNSS signal and it will contribute to the monitoring of the Earth atmosphere through radio-occultation experiments. The expected performance is a time stability of 10 ps over ten days (1 ps for common view comparisons between ground clocks) and a frequency accuracy better than $3\times 10^{-16}$. The accuracy and the long term frequency stability are defined by the PHARAO clock, an instrument developed by the French space agency CNES and SYRTE. The clock uses cold cesium atoms, slowly moving through a Ramsey cavity. We have fully tested the engineering model (PHARAO-EM) of the clock [@Laurent2006]. Figure \[fig\_PHARAO\] shows a picture of the PHARAO-EM under test at the CNES assembly room in Toulouse. For nominal operation the frequency stability is $3.3\times 10^{-13} \tau^{-1/2}$. The central Ramsey resonance linewidth is 5.6 Hz and $10^{6}$ atoms are detected. The contrast is only $93\%$ (Fig. \[fig\_fringesPHARAO\]) since, on Earth, the atoms do not spend the same time in the two interaction zones of the Ramsey cavity. On Earth the clock cycle duration lasts 790 ms with a microwave interrogation time of $90$ ms. Consequently the phase noise of the microwave signal, synthesized from a quartz oscillator, contributes to the frequency instability at a level of $2.2\times 10^{-13}\tau^{-1/2}$. We have performed an accuracy evaluation of the clock. The main frequency corrections and their uncertainties are shown in table \[tab\_accuracyPHARAO\]. The total frequency uncertainty is $1.9\times 10^{-15}$.The largest uncertainty is the quadratic Zeeman shift because the magnetic field along the atomic trajectories is not homogeneous due to a problem with the innermost magnetic shield. The second largest contribution is the cold collision frequency shift. The frequency shift per detected atom is similar to other cold cesium clocks and the uncertainty is mainly limited by the measurement duration (only two weeks). The blackbody radiation shift is deduced from measurements of the ambient temperature. A technical problem prevented us from using the calibrated temperature probes located next to the Ramsey cavity. Finally, the first order Doppler frequency shift has a large value since the atom’s axial velocity decreases due to gravity field, which causes the longitudinal DCP shifts from the 2 interactions to not cancel. This gives a large DCP shift that has a large dependance on the microwave field amplitude. -------------------------------------- ---------------- ----------------- [Correction]{} [Uncertainty]{} $-5633$ $14$ [Blackbody radiation]{} $162.5$ $2.2$ [Cold collisions & cavity pulling]{} $63$ $11$ [First order Doppler]{} $-33$ $5$ [Total]{} $-5440$ $19$ -------------------------------------- ---------------- ----------------- : Systematic fractional frequency corrections for the PHARAO engineering model during the 2009 tests, in units of $10^{-16}$.[]{data-label="tab_accuracyPHARAO"} The PHARAO-EM clock, operated in several configurations, was compared to the transportable fountain FOM. After applying all systematic corrections (Table \[tab\_accuracy\] and \[tab\_accuracyPHARAO\]), the frequency differences were smaller than $2\times 10^{-15}$, with a resolution of $5\times 10^{-16}$. During these measurements, the accuracies of the clocks were $8.1\times 10^{-16}$ for FOM and $1.9\times 10^{-15}$ for PHARAO-EM. Consequently, there is no significant frequency offset between the clocks. Testing the engineering model demonstrated the full operation of the PHARAO clock architecture and highlighted some improvements to implement in the flight model. The PHARAO flight model is now being assembled and it will be delivered in 2012 together with the other ACES instruments. The launch date will be 2 years after the delivery of all the instruments. SYRTE will be a master station to survey and to analyze the ACES signal. One of the microwave link station developed for ACES will be installed on the roof of a building at the Observatoire de Paris. Hydrogen masers steered by the atomic fountains and UTC(OP) will synchronize the station, forming the ACES ground segment at SYRTE. Strontium and mercury optical lattice clocks at SYRTE [@Westergaard2011][@Yi2011], which are connected to the fountain ensemble *via* optical frequency combs, will also contribute to the ACES mission. Notably, ACES will provide ground-to-ground comparisons of these optical clocks to similar clocks at other institutes around the world, demonstrating relativistic geodesy. Conclusions =========== This paper gives an extensive overview of the development of the LNE-SYRTE fountain ensemble. We can also identify some priorities for the future. For fountains, it will be important to further improve inter-comparisons to reduce inaccuracies to 1 to 2 parts in $10^{16}$, and to possibly remove yet unidentified sources of uncertainty or instability. Implementing more sophisticated microwave cavity designs is one interesting approach. Simultaneously, the reliability of the fountain ensemble must further improve. This is required by the ground segment of the ACES mission - quasi-continuous operation at the best accuracy of the three fountains for at least 18 months. This will be achieved by a number of technical improvements in the fountains (notably the laser systems), the fountain environment, and by complementing the cryogenic sapphire oscillator with a second low-noise microwave source derived from an ultra-stable laser. This system will also ensure a permanent link to the optical clocks. Achieving these goals will enable new or improved applications: realization of an ultra-stable local timescale, continuously steered to the atomic fountains, even more regular calibrations of TAI, improved measurements of optical frequencies, enhanced tests of fundamental physics, measurement of the gravitational red shift within the ACES project, and remote clock comparisons *via* ACES. Acknowledgments {#acknowledgments .unnumbered} =============== SYRTE is SYstèmes de Référence Temps-Espace. SYRTE is UMR CNRS 8630 between Centre National de la Recherche Scientifique (CNRS), Université Pierre et Marie Curie (UPMC) and Observatoire de Paris. LNE, Laboratoire National de Métrologie et d’Essais, is the French National Metrology Institute. SYRTE is a member of the Institut Francilien de Recherche sur les Atomes Froids (IFRAF) and of the C’Nano network of the Région Île de France. K. Gibble acknowledges support from the NSF, Penn State, and la Ville de Paris. We acknowledge a tremendous number of past contributions over 2 decades of development of the LNE-SYRTE fountain ensemble. We specifically acknowledge the contribution of SYRTE technical services. The ACES mission is an ESA project. Within ACES, the PHARAO cold atom space clock is developed and funded by CNES. The transportable fountain FOM is also funded by CNES to a large extent. [100]{} \[1\][\#1]{} url@samestyle \[2\][\#2]{} \[2\][[l@\#1=l@\#1\#2]{}]{} A. Clairon, P. Laurent, G. Santarelli, S. Ghezali, S. Lea, and M. Bahoura, “A cesium fountain frequency standard: recent results,” *IEEE Trans. on Inst. and Meas.*, vol. 44, no. 2, p. 128, 1995. P. Laurent, P. Lemonde, E. Simon, G. Santarelli, A. Clairon, N. Dimarcq, P. Petit, C. Audoin, and C. Salomon, “A cold atom clock in absence of gravity,” *Euro. Phys. J. D*, vol. 3, p. 201, 1998. J. Gu[é]{}na, P. Rosenbusch, P. Laurent, M. Abgrall, D. Rovera, G. Santarelli, M. Tobar, S. Bize, and A. Clairon, “Demonstration of a dual alkali [R]{}b/[C]{}s fountain clock,” *Ultrasonics, Ferroelectrics and Frequency Control, IEEE Transactions on*, vol. 57, no. 3, p. 647, 2010. K. Gibble and S. Chu, “A laser cooled [Cs]{} frequency standard and a measurement of the frequency shift due to ultra-cold collisions,” *Phys. Rev. Lett.*, vol. 70, p. 177, 1993. X. Baillard, A. Gauguet, S. Bize, P. Lemonde, P. Laurent, A. Clairon, and P. Rosenbusch, “Interference-filter-stabilized external-cavity diode lasers,” *Opt. Commun.*, vol. 266, p. 609, 2006. A. Clairon, C. Salomon, S. Guellati, and W. Phillips, “[Ramsey resonance in a Zacharias fountain]{},” *Europhys. Lett.*, vol. 16, p. 165, 1991. D. Chambon, M. Lours, F. Chapelet, S. Bize, M. E. Tobar, A. Clairon, and G. Santarelli, “Design and metrological features of microwave synthesizers for atomic fountain frequency standard,” *IEEE Trans. Ultr. Ferr. Freq. Contr.*, vol. 54, p. 729, 2007. G. Santarelli, P. Laurent, P. Lemonde, A. Clairon, A. G. Mann, S. Chang, A. N. Luiten, and C. Salomon, “Quantum projection noise in an atomic fountain: A high stability cesium frequency standard,” *Phys. Rev. Lett.*, vol. 82, no. 23, p. 4619, 1999. C. Vian, P. Rosenbusch, H. Marion, S. Bize, L. Cacciapuoti, S. Zhang, M. Abgrall, D. Chambon, I. Maksimovic, P. Laurent, G. Santarelli, A. Clairon, A. Luiten, M. Tobar, and C. Salomon, “[BNM-SYRTE]{} [F]{}ountains: [R]{}ecent [R]{}esults,” *IEEE Trans. Instrum. Meas.*, vol. 54, p. 833, 2005. G. Santarelli, G. Governatori, D. Chambon, M. Lours, P. Rosenbusch, J. Gu[é]{}na, F. Chapelet, S. Bize, M. Tobar, P. Laurent, T. Potier, and A. Clairon, “Switching atomic fountain clock microwave interrogation signal and high-resolution phase measurements,” *Ultrasonics, Ferroelectrics and Frequency Control, IEEE Transactions on*, vol. 56, no. 7, pp. 1319–1326, July 2009. K. Dieckmann, R. J. C. Spreeuw, M. Weidemüller, and J. T. M. Walraven, “Two-dimensional magneto-optical trap as a source of slow atoms,” *Phys. Rev. A*, vol. 58, no. 5, p. 3891, Nov 1998. F. [Pereira Dos Santos]{}, H. Marion, S. Bize, Y. Sortais, A. Clairon, and C. Salomon, “Controlling the cold collision shift in high precision atomic interferometry,” *Phys. Rev. Lett.*, vol. 89, p. 233004, 2002. F. Chapelet, “Fontaine atomique double de [Cé]{}sium et de [R]{}ubidium avec une exactitude de quelques $10^{-16}$ et applications,” Ph.D. dissertation, department of physics, Universit[é]{} de Paris XI, 2008. Y. Sortais, “Construction d’une fontaine double à atomes froids de $^{87}$[Rb]{} et de $^{133}$[Cs]{}, étude des effets dépendant du nombre d’atomes dans une fontaine.” Ph.D. dissertation, department of physics, Universit[é]{} Pierre et Marie Curie, Paris VI, 2001. S. Bize, “Tests fondamentaux [à]{} l’aide d’horloges [à]{} atomes froids de rubidium et de cesium,” Ph.D. dissertation, department of physics, Universit[é]{} Pierre et Marie Curie, Paris VI, 2001. R. Legere and K. Gibble, “Quantum scattering in a juggling atomic fountain,” *Phys. Rev. Lett.*, vol. 81, p. 5780, 1998. \[Online\]. Available: <http://link.aps.org/doi/10.1103/PhysRevLett.81.5780> R. A. Hart, X. Xu, R. Legere, and K. Gibble, “A quantum scattering interferometer,” *Nature*, vol. 446, no. 7138, pp. 892–895, Apr. 2007. \[Online\]. Available: <http://dx.doi.org/10.1038/nature05680> F. Allard, I. Maksimovic, M. Abgrall, and P. Laurent, “Automatic system to control the operation of an extended cavity diode laser,” *Rev. Sci. Instrum.*, vol. 75, p. 54, 2004. \[Online\]. Available: <http://link.aip.org/link/?RSI/75/54/1> A. Luiten, A. Mann, M. Costa, and D. Blair, “Power stabilized cryogenic sapphire oscillator,” *IEEE Trans. Instr. Meas.*, vol. 44, p. 132, 1995. A. Mann, C. Sheng, and A. Luiten, “Cryogenic sapphire oscillator with exceptionally high frequency stability,” *IEEE Trans. on Inst. and Meas.*, vol. 50, p. 519, 2001. D. Chambon, S. Bize, M. Lours, F. Narbonneau, H. Marion, A. Clairon, G. Santarelli, A. Luiten, and M. Tobar, “Design and realization of a flywheel oscillator for advanced time and frequency metrology,” *Rev. Sci. Instrum.*, vol. 76, p. 094704, 2005. S. Dawkins, R. Chicireanu, M. Petersen, J. Millo, D. Magalhães, C. Mandache, Y. [Le Coq]{}, and S. Bize, “An ultra-stable referenced interrogation system in the deep ultraviolet for a mercury optical lattice clock,” *Appl. Phys. B: Lasers and Optics*, vol. 99, p. 41, 2010. \[Online\]. Available: <http://dx.doi.org/10.1007/s00340-009-3830-3> C. Daussy, O. Lopez, A. Amy-Klein, A. Goncharov, M. Guinet, C. Chardonnet, F. Narbonneau, M. Lours, D. Chambon, S. Bize, A. Clairon, G. Santarelli, M. E. Tobar, and A. N. Luiten, “Long-distance frequency dissemination with a resolution of $10^{-17}$,” *Phys. Rev. Lett.*, vol. 94, no. 20, p. 203904, 2005. \[Online\]. Available: <http://link.aps.org/abstract/PRL/v94/e203904> J. Hartnett and N. Nand, “Ultra-low vibration pulse-tube cryocooler stabilized cryogenic sapphire oscillator with $10^{-16}$ fractional frequency stability,” *Microwave Theory and Techniques, IEEE Transactions on*, vol. 58, no. 12, pp. 3580 –3586, dec. 2010. J. Hartnett, N. Nand, C. Wang, and J.-M. Le Floch, “Cryogenic sapphire oscillator using a low-vibration design pulse-tube cryocooler: first results,” *Ultrasonics, Ferroelectrics and Frequency Control, IEEE Transactions on*, vol. 57, no. 5, pp. 1034 –1038, may 2010. S. Grop, P. Y. Bourgeois, N. Bazin, Y. Kersal[é]{}, E. Rubiola, C. Langham, M. Oxborrow, D. Clapton, S. Walker, J. D. Vicente, and V. Giordano, “[ELISA]{}: A cryocooled 10 [GH]{}z oscillator with $10^{-15}$ frequency stability,” *Rev. Sci. Instrum.*, vol. 81, no. 2, p. 025102, 2010. \[Online\]. Available: <http://link.aip.org/link/?RSI/81/025102/1> S. Grop, V. Giordano, P. Bourgeois, N. Bazin, Y. Kersale, M. Oxborrow, G. Marra, C. Langham, E. Rubiola, and J. DeVincente, “[ELISA]{}: An ultra-stable oscillator for [ESA]{} deep space antennas,” in *Frequency Control Symposium, 2009 Joint with the 22nd European Frequency and Time forum. IEEE International*, april 2009, pp. 376 –380. B. Young, F. Cruz, W. Itano, and J. Bergquist, “Visible lasers with subhertz linewidths,” *Phys. Rev. Lett.*, vol. 82, p. 3799, 1999. J. Millo, D. V. Magalhaes, C. Mandache, Y. L. Coq, E. M. L. English, P. G. Westergaard, J. Lodewyck, S. Bize, P. Lemonde, and G. Santarelli, “Ultrastable lasers based on vibration insensitive cavities,” *Phys. Rev. A*, vol. 79, no. 5, p. 053829, 2009. \[Online\]. Available: <http://link.aps.org/abstract/PRA/v79/e053829> Y. Y. Jiang, A. D. Ludlow, N. D. Lemke, R. W. Fox, J. A. Sherman, L.-S. Ma, and C. W. Oates, “Making optical atomic clocks more stable with $10^{-16}$-level laser stabilization,” *Nat. Photon.*, vol. 5, no. 3, pp. 158–161, 2011. \[Online\]. Available: <http://dx.doi.org/10.1038/nphoton.2010.313> B. Lipphardt, G. Grosche, U. Sterr, C. Tamm, S. Weyers, and H. Schnatz, “The stability of an optical clock laser transferred to the interrogation oscillator for a [C]{}s fountain,” *IEEE Trans. Instrum. Meas.*, 2008. J. Millo, R. Boudot, M. Lours, P. Y. Bourgeois, A. N. Luiten, Y. L. Coq, Y. Kersalé, and G. Santarelli, “Ultra-low-noise microwave extraction from fiber-based optical frequency comb,” *Opt. Lett.*, vol. 34, no. 23, pp. 3707–3709, 2009. \[Online\]. Available: <http://ol.osa.org/abstract.cfm?URI=ol-34-23-3707> J. Millo, M. Abgrall, M. Lours, E. M. L. English, H. Jiang, J. Gu[é]{}na, A. Clairon, M. E. Tobar, S. Bize, Y. L. Coq, and G. Santarelli, “Ultralow noise microwave generation with fiber-based optical frequency comb and application to atomic fountain clock,” *Applied Physics Letters*, vol. 94, no. 14, p. 141105, 2009. \[Online\]. Available: <http://link.aip.org/link/?APL/94/141105/1> S. Weyers, B. Lipphardt, and H. Schnatz, “Reaching the quantum limit in a fountain clock using a microwave oscillator phase locked to an ultrastable laser,” *Phys. Rev. A*, vol. 79, no. 3, p. 031803, 2009. \[Online\]. Available: <http://link.aps.org/abstract/PRA/v79/e031803> J. Gu[é]{}na, G. Dudle, and P. Thomann, “An experimental study of intermodulation effects in an atomic fountain frequency standard,” *The European Physical Journal - Applied Physics*, vol. 38, no. 02, pp. 183–189, 2007. \[Online\]. Available: <http://dx.doi.org/10.1051/epjap:2007072> G. Di Domenico, L. Devenoges, A. Joyet, A. Stefanov, and P. Thomann, “Uncertainty evaluation of the continuous cesium fountain frequency standard focs-2,” in *Frequency Control and the European Frequency and Time Forum (FCS), 2011 Joint Conference of the IEEE International*, may 2011, p. 51. R. Li and K. Gibble, “Phase variations in microwave cavities for atomic clocks,” *Metrologia*, vol. 41, no. 6, pp. 376–386, 2004. \[Online\]. Available: <http://stacks.iop.org/0026-1394/41/376> ——, “Evaluating and minimizing distributed cavity phase errors in atomic clocks,” *Metrologia*, vol. 47, no. 5, p. 534, 2010. \[Online\]. Available: <http://stacks.iop.org/0026-1394/47/i=5/a=004> J. Gu[é]{}na, R. Li, K. Gibble, S. Bize, and A. Clairon, “Evaluation of [D]{}oppler shifts to improve the accuracy of primary atomic fountain clocks,” *Phys. Rev. Lett.*, vol. 106, no. 13, p. 130801, Apr 2011. R. Li, K. Gibble, and K. Szymaniec, “Improved accuracy of the [NPL]{}-[CsF2]{} primary frequency standard: evaluation of distributed cavity phase and microwave lensing frequency shifts,” *Metrologia*, vol. 48, no. 5, p. 283, 2011. \[Online\]. Available: <http://stacks.iop.org/0026-1394/48/i=5/a=007> S. Weyers, V. Gerginov, N. Nemitz, R. Li, and K. Gibble, “Distributed cavity phase frequency shifts of the caesium fountain ptb-csf2,” *Metrologia*, vol. 49, no. 1, p. 82, 2012. \[Online\]. Available: <http://stacks.iop.org/0026-1394/49/i=1/a=012> J. Gu[é]{}na, S. Bize, A. Clairon, R. Li, and K. Gibble, “Quantitative evaluation of distributed cavity phase shifts to improve the accuracy of [SYRTE]{} [FO2]{},” in *Frequency Control and the European Frequency and Time Forum (FCS), 2011 Joint Conference of the IEEE International*, may 2011, p. 61. P. Wolf and C. J. Bord[é]{}, “Recoil effects in microwave ramsey spectroscopy,” *arXiv:quant-ph/0403194v1*, 2004. J. L. Hall, C. J. Bordé, and K. Uehara, “Direct optical resolution of the recoil effect using saturated absorption spectroscopy,” *Phys. Rev. Lett.*, vol. 37, pp. 1339–1342, 1976. K. Gibble, “Difference between a photon’s momentum and an atom’s recoil,” *Phys. Rev. Lett.*, vol. 97, no. 7, p. 073002, 2006. \[Online\]. Available: <http://link.aps.org/abstract/PRL/v97/e073002> W. Itano, L. Lewis, and D. Wineland, “Shift of $^{2}$[S]{}$_{1/2}$ hyperfine splittings due to blackbody radiation,” *Phys. Rev. A.*, vol. 25, p. 1233, 1982. E. Simon, P. Laurent, and A. Clairon, “Measurement of the [Stark]{} shift of the [Cs]{} hyperfine splitting in an atomic fountain,” *Phys. Rev. A.*, vol. 57, p. 436, 1998. S. Micalizio, A. Godone, D. Calonico, F. Levi, and L. Lorini, “Blackbody radiation shift of the $^{133}$[C]{}s hyperfine transition frequency,” *Phys. Rev. A*, vol. 69, no. 5, p. 053401, May 2004. F. Levi, D. Calonico, L. Lorini, S. Micalizio, and A. Godone, “Measurement of the blackbody radiation shift of the $^{133}$cs hyperfine transition in an atomic fountain,” *Phys. Rev. A*, vol. 70, no. 3, p. 033412, 2004. A. Godone, D. Calonico, F. Levi, S. Micalizio, and C. Calosso, “[S]{}tark-shift measurement of the $^2$[S]{}$_{1/2} $, ${F}=3\rightarrow {F}=4 $ hyperfine transition of $^{133}$[C]{}s,” *Phys. Rev. A*, vol. 71, no. 6, p. 063401, 2005. P. Rosenbusch, S. Zhang, and A. Clairon, “Blackbody radiation shift in primary frequency standards,” in *Frequency Control Symposium, 2007 Joint with the 21st European Frequency and Time Forum. IEEE International*, 29 2007-june 1 2007, pp. 1060 –1063. S. Ulzega, A. Hofer, P. Moroshkin, and A. Weis, “Reconciliation of experimental and theoretical electric tensor polarizabilities of the cesium ground state,” *Eur. Phys. Lett.*, vol. 76, no. 6, p. 1074, 2006. \[Online\]. Available: <http://stacks.iop.org/0295-5075/76/i=6/a=1074> P. G. H. Sandars, “Differential polarizability in the ground state of the hydrogen atom,” *Proceedings of the Physical Society*, vol. 92, no. 4, p. 857, 1967. \[Online\]. Available: <http://stacks.iop.org/0370-1328/92/i=4/a=303> J. R. P. Angel and P. G. H. Sandars, “The hyperfine structure [S]{}tark effect. [I]{}. theory,” *Proceedings of the Royal Society of London. Series A. Mathematical and Physical Sciences*, vol. 305, p. 125, 1968. \[Online\]. Available: <http://rspa.royalsocietypublishing.org/content/305/1480/125.abstract> S. Ulzega, A. Hofer, P. Moroshkin, R. Müller-Siebert, D. Nettels, and A. Weis, “Measurement of the forbidden electric tensor polarizability of [C]{}s atoms trapped in solid $^4$[H]{}e,” *Phys. Rev. A*, vol. 75, no. 4, p. 042505, 2007. E. J. Angstmann, V. A. Dzuba, and V. V. Flambaum, “Frequency shift of the cesium clock transition due to blackbody radiation,” *Phys. Rev. Lett.*, vol. 97, p. 040802, Jul 2006. \[Online\]. Available: <http://link.aps.org/doi/10.1103/PhysRevLett.97.040802> K. Beloy, U. I. Safronova, and A. Derevianko, “High-accuracy calculation of the blackbody radiation shift in the $^{133}$[Cs]{} primary frequency standard,” *Phys. Rev. Lett.*, vol. 97, p. 040801, 2006. V. G. Pal’chikov, Y. S. Domnin, and A. V. Novoselov, “Black-body radiation effects and light shifts in atomic frequency standards,” *Journal of Optics B: Quantum and Semiclassical Optics*, vol. 5, no. 2, p. S131, 2003. \[Online\]. Available: <http://stacks.iop.org/1464-4266/5/i=2/a=370> J. D. Feichtner, M. E. Hoover, and M. Mizushima, “Stark effect of the hyperfine structure of cesium-133,” *Phys. Rev.*, vol. 137, no. 3A, pp. A702–A708, Feb 1965. T. Lee, T. P. Das, and R. M. Sternheimer, “Perturbation theory for the [S]{}tark effect in the hyperfine structure of alkali-metal atoms,” *Phys. Rev. A*, vol. 11, no. 6, pp. 1784–1786, Jun 1975. J.-L. Robyr, P. Knowles, and A. Weis, “[CPT]{}-pump-probe measurement of the [Cs]{} clock transition [DC]{} [S]{}tark shift,” in *Frequency Control and the European Frequency and Time Forum (FCS), 2011 Joint Conference of the IEEE International*, may 2011, pp. 1 –4. S. Zhang, “D[é]{}placement de fr[é]{}quence d[û]{} au rayonnement du corps noir dans une fontaine atomique [à]{} c[é]{}sium et am[é]{}lioration des performances de l’horloge,” Ph.D. dissertation, department of physics, Universit[é]{} Pierre et Marie Curie, Paris VI, 2004. T. P. Heavner, T. E. Parker, J. H. Shirley, L. Donley, S. R. Jefferts, F. Levi, D. Calonico, C. Calosso, G. Costanzo, and B. Mongino, “Comparing room temperature and cryogenic cesium fountains,” in *Frequency Control and the European Frequency and Time Forum (FCS), 2011 Joint Conference of the IEEE International*, may 2011, pp. 1 –3. S. Bize, Y. Sortais, M. Santos, C. Mandache, A. Clairon, and C. Salomon, “High-accuracy measurement of the $^{87}$[Rb]{} ground-state hyperfine splitting in an atomic fountain,” *Europhys. Lett.*, vol. 45, no. 5, p. 558, 1999. E. J. Angstmann, V. A. Dzuba, and V. V. Flambaum, “Frequency shift of hyperfine transitions due to blackbody radiation,” *Phys. Rev. A*, vol. 74, p. 023405, 2006. M. S. Safronova, D. Jiang, and U. I. Safronova, “Blackbody radiation shift in the $^{87}$[R]{}b frequency standard,” *Phys. Rev. A*, vol. 82, no. 2, p. 022510, 2010. R. S. S. Weyers and R. Wynands, “Effects of microwave leakage in caesium clocks: theoretical and experimental results,” *Proceedings of the 2006 EFTF*, 2006. S. Jefferts, J. Shirley, N. Ashby, T. Heavner, E. Donley, and F. Levi, “On the power dependence of extraneous microwave fields in atomic frequency standards,” in *Frequency Control Symposium and Exposition, 2005. Proceedings of the 2005 IEEE International*, August 2005, p. 6. S. Bize, P. Laurent, M. Abgrall, H. Marion, I. Maksimovic, L. Cacciapuoti, J. Grünert, C. Vian, F. [Pereira dos Santos]{}, P. Rosenbusch, P. Lemonde, G. Santarelli, P. Wolf, A. Clairon, A. Luiten, M. Tobar, and C. Salomon, “Advances in $^{133}$[C]{}s fountains,” *C. R. Physique*, vol. 5, p. 829, 2004. ——, “Cold atom clocks and applications,” *J. Phys. B: Atomic, Molecular and Optical Physics*, vol. 38, no. 9, pp. S449–S468, 2005. \[Online\]. Available: <http://stacks.iop.org/0953-4075/38/S449> C. Fertig and K. Gibble, “Measurement and cancellation of the cold collision frequency shift in an $^{87}$[R]{}b fountain clock,” *Phys. Rev. Lett.*, vol. 85, p. 1622, 2000. Y. Sortais, S. Bize, C. Nicolas, A. Clairon, C. Salomon, and C. Williams, “Cold collision frequency shifts in a $^{87}$[Rb]{} fountain,” *Phys. Rev. Lett.*, vol. 85, p. 3117, 2000. S. Bize, Y. Sortais, C. Mandache, A. Clairon, and C. Salomon, “Cavity frequency pulling in cold atom fountains,” *IEEE Trans. on Instr. and Meas.*, vol. 50, p. 503, 2001. K. Szymaniec, W. Chałupczak, E. Tiesinga, C. J. Williams, S. Weyers, and R. Wynands, “Cancellation of the collisional frequency shift in caesium fountain clocks,” *Phys. Rev. Lett.*, vol. 98, no. 15, p. 153002, 2007. K. Szymaniec and S. E. Park, “Primary frequency standard [NPL]{}-[CsF2]{}: Optimized operation near the collisional shift cancellation point,” *Instrumentation and Measurement, IEEE Transactions on*, vol. 60, no. 7, pp. 2475 –2481, 2011. Prior to 2011, the DCP correction was $(0.0\pm 3.2)\times 10^{-16}$ for FO1, $(0.0\pm 3.0)\times 10^{-16}$ for FO2-Cs, $(0.0\pm 2.5)\times 10^{-16}$ for FO2-Rb. For FOM, the status of theory and experiment was such that the DCP effects and the effects related to spectral purity and leakage could not be disentangled, yielding a combined correction of $(0 \pm 6.0)\times 10^{-16}$. The microwave lensing correction was $(0.0\pm 1.4)\times 10^{-16}$ for FO1, FO2-Cs and FOM, and $(0.0\pm 1.2)\times 10^{-16}$ for FO2-Rb. *Numerical Recipes in C, Second Edition*, 1992, p. 663. \[Online\]. Available: <http://apps.nrbook.com/c/index.html> T. E. Parker, “Long-term comparison of caesium fountain primary frequency standards,” *Metrologia*, vol. 47, p. 1, 2010. \[Online\]. Available: <http://stacks.iop.org/0026-1394/47/i=1/a=001> H. Marion, F. [Pereira Dos Santos]{}, M. Abgrall, S. Zhang, Y. Sortais, S. Bize, I. Maksimovic, D. Calonico, J. Grünert, C. Mandache, P. Lemonde, G. Santarelli, P. Laurent, A. Clairon, and C. Salomon, “Search for variations of fundamental constants using atomic fountain clocks,” *Phys. Rev. Lett.*, vol. 90, p. 150801, 2003. BIPM-CCTF 2004, “Recommendation CCTF-1 2004 concerning secondary representations of the second”, p. 38 (2004). Y. Ovchinnikov and G. Marra, “Accurate rubidium atomic fountain frequency standard,” *Metrologia*, vol. 48, no. 3, p. 87, 2011. \[Online\]. Available: <http://stacks.iop.org/0026-1394/48/i=3/a=003> R. Li and K. Gibble, “Comment on ’accurate rubidium atomic fountain frequency standard’,” *Metrologia*, vol. 48, no. 5, p. 446, 2011. \[Online\]. Available: <http://stacks.iop.org/0026-1394/48/i=5/a=N03> Y. Ovchinnikov and G. Marra, “Reply to comment on ’accurate rubidium atomic fountain frequency standard’,” *Metrologia*, vol. 48, no. 5, p. 448, 2011. \[Online\]. Available: <http://stacks.iop.org/0026-1394/48/i=5/a=N04> S. Peil, S. Crane, T. Swanson, and C. R. Ekstrom, “The [USNO]{} rubidium fountain,” in *International Frequency Control Symposium and Exposition, 2006 IEEE*, june 2006, pp. 304 –306. S. Peil, S. Crane, J. Hanssen, T. B. Swanson, and C. R. Ekstrom, “Measurements with multiple operational fountain clocks,” in *Frequency Control and the European Frequency and Time Forum (FCS), 2011 Joint Conference of the IEEE International*, may 2011, p. 58. See *Circular T* and fountain reports on the BIPM website. \[Online\]. Available: [www.bipm.org/jsp/en/TimeFtp.jsp](www.bipm.org/jsp/en/TimeFtp.jsp) P. Uhrich, D. Valat, and M. Abgrall, “Steering of the french time scale [TA(F)]{} towards the [LNE]{}-[SYRTE]{} primary frequency standards,” *Metrologia*, vol. 45, no. 6, p. S42, 2008. \[Online\]. Available: <http://stacks.iop.org/0026-1394/45/i=6/a=S07> C. W. Chou, D. B. Hume, J. C. J. Koelemeij, D. J. Wineland, and T. Rosenband, “Frequency comparison of two high-accuracy [A]{}l$^{+}$ optical clocks,” *Phys. Rev. Lett.*, vol. 104, no. 7, p. 070802, Feb 2010. N. D. Lemke, A. D. Ludlow, Z. W. Barber, T. M. Fortier, S. A. Diddams, Y. Jiang, S. R. Jefferts, T. P. Heavner, T. E. Parker, and C. W. Oates, “Spin-1/2 optical lattice clock,” *Phys. Rev. Lett.*, vol. 103, no. 6, p. 063001, 2009. \[Online\]. Available: <http://link.aps.org/abstract/PRL/v103/e063001> T. Rosenband, D. B. Hume, P. O. Schmidt, C. W. Chou, A. Brusch, L. Lorini, W. H. Oskay, R. E. Drullinger, T. M. Fortier, J. E. Stalnaker, S. A. Diddams, W. C. Swann, N. R. Newbury, W. M. Itano, D. J. Wineland, and J. C. Bergquist, “Frequency ratio of [A]{}l$^+$ and [H]{}g$^+$ single-ion optical clocks; metrology at the 17th decimal place,” *Science*, vol. 319, p. 1808, 2008. A. D. Ludlow, T. Zelevinsky, G. K. Campbell, S. Blatt, M. M. Boyd, M. H. G. [de Miranda]{}, M. J. Martin, J. W. Thomsen, S. M. Foreman, J. Ye, T. M. Fortier, J. E. Stalnaker, S. A. Diddams, Y. L. Coq, Z. W. Barber, N. Poli, N. D. Lemke, K. M. Beck, and C. W. Oates, “Sr lattice clock at $1\times 10^{-16}$ fractional uncertainty by remote optical evaluation with a [C]{}a clock,” *Science*, vol. 319, p. 1805, 2008. Consultative Committee for Time and Frequency (CCTF), Report of the 18th meeting (June 2009) to the International Committee for Weights and Measures. X. Baillard, M. Fouché, R. L. Targat, P. G. Westergaard, A. Lecallier, F. Chapelet, M. Abgrall, G. D. Rovera, P. Laurent, P. Rosenbusch, S. Bize, G. Santarelli, A. Clairon, P. Lemonde, G. Grosche, B. Lipphardt, and H. Schnatz, “An optical lattice clock with spin-polarized $^{87}$[S]{}r atoms,” *Eur. Phys. J. D*, vol. 48, p. 11, 2008. S. Blatt, A. D. Ludlow, G. K. Campbell, J. W. Thomsen, T. Zelevinsky, M. M. Boyd, J. Ye, X. Baillard, M. Fouché, R. L. Targat, A. Brusch, P. Lemonde, M. Takamoto, F.-L. Hong, H. Katori, and V. V. Flambaum, “New limits on coupling of fundamental constants to gravity using $^{87}$[S]{}r optical lattice clocks,” *Phys. Rev. Lett.*, vol. 100, p. 140801, 2008. M. Fischer, N. Kolachevsky, M. Zimmermann, R. Holzwarth, T. Udem, T. W. Hänsch, M. Abgrall, J. Grunert, I. Maksimovic, S. Bize, H. Marion, F. P. D. Santos, P. Lemonde, G. Santarelli, P. Laurent, A. Clairon, C. Salomon, M. Haas, U. D. Jentschura, and C. H. Keitel, “New limits on the drift of fundamental constants from laboratory measurements,” *Phys. Rev. Lett.*, vol. 92, p. 230802, 2004. T. M. Fortier, N. Ashby, J. C. Bergquist, M. J. Delaney, S. A. Diddams, T. P. Heavner, L. Hollberg, W. M. Itano, S. R. Jefferts, K. Kim, F. Levi, L. Lorini, W. H. Oskay, T. E. Parker, J. Shirley, and J. E. Stalnaker, “Precision atomic spectroscopy for improved limits on variation of the fine structure constant and local position invariance,” *Phys. Rev. Lett.*, vol. 98, p. 070801, 2007. E. Peik, B. Lipphardt, H. Schnatz, T. Schneider, C. Tamm, and S. G. Karshenboïm, “Limit on the present temporal variation of the fine structure constant,” *Phys. Rev. Lett.*, vol. 93, p. 170801, 2004. P. J. Mohr, B. N. Taylor, and D. B. Newell, “[CODATA]{} recommended values of the fundamental physical constants: 2006,” *Rev. Mod. Phys.*, vol. 80, no. 2, p. 633, 2008. \[Online\]. Available: <http://link.aps.org/abstract/RMP/v80/p633> S. G. Karshenboïm, “Precision physics of simple atoms: [QED]{} tests, nuclear structure and fundamental constants,” *Physics Reports*, vol. 422, no. 1-2, pp. 1 – 63, 2005. \[Online\]. Available: <http://www.sciencedirect.com/science/article/pii/S0370157305003637> M. Niering, R. Holzwarth, J. Reichert, P. Pokasov, T. Udem, M. Weitz, T. W. Hänsch, P. Lemonde, G. Santarelli, M. Abgrall, P. Laurent, C. Salomon, and A. Clairon, “Measurement of the hydrogen 1[S]{}-2[S]{} transition frequency by phase coherent comparison with a microwave cesium fountain clock,” *Phys. Rev. Lett.*, vol. 84, p. 5496, 2000. C. G. Parthey, A. Matveev, J. Alnis, B. Bernhardt, A. Beyer, R. Holzwarth, A. Maistrou, R. Pohl, K. Predehl, T. Udem, T. Wilken, N. Kolachevsky, M. Abgrall, D. Rovera, C. Salomon, P. Laurent, and T. W. Hänsch, “Improved measurement of the hydrogen [1S-2S]{} transition frequency,” *Phys. Rev. Lett.*, vol. 107, p. 203001, Nov 2011. \[Online\]. Available: <http://link.aps.org/doi/10.1103/PhysRevLett.107.203001> R. Pohl, A. Antognini, F. Nez, F. D. Amaro, F. Biraben, J. M. R. Cardoso, D. S. Covita, A. Dax, S. Dhawan, L. M. P. Fernandes, A. Giesen, T. Graf, T. W. H[ä]{}nsch, P. Indelicato, L. Julien, C.-Y. Kao, P. Knowles, E.-O. Le Bigot, Y.-W. Liu, J. A. M. Lopes, L. Ludhova, C. M. B. Monteiro, F. Mulhauser, T. Nebel, P. Rabinowitz, J. M. F. dos Santos, L. A. Schaller, K. Schuhmann, C. Schwob, D. Taqqu, J. F. C. A. Veloso, and F. Kottmann, “The size of the proton,” *Nature*, vol. 466, no. 7303, pp. 213–216, Jul. 2010. \[Online\]. Available: <http://dx.doi.org/10.1038/nature09250> , “Confinement of antihydrogen for 1,000 seconds,” *Nature Physics*, vol. 7, p. 558, 2011. M. Chwalla, J. Benhelm, K. Kim, G. Kirchmair, T. Monz, M. Riebe, P. Schindler, A. S. Villar, W. Hansel, C. F. Roos, R. Blatt, M. Abgrall, G. Santarelli, G. D. Rovera, and P. Laurent, “Absolute frequency measurement of the $^{40}$[C]{}a$^{+}$ 4s $^{2}$[S]{}$_{1/2}$-3d $^{2}$[D]{}$_{5/2}$ clock transition,” *Phys. Rev. Lett.*, vol. 102, no. 2, p. 023002, 2009. \[Online\]. Available: <http://link.aps.org/abstract/PRL/v102/e023002> X. Baillard, M. Fouché, R. L. Targat, P. G. Westergaard, A. Lecallier, Y. L. Coq, G. D. Rovera, S. Bize, and P. Lemonde, “Accuracy evaluation of an optical lattice clock with bosonic atoms,” *Opt. Lett.*, vol. 32, p. 1812, 2007. R. Le Targat, X. Baillard, M. Fouché, A. Brusch, O. Tcherbakoff, G. D. Rovera, and P. Lemonde, “Accurate optical lattice clock with $^{87}$[S]{}r atoms,” *Phys. Rev. Lett.*, vol. 97, no. 13, p. 130801, 2006. I. Courtillot, A. Quessada, R. Kovacich, A. Brusch, D. Kolker, J.-J. Zondy, G. Rovera, and P. Lemonde, “Clock transition for a future optical frequency standard with trapped atoms,” *Phys. Rev. A*, vol. 68, p. 030501, 2003. M. Petersen, R. Chicireanu, S. T. Dawkins, D. V. Magalhaes, C. Mandache, Y. L. Coq, A. Clairon, and S. Bize, “Doppler-free spectroscopy of the $^1$[S]{}$_0$-$^3$[P]{}$_0$ optical clock transition in laser-cooled fermionic isotopes of neutral mercury,” *Phys. Rev. Lett.*, vol. 101, no. 18, p. 183004, 2008. \[Online\]. Available: <http://link.aps.org/abstract/PRL/v101/e183004> L. Yi, S. Mejri, J. J. McFerran, Y. Le Coq, and S. Bize, “Optical lattice trapping of $^{199}$[H]{}g and determination of the magic wavelength for the ultraviolet $^{1}{S}_{0}\leftrightarrow{}^{3}{P}_{0}$ clock transition,” *Phys. Rev. Lett.*, vol. 106, no. 7, p. 073005, 2011. P. Laurent, M. Abgrall, C. Jentsch, P. Lemonde, G. Santarelli, A. Clairon, I. Maksimovic, S. Bize, C. Salomon, D. Blonde, J. Vega, O. Grosjean, F. Picard, M. Saccoccio, M. Chaubet, N. Ladiette, L. Guillet, I. Zenone, C. Delaroche, and C. Sirmain, “Design of the cold atom pharao space clock and initial test results,” *Applied Physics B: Lasers and Optics*, vol. 84, pp. 683–690, 2006, 10.1007/s00340-006-2396-6. \[Online\]. Available: <http://dx.doi.org/10.1007/s00340-006-2396-6> L. Cacciapuoti, N. Dimarcq, G. Santarelli, P. Laurent, P. Lemonde, A. Clairon, P. Berthoud, A. Jornod, F. Reina, S. Feltham, and C. Salomon, “Atomic clock ensemble in space: Scientific objectives and mission status,” *Nuclear Physics B - Proceedings Supplements*, vol. 166, pp. 303 – 306, 2007, proceedings of the Third International Conference on Particle and Fundamental Physics in Space, Proceedings of the Third International Conference on Particle and Fundamental Physics in Space. \[Online\]. Available: <http://www.sciencedirect.com/science/article/pii/S0920563206010425> P. Guillemot, E. Samain, P. Vrancken, P. Exertier, and S. Leon, “Time transfer by laser link - [T2L2]{}: An opportunity to calibrate [RF]{} links,” in *Proceedings of the 40th PTTI meeting*, 2008, pp. 95–106. E. Samain, P. Exertier, P. Guillemot, P. Laurent, F. Pierron, D. Rovera, J.-M. Torre, M. Abgrall, J. Achkar, D. Albanese, C. Courde, K. Djeroud, M. L. Bourez, S. Leon, H. Mariey, G. Martinot-Lagarde, J. L. Oneto, J. Paris, M. Pierron, and H. Viot, “Time transfer by laser link - [T2L2]{}: Current status and future experiments,” in *Frequency Control and the European Frequency and Time Forum (FCS), 2011 Joint Conference of the IEEE International*, May 2011, pp. 1–6. G. D. Rovera, M. Abgrall, and P. Laurent, “Contribution to [TAI]{} frequency by a travelling primary frequency standard,” in *Frequency Control and the European Frequency and Time Forum (FCS), 2011 Joint Conference of the IEEE International*, may 2011. S. G. Karshenboïm, “Some possibilities for laboratory searches for variations of fundamental constants,” *Can. J. Phys.*, vol. 47, p. 639, 2000. V. V. Flambaum, “Variation of fundamental constants in space and time: Theory and observations,” *Eur. Phys. J. Special Topics*, vol. 163, pp. 159–171, oct 2008. \[Online\]. Available: <http://dx.doi.org/10.1140/epjst/e2008-00817-5> J. Uzan, “The fundamental constants and their variation: observational status and theoretical motivations,” *Rev. Mod. Phys*, vol. 75, p. 403, 2003. J.-P. Uzan, “Varying constants, gravitation and cosmology,” *Living Reviews in Relativity*, vol. 14, no. 2, 2011. \[Online\]. Available: <http://www.livingreviews.org/lrr-2011-2> J. Gu[é]{}na, F. Chapelet, P. Rosenbusch, P. Laurent, M. Abgrall, G. Rovera, G. Santarelli, S. Bize, A. Clairon, and M. Tobar, “New measurement of the rubidium hyperfine frequency using [LNE]{}-[SYRTE]{} fountain ensemble,” in *Frequency Control Symposium, 2008 IEEE International*, may 2008, pp. 366 –370. V. V. Flambaum, D. B. Leinweber, A. W. Thomas, and R. D. Young, “Limits on variations of the quark masses, qcd scale, and fine structure constant,” *Phys. Rev. D*, vol. 69, no. 11, p. 115006, Jun 2004. V. V. Flambaum and A. F. Tedesco, “Dependence of nuclear magnetic moments on quark masses and limits on temporal variation of fundamental constants from atomic clock experiments,” *Phys. Rev. C*, vol. 73, no. 5, p. 055501, 2006. \[Online\]. Available: <http://link.aps.org/abstract/PRC/v73/e055501> T. Parker and J. Levine, “Impact of new high stability frequency standards on the performance of the nist at1 time scale,” *Ultrasonics, Ferroelectrics and Frequency Control, IEEE Transactions on*, vol. 44, no. 6, pp. 1239 –1244, nov 1997. N. Ashby, T. P. Heavner, S. R. Jefferts, T. E. Parker, A. G. Radnaev, and Y. O. Dudin, “Testing [L]{}ocal [P]{}osition [I]{}nvariance with four cesium-fountain primary frequency standards and four [NIST]{} hydrogen masers,” *Phys. Rev. Lett.*, vol. 98, p. 070802, 2007. \[Online\]. Available: <http://link.aps.org/abstract/PRL/v98/e070802> C. M. Will, “The confrontation between general relativity and experiment,” *Living Rev. Relativity*, vol. 9, p. 3, 2006. \[Online\]. Available: <http://www.livingreviews.org/lrr-2006-3> P. Wolf, F. Chapelet, S. Bize, and A. Clairon, “Cold atom clock test of [L]{}orentz [I]{}nvariance in the matter sector,” *Phys. Rev. Lett.*, vol. 96, no. 6, p. 060801, 2006. \[Online\]. Available: <http://link.aps.org/abstract/PRL/v96/e060801> M. E. Tobar, P. Wolf, S. Bize, G. Santarelli, and V. Flambaum, “Testing local lorentz and position invariance and variation of fundamental constants by searching the derivative of the comparison frequency between a cryogenic sapphire oscillator and hydrogen maser,” *Phys. Rev. D*, vol. 81, no. 2, p. 022003, Jan 2010. P. Wolf, S. Bize, A. Clairon, G. Santarelli, M. E. Tobar, and A. N. Luiten, “Improved test of lorentz invariance in electrodynamics,” *Phys. Rev. D*, vol. 70, no. 5, p. 051902, Sep 2004. P. G. Westergaard, J. Lodewyck, L. Lorini, A. Lecallier, E. A. Burt, M. Zawada, J. Millo, and P. Lemonde, “Lattice-induced frequency shifts in [S]{}r optical lattice clocks at the $10^{-17}$ level,” *Phys. Rev. Lett.*, vol. 106, no. 21, p. 210801, May 2011. [^1]: Manuscript received September 15, 2011; accepted January 23, 2012. [^2]: This work was supported by SYstèmes de Référence Temps-Espace (SYRTE), the Laboratoire National de Métrologie et d’Essais (LNE), the Centre National de la Recherche Scientifique (CNRS), Université Pierre et Marie curie (UPMC), the Observatoire de Paris, the Institut Francilien de Recherche sur les Atomes Froids (IFRAF), the National Science Foundation (NSF), Penn state, the Ville de Paris, the European Space Agency (ESA), and the Centre National d’Etudes Spatiales (CNES). [^3]: J. Guéna, M. Abgrall, D. Rovera, P. Laurent, B. Chupin, M. Lours, G. Santarelli, P. Rosenbusch, A. Clairon and S. Bize are affiliated to: LNE-SYRTE, Observatoire de Paris, CNRS, UPMC, 61 avenue de l’Observatoire, 75014 Paris, France [^4]: M. E. Tobar is affiliated to: The School of Physics, University of Western Australia, Crawley, Western Australia [^5]: R. Li and K. Gibble are affiliated to: The Department of Physics, The Pennsylvania State University, University Park, Pennsylvania 16802, USA [^6]: This work was supported by SYRTE, LNE, CNRS, UPMC, Observatoire de Paris, IFRAF, NSF, Penn State, Ville de Paris, ESA and CNES [^7]: DOI: http://dx.doi.org/10.1109/TUFFC.2012.2208 [^8]: PHARAO stands for “Projet d’Horloge Atomique par Refroidissement d’Atomes en Orbite” [^9]: In this configuration, the 3 pairs of counter-propagating laser beams for the molasses are aligned along the axes of a 3 dimensional orthonormal basis, where the (111) direction is along the vertical direction, as sketched in Fig. \[fig\_FontaineRbCs\] [^10]: with a computer controlled Direct Digital Synthesizer [^11]: The goodness-of-fit $Q$ is the probability that a value of chi-square as poor as the value found with the data should occur by chance. Quoting [@NR1992]: “If $Q$ is larger than, say, 0.1, then the goodness-of-fit is believable". [^12]: For instance, the gravitational potential created by a point mass $m$ at a distance $r$ is: $U(r)=G m/r$ where $G$ is the Newtonian constant of gravitation.

ArXiv

--- abstract: | We consider the lowest order radiative corrections for the decay $K^{\pm}\to \pi^0 e^{\pm} \nu$, usually referred as $K_{e3}^{\pm}$ decay. This decay is the best way to extract the value of the $V_{us}$ element of the CKM matrix. The radiative corrections become crucial if one wants a precise value of $V_{us}$. The existing calculations were performed in the late 60’s [@B; @G] and are in disagreement. The calculation in [@G] turns out to be ultraviolet cutoff sensitive. The necessity of precise knowledge of $V_{us}$ and the contradiction between the existing results constitute the motivation of our paper. We remove the ultraviolet cutoff dependence by using A.Sirlin’s prescription; we set it equal to the $W$ mass. We establish the whole character of small lepton mass dependence based on the renormalization group approach. In this way we can provide a simple explanation of Kinoshita–Lee–Nauenberg cancellation of singularities in the lepton mass terms in the total width and pion spectrum. We give an explicit evaluation of the structure–dependent photon emission based on ChPT in the lowest order. We estimate the accuracy of our results to be at the level of $1\%$. The corrected total width is $\Gamma=\Gamma_0(1+\delta)$ with $\delta=0.02\pm0.0002$. Using the formfactor value $f_+(0)=0.9842\pm 0.0084$ calculated in [@CKNRT] leads to $|V_{us}|=0.2172 \pm0.0055$. author: - | V. Bytev$^1$, E. Kuraev$^1$,\ A. Baratt$^2$, J. Thompson$^2$ title: ' Radiative Corrections to the $K_{e3}^{\pm}$ Decay Revised ' --- $^1$ Bogoliubov Laboratory of Theoretical Physics, JINR, Dubna, Moscow region, 141980, Russia\ $^2$ Department of Physics and Astronomy, University of Pittsburgh, PA, U.S.A\ Motivation ========== [virt1]{} (100,80) (100,80) (100,80) \ [virt2]{} (100,80) (100,80) (100,80) \ [second]{} (100,80) (100,80) (100,80) \ For corrections due to virtual photons see figure \[fig:virt\], for corrections due to real photons see figure \[fig:real\]. The $K_{e3}$ decay is important since it is the cleanest way to measure the $V_{us}$ matrix element of the CKM matrix. If one uses the current values for $V_{ud}$, $V_{us}$, and $V_{ub}$ taken from the PDG then $|V_{ud}|^2+|V_{us}|^2+|V_{ub}|^2$ misses unity by $2.2$ standard deviations which contradicts the unitarity of the CKM matrix and might indicate physics beyond the Standard Model. The uncertainty brought to the above expression by $V_{us}$ is about the same as uncertainty that comes from $V_{ud}$. Therefore reducing the error in the $V_{us}$ matrix element would reduce substantially the error in the whole unitarity equation. Reliable radiative corrections, potentially of the order of a few percent are necessary to extract the $V_{us}$ matrix element from the $K_{e3}$ decay width with high precision. Calculations of the radiative corrections to the $K_{e3}$ decay were performed independently by E.S.Ginsberg and T.Becherrawy in the late 60’s [@G; @B]. Their results for corrections to the decay rate, Dalitz plot, pion and positron spectra disagree, in some places quite sharply; for example Ginsberg’s correction to the decay rate is $-0.45\%$ while that of Becherrawy is $-2\%$ (corresponding to corrections to the total width $\Gamma$ of 0.45 and 2 respectively). We have decided to perform a new calculation since results of the experiments will become available soon and to explore the causes of the discrepancies in the previous calculations. Recently a revision of E. Ginsberg’s paper, with numerical estimation of the radiation corrections [@CKNRT] was published. Our paper is organized as follows. The introduction (Section 2) is devoted to the short review of kinematics of the elastic decay process (without emission of real photon). In Section 3 we put the results concerning the virtual and soft real photons’ emission contribution to the differential width. In Section 4 we consider the hard photon emission including both the inner bremsstrahlung (IB) and the structure-dependent (SD) contributions and derive an expression for the differential width by starting with the Born width and adding the known structure functions in the leading logarithmical approximation (the so-called Drell-Yan picture of the process). We give the explicit expressions for the non-leading contributions. In Section 5, we summarize our results and compare them with those in the previously published papers. Appendix A contains the details of calculations of virtual and real soft photons emission. Appendix B contains the details of description of hard photon emission both by IB and SD mechanisms. Our approach to study the hard photon emission differs technically from the ones used in papers [@B; @G]. Appendix C contains the explicit formulae for description of SD emission including the interference of IB and SD amplitudes. Appendix D is devoted to analysis of Dalitz-plot distribution and the properties of Drell-Yan conversion mentioned above. Appendix E contains the list of the formulae used for the numerical integration. Appendix F contains the details of kinematics of radiative kaon decay and,besides the analysis of relations of our and paper [@G] technical approaches. In tables 1,2 and graphics (fig. (3,4,5,6)) the result of numerical estimation of Born values and the correction to Dalitz-plot distribution and pion and positron spectra are given. Introduction ============ The lowest order perturbation theory (PT) matrix element of the process $K^+(p) \to \pi^0(p')+e^+(p_e)+\nu(p_\nu)$ has the form $$M=\frac{G_F}{\sqrt{2}}V^*_{us} F_\nu(t)\bar{u}(p_\nu)\gamma_\nu(1+\gamma_5) v(p_e)$$ where $F_\nu(t)=\frac{1}{\sqrt{2}}(p+p')_{\nu}f_+(t)$. Dalitz plot density which takes into account the radiative corrections (RC) of the lowest order PT is $$\begin{aligned} \frac{d^2\Gamma}{dydz}=\mathcal{C} a_0(y,z)(1+\delta(y,z)) \left(1+\lambda_+\frac{t}{m_\pi^2} \right)^2 = \frac{d^2 \Gamma_0(y,z)} {dy dz}(1+\delta(y,z)), \\ \nonumber \Gamma_0=\int\frac{d^2\Gamma_0}{d y d z}d yd z, \quad \Gamma=\int\frac{d^2\Gamma}{d y d z}d yd z,\end{aligned}$$ where the momentum transfer squared between kaon and pion is: $$t=(p-p')^2=M_K^2(1+r_\pi-z)=M_K^2R(z) \ .$$ We accept here the following form for the strong interactions induced form factor $f_+(t)$: $$f_+(t)=f_+(0) \left(1+\lambda_+ \frac{t}{m_\pi^2} \right) \ ,$$ according to PDG $\lambda_+=0.0276 \pm 0.0021$. From now on we’ll use $M^2$ instead of $M_K^2$. We define $$\mathcal{C}=\frac{M^5 G^2_F|V_{us}|^2}{64\pi^3}|f_+(0)|^2 \ ,$$ and $$a_0(y,z)=(z+y-1)(1-y)-r_{\pi} + O(r_e) \ .$$ Here we follow the notation of [@BCEG]: $$r_e \equiv m_e^2/M^2, \hspace{0.2cm} r_\pi \equiv m_\pi^2/M^2 \ ;$$ where $m_e$, $m_\pi$, and $M_K$ are the masses of electron, pion, and kaon; two convenient kinematical variables are $$y \equiv 2pp_e/M^2, \hspace{0.2cm} z \equiv 2pp'/M^2 \ .$$ In the kaon’s rest frame, which we’ll imply throughout the paper, $y$ and $z$ are the energy fractions of the positron and pion: $$y=2E_e/M, \hspace{0.2cm} z=2E_\pi/M \ .$$ The region of $y,z$-plane where $a_0(y,z)>0$ we will named as a region $D$. Later, when dealing with real photons we’ll also use $$x=2\omega/M \ .$$ with $\omega$-real photon energy. The physical region for $y$ and $z$ (further called $D$ -region) is [@BCEG] $$\begin{aligned} 2 \sqrt{r_e} \le y \le 1+r_e-r_\pi \ , \nonumber \\ F_1(y)-F_2(y) \le z \le F_1(y)+F_2(y) \ , \nonumber \\ F_1(y)=(2-y)(1+r_e+r_\pi-y)/\left[2(1+r_e-y)\right] \ , \nonumber \\ F_2(y)=\sqrt{y^2-4r_e}(1+r_e-r_\pi-y)/\left[2(1+r_e-y)\right] \ ;\end{aligned}$$ or, equivalently, $$\begin{aligned} 2 \sqrt{r_\pi} \le z \le 1+r_\pi-r_e \ , \nonumber \\ F_3(z)-F_4(z) \le y \le F_3(z)+F_4(z) \ , \nonumber \\ F_3(z)=(2-z)(1+r_\pi+r_e-z)/[2(1+r_\pi-z)] \ , \nonumber \\ F_4(z)=\sqrt{z^2-4r_\pi}(1+r_\pi-r_e-z)/[2(1+r_\pi-z)] \ .\end{aligned}$$ For our aims we use the simplified form of physical region (omitting the terms of the order of $r_e$): $$2\sqrt{r_e} \le y \le 1-r_\pi \ ,\hspace{0.2cm} c(y) \le z \le 1+r_\pi$$ with $$c(y)=1-y+\frac{r_\pi}{1-y}$$ or, $$2\sqrt{r_\pi} \le z \le 1+r_\pi \ , \hspace{0.2cm} b_-(z) \le y \le b(z)$$ with $$b_-(z)=1-\frac{1}{2}\left(z+\sqrt{z^2-4r_\pi} \right) \ ,$$ $$b(z)=1-\frac{1}{2}\left(z-\sqrt{z^2-4r_\pi} \right) \ .$$ For definiteness we give here the numerical value for Born total width. It is: $$\begin{gathered} \frac{G_F^2M_K^5|V_{us}f_+(0)|^2}{64\pi^3}\int dy\int dz a_0(y,z)(1+\lambda_+\frac{t}{m_\pi^2})^2 \\ =5.36|V_{us}f_+(0)|^2\times 10^{-14} MeV.\end{gathered}$$ Comparing this value with PDG result: $(\Gamma_{K_+e3})_{exp}=(2.56\pm 0.03)\times 10^{-15}MeV$ we conclude $$(V_{us}f_+(0)|)_{\alpha=0}=0.218\pm0.002.$$ Virtual and soft real photon emission ===================================== Taking into account the accuracy level of 0.1% for determination of $\rho/\rho_0$ we will drop terms of order $r_e$. We will distinguish 3 kinds of contributions to $\delta$: from emission of virtual, soft real, and hard real photons in the rest frame of kaon: $\delta=\delta_V+\delta_S+\delta_H $. Standard calculation (see Appendix A for details) allows one to obtain the following contributions:\ from the soft real photons $$\label{ds} \delta_S=\frac{\alpha}{\pi}\left\{\left(L_e-2\right) \ln \frac{2\Delta\epsilon}{\lambda} + \frac{1}{2} L_e - \frac{1}{4} L_e^2 +1 - \frac{\pi^2}{6}\right\} \left(1+O(r_e) \right) \ ,$$ from the virtual photons $\delta_V=\delta_C+\delta_{PLM} $ that make up charged fermion’s renormalization, $\delta_C$ (throughout this paper we use Feynman gauge): $$\label{dc} \delta_C= \frac{\alpha}{2\pi} \left\{ \left[ - \frac{1}{2}L_\Lambda+ \frac{3}{2} \ln r_e + \ln \frac{M^2}{\lambda^2}- \frac{9}{4} \right]+ \left[L_\Lambda+ \ln \frac{M^2} {\lambda^2}- \frac{3}{4} \right] \right\} \ ,$$ here $L_\Lambda=\ln(\Lambda^2/M^2)$, $\Lambda$ is ultraviolet momentum cutoff, the first term in the curly braces comes from positron, the second one from kaon; and for the diagram in fig1(f) in the point like meson (PLM) approximation, $\delta_{PLM}$: $$\begin{gathered} \label{plm} \delta_{PLM}= \\ -\frac{\alpha}{2\pi} \left\{ -L_\Lambda-\frac{1}{2} \ln^2 r_e - 2 L_e+ \ln \frac{M^2}{\lambda^2} L_e -1+2 \ln^2y+ 2 \ln y+2Li_2(1-y) \right\} \ .\end{gathered}$$ When these contributions are grouped all together the dependence on $\lambda$ (fictitious “photon mass”) disappears. According to Sirlin’s prescription [@S] we set $\Lambda=M_W$. The result can be written in the form: $$\begin{gathered} \label{scplm} 1+\delta_S+\delta_C+\delta_{PLM}= S_W [1+\frac{\alpha}{\pi} \biggl[ \left( L_e-1 \right)\left(\ln \Delta +\frac{3}{4} \right) -\ln \Delta \\ -\frac{\pi^2}{6}+\frac{3}{4}-Li_2(1-y)-\frac{3}{2}\ln y \biggr]] , \qquad S_W=1+\frac{3\alpha}{4\pi} L_W.\end{gathered}$$ In the above equations $L_e=2\ln y +\ln(1/r_e)$, and $L_W=\ln(M_W^2/M^2)$; $M_W$ is the mass of $W^\pm$, $\Delta=\Delta\epsilon/E_e$, and $\Delta\epsilon$ is the maximal energy (in the rest frame of kaon) of a real soft photon. We imply $\Delta\epsilon \ll M/2$. For the details of eqs (\[ds\]), (\[dc\]), (\[plm\]), and (\[scplm\]) see Appendix A. Contribution from soft photon emission from structure–dependent part (such as for example, interaction with resonances and intermediate $W^\pm$) is small, of the order $$\frac{\alpha}{\pi} \frac{\Delta \epsilon}{M} \ll 1 \$$ and thus is also neglected. Hard photon emission. Structure function approach ================================================= Next we need to calculate contributions from hard photons. We have to distinguish between inner bremsstrahlung (IB) and the structure-dependent (SD) contributions: $\delta_H=\delta_{IB}+\delta_{int}+\delta_{SD}$, where $\delta_{int}$ is the interference term between the two. The terms $\delta_{int}$ and $\delta_{SD}$ are considered in the framework of the chiral perturbation theory (ChPT) to the orders of $(p^2)$ and $(p^4)$ and find their contribution to be at the level of $0.2\%$ (see appendix C). $$\begin{gathered} \label{deltaH} \delta_H =\frac{\alpha}{2\pi a_0(y,z)} \left\{ \right. \left(L_e-1 \right)\left(\Psi(y,z)-a_0(y,z) \left(2\ln \Delta + \frac{3}{2}\right) \right) - \\ \left. 2a_0(y,z) \ln \frac{b(z)-y}{y\Delta} \right\}+ \delta^{hard} \ ,\end{gathered}$$ with $\delta^{hard}$ given below. Extracting the short-distances contributions in form of replacement $\mathcal{C}\to \mathcal{C}S_W$ it is useful to split $\delta$ (see eq.(2))in the form $$\label{split} \delta(y,z)=\delta_L+\delta_{NL},$$ where $\delta_L$ is the leading order contribution, it contains the ’large logarithm’ $L_e$ and $\delta_{NL}$ is the non–leading contribution, it contains the rest of the terms. $\delta_L$ contains terms from $\delta_C$, $\delta_S$, $\delta_{PLM}$, and the contribution from the collinear configuration of hard IB emission (in the collinear configuration the angle between the positron and the emitted photon is small). $\delta_L$ turns out to be $$\label{dlp} \delta_L=\frac{\alpha(L_e-1)}{2\pi a_0(y,z)}\Psi(y,z) \ .$$ First we note that the kinematics of hard photon emission does not coincide with the elastic process (Region D, the strictly allowed boundaries of the Dalitz plot). In hard photon emission an additional region in the $y,z$ plane, namely $y<b_-(z)$ appears. The nature of this phenomenon is the same as the known phenomenon of the radiative tail in the process of hadron production at colliding $e^+e^-$ beams. The quantity $\Psi(y,z)$ has a different form for Region $D$ and outside it: $$\Psi(y,z)=\Psi_>(y,z), z>c(y), 2\sqrt{r_e}<y<1-r_\pi \ ;$$ and $$\Psi(y,z)=\Psi_<(y,z), z<c(y), 2\sqrt{r_e}<y<1-\sqrt{r_\pi} \ .$$ $\Psi_<(y,z)=0$ when $ y>1-\sqrt{r_\pi}$ . Functions $\Psi_<,\Psi_>$ are studied in Appendix D. $\delta_{NL}$ contains contributions from $\delta_C$, $\delta_{PLM}$, from SD part of hard photons and from the interference term of SD and IB parts of hard radiation. $$\label{nl} \delta_{NL}=\frac{\alpha}{\pi}\eta(z,y) \ ,$$ where $$\frac{\alpha}{\pi} \eta(y,z) = \delta^{hard} + \frac{\alpha}{\pi} \left[ \frac{3}{4} -\frac{\pi^2}{6} -Li_2(1-y)- \frac{3}{2} \ln y -\ln ((b(z)-y)/y) \right] \ ,$$ and for the case when the variables $y,z$ are inside $D$ region: $$\label{deltahard} \delta^{hard}=\frac{\alpha}{2\pi a_0(y,z)}Z_2(y,z) \ ;$$ $$\label{Z2} % Z_2(y,z)=-2 \tilde{\phi_1}(y,z) + \tilde{\phi_2}(y,z)+ \int \limits_0^{b(z)-y} dx Z_2(y,z)=-2 Rphot_{1D}(y,z) + Rphot_{2D}(y,z)+ \int \limits_0^{b(z)-y} dx \mathcal{J}(x,y,z) \ .$$ Explicit expressions for $Rphot_{1,2}$ and $\mathcal{J}$ are given in appendix B. The Born value and the correction to Dalitz-plot distribution $\Delta(y,z)= \delta(y,z) a_0(y,z)$ is illustrated in tables 1,2. We see that the leading contribution from virtual and soft photon emission is associated with the so called $\delta$–part of the evolution equation kernel: $$(\delta_C + \delta_S + \delta_{PLM})^{leading} = \frac{\alpha}{2\pi} \left(L_e-1 \right) \int \frac{a_0(t,z)}{a_0(y,z)} P_\delta^{(1)}\left(\frac{y}{t}\right) \frac{dt}{t}$$ where $$P_\delta^{(1)}(t)=\delta(1-t)\left(2 \ln \Delta + \frac{3}{2} \right) \ .$$ The contribution of hard photon kinematics in the leading order can be found with the method of quasireal electrons [@BFK] as a convolution of the Born approximation with the $\theta$–part of evolution equation kernel $P_\theta(z)$: $$\delta_H^{leading} \sim \frac{\alpha}{2\pi} \left(L_e-1 \right) \int \frac{dt}{t} \frac{a_0(t,z)}{a_0(y,z)} P_\theta^{(1)}\left(\frac{y}{t}\right)$$ where $$P_\theta^{(1)}(z)=\frac{1+z^2}{1-z}\theta(1-z-\Delta).$$ In such a way the whole leading order contribution can be expressed in terms of convolution of the width in the Born approximation with the whole kernel of the evolution equation: $$P^{(1)}(z)= \lim_{\Delta \to 0} \left( P_\delta^{(1)}(z)+ P_\theta^{(1)}(z) \right) \ .$$ This approach can be extended to use nonsinglet structure functions $D(t,y)$ [@KF]: $$\begin{aligned} d\Gamma^{LO}(y,z)=\int\limits_{max[y,b_-(z)]}^{b(z)} \frac{dt}{t} \ d\Gamma_0 \left( t,z \right) D \left( \frac{y}{t}, \ L_e \right), \ t=x+y \ ,\\ \nonumber %\end{equation} % %\begin{equation} \label{D} D(z,L)=\delta(1-z)+\frac{\alpha}{2\pi} L P^{(1)}(z)+ \frac{1}{2!} \left( \frac{\alpha L }{2\pi} \right)^2P^{(2)}(z)+... \ , \\ \nonumber %\end{equation} % %\begin{equation} % P^{(1)}(z)=\left[ \frac{1+z^2}{1-z}\theta \left( 1-z-\Delta \right) % +\delta \left( 1-z \right) \left(2\ln\Delta+ % \frac{3}{2} \right) \right]_{\Delta\to 0} \ ,\nonumber %\end{equation} % %\begin{equation} \label{Pi} P^{(i)}(z)=\int\limits_z^1\frac{dx}{x}P^{(1)}(x)P^{(i-1)}(\frac{z}{x}) \ , \ i=2,3,... \ .\end{aligned}$$ One can check the validity of the useful relation: $$\int\limits_0^1 dz P^{(1)}(z)=0 \ .$$ The above makes it easy to see that in the limit $m_e \to 0$ terms that contain $m_e$ do not contribute to the total width in correspondence with Kinoshita–Lee–Nauenberg (KLN) theorem [@KLN] as well as with results of E.Ginsberg [@G]. Keeping in mind the representation $$\label{psii} \Psi(y,z) = \int \limits_{max[y,b_-(z)]}^{b(z)} \frac{dt}{t} \ a_0(t,z) P^{(1)} \left( \frac{y}{t} \right) \ ,$$ one can get convinced (see Appendix D) that the leading logarithmical contribution to the total width as well as one to the pion spectrum is zero due to: $$\int\limits_{2\sqrt{r_\pi}}^{1+r_\pi} dz\int\limits_0^{b(z)}dy\Psi(y,z)=0.$$ Using the general properties of the evolution equations kernels, eq(\[D\]) one can see that KLN cancellation will take place in all orders of the perturbation theory. The spectra in Born approximation are (we omit terms $O(r_e) \sim 10^{-6}$):\ For pion $$\begin{aligned} &&\frac{1}{\mathcal{C}} \frac{d \Gamma_0}{dz} = \phi_0(z) \ , \\ \nonumber &&\phi_0(z)=\left(1+\frac{\lambda_+}{r_\pi}R(z) \right)^2 \int \limits_{b_-(z)}^{b(z)} dy a_0(y,z)= \left(1+\frac{\lambda_+}{r_\pi}R(z) \right)^2 \frac{1}{6}\left(z^2-4r_\pi \right)^{3/2} \ ,\end{aligned}$$ and for positron $$\begin{gathered} \label{39} f(y) = \frac{1}{\mathcal{C}} \frac{d \Gamma_0}{dy} = \\ f_0(y) \left[1+\frac{2}{3} \left(\frac{\lambda_+}{r_\pi} \right) \frac{y(1-r_\pi-y)} {1-y} + \frac{1}{6} \left(\frac{\lambda_+}{r_\pi} \right)^2 \frac{y^2(1-r_\pi-y)^2}{(1-y)^2} \right] \ , \\ f_0(y)=\frac{y^2(1-r_\pi-y)^2}{2(1-y)}.\end{gathered}$$ The corrected pion spectrum in the inclusive set-up of experiment when integrating over the whole region for $y$ ($0<y<b(z)$) have a form $\phi_0(z)+(\alpha/\pi)\phi_1(z)$ with $$\begin{gathered} \label{rcpions} \phi_1(z)= \left(1+\frac{\lambda_+}{r_\pi}R(z)\right)^2 \biggl[\int\limits_0^{b_-(z)}dy [\Psi_<(y,z)\ln y-% \right. a_0(y,z)\ln\frac{b(z)-y}{b_-(z)-y}+\\ \frac{1}{2}\tilde{Z}_2(y,z)]+ \int\limits_{b_-(z)}^{b(z)}dy [\Psi_>(y,z)\ln y+a_0(y,z)Z_1(y,z)+ \frac{1}{2}Z_2(y,z)] \biggr],%\left.\end{gathered}$$ the quantities $Z_1,\tilde{Z}_2$ explained in Appendix E. This function do not depend on $\ln(1/r_e)$. Pion spectrum in the exclusive set-up ($y,z$ in the region $D$) will depend on $L_e$. It’s expression is given in the Appendix E. Numerical estimation of pion spectrum is illustrated in figure (3,5). The inclusive positron spectrum with the correction of the lowest order is $f(y)+(\alpha/\pi) f_1(y)$ with $f(y)$ given above and: $$\begin{gathered} \label{rcposis} f_1(y)= \frac{1}{2} \left( L_e-1 \right)I(y) -\int\limits_{c(y)}^{1+r_\pi}a_0(y,z) \left(1+\frac{\lambda_+}{r_\pi} R(z) \right)^2 \ln ((b(z)-y)/y) dz+ \\ % \frac{3}{4}L_W + \left(\frac{3}{4}-\frac{\pi^2}{6}-\frac{3}{2}\ln y-Li_2(1-y) \right) f(y) + \frac{1}{2} \int\limits_{c(y)}^{1+r_\pi} Z_2(y,z) \left(1+ \frac{\lambda_+}{r_\pi} R(z)\right)^2 dz + \\ \theta(1-\sqrt{r_\pi}-y)\int\limits_{2\sqrt{r_\pi}}^{c(y)} dz(1+\frac{\lambda_+}{r_\pi}R(z))^2 [(1/2)\tilde{Z}_2-a_0(y,z) \ln\frac{b(z)-y}{b_-(z)-y}] ,\end{gathered}$$ with $$\label{fullI} I(y)=j_0(y)+\left(\frac{\lambda_+}{r_\pi} \right)j_1(y)+ \left(\frac{\lambda_+}{r_\pi} \right)^2 j_2(y) \ ,$$ $$\begin{gathered} j_0(y)=\int \limits_y^{1-r_\pi}\frac{d t}{t} \int \limits_{c(t)}^{1+r_\pi}dz a_0(t,z)P^{(1)}(\frac{y}{t})= \\ (2\ln\frac{1-r_\pi-y}{y}+\frac{3}{2})f_0(y)+ \frac{r_\pi^2(1+y^2)}{2(1-y)}\ln\frac{1-y}{r_\pi}+ \\ \frac{1}{12}(1-r_\pi-y)[1-5r_\pi-2r_\pi^2+y(4-13r_\pi)-17y^2] \ ;\end{gathered}$$ explicit expressions for $j_1(y)$ and $j_2(y)$ are given in Appendix D. Numerical estimation of positron spectrum is illustrated in figure (4,6). One can check the fulfillment of KLN cancellation of singular in the limit $m_e \to 0$ terms for the total width correction:$\int_0^{1-r_\pi}I(y)dy=0$. The expression for $j_0(y)$ agree with A(2) from the 1966 year paper of Eduard Ginsberg [@G]. We put here the general expression for differential width of hard photon emission which might be useful for construction of Monte Carlo simulation of real photon emission in $K_{e3}$: $$d \Gamma_\gamma^{hard} = d \Gamma_0 \frac{\alpha}{2\pi}\frac{dx}{x} \frac{dO_\gamma}{2\pi a_0(y,z)} T \ ,$$ with $$x=\frac{2\omega}{M} > \frac{2\Delta\epsilon}{M}= y\frac{\Delta\epsilon}{E_e}, \ \frac{\Delta\epsilon}{E_e} \ll 1 \ ;$$ and $dO_\gamma$ is an element of photon’s solid angle. The quantity $T$ is explained in Appendix B. For soft photon emission we have $$\label{dgammasoft} d \Gamma_\gamma^{soft} = d \Gamma_0 \frac{\alpha}{2\pi}\frac{dx}{x} \frac{dO_\gamma}{2\pi} \left[ -1-\frac{r_e}{(1-\beta_e C_e)^2}+ \frac{y}{1-\beta_e C_e} \right], \ x<y\frac{\Delta\epsilon}{E_e} \ .$$ Integrating over angles within phase volume of hard photon we obtain the spectral distribution of radiative kaon decay: $$\begin{gathered} \label{47} \frac{d \Gamma}{d \Gamma_0 dx}=\frac{\alpha}{2\pi} \frac{1}{a_0(y,z)} \biggl[\frac{a_0(x+y,z)}{x(x+y)^2} %\right. \left((y^2+(x+y)^2)(L_e-1)+x^2 \right) \\ \nonumber -\frac{2}{x}a_0(y,z)-2(\frac{R(z)}{x+y}-y) +\mathcal{J}(x,y,z) \biggr], \\ y\Delta<x<b(z)-y , \qquad \Delta=\frac{\Delta \epsilon}{E_e} \ll 1.\end{gathered}$$ Discussion ========== Structure–dependent contribution to emission of virtual photons (see Fig 1 d), e)) can be interpreted as a correction to the strong formfactor of $K \pi$ transition, $f_+(t)$. We assume that this formfactor can be extracted from experiment and thus do not consider it. The problem of calculation of RC to $K_{e3}$ and especially the formfactors in the framework of CHpT with virtual photons was considered in a recent paper [@CKNRT]. As in paper [@G] we assume a phenomenological form for the hadronic contribution to the $K-\pi$ vertex, but here we use explicitly the dependence of the form factor in the form $$f_+(t)=f_+(0) \left(1+\frac{\lambda_+}{r_\pi} R(z) \right) \ .$$ We assume that the effect of higher order ChPT as well as RC to the formfactors can be taken as a multiplicative scaling factor for $f_+(0)$, which we take from of a recent paper [@CKNRT]. We assume such an experiment in which only one positron in the final state is present, but place no limits on the number of photons. The ratio of the LO contributions in the first order to the Born contribution is a few percent, and for the second order it is about $$\begin{aligned} \left( \frac{\alpha L_e}{2 \pi} \right)^2 \le 0.03\% \ .\end{aligned}$$ Due to non-definite sign structure of the leading logarithm contribution (see eq(\[dlp\])) there are regions in the kinematically allowed area where $|\Psi(y,z)|$ is close to zero. In these regions the non-leading contributions become dominant. The contribution of the channels with more than 1 charged lepton in the final state as well as the vacuum polarization effects in higher orders may be taken into account by introducing the singlet contribution to the structure functions. The effect will be at the level of 0.03% and we omit them within the precision of our calculation. The contribution of the $O(p^4)$ terms [@BEG] turns out to be small. Indeed, one can see that they are of the order $O(\alpha L^r_9, (\bar{p}/\Lambda_c)^2) \le O(10^{-2}\%)$, $\Lambda_c=4 \pi F_{\pi} \approx 1.2 GeV$ ($F_\pi=93 MeV$ is the pion life time constant), where $\bar{p}$ is the characteristic momentum of a final particle in the given reaction, $\bar{p}^2 \le M^2/16 \sim F^2_\pi$. So the terms of the orders $O(p^4)$ and $O(p^6)$ can be omitted within the accuracy of $O \left( \frac{\alpha}{\pi} \times 10^{-2} \right) \le O \left( 10^{-4} \right)$. Our main results are given in formulae (2,21,22,26-28) for Dalitz plot distribution ; (38-41) for pion and positron spectra;(47) for hard photon emission;(53) for the value $|V_{us}|$, in the tables and figures. The accuracy of these formulas is determined by the following: 1. we don’t account higher order terms in PT, the ones of the order of $(\alpha L_e/\pi)^n, n \ge 2$ which is smaller than $0.03 \%$ 2. structure–dependent real hard photon emission contribution to RC we estimate to be at the level of $0.0005$. 3. higher order CHPT contributions to the structure dependent part are of the order $0.05\%$ [@BCEG; @BEG] All the percentages are taken with respect to the Born width. All together we believe the accuracy of the results to be at the level of $0.01$. So the final result of our calculation may be written in the form $$\label{fr} \frac{\Gamma}{\Gamma_0}=\left( 1+\delta(1\pm0.01) \right) %\left( 1+O \left( 5 \times 10^{-4} \right) \right).$$ the terms on the RHS are given in (\[split\], \[dlp\], \[nl\]). Here is the list of improvements comparing with the older calculations [@B; @G]: 1. we eliminate the ultra–violet cutoff dependence by choosing $\Lambda=M_W$, 2. we describe the dependence on the lepton mass logarithm $L_e$ in all orders of the perturbation theory and explain why the correction to the total width does not depend on $m_e$, 3. we treat the strong interaction effects by the means of CHPT in its lowest order $O(P^2)$ and show that the next order contribution is small, 4. we give an explicit formulas for the total differential cross section and explicit results for corrections to the Dalitz plot and particle spectra that might be used in experimental analysis. In the papers of E.Ginsberg the structure-dependent emission was not considered. T.Becherrawy, on the other hand, did include it, and this will give rise to differences in the Dalitz plot. In addition, Ginsberg used the proton mass as the momentum cutoff parameter. We do not consider the evolution of coupling constant effects in the regions of virtual photon momentum modulo square from the quantities of order $M_\rho^2$ up to $M_Z^2$,which can be taken into account [@MS] (and for details see [@CKNRT]) by replacing the quantity $S_W$ by the $S_{EW}=1+(\alpha/\pi)\ln(M_Z^2/M_\rho^2)=1.0232$. Taking this replacement into account our result for the correction to the total width is $$\frac{\Gamma}{\Gamma_0}=1+\delta=1.02,$$ which results in $$|V_{us}f_+(0)|=0.214\pm0.002.$$ So the correction to the total width is $+2 \%$ while Ginsberg’s result is $-0.45 \%$ and Becherrawy’s is $-2 \%$. Neither Ginsberg nor Becherrawy used the factor $S_{EW}$, and this factor (1.023) accounts for most of the difference between Ginsberg’s and our result. Electromagnetic corrections become negative and have an order of $10^{-3}$. The effect of the SD part, which E.Ginsberg did not consider, is small, of the order of 0.1%. We use the value of formfactor $f_+(0)=0.9842\pm0.0084$ calculated in the paper [@CKNRT] in the framework of ChPT, including virtual photonic loops and terms of order $O(p^6)$ of ChPT. To avoid double counting we use the mesonic contribution to $f_+^{mes}(0)=1.0002\pm0.0022$ and the $p^6$ terms one $f_+(0)|_{p^6}=-0.016\pm0.0008$). Our final result is: $$|V_{us}|=0.21715\pm0.0055.$$ In estimating the uncertainty we take into account the uncertainties arising from structure-dependent emission $\pm 0.005$, theoretical errors of order $\pm 0.0003$, experimental error $\pm 0.0022$ and the ChPT error in $p^6$ terms $0.0008$. In tables 1,2 we give corrections to the distributions in the Dalitz-plot $d\Gamma/(dy dz) \sim a_0(y,z)+(\alpha/\pi)\Delta(y,z)$. In figures (3-6) we illustrate the corrections to the pion and positron spectra. Here we see qualitative agreement for the positron spectrum and disagreement with the pion spectrum obtained by E.Ginsberg. Acknowledgements ================ We are grateful to S.I.Eidelman for interest to this problem, and to E. Swanson for constructive critical discussions. One of us (E.K.) is grateful to the Department of Physics and Astronomy of the University of Pittsburgh for the warm hospitality during the last stages of this work, A.Ali for valuable discussion and V.N.Samoilov for support. The work was supported partly by grants RFFI 01-02-17437 and INTAS 00366 and by USDOE DE FG02 91ER40646. $z/y$ 0.07 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 ------- ------ ------ ------ ------ ------- ------- ------- ------- ------- 1.025 3.83 4.37 3.71 2.01 -0.21 -2.47 -4.31 -5.11 -3.81 0.975 3.76 3.49 2.07 0.05 -2.07 -3.83 -4.61 -3.35 0.925 3.26 2.13 0.32 -1.67 -3.35 -4.11 -2.88 0.875 3.04 2.18 0.58 -1.26 -2.86 -3.60 -2.39 0.825 2.25 0.85 -0.86 -2.37 -3.08 -1.88 0.775 1.14 -0.41 -1.83 -2.51 -1.28 0.725 1.39 -0.04 -1.39 -2.03 -0.72 0.675 0.38 -0.86 -1.48 0.625 -0.35 -0.89 0.580 -0.23 Table 1. Correction to Dalitz plot distribution $\Delta(y,z)=a_0(y,z)\delta(y,z)\times 10^{3}$ (see eq (2)). $z/y$ 0.07 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 ------- -------- ------- ------- -------- ------- ------- ------- ------- -------- 1.025 0.0084 0.069 0.126 0.163 0.181 0.178 0.156 0.114 0.051 0.975 0.026 0.089 0.131 0.153 0.156 0.139 0.101 0.0437 0.925 0.051 0.099 0.126 0.134 0.121 0.088 0.036 0.875 0.014 0.066 0.099 0.111 0.104 0.076 0.029 0.825 0.0337 0.071 0.089 0.086 0.064 0.021 0.775 0.043 0.066 0.069 0.051 0.014 0.725 0.016 0.044 0.051 0.039 0.006 0.675 0.021 0.034 0.026 0.625 0.016 0.014 0.580 0.003 Table 2. Dalitz plot distribution in Born approximation $a_0(y,z)$. \[fig3\] \[fig4\] \[fig5\] \[fig6\] Appendix A ========== Here we explain how to calculate $\delta_S$, $\delta_C$, $\delta_{PLM}$ and how to group them into eq(\[scplm\]). Contribution from emission of a soft real photon can be written in a standard form in terms of the classical currents: $$\delta_S=- \frac{4\pi\alpha}{(2\pi)^3} \int \frac{d^3q}{2\omega} \left( \frac{p}{p \cdot q}-\frac{p_e}{p_e \cdot q} \right)^2 \biggm|_{\omega=\sqrt{\bar{q}^2+\lambda^2} < \Delta\epsilon},$$ where $\lambda$ is fictitious mass of photon. We use the following formulas: $$\frac{1}{2\pi} \int \frac{d^3q}{2\omega} \left( \frac{p}{p \cdot q} \right)^2= \ln \left( \frac{2\Delta\epsilon}{\lambda} \right)-1 \; ;$$ $$\frac{1}{2\pi} \int \frac{d^3q}{2\omega} \left( \frac{p_e}{p_e \cdot q} \right)^2= \ln \left( \frac{2\Delta\epsilon}{\lambda} \right)-\frac{1}{2}L_e \ ;$$ $$\frac{1}{2\pi} \int \frac{d^3q}{2\omega} \frac{2(p \cdot p_e)}{(p \cdot q)( p_e \cdot q)} = L_e \ln \left( \frac{2\Delta\epsilon}{\lambda} \right)- \frac{\pi^2}{6}-\frac{1}{4}L^2_e \ .$$ From them we obtain the eq(\[ds\]). Consider now radiative corrections that arise from emission of virtual photons (excluding SD virtual photons). Feynman graphs containing self–energy insertion to positron and kaon Green functions (Fig.1,b,c) can be taken into account by introducing the wave function renormalization constants $Z_e$ and $Z_K$: $M_0 \to M_0(Z_KZ_e)^{1/2}$. We use the expression for $Z_e$ given in the textbooks [@AB]; the expression for $Z_K$ is given in the paper [@Chang]. The result is eq(\[dc\]). Now consider the Feynman graph in which a virtual photon is emitted by kaon and absorbed by positron or by $W$–boson in the intermediate state (Fig.1,d,e,f). This long distance contribution is calculated using a phenomenological model with point–like mesons as a relevant degrees of freedom. To calculate the contribution from the region $|k|^2 < \Lambda^2$ ($\Lambda$ is the ultra violet cutoff) we use the following expressions for loop momenta scalar, vector, and tensor integrals: $$Re \int \frac{d^4 k}{i\pi^2} \frac{1, \ k^\mu, \ k^2} {(k^2-\lambda^2)((k-p)^2-M^2)((k-p_e)^2-m_e^2)} = I, \ I^\mu, \ J \ .$$ A standard calculation yields: $$\begin{aligned} I=\frac{-1}{yM^2} \left\{ \frac{1}{2} \ln \frac{M^2}{\lambda^2} L_e+ \ln^2y + Li_2(1-y)- \frac{1}{4} \ln^2r_e \right\} \ ;\\ \nonumber %\end{equation} % %\begin{equation} I^\mu=\frac{-1}{yM^2} \left\{ \frac{-y \ln y}{1-y}p^\mu+ p_e^\mu \left( \frac{y \ln y}{1-y}+ L_e \right) \right\} \ ;\\ \nonumber %\end{equation} % %\begin{equation} J = L_\Lambda + \frac{y \ln y}{1-y} + 1 \ .\end{aligned}$$ where $L_\Lambda=\ln(\Lambda^2/M^2)$ and we omitted terms of the order of $O(m_e^2/M^2)$. As a result we obtain $$\begin{gathered} \int \frac{d^4 k}{i\pi^2} \frac{(1/4)Sp \ p_\nu(p+p')(-\hat{p}_e+\hat{k}) (2\hat{p}-\hat{k})p_e(p+p')}{(k^2-\lambda^2)((k-p)^2-M^2) ((k-p_e)^2-m_e^2)}= 2M^4a_0(y,z) \times \\ \left\{ -L_\Lambda-\frac{1}{2} \ln^2 r_e - 2 L_e+ \ln \frac{M^2}{\lambda^2} L_e -1+2 \ln^2y+ 2 \ln y+2Li_2(1-y) \right\} \ .\end{gathered}$$ In a series of papers [@S] A.Sirlin has conducted a detailed analysis of UV behaviour of amplitudes of processes with hadrons in 1–loop level. He showed that they are UV finite (if considered on the quark level), but the effective cutoff scale on loop momenta is of the order $M_W$. For this reason we choose $$L_\Lambda = \ln \frac{M_W^2}{M^2} \ .$$ The sum $\delta_S+\delta_C+\delta_{PLM}$ yields eq(\[scplm\]) Appendix B ========== The matrix element of the radiative $K_{e3}$ decay $$K^+(p) \to \pi^0(p')+e^+(p_e)+\nu(p_\nu)+\gamma(q)$$ with terms up to $O(p^2)$ in CHPT [@GL; @BCEG; @BEG; @H] has the form $$M^{hard} = \frac{G}{2} f_+ V_{us}^* \sqrt{4\pi\alpha} \bar{u}(p_\nu) Q_\mu^{hard}(1+\gamma_5) v(p_e) \epsilon^\mu(q) \ ,$$ where $$\begin{aligned} Q_\mu^{hard} = Q_\mu^e + Q_\mu^\pi + Q_\mu^{SD} = Q_\mu^{IB} + Q_\mu^{SD} \ ;\\ \nonumber %\end{equation} %\begin{equation} Q_\mu^{IB}=\left(\hat{p}+\hat{p}'\right) \left[ \frac{(-\hat{p}_e- \hat{q}+m_e)\gamma_\mu} {2p_eq} + \frac{p_\mu}{pq} \right] \ ;\\ \nonumber %\end{equation} %\begin{equation} Q_\mu^{SD}=\gamma_\nu R_{\mu\nu} \ .\end{aligned}$$ where tensor $R_{\mu\nu}$ describes (see eq(4.17) in [@BCEG]) the structure–dependent emission (fig 2(c)). $$R_{\mu\nu} = g_{\mu\nu} - \frac{q_\nu p_\mu}{pq} \ .$$ Singular at $\chi=2p_e q\to 0$ terms which provide contribution containing large logarithm $L_e$ arises only from $Q_\mu^e$. To extract the corresponding terms we introduce four–vector $v=(x/y)p_e-q$, and $x$ – is the energy fraction of the photon (9). Note that $v \to 0$ when $\chi \to 0$. Separating leading and non–leading terms and letting $f_+(t)=1$, i.e. neglecting form factor’s momentum dependence, we obtain: $$\delta_H=\frac{d\Gamma^{hard}}{d\Gamma_0}= \frac{\alpha}{2\pi a_0(y,z)} \int \frac{dx}{x} \int \frac{dO_\gamma}{2\pi}T \ , x>y\Delta \ .$$ where $$\begin{gathered} \label{T} T= \frac{x^2}{8} \sum_{spins} \Bigm|\bar{u}(p_\nu) \left(Q_{IB}^{hard}+ Q_{SD}^{hard} \right)(1+\gamma_5) v(p_e) \Bigm|^2 = \\ \frac{y a_0(x+y,z)}{x+y} \left[ \frac{y^2+(x+y)^2}{y^2(1-\beta_e C_e)} - 2\frac{(1-\beta_e)(x+y)}{y(1-\beta_e C_e)^2} \right]- \frac{ya_0(x+y,z)}{x+y}+ \mathcal{P} \ .\end{gathered}$$ The quantity $\mathcal{P}$ contains some non–leading contributions from the IB part and the ones that arise from the structure–dependent part: $$\begin{gathered} \mathcal{P}= \left( \frac{p_e q}{M^2} \left( \frac{p_\nu q}{M^2} + z - \frac{2y}{x+y}\left( 1-x-y\right) \right) + \frac{p'v}{M^2} \frac{y(2-x-y)}{x+y} \right) \\ \left(\frac{xM^2}{4 y p_e q} \left(y^2+(x+y)^2) -1 \right) \right)- \frac{M^2x^2}{8p_eq} \left(T_v+\frac{2}{x}T_{1v} \right)- \frac{x^2}{8} \left(T_{RR}+2T_R \right) \ ,\end{gathered}$$ with $$T_v=\frac{1}{4 M^4}Sp \ (\hat{p}+\hat{p}')\hat{p}_\nu(\hat{p}+ \hat{p}')\hat{v} \ ;$$ $$T_{1v}=\frac{1}{4 M^6}Sp \ (\hat{p}+\hat{p}') \hat{p}_\nu(\hat{p}+\hat{p}') \hat{v} \hat{p} \hat{p}_e \ ;$$ $$T_{RR}=R_{\mu\lambda}R_{\mu\sigma}\frac{1}{4M^2}Sp \ \hat{p}_\nu\gamma_\lambda \hat{p}_e\gamma_\sigma;$$ $$T_R=R_{\mu\lambda}\frac{1}{4M^2}Sp \ \hat{p}_\nu(\hat{p}+\hat{p}') \left[ \frac{p_\mu}{pq}- \frac{(\hat{p}_e+\hat{q})\gamma_\mu}{\chi} \right] \hat{p}_e\gamma_\lambda \ .$$ To calculate these traces we use the following expressions for the scalar products of the 4–momenta (in units $M$): $$p^2=1, \hspace{0.25cm} q^2=0, \hspace{0.25cm} p_\nu^2=0, \hspace{0.25cm} p^{'2}=r_\pi , \hspace{0.25cm} p_e^2 = 0, \hspace{0.25cm} p p_e = \frac{y}{2},$$ $$p p'= \frac{z}{2}, \hspace{0.25cm} p q= \frac{x}{2}, \hspace{0.25cm} p p_\nu= \frac{1}{2}\left( 2-y-z-x \right),$$ $$p' p_\nu = \frac{1}{2} \left(1-x-y-r_\pi+A_e \right), \hspace{0.25cm} p' q= \frac{1}{2} \left( x-A_e-A_\nu \right),$$ $$p' p_e =\frac{1}{2} \left(y-R(z)+A_\nu \right), \hspace{0.25cm} p_\nu q = \frac{1}{2} A_\nu , \hspace{0.25cm} p_e q = \frac{1}{2} A_e ,$$ $$p_e p_\nu = \frac{1}{2} \left(R(z)-A_e-A_\nu \right), \hspace{0.25cm} p v =0, \hspace{0.25cm} p_e v =-\frac{1}{2} A_e ,$$ $$q v = \frac{1}{2} \frac{x}{y} A_e , \hspace{0.25cm} p'v = \frac{1}{2} \left( \frac{x+y}{y} \tilde{A}_\nu + A_e\right) ,$$ $$p_\nu v = -\frac{1}{2y}\left(x A_e + (x+y) \tilde{A}_\nu \right) \ ,$$ $$\tilde{A}_\nu = A_\nu -\frac{x}{x+y} R(z) \ .$$ Three terms in the rhs of (\[T\]) have a completely different behavior. The first one corresponds to the kinematic region of collinear emission, when photon is emitted along positron’s momentum. The relevant phase volume has essentially 3–particles form: $$\begin{gathered} (d \phi_4)^{coll} = \left(\frac{d^3 p_e}{2\epsilon_e} \frac{d^3 q}{2 \omega} \frac{d^3 p'}{2 \epsilon'} d^4 p_\nu \delta(p_\nu^2)\delta^4(p-p_e-p_\nu-p'-q) \right)^{coll}= \\ M^4 \frac{\pi^2}{64} \beta_\pi z dz y dy x dx d O_\gamma d C_{e\pi} \delta(1-x-y-z+r_\pi+\frac{x+y}{y} \frac{zy}{2} (1-\beta_\pi C_{e\pi})+ \frac{2 p_e q}{M^2})= \\ \frac{y}{x+y} M^4 \frac{\pi^2}{32} dO_\gamma x dx dy dz \ .\end{gathered}$$ The limits of photon’s energy fraction variation are $y\Delta<x<b(z)-y$. The upper limit is imposed by the Born structure of the width in this kinematical region. The second term corresponds to the contribution from emission by kaon. The relevant kinematics is isotropic. The kinematics of radiative kaon decay and the comparison of our and E.Ginsbergs approaches is given in Appendix F. The third term corresponds to the rest of the contributions which contain neither collinear nor infrared singularities. Performing the integration over photon’s phase volume provided $y,z$ are in the $D$ region we obtain: $$\begin{gathered} \int\frac{dx}{x}\int\frac{dO_\gamma}{2\pi}T=\int\limits_{y\Delta}^{b(z)-y} \frac{dx}{x} \frac{y^2}{(y+x)^2}a_0(y+x,z)[\frac{y^2+(y+x)^2}{y^2}(L_e-1)+ \frac{x^2}{y^2}] \ \\ -2\int\limits_{y\Delta}^{b(z)-y}\frac{dx}{x}[a_0(y,z)+x(\frac{R(z)}{x+y}-y)]+ \int\limits_0^{b(z)-y} dx\mathcal{J},\end{gathered}$$ we obtain (\[deltahard\]) with $$%\tilde{\phi_1}=\int\limits_0^{b(z)-y} dx(\frac{R(z)}{x+y}-y)=R(z)\ln\frac{b(z)}{y}-y(b(z)-y); Rphot_{1D}=\int\limits_0^{b(z)-y} dx(\frac{R(z)}{x+y}-y)=R(z)\ln\frac{b(z)}{y}-y(b(z)-y);$$ $$\begin{gathered} % \tilde{\phi_2}(y,z)=\int\limits_0^{b(z)-y}dx\frac{xa_0(y+x,z)}{(y+x)^2}=-\left(R(z)+y(2-z)\right) \ln \frac{b(z)}{y} + \\ \frac{1}{2} Rphot_{_2D}(y,z)=\int\limits_0^{b(z)-y}dx\frac{xa_0(y+x,z)}{(y+x)^2}=-\left(R(z)+y(2-z)\right) \ln \frac{b(z)}{y} + \\ \frac{1}{2} \left(b(z)-y \right) \left(2 \frac{R(z)}{b(z)} +4 -2z -b(z) +y \right) \ ,\end{gathered}$$ and $$\mathcal{J}(x,y,z)= \frac{1}{x} \int \frac{dO_\gamma}{2\pi} \mathcal{P} \ .$$ One can check that the sum of RC arising from hard, soft and virtual photons do not depend on the auxiliary parameter $\Delta$. We note that the leading contribution from hard part of photon spectra can be reproduced using the method of quasi real electrons [@BFK]. [^1] Now we concentrate on the contribution of the third term in the RHS of (\[T\]). To perform the integration over the phase volume of final states it is convenient to use the following parameterization (see Appendix F): $$d\phi_4=\frac{d^3p'd^3p_ed^3p_\nu d^3q} {2\epsilon' 2E_e 2\epsilon_\nu 2\omega} \delta^4 \left( p-p'-p_e-p_\nu-q \right) = \beta_\pi \frac{\pi^2}{16}M^4 dy dz x dx \frac{d C_e d C_\pi}{\sqrt{D}} \ ,$$ with $$\begin{aligned} D=\beta_\pi^2(1-C^2-C_\pi^2-C_e^2+2C C_\pi C_e) \ , \ \beta_\pi= \sqrt{1-\frac{4r_\pi}{z^2}} \ ,\\ \nonumber %\end{equation} %\begin{equation} C=\cos(\vec{p}_e,\vec{p'}), \ C_e=\cos(\vec{q},\vec{p}_e), \ C_\pi=\cos(\vec{q},\vec{p'}) \ .\nonumber\end{aligned}$$ The neutrino on–mass shell (NMS) condition provides the relation $$1-\beta_\pi C = \frac{2}{yz} \left[ x+y+z-1-r_\pi-\frac{xz}{2} \left( 1-\beta_\pi C_\pi \right) -\frac{xy}{2} \left( 1-C_e \right) \right] \ .$$ For the aim of further integration of $\mathcal{P}$ over angular variables we put it in the form: $$\begin{aligned} \mathcal{P}=x P_1 \frac{\tilde{A}_\nu}{A_e} + x P_2 + P_3 A_e + P_4 A_\nu + P_5 A_\nu A_e \ , \\ \nonumber %\] %\[ A_e=\frac{xy}{2}\left(1-C_e \right) \ ,\\ \nonumber %\] %\[ A_\nu=x-A_e-\frac{xz}{2}\left(1-\beta_\pi C_\pi \right) \ . %\]\end{aligned}$$ and $$\begin{aligned} P_1 = \frac{y}{2} \left( 1-x-y \right) \ ;\\ \nonumber %\end{equation} %\begin{equation} P_2 = \frac{R(z)}{x+y} + \frac{1}{2} \left(z(2x+3y+1)+2x^2+4xy+3y^2 -2x-3y-2 \right);\\ \nonumber %\end{equation} %\begin{equation} P_3 = 1-z-y-\frac{1}{2}x\left(x+y+z \right) \ ;\\ \nonumber %\end{equation} %\begin{equation} P_4 = -1+x+y+\frac{1}{2}xy \ ;\nonumber %\end{equation} %\begin{equation} P_5 = -1 \ .\end{aligned}$$ Angular integration can be performed explicitly, we have $$\int \frac{\beta_\pi dC_\pi}{\pi \sqrt{D}}= \frac{y}{\sqrt{A}} \ , \int \frac{\beta_\pi C_\pi dC_\pi}{\pi \sqrt{D}} = \frac{y(x+y-yt)}{z \beta_\pi A^{3/2}} \left[ 2R(z)-(x+y)(2-z)+xyt) \right] \ ,$$ with $$A=(x+y)^2-2xyt, \; t=1-C_e \ .$$ Performing the integration over $C_\pi$ we have: $$\begin{gathered} \frac{1}{x}\int \frac{d C_\pi \beta_\pi}{\pi \sqrt{D}}\mathcal{P} = \frac{2y}{A^{3/2}} \left( \left(y-x \right)\left(1-\frac{z}{2}- \frac{R}{x+y} \right) -\frac{1}{2}y \left(x+y-xt \right) \right) P_1+ \\ \frac{y}{A^{1/2}}\left( P_2 + \frac{y}{2}t P_3 \right)+ \left(P_4+\frac{xy}{2}t P_5 \right) \times \\ \left\{ \frac{y}{A^{1/2}} \left(1-\frac{z}{2}-\frac{y}{2}t \right)+ \frac{y}{A^{3/2}} \left(x+y-yt \right)\left(R-(x+y)\left(1-\frac{z}{2} \right)+\frac{xy}{2}t \right) \right\} \ .\end{gathered}$$ The following integrals are helpful in integrating the above expression. We define $$I^m_{n}=\int_0^2 \frac{dt t^m}{\sqrt{A^n}}, \ m=0,1,2,3; \ n=1,3.$$ Then $$\begin{aligned} I^0_1=\frac{4}{\sigma},\qquad I^1_1=\frac{8 (x+y+\sigma)}{3 \sigma^2},\nonumber \\ I^2_1= \frac{16}{15 \sigma^3} \left(3 \sigma^2+3 (x+y) \rho+5(x+y)^2 \right),\nonumber \\ I^0_3=\frac{4}{\rho \sigma (x+y)}, \qquad I^1_3=\frac{8}{\rho \sigma^2}, \qquad I^2_3=\frac{16}{3 \rho \sigma^3} \left(2(x+y)+\sigma \right), \nonumber \\ I^3_3=\frac{32}{5\rho\sigma^4} \left( \sigma^2+2 (x+y) \rho +4(x+y)^2 \right), \nonumber \\\end{aligned}$$ where $\rho=|x-y|$ and $\sigma=x+y+\rho$. The first term in $d\Gamma^{hard}$ together with the leading contributions from virtual and soft real photons was given in the form required by RG approach eq(\[psii\]). The non–leading contributions, $\delta^{hard}$ from hard photon emission, include SD emission, IB of point–like mesons as well as the interference terms. It is free from infrared and mass singularities and given above (\[deltahard\]) with $$\mathcal{J}(x,y,z)=P_1 R_1 + P_2 y I^0_1 + P_3 \frac{y^2}{2} I^1_1 + \frac{y}{2} P_4 R_4 + \frac{xy^2}{4} P_5 R_5 \ ,$$ and $$\begin{aligned} \nonumber R_1 = \frac{y}{x+y} \left(y-x \right) \left((2-z)(x+y)-2R(z) \right)I^0_3-y^2((x+y) I^0_3-xI^1_3) \ , \\ \nonumber %\end{equation} % %\begin{multline} R_4 = \left( 2-z\right) I^0_1 - y I^1_1 + (2R(z)-(x+y)(2-z))((x+y)I^0_3 - yI^1_3)+ \\ xy((x+y)I^1_3-yI^2_3) \ ,\nonumber \\ \nonumber %\end{multline} % %\begin{multline} R_5 = \left(2-z \right) I^1_1 -yI^2_1 + (2R(z)-(x+y)(2-z))((x+y)I^1_3-yI^2_3)+ \\ \nonumber xy((x+y)I^2_3-yI^3_3) \ .\end{aligned}$$ Appendix C ========== The contribution to $\delta^{hard}$ from SD emission have the form: $$\delta_{SD}^{hard}= \frac{\alpha}{2 \pi a_0(y,z)} \int \limits_0^{b(z)} dx J^{SD}(x,y,z) \ ,$$ where $$J^{SD}(x,y,z) = Q_1 R_1 + y Q_2 I^0_1 + \frac{y^2}{2} Q_3 I_1^1 + \frac{y}{2} Q_4 R_4 + \frac{x y^2}{4} Q_5 R_5 \ ,$$ with $R_i$ given in Appendix B and $$\begin{aligned} Q_1=-\frac{1}{4}y\left(x+y\right) \ ,\nonumber \\ Q_2=\frac{1}{4} \left[2x(x+2y+z-2)+3y(y+z-2) \right] \ ,\nonumber \\ Q_3=-\frac{1}{8} \left[-8+(z+y)(4+3x)-2x+3x^2 \right] \ ,\nonumber \\ Q_4= \frac{1}{8} \left[ 4y+4x+3xy \right] , \qquad Q_5=-\frac{3}{4} \ .\end{aligned}$$ The contribution to the total width have a form: $$\delta^{SD}=\frac{\alpha}{2\pi}\frac{\int\int dy dz (1+\frac{\lambda_+}{r_\pi}R(z))^2\int\limits_0^N dx J^{SD}(x,y,z)}{\int\int dy dz a_0(y,z) (1+\frac{\lambda_+}{r_\pi}R(z))^2} \ .$$ Numerical estimation gives: $$\delta^{SD}=-0.005.$$ Appendix D ========== The function $\Psi$, defined as $$\Psi(y,z)=\int\limits_{b_-(z)}^{b(z)}\frac{dt}{t}a_0(t,z)P^{(1)})(\frac{y}{t}),$$ contains a restriction on the domain of integration, namely $t$ exceed $y$ or equal to it,which is implied by the kernel $P^{(1)}(y/t)$. Explicit calculations give: $$\begin{gathered} \Psi_<(y,z)=\int\limits_{b_-(z)}^{b(z)} \frac{dt}{t}a_0(t,z) \frac{y^2+t^2}{t(t-y)}=[R(z)-y(2-z)]\ln\frac{b(z)}{b_-(z)}+ \\ 2a_0(y,z)\ln\frac{b(z)-y}{b_-(z)-y}+ \frac{1}{2}(b(z)^2-b_-(z)^2) \ , \\ 2\sqrt{r_e}<y<1-\sqrt{r_\pi},\qquad \Psi_<(y,z)=0, y>1-\sqrt{r_\pi} \ .\end{gathered}$$ and $$\begin{gathered} \Psi_>(y,z)= \int\limits_y^{b(z)} \frac{dt}{t}a_0(t,z)P^{(1)}(\frac{y}{t})= a_0(y,z)[2\ln\frac{b(z)-y}{y}+\frac{3}{2}]- \\ \frac{1}{2}(b(z)^2-y^2)+ (b(z)-y)(2-y-z+b_-(z))+[R(z)-y(2-z)]\ln\frac{b(z)}{y}.\end{gathered}$$ One can convince the validity of the relations: $$j_0(y)=\int\limits_{2\sqrt{r_\pi}}^{c(y)}dz\Psi_<(y,z)+\int\limits_{c(y)}^{1+r_\pi}dz\Psi_>(y,z);$$ and $$\int\limits_0^{b_-(z)}dy\Psi_<(y,z)+\int\limits_{b_-(z)}^{b(z)}dy\Psi_>(y,z)=0.$$ The last relation demonstrates the KLN cancellation for the pion spectrum obtained by integration of the corrections over $y$ in the interval $0<y<b(z)$. The explicit expressions for $j_1(y)$ and $j_2(y)$ are:(for $j_0(y)$ see (\[fullI\])). $$\begin{gathered} j_1(y)=\frac{y^3(1-r_\pi-y)^3}{3(1-y)^2} \left(2 \ln \frac{1-r_\pi-y}{y} + \frac{3}{2} \right) + \\ \frac{r_\pi^2}{3(1-y)^2} \left[ 3(1-y)(1+y^2)+r_\pi(y^3+3y-2) \right] \ln \frac{1-y}{r_\pi} - \\ \frac{1-r_\pi-y}{36(1-y)^2} \left[ (1-y)^2 \left(43 y^3 -15 y^2 -3y-1 \right. +r_\pi(83y^2+26y+11) +3r_\pi^3 \right) + \\ r_\pi^2(31 y^3 -15y^2-39y+47) \left. \right] \ ,\end{gathered}$$ $$\begin{gathered} j_2(y)=\frac{y^4(1-r_\pi-y)^4}{12(1-y)^3} \left(2 \ln \frac{1-r_\pi-y}{y} + \frac{3}{2} \right) + \\ \frac{r_\pi^2}{12(1-y)^3} \ln \frac{1-y}{r_\pi} \biggl[ 6(1+y^2)(1-y)^2 -4 r_\pi(1-y)(2y^3-y^2+4y-3)+ \\ r_\pi^2(y^4+6y^2-8y+3) \biggr] + \frac{1-r_\pi-y}{720(1-y)^3} \biggl[-(1-y)^3(247y^4-88y^3-28y^2-8y-3)- \\ %\right. r_\pi(1-y)^3(733y^3+341y^2+129y+57)-\\ r_\pi^2(1-y)(707y^4-808y^3+212y^2- 408y+717)\\ +r_\pi^3(173y^4-72y^3-492y^2+1048y-477) -12 r_\pi^4(1-y)^3 \biggr].\end{gathered}$$ Appendix E ========== Collection of the relevant formulae. The Dalitz-plot distribution in the region $D$: $$\begin{gathered} \frac{1}{\mathcal{C}S_{EW}} \frac{d\Gamma}{dy dz}=\left(1+\lambda_+\frac{t}{m_\pi^2} \right)^2\biggl(a_0(y,z)+ \frac{\alpha}{\pi}[\frac{1}{2}(L_e-1) \Psi_>(y,z)+a_0(y,z)Z_1+\frac{1}{2}Z_2]\biggr),\\ Z_1=\frac{3}{4}-\frac{\pi^2}{6}-\frac{3}{2}\ln y- \ln (({b(z)-y})/y) - Li_2(1-y) \ .\end{gathered}$$ The function $Z_2$ is defined in (\[Z2\]). Correction to the total width (we include the contribution of the region outside the region $D$), $\Gamma=\Gamma_0(1+\delta)$: $$\begin{gathered} 1+\delta=S_{EW}+\frac{\alpha}{\pi} \frac{1}{\int \int dz dy a_0(y,z) \left(1+\frac{\lambda_+}{r_\pi}R(z) \right)^2} %\int\limits_{2\sqrt{r_\pi}}^{1+r_\pi}\phi_1(z)dz \biggl[ \int\limits_0^{1-r_\pi} I(y)\ln y dy+ \\ \int\limits_{2\sqrt{r_\pi}}^{1+r_\pi} dz (1+\frac{\lambda_+}{r_\pi}R(z))^2 [\int\limits_0^{b_-(z)}dy[-a_0(y,z) \ln\frac{{b(z)-y}}{b_-(z)-y}+(1/2)\tilde{Z}_2(y,z)] + \\ \int\limits_{b_-(z)}^{b(z)}dy[a_0(y,z)Z_1+(1/2)Z_2]\biggr] \ ,\end{gathered}$$ with $$\begin{gathered} % \tilde{Z}_2(y,z)=\phi_{2A}(y,z)-2\phi_{1A}(y,z)+\int_{b_-(z)-y}^{b(z)-y} dx \mathcal{J}(x,y,z) ; \\ % \phi_{1A}(y,z)=R(z)\ln\frac{b(z)}{b_-(z)}-y(b(z)-b_-(z)) \\ % \phi_{2A}(y,z)=\int_{b_-(z)-y}^{b(z)-y}\frac{ dx x}{(x+y)^2}a_0(x+y,z)= \\ \tilde{Z}_2(y,z)=Rphot_{2A}(y,z)-2Rphot_{1A}(y,z)+\int_{b_-(z)-y}^{b(z)-y} dx \mathcal{J}(x,y,z) ; \\ Rphot_{1A}(y,z)=R(z)\ln\frac{b(z)}{b_-(z)}-y(b(z)-b_-(z)) \\ Rphot_{2A}(y,z)=\int_{b_-(z)-y}^{b(z)-y}\frac{ dx x}{(x+y)^2}a_0(x+y,z)= \\ (b(z)-b_-(z))(1-\frac{z}{2}+2y)-(y(2-z)+R(z))\ln\frac{b(z)}{b_-(z)}.\end{gathered}$$ The expression in big square brackets in right-hand side of (95) can be put in the form: $$\begin{gathered} \int\limits_{2\sqrt{r_\pi}}^{1+r_\pi}\phi_1(z)dz=\int\limits_{2\sqrt{r_e}}^{1-r_\pi}f_1(y)dy=-0.035\end{gathered}$$ which results in $\delta=0.02$. For the aim of comparison with E. Ginsberg result we must put here $$\lambda_+=0, \qquad I(y)=j_0(y), \qquad M_W=M_p,$$ as was mentioned above we have reasonable agreement with E. Ginsberg results. For the inclusive set-up of experiment (energy fraction of positron is not measured) we have for pion energy spectrum given above (\[rcpions\]). When we restrict ourselves only by the region $D$ the spectrum becomes dependent on $\ln(1/r_e)$: $$\begin{gathered} \frac{1}{\mathcal{C} S_{EW}}\frac{d\Gamma}{dz}= \left\{ \phi_0(z)+ \right. \frac{\alpha}{\pi}[(1/2)P(z)(\ln(1/r_e)-1)+ \\ \int\limits_{b_-(z)}^{b(z)}dy[\Psi_>(y,z)\ln y+a_0(y,z)Z_1+ \left. \frac{1}{2}Z_2]] \right\} \left(1+\frac{\lambda_+}{r_\pi}R(z) \right)^2 \ ,\end{gathered}$$ with $$\begin{gathered} P(z)=\frac{1}{6}b_-(z)^2(3b(z)+b_-(z))\ln\frac{b(z)}{b_-(z)}+ \frac{1}{3}(b(z)-b_-(z))^3\ln\frac{b(z)-b_-(z)}{b(z)}- \\ \frac{1}{6}b_-(z)(b(z)-b_-(z))(3b_-(z)+b(z)).\end{gathered}$$ Appendix F ========== Our approach to study the radiative kaon decay has an advantage compared to the one used by E. Ginsberg – it has a simple interpretation of electron mass singularities based on Drell-Yan picture. The [@G] approach to study noncollinear kinematics is more transparent than ours one. We remind the reader of some some topics of [@G] paper. One can introduce the missing mass square variable $$l=(p_\nu+k)^2/M^2=A_\nu=(M-E_\pi-E_e)^2/M^2-(\vec{p}_\pi+\vec{p}_e)^2/M^2.$$ the limits of this quantity variation at fixed $y,z$ are pu by the last term: for collinear or anticollinear kinematics of pion and positron 3-momenta. Being expressed in terms of $y,z$ they are (we consider the general point of Dalitz-plot and omit positron mass dependence): $$0<l<b_-(z)(b(z)-y),$$ for the $y,z$ in the $D$ region and $$b(z)(b_-(z)-y)<l<b_-(z)(b(z)-y),$$ for the case when they are in the region $A$ outside $D$: $$0<y<b_-(z),2\sqrt{r_\pi}<z<1+r_\pi.$$ For our approach with separating the case of soft and hard photon emission we must modify the lower bound for $l$ in the region $D$. It can be done using another representation of $l$: $$l=x[1-(y/2)(1-C_e)-(z/2)(1-\beta C_\pi)],$$ with $C_e,C_\pi$-the cosine of the angles between photon 3-momentum and positron and pion ones, $\beta=\sqrt{1-m^2/E_\pi^2}$-is the pion velocity. Maximum of this quantity is $b(z)$. Taking this into account we obtain for the region oh hard photon $$x>2\Delta\varepsilon/M=y\Delta,\Delta=\Delta\varepsilon/E_e<<1,$$ for the region $D$: $$\begin{gathered} y\Delta<x<b(z)-y,\qquad yb(z)\Delta<l<b_-(z)(b(z)-y);\end{gathered}$$ and for region $A$: $$\begin{gathered} b_-(z)-y<x<{b(z)-y},\qquad b(z)(b_-(z)-y)<l<b_-(z)(b(z)-y).\end{gathered}$$ In particular for the collinear case we must choice $C_e=1;C_\pi=-1$, which corresponds to $x+y<b(z)$. Let infer this condition using the NMS condition : $$\begin{gathered} (P_k-p_e-p_\pi-k)^2/M^2=R(z)-x-y+(xy/2)(1-C_e)+\\ (xz/2)(1-\beta C_\pi)+ (yz/2)(1-\beta C_{e\pi})=0.\end{gathered}$$ In collinear case we have $C_e=1;C_\pi=C_{e\pi}$.From NMS condition we obtain $1-\beta C_\pi=(2/z(x+y))(x+y-R(z))$. Using this value we obtain $l_{coll}=R(z)x/(x+y)$. Using further the relation $R(z)=b(z)b_-(z)$ we obtain again $x<b(z)-y$ in the case of emission along positron. Comparing the phase volumes in general case calculated in our approach with using NMS condition with [@G] approach we obtain the relation: $$\int x dx\int\frac{dO_\gamma}{4\pi}=\int dl\int d\gamma, \int d\gamma=\int\frac{d^3 p_\nu}{E_\nu}\frac{d^3k}{k_0} \frac{\delta^4(P-p_\nu-k)}{2\pi}.$$ The non–leading contribution arising from hard photon emission considered above: $$I_{IB}=\int\frac{dx}{x}\int\frac{dO_\gamma}{4\pi}\mathcal{P}_{IB},$$ with $$\begin{gathered} \mathcal{P}_{IB}=xG_1\frac{\tilde{A}_\nu}{A_e}+xG_2+G_3A_e+G_4A_\nu+ G_5A_eA_\nu, \\ G_1=\frac{y}{4}(2-y-x),\\G_2=\frac{R(z)}{2(x+y)}+\frac{x^2}{2}+\frac{1}{2}x(z+2y)+\frac{1}{4} (2z+3y(y+z))-1;\\ G_3=-\frac{1}{8}x^2-\frac{1}{8}x(2+z+y)-\frac{1}{2}(y+z); \qquad G_4=\frac{1}{8}x(4+y)+\frac{1}{8}y-1;\\ G_5=-\frac{1}{4},\end{gathered}$$ (note that $G_i+Q_i=P_i$, see appendix B and C) can be transformed to the form: $$\begin{gathered} I_{IB}=(1/4)\int dl[4-2y-4z-(1/4)R(z)+(1/4)l+y\ln\frac{(R(z)-l)^2}{l}- \\ 2\ln\frac{y^2R(z)^2}{l(l+y(2-z))}+[z+(3/2)y(y+z)-2+(1/4)l(4+y)]I_{10}- \\ (1/2)I_{1-1}-((1/2)l+y+z)I_{2-1}+I_z].\end{gathered}$$ Here we use the list of integrals obtained in the paper of [@G]: $$\begin{aligned} I_{mn}=\int d\gamma\frac{1}{(kP_K/M^2)^m(kp_e/M^2)^n}; \\ \nonumber I_{10}=\frac{2}{s}\ln\frac{2-y-z+s}{2-y-z-s}; \qquad I_{20}=4/l; I_{00}=1;\\ \nonumber I_{-1,0}=(2-y-z)/2;\qquad I_{11}=\frac{4}{yl}\ln\frac{y^2}{l};\\ \nonumber I_{01}=\frac{2}{R(z)-l}\ln\frac{(R(z)-l)^2}{lr_e};\\ \nonumber I_{1-1}=\frac{R(z)(2-y-z)-(2+y-z)l}{s^2}+ \\ \nonumber \frac{2l(y(2-y-z)-2R(z)+2l)}{s^3}\ln\frac{2-y-z+s}{2-y-z-s};\\ \nonumber I_{2-1}=\frac{2(y(2-y-z)+2l-2R(z))}{s^2}+ \\ \nonumber \frac{R(z)(2-y-z)-(2+y-z)l}{s^3}\ln\frac{2-y-z+s}{2-y-z-s},\\ \nonumber s=\sqrt{(2-y-z)^2-4l}.\end{aligned}$$ Besides we need two additional ones: $$\begin{aligned} I_e=\int d\gamma\frac{1}{(kp_e/M^2)(2(kP_K/M^2)+y}=\frac{2}{yR(z)}\ln \frac{y^2R(z)^2}{l(l+y(2-z))r_e}; \\ \nonumber I_z=\int d\gamma\frac{1}{(kP_K/M^2)(2(kP_K/M^2)+y}=\frac{4}{ys}\ln \frac{2l+ys+y(2-y-z)}{2l+ys-y(2-y-z)}.\end{aligned}$$ One can see the cancellation of mass singularities (terms containing $\ln(1/r_e)$) in the expression $I_{IB}$. Numerical calculations in agreement (within few percent) of this and the given above expressions. [99]{} T.Becherrawy, Phys. Rev. D1 (1970), 1452 E.Ginsberg, Phys. Rev. 162(1967), 1570; Phys. Rev. 142(1966), 1035; Phys. Rev. 171(1968), 1675; Phys. Rev. 187(1968), 2280 J.Gasser and H.Leutwyller, Nucl. Phys. B250(1985), 464, 517; J.Bijnens, G.Colangelo, G.Ecker, and J.Gasser, The Second DA$\Phi$NE Physics Handbook, vol.1, 313 J.Bijnens, G.Ecker, and J.Gasser, Nucl. Phys. B396(1993), 81; B.Holstein Phys. Rev. D41(1990), 2829; A.Sirlin, Phys. Rev. D5(1972), 436; Rev. Mod. Phys. 50(1978), 573; Nucl. Phys. B196(1982), 83; M.V.Terent’ev, Yad. Fiz. 18(1973), 870; Sov. J. Nucl. Phys. 18(1974), 449 E.A.Kuraev, V.S.Fadin, Yad. Fiz. 41(1985), 466 V.N.Baier, V.S.Fadin, V.A.Khoze, Nucl. Phys. B65(1973), 381 W.Marciano and A.Sirlin, Phys. Rev. Lett. 71(1993), 3629 T.Kinoshita, J. Math. Phys. 3(1972), 650;\ T.D.Lee and M.Nauenberg, Phys. Rev. B133(1964), 1549 R.Decker and M.Finkemeier, Nucl. Phys. B438(1995), 17 V.Cirigliano, M.Knecht, H.Neufeld, H.Rupertsbergen, P.Talavera,\ hep-ph/0110153 N-P. Chang, Phys. Rev. 131(1963),1272 A.I.Akhiezer, V.B.Berestetski, ‘Quantum Electrodynamics‘, Moscow, 1981;\ V.B.Berestetski, E.Lifshitz, L.Pitaevski,‘Quantum Electrodynamics‘, Moscow, 1989 E.S. Ginsberg, Phys. Rev. D 1, 229 (1970) [^1]: the formula (10) in [@BFK] should read $$d\Gamma_b = \frac{2\epsilon' d^3 \sigma_{0b}}{d^3 p'} \biggm|_{\vec{p}'=\vec{p}_3+\vec{k}} dW_{\vec{p}_3+\vec{k}} \ (k) \ \frac{d^3p_3}{2\epsilon_3 \ .}$$

ArXiv

--- author: - 'L. Marion, O. Absil[^1], S. Ertel, J.-B. Le Bouquin, J.-C. Augereau, N. Blind, D. Defrère, J. Lebreton, J. Milli' bibliography: - 'PIONIER\_companions\_rev.bib' date: 'Received xxx; accepted xxx' subtitle: 'II. 92 main sequence stars from the Exozodi survey[^2]' title: 'Searching for faint companions with VLTI/PIONIER' --- [The <span style="font-variant:small-caps;">Exozodi</span> survey aims to determine the occurrence rate of bright exozodiacal discs around nearby main sequence stars using infrared interferometry. Although the <span style="font-variant:small-caps;">Exozodi</span> survey targets have been carefully selected to avoid the presence of binary stars, the results of this survey can still be biased by the presence of unidentified stellar companions.]{} [Using the PIONIER data set collected within the <span style="font-variant:small-caps;">Exozodi</span> survey in 2012, we aim to search for the signature of point-like companions around the <span style="font-variant:small-caps;">Exozodi</span> target stars.]{} [We make use of both the closure phases and squared visibilities collected by PIONIER to search for companions within the $\sim$100 mas interferometric field of view. The presence of a companion is assessed by computing the goodness of fit to the data for a series of binary models with various separations and contrasts.]{} [Five stellar companions are resolved for the first time around five A-type stars: HD 4150, HD 16555, HD 29388, HD 202730, and HD 224392 (although the companion to HD 16555 was independently resolved by speckle interferometry while we were carrying out the survey). In the most likely case of main sequence companions, their spectral types range from A5V to K4V. Three of these stars were already suspected to be binaries from Hipparcos astrometric measurements, although no information was available on the companions themselves so far. In addition to debiasing the statistics of the <span style="font-variant:small-caps;">Exozodi</span> survey, these results can also be used to revise the fraction of visual binaries among A-type stars, suggesting that an extra $\sim$13% A-type stars are visual binaries in addition to the ones detected in previous direct imaging surveys.]{} [We estimate that about half the population of nearby A-type stars could be resolved as visual binaries using a combination of state-of-the-art interferometry and single-aperture imaging, and we suggest that a significant fraction of these binaries remains undetected to date.]{} Introduction ============ Historically, binary stars have been detected and studied in two main ways: (i) using the orbital motion around the centre of mass that can be measured with spectroscopy (or possibly with astrometry), or (ii) using direct (visual) observations. These two classes of techniques, which can be declined in various flavours, have discovery spaces that may or may not overlap depending on the considered target. In particular, the longest periods reachable by spectroscopy frequently correspond to angular separations too short to be reached with classical imaging observations. This can lead to significant incompleteness when taking a census of binary stars. High angular resolution observations, such as those provided by infrared interferometry, may help bridge the gap between spectroscopy and classical imaging techniques. Here, we aim to search for unknown companions around a specific sample of main sequence stars that was built in the context of an unbiased interferometric search for hot debris discs (also known as exozodiacal discs). The main purpose of this survey, presented in a companion paper [@Ertel14][^3], is to obtain statistically significant information on the occurrence of bright exozodiacal discs around nearby main sequence stars. Binary stars were removed from this sample to prevent contamination of the interferometric observations with spurious circumstellar emission, which would make faint exozodiacal disc detection much more difficult. This selection criterion also avoids gravitational interactions with nearby, massive bodies, which may induce biases by enhancing or disrupting the dusty discs around the target stars. The selection of single stars was based on the literature, which, as explained above, may not be complete regarding companions – especially faint and close ones – around these stars. To make sure that the stars observed within the so-called <span style="font-variant:small-caps;">Exozodi</span> survey, carried out in the southern hemisphere with the Precision Integrated Optics Near Infared Experiment [PIONIER, @LeBouquin11] visitor instrument at the VLTI, are indeed single stars, we use the full information delivered by PIONIER (including closure phases) to systematically search for companions around all the stars observed within the survey. Doing so, we will be able to determine whether the small near-infrared excess detected around several stars within the survey are effectively due to an extended, mostly symmetric source, or to a point-like companion. This is of utmost importance in order to derive unbiased statistics on the occurrence rate of exozodiacal discs. After detailing the stellar sample, the observations and data reduction in Sect. \[sec:obsandred\], we present in Sect. \[sec:searchcomp\] the search method and the detection criterion. The results in terms of newly detected companions are given in that same section. A discussion of each new binary system is given in Sect. \[sec:properties\]. We finally discuss the implications of this study in terms of revised binary fractions and the PIONIER sensitivity to faint companions in Sect. \[sec:discussion\]. Observations and data reduction {#sec:obsandred} =============================== The choice of the <span style="font-variant:small-caps;">Exozodi</span> stellar sample was driven by the goal to assess possible correlations between the presence of cold and hot dust around main sequence stars [@Ertel14]. The targets were selected from a magnitude limited sample of stars known to harbour cold dust in their outer planetary system. An equivalent sample of stars without any detectable cold dust emission (to within the sensitivity limit of far-infrared space-based observatories) was then selected, resulting in an all-sky sample of more than 200 stars. Because of the selection process, the overall sample is biased towards dusty stars (which are less frequent than non-dusty stars in our Galaxy at the sensitivity level of current far-infrared space observatories), although each of the two sub-samples is designed to be free from bias. The stars are more or less equally spread between three spectral type categories: A-type, F-type and G-K-type. All stars have been thoroughly checked for the presence of known (sub-)stellar companions in the literature; any star with a companion more massive than the planet-brown dwarf boundary and at a separation smaller than $5\arcsec$ was discarded. A more detailed description of the sample can be found in @Ertel14. Ninety-two southern stars from the <span style="font-variant:small-caps;">Exozodi</span> sample were observed with the PIONIER visitor instrument operated on the 1.8-m Auxiliary Telescopes (AT) of the VLTI during four observing campaigns of three nights each from April to December 2012. The observations were conducted in H band on the most compact AT configuration (A1–B2–C1–D0), using the *SMALL* spectral dispersion that splits the light into three spectral channels on the detector. Three consecutive observing blocks (OBs) were obtained for most stars, and care was taken to observe as many stars from each of the dusty and non-dusty categories. The optical path length (OPD) scans ranged from $60~\mu$m to $180~\mu$m depending on the baseline, and 1024 samples were read per scan to ensure a proper sampling of the fringes. Only the A and C outputs of the ABCD fringe coding scheme were read to speed up the readout. We used the same data reduction scheme as presented in @Ertel14, based on the `pndrs` pipeline and including the specific calibration strategy developed for the <span style="font-variant:small-caps;">Exozodi</span> survey to optimise the accuracy of the squared visibilities. As a sanity check, we also carried out our study based on a more standard calibration scheme, where the transfer function is averaged over each individual CAL-SCI-CAL-SCI-CAL-SCI-CAL sequence with a weight inversely related to the time separation, as described in @LeBouquin11. The final results did not show any significant change. Searching for companions {#sec:searchcomp} ======================== Principle of the search {#sub:principle} ----------------------- To detect the presence of a companion, we use the closure phases and the squared visibilities in a combined way. Before explaining the details of the method used here, we will summarise the principle of the method used by @Absil11, which was also used as a first step in our analysis. The search for companions in @Absil11 was based only on the closure phases. Indeed, the closure phase (CP) is known to be free of instrumental and atmospheric errors to the first order [@LeBouquin12], and to be sensitive to asymmetries. For a centro-symmetric source, the CP will be strictly zero (or possibly a constant offset from zero due to imperfect CP calibration), while the presence of an off-axis companion will show up as a non-zero closure phase, which varies as a function of time and wavelength. @Absil11 define a field of view and compute the CP associated with a binary model considering the primary star at the centre of the search region with an off-axis companion of various contrast $r$ starting at $r=0$, at each point $(x,y)$ of the field of view in the search region. The modelled CP is then compared to the observations to derive the $\chi^2$ goodness of fit for each binary model, $$\chi^2_{\rm CP} = \sum\frac{({\rm CP}_{mod} - {\rm CP}_{data})^2}{\sigma^2_{\rm CP}} \; ,$$ thereby creating a $\chi^2$ cube. The cube is then normalised (by division) so that its minimum equals 1, and collapsed along the contrast axis to keep only the best-fitting companion contrast (i.e. minimum $\chi^2$ value) at each position in the search region. The resulting $\chi^2$ map can then be used to derive the probability for the single-star model to adequately represent the data, based on the $\chi^2$ distribution under Gaussian noise assumption [see @Absil11 for details]. If this probability is below a pre-defined threshold, the single-star model can be rejected and the best-fit binary solution is then considered as statistically significant. In @Absil11, the threshold is fixed at a $3\sigma$ level, i.e. at a probability of $0.27\%$ under the Gaussian noise assumption. In the present study, we noticed that the use of the sole CP leads to a large number of barely significant detections, with significance levels (i.e. signal-to-noise ratios) between 3 and 4$\sigma$. It was not obvious to discriminate whether these were true detections or false positives. As a consequence, in addition to the CP, we use the squared visibilities ($V^2$), which are expected to show deviations from the single-star model if a companion is present. As in @Absil11, we compute a binary model considering the primary star at the centre of the search region with an off-axis companion of various contrast $r$ (starting at $r<0$ as explained in Sect. \[sub:detecriteria\]) at each point $(x,y)$ of the search region. In the present case, we can safely assume that both the primary and the secondary stars are unresolved, since most of our targets have an angular diameter smaller than 1 mas [@Ertel14] and as the observations are carried out on the most compact AT configuration, which provides an angular resolution of the order of 5 mas. Then, we compute the CP and the $V^2$ for each model and derive a combined goodness of fit, $$\chi^2 = \underbrace{\sum\frac{(V^2_{mod} - V^2_{data})^2}{\sigma^2_{V^2}}}_{\chi^2_{V^2}} + \underbrace{\sum\frac{({\rm CP}_{mod} - {\rm CP}_{data})^2}{\sigma^2_{\rm CP}}}_{\chi^2_{\rm CP}} \; .$$ Here again, this is done for each binary model to obtain a $\chi^2(x,y,r)$ cube. The same procedure as in @Absil11 is then repeated to assess the presence of a companion. This new method is more robust as we need to have a signature in both the CP and $V^2$ to detect a companion. We note however that the $V^2$ is also sensitive to centro-symmetric circumstellar emission, which creates a drop in visibility at all baselines [@DiFolco07]. Sometimes, the signature of a disc can be so strong that the combined $\chi^2$ is relatively high. To discriminate this kind of situation and identify bona fide point-like companions, we inspect the $\chi^2$ maps individually for the CP and the $V^2$. Indeed, if a companion is present, its signature will be seen in both maps (generally at the same position, except for marginal detections), while a disc will only show up in the $\chi^2_{V^2}$ map (assuming a symmetric disc). Furthermore, a small offset in the CP due to imperfect calibration could simulate a companion (false positive detection). In this case, the detection will generally not show up in the $\chi^2_{V^2}$ map. In some cases, the $\chi^2$ maps for both the CP and the $V^2$ show a significant detection, but not at the same position. These cases need to be investigated more carefully (e.g. by looking at secondary peaks in the $\chi^2$ maps) to draw definitive conclusions, when possible. Defining the search region -------------------------- ----------- ------------- -------------- -------------- -------------- ----------------------------- Name Date Significance Significance Significance Nature (CP+$V^2$) (CP only) ($V^2$ only) HD 2262 2012 Oct 16 3.9 2.9 5.0 disc HD 4150 2012 Dec 17 7.7 3.7 7.9 companion HD 15008 2012 Jul 24 4.9 2.3 5.4 disc HD 15798 2012 Oct 16 6.3 2.0 14.0 (see Sect. \[sub:results\]) HD 16555 2012 Dec 18 105.3 26.1 212.7 companion HD 20794 2012 Oct 15 2.2 4.5 2.1 disc $\cdots$ 2012 Dec 17 4.0 1.8 6.0 $\cdots$ HD 23249 2012 Oct 15 2.9 2.7 5.9 disc $\cdots$ 2012 Dec 16 4.8 5.0 7.8 $\cdots$ HD 28355 2012 Dec 15 5.2 2.4 7.9 disc HD 29388 2012 Dec 16 111.7 53.6 130.1 companion HD 39060 2012 Oct 16 5.8 2.3 7.8 disc HD 108767 2012 Apr 30 3.1 1.7 3.6 disc HD 202730 2012 Jul 24 12.2 6.7 21.3 companion HD 224392 2012 Jul 26 15.1 20.4 7.7 companion ----------- ------------- -------------- -------------- -------------- ----------------------------- There are two main limitations to the detection of faint companions with interferometry: the dynamic range in the observations, and the limited field of view. Before starting our search for companions, we first need to define a suitable search region. This region is limited, as explained in @Absil11, by three factors. The first relates to the beams being injected into single-mode fibres for modal filtering and beam combination. As explained in @Absil11, we can consider that the PIONIER fibres have a Gaussian transmission profile with a full width at half maximum of 420 mas under typical turbulence conditions at Paranal. Second, the fringe packets associated with the primary star and its companion need to be within the OPD scan length and to partially overlap in order to contribute to the coherent signal measured by the PIONIER data reduction pipeline. The shortest scans performed by PIONIER are about $60~\mu$m long, and we estimate the size of one fringe packet to be $\lambda^2/\Delta\lambda \simeq 27\mu$m when the H band signal is dispersed onto three spectral channels. The maximum OPD separation ($\Delta$OPD) between the two packets can therefore not exceed $27~\mu$m. This separation is related to the angular separation $\Delta\theta$ of the two objects in the sky: $\Delta {\rm OPD} = B \Delta\theta \cos \alpha$, with $B \cos \alpha$ the projected baseline. Considering a maximum baseline $B = 36$ m, we find a maximum angular separation $\Delta\theta_{\rm max} \simeq 150$ mas. The last limitation to the field of view comes from the need to properly sample the closure phase variations as a function of wavelength: companions located too far away can create aliasing inside the search region if their closure phase signal has a period shorter than four times the spectral channel size (for Nyquist sampling). This phenomenon can be partly mitigated by repeating the observations in time. Here, we only have three consecutive OBs for most of our targets. We will therefore compute a pessimistic non-ambiguous field of view as if we only had one sample in time for each star. As already detailed by @Absil11, the period in the closure phase signal is roughly given by[^4] $P_{\lambda} \simeq \lambda^2/(B\Delta\theta -\lambda)$, and must be larger than four times the spectral channel size ($0.1~\mu$m here). This lead to $\Delta\theta_{\rm max} \simeq 87$ mas for our mean baseline $B \simeq 20$ m. We will therefore consider a search region of about 100 mas in this study. It must be noted that companions located up to 150 mas can nonetheless have a significant signature in our data. Defining the detection criterion {#sub:detecriteria} -------------------------------- ![Statistics of the (signed) significance level for the 92 stars based on the combined $\chi^2$, taking into account the CP and $V^2$. Five stars with a significance level higher than $10\sigma$ are not represented here for the sake of clarity.[]{data-label="fig:histosignif"}](histosignleveltotneg.eps)   \   \   \   \    To determine whether or not a companion is detected, we use, as explained in Sect. \[sub:principle\], the significance level of the detection, which can be associated with a given confidence level if the underlying probability distribution function is known. The threshold above which a detection will be identified must however be chosen with care. A 3$\sigma$ significance level was used by @Absil11, following common practice in presence of Gaussian noise. This choice is however not straightforward, and can be backed up by studying the noise properties of the data set. This can conveniently be done by including negative contrasts for the off-axis companions in the definition of our combined $\chi^2$ cube. While non physical, negative companions can be used to represent positive fluctuations in the $V^2$ (i.e. situations where the measured $V^2$ is higher than the expected $V^2$ from the photosphere). Negative companions do not significantly affect the CP, which can take both positive and negative values. A negative companion would indeed produce the same CP signature as a positive companion located on the opposite side of the star. In the following, we associate negative significance levels with negative companions. The histogram of the significance levels for our complete sample (listed in Table \[tab:all\]) is illustrated in Fig. \[fig:histosignif\], where the range of the plot has been limited to $[-10,10]$ for the sake of clarity. The negative part of the histogram does not contain any astrophysical signal, and can therefore be used as a reference to study the noise properties of our sample. The absence of significance levels close to 0$\sigma$ in the histogram can be explained by the fact that, in presence of noise and because of the limited number of observations, it is always possible to obtain a better fit to our data sets by inserting a companion somewhere in the field of view rather than using the single-star model (a significance level of 0 would only happen if the data were in perfect agreement with the single-star model). ----------- ---------- ------------- ---------- -- ---------- --------------- -- ---------------- ----------------- --------------- Name R.A. Dec. Distance Spectral $H$ Separation P.A. Contrast (pc) type (mag) (mas) (deg) (%) HD 4150 00:43:21 $-$57:27:47 75.5 A0IV $4.35\pm0.08$ $90.5 \pm 2.2$ $84.0\pm 2.2$ $2.3 \pm 0.4$ HD 16555 02:37:24 $-$52:32:35 45.6 A6V $4.60\pm0.02$ $78.7 \pm 1.6$ $40.9\pm 0.3$ $51 \pm 4$ HD 29388 04:38:09 $+$12:30:39 47.1 A6V $3.94\pm0.01$ $11.1 \pm 0.2$ $71.6 \pm 0.05$ $3.1 \pm 0.2$ HD 202730 21:19:51 $-$53:26:57 30.3 A5V $4.22\pm0.07$ $61.7 \pm 1.2$ $-21.4 \pm 0.8$ $87 \pm 14$ ----------- ---------- ------------- ---------- -- ---------- --------------- -- ---------------- ----------------- --------------- Out of the 92 stars in our sample, 38 show a negative significance level, while 54 show a positive significance level (see Fig. \[fig:histosignif\] and Table \[tab:all\]). This suggests that around 16 stars in our sample have circumstellar emission emanating from a disc or a companion. This is consistent with the nine hot exozodiacal discs identified by @Ertel14 and the six companions described in the next section. From the 38 stars with a negative significance level, only three are located below $-3 \sigma$. While the distribution of the significance level looks far from Gaussian in Fig. \[fig:histosignif\], we decide to use $3 \sigma$ as our detection criterion. From the negative part of the histogram, we expect that three of the thirteen stars showing a significance level higher than $3 \sigma$ could be false positive detections. All thirteen stars will be carefully inspected to assess the possible presence of such false positives. Results of the search {#sub:results} --------------------- Table \[tab:detections\] lists all the stars that have a significance level higher than $3\sigma$ for the combined $\chi^2$ analysis. As already mentioned in Sect. \[sub:principle\], by looking separately at the $\chi^2$ maps for the CP and $V^2$, we can readily discriminate between a companion and a disc. This is done in Table \[tab:detections\], where the significance levels are given based on the analysis of the CP and of the $V^2$ separately. The eight stars having a significance level below $3\sigma$ in the analysis of the CP, while showing a significant detection in the $V^2$ analysis, are understood to be surrounded by a circumstellar disc rather than a companion. Seven of these stars are indeed among nine stars identified by @Ertel14 as having a bright exozodiacal disc. A special note must be added to HD 15798, which was very recently shown by @Tokovinin14 to host a faint companion at an angular distance of $0\farcs21$, thanks to speckle interferometry observations. This companion, located outside the PIONIER search region, could not have been detected with the search method used here, as the fringe packets of the two stars are separated on most baselines (or even not within the OPD scanning range on some baselines). Its contribution is therefore considered mostly as incoherent light in the PIONIER data reduction pipeline. With an I-band contrast of four magnitudes, this companion is most probably at the origin of the $V^2$ drop observed by PIONIER for HD 15798. It is difficult to derive the H-band contrast of this companion from our observations, because of the way its contribution to the coherent flux is handled by the PIONIER pipeline, and because it is significantly affected by the off-axis transmission profile of the PIONIER instrument. The five remaining stars (HD 4150, HD 16555, HD 29388, HD 202730 and HD 224392) are identified as having an off-axis companion. This classification is backed up by a careful inspection of the $\chi^2$ maps separately for the CP and $V^2$: to validate the detection of an off-axis companion, the two maps should have their minima (or sufficiently deep local minima) at the same position. This is the case for four of our five newly identified binary stars (see Fig. \[fig:chi2maps\]). Only HD 224392 shows a somewhat suspicious behaviour, where the various $\chi^2$ maps do not show local minima at the same position (see bottom row of Fig. \[fig:chi2maps\]). This star will be the subject of a forthcoming study (S. Borgniet et al., in prep), where the binarity of the source is confirmed by further observations. We will therefore not make further comments on this star in the present manuscript. Two additional stars show a potential signature of companion: HD 20794 and HD 23249. They both have a significance higher than $3\sigma$ in their combined $\chi^2$ map for one of two epochs. A closer inspection of the individual $\chi^2$ maps reveals that the positions of the minima do not match, which indicate that the excess emission is most probably due to a circumstellar disc rather than a point-like companion.    A summary of all the newly found companions (excluding HD 224392) can be found in Table \[tab:companions\], where we also give the contrast and position of the detected companions. Surprisingly, two of them show a best-fit contrast of 0.5 or larger. The reason why such bright companions had not been revealed to date by other observations will be discussed in Sects. \[sec:properties\] and \[sec:discussion\]. Properties of the detected companions {#sec:properties} ===================================== Here, we describe the properties of the newly found companions and of their host stars, and we give a rough estimation of the orbital period. It is indeed most probable that the detected companions are gravitationally bound to their host star. The probability of finding such bright background (or foreground) objects inside the tiny PIONIER field of view can be computed by evaluating the density of stars of magnitude inferior or equal to the detected companions in the immediate neighbourhood of the primary stars. We estimate this density by querying the 2MASS catalogue [@Skrutskie06] within a radius of $1\degr$ around the four stars: the probability of finding an unrelated background object as bright as the detected companions within the PIONIER field of view is never larger than a few $10^{-8}$. HD 4150 ------- Also known as eta Phe, HD 4150 (HIP 3405, HR 191) is an A0IV star surrounded by a cold debris disc [@Su06]. It belongs to a visual double system according to the WDS [@Mason01] and the CCDM [@Dommanget02] catalogues. The previously detected visual companion is faint [$V=11.5$, @Dommanget00] and located at about $20\arcsec$ from the primary. It is therefore not expected to affect our interferometric measurements in any way. In addition, HD 4150 is classified as an astrometric binary based on the comparison of the Tycho-2 and Hipparcos astrometric catalogues [@Makarov05; @Frankowski07], which show discrepant proper motions. The star also shows an acceleration of $-6.41\pm0.99$ mas/yr$^2$ in right ascension and $1.59 \pm 0.98$ mas/yr$^2$ in declination as measured by Hipparcos [@Makarov05]. No orbital solution is given, however, and the nature of the companion is unknown. Two important properties can be estimated from the contrast and angular separation found in Sect. \[sub:results\]: the spectral type and the orbital period. We derive the spectral type by assuming a main sequence companion of lower mass than the primary. This is most likely, as a more massive, post-main sequence companion would be brighter than the primary, while a white dwarf would be much fainter and would remain undetected. The H-band flux ratio amounts to $0.023 \pm 0.004$, which corresponds to $\Delta H = 4.1 \pm 0.2$. Taking the magnitude and distance of HD 4150 into account, the companion has an absolute magnitude $M_H=4.1 \pm 0.2$, which corresponds roughly to a K1V star according to @Allen00. A lower bound can then be given on the period of the orbit, assuming a face-on system. The measured angular separation corresponds to a linear separation of 6.8 AU. Using Kepler’s third law and an estimated mass of 2.82 $M_{\odot}$ for the primary [@Gerbaldi99], we derive a minimum period of about 9.5 years. This estimated period is consistent with an astrometric detection of the companion [@Makarov05]. It is therefore likely that the companion detected within the present study is the source of the astrometric signature. To monitor the orbital movement of the companion, a second observation of the system was obtained on 9 August 2013 with PIONIER operated at K band on the same AT configuration (programme ID 091.C-0597). The companion is recovered with a contrast of $0.041\pm0.004$ at an angular separation of $96.8\pm2.4$ mas and position angle of $-99.2\degr \pm 1.1\degr$ (see Fig. \[fig:HD4150\]). The increase in the angular separation reveals that the orbit is not face on. Nevertheless, assuming a face-on orbit and a minimum period of 9.5 years, we estimate that the companion should have moved by about $24\degr$ on its orbit. This is of the same order as the difference in position angle between the two epoch ($15\fdg2$). We conclude that the companion is most probably on a slightly inclined and/or eccentric orbit with a period slightly longer than 10 years. We note that the measured K-band contrast points towards a slightly more massive companion, with a spectral type around G5V. More observations will need to be performed to nail down the orbital parameters and spectral type of the companion. HD 16555 -------- Eta Hor (HD 16555, HIP 12225, HR 778) is an A6V star classified as a stochastic astrometric binary by @Frankowski07. Despite the stochastic character of the astrometry (i.e. the impossibility of finding a suitable orbital solution based on the Hipparcos astrometry), @Goldin07 propose an orbital solution with a semi-major axis of $23.1^{+1.6}_{-1.0}$ mas and period of 3.0 yr. No information can however be derived on the nature of the companion. Later on, eta Hor was observed with adaptive optics, showing only a possible (yet unconfirmed) faint companion at $4\farcs9$ [@Ehrenreich10], which is not expected to affect our interferometric observations in any way. More recently, the star was resolved as a binary by speckle interferometry [@Hartkopf12], showing the presence of a bright companion ($\Delta I=0.7$ mag, $\Delta y=1.4$ mag) at an angular separation of 70.3 mas and a position angle of $60\fdg4$ (Besselian epoch of observation: 2011.0393 yr). This publication came out after we finalised the sample selection for the <span style="font-variant:small-caps;">Exozodi</span> survey, which explains why this star was nonetheless observed during our survey. The companion found in our observations, showing a contrast $\Delta H \simeq 0.7$ at an angular separation of 78.7 mas (i.e. projected linear separation of 3.6 AU) and position angle of $40\fdg9$, seems to match well the @Hartkopf12 discovery. It however hints at a significantly eccentric and/or inclined orbit. Although the current astrometric information is too scarce to confirm it, this visual binary could also match the stochastic astrometric solution proposed by @Goldin07. A face-on, circular orbit with a semi-major axis of 3.6 AU would indeed have a period of about 3.7 yr, which is in line with the 3.0 yr period found by @Goldin07. Finally, we can infer the spectral type of the companion from the measured contrast. With an absolute H-band absolute magnitude $M_H=1.31 \pm 0.02$ for the primary and a contrast $\Delta H = 0.73\pm 0.08$, the companion has an estimated absolute magnitude $M_H=2.04 \pm 0.08$, which gives an F0V spectral type according to @Allen00. HD 29388 -------- 90 Tau (HD 29388, HIP 21589, HR 1473) is an A6V member of the Hyades cluster. It hosts a cold debris disc [@Su06]. Although it was observed by speckle interferometry [@Patience98], no sign of binarity was found so far around that star to our knowledge, despite being extensively observed since the 1950s as a chemically peculiar member of an open cluster. 90 Tau is nevertheless listed as a member of a double system in Simbad, but with only a faint visual companion at a distance of about $120\arcsec$. The companion discovered here is thus completely new. With an absolute magnitude $M_H=0.71\pm0.01$ for its primary star and a contrast $\Delta H=3.77\pm0.07$, the companion has an absolute magnitude $M_H=4.48\pm0.07$, which corresponds to a K4V star according to @Allen00. Assuming a face-on, circular orbit with a semi-major axis of 11.1 mas (0.52 AU), the period would be around 84 days. This can be considered as a minimal period for the system. HD 202730 --------- Tet Ind (HD 202730, HIP 105319, HR 8140) has been known to host a visual G0V companion (GJ 9733 B) since the 19th century. Its angular separation has been measured over 150 years and is well approximated with a linear motion in the Catalog of Rectilinear Elements[^5]. In July 2012, the companion is expected to be separated by about $7\farcs2$ from the primary, which is comfortably outside the PIONIER field of view (and even outside the AT field of view). Tet Ind has also been shown to have a constant radial velocity in the survey of @Lagrange09. The discovery of a nearly equal flux companion therefore comes as a surprise. Based on the measured contrast ($\Delta H = 0.15 \pm 0.17$), we estimate the companion to have an A5V spectral type like the primary. The odds for a face-on orbit are high (no radial velocity variations detected). Assuming a circular orbit, the period would then be around 1.3 years. Discussion {#sec:discussion} ========== A population of undetected stellar companions? ---------------------------------------------- Five binary stars were identified in this study. In addition to the four stars discussed in Sect. \[sec:properties\], we include here HD 224392, whose companion will be discussed in a forthcoming paper (S. Borgniet et al., in prep.). This is an A-type star, just as the four others are. The fact that five stars among the 30 supposedly single A-type stars[^6] of the <span style="font-variant:small-caps;">Exozodi</span> sample were revealed to be binaries deserves some further discussion. \ \ Historically, many dedicated studies have been carried out to quantify the number of binaries in the solar neighbourhood, showing higher multiplicity rates around more massive stars [see @Duchene13 for a recent review]. A lot of papers were published regarding the multiplicity of solar-type stars (F-G stars), including e.g. censuses of binaries within 25 pc [@Raghavan10] and within 67 pc [@Tokovinin14I; @Tokovinin14II]. Later spectral types have also been extensively addressed, including a quasi-complete sample within 10 pc in @Henry06. Even brown dwarfs have had their share of multiplicity studies [see @Burgasser07 for a review]. However, when it comes to earlier spectral types, and A-type stars in particular, assessing the multiplicity is much more challenging. One of the main reasons is that classical spectroscopic methods do not work very well in the case of A-type stars, for which the spectral lines are generally broadened by high rotational velocities (typically 100–200 km s$^{-1}$) and usually blended. Spectroscopic surveys of A-type stars are therefore much less sensitive and complete than for solar-type stars. They have mostly targeted chemically peculiar A-type stars so far [e.g. @Carrier02; @Carquillat07]. A new method for radial velocity measurements has only recently been proposed to survey intermediate-mass stars [@Galland05], enabling the search for substellar companions, although with an RV accuracy much worse than for solar-type stars. Systematic searches for visual companions around A-type stars also face a significant challenge, i.e. the steep mass-luminosity relationship, which makes low-mass companions hard to detect. Only recently have large-scale, dedicated (AO-assisted) surveys been conducted, including most notably the VAST survey [@DeRosa11]. The first results of this survey, targeting an incomplete sample of 435 A-type stars out to 75 pc, show that $33.8\pm2.6\%$ of A-type stars are visual binaries in the separation range between 30 to 10000 AU (i.e. about $0\farcs4$ to $140\arcsec$). A significant gap however remains between the realm of spectroscopic binaries and the visual binary domain. The study of astrometric binaries from the Hipparcos catalogue partly fills that gap, but the nature of the detected companions, and even their orbital parameters, are generally not well constrained. Interferometry offers an attractive way to fill that gap, as recently proposed for massive stars [@Sana14], before GAIA revolutionises the field. With the present study, we address a separation range (1–100 mas) that has only been partially scraped by speckle interferometry in the past [e.g. @Mason13; @Tokovinin14]. Among the five stars resolved as binaries within the present survey, only one was unambiguously identified as a visual, gravitationally bound, binary before the start of our survey (HD 202730, which has a companion at $7\farcs2$). The four others are new visual companions, although the companion to HD 16555 was independently published by @Hartkopf12. We can therefore consider that four out of 30 (i.e. $13.3_{-4.0}^{+8.6}\%$) A-type stars in our sample revealed to be unknown visual binaries. Adding this to the previous result of @DeRosa14 suggest that about 47% of A-type stars are visual binaries for angular separations ranging from 1 mas to $140\arcsec$, although a gap remains between the outer edge of the PIONIER field of view ($\sim 100$ mas) and the inner edge of the @DeRosa14 survey ($0\farcs4$). This gap could adequately be covered by more systematic speckle interferometry or sparse aperture masking observations. For instance, as explained in Sect. \[sub:results\], the F5V star HD 15798 was recently shown to host a faint companion at an angular distance of $0\farcs21$ by @Tokovinin14. Based on this discussion, it is legitimate to ask oneself how many more close companions remain undiscovered around A-type stars in the solar neighbourhood. Our study suggests that, even after a careful target selection, five out of 30 (i.e. 17%) A-type stars still have a stellar companion within an angular separation range that critically affects interferometric observations. Using A-type stars as calibrators for interferometric observations should therefore be avoided as much as possible. We also note that, despite our efforts to search the literature for the signs of faint companions, some of the hot exozodiacal disc detections claimed within our CHARA/FLUOR survey of debris disc stars [@Absil13] could also be due to unknown, close companions. Assuming that up to 17% of the observed A-type stars in the CHARA/FLUOR survey have a unknown faint companion within the FLUOR field of view (similar to the PIONIER fleld of view) would still preserve a hot exozodiacal disc occurrence rate significantly larger around A-type stars ($\sim$33%) than solar-type stars ($\sim$18%) in that survey. On the PIONIER sensitivity {#sub:pionier} -------------------------- In the case of non detections (79 stars out of 92 in our sample), we can determine an upper limit on the presence of a companion as a function of the position in the field of view, using the $\chi^2$ statistics as explained in @Absil11. The resulting sensitivity map can then be used to derive the median sensitivity (i.e. sensitivity achieved for 50% of the positions within the search region). We can also define the sensitivity at a higher completeness level, e.g. the sensitivity reached for 90% of the positions within the search region. In order to deduce the typical sensitivity of PIONIER in survey mode (3 OBs per target), we produced the histograms of the sensitivity levels for all stars showing no near-infrared excess. These histograms give us the sensitivity to companions in 50% or 90% of the search region at a significance level of $3\sigma$ for the CP, the $V^2$, and the combination of the two. They are illustrated in Fig. \[fig:sensitivitydet\]. We deduce from this figure that the sensitivity of PIONIER is around 1% when using the CP and the $V^2$ in a combined way (median sensitivity of 0.7%, percentile 90 sensitivity of 1.1%). \  Figure \[fig:sensitivitydet\] also gives interesting information on the sensitivities of the CP and $V^2$ individually. They show a similar median sensitivity level ($\sim 1$%), but the situation changes when considering a completeness level of 90%, for which the sensitivity of the CP (2.7%) degrades much more than for the $V^2$ (1.6%). This is somewhat surprising as the closure phases are generally thought to be more sensitive to the presence of faint companions. To investigate the origin of this unexpected behaviour, we make use of the magnification factor defined by @LeBouquin12 for the CP, $$m_{\rm CP} = sin(\alpha_{12})+sin(\alpha_{23})-sin(\alpha_{12}+\alpha_{23}) \; ,$$ where $\alpha_{ij}=2\pi \vec{B}_{ij}\cdot\vec{\Delta}/ \lambda$, with $\vec{B}_{ij}$ the baseline vector, $\vec{\Delta}$ the apparent binary separation vector, and $\cdot$ the scalar product. This purely geometric factor captures the amplification of the CP created by an off-axis companion as a function of its position in the field of view. It is possible to define an equivalent magnification factor for the visibilities, assuming an unresolved primary star, $$m_{V^2} = \sum_{ij}(1-V^2_{ij}) \; ,$$ with $V^2_{ij}$ the squared visibility for baseline $ij$. This definition gives the cumulated drop of visibility created by the companion on all baselines; the larger the drop, the more conspicuous the companion. The magnification maps for two different three-telescope configurations of the VLTI sub-array are given in Fig. \[fig:magmaps\]. The absolute value of the magnification factor is plotted to reveal more clearly the blind spots (green-black regions). We can easily note that, in both configurations, the closure phase magnification map comprises many more regions where the magnification factor is close to zero. This is probably related to the fact that the magnification for the CP ranges between negative and positive values, while the magnification for the $V^2$ is always positive. The closure phase therefore shows many more blind spots, where the presence of a companion would not show up in the data. This becomes even clearer in the histograms plotted in Fig. \[fig:magmaps\] (right), where we see that most of the occurrences for $m_{\rm CP}$ are close to zero. Even though the CP are arguably more robust than the $V^2$ for the detection of a companion, the poor coverage of the field of view limits the completeness of the search at a given contrast level. We conclude that taking into account the $V^2$ is highly recommended when searching for faint companions, despite the possible presence of false positives in the combined $\chi^2$ related to circumstellar discs. Conclusions {#sec:concl} =========== In this paper, we have systematically searched for companions around the 92 stars observed within the <span style="font-variant:small-caps;">Exozodi</span> survey, which aims to unveil the occurrence rate of bright exozodiacal discs around nearby main sequence stars using infrared interferometry. Based on our VLTI/PIONIER observations, five new companions are resolved around HD 4150, HD 16555, HD 29388, HD 202730, and HD 224392, although three of these stars (HD 4150, HD 16555 and HD 224392) were already suspected to be binaries based on Hipparcos astrometry, and one of the binaries (HD 16555) was independently resolved by speckle interferometry while we were carrying out our survey. All these companions happen to be detected around A-type stars. They have estimated spectral types ranging from K4V to A5V, assuming that they are all on the main sequence. The fact that such bright companions have remained undiscovered to date led us to discuss how our observations affect the current estimation of binary fraction around A-type stars. In particular, based on our discoveries, the fraction of visual binaries increases from about 34% to about 47%. These serendipitous discoveries suggest that a significant fraction of supposedly single A-type stars are still undetected binaries. The <span style="font-variant:small-caps;">Exozodi</span> data set also allowed us to study the sensitivity of PIONIER to off-axis companions in its survey mode (3 OBs per star), showing a typical contrast limit of 1%. Finally, we conclude that using the squared visibilities together with the closure phases in the search for companions is crucial for maximising the completeness of the search. The authors thank the French National Research Agency (ANR, contract ANR-2010 BLAN-0505-01, EXOZODI) for financial support. L.M. acknowledges the F.R.S.-FNRS for financial support through a FRIA PhD fellowship. We thank S. Borgniet and A.-M. Lagrange for helpful discussions on HD 224392 and on radial velocity surveys of early-type stars. This work made use of the Smithsonian/NASA Astrophysics Data System (ADS) and of the Centre de Données astronomiques de Strasbourg (CDS). [^1]: F.R.S.-FNRS Research Associate [^2]: Based on observations made with ESO Telescopes at the La Silla Paranal Observatory under programme IDs 089.C-0365 and 090.C-0526. [^3]: See also @Absil13 for a description of the first part of the <span style="font-variant:small-caps;">Exozodi</span> survey, carried out in the northern hemisphere with the FLUOR instrument on the CHARA array. [^4]: This formula can be obtained if we assume that the periodicity in the $V^2$ and in the CP are the same. Using a mean baseline $B$, we can then determine $\Delta\lambda$ using the $2\pi$ periodicity of the $\sin(2\pi B\theta/\lambda)$ term in the $V^2$ formula, i.e. $2\pi B\theta/\lambda= 2\pi B\theta /(\lambda +\Delta \lambda) + 2 \pi$. [^5]: http://www.usno.navy.mil/USNO/astrometry/optical-IR-prod/wds/lin1 [^6]: although we recognise that two of them have been classified as astrometric binaries in the literature

ArXiv

--- abstract: 'In empirical studies of friendship networks participants are typically asked, in interviews or questionnaires, to identify some or all of their close friends, resulting in a directed network in which friendships can, and often do, run in only one direction between a pair of individuals. Here we analyze a large collection of such networks representing friendships among students at US high and junior-high schools and show that the pattern of unreciprocated friendships is far from random. In every network, without exception, we find that there exists a ranking of participants, from low to high, such that almost all unreciprocated friendships consist of a lower-ranked individual claiming friendship with a higher-ranked one. We present a maximum-likelihood method for deducing such rankings from observed network data and conjecture that the rankings produced reflect a measure of social status. We note in particular that reciprocated and unreciprocated friendships obey different statistics, suggesting different formation processes, and that rankings are correlated with other characteristics of the participants that are traditionally associated with status, such as age and overall popularity as measured by total number of friends.' author: - Brian Ball - 'M. E. J. Newman' title: Friendship networks and social status --- Introduction {#introduction .unnumbered} ------------ A social network, in the most general sense of the term, consists of a group of people, variously referred to as nodes or actors, connected by social interactions or ties of some kind [@WF94]. In this paper we consider networks in which the ties represent friendship. Friendship networks have been the subject of scientific study since at least the 1930s. A classic example can be found in the studies by Rapoport and collaborators of friendship among schoolchildren in the town of Ann Arbor, MI in the 1950s and 60s [@RH61], in which the investigators circulated questionnaires among the students in a school asking them to name their friends. Many similar studies have been done since then, with varying degrees of sophistication, but most employ a similar questionnaire-based methodology. A counterintuitive aspect of the resulting networks is that they are directed. Person A states that person B is their friend and hence there is a direction to the ties between individuals. It may also be that person B states that person A is their friend, but it does not have to be the case, and in practice it turns out that a remarkably high fraction of claimed friendships are not reciprocated. In the networks we study in this paper the fraction of reciprocated ties rarely exceeds 50% and can be as low as 30%. This could be seen as a problem for the experimenter. One thinks of friendship as a two-way street—a friendship that goes in only one direction is no friendship at all. How then are we to interpret the many unreciprocated connections in these networks? Are the individuals in question friends or are they not? One common approach is simply to disregard the directions altogether and consider two individuals to be friends if they are connected in either direction (or both) [@Airoldi11]. In this paper, however, we take a different view and consider what we can learn from the unreciprocated connections. It has been conjectured that, rather than being an error or an annoyance, the pattern of connections might reflect underlying features in the structure or dynamics of the community under study [@Homans50; @Davis72; @DBF00]. Working with a large collection of friendship networks from US schools, we find that in every network there is a clear ranking of individuals from low to high such that almost all friendships that run in only one direction consist of a lower-ranked individual claiming friendship with a higher-ranked one. We conjecture that these rankings reflect a measure of social status and present a number of results in support of this idea. For instance, we find that a large majority of reciprocated friendships are between individuals of closely similar rank, while a significant fraction of unreciprocated friendships are between very different ranks, an observation consistent with qualitative results in the sociological literature going back several decades [@Davis72]. We also investigate correlations between rank and other individual characteristics, finding, for example, that there is a strong positive correlation between rank and age, older students having higher rank on average, and between rank and overall popularity, as measured by total number of friends. The outline of the paper is as follows. First, we describe our method of analysis, which uses a maximum-likelihood technique in combination with an expectation–maximization algorithm to extract rankings from directed network data. Then we apply this method to school friendship networks, revealing a surprisingly universal pattern of connections between individuals in different schools. We also present results showing how rank correlates with other measures. Finally we give our conclusions and discuss possible avenues for future research. Inference of rank from network structure {#Sec:Math .unnumbered} ---------------------------------------- Consider a directed network of friendships between $n$ individuals in which a connection running from person A to person B indicates that A claims B as a friend. Suppose that, while some of the friendships in the network may be reciprocated or bidirectional, a significant fraction are unreciprocated, running in one direction only, and suppose we believe there to be a ranking of the individuals implied by the pattern of the unreciprocated friendships so that most such friendships run from lower to higher rank. One possible way to infer that ranking would be simply to ignore any reciprocated friendships and then construct a minimum violations ranking of the remaining network [@Reinelt85; @ACK86]. That is, we find the ranking of the network nodes that minimizes the number of connections running from higher ranked nodes to lower ranked ones. In practice this approach works quite well: for the networks studied in this paper the minimum violations rankings have an average of 98% of their unreciprocated friendships running from lower to higher ranks and only 2% running the other way. By contrast, versions of the same networks in which edge directions have been randomized typically have about 10% of edges running the wrong way. (Statistical errors in either case are 1% or less, so these observations are highly unlikely to be the results of chance.) The minimum violations ranking, however, misses important network features because it focuses only on unreciprocated friendships. In most cases there are a substantial number of reciprocated friendships as well, as many as a half of the total, and they contain significant information about network structure and ranking. For example, as we will see, pairs of individuals who report a reciprocated friendship are almost always closely similar in rank. To make use of this information we need a more flexible and general method for associating rankings with network structure. In this paper we use a maximum likelihood approach defined as follows. Mathematically we represent the distinction between reciprocated and unreciprocated friendships in the network using two separate matrices. The symmetric matrix ${\mathbf}{S}$ will represent the reciprocated connections—undirected edges in graph theory terms—such that $S_{ij}=S_{ji}=1$ if there are connections both ways between nodes $i$ and $j$, and zero otherwise. The asymmetric matrix ${\mathbf}{T}$ will represent the unreciprocated (directed) edges with $T_{ij}=1$ if there is a connection to node $i$ from node $j$ (but not *vice versa*), and zero otherwise. The matrices ${\mathbf}{S}$ and ${\mathbf}{T}$ are related to the conventional adjacency matrix ${\mathbf}{A}$ of the network by ${\mathbf}{A}={\mathbf}{S}+{\mathbf}{T}$. Now suppose that there exists some ranking of the individuals, from low to high, which we will represent by giving each individual a unique integer rank in the range 1 to $n$. We will denote the rank of node $i$ by $r_i$ and the complete set of ranks by $R$. We have found it to be a good approximation to assume that the probability of friendship between two individuals is a function only of the difference between their ranks. We specifically allow the probability to be different for reciprocated and unreciprocated friendships, which acknowledges the possibility that the two may represent different types of relationships, as conjectured for instance in [@Davis72; @Dijkstra10]. We define a function $\alpha(r_i-r_j)$ to represent the probability of an undirected edge between $i$ and $j$ and another $\beta(r_i-r_j)$ for a directed edge to $i$ from $j$. Since $\alpha(r)$ describes undirected edges it must be symmetric $\alpha(-r)=\alpha(r)$, but $\beta(r)$ need not be symmetric. If we were not given a network but we were given the probability functions $\alpha$ and $\beta$ and a complete set of rankings on $n$ vertices, then we could use this model to generate—for instance on a computer—a hypothetical but plausible network in which edges appeared with the appropriate probabilities. In effect, we have a random graph model that incorporates rankings. In this paper, however, we want to perform the reverse operation: given a network we want to deduce the rankings of the nodes and the values of the functions $\alpha$ and $\beta$. To put that another way, if we are given a network and we assume that it is generated from our model, what values of the rankings and probability functions are most likely to have generated the network we observe? This question leads us to a maximum likelihood formulation of our problem, which we treat using an expectation–maximization (EM) approach in which the ranks $R$ are considered hidden variables to be determined and the functions $\alpha$ and $\beta$ are parameters of the model. Using a Poisson formulation of the random network generation process, we can write the probability of generation of a network $G$ with rankings $R$, given the functions $\alpha$ and $\beta$, as $$\begin{aligned} P(G,R|\alpha,\beta) &= \prod_{i>j} {[\alpha(r_i-r_j)]^{S_{ij}}\over S_{ij}!}\,{\mathrm{e}}^{-\alpha(r_i-r_j)} \nonumber\\ & \quad{}\times \prod_{i\ne j} {[\beta(r_i-r_j)]^{T_{ij}}\over T_{ij}!}\, {\mathrm{e}}^{-\beta(r_i-r_j)}. \label{eq:likelihood}\end{aligned}$$ Note that we have excluded self-edges here, since individuals cannot name themselves as friends. We have also assumed that the prior probability of $R$ is uniform over all sets of rankings. The most likely values of the parameter functions $\alpha$ and $\beta$ are now given by maximizing the marginal likelihood $P(G|\alpha,\beta) = \sum_R P(G,R|\alpha,\beta)$, or equivalently maximizing its logarithm, which is more convenient. The logarithm satisfies the Jensen inequality $$\log \sum_R P(G,R|\alpha,\beta) \ge \sum_R q(R) \log {P(G,R|\alpha,\beta)\over q(R)}, \label{eq:jensen}$$ for any set of probabilities $q(R)$ such that $\sum_R q(R) = 1$, with the equality being recovered when $$q(R) = {P(G,R|\alpha,\beta)\over \sum_R P(G,R|\alpha,\beta)}. \label{eq:estep}$$ This implies that the maximization of the log-likelihood on the left side of  is equivalent to the double maximization of the right side, first with respect to $q(R)$, which makes the right side equal to the left, and then with respect to $\alpha$ and $\beta$, which gives us the answer we are looking for. It may appear that expressing the problem as a double maximization in this way, rather than as the original single one, makes it harder, but in fact that’s not the case. The right-hand side of  can be written as $\sum_R q(R) \log P(G,R|\alpha,\beta) - \sum_R q(R) \log q(R)$, but the second term does not depend on $\alpha$ or $\beta$, so as far as $\alpha$ and $\beta$ are concerned we need consider only the first term, which is simply the average $\overline{\mathcal{L}}$ of the log-likelihood over the distribution $q(R)$: $$\overline{\mathcal{L}} = \sum_R q(R) \log P(G,R|\alpha,\beta).$$ Making use of Eq.  and neglecting an unimportant overall constant, we then have $$\begin{aligned} \overline{\mathcal{L}} &= \sum_R q(R) \sum_{i\ne j} \Bigl[ {\mbox{$\frac12$}}S_{ij} \log \alpha(r_i-r_j) + T_{ij} \log \beta(r_i-r_j) \nonumber\\ &\hspace{8em}{} - {\mbox{$\frac12$}}\alpha(r_i-r_j) - \beta(r_i-r_j) \Bigr], \label{eq:mstep1}\end{aligned}$$ where we have used the fact that $\alpha(r)$ is a symmetric function. This expression can be simplified further. The first term in the sum is $$\begin{aligned} & {\mbox{$\frac12$}}\sum_R q(R) \sum_{i\ne j} S_{ij} \log \alpha(r_i-r_j) \nonumber\\ &\hspace{3em}{} = {\mbox{$\frac12$}}\sum_z \sum_{i\ne j} S_{ij} q(r_i-r_j=z) \log \alpha(z), \label{eq:msimplify1}\end{aligned}$$ where $q(r_i-r_j=z)$ means the probability within the distribution $q(R)$ that $r_i-r_j=z$. We can define $$a(z) = {1\over n-|z|} \sum_{i\ne j} S_{ij} q(r_i-r_j=z), \label{eq:defsnu}$$ which is the expected number of undirected edges in the observed network between pairs of nodes with rank difference $z$. It is the direct equivalent in the observed network of the quantity $\alpha(z)$, which is the expected number of edges in the model. The quantity $a(z)$, like $\alpha(z)$, is necessarily symmetric, $a(z)=a(-z)$, and hence  can be written as $${\mbox{$\frac12$}}\sum_R q(R) \sum_{i\ne j} S_{ij} \log \alpha(r_i-r_j) = \sum_{z=1}^{n-1} (n-z) a(z) \log\alpha(z).$$ Similarly, we can define $$b(z) = {1\over n-|z|} \sum_{i\ne j} T_{ij} q(r_i-r_j=z) \label{eq:defsnd}$$ and $$\begin{aligned} &\sum_R q(R) \sum_{i\ne j} T_{ij} \log \beta(r_i-r_j) \nonumber\\ &\hspace{2em}{} = \sum_{z=1}^{n-1} (n-z) \bigl[ b(z) \log\beta(z) + b(-z) \log\beta(-z) \bigr],\end{aligned}$$ where $b(z)$ is the expected number of directed edges between a pair of nodes with rank difference $z$. Our final expression for $\overline{\mathcal{L}}$ is $$\begin{aligned} \overline{\mathcal{L}} &= \sum_{z=1}^{n-1} (n-z) \bigl[ a(z) \log\alpha(z) - \alpha(z) \nonumber\\ &\quad{} + b(z) \log\beta(z) - \beta(z) + b(-z) \log\beta(-z) - \beta(-z) \bigr].\end{aligned}$$ Our approach involves maximizing this expression with respect to $\alpha(z)$ and $\beta(z)$ for given $a(z)$ and $b(z)$, which can be done using standard numerical methods. (Note that the expression separates into terms for the directed and undirected edges, so the two can be maximized independently.) The values of $a(z)$ and $b(z)$ in turn are calculated from Eqs. , , and , leading to an iterative method in which we first guess values for $\alpha(z)$ and $\beta(z)$, use them to calculate $q(R)$ and hence $a(z)$ and $b(z)$, then maximize $\overline{\mathcal{L}}$ to derive new values of $\alpha$ and $\beta$, and repeat to convergence. This is the classic expectation–maximization approach to model fitting. Two further elements are needed to put this scheme into practice. First, we need to specify a parametrization for the functions $\alpha$ and $\beta$. We have found the results to be robust to the choice of parametrization, but in the results reported here we find $\alpha$ to be well represented by a Gaussian centered at the origin. The function $\beta$ takes a more complicated form which we parametrize as a Fourier cosine series, keeping five terms and squaring to enforce nonnegativity, plus an additional Gaussian peak at the origin. Second, the sum in the denominator of Eq.  is too large to be numerically tractable, so we approximate it using a Markov chain Monte Carlo method—we generate complete rankings $R$ in proportion to the probability $q(R)$ given by Eq.  and average over them to calculate $a(z)$ and $b(z)$. Results {#Sec:Results .unnumbered} ------- We have applied the method of the previous section to the analysis of data from the US National Longitudinal Study of Adolescent Health (the “AddHealth” study), a large-scale multi-year study of social conditions for school students and young adults in the United States [@note1]. Using results from surveys conducted in 1994 and 1995, the study compiled friendship networks for over $90\,000$ students in schools covering US school grades 7 to 12 (ages 12 to 18 years). Schools were chosen to represent a broad range of socioeconomic conditions. High schools (grades 9 to 12) were paired with “feeder” middle schools (grades 7 and 8) so that networks spanning schools could be constructed. To create the networks, each student was asked to select, from a list of students attending the same middle/high school combination, up to ten people with whom they were friends, with a maximum of five being male and five female. From these selections, 84 friendship networks were constructed ranging in size from tens to thousands of students, one for each middle/high school pair, along with accompanying data on the participants, including school grade, sex, and ethnicity. Some of the networks divide into more than one strongly connected component, in which case we restrict our analysis to the largest component only. We perform the EM analysis of the previous section on each network separately, repeating the iterative procedure until the rankings no longer change. ![(a) Probability of reciprocated friendships as a function of rank difference (normalized to run from $-1$ to 1). The histogram shows empirical results for a single example network; the solid curve is the fitted function $\alpha(z)$. (b) The equivalent plot for unreciprocated friendships.[]{data-label="plot:p_data_show"}](fig1.eps){width="\columnwidth"} Figure \[plot:p\_data\_show\] shows results for a typical network. In panel (a), the histogram shows the measured value of the quantity $a(z)$, Eq. , the empirical probability of a reciprocated friendship (technically the expected number of undirected edges) between a vertex pair with rank difference $z$, with the horizontal axis rescaled to run from $-1$ to $1$ (rather than $-n$ to $n$). As the figure shows the probability is significantly different from zero only for small values of $z$, with a strong peak centered on the origin. The solid curve shows the fit of this peak by the Gaussian function $\alpha(z)$, which appears good. The fit is similarly good for most networks. The form of $a(z)$ tells us that most reciprocated friendships fall between individuals of closely similar rank: there is a good chance that two people with roughly equal rank will both claim the other as a friend, but very little chance that two people with very different ranks will do so. This result seems at first surprising, implying as it does that people must be able to determine their own and others’ rank with high accuracy in order to form friendships, but a number of previous studies have suggested that indeed this is true [@Anderson06]. Panel (b) of Fig. \[plot:p\_data\_show\] shows $b(z)$, Eq. , for the same network, which is the probability of a directed edge between nodes with rank difference $z$. Again there is a strong central peak to the distribution, of width similar to that for the undirected edges, indicating that many unreciprocated friendships are between individuals of closely similar rank. However, the distribution also has a substantial asymmetric tail for positive values of the rank difference, indicating that in a significant fraction of cases individuals claim friendship with those ranked higher than themselves, but that those claims are not reciprocated. The black curve in the panel shows the best fit to the function $\beta(z)$ in the maximum-likelihood calculation. ![The fitted central peak of the friendship probability distributions for (a) reciprocated and (b) unreciprocated friendships. The horizontal axes are measured in units of absolute (unrescaled) rank difference divided by average network degree. Each blue curve is a network. The bold black curves represent the mean.[]{data-label="plot:real_distributions"}](fig2.eps){width="0.98\columnwidth"} The general forms of these distributions are similar across networks from different schools. They also show interesting scaling behavior. The widths of the central peaks for both undirected and directed edges, when measured in terms of raw (unrescaled) rank difference are, to a good approximation, simply proportional to the average degree of a vertex in the network. Figure \[plot:real\_distributions\] shows these peaks for 78 of the 84 networks on two plots, for undirected edges (panel (a)) and directed edges (panel (b)), rescaled by average degree, and the approximately constant width is clear. (The six networks not shown are all small enough that the central peaks for the directed edges can be fit by the other parameters of the model and thus a direct comparison is not appropriate.) This result indicates that individuals have, roughly speaking, a fixed probability of being friends with others close to them in rank, regardless of the size of the community as a whole—as the average number of friends increases, individuals look proportionately further afield in terms of rank to find their friends, but are no more likely to be friends with any particular individual of nearby rank. Outside of the central peak, i.e., for friendships between individuals with markedly different ranks, there are, to a good approximation, only unreciprocated friendships, and for these the shape of the probability distribution appears by contrast to be roughly constant when measured in terms of the rescaled rank of Fig. \[plot:p\_data\_show\], which runs from $-1$ to 1. This probability, which is equal to the function $\beta(z)$ with the central Gaussian peak subtracted, is shown in Fig. \[plot:supposed\_distributions\] for the same 78 networks, rescaled vertically by the average probability of an edge to account for differing network sizes, and again the similarity of the functional form across networks is apparent, with low probability in the left half of the plot, indicating few claimed friendships with lower-ranked individuals, and higher probability on the right. The roughly constant shape suggests that, among the unreciprocated friendships, there is, for example, a roughly constant probability of the lowest-ranked student in the school claiming friendship with the highest-ranked, relative to other students, no matter how large the school may be. The emerging picture of friendship patterns in these networks is one in which reciprocated friendships appear to fall almost entirely between individuals of closely similar rank. A significant fraction of the unreciprocated ones do the same, and moreover show similar scaling to their reciprocated counterparts, but the remainder seem to show a quite different behavior characterized by different scaling and by claims of friendship by lower-ranked individuals with substantially higher-ranked ones. ![The fitted probability function for unreciprocated friendships, minus its central peak. The horizontal axis measures rank difference rescaled to run from $-1$ to 1. Each blue curve is a network. The bold black curve is the mean.[]{data-label="plot:supposed_distributions"}](onetail.eps){width="6.5cm"} Discussion {#Sec:Discussion .unnumbered} ---------- Taking the results of the previous section as a whole, we conjecture that the rankings discovered by the analysis correlate, at least approximately, with social status. If we assume that reciprocated friendships—almost all of which fall in the central peak—correspond to friendships in the conventional sense of mutual interaction, then a further conjecture, on the basis of similar statistics, is that the unreciprocated friendships in the central peak are also mutual but, for one reason or another, only one side of the relationship is represented in the data. One explanation why one side might be missing is that respondents in the surveys were limited to listing only five male and five female friends, and so might not have been able to list all of their friendships. On the other hand, one might conjecture that the unreciprocated claims of friendship with higher-ranked individuals, those in the tail of the distribution in Fig. \[plot:p\_data\_show\]b, correspond to “aspirational” friendships, hopes of friendship with higher-ranked individuals that are, at present at least, not returned. Note also how the tail falls off with increasing rank difference: individuals are more likely to claim friendship with others of only modestly higher rank, not vastly higher. One way to test these conjectures is to look for correlations between the rankings and other characteristics of individuals in the networks. For instance, it is generally thought that social status is positively correlated with the number of people who claim you as a friend [@Hallinan88; @Dijkstra10]. Figure \[plot:degree\_correlations\]a tests this by plotting average rank over all individuals in all networks (averaged in the posterior distribution of Eq. ) as a function of network in-degree (the number of others who claim an individual as a friend). As the figure shows, there is a strong positive slope to the curve, with the most popular individuals being nearly twice as highly ranked on average as the least popular. Figure \[plot:degree\_correlations\]b shows the corresponding plot for out-degree, the number of individuals one claims as a friend, and here the connection is weaker, as one might expect—claiming many others as friends does not automatically confer high status upon an individual—although the correlation is still statistically significant. Figure \[plot:degree\_correlations\]c shows rank as a function of total degree, in-degree plus out-degree, which could be taken as a measure of total social activity, and here again the correlation is strong. For all three panels the correlations are significant, with $p$-values less than $0.001$. ![Plots of rescaled rank versus degree, averaged over all individuals in all networks for (a) in-degree, (b) out-degree, and (c) the sum of degrees. Measurement errors are comparable with or smaller than the sizes of the data points and are not shown.[]{data-label="plot:degree_correlations"}](fig4.eps){width="\columnwidth"} In addition to the network structure itself, we have additional data about each of the participants, including their age (school grade), sex, and ethnicity. The distributions of rank for each sex and for individual ethnicities turn out to be close to uniform—a member of either sex or any ethnic group is, to a good approximation, equally likely to receive any rank from 1 to $n$, indicating that there is essentially no effect of sex or ethnicity on rank. (A Kolmogorov–Smirnov test does reveal deviations from uniformity in some cases, but the deviations are small, with KS statistics $D<0.08$ in all instances.) Age, however, is a different story. Figure \[plot:meta\_ranks\] shows the rescaled rank of individuals in each grade from 7 to 12, averaged over all individuals in all networks, and here there is a clear correlation. Average rank increases by more than a factor of two from the youngest students to the oldest (a one-way ANOVA gives $p<0.001$). Since older students are generally acknowledged to have higher social status [@Coleman61], this result lends support to the identification of rank with status. A further interesting wrinkle can be seen in the results for the 8th and 9th grades. Unlike other pairs of consecutive grades, these two do not have a statistically significant difference in average rank (a $t$-test gives $p>0.95$). This may reflect the fact that the 8th grade is the most senior grade in the feeder junior-high schools, before students move up to high school. When they are in the 8th grade, students are temporarily the oldest (and therefore highest status) students in school and hence may have a higher rank than would be expected were all students in a single school together. ![Rescaled rank as a function of school grade, averaged over all individuals in all schools.[]{data-label="plot:meta_ranks"}](rank_vs_grade.eps){width="\columnwidth"} Finally, in Fig. \[plot:network\] we show an actual example of one of the networks, with nodes arranged vertically on a scale of inferred rank and colored according to grade. The increase of rank with grade is clearly visible, as is the fact that most undirected edges run between individuals of similar rank (and hence run horizontally in the figure). ![A sample network with (rescaled) rank on the vertical axis, vertices colored according to grade, and undirected edges colored differently from directed edges. Rank is calculated as an average within the Monte Carlo calculation (i.e., an average over the posterior distribution of the model), rather than merely the maximum-likelihood ranking. Note the clear correlation between rank and grade in the network.[]{data-label="plot:network"}](comm71-meta.location.eps){width="\columnwidth"} Conclusions {#Sec:Conclusion .unnumbered} ----------- In this paper, we have analyzed a large set of networks of friendships between students in American high and junior-high schools, focusing particularly on the distinction between friendships claimed by both participating individuals and friendships claimed by only one individual. We find that students can be ranked from low to high such that most unreciprocated friendships consist of a lower-ranked individual claiming friendship with a higher-ranked one. We have developed a maximum-likelihood method for inferring such ranks from complete networks, taking both reciprocated and unreciprocated friendships into account, and we find that the rankings so derived correlate significantly with traditional measures of social status such as age and overall popularity, suggesting that the rankings may correspond to status. On the other hand, rankings seem to be essentially independent on average of other characteristics of the individuals involved such as sex or ethnicity. There are a number of questions unanswered by our analysis. We have only limited data on the personal characteristics of participants. It would be interesting to test for correlation with other characteristics. Are rankings correlated, for instance, with academic achievement, number of siblings or birth order, number of Facebook friends, after-school activities, personality type, body mass index, wealth, or future career success? There is also the question of why a significant number of apparently close friendships are unreciprocated. One idea that has appeared in the literature is that some directed edges may correspond to new, temporary, or unstable friendships, which are either in the process of forming and will become reciprocated in the future, or will disappear over time [@Hallinan88; @Sorensen76]. Evidence suggests that in practice about a half of the unreciprocated friendships do the former and a half the latter, and it is possible that the two behaviors correspond to the two classes of directed edges we identify in our analysis. A test of this hypothesis, however, would require time-resolved data—successive measurements of friendship patterns among the same group of individuals—data which at present we do not possess. Finally, there are potential applications of the statistical methods developed here to other directed networks in which direction might be correlated with ranking, such as networks of team or individual competition [@Stefani97; @PN05b] or dominance hierarchies in animal communities [@Drews93; @DeVries98]. The authors thank Carrie Ferrario, Brian Karrer, Cris Moore, Jason Owen-Smith, Bethany Percha, and Claire Whitlinger for useful comments and suggestions. This work was funded in part by the National Science Foundation under grant DMS–1107796 and by the James S. McDonnell Foundation. [10]{} url \#1[`#1`]{}urlprefix S. Wasserman and K. Faust, *Social Network Analysis*. Cambridge University Press, Cambridge (1994). A. Rapoport and W. J. Horvath, A study of a large sociogram. *Behavioral Science* **6**, 279–291 (1961). E. M. Airoldi, D. S. Choi, and P. J. Wolfe, Confidence sets for network structure. *Statistical Analysis and Data Mining* **4**, 461–469 (2011). G. C. Homans, *The Human Group*. Harcourt, Brace, and Co., New York (1950). J. A. Davis and S. Leinhardt, The structure of positive interpersonal relations in small groups. *Sociological Theories in Progress* **2**, 218–251 (1972). P. Doreian, V. Batagelj, and A. Ferligoj, Symmetric-acyclic decompositions of networks. *Journal of Classification* **17**, 3–28 (2000). G. Reinelt, *The Linear Ordering Problem: Algorithms and Applications*. Heldermann, Berlin (1985). I. Ali, W. D. Cook, and M. Kress, On the minimum violations ranking of a tournament. *Management Science* **32**, 660–672 (1986). J. K. Dijkstra, A. H. N. Cillessen, S. Lindenberg, and R. Veenstra, Basking in reflected glory and its limits: Why adolescents hang out with popular peers. *Journal of Research on Adolescents* **20**, 942–958 (2010). This work uses data from Add Health, a program project designed by J. Richard Udry, Peter S. Bearman, and Kathleen Mullan Harris, and funded by a grant P01-HD31921 from the National Institute of Child Health and Human Development, with cooperative funding from 17 other agencies. Special acknowledgment is due Ronald R. Rindfuss and Barbara Entwisle for assistance in the original design. Persons interested in obtaining data files from Add Health should contact Add Health, Carolina Population Center, 123 W. Franklin Street, Chapel Hill, NC 27516–2524 ([email protected]). C. Anderson, S. Srivastava, J. S. Beer, S. E. Spataro, and J. A. Chatman, Knowing your place: Self-perceptions of status in face-to-face groups. *Journal of Personality and Social Psychology* **91**, 1094–1110 (2006). M. T. Hallinan and W. N. Kubitschek, The effect of individual and structural characteristics on intransitivity in social networks. *Social Psychology Quarterly* **51**, 81–92 (1988). J. S. Coleman, *The Adolescent Society: The Social Life of the Teenager and its Impact on Education*. Greenwood Press, Westport, Conn. (1961). A. B. S[ø]{}rensen and M. T. Hallinan, A stochastic model for change in group structure. *Social Science Research* **5**, 43–61 (1976). R. Stefani, Survey of the major world sports rating systems. *Journal of Applied Statistics* **24**, 635–646 (1997). J. Park and M. E. J. Newman, A network-based ranking system for [A]{}merican college football. *J. Stat. Mech.* P10014 (2005). C. Drews, The concept and definition of dominance in animal behaviour. *Behaviour* **125**, 283–313 (1993). H. De Vries, Finding a dominance order most consistent with a linear hierarchy: A new procedure and review. *Animal Behaviour* **55**, 827–843 (1998).

ArXiv

--- abstract: 'We use a Hamiltonian interacting particle system to derive a stochastic mean field system whose McKean-Vlasov equation yields the incompressible Navier Stokes equation. Since the system is Hamiltonian, the particle relabeling symmetry implies a Kelvin Circulation Theorem along stochastic Lagrangian paths. Moreover, issues of energy dissipation are discussed and the model is connected to other approaches in the literature.' address: 'Austrian Financial Market Authority (FMA), Otto-Wagner Platz 5, A-1090 Vienna ' author: - Simon Hochgerner date: 'September 2, 2018' title: 'A Hamiltonian Mean-Field System for the Navier-Stokes Equation' --- Introduction {#introduction .unnumbered} ============ Stochastic fluid dynamics can be discussed from different perspectives: 1. Multi-scale approach, stochastic dynamics for modelling fluid flow under uncertainties: The idea is to separate the dynamics into a slow (deterministic) and a fast (stochastic) component. The result is a stochastic system and the goal is generally to study the corresponding S(P)DE as a realistic model of fluid motion. Representatives of this approach are [@CGD17; @CFH17; @Holm15; @Mem14; @RMC17; @RMC17a; @RMC17b]. 2. Stochastic approaches to deterministic fluid mechanics: Again one uses a stochastic perturbation to capture fine-scale effects. But, in contrast to (1), the goal is to average over the stochastic system to gain information about (or, a solution of) the resulting deterministic model. Representatives of this approach are [@CC07; @CI05; @C; @CS09; @E10; @H17; @I06b; @IM08; @IN11; @NS17; @Y83]. The present paper belongs to the second category. We are concerned with the incompressible Navier-Stokes equation on the $n$-dimensional torus $M=T^n = \R^n/\mathbb{Z}^n$: $$\begin{aligned} {\mbox{$\frac{\partial }{\partial t}$}}u = -\nabla_u u + \eta\Delta u - \nabla p,\; \d\,u = 0 ,\; u(0,x) = u_0(x) \end{aligned}$$ where $\eta$ is the viscosity and $u = u(t,x)$ and $p=p(t,x)$ are Eulerian velocity and pressure, respectively. Our approach is based on the following “separation of dynamics” idea: We assume that each fluid parcel consists of a large number of identical particles. The corresponding dynamics is then derived as a stochastic Hamiltonian system with respect to an energy that consists of two components: 1. a deterministic part due to the total momentum of the ensemble of fluid particles; 2. a stochastic part due to independent random impacts on the individual particles. Thus the particles interact deterministically via the total ensemble momentum. In the limit, as the number of particles goes to infinity, this interacting particle system (IPS) tends to a mean field stochastic differential equation . Averaging over solutions to this mean field SDE yields a deterministic PDE that describes the deterministic dynamics of the original fluid parcel at the macroscopic level. Essentially only assuming that the random impacts are independent and Gaussian, we show that this PDE is the incompressible Navier-Stokes equation. The incompressibility condition follows because we use the space of volume preserving diffeomorphisms as the configuration space for our system. ### Description of contents Section \[sec:prel\] contains notation and preliminaries. We give a detailed exposition of the Hamiltonian structure used throughout the paper. Section \[sec:ham-ips\] describes the above mentioned Hamiltonian IPS for fluid dynamics. In order to have a simple and tractable picture, we start with a system of identical particles on the real line. Then we use the Hamiltonian structure introduced in Section \[sec:prel\] to transfer the construction to the phase space of incompressible fluid mechanics. Section \[sec:main\] contains our main Equation  and shows how this is obtained as a mean field limit from the IPS in Section \[sec:ham-ips\]. We remark that this passage to the mean field equation is carried out under the assumption that the limit exists. Theorem \[thm:main\] then shows that averaging over solutions of leads to solutions of the Navier-Stokes equation for incompressible flow. Thus the Navier-Stokes equation is obtained from the McKean-Vlasov equation for . A converse is proved as-well: a Gaussian stochastic perturbation of Lagrangian trajectories, corresponding to a solution of the Navier-Stokes equation, yields a mean-field SDE whose mean field coincides with the original solution. Due to the Hamiltonian structure we also obtain a Kelvin Circulation Theorem (Proposition \[prop:KC\]) that holds along stochastic Lagrangian paths. Section \[sec:energy\] is concerned with issues of energy dissipation. It is shown that the system  neither conserves stochastic energy, nor does the average over the stochastic energy dissipate. However, considering a slight modification of the equation proposed in [@H17], one does obtain a stochastic energy whose average dissipates and bounds the energy of corresponding solutions to the Navier-Stokes equation. Section \[sec:IPS\] discusses energy dissipation for the empirical mean of the corresponding IPS. This is carried out, as in Section \[sec:energy\], for Equation  and for the system proposed in [@H17]. It is found that the average over the (stochastic) energy of the empirical mean need not dissipate. While this is contrary to intuition, since one would expect the energy of the empirical mean to behave as that of the deterministic solution, this result is consistent with [@IM08; @IN11; @NS17]. Section \[sec:comp\] offers a discussion of connections with other approaches to stochastic fluid dynamics. We compare our model  to representatives of both of the categories mentioned at the beginning of this introduction. Moreover, in Section \[sec:H17\] it is explained why complying with the Kelvin Circulation Theorem is a desirable property. Section \[sec:conc\] recapitulates the results and draws conclusions. **Acknowledgements.** I am grateful to Darryl Holm for useful remarks and explanations. The referee reports are also gratefully acknowledged, in particular for pointing out references [@Holm02a; @HT12; @Mem14; @RMC17; @RMC17a; @RMC17b]. Preliminaries: phase space and Hamiltonian structure {#sec:prel} ==================================================== Let $({\Omega},\F,(\F_t)_{t\in[0,T]},P)$ be a filtered probability space satisfying the usual assumptions. In the following, all stochastic processes shall be understood to be adapted to this filtration. Volume preserving diffeomorphisms --------------------------------- Let $M=T^n=\R^n/\mathbb{Z}^n$. We fix $s>1+n/2$ and let $G^s$ denote the infinite dimensional ${\mbox{$C^{\infty}$}}$-manifold of $H^s$-diffeomorphisms on $M$. Further, $G^s_0$ denotes the submanifold of volume preserving diffeomorphisms of Sobolev class $H^s$. Both, $G^s$ and $G^s_0$, are topological groups but not Lie groups since left composition is only continuous but not smooth. Right composition is smooth. The tangent space of $G^s$ (resp. $G^s_0$) at the identity $e$ shall be denoted by ${\mathfrak{g}}^s$ (resp. ${\mathfrak{g}}^s_0$). Let ${\mbox{$\mathcal{X}$}}^s(M)$ denote the vector fields on $M$ of class $H^s$ and ${\mbox{$\mathcal{X}$}}_0^s(M)$ denote the subspace of divergence free vector fields of class $H^s$. We have ${\mathfrak{g}}^s_0 = {\mbox{$\mathcal{X}$}}^s_{0}(M)$ and ${\mathfrak{g}}^s={\mbox{$\mathcal{X}$}}^s(M)$. See [@EM70; @MEF]. We will keep track of the superscript $s$ only when it is needed for clarification. Otherwise, we shall assume $s$ to be fixed throughout and write $G$, $G_0$, ${\mathfrak{g}}$, ${\mathfrak{g}}_0$ instead of $G^s$, $G^s_0$, ${\mathfrak{g}}^s$, ${\mathfrak{g}}^s_0$. We use right multiplication $R^g: G_0\to G_0$, $k\mapsto k\circ g = kg$ to trivialize the tangent bundle $TG_0\cong G_0\times{\mathfrak{g}}_0$, $\xi_g\mapsto(g,(TR^g)^{-1}\xi_g)$. By right multiplication we extend the $L^2$ inner product $\ww<.,.>$ on ${\mathfrak{g}}_0$ to a Riemannian metric $\mu$ on $G_0$, that is $$\mu_g(\xi_g,\eta_g) = \int_M\vv<(TR^g)^{-1}\xi_g(x),(TR^g)^{-1}\eta_g(x)>\, dx = \ww<(TR^g)^{-1}\xi_g, (TR^g)^{-1}\eta_g>$$ for $\xi_g,\eta_g\in T_gG$, where $dx$ is the standard volume element in $M$. We shall subsequently use $\vv<.,.>$ to denote both, the standard inner product in $\R^n$ and the duality pairing . Leray-Hodge projection $P$ -------------------------- The Leray-Hodge projection operator is defined as follows (see [@MEF Corollary 1.4.4]): Consider an $H^s$ vector field $\xi$ on $M$. Then there is a unique divergence free vector field $\eta$ of class $H^s$ and a function $f$ on $M$ such that $\xi=\eta+\nabla f$. Setting $\eta=P\xi$ thus defines a bounded linear operator $P: {\mbox{$\mathcal{X}$}}^s(M)\to{\mbox{$\mathcal{X}$}}_0^s(M)$ (for arbitrary $s\ge0$). Further, $\ww<\eta,\nabla f> = 0$. Inertia operator {#sec:InOp} ---------------- Let ${\mathfrak{g}}_0'$ denote the space of continuous linear functionals on ${\mathfrak{g}}_0$. Since $\mu$ is a weak Riemannian metric, we do not obtain an isomorphism between ${\mathfrak{g}}_0$ and ${\mathfrak{g}}_0'$. Following [@AK98], define the smooth dual ${\mathfrak{g}}_0^*$ as $${\mathfrak{g}}_0^* = {\Omega}^1(M)^s/{\mbox{$\{\psi\in{\Omega}^1(M)^s: \psi = df\}$}}$$ which is the space of $1$-forms of Sobolev class $H^s$ modulo exact $1$-forms of class $H^s$. The *inertia tensor* is defined as the isomorphism $$\label{equ:I} \check{\mu}: {\mathfrak{g}}_0{\longrightarrow}{\mathfrak{g}}_0^*, \quad \xi\longmapsto[\xi^{\flat}]$$ where $\xi^{\flat}$ is the $1$-form associated to $\xi$ (raising indices) and $[\cdot]$ denotes the equivalence class. The inverse of $\check{\mu}$ is given by $\check{\mu}^{-1}[\psi+df] = P\psi^{\sharp}$ where $\psi$ is a representative of the class $[\psi+df]\in{\mathfrak{g}}_0^*$ and $\sharp$ is the inverse to $\flat$. The pairing between $[\xi^{\flat}]\in{\mathfrak{g}}_0^*$ and $\eta\in{\mathfrak{g}}_0$ is $$\label{e:dual-p} \vv<[\xi^{\flat}],\eta> = \ww<\xi,\eta> = \int_M \vv<\xi(x),\eta(x)>\,dx$$ which is well-defined independently of the representative of $[\xi^{\flat}]$. We define the bracket $[.,.]$ to be the negative of the usual Lie bracket: $ [\xi,\eta] := -\nabla_{\xi}\eta + \nabla_{\eta}\xi $ where $\nabla_{\xi}\eta=\vv<\xi,\nabla>\eta$. This choice of sign is compatible with [@A66; @AK98; @Michor06]. Define the operators ${\mbox{$\text{\upshape{ad}}$}}$ and ${\mbox{$\text{\upshape{ad}}$}}^*$ by ${\mbox{$\text{\upshape{ad}}$}}(\xi).\eta = [\xi,\eta]$ and ${\mbox{$\text{\upshape{ad}}$}}^*(\xi).[\eta^{\flat}] = [-\eta^{\flat}\circ{\mbox{$\text{\upshape{ad}}$}}(\xi)]$ for $\xi,\eta\in{\mathfrak{g}}_0$. Further, define ${\mbox{$\text{\upshape{ad}}$}}^{\top}$ by ${\mbox{$\text{\upshape{ad}}$}}^{\top}(\xi) := -{\mbox{$\text{\upshape{ad}}$}}(\xi)^{\top}$ which is the transpose with respect to the weak inner product $\ww<.,.>$: $$\ww<{\mbox{$\text{\upshape{ad}}$}}(\xi)^{\top}.\eta, \zeta> = \ww<\eta,{\mbox{$\text{\upshape{ad}}$}}(\xi).\zeta>.$$ Sobolev spaces are not closed under the Lie bracket, whence the operations ${\mbox{$\text{\upshape{ad}}$}}$, ${\mbox{$\text{\upshape{ad}}$}}^{\top}$ and ${\mbox{$\text{\upshape{ad}}$}}^*$ lose one Sobolev index, that is ${\mbox{$\text{\upshape{ad}}$}}(\xi).\eta=[\xi,\eta]\in{\mathfrak{g}}_0^{s-1}$ for $\xi,\eta\in{\mathfrak{g}}_0={\mathfrak{g}}_0^s$. Thus ${\mathfrak{g}}_0$ is not a Lie algebra. \[lem:I\] We have ${\mbox{$\text{\upshape{ad}}$}}^*(\xi).\check{\mu}(\eta) = -\check{\mu}(\nabla_{\xi}\eta + \xi'\otimes\eta)$ and ${\mbox{$\text{\upshape{ad}}$}}^{\top}(\xi).\eta = -P(\nabla_{\xi}\eta + \xi'\otimes\eta)$. Let $\xi,\eta,\zeta\in{\mathfrak{g}}_0$ $$\begin{aligned} ({\mbox{$\text{\upshape{ad}}$}}^*(\xi).\check{\mu}(\eta))(\zeta) &= -\int_M\vv<\eta , [\xi,\zeta]>\,dx = \int_M \vv<\eta, \nabla_{\xi}\zeta - \nabla_{\zeta}\xi>\, dx \\ &= \int_M\Big(\d(\vv<\eta,\zeta>\xi) - \vv<\nabla_{\xi} \eta , \zeta> - \vv<\xi'\otimes \eta , \zeta> \Big)\,dx\end{aligned}$$ Symplectic structure on $G_0\times{\mathfrak{g}}_0$ --------------------------------------------------- Due to the identification $T^*G_0\cong G_0\times{\mathfrak{g}}_0^*\cong G_0\times{\mathfrak{g}}_0$, the space $G_0\times{\mathfrak{g}}_0$ carries a natural symplectic structure. We follow, and slightly adapt, the exposition of [@Michor06 Section 4] to describe this structure. The second tangent bundle is expressed, again via right multiplication, as $ T(G_0\times{\mathfrak{g}}_0) = G_0\times{\mathfrak{g}}_0\times{\mathfrak{g}}_0\times{\mathfrak{g}}_0 = {\mbox{$\{(g,\xi,\dot{g},\dot{\xi})\}$}} $. The natural exact symplectic form ${\Omega}$ on $G_0\times{\mathfrak{g}}_0$ is given by $$\label{e:sp_form} {\Omega}_{(g,\xi)}\Big((\dot{g}_1,\dot{\xi}_1),(\dot{g}_2,\dot{\xi}_2)\Big) = \ww<\dot{\xi}_2,\dot{g}_1> + \ww<-\dot{\xi}_1 + {\mbox{$\text{\upshape{ad}}$}}^{\top}(\dot{g}_1).\xi, \dot{g}_2>.$$ The Hamiltonian vector field $X_f$ of a function $f$ is defined by $i(X_f){\Omega}= df$. It is given explicitly by $$\label{e:ham_vf} X_f(g,\xi) = \Big( \textup{grad}_2 f\,(g,\xi), {\mbox{$\text{\upshape{ad}}$}}^{\top}(\textup{grad}_2 f\,(g,\xi)).\xi - \textup{grad}_1 f\,(g,\xi) \Big)$$ where $\textup{grad}_1 f$ and $\textup{grad}_2 f$ are the partial gradients. Since $\mu$ is only a weak Riemannian metric, the gradients need not exist. But if they do, then so does the Hamiltonian vector field, and it is given by the above formula. The momentum map ---------------- The particle relabeling symmetry group $G_0$ acts on itself by right multiplication $R^g: G^s_0\to G^s_0$. The tangent lifted action acts on $G^s_0\times{\mathfrak{g}}^s_0$ via $TR^g: (k,\xi)\mapsto(kg,\xi)$. The associated momentum map is given by $$\label{e:momap} J: G^s_0\times{\mathfrak{g}}^s_0{\longrightarrow}({\mathfrak{g}}^{s-1}_0)^*, \; (g,\xi)\longmapsto {\mbox{$\text{\upshape{Ad}}$}}( g)^*[\xi^\flat] = [g^*\xi^{\flat}]$$ where ${\mbox{$\text{\upshape{Ad}}$}}( g)\xi = TL_{ g}.(TR^{ g})^{-1}.\xi = T g\circ\xi\circ g^{-1}$, $L_{ g}$ is left multiplication by $g$, and ${\mbox{$\text{\upshape{Ad}}$}}(g)^*=:{\mbox{$\text{\upshape{Ad}}$}}^*(g^{-1})$ is the adjoint of ${\mbox{$\text{\upshape{Ad}}$}}(g)$ with respect to the dual pairing. Further, $g^*\xi^{\flat}$ is the pullback of the one-form $\xi^{\flat}$ by the diffeomorphism $g$. Since the pullback involves differentiation, the result is only of Sobolev class $H^{s-1}$. Thus $J$ is not a momentum map in the classical sense. Nevertheless, we can use it as long as the expression $J(g,\xi)$ makes sense and is of sufficient regularity. \[prop:kct\_gen\] Let $\mathcal{N}\subset\mathbb{N}$ and consider Hamiltonian functions $h, f^a: G_0^s\times{\mathfrak{g}}_0^s\to\R$ for $a\in\mathcal{N}$. Let $(W^a)$ for $a\in\mathcal{N}$ be a sequence of pairwise independent Brownian motions. Assume $\gamma_t = (g_t,\xi_t)$ solves the Stratonovich Hamiltonian SDE $$\delta\gamma_t = X_h(\gamma_t)\,\delta t + \sum_{a\in\mathcal{N}}X_{f^a}(\gamma_t)\;\delta W^a_t$$ for $t\in[0,T]$. If $h$ and $f^a$ are invariant under $G^s_0$, then, for any closed smooth curve $C$ in $M$, $$\delta\int_{(g_t)_*C}\xi_t^{\flat} = 0.$$ By the stochastic Noether Theorem ([@LCO08]), $J(g_t,\xi_t)$ is constant in $t$. Thus $$\delta ({\mbox{$\text{\upshape{Ad}}$}}( g_t)^*.[\xi_t^{\flat}]) = 0 \in{\mathfrak{g}}_0^*,$$ whence there is a time dependent function $p_t$ such that $\delta(\xi_t\circ{\mbox{$\text{\upshape{Ad}}$}}( g_t)) = d p_t$. [Full]{} differentials vanish when integrated over a closed loop $C$. Therefore, $$0 = \int_C \delta( \xi_t^{\flat}\circ{\mbox{$\text{\upshape{Ad}}$}}( g_t) ) = \delta \int_C g_t^*\xi^{\flat} = \delta \int_{( g_t)_*C}\xi_t^{\flat}.$$ The *mechanical connection* is defined by $A := \check{\mu}^{-1}\circ J$, that is $$\label{e:mech_con} A: G^s_0\times{\mathfrak{g}}^s_0{\longrightarrow}{\mathfrak{g}}^{s-1}_0, \; (g,\xi)\longmapsto \check{\mu}^{-1}\Big({{\mbox{$\text{\upshape{Ad}}$}}^*(g^{-1})}[\xi^\flat]\Big) =: {\mbox{$\text{\upshape{Ad}}$}}^{\top}(g^{-1})(\xi) .$$ A Hamiltonian interacting particle system for fluid dynamics {#sec:ham-ips} ============================================================ Translational kinetic energy of a particle ensemble {#sec:det_dyn} --------------------------------------------------- Consider a system of $N$ identical particles in $\R$ with positions $q_1,\dots,q_N\in\R$ and total mass $M=1$. Suppose that the dynamics of the system is determined entirely by the translational kinetic energy of the particle ensemble. The phase space of the system is $T^*\R^N$ and we identify $T^*\R^N=T\R^N$ via the standard inner product $\vv<.,.>$. Let ${\mbox{$\widetilde{{\omega}}$}} = {\mbox{$\frac{1}{N}$}}{\omega}$ denote the natural symplectic form on $T\R$. Then the product symplectic form is $${\mbox{$\widetilde{{\omega}^N}$}} := {\mbox{$\frac{1}{N}$}}{\omega}^N := \sum_{j=1}^N\pi_j^*{\mbox{$\widetilde{{\omega}}$}} = {\mbox{$\frac{1}{N}$}}\sum_{j=1}^N dq^j\wedge dv^j$$ where $(q,v) = (q^1,\dots,q^N,v^1,\dots,v^N)$ are coordinates on $T\R^N$ and $\pi_j: T\R^N\to T\R$ is the projection onto the $j$-th factor. Since ${\mbox{$\frac{1}{N}$}}\sum_{j=1}^N v^j$ is the translational momentum of the system, the kinetic energy is given by $${\mbox{$\widetilde{\mathcal{H}}$}}(q,v) := {\mbox{$\frac{1}{N}$}}\mathcal{H}(q,v) := {\mbox{$\frac{1}{2N^2}$}}\sum_{j,k=1}^N\vv<v^j,v^k> = {\mbox{$\frac{1}{2N^2}$}}\sum_{j=1}^N\Big(\vv<v^j,v^j> + \sum_{k\neq j}\vv<v^j,v^k>\Big).$$ Since $i(X){\mbox{$\widetilde{{\omega}^N}$}} = d{\mbox{$\widetilde{\mathcal{H}}$}}$ if and only if $i(X){\omega}^N = d\mathcal{H}$, the dynamics of the system is given by the Hamiltonian vector field $X=X_{\mathcal{H}}$ with respect to ${\omega}^N$. Denote the $i$-th component of the Hamiltonian vector field by $X^{i}_{\mathcal{H}}(q,v) = T\pi_i.X_{\mathcal{H}}(q,v)$. We remark that $X^{i}_{\mathcal{H}}(q,v)$ is not a vector field on $T\R$. Since ${\omega}^N = \sum_{j=1}^N\pi_j^*{\omega}$, it follows that the $i$-th component $\gamma^i(t)$ of the integral curve $\gamma(t) = (q(t),v(t))$ is determined by $$\label{e:det_dyn} {\mbox{$\frac{\partial }{\partial t}$}}\gamma^i(t) = X_{\mathcal{H}}^{i}(\gamma(t)) = \left( \begin{matrix} {\mbox{$\frac{\partial }{\partial v^i}$}}\mathcal{H}(\gamma(t)) \\ -{\mbox{$\frac{\partial }{\partial q^i}$}}\mathcal{H}(\gamma(t)) \end{matrix} \right) = \left( \begin{matrix} {\mbox{$\frac{1}{N}$}}\sum_{j=1}^N v^j(t) \\ 0 \end{matrix} \right)$$ The Hamiltonian construction of Brownian motion in finite dimensions -------------------------------------------------------------------- Let $(Q,\mu^Q)$ be a Riemannian manifold with dimension $\dim Q=m$. For a vector field $X\in{\mbox{$\mathcal{X}$}}(Q)$ we define the *momentum function* $F_X: TQ\to\R$ by $$\label{e:mom_fun} F_X(Y) = \mu^Q(X(\tau(Y)), Y)$$ where $\tau: TQ\to Q$ is the tangent projection. We use the metric isomorphsim $\check{\mu}^Q:TQ\to T^*Q$ to transfer the natural symplectic form ${\omega}^{T^*Q}$ on $T^*Q$ to a symplectic form ${\omega}^{TQ} = (\check{\mu}^Q)^*{\omega}^{T^*Q}$ on $TQ$. Let $X_f$ denote the Hamiltonian vector field on $TQ$ with respect to ${\omega}^{TQ}$. Assume now that there is a global orthonormal frame $(x_{\alpha})_{\alpha=1}^m$ on $Q$ such that $\nabla_{x_{\alpha}}x_{\alpha}=0$. We shall abbreviate $F^{\alpha}:=F_{x_{\alpha}}$. Consider Brownian motion $(W^{\alpha})_{\alpha=1}^m$ in $\R^m$ and assume that the stochastic process $\gamma$ in $TQ$ is a solution to the Stratonovich SDE $$\label{e:ham_bm} \delta\gamma = \sum X_{F^{\alpha}}(\gamma)\,\delta W^{\alpha}.$$ It is shown in [@LCO08 Section 3.4] that then $\tau\circ\gamma$ is Brownian motion in $Q$. Stochastic perturbation of --------------------------- Let us consider the case $Q=\R$ and ${\omega}^{T\mathbb{R}} = {\omega}= dq\wedge dv$ where $(q,v)$ are coordinates on $T\R$. Let ${\mbox{$\widetilde{F}$}}: T\R\to\R$ $(q,v)\mapsto {\mbox{$\frac{1}{N}$}}e v =: {\mbox{$\frac{1}{N}$}}F(q,v)$ where $e=1$ is viewed as the standard basis vector in $\R$. Notice, as in Section \[sec:det\_dyn\], that $i(X){\mbox{$\widetilde{{\omega}}$}} = d{\mbox{$\widetilde{F}$}}$ if and only if $i(X){\omega}= dF$. Let $X_F$ be the Hamiltonian vector field of $F$ with respect to ${\omega}$. The stochastic perturbation of Equation  is therefore given by $$\label{e:stoch_dyn} \delta \gamma^i_t = X_{\mathcal{H}}^i(\gamma_t)\,\delta t + X_{F}(\gamma_t^{{i}})\,\delta W_t^i$$ where $W_t^1,\dots,W_t^N$ are independent copies of Brownian motion in $\R$ and $\gamma_t=(\gamma^i_t)_i$. This describes a particle in an ensemble, where the energy of each (identical) particle is given as a sum of two components: 1. a deterministic part due to the translational momentum of the ensemble; 2. an internal energy that is modeled as a Brownian motion. The corresponding model for the full system of particles is given by the stochastic Hamiltonian equation $$\label{e:stoch_dyn2} \delta \gamma_t = X_{\mathcal{H}}(\gamma(t))\,\delta t + \sum_{i=1}^N X_{\pi_i^*F}(\gamma_t)\,\delta W_t^i$$ on $T\R^N$, where $X_{\pi_i^*F}$ is the Hamiltonian vector field of the pullback $\pi_i^*F$ to $T\R^N$. We have $T\pi_i.X_{\pi_i^*F} = X_F\circ\pi_i$. Equation is the motivation for below. An orthogonal system on ${\mathfrak{g}}_0$ ------------------------------------------ Let $$\mathbb{Z}_n^+ := {\mbox{$\{k\in\mathbb{Z}_n: k_1>0 \textup{ or, for } i=2,\ldots,n, k_1=\ldots=k_{i-1}=0, k_i>0 \}$}}.$$ For $k\in\mathbb{Z}^+_n$ let $k_1^{\bot},\ldots,k_{n-1}^{\bot}$ denote a choice of pairwise orthogonal vectors in $\R^n$ such that $|k_i^{\bot}|=|k|$ and $\vv<k_i^{\bot},k>=0$ for all $i=1,\ldots,n-1$. In [@CS09 Appendix] the following system of vectors is introduced: $$A_{(k,i)} = \frac{1}{|k|^{s+1}}\cos\vv<k,x>k^{\bot}_i,\; B_{(k,i)} = \frac{1}{|k|^{s+1}}\sin\vv<k,x>k^{\bot}_i,\; A_{(0,j)} = e_j$$ Here $e_j\in\R^n$ is the $j$th standard vector. By slight abuse of notation we identify these vectors with their corresponding right invariant vector fields on $G_0$. Further, we shall make use of the multi-index notation $\alpha = (k,i,a)$ where $k\in\mathbb{Z}_n^+$ and $a=0,1,2$ such that $$\begin{aligned} X_{\alpha} &= A_{(0,i)} \textup{ with } i=1,\ldots,n \textup{ if } a=0\\ X_{\alpha} &= A_{(k,i)} \textup{ with } i=1,\ldots,n-1 \textup{ if } a=1\\ X_{\alpha} &= B_{(k,i)} \textup{ with } i=1,\ldots,n-1 \textup{ if } a=2 \end{aligned}$$ Thus by a sum over $\alpha$ we shall mean a sum over these multi-indices, and this notation for $X_{\alpha}$ will be used throughout the rest of the paper. It is shown in [@CS09 Appendix] that the $X_{\alpha}$ form an orthogonal system of basis vectors in ${\mathfrak{g}}_0$ such that $$\label{e:nablaXX} \nabla_{X_{\alpha}}X_{\alpha} = \vv<X_{\alpha},\nabla>X_{\alpha} = 0.$$ and, for $\xi\in{\mbox{$\mathcal{X}$}}(M)$, $$\label{e:Delta} \sum \nabla_{X_{\alpha}}\nabla_{X_{\alpha}}\xi = c^s\Delta\xi$$ where $c^s>0$ is a constant and $\Delta$ is the vector Laplacian. \[prop:bm\] Let $W_t = \sum X_{\alpha} W_t^{\alpha}$, where $W_t^{\alpha}$ are independent copies of Brownian motion in $\R$. Then $W$ defines (a version of) Brownian motion (i.e., cylindrical Wiener process) in ${\mathfrak{g}}_0$. The Hamiltonian Interacting Particle System (IPS) ------------------------------------------------- We consider the phase space $(G_0\times{\mathfrak{g}}_0)^N$ for (large) $N\in\mathbb{N}$. Let $\pi_i: (G_0\times{\mathfrak{g}}_0)^N\to G_0\times{\mathfrak{g}}_0$ be the projection onto the $i$-th factor. Let ${\Omega}$, from now on, be the symplectic form on $G_0\times{\mathfrak{g}}_0$ and ${\Omega}^N := \sum_{i=1}^N\pi_i^*{\Omega}$ the product symplectic form. Following we define the vector valued Hamiltonian $(\mathcal{H}^N,F_i^{\alpha})_{i,\alpha}$ by $$\begin{aligned} \mathcal{H}^N(\Gamma) &= {\mbox{$\frac{1}{2N}$}}\sum_{i,j=1}^N\ww<\xi^i,\xi^j> \\ F_i^{\alpha}(\Gamma) &= \ww<\xi^i, X_{\alpha}> \end{aligned}$$ where $\Gamma = (\Gamma_i)_{i=1}^N = (g^i,\xi^i)_{i=1}^N\in(G_0\times{\mathfrak{g}}_0)^N$. For each $\alpha$ let $(W^{\alpha,i})_{i=1}^N$ be Brownian motion in $\R^N$ such that $W^{\alpha,i}$ and $W^{\beta,j}$ are independent for $(\alpha,i)\neq(\beta,j)$. For $\eta>0$ we use $\nu=\sqrt{{\mbox{$\frac{2\eta}{c^s}$}}}$ to scale the white noise term. The resulting Stratonovich SDE in $(G_0\times{\mathfrak{g}}_0)^N$ associated to the collection $(\mathcal{H}^N,F_i^{\alpha})_{i,\alpha}$ is $$\label{e:ham_ips1} \delta \Gamma_t = X_{\mathcal{H}^N}(\Gamma_t)\,\delta t + \nu\sum_{{i},\alpha}X_{\pi_i^*F_i^{\alpha}}(\Gamma_t)\,\delta W^{\alpha,i}_t$$ Note that $T\pi_i.X_{\pi_i^*F_i^{\alpha}}(\Gamma) = X_{F_i^{\alpha}}(\pi_i(\Gamma))$ where $X_{\pi_i^*F_i^{\alpha}}$ is the Hamiltonian vector field on $(G_0\times{\mathfrak{g}}_0)^N$. The term $\sum_{{i},\alpha}X_{\pi_i^* F_i^{\alpha}}(\Gamma_t)\,\delta W^{\alpha,i}_t$ is an infinite dimensional version of Equation . This construction is justified by Proposition \[prop:bm\]. Equation  yields $$\begin{aligned} T\pi_i . X_{\mathcal{H}^N}(\Gamma) &= \Big( {\mbox{$\frac{1}{N}$}}\sum_{j=1}^N\xi^j, {\mbox{$\text{\upshape{ad}}$}}^{\top}({\mbox{$\frac{1}{N}$}}\sum_{j=1}^N\xi^j).\xi^i \Big)\\ T\pi_i . X_{F_i^{\alpha}}(\Gamma) &= \Big( X_{\alpha}, {\mbox{$\text{\upshape{ad}}$}}^{\top}(X_{\alpha}).\xi^i \Big) \end{aligned}$$ Therefore, the Hamiltonian equations of motion for $\Gamma_t = (g^i_t,\xi^i_t)_{i=1}^N$ can be written as $$\begin{aligned} \label{e:ham_ips2} \delta g^i_t &= TR^{g^i_t}\Big( {\mbox{$\frac{1}{N}$}}\sum_{j=1}^N\xi^j_t \,\delta t + \nu\sum_{\alpha} X_{\alpha}\,\delta W^{\alpha,i} \Big)\\ \label{e:ham_ips3} \delta \xi^i_t &= {\mbox{$\text{\upshape{ad}}$}}^{\top}({\mbox{$\frac{1}{N}$}}\sum_{j=1}^N\xi^j_t).\xi^i_t\,\delta t + \nu\sum_{\alpha}{\mbox{$\text{\upshape{ad}}$}}^{\top}(X_{\alpha}).\xi^i_t\,\delta W^{\alpha,i} \end{aligned}$$ The mean field system {#sec:main} ===================== The mean field limit -------------------- Stochastic mean field equations in a Hilbert space setting are treated in [@AD95]. See also [@Sni91; @Mel96] for finite dimensional convergence results. Consider equations  and together with a deterministic initial condition $\Gamma^i_0=(e,u_0)$ for $i=1,\dots,N$ and where $e\in G_0$ is the identity in the group. For the following we assume there exists a $T>0$, independent of $N$ but with a possible dependence on the initial condition, such that the system  and has a strong solution on $[0,T]$ in the usual SDE sense. Further, we assume that the mean field limit of and exists in the sense of [@AD95]. This means: 1. The empirical mean ${\mbox{$\frac{1}{N}$}}\sum_{j=1}^N\xi^j$ converges in probability to a smooth curve $u: [0,T]\to{\mathfrak{g}}_0$ as $N\to\infty$. That is, $u=u(t,x)=u_t(x)$ is a time dependent vector field. 2. The stochastic process $(g^1_t,\xi^1_t)$ converges in the appropriate norm, for $N\to\infty$, to a stochastic process $(g_t,\xi_t)$, and the limiting process solves $$\begin{aligned} \label{e:ham_mf1} \delta g_t &= TR^{g_t}\Big( u_t \,\delta t + \nu\sum_{\alpha} X_{\alpha}\,\delta W^{\alpha} \Big)\\ \label{e:ham_mf2} \delta \xi_t &= {\mbox{$\text{\upshape{ad}}$}}^{\top}(u_t).\xi_t\,\delta t + \nu\sum_{\alpha}{\mbox{$\text{\upshape{ad}}$}}^{\top}(X_{\alpha}).\xi_t\,\delta W^{\alpha} \end{aligned}$$ where $W^{\alpha} = W^{\alpha,1}$ is a sequence of independent Brownian motions. Since the particles are identical there is no loss in generality in considering the limit of $(g^1(t),\xi^1(t))$. 3. $u_t = E[\xi_t]$. Equations  and are mean field SDEs and their solutions are to be understood in the sense of Definition \[def:mf\_sol\]. \[def:mf\_sol\] Let $V$ be a Hilbert space and $W$ a cylindrical Wiener process in $V$. A $V$-valued stochastic process $\xi_t$ with $t\in[0,T]$ is a strong solution to the mean-field or McKean-Vlasov Stratonovich SDE $$\delta \xi_t = f(\xi_t,\mu_t)\,\delta t + g(\xi_t)\,\delta W_t$$ if 1. $\mu_t$ is the law of $\xi_t$, i.e., $\mu_t = (\xi_t)_*P$, 2. $\xi_t$ is adapted to $(\F_t)$, 3. $t\mapsto\xi_t$ is continuous $P$-a.s., 4. $\xi_t = \xi_0 + \int_0^t f(\xi_s,\mu_s) \,ds + \int_0^t g(\xi_s)\,\delta W_s $ for all $t\in[0,T]$ $P$-a.s. Notice that, once the prescription $t\mapsto\mu_t$ of the law of $\xi_t$ is found, the concept of a mean field Stratonovich SDE is not different from that of a time dependent Stratonovich SDE. The Stratonovich integral above is hence to be understood as $$\int_0^t g(\xi_s)\,\delta W_s = \int_0^t g(\xi_s)\, d W_s + {\mbox{$\frac{1}{2}$}}[g(\xi),W]_t.$$ See e.g. [@Pro Page 82]. Equations  and are defined on the phase space $G_0\times{\mathfrak{g}}_0$. They are Hamiltonian with respect to the symplectic structure : Consider the time dependent Hamiltonian $H^t(g,\xi) = \ww<\xi,u_t>$ and the associated time dependent Hamiltonian vector field $X_{H^t}$ given by $i(X_{H^t}){\Omega}= dH^t$. Equations  and are thus equivalent to $$\label{e:ham_mf3} \delta\Gamma_t = X_{H^t}(\Gamma_t)\,\delta t + \sum_{\alpha}X_{F^{\alpha}}(\Gamma_t)\,\delta W^{\alpha}_t$$ where $\Gamma_t = (g_t,\xi_t)$ and $F^{\alpha}(g,\xi) = \ww<\xi,X_{\alpha}>$. By right invariance of the Hamiltonian system  and , we can pass to ${\mathfrak{g}}_0$, which is the Lie-Poisson reduction of $G_0\times{\mathfrak{g}}_0$. Therefore, using the expression for ${\mbox{$\text{\upshape{ad}}$}}^{\top}$ in Lemma \[lem:I\], we consider, for $\eta>0$ and $\nu=\sqrt{{\mbox{$\frac{2\eta}{c^s}$}}}$, the mean-field Stratonovich equation in ${\mathfrak{g}}_0$ $$\begin{aligned} \label{e:sdeP} \delta \xi_t &= - P\Big(\nabla_{u_t}\xi_t + u_t'\otimes\xi_t \Big)\,\delta t - \nu \sum P\Big(\nabla_{X_{\alpha}}\xi_t + X_{\alpha}'\otimes\xi_t \Big) \,\delta W_t^{\alpha}\\ u_t &= E[\xi_t] \notag \\ \xi_0 &= u_0 \notag \end{aligned}$$ where $u_0\in{\mathfrak{g}}_0$ is the initial condition. In Proposition \[prop:KC\], Equation  will be used as a reconstruction equation to pass again to the full phase space $G_0\times{\mathfrak{g}}_0$. The stochastic velocity along a path of the fluid motion is given, in the Euler picture, by Equation  as $(TR^{g_t})^{-1}\delta g_t$. Since $(TR^{g_t})^{-1}\delta g_t$ and $\xi_t$ do not coincide, we refer to the latter as *stochastic specific momentum in velocity space* or, more briefly, as *stochastic momentum*. This terminology is justified, because $\check{\mu}(\xi_t) = \rho\check{\mu}(\xi_t)\in{\mathfrak{g}}_0^*$ is the momentum density, where $\check{\mu}$ is the inertia tensor  and $\rho=1$ is the fluid density. See also [@CFH17 Remark 5]. The distinction between velocity and momentum in velocity space is necessary, because the vector valued Hamiltonian, $(H^t,F^{\alpha})$ in , is not the kinetic energy Hamiltonian associated to $\mu$. Hamiltonian mean field system for the Navier-Stokes equation ------------------------------------------------------------ For $x\in M$ we define the (point) evaluation map by ${\mbox{$\text{{\upshape}{ev}}$}}_x: {\mathfrak{g}}\to T_x M = \R^n$, $\xi\mapsto\xi(x)$. \[lem:1\] Consider $\hat{X_{\alpha}}: {\mathfrak{g}}^s\to{\mathfrak{g}}^{s-1}$, $\xi\mapsto \nabla_{X_{\alpha}}\xi + X_{\alpha}'\otimes\xi$ and $\hat{Y_{\alpha}}: {\mathfrak{g}}^s\to{\mathfrak{g}}_0^{s-1}$, $\xi\mapsto P(\nabla_{X_{\alpha}}\xi + X_{\alpha}'\otimes\xi)$. Then $$\sum\hat{X_{\alpha}}\hat{X_{\alpha}}({\mbox{$\text{{\upshape}{ev}}$}}_x)(\xi) = c^s\Delta\xi(x) \textup{ and } \sum\hat{Y_{\alpha}}\hat{Y_{\alpha}}({\mbox{$\text{{\upshape}{ev}}$}}_x)(\xi) = c^s\Delta P \xi(x) .$$ Note that $\hat{Y_{\alpha}}(\xi) = -{\mbox{$\text{\upshape{ad}}$}}^{\top}(X_{\alpha}).\xi$ for $\xi\in{\mathfrak{g}}_0$. The directional derivative of ${\mbox{$\text{{\upshape}{ev}}$}}_x$ along $\hat{X_{\alpha}}$ is not necessarily well-defined, since the latter is not a proper vector field on ${\mathfrak{g}}^s$ but takes values in ${\mathfrak{g}}^{s-1}$. However, since ${\mbox{$\text{{\upshape}{ev}}$}}_x$ is linear, the Gâteaux derivative is $$\hat{X_{\alpha}}({\mbox{$\text{{\upshape}{ev}}$}}_x)(\xi) = d\,{\mbox{$\text{{\upshape}{ev}}$}}_x(\hat{X_{\alpha}})(\xi) = {\mbox{${\mbox{$\frac{\partial }{\partial t}$}}|_{0}$}}{\mbox{$\text{{\upshape}{ev}}$}}_x(\xi + t\hat{X_{\alpha}}(\xi)) = \hat{X_{\alpha}}(\xi)(x)$$ which clearly exists. Similarly, $$\begin{aligned} \sum\hat{X_{\alpha}}\hat{X_{\alpha}}({\mbox{$\text{{\upshape}{ev}}$}}_x)(\xi) &= {\mbox{$\text{{\upshape}{ev}}$}}_x\sum\Big( \nabla_{X_{\alpha}}\nabla_{X_{\alpha}}\xi + \nabla_{X_{\alpha}}(X_{\alpha}'\otimes\xi) + X_{\alpha}'\otimes\nabla_{X_{\alpha}}\xi + X_{\alpha}'\otimes X_{\alpha}'\otimes\xi \Big) $$ By the first term becomes $\sum\nabla_{X_{\alpha}}\nabla_{X_{\alpha}}\xi = c^s\Delta\xi$. The other terms cancel. Indeed, for $\alpha=(k,i,0)$ this is clear. Note that $$\begin{aligned} \label{e:A'} A_{(k,i)}' &= -B_{(k,i)}\otimes k \textup{ and } B_{(k,i)}' = A_{(k,i)}\otimes k. \end{aligned}$$ Let us consider $\alpha=(k,i,1)$, the case $\alpha=(k,i,2)$ being analogous. Now, $$\nabla_{A_{(k,i)}}(A^{{'}}_{(k,i)}\otimes\xi) = \nabla_{A_{(k,i)}}(-\vv<k,\xi>B_{(k,i)}) = -\vv<k,\xi>\nabla_{A_{(k,i)}}B_{(k,i)} - \vv<k,\nabla_{A_{(k,i)}}\xi> B_{(k,i)}$$ where the first term vanishes because of and the second term cancels in the sum because of the asymmetric sign change in . We have thus shown that $$\label{e:alpha_sum} \sum \nabla_{X_{\alpha}}(X_{\alpha}'\otimes\xi) = 0.$$ The terms $$A_{(k,i)}'\otimes\nabla_{A_{(k,i)}}\xi \textup{ and } B_{(k,i)}'\otimes\nabla_{B_{(k,i)}}\xi = -A_{(k,i)}'\otimes\nabla_{A_{(k,i)}}\xi$$ also cancel in the sum. Finally, $$\begin{aligned} A_{(k,i)}'\otimes A_{(k,i)}'\otimes\xi &= A_{(k,i)}'\otimes(-B_{(k,i)}\otimes k\otimes\xi) = \frac{1}{|k|^{2s+2}}\sin^2\vv<k,x> k_i^{\bot} k^t k_i^{\bot} k^t \xi = 0\end{aligned}$$ since $k^t k_i^{\bot} = 0$ by construction, where $k^t$ is the transpose of $k$ . For the second statement we note that $$\begin{aligned} &\sum\hat{Y_{\alpha}}\hat{Y_{\alpha}}({\mbox{$\text{{\upshape}{ev}}$}}_x) = {\mbox{$\text{{\upshape}{ev}}$}}_x\circ P \Big(\hat{X_{\alpha}}(\hat{X_{\alpha}}(\xi)) - \hat{X_{\alpha}}(\nabla f_{\alpha}^{\xi}) \Big)\\ &= {\mbox{$\text{{\upshape}{ev}}$}}_x\circ P\Big( c^s\Delta\xi - \nabla_{X_{\alpha}}(\nabla f_{\alpha}^{\xi}) - X_{\alpha}'\otimes\nabla f_{\alpha}^{\xi} \Big) = {\mbox{$\text{{\upshape}{ev}}$}}_x\circ P\Big( c^s\Delta\xi - \nabla \nabla_{X_{\alpha}}f_{\alpha}^{\xi} \Big) = c^s\Delta P\xi (x)\end{aligned}$$ where $f_{\alpha}^{\xi} = \Delta^{-1}\d(\hat{X_{\alpha}}(\xi))$. \[thm:main\] 1. If $\xi_t\in{\mathfrak{g}}_0$ is a strong solution to on $[0,T]$ such that $\xi_0=u_0\in{\mathfrak{g}}_0$, then $u(t,x) = E[{\mbox{$\text{{\upshape}{ev}}$}}_x(\xi_t)]$ satisfies the Navier-Stokes equation with $u(0,x)=u_0(x)$ for incompressible flow in $M$ on $[0,T]$: $$\begin{aligned} \label{e:nse} {\mbox{$\frac{\partial }{\partial t}$}} u &= -P\nabla_u u + \eta\Delta u \textup{ and } \d\,u = 0\end{aligned}$$ 2. Conversely, suppose $u$ is a smooth solution of on $[0,T]$ with initial condition $u_0$, and $g^u$ is a strong solution of on $[0,T]$ with $g^u_0=e$. Let $\zeta_t := Ad^{\top}(g^u_t).u_0$ where ${\mbox{$\text{\upshape{Ad}}$}}^{\top}$ is defined in . It follows that $u=E[\zeta]$, and $(g_t^u, \zeta_t)$ is a solution of . In particular, $\zeta_t$ solves the mean field system . In Part (1), Equation  is thus the McKean-Vlasov equation for ${\mbox{$\text{{\upshape}{ev}}$}}_x: {\mathfrak{g}}_{{0}}\to\R^n$ corresponding to the non-linear generator of $\xi_t$. In Part (2), we regard $g^u_t$ as a *Gaussian perturbation* of the Lagrangian trajectory defined by $u$. This is justified by Proposition \[prop:bm\]. Ad (1). Note that $\d\,u_t = E[\d\,\xi_t] = 0$. Let us write the Ito version of as $d\xi_t = b(u_t,\xi_t)\,dt - \nu\sum\hat{Y_{\alpha}}(\xi_t)\,dW_t^{\alpha}$ where we use the notation from Lemma \[lem:1\]. By linearity of $\hat{Y_{\alpha}}$ and Lemma \[lem:1\], we have for the quadratic variation $$\begin{aligned} \sum\Big[-\nu\hat{Y_{\alpha}}(\xi_.), W^{\alpha}_. \Big]_t &= -\nu\sum\Big[ \int_0^{\cdot}\hat{Y_{\alpha}}(b(u_s,\xi_s))\, ds - \nu\int_0^{\cdot}\hat{Y_{\alpha}}(\hat{Y_{\alpha}}(\xi_s))\,dW_s^{\beta}, W^{\alpha}_. \Big]_t \\ & = \nu^2\sum\delta_{\alpha\beta}\int_0^t \hat{Y_{\alpha}}(\hat{Y_{\alpha}}(\xi_s)) \;ds = c^s\nu^2\int_0^t \Delta\xi_s\, ds\end{aligned}$$ where $\delta_{\alpha\beta}$ is the Kronecker delta. (See [@Pro] or [@DZ Section 3.4].) Therefore, the Ito version of is $$\label{e:sdeP-Ito} d \xi_t = \Big(-P(\nabla_{u_t}\xi_t + u_t'\otimes\xi_t) + c^s{\mbox{$\frac{\nu^2}{2}$}}\Delta\xi_t \Big)\,dt - \nu \sum P\Big(\nabla_{X_{\alpha}}\xi_t + X_{\alpha}'\otimes\xi\Big) \, dW_t^{\alpha}.$$ Since $\eta = c^s{\mbox{$\frac{\nu^2}{2}$}}$, $P(u'\otimes u) = {\mbox{$\frac{1}{2}$}}P\nabla\vv<u,u> = 0$, and by linearity of $\xi\mapsto b(u,\xi)$, it follows that $$u_t = u_0 + E\Big[\int_0^t\Big( - P(\nabla_{u_s}\xi_s + u_s'\otimes\xi_s ) + \eta\Delta\xi_s \Big)\,ds \Big] = u_0 + \int_0^t\Big(-P\nabla_{u_s}u_s + \eta\Delta u_s\Big)\,ds$$ solves the Navier-Stokes equation on $[0,T]$. Ad (2). Using the notation from , we obtain $A(g_t^u,\zeta_t) = {\mbox{$\text{\upshape{Ad}}$}}^{\top}((g^u_t)^{-1}).{\mbox{$\text{\upshape{Ad}}$}}(g^u_t).u_0 = u_0$. Therefore, $$\begin{aligned} 0 &= \delta \ww<{\mbox{$\text{\upshape{Ad}}$}}((g_t^u)^{-1}).\zeta_t,\eta> = \ww<\delta \zeta_t, {\mbox{$\text{\upshape{Ad}}$}}(g_t^u).\eta> + \ww< \zeta_t, \delta{\mbox{$\text{\upshape{Ad}}$}}(g_t^u).\eta> \notag \\ &=\ww<\delta \zeta_t, {\mbox{$\text{\upshape{Ad}}$}}(g_t^u).\eta> + \ww<\zeta_t, {\mbox{$\text{\upshape{ad}}$}}(u_t\,\delta t + \sum X_{\alpha}\,\delta W_t^{\alpha}).{\mbox{$\text{\upshape{Ad}}$}}(g_t^u).\eta> \notag \\ &=\ww<{\mbox{$\text{\upshape{Ad}}$}}^{\top}((g^u_t)^{-1}). \Big(\delta \zeta_t - {\mbox{$\text{\upshape{ad}}$}}^{\top}(u_t\,\delta t + \sum X_{\alpha}\,\delta W_t^{\alpha}).\zeta_t\Big) , \eta > \label{e:proof-main}\end{aligned}$$ for arbitrary $\eta\in{\mathfrak{g}}_0$, whence $\zeta_t$ satisfies . It remains to show that $E[\zeta_t] = u_t$. Let $v_t = E[\zeta_t]$ and $w = v-u$. Since implies , it follows that $$\label{e:v1} {\mbox{$\frac{\partial }{\partial t}$}}v = -P(\nabla_u v + u'\otimes v) + \eta\Delta v.$$ Notice that $\d\,v = \d\,w = 0$. Subtracting the Navier Stokes equation from yields with $v$ replaced by $w$. Therefore, we have the energy identity $$\label{e:v2} {\mbox{$\frac{\partial }{\partial t}$}}{\mbox{$\frac{1}{2}$}}\ww<w,w> = \ww<-\nabla_u w - u'\otimes w + \eta\Delta w , w> = -\ww<u'\otimes w, w> - \eta\ww<\nabla w, \nabla w>$$ because $\ww<\nabla_u w, w> = 0$ by partial integration. This gives $${\mbox{$\frac{\partial }{\partial t}$}}||w_t||_2^2 \le 2|\ww<u_t'\otimes w_t, w_t>| \le 2\sup_{0\le s\le T}|u'_s|_{\infty}\cdot ||w_t||_2^2$$ where $|u'_s|_{\infty}$ is the supremum norm and $||{w}||_2^2 = \ww<w,w>$ is the $L^2$-norm. The Gronwall Lemma now implies that $||w_t||_2^2 = ||w_0||_2^2 = 0$, whence $v=u$. \[cor:main\] Let $\Phi: C^1([0,T],{\mathfrak{g}}_0)\to C^1([0,T],{\mathfrak{g}}_0)$ be defined by $\Phi: u\mapsto E[{\mbox{$\text{\upshape{Ad}}$}}^{\top}(g^u).u_0]$. Then $u$ is a solution to the Navier-Stokes Equation  if and only if $u$ is a fixed point of $\Phi$. If $u$ is a solution to then $u_t = E[{\mbox{$\text{\upshape{Ad}}$}}^{\top}(g_t^u).u_0]$ by Theorem (2). Conversely, if $\xi_t := {\mbox{$\text{\upshape{Ad}}$}}^{\top}(g_t^u).u_0$ and $u_t = E[\xi_t]$ then $\xi_t$ satisfies the mean field Equation , whence $u$ is a solution of by Theorem (1). Kelvin Circulation Theorem -------------------------- \[prop:KC\] Suppose $\xi_t$ is a solution of with $u_t=E[\xi_t]$. Define $g_t$ in $G_0$ through the reconstruction Equation , that is $$\label{e:del_gamma} \delta g_t = \Big(u_t\delta t + \sum X_{\alpha} \delta W_t^{\alpha}\Big)\circ g_t.$$ Then, for any closed smooth curve $C$ in $M$, $$\label{e:KC} \delta \int_{( g_t)_*C}\xi_t^{\flat} = 0.$$ Since $H^t$ and $F^{\alpha}$ in are invariant under the $G_0$-action, this follows immediately from Proposition \[prop:kct\_gen\]. However, we can also give an explicit proof: $$\begin{aligned} \label{e:kct1} \delta \int_{( g_t)_*C}\xi_t^{\flat} &= \delta\int_C g_t^*\xi_t^{\flat} = \delta\int_C(\xi_t\circ g_t)^{\flat}\circ T g_t = \delta\int_0^1\vv<\xi_t\circ g_t, T g_t.c'(s)>\,ds \\ &= \int_0^1\Big(\vv<\delta(\xi_t\circ g_t),T g_t.c'(s)> + \vv<\xi_t\circ g_t,(Tu_t\,\delta t + \sum TX_{\alpha}\,\delta W_t^{\alpha} )T g_t.c'(s)> \Big)\,ds \notag \\ &= \int_0^1 \vv<\Big( \delta\xi_t + \nabla_{u_t\,\delta t + \sum X_{\alpha}\,\delta W_t^{\alpha}}\xi_t + u_t'\otimes\xi_t \,\delta t + \sum X_{\alpha}'\otimes\xi_t \,\delta W_t^{\alpha} \Big)\circ g_t , T g_t.c'(s) > \,ds \notag \\ &= \int_{( g_t)_*C} (\nabla p_t)^{\flat} = 0 \notag \end{aligned}$$ where $c(s)$ is a parametrization of $C$. Since $$\label{e:line-stretching} u_t'\otimes\xi_t \textup{ and } X_{\alpha}'\otimes\xi_t$$ are obtained by transposing $Tu_t$ and $TX_{\alpha}$, we refer to these as the *line stretching terms*. Compare with [@CFH17 Equ. (2.37)]. Energy dissipation {#sec:energy} ================== The energy $\mathcal{E}^d$ associated to a solution $u$ of the Navier-Stokes equation is $$\label{equ:enedet} \mathcal{E}_t^d = {\mbox{$\frac{1}{2}$}}\ww<u_t, u_t> = {\mbox{$\frac{1}{2}$}}\int_M \vv<u(t,x),u(t,x)>\, dx.$$ The superscript is to emphasize that this is the energy associated with the deterministic solution $u$. Since ${\mbox{$\frac{\partial }{\partial t}$}}\mathcal{E}^d_t = -\eta\ww<\nabla u(t,x),\nabla u(t,x)>$, it follows that energy dissipates for $\eta>0$. For $\xi\in{\mathfrak{g}}_0$ let $$\mathcal{H}_0(\xi) = {\mbox{$\frac{1}{2}$}}\ww<\xi,\xi>$$ We define the energy of a solution $\xi_t$ of Equation  as $$\label{e:ensto} \mathcal{E}^s_t = E[\mathcal{H}_0(\xi_t)].$$ The superscript $s$ stands for stochastic, even though $\mathcal{E}^s_t$ is not a random variable. This section is concerned with two questions: 1. Do conservation or dissipation for $\mathcal{H}_0$ or $\mathcal{E}^s$ hold? 2. Does dissipation of $\mathcal{E}^d$ follow from the stochastic formulation? Concerning the first question, note that Equation  yields $$\begin{aligned} \label{e:non-cons} \delta{\mbox{$\frac{1}{2}$}}\ww<\xi_t,\xi_t> = \ww<{\mbox{$\text{\upshape{ad}}$}}^{\top} (u_t\,\delta t + \nu\sum X_{\alpha}\,\delta W_t^{\alpha}).\xi_t, \xi_t> = -\ww<\xi_t, [u_t, \xi_t]\,\delta t + \nu\sum [X_{\alpha}, \xi_t]\,\delta W_t^{\alpha} >\end{aligned}$$ whence commutation of $\xi_t$ with $u_t$ and $X_{\alpha}$ is an obstruction to stochastic energy conservation. Thus $\mathcal{H}_0$ is not conserved along $\xi_t$. Non-conservation of stochastic energy is ubiquitous stochastic geometric mechanics. This is also discussed in [@CGD17] from a perspective of multi-scale analysis and, in finite dimensions, in [@ACH16]. Non-dissipation {#sec:non-diss} --------------- Suppose $\xi_t$ is a solution of Equation , which is the Ito formulation of . Let $\hat{Y}_{\alpha}(\xi)$ as in Lemma \[lem:1\]. Then $$\label{e:dH} d\ww<\xi_t,\xi_t> = 2\ww<\xi_t,d\xi_t> + \nu^2\sum\ww<\hat{Y}_{\alpha}(\xi_t),\hat{Y}_{\alpha}(\xi_t)>\,dt.$$ Using $\ww<\xi,\nabla f> = 0$, which follows by partial integration from $\d(\xi)=0$, we note that $$\begin{aligned} \ww<\xi_t,d\xi_t> &= \ww<\xi_t, -(\nabla_{u_t}\xi_t + u_t'\otimes\xi_t - \eta\Delta\xi_t)\,dt + \sum\hat{Y}_{\alpha}(\xi_t)\,dW_t^{\alpha} > \\ &= \ww<\xi_t, (-u_t'\otimes\xi_t + \eta\Delta\xi_t )\,dt + \sum\hat{Y}_{\alpha}(\xi_t)\,dW_t^{\alpha} >. \end{aligned}$$ Note that $\ww<\xi_t, u_t'\otimes\xi_t> = \ww<\xi_t, [u_t,\xi_t]>$. Furthermore, $$\begin{aligned} \label{e:crossterms} &\sum\ww<\hat{Y}_{\alpha}(\xi_t),\hat{Y}_{\alpha}(\xi_t)> + \sum\ww<\nabla_{X_{\alpha}}\nabla_{X_{\alpha}}\xi , \xi>\\ &= \sum\ww< X_{\alpha}'\otimes\xi , X_{\alpha}'\otimes\xi> + \sum\ww<-\nabla f_{\alpha}^{\xi} , \nabla_{X_{\alpha}}\xi_t+X_{\alpha}'\otimes\xi_t > \notag \end{aligned}$$ since the cross terms $\sum\ww<\nabla_{X_{\alpha}}\xi, X'_{\alpha}\otimes\xi> = -\ww<\xi,\nabla_{X_{\alpha}}(X'_{\alpha}\otimes\xi)> = 0$ vanish by Equation  and where $\Delta f_{\alpha}^{\xi} = \d(\nabla_{X_{\alpha}}\xi_t+X_{\alpha}'\otimes\xi_t)$. Now, $$\sum\ww<-\nabla f_{\alpha}^{\xi} , \nabla_{X_{\alpha}}\xi_t+X_{\alpha}'\otimes\xi_t > = - \sum \ww<\nabla f_{\alpha}^{\xi} , \nabla f_{\alpha}^{\xi}>$$ and $$\begin{aligned} \label{e:X_alpha} &\sum\ww< X_{\alpha}'\otimes\xi , X_{\alpha}'\otimes\xi>\\ &= \sum_{k\in\mathbb{Z}_n^+}\sum_{i=1}^{n-1}\Big( \ww<B_{(k,i)}\otimes k\otimes\xi , B_{(k,i)}\otimes k\otimes\xi> +\ww<A_{(k,i)}\otimes k\otimes\xi,A_{(k,i)}\otimes k\otimes\xi> \Big) \notag \\ &= \sum_{k\in\mathbb{Z}_n^+}\sum_{i=1}^{n-1} \int_M \frac{\sin^2\vv<k,x>+\cos^2\vv<k,x>}{|k|^{2s+2}} \vv<k_i^{\bot}k^t\xi(x),k_i^{\bot}k^t\xi(x)> \,dx \notag \\ &= \sum_{k\in\mathbb{Z}_n^+}\sum_{i=1}^{n-1} \int_M \frac{\vv<k,\xi(x)>^2|k_i^{\bot}|^2}{|k|^{2s+2}} \,dx = (n-1)\sum_{k\in\mathbb{Z}_n^+} \frac{1}{|k|^{2s}} \int_M \vv<k,\xi(x)>^2 \,dx \notag \end{aligned}$$ where we use that $|k_i^{\bot}| = |k|$. Equation  now yields $$\begin{aligned} {\mbox{$\frac{\partial }{\partial t}$}}\mathcal{E}_t^s(\xi) &= {\mbox{$\frac{1}{2}$}}E\Big[-\ww<\xi_t, [u_t, \xi_t]> + \nu^2(n-1)\sum_{k\in\mathbb{Z}_n^+} \frac{1}{|k|^{2s}} \int_M \vv<k,\xi_t(x)>^2 \,dx - \nu^2\sum \ww<\nabla f_{\alpha}^{\xi} , \nabla f_{\alpha}^{\xi}>\Big] $$ which certainly is not negative in general. We note that the non-dissipation of $\mathcal{E}^s$ is due to the line stretching terms , which are precisely the terms needed to make the Kelvin Circulation Theorem \[prop:KC\] hold. Dissipation {#sec:diss} ----------- Let us now perform the same analysis as in Section \[sec:non-diss\], but with respect to the formulation : Consider, with $\nu=\sqrt{{\mbox{$\frac{2\eta}{c^s}$}}}$ and $u_0\in{\mathfrak{g}}_0$ as above, the mean-field equation $$\begin{aligned} \label{e:sde1} \delta {\mbox{$\widetilde{\zeta}$}}_t &= - \Big(P\nabla_{u_t}{\mbox{$\widetilde{\zeta}$}}_t + \eta\nabla\d\,{\mbox{$\widetilde{\zeta}$}}_t \Big)\,\delta t - \nu \sum \nabla_{X_{\alpha}}{\mbox{$\widetilde{\zeta}$}}_t \,\delta W_t^{\alpha}\\ {\mbox{$\widetilde{\zeta}$}}_0 &= u_0 \notag \\ u_t &= E[{\mbox{$\widetilde{\zeta}$}}_t] \notag\end{aligned}$$ This is now an equation in ${\mathfrak{g}}$. Its solutions are not necessarily divergence free. \[thm:mainH17\] If ${\mbox{$\widetilde{\zeta}$}}_t\in{\mathfrak{g}}$ is a strong solution to on $[0,T]$ such that ${\mbox{$\widetilde{\zeta}$}}_0=u_0\in{\mathfrak{g}}_0$ then $u(t,x) = E[{\mbox{$\text{{\upshape}{ev}}$}}_x({\mbox{$\widetilde{\zeta}$}}_t)]$ satisfies the Navier-Stokes equation with $u(0,.)=u_0$ for $t\in[0,T]$. Further, the Ito version of is $$\label{e:sde2} d {\mbox{$\widetilde{\zeta}$}}_t = \Big(-P\nabla_{u_t}{\mbox{$\widetilde{\zeta}$}}_t - \eta\nabla\d\,{\mbox{$\widetilde{\zeta}$}}_t + \eta\Delta{\mbox{$\widetilde{\zeta}$}}_t \Big)\,dt - \nu \sum \nabla_{X_{\alpha}}{\mbox{$\widetilde{\zeta}$}}_t \, dW_t^{\alpha}.$$ In order to obtain a divergence free solution, we modify Equation  as $$\label{e:sde3} d \zeta_t = (-P\nabla_{u_t}\zeta_t + \eta\Delta\zeta_t )\,dt - \nu \sum P\nabla_{X_{\alpha}}\zeta_t \, dW_t^{\alpha}$$ with $u_t=E[\zeta_t]$ and the same initial condition $\zeta_0=u_0$. Now we have an equation in ${\mathfrak{g}}_0$, such that $ \d\,\zeta = 0 $ and $u_t$ still solves the Navier-Stokes Equation . Indeed, if ${\mbox{$\widetilde{\zeta}$}}_t$ is as in Theorem \[thm:mainH17\], then $\zeta_t = P{\mbox{$\widetilde{\zeta}$}}_t$ and $u_t = E[{\mbox{$\widetilde{\zeta}$}}_t] = PE[{\mbox{$\widetilde{\zeta}$}}_t] = E[\zeta_t]$. \[prop:EnergyDiss\] Assume $\zeta_t$ is a solution of . Let $P\nabla_u\zeta = \nabla_u\zeta + \nabla p$, $P\nabla_{X_{\alpha}}\zeta = \nabla_{X_{\alpha}}\zeta + \nabla q^{\alpha}$ and $P\nabla_{X_{\alpha}}u = \nabla_{X_{\alpha}}u + \nabla q_u^{\alpha}$. Let $\mathcal{E}^s_t = {\mbox{$\frac{1}{2}$}}E[\ww<\zeta_t,\zeta_t>]$. 1. $\mathcal{E}^s_t = E[\mathcal{H}_0(\zeta_t)]$ is dissipative: ${\mbox{$\frac{\partial }{\partial t}$}}\mathcal{E}_t^s = -{\mbox{$\frac{\nu^2}{2}$}}\sum_{\alpha}\ww<\nabla q^{\alpha},\nabla q^{\alpha}> \le -{\mbox{$\frac{\nu^2}{2}$}}\sum_{\alpha}\ww<\nabla q_u^{\alpha},\nabla q_u^{\alpha}> < 0$, for $\nu>0$. 2. $\mathcal{E}^s_t\ge {\mbox{$\frac{1}{2}$}}\ww<E[\zeta_t],E[\zeta_t]> = \mathcal{E}_t^d$. 3. $\mathcal{E}^d_t$ is monotone decreasing in $t$. In particular, (deterministic) energy dissipation for the Navier-Stokes equation follows from the stochastic formulation , without using the PDE setting. Ad (1). Observe that $$\ww<\zeta, P\nabla_u\zeta> = \ww<\zeta, \nabla_u\zeta + \nabla p> = \int_M\Big({\mbox{$\frac{1}{2}$}}\d(\vv<\zeta,\zeta>u)+\d(p\zeta)\Big)\,dx = 0$$ and $$\begin{aligned} &\nu^2\ww<P\nabla_{X_{\alpha}}\zeta,P\nabla_{X_{\alpha}}\zeta> = \nu^2\ww<\nabla_{X_{\alpha}}\zeta, \nabla_{X_{\alpha}}\zeta + \nabla q^{\alpha}>\\ &= \nu^2\int_M\Big(\d\Big(\vv<\zeta,\nabla_{X_{\alpha}}>X_{\alpha}\Big) - \vv<\zeta,\nabla_{X_{\alpha}}\nabla_{X_{\alpha}}\zeta>\Big)\,dx +\nu^2\int_M\Big(\d\Big(q_{\alpha}\nabla_{X_{\alpha}}\zeta\Big) - q_{\alpha}\d\Big(\nabla_{X_{\alpha}}\zeta\Big)\Big)\,dx \\ &= -2\eta\ww<\zeta,\Delta\zeta> + \nu^2\ww<q_{\alpha},\Delta q_{\alpha}>.\end{aligned}$$ It follows that $$\begin{aligned} \label{e:diss1} &d\ww<\zeta_t,\zeta_t> = 2\ww<\zeta_t,d\zeta_t> + \sum_{\alpha}\ww<P\nabla_{X_{\alpha}}\zeta,P\nabla_{X_{\alpha}}\zeta>dt \\ &= 2\ww<\zeta_t, (-P\nabla_{u_t}\zeta_t + \eta\Delta\zeta_t)dt - \sum_{\alpha}P\nabla_{X_{\alpha}}\zeta\,dW_t^{\alpha}> - 2\eta\ww<\zeta_t,\Delta\zeta_t> dt - \nu^2\sum_{\alpha}\ww<\nabla q^{\alpha},\nabla q^{\alpha}>dt \notag\end{aligned}$$ and, therefore, $d\ww<\zeta_t,\zeta_t> = - \nu^2\sum_{\alpha}\ww<\nabla q^{\alpha},\nabla q^{\alpha}>dt$ as-well as $${\mbox{$\frac{\partial }{\partial t}$}}\mathcal{E}_t^s = -{\mbox{$\frac{\nu^2}{2}$}}\sum_{\alpha}E[\ww<\nabla q^{\alpha},\nabla q^{\alpha}>].$$ Now, the Jensen inequality implies that $$E[\ww<\nabla q^{\alpha},\nabla q^{\alpha}>] \ge \ww<E[\nabla q^{\alpha}],\nabla E[q^{\alpha}]> = \ww<\nabla q_u^{\alpha},\nabla q_u^{\alpha}>$$ which is non-zero since $\d\,\nabla_{X_{\alpha}}u = {\mbox{$\textup{Tr}$}}(u'\otimes X_{\alpha}')\neq0$ for non-trivial solutions $u$ and general $\alpha$. Ad (2). This follows from the Jensen inequality. Ad (3). This has to be proved separately, since being bounded by a decreasing function does not imply being (locally) decreasing. Let $t_0<t_1$ (such that solutions $\zeta$ and $u$ exist on an interval containing these two points). We can condition to start at time $t_0$ with initial condition $\zeta_{t_0} = u_{t_0}$. Then $\mathcal{E}^d_{t_0}=\mathcal{E}^s_{t_0} > \mathcal{E}^s_{t_1}\ge\mathcal{E}^d_{t_1}$. We remark that the dissipation of the stochastic energy $\mathcal{E}^s$ follows not, as in the deterministic case, from the Laplacian in the drift, but from the quadratic variation of the martingale part in . One may ask whether $\mathcal{E}^s$ decreases as quickly as $\mathcal{E}^d$. To this end, note that ${\mbox{$\frac{\nu^2}{2}$}}$ equals $\eta$ up to the $c^s$-factor, and $$\begin{aligned} \ww<\nabla q_u^{\alpha},\nabla q_u^{\alpha}> &= \ww<q_u^{\alpha},\d\,\nabla_{X{\alpha}}u> = \ww<\Delta^{-1}({\mbox{$\textup{Tr}$}}(u'\otimes X_{\alpha}'), {\mbox{$\textup{Tr}$}}(u'\otimes X_{\alpha}')>.\end{aligned}$$ Hence the stochastic energy dissipates at a rate that is proportional to the $L^2$-square of $u'$, just as in the deterministic case. Interacting particle system (IPS) {#sec:IPS} ================================= Numerical algorithms for the simulation of mean field equations can be devised by means of IPS approximations. This is true, both, in the finite and infinite dimensional setting. See [@AD95; @delmoral; @Gov]. The mean field formulation of [@CI05] has been used in [@IM08; @IN11; @NS17] to derive and study Lagrangian formulations of associated interacting particle systems. One of the key issues in these articles is the fact that the energy of the approximating particle system does not completely dissipate (even if the solution is defined globally in time). Hence [@NS17] use a resetting technique, inspired by [@IN11], in their approximation scheme. Below we use equations and to discuss our version of an approximating interacting particle system. Version 1 – ------------ Fix a large integer $N$. The approximating IPS for the mean field Equation  is . The Ito version of is $$\begin{aligned} \label{e:ips1} d\xi^i &= -P\Big(\nabla_{u^N}\xi^i + (u^N)'\otimes\xi^i\Big)\,d t +\eta\Delta\xi^i\,d t - \nu\sum P\Big(\nabla_{X_{\alpha}}\xi^i + X_{\alpha}'\otimes\xi^i\Big)\,d W^{\alpha,i} \\ u^N &= {\mbox{$\frac{1}{N}$}}\sum\xi^i \notag\\ \xi^i(0) &= u_0 \notag \end{aligned}$$ where $W^{i,\alpha}$ is a sequence of pairwise independent Brownian motions and $i=1,\dots,N$. Equation  is therefore the approximating IPS for the mean field system . We have $\d(\xi^i)=0=\d(u^N)$. Note that the empirical average $u^N_t$ is a stochastic process. Let $R_t$ denote the martingale part of $\mathcal{H}_0(u^N_t) = {\mbox{$\frac{1}{2}$}}\ww<u^N_t,u^N_t>$. Equation  applied to $u^N$ yields $$\begin{aligned} \label{e:version1} &d\mathcal{H}_0(u^N) - dR\\ &= \ww<-P\Big(\nabla_{u^N}u^N + (u^N)'\otimes u^N\Big) + \eta\Delta u^N, u^N>\,dt + {\mbox{$\frac{\nu^2}{2}$}}\sum\ww<\hat{Y}_{\alpha}(u^N),\hat{Y}_{\alpha}(u^N)> \,dt \notag\\ &= \ww<u^N,\eta\Delta u^N>\,dt + {\mbox{$\frac{\nu^2}{2N^2}$}}\sum_{i}\Big( \ww<-c^s\Delta\xi^i,\xi^i> + \sum\ww<X_{\alpha}' \otimes\xi^i, X_{\alpha}' \otimes\xi^i> - \sum\ww<\nabla f_{\alpha}^i,\nabla f_{\alpha}^i> \Big) \,dt \notag\\ &= -{\mbox{$\frac{\eta}{N^2}$}}\sum_{i\neq j} \ww<\nabla\xi^i,\nabla\xi^j >\,dt + {\mbox{$\frac{\nu^2}{2N^2}$}}\sum_{\alpha,i}\Big( \ww<X_{\alpha}' \otimes\xi^i, X_{\alpha}' \otimes\xi^i> - \ww<\nabla f_{\alpha}^i,\nabla f_{\alpha}^i> \Big)\,dt \notag\end{aligned}$$ where $\Delta f_{\alpha}^i = -\d(\nabla_{X_{\alpha}}\xi^i + X_{\alpha}'\otimes\xi^i)$. The term $-{\mbox{$\frac{\eta}{N^2}$}}\sum_{i\neq j} \ww<\nabla\xi^i,\nabla\xi^j >$ corresponds to [@NS17 Equation (2.8)]. The other two terms are due to the fact that the $X_{\alpha}$ are non-constant, which is a difference between and [@IM08 Lemma 4.2] that is due to the infinite dimensional approach. An explicit formula for the middle term of the right hand side is given in . Version 2 – ------------ Consider now the approximating IPS for Equation : $$\begin{aligned} \label{e:ips2} \delta\zeta^i &= -P\nabla_{u^N}\zeta^i \,dt +\eta\Delta\zeta^i\,dt - \nu\sum P\nabla_{X_{\alpha}}\zeta^i \,dW^{i,\alpha} \\ u^N &= {\mbox{$\frac{1}{N}$}}\sum\zeta^i \notag\\ \zeta^i(0) &= u_0 \notag \end{aligned}$$ where $W^{i,\alpha}$ is a sequence of pairwise independent Brownian motions and $i=1,\dots,N$. Again we have $\d(\zeta^i)=0=\d(u^N)$. Let $R_t$ denote the martingale part of $\mathcal{H}_0(u^N_t) = {\mbox{$\frac{1}{2}$}}\ww<u^N_t,u^N_t>$. The calculation  now implies $$\begin{aligned} \label{e:version2} d\mathcal{H}_0(u^N) - dR &= -{\mbox{$\frac{\eta}{N^2}$}}\sum_{i\neq j} \ww<\nabla\zeta^i,\nabla\zeta^j >\,dt - {\mbox{$\frac{\nu^2}{2N^2}$}}\sum_{\alpha,i} \ww<\nabla f_{\alpha}^i,\nabla f_{\alpha}^i> \,dt \end{aligned}$$ where $\Delta f_{\alpha}^m = -\d(\nabla_{X_{\alpha}}{\zeta}^m)$. Both formulations, and , involve cross terms of the type $\sum_{i\neq j} \ww<\nabla\zeta^i,\nabla\zeta^j >$. This leads to the non-dissipation of the expectation $E[\mathcal{H}_0(u_t^N)]$ that has been observed in [@IM08; @IN11; @NS17]. Comparison with other approaches {#sec:comp} ================================ Stochastic Lagrangian least action principle – Cipriano and Cruzeiro [@CC07; @C] {#sec:least-action} -------------------------------------------------------------------------------- Stochastic variational principles as a means to characterize solutions to the Navier-Stokes equation have already been considered by [@Y83]. In [@CC07; @C] solutions of the Navier-Stokes equation are characterized as solutions to a stochastic variational principle defined on the group of volume preserving homeomorphisms $G^0_0$. They consider Lagrangian curves of the form $$\tag{\cite[Equ.~(3.1)]{CC07}} \delta g^u_t = TR^{g^u_t} \Big(u_t \,\delta t + \nu X_{\alpha}\,\delta W_t^{\alpha}\Big), \; g^u_0 = e$$ where $u_t$ is a time dependent vector field, and prove that $u$ solves the Navier-Stokes equation if and only if $g^u_t$ is a solution to a stochastic variational principle ([@CC07 Theorem 4.1]). Note that [@CC07 Equ. (3.1)] coincides with our Equation . By Theorem \[thm:main\], $u$ solves the Navier-Stokes equation if and only if $(g_t^u,{\mbox{$\text{\upshape{Ad}}$}}^{\top}(g^u_t).u_u)$ is a solution to the Hamiltonian Equation . Therefore, the stochastic variational principle of [@CC07] is equivalent to the Hamiltonian mean field formulation \[e:sdeP\]. However, we remark that the regularity assumptions of Theorem \[thm:main\] are stronger than those of [@CC07 Theorem 4.1]. Thus we have proved this equivalence only under the stronger regularity assumptions on the solutions $u$ and $g^u$. Comparison with Constantin and Iyer [@CI05] {#sec:CI} ------------------------------------------- In [@CI05] the stochastic vector $w = A'\otimes(u_0\circ A)$ is defined; in this definition $A=X^{-1}$, where $X=X(t,x)$ is the Lagrangian flow map, i.e. a volume preserving diffeomorphism for each $t$, and the divergence free vector field $u_0$ is the initial condition. As before, we make use of the notation $ A'\otimes v = \sum({\partial}_i A^j)v^j e_i = (\nabla^t A)v $ where $(e_i)$ is the standard basis in $\R^3$. Let $B$ denote three-dimensional Brownian motion. In [@CI05 Theorem 2.2] it is assumed that the pair $(X,u)$ satisfies the mean field Ito SDE $$dX = u\,dt + \sqrt{2\eta}\,dB,\; u = E[Pw],\; X(0,x) = x$$ and concluded that $u$ is a solution of the incompressible Navier-Stokes equation with initial data $u_0$ and viscosity coefficient $\eta$. In order to do so, they derive the system $$\tag{\cite[Equation~4.5]{CI05}} dw_t = \Big(-\nabla_{u_t} w_t + \eta \Delta w_t - u_t'\otimes w_t\Big)\,dt + \sqrt{2\eta}\nabla w_t\,d{B}_t.$$ Now, [@CI05] construct the mean Eulerian velocity as $u = PE[w]$. It is further shown that a Kelvin Circulation Theorem holds along stochastic Lagrangian paths: $$\tag{\cite[Prop.~2.9]{CI05}} \oint_{X_t(C)}Pw_t = \oint_C u_0$$ where $C$ is a closed curve in $M$. The term $u'\otimes\xi$ in is also present in [@CI05 Equation (4.5)] and it is this term which makes the Kelvin Circulation Theorem work. In our approach we also have to include the $X_{\alpha}'\otimes\xi$ which is not needed in the finite dimensional setting of [@CI05]. Thus Equation  is an infinite dimensional analogue of [@CI05 Equation (4.5)]. Comparison with Holm et al. [@Holm15; @CFH17] {#sec:Holm} --------------------------------------------- Consider vector fields $u_t$ and ${\mbox{$\widetilde{X}$}}_{\alpha}$ on $\mathbb{T}^3$ where $u_t$ is time dependent. Let, as above, $W^{\alpha}$ be a sequence of mutually independent Brownian motions. Consider the Stratonovich SDE $$\tag{\cite[Equ.~(2.29)]{CFH17}} \delta y_t = u_t\,\delta t + \nu\sum \widetilde{X}_{\alpha} \,\delta W_t^{\alpha}$$ where we use the notation $y_t$ to be consistent with [@CFH17] and add the parameter $\nu$. We remark that the $\widetilde{X}_{\alpha}$ correspond to the deterministic vector fields $\xi_i$ in [@CFH17 Equ. (2.29)], while we use $\xi_t$ in this paper to denote the stochastic [momentum]{} field. Let $dx$ denote the standard volume element in ${\mathbb{T}}^3$. Then [@CFH17] use a stochastic version of Newton’s second law to derive the $3D$ *stochastic Euler equation* $$\tag{\cite[Equ.~(2.40)]{CFH17}} (\delta + {\mbox{$\text{\upshape{ad}}$}}^*(\delta y_t)).\psi_t\otimes dx = -dp_t\otimes dx \,\delta t.$$ where $\psi_t\otimes dx$ is now a stochastic curve of one-form densities and $p_t$ is the pressure. For $\eta\in{\mathfrak{g}}$ we have the dual pairing $\vv<\psi_t\otimes dx,\eta> = \int\vv<\psi_t(x),\eta(x)>\,dx$, and ${\mbox{$\text{\upshape{ad}}$}}^*$ is defined with respect to this pairing as in Section \[sec:InOp\]. Since $dx$ is constant, we can reformulate the stochastic Euler equation as an equation in ${\mathfrak{g}}_0^*$: $$\label{e:stoch_euler_equ} (\delta + {\mbox{$\text{\upshape{ad}}$}}^*(\delta y_t)).[\psi_t] = 0$$ where we use again the notation from Section \[sec:InOp\]. Assume that ${\mbox{$\widetilde{X}$}}_{\alpha} = X_{\alpha}$. Note that this means that $\delta y_t = \delta g_t \circ g_t^{-1}$ in the terminology of . Therefore, under these assumptions, the following are equivalent for a stochastic process $\zeta_t$ in ${\mathfrak{g}}_0$ with $E[\zeta_t] = v_t$: 1. The stochastic Newton law [@CFH17 Equ. (2.36)], with constant density $\rho=1$ and where the only force on the fluid is the gradient of the pressure, holds as a one-form relation for $\zeta_t^{\flat}$. 2. The stochastic Euler equation [@CFH17 Equ. (2.40)] holds for $\zeta_t^{\flat}\otimes dx$ in the space of one-form densities. 3. Equation  holds with respect to $u$: $\delta \zeta_t = {\mbox{$\text{\upshape{ad}}$}}^{\top}(u_t).\zeta_t\,\delta t + \nu\sum_{\alpha}{\mbox{$\text{\upshape{ad}}$}}^{\top}(X_{\alpha}).\zeta_t\,\delta W_t^{\alpha}$. 4. $\zeta_t = {\mbox{$\text{\upshape{Ad}}$}}^{\top}(g).u_0$ is the stochastic [momentum]{} along $g_t$. Indeed, the equivalence of (1) and (2) is shown by [@CFH17 Equ. (2.37)]. Equivalence of (2) and (3) follows from the reformulation and the identity $\check{\mu}\circ{\mbox{$\text{\upshape{ad}}$}}^{\top} = {\mbox{$\text{\upshape{ad}}$}}^*\circ\check{\mu}$. Equivalence of (3) and (4) is Equation . If (1) - (4) hold then, by Corollary \[cor:main\], $u$ is a solution to the Navier-Stokes Equation  if and only if $u=v$. This is the case if and only if (2) and (3) are mean field equations, where (3) coincides with the system . In particular, Proposition \[prop:KC\] can also be seen as a direct consequence of the stochastic Kelvin Circulation Theorem in [@Holm15]. Comparison with [@H17] {#sec:H17} ---------------------- Equations and are, at first sight, very similar, and so are the respective conclusions of Theorems \[thm:mainH17\] and \[thm:main\]. However, carrying out the calculation with respect to $\zeta_t$, a solution of , immediately shows that the line stretching terms $u'\otimes\zeta$ and $X'_{\alpha}\otimes\zeta$ do not have a counterpart in $\delta \zeta$, whence there is no cancellation. We conclude that the Kelvin Circulation Theorem does not hold for $\zeta$. Furthermore, is an equation in ${\mathfrak{g}}$. Thus, from a structural point of view, and are very different. In [@Holm02] it is argued that conservation of circulation is a quality criterion for any proposed model of fluid mechanics. Given Proposition \[prop:kct\_gen\], we can extend this argument as follows: The natural phase space of incompressible fluid mechanics is $G_0\times{\mathfrak{g}}_0$. If the model is Hamiltonian with respect to a (possibly vector valued) function $\mathcal{H}$, then $\mathcal{H}$ should be invariant under the particle relabeling symmetry, which is given by right action of $G_0$ on itself. Thus, by Proposition \[prop:kct\_gen\], if a model does not satisfy the Kelvin Circulation Theorem it cannot be a Hamiltonian system. This is why we prefer compared to . In fact, the Hamiltonian structure is given by . We call this structure the *Hamiltonian approach*. Equation  was derived in [@H17 Sec. 1] by using a stochastic version of the material derivative along Lagrangian paths. While this is mathematically correct, the resulting structural properties and physics are different, as argued in the previous paragraph. In this regard it is worth noticing that this *material derivative approach* and the Hamiltonian approach are indeed equivalent in the deterministic case: This follows immediately from the observation that the deterministic version of the line stretching terms is $u'\otimes u = {\mbox{$\frac{1}{2}$}}d\vv<u,u>$, which vanishes upon integration over closed loops. Thus the two approaches differ, in the deterministic case, by a full differential, and this difference is absorbed by the pressure term. See also Section \[sec:memin\]. Comparison with Mémin et al. [@Mem14; @RMC17] {#sec:memin} --------------------------------------------- The approach of [@Mem14; @RMC17] is also based on the material derivative approach. Analogously to Equation , they assume Lagrangian particle trajectories given by an Ito SDE of the form $$\tag{\cite[Equ.~(5)]{RMC17}} dX_t = w(t,X_t)\,dt + \sigma(t,X_t)\,dB_t$$ where $B_t$ is $n$-dimensional Brownian motion. However, their material derivative approach is not derived from a direct stochastic perturbation of the deterministic argument as in [@H17 Sec. 1], but from a stochastic Reynold’s Transport Theorem applied to scalar quantities. Since the transformation of scalar quantities does not involve line stretching terms, the difference between the Hamiltonian and the material derivative approach disappears, as explained in Section \[sec:H17\]. Indeed, various versions of a stochastic Navier-Stokes equation are derived in this setting ([@Mem14 Equ. (43)] and [@RMC17 Equ. (38)]). Compared to the Hamiltonian approach  there are the following differences: 1. The Kelvin Circulation Theorem does not hold for [@Mem14 Equ. (43)] or [@RMC17 Equ. (38)]. This is due to the same reasons that also apply to the model , as explained in Section \[sec:H17\]. 2. [@Mem14 Sec. 5] uses a stochastic version of the stress tensor to model shear forces, whereas is derived without the stress tensor. 3. Energy is conserved exactly along stochastic paths for the model of [@Mem14; @RMC17]. However, this does not hold for Equation , due to the line stretching terms as shown by . Nor does it hold for or , due to the non-solenoidal character of $\nabla_{X_{\alpha}}\zeta$ as shown by . Conclusions {#sec:conc} =========== We have constructed a Hamiltonian interacting particle system  for the description of an ensemble of fluid particles from the assumption that the energy separates into a deterministic and a stochastic part. The deterministic component is responsible for the overall drift of the ensemble, while the stochastic part models the molecular transfer of momentum between fluid layers of different velocities. It is this molecular transfer that is responsible for shear forces which are thus accounted for in the interacting particle system (IPS). In the continuum limit, as the number of particles goes to infinity, this IPS yields a mean field [SDE]{} . The mean field system has the following properties: - It is equivalent to a stochastic Hamiltonian system  with respect to the natural symplectic structure on the phase space $G_0\times{\mathfrak{g}}_0$ of incompressible fluid mechanics. - Averaging over solutions of , which are *stochastic [momenta identified with elements in velocity space]{}*, implies a solution to the Navier-Stokes Equation . Moreover, any Gaussian perturbation  of a Lagrangian trajectory corresponding to a solution of the Navier-Stokes equation yields a mean-field system of the form , and the mean field coincides with the original solution to the Navier-Stokes equation. (Theorem \[thm:main\]) - Consider the average over stochastic [momenta]{} along a Gaussian perturbation  of a divergence-free time dependent vector field. By Corollary \[cor:main\], solutions to the Navier-Stokes equation can be characterized as a fixed point of this assignment. - There are the following analogies to the case of ideal fluids: - The Hamiltonian structure implies conservation of circulation along stochastic paths in the phase space where one jointly considers configuration and [momentum]{} variables. Since these variables cannot be separated in the formulation of the Kelvin Circulation Theorem, this does not imply conservation of circulation for the expectation of the stochastic [momentum]{} – i.e., for a solution of the Navier-Stokes equation. - The Navier-Stokes equation is derived from a (stochastic) Hamiltonian system without use of a stress tensor. - Energy is neither conserved nor dissipating (Section \[sec:non-diss\]). However, the approach of [@H17] can be slightly modified to yield a system which is dissipative (Section \[sec:diss\]). Finally, we have contrasted the model  with existing approaches from the literature and established several connections. [Our model should also be compared with [@Holm02a; @HT12] where (deterministic) advection of fluid microstructure is studied. The *Kinematic Sweeping Ansatz* of [@HT12] is to consider deterministic turbulence parameters which are swept along a mean flow, while is a model for stochastic particles along a mean flow. A detailed comparison of these approaches is a task for future research.]{} [99]{} V.I. Arnold, *Sur la géométrie différentielle de groupes de Lie de dimension infinie et ses applications à l’hydrodynamique des fluides parfaits*, Annales de l’Institut Fourier **16** No. 1 (1966), p. 319-361. V. Arnold, B. Khesin, *Topological Methods in Hydrodynamics*, Springer 1998. A. Arnaudon, A.L. De Castro, D.D. Holm, *Noise and dissipation on coadjoint orbits* J. Nonlinear Sci. **28**, No. 1, pp. 91-145 (2018). N.U. Ahmed, X. Ding, *A semilinear McKean-Vlasov stochastic evolution equation in Hilbert space*, Stoch. Proc. Appl. **60** (1995) pp. 65-85. A. Chorin, J. Marsden, *Mathematical introduction to fluid mechanics*, Springer 1993. C. J. Cotter, G. A. Gottwald, D. D. Holm, *Stochastic partial differential fluid equations as a diffusive limit of deterministic Lagrangian multi-time dynamics* Proc. R. Soc. A **473** (2017) 20170388. F. Cipriano, A.B. Cruzeiro, *Navier-Stokes equation and diffusions on the group of homeomorphisms of the torus*, Comm. Math. Phys., **275** Issue 1 (2007), pp 255-269. P. Constantin, G. Iyer, *A stochastic Lagrangian representation of the three-dimensional incompressible Navier-Stokes equations*, Comm. Pure Appl. Math. **61** (2008), no. 3, 330-345. D. Crisan, F. Flandoli, D.D. Holm, *Solution properties of a 3D stochastic Euler fluid equation*, `arXiv:1704.06989`. A.B. Cruzeiro, *Hydrodynamics, probability and the geometry of the diffeomorphism group*, Seminar on Stochastic Analysis, Random Fields and Applications VI pp. 83-93 (2011). A.B. Cruzeiro, E. Shamarova, *Navier-Stokes equations and [forward-backward]{} SDEs on the group of diffeomorphisms of a torus*, **119** Issue 12 (2009), pp. 4034-4060. G. Da Prato, J. Zabczyk, *Stochastic equations in infinite dimensions*, CUP, 2nd Ed. 2014. P. Del Moral, *Mean-field simulation for Monte-Carlo integration*, Chapman Hall 2016. G.L. Eyink, *Stochastic least-action principle for the incompressible Navier-Stokes equation*, Physica D **239**, No. 14, 1236-1240 (2010). D. Ebin, J.E. Marsden, *Groups of diffeomorphisms and the motion of an incompressible fluid*, Ann. of Math. **92**(1) (1970), pp. 1037-1041. T. Govindan, *Yosida Approximations of stochastic differential equations in infinite dimensions and applications*, Springer 2016. S. Hochgerner, *Stochastic mean-field approach to fluid dynamics*, J. Nonlinear Science **28** No. 2 (2018) pp. 725-737. [ D. Holm, *Euler-Poincaré dynamics of perfect complex fluids*, In *Geometry, Mechanics, and Dynamics: in honor of the 60th birthday of Jerrold E. Marsden* edited by P. Newton, P. Holmes and A. Weinstein, Springer, pp. 113-167 (2002). ]{} , *The Euler-Poincaré variational framework for modeling fluid dynamics*, In: *Geometric Mechanics and Symmetry: The Peyresq Lectures*, edited by: J. Montaldi and T. Ratiu, LMS LNS 306, CUP 2005. , *Variational principles for stochastic fluid dynamics* Proc. R. Soc. A **471** (2015) 20140963. [ D. Holm and C. Tronci, *Multiscale turbulence models based on convected fluid microstructure*, J. Math. Phys. 53: 115614 (2012). ]{} G. Iyer, *A stochastic perturbation of inviscid flows*, Commun. Math. Phys. **266**, No. 3, 631-645 (2006). G. Iyer, J. C. Mattingly, *A stochastic-Lagrangian particle system for the Navier-Stokes equations*, Nonlinearity **21** pp. 2537-2553 (2008). G. Iyer, A. Novikov, *The regularizing effects of resetting in a particle system for the Burgers’ equation*, Ann. Probab. **39** No. 4, pp. 1468-1501 (2011). J.-A. Lazaro-Cami, J.-P. Ortega, *Stochastic Hamiltonian dynamical systems*, Rep. Math. Phys. **61**, Issue 1 (2008), Pages 65-122. J. Marsden, D. Ebin, A. Fischer, *Diffeomorphism groups, hydrodynamics and relativity*, Proc. of the 13th Biennial Seminar of Canadian Mathematical Congress, (J. Vanstone, ed.), (1972), 135-279 S. Méléard, *Asymptotic behaviour of some interacting particle systems; McKean-Vlasov and Boltzmann models* In: Talay D., Tubaro L. (eds) Probabilistic Models for Nonlinear Partial Differential Equations. Lecture Notes in Mathematics, vol 1627. Springer (1996). E. Mémin, *Fluid flow dynamics under location uncertainty*, Geophysical & Astrophysical Fluid Dynamics, **108**(2): pp. 119-146, (2014). Peter W. Michor, *Some Geometric Evolution Equations Arising as Geodesic Equations on Groups of Diffeomorphism, Including the Hamiltonian Approach.* IN: Phase space analysis of Partial Differential Equations. Series: Progress in Non Linear Differential Equations and Their Applications, Vol. 69. Bove, Antonio; Colombini, Ferruccio; Santo, Daniele Del (Eds.). Birkhauser Verlag 2006. Pages 133-215. A. Novikov, K. Shikh Khalil, *A Stochastic Lagrangian particle system for the Navier-Stokes equations*, `arXiv:1709.01536`. P.E. Protter, *Stochastic integration and differential equations*, Stochastic Modelling and Applied Probability, 2nd Ed., Springer 2005. V. Resseguier, E. Mémin, B. Chapron, *Geophysical flows under location uncertainty, Part I: Random transport and general models*, Geophysical & Astrophysical Fluid Dynamics **111**(3): pp. 149-176 (2017). V. Resseguier, E. Mmin, B. Chapron, *Geophysical flows under location uncertainty, Part II: Quasigeostrophic models and efficient ensemble spreading*, Geophysical & Astrophysical Fluid Dynamics **111**(3): pp. 177-208 (2017). V. Resseguier, E. Mémin, B. Chapron, *Geophysical flows under location uncertainty, Part III: SQG and frontal dynamics under strong turbulence*, Geophysical & Astrophysical Fluid Dynamics **111**(3): pp. 209-227 (2017). A. Sznitman, *Topics in propagation of chaos*, In: Hennequin (ed.) Ecole d’Eté de Probabilités de Saint-Flour XIX — 1989. LN in Math., vol 1464. Springer (1991). K. Yasue, *A variational priciple for the Navier-Stokes equation*, J. Funct. Anal. **51**(2) (1983) pp. 133-141.

ArXiv

--- author: - 'S. Anderl, V. Guillet, G. Pineau des Forêts' - 'D. R. Flower' bibliography: - 'GG\_biblio.bib' date: 'Received 4 March 2013; accepted 19 June 2013' subtitle: 'IV. Effects of grain-grain processing on molecular line emission' title: Shocks in dense clouds --- [Grain-grain processing has been shown to be an indispensable ingredient of shock modelling in high density environments. For densities higher than $\sim$10${}^5$ cm${}^{-3}$, shattering becomes a self-enhanced process that imposes severe chemical and dynamical consequences on the shock characteristics. Shattering is accompanied by the vaporization of grains, which can, in addition to sputtering, directly release SiO to the gas phase. Given that SiO rotational line radiation is used as a major tracer of shocks in dense clouds, it is crucial to understand the influence of vaporization on SiO line emission.]{} [We extend our study of the impact of grain-grain processing on C-type shocks in dense clouds. Various values of the magnetic field are explored. We study the corresponding consequences for molecular line emission and, in particular, investigate the influence of shattering and related vaporization on the rotational line emission of SiO.]{} [We have developed a recipe for implementing the effects of shattering and vaporization into a 2-fluid shock model, resulting in a reduction of computation time by a factor $\sim$100 compared to a multi-fluid modelling approach. This implementation was combined with an LVG-based modelling of molecular line radiation transport. Using this combined model we calculated grids of shock models to explore the consequences of different dust-processing scenarios. ]{} Grain-grain processing is shown to have a strong influence on C-type shocks for a broad range of magnetic fields: the shocks become hotter and thinner. The reduction in column density of shocked gas lowers the intensity of molecular lines, at the same time as higher peak temperatures increase the intensity of highly excited transitions compared to shocks without grain-grain processing. For OH the net effect is an increase in line intensities, while for CO and H${}_2$O it is the contrary. The intensity of H${}_2$ emission is decreased in low transitions and increased for highly excited lines. For all molecules, the highly excited lines become sensitive to the value of the magnetic field. Although vaporization increases the intensity of SiO rotational lines, this effect is weakened by the reduced shock width. The release of SiO early in the hot shock changes the excitation characteristics of SiO radiation, although it does not yield an increase in width for the line profiles. To significantly increase the intensity and width of SiO rotational lines, SiO needs to be present in grain mantles. Introduction ============ Shocks are ubiquitous in the interstellar medium, occurring when matter moves into a more rarefied medium at a velocity that exceeds the local sound speed. Depending on the value of the local magnetosonic speed, different types of shocks can be distinguished. The classical shocks are faster than any signal speed in the shocked medium, so the preshock medium is not able to dynamically respond to the shockwave before it arrives. This type of shock is called “J-type” (see e.g. @Hollenbach:1979p19151 [@McKee:1980p19192]). A different situation can occur if a low degree of ionization and the presence of a magnetic field allow magnetosonic waves to precede the shock. Ions then decouple from the neutrals and are already accelerated in the preshock gas, so that there is no longer a discontinuity in the flow of the ion fluid. In the preshock gas, the ions heat and accelerate the neutrals and broaden the heating region, so that heating and cooling take place simultaneously. The shock transition can then become continuous in the neutral fluid (“C-type shocks”) also. Shocks of this type are thicker and less hot than J-type shocks (see @Draine:1980p539 [@Draine:1993p1759]). Observations often reveal shocks as bow-shaped structures, with the ambient material being compressed and pushed aside (e.g. @Nissen:2007p19095 [@Davis:2009p19094]). Shocks play an important role in the energy budget of the interstellar medium by determining the energetic feedback of events such as supernovae, stellar winds, cloud-cloud collisions, or expanding HII regions. On the other hand, shocks have a major influence on the chemistry of the interstellar medium. Among the most characteristic chemical tracers of shock waves are species that are typically heavily depleted on dust grains, such as Fe, Mg, or Si. Dust processing occurring in the violent environment of shocked media is able to release these species into the gas phase (e.g. @Liffman:1989p17725 [@ODonnell:1997p19216]). The understanding of dust processing in shocks is therefore intimately linked with the theoretical interpretation of characteristic emission lines in environments where shocks occur. The processing of dust in shocks can have two different consequences. It either changes the dust-to-gas mass ratio or alters the dust size distribution. The former occurs in the processes of sputtering and vaporization, while the latter is also found with shattering. Sputtering denotes energetic impacts of gas particles on grains that can release species from the grain surfaces into the gas phase. This can happen either at very high temperatures (thermal sputtering) or at high relative gas–grain velocities (inertial sputtering). Sputtering of dust grains has been the subject of many theoretical studies (@Barlow:1978p17402 [@Tielens:1994p2924; @Jones:1994p2862; @May:2000p14538; @VanLoo:2012p18556]) and has become an established ingredient of shock models. Shattering, which is the fragmentation of grains due to grain-grain collisions, has been identified as a crucial process for determining the grain-size spectrum of interstellar grains (@Biermann:1980p17507 [@Borkowski:1995p17506]). Together with vaporization, which describes the return of grain material to the gas phase following grain-grain collisions, shattering needs to be included in order to account for UV extinction curves (@Seab:1983p18312). Moreover, recent infrared and sub-mm observations hint at an overabundance of small dust grains relative to the expected size-distribution in parts of the interstellar medium and thereby stress the importance of dust processing (@Andersen:2011p17918 [@PlanckCollaboration:2011p17886]), which has also been associated with MHD turbulence acceleration of dust grains or charge-fluctuation induced acceleration (e.g. @Ivlev:2010p17923 [@Hirashita:2010p17926]). A rigorous theoretical description of grain-grain collisions requires a detailed description of shock waves in solids, which had not been undertaken before the work of @Tielens:1994p2924, earlier studies having relied on much simplified models (e.g. @Seab:1983p18312). The effect of the microphysics of grain shattering and its effects on the grain size distribution in J-type shocks in the warm intercloud medium was studied by @Jones:1996p2864. @Slavin:2004p3051 extended this work by explicitly following individual trajectories of the grains, considered as test particles. In a series of papers, @Guillet:2007p2767 [@Guillet:2009p3498; @Guillet:2011p15512 hereafter Papers I, II and III] have shown that a multi-fluid approach to the dust dynamics, together with a detailed calculation of the grain charge distribution, shows shattering and the accompanying vaporization to be indispensable ingredients of shock models. For C-type shocks at preshock densities higher than $\sim$ 10${}^5$ cm${}^{-3}$, shattering becomes dramatically self-enhanced, due to feedback processes: electrons are heavily depleted on to small grain fragments, and the lack of electrons in the gas phase affects the grain dynamics, resulting in even more shattering and production of small grains. Owing to the increase of the total geometrical grain cross-section, these shocks become much hotter and narrower, and vaporization becomes important for the release of depleted species into the gas phase. Both the dynamical and the chemical consequences of shattering affect the predicted observational characteristics, compared with models in which shattering and vaporization are neglected. Given that conditions favoring C-shocks are frequently found in dense clouds and Bok globules, it is necessary to evaluate the observational consequences of shock models including grain-grain processing for a proper understanding of massive star formation and interactions of supernova remnants with molecular cloud cores (@Cabrit:2012p18561). The aforementioned studies leave open a series of issues that we aim to address in this paper. The three main questions are: - How do the effects of shattering and vaporization change if the magnetic field strength is varied? - What are the consequences of shattering for molecular line emission? - How do shattering and associated vaporization of SiO and C influence the SiO and atomic carbon line emission? In order to answer the last two questions, it is necessary to introduce a detailed treatment of radiative transport. However, due to the numerical complexity of the multifluid treatment of Papers I–III, a self-consistent merging of this multifluid model with an LVG treatment of molecular line emission would be technically difficult. It would also prevent the resulting model from being a practical analysis tool whenever large grids of models are required. Therefore, we have developed a method for implementing the effects of shattering and vaporization into a 2-fluid shock model that is sufficiently general to be applied to any similar model. We incorporated the main features, neglecting all the minor details of grain-grain processing; the resulting saving in computation time amounts to a factor of $\sim$100. With this model, the consequences of shattering and vaporization on the molecular line emission of C-type shocks can be studied in detail, using a self-consistent treatment of the line transfer (@Flower:2010p18961). In the context of the emission of SiO from C-type shocks (@Schilke:1997p2037 [@Gusdorf:2008p974; @Gusdorf:2008p1617]), the inclusion of grain shattering and vaporization, along with the line transfer, should improve significantly our ability to interpret the observations correctly. Our paper is structured as follows. In Sect. \[Sect2\], we introduce our model, which extends the work of @Flower:2010p18961 by including the shattering and vaporization of grains and molecular line transfer. We use this model to demonstrate the effect of shattering on the shock structure for different values of the magnetic field in the preshock gas (Sect. \[Sect3\]). In Sect. \[Sect4\], we consider observational diagnostics, with the main emphasis being on the rotational line emission of SiO. In addition, we investigate the influence of vaporization on the \[C I\] emission lines. The results are discussed and summarized in Sect. \[Sect5\]. Our model {#Sect2} ========= In order to study the effect of shattering and vaporization on molecular line emission, we have built on the findings of Papers I - III, where shattering and vaporization are described within a multi-fluid formalism for the dust grains. These results needed to be transferred to a 2-fluid formulation, as used in the model of @Flower:2010p18961 (hereafter FPdF10), whose model includes a detailed treatment of molecular line radiative transfer, in the presence of the cosmic microwave background radiation. In the present Section, we first describe the treatment of dust in the model of FPdF10 and then summarize the multi-fluid treatment of dust and its implications, finally outlining how we introduced the effects of shattering and vaporization into the model of FPdF10. Two-fluid treatment of dust {#subsec2.1} --------------------------- The way in which dust is treated in the LVG-model of FPdF10 derives from the work of @Flower:2003p1558 (hereafter FPdF03). This one-dimensional, steady-state, 2-fluid model[^1] of plane–parallel C-type shocks solves the magneto-hydrodynamical equations in parallel with a large chemical network, comprising more than 100 species and approximately 1000 reactions, including ion-neutral and neutral-neutral reactions, charge transfer, radiative and dissociative recombination. Furthermore the populations of the H${}_2$ ro-vibrational levels are computed at each integration step, along with the populations of the rotational levels of other important coolants, such as CO, OH and H${}_2$O. In these cases, the molecular line transfer is treated by means of the large velocity gradient approximation (see also Sect. \[Sect4\]). ‘Dust’ is included in the forms of polycyclic aromatic hydrocarbons (PAHs), represented by C${}_{54}$H${}_{18}$, and large grains from bulk carbonaceous material and silicates (specifically olivine, MgFeSiO${}_4$). The large grains are assumed to have a power law size distribution, $\dr n_{\rm g}(a)/\dr a \propto a^{-3.5}$ (Mathis et al. 1977), and radii in the range 100 to 3000 (0.3 ${\rm \mu}$m). In the preshock medium, the grain cores are covered by ice mantles, consisting of the chemical species listed in Table 2 of FPdF03 and including H${}_2$O, CO and CO${}_2$. Both PAHs and large grains exist as neutral, singly positively and singly negatively charged species. The physical treatment of dust comprises: - determination of the grain charge distribution, limited to $Z~=-1, 0, 1$; - sputtering of grain cores and mantles due to grain-gas collisions; - removal of mantles by cosmic ray desorption; - build-up of mantles by adsorption of gas-phase species in the postshock gas. Multi-fluid treatment of dust {#subsec2.2} ----------------------------- To properly account for the effects of grain-grain collisions in shocks, it is necessary to introduce a multi-fluid description of the dust dynamics, in which dust grains are treated as test particles. In Papers I-III the dust size distribution was modeled by the use of discrete bins, with grain sizes ranging from 5 to 3000 . PAHs were not included as separate species, distinct from the dust grains (as in the two-fluid model), but were incorporated into the dust size distribution. The equations describing the 2-D grain dynamics and complete charge distribution were integrated for each individual bin size, independently for silicate and carbon grains. While the shock structure is not much affected by this more accurate treatment of the grain dynamics and charging (Paper I, Appendix D[^2]), the introduction of dust shattering in grain-grain collisions and the corresponding changes to the grain size distribution have major effects on shocks in dense clouds with a low degree of ionization (Paper III). When grains collide with a velocity greater than 1.2 (carbon grains) or 2.7 (silicate grains), a fraction of their mass, which increases with velocity, is shattered into smaller fragments. At higher velocities, another fraction is vaporized and released into the gas phase (the numerical treatment follows the models of @Tielens:1994p2924 and @Jones:1996p2864, where the vaporization threshold is $\sim$ 19 kms${}^{-1}$). ![Temperature and density (upper panel) and velocity profiles in the shock frame (lower panel) for a 30 km s$^{-1}$ C-type shock with a magnetic field parameter $b=1.5$ and a preshock density of $n{}_{\rm H}$ = 10${}^5$ cm${}^{-3}$. The full curves show the results obtained when grain-grain processing is treated as in Paper III; the broken curves are obtained by following the approach of FPdF03. []{data-label="Fig1_dist"}](Fig1.pdf){width="9cm"} Paper III demonstrated that there is a marked shift of the dust size distribution towards smaller grains when the grain dynamics, charging and evolution are coupled self-consistently with the shock dynamics and chemistry. The most dramatic consequence is an increase of the total geometrical grain cross-section – which governs the coupling between the neutral and charged fluids – by a factor[^3] of $\sim$10. The shock becomes narrower by a factor of $\sim$4, and the peak temperature increases by a factor of 1.5–2 (see Fig. \[Fig1\_dist\]). The small grain fragments deplete ions and electrons from the gas until grains become the dominant charge carriers, a situation known as a [*dusty plasma*]{} (@Fortov:2005). The strong effect of shattering, for preshock densities of 10${}^5$ cm${}^{-3}$ and higher, can be understood in terms of the feedback processes described in Paper III. At these high preshock densities, it turns out that vaporization related to shattering becomes important, if not dominant, compared to sputtering. ![Peak temperature of the neutral fluid and the shock width (in cm), to $T = 100$ K in the cooling flow, for the grid of models without (broken lines) and with (full lines) grain-grain processing. The preshock density is 10${}^5$ cm${}^{-3}$, and the shock velocities are 20 (blue), 30 (green) and 40 (red) for various values of the magnetic field parameter, $b$, in $B(\mu {\rm G})=b\sqrt{n_{\rm H}({\rm cm}^{-3})}$. []{data-label="Fig2_Twidth"}](Fig2.pdf){width="9cm"} Implementation of shattering {#subsec2.3} ---------------------------- As described in Sect. \[subsec2.2\], there are two ways in which shattering affects the shock wave: first it increases the collisional dust cross-section, and second it lowers the degree of ionization of the gas. In the model of FPdF10, dust as a dynamical species and dust as a chemically reacting species were treated separately. Our implementation and validation of shattering in the FPdF10 model are described in Appendix A.1. Here we only summarize our basic approach. - The dynamical effects of shattering can be simulated by incorporating the change in the total grain cross section $\langle n\sigma \rangle$, computed with the model of Paper III; this is done by multiplying $\langle n\sigma \rangle$ by an additional factor that varies with the spatial coordinate, $z$, through the shock wave. This factor is modelled as an implicit function of the compression of the ionized fluid; its value is 1 in the preshock gas, and its maximum value – derived from the multi-fluid model – is attained when shattering is completed. It is possible to use linear fits of the parameters that are required. - The chemical effects of shattering, i.e. the consequences for grains as charged species, as determined by grain-charging reactions, need to be considered separately. In order to model the influence of shattering on the abundances of charged grains, a shattering source-term is introduced, whose value is consistent with $\langle n\sigma \rangle$. Implementation of vaporization {#subsec2.4} ------------------------------ Grain-grain collisions lead not only to shattering of grains, and consequent changes in the shock structure and degree of ionization, but also to their vaporization, when the impact velocity is higher than the vaporization threshold of $\sim$ 19 kms${}^{-1}$. We assume that the silicon released through vaporization is in the form of SiO (@1996GeCoA..60.1445N [@Wang:1999]). We have implemented the effect of SiO vaporization in a similar way as for shattering[^4]; the details are given in Appendix A.2. Our procedure involves introducing an additional term in the rate of creation of gas-phase SiO from Si and O in the grain cores. Thus, vaporization is incorporated as a new type of pseudo-chemical reaction. The rate of creation of SiO through vaporization is determined by parameterizing the results of the multi-fluid models. Another modification of the FPdF10 model concerns the mantle thickness, which is now calculated as described in Paper I, Appendix B. Our simplified, but self-consistent, treatment of shattering and vaporization reduces the computation time by a factor of more than 100. The influence of grain-grain processing on the shock structure {#Sect3} ============================================================== The grid of models {#subsec3.1} ------------------ Having implemented shattering and vaporization in the model of FPdF10, we first study the dependence of grain-grain processing and feedback on the strength of the transverse magnetic field. The analysis of Paper III was restricted to shocks propagating at the critical velocity, which defines the fastest possible C-type shock for a given magnetic field (or, conversely, the lowest value of the magnetic field possible for a given shock velocity). This critical velocity is determined by the condition that the shock velocity should be only slightly smaller than the velocity of magnetosonic waves, $$V_{\rm ms}=\frac{B}{\sqrt{4\pi \rho_{\rm c}}}$$ where $\rho_{\rm c}$ is the mass density of matter that is strongly coupled to the magnetic field. We assumed that the magnetic field strength, $B$, in dense clouds scales with the total proton density, $n{}_{\rm H}$, as $B(\mu {\rm G})=b\sqrt{n_{\rm H}({\rm cm}^{-3})}$ (@Crutcher:1999p19138), which implies that the magnetic energy density is proportional to $n{}_{\rm H}$. The assumed power-law exponent of 0.5 is somewhat lower than the more recent value of 0.65$\pm$0.05, given by @Crutcher:2012p19030, but is probably consistent with the uncertainties associated with its deduction from measurements of Zeeman splitting. The restriction to critical shocks was justified in Paper III by the fact that the observations will be dominated by the fastest shocks, if present. However, to assess the importance of shattering in the dense interstellar medium, it is necessary to establish whether shattering is still a relevant factor when the magnetic field is higher and the (charged) grains are better protected by their coupling to the magnetic field. Shattering is sensitive to the preshock density because the drag force becomes more important, relative to the Lorentz force, as the density increases. At high densities, large grains decouple from the magnetic field and exhibit a large velocity dispersion in the initial part of the shock wave, where they undergo destruction. Furthermore, the degree of ionization is lower in higher density gas, so the feedback due to the depletion of electrons is enhanced. We have studied shocks with a preshock density of 10$^5$ cm${}^{-3}$. We consider only this value because, on the one hand, it was shown that the change in the grain size distribution due to shattering is negligible for a preshock density of 10$^4$ cm${}^{-3}$, and, on the other hand, the strength of the feedback from shattering at higher densities prevents the multi-fluid model of Paper III from converging at a preshock density of 10$^6$ cm${}^{-3}$, and hence there are no results with which to compare at this density. We have restricted our calculations to shocks that do not fully dissociate H${}_2$ (shock velocity $\Vs \le$ 50kms${}^{-1}$), in practice to 20, 30 and 40 kms${}^{-1}$. The values of the magnetic field that we considered were determined by two considerations. First, in order to study critical shocks, we adopted the corresponding (minimum) values of $b$ (see Paper III); and, second, we varied $b$ for a given shock velocity in order to study the influence of variations in the magnetic field strength. Thus, we consider three values of the $b$ parameter for each velocity, in steps of 0.5, where the lowest value is close to that for a critical shock; the lowest values are $b=1.0$ for 20 kms${}^{-1}$, $b=1.5$ for 30 kms${}^{-1}$, and $b=2.0$ for 40 kms${}^{-1}$. The corresponding grid of nine models is summarized in Table \[t1\]. We note that a comparison of models with different shock velocities and the same magnetic field is possible for the case $b=2.0$. [llll]{} $\Vs$ \[kms${}^{-1}$\] & $b$ & $B$ \[$\mu $G\] & $n{}_{\rm H}$ \[cm$^{-3}$\]\ 20 & 1.0 & 316 & 10${}^{5}$\ 20 & 1.5 & 474 & 10${}^{5}$\ 20 & 2.0 & 632 & 10${}^{5}$\ 30 & 1.5 &474 & 10${}^{5}$\ 30 & 2.0 & 632 & 10${}^{5}$\ 30 & 2.5 &791 & 10${}^{5}$\ 40 & 2.0 & 632 & 10${}^{5}$\ 40 & 2.5 & 791 & 10${}^{5}$\ 40 & 3.0 & 949 & 10${}^{5}$\ \[t1\] Hotter and thinner {#subsec3.2} ------------------ The most dramatic effect of shattering is the change in the overall shock structure, owing to the increase in the collision cross-section of the dust and the corresponding increase in the coupling between the neutral and charged fluids. Figure 2 shows the peak temperature of the neutral fluid and the shock width (up to the point at which the temperature has fallen to 100 K) for our grid of nine models, as compared with the results obtained without grain-grain processing. The peak temperature of the models that include shattering increases by a factor of 1.5–2 for the shocks closest to the critical velocity, as in Paper III. Our models show that this increase is smaller for slower shocks, and for shocks with higher magnetic field strengths, because the charged grains are more strongly bound to and protected by the magnetic field. Thus, shocks with $\Vs$ = 20 and $b$=2 have a peak temperature comparable to that found without grain-grain processing. However, the shock width, as determined at the point in the cooling flow at which the gas temperature has fallen to 100 K, is always much smaller for models in which grain-grain processing is included. For the shocks closest to the critical velocity, the width is reduced by a factor of 4–5. The trend is the same as for the peak temperature: the shock widths, as computed with and without grain-grain processing, are most similar for lower velocities and higher magnetic fields, differing by a factor of only 1.7 for $\Vs$ = 20 and $b$=2. It is interesting to see the dependence of the shock structure on the transverse magnetic field strength. In models that neglect grain-grain processing, a change in the magnetic field modifies the shock width at almost constant peak temperature, whereas, when grain-grain processing is included, a variation in the magnetic field has a strong affect on the peak temperature. We return to this point in Sect. 4, where we consider the differences between the spectra predicted by these two categories of model; but it is already clear that shattering is significant, even for non-critical shocks, and should be included in steady-state C-type shock models with high preshock densities ($n{}_H\ge$ 10$^5$ cm${}^{-3}$). Observational consequences {#Sect4} ========================== In our model, the molecular line transfer is treated using the large velocity gradient (LVG) approximation (e.g. ), allowing for self-absorption via the escape probability formalism. The LVG method is well adapted to the conditions of shocks, where flow velocities change rapidly. The computation of the molecular energy level populations is performed in parallel with the integration of the chemical and dynamical rate equations, as introduced for CO by @Flower:2009p18313. The molecular line transfer of H${}_2$O, CH${}_3$OH, NH${}_3$, OH and SiO has since been added (@Flower:2010p19006 [@Flower:2010p18961; @Flower:2012p18432]. The modelling of interstellar shock waves requires many molecular energy levels to be considered. As described in FPdF10, we include levels of H${}_2$O up to an energy of approximately 2000 K above ground. Although this is less than the maximum temperatures that are attained, the high values of the radiative (electric dipole) transition probabilities ensure that the populations of higher, neglected levels remain small, and hence the associated errors in the computed line intensities are modest. Above the maximum temperatures for which the rotational de-excitation coefficients have been calculated, their values are assumed to remain constant (cf. @Flower:2012p18432, Appendix A). Rate coefficients for excitation are obtained from the detailed balance relation. The consequences of grain-grain processes for the molecular line radiation of H${}_2$, H${}_2$O and OH for representative shocks of $\Vs$ = 30 are briefly discussed in Sect. \[subsec4.1\] and in more detail in Appendix B. In Sect. \[subsec4.2\], we consider the rotational line emission of SiO and, in Sect. \[subsec4.3\], the vaporization of carbon. Molecular line emission {#subsec4.1} ----------------------- We have seen that including grain-grain processing leads to an increase in the peak shock temperature; this affects the chemistry and gives rise to higher fractional abundances of molecules in excited states. The intensities of lines emitted by highly excited states are thereby enhanced. At the same time, the column density of shocked gas decreases, which tends to reduce the intensity of molecular line emission. The net effect depends on the chemical and spectroscopic properties of the individual molecules. In Appendix B, we show that the intensities of all transitions of OH increase in shocks faster than 30 that incorporate grain-grain processing; this is due to the temperature sensitivity of OH formation. On the other hand, the intensities of the emission lines of H${}_2$, CO, and H${}_2$O decrease because of the reduction in the column density of shocked material. Furthermore, the inclusion of grain-grain processing introduces a dependence of the intensities of lines from highly excited states on the magnetic field, owing to the variation of the peak shock temperature with the field strength (see Sect. \[subsec3.2\] and Fig. \[Fig2\_Twidth\]). The effect of vaporization on SiO emission {#subsec4.2} ------------------------------------------ ![ The fractions of Si (left panel) and C (right panel) released from grain cores in model M1 of Table \[t2\] by vaporization (full lines) and sputtering (broken lines), for $\Vs$ = 20 (blue), $\Vs$ = 30 (green), and $\Vs$ = 40 (red), as functions of the magnetic field parameter, $b$, in $B(\mu {\rm G})=b\sqrt{n_{\rm H}({\rm cm}^{-3})}$. []{data-label="Fig3_CSifrac"}](Fig3.pdf){width="9cm"} SiO is a prominent indicator of shock processing in dense clouds associated with jets and molecular outflows (e.g. ). The first chemically adequate theoretical study of SiO production by sputtering in C-type shocks was conducted by @Schilke:1997p2037; this work was pursued subsequently by @Gusdorf:2008p974 [@Gusdorf:2008p1617]. These studies considered the release, by sputtering, of Si from grain cores and SiO from grain mantles, but not the process of vaporization, which could modify the predicted SiO rotational line profiles. We note that @Cabrit:2012p18561 found that, in the protostellar jet HH212, at least 10% of elemental Si could be present as gas-phase SiO, if the wind is dusty. To explain such a high value, they hint at the possible importance of grain-grain processing and the release of SiO by vaporization. The sputtering of Si from the grain cores seems unable to account for the observations in this case, as the dynamical timescale of 25 yr is too short for the chemical conversion, in the gas phase, of the sputtered Si into SiO. [llll]{} Scenario & Si in cores & SiO in mantles & Grain-grain processing\ M1 & yes & no & yes\ M2 & yes & no & no\ M3 & yes & 10% & no\ \[t2\] The influence of grain-grain processing on the SiO line emission is twofold. On the one hand, vaporization strongly increases the amount of SiO released from grain cores, at sufficiently high shock speeds. On the other hand, as the process of vaporization is necessarily related to shattering of dust grains, shocks with grain-grain processing have a different structure (cf. Sect. \[subsec3.2\]). The consequences of the latter effect for the emission by molecules whose abundances are not directly influenced by vaporization are summarized in Sect. \[subsec4.1\]. In order to address the question of how the related effects of shattering and vaporization affect the emission of SiO in C-type shocks, we compare results of shock models, obtained including and excluding grain-grain processing, for three different scenarios, summarized in Table \[t2\]. Models corresponding to scenario 1 (M1) were calculated using the implementation of shattering and vaporization described in Sect. 2.3 and 2.4 with the parameters of Table \[ta1\]. Models corresponding to scenario 2 (M2) do not include grain-grain processing, but Si is still released by the sputtering of grain cores. Scenario 3 (M3) differs from M2 in that 10% of the elemental Si is assumed to be in the form of SiO in the grain mantles. In total, there are $9\times3=27$ individual models to be computed. ### The release of SiO through dust processing {#subsubsec4.2.1} In order to compare the relative importance of vaporization and sputtering, we investigate the release of the Si in grain cores into the gas phase by each of these processes during the passage of a shock wave. The left-hand panel of Figure \[Fig3\_CSifrac\] shows the fraction of Si eroded from grain cores through sputtering and vaporization in model M1. There is negligible sputtering for $\Vs \le20$ kms${}^{-1}$, because the adopted sputtering threshold for refractory grain material is $\sim$ 25 (@May:2000p14538). Both the sputtering of Si and the vaporization of SiO are inhibited by the magnetic field. In the case of sputtering, the maximum ion-neutral drift velocity decreases with increasing field strength, whereas, for vaporization, the charged grains are more strongly coupled to the field and hence to the charged fluid (we assume that the magnetic field is ‘frozen’ in the charged fluid). For models M3 (SiO in mantles) with velocities sufficient for mantle sputtering, all the SiO initially in the mantles is released into the gas phase. The rotational line emission of SiO depends not only on the amount of silicon released from the grains but also on the location of its release and the physical conditions prevailing where SiO is present within the shock wave. ![ Upper panel: Evolution of the fractional abundance of SiO in the gas phase (left ordinate, full curves) and temperature of the neutral fluid (right ordinate, broken curves). From left to right: dust modelling scenarios M1, M2, and M3. The models shown in all three panels are for $\Vs$ = 20 with $b$ = 1.0 (blue), $\Vs$ = 30 with $b$ = 1.5 (green), and $\Vs$ = 40 with $b$ = 2.0 (red). The fractional abundance of SiO in the gas phase is negligible for $\Vs$ = 20 with $b$ = 1.0 in scenario M2. Lower panel: variation through the shock wave of the gas-phase fractional abundances of O${}_2$ (left ordinate, full blue curves) and OH (left ordinate, full red curves), together with the temperature of the neutral fluid (right ordinate, black broken curves), for $\Vs$ = 30 and $b$ = 1.5 and each of the dust models M1, M2 and M3 (from left to right). []{data-label="Fig4_xSiO"}](Fig4.pdf){width="9cm"} The upper panel of Figure \[Fig4\_xSiO\] shows the fractional abundance of SiO in the gas phase, for dust models M1–3 of Table \[t2\] and three different values of the shock speed. In model M1, vaporization releases SiO directly into the gas phase in the region where the temperature of the neutral fluid is rising steeply. The higher the shock speed, the more SiO is produced. The large increase in the total grain cross section in the postshock gas enhances the rate of formation of mantles and removes SiO from the gas phase. The variation of the fractional abundance of SiO is very different if only the sputtering of grain cores is considered (M2), because the Si that is released by sputtering has to be transformed into SiO by gas-phase chemical reactions, predominantly oxidation by O${}_2$ and OH. The corresponding reactions are $${\rm Si + O}_2 \longrightarrow {\rm SiO + O}$$ $${\rm Si + OH} \longrightarrow {\rm SiO + H} .$$ For reaction (1) the rate coefficient (cm${}^3$ s${}^{-1}$) $$k = 1.72 \times 10^{-10} (T/300)^{-0.53}\exp(-17/T)$$ was measured by @LePicard:2001p19024. The rate coefficient for reaction (2), which is not measured, was adopted to be the same. Thus, the fractional abundance of SiO in the gas phase depends not only on the amount of Si sputtered from the grain cores but also on the abundances of O${}_2$ and OH, displayed in the lower panel of Fig. \[Fig4\_xSiO\]. The chemical delay in SiO production is apparent in the upper panel of Fig. \[Fig4\_xSiO\]; the abundance of SiO peaks in the cool and dense postshock region. If SiO is initially in the grain mantles (M3), its release is rapid and complete even before the temperature of the neutral fluid rises significantly.[^5] For the shock models considered, all the SiO in the mantles is released into the gas phase. We have assumed that 10% of elemental silicon is initially in the form SiO in the mantles; this is the largest of the values considered by @Gusdorf:2008p974. If the same fraction of silicon is initially in the form of SiH${}_4$ in the mantles, there occurs the same chemical delay in the production of SiO in the gas phase as in models M2. ### Excitation of the SiO rotational lines {#subsubsec4.2.2} The differences in the SiO abundance profiles in the three cases M1–3 – specifically, whether SiO is already present in the hot gas, early in the shock wave, or only in the cold postshock gas – have consequences for the relative intensities of the SiO rotational lines. To illustrate this point, we compare, in Fig. \[Fig5\_Tmaxrel\], the peak temperatures of the rotational lines of SiO, relative to the $J = 5$–$4$ transition, calculated for each of these three cases. The results in Fig. \[Fig5\_Tmaxrel\] may be better understood by referring to the line profiles, shown for the $\Vs$ = 30 shock in Fig. \[Fig7\_SiOlinesorigin\]. The first thing to note is the similarity of the relative peak line temperatures for the $\Vs$ = 30 and $\Vs$ = 40 models with only core sputtering (M2); this can be ascribed to the chemical delay in SiO formation. The rotational lines up to the 8–9 transition in these models mostly stem from the cold postshock gas, where almost maximum compression is reached. Therefore, the lines have large optical depths and near-LTE excitation conditions, which makes the relative peak temperatures of the transitions displayed independent of the shock velocity. The models with vaporization (M1), on the other hand, show a clear variation of excitation conditions with shock speed. The lines are formed mainly in the cooling flow, where the temperature of the neutral fluid is a few hundred K (the case for $\Vs$ = 30 kms${}^{-1}$, $J_{\rm up}$ $>$ 4 and for $\Vs$ = 40 kms${}^{-1}$, $J_{\rm up} > 2$), and are sensitive to the velocity dependence of both the temperature in the cooling flow and the amount of SiO that is produced by vaporization. For the models with SiO in the mantles (M3), the differences in the relative peak line temperatures are much less between shocks of different velocities than is the case for models with vaporization. The same amount of SiO is released for both velocities and the column density of the cooling gas is greater than in models that include vaporization. Most of the radiation arises in the cooling flow, where the temperature profiles are very similar and the optical depths are large: for model M2, with $\Vs$ = 30 and $b$ = 1.5, the transitions up to $J = 7$–$6$ become optically thick, with optical depths at the line centres reaching $1 \lesssim \tau \lesssim 6$, depending on the transition. We conclude that clear variations of the relative line intensities with the shock velocity are characteristic of those models in which grain vaporization occurs. ![ The peak temperatures (K) of the rotational emission lines of SiO, relative to the $J = 5$–$4$ transition, as functions of the rotational quantum number of the upper level of the transition $J_{\rm up}$. Displayed are the shock models in our grid that are most strongly influenced by vaporization, i.e. $\Vs$ = 20 with $b$ = 1 (blue), $\Vs$ = 30 with $b$ = 1.5 (green), and $\Vs$ = 40 with $b$ = 2 (red). Full lines: M1; dotted lines: M2; broken lines: M3. []{data-label="Fig5_Tmaxrel"}](Fig5.pdf){width="9cm"} {width="\textwidth"} ![ Temperature profiles of the neutral fluid (left ordinate, broken curves) and SiO line temperatures (right ordinate, full curves) of the rotational transitions $J = 2$–$1$ (blue), 5–4 (green), and 9–8 (red) for $\Vs$ = 30 , $b$ = 1.5; models M1 (left panel), M2 (centre panel), and M3 (right panel). The abscissa is the distance through the shock wave. []{data-label="Fig7_SiOlinesorigin"}](Fig7.pdf){width="9cm"} ### SiO rotational line profiles {#subsubsec4.2.3} Given that the excitation conditions of the SiO lines differ in the three scenarios M1–3, we expect the SiO spectral line profiles differ also, with respect to line width, location of the peak, and integrated flux. Line profiles for $\Vs$ = 30 kms${}^{-1}$, divided by the peak temperature of the $J = 5$–$4$ transition (except for the 5–4 transition itself), are shown in Fig. \[Fig6\_SiOlines\]. Model M1 predicts narrower lines for low-$J$ transitions, but similar widths for high-$J$ transitions, as model M2, and much narrower lines than model M3. There is a weak variation with $J_{\rm up}$ of the location of the line peak in models M1 and M2. For M1, the peak in the profile of the lowest transition occurs in the cold posthock gas, where the neutral fluid is moving at a velocity $\Vn$ $\sim$ 3 kms${}^{-1}$, whereas, for higher transitions, the peak moves to hotter gas, at $\Vn$ $\sim$ 4 kms${}^{-1}$. A similar trend is seen for model M2, but, in this case, it is only the highest (and very weak) transitions that peak earlier in the shock wave. In contrast, the line peaks for model M3 are always located at the same velocity of $\Vn$ $\sim$ 24 kms${}^{-1}$. Thus, the differences in shock structure and spatial distribution of SiO between models M1 and M2 do not have a strong effect on the SiO rotational line profiles. *In particular, the release of SiO through vaporization early in the shock wave does not lead to significant broadening of the lines*. However, the lines are strongly broadened if SiO is present in the grain mantles (scenario M3), and this seems to be the only way of accounting for observed line widths of several tens of if they are to be explained by one single shock. Alternatively, broad SiO lines can be explained by the existence of several shocks inside the telescope beam, such that the individual narrow line profiles appear spread out in radial velocity due to different velocities and inclination angles, in the observer’s frame. Similarly, the lines would be broadened by the velocity profile associated with a bow shock (e.g. @1989MNRAS.237.1009B). ![Integrated intensities of the rotational transitions $J_{\rm up} \rightarrow J_{\rm up}-1$ of SiO for shocks with $\Vs$ = 20 kms${}^{-1}$, $b$ = 1 (blue), $\Vs$ = 30 kms${}^{-1}$, $b$ = 1.5 (green), and $\Vs$ = 40 kms${}^{-1}$, $b$ = 2 (red). Full lines: model M1; dotted lines: model M2; broken lines: model M3. []{data-label="Fig8_SiOTdV"}](Fig8.pdf){width="9cm"} ![Integrated intensities of the rotational transitions $J_{\rm up} \rightarrow J_{\rm up}-1$ of SiO, relative to the intensity of the transition with $ J_{up}$ = 5. The shock velocity is $\Vs$ = 30 kms${}^{-1}$, with values of the magnetic field parameter $b$ = 1.5 (red), $b$ = 2.0 (green) and $b$ = 2.5 (blue). Full curves denote model M1, dotted curves model M2. []{data-label="Fig9_SiOTdVrel"}](Fig9.pdf){width="9cm"} ### Integrated SiO rotational line intensities {#subsubsec4.2.4} An interesting question is whether the release of SiO due to vaporization significantly increases the integrated (along the $z$-direction) SiO rotational line intensities. Fig. \[Fig8\_SiOTdV\] shows, that indeed there is an increase in the intensities of the highly excited transitions. Furthermore, the slope of the integrated intensity curves differs between models M1 and M2 for transitions $3 \le J_{\rm up} \le 10$. The integrated intensities of the lowest transitions – up to $J_{\rm up} \sim 7$ – are dependent on the timescale for accretion of gas-phase species on to grains in the cooling flow; this timescale depends on the (uncertain) accretion rates. In addition, 1-dimensional models tend to overestimate the compression of the postshock gas. Previous studies (@Gusdorf:2008p1617) have found the accretion timescale to be unimportant for the molecular line emission. However, owing to the large increase in the total grain cross-section, induced by vaporization, we see, for the first time, an observational consequence that may be related to this timescale. Independent of the accretion timescale, the integrated intensities of, in particular, the lowest transitions increase significantly for model M3; but we recall that the amount of SiO in the mantles is treated as a free parameter, and the absolute intensities can be varied accordingly. The intensity curves corresponding to model M3 have similar slopes to the curves of model M1, for transitions from $J_{\rm up} \gtrsim 10$; they peak at $J_{\rm up} = 3$. On the other hand, the peak of the curve corresponding to model M1 at $\Vs$ = 40 is displaced to $J_{\rm up} = 4$, owing to the higher maximum temperature. As may be seen from Fig. \[Fig8\_SiOTdV\], the integrated line intensities decrease by a factor $\sim$100 between $5 \le J_{\rm up} \le 10$ for model M2, which sets it apart from models M1 and M3. Fig. \[Fig9\_SiOTdVrel\] shows the effect of variations in the magnetic field on the integrated SiO rotational line intensities. In this Figure, the integrated intensities are expressed relative to the 5–4 transition, in order to focus attention on the excitation conditions, rather than the release of SiO into the gas phase. For model M1, there is a clear variation with magnetic field strength, which is not present for model M2. This difference can be understood from Fig. \[Fig2\_Twidth\], which shows that the maximum temperature of the neutral fluid hardly varies with the magnetic field strength for model M2, whereas the maximum value changes for model M1. Thus, Fig. \[Fig9\_SiOTdVrel\] demonstrates that the excitation conditions vary with the strength of the magnetic field when the effects of grain-grain collisions are incorporated. The effect of vaporization on \[C I\] emission {#subsec4.3} ---------------------------------------------- Our model, and the models of Papers I–III, includes two different populations of dust grains: silicate and carbon grains. While emission of SiO could be used as a tracer of vaporization in shocks because of the strong depletion of Si and SiO in quiescent gas, the effects of vaporization on \[C I\] emission will not be seen as clearly. Nonetheless, the vaporization of graphite grains modifies the emission of atomic carbon. In order to predict the magnitude of this effect, we have used the multi-fluid model of Paper III. Fig. \[Fig3\_CSifrac\] (right-hand panel) shows the fraction of carbon released from grain cores by sputtering (M2) and vaporization (M1). At low velocities, the destruction of carbon grains is less than for silicates because the gyration of large carbon grains is damped more rapidly, owing to their lower specific density. At higher velocities, for which small grains also contribute significantly to vaporization, the fraction of carbon released from graphite grains is higher than the fraction of silicon from silicates; graphite grains are more strongly affected by shattering. Table B.1 shows the intensities of two \[C I\] forbidden lines, at 609.8 $\mu$m and 370.4 $\mu$m. The values in these Tables confirm that the effect of vaporization on the \[C I\] lines is less than on the SiO rotational transitions. Concluding remarks {#Sect5} =================== In this series of papers, the consequences of grain-grain collisions in shock waves have been investigated; such processes had been ignored in previous studies of C-type shocks. Shattering is a key factor in the production of the large populations of very small grains (mostly carbonaceous: see, for example, @1984ApJ...285...89D) in the turbulent ISM (@Hirashita:2010p17926). It has been shown, in earlier papers in the series, that C-type shock waves provide similar dynamical conditions to turbulence, and so shattering needs to be included in shock models. Our study relies on the shattering model that was developed in Paper III, and our results inherit the dependence of this model on parameters such as the size of the smallest fragments (5Å  in Paper III), the slope of the size distribution of the fragments, their charge distribution, and the composition of the dust grains. Whilst shattering almost certainly occurs in the ISM, as a direct consequence of the weak coupling of large grains to the magnetic field in high density clouds, the threshold density at which shattering starts to have a significant effect on the shock dynamics is more uncertain. In our model, the processes of shattering and vaporization are linked, so that the threshold density for shattering applies also to vaporization. The parameters of the shattering model in Paper III were chosen to minimize the amount of shattering and thereby yield a conservative estimate of the threshold density. It was shown that strong feedback on the shock dynamics was expected only for densities higher than $\sim10^5$ cm${}^{-3}$. The results of the present paper include some observational predictions that should help to constrain the grain-shattering model. The main conclusions of our study are as follows: 1. The influence of grain-grain processing on the overall shock structure was found to be significant for the full range of magnetic field strengths that we studied. The maximum temperature increases by a factor of 1.5–2, and the shock width is reduced by a factor 4–5. The inclusion of grain-grain processing changes the dependence of the shock structure on the magnetic field strength. While for shocks without grain-grain processing there is only a weak dependence of the peak temperature on the magnetic field, the peak temperature becomes strongly dependent on the magnetic field when grain-shattering and vaporization are incorporated. Consequently, the intensities of, in particular, highly excited molecular transitions become dependent on the strength of the magnetic field. While shattering is shown to be important for all models of our grid, the vaporization of SiO from silicate grain cores is significant only for fast shocks and low magnetic fields. 2. There are two consequences of grain-grain processing for the molecular line emission: the reduced shock width results in a lower column density of shock-heated gas, whereas the higher peak temperature can modify the chemistry and enhance the fractional abundances of molecules in highly excited states. Which of these tendencies prevails is decided by the chemical and physical properties of the individual molecules and their transitions. The intensities of all lines of OH increase in shocks with velocities greater than 30 kms${}^{-1}$, when grain-grain processing is included, owing to the temperature sensitivity of OH formation. On the other hand, the intensities of the emission of CO and H${}_2$O decrease because of the reduced column density of shocked material. In the case of H${}_2$, the intensities of highly-excited transitions increase, whilst the intensities of lines of lower excitation decrease. 3. The release of SiO through collisional vaporization of silicate grain cores enhances the integrated intensities of SiO rotational lines, mainly from highly-excited levels. However, this effect is counteracted by the reduction in shock width. To obtain significantly higher line intensities, it is necessary to introduce SiO into the grain mantles. The situation is similar with respect to the widths of the SiO rotational lines. Although vaporizations releases SiO early, in the hot part of the shock wave, the reduction in the shock width prevents the lines from becoming significantly broader than in models that neglect grain-grain processing. Therefore, vaporization alone cannot account for broad lines if only one single shock is considered. To obtain broad profiles, it seems necessary that SiO should be present in the grain mantles, such that mantle sputtering releases SiO, also in the early part of the shock wave. It is essential that SiO is released directly, thereby eliminating the chemical delay that would be associated with its production, in gas-phase reactions, from Si or SiH${}_4$. Due to the fact, that most of the SiO emission stems from the early part of the shock, where the temperature profile depends on the shock velocity, the SiO rotational line ratios vary with the shock velocity; this variation is not present if vaporization is ignored. The effect of vaporization on \[C I\] emission lines was found to be less than for the SiO rotational lines. We are grateful to an anonymous referee for useful comments that helped to strengthen the paper. S. Anderl acknowledges support by the DFG SFB 956, the International Max Planck Research School (IMPRS) for Astronomy and Astrophysics, and the Bonn-Cologne Graduate School of Physics and Astronomy. Shattering and vaporization: their implementation and validation ================================================================ Shattering ---------- ### Dynamical effects of shattering The dynamical effects of shattering on the shock structure can be simulated by ensuring that the total grain cross section, $\langle n\sigma \rangle$, changes in accord with the results of Paper III. We found that it is possible to model the increase of $\langle n\sigma \rangle$ due to shattering by multiplying the total grain cross section, in the absence of grain-grain processing, by an additional factor, which varies with the spatial coordinate, $z$, through the shock wave. This factor is modelled as an intrinsic function of the compression of the ion fluid, normalized to the theoretical postshock compression at infinity, as predicted by the Rankine-Hugoniot relations. This normalized ion compression parameter constitutes a function, $\eta(z)$, varying from $0$ in the preshock to $1$ in the postshock medium. The preshock (medium 1) and postshock (medium 2) kinetic temperatures are approximately equal in the C-type shock models of our grid. Thus, the Rankine-Hugoniot continuity relations may be applied across the shock wave, replacing the relation of conservation of energy flux by the isothermal condition, $T_1 = T_2$. Then, setting the ratio of specific heats $\gamma = 1$, the expression for the compression ratio, $\rho _2/\rho _1$, across the shock wave may be derived (cf. @Draine:1993p1759). Under the conditions of our models, which are such that M$_{\rm s} = V_{\rm s}/c_1 \gg~$M$_{\rm A} = V_{\rm s}/V_{\rm A} \gg 1$, where M$_{\rm s}$ is the sonic Mach number of the flow, evaluated in the preshock gas, $c_1$ is the sound speed, and M$_{\rm A}$ is the Alfvénic Mach number, also in the preshock gas, the compression ratio reduces to $\rho _2/\rho _1 \approx \sqrt {2}$M$_{\rm A}$, whence $$V_{\rm postshock} = \frac{V_{\rm A}}{\sqrt{2}} = \frac{B}{\sqrt{2\times 4\pi ~1.4~ m_{\rm H} n_{\rm H}}} ~ {\rm ,}$$ where $V_{\rm postshock}$ is the flow speed in the postshock gas, $m_{\rm H}$ the proton mass and $n_{\rm H}~=~n({\rm H})+2~n({\rm H}_2)$ the proton density in the preshock gas. Using $$\eta(z)=\frac{\Vs/\Vi(z)-1}{\Vs/V_{\rm postshock}-1}$$ (where $\Vs$ is the shock velocity and $\Vi$ the velocity of the ion fluid in the shock frame, respectively), the factor by which the total grain cross section is enhanced is given by $$\Sigma(z)=\eta(z) \cdot(\Sigma_{\rm max}-1)+1 ~{\rm .}$$ The extent of the increase of $\langle n\sigma \rangle$ due to shattering is given by the final value of the shattering-factor, $\Sigma{}_{\rm max}$. This value depends on the shock velocity and the magnetic field and is an external parameter, which needs to be extracted from the multi-fluid models. However, this simple functional form did not reproduce satisfactorily the onset of shattering in the shock. We therefore propose the following refined expression $$\Sigma=\Big( \eta^{\beta}-\frac{\sin{(2\pi~\eta^{\beta})}}{2\pi}\Big)\cdot \Big(\Sigma_{max}-1\Big)+1$$ that introduces another parameter, $\beta$, which needs to be extracted from the corresponding multi-fluid model. $\beta$ describes the delay in the shattering feedback, relative to the ion compression, and is only weakly dependent on the shock velocity. The parameters, $\beta$ and $\Sigma_{\rm max}$, which correspond to the grid of models introduced in Section \[subsec3.1\], are listed in Table \[ta1\]. Linear fits in the shock velocity $\Vs$ and the magnetic field parameter $b$ are $$\Sigma_{max} = 9.5 + 0.4 \cdot \Vs - 4.6\cdot b$$ and $$\beta = 0.95 - 0.025\cdot \Vs + 0.4\cdot b \textrm{ .}$$ The increase of the total grain collisional cross-section needs to be consistent with the mean square radius of the grains and with their total number density, following the compression of the ions. Analyzing the multi-fluid computations corresponding to Paper III, we find a reasonable approximation to the behaviour of the total grain number density, $n_{\rm G}$, and the mean square radius, $\langle \sigma \rangle_{\rm G}$: the former increases as $\Sigma{}^4$, the latter decreases as $\Sigma{}^{-3}$, relative to the corresponding values without shattering. These changes affect the rates of grain-catalyzed reactions, adsorption to the grain mantles, excitation of H$_2$ in collisions with grains, and transfer of momentum and thermal and kinetic energy between the neutral and the charged fluids. Figure \[FigA1\_shattering\] shows a comparison between the multi-fluid model of Paper III and our current model[^6] for a representative shock for which $n_{\rm H} = 10^5$ cm$^{-3}$, $\Vs$ = 30 and $b$ = 1.5. The temperature profiles of the neutral fluid agree well. Furthermore, the variation of the total grain-core cross-section is reproduced by our simulation, as may be seen from the lower panel of Figure \[FigA1\_shattering\]. The total grain cross-section (including mantles) increases somewhat later in our current model, due to small differences – independent of grain shattering –with the model of Paper III. The discrepancy in the value of the grain cross-section in the postshock medium arises because our simplified treatment of shattering tends to underestimate the final number density of small grains. However, this simplification introduces only small deviations in the hydrodynamic parameters, such that the values of both the peak temperature of the neutral fluid and the shock width (up to 100 K in the cooling flow) agree to within $\pm \sim$15 % for the entire grid of models incorporating shattering. ![Upper panel: temperature profiles of the neutral fluid for a shock with $n_{\rm H}~=~10^5$ cm$^{-3}$, $\Vs$ = 30 and $b$ = 1.5 corresponding to Paper III (in blue) and our current model (in red). Lower panel: evolution of the total grain cross section, $\langle n\sigma \rangle$, for the same shock models, with and without taking into account the grain mantles and normalized to the preshock values of $\langle n\sigma \rangle$. []{data-label="FigA1_shattering"}](FigA1.pdf){width="9cm"} [llllll]{} $\Vs$ \[kms${}^{-1}$\] & $b$ & $B$ \[$\mu {\rm G}$\] & $\Sigma{}_{\rm max}$ & $\beta$ & $x$(SiO${}_{\rm peak}$)\ 20 & 1.0 & 316 & 11.7 & 0.8 & 1.53 $\cdot$ 10${}^{-8}$\ 20 & 1.5 &474 & 10.5 & 1.1 & 1.99 $\cdot$ 10${}^{-17}$\ 20 & 2.0 & 632 & 8.5 & 2.0 & 2.16 $\cdot$ 10${}^{-19}$\ 30 & 1.5 & 474 & 14.2 & 0.8& 1.55 $\cdot$ 10${}^{-7}$\ 30 & 2.0 & 632 & 11.7 & 0.9& 7.92 $\cdot$ 10${}^{-9}$\ 30 & 2.5 & 791 & 10.5 & 1.2 & 5.63 $\cdot$ 10${}^{-13}$\ 40 & 2.0 & 632 & 18.4 & 0.8& 9.13 $\cdot$ 10${}^{-7}$\ 40 & 2.5 & 791 & 13.7 & 0.9 & 3.43 $\cdot$ 10${}^{-7}$\ 40 & 3.0 & 949 & 11.4 & 1.1 & 1.34 $\cdot$ 10${}^{-7}$\ \[ta1\] ### Chemical effect of shattering The change in the total grain cross-section, $\langle n\sigma \rangle$, owing to shattering, has consequences for the rates of grain-charging reactions. In order to model the effect of shattering on the abundances of charged grains, the corresponding chemical source term needs to be introduced. The charge distribution of the fragments is essentially unknown and cannot be numerically integrated separately from that of grains already present in the medium. In the present paper[^7], the charge distribution of fragments is designed to ensure charge conservation. We fitted the fragment charge distribution by aligning the shock widths (see Fig. \[FigA1\_shattering\], upper panel). This procedure yielded a charge distribution in which 1/2 of the fragments were neutral, and 1/4 were positively and 1/4 negatively charged, following shattering. Subsequently, the grain charge distribution evolved on a timescale that can be long – of the order of the flow time through the shock wave when the mean grain size becomes very small. The source term for the creation of grains through shattering can be derived from the equation of conservation of the total flux of grains, in the absence of shattering, $$\label{eqnA7} \sum_{j ~\in \{\rm G0, G+, G-\}} n_{ j} \cdot V_j \cdot \Sigma^{-4}(z) = {\rm constant}$$ where $n_{\rm G0}$ is the number density of neutral grains, $n_{\rm G+,G-}$ are the number densities of positively and negatively charged grains, respectively, $V_{\rm G0} = V_{\rm n}$ is the velocity of the neutral grains, in the shock frame, and $V_{\rm G+, G-} = V_{\rm i}$ is the velocity of the charged grains, in the shock frame. Differentiation of (\[eqnA7\]), subject to the charge distribution of the fragments, then yields $$\frac{\dr n_{\rm G0}}{\dr z}\bigg|_{\rm shat}=\frac{1}{2}\cdot4\cdot n_{\rm G,preshock}\frac{\Vs}{\Vn}\cdot \Sigma^3\cdot \frac{\dr \Sigma}{\dr z}$$ for the neutral grains, and $$\frac{\dr n_{\rm G+,G-}}{\dr z}\bigg|_{\rm shat}=\frac{1}{4}\cdot4\cdot n_{\rm G,preshock}\frac{\Vs}{\Vi}\cdot \Sigma^3\cdot \frac{\dr \Sigma}{\dr z}$$ for the charged grains, where $n_{\rm G,preshock}$ is the total (charged and neutral) number density of grains in the preshock gas. The derivative of $\Sigma$ is given by $$\frac{\dr \Sigma}{\dr z} = \Big( \frac{\dr}{\dr z} \eta^\beta - \cos\big(2\pi \eta^\beta\big)\frac{\dr}{\dr z} \eta^\beta \Big) \Big(\Sigma_{\rm max} - 1 \Big)$$ where $$\frac{\rm d}{\dr z} \eta^\beta= \beta~ \eta^{(\beta -1)} \frac{\dr \eta}{\dr z}$$ and $$\frac{\dr \eta}{\dr z} = -\frac{1}{(\Vs/V_{\rm postshock}-1)}\frac{\Vs}{\Vi^2}\frac{\dr \Vi}{\dr z}.$$ With these parameters, our model is able to reproduce the main effects of the increase in the grain cross section, reported in Paper III: the effective rate of recombination of electrons and ions is enhanced; the fractional abundance of free electrons falls by three orders of magnitude; and dust grains become the dominant charge carriers, with equal numbers of positively and negatively charged grains being produced. Vaporization ------------ The effect of vaporization is modelled as an additional term in the creation rate of gas-phase SiO, from Si and O in grain cores (denoted by \*\*), corresponding to a new type of pseudo-chemical reaction: Si** + O** = SiO + GRAIN Because vaporization sets in suddenly, when the vaporization threshold is reached, the function $$\Omega(z) = \frac{1}{1+\exp \bigl(-10^3\cdot \bigl(\Vs/\Vi(z)-\Vs/6 V_{\rm postshock}\bigr)\bigr)}$$ can be used to approximate the rate of creation of SiO. The function $\Omega(z)$ is centred at the point where the compression of the charged fluid reaches $1/6$ of its final value, as determined by our fitting procedure. Similarly, the factor $10^3$ in the exponent, which determines the steepness of the function, derives from a fit to the numerical results of Paper III. Using this function, the creation rate (${\rm cm}^{-3}~{\rm s}^{-1}$) can be expressed as $$\begin{aligned} \frac{\dr n({\rm SiO})}{\dr t}\bigg|_{\rm vapo}&=& \frac{\dr \Omega(z)}{\dr t} \cdot n_{\rm H}\cdot x({\rm SiO})_{\rm peak} \\ &=& \frac{\dr \Omega(z)}{\dr z} \cdot V_{\rm n} \cdot n_{\rm H}\cdot x({\rm SiO})_{\rm peak} \\ &=& \frac{\dr \Omega(z)}{\dr z} \cdot V_{\rm s} \cdot n_{\rm H,preshock}\cdot x({\rm SiO})_{\rm peak} ~{\rm ,}\end{aligned}$$ where use is made of the conservation of the flux of $n_{\rm H}$, and where $$\frac{\dr}{\dr z}\Omega(z) = -\Omega(z) \cdot \big( 1 - \Omega(z) \big)\cdot 10^3 \frac{\Vs}{\Vi^2}\cdot \frac{\dr \Vi}{\dr z}~ {\rm .}$$ The spatial change in number density of SiO, owing to vaporization, is then given by $$\frac{\dr n({\rm SiO})}{\dr z}\bigg|_{\rm vapo}= \bigg( \frac{\dr n({\rm SiO})}{\dr t}\bigg|_{\rm vapo} - n({\rm SiO})\frac{\dr \Vn}{\dr z}\bigg)\frac{1}{\Vn}$$ The peak fractional abundance $x$(SiO)${}_{\rm peak}$ needs to be computed with the multi-fluid model and constitutes a free parameter, given in Table \[ta1\] for our grid of models. As can be seen in Table \[ta1\], vaporization of SiO is relevant only for the shocks with high velocity and low magnetic field; it can be neglected for the shocks with $\Vs$ = 20 and $b$ = 1.5, $b$= 2, as well as for the model with $\Vs$ = 30 and $b$ = 2.5 . ![The fractional abundance of SiO (full curves), as determined when including grain-grain processing, using the present model (in red) and the multi-fluid model of Paper III (in blue); the shock parameters are $n_{\rm H}~=~10^5$ cm$^{-3}$, $\Vs$ = 30 and $b$ = 1.5. The temperature of the neutral fluid is shown also (broken curves; right-hand ordinate).[]{data-label="FigA2_vaporization"}](FigA2.pdf){width="9cm"} Figure \[FigA2\_vaporization\] shows the result of our implementation of vaporization. We assume that, initially (in the preshock medium), there are no Si-bearing species, either in the gas phase or in the grain mantles; all the Si is contained in olivine (MgFeSiO${}_4$) grain cores. This Figure compares the fractional abundance of SiO, as computed with our current model (incorporating grain-grain processing) and as predicted by the multi-fluid model of Paper III. By construction, the peaks of the fractional abundance of SiO agree, whereas the timescale for its accretion on to grains differs between the two models, as is visible in the plot; this discrepancy relates to the imperfect agreement of the total cross-section, $\langle n \sigma \rangle$, in the postshock medium (see Fig. \[FigA1\_shattering\]). However, we have verified that the timescale for SiO accretion is not critical to our analysis: the complete neglect of accretion on to grains leads to increases in the integrated intensities of the lowest rotational transitions of SiO and CO, by factors of $\sim$2 and $\sim$3, respectively. We note that the intensities of these transitions are, in any case, affected by the foreground emission of ambient, non-shocked gas. Molecular line emission {#molecular-line-emission} ======================= ![Computed H${}_2$ excitation diagrams for rovibrational levels with energies $E_{{\rm v},J} \le 10~000$ K and for shocks with $n_{\rm H}~=~10^5$ cm$^{-3}$, $\Vs$ = 30 and $b$ = 1.5 (red), $b$ = 2.0 (green) and $b$ = 2.5 (blue). Full lines: model M1; dotted lines: model M2. []{data-label="FigB1_H2"}](FigB1.pdf){width="8.5cm"} ![Integrated intensities of the rotational transitions $J_{\rm up} \rightarrow J_{\rm up}-1$ of CO for shocks with $n_{\rm H}~=~10^5$ cm$^{-3}$, $\Vs$ = 30 and $b$ = 1.5 (red), $b$ = 2.0 (green) and $b$ = 2.5 (blue). Full lines: model M1; dotted lines: model M2. []{data-label="FigB2_CO"}](FigB2.pdf){width="8.5cm"} ![Integrated intensities of selected rotational transitions of ortho-H${}_2$O plotted against the excitation energy of the emitting level, expressed relative to the energy of the $J = 0 = K$ ground state of para-H${}_2$O. Results are shown for shocks with $n_{\rm H}~=~10^5$ cm$^{-3}$, $\Vs$ = 30 and $b$ = 1.5 (red), $b$ = 2.0 (green) and $b$ = 2.5 (blue). Full circles: model M1; open circles: model M2. []{data-label="FigB3_oH2O"}](FigB3.pdf){width="8.5cm"} ![Integrated intensities of the rotational transitions of OH for emitting levels of negative parity, plotted against the excitation energy of the upper level. Results are shown for shocks with $\Vs$ = 30 and $b$ = 1.5 (red), $b$ = 2.0 (green) and $b$ = 2.5 (blue). Full circles: model M1; open circles: model M2. []{data-label="FigB4_OH"}](FigB4.pdf){width="8.5cm"} The introduction of shattering leads to a reduction of the shock width, and hence to lower column densities of shocked material, and to higher peak temperatures, which affect the chemistry and enhance the fractional abundances of molecules in excited states. Which of these effects prevails is determined by the chemical and spectroscopic properties of the individual molecular species. H${}_2$ ------- The intensities of pure rotational and ro-vibrational lines of H${}_2$ contain key information on the structure of C-type shock waves, as was demonstrated, for example, by @Wilgenbus:2000p1267. These lines are optically thin, and their intensities are integrated in parallel with the shock structure, neglecting radiative transfer. As may be seen from Fig. \[FigB1\_H2\], the introduction of shattering leads to a reduction in the computed column densities of the lowest rotational levels of H${}_2$, by approximately an order of magnitude. With increasing energy of the emitting level, the effect of the decrease in the shock width is compensated by the higher peak temperature, and the column densities predicted by the models that include shattering eventually exceed those calculated neglecting shattering. The change-over occurs at energies of the emitting level of $\gtrsim$ 6000 K for $\Vs$ = 20 kms${}^{-1}$, $\gtrsim$ 7000 K for $\Vs$ = 30 kms${}^{-1}$, and $\gtrsim$ 10,000 K for $\Vs$ = 40 kms${}^{-1}$, with the exact values depending on the magnetic field strength. The computed intensities of selected lines of H${}_2$ are given in Table \[tb1\], together with the intensities of forbidden lines of atomic oxygen (63 $\mu$m and 147 $\mu$m), of atomic sulfur (25 $\mu$m) and of \[C I\] (610 $\mu$m and 370 $\mu$m). CO, H${}_2$O and OH ------------------- Figure \[FigB2\_CO\] shows the integrated line intensities, $TdV$, of the rotational transitions of CO, plotted against the rotational quantum number, $J$, of the emitting level for shocks with $n_{\rm H}~=~10^5$ cm$^{-3}$ and $\Vs$ = 30 km s${}^{-1}$. The line intensities, which are listed in Tables \[tbCO20\]–\[tbCO40\], are lower when shattering is included, owing to the reduction in the shock width; this effect is most pronounced at low magnetic field strengths. While the intensities computed with models M2 peak at around $J_{\rm up}$ = 7, those of models M1 peak at higher values of $J_{\rm up}$ and exhibit a plateau extending to $J_{\rm up} \approx 12$. These differences reflect the corresponding excitation conditions. As in the case of SiO (see Sect. 4.2), the peak temperature shows a stronger dependence on the strength of the magnetic field in models that include grain-grain processing. Accordingly, the integrated intensities of highly excited transitions vary with $b$ for models M1. Similarly to CO, the intensities of lines of H$_2$O also become weaker when shattering is included, owing to the reduced shock width. Figure \[FigB3\_oH2O\] shows the computed intensities of the lines of ortho-H${}_2$O as a function of the excitation energy of the emitting level. The intensities of all the lines of ortho- and para-H${}_2$O that fall in the Herschel/PACS/HIFI bands are listed in Tables \[tb3\]–\[tb8\]. ![ The rate of cooling by the principal molecular coolants, H${}_2$ (mauve), H${}_2$O (green), CO (blue) and OH (red) for $\Vs$ = 40 and $b$ = 2 for model M1 (left panel) and M2/M3 (right panel), which are shown together because the presence of SiO in grain mantles does not affect the cooling of the shock wave. Note the different distance intervals on the $x$-axes.[]{data-label="Fig_Cooling"}](FigB5.pdf){width="9cm"} Contrary to the behaviour of CO and H${}_2$O, the line intensities of OH become stronger for the 30 and 40 shocks when shattering is included, as may be seen for $\Vs$ = 30 in Figure \[FigB4\_OH\]. The lines displayed fall within the Herschel/PACS band; their emitting levels have negative parity. The intensities of all transitions of OH observable with Herschel are listed in Table \[tb2\]. The increase in the integrated intensities in models M1 is due to the higher peak temperatures, which favour the conversion of the gas-phase oxygen that is not bound in CO into OH. Again, we see a variation of the integrated line intensities with the magnetic field strength in models M1, associated with the temperature-dependent rate of OH formation. On the basis of these findings, it is interesting to ask how the allowance for grain-grain processing affects the radiative cooling of the medium. Although the narrower shock width, and hence larger velocity gradients, in scenario M1 might be expected to modify the optical thickness of the lines, and thereby their rate of cooling, we do not detect such an effect in our models, as is demonstrated by Fig. \[Fig\_Cooling\]. Instead, we see an increase in the contribution of OH to the rate of cooling, owing to the enhanced abundance of gas-phase OH in the hot shocked medium (see the lower panel of Fig. \[Fig4\_xSiO\]). $$ [llllllll]{} Transition & $\lambda$ ($\mu$m) & v20b1, M1 & (v20b1, M2) & v20b1.5, M1 & (v20b1.5, M2) & v20b2, M1 & (v20b2, M2)\ H${}_2$ 0-0 S(0) & 28.22 & 1.87e-06 & (6.80e-06) & 3.33e-06 & (9.44e-06) & 5.98e-06 & (1.01e-05)\ H${}_2$ 0-0 S(1) & 17.04 & 9.59e-05 & (3.48e-04) & 1.74e-04 & (4.50e-04) & 3.31e-04 & (4.84e-04)\ H${}_2$ 0-0 S(2) & 12.28 & 1.69e-04 & (5.85e-04) & 3.06e-04 & (6.96e-04) & 5.88e-04 & (7.45e-04)\ H${}_2$ 0-0 S(3) & 9.66 & 1.55e-03 & (4.79e-03) & 2.73e-03 & (5.17e-03) & 4.92e-03 & (5.46e-03)\ H${}_2$ 0-0 S(4) & 8.02 & 1.11e-03 & (2.89e-03) & 1.85e-03 & (2.79e-03) & 2.93e-03 & (2.85e-03)\ H${}_2$ 0-0 S(5) & 6.91 & 5.64e-03 & (1.14e-02) & 8.48e-03 & (9.63e-03) & 1.07e-02 & (9.23e-03)\ H${}_2$ 0-0 S(6) & 6.11 & 2.57e-03 & (3.65e-03) & 3.33e-03 & (2.66e-03) & 2.97e-03 & (2.26e-03)\ H${}_2$ 0-0 S(7) & 5.51 & 8.75e-03 & (7.71e-03) & 9.09e-03 & (4.78e-03) & 4.90e-03 & (3.35e-03)\ H${}_2$ 1-0 S(1) & 2.12 & 1.02e-03 & (2.17e-04) & 4.12e-04 & (1.00e-04) & 8.82e-05 & (5.43e-05)\ OI 2p${}^4$ ${}^3$P${}_1$ $\rightarrow$ 2p${}^4$ ${}^3$P${}_2$ & 63.1 & 3.72e-06 & (3.93e-06) & 1.03e-05 & (7.57e-06) & 2.02e-05 & (1.27e-05)\ OI 2p${}^4$ ${}^3$P${}_0$ $\rightarrow$ 2p${}^4$ ${}^3$P${}_1$ & 145.5 & 2.66e-07 & (2.74e-07) & 7.44e-07 & (5.32e-07) & 1.45e-06 & (9.06e-07)\ CI ${}^3$P${}_1$ $\rightarrow$ ${}^3$P${}_0$ & 609.75 & 4.84e-11 & (1.11e-10) & 2.21e-10 & (3.36e-10) & 4.68e-10 & (6.45e-10)\ CI ${}^3$P${}_2$ $\rightarrow$ ${}^3$P${}_1$ & 370.37 & 4.05e-11 & (3.40e-10) & 1.23e-10 & (7.04e-10) & 3.12e-10 & (1.24e-09)\ SI 3P J=1 $\rightarrow$ 3P J=2 & 25.25 & 2.81e-05& (7.03e-05) & 3.71e-05 & (9.61e-05) & 8.44e-05 & (1.21e-04)\ $$\end{center} \label{tb1} \end{table*} %--------------------H2 v30 \begin{table*} \begin{center}$$ [llllllll]{} Transition & $\lambda$ ($\mu$m) & v30b1.5, M1& (v30b1.5, M2) & v30b2, M1& (v30b2, M2) & v30b2.5, M1 & (v30b2.5, M2)\ H${}_2$ 0-0 S(0) & 28.22 & 2.13e-06 & (8.83e-06) & 3.34e-06 & (1.19e-05) & 5.19e-06 & (1.52e-05)\ H${}_2$ 0-0 S(1) & 17.04 & 1.08e-04 & (4.77e-04) & 1.76e-04 & (6.31e-04) & 2.90e-04 & (7.91e-04)\ H${}_2$ 0-0 S(2) & 12.28 & 1.90e-04 & (8.79e-04) & 3.24e-04 & (1.13e-03) & 5.54e-04 & (1.38e-03)\ H${}_2$ 0-0 S(3) & 9.66 & 1.77e-03 & (8.14e-03) & 3.13e-03 & (1.01e-02) & 5.43e-03 & (1.19e-02)\ H${}_2$ 0-0 S(4) & 8.02 & 1.29e-03& (5.68e-03) & 2.34e-03 & (6.78e-03) & 4.03e-03 & (7.57e-03)\ H${}_2$ 0-0 S(5) & 6.91 & 6.87e-03 & (2.71e-02) & 1.26e-02 & (3.04e-02) & 2.06e-02 & (3.15e-02)\ H${}_2$ 0-0 S(6) & 6.11 & 3.31e-03 & (1.10e-02) & 5.96e-03 & (1.13e-02) & 8.99e-03 & (1.06e-02)\ H${}_2$ 0-0 S(7) & 5.51 & 1.27e-02 & (3.20e-02) & 2.15e-02 & (2.88e-02) & 2.83e-02 & (2.32e-02)\ H${}_2$ 1-0 S(1) & 2.12 & 6.92e-03& (3.42e-03) & 4.82e-03 & (2.32e-03) & 2.93e-03 & (1.40e-03)\ OI 2p${}^4$ ${}^3$P${}_1$ $\rightarrow$ 2p${}^4$ ${}^3$P${}_2$ & 63.1 & 4.00e-06 & (4.40e-06) & 8.05e-06 & (7.18e-06) & 1.40e-05 & (1.13e-05)\ OI 2p${}^4$ ${}^3$P${}_0$ $\rightarrow$ 2p${}^4$ ${}^3$P${}_1$ & 145.5 & 2.87e-07 & (2.79e-07) & 5.76e-07 & (4.67e-07) & 1.01e-06 & (7.37e-07)\ CI ${}^3$P${}_1$ $\rightarrow$ ${}^3$P${}_0$ & 609.75 & 1.88e-10 & (1.24e-10) & 1.85e-10 & (3.04e-10) & 3.66e-10 & (5.24e-10)\ CI ${}^3$P${}_2$ $\rightarrow$ ${}^3$P${}_1$ & 370.37 & 1.18e-09 & (3.98e-10) & 1.09e-10 & (6.58e-10) & 2.01e-10 & (9.67e-10)\ SI 3P J=1 $\rightarrow$ 3P J=2 & 25.25 & 3.93e-04& (6.10e-04) & 2.56e-04 & (7.34e-04) & 1.56e-04 & (8.24e-04)\ $$\end{center} %\label{t4a} \end{table*} %----------------------H2 v40 \begin{table*} \begin{center}$$ [llllllll]{} Transition & $\lambda$ ($\mu$m) & v40b2, M1 & (v40b2, M2) & v40b2.5, M1 & (v40b2.5, M2) & v40b3, M1 & (v40b3, M2)\ H${}_2$ 0-0 S(0) & 28.22 & 3.54e-06 & (1.08e-05) & 3.60e-06 & (1.35e-05) & 5.02e-06 & (1.65e-05)\ H${}_2$ 0-0 S(1) & 17.04 & 1.46e-04 & (5.81e-04) & 1.81e-04 & (7.26e-04) & 2.73e-04 & (8.79e-04)\ H${}_2$ 0-0 S(2) & 12.28 & 2.03e-04 & (1.07e-03) & 3.21e-04 & (1.34e-03) & 5.21e-04 & (1.62e-03)\ H${}_2$ 0-0 S(3) & 9.66 & 1.59e-03 & (1.02e-02) & 3.04e-03 & (1.27e-02) & 5.21e-04 & (1.51e-02)\ H${}_2$ 0-0 S(4) & 8.02 & 1.09e-03& (7.27e-03) & 2.25e-03 & (9.10e-03) & 4.00e-03 & (1.07e-02)\ H${}_2$ 0-0 S(5) & 6.91 & 5.78e-03 & (3.65e-02) & 1.23e-02 & (4.54e-02) & 2.20e-02 & (5.19e-02)\ H${}_2$ 0-0 S(6) & 6.11 &2.97e-03 & (1.59e-02) &6.03e-03 & (1.95e-02) & 1.06e-02 & (2.14e-02)\ H${}_2$ 0-0 S(7) & 5.51 & 1.24e-02 & (5.21e-02) & 2.36e-02 & (6.15e-02) & 3.92e-02 & (6.35e-02)\ H${}_2$ 1-0 S(1) & 2.12 & 1.34e-02& (1.58e-02) & 1.59e-02 & (1.15e-02) & 1.31e-02 & (8.75e-03)\ OI 2p${}^4$ ${}^3$P${}_1$ $\rightarrow$ 2p${}^4$ ${}^3$P${}_2$ & 63.1 & 3.72e-05 & (3.48e-06) & 7.41e-06 & (5.98e-06) & 1.17e-05 & (9.20e-06)\ OI 2p${}^4$ ${}^3$P${}_0$ $\rightarrow$ 2p${}^4$ ${}^3$P${}_1$ & 145.5 & 1.36e-06 & (2.44e-07) & 5.34e-07 & (3.96e-07) & 8.42e-07& (5.97e-07)\ CI ${}^3$P${}_1$ $\rightarrow$ ${}^3$P${}_0$ & 609.75 & 5.50e-08 & (6.66e-10) & 1.04e-08 & (4.27e-10) & 5.34e-10 & (5.95e-10)\ CI ${}^3$P${}_2$ $\rightarrow$ ${}^3$P${}_1$ & 370.37 & 4.73e-07 & (2.26e-09) & 8.16e-08 & (1.04e-09) & 1.83e-09 & (1.27e-09)\ SI 3P J=1 $\rightarrow$ 3P J=2 & 25.25 & 1.35e-03& (1.43e-03) & 7.07e-04 & (1.37e-03) & 5.97e-04 & (1.40e-03)\ $$\end{center} %\label{t4a} \end{table*} %--------------------CO v20 \begin{table*} \caption[]{ Intensities of CO lines (erg\,cm${}^{-2}$\,s${}^{-1}$\,sr${}^{-1}$) for shocks with velocity $\Vs$~=~20~\kms and the magnetic field strengths listed in Table~1. Results are given for models M1, which include grain-grain processing, and M2 (in parentheses), which neglect grain-grain processing. The preshock density is $n{}_{\rm H}$ = 10${}^5$ cm${}^{-3}$.} \begin{center}$$ [lllllllll]{} Transition & Freq. (GHz) & $\lambda$ ($\mu$m) & v20b1, M1 & (v20b1, M2) & v20b1.5, M1 & (v20b1.5, M2) & v20b2, M1 &(v20b2, M2)\ CO (1–0) & 115.27 & 2600.7 & 3.47e-08 & (1.85e-07) & 9.49e-08 & (2.63e-07) & 1.47e-07 & (2.69e-07)\ CO (2–1) & 230.54 & 1300.4 & 6.39e-07 & (2.41e-06) & 1.28e-06 & (3.23e-06) & 1.76e-06 & (3.14e-06)\ CO (3–2) & 345.80 & 866.96 & 2.72e-06 & (1.01e-05) & 4.69e-06 & (1.30e-05) & 6.33e-06 & (1.22e-05)\ CO (4–3) & 461.04 & 650.25 & 7.01e-06 & (2.73e-05) & 1.15e-05 & (3.50e-05) & 1.57e-05 & (3.24e-05)\ CO (5–4) & 576.27 & 520.23 & 1.41e-05 & (5.78e-05) & 2.36e-05 & (7.55e-05) & 3.40e-05 & (7.07e-05)\ CO (6–5) & 691.47 & 433.55 & 2.42e-05 & (1.00e-04) & 4.22e-05 & (1.36e-04) & 6.46e-05 & (1.30e-04)\ CO (7–6) & 806.65 & 371.65 & 3.76e-05 & (1.51e-04) & 6.71e-05 & (2.09e-04) & 1.09e-04 & (2.06e-04)\ CO (8–7) & 921.80 & 325.22 & 5.51e-05 & (2.13e-04) & 9.84e-05 & (2.91e-04) & 1.66e-04 & (2.94e-04)\ CO (9–8) & 1036.9 & 289.12 & 7.68e-05 & (2.83e-04) & 1.36e-04 & (3.76e-04) & 2.32e-04 & (3.86e-04)\ CO (10–9) & 1152.0 & 260.24 & 1.02e-04 & (3.54e-04) & 1.76e-04 & (4.52e-04) & 3.01e-04 & (4.70e-04)\ CO (11–10) & 1267.0 & 236.61 & 1.29e-04 & (4.16e-04) & 2.17e-04 & (5.09e-04) & 3.63e-04 & (5.31e-04)\ CO (12–11) & 1382.0 & 216.93 & 1.56e-04 & (4.57e-04) & 2.52e-04 & (5.38e-04) & 4.09e-04 & (5.62e-04)\ CO (13–12) & 1496.9 & 200.27 & 1.78e-04 & (4.72e-04) & 2.75e-04 & (5.33e-04) & 4.32e-04 & (5.58e-04)\ CO (14–13) & 1611.8 & 186.00 & 1.94e-04 & (4.64e-04) & 2.87e-04 & (5.03e-04) & 4.34e-04 & (5.25e-04)\ CO (15–14) & 1726.6 & 173.63 & 2.00e-04 & (4.29e-04) & 2.83e-04 & (4.47e-04) & 4.11e-04 & (4.65e-04)\ CO (16–15) & 1841.3 & 162.81 & 2.02e-04 & (3.94e-04) & 2.75e-04 & (3.95e-04) & 3.84e-04 & (4.10e-04)\ CO (17–16) & 1956.0 & 153.27 & 1.98e-04 & (3.52e-04) & 2.59e-04 & (3.42e-04) & 3.49e-04 & (3.52e-04)\ CO (18–17) & 2070.6 & 144.78 & 1.89e-04 & (3.09e-04) & 2.39e-04 & (2.91e-04) & 3.09e-04 & (2.97e-04)\ CO (19–18) & 2185.1 & 137.20 & 1.78e-04 & (2.67e-04) & 2.17e-04 & (2.44e-04) & 2.70e-04 & (2.48e-04)\ CO (20–19) & 2299.6 & 130.37 & 1.64e-04 & (2.28e-04) & 1.94e-04 & (2.03e-04) & 2.32e-04 & (2.04e-04)\ CO (21–20) & 2413.9 & 124.19 & 1.47e-04 & (1.91e-04) & 1.70e-04 & (1.66e-04) & 1.96e-04 & (1.65e-04)\ CO (22–21) & 2528.2 & 118.58 & 1.31e-04 & (1.59e-04) & 1.47e-04 & (1.34e-04) & 1.63e-04 & (1.33e-04)\ CO (23–22) & 2642.3 & 113.46 & 1.15e-04 & (1.31e-04) & 1.26e-04 & (1.08e-04) & 1.34e-04 & (1.06e-04)\ CO (24–23) & 2756.4 & 108.76 & 9.98e-05 & (1.07e-04) & 1.07e-04 & (8.63e-05) & 1.09e-04 & (8.34e-05)\ CO (25–24) & 2870.3 & 104.44 & 8.58e-05 & (8.66e-05) & 8.98e-05 & (6.85e-05) & 8.83e-05 & (6.54e-05)\ CO (26–25) & 2984.2 & 100.46 & 7.32e-05 & (6.98e-05) & 7.50e-05 & (5.41e-05) & 7.08e-05 & (5.10e-05)\ CO (27–26) & 3097.9 & 96.772 & 6.20e-05 & (5.60e-05) & 6.23e-05 & (4.25e-05) & 5.63e-05 & (3.95e-05)\ CO (28–27) & 3211.5 & 93.348 & 5.21e-05 & (4.46e-05) & 5.13e-05 & (3.32e-05) & 4.44e-05 & (3.04e-05)\ CO (29–28) & 3325.0 & 90.162 & 4.35e-05 & (3.53e-05) & 4.20e-05 & (2.58e-05) & 3.48e-05 & (2.33e-05)\ CO (30–29) & 3438.4 & 87.190 & 3.59e-05 & (2.78e-05) & 3.40e-05 & (1.99e-05) & 2.70e-05 & (1.77e-05)\ CO (31–30) & 3551.6 & 84.410 & 2.95e-05 & (2.17e-05) & 2.74e-05 & (1.52e-05) & 2.08e-05 & (1.33e-05)\ CO (32–31) & 3664.7 & 81.805 & 2.40e-05 & (1.68e-05) & 2.19e-05 & (1.16e-05) & 1.58e-05 & (9.96e-06)\ CO (33–32) & 3777.6 & 79.359 & 1.92e-05 & (1.29e-05) & 1.72e-05 & (8.71e-06) & 1.19e-05 & (7.37e-06)\ CO (34–33) & 3890.4 & 77.058 & 1.52e-05 & (9.72e-06) & 1.34e-05 & (6.48e-06) & 8.87e-06 & (5.39e-06)\ CO (35–34) & 4003.1 & 74.889 & 1.18e-05 & (7.21e-06) & 1.02e-05 & (4.73e-06) & 6.46e-06 & (3.87e-06)\ CO (36–35) & 4115.6 & 72.842 & 8.96e-06 & (5.22e-06) & 7.60e-06 & (3.37e-06) & 4.60e-06 & (2.71e-06)\ CO (37–36) & 4228.0 & 70.907 & 6.52e-06 & (3.64e-06) & 5.43e-06 & (2.32e-06) & 3.15e-06 & (1.84e-06)\ CO (38–37) & 4340.1 & 69.074 & 4.46e-06 & (2.39e-06) & 3.65e-06 & (1.50e-06) & 2.03e-06 & (1.17e-06)\ CO (39–38) & 4452.2 & 67.336 & 2.72e-06 & (1.41e-06) & 2.19e-06 & (8.72e-07) & 1.17e-06 & (6.69e-07)\ CO (40–39) & 4564.0 & 65.686 & 1.25e-06 & (6.22e-07) & 9.89e-07 & (3.81e-07) & 5.10e-07 & (2.88e-07)\ $$\end{center} \label{tbCO20} \end{table*} %-----------------------------------------COv30 \begin{table*} \caption[]{ Intensities of CO lines (erg\,cm${}^{-2}$\,s${}^{-1}$\,sr${}^{-1}$) for shocks with velocity $\Vs$~=~30~\kms and the magnetic field strengths listed in Table~1. Results are given for models M1, which include grain-grain processing, and M2 (in parentheses), which neglect grain-grain processing. The preshock density is $n{}_{\rm H}$ = 10${}^5$ cm${}^{-3}$.} \begin{center}$$ [lllllllll]{} Transition & Freq. (GHz) & $\lambda$ ($\mu$m) & v30b1.5, M1 & (v30b1.5, M2) & v30b2, M1 & (v30b2, M2) & v30b2.5, M1 & (v30b2.5, M2)\ CO (1–0) & 115.27 & 2600.7 & 4.01e-08 & (2.64e-07) & 9.93e-08 & (3.42e-07) & 1.65e-07 & (4.14e-07)\ CO (2–1) & 230.54 & 1300.4 & 7.45e-07 & (3.35e-06) & 1.41e-06 & (4.14e-06) & 2.05e-06 & (4.88e-06)\ CO (3–2) & 345.80 & 866.96 & 3.19e-06 & (1.37e-05) & 5.06e-06 & (1.65e-05) & 7.19e-06 & (1.92e-05)\ CO (4–3) & 461.04 & 650.25 & 8.29e-06 & (3.68e-05) & 1.22e-05 & (4.39e-05) & 1.72e-05 & (5.12e-05)\ CO (5–4) & 576.27 & 520.23 & 1.68e-05 & (7.71e-05) & 2.49e-05 & (9.47e-05) & 3.57e-05 & (1.12e-04)\ CO (6–5) & 691.47 & 433.55 & 2.89e-05 & (1.33e-04) & 4.42e-05 & (1.71e-04) & 6.54e-05 & (2.07e-04)\ CO (7–6) & 806.65 & 371.65 & 4.49e-05 & (2.01e-04) & 7.03e-05 & (2.69e-04) & 1.07e-04 & (3.30e-04)\ CO (8–7) & 921.80 & 325.22 & 6.56e-05 & (2.86e-04) & 1.04e-04 & (3.88e-04) & 1.61e-04 & (4.79e-04)\ CO (9–8) & 1036.9 & 289.12 & 9.13e-05 & (3.90e-04) & 1.44e-04 & (5.26e-04) & 2.26e-04 & (6.46e-04)\ CO (10–9) & 1152.0 & 260.24 & 1.22e-04 & (5.06e-04) & 1.92e-04 & (6.74e-04) & 3.00e-04 & (8.15e-04)\ CO (11–10) & 1267.0 & 236.61 & 1.55e-04 & (6.22e-04) & 2.43e-04 & (8.13e-04) & 3.77e-04 & (9.65e-04)\ CO (12–11) & 1382.0 & 216.93 & 1.88e-04 & (7.22e-04) & 2.93e-04 & (9.23e-04) & 4.48e-04 & (1.07e-03)\ CO (13–12) & 1496.9 & 200.27 & 2.18e-04 & (7.91e-04) & 3.35e-04 & (9.86e-04) & 5.03e-04 & (1.12e-03)\ CO (14–13) & 1611.8 & 186.00 & 2.42e-04 & (8.26e-04) & 3.66e-04 & (1.00e-03) & 5.40e-04 & (1.12e-03)\ CO (15–14) & 1726.6 & 173.63 & 2.55e-04 & (8.14e-04) & 3.80e-04 & (9.65e-04) & 5.50e-04 & (1.05e-03)\ CO (16–15) & 1841.3 & 162.81 & 2.63e-04 & (7.89e-04) & 3.86e-04 & (9.15e-04) & 5.49e-04 & (9.77e-04)\ CO (17–16) & 1956.0 & 153.27 & 2.62e-04 & (7.45e-04) & 3.81e-04 & (8.46e-04) & 5.34e-04 & (8.86e-04)\ CO (18–17) & 2070.6 & 144.78 & 2.55e-04 & (6.87e-04) & 3.68e-04 & (7.65e-04) & 5.07e-04 & (7.87e-04)\ CO (19–18) & 2185.1 & 137.20 & 2.43e-04 & (6.23e-04) & 3.48e-04 & (6.81e-04) & 4.73e-04 & (6.89e-04)\ CO (20–19) & 2299.6 & 130.37 & 2.27e-04 & (5.57e-04) & 3.24e-04 & (5.98e-04) & 4.35e-04 & (5.96e-04)\ CO (21–20) & 2413.9 & 124.19 & 2.07e-04 & (4.88e-04) & 2.94e-04 & (5.16e-04) & 3.91e-04 & (5.06e-04)\ CO (22–21) & 2528.2 & 118.58 & 1.86e-04 & (4.23e-04) & 2.64e-04 & (4.40e-04) & 3.47e-04 & (4.25e-04)\ CO (23–22) & 2642.3 & 113.46 & 1.65e-04 & (3.62e-04) & 2.34e-04 & (3.72e-04) & 3.05e-04 & (3.54e-04)\ CO (24–23) & 2756.4 & 108.76 & 1.45e-04 & (3.08e-04) & 2.05e-04 & (3.11e-04) & 2.65e-04 & (2.93e-04)\ CO (25–24) & 2870.3 & 104.44 & 1.25e-04 & (2.60e-04) & 1.78e-04 & (2.59e-04) & 2.28e-04 & (2.40e-04)\ CO (26–25) & 2984.2 & 100.46 & 1.08e-04 & (2.18e-04) & 1.54e-04 & (2.14e-04) & 1.95e-04 & (1.96e-04)\ CO (27–26) & 3097.9 & 96.772 & 9.19e-05 & (1.81e-04) & 1.32e-04 & (1.76e-04) & 1.66e-04 & (1.58e-04)\ CO (28–27) & 3211.5 & 93.348 & 7.78e-05 & (1.50e-04) & 1.12e-04 & (1.43e-04) & 1.39e-04 & (1.27e-04)\ CO (29–28) & 3325.0 & 90.162 & 6.53e-05 & (1.23e-04) & 9.40e-05 & (1.16e-04) & 1.17e-04 & (1.02e-04)\ CO (30–29) & 3438.4 & 87.190 & 5.43e-05 & (1.00e-04) & 7.85e-05 & (9.33e-05) & 9.66e-05 & (8.07e-05)\ CO (31–30) & 3551.6 & 84.410 & 4.49e-05 & (8.08e-05) & 6.51e-05 & (7.44e-05) & 7.94e-05 & (6.35e-05)\ CO (32–31) & 3664.7 & 81.805 & 3.67e-05 & (6.47e-05) & 5.34e-05 & (5.89e-05) & 6.46e-05 & (4.95e-05)\ CO (33–32) & 3777.6 & 79.359 & 2.97e-05 & (5.12e-05) & 4.33e-05 & (4.60e-05) & 5.19e-05 & (3.81e-05)\ CO (34–33) & 3890.4 & 77.058 & 2.37e-05 & (4.00e-05) & 3.45e-05 & (3.55e-05) & 4.11e-05 & (2.90e-05)\ CO (35–34) & 4003.1 & 74.889 & 1.85e-05 & (3.06e-05) & 2.70e-05 & (2.68e-05) & 3.19e-05 & (2.16e-05)\ CO (36–35) & 4115.6 & 72.842 & 1.41e-05 & (2.29e-05) & 2.06e-05 & (1.98e-05) & 2.42e-05 & (1.57e-05)\ CO (37–36) & 4228.0 & 70.907 & 1.03e-05 & (1.64e-05) & 1.51e-05 & (1.40e-05) & 1.76e-05 & (1.10e-05)\ CO (38–37) & 4340.1 & 69.074 & 7.11e-06 & (1.11e-05) & 1.04e-05 & (9.36e-06) & 1.20e-05 & (7.26e-06)\ CO (39–38) & 4452.2 & 67.336 & 4.37e-06 & (6.68e-06) & 6.39e-06 & (5.58e-06) & 7.31e-06 & (4.27e-06)\ CO (40–39) & 4564.0 & 65.686 & 2.02e-06 & (3.03e-06) & 2.95e-06 & (2.50e-06) & 3.35e-06 & (1.89e-06)\ $$\end{center} \label{tbCO30} \end{table*} %----------------------------------------- %-----------------------------------------COv40 \begin{table*} \caption[]{ Intensities of CO lines (erg\,cm${}^{-2}$\,s${}^{-1}$\,sr${}^{-1}$) for shocks with velocity $\Vs$~=~40~\kms and the magnetic field strengths listed in Table~1. Results are given for models M1, which include grain-grain processing, and M2 (in parentheses), which neglect grain-grain processing. The preshock density is $n{}_{\rm H}$ = 10${}^5$ cm${}^{-3}$.} \begin{center}$$ [lllllllll]{} Transition & Freq. (GHz) & $\lambda$ ($\mu$m) & v40b2, M1 & (v40b2, M2) & v40b2.5, M1 & (v40b2.5, M2) & v40b3, M1 & (v40b3, M2)\ CO (1–0) & 115.27 & 2600.7 & 3.42e-08 & (3.41e-07) & 1.00e-07 & (4.18e-07) & 1.73e-07 & (4.86e-07)\ CO (2–1) & 230.54 & 1300.4 & 6.14e-07 & (4.33e-06) & 1.45e-06 & (5.02e-06) & 2.22e-06 & (5.69e-06)\ CO (3–2) & 345.80 & 866.96 & 2.73e-06 & (1.76e-05) & 5.30e-06 & (1.98e-05) & 7.52e-06 & (2.21e-05)\ CO (4–3) & 461.04 & 650.25 & 7.90e-06 & (4.67e-05) & 1.30e-05 & (5.22e-05) & 1.76e-05 & (5.82e-05)\ CO (5–4) & 576.27 & 520.23 & 1.81e-05 & (9.75e-05) & 2.68e-05 & (1.11e-04) & 3.63e-05 & (1.25e-04)\ CO (6–5) & 691.47 & 433.55 & 3.51e-05 & (1.68e-04) & 4.79e-05 & (1.99e-04) & 6.58e-05 & (2.30e-04)\ CO (7–6) & 806.65 & 371.65 & 6.04e-05 & (2.53e-04) & 7.70e-05 & (3.11e-04) & 1.07e-04 & (3.66e-04)\ CO (8–7) & 921.80 & 325.22 & 9.53e-05 & (3.58e-04) & 1.14e-04 & (4.49e-04) & 1.61e-04 & (5.35e-04)\ CO (9–8) & 1036.9 & 289.12 & 1.40e-04 & (4.91e-04) & 1.59e-04 & (6.17e-04) & 2.27e-04 & (7.36e-04)\ CO (10–9) & 1152.0 & 260.24 & 1.93e-04 & (6.46e-04) & 2.11e-04 & (8.08e-04) & 3.03e-04 & (9.57e-04)\ CO (11–10) & 1267.0 & 236.61 & 2.51e-04 & (8.09e-04) & 2.67e-04 & (1.00e-03) & 3.84e-04 & (1.18e-03)\ CO (12–11) & 1382.0 & 216.93 & 3.11e-04 & (9.61e-04) & 3.23e-04 & (1.18e-03) & 4.64e-04 & (1.37e-03)\ CO (13–12) & 1496.9 & 200.27 & 3.66e-04 & (1.08e-03) & 3.71e-04 & (1.31e-03) & 5.32e-04 & (1.50e-03)\ CO (14–13) & 1611.8 & 186.00 & 4.12e-04 & (1.15e-03) & 4.09e-04 & (1.38e-03) & 5.84e-04 & (1.56e-03)\ CO (15–14) & 1726.6 & 173.63 & 4.39e-04 & (1.17e-03) & 4.29e-04 & (1.38e-03) & 6.09e-04 & (1.54e-03)\ CO (16–15) & 1841.3 & 162.81 & 4.57e-04 & (1.16e-03) & 4.40e-04 & (1.36e-03) & 6.22e-04 & (1.50e-03)\ CO (17–16) & 1956.0 & 153.27 & 4.61e-04 & (1.11e-03) & 4.39e-04 & (1.30e-03) & 6.17e-04 & (1.42e-03)\ CO (18–17) & 2070.6 & 144.78 & 4.52e-04 & (1.04e-03) & 4.26e-04 & (1.21e-03) & 5.98e-04 & (1.31e-03)\ CO (19–18) & 2185.1 & 137.20 & 4.33e-04 & (9.61e-04) & 4.05e-04 & (1.11e-03) & 5.68e-04 & (1.19e-03)\ CO (20–19) & 2299.6 & 130.37 & 4.07e-04 & (8.71e-04) & 3.79e-04 & (1.00e-03) & 5.31e-04 & (1.07e-03)\ CO (21–20) & 2413.9 & 124.19 & 3.74e-04 & (7.73e-04) & 3.46e-04 & (8.86e-04) & 4.85e-04 & (9.36e-04)\ CO (22–21) & 2528.2 & 118.58 & 3.37e-04 & (6.76e-04) & 3.11e-04 & (7.74e-04) & 4.37e-04 & (8.13e-04)\ CO (23–22) & 2642.3 & 113.46 & 3.01e-04 & (5.86e-04) & 2.76e-04 & (6.69e-04) & 3.89e-04 & (6.99e-04)\ CO (24–23) & 2756.4 & 108.76 & 2.65e-04 & (5.02e-04) & 2.42e-04 & (5.74e-04) & 3.43e-04 & (5.96e-04)\ CO (25–24) & 2870.3 & 104.44 & 2.30e-04 & (4.27e-04) & 2.11e-04 & (4.88e-04) & 2.99e-04 & (5.04e-04)\ CO (26–25) & 2984.2 & 100.46 & 1.99e-04 & (3.61e-04) & 1.82e-04 & (4.12e-04) & 2.59e-04 & (4.23e-04)\ CO (27–26) & 3097.9 & 96.772 & 1.70e-04 & (3.03e-04) & 1.56e-04 & (3.45e-04) & 2.23e-04 & (3.53e-04)\ CO (28–27) & 3211.5 & 93.348 & 1.45e-04 & (2.52e-04) & 1.32e-04 & (2.88e-04) & 1.90e-04 & (2.93e-04)\ CO (29–28) & 3325.0 & 90.162 & 1.22e-04 & (2.08e-04) & 1.11e-04 & (2.38e-04) & 1.61e-04 & (2.41e-04)\ CO (30–29) & 3438.4 & 87.190 & 1.03e-04 & (1.71e-04) & 9.31e-05 & (1.95e-04) & 1.35e-04 & (1.97e-04)\ CO (31–30) & 3551.6 & 84.410 & 8.53e-05 & (1.39e-04) & 7.72e-05 & (1.59e-04) & 1.12e-04 & (1.60e-04)\ CO (32–31) & 3664.7 & 81.805 & 7.03e-05 & (1.12e-04) & 6.34e-05 & (1.28e-04) & 9.22e-05 & (1.28e-04)\ CO (33–32) & 3777.6 & 79.359 & 5.73e-05 & (8.96e-05) & 5.15e-05 & (1.02e-04) & 7.50e-05 & (1.02e-04)\ CO (34–33) & 3890.4 & 77.058 & 4.61e-05 & (7.05e-05) & 4.12e-05 & (8.06e-05) & 6.01e-05 & (7.99e-05)\ CO (35–34) & 4003.1 & 74.889 & 3.63e-05 & (5.44e-05) & 3.23e-05 & (6.22e-05) & 4.72e-05 & (6.13e-05)\ CO (36–35) & 4115.6 & 72.842 & 2.79e-05 & (4.09e-05) & 2.47e-05 & (4.68e-05) & 3.61e-05 & (4.59e-05)\ CO (37–36) & 4228.0 & 70.907 & 2.06e-05 & (2.96e-05) & 1.82e-05 & (3.38e-05) & 2.66e-05 & (3.31e-05)\ CO (38–37) & 4340.1 & 69.074 & 1.43e-05 & (2.02e-05) & 1.26e-05 & (2.30e-05) & 1.84e-05 & (2.24e-05)\ CO (39–38) & 4452.2 & 67.336 & 8.88e-06 & (1.22e-05) & 7.73e-06 & (1.40e-05) & 1.13e-05 & (1.35e-05)\ CO (40–39) & 4564.0 & 65.686 & 4.13e-06 & (5.58e-06) & 3.58e-06 & (6.36e-06) & 5.22e-06 & (6.14e-06)\ $$\end{center} \label{tbCO40} \end{table*} %----------------------------------------- %------------------------------------------------ OH v20 \begin{table*} \caption[]{Intensities of OH lines (erg\,cm${}^{-2}$\,s${}^{-1}$\,sr${}^{-1}$) observable with the PACS instrument on the Herschel Space Observatory, for shocks with velocities $\Vs$~=~20~\kms (top), $\Vs$~=~30~\kms (middle), and $\Vs$~=~40~\kms (bottom) and the magnetic field strengths listed in Table 1. Results are given for models M1, which include grain-grain processing, and M2 (in parentheses), which neglect grain-grain processing. The preshock density is $n{}_{\rm H}$ = 10${}^5$ cm${}^{-3}$.} \begin{center}$$ [lllllllll]{} Transition & $\lambda$ ($\mu$m) & $ E_{\rm up}$ (K)& v20b1, M1 & (v20b1, M2) & v20b1.5, M1 & (v20b1.5, M2) & v20b2, M1 & (v20b2, M2)\ 1834.8 GHz & 163.39 & 269.8 &1.65e-08 & (1.76e-08) & 2.58e-08& (2.51e-08) & 3.92e-08& (3.03e-08)\ 1839.0 GHz& 163.01 & 270.2 & 2.42e-08& (2.54e-08) & 3.75e-08& (3.58e-08) & 5.65e-08& (4.35e-08)\ 2510.0 GHz & 119.44 & 120.5 & 1.79e-07& (2.71e-07) & 2.88e-07& (4.89e-07) & 4.38e-07& (4.20e-07)\ 2514.0 GHz& 119.23 & 120.7 & 9.78e-08& (1.65e-07) & 1.60e-07& (3.14e-07) & 2.44e-07& (2.50e-07)\ 3035.0 GHz& 98.76 & 415.9 & 3.11e-08& (3.08e-08)& 4.61e-08& (4.01e-08) & 6.78e-08& (5.11e-08)\ 3036.0 GHz& 98.74 & 415.5 & 1.00e-08& (1.00e-08) & 1.50e-08& (1.31e-08) & 2.22e-08& (1.68e-08)\ 3111.0 GHz& 96.36 & 269.8 & 4.01e-09& (4.28e-09) & 6.28e-09& (6.11e-09) & 9.54e-09& (7.36e-09)\ 3114.0 GHz& 96.27 & 270.2 & 5.86e-09& (6.16e-09) & 9.07e-09& (8.67e-09) & 1.37e-08& (1.05e-08)\ 3544.0 GHz& 84.60 & 290.5 & 8.10e-08& (8.30e-08) & 1.24e-07& (1.13e-07) & 1.86e-07& (1.41e-07)\ 3551.0 GHz& 84.42 & 291.2 & 1.94e-08& (2.01e-08) & 3.00e-08& (2.76e-08) & 4.52e-08& (3.44e-08)\ 3786.0 GHz& 79.18 & 181.7 & 1.01e-07& (1.19e-07) & 1.61e-07& (1.88e-07) & 2.47e-07& (2.03e-07)\ 3789.0 GHz& 79.12 & 181.9 & 1.11e-07& (1.27e-07) & 1.76e-07& (1.96e-07) & 2.68e-07& (2.17e-07)\ 4210.0 GHz& 71.22 & 617.9 & 1.51e-08& (1.45e-08) & 2.13e-08& (1.81e-08) & 3.04e-08& (2.31e-08)\ 4212.0 GHz& 71.17 & 617.6 & 2.92e-09& (2.81e-09) & 4.17e-09& (3.53e-09) & 5.96e-09& (4.52e-09)\ 4592.0 GHz& 65.28 & 510.9 & 2.78e-08& (2.71e-08) & 4.05e-08& (3.46e-08) & 5.87e-08& (4.43e-08)\ 4603.0 GHz& 65.13 & 512.1 & 4.03e-09& (3.94e-09) & 5.90e-09& (5.05e-09) & 8.58e-09& (6.47e-09)\ $$\end{center} \label{tb2} \end{table*} %-----------------------OH v30 \begin{table*} \begin{center}$$ [lllllllll]{} Transition & $\lambda$ ($\mu$m)& $E_{\rm up}$ (K) & v30b1.5, M1& (v30b1.5, M2) & v30b2, M1& (v30b2, M2) & v30b2.5, M1 & (v30b2.5, M2)\ 1834.8 GHz & 163.39 & 269.8 & 1.33e-07& (5.28e-08) & 6.54e-08& (5.38e-08) & 5.46e-08& (5.24e-08)\ 1839.0 GHz& 163.01 & 270.2 & 1.98e-07& (7.76e-08) & 9.67e-08& (7.86e-08) &8.00e-08 & (7.62e-08)\ 2510.0 GHz& 119.44 & 120.5 & 4.38e-07& (6.17e-07) & 6.95e-07& (6.58e-07) & 5.94e-07& (6.68e-07)\ 2514.0 GHz& 119.23 & 120.7 & 7.47e-07& (3.46e-07) & 3.77e-07& (3.76e-07) & 3.26e-07& (3.88e-07)\ 3035.0 GHz& 98.76 & 415.9 & 2.66e-07& (1.00e-07)& 1.27e-07& (9.94e-08) & 1.02e-07& (9.38e-08)\ 3036.0 GHz& 98.74 & 415.5 & 8.54e-08& (3.23e-08) & 4.08e-08& (3.21e-08) & 3.28e-08& (3.04e-08)\ 3111.0 GHz& 96.36 & 269.8 & 3.24e-08& (1.28e-08) & 1.59e-08& (1.31e-08) & 1.34e-08& (1.27e-08)\ 3114.0 GHz& 96.27 & 270.2 & 4.80e-08& (1.88e-08) & 2.34e-08& (1.90e-08) & 1.94e-08& (1.84e-08)\ 3544.0 GHz& 84.60 & 290.5 & 6.69e-07& (2.59e-07) & 3.24e-07& (2.61e-07) & 2.66e-07& (2.50e-07)\ 3551.0 GHz& 84.42 & 291.2 & 1.59e-07& (6.20e-08) & 7.73e-08& (6.26e-08) & 6.38e-08& (6.04e-08)\ 3786.0 GHz& 79.18 & 181.7 & 7.87e-07& (3.27e-07) & 3.93e-07& (3.40e-07) & 3.34e-07& (3.39e-07)\ 3789.0 GHz& 79.12 & 181.9 & 8.78e-07& (3.59e-07) & 4.36e-07& (3.71e-07) & 3.67e-07& (3.67e-07)\ 4210.0 GHz& 71.22 & 617.9 & 1.37e-07& (4.97e-08) & 6.34e-08& (4.83e-08) & 4.89e-08& (4.45e-08)\ 4212.0 GHz& 71.17 & 617.6 & 2.64e-08& (9.61e-09) & 1.22e-08& (9.35e-09) & 9.50e-09& (8.65e-09)\ 4592.0 GHz& 65.28& 510.9 & 2.45e-07& (9.07e-08) & 1.15e-07& (8.90e-08) & 9.07e-08& (8.31e-08)\ 4603.0 GHz& 65.13 & 512.1 & 3.53e-08& (1.31e-08) & 1.66e-08& (1.29e-08) & 1.31e-08& (1.21e-08)\ $$\end{center} %\label{t4a} \end{table*} %--------------------OH v40 \begin{table*} \begin{center}$$ [lllllllll]{} Transition & $\lambda$ ($\mu$m)& $ E_{\rm up}$ (K)& v40b2, M1 & (v40b2, M2) & v40b2.5, M1 & (v40b2.5, M2) & v40b3, M1 & (v40b3, M2)\ 1834.8 GHz & 163.39 & 269.8 & 6.25e-06& (2.49e-07) & 4.66e-07& (1.67e-07) & 2.05e-07& (1.36e-07)\ 1839.0 GHz& 163.01 & 270.2 & 9.23e-06& (3.69e-07) & 6.95e-07& (2.47e-07) & 3.05e-07& (2.01e-07)\ 2510.0 GHz& 119.44 & 120.5 & 6.49e-05& (2.67e-06) & 4.84e-06& (1.86e-06) & 2.15e-06& (1.58e-06)\ 2514.0 GHz& 119.23 & 120.7 & 3.68e-05& (1.45e-06) & 2.60e-06& (1.02e-06) & 1.16e-06& (8.84e-07)\ 3035.0 GHz& 98.76 & 415.9 & 1.23e-05& (4.90e-07)& 9.37e-07& (3.25e-07) & 4.06e-07& (2.60e-07)\ 3036.0 GHz& 98.74 & 415.5 & 4.21e-06& (1.57e-07) & 3.01e-07& (1.04e-07) & 1.30e-07& (8.37e-08)\ 3111.0 GHz& 96.36 & 269.8 & 1.52e-06& (6.06e-08) & 1.13e-07& (4.07e-08) & 4.99e-08& (3.32e-08)\ 3114.0 GHz& 96.27 & 270.2 & 2.24e-06& (8.94e-08) & 1.68e-07& (5.99e-08) & 7.38e-08& (4.87e-08)\ 3544.0 GHz& 84.60 & 290.5 & 3.11e-05& (1.24e-06) & 2.35e-06& (8.30e-07) &1.03e-06& (6.70e-07)\ 3551.0 GHz& 84.42 & 291.2 & 7.93e-06& (2.96e-07) & 5.59e-07& (1.98e-07) & 2.44e-07& (1.60e-07)\ 3786.0 GHz& 79.18 & 181.7 & 3.54e-05& (1.49e-06) & 2.74e-06& (1.02e-06) & 1.22e-06& (8.42e-07)\ 3789.0 GHz& 79.12 & 181.9 & 3.90e-05& (1.66e-06) & 3.06e-06& (1.12e-06) & 1.36e-06& (9.26e-07)\ 4210.0 GHz& 71.22 & 617.9 & 6.52e-06& (2.48e-07) & 4.85-07& (1.63e-07) & 2.07e-07& (1.29e-07)\ 4212.0 GHz& 71.17 & 617.6 & 1.29e-06& (4.79e-08) &9.34e-08 & (3.15e-08) &4.00e-08 & (2.50e-08)\ 4592.0 GHz& 65.28& 510.9 & 1.18e-05& (4.79e-08) & 8.66e-07& (2.96e-07) & 3.73e-07& (2.36e-07)\ 4603.0 GHz& 65.13 & 512.1 & 1.79e-06& (6.46e-08) & 1.25e-07& (4.27e-08) & 5.37e-08& (3.40e-08)\ $$\end{center} %\label{t4a} \end{table*} %---------------------------------------------------------H2O ortho v20 \begin{table*} \caption[]{Intensities of ortho-H${}_2$O lines (erg\,cm${}^{-2}$\,s${}^{-1}$\,sr${}^{-1}$) observable with the PACS (top) and HIFI (bottom) instruments on the Herschel Space Observatory, for shocks with velocity $\Vs$~=~20~\kms and the magnetic field strengths listed in Table 1. Results are given for models M1, which include grain-grain processing, and M2 (in parentheses), which neglect grain-grain processing. The preshock density is $n{}_{\rm H}$ = 10${}^5$ cm${}^{-3}$.} \begin{center}$$ [lllllllll]{} Transition & $\lambda$ ($\mu$m) & $ E_{\rm up}$ (K)& v20b1, M1 & (v20b1, M2) & v20b1.5, M1 & (v20b1.5, M2) & v20b2, M1 & (v20b2, M2)\ 5500.1 GHz& 54.506 & 732.1 &1.31e-06 & (2.25e-06) & 1.56e-06 & (2.12e-06) & 1.38e-06 & (1.44e-06)\ 5437.8 GHz& 55.131 & 1274. & 5.49e-06 & (8.49e-06) & 6.44e-06 & (7.07e-06) & 5.49e-06 & (4.75e-06)\ 5276.5 GHz& 56.816 & 1324 & 1.97e-05 & (3.03e-05) &2.30e-05 & (2.43e-05) & 1.95e-05 & (1.64e-05)\ 5107.3 GHz& 58.699 & 550.4 & 1.10e-04 & (1.98e-04) & 1.29e-04 & (1.94e-04) & 1.16e-04 & (1.33e-04)\ 4764.0 GHz& 62.928 & 1553 & 2.74e-07 & (4.08e-07) & 3.20e-07 & (3.36e-07) & 2.67e-07 & (2.24e-07)\ 4734.3 GHz& 63.323 & 1071 & 3.62e-05 & (5.86e-05) & 4.26e-05 & (4.87e-05) & 3.72e-05 & (3.32e-05)\ 4600.4 GHz& 65.166 & 795.5 & 2.57e-05 & (4.19e-05) & 3.03e-05 & (3.89e-05) & 2.61e-05 & (2.61e-05)\ 4535.9 GHz& 66.092 & 1013 & 9.11e-06 & (1.44e-05) & 1.07e-05 & (1.29e-05) & 9.17e-06 & (8.64e-06)\ 4512.4 GHz& 66.437 & 410.7 & 1.54e-04 & (2.99e-04) & 1.76e-04 & (2.47e-04) & 1.74e-04 & (1.85e-04)\ 4456.6 GHz& 67.268 & 410.7 & 7.77e-06 & (1.24e-05) & 1.08e-05 & (2.21e-05) & 6.45e-06 & (1.24e-05)\ 4240.2 GHz& 70.702 & 1274 & 7.40e-07 & (1.14e-06) & 8.68e-07 & (9.53e-07) & 7.39e-07 & (6.40e-07)\ 4166.9 GHz& 71.946 & 843.5 & 6.01e-05 & (9.94e-05) & 7.09e-05 & (8.79e-05) & 6.23e-05 & (5.96e-05)\ 4000.2 GHz& 74.944 & 1126 & 2.34e-06 & (3.70e-06) & 2.75e-06 & (3.13e-06) & 2.36e-06 & (2.11e-06)\ 3977.0 GHz& 75.380 & 305.3 & 2.90e-04 & (4.45e-04) & 3.37e-04 & (4.79e-04) & 2.63e-04 & (3.07e-04)\ 3971.0 GHz& 75.495 & 1806& 3.32e-07 & (4.91e-07) & 3.83e-07 & (3.59e-07) & 3.20e-07 & (2.45e-07)\ 3807.3 GHz& 78.742 & 432.2& 1.30e-04 & (2.33e-04) & 1.53e-04 & (2.21e-04) & 1.36e-04 & (1.53e-04)\ 3654.6 GHz& 82.031 & 643.5& 1.07e-04 & (1.81e-04) & 1.26e-04 & (1.70e-04) & 1.11e-04 & (1.15e-04)\ 3536.7 GHz& 84.766 & 1013& 1.61e-06 & (2.56e-06) & 1.90e-06 & (2.29e-06) & 1.62e-06 & (1.53e-06)\ 3167.6 GHz& 94.643 & 795.5& 3.28e-06 & (5.34e-06) &3.87e-06 & (4.97e-06) & 3.34e-06 & (3.33e-06)\ 3165.5 GHz& 94.704 & 702.3& 1.26e-06 & (2.24e-06) & 1.51e-06 & (2.23e-06) & 1.34e-06 & (1.50e-06)\ 3013.2 GHz& 99.492 & 468.1& 1.97e-04 & (3.45e-04) & 2.33e-04 & (3.41e-04) & 2.04e-04 & (2.31e-04)\ 2970.8 GHz& 100.91 & 574.7& 2.80e-05 & (4.71e-05) & 3.32e-05 & (4.70e-05) & 2.84e-05 & (3.13e-05)\ 2774.0 GHz& 108.07 & 194.1& 6.17e-04 & (1.30e-03) & 7.25e-04 & (1.52e-03) & 6.04e-04 & (1.00e-03)\ 2664.6 GHz& 112.51 & 1340& 3.64e-07 & (5.72e-07) & 4.24e-07 & (4.42e-07) & 3.66e-07 & (3.04e-07)\ 2640.5 GHz& 113.54 & 323.5& 4.32e-04 & (7.79e-04) & 5.14e-04 & (8.34e-04) & 4.39e-04 & (5.55e-04)\ 2462.9 GHz& 121.72 & 550.4& 5.88e-06 & (9.46e-06) & 7.77e-06 & (1.16e-05) & 5.79e-06 & (7.11e-06)\ 2344.3 GHz& 127.88 & 1126& 7.24e-07 & (1.14e-06) & 8.51e-07 & (9.68e-07) & 7.30e-07 & (6.52e-07)\ 2264.1 GHz& 132.41 & 432.2& 2.81e-05 & (4.18e-05) & 3.53e-05 & (6.21e-05) & 2.45e-05 & (3.65e-05)\ 2221.8 GHz& 134.93 & 574.7& 1.01e-05 & (1.70e-05) & 1.20e-05 & (1.70e-05) & 1.03e-05 & (1.13e-05)\ 2196.3 GHz& 136.49 & 410.7& 2.79e-05 & (4.79e-05) & 3.62e-05 & (8.13e-05) & 2.36e-05 & (4.71e-05)\ 1918.5 GHz& 156.26 & 642.4& 6.92e-07 & (1.17e-06) & 8.33e-07 & (1.13e-06) & 7.12e-07 & (7.49e-07)\ 1867.7 GHz& 160.51 & 732.1& 9.50e-07 & (1.63e-06) & 1.13e-06 & (1.54e-06) & 9.99e-07 & (1.04e-06)\ 1716.8 GHz& 174.62 & 196.8& 6.14e-04 & (1.34e-03) & 7.21e-04 & (1.31e-03) & 6.77e-04 & (9.16e-04)\ 1669.9 GHz& 179.53 & 114.4& 9.57e-04 & (2.22e-03) & 1.09e-03 & (2.06e-03) & 1.05e-03 & (1.49e-03)\ 1661.0 GHz& 180.49 & 194.1& 1.93e-04 & (4.63e-04) & 2.30e-04 & (5.06e-04) & 2.03e-04 & (3.41e-04)\ 1162.9 GHz& 257.79 & 305.3& 5.65e-05 & (1.39e-04) & 7.12e-05 & (1.64e-04) & 6.68e-05 & (1.11e-04)\ 1153.1 GHz& 259.98 & 249.4& 1.22e-04 & (2.54e-04) & 1.49e-04 & (3.43e-04) & 1.12e-04 & (2.17e-04)\ 1097.4 GHz& 273.19 & 249.4& 8.72e-05 & (1.96e-04) & 1.04e-04 & (1.71e-04) & 1.07e-04 & (1.25e-04)\ 556.94 GHz& 538.28 & 61.0& 5.61e-05 & (1.32e-04) & 6.01e-05 & (1.21e-04) & 5.73e-05 & (8.89e-05)\ $$\end{center} \label{tb3} \end{table*} %--------------------------------ortho H2O v30 \begin{table*} \caption[]{Intensities of ortho-H${}_2$O lines (erg\,cm${}^{-2}$\,s${}^{-1}$\,sr${}^{-1}$) observable with the PACS (top) and HIFI (bottom) instruments on the Herschel Space Observatory, for shocks with velocity $\Vs$~=~30~\kms and the magnetic field strengths listed in Table 1. Results are given for models M1, which include grain-grain processing, and M2 (in parentheses), which neglect grain-grain processing. The preshock density is $n{}_{\rm H}$ = 10${}^5$ cm${}^{-3}$. } \begin{center}$$ [lllllllll]{} Transition & $\lambda$ ($\mu$m) & $ E_{\rm up}$ (K)& v30b1.5, M1& (v30b1.5, M2) & v30b2, M1& (v30b2, M2) & v30b2.5, M1 & (v30b2.5, M2)\ 5500.1 GHz& 54.506 & 732.1 & 2.16e-06 & (5.36e-06) & 2.80e-06 & (6.34e-06) & 3.60e-06 & (6.74e-06)\ 5437.8 GHz& 55.131& 1274 & 8.87e-06 & (2.26e-05) & 1.21e-05 & (2.60e-05) & 1.60e-05 & (2.68e-05)\ 5276.5 GHz& 56.816 & 1324 & 3.13e-05 & (8.00e-05) & 4.30e-05 & (8.99e-05) & 5.69e-05 & (9.14e-05)\ 5107.3 GHz& 58.699 & 550.4 & 1.75e-04 & (4.39e-04) & 2.13e-04 & (4.97e-04) & 2.72e-04 & (5.15e-04)\ 4764.0 GHz& 62.928 & 1553 & 4.44e-07 & (1.13e-06) & 6.15e-07 & (1.32e-06) & 8.16e-07 & (1.36e-06)\ 4734.3 GHz& 63.323 & 1071 & 5.78e-05 & (1.47e-04) & 7.78e-05 & (1.66e-04) & 1.02e-04 & (1.70e-04)\ 4600.4 GHz& 65.166 & 795.5 & 4.28e-05 & (1.05e-04) & 5.60e-05 & (1.25e-04) & 7.23e-05 & (1.33e-04)\ 4535.9 GHz& 66.092 & 1013 & 1.50e-05 & (3.76e-05) & 2.00e-05 & (4.43e-05) & 2.62e-05 & (4.65e-05)\ 4512.4 GHz& 66.437 & 410.7 & 2.35e-04 & (5.69e-04) & 2.82e-04 & (5.97e-04) & 3.60e-04 & (5.97e-04)\ 4456.6 GHz& 67.268 & 410.7 & 1.89e-05 & (3.95e-05) & 2.66e-05 & (6.85e-05) & 3.09e-05 & (8.70e-05)\ 4240.2 GHz& 70.702 & 1274 & 1.20e-06 & (3.04e-06) & 1.64e-06 & (3.51e-06) & 2.16e-06 & (3.62e-06)\ 4166.9 GHz& 71.946 & 843.5 & 9.74e-05 & (2.45e-04) & 1.29e-04 & (2.83e-04) & 1.68e-04 & (2.95e-04)\ 4000.2 GHz& 74.944 & 1126& 3.76e-06 & (9.59e-06) & 5.09e-06 & (1.10e-05) & 6.71e-06 & (1.13e-05)\ 3977.0 GHz& 75.380 & 305.3 & 5.48e-04 & (1.11e-03) & 6.56e-04 & (1.35e-03) & 7.45e-04 & (1.49e-03)\ 3971.0 GHz& 75.495 & 1806 & 5.24e-07 & (1.32e-06) & 7.29e-07 & (1.42e-06) & 9.63e-07 & (1.41e-06)\ 3807.3 GHz& 78.742 & 432.2 & 2.22e-04 & (5.14e-04) & 2.80e-04 & (6.19e-04) & 3.49e-04 & (6.68e-04)\ 3654.6 GHz& 82.031 & 643.5 & 1.76e-04 & (4.35e-04) & 2.27e-04 & (5.11e-04) & 2.92e-04 & (5.41e-04)\ 3536.7 GHz& 84.766 & 1013 & 2.65e-06 & (6.66e-06) & 3.55e-06 & (7.86e-06) & 4.64e-06 & (8.23e-06)\ 3167.6 GHz& 94.643 & 795.5 & 5.47e-06 & (1.35e-05) & 7.15e-06 & (1.60e-05) & 9.23e-06 & (1.69e-05)\ 3165.5 GHz& 94.704 & 702.3 & 2.11e-06 & (5.21e-06) & 2.69e-06 & (6.29e-06) & 3.43e-06 & (6.79e-06)\ 3013.2 GHz& 99.492 & 468.1 & 3.31e-04 & (7.96e-04) & 4.16e-04 & (9.46e-04) & 5.23e-04 & (1.01e-03)\ 2970.8 GHz& 100.91 & 574.7 & 4.87e-05 & (1.14e-04) & 6.22e-05 & (1.40e-04) & 7.79e-05 & (1.52e-04)\ 2774.0 GHz& 108.07 & 194.1 & 1.01e-03 & (2.62e-03) & 1.15e-03 & (3.14e-03) & 1.39e-03 & (3.42e-03)\ 2664.6 GHz& 112.51 & 1340 & 5.73e-07 & (1.46e-06) & 7.81e-07 & (1.58e-06) & 1.03e-06 & (1.58e-06)\ 2640.5 GHz& 113.54 & 323.5 & 7.37e-04 & (1.76e-03) & 8.96e-04 & (2.12e-03) & 1.11e-03 & (2.29e-03)\ 2462.9 GHz& 121.72 & 550.4 & 1.30e-05 & (2.68e-05) & 1.89e-05 & (4.43e-05) & 2.29e-05 & (5.58e-05)\ 2344.3 GHz& 127.88 & 1126 & 1.16e-06 & (2.96e-06) & 1.57e-06 & (3.39e-06) & 2.07e-06 & (3.49e-06)\ 2264.1 GHz& 132.41 & 432.2 & 5.71e-05 & (1.27e-04) & 7.08e-05 & (1.80e-04) & 8.42e-05 & (2.07e-04)\ 2221.8 GHz& 134.93 & 574.7 & 1.76e-05 & (4.13e-05) & 2.24e-05 & (5.05e-05) & 2.81e-05 & (5.47e-05)\ 2196.3 GHz& 136.49 & 410.7 & 5.57e-05 & (1.36e-04) & 6.77e-05 & (1.98e-04) & 7.99e-05 & (2.33e-04)\ 1918.5 GHz& 156.26 & 642.4 & 1.23e-06 & (2.87e-06) & 1.66e-06 & (3.67e-06) & 2.08e-06 & (4.13e-06)\ 1867.7 GHz& 160.51 & 732.1 & 1.57e-06 & (3.89e-06) & 2.03e-06 & (4.60e-06) & 2.62e-06 & (4.90e-06)\ 1716.8 GHz& 174.62 & 196.8 & 9.24e-04 & (2.48e-03) & 1.09e-03 & (2.70e-03) & 1.37e-03 & (2.79e-03)\ 1669.9 GHz& 179.53 & 114.4 & 1.38e-03 & (3.73e-03) & 1.58e-03 & (3.81e-03) & 1.96e-03 & (3.89e-03)\ 1661.0 GHz& 180.49 & 194.1 & 2.93e-04 & (8.45e-04) & 3.41e-04 & (9.58e-04) & 4.19e-04 & (1.03e-03)\ 1162.9 GHz& 257.79 & 305.3 & 8.11e-05 & (2.61e-04) & 9.72e-05 & (3.20e-04) & 1.35e-04 & (3.48e-04)\ 1153.1 GHz& 259.98 & 249.4 & 2.09e-04 & (5.63e-04) & 2.44e-04 & (7.36e-04) & 2.97e-04 & (8.35e-04)\ 1097.4 GHz& 273.19 & 249.4 & 1.27e-04 & (3.43e-04) & 1.52e-04 & (3.55e-04) & 2.00e-04 & (3.50e-04)\ 556.94 GHz& 538.28 & 61.0 & 7.98e-05 & (2.06e-04) & 8.45e-05 & (2.00e-04) & 9.89e-05 & (2.04e-04)\ $$\end{center} \label{tb4} \end{table*} %-------------------------------orthoH2O v40 \begin{table*} \caption[]{Intensities of ortho-H${}_2$O lines (erg\,cm${}^{-2}$\,s${}^{-1}$\,sr${}^{-1}$) observable with the PACS (top) and HIFI (bottom) instruments on the Herschel Space Observatory, for shocks with velocity $\Vs$~=~40~\kms and the magnetic field strengths listed in Table 1. Results are given for models M1, which include grain-grain processing, and M2 (in parentheses), which neglect grain-grain processing. The preshock density is $n{}_{\rm H}$ = 10${}^5$ cm${}^{-3}$.} \begin{center}$$ [lllllllll]{} Transition & $\lambda$ ($\mu$m) & $E_{\rm up}$ (K)& v40b2, M1 & (v40b2, M2) & v40b2.5, M1 & (v40b2.5, M2) & v40b3, M1 & (v40b3, M2)\ 5500.1 GHz& 54.506 & 732.1 & 5.58e-06 & (9.15e-06) & 3.78e-06 & (1.08e-05) & 4.89e-06 & (1.19e-05)\ 5437.8 GHz& 55.131& 1274 & 1.47e-05 & (3.87e-05) & 1.53e-05 & (4.56e-05) & 2.14e-05 & (4.94e-05)\ 5276.5 GHz& 56.816 & 1324 & 4.82e-05 & (1.35e-04) & 5.37e-05 & (1.56e-04) & 7.51e-05 & (1.67e-04)\ 5107.3 GHz& 58.699 & 550.4 & 3.39e-04 & (7.11e-04) & 2.75e-04 & (7.70e-04) & 3.44e-04 & (8.04e-04)\ 4764.0 GHz& 62.928 & 1553 & 7.07e-07 & (1.96e-06) & 7.73e-07 & (2.35e-06) & 1.09e-06 & (2.55e-06)\ 4734.3 GHz& 63.323 & 1071 & 9.55e-05 & (2.46e-04) & 9.84e-05 & (2.83e-04) & 1.35e-04 & (3.03e-04)\ 4600.4 GHz& 65.166 & 795.5 & 9.22e-05 & (1.83e-04) & 7.40e-05 & (2.17e-04) & 9.76e-05 & (2.37e-04)\ 4535.9 GHz& 66.092 & 1013 & 2.84e-05 & (6.51e-05) & 2.59e-05 & (7.73e-05) & 3.51e-05 & (8.41e-05)\ 4512.4 GHz& 66.437 & 410.7 & 5.35e-04 & (8.84e-04) & 3.80e-04 & (9.42e-04) & 4.63e-04 & (9.85e-04)\ 4456.6 GHz& 67.268 & 410.7 & 1.27e-04 & (8.96e-05) & 4.90e-05 & (1.41e-04) & 5.51e-05 & (1.74e-04)\ 4240.2 GHz& 70.702 & 1274 & 1.98e-06 & (5.21e-06) & 2.07e-06 & (6.15e-06) & 2.88e-06 & (6.65e-06)\ 4166.9 GHz& 71.946 & 843.5 & 1.83e-04 & (4.15e-04) & 1.66e-04 & (4.81e-04) & 2.22e-04 & (5.19e-04)\ 4000.2 GHz& 74.944 & 1126& 6.26e-06 & (1.62e-05) & 6.44e-06 & (1.89e-05) & 8.87e-06 & (2.03e-05)\ 3977.0 GHz& 75.380 & 305.3 & 1.64e-03 & (2.05e-03) & 9.85e-04 & (2.40e-03) & 1.10e-03 & (2.62e-03)\ 3971.0 GHz& 75.495 & 1806 & 7.12e-07 & (2.20e-06) & 8.98e-07 & (2.48e-06) & 1.27e-06 & (2.62e-06)\ 3807.3 GHz& 78.742 & 432.2 & 6.95e-04 & (8.97e-04) & 4.06e-04 & (1.09e-03) & 4.95e-04 & (1.21e-03)\ 3654.6 GHz& 82.031 & 643.5 & 3.82e-04 & (7.41e-04) & 3.01e-04 & (8.65e-04) & 3.90e-04 & (9.37e-04)\ 3536.7 GHz& 84.766 & 1013 & 5.03e-06 & (1.15e-05) & 4.58e-06 & (1.37e-05) & 6.22e-06 & (1.49e-05)\ 3167.6 GHz& 94.643 & 795.5 & 1.18e-05 & (2.34e-05) & 9.46e-06 & (2.78e-05) & 1.25e-05 & (3.02e-05)\ 3165.5 GHz& 94.704 & 702.3 & 5.90e-06 & (8.94e-06) & 3.70e-06 & (1.07e-05) & 4.67e-06 & (1.18e-05)\ 3013.2 GHz& 99.492 & 468.1 & 8.28e-04 & (1.36e-03) & 5.70e-04 & (1.59e-03) & 7.08e-04 & (1.73e-03)\ 2970.8 GHz& 100.91 & 574.7 & 1.47e-04 & (2.03e-04) & 8.82e-05 & (2.48e-04) & 1.10e-04 & (2.76e-04)\ 2774.0 GHz& 108.07 & 194.1 & 2.17e-03 & (4.21e-03) & 1.56e-03 & (4.71e-03) & 1.79e-03 & (5.06e-03)\ 2664.6 GHz& 112.51 & 1340 & 8.39e-07 & (2.40e-06) & 9.72e-07 & (2.69e-06) & 1.35e-06 & (2.85e-06)\ 2640.5 GHz& 113.54 & 323.5 & 1.78e-03 & (3.00e-03) & 1.24e-03 & (3.45e-03) & 1.48e-03 & (3.73e-03)\ 2462.9 GHz& 121.72 & 550.4 & 7.19e-05 & (6.04e-05) & 3.27e-05 & (9.70e-05) & 3.95e-05 & (1.21e-04)\ 2344.3 GHz& 127.88 & 1126 & 1.94e-06 & (5.02e-06) & 1.99e-06 & (5.84e-06) & 2.74e-06 & (6.28e-06)\ 2264.1 GHz& 132.41 & 432.2 & 1.94e-04 & (2.46e-04) & 1.06e-04 & (3.13e-04) & 1.23e-04 & (3.53e-04)\ 2221.8 GHz& 134.93 & 574.7 & 5.05e-05 & (7.31e-05) & 3.16e-05 & (8.86e-05) & 3.94e-05 & (9.80e-05)\ 2196.3 GHz& 136.49 & 410.7 & 1.60e-04 & (2.56e-04) & 9.58e-05 & (3.20e-04) & 1.10e-04 & (3.57e-04)\ 1918.5 GHz& 156.26 & 642.4 & 5.16e-06 & (5.30e-06) & 2.55e-06 & (7.21e-06) & 3.22e-06 & (8.64e-06)\ 1867.7 GHz& 160.51 & 732.1 & 4.07e-06 & (6.64e-06) & 2.75e-06 & (7.87e-06) & 3.55e-06 & (8.66e-06)\ 1716.8 GHz& 174.62 & 196.8 & 1.62e-03 & (3.74e-03) & 1.39e-03 & (3.94e-03) & 1.69e-03 & (4.10e-03)\ 1669.9 GHz& 179.53 & 114.4 & 2.15e-03 & (5.35e-03) & 1.98e-03 & (5.35e-03) & 2.38e-03 & (5.47e-03)\ 1661.0 GHz& 180.49 & 194.1 & 5.39e-04 & (1.28e-03) & 4.43e-04 & (1.37e-03) & 5.20e-04 & (1.45e-03)\ 1162.9 GHz& 257.79 & 305.3 & 1.37e-04 & (3.89e-04) & 1.20e-04 & (4.39e-04) & 1.52e-04 & (4.74e-04)\ 1153.1 GHz& 259.98 & 249.4 & 5.07e-04 & (9.47e-04) & 3.38e-04 & (1.12e-03) & 3.89e-04 & (1.24e-03)\ 1097.4 GHz& 273.19 & 249.4 & 2.20e-04 & (5.04e-04) & 1.97e-04 & (5.22e-04) & 2.42e-04 & (5.29e-04)\ 556.94 GHz& 538.28 & 61.0 & 1.14e-04 & (2.84e-04) & 1.06e-04 & (2.71e-04) & 1.20e-04 & (2.74e-04)\ $$\end{center} \label{tb5} \end{table*} %--------------------paraH2O v20 \begin{table*} \caption[]{Intensities of para-H${}_2$O lines (erg\,cm${}^{-2}$\,s${}^{-1}$\,sr${}^{-1}$) observable with the PACS (top) and HIFI (bottom) instruments on the Herschel Space Observatory, for shocks with velocity $\Vs$~=~20~\kms and the magnetic field strengths listed in Table 1. Results are given for models M1, which include grain-grain processing, and M2 (in parentheses), which neglect grain-grain processing. The preshock density is $n{}_{\rm H}$ = 10${}^5$ cm${}^{-3}$.} \begin{center}$$ [lllllllll]{} Transition & $\lambda$ ($\mu$m) & $ E_{\rm up}$ (K)& v20b1, M1 & (v20b1, M2) & v20b1.5, M1 & (v20b1.5, M2) & v20b2, M1 & (v20b2, M2)\ 5322.5 GHz & 56.325 & 552.3 & 2.46e-05 & (4.10e-05) & 2.91e-05 & (4.81e-05) & 2.37e-05 & (3.10e-05)\ 5280.7 GHz & 56.771 & 1324 & 4.69e-06 & (7.15e-06) & 5.49e-06 & (5.91e-06) & 4.63e-06 & (3.96e-06)\ 5201.4 GHz & 57.636 & 454.3 & 6.05e-05 & (1.10e-04) & 7.08e-05 & (1.04e-04) & 6.41e-05 & (7.16e-05)\ 5194.9 GHz & 57.709 & 1270 & 9.33e-07 & (1.41e-06) & 1.09e-06 & (1.25e-06) & 9.06e-07 & (8.23e-07)\ 4997.6 GHz & 59.987 & 1021 & 1.76e-06 & (2.71e-06) & 2.06e-06 & (2.59e-06) & 1.71e-06 & (1.69e-06)\ 4850.3 GHz & 61.808 & 552.3 & 3.85e-07 & (6.35e-07) & 4.60e-07 & (7.56e-07) & 3.68e-07 & (4.84e-07)\ 4724.0 GHz & 63.457 & 1070 & 8.78e-06 & (1.40e-05) & 1.03e-05 & (1.18e-05) & 8.87e-06 & (8.00e-06)\ 4468.6 GHz & 67.089 & 410.4 & 6.05e-05 & (1.16e-04) & 7.04e-05 & (1.13e-04) & 6.54e-05 & (7.90e-05)\ 4218.4 GHz & 71.067 & 598.8 & 1.54e-05 & (2.56e-05) & 1.82e-05 & (2.67e-05) & 1.54e-05 & (1.76e-05)\ 4190.6 GHz & 71.539 & 843.8 & 1.76e-05 & (2.95e-05) & 2.08e-05 & (2.53e-05) & 1.85e-05 & (1.73e-05)\ 3798.3 GHz & 78.928 & 781.1 & 3.35e-06 & (5.31e-06) & 3.95e-06 & (5.66e-06) & 3.27e-06 & (3.66e-06)\ 3691.3 GHz & 81.215 & 1021 & 2.36e-07 & (3.63e-07) & 2.77e-07 & (3.48e-07) & 2.29e-07 & (2.27e-07)\ 3599.6 GHz & 83.283 & 642.7 & 3.16e-05 & (5.37e-05) & 3.74e-05 & (5.04e-05) & 3.29e-05 & (3.40e-05)\ 3331.5 GHz & 89.988 & 296.8 & 5.73e-05 & (9.59e-05) & 6.85e-05 & (1.03e-04) & 5.74e-05 & (6.87e-05)\ 3182.2 GHz & 94.209 & 877.8 & 2.05e-07 & (3.32e-07) & 2.44e-07 & (3.57e-07) & 2.05e-07 & (2.32e-07)\ 3135.0 GHz & 95.626 & 469.9 & 5.68e-05 & (9.94e-05) & 6.73e-05 & (1.01e-04) & 5.86e-05 & (6.75e-05)\ 2968.7 GHz & 100.98 & 195.9 & 1.86e-04 & (3.60e-04) & 2.14e-04 & (4.13e-04) & 1.68e-04 & (2.66e-04)\ 2884.3 GHz & 103.94 & 781.1 & 7.63e-07 & (1.21e-06) & 9.01e-07 & (1.29e-06) & 7.45e-07 & (8.34e-07)\ 2685.6 GHz & 111.63 & 598.8 & 1.77e-06 & (2.94e-06) & 2.09e-06 & (3.08e-06) & 1.77e-06 & (2.02e-06)\ 2631.0 GHz & 113.95 & 725.1 & 1.35e-06 & (2.40e-06) & 1.60e-06 & (1.98e-06) & 1.47e-06 & (1.40e-06)\ 2391.6 GHz & 125.35 & 319.5 & 1.12e-04 & (2.04e-04) & 1.33e-04 & (2.18e-04) & 1.14e-04 & (1.46e-04)\ 2365.9 GHz & 126.71 & 410.4 & 3.25e-06 & (5.59e-06) & 4.66e-06 & (8.41e-06) & 3.33e-06 & (4.95e-06)\ 2164.1 GHz & 138.53 & 204.7 & 2.83e-04 & (5.69e-04) & 3.36e-04 & (6.40e-04) & 2.82e-04 & (4.21e-04)\ 2074.4 GHz & 144.52 & 396.4 & 1.10e-05 & (1.84e-05) & 1.31e-05 & (2.17e-05) & 1.04e-05 & (1.39e-05)\ 2040.5 GHz & 146.92 & 552.3 & 5.41e-07 & (8.92e-07) & 6.47e-07 & (1.06e-06) & 5.18e-07 & (6.80e-07)\ 1919.4 GHz & 156.19 & 296.8 & 2.97e-05 & (5.47e-05) & 3.59e-05 & (7.92e-05) & 2.50e-05 & (4.86e-05)\ 1794.8 GHz & 167.03 & 867.3 & 1.81e-07 & (3.03e-07) & 2.16e-07 & (3.08e-07) & 1.87e-07 & (2.04e-07)\ 1602.2 GHz & 187.11 & 396.4 & 9.43e-06 & (1.59e-05) & 1.12e-05 & (1.86e-05) & 8.99e-06 & (1.20e-05)\ 1228.8 GHz & 243.97 & 195.9 & 3.48e-05 & (9.50e-05) & 4.45e-05 & (9.36e-05) & 4.62e-05 & (6.63e-05)\ 1113.3 GHz & 269.27 & 53.4 & 2.64e-04 & (6.40e-04) & 2.97e-04 & (5.82e-04) & 2.89e-04 & (4.25e-04)\ 987.93 GHz & 303.45 & 100.8 & 2.10e-04 & (5.15e-04) & 2.42e-04 & (4.97e-04) & 2.33e-04 & (3.55e-04)\ 752.03 GHz & 398.64 & 136.9 & 6.59e-05 & (1.75e-04) & 7.72e-05 & (1.71e-04) & 7.54e-05 & (1.21e-04)\ $$\end{center} \label{tb6} \end{table*} %----------------------------------------- %------------------------------------------------------------------------------------ table v30 %---------------------------- %-------------------------------------------------------------------- \begin{table*} \caption[]{Intensities of para-H${}_2$O lines (erg\,cm${}^{-2}$\,s${}^{-1}$\,sr${}^{-1}$) observable with the PACS (top) and HIFI (bottom) instruments on the Herschel Space Observatory, for shocks with velocity $\Vs$~=~30~\kms and the magnetic field strengths listed in Table 1. Results are given for models M1, which include grain-grain processing, and M2 (in parentheses), which neglect grain-grain processing. The preshock density is $n{}_{\rm H}$ = 10${}^5$ cm${}^{-3}$.} \begin{center}$$ [lllllllll]{} Transition & $\lambda$ ($\mu$m)& $ E_{\rm up}$ (K) & v30b1.5, M1& (v30b1.5, M2) & v30b2, M1& (v30b2, M2) & v30b2.5, M1 & (v30b2.5, M2)\ 5322.5 GHz & 56.325 & 552.30 & 4.33e-05 & (1.03e-04) & 5.31e-05 & (1.34e-04) & 6.64e-05 & (1.48e-04)\ 5280.7 GHz & 56.771 & 1324.0 & 7.51e-06 & (1.93e-05) & 1.03e-05 & (2.21e-05) & 1.38e-05 & (2.26e-05)\ 5201.4 GHz & 57.636 & 454.30 & 9.73e-05 & (2.39e-04) & 1.19e-04 & (2.65e-04) & 1.50e-04 & (2.74e-04)\ 5194.9 GHz & 57.709 & 1270.3 & 1.54e-06 & (3.88e-06) & 2.08e-06 & (4.62e-06) & 2.74e-06 & (4.84e-06)\ 4997.6 GHz & 59.987 & 1021.0 & 2.97e-06 & (7.39e-06) & 3.94e-06 & (9.09e-06) & 5.15e-06 & (9.70e-06)\ 4850.3 GHz & 61.808 & 552.30 & 7.05e-07 & (1.62e-06) & 8.95e-07 & (2.20e-06) & 1.11e-06 & (2.51e-06)\ 4724.0 GHz & 63.457 & 1070.0 & 1.41e-05 & (3.60e-05) & 1.90e-05 & (4.10e-05) & 2.51e-05 & (4.21e-05)\ 4468.6 GHz & 67.089 & 410.40 & 9.46e-05 & (2.37e-04) & 1.12e-04 & (2.61e-04) & 1.42e-04 & (2.70e-04)\ 4218.4 GHz & 71.067 & 598.80 & 2.62e-05 & (6.37e-05) & 3.32e-05 & (7.92e-05) & 4.23e-05 & (8.59e-05)\ 4190.6 GHz & 71.539 & 843.80 & 2.82e-05 & (7.16e-05) & 3.73e-05 & (8.08e-05) & 4.87e-05 & (8.35e-05)\ 3798.3 GHz & 78.928 & 781.10 & 5.73e-06 & (1.41e-05) & 7.39e-06 & (1.81e-05) & 9.56e-06 & (1.97e-05)\ 3691.3 GHz & 81.215 & 1021.0 & 3.98e-07 & (9.91e-07) & 5.29e-07 & (1.22e-06) & 6.91e-07 & (1.30e-06)\ 3599.6 GHz & 83.283 & 642.70 & 5.17e-05 & (1.29e-04) & 6.68e-05 & (1.51e-04) & 8.63e-05 & (1.60e-04)\ 3331.5 GHz & 89.988 & 296.80 & 1.09e-04 & (2.28e-04) & 1.33e-04 & (2.99e-04) & 1.60e-04 & (3.30e-04)\ 3182.2 GHz & 94.209 & 877.80 & 3.47e-07 & (8.74e-07) & 4.48e-07 & (1.12e-06) & 5.89e-07 & (1.22e-06)\ 3135.0 GHz & 95.626 & 469.90 & 9.47e-05 & (2.32e-04) & 1.19e-04 & (2.78e-04) & 1.51e-04 & (2.99e-04)\ 2968.7 GHz & 100.98 & 195.90 & 3.21e-04 & (7.66e-04) & 3.69e-04 & (8.96e-04) & 4.19e-04 & (9.80e-04)\ 2884.3 GHz & 103.94 & 781.10 & 1.31e-06 & (3.23e-06) & 1.69e-06 & (4.12e-06) & 2.18e-06 & (4.49e-06)\ 2685.6 GHz & 111.63 & 598.80 & 3.02e-06 & (7.33e-06) & 3.83e-06 & (9.14e-06) & 4.87e-06 & (9.91e-06)\ 2631.0 GHz & 113.95 & 725.10 & 2.11e-06 & (5.34e-06) & 2.74e-06 & (5.74e-06) & 3.55e-06 & (5.85e-06)\ 2391.6 GHz & 125.35 & 319.50 & 1.92e-04 & (4.57e-04) & 2.34e-04 & (5.55e-04) & 2.89e-04 & (6.03e-04)\ 2365.9 GHz & 126.71 & 410.40 & 7.21e-06 & (1.60e-05) & 1.03e-05 & (2.66e-05) & 1.25e-05 & (3.32e-05)\ 2164.1 GHz & 138.53 & 204.70 & 4.71e-04 & (1.19e-03) & 5.54e-04 & (1.42e-03) & 6.74e-04 & (1.54e-03)\ 2074.4 GHz & 144.52 & 396.40 & 2.05e-05 & (4.55e-05) & 2.57e-05 & (6.10e-05) & 3.10e-05 & (6.94e-05)\ 2040.5 GHz & 146.92 & 552.30 & 9.90e-07 & (2.28e-06) & 1.26e-06 & (3.10e-06) & 1.56e-06 & (3.54e-06)\ 1919.4 GHz & 156.19 & 296.80 & 5.46e-05 & (1.33e-04) & 6.41e-05 & (1.80e-04) & 7.51e-05 & (2.08e-04)\ 1794.8 GHz & 167.03 & 867.30 & 3.01e-07 & (7.63e-07) & 3.89e-07 & (9.46e-07) & 5.08e-07 & (1.02e-06)\ 1602.2 GHz & 187.11 & 396.40 & 1.72e-05 & (3.89e-05) & 2.11e-05 & (5.08e-05) & 2.56e-05 & (5.67e-05)\ 1228.8 GHz & 243.97 & 195.90 & 4.90e-05 & (1.61e-04) & 5.99e-05 & (1.80e-04) & 8.46e-05 & (1.84e-04)\ 1113.3 GHz & 269.27 & 53.400 & 3.72e-04 & (1.03e-03) & 4.20e-04 & (1.01e-03) & 5.13e-04 & (1.02e-03)\ 987.93 GHz & 303.45 & 100.80 & 2.97e-04 & (8.56e-04) & 3.40e-04 & (8.95e-04) & 4.31e-04 & (9.18e-04)\ 752.03 GHz & 398.64 & 136.90 & 9.10e-05 & (2.81e-04) & 1.03e-04 & (2.98e-04) & 1.35e-04 & (3.04e-04)\ $$\end{center} \label{tb7} \end{table*} %____________________________ %__________________________________________________________________ %------------------------------------------------------------------------------------ table v40 %------------------------------------------------ %-------------------------------------- \begin{table*} \caption[]{ Intensities of para-H${}_2$O lines (erg\,cm${}^{-2}$\,s${}^{-1}$\,sr${}^{-1}$) observable with the PACS (top) and HIFI (bottom) instruments on the Herschel Space Observatory, for shocks with velocity $\Vs$~=~40~\kms and the magnetic field strengths listed in Table 1. Results are given for models M1, which include grain-grain processing, and M2 (in parentheses), which neglect grain-grain processing. The preshock density is $n{}_{\rm H}$ = 10${}^5$ cm${}^{-3}$.} \begin{center}$$ [lllllllll]{} Transition & $\lambda$ ($\mu$m)& $ E_{\rm up}$ (K) & v40b2, M1 & (v40b2, M2) & v40b2.5, M1 & (v40b2.5, M2) & v40b3, M1 & (v40b3, M2)\ 5322.5 GHz & 56.325 & 552.30 & 1.19e-04 & (1.86e-04) & 7.42e-05 & (2.27e-04) & 9.05e-05 & (2.50e-04)\ 5280.7 GHz & 56.771 & 1324.0 & 1.25e-05 & (3.28e-05) & 1.30e-05 & (3.84e-05) & 1.82e-05 & (4.14e-05)\ 5201.4 GHz & 57.636 & 454.30 & 1.94e-04 & (3.86e-04) & 1.55e-04 & (4.16e-04) & 1.93e-04 & (4.33e-04)\ 5194.9 GHz & 57.709 & 1270.3 & 2.91e-06 & (6.81e-06) & 2.67e-06 & (8.18e-06) & 3.68e-06 & (8.93e-06)\ 4997.6 GHz & 59.987 & 1021.0 & 6.42e-06 & (1.32e-05) & 5.17e-06 & (1.61e-05) & 6.97e-06 & (1.77e-05)\ 4850.3 GHz & 61.808 & 552.30 & 3.27e-06 & (3.06e-06) & 1.40e-06 & (4.07e-06) & 1.68e-06 & (4.80e-06)\ 4724.0 GHz & 63.457 & 1070.0 & 2.47e-05 & (6.05e-05) & 2.40e-05 & (7.00e-05) & 3.29e-05 & (7.51e-05)\ 4468.6 GHz & 67.089 & 410.40 & 1.96e-04 & (3.74e-04) & 1.46e-04 & (3.99e-04) & 1.79e-04 & (4.15e-04)\ 4218.4 GHz & 71.067 & 598.80 & 6.55e-05 & (1.12e-04) & 4.51e-05 & (1.35e-04) & 5.73e-05 & (1.49e-04)\ 4190.6 GHz & 71.539 & 843.80 & 5.10e-05 & (1.19e-04) & 4.78e-05 & (1.36e-04) & 6.40e-05 & (1.46e-04)\ 3798.3 GHz & 78.928 & 781.10 & 1.37e-05 & (2.56e-05) & 9.94e-06 & (3.15e-05) & 1.30e-05 & (3.48e-05)\ 3691.3 GHz & 81.215 & 1021.0 & 8.61e-07 & (1.77e-06) & 6.94e-07 & (2.16e-06) & 9.36e-07 & (2.38e-06)\ 3599.6 GHz & 83.283 & 642.70 & 1.07e-04 & (2.19e-04) & 8.77e-05 & (2.54e-04) & 1.14e-04 & (2.75e-04)\ 3331.5 GHz & 89.988 & 296.80 & 4.09e-04 & (4.30e-04) & 2.10e-04 & (5.44e-04) & 2.38e-04 & (6.11e-04)\ 3182.2 GHz & 94.209 & 877.80 & 8.47e-07 & (1.57e-06) & 5.99e-07 & (1.93e-06) & 7.86e-07 & (2.14e-06)\ 3135.0 GHz & 95.626 & 469.90 & 2.28e-04 & (3.96e-04) & 1.61e-04 & (4.63e-04) & 2.02e-04 & (5.04e-04)\ 2968.7 GHz & 100.98 & 195.90 & 7.60e-04 & (1.28e-03) & 5.30e-04 & (1.41e-03) & 5.83e-04 & (1.52e-03)\ 2884.3 GHz & 103.94 & 781.10 & 3.14e-06 & (5.83e-06) & 2.27e-06 & (7.18e-06) & 2.96e-06 & (7.95e-06)\ 2685.6 GHz & 111.63 & 598.80 & 7.73e-06 & (1.30e-05) & 5.23e-06 & (1.57e-05) & 6.63e-06 & (1.72e-05)\ 2631.0 GHz & 113.95 & 725.10 & 3.67e-06 & (8.53e-06) & 3.51e-06 & (9.32e-06) & 4.60e-06 & (9.83e-06)\ 2391.6 GHz & 125.35 & 319.50 & 5.05e-04 & (7.82e-04) & 3.29e-04 & (9.15e-04) & 3.92e-04 & (9.97e-04)\ 2365.9 GHz & 126.71 & 410.40 & 3.46e-05 & (3.41e-05) & 1.72e-05 & (5.11e-05) & 2.00e-05 & (6.20e-05)\ 2164.1 GHz & 138.53 & 204.70 & 1.05e-03 & (1.96e-03) & 7.60e-04 & (2.20e-03) & 8.84e-04 & (2.36e-03)\ 2074.4 GHz & 144.52 & 396.40 & 8.42e-05 & (8.54e-05) & 4.01e-05 & (1.11e-04) & 4.69e-05 & (1.29e-04)\ 2040.5 GHz & 146.92 & 552.30 & 4.64e-06 & (4.30e-06) & 1.96e-06 & (5.72e-06) & 2.37e-06 & (6.76e-06)\ 1919.4 GHz & 156.19 & 296.80 & 1.61e-04 & (2.38e-04) & 9.42e-05 & (2.91e-04) & 1.05e-04 & (3.26e-04)\ 1794.8 GHz & 167.03 & 867.30 & 6.52e-07 & (1.33e-06) & 5.12e-07 & (1.60e-06) & 6.72e-07 & (1.75e-06)\ 1602.2 GHz & 187.11 & 396.40 & 5.10e-05 & (7.11e-05) & 3.06e-05 & (8.73e-05) & 3.60e-05 & (9.67e-05)\ 1228.8 GHz & 243.97 & 195.90 & 8.03e-05 & (2.34e-04) & 7.51e-05 & (2.53e-04) & 9.65e-05 & (2.63e-04)\ 1113.3 GHz & 269.27 & 53.400 & 5.30e-04 & (1.43e-03) & 5.21e-04 & (1.39e-03) & 6.21e-04 & (1.40e-03)\ 987.93 GHz & 303.45 & 100.80 & 4.52e-04 & (1.22e-03) & 4.20e-04 & (1.24e-03) & 5.10e-04 & (1.27e-03)\ 752.03 GHz & 398.64 & 136.90 & 1.37e-04 & (3.95e-04) & 1.26e-04 & (4.05e-04) & 1.54e-04 & (4.15e-04)\ $$ \[tb8\] [^1]: Although electrons and ions are treated as one dynamical fluid, their temperatures are calculated separately. [^2]: Although changes to the re-accretion of mantle species in the multi-fluid treatment give rise to a slightly different temperature profile in the postshock gas. [^3]: This factor refers to the total cross-section of grain cores. The re-accretion of grain mantles in the postshock gas leads to a much larger increase in the total cross-section, by a factor of $\sim$100 (see. Fig. \[FigA1\_shattering\]). However, this large increase is irrelevant to the shock dynamics, as it happens where the momentum transfer between the ionized and the neutral fluids is almost complete ($T\sim 100\,$K). [^4]: We emphasize the effects of SiO vaporization because the fine-structure lines of atomic carbon are optically thin and do not require an LVG treatment. [^5]: The threshold velocity for sputtering of the mantles is $\sim 10 \mu ^{-\frac {1}{2}}$kms${}^{-1}$, where $\mu$ is the reduced mass of the colliding species, relative to atomic hydrogen (@Barlow:1978p17402). [^6]: In order to make a direct comparison with the results of Paper III, it is necessary to disable the LVG treatment of molecular line transfer in our current model. [^7]: The situation was more complicated in Paper III, which dealt with a full grain size distribution. Fragments were allocated to the size bins corresponding to their individual mass, while the corresponding mass was removed from the projectile and target size bins. This procedure does not allow for charge conservation, because small grains carry much more (negative) charge per unit mass than large grains. To compensate for this charge excess, the charge distributions of all grain sizes was shifted infinitesimally at each shattering event to ensure charge conservation.

ArXiv

--- abstract: | We study the charmonium $p \bar{p} \to \psi(3770) \pi^0$ reaction using effective lagrangian approach where the contributions from well established $N^*$ states are considered, and all parameters are fixed in the process of $e^+e^- \to p \bar{p}\pi^0$ at center of mass energy $\sqrt{s} = 3.773$ GeV. The experimental data on the line shape of the mass distribution of the $e^+e^- \to p\bar{p}\pi^0$ can be well reproduced. Based on the studying of $e^+e^- \to p \bar{p}\pi^0$, the total and differential cross sections of the $p \bar{p} \to \psi(3770) \pi^0$ reaction are predicted. At the same time we evaluated also the cross sections of the $p \bar{p} \to \psi(3686) \pi^0$ reaction. It is shown that the contribution of nucleon pole to this reaction is largest close to the reaction threshold. However, the interference between nucleon pole and the other nucleon resonance can still change the angle distributions significantly. Those theoretical results may be test by the future experiments at $\overline{\mbox{P}}$ANDA. author: - 'Hao Xu$^{1,2}$' - 'Ju-Jun Xie$^{1,3,4}$' - 'Xiang Liu$^{1,2,4}$[^1]' title: 'Implication of the observed $e^+e^-\to p\bar{p}\pi^0$ for studying the $p\bar{p}\to \psi(3770)\pi^0$ process' --- introduction {#sec1} ============ As a forthcoming facility in future, the Anti-Proton Annihilations at Darmstadt ($\overline{\mbox{P}}$ANDA) experiment will focus on the production of charmonium, which is govern by nonperturbative effect of quantum chromodynamics (QCD) [@Lutz:2009ff]. Before $\overline{\mbox{P}}$ANDA run, there were pioneering theoretical studies of the charmonium production in the $p\bar{p}$ annihilation processes [@Gaillard:1982zm; @Lundborg:2005am; @Barnes:2006ck; @Barnes:2007ub; @Barnes:2010yb; @Lin:2012ru; @Pire:2013jva; @Wiele:2013vla]. By calculating two hadron-level diagrams introduced by the Born approximation, Gaillard and Maiani firstly studied the differential cross section of the charmonium production plus a soft pion in the $p \bar{p}$ reaction [@Gaillard:1982zm]. In Ref. [@Lundborg:2005am], the cross sections of the chamonium ($\Psi$) production accompanied by a light meson ($m$) from the process of $p\bar{p} \to \Psi + m$ was calculated by combing with the measured partial decay widths of charmonium decay into $p \bar{p} m$. And then, Barnes and Li proposed an initial state light meson emission model for the near threshold associated charmonium production processes $p \bar p \to \pi^0 \Psi$ ($\Psi=\eta_c,\,J/\psi,\,\psi^\prime,\,\chi_{c0},\,\chi_{c1}$), and the total and differential cross sections for these reactions were evaluated [@Barnes:2006ck; @Barnes:2007ub; @Barnes:2010yb]. It is also found that the cross section of $p \bar p\to \pi^0 \Psi$ near threshold may be affected by the Pauli $J/\psi p\bar p$ coupling [@Barnes:2007ub]. Furthermore, Lin, Xu and Liu revisited the issue of the production of charmonium plus a light meson at $\overline{\mbox{P}}$ANDA, where the contribution of form factors (FFs) to these processes are included [@Lin:2012ru]. Recently, Pire [*et al.*]{} studied the associated production of a $J/\psi$ and a pion in antiproton-nucleon annihilation in the framework of QCD collinear factorization [@Pire:2013jva], while in Ref. [@Wiele:2013vla], the exclusive charmonium production process $p\bar p\to \pi^0 J/\psi$ was studied within a nucleon-pole exchange model by including off-shell hadronic FFs and a complete Lorentz structure with a $\bar p p J/\psi$ Pauli strong coupling. The contributions from the intermediate $N^*$ states are also studied in Ref. [@Wiele:2013vla], and it was found that one can not ignore the contributions of the $N^*$ resonances in the $\bar{p} p \to \pi^0 J/\psi$ reaction. The experimental activity on the charmonium decays have run in parallel. These decays are of interest because they can be used to study the associated charomonium production in $p \bar p$ annihilation. In 2014, the BESIII Collaboration reported the analysis of $e^+e^- \to p \bar p \pi^0$ in the vicinity of $\psi (3770)$ [@Ablikim:2014kxa]. In addition to the Born cross section of $e^+e^-\to p \bar{p}\pi^0$, the corresponding $p\pi^0$ and $\bar{p} \pi^0$ invariant mass distributions of $e^+e^- \to p\bar{p}\pi^0$ process are also measured [@Ablikim:2014kxa]. These new experimental information in Ref. [@Ablikim:2014kxa] allows us to further perform a comprehensive study of $e^+e^- \to \psi(3770) \to p \bar p \pi^0$, which stimulates our interest to study the contribution of excited nucleon resonances ($N^*$) to $e^+e^- \to \psi(3770)\to p \bar p \pi^0$ and $\psi(3770)$ production from $p \bar p \to \psi(3770) \pi^0$ reaction. The nucleon is the simplest system in which the three colors of QCD can combine to form a colorless object, thus it is important to understand the internal quark-gluon structure of the nucleon and its excited $N^*$ states, and the study of excited $N^*$ states is an interested research field of hadron physics [@Klempt:2009pi], which can make our knowledge of hadron spectrum abundant. A very important source of information for the nucleon internal structure is the $N^*$ mass spectrum as well as its various production and decay rates, while the charmonium decay into $p \bar{p} \pi^0$ is an ideal platform to study excited $N^*$ nucleon resonances, because it provides an effective isospin 1/2 filter for the $\pi N$ system due to isospin conservation [@Ablikim:2004ug; @Barnes:2011pm; @Ablikim:2012zk]. In this work, we introduce excited $N^*$ nucleon resonances in the process of $e^+e^- \to \psi(3770) \to p \bar p \pi^0$. By fitting the $p\pi^0$ and $\bar{p}\pi^0$ invariant mass distributions of the cross section of $e^+e^-\to p\bar{p}\pi^0$, we extract the information of couplings of $N^* N \pi$ and $\psi(3770) N^* \bar{N}$, which not only reflects the inner features of discussed $N^*$, but also helps us to learn the role played by $N^*$ in the $e^+e^- \to \psi(3770)\to p \bar p \pi^0$. Based on our studying on the $e^+e^- \to \psi(3770) \to p \bar{p} \pi^0$ process, we move forward to study the $p \bar p \to \psi(3770) \pi^0$ reaction, which is due to the cross relation between the $\psi(3770) \to p \bar{p} \pi^0$ decay and the $p \bar p\to \psi(3770) \pi^0$ reaction [@Barnes:2011pm]. Here, these extracted parameters from our study of $e^+e^- \to \psi(3770)\to p \bar p \pi^0$ will be employed to estimate the production rate of $p \bar p\to \psi(3770)\pi^0$ and relevant features. We calculate the total and differential cross sections of the $p \bar{p} \to \psi(3770) \pi^0$ reaction. It is shown that the contribution of nucleon pole to this reaction is the largest close to the reaction threshold. However, the interference between nucleon pole and the other nucleon resonance affects significantly and could change the angle distributions clearly. Additionally, there were abundant experimental data of $\psi(3686)\to p\bar{p} \pi^0$ given by BESIII [@Ablikim:2012zk], where BESIII released the branching ratio $B(\psi(3686)\to p \bar p \pi^0) = (1.65 \pm 0.03 \pm 0.15) \times 10^{-4}$ and the measured $p \pi^0$ and $\bar p \pi^0$ invariant mass spectra [@Ablikim:2012zk]. This experimental status related to $\psi(3686)$ makes us extend the above study to the $\psi(3686)\to p \bar{p}\pi^0$ decay, and also the $p \bar p \to \psi(3686) \pi^0$ reaction. Our studies provide valuable information to future experimental exploration of $\psi(3770)$ and $\psi(3686)$ productions plus a pion through the $p \bar p$ interaction at $\overline{\mbox{P}}$ANDA. This paper is organized as follows. After introduction in Sec. \[sec1\], we present the detailed study of $e^+e^- \to p\bar{p}\pi^0$ by including the excited $N^*$ nucleon resonances (see Sec. \[sec2\]). In Sec. \[sec3\], we further calculate $p\bar p\to \psi(3770)\pi^0$ by combining with these results obtained in Sec. \[sec2\]. In Sec. \[sec4\], we adopt the similar approach to study $\psi(3686)\to p\bar{p}\pi^0$ decay and the $p\bar p\to \psi(3686)\pi^0$ process. The paper ends with a discussion and conclusion. Excited $N^*$ nucleon resonance contributions to $e^+e^- \to \psi(3770)\to p\bar{p}\pi^0$ {#sec2} ========================================================================================= First, we study the process $e^+e^- \to p \bar{p} \pi^0$ with an effective Lagrangian approach. In hadron level, the process $e^+e^- \to p \bar{p} \pi^0$ in the vicinity of $\psi(3770)$ is described by the diagrams shown in Fig. \[dec\]. In Fig. \[dec\] (a), $e^+$ and $e^-$ annihilate into photon, which couples with charmonium $\psi(3770)$. And then, $\psi(3770)$ interacts with final states, where we consider the contributions from nucleon-pole ($ \equiv P_{11}$) with $J^{P} = \frac{1}{2}^+$ and five $N^*$ states that are well established [@Agashe:2014kda]: $N(1440)$ ($ \equiv P_{11}$) with $J^{P}=\frac{1}{2}^+$, $N(1520)$ ($ \equiv D_{13}$) with $J^{P}=\frac{3}{2}^-$, $N(1535)$ ($ \equiv S_{11}$) with $J^{P}=\frac{1}{2}^-$, $N(1650)$ ($ \equiv S_{11}$) with $J^{P}=\frac{1}{2}^-$, and $N(1720)$ ($ \equiv P_{13}$) with $J^{P}=\frac{3}{2}^+$. Additionally, we also consider the background contribution, where the $e^+e^-$ annihilation directly into $p \bar{p} \pi$ without intermediate $\psi(3770)$, which is shown in Fig. \[dec\] (b). ![(color online). The Feynman diagrams for the process $e^+e^- \to p \bar{p} \pi^0$ in the vicinity of $\psi(3770)$.[]{data-label="dec"}](d2.eps "fig:") ![(color online). The Feynman diagrams for the process $e^+e^- \to p \bar{p} \pi^0$ in the vicinity of $\psi(3770)$.[]{data-label="dec"}](d2b.eps "fig:") To compute the contributions of these terms, we use the effective interaction Lagrangian densities for each vertex. For the $\gamma \psi(3770)$ coupling, we adopt the vector meson dominant (VMD) model, where a vector meson couples to a photon is described by [@Lin:2013mka] $$\begin{aligned} \mathcal {L}_{V\gamma} = - \frac{e M^2_V }{f_V} V_\mu A^\mu.\end{aligned}$$ In above expression, $M_V$ and $f_V$ are the mass and the decay constant of the vector meson, respectively. The decay constant $e/f_V$ can be fitted through $V \rightarrow e^+ e^-$: $$\begin{aligned} e/f_V = \left[\frac{ 3 \Gamma_{V \rightarrow e^+ e^- } M_V^2} { 8 \alpha |\vec{p}|^3}\right]^{1/2} \simeq \left[\frac{3\Gamma_{V\to e^+ e^-}}{\alpha M_V}\right]^{1/2} \ ,\end{aligned}$$ where $|\vec{p}|=(M_V^2-4m_e^2)^{1/2}/2\simeq M_V/2$ is three momentum of an electron in the rest frame of the vector meson. $\alpha = e^2/(4\pi) = 1/137$. Using $B(\psi(3770)\to e^+e^-)=(9.6\pm 0.7)\times 10^{-6}$ [@Agashe:2014kda], we obtain $e/f_{\psi(3770)}= 0.0053$. The $J/\psi N \bar{N}$ and $N N\pi$ couplings are described by: $$\begin{aligned} \label{npi} \mathcal {L}_{\pi NN}&=&-\frac{g_{\pi NN}}{2m_N}\bar{N}\gamma_5\gamma_\mu\tau\cdot\partial^\mu\pi N , \label{pinn} \\ \mathcal {L}_{\psi NN}&=&-g_{\psi NN}\bar{N}\gamma_\mu V^\mu N, \label{vnn}\end{aligned}$$ where $V^\mu$ stands for the vector field of $\psi(3770)$. We take $g_{\pi NN} = 13.45$. For the $N^* N \pi$ and $\psi N^* \bar N$ vertexes, we adopt the Lagrangian densities as used in Refs. [@Tsushima:1996xc; @Tsushima:1998jz; @Ouyang:2009kv; @Wu:2009md; @Cao:2010km; @Cao:2010ji; @Zou:2002yy]: $$\begin{aligned} \mathcal {L}_{\pi NP_{11}}&=&-\frac{g_{\pi NP_{11}}}{2m_N}\bar{N}\gamma_5\gamma_\mu\tau\cdot\partial^\mu\pi R_{P_{11}}+h.c. ,\label{nspi} \\ \mathcal {L}_{\pi NS_{11}}&=&-g_{\pi NS_{11}}\bar{N}\tau\cdot\pi R_{S_{11}}+h.c., \\ \mathcal {L}_{\pi NP_{13}}&=&-\frac{g_{\pi NP_{13}}}{m_N}\bar{N}\tau\cdot\partial_\mu\pi R^\mu_{P_{13}}+h.c., \\ \mathcal {L}_{\pi ND_{13}}&=&-\frac{g_{\pi ND_{13}}}{m_N^2}\bar{N}\gamma_5\gamma^\mu\tau\cdot\partial_\mu\partial_\nu\pi R^\nu_{D_{13}}+h.c., \\ \mathcal {L}_{\psi NP_{11}}&=&-g_{\psi NP_{11}}\bar{N}\gamma_\mu V^\mu R_{P_{11}}+h.c. , \\ \mathcal {L}_{\psi NS_{11}}&=&-g_{\psi NS_{11}}\bar{N}\gamma_5 \gamma_\mu V^\mu R_{S_{11}}+h.c., \\ \mathcal {L}_{\psi NP_{13}}&=&-ig_{\psi NP_{13}}\bar{N}\gamma_5 V_\mu R^\mu_{P_{13}}+h.c., \\ \mathcal {L}_{\psi ND_{13}}&=&-g_{\psi ND_{13}}\bar{N}V_\mu R^\mu_{D_{13}}+h.c. ,\label{nsv}\end{aligned}$$ where $R$ is a $N^*$ field. For the intermediate nucleon-pole or $N^*$ state, a Breit-Wigner form of its propagator $G_J (q)$ can be written as [@Huang:2005js] $$\begin{aligned} \label{p12} G_{\frac{1}{2}}(q)=i\frac{\slashed{q}+M_{N^*}}{q^2-M^2_{N^*}+iM_{N^*} \Gamma_{N^*} }\end{aligned}$$ for $J = \frac{1}{2}$, and $$\begin{aligned} \label{p32} G^{\mu\nu}_{\frac{3}{2}}(q) &=& i\frac{\slashed{q}+M_{N^*}}{q^2-M^2_{N^*}+iM_{N^*}\Gamma_{N^*} }\Bigg(-g_{\mu\nu}+\frac{1}{3}\gamma_\mu\gamma_\nu \nonumber \\ & &+\frac{1}{3m}(\gamma_\mu q_\nu-\gamma_\nu q_\mu)+\frac{2}{3}\frac{q_\mu q_\nu}{q^2}\Bigg)\end{aligned}$$ for $J=\frac{3}{2}$. In Eqs. (\[p12\]) and (\[p32\]), $M_{N^*}$ and $\Gamma_{N^*}$ are the masses and widths of these intermediate $N^*$ states, respectively. The values used in the present work for $M_{N^*}$ and $\Gamma_{N^*}$ are summarized in Table. \[table:Nstar\]. $N^*$ $M_{N^*}$ (MeV) $\Gamma_{N^*}$ (MeV) ----------- ----------------- ---------------------- $N(938)$ $938$ $0$ $N(1440)$ $1430$ $350$ $N(1520)$ $1515$ $115$ $N(1535)$ $1535$ $150$ $N(1650)$ $1655$ $140$ $N(1720)$ $1720$ $250$ : Relevant resonant parameters for $N^*$ states. The values are taken from Particle Data Book [@Agashe:2014kda].[]{data-label="table:Nstar"} On the other hand, we also need to introduce the form factors for these intermediate off-shell $N^*$ ($N$), which are taken as in Refs. [@Feuster:1997pq; @Haberzettl:1998eq; @Yoshimoto:1999dr; @Oh:2000zi]: $$\begin{aligned} \label{ff} F(q^2)=\frac{\Lambda^4}{\Lambda^4 + (q^2-M_{N^*}^2)^2},\end{aligned}$$ where the cutoff parameter $\Lambda$ can be parameterized as $$\begin{aligned} \label{cut} \Lambda = M_{N^*} + \beta \Lambda_{QCD},\end{aligned}$$ with $\Lambda_{QCD} = 220$ MeV. The parameter $\beta$ will be determined by fitting the experimental data. For the background contribution depicted in Fig. \[dec\] (b), we construct the amplitude in analogy of Ref. [@Chen:2010nv]: $$\begin{aligned} \label{directamplitude} \mathcal{M}_{NoR}=g_{NoR} \bar{v}(k_2) e \gamma^\mu u(k_1) \frac{1}{s} \bar{u}(p_2) \gamma^\mu \gamma_5 v(p_3) \mathcal{F}_{NoR}(s),\end{aligned}$$ with $\mathcal{F}_{NoR}(s) = \textnormal{exp} ( -a ( \sqrt{s} - \sum_f m_f )^2)$, where $ \sum_f m_f$ means the mass of the final states are summed over. The parameter $a$ will be fitted to the experimental measurements, and $s$ is the invariant mass square of the $e^+ e^-$ system. In the phenomenological Lagrangian approaches, the relative phases between amplitudes from different diagrams are not fixed. Generally, we should introduce a relative phase between different amplitudes as free parameters, and the total amplitude can be written as: $$\begin{aligned} \label{mpppi} &&\mathcal{M}_{e^+ e^- \rightarrow p \bar{p} \pi^0} \nonumber \\ &=& \mathcal{M}_{NoR}e^{i\phi_{NoR}} + \bar{v}(k_2) e \gamma^\mu u(k_1) \frac{-g^{\mu\nu}}{s} e m^2_{\psi} /f_{\psi} \nonumber \\ && \times \frac{ -g_{\nu\alpha}+\frac{p_{\psi\nu} p_{\psi\alpha}}{m^2_\psi} } {s-m^2_\psi+im_\psi\Gamma_\psi}\left(\mathcal{M}_{N}^\alpha+\sum_{N^*}\mathcal{M}_{N^*}^\alpha e^{i\phi_{N^*}}\right) ,\end{aligned}$$ where $\mathcal{M}_{N^*(N)}^\alpha$ describing the subprocesses $\psi(3770) \to p\bar{p}\pi^0$ are given completely in appendix. The differential cross section is given by [@Xie:2015zga] $$\begin{aligned} \label{gpppi} d\sigma_{e^+ e^- \rightarrow p \bar{p} \pi^0}=\frac{(2\pi)^4 \sum |\mathcal{M}_{e^+ e^- \rightarrow p \bar{p} \pi^0}|^2}{4\sqrt{(k_1 \cdot k_2)^2}}d\Phi_3,\end{aligned}$$ and the phase space factor is given by $$\begin{aligned} d\Phi_3=\frac{1}{(2\pi)^9}\frac{1}{8\sqrt{s}}|\vec p^*_3||\vec p_2|d\Omega^*_3 d\Omega_2 dm_{\bar{p}\pi},\end{aligned}$$ with $\sum |{\cal M}|^2$ averaging over the spins of the initial $e^+e^-$ and summing over the polarizations of the final states $p \bar{p}$. As we can see in the appendix, in the tree-level approximation, only the products like $g_{N^*} \equiv g_{V N N^*}g_{\pi N N^*}$ enter in the invariant amplitudes. They are determined with the use of MINUIT, by fitting to the low energy experimental data on mass distribution of $e^+e^- \to p\bar{p} \pi^0$ at $\sqrt{s} = 3.773$ GeV [@Ablikim:2014kxa]. So far we have fifteen unknown parameters: six $g_{N^*}$, six phase angles $\phi_{N^*}$ and $\phi_{NoR}$, one cutoff $\beta$ in the form factors and two parameters $g_{NoR}$ and $a$ in direct production amplitude Eq. (\[directamplitude\]). We perform those fifteen-parameter $\chi^2$ fits to the BESIII experiment data on the invariant mass distribution at $3.773$ GeV below $1.8$ GeV, and make use of the total cross section information in Ref. [@Ablikim:2014kxa]. Here, we do not consider the invariant mass region beyond $1.8$ GeV, where contains large contribution from higher mass $N^*$ states and other complicated resonance which decays to $p\bar{p}$. In Ref. [@Wiele:2013vla], it was pointed that in the case of $\bar{p} p \to \pi^0 J/\psi$ reaction the higher mass $N^*$ resonances are needed. Indeed, in the present case, if we go beyond $1.8$ GeV, we need also the higher mass $N^*$ states. On the other hand, we did also another calculation including the contributions of higher spin nuclear excited states, $N(1675)5/2^-$ and $N(1680)5/2^+$. It is find that their contributions are quite small and the fitted parameters for the other nuclear resonance are little changed. Thus, we will not include the contributions of this two states in this work. We get a minimal $\chi^2/dof = 1.03$ with the fitted cut-off parameter $\beta =6.2 \pm 3.5$. The parameters appearing in direct amplitude Eq. (\[directamplitude\]) are $g_{NoR}=0.45\pm0.02$, $\phi_{NoR}=4.84\pm0.20$ Rad and $a=0.84\pm0.02$. The other fitted parameters are compiled in Table \[ngphi\]. The fitted results are shown in Fig. \[figdec2\] compared with the experimental data taken from Ref. [@Ablikim:2014kxa], where the green dashed line stands for the background contribution, the orange doted line stands for the nucleon-pole contribution, the red line is the full result, and other lines show the contributions from different $N^*$ resonances. Notice that we have converted the experimental event to physical differential cross section using the experimental value $\sigma_{total}=7.71$ pb at $3.773$ GeV [@Ablikim:2014kxa]. Our results can describe the two clear peaks around 1.5 GeV and 1.7 GeV, thanks to the contributions from $N(1520)$, $N(1535)$ and $N(1650)$ resonances. The contribution from the nucleon pole is small, while the background contribution is quite large. In Fig. \[figdec2\], it is interesting to see large interfering effects between different contributions. At the low $M_{\bar{p}\pi}$ region around $1.1-1.3$ GeV, large cancelation between the nucleon pole and the background leads to quiet suppressed spectrum, and the bump structure from the nucleon pole just disappears. From the two-peak region around $1.4-1.8$ GeV, we can directly see that, the background contribution plus $N^*$ contribution (means without interfering contribution) is not able to reach the data peak, it indicates a large enhancement between the background contribution and $N^*$ contribution thanks to the interfering effect. . \[ngphi\] $N^*$ $g_{N^*}\,(\times10^{-3})$ $\phi_{N^*} \,(\textnormal{rad})$ ----------- ---------------------------- ----------------------------------- $N(938)$ $8.00 \pm 0.46$ $--$ $N(1440)$ $1.92 \pm 0.98$ $6.09\pm0.38$ $N(1520)$ $0.28 \pm 0.24$ $3.74\pm1.07$ $N(1535)$ $1.74 \pm 1.34$ $2.99\pm0.67$ $N(1650)$ $1.99 \pm 0.18$ $2.17\pm0.19$ $N(1720)$ $1.14 \pm 0.63$ $6.02\pm0.71$ : The fitted parameters in the process $e^+e^- \to p \bar{p}\pi^0$, where $g_{N^*} = g_{\psi(3770)NN^*}g_{\pi N N^*}$. For nucleon, $g_{N}$ is defined as $g_{N}=g_{\psi(3770)NN}g_{\pi N N}$ ![(color online). The fitted mass spectrum of the process $e^+e^-\rightarrow p\bar{p}\pi^0$ at $\sqrt{s} = 3.773$ GeV comparing to the experiment data. The experiment data are taken from Ref. [@Ablikim:2014kxa]. The green dashed line stands for the background contribution, the orange doted line stands for the nucleon-pole contribution, the red line is the full result, and other lines show the contributions from different $N^*$ resonances. Notice that the experimental event is converted to physical differential cross section using the experimental total cross section at $\sqrt{s} = 3.773$ GeV [@Ablikim:2014kxa].[]{data-label="figdec2"}](fit2.eps) The $\psi(3770)$ production in the process $p\bar p\to \psi(3770)\pi^0$ {#sec3} ======================================================================= A charmonium plus a light meson $\pi$ can produced by the low energy $p\bar p$ annihilation process. The tree level diagrams for the $p\bar p \to \psi(3770) \pi^0$ reaction are depicted in Fig. \[cro\]. It is worth to mention that the effect of the $N^*$ resonances in the cross channel of Fig. \[cro\] has been studied firstly in the $\bar{p} p \to \pi^0 J/\psi$ reaction [@Wiele:2013vla]. It was found that the contributions from the $N^*$ resonances in the $\bar{p} p \to \pi^0 J/\psi$ reaction are important. In the present work, we extend the model of Ref. [@Wiele:2013vla] to the process of the higher charmonium states \[$\psi(3770)$ and $\psi(3686)$\] production. [^2] ![(color online). The typical Feynman diagrams for the process $p \bar p \to \pi^0 \psi(3770)$.[]{data-label="cro"}](d1.eps) The differential cross section of the $p \bar p \rightarrow \pi^0 \psi(3770)$ reaction at center of mass (c.m.) frame can be expressed as [@Agashe:2014kda] $$\begin{aligned} \label{gcro} \frac{d\sigma_{p \bar p \to \pi^0 \psi(3770)}}{dcos\theta} =\frac{1}{32 \pi s}\frac{|\vec{p}^{~\rm cm}_{3}|}{|\vec{p}^{~\rm cm}_{1}|} \sum \overline{|\mathcal{M}|^2},\end{aligned}$$ where $\theta$ denotes the angle of the outgoing $\pi^0$ relative to beam direction in the $\rm c.m.$ frame, $\vec{p}^{~\rm cm}_{1}$ and $\vec{p}^{~\rm cm}_3$ are the three-momentum of the proton and $\psi(3770)$ in c.m. frame, respectively, while the total invariant scattering amplitude $\cal M$ is given in appendix using cross symmetry. With the parameters determined from the process of $e^+e^- \to p \bar p \pi^0$, we calculate the total and differential cross sections of $p\bar{p} \to \pi^0 \psi(3770)$ reaction. In Fig. \[figcros21\], we show our results for the total cross section of the $p\bar{p} \to \pi^0 \psi(3770)$ reaction as a function of the invariant mass ($E_{\rm cm}$) of $\bar p p$ system. At $E_{\rm cm} = 5.26$ GeV, the total cross section is $0.056$ nb, and it is under the upper limit of the value obtained in Ref. [@Ablikim:2014kxa]. ![(color online). Total cross section of the $p \bar p \to \pi^0 \psi(3770)$ reaction. The black line is total result, and other lines show the contributions from different $N^*$ resonances.[]{data-label="figcros21"}](cross21.eps) From Fig. \[figcros21\], we see that the nucleon pole gives largest contribution, and becomes dominant in the region $E_{\rm cm} > 5.0$ GeV. This is because in the reaction of $ p \bar p \to \psi(3770) \pi^0$, the four momentum square, $q^2$, of nucleon or other nucleon resonance is smaller than 0, and the propagator $\frac{1}{q^2 - M^2}$ will increase the contribution of nucleon because of its small mass. Besides, it is found that the contributions from $N^*$ state with different quantum numbers have quite different behavior. The contributions from $N(1535)$ and $N(1650)$ with $J^{P} = \frac{1}{2}^-$ decrease at $E_{\rm cm}$ around $4.2$ GeV, while the others increase all the time. Overall, the total cross section become quiet flat while $E_{\rm cm}>4.1$ GeV. In addition, we also calculate the angular distribution of the $p \bar p \rightarrow \pi^0 \psi(3770)$ reaction at $E_{\rm cm} = 4.0$, $4.25$, $4.5$, $4.75$ and $5.0$ GeV. The numerical results are shown in Fig. \[figangle1\]. We can see that there emerges an obvious peak at the backward angles (around ${\rm cos}\theta \sim -0.8$) at $E_{\rm cm} \ge 4.25$ GeV produced by the contributions of nucleon results in the $u$-channel, while the larger results at the forward angles is due to the $t$-channel nucleon resonances contributions. ![(color online). Angular distributions of the $p \bar p \rightarrow \pi^0 \psi(3770)$ reaction with full contribution.[]{data-label="figangle1"}](angle1.eps) In Fig. \[figangle1n\], we show the numerical results of the angular distributions by only considering the contribution from the nucleon pole. We can see that the angular distributions are symmetry between the backward and forward angles. Comparing Fig. \[figangle1\] with Fig. \[figangle1n\], we see that, there is a big difference between the full contribution and the only nucleon contribution. Our model predictions may be tested by the future experiments. ![(color online). Angular distributions of the $p \bar p \rightarrow \pi^0 \psi(3770)$ reaction considered only the contribution from the nucleon pole.[]{data-label="figangle1n"}](angle1n.eps) Note that the exchanged nuclear resonances in Fig. \[cro\] are far off mass shell, and the form factors for exchanged nuclear resonances here should be different with those that have been used for the $e^-e^+\to \psi(3770) \to p \bar{p} \pi^0$ reaction. We know that the form factors can be directly related to the hadron structure. However, the question of hadron structure is still very open, we have to adjust the form factor to fit the experimental data, and the hadronic form factors are commonly used phenomenologically [@Feuster:1997pq; @Haberzettl:1998eq; @Yoshimoto:1999dr; @Oh:2000zi]. The effects of these form factors could substantially change the predicted cross sections. Because of the lack of the available experimental measurements, we can not determine the form factors without ambiguities. In the present work, we take the same form factors for both $\bar{p} p \to \psi(3770) \pi^0$ reaction and $e^+ e^- \to \psi(3770) \to \bar{p} p \pi^0$ reaction. The implication for $\psi(3686)\to p\bar{p}\pi^0$ and $p\bar p\to \psi(3686)\pi^0$ {#sec4} ================================================================================== For the process $\psi(3686)\to p\bar{p}\pi^0$, we first determine the coupling constant $g_{\psi(3686)NN}$, i.e., by using the Lagrangian in Eq. (\[vnn\]), $g_{\psi(3686)NN}$ can be fitted through the process $\psi(3686)\rightarrow p\bar{p}$. With the experimental value [@Agashe:2014kda] $B(\psi(3686)\rightarrow p\bar{p}$)$=2.8\times10^{-4}$, $g_{\psi(3686)NN}$ is determined to be $$\begin{aligned} \label{g3686nn} g_{\psi(3686)NN} = 9.4 \times 10^{-4},\end{aligned}$$ which is consistent with that given in Ref. [@Barnes:2006ck]. ![(color online). The fitted $p \pi$ invariant mass spectrum of the process of $\psi(3686) \to p\bar{p}\pi^0$. The dashed green curve stands for the contribution of the nucleon pole, the solid red line stands for the full contributions, and other lines show the contributions from different $N^*$ resonances. The experiment data are taken from Ref. [@Ablikim:2012zk].[]{data-label="figdec"}](fit1.eps) In Ref. [@Ablikim:2012zk], BESIII released the $p\pi$ invariant mass spectrum of the process $\psi(3686) \to p\bar{p}\pi^0$ and decay width $\Gamma(\psi(3686)\rightarrow p\bar{p}\pi^0)=(1.65\pm0.03\pm0.15)\times10^{-5}$. Similar to the case of $\psi(3770)$, we fit five coupling constants $g_{N^*}$, five phase angles and a cut off parameter $\beta$ to the experimental data. The fitted results are shown in Fig. \[figdec\]. Here, one gets $\chi^2/d.o.f=2.90$ and $\beta=3.28 \pm 2.23$, while the fitted coupling constants $g_{N^*}$ and phase angles are listed in Table \[ngphi2\]. . \[ngphi2\] $N^*$ $g_{N^*}\,(\times10^{-3})$ $\phi_{N^*}\,(\textnormal{rad})$ ----------- ---------------------------- ---------------------------------- $N(1440)$ $5.10 \pm 0.86$ $3.40 \pm 0.22$ $N(1520)$ $2.27 \pm 0.39$ $4.96 \pm 1.10$ $N(1535)$ $0.51 \pm 0.36$ $0.75 \pm 0.64$ $N(1650)$ $0.76 \pm 0.19$ $5.35 \pm 0.92$ $N(1720)$ $0.98 \pm 0.42$ $1.77 \pm 0.99$ : Fitted coupling constants $g_{N^*}$ and phase angles $\phi_{N^*}$ in the process $\psi(3686) \to p\bar{p}\pi^0$, where $g_{N^*} = g_{\psi(3686)NN^*}g_{\pi N N^*}$ In Fig. \[figdec\], the dashed curve stands for the contribution of the nucleon pole, the solid line stands for the full contributions, and other lines show the contributions from different $N^*$ resonances. We see that we can describe the experimental data fairly well. Furthermore, we find that the peak between 1.6 GeV and 1.7 GeV mainly comes from the contribution of $N(1650)$. There also exist quiet obvious interfering effects between different $N^*$ contributions in Fig. \[figdec\]. Close to $M_{p \pi} = 1.6$ GeV, comparing the $N(1440)$ contribution to the total contribution, one can see the $N(1440)$ contribution is “digged out” a valley by other $N^*$ contributions. In the region of $M_{p \pi} > 1.7$ GeV, the total contribution is smaller than the $N(1440)$ contribution, i.e., the total contribution is suppressed by interfering terms. So, from Fig. \[figdec2\] and Fig. \[figdec\], one can see how important the interference effect is. We will not be able to get a good fit without interfering terms and arbitrary phase angles. Additionally, we also calculated the branch fractions of $\psi(3686)\rightarrow (N^*\bar{p}+c.c.)\rightarrow p\bar{p}\pi^0$ from individual intermediate $N^*$ (or $p$) state, with the fitted coupling constants listed in the Table \[ngphi2\]. Our results are shown in Table \[bf\]. The errors of our theoretical results are obtained from the errors of those fitted coupling constants of $g_{N^*}$. We also notice that in Ref. [@Ablikim:2012zk] BESIII also extracted the corresponding branching fractions without considering the interference of different intermediate $N^*$ (or $p$) states, which is different from the treatment in the present work. Thus, in Table \[bf\] we further compare our result with the experimental results [@Ablikim:2012zk], we see that our results are in agreement within errors with that given in Ref. [@Ablikim:2012zk]. Our results The results in Ref. [@Ablikim:2012zk] ----------- --------------- --------------------------------------- $N$ $7.5$ $6.42^{+0.20+1.78}_{-0.20-1.28}$ $N(1440)$ $14 \pm 4.5$ $3.58^{+0.25+1.59}_{-0.25-0.84}$ $N(1520)$ $2.8 \pm 0.8$ $0.64^{+0.05+0.22}_{-0.05-0.17}$ $N(1535)$ $2.1 \pm 3.0$ $2.47^{+0.28+0.99}_{-0.28-0.97}$ $N(1650)$ $4.9 \pm 2.5$ $3.76^{+0.28+1.37}_{-0.28-1.66}$ $N(1720)$ $1.3 \pm 1.0$ $1.79^{+0.10+0.24}_{-0.10-0.71}$ : The calculated branching fractions $\psi(3686) \rightarrow p\bar{p}\pi^0$ if considering individual intermediate $N^*$ (or $N$) contribution, and the comparison with the experiment values of Ref. [@Ablikim:2012zk]. Here, all values are in the unit of $10^{-5}$.[]{data-label="bf"} ![(color online). The cross section of the process $p \bar p \rightarrow \pi^0 \psi(3686)$. The black line is total result, and other lines shows the $N^*$ contribution.[]{data-label="figcros11"}](cross11.eps) With these fitted parameters, we calculate the cross section of the process $p \bar p \rightarrow \pi^0 \psi(3686)$ with cross symmetry. The results are shown in Fig. \[figcros11\]. One can see that the nucleon pole contribution is predominant in the whole energy region, while the contributions from other $N^*$ states are small. In the higher energy region, the nucleon pole contribution is starting to decrease, while the full contribution increases slowly, this behavior resembles the process $p \bar p \rightarrow \pi^0 \psi(3770)$. Furthermore, it is noticed that, the discrepancy between the total result and the nucleon contribution is smaller than the case of $p \bar p \to \pi^0 \psi(3770)$. Finally, we show the angular distributions of the process $p \bar p \rightarrow \pi^0 \psi(3686)$ in Figs. \[figangle2\] and \[figangle2n\]. Similar to Fig. \[figangle1\], there is a peak in backward angle and a valley close to $\cos\theta=0$. Comparing to the angular distribution with the nucleon contribution in Fig. \[figangle2n\], there exits obvious difference, since the nucleon contribution only is symmetry while total contribution is asymmetry. ![(color online). the angular distribution of the process $p \bar p \rightarrow \pi^0 \psi(3686)$. Each line shows the different c.m. energy.[]{data-label="figangle2"}](angle2.eps) ![(color online). the angular distribution of the process $p \bar p \rightarrow \pi^0 \psi(3686)$ only considering the nucleon contribution. Each line shows the different c.m. energy.[]{data-label="figangle2n"}](angle2n.eps) discussion and conclusion {#sec5} ========================= We have studied the $e^+e^- \to p \bar{p} \pi^0$ at $3.773$ GeV c.m. energy and $p \bar p \to \pi^0 \psi(3770)$ reaction within an effective lagrangian approach. The $e^+e^- \to p \bar{p} \pi^0$ process is a good platform to study excited $N^*$ nucleon resonances. We consider contributions from nucleon pole and five well established $N^*$ states. First, we perform a $\chi^2$-fit to the experimental data on the mass distribution of the $e^+e^- \to p \bar{p} \pi^0$ , from where we obtain the couplings of $\psi(3770)$ to these $N^*$ states. It is shown that we can describe the experimental data quite well. In particular, the two bumps around $1.5$ and $1.7$ GeV can be well reproduced. We also find that the contribution of the nucleon pole is small comparing to the background contribution, and there exists large cancellation in low $M_{\bar{p}\pi}$ region. Second, based on our results of the $e^+e^- \to p \bar{p} \pi^0$, we study the $p \bar p \to \pi^0 \psi(3770)$ reaction with cross symmetry. We evaluate the total and differential cross sections of the $p \bar p \to \pi^0 \psi(3770)$ reaction. The nucleon pole gives largest contribution to the $p \bar p \to \pi^0 \psi(3770)$ reaction close to threshold. However, the interference terms between nucleon pole and the other nucleon resonance affects significantly and could change the angle distributions clearly. Our studies provide valuable information to future experimental exploration the $\psi(3770) \pi^0$ production through the $p \bar p$ interaction. Additionally, we also study the $\psi(3686)$ production through the process $p \bar p \to \pi^0 \psi(3686)$. Similarly to the case of $e^+e^- \to \psi(3770) \to p \bar{p} \pi^0$, we study firstly the decay process of $\psi(3686) \to p \bar{p} \pi^0$ to extract the parameters we needed. Then we study the $p \bar p \to \pi^0 \psi(3686)$ reaction. We find that the contribution from the nucleon pole is dominant, while the angular distributions show a quite discrepancy induced by the $N^*$ states. We hope and expect that future experiments at $\overline{\mbox{P}}$ANDA will provide a test to our model and give more constraints on our theoretical study. Acknowledgments {#acknowledgments .unnumbered} =============== This project is supported by the National Natural Science Foundation of China under Grants No. 11222547, No. 11175073, and No. 11475227, the Ministry of Education of China (SRFDP under Grant No. 2012021111000), and the Fok Ying Tung Education Foundation (Grant No. 131006). This work is also supported by the Open Project Program of State Key Laboratory of Theoretical Physics, Institute of Theoretical Physics, Chinese Academy of Sciences, China (No.Y5KF151CJ1). Appendix: Scattering amplitudes of the subprocess $\psi(3770) \to p\bar{p}\pi^0$ and the process $p \bar p \rightarrow \pi^0 \psi(3770)$. {#appendix-scattering-amplitudes-of-the-subprocess-psi3770-to-pbarppi0-and-the-process-p-bar-p-rightarrow-pi0-psi3770. .unnumbered} =========================================================================================================================================== The tree level diagrams of the subprocess $\psi(3770) \to p \bar{p}\pi^0$ is depicted in part of Fig. \[dec\] (a). According to the feynmann diagrams shown in Fig. \[dec\], the scattering amplitudes $\mathcal{M}_{J^P}$ with a exchanged $N^*(J^P)$ (including $N$) are given by: $$\begin{aligned} \mathcal{M}_{\frac{1}{2}^+} &=& \frac{g_{\pi NP_{11}}}{2m_N} g_{VNP_{11}} \epsilon^\mu(p_1) \overline{u}(p_2) \left[\gamma_\mu \frac{-\slashed{p_t}+m_{N^*}}{t-m_{N^*}^2} \gamma_5 (i\slashed{p_4}) F(t) + \gamma_5 (i\slashed{p_4}) \frac{\slashed{p_u} + m_{N^*}}{u-m_{N^*}^2} \gamma_\mu F(u) \right] v(p_3), \label{amp0} \\ \mathcal{M}_{\frac{1}{2}^-} &=& g_{\pi NS_{11}} g_{VNS_{11}} \epsilon^\mu(p_1) \overline{u}(p_2) \left[\gamma_5 \gamma_\mu \frac{-\slashed{p_t}+m_{N^*}}{t-m_{N^*}^2} F(t) + \frac{\slashed{p_u} +m_{N^*}}{u-m_{N^*}^2} \gamma_5 \gamma_\mu F(u) \right] v(p_3), \\ \mathcal{M}_{\frac{3}{2}^+} &=& \frac{g_{\pi NP_{13}}}{m_N} i g_{VNP_{13}} \epsilon^\mu(p_1) \overline{u}(p_2) \left[\gamma_5 \frac{-\slashed{p_t}+m_{N^*}}{t-m_{N^*}^2} G_{\mu\nu}(-p_t) (i p^\nu_4) F(t) + (i p^\nu_4) \frac{\slashed{p_u} +m_{N^*}}{u-m_{N^*}^2} G_{\nu\mu}(p_u) \gamma_5 F(u) \right] v(p_3), \\ \mathcal{M}_{\frac{3}{2}^-}&=& \frac{g_{\pi ND_{13}}}{m_N^2} g_{VND_{13}} \epsilon^\mu(p_1) \overline{u}(p_2) \left[\frac{-\slashed{p_t}+m_{N^*}}{t-m_{N^*}^2} G_{\mu\nu}(-p_t) \gamma_5 (i \slashed{p_4}) (i p^\nu_4) F(t) + \gamma_5 (i \slashed{p_4}) (i p^\nu_4) \frac{\slashed{p_u} +m_{N^*}}{u-m_{N^*}^2} G_{\nu\mu}(p_u) F(u) \right] v(p_3)\label{amp1}\end{aligned}$$ with $t = p^2_t = (p_1-p_2)^2$ and $u = p^2_u = (p_1 - p_3)^2$, and $G_{\mu\nu}$ is $$\begin{aligned} G_{\mu\nu}(p) = (-g_{\mu\nu}+\frac{1}{3} \gamma_\mu \gamma_\nu + \frac{1}{3m_{N^*}}(\gamma_\mu p_\nu- \gamma_\nu p_\mu) + \frac{2}{3} \frac{p_\mu p_\nu}{m_{N^*}^2}).\end{aligned}$$ For $\mathcal{M}_{N^*(N)}^\alpha$ in Eq. (\[mpppi\]), just drop the polarization vector $\epsilon^\mu(p_1)$. For $p \bar p \to \pi^0 \psi(3770)$ reaction, the scattering amplitudes can be easily obtained just applying the substitution to Eqs. (\[amp0\]-\[amp1\]): $$\begin{aligned} p_1\rightarrow-p_3, \quad p_2\rightarrow-p_2, \quad p_3\rightarrow-p_1, \quad p_t\rightarrow -p_u.\end{aligned}$$ The amplitude of the process $\psi(3686) \to p\bar{p}\pi^0$ and $p \bar p \rightarrow \pi^0 \psi(3686)$ is exactly the same. [99]{} M. F. M. Lutz [*et al.*]{} \[PANDA Collaboration\], arXiv:0903.3905 \[hep-ex\]. M. K. Gaillard, L. Maiani and R. Petronzio, Phys. Lett. B [**110**]{}, 489 (1982). A. Lundborg, T. Barnes and U. Wiedner, Phys. Rev. D [**73**]{}, 096003 (2006). T. Barnes and X. Li, Phys. Rev. D [**75**]{}, 054018 (2007). T. Barnes, X. Li and W. Roberts, Phys. Rev. D [**77**]{}, 056001 (2008). T. Barnes, X. Li and W. Roberts, Phys. Rev. D [**81**]{}, 034025 (2010). Q. Y. Lin, H. S. Xu and X. Liu, Phys. Rev. D [**86**]{}, 034007 (2012). B. Pire, K. Semenov-Tian-Shansky and L. Szymanowski, Phys. Lett. B [**724**]{}, 99 (2013). J. Van de Wiele and S. Ong, Eur. Phys. J. C [**73**]{}, 2640 (2013). M. Ablikim [*et al.*]{} \[BESIII Collaboration\], Phys. Rev. D [**90**]{}, 032007 (2014). E. Klempt and J. M. Richard, Rev. Mod. Phys.  [**82**]{}, 1095 (2010). M. Ablikim [*et al.*]{} \[BES Collaboration\], Phys. Rev. Lett.  [**97**]{}, 062001 (2006). T. Barnes, Int. J. Mod. Phys. Conf. Ser.  [**02**]{}, 193 (2011). M. Ablikim [*et al.*]{} \[BESIII Collaboration\], Phys. Rev. Lett.  [**110**]{}, 022001 (2013). K. A. Olive [*et al.*]{} \[Particle Data Group Collaboration\], Chin. Phys. C [**38**]{}, 090001 (2014). Q. Y. Lin, X. Liu and H. S. Xu, Phys. Rev. D [**88**]{}, 114009 (2013). K. Tsushima, A. Sibirtsev and A. W. Thomas, Phys. Lett. B [**390**]{}, 29 (1997). K. Tsushima, A. Sibirtsev, A. W. Thomas and G. Q. Li, Phys. Rev. C [**59**]{} (1999) 369 \[Phys. Rev. C [**61**]{} (2000) 029903\]. B. S. Zou and F. Hussain, Phys. Rev. C [**67**]{}, 015204 (2003). Z. Ouyang, J. J. Xie, B. S. Zou and H. S. Xu, Int. J. Mod. Phys. E [**18**]{}, 281 (2009). J. J. Wu, Z. Ouyang and B. S. Zou, Phys. Rev. C [**80**]{}, 045211 (2009). X. Cao, B. S. Zou and H. S. Xu, Phys. Rev. C [**81**]{}, 065201 (2010). X. Cao, B. S. Zou and H. S. Xu, Nucl. Phys. A [**861**]{} (2011) 23. S. Z. Huang, P. F. Zhang, T. N. Ruan, Y. C. Zhu and Z. P. Zheng, Eur. Phys. J. C [**42**]{}, 375 (2005). T. Feuster and U. Mosel, Phys. Rev. C [**58**]{}, 457 (1998). H. Haberzettl, C. Bennhold, T. Mart and T. Feuster, Phys. Rev. C [**58**]{} (1998) 40. T. Yoshimoto, T. Sato, M. Arima and T. S. H. Lee, Phys. Rev. C [**61**]{}, 065203 (2000). Y. s. Oh, A. I. Titov and T. S. H. Lee, Phys. Rev. C [**63**]{}, 025201 (2001). D. Y. Chen, J. He and X. Liu, Phys. Rev. D [**83**]{}, 054021 (2011) \[arXiv:1012.5362 \[hep-ph\]\]. J. J. Xie, Y. B. Dong and X. Cao, Phys. Rev. D [**92**]{}, 034029 (2015). [^1]: Corresponding author [^2]: We mention that the Regge exchange may be important, unfortunately, the information of the Regge propagators are scarce and we hope we can include the Regge contribution in the future.

ArXiv